Skip to contentSkip to conversation

drm/ttm: remove ttm_bo_(un)lock_delayed_workqueue

PATCH In review Christian König 2 participants 4 messages Nov 24, 2022

Subsystems (5)
lore
Share & tools
lore
#0 lore raw

PATCH 1/3 drm/ttm: remove ttm_bo_(un)lock_delayed_workqueue

Those functions never worked correctly since it is still perfectly possible that a buffer object is released and the background worker restarted even after calling them.

---
 drivers/gpu/drm/amd/amdgpu/amdgpu_debugfs.c |  6 +-----
 drivers/gpu/drm/amd/amdgpu/amdgpu_device.c  |  4 +---
 drivers/gpu/drm/radeon/radeon_device.c      |  5 -----
 drivers/gpu/drm/radeon/radeon_pm.c          |  4 +---
 drivers/gpu/drm/ttm/ttm_bo.c                | 14 --------------
 include/drm/ttm/ttm_bo_api.h                | 16 ----------------
 6 files changed, 3 insertions(+), 46 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_debugfs.c +1 -5
index de61a85c4b02..8c57a5b7ecba 100644
@@ -1717,7 +1717,7 @@ static void amdgpu_ib_preempt_mark_partial_job(struct amdgpu_ring *ring)
 
 static int amdgpu_debugfs_ib_preempt(void *data, u64 val)
 {
-	int r, resched, length;
+	int r, length;
 	struct amdgpu_ring *ring;
 	struct dma_fence **fences = NULL;
 	struct amdgpu_device *adev = (struct amdgpu_device *)data;
@@ -1747,8 +1747,6 @@ static int amdgpu_debugfs_ib_preempt(void *data, u64 val)
 	/* stop the scheduler */
 	kthread_park(ring->sched.thread);
 
-	resched = ttm_bo_lock_delayed_workqueue(&adev->mman.bdev);
-
 	/* preempt the IB */
 	r = amdgpu_ring_preempt_ib(ring);
 	if (r) {
@@ -1785,8 +1783,6 @@ static int amdgpu_debugfs_ib_preempt(void *data, u64 val)
 
 	up_read(&adev->reset_domain->sem);
 
-	ttm_bo_unlock_delayed_workqueue(&adev->mman.bdev, resched);
-
 pro_end:
 	kfree(fences);
 
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +1 -3
index cc65df9f2419..ff2ae0be2c28 100644
@@ -3965,10 +3965,8 @@ void amdgpu_device_fini_hw(struct amdgpu_device *adev)
 	}
 	amdgpu_fence_driver_hw_fini(adev);
 
-	if (adev->mman.initialized) {
+	if (adev->mman.initialized)
 		flush_delayed_work(&adev->mman.bdev.wq);
-		ttm_bo_lock_delayed_workqueue(&adev->mman.bdev);
-	}
 
 	if (adev->pm_sysfs_en)
 		amdgpu_pm_sysfs_fini(adev);
drivers/gpu/drm/radeon/radeon_device.c +0 -5
index a556b6be1137..ea10306809cf 100644
@@ -1768,7 +1768,6 @@ int radeon_gpu_reset(struct radeon_device *rdev)
 	bool saved = false;
 
 	int i, r;
-	int resched;
 
 	down_write(&rdev->exclusive_lock);
 
@@ -1780,8 +1779,6 @@ int radeon_gpu_reset(struct radeon_device *rdev)
 	atomic_inc(&rdev->gpu_reset_counter);
 
 	radeon_save_bios_scratch_regs(rdev);
-	/* block TTM */
-	resched = ttm_bo_lock_delayed_workqueue(&rdev->mman.bdev);
 	radeon_suspend(rdev);
 	radeon_hpd_fini(rdev);
 
@@ -1840,8 +1837,6 @@ int radeon_gpu_reset(struct radeon_device *rdev)
 	/* reset hpd state */
 	radeon_hpd_init(rdev);
 
-	ttm_bo_unlock_delayed_workqueue(&rdev->mman.bdev, resched);
-
 	rdev->in_reset = true;
 	rdev->needs_reset = false;
 
drivers/gpu/drm/radeon/radeon_pm.c +1 -3
index 04c693ca419a..cbc554928bcc 100644
@@ -1853,11 +1853,10 @@ static bool radeon_pm_debug_check_in_vbl(struct radeon_device *rdev, bool finish
 static void radeon_dynpm_idle_work_handler(struct work_struct *work)
 {
 	struct radeon_device *rdev;
-	int resched;
+
 	rdev = container_of(work, struct radeon_device,
 				pm.dynpm_idle_work.work);
 
-	resched = ttm_bo_lock_delayed_workqueue(&rdev->mman.bdev);
 	mutex_lock(&rdev->pm.mutex);
 	if (rdev->pm.dynpm_state == DYNPM_STATE_ACTIVE) {
 		int not_processed = 0;
@@ -1908,7 +1907,6 @@ static void radeon_dynpm_idle_work_handler(struct work_struct *work)
 				      msecs_to_jiffies(RADEON_IDLE_LOOP_MS));
 	}
 	mutex_unlock(&rdev->pm.mutex);
-	ttm_bo_unlock_delayed_workqueue(&rdev->mman.bdev, resched);
 }
 
 /*
drivers/gpu/drm/ttm/ttm_bo.c +0 -14
index c3f4b33136e5..b77262a623e0 100644
@@ -418,20 +418,6 @@ void ttm_bo_put(struct ttm_buffer_object *bo)
 }
 EXPORT_SYMBOL(ttm_bo_put);
 
-int ttm_bo_lock_delayed_workqueue(struct ttm_device *bdev)
-{
-	return cancel_delayed_work_sync(&bdev->wq);
-}
-EXPORT_SYMBOL(ttm_bo_lock_delayed_workqueue);
-
-void ttm_bo_unlock_delayed_workqueue(struct ttm_device *bdev, int resched)
-{
-	if (resched)
-		schedule_delayed_work(&bdev->wq,
-				      ((HZ / 100) < 1) ? 1 : HZ / 100);
-}
-EXPORT_SYMBOL(ttm_bo_unlock_delayed_workqueue);
-
 static int ttm_bo_bounce_temp_buffer(struct ttm_buffer_object *bo,
 				     struct ttm_resource **mem,
 				     struct ttm_operation_ctx *ctx,
include/drm/ttm/ttm_bo_api.h +0 -16
index 44a538ee5e2a..7758347c461c 100644
@@ -290,22 +290,6 @@ void ttm_bo_move_to_lru_tail(struct ttm_buffer_object *bo);
 void ttm_bo_set_bulk_move(struct ttm_buffer_object *bo,
 			  struct ttm_lru_bulk_move *bulk);
 
-/**
- * ttm_bo_lock_delayed_workqueue
- *
- * Prevent the delayed workqueue from running.
- * Returns
- * True if the workqueue was queued at the time
- */
-int ttm_bo_lock_delayed_workqueue(struct ttm_device *bdev);
-
-/**
- * ttm_bo_unlock_delayed_workqueue
- *
- * Allows the delayed workqueue to run.
- */
-void ttm_bo_unlock_delayed_workqueue(struct ttm_device *bdev, int resched);
-
 /**
  * ttm_bo_eviction_valuable
  *
Signature · 1 line
2.34.1
#1 lore raw

PATCH 2/3 drm/ttm: use per BO cleanup workers

Instead of a single worker going over the list of delete BOs in regular intervals use a per BO worker which blocks for the resv object and locking of the BO.

This not only simplifies the handling massively, but also results in much better response time when cleaning up buffers.

---
 drivers/gpu/drm/amd/amdgpu/amdgpu_device.c |   2 +-
 drivers/gpu/drm/i915/i915_gem.c            |   2 +-
 drivers/gpu/drm/i915/intel_region_ttm.c    |   2 +-
 drivers/gpu/drm/ttm/ttm_bo.c               | 112 ++++++++-------------
 drivers/gpu/drm/ttm/ttm_bo_util.c          |   1 -
 drivers/gpu/drm/ttm/ttm_device.c           |  24 ++---
 include/drm/ttm/ttm_bo_api.h               |  18 +---
 include/drm/ttm/ttm_device.h               |   7 +-
 8 files changed, 57 insertions(+), 111 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +1 -1
index ff2ae0be2c28..05f458d67eb8 100644
@@ -3966,7 +3966,7 @@ void amdgpu_device_fini_hw(struct amdgpu_device *adev)
 	amdgpu_fence_driver_hw_fini(adev);
 
 	if (adev->mman.initialized)
-		flush_delayed_work(&adev->mman.bdev.wq);
+		drain_workqueue(adev->mman.bdev.wq);
 
 	if (adev->pm_sysfs_en)
 		amdgpu_pm_sysfs_fini(adev);
drivers/gpu/drm/i915/i915_gem.c +1 -1
index 299f94a9fb87..413c9dbe5be1 100644
@@ -1099,7 +1099,7 @@ void i915_gem_drain_freed_objects(struct drm_i915_private *i915)
 {
 	while (atomic_read(&i915->mm.free_count)) {
 		flush_work(&i915->mm.free_work);
-		flush_delayed_work(&i915->bdev.wq);
+		drain_workqueue(i915->bdev.wq);
 		rcu_barrier();
 	}
 }
drivers/gpu/drm/i915/intel_region_ttm.c +1 -1
index cf89d0c2a2d9..657bbc16a48a 100644
@@ -132,7 +132,7 @@ int intel_region_ttm_fini(struct intel_memory_region *mem)
 			break;
 
 		msleep(20);
-		flush_delayed_work(&mem->i915->bdev.wq);
+		drain_workqueue(mem->i915->bdev.wq);
 	}
 
 	/* If we leaked objects, Don't free the region causing use after free */
drivers/gpu/drm/ttm/ttm_bo.c +40 -72

152 lines not on this page · raw message

drivers/gpu/drm/ttm/ttm_bo_util.c +0 -1
index ba3aa0a0fc43..ae4b7922ee1a 100644
@@ -230,7 +230,6 @@ static int ttm_buffer_object_transfer(struct ttm_buffer_object *bo,
 	 */
 
 	atomic_inc(&ttm_glob.bo_count);
-	INIT_LIST_HEAD(&fbo->base.ddestroy);
 	drm_vma_node_reset(&fbo->base.base.vma_node);
 
 	kref_init(&fbo->base.kref);
drivers/gpu/drm/ttm/ttm_device.c +8 -16
index e7147e304637..e9bedca4dfdc 100644
@@ -175,16 +175,6 @@ int ttm_device_swapout(struct ttm_device *bdev, struct ttm_operation_ctx *ctx,
 }
 EXPORT_SYMBOL(ttm_device_swapout);
 
-static void ttm_device_delayed_workqueue(struct work_struct *work)
-{
-	struct ttm_device *bdev =
-		container_of(work, struct ttm_device, wq.work);
-
-	if (!ttm_bo_delayed_delete(bdev, false))
-		schedule_delayed_work(&bdev->wq,
-				      ((HZ / 100) < 1) ? 1 : HZ / 100);
-}
-
 /**
  * ttm_device_init
  *
@@ -215,15 +205,19 @@ int ttm_device_init(struct ttm_device *bdev, struct ttm_device_funcs *funcs,
 	if (ret)
 		return ret;
 
+	bdev->wq = alloc_workqueue("ttm", WQ_MEM_RECLAIM | WQ_HIGHPRI, 16);
+	if (!bdev->wq) {
+		ttm_global_release();
+		return -ENOMEM;
+	}
+
 	bdev->funcs = funcs;
 
 	ttm_sys_man_init(bdev);
 	ttm_pool_init(&bdev->pool, dev, use_dma_alloc, use_dma32);
 
 	bdev->vma_manager = vma_manager;
-	INIT_DELAYED_WORK(&bdev->wq, ttm_device_delayed_workqueue);
 	spin_lock_init(&bdev->lru_lock);
-	INIT_LIST_HEAD(&bdev->ddestroy);
 	INIT_LIST_HEAD(&bdev->pinned);
 	bdev->dev_mapping = mapping;
 	mutex_lock(&ttm_global_mutex);
@@ -247,10 +241,8 @@ void ttm_device_fini(struct ttm_device *bdev)
 	list_del(&bdev->device_list);
 	mutex_unlock(&ttm_global_mutex);
 
-	cancel_delayed_work_sync(&bdev->wq);
-
-	if (ttm_bo_delayed_delete(bdev, true))
-		pr_debug("Delayed destroy list was clean\n");
+	drain_workqueue(bdev->wq);
+	destroy_workqueue(bdev->wq);
 
 	spin_lock(&bdev->lru_lock);
 	for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i)
include/drm/ttm/ttm_bo_api.h +5 -13
index 7758347c461c..69e62bbb01e3 100644
@@ -92,7 +92,6 @@ struct ttm_tt;
  * @ttm: TTM structure holding system pages.
  * @evicted: Whether the object was evicted without user-space knowing.
  * @deleted: True if the object is only a zombie and already deleted.
- * @ddestroy: List head for the delayed destroy list.
  * @swap: List head for swap LRU list.
  * @offset: The current GPU offset, which can have different meanings
  * depending on the memory type. For SYSTEM type memory, it should be 0.
@@ -135,19 +134,14 @@ struct ttm_buffer_object {
 	struct ttm_tt *ttm;
 	bool deleted;
 	struct ttm_lru_bulk_move *bulk_move;
+	unsigned priority;
+	unsigned pin_count;
 
 	/**
-	 * Members protected by the bdev::lru_lock.
-	 */
-
-	struct list_head ddestroy;
-
-	/**
-	 * Members protected by a bo reservation.
+	 * @delayed_delete: Work item used when we can't delete the BO
+	 * immediately
 	 */
-
-	unsigned priority;
-	unsigned pin_count;
+	struct work_struct delayed_delete;
 
 	/**
 	 * Special members that are protected by the reserve lock
@@ -448,8 +442,6 @@ void ttm_bo_vm_close(struct vm_area_struct *vma);
 
 int ttm_bo_vm_access(struct vm_area_struct *vma, unsigned long addr,
 		     void *buf, int len, int write);
-bool ttm_bo_delayed_delete(struct ttm_device *bdev, bool remove_all);
-
 vm_fault_t ttm_bo_vm_dummy_page(struct vm_fault *vmf, pgprot_t prot);
 
 #endif
include/drm/ttm/ttm_device.h +1 -6
index 95b3c04b1ab9..4f3e81eac6f3 100644
@@ -251,11 +251,6 @@ struct ttm_device {
 	 */
 	spinlock_t lru_lock;
 
-	/**
-	 * @ddestroy: Destroyed but not yet cleaned up buffer objects.
-	 */
-	struct list_head ddestroy;
-
 	/**
 	 * @pinned: Buffer objects which are pinned and so not on any LRU list.
 	 */
@@ -270,7 +265,7 @@ struct ttm_device {
 	/**
 	 * @wq: Work queue structure for the delayed delete workqueue.
 	 */
-	struct delayed_work wq;
+	struct workqueue_struct *wq;
 };
 
 int ttm_global_swapout(struct ttm_operation_ctx *ctx, gfp_t gfp_flags);
Signature · 1 line
2.34.1
#2 lore raw

PATCH 3/3 drm/amdgpu: generally allow over-commit during BO allocation

In reply to Christian König

We already fallback to a dummy BO with no backing store when we allocate GDS,GWS and OA resources and to GTT when we allocate VRAM.

Drop all those workarounds and generalize this for GTT as well. This fixes ENOMEM issues with runaway applications which try to allocate/free GTT in a loop and are otherwise only limited by the CPU speed.

The CS will wait for the cleanup of freed up BOs to satisfy the various domain specific limits and so effectively throttle those buggy applications down to a sane allocation behavior again.

---
 drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c    | 16 +++-------------
 drivers/gpu/drm/amd/amdgpu/amdgpu_object.c |  6 +-----
 2 files changed, 4 insertions(+), 18 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c +3 -13
index 8ef31d687ef3..286509e5e17e 100644
@@ -112,7 +112,7 @@ int amdgpu_gem_object_create(struct amdgpu_device *adev, unsigned long size,
 	bp.resv = resv;
 	bp.preferred_domain = initial_domain;
 	bp.flags = flags;
-	bp.domain = initial_domain;
+	bp.domain = initial_domain | AMDGPU_GEM_DOMAIN_CPU;
 	bp.bo_ptr_size = sizeof(struct amdgpu_bo);
 
 	r = amdgpu_bo_create_user(adev, &bp, &ubo);
@@ -331,20 +331,10 @@ int amdgpu_gem_create_ioctl(struct drm_device *dev, void *data,
 	}
 
 	initial_domain = (u32)(0xffffffff & args->in.domains);
-retry:
 	r = amdgpu_gem_object_create(adev, size, args->in.alignment,
-				     initial_domain,
-				     flags, ttm_bo_type_device, resv, &gobj);
+				     initial_domain, flags, ttm_bo_type_device,
+				     resv, &gobj);
 	if (r && r != -ERESTARTSYS) {
-		if (flags & AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED) {
-			flags &= ~AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED;
-			goto retry;
-		}
-
-		if (initial_domain == AMDGPU_GEM_DOMAIN_VRAM) {
-			initial_domain |= AMDGPU_GEM_DOMAIN_GTT;
-			goto retry;
-		}
 		DRM_DEBUG("Failed to allocate GEM object (%llu, %d, %llu, %d)\n",
 				size, initial_domain, args->in.alignment, r);
 	}
drivers/gpu/drm/amd/amdgpu/amdgpu_object.c +1 -5
index 974e85d8b6cc..919bbea2e3ac 100644
@@ -581,11 +581,7 @@ int amdgpu_bo_create(struct amdgpu_device *adev,
 		bo->flags |= AMDGPU_GEM_CREATE_VRAM_WIPE_ON_RELEASE;
 
 	bo->tbo.bdev = &adev->mman.bdev;
-	if (bp->domain & (AMDGPU_GEM_DOMAIN_GWS | AMDGPU_GEM_DOMAIN_OA |
-			  AMDGPU_GEM_DOMAIN_GDS))
-		amdgpu_bo_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_CPU);
-	else
-		amdgpu_bo_placement_from_domain(bo, bp->domain);
+	amdgpu_bo_placement_from_domain(bo, bp->domain);
 	if (bp->type == ttm_bo_type_kernel)
 		bo->tbo.priority = 1;
 
Signature · 1 line
2.34.1
#3 lore raw
In reply to Christian König
Quoted 6 lines · Christian König

Those functions never worked correctly since it is still perfectly possible that a buffer object is released and the background worker restarted even after calling them.

I know you usually do, but just a friendly reminder to Cc: intel-gfx on the next revision or before merging, just so our CI can give the series a quick test. Thanks.