From: Eric Curtin <ericcurtin17@gmail.com>
To: linux-hyperv@vger.kernel.org
Cc: linux-kernel@vger.kernel.org, iourit@linux.microsoft.com,
wei.liu@kernel.org, decui@microsoft.com, haiyangz@microsoft.com
Subject: [PATCH 28/55] drivers: hv: dxgkrnl: Add support to map guest pages by host
Date: Thu, 19 Mar 2026 20:24:42 +0000 [thread overview]
Message-ID: <20260319202509.63802-29-eric.curtin@docker.com> (raw)
In-Reply-To: <20260319202509.63802-1-eric.curtin@docker.com>
From: Iouri Tarassov <iourit@linux.microsoft.com>
Implement support for mapping guest memory pages by the host.
This removes hyper-v limitations of using GPADL (guest physical
address list).
Dxgkrnl uses hyper-v GPADLs to share guest system memory with the
host. This method has limitations:
- a single GPADL can represent only ~32MB of memory
- there is a limit of how much memory the total size of GPADLs
in a VM can represent.
To avoid these limitations the host implemented mapping guest memory
pages. Presence of this support is determined by reading PCI config
space. When the support is enabled, dxgkrnl does not use GPADLs and
instead uses the following code flow:
- memory pages of an existing system memory buffer are pinned
- PFNs of the pages are sent to the host via a VM bus message
- the host maps the PFNs to get access to the memory
Signed-off-by: Iouri Tarassov <iourit@linux.microsoft.com>
[kms: forward port to 6.6 from 6.1. No code changes made.]
Signed-off-by: Kelsey Steele <kelseysteele@microsoft.com>
---
drivers/hv/dxgkrnl/Makefile | 2 +-
drivers/hv/dxgkrnl/dxgkrnl.h | 1 +
drivers/hv/dxgkrnl/dxgmodule.c | 33 +++++++++-
drivers/hv/dxgkrnl/dxgvmbus.c | 117 ++++++++++++++++++++++++---------
drivers/hv/dxgkrnl/dxgvmbus.h | 10 +++
drivers/hv/dxgkrnl/misc.c | 1 +
6 files changed, 129 insertions(+), 35 deletions(-)
diff --git a/drivers/hv/dxgkrnl/Makefile b/drivers/hv/dxgkrnl/Makefile
index 9d821e83448a..fc85a47a6ad5 100644
--- a/drivers/hv/dxgkrnl/Makefile
+++ b/drivers/hv/dxgkrnl/Makefile
@@ -2,4 +2,4 @@
# Makefile for the hyper-v compute device driver (dxgkrnl).
obj-$(CONFIG_DXGKRNL) += dxgkrnl.o
-dxgkrnl-y := dxgmodule.o hmgr.o misc.o dxgadapter.o ioctl.o dxgvmbus.o dxgprocess.o
+dxgkrnl-y := dxgmodule.o hmgr.o misc.o dxgadapter.o ioctl.o dxgvmbus.o dxgprocess.o
diff --git a/drivers/hv/dxgkrnl/dxgkrnl.h b/drivers/hv/dxgkrnl/dxgkrnl.h
index 93bc9b41aa41..091dbe999d33 100644
--- a/drivers/hv/dxgkrnl/dxgkrnl.h
+++ b/drivers/hv/dxgkrnl/dxgkrnl.h
@@ -316,6 +316,7 @@ struct dxgglobal {
bool misc_registered;
bool pci_registered;
bool vmbus_registered;
+ bool map_guest_pages_enabled;
};
static inline struct dxgglobal *dxggbl(void)
diff --git a/drivers/hv/dxgkrnl/dxgmodule.c b/drivers/hv/dxgkrnl/dxgmodule.c
index 5c364a46b65f..b1b612b90fc1 100644
--- a/drivers/hv/dxgkrnl/dxgmodule.c
+++ b/drivers/hv/dxgkrnl/dxgmodule.c
@@ -147,7 +147,7 @@ void dxgglobal_remove_host_event(struct dxghostevent *event)
void signal_host_cpu_event(struct dxghostevent *eventhdr)
{
- struct dxghosteventcpu *event = (struct dxghosteventcpu *)eventhdr;
+ struct dxghosteventcpu *event = (struct dxghosteventcpu *)eventhdr;
if (event->remove_from_list ||
event->destroy_after_signal) {
@@ -426,7 +426,11 @@ const struct file_operations dxgk_fops = {
#define DXGK_VMBUS_VGPU_LUID_OFFSET (DXGK_VMBUS_VERSION_OFFSET + \
sizeof(u32))
-/* The guest writes its capabilities to this address */
+/* The host caps (dxgk_vmbus_hostcaps) */
+#define DXGK_VMBUS_HOSTCAPS_OFFSET (DXGK_VMBUS_VGPU_LUID_OFFSET + \
+ sizeof(struct winluid))
+
+/* The guest writes its capavilities to this adderss */
#define DXGK_VMBUS_GUESTCAPS_OFFSET (DXGK_VMBUS_VERSION_OFFSET + \
sizeof(u32))
@@ -441,6 +445,23 @@ struct dxgk_vmbus_guestcaps {
};
};
+/*
+ * The structure defines features, supported by the host.
+ *
+ * map_guest_memory
+ * Host can map guest memory pages, so the guest can avoid using GPADLs
+ * to represent existing system memory allocations.
+ */
+struct dxgk_vmbus_hostcaps {
+ union {
+ struct {
+ u32 map_guest_memory : 1;
+ u32 reserved : 31;
+ };
+ u32 host_caps;
+ };
+};
+
/*
* A helper function to read PCI config space.
*/
@@ -475,6 +496,7 @@ static int dxg_pci_probe_device(struct pci_dev *dev,
struct winluid vgpu_luid = {};
struct dxgk_vmbus_guestcaps guest_caps = {.wsl2 = 1};
struct dxgglobal *dxgglobal = dxggbl();
+ struct dxgk_vmbus_hostcaps host_caps = {};
mutex_lock(&dxgglobal->device_mutex);
@@ -503,6 +525,13 @@ static int dxg_pci_probe_device(struct pci_dev *dev,
if (ret)
goto cleanup;
+ ret = pci_read_config_dword(dev, DXGK_VMBUS_HOSTCAPS_OFFSET,
+ &host_caps.host_caps);
+ if (ret == 0) {
+ if (host_caps.map_guest_memory)
+ dxgglobal->map_guest_pages_enabled = true;
+ }
+
if (dxgglobal->vmbus_ver > DXGK_VMBUS_INTERFACE_VERSION)
dxgglobal->vmbus_ver = DXGK_VMBUS_INTERFACE_VERSION;
}
diff --git a/drivers/hv/dxgkrnl/dxgvmbus.c b/drivers/hv/dxgkrnl/dxgvmbus.c
index 425a1ab87bd6..4d7807909284 100644
--- a/drivers/hv/dxgkrnl/dxgvmbus.c
+++ b/drivers/hv/dxgkrnl/dxgvmbus.c
@@ -1383,15 +1383,19 @@ int create_existing_sysmem(struct dxgdevice *device,
void *kmem = NULL;
int ret = 0;
struct dxgkvmb_command_setexistingsysmemstore *set_store_command;
+ struct dxgkvmb_command_setexistingsysmempages *set_pages_command;
u64 alloc_size = host_alloc->allocation_size;
u32 npages = alloc_size >> PAGE_SHIFT;
struct dxgvmbusmsg msg = {.hdr = NULL};
-
- ret = init_message(&msg, device->adapter, device->process,
- sizeof(*set_store_command));
- if (ret)
- goto cleanup;
- set_store_command = (void *)msg.msg;
+ const u32 max_pfns_in_message =
+ (DXG_MAX_VM_BUS_PACKET_SIZE - sizeof(*set_pages_command) -
+ PAGE_SIZE) / sizeof(__u64);
+ u32 alloc_offset_in_pages = 0;
+ struct page **page_in;
+ u64 *pfn;
+ u32 pages_to_send;
+ u32 i;
+ struct dxgglobal *dxgglobal = dxggbl();
/*
* Create a guest physical address list and set it as the allocation
@@ -1402,6 +1406,7 @@ int create_existing_sysmem(struct dxgdevice *device,
DXG_TRACE("Alloc size: %lld", alloc_size);
dxgalloc->cpu_address = (void *)sysmem;
+
dxgalloc->pages = vzalloc(npages * sizeof(void *));
if (dxgalloc->pages == NULL) {
DXG_ERR("failed to allocate pages");
@@ -1419,39 +1424,87 @@ int create_existing_sysmem(struct dxgdevice *device,
ret = -ENOMEM;
goto cleanup;
}
- kmem = vmap(dxgalloc->pages, npages, VM_MAP, PAGE_KERNEL);
- if (kmem == NULL) {
- DXG_ERR("vmap failed");
- ret = -ENOMEM;
- goto cleanup;
- }
- ret1 = vmbus_establish_gpadl(dxgglobal_get_vmbus(), kmem,
- alloc_size, &dxgalloc->gpadl);
- if (ret1) {
- DXG_ERR("establish_gpadl failed: %d", ret1);
- ret = -ENOMEM;
- goto cleanup;
- }
+ if (!dxgglobal->map_guest_pages_enabled) {
+ ret = init_message(&msg, device->adapter, device->process,
+ sizeof(*set_store_command));
+ if (ret)
+ goto cleanup;
+ set_store_command = (void *)msg.msg;
+
+ kmem = vmap(dxgalloc->pages, npages, VM_MAP, PAGE_KERNEL);
+ if (kmem == NULL) {
+ DXG_ERR("vmap failed");
+ ret = -ENOMEM;
+ goto cleanup;
+ }
+ ret1 = vmbus_establish_gpadl(dxgglobal_get_vmbus(), kmem,
+ alloc_size, &dxgalloc->gpadl);
+ if (ret1) {
+ DXG_ERR("establish_gpadl failed: %d", ret1);
+ ret = -ENOMEM;
+ goto cleanup;
+ }
#ifdef _MAIN_KERNEL_
- DXG_TRACE("New gpadl %d", dxgalloc->gpadl.gpadl_handle);
+ DXG_TRACE("New gpadl %d", dxgalloc->gpadl.gpadl_handle);
#else
- DXG_TRACE("New gpadl %d", dxgalloc->gpadl);
+ DXG_TRACE("New gpadl %d", dxgalloc->gpadl);
#endif
- command_vgpu_to_host_init2(&set_store_command->hdr,
- DXGK_VMBCOMMAND_SETEXISTINGSYSMEMSTORE,
- device->process->host_handle);
- set_store_command->device = device->handle;
- set_store_command->device = device->handle;
- set_store_command->allocation = host_alloc->allocation;
+ command_vgpu_to_host_init2(&set_store_command->hdr,
+ DXGK_VMBCOMMAND_SETEXISTINGSYSMEMSTORE,
+ device->process->host_handle);
+ set_store_command->device = device->handle;
+ set_store_command->allocation = host_alloc->allocation;
#ifdef _MAIN_KERNEL_
- set_store_command->gpadl = dxgalloc->gpadl.gpadl_handle;
+ set_store_command->gpadl = dxgalloc->gpadl.gpadl_handle;
#else
- set_store_command->gpadl = dxgalloc->gpadl;
+ set_store_command->gpadl = dxgalloc->gpadl;
#endif
- ret = dxgvmb_send_sync_msg_ntstatus(msg.channel, msg.hdr, msg.size);
- if (ret < 0)
- DXG_ERR("failed to set existing store: %x", ret);
+ ret = dxgvmb_send_sync_msg_ntstatus(msg.channel, msg.hdr,
+ msg.size);
+ if (ret < 0)
+ DXG_ERR("failed set existing store: %x", ret);
+ } else {
+ /*
+ * Send the list of the allocation PFNs to the host. The host
+ * will map the pages for GPU access.
+ */
+
+ ret = init_message(&msg, device->adapter, device->process,
+ sizeof(*set_pages_command) +
+ max_pfns_in_message * sizeof(u64));
+ if (ret)
+ goto cleanup;
+ set_pages_command = (void *)msg.msg;
+ command_vgpu_to_host_init2(&set_pages_command->hdr,
+ DXGK_VMBCOMMAND_SETEXISTINGSYSMEMPAGES,
+ device->process->host_handle);
+ set_pages_command->device = device->handle;
+ set_pages_command->allocation = host_alloc->allocation;
+
+ page_in = dxgalloc->pages;
+ while (alloc_offset_in_pages < npages) {
+ pfn = (u64 *)((char *)msg.msg +
+ sizeof(*set_pages_command));
+ pages_to_send = min(npages - alloc_offset_in_pages,
+ max_pfns_in_message);
+ set_pages_command->num_pages = pages_to_send;
+ set_pages_command->alloc_offset_in_pages =
+ alloc_offset_in_pages;
+
+ for (i = 0; i < pages_to_send; i++)
+ *pfn++ = page_to_pfn(*page_in++);
+
+ ret = dxgvmb_send_sync_msg_ntstatus(msg.channel,
+ msg.hdr,
+ msg.size);
+ if (ret < 0) {
+ DXG_ERR("failed set existing pages: %x", ret);
+ break;
+ }
+ alloc_offset_in_pages += pages_to_send;
+ }
+ }
cleanup:
if (kmem)
diff --git a/drivers/hv/dxgkrnl/dxgvmbus.h b/drivers/hv/dxgkrnl/dxgvmbus.h
index 88967ff6a505..b4a98f7c2522 100644
--- a/drivers/hv/dxgkrnl/dxgvmbus.h
+++ b/drivers/hv/dxgkrnl/dxgvmbus.h
@@ -234,6 +234,16 @@ struct dxgkvmb_command_setexistingsysmemstore {
u32 gpadl;
};
+/* Returns ntstatus */
+struct dxgkvmb_command_setexistingsysmempages {
+ struct dxgkvmb_command_vgpu_to_host hdr;
+ struct d3dkmthandle device;
+ struct d3dkmthandle allocation;
+ u32 num_pages;
+ u32 alloc_offset_in_pages;
+ /* u64 pfn_array[num_pages] */
+};
+
struct dxgkvmb_command_createprocess {
struct dxgkvmb_command_vm_to_host hdr;
void *process;
diff --git a/drivers/hv/dxgkrnl/misc.c b/drivers/hv/dxgkrnl/misc.c
index cb1e0635bebc..4a1309d80ee5 100644
--- a/drivers/hv/dxgkrnl/misc.c
+++ b/drivers/hv/dxgkrnl/misc.c
@@ -35,3 +35,4 @@ u16 *wcsncpy(u16 *dest, const u16 *src, size_t n)
dest[i - 1] = 0;
return dest;
}
+
next prev parent reply other threads:[~2026-03-19 20:25 UTC|newest]
Thread overview: 56+ messages / expand[flat|nested] mbox.gz Atom feed top
2026-03-19 20:24 [PATCH v4 00/55] drivers: hv: dxgkrnl: Driver for Hyper-V virtual compute device Eric Curtin
2026-03-19 20:24 ` [PATCH 01/55] drivers: hv: dxgkrnl: Driver initialization and loading Eric Curtin
2026-03-19 20:24 ` [PATCH 02/55] drivers: hv: dxgkrnl: Add VMBus message support, initialize VMBus channels Eric Curtin
2026-03-19 20:24 ` [PATCH 03/55] drivers: hv: dxgkrnl: Creation of dxgadapter object Eric Curtin
2026-03-19 20:24 ` [PATCH 04/55] drivers: hv: dxgkrnl: Opening of /dev/dxg device and dxgprocess creation Eric Curtin
2026-03-19 20:24 ` [PATCH 05/55] drivers: hv: dxgkrnl: Enumerate and open dxgadapter objects Eric Curtin
2026-03-19 20:24 ` [PATCH 06/55] drivers: hv: dxgkrnl: Creation of dxgdevice objects Eric Curtin
2026-03-19 20:24 ` [PATCH 07/55] drivers: hv: dxgkrnl: Creation of dxgcontext objects Eric Curtin
2026-03-19 20:24 ` [PATCH 08/55] drivers: hv: dxgkrnl: Creation of compute device allocations and resources Eric Curtin
2026-03-19 20:24 ` [PATCH 09/55] drivers: hv: dxgkrnl: Creation of compute device sync objects Eric Curtin
2026-03-19 20:24 ` [PATCH 10/55] drivers: hv: dxgkrnl: Operations using " Eric Curtin
2026-03-19 20:24 ` [PATCH 11/55] drivers: hv: dxgkrnl: Sharing of dxgresource objects Eric Curtin
2026-03-19 20:24 ` [PATCH 12/55] drivers: hv: dxgkrnl: Sharing of sync objects Eric Curtin
2026-03-19 20:24 ` [PATCH 13/55] drivers: hv: dxgkrnl: Creation of paging queue objects Eric Curtin
2026-03-19 20:24 ` [PATCH 14/55] drivers: hv: dxgkrnl: Submit execution commands to the compute device Eric Curtin
2026-03-19 20:24 ` [PATCH 15/55] drivers: hv: dxgkrnl: Share objects with the host Eric Curtin
2026-03-19 20:24 ` [PATCH 16/55] drivers: hv: dxgkrnl: Query the dxgdevice state Eric Curtin
2026-03-19 20:24 ` [PATCH 17/55] drivers: hv: dxgkrnl: Map(unmap) CPU address to device allocation Eric Curtin
2026-03-19 20:24 ` [PATCH 18/55] drivers: hv: dxgkrnl: Manage device allocation properties Eric Curtin
2026-03-19 20:24 ` [PATCH 19/55] drivers: hv: dxgkrnl: Flush heap transitions Eric Curtin
2026-03-19 20:24 ` [PATCH 20/55] drivers: hv: dxgkrnl: Query video memory information Eric Curtin
2026-03-19 20:24 ` [PATCH 21/55] drivers: hv: dxgkrnl: The escape ioctl Eric Curtin
2026-03-19 20:24 ` [PATCH 22/55] drivers: hv: dxgkrnl: Ioctl to put device to error state Eric Curtin
2026-03-19 20:24 ` [PATCH 23/55] drivers: hv: dxgkrnl: Ioctls to query statistics and clock calibration Eric Curtin
2026-03-19 20:24 ` [PATCH 24/55] drivers: hv: dxgkrnl: Offer and reclaim allocations Eric Curtin
2026-03-19 20:24 ` [PATCH 25/55] drivers: hv: dxgkrnl: Ioctls to manage scheduling priority Eric Curtin
2026-03-19 20:24 ` [PATCH 26/55] drivers: hv: dxgkrnl: Manage residency of allocations Eric Curtin
2026-03-19 20:24 ` [PATCH 27/55] drivers: hv: dxgkrnl: Manage compute device virtual addresses Eric Curtin
2026-03-19 20:24 ` Eric Curtin [this message]
2026-03-19 20:24 ` [PATCH 29/55] drivers: hv: dxgkrnl: Removed struct vmbus_gpadl, which was defined in the main linux branch Eric Curtin
2026-03-19 20:24 ` [PATCH 30/55] drivers: hv: dxgkrnl: Remove dxgk_init_ioctls Eric Curtin
2026-03-19 20:24 ` [PATCH 31/55] drivers: hv: dxgkrnl: Creation of dxgsyncfile objects Eric Curtin
2026-03-19 20:24 ` [PATCH 32/55] drivers: hv: dxgkrnl: Use tracing instead of dev_dbg Eric Curtin
2026-03-19 20:24 ` [PATCH 33/55] drivers: hv: dxgkrnl: Implement D3DKMTWaitSyncFile Eric Curtin
2026-03-19 20:24 ` [PATCH 34/55] drivers: hv: dxgkrnl: Improve tracing and return values from copy from user Eric Curtin
2026-03-19 20:24 ` [PATCH 35/55] drivers: hv: dxgkrnl: Fix synchronization locks Eric Curtin
2026-03-19 20:24 ` [PATCH 36/55] drivers: hv: dxgkrnl: Close shared file objects in case of a failure Eric Curtin
2026-03-19 20:24 ` [PATCH 37/55] drivers: hv: dxgkrnl: Added missed NULL check for resource object Eric Curtin
2026-03-19 20:24 ` [PATCH 38/55] drivers: hv: dxgkrnl: Fixed dxgkrnl to build for the 6.1 kernel Eric Curtin
2026-03-19 20:24 ` [PATCH 39/55] drivers: hv: dxgkrnl: Added support for compute only adapters Eric Curtin
2026-03-19 20:24 ` [PATCH 40/55] drivers: hv: dxgkrnl: Added implementation for D3DKMTInvalidateCache Eric Curtin
2026-03-19 20:24 ` [PATCH 41/55] drivers: hv: dxgkrnl: Handle process ID in D3DKMTQueryStatistics Eric Curtin
2026-03-19 20:24 ` [PATCH 42/55] drivers: hv: dxgkrnl: Implement the D3DKMTEnumProcesses API Eric Curtin
2026-03-19 20:24 ` [PATCH 43/55] drivers: hv: dxgkrnl: Implement D3DDKMTIsFeatureEnabled API Eric Curtin
2026-03-19 20:24 ` [PATCH 44/55] drivers: hv: dxgkrnl: Implement known escapes Eric Curtin
2026-03-19 20:24 ` [PATCH 45/55] drivers: hv: dxgkrnl: Fixed coding style issues Eric Curtin
2026-03-19 20:25 ` [PATCH 46/55] drivers: hv: dxgkrnl: Fixed the implementation of D3DKMTQueryClockCalibration Eric Curtin
2026-03-19 20:25 ` [PATCH 47/55] drivers: hv: dxgkrnl: Retry sending a VM bus packet when there is no place in the ring buffer Eric Curtin
2026-03-19 20:25 ` [PATCH 48/55] drivers: hv: dxgkrnl: Add support for locking a shared allocation by not the owner Eric Curtin
2026-03-19 20:25 ` [PATCH 49/55] drivers: hv: dxgkrnl: Fix build breaks when switching to 6.6 kernel due to hv_driver remove callback change Eric Curtin
2026-03-19 20:25 ` [PATCH 50/55] drivers: hv: dxgkrnl: Fix build breaks when switching to 6.6 kernel due to removed uuid_le_cmp Eric Curtin
2026-03-19 20:25 ` [PATCH 51/55] drivers: hv: dxgkrnl: Implement D3DKMTEnumProcesses to match the Windows implementation Eric Curtin
2026-03-19 20:25 ` [PATCH 52/55] drivers: hv: dxgkrnl: Use pin_user_pages instead of get_user_pages for DMA accessible memory Eric Curtin
2026-03-19 20:25 ` [PATCH 53/55] drivers: hv: dxgkrnl: Do not print error messages when virtual GPU is not present Eric Curtin
2026-03-19 20:25 ` [PATCH 54/55] drivers: hv: dxgkrnl: Fix crash at hmgrtable_free_handle Eric Curtin
2026-03-19 20:25 ` [PATCH 55/55] drivers: hv: dxgkrnl: Code cleanup for upstream submission Eric Curtin
Reply instructions:
You may reply publicly to this message via plain-text email
using any one of the following methods:
* Save the following mbox file, import it into your mail client,
and reply-to-all from there: mbox
Avoid top-posting and favor interleaved quoting:
https://en.wikipedia.org/wiki/Posting_style#Interleaved_style
* Reply using the --to, --cc, and --in-reply-to
switches of git-send-email(1):
git send-email \
--in-reply-to=20260319202509.63802-29-eric.curtin@docker.com \
--to=ericcurtin17@gmail.com \
--cc=decui@microsoft.com \
--cc=haiyangz@microsoft.com \
--cc=iourit@linux.microsoft.com \
--cc=linux-hyperv@vger.kernel.org \
--cc=linux-kernel@vger.kernel.org \
--cc=wei.liu@kernel.org \
/path/to/YOUR_REPLY
https://kernel.org/pub/software/scm/git/docs/git-send-email.html
* If your mail client supports setting the In-Reply-To header
via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line
before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox