1 files changed, 69 insertions, 42 deletions

diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_svm.c b/drivers/gpu/drm/amd/amdkfd/kfd_svm.c
index 16137c4247bb..58b89b53ebe6 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_svm.c
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_svm.c
@@ -1968,10 +1968,16 @@ 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)
@@ -1983,6 +1989,8 @@ 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)
@@ -1990,43 +1998,41 @@ 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);
 
-		mm = prange->work_item.mm;
-retry:
-		mmap_write_lock(mm);
 		mutex_lock(&svms->lock);
-
-		/* Checking for the need to drain retry faults must be in
-		 * mmap write lock to serialize with munmap notifiers.
-		 *
-		 * 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);
-		if (unlikely(svms->drain_pagefaults)) {
-			svms->drain_pagefaults = false;
-			spin_unlock(&svms->deferred_list_lock);
-			mutex_unlock(&svms->lock);
-			mmap_write_unlock(mm);
-			svm_range_drain_retry_fault(svms);
-			goto retry;
-		}
-		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;
@@ -2042,12 +2048,13 @@ retry:
 
 		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);
 }
 
@@ -2056,12 +2063,6 @@ svm_range_add_list_work(struct svm_range_list *svms, struct svm_range *prange,
 			struct mm_struct *mm, enum svm_work_list_ops op)
 {
 	spin_lock(&svms->deferred_list_lock);
-	/* 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".
-	 */
-	if (op == SVM_OP_UNMAP_RANGE)
-		svms->drain_pagefaults = true;
 	/* if prange is on the deferred list */
 	if (!list_empty(&prange->deferred_list)) {
 		pr_debug("update exist prange 0x%p work op %d\n", prange, op);
@@ -2140,6 +2141,12 @@ 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) {
@@ -2559,20 +2566,13 @@ 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;
@@ -2590,6 +2590,7 @@ 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)) {
@@ -2600,7 +2601,7 @@ 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);
@@ -2611,10 +2612,17 @@ 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);
+		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;
 		goto out;
 	}
 
@@ -2663,7 +2671,17 @@ retry_write_locked:
 		goto out_unlock_range;
 	}
 
-	if (!svm_fault_allowed(mm, addr, write_fault)) {
+	/* __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(vma, write_fault)) {
 		pr_debug("fault addr 0x%llx no %s permission\n", addr,
 			write_fault ? "write" : "read");
 		r = -EPERM;
@@ -2741,6 +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);
@@ -2761,6 +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);