[PATCH 20/26] drm/i915: Add i915_vma_unbind_unlocked, and take obj lock for i915_vma_unbind

Maarten Lankhorst maarten.lankhorst at linux.intel.com
Wed Sep 22 13:21:43 UTC 2021


We want to remove more members of i915_vma, which requires the locking to be
held more often.

Start requiring gem object lock for i915_vma_unbind, as it's one of the
callers that may unpin pages.

Some special care is needed when evicting, because the last reference to the
object may be held by the VMA, so after __i915_vma_unbind, vma may be garbage,
and we need to cache vma->obj before unlocking.

Signed-off-by: Maarten Lankhorst <maarten.lankhorst at linux.intel.com>
---
 drivers/gpu/drm/i915/display/intel_display.c  |  2 +-
 .../gpu/drm/i915/gem/selftests/huge_pages.c   |  4 +-
 .../i915/gem/selftests/i915_gem_client_blt.c  |  2 +-
 .../drm/i915/gem/selftests/i915_gem_mman.c    |  6 +++
 drivers/gpu/drm/i915/gt/intel_ggtt.c          | 46 ++++++++++++++++---
 drivers/gpu/drm/i915/i915_gem.c               |  2 +
 drivers/gpu/drm/i915/i915_vma.c               | 27 ++++++++++-
 drivers/gpu/drm/i915/i915_vma.h               |  1 +
 drivers/gpu/drm/i915/selftests/i915_gem_gtt.c | 22 ++++-----
 drivers/gpu/drm/i915/selftests/i915_vma.c     |  2 +-
 10 files changed, 90 insertions(+), 24 deletions(-)

diff --git a/drivers/gpu/drm/i915/display/intel_display.c b/drivers/gpu/drm/i915/display/intel_display.c
index f6c0c595f631..5c936ca860f0 100644
--- a/drivers/gpu/drm/i915/display/intel_display.c
+++ b/drivers/gpu/drm/i915/display/intel_display.c
@@ -941,7 +941,7 @@ intel_pin_fb_obj_dpt(struct drm_framebuffer *fb,
 		goto err;
 
 	if (i915_vma_misplaced(vma, 0, alignment, 0)) {
-		ret = i915_vma_unbind(vma);
+		ret = i915_vma_unbind_unlocked(vma);
 		if (ret) {
 			vma = ERR_PTR(ret);
 			goto err;
diff --git a/drivers/gpu/drm/i915/gem/selftests/huge_pages.c b/drivers/gpu/drm/i915/gem/selftests/huge_pages.c
index b74ed3887994..94667ec76340 100644
--- a/drivers/gpu/drm/i915/gem/selftests/huge_pages.c
+++ b/drivers/gpu/drm/i915/gem/selftests/huge_pages.c
@@ -626,7 +626,7 @@ static int igt_mock_ppgtt_misaligned_dma(void *arg)
 		 * pages.
 		 */
 		for (offset = 4096; offset < page_size; offset += 4096) {
-			err = i915_vma_unbind(vma);
+			err = i915_vma_unbind_unlocked(vma);
 			if (err)
 				goto out_unpin;
 
@@ -1077,7 +1077,7 @@ static int __igt_write_huge(struct intel_context *ce,
 	if (IS_ERR(vma))
 		return PTR_ERR(vma);
 
-	err = i915_vma_unbind(vma);
+	err = i915_vma_unbind_unlocked(vma);
 	if (err)
 		return err;
 
diff --git a/drivers/gpu/drm/i915/gem/selftests/i915_gem_client_blt.c b/drivers/gpu/drm/i915/gem/selftests/i915_gem_client_blt.c
index ecbcbb86ae1e..d132933c5054 100644
--- a/drivers/gpu/drm/i915/gem/selftests/i915_gem_client_blt.c
+++ b/drivers/gpu/drm/i915/gem/selftests/i915_gem_client_blt.c
@@ -311,7 +311,7 @@ static int pin_buffer(struct i915_vma *vma, u64 addr)
 	int err;
 
 	if (drm_mm_node_allocated(&vma->node) && vma->node.start != addr) {
-		err = i915_vma_unbind(vma);
+		err = i915_vma_unbind_unlocked(vma);
 		if (err)
 			return err;
 	}
diff --git a/drivers/gpu/drm/i915/gem/selftests/i915_gem_mman.c b/drivers/gpu/drm/i915/gem/selftests/i915_gem_mman.c
index 6d30cdfa80f3..e69e8861352d 100644
--- a/drivers/gpu/drm/i915/gem/selftests/i915_gem_mman.c
+++ b/drivers/gpu/drm/i915/gem/selftests/i915_gem_mman.c
@@ -165,7 +165,9 @@ static int check_partial_mapping(struct drm_i915_gem_object *obj,
 	kunmap(p);
 
 out:
+	i915_gem_object_lock(obj, NULL);
 	__i915_vma_put(vma);
+	i915_gem_object_unlock(obj);
 	return err;
 }
 
@@ -259,7 +261,9 @@ static int check_partial_mappings(struct drm_i915_gem_object *obj,
 		if (err)
 			return err;
 
+		i915_gem_object_lock(obj, NULL);
 		__i915_vma_put(vma);
+		i915_gem_object_unlock(obj);
 
 		if (igt_timeout(end_time,
 				"%s: timed out after tiling=%d stride=%d\n",
@@ -1349,7 +1353,9 @@ static int __igt_mmap_revoke(struct drm_i915_private *i915,
 	 * for other objects. Ergo we have to revoke the previous mmap PTE
 	 * access as it no longer points to the same object.
 	 */
+	i915_gem_object_lock(obj, NULL);
 	err = i915_gem_object_unbind(obj, I915_GEM_OBJECT_UNBIND_ACTIVE);
+	i915_gem_object_unlock(obj);
 	if (err) {
 		pr_err("Failed to unbind object!\n");
 		goto out_unmap;
diff --git a/drivers/gpu/drm/i915/gt/intel_ggtt.c b/drivers/gpu/drm/i915/gt/intel_ggtt.c
index edff2bfc9867..9de619489765 100644
--- a/drivers/gpu/drm/i915/gt/intel_ggtt.c
+++ b/drivers/gpu/drm/i915/gt/intel_ggtt.c
@@ -118,22 +118,44 @@ void i915_ggtt_suspend(struct i915_ggtt *ggtt)
 	struct i915_vma *vma, *vn;
 	int open;
 
+retry:
 	mutex_lock(&ggtt->vm.mutex);
 
 	/* Skip rewriting PTE on VMA unbind. */
 	open = atomic_xchg(&ggtt->vm.open, 0);
 
 	list_for_each_entry_safe(vma, vn, &ggtt->vm.bound_list, vm_link) {
+		struct drm_i915_gem_object *obj = vma->obj;
+
 		GEM_BUG_ON(!drm_mm_node_allocated(&vma->node));
-		i915_vma_wait_for_bind(vma);
 
-		if (i915_vma_is_pinned(vma))
+		if (i915_vma_is_pinned(vma) || !i915_vma_is_bound(vma, I915_VMA_GLOBAL_BIND))
 			continue;
 
-		if (!i915_vma_is_bound(vma, I915_VMA_GLOBAL_BIND)) {
-			__i915_vma_evict(vma);
-			drm_mm_remove_node(&vma->node);
+		i915_gem_object_get(obj);
+
+		/* unlikely to race when GPU is idle, so no worry about slowpath.. */
+		if (!i915_gem_object_trylock(obj, NULL)) {
+			atomic_set(&ggtt->vm.open, open);
+			mutex_unlock(&ggtt->vm.mutex);
+
+			i915_gem_object_lock(obj, NULL);
+			open = i915_vma_unbind(vma);
+			i915_gem_object_unlock(obj);
+
+			GEM_WARN_ON(open);
+
+			i915_gem_object_put(obj);
+			goto retry;
 		}
+
+		i915_vma_wait_for_bind(vma);
+
+		__i915_vma_evict(vma);
+		drm_mm_remove_node(&vma->node);
+
+		i915_gem_object_unlock(obj);
+		i915_gem_object_put(obj);
 	}
 
 	ggtt->vm.clear_range(&ggtt->vm, 0, ggtt->vm.total);
@@ -728,9 +750,21 @@ static void ggtt_cleanup_hw(struct i915_ggtt *ggtt)
 
 	mutex_lock(&ggtt->vm.mutex);
 
-	list_for_each_entry_safe(vma, vn, &ggtt->vm.bound_list, vm_link)
+	list_for_each_entry_safe(vma, vn, &ggtt->vm.bound_list, vm_link) {
+		struct drm_i915_gem_object *obj = vma->obj;
+		bool trylock;
+
+		i915_gem_object_get(obj);
+		trylock = i915_gem_object_trylock(obj, NULL);
+
+		WARN_ON(!trylock);
 		WARN_ON(__i915_vma_unbind(vma));
 
+		if (trylock)
+			i915_gem_object_unlock(obj);
+		i915_gem_object_put(obj);
+	}
+
 	if (drm_mm_node_allocated(&ggtt->error_capture))
 		drm_mm_remove_node(&ggtt->error_capture);
 	mutex_destroy(&ggtt->error_mutex);
diff --git a/drivers/gpu/drm/i915/i915_gem.c b/drivers/gpu/drm/i915/i915_gem.c
index b59117690960..8db60cf677bf 100644
--- a/drivers/gpu/drm/i915/i915_gem.c
+++ b/drivers/gpu/drm/i915/i915_gem.c
@@ -118,6 +118,8 @@ int i915_gem_object_unbind(struct drm_i915_gem_object *obj,
 	struct i915_vma *vma;
 	int ret;
 
+	assert_object_held(obj);
+
 	if (list_empty(&obj->vma.list))
 		return 0;
 
diff --git a/drivers/gpu/drm/i915/i915_vma.c b/drivers/gpu/drm/i915/i915_vma.c
index 69c82ee610f0..97daf871f239 100644
--- a/drivers/gpu/drm/i915/i915_vma.c
+++ b/drivers/gpu/drm/i915/i915_vma.c
@@ -1480,8 +1480,16 @@ void i915_vma_parked(struct intel_gt *gt)
 		struct drm_i915_gem_object *obj = vma->obj;
 		struct i915_address_space *vm = vma->vm;
 
-		INIT_LIST_HEAD(&vma->closed_link);
-		__i915_vma_put(vma);
+		if (i915_gem_object_trylock(obj, NULL)) {
+			INIT_LIST_HEAD(&vma->closed_link);
+			__i915_vma_put(vma);
+			i915_gem_object_unlock(obj);
+		} else {
+			/* back you go.. */
+			spin_lock_irq(&gt->closed_lock);
+			list_move(&vma->closed_link, &gt->closed_vma);
+			spin_unlock_irq(&gt->closed_lock);
+		}
 
 		i915_gem_object_put(obj);
 		i915_vm_close(vm);
@@ -1592,6 +1600,7 @@ int i915_vma_move_to_active(struct i915_vma *vma,
 void __i915_vma_evict(struct i915_vma *vma)
 {
 	GEM_BUG_ON(i915_vma_is_pinned(vma));
+	assert_object_held_shared(vma->obj);
 
 	if (i915_vma_is_map_and_fenceable(vma)) {
 		/* Force a pagefault for domain tracking on next user access */
@@ -1637,6 +1646,7 @@ int __i915_vma_unbind(struct i915_vma *vma)
 	int ret;
 
 	lockdep_assert_held(&vma->vm->mutex);
+	assert_object_held_shared(vma->obj);
 
 	if (!drm_mm_node_allocated(&vma->node))
 		return 0;
@@ -1668,6 +1678,8 @@ int i915_vma_unbind(struct i915_vma *vma)
 	intel_wakeref_t wakeref = 0;
 	int err;
 
+	assert_object_held_shared(vma->obj);
+
 	/* Optimistic wait before taking the mutex */
 	err = i915_vma_sync(vma);
 	if (err)
@@ -1698,6 +1710,17 @@ int i915_vma_unbind(struct i915_vma *vma)
 	return err;
 }
 
+int i915_vma_unbind_unlocked(struct i915_vma *vma)
+{
+	int err;
+
+	i915_gem_object_lock(vma->obj, NULL);
+	err = i915_vma_unbind(vma);
+	i915_gem_object_unlock(vma->obj);
+
+	return err;
+}
+
 struct i915_vma *i915_vma_make_unshrinkable(struct i915_vma *vma)
 {
 	i915_gem_object_make_unshrinkable(vma->obj);
diff --git a/drivers/gpu/drm/i915/i915_vma.h b/drivers/gpu/drm/i915/i915_vma.h
index e4dc5187c8ad..74357ea71a9f 100644
--- a/drivers/gpu/drm/i915/i915_vma.h
+++ b/drivers/gpu/drm/i915/i915_vma.h
@@ -207,6 +207,7 @@ void i915_vma_revoke_mmap(struct i915_vma *vma);
 void __i915_vma_evict(struct i915_vma *vma);
 int __i915_vma_unbind(struct i915_vma *vma);
 int __must_check i915_vma_unbind(struct i915_vma *vma);
+int __must_check i915_vma_unbind_unlocked(struct i915_vma *vma);
 void i915_vma_unlink_ctx(struct i915_vma *vma);
 void i915_vma_close(struct i915_vma *vma);
 void i915_vma_reopen(struct i915_vma *vma);
diff --git a/drivers/gpu/drm/i915/selftests/i915_gem_gtt.c b/drivers/gpu/drm/i915/selftests/i915_gem_gtt.c
index 4ffd59ec67da..de850e4c782a 100644
--- a/drivers/gpu/drm/i915/selftests/i915_gem_gtt.c
+++ b/drivers/gpu/drm/i915/selftests/i915_gem_gtt.c
@@ -385,7 +385,7 @@ static void close_object_list(struct list_head *objects,
 
 		vma = i915_vma_instance(obj, vm, NULL);
 		if (!IS_ERR(vma))
-			ignored = i915_vma_unbind(vma);
+			ignored = i915_vma_unbind_unlocked(vma);
 
 		list_del(&obj->st_link);
 		i915_gem_object_put(obj);
@@ -496,7 +496,7 @@ static int fill_hole(struct i915_address_space *vm,
 						goto err;
 					}
 
-					err = i915_vma_unbind(vma);
+					err = i915_vma_unbind_unlocked(vma);
 					if (err) {
 						pr_err("%s(%s) (forward) unbind of vma.node=%llx + %llx failed with err=%d\n",
 						       __func__, p->name, vma->node.start, vma->node.size,
@@ -569,7 +569,7 @@ static int fill_hole(struct i915_address_space *vm,
 						goto err;
 					}
 
-					err = i915_vma_unbind(vma);
+					err = i915_vma_unbind_unlocked(vma);
 					if (err) {
 						pr_err("%s(%s) (backward) unbind of vma.node=%llx + %llx failed with err=%d\n",
 						       __func__, p->name, vma->node.start, vma->node.size,
@@ -655,7 +655,7 @@ static int walk_hole(struct i915_address_space *vm,
 				goto err_put;
 			}
 
-			err = i915_vma_unbind(vma);
+			err = i915_vma_unbind_unlocked(vma);
 			if (err) {
 				pr_err("%s unbind failed at %llx + %llx  with err=%d\n",
 				       __func__, addr, vma->size, err);
@@ -732,13 +732,13 @@ static int pot_hole(struct i915_address_space *vm,
 				pr_err("%s incorrect at %llx + %llx\n",
 				       __func__, addr, vma->size);
 				i915_vma_unpin(vma);
-				err = i915_vma_unbind(vma);
+				err = i915_vma_unbind_unlocked(vma);
 				err = -EINVAL;
 				goto err_obj;
 			}
 
 			i915_vma_unpin(vma);
-			err = i915_vma_unbind(vma);
+			err = i915_vma_unbind_unlocked(vma);
 			GEM_BUG_ON(err);
 		}
 
@@ -832,13 +832,13 @@ static int drunk_hole(struct i915_address_space *vm,
 				pr_err("%s incorrect at %llx + %llx\n",
 				       __func__, addr, BIT_ULL(size));
 				i915_vma_unpin(vma);
-				err = i915_vma_unbind(vma);
+				err = i915_vma_unbind_unlocked(vma);
 				err = -EINVAL;
 				goto err_obj;
 			}
 
 			i915_vma_unpin(vma);
-			err = i915_vma_unbind(vma);
+			err = i915_vma_unbind_unlocked(vma);
 			GEM_BUG_ON(err);
 
 			if (igt_timeout(end_time,
@@ -906,7 +906,7 @@ static int __shrink_hole(struct i915_address_space *vm,
 			pr_err("%s incorrect at %llx + %llx\n",
 			       __func__, addr, size);
 			i915_vma_unpin(vma);
-			err = i915_vma_unbind(vma);
+			err = i915_vma_unbind_unlocked(vma);
 			err = -EINVAL;
 			break;
 		}
@@ -1465,7 +1465,7 @@ static int igt_gtt_reserve(void *arg)
 			goto out;
 		}
 
-		err = i915_vma_unbind(vma);
+		err = i915_vma_unbind_unlocked(vma);
 		if (err) {
 			pr_err("i915_vma_unbind failed with err=%d!\n", err);
 			goto out;
@@ -1647,7 +1647,7 @@ static int igt_gtt_insert(void *arg)
 		GEM_BUG_ON(!drm_mm_node_allocated(&vma->node));
 		offset = vma->node.start;
 
-		err = i915_vma_unbind(vma);
+		err = i915_vma_unbind_unlocked(vma);
 		if (err) {
 			pr_err("i915_vma_unbind failed with err=%d!\n", err);
 			goto out;
diff --git a/drivers/gpu/drm/i915/selftests/i915_vma.c b/drivers/gpu/drm/i915/selftests/i915_vma.c
index 1f10fe36619b..e92961cb737d 100644
--- a/drivers/gpu/drm/i915/selftests/i915_vma.c
+++ b/drivers/gpu/drm/i915/selftests/i915_vma.c
@@ -340,7 +340,7 @@ static int igt_vma_pin1(void *arg)
 
 		if (!err) {
 			i915_vma_unpin(vma);
-			err = i915_vma_unbind(vma);
+			err = i915_vma_unbind_unlocked(vma);
 			if (err) {
 				pr_err("Failed to unbind single page from GGTT, err=%d\n", err);
 				goto out;
-- 
2.33.0



More information about the Intel-gfx-trybot mailing list