1 files changed, 110 insertions, 46 deletions

diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_svm.c b/drivers/gpu/drm/amd/amdkfd/kfd_svm.c
index b691c8495d66..3cb4681c5f53 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_svm.c
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_svm.c
@@ -1496,9 +1496,11 @@ static int svm_range_validate_and_map(struct mm_struct *mm,
 
 		next = min(vma->vm_end, end);
 		npages = (next - addr) >> PAGE_SHIFT;
+		WRITE_ONCE(p->svms.faulting_task, current);
 		r = amdgpu_hmm_range_get_pages(&prange->notifier, mm, NULL,
 					       addr, npages, &hmm_range,
 					       readonly, true, owner);
+		WRITE_ONCE(p->svms.faulting_task, NULL);
 		if (r) {
 			pr_debug("failed %d to get svm range pages\n", r);
 			goto unreserve_out;
@@ -1572,7 +1574,6 @@ retry_flush_work:
 static void svm_range_restore_work(struct work_struct *work)
 {
 	struct delayed_work *dwork = to_delayed_work(work);
-	struct amdkfd_process_info *process_info;
 	struct svm_range_list *svms;
 	struct svm_range *prange;
 	struct kfd_process *p;
@@ -1592,12 +1593,10 @@ static void svm_range_restore_work(struct work_struct *work)
 	 * the lifetime of this thread, kfd_process and mm will be valid.
 	 */
 	p = container_of(svms, struct kfd_process, svms);
-	process_info = p->kgd_process_info;
 	mm = p->mm;
 	if (!mm)
 		return;
 
-	mutex_lock(&process_info->lock);
 	svm_range_list_lock_and_flush_work(svms, mm);
 	mutex_lock(&svms->lock);
 
@@ -1650,7 +1649,6 @@ static void svm_range_restore_work(struct work_struct *work)
 out_reschedule:
 	mutex_unlock(&svms->lock);
 	mmap_write_unlock(mm);
-	mutex_unlock(&process_info->lock);
 
 	/* If validation failed, reschedule another attempt */
 	if (evicted_ranges) {
@@ -1966,10 +1964,16 @@ static void svm_range_drain_retry_fault(struct svm_range_list *svms)
 	struct kfd_process_device *pdd;
 	struct amdgpu_device *adev;
 	struct kfd_process *p;
+	int drain;
 	uint32_t i;
 
 	p = container_of(svms, struct kfd_process, svms);
 
+restart:
+	drain = atomic_read(&svms->drain_pagefaults);
+	if (!drain)
+		return;
+
 	for_each_set_bit(i, svms->bitmap_supported, p->n_pdds) {
 		pdd = p->pdds[i];
 		if (!pdd)
@@ -1981,6 +1985,8 @@ static void svm_range_drain_retry_fault(struct svm_range_list *svms)
 		amdgpu_ih_wait_on_checkpoint_process(adev, &adev->irq.ih1);
 		pr_debug("drain retry fault gpu %d svms 0x%p done\n", i, svms);
 	}
+	if (atomic_cmpxchg(&svms->drain_pagefaults, drain, 0) != drain)
+		goto restart;
 }
 
 static void svm_range_deferred_list_work(struct work_struct *work)
@@ -1988,35 +1994,41 @@ static void svm_range_deferred_list_work(struct work_struct *work)
 	struct svm_range_list *svms;
 	struct svm_range *prange;
 	struct mm_struct *mm;
+	struct kfd_process *p;
 
 	svms = container_of(work, struct svm_range_list, deferred_list_work);
 	pr_debug("enter svms 0x%p\n", svms);
 
+	p = container_of(svms, struct kfd_process, svms);
+	/* Avoid mm is gone when inserting mmu notifier */
+	mm = get_task_mm(p->lead_thread);
+	if (!mm) {
+		pr_debug("svms 0x%p process mm gone\n", svms);
+		return;
+	}
+retry:
+	mmap_write_lock(mm);
+
+	/* Checking for the need to drain retry faults must be inside
+	 * mmap write lock to serialize with munmap notifiers.
+	 */
+	if (unlikely(atomic_read(&svms->drain_pagefaults))) {
+		mmap_write_unlock(mm);
+		svm_range_drain_retry_fault(svms);
+		goto retry;
+	}
+
 	spin_lock(&svms->deferred_list_lock);
 	while (!list_empty(&svms->deferred_range_list)) {
 		prange = list_first_entry(&svms->deferred_range_list,
 					  struct svm_range, deferred_list);
+		list_del_init(&prange->deferred_list);
 		spin_unlock(&svms->deferred_list_lock);
+
 		pr_debug("prange 0x%p [0x%lx 0x%lx] op %d\n", prange,
 			 prange->start, prange->last, prange->work_item.op);
 
-		/* Make sure no stale retry fault coming after range is freed */
-		if (prange->work_item.op == SVM_OP_UNMAP_RANGE)
-			svm_range_drain_retry_fault(prange->svms);
-
-		mm = prange->work_item.mm;
-		mmap_write_lock(mm);
 		mutex_lock(&svms->lock);
-
-		/* Remove from deferred_list must be inside mmap write lock,
-		 * otherwise, svm_range_list_lock_and_flush_work may hold mmap
-		 * write lock, and continue because deferred_list is empty, then
-		 * deferred_list handle is blocked by mmap write lock.
-		 */
-		spin_lock(&svms->deferred_list_lock);
-		list_del_init(&prange->deferred_list);
-		spin_unlock(&svms->deferred_list_lock);
-
 		mutex_lock(&prange->migrate_mutex);
 		while (!list_empty(&prange->child_list)) {
 			struct svm_range *pchild;
@@ -2032,12 +2044,13 @@ static void svm_range_deferred_list_work(struct work_struct *work)
 
 		svm_range_handle_list_op(svms, prange);
 		mutex_unlock(&svms->lock);
-		mmap_write_unlock(mm);
 
 		spin_lock(&svms->deferred_list_lock);
 	}
 	spin_unlock(&svms->deferred_list_lock);
 
+	mmap_write_unlock(mm);
+	mmput(mm);
 	pr_debug("exit svms 0x%p\n", svms);
 }
 
@@ -2124,6 +2137,12 @@ svm_range_unmap_from_cpu(struct mm_struct *mm, struct svm_range *prange,
 	pr_debug("svms 0x%p prange 0x%p [0x%lx 0x%lx] [0x%lx 0x%lx]\n", svms,
 		 prange, prange->start, prange->last, start, last);
 
+	/* Make sure pending page faults are drained in the deferred worker
+	 * before the range is freed to avoid straggler interrupts on
+	 * unmapped memory causing "phantom faults".
+	 */
+	atomic_inc(&svms->drain_pagefaults);
+
 	unmap_parent = start <= prange->start && last >= prange->last;
 
 	list_for_each_entry(pchild, &prange->child_list, child_list) {
@@ -2261,7 +2280,7 @@ svm_range_from_addr(struct svm_range_list *svms, unsigned long addr,
  * migration if actual loc is not best location, then update GPU page table
  * mapping to the best location.
  *
- * If vm fault gpu is range preferred loc, the best_loc is preferred loc.
+ * If the preferred loc is accessible by faulting GPU, use preferred loc.
  * If vm fault gpu idx is on range ACCESSIBLE bitmap, best_loc is vm fault gpu
  * If vm fault gpu idx is on range ACCESSIBLE_IN_PLACE bitmap, then
  *    if range actual loc is cpu, best_loc is cpu
@@ -2278,7 +2297,7 @@ svm_range_best_restore_location(struct svm_range *prange,
 				struct amdgpu_device *adev,
 				int32_t *gpuidx)
 {
-	struct amdgpu_device *bo_adev;
+	struct amdgpu_device *bo_adev, *preferred_adev;
 	struct kfd_process *p;
 	uint32_t gpuid;
 	int r;
@@ -2291,8 +2310,16 @@ svm_range_best_restore_location(struct svm_range *prange,
 		return -1;
 	}
 
-	if (prange->preferred_loc == gpuid)
+	if (prange->preferred_loc == gpuid ||
+	    prange->preferred_loc == KFD_IOCTL_SVM_LOCATION_SYSMEM) {
 		return prange->preferred_loc;
+	} else if (prange->preferred_loc != KFD_IOCTL_SVM_LOCATION_UNDEFINED) {
+		preferred_adev = svm_range_get_adev_by_id(prange,
+							prange->preferred_loc);
+		if (amdgpu_xgmi_same_hive(adev, preferred_adev))
+			return prange->preferred_loc;
+		/* fall through */
+	}
 
 	if (test_bit(*gpuidx, prange->bitmap_access))
 		return gpuid;
@@ -2313,7 +2340,8 @@ svm_range_best_restore_location(struct svm_range *prange,
 
 static int
 svm_range_get_range_boundaries(struct kfd_process *p, int64_t addr,
-				unsigned long *start, unsigned long *last)
+			       unsigned long *start, unsigned long *last,
+			       bool *is_heap_stack)
 {
 	struct vm_area_struct *vma;
 	struct interval_tree_node *node;
@@ -2324,6 +2352,12 @@ svm_range_get_range_boundaries(struct kfd_process *p, int64_t addr,
 		pr_debug("VMA does not exist in address [0x%llx]\n", addr);
 		return -EFAULT;
 	}
+
+	*is_heap_stack = (vma->vm_start <= vma->vm_mm->brk &&
+			  vma->vm_end >= vma->vm_mm->start_brk) ||
+			 (vma->vm_start <= vma->vm_mm->start_stack &&
+			  vma->vm_end >= vma->vm_mm->start_stack);
+
 	start_limit = max(vma->vm_start >> PAGE_SHIFT,
 		      (unsigned long)ALIGN_DOWN(addr, 2UL << 8));
 	end_limit = min(vma->vm_end >> PAGE_SHIFT,
@@ -2353,9 +2387,9 @@ svm_range_get_range_boundaries(struct kfd_process *p, int64_t addr,
 	*start = start_limit;
 	*last = end_limit - 1;
 
-	pr_debug("vma start: 0x%lx start: 0x%lx vma end: 0x%lx last: 0x%lx\n",
-		  vma->vm_start >> PAGE_SHIFT, *start,
-		  vma->vm_end >> PAGE_SHIFT, *last);
+	pr_debug("vma [0x%lx 0x%lx] range [0x%lx 0x%lx] is_heap_stack %d\n",
+		 vma->vm_start >> PAGE_SHIFT, vma->vm_end >> PAGE_SHIFT,
+		 *start, *last, *is_heap_stack);
 
 	return 0;
 }
@@ -2420,11 +2454,13 @@ svm_range *svm_range_create_unregistered_range(struct amdgpu_device *adev,
 	struct svm_range *prange = NULL;
 	unsigned long start, last;
 	uint32_t gpuid, gpuidx;
+	bool is_heap_stack;
 	uint64_t bo_s = 0;
 	uint64_t bo_l = 0;
 	int r;
 
-	if (svm_range_get_range_boundaries(p, addr, &start, &last))
+	if (svm_range_get_range_boundaries(p, addr, &start, &last,
+					   &is_heap_stack))
 		return NULL;
 
 	r = svm_range_check_vm(p, start, last, &bo_s, &bo_l);
@@ -2451,6 +2487,9 @@ svm_range *svm_range_create_unregistered_range(struct amdgpu_device *adev,
 		return NULL;
 	}
 
+	if (is_heap_stack)
+		prange->preferred_loc = KFD_IOCTL_SVM_LOCATION_SYSMEM;
+
 	svm_range_add_to_svms(prange);
 	svm_range_add_notifier_locked(mm, prange);
 
@@ -2523,20 +2562,13 @@ svm_range_count_fault(struct amdgpu_device *adev, struct kfd_process *p,
 }
 
 static bool
-svm_fault_allowed(struct mm_struct *mm, uint64_t addr, bool write_fault)
+svm_fault_allowed(struct vm_area_struct *vma, bool write_fault)
 {
 	unsigned long requested = VM_READ;
-	struct vm_area_struct *vma;
 
 	if (write_fault)
 		requested |= VM_WRITE;
 
-	vma = find_vma(mm, addr << PAGE_SHIFT);
-	if (!vma || (addr << PAGE_SHIFT) < vma->vm_start) {
-		pr_debug("address 0x%llx VMA is removed\n", addr);
-		return true;
-	}
-
 	pr_debug("requested 0x%lx, vma permission flags 0x%lx\n", requested,
 		vma->vm_flags);
 	return (vma->vm_flags & requested) == requested;
@@ -2554,6 +2586,7 @@ svm_range_restore_pages(struct amdgpu_device *adev, unsigned int pasid,
 	int32_t best_loc;
 	int32_t gpuidx = MAX_GPU_INSTANCE;
 	bool write_locked = false;
+	struct vm_area_struct *vma;
 	int r = 0;
 
 	if (!KFD_IS_SVM_API_SUPPORTED(adev->kfd.dev)) {
@@ -2564,7 +2597,7 @@ svm_range_restore_pages(struct amdgpu_device *adev, unsigned int pasid,
 	p = kfd_lookup_process_by_pasid(pasid);
 	if (!p) {
 		pr_debug("kfd process not founded pasid 0x%x\n", pasid);
-		return -ESRCH;
+		return 0;
 	}
 	if (!p->xnack_enabled) {
 		pr_debug("XNACK not enabled for pasid 0x%x\n", pasid);
@@ -2575,10 +2608,19 @@ svm_range_restore_pages(struct amdgpu_device *adev, unsigned int pasid,
 
 	pr_debug("restoring svms 0x%p fault address 0x%llx\n", svms, addr);
 
+	if (atomic_read(&svms->drain_pagefaults)) {
+		pr_debug("draining retry fault, drop fault 0x%llx\n", addr);
+		r = 0;
+		goto out;
+	}
+
+	/* p->lead_thread is available as kfd_process_wq_release flush the work
+	 * before releasing task ref.
+	 */
 	mm = get_task_mm(p->lead_thread);
 	if (!mm) {
 		pr_debug("svms 0x%p failed to get mm\n", svms);
-		r = -ESRCH;
+		r = 0;
 		goto out;
 	}
 
@@ -2616,6 +2658,7 @@ retry_write_locked:
 
 	if (svm_range_skip_recover(prange)) {
 		amdgpu_gmc_filter_faults_remove(adev, addr, pasid);
+		r = 0;
 		goto out_unlock_range;
 	}
 
@@ -2624,10 +2667,21 @@ retry_write_locked:
 	if (timestamp < AMDGPU_SVM_RANGE_RETRY_FAULT_PENDING) {
 		pr_debug("svms 0x%p [0x%lx %lx] already restored\n",
 			 svms, prange->start, prange->last);
+		r = 0;
+		goto out_unlock_range;
+	}
+
+	/* __do_munmap removed VMA, return success as we are handling stale
+	 * retry fault.
+	 */
+	vma = find_vma(mm, addr << PAGE_SHIFT);
+	if (!vma || (addr << PAGE_SHIFT) < vma->vm_start) {
+		pr_debug("address 0x%llx VMA is removed\n", addr);
+		r = 0;
 		goto out_unlock_range;
 	}
 
-	if (!svm_fault_allowed(mm, addr, write_fault)) {
+	if (!svm_fault_allowed(vma, write_fault)) {
 		pr_debug("fault addr 0x%llx no %s permission\n", addr,
 			write_fault ? "write" : "read");
 		r = -EPERM;
@@ -2705,6 +2759,14 @@ void svm_range_list_fini(struct kfd_process *p)
 	/* Ensure list work is finished before process is destroyed */
 	flush_work(&p->svms.deferred_list_work);
 
+	/*
+	 * Ensure no retry fault comes in afterwards, as page fault handler will
+	 * not find kfd process and take mm lock to recover fault.
+	 */
+	atomic_inc(&p->svms.drain_pagefaults);
+	svm_range_drain_retry_fault(&p->svms);
+
+
 	list_for_each_entry_safe(prange, next, &p->svms.list, list) {
 		svm_range_unlink(prange);
 		svm_range_remove_notifier(prange);
@@ -2725,6 +2787,7 @@ int svm_range_list_init(struct kfd_process *p)
 	mutex_init(&svms->lock);
 	INIT_LIST_HEAD(&svms->list);
 	atomic_set(&svms->evicted_ranges, 0);
+	atomic_set(&svms->drain_pagefaults, 0);
 	INIT_DELAYED_WORK(&svms->restore_work, svm_range_restore_work);
 	INIT_WORK(&svms->deferred_list_work, svm_range_deferred_list_work);
 	INIT_LIST_HEAD(&svms->deferred_range_list);
@@ -3076,6 +3139,8 @@ static void svm_range_evict_svm_bo_worker(struct work_struct *work)
 		struct svm_range *prange =
 				list_first_entry(&svm_bo->range_list,
 						struct svm_range, svm_bo_list);
+		int retries = 3;
+
 		list_del_init(&prange->svm_bo_list);
 		spin_unlock(&svm_bo->list_lock);
 
@@ -3083,7 +3148,11 @@ static void svm_range_evict_svm_bo_worker(struct work_struct *work)
 			 prange->start, prange->last);
 
 		mutex_lock(&prange->migrate_mutex);
-		svm_migrate_vram_to_ram(prange, svm_bo->eviction_fence->mm);
+		do {
+			svm_migrate_vram_to_ram(prange,
+						svm_bo->eviction_fence->mm);
+		} while (prange->actual_loc && --retries);
+		WARN(prange->actual_loc, "Migration failed during eviction");
 
 		mutex_lock(&prange->lock);
 		prange->svm_bo = NULL;
@@ -3108,7 +3177,6 @@ static int
 svm_range_set_attr(struct kfd_process *p, uint64_t start, uint64_t size,
 		   uint32_t nattr, struct kfd_ioctl_svm_attribute *attrs)
 {
-	struct amdkfd_process_info *process_info = p->kgd_process_info;
 	struct mm_struct *mm = current->mm;
 	struct list_head update_list;
 	struct list_head insert_list;
@@ -3127,8 +3195,6 @@ svm_range_set_attr(struct kfd_process *p, uint64_t start, uint64_t size,
 
 	svms = &p->svms;
 
-	mutex_lock(&process_info->lock);
-
 	svm_range_list_lock_and_flush_work(svms, mm);
 
 	r = svm_range_is_valid(p, start, size);
@@ -3204,8 +3270,6 @@ out_unlock_range:
 	mutex_unlock(&svms->lock);
 	mmap_read_unlock(mm);
 out:
-	mutex_unlock(&process_info->lock);
-
 	pr_debug("pasid 0x%x svms 0x%p [0x%llx 0x%llx] done, r=%d\n", p->pasid,
 		 &p->svms, start, start + size - 1, r);