From: Gustavo Sousa <gustavo.sousa@intel.com>
To: Tejas Upadhyay <tejas.upadhyay@intel.com>,
<intel-xe@lists.freedesktop.org>
Cc: <nitin.r.gote@intel.com>, <matthew.d.roper@intel.com>,
Tejas Upadhyay <tejas.upadhyay@intel.com>
Subject: Re: [PATCH V2] drm/xe/nvl-p: Add Wa_14028564128
Date: Fri, 20 Nov 2026 20:58:03 -0300 [thread overview]
Message-ID: <877bi7qcms.fsf@intel.com> (raw)
In-Reply-To: <20261119075242.3292552-2-tejas.upadhyay@intel.com>
Tejas Upadhyay <tejas.upadhyay@intel.com> writes:
> Wa_14028564128 applies to 35.10
>
> V2:
> - Add missing pat indexes to no_promote as asked by WA
>
> Signed-off-by: Tejas Upadhyay <tejas.upadhyay@intel.com>
> ---
> drivers/gpu/drm/xe/xe_pat.c | 39 ++++++++++++++++++++++++++++++
> drivers/gpu/drm/xe/xe_pat.h | 14 +++++++++++
> drivers/gpu/drm/xe/xe_vm.c | 6 +++--
> drivers/gpu/drm/xe/xe_vm_madvise.c | 11 ++++++---
> drivers/gpu/drm/xe/xe_wa_oob.rules | 1 +
> 5 files changed, 66 insertions(+), 5 deletions(-)
>
> diff --git a/drivers/gpu/drm/xe/xe_pat.c b/drivers/gpu/drm/xe/xe_pat.c
> index 5938fef22579..24caed262642 100644
> --- a/drivers/gpu/drm/xe/xe_pat.c
> +++ b/drivers/gpu/drm/xe/xe_pat.c
> @@ -337,6 +337,18 @@ bool xe_pat_wa_14026539277_reserved(struct xe_device *xe, u16 pat_index)
> }
> }
>
> + /*
> + * Wa_14028564128 remaps every PAT index that is statically 2-way
> + * coherent (indices 2, 7, 23, 27 and 31 in this table) to a safe
> + * no-promote/UC/1-way value, so none of them are subject to this
> + * restriction once that WA is enabled.
> + */
> + for_each_gt(gt, xe, id) {
> + if (XE_GT_WA(gt, 14028564128) &&
> + xe_pat_index_get_coh_mode(xe, pat_index) == XE_COH_2WAY)
> + return false;
> + }
> +
> /*
> * Disallow 2-way coherency for any L3 cache policy because
> * XE_L3_POLICY_UC could still be promoted to a cached policy
> @@ -345,6 +357,21 @@ bool xe_pat_wa_14026539277_reserved(struct xe_device *xe, u16 pat_index)
> return has_wa && xe_pat_index_get_coh_mode(xe, pat_index) == XE_COH_2WAY;
> }
>
> +bool xe_pat_wa_14028564128_safe_userptr_index(struct xe_device *xe, u16 pat_index)
> +{
> + struct xe_gt *gt;
> + u8 id;
> +
> + if (pat_index != 2)
> + return false;
> +
> + for_each_gt(gt, xe, id)
> + if (XE_GT_WA(gt, 14028564128))
> + return true;
> +
> + return false;
> +}
Is this change really related to the workaround? If it is just a happy
coincidence that fulfills some requirement associated to userptr usage, I
think we could make it a separate patch and add some rationale in the
commit message.
> +
> static u32 wa_14026539277_fixup_pat_value(struct xe_gt *gt, u32 value)
> {
> if (XE_GT_WA(gt, 14026539277)) {
> @@ -358,6 +385,15 @@ static u32 wa_14026539277_fixup_pat_value(struct xe_gt *gt, u32 value)
> return value;
> }
>
> +static u32 wa_14028564128_fixup_pat_value(struct xe_gt *gt, u32 orig_value, u32 value)
> +{
> + if (XE_GT_WA(gt, 14028564128) &&
> + REG_FIELD_GET(XE2_COH_MODE, orig_value) == XE_COH_2WAY)
> + value |= XE2_NO_PROMOTE;
This is just setting the XE2_NO_PROMOTE flag and is not setting the
other required ones. Is this is implemented this way because of
Wa_14026539277?
Because Wa_14026539277 applies only to A* steppings, we would wind up
with the wrong PAT bits on B* steppings.
Note that Wa_14028564128 must be applied to all steppings of graphics
version 35.10.
I propose we make this function set all the required bits and ensure
that wa_14028564128_fixup_pat_value() is always called before
wa_14026539277_fixup_pat_value() (and with a comment noting the ordering
dependency), because the modifications done by Wa_14028564128 are a
superset of the ones done by Wa_14026539277.
> +
> + return value;
> +}
> +
> static const struct xe_pat_table_entry *gt_pta_entry(struct xe_gt *gt)
> {
> struct xe_device *xe = gt_to_xe(gt);
> @@ -420,6 +456,8 @@ static void program_pat_mcr(struct xe_gt *gt, const struct xe_pat_table_entry ta
> struct xe_reg_mcr reg_mcr = XE_REG_MCR(_PAT_INDEX(i));
> u32 pat = wa_14026539277_fixup_pat_value(gt, table[i].value);
>
> + pat = wa_14028564128_fixup_pat_value(gt, table[i].value, pat);
> +
> xe_gt_mcr_multicast_write(gt, reg_mcr, pat);
> }
>
> @@ -810,6 +848,7 @@ int xe_pat_dump_sw_config(struct xe_gt *gt, struct drm_printer *p)
> u32 pat = xe->pat.table[i].value;
>
> pat = wa_14026539277_fixup_pat_value(gt, pat);
> + pat = wa_14028564128_fixup_pat_value(gt, xe->pat.table[i].value, pat);
>
> if (GRAPHICS_VER(xe) >= 20) {
> bool rsvd = !xe->pat.table[i].valid ||
> diff --git a/drivers/gpu/drm/xe/xe_pat.h b/drivers/gpu/drm/xe/xe_pat.h
> index 2ac6ddc66587..247dc0faa095 100644
> --- a/drivers/gpu/drm/xe/xe_pat.h
> +++ b/drivers/gpu/drm/xe/xe_pat.h
> @@ -105,6 +105,20 @@ u16 xe_pat_index_get_l3_policy(struct xe_device *xe, u16 pat_index);
> */
> bool xe_pat_wa_14026539277_reserved(struct xe_device *xe, u16 pat_index);
>
> +/**
> + * xe_pat_wa_14028564128_safe_userptr_index - Is this PAT index safe for
> + * misaligned host-pointer userptr/SVM binds due to Wa_14028564128?
> + * @xe: xe device
> + * @pat_index: The pat_index to query
> + *
> + * Wa_14028564128 reprograms PAT index 2 to the no-promote/UC/1-way policy
> + * required for misaligned host pointers; once that WA is enabled, index 2
> + * should be accepted anywhere PAT 19 or a 2-way coherent index already is.
> + *
> + * Return: a boolean indicating whether the PAT index is safe to use.
> + */
> +bool xe_pat_wa_14028564128_safe_userptr_index(struct xe_device *xe, u16 pat_index);
> +
> #define xe_cache_pat_idx(xe, cache_mode) ({ \
> const struct xe_device *__xedev = (xe); \
> enum xe_cache_level __mode = (cache_mode); \
> diff --git a/drivers/gpu/drm/xe/xe_vm.c b/drivers/gpu/drm/xe/xe_vm.c
> index dd855876f571..80ccaa44720c 100644
> --- a/drivers/gpu/drm/xe/xe_vm.c
> +++ b/drivers/gpu/drm/xe/xe_vm.c
> @@ -3888,7 +3888,8 @@ static int vm_bind_ioctl_check_args(struct xe_device *xe, struct xe_vm *vm,
> XE_IOCTL_DBG(xe, xe_device_is_l2_flush_optimized(xe) &&
> (op == DRM_XE_VM_BIND_OP_MAP_USERPTR ||
> is_cpu_addr_mirror) &&
> - (pat_index != 19 && coh_mode != XE_COH_2WAY)) ||
> + (pat_index != 19 && coh_mode != XE_COH_2WAY &&
> + !xe_pat_wa_14028564128_safe_userptr_index(xe, pat_index))) ||
> 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 &&
> @@ -4033,7 +4034,8 @@ static int xe_vm_bind_ioctl_validate_bo(struct xe_device *xe, struct xe_bo *bo,
> return -EINVAL;
>
> if (XE_IOCTL_DBG(xe, bo->ttm.base.import_attach && xe_device_is_l2_flush_optimized(xe) &&
> - (pat_index != 19 && coh_mode != XE_COH_2WAY)))
> + (pat_index != 19 && coh_mode != XE_COH_2WAY &&
> + !xe_pat_wa_14028564128_safe_userptr_index(xe, pat_index))))
> return -EINVAL;
>
> /* If a BO is protected it can only be mapped if the key is still valid */
> diff --git a/drivers/gpu/drm/xe/xe_vm_madvise.c b/drivers/gpu/drm/xe/xe_vm_madvise.c
> index 31245c6c96d4..80dedeaae579 100644
> --- a/drivers/gpu/drm/xe/xe_vm_madvise.c
> +++ b/drivers/gpu/drm/xe/xe_vm_madvise.c
> @@ -594,6 +594,7 @@ int xe_vm_madvise_ioctl(struct drm_device *dev, void *data, struct drm_file *fil
> };
> struct xe_madvise_details details;
> u16 pat_index, coh_mode;
> + bool safe_pat;
> struct xe_vm *vm;
> struct drm_exec exec;
> int err, attr_type;
> @@ -653,9 +654,11 @@ int xe_vm_madvise_ioctl(struct drm_device *dev, void *data, struct drm_file *fil
> if (args->type == DRM_XE_MEM_RANGE_ATTR_PAT) {
> pat_index = array_index_nospec(args->pat_index.val, xe->pat.n_entries);
> coh_mode = xe_pat_index_get_coh_mode(xe, pat_index);
> + safe_pat = xe_pat_wa_14028564128_safe_userptr_index(xe, pat_index);
> if (XE_IOCTL_DBG(xe, madvise_range.has_svm_userptr_vmas &&
> xe_device_is_l2_flush_optimized(xe) &&
> - (pat_index != 19 && coh_mode != XE_COH_2WAY))) {
> + (pat_index != 19 && coh_mode != XE_COH_2WAY &&
> + !safe_pat))) {
> err = -EINVAL;
> goto free_vmas;
> }
> @@ -685,10 +688,12 @@ int xe_vm_madvise_ioctl(struct drm_device *dev, void *data, struct drm_file *fil
> continue;
>
> if (args->type == DRM_XE_MEM_RANGE_ATTR_PAT) {
> + bool invalid_pat = pat_index != 19 &&
> + coh_mode != XE_COH_2WAY && !safe_pat;
> +
> if (XE_IOCTL_DBG(xe, bo->ttm.base.import_attach &&
> xe_device_is_l2_flush_optimized(xe) &&
> - (pat_index != 19 &&
> - coh_mode != XE_COH_2WAY))) {
> + invalid_pat)) {
> err = -EINVAL;
> goto err_fini;
> }
> diff --git a/drivers/gpu/drm/xe/xe_wa_oob.rules b/drivers/gpu/drm/xe/xe_wa_oob.rules
> index 03b713f55dfb..6a48b2c716ea 100644
> --- a/drivers/gpu/drm/xe/xe_wa_oob.rules
> +++ b/drivers/gpu/drm/xe/xe_wa_oob.rules
> @@ -74,6 +74,7 @@
> 16029897822 MEDIA_VERSION(3500)
> GRAPHICS_VERSION(3510)
> 14026539277 GRAPHICS_VERSION(3510), GRAPHICS_STEP(A0, B0)
> +14028564128 GRAPHICS_VERSION(3510), GRAPHICS_STEP(A0, B0)
This is a temporary workaround for graphics version 35.10, meaning that
we should drop GRAPHICS_STEP(A0, B0) here.
--
Gustavo Sousa
> 14027054324 GRAPHICS_VERSION(3511)
> 14025941587 GRAPHICS_VERSION_RANGE(2001, 3511)
> MEDIA_VERSION_RANGE(1301, 3503)
> --
> 2.52.0
prev parent reply other threads:[~2026-11-20 23:58 UTC|newest]
Thread overview: 6+ messages / expand[flat|nested] mbox.gz Atom feed top
2026-11-19 7:52 [PATCH V2] drm/xe/nvl-p: Add Wa_14028564128 Tejas Upadhyay
2026-11-19 8:08 ` sashiko-bot
2026-11-19 9:48 ` ✓ CI.KUnit: success for drm/xe/nvl-p: Add Wa_14028564128 (rev2) Patchwork
2026-11-19 10:47 ` ✓ Xe.CI.BAT: " Patchwork
2026-11-20 15:30 ` ✗ Xe.CI.FULL: failure " Patchwork
2026-11-20 23:58 ` Gustavo Sousa [this message]
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=877bi7qcms.fsf@intel.com \
--to=gustavo.sousa@intel.com \
--cc=intel-xe@lists.freedesktop.org \
--cc=matthew.d.roper@intel.com \
--cc=nitin.r.gote@intel.com \
--cc=tejas.upadhyay@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