From: Matthew Auld <matthew.auld@intel.com>
To: Xin Wang <x.wang@intel.com>, intel-xe@lists.freedesktop.org
Cc: Matt Roper <matthew.d.roper@intel.com>
Subject: Re: [PATCH v3] drm/xe: Allow compressible surfaces to be 1-way coherent
Date: Wed, 7 Jan 2026 09:45:26 +0000 [thread overview]
Message-ID: <38e7e2b5-bf30-4353-b97b-272e6ac3bbc1@intel.com> (raw)
In-Reply-To: <20260106185501.964562-1-x.wang@intel.com>
On 06/01/2026 18:55, Xin Wang wrote:
> Previously, compressible surfaces were required to be non-coherent (allocated
> as WC) because compression and coherency were mutually exclusive. Starting
> with Xe3, hardware supports combining compression with 1-way coherency,
> allowing compressible surfaces to be allocated as WB memory. This provides
> applications with more efficient memory allocation by avoiding WC allocation
> overhead that can cause system stuttering and memory management challenges.
>
> The implementation adds support for compressed+coherent PAT entry for the
> xe3_lpg devices and updates the driver logic to handle the new compression
> capabilities.
>
> v2: (Matthew Auld)
> - Improved error handling with XE_IOCTL_DBG()
> - Enhanced documentation and comments
> - Fixed xe_bo_needs_ccs_pages() outdated compression assumptions
>
> v3:
> - Improve WB compression support detection by checking PAT table instead
> of version check
>
> Bspec: 71582, 59361, 59399
> Cc: Matthew Auld <matthew.auld@intel.com>
> Cc: Matt Roper <matthew.d.roper@intel.com>
> Signed-off-by: Xin Wang <x.wang@intel.com>
> ---
> drivers/gpu/drm/xe/regs/xe_gt_regs.h | 6 ++++
> drivers/gpu/drm/xe/xe_bo.c | 41 ++++++++++++++++++------
> drivers/gpu/drm/xe/xe_gt.c | 32 +++++++++++++++++++
> drivers/gpu/drm/xe/xe_pat.c | 47 ++++++++++++++++++++++++----
> drivers/gpu/drm/xe/xe_vm.c | 13 ++++++++
> 5 files changed, 124 insertions(+), 15 deletions(-)
>
> diff --git a/drivers/gpu/drm/xe/regs/xe_gt_regs.h b/drivers/gpu/drm/xe/regs/xe_gt_regs.h
> index 93643da57428..24fc64fc832e 100644
> --- a/drivers/gpu/drm/xe/regs/xe_gt_regs.h
> +++ b/drivers/gpu/drm/xe/regs/xe_gt_regs.h
> @@ -89,6 +89,7 @@
> #define UNIFIED_COMPRESSION_FORMAT REG_GENMASK(3, 0)
>
> #define XE2_GAMREQSTRM_CTRL XE_REG_MCR(0x4194)
> +#define EN_CMP_1WCOH REG_BIT(15)
> #define CG_DIS_CNTLBUS REG_BIT(6)
>
> #define CCS_AUX_INV XE_REG(0x4208)
> @@ -101,6 +102,11 @@
>
> #define XE2_LMEM_CFG XE_REG(0x48b0)
>
> +#define XE2_GAMWALK_CTRL 0x47e4
> +#define XE2_GAMWALK_CTRL_MEDIA XE_REG(XE2_GAMWALK_CTRL + MEDIA_GT_GSI_OFFSET)
> +#define XE2_GAMWALK_CTRL_3D XE_REG_MCR(XE2_GAMWALK_CTRL)
> +#define EN_CMP_1WCOH_GW REG_BIT(14)
> +
> #define XEHP_FLAT_CCS_BASE_ADDR XE_REG_MCR(0x4910)
> #define XEHP_FLAT_CCS_PTR REG_GENMASK(31, 8)
>
> diff --git a/drivers/gpu/drm/xe/xe_bo.c b/drivers/gpu/drm/xe/xe_bo.c
> index 8b6474cd3eaf..efd199557f67 100644
> --- a/drivers/gpu/drm/xe/xe_bo.c
> +++ b/drivers/gpu/drm/xe/xe_bo.c
> @@ -29,6 +29,7 @@
> #include "xe_gt.h"
> #include "xe_map.h"
> #include "xe_migrate.h"
> +#include "xe_pat.h"
> #include "xe_pm.h"
> #include "xe_preempt_fence.h"
> #include "xe_pxp.h"
> @@ -3517,17 +3518,39 @@ bool xe_bo_needs_ccs_pages(struct xe_bo *bo)
> if (IS_DGFX(xe) && (bo->flags & XE_BO_FLAG_SYSTEM))
> return false;
>
> + /* Check if userspace explicitly requested no compression */
> + if (bo->flags & XE_BO_FLAG_NO_COMPRESSION)
> + return false;
> +
> /*
> - * Compression implies coh_none, therefore we know for sure that WB
> - * memory can't currently use compression, which is likely one of the
> - * common cases.
> - * Additionally, userspace may explicitly request no compression via the
> - * DRM_XE_GEM_CREATE_FLAG_NO_COMPRESSION flag, which should also disable
> - * CCS usage.
> + * For WB (Write-Back) CPU caching mode, check if compression is
> + * supported through any available PAT index. If not, FlatCCS
> + * can't be used.
> */
> - if (bo->cpu_caching == DRM_XE_GEM_CPU_CACHING_WB ||
> - bo->flags & XE_BO_FLAG_NO_COMPRESSION)
> - return false;
> + if (bo->cpu_caching == DRM_XE_GEM_CPU_CACHING_WB) {
> + bool wb_comp_supported = false;
> +
> + /*
> + * Compression for WB caching was introduced in
> + * GRAPHICS_VER 30 (Xe2). Earlier versions do not
> + * support it.
> + */
> + if (GRAPHICS_VER(xe) < 30)
> + return false;
> +
> + for (int i = 0; i < xe->pat.n_entries; i++) {
> + if (!xe->pat.table[i].valid)
> + continue;
> + if (xe_pat_index_get_comp_en(xe, i) &&
> + xe_pat_index_get_coh_mode(xe, i) != XE_COH_NONE) {
> + wb_comp_supported = true;
> + break;
> + }
> + }
> +
> + if (!wb_comp_supported)
> + return false;
> + }
Would it be cleaner to make this a feature flag instead of checking this
every time, if you want to avoid the version check? info.wb_comp_supported?
>
> return true;
> }
> diff --git a/drivers/gpu/drm/xe/xe_gt.c b/drivers/gpu/drm/xe/xe_gt.c
> index 313ce83ab0e5..04dbf995a18b 100644
> --- a/drivers/gpu/drm/xe/xe_gt.c
> +++ b/drivers/gpu/drm/xe/xe_gt.c
> @@ -140,6 +140,36 @@ static void xe_gt_disable_host_l2_vram(struct xe_gt *gt)
> xe_gt_mcr_multicast_write(gt, XE2_GAMREQSTRM_CTRL, reg);
> }
>
> +static void xe_gt_enable_comp_1wcoh(struct xe_gt *gt)
> +{
> + struct xe_device *xe = gt_to_xe(gt);
> + unsigned int fw_ref;
> + u32 reg;
> +
> + if (IS_SRIOV_VF(xe))
> + return;
> +
> + if (GRAPHICS_VER(xe) >= 30 && xe->info.has_flat_ccs) {
> + fw_ref = xe_force_wake_get(gt_to_fw(gt), XE_FW_GT);
> + if (!fw_ref)
> + return;
> +
> + reg = xe_gt_mcr_unicast_read_any(gt, XE2_GAMREQSTRM_CTRL);
> + reg |= EN_CMP_1WCOH;
> + xe_gt_mcr_multicast_write(gt, XE2_GAMREQSTRM_CTRL, reg);
> +
> + if (xe_gt_is_media_type(gt)) {
> + xe_mmio_rmw32(>->mmio, XE2_GAMWALK_CTRL_MEDIA, 0, EN_CMP_1WCOH_GW);
> + } else {
> + reg = xe_gt_mcr_unicast_read_any(gt, XE2_GAMWALK_CTRL_3D);
> + reg |= EN_CMP_1WCOH_GW;
> + xe_gt_mcr_multicast_write(gt, XE2_GAMWALK_CTRL_3D, reg);
> + }
> +
> + xe_force_wake_put(gt_to_fw(gt), fw_ref);
> + }
> +}
> +
> static void gt_reset_worker(struct work_struct *w);
>
> static int emit_job_sync(struct xe_exec_queue *q, struct xe_bb *bb,
> @@ -466,6 +496,7 @@ static int gt_init_with_gt_forcewake(struct xe_gt *gt)
> xe_gt_topology_init(gt);
> xe_gt_mcr_init(gt);
> xe_gt_enable_host_l2_vram(gt);
> + xe_gt_enable_comp_1wcoh(gt);
>
> if (xe_gt_is_main_type(gt)) {
> err = xe_ggtt_init(gt_to_tile(gt)->mem.ggtt);
> @@ -745,6 +776,7 @@ static int do_gt_restart(struct xe_gt *gt)
> xe_pat_init(gt);
>
> xe_gt_enable_host_l2_vram(gt);
> + xe_gt_enable_comp_1wcoh(gt);
>
> xe_gt_mcr_set_implicit_defaults(gt);
> xe_reg_sr_apply_mmio(>->reg_sr, gt);
> diff --git a/drivers/gpu/drm/xe/xe_pat.c b/drivers/gpu/drm/xe/xe_pat.c
> index 2c3375e0250b..440a9013dc04 100644
> --- a/drivers/gpu/drm/xe/xe_pat.c
> +++ b/drivers/gpu/drm/xe/xe_pat.c
> @@ -132,9 +132,10 @@ static const struct xe_pat_table_entry xelpg_pat_table[] = {
> * in the table.
> *
> * Note: There is an implicit assumption in the driver that compression and
> - * coh_1way+ are mutually exclusive. If this is ever not true then userptr
> - * and imported dma-buf from external device will have uncleared ccs state. See
> - * also xe_bo_needs_ccs_pages().
> + * coh_1way+ are mutually exclusive for platforms prior to Xe3. Starting
> + * with Xe3, compression can be combined with coherency. If using compression
> + * with coherency, userptr and imported dma-buf from external device will
> + * have uncleared ccs state. See also xe_bo_needs_ccs_pages().
> */
> #define XE2_PAT(no_promote, comp_en, l3clos, l3_policy, l4_policy, __coh_mode) \
> { \
> @@ -144,8 +145,7 @@ static const struct xe_pat_table_entry xelpg_pat_table[] = {
> REG_FIELD_PREP(XE2_L3_POLICY, l3_policy) | \
> REG_FIELD_PREP(XE2_L4_POLICY, l4_policy) | \
> REG_FIELD_PREP(XE2_COH_MODE, __coh_mode), \
> - .coh_mode = (BUILD_BUG_ON_ZERO(__coh_mode && comp_en) || __coh_mode) ? \
> - XE_COH_AT_LEAST_1WAY : XE_COH_NONE, \
> + .coh_mode = __coh_mode ? XE_COH_AT_LEAST_1WAY : XE_COH_NONE, \
> .valid = 1 \
> }
>
> @@ -181,6 +181,38 @@ static const struct xe_pat_table_entry xe2_pat_table[] = {
> [31] = XE2_PAT( 0, 0, 3, 0, 3, 3 ),
> };
>
> +static const struct xe_pat_table_entry xe3_lpg_pat_table[] = {
> + [ 0] = XE2_PAT( 0, 0, 0, 0, 3, 0 ),
> + [ 1] = XE2_PAT( 0, 0, 0, 0, 3, 2 ),
> + [ 2] = XE2_PAT( 0, 0, 0, 0, 3, 3 ),
> + [ 3] = XE2_PAT( 0, 0, 0, 3, 3, 0 ),
> + [ 4] = XE2_PAT( 0, 0, 0, 3, 0, 2 ),
> + [ 5] = XE2_PAT( 0, 0, 0, 3, 3, 2 ),
> + [ 6] = XE2_PAT( 1, 0, 0, 1, 3, 0 ),
> + [ 7] = XE2_PAT( 0, 0, 0, 3, 0, 3 ),
> + [ 8] = XE2_PAT( 0, 0, 0, 3, 0, 0 ),
> + [ 9] = XE2_PAT( 0, 1, 0, 0, 3, 0 ),
> + [10] = XE2_PAT( 0, 1, 0, 3, 0, 0 ),
> + [11] = XE2_PAT( 1, 1, 0, 1, 3, 0 ),
> + [12] = XE2_PAT( 0, 1, 0, 3, 3, 0 ),
> + [13] = XE2_PAT( 0, 0, 0, 0, 0, 0 ),
> + [14] = XE2_PAT( 0, 1, 0, 0, 0, 0 ),
> + [15] = XE2_PAT( 1, 1, 0, 1, 1, 0 ),
> + [16] = XE2_PAT( 0, 1, 0, 0, 3, 2 ),
> + /* 17..19 are reserved; leave set to all 0's */
> + [20] = XE2_PAT( 0, 0, 1, 0, 3, 0 ),
> + [21] = XE2_PAT( 0, 1, 1, 0, 3, 0 ),
> + [22] = XE2_PAT( 0, 0, 1, 0, 3, 2 ),
> + [23] = XE2_PAT( 0, 0, 1, 0, 3, 3 ),
> + [24] = XE2_PAT( 0, 0, 2, 0, 3, 0 ),
> + [25] = XE2_PAT( 0, 1, 2, 0, 3, 0 ),
> + [26] = XE2_PAT( 0, 0, 2, 0, 3, 2 ),
> + [27] = XE2_PAT( 0, 0, 2, 0, 3, 3 ),
> + [28] = XE2_PAT( 0, 0, 3, 0, 3, 0 ),
> + [29] = XE2_PAT( 0, 1, 3, 0, 3, 0 ),
> + [30] = XE2_PAT( 0, 0, 3, 0, 3, 2 ),
> + [31] = XE2_PAT( 0, 0, 3, 0, 3, 3 ),
> +};
> /* Special PAT values programmed outside the main table */
> static const struct xe_pat_table_entry xe2_pat_ats = XE2_PAT( 0, 0, 0, 0, 3, 3 );
> static const struct xe_pat_table_entry xe2_pat_pta = XE2_PAT( 0, 0, 0, 0, 3, 0 );
> @@ -501,7 +533,10 @@ void xe_pat_init_early(struct xe_device *xe)
> xe->pat.idx[XE_CACHE_WB] = 2;
> } else if (GRAPHICS_VER(xe) == 30 || GRAPHICS_VER(xe) == 20) {
> xe->pat.ops = &xe2_pat_ops;
> - xe->pat.table = xe2_pat_table;
> + if (GRAPHICS_VER(xe) == 30)
> + xe->pat.table = xe3_lpg_pat_table;
> + else
> + xe->pat.table = xe2_pat_table;
> xe->pat.pat_ats = &xe2_pat_ats;
> if (IS_DGFX(xe))
> xe->pat.pat_pta = &xe2_pat_pta;
> diff --git a/drivers/gpu/drm/xe/xe_vm.c b/drivers/gpu/drm/xe/xe_vm.c
> index a07d8b53de66..481ee7763b09 100644
> --- a/drivers/gpu/drm/xe/xe_vm.c
> +++ b/drivers/gpu/drm/xe/xe_vm.c
> @@ -3405,6 +3405,7 @@ static int vm_bind_ioctl_check_args(struct xe_device *xe, struct xe_vm *vm,
> DRM_XE_VM_BIND_FLAG_CPU_ADDR_MIRROR;
> u16 pat_index = (*bind_ops)[i].pat_index;
> u16 coh_mode;
> + bool comp_en;
>
> if (XE_IOCTL_DBG(xe, is_cpu_addr_mirror &&
> (!xe_vm_in_fault_mode(vm) ||
> @@ -3421,6 +3422,7 @@ static int vm_bind_ioctl_check_args(struct xe_device *xe, struct xe_vm *vm,
> pat_index = array_index_nospec(pat_index, xe->pat.n_entries);
> (*bind_ops)[i].pat_index = pat_index;
> coh_mode = xe_pat_index_get_coh_mode(xe, pat_index);
> + comp_en = xe_pat_index_get_comp_en(xe, pat_index);
> if (XE_IOCTL_DBG(xe, !coh_mode)) { /* hw reserved */
> err = -EINVAL;
> goto free_bind_ops;
> @@ -3451,6 +3453,8 @@ static int vm_bind_ioctl_check_args(struct xe_device *xe, struct xe_vm *vm,
> op == DRM_XE_VM_BIND_OP_MAP_USERPTR) ||
> XE_IOCTL_DBG(xe, coh_mode == XE_COH_NONE &&
> op == DRM_XE_VM_BIND_OP_MAP_USERPTR) ||
> + XE_IOCTL_DBG(xe, comp_en &&
> + op == DRM_XE_VM_BIND_OP_MAP_USERPTR) ||
> XE_IOCTL_DBG(xe, op == DRM_XE_VM_BIND_OP_MAP_USERPTR &&
> !IS_ENABLED(CONFIG_DRM_GPUSVM)) ||
> XE_IOCTL_DBG(xe, obj &&
> @@ -3529,6 +3533,7 @@ static int xe_vm_bind_ioctl_validate_bo(struct xe_device *xe, struct xe_bo *bo,
> u16 pat_index, u32 op, u32 bind_flags)
> {
> u16 coh_mode;
> + bool comp_en;
>
> if (XE_IOCTL_DBG(xe, (bo->flags & XE_BO_FLAG_NO_COMPRESSION) &&
> xe_pat_index_get_comp_en(xe, pat_index)))
> @@ -3574,6 +3579,14 @@ static int xe_vm_bind_ioctl_validate_bo(struct xe_device *xe, struct xe_bo *bo,
> return -EINVAL;
> }
>
> + /*
> + * Ensures that imported buffer objects (dma-bufs) are not mapped
> + * with a PAT index that enables compression.
> + */
> + comp_en = xe_pat_index_get_comp_en(xe, pat_index);
> + if (XE_IOCTL_DBG(xe, bo->ttm.base.import_attach && comp_en))
> + return -EINVAL;
> +
> /* If a BO is protected it can only be mapped if the key is still valid */
> if ((bind_flags & DRM_XE_VM_BIND_FLAG_CHECK_PXP) && xe_bo_is_protected(bo) &&
> op != DRM_XE_VM_BIND_OP_UNMAP && op != DRM_XE_VM_BIND_OP_UNMAP_ALL)
next prev parent reply other threads:[~2026-01-07 9:45 UTC|newest]
Thread overview: 8+ messages / expand[flat|nested] mbox.gz Atom feed top
2026-01-06 18:55 [PATCH v3] drm/xe: Allow compressible surfaces to be 1-way coherent Xin Wang
2026-01-06 19:01 ` ✗ CI.checkpatch: warning for drm/xe: Allow compressible surfaces to be 1-way coherent (rev4) Patchwork
2026-01-06 19:02 ` ✓ CI.KUnit: success " Patchwork
2026-01-06 19:45 ` ✓ Xe.CI.BAT: " Patchwork
2026-01-06 21:22 ` ✗ Xe.CI.Full: failure " Patchwork
2026-01-07 9:45 ` Matthew Auld [this message]
2026-01-08 1:20 ` [PATCH v3] drm/xe: Allow compressible surfaces to be 1-way coherent Wang, X
-- strict thread matches above, loose matches on Subject: below --
2025-11-04 19:17 [PATCH] " Xin Wang
2026-01-06 18:40 ` [PATCH v3] " Xin Wang
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=38e7e2b5-bf30-4353-b97b-272e6ac3bbc1@intel.com \
--to=matthew.auld@intel.com \
--cc=intel-xe@lists.freedesktop.org \
--cc=matthew.d.roper@intel.com \
--cc=x.wang@intel.com \
/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