From 92f3403a7ca5d186b2bad342603410cc6adc95a3 Mon Sep 17 00:00:00 2001 From: Arvind Yadav Date: Wed, 6 May 2026 18:50:27 +0530 Subject: [PATCH 1/5] drm/xe/madvise: Track purgeability with BO-local counters MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit xe_bo_recompute_purgeable_state() walks all VMAs of a BO to determine whether the BO can be made purgeable. This makes VMA create/destroy and madvise updates O(n) in the number of mappings. Replace the walk with BO-local counters protected by the BO dma-resv lock: - vma_count tracks the number of VMAs mapping the BO. - willneed_count tracks active WILLNEED holders, including WILLNEED VMAs and active dma-buf exports for non-imported BOs. A DONTNEED BO is promoted back to WILLNEED on a 0->1 transition of willneed_count. A BO is demoted to DONTNEED on a 1->0 transition only when it still has VMAs, preserving the previous behaviour where a BO with no mappings keeps its current madvise state. PURGED remains terminal, preserving the existing "once purged, always purged" rule. Fixes: 4f44961eab84 ("drm/xe/vm: Prevent binding of purged buffer objects") v2: - Use early return for imported BOs in all four helpers to avoid nesting (Matt B). - Group purgeability state into a purgeable sub-struct on struct xe_bo (Matt B). - Reword xe_bo_willneed_put_locked() kernel-doc to explain that a 1->0 transition means all remaining active VMAs are DONTNEED (Matt B). v3: - Move DONTNEED/PURGED reject from vma_lock_and_validate() into xe_vma_create(), gated on attr->purgeable_state == WILLNEED. Fixes vm_bind bypass and partial-unbind rejection on DONTNEED BOs (Matt B). - Drop .check_purged from MAP and REMAP; keep it for PREFETCH and add a comment why (Matt B). - Skip BO validation in vma_lock_and_validate() for non-WILLNEED VMA remnants so cleanup/remap paths do not repopulate DONTNEED/PURGED BOs. Suggested-by: Thomas Hellström Cc: Matthew Brost Cc: Thomas Hellström Cc: Himal Prasad Ghimiray Signed-off-by: Arvind Yadav Reviewed-by: Matthew Brost Link: https://patch.msgid.link/20260506132027.2556046-1-arvind.yadav@intel.com Signed-off-by: Himal Prasad Ghimiray (cherry picked from commit 23fb2ea56cb4fa2587bc072b04e4e698687a48e4) Signed-off-by: Rodrigo Vivi --- drivers/gpu/drm/xe/xe_bo.c | 6 +- drivers/gpu/drm/xe/xe_bo.h | 88 +++++++++++++++- drivers/gpu/drm/xe/xe_bo_types.h | 28 ++++- drivers/gpu/drm/xe/xe_dma_buf.c | 28 ++++- drivers/gpu/drm/xe/xe_vm.c | 51 +++++++-- drivers/gpu/drm/xe/xe_vm_madvise.c | 162 ++--------------------------- drivers/gpu/drm/xe/xe_vm_madvise.h | 2 - 7 files changed, 190 insertions(+), 175 deletions(-) diff --git a/drivers/gpu/drm/xe/xe_bo.c b/drivers/gpu/drm/xe/xe_bo.c index 4075edf97421..6b518858538f 100644 --- a/drivers/gpu/drm/xe/xe_bo.c +++ b/drivers/gpu/drm/xe/xe_bo.c @@ -897,10 +897,10 @@ void xe_bo_set_purgeable_state(struct xe_bo *bo, new_state == XE_MADV_PURGEABLE_PURGED); /* Once purged, always purged - cannot transition out */ - xe_assert(xe, !(bo->madv_purgeable == XE_MADV_PURGEABLE_PURGED && + xe_assert(xe, !(bo->purgeable.state == XE_MADV_PURGEABLE_PURGED && new_state != XE_MADV_PURGEABLE_PURGED)); - bo->madv_purgeable = new_state; + bo->purgeable.state = new_state; xe_bo_set_purgeable_shrinker(bo, new_state); } @@ -2368,7 +2368,7 @@ struct xe_bo *xe_bo_init_locked(struct xe_device *xe, struct xe_bo *bo, INIT_LIST_HEAD(&bo->vram_userfault_link); /* Initialize purge advisory state */ - bo->madv_purgeable = XE_MADV_PURGEABLE_WILLNEED; + bo->purgeable.state = XE_MADV_PURGEABLE_WILLNEED; drm_gem_private_object_init(&xe->drm, &bo->ttm.base, size); diff --git a/drivers/gpu/drm/xe/xe_bo.h b/drivers/gpu/drm/xe/xe_bo.h index 68dea7d25a6b..6340317f7d2e 100644 --- a/drivers/gpu/drm/xe/xe_bo.h +++ b/drivers/gpu/drm/xe/xe_bo.h @@ -251,7 +251,7 @@ static inline bool xe_bo_is_protected(const struct xe_bo *bo) static inline bool xe_bo_is_purged(struct xe_bo *bo) { xe_bo_assert_held(bo); - return bo->madv_purgeable == XE_MADV_PURGEABLE_PURGED; + return bo->purgeable.state == XE_MADV_PURGEABLE_PURGED; } /** @@ -268,11 +268,95 @@ static inline bool xe_bo_is_purged(struct xe_bo *bo) static inline bool xe_bo_madv_is_dontneed(struct xe_bo *bo) { xe_bo_assert_held(bo); - return bo->madv_purgeable == XE_MADV_PURGEABLE_DONTNEED; + return bo->purgeable.state == XE_MADV_PURGEABLE_DONTNEED; } void xe_bo_set_purgeable_state(struct xe_bo *bo, enum xe_madv_purgeable_state new_state); +/** + * xe_bo_willneed_get_locked() - Acquire a WILLNEED holder on a BO + * @bo: Buffer object + * + * Increments willneed_count and, on a 0->1 transition, promotes the BO + * from DONTNEED to WILLNEED. PURGED is terminal and is never modified. + * + * Caller must hold the BO's dma-resv lock. + */ +static inline void xe_bo_willneed_get_locked(struct xe_bo *bo) +{ + xe_bo_assert_held(bo); + + /* Imported BOs are owned externally; do not track purgeability. */ + if (drm_gem_is_imported(&bo->ttm.base)) + return; + + if (bo->purgeable.willneed_count++ == 0 && xe_bo_madv_is_dontneed(bo)) + xe_bo_set_purgeable_state(bo, XE_MADV_PURGEABLE_WILLNEED); +} + +/** + * xe_bo_willneed_put_locked() - Release a WILLNEED holder on a BO + * @bo: Buffer object + * + * Decrements willneed_count and, on a 1->0 transition, marks the BO + * DONTNEED only if it still has VMAs (implying all active VMAs are + * DONTNEED). If the last VMA is being removed, preserve the current BO + * state to match the previous VMA-walk semantics. + * + * PURGED is terminal and the BO state is never modified. + * + * Caller must hold the BO's dma-resv lock. + */ +static inline void xe_bo_willneed_put_locked(struct xe_bo *bo) +{ + xe_bo_assert_held(bo); + + if (drm_gem_is_imported(&bo->ttm.base)) + return; + + xe_assert(xe_bo_device(bo), bo->purgeable.willneed_count > 0); + if (--bo->purgeable.willneed_count == 0 && bo->purgeable.vma_count > 0 && + !xe_bo_is_purged(bo)) + xe_bo_set_purgeable_state(bo, XE_MADV_PURGEABLE_DONTNEED); +} + +/** + * xe_bo_vma_count_inc_locked() - Account a new VMA on a BO + * @bo: Buffer object + * + * Increments vma_count. + * + * Caller must hold the BO's dma-resv lock. + */ +static inline void xe_bo_vma_count_inc_locked(struct xe_bo *bo) +{ + xe_bo_assert_held(bo); + + if (drm_gem_is_imported(&bo->ttm.base)) + return; + + bo->purgeable.vma_count++; +} + +/** + * xe_bo_vma_count_dec_locked() - Account a VMA removal on a BO + * @bo: Buffer object + * + * Decrements vma_count. + * + * Caller must hold the BO's dma-resv lock. + */ +static inline void xe_bo_vma_count_dec_locked(struct xe_bo *bo) +{ + xe_bo_assert_held(bo); + + if (drm_gem_is_imported(&bo->ttm.base)) + return; + + xe_assert(xe_bo_device(bo), bo->purgeable.vma_count > 0); + bo->purgeable.vma_count--; +} + static inline void xe_bo_unpin_map_no_vm(struct xe_bo *bo) { if (likely(bo)) { diff --git a/drivers/gpu/drm/xe/xe_bo_types.h b/drivers/gpu/drm/xe/xe_bo_types.h index 9d19940b8fc0..077e35b4cdce 100644 --- a/drivers/gpu/drm/xe/xe_bo_types.h +++ b/drivers/gpu/drm/xe/xe_bo_types.h @@ -111,10 +111,32 @@ struct xe_bo { u64 min_align; /** - * @madv_purgeable: user space advise on BO purgeability, protected - * by BO's dma-resv lock. + * @purgeable: Purgeability state and accounting. + * + * All fields are protected by the BO's dma-resv lock. */ - u32 madv_purgeable; + struct { + /** + * @purgeable.state: BO purgeability state + * (WILLNEED/DONTNEED/PURGED). + */ + u32 state; + + /** + * @purgeable.vma_count: Number of VMAs currently mapping this BO. + */ + u32 vma_count; + + /** + * @purgeable.willneed_count: Number of active WILLNEED holders. + * + * Counts WILLNEED VMAs plus active dma-buf exports for + * non-imported BOs. The BO flips to DONTNEED on a 1->0 + * transition only when VMAs still exist; if the last VMA is + * removed, the previous BO state is preserved. + */ + u32 willneed_count; + } purgeable; }; #endif diff --git a/drivers/gpu/drm/xe/xe_dma_buf.c b/drivers/gpu/drm/xe/xe_dma_buf.c index b9828da15897..855d32ba314d 100644 --- a/drivers/gpu/drm/xe/xe_dma_buf.c +++ b/drivers/gpu/drm/xe/xe_dma_buf.c @@ -193,6 +193,18 @@ static int xe_dma_buf_begin_cpu_access(struct dma_buf *dma_buf, return 0; } +static void xe_dma_buf_release(struct dma_buf *dmabuf) +{ + struct drm_gem_object *obj = dmabuf->priv; + struct xe_bo *bo = gem_to_xe_bo(obj); + + xe_bo_lock(bo, false); + xe_bo_willneed_put_locked(bo); + xe_bo_unlock(bo); + + drm_gem_dmabuf_release(dmabuf); +} + static const struct dma_buf_ops xe_dmabuf_ops = { .attach = xe_dma_buf_attach, .detach = xe_dma_buf_detach, @@ -200,7 +212,7 @@ static const struct dma_buf_ops xe_dmabuf_ops = { .unpin = xe_dma_buf_unpin, .map_dma_buf = xe_dma_buf_map, .unmap_dma_buf = xe_dma_buf_unmap, - .release = drm_gem_dmabuf_release, + .release = xe_dma_buf_release, .begin_cpu_access = xe_dma_buf_begin_cpu_access, .mmap = drm_gem_dmabuf_mmap, .vmap = drm_gem_dmabuf_vmap, @@ -241,18 +253,26 @@ struct dma_buf *xe_gem_prime_export(struct drm_gem_object *obj, int flags) ret = -EINVAL; goto out_unlock; } + + xe_bo_willneed_get_locked(bo); xe_bo_unlock(bo); ret = ttm_bo_setup_export(&bo->ttm, &ctx); if (ret) - return ERR_PTR(ret); + goto out_put; buf = drm_gem_prime_export(obj, flags); - if (!IS_ERR(buf)) - buf->ops = &xe_dmabuf_ops; + if (IS_ERR(buf)) { + ret = PTR_ERR(buf); + goto out_put; + } + buf->ops = &xe_dmabuf_ops; return buf; +out_put: + xe_bo_lock(bo, false); + xe_bo_willneed_put_locked(bo); out_unlock: xe_bo_unlock(bo); return ERR_PTR(ret); diff --git a/drivers/gpu/drm/xe/xe_vm.c b/drivers/gpu/drm/xe/xe_vm.c index a717a2b8dea3..ab6cc1f0a789 100644 --- a/drivers/gpu/drm/xe/xe_vm.c +++ b/drivers/gpu/drm/xe/xe_vm.c @@ -1120,6 +1120,25 @@ static struct xe_vma *xe_vma_create(struct xe_vm *vm, xe_bo_assert_held(bo); + /* + * Reject only WILLNEED mappings on DONTNEED/PURGED BOs. This + * gates new vm_bind ioctls (user supplies WILLNEED) while + * still allowing partial-unbind / remap splits whose new VMAs + * inherit the parent's DONTNEED attr. It must also run before + * xe_bo_willneed_get_locked() below so a 0->1 holder bump + * cannot silently promote DONTNEED back to WILLNEED. + */ + if (vma->attr.purgeable_state == XE_MADV_PURGEABLE_WILLNEED) { + if (xe_bo_madv_is_dontneed(bo)) { + xe_vma_free(vma); + return ERR_PTR(-EBUSY); + } + if (xe_bo_is_purged(bo)) { + xe_vma_free(vma); + return ERR_PTR(-EINVAL); + } + } + vm_bo = drm_gpuvm_bo_obtain_locked(vma->gpuva.vm, &bo->ttm.base); if (IS_ERR(vm_bo)) { xe_vma_free(vma); @@ -1131,6 +1150,10 @@ static struct xe_vma *xe_vma_create(struct xe_vm *vm, vma->gpuva.gem.offset = bo_offset_or_userptr; drm_gpuva_link(&vma->gpuva, vm_bo); drm_gpuvm_bo_put(vm_bo); + + xe_bo_vma_count_inc_locked(bo); + if (vma->attr.purgeable_state == XE_MADV_PURGEABLE_WILLNEED) + xe_bo_willneed_get_locked(bo); } else /* userptr or null */ { if (!is_null && !is_cpu_addr_mirror) { struct xe_userptr_vma *uvma = to_userptr_vma(vma); @@ -1208,7 +1231,10 @@ static void xe_vma_destroy(struct xe_vma *vma, struct dma_fence *fence) xe_bo_assert_held(bo); drm_gpuva_unlink(&vma->gpuva); - xe_bo_recompute_purgeable_state(bo); + + xe_bo_vma_count_dec_locked(bo); + if (vma->attr.purgeable_state == XE_MADV_PURGEABLE_WILLNEED) + xe_bo_willneed_put_locked(bo); } xe_vm_assert_held(vm); @@ -3016,7 +3042,7 @@ static void vm_bind_ioctl_ops_unwind(struct xe_vm *vm, * @res_evict: Allow evicting resources during validation * @validate: Perform BO validation * @request_decompress: Request BO decompression - * @check_purged: Reject operation if BO is purged + * @check_purged: Reject operation if BO is DONTNEED or PURGED */ struct xe_vma_lock_and_validate_flags { u32 res_evict : 1; @@ -3030,6 +3056,7 @@ static int vma_lock_and_validate(struct drm_exec *exec, struct xe_vma *vma, { struct xe_bo *bo = xe_vma_bo(vma); struct xe_vm *vm = xe_vma_vm(vma); + bool validate_bo = flags.validate; int err = 0; if (bo) { @@ -3044,7 +3071,11 @@ static int vma_lock_and_validate(struct drm_exec *exec, struct xe_vma *vma, err = -EINVAL; /* BO already purged */ } - if (!err && flags.validate) + /* Don't validate the BO for DONTNEED/PURGED remap remnants. */ + if (vma->attr.purgeable_state != XE_MADV_PURGEABLE_WILLNEED) + validate_bo = false; + + if (!err && validate_bo) err = xe_bo_validate(bo, vm, xe_vm_allow_vm_eviction(vm) && flags.res_evict, exec); @@ -3152,7 +3183,7 @@ static int op_lock_and_prep(struct drm_exec *exec, struct xe_vm *vm, op->map.immediate, .request_decompress = op->map.request_decompress, - .check_purged = true, + .check_purged = false, }); break; case DRM_GPUVA_OP_REMAP: @@ -3174,7 +3205,7 @@ static int op_lock_and_prep(struct drm_exec *exec, struct xe_vm *vm, .res_evict = res_evict, .validate = true, .request_decompress = false, - .check_purged = true, + .check_purged = false, }); if (!err && op->remap.next) err = vma_lock_and_validate(exec, op->remap.next, @@ -3182,7 +3213,7 @@ static int op_lock_and_prep(struct drm_exec *exec, struct xe_vm *vm, .res_evict = res_evict, .validate = true, .request_decompress = false, - .check_purged = true, + .check_purged = false, }); break; case DRM_GPUVA_OP_UNMAP: @@ -3211,9 +3242,11 @@ static int op_lock_and_prep(struct drm_exec *exec, struct xe_vm *vm, } /* - * Prefetch attempts to migrate BO's backing store without - * repopulating it first. Purged BOs have no backing store - * to migrate, so reject the operation. + * PREFETCH is the only op that still gates on BO purge state. + * MAP/REMAP handle this inside xe_vma_create() so partial + * unbind on a DONTNEED BO still works. PREFETCH skips + * xe_vma_create() and would migrate a BO with no backing + * store, so reject DONTNEED/PURGED here. */ err = vma_lock_and_validate(exec, gpuva_to_vma(op->base.prefetch.va), diff --git a/drivers/gpu/drm/xe/xe_vm_madvise.c b/drivers/gpu/drm/xe/xe_vm_madvise.c index c78906dea82b..c4fb29004195 100644 --- a/drivers/gpu/drm/xe/xe_vm_madvise.c +++ b/drivers/gpu/drm/xe/xe_vm_madvise.c @@ -185,147 +185,6 @@ static void madvise_pat_index(struct xe_device *xe, struct xe_vm *vm, } } -/** - * xe_bo_is_dmabuf_shared() - Check if BO is shared via dma-buf - * @bo: Buffer object - * - * Prevent marking imported or exported dma-bufs as purgeable. - * For imported BOs, Xe doesn't own the backing store and cannot - * safely reclaim pages (exporter or other devices may still be - * using them). For exported BOs, external devices may have active - * mappings we cannot track. - * - * Return: true if BO is imported or exported, false otherwise - */ -static bool xe_bo_is_dmabuf_shared(struct xe_bo *bo) -{ - struct drm_gem_object *obj = &bo->ttm.base; - - /* Imported: exporter owns backing store */ - if (drm_gem_is_imported(obj)) - return true; - - /* Exported: external devices may be accessing */ - if (obj->dma_buf) - return true; - - return false; -} - -/** - * enum xe_bo_vmas_purge_state - VMA purgeable state aggregation - * - * Distinguishes whether a BO's VMAs are all DONTNEED, have at least - * one WILLNEED, or have no VMAs at all. - * - * Enum values align with XE_MADV_PURGEABLE_* states for consistency. - */ -enum xe_bo_vmas_purge_state { - /** @XE_BO_VMAS_STATE_WILLNEED: At least one VMA is WILLNEED */ - XE_BO_VMAS_STATE_WILLNEED = 0, - /** @XE_BO_VMAS_STATE_DONTNEED: All VMAs are DONTNEED */ - XE_BO_VMAS_STATE_DONTNEED = 1, - /** @XE_BO_VMAS_STATE_NO_VMAS: BO has no VMAs */ - XE_BO_VMAS_STATE_NO_VMAS = 2, -}; - -/* - * xe_bo_recompute_purgeable_state() casts between xe_bo_vmas_purge_state and - * xe_madv_purgeable_state. Enforce that WILLNEED=0 and DONTNEED=1 match across - * both enums so the single-line cast is always valid. - */ -static_assert(XE_BO_VMAS_STATE_WILLNEED == (int)XE_MADV_PURGEABLE_WILLNEED, - "VMA purge state WILLNEED must equal madv purgeable WILLNEED"); -static_assert(XE_BO_VMAS_STATE_DONTNEED == (int)XE_MADV_PURGEABLE_DONTNEED, - "VMA purge state DONTNEED must equal madv purgeable DONTNEED"); - -/** - * xe_bo_all_vmas_dontneed() - Determine BO VMA purgeable state - * @bo: Buffer object - * - * Check all VMAs across all VMs to determine aggregate purgeable state. - * Shared BOs require unanimous DONTNEED state from all mappings. - * - * Caller must hold BO dma-resv lock. - * - * Return: XE_BO_VMAS_STATE_DONTNEED if all VMAs are DONTNEED, - * XE_BO_VMAS_STATE_WILLNEED if at least one VMA is not DONTNEED, - * XE_BO_VMAS_STATE_NO_VMAS if BO has no VMAs - */ -static enum xe_bo_vmas_purge_state 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; - - xe_bo_assert_held(bo); - - /* Shared dma-bufs cannot be purgeable */ - if (xe_bo_is_dmabuf_shared(bo)) - return XE_BO_VMAS_STATE_WILLNEED; - - 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 (vma->attr.purgeable_state != XE_MADV_PURGEABLE_DONTNEED) - return XE_BO_VMAS_STATE_WILLNEED; - } - } - - /* - * No VMAs => preserve existing BO purgeable state. - * Avoids incorrectly flipping DONTNEED -> WILLNEED when last VMA unmapped. - */ - if (!has_vmas) - return XE_BO_VMAS_STATE_NO_VMAS; - - return XE_BO_VMAS_STATE_DONTNEED; -} - -/** - * xe_bo_recompute_purgeable_state() - Recompute BO purgeable state from VMAs - * @bo: Buffer object - * - * Walk all VMAs to determine if BO should be purgeable or not. - * Shared BOs require unanimous DONTNEED state from all mappings. - * If the BO has no VMAs the existing state is preserved. - * - * Locking: Caller must hold BO dma-resv lock. When iterating GPUVM lists, - * VM lock must also be held (write) to prevent concurrent VMA modifications. - * This is satisfied at both call sites: - * - xe_vma_destroy(): holds vm->lock write - * - madvise_purgeable(): holds vm->lock write (from madvise ioctl path) - * - * Return: nothing - */ -void xe_bo_recompute_purgeable_state(struct xe_bo *bo) -{ - enum xe_bo_vmas_purge_state vma_state; - - if (!bo) - return; - - xe_bo_assert_held(bo); - - /* - * 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; - - vma_state = xe_bo_all_vmas_dontneed(bo); - - if (vma_state != (enum xe_bo_vmas_purge_state)bo->madv_purgeable && - vma_state != XE_BO_VMAS_STATE_NO_VMAS) - xe_bo_set_purgeable_state(bo, (enum xe_madv_purgeable_state)vma_state); -} - /** * madvise_purgeable - Handle purgeable buffer object advice * @xe: XE device @@ -359,12 +218,6 @@ static void madvise_purgeable(struct xe_device *xe, struct xe_vm *vm, /* BO must be locked before modifying madv state */ xe_bo_assert_held(bo); - /* Skip shared dma-bufs - no PTEs to zap */ - if (xe_bo_is_dmabuf_shared(bo)) { - vmas[i]->skip_invalidation = true; - continue; - } - /* * Once purged, always purged. Cannot transition back to WILLNEED. * This matches i915 semantics where purged BOs are permanently invalid. @@ -377,13 +230,14 @@ static void madvise_purgeable(struct xe_device *xe, struct xe_vm *vm, switch (op->purge_state_val.val) { case DRM_XE_VMA_PURGEABLE_STATE_WILLNEED: - vmas[i]->attr.purgeable_state = XE_MADV_PURGEABLE_WILLNEED; vmas[i]->skip_invalidation = true; - - xe_bo_recompute_purgeable_state(bo); + /* Only act on a real DONTNEED -> WILLNEED transition. */ + if (vmas[i]->attr.purgeable_state == XE_MADV_PURGEABLE_DONTNEED) { + vmas[i]->attr.purgeable_state = XE_MADV_PURGEABLE_WILLNEED; + xe_bo_willneed_get_locked(bo); + } break; case DRM_XE_VMA_PURGEABLE_STATE_DONTNEED: - vmas[i]->attr.purgeable_state = XE_MADV_PURGEABLE_DONTNEED; /* * Don't zap PTEs at DONTNEED time -- pages are still * alive. The zap happens in xe_bo_move_notify() right @@ -391,7 +245,11 @@ static void madvise_purgeable(struct xe_device *xe, struct xe_vm *vm, */ vmas[i]->skip_invalidation = true; - xe_bo_recompute_purgeable_state(bo); + /* Only act on a real WILLNEED -> DONTNEED transition. */ + if (vmas[i]->attr.purgeable_state == XE_MADV_PURGEABLE_WILLNEED) { + vmas[i]->attr.purgeable_state = XE_MADV_PURGEABLE_DONTNEED; + xe_bo_willneed_put_locked(bo); + } break; default: /* Should never hit - values validated in madvise_args_are_sane() */ diff --git a/drivers/gpu/drm/xe/xe_vm_madvise.h b/drivers/gpu/drm/xe/xe_vm_madvise.h index 39acd2689ca0..a3078f634c7e 100644 --- a/drivers/gpu/drm/xe/xe_vm_madvise.h +++ b/drivers/gpu/drm/xe/xe_vm_madvise.h @@ -13,6 +13,4 @@ struct xe_bo; int xe_vm_madvise_ioctl(struct drm_device *dev, void *data, struct drm_file *file); -void xe_bo_recompute_purgeable_state(struct xe_bo *bo); - #endif From 4568dfbb8363a153e7cef384d23cc6d6563ff316 Mon Sep 17 00:00:00 2001 From: Matt Roper Date: Thu, 7 May 2026 14:00:15 -0700 Subject: [PATCH 2/5] drm/xe: Make decision to use Xe2-style blitter instructions a feature flag The blitter engines' MEM_COPY and MEM_SET instructions were added as part of the same hardware change that introduced service copy engines (i.e., BCS1-BCS8) which is why the driver checks for service copy engine presence when deciding whether to use these instructions or the older XY_* instructions. However when making this decision the driver should consider which engines are part of the hardware architecture, not which engines are present/usable on the current device. For graphics IP versions that architecturally include service copy engines (i.e., everything Xe2 and later, plus PVC's Xe_HPC) we should use MEM_SET and MEM_COPY even in if all of the service copy engines wind up getting fused off. I.e., we need to decide based on whether the platform's graphics descriptor contains these engines, rather than whether the usable engine mask contains them. This logic got broken when gt->info.__engine_mask was removed, although in practice that mistake has been harmless so far because there haven't been any hardware SKUs that fuse off all of the service copy engines yet. Replace the incorrect has_service_copy_support() function with a GT feature flag that tracks more accurately whether the new blitter instructions are usable. In addition to fixing incorrect logic if all service copies are fused off, the flag also makes it more obvious what the calling code is trying to do; previously it wasn't terribly obvious why "has service copy engines" was being used as the condition for using different instructions on all copy engine types. The new feature flag is named 'has_xe2_blt_instructions' because we expect this flag to be set for all Xe2 and later platforms (i.e., everything officially supported by the Xe driver). Technically there's also one Xe1-era platform (PVC) that supports these engines/instructions and will set this flag, but this still seems to be the most clear and understandable name for the flag. Fixes: 61549a2ee594 ("drm/xe: Drop __engine_mask") Cc: Balasubramani Vivekanandan Reviewed-by: Balasubramani Vivekanandan Link: https://patch.msgid.link/20260507-xe2_copy-v1-1-26506381b821@intel.com Signed-off-by: Matt Roper (cherry picked from commit 09b399842907565a64e351fb22da790b4c673ffb) Signed-off-by: Rodrigo Vivi --- drivers/gpu/drm/xe/xe_gt_types.h | 7 +++++++ drivers/gpu/drm/xe/xe_migrate.c | 18 ++---------------- drivers/gpu/drm/xe/xe_pci.c | 9 +++++++++ 3 files changed, 18 insertions(+), 16 deletions(-) diff --git a/drivers/gpu/drm/xe/xe_gt_types.h b/drivers/gpu/drm/xe/xe_gt_types.h index 8b55cf25a75f..fffb5d631b69 100644 --- a/drivers/gpu/drm/xe/xe_gt_types.h +++ b/drivers/gpu/drm/xe/xe_gt_types.h @@ -144,6 +144,13 @@ struct xe_gt { u8 id; /** @info.has_indirect_ring_state: GT has indirect ring state support */ u8 has_indirect_ring_state:1; + /** + * @info.has_xe2_blt_instructions: GT supports Xe2-style MEM_SET + * and MEM_COPY blitter functionality. Note that despite the + * name, some Xe1 platforms may also support this "Xe2-style" + * feature. + */ + u8 has_xe2_blt_instructions:1; /** * @info.num_geometry_xecore_fuse_regs: Number of 32b-bit fuse * registers the geometry XeCore mask spans. diff --git a/drivers/gpu/drm/xe/xe_migrate.c b/drivers/gpu/drm/xe/xe_migrate.c index 5fdc89ed5256..a22413f892a0 100644 --- a/drivers/gpu/drm/xe/xe_migrate.c +++ b/drivers/gpu/drm/xe/xe_migrate.c @@ -1524,23 +1524,9 @@ static void emit_clear_main_copy(struct xe_gt *gt, struct xe_bb *bb, bb->len += len; } -static bool has_service_copy_support(struct xe_gt *gt) -{ - /* - * What we care about is whether the architecture was designed with - * service copy functionality (specifically the new MEM_SET / MEM_COPY - * instructions) so check the architectural engine list rather than the - * actual list since these instructions are usable on BCS0 even if - * all of the actual service copy engines (BCS1-BCS8) have been fused - * off. - */ - return gt->info.engine_mask & GENMASK(XE_HW_ENGINE_BCS8, - XE_HW_ENGINE_BCS1); -} - static u32 emit_clear_cmd_len(struct xe_gt *gt) { - if (has_service_copy_support(gt)) + if (gt->info.has_xe2_blt_instructions) return PVC_MEM_SET_CMD_LEN_DW; else return XY_FAST_COLOR_BLT_DW; @@ -1549,7 +1535,7 @@ static u32 emit_clear_cmd_len(struct xe_gt *gt) static void emit_clear(struct xe_gt *gt, struct xe_bb *bb, u64 src_ofs, u32 size, u32 pitch, bool is_vram) { - if (has_service_copy_support(gt)) + if (gt->info.has_xe2_blt_instructions) emit_clear_link_copy(gt, bb, src_ofs, size, pitch); else emit_clear_main_copy(gt, bb, src_ofs, size, pitch, diff --git a/drivers/gpu/drm/xe/xe_pci.c b/drivers/gpu/drm/xe/xe_pci.c index 9f98d0334164..c2ecd27ec770 100644 --- a/drivers/gpu/drm/xe/xe_pci.c +++ b/drivers/gpu/drm/xe/xe_pci.c @@ -851,6 +851,15 @@ static struct xe_gt *alloc_primary_gt(struct xe_tile *tile, gt->info.num_geometry_xecore_fuse_regs = graphics_desc->num_geometry_xecore_fuse_regs; gt->info.num_compute_xecore_fuse_regs = graphics_desc->num_compute_xecore_fuse_regs; + /* + * Even if the service copy engines wind up being fused off, their + * presence in the IP descriptor indicates that the platform supports + * Xe2-style MEM_SET and MEM_COPY functionality. + */ + if (graphics_desc->hw_engine_mask & GENMASK(XE_HW_ENGINE_BCS8, + XE_HW_ENGINE_BCS1)) + gt->info.has_xe2_blt_instructions = true; + /* * Before media version 13, the media IP was part of the primary GT * so we need to add the media engines to the primary GT's engine list. From 981bedbbe61364fcc3a3b87ebaf648a66cd07108 Mon Sep 17 00:00:00 2001 From: Matthew Auld Date: Fri, 8 May 2026 11:26:36 +0100 Subject: [PATCH 3/5] drm/xe/dma-buf: handle empty bo and UAF races MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit There look to be some nasty races here when triggering the invalidate_mappings hook: 1) We do xe_bo_alloc() followed by the attach, before the actual full bo init step in xe_dma_buf_init_obj(). However the bo is visible on the attachments list after the attach. This is bad since exporter driver, say amdgpu, can at any time call back into our invalidate_mappings hook, with an empty/bogus bo, leading to potential bugs/crashes. 2) Similar to 1) but here we get a UAF, when the invalidate_mappings hook is triggered. For example, we get as far as xe_bo_init_locked() but this fails in some way. But here the bo will be freed on error, but we still have it attached from dma-buf pov, so if the invalidate_mappings is now triggered then the bo we access is gone and we trigger UAF and more bugs/crashes. To fix this, move the attach step until after we actually have a fully set up buffer object. Note that the bo is not published to userspace until later, so not sure what the comment "Don't publish the bo until we have a valid attachment", is referring to. We have at least two different customers reporting hitting a NULL ptr deref in evict_flags when importing something from amdgpu, followed by triggering the evict flow. Hit rate is also pretty low, which would hint at some kind of race, so something like 1) or 2) might explain this. v2: - Shuffle the order of the ops slightly (no functional change) - Improve the comment to better explain the ordering (Matt B) Assisted-by: Gemini:gemini-3 #debug Link: https://gitlab.freedesktop.org/drm/xe/kernel/-/work_items/7903 Link: https://gitlab.freedesktop.org/drm/xe/kernel/-/work_items/4055 Fixes: dd08ebf6c352 ("drm/xe: Introduce a new DRM driver for Intel GPUs") Signed-off-by: Matthew Auld Cc: Thomas Hellström Cc: Matthew Brost Cc: # v6.8+ Reviewed-by: Matthew Brost Acked-by: Thomas Hellström Link: https://patch.msgid.link/20260508102635.149172-3-matthew.auld@intel.com (cherry picked from commit af1f2ad0c59fe4e2f924c526f66e968289d77971) Signed-off-by: Rodrigo Vivi --- drivers/gpu/drm/xe/xe_dma_buf.c | 31 ++++++++++++++++--------------- 1 file changed, 16 insertions(+), 15 deletions(-) diff --git a/drivers/gpu/drm/xe/xe_dma_buf.c b/drivers/gpu/drm/xe/xe_dma_buf.c index 855d32ba314d..f7b47e9d1a4c 100644 --- a/drivers/gpu/drm/xe/xe_dma_buf.c +++ b/drivers/gpu/drm/xe/xe_dma_buf.c @@ -377,15 +377,25 @@ struct drm_gem_object *xe_gem_prime_import(struct drm_device *dev, } } - /* - * Don't publish the bo until we have a valid attachment, and a - * valid attachment needs the bo address. So pre-create a bo before - * creating the attachment and publish. - */ bo = xe_bo_alloc(); if (IS_ERR(bo)) return ERR_CAST(bo); + /* + * xe_dma_buf_init_obj() takes ownership of the raw bo, so do not touch + * on fail, since it will already take care of cleanup. On success we + * still need to drop the ref, if something later fails. + * + * In addition this needs to happen before the attach, since + * it will create a new attachment for this, and add it to the list of + * attachments, at which point it is globally visible, and at any point + * the export side can call into on invalidate_mappings callback, which + * require a working object. + */ + obj = xe_dma_buf_init_obj(dev, bo, dma_buf); + if (IS_ERR(obj)) + return obj; + attach_ops = &xe_dma_buf_attach_ops; #if IS_ENABLED(CONFIG_DRM_XE_KUNIT_TEST) if (test) @@ -398,21 +408,12 @@ struct drm_gem_object *xe_gem_prime_import(struct drm_device *dev, goto out_err; } - /* - * xe_dma_buf_init_obj() takes ownership of bo on both success - * and failure, so we must not touch bo after this call. - */ - obj = xe_dma_buf_init_obj(dev, bo, dma_buf); - if (IS_ERR(obj)) { - dma_buf_detach(dma_buf, attach); - return obj; - } get_dma_buf(dma_buf); obj->import_attach = attach; return obj; out_err: - xe_bo_free(bo); + xe_bo_put(bo); return obj; } From 155a372a1cc50fa93387c5d3cdfd614a61e1afd1 Mon Sep 17 00:00:00 2001 From: Matthew Auld Date: Fri, 8 May 2026 11:26:37 +0100 Subject: [PATCH 4/5] drm/xe/dma-buf: fix UAF with retry loop MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit Retry doesn't work here, since bo will be freed on error, leading to UAF. However, now that we do the alloc & init before the attach, we can now combine this as one unit and have the init do the alloc for us. This should make the retry safe. Reported by Sashiko. v2: Fix up the error unwind (CI) Closes: https://sashiko.dev/#/patchset/20260506184332.86743-2-matthew.auld%40intel.com Fixes: eb289a5f6cc6 ("drm/xe: Convert xe_dma_buf.c for exhaustive eviction") Signed-off-by: Matthew Auld Cc: Thomas Hellström Cc: Matthew Brost Cc: # v6.18+ Reviewed-by: Thomas Hellström Link: https://patch.msgid.link/20260508102635.149172-4-matthew.auld@intel.com (cherry picked from commit 479669418253e0f27f8cf5db01a731352ea592e7) Signed-off-by: Rodrigo Vivi --- drivers/gpu/drm/xe/xe_dma_buf.c | 49 ++++++++------------------------- 1 file changed, 12 insertions(+), 37 deletions(-) diff --git a/drivers/gpu/drm/xe/xe_dma_buf.c b/drivers/gpu/drm/xe/xe_dma_buf.c index f7b47e9d1a4c..8a920e58245c 100644 --- a/drivers/gpu/drm/xe/xe_dma_buf.c +++ b/drivers/gpu/drm/xe/xe_dma_buf.c @@ -278,16 +278,8 @@ out_unlock: return ERR_PTR(ret); } -/* - * Takes ownership of @storage: on success it is transferred to the returned - * drm_gem_object; on failure it is freed before returning the error. - * This matches the contract of xe_bo_init_locked() which frees @storage on - * its error paths, so callers need not (and must not) free @storage after - * this call. - */ static struct drm_gem_object * -xe_dma_buf_init_obj(struct drm_device *dev, struct xe_bo *storage, - struct dma_buf *dma_buf) +xe_dma_buf_create_obj(struct drm_device *dev, struct dma_buf *dma_buf) { struct dma_resv *resv = dma_buf->resv; struct xe_device *xe = to_xe_device(dev); @@ -298,10 +290,8 @@ xe_dma_buf_init_obj(struct drm_device *dev, struct xe_bo *storage, int ret = 0; dummy_obj = drm_gpuvm_resv_object_alloc(&xe->drm); - if (!dummy_obj) { - xe_bo_free(storage); + if (!dummy_obj) return ERR_PTR(-ENOMEM); - } dummy_obj->resv = resv; xe_validation_guard(&ctx, &xe->val, &exec, (struct xe_val_flags) {}, ret) { @@ -310,8 +300,7 @@ xe_dma_buf_init_obj(struct drm_device *dev, struct xe_bo *storage, if (ret) break; - /* xe_bo_init_locked() frees storage on error */ - bo = xe_bo_init_locked(xe, storage, NULL, resv, NULL, dma_buf->size, + bo = xe_bo_init_locked(xe, NULL, NULL, resv, NULL, dma_buf->size, 0, /* Will require 1way or 2way for vm_bind */ ttm_bo_type_sg, XE_BO_FLAG_SYSTEM, &exec); drm_exec_retry_on_contention(&exec); @@ -362,7 +351,6 @@ struct drm_gem_object *xe_gem_prime_import(struct drm_device *dev, const struct dma_buf_attach_ops *attach_ops; struct dma_buf_attachment *attach; struct drm_gem_object *obj; - struct xe_bo *bo; if (dma_buf->ops == &xe_dmabuf_ops) { obj = dma_buf->priv; @@ -377,22 +365,14 @@ struct drm_gem_object *xe_gem_prime_import(struct drm_device *dev, } } - bo = xe_bo_alloc(); - if (IS_ERR(bo)) - return ERR_CAST(bo); - /* - * xe_dma_buf_init_obj() takes ownership of the raw bo, so do not touch - * on fail, since it will already take care of cleanup. On success we - * still need to drop the ref, if something later fails. - * - * In addition this needs to happen before the attach, since - * it will create a new attachment for this, and add it to the list of - * attachments, at which point it is globally visible, and at any point - * the export side can call into on invalidate_mappings callback, which - * require a working object. + * This needs to happen before the attach, since it will create a new + * attachment for this, and add it to the list of attachments, at which + * point it is globally visible, and at any point the export side can + * call into on invalidate_mappings callback, which require a working + * object. */ - obj = xe_dma_buf_init_obj(dev, bo, dma_buf); + obj = xe_dma_buf_create_obj(dev, dma_buf); if (IS_ERR(obj)) return obj; @@ -402,20 +382,15 @@ struct drm_gem_object *xe_gem_prime_import(struct drm_device *dev, attach_ops = test->attach_ops; #endif - attach = dma_buf_dynamic_attach(dma_buf, dev->dev, attach_ops, &bo->ttm.base); + attach = dma_buf_dynamic_attach(dma_buf, dev->dev, attach_ops, obj); if (IS_ERR(attach)) { - obj = ERR_CAST(attach); - goto out_err; + xe_bo_put(gem_to_xe_bo(obj)); + return ERR_CAST(attach); } get_dma_buf(dma_buf); obj->import_attach = attach; return obj; - -out_err: - xe_bo_put(bo); - - return obj; } #if IS_ENABLED(CONFIG_DRM_XE_KUNIT_TEST) From 8fb70afe671cc8c1f6237a39aabd50714fcd1189 Mon Sep 17 00:00:00 2001 From: Michal Wajdeczko Date: Sun, 10 May 2026 22:56:05 +0200 Subject: [PATCH 5/5] drm/xe: Drop unused ggtt_balloon field During recent GGTT refactoring we missed to drop now unused field from the xe_tile. Drop it now. Fixes: e904c56ba6e0 ("drm/xe: Rewrite GGTT VF initialization") Signed-off-by: Michal Wajdeczko Reviewed-by: Maarten Lankhorst Link: https://patch.msgid.link/20260510205605.642-1-michal.wajdeczko@intel.com (cherry picked from commit 21d5a871f57909dc4d8e4f5d3bf92f9ccf2597b2) Signed-off-by: Rodrigo Vivi --- drivers/gpu/drm/xe/xe_tile_types.h | 2 -- 1 file changed, 2 deletions(-) diff --git a/drivers/gpu/drm/xe/xe_tile_types.h b/drivers/gpu/drm/xe/xe_tile_types.h index 33932fd547d7..0048100ccb72 100644 --- a/drivers/gpu/drm/xe/xe_tile_types.h +++ b/drivers/gpu/drm/xe/xe_tile_types.h @@ -106,8 +106,6 @@ struct xe_tile { struct xe_lmtt lmtt; } pf; struct { - /** @sriov.vf.ggtt_balloon: GGTT regions excluded from use. */ - struct xe_ggtt_node *ggtt_balloon[2]; /** @sriov.vf.self_config: VF configuration data */ struct xe_tile_sriov_vf_selfconfig self_config; } vf;