From: Matthew Brost <matthew.brost@intel.com>
To: Arvind Yadav <arvind.yadav@intel.com>
Cc: <intel-xe@lists.freedesktop.org>,
<himal.prasad.ghimiray@intel.com>,
<thomas.hellstrom@linux.intel.com>, <pallavi.mishra@intel.com>
Subject: Re: [PATCH v4 6/8] drm/xe/madvise: Implement per-VMA purgeable state tracking
Date: Tue, 20 Jan 2026 09:41:22 -0800 [thread overview]
Message-ID: <aW++QuG51rGP1Ru4@lstrano-desk.jf.intel.com> (raw)
In-Reply-To: <20260120060900.3137984-7-arvind.yadav@intel.com>
On Tue, Jan 20, 2026 at 11:38:52AM +0530, Arvind Yadav wrote:
> Track purgeable state per-VMA instead of using a coarse shared
> BO check. This prevents purging shared BOs until all VMAs across
> all VMs are marked DONTNEED.
>
> Add xe_bo_all_vmas_dontneed() to check all VMAs before marking
> a BO purgeable. Add xe_bo_recheck_purgeable_on_vma_unbind() to
> handle state transitions when VMAs are destroyed - if all
> remaining VMAs are DONTNEED the BO can become purgeable, or if
> no VMAs remain it transitions to WILLNEED.
>
> The per-VMA purgeable_state field stores the madvise hint for
> each mapping. Shared BOs can only be purged when all VMAs
> unanimously indicate DONTNEED.
>
> v3:
> - This addresses Thomas Hellström's feedback: "loop over all vmas
> attached to the bo and check that they all say WONTNEED. This will
> also need a check at VMA unbinding"
>
> v4:
> - @madv_purgeable atomic_t → u32 change across all relevant patches. (Matt)
>
> Cc: Matthew Brost <matthew.brost@intel.com>
> Cc: Thomas Hellström <thomas.hellstrom@linux.intel.com>
> Cc: Himal Prasad Ghimiray <himal.prasad.ghimiray@intel.com>
> Signed-off-by: Arvind Yadav <arvind.yadav@intel.com>
> ---
> drivers/gpu/drm/xe/xe_vm.c | 15 +++++-
> drivers/gpu/drm/xe/xe_vm_madvise.c | 84 +++++++++++++++++++++++++++++-
> drivers/gpu/drm/xe/xe_vm_madvise.h | 3 ++
> drivers/gpu/drm/xe/xe_vm_types.h | 11 ++++
> 4 files changed, 111 insertions(+), 2 deletions(-)
>
> diff --git a/drivers/gpu/drm/xe/xe_vm.c b/drivers/gpu/drm/xe/xe_vm.c
> index f250daae3012..9543960b5613 100644
> --- a/drivers/gpu/drm/xe/xe_vm.c
> +++ b/drivers/gpu/drm/xe/xe_vm.c
> @@ -40,6 +40,7 @@
> #include "xe_tile.h"
> #include "xe_tlb_inval.h"
> #include "xe_trace_bo.h"
> +#include "xe_vm_madvise.h"
> #include "xe_wa.h"
>
> static struct drm_gem_object *xe_vm_obj(struct xe_vm *vm)
> @@ -1079,12 +1080,18 @@ static struct xe_vma *xe_vma_create(struct xe_vm *vm,
> static void xe_vma_destroy_late(struct xe_vma *vma)
> {
> struct xe_vm *vm = xe_vma_vm(vma);
> + struct xe_bo *bo = NULL;
>
> if (vma->ufence) {
> xe_sync_ufence_put(vma->ufence);
> vma->ufence = NULL;
> }
>
> + /* Get BO reference for purgeable state re-check */
> + if (!xe_vma_is_userptr(vma) && !xe_vma_is_null(vma) &&
> + !xe_vma_is_cpu_addr_mirror(vma))
> + bo = xe_vma_bo(vma);
I think xe_vma_bo just returns NULL if any of the above conditions are
met, so I believe is ok to just blindly call xe_vma_bo as you have a
NULL check on the BO below.
> +
> if (xe_vma_is_userptr(vma)) {
> struct xe_userptr_vma *uvma = to_userptr_vma(vma);
>
> @@ -1093,7 +1100,13 @@ static void xe_vma_destroy_late(struct xe_vma *vma)
> } else if (xe_vma_is_null(vma) || xe_vma_is_cpu_addr_mirror(vma)) {
> xe_vm_put(vm);
> } else {
> - xe_bo_put(xe_vma_bo(vma));
> + /* Trylock safe for async context; madvise corrects failures */
> + if (bo && dma_resv_trylock(bo->ttm.base.resv)) {
> + xe_bo_recheck_purgeable_on_vma_unbind(bo);
Also I don't think the correct place to call this.
I believe you can call this function in xe_vma_destroy after
drm_gpuva_unlink. You also have BO lock there too so no need from this
trylock path.
> + dma_resv_unlock(bo->ttm.base.resv);
> + }
> +
> + xe_bo_put(bo);
> }
>
> xe_vma_free(vma);
> diff --git a/drivers/gpu/drm/xe/xe_vm_madvise.c b/drivers/gpu/drm/xe/xe_vm_madvise.c
> index dfeab9e24a09..27b6ad65b314 100644
> --- a/drivers/gpu/drm/xe/xe_vm_madvise.c
> +++ b/drivers/gpu/drm/xe/xe_vm_madvise.c
> @@ -12,6 +12,7 @@
> #include "xe_pat.h"
> #include "xe_pt.h"
> #include "xe_svm.h"
> +#include "xe_vm.h"
>
> struct xe_vmas_in_madvise_range {
> u64 addr;
> @@ -179,6 +180,80 @@ static void madvise_pat_index(struct xe_device *xe, struct xe_vm *vm,
> }
> }
>
> +/**
> + * xe_bo_all_vmas_dontneed() - Check if all VMAs of a BO are marked DONTNEED
> + * @bo: Buffer object
> + *
> + * Check all VMAs across all VMs to determine if BO can be purged.
> + * Shared BOs require unanimous DONTNEED state from all mappings.
> + *
> + * Caller must hold BO dma-resv lock.
> + *
> + * Return: true if all VMAs are DONTNEED, false otherwise
> + */
> +static bool xe_bo_all_vmas_dontneed(struct xe_bo *bo)
> +{
> + struct drm_gpuvm_bo *vm_bo;
> + struct drm_gpuva *gpuva;
> + struct drm_gem_object *obj = &bo->ttm.base;
> + bool has_vmas = false;
> +
> + dma_resv_assert_held(bo->ttm.base.resv);
> +
> + drm_gem_for_each_gpuvm_bo(vm_bo, obj) {
> + drm_gpuvm_bo_for_each_va(gpuva, vm_bo) {
> + struct xe_vma *vma = gpuva_to_vma(gpuva);
> +
> + has_vmas = true;
> +
> + /* Any non-DONTNEED VMA prevents purging */
> + if (READ_ONCE(vma->purgeable_state) != XE_MADV_PURGEABLE_DONTNEED)
You don't need the READ_ONCE as purgeable_state is only accessed under
the dma-resv lock, also there isn't a WRITE_ONCE anywhere.
> + return false;
> + }
> + }
> +
> + /* No VMAs means not purgeable */
No VMAs means it is purgeable, right?
> + if (!has_vmas)
> + return false;
> +
> + return true;
> +}
> +
> +/**
> + * xe_bo_recheck_purgeable_on_vma_unbind() - Re-evaluate BO purgeable state after VMA unbind
> + * @bo: Buffer object
> + *
> + * When a VMA is unbound, re-check if the BO's purgeable state should change.
> + * Destroyed VMAs may allow the BO to become purgeable if all remaining VMAs
> + * are DONTNEED, or require transition to WILLNEED if no VMAs remain.
> + *
> + * Called from VMA destruction path with BO dma-resv lock held.
> + */
> +void xe_bo_recheck_purgeable_on_vma_unbind(struct xe_bo *bo)
> +{
> + if (!bo)
> + return;
> +
> + dma_resv_assert_held(bo->ttm.base.resv);
> +
> + /*
> + * Once purged, always purged. Cannot transition back to WILLNEED.
> + * This matches i915 semantics where purged BOs are permanently invalid.
> + */
> + if (bo->madv_purgeable == XE_MADV_PURGEABLE_PURGED)
> + return;
> +
> + if (xe_bo_all_vmas_dontneed(bo)) {
> + /* All VMAs are DONTNEED - mark BO purgeable */
> + if (bo->madv_purgeable != XE_MADV_PURGEABLE_DONTNEED)
> + bo->madv_purgeable = XE_MADV_PURGEABLE_DONTNEED;
> + } else {
> + /* At least one VMA is WILLNEED - BO must not be purgeable */
> + if (bo->madv_purgeable != XE_MADV_PURGEABLE_WILLNEED)
> + bo->madv_purgeable = XE_MADV_PURGEABLE_WILLNEED;
> + }
> +}
> +
> /*
> * Handle purgeable buffer object advice for DONTNEED/WILLNEED/PURGED.
> * Returns true if any BO was purged, false otherwise.
> @@ -213,10 +288,17 @@ static bool xe_vm_madvise_purgeable_bo(struct xe_device *xe, struct xe_vm *vm,
>
> switch (op->purge_state_val.val) {
> case DRM_XE_VMA_PURGEABLE_STATE_WILLNEED:
> + vmas[i]->purgeable_state = XE_MADV_PURGEABLE_WILLNEED;
> +
> + /* Mark VMA WILLNEED - BO becomes non-purgeable immediately */
> bo->madv_purgeable = XE_MADV_PURGEABLE_WILLNEED;
> break;
> case DRM_XE_VMA_PURGEABLE_STATE_DONTNEED:
> - bo->madv_purgeable = XE_MADV_PURGEABLE_DONTNEED;
> + vmas[i]->purgeable_state = XE_MADV_PURGEABLE_DONTNEED;
> +
> + /* Mark BO purgeable only if all VMAs are DONTNEED */
> + if (xe_bo_all_vmas_dontneed(bo))
> + bo->madv_purgeable = XE_MADV_PURGEABLE_DONTNEED;
> break;
> default:
> drm_warn(&vm->xe->drm, "Invalid madvice value = %d\n",
> diff --git a/drivers/gpu/drm/xe/xe_vm_madvise.h b/drivers/gpu/drm/xe/xe_vm_madvise.h
> index b0e1fc445f23..61868f851949 100644
> --- a/drivers/gpu/drm/xe/xe_vm_madvise.h
> +++ b/drivers/gpu/drm/xe/xe_vm_madvise.h
> @@ -8,8 +8,11 @@
>
> struct drm_device;
> struct drm_file;
> +struct xe_bo;
>
> int xe_vm_madvise_ioctl(struct drm_device *dev, void *data,
> struct drm_file *file);
>
> +void xe_bo_recheck_purgeable_on_vma_unbind(struct xe_bo *bo);
> +
> #endif
> diff --git a/drivers/gpu/drm/xe/xe_vm_types.h b/drivers/gpu/drm/xe/xe_vm_types.h
> index 437f64202f3b..94ca9d033b06 100644
> --- a/drivers/gpu/drm/xe/xe_vm_types.h
> +++ b/drivers/gpu/drm/xe/xe_vm_types.h
> @@ -150,6 +150,17 @@ struct xe_vma {
> */
> bool skip_invalidation;
>
> + /**
> + * @purgeable_state: Purgeable hint for this VMA mapping
> + *
> + * Per-VMA purgeable state from madvise. Valid states are WILLNEED (0)
> + * or DONTNEED (1). Shared BOs require all VMAs to be DONTNEED before
> + * the BO can be purged. PURGED state exists only at BO level.
> + *
> + * Protected by BO dma-resv lock. Set via DRM_IOCTL_XE_MADVISE.
> + */
> + u32 purgeable_state;
> +
I think xe_vma_mem_attr is a better place for this field.
Matt
> /**
> * @ufence: The user fence that was provided with MAP.
> * Needs to be signalled before UNMAP can be processed.
> --
> 2.43.0
>
next prev parent reply other threads:[~2026-01-20 17:41 UTC|newest]
Thread overview: 37+ messages / expand[flat|nested] mbox.gz Atom feed top
2026-01-20 6:08 [PATCH v4 0/8] drm/xe/madvise: Add support for purgeable buffer objects Arvind Yadav
2026-01-20 6:08 ` [PATCH v4 1/8] drm/xe/uapi: Add UAPI " Arvind Yadav
2026-01-20 17:20 ` Matthew Brost
2026-01-21 18:42 ` Vivi, Rodrigo
2026-01-20 6:08 ` [PATCH v4 2/8] drm/xe/bo: Add purgeable bo state tracking and field madv to xe_bo Arvind Yadav
2026-01-20 17:45 ` Matthew Brost
2026-01-21 5:30 ` Yadav, Arvind
2026-01-22 15:05 ` Thomas Hellström
2026-01-20 6:08 ` [PATCH v4 3/8] drm/xe/madvise: Implement purgeable buffer object support Arvind Yadav
2026-01-20 16:58 ` Matthew Brost
2026-01-20 17:15 ` Matthew Brost
2026-01-21 8:24 ` Yadav, Arvind
2026-01-22 15:30 ` Thomas Hellström
2026-01-30 8:13 ` Yadav, Arvind
2026-01-20 17:44 ` Matthew Brost
2026-01-20 6:08 ` [PATCH v4 4/8] drm/xe/bo: Handle CPU faults on purged buffer objects Arvind Yadav
2026-01-20 17:23 ` Matthew Brost
2026-01-22 15:54 ` Thomas Hellström
2026-01-20 6:08 ` [PATCH v4 5/8] drm/xe/vm: Prevent binding of " Arvind Yadav
2026-01-20 17:27 ` Matthew Brost
2026-01-23 5:41 ` Yadav, Arvind
2026-01-23 12:37 ` Thomas Hellström
2026-01-30 8:17 ` Yadav, Arvind
2026-01-20 6:08 ` [PATCH v4 6/8] drm/xe/madvise: Implement per-VMA purgeable state tracking Arvind Yadav
2026-01-20 17:41 ` Matthew Brost [this message]
2026-01-21 5:11 ` Yadav, Arvind
2026-01-23 13:07 ` Thomas Hellström
2026-01-20 6:08 ` [PATCH v4 7/8] drm/xe/madvise: Block imported and exported dma-bufs Arvind Yadav
2026-01-20 17:51 ` Matthew Brost
2026-01-23 13:31 ` Thomas Hellström
2026-01-30 8:22 ` Yadav, Arvind
2026-01-30 8:59 ` Thomas Hellström
2026-01-20 6:08 ` [PATCH v4 8/8] drm/xe/bo: Add purgeable shrinker state helpers Arvind Yadav
2026-01-20 17:58 ` Matthew Brost
2026-01-23 13:42 ` Thomas Hellström
2026-01-20 6:14 ` ✗ CI.checkpatch: warning for drm/xe/madvise: Add support for purgeable buffer objects (rev5) Patchwork
2026-01-20 6:16 ` ✗ CI.KUnit: failure " Patchwork
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=aW++QuG51rGP1Ru4@lstrano-desk.jf.intel.com \
--to=matthew.brost@intel.com \
--cc=arvind.yadav@intel.com \
--cc=himal.prasad.ghimiray@intel.com \
--cc=intel-xe@lists.freedesktop.org \
--cc=pallavi.mishra@intel.com \
--cc=thomas.hellstrom@linux.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