[PATCH v3 00/28] vfio/xe: Add driver variant for Xe VF migration

[PATCH v3 00/28] vfio/xe: Add driver variant for Xe VF migration

[Michał Winiarski]

Hi,

This is a third round of patches introducing support for Xe SR-IOV VF
migration.
Thanks for all the review feedback.
On VFIO side, the biggest changes are related to addition of
state_mutex, pci_iov_get_pf_drvdata usage and PCI driver match (where
we're now using specific device ids).
There's also interface change - PF struct xe_device is now used across
the entire API between xe-vfio-pci and xe (excluding the wrapper around
pci_iov_get_pf_drvdata).
On Xe side, the biggest changes are related to PF control state machine,
where SAVE and RESTORE states are now symmetric and individual
resource handling is now tracked outside of control state machine.

Full changelog can be found below.

Cover letter from the previous revision:

Xe is a DRM driver supporting Intel GPUs and for SR-IOV capable
devices, it enables the creation of SR-IOV VFs.
This series adds xe-vfio-pci driver variant that interacts with Xe
driver to control VF device state and read/write migration data,
allowing it to extend regular vfio-pci functionality with VFIO migration
capability.
The driver doesn't expose PRE_COPY support, as currently supported
hardware lacks the capability to track dirty pages.

While Xe driver already had the capability to manage VF device state,
management of migration data was something that needed to be implemented
and constitutes the majority of the series.

The migration data is processed asynchronously by the Xe driver, and is
organized into multiple migration data packet types representing the
hardware interfaces of the device (GGTT / MMIO / GuC FW / VRAM).
Since the VRAM can potentially be larger than available system memory,
it is copied in multiple chunks. The metadata needed for migration
compatibility decisions is added as part of descriptor packet (currently
limited to PCI device ID / revision).
Xe driver abstracts away the internals of packet processing and takes
care of tracking the position within individual packets.
The API exported to VFIO is similar to API exported by VFIO to
userspace, a simple .read()/.write().

Note that some of the VF resources are not virtualized (e.g. GGTT - the
GFX device global virtual address space). This means that the VF driver
needs to be aware that migration has occurred in order to properly
relocate (patching or reemiting data that contains references to GGTT
addresses) before resuming operation.
The code to handle that is already present in upstream Linux and in
production VF drivers for other OSes.

Links to previous revisions for reference.
v1:
https://lore.kernel.org/lkml/20251011193847.1836454-1-michal.winiarski@intel.com/
v2:
https://lore.kernel.org/lkml/20251021224133.577765-1-michal.winiarski@intel.com/

v2 -> v3:
* Bind xe-vfio-pci to specific devices instead of using vendor and
  class (Christoph Hellwig / Jason Gunthorpe)
* Don't refer to the driver as "vendor specific" (Christoph)
* Use pci_iov_get_pf_drvdata and change the interface to take xe_device
  (Jason)
* Update the RUNNING_P2P comment (Jason / Kevin Tian)
* Add state_mutex to protect device state transitions (Kevin)
* Implement .error_detected (Kevin)
* Drop redundant comments (Kevin)
* Explain 1-based indexing and wait_flr_done (Kevin)
* Add a missing get_file() (Kevin)
* Drop redundant state transitions when p2p is supported (Kevin)
* Update run/stop naming to match other drivers (Kevin)
* Fix error state handling (Kevin)
* Fix SAVE state diagram rendering (Michał Wajdeczko)
* Control state machine flipping PROCESS / WAIT logic (Michał Wajdeczko)
* Drop GUC / GGTT / MMIO / VRAM from SAVE control state machine
* Use devm instead of drmm for migration-related allocations (Michał)
* Use GGTT node for size calculations (Michał)
* Use mutex guards consistently (Michał)
* Fix build break on 32-bit (lkp)
* Kernel-doc updates (Michał)
* And other, more minor changes

v1 -> v2:
* Do not require debug flag to support migration on PTL/BMG
* Fix PCI class match on VFIO side
* Reorganized PF Control state machine (Michał Wajdeczko)
* Kerneldoc tidying (Michał Wajdeczko)
* Return NULL instead of -ENODATA for produce/consume (Michał Wajdeczko)
* guc_buf s/sync/sync_read (Matt Brost)
* Squash patch 03 (Matt Brost)
* Assert on PM ref instead of taking it (Matt Brost)
* Remove CCS completely (Matt Brost)
* Return ptr on guc_buf_sync_read (Michał Wajdeczko)
* Define default guc_buf size (Michał Wajdeczko)
* Drop CONFIG_PCI_IOV=n stubs where not needed (Michał Wajdeczko)
* And other, more minor changes

Lukasz Laguna (2):
  drm/xe/pf: Add helper to retrieve VF's LMEM object
  drm/xe/migrate: Add function to copy of VRAM data in chunks

Michał Winiarski (26):
  drm/xe/pf: Remove GuC version check for migration support
  drm/xe: Move migration support to device-level struct
  drm/xe/pf: Convert control state to bitmap
  drm/xe/pf: Add save/restore control state stubs and connect to debugfs
  drm/xe/pf: Add data structures and handlers for migration rings
  drm/xe/pf: Add helpers for migration data allocation / free
  drm/xe/pf: Add support for encap/decap of bitstream to/from packet
  drm/xe/pf: Add minimalistic migration descriptor
  drm/xe/pf: Expose VF migration data size over debugfs
  drm/xe: Add sa/guc_buf_cache sync interface
  drm/xe: Allow the caller to pass guc_buf_cache size
  drm/xe/pf: Increase PF GuC Buffer Cache size and use it for VF
    migration
  drm/xe/pf: Remove GuC migration data save/restore from GT debugfs
  drm/xe/pf: Don't save GuC VF migration data on pause
  drm/xe/pf: Switch VF migration GuC save/restore to struct migration
    data
  drm/xe/pf: Handle GuC migration data as part of PF control
  drm/xe/pf: Add helpers for VF GGTT migration data handling
  drm/xe/pf: Handle GGTT migration data as part of PF control
  drm/xe/pf: Handle MMIO migration data as part of PF control
  drm/xe/pf: Handle VRAM migration data as part of PF control
  drm/xe/pf: Add wait helper for VF FLR
  drm/xe/pf: Enable SR-IOV VF migration
  drm/xe/pci: Introduce a helper to allow VF access to PF xe_device
  drm/xe/pf: Export helpers for VFIO
  drm/intel/pciids: Add match with VFIO override
  vfio/xe: Add device specific vfio_pci driver variant for Intel
    graphics

 MAINTAINERS                                   |   7 +
 drivers/gpu/drm/xe/Makefile                   |   4 +
 drivers/gpu/drm/xe/xe_ggtt.c                  | 104 ++
 drivers/gpu/drm/xe/xe_ggtt.h                  |   4 +
 drivers/gpu/drm/xe/xe_ggtt_types.h            |   2 +
 drivers/gpu/drm/xe/xe_gt_sriov_pf.h           |   2 +
 drivers/gpu/drm/xe/xe_gt_sriov_pf_config.c    |  78 ++
 drivers/gpu/drm/xe/xe_gt_sriov_pf_config.h    |   6 +
 drivers/gpu/drm/xe/xe_gt_sriov_pf_control.c   | 599 ++++++++++-
 drivers/gpu/drm/xe/xe_gt_sriov_pf_control.h   |   8 +
 .../gpu/drm/xe/xe_gt_sriov_pf_control_types.h |  33 +-
 drivers/gpu/drm/xe/xe_gt_sriov_pf_debugfs.c   |  47 -
 drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.c | 985 ++++++++++++++----
 drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.h |  50 +-
 .../drm/xe/xe_gt_sriov_pf_migration_types.h   |  34 +-
 drivers/gpu/drm/xe/xe_gt_sriov_pf_types.h     |   5 +-
 drivers/gpu/drm/xe/xe_guc.c                   |  12 +-
 drivers/gpu/drm/xe/xe_guc_buf.c               |  57 +-
 drivers/gpu/drm/xe/xe_guc_buf.h               |   2 +
 drivers/gpu/drm/xe/xe_migrate.c               | 128 ++-
 drivers/gpu/drm/xe/xe_migrate.h               |   8 +
 drivers/gpu/drm/xe/xe_pci.c                   |  17 +
 drivers/gpu/drm/xe/xe_pci.h                   |   3 +
 drivers/gpu/drm/xe/xe_sa.c                    |  21 +
 drivers/gpu/drm/xe/xe_sa.h                    |   1 +
 drivers/gpu/drm/xe/xe_sriov_migration_data.c  | 550 ++++++++++
 drivers/gpu/drm/xe/xe_sriov_migration_data.h  |  37 +
 drivers/gpu/drm/xe/xe_sriov_pf.c              |   5 +
 drivers/gpu/drm/xe/xe_sriov_pf_control.c      | 128 +++
 drivers/gpu/drm/xe/xe_sriov_pf_control.h      |   5 +
 drivers/gpu/drm/xe/xe_sriov_pf_debugfs.c      | 102 ++
 drivers/gpu/drm/xe/xe_sriov_pf_migration.c    | 275 +++++
 drivers/gpu/drm/xe/xe_sriov_pf_migration.h    |  24 +
 .../gpu/drm/xe/xe_sriov_pf_migration_types.h  |  67 ++
 drivers/gpu/drm/xe/xe_sriov_pf_types.h        |   9 +
 drivers/gpu/drm/xe/xe_sriov_vfio.c            | 248 +++++
 drivers/vfio/pci/Kconfig                      |   2 +
 drivers/vfio/pci/Makefile                     |   2 +
 drivers/vfio/pci/xe/Kconfig                   |  12 +
 drivers/vfio/pci/xe/Makefile                  |   3 +
 drivers/vfio/pci/xe/main.c                    | 552 ++++++++++
 include/drm/intel/pciids.h                    |   7 +
 include/drm/intel/xe_sriov_vfio.h             |  30 +
 43 files changed, 3948 insertions(+), 327 deletions(-)
 create mode 100644 drivers/gpu/drm/xe/xe_sriov_migration_data.c
 create mode 100644 drivers/gpu/drm/xe/xe_sriov_migration_data.h
 create mode 100644 drivers/gpu/drm/xe/xe_sriov_pf_migration.c
 create mode 100644 drivers/gpu/drm/xe/xe_sriov_pf_migration.h
 create mode 100644 drivers/gpu/drm/xe/xe_sriov_pf_migration_types.h
 create mode 100644 drivers/gpu/drm/xe/xe_sriov_vfio.c
 create mode 100644 drivers/vfio/pci/xe/Kconfig
 create mode 100644 drivers/vfio/pci/xe/Makefile
 create mode 100644 drivers/vfio/pci/xe/main.c
 create mode 100644 include/drm/intel/xe_sriov_vfio.h

-- 
2.50.1

[PATCH v3 01/28] drm/xe/pf: Remove GuC version check for migration support

[Michał Winiarski]

Since commit 4eb0aab6e4434 ("drm/xe/guc: Bump minimum required GuC
version to v70.29.2"), the minimum GuC version required by the driver
is v70.29.2, which should already include everything that we need for
migration.
Remove the version check.

Suggested-by: Michal Wajdeczko <michal.wajdeczko@intel.com>
Signed-off-by: Michał Winiarski <michal.winiarski@intel.com>
Reviewed-by: Michal Wajdeczko <michal.wajdeczko@intel.com>
---
 drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.c | 3 ---
 1 file changed, 3 deletions(-)

diff --git a/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.c b/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.c
index 44cc612b0a752..a5bf327ef8889 100644
--- a/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.c
+++ b/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.c
@@ -384,9 +384,6 @@ ssize_t xe_gt_sriov_pf_migration_write_guc_state(struct xe_gt *gt, unsigned int
 
 static bool pf_check_migration_support(struct xe_gt *gt)
 {
-	/* GuC 70.25 with save/restore v2 is required */
-	xe_gt_assert(gt, GUC_FIRMWARE_VER(&gt->uc.guc) >= MAKE_GUC_VER(70, 25, 0));
-
 	/* XXX: for now this is for feature enabling only */
 	return IS_ENABLED(CONFIG_DRM_XE_DEBUG);
 }
-- 
2.50.1

[PATCH v3 02/28] drm/xe: Move migration support to device-level struct

[Michał Winiarski]

Upcoming changes will allow users to control VF state and obtain its
migration data with a device-level granularity (not tile/gt).
Change the data structures to reflect that and move the GT-level
migration init to happen after device-level init.

Signed-off-by: Michał Winiarski <michal.winiarski@intel.com>
Reviewed-by: Michal Wajdeczko <michal.wajdeczko@intel.com>
---
 drivers/gpu/drm/xe/Makefile                   |  1 +
 drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.c | 12 +-----
 .../drm/xe/xe_gt_sriov_pf_migration_types.h   |  3 --
 drivers/gpu/drm/xe/xe_sriov_pf.c              |  5 +++
 drivers/gpu/drm/xe/xe_sriov_pf_migration.c    | 41 +++++++++++++++++++
 drivers/gpu/drm/xe/xe_sriov_pf_migration.h    | 16 ++++++++
 .../gpu/drm/xe/xe_sriov_pf_migration_types.h  |  0
 drivers/gpu/drm/xe/xe_sriov_pf_types.h        |  6 +++
 8 files changed, 71 insertions(+), 13 deletions(-)
 create mode 100644 drivers/gpu/drm/xe/xe_sriov_pf_migration.c
 create mode 100644 drivers/gpu/drm/xe/xe_sriov_pf_migration.h
 create mode 100644 drivers/gpu/drm/xe/xe_sriov_pf_migration_types.h

diff --git a/drivers/gpu/drm/xe/Makefile b/drivers/gpu/drm/xe/Makefile
index 82c6b3d296769..89e5b26c27975 100644
--- a/drivers/gpu/drm/xe/Makefile
+++ b/drivers/gpu/drm/xe/Makefile
@@ -176,6 +176,7 @@ xe-$(CONFIG_PCI_IOV) += \
 	xe_sriov_pf.o \
 	xe_sriov_pf_control.o \
 	xe_sriov_pf_debugfs.o \
+	xe_sriov_pf_migration.o \
 	xe_sriov_pf_provision.o \
 	xe_sriov_pf_service.o \
 	xe_tile_sriov_pf_debugfs.o
diff --git a/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.c b/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.c
index a5bf327ef8889..ca28f45aaf481 100644
--- a/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.c
+++ b/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.c
@@ -13,6 +13,7 @@
 #include "xe_guc.h"
 #include "xe_guc_ct.h"
 #include "xe_sriov.h"
+#include "xe_sriov_pf_migration.h"
 
 /* Return: number of dwords saved/restored/required or a negative error code on failure */
 static int guc_action_vf_save_restore(struct xe_guc *guc, u32 vfid, u32 opcode,
@@ -115,8 +116,7 @@ static int pf_send_guc_restore_vf_state(struct xe_gt *gt, unsigned int vfid,
 
 static bool pf_migration_supported(struct xe_gt *gt)
 {
-	xe_gt_assert(gt, IS_SRIOV_PF(gt_to_xe(gt)));
-	return gt->sriov.pf.migration.supported;
+	return xe_sriov_pf_migration_supported(gt_to_xe(gt));
 }
 
 static struct mutex *pf_migration_mutex(struct xe_gt *gt)
@@ -382,12 +382,6 @@ ssize_t xe_gt_sriov_pf_migration_write_guc_state(struct xe_gt *gt, unsigned int
 }
 #endif /* CONFIG_DEBUG_FS */
 
-static bool pf_check_migration_support(struct xe_gt *gt)
-{
-	/* XXX: for now this is for feature enabling only */
-	return IS_ENABLED(CONFIG_DRM_XE_DEBUG);
-}
-
 /**
  * xe_gt_sriov_pf_migration_init() - Initialize support for VF migration.
  * @gt: the &xe_gt
@@ -403,8 +397,6 @@ int xe_gt_sriov_pf_migration_init(struct xe_gt *gt)
 
 	xe_gt_assert(gt, IS_SRIOV_PF(xe));
 
-	gt->sriov.pf.migration.supported = pf_check_migration_support(gt);
-
 	if (!pf_migration_supported(gt))
 		return 0;
 
diff --git a/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration_types.h b/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration_types.h
index 1f3110b6d44fa..9d672feac5f04 100644
--- a/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration_types.h
+++ b/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration_types.h
@@ -30,9 +30,6 @@ struct xe_gt_sriov_state_snapshot {
  * Used by the PF driver to maintain non-VF specific per-GT data.
  */
 struct xe_gt_sriov_pf_migration {
-	/** @supported: indicates whether the feature is supported */
-	bool supported;
-
 	/** @snapshot_lock: protects all VFs snapshots */
 	struct mutex snapshot_lock;
 };
diff --git a/drivers/gpu/drm/xe/xe_sriov_pf.c b/drivers/gpu/drm/xe/xe_sriov_pf.c
index bc1ab9ee31d92..95743c7af8050 100644
--- a/drivers/gpu/drm/xe/xe_sriov_pf.c
+++ b/drivers/gpu/drm/xe/xe_sriov_pf.c
@@ -15,6 +15,7 @@
 #include "xe_sriov.h"
 #include "xe_sriov_pf.h"
 #include "xe_sriov_pf_helpers.h"
+#include "xe_sriov_pf_migration.h"
 #include "xe_sriov_pf_service.h"
 #include "xe_sriov_printk.h"
 
@@ -101,6 +102,10 @@ int xe_sriov_pf_init_early(struct xe_device *xe)
 	if (err)
 		return err;
 
+	err = xe_sriov_pf_migration_init(xe);
+	if (err)
+		return err;
+
 	xe_sriov_pf_service_init(xe);
 
 	return 0;
diff --git a/drivers/gpu/drm/xe/xe_sriov_pf_migration.c b/drivers/gpu/drm/xe/xe_sriov_pf_migration.c
new file mode 100644
index 0000000000000..8c523c392f98b
--- /dev/null
+++ b/drivers/gpu/drm/xe/xe_sriov_pf_migration.c
@@ -0,0 +1,41 @@
+// SPDX-License-Identifier: MIT
+/*
+ * Copyright © 2025 Intel Corporation
+ */
+
+#include "xe_sriov.h"
+#include "xe_sriov_pf_migration.h"
+
+/**
+ * xe_sriov_pf_migration_supported() - Check if SR-IOV VF migration is supported by the device
+ * @xe: the &xe_device
+ *
+ * Return: true if migration is supported, false otherwise
+ */
+bool xe_sriov_pf_migration_supported(struct xe_device *xe)
+{
+	xe_assert(xe, IS_SRIOV_PF(xe));
+
+	return xe->sriov.pf.migration.supported;
+}
+
+static bool pf_check_migration_support(struct xe_device *xe)
+{
+	/* XXX: for now this is for feature enabling only */
+	return IS_ENABLED(CONFIG_DRM_XE_DEBUG);
+}
+
+/**
+ * xe_sriov_pf_migration_init() - Initialize support for SR-IOV VF migration.
+ * @xe: the &xe_device
+ *
+ * Return: 0 on success or a negative error code on failure.
+ */
+int xe_sriov_pf_migration_init(struct xe_device *xe)
+{
+	xe_assert(xe, IS_SRIOV_PF(xe));
+
+	xe->sriov.pf.migration.supported = pf_check_migration_support(xe);
+
+	return 0;
+}
diff --git a/drivers/gpu/drm/xe/xe_sriov_pf_migration.h b/drivers/gpu/drm/xe/xe_sriov_pf_migration.h
new file mode 100644
index 0000000000000..d2b4a24165438
--- /dev/null
+++ b/drivers/gpu/drm/xe/xe_sriov_pf_migration.h
@@ -0,0 +1,16 @@
+/* SPDX-License-Identifier: MIT */
+/*
+ * Copyright © 2025 Intel Corporation
+ */
+
+#ifndef _XE_SRIOV_PF_MIGRATION_H_
+#define _XE_SRIOV_PF_MIGRATION_H_
+
+#include <linux/types.h>
+
+struct xe_device;
+
+int xe_sriov_pf_migration_init(struct xe_device *xe);
+bool xe_sriov_pf_migration_supported(struct xe_device *xe);
+
+#endif
diff --git a/drivers/gpu/drm/xe/xe_sriov_pf_migration_types.h b/drivers/gpu/drm/xe/xe_sriov_pf_migration_types.h
new file mode 100644
index 0000000000000..e69de29bb2d1d
diff --git a/drivers/gpu/drm/xe/xe_sriov_pf_types.h b/drivers/gpu/drm/xe/xe_sriov_pf_types.h
index c753cd59aed2b..24d22afeececa 100644
--- a/drivers/gpu/drm/xe/xe_sriov_pf_types.h
+++ b/drivers/gpu/drm/xe/xe_sriov_pf_types.h
@@ -39,6 +39,12 @@ struct xe_device_pf {
 	/** @provision: device level provisioning data. */
 	struct xe_sriov_pf_provision provision;
 
+	/** @migration: device level VF migration data */
+	struct {
+		/** @migration.supported: indicates whether VF migration feature is supported */
+		bool supported;
+	} migration;
+
 	/** @service: device level service data. */
 	struct xe_sriov_pf_service service;
 
-- 
2.50.1

[PATCH v3 03/28] drm/xe/pf: Convert control state to bitmap

[Michał Winiarski]

In upcoming changes, the number of states will increase as a result of
introducing SAVE and RESTORE states.
This means that using unsigned long as underlying storage won't work on
32-bit architectures, as we'll run out of bits.
Use bitmap instead.

Reported-by: kernel test robot <lkp@intel.com>
Closes: https://lore.kernel.org/oe-kbuild-all/202510231918.XlOqymLC-lkp@intel.com/
Signed-off-by: Michał Winiarski <michal.winiarski@intel.com>
---
 drivers/gpu/drm/xe/xe_gt_sriov_pf_control.c       | 2 +-
 drivers/gpu/drm/xe/xe_gt_sriov_pf_control_types.h | 5 +++--
 2 files changed, 4 insertions(+), 3 deletions(-)

diff --git a/drivers/gpu/drm/xe/xe_gt_sriov_pf_control.c b/drivers/gpu/drm/xe/xe_gt_sriov_pf_control.c
index 9de05db1f0905..8a2577fda4198 100644
--- a/drivers/gpu/drm/xe/xe_gt_sriov_pf_control.c
+++ b/drivers/gpu/drm/xe/xe_gt_sriov_pf_control.c
@@ -225,7 +225,7 @@ static unsigned long *pf_peek_vf_state(struct xe_gt *gt, unsigned int vfid)
 {
 	struct xe_gt_sriov_control_state *cs = pf_pick_vf_control(gt, vfid);
 
-	return &cs->state;
+	return cs->state;
 }
 
 static bool pf_check_vf_state(struct xe_gt *gt, unsigned int vfid,
diff --git a/drivers/gpu/drm/xe/xe_gt_sriov_pf_control_types.h b/drivers/gpu/drm/xe/xe_gt_sriov_pf_control_types.h
index c80b7e77f1ad2..3ba6ad886c939 100644
--- a/drivers/gpu/drm/xe/xe_gt_sriov_pf_control_types.h
+++ b/drivers/gpu/drm/xe/xe_gt_sriov_pf_control_types.h
@@ -73,7 +73,8 @@ enum xe_gt_sriov_control_bits {
 	XE_GT_SRIOV_STATE_STOP_FAILED,
 	XE_GT_SRIOV_STATE_STOPPED,
 
-	XE_GT_SRIOV_STATE_MISMATCH = BITS_PER_LONG - 1,
+	XE_GT_SRIOV_STATE_MISMATCH,
+	XE_GT_SRIOV_STATE_MAX,
 };
 
 /**
@@ -83,7 +84,7 @@ enum xe_gt_sriov_control_bits {
  */
 struct xe_gt_sriov_control_state {
 	/** @state: VF state bits */
-	unsigned long state;
+	DECLARE_BITMAP(state, XE_GT_SRIOV_STATE_MAX);
 
 	/** @done: completion of async operations */
 	struct completion done;
-- 
2.50.1

[PATCH v3 04/28] drm/xe/pf: Add save/restore control state stubs and connect to debugfs

[Michał Winiarski]

The states will be used by upcoming changes to produce (in case of save)
or consume (in case of resume) the VF migration data.

Signed-off-by: Michał Winiarski <michal.winiarski@intel.com>
Reviewed-by: Michal Wajdeczko <michal.wajdeczko@intel.com>
---
 drivers/gpu/drm/xe/xe_gt_sriov_pf_control.c   | 248 ++++++++++++++++++
 drivers/gpu/drm/xe/xe_gt_sriov_pf_control.h   |   6 +
 .../gpu/drm/xe/xe_gt_sriov_pf_control_types.h |  14 +
 drivers/gpu/drm/xe/xe_sriov_pf_control.c      |  96 +++++++
 drivers/gpu/drm/xe/xe_sriov_pf_control.h      |   4 +
 drivers/gpu/drm/xe/xe_sriov_pf_debugfs.c      |  38 +++
 6 files changed, 406 insertions(+)

diff --git a/drivers/gpu/drm/xe/xe_gt_sriov_pf_control.c b/drivers/gpu/drm/xe/xe_gt_sriov_pf_control.c
index 8a2577fda4198..a571e1c02a3b4 100644
--- a/drivers/gpu/drm/xe/xe_gt_sriov_pf_control.c
+++ b/drivers/gpu/drm/xe/xe_gt_sriov_pf_control.c
@@ -184,6 +184,12 @@ static const char *control_bit_to_string(enum xe_gt_sriov_control_bits bit)
 	CASE2STR(PAUSE_SAVE_GUC);
 	CASE2STR(PAUSE_FAILED);
 	CASE2STR(PAUSED);
+	CASE2STR(SAVE_WIP);
+	CASE2STR(SAVE_FAILED);
+	CASE2STR(SAVED);
+	CASE2STR(RESTORE_WIP);
+	CASE2STR(RESTORE_FAILED);
+	CASE2STR(RESTORED);
 	CASE2STR(RESUME_WIP);
 	CASE2STR(RESUME_SEND_RESUME);
 	CASE2STR(RESUME_FAILED);
@@ -208,6 +214,8 @@ static unsigned long pf_get_default_timeout(enum xe_gt_sriov_control_bits bit)
 	case XE_GT_SRIOV_STATE_FLR_WIP:
 	case XE_GT_SRIOV_STATE_FLR_RESET_CONFIG:
 		return 5 * HZ;
+	case XE_GT_SRIOV_STATE_RESTORE_WIP:
+		return 20 * HZ;
 	default:
 		return HZ;
 	}
@@ -329,6 +337,8 @@ static void pf_exit_vf_mismatch(struct xe_gt *gt, unsigned int vfid)
 	pf_exit_vf_state(gt, vfid, XE_GT_SRIOV_STATE_PAUSE_FAILED);
 	pf_exit_vf_state(gt, vfid, XE_GT_SRIOV_STATE_RESUME_FAILED);
 	pf_exit_vf_state(gt, vfid, XE_GT_SRIOV_STATE_FLR_FAILED);
+	pf_exit_vf_state(gt, vfid, XE_GT_SRIOV_STATE_SAVE_FAILED);
+	pf_exit_vf_state(gt, vfid, XE_GT_SRIOV_STATE_RESTORE_FAILED);
 }
 
 #define pf_enter_vf_state_machine_bug(gt, vfid) ({	\
@@ -359,6 +369,8 @@ static void pf_queue_vf(struct xe_gt *gt, unsigned int vfid)
 
 static void pf_exit_vf_flr_wip(struct xe_gt *gt, unsigned int vfid);
 static void pf_exit_vf_stop_wip(struct xe_gt *gt, unsigned int vfid);
+static void pf_exit_vf_save_wip(struct xe_gt *gt, unsigned int vfid);
+static void pf_exit_vf_restore_wip(struct xe_gt *gt, unsigned int vfid);
 static void pf_exit_vf_pause_wip(struct xe_gt *gt, unsigned int vfid);
 static void pf_exit_vf_resume_wip(struct xe_gt *gt, unsigned int vfid);
 
@@ -380,6 +392,8 @@ static void pf_exit_vf_wip(struct xe_gt *gt, unsigned int vfid)
 
 		pf_exit_vf_flr_wip(gt, vfid);
 		pf_exit_vf_stop_wip(gt, vfid);
+		pf_exit_vf_save_wip(gt, vfid);
+		pf_exit_vf_restore_wip(gt, vfid);
 		pf_exit_vf_pause_wip(gt, vfid);
 		pf_exit_vf_resume_wip(gt, vfid);
 
@@ -399,6 +413,8 @@ static void pf_enter_vf_ready(struct xe_gt *gt, unsigned int vfid)
 	pf_exit_vf_state(gt, vfid, XE_GT_SRIOV_STATE_PAUSED);
 	pf_exit_vf_state(gt, vfid, XE_GT_SRIOV_STATE_STOPPED);
 	pf_exit_vf_state(gt, vfid, XE_GT_SRIOV_STATE_RESUMED);
+	pf_exit_vf_state(gt, vfid, XE_GT_SRIOV_STATE_SAVED);
+	pf_exit_vf_state(gt, vfid, XE_GT_SRIOV_STATE_RESTORED);
 	pf_exit_vf_mismatch(gt, vfid);
 	pf_exit_vf_wip(gt, vfid);
 }
@@ -675,6 +691,8 @@ static void pf_enter_vf_resumed(struct xe_gt *gt, unsigned int vfid)
 {
 	pf_enter_vf_state(gt, vfid, XE_GT_SRIOV_STATE_RESUMED);
 	pf_exit_vf_state(gt, vfid, XE_GT_SRIOV_STATE_PAUSED);
+	pf_exit_vf_state(gt, vfid, XE_GT_SRIOV_STATE_SAVED);
+	pf_exit_vf_state(gt, vfid, XE_GT_SRIOV_STATE_RESTORED);
 	pf_exit_vf_mismatch(gt, vfid);
 	pf_exit_vf_wip(gt, vfid);
 }
@@ -753,6 +771,16 @@ int xe_gt_sriov_pf_control_resume_vf(struct xe_gt *gt, unsigned int vfid)
 		return -EPERM;
 	}
 
+	if (pf_check_vf_state(gt, vfid, XE_GT_SRIOV_STATE_SAVE_WIP)) {
+		xe_gt_sriov_dbg(gt, "VF%u save is in progress!\n", vfid);
+		return -EBUSY;
+	}
+
+	if (pf_check_vf_state(gt, vfid, XE_GT_SRIOV_STATE_RESTORE_WIP)) {
+		xe_gt_sriov_dbg(gt, "VF%u restore is in progress!\n", vfid);
+		return -EBUSY;
+	}
+
 	if (!pf_enter_vf_resume_wip(gt, vfid)) {
 		xe_gt_sriov_dbg(gt, "VF%u resume already in progress!\n", vfid);
 		return -EALREADY;
@@ -776,6 +804,218 @@ int xe_gt_sriov_pf_control_resume_vf(struct xe_gt *gt, unsigned int vfid)
 	return -ECANCELED;
 }
 
+static void pf_exit_vf_save_wip(struct xe_gt *gt, unsigned int vfid)
+{
+	pf_exit_vf_state(gt, vfid, XE_GT_SRIOV_STATE_SAVE_WIP);
+}
+
+static void pf_enter_vf_saved(struct xe_gt *gt, unsigned int vfid)
+{
+	if (!pf_enter_vf_state(gt, vfid, XE_GT_SRIOV_STATE_SAVED))
+		pf_enter_vf_state_machine_bug(gt, vfid);
+
+	xe_gt_sriov_dbg(gt, "VF%u saved!\n", vfid);
+
+	pf_expect_vf_state(gt, vfid, XE_GT_SRIOV_STATE_PAUSED);
+	pf_exit_vf_mismatch(gt, vfid);
+	pf_exit_vf_wip(gt, vfid);
+}
+
+static bool pf_handle_vf_save(struct xe_gt *gt, unsigned int vfid)
+{
+	if (!pf_exit_vf_state(gt, vfid, XE_GT_SRIOV_STATE_SAVE_WIP))
+		return false;
+
+	pf_enter_vf_saved(gt, vfid);
+
+	return true;
+}
+
+static bool pf_enter_vf_save_wip(struct xe_gt *gt, unsigned int vfid)
+{
+	if (pf_enter_vf_state(gt, vfid, XE_GT_SRIOV_STATE_SAVE_WIP)) {
+		pf_enter_vf_wip(gt, vfid);
+		pf_queue_vf(gt, vfid);
+		return true;
+	}
+
+	return false;
+}
+
+/**
+ * xe_gt_sriov_pf_control_trigger_save_vf() - Start an SR-IOV VF migration data save sequence.
+ * @gt: the &xe_gt
+ * @vfid: the VF identifier
+ *
+ * This function is for PF only.
+ *
+ * Return: 0 on success or a negative error code on failure.
+ */
+int xe_gt_sriov_pf_control_trigger_save_vf(struct xe_gt *gt, unsigned int vfid)
+{
+	if (pf_check_vf_state(gt, vfid, XE_GT_SRIOV_STATE_STOPPED)) {
+		xe_gt_sriov_dbg(gt, "VF%u is stopped!\n", vfid);
+		return -EPERM;
+	}
+
+	if (!pf_check_vf_state(gt, vfid, XE_GT_SRIOV_STATE_PAUSED)) {
+		xe_gt_sriov_dbg(gt, "VF%u is not paused!\n", vfid);
+		return -EPERM;
+	}
+
+	if (pf_check_vf_state(gt, vfid, XE_GT_SRIOV_STATE_RESTORE_WIP)) {
+		xe_gt_sriov_dbg(gt, "VF%u restore is in progress!\n", vfid);
+		return -EBUSY;
+	}
+
+	if (!pf_enter_vf_save_wip(gt, vfid)) {
+		xe_gt_sriov_dbg(gt, "VF%u save already in progress!\n", vfid);
+		return -EALREADY;
+	}
+
+	return 0;
+}
+
+/**
+ * xe_gt_sriov_pf_control_finish_save_vf() - Complete a VF migration data save sequence.
+ * @gt: the &xe_gt
+ * @vfid: the VF identifier
+ *
+ * This function is for PF only.
+ *
+ * Return: 0 on success or a negative error code on failure.
+ */
+int xe_gt_sriov_pf_control_finish_save_vf(struct xe_gt *gt, unsigned int vfid)
+{
+	if (!pf_expect_vf_state(gt, vfid, XE_GT_SRIOV_STATE_SAVED)) {
+		pf_enter_vf_mismatch(gt, vfid);
+		return -EIO;
+	}
+
+	pf_expect_vf_state(gt, vfid, XE_GT_SRIOV_STATE_PAUSED);
+
+	return 0;
+}
+
+static void pf_exit_vf_restore_wip(struct xe_gt *gt, unsigned int vfid)
+{
+	pf_exit_vf_state(gt, vfid, XE_GT_SRIOV_STATE_RESTORE_WIP);
+}
+
+static void pf_enter_vf_restored(struct xe_gt *gt, unsigned int vfid)
+{
+	if (!pf_enter_vf_state(gt, vfid, XE_GT_SRIOV_STATE_RESTORED))
+		pf_enter_vf_state_machine_bug(gt, vfid);
+
+	xe_gt_sriov_dbg(gt, "VF%u restored!\n", vfid);
+
+	pf_expect_vf_state(gt, vfid, XE_GT_SRIOV_STATE_PAUSED);
+	pf_exit_vf_mismatch(gt, vfid);
+	pf_exit_vf_wip(gt, vfid);
+}
+
+static bool pf_handle_vf_restore(struct xe_gt *gt, unsigned int vfid)
+{
+	if (!pf_exit_vf_state(gt, vfid, XE_GT_SRIOV_STATE_RESTORE_WIP))
+		return false;
+
+	pf_enter_vf_restored(gt, vfid);
+
+	return true;
+}
+
+static bool pf_enter_vf_restore_wip(struct xe_gt *gt, unsigned int vfid)
+{
+	if (pf_enter_vf_state(gt, vfid, XE_GT_SRIOV_STATE_RESTORE_WIP)) {
+		pf_enter_vf_wip(gt, vfid);
+		pf_queue_vf(gt, vfid);
+		return true;
+	}
+
+	return false;
+}
+
+/**
+ * xe_gt_sriov_pf_control_trigger restore_vf() - Start an SR-IOV VF migration data restore sequence.
+ * @gt: the &xe_gt
+ * @vfid: the VF identifier
+ *
+ * This function is for PF only.
+ *
+ * Return: 0 on success or a negative error code on failure.
+ */
+int xe_gt_sriov_pf_control_trigger_restore_vf(struct xe_gt *gt, unsigned int vfid)
+{
+	if (pf_check_vf_state(gt, vfid, XE_GT_SRIOV_STATE_STOPPED)) {
+		xe_gt_sriov_dbg(gt, "VF%u is stopped!\n", vfid);
+		return -EPERM;
+	}
+
+	if (!pf_check_vf_state(gt, vfid, XE_GT_SRIOV_STATE_PAUSED)) {
+		xe_gt_sriov_dbg(gt, "VF%u is not paused!\n", vfid);
+		return -EPERM;
+	}
+
+	if (pf_check_vf_state(gt, vfid, XE_GT_SRIOV_STATE_SAVE_WIP)) {
+		xe_gt_sriov_dbg(gt, "VF%u save is in progress!\n", vfid);
+		return -EBUSY;
+	}
+
+	if (!pf_enter_vf_restore_wip(gt, vfid)) {
+		xe_gt_sriov_dbg(gt, "VF%u restore already in progress!\n", vfid);
+		return -EALREADY;
+	}
+
+	return 0;
+}
+
+static int pf_wait_vf_restore_done(struct xe_gt *gt, unsigned int vfid)
+{
+	unsigned long timeout = pf_get_default_timeout(XE_GT_SRIOV_STATE_RESTORE_WIP);
+	int err;
+
+	err = pf_wait_vf_wip_done(gt, vfid, timeout);
+	if (err) {
+		xe_gt_sriov_notice(gt, "VF%u RESTORE didn't finish in %u ms (%pe)\n",
+				   vfid, jiffies_to_msecs(timeout), ERR_PTR(err));
+		return err;
+	}
+
+	if (!pf_expect_vf_not_state(gt, vfid, XE_GT_SRIOV_STATE_RESTORE_FAILED))
+		return -EIO;
+
+	return 0;
+}
+
+/**
+ * xe_gt_sriov_pf_control_finish_restore_vf() - Complete a VF migration data restore sequence.
+ * @gt: the &xe_gt
+ * @vfid: the VF identifier
+ *
+ * This function is for PF only.
+ *
+ * Return: 0 on success or a negative error code on failure.
+ */
+int xe_gt_sriov_pf_control_finish_restore_vf(struct xe_gt *gt, unsigned int vfid)
+{
+	int ret;
+
+	if (pf_check_vf_state(gt, vfid, XE_GT_SRIOV_STATE_RESTORE_WIP)) {
+		ret = pf_wait_vf_restore_done(gt, vfid);
+		if (ret)
+			return ret;
+	}
+
+	if (!pf_expect_vf_state(gt, vfid, XE_GT_SRIOV_STATE_RESTORED)) {
+		pf_enter_vf_mismatch(gt, vfid);
+		return -EIO;
+	}
+
+	pf_expect_vf_state(gt, vfid, XE_GT_SRIOV_STATE_PAUSED);
+
+	return 0;
+}
+
 /**
  * DOC: The VF STOP state machine
  *
@@ -817,6 +1057,8 @@ static void pf_enter_vf_stopped(struct xe_gt *gt, unsigned int vfid)
 
 	pf_exit_vf_state(gt, vfid, XE_GT_SRIOV_STATE_RESUMED);
 	pf_exit_vf_state(gt, vfid, XE_GT_SRIOV_STATE_PAUSED);
+	pf_exit_vf_state(gt, vfid, XE_GT_SRIOV_STATE_SAVED);
+	pf_exit_vf_state(gt, vfid, XE_GT_SRIOV_STATE_RESTORED);
 	pf_exit_vf_mismatch(gt, vfid);
 	pf_exit_vf_wip(gt, vfid);
 }
@@ -1463,6 +1705,12 @@ static bool pf_process_vf_state_machine(struct xe_gt *gt, unsigned int vfid)
 	if (pf_exit_vf_pause_save_guc(gt, vfid))
 		return true;
 
+	if (pf_handle_vf_save(gt, vfid))
+		return true;
+
+	if (pf_handle_vf_restore(gt, vfid))
+		return true;
+
 	if (pf_exit_vf_resume_send_resume(gt, vfid))
 		return true;
 
diff --git a/drivers/gpu/drm/xe/xe_gt_sriov_pf_control.h b/drivers/gpu/drm/xe/xe_gt_sriov_pf_control.h
index 8a72ef3778d47..abc233f6302ed 100644
--- a/drivers/gpu/drm/xe/xe_gt_sriov_pf_control.h
+++ b/drivers/gpu/drm/xe/xe_gt_sriov_pf_control.h
@@ -14,8 +14,14 @@ struct xe_gt;
 int xe_gt_sriov_pf_control_init(struct xe_gt *gt);
 void xe_gt_sriov_pf_control_restart(struct xe_gt *gt);
 
+bool xe_gt_sriov_pf_control_check_vf_data_wip(struct xe_gt *gt, unsigned int vfid);
+
 int xe_gt_sriov_pf_control_pause_vf(struct xe_gt *gt, unsigned int vfid);
 int xe_gt_sriov_pf_control_resume_vf(struct xe_gt *gt, unsigned int vfid);
+int xe_gt_sriov_pf_control_trigger_save_vf(struct xe_gt *gt, unsigned int vfid);
+int xe_gt_sriov_pf_control_finish_save_vf(struct xe_gt *gt, unsigned int vfid);
+int xe_gt_sriov_pf_control_trigger_restore_vf(struct xe_gt *gt, unsigned int vfid);
+int xe_gt_sriov_pf_control_finish_restore_vf(struct xe_gt *gt, unsigned int vfid);
 int xe_gt_sriov_pf_control_stop_vf(struct xe_gt *gt, unsigned int vfid);
 int xe_gt_sriov_pf_control_trigger_flr(struct xe_gt *gt, unsigned int vfid);
 int xe_gt_sriov_pf_control_sync_flr(struct xe_gt *gt, unsigned int vfid, bool sync);
diff --git a/drivers/gpu/drm/xe/xe_gt_sriov_pf_control_types.h b/drivers/gpu/drm/xe/xe_gt_sriov_pf_control_types.h
index 3ba6ad886c939..ea16ceb39dabb 100644
--- a/drivers/gpu/drm/xe/xe_gt_sriov_pf_control_types.h
+++ b/drivers/gpu/drm/xe/xe_gt_sriov_pf_control_types.h
@@ -31,6 +31,12 @@
  * @XE_GT_SRIOV_STATE_PAUSE_SAVE_GUC: indicates that the PF needs to save the VF GuC state.
  * @XE_GT_SRIOV_STATE_PAUSE_FAILED: indicates that a VF pause operation has failed.
  * @XE_GT_SRIOV_STATE_PAUSED: indicates that the VF is paused.
+ * @XE_GT_SRIOV_STATE_SAVE_WIP: indicates that VF save operation is in progress.
+ * @XE_GT_SRIOV_STATE_SAVE_FAILED: indicates that VF save operation has failed.
+ * @XE_GT_SRIOV_STATE_SAVED: indicates that VF data is saved.
+ * @XE_GT_SRIOV_STATE_RESTORE_WIP: indicates that VF restore operation is in progress.
+ * @XE_GT_SRIOV_STATE_RESTORE_FAILED: indicates that VF restore operation has failed.
+ * @XE_GT_SRIOV_STATE_RESTORED: indicates that VF data is restored.
  * @XE_GT_SRIOV_STATE_RESUME_WIP: indicates the a VF resume operation is in progress.
  * @XE_GT_SRIOV_STATE_RESUME_SEND_RESUME: indicates that the PF is about to send RESUME command.
  * @XE_GT_SRIOV_STATE_RESUME_FAILED: indicates that a VF resume operation has failed.
@@ -63,6 +69,14 @@ enum xe_gt_sriov_control_bits {
 	XE_GT_SRIOV_STATE_PAUSE_FAILED,
 	XE_GT_SRIOV_STATE_PAUSED,
 
+	XE_GT_SRIOV_STATE_SAVE_WIP,
+	XE_GT_SRIOV_STATE_SAVE_FAILED,
+	XE_GT_SRIOV_STATE_SAVED,
+
+	XE_GT_SRIOV_STATE_RESTORE_WIP,
+	XE_GT_SRIOV_STATE_RESTORE_FAILED,
+	XE_GT_SRIOV_STATE_RESTORED,
+
 	XE_GT_SRIOV_STATE_RESUME_WIP,
 	XE_GT_SRIOV_STATE_RESUME_SEND_RESUME,
 	XE_GT_SRIOV_STATE_RESUME_FAILED,
diff --git a/drivers/gpu/drm/xe/xe_sriov_pf_control.c b/drivers/gpu/drm/xe/xe_sriov_pf_control.c
index 416d00a03fbb7..8d8a01faf5291 100644
--- a/drivers/gpu/drm/xe/xe_sriov_pf_control.c
+++ b/drivers/gpu/drm/xe/xe_sriov_pf_control.c
@@ -149,3 +149,99 @@ int xe_sriov_pf_control_sync_flr(struct xe_device *xe, unsigned int vfid)
 
 	return 0;
 }
+
+/**
+ * xe_sriov_pf_control_trigger_save_vf - Start a VF migration data SAVE sequence on all GTs.
+ * @xe: the &xe_device
+ * @vfid: the VF identifier
+ *
+ * This function is for PF only.
+ *
+ * Return: 0 on success or a negative error code on failure.
+ */
+int xe_sriov_pf_control_trigger_save_vf(struct xe_device *xe, unsigned int vfid)
+{
+	struct xe_gt *gt;
+	unsigned int id;
+	int ret;
+
+	for_each_gt(gt, xe, id) {
+		ret = xe_gt_sriov_pf_control_trigger_save_vf(gt, vfid);
+		if (ret)
+			return ret;
+	}
+
+	return 0;
+}
+
+/**
+ * xe_sriov_pf_control_finish_save_vf - Complete a VF migration data SAVE sequence on all GTs.
+ * @xe: the &xe_device
+ * @vfid: the VF identifier
+ *
+ * This function is for PF only.
+ *
+ * Return: 0 on success or a negative error code on failure.
+ */
+int xe_sriov_pf_control_finish_save_vf(struct xe_device *xe, unsigned int vfid)
+{
+	struct xe_gt *gt;
+	unsigned int id;
+	int ret;
+
+	for_each_gt(gt, xe, id) {
+		ret = xe_gt_sriov_pf_control_finish_save_vf(gt, vfid);
+		if (ret)
+			break;
+	}
+
+	return ret;
+}
+
+/**
+ * xe_sriov_pf_control_trigger_restore_vf - Start a VF migration data RESTORE sequence on all GTs.
+ * @xe: the &xe_device
+ * @vfid: the VF identifier
+ *
+ * This function is for PF only.
+ *
+ * Return: 0 on success or a negative error code on failure.
+ */
+int xe_sriov_pf_control_trigger_restore_vf(struct xe_device *xe, unsigned int vfid)
+{
+	struct xe_gt *gt;
+	unsigned int id;
+	int ret;
+
+	for_each_gt(gt, xe, id) {
+		ret = xe_gt_sriov_pf_control_trigger_restore_vf(gt, vfid);
+		if (ret)
+			return ret;
+	}
+
+	return ret;
+}
+
+/**
+ * xe_sriov_pf_control_wait_restore_vf - Complete a VF migration data RESTORE sequence in all GTs.
+ * @xe: the &xe_device
+ * @vfid: the VF identifier
+ *
+ * This function is for PF only.
+ *
+ * Return: 0 on success or a negative error code on failure.
+ */
+int xe_sriov_pf_control_finish_restore_vf(struct xe_device *xe, unsigned int vfid)
+{
+	struct xe_gt *gt;
+	unsigned int id;
+	int ret;
+
+	for_each_gt(gt, xe, id) {
+		ret = xe_gt_sriov_pf_control_finish_restore_vf(gt, vfid);
+		if (ret)
+			break;
+	}
+
+	return ret;
+}
diff --git a/drivers/gpu/drm/xe/xe_sriov_pf_control.h b/drivers/gpu/drm/xe/xe_sriov_pf_control.h
index 2d52d0ac1b28f..30318c1fba34e 100644
--- a/drivers/gpu/drm/xe/xe_sriov_pf_control.h
+++ b/drivers/gpu/drm/xe/xe_sriov_pf_control.h
@@ -13,5 +13,9 @@ int xe_sriov_pf_control_resume_vf(struct xe_device *xe, unsigned int vfid);
 int xe_sriov_pf_control_stop_vf(struct xe_device *xe, unsigned int vfid);
 int xe_sriov_pf_control_reset_vf(struct xe_device *xe, unsigned int vfid);
 int xe_sriov_pf_control_sync_flr(struct xe_device *xe, unsigned int vfid);
+int xe_sriov_pf_control_trigger_save_vf(struct xe_device *xe, unsigned int vfid);
+int xe_sriov_pf_control_finish_save_vf(struct xe_device *xe, unsigned int vfid);
+int xe_sriov_pf_control_trigger_restore_vf(struct xe_device *xe, unsigned int vfid);
+int xe_sriov_pf_control_finish_restore_vf(struct xe_device *xe, unsigned int vfid);
 
 #endif
diff --git a/drivers/gpu/drm/xe/xe_sriov_pf_debugfs.c b/drivers/gpu/drm/xe/xe_sriov_pf_debugfs.c
index a81aa05c55326..e0e6340c49106 100644
--- a/drivers/gpu/drm/xe/xe_sriov_pf_debugfs.c
+++ b/drivers/gpu/drm/xe/xe_sriov_pf_debugfs.c
@@ -136,11 +136,31 @@ static void pf_populate_pf(struct xe_device *xe, struct dentry *pfdent)
  *      │   │   ├── reset
  *      │   │   ├── resume
  *      │   │   ├── stop
+ *      │   │   ├── save
+ *      │   │   ├── restore
  *      │   │   :
  *      │   ├── vf2
  *      │   │   ├── ...
  */
 
+static int from_file_read_to_vf_call(struct seq_file *s,
+				     int (*call)(struct xe_device *, unsigned int))
+{
+	struct dentry *dent = file_dentry(s->file)->d_parent;
+	struct xe_device *xe = extract_xe(dent);
+	unsigned int vfid = extract_vfid(dent);
+	int ret;
+
+	xe_pm_runtime_get(xe);
+	ret = call(xe, vfid);
+	xe_pm_runtime_put(xe);
+
+	if (ret < 0)
+		return ret;
+
+	return 0;
+}
+
 static ssize_t from_file_write_to_vf_call(struct file *file, const char __user *userbuf,
 					  size_t count, loff_t *ppos,
 					  int (*call)(struct xe_device *, unsigned int))
@@ -179,10 +199,26 @@ static ssize_t OP##_write(struct file *file, const char __user *userbuf,	\
 }										\
 DEFINE_SHOW_STORE_ATTRIBUTE(OP)
 
+#define DEFINE_VF_CONTROL_ATTRIBUTE_RW(OP)					\
+static int OP##_show(struct seq_file *s, void *unused)				\
+{										\
+	return from_file_read_to_vf_call(s,					\
+					 xe_sriov_pf_control_finish_##OP);	\
+}										\
+static ssize_t OP##_write(struct file *file, const char __user *userbuf,	\
+			  size_t count, loff_t *ppos)				\
+{										\
+	return from_file_write_to_vf_call(file, userbuf, count, ppos,		\
+					  xe_sriov_pf_control_trigger_##OP);	\
+}										\
+DEFINE_SHOW_STORE_ATTRIBUTE(OP)
+
 DEFINE_VF_CONTROL_ATTRIBUTE(pause_vf);
 DEFINE_VF_CONTROL_ATTRIBUTE(resume_vf);
 DEFINE_VF_CONTROL_ATTRIBUTE(stop_vf);
 DEFINE_VF_CONTROL_ATTRIBUTE(reset_vf);
+DEFINE_VF_CONTROL_ATTRIBUTE_RW(save_vf);
+DEFINE_VF_CONTROL_ATTRIBUTE_RW(restore_vf);
 
 static void pf_populate_vf(struct xe_device *xe, struct dentry *vfdent)
 {
@@ -190,6 +226,8 @@ static void pf_populate_vf(struct xe_device *xe, struct dentry *vfdent)
 	debugfs_create_file("resume", 0200, vfdent, xe, &resume_vf_fops);
 	debugfs_create_file("stop", 0200, vfdent, xe, &stop_vf_fops);
 	debugfs_create_file("reset", 0200, vfdent, xe, &reset_vf_fops);
+	debugfs_create_file("save", 0600, vfdent, xe, &save_vf_fops);
+	debugfs_create_file("restore", 0600, vfdent, xe, &restore_vf_fops);
 }
 
 static void pf_populate_with_tiles(struct xe_device *xe, struct dentry *dent, unsigned int vfid)
-- 
2.50.1

[PATCH v3 05/28] drm/xe/pf: Add data structures and handlers for migration rings

[Michał Winiarski]

Migration data is queued in a per-GT ptr_ring to decouple the worker
responsible for handling the data transfer from the .read() and .write()
syscalls.
Add the data structures and handlers that will be used in future
commits.

Signed-off-by: Michał Winiarski <michal.winiarski@intel.com>
---
 drivers/gpu/drm/xe/xe_gt_sriov_pf_control.c   | 271 +++++++++++++++++-
 drivers/gpu/drm/xe/xe_gt_sriov_pf_control.h   |   6 +-
 .../gpu/drm/xe/xe_gt_sriov_pf_control_types.h |  12 +
 drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.c | 183 ++++++++++++
 drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.h |  14 +
 .../drm/xe/xe_gt_sriov_pf_migration_types.h   |  11 +
 drivers/gpu/drm/xe/xe_gt_sriov_pf_types.h     |   3 +
 drivers/gpu/drm/xe/xe_sriov_pf_migration.c    | 142 +++++++++
 drivers/gpu/drm/xe/xe_sriov_pf_migration.h    |   7 +
 .../gpu/drm/xe/xe_sriov_pf_migration_types.h  |  58 ++++
 drivers/gpu/drm/xe/xe_sriov_pf_types.h        |   3 +
 11 files changed, 695 insertions(+), 15 deletions(-)

diff --git a/drivers/gpu/drm/xe/xe_gt_sriov_pf_control.c b/drivers/gpu/drm/xe/xe_gt_sriov_pf_control.c
index a571e1c02a3b4..ad9dcd3f56453 100644
--- a/drivers/gpu/drm/xe/xe_gt_sriov_pf_control.c
+++ b/drivers/gpu/drm/xe/xe_gt_sriov_pf_control.c
@@ -19,6 +19,7 @@
 #include "xe_guc_ct.h"
 #include "xe_sriov.h"
 #include "xe_sriov_pf_control.h"
+#include "xe_sriov_pf_migration.h"
 #include "xe_sriov_pf_service.h"
 #include "xe_tile.h"
 
@@ -185,9 +186,15 @@ static const char *control_bit_to_string(enum xe_gt_sriov_control_bits bit)
 	CASE2STR(PAUSE_FAILED);
 	CASE2STR(PAUSED);
 	CASE2STR(SAVE_WIP);
+	CASE2STR(SAVE_PROCESS_DATA);
+	CASE2STR(SAVE_WAIT_DATA);
+	CASE2STR(SAVE_DATA_DONE);
 	CASE2STR(SAVE_FAILED);
 	CASE2STR(SAVED);
 	CASE2STR(RESTORE_WIP);
+	CASE2STR(RESTORE_PROCESS_DATA);
+	CASE2STR(RESTORE_WAIT_DATA);
+	CASE2STR(RESTORE_DATA_DONE);
 	CASE2STR(RESTORE_FAILED);
 	CASE2STR(RESTORED);
 	CASE2STR(RESUME_WIP);
@@ -804,9 +811,50 @@ int xe_gt_sriov_pf_control_resume_vf(struct xe_gt *gt, unsigned int vfid)
 	return -ECANCELED;
 }
 
+/**
+ * DOC: The VF SAVE state machine
+ *
+ * SAVE extends the PAUSED state.
+ *
+ * The VF SAVE state machine looks like::
+ *
+ *  ....PAUSED....................................................
+ *  :                                                            :
+ *  :     (O)<---------o                                         :
+ *  :      |            \                                        :
+ *  :    save          (SAVED)    (SAVE_FAILED)                  :
+ *  :      |               ^           ^                         :
+ *  :      |               |           |                         :
+ *  :  ....V...............o...........o......SAVE_WIP.........  :
+ *  :  :   |               |           |                      :  :
+ *  :  :   |             empty         |                      :  :
+ *  :  :   |               |           |                      :  :
+ *  :  :   |               |           |                      :  :
+ *  :  :   |           DATA_DONE       |                      :  :
+ *  :  :   |               ^           |                      :  :
+ *  :  :   |               |        error                     :  :
+ *  :  :   |            no_data       /                       :  :
+ *  :  :   |              /          /                        :  :
+ *  :  :   |             /          /                         :  :
+ *  :  :   |            /          /                          :  :
+ *  :  :   o---------->PROCESS_DATA<----consume               :  :
+ *  :  :                \                      \              :  :
+ *  :  :                 \                      \             :  :
+ *  :  :                  \                      \            :  :
+ *  :  :                   ring_full----->WAIT_DATA           :  :
+ *  :  :                                                      :  :
+ *  :  :......................................................:  :
+ *  :............................................................:
+ *
+ * For the full state machine view, see `The VF state machine`_.
+ */
 static void pf_exit_vf_save_wip(struct xe_gt *gt, unsigned int vfid)
 {
-	pf_exit_vf_state(gt, vfid, XE_GT_SRIOV_STATE_SAVE_WIP);
+	if (pf_exit_vf_state(gt, vfid, XE_GT_SRIOV_STATE_SAVE_WIP)) {
+		pf_escape_vf_state(gt, vfid, XE_GT_SRIOV_STATE_SAVE_PROCESS_DATA);
+		pf_escape_vf_state(gt, vfid, XE_GT_SRIOV_STATE_SAVE_WAIT_DATA);
+		pf_escape_vf_state(gt, vfid, XE_GT_SRIOV_STATE_SAVE_DATA_DONE);
+	}
 }
 
 static void pf_enter_vf_saved(struct xe_gt *gt, unsigned int vfid)
@@ -821,19 +869,54 @@ static void pf_enter_vf_saved(struct xe_gt *gt, unsigned int vfid)
 	pf_exit_vf_wip(gt, vfid);
 }
 
+static void pf_enter_vf_save_failed(struct xe_gt *gt, unsigned int vfid)
+{
+	pf_enter_vf_state(gt, vfid, XE_GT_SRIOV_STATE_SAVE_FAILED);
+	pf_exit_vf_wip(gt, vfid);
+}
+
+static int pf_handle_vf_save_data(struct xe_gt *gt, unsigned int vfid)
+{
+	return 0;
+}
+
 static bool pf_handle_vf_save(struct xe_gt *gt, unsigned int vfid)
 {
-	if (!pf_exit_vf_state(gt, vfid, XE_GT_SRIOV_STATE_SAVE_WIP))
+	int ret;
+
+	if (!pf_exit_vf_state(gt, vfid, XE_GT_SRIOV_STATE_SAVE_PROCESS_DATA))
 		return false;
 
-	pf_enter_vf_saved(gt, vfid);
+	if (xe_gt_sriov_pf_migration_ring_full(gt, vfid)) {
+		pf_enter_vf_state(gt, vfid, XE_GT_SRIOV_STATE_SAVE_WAIT_DATA);
+
+		return true;
+	}
+
+	ret = pf_handle_vf_save_data(gt, vfid);
+	if (ret == -EAGAIN)
+		pf_enter_vf_state(gt, vfid, XE_GT_SRIOV_STATE_SAVE_PROCESS_DATA);
+	else if (ret)
+		pf_enter_vf_save_failed(gt, vfid);
+	else
+		pf_enter_vf_state(gt, vfid, XE_GT_SRIOV_STATE_SAVE_DATA_DONE);
 
 	return true;
 }
 
+static void pf_exit_vf_save_wait_data(struct xe_gt *gt, unsigned int vfid)
+{
+	if (!pf_exit_vf_state(gt, vfid, XE_GT_SRIOV_STATE_SAVE_WAIT_DATA))
+		return;
+
+	pf_enter_vf_state(gt, vfid, XE_GT_SRIOV_STATE_SAVE_PROCESS_DATA);
+	pf_queue_vf(gt, vfid);
+}
+
 static bool pf_enter_vf_save_wip(struct xe_gt *gt, unsigned int vfid)
 {
 	if (pf_enter_vf_state(gt, vfid, XE_GT_SRIOV_STATE_SAVE_WIP)) {
+		pf_enter_vf_state(gt, vfid, XE_GT_SRIOV_STATE_SAVE_PROCESS_DATA);
 		pf_enter_vf_wip(gt, vfid);
 		pf_queue_vf(gt, vfid);
 		return true;
@@ -842,6 +925,32 @@ static bool pf_enter_vf_save_wip(struct xe_gt *gt, unsigned int vfid)
 	return false;
 }
 
+/**
+ * xe_gt_sriov_pf_control_check_save_data_done() - Check if all save migration data was produced.
+ * @gt: the &xe_gt
+ * @vfid: the VF identifier
+ *
+ * This function is for PF only.
+ *
+ * Return: 0 on success or a negative error code on failure.
+ */
+bool xe_gt_sriov_pf_control_check_save_data_done(struct xe_gt *gt, unsigned int vfid)
+{
+	return pf_check_vf_state(gt, vfid, XE_GT_SRIOV_STATE_SAVE_DATA_DONE);
+}
+
+/**
+ * xe_gt_sriov_pf_control_process_save_data() - Queue VF save migration data processing.
+ * @gt: the &xe_gt
+ * @vfid: the VF identifier
+ *
+ * This function is for PF only.
+ */
+void xe_gt_sriov_pf_control_process_save_data(struct xe_gt *gt, unsigned int vfid)
+{
+	pf_exit_vf_save_wait_data(gt, vfid);
+}
+
 /**
  * xe_gt_sriov_pf_control_trigger_save_vf() - Start an SR-IOV VF migration data save sequence.
  * @gt: the &xe_gt
@@ -887,19 +996,62 @@ int xe_gt_sriov_pf_control_trigger_save_vf(struct xe_gt *gt, unsigned int vfid)
  */
 int xe_gt_sriov_pf_control_finish_save_vf(struct xe_gt *gt, unsigned int vfid)
 {
-	if (!pf_expect_vf_state(gt, vfid, XE_GT_SRIOV_STATE_SAVED)) {
-		pf_enter_vf_mismatch(gt, vfid);
+	if (!pf_check_vf_state(gt, vfid, XE_GT_SRIOV_STATE_SAVE_DATA_DONE)) {
+		xe_gt_sriov_err(gt, "VF%u save is still in progress!\n", vfid);
 		return -EIO;
 	}
 
 	pf_expect_vf_state(gt, vfid, XE_GT_SRIOV_STATE_PAUSED);
+	pf_exit_vf_state(gt, vfid, XE_GT_SRIOV_STATE_SAVE_DATA_DONE);
+	pf_enter_vf_saved(gt, vfid);
 
 	return 0;
 }
 
+/**
+ * DOC: The VF RESTORE state machine
+ *
+ * RESTORE extends the PAUSED state.
+ *
+ * The VF RESTORE state machine looks like::
+ *
+ *  ....PAUSED....................................................
+ *  :                                                            :
+ *  :     (O)<---------o                                         :
+ *  :      |            \                                        :
+ *  :    restore      (RESTORED)  (RESTORE_FAILED)               :
+ *  :      |               ^           ^                         :
+ *  :      |               |           |                         :
+ *  :  ....V...............o...........o......RESTORE_WIP......  :
+ *  :  :   |               |           |                      :  :
+ *  :  :   |             empty         |                      :  :
+ *  :  :   |               |           |                      :  :
+ *  :  :   |               |           |                      :  :
+ *  :  :   |           DATA_DONE       |                      :  :
+ *  :  :   |               ^           |                      :  :
+ *  :  :   |               |        error                     :  :
+ *  :  :   |           trailer        /                       :  :
+ *  :  :   |              /          /                        :  :
+ *  :  :   |             /          /                         :  :
+ *  :  :   |            /          /                          :  :
+ *  :  :   o---------->PROCESS_DATA<----produce               :  :
+ *  :  :                \                      \              :  :
+ *  :  :                 \                      \             :  :
+ *  :  :                  \                      \            :  :
+ *  :  :                   ring_empty---->WAIT_DATA           :  :
+ *  :  :                                                      :  :
+ *  :  :......................................................:  :
+ *  :............................................................:
+ *
+ * For the full state machine view, see `The VF state machine`_.
+ */
 static void pf_exit_vf_restore_wip(struct xe_gt *gt, unsigned int vfid)
 {
-	pf_exit_vf_state(gt, vfid, XE_GT_SRIOV_STATE_RESTORE_WIP);
+	if (pf_exit_vf_state(gt, vfid, XE_GT_SRIOV_STATE_RESTORE_WIP)) {
+		pf_escape_vf_state(gt, vfid, XE_GT_SRIOV_STATE_RESTORE_PROCESS_DATA);
+		pf_escape_vf_state(gt, vfid, XE_GT_SRIOV_STATE_RESTORE_WAIT_DATA);
+		pf_escape_vf_state(gt, vfid, XE_GT_SRIOV_STATE_RESTORE_DATA_DONE);
+	}
 }
 
 static void pf_enter_vf_restored(struct xe_gt *gt, unsigned int vfid)
@@ -914,19 +1066,61 @@ static void pf_enter_vf_restored(struct xe_gt *gt, unsigned int vfid)
 	pf_exit_vf_wip(gt, vfid);
 }
 
+static void pf_enter_vf_restore_failed(struct xe_gt *gt, unsigned int vfid)
+{
+	pf_enter_vf_state(gt, vfid, XE_GT_SRIOV_STATE_RESTORE_FAILED);
+	pf_exit_vf_wip(gt, vfid);
+}
+
+static int pf_handle_vf_restore_data(struct xe_gt *gt, unsigned int vfid)
+{
+	struct xe_sriov_migration_data *data = xe_gt_sriov_pf_migration_restore_consume(gt, vfid);
+
+	xe_gt_assert(gt, data);
+
+	xe_gt_sriov_notice(gt, "Skipping VF%u unknown data type: %d\n", vfid, data->type);
+
+	return 0;
+}
+
 static bool pf_handle_vf_restore(struct xe_gt *gt, unsigned int vfid)
 {
-	if (!pf_exit_vf_state(gt, vfid, XE_GT_SRIOV_STATE_RESTORE_WIP))
+	int ret;
+
+	if (!pf_exit_vf_state(gt, vfid, XE_GT_SRIOV_STATE_RESTORE_PROCESS_DATA))
 		return false;
 
-	pf_enter_vf_restored(gt, vfid);
+	if (xe_gt_sriov_pf_migration_ring_empty(gt, vfid)) {
+		if (pf_exit_vf_state(gt, vfid, XE_GT_SRIOV_STATE_RESTORE_DATA_DONE))
+			pf_enter_vf_restored(gt, vfid);
+		else
+			pf_enter_vf_state(gt, vfid, XE_GT_SRIOV_STATE_RESTORE_WAIT_DATA);
+
+		return true;
+	}
+
+	ret = pf_handle_vf_restore_data(gt, vfid);
+	if (ret)
+		pf_enter_vf_restore_failed(gt, vfid);
+	else
+		pf_enter_vf_state(gt, vfid, XE_GT_SRIOV_STATE_RESTORE_PROCESS_DATA);
 
 	return true;
 }
 
+static void pf_exit_vf_restore_wait_data(struct xe_gt *gt, unsigned int vfid)
+{
+	if (!pf_exit_vf_state(gt, vfid, XE_GT_SRIOV_STATE_RESTORE_WAIT_DATA))
+		return;
+
+	pf_enter_vf_state(gt, vfid, XE_GT_SRIOV_STATE_RESTORE_PROCESS_DATA);
+	pf_queue_vf(gt, vfid);
+}
+
 static bool pf_enter_vf_restore_wip(struct xe_gt *gt, unsigned int vfid)
 {
 	if (pf_enter_vf_state(gt, vfid, XE_GT_SRIOV_STATE_RESTORE_WIP)) {
+		pf_enter_vf_state(gt, vfid, XE_GT_SRIOV_STATE_RESTORE_PROCESS_DATA);
 		pf_enter_vf_wip(gt, vfid);
 		pf_queue_vf(gt, vfid);
 		return true;
@@ -935,6 +1129,47 @@ static bool pf_enter_vf_restore_wip(struct xe_gt *gt, unsigned int vfid)
 	return false;
 }
 
+/**
+ * xe_gt_sriov_pf_control_restore_data_done() - Indicate the end of VF migration data stream.
+ * @gt: the &xe_gt
+ * @vfid: the VF identifier
+ *
+ * This function is for PF only.
+ *
+ * Return: 0 on success or a negative error code on failure.
+ */
+int xe_gt_sriov_pf_control_restore_data_done(struct xe_gt *gt, unsigned int vfid)
+{
+	if (!pf_expect_vf_state(gt, vfid, XE_GT_SRIOV_STATE_RESTORE_WIP)) {
+		pf_enter_vf_state_machine_bug(gt, vfid);
+		return -EIO;
+	}
+
+	if (!pf_enter_vf_state(gt, vfid, XE_GT_SRIOV_STATE_RESTORE_DATA_DONE)) {
+		pf_enter_vf_state_machine_bug(gt, vfid);
+		return -EIO;
+	}
+
+	xe_gt_sriov_pf_control_process_restore_data(gt, vfid);
+
+	return 0;
+}
+
+/**
+ * xe_gt_sriov_pf_control_process_restore_data() - Queue VF restore migration data processing.
+ * @gt: the &xe_gt
+ * @vfid: the VF identifier
+ *
+ * This function is for PF only.
+ */
+void xe_gt_sriov_pf_control_process_restore_data(struct xe_gt *gt, unsigned int vfid)
+{
+	if (!pf_expect_vf_state(gt, vfid, XE_GT_SRIOV_STATE_RESTORE_WIP))
+		pf_enter_vf_state_machine_bug(gt, vfid);
+
+	pf_exit_vf_restore_wait_data(gt, vfid);
+}
+
 /**
  * xe_gt_sriov_pf_control_trigger restore_vf() - Start an SR-IOV VF migration data restore sequence.
  * @gt: the &xe_gt
@@ -1000,11 +1235,9 @@ int xe_gt_sriov_pf_control_finish_restore_vf(struct xe_gt *gt, unsigned int vfid
 {
 	int ret;
 
-	if (pf_check_vf_state(gt, vfid, XE_GT_SRIOV_STATE_RESTORE_WIP)) {
-		ret = pf_wait_vf_restore_done(gt, vfid);
-		if (ret)
-			return ret;
-	}
+	ret = pf_wait_vf_restore_done(gt, vfid);
+	if (ret)
+		return ret;
 
 	if (!pf_expect_vf_state(gt, vfid, XE_GT_SRIOV_STATE_RESTORED)) {
 		pf_enter_vf_mismatch(gt, vfid);
@@ -1705,9 +1938,21 @@ static bool pf_process_vf_state_machine(struct xe_gt *gt, unsigned int vfid)
 	if (pf_exit_vf_pause_save_guc(gt, vfid))
 		return true;
 
+	if (pf_check_vf_state(gt, vfid, XE_GT_SRIOV_STATE_SAVE_WAIT_DATA)) {
+		xe_gt_sriov_dbg_verbose(gt, "VF%u in %s\n", vfid,
+					control_bit_to_string(XE_GT_SRIOV_STATE_SAVE_WAIT_DATA));
+		return false;
+	}
+
 	if (pf_handle_vf_save(gt, vfid))
 		return true;
 
+	if (pf_check_vf_state(gt, vfid, XE_GT_SRIOV_STATE_RESTORE_WAIT_DATA)) {
+		xe_gt_sriov_dbg_verbose(gt, "VF%u in %s\n", vfid,
+					control_bit_to_string(XE_GT_SRIOV_STATE_RESTORE_WAIT_DATA));
+		return false;
+	}
+
 	if (pf_handle_vf_restore(gt, vfid))
 		return true;
 
diff --git a/drivers/gpu/drm/xe/xe_gt_sriov_pf_control.h b/drivers/gpu/drm/xe/xe_gt_sriov_pf_control.h
index abc233f6302ed..6b1ab339e3b73 100644
--- a/drivers/gpu/drm/xe/xe_gt_sriov_pf_control.h
+++ b/drivers/gpu/drm/xe/xe_gt_sriov_pf_control.h
@@ -14,12 +14,14 @@ struct xe_gt;
 int xe_gt_sriov_pf_control_init(struct xe_gt *gt);
 void xe_gt_sriov_pf_control_restart(struct xe_gt *gt);
 
-bool xe_gt_sriov_pf_control_check_vf_data_wip(struct xe_gt *gt, unsigned int vfid);
-
 int xe_gt_sriov_pf_control_pause_vf(struct xe_gt *gt, unsigned int vfid);
 int xe_gt_sriov_pf_control_resume_vf(struct xe_gt *gt, unsigned int vfid);
+bool xe_gt_sriov_pf_control_check_save_data_done(struct xe_gt *gt, unsigned int vfid);
+void xe_gt_sriov_pf_control_process_save_data(struct xe_gt *gt, unsigned int vfid);
 int xe_gt_sriov_pf_control_trigger_save_vf(struct xe_gt *gt, unsigned int vfid);
 int xe_gt_sriov_pf_control_finish_save_vf(struct xe_gt *gt, unsigned int vfid);
+int xe_gt_sriov_pf_control_restore_data_done(struct xe_gt *gt, unsigned int vfid);
+void xe_gt_sriov_pf_control_process_restore_data(struct xe_gt *gt, unsigned int vfid);
 int xe_gt_sriov_pf_control_trigger_restore_vf(struct xe_gt *gt, unsigned int vfid);
 int xe_gt_sriov_pf_control_finish_restore_vf(struct xe_gt *gt, unsigned int vfid);
 int xe_gt_sriov_pf_control_stop_vf(struct xe_gt *gt, unsigned int vfid);
diff --git a/drivers/gpu/drm/xe/xe_gt_sriov_pf_control_types.h b/drivers/gpu/drm/xe/xe_gt_sriov_pf_control_types.h
index ea16ceb39dabb..a14c738581ae2 100644
--- a/drivers/gpu/drm/xe/xe_gt_sriov_pf_control_types.h
+++ b/drivers/gpu/drm/xe/xe_gt_sriov_pf_control_types.h
@@ -32,9 +32,15 @@
  * @XE_GT_SRIOV_STATE_PAUSE_FAILED: indicates that a VF pause operation has failed.
  * @XE_GT_SRIOV_STATE_PAUSED: indicates that the VF is paused.
  * @XE_GT_SRIOV_STATE_SAVE_WIP: indicates that VF save operation is in progress.
+ * @XE_GT_SRIOV_STATE_SAVE_PROCESS_DATA: indicates that VF migration data is being produced.
+ * @XE_GT_SRIOV_STATE_SAVE_WAIT_DATA: indicates that PF awaits for space in migration data ring.
+ * @XE_GT_SRIOV_STATE_SAVE_DATA_DONE: indicates that all migration data was produced by Xe.
  * @XE_GT_SRIOV_STATE_SAVE_FAILED: indicates that VF save operation has failed.
  * @XE_GT_SRIOV_STATE_SAVED: indicates that VF data is saved.
  * @XE_GT_SRIOV_STATE_RESTORE_WIP: indicates that VF restore operation is in progress.
+ * @XE_GT_SRIOV_STATE_RESTORE_PROCESS_DATA: indicates that VF migration data is being consumed.
+ * @XE_GT_SRIOV_STATE_RESTORE_WAIT_DATA: indicates that PF awaits for data in migration data ring.
+ * @XE_GT_SRIOV_STATE_RESTORE_DATA_DONE: indicates that all migration data was produced by the user.
  * @XE_GT_SRIOV_STATE_RESTORE_FAILED: indicates that VF restore operation has failed.
  * @XE_GT_SRIOV_STATE_RESTORED: indicates that VF data is restored.
  * @XE_GT_SRIOV_STATE_RESUME_WIP: indicates the a VF resume operation is in progress.
@@ -70,10 +76,16 @@ enum xe_gt_sriov_control_bits {
 	XE_GT_SRIOV_STATE_PAUSED,
 
 	XE_GT_SRIOV_STATE_SAVE_WIP,
+	XE_GT_SRIOV_STATE_SAVE_PROCESS_DATA,
+	XE_GT_SRIOV_STATE_SAVE_WAIT_DATA,
+	XE_GT_SRIOV_STATE_SAVE_DATA_DONE,
 	XE_GT_SRIOV_STATE_SAVE_FAILED,
 	XE_GT_SRIOV_STATE_SAVED,
 
 	XE_GT_SRIOV_STATE_RESTORE_WIP,
+	XE_GT_SRIOV_STATE_RESTORE_PROCESS_DATA,
+	XE_GT_SRIOV_STATE_RESTORE_WAIT_DATA,
+	XE_GT_SRIOV_STATE_RESTORE_DATA_DONE,
 	XE_GT_SRIOV_STATE_RESTORE_FAILED,
 	XE_GT_SRIOV_STATE_RESTORED,
 
diff --git a/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.c b/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.c
index ca28f45aaf481..e61e6f4215864 100644
--- a/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.c
+++ b/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.c
@@ -7,6 +7,7 @@
 
 #include "abi/guc_actions_sriov_abi.h"
 #include "xe_bo.h"
+#include "xe_gt_sriov_pf_control.h"
 #include "xe_gt_sriov_pf_helpers.h"
 #include "xe_gt_sriov_pf_migration.h"
 #include "xe_gt_sriov_printk.h"
@@ -15,6 +16,17 @@
 #include "xe_sriov.h"
 #include "xe_sriov_pf_migration.h"
 
+#define XE_GT_SRIOV_PF_MIGRATION_RING_SIZE 5
+
+static struct xe_gt_sriov_migration_data *pf_pick_gt_migration(struct xe_gt *gt, unsigned int vfid)
+{
+	xe_gt_assert(gt, IS_SRIOV_PF(gt_to_xe(gt)));
+	xe_gt_assert(gt, vfid != PFID);
+	xe_gt_assert(gt, vfid <= xe_sriov_pf_get_totalvfs(gt_to_xe(gt)));
+
+	return &gt->sriov.pf.vfs[vfid].migration;
+}
+
 /* Return: number of dwords saved/restored/required or a negative error code on failure */
 static int guc_action_vf_save_restore(struct xe_guc *guc, u32 vfid, u32 opcode,
 				      u64 addr, u32 ndwords)
@@ -382,6 +394,162 @@ ssize_t xe_gt_sriov_pf_migration_write_guc_state(struct xe_gt *gt, unsigned int
 }
 #endif /* CONFIG_DEBUG_FS */
 
+/**
+ * xe_gt_sriov_pf_migration_ring_empty() - Check if a migration ring is empty.
+ * @gt: the &xe_gt
+ * @vfid: the VF identifier
+ *
+ * Return: true if the ring is empty, otherwise false.
+ */
+bool xe_gt_sriov_pf_migration_ring_empty(struct xe_gt *gt, unsigned int vfid)
+{
+	return ptr_ring_empty(&pf_pick_gt_migration(gt, vfid)->ring);
+}
+
+/**
+ * xe_gt_sriov_pf_migration_ring_full() - Check if a migration ring is full.
+ * @gt: the &xe_gt
+ * @vfid: the VF identifier
+ *
+ * Return: true if the ring is full, otherwise false.
+ */
+bool xe_gt_sriov_pf_migration_ring_full(struct xe_gt *gt, unsigned int vfid)
+{
+	return ptr_ring_full(&pf_pick_gt_migration(gt, vfid)->ring);
+}
+
+/**
+ * xe_gt_sriov_pf_migration_save_produce() - Add VF save data packet to migration ring.
+ * @gt: the &xe_gt
+ * @vfid: the VF identifier
+ * @data: &xe_sriov_migration_data packet
+ *
+ * Called by the save migration data producer (PF SR-IOV Control worker) when
+ * processing migration data.
+ * Wakes up the save migration data consumer (userspace), that is potentially
+ * waiting for data when the ring is empty.
+ *
+ * Return: 0 on success or a negative error code on failure.
+ */
+int xe_gt_sriov_pf_migration_save_produce(struct xe_gt *gt, unsigned int vfid,
+					  struct xe_sriov_migration_data *data)
+{
+	int ret;
+
+	ret = ptr_ring_produce(&pf_pick_gt_migration(gt, vfid)->ring, data);
+	if (ret)
+		return ret;
+
+	wake_up_all(xe_sriov_pf_migration_waitqueue(gt_to_xe(gt), vfid));
+
+	return 0;
+}
+
+/**
+ * xe_gt_sriov_pf_migration_restore_consume() - Get VF restore data packet from migration ring.
+ * @gt: the &xe_gt
+ * @vfid: the VF identifier
+ *
+ * Called by the restore migration data consumer (PF SR-IOV Control worker) when
+ * processing migration data.
+ * Wakes up the restore migration data producer (userspace), that is
+ * potentially waiting to add more data when the ring is full.
+ *
+ * Return: Pointer to &struct xe_sriov_migration_data on success,
+ *	   NULL if ring is empty.
+ */
+struct xe_sriov_migration_data *
+xe_gt_sriov_pf_migration_restore_consume(struct xe_gt *gt, unsigned int vfid)
+{
+	struct xe_gt_sriov_migration_data *migration = pf_pick_gt_migration(gt, vfid);
+	struct wait_queue_head *wq = xe_sriov_pf_migration_waitqueue(gt_to_xe(gt), vfid);
+	struct xe_sriov_migration_data *data;
+
+	data = ptr_ring_consume(&migration->ring);
+	if (data)
+		wake_up_all(wq);
+
+	return data;
+}
+
+/**
+ * xe_gt_sriov_pf_migration_restore_produce() - Add VF restore data packet to migration ring.
+ * @gt: the &xe_gt
+ * @vfid: the VF identifier
+ * @data: &xe_sriov_migration_data packet
+ *
+ * Called by the restore migration data producer (userspace) when processing
+ * migration data.
+ * If the ring is full, waits until there is space.
+ * Queues the restore migration data consumer (PF SR-IOV Control worker), that
+ * is potentially waiting for data when the ring is empty.
+ *
+ * Return: 0 on success or a negative error code on failure.
+ */
+int xe_gt_sriov_pf_migration_restore_produce(struct xe_gt *gt, unsigned int vfid,
+					     struct xe_sriov_migration_data *data)
+{
+	struct wait_queue_head *wq = xe_sriov_pf_migration_waitqueue(gt_to_xe(gt), vfid);
+	struct xe_gt_sriov_migration_data *migration = pf_pick_gt_migration(gt, vfid);
+	int ret;
+
+	xe_gt_assert(gt, data->tile == gt->tile->id);
+	xe_gt_assert(gt, data->gt == gt->info.id);
+
+	for (;;) {
+		ret = ptr_ring_produce(&migration->ring, data);
+		if (!ret)
+			break;
+
+		ret = wait_event_interruptible(*wq, !ptr_ring_full(&migration->ring));
+		if (ret)
+			return ret;
+	}
+
+	xe_gt_sriov_pf_control_process_restore_data(gt, vfid);
+
+	return 0;
+}
+
+/**
+ * xe_gt_sriov_pf_migration_save_consume() - Get VF save data packet from migration ring.
+ * @gt: the &xe_gt
+ * @vfid: the VF identifier
+ *
+ * Called by the save migration data consumer (userspace) when
+ * processing migration data.
+ * Queues the save migration data producer (PF SR-IOV Control worker), that is
+ * potentially waiting to add more data when the ring is full.
+ *
+ * Return: Pointer to &struct xe_sriov_migration_data on success,
+ *	   NULL if ring is empty and there's no more data available,
+ *	   ERR_PTR(-EAGAIN) if the ring is empty, but data is still produced.
+ */
+struct xe_sriov_migration_data *
+xe_gt_sriov_pf_migration_save_consume(struct xe_gt *gt, unsigned int vfid)
+{
+	struct xe_gt_sriov_migration_data *migration = pf_pick_gt_migration(gt, vfid);
+	struct xe_sriov_migration_data *data;
+
+	data = ptr_ring_consume(&migration->ring);
+	if (data) {
+		xe_gt_sriov_pf_control_process_save_data(gt, vfid);
+		return data;
+	}
+
+	if (xe_gt_sriov_pf_control_check_save_data_done(gt, vfid))
+		return NULL;
+
+	return ERR_PTR(-EAGAIN);
+}
+
+static void action_ring_cleanup(void *arg)
+{
+	struct ptr_ring *r = arg;
+
+	ptr_ring_cleanup(r, NULL);
+}
+
 /**
  * xe_gt_sriov_pf_migration_init() - Initialize support for VF migration.
  * @gt: the &xe_gt
@@ -393,6 +561,7 @@ ssize_t xe_gt_sriov_pf_migration_write_guc_state(struct xe_gt *gt, unsigned int
 int xe_gt_sriov_pf_migration_init(struct xe_gt *gt)
 {
 	struct xe_device *xe = gt_to_xe(gt);
+	unsigned int n, totalvfs;
 	int err;
 
 	xe_gt_assert(gt, IS_SRIOV_PF(xe));
@@ -404,5 +573,19 @@ int xe_gt_sriov_pf_migration_init(struct xe_gt *gt)
 	if (err)
 		return err;
 
+	totalvfs = xe_sriov_pf_get_totalvfs(xe);
+	for (n = 1; n <= totalvfs; n++) {
+		struct xe_gt_sriov_migration_data *migration = pf_pick_gt_migration(gt, n);
+
+		err = ptr_ring_init(&migration->ring,
+				    XE_GT_SRIOV_PF_MIGRATION_RING_SIZE, GFP_KERNEL);
+		if (err)
+			return err;
+
+		err = devm_add_action_or_reset(xe->drm.dev, action_ring_cleanup, &migration->ring);
+		if (err)
+			return err;
+	}
+
 	return 0;
 }
diff --git a/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.h b/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.h
index 09faeae00ddbb..9e67f18ded205 100644
--- a/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.h
+++ b/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.h
@@ -9,11 +9,25 @@
 #include <linux/types.h>
 
 struct xe_gt;
+struct xe_sriov_migration_data;
 
 int xe_gt_sriov_pf_migration_init(struct xe_gt *gt);
 int xe_gt_sriov_pf_migration_save_guc_state(struct xe_gt *gt, unsigned int vfid);
 int xe_gt_sriov_pf_migration_restore_guc_state(struct xe_gt *gt, unsigned int vfid);
 
+bool xe_gt_sriov_pf_migration_ring_empty(struct xe_gt *gt, unsigned int vfid);
+bool xe_gt_sriov_pf_migration_ring_full(struct xe_gt *gt, unsigned int vfid);
+
+int xe_gt_sriov_pf_migration_save_produce(struct xe_gt *gt, unsigned int vfid,
+					  struct xe_sriov_migration_data *data);
+struct xe_sriov_migration_data *
+xe_gt_sriov_pf_migration_restore_consume(struct xe_gt *gt, unsigned int vfid);
+
+int xe_gt_sriov_pf_migration_restore_produce(struct xe_gt *gt, unsigned int vfid,
+					     struct xe_sriov_migration_data *data);
+struct xe_sriov_migration_data *
+xe_gt_sriov_pf_migration_save_consume(struct xe_gt *gt, unsigned int vfid);
+
 #ifdef CONFIG_DEBUG_FS
 ssize_t xe_gt_sriov_pf_migration_read_guc_state(struct xe_gt *gt, unsigned int vfid,
 						char __user *buf, size_t count, loff_t *pos);
diff --git a/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration_types.h b/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration_types.h
index 9d672feac5f04..84be6fac16c8b 100644
--- a/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration_types.h
+++ b/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration_types.h
@@ -7,6 +7,7 @@
 #define _XE_GT_SRIOV_PF_MIGRATION_TYPES_H_
 
 #include <linux/mutex.h>
+#include <linux/ptr_ring.h>
 #include <linux/types.h>
 
 /**
@@ -24,6 +25,16 @@ struct xe_gt_sriov_state_snapshot {
 	} guc;
 };
 
+/**
+ * struct xe_gt_sriov_migration_data - GT-level per-VF migration data.
+ *
+ * Used by the PF driver to maintain per-VF migration data.
+ */
+struct xe_gt_sriov_migration_data {
+	/** @ring: queue containing VF save / restore migration data */
+	struct ptr_ring ring;
+};
+
 /**
  * struct xe_gt_sriov_pf_migration - GT-level data.
  *
diff --git a/drivers/gpu/drm/xe/xe_gt_sriov_pf_types.h b/drivers/gpu/drm/xe/xe_gt_sriov_pf_types.h
index a64a6835ad656..812e74d3f8f80 100644
--- a/drivers/gpu/drm/xe/xe_gt_sriov_pf_types.h
+++ b/drivers/gpu/drm/xe/xe_gt_sriov_pf_types.h
@@ -33,6 +33,9 @@ struct xe_gt_sriov_metadata {
 
 	/** @snapshot: snapshot of the VF state data */
 	struct xe_gt_sriov_state_snapshot snapshot;
+
+	/** @migration: per-VF migration data. */
+	struct xe_gt_sriov_migration_data migration;
 };
 
 /**
diff --git a/drivers/gpu/drm/xe/xe_sriov_pf_migration.c b/drivers/gpu/drm/xe/xe_sriov_pf_migration.c
index 8c523c392f98b..7be9f026d80e8 100644
--- a/drivers/gpu/drm/xe/xe_sriov_pf_migration.c
+++ b/drivers/gpu/drm/xe/xe_sriov_pf_migration.c
@@ -3,8 +3,36 @@
  * Copyright © 2025 Intel Corporation
  */
 
+#include <drm/drm_managed.h>
+
+#include "xe_device.h"
+#include "xe_gt_sriov_pf_control.h"
+#include "xe_gt_sriov_pf_migration.h"
+#include "xe_pm.h"
 #include "xe_sriov.h"
+#include "xe_sriov_pf_helpers.h"
 #include "xe_sriov_pf_migration.h"
+#include "xe_sriov_printk.h"
+
+static struct xe_sriov_pf_migration *pf_pick_migration(struct xe_device *xe, unsigned int vfid)
+{
+	xe_assert(xe, IS_SRIOV_PF(xe));
+	xe_assert(xe, vfid <= xe_sriov_pf_get_totalvfs(xe));
+
+	return &xe->sriov.pf.vfs[vfid].migration;
+}
+
+/**
+ * xe_sriov_pf_migration_waitqueue - Get waitqueue for migration.
+ * @xe: the &xe_device
+ * @vfid: the VF identifier
+ *
+ * Return: pointer to the migration waitqueue.
+ */
+wait_queue_head_t *xe_sriov_pf_migration_waitqueue(struct xe_device *xe, unsigned int vfid)
+{
+	return &pf_pick_migration(xe, vfid)->wq;
+}
 
 /**
  * xe_sriov_pf_migration_supported() - Check if SR-IOV VF migration is supported by the device
@@ -33,9 +61,123 @@ static bool pf_check_migration_support(struct xe_device *xe)
  */
 int xe_sriov_pf_migration_init(struct xe_device *xe)
 {
+	unsigned int n, totalvfs;
+
 	xe_assert(xe, IS_SRIOV_PF(xe));
 
 	xe->sriov.pf.migration.supported = pf_check_migration_support(xe);
+	if (!xe_sriov_pf_migration_supported(xe))
+		return 0;
+
+	totalvfs = xe_sriov_pf_get_totalvfs(xe);
+	for (n = 1; n <= totalvfs; n++) {
+		struct xe_sriov_pf_migration *migration = pf_pick_migration(xe, n);
+
+		init_waitqueue_head(&migration->wq);
+	}
 
 	return 0;
 }
+
+static bool pf_migration_data_ready(struct xe_device *xe, unsigned int vfid)
+{
+	struct xe_gt *gt;
+	u8 gt_id;
+
+	for_each_gt(gt, xe, gt_id) {
+		if (!xe_gt_sriov_pf_migration_ring_empty(gt, vfid) ||
+		    xe_gt_sriov_pf_control_check_save_data_done(gt, vfid))
+			return true;
+	}
+
+	return false;
+}
+
+static struct xe_sriov_migration_data *
+pf_migration_consume(struct xe_device *xe, unsigned int vfid)
+{
+	struct xe_sriov_migration_data *data;
+	struct xe_gt *gt;
+	u8 gt_id;
+	bool more_data = false;
+
+	for_each_gt(gt, xe, gt_id) {
+		data = xe_gt_sriov_pf_migration_save_consume(gt, vfid);
+		if (data && PTR_ERR(data) != EAGAIN)
+			return data;
+		if (PTR_ERR(data) == -EAGAIN)
+			more_data = true;
+	}
+
+	if (!more_data)
+		return NULL;
+
+	return ERR_PTR(-EAGAIN);
+}
+
+/**
+ * xe_sriov_pf_migration_save_consume() - Consume a VF migration data packet from the device.
+ * @xe: the &xe_device
+ * @vfid: the VF identifier
+ *
+ * Called by the save migration data consumer (userspace) when
+ * processing migration data.
+ * If there is no migration data to process, wait until more data is available.
+ *
+ * Return: Pointer to &xe_sriov_migration_data on success,
+ *	   NULL if ring is empty and no more migration data is expected,
+ *	   ERR_PTR value in case of error.
+ *
+ * Return: 0 on success or a negative error code on failure.
+ */
+struct xe_sriov_migration_data *
+xe_sriov_pf_migration_save_consume(struct xe_device *xe, unsigned int vfid)
+{
+	struct xe_sriov_pf_migration *migration = pf_pick_migration(xe, vfid);
+	struct xe_sriov_migration_data *data;
+	int ret;
+
+	xe_assert(xe, IS_SRIOV_PF(xe));
+
+	for (;;) {
+		data = pf_migration_consume(xe, vfid);
+		if (PTR_ERR(data) != -EAGAIN)
+			break;
+
+		ret = wait_event_interruptible(migration->wq,
+					       pf_migration_data_ready(xe, vfid));
+		if (ret)
+			return ERR_PTR(ret);
+	}
+
+	return data;
+}
+
+/**
+ * xe_sriov_pf_migration_restore_produce() - Produce a VF migration data packet to the device.
+ * @xe: the &xe_device
+ * @vfid: the VF identifier
+ * @data: Pointer to &xe_sriov_migration_data
+ *
+ * Called by the restore migration data producer (userspace) when processing
+ * migration data.
+ * If the underlying data structure is full, wait until there is space.
+ *
+ * Return: 0 on success or a negative error code on failure.
+ */
+int xe_sriov_pf_migration_restore_produce(struct xe_device *xe, unsigned int vfid,
+					  struct xe_sriov_migration_data *data)
+{
+	struct xe_gt *gt;
+
+	xe_assert(xe, IS_SRIOV_PF(xe));
+
+	gt = xe_device_get_gt(xe, data->gt);
+	if (!gt || data->tile != gt->tile->id) {
+		xe_sriov_err_ratelimited(xe, "VF%d Invalid GT - tile:%u, GT:%u\n",
+					 vfid, data->tile, data->gt);
+		return -EINVAL;
+	}
+
+	return xe_gt_sriov_pf_migration_restore_produce(gt, vfid, data);
+}
diff --git a/drivers/gpu/drm/xe/xe_sriov_pf_migration.h b/drivers/gpu/drm/xe/xe_sriov_pf_migration.h
index d2b4a24165438..df81a540c246a 100644
--- a/drivers/gpu/drm/xe/xe_sriov_pf_migration.h
+++ b/drivers/gpu/drm/xe/xe_sriov_pf_migration.h
@@ -7,10 +7,17 @@
 #define _XE_SRIOV_PF_MIGRATION_H_
 
 #include <linux/types.h>
+#include <linux/wait.h>
 
 struct xe_device;
+struct xe_sriov_migration_data;
 
 int xe_sriov_pf_migration_init(struct xe_device *xe);
 bool xe_sriov_pf_migration_supported(struct xe_device *xe);
+int xe_sriov_pf_migration_restore_produce(struct xe_device *xe, unsigned int vfid,
+					  struct xe_sriov_migration_data *data);
+struct xe_sriov_migration_data *
+xe_sriov_pf_migration_save_consume(struct xe_device *xe, unsigned int vfid);
+wait_queue_head_t *xe_sriov_pf_migration_waitqueue(struct xe_device *xe, unsigned int vfid);
 
 #endif
diff --git a/drivers/gpu/drm/xe/xe_sriov_pf_migration_types.h b/drivers/gpu/drm/xe/xe_sriov_pf_migration_types.h
index e69de29bb2d1d..2a45ee4e3ece8 100644
--- a/drivers/gpu/drm/xe/xe_sriov_pf_migration_types.h
+++ b/drivers/gpu/drm/xe/xe_sriov_pf_migration_types.h
@@ -0,0 +1,58 @@
+/* SPDX-License-Identifier: MIT */
+/*
+ * Copyright © 2025 Intel Corporation
+ */
+
+#ifndef _XE_SRIOV_PF_MIGRATION_TYPES_H_
+#define _XE_SRIOV_PF_MIGRATION_TYPES_H_
+
+#include <linux/types.h>
+#include <linux/wait.h>
+
+/**
+ * struct xe_sriov_migration_data - Xe SR-IOV VF migration data packet
+ */
+struct xe_sriov_migration_data {
+	/** @xe: Xe device */
+	struct xe_device *xe;
+	/** @vaddr: CPU pointer to payload data */
+	void *vaddr;
+	/** @remaining: payload data remaining */
+	size_t remaining;
+	/** @hdr_remaining: header data remaining */
+	size_t hdr_remaining;
+	union {
+		/** @bo: Buffer object with migration data */
+		struct xe_bo *bo;
+		/** @buff: Buffer with migration data */
+		void *buff;
+	};
+	__struct_group(xe_sriov_pf_migration_hdr, hdr, __packed,
+		/** @hdr.version: migration data protocol version */
+		u8 version;
+		/** @hdr.type: migration data type */
+		u8 type;
+		/** @hdr.tile: migration data tile id */
+		u8 tile;
+		/** @hdr.gt: migration data gt id */
+		u8 gt;
+		/** @hdr.flags: migration data flags */
+		u32 flags;
+		/** @hdr.offset: offset into the resource;
+		 * used when multiple packets of given type are used for migration
+		 */
+		u64 offset;
+		/** @hdr.size: migration data size  */
+		u64 size;
+	);
+};
+
+/**
+ * struct xe_sriov_pf_migration - Per VF device-level migration related data
+ */
+struct xe_sriov_pf_migration {
+	/** @wq: waitqueue used to avoid busy-waiting for snapshot production/consumption */
+	wait_queue_head_t wq;
+};
+
+#endif
diff --git a/drivers/gpu/drm/xe/xe_sriov_pf_types.h b/drivers/gpu/drm/xe/xe_sriov_pf_types.h
index 24d22afeececa..c92baaa1694ca 100644
--- a/drivers/gpu/drm/xe/xe_sriov_pf_types.h
+++ b/drivers/gpu/drm/xe/xe_sriov_pf_types.h
@@ -9,6 +9,7 @@
 #include <linux/mutex.h>
 #include <linux/types.h>
 
+#include "xe_sriov_pf_migration_types.h"
 #include "xe_sriov_pf_provision_types.h"
 #include "xe_sriov_pf_service_types.h"
 
@@ -18,6 +19,8 @@
 struct xe_sriov_metadata {
 	/** @version: negotiated VF/PF ABI version */
 	struct xe_sriov_pf_service_version version;
+	/** @migration: migration data */
+	struct xe_sriov_pf_migration migration;
 };
 
 /**
-- 
2.50.1

[PATCH v3 06/28] drm/xe/pf: Add helpers for migration data allocation / free

[Michał Winiarski]

Now that it's possible to free the packets - connect the restore
handling logic with the ring.
The helpers will also be used in upcoming changes that will start producing
migration data packets.

Signed-off-by: Michał Winiarski <michal.winiarski@intel.com>
Reviewed-by: Michal Wajdeczko <michal.wajdeczko@intel.com>
---
 drivers/gpu/drm/xe/Makefile                   |   1 +
 drivers/gpu/drm/xe/xe_gt_sriov_pf_control.c   |   7 +
 drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.c |  29 +++-
 drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.h |   1 +
 drivers/gpu/drm/xe/xe_sriov_migration_data.c  | 126 ++++++++++++++++++
 drivers/gpu/drm/xe/xe_sriov_migration_data.h  |  30 +++++
 6 files changed, 193 insertions(+), 1 deletion(-)
 create mode 100644 drivers/gpu/drm/xe/xe_sriov_migration_data.c
 create mode 100644 drivers/gpu/drm/xe/xe_sriov_migration_data.h

diff --git a/drivers/gpu/drm/xe/Makefile b/drivers/gpu/drm/xe/Makefile
index 89e5b26c27975..3d72db9e528e4 100644
--- a/drivers/gpu/drm/xe/Makefile
+++ b/drivers/gpu/drm/xe/Makefile
@@ -173,6 +173,7 @@ xe-$(CONFIG_PCI_IOV) += \
 	xe_lmtt_2l.o \
 	xe_lmtt_ml.o \
 	xe_pci_sriov.o \
+	xe_sriov_migration_data.o \
 	xe_sriov_pf.o \
 	xe_sriov_pf_control.o \
 	xe_sriov_pf_debugfs.o \
diff --git a/drivers/gpu/drm/xe/xe_gt_sriov_pf_control.c b/drivers/gpu/drm/xe/xe_gt_sriov_pf_control.c
index ad9dcd3f56453..02019b646389c 100644
--- a/drivers/gpu/drm/xe/xe_gt_sriov_pf_control.c
+++ b/drivers/gpu/drm/xe/xe_gt_sriov_pf_control.c
@@ -18,6 +18,7 @@
 #include "xe_gt_sriov_printk.h"
 #include "xe_guc_ct.h"
 #include "xe_sriov.h"
+#include "xe_sriov_migration_data.h"
 #include "xe_sriov_pf_control.h"
 #include "xe_sriov_pf_migration.h"
 #include "xe_sriov_pf_service.h"
@@ -851,6 +852,8 @@ int xe_gt_sriov_pf_control_resume_vf(struct xe_gt *gt, unsigned int vfid)
 static void pf_exit_vf_save_wip(struct xe_gt *gt, unsigned int vfid)
 {
 	if (pf_exit_vf_state(gt, vfid, XE_GT_SRIOV_STATE_SAVE_WIP)) {
+		xe_gt_sriov_pf_migration_ring_free(gt, vfid);
+
 		pf_escape_vf_state(gt, vfid, XE_GT_SRIOV_STATE_SAVE_PROCESS_DATA);
 		pf_escape_vf_state(gt, vfid, XE_GT_SRIOV_STATE_SAVE_WAIT_DATA);
 		pf_escape_vf_state(gt, vfid, XE_GT_SRIOV_STATE_SAVE_DATA_DONE);
@@ -1048,6 +1051,8 @@ int xe_gt_sriov_pf_control_finish_save_vf(struct xe_gt *gt, unsigned int vfid)
 static void pf_exit_vf_restore_wip(struct xe_gt *gt, unsigned int vfid)
 {
 	if (pf_exit_vf_state(gt, vfid, XE_GT_SRIOV_STATE_RESTORE_WIP)) {
+		xe_gt_sriov_pf_migration_ring_free(gt, vfid);
+
 		pf_escape_vf_state(gt, vfid, XE_GT_SRIOV_STATE_RESTORE_PROCESS_DATA);
 		pf_escape_vf_state(gt, vfid, XE_GT_SRIOV_STATE_RESTORE_WAIT_DATA);
 		pf_escape_vf_state(gt, vfid, XE_GT_SRIOV_STATE_RESTORE_DATA_DONE);
@@ -1080,6 +1085,8 @@ static int pf_handle_vf_restore_data(struct xe_gt *gt, unsigned int vfid)
 
 	xe_gt_sriov_notice(gt, "Skipping VF%u unknown data type: %d\n", vfid, data->type);
 
+	xe_sriov_migration_data_free(data);
+
 	return 0;
 }
 
diff --git a/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.c b/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.c
index e61e6f4215864..40b75b5fe9229 100644
--- a/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.c
+++ b/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.c
@@ -14,6 +14,7 @@
 #include "xe_guc.h"
 #include "xe_guc_ct.h"
 #include "xe_sriov.h"
+#include "xe_sriov_migration_data.h"
 #include "xe_sriov_pf_migration.h"
 
 #define XE_GT_SRIOV_PF_MIGRATION_RING_SIZE 5
@@ -418,6 +419,25 @@ bool xe_gt_sriov_pf_migration_ring_full(struct xe_gt *gt, unsigned int vfid)
 	return ptr_ring_full(&pf_pick_gt_migration(gt, vfid)->ring);
 }
 
+/**
+ * xe_gt_sriov_pf_migration_ring_free() - Consume and free all data in migration ring
+ * @gt: the &xe_gt
+ * @vfid: the VF identifier
+ */
+void xe_gt_sriov_pf_migration_ring_free(struct xe_gt *gt, unsigned int vfid)
+{
+	struct xe_gt_sriov_migration_data *migration = pf_pick_gt_migration(gt, vfid);
+	struct xe_sriov_migration_data *data;
+
+	if (ptr_ring_empty(&migration->ring))
+		return;
+
+	xe_gt_sriov_notice(gt, "VF%u unprocessed migration data left in the ring!\n", vfid);
+
+	while ((data = ptr_ring_consume(&migration->ring)))
+		xe_sriov_migration_data_free(data);
+}
+
 /**
  * xe_gt_sriov_pf_migration_save_produce() - Add VF save data packet to migration ring.
  * @gt: the &xe_gt
@@ -543,11 +563,18 @@ xe_gt_sriov_pf_migration_save_consume(struct xe_gt *gt, unsigned int vfid)
 	return ERR_PTR(-EAGAIN);
 }
 
+static void pf_mig_data_destroy(void *ptr)
+{
+	struct xe_sriov_migration_data *data = ptr;
+
+	xe_sriov_migration_data_free(data);
+}
+
 static void action_ring_cleanup(void *arg)
 {
 	struct ptr_ring *r = arg;
 
-	ptr_ring_cleanup(r, NULL);
+	ptr_ring_cleanup(r, pf_mig_data_destroy);
 }
 
 /**
diff --git a/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.h b/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.h
index 9e67f18ded205..1ed2248f0a17e 100644
--- a/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.h
+++ b/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.h
@@ -17,6 +17,7 @@ int xe_gt_sriov_pf_migration_restore_guc_state(struct xe_gt *gt, unsigned int vf
 
 bool xe_gt_sriov_pf_migration_ring_empty(struct xe_gt *gt, unsigned int vfid);
 bool xe_gt_sriov_pf_migration_ring_full(struct xe_gt *gt, unsigned int vfid);
+void xe_gt_sriov_pf_migration_ring_free(struct xe_gt *gt, unsigned int vfid);
 
 int xe_gt_sriov_pf_migration_save_produce(struct xe_gt *gt, unsigned int vfid,
 					  struct xe_sriov_migration_data *data);
diff --git a/drivers/gpu/drm/xe/xe_sriov_migration_data.c b/drivers/gpu/drm/xe/xe_sriov_migration_data.c
new file mode 100644
index 0000000000000..2371ca3e6b9a9
--- /dev/null
+++ b/drivers/gpu/drm/xe/xe_sriov_migration_data.c
@@ -0,0 +1,126 @@
+// SPDX-License-Identifier: MIT
+/*
+ * Copyright © 2025 Intel Corporation
+ */
+
+#include "xe_bo.h"
+#include "xe_device.h"
+#include "xe_sriov_migration_data.h"
+
+static bool data_needs_bo(struct xe_sriov_migration_data *data)
+{
+	return data->type == XE_SRIOV_MIGRATION_DATA_TYPE_VRAM;
+}
+
+/**
+ * xe_sriov_migration_data() - Allocate migration data packet
+ * @xe: the &xe_device
+ *
+ * Only allocates the "outer" structure, without initializing the migration
+ * data backing storage.
+ *
+ * Return: Pointer to &xe_sriov_migration_data on success,
+ *         NULL in case of error.
+ */
+struct xe_sriov_migration_data *xe_sriov_migration_data_alloc(struct xe_device *xe)
+{
+	struct xe_sriov_migration_data *data;
+
+	data = kzalloc(sizeof(*data), GFP_KERNEL);
+	if (!data)
+		return NULL;
+
+	data->xe = xe;
+	data->hdr_remaining = sizeof(data->hdr);
+
+	return data;
+}
+
+/**
+ * xe_sriov_migration_data_free() - Free migration data packet.
+ * @data: the &xe_sriov_migration_data packet
+ */
+void xe_sriov_migration_data_free(struct xe_sriov_migration_data *data)
+{
+	if (data_needs_bo(data))
+		xe_bo_unpin_map_no_vm(data->bo);
+	else
+		kvfree(data->buff);
+
+	kfree(data);
+}
+
+static int mig_data_init(struct xe_sriov_migration_data *data)
+{
+	struct xe_gt *gt = xe_device_get_gt(data->xe, data->gt);
+
+	if (data->size == 0)
+		return 0;
+
+	if (data_needs_bo(data)) {
+		struct xe_bo *bo;
+
+		bo = xe_bo_create_pin_map_novm(data->xe, gt->tile, PAGE_ALIGN(data->size),
+					       ttm_bo_type_kernel,
+					       XE_BO_FLAG_SYSTEM | XE_BO_FLAG_PINNED, false);
+		if (IS_ERR(bo))
+			return PTR_ERR(bo);
+
+		data->bo = bo;
+		data->vaddr = bo->vmap.vaddr;
+	} else {
+		void *buff = kvzalloc(data->size, GFP_KERNEL);
+
+		if (!buff)
+			return -ENOMEM;
+
+		data->buff = buff;
+		data->vaddr = buff;
+	}
+
+	return 0;
+}
+
+#define XE_SRIOV_MIGRATION_DATA_SUPPORTED_VERSION 1
+/**
+ * xe_sriov_migration_data_init() - Initialize the migration data header and backing storage.
+ * @data: the &xe_sriov_migration_data packet
+ * @tile_id: tile identifier
+ * @gt_id: GT identifier
+ * @type: &xe_sriov_migration_data_type
+ * @offset: offset of data packet payload (within wider resource)
+ * @size: size of data packet payload
+ *
+ * Return: 0 on success or a negative error code on failure.
+ */
+int xe_sriov_migration_data_init(struct xe_sriov_migration_data *data, u8 tile_id, u8 gt_id,
+				 enum xe_sriov_migration_data_type type, loff_t offset, size_t size)
+{
+	data->version = XE_SRIOV_MIGRATION_DATA_SUPPORTED_VERSION;
+	data->type = type;
+	data->tile = tile_id;
+	data->gt = gt_id;
+	data->offset = offset;
+	data->size = size;
+	data->remaining = size;
+
+	return mig_data_init(data);
+}
+
+/**
+ * xe_sriov_migration_data_init() - Initialize the migration data backing storage based on header.
+ * @data: the &xe_sriov_migration_data packet
+ *
+ * Header data is expected to be filled prior to calling this function.
+ *
+ * Return: 0 on success or a negative error code on failure.
+ */
+int xe_sriov_migration_data_init_from_hdr(struct xe_sriov_migration_data *data)
+{
+	if (data->version != XE_SRIOV_MIGRATION_DATA_SUPPORTED_VERSION)
+		return -EINVAL;
+
+	data->remaining = data->size;
+
+	return mig_data_init(data);
+}
diff --git a/drivers/gpu/drm/xe/xe_sriov_migration_data.h b/drivers/gpu/drm/xe/xe_sriov_migration_data.h
new file mode 100644
index 0000000000000..3958f58a170f5
--- /dev/null
+++ b/drivers/gpu/drm/xe/xe_sriov_migration_data.h
@@ -0,0 +1,30 @@
+/* SPDX-License-Identifier: MIT */
+/*
+ * Copyright © 2025 Intel Corporation
+ */
+
+#ifndef _XE_SRIOV_MIGRATION_DATA_H_
+#define _XE_SRIOV_MIGRATION_DATA_H_
+
+#include <linux/types.h>
+
+struct xe_device;
+
+enum xe_sriov_migration_data_type {
+	/* Skipping 0 to catch uninitialized data */
+	XE_SRIOV_MIGRATION_DATA_TYPE_DESCRIPTOR = 1,
+	XE_SRIOV_MIGRATION_DATA_TYPE_TRAILER,
+	XE_SRIOV_MIGRATION_DATA_TYPE_GGTT,
+	XE_SRIOV_MIGRATION_DATA_TYPE_MMIO,
+	XE_SRIOV_MIGRATION_DATA_TYPE_GUC,
+	XE_SRIOV_MIGRATION_DATA_TYPE_VRAM,
+};
+
+struct xe_sriov_migration_data *xe_sriov_migration_data_alloc(struct xe_device *xe);
+void xe_sriov_migration_data_free(struct xe_sriov_migration_data *snapshot);
+
+int xe_sriov_migration_data_init(struct xe_sriov_migration_data *data, u8 tile_id, u8 gt_id,
+				 enum xe_sriov_migration_data_type, loff_t offset, size_t size);
+int xe_sriov_migration_data_init_from_hdr(struct xe_sriov_migration_data *snapshot);
+
+#endif
-- 
2.50.1

[PATCH v3 07/28] drm/xe/pf: Add support for encap/decap of bitstream to/from packet

[Michał Winiarski]

Add debugfs handlers for migration state and handle bitstream
.read()/.write() to convert from bitstream to/from migration data
packets.
As descriptor/trailer are handled at this layer - add handling for both
save and restore side.

Signed-off-by: Michał Winiarski <michal.winiarski@intel.com>
---
 drivers/gpu/drm/xe/xe_sriov_migration_data.c  | 337 ++++++++++++++++++
 drivers/gpu/drm/xe/xe_sriov_migration_data.h  |   5 +
 drivers/gpu/drm/xe/xe_sriov_pf_control.c      |   5 +
 drivers/gpu/drm/xe/xe_sriov_pf_debugfs.c      |  35 ++
 drivers/gpu/drm/xe/xe_sriov_pf_migration.c    |  54 +++
 .../gpu/drm/xe/xe_sriov_pf_migration_types.h  |   9 +
 6 files changed, 445 insertions(+)

diff --git a/drivers/gpu/drm/xe/xe_sriov_migration_data.c b/drivers/gpu/drm/xe/xe_sriov_migration_data.c
index 2371ca3e6b9a9..a3f50836adc81 100644
--- a/drivers/gpu/drm/xe/xe_sriov_migration_data.c
+++ b/drivers/gpu/drm/xe/xe_sriov_migration_data.c
@@ -6,6 +6,45 @@
 #include "xe_bo.h"
 #include "xe_device.h"
 #include "xe_sriov_migration_data.h"
+#include "xe_sriov_pf_helpers.h"
+#include "xe_sriov_pf_migration.h"
+#include "xe_sriov_printk.h"
+
+static struct mutex *pf_migration_mutex(struct xe_device *xe, unsigned int vfid)
+{
+	xe_assert(xe, IS_SRIOV_PF(xe));
+	xe_assert(xe, vfid <= xe_sriov_pf_get_totalvfs(xe));
+
+	return &xe->sriov.pf.vfs[vfid].migration.lock;
+}
+
+static struct xe_sriov_migration_data **pf_pick_pending(struct xe_device *xe, unsigned int vfid)
+{
+	xe_assert(xe, IS_SRIOV_PF(xe));
+	xe_assert(xe, vfid <= xe_sriov_pf_get_totalvfs(xe));
+	lockdep_assert_held(pf_migration_mutex(xe, vfid));
+
+	return &xe->sriov.pf.vfs[vfid].migration.pending;
+}
+
+static struct xe_sriov_migration_data **
+pf_pick_descriptor(struct xe_device *xe, unsigned int vfid)
+{
+	xe_assert(xe, IS_SRIOV_PF(xe));
+	xe_assert(xe, vfid <= xe_sriov_pf_get_totalvfs(xe));
+	lockdep_assert_held(pf_migration_mutex(xe, vfid));
+
+	return &xe->sriov.pf.vfs[vfid].migration.descriptor;
+}
+
+static struct xe_sriov_migration_data **pf_pick_trailer(struct xe_device *xe, unsigned int vfid)
+{
+	xe_assert(xe, IS_SRIOV_PF(xe));
+	xe_assert(xe, vfid <= xe_sriov_pf_get_totalvfs(xe));
+	lockdep_assert_held(pf_migration_mutex(xe, vfid));
+
+	return &xe->sriov.pf.vfs[vfid].migration.trailer;
+}
 
 static bool data_needs_bo(struct xe_sriov_migration_data *data)
 {
@@ -42,6 +81,9 @@ struct xe_sriov_migration_data *xe_sriov_migration_data_alloc(struct xe_device *
  */
 void xe_sriov_migration_data_free(struct xe_sriov_migration_data *data)
 {
+	if (IS_ERR_OR_NULL(data))
+		return;
+
 	if (data_needs_bo(data))
 		xe_bo_unpin_map_no_vm(data->bo);
 	else
@@ -124,3 +166,298 @@ int xe_sriov_migration_data_init_from_hdr(struct xe_sriov_migration_data *data)
 
 	return mig_data_init(data);
 }
+
+static ssize_t vf_mig_data_hdr_read(struct xe_sriov_migration_data *data,
+				    char __user *buf, size_t len)
+{
+	loff_t offset = sizeof(data->hdr) - data->hdr_remaining;
+
+	if (!data->hdr_remaining)
+		return -EINVAL;
+
+	if (len > data->hdr_remaining)
+		len = data->hdr_remaining;
+
+	if (copy_to_user(buf, (void *)&data->hdr + offset, len))
+		return -EFAULT;
+
+	data->hdr_remaining -= len;
+
+	return len;
+}
+
+static ssize_t vf_mig_data_read(struct xe_sriov_migration_data *data,
+				char __user *buf, size_t len)
+{
+	if (len > data->remaining)
+		len = data->remaining;
+
+	if (copy_to_user(buf, data->vaddr + (data->size - data->remaining), len))
+		return -EFAULT;
+
+	data->remaining -= len;
+
+	return len;
+}
+
+static ssize_t __vf_mig_data_read_single(struct xe_sriov_migration_data **data,
+					 unsigned int vfid, char __user *buf, size_t len)
+{
+	ssize_t copied = 0;
+
+	if ((*data)->hdr_remaining)
+		copied = vf_mig_data_hdr_read(*data, buf, len);
+	else
+		copied = vf_mig_data_read(*data, buf, len);
+
+	if ((*data)->remaining == 0 && (*data)->hdr_remaining == 0) {
+		xe_sriov_migration_data_free(*data);
+		*data = NULL;
+	}
+
+	return copied;
+}
+
+static struct xe_sriov_migration_data **vf_mig_pick_data(struct xe_device *xe, unsigned int vfid)
+{
+	struct xe_sriov_migration_data **data;
+
+	data = pf_pick_descriptor(xe, vfid);
+	if (*data)
+		return data;
+
+	data = pf_pick_pending(xe, vfid);
+	if (!*data)
+		*data = xe_sriov_pf_migration_save_consume(xe, vfid);
+	if (*data)
+		return data;
+
+	data = pf_pick_trailer(xe, vfid);
+	if (*data)
+		return data;
+
+	return ERR_PTR(-ENODATA);
+}
+
+static ssize_t vf_mig_data_read_single(struct xe_device *xe, unsigned int vfid,
+				       char __user *buf, size_t len)
+{
+	struct xe_sriov_migration_data **data = vf_mig_pick_data(xe, vfid);
+
+	if (IS_ERR(data))
+		return PTR_ERR(data);
+
+	return __vf_mig_data_read_single(data, vfid, buf, len);
+}
+
+/**
+ * xe_sriov_migration_data_read() - Read migration data from the device.
+ * @xe: the &xe_device
+ * @vfid: the VF identifier
+ * @buf: start address of userspace buffer
+ * @len: requested read size from userspace
+ *
+ * Return: number of bytes that has been successfully read,
+ *	   0 if no more migration data is available,
+ *	   -errno on failure.
+ */
+ssize_t xe_sriov_migration_data_read(struct xe_device *xe, unsigned int vfid,
+				     char __user *buf, size_t len)
+{
+	ssize_t ret, consumed = 0;
+
+	xe_assert(xe, IS_SRIOV_PF(xe));
+
+	scoped_cond_guard(mutex_intr, return -EINTR, pf_migration_mutex(xe, vfid)) {
+		while (consumed < len) {
+			ret = vf_mig_data_read_single(xe, vfid, buf, len - consumed);
+			if (ret == -ENODATA)
+				break;
+			if (ret < 0)
+				return ret;
+
+			consumed += ret;
+			buf += ret;
+		}
+	}
+
+	return consumed;
+}
+
+static ssize_t vf_mig_hdr_write(struct xe_sriov_migration_data *data,
+				const char __user *buf, size_t len)
+{
+	loff_t offset = sizeof(data->hdr) - data->hdr_remaining;
+	int ret;
+
+	if (len > data->hdr_remaining)
+		len = data->hdr_remaining;
+
+	if (copy_from_user((void *)&data->hdr + offset, buf, len))
+		return -EFAULT;
+
+	data->hdr_remaining -= len;
+
+	if (!data->hdr_remaining) {
+		ret = xe_sriov_migration_data_init_from_hdr(data);
+		if (ret)
+			return ret;
+	}
+
+	return len;
+}
+
+static ssize_t vf_mig_data_write(struct xe_sriov_migration_data *data,
+				 const char __user *buf, size_t len)
+{
+	if (len > data->remaining)
+		len = data->remaining;
+
+	if (copy_from_user(data->vaddr + (data->size - data->remaining), buf, len))
+		return -EFAULT;
+
+	data->remaining -= len;
+
+	return len;
+}
+
+static ssize_t vf_mig_data_write_single(struct xe_device *xe, unsigned int vfid,
+					const char __user *buf, size_t len)
+{
+	struct xe_sriov_migration_data **data = pf_pick_pending(xe, vfid);
+	int ret;
+	ssize_t copied;
+
+	if (IS_ERR_OR_NULL(*data)) {
+		*data = xe_sriov_migration_data_alloc(xe);
+		if (!*data)
+			return -ENOMEM;
+	}
+
+	if ((*data)->hdr_remaining)
+		copied = vf_mig_hdr_write(*data, buf, len);
+	else
+		copied = vf_mig_data_write(*data, buf, len);
+
+	if ((*data)->hdr_remaining == 0 && (*data)->remaining == 0) {
+		ret = xe_sriov_pf_migration_restore_produce(xe, vfid, *data);
+		if (ret) {
+			xe_sriov_migration_data_free(*data);
+			return ret;
+		}
+
+		*data = NULL;
+	}
+
+	return copied;
+}
+
+/**
+ * xe_sriov_migration_data_write() - Write migration data to the device.
+ * @xe: the &xe_device
+ * @vfid: the VF identifier
+ * @buf: start address of userspace buffer
+ * @len: requested write size from userspace
+ *
+ * Return: number of bytes that has been successfully written,
+ *	   -errno on failure.
+ */
+ssize_t xe_sriov_migration_data_write(struct xe_device *xe, unsigned int vfid,
+				      const char __user *buf, size_t len)
+{
+	ssize_t ret, produced = 0;
+
+	xe_assert(xe, IS_SRIOV_PF(xe));
+
+	scoped_cond_guard(mutex_intr, return -EINTR, pf_migration_mutex(xe, vfid)) {
+		while (produced < len) {
+			ret = vf_mig_data_write_single(xe, vfid, buf, len - produced);
+			if (ret < 0)
+				return ret;
+
+			produced += ret;
+			buf += ret;
+		}
+	}
+
+	return produced;
+}
+
+#define MIGRATION_DESCRIPTOR_DWORDS 0
+static size_t pf_descriptor_init(struct xe_device *xe, unsigned int vfid)
+{
+	struct xe_sriov_migration_data **desc = pf_pick_descriptor(xe, vfid);
+	struct xe_sriov_migration_data *data;
+	int ret;
+
+	data = xe_sriov_migration_data_alloc(xe);
+	if (!data)
+		return -ENOMEM;
+
+	ret = xe_sriov_migration_data_init(data, 0, 0, XE_SRIOV_MIGRATION_DATA_TYPE_DESCRIPTOR,
+					   0, MIGRATION_DESCRIPTOR_DWORDS * sizeof(u32));
+	if (ret) {
+		xe_sriov_migration_data_free(data);
+		return ret;
+	}
+
+	*desc = data;
+
+	return 0;
+}
+
+static void pf_pending_init(struct xe_device *xe, unsigned int vfid)
+{
+	struct xe_sriov_migration_data **data = pf_pick_pending(xe, vfid);
+
+	*data = NULL;
+}
+
+#define MIGRATION_TRAILER_SIZE 0
+static int pf_trailer_init(struct xe_device *xe, unsigned int vfid)
+{
+	struct xe_sriov_migration_data **trailer = pf_pick_trailer(xe, vfid);
+	struct xe_sriov_migration_data *data;
+	int ret;
+
+	data = xe_sriov_migration_data_alloc(xe);
+	if (!data)
+		return -ENOMEM;
+
+	ret = xe_sriov_migration_data_init(data, 0, 0, XE_SRIOV_MIGRATION_DATA_TYPE_TRAILER,
+					   0, MIGRATION_TRAILER_SIZE);
+	if (ret) {
+		xe_sriov_migration_data_free(data);
+		return ret;
+	}
+
+	*trailer = data;
+
+	return 0;
+}
+
+/**
+ * xe_sriov_migration_data_save_init() - Initialize the pending save migration data.
+ * @xe: the &xe_device
+ * @vfid: the VF identifier
+ *
+ * Return: 0 on success, -errno on failure.
+ */
+int xe_sriov_migration_data_save_init(struct xe_device *xe, unsigned int vfid)
+{
+	int ret;
+
+	scoped_cond_guard(mutex_intr, return -EINTR, pf_migration_mutex(xe, vfid)) {
+		ret = pf_descriptor_init(xe, vfid);
+		if (ret)
+			return ret;
+
+		ret = pf_trailer_init(xe, vfid);
+		if (ret)
+			return ret;
+
+		pf_pending_init(xe, vfid);
+	}
+
+	return 0;
+}
diff --git a/drivers/gpu/drm/xe/xe_sriov_migration_data.h b/drivers/gpu/drm/xe/xe_sriov_migration_data.h
index 3958f58a170f5..7ec489c3f28d2 100644
--- a/drivers/gpu/drm/xe/xe_sriov_migration_data.h
+++ b/drivers/gpu/drm/xe/xe_sriov_migration_data.h
@@ -26,5 +26,10 @@ void xe_sriov_migration_data_free(struct xe_sriov_migration_data *snapshot);
 int xe_sriov_migration_data_init(struct xe_sriov_migration_data *data, u8 tile_id, u8 gt_id,
 				 enum xe_sriov_migration_data_type, loff_t offset, size_t size);
 int xe_sriov_migration_data_init_from_hdr(struct xe_sriov_migration_data *snapshot);
+ssize_t xe_sriov_migration_data_read(struct xe_device *xe, unsigned int vfid,
+				     char __user *buf, size_t len);
+ssize_t xe_sriov_migration_data_write(struct xe_device *xe, unsigned int vfid,
+				      const char __user *buf, size_t len);
+int xe_sriov_migration_data_save_init(struct xe_device *xe, unsigned int vfid);
 
 #endif
diff --git a/drivers/gpu/drm/xe/xe_sriov_pf_control.c b/drivers/gpu/drm/xe/xe_sriov_pf_control.c
index 8d8a01faf5291..c2768848daba1 100644
--- a/drivers/gpu/drm/xe/xe_sriov_pf_control.c
+++ b/drivers/gpu/drm/xe/xe_sriov_pf_control.c
@@ -5,6 +5,7 @@
 
 #include "xe_device.h"
 #include "xe_gt_sriov_pf_control.h"
+#include "xe_sriov_migration_data.h"
 #include "xe_sriov_pf_control.h"
 #include "xe_sriov_printk.h"
 
@@ -165,6 +166,10 @@ int xe_sriov_pf_control_trigger_save_vf(struct xe_device *xe, unsigned int vfid)
 	unsigned int id;
 	int ret;
 
+	ret = xe_sriov_migration_data_save_init(xe, vfid);
+	if (ret)
+		return ret;
+
 	for_each_gt(gt, xe, id) {
 		ret = xe_gt_sriov_pf_control_trigger_save_vf(gt, vfid);
 		if (ret)
diff --git a/drivers/gpu/drm/xe/xe_sriov_pf_debugfs.c b/drivers/gpu/drm/xe/xe_sriov_pf_debugfs.c
index e0e6340c49106..a9a28aec22421 100644
--- a/drivers/gpu/drm/xe/xe_sriov_pf_debugfs.c
+++ b/drivers/gpu/drm/xe/xe_sriov_pf_debugfs.c
@@ -9,6 +9,7 @@
 #include "xe_device.h"
 #include "xe_device_types.h"
 #include "xe_pm.h"
+#include "xe_sriov_migration_data.h"
 #include "xe_sriov_pf.h"
 #include "xe_sriov_pf_control.h"
 #include "xe_sriov_pf_debugfs.h"
@@ -132,6 +133,7 @@ static void pf_populate_pf(struct xe_device *xe, struct dentry *pfdent)
  *      /sys/kernel/debug/dri/BDF/
  *      ├── sriov
  *      │   ├── vf1
+ *      │   │   ├── migration_data
  *      │   │   ├── pause
  *      │   │   ├── reset
  *      │   │   ├── resume
@@ -220,6 +222,38 @@ DEFINE_VF_CONTROL_ATTRIBUTE(reset_vf);
 DEFINE_VF_CONTROL_ATTRIBUTE_RW(save_vf);
 DEFINE_VF_CONTROL_ATTRIBUTE_RW(restore_vf);
 
+static ssize_t data_write(struct file *file, const char __user *buf, size_t count, loff_t *pos)
+{
+	struct dentry *dent = file_dentry(file)->d_parent;
+	struct xe_device *xe = extract_xe(dent);
+	unsigned int vfid = extract_vfid(dent);
+
+	if (*pos)
+		return -ESPIPE;
+
+	return xe_sriov_migration_data_write(xe, vfid, buf, count);
+}
+
+static ssize_t data_read(struct file *file, char __user *buf, size_t count, loff_t *ppos)
+{
+	struct dentry *dent = file_dentry(file)->d_parent;
+	struct xe_device *xe = extract_xe(dent);
+	unsigned int vfid = extract_vfid(dent);
+
+	if (*ppos)
+		return -ESPIPE;
+
+	return xe_sriov_migration_data_read(xe, vfid, buf, count);
+}
+
+static const struct file_operations data_vf_fops = {
+	.owner		= THIS_MODULE,
+	.open		= simple_open,
+	.write		= data_write,
+	.read		= data_read,
+	.llseek		= default_llseek,
+};
+
 static void pf_populate_vf(struct xe_device *xe, struct dentry *vfdent)
 {
 	debugfs_create_file("pause", 0200, vfdent, xe, &pause_vf_fops);
@@ -228,6 +262,7 @@ static void pf_populate_vf(struct xe_device *xe, struct dentry *vfdent)
 	debugfs_create_file("reset", 0200, vfdent, xe, &reset_vf_fops);
 	debugfs_create_file("save", 0600, vfdent, xe, &save_vf_fops);
 	debugfs_create_file("restore", 0600, vfdent, xe, &restore_vf_fops);
+	debugfs_create_file("migration_data", 0600, vfdent, xe, &data_vf_fops);
 }
 
 static void pf_populate_with_tiles(struct xe_device *xe, struct dentry *dent, unsigned int vfid)
diff --git a/drivers/gpu/drm/xe/xe_sriov_pf_migration.c b/drivers/gpu/drm/xe/xe_sriov_pf_migration.c
index 7be9f026d80e8..8ea531d36f53b 100644
--- a/drivers/gpu/drm/xe/xe_sriov_pf_migration.c
+++ b/drivers/gpu/drm/xe/xe_sriov_pf_migration.c
@@ -10,6 +10,7 @@
 #include "xe_gt_sriov_pf_migration.h"
 #include "xe_pm.h"
 #include "xe_sriov.h"
+#include "xe_sriov_migration_data.h"
 #include "xe_sriov_pf_helpers.h"
 #include "xe_sriov_pf_migration.h"
 #include "xe_sriov_printk.h"
@@ -53,6 +54,15 @@ static bool pf_check_migration_support(struct xe_device *xe)
 	return IS_ENABLED(CONFIG_DRM_XE_DEBUG);
 }
 
+static void pf_migration_cleanup(void *arg)
+{
+	struct xe_sriov_pf_migration *migration = arg;
+
+	xe_sriov_migration_data_free(migration->pending);
+	xe_sriov_migration_data_free(migration->trailer);
+	xe_sriov_migration_data_free(migration->descriptor);
+}
+
 /**
  * xe_sriov_pf_migration_init() - Initialize support for SR-IOV VF migration.
  * @xe: the &xe_device
@@ -62,6 +72,7 @@ static bool pf_check_migration_support(struct xe_device *xe)
 int xe_sriov_pf_migration_init(struct xe_device *xe)
 {
 	unsigned int n, totalvfs;
+	int err;
 
 	xe_assert(xe, IS_SRIOV_PF(xe));
 
@@ -73,7 +84,15 @@ int xe_sriov_pf_migration_init(struct xe_device *xe)
 	for (n = 1; n <= totalvfs; n++) {
 		struct xe_sriov_pf_migration *migration = pf_pick_migration(xe, n);
 
+		err = devm_mutex_init(xe->drm.dev, &migration->lock);
+		if (err)
+			return err;
+
 		init_waitqueue_head(&migration->wq);
+
+		err = devm_add_action_or_reset(xe->drm.dev, pf_migration_cleanup, migration);
+		if (err)
+			return err;
 	}
 
 	return 0;
@@ -153,6 +172,36 @@ xe_sriov_pf_migration_save_consume(struct xe_device *xe, unsigned int vfid)
 	return data;
 }
 
+static int pf_handle_descriptor(struct xe_device *xe, unsigned int vfid,
+				struct xe_sriov_migration_data *data)
+{
+	if (data->tile != 0 || data->gt != 0)
+		return -EINVAL;
+
+	xe_sriov_migration_data_free(data);
+
+	return 0;
+}
+
+static int pf_handle_trailer(struct xe_device *xe, unsigned int vfid,
+			     struct xe_sriov_migration_data *data)
+{
+	struct xe_gt *gt;
+	u8 gt_id;
+
+	if (data->tile != 0 || data->gt != 0)
+		return -EINVAL;
+	if (data->offset != 0 || data->size != 0 || data->buff || data->bo)
+		return -EINVAL;
+
+	xe_sriov_migration_data_free(data);
+
+	for_each_gt(gt, xe, gt_id)
+		xe_gt_sriov_pf_control_restore_data_done(gt, vfid);
+
+	return 0;
+}
+
 /**
  * xe_sriov_pf_migration_restore_produce() - Produce a VF migration data packet to the device.
  * @xe: the &xe_device
@@ -172,6 +221,11 @@ int xe_sriov_pf_migration_restore_produce(struct xe_device *xe, unsigned int vfi
 
 	xe_assert(xe, IS_SRIOV_PF(xe));
 
+	if (data->type == XE_SRIOV_MIGRATION_DATA_TYPE_DESCRIPTOR)
+		return pf_handle_descriptor(xe, vfid, data);
+	if (data->type == XE_SRIOV_MIGRATION_DATA_TYPE_TRAILER)
+		return pf_handle_trailer(xe, vfid, data);
+
 	gt = xe_device_get_gt(xe, data->gt);
 	if (!gt || data->tile != gt->tile->id) {
 		xe_sriov_err_ratelimited(xe, "VF%d Invalid GT - tile:%u, GT:%u\n",
diff --git a/drivers/gpu/drm/xe/xe_sriov_pf_migration_types.h b/drivers/gpu/drm/xe/xe_sriov_pf_migration_types.h
index 2a45ee4e3ece8..8468e5eeb6d66 100644
--- a/drivers/gpu/drm/xe/xe_sriov_pf_migration_types.h
+++ b/drivers/gpu/drm/xe/xe_sriov_pf_migration_types.h
@@ -7,6 +7,7 @@
 #define _XE_SRIOV_PF_MIGRATION_TYPES_H_
 
 #include <linux/types.h>
+#include <linux/mutex_types.h>
 #include <linux/wait.h>
 
 /**
@@ -53,6 +54,14 @@ struct xe_sriov_migration_data {
 struct xe_sriov_pf_migration {
 	/** @wq: waitqueue used to avoid busy-waiting for snapshot production/consumption */
 	wait_queue_head_t wq;
+	/** @lock: Mutex protecting the migration data */
+	struct mutex lock;
+	/** @pending: currently processed data packet of VF resource */
+	struct xe_sriov_migration_data *pending;
+	/** @trailer: data packet used to indicate the end of stream */
+	struct xe_sriov_migration_data *trailer;
+	/** @descriptor: data packet containing the metadata describing the device */
+	struct xe_sriov_migration_data *descriptor;
 };
 
 #endif
-- 
2.50.1

[PATCH v3 08/28] drm/xe/pf: Add minimalistic migration descriptor

[Michał Winiarski]

The descriptor reuses the KLV format used by GuC and contains metadata
that can be used to quickly fail migration when source is incompatible
with destination.

Signed-off-by: Michał Winiarski <michal.winiarski@intel.com>
---
 drivers/gpu/drm/xe/xe_sriov_migration_data.c | 89 +++++++++++++++++++-
 drivers/gpu/drm/xe/xe_sriov_migration_data.h |  2 +
 drivers/gpu/drm/xe/xe_sriov_pf_migration.c   |  6 ++
 3 files changed, 96 insertions(+), 1 deletion(-)

diff --git a/drivers/gpu/drm/xe/xe_sriov_migration_data.c b/drivers/gpu/drm/xe/xe_sriov_migration_data.c
index a3f50836adc81..18e17706772fc 100644
--- a/drivers/gpu/drm/xe/xe_sriov_migration_data.c
+++ b/drivers/gpu/drm/xe/xe_sriov_migration_data.c
@@ -5,6 +5,7 @@
 
 #include "xe_bo.h"
 #include "xe_device.h"
+#include "xe_guc_klv_helpers.h"
 #include "xe_sriov_migration_data.h"
 #include "xe_sriov_pf_helpers.h"
 #include "xe_sriov_pf_migration.h"
@@ -383,11 +384,19 @@ ssize_t xe_sriov_migration_data_write(struct xe_device *xe, unsigned int vfid,
 	return produced;
 }
 
-#define MIGRATION_DESCRIPTOR_DWORDS 0
+#define MIGRATION_KLV_DEVICE_DEVID_KEY	0xf001u
+#define MIGRATION_KLV_DEVICE_DEVID_LEN	1u
+#define MIGRATION_KLV_DEVICE_REVID_KEY	0xf002u
+#define MIGRATION_KLV_DEVICE_REVID_LEN	1u
+
+#define MIGRATION_DESCRIPTOR_DWORDS	(GUC_KLV_LEN_MIN + MIGRATION_KLV_DEVICE_DEVID_LEN + \
+					 GUC_KLV_LEN_MIN + MIGRATION_KLV_DEVICE_REVID_LEN)
 static size_t pf_descriptor_init(struct xe_device *xe, unsigned int vfid)
 {
 	struct xe_sriov_migration_data **desc = pf_pick_descriptor(xe, vfid);
 	struct xe_sriov_migration_data *data;
+	unsigned int len = 0;
+	u32 *klvs;
 	int ret;
 
 	data = xe_sriov_migration_data_alloc(xe);
@@ -401,11 +410,89 @@ static size_t pf_descriptor_init(struct xe_device *xe, unsigned int vfid)
 		return ret;
 	}
 
+	klvs = data->vaddr;
+	klvs[len++] = PREP_GUC_KLV_CONST(MIGRATION_KLV_DEVICE_DEVID_KEY,
+					 MIGRATION_KLV_DEVICE_DEVID_LEN);
+	klvs[len++] = xe->info.devid;
+	klvs[len++] = PREP_GUC_KLV_CONST(MIGRATION_KLV_DEVICE_REVID_KEY,
+					 MIGRATION_KLV_DEVICE_REVID_LEN);
+	klvs[len++] = xe->info.revid;
+
+	xe_assert(xe, len == MIGRATION_DESCRIPTOR_DWORDS);
+
 	*desc = data;
 
 	return 0;
 }
 
+/**
+ * xe_sriov_migration_data_process_descriptor() - Process migration data descriptor.
+ * @xe: the &xe_device
+ * @vfid: the VF identifier
+ * @data: the &struct xe_sriov_pf_migration_data containing the descriptor
+ *
+ * The descriptor uses the same KLV format as GuC, and contains metadata used for
+ * checking migration data compatibility.
+ *
+ * Return: 0 on success, -errno on failure.
+ */
+int xe_sriov_migration_data_process_descriptor(struct xe_device *xe, unsigned int vfid,
+					       struct xe_sriov_migration_data *data)
+{
+	u32 num_dwords = data->size / sizeof(u32);
+	u32 *klvs = data->vaddr;
+
+	xe_assert(xe, data->type == XE_SRIOV_MIGRATION_DATA_TYPE_DESCRIPTOR);
+
+	if (data->size % sizeof(u32)) {
+		xe_sriov_warn(xe, "Aborting migration, descriptor not in KLV format (size=%llu)\n",
+			      data->size);
+		return -EINVAL;
+	}
+
+	while (num_dwords >= GUC_KLV_LEN_MIN) {
+		u32 key = FIELD_GET(GUC_KLV_0_KEY, klvs[0]);
+		u32 len = FIELD_GET(GUC_KLV_0_LEN, klvs[0]);
+
+		klvs += GUC_KLV_LEN_MIN;
+		num_dwords -= GUC_KLV_LEN_MIN;
+
+		if (len > num_dwords)
+			return -EINVAL;
+
+		switch (key) {
+		case MIGRATION_KLV_DEVICE_DEVID_KEY:
+			if (*klvs != xe->info.devid) {
+				xe_sriov_warn(xe,
+					      "Aborting migration, devid mismatch %#06x!=%#06x\n",
+					      *klvs, xe->info.devid);
+				return -ENODEV;
+			}
+			break;
+		case MIGRATION_KLV_DEVICE_REVID_KEY:
+			if (*klvs != xe->info.revid) {
+				xe_sriov_warn(xe,
+					      "Aborting migration, revid mismatch %#06x!=%#06x\n",
+					      *klvs, xe->info.revid);
+				return -ENODEV;
+			}
+			break;
+		default:
+			xe_sriov_dbg(xe,
+				     "Skipping unknown migration descriptor key %#06x (len=%#06x)\n",
+				     key, len);
+			print_hex_dump_bytes("desc: ", DUMP_PREFIX_OFFSET, klvs,
+					     min(SZ_64, len * sizeof(u32)));
+			break;
+		}
+
+		klvs += len;
+		num_dwords -= len;
+	}
+
+	return 0;
+}
+
 static void pf_pending_init(struct xe_device *xe, unsigned int vfid)
 {
 	struct xe_sriov_migration_data **data = pf_pick_pending(xe, vfid);
diff --git a/drivers/gpu/drm/xe/xe_sriov_migration_data.h b/drivers/gpu/drm/xe/xe_sriov_migration_data.h
index 7ec489c3f28d2..bb4ea5850e5c0 100644
--- a/drivers/gpu/drm/xe/xe_sriov_migration_data.h
+++ b/drivers/gpu/drm/xe/xe_sriov_migration_data.h
@@ -30,6 +30,8 @@ ssize_t xe_sriov_migration_data_read(struct xe_device *xe, unsigned int vfid,
 				     char __user *buf, size_t len);
 ssize_t xe_sriov_migration_data_write(struct xe_device *xe, unsigned int vfid,
 				      const char __user *buf, size_t len);
+int xe_sriov_migration_data_process_descriptor(struct xe_device *xe, unsigned int vfid,
+					       struct xe_sriov_migration_data *data);
 int xe_sriov_migration_data_save_init(struct xe_device *xe, unsigned int vfid);
 
 #endif
diff --git a/drivers/gpu/drm/xe/xe_sriov_pf_migration.c b/drivers/gpu/drm/xe/xe_sriov_pf_migration.c
index 8ea531d36f53b..f0a0c2b027a20 100644
--- a/drivers/gpu/drm/xe/xe_sriov_pf_migration.c
+++ b/drivers/gpu/drm/xe/xe_sriov_pf_migration.c
@@ -175,9 +175,15 @@ xe_sriov_pf_migration_save_consume(struct xe_device *xe, unsigned int vfid)
 static int pf_handle_descriptor(struct xe_device *xe, unsigned int vfid,
 				struct xe_sriov_migration_data *data)
 {
+	int ret;
+
 	if (data->tile != 0 || data->gt != 0)
 		return -EINVAL;
 
+	ret = xe_sriov_migration_data_process_descriptor(xe, vfid, data);
+	if (ret)
+		return ret;
+
 	xe_sriov_migration_data_free(data);
 
 	return 0;
-- 
2.50.1

[PATCH v3 09/28] drm/xe/pf: Expose VF migration data size over debugfs

[Michał Winiarski]

The size is normally used to make a decision on when to stop the device
(mainly when it's in a pre_copy state).

Signed-off-by: Michał Winiarski <michal.winiarski@intel.com>
Reviewed-by: Michal Wajdeczko <michal.wajdeczko@intel.com>
---
 drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.c | 21 +++++++++++++
 drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.h |  2 ++
 drivers/gpu/drm/xe/xe_sriov_pf_debugfs.c      | 29 ++++++++++++++++++
 drivers/gpu/drm/xe/xe_sriov_pf_migration.c    | 30 +++++++++++++++++++
 drivers/gpu/drm/xe/xe_sriov_pf_migration.h    |  1 +
 5 files changed, 83 insertions(+)

diff --git a/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.c b/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.c
index 40b75b5fe9229..30f0e51da49ae 100644
--- a/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.c
+++ b/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.c
@@ -395,6 +395,27 @@ ssize_t xe_gt_sriov_pf_migration_write_guc_state(struct xe_gt *gt, unsigned int
 }
 #endif /* CONFIG_DEBUG_FS */
 
+/**
+ * xe_gt_sriov_pf_migration_size() - Total size of migration data from all components within a GT.
+ * @gt: the &xe_gt
+ * @vfid: the VF identifier (can't be 0)
+ *
+ * This function is for PF only.
+ *
+ * Return: total migration data size in bytes or a negative error code on failure.
+ */
+ssize_t xe_gt_sriov_pf_migration_size(struct xe_gt *gt, unsigned int vfid)
+{
+	ssize_t total = 0;
+
+	xe_gt_assert(gt, IS_SRIOV_PF(gt_to_xe(gt)));
+	xe_gt_assert(gt, vfid != PFID);
+	xe_gt_assert(gt, vfid <= xe_sriov_pf_get_totalvfs(gt_to_xe(gt)));
+
+	/* Nothing to query yet - will be updated once per-GT migration data types are added */
+	return total;
+}
+
 /**
  * xe_gt_sriov_pf_migration_ring_empty() - Check if a migration ring is empty.
  * @gt: the &xe_gt
diff --git a/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.h b/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.h
index 1ed2248f0a17e..e2d41750f863c 100644
--- a/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.h
+++ b/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.h
@@ -15,6 +15,8 @@ int xe_gt_sriov_pf_migration_init(struct xe_gt *gt);
 int xe_gt_sriov_pf_migration_save_guc_state(struct xe_gt *gt, unsigned int vfid);
 int xe_gt_sriov_pf_migration_restore_guc_state(struct xe_gt *gt, unsigned int vfid);
 
+ssize_t xe_gt_sriov_pf_migration_size(struct xe_gt *gt, unsigned int vfid);
+
 bool xe_gt_sriov_pf_migration_ring_empty(struct xe_gt *gt, unsigned int vfid);
 bool xe_gt_sriov_pf_migration_ring_full(struct xe_gt *gt, unsigned int vfid);
 void xe_gt_sriov_pf_migration_ring_free(struct xe_gt *gt, unsigned int vfid);
diff --git a/drivers/gpu/drm/xe/xe_sriov_pf_debugfs.c b/drivers/gpu/drm/xe/xe_sriov_pf_debugfs.c
index a9a28aec22421..bc2d0b0342f22 100644
--- a/drivers/gpu/drm/xe/xe_sriov_pf_debugfs.c
+++ b/drivers/gpu/drm/xe/xe_sriov_pf_debugfs.c
@@ -14,6 +14,7 @@
 #include "xe_sriov_pf_control.h"
 #include "xe_sriov_pf_debugfs.h"
 #include "xe_sriov_pf_helpers.h"
+#include "xe_sriov_pf_migration.h"
 #include "xe_sriov_pf_provision.h"
 #include "xe_sriov_pf_service.h"
 #include "xe_sriov_printk.h"
@@ -254,6 +255,33 @@ static const struct file_operations data_vf_fops = {
 	.llseek		= default_llseek,
 };
 
+static ssize_t size_read(struct file *file, char __user *ubuf, size_t count, loff_t *ppos)
+{
+	struct dentry *dent = file_dentry(file)->d_parent;
+	struct xe_device *xe = extract_xe(dent);
+	unsigned int vfid = extract_vfid(dent);
+	char buf[21];
+	ssize_t ret;
+	int len;
+
+	xe_pm_runtime_get(xe);
+	ret = xe_sriov_pf_migration_size(xe, vfid);
+	xe_pm_runtime_put(xe);
+	if (ret < 0)
+		return ret;
+
+	len = scnprintf(buf, sizeof(buf), "%zd\n", ret);
+
+	return simple_read_from_buffer(ubuf, count, ppos, buf, len);
+}
+
+static const struct file_operations size_vf_fops = {
+	.owner		= THIS_MODULE,
+	.open		= simple_open,
+	.read		= size_read,
+	.llseek		= default_llseek,
+};
+
 static void pf_populate_vf(struct xe_device *xe, struct dentry *vfdent)
 {
 	debugfs_create_file("pause", 0200, vfdent, xe, &pause_vf_fops);
@@ -263,6 +291,7 @@ static void pf_populate_vf(struct xe_device *xe, struct dentry *vfdent)
 	debugfs_create_file("save", 0600, vfdent, xe, &save_vf_fops);
 	debugfs_create_file("restore", 0600, vfdent, xe, &restore_vf_fops);
 	debugfs_create_file("migration_data", 0600, vfdent, xe, &data_vf_fops);
+	debugfs_create_file("migration_size", 0400, vfdent, xe, &size_vf_fops);
 }
 
 static void pf_populate_with_tiles(struct xe_device *xe, struct dentry *dent, unsigned int vfid)
diff --git a/drivers/gpu/drm/xe/xe_sriov_pf_migration.c b/drivers/gpu/drm/xe/xe_sriov_pf_migration.c
index f0a0c2b027a20..6992c227e5a44 100644
--- a/drivers/gpu/drm/xe/xe_sriov_pf_migration.c
+++ b/drivers/gpu/drm/xe/xe_sriov_pf_migration.c
@@ -241,3 +241,33 @@ int xe_sriov_pf_migration_restore_produce(struct xe_device *xe, unsigned int vfi
 
 	return xe_gt_sriov_pf_migration_restore_produce(gt, vfid, data);
 }
+
+/**
+ * xe_sriov_pf_migration_size() - Total size of migration data from all components within a device
+ * @xe: the &xe_device
+ * @vfid: the VF identifier (can't be 0)
+ *
+ * This function is for PF only.
+ *
+ * Return: total migration data size in bytes or a negative error code on failure.
+ */
+ssize_t xe_sriov_pf_migration_size(struct xe_device *xe, unsigned int vfid)
+{
+	size_t size = 0;
+	struct xe_gt *gt;
+	ssize_t ret;
+	u8 gt_id;
+
+	xe_assert(xe, IS_SRIOV_PF(xe));
+	xe_assert(xe, vfid);
+
+	for_each_gt(gt, xe, gt_id) {
+		ret = xe_gt_sriov_pf_migration_size(gt, vfid);
+		if (ret < 0)
+			return ret;
+
+		size += ret;
+	}
+
+	return size;
+}
diff --git a/drivers/gpu/drm/xe/xe_sriov_pf_migration.h b/drivers/gpu/drm/xe/xe_sriov_pf_migration.h
index df81a540c246a..16cb444c36aa6 100644
--- a/drivers/gpu/drm/xe/xe_sriov_pf_migration.h
+++ b/drivers/gpu/drm/xe/xe_sriov_pf_migration.h
@@ -18,6 +18,7 @@ int xe_sriov_pf_migration_restore_produce(struct xe_device *xe, unsigned int vfi
 					  struct xe_sriov_migration_data *data);
 struct xe_sriov_migration_data *
 xe_sriov_pf_migration_save_consume(struct xe_device *xe, unsigned int vfid);
+ssize_t xe_sriov_pf_migration_size(struct xe_device *xe, unsigned int vfid);
 wait_queue_head_t *xe_sriov_pf_migration_waitqueue(struct xe_device *xe, unsigned int vfid);
 
 #endif
-- 
2.50.1

[PATCH v3 10/28] drm/xe: Add sa/guc_buf_cache sync interface

[Michał Winiarski]

In upcoming changes the cached buffers are going to be used to read data
produced by the GuC. Add a counterpart to flush, which synchronizes the
CPU-side of suballocation with the GPU data and propagate the interface
to GuC Buffer Cache.

Signed-off-by: Michał Winiarski <michal.winiarski@intel.com>
Reviewed-by: Michal Wajdeczko <michal.wajdeczko@intel.com>
---
 drivers/gpu/drm/xe/xe_guc_buf.c | 13 +++++++++++++
 drivers/gpu/drm/xe/xe_guc_buf.h |  1 +
 drivers/gpu/drm/xe/xe_sa.c      | 21 +++++++++++++++++++++
 drivers/gpu/drm/xe/xe_sa.h      |  1 +
 4 files changed, 36 insertions(+)

diff --git a/drivers/gpu/drm/xe/xe_guc_buf.c b/drivers/gpu/drm/xe/xe_guc_buf.c
index 502ca3a4ee606..4d8a4712309f4 100644
--- a/drivers/gpu/drm/xe/xe_guc_buf.c
+++ b/drivers/gpu/drm/xe/xe_guc_buf.c
@@ -115,6 +115,19 @@ void xe_guc_buf_release(const struct xe_guc_buf buf)
 		xe_sa_bo_free(buf.sa, NULL);
 }
 
+/**
+ * xe_guc_buf_sync_read() - Copy the data from the GPU memory to the sub-allocation.
+ * @buf: the &xe_guc_buf to sync
+ *
+ * Return: a CPU pointer of the sub-allocation.
+ */
+void *xe_guc_buf_sync_read(const struct xe_guc_buf buf)
+{
+	xe_sa_bo_sync_read(buf.sa);
+
+	return xe_sa_bo_cpu_addr(buf.sa);
+}
+
 /**
  * xe_guc_buf_flush() - Copy the data from the sub-allocation to the GPU memory.
  * @buf: the &xe_guc_buf to flush
diff --git a/drivers/gpu/drm/xe/xe_guc_buf.h b/drivers/gpu/drm/xe/xe_guc_buf.h
index 0d67604d96bdd..c5e0f1fd24d74 100644
--- a/drivers/gpu/drm/xe/xe_guc_buf.h
+++ b/drivers/gpu/drm/xe/xe_guc_buf.h
@@ -30,6 +30,7 @@ static inline bool xe_guc_buf_is_valid(const struct xe_guc_buf buf)
 }
 
 void *xe_guc_buf_cpu_ptr(const struct xe_guc_buf buf);
+void *xe_guc_buf_sync_read(const struct xe_guc_buf buf);
 u64 xe_guc_buf_flush(const struct xe_guc_buf buf);
 u64 xe_guc_buf_gpu_addr(const struct xe_guc_buf buf);
 u64 xe_guc_cache_gpu_addr_from_ptr(struct xe_guc_buf_cache *cache, const void *ptr, u32 size);
diff --git a/drivers/gpu/drm/xe/xe_sa.c b/drivers/gpu/drm/xe/xe_sa.c
index fedd017d6dd36..63a5263dcf1b1 100644
--- a/drivers/gpu/drm/xe/xe_sa.c
+++ b/drivers/gpu/drm/xe/xe_sa.c
@@ -110,6 +110,10 @@ struct drm_suballoc *__xe_sa_bo_new(struct xe_sa_manager *sa_manager, u32 size,
 	return drm_suballoc_new(&sa_manager->base, size, gfp, true, 0);
 }
 
+/**
+ * xe_sa_bo_flush_write() - Copy the data from the sub-allocation to the GPU memory.
+ * @sa_bo: the &drm_suballoc to flush
+ */
 void xe_sa_bo_flush_write(struct drm_suballoc *sa_bo)
 {
 	struct xe_sa_manager *sa_manager = to_xe_sa_manager(sa_bo->manager);
@@ -123,6 +127,23 @@ void xe_sa_bo_flush_write(struct drm_suballoc *sa_bo)
 			 drm_suballoc_size(sa_bo));
 }
 
+/**
+ * xe_sa_bo_sync_read() - Copy the data from GPU memory to the sub-allocation.
+ * @sa_bo: the &drm_suballoc to sync
+ */
+void xe_sa_bo_sync_read(struct drm_suballoc *sa_bo)
+{
+	struct xe_sa_manager *sa_manager = to_xe_sa_manager(sa_bo->manager);
+	struct xe_device *xe = tile_to_xe(sa_manager->bo->tile);
+
+	if (!sa_manager->bo->vmap.is_iomem)
+		return;
+
+	xe_map_memcpy_from(xe, xe_sa_bo_cpu_addr(sa_bo), &sa_manager->bo->vmap,
+			   drm_suballoc_soffset(sa_bo),
+			   drm_suballoc_size(sa_bo));
+}
+
 void xe_sa_bo_free(struct drm_suballoc *sa_bo,
 		   struct dma_fence *fence)
 {
diff --git a/drivers/gpu/drm/xe/xe_sa.h b/drivers/gpu/drm/xe/xe_sa.h
index 99dbf0eea5402..1be7443508361 100644
--- a/drivers/gpu/drm/xe/xe_sa.h
+++ b/drivers/gpu/drm/xe/xe_sa.h
@@ -37,6 +37,7 @@ static inline struct drm_suballoc *xe_sa_bo_new(struct xe_sa_manager *sa_manager
 }
 
 void xe_sa_bo_flush_write(struct drm_suballoc *sa_bo);
+void xe_sa_bo_sync_read(struct drm_suballoc *sa_bo);
 void xe_sa_bo_free(struct drm_suballoc *sa_bo, struct dma_fence *fence);
 
 static inline struct xe_sa_manager *
-- 
2.50.1

[PATCH v3 11/28] drm/xe: Allow the caller to pass guc_buf_cache size

[Michał Winiarski]

An upcoming change will use GuC buffer cache as a place where GuC
migration data will be stored, and the memory requirement for that is
larger than indirect data.
Allow the caller to pass the size based on the intended usecase.

Signed-off-by: Michał Winiarski <michal.winiarski@intel.com>
Reviewed-by: Michal Wajdeczko <michal.wajdeczko@intel.com>
---
 drivers/gpu/drm/xe/xe_guc_buf.c | 44 ++++++++++++++++++++++++---------
 drivers/gpu/drm/xe/xe_guc_buf.h |  1 +
 2 files changed, 34 insertions(+), 11 deletions(-)

diff --git a/drivers/gpu/drm/xe/xe_guc_buf.c b/drivers/gpu/drm/xe/xe_guc_buf.c
index 4d8a4712309f4..3ce4425001303 100644
--- a/drivers/gpu/drm/xe/xe_guc_buf.c
+++ b/drivers/gpu/drm/xe/xe_guc_buf.c
@@ -13,6 +13,8 @@
 #include "xe_guc_buf.h"
 #include "xe_sa.h"
 
+#define XE_GUC_BUF_CACHE_DEFAULT_SIZE SZ_8K
+
 static struct xe_guc *cache_to_guc(struct xe_guc_buf_cache *cache)
 {
 	return container_of(cache, struct xe_guc, buf);
@@ -23,21 +25,12 @@ static struct xe_gt *cache_to_gt(struct xe_guc_buf_cache *cache)
 	return guc_to_gt(cache_to_guc(cache));
 }
 
-/**
- * xe_guc_buf_cache_init() - Initialize the GuC Buffer Cache.
- * @cache: the &xe_guc_buf_cache to initialize
- *
- * The Buffer Cache allows to obtain a reusable buffer that can be used to pass
- * indirect H2G data to GuC without a need to create a ad-hoc allocation.
- *
- * Return: 0 on success or a negative error code on failure.
- */
-int xe_guc_buf_cache_init(struct xe_guc_buf_cache *cache)
+static int guc_buf_cache_init(struct xe_guc_buf_cache *cache, u32 size)
 {
 	struct xe_gt *gt = cache_to_gt(cache);
 	struct xe_sa_manager *sam;
 
-	sam = __xe_sa_bo_manager_init(gt_to_tile(gt), SZ_8K, 0, sizeof(u32));
+	sam = __xe_sa_bo_manager_init(gt_to_tile(gt), size, 0, sizeof(u32));
 	if (IS_ERR(sam))
 		return PTR_ERR(sam);
 	cache->sam = sam;
@@ -48,6 +41,35 @@ int xe_guc_buf_cache_init(struct xe_guc_buf_cache *cache)
 	return 0;
 }
 
+/**
+ * xe_guc_buf_cache_init() - Initialize the GuC Buffer Cache.
+ * @cache: the &xe_guc_buf_cache to initialize
+ *
+ * The Buffer Cache allows to obtain a reusable buffer that can be used to pass
+ * data to GuC or read data from GuC without a need to create a ad-hoc allocation.
+ *
+ * Return: 0 on success or a negative error code on failure.
+ */
+int xe_guc_buf_cache_init(struct xe_guc_buf_cache *cache)
+{
+	return guc_buf_cache_init(cache, XE_GUC_BUF_CACHE_DEFAULT_SIZE);
+}
+
+/**
+ * xe_guc_buf_cache_init_with_size() - Initialize the GuC Buffer Cache.
+ * @cache: the &xe_guc_buf_cache to initialize
+ * @size: size in bytes
+ *
+ * Like xe_guc_buf_cache_init(), except it allows the caller to make the cache
+ * buffer larger, allowing to accommodate larger objects.
+ *
+ * Return: 0 on success or a negative error code on failure.
+ */
+int xe_guc_buf_cache_init_with_size(struct xe_guc_buf_cache *cache, u32 size)
+{
+	return guc_buf_cache_init(cache, max(XE_GUC_BUF_CACHE_DEFAULT_SIZE, size));
+}
+
 /**
  * xe_guc_buf_cache_dwords() - Number of dwords the GuC Buffer Cache supports.
  * @cache: the &xe_guc_buf_cache to query
diff --git a/drivers/gpu/drm/xe/xe_guc_buf.h b/drivers/gpu/drm/xe/xe_guc_buf.h
index c5e0f1fd24d74..e3cca553fb004 100644
--- a/drivers/gpu/drm/xe/xe_guc_buf.h
+++ b/drivers/gpu/drm/xe/xe_guc_buf.h
@@ -12,6 +12,7 @@
 #include "xe_guc_buf_types.h"
 
 int xe_guc_buf_cache_init(struct xe_guc_buf_cache *cache);
+int xe_guc_buf_cache_init_with_size(struct xe_guc_buf_cache *cache, u32 size);
 u32 xe_guc_buf_cache_dwords(struct xe_guc_buf_cache *cache);
 struct xe_guc_buf xe_guc_buf_reserve(struct xe_guc_buf_cache *cache, u32 dwords);
 struct xe_guc_buf xe_guc_buf_from_data(struct xe_guc_buf_cache *cache,
-- 
2.50.1

[PATCH v3 12/28] drm/xe/pf: Increase PF GuC Buffer Cache size and use it for VF migration

[Michał Winiarski]

Contiguous PF GGTT VMAs can be scarce after creating VFs.
Increase the GuC buffer cache size to 8M for PF so that we can fit GuC
migration data (which currently maxes out at just under 4M) and use the
cache instead of allocating fresh BOs.

Signed-off-by: Michał Winiarski <michal.winiarski@intel.com>
---
 drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.c | 46 ++++++-------------
 drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.h |  3 ++
 drivers/gpu/drm/xe/xe_guc.c                   | 12 ++++-
 3 files changed, 28 insertions(+), 33 deletions(-)

diff --git a/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.c b/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.c
index 30f0e51da49ae..a2db127982638 100644
--- a/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.c
+++ b/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.c
@@ -11,7 +11,7 @@
 #include "xe_gt_sriov_pf_helpers.h"
 #include "xe_gt_sriov_pf_migration.h"
 #include "xe_gt_sriov_printk.h"
-#include "xe_guc.h"
+#include "xe_guc_buf.h"
 #include "xe_guc_ct.h"
 #include "xe_sriov.h"
 #include "xe_sriov_migration_data.h"
@@ -57,73 +57,55 @@ static int pf_send_guc_query_vf_state_size(struct xe_gt *gt, unsigned int vfid)
 
 /* Return: number of state dwords saved or a negative error code on failure */
 static int pf_send_guc_save_vf_state(struct xe_gt *gt, unsigned int vfid,
-				     void *buff, size_t size)
+				     void *dst, size_t size)
 {
 	const int ndwords = size / sizeof(u32);
-	struct xe_tile *tile = gt_to_tile(gt);
-	struct xe_device *xe = tile_to_xe(tile);
 	struct xe_guc *guc = &gt->uc.guc;
-	struct xe_bo *bo;
+	CLASS(xe_guc_buf, buf)(&guc->buf, ndwords);
 	int ret;
 
 	xe_gt_assert(gt, size % sizeof(u32) == 0);
 	xe_gt_assert(gt, size == ndwords * sizeof(u32));
 
-	bo = xe_bo_create_pin_map_novm(xe, tile,
-				       ALIGN(size, PAGE_SIZE),
-				       ttm_bo_type_kernel,
-				       XE_BO_FLAG_SYSTEM |
-				       XE_BO_FLAG_GGTT |
-				       XE_BO_FLAG_GGTT_INVALIDATE, false);
-	if (IS_ERR(bo))
-		return PTR_ERR(bo);
+	if (!xe_guc_buf_is_valid(buf))
+		return -ENOBUFS;
+
+	memset(xe_guc_buf_cpu_ptr(buf), 0, size);
 
 	ret = guc_action_vf_save_restore(guc, vfid, GUC_PF_OPCODE_VF_SAVE,
-					 xe_bo_ggtt_addr(bo), ndwords);
+					 xe_guc_buf_flush(buf), ndwords);
 	if (!ret)
 		ret = -ENODATA;
 	else if (ret > ndwords)
 		ret = -EPROTO;
 	else if (ret > 0)
-		xe_map_memcpy_from(xe, buff, &bo->vmap, 0, ret * sizeof(u32));
+		memcpy(dst, xe_guc_buf_sync_read(buf), ret * sizeof(u32));
 
-	xe_bo_unpin_map_no_vm(bo);
 	return ret;
 }
 
 /* Return: number of state dwords restored or a negative error code on failure */
 static int pf_send_guc_restore_vf_state(struct xe_gt *gt, unsigned int vfid,
-					const void *buff, size_t size)
+					const void *src, size_t size)
 {
 	const int ndwords = size / sizeof(u32);
-	struct xe_tile *tile = gt_to_tile(gt);
-	struct xe_device *xe = tile_to_xe(tile);
 	struct xe_guc *guc = &gt->uc.guc;
-	struct xe_bo *bo;
+	CLASS(xe_guc_buf_from_data, buf)(&guc->buf, src, size);
 	int ret;
 
 	xe_gt_assert(gt, size % sizeof(u32) == 0);
 	xe_gt_assert(gt, size == ndwords * sizeof(u32));
 
-	bo = xe_bo_create_pin_map_novm(xe, tile,
-				       ALIGN(size, PAGE_SIZE),
-				       ttm_bo_type_kernel,
-				       XE_BO_FLAG_SYSTEM |
-				       XE_BO_FLAG_GGTT |
-				       XE_BO_FLAG_GGTT_INVALIDATE, false);
-	if (IS_ERR(bo))
-		return PTR_ERR(bo);
-
-	xe_map_memcpy_to(xe, &bo->vmap, 0, buff, size);
+	if (!xe_guc_buf_is_valid(buf))
+		return -ENOBUFS;
 
 	ret = guc_action_vf_save_restore(guc, vfid, GUC_PF_OPCODE_VF_RESTORE,
-					 xe_bo_ggtt_addr(bo), ndwords);
+					 xe_guc_buf_flush(buf), ndwords);
 	if (!ret)
 		ret = -ENODATA;
 	else if (ret > ndwords)
 		ret = -EPROTO;
 
-	xe_bo_unpin_map_no_vm(bo);
 	return ret;
 }
 
diff --git a/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.h b/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.h
index e2d41750f863c..4f2f2783339c3 100644
--- a/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.h
+++ b/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.h
@@ -11,6 +11,9 @@
 struct xe_gt;
 struct xe_sriov_migration_data;
 
+/* TODO: get this information by querying GuC in the future */
+#define XE_GT_SRIOV_PF_MIGRATION_GUC_DATA_MAX_SIZE SZ_8M
+
 int xe_gt_sriov_pf_migration_init(struct xe_gt *gt);
 int xe_gt_sriov_pf_migration_save_guc_state(struct xe_gt *gt, unsigned int vfid);
 int xe_gt_sriov_pf_migration_restore_guc_state(struct xe_gt *gt, unsigned int vfid);
diff --git a/drivers/gpu/drm/xe/xe_guc.c b/drivers/gpu/drm/xe/xe_guc.c
index ecc3e091b89e6..badae9a26220e 100644
--- a/drivers/gpu/drm/xe/xe_guc.c
+++ b/drivers/gpu/drm/xe/xe_guc.c
@@ -24,6 +24,7 @@
 #include "xe_gt_printk.h"
 #include "xe_gt_sriov_vf.h"
 #include "xe_gt_throttle.h"
+#include "xe_gt_sriov_pf_migration.h"
 #include "xe_guc_ads.h"
 #include "xe_guc_buf.h"
 #include "xe_guc_capture.h"
@@ -40,6 +41,7 @@
 #include "xe_mmio.h"
 #include "xe_platform_types.h"
 #include "xe_sriov.h"
+#include "xe_sriov_pf_migration.h"
 #include "xe_uc.h"
 #include "xe_uc_fw.h"
 #include "xe_wa.h"
@@ -821,6 +823,14 @@ static int vf_guc_init_post_hwconfig(struct xe_guc *guc)
 	return 0;
 }
 
+static u32 guc_additional_cache_size(struct xe_device *xe)
+{
+	if (IS_SRIOV_PF(xe) && xe_sriov_pf_migration_supported(xe))
+		return XE_GT_SRIOV_PF_MIGRATION_GUC_DATA_MAX_SIZE;
+	else
+		return 0; /* Fallback to default size */
+}
+
 /**
  * xe_guc_init_post_hwconfig - initialize GuC post hwconfig load
  * @guc: The GuC object
@@ -860,7 +870,7 @@ int xe_guc_init_post_hwconfig(struct xe_guc *guc)
 	if (ret)
 		return ret;
 
-	ret = xe_guc_buf_cache_init(&guc->buf);
+	ret = xe_guc_buf_cache_init_with_size(&guc->buf, guc_additional_cache_size(guc_to_xe(guc)));
 	if (ret)
 		return ret;
 
-- 
2.50.1

[PATCH v3 13/28] drm/xe/pf: Remove GuC migration data save/restore from GT debugfs

[Michał Winiarski]

In upcoming changes, SR-IOV VF migration data will be extended beyond
GuC data and exported to userspace using VFIO interface (with a
vendor-specific variant driver) and a device-level debugfs interface.
Remove the GT-level debugfs.

Signed-off-by: Michał Winiarski <michal.winiarski@intel.com>
Reviewed-by: Michal Wajdeczko <michal.wajdeczko@intel.com>
---
 drivers/gpu/drm/xe/xe_gt_sriov_pf_debugfs.c | 47 ---------------------
 1 file changed, 47 deletions(-)

diff --git a/drivers/gpu/drm/xe/xe_gt_sriov_pf_debugfs.c b/drivers/gpu/drm/xe/xe_gt_sriov_pf_debugfs.c
index 838beb7f6327f..5278ea4fd6552 100644
--- a/drivers/gpu/drm/xe/xe_gt_sriov_pf_debugfs.c
+++ b/drivers/gpu/drm/xe/xe_gt_sriov_pf_debugfs.c
@@ -327,9 +327,6 @@ static const struct {
 	{ "stop", xe_gt_sriov_pf_control_stop_vf },
 	{ "pause", xe_gt_sriov_pf_control_pause_vf },
 	{ "resume", xe_gt_sriov_pf_control_resume_vf },
-#ifdef CONFIG_DRM_XE_DEBUG_SRIOV
-	{ "restore!", xe_gt_sriov_pf_migration_restore_guc_state },
-#endif
 };
 
 static ssize_t control_write(struct file *file, const char __user *buf, size_t count, loff_t *pos)
@@ -393,47 +390,6 @@ static const struct file_operations control_ops = {
 	.llseek		= default_llseek,
 };
 
-/*
- *      /sys/kernel/debug/dri/BDF/
- *      ├── sriov
- *      :   ├── vf1
- *          :   ├── tile0
- *              :   ├── gt0
- *                  :   ├── guc_state
- */
-
-static ssize_t guc_state_read(struct file *file, char __user *buf,
-			      size_t count, loff_t *pos)
-{
-	struct dentry *dent = file_dentry(file);
-	struct dentry *parent = dent->d_parent;
-	struct xe_gt *gt = extract_gt(parent);
-	unsigned int vfid = extract_vfid(parent);
-
-	return xe_gt_sriov_pf_migration_read_guc_state(gt, vfid, buf, count, pos);
-}
-
-static ssize_t guc_state_write(struct file *file, const char __user *buf,
-			       size_t count, loff_t *pos)
-{
-	struct dentry *dent = file_dentry(file);
-	struct dentry *parent = dent->d_parent;
-	struct xe_gt *gt = extract_gt(parent);
-	unsigned int vfid = extract_vfid(parent);
-
-	if (*pos)
-		return -EINVAL;
-
-	return xe_gt_sriov_pf_migration_write_guc_state(gt, vfid, buf, count);
-}
-
-static const struct file_operations guc_state_ops = {
-	.owner		= THIS_MODULE,
-	.read		= guc_state_read,
-	.write		= guc_state_write,
-	.llseek		= default_llseek,
-};
-
 /*
  *      /sys/kernel/debug/dri/BDF/
  *      ├── sriov
@@ -568,9 +524,6 @@ static void pf_populate_gt(struct xe_gt *gt, struct dentry *dent, unsigned int v
 
 		/* for testing/debugging purposes only! */
 		if (IS_ENABLED(CONFIG_DRM_XE_DEBUG)) {
-			debugfs_create_file("guc_state",
-					    IS_ENABLED(CONFIG_DRM_XE_DEBUG_SRIOV) ? 0600 : 0400,
-					    dent, NULL, &guc_state_ops);
 			debugfs_create_file("config_blob",
 					    IS_ENABLED(CONFIG_DRM_XE_DEBUG_SRIOV) ? 0600 : 0400,
 					    dent, NULL, &config_blob_ops);
-- 
2.50.1

[PATCH v3 14/28] drm/xe/pf: Don't save GuC VF migration data on pause

[Michał Winiarski]

In upcoming changes, the GuC VF migration data will be handled as part
of separate SAVE/RESTORE states in VF control state machine.
Remove it from PAUSE state.

Signed-off-by: Michał Winiarski <michal.winiarski@intel.com>
Reviewed-by: Michal Wajdeczko <michal.wajdeczko@intel.com>
---
 drivers/gpu/drm/xe/xe_gt_sriov_pf_control.c   | 39 +------------------
 .../gpu/drm/xe/xe_gt_sriov_pf_control_types.h |  2 -
 2 files changed, 2 insertions(+), 39 deletions(-)

diff --git a/drivers/gpu/drm/xe/xe_gt_sriov_pf_control.c b/drivers/gpu/drm/xe/xe_gt_sriov_pf_control.c
index 02019b646389c..203ba1fbeefcd 100644
--- a/drivers/gpu/drm/xe/xe_gt_sriov_pf_control.c
+++ b/drivers/gpu/drm/xe/xe_gt_sriov_pf_control.c
@@ -183,7 +183,6 @@ static const char *control_bit_to_string(enum xe_gt_sriov_control_bits bit)
 	CASE2STR(PAUSE_SEND_PAUSE);
 	CASE2STR(PAUSE_WAIT_GUC);
 	CASE2STR(PAUSE_GUC_DONE);
-	CASE2STR(PAUSE_SAVE_GUC);
 	CASE2STR(PAUSE_FAILED);
 	CASE2STR(PAUSED);
 	CASE2STR(SAVE_WIP);
@@ -453,8 +452,7 @@ static void pf_enter_vf_ready(struct xe_gt *gt, unsigned int vfid)
  *	:       PAUSE_GUC_DONE                          o-----restart
  *	:        |                                      :
  *	:        |   o---<--busy                        :
- *	:        v  /         /                         :
- *	:       PAUSE_SAVE_GUC                          :
+ *	:       /                                       :
  *	:      /                                        :
  *	:     /                                         :
  *	:....o..............o...............o...........:
@@ -474,7 +472,6 @@ static void pf_exit_vf_pause_wip(struct xe_gt *gt, unsigned int vfid)
 		pf_escape_vf_state(gt, vfid, XE_GT_SRIOV_STATE_PAUSE_SEND_PAUSE);
 		pf_escape_vf_state(gt, vfid, XE_GT_SRIOV_STATE_PAUSE_WAIT_GUC);
 		pf_escape_vf_state(gt, vfid, XE_GT_SRIOV_STATE_PAUSE_GUC_DONE);
-		pf_escape_vf_state(gt, vfid, XE_GT_SRIOV_STATE_PAUSE_SAVE_GUC);
 	}
 }
 
@@ -505,41 +502,12 @@ static void pf_enter_vf_pause_rejected(struct xe_gt *gt, unsigned int vfid)
 	pf_enter_vf_pause_failed(gt, vfid);
 }
 
-static void pf_enter_vf_pause_save_guc(struct xe_gt *gt, unsigned int vfid)
-{
-	if (!pf_enter_vf_state(gt, vfid, XE_GT_SRIOV_STATE_PAUSE_SAVE_GUC))
-		pf_enter_vf_state_machine_bug(gt, vfid);
-}
-
-static bool pf_exit_vf_pause_save_guc(struct xe_gt *gt, unsigned int vfid)
-{
-	int err;
-
-	if (!pf_exit_vf_state(gt, vfid, XE_GT_SRIOV_STATE_PAUSE_SAVE_GUC))
-		return false;
-
-	err = xe_gt_sriov_pf_migration_save_guc_state(gt, vfid);
-	if (err) {
-		/* retry if busy */
-		if (err == -EBUSY) {
-			pf_enter_vf_pause_save_guc(gt, vfid);
-			return true;
-		}
-		/* give up on error */
-		if (err == -EIO)
-			pf_enter_vf_mismatch(gt, vfid);
-	}
-
-	pf_enter_vf_pause_completed(gt, vfid);
-	return true;
-}
-
 static bool pf_exit_vf_pause_guc_done(struct xe_gt *gt, unsigned int vfid)
 {
 	if (!pf_exit_vf_state(gt, vfid, XE_GT_SRIOV_STATE_PAUSE_GUC_DONE))
 		return false;
 
-	pf_enter_vf_pause_save_guc(gt, vfid);
+	pf_enter_vf_pause_completed(gt, vfid);
 	return true;
 }
 
@@ -1942,9 +1910,6 @@ static bool pf_process_vf_state_machine(struct xe_gt *gt, unsigned int vfid)
 	if (pf_exit_vf_pause_guc_done(gt, vfid))
 		return true;
 
-	if (pf_exit_vf_pause_save_guc(gt, vfid))
-		return true;
-
 	if (pf_check_vf_state(gt, vfid, XE_GT_SRIOV_STATE_SAVE_WAIT_DATA)) {
 		xe_gt_sriov_dbg_verbose(gt, "VF%u in %s\n", vfid,
 					control_bit_to_string(XE_GT_SRIOV_STATE_SAVE_WAIT_DATA));
diff --git a/drivers/gpu/drm/xe/xe_gt_sriov_pf_control_types.h b/drivers/gpu/drm/xe/xe_gt_sriov_pf_control_types.h
index a14c738581ae2..c2ec0403b2be7 100644
--- a/drivers/gpu/drm/xe/xe_gt_sriov_pf_control_types.h
+++ b/drivers/gpu/drm/xe/xe_gt_sriov_pf_control_types.h
@@ -28,7 +28,6 @@
  * @XE_GT_SRIOV_STATE_PAUSE_SEND_PAUSE: indicates that the PF is about to send a PAUSE command.
  * @XE_GT_SRIOV_STATE_PAUSE_WAIT_GUC: indicates that the PF awaits for a response from the GuC.
  * @XE_GT_SRIOV_STATE_PAUSE_GUC_DONE: indicates that the PF has received a response from the GuC.
- * @XE_GT_SRIOV_STATE_PAUSE_SAVE_GUC: indicates that the PF needs to save the VF GuC state.
  * @XE_GT_SRIOV_STATE_PAUSE_FAILED: indicates that a VF pause operation has failed.
  * @XE_GT_SRIOV_STATE_PAUSED: indicates that the VF is paused.
  * @XE_GT_SRIOV_STATE_SAVE_WIP: indicates that VF save operation is in progress.
@@ -71,7 +70,6 @@ enum xe_gt_sriov_control_bits {
 	XE_GT_SRIOV_STATE_PAUSE_SEND_PAUSE,
 	XE_GT_SRIOV_STATE_PAUSE_WAIT_GUC,
 	XE_GT_SRIOV_STATE_PAUSE_GUC_DONE,
-	XE_GT_SRIOV_STATE_PAUSE_SAVE_GUC,
 	XE_GT_SRIOV_STATE_PAUSE_FAILED,
 	XE_GT_SRIOV_STATE_PAUSED,
 
-- 
2.50.1

[PATCH v3 15/28] drm/xe/pf: Switch VF migration GuC save/restore to struct migration data

[Michał Winiarski]

In upcoming changes, the GuC VF migration data will be handled as part
of separate SAVE/RESTORE states in VF control state machine.
Now that the data is decoupled from both guc_state debugfs and PAUSE
state, we can safely remove the struct xe_gt_sriov_state_snapshot and
modify the GuC save/restore functions to operate on struct
xe_sriov_migration_data.

Signed-off-by: Michał Winiarski <michal.winiarski@intel.com>
---
 drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.c | 265 +++++-------------
 drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.h |  13 +-
 .../drm/xe/xe_gt_sriov_pf_migration_types.h   |  27 --
 drivers/gpu/drm/xe/xe_gt_sriov_pf_types.h     |   4 -
 4 files changed, 79 insertions(+), 230 deletions(-)

diff --git a/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.c b/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.c
index a2db127982638..4a716e0a29fe4 100644
--- a/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.c
+++ b/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.c
@@ -28,6 +28,17 @@ static struct xe_gt_sriov_migration_data *pf_pick_gt_migration(struct xe_gt *gt,
 	return &gt->sriov.pf.vfs[vfid].migration;
 }
 
+static void pf_dump_mig_data(struct xe_gt *gt, unsigned int vfid,
+			     struct xe_sriov_migration_data *data)
+{
+	if (IS_ENABLED(CONFIG_DRM_XE_DEBUG_SRIOV)) {
+		print_hex_dump_bytes("mig_hdr:  ", DUMP_PREFIX_OFFSET,
+				     &data->hdr, sizeof(data->hdr));
+		print_hex_dump_bytes("mig_data: ", DUMP_PREFIX_OFFSET,
+				     data->vaddr, min(SZ_64, data->size));
+	}
+}
+
 /* Return: number of dwords saved/restored/required or a negative error code on failure */
 static int guc_action_vf_save_restore(struct xe_guc *guc, u32 vfid, u32 opcode,
 				      u64 addr, u32 ndwords)
@@ -47,7 +58,7 @@ static int guc_action_vf_save_restore(struct xe_guc *guc, u32 vfid, u32 opcode,
 }
 
 /* Return: size of the state in dwords or a negative error code on failure */
-static int pf_send_guc_query_vf_state_size(struct xe_gt *gt, unsigned int vfid)
+static int pf_send_guc_query_vf_mig_data_size(struct xe_gt *gt, unsigned int vfid)
 {
 	int ret;
 
@@ -56,8 +67,8 @@ static int pf_send_guc_query_vf_state_size(struct xe_gt *gt, unsigned int vfid)
 }
 
 /* Return: number of state dwords saved or a negative error code on failure */
-static int pf_send_guc_save_vf_state(struct xe_gt *gt, unsigned int vfid,
-				     void *dst, size_t size)
+static int pf_send_guc_save_vf_mig_data(struct xe_gt *gt, unsigned int vfid,
+					void *dst, size_t size)
 {
 	const int ndwords = size / sizeof(u32);
 	struct xe_guc *guc = &gt->uc.guc;
@@ -85,8 +96,8 @@ static int pf_send_guc_save_vf_state(struct xe_gt *gt, unsigned int vfid,
 }
 
 /* Return: number of state dwords restored or a negative error code on failure */
-static int pf_send_guc_restore_vf_state(struct xe_gt *gt, unsigned int vfid,
-					const void *src, size_t size)
+static int pf_send_guc_restore_vf_mig_data(struct xe_gt *gt, unsigned int vfid,
+					   const void *src, size_t size)
 {
 	const int ndwords = size / sizeof(u32);
 	struct xe_guc *guc = &gt->uc.guc;
@@ -114,120 +125,67 @@ static bool pf_migration_supported(struct xe_gt *gt)
 	return xe_sriov_pf_migration_supported(gt_to_xe(gt));
 }
 
-static struct mutex *pf_migration_mutex(struct xe_gt *gt)
+static int pf_save_vf_guc_mig_data(struct xe_gt *gt, unsigned int vfid)
 {
-	xe_gt_assert(gt, IS_SRIOV_PF(gt_to_xe(gt)));
-	return &gt->sriov.pf.migration.snapshot_lock;
-}
-
-static struct xe_gt_sriov_state_snapshot *pf_pick_vf_snapshot(struct xe_gt *gt,
-							      unsigned int vfid)
-{
-	xe_gt_assert(gt, IS_SRIOV_PF(gt_to_xe(gt)));
-	xe_gt_assert(gt, vfid <= xe_sriov_pf_get_totalvfs(gt_to_xe(gt)));
-	lockdep_assert_held(pf_migration_mutex(gt));
-
-	return &gt->sriov.pf.vfs[vfid].snapshot;
-}
-
-static unsigned int pf_snapshot_index(struct xe_gt *gt, struct xe_gt_sriov_state_snapshot *snapshot)
-{
-	return container_of(snapshot, struct xe_gt_sriov_metadata, snapshot) - gt->sriov.pf.vfs;
-}
-
-static void pf_free_guc_state(struct xe_gt *gt, struct xe_gt_sriov_state_snapshot *snapshot)
-{
-	struct xe_device *xe = gt_to_xe(gt);
-
-	drmm_kfree(&xe->drm, snapshot->guc.buff);
-	snapshot->guc.buff = NULL;
-	snapshot->guc.size = 0;
-}
-
-static int pf_alloc_guc_state(struct xe_gt *gt,
-			      struct xe_gt_sriov_state_snapshot *snapshot,
-			      size_t size)
-{
-	struct xe_device *xe = gt_to_xe(gt);
-	void *p;
-
-	pf_free_guc_state(gt, snapshot);
-
-	if (!size)
-		return -ENODATA;
-
-	if (size % sizeof(u32))
-		return -EINVAL;
-
-	if (size > SZ_2M)
-		return -EFBIG;
-
-	p = drmm_kzalloc(&xe->drm, size, GFP_KERNEL);
-	if (!p)
-		return -ENOMEM;
-
-	snapshot->guc.buff = p;
-	snapshot->guc.size = size;
-	return 0;
-}
-
-static void pf_dump_guc_state(struct xe_gt *gt, struct xe_gt_sriov_state_snapshot *snapshot)
-{
-	if (IS_ENABLED(CONFIG_DRM_XE_DEBUG_SRIOV)) {
-		unsigned int vfid __maybe_unused = pf_snapshot_index(gt, snapshot);
-
-		xe_gt_sriov_dbg_verbose(gt, "VF%u GuC state is %zu dwords:\n",
-					vfid, snapshot->guc.size / sizeof(u32));
-		print_hex_dump_bytes("state: ", DUMP_PREFIX_OFFSET,
-				     snapshot->guc.buff, min(SZ_64, snapshot->guc.size));
-	}
-}
-
-static int pf_save_vf_guc_state(struct xe_gt *gt, unsigned int vfid)
-{
-	struct xe_gt_sriov_state_snapshot *snapshot = pf_pick_vf_snapshot(gt, vfid);
+	struct xe_sriov_migration_data *data;
 	size_t size;
 	int ret;
 
-	ret = pf_send_guc_query_vf_state_size(gt, vfid);
+	ret = pf_send_guc_query_vf_mig_data_size(gt, vfid);
 	if (ret < 0)
 		goto fail;
+
 	size = ret * sizeof(u32);
-	xe_gt_sriov_dbg_verbose(gt, "VF%u state size is %d dwords (%zu bytes)\n", vfid, ret, size);
 
-	ret = pf_alloc_guc_state(gt, snapshot, size);
-	if (ret < 0)
+	data = xe_sriov_migration_data_alloc(gt_to_xe(gt));
+	if (!data) {
+		ret = -ENOMEM;
 		goto fail;
+	}
+
+	ret = xe_sriov_migration_data_init(data, gt->tile->id, gt->info.id,
+					   XE_SRIOV_MIGRATION_DATA_TYPE_GUC, 0, size);
+	if (ret)
+		goto fail_free;
 
-	ret = pf_send_guc_save_vf_state(gt, vfid, snapshot->guc.buff, size);
+	ret = pf_send_guc_save_vf_mig_data(gt, vfid, data->vaddr, size);
 	if (ret < 0)
-		goto fail;
+		goto fail_free;
 	size = ret * sizeof(u32);
 	xe_gt_assert(gt, size);
-	xe_gt_assert(gt, size <= snapshot->guc.size);
-	snapshot->guc.size = size;
+	xe_gt_assert(gt, size <= data->size);
+	data->size = size;
+	data->remaining = size;
+
+	xe_gt_sriov_dbg_verbose(gt, "VF%u GuC data save (%zu bytes)\n", vfid, size);
+	pf_dump_mig_data(gt, vfid, data);
+
+	ret = xe_gt_sriov_pf_migration_save_produce(gt, vfid, data);
+	if (ret)
+		goto fail_free;
 
-	pf_dump_guc_state(gt, snapshot);
 	return 0;
 
+fail_free:
+	xe_sriov_migration_data_free(data);
 fail:
-	xe_gt_sriov_dbg(gt, "Unable to save VF%u state (%pe)\n", vfid, ERR_PTR(ret));
-	pf_free_guc_state(gt, snapshot);
+	xe_gt_sriov_err(gt, "Failed to save VF%u GuC data (%pe)\n",
+			vfid, ERR_PTR(ret));
 	return ret;
 }
 
 /**
- * xe_gt_sriov_pf_migration_save_guc_state() - Take a GuC VF state snapshot.
+ * xe_gt_sriov_pf_migration_guc_size() - Get the size of VF GuC migration data.
  * @gt: the &xe_gt
  * @vfid: the VF identifier
  *
  * This function is for PF only.
  *
- * Return: 0 on success or a negative error code on failure.
+ * Return: size in bytes or a negative error code on failure.
  */
-int xe_gt_sriov_pf_migration_save_guc_state(struct xe_gt *gt, unsigned int vfid)
+ssize_t xe_gt_sriov_pf_migration_guc_size(struct xe_gt *gt, unsigned int vfid)
 {
-	int err;
+	ssize_t size;
 
 	xe_gt_assert(gt, IS_SRIOV_PF(gt_to_xe(gt)));
 	xe_gt_assert(gt, vfid != PFID);
@@ -236,37 +194,15 @@ int xe_gt_sriov_pf_migration_save_guc_state(struct xe_gt *gt, unsigned int vfid)
 	if (!pf_migration_supported(gt))
 		return -ENOPKG;
 
-	mutex_lock(pf_migration_mutex(gt));
-	err = pf_save_vf_guc_state(gt, vfid);
-	mutex_unlock(pf_migration_mutex(gt));
-
-	return err;
-}
-
-static int pf_restore_vf_guc_state(struct xe_gt *gt, unsigned int vfid)
-{
-	struct xe_gt_sriov_state_snapshot *snapshot = pf_pick_vf_snapshot(gt, vfid);
-	int ret;
-
-	if (!snapshot->guc.size)
-		return -ENODATA;
-
-	xe_gt_sriov_dbg_verbose(gt, "restoring %zu dwords of VF%u GuC state\n",
-				snapshot->guc.size / sizeof(u32), vfid);
-	ret = pf_send_guc_restore_vf_state(gt, vfid, snapshot->guc.buff, snapshot->guc.size);
-	if (ret < 0)
-		goto fail;
-
-	xe_gt_sriov_dbg_verbose(gt, "restored %d dwords of VF%u GuC state\n", ret, vfid);
-	return 0;
+	size = pf_send_guc_query_vf_mig_data_size(gt, vfid);
+	if (size >= 0)
+		size *= sizeof(u32);
 
-fail:
-	xe_gt_sriov_dbg(gt, "Failed to restore VF%u GuC state (%pe)\n", vfid, ERR_PTR(ret));
-	return ret;
+	return size;
 }
 
 /**
- * xe_gt_sriov_pf_migration_restore_guc_state() - Restore a GuC VF state.
+ * xe_gt_sriov_pf_migration_guc_save() - Save VF GuC migration data.
  * @gt: the &xe_gt
  * @vfid: the VF identifier
  *
@@ -274,10 +210,8 @@ static int pf_restore_vf_guc_state(struct xe_gt *gt, unsigned int vfid)
  *
  * Return: 0 on success or a negative error code on failure.
  */
-int xe_gt_sriov_pf_migration_restore_guc_state(struct xe_gt *gt, unsigned int vfid)
+int xe_gt_sriov_pf_migration_guc_save(struct xe_gt *gt, unsigned int vfid)
 {
-	int ret;
-
 	xe_gt_assert(gt, IS_SRIOV_PF(gt_to_xe(gt)));
 	xe_gt_assert(gt, vfid != PFID);
 	xe_gt_assert(gt, vfid <= xe_sriov_pf_get_totalvfs(gt_to_xe(gt)));
@@ -285,75 +219,43 @@ int xe_gt_sriov_pf_migration_restore_guc_state(struct xe_gt *gt, unsigned int vf
 	if (!pf_migration_supported(gt))
 		return -ENOPKG;
 
-	mutex_lock(pf_migration_mutex(gt));
-	ret = pf_restore_vf_guc_state(gt, vfid);
-	mutex_unlock(pf_migration_mutex(gt));
-
-	return ret;
+	return pf_save_vf_guc_mig_data(gt, vfid);
 }
 
-#ifdef CONFIG_DEBUG_FS
-/**
- * xe_gt_sriov_pf_migration_read_guc_state() - Read a GuC VF state.
- * @gt: the &xe_gt
- * @vfid: the VF identifier
- * @buf: the user space buffer to read to
- * @count: the maximum number of bytes to read
- * @pos: the current position in the buffer
- *
- * This function is for PF only.
- *
- * This function reads up to @count bytes from the saved VF GuC state buffer
- * at offset @pos into the user space address starting at @buf.
- *
- * Return: the number of bytes read or a negative error code on failure.
- */
-ssize_t xe_gt_sriov_pf_migration_read_guc_state(struct xe_gt *gt, unsigned int vfid,
-						char __user *buf, size_t count, loff_t *pos)
+static int pf_restore_vf_guc_state(struct xe_gt *gt, unsigned int vfid,
+				   struct xe_sriov_migration_data *data)
 {
-	struct xe_gt_sriov_state_snapshot *snapshot;
-	ssize_t ret;
+	int ret;
 
-	xe_gt_assert(gt, IS_SRIOV_PF(gt_to_xe(gt)));
-	xe_gt_assert(gt, vfid != PFID);
-	xe_gt_assert(gt, vfid <= xe_sriov_pf_get_totalvfs(gt_to_xe(gt)));
+	xe_gt_assert(gt, data->size);
 
-	if (!pf_migration_supported(gt))
-		return -ENOPKG;
+	xe_gt_sriov_dbg_verbose(gt, "VF%u GuC data restore (%llu bytes)\n", vfid, data->size);
+	pf_dump_mig_data(gt, vfid, data);
 
-	mutex_lock(pf_migration_mutex(gt));
-	snapshot = pf_pick_vf_snapshot(gt, vfid);
-	if (snapshot->guc.size)
-		ret = simple_read_from_buffer(buf, count, pos, snapshot->guc.buff,
-					      snapshot->guc.size);
-	else
-		ret = -ENODATA;
-	mutex_unlock(pf_migration_mutex(gt));
+	ret = pf_send_guc_restore_vf_mig_data(gt, vfid, data->vaddr, data->size);
+	if (ret < 0)
+		goto fail;
+
+	return 0;
 
+fail:
+	xe_gt_sriov_err(gt, "Failed to restore VF%u GuC data (%pe)\n",
+			vfid, ERR_PTR(ret));
 	return ret;
 }
 
 /**
- * xe_gt_sriov_pf_migration_write_guc_state() - Write a GuC VF state.
+ * xe_gt_sriov_pf_migration_guc_restore() - Restore VF GuC migration data.
  * @gt: the &xe_gt
  * @vfid: the VF identifier
- * @buf: the user space buffer with GuC VF state
- * @size: the size of GuC VF state (in bytes)
  *
  * This function is for PF only.
  *
- * This function reads @size bytes of the VF GuC state stored at user space
- * address @buf and writes it into a internal VF state buffer.
- *
- * Return: the number of bytes used or a negative error code on failure.
+ * Return: 0 on success or a negative error code on failure.
  */
-ssize_t xe_gt_sriov_pf_migration_write_guc_state(struct xe_gt *gt, unsigned int vfid,
-						 const char __user *buf, size_t size)
+int xe_gt_sriov_pf_migration_guc_restore(struct xe_gt *gt, unsigned int vfid,
+					 struct xe_sriov_migration_data *data)
 {
-	struct xe_gt_sriov_state_snapshot *snapshot;
-	loff_t pos = 0;
-	ssize_t ret;
-
 	xe_gt_assert(gt, IS_SRIOV_PF(gt_to_xe(gt)));
 	xe_gt_assert(gt, vfid != PFID);
 	xe_gt_assert(gt, vfid <= xe_sriov_pf_get_totalvfs(gt_to_xe(gt)));
@@ -361,21 +263,8 @@ ssize_t xe_gt_sriov_pf_migration_write_guc_state(struct xe_gt *gt, unsigned int
 	if (!pf_migration_supported(gt))
 		return -ENOPKG;
 
-	mutex_lock(pf_migration_mutex(gt));
-	snapshot = pf_pick_vf_snapshot(gt, vfid);
-	ret = pf_alloc_guc_state(gt, snapshot, size);
-	if (!ret) {
-		ret = simple_write_to_buffer(snapshot->guc.buff, size, &pos, buf, size);
-		if (ret < 0)
-			pf_free_guc_state(gt, snapshot);
-		else
-			pf_dump_guc_state(gt, snapshot);
-	}
-	mutex_unlock(pf_migration_mutex(gt));
-
-	return ret;
+	return pf_restore_vf_guc_state(gt, vfid, data);
 }
-#endif /* CONFIG_DEBUG_FS */
 
 /**
  * xe_gt_sriov_pf_migration_size() - Total size of migration data from all components within a GT.
@@ -599,10 +488,6 @@ int xe_gt_sriov_pf_migration_init(struct xe_gt *gt)
 	if (!pf_migration_supported(gt))
 		return 0;
 
-	err = drmm_mutex_init(&xe->drm, &gt->sriov.pf.migration.snapshot_lock);
-	if (err)
-		return err;
-
 	totalvfs = xe_sriov_pf_get_totalvfs(xe);
 	for (n = 1; n <= totalvfs; n++) {
 		struct xe_gt_sriov_migration_data *migration = pf_pick_gt_migration(gt, n);
diff --git a/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.h b/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.h
index 4f2f2783339c3..b3c18e369df79 100644
--- a/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.h
+++ b/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.h
@@ -15,8 +15,10 @@ struct xe_sriov_migration_data;
 #define XE_GT_SRIOV_PF_MIGRATION_GUC_DATA_MAX_SIZE SZ_8M
 
 int xe_gt_sriov_pf_migration_init(struct xe_gt *gt);
-int xe_gt_sriov_pf_migration_save_guc_state(struct xe_gt *gt, unsigned int vfid);
-int xe_gt_sriov_pf_migration_restore_guc_state(struct xe_gt *gt, unsigned int vfid);
+ssize_t xe_gt_sriov_pf_migration_guc_size(struct xe_gt *gt, unsigned int vfid);
+int xe_gt_sriov_pf_migration_guc_save(struct xe_gt *gt, unsigned int vfid);
+int xe_gt_sriov_pf_migration_guc_restore(struct xe_gt *gt, unsigned int vfid,
+					 struct xe_sriov_migration_data *data);
 
 ssize_t xe_gt_sriov_pf_migration_size(struct xe_gt *gt, unsigned int vfid);
 
@@ -34,11 +36,4 @@ int xe_gt_sriov_pf_migration_restore_produce(struct xe_gt *gt, unsigned int vfid
 struct xe_sriov_migration_data *
 xe_gt_sriov_pf_migration_save_consume(struct xe_gt *gt, unsigned int vfid);
 
-#ifdef CONFIG_DEBUG_FS
-ssize_t xe_gt_sriov_pf_migration_read_guc_state(struct xe_gt *gt, unsigned int vfid,
-						char __user *buf, size_t count, loff_t *pos);
-ssize_t xe_gt_sriov_pf_migration_write_guc_state(struct xe_gt *gt, unsigned int vfid,
-						 const char __user *buf, size_t count);
-#endif
-
 #endif
diff --git a/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration_types.h b/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration_types.h
index 84be6fac16c8b..75d8b94cbbefb 100644
--- a/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration_types.h
+++ b/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration_types.h
@@ -6,24 +6,7 @@
 #ifndef _XE_GT_SRIOV_PF_MIGRATION_TYPES_H_
 #define _XE_GT_SRIOV_PF_MIGRATION_TYPES_H_
 
-#include <linux/mutex.h>
 #include <linux/ptr_ring.h>
-#include <linux/types.h>
-
-/**
- * struct xe_gt_sriov_state_snapshot - GT-level per-VF state snapshot data.
- *
- * Used by the PF driver to maintain per-VF migration data.
- */
-struct xe_gt_sriov_state_snapshot {
-	/** @guc: GuC VF state snapshot */
-	struct {
-		/** @guc.buff: buffer with the VF state */
-		u32 *buff;
-		/** @guc.size: size of the buffer (must be dwords aligned) */
-		u32 size;
-	} guc;
-};
 
 /**
  * struct xe_gt_sriov_migration_data - GT-level per-VF migration data.
@@ -35,14 +18,4 @@ struct xe_gt_sriov_migration_data {
 	struct ptr_ring ring;
 };
 
-/**
- * struct xe_gt_sriov_pf_migration - GT-level data.
- *
- * Used by the PF driver to maintain non-VF specific per-GT data.
- */
-struct xe_gt_sriov_pf_migration {
-	/** @snapshot_lock: protects all VFs snapshots */
-	struct mutex snapshot_lock;
-};
-
 #endif
diff --git a/drivers/gpu/drm/xe/xe_gt_sriov_pf_types.h b/drivers/gpu/drm/xe/xe_gt_sriov_pf_types.h
index 812e74d3f8f80..667b8310478d4 100644
--- a/drivers/gpu/drm/xe/xe_gt_sriov_pf_types.h
+++ b/drivers/gpu/drm/xe/xe_gt_sriov_pf_types.h
@@ -31,9 +31,6 @@ struct xe_gt_sriov_metadata {
 	/** @version: negotiated VF/PF ABI version */
 	struct xe_gt_sriov_pf_service_version version;
 
-	/** @snapshot: snapshot of the VF state data */
-	struct xe_gt_sriov_state_snapshot snapshot;
-
 	/** @migration: per-VF migration data. */
 	struct xe_gt_sriov_migration_data migration;
 };
@@ -61,7 +58,6 @@ struct xe_gt_sriov_pf {
 	struct xe_gt_sriov_pf_service service;
 	struct xe_gt_sriov_pf_control control;
 	struct xe_gt_sriov_pf_policy policy;
-	struct xe_gt_sriov_pf_migration migration;
 	struct xe_gt_sriov_spare_config spare;
 	struct xe_gt_sriov_metadata *vfs;
 };
-- 
2.50.1

[PATCH v3 16/28] drm/xe/pf: Handle GuC migration data as part of PF control

[Michał Winiarski]

Connect the helpers to allow save and restore of GuC migration data in
stop_copy / resume device state.

Signed-off-by: Michał Winiarski <michal.winiarski@intel.com>
---
 drivers/gpu/drm/xe/xe_gt_sriov_pf_control.c   | 25 ++++++-
 drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.c | 66 +++++++++++++++----
 drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.h |  8 ++-
 .../drm/xe/xe_gt_sriov_pf_migration_types.h   |  5 ++
 4 files changed, 87 insertions(+), 17 deletions(-)

diff --git a/drivers/gpu/drm/xe/xe_gt_sriov_pf_control.c b/drivers/gpu/drm/xe/xe_gt_sriov_pf_control.c
index 203ba1fbeefcd..cb45e97f4c4d9 100644
--- a/drivers/gpu/drm/xe/xe_gt_sriov_pf_control.c
+++ b/drivers/gpu/drm/xe/xe_gt_sriov_pf_control.c
@@ -848,6 +848,18 @@ static void pf_enter_vf_save_failed(struct xe_gt *gt, unsigned int vfid)
 
 static int pf_handle_vf_save_data(struct xe_gt *gt, unsigned int vfid)
 {
+	int ret;
+
+	if (xe_gt_sriov_pf_migration_save_test(gt, vfid, XE_SRIOV_MIGRATION_DATA_TYPE_GUC)) {
+		ret = xe_gt_sriov_pf_migration_guc_save(gt, vfid);
+		if (ret)
+			return ret;
+
+		xe_gt_sriov_pf_migration_save_clear(gt, vfid, XE_SRIOV_MIGRATION_DATA_TYPE_GUC);
+
+		return -EAGAIN;
+	}
+
 	return 0;
 }
 
@@ -887,6 +899,7 @@ static void pf_exit_vf_save_wait_data(struct xe_gt *gt, unsigned int vfid)
 static bool pf_enter_vf_save_wip(struct xe_gt *gt, unsigned int vfid)
 {
 	if (pf_enter_vf_state(gt, vfid, XE_GT_SRIOV_STATE_SAVE_WIP)) {
+		xe_gt_sriov_pf_migration_save_init(gt, vfid);
 		pf_enter_vf_state(gt, vfid, XE_GT_SRIOV_STATE_SAVE_PROCESS_DATA);
 		pf_enter_vf_wip(gt, vfid);
 		pf_queue_vf(gt, vfid);
@@ -1048,14 +1061,22 @@ static void pf_enter_vf_restore_failed(struct xe_gt *gt, unsigned int vfid)
 static int pf_handle_vf_restore_data(struct xe_gt *gt, unsigned int vfid)
 {
 	struct xe_sriov_migration_data *data = xe_gt_sriov_pf_migration_restore_consume(gt, vfid);
+	int ret = 0;
 
 	xe_gt_assert(gt, data);
 
-	xe_gt_sriov_notice(gt, "Skipping VF%u unknown data type: %d\n", vfid, data->type);
+	switch (data->type) {
+	case XE_SRIOV_MIGRATION_DATA_TYPE_GUC:
+		ret = xe_gt_sriov_pf_migration_guc_restore(gt, vfid, data);
+		break;
+	default:
+		xe_gt_sriov_notice(gt, "Skipping VF%u unknown data type: %d\n", vfid, data->type);
+		break;
+	}
 
 	xe_sriov_migration_data_free(data);
 
-	return 0;
+	return ret;
 }
 
 static bool pf_handle_vf_restore(struct xe_gt *gt, unsigned int vfid)
diff --git a/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.c b/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.c
index 4a716e0a29fe4..fbb451767712a 100644
--- a/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.c
+++ b/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.c
@@ -174,22 +174,10 @@ static int pf_save_vf_guc_mig_data(struct xe_gt *gt, unsigned int vfid)
 	return ret;
 }
 
-/**
- * xe_gt_sriov_pf_migration_guc_size() - Get the size of VF GuC migration data.
- * @gt: the &xe_gt
- * @vfid: the VF identifier
- *
- * This function is for PF only.
- *
- * Return: size in bytes or a negative error code on failure.
- */
-ssize_t xe_gt_sriov_pf_migration_guc_size(struct xe_gt *gt, unsigned int vfid)
+static ssize_t pf_migration_guc_size(struct xe_gt *gt, unsigned int vfid)
 {
 	ssize_t size;
 
-	xe_gt_assert(gt, IS_SRIOV_PF(gt_to_xe(gt)));
-	xe_gt_assert(gt, vfid != PFID);
-	xe_gt_assert(gt, vfid <= xe_sriov_pf_get_totalvfs(gt_to_xe(gt)));
 
 	if (!pf_migration_supported(gt))
 		return -ENOPKG;
@@ -278,12 +266,19 @@ int xe_gt_sriov_pf_migration_guc_restore(struct xe_gt *gt, unsigned int vfid,
 ssize_t xe_gt_sriov_pf_migration_size(struct xe_gt *gt, unsigned int vfid)
 {
 	ssize_t total = 0;
+	ssize_t size;
 
 	xe_gt_assert(gt, IS_SRIOV_PF(gt_to_xe(gt)));
 	xe_gt_assert(gt, vfid != PFID);
 	xe_gt_assert(gt, vfid <= xe_sriov_pf_get_totalvfs(gt_to_xe(gt)));
 
-	/* Nothing to query yet - will be updated once per-GT migration data types are added */
+	size = pf_migration_guc_size(gt, vfid);
+	if (size < 0)
+		return size;
+	if (size > 0)
+		size += sizeof(struct xe_sriov_pf_migration_hdr);
+	total += size;
+
 	return total;
 }
 
@@ -330,6 +325,49 @@ void xe_gt_sriov_pf_migration_ring_free(struct xe_gt *gt, unsigned int vfid)
 		xe_sriov_migration_data_free(data);
 }
 
+/**
+ * xe_gt_sriov_pf_migration_save_init() - Initialize per-GT migration related data.
+ * @gt: the &xe_gt
+ * @vfid: the VF identifier (can't be 0)
+ *
+ * Return: 0 on success or a negative error code on failure.
+ */
+void xe_gt_sriov_pf_migration_save_init(struct xe_gt *gt, unsigned int vfid)
+{
+	struct xe_gt_sriov_migration_data *migration = pf_pick_gt_migration(gt, vfid);
+
+	migration->save.data_remaining = 0;
+
+	xe_gt_assert(gt, pf_migration_guc_size(gt, vfid) > 0);
+	set_bit(XE_SRIOV_MIGRATION_DATA_TYPE_GUC, &migration->save.data_remaining);
+}
+
+/**
+ * xe_gt_sriov_pf_migration_save_test() - Test if migration data type needs to be saved.
+ * @gt: the &xe_gt
+ * @vfid: the VF identifier (can't be 0)
+ * @type: the &xe_sriov_migration_data_type of data to be saved
+ *
+ * Return: true if the data needs saving, otherwise false.
+ */
+bool xe_gt_sriov_pf_migration_save_test(struct xe_gt *gt, unsigned int vfid,
+					enum xe_sriov_migration_data_type type)
+{
+	return test_bit(type, &pf_pick_gt_migration(gt, vfid)->save.data_remaining);
+}
+
+/**
+ * xe_gt_sriov_pf_migration_save_clear() - Clear migration data type from saving.
+ * @gt: the &xe_gt
+ * @vfid: the VF identifier (can't be 0)
+ * @type: the &xe_sriov_migration_data_type of data to be cleared
+ */
+void xe_gt_sriov_pf_migration_save_clear(struct xe_gt *gt, unsigned int vfid,
+					 enum xe_sriov_migration_data_type type)
+{
+	clear_bit(type, &pf_pick_gt_migration(gt, vfid)->save.data_remaining);
+}
+
 /**
  * xe_gt_sriov_pf_migration_save_produce() - Add VF save data packet to migration ring.
  * @gt: the &xe_gt
diff --git a/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.h b/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.h
index b3c18e369df79..66a45d6234245 100644
--- a/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.h
+++ b/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.h
@@ -10,12 +10,12 @@
 
 struct xe_gt;
 struct xe_sriov_migration_data;
+enum xe_sriov_migration_data_type;
 
 /* TODO: get this information by querying GuC in the future */
 #define XE_GT_SRIOV_PF_MIGRATION_GUC_DATA_MAX_SIZE SZ_8M
 
 int xe_gt_sriov_pf_migration_init(struct xe_gt *gt);
-ssize_t xe_gt_sriov_pf_migration_guc_size(struct xe_gt *gt, unsigned int vfid);
 int xe_gt_sriov_pf_migration_guc_save(struct xe_gt *gt, unsigned int vfid);
 int xe_gt_sriov_pf_migration_guc_restore(struct xe_gt *gt, unsigned int vfid,
 					 struct xe_sriov_migration_data *data);
@@ -26,6 +26,12 @@ bool xe_gt_sriov_pf_migration_ring_empty(struct xe_gt *gt, unsigned int vfid);
 bool xe_gt_sriov_pf_migration_ring_full(struct xe_gt *gt, unsigned int vfid);
 void xe_gt_sriov_pf_migration_ring_free(struct xe_gt *gt, unsigned int vfid);
 
+void xe_gt_sriov_pf_migration_save_init(struct xe_gt *gt, unsigned int vfid);
+bool xe_gt_sriov_pf_migration_save_test(struct xe_gt *gt, unsigned int vfid,
+					enum xe_sriov_migration_data_type type);
+void xe_gt_sriov_pf_migration_save_clear(struct xe_gt *gt, unsigned int vfid,
+					 enum xe_sriov_migration_data_type type);
+
 int xe_gt_sriov_pf_migration_save_produce(struct xe_gt *gt, unsigned int vfid,
 					  struct xe_sriov_migration_data *data);
 struct xe_sriov_migration_data *
diff --git a/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration_types.h b/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration_types.h
index 75d8b94cbbefb..9f24878690d9c 100644
--- a/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration_types.h
+++ b/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration_types.h
@@ -16,6 +16,11 @@
 struct xe_gt_sriov_migration_data {
 	/** @ring: queue containing VF save / restore migration data */
 	struct ptr_ring ring;
+	/** @save: structure for currently processed save migration data */
+	struct {
+		/** @save.data_remaining: bitmap of migration types that need to be saved */
+		unsigned long data_remaining;
+	} save;
 };
 
 #endif
-- 
2.50.1

[PATCH v3 17/28] drm/xe/pf: Add helpers for VF GGTT migration data handling

[Michał Winiarski]

In an upcoming change, the VF GGTT migration data will be handled as
part of VF control state machine. Add the necessary helpers to allow the
migration data transfer to/from the HW GGTT resource.

Signed-off-by: Michał Winiarski <michal.winiarski@intel.com>
---
 drivers/gpu/drm/xe/xe_ggtt.c               | 104 +++++++++++++++++++++
 drivers/gpu/drm/xe/xe_ggtt.h               |   4 +
 drivers/gpu/drm/xe/xe_ggtt_types.h         |   2 +
 drivers/gpu/drm/xe/xe_gt_sriov_pf_config.c |  52 +++++++++++
 drivers/gpu/drm/xe/xe_gt_sriov_pf_config.h |   5 +
 5 files changed, 167 insertions(+)

diff --git a/drivers/gpu/drm/xe/xe_ggtt.c b/drivers/gpu/drm/xe/xe_ggtt.c
index 20d226d90c50f..00ddb4f013466 100644
--- a/drivers/gpu/drm/xe/xe_ggtt.c
+++ b/drivers/gpu/drm/xe/xe_ggtt.c
@@ -151,6 +151,14 @@ static void xe_ggtt_set_pte_and_flush(struct xe_ggtt *ggtt, u64 addr, u64 pte)
 	ggtt_update_access_counter(ggtt);
 }
 
+static u64 xe_ggtt_get_pte(struct xe_ggtt *ggtt, u64 addr)
+{
+	xe_tile_assert(ggtt->tile, !(addr & XE_PTE_MASK));
+	xe_tile_assert(ggtt->tile, addr < ggtt->size);
+
+	return readq(&ggtt->gsm[addr >> XE_PTE_SHIFT]);
+}
+
 static void xe_ggtt_clear(struct xe_ggtt *ggtt, u64 start, u64 size)
 {
 	u16 pat_index = tile_to_xe(ggtt->tile)->pat.idx[XE_CACHE_WB];
@@ -233,16 +241,19 @@ void xe_ggtt_might_lock(struct xe_ggtt *ggtt)
 static const struct xe_ggtt_pt_ops xelp_pt_ops = {
 	.pte_encode_flags = xelp_ggtt_pte_flags,
 	.ggtt_set_pte = xe_ggtt_set_pte,
+	.ggtt_get_pte = xe_ggtt_get_pte,
 };
 
 static const struct xe_ggtt_pt_ops xelpg_pt_ops = {
 	.pte_encode_flags = xelpg_ggtt_pte_flags,
 	.ggtt_set_pte = xe_ggtt_set_pte,
+	.ggtt_get_pte = xe_ggtt_get_pte,
 };
 
 static const struct xe_ggtt_pt_ops xelpg_pt_wa_ops = {
 	.pte_encode_flags = xelpg_ggtt_pte_flags,
 	.ggtt_set_pte = xe_ggtt_set_pte_and_flush,
+	.ggtt_get_pte = xe_ggtt_get_pte,
 };
 
 static void __xe_ggtt_init_early(struct xe_ggtt *ggtt, u32 reserved)
@@ -889,6 +900,20 @@ u64 xe_ggtt_largest_hole(struct xe_ggtt *ggtt, u64 alignment, u64 *spare)
 	return max_hole;
 }
 
+/**
+ * xe_ggtt_node_pt_size() - Convert GGTT node size to its page table entries size.
+ * @node: the &xe_ggtt_node
+ *
+ * Return: GGTT node page table entries size in bytes.
+ */
+size_t xe_ggtt_node_pt_size(const struct xe_ggtt_node *node)
+{
+	if (!node)
+		return 0;
+
+	return node->base.size / XE_PAGE_SIZE * sizeof(u64);
+}
+
 #ifdef CONFIG_PCI_IOV
 static u64 xe_encode_vfid_pte(u16 vfid)
 {
@@ -930,6 +955,85 @@ void xe_ggtt_assign(const struct xe_ggtt_node *node, u16 vfid)
 	xe_ggtt_assign_locked(node->ggtt, &node->base, vfid);
 	mutex_unlock(&node->ggtt->lock);
 }
+
+/**
+ * xe_ggtt_node_save() - Save a &xe_ggtt_node to a buffer.
+ * @node: the &xe_ggtt_node to be saved
+ * @dst: destination buffer
+ * @size: destination buffer size in bytes
+ * @vfid: VF identifier
+ *
+ * Return: 0 on success or a negative error code on failure.
+ */
+int xe_ggtt_node_save(struct xe_ggtt_node *node, void *dst, size_t size, u16 vfid)
+{
+	struct xe_ggtt *ggtt;
+	u64 start, end;
+	u64 *buf = dst;
+	u64 pte;
+
+	if (!node)
+		return -ENOENT;
+
+	guard(mutex)(&node->ggtt->lock);
+
+	ggtt = node->ggtt;
+	start = node->base.start;
+	end = start + size - 1;
+
+	if (xe_ggtt_node_pt_size(node) != size)
+		return -EINVAL;
+
+	while (start < end) {
+		pte = ggtt->pt_ops->ggtt_get_pte(ggtt, start);
+		if (vfid != u64_get_bits(pte, GGTT_PTE_VFID))
+			return -EPERM;
+
+		*buf++ = u64_replace_bits(pte, 0, GGTT_PTE_VFID);
+		start += XE_PAGE_SIZE;
+	}
+
+	return 0;
+}
+
+/**
+ * xe_ggtt_node_load() - Load a &xe_ggtt_node from a buffer.
+ * @node: the &xe_ggtt_node to be loaded
+ * @src: source buffer
+ * @size: source buffer size in bytes
+ * @vfid: VF identifier
+ *
+ * Return: 0 on success or a negative error code on failure.
+ */
+int xe_ggtt_node_load(struct xe_ggtt_node *node, const void *src, size_t size, u16 vfid)
+{
+	u64 vfid_pte = xe_encode_vfid_pte(vfid);
+	const u64 *buf = src;
+	struct xe_ggtt *ggtt;
+	u64 start, end;
+
+	if (!node)
+		return -ENOENT;
+
+	guard(mutex)(&node->ggtt->lock);
+
+	ggtt = node->ggtt;
+	start = node->base.start;
+	end = start + size - 1;
+
+	if (xe_ggtt_node_pt_size(node) != size)
+		return -EINVAL;
+
+	while (start < end) {
+		vfid_pte = u64_replace_bits(*buf++, vfid, GGTT_PTE_VFID);
+		ggtt->pt_ops->ggtt_set_pte(ggtt, start, vfid_pte);
+		start += XE_PAGE_SIZE;
+	}
+	xe_ggtt_invalidate(ggtt);
+
+	return 0;
+}
+
 #endif
 
 /**
diff --git a/drivers/gpu/drm/xe/xe_ggtt.h b/drivers/gpu/drm/xe/xe_ggtt.h
index 75fc7a1efea76..1edf27608d39a 100644
--- a/drivers/gpu/drm/xe/xe_ggtt.h
+++ b/drivers/gpu/drm/xe/xe_ggtt.h
@@ -41,8 +41,12 @@ u64 xe_ggtt_largest_hole(struct xe_ggtt *ggtt, u64 alignment, u64 *spare);
 int xe_ggtt_dump(struct xe_ggtt *ggtt, struct drm_printer *p);
 u64 xe_ggtt_print_holes(struct xe_ggtt *ggtt, u64 alignment, struct drm_printer *p);
 
+size_t xe_ggtt_node_pt_size(const struct xe_ggtt_node *node);
+
 #ifdef CONFIG_PCI_IOV
 void xe_ggtt_assign(const struct xe_ggtt_node *node, u16 vfid);
+int xe_ggtt_node_save(struct xe_ggtt_node *node, void *dst, size_t size, u16 vfid);
+int xe_ggtt_node_load(struct xe_ggtt_node *node, const void *src, size_t size, u16 vfid);
 #endif
 
 #ifndef CONFIG_LOCKDEP
diff --git a/drivers/gpu/drm/xe/xe_ggtt_types.h b/drivers/gpu/drm/xe/xe_ggtt_types.h
index c5e999d58ff2a..dacd796f81844 100644
--- a/drivers/gpu/drm/xe/xe_ggtt_types.h
+++ b/drivers/gpu/drm/xe/xe_ggtt_types.h
@@ -78,6 +78,8 @@ struct xe_ggtt_pt_ops {
 	u64 (*pte_encode_flags)(struct xe_bo *bo, u16 pat_index);
 	/** @ggtt_set_pte: Directly write into GGTT's PTE */
 	void (*ggtt_set_pte)(struct xe_ggtt *ggtt, u64 addr, u64 pte);
+	/** @ggtt_get_pte: Directly read from GGTT's PTE */
+	u64 (*ggtt_get_pte)(struct xe_ggtt *ggtt, u64 addr);
 };
 
 #endif
diff --git a/drivers/gpu/drm/xe/xe_gt_sriov_pf_config.c b/drivers/gpu/drm/xe/xe_gt_sriov_pf_config.c
index c0c0215c07036..55444883f2ac3 100644
--- a/drivers/gpu/drm/xe/xe_gt_sriov_pf_config.c
+++ b/drivers/gpu/drm/xe/xe_gt_sriov_pf_config.c
@@ -726,6 +726,58 @@ int xe_gt_sriov_pf_config_set_fair_ggtt(struct xe_gt *gt, unsigned int vfid,
 	return xe_gt_sriov_pf_config_bulk_set_ggtt(gt, vfid, num_vfs, fair);
 }
 
+/**
+ * xe_gt_sriov_pf_config_ggtt_save() - Save a VF provisioned GGTT data into a buffer.
+ * @gt: the &xe_gt
+ * @vfid: VF identifier (can't be 0)
+ * @buf: the GGTT data destination buffer (or NULL to query the buf size)
+ * @size: the size of the buffer (or 0 to query the buf size)
+ *
+ * This function can only be called on PF.
+ *
+ * Return: size of the buffer needed to save GGTT data if querying,
+ *         0 on successful save or a negative error code on failure.
+ */
+ssize_t xe_gt_sriov_pf_config_ggtt_save(struct xe_gt *gt, unsigned int vfid,
+					void *buf, size_t size)
+{
+	struct xe_ggtt_node *node;
+
+	xe_gt_assert(gt, IS_SRIOV_PF(gt_to_xe(gt)));
+	xe_gt_assert(gt, vfid);
+	xe_gt_assert(gt, !(!buf ^ !size));
+
+	guard(mutex)(xe_gt_sriov_pf_master_mutex(gt));
+	node = pf_pick_vf_config(gt, vfid)->ggtt_region;
+
+	if (!buf)
+		return xe_ggtt_node_pt_size(node);
+
+	return xe_ggtt_node_save(node, buf, size, vfid);
+}
+
+/**
+ * xe_gt_sriov_pf_config_ggtt_restore() - Restore a VF provisioned GGTT data from a buffer.
+ * @gt: the &xe_gt
+ * @vfid: VF identifier (can't be 0)
+ * @buf: the GGTT data source buffer
+ * @size: the size of the buffer
+ *
+ * This function can only be called on PF.
+ *
+ * Return: 0 on success or a negative error code on failure.
+ */
+int xe_gt_sriov_pf_config_ggtt_restore(struct xe_gt *gt, unsigned int vfid,
+				       const void *buf, size_t size)
+{
+	xe_gt_assert(gt, IS_SRIOV_PF(gt_to_xe(gt)));
+	xe_gt_assert(gt, vfid);
+
+	guard(mutex)(xe_gt_sriov_pf_master_mutex(gt));
+
+	return xe_ggtt_node_load(pf_pick_vf_config(gt, vfid)->ggtt_region, buf, size, vfid);
+}
+
 static u32 pf_get_min_spare_ctxs(struct xe_gt *gt)
 {
 	/* XXX: preliminary */
diff --git a/drivers/gpu/drm/xe/xe_gt_sriov_pf_config.h b/drivers/gpu/drm/xe/xe_gt_sriov_pf_config.h
index 513e6512a575b..0293ba98eb6df 100644
--- a/drivers/gpu/drm/xe/xe_gt_sriov_pf_config.h
+++ b/drivers/gpu/drm/xe/xe_gt_sriov_pf_config.h
@@ -61,6 +61,11 @@ ssize_t xe_gt_sriov_pf_config_save(struct xe_gt *gt, unsigned int vfid, void *bu
 int xe_gt_sriov_pf_config_restore(struct xe_gt *gt, unsigned int vfid,
 				  const void *buf, size_t size);
 
+ssize_t xe_gt_sriov_pf_config_ggtt_save(struct xe_gt *gt, unsigned int vfid,
+					void *buf, size_t size);
+int xe_gt_sriov_pf_config_ggtt_restore(struct xe_gt *gt, unsigned int vfid,
+				       const void *buf, size_t size);
+
 bool xe_gt_sriov_pf_config_is_empty(struct xe_gt *gt, unsigned int vfid);
 
 int xe_gt_sriov_pf_config_init(struct xe_gt *gt);
-- 
2.50.1

[PATCH v3 18/28] drm/xe/pf: Handle GGTT migration data as part of PF control

[Michał Winiarski]

Connect the helpers to allow save and restore of GGTT migration data in
stop_copy / resume device state.

Signed-off-by: Michał Winiarski <michal.winiarski@intel.com>
---
 drivers/gpu/drm/xe/xe_gt_sriov_pf_control.c   |  13 ++
 drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.c | 113 ++++++++++++++++++
 drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.h |   3 +
 3 files changed, 129 insertions(+)

diff --git a/drivers/gpu/drm/xe/xe_gt_sriov_pf_control.c b/drivers/gpu/drm/xe/xe_gt_sriov_pf_control.c
index cb45e97f4c4d9..e7ea9b88fd246 100644
--- a/drivers/gpu/drm/xe/xe_gt_sriov_pf_control.c
+++ b/drivers/gpu/drm/xe/xe_gt_sriov_pf_control.c
@@ -860,6 +860,16 @@ static int pf_handle_vf_save_data(struct xe_gt *gt, unsigned int vfid)
 		return -EAGAIN;
 	}
 
+	if (xe_gt_sriov_pf_migration_save_test(gt, vfid, XE_SRIOV_MIGRATION_DATA_TYPE_GGTT)) {
+		ret = xe_gt_sriov_pf_migration_ggtt_save(gt, vfid);
+		if (ret)
+			return ret;
+
+		xe_gt_sriov_pf_migration_save_clear(gt, vfid, XE_SRIOV_MIGRATION_DATA_TYPE_GGTT);
+
+		return -EAGAIN;
+	}
+
 	return 0;
 }
 
@@ -1066,6 +1076,9 @@ static int pf_handle_vf_restore_data(struct xe_gt *gt, unsigned int vfid)
 	xe_gt_assert(gt, data);
 
 	switch (data->type) {
+	case XE_SRIOV_MIGRATION_DATA_TYPE_GGTT:
+		ret = xe_gt_sriov_pf_migration_ggtt_restore(gt, vfid, data);
+		break;
 	case XE_SRIOV_MIGRATION_DATA_TYPE_GUC:
 		ret = xe_gt_sriov_pf_migration_guc_restore(gt, vfid, data);
 		break;
diff --git a/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.c b/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.c
index fbb451767712a..6f2ee5820bdd4 100644
--- a/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.c
+++ b/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.c
@@ -7,6 +7,9 @@
 
 #include "abi/guc_actions_sriov_abi.h"
 #include "xe_bo.h"
+#include "xe_ggtt.h"
+#include "xe_gt.h"
+#include "xe_gt_sriov_pf_config.h"
 #include "xe_gt_sriov_pf_control.h"
 #include "xe_gt_sriov_pf_helpers.h"
 #include "xe_gt_sriov_pf_migration.h"
@@ -39,6 +42,106 @@ static void pf_dump_mig_data(struct xe_gt *gt, unsigned int vfid,
 	}
 }
 
+static ssize_t pf_migration_ggtt_size(struct xe_gt *gt, unsigned int vfid)
+{
+	if (!xe_gt_is_main_type(gt))
+		return 0;
+
+	return xe_gt_sriov_pf_config_ggtt_save(gt, vfid, NULL, 0);
+}
+
+static int pf_save_vf_ggtt_mig_data(struct xe_gt *gt, unsigned int vfid)
+{
+	struct xe_sriov_migration_data *data;
+	size_t size;
+	int ret;
+
+	size = pf_migration_ggtt_size(gt, vfid);
+	xe_gt_assert(gt, size);
+
+	data = xe_sriov_migration_data_alloc(gt_to_xe(gt));
+	if (!data)
+		return -ENOMEM;
+
+	ret = xe_sriov_migration_data_init(data, gt->tile->id, gt->info.id,
+					   XE_SRIOV_MIGRATION_DATA_TYPE_GGTT, 0, size);
+	if (ret)
+		goto fail;
+
+	ret = xe_gt_sriov_pf_config_ggtt_save(gt, vfid, data->vaddr, size);
+	if (ret)
+		goto fail;
+
+	xe_gt_sriov_dbg_verbose(gt, "VF%u GGTT data save (%zu bytes)\n", vfid, size);
+	pf_dump_mig_data(gt, vfid, data);
+
+	ret = xe_gt_sriov_pf_migration_save_produce(gt, vfid, data);
+	if (ret)
+		goto fail;
+
+	return 0;
+
+fail:
+	xe_sriov_migration_data_free(data);
+	xe_gt_sriov_err(gt, "Failed to save VF%u GGTT data (%pe)\n", vfid, ERR_PTR(ret));
+	return ret;
+}
+
+static int pf_restore_vf_ggtt_mig_data(struct xe_gt *gt, unsigned int vfid,
+				       struct xe_sriov_migration_data *data)
+{
+	int ret;
+
+	xe_gt_sriov_dbg_verbose(gt, "VF%u GGTT data restore (%llu bytes)\n", vfid, data->size);
+	pf_dump_mig_data(gt, vfid, data);
+
+	ret = xe_gt_sriov_pf_config_ggtt_restore(gt, vfid, data->vaddr, data->size);
+	if (ret) {
+		xe_gt_sriov_err(gt, "Failed to restore VF%u GGTT data (%pe)\n",
+				vfid, ERR_PTR(ret));
+		return ret;
+	}
+
+	return 0;
+}
+
+/**
+ * xe_gt_sriov_pf_migration_ggtt_save() - Save VF GGTT migration data.
+ * @gt: the &xe_gt
+ * @vfid: the VF identifier (can't be 0)
+ *
+ * This function is for PF only.
+ *
+ * Return: 0 on success or a negative error code on failure.
+ */
+int xe_gt_sriov_pf_migration_ggtt_save(struct xe_gt *gt, unsigned int vfid)
+{
+	xe_gt_assert(gt, IS_SRIOV_PF(gt_to_xe(gt)));
+	xe_gt_assert(gt, vfid != PFID);
+	xe_gt_assert(gt, vfid <= xe_sriov_pf_get_totalvfs(gt_to_xe(gt)));
+
+	return pf_save_vf_ggtt_mig_data(gt, vfid);
+}
+
+/**
+ * xe_gt_sriov_pf_migration_ggtt_restore() - Restore VF GGTT migration data.
+ * @gt: the &xe_gt
+ * @vfid: the VF identifier (can't be 0)
+ *
+ * This function is for PF only.
+ *
+ * Return: 0 on success or a negative error code on failure.
+ */
+int xe_gt_sriov_pf_migration_ggtt_restore(struct xe_gt *gt, unsigned int vfid,
+					  struct xe_sriov_migration_data *data)
+{
+	xe_gt_assert(gt, IS_SRIOV_PF(gt_to_xe(gt)));
+	xe_gt_assert(gt, vfid != PFID);
+	xe_gt_assert(gt, vfid <= xe_sriov_pf_get_totalvfs(gt_to_xe(gt)));
+
+	return pf_restore_vf_ggtt_mig_data(gt, vfid, data);
+}
+
 /* Return: number of dwords saved/restored/required or a negative error code on failure */
 static int guc_action_vf_save_restore(struct xe_guc *guc, u32 vfid, u32 opcode,
 				      u64 addr, u32 ndwords)
@@ -279,6 +382,13 @@ ssize_t xe_gt_sriov_pf_migration_size(struct xe_gt *gt, unsigned int vfid)
 		size += sizeof(struct xe_sriov_pf_migration_hdr);
 	total += size;
 
+	size = pf_migration_ggtt_size(gt, vfid);
+	if (size < 0)
+		return size;
+	if (size > 0)
+		size += sizeof(struct xe_sriov_pf_migration_hdr);
+	total += size;
+
 	return total;
 }
 
@@ -340,6 +450,9 @@ void xe_gt_sriov_pf_migration_save_init(struct xe_gt *gt, unsigned int vfid)
 
 	xe_gt_assert(gt, pf_migration_guc_size(gt, vfid) > 0);
 	set_bit(XE_SRIOV_MIGRATION_DATA_TYPE_GUC, &migration->save.data_remaining);
+
+	if (pf_migration_ggtt_size(gt, vfid) > 0)
+		set_bit(XE_SRIOV_MIGRATION_DATA_TYPE_GGTT, &migration->save.data_remaining);
 }
 
 /**
diff --git a/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.h b/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.h
index 66a45d6234245..bc201d8f3147a 100644
--- a/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.h
+++ b/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.h
@@ -19,6 +19,9 @@ int xe_gt_sriov_pf_migration_init(struct xe_gt *gt);
 int xe_gt_sriov_pf_migration_guc_save(struct xe_gt *gt, unsigned int vfid);
 int xe_gt_sriov_pf_migration_guc_restore(struct xe_gt *gt, unsigned int vfid,
 					 struct xe_sriov_migration_data *data);
+int xe_gt_sriov_pf_migration_ggtt_save(struct xe_gt *gt, unsigned int vfid);
+int xe_gt_sriov_pf_migration_ggtt_restore(struct xe_gt *gt, unsigned int vfid,
+					  struct xe_sriov_migration_data *data);
 
 ssize_t xe_gt_sriov_pf_migration_size(struct xe_gt *gt, unsigned int vfid);
 
-- 
2.50.1

[PATCH v3 19/28] drm/xe/pf: Handle MMIO migration data as part of PF control

[Michał Winiarski]

Implement the helpers and use them for save and restore of MMIO
migration data in stop_copy / resume device state.

Signed-off-by: Michał Winiarski <michal.winiarski@intel.com>
---
 drivers/gpu/drm/xe/xe_gt_sriov_pf.h           |   2 +
 drivers/gpu/drm/xe/xe_gt_sriov_pf_control.c   |  13 ++
 drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.c | 158 ++++++++++++++++++
 drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.h |   3 +
 4 files changed, 176 insertions(+)

diff --git a/drivers/gpu/drm/xe/xe_gt_sriov_pf.h b/drivers/gpu/drm/xe/xe_gt_sriov_pf.h
index e7fde3f9937af..c0dcbb1054e4d 100644
--- a/drivers/gpu/drm/xe/xe_gt_sriov_pf.h
+++ b/drivers/gpu/drm/xe/xe_gt_sriov_pf.h
@@ -6,6 +6,8 @@
 #ifndef _XE_GT_SRIOV_PF_H_
 #define _XE_GT_SRIOV_PF_H_
 
+#include <linux/types.h>
+
 struct xe_gt;
 
 #ifdef CONFIG_PCI_IOV
diff --git a/drivers/gpu/drm/xe/xe_gt_sriov_pf_control.c b/drivers/gpu/drm/xe/xe_gt_sriov_pf_control.c
index e7ea9b88fd246..7cd7cae950bc7 100644
--- a/drivers/gpu/drm/xe/xe_gt_sriov_pf_control.c
+++ b/drivers/gpu/drm/xe/xe_gt_sriov_pf_control.c
@@ -870,6 +870,16 @@ static int pf_handle_vf_save_data(struct xe_gt *gt, unsigned int vfid)
 		return -EAGAIN;
 	}
 
+	if (xe_gt_sriov_pf_migration_save_test(gt, vfid, XE_SRIOV_MIGRATION_DATA_TYPE_MMIO)) {
+		ret = xe_gt_sriov_pf_migration_mmio_save(gt, vfid);
+		if (ret)
+			return ret;
+
+		xe_gt_sriov_pf_migration_save_clear(gt, vfid, XE_SRIOV_MIGRATION_DATA_TYPE_MMIO);
+
+		return -EAGAIN;
+	}
+
 	return 0;
 }
 
@@ -1079,6 +1089,9 @@ static int pf_handle_vf_restore_data(struct xe_gt *gt, unsigned int vfid)
 	case XE_SRIOV_MIGRATION_DATA_TYPE_GGTT:
 		ret = xe_gt_sriov_pf_migration_ggtt_restore(gt, vfid, data);
 		break;
+	case XE_SRIOV_MIGRATION_DATA_TYPE_MMIO:
+		ret = xe_gt_sriov_pf_migration_mmio_restore(gt, vfid, data);
+		break;
 	case XE_SRIOV_MIGRATION_DATA_TYPE_GUC:
 		ret = xe_gt_sriov_pf_migration_guc_restore(gt, vfid, data);
 		break;
diff --git a/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.c b/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.c
index 6f2ee5820bdd4..5e90aeafeeb41 100644
--- a/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.c
+++ b/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.c
@@ -5,10 +5,13 @@
 
 #include <drm/drm_managed.h>
 
+#include "regs/xe_guc_regs.h"
+
 #include "abi/guc_actions_sriov_abi.h"
 #include "xe_bo.h"
 #include "xe_ggtt.h"
 #include "xe_gt.h"
+#include "xe_gt_sriov_pf.h"
 #include "xe_gt_sriov_pf_config.h"
 #include "xe_gt_sriov_pf_control.h"
 #include "xe_gt_sriov_pf_helpers.h"
@@ -16,6 +19,7 @@
 #include "xe_gt_sriov_printk.h"
 #include "xe_guc_buf.h"
 #include "xe_guc_ct.h"
+#include "xe_mmio.h"
 #include "xe_sriov.h"
 #include "xe_sriov_migration_data.h"
 #include "xe_sriov_pf_migration.h"
@@ -357,6 +361,150 @@ int xe_gt_sriov_pf_migration_guc_restore(struct xe_gt *gt, unsigned int vfid,
 	return pf_restore_vf_guc_state(gt, vfid, data);
 }
 
+static ssize_t pf_migration_mmio_size(struct xe_gt *gt, unsigned int vfid)
+{
+	if (xe_gt_is_media_type(gt))
+		return MED_VF_SW_FLAG_COUNT * sizeof(u32);
+	else
+		return VF_SW_FLAG_COUNT * sizeof(u32);
+}
+
+static int pf_migration_mmio_save(struct xe_gt *gt, unsigned int vfid, void *buf, size_t size)
+{
+	struct xe_mmio mmio;
+	u32 *regs = buf;
+	int n;
+
+	xe_mmio_init_vf_view(&mmio, &gt->mmio, vfid);
+
+	if (size != pf_migration_mmio_size(gt, vfid))
+		return -EINVAL;
+
+	if (xe_gt_is_media_type(gt))
+		for (n = 0; n < MED_VF_SW_FLAG_COUNT; n++)
+			regs[n] = xe_mmio_read32(&gt->mmio, MED_VF_SW_FLAG(n));
+	else
+		for (n = 0; n < VF_SW_FLAG_COUNT; n++)
+			regs[n] = xe_mmio_read32(&gt->mmio, VF_SW_FLAG(n));
+
+	return 0;
+}
+
+static int pf_migration_mmio_restore(struct xe_gt *gt, unsigned int vfid,
+				     const void *buf, size_t size)
+{
+	const u32 *regs = buf;
+	struct xe_mmio mmio;
+	int n;
+
+	xe_mmio_init_vf_view(&mmio, &gt->mmio, vfid);
+
+	if (size != pf_migration_mmio_size(gt, vfid))
+		return -EINVAL;
+
+	if (xe_gt_is_media_type(gt))
+		for (n = 0; n < MED_VF_SW_FLAG_COUNT; n++)
+			xe_mmio_write32(&gt->mmio, MED_VF_SW_FLAG(n), regs[n]);
+	else
+		for (n = 0; n < VF_SW_FLAG_COUNT; n++)
+			xe_mmio_write32(&gt->mmio, VF_SW_FLAG(n), regs[n]);
+
+	return 0;
+}
+
+static int pf_save_vf_mmio_mig_data(struct xe_gt *gt, unsigned int vfid)
+{
+	struct xe_sriov_migration_data *data;
+	size_t size;
+	int ret;
+
+	size = pf_migration_mmio_size(gt, vfid);
+	xe_gt_assert(gt, size);
+
+	data = xe_sriov_migration_data_alloc(gt_to_xe(gt));
+	if (!data)
+		return -ENOMEM;
+
+	ret = xe_sriov_migration_data_init(data, gt->tile->id, gt->info.id,
+					   XE_SRIOV_MIGRATION_DATA_TYPE_MMIO, 0, size);
+	if (ret)
+		goto fail;
+
+	ret = pf_migration_mmio_save(gt, vfid, data->vaddr, size);
+	if (ret)
+		goto fail;
+
+	xe_gt_sriov_dbg_verbose(gt, "VF%u MMIO data save (%zu bytes)\n", vfid, size);
+	pf_dump_mig_data(gt, vfid, data);
+
+	ret = xe_gt_sriov_pf_migration_save_produce(gt, vfid, data);
+	if (ret)
+		goto fail;
+
+	return 0;
+
+fail:
+	xe_sriov_migration_data_free(data);
+	xe_gt_sriov_err(gt, "Failed to save VF%u MMIO data (%pe)\n", vfid, ERR_PTR(ret));
+	return ret;
+}
+
+static int pf_restore_vf_mmio_mig_data(struct xe_gt *gt, unsigned int vfid,
+				       struct xe_sriov_migration_data *data)
+{
+	int ret;
+
+	xe_gt_sriov_dbg_verbose(gt, "VF%u MMIO data restore (%llu bytes)\n", vfid, data->size);
+	pf_dump_mig_data(gt, vfid, data);
+
+	ret = pf_migration_mmio_restore(gt, vfid, data->vaddr, data->size);
+	if (ret) {
+		xe_gt_sriov_err(gt, "Failed to restore VF%u MMIO data (%pe)\n",
+				vfid, ERR_PTR(ret));
+
+		return ret;
+	}
+
+	return 0;
+}
+
+/**
+ * xe_gt_sriov_pf_migration_mmio_save() - Save VF MMIO migration data.
+ * @gt: the &xe_gt
+ * @vfid: the VF identifier (can't be 0)
+ *
+ * This function is for PF only.
+ *
+ * Return: 0 on success or a negative error code on failure.
+ */
+int xe_gt_sriov_pf_migration_mmio_save(struct xe_gt *gt, unsigned int vfid)
+{
+	xe_gt_assert(gt, IS_SRIOV_PF(gt_to_xe(gt)));
+	xe_gt_assert(gt, vfid != PFID);
+	xe_gt_assert(gt, vfid <= xe_sriov_pf_get_totalvfs(gt_to_xe(gt)));
+
+	return pf_save_vf_mmio_mig_data(gt, vfid);
+}
+
+/**
+ * xe_gt_sriov_pf_migration_mmio_restore() - Restore VF MMIO migration data.
+ * @gt: the &xe_gt
+ * @vfid: the VF identifier (can't be 0)
+ *
+ * This function is for PF only.
+ *
+ * Return: 0 on success or a negative error code on failure.
+ */
+int xe_gt_sriov_pf_migration_mmio_restore(struct xe_gt *gt, unsigned int vfid,
+					  struct xe_sriov_migration_data *data)
+{
+	xe_gt_assert(gt, IS_SRIOV_PF(gt_to_xe(gt)));
+	xe_gt_assert(gt, vfid != PFID);
+	xe_gt_assert(gt, vfid <= xe_sriov_pf_get_totalvfs(gt_to_xe(gt)));
+
+	return pf_restore_vf_mmio_mig_data(gt, vfid, data);
+}
+
 /**
  * xe_gt_sriov_pf_migration_size() - Total size of migration data from all components within a GT.
  * @gt: the &xe_gt
@@ -389,6 +537,13 @@ ssize_t xe_gt_sriov_pf_migration_size(struct xe_gt *gt, unsigned int vfid)
 		size += sizeof(struct xe_sriov_pf_migration_hdr);
 	total += size;
 
+	size = pf_migration_mmio_size(gt, vfid);
+	if (size < 0)
+		return size;
+	if (size > 0)
+		size += sizeof(struct xe_sriov_pf_migration_hdr);
+	total += size;
+
 	return total;
 }
 
@@ -453,6 +608,9 @@ void xe_gt_sriov_pf_migration_save_init(struct xe_gt *gt, unsigned int vfid)
 
 	if (pf_migration_ggtt_size(gt, vfid) > 0)
 		set_bit(XE_SRIOV_MIGRATION_DATA_TYPE_GGTT, &migration->save.data_remaining);
+
+	xe_gt_assert(gt, pf_migration_mmio_size(gt, vfid) > 0);
+	set_bit(XE_SRIOV_MIGRATION_DATA_TYPE_MMIO, &migration->save.data_remaining);
 }
 
 /**
diff --git a/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.h b/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.h
index bc201d8f3147a..b0eec94fea3a6 100644
--- a/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.h
+++ b/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.h
@@ -22,6 +22,9 @@ int xe_gt_sriov_pf_migration_guc_restore(struct xe_gt *gt, unsigned int vfid,
 int xe_gt_sriov_pf_migration_ggtt_save(struct xe_gt *gt, unsigned int vfid);
 int xe_gt_sriov_pf_migration_ggtt_restore(struct xe_gt *gt, unsigned int vfid,
 					  struct xe_sriov_migration_data *data);
+int xe_gt_sriov_pf_migration_mmio_save(struct xe_gt *gt, unsigned int vfid);
+int xe_gt_sriov_pf_migration_mmio_restore(struct xe_gt *gt, unsigned int vfid,
+					  struct xe_sriov_migration_data *data);
 
 ssize_t xe_gt_sriov_pf_migration_size(struct xe_gt *gt, unsigned int vfid);
 
-- 
2.50.1

[PATCH v3 20/28] drm/xe/pf: Add helper to retrieve VF's LMEM object

[Michał Winiarski]

From: Lukasz Laguna <lukasz.laguna@intel.com>

Instead of accessing VF's lmem_obj directly, introduce a helper function
to make the access more convenient.

Signed-off-by: Lukasz Laguna <lukasz.laguna@intel.com>
Signed-off-by: Michał Winiarski <michal.winiarski@intel.com>
Reviewed-by: Michal Wajdeczko <michal.wajdeczko@intel.com>
---
 drivers/gpu/drm/xe/xe_gt_sriov_pf_config.c | 26 ++++++++++++++++++++++
 drivers/gpu/drm/xe/xe_gt_sriov_pf_config.h |  1 +
 2 files changed, 27 insertions(+)

diff --git a/drivers/gpu/drm/xe/xe_gt_sriov_pf_config.c b/drivers/gpu/drm/xe/xe_gt_sriov_pf_config.c
index 55444883f2ac3..0815f761969f0 100644
--- a/drivers/gpu/drm/xe/xe_gt_sriov_pf_config.c
+++ b/drivers/gpu/drm/xe/xe_gt_sriov_pf_config.c
@@ -1651,6 +1651,32 @@ int xe_gt_sriov_pf_config_bulk_set_lmem(struct xe_gt *gt, unsigned int vfid,
 					   "LMEM", n, err);
 }
 
+static struct xe_bo *pf_get_vf_config_lmem_obj(struct xe_gt *gt, unsigned int vfid)
+{
+	struct xe_gt_sriov_config *config = pf_pick_vf_config(gt, vfid);
+
+	return config->lmem_obj;
+}
+
+/**
+ * xe_gt_sriov_pf_config_get_lmem_obj() - Take a reference to the struct &xe_bo backing VF LMEM.
+ * @gt: the &xe_gt
+ * @vfid: the VF identifier (can't be 0)
+ *
+ * This function can only be called on PF.
+ * The caller is responsible for calling xe_bo_put() on the returned object.
+ *
+ * Return: pointer to struct &xe_bo backing VF LMEM (if any).
+ */
+struct xe_bo *xe_gt_sriov_pf_config_get_lmem_obj(struct xe_gt *gt, unsigned int vfid)
+{
+	xe_gt_assert(gt, vfid);
+
+	guard(mutex)(xe_gt_sriov_pf_master_mutex(gt));
+
+	return xe_bo_get(pf_get_vf_config_lmem_obj(gt, vfid));
+}
+
 static u64 pf_query_free_lmem(struct xe_gt *gt)
 {
 	struct xe_tile *tile = gt->tile;
diff --git a/drivers/gpu/drm/xe/xe_gt_sriov_pf_config.h b/drivers/gpu/drm/xe/xe_gt_sriov_pf_config.h
index 0293ba98eb6df..d9fbc6c30b158 100644
--- a/drivers/gpu/drm/xe/xe_gt_sriov_pf_config.h
+++ b/drivers/gpu/drm/xe/xe_gt_sriov_pf_config.h
@@ -36,6 +36,7 @@ int xe_gt_sriov_pf_config_set_lmem(struct xe_gt *gt, unsigned int vfid, u64 size
 int xe_gt_sriov_pf_config_set_fair_lmem(struct xe_gt *gt, unsigned int vfid, unsigned int num_vfs);
 int xe_gt_sriov_pf_config_bulk_set_lmem(struct xe_gt *gt, unsigned int vfid, unsigned int num_vfs,
 					u64 size);
+struct xe_bo *xe_gt_sriov_pf_config_get_lmem_obj(struct xe_gt *gt, unsigned int vfid);
 
 u32 xe_gt_sriov_pf_config_get_exec_quantum(struct xe_gt *gt, unsigned int vfid);
 int xe_gt_sriov_pf_config_set_exec_quantum(struct xe_gt *gt, unsigned int vfid, u32 exec_quantum);
-- 
2.50.1

[PATCH v3 21/28] drm/xe/migrate: Add function to copy of VRAM data in chunks

[Michał Winiarski]

From: Lukasz Laguna <lukasz.laguna@intel.com>

Introduce a new function to copy data between VRAM and sysmem objects.
The existing xe_migrate_copy() is tailored for eviction and restore
operations, which involves additional logic and operates on entire
objects.
The xe_migrate_vram_copy_chunk() allows copying chunks of data to or
from a dedicated buffer object, which is essential in case of VF
migration.

Signed-off-by: Lukasz Laguna <lukasz.laguna@intel.com>
Signed-off-by: Michał Winiarski <michal.winiarski@intel.com>
---
 drivers/gpu/drm/xe/xe_migrate.c | 128 ++++++++++++++++++++++++++++++--
 drivers/gpu/drm/xe/xe_migrate.h |   8 ++
 2 files changed, 131 insertions(+), 5 deletions(-)

diff --git a/drivers/gpu/drm/xe/xe_migrate.c b/drivers/gpu/drm/xe/xe_migrate.c
index 56a5804726e96..dbe9320863ab0 100644
--- a/drivers/gpu/drm/xe/xe_migrate.c
+++ b/drivers/gpu/drm/xe/xe_migrate.c
@@ -29,6 +29,7 @@
 #include "xe_lrc.h"
 #include "xe_map.h"
 #include "xe_mocs.h"
+#include "xe_printk.h"
 #include "xe_pt.h"
 #include "xe_res_cursor.h"
 #include "xe_sa.h"
@@ -1210,6 +1211,128 @@ struct xe_exec_queue *xe_migrate_exec_queue(struct xe_migrate *migrate)
 	return migrate->q;
 }
 
+/**
+ * xe_migrate_vram_copy_chunk() - Copy a chunk of a VRAM buffer object.
+ * @vram_bo: The VRAM buffer object.
+ * @vram_offset: The VRAM offset.
+ * @sysmem_bo: The sysmem buffer object.
+ * @sysmem_offset: The sysmem offset.
+ * @size: The size of VRAM chunk to copy.
+ * @dir: The direction of the copy operation.
+ *
+ * Copies a portion of a buffer object between VRAM and system memory.
+ * On Xe2 platforms that support flat CCS, VRAM data is decompressed when
+ * copying to system memory.
+ *
+ * Return: Pointer to a dma_fence representing the last copy batch, or
+ * an error pointer on failure. If there is a failure, any copy operation
+ * started by the function call has been synced.
+ */
+struct dma_fence *xe_migrate_vram_copy_chunk(struct xe_bo *vram_bo, u64 vram_offset,
+					     struct xe_bo *sysmem_bo, u64 sysmem_offset,
+					     u64 size, enum xe_migrate_copy_dir dir)
+{
+	struct xe_device *xe = xe_bo_device(vram_bo);
+	struct xe_tile *tile = vram_bo->tile;
+	struct xe_gt *gt = tile->primary_gt;
+	struct xe_migrate *m = tile->migrate;
+	struct dma_fence *fence = NULL;
+	struct ttm_resource *vram = vram_bo->ttm.resource;
+	struct ttm_resource *sysmem = sysmem_bo->ttm.resource;
+	struct xe_res_cursor vram_it, sysmem_it;
+	u64 vram_L0_ofs, sysmem_L0_ofs;
+	u32 vram_L0_pt, sysmem_L0_pt;
+	u64 vram_L0, sysmem_L0;
+	bool to_sysmem = (dir == XE_MIGRATE_COPY_TO_SRAM);
+	bool use_comp_pat = to_sysmem &&
+		GRAPHICS_VER(xe) >= 20 && xe_device_has_flat_ccs(xe);
+	int pass = 0;
+	int err;
+
+	xe_assert(xe, IS_ALIGNED(vram_offset | sysmem_offset | size, PAGE_SIZE));
+	xe_assert(xe, xe_bo_is_vram(vram_bo));
+	xe_assert(xe, !xe_bo_is_vram(sysmem_bo));
+	xe_assert(xe, !range_overflows(vram_offset, size, (u64)vram_bo->ttm.base.size));
+	xe_assert(xe, !range_overflows(sysmem_offset, size, (u64)sysmem_bo->ttm.base.size));
+
+	xe_res_first(vram, vram_offset, size, &vram_it);
+	xe_res_first_sg(xe_bo_sg(sysmem_bo), sysmem_offset, size, &sysmem_it);
+
+	while (size) {
+		u32 pte_flags = PTE_UPDATE_FLAG_IS_VRAM;
+		u32 batch_size = 2; /* arb_clear() + MI_BATCH_BUFFER_END */
+		struct xe_sched_job *job;
+		struct xe_bb *bb;
+		u32 update_idx;
+		bool usm = xe->info.has_usm;
+		u32 avail_pts = max_mem_transfer_per_pass(xe) / LEVEL0_PAGE_TABLE_ENCODE_SIZE;
+
+		sysmem_L0 = xe_migrate_res_sizes(m, &sysmem_it);
+		vram_L0 = min(xe_migrate_res_sizes(m, &vram_it), sysmem_L0);
+
+		xe_dbg(xe, "Pass %u, size: %llu\n", pass++, vram_L0);
+
+		pte_flags |= use_comp_pat ? PTE_UPDATE_FLAG_IS_COMP_PTE : 0;
+		batch_size += pte_update_size(m, pte_flags, vram, &vram_it, &vram_L0,
+					      &vram_L0_ofs, &vram_L0_pt, 0, 0, avail_pts);
+
+		batch_size += pte_update_size(m, 0, sysmem, &sysmem_it, &vram_L0, &sysmem_L0_ofs,
+					      &sysmem_L0_pt, 0, avail_pts, avail_pts);
+		batch_size += EMIT_COPY_DW;
+
+		bb = xe_bb_new(gt, batch_size, usm);
+		if (IS_ERR(bb)) {
+			err = PTR_ERR(bb);
+			return ERR_PTR(err);
+		}
+
+		if (xe_migrate_allow_identity(vram_L0, &vram_it))
+			xe_res_next(&vram_it, vram_L0);
+		else
+			emit_pte(m, bb, vram_L0_pt, true, use_comp_pat, &vram_it, vram_L0, vram);
+
+		emit_pte(m, bb, sysmem_L0_pt, false, false, &sysmem_it, vram_L0, sysmem);
+
+		bb->cs[bb->len++] = MI_BATCH_BUFFER_END;
+		update_idx = bb->len;
+
+		if (to_sysmem)
+			emit_copy(gt, bb, vram_L0_ofs, sysmem_L0_ofs, vram_L0, XE_PAGE_SIZE);
+		else
+			emit_copy(gt, bb, sysmem_L0_ofs, vram_L0_ofs, vram_L0, XE_PAGE_SIZE);
+
+		job = xe_bb_create_migration_job(m->q, bb, xe_migrate_batch_base(m, usm),
+						 update_idx);
+		if (IS_ERR(job)) {
+			xe_bb_free(bb, NULL);
+			err = PTR_ERR(job);
+			return ERR_PTR(err);
+		}
+
+		xe_sched_job_add_migrate_flush(job, MI_INVALIDATE_TLB);
+
+		xe_assert(xe, dma_resv_test_signaled(vram_bo->ttm.base.resv,
+						     DMA_RESV_USAGE_BOOKKEEP));
+		xe_assert(xe, dma_resv_test_signaled(sysmem_bo->ttm.base.resv,
+						     DMA_RESV_USAGE_BOOKKEEP));
+
+		scoped_guard(mutex, &m->job_mutex) {
+			xe_sched_job_arm(job);
+			dma_fence_put(fence);
+			fence = dma_fence_get(&job->drm.s_fence->finished);
+			xe_sched_job_push(job);
+
+			dma_fence_put(m->fence);
+			m->fence = dma_fence_get(fence);
+		}
+
+		xe_bb_free(bb, fence);
+		size -= vram_L0;
+	}
+
+	return fence;
+}
+
 static void emit_clear_link_copy(struct xe_gt *gt, struct xe_bb *bb, u64 src_ofs,
 				 u32 size, u32 pitch)
 {
@@ -1912,11 +2035,6 @@ static bool xe_migrate_vram_use_pde(struct drm_pagemap_addr *sram_addr,
 	return true;
 }
 
-enum xe_migrate_copy_dir {
-	XE_MIGRATE_COPY_TO_VRAM,
-	XE_MIGRATE_COPY_TO_SRAM,
-};
-
 #define XE_CACHELINE_BYTES	64ull
 #define XE_CACHELINE_MASK	(XE_CACHELINE_BYTES - 1)
 
diff --git a/drivers/gpu/drm/xe/xe_migrate.h b/drivers/gpu/drm/xe/xe_migrate.h
index 4fad324b62535..d7bcc6ad8464e 100644
--- a/drivers/gpu/drm/xe/xe_migrate.h
+++ b/drivers/gpu/drm/xe/xe_migrate.h
@@ -28,6 +28,11 @@ struct xe_vma;
 
 enum xe_sriov_vf_ccs_rw_ctxs;
 
+enum xe_migrate_copy_dir {
+	XE_MIGRATE_COPY_TO_VRAM,
+	XE_MIGRATE_COPY_TO_SRAM,
+};
+
 /**
  * struct xe_migrate_pt_update_ops - Callbacks for the
  * xe_migrate_update_pgtables() function.
@@ -131,6 +136,9 @@ int xe_migrate_ccs_rw_copy(struct xe_tile *tile, struct xe_exec_queue *q,
 
 struct xe_lrc *xe_migrate_lrc(struct xe_migrate *migrate);
 struct xe_exec_queue *xe_migrate_exec_queue(struct xe_migrate *migrate);
+struct dma_fence *xe_migrate_vram_copy_chunk(struct xe_bo *vram_bo, u64 vram_offset,
+					     struct xe_bo *sysmem_bo, u64 sysmem_offset,
+					     u64 size, enum xe_migrate_copy_dir dir);
 int xe_migrate_access_memory(struct xe_migrate *m, struct xe_bo *bo,
 			     unsigned long offset, void *buf, int len,
 			     int write);
-- 
2.50.1

[PATCH v3 22/28] drm/xe/pf: Handle VRAM migration data as part of PF control

[Michał Winiarski]

Connect the helpers to allow save and restore of VRAM migration data in
stop_copy / resume device state.

Co-developed-by: Lukasz Laguna <lukasz.laguna@intel.com>
Signed-off-by: Lukasz Laguna <lukasz.laguna@intel.com>
Signed-off-by: Michał Winiarski <michal.winiarski@intel.com>
---
 drivers/gpu/drm/xe/xe_gt_sriov_pf_control.c   |  15 ++
 drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.c | 213 ++++++++++++++++++
 drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.h |   5 +
 .../drm/xe/xe_gt_sriov_pf_migration_types.h   |   2 +
 drivers/gpu/drm/xe/xe_sriov_pf_control.c      |   3 +
 5 files changed, 238 insertions(+)

diff --git a/drivers/gpu/drm/xe/xe_gt_sriov_pf_control.c b/drivers/gpu/drm/xe/xe_gt_sriov_pf_control.c
index 7cd7cae950bc7..de0f63610e780 100644
--- a/drivers/gpu/drm/xe/xe_gt_sriov_pf_control.c
+++ b/drivers/gpu/drm/xe/xe_gt_sriov_pf_control.c
@@ -880,6 +880,18 @@ static int pf_handle_vf_save_data(struct xe_gt *gt, unsigned int vfid)
 		return -EAGAIN;
 	}
 
+	if (xe_gt_sriov_pf_migration_save_test(gt, vfid, XE_SRIOV_MIGRATION_DATA_TYPE_VRAM)) {
+		ret = xe_gt_sriov_pf_migration_vram_save(gt, vfid);
+		if (ret == -EAGAIN)
+			return -EAGAIN;
+		else if (ret)
+			return ret;
+
+		xe_gt_sriov_pf_migration_save_clear(gt, vfid, XE_SRIOV_MIGRATION_DATA_TYPE_VRAM);
+
+		return -EAGAIN;
+	}
+
 	return 0;
 }
 
@@ -1095,6 +1107,9 @@ static int pf_handle_vf_restore_data(struct xe_gt *gt, unsigned int vfid)
 	case XE_SRIOV_MIGRATION_DATA_TYPE_GUC:
 		ret = xe_gt_sriov_pf_migration_guc_restore(gt, vfid, data);
 		break;
+	case XE_SRIOV_MIGRATION_DATA_TYPE_VRAM:
+		ret = xe_gt_sriov_pf_migration_vram_restore(gt, vfid, data);
+		break;
 	default:
 		xe_gt_sriov_notice(gt, "Skipping VF%u unknown data type: %d\n", vfid, data->type);
 		break;
diff --git a/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.c b/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.c
index 5e90aeafeeb41..7ccd43545fdac 100644
--- a/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.c
+++ b/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.c
@@ -19,6 +19,7 @@
 #include "xe_gt_sriov_printk.h"
 #include "xe_guc_buf.h"
 #include "xe_guc_ct.h"
+#include "xe_migrate.h"
 #include "xe_mmio.h"
 #include "xe_sriov.h"
 #include "xe_sriov_migration_data.h"
@@ -505,6 +506,207 @@ int xe_gt_sriov_pf_migration_mmio_restore(struct xe_gt *gt, unsigned int vfid,
 	return pf_restore_vf_mmio_mig_data(gt, vfid, data);
 }
 
+static ssize_t pf_migration_vram_size(struct xe_gt *gt, unsigned int vfid)
+{
+	if (!xe_gt_is_main_type(gt))
+		return 0;
+
+	return xe_gt_sriov_pf_config_get_lmem(gt, vfid);
+}
+
+static struct dma_fence *__pf_save_restore_vram(struct xe_gt *gt, unsigned int vfid,
+						struct xe_bo *vram, u64 vram_offset,
+						struct xe_bo *sysmem, u64 sysmem_offset,
+						size_t size, bool save)
+{
+	struct dma_fence *ret = NULL;
+	struct drm_exec exec;
+	int err;
+
+	drm_exec_init(&exec, DRM_EXEC_INTERRUPTIBLE_WAIT, 0);
+	drm_exec_until_all_locked(&exec) {
+		err = drm_exec_lock_obj(&exec, &vram->ttm.base);
+		drm_exec_retry_on_contention(&exec);
+		if (err) {
+			ret = ERR_PTR(err);
+			goto err;
+		}
+
+		err = drm_exec_lock_obj(&exec, &sysmem->ttm.base);
+		drm_exec_retry_on_contention(&exec);
+		if (err) {
+			ret = ERR_PTR(err);
+			goto err;
+		}
+	}
+
+	ret = xe_migrate_vram_copy_chunk(vram, vram_offset, sysmem, sysmem_offset, size,
+					 save ? XE_MIGRATE_COPY_TO_SRAM : XE_MIGRATE_COPY_TO_VRAM);
+
+err:
+	drm_exec_fini(&exec);
+
+	return ret;
+}
+
+#define PF_VRAM_SAVE_RESTORE_TIMEOUT (5 * HZ)
+static int pf_save_vram_chunk(struct xe_gt *gt, unsigned int vfid,
+			      struct xe_bo *src_vram, u64 src_vram_offset,
+			      size_t size)
+{
+	struct xe_sriov_migration_data *data;
+	struct dma_fence *fence;
+	int ret;
+
+	data = xe_sriov_migration_data_alloc(gt_to_xe(gt));
+	if (!data)
+		return -ENOMEM;
+
+	ret = xe_sriov_migration_data_init(data, gt->tile->id, gt->info.id,
+					   XE_SRIOV_MIGRATION_DATA_TYPE_VRAM,
+					   src_vram_offset, size);
+	if (ret)
+		goto fail;
+
+	fence = __pf_save_restore_vram(gt, vfid,
+				       src_vram, src_vram_offset,
+				       data->bo, 0, size, true);
+
+	ret = dma_fence_wait_timeout(fence, false, PF_VRAM_SAVE_RESTORE_TIMEOUT);
+	dma_fence_put(fence);
+	if (!ret) {
+		ret = -ETIME;
+		goto fail;
+	}
+
+	xe_gt_sriov_dbg_verbose(gt, "VF%u VRAM data save (%zu bytes)\n", vfid, size);
+	pf_dump_mig_data(gt, vfid, data);
+
+	ret = xe_gt_sriov_pf_migration_save_produce(gt, vfid, data);
+	if (ret)
+		goto fail;
+
+	return 0;
+
+fail:
+	xe_sriov_migration_data_free(data);
+	return ret;
+}
+
+#define VF_VRAM_STATE_CHUNK_MAX_SIZE SZ_512M
+static int pf_save_vf_vram_mig_data(struct xe_gt *gt, unsigned int vfid)
+{
+	struct xe_gt_sriov_migration_data *migration = pf_pick_gt_migration(gt, vfid);
+	loff_t *offset = &migration->save.vram_offset;
+	struct xe_bo *vram;
+	size_t vram_size, chunk_size;
+	int ret;
+
+	vram = xe_gt_sriov_pf_config_get_lmem_obj(gt, vfid);
+	if (!vram)
+		return -ENXIO;
+
+	vram_size = xe_bo_size(vram);
+
+	xe_gt_assert(gt, *offset < vram_size);
+
+	chunk_size = min(vram_size - *offset, VF_VRAM_STATE_CHUNK_MAX_SIZE);
+
+	ret = pf_save_vram_chunk(gt, vfid, vram, *offset, chunk_size);
+	if (ret)
+		goto fail;
+
+	*offset += chunk_size;
+
+	xe_bo_put(vram);
+
+	if (*offset < vram_size)
+		return -EAGAIN;
+
+	return 0;
+
+fail:
+	xe_bo_put(vram);
+	xe_gt_sriov_err(gt, "Failed to save VF%u VRAM data (%pe)\n", vfid, ERR_PTR(ret));
+	return ret;
+}
+
+static int pf_restore_vf_vram_mig_data(struct xe_gt *gt, unsigned int vfid,
+				       struct xe_sriov_migration_data *data)
+{
+	u64 end = data->hdr.offset + data->hdr.size;
+	struct dma_fence *fence;
+	struct xe_bo *vram;
+	size_t size;
+	int ret = 0;
+
+	vram = xe_gt_sriov_pf_config_get_lmem_obj(gt, vfid);
+	if (!vram)
+		return -ENXIO;
+
+	size = xe_bo_size(vram);
+
+	if (end > size || end < data->hdr.size) {
+		ret = -EINVAL;
+		goto err;
+	}
+
+	xe_gt_sriov_dbg_verbose(gt, "VF%u VRAM data restore (%llu bytes)\n", vfid, data->size);
+	pf_dump_mig_data(gt, vfid, data);
+
+	fence = __pf_save_restore_vram(gt, vfid, vram, data->hdr.offset,
+				       data->bo, 0, data->hdr.size, false);
+	ret = dma_fence_wait_timeout(fence, false, PF_VRAM_SAVE_RESTORE_TIMEOUT);
+	dma_fence_put(fence);
+	if (!ret) {
+		ret = -ETIME;
+		goto err;
+	}
+
+	return 0;
+err:
+	xe_bo_put(vram);
+	xe_gt_sriov_err(gt, "Failed to restore VF%u VRAM data (%pe)\n", vfid, ERR_PTR(ret));
+	return ret;
+}
+
+/**
+ * xe_gt_sriov_pf_migration_vram_save() - Save VF VRAM migration data.
+ * @gt: the &xe_gt
+ * @vfid: the VF identifier (can't be 0)
+ *
+ * This function is for PF only.
+ *
+ * Return: 0 on success or a negative error code on failure.
+ */
+int xe_gt_sriov_pf_migration_vram_save(struct xe_gt *gt, unsigned int vfid)
+{
+	xe_gt_assert(gt, IS_SRIOV_PF(gt_to_xe(gt)));
+	xe_gt_assert(gt, vfid != PFID);
+	xe_gt_assert(gt, vfid <= xe_sriov_pf_get_totalvfs(gt_to_xe(gt)));
+
+	return pf_save_vf_vram_mig_data(gt, vfid);
+}
+
+/**
+ * xe_gt_sriov_pf_migration_vram_restore() - Restore VF VRAM migration data.
+ * @gt: the &xe_gt
+ * @vfid: the VF identifier (can't be 0)
+ *
+ * This function is for PF only.
+ *
+ * Return: 0 on success or a negative error code on failure.
+ */
+int xe_gt_sriov_pf_migration_vram_restore(struct xe_gt *gt, unsigned int vfid,
+					  struct xe_sriov_migration_data *data)
+{
+	xe_gt_assert(gt, IS_SRIOV_PF(gt_to_xe(gt)));
+	xe_gt_assert(gt, vfid != PFID);
+	xe_gt_assert(gt, vfid <= xe_sriov_pf_get_totalvfs(gt_to_xe(gt)));
+
+	return pf_restore_vf_vram_mig_data(gt, vfid, data);
+}
+
 /**
  * xe_gt_sriov_pf_migration_size() - Total size of migration data from all components within a GT.
  * @gt: the &xe_gt
@@ -544,6 +746,13 @@ ssize_t xe_gt_sriov_pf_migration_size(struct xe_gt *gt, unsigned int vfid)
 		size += sizeof(struct xe_sriov_pf_migration_hdr);
 	total += size;
 
+	size = pf_migration_vram_size(gt, vfid);
+	if (size < 0)
+		return size;
+	if (size > 0)
+		size += sizeof(struct xe_sriov_pf_migration_hdr);
+	total += size;
+
 	return total;
 }
 
@@ -602,6 +811,7 @@ void xe_gt_sriov_pf_migration_save_init(struct xe_gt *gt, unsigned int vfid)
 	struct xe_gt_sriov_migration_data *migration = pf_pick_gt_migration(gt, vfid);
 
 	migration->save.data_remaining = 0;
+	migration->save.vram_offset = 0;
 
 	xe_gt_assert(gt, pf_migration_guc_size(gt, vfid) > 0);
 	set_bit(XE_SRIOV_MIGRATION_DATA_TYPE_GUC, &migration->save.data_remaining);
@@ -611,6 +821,9 @@ void xe_gt_sriov_pf_migration_save_init(struct xe_gt *gt, unsigned int vfid)
 
 	xe_gt_assert(gt, pf_migration_mmio_size(gt, vfid) > 0);
 	set_bit(XE_SRIOV_MIGRATION_DATA_TYPE_MMIO, &migration->save.data_remaining);
+
+	if (pf_migration_vram_size(gt, vfid) > 0)
+		set_bit(XE_SRIOV_MIGRATION_DATA_TYPE_VRAM, &migration->save.data_remaining);
 }
 
 /**
diff --git a/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.h b/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.h
index b0eec94fea3a6..85b43b2cceb73 100644
--- a/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.h
+++ b/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.h
@@ -25,6 +25,11 @@ int xe_gt_sriov_pf_migration_ggtt_restore(struct xe_gt *gt, unsigned int vfid,
 int xe_gt_sriov_pf_migration_mmio_save(struct xe_gt *gt, unsigned int vfid);
 int xe_gt_sriov_pf_migration_mmio_restore(struct xe_gt *gt, unsigned int vfid,
 					  struct xe_sriov_migration_data *data);
+int xe_gt_sriov_pf_migration_vram_save(struct xe_gt *gt, unsigned int vfid);
+int xe_gt_sriov_pf_migration_vram_restore(struct xe_gt *gt, unsigned int vfid,
+					  struct xe_sriov_migration_data *data);
+
+void xe_gt_sriov_pf_migration_save_init(struct xe_gt *gt, unsigned int vfid);
 
 ssize_t xe_gt_sriov_pf_migration_size(struct xe_gt *gt, unsigned int vfid);
 
diff --git a/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration_types.h b/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration_types.h
index 9f24878690d9c..f50c64241e9c0 100644
--- a/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration_types.h
+++ b/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration_types.h
@@ -20,6 +20,8 @@ struct xe_gt_sriov_migration_data {
 	struct {
 		/** @save.data_remaining: bitmap of migration types that need to be saved */
 		unsigned long data_remaining;
+		/** @save.vram_offset: last saved offset within VRAM, used for chunked VRAM save */
+		loff_t vram_offset;
 	} save;
 };
 
diff --git a/drivers/gpu/drm/xe/xe_sriov_pf_control.c b/drivers/gpu/drm/xe/xe_sriov_pf_control.c
index c2768848daba1..aac8ecb861545 100644
--- a/drivers/gpu/drm/xe/xe_sriov_pf_control.c
+++ b/drivers/gpu/drm/xe/xe_sriov_pf_control.c
@@ -5,6 +5,7 @@
 
 #include "xe_device.h"
 #include "xe_gt_sriov_pf_control.h"
+#include "xe_gt_sriov_pf_migration.h"
 #include "xe_sriov_migration_data.h"
 #include "xe_sriov_pf_control.h"
 #include "xe_sriov_printk.h"
@@ -171,6 +172,8 @@ int xe_sriov_pf_control_trigger_save_vf(struct xe_device *xe, unsigned int vfid)
 		return ret;
 
 	for_each_gt(gt, xe, id) {
+		xe_gt_sriov_pf_migration_save_init(gt, vfid);
+
 		ret = xe_gt_sriov_pf_control_trigger_save_vf(gt, vfid);
 		if (ret)
 			return ret;
-- 
2.50.1

[PATCH v3 23/28] drm/xe/pf: Add wait helper for VF FLR

[Michał Winiarski]

VF FLR requires additional processing done by PF driver.
The processing is done after FLR is already finished from PCIe
perspective.
In order to avoid a scenario where migration state transitions while
PF processing is still in progress, additional synchronization
point is needed.
Add a helper that will be used as part of VF driver struct
pci_error_handlers .reset_done() callback.

Signed-off-by: Michał Winiarski <michal.winiarski@intel.com>
Reviewed-by: Michal Wajdeczko <michal.wajdeczko@intel.com>
---
 drivers/gpu/drm/xe/xe_sriov_pf_control.c | 24 ++++++++++++++++++++++++
 drivers/gpu/drm/xe/xe_sriov_pf_control.h |  1 +
 2 files changed, 25 insertions(+)

diff --git a/drivers/gpu/drm/xe/xe_sriov_pf_control.c b/drivers/gpu/drm/xe/xe_sriov_pf_control.c
index aac8ecb861545..bed488476706d 100644
--- a/drivers/gpu/drm/xe/xe_sriov_pf_control.c
+++ b/drivers/gpu/drm/xe/xe_sriov_pf_control.c
@@ -123,6 +123,30 @@ int xe_sriov_pf_control_reset_vf(struct xe_device *xe, unsigned int vfid)
 	return result;
 }
 
+/**
+ * xe_sriov_pf_control_wait_flr() - Wait for a VF reset (FLR) to complete.
+ * @xe: the &xe_device
+ * @vfid: the VF identifier
+ *
+ * This function is for PF only.
+ *
+ * Return: 0 on success or a negative error code on failure.
+ */
+int xe_sriov_pf_control_wait_flr(struct xe_device *xe, unsigned int vfid)
+{
+	struct xe_gt *gt;
+	unsigned int id;
+	int result = 0;
+	int err;
+
+	for_each_gt(gt, xe, id) {
+		err = xe_gt_sriov_pf_control_wait_flr(gt, vfid);
+		result = result ? -EUCLEAN : err;
+	}
+
+	return result;
+}
+
 /**
  * xe_sriov_pf_control_sync_flr() - Synchronize a VF FLR between all GTs.
  * @xe: the &xe_device
diff --git a/drivers/gpu/drm/xe/xe_sriov_pf_control.h b/drivers/gpu/drm/xe/xe_sriov_pf_control.h
index 30318c1fba34e..ef9f219b21096 100644
--- a/drivers/gpu/drm/xe/xe_sriov_pf_control.h
+++ b/drivers/gpu/drm/xe/xe_sriov_pf_control.h
@@ -12,6 +12,7 @@ int xe_sriov_pf_control_pause_vf(struct xe_device *xe, unsigned int vfid);
 int xe_sriov_pf_control_resume_vf(struct xe_device *xe, unsigned int vfid);
 int xe_sriov_pf_control_stop_vf(struct xe_device *xe, unsigned int vfid);
 int xe_sriov_pf_control_reset_vf(struct xe_device *xe, unsigned int vfid);
+int xe_sriov_pf_control_wait_flr(struct xe_device *xe, unsigned int vfid);
 int xe_sriov_pf_control_sync_flr(struct xe_device *xe, unsigned int vfid);
 int xe_sriov_pf_control_trigger_save_vf(struct xe_device *xe, unsigned int vfid);
 int xe_sriov_pf_control_finish_save_vf(struct xe_device *xe, unsigned int vfid);
-- 
2.50.1

[PATCH v3 24/28] drm/xe/pf: Enable SR-IOV VF migration

[Michał Winiarski]

All of the necessary building blocks are now in place to support SR-IOV
VF migration.
Enable the feature without the need to pass feature enabling debug flags
for those platforms and rely on .has_memirq presence instead (like with
VF resource fixup).

Signed-off-by: Michał Winiarski <michal.winiarski@intel.com>
---
 drivers/gpu/drm/xe/xe_sriov_pf_migration.c | 6 ++++--
 1 file changed, 4 insertions(+), 2 deletions(-)

diff --git a/drivers/gpu/drm/xe/xe_sriov_pf_migration.c b/drivers/gpu/drm/xe/xe_sriov_pf_migration.c
index 6992c227e5a44..c4e9b0ff5b3ae 100644
--- a/drivers/gpu/drm/xe/xe_sriov_pf_migration.c
+++ b/drivers/gpu/drm/xe/xe_sriov_pf_migration.c
@@ -50,8 +50,10 @@ bool xe_sriov_pf_migration_supported(struct xe_device *xe)
 
 static bool pf_check_migration_support(struct xe_device *xe)
 {
-	/* XXX: for now this is for feature enabling only */
-	return IS_ENABLED(CONFIG_DRM_XE_DEBUG);
+	if (xe_device_has_memirq(xe))
+		return true;
+
+	return false;
 }
 
 static void pf_migration_cleanup(void *arg)
-- 
2.50.1

[PATCH v3 25/28] drm/xe/pci: Introduce a helper to allow VF access to PF xe_device

[Michał Winiarski]

In certain scenarios (such as VF migration), VF driver needs to interact
with PF driver.
Add a helper to allow VF driver access to PF xe_device.

Signed-off-by: Michał Winiarski <michal.winiarski@intel.com>
---
 drivers/gpu/drm/xe/xe_pci.c | 17 +++++++++++++++++
 drivers/gpu/drm/xe/xe_pci.h |  3 +++
 2 files changed, 20 insertions(+)

diff --git a/drivers/gpu/drm/xe/xe_pci.c b/drivers/gpu/drm/xe/xe_pci.c
index aeae675c912b7..3e7b03c847a42 100644
--- a/drivers/gpu/drm/xe/xe_pci.c
+++ b/drivers/gpu/drm/xe/xe_pci.c
@@ -1224,6 +1224,23 @@ static struct pci_driver xe_pci_driver = {
 #endif
 };
 
+/**
+ * xe_pci_get_pf_xe_device() - Get PF &xe_device.
+ * @pdev: the VF &pci_dev device
+ *
+ * Return: pointer to PF &xe_device, NULL otherwise.
+ */
+struct xe_device *xe_pci_get_pf_xe_device(struct pci_dev *pdev)
+{
+	struct drm_device *drm;
+
+	drm = pci_iov_get_pf_drvdata(pdev, &xe_pci_driver);
+	if (IS_ERR(drm))
+		return NULL;
+
+	return to_xe_device(drm);
+}
+
 int xe_register_pci_driver(void)
 {
 	return pci_register_driver(&xe_pci_driver);
diff --git a/drivers/gpu/drm/xe/xe_pci.h b/drivers/gpu/drm/xe/xe_pci.h
index 611c1209b14cc..2bb2e486756db 100644
--- a/drivers/gpu/drm/xe/xe_pci.h
+++ b/drivers/gpu/drm/xe/xe_pci.h
@@ -6,6 +6,9 @@
 #ifndef _XE_PCI_H_
 #define _XE_PCI_H_
 
+struct pci_dev;
+
+struct xe_device *xe_pci_get_pf_xe_device(struct pci_dev *pdev);
 int xe_register_pci_driver(void);
 void xe_unregister_pci_driver(void);
 
-- 
2.50.1

[PATCH v3 26/28] drm/xe/pf: Export helpers for VFIO

[Michał Winiarski]

Vendor-specific VFIO driver for Xe will implement VF migration.
Export everything that's needed for migration ops.

Signed-off-by: Michał Winiarski <michal.winiarski@intel.com>
---
 drivers/gpu/drm/xe/Makefile        |   2 +
 drivers/gpu/drm/xe/xe_sriov_vfio.c | 248 +++++++++++++++++++++++++++++
 include/drm/intel/xe_sriov_vfio.h  |  30 ++++
 3 files changed, 280 insertions(+)
 create mode 100644 drivers/gpu/drm/xe/xe_sriov_vfio.c
 create mode 100644 include/drm/intel/xe_sriov_vfio.h

diff --git a/drivers/gpu/drm/xe/Makefile b/drivers/gpu/drm/xe/Makefile
index 3d72db9e528e4..de3778f51ce7e 100644
--- a/drivers/gpu/drm/xe/Makefile
+++ b/drivers/gpu/drm/xe/Makefile
@@ -182,6 +182,8 @@ xe-$(CONFIG_PCI_IOV) += \
 	xe_sriov_pf_service.o \
 	xe_tile_sriov_pf_debugfs.o
 
+xe-$(CONFIG_XE_VFIO_PCI) += xe_sriov_vfio.o
+
 # include helpers for tests even when XE is built-in
 ifdef CONFIG_DRM_XE_KUNIT_TEST
 xe-y += tests/xe_kunit_helpers.o
diff --git a/drivers/gpu/drm/xe/xe_sriov_vfio.c b/drivers/gpu/drm/xe/xe_sriov_vfio.c
new file mode 100644
index 0000000000000..9b02a86132e8b
--- /dev/null
+++ b/drivers/gpu/drm/xe/xe_sriov_vfio.c
@@ -0,0 +1,248 @@
+// SPDX-License-Identifier: MIT
+/*
+ * Copyright © 2025 Intel Corporation
+ */
+
+#include <drm/intel/xe_sriov_vfio.h>
+
+#include "xe_assert.h"
+#include "xe_pci.h"
+#include "xe_pm.h"
+#include "xe_sriov_migration_data.h"
+#include "xe_sriov_pf_control.h"
+#include "xe_sriov_pf_helpers.h"
+#include "xe_sriov_pf_migration.h"
+
+/**
+ * xe_sriov_vfio_get_xe_device() - Get PF &xe_device.
+ * @pdev: the VF &pci_dev device
+ *
+ * Return: pointer to PF &xe_device, NULL otherwise.
+ */
+struct xe_device *xe_sriov_vfio_get_xe_device(struct pci_dev *pdev)
+{
+	return xe_pci_get_pf_xe_device(pdev);
+}
+EXPORT_SYMBOL_FOR_MODULES(xe_sriov_vfio_get_xe_device, "xe-vfio-pci");
+
+/**
+ * xe_sriov_vfio_migration_supported() - Check if migration is supported.
+ * @xe: the PF &xe_device
+ *
+ * Return: true if migration is supported, false otherwise.
+ */
+bool xe_sriov_vfio_migration_supported(struct xe_device *xe)
+{
+	return xe_sriov_pf_migration_supported(xe);
+}
+EXPORT_SYMBOL_FOR_MODULES(xe_sriov_vfio_migration_supported, "xe-vfio-pci");
+
+/**
+ * xe_sriov_vfio_wait_flr_done() - Wait for VF FLR completion.
+ * @xe: the PF &xe_device
+ * @vfid: the VF identifier (can't be 0)
+ *
+ * This function will wait until VF FLR is processed by PF on all tiles (or
+ * until timeout occurs).
+ *
+ * Return: 0 on success or a negative error code on failure.
+ */
+int xe_sriov_vfio_wait_flr_done(struct xe_device *xe, unsigned int vfid)
+{
+	if (vfid == PFID || vfid > xe_sriov_pf_get_totalvfs(xe))
+		return -EINVAL;
+
+	xe_assert(xe, !xe_pm_runtime_suspended(xe));
+
+	return xe_sriov_pf_control_wait_flr(xe, vfid);
+}
+EXPORT_SYMBOL_FOR_MODULES(xe_sriov_vfio_wait_flr_done, "xe-vfio-pci");
+
+/**
+ * xe_sriov_vfio_suspend_device() - Suspend VF.
+ * @xe: the PF &xe_device
+ * @vfid: the VF identifier (can't be 0)
+ *
+ * This function will pause VF on all tiles/GTs.
+ *
+ * Return: 0 on success or a negative error code on failure.
+ */
+int xe_sriov_vfio_suspend_device(struct xe_device *xe, unsigned int vfid)
+{
+	xe_assert(xe, !xe_pm_runtime_suspended(xe));
+
+	if (vfid == PFID || vfid > xe_sriov_pf_get_totalvfs(xe))
+		return -EINVAL;
+
+	return xe_sriov_pf_control_pause_vf(xe, vfid);
+}
+EXPORT_SYMBOL_FOR_MODULES(xe_sriov_vfio_suspend_device, "xe-vfio-pci");
+
+/**
+ * xe_sriov_vfio_resume_device() - Resume VF.
+ * @xe: the PF &xe_device
+ * @vfid: the VF identifier (can't be 0)
+ *
+ * This function will resume VF on all tiles.
+ *
+ * Return: 0 on success or a negative error code on failure.
+ */
+int xe_sriov_vfio_resume_device(struct xe_device *xe, unsigned int vfid)
+{
+	if (vfid == PFID || vfid > xe_sriov_pf_get_totalvfs(xe))
+		return -EINVAL;
+
+	xe_assert(xe, !xe_pm_runtime_suspended(xe));
+
+	return xe_sriov_pf_control_resume_vf(xe, vfid);
+}
+EXPORT_SYMBOL_FOR_MODULES(xe_sriov_vfio_resume_device, "xe-vfio-pci");
+
+/**
+ * xe_sriov_vfio_stop_copy_enter() - Initiate a VF device migration data save.
+ * @xe: the PF &xe_device
+ * @vfid: the VF identifier (can't be 0)
+ *
+ * Return: 0 on success or a negative error code on failure.
+ */
+int xe_sriov_vfio_stop_copy_enter(struct xe_device *xe, unsigned int vfid)
+{
+	xe_assert(xe, !xe_pm_runtime_suspended(xe));
+
+	if (vfid == PFID || vfid > xe_sriov_pf_get_totalvfs(xe))
+		return -EINVAL;
+
+	return xe_sriov_pf_control_trigger_save_vf(xe, vfid);
+}
+EXPORT_SYMBOL_FOR_MODULES(xe_sriov_vfio_stop_copy_enter, "xe-vfio-pci");
+
+/**
+ * xe_sriov_vfio_stop_copy_exit() - Finish a VF device migration data save.
+ * @xe: the PF &xe_device
+ * @vfid: the VF identifier (can't be 0)
+ *
+ * Return: 0 on success or a negative error code on failure.
+ */
+int xe_sriov_vfio_stop_copy_exit(struct xe_device *xe, unsigned int vfid)
+{
+	if (vfid == PFID || vfid > xe_sriov_pf_get_totalvfs(xe))
+		return -EINVAL;
+
+	xe_assert(xe, !xe_pm_runtime_suspended(xe));
+
+	return xe_sriov_pf_control_finish_save_vf(xe, vfid);
+}
+EXPORT_SYMBOL_FOR_MODULES(xe_sriov_vfio_stop_copy_exit, "xe-vfio-pci");
+
+/**
+ * xe_sriov_vfio_resume_data_enter() - Initiate a VF device migration data restore.
+ * @xe: the PF &xe_device
+ * @vfid: the VF identifier (can't be 0)
+ *
+ * Return: 0 on success or a negative error code on failure.
+ */
+int xe_sriov_vfio_resume_data_enter(struct xe_device *xe, unsigned int vfid)
+{
+	xe_assert(xe, !xe_pm_runtime_suspended(xe));
+
+	if (vfid == PFID || vfid > xe_sriov_pf_get_totalvfs(xe))
+		return -EINVAL;
+
+	return xe_sriov_pf_control_trigger_restore_vf(xe, vfid);
+}
+EXPORT_SYMBOL_FOR_MODULES(xe_sriov_vfio_resume_data_enter, "xe-vfio-pci");
+
+/**
+ * xe_sriov_vfio_resume_data_exit() - Finish a VF device migration data restore.
+ * @xe: the PF &xe_device
+ * @vfid: the VF identifier (can't be 0)
+ *
+ * Return: 0 on success or a negative error code on failure.
+ */
+int xe_sriov_vfio_resume_data_exit(struct xe_device *xe, unsigned int vfid)
+{
+	if (vfid == PFID || vfid > xe_sriov_pf_get_totalvfs(xe))
+		return -EINVAL;
+
+	xe_assert(xe, !xe_pm_runtime_suspended(xe));
+
+	return xe_sriov_pf_control_finish_restore_vf(xe, vfid);
+}
+EXPORT_SYMBOL_FOR_MODULES(xe_sriov_vfio_resume_data_exit, "xe-vfio-pci");
+
+/**
+ * xe_sriov_vfio_error() - Move VF device to error state.
+ * @xe: the PF &xe_device
+ * @vfid: the VF identifier (can't be 0)
+ *
+ * Reset is needed to move it out of error state.
+ *
+ * Return: 0 on success or a negative error code on failure.
+ */
+int xe_sriov_vfio_error(struct xe_device *xe, unsigned int vfid)
+{
+	if (vfid == PFID || vfid > xe_sriov_pf_get_totalvfs(xe))
+		return -EINVAL;
+
+	xe_assert(xe, !xe_pm_runtime_suspended(xe));
+
+	return xe_sriov_pf_control_stop_vf(xe, vfid);
+}
+EXPORT_SYMBOL_FOR_MODULES(xe_sriov_vfio_error, "xe-vfio-pci");
+
+/**
+ * xe_sriov_vfio_data_read() - Read migration data from the VF device.
+ * @xe: the PF &xe_device
+ * @vfid: the VF identifier (can't be 0)
+ * @buf: start address of userspace buffer
+ * @len: requested read size from userspace
+ *
+ * Return: number of bytes that has been successfully read,
+ *	   0 if no more migration data is available, -errno on failure.
+ */
+ssize_t xe_sriov_vfio_data_read(struct xe_device *xe, unsigned int vfid,
+				char __user *buf, size_t len)
+{
+	if (vfid == PFID || vfid > xe_sriov_pf_get_totalvfs(xe))
+		return -EINVAL;
+
+	return xe_sriov_migration_data_read(xe, vfid, buf, len);
+}
+EXPORT_SYMBOL_FOR_MODULES(xe_sriov_vfio_data_read, "xe-vfio-pci");
+
+/**
+ * xe_sriov_vfio_data_write() - Write migration data to the VF device.
+ * @xe: the PF &xe_device
+ * @vfid: the VF identifier (can't be 0)
+ * @buf: start address of userspace buffer
+ * @len: requested write size from userspace
+ *
+ * Return: number of bytes that has been successfully written, -errno on failure.
+ */
+ssize_t xe_sriov_vfio_data_write(struct xe_device *xe, unsigned int vfid,
+				 const char __user *buf, size_t len)
+{
+	if (vfid == PFID || vfid > xe_sriov_pf_get_totalvfs(xe))
+		return -EINVAL;
+
+	return xe_sriov_migration_data_write(xe, vfid, buf, len);
+}
+EXPORT_SYMBOL_FOR_MODULES(xe_sriov_vfio_data_write, "xe-vfio-pci");
+
+/**
+ * xe_sriov_vfio_stop_copy_size() - Get a size estimate of VF device migration data.
+ * @xe: the PF &xe_device
+ * @vfid: the VF identifier (can't be 0)
+ *
+ * Return: migration data size in bytes or a negative error code on failure.
+ */
+ssize_t xe_sriov_vfio_stop_copy_size(struct xe_device *xe, unsigned int vfid)
+{
+	if (vfid == PFID || vfid > xe_sriov_pf_get_totalvfs(xe))
+		return -EINVAL;
+
+	xe_assert(xe, !xe_pm_runtime_suspended(xe));
+
+	return xe_sriov_pf_migration_size(xe, vfid);
+}
+EXPORT_SYMBOL_FOR_MODULES(xe_sriov_vfio_stop_copy_size, "xe-vfio-pci");
diff --git a/include/drm/intel/xe_sriov_vfio.h b/include/drm/intel/xe_sriov_vfio.h
new file mode 100644
index 0000000000000..5b02f66465ed9
--- /dev/null
+++ b/include/drm/intel/xe_sriov_vfio.h
@@ -0,0 +1,30 @@
+/* SPDX-License-Identifier: MIT */
+/*
+ * Copyright © 2025 Intel Corporation
+ */
+
+#ifndef _XE_SRIOV_VFIO_H_
+#define _XE_SRIOV_VFIO_H_
+
+#include <linux/types.h>
+
+struct pci_dev;
+struct xe_device;
+
+struct xe_device *xe_sriov_vfio_get_xe_device(struct pci_dev *pdev);
+bool xe_sriov_vfio_migration_supported(struct xe_device *xe);
+int xe_sriov_vfio_wait_flr_done(struct xe_device *xe, unsigned int vfid);
+int xe_sriov_vfio_suspend_device(struct xe_device *xe, unsigned int vfid);
+int xe_sriov_vfio_resume_device(struct xe_device *xe, unsigned int vfid);
+int xe_sriov_vfio_stop_copy_enter(struct xe_device *xe, unsigned int vfid);
+int xe_sriov_vfio_stop_copy_exit(struct xe_device *xe, unsigned int vfid);
+int xe_sriov_vfio_resume_data_enter(struct xe_device *xe, unsigned int vfid);
+int xe_sriov_vfio_resume_data_exit(struct xe_device *xe, unsigned int vfid);
+int xe_sriov_vfio_error(struct xe_device *xe, unsigned int vfid);
+ssize_t xe_sriov_vfio_data_read(struct xe_device *xe, unsigned int vfid,
+				char __user *buf, size_t len);
+ssize_t xe_sriov_vfio_data_write(struct xe_device *xe, unsigned int vfid,
+				 const char __user *buf, size_t len);
+ssize_t xe_sriov_vfio_stop_copy_size(struct xe_device *xe, unsigned int vfid);
+
+#endif
-- 
2.50.1

[PATCH v3 27/28] drm/intel/pciids: Add match with VFIO override

[Michał Winiarski]

In order to allow VFIO users to choose the right driver override, VFIO
driver variant used for VF migration needs to use Intel Graphics PCI
IDs.
Add INTEL_VGA_VFIO_DEVICE match that sets VFIO override_only.

Signed-off-by: Michał Winiarski <michal.winiarski@intel.com>
---
 include/drm/intel/pciids.h | 7 +++++++
 1 file changed, 7 insertions(+)

diff --git a/include/drm/intel/pciids.h b/include/drm/intel/pciids.h
index b258e79b437ac..d14ce43139a28 100644
--- a/include/drm/intel/pciids.h
+++ b/include/drm/intel/pciids.h
@@ -43,6 +43,13 @@
 	.class = PCI_BASE_CLASS_DISPLAY << 16, .class_mask = 0xff << 16, \
 	.driver_data = (kernel_ulong_t)(_info), \
 }
+
+#define INTEL_VGA_VFIO_DEVICE(_id, _info) { \
+	PCI_DEVICE(PCI_VENDOR_ID_INTEL, (_id)), \
+	.class = PCI_BASE_CLASS_DISPLAY << 16, .class_mask = 0xff << 16, \
+	.driver_data = (kernel_ulong_t)(_info), \
+	.override_only = PCI_ID_F_VFIO_DRIVER_OVERRIDE, \
+}
 #endif
 
 #define INTEL_I810_IDS(MACRO__, ...) \
-- 
2.50.1

[PATCH v3 28/28] vfio/xe: Add device specific vfio_pci driver variant for Intel graphics

[Michał Winiarski]

In addition to generic VFIO PCI functionality, the driver implements
VFIO migration uAPI, allowing userspace to enable migration for Intel
Graphics SR-IOV Virtual Functions.
The driver binds to VF device, and uses API exposed by Xe driver bound
to PF device to control VF device state and transfer the migration data.

Signed-off-by: Michał Winiarski <michal.winiarski@intel.com>
---
 MAINTAINERS                  |   7 +
 drivers/vfio/pci/Kconfig     |   2 +
 drivers/vfio/pci/Makefile    |   2 +
 drivers/vfio/pci/xe/Kconfig  |  12 +
 drivers/vfio/pci/xe/Makefile |   3 +
 drivers/vfio/pci/xe/main.c   | 552 +++++++++++++++++++++++++++++++++++
 6 files changed, 578 insertions(+)
 create mode 100644 drivers/vfio/pci/xe/Kconfig
 create mode 100644 drivers/vfio/pci/xe/Makefile
 create mode 100644 drivers/vfio/pci/xe/main.c

diff --git a/MAINTAINERS b/MAINTAINERS
index b890ff265f03f..d73348c5f3f3e 100644
--- a/MAINTAINERS
+++ b/MAINTAINERS
@@ -27008,6 +27008,13 @@ L:	virtualization@lists.linux.dev
 S:	Maintained
 F:	drivers/vfio/pci/virtio
 
+VFIO XE PCI DRIVER
+M:	Michał Winiarski <michal.winiarski@intel.com>
+L:	kvm@vger.kernel.org
+L:	intel-xe@lists.freedesktop.org
+S:	Supported
+F:	drivers/vfio/pci/xe
+
 VGA_SWITCHEROO
 R:	Lukas Wunner <lukas@wunner.de>
 S:	Maintained
diff --git a/drivers/vfio/pci/Kconfig b/drivers/vfio/pci/Kconfig
index 2b0172f546652..c100f0ab87f2d 100644
--- a/drivers/vfio/pci/Kconfig
+++ b/drivers/vfio/pci/Kconfig
@@ -67,4 +67,6 @@ source "drivers/vfio/pci/nvgrace-gpu/Kconfig"
 
 source "drivers/vfio/pci/qat/Kconfig"
 
+source "drivers/vfio/pci/xe/Kconfig"
+
 endmenu
diff --git a/drivers/vfio/pci/Makefile b/drivers/vfio/pci/Makefile
index cf00c0a7e55c8..f5d46aa9347b9 100644
--- a/drivers/vfio/pci/Makefile
+++ b/drivers/vfio/pci/Makefile
@@ -19,3 +19,5 @@ obj-$(CONFIG_VIRTIO_VFIO_PCI) += virtio/
 obj-$(CONFIG_NVGRACE_GPU_VFIO_PCI) += nvgrace-gpu/
 
 obj-$(CONFIG_QAT_VFIO_PCI) += qat/
+
+obj-$(CONFIG_XE_VFIO_PCI) += xe/
diff --git a/drivers/vfio/pci/xe/Kconfig b/drivers/vfio/pci/xe/Kconfig
new file mode 100644
index 0000000000000..787be88268685
--- /dev/null
+++ b/drivers/vfio/pci/xe/Kconfig
@@ -0,0 +1,12 @@
+# SPDX-License-Identifier: GPL-2.0-only
+config XE_VFIO_PCI
+	tristate "VFIO support for Intel Graphics"
+	depends on DRM_XE
+	select VFIO_PCI_CORE
+	help
+	  This option enables vendor-specific VFIO driver for Intel Graphics.
+	  In addition to generic VFIO PCI functionality, it implements VFIO
+	  migration uAPI allowing userspace to enable migration for
+	  Intel Graphics SR-IOV Virtual Functions supported by the Xe driver.
+
+	  If you don't know what to do here, say N.
diff --git a/drivers/vfio/pci/xe/Makefile b/drivers/vfio/pci/xe/Makefile
new file mode 100644
index 0000000000000..13aa0fd192cd4
--- /dev/null
+++ b/drivers/vfio/pci/xe/Makefile
@@ -0,0 +1,3 @@
+# SPDX-License-Identifier: GPL-2.0-only
+obj-$(CONFIG_XE_VFIO_PCI) += xe-vfio-pci.o
+xe-vfio-pci-y := main.o
diff --git a/drivers/vfio/pci/xe/main.c b/drivers/vfio/pci/xe/main.c
new file mode 100644
index 0000000000000..4892c6648c0a7
--- /dev/null
+++ b/drivers/vfio/pci/xe/main.c
@@ -0,0 +1,552 @@
+// SPDX-License-Identifier: GPL-2.0-only
+/*
+ * Copyright © 2025 Intel Corporation
+ */
+
+#include <linux/anon_inodes.h>
+#include <linux/delay.h>
+#include <linux/file.h>
+#include <linux/module.h>
+#include <linux/pci.h>
+#include <linux/sizes.h>
+#include <linux/types.h>
+#include <linux/vfio.h>
+#include <linux/vfio_pci_core.h>
+
+#include <drm/intel/xe_sriov_vfio.h>
+#include <drm/intel/pciids.h>
+
+struct xe_vfio_pci_migration_file {
+	struct file *filp;
+	/* serializes accesses to migration data */
+	struct mutex lock;
+	struct xe_vfio_pci_core_device *xe_vdev;
+};
+
+struct xe_vfio_pci_core_device {
+	struct vfio_pci_core_device core_device;
+	struct xe_device *xe;
+	/* VF number used by PF, Xe HW/FW components use vfid indexing starting from 1 */
+	unsigned int vfid;
+	u8 migrate_cap:1;
+	u8 deferred_reset:1;
+	/* protects migration state */
+	struct mutex state_mutex;
+	enum vfio_device_mig_state mig_state;
+	/* protects the reset_done flow */
+	spinlock_t reset_lock;
+	struct xe_vfio_pci_migration_file *migf;
+};
+
+#define xe_vdev_to_dev(xe_vdev) (&(xe_vdev)->core_device.pdev->dev)
+
+static void xe_vfio_pci_disable_file(struct xe_vfio_pci_migration_file *migf)
+{
+	struct xe_vfio_pci_core_device *xe_vdev = migf->xe_vdev;
+
+	mutex_lock(&migf->lock);
+	xe_vdev->migf = NULL;
+	mutex_unlock(&migf->lock);
+}
+
+static void xe_vfio_pci_reset(struct xe_vfio_pci_core_device *xe_vdev)
+{
+	if (xe_vdev->migf)
+		xe_vfio_pci_disable_file(xe_vdev->migf);
+
+	xe_vdev->mig_state = VFIO_DEVICE_STATE_RUNNING;
+}
+
+static void xe_vfio_pci_state_mutex_lock(struct xe_vfio_pci_core_device *xe_vdev)
+{
+	mutex_lock(&xe_vdev->state_mutex);
+}
+
+/*
+ * This function is called in all state_mutex unlock cases to
+ * handle a 'deferred_reset' if exists.
+ */
+static void xe_vfio_pci_state_mutex_unlock(struct xe_vfio_pci_core_device *xe_vdev)
+{
+again:
+	spin_lock(&xe_vdev->reset_lock);
+	if (xe_vdev->deferred_reset) {
+		xe_vdev->deferred_reset = false;
+		spin_unlock(&xe_vdev->reset_lock);
+		xe_vfio_pci_reset(xe_vdev);
+		goto again;
+	}
+	mutex_unlock(&xe_vdev->state_mutex);
+	spin_unlock(&xe_vdev->reset_lock);
+}
+
+static void xe_vfio_pci_reset_done(struct pci_dev *pdev)
+{
+	struct xe_vfio_pci_core_device *xe_vdev = pci_get_drvdata(pdev);
+	int ret;
+
+	if (!xe_vdev->vfid)
+		return;
+
+	/*
+	 * VF FLR requires additional processing done by PF driver.
+	 * The processing is done after FLR is already finished from PCIe
+	 * perspective.
+	 * In order to avoid a scenario where VF is used while PF processing
+	 * is still in progress, additional synchronization point is needed.
+	 */
+	ret = xe_sriov_vfio_wait_flr_done(xe_vdev->xe, xe_vdev->vfid);
+	if (ret)
+		dev_err(&pdev->dev, "Failed to wait for FLR: %d\n", ret);
+
+	if (!xe_vdev->migrate_cap)
+		return;
+
+	/*
+	 * As the higher VFIO layers are holding locks across reset and using
+	 * those same locks with the mm_lock we need to prevent ABBA deadlock
+	 * with the state_mutex and mm_lock.
+	 * In case the state_mutex was taken already we defer the cleanup work
+	 * to the unlock flow of the other running context.
+	 */
+	spin_lock(&xe_vdev->reset_lock);
+	xe_vdev->deferred_reset = true;
+	if (!mutex_trylock(&xe_vdev->state_mutex)) {
+		spin_unlock(&xe_vdev->reset_lock);
+		return;
+	}
+	spin_unlock(&xe_vdev->reset_lock);
+	xe_vfio_pci_state_mutex_unlock(xe_vdev);
+
+	xe_vfio_pci_reset(xe_vdev);
+}
+
+static const struct pci_error_handlers xe_vfio_pci_err_handlers = {
+	.reset_done = xe_vfio_pci_reset_done,
+	.error_detected = vfio_pci_core_aer_err_detected,
+};
+
+static int xe_vfio_pci_open_device(struct vfio_device *core_vdev)
+{
+	struct xe_vfio_pci_core_device *xe_vdev =
+		container_of(core_vdev, struct xe_vfio_pci_core_device, core_device.vdev);
+	struct vfio_pci_core_device *vdev = &xe_vdev->core_device;
+	int ret;
+
+	ret = vfio_pci_core_enable(vdev);
+	if (ret)
+		return ret;
+
+	vfio_pci_core_finish_enable(vdev);
+
+	return 0;
+}
+
+static int xe_vfio_pci_release_file(struct inode *inode, struct file *filp)
+{
+	struct xe_vfio_pci_migration_file *migf = filp->private_data;
+
+	xe_vfio_pci_disable_file(migf);
+	mutex_destroy(&migf->lock);
+	kfree(migf);
+
+	return 0;
+}
+
+static ssize_t xe_vfio_pci_save_read(struct file *filp, char __user *buf, size_t len, loff_t *pos)
+{
+	struct xe_vfio_pci_migration_file *migf = filp->private_data;
+	ssize_t ret;
+
+	if (pos)
+		return -ESPIPE;
+
+	mutex_lock(&migf->lock);
+	ret = xe_sriov_vfio_data_read(migf->xe_vdev->xe, migf->xe_vdev->vfid, buf, len);
+	mutex_unlock(&migf->lock);
+
+	return ret;
+}
+
+static const struct file_operations xe_vfio_pci_save_fops = {
+	.owner = THIS_MODULE,
+	.read = xe_vfio_pci_save_read,
+	.release = xe_vfio_pci_release_file,
+	.llseek = noop_llseek,
+};
+
+static ssize_t xe_vfio_pci_resume_write(struct file *filp, const char __user *buf,
+					size_t len, loff_t *pos)
+{
+	struct xe_vfio_pci_migration_file *migf = filp->private_data;
+	ssize_t ret;
+
+	if (pos)
+		return -ESPIPE;
+
+	mutex_lock(&migf->lock);
+	ret = xe_sriov_vfio_data_write(migf->xe_vdev->xe, migf->xe_vdev->vfid, buf, len);
+	mutex_unlock(&migf->lock);
+
+	return ret;
+}
+
+static const struct file_operations xe_vfio_pci_resume_fops = {
+	.owner = THIS_MODULE,
+	.write = xe_vfio_pci_resume_write,
+	.release = xe_vfio_pci_release_file,
+	.llseek = noop_llseek,
+};
+
+static const char *vfio_dev_state_str(u32 state)
+{
+	switch (state) {
+	case VFIO_DEVICE_STATE_RUNNING: return "running";
+	case VFIO_DEVICE_STATE_RUNNING_P2P: return "running_p2p";
+	case VFIO_DEVICE_STATE_STOP_COPY: return "stopcopy";
+	case VFIO_DEVICE_STATE_STOP: return "stop";
+	case VFIO_DEVICE_STATE_RESUMING: return "resuming";
+	case VFIO_DEVICE_STATE_ERROR: return "error";
+	default: return "";
+	}
+}
+
+enum xe_vfio_pci_file_type {
+	XE_VFIO_FILE_SAVE = 0,
+	XE_VFIO_FILE_RESUME,
+};
+
+static struct xe_vfio_pci_migration_file *
+xe_vfio_pci_alloc_file(struct xe_vfio_pci_core_device *xe_vdev,
+		       enum xe_vfio_pci_file_type type)
+{
+	struct xe_vfio_pci_migration_file *migf;
+	const struct file_operations *fops;
+	int flags;
+
+	migf = kzalloc(sizeof(*migf), GFP_KERNEL);
+	if (!migf)
+		return ERR_PTR(-ENOMEM);
+
+	fops = type == XE_VFIO_FILE_SAVE ? &xe_vfio_pci_save_fops : &xe_vfio_pci_resume_fops;
+	flags = type == XE_VFIO_FILE_SAVE ? O_RDONLY : O_WRONLY;
+	migf->filp = anon_inode_getfile("xe_vfio_mig", fops, migf, flags);
+	if (IS_ERR(migf->filp)) {
+		kfree(migf);
+		return ERR_CAST(migf->filp);
+	}
+
+	mutex_init(&migf->lock);
+	migf->xe_vdev = xe_vdev;
+	xe_vdev->migf = migf;
+
+	stream_open(migf->filp->f_inode, migf->filp);
+
+	return migf;
+}
+
+static struct file *
+xe_vfio_set_state(struct xe_vfio_pci_core_device *xe_vdev, u32 new)
+{
+	u32 cur = xe_vdev->mig_state;
+	int ret;
+
+	dev_dbg(xe_vdev_to_dev(xe_vdev),
+		"state: %s->%s\n", vfio_dev_state_str(cur), vfio_dev_state_str(new));
+
+	/*
+	 * "STOP" handling is reused for "RUNNING_P2P", as the device doesn't
+	 * have the capability to selectively block outgoing p2p DMA transfers.
+	 * While the device is allowing BAR accesses when the VF is stopped, it
+	 * is not processing any new workload requests, effectively stopping
+	 * any outgoing DMA transfers (not just p2p).
+	 * Any VRAM / MMIO accesses occurring during "RUNNING_P2P" are kept and
+	 * will be migrated to target VF during stop-copy.
+	 */
+	if (cur == VFIO_DEVICE_STATE_RUNNING && new == VFIO_DEVICE_STATE_RUNNING_P2P) {
+		ret = xe_sriov_vfio_suspend_device(xe_vdev->xe, xe_vdev->vfid);
+		if (ret)
+			goto err;
+
+		return NULL;
+	}
+
+	if ((cur == VFIO_DEVICE_STATE_RUNNING_P2P && new == VFIO_DEVICE_STATE_STOP) ||
+	    (cur == VFIO_DEVICE_STATE_STOP && new == VFIO_DEVICE_STATE_RUNNING_P2P))
+		return NULL;
+
+	if (cur == VFIO_DEVICE_STATE_RUNNING_P2P && new == VFIO_DEVICE_STATE_RUNNING) {
+		ret = xe_sriov_vfio_resume_device(xe_vdev->xe, xe_vdev->vfid);
+		if (ret)
+			goto err;
+
+		return NULL;
+	}
+
+	if (cur == VFIO_DEVICE_STATE_STOP && new == VFIO_DEVICE_STATE_STOP_COPY) {
+		struct xe_vfio_pci_migration_file *migf;
+
+		migf = xe_vfio_pci_alloc_file(xe_vdev, XE_VFIO_FILE_SAVE);
+		if (IS_ERR(migf)) {
+			ret = PTR_ERR(migf);
+			goto err;
+		}
+		get_file(migf->filp);
+
+		ret = xe_sriov_vfio_stop_copy_enter(xe_vdev->xe, xe_vdev->vfid);
+		if (ret) {
+			fput(migf->filp);
+			goto err;
+		}
+
+		return migf->filp;
+	}
+
+	if (cur == VFIO_DEVICE_STATE_STOP_COPY && new == VFIO_DEVICE_STATE_STOP) {
+		if (xe_vdev->migf) {
+			fput(xe_vdev->migf->filp);
+			xe_vfio_pci_disable_file(xe_vdev->migf);
+		}
+
+		ret = xe_sriov_vfio_stop_copy_exit(xe_vdev->xe, xe_vdev->vfid);
+		if (ret)
+			goto err;
+
+		return NULL;
+	}
+
+	if (cur == VFIO_DEVICE_STATE_STOP && new == VFIO_DEVICE_STATE_RESUMING) {
+		struct xe_vfio_pci_migration_file *migf;
+
+		migf = xe_vfio_pci_alloc_file(xe_vdev, XE_VFIO_FILE_RESUME);
+		if (IS_ERR(migf)) {
+			ret = PTR_ERR(migf);
+			goto err;
+		}
+		get_file(migf->filp);
+
+		ret = xe_sriov_vfio_resume_data_enter(xe_vdev->xe, xe_vdev->vfid);
+		if (ret) {
+			fput(migf->filp);
+			goto err;
+		}
+
+		return migf->filp;
+	}
+
+	if (cur == VFIO_DEVICE_STATE_RESUMING && new == VFIO_DEVICE_STATE_STOP) {
+		if (xe_vdev->migf) {
+			fput(xe_vdev->migf->filp);
+			xe_vfio_pci_disable_file(xe_vdev->migf);
+		}
+
+		ret = xe_sriov_vfio_resume_data_exit(xe_vdev->xe, xe_vdev->vfid);
+		if (ret)
+			goto err;
+
+		return NULL;
+	}
+
+	WARN(true, "Unknown state transition %d->%d", cur, new);
+	return ERR_PTR(-EINVAL);
+
+err:
+	dev_dbg(xe_vdev_to_dev(xe_vdev),
+		"Failed to transition state: %s->%s err=%d\n",
+		vfio_dev_state_str(cur), vfio_dev_state_str(new), ret);
+	return ERR_PTR(ret);
+}
+
+static struct file *
+xe_vfio_pci_set_device_state(struct vfio_device *core_vdev,
+			     enum vfio_device_mig_state new_state)
+{
+	struct xe_vfio_pci_core_device *xe_vdev =
+		container_of(core_vdev, struct xe_vfio_pci_core_device, core_device.vdev);
+	enum vfio_device_mig_state next_state;
+	struct file *f = NULL;
+	int ret;
+
+	xe_vfio_pci_state_mutex_lock(xe_vdev);
+	while (new_state != xe_vdev->mig_state) {
+		ret = vfio_mig_get_next_state(core_vdev, xe_vdev->mig_state,
+					      new_state, &next_state);
+		if (ret) {
+			xe_sriov_vfio_error(xe_vdev->xe, xe_vdev->vfid);
+			f = ERR_PTR(ret);
+			break;
+		}
+		f = xe_vfio_set_state(xe_vdev, next_state);
+		if (IS_ERR(f))
+			break;
+
+		xe_vdev->mig_state = next_state;
+
+		/* Multiple state transitions with non-NULL file in the middle */
+		if (f && new_state != xe_vdev->mig_state) {
+			fput(f);
+			f = ERR_PTR(-EINVAL);
+			break;
+		}
+	}
+	xe_vfio_pci_state_mutex_unlock(xe_vdev);
+
+	return f;
+}
+
+static int xe_vfio_pci_get_device_state(struct vfio_device *core_vdev,
+					enum vfio_device_mig_state *curr_state)
+{
+	struct xe_vfio_pci_core_device *xe_vdev =
+		container_of(core_vdev, struct xe_vfio_pci_core_device, core_device.vdev);
+
+	xe_vfio_pci_state_mutex_lock(xe_vdev);
+	*curr_state = xe_vdev->mig_state;
+	xe_vfio_pci_state_mutex_unlock(xe_vdev);
+
+	return 0;
+}
+
+static int xe_vfio_pci_get_data_size(struct vfio_device *vdev,
+				     unsigned long *stop_copy_length)
+{
+	struct xe_vfio_pci_core_device *xe_vdev =
+		container_of(vdev, struct xe_vfio_pci_core_device, core_device.vdev);
+
+	xe_vfio_pci_state_mutex_lock(xe_vdev);
+	*stop_copy_length = xe_sriov_vfio_stop_copy_size(xe_vdev->xe, xe_vdev->vfid);
+	xe_vfio_pci_state_mutex_unlock(xe_vdev);
+
+	return 0;
+}
+
+static const struct vfio_migration_ops xe_vfio_pci_migration_ops = {
+	.migration_set_state = xe_vfio_pci_set_device_state,
+	.migration_get_state = xe_vfio_pci_get_device_state,
+	.migration_get_data_size = xe_vfio_pci_get_data_size,
+};
+
+static void xe_vfio_pci_migration_init(struct xe_vfio_pci_core_device *xe_vdev)
+{
+	struct vfio_device *core_vdev = &xe_vdev->core_device.vdev;
+	struct pci_dev *pdev = to_pci_dev(core_vdev->dev);
+	struct xe_device *xe = xe_sriov_vfio_get_xe_device(pdev);
+	int ret;
+
+	if (!xe)
+		return;
+	if (!xe_sriov_vfio_migration_supported(xe))
+		return;
+
+	ret = pci_iov_vf_id(pdev);
+	if (ret < 0)
+		return;
+
+	mutex_init(&xe_vdev->state_mutex);
+	spin_lock_init(&xe_vdev->reset_lock);
+
+	/* Xe HW/FW components use vfid indexing starting from 1 */
+	xe_vdev->vfid = ret + 1;
+	xe_vdev->xe = xe;
+	xe_vdev->migrate_cap = true;
+
+	core_vdev->migration_flags = VFIO_MIGRATION_STOP_COPY | VFIO_MIGRATION_P2P;
+	core_vdev->mig_ops = &xe_vfio_pci_migration_ops;
+}
+
+static void xe_vfio_pci_migration_fini(struct xe_vfio_pci_core_device *xe_vdev)
+{
+	if (!xe_vdev->migrate_cap)
+		return;
+
+	mutex_destroy(&xe_vdev->state_mutex);
+}
+
+static int xe_vfio_pci_init_dev(struct vfio_device *core_vdev)
+{
+	struct xe_vfio_pci_core_device *xe_vdev =
+		container_of(core_vdev, struct xe_vfio_pci_core_device, core_device.vdev);
+
+	xe_vfio_pci_migration_init(xe_vdev);
+
+	return vfio_pci_core_init_dev(core_vdev);
+}
+
+static void xe_vfio_pci_release_dev(struct vfio_device *core_vdev)
+{
+	struct xe_vfio_pci_core_device *xe_vdev =
+		container_of(core_vdev, struct xe_vfio_pci_core_device, core_device.vdev);
+
+	xe_vfio_pci_migration_fini(xe_vdev);
+}
+
+static const struct vfio_device_ops xe_vfio_pci_ops = {
+	.name = "xe-vfio-pci",
+	.init = xe_vfio_pci_init_dev,
+	.release = xe_vfio_pci_release_dev,
+	.open_device = xe_vfio_pci_open_device,
+	.close_device = vfio_pci_core_close_device,
+	.ioctl = vfio_pci_core_ioctl,
+	.device_feature = vfio_pci_core_ioctl_feature,
+	.read = vfio_pci_core_read,
+	.write = vfio_pci_core_write,
+	.mmap = vfio_pci_core_mmap,
+	.request = vfio_pci_core_request,
+	.match = vfio_pci_core_match,
+	.match_token_uuid = vfio_pci_core_match_token_uuid,
+	.bind_iommufd = vfio_iommufd_physical_bind,
+	.unbind_iommufd = vfio_iommufd_physical_unbind,
+	.attach_ioas = vfio_iommufd_physical_attach_ioas,
+	.detach_ioas = vfio_iommufd_physical_detach_ioas,
+};
+
+static int xe_vfio_pci_probe(struct pci_dev *pdev, const struct pci_device_id *id)
+{
+	struct xe_vfio_pci_core_device *xe_vdev;
+	int ret;
+
+	xe_vdev = vfio_alloc_device(xe_vfio_pci_core_device, core_device.vdev, &pdev->dev,
+				    &xe_vfio_pci_ops);
+	if (IS_ERR(xe_vdev))
+		return PTR_ERR(xe_vdev);
+
+	dev_set_drvdata(&pdev->dev, &xe_vdev->core_device);
+
+	ret = vfio_pci_core_register_device(&xe_vdev->core_device);
+	if (ret) {
+		vfio_put_device(&xe_vdev->core_device.vdev);
+		return ret;
+	}
+
+	return 0;
+}
+
+static void xe_vfio_pci_remove(struct pci_dev *pdev)
+{
+	struct xe_vfio_pci_core_device *xe_vdev = pci_get_drvdata(pdev);
+
+	vfio_pci_core_unregister_device(&xe_vdev->core_device);
+	vfio_put_device(&xe_vdev->core_device.vdev);
+}
+
+static const struct pci_device_id xe_vfio_pci_table[] = {
+	INTEL_PTL_IDS(INTEL_VGA_VFIO_DEVICE, NULL),
+	INTEL_WCL_IDS(INTEL_VGA_VFIO_DEVICE, NULL),
+	INTEL_BMG_IDS(INTEL_VGA_VFIO_DEVICE, NULL),
+	{}
+};
+MODULE_DEVICE_TABLE(pci, xe_vfio_pci_table);
+
+static struct pci_driver xe_vfio_pci_driver = {
+	.name = "xe-vfio-pci",
+	.id_table = xe_vfio_pci_table,
+	.probe = xe_vfio_pci_probe,
+	.remove = xe_vfio_pci_remove,
+	.err_handler = &xe_vfio_pci_err_handlers,
+	.driver_managed_dma = true,
+};
+module_pci_driver(xe_vfio_pci_driver);
+
+MODULE_LICENSE("GPL");
+MODULE_AUTHOR("Michał Winiarski <michal.winiarski@intel.com>");
+MODULE_DESCRIPTION("VFIO PCI driver with migration support for Intel Graphics");
-- 
2.50.1

Re: [PATCH v3 03/28] drm/xe/pf: Convert control state to bitmap

[Michal Wajdeczko]

On 10/30/2025 9:31 PM, Michał Winiarski wrote:
> In upcoming changes, the number of states will increase as a result of
> introducing SAVE and RESTORE states.
> This means that using unsigned long as underlying storage won't work on
> 32-bit architectures, as we'll run out of bits.
> Use bitmap instead.
> 
> Reported-by: kernel test robot <lkp@intel.com>
> Closes: https://lore.kernel.org/oe-kbuild-all/202510231918.XlOqymLC-lkp@intel.com/
> Signed-off-by: Michał Winiarski <michal.winiarski@intel.com>
> ---
>  drivers/gpu/drm/xe/xe_gt_sriov_pf_control.c       | 2 +-
>  drivers/gpu/drm/xe/xe_gt_sriov_pf_control_types.h | 5 +++--
>  2 files changed, 4 insertions(+), 3 deletions(-)
> 
> diff --git a/drivers/gpu/drm/xe/xe_gt_sriov_pf_control.c b/drivers/gpu/drm/xe/xe_gt_sriov_pf_control.c
> index 9de05db1f0905..8a2577fda4198 100644
> --- a/drivers/gpu/drm/xe/xe_gt_sriov_pf_control.c
> +++ b/drivers/gpu/drm/xe/xe_gt_sriov_pf_control.c
> @@ -225,7 +225,7 @@ static unsigned long *pf_peek_vf_state(struct xe_gt *gt, unsigned int vfid)
>  {
>  	struct xe_gt_sriov_control_state *cs = pf_pick_vf_control(gt, vfid);
>  
> -	return &cs->state;
> +	return cs->state;
>  }
>  
>  static bool pf_check_vf_state(struct xe_gt *gt, unsigned int vfid,
> diff --git a/drivers/gpu/drm/xe/xe_gt_sriov_pf_control_types.h b/drivers/gpu/drm/xe/xe_gt_sriov_pf_control_types.h
> index c80b7e77f1ad2..3ba6ad886c939 100644
> --- a/drivers/gpu/drm/xe/xe_gt_sriov_pf_control_types.h
> +++ b/drivers/gpu/drm/xe/xe_gt_sriov_pf_control_types.h
> @@ -73,7 +73,8 @@ enum xe_gt_sriov_control_bits {
>  	XE_GT_SRIOV_STATE_STOP_FAILED,
>  	XE_GT_SRIOV_STATE_STOPPED,
>  
> -	XE_GT_SRIOV_STATE_MISMATCH = BITS_PER_LONG - 1,
> +	XE_GT_SRIOV_STATE_MISMATCH,
> +	XE_GT_SRIOV_STATE_MAX,

while this feels handy, this MAX enumerator is not a real state
and as such shouldn't be passed to any function that expects
"enum"

since we know (and want) to keep MISMATCH state as last one
(aka top bit) then maybe tag it and use separate define:

-	XE_GT_SRIOV_STATE_MISMATCH = BITS_PER_LONG - 1,
+	XE_GT_SRIOV_STATE_MISMATCH /* always keep as last */
+
+ #define XE_GT_SRIOV_NUM_STATES (XE_GT_SRIOV_STATE_MISMATCH + 1)

>  };
>  
>  /**
> @@ -83,7 +84,7 @@ enum xe_gt_sriov_control_bits {
>   */
>  struct xe_gt_sriov_control_state {
>  	/** @state: VF state bits */
> -	unsigned long state;
> +	DECLARE_BITMAP(state, XE_GT_SRIOV_STATE_MAX);
>  
>  	/** @done: completion of async operations */
>  	struct completion done;

Re: [PATCH v3 03/28] drm/xe/pf: Convert control state to bitmap

[Michał Winiarski]

On Thu, Oct 30, 2025 at 11:57:28PM +0100, Michal Wajdeczko wrote:
> 
> 
> On 10/30/2025 9:31 PM, Michał Winiarski wrote:
> > In upcoming changes, the number of states will increase as a result of
> > introducing SAVE and RESTORE states.
> > This means that using unsigned long as underlying storage won't work on
> > 32-bit architectures, as we'll run out of bits.
> > Use bitmap instead.
> > 
> > Reported-by: kernel test robot <lkp@intel.com>
> > Closes: https://lore.kernel.org/oe-kbuild-all/202510231918.XlOqymLC-lkp@intel.com/
> > Signed-off-by: Michał Winiarski <michal.winiarski@intel.com>
> > ---
> >  drivers/gpu/drm/xe/xe_gt_sriov_pf_control.c       | 2 +-
> >  drivers/gpu/drm/xe/xe_gt_sriov_pf_control_types.h | 5 +++--
> >  2 files changed, 4 insertions(+), 3 deletions(-)
> > 
> > diff --git a/drivers/gpu/drm/xe/xe_gt_sriov_pf_control.c b/drivers/gpu/drm/xe/xe_gt_sriov_pf_control.c
> > index 9de05db1f0905..8a2577fda4198 100644
> > --- a/drivers/gpu/drm/xe/xe_gt_sriov_pf_control.c
> > +++ b/drivers/gpu/drm/xe/xe_gt_sriov_pf_control.c
> > @@ -225,7 +225,7 @@ static unsigned long *pf_peek_vf_state(struct xe_gt *gt, unsigned int vfid)
> >  {
> >  	struct xe_gt_sriov_control_state *cs = pf_pick_vf_control(gt, vfid);
> >  
> > -	return &cs->state;
> > +	return cs->state;
> >  }
> >  
> >  static bool pf_check_vf_state(struct xe_gt *gt, unsigned int vfid,
> > diff --git a/drivers/gpu/drm/xe/xe_gt_sriov_pf_control_types.h b/drivers/gpu/drm/xe/xe_gt_sriov_pf_control_types.h
> > index c80b7e77f1ad2..3ba6ad886c939 100644
> > --- a/drivers/gpu/drm/xe/xe_gt_sriov_pf_control_types.h
> > +++ b/drivers/gpu/drm/xe/xe_gt_sriov_pf_control_types.h
> > @@ -73,7 +73,8 @@ enum xe_gt_sriov_control_bits {
> >  	XE_GT_SRIOV_STATE_STOP_FAILED,
> >  	XE_GT_SRIOV_STATE_STOPPED,
> >  
> > -	XE_GT_SRIOV_STATE_MISMATCH = BITS_PER_LONG - 1,
> > +	XE_GT_SRIOV_STATE_MISMATCH,
> > +	XE_GT_SRIOV_STATE_MAX,
> 
> while this feels handy, this MAX enumerator is not a real state
> and as such shouldn't be passed to any function that expects
> "enum"
> 
> since we know (and want) to keep MISMATCH state as last one
> (aka top bit) then maybe tag it and use separate define:
> 
> -	XE_GT_SRIOV_STATE_MISMATCH = BITS_PER_LONG - 1,
> +	XE_GT_SRIOV_STATE_MISMATCH /* always keep as last */
> +
> + #define XE_GT_SRIOV_NUM_STATES (XE_GT_SRIOV_STATE_MISMATCH + 1)

Sure, sounds good.

Thanks,
-Michał

> 
> >  };
> >  
> >  /**
> > @@ -83,7 +84,7 @@ enum xe_gt_sriov_control_bits {
> >   */
> >  struct xe_gt_sriov_control_state {
> >  	/** @state: VF state bits */
> > -	unsigned long state;
> > +	DECLARE_BITMAP(state, XE_GT_SRIOV_STATE_MAX);
> >  
> >  	/** @done: completion of async operations */
> >  	struct completion done;
> 

Re: [PATCH v3 05/28] drm/xe/pf: Add data structures and handlers for migration rings

[Michal Wajdeczko]

On 10/30/2025 9:31 PM, Michał Winiarski wrote:
> Migration data is queued in a per-GT ptr_ring to decouple the worker
> responsible for handling the data transfer from the .read() and .write()
> syscalls.
> Add the data structures and handlers that will be used in future
> commits.
> 
> Signed-off-by: Michał Winiarski <michal.winiarski@intel.com>
> ---
>  drivers/gpu/drm/xe/xe_gt_sriov_pf_control.c   | 271 +++++++++++++++++-
>  drivers/gpu/drm/xe/xe_gt_sriov_pf_control.h   |   6 +-
>  .../gpu/drm/xe/xe_gt_sriov_pf_control_types.h |  12 +
>  drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.c | 183 ++++++++++++
>  drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.h |  14 +
>  .../drm/xe/xe_gt_sriov_pf_migration_types.h   |  11 +
>  drivers/gpu/drm/xe/xe_gt_sriov_pf_types.h     |   3 +
>  drivers/gpu/drm/xe/xe_sriov_pf_migration.c    | 142 +++++++++
>  drivers/gpu/drm/xe/xe_sriov_pf_migration.h    |   7 +
>  .../gpu/drm/xe/xe_sriov_pf_migration_types.h  |  58 ++++
>  drivers/gpu/drm/xe/xe_sriov_pf_types.h        |   3 +
>  11 files changed, 695 insertions(+), 15 deletions(-)
> 
> diff --git a/drivers/gpu/drm/xe/xe_gt_sriov_pf_control.c b/drivers/gpu/drm/xe/xe_gt_sriov_pf_control.c
> index a571e1c02a3b4..ad9dcd3f56453 100644
> --- a/drivers/gpu/drm/xe/xe_gt_sriov_pf_control.c
> +++ b/drivers/gpu/drm/xe/xe_gt_sriov_pf_control.c
> @@ -19,6 +19,7 @@
>  #include "xe_guc_ct.h"
>  #include "xe_sriov.h"
>  #include "xe_sriov_pf_control.h"
> +#include "xe_sriov_pf_migration.h"
>  #include "xe_sriov_pf_service.h"
>  #include "xe_tile.h"
>  
> @@ -185,9 +186,15 @@ static const char *control_bit_to_string(enum xe_gt_sriov_control_bits bit)
>  	CASE2STR(PAUSE_FAILED);
>  	CASE2STR(PAUSED);
>  	CASE2STR(SAVE_WIP);
> +	CASE2STR(SAVE_PROCESS_DATA);
> +	CASE2STR(SAVE_WAIT_DATA);
> +	CASE2STR(SAVE_DATA_DONE);
>  	CASE2STR(SAVE_FAILED);
>  	CASE2STR(SAVED);
>  	CASE2STR(RESTORE_WIP);
> +	CASE2STR(RESTORE_PROCESS_DATA);
> +	CASE2STR(RESTORE_WAIT_DATA);
> +	CASE2STR(RESTORE_DATA_DONE);
>  	CASE2STR(RESTORE_FAILED);
>  	CASE2STR(RESTORED);
>  	CASE2STR(RESUME_WIP);
> @@ -804,9 +811,50 @@ int xe_gt_sriov_pf_control_resume_vf(struct xe_gt *gt, unsigned int vfid)
>  	return -ECANCELED;
>  }
>  
> +/**
> + * DOC: The VF SAVE state machine
> + *
> + * SAVE extends the PAUSED state.
> + *
> + * The VF SAVE state machine looks like::
> + *
> + *  ....PAUSED....................................................
> + *  :                                                            :
> + *  :     (O)<---------o                                         :
> + *  :      |            \                                        :
> + *  :    save          (SAVED)    (SAVE_FAILED)                  :
> + *  :      |               ^           ^                         :
> + *  :      |               |           |                         :
> + *  :  ....V...............o...........o......SAVE_WIP.........  :
> + *  :  :   |               |           |                      :  :
> + *  :  :   |             empty         |                      :  :
> + *  :  :   |               |           |                      :  :
> + *  :  :   |               |           |                      :  :
> + *  :  :   |           DATA_DONE       |                      :  :
> + *  :  :   |               ^           |                      :  :
> + *  :  :   |               |        error                     :  :
> + *  :  :   |            no_data       /                       :  :
> + *  :  :   |              /          /                        :  :
> + *  :  :   |             /          /                         :  :
> + *  :  :   |            /          /                          :  :
> + *  :  :   o---------->PROCESS_DATA<----consume               :  :
> + *  :  :                \                      \              :  :
> + *  :  :                 \                      \             :  :
> + *  :  :                  \                      \            :  :
> + *  :  :                   ring_full----->WAIT_DATA           :  :
> + *  :  :                                                      :  :
> + *  :  :......................................................:  :
> + *  :............................................................:
> + *
> + * For the full state machine view, see `The VF state machine`_.
> + */

nit: since above doc is not strictly for the function below,
add separation line here


>  static void pf_exit_vf_save_wip(struct xe_gt *gt, unsigned int vfid)
>  {
> -	pf_exit_vf_state(gt, vfid, XE_GT_SRIOV_STATE_SAVE_WIP);
> +	if (pf_exit_vf_state(gt, vfid, XE_GT_SRIOV_STATE_SAVE_WIP)) {
> +		pf_escape_vf_state(gt, vfid, XE_GT_SRIOV_STATE_SAVE_PROCESS_DATA);
> +		pf_escape_vf_state(gt, vfid, XE_GT_SRIOV_STATE_SAVE_WAIT_DATA);
> +		pf_escape_vf_state(gt, vfid, XE_GT_SRIOV_STATE_SAVE_DATA_DONE);
> +	}
>  }
>  
>  static void pf_enter_vf_saved(struct xe_gt *gt, unsigned int vfid)
> @@ -821,19 +869,54 @@ static void pf_enter_vf_saved(struct xe_gt *gt, unsigned int vfid)
>  	pf_exit_vf_wip(gt, vfid);
>  }
>  
> +static void pf_enter_vf_save_failed(struct xe_gt *gt, unsigned int vfid)
> +{
> +	pf_enter_vf_state(gt, vfid, XE_GT_SRIOV_STATE_SAVE_FAILED);

nit: is it ok to do not test if we are already in the FAILED state ?

> +	pf_exit_vf_wip(gt, vfid);
> +}
> +
> +static int pf_handle_vf_save_data(struct xe_gt *gt, unsigned int vfid)
> +{
> +	return 0;
> +}
> +
>  static bool pf_handle_vf_save(struct xe_gt *gt, unsigned int vfid)
>  {
> -	if (!pf_exit_vf_state(gt, vfid, XE_GT_SRIOV_STATE_SAVE_WIP))
> +	int ret;
> +
> +	if (!pf_exit_vf_state(gt, vfid, XE_GT_SRIOV_STATE_SAVE_PROCESS_DATA))
>  		return false;
>  
> -	pf_enter_vf_saved(gt, vfid);
> +	if (xe_gt_sriov_pf_migration_ring_full(gt, vfid)) {
> +		pf_enter_vf_state(gt, vfid, XE_GT_SRIOV_STATE_SAVE_WAIT_DATA);
> +

nit: can remove this sep line

> +		return true;
> +	}
> +
> +	ret = pf_handle_vf_save_data(gt, vfid);
> +	if (ret == -EAGAIN)
> +		pf_enter_vf_state(gt, vfid, XE_GT_SRIOV_STATE_SAVE_PROCESS_DATA);
> +	else if (ret)
> +		pf_enter_vf_save_failed(gt, vfid);
> +	else
> +		pf_enter_vf_state(gt, vfid, XE_GT_SRIOV_STATE_SAVE_DATA_DONE);
>  
>  	return true;
>  }
>  
> +static void pf_exit_vf_save_wait_data(struct xe_gt *gt, unsigned int vfid)
> +{
> +	if (!pf_exit_vf_state(gt, vfid, XE_GT_SRIOV_STATE_SAVE_WAIT_DATA))
> +		return;
> +
> +	pf_enter_vf_state(gt, vfid, XE_GT_SRIOV_STATE_SAVE_PROCESS_DATA);
> +	pf_queue_vf(gt, vfid);
> +}
> +
>  static bool pf_enter_vf_save_wip(struct xe_gt *gt, unsigned int vfid)
>  {
>  	if (pf_enter_vf_state(gt, vfid, XE_GT_SRIOV_STATE_SAVE_WIP)) {
> +		pf_enter_vf_state(gt, vfid, XE_GT_SRIOV_STATE_SAVE_PROCESS_DATA);
>  		pf_enter_vf_wip(gt, vfid);
>  		pf_queue_vf(gt, vfid);
>  		return true;
> @@ -842,6 +925,32 @@ static bool pf_enter_vf_save_wip(struct xe_gt *gt, unsigned int vfid)
>  	return false;
>  }
>  
> +/**
> + * xe_gt_sriov_pf_control_check_save_data_done() - Check if all save migration data was produced.
> + * @gt: the &xe_gt
> + * @vfid: the VF identifier
> + *
> + * This function is for PF only.
> + *
> + * Return: 0 on success or a negative error code on failure.

but it's defined as a bool below

> + */
> +bool xe_gt_sriov_pf_control_check_save_data_done(struct xe_gt *gt, unsigned int vfid)
> +{
> +	return pf_check_vf_state(gt, vfid, XE_GT_SRIOV_STATE_SAVE_DATA_DONE);
> +}
> +
> +/**
> + * xe_gt_sriov_pf_control_process_save_data() - Queue VF save migration data processing.
> + * @gt: the &xe_gt
> + * @vfid: the VF identifier
> + *
> + * This function is for PF only.
> + */
> +void xe_gt_sriov_pf_control_process_save_data(struct xe_gt *gt, unsigned int vfid)
> +{
> +	pf_exit_vf_save_wait_data(gt, vfid);

ok to have no check against unexpected state ?

> +}
> +
>  /**
>   * xe_gt_sriov_pf_control_trigger_save_vf() - Start an SR-IOV VF migration data save sequence.
>   * @gt: the &xe_gt
> @@ -887,19 +996,62 @@ int xe_gt_sriov_pf_control_trigger_save_vf(struct xe_gt *gt, unsigned int vfid)
>   */
>  int xe_gt_sriov_pf_control_finish_save_vf(struct xe_gt *gt, unsigned int vfid)
>  {
> -	if (!pf_expect_vf_state(gt, vfid, XE_GT_SRIOV_STATE_SAVED)) {
> -		pf_enter_vf_mismatch(gt, vfid);
> +	if (!pf_check_vf_state(gt, vfid, XE_GT_SRIOV_STATE_SAVE_DATA_DONE)) {
> +		xe_gt_sriov_err(gt, "VF%u save is still in progress!\n", vfid);
>  		return -EIO;
>  	}
>  
>  	pf_expect_vf_state(gt, vfid, XE_GT_SRIOV_STATE_PAUSED);
> +	pf_exit_vf_state(gt, vfid, XE_GT_SRIOV_STATE_SAVE_DATA_DONE);
> +	pf_enter_vf_saved(gt, vfid);
>  
>  	return 0;
>  }
>  
> +/**
> + * DOC: The VF RESTORE state machine
> + *
> + * RESTORE extends the PAUSED state.
> + *
> + * The VF RESTORE state machine looks like::
> + *
> + *  ....PAUSED....................................................
> + *  :                                                            :
> + *  :     (O)<---------o                                         :
> + *  :      |            \                                        :
> + *  :    restore      (RESTORED)  (RESTORE_FAILED)               :
> + *  :      |               ^           ^                         :
> + *  :      |               |           |                         :
> + *  :  ....V...............o...........o......RESTORE_WIP......  :
> + *  :  :   |               |           |                      :  :
> + *  :  :   |             empty         |                      :  :
> + *  :  :   |               |           |                      :  :
> + *  :  :   |               |           |                      :  :
> + *  :  :   |           DATA_DONE       |                      :  :
> + *  :  :   |               ^           |                      :  :
> + *  :  :   |               |        error                     :  :
> + *  :  :   |           trailer        /                       :  :
> + *  :  :   |              /          /                        :  :
> + *  :  :   |             /          /                         :  :
> + *  :  :   |            /          /                          :  :
> + *  :  :   o---------->PROCESS_DATA<----produce               :  :
> + *  :  :                \                      \              :  :
> + *  :  :                 \                      \             :  :
> + *  :  :                  \                      \            :  :
> + *  :  :                   ring_empty---->WAIT_DATA           :  :
> + *  :  :                                                      :  :
> + *  :  :......................................................:  :
> + *  :............................................................:
> + *
> + * For the full state machine view, see `The VF state machine`_.
> + */

nit: like above, add separation line here

>  static void pf_exit_vf_restore_wip(struct xe_gt *gt, unsigned int vfid)
>  {
> -	pf_exit_vf_state(gt, vfid, XE_GT_SRIOV_STATE_RESTORE_WIP);
> +	if (pf_exit_vf_state(gt, vfid, XE_GT_SRIOV_STATE_RESTORE_WIP)) {
> +		pf_escape_vf_state(gt, vfid, XE_GT_SRIOV_STATE_RESTORE_PROCESS_DATA);
> +		pf_escape_vf_state(gt, vfid, XE_GT_SRIOV_STATE_RESTORE_WAIT_DATA);
> +		pf_escape_vf_state(gt, vfid, XE_GT_SRIOV_STATE_RESTORE_DATA_DONE);
> +	}
>  }
>  
>  static void pf_enter_vf_restored(struct xe_gt *gt, unsigned int vfid)
> @@ -914,19 +1066,61 @@ static void pf_enter_vf_restored(struct xe_gt *gt, unsigned int vfid)
>  	pf_exit_vf_wip(gt, vfid);
>  }
>  
> +static void pf_enter_vf_restore_failed(struct xe_gt *gt, unsigned int vfid)
> +{
> +	pf_enter_vf_state(gt, vfid, XE_GT_SRIOV_STATE_RESTORE_FAILED);
> +	pf_exit_vf_wip(gt, vfid);
> +}
> +
> +static int pf_handle_vf_restore_data(struct xe_gt *gt, unsigned int vfid)
> +{
> +	struct xe_sriov_migration_data *data = xe_gt_sriov_pf_migration_restore_consume(gt, vfid);
> +
> +	xe_gt_assert(gt, data);

maybe assert is redundant here, as if data is null, we will get NPD crash call
stack from the below line anyway?

> +
> +	xe_gt_sriov_notice(gt, "Skipping VF%u unknown data type: %d\n", vfid, data->type);
> +
> +	return 0;
> +}
> +
>  static bool pf_handle_vf_restore(struct xe_gt *gt, unsigned int vfid)
>  {
> -	if (!pf_exit_vf_state(gt, vfid, XE_GT_SRIOV_STATE_RESTORE_WIP))
> +	int ret;
> +
> +	if (!pf_exit_vf_state(gt, vfid, XE_GT_SRIOV_STATE_RESTORE_PROCESS_DATA))
>  		return false;
>  
> -	pf_enter_vf_restored(gt, vfid);
> +	if (xe_gt_sriov_pf_migration_ring_empty(gt, vfid)) {
> +		if (pf_exit_vf_state(gt, vfid, XE_GT_SRIOV_STATE_RESTORE_DATA_DONE))
> +			pf_enter_vf_restored(gt, vfid);
> +		else
> +			pf_enter_vf_state(gt, vfid, XE_GT_SRIOV_STATE_RESTORE_WAIT_DATA);
> +
> +		return true;
> +	}
> +
> +	ret = pf_handle_vf_restore_data(gt, vfid);
> +	if (ret)
> +		pf_enter_vf_restore_failed(gt, vfid);
> +	else
> +		pf_enter_vf_state(gt, vfid, XE_GT_SRIOV_STATE_RESTORE_PROCESS_DATA);
>  
>  	return true;
>  }
>  
> +static void pf_exit_vf_restore_wait_data(struct xe_gt *gt, unsigned int vfid)
> +{
> +	if (!pf_exit_vf_state(gt, vfid, XE_GT_SRIOV_STATE_RESTORE_WAIT_DATA))
> +		return;
> +
> +	pf_enter_vf_state(gt, vfid, XE_GT_SRIOV_STATE_RESTORE_PROCESS_DATA);
> +	pf_queue_vf(gt, vfid);
> +}
> +
>  static bool pf_enter_vf_restore_wip(struct xe_gt *gt, unsigned int vfid)
>  {
>  	if (pf_enter_vf_state(gt, vfid, XE_GT_SRIOV_STATE_RESTORE_WIP)) {
> +		pf_enter_vf_state(gt, vfid, XE_GT_SRIOV_STATE_RESTORE_PROCESS_DATA);
>  		pf_enter_vf_wip(gt, vfid);
>  		pf_queue_vf(gt, vfid);
>  		return true;
> @@ -935,6 +1129,47 @@ static bool pf_enter_vf_restore_wip(struct xe_gt *gt, unsigned int vfid)
>  	return false;
>  }
>  
> +/**
> + * xe_gt_sriov_pf_control_restore_data_done() - Indicate the end of VF migration data stream.
> + * @gt: the &xe_gt
> + * @vfid: the VF identifier
> + *
> + * This function is for PF only.
> + *
> + * Return: 0 on success or a negative error code on failure.

save_data_done() is bool
and below you return only 0 or -EIO
maybe it could be bool too?

> + */
> +int xe_gt_sriov_pf_control_restore_data_done(struct xe_gt *gt, unsigned int vfid)
> +{
> +	if (!pf_expect_vf_state(gt, vfid, XE_GT_SRIOV_STATE_RESTORE_WIP)) {
> +		pf_enter_vf_state_machine_bug(gt, vfid);
> +		return -EIO;
> +	}
> +
> +	if (!pf_enter_vf_state(gt, vfid, XE_GT_SRIOV_STATE_RESTORE_DATA_DONE)) {
> +		pf_enter_vf_state_machine_bug(gt, vfid);
> +		return -EIO;
> +	}
> +
> +	xe_gt_sriov_pf_control_process_restore_data(gt, vfid);
> +
> +	return 0;
> +}
> +
> +/**
> + * xe_gt_sriov_pf_control_process_restore_data() - Queue VF restore migration data processing.
> + * @gt: the &xe_gt
> + * @vfid: the VF identifier
> + *
> + * This function is for PF only.
> + */
> +void xe_gt_sriov_pf_control_process_restore_data(struct xe_gt *gt, unsigned int vfid)
> +{
> +	if (!pf_expect_vf_state(gt, vfid, XE_GT_SRIOV_STATE_RESTORE_WIP))
> +		pf_enter_vf_state_machine_bug(gt, vfid);
> +
> +	pf_exit_vf_restore_wait_data(gt, vfid);
> +}
> +
>  /**
>   * xe_gt_sriov_pf_control_trigger restore_vf() - Start an SR-IOV VF migration data restore sequence.
>   * @gt: the &xe_gt
> @@ -1000,11 +1235,9 @@ int xe_gt_sriov_pf_control_finish_restore_vf(struct xe_gt *gt, unsigned int vfid
>  {
>  	int ret;
>  
> -	if (pf_check_vf_state(gt, vfid, XE_GT_SRIOV_STATE_RESTORE_WIP)) {
> -		ret = pf_wait_vf_restore_done(gt, vfid);
> -		if (ret)
> -			return ret;
> -	}
> +	ret = pf_wait_vf_restore_done(gt, vfid);
> +	if (ret)
> +		return ret;
>  
>  	if (!pf_expect_vf_state(gt, vfid, XE_GT_SRIOV_STATE_RESTORED)) {
>  		pf_enter_vf_mismatch(gt, vfid);
> @@ -1705,9 +1938,21 @@ static bool pf_process_vf_state_machine(struct xe_gt *gt, unsigned int vfid)
>  	if (pf_exit_vf_pause_save_guc(gt, vfid))
>  		return true;
>  
> +	if (pf_check_vf_state(gt, vfid, XE_GT_SRIOV_STATE_SAVE_WAIT_DATA)) {
> +		xe_gt_sriov_dbg_verbose(gt, "VF%u in %s\n", vfid,
> +					control_bit_to_string(XE_GT_SRIOV_STATE_SAVE_WAIT_DATA));
> +		return false;
> +	}
> +
>  	if (pf_handle_vf_save(gt, vfid))
>  		return true;
>  
> +	if (pf_check_vf_state(gt, vfid, XE_GT_SRIOV_STATE_RESTORE_WAIT_DATA)) {
> +		xe_gt_sriov_dbg_verbose(gt, "VF%u in %s\n", vfid,
> +					control_bit_to_string(XE_GT_SRIOV_STATE_RESTORE_WAIT_DATA));
> +		return false;
> +	}
> +
>  	if (pf_handle_vf_restore(gt, vfid))
>  		return true;
>  
> diff --git a/drivers/gpu/drm/xe/xe_gt_sriov_pf_control.h b/drivers/gpu/drm/xe/xe_gt_sriov_pf_control.h
> index abc233f6302ed..6b1ab339e3b73 100644
> --- a/drivers/gpu/drm/xe/xe_gt_sriov_pf_control.h
> +++ b/drivers/gpu/drm/xe/xe_gt_sriov_pf_control.h
> @@ -14,12 +14,14 @@ struct xe_gt;
>  int xe_gt_sriov_pf_control_init(struct xe_gt *gt);
>  void xe_gt_sriov_pf_control_restart(struct xe_gt *gt);
>  
> -bool xe_gt_sriov_pf_control_check_vf_data_wip(struct xe_gt *gt, unsigned int vfid);
> -
>  int xe_gt_sriov_pf_control_pause_vf(struct xe_gt *gt, unsigned int vfid);
>  int xe_gt_sriov_pf_control_resume_vf(struct xe_gt *gt, unsigned int vfid);
> +bool xe_gt_sriov_pf_control_check_save_data_done(struct xe_gt *gt, unsigned int vfid);
> +void xe_gt_sriov_pf_control_process_save_data(struct xe_gt *gt, unsigned int vfid);
>  int xe_gt_sriov_pf_control_trigger_save_vf(struct xe_gt *gt, unsigned int vfid);
>  int xe_gt_sriov_pf_control_finish_save_vf(struct xe_gt *gt, unsigned int vfid);
> +int xe_gt_sriov_pf_control_restore_data_done(struct xe_gt *gt, unsigned int vfid);
> +void xe_gt_sriov_pf_control_process_restore_data(struct xe_gt *gt, unsigned int vfid);
>  int xe_gt_sriov_pf_control_trigger_restore_vf(struct xe_gt *gt, unsigned int vfid);
>  int xe_gt_sriov_pf_control_finish_restore_vf(struct xe_gt *gt, unsigned int vfid);
>  int xe_gt_sriov_pf_control_stop_vf(struct xe_gt *gt, unsigned int vfid);
> diff --git a/drivers/gpu/drm/xe/xe_gt_sriov_pf_control_types.h b/drivers/gpu/drm/xe/xe_gt_sriov_pf_control_types.h
> index ea16ceb39dabb..a14c738581ae2 100644
> --- a/drivers/gpu/drm/xe/xe_gt_sriov_pf_control_types.h
> +++ b/drivers/gpu/drm/xe/xe_gt_sriov_pf_control_types.h
> @@ -32,9 +32,15 @@
>   * @XE_GT_SRIOV_STATE_PAUSE_FAILED: indicates that a VF pause operation has failed.
>   * @XE_GT_SRIOV_STATE_PAUSED: indicates that the VF is paused.
>   * @XE_GT_SRIOV_STATE_SAVE_WIP: indicates that VF save operation is in progress.
> + * @XE_GT_SRIOV_STATE_SAVE_PROCESS_DATA: indicates that VF migration data is being produced.
> + * @XE_GT_SRIOV_STATE_SAVE_WAIT_DATA: indicates that PF awaits for space in migration data ring.
> + * @XE_GT_SRIOV_STATE_SAVE_DATA_DONE: indicates that all migration data was produced by Xe.
>   * @XE_GT_SRIOV_STATE_SAVE_FAILED: indicates that VF save operation has failed.
>   * @XE_GT_SRIOV_STATE_SAVED: indicates that VF data is saved.
>   * @XE_GT_SRIOV_STATE_RESTORE_WIP: indicates that VF restore operation is in progress.
> + * @XE_GT_SRIOV_STATE_RESTORE_PROCESS_DATA: indicates that VF migration data is being consumed.
> + * @XE_GT_SRIOV_STATE_RESTORE_WAIT_DATA: indicates that PF awaits for data in migration data ring.
> + * @XE_GT_SRIOV_STATE_RESTORE_DATA_DONE: indicates that all migration data was produced by the user.
>   * @XE_GT_SRIOV_STATE_RESTORE_FAILED: indicates that VF restore operation has failed.
>   * @XE_GT_SRIOV_STATE_RESTORED: indicates that VF data is restored.
>   * @XE_GT_SRIOV_STATE_RESUME_WIP: indicates the a VF resume operation is in progress.
> @@ -70,10 +76,16 @@ enum xe_gt_sriov_control_bits {
>  	XE_GT_SRIOV_STATE_PAUSED,
>  
>  	XE_GT_SRIOV_STATE_SAVE_WIP,
> +	XE_GT_SRIOV_STATE_SAVE_PROCESS_DATA,
> +	XE_GT_SRIOV_STATE_SAVE_WAIT_DATA,
> +	XE_GT_SRIOV_STATE_SAVE_DATA_DONE,
>  	XE_GT_SRIOV_STATE_SAVE_FAILED,
>  	XE_GT_SRIOV_STATE_SAVED,
>  
>  	XE_GT_SRIOV_STATE_RESTORE_WIP,
> +	XE_GT_SRIOV_STATE_RESTORE_PROCESS_DATA,
> +	XE_GT_SRIOV_STATE_RESTORE_WAIT_DATA,
> +	XE_GT_SRIOV_STATE_RESTORE_DATA_DONE,
>  	XE_GT_SRIOV_STATE_RESTORE_FAILED,
>  	XE_GT_SRIOV_STATE_RESTORED,
>  
> diff --git a/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.c b/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.c
> index ca28f45aaf481..e61e6f4215864 100644
> --- a/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.c
> +++ b/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.c
> @@ -7,6 +7,7 @@
>  
>  #include "abi/guc_actions_sriov_abi.h"
>  #include "xe_bo.h"
> +#include "xe_gt_sriov_pf_control.h"
>  #include "xe_gt_sriov_pf_helpers.h"
>  #include "xe_gt_sriov_pf_migration.h"
>  #include "xe_gt_sriov_printk.h"
> @@ -15,6 +16,17 @@
>  #include "xe_sriov.h"
>  #include "xe_sriov_pf_migration.h"
>  
> +#define XE_GT_SRIOV_PF_MIGRATION_RING_SIZE 5
> +
> +static struct xe_gt_sriov_migration_data *pf_pick_gt_migration(struct xe_gt *gt, unsigned int vfid)
> +{
> +	xe_gt_assert(gt, IS_SRIOV_PF(gt_to_xe(gt)));
> +	xe_gt_assert(gt, vfid != PFID);
> +	xe_gt_assert(gt, vfid <= xe_sriov_pf_get_totalvfs(gt_to_xe(gt)));
> +
> +	return &gt->sriov.pf.vfs[vfid].migration;
> +}
> +
>  /* Return: number of dwords saved/restored/required or a negative error code on failure */
>  static int guc_action_vf_save_restore(struct xe_guc *guc, u32 vfid, u32 opcode,
>  				      u64 addr, u32 ndwords)
> @@ -382,6 +394,162 @@ ssize_t xe_gt_sriov_pf_migration_write_guc_state(struct xe_gt *gt, unsigned int
>  }
>  #endif /* CONFIG_DEBUG_FS */
>  
> +/**
> + * xe_gt_sriov_pf_migration_ring_empty() - Check if a migration ring is empty.
> + * @gt: the &xe_gt
> + * @vfid: the VF identifier
> + *
> + * Return: true if the ring is empty, otherwise false.
> + */
> +bool xe_gt_sriov_pf_migration_ring_empty(struct xe_gt *gt, unsigned int vfid)
> +{
> +	return ptr_ring_empty(&pf_pick_gt_migration(gt, vfid)->ring);
> +}
> +
> +/**
> + * xe_gt_sriov_pf_migration_ring_full() - Check if a migration ring is full.
> + * @gt: the &xe_gt
> + * @vfid: the VF identifier
> + *
> + * Return: true if the ring is full, otherwise false.
> + */
> +bool xe_gt_sriov_pf_migration_ring_full(struct xe_gt *gt, unsigned int vfid)
> +{
> +	return ptr_ring_full(&pf_pick_gt_migration(gt, vfid)->ring);
> +}
> +
> +/**
> + * xe_gt_sriov_pf_migration_save_produce() - Add VF save data packet to migration ring.
> + * @gt: the &xe_gt
> + * @vfid: the VF identifier
> + * @data: &xe_sriov_migration_data packet
> + *
> + * Called by the save migration data producer (PF SR-IOV Control worker) when
> + * processing migration data.
> + * Wakes up the save migration data consumer (userspace), that is potentially
> + * waiting for data when the ring is empty.

... was empty ?

> + *
> + * Return: 0 on success or a negative error code on failure.
> + */
> +int xe_gt_sriov_pf_migration_save_produce(struct xe_gt *gt, unsigned int vfid,
> +					  struct xe_sriov_migration_data *data)
> +{
> +	int ret;
> +
> +	ret = ptr_ring_produce(&pf_pick_gt_migration(gt, vfid)->ring, data);
> +	if (ret)
> +		return ret;
> +
> +	wake_up_all(xe_sriov_pf_migration_waitqueue(gt_to_xe(gt), vfid));
> +
> +	return 0;
> +}
> +
> +/**
> + * xe_gt_sriov_pf_migration_restore_consume() - Get VF restore data packet from migration ring.
> + * @gt: the &xe_gt
> + * @vfid: the VF identifier
> + *
> + * Called by the restore migration data consumer (PF SR-IOV Control worker) when
> + * processing migration data.
> + * Wakes up the restore migration data producer (userspace), that is
> + * potentially waiting to add more data when the ring is full.
> + *
> + * Return: Pointer to &struct xe_sriov_migration_data on success,
> + *	   NULL if ring is empty.
> + */
> +struct xe_sriov_migration_data *
> +xe_gt_sriov_pf_migration_restore_consume(struct xe_gt *gt, unsigned int vfid)
> +{
> +	struct xe_gt_sriov_migration_data *migration = pf_pick_gt_migration(gt, vfid);
> +	struct wait_queue_head *wq = xe_sriov_pf_migration_waitqueue(gt_to_xe(gt), vfid);
> +	struct xe_sriov_migration_data *data;
> +
> +	data = ptr_ring_consume(&migration->ring);
> +	if (data)
> +		wake_up_all(wq);
> +
> +	return data;
> +}
> +
> +/**
> + * xe_gt_sriov_pf_migration_restore_produce() - Add VF restore data packet to migration ring.
> + * @gt: the &xe_gt
> + * @vfid: the VF identifier
> + * @data: &xe_sriov_migration_data packet
> + *
> + * Called by the restore migration data producer (userspace) when processing
> + * migration data.
> + * If the ring is full, waits until there is space.
> + * Queues the restore migration data consumer (PF SR-IOV Control worker), that
> + * is potentially waiting for data when the ring is empty.
> + *
> + * Return: 0 on success or a negative error code on failure.
> + */
> +int xe_gt_sriov_pf_migration_restore_produce(struct xe_gt *gt, unsigned int vfid,
> +					     struct xe_sriov_migration_data *data)
> +{
> +	struct wait_queue_head *wq = xe_sriov_pf_migration_waitqueue(gt_to_xe(gt), vfid);
> +	struct xe_gt_sriov_migration_data *migration = pf_pick_gt_migration(gt, vfid);
> +	int ret;
> +
> +	xe_gt_assert(gt, data->tile == gt->tile->id);
> +	xe_gt_assert(gt, data->gt == gt->info.id);
> +
> +	for (;;) {
> +		ret = ptr_ring_produce(&migration->ring, data);
> +		if (!ret)
> +			break;
> +
> +		ret = wait_event_interruptible(*wq, !ptr_ring_full(&migration->ring));
> +		if (ret)
> +			return ret;
> +	}
> +
> +	xe_gt_sriov_pf_control_process_restore_data(gt, vfid);
> +
> +	return 0;
> +}
> +
> +/**
> + * xe_gt_sriov_pf_migration_save_consume() - Get VF save data packet from migration ring.
> + * @gt: the &xe_gt
> + * @vfid: the VF identifier
> + *
> + * Called by the save migration data consumer (userspace) when
> + * processing migration data.
> + * Queues the save migration data producer (PF SR-IOV Control worker), that is
> + * potentially waiting to add more data when the ring is full.
> + *
> + * Return: Pointer to &struct xe_sriov_migration_data on success,
> + *	   NULL if ring is empty and there's no more data available,
> + *	   ERR_PTR(-EAGAIN) if the ring is empty, but data is still produced.
> + */
> +struct xe_sriov_migration_data *
> +xe_gt_sriov_pf_migration_save_consume(struct xe_gt *gt, unsigned int vfid)
> +{
> +	struct xe_gt_sriov_migration_data *migration = pf_pick_gt_migration(gt, vfid);
> +	struct xe_sriov_migration_data *data;
> +
> +	data = ptr_ring_consume(&migration->ring);
> +	if (data) {
> +		xe_gt_sriov_pf_control_process_save_data(gt, vfid);
> +		return data;
> +	}
> +
> +	if (xe_gt_sriov_pf_control_check_save_data_done(gt, vfid))
> +		return NULL;
> +
> +	return ERR_PTR(-EAGAIN);
> +}
> +
> +static void action_ring_cleanup(void *arg)
> +{
> +	struct ptr_ring *r = arg;
> +
> +	ptr_ring_cleanup(r, NULL);
> +}
> +
>  /**
>   * xe_gt_sriov_pf_migration_init() - Initialize support for VF migration.
>   * @gt: the &xe_gt
> @@ -393,6 +561,7 @@ ssize_t xe_gt_sriov_pf_migration_write_guc_state(struct xe_gt *gt, unsigned int
>  int xe_gt_sriov_pf_migration_init(struct xe_gt *gt)
>  {
>  	struct xe_device *xe = gt_to_xe(gt);
> +	unsigned int n, totalvfs;
>  	int err;
>  
>  	xe_gt_assert(gt, IS_SRIOV_PF(xe));
> @@ -404,5 +573,19 @@ int xe_gt_sriov_pf_migration_init(struct xe_gt *gt)
>  	if (err)
>  		return err;
>  
> +	totalvfs = xe_sriov_pf_get_totalvfs(xe);
> +	for (n = 1; n <= totalvfs; n++) {
> +		struct xe_gt_sriov_migration_data *migration = pf_pick_gt_migration(gt, n);
> +
> +		err = ptr_ring_init(&migration->ring,
> +				    XE_GT_SRIOV_PF_MIGRATION_RING_SIZE, GFP_KERNEL);
> +		if (err)
> +			return err;
> +
> +		err = devm_add_action_or_reset(xe->drm.dev, action_ring_cleanup, &migration->ring);
> +		if (err)
> +			return err;
> +	}
> +
>  	return 0;
>  }
> diff --git a/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.h b/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.h
> index 09faeae00ddbb..9e67f18ded205 100644
> --- a/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.h
> +++ b/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.h
> @@ -9,11 +9,25 @@
>  #include <linux/types.h>
>  
>  struct xe_gt;
> +struct xe_sriov_migration_data;
>  
>  int xe_gt_sriov_pf_migration_init(struct xe_gt *gt);
>  int xe_gt_sriov_pf_migration_save_guc_state(struct xe_gt *gt, unsigned int vfid);
>  int xe_gt_sriov_pf_migration_restore_guc_state(struct xe_gt *gt, unsigned int vfid);
>  
> +bool xe_gt_sriov_pf_migration_ring_empty(struct xe_gt *gt, unsigned int vfid);
> +bool xe_gt_sriov_pf_migration_ring_full(struct xe_gt *gt, unsigned int vfid);
> +
> +int xe_gt_sriov_pf_migration_save_produce(struct xe_gt *gt, unsigned int vfid,
> +					  struct xe_sriov_migration_data *data);
> +struct xe_sriov_migration_data *
> +xe_gt_sriov_pf_migration_restore_consume(struct xe_gt *gt, unsigned int vfid);
> +
> +int xe_gt_sriov_pf_migration_restore_produce(struct xe_gt *gt, unsigned int vfid,
> +					     struct xe_sriov_migration_data *data);
> +struct xe_sriov_migration_data *
> +xe_gt_sriov_pf_migration_save_consume(struct xe_gt *gt, unsigned int vfid);
> +
>  #ifdef CONFIG_DEBUG_FS
>  ssize_t xe_gt_sriov_pf_migration_read_guc_state(struct xe_gt *gt, unsigned int vfid,
>  						char __user *buf, size_t count, loff_t *pos);
> diff --git a/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration_types.h b/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration_types.h
> index 9d672feac5f04..84be6fac16c8b 100644
> --- a/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration_types.h
> +++ b/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration_types.h
> @@ -7,6 +7,7 @@
>  #define _XE_GT_SRIOV_PF_MIGRATION_TYPES_H_
>  
>  #include <linux/mutex.h>
> +#include <linux/ptr_ring.h>
>  #include <linux/types.h>
>  
>  /**
> @@ -24,6 +25,16 @@ struct xe_gt_sriov_state_snapshot {
>  	} guc;
>  };
>  
> +/**
> + * struct xe_gt_sriov_migration_data - GT-level per-VF migration data.
> + *
> + * Used by the PF driver to maintain per-VF migration data.
> + */
> +struct xe_gt_sriov_migration_data {
> +	/** @ring: queue containing VF save / restore migration data */
> +	struct ptr_ring ring;
> +};
> +
>  /**
>   * struct xe_gt_sriov_pf_migration - GT-level data.
>   *
> diff --git a/drivers/gpu/drm/xe/xe_gt_sriov_pf_types.h b/drivers/gpu/drm/xe/xe_gt_sriov_pf_types.h
> index a64a6835ad656..812e74d3f8f80 100644
> --- a/drivers/gpu/drm/xe/xe_gt_sriov_pf_types.h
> +++ b/drivers/gpu/drm/xe/xe_gt_sriov_pf_types.h
> @@ -33,6 +33,9 @@ struct xe_gt_sriov_metadata {
>  
>  	/** @snapshot: snapshot of the VF state data */
>  	struct xe_gt_sriov_state_snapshot snapshot;
> +
> +	/** @migration: per-VF migration data. */
> +	struct xe_gt_sriov_migration_data migration;
>  };
>  
>  /**
> diff --git a/drivers/gpu/drm/xe/xe_sriov_pf_migration.c b/drivers/gpu/drm/xe/xe_sriov_pf_migration.c
> index 8c523c392f98b..7be9f026d80e8 100644
> --- a/drivers/gpu/drm/xe/xe_sriov_pf_migration.c
> +++ b/drivers/gpu/drm/xe/xe_sriov_pf_migration.c
> @@ -3,8 +3,36 @@
>   * Copyright © 2025 Intel Corporation
>   */
>  
> +#include <drm/drm_managed.h>
> +
> +#include "xe_device.h"
> +#include "xe_gt_sriov_pf_control.h"
> +#include "xe_gt_sriov_pf_migration.h"
> +#include "xe_pm.h"
>  #include "xe_sriov.h"
> +#include "xe_sriov_pf_helpers.h"
>  #include "xe_sriov_pf_migration.h"
> +#include "xe_sriov_printk.h"
> +
> +static struct xe_sriov_pf_migration *pf_pick_migration(struct xe_device *xe, unsigned int vfid)
> +{
> +	xe_assert(xe, IS_SRIOV_PF(xe));
> +	xe_assert(xe, vfid <= xe_sriov_pf_get_totalvfs(xe));
> +
> +	return &xe->sriov.pf.vfs[vfid].migration;
> +}
> +
> +/**
> + * xe_sriov_pf_migration_waitqueue - Get waitqueue for migration.
> + * @xe: the &xe_device
> + * @vfid: the VF identifier
> + *
> + * Return: pointer to the migration waitqueue.
> + */
> +wait_queue_head_t *xe_sriov_pf_migration_waitqueue(struct xe_device *xe, unsigned int vfid)
> +{
> +	return &pf_pick_migration(xe, vfid)->wq;
> +}
>  
>  /**
>   * xe_sriov_pf_migration_supported() - Check if SR-IOV VF migration is supported by the device
> @@ -33,9 +61,123 @@ static bool pf_check_migration_support(struct xe_device *xe)
>   */
>  int xe_sriov_pf_migration_init(struct xe_device *xe)
>  {
> +	unsigned int n, totalvfs;
> +
>  	xe_assert(xe, IS_SRIOV_PF(xe));
>  
>  	xe->sriov.pf.migration.supported = pf_check_migration_support(xe);
> +	if (!xe_sriov_pf_migration_supported(xe))
> +		return 0;
> +
> +	totalvfs = xe_sriov_pf_get_totalvfs(xe);
> +	for (n = 1; n <= totalvfs; n++) {
> +		struct xe_sriov_pf_migration *migration = pf_pick_migration(xe, n);
> +
> +		init_waitqueue_head(&migration->wq);
> +	}
>  
>  	return 0;
>  }
> +
> +static bool pf_migration_data_ready(struct xe_device *xe, unsigned int vfid)
> +{
> +	struct xe_gt *gt;
> +	u8 gt_id;
> +
> +	for_each_gt(gt, xe, gt_id) {
> +		if (!xe_gt_sriov_pf_migration_ring_empty(gt, vfid) ||
> +		    xe_gt_sriov_pf_control_check_save_data_done(gt, vfid))
> +			return true;
> +	}
> +
> +	return false;
> +}
> +
> +static struct xe_sriov_migration_data *
> +pf_migration_consume(struct xe_device *xe, unsigned int vfid)
> +{
> +	struct xe_sriov_migration_data *data;
> +	struct xe_gt *gt;
> +	u8 gt_id;
> +	bool more_data = false;

nit: vars in rev xmas tree order?

> +
> +	for_each_gt(gt, xe, gt_id) {
> +		data = xe_gt_sriov_pf_migration_save_consume(gt, vfid);
> +		if (data && PTR_ERR(data) != EAGAIN)
> +			return data;
> +		if (PTR_ERR(data) == -EAGAIN)
> +			more_data = true;
> +	}
> +
> +	if (!more_data)
> +		return NULL;
> +
> +	return ERR_PTR(-EAGAIN);
> +}
> +
> +/**
> + * xe_sriov_pf_migration_save_consume() - Consume a VF migration data packet from the device.
> + * @xe: the &xe_device
> + * @vfid: the VF identifier
> + *
> + * Called by the save migration data consumer (userspace) when
> + * processing migration data.
> + * If there is no migration data to process, wait until more data is available.
> + *
> + * Return: Pointer to &xe_sriov_migration_data on success,
> + *	   NULL if ring is empty and no more migration data is expected,
> + *	   ERR_PTR value in case of error.
> + *
> + * Return: 0 on success or a negative error code on failure.
> + */
> +struct xe_sriov_migration_data *
> +xe_sriov_pf_migration_save_consume(struct xe_device *xe, unsigned int vfid)
> +{
> +	struct xe_sriov_pf_migration *migration = pf_pick_migration(xe, vfid);
> +	struct xe_sriov_migration_data *data;
> +	int ret;
> +
> +	xe_assert(xe, IS_SRIOV_PF(xe));
> +
> +	for (;;) {
> +		data = pf_migration_consume(xe, vfid);
> +		if (PTR_ERR(data) != -EAGAIN)
> +			break;
> +
> +		ret = wait_event_interruptible(migration->wq,
> +					       pf_migration_data_ready(xe, vfid));
> +		if (ret)
> +			return ERR_PTR(ret);
> +	}
> +
> +	return data;
> +}
> +
> +/**
> + * xe_sriov_pf_migration_restore_produce() - Produce a VF migration data packet to the device.
> + * @xe: the &xe_device
> + * @vfid: the VF identifier
> + * @data: Pointer to &xe_sriov_migration_data
> + *
> + * Called by the restore migration data producer (userspace) when processing
> + * migration data.
> + * If the underlying data structure is full, wait until there is space.
> + *
> + * Return: 0 on success or a negative error code on failure.
> + */
> +int xe_sriov_pf_migration_restore_produce(struct xe_device *xe, unsigned int vfid,
> +					  struct xe_sriov_migration_data *data)
> +{
> +	struct xe_gt *gt;
> +
> +	xe_assert(xe, IS_SRIOV_PF(xe));
> +
> +	gt = xe_device_get_gt(xe, data->gt);
> +	if (!gt || data->tile != gt->tile->id) {
> +		xe_sriov_err_ratelimited(xe, "VF%d Invalid GT - tile:%u, GT:%u\n",
> +					 vfid, data->tile, data->gt);
> +		return -EINVAL;
> +	}
> +
> +	return xe_gt_sriov_pf_migration_restore_produce(gt, vfid, data);
> +}
> diff --git a/drivers/gpu/drm/xe/xe_sriov_pf_migration.h b/drivers/gpu/drm/xe/xe_sriov_pf_migration.h
> index d2b4a24165438..df81a540c246a 100644
> --- a/drivers/gpu/drm/xe/xe_sriov_pf_migration.h
> +++ b/drivers/gpu/drm/xe/xe_sriov_pf_migration.h
> @@ -7,10 +7,17 @@
>  #define _XE_SRIOV_PF_MIGRATION_H_
>  
>  #include <linux/types.h>
> +#include <linux/wait.h>
>  
>  struct xe_device;
> +struct xe_sriov_migration_data;
>  
>  int xe_sriov_pf_migration_init(struct xe_device *xe);
>  bool xe_sriov_pf_migration_supported(struct xe_device *xe);
> +int xe_sriov_pf_migration_restore_produce(struct xe_device *xe, unsigned int vfid,
> +					  struct xe_sriov_migration_data *data);
> +struct xe_sriov_migration_data *
> +xe_sriov_pf_migration_save_consume(struct xe_device *xe, unsigned int vfid);
> +wait_queue_head_t *xe_sriov_pf_migration_waitqueue(struct xe_device *xe, unsigned int vfid);
>  
>  #endif
> diff --git a/drivers/gpu/drm/xe/xe_sriov_pf_migration_types.h b/drivers/gpu/drm/xe/xe_sriov_pf_migration_types.h
> index e69de29bb2d1d..2a45ee4e3ece8 100644
> --- a/drivers/gpu/drm/xe/xe_sriov_pf_migration_types.h
> +++ b/drivers/gpu/drm/xe/xe_sriov_pf_migration_types.h
> @@ -0,0 +1,58 @@
> +/* SPDX-License-Identifier: MIT */
> +/*
> + * Copyright © 2025 Intel Corporation
> + */
> +
> +#ifndef _XE_SRIOV_PF_MIGRATION_TYPES_H_
> +#define _XE_SRIOV_PF_MIGRATION_TYPES_H_
> +
> +#include <linux/types.h>
> +#include <linux/wait.h>
> +
> +/**
> + * struct xe_sriov_migration_data - Xe SR-IOV VF migration data packet
> + */
> +struct xe_sriov_migration_data {
> +	/** @xe: Xe device */
> +	struct xe_device *xe;
> +	/** @vaddr: CPU pointer to payload data */
> +	void *vaddr;
> +	/** @remaining: payload data remaining */
> +	size_t remaining;
> +	/** @hdr_remaining: header data remaining */
> +	size_t hdr_remaining;
> +	union {
> +		/** @bo: Buffer object with migration data */
> +		struct xe_bo *bo;
> +		/** @buff: Buffer with migration data */
> +		void *buff;
> +	};
> +	__struct_group(xe_sriov_pf_migration_hdr, hdr, __packed,
> +		/** @hdr.version: migration data protocol version */
> +		u8 version;
> +		/** @hdr.type: migration data type */
> +		u8 type;
> +		/** @hdr.tile: migration data tile id */
> +		u8 tile;
> +		/** @hdr.gt: migration data gt id */
> +		u8 gt;
> +		/** @hdr.flags: migration data flags */
> +		u32 flags;
> +		/** @hdr.offset: offset into the resource;
> +		 * used when multiple packets of given type are used for migration
> +		 */
> +		u64 offset;
> +		/** @hdr.size: migration data size  */
> +		u64 size;
> +	);
> +};
> +
> +/**
> + * struct xe_sriov_pf_migration - Per VF device-level migration related data
> + */
> +struct xe_sriov_pf_migration {
> +	/** @wq: waitqueue used to avoid busy-waiting for snapshot production/consumption */
> +	wait_queue_head_t wq;
> +};
> +
> +#endif
> diff --git a/drivers/gpu/drm/xe/xe_sriov_pf_types.h b/drivers/gpu/drm/xe/xe_sriov_pf_types.h
> index 24d22afeececa..c92baaa1694ca 100644
> --- a/drivers/gpu/drm/xe/xe_sriov_pf_types.h
> +++ b/drivers/gpu/drm/xe/xe_sriov_pf_types.h
> @@ -9,6 +9,7 @@
>  #include <linux/mutex.h>
>  #include <linux/types.h>
>  
> +#include "xe_sriov_pf_migration_types.h"
>  #include "xe_sriov_pf_provision_types.h"
>  #include "xe_sriov_pf_service_types.h"
>  
> @@ -18,6 +19,8 @@
>  struct xe_sriov_metadata {
>  	/** @version: negotiated VF/PF ABI version */
>  	struct xe_sriov_pf_service_version version;
> +	/** @migration: migration data */
> +	struct xe_sriov_pf_migration migration;
>  };
>  
>  /**

otherwise LGTM

Re: [PATCH v3 07/28] drm/xe/pf: Add support for encap/decap of bitstream to/from packet

[Michal Wajdeczko]

On 10/30/2025 9:31 PM, Michał Winiarski wrote:
> Add debugfs handlers for migration state and handle bitstream
> .read()/.write() to convert from bitstream to/from migration data
> packets.
> As descriptor/trailer are handled at this layer - add handling for both
> save and restore side.
> 
> Signed-off-by: Michał Winiarski <michal.winiarski@intel.com>
> ---
>  drivers/gpu/drm/xe/xe_sriov_migration_data.c  | 337 ++++++++++++++++++
>  drivers/gpu/drm/xe/xe_sriov_migration_data.h  |   5 +
>  drivers/gpu/drm/xe/xe_sriov_pf_control.c      |   5 +
>  drivers/gpu/drm/xe/xe_sriov_pf_debugfs.c      |  35 ++
>  drivers/gpu/drm/xe/xe_sriov_pf_migration.c    |  54 +++
>  .../gpu/drm/xe/xe_sriov_pf_migration_types.h  |   9 +
>  6 files changed, 445 insertions(+)
> 
> diff --git a/drivers/gpu/drm/xe/xe_sriov_migration_data.c b/drivers/gpu/drm/xe/xe_sriov_migration_data.c
> index 2371ca3e6b9a9..a3f50836adc81 100644
> --- a/drivers/gpu/drm/xe/xe_sriov_migration_data.c
> +++ b/drivers/gpu/drm/xe/xe_sriov_migration_data.c
> @@ -6,6 +6,45 @@
>  #include "xe_bo.h"
>  #include "xe_device.h"
>  #include "xe_sriov_migration_data.h"
> +#include "xe_sriov_pf_helpers.h"
> +#include "xe_sriov_pf_migration.h"
> +#include "xe_sriov_printk.h"
> +
> +static struct mutex *pf_migration_mutex(struct xe_device *xe, unsigned int vfid)
> +{
> +	xe_assert(xe, IS_SRIOV_PF(xe));
> +	xe_assert(xe, vfid <= xe_sriov_pf_get_totalvfs(xe));
> +
> +	return &xe->sriov.pf.vfs[vfid].migration.lock;
> +}
> +
> +static struct xe_sriov_migration_data **pf_pick_pending(struct xe_device *xe, unsigned int vfid)
> +{
> +	xe_assert(xe, IS_SRIOV_PF(xe));
> +	xe_assert(xe, vfid <= xe_sriov_pf_get_totalvfs(xe));
> +	lockdep_assert_held(pf_migration_mutex(xe, vfid));
> +
> +	return &xe->sriov.pf.vfs[vfid].migration.pending;
> +}
> +
> +static struct xe_sriov_migration_data **
> +pf_pick_descriptor(struct xe_device *xe, unsigned int vfid)
> +{
> +	xe_assert(xe, IS_SRIOV_PF(xe));
> +	xe_assert(xe, vfid <= xe_sriov_pf_get_totalvfs(xe));
> +	lockdep_assert_held(pf_migration_mutex(xe, vfid));
> +
> +	return &xe->sriov.pf.vfs[vfid].migration.descriptor;
> +}
> +
> +static struct xe_sriov_migration_data **pf_pick_trailer(struct xe_device *xe, unsigned int vfid)
> +{
> +	xe_assert(xe, IS_SRIOV_PF(xe));
> +	xe_assert(xe, vfid <= xe_sriov_pf_get_totalvfs(xe));
> +	lockdep_assert_held(pf_migration_mutex(xe, vfid));
> +
> +	return &xe->sriov.pf.vfs[vfid].migration.trailer;
> +}
>  
>  static bool data_needs_bo(struct xe_sriov_migration_data *data)
>  {
> @@ -42,6 +81,9 @@ struct xe_sriov_migration_data *xe_sriov_migration_data_alloc(struct xe_device *
>   */
>  void xe_sriov_migration_data_free(struct xe_sriov_migration_data *data)
>  {
> +	if (IS_ERR_OR_NULL(data))
> +		return;

nit: maybe to be always on the safe side, this chunk should be moved to 06/28 ?

> +
>  	if (data_needs_bo(data))
>  		xe_bo_unpin_map_no_vm(data->bo);
>  	else
> @@ -124,3 +166,298 @@ int xe_sriov_migration_data_init_from_hdr(struct xe_sriov_migration_data *data)
>  
>  	return mig_data_init(data);
>  }
> +
> +static ssize_t vf_mig_data_hdr_read(struct xe_sriov_migration_data *data,
> +				    char __user *buf, size_t len)
> +{
> +	loff_t offset = sizeof(data->hdr) - data->hdr_remaining;
> +
> +	if (!data->hdr_remaining)
> +		return -EINVAL;
> +
> +	if (len > data->hdr_remaining)
> +		len = data->hdr_remaining;
> +
> +	if (copy_to_user(buf, (void *)&data->hdr + offset, len))
> +		return -EFAULT;
> +
> +	data->hdr_remaining -= len;
> +
> +	return len;
> +}
> +
> +static ssize_t vf_mig_data_read(struct xe_sriov_migration_data *data,
> +				char __user *buf, size_t len)
> +{
> +	if (len > data->remaining)
> +		len = data->remaining;
> +
> +	if (copy_to_user(buf, data->vaddr + (data->size - data->remaining), len))
> +		return -EFAULT;
> +
> +	data->remaining -= len;
> +
> +	return len;
> +}
> +
> +static ssize_t __vf_mig_data_read_single(struct xe_sriov_migration_data **data,
> +					 unsigned int vfid, char __user *buf, size_t len)
> +{
> +	ssize_t copied = 0;
> +
> +	if ((*data)->hdr_remaining)
> +		copied = vf_mig_data_hdr_read(*data, buf, len);
> +	else
> +		copied = vf_mig_data_read(*data, buf, len);
> +
> +	if ((*data)->remaining == 0 && (*data)->hdr_remaining == 0) {
> +		xe_sriov_migration_data_free(*data);
> +		*data = NULL;
> +	}
> +
> +	return copied;
> +}
> +
> +static struct xe_sriov_migration_data **vf_mig_pick_data(struct xe_device *xe, unsigned int vfid)
> +{
> +	struct xe_sriov_migration_data **data;
> +
> +	data = pf_pick_descriptor(xe, vfid);
> +	if (*data)
> +		return data;
> +
> +	data = pf_pick_pending(xe, vfid);
> +	if (!*data)
> +		*data = xe_sriov_pf_migration_save_consume(xe, vfid);
> +	if (*data)
> +		return data;
> +
> +	data = pf_pick_trailer(xe, vfid);
> +	if (*data)
> +		return data;
> +
> +	return ERR_PTR(-ENODATA);
> +}
> +
> +static ssize_t vf_mig_data_read_single(struct xe_device *xe, unsigned int vfid,
> +				       char __user *buf, size_t len)
> +{
> +	struct xe_sriov_migration_data **data = vf_mig_pick_data(xe, vfid);
> +
> +	if (IS_ERR(data))
> +		return PTR_ERR(data);
> +
> +	return __vf_mig_data_read_single(data, vfid, buf, len);
> +}
> +
> +/**
> + * xe_sriov_migration_data_read() - Read migration data from the device.
> + * @xe: the &xe_device
> + * @vfid: the VF identifier
> + * @buf: start address of userspace buffer
> + * @len: requested read size from userspace
> + *
> + * Return: number of bytes that has been successfully read,
> + *	   0 if no more migration data is available,
> + *	   -errno on failure.
> + */
> +ssize_t xe_sriov_migration_data_read(struct xe_device *xe, unsigned int vfid,
> +				     char __user *buf, size_t len)
> +{
> +	ssize_t ret, consumed = 0;
> +
> +	xe_assert(xe, IS_SRIOV_PF(xe));
> +
> +	scoped_cond_guard(mutex_intr, return -EINTR, pf_migration_mutex(xe, vfid)) {
> +		while (consumed < len) {
> +			ret = vf_mig_data_read_single(xe, vfid, buf, len - consumed);
> +			if (ret == -ENODATA)
> +				break;
> +			if (ret < 0)
> +				return ret;
> +
> +			consumed += ret;
> +			buf += ret;
> +		}
> +	}
> +
> +	return consumed;
> +}
> +
> +static ssize_t vf_mig_hdr_write(struct xe_sriov_migration_data *data,
> +				const char __user *buf, size_t len)
> +{
> +	loff_t offset = sizeof(data->hdr) - data->hdr_remaining;
> +	int ret;
> +
> +	if (len > data->hdr_remaining)
> +		len = data->hdr_remaining;
> +
> +	if (copy_from_user((void *)&data->hdr + offset, buf, len))
> +		return -EFAULT;
> +
> +	data->hdr_remaining -= len;
> +
> +	if (!data->hdr_remaining) {
> +		ret = xe_sriov_migration_data_init_from_hdr(data);
> +		if (ret)
> +			return ret;
> +	}
> +
> +	return len;
> +}
> +
> +static ssize_t vf_mig_data_write(struct xe_sriov_migration_data *data,
> +				 const char __user *buf, size_t len)
> +{
> +	if (len > data->remaining)
> +		len = data->remaining;
> +
> +	if (copy_from_user(data->vaddr + (data->size - data->remaining), buf, len))
> +		return -EFAULT;
> +
> +	data->remaining -= len;
> +
> +	return len;
> +}
> +
> +static ssize_t vf_mig_data_write_single(struct xe_device *xe, unsigned int vfid,
> +					const char __user *buf, size_t len)
> +{
> +	struct xe_sriov_migration_data **data = pf_pick_pending(xe, vfid);
> +	int ret;
> +	ssize_t copied;
> +
> +	if (IS_ERR_OR_NULL(*data)) {
> +		*data = xe_sriov_migration_data_alloc(xe);
> +		if (!*data)
> +			return -ENOMEM;
> +	}
> +
> +	if ((*data)->hdr_remaining)
> +		copied = vf_mig_hdr_write(*data, buf, len);
> +	else
> +		copied = vf_mig_data_write(*data, buf, len);
> +
> +	if ((*data)->hdr_remaining == 0 && (*data)->remaining == 0) {
> +		ret = xe_sriov_pf_migration_restore_produce(xe, vfid, *data);
> +		if (ret) {
> +			xe_sriov_migration_data_free(*data);
> +			return ret;
> +		}
> +
> +		*data = NULL;
> +	}
> +
> +	return copied;
> +}
> +
> +/**
> + * xe_sriov_migration_data_write() - Write migration data to the device.
> + * @xe: the &xe_device
> + * @vfid: the VF identifier
> + * @buf: start address of userspace buffer
> + * @len: requested write size from userspace
> + *
> + * Return: number of bytes that has been successfully written,
> + *	   -errno on failure.
> + */
> +ssize_t xe_sriov_migration_data_write(struct xe_device *xe, unsigned int vfid,
> +				      const char __user *buf, size_t len)
> +{
> +	ssize_t ret, produced = 0;
> +
> +	xe_assert(xe, IS_SRIOV_PF(xe));
> +
> +	scoped_cond_guard(mutex_intr, return -EINTR, pf_migration_mutex(xe, vfid)) {
> +		while (produced < len) {
> +			ret = vf_mig_data_write_single(xe, vfid, buf, len - produced);
> +			if (ret < 0)
> +				return ret;
> +
> +			produced += ret;
> +			buf += ret;
> +		}
> +	}
> +
> +	return produced;
> +}
> +
> +#define MIGRATION_DESCRIPTOR_DWORDS 0
> +static size_t pf_descriptor_init(struct xe_device *xe, unsigned int vfid)
> +{
> +	struct xe_sriov_migration_data **desc = pf_pick_descriptor(xe, vfid);
> +	struct xe_sriov_migration_data *data;
> +	int ret;
> +
> +	data = xe_sriov_migration_data_alloc(xe);
> +	if (!data)
> +		return -ENOMEM;
> +
> +	ret = xe_sriov_migration_data_init(data, 0, 0, XE_SRIOV_MIGRATION_DATA_TYPE_DESCRIPTOR,
> +					   0, MIGRATION_DESCRIPTOR_DWORDS * sizeof(u32));
> +	if (ret) {
> +		xe_sriov_migration_data_free(data);
> +		return ret;
> +	}
> +
> +	*desc = data;
> +
> +	return 0;
> +}
> +
> +static void pf_pending_init(struct xe_device *xe, unsigned int vfid)
> +{
> +	struct xe_sriov_migration_data **data = pf_pick_pending(xe, vfid);
> +
> +	*data = NULL;
> +}
> +
> +#define MIGRATION_TRAILER_SIZE 0
> +static int pf_trailer_init(struct xe_device *xe, unsigned int vfid)
> +{
> +	struct xe_sriov_migration_data **trailer = pf_pick_trailer(xe, vfid);
> +	struct xe_sriov_migration_data *data;
> +	int ret;
> +
> +	data = xe_sriov_migration_data_alloc(xe);
> +	if (!data)
> +		return -ENOMEM;
> +
> +	ret = xe_sriov_migration_data_init(data, 0, 0, XE_SRIOV_MIGRATION_DATA_TYPE_TRAILER,
> +					   0, MIGRATION_TRAILER_SIZE);
> +	if (ret) {
> +		xe_sriov_migration_data_free(data);
> +		return ret;
> +	}
> +
> +	*trailer = data;
> +
> +	return 0;
> +}
> +
> +/**
> + * xe_sriov_migration_data_save_init() - Initialize the pending save migration data.
> + * @xe: the &xe_device
> + * @vfid: the VF identifier
> + *
> + * Return: 0 on success, -errno on failure.
> + */
> +int xe_sriov_migration_data_save_init(struct xe_device *xe, unsigned int vfid)
> +{
> +	int ret;
> +
> +	scoped_cond_guard(mutex_intr, return -EINTR, pf_migration_mutex(xe, vfid)) {
> +		ret = pf_descriptor_init(xe, vfid);
> +		if (ret)
> +			return ret;
> +
> +		ret = pf_trailer_init(xe, vfid);
> +		if (ret)
> +			return ret;
> +
> +		pf_pending_init(xe, vfid);
> +	}
> +
> +	return 0;
> +}
> diff --git a/drivers/gpu/drm/xe/xe_sriov_migration_data.h b/drivers/gpu/drm/xe/xe_sriov_migration_data.h
> index 3958f58a170f5..7ec489c3f28d2 100644
> --- a/drivers/gpu/drm/xe/xe_sriov_migration_data.h
> +++ b/drivers/gpu/drm/xe/xe_sriov_migration_data.h
> @@ -26,5 +26,10 @@ void xe_sriov_migration_data_free(struct xe_sriov_migration_data *snapshot);
>  int xe_sriov_migration_data_init(struct xe_sriov_migration_data *data, u8 tile_id, u8 gt_id,
>  				 enum xe_sriov_migration_data_type, loff_t offset, size_t size);
>  int xe_sriov_migration_data_init_from_hdr(struct xe_sriov_migration_data *snapshot);
> +ssize_t xe_sriov_migration_data_read(struct xe_device *xe, unsigned int vfid,
> +				     char __user *buf, size_t len);
> +ssize_t xe_sriov_migration_data_write(struct xe_device *xe, unsigned int vfid,
> +				      const char __user *buf, size_t len);
> +int xe_sriov_migration_data_save_init(struct xe_device *xe, unsigned int vfid);
>  
>  #endif
> diff --git a/drivers/gpu/drm/xe/xe_sriov_pf_control.c b/drivers/gpu/drm/xe/xe_sriov_pf_control.c
> index 8d8a01faf5291..c2768848daba1 100644
> --- a/drivers/gpu/drm/xe/xe_sriov_pf_control.c
> +++ b/drivers/gpu/drm/xe/xe_sriov_pf_control.c
> @@ -5,6 +5,7 @@
>  
>  #include "xe_device.h"
>  #include "xe_gt_sriov_pf_control.h"
> +#include "xe_sriov_migration_data.h"
>  #include "xe_sriov_pf_control.h"
>  #include "xe_sriov_printk.h"
>  
> @@ -165,6 +166,10 @@ int xe_sriov_pf_control_trigger_save_vf(struct xe_device *xe, unsigned int vfid)
>  	unsigned int id;
>  	int ret;
>  
> +	ret = xe_sriov_migration_data_save_init(xe, vfid);
> +	if (ret)
> +		return ret;
> +
>  	for_each_gt(gt, xe, id) {
>  		ret = xe_gt_sriov_pf_control_trigger_save_vf(gt, vfid);
>  		if (ret)
> diff --git a/drivers/gpu/drm/xe/xe_sriov_pf_debugfs.c b/drivers/gpu/drm/xe/xe_sriov_pf_debugfs.c
> index e0e6340c49106..a9a28aec22421 100644
> --- a/drivers/gpu/drm/xe/xe_sriov_pf_debugfs.c
> +++ b/drivers/gpu/drm/xe/xe_sriov_pf_debugfs.c
> @@ -9,6 +9,7 @@
>  #include "xe_device.h"
>  #include "xe_device_types.h"
>  #include "xe_pm.h"
> +#include "xe_sriov_migration_data.h"
>  #include "xe_sriov_pf.h"
>  #include "xe_sriov_pf_control.h"
>  #include "xe_sriov_pf_debugfs.h"
> @@ -132,6 +133,7 @@ static void pf_populate_pf(struct xe_device *xe, struct dentry *pfdent)
>   *      /sys/kernel/debug/dri/BDF/
>   *      ├── sriov
>   *      │   ├── vf1
> + *      │   │   ├── migration_data
>   *      │   │   ├── pause
>   *      │   │   ├── reset
>   *      │   │   ├── resume
> @@ -220,6 +222,38 @@ DEFINE_VF_CONTROL_ATTRIBUTE(reset_vf);
>  DEFINE_VF_CONTROL_ATTRIBUTE_RW(save_vf);
>  DEFINE_VF_CONTROL_ATTRIBUTE_RW(restore_vf);
>  
> +static ssize_t data_write(struct file *file, const char __user *buf, size_t count, loff_t *pos)
> +{
> +	struct dentry *dent = file_dentry(file)->d_parent;
> +	struct xe_device *xe = extract_xe(dent);
> +	unsigned int vfid = extract_vfid(dent);
> +
> +	if (*pos)
> +		return -ESPIPE;
> +
> +	return xe_sriov_migration_data_write(xe, vfid, buf, count);
> +}
> +
> +static ssize_t data_read(struct file *file, char __user *buf, size_t count, loff_t *ppos)
> +{
> +	struct dentry *dent = file_dentry(file)->d_parent;
> +	struct xe_device *xe = extract_xe(dent);
> +	unsigned int vfid = extract_vfid(dent);
> +
> +	if (*ppos)
> +		return -ESPIPE;
> +
> +	return xe_sriov_migration_data_read(xe, vfid, buf, count);
> +}
> +
> +static const struct file_operations data_vf_fops = {
> +	.owner		= THIS_MODULE,
> +	.open		= simple_open,
> +	.write		= data_write,
> +	.read		= data_read,
> +	.llseek		= default_llseek,
> +};
> +
>  static void pf_populate_vf(struct xe_device *xe, struct dentry *vfdent)
>  {
>  	debugfs_create_file("pause", 0200, vfdent, xe, &pause_vf_fops);
> @@ -228,6 +262,7 @@ static void pf_populate_vf(struct xe_device *xe, struct dentry *vfdent)
>  	debugfs_create_file("reset", 0200, vfdent, xe, &reset_vf_fops);
>  	debugfs_create_file("save", 0600, vfdent, xe, &save_vf_fops);
>  	debugfs_create_file("restore", 0600, vfdent, xe, &restore_vf_fops);
> +	debugfs_create_file("migration_data", 0600, vfdent, xe, &data_vf_fops);
>  }
>  
>  static void pf_populate_with_tiles(struct xe_device *xe, struct dentry *dent, unsigned int vfid)
> diff --git a/drivers/gpu/drm/xe/xe_sriov_pf_migration.c b/drivers/gpu/drm/xe/xe_sriov_pf_migration.c
> index 7be9f026d80e8..8ea531d36f53b 100644
> --- a/drivers/gpu/drm/xe/xe_sriov_pf_migration.c
> +++ b/drivers/gpu/drm/xe/xe_sriov_pf_migration.c
> @@ -10,6 +10,7 @@
>  #include "xe_gt_sriov_pf_migration.h"
>  #include "xe_pm.h"
>  #include "xe_sriov.h"
> +#include "xe_sriov_migration_data.h"
>  #include "xe_sriov_pf_helpers.h"
>  #include "xe_sriov_pf_migration.h"
>  #include "xe_sriov_printk.h"
> @@ -53,6 +54,15 @@ static bool pf_check_migration_support(struct xe_device *xe)
>  	return IS_ENABLED(CONFIG_DRM_XE_DEBUG);
>  }
>  
> +static void pf_migration_cleanup(void *arg)
> +{
> +	struct xe_sriov_pf_migration *migration = arg;
> +
> +	xe_sriov_migration_data_free(migration->pending);
> +	xe_sriov_migration_data_free(migration->trailer);
> +	xe_sriov_migration_data_free(migration->descriptor);
> +}
> +
>  /**
>   * xe_sriov_pf_migration_init() - Initialize support for SR-IOV VF migration.
>   * @xe: the &xe_device
> @@ -62,6 +72,7 @@ static bool pf_check_migration_support(struct xe_device *xe)
>  int xe_sriov_pf_migration_init(struct xe_device *xe)
>  {
>  	unsigned int n, totalvfs;
> +	int err;
>  
>  	xe_assert(xe, IS_SRIOV_PF(xe));
>  
> @@ -73,7 +84,15 @@ int xe_sriov_pf_migration_init(struct xe_device *xe)
>  	for (n = 1; n <= totalvfs; n++) {
>  		struct xe_sriov_pf_migration *migration = pf_pick_migration(xe, n);
>  
> +		err = devm_mutex_init(xe->drm.dev, &migration->lock);

IIRC all software data allocations/inits we are doing as drmm
only actions that interacts with or cleanups the hw (*) use devm

> +		if (err)
> +			return err;
> +
>  		init_waitqueue_head(&migration->wq);
> +
> +		err = devm_add_action_or_reset(xe->drm.dev, pf_migration_cleanup, migration);

(*) like here

> +		if (err)
> +			return err;
>  	}
>  
>  	return 0;
> @@ -153,6 +172,36 @@ xe_sriov_pf_migration_save_consume(struct xe_device *xe, unsigned int vfid)
>  	return data;
>  }
>  
> +static int pf_handle_descriptor(struct xe_device *xe, unsigned int vfid,
> +				struct xe_sriov_migration_data *data)
> +{
> +	if (data->tile != 0 || data->gt != 0)
> +		return -EINVAL;
> +
> +	xe_sriov_migration_data_free(data);
> +
> +	return 0;
> +}
> +
> +static int pf_handle_trailer(struct xe_device *xe, unsigned int vfid,
> +			     struct xe_sriov_migration_data *data)
> +{
> +	struct xe_gt *gt;
> +	u8 gt_id;
> +
> +	if (data->tile != 0 || data->gt != 0)
> +		return -EINVAL;
> +	if (data->offset != 0 || data->size != 0 || data->buff || data->bo)
> +		return -EINVAL;
> +
> +	xe_sriov_migration_data_free(data);
> +
> +	for_each_gt(gt, xe, gt_id)
> +		xe_gt_sriov_pf_control_restore_data_done(gt, vfid);
> +
> +	return 0;
> +}
> +
>  /**
>   * xe_sriov_pf_migration_restore_produce() - Produce a VF migration data packet to the device.
>   * @xe: the &xe_device
> @@ -172,6 +221,11 @@ int xe_sriov_pf_migration_restore_produce(struct xe_device *xe, unsigned int vfi
>  
>  	xe_assert(xe, IS_SRIOV_PF(xe));
>  
> +	if (data->type == XE_SRIOV_MIGRATION_DATA_TYPE_DESCRIPTOR)
> +		return pf_handle_descriptor(xe, vfid, data);
> +	if (data->type == XE_SRIOV_MIGRATION_DATA_TYPE_TRAILER)
> +		return pf_handle_trailer(xe, vfid, data);
> +
>  	gt = xe_device_get_gt(xe, data->gt);
>  	if (!gt || data->tile != gt->tile->id) {
>  		xe_sriov_err_ratelimited(xe, "VF%d Invalid GT - tile:%u, GT:%u\n",
> diff --git a/drivers/gpu/drm/xe/xe_sriov_pf_migration_types.h b/drivers/gpu/drm/xe/xe_sriov_pf_migration_types.h
> index 2a45ee4e3ece8..8468e5eeb6d66 100644
> --- a/drivers/gpu/drm/xe/xe_sriov_pf_migration_types.h
> +++ b/drivers/gpu/drm/xe/xe_sriov_pf_migration_types.h
> @@ -7,6 +7,7 @@
>  #define _XE_SRIOV_PF_MIGRATION_TYPES_H_
>  
>  #include <linux/types.h>
> +#include <linux/mutex_types.h>
>  #include <linux/wait.h>
>  
>  /**
> @@ -53,6 +54,14 @@ struct xe_sriov_migration_data {
>  struct xe_sriov_pf_migration {
>  	/** @wq: waitqueue used to avoid busy-waiting for snapshot production/consumption */
>  	wait_queue_head_t wq;
> +	/** @lock: Mutex protecting the migration data */
> +	struct mutex lock;
> +	/** @pending: currently processed data packet of VF resource */
> +	struct xe_sriov_migration_data *pending;
> +	/** @trailer: data packet used to indicate the end of stream */
> +	struct xe_sriov_migration_data *trailer;
> +	/** @descriptor: data packet containing the metadata describing the device */
> +	struct xe_sriov_migration_data *descriptor;
>  };
>  
>  #endif

with devm/drmm clarified,

	Reviewed-by: Michal Wajdeczko <michal.wajdeczko@intel.com>

Re: [PATCH v3 08/28] drm/xe/pf: Add minimalistic migration descriptor

[Michal Wajdeczko]

On 10/30/2025 9:31 PM, Michał Winiarski wrote:
> The descriptor reuses the KLV format used by GuC and contains metadata
> that can be used to quickly fail migration when source is incompatible
> with destination.
> 
> Signed-off-by: Michał Winiarski <michal.winiarski@intel.com>
> ---
>  drivers/gpu/drm/xe/xe_sriov_migration_data.c | 89 +++++++++++++++++++-
>  drivers/gpu/drm/xe/xe_sriov_migration_data.h |  2 +
>  drivers/gpu/drm/xe/xe_sriov_pf_migration.c   |  6 ++
>  3 files changed, 96 insertions(+), 1 deletion(-)
> 
> diff --git a/drivers/gpu/drm/xe/xe_sriov_migration_data.c b/drivers/gpu/drm/xe/xe_sriov_migration_data.c
> index a3f50836adc81..18e17706772fc 100644
> --- a/drivers/gpu/drm/xe/xe_sriov_migration_data.c
> +++ b/drivers/gpu/drm/xe/xe_sriov_migration_data.c
> @@ -5,6 +5,7 @@
>  
>  #include "xe_bo.h"
>  #include "xe_device.h"
> +#include "xe_guc_klv_helpers.h"
>  #include "xe_sriov_migration_data.h"
>  #include "xe_sriov_pf_helpers.h"
>  #include "xe_sriov_pf_migration.h"
> @@ -383,11 +384,19 @@ ssize_t xe_sriov_migration_data_write(struct xe_device *xe, unsigned int vfid,
>  	return produced;
>  }
>  
> -#define MIGRATION_DESCRIPTOR_DWORDS 0
> +#define MIGRATION_KLV_DEVICE_DEVID_KEY	0xf001u
> +#define MIGRATION_KLV_DEVICE_DEVID_LEN	1u
> +#define MIGRATION_KLV_DEVICE_REVID_KEY	0xf002u
> +#define MIGRATION_KLV_DEVICE_REVID_LEN	1u
> +
> +#define MIGRATION_DESCRIPTOR_DWORDS	(GUC_KLV_LEN_MIN + MIGRATION_KLV_DEVICE_DEVID_LEN + \
> +					 GUC_KLV_LEN_MIN + MIGRATION_KLV_DEVICE_REVID_LEN)
>  static size_t pf_descriptor_init(struct xe_device *xe, unsigned int vfid)
>  {
>  	struct xe_sriov_migration_data **desc = pf_pick_descriptor(xe, vfid);
>  	struct xe_sriov_migration_data *data;
> +	unsigned int len = 0;
> +	u32 *klvs;
>  	int ret;
>  
>  	data = xe_sriov_migration_data_alloc(xe);
> @@ -401,11 +410,89 @@ static size_t pf_descriptor_init(struct xe_device *xe, unsigned int vfid)
>  		return ret;
>  	}
>  
> +	klvs = data->vaddr;
> +	klvs[len++] = PREP_GUC_KLV_CONST(MIGRATION_KLV_DEVICE_DEVID_KEY,
> +					 MIGRATION_KLV_DEVICE_DEVID_LEN);
> +	klvs[len++] = xe->info.devid;
> +	klvs[len++] = PREP_GUC_KLV_CONST(MIGRATION_KLV_DEVICE_REVID_KEY,
> +					 MIGRATION_KLV_DEVICE_REVID_LEN);
> +	klvs[len++] = xe->info.revid;
> +
> +	xe_assert(xe, len == MIGRATION_DESCRIPTOR_DWORDS);
> +
>  	*desc = data;
>  
>  	return 0;
>  }
>  
> +/**
> + * xe_sriov_migration_data_process_descriptor() - Process migration data descriptor.
> + * @xe: the &xe_device
> + * @vfid: the VF identifier
> + * @data: the &struct xe_sriov_pf_migration_data containing the descriptor
> + *
> + * The descriptor uses the same KLV format as GuC, and contains metadata used for
> + * checking migration data compatibility.
> + *
> + * Return: 0 on success, -errno on failure.
> + */
> +int xe_sriov_migration_data_process_descriptor(struct xe_device *xe, unsigned int vfid,
> +					       struct xe_sriov_migration_data *data)
> +{
> +	u32 num_dwords = data->size / sizeof(u32);
> +	u32 *klvs = data->vaddr;
> +
> +	xe_assert(xe, data->type == XE_SRIOV_MIGRATION_DATA_TYPE_DESCRIPTOR);
> +
> +	if (data->size % sizeof(u32)) {
> +		xe_sriov_warn(xe, "Aborting migration, descriptor not in KLV format (size=%llu)\n",
> +			      data->size);
> +		return -EINVAL;
> +	}
> +
> +	while (num_dwords >= GUC_KLV_LEN_MIN) {
> +		u32 key = FIELD_GET(GUC_KLV_0_KEY, klvs[0]);
> +		u32 len = FIELD_GET(GUC_KLV_0_LEN, klvs[0]);
> +
> +		klvs += GUC_KLV_LEN_MIN;
> +		num_dwords -= GUC_KLV_LEN_MIN;
> +
> +		if (len > num_dwords)

nit:
			xe_sriov_warn(xe, "Aborting migration, truncated KLV %#x, len %u\n",


> +			return -EINVAL;
> +
> +		switch (key) {
> +		case MIGRATION_KLV_DEVICE_DEVID_KEY:
> +			if (*klvs != xe->info.devid) {
> +				xe_sriov_warn(xe,
> +					      "Aborting migration, devid mismatch %#06x!=%#06x\n",
> +					      *klvs, xe->info.devid);
> +				return -ENODEV;
> +			}
> +			break;
> +		case MIGRATION_KLV_DEVICE_REVID_KEY:
> +			if (*klvs != xe->info.revid) {
> +				xe_sriov_warn(xe,
> +					      "Aborting migration, revid mismatch %#06x!=%#06x\n",
> +					      *klvs, xe->info.revid);
> +				return -ENODEV;
> +			}
> +			break;
> +		default:
> +			xe_sriov_dbg(xe,
> +				     "Skipping unknown migration descriptor key %#06x (len=%#06x)\n",
> +				     key, len);
> +			print_hex_dump_bytes("desc: ", DUMP_PREFIX_OFFSET, klvs,
> +					     min(SZ_64, len * sizeof(u32)));
> +			break;
> +		}
> +
> +		klvs += len;
> +		num_dwords -= len;
> +	}
> +
> +	return 0;
> +}
> +
>  static void pf_pending_init(struct xe_device *xe, unsigned int vfid)
>  {
>  	struct xe_sriov_migration_data **data = pf_pick_pending(xe, vfid);
> diff --git a/drivers/gpu/drm/xe/xe_sriov_migration_data.h b/drivers/gpu/drm/xe/xe_sriov_migration_data.h
> index 7ec489c3f28d2..bb4ea5850e5c0 100644
> --- a/drivers/gpu/drm/xe/xe_sriov_migration_data.h
> +++ b/drivers/gpu/drm/xe/xe_sriov_migration_data.h
> @@ -30,6 +30,8 @@ ssize_t xe_sriov_migration_data_read(struct xe_device *xe, unsigned int vfid,
>  				     char __user *buf, size_t len);
>  ssize_t xe_sriov_migration_data_write(struct xe_device *xe, unsigned int vfid,
>  				      const char __user *buf, size_t len);
> +int xe_sriov_migration_data_process_descriptor(struct xe_device *xe, unsigned int vfid,
> +					       struct xe_sriov_migration_data *data);
>  int xe_sriov_migration_data_save_init(struct xe_device *xe, unsigned int vfid);
>  
>  #endif
> diff --git a/drivers/gpu/drm/xe/xe_sriov_pf_migration.c b/drivers/gpu/drm/xe/xe_sriov_pf_migration.c
> index 8ea531d36f53b..f0a0c2b027a20 100644
> --- a/drivers/gpu/drm/xe/xe_sriov_pf_migration.c
> +++ b/drivers/gpu/drm/xe/xe_sriov_pf_migration.c
> @@ -175,9 +175,15 @@ xe_sriov_pf_migration_save_consume(struct xe_device *xe, unsigned int vfid)
>  static int pf_handle_descriptor(struct xe_device *xe, unsigned int vfid,
>  				struct xe_sriov_migration_data *data)
>  {
> +	int ret;
> +
>  	if (data->tile != 0 || data->gt != 0)
>  		return -EINVAL;
>  
> +	ret = xe_sriov_migration_data_process_descriptor(xe, vfid, data);
> +	if (ret)
> +		return ret;
> +
>  	xe_sriov_migration_data_free(data);
>  
>  	return 0;

just nit, so

Reviewed-by: Michal Wajdeczko <michal.wajdeczko@intel.com>

Re: [PATCH v3 12/28] drm/xe/pf: Increase PF GuC Buffer Cache size and use it for VF migration

[Michal Wajdeczko]

On 10/30/2025 9:31 PM, Michał Winiarski wrote:
> Contiguous PF GGTT VMAs can be scarce after creating VFs.
> Increase the GuC buffer cache size to 8M for PF so that we can fit GuC
> migration data (which currently maxes out at just under 4M) and use the
> cache instead of allocating fresh BOs.
> 
> Signed-off-by: Michał Winiarski <michal.winiarski@intel.com>
> ---
>  drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.c | 46 ++++++-------------
>  drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.h |  3 ++
>  drivers/gpu/drm/xe/xe_guc.c                   | 12 ++++-
>  3 files changed, 28 insertions(+), 33 deletions(-)
> 
> diff --git a/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.c b/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.c
> index 30f0e51da49ae..a2db127982638 100644
> --- a/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.c
> +++ b/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.c
> @@ -11,7 +11,7 @@
>  #include "xe_gt_sriov_pf_helpers.h"
>  #include "xe_gt_sriov_pf_migration.h"
>  #include "xe_gt_sriov_printk.h"
> -#include "xe_guc.h"
> +#include "xe_guc_buf.h"
>  #include "xe_guc_ct.h"
>  #include "xe_sriov.h"
>  #include "xe_sriov_migration_data.h"
> @@ -57,73 +57,55 @@ static int pf_send_guc_query_vf_state_size(struct xe_gt *gt, unsigned int vfid)
>  
>  /* Return: number of state dwords saved or a negative error code on failure */
>  static int pf_send_guc_save_vf_state(struct xe_gt *gt, unsigned int vfid,
> -				     void *buff, size_t size)
> +				     void *dst, size_t size)
>  {
>  	const int ndwords = size / sizeof(u32);
> -	struct xe_tile *tile = gt_to_tile(gt);
> -	struct xe_device *xe = tile_to_xe(tile);
>  	struct xe_guc *guc = &gt->uc.guc;
> -	struct xe_bo *bo;
> +	CLASS(xe_guc_buf, buf)(&guc->buf, ndwords);
>  	int ret;
>  
>  	xe_gt_assert(gt, size % sizeof(u32) == 0);
>  	xe_gt_assert(gt, size == ndwords * sizeof(u32));
>  
> -	bo = xe_bo_create_pin_map_novm(xe, tile,
> -				       ALIGN(size, PAGE_SIZE),
> -				       ttm_bo_type_kernel,
> -				       XE_BO_FLAG_SYSTEM |
> -				       XE_BO_FLAG_GGTT |
> -				       XE_BO_FLAG_GGTT_INVALIDATE, false);
> -	if (IS_ERR(bo))
> -		return PTR_ERR(bo);
> +	if (!xe_guc_buf_is_valid(buf))
> +		return -ENOBUFS;
> +

nit: I would still add the comment that this to WA some known FW limitation

> +	memset(xe_guc_buf_cpu_ptr(buf), 0, size);
>  
>  	ret = guc_action_vf_save_restore(guc, vfid, GUC_PF_OPCODE_VF_SAVE,
> -					 xe_bo_ggtt_addr(bo), ndwords);
> +					 xe_guc_buf_flush(buf), ndwords);
>  	if (!ret)
>  		ret = -ENODATA;
>  	else if (ret > ndwords)
>  		ret = -EPROTO;
>  	else if (ret > 0)
> -		xe_map_memcpy_from(xe, buff, &bo->vmap, 0, ret * sizeof(u32));
> +		memcpy(dst, xe_guc_buf_sync_read(buf), ret * sizeof(u32));
>  
> -	xe_bo_unpin_map_no_vm(bo);
>  	return ret;
>  }
>  
>  /* Return: number of state dwords restored or a negative error code on failure */
>  static int pf_send_guc_restore_vf_state(struct xe_gt *gt, unsigned int vfid,
> -					const void *buff, size_t size)
> +					const void *src, size_t size)
>  {
>  	const int ndwords = size / sizeof(u32);
> -	struct xe_tile *tile = gt_to_tile(gt);
> -	struct xe_device *xe = tile_to_xe(tile);
>  	struct xe_guc *guc = &gt->uc.guc;
> -	struct xe_bo *bo;
> +	CLASS(xe_guc_buf_from_data, buf)(&guc->buf, src, size);
>  	int ret;
>  
>  	xe_gt_assert(gt, size % sizeof(u32) == 0);
>  	xe_gt_assert(gt, size == ndwords * sizeof(u32));
>  
> -	bo = xe_bo_create_pin_map_novm(xe, tile,
> -				       ALIGN(size, PAGE_SIZE),
> -				       ttm_bo_type_kernel,
> -				       XE_BO_FLAG_SYSTEM |
> -				       XE_BO_FLAG_GGTT |
> -				       XE_BO_FLAG_GGTT_INVALIDATE, false);
> -	if (IS_ERR(bo))
> -		return PTR_ERR(bo);
> -
> -	xe_map_memcpy_to(xe, &bo->vmap, 0, buff, size);
> +	if (!xe_guc_buf_is_valid(buf))
> +		return -ENOBUFS;
>  
>  	ret = guc_action_vf_save_restore(guc, vfid, GUC_PF_OPCODE_VF_RESTORE,
> -					 xe_bo_ggtt_addr(bo), ndwords);
> +					 xe_guc_buf_flush(buf), ndwords);
>  	if (!ret)
>  		ret = -ENODATA;
>  	else if (ret > ndwords)
>  		ret = -EPROTO;
>  
> -	xe_bo_unpin_map_no_vm(bo);
>  	return ret;
>  }
>  
> diff --git a/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.h b/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.h
> index e2d41750f863c..4f2f2783339c3 100644
> --- a/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.h
> +++ b/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.h
> @@ -11,6 +11,9 @@
>  struct xe_gt;
>  struct xe_sriov_migration_data;
>  
> +/* TODO: get this information by querying GuC in the future */
> +#define XE_GT_SRIOV_PF_MIGRATION_GUC_DATA_MAX_SIZE SZ_8M
> +
>  int xe_gt_sriov_pf_migration_init(struct xe_gt *gt);
>  int xe_gt_sriov_pf_migration_save_guc_state(struct xe_gt *gt, unsigned int vfid);
>  int xe_gt_sriov_pf_migration_restore_guc_state(struct xe_gt *gt, unsigned int vfid);
> diff --git a/drivers/gpu/drm/xe/xe_guc.c b/drivers/gpu/drm/xe/xe_guc.c
> index ecc3e091b89e6..badae9a26220e 100644
> --- a/drivers/gpu/drm/xe/xe_guc.c
> +++ b/drivers/gpu/drm/xe/xe_guc.c
> @@ -24,6 +24,7 @@
>  #include "xe_gt_printk.h"
>  #include "xe_gt_sriov_vf.h"
>  #include "xe_gt_throttle.h"
> +#include "xe_gt_sriov_pf_migration.h"
>  #include "xe_guc_ads.h"
>  #include "xe_guc_buf.h"
>  #include "xe_guc_capture.h"
> @@ -40,6 +41,7 @@
>  #include "xe_mmio.h"
>  #include "xe_platform_types.h"
>  #include "xe_sriov.h"
> +#include "xe_sriov_pf_migration.h"
>  #include "xe_uc.h"
>  #include "xe_uc_fw.h"
>  #include "xe_wa.h"
> @@ -821,6 +823,14 @@ static int vf_guc_init_post_hwconfig(struct xe_guc *guc)
>  	return 0;
>  }
>  
> +static u32 guc_additional_cache_size(struct xe_device *xe)
> +{
> +	if (IS_SRIOV_PF(xe) && xe_sriov_pf_migration_supported(xe))
> +		return XE_GT_SRIOV_PF_MIGRATION_GUC_DATA_MAX_SIZE;
> +	else
> +		return 0; /* Fallback to default size */
> +}
> +
>  /**
>   * xe_guc_init_post_hwconfig - initialize GuC post hwconfig load
>   * @guc: The GuC object
> @@ -860,7 +870,7 @@ int xe_guc_init_post_hwconfig(struct xe_guc *guc)
>  	if (ret)
>  		return ret;
>  
> -	ret = xe_guc_buf_cache_init(&guc->buf);
> +	ret = xe_guc_buf_cache_init_with_size(&guc->buf, guc_additional_cache_size(guc_to_xe(guc)));

nit: quite long line ... either split, or choose shorter name for the helper
>  	if (ret)
>  		return ret;
>  

just nits, otherwise LGTM,

Reviewed-by: Michal Wajdeczko <michal.wajdeczko@intel.com>

Re: [PATCH v3 17/28] drm/xe/pf: Add helpers for VF GGTT migration data handling

[Michal Wajdeczko]

On 10/30/2025 9:31 PM, Michał Winiarski wrote:
> In an upcoming change, the VF GGTT migration data will be handled as
> part of VF control state machine. Add the necessary helpers to allow the
> migration data transfer to/from the HW GGTT resource.
> 
> Signed-off-by: Michał Winiarski <michal.winiarski@intel.com>
> ---
>  drivers/gpu/drm/xe/xe_ggtt.c               | 104 +++++++++++++++++++++
>  drivers/gpu/drm/xe/xe_ggtt.h               |   4 +
>  drivers/gpu/drm/xe/xe_ggtt_types.h         |   2 +
>  drivers/gpu/drm/xe/xe_gt_sriov_pf_config.c |  52 +++++++++++
>  drivers/gpu/drm/xe/xe_gt_sriov_pf_config.h |   5 +
>  5 files changed, 167 insertions(+)
> 
> diff --git a/drivers/gpu/drm/xe/xe_ggtt.c b/drivers/gpu/drm/xe/xe_ggtt.c
> index 20d226d90c50f..00ddb4f013466 100644
> --- a/drivers/gpu/drm/xe/xe_ggtt.c
> +++ b/drivers/gpu/drm/xe/xe_ggtt.c
> @@ -151,6 +151,14 @@ static void xe_ggtt_set_pte_and_flush(struct xe_ggtt *ggtt, u64 addr, u64 pte)
>  	ggtt_update_access_counter(ggtt);
>  }
>  
> +static u64 xe_ggtt_get_pte(struct xe_ggtt *ggtt, u64 addr)
> +{
> +	xe_tile_assert(ggtt->tile, !(addr & XE_PTE_MASK));
> +	xe_tile_assert(ggtt->tile, addr < ggtt->size);
> +
> +	return readq(&ggtt->gsm[addr >> XE_PTE_SHIFT]);
> +}
> +
>  static void xe_ggtt_clear(struct xe_ggtt *ggtt, u64 start, u64 size)
>  {
>  	u16 pat_index = tile_to_xe(ggtt->tile)->pat.idx[XE_CACHE_WB];
> @@ -233,16 +241,19 @@ void xe_ggtt_might_lock(struct xe_ggtt *ggtt)
>  static const struct xe_ggtt_pt_ops xelp_pt_ops = {
>  	.pte_encode_flags = xelp_ggtt_pte_flags,
>  	.ggtt_set_pte = xe_ggtt_set_pte,
> +	.ggtt_get_pte = xe_ggtt_get_pte,
>  };
>  
>  static const struct xe_ggtt_pt_ops xelpg_pt_ops = {
>  	.pte_encode_flags = xelpg_ggtt_pte_flags,
>  	.ggtt_set_pte = xe_ggtt_set_pte,
> +	.ggtt_get_pte = xe_ggtt_get_pte,
>  };
>  
>  static const struct xe_ggtt_pt_ops xelpg_pt_wa_ops = {
>  	.pte_encode_flags = xelpg_ggtt_pte_flags,
>  	.ggtt_set_pte = xe_ggtt_set_pte_and_flush,
> +	.ggtt_get_pte = xe_ggtt_get_pte,
>  };
>  
>  static void __xe_ggtt_init_early(struct xe_ggtt *ggtt, u32 reserved)
> @@ -889,6 +900,20 @@ u64 xe_ggtt_largest_hole(struct xe_ggtt *ggtt, u64 alignment, u64 *spare)
>  	return max_hole;
>  }
>  
> +/**
> + * xe_ggtt_node_pt_size() - Convert GGTT node size to its page table entries size.
> + * @node: the &xe_ggtt_node
> + *
> + * Return: GGTT node page table entries size in bytes.
> + */
> +size_t xe_ggtt_node_pt_size(const struct xe_ggtt_node *node)
> +{
> +	if (!node)
> +		return 0;
> +
> +	return node->base.size / XE_PAGE_SIZE * sizeof(u64);
> +}
> +
>  #ifdef CONFIG_PCI_IOV
>  static u64 xe_encode_vfid_pte(u16 vfid)
>  {
> @@ -930,6 +955,85 @@ void xe_ggtt_assign(const struct xe_ggtt_node *node, u16 vfid)
>  	xe_ggtt_assign_locked(node->ggtt, &node->base, vfid);
>  	mutex_unlock(&node->ggtt->lock);
>  }
> +
> +/**
> + * xe_ggtt_node_save() - Save a &xe_ggtt_node to a buffer.
> + * @node: the &xe_ggtt_node to be saved
> + * @dst: destination buffer
> + * @size: destination buffer size in bytes
> + * @vfid: VF identifier
> + *
> + * Return: 0 on success or a negative error code on failure.
> + */
> +int xe_ggtt_node_save(struct xe_ggtt_node *node, void *dst, size_t size, u16 vfid)
> +{
> +	struct xe_ggtt *ggtt;
> +	u64 start, end;
> +	u64 *buf = dst;
> +	u64 pte;
> +
> +	if (!node)
> +		return -ENOENT;
> +
> +	guard(mutex)(&node->ggtt->lock);
> +
> +	ggtt = node->ggtt;
> +	start = node->base.start;
> +	end = start + size - 1;

as you've already noticed, this 'size' (PT size) is the right size to calc the end

> +
> +	if (xe_ggtt_node_pt_size(node) != size)
> +		return -EINVAL;
> +
> +	while (start < end) {
> +		pte = ggtt->pt_ops->ggtt_get_pte(ggtt, start);
> +		if (vfid != u64_get_bits(pte, GGTT_PTE_VFID))
> +			return -EPERM;
> +
> +		*buf++ = u64_replace_bits(pte, 0, GGTT_PTE_VFID);
> +		start += XE_PAGE_SIZE;
> +	}
> +
> +	return 0;
> +}
> +
> +/**
> + * xe_ggtt_node_load() - Load a &xe_ggtt_node from a buffer.
> + * @node: the &xe_ggtt_node to be loaded
> + * @src: source buffer
> + * @size: source buffer size in bytes
> + * @vfid: VF identifier
> + *
> + * Return: 0 on success or a negative error code on failure.
> + */
> +int xe_ggtt_node_load(struct xe_ggtt_node *node, const void *src, size_t size, u16 vfid)
> +{
> +	u64 vfid_pte = xe_encode_vfid_pte(vfid);
> +	const u64 *buf = src;
> +	struct xe_ggtt *ggtt;
> +	u64 start, end;
> +
> +	if (!node)
> +		return -ENOENT;
> +
> +	guard(mutex)(&node->ggtt->lock);
> +
> +	ggtt = node->ggtt;
> +	start = node->base.start;
> +	end = start + size - 1;

ditto

> +
> +	if (xe_ggtt_node_pt_size(node) != size)
> +		return -EINVAL;
> +
> +	while (start < end) {
> +		vfid_pte = u64_replace_bits(*buf++, vfid, GGTT_PTE_VFID);
> +		ggtt->pt_ops->ggtt_set_pte(ggtt, start, vfid_pte);
> +		start += XE_PAGE_SIZE;
> +	}
> +	xe_ggtt_invalidate(ggtt);
> +
> +	return 0;
> +}
> +
>  #endif
>  
>  /**
> diff --git a/drivers/gpu/drm/xe/xe_ggtt.h b/drivers/gpu/drm/xe/xe_ggtt.h
> index 75fc7a1efea76..1edf27608d39a 100644
> --- a/drivers/gpu/drm/xe/xe_ggtt.h
> +++ b/drivers/gpu/drm/xe/xe_ggtt.h
> @@ -41,8 +41,12 @@ u64 xe_ggtt_largest_hole(struct xe_ggtt *ggtt, u64 alignment, u64 *spare);
>  int xe_ggtt_dump(struct xe_ggtt *ggtt, struct drm_printer *p);
>  u64 xe_ggtt_print_holes(struct xe_ggtt *ggtt, u64 alignment, struct drm_printer *p);
>  
> +size_t xe_ggtt_node_pt_size(const struct xe_ggtt_node *node);
> +
>  #ifdef CONFIG_PCI_IOV
>  void xe_ggtt_assign(const struct xe_ggtt_node *node, u16 vfid);
> +int xe_ggtt_node_save(struct xe_ggtt_node *node, void *dst, size_t size, u16 vfid);
> +int xe_ggtt_node_load(struct xe_ggtt_node *node, const void *src, size_t size, u16 vfid);
>  #endif
>  
>  #ifndef CONFIG_LOCKDEP
> diff --git a/drivers/gpu/drm/xe/xe_ggtt_types.h b/drivers/gpu/drm/xe/xe_ggtt_types.h
> index c5e999d58ff2a..dacd796f81844 100644
> --- a/drivers/gpu/drm/xe/xe_ggtt_types.h
> +++ b/drivers/gpu/drm/xe/xe_ggtt_types.h
> @@ -78,6 +78,8 @@ struct xe_ggtt_pt_ops {
>  	u64 (*pte_encode_flags)(struct xe_bo *bo, u16 pat_index);
>  	/** @ggtt_set_pte: Directly write into GGTT's PTE */
>  	void (*ggtt_set_pte)(struct xe_ggtt *ggtt, u64 addr, u64 pte);
> +	/** @ggtt_get_pte: Directly read from GGTT's PTE */
> +	u64 (*ggtt_get_pte)(struct xe_ggtt *ggtt, u64 addr);
>  };
>  
>  #endif
> diff --git a/drivers/gpu/drm/xe/xe_gt_sriov_pf_config.c b/drivers/gpu/drm/xe/xe_gt_sriov_pf_config.c
> index c0c0215c07036..55444883f2ac3 100644
> --- a/drivers/gpu/drm/xe/xe_gt_sriov_pf_config.c
> +++ b/drivers/gpu/drm/xe/xe_gt_sriov_pf_config.c
> @@ -726,6 +726,58 @@ int xe_gt_sriov_pf_config_set_fair_ggtt(struct xe_gt *gt, unsigned int vfid,
>  	return xe_gt_sriov_pf_config_bulk_set_ggtt(gt, vfid, num_vfs, fair);
>  }
>  
> +/**
> + * xe_gt_sriov_pf_config_ggtt_save() - Save a VF provisioned GGTT data into a buffer.
> + * @gt: the &xe_gt
> + * @vfid: VF identifier (can't be 0)
> + * @buf: the GGTT data destination buffer (or NULL to query the buf size)
> + * @size: the size of the buffer (or 0 to query the buf size)
> + *
> + * This function can only be called on PF.
> + *
> + * Return: size of the buffer needed to save GGTT data if querying,
> + *         0 on successful save or a negative error code on failure.
> + */
> +ssize_t xe_gt_sriov_pf_config_ggtt_save(struct xe_gt *gt, unsigned int vfid,
> +					void *buf, size_t size)
> +{
> +	struct xe_ggtt_node *node;
> +
> +	xe_gt_assert(gt, IS_SRIOV_PF(gt_to_xe(gt)));
> +	xe_gt_assert(gt, vfid);
> +	xe_gt_assert(gt, !(!buf ^ !size));
> +
> +	guard(mutex)(xe_gt_sriov_pf_master_mutex(gt));
> +	node = pf_pick_vf_config(gt, vfid)->ggtt_region;
> +
> +	if (!buf)
> +		return xe_ggtt_node_pt_size(node);
> +
> +	return xe_ggtt_node_save(node, buf, size, vfid);
> +}
> +
> +/**
> + * xe_gt_sriov_pf_config_ggtt_restore() - Restore a VF provisioned GGTT data from a buffer.
> + * @gt: the &xe_gt
> + * @vfid: VF identifier (can't be 0)
> + * @buf: the GGTT data source buffer
> + * @size: the size of the buffer
> + *
> + * This function can only be called on PF.
> + *
> + * Return: 0 on success or a negative error code on failure.
> + */
> +int xe_gt_sriov_pf_config_ggtt_restore(struct xe_gt *gt, unsigned int vfid,
> +				       const void *buf, size_t size)
> +{
> +	xe_gt_assert(gt, IS_SRIOV_PF(gt_to_xe(gt)));
> +	xe_gt_assert(gt, vfid);
> +
> +	guard(mutex)(xe_gt_sriov_pf_master_mutex(gt));
> +
> +	return xe_ggtt_node_load(pf_pick_vf_config(gt, vfid)->ggtt_region, buf, size, vfid);
> +}
> +
>  static u32 pf_get_min_spare_ctxs(struct xe_gt *gt)
>  {
>  	/* XXX: preliminary */
> diff --git a/drivers/gpu/drm/xe/xe_gt_sriov_pf_config.h b/drivers/gpu/drm/xe/xe_gt_sriov_pf_config.h
> index 513e6512a575b..0293ba98eb6df 100644
> --- a/drivers/gpu/drm/xe/xe_gt_sriov_pf_config.h
> +++ b/drivers/gpu/drm/xe/xe_gt_sriov_pf_config.h
> @@ -61,6 +61,11 @@ ssize_t xe_gt_sriov_pf_config_save(struct xe_gt *gt, unsigned int vfid, void *bu
>  int xe_gt_sriov_pf_config_restore(struct xe_gt *gt, unsigned int vfid,
>  				  const void *buf, size_t size);
>  
> +ssize_t xe_gt_sriov_pf_config_ggtt_save(struct xe_gt *gt, unsigned int vfid,
> +					void *buf, size_t size);
> +int xe_gt_sriov_pf_config_ggtt_restore(struct xe_gt *gt, unsigned int vfid,
> +				       const void *buf, size_t size);
> +
>  bool xe_gt_sriov_pf_config_is_empty(struct xe_gt *gt, unsigned int vfid);
>  
>  int xe_gt_sriov_pf_config_init(struct xe_gt *gt);

Re: [PATCH v3 24/28] drm/xe/pf: Enable SR-IOV VF migration

[Michal Wajdeczko]

On 10/30/2025 9:31 PM, Michał Winiarski wrote:
> All of the necessary building blocks are now in place to support SR-IOV
> VF migration.
> Enable the feature without the need to pass feature enabling debug flags
> for those platforms and rely on .has_memirq presence instead (like with
> VF resource fixup).
> 
> Signed-off-by: Michał Winiarski <michal.winiarski@intel.com>
> ---
>  drivers/gpu/drm/xe/xe_sriov_pf_migration.c | 6 ++++--
>  1 file changed, 4 insertions(+), 2 deletions(-)
> 
> diff --git a/drivers/gpu/drm/xe/xe_sriov_pf_migration.c b/drivers/gpu/drm/xe/xe_sriov_pf_migration.c
> index 6992c227e5a44..c4e9b0ff5b3ae 100644
> --- a/drivers/gpu/drm/xe/xe_sriov_pf_migration.c
> +++ b/drivers/gpu/drm/xe/xe_sriov_pf_migration.c
> @@ -50,8 +50,10 @@ bool xe_sriov_pf_migration_supported(struct xe_device *xe)
>  
>  static bool pf_check_migration_support(struct xe_device *xe)
>  {
> -	/* XXX: for now this is for feature enabling only */
> -	return IS_ENABLED(CONFIG_DRM_XE_DEBUG);
> +	if (xe_device_has_memirq(xe))
> +		return true;
> +
> +	return false;

maybe simpler:

	/* all platforms with MEMIRQ should work */
	return xe_device_has_memirq(xe)

and shouldn't we also check somewhere that we have the right GuC FW?
here is likely too early, but maybe we need additional check in:

	xe_sriov_pf_init_late(xe) ?

>  }
>  
>  static void pf_migration_cleanup(void *arg)

Re: [PATCH v3 25/28] drm/xe/pci: Introduce a helper to allow VF access to PF xe_device

[Michal Wajdeczko]

On 10/30/2025 9:31 PM, Michał Winiarski wrote:
> In certain scenarios (such as VF migration), VF driver needs to interact
> with PF driver.
> Add a helper to allow VF driver access to PF xe_device.
> 
> Signed-off-by: Michał Winiarski <michal.winiarski@intel.com>
> ---
>  drivers/gpu/drm/xe/xe_pci.c | 17 +++++++++++++++++
>  drivers/gpu/drm/xe/xe_pci.h |  3 +++
>  2 files changed, 20 insertions(+)
> 
> diff --git a/drivers/gpu/drm/xe/xe_pci.c b/drivers/gpu/drm/xe/xe_pci.c
> index aeae675c912b7..3e7b03c847a42 100644
> --- a/drivers/gpu/drm/xe/xe_pci.c
> +++ b/drivers/gpu/drm/xe/xe_pci.c
> @@ -1224,6 +1224,23 @@ static struct pci_driver xe_pci_driver = {
>  #endif
>  };
>  
> +/**
> + * xe_pci_get_pf_xe_device() - Get PF &xe_device.
> + * @pdev: the VF &pci_dev device
> + *
> + * Return: pointer to PF &xe_device, NULL otherwise.
> + */
> +struct xe_device *xe_pci_get_pf_xe_device(struct pci_dev *pdev)
> +{
> +	struct drm_device *drm;
> +
> +	drm = pci_iov_get_pf_drvdata(pdev, &xe_pci_driver);
> +	if (IS_ERR(drm))
> +		return NULL;
> +
> +	return to_xe_device(drm);
> +}

I would rather expose xe_pci_driver from xe_pci.c (*) and then
make this a static helper as part of the vfio export file

	static struct xe_device *to_xe_pf(struct pci_dev *pdev)

but maybe Lucas or Matt has a different opinion

(*) this will also help us to move our live-kunit generator to a better place 


> +
>  int xe_register_pci_driver(void)
>  {
>  	return pci_register_driver(&xe_pci_driver);
> diff --git a/drivers/gpu/drm/xe/xe_pci.h b/drivers/gpu/drm/xe/xe_pci.h
> index 611c1209b14cc..2bb2e486756db 100644
> --- a/drivers/gpu/drm/xe/xe_pci.h
> +++ b/drivers/gpu/drm/xe/xe_pci.h
> @@ -6,6 +6,9 @@
>  #ifndef _XE_PCI_H_
>  #define _XE_PCI_H_
>  
> +struct pci_dev;
> +
> +struct xe_device *xe_pci_get_pf_xe_device(struct pci_dev *pdev);
>  int xe_register_pci_driver(void);
>  void xe_unregister_pci_driver(void);
>  

Re: [PATCH v3 16/28] drm/xe/pf: Handle GuC migration data as part of PF control

[Michal Wajdeczko]

On 10/30/2025 9:31 PM, Michał Winiarski wrote:
> Connect the helpers to allow save and restore of GuC migration data in
> stop_copy / resume device state.
> 
> Signed-off-by: Michał Winiarski <michal.winiarski@intel.com>
> ---
>  drivers/gpu/drm/xe/xe_gt_sriov_pf_control.c   | 25 ++++++-
>  drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.c | 66 +++++++++++++++----
>  drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.h |  8 ++-
>  .../drm/xe/xe_gt_sriov_pf_migration_types.h   |  5 ++
>  4 files changed, 87 insertions(+), 17 deletions(-)
> 
> diff --git a/drivers/gpu/drm/xe/xe_gt_sriov_pf_control.c b/drivers/gpu/drm/xe/xe_gt_sriov_pf_control.c
> index 203ba1fbeefcd..cb45e97f4c4d9 100644
> --- a/drivers/gpu/drm/xe/xe_gt_sriov_pf_control.c
> +++ b/drivers/gpu/drm/xe/xe_gt_sriov_pf_control.c
> @@ -848,6 +848,18 @@ static void pf_enter_vf_save_failed(struct xe_gt *gt, unsigned int vfid)
>  
>  static int pf_handle_vf_save_data(struct xe_gt *gt, unsigned int vfid)
>  {
> +	int ret;
> +
> +	if (xe_gt_sriov_pf_migration_save_test(gt, vfid, XE_SRIOV_MIGRATION_DATA_TYPE_GUC)) {
> +		ret = xe_gt_sriov_pf_migration_guc_save(gt, vfid);
> +		if (ret)
> +			return ret;
> +
> +		xe_gt_sriov_pf_migration_save_clear(gt, vfid, XE_SRIOV_MIGRATION_DATA_TYPE_GUC);
> +
> +		return -EAGAIN;
> +	}
> +
>  	return 0;
>  }
>  
> @@ -887,6 +899,7 @@ static void pf_exit_vf_save_wait_data(struct xe_gt *gt, unsigned int vfid)
>  static bool pf_enter_vf_save_wip(struct xe_gt *gt, unsigned int vfid)
>  {
>  	if (pf_enter_vf_state(gt, vfid, XE_GT_SRIOV_STATE_SAVE_WIP)) {
> +		xe_gt_sriov_pf_migration_save_init(gt, vfid);
>  		pf_enter_vf_state(gt, vfid, XE_GT_SRIOV_STATE_SAVE_PROCESS_DATA);
>  		pf_enter_vf_wip(gt, vfid);
>  		pf_queue_vf(gt, vfid);
> @@ -1048,14 +1061,22 @@ static void pf_enter_vf_restore_failed(struct xe_gt *gt, unsigned int vfid)
>  static int pf_handle_vf_restore_data(struct xe_gt *gt, unsigned int vfid)
>  {
>  	struct xe_sriov_migration_data *data = xe_gt_sriov_pf_migration_restore_consume(gt, vfid);
> +	int ret = 0;
>  
>  	xe_gt_assert(gt, data);
>  
> -	xe_gt_sriov_notice(gt, "Skipping VF%u unknown data type: %d\n", vfid, data->type);
> +	switch (data->type) {
> +	case XE_SRIOV_MIGRATION_DATA_TYPE_GUC:
> +		ret = xe_gt_sriov_pf_migration_guc_restore(gt, vfid, data);
> +		break;
> +	default:
> +		xe_gt_sriov_notice(gt, "Skipping VF%u unknown data type: %d\n", vfid, data->type);
> +		break;
> +	}
>  
>  	xe_sriov_migration_data_free(data);
>  
> -	return 0;
> +	return ret;
>  }
>  
>  static bool pf_handle_vf_restore(struct xe_gt *gt, unsigned int vfid)
> diff --git a/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.c b/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.c
> index 4a716e0a29fe4..fbb451767712a 100644
> --- a/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.c
> +++ b/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.c
> @@ -174,22 +174,10 @@ static int pf_save_vf_guc_mig_data(struct xe_gt *gt, unsigned int vfid)
>  	return ret;
>  }
>  
> -/**
> - * xe_gt_sriov_pf_migration_guc_size() - Get the size of VF GuC migration data.
> - * @gt: the &xe_gt
> - * @vfid: the VF identifier
> - *
> - * This function is for PF only.
> - *
> - * Return: size in bytes or a negative error code on failure.
> - */
> -ssize_t xe_gt_sriov_pf_migration_guc_size(struct xe_gt *gt, unsigned int vfid)
> +static ssize_t pf_migration_guc_size(struct xe_gt *gt, unsigned int vfid)
>  {
>  	ssize_t size;
>  
> -	xe_gt_assert(gt, IS_SRIOV_PF(gt_to_xe(gt)));
> -	xe_gt_assert(gt, vfid != PFID);
> -	xe_gt_assert(gt, vfid <= xe_sriov_pf_get_totalvfs(gt_to_xe(gt)));
>  
>  	if (!pf_migration_supported(gt))
>  		return -ENOPKG;
> @@ -278,12 +266,19 @@ int xe_gt_sriov_pf_migration_guc_restore(struct xe_gt *gt, unsigned int vfid,
>  ssize_t xe_gt_sriov_pf_migration_size(struct xe_gt *gt, unsigned int vfid)
>  {
>  	ssize_t total = 0;
> +	ssize_t size;
>  
>  	xe_gt_assert(gt, IS_SRIOV_PF(gt_to_xe(gt)));
>  	xe_gt_assert(gt, vfid != PFID);
>  	xe_gt_assert(gt, vfid <= xe_sriov_pf_get_totalvfs(gt_to_xe(gt)));
>  
> -	/* Nothing to query yet - will be updated once per-GT migration data types are added */
> +	size = pf_migration_guc_size(gt, vfid);
> +	if (size < 0)
> +		return size;
> +	if (size > 0)
> +		size += sizeof(struct xe_sriov_pf_migration_hdr);
> +	total += size;
> +
>  	return total;
>  }
>  
> @@ -330,6 +325,49 @@ void xe_gt_sriov_pf_migration_ring_free(struct xe_gt *gt, unsigned int vfid)
>  		xe_sriov_migration_data_free(data);
>  }
>  
> +/**
> + * xe_gt_sriov_pf_migration_save_init() - Initialize per-GT migration related data.
> + * @gt: the &xe_gt
> + * @vfid: the VF identifier (can't be 0)
> + *
> + * Return: 0 on success or a negative error code on failure.

but it's void

> + */
> +void xe_gt_sriov_pf_migration_save_init(struct xe_gt *gt, unsigned int vfid)
> +{
> +	struct xe_gt_sriov_migration_data *migration = pf_pick_gt_migration(gt, vfid);
> +
> +	migration->save.data_remaining = 0;
> +
> +	xe_gt_assert(gt, pf_migration_guc_size(gt, vfid) > 0);
> +	set_bit(XE_SRIOV_MIGRATION_DATA_TYPE_GUC, &migration->save.data_remaining);
> +}
> +
> +/**
> + * xe_gt_sriov_pf_migration_save_test() - Test if migration data type needs to be saved.
> + * @gt: the &xe_gt
> + * @vfid: the VF identifier (can't be 0)
> + * @type: the &xe_sriov_migration_data_type of data to be saved

.. to check

> + *
> + * Return: true if the data needs saving, otherwise false.
> + */
> +bool xe_gt_sriov_pf_migration_save_test(struct xe_gt *gt, unsigned int vfid,
> +					enum xe_sriov_migration_data_type type)
> +{
> +	return test_bit(type, &pf_pick_gt_migration(gt, vfid)->save.data_remaining);
> +}
> +
> +/**
> + * xe_gt_sriov_pf_migration_save_clear() - Clear migration data type from saving.
> + * @gt: the &xe_gt
> + * @vfid: the VF identifier (can't be 0)
> + * @type: the &xe_sriov_migration_data_type of data to be cleared

.. to mark as done/saved

> + */
> +void xe_gt_sriov_pf_migration_save_clear(struct xe_gt *gt, unsigned int vfid,
> +					 enum xe_sriov_migration_data_type type)
> +{
> +	clear_bit(type, &pf_pick_gt_migration(gt, vfid)->save.data_remaining);
> +}
> +
>  /**
>   * xe_gt_sriov_pf_migration_save_produce() - Add VF save data packet to migration ring.
>   * @gt: the &xe_gt
> diff --git a/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.h b/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.h
> index b3c18e369df79..66a45d6234245 100644
> --- a/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.h
> +++ b/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.h
> @@ -10,12 +10,12 @@
>  
>  struct xe_gt;
>  struct xe_sriov_migration_data;
> +enum xe_sriov_migration_data_type;
>  
>  /* TODO: get this information by querying GuC in the future */
>  #define XE_GT_SRIOV_PF_MIGRATION_GUC_DATA_MAX_SIZE SZ_8M
>  
>  int xe_gt_sriov_pf_migration_init(struct xe_gt *gt);
> -ssize_t xe_gt_sriov_pf_migration_guc_size(struct xe_gt *gt, unsigned int vfid);
>  int xe_gt_sriov_pf_migration_guc_save(struct xe_gt *gt, unsigned int vfid);
>  int xe_gt_sriov_pf_migration_guc_restore(struct xe_gt *gt, unsigned int vfid,
>  					 struct xe_sriov_migration_data *data);
> @@ -26,6 +26,12 @@ bool xe_gt_sriov_pf_migration_ring_empty(struct xe_gt *gt, unsigned int vfid);
>  bool xe_gt_sriov_pf_migration_ring_full(struct xe_gt *gt, unsigned int vfid);
>  void xe_gt_sriov_pf_migration_ring_free(struct xe_gt *gt, unsigned int vfid);
>  
> +void xe_gt_sriov_pf_migration_save_init(struct xe_gt *gt, unsigned int vfid);
> +bool xe_gt_sriov_pf_migration_save_test(struct xe_gt *gt, unsigned int vfid,
> +					enum xe_sriov_migration_data_type type);
> +void xe_gt_sriov_pf_migration_save_clear(struct xe_gt *gt, unsigned int vfid,
> +					 enum xe_sriov_migration_data_type type);

nit: "clear" is little confusing, maybe

	xe_gt_sriov_pf_migration_save_data_done(gt, vfid, type);	// instead of save_clear
	xe_gt_sriov_pf_migration_save_data_pending(gt, vfid, type);	// instead of save_test
and 
	pf_migration_save_data_todo(gt, vfid, type);	// instead of set_bit


> +
>  int xe_gt_sriov_pf_migration_save_produce(struct xe_gt *gt, unsigned int vfid,
>  					  struct xe_sriov_migration_data *data);
>  struct xe_sriov_migration_data *
> diff --git a/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration_types.h b/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration_types.h
> index 75d8b94cbbefb..9f24878690d9c 100644
> --- a/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration_types.h
> +++ b/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration_types.h
> @@ -16,6 +16,11 @@
>  struct xe_gt_sriov_migration_data {
>  	/** @ring: queue containing VF save / restore migration data */
>  	struct ptr_ring ring;
> +	/** @save: structure for currently processed save migration data */
> +	struct {
> +		/** @save.data_remaining: bitmap of migration types that need to be saved */
> +		unsigned long data_remaining;
> +	} save;
>  };
>  
>  #endif

otherwise LGTM, so with kernel-doc fixed, and (hopefully) with better names,

Reviewed-by: Michal Wajdeczko <michal.wajdeczko@intel.com>

Re: [PATCH v3 18/28] drm/xe/pf: Handle GGTT migration data as part of PF control

[Michal Wajdeczko]

On 10/30/2025 9:31 PM, Michał Winiarski wrote:
> Connect the helpers to allow save and restore of GGTT migration data in
> stop_copy / resume device state.
> 
> Signed-off-by: Michał Winiarski <michal.winiarski@intel.com>
> ---
>  drivers/gpu/drm/xe/xe_gt_sriov_pf_control.c   |  13 ++
>  drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.c | 113 ++++++++++++++++++
>  drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.h |   3 +
>  3 files changed, 129 insertions(+)
> 
> diff --git a/drivers/gpu/drm/xe/xe_gt_sriov_pf_control.c b/drivers/gpu/drm/xe/xe_gt_sriov_pf_control.c
> index cb45e97f4c4d9..e7ea9b88fd246 100644
> --- a/drivers/gpu/drm/xe/xe_gt_sriov_pf_control.c
> +++ b/drivers/gpu/drm/xe/xe_gt_sriov_pf_control.c
> @@ -860,6 +860,16 @@ static int pf_handle_vf_save_data(struct xe_gt *gt, unsigned int vfid)
>  		return -EAGAIN;
>  	}
>  
> +	if (xe_gt_sriov_pf_migration_save_test(gt, vfid, XE_SRIOV_MIGRATION_DATA_TYPE_GGTT)) {
> +		ret = xe_gt_sriov_pf_migration_ggtt_save(gt, vfid);
> +		if (ret)
> +			return ret;
> +
> +		xe_gt_sriov_pf_migration_save_clear(gt, vfid, XE_SRIOV_MIGRATION_DATA_TYPE_GGTT);
> +
> +		return -EAGAIN;
> +	}
> +
>  	return 0;
>  }
>  
> @@ -1066,6 +1076,9 @@ static int pf_handle_vf_restore_data(struct xe_gt *gt, unsigned int vfid)
>  	xe_gt_assert(gt, data);
>  
>  	switch (data->type) {
> +	case XE_SRIOV_MIGRATION_DATA_TYPE_GGTT:
> +		ret = xe_gt_sriov_pf_migration_ggtt_restore(gt, vfid, data);
> +		break;
>  	case XE_SRIOV_MIGRATION_DATA_TYPE_GUC:
>  		ret = xe_gt_sriov_pf_migration_guc_restore(gt, vfid, data);
>  		break;
> diff --git a/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.c b/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.c
> index fbb451767712a..6f2ee5820bdd4 100644
> --- a/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.c
> +++ b/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.c
> @@ -7,6 +7,9 @@
>  
>  #include "abi/guc_actions_sriov_abi.h"
>  #include "xe_bo.h"
> +#include "xe_ggtt.h"
> +#include "xe_gt.h"
> +#include "xe_gt_sriov_pf_config.h"
>  #include "xe_gt_sriov_pf_control.h"
>  #include "xe_gt_sriov_pf_helpers.h"
>  #include "xe_gt_sriov_pf_migration.h"
> @@ -39,6 +42,106 @@ static void pf_dump_mig_data(struct xe_gt *gt, unsigned int vfid,
>  	}
>  }
>  
> +static ssize_t pf_migration_ggtt_size(struct xe_gt *gt, unsigned int vfid)
> +{
> +	if (!xe_gt_is_main_type(gt))
> +		return 0;
> +
> +	return xe_gt_sriov_pf_config_ggtt_save(gt, vfid, NULL, 0);
> +}
> +
> +static int pf_save_vf_ggtt_mig_data(struct xe_gt *gt, unsigned int vfid)
> +{
> +	struct xe_sriov_migration_data *data;
> +	size_t size;
> +	int ret;
> +
> +	size = pf_migration_ggtt_size(gt, vfid);
> +	xe_gt_assert(gt, size);
> +
> +	data = xe_sriov_migration_data_alloc(gt_to_xe(gt));
> +	if (!data)
> +		return -ENOMEM;
> +
> +	ret = xe_sriov_migration_data_init(data, gt->tile->id, gt->info.id,
> +					   XE_SRIOV_MIGRATION_DATA_TYPE_GGTT, 0, size);
> +	if (ret)
> +		goto fail;
> +
> +	ret = xe_gt_sriov_pf_config_ggtt_save(gt, vfid, data->vaddr, size);
> +	if (ret)
> +		goto fail;
> +
> +	xe_gt_sriov_dbg_verbose(gt, "VF%u GGTT data save (%zu bytes)\n", vfid, size);
> +	pf_dump_mig_data(gt, vfid, data);
> +
> +	ret = xe_gt_sriov_pf_migration_save_produce(gt, vfid, data);
> +	if (ret)
> +		goto fail;
> +
> +	return 0;
> +
> +fail:
> +	xe_sriov_migration_data_free(data);
> +	xe_gt_sriov_err(gt, "Failed to save VF%u GGTT data (%pe)\n", vfid, ERR_PTR(ret));
> +	return ret;
> +}
> +
> +static int pf_restore_vf_ggtt_mig_data(struct xe_gt *gt, unsigned int vfid,
> +				       struct xe_sriov_migration_data *data)
> +{
> +	int ret;
> +
> +	xe_gt_sriov_dbg_verbose(gt, "VF%u GGTT data restore (%llu bytes)\n", vfid, data->size);
> +	pf_dump_mig_data(gt, vfid, data);
> +
> +	ret = xe_gt_sriov_pf_config_ggtt_restore(gt, vfid, data->vaddr, data->size);
> +	if (ret) {
> +		xe_gt_sriov_err(gt, "Failed to restore VF%u GGTT data (%pe)\n",
> +				vfid, ERR_PTR(ret));
> +		return ret;
> +	}
> +
> +	return 0;
> +}
> +
> +/**
> + * xe_gt_sriov_pf_migration_ggtt_save() - Save VF GGTT migration data.
> + * @gt: the &xe_gt
> + * @vfid: the VF identifier (can't be 0)
> + *
> + * This function is for PF only.
> + *
> + * Return: 0 on success or a negative error code on failure.
> + */
> +int xe_gt_sriov_pf_migration_ggtt_save(struct xe_gt *gt, unsigned int vfid)
> +{
> +	xe_gt_assert(gt, IS_SRIOV_PF(gt_to_xe(gt)));
> +	xe_gt_assert(gt, vfid != PFID);
> +	xe_gt_assert(gt, vfid <= xe_sriov_pf_get_totalvfs(gt_to_xe(gt)));
> +
> +	return pf_save_vf_ggtt_mig_data(gt, vfid);
> +}
> +
> +/**
> + * xe_gt_sriov_pf_migration_ggtt_restore() - Restore VF GGTT migration data.
> + * @gt: the &xe_gt
> + * @vfid: the VF identifier (can't be 0)
> + *
> + * This function is for PF only.
> + *
> + * Return: 0 on success or a negative error code on failure.
> + */
> +int xe_gt_sriov_pf_migration_ggtt_restore(struct xe_gt *gt, unsigned int vfid,
> +					  struct xe_sriov_migration_data *data)
> +{
> +	xe_gt_assert(gt, IS_SRIOV_PF(gt_to_xe(gt)));
> +	xe_gt_assert(gt, vfid != PFID);
> +	xe_gt_assert(gt, vfid <= xe_sriov_pf_get_totalvfs(gt_to_xe(gt)));
> +
> +	return pf_restore_vf_ggtt_mig_data(gt, vfid, data);
> +}
> +
>  /* Return: number of dwords saved/restored/required or a negative error code on failure */
>  static int guc_action_vf_save_restore(struct xe_guc *guc, u32 vfid, u32 opcode,
>  				      u64 addr, u32 ndwords)
> @@ -279,6 +382,13 @@ ssize_t xe_gt_sriov_pf_migration_size(struct xe_gt *gt, unsigned int vfid)
>  		size += sizeof(struct xe_sriov_pf_migration_hdr);
>  	total += size;
>  
> +	size = pf_migration_ggtt_size(gt, vfid);
> +	if (size < 0)
> +		return size;
> +	if (size > 0)
> +		size += sizeof(struct xe_sriov_pf_migration_hdr);
> +	total += size;
> +
>  	return total;
>  }
>  
> @@ -340,6 +450,9 @@ void xe_gt_sriov_pf_migration_save_init(struct xe_gt *gt, unsigned int vfid)
>  
>  	xe_gt_assert(gt, pf_migration_guc_size(gt, vfid) > 0);
>  	set_bit(XE_SRIOV_MIGRATION_DATA_TYPE_GUC, &migration->save.data_remaining);
> +
> +	if (pf_migration_ggtt_size(gt, vfid) > 0)
> +		set_bit(XE_SRIOV_MIGRATION_DATA_TYPE_GGTT, &migration->save.data_remaining);

btw, does it make sense to save GuC data if there is no GGTT data on the main GT?

>  }
>  
>  /**
> diff --git a/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.h b/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.h
> index 66a45d6234245..bc201d8f3147a 100644
> --- a/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.h
> +++ b/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.h
> @@ -19,6 +19,9 @@ int xe_gt_sriov_pf_migration_init(struct xe_gt *gt);
>  int xe_gt_sriov_pf_migration_guc_save(struct xe_gt *gt, unsigned int vfid);
>  int xe_gt_sriov_pf_migration_guc_restore(struct xe_gt *gt, unsigned int vfid,
>  					 struct xe_sriov_migration_data *data);
> +int xe_gt_sriov_pf_migration_ggtt_save(struct xe_gt *gt, unsigned int vfid);
> +int xe_gt_sriov_pf_migration_ggtt_restore(struct xe_gt *gt, unsigned int vfid,
> +					  struct xe_sriov_migration_data *data);
>  
>  ssize_t xe_gt_sriov_pf_migration_size(struct xe_gt *gt, unsigned int vfid);
>  

only one nit question about corner case, so

Reviewed-by: Michal Wajdeczko <michal.wajdeczko@intel.com>

Re: [PATCH v3 19/28] drm/xe/pf: Handle MMIO migration data as part of PF control

[Michal Wajdeczko]

On 10/30/2025 9:31 PM, Michał Winiarski wrote:
> Implement the helpers and use them for save and restore of MMIO
> migration data in stop_copy / resume device state.
> 
> Signed-off-by: Michał Winiarski <michal.winiarski@intel.com>
> ---
>  drivers/gpu/drm/xe/xe_gt_sriov_pf.h           |   2 +
>  drivers/gpu/drm/xe/xe_gt_sriov_pf_control.c   |  13 ++
>  drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.c | 158 ++++++++++++++++++
>  drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.h |   3 +
>  4 files changed, 176 insertions(+)
> 
> diff --git a/drivers/gpu/drm/xe/xe_gt_sriov_pf.h b/drivers/gpu/drm/xe/xe_gt_sriov_pf.h
> index e7fde3f9937af..c0dcbb1054e4d 100644
> --- a/drivers/gpu/drm/xe/xe_gt_sriov_pf.h
> +++ b/drivers/gpu/drm/xe/xe_gt_sriov_pf.h
> @@ -6,6 +6,8 @@
>  #ifndef _XE_GT_SRIOV_PF_H_
>  #define _XE_GT_SRIOV_PF_H_
>  
> +#include <linux/types.h>

?

> +
>  struct xe_gt;
>  
>  #ifdef CONFIG_PCI_IOV
> diff --git a/drivers/gpu/drm/xe/xe_gt_sriov_pf_control.c b/drivers/gpu/drm/xe/xe_gt_sriov_pf_control.c
> index e7ea9b88fd246..7cd7cae950bc7 100644
> --- a/drivers/gpu/drm/xe/xe_gt_sriov_pf_control.c
> +++ b/drivers/gpu/drm/xe/xe_gt_sriov_pf_control.c
> @@ -870,6 +870,16 @@ static int pf_handle_vf_save_data(struct xe_gt *gt, unsigned int vfid)
>  		return -EAGAIN;
>  	}
>  
> +	if (xe_gt_sriov_pf_migration_save_test(gt, vfid, XE_SRIOV_MIGRATION_DATA_TYPE_MMIO)) {
> +		ret = xe_gt_sriov_pf_migration_mmio_save(gt, vfid);
> +		if (ret)
> +			return ret;
> +
> +		xe_gt_sriov_pf_migration_save_clear(gt, vfid, XE_SRIOV_MIGRATION_DATA_TYPE_MMIO);
> +
> +		return -EAGAIN;
> +	}
> +
>  	return 0;
>  }
>  
> @@ -1079,6 +1089,9 @@ static int pf_handle_vf_restore_data(struct xe_gt *gt, unsigned int vfid)
>  	case XE_SRIOV_MIGRATION_DATA_TYPE_GGTT:
>  		ret = xe_gt_sriov_pf_migration_ggtt_restore(gt, vfid, data);
>  		break;
> +	case XE_SRIOV_MIGRATION_DATA_TYPE_MMIO:
> +		ret = xe_gt_sriov_pf_migration_mmio_restore(gt, vfid, data);
> +		break;
>  	case XE_SRIOV_MIGRATION_DATA_TYPE_GUC:
>  		ret = xe_gt_sriov_pf_migration_guc_restore(gt, vfid, data);
>  		break;
> diff --git a/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.c b/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.c
> index 6f2ee5820bdd4..5e90aeafeeb41 100644
> --- a/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.c
> +++ b/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.c
> @@ -5,10 +5,13 @@
>  
>  #include <drm/drm_managed.h>
>  
> +#include "regs/xe_guc_regs.h"
> +
>  #include "abi/guc_actions_sriov_abi.h"
>  #include "xe_bo.h"
>  #include "xe_ggtt.h"
>  #include "xe_gt.h"
> +#include "xe_gt_sriov_pf.h"
>  #include "xe_gt_sriov_pf_config.h"
>  #include "xe_gt_sriov_pf_control.h"
>  #include "xe_gt_sriov_pf_helpers.h"
> @@ -16,6 +19,7 @@
>  #include "xe_gt_sriov_printk.h"
>  #include "xe_guc_buf.h"
>  #include "xe_guc_ct.h"
> +#include "xe_mmio.h"
>  #include "xe_sriov.h"
>  #include "xe_sriov_migration_data.h"
>  #include "xe_sriov_pf_migration.h"
> @@ -357,6 +361,150 @@ int xe_gt_sriov_pf_migration_guc_restore(struct xe_gt *gt, unsigned int vfid,
>  	return pf_restore_vf_guc_state(gt, vfid, data);
>  }
>  
> +static ssize_t pf_migration_mmio_size(struct xe_gt *gt, unsigned int vfid)
> +{
> +	if (xe_gt_is_media_type(gt))
> +		return MED_VF_SW_FLAG_COUNT * sizeof(u32);
> +	else
> +		return VF_SW_FLAG_COUNT * sizeof(u32);
> +}
> +
> +static int pf_migration_mmio_save(struct xe_gt *gt, unsigned int vfid, void *buf, size_t size)
> +{
> +	struct xe_mmio mmio;
> +	u32 *regs = buf;
> +	int n;
> +
> +	xe_mmio_init_vf_view(&mmio, &gt->mmio, vfid);
> +
> +	if (size != pf_migration_mmio_size(gt, vfid))
> +		return -EINVAL;

you may want to check that first (before init vf view)

> +
> +	if (xe_gt_is_media_type(gt))
> +		for (n = 0; n < MED_VF_SW_FLAG_COUNT; n++)
> +			regs[n] = xe_mmio_read32(&gt->mmio, MED_VF_SW_FLAG(n));
> +	else
> +		for (n = 0; n < VF_SW_FLAG_COUNT; n++)
> +			regs[n] = xe_mmio_read32(&gt->mmio, VF_SW_FLAG(n));
> +
> +	return 0;
> +}
> +
> +static int pf_migration_mmio_restore(struct xe_gt *gt, unsigned int vfid,
> +				     const void *buf, size_t size)
> +{
> +	const u32 *regs = buf;
> +	struct xe_mmio mmio;
> +	int n;
> +
> +	xe_mmio_init_vf_view(&mmio, &gt->mmio, vfid);
> +
> +	if (size != pf_migration_mmio_size(gt, vfid))
> +		return -EINVAL;

ditto

> +
> +	if (xe_gt_is_media_type(gt))
> +		for (n = 0; n < MED_VF_SW_FLAG_COUNT; n++)
> +			xe_mmio_write32(&gt->mmio, MED_VF_SW_FLAG(n), regs[n]);
> +	else
> +		for (n = 0; n < VF_SW_FLAG_COUNT; n++)
> +			xe_mmio_write32(&gt->mmio, VF_SW_FLAG(n), regs[n]);
> +
> +	return 0;
> +}
> +
> +static int pf_save_vf_mmio_mig_data(struct xe_gt *gt, unsigned int vfid)
> +{
> +	struct xe_sriov_migration_data *data;
> +	size_t size;
> +	int ret;
> +
> +	size = pf_migration_mmio_size(gt, vfid);
> +	xe_gt_assert(gt, size);
> +
> +	data = xe_sriov_migration_data_alloc(gt_to_xe(gt));
> +	if (!data)
> +		return -ENOMEM;
> +
> +	ret = xe_sriov_migration_data_init(data, gt->tile->id, gt->info.id,
> +					   XE_SRIOV_MIGRATION_DATA_TYPE_MMIO, 0, size);
> +	if (ret)
> +		goto fail;
> +
> +	ret = pf_migration_mmio_save(gt, vfid, data->vaddr, size);
> +	if (ret)
> +		goto fail;
> +
> +	xe_gt_sriov_dbg_verbose(gt, "VF%u MMIO data save (%zu bytes)\n", vfid, size);

maybe make it part of the pf_dump_mig_data() ?

	pf_dump_mig_data(gt, vfid, data, "MMIO saved");

> +	pf_dump_mig_data(gt, vfid, data);
> +
> +	ret = xe_gt_sriov_pf_migration_save_produce(gt, vfid, data);
> +	if (ret)
> +		goto fail;
> +
> +	return 0;
> +
> +fail:
> +	xe_sriov_migration_data_free(data);
> +	xe_gt_sriov_err(gt, "Failed to save VF%u MMIO data (%pe)\n", vfid, ERR_PTR(ret));
> +	return ret;
> +}
> +
> +static int pf_restore_vf_mmio_mig_data(struct xe_gt *gt, unsigned int vfid,
> +				       struct xe_sriov_migration_data *data)
> +{
> +	int ret;
> +
> +	xe_gt_sriov_dbg_verbose(gt, "VF%u MMIO data restore (%llu bytes)\n", vfid, data->size);

and here:

	pf_dump_mig_data(gt, vfid, data, "restoring MMIO");
> +	pf_dump_mig_data(gt, vfid, data);
> +
> +	ret = pf_migration_mmio_restore(gt, vfid, data->vaddr, data->size);
> +	if (ret) {
> +		xe_gt_sriov_err(gt, "Failed to restore VF%u MMIO data (%pe)\n",
> +				vfid, ERR_PTR(ret));
> +
> +		return ret;
> +	}
> +
> +	return 0;
> +}
> +
> +/**
> + * xe_gt_sriov_pf_migration_mmio_save() - Save VF MMIO migration data.
> + * @gt: the &xe_gt
> + * @vfid: the VF identifier (can't be 0)
> + *
> + * This function is for PF only.
> + *
> + * Return: 0 on success or a negative error code on failure.
> + */
> +int xe_gt_sriov_pf_migration_mmio_save(struct xe_gt *gt, unsigned int vfid)
> +{
> +	xe_gt_assert(gt, IS_SRIOV_PF(gt_to_xe(gt)));
> +	xe_gt_assert(gt, vfid != PFID);
> +	xe_gt_assert(gt, vfid <= xe_sriov_pf_get_totalvfs(gt_to_xe(gt)));
> +
> +	return pf_save_vf_mmio_mig_data(gt, vfid);
> +}
> +
> +/**
> + * xe_gt_sriov_pf_migration_mmio_restore() - Restore VF MMIO migration data.
> + * @gt: the &xe_gt
> + * @vfid: the VF identifier (can't be 0)
> + *
> + * This function is for PF only.
> + *
> + * Return: 0 on success or a negative error code on failure.
> + */
> +int xe_gt_sriov_pf_migration_mmio_restore(struct xe_gt *gt, unsigned int vfid,
> +					  struct xe_sriov_migration_data *data)
> +{
> +	xe_gt_assert(gt, IS_SRIOV_PF(gt_to_xe(gt)));
> +	xe_gt_assert(gt, vfid != PFID);
> +	xe_gt_assert(gt, vfid <= xe_sriov_pf_get_totalvfs(gt_to_xe(gt)));
> +
> +	return pf_restore_vf_mmio_mig_data(gt, vfid, data);
> +}
> +
>  /**
>   * xe_gt_sriov_pf_migration_size() - Total size of migration data from all components within a GT.
>   * @gt: the &xe_gt
> @@ -389,6 +537,13 @@ ssize_t xe_gt_sriov_pf_migration_size(struct xe_gt *gt, unsigned int vfid)
>  		size += sizeof(struct xe_sriov_pf_migration_hdr);
>  	total += size;
>  
> +	size = pf_migration_mmio_size(gt, vfid);
> +	if (size < 0)
> +		return size;
> +	if (size > 0)
> +		size += sizeof(struct xe_sriov_pf_migration_hdr);
> +	total += size;
> +
>  	return total;
>  }
>  
> @@ -453,6 +608,9 @@ void xe_gt_sriov_pf_migration_save_init(struct xe_gt *gt, unsigned int vfid)
>  
>  	if (pf_migration_ggtt_size(gt, vfid) > 0)
>  		set_bit(XE_SRIOV_MIGRATION_DATA_TYPE_GGTT, &migration->save.data_remaining);
> +
> +	xe_gt_assert(gt, pf_migration_mmio_size(gt, vfid) > 0);
> +	set_bit(XE_SRIOV_MIGRATION_DATA_TYPE_MMIO, &migration->save.data_remaining);
>  }
>  
>  /**
> diff --git a/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.h b/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.h
> index bc201d8f3147a..b0eec94fea3a6 100644
> --- a/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.h
> +++ b/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.h
> @@ -22,6 +22,9 @@ int xe_gt_sriov_pf_migration_guc_restore(struct xe_gt *gt, unsigned int vfid,
>  int xe_gt_sriov_pf_migration_ggtt_save(struct xe_gt *gt, unsigned int vfid);
>  int xe_gt_sriov_pf_migration_ggtt_restore(struct xe_gt *gt, unsigned int vfid,
>  					  struct xe_sriov_migration_data *data);
> +int xe_gt_sriov_pf_migration_mmio_save(struct xe_gt *gt, unsigned int vfid);
> +int xe_gt_sriov_pf_migration_mmio_restore(struct xe_gt *gt, unsigned int vfid,
> +					  struct xe_sriov_migration_data *data);
>  
>  ssize_t xe_gt_sriov_pf_migration_size(struct xe_gt *gt, unsigned int vfid);
>  

but patch looks ok, so with #include fixed (and maybe with better dump helper),

Reviewed-by: Michal Wajdeczko <michal.wajdeczko@intel.com>

Re: [PATCH v3 15/28] drm/xe/pf: Switch VF migration GuC save/restore to struct migration data

[Michal Wajdeczko]

On 10/30/2025 9:31 PM, Michał Winiarski wrote:
> In upcoming changes, the GuC VF migration data will be handled as part
> of separate SAVE/RESTORE states in VF control state machine.
> Now that the data is decoupled from both guc_state debugfs and PAUSE
> state, we can safely remove the struct xe_gt_sriov_state_snapshot and
> modify the GuC save/restore functions to operate on struct
> xe_sriov_migration_data.

hmm, that reminded me that maybe instead

	xe_sriov_migration_data

better name for this "data" struct could be

	xe_sriov_migration_packet

to make more distinguished from

	xe_gt_sriov_migration_data

which has completely different usage

> 
> Signed-off-by: Michał Winiarski <michal.winiarski@intel.com>
> ---
>  drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.c | 265 +++++-------------
>  drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.h |  13 +-
>  .../drm/xe/xe_gt_sriov_pf_migration_types.h   |  27 --
>  drivers/gpu/drm/xe/xe_gt_sriov_pf_types.h     |   4 -
>  4 files changed, 79 insertions(+), 230 deletions(-)
> 
> diff --git a/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.c b/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.c
> index a2db127982638..4a716e0a29fe4 100644
> --- a/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.c
> +++ b/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.c
> @@ -28,6 +28,17 @@ static struct xe_gt_sriov_migration_data *pf_pick_gt_migration(struct xe_gt *gt,
>  	return &gt->sriov.pf.vfs[vfid].migration;
>  }
>  
> +static void pf_dump_mig_data(struct xe_gt *gt, unsigned int vfid,
> +			     struct xe_sriov_migration_data *data)
> +{
> +	if (IS_ENABLED(CONFIG_DRM_XE_DEBUG_SRIOV)) {
> +		print_hex_dump_bytes("mig_hdr:  ", DUMP_PREFIX_OFFSET,
> +				     &data->hdr, sizeof(data->hdr));
> +		print_hex_dump_bytes("mig_data: ", DUMP_PREFIX_OFFSET,
> +				     data->vaddr, min(SZ_64, data->size));
> +	}

nit: maybe this function should be based on the drm_printer/drm_print_hex_dump
then we will get proper GTn: prefix 

> +}
> +
>  /* Return: number of dwords saved/restored/required or a negative error code on failure */
>  static int guc_action_vf_save_restore(struct xe_guc *guc, u32 vfid, u32 opcode,
>  				      u64 addr, u32 ndwords)
> @@ -47,7 +58,7 @@ static int guc_action_vf_save_restore(struct xe_guc *guc, u32 vfid, u32 opcode,
>  }
>  
>  /* Return: size of the state in dwords or a negative error code on failure */
> -static int pf_send_guc_query_vf_state_size(struct xe_gt *gt, unsigned int vfid)
> +static int pf_send_guc_query_vf_mig_data_size(struct xe_gt *gt, unsigned int vfid)
>  {
>  	int ret;
>  
> @@ -56,8 +67,8 @@ static int pf_send_guc_query_vf_state_size(struct xe_gt *gt, unsigned int vfid)
>  }
>  
>  /* Return: number of state dwords saved or a negative error code on failure */
> -static int pf_send_guc_save_vf_state(struct xe_gt *gt, unsigned int vfid,
> -				     void *dst, size_t size)
> +static int pf_send_guc_save_vf_mig_data(struct xe_gt *gt, unsigned int vfid,
> +					void *dst, size_t size)
>  {
>  	const int ndwords = size / sizeof(u32);
>  	struct xe_guc *guc = &gt->uc.guc;
> @@ -85,8 +96,8 @@ static int pf_send_guc_save_vf_state(struct xe_gt *gt, unsigned int vfid,
>  }
>  
>  /* Return: number of state dwords restored or a negative error code on failure */
> -static int pf_send_guc_restore_vf_state(struct xe_gt *gt, unsigned int vfid,
> -					const void *src, size_t size)
> +static int pf_send_guc_restore_vf_mig_data(struct xe_gt *gt, unsigned int vfid,
> +					   const void *src, size_t size)
>  {
>  	const int ndwords = size / sizeof(u32);
>  	struct xe_guc *guc = &gt->uc.guc;
> @@ -114,120 +125,67 @@ static bool pf_migration_supported(struct xe_gt *gt)
>  	return xe_sriov_pf_migration_supported(gt_to_xe(gt));
>  }
>  
> -static struct mutex *pf_migration_mutex(struct xe_gt *gt)
> +static int pf_save_vf_guc_mig_data(struct xe_gt *gt, unsigned int vfid)
>  {
> -	xe_gt_assert(gt, IS_SRIOV_PF(gt_to_xe(gt)));
> -	return &gt->sriov.pf.migration.snapshot_lock;
> -}
> -
> -static struct xe_gt_sriov_state_snapshot *pf_pick_vf_snapshot(struct xe_gt *gt,
> -							      unsigned int vfid)
> -{
> -	xe_gt_assert(gt, IS_SRIOV_PF(gt_to_xe(gt)));
> -	xe_gt_assert(gt, vfid <= xe_sriov_pf_get_totalvfs(gt_to_xe(gt)));
> -	lockdep_assert_held(pf_migration_mutex(gt));
> -
> -	return &gt->sriov.pf.vfs[vfid].snapshot;
> -}
> -
> -static unsigned int pf_snapshot_index(struct xe_gt *gt, struct xe_gt_sriov_state_snapshot *snapshot)
> -{
> -	return container_of(snapshot, struct xe_gt_sriov_metadata, snapshot) - gt->sriov.pf.vfs;
> -}
> -
> -static void pf_free_guc_state(struct xe_gt *gt, struct xe_gt_sriov_state_snapshot *snapshot)
> -{
> -	struct xe_device *xe = gt_to_xe(gt);
> -
> -	drmm_kfree(&xe->drm, snapshot->guc.buff);
> -	snapshot->guc.buff = NULL;
> -	snapshot->guc.size = 0;
> -}
> -
> -static int pf_alloc_guc_state(struct xe_gt *gt,
> -			      struct xe_gt_sriov_state_snapshot *snapshot,
> -			      size_t size)
> -{
> -	struct xe_device *xe = gt_to_xe(gt);
> -	void *p;
> -
> -	pf_free_guc_state(gt, snapshot);
> -
> -	if (!size)
> -		return -ENODATA;
> -
> -	if (size % sizeof(u32))
> -		return -EINVAL;
> -
> -	if (size > SZ_2M)
> -		return -EFBIG;
> -
> -	p = drmm_kzalloc(&xe->drm, size, GFP_KERNEL);
> -	if (!p)
> -		return -ENOMEM;
> -
> -	snapshot->guc.buff = p;
> -	snapshot->guc.size = size;
> -	return 0;
> -}
> -
> -static void pf_dump_guc_state(struct xe_gt *gt, struct xe_gt_sriov_state_snapshot *snapshot)
> -{
> -	if (IS_ENABLED(CONFIG_DRM_XE_DEBUG_SRIOV)) {
> -		unsigned int vfid __maybe_unused = pf_snapshot_index(gt, snapshot);
> -
> -		xe_gt_sriov_dbg_verbose(gt, "VF%u GuC state is %zu dwords:\n",
> -					vfid, snapshot->guc.size / sizeof(u32));
> -		print_hex_dump_bytes("state: ", DUMP_PREFIX_OFFSET,
> -				     snapshot->guc.buff, min(SZ_64, snapshot->guc.size));
> -	}
> -}
> -
> -static int pf_save_vf_guc_state(struct xe_gt *gt, unsigned int vfid)
> -{
> -	struct xe_gt_sriov_state_snapshot *snapshot = pf_pick_vf_snapshot(gt, vfid);
> +	struct xe_sriov_migration_data *data;
>  	size_t size;
>  	int ret;
>  
> -	ret = pf_send_guc_query_vf_state_size(gt, vfid);
> +	ret = pf_send_guc_query_vf_mig_data_size(gt, vfid);
>  	if (ret < 0)
>  		goto fail;
> +
>  	size = ret * sizeof(u32);
> -	xe_gt_sriov_dbg_verbose(gt, "VF%u state size is %d dwords (%zu bytes)\n", vfid, ret, size);
>  
> -	ret = pf_alloc_guc_state(gt, snapshot, size);
> -	if (ret < 0)
> +	data = xe_sriov_migration_data_alloc(gt_to_xe(gt));
> +	if (!data) {
> +		ret = -ENOMEM;
>  		goto fail;
> +	}
> +
> +	ret = xe_sriov_migration_data_init(data, gt->tile->id, gt->info.id,
> +					   XE_SRIOV_MIGRATION_DATA_TYPE_GUC, 0, size);
> +	if (ret)
> +		goto fail_free;
>  
> -	ret = pf_send_guc_save_vf_state(gt, vfid, snapshot->guc.buff, size);
> +	ret = pf_send_guc_save_vf_mig_data(gt, vfid, data->vaddr, size);
>  	if (ret < 0)
> -		goto fail;
> +		goto fail_free;
>  	size = ret * sizeof(u32);
>  	xe_gt_assert(gt, size);
> -	xe_gt_assert(gt, size <= snapshot->guc.size);
> -	snapshot->guc.size = size;
> +	xe_gt_assert(gt, size <= data->size);
> +	data->size = size;
> +	data->remaining = size;
> +
> +	xe_gt_sriov_dbg_verbose(gt, "VF%u GuC data save (%zu bytes)\n", vfid, size);
> +	pf_dump_mig_data(gt, vfid, data);

as already commented elsewhere, these two lines are always together,
we can combine them into improved pf_dump_mig_data(gt, vfid, data, what)

> +
> +	ret = xe_gt_sriov_pf_migration_save_produce(gt, vfid, data);
> +	if (ret)
> +		goto fail_free;
>  
> -	pf_dump_guc_state(gt, snapshot);
>  	return 0;
>  
> +fail_free:
> +	xe_sriov_migration_data_free(data);
>  fail:
> -	xe_gt_sriov_dbg(gt, "Unable to save VF%u state (%pe)\n", vfid, ERR_PTR(ret));
> -	pf_free_guc_state(gt, snapshot);
> +	xe_gt_sriov_err(gt, "Failed to save VF%u GuC data (%pe)\n",
> +			vfid, ERR_PTR(ret));
>  	return ret;
>  }
>  
>  /**
> - * xe_gt_sriov_pf_migration_save_guc_state() - Take a GuC VF state snapshot.
> + * xe_gt_sriov_pf_migration_guc_size() - Get the size of VF GuC migration data.
>   * @gt: the &xe_gt
>   * @vfid: the VF identifier
>   *
>   * This function is for PF only.
>   *
> - * Return: 0 on success or a negative error code on failure.
> + * Return: size in bytes or a negative error code on failure.
>   */
> -int xe_gt_sriov_pf_migration_save_guc_state(struct xe_gt *gt, unsigned int vfid)
> +ssize_t xe_gt_sriov_pf_migration_guc_size(struct xe_gt *gt, unsigned int vfid)
>  {
> -	int err;
> +	ssize_t size;
>  
>  	xe_gt_assert(gt, IS_SRIOV_PF(gt_to_xe(gt)));
>  	xe_gt_assert(gt, vfid != PFID);
> @@ -236,37 +194,15 @@ int xe_gt_sriov_pf_migration_save_guc_state(struct xe_gt *gt, unsigned int vfid)
>  	if (!pf_migration_supported(gt))
>  		return -ENOPKG;
>  
> -	mutex_lock(pf_migration_mutex(gt));
> -	err = pf_save_vf_guc_state(gt, vfid);
> -	mutex_unlock(pf_migration_mutex(gt));
> -
> -	return err;
> -}
> -
> -static int pf_restore_vf_guc_state(struct xe_gt *gt, unsigned int vfid)
> -{
> -	struct xe_gt_sriov_state_snapshot *snapshot = pf_pick_vf_snapshot(gt, vfid);
> -	int ret;
> -
> -	if (!snapshot->guc.size)
> -		return -ENODATA;
> -
> -	xe_gt_sriov_dbg_verbose(gt, "restoring %zu dwords of VF%u GuC state\n",
> -				snapshot->guc.size / sizeof(u32), vfid);
> -	ret = pf_send_guc_restore_vf_state(gt, vfid, snapshot->guc.buff, snapshot->guc.size);
> -	if (ret < 0)
> -		goto fail;
> -
> -	xe_gt_sriov_dbg_verbose(gt, "restored %d dwords of VF%u GuC state\n", ret, vfid);
> -	return 0;
> +	size = pf_send_guc_query_vf_mig_data_size(gt, vfid);
> +	if (size >= 0)
> +		size *= sizeof(u32);
>  
> -fail:
> -	xe_gt_sriov_dbg(gt, "Failed to restore VF%u GuC state (%pe)\n", vfid, ERR_PTR(ret));
> -	return ret;
> +	return size;
>  }
>  
>  /**
> - * xe_gt_sriov_pf_migration_restore_guc_state() - Restore a GuC VF state.
> + * xe_gt_sriov_pf_migration_guc_save() - Save VF GuC migration data.
>   * @gt: the &xe_gt
>   * @vfid: the VF identifier
>   *
> @@ -274,10 +210,8 @@ static int pf_restore_vf_guc_state(struct xe_gt *gt, unsigned int vfid)
>   *
>   * Return: 0 on success or a negative error code on failure.
>   */
> -int xe_gt_sriov_pf_migration_restore_guc_state(struct xe_gt *gt, unsigned int vfid)
> +int xe_gt_sriov_pf_migration_guc_save(struct xe_gt *gt, unsigned int vfid)
>  {
> -	int ret;
> -
>  	xe_gt_assert(gt, IS_SRIOV_PF(gt_to_xe(gt)));
>  	xe_gt_assert(gt, vfid != PFID);
>  	xe_gt_assert(gt, vfid <= xe_sriov_pf_get_totalvfs(gt_to_xe(gt)));
> @@ -285,75 +219,43 @@ int xe_gt_sriov_pf_migration_restore_guc_state(struct xe_gt *gt, unsigned int vf
>  	if (!pf_migration_supported(gt))
>  		return -ENOPKG;
>  
> -	mutex_lock(pf_migration_mutex(gt));
> -	ret = pf_restore_vf_guc_state(gt, vfid);
> -	mutex_unlock(pf_migration_mutex(gt));
> -
> -	return ret;
> +	return pf_save_vf_guc_mig_data(gt, vfid);
>  }
>  
> -#ifdef CONFIG_DEBUG_FS
> -/**
> - * xe_gt_sriov_pf_migration_read_guc_state() - Read a GuC VF state.
> - * @gt: the &xe_gt
> - * @vfid: the VF identifier
> - * @buf: the user space buffer to read to
> - * @count: the maximum number of bytes to read
> - * @pos: the current position in the buffer
> - *
> - * This function is for PF only.
> - *
> - * This function reads up to @count bytes from the saved VF GuC state buffer
> - * at offset @pos into the user space address starting at @buf.
> - *
> - * Return: the number of bytes read or a negative error code on failure.
> - */
> -ssize_t xe_gt_sriov_pf_migration_read_guc_state(struct xe_gt *gt, unsigned int vfid,
> -						char __user *buf, size_t count, loff_t *pos)
> +static int pf_restore_vf_guc_state(struct xe_gt *gt, unsigned int vfid,
> +				   struct xe_sriov_migration_data *data)
>  {
> -	struct xe_gt_sriov_state_snapshot *snapshot;
> -	ssize_t ret;
> +	int ret;
>  
> -	xe_gt_assert(gt, IS_SRIOV_PF(gt_to_xe(gt)));
> -	xe_gt_assert(gt, vfid != PFID);
> -	xe_gt_assert(gt, vfid <= xe_sriov_pf_get_totalvfs(gt_to_xe(gt)));
> +	xe_gt_assert(gt, data->size);
>  
> -	if (!pf_migration_supported(gt))
> -		return -ENOPKG;
> +	xe_gt_sriov_dbg_verbose(gt, "VF%u GuC data restore (%llu bytes)\n", vfid, data->size);
> +	pf_dump_mig_data(gt, vfid, data);
>  
> -	mutex_lock(pf_migration_mutex(gt));
> -	snapshot = pf_pick_vf_snapshot(gt, vfid);
> -	if (snapshot->guc.size)
> -		ret = simple_read_from_buffer(buf, count, pos, snapshot->guc.buff,
> -					      snapshot->guc.size);
> -	else
> -		ret = -ENODATA;
> -	mutex_unlock(pf_migration_mutex(gt));
> +	ret = pf_send_guc_restore_vf_mig_data(gt, vfid, data->vaddr, data->size);
> +	if (ret < 0)
> +		goto fail;
> +
> +	return 0;
>  
> +fail:
> +	xe_gt_sriov_err(gt, "Failed to restore VF%u GuC data (%pe)\n",
> +			vfid, ERR_PTR(ret));
>  	return ret;
>  }
>  
>  /**
> - * xe_gt_sriov_pf_migration_write_guc_state() - Write a GuC VF state.
> + * xe_gt_sriov_pf_migration_guc_restore() - Restore VF GuC migration data.
>   * @gt: the &xe_gt
>   * @vfid: the VF identifier
> - * @buf: the user space buffer with GuC VF state
> - * @size: the size of GuC VF state (in bytes)
>   *
>   * This function is for PF only.
>   *
> - * This function reads @size bytes of the VF GuC state stored at user space
> - * address @buf and writes it into a internal VF state buffer.
> - *
> - * Return: the number of bytes used or a negative error code on failure.
> + * Return: 0 on success or a negative error code on failure.
>   */
> -ssize_t xe_gt_sriov_pf_migration_write_guc_state(struct xe_gt *gt, unsigned int vfid,
> -						 const char __user *buf, size_t size)
> +int xe_gt_sriov_pf_migration_guc_restore(struct xe_gt *gt, unsigned int vfid,
> +					 struct xe_sriov_migration_data *data)
>  {
> -	struct xe_gt_sriov_state_snapshot *snapshot;
> -	loff_t pos = 0;
> -	ssize_t ret;
> -
>  	xe_gt_assert(gt, IS_SRIOV_PF(gt_to_xe(gt)));
>  	xe_gt_assert(gt, vfid != PFID);
>  	xe_gt_assert(gt, vfid <= xe_sriov_pf_get_totalvfs(gt_to_xe(gt)));
> @@ -361,21 +263,8 @@ ssize_t xe_gt_sriov_pf_migration_write_guc_state(struct xe_gt *gt, unsigned int
>  	if (!pf_migration_supported(gt))
>  		return -ENOPKG;
>  
> -	mutex_lock(pf_migration_mutex(gt));
> -	snapshot = pf_pick_vf_snapshot(gt, vfid);
> -	ret = pf_alloc_guc_state(gt, snapshot, size);
> -	if (!ret) {
> -		ret = simple_write_to_buffer(snapshot->guc.buff, size, &pos, buf, size);
> -		if (ret < 0)
> -			pf_free_guc_state(gt, snapshot);
> -		else
> -			pf_dump_guc_state(gt, snapshot);
> -	}
> -	mutex_unlock(pf_migration_mutex(gt));
> -
> -	return ret;
> +	return pf_restore_vf_guc_state(gt, vfid, data);
>  }
> -#endif /* CONFIG_DEBUG_FS */
>  
>  /**
>   * xe_gt_sriov_pf_migration_size() - Total size of migration data from all components within a GT.
> @@ -599,10 +488,6 @@ int xe_gt_sriov_pf_migration_init(struct xe_gt *gt)
>  	if (!pf_migration_supported(gt))
>  		return 0;
>  
> -	err = drmm_mutex_init(&xe->drm, &gt->sriov.pf.migration.snapshot_lock);
> -	if (err)
> -		return err;
> -
>  	totalvfs = xe_sriov_pf_get_totalvfs(xe);
>  	for (n = 1; n <= totalvfs; n++) {
>  		struct xe_gt_sriov_migration_data *migration = pf_pick_gt_migration(gt, n);
> diff --git a/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.h b/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.h
> index 4f2f2783339c3..b3c18e369df79 100644
> --- a/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.h
> +++ b/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.h
> @@ -15,8 +15,10 @@ struct xe_sriov_migration_data;
>  #define XE_GT_SRIOV_PF_MIGRATION_GUC_DATA_MAX_SIZE SZ_8M
>  
>  int xe_gt_sriov_pf_migration_init(struct xe_gt *gt);
> -int xe_gt_sriov_pf_migration_save_guc_state(struct xe_gt *gt, unsigned int vfid);
> -int xe_gt_sriov_pf_migration_restore_guc_state(struct xe_gt *gt, unsigned int vfid);
> +ssize_t xe_gt_sriov_pf_migration_guc_size(struct xe_gt *gt, unsigned int vfid);
> +int xe_gt_sriov_pf_migration_guc_save(struct xe_gt *gt, unsigned int vfid);
> +int xe_gt_sriov_pf_migration_guc_restore(struct xe_gt *gt, unsigned int vfid,
> +					 struct xe_sriov_migration_data *data);
>  
>  ssize_t xe_gt_sriov_pf_migration_size(struct xe_gt *gt, unsigned int vfid);
>  
> @@ -34,11 +36,4 @@ int xe_gt_sriov_pf_migration_restore_produce(struct xe_gt *gt, unsigned int vfid
>  struct xe_sriov_migration_data *
>  xe_gt_sriov_pf_migration_save_consume(struct xe_gt *gt, unsigned int vfid);
>  
> -#ifdef CONFIG_DEBUG_FS
> -ssize_t xe_gt_sriov_pf_migration_read_guc_state(struct xe_gt *gt, unsigned int vfid,
> -						char __user *buf, size_t count, loff_t *pos);
> -ssize_t xe_gt_sriov_pf_migration_write_guc_state(struct xe_gt *gt, unsigned int vfid,
> -						 const char __user *buf, size_t count);
> -#endif
> -
>  #endif
> diff --git a/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration_types.h b/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration_types.h
> index 84be6fac16c8b..75d8b94cbbefb 100644
> --- a/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration_types.h
> +++ b/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration_types.h
> @@ -6,24 +6,7 @@
>  #ifndef _XE_GT_SRIOV_PF_MIGRATION_TYPES_H_
>  #define _XE_GT_SRIOV_PF_MIGRATION_TYPES_H_
>  
> -#include <linux/mutex.h>
>  #include <linux/ptr_ring.h>
> -#include <linux/types.h>
> -
> -/**
> - * struct xe_gt_sriov_state_snapshot - GT-level per-VF state snapshot data.
> - *
> - * Used by the PF driver to maintain per-VF migration data.
> - */
> -struct xe_gt_sriov_state_snapshot {
> -	/** @guc: GuC VF state snapshot */
> -	struct {
> -		/** @guc.buff: buffer with the VF state */
> -		u32 *buff;
> -		/** @guc.size: size of the buffer (must be dwords aligned) */
> -		u32 size;
> -	} guc;
> -};
>  
>  /**
>   * struct xe_gt_sriov_migration_data - GT-level per-VF migration data.
> @@ -35,14 +18,4 @@ struct xe_gt_sriov_migration_data {
>  	struct ptr_ring ring;
>  };
>  
> -/**
> - * struct xe_gt_sriov_pf_migration - GT-level data.
> - *
> - * Used by the PF driver to maintain non-VF specific per-GT data.
> - */
> -struct xe_gt_sriov_pf_migration {
> -	/** @snapshot_lock: protects all VFs snapshots */
> -	struct mutex snapshot_lock;
> -};
> -
>  #endif
> diff --git a/drivers/gpu/drm/xe/xe_gt_sriov_pf_types.h b/drivers/gpu/drm/xe/xe_gt_sriov_pf_types.h
> index 812e74d3f8f80..667b8310478d4 100644
> --- a/drivers/gpu/drm/xe/xe_gt_sriov_pf_types.h
> +++ b/drivers/gpu/drm/xe/xe_gt_sriov_pf_types.h
> @@ -31,9 +31,6 @@ struct xe_gt_sriov_metadata {
>  	/** @version: negotiated VF/PF ABI version */
>  	struct xe_gt_sriov_pf_service_version version;
>  
> -	/** @snapshot: snapshot of the VF state data */
> -	struct xe_gt_sriov_state_snapshot snapshot;
> -
>  	/** @migration: per-VF migration data. */
>  	struct xe_gt_sriov_migration_data migration;
>  };
> @@ -61,7 +58,6 @@ struct xe_gt_sriov_pf {
>  	struct xe_gt_sriov_pf_service service;
>  	struct xe_gt_sriov_pf_control control;
>  	struct xe_gt_sriov_pf_policy policy;
> -	struct xe_gt_sriov_pf_migration migration;
>  	struct xe_gt_sriov_spare_config spare;
>  	struct xe_gt_sriov_metadata *vfs;
>  };

otherwise LGTM, so with dump helper improved

	Reviewed-by: Michal Wajdeczko <michal.wajdeczko@intel.com>

Re: [PATCH v3 02/28] drm/xe: Move migration support to device-level struct

[Michal Wajdeczko]

On 10/30/2025 9:31 PM, Michał Winiarski wrote:
> Upcoming changes will allow users to control VF state and obtain its
> migration data with a device-level granularity (not tile/gt).
> Change the data structures to reflect that and move the GT-level
> migration init to happen after device-level init.
> 
> Signed-off-by: Michał Winiarski <michal.winiarski@intel.com>
> Reviewed-by: Michal Wajdeczko <michal.wajdeczko@intel.com>
> ---

...

> diff --git a/drivers/gpu/drm/xe/xe_sriov_pf_migration_types.h b/drivers/gpu/drm/xe/xe_sriov_pf_migration_types.h
> new file mode 100644
> index 0000000000000..e69de29bb2d1d
> diff --git a/drivers/gpu/drm/xe/xe_sriov_pf_types.h b/drivers/gpu/drm/xe/xe_sriov_pf_types.h
> index c753cd59aed2b..24d22afeececa 100644
> --- a/drivers/gpu/drm/xe/xe_sriov_pf_types.h
> +++ b/drivers/gpu/drm/xe/xe_sriov_pf_types.h
> @@ -39,6 +39,12 @@ struct xe_device_pf {
>  	/** @provision: device level provisioning data. */
>  	struct xe_sriov_pf_provision provision;
>  
> +	/** @migration: device level VF migration data */
> +	struct {
> +		/** @migration.supported: indicates whether VF migration feature is supported */
> +		bool supported;
> +	} migration;
> +

late notice: all our other sub-components (except sysfs *) to hold its private fields define its own struct that matches component name

but here, you use anonymous struct instead and what worse,
later on you start using xe_sriov_pf_migration name to hold per-VF data,
which breaks the above naming pattern even more

can you add to xe_sriov_pf_migration_types.h:

+struct xe_sriov_pf_migration {
+	/** @supported: indicates whether VF migration feature is supported */
+	bool supported;
+};

and rename per-VF struct to something else? like:

	struct xe_sriov_migration_state - Per VF device-level migration related data


*) sysfs doesn't define/use any custom data types, just kobject

>  	/** @service: device level service data. */
>  	struct xe_sriov_pf_service service;
>  

Re: [PATCH v3 27/28] drm/intel/pciids: Add match with VFIO override

[Lucas De Marchi]

On Thu, Oct 30, 2025 at 09:31:34PM +0100, Michał Winiarski wrote:
>In order to allow VFIO users to choose the right driver override, VFIO
>driver variant used for VF migration needs to use Intel Graphics PCI
>IDs.
>Add INTEL_VGA_VFIO_DEVICE match that sets VFIO override_only.
>
>Signed-off-by: Michał Winiarski <michal.winiarski@intel.com>
>---
> include/drm/intel/pciids.h | 7 +++++++
> 1 file changed, 7 insertions(+)
>
>diff --git a/include/drm/intel/pciids.h b/include/drm/intel/pciids.h
>index b258e79b437ac..d14ce43139a28 100644
>--- a/include/drm/intel/pciids.h
>+++ b/include/drm/intel/pciids.h
>@@ -43,6 +43,13 @@
> 	.class = PCI_BASE_CLASS_DISPLAY << 16, .class_mask = 0xff << 16, \
> 	.driver_data = (kernel_ulong_t)(_info), \
> }
>+
>+#define INTEL_VGA_VFIO_DEVICE(_id, _info) { \
>+	PCI_DEVICE(PCI_VENDOR_ID_INTEL, (_id)), \
>+	.class = PCI_BASE_CLASS_DISPLAY << 16, .class_mask = 0xff << 16, \
>+	.driver_data = (kernel_ulong_t)(_info), \
>+	.override_only = PCI_ID_F_VFIO_DRIVER_OVERRIDE, \

why do we need this and can't use PCI_DRIVER_OVERRIDE_DEVICE_VFIO()
directly? Note that there are GPUs that wouldn't match the display class
above.

	edb660ad79ff ("drm/intel/pciids: Add match on vendor/id only")
	5e0de2dfbc1b ("drm/xe/cri: Add CRI platform definition")

Lucas De Marchi

>+}
> #endif
>
> #define INTEL_I810_IDS(MACRO__, ...) \
>-- 
>2.50.1
>

Re: [PATCH v3 21/28] drm/xe/migrate: Add function to copy of VRAM data in chunks

[Matthew Brost]

On Thu, Oct 30, 2025 at 09:31:28PM +0100, Michał Winiarski wrote:
> From: Lukasz Laguna <lukasz.laguna@intel.com>
> 
> Introduce a new function to copy data between VRAM and sysmem objects.
> The existing xe_migrate_copy() is tailored for eviction and restore
> operations, which involves additional logic and operates on entire
> objects.
> The xe_migrate_vram_copy_chunk() allows copying chunks of data to or
> from a dedicated buffer object, which is essential in case of VF
> migration.
> 
> Signed-off-by: Lukasz Laguna <lukasz.laguna@intel.com>
> Signed-off-by: Michał Winiarski <michal.winiarski@intel.com>

Reviewed-by: Matthew Brost <matthew.brost@intel.com>

> ---
>  drivers/gpu/drm/xe/xe_migrate.c | 128 ++++++++++++++++++++++++++++++--
>  drivers/gpu/drm/xe/xe_migrate.h |   8 ++
>  2 files changed, 131 insertions(+), 5 deletions(-)
> 
> diff --git a/drivers/gpu/drm/xe/xe_migrate.c b/drivers/gpu/drm/xe/xe_migrate.c
> index 56a5804726e96..dbe9320863ab0 100644
> --- a/drivers/gpu/drm/xe/xe_migrate.c
> +++ b/drivers/gpu/drm/xe/xe_migrate.c
> @@ -29,6 +29,7 @@
>  #include "xe_lrc.h"
>  #include "xe_map.h"
>  #include "xe_mocs.h"
> +#include "xe_printk.h"
>  #include "xe_pt.h"
>  #include "xe_res_cursor.h"
>  #include "xe_sa.h"
> @@ -1210,6 +1211,128 @@ struct xe_exec_queue *xe_migrate_exec_queue(struct xe_migrate *migrate)
>  	return migrate->q;
>  }
>  
> +/**
> + * xe_migrate_vram_copy_chunk() - Copy a chunk of a VRAM buffer object.
> + * @vram_bo: The VRAM buffer object.
> + * @vram_offset: The VRAM offset.
> + * @sysmem_bo: The sysmem buffer object.
> + * @sysmem_offset: The sysmem offset.
> + * @size: The size of VRAM chunk to copy.
> + * @dir: The direction of the copy operation.
> + *
> + * Copies a portion of a buffer object between VRAM and system memory.
> + * On Xe2 platforms that support flat CCS, VRAM data is decompressed when
> + * copying to system memory.
> + *
> + * Return: Pointer to a dma_fence representing the last copy batch, or
> + * an error pointer on failure. If there is a failure, any copy operation
> + * started by the function call has been synced.
> + */
> +struct dma_fence *xe_migrate_vram_copy_chunk(struct xe_bo *vram_bo, u64 vram_offset,
> +					     struct xe_bo *sysmem_bo, u64 sysmem_offset,
> +					     u64 size, enum xe_migrate_copy_dir dir)
> +{
> +	struct xe_device *xe = xe_bo_device(vram_bo);
> +	struct xe_tile *tile = vram_bo->tile;
> +	struct xe_gt *gt = tile->primary_gt;
> +	struct xe_migrate *m = tile->migrate;
> +	struct dma_fence *fence = NULL;
> +	struct ttm_resource *vram = vram_bo->ttm.resource;
> +	struct ttm_resource *sysmem = sysmem_bo->ttm.resource;
> +	struct xe_res_cursor vram_it, sysmem_it;
> +	u64 vram_L0_ofs, sysmem_L0_ofs;
> +	u32 vram_L0_pt, sysmem_L0_pt;
> +	u64 vram_L0, sysmem_L0;
> +	bool to_sysmem = (dir == XE_MIGRATE_COPY_TO_SRAM);
> +	bool use_comp_pat = to_sysmem &&
> +		GRAPHICS_VER(xe) >= 20 && xe_device_has_flat_ccs(xe);
> +	int pass = 0;
> +	int err;
> +
> +	xe_assert(xe, IS_ALIGNED(vram_offset | sysmem_offset | size, PAGE_SIZE));
> +	xe_assert(xe, xe_bo_is_vram(vram_bo));
> +	xe_assert(xe, !xe_bo_is_vram(sysmem_bo));
> +	xe_assert(xe, !range_overflows(vram_offset, size, (u64)vram_bo->ttm.base.size));
> +	xe_assert(xe, !range_overflows(sysmem_offset, size, (u64)sysmem_bo->ttm.base.size));
> +
> +	xe_res_first(vram, vram_offset, size, &vram_it);
> +	xe_res_first_sg(xe_bo_sg(sysmem_bo), sysmem_offset, size, &sysmem_it);
> +
> +	while (size) {
> +		u32 pte_flags = PTE_UPDATE_FLAG_IS_VRAM;
> +		u32 batch_size = 2; /* arb_clear() + MI_BATCH_BUFFER_END */
> +		struct xe_sched_job *job;
> +		struct xe_bb *bb;
> +		u32 update_idx;
> +		bool usm = xe->info.has_usm;
> +		u32 avail_pts = max_mem_transfer_per_pass(xe) / LEVEL0_PAGE_TABLE_ENCODE_SIZE;
> +
> +		sysmem_L0 = xe_migrate_res_sizes(m, &sysmem_it);
> +		vram_L0 = min(xe_migrate_res_sizes(m, &vram_it), sysmem_L0);
> +
> +		xe_dbg(xe, "Pass %u, size: %llu\n", pass++, vram_L0);
> +
> +		pte_flags |= use_comp_pat ? PTE_UPDATE_FLAG_IS_COMP_PTE : 0;
> +		batch_size += pte_update_size(m, pte_flags, vram, &vram_it, &vram_L0,
> +					      &vram_L0_ofs, &vram_L0_pt, 0, 0, avail_pts);
> +
> +		batch_size += pte_update_size(m, 0, sysmem, &sysmem_it, &vram_L0, &sysmem_L0_ofs,
> +					      &sysmem_L0_pt, 0, avail_pts, avail_pts);
> +		batch_size += EMIT_COPY_DW;
> +
> +		bb = xe_bb_new(gt, batch_size, usm);
> +		if (IS_ERR(bb)) {
> +			err = PTR_ERR(bb);
> +			return ERR_PTR(err);
> +		}
> +
> +		if (xe_migrate_allow_identity(vram_L0, &vram_it))
> +			xe_res_next(&vram_it, vram_L0);
> +		else
> +			emit_pte(m, bb, vram_L0_pt, true, use_comp_pat, &vram_it, vram_L0, vram);
> +
> +		emit_pte(m, bb, sysmem_L0_pt, false, false, &sysmem_it, vram_L0, sysmem);
> +
> +		bb->cs[bb->len++] = MI_BATCH_BUFFER_END;
> +		update_idx = bb->len;
> +
> +		if (to_sysmem)
> +			emit_copy(gt, bb, vram_L0_ofs, sysmem_L0_ofs, vram_L0, XE_PAGE_SIZE);
> +		else
> +			emit_copy(gt, bb, sysmem_L0_ofs, vram_L0_ofs, vram_L0, XE_PAGE_SIZE);
> +
> +		job = xe_bb_create_migration_job(m->q, bb, xe_migrate_batch_base(m, usm),
> +						 update_idx);
> +		if (IS_ERR(job)) {
> +			xe_bb_free(bb, NULL);
> +			err = PTR_ERR(job);
> +			return ERR_PTR(err);
> +		}
> +
> +		xe_sched_job_add_migrate_flush(job, MI_INVALIDATE_TLB);
> +
> +		xe_assert(xe, dma_resv_test_signaled(vram_bo->ttm.base.resv,
> +						     DMA_RESV_USAGE_BOOKKEEP));
> +		xe_assert(xe, dma_resv_test_signaled(sysmem_bo->ttm.base.resv,
> +						     DMA_RESV_USAGE_BOOKKEEP));
> +
> +		scoped_guard(mutex, &m->job_mutex) {
> +			xe_sched_job_arm(job);
> +			dma_fence_put(fence);
> +			fence = dma_fence_get(&job->drm.s_fence->finished);
> +			xe_sched_job_push(job);
> +
> +			dma_fence_put(m->fence);
> +			m->fence = dma_fence_get(fence);
> +		}
> +
> +		xe_bb_free(bb, fence);
> +		size -= vram_L0;
> +	}
> +
> +	return fence;
> +}
> +
>  static void emit_clear_link_copy(struct xe_gt *gt, struct xe_bb *bb, u64 src_ofs,
>  				 u32 size, u32 pitch)
>  {
> @@ -1912,11 +2035,6 @@ static bool xe_migrate_vram_use_pde(struct drm_pagemap_addr *sram_addr,
>  	return true;
>  }
>  
> -enum xe_migrate_copy_dir {
> -	XE_MIGRATE_COPY_TO_VRAM,
> -	XE_MIGRATE_COPY_TO_SRAM,
> -};
> -
>  #define XE_CACHELINE_BYTES	64ull
>  #define XE_CACHELINE_MASK	(XE_CACHELINE_BYTES - 1)
>  
> diff --git a/drivers/gpu/drm/xe/xe_migrate.h b/drivers/gpu/drm/xe/xe_migrate.h
> index 4fad324b62535..d7bcc6ad8464e 100644
> --- a/drivers/gpu/drm/xe/xe_migrate.h
> +++ b/drivers/gpu/drm/xe/xe_migrate.h
> @@ -28,6 +28,11 @@ struct xe_vma;
>  
>  enum xe_sriov_vf_ccs_rw_ctxs;
>  
> +enum xe_migrate_copy_dir {
> +	XE_MIGRATE_COPY_TO_VRAM,
> +	XE_MIGRATE_COPY_TO_SRAM,
> +};
> +
>  /**
>   * struct xe_migrate_pt_update_ops - Callbacks for the
>   * xe_migrate_update_pgtables() function.
> @@ -131,6 +136,9 @@ int xe_migrate_ccs_rw_copy(struct xe_tile *tile, struct xe_exec_queue *q,
>  
>  struct xe_lrc *xe_migrate_lrc(struct xe_migrate *migrate);
>  struct xe_exec_queue *xe_migrate_exec_queue(struct xe_migrate *migrate);
> +struct dma_fence *xe_migrate_vram_copy_chunk(struct xe_bo *vram_bo, u64 vram_offset,
> +					     struct xe_bo *sysmem_bo, u64 sysmem_offset,
> +					     u64 size, enum xe_migrate_copy_dir dir);
>  int xe_migrate_access_memory(struct xe_migrate *m, struct xe_bo *bo,
>  			     unsigned long offset, void *buf, int len,
>  			     int write);
> -- 
> 2.50.1
> 

Re: [PATCH v3 22/28] drm/xe/pf: Handle VRAM migration data as part of PF control

[Matthew Brost]

On Thu, Oct 30, 2025 at 09:31:29PM +0100, Michał Winiarski wrote:
> Connect the helpers to allow save and restore of VRAM migration data in
> stop_copy / resume device state.
> 
> Co-developed-by: Lukasz Laguna <lukasz.laguna@intel.com>
> Signed-off-by: Lukasz Laguna <lukasz.laguna@intel.com>
> Signed-off-by: Michał Winiarski <michal.winiarski@intel.com>
> ---
>  drivers/gpu/drm/xe/xe_gt_sriov_pf_control.c   |  15 ++
>  drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.c | 213 ++++++++++++++++++
>  drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.h |   5 +
>  .../drm/xe/xe_gt_sriov_pf_migration_types.h   |   2 +
>  drivers/gpu/drm/xe/xe_sriov_pf_control.c      |   3 +
>  5 files changed, 238 insertions(+)
> 
> diff --git a/drivers/gpu/drm/xe/xe_gt_sriov_pf_control.c b/drivers/gpu/drm/xe/xe_gt_sriov_pf_control.c
> index 7cd7cae950bc7..de0f63610e780 100644
> --- a/drivers/gpu/drm/xe/xe_gt_sriov_pf_control.c
> +++ b/drivers/gpu/drm/xe/xe_gt_sriov_pf_control.c
> @@ -880,6 +880,18 @@ static int pf_handle_vf_save_data(struct xe_gt *gt, unsigned int vfid)
>  		return -EAGAIN;
>  	}
>  
> +	if (xe_gt_sriov_pf_migration_save_test(gt, vfid, XE_SRIOV_MIGRATION_DATA_TYPE_VRAM)) {
> +		ret = xe_gt_sriov_pf_migration_vram_save(gt, vfid);
> +		if (ret == -EAGAIN)
> +			return -EAGAIN;
> +		else if (ret)
> +			return ret;
> +
> +		xe_gt_sriov_pf_migration_save_clear(gt, vfid, XE_SRIOV_MIGRATION_DATA_TYPE_VRAM);
> +
> +		return -EAGAIN;
> +	}
> +
>  	return 0;
>  }
>  
> @@ -1095,6 +1107,9 @@ static int pf_handle_vf_restore_data(struct xe_gt *gt, unsigned int vfid)
>  	case XE_SRIOV_MIGRATION_DATA_TYPE_GUC:
>  		ret = xe_gt_sriov_pf_migration_guc_restore(gt, vfid, data);
>  		break;
> +	case XE_SRIOV_MIGRATION_DATA_TYPE_VRAM:
> +		ret = xe_gt_sriov_pf_migration_vram_restore(gt, vfid, data);
> +		break;
>  	default:
>  		xe_gt_sriov_notice(gt, "Skipping VF%u unknown data type: %d\n", vfid, data->type);
>  		break;
> diff --git a/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.c b/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.c
> index 5e90aeafeeb41..7ccd43545fdac 100644
> --- a/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.c
> +++ b/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.c
> @@ -19,6 +19,7 @@
>  #include "xe_gt_sriov_printk.h"
>  #include "xe_guc_buf.h"
>  #include "xe_guc_ct.h"
> +#include "xe_migrate.h"
>  #include "xe_mmio.h"
>  #include "xe_sriov.h"
>  #include "xe_sriov_migration_data.h"
> @@ -505,6 +506,207 @@ int xe_gt_sriov_pf_migration_mmio_restore(struct xe_gt *gt, unsigned int vfid,
>  	return pf_restore_vf_mmio_mig_data(gt, vfid, data);
>  }
>  
> +static ssize_t pf_migration_vram_size(struct xe_gt *gt, unsigned int vfid)
> +{
> +	if (!xe_gt_is_main_type(gt))
> +		return 0;
> +
> +	return xe_gt_sriov_pf_config_get_lmem(gt, vfid);
> +}
> +
> +static struct dma_fence *__pf_save_restore_vram(struct xe_gt *gt, unsigned int vfid,
> +						struct xe_bo *vram, u64 vram_offset,
> +						struct xe_bo *sysmem, u64 sysmem_offset,
> +						size_t size, bool save)
> +{
> +	struct dma_fence *ret = NULL;
> +	struct drm_exec exec;
> +	int err;
> +
> +	drm_exec_init(&exec, DRM_EXEC_INTERRUPTIBLE_WAIT, 0);

It is a bit incongruent that the dma-resv lock is interruptible here.

> +	drm_exec_until_all_locked(&exec) {
> +		err = drm_exec_lock_obj(&exec, &vram->ttm.base);
> +		drm_exec_retry_on_contention(&exec);
> +		if (err) {
> +			ret = ERR_PTR(err);
> +			goto err;
> +		}
> +
> +		err = drm_exec_lock_obj(&exec, &sysmem->ttm.base);
> +		drm_exec_retry_on_contention(&exec);
> +		if (err) {
> +			ret = ERR_PTR(err);
> +			goto err;
> +		}
> +	}
> +
> +	ret = xe_migrate_vram_copy_chunk(vram, vram_offset, sysmem, sysmem_offset, size,
> +					 save ? XE_MIGRATE_COPY_TO_SRAM : XE_MIGRATE_COPY_TO_VRAM);
> +
> +err:
> +	drm_exec_fini(&exec);
> +
> +	return ret;
> +}
> +
> +#define PF_VRAM_SAVE_RESTORE_TIMEOUT (5 * HZ)
> +static int pf_save_vram_chunk(struct xe_gt *gt, unsigned int vfid,
> +			      struct xe_bo *src_vram, u64 src_vram_offset,
> +			      size_t size)
> +{
> +	struct xe_sriov_migration_data *data;
> +	struct dma_fence *fence;
> +	int ret;
> +
> +	data = xe_sriov_migration_data_alloc(gt_to_xe(gt));
> +	if (!data)
> +		return -ENOMEM;
> +
> +	ret = xe_sriov_migration_data_init(data, gt->tile->id, gt->info.id,
> +					   XE_SRIOV_MIGRATION_DATA_TYPE_VRAM,
> +					   src_vram_offset, size);
> +	if (ret)
> +		goto fail;
> +
> +	fence = __pf_save_restore_vram(gt, vfid,
> +				       src_vram, src_vram_offset,
> +				       data->bo, 0, size, true);
> +
> +	ret = dma_fence_wait_timeout(fence, false, PF_VRAM_SAVE_RESTORE_TIMEOUT);

Then here the dma-fence wait is non-interruptible.

> +	dma_fence_put(fence);
> +	if (!ret) {
> +		ret = -ETIME;
> +		goto fail;
> +	}
> +
> +	xe_gt_sriov_dbg_verbose(gt, "VF%u VRAM data save (%zu bytes)\n", vfid, size);
> +	pf_dump_mig_data(gt, vfid, data);
> +
> +	ret = xe_gt_sriov_pf_migration_save_produce(gt, vfid, data);
> +	if (ret)
> +		goto fail;
> +
> +	return 0;
> +
> +fail:
> +	xe_sriov_migration_data_free(data);
> +	return ret;
> +}
> +
> +#define VF_VRAM_STATE_CHUNK_MAX_SIZE SZ_512M
> +static int pf_save_vf_vram_mig_data(struct xe_gt *gt, unsigned int vfid)
> +{
> +	struct xe_gt_sriov_migration_data *migration = pf_pick_gt_migration(gt, vfid);
> +	loff_t *offset = &migration->save.vram_offset;
> +	struct xe_bo *vram;
> +	size_t vram_size, chunk_size;
> +	int ret;
> +
> +	vram = xe_gt_sriov_pf_config_get_lmem_obj(gt, vfid);
> +	if (!vram)
> +		return -ENXIO;
> +
> +	vram_size = xe_bo_size(vram);
> +
> +	xe_gt_assert(gt, *offset < vram_size);
> +
> +	chunk_size = min(vram_size - *offset, VF_VRAM_STATE_CHUNK_MAX_SIZE);
> +
> +	ret = pf_save_vram_chunk(gt, vfid, vram, *offset, chunk_size);
> +	if (ret)
> +		goto fail;
> +
> +	*offset += chunk_size;
> +
> +	xe_bo_put(vram);
> +
> +	if (*offset < vram_size)
> +		return -EAGAIN;
> +
> +	return 0;
> +
> +fail:
> +	xe_bo_put(vram);
> +	xe_gt_sriov_err(gt, "Failed to save VF%u VRAM data (%pe)\n", vfid, ERR_PTR(ret));
> +	return ret;
> +}
> +
> +static int pf_restore_vf_vram_mig_data(struct xe_gt *gt, unsigned int vfid,
> +				       struct xe_sriov_migration_data *data)
> +{
> +	u64 end = data->hdr.offset + data->hdr.size;
> +	struct dma_fence *fence;
> +	struct xe_bo *vram;
> +	size_t size;
> +	int ret = 0;
> +
> +	vram = xe_gt_sriov_pf_config_get_lmem_obj(gt, vfid);
> +	if (!vram)
> +		return -ENXIO;
> +
> +	size = xe_bo_size(vram);
> +
> +	if (end > size || end < data->hdr.size) {
> +		ret = -EINVAL;
> +		goto err;
> +	}
> +
> +	xe_gt_sriov_dbg_verbose(gt, "VF%u VRAM data restore (%llu bytes)\n", vfid, data->size);
> +	pf_dump_mig_data(gt, vfid, data);
> +
> +	fence = __pf_save_restore_vram(gt, vfid, vram, data->hdr.offset,
> +				       data->bo, 0, data->hdr.size, false);
> +	ret = dma_fence_wait_timeout(fence, false, PF_VRAM_SAVE_RESTORE_TIMEOUT);

And also here the dma-fence wait is non-interruptible.

Aside from the incongruence, the usage of drm-exec, dma-fences looks correct.

Matt

> +	dma_fence_put(fence);
> +	if (!ret) {
> +		ret = -ETIME;
> +		goto err;
> +	}
> +
> +	return 0;
> +err:
> +	xe_bo_put(vram);
> +	xe_gt_sriov_err(gt, "Failed to restore VF%u VRAM data (%pe)\n", vfid, ERR_PTR(ret));
> +	return ret;
> +}
> +
> +/**
> + * xe_gt_sriov_pf_migration_vram_save() - Save VF VRAM migration data.
> + * @gt: the &xe_gt
> + * @vfid: the VF identifier (can't be 0)
> + *
> + * This function is for PF only.
> + *
> + * Return: 0 on success or a negative error code on failure.
> + */
> +int xe_gt_sriov_pf_migration_vram_save(struct xe_gt *gt, unsigned int vfid)
> +{
> +	xe_gt_assert(gt, IS_SRIOV_PF(gt_to_xe(gt)));
> +	xe_gt_assert(gt, vfid != PFID);
> +	xe_gt_assert(gt, vfid <= xe_sriov_pf_get_totalvfs(gt_to_xe(gt)));
> +
> +	return pf_save_vf_vram_mig_data(gt, vfid);
> +}
> +
> +/**
> + * xe_gt_sriov_pf_migration_vram_restore() - Restore VF VRAM migration data.
> + * @gt: the &xe_gt
> + * @vfid: the VF identifier (can't be 0)
> + *
> + * This function is for PF only.
> + *
> + * Return: 0 on success or a negative error code on failure.
> + */
> +int xe_gt_sriov_pf_migration_vram_restore(struct xe_gt *gt, unsigned int vfid,
> +					  struct xe_sriov_migration_data *data)
> +{
> +	xe_gt_assert(gt, IS_SRIOV_PF(gt_to_xe(gt)));
> +	xe_gt_assert(gt, vfid != PFID);
> +	xe_gt_assert(gt, vfid <= xe_sriov_pf_get_totalvfs(gt_to_xe(gt)));
> +
> +	return pf_restore_vf_vram_mig_data(gt, vfid, data);
> +}
> +
>  /**
>   * xe_gt_sriov_pf_migration_size() - Total size of migration data from all components within a GT.
>   * @gt: the &xe_gt
> @@ -544,6 +746,13 @@ ssize_t xe_gt_sriov_pf_migration_size(struct xe_gt *gt, unsigned int vfid)
>  		size += sizeof(struct xe_sriov_pf_migration_hdr);
>  	total += size;
>  
> +	size = pf_migration_vram_size(gt, vfid);
> +	if (size < 0)
> +		return size;
> +	if (size > 0)
> +		size += sizeof(struct xe_sriov_pf_migration_hdr);
> +	total += size;
> +
>  	return total;
>  }
>  
> @@ -602,6 +811,7 @@ void xe_gt_sriov_pf_migration_save_init(struct xe_gt *gt, unsigned int vfid)
>  	struct xe_gt_sriov_migration_data *migration = pf_pick_gt_migration(gt, vfid);
>  
>  	migration->save.data_remaining = 0;
> +	migration->save.vram_offset = 0;
>  
>  	xe_gt_assert(gt, pf_migration_guc_size(gt, vfid) > 0);
>  	set_bit(XE_SRIOV_MIGRATION_DATA_TYPE_GUC, &migration->save.data_remaining);
> @@ -611,6 +821,9 @@ void xe_gt_sriov_pf_migration_save_init(struct xe_gt *gt, unsigned int vfid)
>  
>  	xe_gt_assert(gt, pf_migration_mmio_size(gt, vfid) > 0);
>  	set_bit(XE_SRIOV_MIGRATION_DATA_TYPE_MMIO, &migration->save.data_remaining);
> +
> +	if (pf_migration_vram_size(gt, vfid) > 0)
> +		set_bit(XE_SRIOV_MIGRATION_DATA_TYPE_VRAM, &migration->save.data_remaining);
>  }
>  
>  /**
> diff --git a/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.h b/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.h
> index b0eec94fea3a6..85b43b2cceb73 100644
> --- a/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.h
> +++ b/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.h
> @@ -25,6 +25,11 @@ int xe_gt_sriov_pf_migration_ggtt_restore(struct xe_gt *gt, unsigned int vfid,
>  int xe_gt_sriov_pf_migration_mmio_save(struct xe_gt *gt, unsigned int vfid);
>  int xe_gt_sriov_pf_migration_mmio_restore(struct xe_gt *gt, unsigned int vfid,
>  					  struct xe_sriov_migration_data *data);
> +int xe_gt_sriov_pf_migration_vram_save(struct xe_gt *gt, unsigned int vfid);
> +int xe_gt_sriov_pf_migration_vram_restore(struct xe_gt *gt, unsigned int vfid,
> +					  struct xe_sriov_migration_data *data);
> +
> +void xe_gt_sriov_pf_migration_save_init(struct xe_gt *gt, unsigned int vfid);
>  
>  ssize_t xe_gt_sriov_pf_migration_size(struct xe_gt *gt, unsigned int vfid);
>  
> diff --git a/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration_types.h b/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration_types.h
> index 9f24878690d9c..f50c64241e9c0 100644
> --- a/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration_types.h
> +++ b/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration_types.h
> @@ -20,6 +20,8 @@ struct xe_gt_sriov_migration_data {
>  	struct {
>  		/** @save.data_remaining: bitmap of migration types that need to be saved */
>  		unsigned long data_remaining;
> +		/** @save.vram_offset: last saved offset within VRAM, used for chunked VRAM save */
> +		loff_t vram_offset;
>  	} save;
>  };
>  
> diff --git a/drivers/gpu/drm/xe/xe_sriov_pf_control.c b/drivers/gpu/drm/xe/xe_sriov_pf_control.c
> index c2768848daba1..aac8ecb861545 100644
> --- a/drivers/gpu/drm/xe/xe_sriov_pf_control.c
> +++ b/drivers/gpu/drm/xe/xe_sriov_pf_control.c
> @@ -5,6 +5,7 @@
>  
>  #include "xe_device.h"
>  #include "xe_gt_sriov_pf_control.h"
> +#include "xe_gt_sriov_pf_migration.h"
>  #include "xe_sriov_migration_data.h"
>  #include "xe_sriov_pf_control.h"
>  #include "xe_sriov_printk.h"
> @@ -171,6 +172,8 @@ int xe_sriov_pf_control_trigger_save_vf(struct xe_device *xe, unsigned int vfid)
>  		return ret;
>  
>  	for_each_gt(gt, xe, id) {
> +		xe_gt_sriov_pf_migration_save_init(gt, vfid);
> +
>  		ret = xe_gt_sriov_pf_control_trigger_save_vf(gt, vfid);
>  		if (ret)
>  			return ret;
> -- 
> 2.50.1
> 

Re: [PATCH v3 05/28] drm/xe/pf: Add data structures and handlers for migration rings

[Michał Winiarski]

On Fri, Oct 31, 2025 at 05:17:26PM +0100, Michal Wajdeczko wrote:
> 
> 
> On 10/30/2025 9:31 PM, Michał Winiarski wrote:
> > Migration data is queued in a per-GT ptr_ring to decouple the worker
> > responsible for handling the data transfer from the .read() and .write()
> > syscalls.
> > Add the data structures and handlers that will be used in future
> > commits.
> > 
> > Signed-off-by: Michał Winiarski <michal.winiarski@intel.com>
> > ---
> >  drivers/gpu/drm/xe/xe_gt_sriov_pf_control.c   | 271 +++++++++++++++++-
> >  drivers/gpu/drm/xe/xe_gt_sriov_pf_control.h   |   6 +-
> >  .../gpu/drm/xe/xe_gt_sriov_pf_control_types.h |  12 +
> >  drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.c | 183 ++++++++++++
> >  drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.h |  14 +
> >  .../drm/xe/xe_gt_sriov_pf_migration_types.h   |  11 +
> >  drivers/gpu/drm/xe/xe_gt_sriov_pf_types.h     |   3 +
> >  drivers/gpu/drm/xe/xe_sriov_pf_migration.c    | 142 +++++++++
> >  drivers/gpu/drm/xe/xe_sriov_pf_migration.h    |   7 +
> >  .../gpu/drm/xe/xe_sriov_pf_migration_types.h  |  58 ++++
> >  drivers/gpu/drm/xe/xe_sriov_pf_types.h        |   3 +
> >  11 files changed, 695 insertions(+), 15 deletions(-)
> > 
> > diff --git a/drivers/gpu/drm/xe/xe_gt_sriov_pf_control.c b/drivers/gpu/drm/xe/xe_gt_sriov_pf_control.c
> > index a571e1c02a3b4..ad9dcd3f56453 100644
> > --- a/drivers/gpu/drm/xe/xe_gt_sriov_pf_control.c
> > +++ b/drivers/gpu/drm/xe/xe_gt_sriov_pf_control.c
> > @@ -19,6 +19,7 @@
> >  #include "xe_guc_ct.h"
> >  #include "xe_sriov.h"
> >  #include "xe_sriov_pf_control.h"
> > +#include "xe_sriov_pf_migration.h"
> >  #include "xe_sriov_pf_service.h"
> >  #include "xe_tile.h"
> >  
> > @@ -185,9 +186,15 @@ static const char *control_bit_to_string(enum xe_gt_sriov_control_bits bit)
> >  	CASE2STR(PAUSE_FAILED);
> >  	CASE2STR(PAUSED);
> >  	CASE2STR(SAVE_WIP);
> > +	CASE2STR(SAVE_PROCESS_DATA);
> > +	CASE2STR(SAVE_WAIT_DATA);
> > +	CASE2STR(SAVE_DATA_DONE);
> >  	CASE2STR(SAVE_FAILED);
> >  	CASE2STR(SAVED);
> >  	CASE2STR(RESTORE_WIP);
> > +	CASE2STR(RESTORE_PROCESS_DATA);
> > +	CASE2STR(RESTORE_WAIT_DATA);
> > +	CASE2STR(RESTORE_DATA_DONE);
> >  	CASE2STR(RESTORE_FAILED);
> >  	CASE2STR(RESTORED);
> >  	CASE2STR(RESUME_WIP);
> > @@ -804,9 +811,50 @@ int xe_gt_sriov_pf_control_resume_vf(struct xe_gt *gt, unsigned int vfid)
> >  	return -ECANCELED;
> >  }
> >  
> > +/**
> > + * DOC: The VF SAVE state machine
> > + *
> > + * SAVE extends the PAUSED state.
> > + *
> > + * The VF SAVE state machine looks like::
> > + *
> > + *  ....PAUSED....................................................
> > + *  :                                                            :
> > + *  :     (O)<---------o                                         :
> > + *  :      |            \                                        :
> > + *  :    save          (SAVED)    (SAVE_FAILED)                  :
> > + *  :      |               ^           ^                         :
> > + *  :      |               |           |                         :
> > + *  :  ....V...............o...........o......SAVE_WIP.........  :
> > + *  :  :   |               |           |                      :  :
> > + *  :  :   |             empty         |                      :  :
> > + *  :  :   |               |           |                      :  :
> > + *  :  :   |               |           |                      :  :
> > + *  :  :   |           DATA_DONE       |                      :  :
> > + *  :  :   |               ^           |                      :  :
> > + *  :  :   |               |        error                     :  :
> > + *  :  :   |            no_data       /                       :  :
> > + *  :  :   |              /          /                        :  :
> > + *  :  :   |             /          /                         :  :
> > + *  :  :   |            /          /                          :  :
> > + *  :  :   o---------->PROCESS_DATA<----consume               :  :
> > + *  :  :                \                      \              :  :
> > + *  :  :                 \                      \             :  :
> > + *  :  :                  \                      \            :  :
> > + *  :  :                   ring_full----->WAIT_DATA           :  :
> > + *  :  :                                                      :  :
> > + *  :  :......................................................:  :
> > + *  :............................................................:
> > + *
> > + * For the full state machine view, see `The VF state machine`_.
> > + */
> 
> nit: since above doc is not strictly for the function below,
> add separation line here

Ok.

> 
> 
> >  static void pf_exit_vf_save_wip(struct xe_gt *gt, unsigned int vfid)
> >  {
> > -	pf_exit_vf_state(gt, vfid, XE_GT_SRIOV_STATE_SAVE_WIP);
> > +	if (pf_exit_vf_state(gt, vfid, XE_GT_SRIOV_STATE_SAVE_WIP)) {
> > +		pf_escape_vf_state(gt, vfid, XE_GT_SRIOV_STATE_SAVE_PROCESS_DATA);
> > +		pf_escape_vf_state(gt, vfid, XE_GT_SRIOV_STATE_SAVE_WAIT_DATA);
> > +		pf_escape_vf_state(gt, vfid, XE_GT_SRIOV_STATE_SAVE_DATA_DONE);
> > +	}
> >  }
> >  
> >  static void pf_enter_vf_saved(struct xe_gt *gt, unsigned int vfid)
> > @@ -821,19 +869,54 @@ static void pf_enter_vf_saved(struct xe_gt *gt, unsigned int vfid)
> >  	pf_exit_vf_wip(gt, vfid);
> >  }
> >  
> > +static void pf_enter_vf_save_failed(struct xe_gt *gt, unsigned int vfid)
> > +{
> > +	pf_enter_vf_state(gt, vfid, XE_GT_SRIOV_STATE_SAVE_FAILED);
> 
> nit: is it ok to do not test if we are already in the FAILED state ?

I'll add:
	if (!pf_enter_vf_state(gt, vfid, XE_GT_SRIOV_STATE_SAVE_FAILED))
		pf_enter_vf_state_machine_bug(gt, vfid);

To both save and restore path.

> 
> > +	pf_exit_vf_wip(gt, vfid);
> > +}
> > +
> > +static int pf_handle_vf_save_data(struct xe_gt *gt, unsigned int vfid)
> > +{
> > +	return 0;
> > +}
> > +
> >  static bool pf_handle_vf_save(struct xe_gt *gt, unsigned int vfid)
> >  {
> > -	if (!pf_exit_vf_state(gt, vfid, XE_GT_SRIOV_STATE_SAVE_WIP))
> > +	int ret;
> > +
> > +	if (!pf_exit_vf_state(gt, vfid, XE_GT_SRIOV_STATE_SAVE_PROCESS_DATA))
> >  		return false;
> >  
> > -	pf_enter_vf_saved(gt, vfid);
> > +	if (xe_gt_sriov_pf_migration_ring_full(gt, vfid)) {
> > +		pf_enter_vf_state(gt, vfid, XE_GT_SRIOV_STATE_SAVE_WAIT_DATA);
> > +
> 
> nit: can remove this sep line

Ok.

> 
> > +		return true;
> > +	}
> > +
> > +	ret = pf_handle_vf_save_data(gt, vfid);
> > +	if (ret == -EAGAIN)
> > +		pf_enter_vf_state(gt, vfid, XE_GT_SRIOV_STATE_SAVE_PROCESS_DATA);
> > +	else if (ret)
> > +		pf_enter_vf_save_failed(gt, vfid);
> > +	else
> > +		pf_enter_vf_state(gt, vfid, XE_GT_SRIOV_STATE_SAVE_DATA_DONE);
> >  
> >  	return true;
> >  }
> >  
> > +static void pf_exit_vf_save_wait_data(struct xe_gt *gt, unsigned int vfid)
> > +{
> > +	if (!pf_exit_vf_state(gt, vfid, XE_GT_SRIOV_STATE_SAVE_WAIT_DATA))
> > +		return;
> > +
> > +	pf_enter_vf_state(gt, vfid, XE_GT_SRIOV_STATE_SAVE_PROCESS_DATA);
> > +	pf_queue_vf(gt, vfid);
> > +}
> > +
> >  static bool pf_enter_vf_save_wip(struct xe_gt *gt, unsigned int vfid)
> >  {
> >  	if (pf_enter_vf_state(gt, vfid, XE_GT_SRIOV_STATE_SAVE_WIP)) {
> > +		pf_enter_vf_state(gt, vfid, XE_GT_SRIOV_STATE_SAVE_PROCESS_DATA);
> >  		pf_enter_vf_wip(gt, vfid);
> >  		pf_queue_vf(gt, vfid);
> >  		return true;
> > @@ -842,6 +925,32 @@ static bool pf_enter_vf_save_wip(struct xe_gt *gt, unsigned int vfid)
> >  	return false;
> >  }
> >  
> > +/**
> > + * xe_gt_sriov_pf_control_check_save_data_done() - Check if all save migration data was produced.
> > + * @gt: the &xe_gt
> > + * @vfid: the VF identifier
> > + *
> > + * This function is for PF only.
> > + *
> > + * Return: 0 on success or a negative error code on failure.
> 
> but it's defined as a bool below

Ok.

> 
> > + */
> > +bool xe_gt_sriov_pf_control_check_save_data_done(struct xe_gt *gt, unsigned int vfid)
> > +{
> > +	return pf_check_vf_state(gt, vfid, XE_GT_SRIOV_STATE_SAVE_DATA_DONE);
> > +}
> > +
> > +/**
> > + * xe_gt_sriov_pf_control_process_save_data() - Queue VF save migration data processing.
> > + * @gt: the &xe_gt
> > + * @vfid: the VF identifier
> > + *
> > + * This function is for PF only.
> > + */
> > +void xe_gt_sriov_pf_control_process_save_data(struct xe_gt *gt, unsigned int vfid)
> > +{
> > +	pf_exit_vf_save_wait_data(gt, vfid);
> 
> ok to have no check against unexpected state ?

It is OK, as the worker can go into SAVE_FAILED state at any point in
time, so there's not much we can check here.

However, we're missing FAILURE state checks on data consume (for save)
and produce (for restore) path.
I'll add it here (and in other relevant paths).

> 
> > +}
> > +
> >  /**
> >   * xe_gt_sriov_pf_control_trigger_save_vf() - Start an SR-IOV VF migration data save sequence.
> >   * @gt: the &xe_gt
> > @@ -887,19 +996,62 @@ int xe_gt_sriov_pf_control_trigger_save_vf(struct xe_gt *gt, unsigned int vfid)
> >   */
> >  int xe_gt_sriov_pf_control_finish_save_vf(struct xe_gt *gt, unsigned int vfid)
> >  {
> > -	if (!pf_expect_vf_state(gt, vfid, XE_GT_SRIOV_STATE_SAVED)) {
> > -		pf_enter_vf_mismatch(gt, vfid);
> > +	if (!pf_check_vf_state(gt, vfid, XE_GT_SRIOV_STATE_SAVE_DATA_DONE)) {
> > +		xe_gt_sriov_err(gt, "VF%u save is still in progress!\n", vfid);
> >  		return -EIO;
> >  	}
> >  
> >  	pf_expect_vf_state(gt, vfid, XE_GT_SRIOV_STATE_PAUSED);
> > +	pf_exit_vf_state(gt, vfid, XE_GT_SRIOV_STATE_SAVE_DATA_DONE);
> > +	pf_enter_vf_saved(gt, vfid);
> >  
> >  	return 0;
> >  }
> >  
> > +/**
> > + * DOC: The VF RESTORE state machine
> > + *
> > + * RESTORE extends the PAUSED state.
> > + *
> > + * The VF RESTORE state machine looks like::
> > + *
> > + *  ....PAUSED....................................................
> > + *  :                                                            :
> > + *  :     (O)<---------o                                         :
> > + *  :      |            \                                        :
> > + *  :    restore      (RESTORED)  (RESTORE_FAILED)               :
> > + *  :      |               ^           ^                         :
> > + *  :      |               |           |                         :
> > + *  :  ....V...............o...........o......RESTORE_WIP......  :
> > + *  :  :   |               |           |                      :  :
> > + *  :  :   |             empty         |                      :  :
> > + *  :  :   |               |           |                      :  :
> > + *  :  :   |               |           |                      :  :
> > + *  :  :   |           DATA_DONE       |                      :  :
> > + *  :  :   |               ^           |                      :  :
> > + *  :  :   |               |        error                     :  :
> > + *  :  :   |           trailer        /                       :  :
> > + *  :  :   |              /          /                        :  :
> > + *  :  :   |             /          /                         :  :
> > + *  :  :   |            /          /                          :  :
> > + *  :  :   o---------->PROCESS_DATA<----produce               :  :
> > + *  :  :                \                      \              :  :
> > + *  :  :                 \                      \             :  :
> > + *  :  :                  \                      \            :  :
> > + *  :  :                   ring_empty---->WAIT_DATA           :  :
> > + *  :  :                                                      :  :
> > + *  :  :......................................................:  :
> > + *  :............................................................:
> > + *
> > + * For the full state machine view, see `The VF state machine`_.
> > + */
> 
> nit: like above, add separation line here

Ok.

> 
> >  static void pf_exit_vf_restore_wip(struct xe_gt *gt, unsigned int vfid)
> >  {
> > -	pf_exit_vf_state(gt, vfid, XE_GT_SRIOV_STATE_RESTORE_WIP);
> > +	if (pf_exit_vf_state(gt, vfid, XE_GT_SRIOV_STATE_RESTORE_WIP)) {
> > +		pf_escape_vf_state(gt, vfid, XE_GT_SRIOV_STATE_RESTORE_PROCESS_DATA);
> > +		pf_escape_vf_state(gt, vfid, XE_GT_SRIOV_STATE_RESTORE_WAIT_DATA);
> > +		pf_escape_vf_state(gt, vfid, XE_GT_SRIOV_STATE_RESTORE_DATA_DONE);
> > +	}
> >  }
> >  
> >  static void pf_enter_vf_restored(struct xe_gt *gt, unsigned int vfid)
> > @@ -914,19 +1066,61 @@ static void pf_enter_vf_restored(struct xe_gt *gt, unsigned int vfid)
> >  	pf_exit_vf_wip(gt, vfid);
> >  }
> >  
> > +static void pf_enter_vf_restore_failed(struct xe_gt *gt, unsigned int vfid)
> > +{
> > +	pf_enter_vf_state(gt, vfid, XE_GT_SRIOV_STATE_RESTORE_FAILED);
> > +	pf_exit_vf_wip(gt, vfid);
> > +}
> > +
> > +static int pf_handle_vf_restore_data(struct xe_gt *gt, unsigned int vfid)
> > +{
> > +	struct xe_sriov_migration_data *data = xe_gt_sriov_pf_migration_restore_consume(gt, vfid);
> > +
> > +	xe_gt_assert(gt, data);
> 
> maybe assert is redundant here, as if data is null, we will get NPD crash call
> stack from the below line anyway?

Ok.

> 
> > +
> > +	xe_gt_sriov_notice(gt, "Skipping VF%u unknown data type: %d\n", vfid, data->type);
> > +
> > +	return 0;
> > +}
> > +
> >  static bool pf_handle_vf_restore(struct xe_gt *gt, unsigned int vfid)
> >  {
> > -	if (!pf_exit_vf_state(gt, vfid, XE_GT_SRIOV_STATE_RESTORE_WIP))
> > +	int ret;
> > +
> > +	if (!pf_exit_vf_state(gt, vfid, XE_GT_SRIOV_STATE_RESTORE_PROCESS_DATA))
> >  		return false;
> >  
> > -	pf_enter_vf_restored(gt, vfid);
> > +	if (xe_gt_sriov_pf_migration_ring_empty(gt, vfid)) {
> > +		if (pf_exit_vf_state(gt, vfid, XE_GT_SRIOV_STATE_RESTORE_DATA_DONE))
> > +			pf_enter_vf_restored(gt, vfid);
> > +		else
> > +			pf_enter_vf_state(gt, vfid, XE_GT_SRIOV_STATE_RESTORE_WAIT_DATA);
> > +
> > +		return true;
> > +	}
> > +
> > +	ret = pf_handle_vf_restore_data(gt, vfid);
> > +	if (ret)
> > +		pf_enter_vf_restore_failed(gt, vfid);
> > +	else
> > +		pf_enter_vf_state(gt, vfid, XE_GT_SRIOV_STATE_RESTORE_PROCESS_DATA);
> >  
> >  	return true;
> >  }
> >  
> > +static void pf_exit_vf_restore_wait_data(struct xe_gt *gt, unsigned int vfid)
> > +{
> > +	if (!pf_exit_vf_state(gt, vfid, XE_GT_SRIOV_STATE_RESTORE_WAIT_DATA))
> > +		return;
> > +
> > +	pf_enter_vf_state(gt, vfid, XE_GT_SRIOV_STATE_RESTORE_PROCESS_DATA);
> > +	pf_queue_vf(gt, vfid);
> > +}
> > +
> >  static bool pf_enter_vf_restore_wip(struct xe_gt *gt, unsigned int vfid)
> >  {
> >  	if (pf_enter_vf_state(gt, vfid, XE_GT_SRIOV_STATE_RESTORE_WIP)) {
> > +		pf_enter_vf_state(gt, vfid, XE_GT_SRIOV_STATE_RESTORE_PROCESS_DATA);
> >  		pf_enter_vf_wip(gt, vfid);
> >  		pf_queue_vf(gt, vfid);
> >  		return true;
> > @@ -935,6 +1129,47 @@ static bool pf_enter_vf_restore_wip(struct xe_gt *gt, unsigned int vfid)
> >  	return false;
> >  }
> >  
> > +/**
> > + * xe_gt_sriov_pf_control_restore_data_done() - Indicate the end of VF migration data stream.
> > + * @gt: the &xe_gt
> > + * @vfid: the VF identifier
> > + *
> > + * This function is for PF only.
> > + *
> > + * Return: 0 on success or a negative error code on failure.
> 
> save_data_done() is bool
> and below you return only 0 or -EIO
> maybe it could be bool too?

It's check_save_data_done(), which is why it is bool.
Here we're telling the worker to go into data_done state.

> 
> > + */
> > +int xe_gt_sriov_pf_control_restore_data_done(struct xe_gt *gt, unsigned int vfid)
> > +{
> > +	if (!pf_expect_vf_state(gt, vfid, XE_GT_SRIOV_STATE_RESTORE_WIP)) {
> > +		pf_enter_vf_state_machine_bug(gt, vfid);
> > +		return -EIO;
> > +	}
> > +
> > +	if (!pf_enter_vf_state(gt, vfid, XE_GT_SRIOV_STATE_RESTORE_DATA_DONE)) {
> > +		pf_enter_vf_state_machine_bug(gt, vfid);
> > +		return -EIO;
> > +	}
> > +
> > +	xe_gt_sriov_pf_control_process_restore_data(gt, vfid);
> > +
> > +	return 0;
> > +}
> > +
> > +/**
> > + * xe_gt_sriov_pf_control_process_restore_data() - Queue VF restore migration data processing.
> > + * @gt: the &xe_gt
> > + * @vfid: the VF identifier
> > + *
> > + * This function is for PF only.
> > + */
> > +void xe_gt_sriov_pf_control_process_restore_data(struct xe_gt *gt, unsigned int vfid)
> > +{
> > +	if (!pf_expect_vf_state(gt, vfid, XE_GT_SRIOV_STATE_RESTORE_WIP))
> > +		pf_enter_vf_state_machine_bug(gt, vfid);
> > +
> > +	pf_exit_vf_restore_wait_data(gt, vfid);
> > +}
> > +
> >  /**
> >   * xe_gt_sriov_pf_control_trigger restore_vf() - Start an SR-IOV VF migration data restore sequence.
> >   * @gt: the &xe_gt
> > @@ -1000,11 +1235,9 @@ int xe_gt_sriov_pf_control_finish_restore_vf(struct xe_gt *gt, unsigned int vfid
> >  {
> >  	int ret;
> >  
> > -	if (pf_check_vf_state(gt, vfid, XE_GT_SRIOV_STATE_RESTORE_WIP)) {
> > -		ret = pf_wait_vf_restore_done(gt, vfid);
> > -		if (ret)
> > -			return ret;
> > -	}
> > +	ret = pf_wait_vf_restore_done(gt, vfid);
> > +	if (ret)
> > +		return ret;
> >  
> >  	if (!pf_expect_vf_state(gt, vfid, XE_GT_SRIOV_STATE_RESTORED)) {
> >  		pf_enter_vf_mismatch(gt, vfid);
> > @@ -1705,9 +1938,21 @@ static bool pf_process_vf_state_machine(struct xe_gt *gt, unsigned int vfid)
> >  	if (pf_exit_vf_pause_save_guc(gt, vfid))
> >  		return true;
> >  
> > +	if (pf_check_vf_state(gt, vfid, XE_GT_SRIOV_STATE_SAVE_WAIT_DATA)) {
> > +		xe_gt_sriov_dbg_verbose(gt, "VF%u in %s\n", vfid,
> > +					control_bit_to_string(XE_GT_SRIOV_STATE_SAVE_WAIT_DATA));
> > +		return false;
> > +	}
> > +
> >  	if (pf_handle_vf_save(gt, vfid))
> >  		return true;
> >  
> > +	if (pf_check_vf_state(gt, vfid, XE_GT_SRIOV_STATE_RESTORE_WAIT_DATA)) {
> > +		xe_gt_sriov_dbg_verbose(gt, "VF%u in %s\n", vfid,
> > +					control_bit_to_string(XE_GT_SRIOV_STATE_RESTORE_WAIT_DATA));
> > +		return false;
> > +	}
> > +
> >  	if (pf_handle_vf_restore(gt, vfid))
> >  		return true;
> >  
> > diff --git a/drivers/gpu/drm/xe/xe_gt_sriov_pf_control.h b/drivers/gpu/drm/xe/xe_gt_sriov_pf_control.h
> > index abc233f6302ed..6b1ab339e3b73 100644
> > --- a/drivers/gpu/drm/xe/xe_gt_sriov_pf_control.h
> > +++ b/drivers/gpu/drm/xe/xe_gt_sriov_pf_control.h
> > @@ -14,12 +14,14 @@ struct xe_gt;
> >  int xe_gt_sriov_pf_control_init(struct xe_gt *gt);
> >  void xe_gt_sriov_pf_control_restart(struct xe_gt *gt);
> >  
> > -bool xe_gt_sriov_pf_control_check_vf_data_wip(struct xe_gt *gt, unsigned int vfid);
> > -
> >  int xe_gt_sriov_pf_control_pause_vf(struct xe_gt *gt, unsigned int vfid);
> >  int xe_gt_sriov_pf_control_resume_vf(struct xe_gt *gt, unsigned int vfid);
> > +bool xe_gt_sriov_pf_control_check_save_data_done(struct xe_gt *gt, unsigned int vfid);
> > +void xe_gt_sriov_pf_control_process_save_data(struct xe_gt *gt, unsigned int vfid);
> >  int xe_gt_sriov_pf_control_trigger_save_vf(struct xe_gt *gt, unsigned int vfid);
> >  int xe_gt_sriov_pf_control_finish_save_vf(struct xe_gt *gt, unsigned int vfid);
> > +int xe_gt_sriov_pf_control_restore_data_done(struct xe_gt *gt, unsigned int vfid);
> > +void xe_gt_sriov_pf_control_process_restore_data(struct xe_gt *gt, unsigned int vfid);
> >  int xe_gt_sriov_pf_control_trigger_restore_vf(struct xe_gt *gt, unsigned int vfid);
> >  int xe_gt_sriov_pf_control_finish_restore_vf(struct xe_gt *gt, unsigned int vfid);
> >  int xe_gt_sriov_pf_control_stop_vf(struct xe_gt *gt, unsigned int vfid);
> > diff --git a/drivers/gpu/drm/xe/xe_gt_sriov_pf_control_types.h b/drivers/gpu/drm/xe/xe_gt_sriov_pf_control_types.h
> > index ea16ceb39dabb..a14c738581ae2 100644
> > --- a/drivers/gpu/drm/xe/xe_gt_sriov_pf_control_types.h
> > +++ b/drivers/gpu/drm/xe/xe_gt_sriov_pf_control_types.h
> > @@ -32,9 +32,15 @@
> >   * @XE_GT_SRIOV_STATE_PAUSE_FAILED: indicates that a VF pause operation has failed.
> >   * @XE_GT_SRIOV_STATE_PAUSED: indicates that the VF is paused.
> >   * @XE_GT_SRIOV_STATE_SAVE_WIP: indicates that VF save operation is in progress.
> > + * @XE_GT_SRIOV_STATE_SAVE_PROCESS_DATA: indicates that VF migration data is being produced.
> > + * @XE_GT_SRIOV_STATE_SAVE_WAIT_DATA: indicates that PF awaits for space in migration data ring.
> > + * @XE_GT_SRIOV_STATE_SAVE_DATA_DONE: indicates that all migration data was produced by Xe.
> >   * @XE_GT_SRIOV_STATE_SAVE_FAILED: indicates that VF save operation has failed.
> >   * @XE_GT_SRIOV_STATE_SAVED: indicates that VF data is saved.
> >   * @XE_GT_SRIOV_STATE_RESTORE_WIP: indicates that VF restore operation is in progress.
> > + * @XE_GT_SRIOV_STATE_RESTORE_PROCESS_DATA: indicates that VF migration data is being consumed.
> > + * @XE_GT_SRIOV_STATE_RESTORE_WAIT_DATA: indicates that PF awaits for data in migration data ring.
> > + * @XE_GT_SRIOV_STATE_RESTORE_DATA_DONE: indicates that all migration data was produced by the user.
> >   * @XE_GT_SRIOV_STATE_RESTORE_FAILED: indicates that VF restore operation has failed.
> >   * @XE_GT_SRIOV_STATE_RESTORED: indicates that VF data is restored.
> >   * @XE_GT_SRIOV_STATE_RESUME_WIP: indicates the a VF resume operation is in progress.
> > @@ -70,10 +76,16 @@ enum xe_gt_sriov_control_bits {
> >  	XE_GT_SRIOV_STATE_PAUSED,
> >  
> >  	XE_GT_SRIOV_STATE_SAVE_WIP,
> > +	XE_GT_SRIOV_STATE_SAVE_PROCESS_DATA,
> > +	XE_GT_SRIOV_STATE_SAVE_WAIT_DATA,
> > +	XE_GT_SRIOV_STATE_SAVE_DATA_DONE,
> >  	XE_GT_SRIOV_STATE_SAVE_FAILED,
> >  	XE_GT_SRIOV_STATE_SAVED,
> >  
> >  	XE_GT_SRIOV_STATE_RESTORE_WIP,
> > +	XE_GT_SRIOV_STATE_RESTORE_PROCESS_DATA,
> > +	XE_GT_SRIOV_STATE_RESTORE_WAIT_DATA,
> > +	XE_GT_SRIOV_STATE_RESTORE_DATA_DONE,
> >  	XE_GT_SRIOV_STATE_RESTORE_FAILED,
> >  	XE_GT_SRIOV_STATE_RESTORED,
> >  
> > diff --git a/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.c b/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.c
> > index ca28f45aaf481..e61e6f4215864 100644
> > --- a/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.c
> > +++ b/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.c
> > @@ -7,6 +7,7 @@
> >  
> >  #include "abi/guc_actions_sriov_abi.h"
> >  #include "xe_bo.h"
> > +#include "xe_gt_sriov_pf_control.h"
> >  #include "xe_gt_sriov_pf_helpers.h"
> >  #include "xe_gt_sriov_pf_migration.h"
> >  #include "xe_gt_sriov_printk.h"
> > @@ -15,6 +16,17 @@
> >  #include "xe_sriov.h"
> >  #include "xe_sriov_pf_migration.h"
> >  
> > +#define XE_GT_SRIOV_PF_MIGRATION_RING_SIZE 5
> > +
> > +static struct xe_gt_sriov_migration_data *pf_pick_gt_migration(struct xe_gt *gt, unsigned int vfid)
> > +{
> > +	xe_gt_assert(gt, IS_SRIOV_PF(gt_to_xe(gt)));
> > +	xe_gt_assert(gt, vfid != PFID);
> > +	xe_gt_assert(gt, vfid <= xe_sriov_pf_get_totalvfs(gt_to_xe(gt)));
> > +
> > +	return &gt->sriov.pf.vfs[vfid].migration;
> > +}
> > +
> >  /* Return: number of dwords saved/restored/required or a negative error code on failure */
> >  static int guc_action_vf_save_restore(struct xe_guc *guc, u32 vfid, u32 opcode,
> >  				      u64 addr, u32 ndwords)
> > @@ -382,6 +394,162 @@ ssize_t xe_gt_sriov_pf_migration_write_guc_state(struct xe_gt *gt, unsigned int
> >  }
> >  #endif /* CONFIG_DEBUG_FS */
> >  
> > +/**
> > + * xe_gt_sriov_pf_migration_ring_empty() - Check if a migration ring is empty.
> > + * @gt: the &xe_gt
> > + * @vfid: the VF identifier
> > + *
> > + * Return: true if the ring is empty, otherwise false.
> > + */
> > +bool xe_gt_sriov_pf_migration_ring_empty(struct xe_gt *gt, unsigned int vfid)
> > +{
> > +	return ptr_ring_empty(&pf_pick_gt_migration(gt, vfid)->ring);
> > +}
> > +
> > +/**
> > + * xe_gt_sriov_pf_migration_ring_full() - Check if a migration ring is full.
> > + * @gt: the &xe_gt
> > + * @vfid: the VF identifier
> > + *
> > + * Return: true if the ring is full, otherwise false.
> > + */
> > +bool xe_gt_sriov_pf_migration_ring_full(struct xe_gt *gt, unsigned int vfid)
> > +{
> > +	return ptr_ring_full(&pf_pick_gt_migration(gt, vfid)->ring);
> > +}
> > +
> > +/**
> > + * xe_gt_sriov_pf_migration_save_produce() - Add VF save data packet to migration ring.
> > + * @gt: the &xe_gt
> > + * @vfid: the VF identifier
> > + * @data: &xe_sriov_migration_data packet
> > + *
> > + * Called by the save migration data producer (PF SR-IOV Control worker) when
> > + * processing migration data.
> > + * Wakes up the save migration data consumer (userspace), that is potentially
> > + * waiting for data when the ring is empty.
> 
> ... was empty ?

Ok.

> 
> > + *
> > + * Return: 0 on success or a negative error code on failure.
> > + */
> > +int xe_gt_sriov_pf_migration_save_produce(struct xe_gt *gt, unsigned int vfid,
> > +					  struct xe_sriov_migration_data *data)
> > +{
> > +	int ret;
> > +
> > +	ret = ptr_ring_produce(&pf_pick_gt_migration(gt, vfid)->ring, data);
> > +	if (ret)
> > +		return ret;
> > +
> > +	wake_up_all(xe_sriov_pf_migration_waitqueue(gt_to_xe(gt), vfid));
> > +
> > +	return 0;
> > +}
> > +
> > +/**
> > + * xe_gt_sriov_pf_migration_restore_consume() - Get VF restore data packet from migration ring.
> > + * @gt: the &xe_gt
> > + * @vfid: the VF identifier
> > + *
> > + * Called by the restore migration data consumer (PF SR-IOV Control worker) when
> > + * processing migration data.
> > + * Wakes up the restore migration data producer (userspace), that is
> > + * potentially waiting to add more data when the ring is full.
> > + *
> > + * Return: Pointer to &struct xe_sriov_migration_data on success,
> > + *	   NULL if ring is empty.
> > + */
> > +struct xe_sriov_migration_data *
> > +xe_gt_sriov_pf_migration_restore_consume(struct xe_gt *gt, unsigned int vfid)
> > +{
> > +	struct xe_gt_sriov_migration_data *migration = pf_pick_gt_migration(gt, vfid);
> > +	struct wait_queue_head *wq = xe_sriov_pf_migration_waitqueue(gt_to_xe(gt), vfid);
> > +	struct xe_sriov_migration_data *data;
> > +
> > +	data = ptr_ring_consume(&migration->ring);
> > +	if (data)
> > +		wake_up_all(wq);
> > +
> > +	return data;
> > +}
> > +
> > +/**
> > + * xe_gt_sriov_pf_migration_restore_produce() - Add VF restore data packet to migration ring.
> > + * @gt: the &xe_gt
> > + * @vfid: the VF identifier
> > + * @data: &xe_sriov_migration_data packet
> > + *
> > + * Called by the restore migration data producer (userspace) when processing
> > + * migration data.
> > + * If the ring is full, waits until there is space.
> > + * Queues the restore migration data consumer (PF SR-IOV Control worker), that
> > + * is potentially waiting for data when the ring is empty.
> > + *
> > + * Return: 0 on success or a negative error code on failure.
> > + */
> > +int xe_gt_sriov_pf_migration_restore_produce(struct xe_gt *gt, unsigned int vfid,
> > +					     struct xe_sriov_migration_data *data)
> > +{
> > +	struct wait_queue_head *wq = xe_sriov_pf_migration_waitqueue(gt_to_xe(gt), vfid);
> > +	struct xe_gt_sriov_migration_data *migration = pf_pick_gt_migration(gt, vfid);
> > +	int ret;
> > +
> > +	xe_gt_assert(gt, data->tile == gt->tile->id);
> > +	xe_gt_assert(gt, data->gt == gt->info.id);
> > +
> > +	for (;;) {
> > +		ret = ptr_ring_produce(&migration->ring, data);
> > +		if (!ret)
> > +			break;
> > +
> > +		ret = wait_event_interruptible(*wq, !ptr_ring_full(&migration->ring));
> > +		if (ret)
> > +			return ret;
> > +	}
> > +
> > +	xe_gt_sriov_pf_control_process_restore_data(gt, vfid);
> > +
> > +	return 0;
> > +}
> > +
> > +/**
> > + * xe_gt_sriov_pf_migration_save_consume() - Get VF save data packet from migration ring.
> > + * @gt: the &xe_gt
> > + * @vfid: the VF identifier
> > + *
> > + * Called by the save migration data consumer (userspace) when
> > + * processing migration data.
> > + * Queues the save migration data producer (PF SR-IOV Control worker), that is
> > + * potentially waiting to add more data when the ring is full.
> > + *
> > + * Return: Pointer to &struct xe_sriov_migration_data on success,
> > + *	   NULL if ring is empty and there's no more data available,
> > + *	   ERR_PTR(-EAGAIN) if the ring is empty, but data is still produced.
> > + */
> > +struct xe_sriov_migration_data *
> > +xe_gt_sriov_pf_migration_save_consume(struct xe_gt *gt, unsigned int vfid)
> > +{
> > +	struct xe_gt_sriov_migration_data *migration = pf_pick_gt_migration(gt, vfid);
> > +	struct xe_sriov_migration_data *data;
> > +
> > +	data = ptr_ring_consume(&migration->ring);
> > +	if (data) {
> > +		xe_gt_sriov_pf_control_process_save_data(gt, vfid);
> > +		return data;
> > +	}
> > +
> > +	if (xe_gt_sriov_pf_control_check_save_data_done(gt, vfid))
> > +		return NULL;
> > +
> > +	return ERR_PTR(-EAGAIN);
> > +}
> > +
> > +static void action_ring_cleanup(void *arg)
> > +{
> > +	struct ptr_ring *r = arg;
> > +
> > +	ptr_ring_cleanup(r, NULL);
> > +}
> > +
> >  /**
> >   * xe_gt_sriov_pf_migration_init() - Initialize support for VF migration.
> >   * @gt: the &xe_gt
> > @@ -393,6 +561,7 @@ ssize_t xe_gt_sriov_pf_migration_write_guc_state(struct xe_gt *gt, unsigned int
> >  int xe_gt_sriov_pf_migration_init(struct xe_gt *gt)
> >  {
> >  	struct xe_device *xe = gt_to_xe(gt);
> > +	unsigned int n, totalvfs;
> >  	int err;
> >  
> >  	xe_gt_assert(gt, IS_SRIOV_PF(xe));
> > @@ -404,5 +573,19 @@ int xe_gt_sriov_pf_migration_init(struct xe_gt *gt)
> >  	if (err)
> >  		return err;
> >  
> > +	totalvfs = xe_sriov_pf_get_totalvfs(xe);
> > +	for (n = 1; n <= totalvfs; n++) {
> > +		struct xe_gt_sriov_migration_data *migration = pf_pick_gt_migration(gt, n);
> > +
> > +		err = ptr_ring_init(&migration->ring,
> > +				    XE_GT_SRIOV_PF_MIGRATION_RING_SIZE, GFP_KERNEL);
> > +		if (err)
> > +			return err;
> > +
> > +		err = devm_add_action_or_reset(xe->drm.dev, action_ring_cleanup, &migration->ring);
> > +		if (err)
> > +			return err;
> > +	}
> > +
> >  	return 0;
> >  }
> > diff --git a/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.h b/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.h
> > index 09faeae00ddbb..9e67f18ded205 100644
> > --- a/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.h
> > +++ b/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.h
> > @@ -9,11 +9,25 @@
> >  #include <linux/types.h>
> >  
> >  struct xe_gt;
> > +struct xe_sriov_migration_data;
> >  
> >  int xe_gt_sriov_pf_migration_init(struct xe_gt *gt);
> >  int xe_gt_sriov_pf_migration_save_guc_state(struct xe_gt *gt, unsigned int vfid);
> >  int xe_gt_sriov_pf_migration_restore_guc_state(struct xe_gt *gt, unsigned int vfid);
> >  
> > +bool xe_gt_sriov_pf_migration_ring_empty(struct xe_gt *gt, unsigned int vfid);
> > +bool xe_gt_sriov_pf_migration_ring_full(struct xe_gt *gt, unsigned int vfid);
> > +
> > +int xe_gt_sriov_pf_migration_save_produce(struct xe_gt *gt, unsigned int vfid,
> > +					  struct xe_sriov_migration_data *data);
> > +struct xe_sriov_migration_data *
> > +xe_gt_sriov_pf_migration_restore_consume(struct xe_gt *gt, unsigned int vfid);
> > +
> > +int xe_gt_sriov_pf_migration_restore_produce(struct xe_gt *gt, unsigned int vfid,
> > +					     struct xe_sriov_migration_data *data);
> > +struct xe_sriov_migration_data *
> > +xe_gt_sriov_pf_migration_save_consume(struct xe_gt *gt, unsigned int vfid);
> > +
> >  #ifdef CONFIG_DEBUG_FS
> >  ssize_t xe_gt_sriov_pf_migration_read_guc_state(struct xe_gt *gt, unsigned int vfid,
> >  						char __user *buf, size_t count, loff_t *pos);
> > diff --git a/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration_types.h b/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration_types.h
> > index 9d672feac5f04..84be6fac16c8b 100644
> > --- a/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration_types.h
> > +++ b/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration_types.h
> > @@ -7,6 +7,7 @@
> >  #define _XE_GT_SRIOV_PF_MIGRATION_TYPES_H_
> >  
> >  #include <linux/mutex.h>
> > +#include <linux/ptr_ring.h>
> >  #include <linux/types.h>
> >  
> >  /**
> > @@ -24,6 +25,16 @@ struct xe_gt_sriov_state_snapshot {
> >  	} guc;
> >  };
> >  
> > +/**
> > + * struct xe_gt_sriov_migration_data - GT-level per-VF migration data.
> > + *
> > + * Used by the PF driver to maintain per-VF migration data.
> > + */
> > +struct xe_gt_sriov_migration_data {
> > +	/** @ring: queue containing VF save / restore migration data */
> > +	struct ptr_ring ring;
> > +};
> > +
> >  /**
> >   * struct xe_gt_sriov_pf_migration - GT-level data.
> >   *
> > diff --git a/drivers/gpu/drm/xe/xe_gt_sriov_pf_types.h b/drivers/gpu/drm/xe/xe_gt_sriov_pf_types.h
> > index a64a6835ad656..812e74d3f8f80 100644
> > --- a/drivers/gpu/drm/xe/xe_gt_sriov_pf_types.h
> > +++ b/drivers/gpu/drm/xe/xe_gt_sriov_pf_types.h
> > @@ -33,6 +33,9 @@ struct xe_gt_sriov_metadata {
> >  
> >  	/** @snapshot: snapshot of the VF state data */
> >  	struct xe_gt_sriov_state_snapshot snapshot;
> > +
> > +	/** @migration: per-VF migration data. */
> > +	struct xe_gt_sriov_migration_data migration;
> >  };
> >  
> >  /**
> > diff --git a/drivers/gpu/drm/xe/xe_sriov_pf_migration.c b/drivers/gpu/drm/xe/xe_sriov_pf_migration.c
> > index 8c523c392f98b..7be9f026d80e8 100644
> > --- a/drivers/gpu/drm/xe/xe_sriov_pf_migration.c
> > +++ b/drivers/gpu/drm/xe/xe_sriov_pf_migration.c
> > @@ -3,8 +3,36 @@
> >   * Copyright © 2025 Intel Corporation
> >   */
> >  
> > +#include <drm/drm_managed.h>
> > +
> > +#include "xe_device.h"
> > +#include "xe_gt_sriov_pf_control.h"
> > +#include "xe_gt_sriov_pf_migration.h"
> > +#include "xe_pm.h"
> >  #include "xe_sriov.h"
> > +#include "xe_sriov_pf_helpers.h"
> >  #include "xe_sriov_pf_migration.h"
> > +#include "xe_sriov_printk.h"
> > +
> > +static struct xe_sriov_pf_migration *pf_pick_migration(struct xe_device *xe, unsigned int vfid)
> > +{
> > +	xe_assert(xe, IS_SRIOV_PF(xe));
> > +	xe_assert(xe, vfid <= xe_sriov_pf_get_totalvfs(xe));
> > +
> > +	return &xe->sriov.pf.vfs[vfid].migration;
> > +}
> > +
> > +/**
> > + * xe_sriov_pf_migration_waitqueue - Get waitqueue for migration.
> > + * @xe: the &xe_device
> > + * @vfid: the VF identifier
> > + *
> > + * Return: pointer to the migration waitqueue.
> > + */
> > +wait_queue_head_t *xe_sriov_pf_migration_waitqueue(struct xe_device *xe, unsigned int vfid)
> > +{
> > +	return &pf_pick_migration(xe, vfid)->wq;
> > +}
> >  
> >  /**
> >   * xe_sriov_pf_migration_supported() - Check if SR-IOV VF migration is supported by the device
> > @@ -33,9 +61,123 @@ static bool pf_check_migration_support(struct xe_device *xe)
> >   */
> >  int xe_sriov_pf_migration_init(struct xe_device *xe)
> >  {
> > +	unsigned int n, totalvfs;
> > +
> >  	xe_assert(xe, IS_SRIOV_PF(xe));
> >  
> >  	xe->sriov.pf.migration.supported = pf_check_migration_support(xe);
> > +	if (!xe_sriov_pf_migration_supported(xe))
> > +		return 0;
> > +
> > +	totalvfs = xe_sriov_pf_get_totalvfs(xe);
> > +	for (n = 1; n <= totalvfs; n++) {
> > +		struct xe_sriov_pf_migration *migration = pf_pick_migration(xe, n);
> > +
> > +		init_waitqueue_head(&migration->wq);
> > +	}
> >  
> >  	return 0;
> >  }
> > +
> > +static bool pf_migration_data_ready(struct xe_device *xe, unsigned int vfid)
> > +{
> > +	struct xe_gt *gt;
> > +	u8 gt_id;
> > +
> > +	for_each_gt(gt, xe, gt_id) {
> > +		if (!xe_gt_sriov_pf_migration_ring_empty(gt, vfid) ||
> > +		    xe_gt_sriov_pf_control_check_save_data_done(gt, vfid))
> > +			return true;
> > +	}
> > +
> > +	return false;
> > +}
> > +
> > +static struct xe_sriov_migration_data *
> > +pf_migration_consume(struct xe_device *xe, unsigned int vfid)
> > +{
> > +	struct xe_sriov_migration_data *data;
> > +	struct xe_gt *gt;
> > +	u8 gt_id;
> > +	bool more_data = false;
> 
> nit: vars in rev xmas tree order?

Ok.

> 
> > +
> > +	for_each_gt(gt, xe, gt_id) {
> > +		data = xe_gt_sriov_pf_migration_save_consume(gt, vfid);
> > +		if (data && PTR_ERR(data) != EAGAIN)
> > +			return data;
> > +		if (PTR_ERR(data) == -EAGAIN)
> > +			more_data = true;
> > +	}
> > +
> > +	if (!more_data)
> > +		return NULL;
> > +
> > +	return ERR_PTR(-EAGAIN);
> > +}
> > +
> > +/**
> > + * xe_sriov_pf_migration_save_consume() - Consume a VF migration data packet from the device.
> > + * @xe: the &xe_device
> > + * @vfid: the VF identifier
> > + *
> > + * Called by the save migration data consumer (userspace) when
> > + * processing migration data.
> > + * If there is no migration data to process, wait until more data is available.
> > + *
> > + * Return: Pointer to &xe_sriov_migration_data on success,
> > + *	   NULL if ring is empty and no more migration data is expected,
> > + *	   ERR_PTR value in case of error.
> > + *
> > + * Return: 0 on success or a negative error code on failure.
> > + */
> > +struct xe_sriov_migration_data *
> > +xe_sriov_pf_migration_save_consume(struct xe_device *xe, unsigned int vfid)
> > +{
> > +	struct xe_sriov_pf_migration *migration = pf_pick_migration(xe, vfid);
> > +	struct xe_sriov_migration_data *data;
> > +	int ret;
> > +
> > +	xe_assert(xe, IS_SRIOV_PF(xe));
> > +
> > +	for (;;) {
> > +		data = pf_migration_consume(xe, vfid);
> > +		if (PTR_ERR(data) != -EAGAIN)
> > +			break;
> > +
> > +		ret = wait_event_interruptible(migration->wq,
> > +					       pf_migration_data_ready(xe, vfid));
> > +		if (ret)
> > +			return ERR_PTR(ret);
> > +	}
> > +
> > +	return data;
> > +}
> > +
> > +/**
> > + * xe_sriov_pf_migration_restore_produce() - Produce a VF migration data packet to the device.
> > + * @xe: the &xe_device
> > + * @vfid: the VF identifier
> > + * @data: Pointer to &xe_sriov_migration_data
> > + *
> > + * Called by the restore migration data producer (userspace) when processing
> > + * migration data.
> > + * If the underlying data structure is full, wait until there is space.
> > + *
> > + * Return: 0 on success or a negative error code on failure.
> > + */
> > +int xe_sriov_pf_migration_restore_produce(struct xe_device *xe, unsigned int vfid,
> > +					  struct xe_sriov_migration_data *data)
> > +{
> > +	struct xe_gt *gt;
> > +
> > +	xe_assert(xe, IS_SRIOV_PF(xe));
> > +
> > +	gt = xe_device_get_gt(xe, data->gt);
> > +	if (!gt || data->tile != gt->tile->id) {
> > +		xe_sriov_err_ratelimited(xe, "VF%d Invalid GT - tile:%u, GT:%u\n",
> > +					 vfid, data->tile, data->gt);
> > +		return -EINVAL;
> > +	}
> > +
> > +	return xe_gt_sriov_pf_migration_restore_produce(gt, vfid, data);
> > +}
> > diff --git a/drivers/gpu/drm/xe/xe_sriov_pf_migration.h b/drivers/gpu/drm/xe/xe_sriov_pf_migration.h
> > index d2b4a24165438..df81a540c246a 100644
> > --- a/drivers/gpu/drm/xe/xe_sriov_pf_migration.h
> > +++ b/drivers/gpu/drm/xe/xe_sriov_pf_migration.h
> > @@ -7,10 +7,17 @@
> >  #define _XE_SRIOV_PF_MIGRATION_H_
> >  
> >  #include <linux/types.h>
> > +#include <linux/wait.h>
> >  
> >  struct xe_device;
> > +struct xe_sriov_migration_data;
> >  
> >  int xe_sriov_pf_migration_init(struct xe_device *xe);
> >  bool xe_sriov_pf_migration_supported(struct xe_device *xe);
> > +int xe_sriov_pf_migration_restore_produce(struct xe_device *xe, unsigned int vfid,
> > +					  struct xe_sriov_migration_data *data);
> > +struct xe_sriov_migration_data *
> > +xe_sriov_pf_migration_save_consume(struct xe_device *xe, unsigned int vfid);
> > +wait_queue_head_t *xe_sriov_pf_migration_waitqueue(struct xe_device *xe, unsigned int vfid);
> >  
> >  #endif
> > diff --git a/drivers/gpu/drm/xe/xe_sriov_pf_migration_types.h b/drivers/gpu/drm/xe/xe_sriov_pf_migration_types.h
> > index e69de29bb2d1d..2a45ee4e3ece8 100644
> > --- a/drivers/gpu/drm/xe/xe_sriov_pf_migration_types.h
> > +++ b/drivers/gpu/drm/xe/xe_sriov_pf_migration_types.h
> > @@ -0,0 +1,58 @@
> > +/* SPDX-License-Identifier: MIT */
> > +/*
> > + * Copyright © 2025 Intel Corporation
> > + */
> > +
> > +#ifndef _XE_SRIOV_PF_MIGRATION_TYPES_H_
> > +#define _XE_SRIOV_PF_MIGRATION_TYPES_H_
> > +
> > +#include <linux/types.h>
> > +#include <linux/wait.h>
> > +
> > +/**
> > + * struct xe_sriov_migration_data - Xe SR-IOV VF migration data packet
> > + */
> > +struct xe_sriov_migration_data {
> > +	/** @xe: Xe device */
> > +	struct xe_device *xe;
> > +	/** @vaddr: CPU pointer to payload data */
> > +	void *vaddr;
> > +	/** @remaining: payload data remaining */
> > +	size_t remaining;
> > +	/** @hdr_remaining: header data remaining */
> > +	size_t hdr_remaining;
> > +	union {
> > +		/** @bo: Buffer object with migration data */
> > +		struct xe_bo *bo;
> > +		/** @buff: Buffer with migration data */
> > +		void *buff;
> > +	};
> > +	__struct_group(xe_sriov_pf_migration_hdr, hdr, __packed,
> > +		/** @hdr.version: migration data protocol version */
> > +		u8 version;
> > +		/** @hdr.type: migration data type */
> > +		u8 type;
> > +		/** @hdr.tile: migration data tile id */
> > +		u8 tile;
> > +		/** @hdr.gt: migration data gt id */
> > +		u8 gt;
> > +		/** @hdr.flags: migration data flags */
> > +		u32 flags;
> > +		/** @hdr.offset: offset into the resource;
> > +		 * used when multiple packets of given type are used for migration
> > +		 */
> > +		u64 offset;
> > +		/** @hdr.size: migration data size  */
> > +		u64 size;
> > +	);
> > +};
> > +
> > +/**
> > + * struct xe_sriov_pf_migration - Per VF device-level migration related data
> > + */
> > +struct xe_sriov_pf_migration {
> > +	/** @wq: waitqueue used to avoid busy-waiting for snapshot production/consumption */
> > +	wait_queue_head_t wq;
> > +};
> > +
> > +#endif
> > diff --git a/drivers/gpu/drm/xe/xe_sriov_pf_types.h b/drivers/gpu/drm/xe/xe_sriov_pf_types.h
> > index 24d22afeececa..c92baaa1694ca 100644
> > --- a/drivers/gpu/drm/xe/xe_sriov_pf_types.h
> > +++ b/drivers/gpu/drm/xe/xe_sriov_pf_types.h
> > @@ -9,6 +9,7 @@
> >  #include <linux/mutex.h>
> >  #include <linux/types.h>
> >  
> > +#include "xe_sriov_pf_migration_types.h"
> >  #include "xe_sriov_pf_provision_types.h"
> >  #include "xe_sriov_pf_service_types.h"
> >  
> > @@ -18,6 +19,8 @@
> >  struct xe_sriov_metadata {
> >  	/** @version: negotiated VF/PF ABI version */
> >  	struct xe_sriov_pf_service_version version;
> > +	/** @migration: migration data */
> > +	struct xe_sriov_pf_migration migration;
> >  };
> >  
> >  /**
> 
> otherwise LGTM

Thanks,
-Michał

> 
> 

Re: [PATCH v3 07/28] drm/xe/pf: Add support for encap/decap of bitstream to/from packet

[Michał Winiarski]

On Fri, Oct 31, 2025 at 05:31:21PM +0100, Michal Wajdeczko wrote:
> 
> 
> On 10/30/2025 9:31 PM, Michał Winiarski wrote:
> > Add debugfs handlers for migration state and handle bitstream
> > .read()/.write() to convert from bitstream to/from migration data
> > packets.
> > As descriptor/trailer are handled at this layer - add handling for both
> > save and restore side.
> > 
> > Signed-off-by: Michał Winiarski <michal.winiarski@intel.com>
> > ---
> >  drivers/gpu/drm/xe/xe_sriov_migration_data.c  | 337 ++++++++++++++++++
> >  drivers/gpu/drm/xe/xe_sriov_migration_data.h  |   5 +
> >  drivers/gpu/drm/xe/xe_sriov_pf_control.c      |   5 +
> >  drivers/gpu/drm/xe/xe_sriov_pf_debugfs.c      |  35 ++
> >  drivers/gpu/drm/xe/xe_sriov_pf_migration.c    |  54 +++
> >  .../gpu/drm/xe/xe_sriov_pf_migration_types.h  |   9 +
> >  6 files changed, 445 insertions(+)
> > 
> > diff --git a/drivers/gpu/drm/xe/xe_sriov_migration_data.c b/drivers/gpu/drm/xe/xe_sriov_migration_data.c
> > index 2371ca3e6b9a9..a3f50836adc81 100644
> > --- a/drivers/gpu/drm/xe/xe_sriov_migration_data.c
> > +++ b/drivers/gpu/drm/xe/xe_sriov_migration_data.c
> > @@ -6,6 +6,45 @@
> >  #include "xe_bo.h"
> >  #include "xe_device.h"
> >  #include "xe_sriov_migration_data.h"
> > +#include "xe_sriov_pf_helpers.h"
> > +#include "xe_sriov_pf_migration.h"
> > +#include "xe_sriov_printk.h"
> > +
> > +static struct mutex *pf_migration_mutex(struct xe_device *xe, unsigned int vfid)
> > +{
> > +	xe_assert(xe, IS_SRIOV_PF(xe));
> > +	xe_assert(xe, vfid <= xe_sriov_pf_get_totalvfs(xe));
> > +
> > +	return &xe->sriov.pf.vfs[vfid].migration.lock;
> > +}
> > +
> > +static struct xe_sriov_migration_data **pf_pick_pending(struct xe_device *xe, unsigned int vfid)
> > +{
> > +	xe_assert(xe, IS_SRIOV_PF(xe));
> > +	xe_assert(xe, vfid <= xe_sriov_pf_get_totalvfs(xe));
> > +	lockdep_assert_held(pf_migration_mutex(xe, vfid));
> > +
> > +	return &xe->sriov.pf.vfs[vfid].migration.pending;
> > +}
> > +
> > +static struct xe_sriov_migration_data **
> > +pf_pick_descriptor(struct xe_device *xe, unsigned int vfid)
> > +{
> > +	xe_assert(xe, IS_SRIOV_PF(xe));
> > +	xe_assert(xe, vfid <= xe_sriov_pf_get_totalvfs(xe));
> > +	lockdep_assert_held(pf_migration_mutex(xe, vfid));
> > +
> > +	return &xe->sriov.pf.vfs[vfid].migration.descriptor;
> > +}
> > +
> > +static struct xe_sriov_migration_data **pf_pick_trailer(struct xe_device *xe, unsigned int vfid)
> > +{
> > +	xe_assert(xe, IS_SRIOV_PF(xe));
> > +	xe_assert(xe, vfid <= xe_sriov_pf_get_totalvfs(xe));
> > +	lockdep_assert_held(pf_migration_mutex(xe, vfid));
> > +
> > +	return &xe->sriov.pf.vfs[vfid].migration.trailer;
> > +}
> >  
> >  static bool data_needs_bo(struct xe_sriov_migration_data *data)
> >  {
> > @@ -42,6 +81,9 @@ struct xe_sriov_migration_data *xe_sriov_migration_data_alloc(struct xe_device *
> >   */
> >  void xe_sriov_migration_data_free(struct xe_sriov_migration_data *data)
> >  {
> > +	if (IS_ERR_OR_NULL(data))
> > +		return;
> 
> nit: maybe to be always on the safe side, this chunk should be moved to 06/28 ?

Ok.

> 
> > +
> >  	if (data_needs_bo(data))
> >  		xe_bo_unpin_map_no_vm(data->bo);
> >  	else
> > @@ -124,3 +166,298 @@ int xe_sriov_migration_data_init_from_hdr(struct xe_sriov_migration_data *data)
> >  
> >  	return mig_data_init(data);
> >  }
> > +
> > +static ssize_t vf_mig_data_hdr_read(struct xe_sriov_migration_data *data,
> > +				    char __user *buf, size_t len)
> > +{
> > +	loff_t offset = sizeof(data->hdr) - data->hdr_remaining;
> > +
> > +	if (!data->hdr_remaining)
> > +		return -EINVAL;
> > +
> > +	if (len > data->hdr_remaining)
> > +		len = data->hdr_remaining;
> > +
> > +	if (copy_to_user(buf, (void *)&data->hdr + offset, len))
> > +		return -EFAULT;
> > +
> > +	data->hdr_remaining -= len;
> > +
> > +	return len;
> > +}
> > +
> > +static ssize_t vf_mig_data_read(struct xe_sriov_migration_data *data,
> > +				char __user *buf, size_t len)
> > +{
> > +	if (len > data->remaining)
> > +		len = data->remaining;
> > +
> > +	if (copy_to_user(buf, data->vaddr + (data->size - data->remaining), len))
> > +		return -EFAULT;
> > +
> > +	data->remaining -= len;
> > +
> > +	return len;
> > +}
> > +
> > +static ssize_t __vf_mig_data_read_single(struct xe_sriov_migration_data **data,
> > +					 unsigned int vfid, char __user *buf, size_t len)
> > +{
> > +	ssize_t copied = 0;
> > +
> > +	if ((*data)->hdr_remaining)
> > +		copied = vf_mig_data_hdr_read(*data, buf, len);
> > +	else
> > +		copied = vf_mig_data_read(*data, buf, len);
> > +
> > +	if ((*data)->remaining == 0 && (*data)->hdr_remaining == 0) {
> > +		xe_sriov_migration_data_free(*data);
> > +		*data = NULL;
> > +	}
> > +
> > +	return copied;
> > +}
> > +
> > +static struct xe_sriov_migration_data **vf_mig_pick_data(struct xe_device *xe, unsigned int vfid)
> > +{
> > +	struct xe_sriov_migration_data **data;
> > +
> > +	data = pf_pick_descriptor(xe, vfid);
> > +	if (*data)
> > +		return data;
> > +
> > +	data = pf_pick_pending(xe, vfid);
> > +	if (!*data)
> > +		*data = xe_sriov_pf_migration_save_consume(xe, vfid);
> > +	if (*data)
> > +		return data;
> > +
> > +	data = pf_pick_trailer(xe, vfid);
> > +	if (*data)
> > +		return data;
> > +
> > +	return ERR_PTR(-ENODATA);
> > +}
> > +
> > +static ssize_t vf_mig_data_read_single(struct xe_device *xe, unsigned int vfid,
> > +				       char __user *buf, size_t len)
> > +{
> > +	struct xe_sriov_migration_data **data = vf_mig_pick_data(xe, vfid);
> > +
> > +	if (IS_ERR(data))
> > +		return PTR_ERR(data);
> > +
> > +	return __vf_mig_data_read_single(data, vfid, buf, len);
> > +}
> > +
> > +/**
> > + * xe_sriov_migration_data_read() - Read migration data from the device.
> > + * @xe: the &xe_device
> > + * @vfid: the VF identifier
> > + * @buf: start address of userspace buffer
> > + * @len: requested read size from userspace
> > + *
> > + * Return: number of bytes that has been successfully read,
> > + *	   0 if no more migration data is available,
> > + *	   -errno on failure.
> > + */
> > +ssize_t xe_sriov_migration_data_read(struct xe_device *xe, unsigned int vfid,
> > +				     char __user *buf, size_t len)
> > +{
> > +	ssize_t ret, consumed = 0;
> > +
> > +	xe_assert(xe, IS_SRIOV_PF(xe));
> > +
> > +	scoped_cond_guard(mutex_intr, return -EINTR, pf_migration_mutex(xe, vfid)) {
> > +		while (consumed < len) {
> > +			ret = vf_mig_data_read_single(xe, vfid, buf, len - consumed);
> > +			if (ret == -ENODATA)
> > +				break;
> > +			if (ret < 0)
> > +				return ret;
> > +
> > +			consumed += ret;
> > +			buf += ret;
> > +		}
> > +	}
> > +
> > +	return consumed;
> > +}
> > +
> > +static ssize_t vf_mig_hdr_write(struct xe_sriov_migration_data *data,
> > +				const char __user *buf, size_t len)
> > +{
> > +	loff_t offset = sizeof(data->hdr) - data->hdr_remaining;
> > +	int ret;
> > +
> > +	if (len > data->hdr_remaining)
> > +		len = data->hdr_remaining;
> > +
> > +	if (copy_from_user((void *)&data->hdr + offset, buf, len))
> > +		return -EFAULT;
> > +
> > +	data->hdr_remaining -= len;
> > +
> > +	if (!data->hdr_remaining) {
> > +		ret = xe_sriov_migration_data_init_from_hdr(data);
> > +		if (ret)
> > +			return ret;
> > +	}
> > +
> > +	return len;
> > +}
> > +
> > +static ssize_t vf_mig_data_write(struct xe_sriov_migration_data *data,
> > +				 const char __user *buf, size_t len)
> > +{
> > +	if (len > data->remaining)
> > +		len = data->remaining;
> > +
> > +	if (copy_from_user(data->vaddr + (data->size - data->remaining), buf, len))
> > +		return -EFAULT;
> > +
> > +	data->remaining -= len;
> > +
> > +	return len;
> > +}
> > +
> > +static ssize_t vf_mig_data_write_single(struct xe_device *xe, unsigned int vfid,
> > +					const char __user *buf, size_t len)
> > +{
> > +	struct xe_sriov_migration_data **data = pf_pick_pending(xe, vfid);
> > +	int ret;
> > +	ssize_t copied;
> > +
> > +	if (IS_ERR_OR_NULL(*data)) {
> > +		*data = xe_sriov_migration_data_alloc(xe);
> > +		if (!*data)
> > +			return -ENOMEM;
> > +	}
> > +
> > +	if ((*data)->hdr_remaining)
> > +		copied = vf_mig_hdr_write(*data, buf, len);
> > +	else
> > +		copied = vf_mig_data_write(*data, buf, len);
> > +
> > +	if ((*data)->hdr_remaining == 0 && (*data)->remaining == 0) {
> > +		ret = xe_sriov_pf_migration_restore_produce(xe, vfid, *data);
> > +		if (ret) {
> > +			xe_sriov_migration_data_free(*data);
> > +			return ret;
> > +		}
> > +
> > +		*data = NULL;
> > +	}
> > +
> > +	return copied;
> > +}
> > +
> > +/**
> > + * xe_sriov_migration_data_write() - Write migration data to the device.
> > + * @xe: the &xe_device
> > + * @vfid: the VF identifier
> > + * @buf: start address of userspace buffer
> > + * @len: requested write size from userspace
> > + *
> > + * Return: number of bytes that has been successfully written,
> > + *	   -errno on failure.
> > + */
> > +ssize_t xe_sriov_migration_data_write(struct xe_device *xe, unsigned int vfid,
> > +				      const char __user *buf, size_t len)
> > +{
> > +	ssize_t ret, produced = 0;
> > +
> > +	xe_assert(xe, IS_SRIOV_PF(xe));
> > +
> > +	scoped_cond_guard(mutex_intr, return -EINTR, pf_migration_mutex(xe, vfid)) {
> > +		while (produced < len) {
> > +			ret = vf_mig_data_write_single(xe, vfid, buf, len - produced);
> > +			if (ret < 0)
> > +				return ret;
> > +
> > +			produced += ret;
> > +			buf += ret;
> > +		}
> > +	}
> > +
> > +	return produced;
> > +}
> > +
> > +#define MIGRATION_DESCRIPTOR_DWORDS 0
> > +static size_t pf_descriptor_init(struct xe_device *xe, unsigned int vfid)
> > +{
> > +	struct xe_sriov_migration_data **desc = pf_pick_descriptor(xe, vfid);
> > +	struct xe_sriov_migration_data *data;
> > +	int ret;
> > +
> > +	data = xe_sriov_migration_data_alloc(xe);
> > +	if (!data)
> > +		return -ENOMEM;
> > +
> > +	ret = xe_sriov_migration_data_init(data, 0, 0, XE_SRIOV_MIGRATION_DATA_TYPE_DESCRIPTOR,
> > +					   0, MIGRATION_DESCRIPTOR_DWORDS * sizeof(u32));
> > +	if (ret) {
> > +		xe_sriov_migration_data_free(data);
> > +		return ret;
> > +	}
> > +
> > +	*desc = data;
> > +
> > +	return 0;
> > +}
> > +
> > +static void pf_pending_init(struct xe_device *xe, unsigned int vfid)
> > +{
> > +	struct xe_sriov_migration_data **data = pf_pick_pending(xe, vfid);
> > +
> > +	*data = NULL;
> > +}
> > +
> > +#define MIGRATION_TRAILER_SIZE 0
> > +static int pf_trailer_init(struct xe_device *xe, unsigned int vfid)
> > +{
> > +	struct xe_sriov_migration_data **trailer = pf_pick_trailer(xe, vfid);
> > +	struct xe_sriov_migration_data *data;
> > +	int ret;
> > +
> > +	data = xe_sriov_migration_data_alloc(xe);
> > +	if (!data)
> > +		return -ENOMEM;
> > +
> > +	ret = xe_sriov_migration_data_init(data, 0, 0, XE_SRIOV_MIGRATION_DATA_TYPE_TRAILER,
> > +					   0, MIGRATION_TRAILER_SIZE);
> > +	if (ret) {
> > +		xe_sriov_migration_data_free(data);
> > +		return ret;
> > +	}
> > +
> > +	*trailer = data;
> > +
> > +	return 0;
> > +}
> > +
> > +/**
> > + * xe_sriov_migration_data_save_init() - Initialize the pending save migration data.
> > + * @xe: the &xe_device
> > + * @vfid: the VF identifier
> > + *
> > + * Return: 0 on success, -errno on failure.
> > + */
> > +int xe_sriov_migration_data_save_init(struct xe_device *xe, unsigned int vfid)
> > +{
> > +	int ret;
> > +
> > +	scoped_cond_guard(mutex_intr, return -EINTR, pf_migration_mutex(xe, vfid)) {
> > +		ret = pf_descriptor_init(xe, vfid);
> > +		if (ret)
> > +			return ret;
> > +
> > +		ret = pf_trailer_init(xe, vfid);
> > +		if (ret)
> > +			return ret;
> > +
> > +		pf_pending_init(xe, vfid);
> > +	}
> > +
> > +	return 0;
> > +}
> > diff --git a/drivers/gpu/drm/xe/xe_sriov_migration_data.h b/drivers/gpu/drm/xe/xe_sriov_migration_data.h
> > index 3958f58a170f5..7ec489c3f28d2 100644
> > --- a/drivers/gpu/drm/xe/xe_sriov_migration_data.h
> > +++ b/drivers/gpu/drm/xe/xe_sriov_migration_data.h
> > @@ -26,5 +26,10 @@ void xe_sriov_migration_data_free(struct xe_sriov_migration_data *snapshot);
> >  int xe_sriov_migration_data_init(struct xe_sriov_migration_data *data, u8 tile_id, u8 gt_id,
> >  				 enum xe_sriov_migration_data_type, loff_t offset, size_t size);
> >  int xe_sriov_migration_data_init_from_hdr(struct xe_sriov_migration_data *snapshot);
> > +ssize_t xe_sriov_migration_data_read(struct xe_device *xe, unsigned int vfid,
> > +				     char __user *buf, size_t len);
> > +ssize_t xe_sriov_migration_data_write(struct xe_device *xe, unsigned int vfid,
> > +				      const char __user *buf, size_t len);
> > +int xe_sriov_migration_data_save_init(struct xe_device *xe, unsigned int vfid);
> >  
> >  #endif
> > diff --git a/drivers/gpu/drm/xe/xe_sriov_pf_control.c b/drivers/gpu/drm/xe/xe_sriov_pf_control.c
> > index 8d8a01faf5291..c2768848daba1 100644
> > --- a/drivers/gpu/drm/xe/xe_sriov_pf_control.c
> > +++ b/drivers/gpu/drm/xe/xe_sriov_pf_control.c
> > @@ -5,6 +5,7 @@
> >  
> >  #include "xe_device.h"
> >  #include "xe_gt_sriov_pf_control.h"
> > +#include "xe_sriov_migration_data.h"
> >  #include "xe_sriov_pf_control.h"
> >  #include "xe_sriov_printk.h"
> >  
> > @@ -165,6 +166,10 @@ int xe_sriov_pf_control_trigger_save_vf(struct xe_device *xe, unsigned int vfid)
> >  	unsigned int id;
> >  	int ret;
> >  
> > +	ret = xe_sriov_migration_data_save_init(xe, vfid);
> > +	if (ret)
> > +		return ret;
> > +
> >  	for_each_gt(gt, xe, id) {
> >  		ret = xe_gt_sriov_pf_control_trigger_save_vf(gt, vfid);
> >  		if (ret)
> > diff --git a/drivers/gpu/drm/xe/xe_sriov_pf_debugfs.c b/drivers/gpu/drm/xe/xe_sriov_pf_debugfs.c
> > index e0e6340c49106..a9a28aec22421 100644
> > --- a/drivers/gpu/drm/xe/xe_sriov_pf_debugfs.c
> > +++ b/drivers/gpu/drm/xe/xe_sriov_pf_debugfs.c
> > @@ -9,6 +9,7 @@
> >  #include "xe_device.h"
> >  #include "xe_device_types.h"
> >  #include "xe_pm.h"
> > +#include "xe_sriov_migration_data.h"
> >  #include "xe_sriov_pf.h"
> >  #include "xe_sriov_pf_control.h"
> >  #include "xe_sriov_pf_debugfs.h"
> > @@ -132,6 +133,7 @@ static void pf_populate_pf(struct xe_device *xe, struct dentry *pfdent)
> >   *      /sys/kernel/debug/dri/BDF/
> >   *      ├── sriov
> >   *      │   ├── vf1
> > + *      │   │   ├── migration_data
> >   *      │   │   ├── pause
> >   *      │   │   ├── reset
> >   *      │   │   ├── resume
> > @@ -220,6 +222,38 @@ DEFINE_VF_CONTROL_ATTRIBUTE(reset_vf);
> >  DEFINE_VF_CONTROL_ATTRIBUTE_RW(save_vf);
> >  DEFINE_VF_CONTROL_ATTRIBUTE_RW(restore_vf);
> >  
> > +static ssize_t data_write(struct file *file, const char __user *buf, size_t count, loff_t *pos)
> > +{
> > +	struct dentry *dent = file_dentry(file)->d_parent;
> > +	struct xe_device *xe = extract_xe(dent);
> > +	unsigned int vfid = extract_vfid(dent);
> > +
> > +	if (*pos)
> > +		return -ESPIPE;
> > +
> > +	return xe_sriov_migration_data_write(xe, vfid, buf, count);
> > +}
> > +
> > +static ssize_t data_read(struct file *file, char __user *buf, size_t count, loff_t *ppos)
> > +{
> > +	struct dentry *dent = file_dentry(file)->d_parent;
> > +	struct xe_device *xe = extract_xe(dent);
> > +	unsigned int vfid = extract_vfid(dent);
> > +
> > +	if (*ppos)
> > +		return -ESPIPE;
> > +
> > +	return xe_sriov_migration_data_read(xe, vfid, buf, count);
> > +}
> > +
> > +static const struct file_operations data_vf_fops = {
> > +	.owner		= THIS_MODULE,
> > +	.open		= simple_open,
> > +	.write		= data_write,
> > +	.read		= data_read,
> > +	.llseek		= default_llseek,
> > +};
> > +
> >  static void pf_populate_vf(struct xe_device *xe, struct dentry *vfdent)
> >  {
> >  	debugfs_create_file("pause", 0200, vfdent, xe, &pause_vf_fops);
> > @@ -228,6 +262,7 @@ static void pf_populate_vf(struct xe_device *xe, struct dentry *vfdent)
> >  	debugfs_create_file("reset", 0200, vfdent, xe, &reset_vf_fops);
> >  	debugfs_create_file("save", 0600, vfdent, xe, &save_vf_fops);
> >  	debugfs_create_file("restore", 0600, vfdent, xe, &restore_vf_fops);
> > +	debugfs_create_file("migration_data", 0600, vfdent, xe, &data_vf_fops);
> >  }
> >  
> >  static void pf_populate_with_tiles(struct xe_device *xe, struct dentry *dent, unsigned int vfid)
> > diff --git a/drivers/gpu/drm/xe/xe_sriov_pf_migration.c b/drivers/gpu/drm/xe/xe_sriov_pf_migration.c
> > index 7be9f026d80e8..8ea531d36f53b 100644
> > --- a/drivers/gpu/drm/xe/xe_sriov_pf_migration.c
> > +++ b/drivers/gpu/drm/xe/xe_sriov_pf_migration.c
> > @@ -10,6 +10,7 @@
> >  #include "xe_gt_sriov_pf_migration.h"
> >  #include "xe_pm.h"
> >  #include "xe_sriov.h"
> > +#include "xe_sriov_migration_data.h"
> >  #include "xe_sriov_pf_helpers.h"
> >  #include "xe_sriov_pf_migration.h"
> >  #include "xe_sriov_printk.h"
> > @@ -53,6 +54,15 @@ static bool pf_check_migration_support(struct xe_device *xe)
> >  	return IS_ENABLED(CONFIG_DRM_XE_DEBUG);
> >  }
> >  
> > +static void pf_migration_cleanup(void *arg)
> > +{
> > +	struct xe_sriov_pf_migration *migration = arg;
> > +
> > +	xe_sriov_migration_data_free(migration->pending);
> > +	xe_sriov_migration_data_free(migration->trailer);
> > +	xe_sriov_migration_data_free(migration->descriptor);
> > +}
> > +
> >  /**
> >   * xe_sriov_pf_migration_init() - Initialize support for SR-IOV VF migration.
> >   * @xe: the &xe_device
> > @@ -62,6 +72,7 @@ static bool pf_check_migration_support(struct xe_device *xe)
> >  int xe_sriov_pf_migration_init(struct xe_device *xe)
> >  {
> >  	unsigned int n, totalvfs;
> > +	int err;
> >  
> >  	xe_assert(xe, IS_SRIOV_PF(xe));
> >  
> > @@ -73,7 +84,15 @@ int xe_sriov_pf_migration_init(struct xe_device *xe)
> >  	for (n = 1; n <= totalvfs; n++) {
> >  		struct xe_sriov_pf_migration *migration = pf_pick_migration(xe, n);
> >  
> > +		err = devm_mutex_init(xe->drm.dev, &migration->lock);
> 
> IIRC all software data allocations/inits we are doing as drmm
> only actions that interacts with or cleanups the hw (*) use devm

I'll use drmm for the mutex.

> 
> > +		if (err)
> > +			return err;
> > +
> >  		init_waitqueue_head(&migration->wq);
> > +
> > +		err = devm_add_action_or_reset(xe->drm.dev, pf_migration_cleanup, migration);
> 
> (*) like here
> 
> > +		if (err)
> > +			return err;
> >  	}
> >  
> >  	return 0;
> > @@ -153,6 +172,36 @@ xe_sriov_pf_migration_save_consume(struct xe_device *xe, unsigned int vfid)
> >  	return data;
> >  }
> >  
> > +static int pf_handle_descriptor(struct xe_device *xe, unsigned int vfid,
> > +				struct xe_sriov_migration_data *data)
> > +{
> > +	if (data->tile != 0 || data->gt != 0)
> > +		return -EINVAL;
> > +
> > +	xe_sriov_migration_data_free(data);
> > +
> > +	return 0;
> > +}
> > +
> > +static int pf_handle_trailer(struct xe_device *xe, unsigned int vfid,
> > +			     struct xe_sriov_migration_data *data)
> > +{
> > +	struct xe_gt *gt;
> > +	u8 gt_id;
> > +
> > +	if (data->tile != 0 || data->gt != 0)
> > +		return -EINVAL;
> > +	if (data->offset != 0 || data->size != 0 || data->buff || data->bo)
> > +		return -EINVAL;
> > +
> > +	xe_sriov_migration_data_free(data);
> > +
> > +	for_each_gt(gt, xe, gt_id)
> > +		xe_gt_sriov_pf_control_restore_data_done(gt, vfid);
> > +
> > +	return 0;
> > +}
> > +
> >  /**
> >   * xe_sriov_pf_migration_restore_produce() - Produce a VF migration data packet to the device.
> >   * @xe: the &xe_device
> > @@ -172,6 +221,11 @@ int xe_sriov_pf_migration_restore_produce(struct xe_device *xe, unsigned int vfi
> >  
> >  	xe_assert(xe, IS_SRIOV_PF(xe));
> >  
> > +	if (data->type == XE_SRIOV_MIGRATION_DATA_TYPE_DESCRIPTOR)
> > +		return pf_handle_descriptor(xe, vfid, data);
> > +	if (data->type == XE_SRIOV_MIGRATION_DATA_TYPE_TRAILER)
> > +		return pf_handle_trailer(xe, vfid, data);
> > +
> >  	gt = xe_device_get_gt(xe, data->gt);
> >  	if (!gt || data->tile != gt->tile->id) {
> >  		xe_sriov_err_ratelimited(xe, "VF%d Invalid GT - tile:%u, GT:%u\n",
> > diff --git a/drivers/gpu/drm/xe/xe_sriov_pf_migration_types.h b/drivers/gpu/drm/xe/xe_sriov_pf_migration_types.h
> > index 2a45ee4e3ece8..8468e5eeb6d66 100644
> > --- a/drivers/gpu/drm/xe/xe_sriov_pf_migration_types.h
> > +++ b/drivers/gpu/drm/xe/xe_sriov_pf_migration_types.h
> > @@ -7,6 +7,7 @@
> >  #define _XE_SRIOV_PF_MIGRATION_TYPES_H_
> >  
> >  #include <linux/types.h>
> > +#include <linux/mutex_types.h>
> >  #include <linux/wait.h>
> >  
> >  /**
> > @@ -53,6 +54,14 @@ struct xe_sriov_migration_data {
> >  struct xe_sriov_pf_migration {
> >  	/** @wq: waitqueue used to avoid busy-waiting for snapshot production/consumption */
> >  	wait_queue_head_t wq;
> > +	/** @lock: Mutex protecting the migration data */
> > +	struct mutex lock;
> > +	/** @pending: currently processed data packet of VF resource */
> > +	struct xe_sriov_migration_data *pending;
> > +	/** @trailer: data packet used to indicate the end of stream */
> > +	struct xe_sriov_migration_data *trailer;
> > +	/** @descriptor: data packet containing the metadata describing the device */
> > +	struct xe_sriov_migration_data *descriptor;
> >  };
> >  
> >  #endif
> 
> with devm/drmm clarified,
> 
> 	Reviewed-by: Michal Wajdeczko <michal.wajdeczko@intel.com>
> 

Thanks,
-Michał

Re: [PATCH v3 16/28] drm/xe/pf: Handle GuC migration data as part of PF control

[Michał Winiarski]

On Fri, Oct 31, 2025 at 07:15:18PM +0100, Michal Wajdeczko wrote:
> 
> 
> On 10/30/2025 9:31 PM, Michał Winiarski wrote:
> > Connect the helpers to allow save and restore of GuC migration data in
> > stop_copy / resume device state.
> > 
> > Signed-off-by: Michał Winiarski <michal.winiarski@intel.com>
> > ---
> >  drivers/gpu/drm/xe/xe_gt_sriov_pf_control.c   | 25 ++++++-
> >  drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.c | 66 +++++++++++++++----
> >  drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.h |  8 ++-
> >  .../drm/xe/xe_gt_sriov_pf_migration_types.h   |  5 ++
> >  4 files changed, 87 insertions(+), 17 deletions(-)
> > 
> > diff --git a/drivers/gpu/drm/xe/xe_gt_sriov_pf_control.c b/drivers/gpu/drm/xe/xe_gt_sriov_pf_control.c
> > index 203ba1fbeefcd..cb45e97f4c4d9 100644
> > --- a/drivers/gpu/drm/xe/xe_gt_sriov_pf_control.c
> > +++ b/drivers/gpu/drm/xe/xe_gt_sriov_pf_control.c
> > @@ -848,6 +848,18 @@ static void pf_enter_vf_save_failed(struct xe_gt *gt, unsigned int vfid)
> >  
> >  static int pf_handle_vf_save_data(struct xe_gt *gt, unsigned int vfid)
> >  {
> > +	int ret;
> > +
> > +	if (xe_gt_sriov_pf_migration_save_test(gt, vfid, XE_SRIOV_MIGRATION_DATA_TYPE_GUC)) {
> > +		ret = xe_gt_sriov_pf_migration_guc_save(gt, vfid);
> > +		if (ret)
> > +			return ret;
> > +
> > +		xe_gt_sriov_pf_migration_save_clear(gt, vfid, XE_SRIOV_MIGRATION_DATA_TYPE_GUC);
> > +
> > +		return -EAGAIN;
> > +	}
> > +
> >  	return 0;
> >  }
> >  
> > @@ -887,6 +899,7 @@ static void pf_exit_vf_save_wait_data(struct xe_gt *gt, unsigned int vfid)
> >  static bool pf_enter_vf_save_wip(struct xe_gt *gt, unsigned int vfid)
> >  {
> >  	if (pf_enter_vf_state(gt, vfid, XE_GT_SRIOV_STATE_SAVE_WIP)) {
> > +		xe_gt_sriov_pf_migration_save_init(gt, vfid);
> >  		pf_enter_vf_state(gt, vfid, XE_GT_SRIOV_STATE_SAVE_PROCESS_DATA);
> >  		pf_enter_vf_wip(gt, vfid);
> >  		pf_queue_vf(gt, vfid);
> > @@ -1048,14 +1061,22 @@ static void pf_enter_vf_restore_failed(struct xe_gt *gt, unsigned int vfid)
> >  static int pf_handle_vf_restore_data(struct xe_gt *gt, unsigned int vfid)
> >  {
> >  	struct xe_sriov_migration_data *data = xe_gt_sriov_pf_migration_restore_consume(gt, vfid);
> > +	int ret = 0;
> >  
> >  	xe_gt_assert(gt, data);
> >  
> > -	xe_gt_sriov_notice(gt, "Skipping VF%u unknown data type: %d\n", vfid, data->type);
> > +	switch (data->type) {
> > +	case XE_SRIOV_MIGRATION_DATA_TYPE_GUC:
> > +		ret = xe_gt_sriov_pf_migration_guc_restore(gt, vfid, data);
> > +		break;
> > +	default:
> > +		xe_gt_sriov_notice(gt, "Skipping VF%u unknown data type: %d\n", vfid, data->type);
> > +		break;
> > +	}
> >  
> >  	xe_sriov_migration_data_free(data);
> >  
> > -	return 0;
> > +	return ret;
> >  }
> >  
> >  static bool pf_handle_vf_restore(struct xe_gt *gt, unsigned int vfid)
> > diff --git a/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.c b/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.c
> > index 4a716e0a29fe4..fbb451767712a 100644
> > --- a/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.c
> > +++ b/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.c
> > @@ -174,22 +174,10 @@ static int pf_save_vf_guc_mig_data(struct xe_gt *gt, unsigned int vfid)
> >  	return ret;
> >  }
> >  
> > -/**
> > - * xe_gt_sriov_pf_migration_guc_size() - Get the size of VF GuC migration data.
> > - * @gt: the &xe_gt
> > - * @vfid: the VF identifier
> > - *
> > - * This function is for PF only.
> > - *
> > - * Return: size in bytes or a negative error code on failure.
> > - */
> > -ssize_t xe_gt_sriov_pf_migration_guc_size(struct xe_gt *gt, unsigned int vfid)
> > +static ssize_t pf_migration_guc_size(struct xe_gt *gt, unsigned int vfid)
> >  {
> >  	ssize_t size;
> >  
> > -	xe_gt_assert(gt, IS_SRIOV_PF(gt_to_xe(gt)));
> > -	xe_gt_assert(gt, vfid != PFID);
> > -	xe_gt_assert(gt, vfid <= xe_sriov_pf_get_totalvfs(gt_to_xe(gt)));
> >  
> >  	if (!pf_migration_supported(gt))
> >  		return -ENOPKG;
> > @@ -278,12 +266,19 @@ int xe_gt_sriov_pf_migration_guc_restore(struct xe_gt *gt, unsigned int vfid,
> >  ssize_t xe_gt_sriov_pf_migration_size(struct xe_gt *gt, unsigned int vfid)
> >  {
> >  	ssize_t total = 0;
> > +	ssize_t size;
> >  
> >  	xe_gt_assert(gt, IS_SRIOV_PF(gt_to_xe(gt)));
> >  	xe_gt_assert(gt, vfid != PFID);
> >  	xe_gt_assert(gt, vfid <= xe_sriov_pf_get_totalvfs(gt_to_xe(gt)));
> >  
> > -	/* Nothing to query yet - will be updated once per-GT migration data types are added */
> > +	size = pf_migration_guc_size(gt, vfid);
> > +	if (size < 0)
> > +		return size;
> > +	if (size > 0)
> > +		size += sizeof(struct xe_sriov_pf_migration_hdr);
> > +	total += size;
> > +
> >  	return total;
> >  }
> >  
> > @@ -330,6 +325,49 @@ void xe_gt_sriov_pf_migration_ring_free(struct xe_gt *gt, unsigned int vfid)
> >  		xe_sriov_migration_data_free(data);
> >  }
> >  
> > +/**
> > + * xe_gt_sriov_pf_migration_save_init() - Initialize per-GT migration related data.
> > + * @gt: the &xe_gt
> > + * @vfid: the VF identifier (can't be 0)
> > + *
> > + * Return: 0 on success or a negative error code on failure.
> 
> but it's void

Ok.

> 
> > + */
> > +void xe_gt_sriov_pf_migration_save_init(struct xe_gt *gt, unsigned int vfid)
> > +{
> > +	struct xe_gt_sriov_migration_data *migration = pf_pick_gt_migration(gt, vfid);
> > +
> > +	migration->save.data_remaining = 0;
> > +
> > +	xe_gt_assert(gt, pf_migration_guc_size(gt, vfid) > 0);
> > +	set_bit(XE_SRIOV_MIGRATION_DATA_TYPE_GUC, &migration->save.data_remaining);
> > +}
> > +
> > +/**
> > + * xe_gt_sriov_pf_migration_save_test() - Test if migration data type needs to be saved.
> > + * @gt: the &xe_gt
> > + * @vfid: the VF identifier (can't be 0)
> > + * @type: the &xe_sriov_migration_data_type of data to be saved
> 
> .. to check

Ok.

> 
> > + *
> > + * Return: true if the data needs saving, otherwise false.
> > + */
> > +bool xe_gt_sriov_pf_migration_save_test(struct xe_gt *gt, unsigned int vfid,
> > +					enum xe_sriov_migration_data_type type)
> > +{
> > +	return test_bit(type, &pf_pick_gt_migration(gt, vfid)->save.data_remaining);
> > +}
> > +
> > +/**
> > + * xe_gt_sriov_pf_migration_save_clear() - Clear migration data type from saving.
> > + * @gt: the &xe_gt
> > + * @vfid: the VF identifier (can't be 0)
> > + * @type: the &xe_sriov_migration_data_type of data to be cleared
> 
> .. to mark as done/saved

Ok.

> 
> > + */
> > +void xe_gt_sriov_pf_migration_save_clear(struct xe_gt *gt, unsigned int vfid,
> > +					 enum xe_sriov_migration_data_type type)
> > +{
> > +	clear_bit(type, &pf_pick_gt_migration(gt, vfid)->save.data_remaining);
> > +}
> > +
> >  /**
> >   * xe_gt_sriov_pf_migration_save_produce() - Add VF save data packet to migration ring.
> >   * @gt: the &xe_gt
> > diff --git a/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.h b/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.h
> > index b3c18e369df79..66a45d6234245 100644
> > --- a/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.h
> > +++ b/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.h
> > @@ -10,12 +10,12 @@
> >  
> >  struct xe_gt;
> >  struct xe_sriov_migration_data;
> > +enum xe_sriov_migration_data_type;
> >  
> >  /* TODO: get this information by querying GuC in the future */
> >  #define XE_GT_SRIOV_PF_MIGRATION_GUC_DATA_MAX_SIZE SZ_8M
> >  
> >  int xe_gt_sriov_pf_migration_init(struct xe_gt *gt);
> > -ssize_t xe_gt_sriov_pf_migration_guc_size(struct xe_gt *gt, unsigned int vfid);
> >  int xe_gt_sriov_pf_migration_guc_save(struct xe_gt *gt, unsigned int vfid);
> >  int xe_gt_sriov_pf_migration_guc_restore(struct xe_gt *gt, unsigned int vfid,
> >  					 struct xe_sriov_migration_data *data);
> > @@ -26,6 +26,12 @@ bool xe_gt_sriov_pf_migration_ring_empty(struct xe_gt *gt, unsigned int vfid);
> >  bool xe_gt_sriov_pf_migration_ring_full(struct xe_gt *gt, unsigned int vfid);
> >  void xe_gt_sriov_pf_migration_ring_free(struct xe_gt *gt, unsigned int vfid);
> >  
> > +void xe_gt_sriov_pf_migration_save_init(struct xe_gt *gt, unsigned int vfid);
> > +bool xe_gt_sriov_pf_migration_save_test(struct xe_gt *gt, unsigned int vfid,
> > +					enum xe_sriov_migration_data_type type);
> > +void xe_gt_sriov_pf_migration_save_clear(struct xe_gt *gt, unsigned int vfid,
> > +					 enum xe_sriov_migration_data_type type);
> 
> nit: "clear" is little confusing, maybe
> 
> 	xe_gt_sriov_pf_migration_save_data_done(gt, vfid, type);	// instead of save_clear
> 	xe_gt_sriov_pf_migration_save_data_pending(gt, vfid, type);	// instead of save_test
> and 
> 	pf_migration_save_data_todo(gt, vfid, type);	// instead of set_bit

I'll go with:
xe_gt_sriov_pf_migration_save_data_pending 
xe_gt_sriov_pf_migration_save_data_complete

> 
> 
> > +
> >  int xe_gt_sriov_pf_migration_save_produce(struct xe_gt *gt, unsigned int vfid,
> >  					  struct xe_sriov_migration_data *data);
> >  struct xe_sriov_migration_data *
> > diff --git a/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration_types.h b/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration_types.h
> > index 75d8b94cbbefb..9f24878690d9c 100644
> > --- a/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration_types.h
> > +++ b/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration_types.h
> > @@ -16,6 +16,11 @@
> >  struct xe_gt_sriov_migration_data {
> >  	/** @ring: queue containing VF save / restore migration data */
> >  	struct ptr_ring ring;
> > +	/** @save: structure for currently processed save migration data */
> > +	struct {
> > +		/** @save.data_remaining: bitmap of migration types that need to be saved */
> > +		unsigned long data_remaining;
> > +	} save;
> >  };
> >  
> >  #endif
> 
> otherwise LGTM, so with kernel-doc fixed, and (hopefully) with better names,
> 
> Reviewed-by: Michal Wajdeczko <michal.wajdeczko@intel.com>
> 

Thanks,
-Michał

Re: [PATCH v3 18/28] drm/xe/pf: Handle GGTT migration data as part of PF control

[Michał Winiarski]

On Fri, Oct 31, 2025 at 07:26:29PM +0100, Michal Wajdeczko wrote:
> 
> 
> On 10/30/2025 9:31 PM, Michał Winiarski wrote:
> > Connect the helpers to allow save and restore of GGTT migration data in
> > stop_copy / resume device state.
> > 
> > Signed-off-by: Michał Winiarski <michal.winiarski@intel.com>
> > ---
> >  drivers/gpu/drm/xe/xe_gt_sriov_pf_control.c   |  13 ++
> >  drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.c | 113 ++++++++++++++++++
> >  drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.h |   3 +
> >  3 files changed, 129 insertions(+)
> > 
> > diff --git a/drivers/gpu/drm/xe/xe_gt_sriov_pf_control.c b/drivers/gpu/drm/xe/xe_gt_sriov_pf_control.c
> > index cb45e97f4c4d9..e7ea9b88fd246 100644
> > --- a/drivers/gpu/drm/xe/xe_gt_sriov_pf_control.c
> > +++ b/drivers/gpu/drm/xe/xe_gt_sriov_pf_control.c
> > @@ -860,6 +860,16 @@ static int pf_handle_vf_save_data(struct xe_gt *gt, unsigned int vfid)
> >  		return -EAGAIN;
> >  	}
> >  
> > +	if (xe_gt_sriov_pf_migration_save_test(gt, vfid, XE_SRIOV_MIGRATION_DATA_TYPE_GGTT)) {
> > +		ret = xe_gt_sriov_pf_migration_ggtt_save(gt, vfid);
> > +		if (ret)
> > +			return ret;
> > +
> > +		xe_gt_sriov_pf_migration_save_clear(gt, vfid, XE_SRIOV_MIGRATION_DATA_TYPE_GGTT);
> > +
> > +		return -EAGAIN;
> > +	}
> > +
> >  	return 0;
> >  }
> >  
> > @@ -1066,6 +1076,9 @@ static int pf_handle_vf_restore_data(struct xe_gt *gt, unsigned int vfid)
> >  	xe_gt_assert(gt, data);
> >  
> >  	switch (data->type) {
> > +	case XE_SRIOV_MIGRATION_DATA_TYPE_GGTT:
> > +		ret = xe_gt_sriov_pf_migration_ggtt_restore(gt, vfid, data);
> > +		break;
> >  	case XE_SRIOV_MIGRATION_DATA_TYPE_GUC:
> >  		ret = xe_gt_sriov_pf_migration_guc_restore(gt, vfid, data);
> >  		break;
> > diff --git a/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.c b/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.c
> > index fbb451767712a..6f2ee5820bdd4 100644
> > --- a/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.c
> > +++ b/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.c
> > @@ -7,6 +7,9 @@
> >  
> >  #include "abi/guc_actions_sriov_abi.h"
> >  #include "xe_bo.h"
> > +#include "xe_ggtt.h"
> > +#include "xe_gt.h"
> > +#include "xe_gt_sriov_pf_config.h"
> >  #include "xe_gt_sriov_pf_control.h"
> >  #include "xe_gt_sriov_pf_helpers.h"
> >  #include "xe_gt_sriov_pf_migration.h"
> > @@ -39,6 +42,106 @@ static void pf_dump_mig_data(struct xe_gt *gt, unsigned int vfid,
> >  	}
> >  }
> >  
> > +static ssize_t pf_migration_ggtt_size(struct xe_gt *gt, unsigned int vfid)
> > +{
> > +	if (!xe_gt_is_main_type(gt))
> > +		return 0;
> > +
> > +	return xe_gt_sriov_pf_config_ggtt_save(gt, vfid, NULL, 0);
> > +}
> > +
> > +static int pf_save_vf_ggtt_mig_data(struct xe_gt *gt, unsigned int vfid)
> > +{
> > +	struct xe_sriov_migration_data *data;
> > +	size_t size;
> > +	int ret;
> > +
> > +	size = pf_migration_ggtt_size(gt, vfid);
> > +	xe_gt_assert(gt, size);
> > +
> > +	data = xe_sriov_migration_data_alloc(gt_to_xe(gt));
> > +	if (!data)
> > +		return -ENOMEM;
> > +
> > +	ret = xe_sriov_migration_data_init(data, gt->tile->id, gt->info.id,
> > +					   XE_SRIOV_MIGRATION_DATA_TYPE_GGTT, 0, size);
> > +	if (ret)
> > +		goto fail;
> > +
> > +	ret = xe_gt_sriov_pf_config_ggtt_save(gt, vfid, data->vaddr, size);
> > +	if (ret)
> > +		goto fail;
> > +
> > +	xe_gt_sriov_dbg_verbose(gt, "VF%u GGTT data save (%zu bytes)\n", vfid, size);
> > +	pf_dump_mig_data(gt, vfid, data);
> > +
> > +	ret = xe_gt_sriov_pf_migration_save_produce(gt, vfid, data);
> > +	if (ret)
> > +		goto fail;
> > +
> > +	return 0;
> > +
> > +fail:
> > +	xe_sriov_migration_data_free(data);
> > +	xe_gt_sriov_err(gt, "Failed to save VF%u GGTT data (%pe)\n", vfid, ERR_PTR(ret));
> > +	return ret;
> > +}
> > +
> > +static int pf_restore_vf_ggtt_mig_data(struct xe_gt *gt, unsigned int vfid,
> > +				       struct xe_sriov_migration_data *data)
> > +{
> > +	int ret;
> > +
> > +	xe_gt_sriov_dbg_verbose(gt, "VF%u GGTT data restore (%llu bytes)\n", vfid, data->size);
> > +	pf_dump_mig_data(gt, vfid, data);
> > +
> > +	ret = xe_gt_sriov_pf_config_ggtt_restore(gt, vfid, data->vaddr, data->size);
> > +	if (ret) {
> > +		xe_gt_sriov_err(gt, "Failed to restore VF%u GGTT data (%pe)\n",
> > +				vfid, ERR_PTR(ret));
> > +		return ret;
> > +	}
> > +
> > +	return 0;
> > +}
> > +
> > +/**
> > + * xe_gt_sriov_pf_migration_ggtt_save() - Save VF GGTT migration data.
> > + * @gt: the &xe_gt
> > + * @vfid: the VF identifier (can't be 0)
> > + *
> > + * This function is for PF only.
> > + *
> > + * Return: 0 on success or a negative error code on failure.
> > + */
> > +int xe_gt_sriov_pf_migration_ggtt_save(struct xe_gt *gt, unsigned int vfid)
> > +{
> > +	xe_gt_assert(gt, IS_SRIOV_PF(gt_to_xe(gt)));
> > +	xe_gt_assert(gt, vfid != PFID);
> > +	xe_gt_assert(gt, vfid <= xe_sriov_pf_get_totalvfs(gt_to_xe(gt)));
> > +
> > +	return pf_save_vf_ggtt_mig_data(gt, vfid);
> > +}
> > +
> > +/**
> > + * xe_gt_sriov_pf_migration_ggtt_restore() - Restore VF GGTT migration data.
> > + * @gt: the &xe_gt
> > + * @vfid: the VF identifier (can't be 0)
> > + *
> > + * This function is for PF only.
> > + *
> > + * Return: 0 on success or a negative error code on failure.
> > + */
> > +int xe_gt_sriov_pf_migration_ggtt_restore(struct xe_gt *gt, unsigned int vfid,
> > +					  struct xe_sriov_migration_data *data)
> > +{
> > +	xe_gt_assert(gt, IS_SRIOV_PF(gt_to_xe(gt)));
> > +	xe_gt_assert(gt, vfid != PFID);
> > +	xe_gt_assert(gt, vfid <= xe_sriov_pf_get_totalvfs(gt_to_xe(gt)));
> > +
> > +	return pf_restore_vf_ggtt_mig_data(gt, vfid, data);
> > +}
> > +
> >  /* Return: number of dwords saved/restored/required or a negative error code on failure */
> >  static int guc_action_vf_save_restore(struct xe_guc *guc, u32 vfid, u32 opcode,
> >  				      u64 addr, u32 ndwords)
> > @@ -279,6 +382,13 @@ ssize_t xe_gt_sriov_pf_migration_size(struct xe_gt *gt, unsigned int vfid)
> >  		size += sizeof(struct xe_sriov_pf_migration_hdr);
> >  	total += size;
> >  
> > +	size = pf_migration_ggtt_size(gt, vfid);
> > +	if (size < 0)
> > +		return size;
> > +	if (size > 0)
> > +		size += sizeof(struct xe_sriov_pf_migration_hdr);
> > +	total += size;
> > +
> >  	return total;
> >  }
> >  
> > @@ -340,6 +450,9 @@ void xe_gt_sriov_pf_migration_save_init(struct xe_gt *gt, unsigned int vfid)
> >  
> >  	xe_gt_assert(gt, pf_migration_guc_size(gt, vfid) > 0);
> >  	set_bit(XE_SRIOV_MIGRATION_DATA_TYPE_GUC, &migration->save.data_remaining);
> > +
> > +	if (pf_migration_ggtt_size(gt, vfid) > 0)
> > +		set_bit(XE_SRIOV_MIGRATION_DATA_TYPE_GGTT, &migration->save.data_remaining);
> 
> btw, does it make sense to save GuC data if there is no GGTT data on the main GT?

Not in a sense that such provisioning is usable, but we do not want to
have any dependencies between data types, and we do need to handle this
corner case (not fail the migration if GGTT size is set to 0).

> 
> >  }
> >  
> >  /**
> > diff --git a/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.h b/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.h
> > index 66a45d6234245..bc201d8f3147a 100644
> > --- a/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.h
> > +++ b/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.h
> > @@ -19,6 +19,9 @@ int xe_gt_sriov_pf_migration_init(struct xe_gt *gt);
> >  int xe_gt_sriov_pf_migration_guc_save(struct xe_gt *gt, unsigned int vfid);
> >  int xe_gt_sriov_pf_migration_guc_restore(struct xe_gt *gt, unsigned int vfid,
> >  					 struct xe_sriov_migration_data *data);
> > +int xe_gt_sriov_pf_migration_ggtt_save(struct xe_gt *gt, unsigned int vfid);
> > +int xe_gt_sriov_pf_migration_ggtt_restore(struct xe_gt *gt, unsigned int vfid,
> > +					  struct xe_sriov_migration_data *data);
> >  
> >  ssize_t xe_gt_sriov_pf_migration_size(struct xe_gt *gt, unsigned int vfid);
> >  
> 
> only one nit question about corner case, so
> 
> Reviewed-by: Michal Wajdeczko <michal.wajdeczko@intel.com>
> 

Thanks,
-Michał

Re: [PATCH v3 19/28] drm/xe/pf: Handle MMIO migration data as part of PF control

[Michał Winiarski]

On Fri, Oct 31, 2025 at 07:39:58PM +0100, Michal Wajdeczko wrote:
> 
> 
> On 10/30/2025 9:31 PM, Michał Winiarski wrote:
> > Implement the helpers and use them for save and restore of MMIO
> > migration data in stop_copy / resume device state.
> > 
> > Signed-off-by: Michał Winiarski <michal.winiarski@intel.com>
> > ---
> >  drivers/gpu/drm/xe/xe_gt_sriov_pf.h           |   2 +
> >  drivers/gpu/drm/xe/xe_gt_sriov_pf_control.c   |  13 ++
> >  drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.c | 158 ++++++++++++++++++
> >  drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.h |   3 +
> >  4 files changed, 176 insertions(+)
> > 
> > diff --git a/drivers/gpu/drm/xe/xe_gt_sriov_pf.h b/drivers/gpu/drm/xe/xe_gt_sriov_pf.h
> > index e7fde3f9937af..c0dcbb1054e4d 100644
> > --- a/drivers/gpu/drm/xe/xe_gt_sriov_pf.h
> > +++ b/drivers/gpu/drm/xe/xe_gt_sriov_pf.h
> > @@ -6,6 +6,8 @@
> >  #ifndef _XE_GT_SRIOV_PF_H_
> >  #define _XE_GT_SRIOV_PF_H_
> >  
> > +#include <linux/types.h>
> 
> ?

Oops, leftover from previous iterations.

> 
> > +
> >  struct xe_gt;
> >  
> >  #ifdef CONFIG_PCI_IOV
> > diff --git a/drivers/gpu/drm/xe/xe_gt_sriov_pf_control.c b/drivers/gpu/drm/xe/xe_gt_sriov_pf_control.c
> > index e7ea9b88fd246..7cd7cae950bc7 100644
> > --- a/drivers/gpu/drm/xe/xe_gt_sriov_pf_control.c
> > +++ b/drivers/gpu/drm/xe/xe_gt_sriov_pf_control.c
> > @@ -870,6 +870,16 @@ static int pf_handle_vf_save_data(struct xe_gt *gt, unsigned int vfid)
> >  		return -EAGAIN;
> >  	}
> >  
> > +	if (xe_gt_sriov_pf_migration_save_test(gt, vfid, XE_SRIOV_MIGRATION_DATA_TYPE_MMIO)) {
> > +		ret = xe_gt_sriov_pf_migration_mmio_save(gt, vfid);
> > +		if (ret)
> > +			return ret;
> > +
> > +		xe_gt_sriov_pf_migration_save_clear(gt, vfid, XE_SRIOV_MIGRATION_DATA_TYPE_MMIO);
> > +
> > +		return -EAGAIN;
> > +	}
> > +
> >  	return 0;
> >  }
> >  
> > @@ -1079,6 +1089,9 @@ static int pf_handle_vf_restore_data(struct xe_gt *gt, unsigned int vfid)
> >  	case XE_SRIOV_MIGRATION_DATA_TYPE_GGTT:
> >  		ret = xe_gt_sriov_pf_migration_ggtt_restore(gt, vfid, data);
> >  		break;
> > +	case XE_SRIOV_MIGRATION_DATA_TYPE_MMIO:
> > +		ret = xe_gt_sriov_pf_migration_mmio_restore(gt, vfid, data);
> > +		break;
> >  	case XE_SRIOV_MIGRATION_DATA_TYPE_GUC:
> >  		ret = xe_gt_sriov_pf_migration_guc_restore(gt, vfid, data);
> >  		break;
> > diff --git a/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.c b/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.c
> > index 6f2ee5820bdd4..5e90aeafeeb41 100644
> > --- a/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.c
> > +++ b/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.c
> > @@ -5,10 +5,13 @@
> >  
> >  #include <drm/drm_managed.h>
> >  
> > +#include "regs/xe_guc_regs.h"
> > +
> >  #include "abi/guc_actions_sriov_abi.h"
> >  #include "xe_bo.h"
> >  #include "xe_ggtt.h"
> >  #include "xe_gt.h"
> > +#include "xe_gt_sriov_pf.h"
> >  #include "xe_gt_sriov_pf_config.h"
> >  #include "xe_gt_sriov_pf_control.h"
> >  #include "xe_gt_sriov_pf_helpers.h"
> > @@ -16,6 +19,7 @@
> >  #include "xe_gt_sriov_printk.h"
> >  #include "xe_guc_buf.h"
> >  #include "xe_guc_ct.h"
> > +#include "xe_mmio.h"
> >  #include "xe_sriov.h"
> >  #include "xe_sriov_migration_data.h"
> >  #include "xe_sriov_pf_migration.h"
> > @@ -357,6 +361,150 @@ int xe_gt_sriov_pf_migration_guc_restore(struct xe_gt *gt, unsigned int vfid,
> >  	return pf_restore_vf_guc_state(gt, vfid, data);
> >  }
> >  
> > +static ssize_t pf_migration_mmio_size(struct xe_gt *gt, unsigned int vfid)
> > +{
> > +	if (xe_gt_is_media_type(gt))
> > +		return MED_VF_SW_FLAG_COUNT * sizeof(u32);
> > +	else
> > +		return VF_SW_FLAG_COUNT * sizeof(u32);
> > +}
> > +
> > +static int pf_migration_mmio_save(struct xe_gt *gt, unsigned int vfid, void *buf, size_t size)
> > +{
> > +	struct xe_mmio mmio;
> > +	u32 *regs = buf;
> > +	int n;
> > +
> > +	xe_mmio_init_vf_view(&mmio, &gt->mmio, vfid);
> > +
> > +	if (size != pf_migration_mmio_size(gt, vfid))
> > +		return -EINVAL;
> 
> you may want to check that first (before init vf view)

Ok.

> 
> > +
> > +	if (xe_gt_is_media_type(gt))
> > +		for (n = 0; n < MED_VF_SW_FLAG_COUNT; n++)
> > +			regs[n] = xe_mmio_read32(&gt->mmio, MED_VF_SW_FLAG(n));
> > +	else
> > +		for (n = 0; n < VF_SW_FLAG_COUNT; n++)
> > +			regs[n] = xe_mmio_read32(&gt->mmio, VF_SW_FLAG(n));
> > +
> > +	return 0;
> > +}
> > +
> > +static int pf_migration_mmio_restore(struct xe_gt *gt, unsigned int vfid,
> > +				     const void *buf, size_t size)
> > +{
> > +	const u32 *regs = buf;
> > +	struct xe_mmio mmio;
> > +	int n;
> > +
> > +	xe_mmio_init_vf_view(&mmio, &gt->mmio, vfid);
> > +
> > +	if (size != pf_migration_mmio_size(gt, vfid))
> > +		return -EINVAL;
> 
> ditto

Ok.

> 
> > +
> > +	if (xe_gt_is_media_type(gt))
> > +		for (n = 0; n < MED_VF_SW_FLAG_COUNT; n++)
> > +			xe_mmio_write32(&gt->mmio, MED_VF_SW_FLAG(n), regs[n]);
> > +	else
> > +		for (n = 0; n < VF_SW_FLAG_COUNT; n++)
> > +			xe_mmio_write32(&gt->mmio, VF_SW_FLAG(n), regs[n]);
> > +
> > +	return 0;
> > +}
> > +
> > +static int pf_save_vf_mmio_mig_data(struct xe_gt *gt, unsigned int vfid)
> > +{
> > +	struct xe_sriov_migration_data *data;
> > +	size_t size;
> > +	int ret;
> > +
> > +	size = pf_migration_mmio_size(gt, vfid);
> > +	xe_gt_assert(gt, size);
> > +
> > +	data = xe_sriov_migration_data_alloc(gt_to_xe(gt));
> > +	if (!data)
> > +		return -ENOMEM;
> > +
> > +	ret = xe_sriov_migration_data_init(data, gt->tile->id, gt->info.id,
> > +					   XE_SRIOV_MIGRATION_DATA_TYPE_MMIO, 0, size);
> > +	if (ret)
> > +		goto fail;
> > +
> > +	ret = pf_migration_mmio_save(gt, vfid, data->vaddr, size);
> > +	if (ret)
> > +		goto fail;
> > +
> > +	xe_gt_sriov_dbg_verbose(gt, "VF%u MMIO data save (%zu bytes)\n", vfid, size);
> 
> maybe make it part of the pf_dump_mig_data() ?
> 
> 	pf_dump_mig_data(gt, vfid, data, "MMIO saved");

Different patch - but I'll merge the two.

> 
> > +	pf_dump_mig_data(gt, vfid, data);
> > +
> > +	ret = xe_gt_sriov_pf_migration_save_produce(gt, vfid, data);
> > +	if (ret)
> > +		goto fail;
> > +
> > +	return 0;
> > +
> > +fail:
> > +	xe_sriov_migration_data_free(data);
> > +	xe_gt_sriov_err(gt, "Failed to save VF%u MMIO data (%pe)\n", vfid, ERR_PTR(ret));
> > +	return ret;
> > +}
> > +
> > +static int pf_restore_vf_mmio_mig_data(struct xe_gt *gt, unsigned int vfid,
> > +				       struct xe_sriov_migration_data *data)
> > +{
> > +	int ret;
> > +
> > +	xe_gt_sriov_dbg_verbose(gt, "VF%u MMIO data restore (%llu bytes)\n", vfid, data->size);
> 
> and here:

Ok.

> 
> 	pf_dump_mig_data(gt, vfid, data, "restoring MMIO");
> > +	pf_dump_mig_data(gt, vfid, data);
> > +
> > +	ret = pf_migration_mmio_restore(gt, vfid, data->vaddr, data->size);
> > +	if (ret) {
> > +		xe_gt_sriov_err(gt, "Failed to restore VF%u MMIO data (%pe)\n",
> > +				vfid, ERR_PTR(ret));
> > +
> > +		return ret;
> > +	}
> > +
> > +	return 0;
> > +}
> > +
> > +/**
> > + * xe_gt_sriov_pf_migration_mmio_save() - Save VF MMIO migration data.
> > + * @gt: the &xe_gt
> > + * @vfid: the VF identifier (can't be 0)
> > + *
> > + * This function is for PF only.
> > + *
> > + * Return: 0 on success or a negative error code on failure.
> > + */
> > +int xe_gt_sriov_pf_migration_mmio_save(struct xe_gt *gt, unsigned int vfid)
> > +{
> > +	xe_gt_assert(gt, IS_SRIOV_PF(gt_to_xe(gt)));
> > +	xe_gt_assert(gt, vfid != PFID);
> > +	xe_gt_assert(gt, vfid <= xe_sriov_pf_get_totalvfs(gt_to_xe(gt)));
> > +
> > +	return pf_save_vf_mmio_mig_data(gt, vfid);
> > +}
> > +
> > +/**
> > + * xe_gt_sriov_pf_migration_mmio_restore() - Restore VF MMIO migration data.
> > + * @gt: the &xe_gt
> > + * @vfid: the VF identifier (can't be 0)
> > + *
> > + * This function is for PF only.
> > + *
> > + * Return: 0 on success or a negative error code on failure.
> > + */
> > +int xe_gt_sriov_pf_migration_mmio_restore(struct xe_gt *gt, unsigned int vfid,
> > +					  struct xe_sriov_migration_data *data)
> > +{
> > +	xe_gt_assert(gt, IS_SRIOV_PF(gt_to_xe(gt)));
> > +	xe_gt_assert(gt, vfid != PFID);
> > +	xe_gt_assert(gt, vfid <= xe_sriov_pf_get_totalvfs(gt_to_xe(gt)));
> > +
> > +	return pf_restore_vf_mmio_mig_data(gt, vfid, data);
> > +}
> > +
> >  /**
> >   * xe_gt_sriov_pf_migration_size() - Total size of migration data from all components within a GT.
> >   * @gt: the &xe_gt
> > @@ -389,6 +537,13 @@ ssize_t xe_gt_sriov_pf_migration_size(struct xe_gt *gt, unsigned int vfid)
> >  		size += sizeof(struct xe_sriov_pf_migration_hdr);
> >  	total += size;
> >  
> > +	size = pf_migration_mmio_size(gt, vfid);
> > +	if (size < 0)
> > +		return size;
> > +	if (size > 0)
> > +		size += sizeof(struct xe_sriov_pf_migration_hdr);
> > +	total += size;
> > +
> >  	return total;
> >  }
> >  
> > @@ -453,6 +608,9 @@ void xe_gt_sriov_pf_migration_save_init(struct xe_gt *gt, unsigned int vfid)
> >  
> >  	if (pf_migration_ggtt_size(gt, vfid) > 0)
> >  		set_bit(XE_SRIOV_MIGRATION_DATA_TYPE_GGTT, &migration->save.data_remaining);
> > +
> > +	xe_gt_assert(gt, pf_migration_mmio_size(gt, vfid) > 0);
> > +	set_bit(XE_SRIOV_MIGRATION_DATA_TYPE_MMIO, &migration->save.data_remaining);
> >  }
> >  
> >  /**
> > diff --git a/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.h b/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.h
> > index bc201d8f3147a..b0eec94fea3a6 100644
> > --- a/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.h
> > +++ b/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.h
> > @@ -22,6 +22,9 @@ int xe_gt_sriov_pf_migration_guc_restore(struct xe_gt *gt, unsigned int vfid,
> >  int xe_gt_sriov_pf_migration_ggtt_save(struct xe_gt *gt, unsigned int vfid);
> >  int xe_gt_sriov_pf_migration_ggtt_restore(struct xe_gt *gt, unsigned int vfid,
> >  					  struct xe_sriov_migration_data *data);
> > +int xe_gt_sriov_pf_migration_mmio_save(struct xe_gt *gt, unsigned int vfid);
> > +int xe_gt_sriov_pf_migration_mmio_restore(struct xe_gt *gt, unsigned int vfid,
> > +					  struct xe_sriov_migration_data *data);
> >  
> >  ssize_t xe_gt_sriov_pf_migration_size(struct xe_gt *gt, unsigned int vfid);
> >  
> 
> but patch looks ok, so with #include fixed (and maybe with better dump helper),
> 
> Reviewed-by: Michal Wajdeczko <michal.wajdeczko@intel.com>
> 

Thanks,
-Michał

Re: [PATCH v3 22/28] drm/xe/pf: Handle VRAM migration data as part of PF control

[Michał Winiarski]

On Mon, Nov 03, 2025 at 02:37:29PM -0800, Matthew Brost wrote:
> On Thu, Oct 30, 2025 at 09:31:29PM +0100, Michał Winiarski wrote:
> > Connect the helpers to allow save and restore of VRAM migration data in
> > stop_copy / resume device state.
> > 
> > Co-developed-by: Lukasz Laguna <lukasz.laguna@intel.com>
> > Signed-off-by: Lukasz Laguna <lukasz.laguna@intel.com>
> > Signed-off-by: Michał Winiarski <michal.winiarski@intel.com>
> > ---
> >  drivers/gpu/drm/xe/xe_gt_sriov_pf_control.c   |  15 ++
> >  drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.c | 213 ++++++++++++++++++
> >  drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.h |   5 +
> >  .../drm/xe/xe_gt_sriov_pf_migration_types.h   |   2 +
> >  drivers/gpu/drm/xe/xe_sriov_pf_control.c      |   3 +
> >  5 files changed, 238 insertions(+)
> > 
> > diff --git a/drivers/gpu/drm/xe/xe_gt_sriov_pf_control.c b/drivers/gpu/drm/xe/xe_gt_sriov_pf_control.c
> > index 7cd7cae950bc7..de0f63610e780 100644
> > --- a/drivers/gpu/drm/xe/xe_gt_sriov_pf_control.c
> > +++ b/drivers/gpu/drm/xe/xe_gt_sriov_pf_control.c
> > @@ -880,6 +880,18 @@ static int pf_handle_vf_save_data(struct xe_gt *gt, unsigned int vfid)
> >  		return -EAGAIN;
> >  	}
> >  
> > +	if (xe_gt_sriov_pf_migration_save_test(gt, vfid, XE_SRIOV_MIGRATION_DATA_TYPE_VRAM)) {
> > +		ret = xe_gt_sriov_pf_migration_vram_save(gt, vfid);
> > +		if (ret == -EAGAIN)
> > +			return -EAGAIN;
> > +		else if (ret)
> > +			return ret;
> > +
> > +		xe_gt_sriov_pf_migration_save_clear(gt, vfid, XE_SRIOV_MIGRATION_DATA_TYPE_VRAM);
> > +
> > +		return -EAGAIN;
> > +	}
> > +
> >  	return 0;
> >  }
> >  
> > @@ -1095,6 +1107,9 @@ static int pf_handle_vf_restore_data(struct xe_gt *gt, unsigned int vfid)
> >  	case XE_SRIOV_MIGRATION_DATA_TYPE_GUC:
> >  		ret = xe_gt_sriov_pf_migration_guc_restore(gt, vfid, data);
> >  		break;
> > +	case XE_SRIOV_MIGRATION_DATA_TYPE_VRAM:
> > +		ret = xe_gt_sriov_pf_migration_vram_restore(gt, vfid, data);
> > +		break;
> >  	default:
> >  		xe_gt_sriov_notice(gt, "Skipping VF%u unknown data type: %d\n", vfid, data->type);
> >  		break;
> > diff --git a/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.c b/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.c
> > index 5e90aeafeeb41..7ccd43545fdac 100644
> > --- a/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.c
> > +++ b/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.c
> > @@ -19,6 +19,7 @@
> >  #include "xe_gt_sriov_printk.h"
> >  #include "xe_guc_buf.h"
> >  #include "xe_guc_ct.h"
> > +#include "xe_migrate.h"
> >  #include "xe_mmio.h"
> >  #include "xe_sriov.h"
> >  #include "xe_sriov_migration_data.h"
> > @@ -505,6 +506,207 @@ int xe_gt_sriov_pf_migration_mmio_restore(struct xe_gt *gt, unsigned int vfid,
> >  	return pf_restore_vf_mmio_mig_data(gt, vfid, data);
> >  }
> >  
> > +static ssize_t pf_migration_vram_size(struct xe_gt *gt, unsigned int vfid)
> > +{
> > +	if (!xe_gt_is_main_type(gt))
> > +		return 0;
> > +
> > +	return xe_gt_sriov_pf_config_get_lmem(gt, vfid);
> > +}
> > +
> > +static struct dma_fence *__pf_save_restore_vram(struct xe_gt *gt, unsigned int vfid,
> > +						struct xe_bo *vram, u64 vram_offset,
> > +						struct xe_bo *sysmem, u64 sysmem_offset,
> > +						size_t size, bool save)
> > +{
> > +	struct dma_fence *ret = NULL;
> > +	struct drm_exec exec;
> > +	int err;
> > +
> > +	drm_exec_init(&exec, DRM_EXEC_INTERRUPTIBLE_WAIT, 0);
> 
> It is a bit incongruent that the dma-resv lock is interruptible here.
> 
> > +	drm_exec_until_all_locked(&exec) {
> > +		err = drm_exec_lock_obj(&exec, &vram->ttm.base);
> > +		drm_exec_retry_on_contention(&exec);
> > +		if (err) {
> > +			ret = ERR_PTR(err);
> > +			goto err;
> > +		}
> > +
> > +		err = drm_exec_lock_obj(&exec, &sysmem->ttm.base);
> > +		drm_exec_retry_on_contention(&exec);
> > +		if (err) {
> > +			ret = ERR_PTR(err);
> > +			goto err;
> > +		}
> > +	}
> > +
> > +	ret = xe_migrate_vram_copy_chunk(vram, vram_offset, sysmem, sysmem_offset, size,
> > +					 save ? XE_MIGRATE_COPY_TO_SRAM : XE_MIGRATE_COPY_TO_VRAM);
> > +
> > +err:
> > +	drm_exec_fini(&exec);
> > +
> > +	return ret;
> > +}
> > +
> > +#define PF_VRAM_SAVE_RESTORE_TIMEOUT (5 * HZ)
> > +static int pf_save_vram_chunk(struct xe_gt *gt, unsigned int vfid,
> > +			      struct xe_bo *src_vram, u64 src_vram_offset,
> > +			      size_t size)
> > +{
> > +	struct xe_sriov_migration_data *data;
> > +	struct dma_fence *fence;
> > +	int ret;
> > +
> > +	data = xe_sriov_migration_data_alloc(gt_to_xe(gt));
> > +	if (!data)
> > +		return -ENOMEM;
> > +
> > +	ret = xe_sriov_migration_data_init(data, gt->tile->id, gt->info.id,
> > +					   XE_SRIOV_MIGRATION_DATA_TYPE_VRAM,
> > +					   src_vram_offset, size);
> > +	if (ret)
> > +		goto fail;
> > +
> > +	fence = __pf_save_restore_vram(gt, vfid,
> > +				       src_vram, src_vram_offset,
> > +				       data->bo, 0, size, true);
> > +
> > +	ret = dma_fence_wait_timeout(fence, false, PF_VRAM_SAVE_RESTORE_TIMEOUT);
> 
> Then here the dma-fence wait is non-interruptible.
> 
> > +	dma_fence_put(fence);
> > +	if (!ret) {
> > +		ret = -ETIME;
> > +		goto fail;
> > +	}
> > +
> > +	xe_gt_sriov_dbg_verbose(gt, "VF%u VRAM data save (%zu bytes)\n", vfid, size);
> > +	pf_dump_mig_data(gt, vfid, data);
> > +
> > +	ret = xe_gt_sriov_pf_migration_save_produce(gt, vfid, data);
> > +	if (ret)
> > +		goto fail;
> > +
> > +	return 0;
> > +
> > +fail:
> > +	xe_sriov_migration_data_free(data);
> > +	return ret;
> > +}
> > +
> > +#define VF_VRAM_STATE_CHUNK_MAX_SIZE SZ_512M
> > +static int pf_save_vf_vram_mig_data(struct xe_gt *gt, unsigned int vfid)
> > +{
> > +	struct xe_gt_sriov_migration_data *migration = pf_pick_gt_migration(gt, vfid);
> > +	loff_t *offset = &migration->save.vram_offset;
> > +	struct xe_bo *vram;
> > +	size_t vram_size, chunk_size;
> > +	int ret;
> > +
> > +	vram = xe_gt_sriov_pf_config_get_lmem_obj(gt, vfid);
> > +	if (!vram)
> > +		return -ENXIO;
> > +
> > +	vram_size = xe_bo_size(vram);
> > +
> > +	xe_gt_assert(gt, *offset < vram_size);
> > +
> > +	chunk_size = min(vram_size - *offset, VF_VRAM_STATE_CHUNK_MAX_SIZE);
> > +
> > +	ret = pf_save_vram_chunk(gt, vfid, vram, *offset, chunk_size);
> > +	if (ret)
> > +		goto fail;
> > +
> > +	*offset += chunk_size;
> > +
> > +	xe_bo_put(vram);
> > +
> > +	if (*offset < vram_size)
> > +		return -EAGAIN;
> > +
> > +	return 0;
> > +
> > +fail:
> > +	xe_bo_put(vram);
> > +	xe_gt_sriov_err(gt, "Failed to save VF%u VRAM data (%pe)\n", vfid, ERR_PTR(ret));
> > +	return ret;
> > +}
> > +
> > +static int pf_restore_vf_vram_mig_data(struct xe_gt *gt, unsigned int vfid,
> > +				       struct xe_sriov_migration_data *data)
> > +{
> > +	u64 end = data->hdr.offset + data->hdr.size;
> > +	struct dma_fence *fence;
> > +	struct xe_bo *vram;
> > +	size_t size;
> > +	int ret = 0;
> > +
> > +	vram = xe_gt_sriov_pf_config_get_lmem_obj(gt, vfid);
> > +	if (!vram)
> > +		return -ENXIO;
> > +
> > +	size = xe_bo_size(vram);
> > +
> > +	if (end > size || end < data->hdr.size) {
> > +		ret = -EINVAL;
> > +		goto err;
> > +	}
> > +
> > +	xe_gt_sriov_dbg_verbose(gt, "VF%u VRAM data restore (%llu bytes)\n", vfid, data->size);
> > +	pf_dump_mig_data(gt, vfid, data);
> > +
> > +	fence = __pf_save_restore_vram(gt, vfid, vram, data->hdr.offset,
> > +				       data->bo, 0, data->hdr.size, false);
> > +	ret = dma_fence_wait_timeout(fence, false, PF_VRAM_SAVE_RESTORE_TIMEOUT);
> 
> And also here the dma-fence wait is non-interruptible.
> 
> Aside from the incongruence, the usage of drm-exec, dma-fences looks correct.

All of this code is executed from the PF control state worker, so using
non-interruptible sleep is fine.
I'll drop the DRM_EXEC_INTERRUPTIBLE_WAIT flag from drm_exec_init().

Thanks,
-Michał

> 
> Matt
> 
> > +	dma_fence_put(fence);
> > +	if (!ret) {
> > +		ret = -ETIME;
> > +		goto err;
> > +	}
> > +
> > +	return 0;
> > +err:
> > +	xe_bo_put(vram);
> > +	xe_gt_sriov_err(gt, "Failed to restore VF%u VRAM data (%pe)\n", vfid, ERR_PTR(ret));
> > +	return ret;
> > +}
> > +
> > +/**
> > + * xe_gt_sriov_pf_migration_vram_save() - Save VF VRAM migration data.
> > + * @gt: the &xe_gt
> > + * @vfid: the VF identifier (can't be 0)
> > + *
> > + * This function is for PF only.
> > + *
> > + * Return: 0 on success or a negative error code on failure.
> > + */
> > +int xe_gt_sriov_pf_migration_vram_save(struct xe_gt *gt, unsigned int vfid)
> > +{
> > +	xe_gt_assert(gt, IS_SRIOV_PF(gt_to_xe(gt)));
> > +	xe_gt_assert(gt, vfid != PFID);
> > +	xe_gt_assert(gt, vfid <= xe_sriov_pf_get_totalvfs(gt_to_xe(gt)));
> > +
> > +	return pf_save_vf_vram_mig_data(gt, vfid);
> > +}
> > +
> > +/**
> > + * xe_gt_sriov_pf_migration_vram_restore() - Restore VF VRAM migration data.
> > + * @gt: the &xe_gt
> > + * @vfid: the VF identifier (can't be 0)
> > + *
> > + * This function is for PF only.
> > + *
> > + * Return: 0 on success or a negative error code on failure.
> > + */
> > +int xe_gt_sriov_pf_migration_vram_restore(struct xe_gt *gt, unsigned int vfid,
> > +					  struct xe_sriov_migration_data *data)
> > +{
> > +	xe_gt_assert(gt, IS_SRIOV_PF(gt_to_xe(gt)));
> > +	xe_gt_assert(gt, vfid != PFID);
> > +	xe_gt_assert(gt, vfid <= xe_sriov_pf_get_totalvfs(gt_to_xe(gt)));
> > +
> > +	return pf_restore_vf_vram_mig_data(gt, vfid, data);
> > +}
> > +
> >  /**
> >   * xe_gt_sriov_pf_migration_size() - Total size of migration data from all components within a GT.
> >   * @gt: the &xe_gt
> > @@ -544,6 +746,13 @@ ssize_t xe_gt_sriov_pf_migration_size(struct xe_gt *gt, unsigned int vfid)
> >  		size += sizeof(struct xe_sriov_pf_migration_hdr);
> >  	total += size;
> >  
> > +	size = pf_migration_vram_size(gt, vfid);
> > +	if (size < 0)
> > +		return size;
> > +	if (size > 0)
> > +		size += sizeof(struct xe_sriov_pf_migration_hdr);
> > +	total += size;
> > +
> >  	return total;
> >  }
> >  
> > @@ -602,6 +811,7 @@ void xe_gt_sriov_pf_migration_save_init(struct xe_gt *gt, unsigned int vfid)
> >  	struct xe_gt_sriov_migration_data *migration = pf_pick_gt_migration(gt, vfid);
> >  
> >  	migration->save.data_remaining = 0;
> > +	migration->save.vram_offset = 0;
> >  
> >  	xe_gt_assert(gt, pf_migration_guc_size(gt, vfid) > 0);
> >  	set_bit(XE_SRIOV_MIGRATION_DATA_TYPE_GUC, &migration->save.data_remaining);
> > @@ -611,6 +821,9 @@ void xe_gt_sriov_pf_migration_save_init(struct xe_gt *gt, unsigned int vfid)
> >  
> >  	xe_gt_assert(gt, pf_migration_mmio_size(gt, vfid) > 0);
> >  	set_bit(XE_SRIOV_MIGRATION_DATA_TYPE_MMIO, &migration->save.data_remaining);
> > +
> > +	if (pf_migration_vram_size(gt, vfid) > 0)
> > +		set_bit(XE_SRIOV_MIGRATION_DATA_TYPE_VRAM, &migration->save.data_remaining);
> >  }
> >  
> >  /**
> > diff --git a/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.h b/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.h
> > index b0eec94fea3a6..85b43b2cceb73 100644
> > --- a/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.h
> > +++ b/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration.h
> > @@ -25,6 +25,11 @@ int xe_gt_sriov_pf_migration_ggtt_restore(struct xe_gt *gt, unsigned int vfid,
> >  int xe_gt_sriov_pf_migration_mmio_save(struct xe_gt *gt, unsigned int vfid);
> >  int xe_gt_sriov_pf_migration_mmio_restore(struct xe_gt *gt, unsigned int vfid,
> >  					  struct xe_sriov_migration_data *data);
> > +int xe_gt_sriov_pf_migration_vram_save(struct xe_gt *gt, unsigned int vfid);
> > +int xe_gt_sriov_pf_migration_vram_restore(struct xe_gt *gt, unsigned int vfid,
> > +					  struct xe_sriov_migration_data *data);
> > +
> > +void xe_gt_sriov_pf_migration_save_init(struct xe_gt *gt, unsigned int vfid);
> >  
> >  ssize_t xe_gt_sriov_pf_migration_size(struct xe_gt *gt, unsigned int vfid);
> >  
> > diff --git a/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration_types.h b/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration_types.h
> > index 9f24878690d9c..f50c64241e9c0 100644
> > --- a/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration_types.h
> > +++ b/drivers/gpu/drm/xe/xe_gt_sriov_pf_migration_types.h
> > @@ -20,6 +20,8 @@ struct xe_gt_sriov_migration_data {
> >  	struct {
> >  		/** @save.data_remaining: bitmap of migration types that need to be saved */
> >  		unsigned long data_remaining;
> > +		/** @save.vram_offset: last saved offset within VRAM, used for chunked VRAM save */
> > +		loff_t vram_offset;
> >  	} save;
> >  };
> >  
> > diff --git a/drivers/gpu/drm/xe/xe_sriov_pf_control.c b/drivers/gpu/drm/xe/xe_sriov_pf_control.c
> > index c2768848daba1..aac8ecb861545 100644
> > --- a/drivers/gpu/drm/xe/xe_sriov_pf_control.c
> > +++ b/drivers/gpu/drm/xe/xe_sriov_pf_control.c
> > @@ -5,6 +5,7 @@
> >  
> >  #include "xe_device.h"
> >  #include "xe_gt_sriov_pf_control.h"
> > +#include "xe_gt_sriov_pf_migration.h"
> >  #include "xe_sriov_migration_data.h"
> >  #include "xe_sriov_pf_control.h"
> >  #include "xe_sriov_printk.h"
> > @@ -171,6 +172,8 @@ int xe_sriov_pf_control_trigger_save_vf(struct xe_device *xe, unsigned int vfid)
> >  		return ret;
> >  
> >  	for_each_gt(gt, xe, id) {
> > +		xe_gt_sriov_pf_migration_save_init(gt, vfid);
> > +
> >  		ret = xe_gt_sriov_pf_control_trigger_save_vf(gt, vfid);
> >  		if (ret)
> >  			return ret;
> > -- 
> > 2.50.1
> > 

Re: [PATCH v3 27/28] drm/intel/pciids: Add match with VFIO override

[Michał Winiarski]

On Mon, Nov 03, 2025 at 03:30:49PM -0600, Lucas De Marchi wrote:
> On Thu, Oct 30, 2025 at 09:31:34PM +0100, Michał Winiarski wrote:
> > In order to allow VFIO users to choose the right driver override, VFIO
> > driver variant used for VF migration needs to use Intel Graphics PCI
> > IDs.
> > Add INTEL_VGA_VFIO_DEVICE match that sets VFIO override_only.
> > 
> > Signed-off-by: Michał Winiarski <michal.winiarski@intel.com>
> > ---
> > include/drm/intel/pciids.h | 7 +++++++
> > 1 file changed, 7 insertions(+)
> > 
> > diff --git a/include/drm/intel/pciids.h b/include/drm/intel/pciids.h
> > index b258e79b437ac..d14ce43139a28 100644
> > --- a/include/drm/intel/pciids.h
> > +++ b/include/drm/intel/pciids.h
> > @@ -43,6 +43,13 @@
> > 	.class = PCI_BASE_CLASS_DISPLAY << 16, .class_mask = 0xff << 16, \
> > 	.driver_data = (kernel_ulong_t)(_info), \
> > }
> > +
> > +#define INTEL_VGA_VFIO_DEVICE(_id, _info) { \
> > +	PCI_DEVICE(PCI_VENDOR_ID_INTEL, (_id)), \
> > +	.class = PCI_BASE_CLASS_DISPLAY << 16, .class_mask = 0xff << 16, \
> > +	.driver_data = (kernel_ulong_t)(_info), \
> > +	.override_only = PCI_ID_F_VFIO_DRIVER_OVERRIDE, \
> 
> why do we need this and can't use PCI_DRIVER_OVERRIDE_DEVICE_VFIO()
> directly? Note that there are GPUs that wouldn't match the display class
> above.
> 
> 	edb660ad79ff ("drm/intel/pciids: Add match on vendor/id only")
> 	5e0de2dfbc1b ("drm/xe/cri: Add CRI platform definition")
> 
> Lucas De Marchi
> 

I'll define it on xe-vfio-pci side and use
PCI_DRIVER_OVERRIDE_DEVICE_VFIO() internally.

Thanks,
-Michał

> > +}
> > #endif
> > 
> > #define INTEL_I810_IDS(MACRO__, ...) \
> > -- 
> > 2.50.1
> > 

Re: [PATCH v3 27/28] drm/intel/pciids: Add match with VFIO override

[Lucas De Marchi]

On Tue, Nov 04, 2025 at 01:59:45PM +0100, Michał Winiarski wrote:
>On Mon, Nov 03, 2025 at 03:30:49PM -0600, Lucas De Marchi wrote:
>> On Thu, Oct 30, 2025 at 09:31:34PM +0100, Michał Winiarski wrote:
>> > In order to allow VFIO users to choose the right driver override, VFIO
>> > driver variant used for VF migration needs to use Intel Graphics PCI
>> > IDs.
>> > Add INTEL_VGA_VFIO_DEVICE match that sets VFIO override_only.
>> >
>> > Signed-off-by: Michał Winiarski <michal.winiarski@intel.com>
>> > ---
>> > include/drm/intel/pciids.h | 7 +++++++
>> > 1 file changed, 7 insertions(+)
>> >
>> > diff --git a/include/drm/intel/pciids.h b/include/drm/intel/pciids.h
>> > index b258e79b437ac..d14ce43139a28 100644
>> > --- a/include/drm/intel/pciids.h
>> > +++ b/include/drm/intel/pciids.h
>> > @@ -43,6 +43,13 @@
>> > 	.class = PCI_BASE_CLASS_DISPLAY << 16, .class_mask = 0xff << 16, \
>> > 	.driver_data = (kernel_ulong_t)(_info), \
>> > }
>> > +
>> > +#define INTEL_VGA_VFIO_DEVICE(_id, _info) { \
>> > +	PCI_DEVICE(PCI_VENDOR_ID_INTEL, (_id)), \
>> > +	.class = PCI_BASE_CLASS_DISPLAY << 16, .class_mask = 0xff << 16, \
>> > +	.driver_data = (kernel_ulong_t)(_info), \
>> > +	.override_only = PCI_ID_F_VFIO_DRIVER_OVERRIDE, \
>>
>> why do we need this and can't use PCI_DRIVER_OVERRIDE_DEVICE_VFIO()
>> directly? Note that there are GPUs that wouldn't match the display class
>> above.
>>
>> 	edb660ad79ff ("drm/intel/pciids: Add match on vendor/id only")
>> 	5e0de2dfbc1b ("drm/xe/cri: Add CRI platform definition")
>>
>> Lucas De Marchi
>>
>
>I'll define it on xe-vfio-pci side and use

but no matter where it's defined, why do you need it to match on the
class? The vid/devid should be sufficient.

Lucas De Marchi

Re: [PATCH v3 27/28] drm/intel/pciids: Add match with VFIO override

[Jason Gunthorpe]

On Tue, Nov 04, 2025 at 11:41:53AM -0600, Lucas De Marchi wrote:

> > > > +#define INTEL_VGA_VFIO_DEVICE(_id, _info) { \
> > > > +	PCI_DEVICE(PCI_VENDOR_ID_INTEL, (_id)), \
> > > > +	.class = PCI_BASE_CLASS_DISPLAY << 16, .class_mask = 0xff << 16, \
> > > > +	.driver_data = (kernel_ulong_t)(_info), \
> > > > +	.override_only = PCI_ID_F_VFIO_DRIVER_OVERRIDE, \
> > > 
> > > why do we need this and can't use PCI_DRIVER_OVERRIDE_DEVICE_VFIO()
> > > directly? Note that there are GPUs that wouldn't match the display class
> > > above.
> > > 
> > > 	edb660ad79ff ("drm/intel/pciids: Add match on vendor/id only")
> > > 	5e0de2dfbc1b ("drm/xe/cri: Add CRI platform definition")
> > > 
> > > Lucas De Marchi
> > > 
> > 
> > I'll define it on xe-vfio-pci side and use
> 
> but no matter where it's defined, why do you need it to match on the
> class? The vid/devid should be sufficient.

+1

Jason

Re: [PATCH v3 27/28] drm/intel/pciids: Add match with VFIO override

[Michał Winiarski]

On Tue, Nov 04, 2025 at 03:27:14PM -0400, Jason Gunthorpe wrote:
> On Tue, Nov 04, 2025 at 11:41:53AM -0600, Lucas De Marchi wrote:
> 
> > > > > +#define INTEL_VGA_VFIO_DEVICE(_id, _info) { \
> > > > > +	PCI_DEVICE(PCI_VENDOR_ID_INTEL, (_id)), \
> > > > > +	.class = PCI_BASE_CLASS_DISPLAY << 16, .class_mask = 0xff << 16, \
> > > > > +	.driver_data = (kernel_ulong_t)(_info), \
> > > > > +	.override_only = PCI_ID_F_VFIO_DRIVER_OVERRIDE, \
> > > > 
> > > > why do we need this and can't use PCI_DRIVER_OVERRIDE_DEVICE_VFIO()
> > > > directly? Note that there are GPUs that wouldn't match the display class
> > > > above.
> > > > 
> > > > 	edb660ad79ff ("drm/intel/pciids: Add match on vendor/id only")
> > > > 	5e0de2dfbc1b ("drm/xe/cri: Add CRI platform definition")
> > > > 
> > > > Lucas De Marchi
> > > > 
> > > 
> > > I'll define it on xe-vfio-pci side and use
> > 
> > but no matter where it's defined, why do you need it to match on the
> > class? The vid/devid should be sufficient.
> 
> +1
> 
> Jason

I don't need to match on class.

With PCI_DRIVER_OVERRIDE_DEVICE_VFIO it just becomes:
#define INTEL_PCI_VFIO_DEVICE(_id) { \
	PCI_DRIVER_OVERRIDE_DEVICE_VFIO(PCI_VENDOR_ID_INTEL, (_id)) \
}

static const struct pci_device_id xe_vfio_pci_table[] = {
	INTEL_PTL_IDS(INTEL_PCI_VFIO_DEVICE),
	INTEL_WCL_IDS(INTEL_PCI_VFIO_DEVICE),
	INTEL_BMG_IDS(INTEL_PCI_VFIO_DEVICE),
	{}
};

So, no matching on class, but I still do need a helper macro.

Thanks,
-Michał

Re: [PATCH v3 27/28] drm/intel/pciids: Add match with VFIO override

[Jason Gunthorpe]

On Wed, Nov 05, 2025 at 04:20:33PM +0100, Michał Winiarski wrote:
> On Tue, Nov 04, 2025 at 03:27:14PM -0400, Jason Gunthorpe wrote:
> > On Tue, Nov 04, 2025 at 11:41:53AM -0600, Lucas De Marchi wrote:
> > 
> > > > > > +#define INTEL_VGA_VFIO_DEVICE(_id, _info) { \
> > > > > > +	PCI_DEVICE(PCI_VENDOR_ID_INTEL, (_id)), \
> > > > > > +	.class = PCI_BASE_CLASS_DISPLAY << 16, .class_mask = 0xff << 16, \
> > > > > > +	.driver_data = (kernel_ulong_t)(_info), \
> > > > > > +	.override_only = PCI_ID_F_VFIO_DRIVER_OVERRIDE, \
> > > > > 
> > > > > why do we need this and can't use PCI_DRIVER_OVERRIDE_DEVICE_VFIO()
> > > > > directly? Note that there are GPUs that wouldn't match the display class
> > > > > above.
> > > > > 
> > > > > 	edb660ad79ff ("drm/intel/pciids: Add match on vendor/id only")
> > > > > 	5e0de2dfbc1b ("drm/xe/cri: Add CRI platform definition")
> > > > > 
> > > > > Lucas De Marchi
> > > > > 
> > > > 
> > > > I'll define it on xe-vfio-pci side and use
> > > 
> > > but no matter where it's defined, why do you need it to match on the
> > > class? The vid/devid should be sufficient.
> > 
> > +1
> > 
> > Jason
> 
> I don't need to match on class.
> 
> With PCI_DRIVER_OVERRIDE_DEVICE_VFIO it just becomes:
> #define INTEL_PCI_VFIO_DEVICE(_id) { \
> 	PCI_DRIVER_OVERRIDE_DEVICE_VFIO(PCI_VENDOR_ID_INTEL, (_id)) \
> }
> 
> static const struct pci_device_id xe_vfio_pci_table[] = {
> 	INTEL_PTL_IDS(INTEL_PCI_VFIO_DEVICE),
> 	INTEL_WCL_IDS(INTEL_PCI_VFIO_DEVICE),
> 	INTEL_BMG_IDS(INTEL_PCI_VFIO_DEVICE),
> 	{}
> };
> 
> So, no matching on class, but I still do need a helper macro.

Yes, that looks right to me.

Jason

Re: [PATCH v3 27/28] drm/intel/pciids: Add match with VFIO override

[Lucas De Marchi]

On Wed, Nov 05, 2025 at 04:20:33PM +0100, Michał Winiarski wrote:
>On Tue, Nov 04, 2025 at 03:27:14PM -0400, Jason Gunthorpe wrote:
>> On Tue, Nov 04, 2025 at 11:41:53AM -0600, Lucas De Marchi wrote:
>>
>> > > > > +#define INTEL_VGA_VFIO_DEVICE(_id, _info) { \
>> > > > > +	PCI_DEVICE(PCI_VENDOR_ID_INTEL, (_id)), \
>> > > > > +	.class = PCI_BASE_CLASS_DISPLAY << 16, .class_mask = 0xff << 16, \
>> > > > > +	.driver_data = (kernel_ulong_t)(_info), \
>> > > > > +	.override_only = PCI_ID_F_VFIO_DRIVER_OVERRIDE, \
>> > > >
>> > > > why do we need this and can't use PCI_DRIVER_OVERRIDE_DEVICE_VFIO()
>> > > > directly? Note that there are GPUs that wouldn't match the display class
>> > > > above.
>> > > >
>> > > > 	edb660ad79ff ("drm/intel/pciids: Add match on vendor/id only")
>> > > > 	5e0de2dfbc1b ("drm/xe/cri: Add CRI platform definition")
>> > > >
>> > > > Lucas De Marchi
>> > > >
>> > >
>> > > I'll define it on xe-vfio-pci side and use
>> >
>> > but no matter where it's defined, why do you need it to match on the
>> > class? The vid/devid should be sufficient.
>>
>> +1
>>
>> Jason
>
>I don't need to match on class.
>
>With PCI_DRIVER_OVERRIDE_DEVICE_VFIO it just becomes:
>#define INTEL_PCI_VFIO_DEVICE(_id) { \
>	PCI_DRIVER_OVERRIDE_DEVICE_VFIO(PCI_VENDOR_ID_INTEL, (_id)) \
>}
>
>static const struct pci_device_id xe_vfio_pci_table[] = {
>	INTEL_PTL_IDS(INTEL_PCI_VFIO_DEVICE),
>	INTEL_WCL_IDS(INTEL_PCI_VFIO_DEVICE),
>	INTEL_BMG_IDS(INTEL_PCI_VFIO_DEVICE),
>	{}
>};
>
>So, no matching on class, but I still do need a helper macro.

yeah, that lgtm

thanks
Lucas De Marchi

>
>Thanks,
>-Michał

RE: [PATCH v3 28/28] vfio/xe: Add device specific vfio_pci driver variant for Intel graphics

[Muqthyar Ahmed, Syed Abdul]

> -----Original Message-----
> From: Intel-xe <intel-xe-bounces@lists.freedesktop.org> On Behalf Of Michał
> Winiarski
> Sent: Friday, October 31, 2025 2:02 AM
> To: Alex Williamson <alex@shazbot.org>; De Marchi, Lucas
> <lucas.demarchi@intel.com>; Thomas Hellström
> <thomas.hellstrom@linux.intel.com>; Vivi, Rodrigo <rodrigo.vivi@intel.com>;
> Jason Gunthorpe <jgg@ziepe.ca>; Yishai Hadas <yishaih@nvidia.com>; Tian,
> Kevin <kevin.tian@intel.com>; Shameer Kolothum
> <skolothumtho@nvidia.com>; intel-xe@lists.freedesktop.org; linux-
> kernel@vger.kernel.org; kvm@vger.kernel.org; Brost, Matthew
> <matthew.brost@intel.com>; Wajdeczko, Michal
> <Michal.Wajdeczko@intel.com>
> Cc: dri-devel@lists.freedesktop.org; Jani Nikula <jani.nikula@linux.intel.com>;
> Joonas Lahtinen <joonas.lahtinen@linux.intel.com>; Tvrtko Ursulin
> <tursulin@ursulin.net>; David Airlie <airlied@gmail.com>; Simona Vetter
> <simona@ffwll.ch>; Laguna, Lukasz <lukasz.laguna@intel.com>; Christoph
> Hellwig <hch@infradead.org>; Winiarski, Michal
> <michal.winiarski@intel.com>
> Subject: [PATCH v3 28/28] vfio/xe: Add device specific vfio_pci driver variant
> for Intel graphics
> 
> In addition to generic VFIO PCI functionality, the driver implements VFIO
> migration uAPI, allowing userspace to enable migration for Intel Graphics SR-
> IOV Virtual Functions.
> The driver binds to VF device, and uses API exposed by Xe driver bound to PF
> device to control VF device state and transfer the migration data.
> 
> Signed-off-by: Michał Winiarski <michal.winiarski@intel.com>
> ---
>  MAINTAINERS                  |   7 +
>  drivers/vfio/pci/Kconfig     |   2 +
>  drivers/vfio/pci/Makefile    |   2 +
>  drivers/vfio/pci/xe/Kconfig  |  12 +
>  drivers/vfio/pci/xe/Makefile |   3 +
>  drivers/vfio/pci/xe/main.c   | 552 +++++++++++++++++++++++++++++++++++
>  6 files changed, 578 insertions(+)
>  create mode 100644 drivers/vfio/pci/xe/Kconfig  create mode 100644
> drivers/vfio/pci/xe/Makefile  create mode 100644 drivers/vfio/pci/xe/main.c
> 
> diff --git a/MAINTAINERS b/MAINTAINERS
> index b890ff265f03f..d73348c5f3f3e 100644
> --- a/MAINTAINERS
> +++ b/MAINTAINERS
> @@ -27008,6 +27008,13 @@ L:	virtualization@lists.linux.dev
>  S:	Maintained
>  F:	drivers/vfio/pci/virtio
> 
> +VFIO XE PCI DRIVER
> +M:	Michał Winiarski <michal.winiarski@intel.com>
> +L:	kvm@vger.kernel.org
> +L:	intel-xe@lists.freedesktop.org
> +S:	Supported
> +F:	drivers/vfio/pci/xe
> +
>  VGA_SWITCHEROO
>  R:	Lukas Wunner <lukas@wunner.de>
>  S:	Maintained
> diff --git a/drivers/vfio/pci/Kconfig b/drivers/vfio/pci/Kconfig index
> 2b0172f546652..c100f0ab87f2d 100644
> --- a/drivers/vfio/pci/Kconfig
> +++ b/drivers/vfio/pci/Kconfig
> @@ -67,4 +67,6 @@ source "drivers/vfio/pci/nvgrace-gpu/Kconfig"
> 
>  source "drivers/vfio/pci/qat/Kconfig"
> 
> +source "drivers/vfio/pci/xe/Kconfig"
> +
>  endmenu
> diff --git a/drivers/vfio/pci/Makefile b/drivers/vfio/pci/Makefile index
> cf00c0a7e55c8..f5d46aa9347b9 100644
> --- a/drivers/vfio/pci/Makefile
> +++ b/drivers/vfio/pci/Makefile
> @@ -19,3 +19,5 @@ obj-$(CONFIG_VIRTIO_VFIO_PCI) += virtio/
>  obj-$(CONFIG_NVGRACE_GPU_VFIO_PCI) += nvgrace-gpu/
> 
>  obj-$(CONFIG_QAT_VFIO_PCI) += qat/
> +
> +obj-$(CONFIG_XE_VFIO_PCI) += xe/
> diff --git a/drivers/vfio/pci/xe/Kconfig b/drivers/vfio/pci/xe/Kconfig new file
> mode 100644 index 0000000000000..787be88268685
> --- /dev/null
> +++ b/drivers/vfio/pci/xe/Kconfig
> @@ -0,0 +1,12 @@
> +# SPDX-License-Identifier: GPL-2.0-only config XE_VFIO_PCI
> +	tristate "VFIO support for Intel Graphics"
> +	depends on DRM_XE
> +	select VFIO_PCI_CORE
Shall we make it default enabled with "default m " so that config gets enabled for first timers and no need to enable explicitly. 

> +	help
> +	  This option enables vendor-specific VFIO driver for Intel Graphics.
> +	  In addition to generic VFIO PCI functionality, it implements VFIO
> +	  migration uAPI allowing userspace to enable migration for
> +	  Intel Graphics SR-IOV Virtual Functions supported by the Xe driver.
> +
> +	  If you don't know what to do here, say N.
> diff --git a/drivers/vfio/pci/xe/Makefile b/drivers/vfio/pci/xe/Makefile new file
> mode 100644 index 0000000000000..13aa0fd192cd4
> --- /dev/null
> +++ b/drivers/vfio/pci/xe/Makefile
> @@ -0,0 +1,3 @@
> +# SPDX-License-Identifier: GPL-2.0-only
> +obj-$(CONFIG_XE_VFIO_PCI) += xe-vfio-pci.o xe-vfio-pci-y := main.o
> diff --git a/drivers/vfio/pci/xe/main.c b/drivers/vfio/pci/xe/main.c new file
> mode 100644 index 0000000000000..4892c6648c0a7
> --- /dev/null
> +++ b/drivers/vfio/pci/xe/main.c
> @@ -0,0 +1,552 @@
> +// SPDX-License-Identifier: GPL-2.0-only
> +/*
> + * Copyright © 2025 Intel Corporation
> + */
> +
> +#include <linux/anon_inodes.h>
> +#include <linux/delay.h>
> +#include <linux/file.h>
> +#include <linux/module.h>
> +#include <linux/pci.h>
> +#include <linux/sizes.h>
> +#include <linux/types.h>
> +#include <linux/vfio.h>
> +#include <linux/vfio_pci_core.h>
> +
> +#include <drm/intel/xe_sriov_vfio.h>
> +#include <drm/intel/pciids.h>
> +
> +struct xe_vfio_pci_migration_file {
> +	struct file *filp;
> +	/* serializes accesses to migration data */
> +	struct mutex lock;
> +	struct xe_vfio_pci_core_device *xe_vdev; };
> +
> +struct xe_vfio_pci_core_device {
> +	struct vfio_pci_core_device core_device;
> +	struct xe_device *xe;
> +	/* VF number used by PF, Xe HW/FW components use vfid indexing
> starting from 1 */
> +	unsigned int vfid;
> +	u8 migrate_cap:1;
> +	u8 deferred_reset:1;
> +	/* protects migration state */
> +	struct mutex state_mutex;
> +	enum vfio_device_mig_state mig_state;
> +	/* protects the reset_done flow */
> +	spinlock_t reset_lock;
> +	struct xe_vfio_pci_migration_file *migf; };
> +
> +#define xe_vdev_to_dev(xe_vdev) (&(xe_vdev)->core_device.pdev->dev)
> +
> +static void xe_vfio_pci_disable_file(struct xe_vfio_pci_migration_file
> +*migf) {
> +	struct xe_vfio_pci_core_device *xe_vdev = migf->xe_vdev;
> +
> +	mutex_lock(&migf->lock);
> +	xe_vdev->migf = NULL;
> +	mutex_unlock(&migf->lock);
> +}
> +
> +static void xe_vfio_pci_reset(struct xe_vfio_pci_core_device *xe_vdev)
> +{
> +	if (xe_vdev->migf)
> +		xe_vfio_pci_disable_file(xe_vdev->migf);
> +
> +	xe_vdev->mig_state = VFIO_DEVICE_STATE_RUNNING; }
> +
> +static void xe_vfio_pci_state_mutex_lock(struct xe_vfio_pci_core_device
> +*xe_vdev) {
> +	mutex_lock(&xe_vdev->state_mutex);
> +}
> +
> +/*
> + * This function is called in all state_mutex unlock cases to
> + * handle a 'deferred_reset' if exists.
> + */
> +static void xe_vfio_pci_state_mutex_unlock(struct
> +xe_vfio_pci_core_device *xe_vdev) {
> +again:
> +	spin_lock(&xe_vdev->reset_lock);
> +	if (xe_vdev->deferred_reset) {
> +		xe_vdev->deferred_reset = false;
> +		spin_unlock(&xe_vdev->reset_lock);
> +		xe_vfio_pci_reset(xe_vdev);
> +		goto again;
> +	}
> +	mutex_unlock(&xe_vdev->state_mutex);
> +	spin_unlock(&xe_vdev->reset_lock);
> +}
> +
> +static void xe_vfio_pci_reset_done(struct pci_dev *pdev) {
> +	struct xe_vfio_pci_core_device *xe_vdev = pci_get_drvdata(pdev);
> +	int ret;
> +
> +	if (!xe_vdev->vfid)
> +		return;
> +
> +	/*
> +	 * VF FLR requires additional processing done by PF driver.
> +	 * The processing is done after FLR is already finished from PCIe
> +	 * perspective.
> +	 * In order to avoid a scenario where VF is used while PF processing
> +	 * is still in progress, additional synchronization point is needed.
> +	 */
> +	ret = xe_sriov_vfio_wait_flr_done(xe_vdev->xe, xe_vdev->vfid);
> +	if (ret)
> +		dev_err(&pdev->dev, "Failed to wait for FLR: %d\n", ret);
> +
> +	if (!xe_vdev->migrate_cap)
> +		return;
> +
> +	/*
> +	 * As the higher VFIO layers are holding locks across reset and using
> +	 * those same locks with the mm_lock we need to prevent ABBA
> deadlock
> +	 * with the state_mutex and mm_lock.
> +	 * In case the state_mutex was taken already we defer the cleanup
> work
> +	 * to the unlock flow of the other running context.
> +	 */
> +	spin_lock(&xe_vdev->reset_lock);
> +	xe_vdev->deferred_reset = true;
> +	if (!mutex_trylock(&xe_vdev->state_mutex)) {
> +		spin_unlock(&xe_vdev->reset_lock);
> +		return;
> +	}
> +	spin_unlock(&xe_vdev->reset_lock);
> +	xe_vfio_pci_state_mutex_unlock(xe_vdev);
> +
> +	xe_vfio_pci_reset(xe_vdev);
> +}
> +
> +static const struct pci_error_handlers xe_vfio_pci_err_handlers = {
> +	.reset_done = xe_vfio_pci_reset_done,
> +	.error_detected = vfio_pci_core_aer_err_detected, };
> +
> +static int xe_vfio_pci_open_device(struct vfio_device *core_vdev) {
> +	struct xe_vfio_pci_core_device *xe_vdev =
> +		container_of(core_vdev, struct xe_vfio_pci_core_device,
> core_device.vdev);
> +	struct vfio_pci_core_device *vdev = &xe_vdev->core_device;
> +	int ret;
> +
> +	ret = vfio_pci_core_enable(vdev);
> +	if (ret)
> +		return ret;
> +
> +	vfio_pci_core_finish_enable(vdev);
> +
> +	return 0;
> +}
> +
> +static int xe_vfio_pci_release_file(struct inode *inode, struct file
> +*filp) {
> +	struct xe_vfio_pci_migration_file *migf = filp->private_data;
> +
> +	xe_vfio_pci_disable_file(migf);
> +	mutex_destroy(&migf->lock);
> +	kfree(migf);
> +
> +	return 0;
> +}
> +
> +static ssize_t xe_vfio_pci_save_read(struct file *filp, char __user
> +*buf, size_t len, loff_t *pos) {
> +	struct xe_vfio_pci_migration_file *migf = filp->private_data;
> +	ssize_t ret;
> +
> +	if (pos)
> +		return -ESPIPE;
> +
> +	mutex_lock(&migf->lock);
> +	ret = xe_sriov_vfio_data_read(migf->xe_vdev->xe, migf->xe_vdev-
> >vfid, buf, len);
> +	mutex_unlock(&migf->lock);
> +
> +	return ret;
> +}
> +
> +static const struct file_operations xe_vfio_pci_save_fops = {
> +	.owner = THIS_MODULE,
> +	.read = xe_vfio_pci_save_read,
> +	.release = xe_vfio_pci_release_file,
> +	.llseek = noop_llseek,
> +};
> +
> +static ssize_t xe_vfio_pci_resume_write(struct file *filp, const char __user
> *buf,
> +					size_t len, loff_t *pos)
> +{
> +	struct xe_vfio_pci_migration_file *migf = filp->private_data;
> +	ssize_t ret;
> +
> +	if (pos)
> +		return -ESPIPE;
> +
> +	mutex_lock(&migf->lock);
> +	ret = xe_sriov_vfio_data_write(migf->xe_vdev->xe, migf->xe_vdev-
> >vfid, buf, len);
> +	mutex_unlock(&migf->lock);
> +
> +	return ret;
> +}
> +
> +static const struct file_operations xe_vfio_pci_resume_fops = {
> +	.owner = THIS_MODULE,
> +	.write = xe_vfio_pci_resume_write,
> +	.release = xe_vfio_pci_release_file,
> +	.llseek = noop_llseek,
> +};
> +
> +static const char *vfio_dev_state_str(u32 state) {
> +	switch (state) {
> +	case VFIO_DEVICE_STATE_RUNNING: return "running";
> +	case VFIO_DEVICE_STATE_RUNNING_P2P: return "running_p2p";
> +	case VFIO_DEVICE_STATE_STOP_COPY: return "stopcopy";
> +	case VFIO_DEVICE_STATE_STOP: return "stop";
> +	case VFIO_DEVICE_STATE_RESUMING: return "resuming";
> +	case VFIO_DEVICE_STATE_ERROR: return "error";
> +	default: return "";
> +	}
> +}
> +
> +enum xe_vfio_pci_file_type {
> +	XE_VFIO_FILE_SAVE = 0,
> +	XE_VFIO_FILE_RESUME,
> +};
> +
> +static struct xe_vfio_pci_migration_file *
> +xe_vfio_pci_alloc_file(struct xe_vfio_pci_core_device *xe_vdev,
> +		       enum xe_vfio_pci_file_type type) {
> +	struct xe_vfio_pci_migration_file *migf;
> +	const struct file_operations *fops;
> +	int flags;
> +
> +	migf = kzalloc(sizeof(*migf), GFP_KERNEL);
> +	if (!migf)
> +		return ERR_PTR(-ENOMEM);
> +
> +	fops = type == XE_VFIO_FILE_SAVE ? &xe_vfio_pci_save_fops :
> &xe_vfio_pci_resume_fops;
> +	flags = type == XE_VFIO_FILE_SAVE ? O_RDONLY : O_WRONLY;
> +	migf->filp = anon_inode_getfile("xe_vfio_mig", fops, migf, flags);
> +	if (IS_ERR(migf->filp)) {
> +		kfree(migf);
> +		return ERR_CAST(migf->filp);
> +	}
> +
> +	mutex_init(&migf->lock);
> +	migf->xe_vdev = xe_vdev;
> +	xe_vdev->migf = migf;
> +
> +	stream_open(migf->filp->f_inode, migf->filp);
> +
> +	return migf;
> +}
> +
> +static struct file *
> +xe_vfio_set_state(struct xe_vfio_pci_core_device *xe_vdev, u32 new) {
> +	u32 cur = xe_vdev->mig_state;
> +	int ret;
> +
> +	dev_dbg(xe_vdev_to_dev(xe_vdev),
> +		"state: %s->%s\n", vfio_dev_state_str(cur),
> vfio_dev_state_str(new));
> +
> +	/*
> +	 * "STOP" handling is reused for "RUNNING_P2P", as the device
> doesn't
> +	 * have the capability to selectively block outgoing p2p DMA transfers.
> +	 * While the device is allowing BAR accesses when the VF is stopped,
> it
> +	 * is not processing any new workload requests, effectively stopping
> +	 * any outgoing DMA transfers (not just p2p).
> +	 * Any VRAM / MMIO accesses occurring during "RUNNING_P2P" are
> kept and
> +	 * will be migrated to target VF during stop-copy.
> +	 */
> +	if (cur == VFIO_DEVICE_STATE_RUNNING && new ==
> VFIO_DEVICE_STATE_RUNNING_P2P) {
> +		ret = xe_sriov_vfio_suspend_device(xe_vdev->xe, xe_vdev-
> >vfid);
> +		if (ret)
> +			goto err;
> +
> +		return NULL;
> +	}
> +
> +	if ((cur == VFIO_DEVICE_STATE_RUNNING_P2P && new ==
> VFIO_DEVICE_STATE_STOP) ||
> +	    (cur == VFIO_DEVICE_STATE_STOP && new ==
> VFIO_DEVICE_STATE_RUNNING_P2P))
> +		return NULL;
> +
> +	if (cur == VFIO_DEVICE_STATE_RUNNING_P2P && new ==
> VFIO_DEVICE_STATE_RUNNING) {
> +		ret = xe_sriov_vfio_resume_device(xe_vdev->xe, xe_vdev-
> >vfid);
> +		if (ret)
> +			goto err;
> +
> +		return NULL;
> +	}
> +
> +	if (cur == VFIO_DEVICE_STATE_STOP && new ==
> VFIO_DEVICE_STATE_STOP_COPY) {
> +		struct xe_vfio_pci_migration_file *migf;
> +
> +		migf = xe_vfio_pci_alloc_file(xe_vdev, XE_VFIO_FILE_SAVE);
> +		if (IS_ERR(migf)) {
> +			ret = PTR_ERR(migf);
> +			goto err;
> +		}
> +		get_file(migf->filp);
> +
> +		ret = xe_sriov_vfio_stop_copy_enter(xe_vdev->xe, xe_vdev-
> >vfid);
> +		if (ret) {
> +			fput(migf->filp);
> +			goto err;
> +		}
> +
> +		return migf->filp;
> +	}
> +
> +	if (cur == VFIO_DEVICE_STATE_STOP_COPY && new ==
> VFIO_DEVICE_STATE_STOP) {
> +		if (xe_vdev->migf) {
> +			fput(xe_vdev->migf->filp);
> +			xe_vfio_pci_disable_file(xe_vdev->migf);
> +		}
> +
> +		ret = xe_sriov_vfio_stop_copy_exit(xe_vdev->xe, xe_vdev-
> >vfid);
> +		if (ret)
> +			goto err;
> +
> +		return NULL;
> +	}
> +
> +	if (cur == VFIO_DEVICE_STATE_STOP && new ==
> VFIO_DEVICE_STATE_RESUMING) {
> +		struct xe_vfio_pci_migration_file *migf;
> +
> +		migf = xe_vfio_pci_alloc_file(xe_vdev,
> XE_VFIO_FILE_RESUME);
> +		if (IS_ERR(migf)) {
> +			ret = PTR_ERR(migf);
> +			goto err;
> +		}
> +		get_file(migf->filp);
> +
> +		ret = xe_sriov_vfio_resume_data_enter(xe_vdev->xe,
> xe_vdev->vfid);
> +		if (ret) {
> +			fput(migf->filp);
> +			goto err;
> +		}
> +
> +		return migf->filp;
> +	}
> +
> +	if (cur == VFIO_DEVICE_STATE_RESUMING && new ==
> VFIO_DEVICE_STATE_STOP) {
> +		if (xe_vdev->migf) {
> +			fput(xe_vdev->migf->filp);
> +			xe_vfio_pci_disable_file(xe_vdev->migf);
> +		}
> +
> +		ret = xe_sriov_vfio_resume_data_exit(xe_vdev->xe, xe_vdev-
> >vfid);
> +		if (ret)
> +			goto err;
> +
> +		return NULL;
> +	}
> +
> +	WARN(true, "Unknown state transition %d->%d", cur, new);
> +	return ERR_PTR(-EINVAL);
> +
> +err:
> +	dev_dbg(xe_vdev_to_dev(xe_vdev),
> +		"Failed to transition state: %s->%s err=%d\n",
> +		vfio_dev_state_str(cur), vfio_dev_state_str(new), ret);
> +	return ERR_PTR(ret);
> +}
> +
> +static struct file *
> +xe_vfio_pci_set_device_state(struct vfio_device *core_vdev,
> +			     enum vfio_device_mig_state new_state) {
> +	struct xe_vfio_pci_core_device *xe_vdev =
> +		container_of(core_vdev, struct xe_vfio_pci_core_device,
> core_device.vdev);
> +	enum vfio_device_mig_state next_state;
> +	struct file *f = NULL;
> +	int ret;
> +
> +	xe_vfio_pci_state_mutex_lock(xe_vdev);
> +	while (new_state != xe_vdev->mig_state) {
> +		ret = vfio_mig_get_next_state(core_vdev, xe_vdev->mig_state,
> +					      new_state, &next_state);
> +		if (ret) {
> +			xe_sriov_vfio_error(xe_vdev->xe, xe_vdev->vfid);
> +			f = ERR_PTR(ret);
> +			break;
> +		}
> +		f = xe_vfio_set_state(xe_vdev, next_state);
> +		if (IS_ERR(f))
> +			break;
> +
> +		xe_vdev->mig_state = next_state;
> +
> +		/* Multiple state transitions with non-NULL file in the middle
> */
> +		if (f && new_state != xe_vdev->mig_state) {
> +			fput(f);
> +			f = ERR_PTR(-EINVAL);
> +			break;
> +		}
> +	}
> +	xe_vfio_pci_state_mutex_unlock(xe_vdev);
> +
> +	return f;
> +}
> +
> +static int xe_vfio_pci_get_device_state(struct vfio_device *core_vdev,
> +					enum vfio_device_mig_state
> *curr_state) {
> +	struct xe_vfio_pci_core_device *xe_vdev =
> +		container_of(core_vdev, struct xe_vfio_pci_core_device,
> +core_device.vdev);
> +
> +	xe_vfio_pci_state_mutex_lock(xe_vdev);
> +	*curr_state = xe_vdev->mig_state;
> +	xe_vfio_pci_state_mutex_unlock(xe_vdev);
> +
> +	return 0;
> +}
> +
> +static int xe_vfio_pci_get_data_size(struct vfio_device *vdev,
> +				     unsigned long *stop_copy_length) {
> +	struct xe_vfio_pci_core_device *xe_vdev =
> +		container_of(vdev, struct xe_vfio_pci_core_device,
> core_device.vdev);
> +
> +	xe_vfio_pci_state_mutex_lock(xe_vdev);
> +	*stop_copy_length = xe_sriov_vfio_stop_copy_size(xe_vdev->xe,
> xe_vdev->vfid);
> +	xe_vfio_pci_state_mutex_unlock(xe_vdev);
> +
> +	return 0;
> +}
> +
> +static const struct vfio_migration_ops xe_vfio_pci_migration_ops = {
> +	.migration_set_state = xe_vfio_pci_set_device_state,
> +	.migration_get_state = xe_vfio_pci_get_device_state,
> +	.migration_get_data_size = xe_vfio_pci_get_data_size, };
> +
> +static void xe_vfio_pci_migration_init(struct xe_vfio_pci_core_device
> +*xe_vdev) {
> +	struct vfio_device *core_vdev = &xe_vdev->core_device.vdev;
> +	struct pci_dev *pdev = to_pci_dev(core_vdev->dev);
> +	struct xe_device *xe = xe_sriov_vfio_get_xe_device(pdev);
> +	int ret;
> +
> +	if (!xe)
> +		return;
> +	if (!xe_sriov_vfio_migration_supported(xe))
> +		return;
> +
> +	ret = pci_iov_vf_id(pdev);
> +	if (ret < 0)
> +		return;
> +
> +	mutex_init(&xe_vdev->state_mutex);
> +	spin_lock_init(&xe_vdev->reset_lock);
> +
> +	/* Xe HW/FW components use vfid indexing starting from 1 */
> +	xe_vdev->vfid = ret + 1;
> +	xe_vdev->xe = xe;
> +	xe_vdev->migrate_cap = true;
> +
> +	core_vdev->migration_flags = VFIO_MIGRATION_STOP_COPY |
> VFIO_MIGRATION_P2P;
> +	core_vdev->mig_ops = &xe_vfio_pci_migration_ops; }
> +
> +static void xe_vfio_pci_migration_fini(struct xe_vfio_pci_core_device
> +*xe_vdev) {
> +	if (!xe_vdev->migrate_cap)
> +		return;
> +
> +	mutex_destroy(&xe_vdev->state_mutex);
> +}
> +
> +static int xe_vfio_pci_init_dev(struct vfio_device *core_vdev) {
> +	struct xe_vfio_pci_core_device *xe_vdev =
> +		container_of(core_vdev, struct xe_vfio_pci_core_device,
> +core_device.vdev);
> +
> +	xe_vfio_pci_migration_init(xe_vdev);
> +
> +	return vfio_pci_core_init_dev(core_vdev);
> +}
> +
> +static void xe_vfio_pci_release_dev(struct vfio_device *core_vdev) {
> +	struct xe_vfio_pci_core_device *xe_vdev =
> +		container_of(core_vdev, struct xe_vfio_pci_core_device,
> +core_device.vdev);
> +
> +	xe_vfio_pci_migration_fini(xe_vdev);
> +}
> +
> +static const struct vfio_device_ops xe_vfio_pci_ops = {
> +	.name = "xe-vfio-pci",
> +	.init = xe_vfio_pci_init_dev,
> +	.release = xe_vfio_pci_release_dev,
> +	.open_device = xe_vfio_pci_open_device,
> +	.close_device = vfio_pci_core_close_device,
> +	.ioctl = vfio_pci_core_ioctl,
> +	.device_feature = vfio_pci_core_ioctl_feature,
> +	.read = vfio_pci_core_read,
> +	.write = vfio_pci_core_write,
> +	.mmap = vfio_pci_core_mmap,
> +	.request = vfio_pci_core_request,
> +	.match = vfio_pci_core_match,
> +	.match_token_uuid = vfio_pci_core_match_token_uuid,
> +	.bind_iommufd = vfio_iommufd_physical_bind,
> +	.unbind_iommufd = vfio_iommufd_physical_unbind,
> +	.attach_ioas = vfio_iommufd_physical_attach_ioas,
> +	.detach_ioas = vfio_iommufd_physical_detach_ioas,
> +};
> +
> +static int xe_vfio_pci_probe(struct pci_dev *pdev, const struct
> +pci_device_id *id) {
> +	struct xe_vfio_pci_core_device *xe_vdev;
> +	int ret;
> +
> +	xe_vdev = vfio_alloc_device(xe_vfio_pci_core_device,
> core_device.vdev, &pdev->dev,
> +				    &xe_vfio_pci_ops);
> +	if (IS_ERR(xe_vdev))
> +		return PTR_ERR(xe_vdev);
> +
> +	dev_set_drvdata(&pdev->dev, &xe_vdev->core_device);
> +
> +	ret = vfio_pci_core_register_device(&xe_vdev->core_device);
> +	if (ret) {
> +		vfio_put_device(&xe_vdev->core_device.vdev);
> +		return ret;
> +	}
> +
> +	return 0;
> +}
> +
> +static void xe_vfio_pci_remove(struct pci_dev *pdev) {
> +	struct xe_vfio_pci_core_device *xe_vdev = pci_get_drvdata(pdev);
> +
> +	vfio_pci_core_unregister_device(&xe_vdev->core_device);
> +	vfio_put_device(&xe_vdev->core_device.vdev);
> +}
> +
> +static const struct pci_device_id xe_vfio_pci_table[] = {
> +	INTEL_PTL_IDS(INTEL_VGA_VFIO_DEVICE, NULL),
> +	INTEL_WCL_IDS(INTEL_VGA_VFIO_DEVICE, NULL),
> +	INTEL_BMG_IDS(INTEL_VGA_VFIO_DEVICE, NULL),
> +	{}
> +};
> +MODULE_DEVICE_TABLE(pci, xe_vfio_pci_table);
> +
> +static struct pci_driver xe_vfio_pci_driver = {
> +	.name = "xe-vfio-pci",
> +	.id_table = xe_vfio_pci_table,
> +	.probe = xe_vfio_pci_probe,
> +	.remove = xe_vfio_pci_remove,
> +	.err_handler = &xe_vfio_pci_err_handlers,
> +	.driver_managed_dma = true,
> +};
> +module_pci_driver(xe_vfio_pci_driver);
> +
> +MODULE_LICENSE("GPL");
> +MODULE_AUTHOR("Michał Winiarski <michal.winiarski@intel.com>");
> +MODULE_DESCRIPTION("VFIO PCI driver with migration support for Intel
> +Graphics");
> --
> 2.50.1

Re: [PATCH v3 28/28] vfio/xe: Add device specific vfio_pci driver variant for Intel graphics

[Winiarski, Michal]

On Fri, Nov 07, 2025 at 10:38:05AM +0100, Muqthyar Ahmed, Syed Abdul wrote:
> 
> 
> > -----Original Message-----
> > From: Intel-xe <intel-xe-bounces@lists.freedesktop.org> On Behalf Of Michał
> > Winiarski
> > Sent: Friday, October 31, 2025 2:02 AM
> > To: Alex Williamson <alex@shazbot.org>; De Marchi, Lucas
> > <lucas.demarchi@intel.com>; Thomas Hellström
> > <thomas.hellstrom@linux.intel.com>; Vivi, Rodrigo <rodrigo.vivi@intel.com>;
> > Jason Gunthorpe <jgg@ziepe.ca>; Yishai Hadas <yishaih@nvidia.com>; Tian,
> > Kevin <kevin.tian@intel.com>; Shameer Kolothum
> > <skolothumtho@nvidia.com>; intel-xe@lists.freedesktop.org; linux-
> > kernel@vger.kernel.org; kvm@vger.kernel.org; Brost, Matthew
> > <matthew.brost@intel.com>; Wajdeczko, Michal
> > <Michal.Wajdeczko@intel.com>
> > Cc: dri-devel@lists.freedesktop.org; Jani Nikula <jani.nikula@linux.intel.com>;
> > Joonas Lahtinen <joonas.lahtinen@linux.intel.com>; Tvrtko Ursulin
> > <tursulin@ursulin.net>; David Airlie <airlied@gmail.com>; Simona Vetter
> > <simona@ffwll.ch>; Laguna, Lukasz <lukasz.laguna@intel.com>; Christoph
> > Hellwig <hch@infradead.org>; Winiarski, Michal
> > <michal.winiarski@intel.com>
> > Subject: [PATCH v3 28/28] vfio/xe: Add device specific vfio_pci driver variant
> > for Intel graphics
> > 
> > In addition to generic VFIO PCI functionality, the driver implements VFIO
> > migration uAPI, allowing userspace to enable migration for Intel Graphics SR-
> > IOV Virtual Functions.
> > The driver binds to VF device, and uses API exposed by Xe driver bound to PF
> > device to control VF device state and transfer the migration data.
> > 
> > Signed-off-by: Michał Winiarski <michal.winiarski@intel.com>
> > ---
> >  MAINTAINERS                  |   7 +
> >  drivers/vfio/pci/Kconfig     |   2 +
> >  drivers/vfio/pci/Makefile    |   2 +
> >  drivers/vfio/pci/xe/Kconfig  |  12 +
> >  drivers/vfio/pci/xe/Makefile |   3 +
> >  drivers/vfio/pci/xe/main.c   | 552 +++++++++++++++++++++++++++++++++++
> >  6 files changed, 578 insertions(+)
> >  create mode 100644 drivers/vfio/pci/xe/Kconfig  create mode 100644
> > drivers/vfio/pci/xe/Makefile  create mode 100644 drivers/vfio/pci/xe/main.c
> > 
> > diff --git a/MAINTAINERS b/MAINTAINERS
> > index b890ff265f03f..d73348c5f3f3e 100644
> > --- a/MAINTAINERS
> > +++ b/MAINTAINERS
> > @@ -27008,6 +27008,13 @@ L:	virtualization@lists.linux.dev
> >  S:	Maintained
> >  F:	drivers/vfio/pci/virtio
> > 
> > +VFIO XE PCI DRIVER
> > +M:	Michał Winiarski <michal.winiarski@intel.com>
> > +L:	kvm@vger.kernel.org
> > +L:	intel-xe@lists.freedesktop.org
> > +S:	Supported
> > +F:	drivers/vfio/pci/xe
> > +
> >  VGA_SWITCHEROO
> >  R:	Lukas Wunner <lukas@wunner.de>
> >  S:	Maintained
> > diff --git a/drivers/vfio/pci/Kconfig b/drivers/vfio/pci/Kconfig index
> > 2b0172f546652..c100f0ab87f2d 100644
> > --- a/drivers/vfio/pci/Kconfig
> > +++ b/drivers/vfio/pci/Kconfig
> > @@ -67,4 +67,6 @@ source "drivers/vfio/pci/nvgrace-gpu/Kconfig"
> > 
> >  source "drivers/vfio/pci/qat/Kconfig"
> > 
> > +source "drivers/vfio/pci/xe/Kconfig"
> > +
> >  endmenu
> > diff --git a/drivers/vfio/pci/Makefile b/drivers/vfio/pci/Makefile index
> > cf00c0a7e55c8..f5d46aa9347b9 100644
> > --- a/drivers/vfio/pci/Makefile
> > +++ b/drivers/vfio/pci/Makefile
> > @@ -19,3 +19,5 @@ obj-$(CONFIG_VIRTIO_VFIO_PCI) += virtio/
> >  obj-$(CONFIG_NVGRACE_GPU_VFIO_PCI) += nvgrace-gpu/
> > 
> >  obj-$(CONFIG_QAT_VFIO_PCI) += qat/
> > +
> > +obj-$(CONFIG_XE_VFIO_PCI) += xe/
> > diff --git a/drivers/vfio/pci/xe/Kconfig b/drivers/vfio/pci/xe/Kconfig new file
> > mode 100644 index 0000000000000..787be88268685
> > --- /dev/null
> > +++ b/drivers/vfio/pci/xe/Kconfig
> > @@ -0,0 +1,12 @@
> > +# SPDX-License-Identifier: GPL-2.0-only config XE_VFIO_PCI
> > +	tristate "VFIO support for Intel Graphics"
> > +	depends on DRM_XE
> > +	select VFIO_PCI_CORE
> Shall we make it default enabled with "default m " so that config gets enabled for first timers and no need to enable explicitly. 

No. The disable-by-default rule is intentional.

VFIO driver variants are generally not part of "Hardware or
infrastructure that everybody expects, such as CONFIG_NET or
CONFIG_BLOCK".
There's nothing about xe-vfio-pci that would make it an exception.

For more information, see:
https://docs.kernel.org/kbuild/kconfig-language.html#menu-attributes

Thanks,
-Michał