Merge tag 'drm-xe-fixes-2025-03-06' of https://gitlab.freedesktop.org/drm/xe/kernel into drm-fixes

- Remove double page flip on initial plane (Maarten)
- Properly setup userptr pfn_flags_mask (Auld)
- Fix GT "for each engine" workarounds (Tvrtko)
- Fix userptr races and missed validations (Thomas, Brost)
- Userptr invalid page access fixes (Thomas)

Signed-off-by: Dave Airlie <airlied@redhat.com>

From: Rodrigo Vivi <rodrigo.vivi@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/Z8ni6w3tskCFL11O@intel.com

10 files changed, 289 insertions, 141 deletions

diff --git a/drivers/gpu/drm/xe/display/xe_plane_initial.c b/drivers/gpu/drm/xe/display/xe_plane_initial.c
index 2eb9633f163a..2a2f250fa495 100644
--- a/drivers/gpu/drm/xe/display/xe_plane_initial.c
+++ b/drivers/gpu/drm/xe/display/xe_plane_initial.c
@@ -194,8 +194,6 @@ intel_find_initial_plane_obj(struct intel_crtc *crtc,
 		to_intel_plane(crtc->base.primary);
 	struct intel_plane_state *plane_state =
 		to_intel_plane_state(plane->base.state);
-	struct intel_crtc_state *crtc_state =
-		to_intel_crtc_state(crtc->base.state);
 	struct drm_framebuffer *fb;
 	struct i915_vma *vma;
 
@@ -241,14 +239,6 @@ intel_find_initial_plane_obj(struct intel_crtc *crtc,
 	atomic_or(plane->frontbuffer_bit, &to_intel_frontbuffer(fb)->bits);
 
 	plane_config->vma = vma;
-
-	/*
-	 * Flip to the newly created mapping ASAP, so we can re-use the
-	 * first part of GGTT for WOPCM, prevent flickering, and prevent
-	 * the lookup of sysmem scratch pages.
-	 */
-	plane->check_plane(crtc_state, plane_state);
-	plane->async_flip(NULL, plane, crtc_state, plane_state, true);
 	return;
 
 nofb:
diff --git a/drivers/gpu/drm/xe/xe_gt.c b/drivers/gpu/drm/xe/xe_gt.c
index 5d6fb79957b6..9f4f27d1ef4a 100644
--- a/drivers/gpu/drm/xe/xe_gt.c
+++ b/drivers/gpu/drm/xe/xe_gt.c
@@ -380,9 +380,7 @@ int xe_gt_init_early(struct xe_gt *gt)
 	if (err)
 		return err;
 
-	xe_wa_process_gt(gt);
 	xe_wa_process_oob(gt);
-	xe_tuning_process_gt(gt);
 
 	xe_force_wake_init_gt(gt, gt_to_fw(gt));
 	spin_lock_init(&gt->global_invl_lock);
@@ -474,6 +472,8 @@ static int all_fw_domain_init(struct xe_gt *gt)
 	}
 
 	xe_gt_mcr_set_implicit_defaults(gt);
+	xe_wa_process_gt(gt);
+	xe_tuning_process_gt(gt);
 	xe_reg_sr_apply_mmio(&gt->reg_sr, gt);
 
 	err = xe_gt_clock_init(gt);
diff --git a/drivers/gpu/drm/xe/xe_hmm.c b/drivers/gpu/drm/xe/xe_hmm.c
index 089834467880..392102515f3d 100644
--- a/drivers/gpu/drm/xe/xe_hmm.c
+++ b/drivers/gpu/drm/xe/xe_hmm.c
@@ -19,11 +19,10 @@ static u64 xe_npages_in_range(unsigned long start, unsigned long end)
 	return (end - start) >> PAGE_SHIFT;
 }
 
-/*
+/**
  * xe_mark_range_accessed() - mark a range is accessed, so core mm
  * have such information for memory eviction or write back to
  * hard disk
- *
  * @range: the range to mark
  * @write: if write to this range, we mark pages in this range
  * as dirty
@@ -43,15 +42,51 @@ static void xe_mark_range_accessed(struct hmm_range *range, bool write)
 	}
 }
 
-/*
+static int xe_alloc_sg(struct xe_device *xe, struct sg_table *st,
+		       struct hmm_range *range, struct rw_semaphore *notifier_sem)
+{
+	unsigned long i, npages, hmm_pfn;
+	unsigned long num_chunks = 0;
+	int ret;
+
+	/* HMM docs says this is needed. */
+	ret = down_read_interruptible(notifier_sem);
+	if (ret)
+		return ret;
+
+	if (mmu_interval_read_retry(range->notifier, range->notifier_seq)) {
+		up_read(notifier_sem);
+		return -EAGAIN;
+	}
+
+	npages = xe_npages_in_range(range->start, range->end);
+	for (i = 0; i < npages;) {
+		unsigned long len;
+
+		hmm_pfn = range->hmm_pfns[i];
+		xe_assert(xe, hmm_pfn & HMM_PFN_VALID);
+
+		len = 1UL << hmm_pfn_to_map_order(hmm_pfn);
+
+		/* If order > 0 the page may extend beyond range->start */
+		len -= (hmm_pfn & ~HMM_PFN_FLAGS) & (len - 1);
+		i += len;
+		num_chunks++;
+	}
+	up_read(notifier_sem);
+
+	return sg_alloc_table(st, num_chunks, GFP_KERNEL);
+}
+
+/**
  * xe_build_sg() - build a scatter gather table for all the physical pages/pfn
  * in a hmm_range. dma-map pages if necessary. dma-address is save in sg table
  * and will be used to program GPU page table later.
- *
  * @xe: the xe device who will access the dma-address in sg table
  * @range: the hmm range that we build the sg table from. range->hmm_pfns[]
  * has the pfn numbers of pages that back up this hmm address range.
  * @st: pointer to the sg table.
+ * @notifier_sem: The xe notifier lock.
  * @write: whether we write to this range. This decides dma map direction
  * for system pages. If write we map it bi-diretional; otherwise
  * DMA_TO_DEVICE
@@ -78,43 +113,84 @@ static void xe_mark_range_accessed(struct hmm_range *range, bool write)
  * Returns 0 if successful; -ENOMEM if fails to allocate memory
  */
 static int xe_build_sg(struct xe_device *xe, struct hmm_range *range,
-		       struct sg_table *st, bool write)
+		       struct sg_table *st,
+		       struct rw_semaphore *notifier_sem,
+		       bool write)
 {
+	unsigned long npages = xe_npages_in_range(range->start, range->end);
 	struct device *dev = xe->drm.dev;
-	struct page **pages;
-	u64 i, npages;
-	int ret;
+	struct scatterlist *sgl;
+	struct page *page;
+	unsigned long i, j;
 
-	npages = xe_npages_in_range(range->start, range->end);
-	pages = kvmalloc_array(npages, sizeof(*pages), GFP_KERNEL);
-	if (!pages)
-		return -ENOMEM;
+	lockdep_assert_held(notifier_sem);
 
-	for (i = 0; i < npages; i++) {
-		pages[i] = hmm_pfn_to_page(range->hmm_pfns[i]);
-		xe_assert(xe, !is_device_private_page(pages[i]));
+	i = 0;
+	for_each_sg(st->sgl, sgl, st->nents, j) {
+		unsigned long hmm_pfn, size;
+
+		hmm_pfn = range->hmm_pfns[i];
+		page = hmm_pfn_to_page(hmm_pfn);
+		xe_assert(xe, !is_device_private_page(page));
+
+		size = 1UL << hmm_pfn_to_map_order(hmm_pfn);
+		size -= page_to_pfn(page) & (size - 1);
+		i += size;
+
+		if (unlikely(j == st->nents - 1)) {
+			if (i > npages)
+				size -= (i - npages);
+			sg_mark_end(sgl);
+		}
+		sg_set_page(sgl, page, size << PAGE_SHIFT, 0);
 	}
+	xe_assert(xe, i == npages);
 
-	ret = sg_alloc_table_from_pages_segment(st, pages, npages, 0, npages << PAGE_SHIFT,
-						xe_sg_segment_size(dev), GFP_KERNEL);
-	if (ret)
-		goto free_pages;
+	return dma_map_sgtable(dev, st, write ? DMA_BIDIRECTIONAL : DMA_TO_DEVICE,
+			       DMA_ATTR_SKIP_CPU_SYNC | DMA_ATTR_NO_KERNEL_MAPPING);
+}
+
+static void xe_hmm_userptr_set_mapped(struct xe_userptr_vma *uvma)
+{
+	struct xe_userptr *userptr = &uvma->userptr;
+	struct xe_vm *vm = xe_vma_vm(&uvma->vma);
+
+	lockdep_assert_held_write(&vm->lock);
+	lockdep_assert_held(&vm->userptr.notifier_lock);
+
+	mutex_lock(&userptr->unmap_mutex);
+	xe_assert(vm->xe, !userptr->mapped);
+	userptr->mapped = true;
+	mutex_unlock(&userptr->unmap_mutex);
+}
+
+void xe_hmm_userptr_unmap(struct xe_userptr_vma *uvma)
+{
+	struct xe_userptr *userptr = &uvma->userptr;
+	struct xe_vma *vma = &uvma->vma;
+	bool write = !xe_vma_read_only(vma);
+	struct xe_vm *vm = xe_vma_vm(vma);
+	struct xe_device *xe = vm->xe;
 
-	ret = dma_map_sgtable(dev, st, write ? DMA_BIDIRECTIONAL : DMA_TO_DEVICE,
-			      DMA_ATTR_SKIP_CPU_SYNC | DMA_ATTR_NO_KERNEL_MAPPING);
-	if (ret) {
-		sg_free_table(st);
-		st = NULL;
+	if (!lockdep_is_held_type(&vm->userptr.notifier_lock, 0) &&
+	    !lockdep_is_held_type(&vm->lock, 0) &&
+	    !(vma->gpuva.flags & XE_VMA_DESTROYED)) {
+		/* Don't unmap in exec critical section. */
+		xe_vm_assert_held(vm);
+		/* Don't unmap while mapping the sg. */
+		lockdep_assert_held(&vm->lock);
 	}
 
-free_pages:
-	kvfree(pages);
-	return ret;
+	mutex_lock(&userptr->unmap_mutex);
+	if (userptr->sg && userptr->mapped)
+		dma_unmap_sgtable(xe->drm.dev, userptr->sg,
+				  write ? DMA_BIDIRECTIONAL : DMA_TO_DEVICE, 0);
+	userptr->mapped = false;
+	mutex_unlock(&userptr->unmap_mutex);
 }
 
-/*
+/**
  * xe_hmm_userptr_free_sg() - Free the scatter gather table of userptr
- *
  * @uvma: the userptr vma which hold the scatter gather table
  *
  * With function xe_userptr_populate_range, we allocate storage of
@@ -124,16 +200,9 @@ free_pages:
 void xe_hmm_userptr_free_sg(struct xe_userptr_vma *uvma)
 {
 	struct xe_userptr *userptr = &uvma->userptr;
-	struct xe_vma *vma = &uvma->vma;
-	bool write = !xe_vma_read_only(vma);
-	struct xe_vm *vm = xe_vma_vm(vma);
-	struct xe_device *xe = vm->xe;
-	struct device *dev = xe->drm.dev;
-
-	xe_assert(xe, userptr->sg);
-	dma_unmap_sgtable(dev, userptr->sg,
-			  write ? DMA_BIDIRECTIONAL : DMA_TO_DEVICE, 0);
 
+	xe_assert(xe_vma_vm(&uvma->vma)->xe, userptr->sg);
+	xe_hmm_userptr_unmap(uvma);
 	sg_free_table(userptr->sg);
 	userptr->sg = NULL;
 }
@@ -166,13 +235,20 @@ int xe_hmm_userptr_populate_range(struct xe_userptr_vma *uvma,
 {
 	unsigned long timeout =
 		jiffies + msecs_to_jiffies(HMM_RANGE_DEFAULT_TIMEOUT);
-	unsigned long *pfns, flags = HMM_PFN_REQ_FAULT;
+	unsigned long *pfns;
 	struct xe_userptr *userptr;
 	struct xe_vma *vma = &uvma->vma;
 	u64 userptr_start = xe_vma_userptr(vma);
 	u64 userptr_end = userptr_start + xe_vma_size(vma);
 	struct xe_vm *vm = xe_vma_vm(vma);
-	struct hmm_range hmm_range;
+	struct hmm_range hmm_range = {
+		.pfn_flags_mask = 0, /* ignore pfns */
+		.default_flags = HMM_PFN_REQ_FAULT,
+		.start = userptr_start,
+		.end = userptr_end,
+		.notifier = &uvma->userptr.notifier,
+		.dev_private_owner = vm->xe,
+	};
 	bool write = !xe_vma_read_only(vma);
 	unsigned long notifier_seq;
 	u64 npages;
@@ -199,19 +275,14 @@ int xe_hmm_userptr_populate_range(struct xe_userptr_vma *uvma,
 		return -ENOMEM;
 
 	if (write)
-		flags |= HMM_PFN_REQ_WRITE;
+		hmm_range.default_flags |= HMM_PFN_REQ_WRITE;
 
 	if (!mmget_not_zero(userptr->notifier.mm)) {
 		ret = -EFAULT;
 		goto free_pfns;
 	}
 
-	hmm_range.default_flags = flags;
 	hmm_range.hmm_pfns = pfns;
-	hmm_range.notifier = &userptr->notifier;
-	hmm_range.start = userptr_start;
-	hmm_range.end = userptr_end;
-	hmm_range.dev_private_owner = vm->xe;
 
 	while (true) {
 		hmm_range.notifier_seq = mmu_interval_read_begin(&userptr->notifier);
@@ -238,16 +309,37 @@ int xe_hmm_userptr_populate_range(struct xe_userptr_vma *uvma,
 	if (ret)
 		goto free_pfns;
 
-	ret = xe_build_sg(vm->xe, &hmm_range, &userptr->sgt, write);
+	ret = xe_alloc_sg(vm->xe, &userptr->sgt, &hmm_range, &vm->userptr.notifier_lock);
 	if (ret)
 		goto free_pfns;
 
+	ret = down_read_interruptible(&vm->userptr.notifier_lock);
+	if (ret)
+		goto free_st;
+
+	if (mmu_interval_read_retry(hmm_range.notifier, hmm_range.notifier_seq)) {
+		ret = -EAGAIN;
+		goto out_unlock;
+	}
+
+	ret = xe_build_sg(vm->xe, &hmm_range, &userptr->sgt,
+			  &vm->userptr.notifier_lock, write);
+	if (ret)
+		goto out_unlock;
+
 	xe_mark_range_accessed(&hmm_range, write);
 	userptr->sg = &userptr->sgt;
+	xe_hmm_userptr_set_mapped(uvma);
 	userptr->notifier_seq = hmm_range.notifier_seq;
+	up_read(&vm->userptr.notifier_lock);
+	kvfree(pfns);
+	return 0;
 
+out_unlock:
+	up_read(&vm->userptr.notifier_lock);
+free_st:
+	sg_free_table(&userptr->sgt);
 free_pfns:
 	kvfree(pfns);
 	return ret;
 }
-
diff --git a/drivers/gpu/drm/xe/xe_hmm.h b/drivers/gpu/drm/xe/xe_hmm.h
index 909dc2bdcd97..0ea98d8e7bbc 100644
--- a/drivers/gpu/drm/xe/xe_hmm.h
+++ b/drivers/gpu/drm/xe/xe_hmm.h
@@ -3,9 +3,16 @@
  * Copyright © 2024 Intel Corporation
  */
 
+#ifndef _XE_HMM_H_
+#define _XE_HMM_H_
+
 #include <linux/types.h>
 
 struct xe_userptr_vma;
 
 int xe_hmm_userptr_populate_range(struct xe_userptr_vma *uvma, bool is_mm_mmap_locked);
+
 void xe_hmm_userptr_free_sg(struct xe_userptr_vma *uvma);
+
+void xe_hmm_userptr_unmap(struct xe_userptr_vma *uvma);
+#endif
diff --git a/drivers/gpu/drm/xe/xe_pt.c b/drivers/gpu/drm/xe/xe_pt.c
index 1ddcc7e79a93..dc24baa84092 100644
--- a/drivers/gpu/drm/xe/xe_pt.c
+++ b/drivers/gpu/drm/xe/xe_pt.c
@@ -28,6 +28,8 @@ struct xe_pt_dir {
 	struct xe_pt pt;
 	/** @children: Array of page-table child nodes */
 	struct xe_ptw *children[XE_PDES];
+	/** @staging: Array of page-table staging nodes */
+	struct xe_ptw *staging[XE_PDES];
 };
 
 #if IS_ENABLED(CONFIG_DRM_XE_DEBUG_VM)
@@ -48,9 +50,10 @@ static struct xe_pt_dir *as_xe_pt_dir(struct xe_pt *pt)
 	return container_of(pt, struct xe_pt_dir, pt);
 }
 
-static struct xe_pt *xe_pt_entry(struct xe_pt_dir *pt_dir, unsigned int index)
+static struct xe_pt *
+xe_pt_entry_staging(struct xe_pt_dir *pt_dir, unsigned int index)
 {
-	return container_of(pt_dir->children[index], struct xe_pt, base);
+	return container_of(pt_dir->staging[index], struct xe_pt, base);
 }
 
 static u64 __xe_pt_empty_pte(struct xe_tile *tile, struct xe_vm *vm,
@@ -125,6 +128,7 @@ struct xe_pt *xe_pt_create(struct xe_vm *vm, struct xe_tile *tile,
 	}
 	pt->bo = bo;
 	pt->base.children = level ? as_xe_pt_dir(pt)->children : NULL;
+	pt->base.staging = level ? as_xe_pt_dir(pt)->staging : NULL;
 
 	if (vm->xef)
 		xe_drm_client_add_bo(vm->xef->client, pt->bo);
@@ -206,8 +210,8 @@ void xe_pt_destroy(struct xe_pt *pt, u32 flags, struct llist_head *deferred)
 		struct xe_pt_dir *pt_dir = as_xe_pt_dir(pt);
 
 		for (i = 0; i < XE_PDES; i++) {
-			if (xe_pt_entry(pt_dir, i))
-				xe_pt_destroy(xe_pt_entry(pt_dir, i), flags,
+			if (xe_pt_entry_staging(pt_dir, i))
+				xe_pt_destroy(xe_pt_entry_staging(pt_dir, i), flags,
 					      deferred);
 		}
 	}
@@ -376,8 +380,10 @@ xe_pt_insert_entry(struct xe_pt_stage_bind_walk *xe_walk, struct xe_pt *parent,
 		/* Continue building a non-connected subtree. */
 		struct iosys_map *map = &parent->bo->vmap;
 
-		if (unlikely(xe_child))
+		if (unlikely(xe_child)) {
 			parent->base.children[offset] = &xe_child->base;
+			parent->base.staging[offset] = &xe_child->base;
+		}
 
 		xe_pt_write(xe_walk->vm->xe, map, offset, pte);
 		parent->num_live++;
@@ -614,6 +620,7 @@ xe_pt_stage_bind(struct xe_tile *tile, struct xe_vma *vma,
 			.ops = &xe_pt_stage_bind_ops,
 			.shifts = xe_normal_pt_shifts,
 			.max_level = XE_PT_HIGHEST_LEVEL,
+			.staging = true,
 		},
 		.vm = xe_vma_vm(vma),
 		.tile = tile,
@@ -873,7 +880,7 @@ static void xe_pt_cancel_bind(struct xe_vma *vma,
 	}
 }
 
-static void xe_pt_commit_locks_assert(struct xe_vma *vma)
+static void xe_pt_commit_prepare_locks_assert(struct xe_vma *vma)
 {
 	struct xe_vm *vm = xe_vma_vm(vma);
 
@@ -885,6 +892,16 @@ static void xe_pt_commit_locks_assert(struct xe_vma *vma)
 	xe_vm_assert_held(vm);
 }
 
+static void xe_pt_commit_locks_assert(struct xe_vma *vma)
+{
+	struct xe_vm *vm = xe_vma_vm(vma);
+
+	xe_pt_commit_prepare_locks_assert(vma);
+
+	if (xe_vma_is_userptr(vma))
+		lockdep_assert_held_read(&vm->userptr.notifier_lock);
+}
+
 static void xe_pt_commit(struct xe_vma *vma,
 			 struct xe_vm_pgtable_update *entries,
 			 u32 num_entries, struct llist_head *deferred)
@@ -895,13 +912,17 @@ static void xe_pt_commit(struct xe_vma *vma,
 
 	for (i = 0; i < num_entries; i++) {
 		struct xe_pt *pt = entries[i].pt;
+		struct xe_pt_dir *pt_dir;
 
 		if (!pt->level)
 			continue;
 
+		pt_dir = as_xe_pt_dir(pt);
 		for (j = 0; j < entries[i].qwords; j++) {
 			struct xe_pt *oldpte = entries[i].pt_entries[j].pt;
+			int j_ = j + entries[i].ofs;
 
+			pt_dir->children[j_] = pt_dir->staging[j_];
 			xe_pt_destroy(oldpte, xe_vma_vm(vma)->flags, deferred);
 		}
 	}
@@ -913,7 +934,7 @@ static void xe_pt_abort_bind(struct xe_vma *vma,
 {
 	int i, j;
 
-	xe_pt_commit_locks_assert(vma);
+	xe_pt_commit_prepare_locks_assert(vma);
 
 	for (i = num_entries - 1; i >= 0; --i) {
 		struct xe_pt *pt = entries[i].pt;
@@ -928,10 +949,10 @@ static void xe_pt_abort_bind(struct xe_vma *vma,
 		pt_dir = as_xe_pt_dir(pt);
 		for (j = 0; j < entries[i].qwords; j++) {
 			u32 j_ = j + entries[i].ofs;
-			struct xe_pt *newpte = xe_pt_entry(pt_dir, j_);
+			struct xe_pt *newpte = xe_pt_entry_staging(pt_dir, j_);
 			struct xe_pt *oldpte = entries[i].pt_entries[j].pt;
 
-			pt_dir->children[j_] = oldpte ? &oldpte->base : 0;
+			pt_dir->staging[j_] = oldpte ? &oldpte->base : 0;
 			xe_pt_destroy(newpte, xe_vma_vm(vma)->flags, NULL);
 		}
 	}
@@ -943,7 +964,7 @@ static void xe_pt_commit_prepare_bind(struct xe_vma *vma,
 {
 	u32 i, j;
 
-	xe_pt_commit_locks_assert(vma);
+	xe_pt_commit_prepare_locks_assert(vma);
 
 	for (i = 0; i < num_entries; i++) {
 		struct xe_pt *pt = entries[i].pt;
@@ -961,10 +982,10 @@ static void xe_pt_commit_prepare_bind(struct xe_vma *vma,
 			struct xe_pt *newpte = entries[i].pt_entries[j].pt;
 			struct xe_pt *oldpte = NULL;
 
-			if (xe_pt_entry(pt_dir, j_))
-				oldpte = xe_pt_entry(pt_dir, j_);
+			if (xe_pt_entry_staging(pt_dir, j_))
+				oldpte = xe_pt_entry_staging(pt_dir, j_);
 
-			pt_dir->children[j_] = &newpte->base;
+			pt_dir->staging[j_] = &newpte->base;
 			entries[i].pt_entries[j].pt = oldpte;
 		}
 	}
@@ -1213,42 +1234,22 @@ static int vma_check_userptr(struct xe_vm *vm, struct xe_vma *vma,
 		return 0;
 
 	uvma = to_userptr_vma(vma);
-	notifier_seq = uvma->userptr.notifier_seq;
+	if (xe_pt_userptr_inject_eagain(uvma))
+		xe_vma_userptr_force_invalidate(uvma);
 
-	if (uvma->userptr.initial_bind && !xe_vm_in_fault_mode(vm))
-		return 0;
+	notifier_seq = uvma->userptr.notifier_seq;
 
 	if (!mmu_interval_read_retry(&uvma->userptr.notifier,
-				     notifier_seq) &&
-	    !xe_pt_userptr_inject_eagain(uvma))
+				     notifier_seq))
 		return 0;
 
-	if (xe_vm_in_fault_mode(vm)) {
+	if (xe_vm_in_fault_mode(vm))
 		return -EAGAIN;
-	} else {
-		spin_lock(&vm->userptr.invalidated_lock);
-		list_move_tail(&uvma->userptr.invalidate_link,
-			       &vm->userptr.invalidated);
-		spin_unlock(&vm->userptr.invalidated_lock);
-
-		if (xe_vm_in_preempt_fence_mode(vm)) {
-			struct dma_resv_iter cursor;
-			struct dma_fence *fence;
-			long err;
-
-			dma_resv_iter_begin(&cursor, xe_vm_resv(vm),
-					    DMA_RESV_USAGE_BOOKKEEP);
-			dma_resv_for_each_fence_unlocked(&cursor, fence)
-				dma_fence_enable_sw_signaling(fence);
-			dma_resv_iter_end(&cursor);
-
-			err = dma_resv_wait_timeout(xe_vm_resv(vm),
-						    DMA_RESV_USAGE_BOOKKEEP,
-						    false, MAX_SCHEDULE_TIMEOUT);
-			XE_WARN_ON(err <= 0);
-		}
-	}
 
+	/*
+	 * Just continue the operation since exec or rebind worker
+	 * will take care of rebinding.
+	 */
 	return 0;
 }
 
@@ -1514,6 +1515,7 @@ static unsigned int xe_pt_stage_unbind(struct xe_tile *tile, struct xe_vma *vma,
 			.ops = &xe_pt_stage_unbind_ops,
 			.shifts = xe_normal_pt_shifts,
 			.max_level = XE_PT_HIGHEST_LEVEL,
+			.staging = true,
 		},
 		.tile = tile,
 		.modified_start = xe_vma_start(vma),
@@ -1555,7 +1557,7 @@ static void xe_pt_abort_unbind(struct xe_vma *vma,
 {
 	int i, j;
 
-	xe_pt_commit_locks_assert(vma);
+	xe_pt_commit_prepare_locks_assert(vma);
 
 	for (i = num_entries - 1; i >= 0; --i) {
 		struct xe_vm_pgtable_update *entry = &entries[i];
@@ -1568,7 +1570,7 @@ static void xe_pt_abort_unbind(struct xe_vma *vma,
 			continue;
 
 		for (j = entry->ofs; j < entry->ofs + entry->qwords; j++)
-			pt_dir->children[j] =
+			pt_dir->staging[j] =
 				entries[i].pt_entries[j - entry->ofs].pt ?
 				&entries[i].pt_entries[j - entry->ofs].pt->base : NULL;
 	}
@@ -1581,7 +1583,7 @@ xe_pt_commit_prepare_unbind(struct xe_vma *vma,
 {
 	int i, j;
 
-	xe_pt_commit_locks_assert(vma);
+	xe_pt_commit_prepare_locks_assert(vma);
 
 	for (i = 0; i < num_entries; ++i) {
 		struct xe_vm_pgtable_update *entry = &entries[i];
@@ -1595,8 +1597,8 @@ xe_pt_commit_prepare_unbind(struct xe_vma *vma,
 		pt_dir = as_xe_pt_dir(pt);
 		for (j = entry->ofs; j < entry->ofs + entry->qwords; j++) {
 			entry->pt_entries[j - entry->ofs].pt =
-				xe_pt_entry(pt_dir, j);
-			pt_dir->children[j] = NULL;
+				xe_pt_entry_staging(pt_dir, j);
+			pt_dir->staging[j] = NULL;
 		}
 	}
 }
diff --git a/drivers/gpu/drm/xe/xe_pt_walk.c b/drivers/gpu/drm/xe/xe_pt_walk.c
index b8b3d2aea492..be602a763ff3 100644
--- a/drivers/gpu/drm/xe/xe_pt_walk.c
+++ b/drivers/gpu/drm/xe/xe_pt_walk.c
@@ -74,7 +74,8 @@ int xe_pt_walk_range(struct xe_ptw *parent, unsigned int level,
 		     u64 addr, u64 end, struct xe_pt_walk *walk)
 {
 	pgoff_t offset = xe_pt_offset(addr, level, walk);
-	struct xe_ptw **entries = parent->children ? parent->children : NULL;
+	struct xe_ptw **entries = walk->staging ? (parent->staging ?: NULL) :
+		(parent->children ?: NULL);
 	const struct xe_pt_walk_ops *ops = walk->ops;
 	enum page_walk_action action;
 	struct xe_ptw *child;
diff --git a/drivers/gpu/drm/xe/xe_pt_walk.h b/drivers/gpu/drm/xe/xe_pt_walk.h
index 5ecc4d2f0f65..5c02c244f7de 100644
--- a/drivers/gpu/drm/xe/xe_pt_walk.h
+++ b/drivers/gpu/drm/xe/xe_pt_walk.h
@@ -11,12 +11,14 @@
 /**
  * struct xe_ptw - base class for driver pagetable subclassing.
  * @children: Pointer to an array of children if any.
+ * @staging: Pointer to an array of staging if any.
  *
  * Drivers could subclass this, and if it's a page-directory, typically
  * embed an array of xe_ptw pointers.
  */
 struct xe_ptw {
 	struct xe_ptw **children;
+	struct xe_ptw **staging;
 };
 
 /**
@@ -41,6 +43,8 @@ struct xe_pt_walk {
 	 * as shared pagetables.
 	 */
 	bool shared_pt_mode;
+	/** @staging: Walk staging PT structure */
+	bool staging;
 };
 
 /**
diff --git a/drivers/gpu/drm/xe/xe_vm.c b/drivers/gpu/drm/xe/xe_vm.c
index 30259eba450b..ec6ec18ab3fa 100644
--- a/drivers/gpu/drm/xe/xe_vm.c
+++ b/drivers/gpu/drm/xe/xe_vm.c
@@ -579,51 +579,26 @@ out_unlock_outer:
 	trace_xe_vm_rebind_worker_exit(vm);
 }
 
-static bool vma_userptr_invalidate(struct mmu_interval_notifier *mni,
-				   const struct mmu_notifier_range *range,
-				   unsigned long cur_seq)
+static void __vma_userptr_invalidate(struct xe_vm *vm, struct xe_userptr_vma *uvma)
 {
-	struct xe_userptr *userptr = container_of(mni, typeof(*userptr), notifier);
-	struct xe_userptr_vma *uvma = container_of(userptr, typeof(*uvma), userptr);
+	struct xe_userptr *userptr = &uvma->userptr;
 	struct xe_vma *vma = &uvma->vma;
-	struct xe_vm *vm = xe_vma_vm(vma);
 	struct dma_resv_iter cursor;
 	struct dma_fence *fence;
 	long err;
 
-	xe_assert(vm->xe, xe_vma_is_userptr(vma));
-	trace_xe_vma_userptr_invalidate(vma);
-
-	if (!mmu_notifier_range_blockable(range))
-		return false;
-
-	vm_dbg(&xe_vma_vm(vma)->xe->drm,
-	       "NOTIFIER: addr=0x%016llx, range=0x%016llx",
-		xe_vma_start(vma), xe_vma_size(vma));
-
-	down_write(&vm->userptr.notifier_lock);
-	mmu_interval_set_seq(mni, cur_seq);
-
-	/* No need to stop gpu access if the userptr is not yet bound. */
-	if (!userptr->initial_bind) {
-		up_write(&vm->userptr.notifier_lock);
-		return true;
-	}
-
 	/*
 	 * Tell exec and rebind worker they need to repin and rebind this
 	 * userptr.
 	 */
 	if (!xe_vm_in_fault_mode(vm) &&
-	    !(vma->gpuva.flags & XE_VMA_DESTROYED) && vma->tile_present) {
+	    !(vma->gpuva.flags & XE_VMA_DESTROYED)) {
 		spin_lock(&vm->userptr.invalidated_lock);
 		list_move_tail(&userptr->invalidate_link,
 			       &vm->userptr.invalidated);
 		spin_unlock(&vm->userptr.invalidated_lock);
 	}
 
-	up_write(&vm->userptr.notifier_lock);
-
 	/*
 	 * Preempt fences turn into schedule disables, pipeline these.
 	 * Note that even in fault mode, we need to wait for binds and
@@ -641,11 +616,37 @@ static bool vma_userptr_invalidate(struct mmu_interval_notifier *mni,
 				    false, MAX_SCHEDULE_TIMEOUT);
 	XE_WARN_ON(err <= 0);
 
-	if (xe_vm_in_fault_mode(vm)) {
+	if (xe_vm_in_fault_mode(vm) && userptr->initial_bind) {
 		err = xe_vm_invalidate_vma(vma);
 		XE_WARN_ON(err);
 	}
 
+	xe_hmm_userptr_unmap(uvma);
+}
+
+static bool vma_userptr_invalidate(struct mmu_interval_notifier *mni,
+				   const struct mmu_notifier_range *range,
+				   unsigned long cur_seq)
+{
+	struct xe_userptr_vma *uvma = container_of(mni, typeof(*uvma), userptr.notifier);
+	struct xe_vma *vma = &uvma->vma;
+	struct xe_vm *vm = xe_vma_vm(vma);
+
+	xe_assert(vm->xe, xe_vma_is_userptr(vma));
+	trace_xe_vma_userptr_invalidate(vma);
+
+	if (!mmu_notifier_range_blockable(range))
+		return false;
+
+	vm_dbg(&xe_vma_vm(vma)->xe->drm,
+	       "NOTIFIER: addr=0x%016llx, range=0x%016llx",
+		xe_vma_start(vma), xe_vma_size(vma));
+
+	down_write(&vm->userptr.notifier_lock);
+	mmu_interval_set_seq(mni, cur_seq);
+
+	__vma_userptr_invalidate(vm, uvma);
+	up_write(&vm->userptr.notifier_lock);
 	trace_xe_vma_userptr_invalidate_complete(vma);
 
 	return true;
@@ -655,6 +656,34 @@ static const struct mmu_interval_notifier_ops vma_userptr_notifier_ops = {
 	.invalidate = vma_userptr_invalidate,
 };
 
+#if IS_ENABLED(CONFIG_DRM_XE_USERPTR_INVAL_INJECT)
+/**
+ * xe_vma_userptr_force_invalidate() - force invalidate a userptr
+ * @uvma: The userptr vma to invalidate
+ *
+ * Perform a forced userptr invalidation for testing purposes.
+ */
+void xe_vma_userptr_force_invalidate(struct xe_userptr_vma *uvma)
+{
+	struct xe_vm *vm = xe_vma_vm(&uvma->vma);
+
+	/* Protect against concurrent userptr pinning */
+	lockdep_assert_held(&vm->lock);
+	/* Protect against concurrent notifiers */
+	lockdep_assert_held(&vm->userptr.notifier_lock);
+	/*
+	 * Protect against concurrent instances of this function and
+	 * the critical exec sections
+	 */
+	xe_vm_assert_held(vm);
+
+	if (!mmu_interval_read_retry(&uvma->userptr.notifier,
+				     uvma->userptr.notifier_seq))
+		uvma->userptr.notifier_seq -= 2;
+	__vma_userptr_invalidate(vm, uvma);
+}
+#endif
+
 int xe_vm_userptr_pin(struct xe_vm *vm)
 {
 	struct xe_userptr_vma *uvma, *next;
@@ -1012,6 +1041,7 @@ static struct xe_vma *xe_vma_create(struct xe_vm *vm,
 			INIT_LIST_HEAD(&userptr->invalidate_link);
 			INIT_LIST_HEAD(&userptr->repin_link);
 			vma->gpuva.gem.offset = bo_offset_or_userptr;
+			mutex_init(&userptr->unmap_mutex);
 
 			err = mmu_interval_notifier_insert(&userptr->notifier,
 							   current->mm,
@@ -1053,6 +1083,7 @@ static void xe_vma_destroy_late(struct xe_vma *vma)
 		 * them anymore
 		 */
 		mmu_interval_notifier_remove(&userptr->notifier);
+		mutex_destroy(&userptr->unmap_mutex);
 		xe_vm_put(vm);
 	} else if (xe_vma_is_null(vma)) {
 		xe_vm_put(vm);
@@ -2286,8 +2317,17 @@ static int vm_bind_ioctl_ops_parse(struct xe_vm *vm, struct drm_gpuva_ops *ops,
 			break;
 		}
 		case DRM_GPUVA_OP_UNMAP:
+			xe_vma_ops_incr_pt_update_ops(vops, op->tile_mask);
+			break;
 		case DRM_GPUVA_OP_PREFETCH:
-			/* FIXME: Need to skip some prefetch ops */
+			vma = gpuva_to_vma(op->base.prefetch.va);
+
+			if (xe_vma_is_userptr(vma)) {
+				err = xe_vma_userptr_pin_pages(to_userptr_vma(vma));
+				if (err)
+					return err;
+			}
+
 			xe_vma_ops_incr_pt_update_ops(vops, op->tile_mask);
 			break;
 		default:
diff --git a/drivers/gpu/drm/xe/xe_vm.h b/drivers/gpu/drm/xe/xe_vm.h
index 23adb7442881..b882bfb31bd0 100644
--- a/drivers/gpu/drm/xe/xe_vm.h
+++ b/drivers/gpu/drm/xe/xe_vm.h
@@ -274,9 +274,17 @@ static inline void vm_dbg(const struct drm_device *dev,
 			  const char *format, ...)
 { /* noop */ }
 #endif
-#endif
 
 struct xe_vm_snapshot *xe_vm_snapshot_capture(struct xe_vm *vm);
 void xe_vm_snapshot_capture_delayed(struct xe_vm_snapshot *snap);
 void xe_vm_snapshot_print(struct xe_vm_snapshot *snap, struct drm_printer *p);
 void xe_vm_snapshot_free(struct xe_vm_snapshot *snap);
+
+#if IS_ENABLED(CONFIG_DRM_XE_USERPTR_INVAL_INJECT)
+void xe_vma_userptr_force_invalidate(struct xe_userptr_vma *uvma);
+#else
+static inline void xe_vma_userptr_force_invalidate(struct xe_userptr_vma *uvma)
+{
+}
+#endif
+#endif
diff --git a/drivers/gpu/drm/xe/xe_vm_types.h b/drivers/gpu/drm/xe/xe_vm_types.h
index 7f9a303e51d8..a4b4091cfd0d 100644
--- a/drivers/gpu/drm/xe/xe_vm_types.h
+++ b/drivers/gpu/drm/xe/xe_vm_types.h
@@ -59,12 +59,16 @@ struct xe_userptr {
 	struct sg_table *sg;
 	/** @notifier_seq: notifier sequence number */
 	unsigned long notifier_seq;
+	/** @unmap_mutex: Mutex protecting dma-unmapping */
+	struct mutex unmap_mutex;
 	/**
 	 * @initial_bind: user pointer has been bound at least once.
 	 * write: vm->userptr.notifier_lock in read mode and vm->resv held.
 	 * read: vm->userptr.notifier_lock in write mode or vm->resv held.
 	 */
 	bool initial_bind;
+	/** @mapped: Whether the @sgt sg-table is dma-mapped. Protected by @unmap_mutex. */
+	bool mapped;
 #if IS_ENABLED(CONFIG_DRM_XE_USERPTR_INVAL_INJECT)
 	u32 divisor;
 #endif
@@ -227,8 +231,8 @@ struct xe_vm {
 		 * up for revalidation. Protected from access with the
 		 * @invalidated_lock. Removing items from the list
 		 * additionally requires @lock in write mode, and adding
-		 * items to the list requires the @userptr.notifer_lock in
-		 * write mode.
+		 * items to the list requires either the @userptr.notifer_lock in
+		 * write mode, OR @lock in write mode.
 		 */
 		struct list_head invalidated;
 	} userptr;