drm/amdgpu: Add notifier lock for KFD userptrs

Add a per-process MMU notifier lock for processing notifiers from
userptrs. Use that lock to properly synchronize page table updates with
MMU notifiers.

Signed-off-by: Felix Kuehling <Felix.Kuehling@amd.com>
Reviewed-by: Xiaogang Chen<Xiaogang.Chen@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

1 files changed, 140 insertions, 72 deletions

diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c
index 8782916e64a0..0a854bb8b47e 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c
@@ -964,7 +964,9 @@ static int init_user_pages(struct kgd_mem *mem, uint64_t user_addr,
 		 * later stage when it is scheduled by another ioctl called by
 		 * CRIU master process for the target pid for restore.
 		 */
-		atomic_inc(&mem->invalid);
+		mutex_lock(&process_info->notifier_lock);
+		mem->invalid++;
+		mutex_unlock(&process_info->notifier_lock);
 		mutex_unlock(&process_info->lock);
 		return 0;
 	}
@@ -1301,6 +1303,7 @@ static int init_kfd_vm(struct amdgpu_vm *vm, void **process_info,
 			return -ENOMEM;
 
 		mutex_init(&info->lock);
+		mutex_init(&info->notifier_lock);
 		INIT_LIST_HEAD(&info->vm_list_head);
 		INIT_LIST_HEAD(&info->kfd_bo_list);
 		INIT_LIST_HEAD(&info->userptr_valid_list);
@@ -1317,7 +1320,6 @@ static int init_kfd_vm(struct amdgpu_vm *vm, void **process_info,
 		}
 
 		info->pid = get_task_pid(current->group_leader, PIDTYPE_PID);
-		atomic_set(&info->evicted_bos, 0);
 		INIT_DELAYED_WORK(&info->restore_userptr_work,
 				  amdgpu_amdkfd_restore_userptr_worker);
 
@@ -1372,6 +1374,7 @@ reserve_pd_fail:
 		put_pid(info->pid);
 create_evict_fence_fail:
 		mutex_destroy(&info->lock);
+		mutex_destroy(&info->notifier_lock);
 		kfree(info);
 	}
 	return ret;
@@ -1496,6 +1499,7 @@ void amdgpu_amdkfd_gpuvm_destroy_cb(struct amdgpu_device *adev,
 		cancel_delayed_work_sync(&process_info->restore_userptr_work);
 		put_pid(process_info->pid);
 		mutex_destroy(&process_info->lock);
+		mutex_destroy(&process_info->notifier_lock);
 		kfree(process_info);
 	}
 }
@@ -1548,7 +1552,9 @@ int amdgpu_amdkfd_criu_resume(void *p)
 
 	mutex_lock(&pinfo->lock);
 	pr_debug("scheduling work\n");
-	atomic_inc(&pinfo->evicted_bos);
+	mutex_lock(&pinfo->notifier_lock);
+	pinfo->evicted_bos++;
+	mutex_unlock(&pinfo->notifier_lock);
 	if (!READ_ONCE(pinfo->block_mmu_notifications)) {
 		ret = -EINVAL;
 		goto out_unlock;
@@ -1773,8 +1779,13 @@ int amdgpu_amdkfd_gpuvm_free_memory_of_gpu(
 	list_del(&bo_list_entry->head);
 	mutex_unlock(&process_info->lock);
 
-	/* No more MMU notifiers */
-	amdgpu_hmm_unregister(mem->bo);
+	/* Cleanup user pages and MMU notifiers */
+	if (amdgpu_ttm_tt_get_usermm(mem->bo->tbo.ttm)) {
+		amdgpu_hmm_unregister(mem->bo);
+		mutex_lock(&process_info->notifier_lock);
+		amdgpu_ttm_tt_discard_user_pages(mem->bo->tbo.ttm, mem->range);
+		mutex_unlock(&process_info->notifier_lock);
+	}
 
 	ret = reserve_bo_and_cond_vms(mem, NULL, BO_VM_ALL, &ctx);
 	if (unlikely(ret))
@@ -1864,6 +1875,16 @@ int amdgpu_amdkfd_gpuvm_map_memory_to_gpu(
 	 */
 	mutex_lock(&mem->process_info->lock);
 
+	/* Lock notifier lock. If we find an invalid userptr BO, we can be
+	 * sure that the MMU notifier is no longer running
+	 * concurrently and the queues are actually stopped
+	 */
+	if (amdgpu_ttm_tt_get_usermm(bo->tbo.ttm)) {
+		mutex_lock(&mem->process_info->notifier_lock);
+		is_invalid_userptr = !!mem->invalid;
+		mutex_unlock(&mem->process_info->notifier_lock);
+	}
+
 	mutex_lock(&mem->lock);
 
 	domain = mem->domain;
@@ -2241,34 +2262,38 @@ int amdgpu_amdkfd_gpuvm_import_dmabuf(struct amdgpu_device *adev,
  *
  * Runs in MMU notifier, may be in RECLAIM_FS context. This means it
  * cannot do any memory allocations, and cannot take any locks that
- * are held elsewhere while allocating memory. Therefore this is as
- * simple as possible, using atomic counters.
+ * are held elsewhere while allocating memory.
  *
  * It doesn't do anything to the BO itself. The real work happens in
  * restore, where we get updated page addresses. This function only
  * ensures that GPU access to the BO is stopped.
  */
-int amdgpu_amdkfd_evict_userptr(struct kgd_mem *mem,
-				struct mm_struct *mm)
+int amdgpu_amdkfd_evict_userptr(struct mmu_interval_notifier *mni,
+				unsigned long cur_seq, struct kgd_mem *mem)
 {
 	struct amdkfd_process_info *process_info = mem->process_info;
-	int evicted_bos;
 	int r = 0;
 
-	/* Do not process MMU notifications until stage-4 IOCTL is received */
+	/* Do not process MMU notifications during CRIU restore until
+	 * KFD_CRIU_OP_RESUME IOCTL is received
+	 */
 	if (READ_ONCE(process_info->block_mmu_notifications))
 		return 0;
 
-	atomic_inc(&mem->invalid);
-	evicted_bos = atomic_inc_return(&process_info->evicted_bos);
-	if (evicted_bos == 1) {
+	mutex_lock(&process_info->notifier_lock);
+	mmu_interval_set_seq(mni, cur_seq);
+
+	mem->invalid++;
+	if (++process_info->evicted_bos == 1) {
 		/* First eviction, stop the queues */
-		r = kgd2kfd_quiesce_mm(mm, KFD_QUEUE_EVICTION_TRIGGER_USERPTR);
+		r = kgd2kfd_quiesce_mm(mni->mm,
+				       KFD_QUEUE_EVICTION_TRIGGER_USERPTR);
 		if (r)
 			pr_err("Failed to quiesce KFD\n");
 		schedule_delayed_work(&process_info->restore_userptr_work,
 			msecs_to_jiffies(AMDGPU_USERPTR_RESTORE_DELAY_MS));
 	}
+	mutex_unlock(&process_info->notifier_lock);
 
 	return r;
 }
@@ -2285,54 +2310,58 @@ static int update_invalid_user_pages(struct amdkfd_process_info *process_info,
 	struct kgd_mem *mem, *tmp_mem;
 	struct amdgpu_bo *bo;
 	struct ttm_operation_ctx ctx = { false, false };
-	int invalid, ret;
+	uint32_t invalid;
+	int ret = 0;
 
-	/* Move all invalidated BOs to the userptr_inval_list and
-	 * release their user pages by migration to the CPU domain
-	 */
+	mutex_lock(&process_info->notifier_lock);
+
+	/* Move all invalidated BOs to the userptr_inval_list */
 	list_for_each_entry_safe(mem, tmp_mem,
 				 &process_info->userptr_valid_list,
-				 validate_list.head) {
-		if (!atomic_read(&mem->invalid))
-			continue; /* BO is still valid */
-
-		bo = mem->bo;
-
-		if (amdgpu_bo_reserve(bo, true))
-			return -EAGAIN;
-		amdgpu_bo_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_CPU);
-		ret = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
-		amdgpu_bo_unreserve(bo);
-		if (ret) {
-			pr_err("%s: Failed to invalidate userptr BO\n",
-			       __func__);
-			return -EAGAIN;
-		}
-
-		list_move_tail(&mem->validate_list.head,
-			       &process_info->userptr_inval_list);
-	}
-
-	if (list_empty(&process_info->userptr_inval_list))
-		return 0; /* All evicted userptr BOs were freed */
+				 validate_list.head)
+		if (mem->invalid)
+			list_move_tail(&mem->validate_list.head,
+				       &process_info->userptr_inval_list);
 
 	/* Go through userptr_inval_list and update any invalid user_pages */
 	list_for_each_entry(mem, &process_info->userptr_inval_list,
 			    validate_list.head) {
-		struct hmm_range *range;
-
-		invalid = atomic_read(&mem->invalid);
+		invalid = mem->invalid;
 		if (!invalid)
 			/* BO hasn't been invalidated since the last
-			 * revalidation attempt. Keep its BO list.
+			 * revalidation attempt. Keep its page list.
 			 */
 			continue;
 
 		bo = mem->bo;
 
+		amdgpu_ttm_tt_discard_user_pages(bo->tbo.ttm, mem->range);
+		mem->range = NULL;
+
+		/* BO reservations and getting user pages (hmm_range_fault)
+		 * must happen outside the notifier lock
+		 */
+		mutex_unlock(&process_info->notifier_lock);
+
+		/* Move the BO to system (CPU) domain if necessary to unmap
+		 * and free the SG table
+		 */
+		if (bo->tbo.resource->mem_type != TTM_PL_SYSTEM) {
+			if (amdgpu_bo_reserve(bo, true))
+				return -EAGAIN;
+			amdgpu_bo_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_CPU);
+			ret = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
+			amdgpu_bo_unreserve(bo);
+			if (ret) {
+				pr_err("%s: Failed to invalidate userptr BO\n",
+				       __func__);
+				return -EAGAIN;
+			}
+		}
+
 		/* Get updated user pages */
 		ret = amdgpu_ttm_tt_get_user_pages(bo, bo->tbo.ttm->pages,
-						   &range);
+						   &mem->range);
 		if (ret) {
 			pr_debug("Failed %d to get user pages\n", ret);
 
@@ -2345,30 +2374,32 @@ static int update_invalid_user_pages(struct amdkfd_process_info *process_info,
 			 */
 			if (ret != -EFAULT)
 				return ret;
-		} else {
 
-			/*
-			 * FIXME: Cannot ignore the return code, must hold
-			 * notifier_lock
-			 */
-			amdgpu_ttm_tt_get_user_pages_done(bo->tbo.ttm, range);
+			ret = 0;
 		}
 
+		mutex_lock(&process_info->notifier_lock);
+
 		/* Mark the BO as valid unless it was invalidated
 		 * again concurrently.
 		 */
-		if (atomic_cmpxchg(&mem->invalid, invalid, 0) != invalid)
-			return -EAGAIN;
+		if (mem->invalid != invalid) {
+			ret = -EAGAIN;
+			goto unlock_out;
+		}
+		mem->invalid = 0;
 	}
 
-	return 0;
+unlock_out:
+	mutex_unlock(&process_info->notifier_lock);
+
+	return ret;
 }
 
 /* Validate invalid userptr BOs
  *
- * Validates BOs on the userptr_inval_list, and moves them back to the
- * userptr_valid_list. Also updates GPUVM page tables with new page
- * addresses and waits for the page table updates to complete.
+ * Validates BOs on the userptr_inval_list. Also updates GPUVM page tables
+ * with new page addresses and waits for the page table updates to complete.
  */
 static int validate_invalid_user_pages(struct amdkfd_process_info *process_info)
 {
@@ -2439,9 +2470,6 @@ static int validate_invalid_user_pages(struct amdkfd_process_info *process_info)
 			}
 		}
 
-		list_move_tail(&mem->validate_list.head,
-			       &process_info->userptr_valid_list);
-
 		/* Update mapping. If the BO was not validated
 		 * (because we couldn't get user pages), this will
 		 * clear the page table entries, which will result in
@@ -2457,7 +2485,9 @@ static int validate_invalid_user_pages(struct amdkfd_process_info *process_info)
 			if (ret) {
 				pr_err("%s: update PTE failed\n", __func__);
 				/* make sure this gets validated again */
-				atomic_inc(&mem->invalid);
+				mutex_lock(&process_info->notifier_lock);
+				mem->invalid++;
+				mutex_unlock(&process_info->notifier_lock);
 				goto unreserve_out;
 			}
 		}
@@ -2477,6 +2507,36 @@ out_no_mem:
 	return ret;
 }
 
+/* Confirm that all user pages are valid while holding the notifier lock
+ *
+ * Moves valid BOs from the userptr_inval_list back to userptr_val_list.
+ */
+static int confirm_valid_user_pages_locked(struct amdkfd_process_info *process_info)
+{
+	struct kgd_mem *mem, *tmp_mem;
+	int ret = 0;
+
+	list_for_each_entry_safe(mem, tmp_mem,
+				 &process_info->userptr_inval_list,
+				 validate_list.head) {
+		bool valid = amdgpu_ttm_tt_get_user_pages_done(
+				mem->bo->tbo.ttm, mem->range);
+
+		mem->range = NULL;
+		if (!valid) {
+			WARN(!mem->invalid, "Invalid BO not marked invalid");
+			ret = -EAGAIN;
+			continue;
+		}
+		WARN(mem->invalid, "Valid BO is marked invalid");
+
+		list_move_tail(&mem->validate_list.head,
+			       &process_info->userptr_valid_list);
+	}
+
+	return ret;
+}
+
 /* Worker callback to restore evicted userptr BOs
  *
  * Tries to update and validate all userptr BOs. If successful and no
@@ -2491,9 +2551,11 @@ static void amdgpu_amdkfd_restore_userptr_worker(struct work_struct *work)
 			     restore_userptr_work);
 	struct task_struct *usertask;
 	struct mm_struct *mm;
-	int evicted_bos;
+	uint32_t evicted_bos;
 
-	evicted_bos = atomic_read(&process_info->evicted_bos);
+	mutex_lock(&process_info->notifier_lock);
+	evicted_bos = process_info->evicted_bos;
+	mutex_unlock(&process_info->notifier_lock);
 	if (!evicted_bos)
 		return;
 
@@ -2516,9 +2578,6 @@ static void amdgpu_amdkfd_restore_userptr_worker(struct work_struct *work)
 	 * and we can just restart the queues.
 	 */
 	if (!list_empty(&process_info->userptr_inval_list)) {
-		if (atomic_read(&process_info->evicted_bos) != evicted_bos)
-			goto unlock_out; /* Concurrent eviction, try again */
-
 		if (validate_invalid_user_pages(process_info))
 			goto unlock_out;
 	}
@@ -2527,10 +2586,17 @@ static void amdgpu_amdkfd_restore_userptr_worker(struct work_struct *work)
 	 * be a first eviction that calls quiesce_mm. The eviction
 	 * reference counting inside KFD will handle this case.
 	 */
-	if (atomic_cmpxchg(&process_info->evicted_bos, evicted_bos, 0) !=
-	    evicted_bos)
-		goto unlock_out;
-	evicted_bos = 0;
+	mutex_lock(&process_info->notifier_lock);
+	if (process_info->evicted_bos != evicted_bos)
+		goto unlock_notifier_out;
+
+	if (confirm_valid_user_pages_locked(process_info)) {
+		WARN(1, "User pages unexpectedly invalid");
+		goto unlock_notifier_out;
+	}
+
+	process_info->evicted_bos = evicted_bos = 0;
+
 	if (kgd2kfd_resume_mm(mm)) {
 		pr_err("%s: Failed to resume KFD\n", __func__);
 		/* No recovery from this failure. Probably the CP is
@@ -2538,6 +2604,8 @@ static void amdgpu_amdkfd_restore_userptr_worker(struct work_struct *work)
 		 */
 	}
 
+unlock_notifier_out:
+	mutex_unlock(&process_info->notifier_lock);
 unlock_out:
 	mutex_unlock(&process_info->lock);