[PATCH 01/19] fence: add debugging lines to fence_is_signaled for the callback

[PATCH 01/19] fence: add debugging lines to fence_is_signaled for the callback

[Maarten Lankhorst]

fence_is_signaled callback should support being run in
atomic context, but not in irq context.

Signed-off-by: Maarten Lankhorst <maarten.lankhorst@canonical.com>
---
 include/linux/fence.h |   23 +++++++++++++++++++----
 1 file changed, 19 insertions(+), 4 deletions(-)

diff --git a/include/linux/fence.h b/include/linux/fence.h
index d174585b874b..c1a4519ba2f5 100644
--- a/include/linux/fence.h
+++ b/include/linux/fence.h
@@ -143,6 +143,7 @@ struct fence_cb {
  * the second time will be a noop since it was already signaled.
  *
  * Notes on signaled:
+ * Called with interrupts enabled, and never from interrupt context.
  * May set fence->status if returning true.
  *
  * Notes on wait:
@@ -268,15 +269,29 @@ fence_is_signaled_locked(struct fence *fence)
 static inline bool
 fence_is_signaled(struct fence *fence)
 {
+	bool ret;
+
 	if (test_bit(FENCE_FLAG_SIGNALED_BIT, &fence->flags))
 		return true;
 
-	if (fence->ops->signaled && fence->ops->signaled(fence)) {
+	if (!fence->ops->signaled)
+		return false;
+
+	if (config_enabled(CONFIG_PROVE_LOCKING))
+		WARN_ON(in_interrupt() || irqs_disabled());
+
+	if (config_enabled(CONFIG_DEBUG_ATOMIC_SLEEP))
+		preempt_disable();
+
+	ret = fence->ops->signaled(fence);
+
+	if (config_enabled(CONFIG_DEBUG_ATOMIC_SLEEP))
+		preempt_enable();
+
+	if (ret)
 		fence_signal(fence);
-		return true;
-	}
 
-	return false;
+	return ret;
 }
 
 /**

[PATCH 02/19] drm/ttm: add interruptible parameter to ttm_eu_reserve_buffers

[Maarten Lankhorst]

It seems some drivers really want this as a parameter,
like vmwgfx.

Signed-off-by: Maarten Lankhorst <maarten.lankhorst@canonical.com>
---
 drivers/gpu/drm/qxl/qxl_release.c        |    2 +-
 drivers/gpu/drm/radeon/radeon_object.c   |    2 +-
 drivers/gpu/drm/radeon/radeon_uvd.c      |    2 +-
 drivers/gpu/drm/radeon/radeon_vm.c       |    2 +-
 drivers/gpu/drm/ttm/ttm_execbuf_util.c   |   22 +++++++++++++---------
 drivers/gpu/drm/vmwgfx/vmwgfx_execbuf.c  |    7 ++-----
 drivers/gpu/drm/vmwgfx/vmwgfx_resource.c |    2 +-
 include/drm/ttm/ttm_execbuf_util.h       |    9 +++++----
 8 files changed, 25 insertions(+), 23 deletions(-)

diff --git a/drivers/gpu/drm/qxl/qxl_release.c b/drivers/gpu/drm/qxl/qxl_release.c
index 14e776f1d14e..2b43e5deb051 100644
--- a/drivers/gpu/drm/qxl/qxl_release.c
+++ b/drivers/gpu/drm/qxl/qxl_release.c
@@ -159,7 +159,7 @@ int qxl_release_reserve_list(struct qxl_release *release, bool no_intr)
 	if (list_is_singular(&release->bos))
 		return 0;
 
-	ret = ttm_eu_reserve_buffers(&release->ticket, &release->bos);
+	ret = ttm_eu_reserve_buffers(&release->ticket, &release->bos, !no_intr);
 	if (ret)
 		return ret;
 
diff --git a/drivers/gpu/drm/radeon/radeon_object.c b/drivers/gpu/drm/radeon/radeon_object.c
index 6c717b257d6d..a3ed725ea641 100644
--- a/drivers/gpu/drm/radeon/radeon_object.c
+++ b/drivers/gpu/drm/radeon/radeon_object.c
@@ -438,7 +438,7 @@ int radeon_bo_list_validate(struct radeon_device *rdev,
 	u64 bytes_moved = 0, initial_bytes_moved;
 	u64 bytes_moved_threshold = radeon_bo_get_threshold_for_moves(rdev);
 
-	r = ttm_eu_reserve_buffers(ticket, head);
+	r = ttm_eu_reserve_buffers(ticket, head, true);
 	if (unlikely(r != 0)) {
 		return r;
 	}
diff --git a/drivers/gpu/drm/radeon/radeon_uvd.c b/drivers/gpu/drm/radeon/radeon_uvd.c
index a4ad270e8261..67b2a367df40 100644
--- a/drivers/gpu/drm/radeon/radeon_uvd.c
+++ b/drivers/gpu/drm/radeon/radeon_uvd.c
@@ -620,7 +620,7 @@ static int radeon_uvd_send_msg(struct radeon_device *rdev,
 	INIT_LIST_HEAD(&head);
 	list_add(&tv.head, &head);
 
-	r = ttm_eu_reserve_buffers(&ticket, &head);
+	r = ttm_eu_reserve_buffers(&ticket, &head, true);
 	if (r)
 		return r;
 
diff --git a/drivers/gpu/drm/radeon/radeon_vm.c b/drivers/gpu/drm/radeon/radeon_vm.c
index 725d3669014f..0a0d0291af0d 100644
--- a/drivers/gpu/drm/radeon/radeon_vm.c
+++ b/drivers/gpu/drm/radeon/radeon_vm.c
@@ -365,7 +365,7 @@ static int radeon_vm_clear_bo(struct radeon_device *rdev,
         INIT_LIST_HEAD(&head);
         list_add(&tv.head, &head);
 
-        r = ttm_eu_reserve_buffers(&ticket, &head);
+        r = ttm_eu_reserve_buffers(&ticket, &head, true);
         if (r)
 		return r;
 
diff --git a/drivers/gpu/drm/ttm/ttm_execbuf_util.c b/drivers/gpu/drm/ttm/ttm_execbuf_util.c
index e8dac8758528..39a11bbd2bac 100644
--- a/drivers/gpu/drm/ttm/ttm_execbuf_util.c
+++ b/drivers/gpu/drm/ttm/ttm_execbuf_util.c
@@ -112,7 +112,7 @@ EXPORT_SYMBOL(ttm_eu_backoff_reservation);
  */
 
 int ttm_eu_reserve_buffers(struct ww_acquire_ctx *ticket,
-			   struct list_head *list)
+			   struct list_head *list, bool intr)
 {
 	struct ttm_bo_global *glob;
 	struct ttm_validate_buffer *entry;
@@ -140,7 +140,7 @@ retry:
 		if (entry->reserved)
 			continue;
 
-		ret = __ttm_bo_reserve(bo, true, (ticket == NULL), true,
+		ret = __ttm_bo_reserve(bo, intr, (ticket == NULL), true,
 				       ticket);
 
 		if (ret == -EDEADLK) {
@@ -153,13 +153,17 @@ retry:
 			ttm_eu_backoff_reservation_locked(list);
 			spin_unlock(&glob->lru_lock);
 			ttm_eu_list_ref_sub(list);
-			ret = ww_mutex_lock_slow_interruptible(&bo->resv->lock,
-							       ticket);
-			if (unlikely(ret != 0)) {
-				if (ret == -EINTR)
-					ret = -ERESTARTSYS;
-				goto err_fini;
-			}
+
+			if (intr) {
+				ret = ww_mutex_lock_slow_interruptible(&bo->resv->lock,
+								       ticket);
+				if (unlikely(ret != 0)) {
+					if (ret == -EINTR)
+						ret = -ERESTARTSYS;
+					goto err_fini;
+				}
+			} else
+				ww_mutex_lock_slow(&bo->resv->lock, ticket);
 
 			entry->reserved = true;
 			if (unlikely(atomic_read(&bo->cpu_writers) > 0)) {
diff --git a/drivers/gpu/drm/vmwgfx/vmwgfx_execbuf.c b/drivers/gpu/drm/vmwgfx/vmwgfx_execbuf.c
index 7bfdaa163a33..24f067bf438d 100644
--- a/drivers/gpu/drm/vmwgfx/vmwgfx_execbuf.c
+++ b/drivers/gpu/drm/vmwgfx/vmwgfx_execbuf.c
@@ -2496,7 +2496,7 @@ int vmw_execbuf_process(struct drm_file *file_priv,
 	if (unlikely(ret != 0))
 		goto out_err_nores;
 
-	ret = ttm_eu_reserve_buffers(&ticket, &sw_context->validate_nodes);
+	ret = ttm_eu_reserve_buffers(&ticket, &sw_context->validate_nodes, true);
 	if (unlikely(ret != 0))
 		goto out_err;
 
@@ -2684,10 +2684,7 @@ void __vmw_execbuf_release_pinned_bo(struct vmw_private *dev_priv,
 	query_val.bo = ttm_bo_reference(dev_priv->dummy_query_bo);
 	list_add_tail(&query_val.head, &validate_list);
 
-	do {
-		ret = ttm_eu_reserve_buffers(&ticket, &validate_list);
-	} while (ret == -ERESTARTSYS);
-
+	ret = ttm_eu_reserve_buffers(&ticket, &validate_list, false);
 	if (unlikely(ret != 0)) {
 		vmw_execbuf_unpin_panic(dev_priv);
 		goto out_no_reserve;
diff --git a/drivers/gpu/drm/vmwgfx/vmwgfx_resource.c b/drivers/gpu/drm/vmwgfx/vmwgfx_resource.c
index a432c0db257c..9d8c0a3f391c 100644
--- a/drivers/gpu/drm/vmwgfx/vmwgfx_resource.c
+++ b/drivers/gpu/drm/vmwgfx/vmwgfx_resource.c
@@ -1215,7 +1215,7 @@ vmw_resource_check_buffer(struct vmw_resource *res,
 	INIT_LIST_HEAD(&val_list);
 	val_buf->bo = ttm_bo_reference(&res->backup->base);
 	list_add_tail(&val_buf->head, &val_list);
-	ret = ttm_eu_reserve_buffers(NULL, &val_list);
+	ret = ttm_eu_reserve_buffers(NULL, &val_list, interruptible);
 	if (unlikely(ret != 0))
 		goto out_no_reserve;
 
diff --git a/include/drm/ttm/ttm_execbuf_util.h b/include/drm/ttm/ttm_execbuf_util.h
index 16db7d01a336..fd95fd569ca3 100644
--- a/include/drm/ttm/ttm_execbuf_util.h
+++ b/include/drm/ttm/ttm_execbuf_util.h
@@ -73,6 +73,7 @@ extern void ttm_eu_backoff_reservation(struct ww_acquire_ctx *ticket,
  * @ticket:  [out] ww_acquire_ctx filled in by call, or NULL if only
  *           non-blocking reserves should be tried.
  * @list:    thread private list of ttm_validate_buffer structs.
+ * @intr:    should the wait be interruptible
  *
  * Tries to reserve bos pointed to by the list entries for validation.
  * If the function returns 0, all buffers are marked as "unfenced",
@@ -84,9 +85,9 @@ extern void ttm_eu_backoff_reservation(struct ww_acquire_ctx *ticket,
  * CPU write reservations to be cleared, and for other threads to
  * unreserve their buffers.
  *
- * This function may return -ERESTART or -EAGAIN if the calling process
- * receives a signal while waiting. In that case, no buffers on the list
- * will be reserved upon return.
+ * If intr is set to true, this function may return -ERESTARTSYS if the
+ * calling process receives a signal while waiting. In that case, no
+ * buffers on the list will be reserved upon return.
  *
  * Buffers reserved by this function should be unreserved by
  * a call to either ttm_eu_backoff_reservation() or
@@ -95,7 +96,7 @@ extern void ttm_eu_backoff_reservation(struct ww_acquire_ctx *ticket,
  */
 
 extern int ttm_eu_reserve_buffers(struct ww_acquire_ctx *ticket,
-				  struct list_head *list);
+				  struct list_head *list, bool intr);
 
 /**
  * function ttm_eu_fence_buffer_objects.

[PATCH 03/19] drm/ttm: kill off some members to ttm_validate_buffer

[Maarten Lankhorst]

This reorders the list to keep track of what buffers are reserved,
so previous members are always unreserved.

This gets rid of some bookkeeping that's no longer needed,
while simplifying the code some.

Signed-off-by: Maarten Lankhorst <maarten.lankhorst@canonical.com>
---
 drivers/gpu/drm/qxl/qxl_release.c       |    1 
 drivers/gpu/drm/ttm/ttm_execbuf_util.c  |  142 +++++++++++--------------------
 drivers/gpu/drm/vmwgfx/vmwgfx_execbuf.c |    1 
 include/drm/ttm/ttm_execbuf_util.h      |    3 -
 4 files changed, 50 insertions(+), 97 deletions(-)

diff --git a/drivers/gpu/drm/qxl/qxl_release.c b/drivers/gpu/drm/qxl/qxl_release.c
index 2b43e5deb051..e85c4d274dc0 100644
--- a/drivers/gpu/drm/qxl/qxl_release.c
+++ b/drivers/gpu/drm/qxl/qxl_release.c
@@ -350,7 +350,6 @@ void qxl_release_fence_buffer_objects(struct qxl_release *release)
 
 		ttm_bo_add_to_lru(bo);
 		__ttm_bo_unreserve(bo);
-		entry->reserved = false;
 	}
 	spin_unlock(&bdev->fence_lock);
 	spin_unlock(&glob->lru_lock);
diff --git a/drivers/gpu/drm/ttm/ttm_execbuf_util.c b/drivers/gpu/drm/ttm/ttm_execbuf_util.c
index 39a11bbd2bac..6db47a72667e 100644
--- a/drivers/gpu/drm/ttm/ttm_execbuf_util.c
+++ b/drivers/gpu/drm/ttm/ttm_execbuf_util.c
@@ -32,20 +32,12 @@
 #include <linux/sched.h>
 #include <linux/module.h>
 
-static void ttm_eu_backoff_reservation_locked(struct list_head *list)
+static void ttm_eu_backoff_reservation_reverse(struct list_head *list,
+					      struct ttm_validate_buffer *entry)
 {
-	struct ttm_validate_buffer *entry;
-
-	list_for_each_entry(entry, list, head) {
+	list_for_each_entry_continue_reverse(entry, list, head) {
 		struct ttm_buffer_object *bo = entry->bo;
-		if (!entry->reserved)
-			continue;
 
-		entry->reserved = false;
-		if (entry->removed) {
-			ttm_bo_add_to_lru(bo);
-			entry->removed = false;
-		}
 		__ttm_bo_unreserve(bo);
 	}
 }
@@ -56,27 +48,9 @@ static void ttm_eu_del_from_lru_locked(struct list_head *list)
 
 	list_for_each_entry(entry, list, head) {
 		struct ttm_buffer_object *bo = entry->bo;
-		if (!entry->reserved)
-			continue;
+		unsigned put_count = ttm_bo_del_from_lru(bo);
 
-		if (!entry->removed) {
-			entry->put_count = ttm_bo_del_from_lru(bo);
-			entry->removed = true;
-		}
-	}
-}
-
-static void ttm_eu_list_ref_sub(struct list_head *list)
-{
-	struct ttm_validate_buffer *entry;
-
-	list_for_each_entry(entry, list, head) {
-		struct ttm_buffer_object *bo = entry->bo;
-
-		if (entry->put_count) {
-			ttm_bo_list_ref_sub(bo, entry->put_count, true);
-			entry->put_count = 0;
-		}
+		ttm_bo_list_ref_sub(bo, put_count, true);
 	}
 }
 
@@ -91,11 +65,18 @@ void ttm_eu_backoff_reservation(struct ww_acquire_ctx *ticket,
 
 	entry = list_first_entry(list, struct ttm_validate_buffer, head);
 	glob = entry->bo->glob;
+
 	spin_lock(&glob->lru_lock);
-	ttm_eu_backoff_reservation_locked(list);
+	list_for_each_entry(entry, list, head) {
+		struct ttm_buffer_object *bo = entry->bo;
+
+		ttm_bo_add_to_lru(bo);
+		__ttm_bo_unreserve(bo);
+	}
+	spin_unlock(&glob->lru_lock);
+
 	if (ticket)
 		ww_acquire_fini(ticket);
-	spin_unlock(&glob->lru_lock);
 }
 EXPORT_SYMBOL(ttm_eu_backoff_reservation);
 
@@ -121,64 +102,55 @@ int ttm_eu_reserve_buffers(struct ww_acquire_ctx *ticket,
 	if (list_empty(list))
 		return 0;
 
-	list_for_each_entry(entry, list, head) {
-		entry->reserved = false;
-		entry->put_count = 0;
-		entry->removed = false;
-	}
-
 	entry = list_first_entry(list, struct ttm_validate_buffer, head);
 	glob = entry->bo->glob;
 
 	if (ticket)
 		ww_acquire_init(ticket, &reservation_ww_class);
-retry:
+
 	list_for_each_entry(entry, list, head) {
 		struct ttm_buffer_object *bo = entry->bo;
 
-		/* already slowpath reserved? */
-		if (entry->reserved)
-			continue;
-
 		ret = __ttm_bo_reserve(bo, intr, (ticket == NULL), true,
 				       ticket);
+		if (!ret && unlikely(atomic_read(&bo->cpu_writers) > 0)) {
+			__ttm_bo_unreserve(bo);
 
-		if (ret == -EDEADLK) {
-			/* uh oh, we lost out, drop every reservation and try
-			 * to only reserve this buffer, then start over if
-			 * this succeeds.
-			 */
-			BUG_ON(ticket == NULL);
-			spin_lock(&glob->lru_lock);
-			ttm_eu_backoff_reservation_locked(list);
-			spin_unlock(&glob->lru_lock);
-			ttm_eu_list_ref_sub(list);
-
-			if (intr) {
-				ret = ww_mutex_lock_slow_interruptible(&bo->resv->lock,
-								       ticket);
-				if (unlikely(ret != 0)) {
-					if (ret == -EINTR)
-						ret = -ERESTARTSYS;
-					goto err_fini;
-				}
-			} else
-				ww_mutex_lock_slow(&bo->resv->lock, ticket);
-
-			entry->reserved = true;
-			if (unlikely(atomic_read(&bo->cpu_writers) > 0)) {
-				ret = -EBUSY;
-				goto err;
-			}
-			goto retry;
-		} else if (ret)
-			goto err;
-
-		entry->reserved = true;
-		if (unlikely(atomic_read(&bo->cpu_writers) > 0)) {
 			ret = -EBUSY;
-			goto err;
 		}
+
+		if (!ret)
+			continue;
+
+		/* uh oh, we lost out, drop every reservation and try
+		 * to only reserve this buffer, then start over if
+		 * this succeeds.
+		 */
+		ttm_eu_backoff_reservation_reverse(list, entry);
+
+		if (ret == -EDEADLK && intr) {
+			ret = ww_mutex_lock_slow_interruptible(&bo->resv->lock,
+							       ticket);
+		} else if (ret == -EDEADLK) {
+			ww_mutex_lock_slow(&bo->resv->lock, ticket);
+			ret = 0;
+		}
+
+		if (unlikely(ret != 0)) {
+			if (ret == -EINTR)
+				ret = -ERESTARTSYS;
+			if (ticket) {
+				ww_acquire_done(ticket);
+				ww_acquire_fini(ticket);
+			}
+			return ret;
+		}
+
+		/* move this item to the front of the list,
+		 * forces correct iteration of the loop without keeping track
+		 */
+		list_del(&entry->head);
+		list_add(&entry->head, list);
 	}
 
 	if (ticket)
@@ -186,20 +158,7 @@ retry:
 	spin_lock(&glob->lru_lock);
 	ttm_eu_del_from_lru_locked(list);
 	spin_unlock(&glob->lru_lock);
-	ttm_eu_list_ref_sub(list);
 	return 0;
-
-err:
-	spin_lock(&glob->lru_lock);
-	ttm_eu_backoff_reservation_locked(list);
-	spin_unlock(&glob->lru_lock);
-	ttm_eu_list_ref_sub(list);
-err_fini:
-	if (ticket) {
-		ww_acquire_done(ticket);
-		ww_acquire_fini(ticket);
-	}
-	return ret;
 }
 EXPORT_SYMBOL(ttm_eu_reserve_buffers);
 
@@ -229,7 +188,6 @@ void ttm_eu_fence_buffer_objects(struct ww_acquire_ctx *ticket,
 		bo->sync_obj = driver->sync_obj_ref(sync_obj);
 		ttm_bo_add_to_lru(bo);
 		__ttm_bo_unreserve(bo);
-		entry->reserved = false;
 	}
 	spin_unlock(&bdev->fence_lock);
 	spin_unlock(&glob->lru_lock);
diff --git a/drivers/gpu/drm/vmwgfx/vmwgfx_execbuf.c b/drivers/gpu/drm/vmwgfx/vmwgfx_execbuf.c
index 24f067bf438d..b19b2b980cb4 100644
--- a/drivers/gpu/drm/vmwgfx/vmwgfx_execbuf.c
+++ b/drivers/gpu/drm/vmwgfx/vmwgfx_execbuf.c
@@ -346,7 +346,6 @@ static int vmw_bo_to_validate_list(struct vmw_sw_context *sw_context,
 		++sw_context->cur_val_buf;
 		val_buf = &vval_buf->base;
 		val_buf->bo = ttm_bo_reference(bo);
-		val_buf->reserved = false;
 		list_add_tail(&val_buf->head, &sw_context->validate_nodes);
 		vval_buf->validate_as_mob = validate_as_mob;
 	}
diff --git a/include/drm/ttm/ttm_execbuf_util.h b/include/drm/ttm/ttm_execbuf_util.h
index fd95fd569ca3..8490cb8ee0d8 100644
--- a/include/drm/ttm/ttm_execbuf_util.h
+++ b/include/drm/ttm/ttm_execbuf_util.h
@@ -48,9 +48,6 @@
 struct ttm_validate_buffer {
 	struct list_head head;
 	struct ttm_buffer_object *bo;
-	bool reserved;
-	bool removed;
-	int put_count;
 	void *old_sync_obj;
 };
 

[PATCH 04/19] drm/nouveau: add reservation to nouveau_gem_ioctl_cpu_prep

[Maarten Lankhorst]

Apart from some code inside ttm itself and nouveau_bo_vma_del,
this is the only place where ttm_bo_wait is used without a reservation.
Fix this so we can remove the fence_lock later on.

After the switch to rcu the reservation lock will be
removed again.

Signed-off-by: Maarten Lankhorst <maarten.lankhorst@canonical.com>
---
 drivers/gpu/drm/nouveau/nouveau_gem.c |   22 ++++++++++++++++++----
 1 file changed, 18 insertions(+), 4 deletions(-)

diff --git a/drivers/gpu/drm/nouveau/nouveau_gem.c b/drivers/gpu/drm/nouveau/nouveau_gem.c
index df9d451afdcd..0da3740f150f 100644
--- a/drivers/gpu/drm/nouveau/nouveau_gem.c
+++ b/drivers/gpu/drm/nouveau/nouveau_gem.c
@@ -886,17 +886,31 @@ nouveau_gem_ioctl_cpu_prep(struct drm_device *dev, void *data,
 	struct drm_gem_object *gem;
 	struct nouveau_bo *nvbo;
 	bool no_wait = !!(req->flags & NOUVEAU_GEM_CPU_PREP_NOWAIT);
-	int ret = -EINVAL;
+	int ret;
+	struct nouveau_fence *fence = NULL;
 
 	gem = drm_gem_object_lookup(dev, file_priv, req->handle);
 	if (!gem)
 		return -ENOENT;
 	nvbo = nouveau_gem_object(gem);
 
-	spin_lock(&nvbo->bo.bdev->fence_lock);
-	ret = ttm_bo_wait(&nvbo->bo, true, true, no_wait);
-	spin_unlock(&nvbo->bo.bdev->fence_lock);
+	ret = ttm_bo_reserve(&nvbo->bo, true, false, false, 0);
+	if (!ret) {
+		spin_lock(&nvbo->bo.bdev->fence_lock);
+		ret = ttm_bo_wait(&nvbo->bo, true, true, true);
+		if (!no_wait && ret)
+			fence = nouveau_fence_ref(nvbo->bo.sync_obj);
+		spin_unlock(&nvbo->bo.bdev->fence_lock);
+
+		ttm_bo_unreserve(&nvbo->bo);
+	}
 	drm_gem_object_unreference_unlocked(gem);
+
+	if (fence) {
+		ret = nouveau_fence_wait(fence, true, no_wait);
+		nouveau_fence_unref(&fence);
+	}
+
 	return ret;
 }
 

[PATCH 05/19] drm/nouveau: require reservations for nouveau_fence_sync and nouveau_bo_fence

[Maarten Lankhorst]

This will ensure we always hold the required lock when calling those functions.
---
 drivers/gpu/drm/nouveau/nouveau_bo.c      |    2 ++
 drivers/gpu/drm/nouveau/nouveau_display.c |   17 +++++++++++++----
 2 files changed, 15 insertions(+), 4 deletions(-)

diff --git a/drivers/gpu/drm/nouveau/nouveau_bo.c b/drivers/gpu/drm/nouveau/nouveau_bo.c
index ba29a701ca1d..62d79492c7c5 100644
--- a/drivers/gpu/drm/nouveau/nouveau_bo.c
+++ b/drivers/gpu/drm/nouveau/nouveau_bo.c
@@ -1431,6 +1431,8 @@ nouveau_bo_fence(struct nouveau_bo *nvbo, struct nouveau_fence *fence)
 	struct nouveau_fence *new_fence = nouveau_fence_ref(fence);
 	struct nouveau_fence *old_fence = NULL;
 
+	lockdep_assert_held(&nvbo->bo.resv->lock.base);
+
 	spin_lock(&nvbo->bo.bdev->fence_lock);
 	old_fence = nvbo->bo.sync_obj;
 	nvbo->bo.sync_obj = new_fence;
diff --git a/drivers/gpu/drm/nouveau/nouveau_display.c b/drivers/gpu/drm/nouveau/nouveau_display.c
index 47ad74255bf1..826b66c44235 100644
--- a/drivers/gpu/drm/nouveau/nouveau_display.c
+++ b/drivers/gpu/drm/nouveau/nouveau_display.c
@@ -716,6 +716,9 @@ nouveau_crtc_page_flip(struct drm_crtc *crtc, struct drm_framebuffer *fb,
 	}
 
 	mutex_lock(&chan->cli->mutex);
+	ret = ttm_bo_reserve(&new_bo->bo, true, false, false, NULL);
+	if (ret)
+		goto fail_unpin;
 
 	/* synchronise rendering channel with the kernel's channel */
 	spin_lock(&new_bo->bo.bdev->fence_lock);
@@ -723,12 +726,18 @@ nouveau_crtc_page_flip(struct drm_crtc *crtc, struct drm_framebuffer *fb,
 	spin_unlock(&new_bo->bo.bdev->fence_lock);
 	ret = nouveau_fence_sync(fence, chan);
 	nouveau_fence_unref(&fence);
-	if (ret)
+	if (ret) {
+		ttm_bo_unreserve(&new_bo->bo);
 		goto fail_unpin;
+	}
 
-	ret = ttm_bo_reserve(&old_bo->bo, true, false, false, NULL);
-	if (ret)
-		goto fail_unpin;
+	if (new_bo != old_bo) {
+		ttm_bo_unreserve(&new_bo->bo);
+
+		ret = ttm_bo_reserve(&old_bo->bo, true, false, false, NULL);
+		if (ret)
+			goto fail_unpin;
+	}
 
 	/* Initialize a page flip struct */
 	*s = (struct nouveau_page_flip_state)

[PATCH 06/19] drm/ttm: call ttm_bo_wait while inside a reservation

[Maarten Lankhorst]

This is the last remaining function that doesn't use the reservation
lock completely to fence off access to a buffer.
---
 drivers/gpu/drm/ttm/ttm_bo.c |   25 ++++++++++++-------------
 1 file changed, 12 insertions(+), 13 deletions(-)

diff --git a/drivers/gpu/drm/ttm/ttm_bo.c b/drivers/gpu/drm/ttm/ttm_bo.c
index 3da89d5dab60..df08f18f05ac 100644
--- a/drivers/gpu/drm/ttm/ttm_bo.c
+++ b/drivers/gpu/drm/ttm/ttm_bo.c
@@ -502,17 +502,6 @@ static int ttm_bo_cleanup_refs_and_unlock(struct ttm_buffer_object *bo,
 		if (ret)
 			return ret;
 
-		/*
-		 * remove sync_obj with ttm_bo_wait, the wait should be
-		 * finished, and no new wait object should have been added.
-		 */
-		spin_lock(&bdev->fence_lock);
-		ret = ttm_bo_wait(bo, false, false, true);
-		WARN_ON(ret);
-		spin_unlock(&bdev->fence_lock);
-		if (ret)
-			return ret;
-
 		spin_lock(&glob->lru_lock);
 		ret = __ttm_bo_reserve(bo, false, true, false, NULL);
 
@@ -528,8 +517,16 @@ static int ttm_bo_cleanup_refs_and_unlock(struct ttm_buffer_object *bo,
 			spin_unlock(&glob->lru_lock);
 			return 0;
 		}
-	} else
-		spin_unlock(&bdev->fence_lock);
+
+		/*
+		 * remove sync_obj with ttm_bo_wait, the wait should be
+		 * finished, and no new wait object should have been added.
+		 */
+		spin_lock(&bdev->fence_lock);
+		ret = ttm_bo_wait(bo, false, false, true);
+		WARN_ON(ret);
+	}
+	spin_unlock(&bdev->fence_lock);
 
 	if (ret || unlikely(list_empty(&bo->ddestroy))) {
 		__ttm_bo_unreserve(bo);
@@ -1539,6 +1536,8 @@ int ttm_bo_wait(struct ttm_buffer_object *bo,
 	void *sync_obj;
 	int ret = 0;
 
+	lockdep_assert_held(&bo->resv->lock.base);
+
 	if (likely(bo->sync_obj == NULL))
 		return 0;
 

[PATCH 07/19] drm/ttm: kill fence_lock

[Maarten Lankhorst]

No users are left, kill it off! :D
Conversion to the reservation api is next on the list, after
that the functionality can be restored with rcu.

Signed-off-by: Maarten Lankhorst <maarten.lankhorst@canonical.com>
---
 drivers/gpu/drm/nouveau/nouveau_bo.c      |   25 +++-------
 drivers/gpu/drm/nouveau/nouveau_display.c |    6 --
 drivers/gpu/drm/nouveau/nouveau_gem.c     |   16 +-----
 drivers/gpu/drm/qxl/qxl_cmd.c             |    2 -
 drivers/gpu/drm/qxl/qxl_fence.c           |    4 --
 drivers/gpu/drm/qxl/qxl_object.h          |    2 -
 drivers/gpu/drm/qxl/qxl_release.c         |    2 -
 drivers/gpu/drm/radeon/radeon_display.c   |    9 ++-
 drivers/gpu/drm/radeon/radeon_object.c    |    2 -
 drivers/gpu/drm/ttm/ttm_bo.c              |   75 +++++++----------------------
 drivers/gpu/drm/ttm/ttm_bo_util.c         |    5 --
 drivers/gpu/drm/ttm/ttm_bo_vm.c           |    3 -
 drivers/gpu/drm/ttm/ttm_execbuf_util.c    |    2 -
 drivers/gpu/drm/vmwgfx/vmwgfx_buffer.c    |    4 --
 drivers/gpu/drm/vmwgfx/vmwgfx_resource.c  |   17 ++-----
 include/drm/ttm/ttm_bo_api.h              |    5 --
 include/drm/ttm/ttm_bo_driver.h           |    3 -
 17 files changed, 40 insertions(+), 142 deletions(-)

diff --git a/drivers/gpu/drm/nouveau/nouveau_bo.c b/drivers/gpu/drm/nouveau/nouveau_bo.c
index 62d79492c7c5..2cf5b0572f66 100644
--- a/drivers/gpu/drm/nouveau/nouveau_bo.c
+++ b/drivers/gpu/drm/nouveau/nouveau_bo.c
@@ -1196,9 +1196,7 @@ nouveau_bo_move(struct ttm_buffer_object *bo, bool evict, bool intr,
 	}
 
 	/* Fallback to software copy. */
-	spin_lock(&bo->bdev->fence_lock);
 	ret = ttm_bo_wait(bo, true, intr, no_wait_gpu);
-	spin_unlock(&bo->bdev->fence_lock);
 	if (ret == 0)
 		ret = ttm_bo_move_memcpy(bo, evict, no_wait_gpu, new_mem);
 
@@ -1425,26 +1423,19 @@ nouveau_ttm_tt_unpopulate(struct ttm_tt *ttm)
 	ttm_pool_unpopulate(ttm);
 }
 
+static void
+nouveau_bo_fence_unref(void **sync_obj)
+{
+	nouveau_fence_unref((struct nouveau_fence **)sync_obj);
+}
+
 void
 nouveau_bo_fence(struct nouveau_bo *nvbo, struct nouveau_fence *fence)
 {
-	struct nouveau_fence *new_fence = nouveau_fence_ref(fence);
-	struct nouveau_fence *old_fence = NULL;
-
 	lockdep_assert_held(&nvbo->bo.resv->lock.base);
 
-	spin_lock(&nvbo->bo.bdev->fence_lock);
-	old_fence = nvbo->bo.sync_obj;
-	nvbo->bo.sync_obj = new_fence;
-	spin_unlock(&nvbo->bo.bdev->fence_lock);
-
-	nouveau_fence_unref(&old_fence);
-}
-
-static void
-nouveau_bo_fence_unref(void **sync_obj)
-{
-	nouveau_fence_unref((struct nouveau_fence **)sync_obj);
+	nouveau_bo_fence_unref(&nvbo->bo.sync_obj);
+	nvbo->bo.sync_obj = nouveau_fence_ref(fence);
 }
 
 static void *
diff --git a/drivers/gpu/drm/nouveau/nouveau_display.c b/drivers/gpu/drm/nouveau/nouveau_display.c
index 826b66c44235..7928f8f07334 100644
--- a/drivers/gpu/drm/nouveau/nouveau_display.c
+++ b/drivers/gpu/drm/nouveau/nouveau_display.c
@@ -721,11 +721,7 @@ nouveau_crtc_page_flip(struct drm_crtc *crtc, struct drm_framebuffer *fb,
 		goto fail_unpin;
 
 	/* synchronise rendering channel with the kernel's channel */
-	spin_lock(&new_bo->bo.bdev->fence_lock);
-	fence = nouveau_fence_ref(new_bo->bo.sync_obj);
-	spin_unlock(&new_bo->bo.bdev->fence_lock);
-	ret = nouveau_fence_sync(fence, chan);
-	nouveau_fence_unref(&fence);
+	ret = nouveau_fence_sync(new_bo->bo.sync_obj, chan);
 	if (ret) {
 		ttm_bo_unreserve(&new_bo->bo);
 		goto fail_unpin;
diff --git a/drivers/gpu/drm/nouveau/nouveau_gem.c b/drivers/gpu/drm/nouveau/nouveau_gem.c
index 0da3740f150f..474f67238cef 100644
--- a/drivers/gpu/drm/nouveau/nouveau_gem.c
+++ b/drivers/gpu/drm/nouveau/nouveau_gem.c
@@ -105,9 +105,7 @@ nouveau_gem_object_unmap(struct nouveau_bo *nvbo, struct nouveau_vma *vma)
 	list_del(&vma->head);
 
 	if (mapped) {
-		spin_lock(&nvbo->bo.bdev->fence_lock);
 		fence = nouveau_fence_ref(nvbo->bo.sync_obj);
-		spin_unlock(&nvbo->bo.bdev->fence_lock);
 	}
 
 	if (fence) {
@@ -432,17 +430,11 @@ retry:
 static int
 validate_sync(struct nouveau_channel *chan, struct nouveau_bo *nvbo)
 {
-	struct nouveau_fence *fence = NULL;
+	struct nouveau_fence *fence = nvbo->bo.sync_obj;
 	int ret = 0;
 
-	spin_lock(&nvbo->bo.bdev->fence_lock);
-	fence = nouveau_fence_ref(nvbo->bo.sync_obj);
-	spin_unlock(&nvbo->bo.bdev->fence_lock);
-
-	if (fence) {
+	if (fence)
 		ret = nouveau_fence_sync(fence, chan);
-		nouveau_fence_unref(&fence);
-	}
 
 	return ret;
 }
@@ -661,9 +653,7 @@ nouveau_gem_pushbuf_reloc_apply(struct nouveau_cli *cli,
 				data |= r->vor;
 		}
 
-		spin_lock(&nvbo->bo.bdev->fence_lock);
 		ret = ttm_bo_wait(&nvbo->bo, false, false, false);
-		spin_unlock(&nvbo->bo.bdev->fence_lock);
 		if (ret) {
 			NV_ERROR(cli, "reloc wait_idle failed: %d\n", ret);
 			break;
@@ -896,11 +886,9 @@ nouveau_gem_ioctl_cpu_prep(struct drm_device *dev, void *data,
 
 	ret = ttm_bo_reserve(&nvbo->bo, true, false, false, 0);
 	if (!ret) {
-		spin_lock(&nvbo->bo.bdev->fence_lock);
 		ret = ttm_bo_wait(&nvbo->bo, true, true, true);
 		if (!no_wait && ret)
 			fence = nouveau_fence_ref(nvbo->bo.sync_obj);
-		spin_unlock(&nvbo->bo.bdev->fence_lock);
 
 		ttm_bo_unreserve(&nvbo->bo);
 	}
diff --git a/drivers/gpu/drm/qxl/qxl_cmd.c b/drivers/gpu/drm/qxl/qxl_cmd.c
index eb89653a7a17..45fad7b45486 100644
--- a/drivers/gpu/drm/qxl/qxl_cmd.c
+++ b/drivers/gpu/drm/qxl/qxl_cmd.c
@@ -628,9 +628,7 @@ static int qxl_reap_surf(struct qxl_device *qdev, struct qxl_bo *surf, bool stal
 	if (stall)
 		mutex_unlock(&qdev->surf_evict_mutex);
 
-	spin_lock(&surf->tbo.bdev->fence_lock);
 	ret = ttm_bo_wait(&surf->tbo, true, true, !stall);
-	spin_unlock(&surf->tbo.bdev->fence_lock);
 
 	if (stall)
 		mutex_lock(&qdev->surf_evict_mutex);
diff --git a/drivers/gpu/drm/qxl/qxl_fence.c b/drivers/gpu/drm/qxl/qxl_fence.c
index ae59e91cfb9a..c7248418117d 100644
--- a/drivers/gpu/drm/qxl/qxl_fence.c
+++ b/drivers/gpu/drm/qxl/qxl_fence.c
@@ -60,9 +60,6 @@ int qxl_fence_remove_release(struct qxl_fence *qfence, uint32_t rel_id)
 {
 	void *ret;
 	int retval = 0;
-	struct qxl_bo *bo = container_of(qfence, struct qxl_bo, fence);
-
-	spin_lock(&bo->tbo.bdev->fence_lock);
 
 	ret = radix_tree_delete(&qfence->tree, rel_id);
 	if (ret == qfence)
@@ -71,7 +68,6 @@ int qxl_fence_remove_release(struct qxl_fence *qfence, uint32_t rel_id)
 		DRM_DEBUG("didn't find fence in radix tree for %d\n", rel_id);
 		retval = -ENOENT;
 	}
-	spin_unlock(&bo->tbo.bdev->fence_lock);
 	return retval;
 }
 
diff --git a/drivers/gpu/drm/qxl/qxl_object.h b/drivers/gpu/drm/qxl/qxl_object.h
index 83a423293afd..1edaf5768086 100644
--- a/drivers/gpu/drm/qxl/qxl_object.h
+++ b/drivers/gpu/drm/qxl/qxl_object.h
@@ -76,12 +76,10 @@ static inline int qxl_bo_wait(struct qxl_bo *bo, u32 *mem_type,
 		}
 		return r;
 	}
-	spin_lock(&bo->tbo.bdev->fence_lock);
 	if (mem_type)
 		*mem_type = bo->tbo.mem.mem_type;
 	if (bo->tbo.sync_obj)
 		r = ttm_bo_wait(&bo->tbo, true, true, no_wait);
-	spin_unlock(&bo->tbo.bdev->fence_lock);
 	ttm_bo_unreserve(&bo->tbo);
 	return r;
 }
diff --git a/drivers/gpu/drm/qxl/qxl_release.c b/drivers/gpu/drm/qxl/qxl_release.c
index e85c4d274dc0..4045ba873ab8 100644
--- a/drivers/gpu/drm/qxl/qxl_release.c
+++ b/drivers/gpu/drm/qxl/qxl_release.c
@@ -337,7 +337,6 @@ void qxl_release_fence_buffer_objects(struct qxl_release *release)
 	glob = bo->glob;
 
 	spin_lock(&glob->lru_lock);
-	spin_lock(&bdev->fence_lock);
 
 	list_for_each_entry(entry, &release->bos, head) {
 		bo = entry->bo;
@@ -351,7 +350,6 @@ void qxl_release_fence_buffer_objects(struct qxl_release *release)
 		ttm_bo_add_to_lru(bo);
 		__ttm_bo_unreserve(bo);
 	}
-	spin_unlock(&bdev->fence_lock);
 	spin_unlock(&glob->lru_lock);
 	ww_acquire_fini(&release->ticket);
 }
diff --git a/drivers/gpu/drm/radeon/radeon_display.c b/drivers/gpu/drm/radeon/radeon_display.c
index bf25061c8ac4..3402540aa641 100644
--- a/drivers/gpu/drm/radeon/radeon_display.c
+++ b/drivers/gpu/drm/radeon/radeon_display.c
@@ -462,11 +462,6 @@ static int radeon_crtc_page_flip(struct drm_crtc *crtc,
 	obj = new_radeon_fb->obj;
 	new_rbo = gem_to_radeon_bo(obj);
 
-	spin_lock(&new_rbo->tbo.bdev->fence_lock);
-	if (new_rbo->tbo.sync_obj)
-		work->fence = radeon_fence_ref(new_rbo->tbo.sync_obj);
-	spin_unlock(&new_rbo->tbo.bdev->fence_lock);
-
 	/* pin the new buffer */
 	DRM_DEBUG_DRIVER("flip-ioctl() cur_rbo = %p, new_rbo = %p\n",
 			 work->old_rbo, new_rbo);
@@ -485,6 +480,10 @@ static int radeon_crtc_page_flip(struct drm_crtc *crtc,
 		DRM_ERROR("failed to pin new rbo buffer before flip\n");
 		goto cleanup;
 	}
+
+	if (new_rbo->tbo.sync_obj)
+		work->fence = radeon_fence_ref(new_rbo->tbo.sync_obj);
+
 	radeon_bo_get_tiling_flags(new_rbo, &tiling_flags, NULL);
 	radeon_bo_unreserve(new_rbo);
 
diff --git a/drivers/gpu/drm/radeon/radeon_object.c b/drivers/gpu/drm/radeon/radeon_object.c
index a3ed725ea641..8538aebb6580 100644
--- a/drivers/gpu/drm/radeon/radeon_object.c
+++ b/drivers/gpu/drm/radeon/radeon_object.c
@@ -734,12 +734,10 @@ int radeon_bo_wait(struct radeon_bo *bo, u32 *mem_type, bool no_wait)
 	r = ttm_bo_reserve(&bo->tbo, true, no_wait, false, NULL);
 	if (unlikely(r != 0))
 		return r;
-	spin_lock(&bo->tbo.bdev->fence_lock);
 	if (mem_type)
 		*mem_type = bo->tbo.mem.mem_type;
 	if (bo->tbo.sync_obj)
 		r = ttm_bo_wait(&bo->tbo, true, true, no_wait);
-	spin_unlock(&bo->tbo.bdev->fence_lock);
 	ttm_bo_unreserve(&bo->tbo);
 	return r;
 }
diff --git a/drivers/gpu/drm/ttm/ttm_bo.c b/drivers/gpu/drm/ttm/ttm_bo.c
index df08f18f05ac..c441ac6d546b 100644
--- a/drivers/gpu/drm/ttm/ttm_bo.c
+++ b/drivers/gpu/drm/ttm/ttm_bo.c
@@ -414,24 +414,20 @@ static void ttm_bo_cleanup_refs_or_queue(struct ttm_buffer_object *bo)
 	spin_lock(&glob->lru_lock);
 	ret = __ttm_bo_reserve(bo, false, true, false, NULL);
 
-	spin_lock(&bdev->fence_lock);
-	(void) ttm_bo_wait(bo, false, false, true);
-	if (!ret && !bo->sync_obj) {
-		spin_unlock(&bdev->fence_lock);
-		put_count = ttm_bo_del_from_lru(bo);
+	if (!ret) {
+		(void) ttm_bo_wait(bo, false, false, true);
 
-		spin_unlock(&glob->lru_lock);
-		ttm_bo_cleanup_memtype_use(bo);
+		if (!bo->sync_obj) {
+			put_count = ttm_bo_del_from_lru(bo);
 
-		ttm_bo_list_ref_sub(bo, put_count, true);
+			spin_unlock(&glob->lru_lock);
+			ttm_bo_cleanup_memtype_use(bo);
 
-		return;
-	}
-	if (bo->sync_obj)
-		sync_obj = driver->sync_obj_ref(bo->sync_obj);
-	spin_unlock(&bdev->fence_lock);
+			ttm_bo_list_ref_sub(bo, put_count, true);
 
-	if (!ret) {
+			return;
+		}
+		sync_obj = driver->sync_obj_ref(bo->sync_obj);
 
 		/*
 		 * Make NO_EVICT bos immediately available to
@@ -480,7 +476,6 @@ static int ttm_bo_cleanup_refs_and_unlock(struct ttm_buffer_object *bo,
 	int put_count;
 	int ret;
 
-	spin_lock(&bdev->fence_lock);
 	ret = ttm_bo_wait(bo, false, false, true);
 
 	if (ret && !no_wait_gpu) {
@@ -492,7 +487,6 @@ static int ttm_bo_cleanup_refs_and_unlock(struct ttm_buffer_object *bo,
 		 * no new sync objects can be attached.
 		 */
 		sync_obj = driver->sync_obj_ref(bo->sync_obj);
-		spin_unlock(&bdev->fence_lock);
 
 		__ttm_bo_unreserve(bo);
 		spin_unlock(&glob->lru_lock);
@@ -522,11 +516,9 @@ static int ttm_bo_cleanup_refs_and_unlock(struct ttm_buffer_object *bo,
 		 * remove sync_obj with ttm_bo_wait, the wait should be
 		 * finished, and no new wait object should have been added.
 		 */
-		spin_lock(&bdev->fence_lock);
 		ret = ttm_bo_wait(bo, false, false, true);
 		WARN_ON(ret);
 	}
-	spin_unlock(&bdev->fence_lock);
 
 	if (ret || unlikely(list_empty(&bo->ddestroy))) {
 		__ttm_bo_unreserve(bo);
@@ -664,9 +656,7 @@ static int ttm_bo_evict(struct ttm_buffer_object *bo, bool interruptible,
 	struct ttm_placement placement;
 	int ret = 0;
 
-	spin_lock(&bdev->fence_lock);
 	ret = ttm_bo_wait(bo, false, interruptible, no_wait_gpu);
-	spin_unlock(&bdev->fence_lock);
 
 	if (unlikely(ret != 0)) {
 		if (ret != -ERESTARTSYS) {
@@ -963,7 +953,6 @@ static int ttm_bo_move_buffer(struct ttm_buffer_object *bo,
 {
 	int ret = 0;
 	struct ttm_mem_reg mem;
-	struct ttm_bo_device *bdev = bo->bdev;
 
 	lockdep_assert_held(&bo->resv->lock.base);
 
@@ -972,9 +961,7 @@ static int ttm_bo_move_buffer(struct ttm_buffer_object *bo,
 	 * Have the driver move function wait for idle when necessary,
 	 * instead of doing it here.
 	 */
-	spin_lock(&bdev->fence_lock);
 	ret = ttm_bo_wait(bo, false, interruptible, no_wait_gpu);
-	spin_unlock(&bdev->fence_lock);
 	if (ret)
 		return ret;
 	mem.num_pages = bo->num_pages;
@@ -1474,7 +1461,6 @@ int ttm_bo_device_init(struct ttm_bo_device *bdev,
 	bdev->glob = glob;
 	bdev->need_dma32 = need_dma32;
 	bdev->val_seq = 0;
-	spin_lock_init(&bdev->fence_lock);
 	mutex_lock(&glob->device_list_mutex);
 	list_add_tail(&bdev->device_list, &glob->device_list);
 	mutex_unlock(&glob->device_list_mutex);
@@ -1532,7 +1518,6 @@ int ttm_bo_wait(struct ttm_buffer_object *bo,
 		bool lazy, bool interruptible, bool no_wait)
 {
 	struct ttm_bo_driver *driver = bo->bdev->driver;
-	struct ttm_bo_device *bdev = bo->bdev;
 	void *sync_obj;
 	int ret = 0;
 
@@ -1541,53 +1526,33 @@ int ttm_bo_wait(struct ttm_buffer_object *bo,
 	if (likely(bo->sync_obj == NULL))
 		return 0;
 
-	while (bo->sync_obj) {
-
+	if (bo->sync_obj) {
 		if (driver->sync_obj_signaled(bo->sync_obj)) {
-			void *tmp_obj = bo->sync_obj;
-			bo->sync_obj = NULL;
+			driver->sync_obj_unref(&bo->sync_obj);
 			clear_bit(TTM_BO_PRIV_FLAG_MOVING, &bo->priv_flags);
-			spin_unlock(&bdev->fence_lock);
-			driver->sync_obj_unref(&tmp_obj);
-			spin_lock(&bdev->fence_lock);
-			continue;
+			return 0;
 		}
 
 		if (no_wait)
 			return -EBUSY;
 
 		sync_obj = driver->sync_obj_ref(bo->sync_obj);
-		spin_unlock(&bdev->fence_lock);
 		ret = driver->sync_obj_wait(sync_obj,
 					    lazy, interruptible);
-		if (unlikely(ret != 0)) {
-			driver->sync_obj_unref(&sync_obj);
-			spin_lock(&bdev->fence_lock);
-			return ret;
-		}
-		spin_lock(&bdev->fence_lock);
-		if (likely(bo->sync_obj == sync_obj)) {
-			void *tmp_obj = bo->sync_obj;
-			bo->sync_obj = NULL;
+
+		if (likely(ret == 0)) {
 			clear_bit(TTM_BO_PRIV_FLAG_MOVING,
 				  &bo->priv_flags);
-			spin_unlock(&bdev->fence_lock);
-			driver->sync_obj_unref(&sync_obj);
-			driver->sync_obj_unref(&tmp_obj);
-			spin_lock(&bdev->fence_lock);
-		} else {
-			spin_unlock(&bdev->fence_lock);
-			driver->sync_obj_unref(&sync_obj);
-			spin_lock(&bdev->fence_lock);
+			driver->sync_obj_unref(&bo->sync_obj);
 		}
+		driver->sync_obj_unref(&sync_obj);
 	}
-	return 0;
+	return ret;
 }
 EXPORT_SYMBOL(ttm_bo_wait);
 
 int ttm_bo_synccpu_write_grab(struct ttm_buffer_object *bo, bool no_wait)
 {
-	struct ttm_bo_device *bdev = bo->bdev;
 	int ret = 0;
 
 	/*
@@ -1597,9 +1562,7 @@ int ttm_bo_synccpu_write_grab(struct ttm_buffer_object *bo, bool no_wait)
 	ret = ttm_bo_reserve(bo, true, no_wait, false, NULL);
 	if (unlikely(ret != 0))
 		return ret;
-	spin_lock(&bdev->fence_lock);
 	ret = ttm_bo_wait(bo, false, true, no_wait);
-	spin_unlock(&bdev->fence_lock);
 	if (likely(ret == 0))
 		atomic_inc(&bo->cpu_writers);
 	ttm_bo_unreserve(bo);
@@ -1656,9 +1619,7 @@ static int ttm_bo_swapout(struct ttm_mem_shrink *shrink)
 	 * Wait for GPU, then move to system cached.
 	 */
 
-	spin_lock(&bo->bdev->fence_lock);
 	ret = ttm_bo_wait(bo, false, false, false);
-	spin_unlock(&bo->bdev->fence_lock);
 
 	if (unlikely(ret != 0))
 		goto out;
diff --git a/drivers/gpu/drm/ttm/ttm_bo_util.c b/drivers/gpu/drm/ttm/ttm_bo_util.c
index 30e5d90cb7bc..495aebf0f9c3 100644
--- a/drivers/gpu/drm/ttm/ttm_bo_util.c
+++ b/drivers/gpu/drm/ttm/ttm_bo_util.c
@@ -466,12 +466,10 @@ static int ttm_buffer_object_transfer(struct ttm_buffer_object *bo,
 	drm_vma_node_reset(&fbo->vma_node);
 	atomic_set(&fbo->cpu_writers, 0);
 
-	spin_lock(&bdev->fence_lock);
 	if (bo->sync_obj)
 		fbo->sync_obj = driver->sync_obj_ref(bo->sync_obj);
 	else
 		fbo->sync_obj = NULL;
-	spin_unlock(&bdev->fence_lock);
 	kref_init(&fbo->list_kref);
 	kref_init(&fbo->kref);
 	fbo->destroy = &ttm_transfered_destroy;
@@ -657,7 +655,6 @@ int ttm_bo_move_accel_cleanup(struct ttm_buffer_object *bo,
 	struct ttm_buffer_object *ghost_obj;
 	void *tmp_obj = NULL;
 
-	spin_lock(&bdev->fence_lock);
 	if (bo->sync_obj) {
 		tmp_obj = bo->sync_obj;
 		bo->sync_obj = NULL;
@@ -665,7 +662,6 @@ int ttm_bo_move_accel_cleanup(struct ttm_buffer_object *bo,
 	bo->sync_obj = driver->sync_obj_ref(sync_obj);
 	if (evict) {
 		ret = ttm_bo_wait(bo, false, false, false);
-		spin_unlock(&bdev->fence_lock);
 		if (tmp_obj)
 			driver->sync_obj_unref(&tmp_obj);
 		if (ret)
@@ -688,7 +684,6 @@ int ttm_bo_move_accel_cleanup(struct ttm_buffer_object *bo,
 		 */
 
 		set_bit(TTM_BO_PRIV_FLAG_MOVING, &bo->priv_flags);
-		spin_unlock(&bdev->fence_lock);
 		if (tmp_obj)
 			driver->sync_obj_unref(&tmp_obj);
 
diff --git a/drivers/gpu/drm/ttm/ttm_bo_vm.c b/drivers/gpu/drm/ttm/ttm_bo_vm.c
index 0ce48e5a9cb4..d05437f219e9 100644
--- a/drivers/gpu/drm/ttm/ttm_bo_vm.c
+++ b/drivers/gpu/drm/ttm/ttm_bo_vm.c
@@ -45,10 +45,8 @@ static int ttm_bo_vm_fault_idle(struct ttm_buffer_object *bo,
 				struct vm_area_struct *vma,
 				struct vm_fault *vmf)
 {
-	struct ttm_bo_device *bdev = bo->bdev;
 	int ret = 0;
 
-	spin_lock(&bdev->fence_lock);
 	if (likely(!test_bit(TTM_BO_PRIV_FLAG_MOVING, &bo->priv_flags)))
 		goto out_unlock;
 
@@ -82,7 +80,6 @@ static int ttm_bo_vm_fault_idle(struct ttm_buffer_object *bo,
 			VM_FAULT_NOPAGE;
 
 out_unlock:
-	spin_unlock(&bdev->fence_lock);
 	return ret;
 }
 
diff --git a/drivers/gpu/drm/ttm/ttm_execbuf_util.c b/drivers/gpu/drm/ttm/ttm_execbuf_util.c
index 6db47a72667e..108730e9147b 100644
--- a/drivers/gpu/drm/ttm/ttm_execbuf_util.c
+++ b/drivers/gpu/drm/ttm/ttm_execbuf_util.c
@@ -180,7 +180,6 @@ void ttm_eu_fence_buffer_objects(struct ww_acquire_ctx *ticket,
 	glob = bo->glob;
 
 	spin_lock(&glob->lru_lock);
-	spin_lock(&bdev->fence_lock);
 
 	list_for_each_entry(entry, list, head) {
 		bo = entry->bo;
@@ -189,7 +188,6 @@ void ttm_eu_fence_buffer_objects(struct ww_acquire_ctx *ticket,
 		ttm_bo_add_to_lru(bo);
 		__ttm_bo_unreserve(bo);
 	}
-	spin_unlock(&bdev->fence_lock);
 	spin_unlock(&glob->lru_lock);
 	if (ticket)
 		ww_acquire_fini(ticket);
diff --git a/drivers/gpu/drm/vmwgfx/vmwgfx_buffer.c b/drivers/gpu/drm/vmwgfx/vmwgfx_buffer.c
index 6327cfc36805..4a36bb1dc525 100644
--- a/drivers/gpu/drm/vmwgfx/vmwgfx_buffer.c
+++ b/drivers/gpu/drm/vmwgfx/vmwgfx_buffer.c
@@ -829,11 +829,7 @@ static void vmw_move_notify(struct ttm_buffer_object *bo,
  */
 static void vmw_swap_notify(struct ttm_buffer_object *bo)
 {
-	struct ttm_bo_device *bdev = bo->bdev;
-
-	spin_lock(&bdev->fence_lock);
 	ttm_bo_wait(bo, false, false, false);
-	spin_unlock(&bdev->fence_lock);
 }
 
 
diff --git a/drivers/gpu/drm/vmwgfx/vmwgfx_resource.c b/drivers/gpu/drm/vmwgfx/vmwgfx_resource.c
index 9d8c0a3f391c..936777e469bb 100644
--- a/drivers/gpu/drm/vmwgfx/vmwgfx_resource.c
+++ b/drivers/gpu/drm/vmwgfx/vmwgfx_resource.c
@@ -567,12 +567,12 @@ static int vmw_user_dmabuf_synccpu_grab(struct vmw_user_dma_buffer *user_bo,
 	int ret;
 
 	if (flags & drm_vmw_synccpu_allow_cs) {
-		struct ttm_bo_device *bdev = bo->bdev;
-
-		spin_lock(&bdev->fence_lock);
-		ret = ttm_bo_wait(bo, false, true,
-				  !!(flags & drm_vmw_synccpu_dontblock));
-		spin_unlock(&bdev->fence_lock);
+		ret = ttm_bo_reserve(bo, true, !!(flags & drm_vmw_synccpu_dontblock), false, 0);
+		if (!ret) {
+			ret = ttm_bo_wait(bo, false, true,
+					  !!(flags & drm_vmw_synccpu_dontblock));
+			ttm_bo_unreserve(bo);
+		}
 		return ret;
 	}
 
@@ -1429,12 +1429,10 @@ void vmw_fence_single_bo(struct ttm_buffer_object *bo,
 	else
 		driver->sync_obj_ref(fence);
 
-	spin_lock(&bdev->fence_lock);
 
 	old_fence_obj = bo->sync_obj;
 	bo->sync_obj = fence;
 
-	spin_unlock(&bdev->fence_lock);
 
 	if (old_fence_obj)
 		vmw_fence_obj_unreference(&old_fence_obj);
@@ -1475,7 +1473,6 @@ void vmw_resource_move_notify(struct ttm_buffer_object *bo,
 
 	if (mem->mem_type != VMW_PL_MOB) {
 		struct vmw_resource *res, *n;
-		struct ttm_bo_device *bdev = bo->bdev;
 		struct ttm_validate_buffer val_buf;
 
 		val_buf.bo = bo;
@@ -1491,9 +1488,7 @@ void vmw_resource_move_notify(struct ttm_buffer_object *bo,
 			list_del_init(&res->mob_head);
 		}
 
-		spin_lock(&bdev->fence_lock);
 		(void) ttm_bo_wait(bo, false, false, false);
-		spin_unlock(&bdev->fence_lock);
 	}
 }
 
diff --git a/include/drm/ttm/ttm_bo_api.h b/include/drm/ttm/ttm_bo_api.h
index 7526c5bf5610..67df9d7c06cc 100644
--- a/include/drm/ttm/ttm_bo_api.h
+++ b/include/drm/ttm/ttm_bo_api.h
@@ -227,10 +227,7 @@ struct ttm_buffer_object {
 	struct list_head io_reserve_lru;
 
 	/**
-	 * Members protected by struct buffer_object_device::fence_lock
-	 * In addition, setting sync_obj to anything else
-	 * than NULL requires bo::reserved to be held. This allows for
-	 * checking NULL while reserved but not holding the mentioned lock.
+	 * Members protected by a bo reservation.
 	 */
 
 	void *sync_obj;
diff --git a/include/drm/ttm/ttm_bo_driver.h b/include/drm/ttm/ttm_bo_driver.h
index 202f0a7171e8..ff725df48595 100644
--- a/include/drm/ttm/ttm_bo_driver.h
+++ b/include/drm/ttm/ttm_bo_driver.h
@@ -520,8 +520,6 @@ struct ttm_bo_global {
  *
  * @driver: Pointer to a struct ttm_bo_driver struct setup by the driver.
  * @man: An array of mem_type_managers.
- * @fence_lock: Protects the synchronizing members on *all* bos belonging
- * to this device.
  * @vma_manager: Address space manager
  * lru_lock: Spinlock that protects the buffer+device lru lists and
  * ddestroy lists.
@@ -541,7 +539,6 @@ struct ttm_bo_device {
 	struct ttm_bo_global *glob;
 	struct ttm_bo_driver *driver;
 	struct ttm_mem_type_manager man[TTM_NUM_MEM_TYPES];
-	spinlock_t fence_lock;
 
 	/*
 	 * Protected by internal locks.

[PATCH 08/19] drm/nouveau: rework to new fence interface

[Maarten Lankhorst]

From: Maarten Lankhorst <maarten.lankhorst@ubuntu.com>

Signed-off-by: Maarten Lankhorst <maarten.lankhorst@canonical.com>
---
 drivers/gpu/drm/nouveau/core/core/event.c |    4 
 drivers/gpu/drm/nouveau/nouveau_bo.c      |    6 
 drivers/gpu/drm/nouveau/nouveau_display.c |    4 
 drivers/gpu/drm/nouveau/nouveau_fence.c   |  435 ++++++++++++++++++++---------
 drivers/gpu/drm/nouveau/nouveau_fence.h   |   20 +
 drivers/gpu/drm/nouveau/nouveau_gem.c     |   17 -
 drivers/gpu/drm/nouveau/nv04_fence.c      |    4 
 drivers/gpu/drm/nouveau/nv10_fence.c      |    4 
 drivers/gpu/drm/nouveau/nv17_fence.c      |    2 
 drivers/gpu/drm/nouveau/nv50_fence.c      |    2 
 drivers/gpu/drm/nouveau/nv84_fence.c      |   11 -
 11 files changed, 330 insertions(+), 179 deletions(-)

diff --git a/drivers/gpu/drm/nouveau/core/core/event.c b/drivers/gpu/drm/nouveau/core/core/event.c
index ae81d3b5d8b7..5ddc28ec7660 100644
--- a/drivers/gpu/drm/nouveau/core/core/event.c
+++ b/drivers/gpu/drm/nouveau/core/core/event.c
@@ -139,14 +139,14 @@ nouveau_event_ref(struct nouveau_eventh *handler, struct nouveau_eventh **ref)
 void
 nouveau_event_trigger(struct nouveau_event *event, u32 types, int index)
 {
-	struct nouveau_eventh *handler;
+	struct nouveau_eventh *handler, *next;
 	unsigned long flags;
 
 	if (WARN_ON(index >= event->index_nr))
 		return;
 
 	spin_lock_irqsave(&event->list_lock, flags);
-	list_for_each_entry(handler, &event->list[index], head) {
+	list_for_each_entry_safe(handler, next, &event->list[index], head) {
 		if (!test_bit(NVKM_EVENT_ENABLE, &handler->flags))
 			continue;
 		if (!(handler->types & types))
diff --git a/drivers/gpu/drm/nouveau/nouveau_bo.c b/drivers/gpu/drm/nouveau/nouveau_bo.c
index 2cf5b0572f66..6e71d3dc2d87 100644
--- a/drivers/gpu/drm/nouveau/nouveau_bo.c
+++ b/drivers/gpu/drm/nouveau/nouveau_bo.c
@@ -959,7 +959,7 @@ nouveau_bo_move_m2mf(struct ttm_buffer_object *bo, int evict, bool intr,
 	}
 
 	mutex_lock_nested(&chan->cli->mutex, SINGLE_DEPTH_NESTING);
-	ret = nouveau_fence_sync(bo->sync_obj, chan);
+	ret = nouveau_fence_sync(nouveau_bo(bo), chan);
 	if (ret == 0) {
 		ret = drm->ttm.move(chan, bo, &bo->mem, new_mem);
 		if (ret == 0) {
@@ -1432,10 +1432,12 @@ nouveau_bo_fence_unref(void **sync_obj)
 void
 nouveau_bo_fence(struct nouveau_bo *nvbo, struct nouveau_fence *fence)
 {
-	lockdep_assert_held(&nvbo->bo.resv->lock.base);
+	struct reservation_object *resv = nvbo->bo.resv;
 
 	nouveau_bo_fence_unref(&nvbo->bo.sync_obj);
 	nvbo->bo.sync_obj = nouveau_fence_ref(fence);
+
+	reservation_object_add_excl_fence(resv, &fence->base);
 }
 
 static void *
diff --git a/drivers/gpu/drm/nouveau/nouveau_display.c b/drivers/gpu/drm/nouveau/nouveau_display.c
index 7928f8f07334..2c4798750b20 100644
--- a/drivers/gpu/drm/nouveau/nouveau_display.c
+++ b/drivers/gpu/drm/nouveau/nouveau_display.c
@@ -660,7 +660,7 @@ nouveau_page_flip_emit(struct nouveau_channel *chan,
 	spin_unlock_irqrestore(&dev->event_lock, flags);
 
 	/* Synchronize with the old framebuffer */
-	ret = nouveau_fence_sync(old_bo->bo.sync_obj, chan);
+	ret = nouveau_fence_sync(old_bo, chan);
 	if (ret)
 		goto fail;
 
@@ -721,7 +721,7 @@ nouveau_crtc_page_flip(struct drm_crtc *crtc, struct drm_framebuffer *fb,
 		goto fail_unpin;
 
 	/* synchronise rendering channel with the kernel's channel */
-	ret = nouveau_fence_sync(new_bo->bo.sync_obj, chan);
+	ret = nouveau_fence_sync(new_bo, chan);
 	if (ret) {
 		ttm_bo_unreserve(&new_bo->bo);
 		goto fail_unpin;
diff --git a/drivers/gpu/drm/nouveau/nouveau_fence.c b/drivers/gpu/drm/nouveau/nouveau_fence.c
index ab5ea3b0d666..d24f8ce4341a 100644
--- a/drivers/gpu/drm/nouveau/nouveau_fence.c
+++ b/drivers/gpu/drm/nouveau/nouveau_fence.c
@@ -32,91 +32,139 @@
 #include "nouveau_drm.h"
 #include "nouveau_dma.h"
 #include "nouveau_fence.h"
+#include <trace/events/fence.h>
 
 #include <engine/fifo.h>
 
-struct fence_work {
-	struct work_struct base;
-	struct list_head head;
-	void (*func)(void *);
-	void *data;
-};
+static const struct fence_ops nouveau_fence_ops_uevent;
+static const struct fence_ops nouveau_fence_ops_legacy;
 
 static void
 nouveau_fence_signal(struct nouveau_fence *fence)
 {
-	struct fence_work *work, *temp;
+	fence_signal_locked(&fence->base);
+	list_del(&fence->head);
+
+	if (fence->base.ops == &nouveau_fence_ops_uevent &&
+	    fence->event.head.next) {
+		struct nouveau_event *event;
 
-	list_for_each_entry_safe(work, temp, &fence->work, head) {
-		schedule_work(&work->base);
-		list_del(&work->head);
+		list_del(&fence->event.head);
+		fence->event.head.next = NULL;
+
+		event = container_of(fence->base.lock, typeof(*event), list_lock);
+		nouveau_event_put(&fence->event);
 	}
 
-	fence->channel = NULL;
-	list_del(&fence->head);
+	fence_put(&fence->base);
+}
+
+static struct nouveau_fence *
+nouveau_local_fence(struct fence *fence, struct nouveau_drm *drm) {
+	struct nouveau_fence_priv *priv = (void*)drm->fence;
+	struct nouveau_fence *f = container_of(fence,
+					       struct nouveau_fence,
+					       base);
+
+	if (fence->ops != &nouveau_fence_ops_legacy &&
+	    fence->ops != &nouveau_fence_ops_uevent)
+		return NULL;
+
+	if (fence->context < priv->context_base ||
+	    fence->context >= priv->context_base + priv->contexts)
+		return NULL;
+
+	return f;
 }
 
 void
 nouveau_fence_context_del(struct nouveau_fence_chan *fctx)
 {
 	struct nouveau_fence *fence, *fnext;
-	spin_lock(&fctx->lock);
-	list_for_each_entry_safe(fence, fnext, &fctx->pending, head) {
+
+	spin_lock_irq(fctx->lock);
+	list_for_each_entry_safe(fence, fnext, &fctx->pending, head)
 		nouveau_fence_signal(fence);
-	}
-	spin_unlock(&fctx->lock);
+	spin_unlock_irq(fctx->lock);
 }
 
 void
-nouveau_fence_context_new(struct nouveau_fence_chan *fctx)
+nouveau_fence_context_new(struct nouveau_channel *chan, struct nouveau_fence_chan *fctx)
 {
+	struct nouveau_fifo *pfifo = nouveau_fifo(chan->drm->device);
+	struct nouveau_fifo_chan *fifo = (void*)chan->object;
+
+	fctx->lock = &pfifo->uevent->list_lock;
 	INIT_LIST_HEAD(&fctx->flip);
 	INIT_LIST_HEAD(&fctx->pending);
-	spin_lock_init(&fctx->lock);
+
+	snprintf(fctx->name, sizeof(fctx->name) - 1, "nouveau channel %i", fifo->chid);
 }
 
+struct nouveau_fence_work {
+	struct work_struct work;
+	struct fence_cb cb;
+	void (*func)(void *);
+	void *data;
+};
+
 static void
 nouveau_fence_work_handler(struct work_struct *kwork)
 {
-	struct fence_work *work = container_of(kwork, typeof(*work), base);
+	struct nouveau_fence_work *work = container_of(kwork, typeof(*work), work);
 	work->func(work->data);
 	kfree(work);
 }
 
+static void nouveau_fence_work_cb(struct fence *fence, struct fence_cb *cb)
+{
+	struct nouveau_fence_work *work = container_of(cb, typeof(*work), cb);
+
+	schedule_work(&work->work);
+}
+
+/*
+ * In an ideal world, read would not assume the channel context is still alive.
+ * This function may be called from another device, running into free memory as a
+ * result. The drm node should still be there, so we can derive the index from
+ * the fence context.
+ */
+static bool nouveau_fence_is_signaled(struct fence *f)
+{
+	struct nouveau_fence *fence = container_of(f, struct nouveau_fence, base);
+	struct nouveau_channel *chan = fence->channel;
+	struct nouveau_fence_chan *fctx = chan->fence;
+
+	return (int)(fctx->read(chan) - fence->base.seqno) >= 0;
+}
+
 void
 nouveau_fence_work(struct nouveau_fence *fence,
 		   void (*func)(void *), void *data)
 {
-	struct nouveau_channel *chan = fence->channel;
-	struct nouveau_fence_chan *fctx;
-	struct fence_work *work = NULL;
+	struct nouveau_fence_work *work;
 
-	if (nouveau_fence_done(fence)) {
-		func(data);
-		return;
-	}
+	if (fence_is_signaled(&fence->base))
+		goto err;
 
-	fctx = chan->fence;
 	work = kmalloc(sizeof(*work), GFP_KERNEL);
 	if (!work) {
 		WARN_ON(nouveau_fence_wait(fence, false, false));
-		func(data);
-		return;
+		goto err;
 	}
 
-	spin_lock(&fctx->lock);
-	if (!fence->channel) {
-		spin_unlock(&fctx->lock);
-		kfree(work);
-		func(data);
-		return;
-	}
-
-	INIT_WORK(&work->base, nouveau_fence_work_handler);
+	INIT_WORK(&work->work, nouveau_fence_work_handler);
 	work->func = func;
 	work->data = data;
-	list_add(&work->head, &fence->work);
-	spin_unlock(&fctx->lock);
+
+	if (fence_add_callback(&fence->base, &work->cb, nouveau_fence_work_cb) < 0)
+		goto err_free;
+	return;
+
+err_free:
+	kfree(work);
+err:
+	func(data);
 }
 
 static void
@@ -125,33 +173,45 @@ nouveau_fence_update(struct nouveau_channel *chan)
 	struct nouveau_fence_chan *fctx = chan->fence;
 	struct nouveau_fence *fence, *fnext;
 
-	spin_lock(&fctx->lock);
+	u32 seq = fctx->read(chan);
+
 	list_for_each_entry_safe(fence, fnext, &fctx->pending, head) {
-		if (fctx->read(chan) < fence->sequence)
+		if ((int)(seq - fence->base.seqno) < 0)
 			break;
 
 		nouveau_fence_signal(fence);
-		nouveau_fence_unref(&fence);
 	}
-	spin_unlock(&fctx->lock);
 }
 
 int
 nouveau_fence_emit(struct nouveau_fence *fence, struct nouveau_channel *chan)
 {
 	struct nouveau_fence_chan *fctx = chan->fence;
+	struct nouveau_fifo *pfifo = nouveau_fifo(chan->drm->device);
+	struct nouveau_fifo_chan *fifo = (void*)chan->object;
+	struct nouveau_fence_priv *priv = (void*)chan->drm->fence;
 	int ret;
 
 	fence->channel  = chan;
 	fence->timeout  = jiffies + (15 * HZ);
-	fence->sequence = ++fctx->sequence;
 
+	if (priv->uevent)
+		fence_init(&fence->base, &nouveau_fence_ops_uevent,
+			   &pfifo->uevent->list_lock,
+			   priv->context_base + fifo->chid, ++fctx->sequence);
+	else
+		fence_init(&fence->base, &nouveau_fence_ops_legacy,
+			   &pfifo->uevent->list_lock,
+			   priv->context_base + fifo->chid, ++fctx->sequence);
+
+	trace_fence_emit(&fence->base);
 	ret = fctx->emit(fence);
 	if (!ret) {
-		kref_get(&fence->kref);
-		spin_lock(&fctx->lock);
+		fence_get(&fence->base);
+		spin_lock_irq(fctx->lock);
+		nouveau_fence_update(chan);
 		list_add_tail(&fence->head, &fctx->pending);
-		spin_unlock(&fctx->lock);
+		spin_unlock_irq(fctx->lock);
 	}
 
 	return ret;
@@ -160,104 +220,71 @@ nouveau_fence_emit(struct nouveau_fence *fence, struct nouveau_channel *chan)
 bool
 nouveau_fence_done(struct nouveau_fence *fence)
 {
-	if (fence->channel)
+	if (fence->base.ops == &nouveau_fence_ops_legacy ||
+	    fence->base.ops == &nouveau_fence_ops_uevent) {
+		struct nouveau_fence_chan *fctx;
+		unsigned long flags;
+
+		if (test_bit(FENCE_FLAG_SIGNALED_BIT, &fence->base.flags))
+			return true;
+
+		fctx = fence->channel->fence;
+		spin_lock_irqsave(fctx->lock, flags);
 		nouveau_fence_update(fence->channel);
-	return !fence->channel;
+		spin_unlock_irqrestore(fctx->lock, flags);
+	}
+	return fence_is_signaled(&fence->base);
 }
 
-static int
-nouveau_fence_wait_uevent_handler(void *data, u32 type, int index)
+static long
+nouveau_fence_wait_legacy(struct fence *f, bool intr, long wait)
 {
-	struct nouveau_fence_priv *priv = data;
-	wake_up_all(&priv->waiting);
-	return NVKM_EVENT_KEEP;
-}
+	struct nouveau_fence *fence = container_of(f, typeof(*fence), base);
+	unsigned long sleep_time = NSEC_PER_MSEC / 1000;
+	unsigned long t = jiffies, timeout = t + wait;
 
-static int
-nouveau_fence_wait_uevent(struct nouveau_fence *fence, bool intr)
+	while (!nouveau_fence_done(fence)) {
+		ktime_t kt;
 
-{
-	struct nouveau_channel *chan = fence->channel;
-	struct nouveau_fifo *pfifo = nouveau_fifo(chan->drm->device);
-	struct nouveau_fence_priv *priv = chan->drm->fence;
-	struct nouveau_eventh *handler;
-	int ret = 0;
+		t = jiffies;
 
-	ret = nouveau_event_new(pfifo->uevent, 1, 0,
-				nouveau_fence_wait_uevent_handler,
-				priv, &handler);
-	if (ret)
-		return ret;
+		if (wait != MAX_SCHEDULE_TIMEOUT && time_after_eq(t, timeout)) {
+			__set_current_state(TASK_RUNNING);
+			return 0;
+		}
 
-	nouveau_event_get(handler);
+		__set_current_state(intr ? TASK_INTERRUPTIBLE :
+					   TASK_UNINTERRUPTIBLE);
 
-	if (fence->timeout) {
-		unsigned long timeout = fence->timeout - jiffies;
-
-		if (time_before(jiffies, fence->timeout)) {
-			if (intr) {
-				ret = wait_event_interruptible_timeout(
-						priv->waiting,
-						nouveau_fence_done(fence),
-						timeout);
-			} else {
-				ret = wait_event_timeout(priv->waiting,
-						nouveau_fence_done(fence),
-						timeout);
-			}
-		}
+		kt = ktime_set(0, sleep_time);
+		schedule_hrtimeout(&kt, HRTIMER_MODE_REL);
+		sleep_time *= 2;
+		if (sleep_time > NSEC_PER_MSEC)
+			sleep_time = NSEC_PER_MSEC;
 
-		if (ret >= 0) {
-			fence->timeout = jiffies + ret;
-			if (time_after_eq(jiffies, fence->timeout))
-				ret = -EBUSY;
-		}
-	} else {
-		if (intr) {
-			ret = wait_event_interruptible(priv->waiting,
-					nouveau_fence_done(fence));
-		} else {
-			wait_event(priv->waiting, nouveau_fence_done(fence));
-		}
+		if (intr && signal_pending(current))
+			return -ERESTARTSYS;
 	}
 
-	nouveau_event_ref(NULL, &handler);
-	if (unlikely(ret < 0))
-		return ret;
+	__set_current_state(TASK_RUNNING);
 
-	return 0;
+	return timeout - t;
 }
 
-int
-nouveau_fence_wait(struct nouveau_fence *fence, bool lazy, bool intr)
+static int
+nouveau_fence_wait_busy(struct nouveau_fence *fence, bool intr)
 {
-	struct nouveau_channel *chan = fence->channel;
-	struct nouveau_fence_priv *priv = chan ? chan->drm->fence : NULL;
-	unsigned long sleep_time = NSEC_PER_MSEC / 1000;
-	ktime_t t;
 	int ret = 0;
 
-	while (priv && priv->uevent && lazy && !nouveau_fence_done(fence)) {
-		ret = nouveau_fence_wait_uevent(fence, intr);
-		if (ret < 0)
-			return ret;
-	}
-
 	while (!nouveau_fence_done(fence)) {
-		if (fence->timeout && time_after_eq(jiffies, fence->timeout)) {
+		if (time_after_eq(jiffies, fence->timeout)) {
 			ret = -EBUSY;
 			break;
 		}
 
-		__set_current_state(intr ? TASK_INTERRUPTIBLE :
-					   TASK_UNINTERRUPTIBLE);
-		if (lazy) {
-			t = ktime_set(0, sleep_time);
-			schedule_hrtimeout(&t, HRTIMER_MODE_REL);
-			sleep_time *= 2;
-			if (sleep_time > NSEC_PER_MSEC)
-				sleep_time = NSEC_PER_MSEC;
-		}
+		__set_current_state(intr ?
+				    TASK_INTERRUPTIBLE :
+				    TASK_UNINTERRUPTIBLE);
 
 		if (intr && signal_pending(current)) {
 			ret = -ERESTARTSYS;
@@ -270,36 +297,79 @@ nouveau_fence_wait(struct nouveau_fence *fence, bool lazy, bool intr)
 }
 
 int
-nouveau_fence_sync(struct nouveau_fence *fence, struct nouveau_channel *chan)
+nouveau_fence_wait(struct nouveau_fence *fence, bool lazy, bool intr)
+{
+	long ret;
+
+	if (!lazy)
+		return nouveau_fence_wait_busy(fence, intr);
+
+	ret = fence_wait_timeout(&fence->base, intr, 15 * HZ);
+	if (ret < 0)
+		return ret;
+	else if (!ret)
+		return -EBUSY;
+	else
+		return 0;
+}
+
+int
+nouveau_fence_sync(struct nouveau_bo *nvbo, struct nouveau_channel *chan)
 {
 	struct nouveau_fence_chan *fctx = chan->fence;
-	struct nouveau_channel *prev;
-	int ret = 0;
+	struct fence *fence = NULL;
+	struct reservation_object *resv = nvbo->bo.resv;
+	struct reservation_object_list *fobj;
+	int ret = 0, i;
+
+	fence = nvbo->bo.sync_obj;
+	if (fence && fence_is_signaled(fence)) {
+		nouveau_fence_unref((struct nouveau_fence **)
+				    &nvbo->bo.sync_obj);
+		fence = NULL;
+	}
+
+	if (fence) {
+		struct nouveau_fence *f = container_of(fence,
+						       struct nouveau_fence,
+						       base);
+		struct nouveau_channel *prev = f->channel;
 
-	prev = fence ? fence->channel : NULL;
-	if (prev) {
-		if (unlikely(prev != chan && !nouveau_fence_done(fence))) {
-			ret = fctx->sync(fence, prev, chan);
+		if (prev != chan) {
+			ret = fctx->sync(f, prev, chan);
 			if (unlikely(ret))
-				ret = nouveau_fence_wait(fence, true, false);
+				ret = nouveau_fence_wait(f, true, true);
 		}
 	}
 
-	return ret;
-}
+	if (ret)
+		return ret;
 
-static void
-nouveau_fence_del(struct kref *kref)
-{
-	struct nouveau_fence *fence = container_of(kref, typeof(*fence), kref);
-	kfree(fence);
+	fence = reservation_object_get_excl(resv);
+	if (fence && !nouveau_local_fence(fence, chan->drm))
+		ret = fence_wait(fence, true);
+
+	fobj = reservation_object_get_list(resv);
+	if (!fobj || ret)
+		return ret;
+
+	for (i = 0; i < fobj->shared_count && !ret; ++i) {
+		fence = rcu_dereference_protected(fobj->shared[i],
+						reservation_object_held(resv));
+
+		/* should always be true, for now */
+		if (!nouveau_local_fence(fence, chan->drm))
+			ret = fence_wait(fence, true);
+	}
+
+	return ret;
 }
 
 void
 nouveau_fence_unref(struct nouveau_fence **pfence)
 {
 	if (*pfence)
-		kref_put(&(*pfence)->kref, nouveau_fence_del);
+		fence_put(&(*pfence)->base);
 	*pfence = NULL;
 }
 
@@ -307,7 +377,7 @@ struct nouveau_fence *
 nouveau_fence_ref(struct nouveau_fence *fence)
 {
 	if (fence)
-		kref_get(&fence->kref);
+		fence_get(&fence->base);
 	return fence;
 }
 
@@ -325,9 +395,7 @@ nouveau_fence_new(struct nouveau_channel *chan, bool sysmem,
 	if (!fence)
 		return -ENOMEM;
 
-	INIT_LIST_HEAD(&fence->work);
 	fence->sysmem = sysmem;
-	kref_init(&fence->kref);
 
 	ret = nouveau_fence_emit(fence, chan);
 	if (ret)
@@ -336,3 +404,86 @@ nouveau_fence_new(struct nouveau_channel *chan, bool sysmem,
 	*pfence = fence;
 	return ret;
 }
+
+
+static bool nouveau_fence_no_signaling(struct fence *f)
+{
+	/*
+	 * This needs uevents to work correctly, but fence_add_callback relies on
+	 * being able to enable signaling. It will still get signaled eventually,
+	 * just not right away.
+	 */
+	if (nouveau_fence_is_signaled(f))
+		return false;
+
+	return true;
+}
+
+static const char *nouveau_fence_get_get_driver_name(struct fence *fence)
+{
+	return "nouveau";
+}
+
+static const char *nouveau_fence_get_timeline_name(struct fence *f)
+{
+	struct nouveau_fence *fence =
+		container_of(f, struct nouveau_fence, base);
+	struct nouveau_fence_chan *fctx = fence->channel->fence;
+
+	return fctx ? fctx->name : "dead channel";
+}
+
+static const struct fence_ops nouveau_fence_ops_legacy = {
+	.get_driver_name = nouveau_fence_get_get_driver_name,
+	.get_timeline_name = nouveau_fence_get_timeline_name,
+	.enable_signaling = nouveau_fence_no_signaling,
+	.signaled = nouveau_fence_is_signaled,
+	.wait = nouveau_fence_wait_legacy,
+	.release = NULL
+};
+
+static int
+nouveau_fence_wait_uevent_handler(void *priv, u32 types, int index)
+{
+	struct nouveau_fence *fence = priv;
+
+	if (nouveau_fence_is_signaled(&fence->base))
+		nouveau_fence_signal(fence);
+
+	/*
+	 * NVKM_EVENT_DROP is never appropriate here, nouveau_fence_signal
+	 * will unlink and free the event if needed.
+	 */
+	return NVKM_EVENT_KEEP;
+}
+
+static bool nouveau_fence_enable_signaling(struct fence *f)
+{
+	struct nouveau_fence *fence = container_of(f, struct nouveau_fence, base);
+	struct nouveau_event *event = container_of(f->lock, struct nouveau_event, list_lock);
+	struct nouveau_eventh *handler = &fence->event;
+
+	handler->event = event;
+	handler->func = nouveau_fence_wait_uevent_handler;
+	handler->priv = fence;
+	handler->types = 1;
+
+	nouveau_event_get(handler);
+	if (nouveau_fence_is_signaled(f)) {
+		nouveau_event_put(handler);
+		return false;
+	}
+
+	list_add_tail(&handler->head, &event->list[0]);
+
+	return true;
+}
+
+static const struct fence_ops nouveau_fence_ops_uevent = {
+	.get_driver_name = nouveau_fence_get_get_driver_name,
+	.get_timeline_name = nouveau_fence_get_timeline_name,
+	.enable_signaling = nouveau_fence_enable_signaling,
+	.signaled = nouveau_fence_is_signaled,
+	.wait = fence_default_wait,
+	.release = NULL
+};
diff --git a/drivers/gpu/drm/nouveau/nouveau_fence.h b/drivers/gpu/drm/nouveau/nouveau_fence.h
index c57bb61da58c..1989ec22e66e 100644
--- a/drivers/gpu/drm/nouveau/nouveau_fence.h
+++ b/drivers/gpu/drm/nouveau/nouveau_fence.h
@@ -1,18 +1,21 @@
 #ifndef __NOUVEAU_FENCE_H__
 #define __NOUVEAU_FENCE_H__
 
+#include <linux/fence.h>
+
 struct nouveau_drm;
+struct nouveau_bo;
 
 struct nouveau_fence {
+	struct fence base;
+
 	struct list_head head;
-	struct list_head work;
-	struct kref kref;
+	struct nouveau_eventh event;
 
 	bool sysmem;
 
 	struct nouveau_channel *channel;
 	unsigned long timeout;
-	u32 sequence;
 };
 
 int  nouveau_fence_new(struct nouveau_channel *, bool sysmem,
@@ -25,7 +28,7 @@ int  nouveau_fence_emit(struct nouveau_fence *, struct nouveau_channel *);
 bool nouveau_fence_done(struct nouveau_fence *);
 void nouveau_fence_work(struct nouveau_fence *, void (*)(void *), void *);
 int  nouveau_fence_wait(struct nouveau_fence *, bool lazy, bool intr);
-int  nouveau_fence_sync(struct nouveau_fence *, struct nouveau_channel *);
+int  nouveau_fence_sync(struct nouveau_bo *, struct nouveau_channel *);
 
 struct nouveau_fence_chan {
 	struct list_head pending;
@@ -38,8 +41,10 @@ struct nouveau_fence_chan {
 	int  (*emit32)(struct nouveau_channel *, u64, u32);
 	int  (*sync32)(struct nouveau_channel *, u64, u32);
 
-	spinlock_t lock;
+	spinlock_t *lock;
 	u32 sequence;
+	u32 context;
+	char name[24];
 };
 
 struct nouveau_fence_priv {
@@ -49,13 +54,14 @@ struct nouveau_fence_priv {
 	int  (*context_new)(struct nouveau_channel *);
 	void (*context_del)(struct nouveau_channel *);
 
-	wait_queue_head_t waiting;
 	bool uevent;
+
+	u32 contexts, context_base;
 };
 
 #define nouveau_fence(drm) ((struct nouveau_fence_priv *)(drm)->fence)
 
-void nouveau_fence_context_new(struct nouveau_fence_chan *);
+void nouveau_fence_context_new(struct nouveau_channel *, struct nouveau_fence_chan *);
 void nouveau_fence_context_del(struct nouveau_fence_chan *);
 
 int nv04_fence_create(struct nouveau_drm *);
diff --git a/drivers/gpu/drm/nouveau/nouveau_gem.c b/drivers/gpu/drm/nouveau/nouveau_gem.c
index 474f67238cef..2bd62f4ecf6a 100644
--- a/drivers/gpu/drm/nouveau/nouveau_gem.c
+++ b/drivers/gpu/drm/nouveau/nouveau_gem.c
@@ -428,18 +428,6 @@ retry:
 }
 
 static int
-validate_sync(struct nouveau_channel *chan, struct nouveau_bo *nvbo)
-{
-	struct nouveau_fence *fence = nvbo->bo.sync_obj;
-	int ret = 0;
-
-	if (fence)
-		ret = nouveau_fence_sync(fence, chan);
-
-	return ret;
-}
-
-static int
 validate_list(struct nouveau_channel *chan, struct nouveau_cli *cli,
 	      struct list_head *list, struct drm_nouveau_gem_pushbuf_bo *pbbo,
 	      uint64_t user_pbbo_ptr)
@@ -468,9 +456,10 @@ validate_list(struct nouveau_channel *chan, struct nouveau_cli *cli,
 			return ret;
 		}
 
-		ret = validate_sync(chan, nvbo);
+		ret = nouveau_fence_sync(nvbo, chan);
 		if (unlikely(ret)) {
-			NV_ERROR(cli, "fail post-validate sync\n");
+			if (ret != -ERESTARTSYS)
+				NV_ERROR(cli, "fail post-validate sync\n");
 			return ret;
 		}
 
diff --git a/drivers/gpu/drm/nouveau/nv04_fence.c b/drivers/gpu/drm/nouveau/nv04_fence.c
index 94eadd1dd10a..997c54122ed9 100644
--- a/drivers/gpu/drm/nouveau/nv04_fence.c
+++ b/drivers/gpu/drm/nouveau/nv04_fence.c
@@ -43,7 +43,7 @@ nv04_fence_emit(struct nouveau_fence *fence)
 	int ret = RING_SPACE(chan, 2);
 	if (ret == 0) {
 		BEGIN_NV04(chan, NvSubSw, 0x0150, 1);
-		OUT_RING  (chan, fence->sequence);
+		OUT_RING  (chan, fence->base.seqno);
 		FIRE_RING (chan);
 	}
 	return ret;
@@ -77,7 +77,7 @@ nv04_fence_context_new(struct nouveau_channel *chan)
 {
 	struct nv04_fence_chan *fctx = kzalloc(sizeof(*fctx), GFP_KERNEL);
 	if (fctx) {
-		nouveau_fence_context_new(&fctx->base);
+		nouveau_fence_context_new(chan, &fctx->base);
 		fctx->base.emit = nv04_fence_emit;
 		fctx->base.sync = nv04_fence_sync;
 		fctx->base.read = nv04_fence_read;
diff --git a/drivers/gpu/drm/nouveau/nv10_fence.c b/drivers/gpu/drm/nouveau/nv10_fence.c
index 06f434f03fba..e8f73f7f31ef 100644
--- a/drivers/gpu/drm/nouveau/nv10_fence.c
+++ b/drivers/gpu/drm/nouveau/nv10_fence.c
@@ -36,7 +36,7 @@ nv10_fence_emit(struct nouveau_fence *fence)
 	int ret = RING_SPACE(chan, 2);
 	if (ret == 0) {
 		BEGIN_NV04(chan, 0, NV10_SUBCHAN_REF_CNT, 1);
-		OUT_RING  (chan, fence->sequence);
+		OUT_RING  (chan, fence->base.seqno);
 		FIRE_RING (chan);
 	}
 	return ret;
@@ -74,7 +74,7 @@ nv10_fence_context_new(struct nouveau_channel *chan)
 	if (!fctx)
 		return -ENOMEM;
 
-	nouveau_fence_context_new(&fctx->base);
+	nouveau_fence_context_new(chan, &fctx->base);
 	fctx->base.emit = nv10_fence_emit;
 	fctx->base.read = nv10_fence_read;
 	fctx->base.sync = nv10_fence_sync;
diff --git a/drivers/gpu/drm/nouveau/nv17_fence.c b/drivers/gpu/drm/nouveau/nv17_fence.c
index 22aa9963ea6f..e404bab31e9d 100644
--- a/drivers/gpu/drm/nouveau/nv17_fence.c
+++ b/drivers/gpu/drm/nouveau/nv17_fence.c
@@ -83,7 +83,7 @@ nv17_fence_context_new(struct nouveau_channel *chan)
 	if (!fctx)
 		return -ENOMEM;
 
-	nouveau_fence_context_new(&fctx->base);
+	nouveau_fence_context_new(chan, &fctx->base);
 	fctx->base.emit = nv10_fence_emit;
 	fctx->base.read = nv10_fence_read;
 	fctx->base.sync = nv17_fence_sync;
diff --git a/drivers/gpu/drm/nouveau/nv50_fence.c b/drivers/gpu/drm/nouveau/nv50_fence.c
index 0ee363840035..19f6fccb84a1 100644
--- a/drivers/gpu/drm/nouveau/nv50_fence.c
+++ b/drivers/gpu/drm/nouveau/nv50_fence.c
@@ -47,7 +47,7 @@ nv50_fence_context_new(struct nouveau_channel *chan)
 	if (!fctx)
 		return -ENOMEM;
 
-	nouveau_fence_context_new(&fctx->base);
+	nouveau_fence_context_new(chan, &fctx->base);
 	fctx->base.emit = nv10_fence_emit;
 	fctx->base.read = nv10_fence_read;
 	fctx->base.sync = nv17_fence_sync;
diff --git a/drivers/gpu/drm/nouveau/nv84_fence.c b/drivers/gpu/drm/nouveau/nv84_fence.c
index 9fd475c89820..8a06727b23d1 100644
--- a/drivers/gpu/drm/nouveau/nv84_fence.c
+++ b/drivers/gpu/drm/nouveau/nv84_fence.c
@@ -89,7 +89,7 @@ nv84_fence_emit(struct nouveau_fence *fence)
 	else
 		addr += fctx->vma.offset;
 
-	return fctx->base.emit32(chan, addr, fence->sequence);
+	return fctx->base.emit32(chan, addr, fence->base.seqno);
 }
 
 static int
@@ -105,7 +105,7 @@ nv84_fence_sync(struct nouveau_fence *fence,
 	else
 		addr += fctx->vma.offset;
 
-	return fctx->base.sync32(chan, addr, fence->sequence);
+	return fctx->base.sync32(chan, addr, fence->base.seqno);
 }
 
 static u32
@@ -149,12 +149,14 @@ nv84_fence_context_new(struct nouveau_channel *chan)
 	if (!fctx)
 		return -ENOMEM;
 
-	nouveau_fence_context_new(&fctx->base);
+	nouveau_fence_context_new(chan, &fctx->base);
 	fctx->base.emit = nv84_fence_emit;
 	fctx->base.sync = nv84_fence_sync;
 	fctx->base.read = nv84_fence_read;
 	fctx->base.emit32 = nv84_fence_emit32;
 	fctx->base.sync32 = nv84_fence_sync32;
+	fctx->base.sequence = nv84_fence_read(chan);
+	fctx->base.context = priv->base.context_base + fifo->chid;
 
 	ret = nouveau_bo_vma_add(priv->bo, client->vm, &fctx->vma);
 	if (ret == 0) {
@@ -239,7 +241,8 @@ nv84_fence_create(struct nouveau_drm *drm)
 	priv->base.context_new = nv84_fence_context_new;
 	priv->base.context_del = nv84_fence_context_del;
 
-	init_waitqueue_head(&priv->base.waiting);
+	priv->base.contexts = pfifo->max + 1;
+	priv->base.context_base = fence_context_alloc(priv->base.contexts);
 	priv->base.uevent = true;
 
 	ret = nouveau_bo_new(drm->dev, 16 * (pfifo->max + 1), 0,

[PATCH 09/19] drm/radeon: handle lockup in delayed work, v2

[Maarten Lankhorst]

Signed-off-by: Maarten Lankhorst <maarten.lankhorst@canonical.com>
---
V1 had a nasty bug breaking gpu lockup recovery. The fix is not
allowing radeon_fence_driver_check_lockup to take exclusive_lock,
and kill it during lockup recovery instead.
---
 drivers/gpu/drm/radeon/radeon.h        |    3 +
 drivers/gpu/drm/radeon/radeon_device.c |    5 +
 drivers/gpu/drm/radeon/radeon_fence.c  |  124 ++++++++++++++++++--------------
 drivers/gpu/drm/radeon/radeon_ring.c   |    1 
 4 files changed, 77 insertions(+), 56 deletions(-)

diff --git a/drivers/gpu/drm/radeon/radeon.h b/drivers/gpu/drm/radeon/radeon.h
index 60c47f829122..b01d88fc10cb 100644
--- a/drivers/gpu/drm/radeon/radeon.h
+++ b/drivers/gpu/drm/radeon/radeon.h
@@ -347,6 +347,8 @@ struct radeon_fence_driver {
 	uint64_t			sync_seq[RADEON_NUM_RINGS];
 	atomic64_t			last_seq;
 	bool				initialized;
+	struct delayed_work	fence_check_work;
+	struct radeon_device	*rdev;
 };
 
 struct radeon_fence {
@@ -360,6 +362,7 @@ struct radeon_fence {
 
 int radeon_fence_driver_start_ring(struct radeon_device *rdev, int ring);
 int radeon_fence_driver_init(struct radeon_device *rdev);
+void radeon_fence_cancel_delayed_check(struct radeon_device *rdev, int ring);
 void radeon_fence_driver_fini(struct radeon_device *rdev);
 void radeon_fence_driver_force_completion(struct radeon_device *rdev);
 int radeon_fence_emit(struct radeon_device *rdev, struct radeon_fence **fence, int ring);
diff --git a/drivers/gpu/drm/radeon/radeon_device.c b/drivers/gpu/drm/radeon/radeon_device.c
index 697add2cd4e3..21efd32d07ee 100644
--- a/drivers/gpu/drm/radeon/radeon_device.c
+++ b/drivers/gpu/drm/radeon/radeon_device.c
@@ -1637,6 +1637,11 @@ int radeon_gpu_reset(struct radeon_device *rdev)
 	radeon_save_bios_scratch_regs(rdev);
 	/* block TTM */
 	resched = ttm_bo_lock_delayed_workqueue(&rdev->mman.bdev);
+
+	/* kill all the hangcheck tests too */
+	for (i = 0; i < RADEON_NUM_RINGS; ++i)
+		radeon_fence_cancel_delayed_check(rdev, i);
+
 	radeon_pm_suspend(rdev);
 	radeon_suspend(rdev);
 
diff --git a/drivers/gpu/drm/radeon/radeon_fence.c b/drivers/gpu/drm/radeon/radeon_fence.c
index 913787085dfa..c055acc3a271 100644
--- a/drivers/gpu/drm/radeon/radeon_fence.c
+++ b/drivers/gpu/drm/radeon/radeon_fence.c
@@ -125,16 +125,7 @@ int radeon_fence_emit(struct radeon_device *rdev,
 	return 0;
 }
 
-/**
- * radeon_fence_process - process a fence
- *
- * @rdev: radeon_device pointer
- * @ring: ring index the fence is associated with
- *
- * Checks the current fence value and wakes the fence queue
- * if the sequence number has increased (all asics).
- */
-void radeon_fence_process(struct radeon_device *rdev, int ring)
+static bool __radeon_fence_process(struct radeon_device *rdev, int ring)
 {
 	uint64_t seq, last_seq, last_emitted;
 	unsigned count_loop = 0;
@@ -190,7 +181,51 @@ void radeon_fence_process(struct radeon_device *rdev, int ring)
 		}
 	} while (atomic64_xchg(&rdev->fence_drv[ring].last_seq, seq) > seq);
 
-	if (wake)
+	if (seq < last_emitted)
+		mod_delayed_work(system_power_efficient_wq,
+				 &rdev->fence_drv[ring].fence_check_work,
+				 RADEON_FENCE_JIFFIES_TIMEOUT);
+
+	return wake;
+}
+
+static void radeon_fence_driver_check_lockup(struct work_struct *work)
+{
+	struct radeon_fence_driver *fence_drv;
+	struct radeon_device *rdev;
+	unsigned long iring;
+
+	fence_drv = container_of(work, struct radeon_fence_driver, fence_check_work.work);
+	rdev = fence_drv->rdev;
+	iring = fence_drv - &rdev->fence_drv[0];
+
+	if (__radeon_fence_process(rdev, iring))
+		wake_up_all(&rdev->fence_queue);
+	else if (radeon_ring_is_lockup(rdev, iring, &rdev->ring[iring])) {
+		/* good news we believe it's a lockup */
+		dev_warn(rdev->dev, "GPU lockup (current fence id "
+			 "0x%016llx last fence id 0x%016llx on ring %ld)\n",
+			 (uint64_t)atomic64_read(&fence_drv->last_seq),
+			 fence_drv->sync_seq[iring], iring);
+
+		/* remember that we need an reset */
+		rdev->needs_reset = true;
+		wake_up_all(&rdev->fence_queue);
+	}
+}
+
+/**
+ * radeon_fence_process - process a fence
+ *
+ * @rdev: radeon_device pointer
+ * @ring: ring index the fence is associated with
+ *
+ * Checks the current fence value and wakes the fence queue
+ * if the sequence number has increased (all asics).
+ */
+void radeon_fence_process(struct radeon_device *rdev, int ring)
+{
+	if (__radeon_fence_process(rdev, ring))
 		wake_up_all(&rdev->fence_queue);
 }
 
@@ -302,9 +337,10 @@ static int radeon_fence_wait_seq(struct radeon_device *rdev, u64 *target_seq,
 {
 	uint64_t last_seq[RADEON_NUM_RINGS];
 	bool signaled;
-	int i, r;
+	int i;
 
 	while (!radeon_fence_any_seq_signaled(rdev, target_seq)) {
+		long r;
 
 		/* Save current sequence values, used to check for GPU lockups */
 		for (i = 0; i < RADEON_NUM_RINGS; ++i) {
@@ -319,11 +355,11 @@ static int radeon_fence_wait_seq(struct radeon_device *rdev, u64 *target_seq,
 		if (intr) {
 			r = wait_event_interruptible_timeout(rdev->fence_queue, (
 				(signaled = radeon_fence_any_seq_signaled(rdev, target_seq))
-				 || rdev->needs_reset), RADEON_FENCE_JIFFIES_TIMEOUT);
+				 || rdev->needs_reset), MAX_SCHEDULE_TIMEOUT);
 		} else {
 			r = wait_event_timeout(rdev->fence_queue, (
 				(signaled = radeon_fence_any_seq_signaled(rdev, target_seq))
-				 || rdev->needs_reset), RADEON_FENCE_JIFFIES_TIMEOUT);
+				 || rdev->needs_reset), MAX_SCHEDULE_TIMEOUT);
 		}
 
 		for (i = 0; i < RADEON_NUM_RINGS; ++i) {
@@ -334,50 +370,11 @@ static int radeon_fence_wait_seq(struct radeon_device *rdev, u64 *target_seq,
 			trace_radeon_fence_wait_end(rdev->ddev, i, target_seq[i]);
 		}
 
-		if (unlikely(r < 0))
+		if (r < 0)
 			return r;
 
-		if (unlikely(!signaled)) {
-			if (rdev->needs_reset)
-				return -EDEADLK;
-
-			/* we were interrupted for some reason and fence
-			 * isn't signaled yet, resume waiting */
-			if (r)
-				continue;
-
-			for (i = 0; i < RADEON_NUM_RINGS; ++i) {
-				if (!target_seq[i])
-					continue;
-
-				if (last_seq[i] != atomic64_read(&rdev->fence_drv[i].last_seq))
-					break;
-			}
-
-			if (i != RADEON_NUM_RINGS)
-				continue;
-
-			for (i = 0; i < RADEON_NUM_RINGS; ++i) {
-				if (!target_seq[i])
-					continue;
-
-				if (radeon_ring_is_lockup(rdev, i, &rdev->ring[i]))
-					break;
-			}
-
-			if (i < RADEON_NUM_RINGS) {
-				/* good news we believe it's a lockup */
-				dev_warn(rdev->dev, "GPU lockup (waiting for "
-					 "0x%016llx last fence id 0x%016llx on"
-					 " ring %d)\n",
-					 target_seq[i], last_seq[i], i);
-
-				/* remember that we need an reset */
-				rdev->needs_reset = true;
-				wake_up_all(&rdev->fence_queue);
-				return -EDEADLK;
-			}
-		}
+		if (rdev->needs_reset)
+			return -EDEADLK;
 	}
 	return 0;
 }
@@ -711,6 +708,8 @@ static void radeon_fence_driver_init_ring(struct radeon_device *rdev, int ring)
 		rdev->fence_drv[ring].sync_seq[i] = 0;
 	atomic64_set(&rdev->fence_drv[ring].last_seq, 0);
 	rdev->fence_drv[ring].initialized = false;
+	INIT_DELAYED_WORK(&rdev->fence_drv[ring].fence_check_work, radeon_fence_driver_check_lockup);
+	rdev->fence_drv[ring].rdev = rdev;
 }
 
 /**
@@ -740,6 +739,17 @@ int radeon_fence_driver_init(struct radeon_device *rdev)
 }
 
 /**
+ * radeon_fence_cancel_delayed_check - cancel all delayed checks on a ring during lockup
+ *
+ * This prevents the lockup check from being done while suspend is running
+ * during a recovery from a lockup.
+ */
+void radeon_fence_cancel_delayed_check(struct radeon_device *rdev, int ring)
+{
+	cancel_delayed_work_sync(&rdev->fence_drv[ring].fence_check_work);
+}
+
+/**
  * radeon_fence_driver_fini - tear down the fence driver
  * for all possible rings.
  *
@@ -755,6 +765,8 @@ void radeon_fence_driver_fini(struct radeon_device *rdev)
 	for (ring = 0; ring < RADEON_NUM_RINGS; ring++) {
 		if (!rdev->fence_drv[ring].initialized)
 			continue;
+
+		cancel_delayed_work_sync(&rdev->fence_drv[ring].fence_check_work);
 		r = radeon_fence_wait_empty(rdev, ring);
 		if (r) {
 			/* no need to trigger GPU reset as we are unloading */
diff --git a/drivers/gpu/drm/radeon/radeon_ring.c b/drivers/gpu/drm/radeon/radeon_ring.c
index f8050f5429e2..594d871a6fce 100644
--- a/drivers/gpu/drm/radeon/radeon_ring.c
+++ b/drivers/gpu/drm/radeon/radeon_ring.c
@@ -261,6 +261,7 @@ int radeon_ib_ring_tests(struct radeon_device *rdev)
 
 		r = radeon_ib_test(rdev, i, ring);
 		if (r) {
+			radeon_fence_cancel_delayed_check(rdev, i);
 			ring->ready = false;
 			rdev->needs_reset = false;
 

Re: [PATCH 09/19] drm/radeon: handle lockup in delayed work, v2

[Christian König]

Am 31.07.2014 um 17:33 schrieb Maarten Lankhorst:
> Signed-off-by: Maarten Lankhorst <maarten.lankhorst@canonical.com>
> ---
> V1 had a nasty bug breaking gpu lockup recovery. The fix is not
> allowing radeon_fence_driver_check_lockup to take exclusive_lock,
> and kill it during lockup recovery instead.

That looks like the delayed work starts running as soon as we submit a 
fence, and not when it's needed for waiting.

Since it's a backup for failing IRQs I would rather put it into 
radeon_irq_kms.c and start/stop it when the IRQs are started/stoped.

Christian.

> ---
>   drivers/gpu/drm/radeon/radeon.h        |    3 +
>   drivers/gpu/drm/radeon/radeon_device.c |    5 +
>   drivers/gpu/drm/radeon/radeon_fence.c  |  124 ++++++++++++++++++--------------
>   drivers/gpu/drm/radeon/radeon_ring.c   |    1
>   4 files changed, 77 insertions(+), 56 deletions(-)
>
> diff --git a/drivers/gpu/drm/radeon/radeon.h b/drivers/gpu/drm/radeon/radeon.h
> index 60c47f829122..b01d88fc10cb 100644
> --- a/drivers/gpu/drm/radeon/radeon.h
> +++ b/drivers/gpu/drm/radeon/radeon.h
> @@ -347,6 +347,8 @@ struct radeon_fence_driver {
>   	uint64_t			sync_seq[RADEON_NUM_RINGS];
>   	atomic64_t			last_seq;
>   	bool				initialized;
> +	struct delayed_work	fence_check_work;
> +	struct radeon_device	*rdev;
>   };
>   
>   struct radeon_fence {
> @@ -360,6 +362,7 @@ struct radeon_fence {
>   
>   int radeon_fence_driver_start_ring(struct radeon_device *rdev, int ring);
>   int radeon_fence_driver_init(struct radeon_device *rdev);
> +void radeon_fence_cancel_delayed_check(struct radeon_device *rdev, int ring);
>   void radeon_fence_driver_fini(struct radeon_device *rdev);
>   void radeon_fence_driver_force_completion(struct radeon_device *rdev);
>   int radeon_fence_emit(struct radeon_device *rdev, struct radeon_fence **fence, int ring);
> diff --git a/drivers/gpu/drm/radeon/radeon_device.c b/drivers/gpu/drm/radeon/radeon_device.c
> index 697add2cd4e3..21efd32d07ee 100644
> --- a/drivers/gpu/drm/radeon/radeon_device.c
> +++ b/drivers/gpu/drm/radeon/radeon_device.c
> @@ -1637,6 +1637,11 @@ int radeon_gpu_reset(struct radeon_device *rdev)
>   	radeon_save_bios_scratch_regs(rdev);
>   	/* block TTM */
>   	resched = ttm_bo_lock_delayed_workqueue(&rdev->mman.bdev);
> +
> +	/* kill all the hangcheck tests too */
> +	for (i = 0; i < RADEON_NUM_RINGS; ++i)
> +		radeon_fence_cancel_delayed_check(rdev, i);
> +
>   	radeon_pm_suspend(rdev);
>   	radeon_suspend(rdev);
>   
> diff --git a/drivers/gpu/drm/radeon/radeon_fence.c b/drivers/gpu/drm/radeon/radeon_fence.c
> index 913787085dfa..c055acc3a271 100644
> --- a/drivers/gpu/drm/radeon/radeon_fence.c
> +++ b/drivers/gpu/drm/radeon/radeon_fence.c
> @@ -125,16 +125,7 @@ int radeon_fence_emit(struct radeon_device *rdev,
>   	return 0;
>   }
>   
> -/**
> - * radeon_fence_process - process a fence
> - *
> - * @rdev: radeon_device pointer
> - * @ring: ring index the fence is associated with
> - *
> - * Checks the current fence value and wakes the fence queue
> - * if the sequence number has increased (all asics).
> - */
> -void radeon_fence_process(struct radeon_device *rdev, int ring)
> +static bool __radeon_fence_process(struct radeon_device *rdev, int ring)
>   {
>   	uint64_t seq, last_seq, last_emitted;
>   	unsigned count_loop = 0;
> @@ -190,7 +181,51 @@ void radeon_fence_process(struct radeon_device *rdev, int ring)
>   		}
>   	} while (atomic64_xchg(&rdev->fence_drv[ring].last_seq, seq) > seq);
>   
> -	if (wake)
> +	if (seq < last_emitted)
> +		mod_delayed_work(system_power_efficient_wq,
> +				 &rdev->fence_drv[ring].fence_check_work,
> +				 RADEON_FENCE_JIFFIES_TIMEOUT);
> +
> +	return wake;
> +}
> +
> +static void radeon_fence_driver_check_lockup(struct work_struct *work)
> +{
> +	struct radeon_fence_driver *fence_drv;
> +	struct radeon_device *rdev;
> +	unsigned long iring;
> +
> +	fence_drv = container_of(work, struct radeon_fence_driver, fence_check_work.work);
> +	rdev = fence_drv->rdev;
> +	iring = fence_drv - &rdev->fence_drv[0];
> +
> +	if (__radeon_fence_process(rdev, iring))
> +		wake_up_all(&rdev->fence_queue);
> +	else if (radeon_ring_is_lockup(rdev, iring, &rdev->ring[iring])) {
> +		/* good news we believe it's a lockup */
> +		dev_warn(rdev->dev, "GPU lockup (current fence id "
> +			 "0x%016llx last fence id 0x%016llx on ring %ld)\n",
> +			 (uint64_t)atomic64_read(&fence_drv->last_seq),
> +			 fence_drv->sync_seq[iring], iring);
> +
> +		/* remember that we need an reset */
> +		rdev->needs_reset = true;
> +		wake_up_all(&rdev->fence_queue);
> +	}
> +}
> +
> +/**
> + * radeon_fence_process - process a fence
> + *
> + * @rdev: radeon_device pointer
> + * @ring: ring index the fence is associated with
> + *
> + * Checks the current fence value and wakes the fence queue
> + * if the sequence number has increased (all asics).
> + */
> +void radeon_fence_process(struct radeon_device *rdev, int ring)
> +{
> +	if (__radeon_fence_process(rdev, ring))
>   		wake_up_all(&rdev->fence_queue);
>   }
>   
> @@ -302,9 +337,10 @@ static int radeon_fence_wait_seq(struct radeon_device *rdev, u64 *target_seq,
>   {
>   	uint64_t last_seq[RADEON_NUM_RINGS];
>   	bool signaled;
> -	int i, r;
> +	int i;
>   
>   	while (!radeon_fence_any_seq_signaled(rdev, target_seq)) {
> +		long r;
>   
>   		/* Save current sequence values, used to check for GPU lockups */
>   		for (i = 0; i < RADEON_NUM_RINGS; ++i) {
> @@ -319,11 +355,11 @@ static int radeon_fence_wait_seq(struct radeon_device *rdev, u64 *target_seq,
>   		if (intr) {
>   			r = wait_event_interruptible_timeout(rdev->fence_queue, (
>   				(signaled = radeon_fence_any_seq_signaled(rdev, target_seq))
> -				 || rdev->needs_reset), RADEON_FENCE_JIFFIES_TIMEOUT);
> +				 || rdev->needs_reset), MAX_SCHEDULE_TIMEOUT);
>   		} else {
>   			r = wait_event_timeout(rdev->fence_queue, (
>   				(signaled = radeon_fence_any_seq_signaled(rdev, target_seq))
> -				 || rdev->needs_reset), RADEON_FENCE_JIFFIES_TIMEOUT);
> +				 || rdev->needs_reset), MAX_SCHEDULE_TIMEOUT);
>   		}
>   
>   		for (i = 0; i < RADEON_NUM_RINGS; ++i) {
> @@ -334,50 +370,11 @@ static int radeon_fence_wait_seq(struct radeon_device *rdev, u64 *target_seq,
>   			trace_radeon_fence_wait_end(rdev->ddev, i, target_seq[i]);
>   		}
>   
> -		if (unlikely(r < 0))
> +		if (r < 0)
>   			return r;
>   
> -		if (unlikely(!signaled)) {
> -			if (rdev->needs_reset)
> -				return -EDEADLK;
> -
> -			/* we were interrupted for some reason and fence
> -			 * isn't signaled yet, resume waiting */
> -			if (r)
> -				continue;
> -
> -			for (i = 0; i < RADEON_NUM_RINGS; ++i) {
> -				if (!target_seq[i])
> -					continue;
> -
> -				if (last_seq[i] != atomic64_read(&rdev->fence_drv[i].last_seq))
> -					break;
> -			}
> -
> -			if (i != RADEON_NUM_RINGS)
> -				continue;
> -
> -			for (i = 0; i < RADEON_NUM_RINGS; ++i) {
> -				if (!target_seq[i])
> -					continue;
> -
> -				if (radeon_ring_is_lockup(rdev, i, &rdev->ring[i]))
> -					break;
> -			}
> -
> -			if (i < RADEON_NUM_RINGS) {
> -				/* good news we believe it's a lockup */
> -				dev_warn(rdev->dev, "GPU lockup (waiting for "
> -					 "0x%016llx last fence id 0x%016llx on"
> -					 " ring %d)\n",
> -					 target_seq[i], last_seq[i], i);
> -
> -				/* remember that we need an reset */
> -				rdev->needs_reset = true;
> -				wake_up_all(&rdev->fence_queue);
> -				return -EDEADLK;
> -			}
> -		}
> +		if (rdev->needs_reset)
> +			return -EDEADLK;
>   	}
>   	return 0;
>   }
> @@ -711,6 +708,8 @@ static void radeon_fence_driver_init_ring(struct radeon_device *rdev, int ring)
>   		rdev->fence_drv[ring].sync_seq[i] = 0;
>   	atomic64_set(&rdev->fence_drv[ring].last_seq, 0);
>   	rdev->fence_drv[ring].initialized = false;
> +	INIT_DELAYED_WORK(&rdev->fence_drv[ring].fence_check_work, radeon_fence_driver_check_lockup);
> +	rdev->fence_drv[ring].rdev = rdev;
>   }
>   
>   /**
> @@ -740,6 +739,17 @@ int radeon_fence_driver_init(struct radeon_device *rdev)
>   }
>   
>   /**
> + * radeon_fence_cancel_delayed_check - cancel all delayed checks on a ring during lockup
> + *
> + * This prevents the lockup check from being done while suspend is running
> + * during a recovery from a lockup.
> + */
> +void radeon_fence_cancel_delayed_check(struct radeon_device *rdev, int ring)
> +{
> +	cancel_delayed_work_sync(&rdev->fence_drv[ring].fence_check_work);
> +}
> +
> +/**
>    * radeon_fence_driver_fini - tear down the fence driver
>    * for all possible rings.
>    *
> @@ -755,6 +765,8 @@ void radeon_fence_driver_fini(struct radeon_device *rdev)
>   	for (ring = 0; ring < RADEON_NUM_RINGS; ring++) {
>   		if (!rdev->fence_drv[ring].initialized)
>   			continue;
> +
> +		cancel_delayed_work_sync(&rdev->fence_drv[ring].fence_check_work);
>   		r = radeon_fence_wait_empty(rdev, ring);
>   		if (r) {
>   			/* no need to trigger GPU reset as we are unloading */
> diff --git a/drivers/gpu/drm/radeon/radeon_ring.c b/drivers/gpu/drm/radeon/radeon_ring.c
> index f8050f5429e2..594d871a6fce 100644
> --- a/drivers/gpu/drm/radeon/radeon_ring.c
> +++ b/drivers/gpu/drm/radeon/radeon_ring.c
> @@ -261,6 +261,7 @@ int radeon_ib_ring_tests(struct radeon_device *rdev)
>   
>   		r = radeon_ib_test(rdev, i, ring);
>   		if (r) {
> +			radeon_fence_cancel_delayed_check(rdev, i);
>   			ring->ready = false;
>   			rdev->needs_reset = false;
>   
>

Re: [PATCH 09/19] drm/radeon: handle lockup in delayed work, v2

[Maarten Lankhorst]

On 01-08-14 18:35, Christian König wrote:
> Am 31.07.2014 um 17:33 schrieb Maarten Lankhorst:
>> Signed-off-by: Maarten Lankhorst <maarten.lankhorst@canonical.com>
>> ---
>> V1 had a nasty bug breaking gpu lockup recovery. The fix is not
>> allowing radeon_fence_driver_check_lockup to take exclusive_lock,
>> and kill it during lockup recovery instead.
> 
> That looks like the delayed work starts running as soon as we submit a fence, and not when it's needed for waiting.
> 
> Since it's a backup for failing IRQs I would rather put it into radeon_irq_kms.c and start/stop it when the IRQs are started/stoped.

The delayed work is not just for failing irq's, it's also the handler that's used to detect lockups, which is why I trigger after processing fences, and reset the timer after processing.

Specifically what happened was this scenario:

- lock up occurs
- write lock taken by gpu_reset
- delayed work runs, tries to acquire read lock, blocks
- gpu_reset tries to cancel delayed work synchronously
- has to wait for delayed work to finish -> deadlock

~Maarten

Re: [PATCH 09/19] drm/radeon: handle lockup in delayed work, v2

[Christian König]

Hi Maarten,

Sorry for the delay. I've got way to much todo recently.

Am 01.08.2014 um 19:46 schrieb Maarten Lankhorst:
>
> On 01-08-14 18:35, Christian König wrote:
>> Am 31.07.2014 um 17:33 schrieb Maarten Lankhorst:
>>> Signed-off-by: Maarten Lankhorst <maarten.lankhorst@canonical.com>
>>> ---
>>> V1 had a nasty bug breaking gpu lockup recovery. The fix is not
>>> allowing radeon_fence_driver_check_lockup to take exclusive_lock,
>>> and kill it during lockup recovery instead.
>> That looks like the delayed work starts running as soon as we submit a fence, and not when it's needed for waiting.
>>
>> Since it's a backup for failing IRQs I would rather put it into radeon_irq_kms.c and start/stop it when the IRQs are started/stoped.
> The delayed work is not just for failing irq's, it's also the handler that's used to detect lockups, which is why I trigger after processing fences, and reset the timer after processing.

The idea was turning the delayed work on and off when we turn the irq on 
and off as well, processing of the delayed work handler can still happen 
in radeon_fence.c

>
> Specifically what happened was this scenario:
>
> - lock up occurs
> - write lock taken by gpu_reset
> - delayed work runs, tries to acquire read lock, blocks
> - gpu_reset tries to cancel delayed work synchronously
> - has to wait for delayed work to finish -> deadlock

Why do you want to disable the work item from the lockup handler in the 
first place?

Just take the exclusive lock in the work item, when it concurrently runs 
with the lockup handler it will just block for the lockup handler to 
complete.

Regards,
Christian.

> ~Maarten

Re: [PATCH 09/19] drm/radeon: handle lockup in delayed work, v2

[Maarten Lankhorst]

op 04-08-14 10:36, Christian König schreef:
> Hi Maarten,
>
> Sorry for the delay. I've got way to much todo recently.
>
> Am 01.08.2014 um 19:46 schrieb Maarten Lankhorst:
>>
>> On 01-08-14 18:35, Christian König wrote:
>>> Am 31.07.2014 um 17:33 schrieb Maarten Lankhorst:
>>>> Signed-off-by: Maarten Lankhorst <maarten.lankhorst@canonical.com>
>>>> ---
>>>> V1 had a nasty bug breaking gpu lockup recovery. The fix is not
>>>> allowing radeon_fence_driver_check_lockup to take exclusive_lock,
>>>> and kill it during lockup recovery instead.
>>> That looks like the delayed work starts running as soon as we submit a fence, and not when it's needed for waiting.
>>>
>>> Since it's a backup for failing IRQs I would rather put it into radeon_irq_kms.c and start/stop it when the IRQs are started/stoped.
>> The delayed work is not just for failing irq's, it's also the handler that's used to detect lockups, which is why I trigger after processing fences, and reset the timer after processing.
>
> The idea was turning the delayed work on and off when we turn the irq on and off as well, processing of the delayed work handler can still happen in radeon_fence.c
>
>>
>> Specifically what happened was this scenario:
>>
>> - lock up occurs
>> - write lock taken by gpu_reset
>> - delayed work runs, tries to acquire read lock, blocks
>> - gpu_reset tries to cancel delayed work synchronously
>> - has to wait for delayed work to finish -> deadlock
>
> Why do you want to disable the work item from the lockup handler in the first place?
>
> Just take the exclusive lock in the work item, when it concurrently runs with the lockup handler it will just block for the lockup handler to complete.
With the delayed work radeon_fence_wait no longer handles unreliable interrupts itself, so it has to run from the lockup handler.
But an alternative solution could be adding a radeon_fence_wait_timeout, ignore the timeout and check if fence is signaled on timeout.
This would probably be a cleaner solution.

~Maarten

Re: [PATCH 09/19] drm/radeon: handle lockup in delayed work, v2

[Christian König]

Am 04.08.2014 um 10:55 schrieb Maarten Lankhorst:
> op 04-08-14 10:36, Christian König schreef:
>> Hi Maarten,
>>
>> Sorry for the delay. I've got way to much todo recently.
>>
>> Am 01.08.2014 um 19:46 schrieb Maarten Lankhorst:
>>> On 01-08-14 18:35, Christian König wrote:
>>>> Am 31.07.2014 um 17:33 schrieb Maarten Lankhorst:
>>>>> Signed-off-by: Maarten Lankhorst <maarten.lankhorst@canonical.com>
>>>>> ---
>>>>> V1 had a nasty bug breaking gpu lockup recovery. The fix is not
>>>>> allowing radeon_fence_driver_check_lockup to take exclusive_lock,
>>>>> and kill it during lockup recovery instead.
>>>> That looks like the delayed work starts running as soon as we submit a fence, and not when it's needed for waiting.
>>>>
>>>> Since it's a backup for failing IRQs I would rather put it into radeon_irq_kms.c and start/stop it when the IRQs are started/stoped.
>>> The delayed work is not just for failing irq's, it's also the handler that's used to detect lockups, which is why I trigger after processing fences, and reset the timer after processing.
>> The idea was turning the delayed work on and off when we turn the irq on and off as well, processing of the delayed work handler can still happen in radeon_fence.c
>>
>>> Specifically what happened was this scenario:
>>>
>>> - lock up occurs
>>> - write lock taken by gpu_reset
>>> - delayed work runs, tries to acquire read lock, blocks
>>> - gpu_reset tries to cancel delayed work synchronously
>>> - has to wait for delayed work to finish -> deadlock
>> Why do you want to disable the work item from the lockup handler in the first place?
>>
>> Just take the exclusive lock in the work item, when it concurrently runs with the lockup handler it will just block for the lockup handler to complete.
> With the delayed work radeon_fence_wait no longer handles unreliable interrupts itself, so it has to run from the lockup handler.
> But an alternative solution could be adding a radeon_fence_wait_timeout, ignore the timeout and check if fence is signaled on timeout.
> This would probably be a cleaner solution.

Yeah, agree. Manually specifying a timeout in the fence wait on lockup 
handling sounds like the best alternative to me.

Christian.

>
> ~Maarten
>

Re: [PATCH 09/19] drm/radeon: handle lockup in delayed work, v2

[Maarten Lankhorst]

Hey,

op 04-08-14 13:57, Christian König schreef:
> Am 04.08.2014 um 10:55 schrieb Maarten Lankhorst:
>> op 04-08-14 10:36, Christian König schreef:
>>> Hi Maarten,
>>>
>>> Sorry for the delay. I've got way to much todo recently.
>>>
>>> Am 01.08.2014 um 19:46 schrieb Maarten Lankhorst:
>>>> On 01-08-14 18:35, Christian König wrote:
>>>>> Am 31.07.2014 um 17:33 schrieb Maarten Lankhorst:
>>>>>> Signed-off-by: Maarten Lankhorst <maarten.lankhorst@canonical.com>
>>>>>> ---
>>>>>> V1 had a nasty bug breaking gpu lockup recovery. The fix is not
>>>>>> allowing radeon_fence_driver_check_lockup to take exclusive_lock,
>>>>>> and kill it during lockup recovery instead.
>>>>> That looks like the delayed work starts running as soon as we submit a fence, and not when it's needed for waiting.
>>>>>
>>>>> Since it's a backup for failing IRQs I would rather put it into radeon_irq_kms.c and start/stop it when the IRQs are started/stoped.
>>>> The delayed work is not just for failing irq's, it's also the handler that's used to detect lockups, which is why I trigger after processing fences, and reset the timer after processing.
>>> The idea was turning the delayed work on and off when we turn the irq on and off as well, processing of the delayed work handler can still happen in radeon_fence.c
>>>
>>>> Specifically what happened was this scenario:
>>>>
>>>> - lock up occurs
>>>> - write lock taken by gpu_reset
>>>> - delayed work runs, tries to acquire read lock, blocks
>>>> - gpu_reset tries to cancel delayed work synchronously
>>>> - has to wait for delayed work to finish -> deadlock
>>> Why do you want to disable the work item from the lockup handler in the first place?
>>>
>>> Just take the exclusive lock in the work item, when it concurrently runs with the lockup handler it will just block for the lockup handler to complete.
>> With the delayed work radeon_fence_wait no longer handles unreliable interrupts itself, so it has to run from the lockup handler.
>> But an alternative solution could be adding a radeon_fence_wait_timeout, ignore the timeout and check if fence is signaled on timeout.
>> This would probably be a cleaner solution.
>
> Yeah, agree. Manually specifying a timeout in the fence wait on lockup handling sounds like the best alternative to me.
It'a pain to deal with gpu reset.

I've now tried other solutions but that would mean reverting to the old style during gpu lockup recovery, and only running the delayed work when !lockup.
But this meant that the timeout was useless to add. I think the cleanest is keeping the v2 patch, because potentially any waiting code can be called during lockup recovery.

~Maarten

Re: [PATCH 09/19] drm/radeon: handle lockup in delayed work, v2

[Christian König]

> It'a pain to deal with gpu reset.

Yeah, well that's nothing new.

> I've now tried other solutions but that would mean reverting to the old style during gpu lockup recovery, and only running the delayed work when !lockup.
> But this meant that the timeout was useless to add. I think the cleanest is keeping the v2 patch, because potentially any waiting code can be called during lockup recovery.

The lockup code itself should never call any waiting code and V2 doesn't 
seem to handle a couple of cases correctly either.

How about moving the fence waiting out of the reset code?

Christian.

Am 04.08.2014 um 15:34 schrieb Maarten Lankhorst:
> Hey,
>
> op 04-08-14 13:57, Christian König schreef:
>> Am 04.08.2014 um 10:55 schrieb Maarten Lankhorst:
>>> op 04-08-14 10:36, Christian König schreef:
>>>> Hi Maarten,
>>>>
>>>> Sorry for the delay. I've got way to much todo recently.
>>>>
>>>> Am 01.08.2014 um 19:46 schrieb Maarten Lankhorst:
>>>>> On 01-08-14 18:35, Christian König wrote:
>>>>>> Am 31.07.2014 um 17:33 schrieb Maarten Lankhorst:
>>>>>>> Signed-off-by: Maarten Lankhorst <maarten.lankhorst@canonical.com>
>>>>>>> ---
>>>>>>> V1 had a nasty bug breaking gpu lockup recovery. The fix is not
>>>>>>> allowing radeon_fence_driver_check_lockup to take exclusive_lock,
>>>>>>> and kill it during lockup recovery instead.
>>>>>> That looks like the delayed work starts running as soon as we submit a fence, and not when it's needed for waiting.
>>>>>>
>>>>>> Since it's a backup for failing IRQs I would rather put it into radeon_irq_kms.c and start/stop it when the IRQs are started/stoped.
>>>>> The delayed work is not just for failing irq's, it's also the handler that's used to detect lockups, which is why I trigger after processing fences, and reset the timer after processing.
>>>> The idea was turning the delayed work on and off when we turn the irq on and off as well, processing of the delayed work handler can still happen in radeon_fence.c
>>>>
>>>>> Specifically what happened was this scenario:
>>>>>
>>>>> - lock up occurs
>>>>> - write lock taken by gpu_reset
>>>>> - delayed work runs, tries to acquire read lock, blocks
>>>>> - gpu_reset tries to cancel delayed work synchronously
>>>>> - has to wait for delayed work to finish -> deadlock
>>>> Why do you want to disable the work item from the lockup handler in the first place?
>>>>
>>>> Just take the exclusive lock in the work item, when it concurrently runs with the lockup handler it will just block for the lockup handler to complete.
>>> With the delayed work radeon_fence_wait no longer handles unreliable interrupts itself, so it has to run from the lockup handler.
>>> But an alternative solution could be adding a radeon_fence_wait_timeout, ignore the timeout and check if fence is signaled on timeout.
>>> This would probably be a cleaner solution.
>> Yeah, agree. Manually specifying a timeout in the fence wait on lockup handling sounds like the best alternative to me.
> It'a pain to deal with gpu reset.
>
> I've now tried other solutions but that would mean reverting to the old style during gpu lockup recovery, and only running the delayed work when !lockup.
> But this meant that the timeout was useless to add. I think the cleanest is keeping the v2 patch, because potentially any waiting code can be called during lockup recovery.
>
> ~Maarten
>

Re: [PATCH 09/19] drm/radeon: handle lockup in delayed work, v2

[Maarten Lankhorst]

op 04-08-14 16:37, Christian König schreef:
>> It'a pain to deal with gpu reset.
>
> Yeah, well that's nothing new.
>
>> I've now tried other solutions but that would mean reverting to the old style during gpu lockup recovery, and only running the delayed work when !lockup.
>> But this meant that the timeout was useless to add. I think the cleanest is keeping the v2 patch, because potentially any waiting code can be called during lockup recovery.
>
> The lockup code itself should never call any waiting code and V2 doesn't seem to handle a couple of cases correctly either.
>
> How about moving the fence waiting out of the reset code?
What cases did I miss then?

I'm curious how you want to move the fence waiting out of reset, when there are so many places that could potentially wait, like radeon_ib_get can call radeon_sa_bo_new which can do a wait, or radeon_ring_alloc that can wait on radeon_fence_wait_next, etc.

Re: [PATCH 09/19] drm/radeon: handle lockup in delayed work, v2

[Christian König]

Am 04.08.2014 um 16:40 schrieb Maarten Lankhorst:
> op 04-08-14 16:37, Christian König schreef:
>>> It'a pain to deal with gpu reset.
>> Yeah, well that's nothing new.
>>
>>> I've now tried other solutions but that would mean reverting to the old style during gpu lockup recovery, and only running the delayed work when !lockup.
>>> But this meant that the timeout was useless to add. I think the cleanest is keeping the v2 patch, because potentially any waiting code can be called during lockup recovery.
>> The lockup code itself should never call any waiting code and V2 doesn't seem to handle a couple of cases correctly either.
>>
>> How about moving the fence waiting out of the reset code?
> What cases did I miss then?
>
> I'm curious how you want to move the fence waiting out of reset, when there are so many places that could potentially wait, like radeon_ib_get can call radeon_sa_bo_new which can do a wait, or radeon_ring_alloc that can wait on radeon_fence_wait_next, etc.

The IB test itself doesn't needs to be protected by the exclusive lock. 
Only everything between radeon_save_bios_scratch_regs and 
radeon_ring_restore.

Christian.

Re: [PATCH 09/19] drm/radeon: handle lockup in delayed work, v2

[Maarten Lankhorst]

op 04-08-14 16:45, Christian König schreef:
> Am 04.08.2014 um 16:40 schrieb Maarten Lankhorst:
>> op 04-08-14 16:37, Christian König schreef:
>>>> It'a pain to deal with gpu reset.
>>> Yeah, well that's nothing new.
>>>
>>>> I've now tried other solutions but that would mean reverting to the old style during gpu lockup recovery, and only running the delayed work when !lockup.
>>>> But this meant that the timeout was useless to add. I think the cleanest is keeping the v2 patch, because potentially any waiting code can be called during lockup recovery.
>>> The lockup code itself should never call any waiting code and V2 doesn't seem to handle a couple of cases correctly either.
>>>
>>> How about moving the fence waiting out of the reset code?
>> What cases did I miss then?
>>
>> I'm curious how you want to move the fence waiting out of reset, when there are so many places that could potentially wait, like radeon_ib_get can call radeon_sa_bo_new which can do a wait, or radeon_ring_alloc that can wait on radeon_fence_wait_next, etc.
>
> The IB test itself doesn't needs to be protected by the exclusive lock. Only everything between radeon_save_bios_scratch_regs and radeon_ring_restore.
I'm not sure about that, what do you want to do if the ring tests fail? Do you have to retake the exclusive lock?

~Maarten

Re: [PATCH 09/19] drm/radeon: handle lockup in delayed work, v2

[Christian König]

Am 04.08.2014 um 16:58 schrieb Maarten Lankhorst:
> op 04-08-14 16:45, Christian König schreef:
>> Am 04.08.2014 um 16:40 schrieb Maarten Lankhorst:
>>> op 04-08-14 16:37, Christian König schreef:
>>>>> It'a pain to deal with gpu reset.
>>>> Yeah, well that's nothing new.
>>>>
>>>>> I've now tried other solutions but that would mean reverting to the old style during gpu lockup recovery, and only running the delayed work when !lockup.
>>>>> But this meant that the timeout was useless to add. I think the cleanest is keeping the v2 patch, because potentially any waiting code can be called during lockup recovery.
>>>> The lockup code itself should never call any waiting code and V2 doesn't seem to handle a couple of cases correctly either.
>>>>
>>>> How about moving the fence waiting out of the reset code?
>>> What cases did I miss then?
>>>
>>> I'm curious how you want to move the fence waiting out of reset, when there are so many places that could potentially wait, like radeon_ib_get can call radeon_sa_bo_new which can do a wait, or radeon_ring_alloc that can wait on radeon_fence_wait_next, etc.
>> The IB test itself doesn't needs to be protected by the exclusive lock. Only everything between radeon_save_bios_scratch_regs and radeon_ring_restore.
> I'm not sure about that, what do you want to do if the ring tests fail? Do you have to retake the exclusive lock?

Just set need_reset again and return -EAGAIN, that should have mostly 
the same effect as what we are doing right now.

Christian.

>
> ~Maarten
>

Re: [PATCH 09/19] drm/radeon: handle lockup in delayed work, v2

[Maarten Lankhorst]

op 04-08-14 17:04, Christian König schreef:
> Am 04.08.2014 um 16:58 schrieb Maarten Lankhorst:
>> op 04-08-14 16:45, Christian König schreef:
>>> Am 04.08.2014 um 16:40 schrieb Maarten Lankhorst:
>>>> op 04-08-14 16:37, Christian König schreef:
>>>>>> It'a pain to deal with gpu reset.
>>>>> Yeah, well that's nothing new.
>>>>>
>>>>>> I've now tried other solutions but that would mean reverting to the old style during gpu lockup recovery, and only running the delayed work when !lockup.
>>>>>> But this meant that the timeout was useless to add. I think the cleanest is keeping the v2 patch, because potentially any waiting code can be called during lockup recovery.
>>>>> The lockup code itself should never call any waiting code and V2 doesn't seem to handle a couple of cases correctly either.
>>>>>
>>>>> How about moving the fence waiting out of the reset code?
>>>> What cases did I miss then?
>>>>
>>>> I'm curious how you want to move the fence waiting out of reset, when there are so many places that could potentially wait, like radeon_ib_get can call radeon_sa_bo_new which can do a wait, or radeon_ring_alloc that can wait on radeon_fence_wait_next, etc.
>>> The IB test itself doesn't needs to be protected by the exclusive lock. Only everything between radeon_save_bios_scratch_regs and radeon_ring_restore.
>> I'm not sure about that, what do you want to do if the ring tests fail? Do you have to retake the exclusive lock?
>
> Just set need_reset again and return -EAGAIN, that should have mostly the same effect as what we are doing right now.
Yeah, except for the locking the ttm delayed workqueue, but that bool should be easy to save/restore.
I think this could work.

~Maarten

Re: [PATCH 09/19] drm/radeon: handle lockup in delayed work, v2

[Christian König]

Am 04.08.2014 um 17:09 schrieb Maarten Lankhorst:
> op 04-08-14 17:04, Christian König schreef:
>> Am 04.08.2014 um 16:58 schrieb Maarten Lankhorst:
>>> op 04-08-14 16:45, Christian König schreef:
>>>> Am 04.08.2014 um 16:40 schrieb Maarten Lankhorst:
>>>>> op 04-08-14 16:37, Christian König schreef:
>>>>>>> It'a pain to deal with gpu reset.
>>>>>> Yeah, well that's nothing new.
>>>>>>
>>>>>>> I've now tried other solutions but that would mean reverting to the old style during gpu lockup recovery, and only running the delayed work when !lockup.
>>>>>>> But this meant that the timeout was useless to add. I think the cleanest is keeping the v2 patch, because potentially any waiting code can be called during lockup recovery.
>>>>>> The lockup code itself should never call any waiting code and V2 doesn't seem to handle a couple of cases correctly either.
>>>>>>
>>>>>> How about moving the fence waiting out of the reset code?
>>>>> What cases did I miss then?
>>>>>
>>>>> I'm curious how you want to move the fence waiting out of reset, when there are so many places that could potentially wait, like radeon_ib_get can call radeon_sa_bo_new which can do a wait, or radeon_ring_alloc that can wait on radeon_fence_wait_next, etc.
>>>> The IB test itself doesn't needs to be protected by the exclusive lock. Only everything between radeon_save_bios_scratch_regs and radeon_ring_restore.
>>> I'm not sure about that, what do you want to do if the ring tests fail? Do you have to retake the exclusive lock?
>> Just set need_reset again and return -EAGAIN, that should have mostly the same effect as what we are doing right now.
> Yeah, except for the locking the ttm delayed workqueue, but that bool should be easy to save/restore.
> I think this could work.

Actually you could activate the delayed workqueue much earlier as well.

Thinking more about it that sounds like a bug in the current code, 
because we probably want the workqueue activated before waiting for the 
fence.

Christian.

>
> ~Maarten
>

Re: [PATCH 09/19] drm/radeon: handle lockup in delayed work, v2

[Daniel Vetter]

On Mon, Aug 04, 2014 at 07:04:46PM +0200, Christian König wrote:
> Am 04.08.2014 um 17:09 schrieb Maarten Lankhorst:
> >op 04-08-14 17:04, Christian König schreef:
> >>Am 04.08.2014 um 16:58 schrieb Maarten Lankhorst:
> >>>op 04-08-14 16:45, Christian König schreef:
> >>>>Am 04.08.2014 um 16:40 schrieb Maarten Lankhorst:
> >>>>>op 04-08-14 16:37, Christian König schreef:
> >>>>>>>It'a pain to deal with gpu reset.
> >>>>>>Yeah, well that's nothing new.
> >>>>>>
> >>>>>>>I've now tried other solutions but that would mean reverting to the old style during gpu lockup recovery, and only running the delayed work when !lockup.
> >>>>>>>But this meant that the timeout was useless to add. I think the cleanest is keeping the v2 patch, because potentially any waiting code can be called during lockup recovery.
> >>>>>>The lockup code itself should never call any waiting code and V2 doesn't seem to handle a couple of cases correctly either.
> >>>>>>
> >>>>>>How about moving the fence waiting out of the reset code?
> >>>>>What cases did I miss then?
> >>>>>
> >>>>>I'm curious how you want to move the fence waiting out of reset, when there are so many places that could potentially wait, like radeon_ib_get can call radeon_sa_bo_new which can do a wait, or radeon_ring_alloc that can wait on radeon_fence_wait_next, etc.
> >>>>The IB test itself doesn't needs to be protected by the exclusive lock. Only everything between radeon_save_bios_scratch_regs and radeon_ring_restore.
> >>>I'm not sure about that, what do you want to do if the ring tests fail? Do you have to retake the exclusive lock?
> >>Just set need_reset again and return -EAGAIN, that should have mostly the same effect as what we are doing right now.
> >Yeah, except for the locking the ttm delayed workqueue, but that bool should be easy to save/restore.
> >I think this could work.
> 
> Actually you could activate the delayed workqueue much earlier as well.
> 
> Thinking more about it that sounds like a bug in the current code, because
> we probably want the workqueue activated before waiting for the fence.

We've actually had a similar issue on i915 where when userspace never
waited for rendering (some shitty userspace drivers did that way back) we
never noticed that the gpu died. So launching the hangcheck/stuck wait
worker (we have both too) right away is what we do now.
-Daniel
-- 
Daniel Vetter
Software Engineer, Intel Corporation
+41 (0) 79 365 57 48 - http://blog.ffwll.ch

Re: [PATCH 09/19] drm/radeon: handle lockup in delayed work, v2

[Maarten Lankhorst]

op 04-08-14 19:04, Christian König schreef:
> Am 04.08.2014 um 17:09 schrieb Maarten Lankhorst:
>> op 04-08-14 17:04, Christian König schreef:
>>> Am 04.08.2014 um 16:58 schrieb Maarten Lankhorst:
>>>> op 04-08-14 16:45, Christian König schreef:
>>>>> Am 04.08.2014 um 16:40 schrieb Maarten Lankhorst:
>>>>>> op 04-08-14 16:37, Christian König schreef:
>>>>>>>> It'a pain to deal with gpu reset.
>>>>>>> Yeah, well that's nothing new.
>>>>>>>
>>>>>>>> I've now tried other solutions but that would mean reverting to the old style during gpu lockup recovery, and only running the delayed work when !lockup.
>>>>>>>> But this meant that the timeout was useless to add. I think the cleanest is keeping the v2 patch, because potentially any waiting code can be called during lockup recovery.
>>>>>>> The lockup code itself should never call any waiting code and V2 doesn't seem to handle a couple of cases correctly either.
>>>>>>>
>>>>>>> How about moving the fence waiting out of the reset code?
>>>>>> What cases did I miss then?
>>>>>>
>>>>>> I'm curious how you want to move the fence waiting out of reset, when there are so many places that could potentially wait, like radeon_ib_get can call radeon_sa_bo_new which can do a wait, or radeon_ring_alloc that can wait on radeon_fence_wait_next, etc.
>>>>> The IB test itself doesn't needs to be protected by the exclusive lock. Only everything between radeon_save_bios_scratch_regs and radeon_ring_restore.
>>>> I'm not sure about that, what do you want to do if the ring tests fail? Do you have to retake the exclusive lock?
>>> Just set need_reset again and return -EAGAIN, that should have mostly the same effect as what we are doing right now.
>> Yeah, except for the locking the ttm delayed workqueue, but that bool should be easy to save/restore.
>> I think this could work.
>
> Actually you could activate the delayed workqueue much earlier as well.
>
> Thinking more about it that sounds like a bug in the current code, because we probably want the workqueue activated before waiting for the fence.

Ok, how about this?

Because of the downgrade_write, a second gpu reset can't be started until the first finishes.

I'm uncertain about it, I think I might either have to stop calling radeon_restore_bios_scratch_regs a second time,
or I should call save_bios_scratch_regs the second time around.

Also it feels like drm_helper_resume_force_mode should be called before downgrading exclusive_lock, but it might depend on PM restore.
Tough! Maybe move both calls to before downgrade_write?
 
commit 3644aae8581a15e3a935279287c397f7eab400ff
Author: Maarten Lankhorst <maarten.lankhorst@canonical.com>
Date:   Tue Aug 5 10:29:23 2014 +0200

    drm/radeon: take exclusive_lock in read mode during ring tests
    
    Signed-off-by: Maarten Lankhorst <maarten.lankhorst@canonical.com>

diff --git a/drivers/gpu/drm/radeon/radeon.h b/drivers/gpu/drm/radeon/radeon.h
index 29d9cc04c04e..de14e35da002 100644
--- a/drivers/gpu/drm/radeon/radeon.h
+++ b/drivers/gpu/drm/radeon/radeon.h
@@ -2286,7 +2286,7 @@ struct radeon_device {
 	bool				need_dma32;
 	bool				accel_working;
 	bool				fastfb_working; /* IGP feature*/
-	bool				needs_reset;
+	bool				needs_reset, in_reset;
 	struct radeon_surface_reg surface_regs[RADEON_GEM_MAX_SURFACES];
 	const struct firmware *me_fw;	/* all family ME firmware */
 	const struct firmware *pfp_fw;	/* r6/700 PFP firmware */
diff --git a/drivers/gpu/drm/radeon/radeon_device.c b/drivers/gpu/drm/radeon/radeon_device.c
index 03686fab842d..6317b8a2ef20 100644
--- a/drivers/gpu/drm/radeon/radeon_device.c
+++ b/drivers/gpu/drm/radeon/radeon_device.c
@@ -1620,37 +1620,44 @@ int radeon_gpu_reset(struct radeon_device *rdev)
 	unsigned ring_sizes[RADEON_NUM_RINGS];
 	uint32_t *ring_data[RADEON_NUM_RINGS];
 
-	bool saved = false;
+	bool saved;
 
 	int i, r;
 	int resched;
 
+retry:
+	saved = false;
 	down_write(&rdev->exclusive_lock);
 
 	if (!rdev->needs_reset) {
+		WARN_ON(rdev->in_reset);
 		up_write(&rdev->exclusive_lock);
 		return 0;
 	}
 
 	rdev->needs_reset = false;
-
-	radeon_save_bios_scratch_regs(rdev);
-	/* block TTM */
 	resched = ttm_bo_lock_delayed_workqueue(&rdev->mman.bdev);
-	radeon_pm_suspend(rdev);
-	radeon_suspend(rdev);
 
-	for (i = 0; i < RADEON_NUM_RINGS; ++i) {
-		ring_sizes[i] = radeon_ring_backup(rdev, &rdev->ring[i],
-						   &ring_data[i]);
-		if (ring_sizes[i]) {
-			saved = true;
-			dev_info(rdev->dev, "Saved %d dwords of commands "
-				 "on ring %d.\n", ring_sizes[i], i);
+	if (!rdev->in_reset) {
+		rdev->in_reset = true;
+
+		radeon_save_bios_scratch_regs(rdev);
+		/* block TTM */
+		radeon_pm_suspend(rdev);
+		radeon_suspend(rdev);
+
+		for (i = 0; i < RADEON_NUM_RINGS; ++i) {
+			ring_sizes[i] = radeon_ring_backup(rdev, &rdev->ring[i],
+							   &ring_data[i]);
+			if (ring_sizes[i]) {
+				saved = true;
+				dev_info(rdev->dev, "Saved %d dwords of commands "
+					 "on ring %d.\n", ring_sizes[i], i);
+			}
 		}
-	}
+	} else
+		memset(ring_data, 0, sizeof(ring_data));
 
-retry:
 	r = radeon_asic_reset(rdev);
 	if (!r) {
 		dev_info(rdev->dev, "GPU reset succeeded, trying to resume\n");
@@ -1659,40 +1666,46 @@ retry:
 
 	radeon_restore_bios_scratch_regs(rdev);
 
-	if (!r) {
+	if (!r && saved) {
 		for (i = 0; i < RADEON_NUM_RINGS; ++i) {
 			radeon_ring_restore(rdev, &rdev->ring[i],
 					    ring_sizes[i], ring_data[i]);
-			ring_sizes[i] = 0;
 			ring_data[i] = NULL;
 		}
+	} else {
+		radeon_fence_driver_force_completion(rdev);
+
+		for (i = 0; i < RADEON_NUM_RINGS; ++i) {
+			kfree(ring_data[i]);
+		}
+	}
+	downgrade_write(&rdev->exclusive_lock);
+	ttm_bo_unlock_delayed_workqueue(&rdev->mman.bdev, resched);
 
+	if (!r) {
 		r = radeon_ib_ring_tests(rdev);
 		if (r) {
 			dev_err(rdev->dev, "ib ring test failed (%d).\n", r);
 			if (saved) {
-				saved = false;
+				/* if reset fails, try without saving data */
+				rdev->needs_reset = true;
 				radeon_suspend(rdev);
+				up_read(&rdev->exclusive_lock);
 				goto retry;
 			}
 		}
-	} else {
-		radeon_fence_driver_force_completion(rdev);
-		for (i = 0; i < RADEON_NUM_RINGS; ++i) {
-			kfree(ring_data[i]);
-		}
 	}
 
 	radeon_pm_resume(rdev);
 	drm_helper_resume_force_mode(rdev->ddev);
 
-	ttm_bo_unlock_delayed_workqueue(&rdev->mman.bdev, resched);
 	if (r) {
 		/* bad news, how to tell it to userspace ? */
 		dev_info(rdev->dev, "GPU reset failed\n");
 	}
 
-	up_write(&rdev->exclusive_lock);
+	rdev->in_reset = false;
+	up_read(&rdev->exclusive_lock);
 	return r;
 }
 

[PATCH 10/19] drm/radeon: add timeout argument to radeon_fence_wait_seq

[Maarten Lankhorst]

This makes it possible to wait for a specific amount of time,
rather than wait until infinity.

Signed-off-by: Maarten Lankhorst <maarten.lankhorst@canonical.com>
Reviewed-by: Christian König <deathsimple@vodafone.de>
---
 drivers/gpu/drm/radeon/radeon_fence.c |   50 ++++++++++++++++++++-------------
 1 file changed, 30 insertions(+), 20 deletions(-)

diff --git a/drivers/gpu/drm/radeon/radeon_fence.c b/drivers/gpu/drm/radeon/radeon_fence.c
index c055acc3a271..f79e0b34582b 100644
--- a/drivers/gpu/drm/radeon/radeon_fence.c
+++ b/drivers/gpu/drm/radeon/radeon_fence.c
@@ -318,22 +318,25 @@ static bool radeon_fence_any_seq_signaled(struct radeon_device *rdev, u64 *seq)
 }
 
 /**
- * radeon_fence_wait_seq - wait for a specific sequence numbers
+ * radeon_fence_wait_seq_timeout - wait for a specific sequence numbers
  *
  * @rdev: radeon device pointer
  * @target_seq: sequence number(s) we want to wait for
  * @intr: use interruptable sleep
+ * @timeout: maximum time to wait, or MAX_SCHEDULE_TIMEOUT for infinite wait
  *
  * Wait for the requested sequence number(s) to be written by any ring
  * (all asics).  Sequnce number array is indexed by ring id.
  * @intr selects whether to use interruptable (true) or non-interruptable
  * (false) sleep when waiting for the sequence number.  Helper function
  * for radeon_fence_wait_*().
- * Returns 0 if the sequence number has passed, error for all other cases.
+ * Returns remaining time if the sequence number has passed, 0 when
+ * the wait timeout, or an error for all other cases.
  * -EDEADLK is returned when a GPU lockup has been detected.
  */
-static int radeon_fence_wait_seq(struct radeon_device *rdev, u64 *target_seq,
-				 bool intr)
+static long radeon_fence_wait_seq_timeout(struct radeon_device *rdev,
+					  u64 *target_seq, bool intr,
+					  long timeout)
 {
 	uint64_t last_seq[RADEON_NUM_RINGS];
 	bool signaled;
@@ -355,11 +358,11 @@ static int radeon_fence_wait_seq(struct radeon_device *rdev, u64 *target_seq,
 		if (intr) {
 			r = wait_event_interruptible_timeout(rdev->fence_queue, (
 				(signaled = radeon_fence_any_seq_signaled(rdev, target_seq))
-				 || rdev->needs_reset), MAX_SCHEDULE_TIMEOUT);
+				 || rdev->needs_reset), timeout);
 		} else {
 			r = wait_event_timeout(rdev->fence_queue, (
 				(signaled = radeon_fence_any_seq_signaled(rdev, target_seq))
-				 || rdev->needs_reset), MAX_SCHEDULE_TIMEOUT);
+				 || rdev->needs_reset), timeout);
 		}
 
 		for (i = 0; i < RADEON_NUM_RINGS; ++i) {
@@ -370,20 +373,22 @@ static int radeon_fence_wait_seq(struct radeon_device *rdev, u64 *target_seq,
 			trace_radeon_fence_wait_end(rdev->ddev, i, target_seq[i]);
 		}
 
-		if (r < 0)
+		if (r <= 0)
 			return r;
 
 		if (rdev->needs_reset)
 			return -EDEADLK;
+
+		timeout = r;
 	}
-	return 0;
+	return timeout;
 }
 
 /**
  * radeon_fence_wait - wait for a fence to signal
  *
  * @fence: radeon fence object
- * @intr: use interruptable sleep
+ * @intr: use interruptible sleep
  *
  * Wait for the requested fence to signal (all asics).
  * @intr selects whether to use interruptable (true) or non-interruptable
@@ -393,7 +398,7 @@ static int radeon_fence_wait_seq(struct radeon_device *rdev, u64 *target_seq,
 int radeon_fence_wait(struct radeon_fence *fence, bool intr)
 {
 	uint64_t seq[RADEON_NUM_RINGS] = {};
-	int r;
+	long r;
 
 	if (fence == NULL) {
 		WARN(1, "Querying an invalid fence : %p !\n", fence);
@@ -404,9 +409,10 @@ int radeon_fence_wait(struct radeon_fence *fence, bool intr)
 	if (seq[fence->ring] == RADEON_FENCE_SIGNALED_SEQ)
 		return 0;
 
-	r = radeon_fence_wait_seq(fence->rdev, seq, intr);
-	if (r)
+	r = radeon_fence_wait_seq_timeout(fence->rdev, seq, intr, MAX_SCHEDULE_TIMEOUT);
+	if (r < 0) {
 		return r;
+	}
 
 	fence->seq = RADEON_FENCE_SIGNALED_SEQ;
 	return 0;
@@ -431,7 +437,7 @@ int radeon_fence_wait_any(struct radeon_device *rdev,
 {
 	uint64_t seq[RADEON_NUM_RINGS];
 	unsigned i, num_rings = 0;
-	int r;
+	long r;
 
 	for (i = 0; i < RADEON_NUM_RINGS; ++i) {
 		seq[i] = 0;
@@ -452,8 +458,8 @@ int radeon_fence_wait_any(struct radeon_device *rdev,
 	if (num_rings == 0)
 		return -ENOENT;
 
-	r = radeon_fence_wait_seq(rdev, seq, intr);
-	if (r) {
+	r = radeon_fence_wait_seq_timeout(rdev, seq, intr, MAX_SCHEDULE_TIMEOUT);
+	if (r < 0) {
 		return r;
 	}
 	return 0;
@@ -472,6 +478,7 @@ int radeon_fence_wait_any(struct radeon_device *rdev,
 int radeon_fence_wait_next(struct radeon_device *rdev, int ring)
 {
 	uint64_t seq[RADEON_NUM_RINGS] = {};
+	long r;
 
 	seq[ring] = atomic64_read(&rdev->fence_drv[ring].last_seq) + 1ULL;
 	if (seq[ring] >= rdev->fence_drv[ring].sync_seq[ring]) {
@@ -479,7 +486,10 @@ int radeon_fence_wait_next(struct radeon_device *rdev, int ring)
 		   already the last emited fence */
 		return -ENOENT;
 	}
-	return radeon_fence_wait_seq(rdev, seq, false);
+	r = radeon_fence_wait_seq_timeout(rdev, seq, false, MAX_SCHEDULE_TIMEOUT);
+	if (r < 0)
+		return r;
+	return 0;
 }
 
 /**
@@ -495,18 +505,18 @@ int radeon_fence_wait_next(struct radeon_device *rdev, int ring)
 int radeon_fence_wait_empty(struct radeon_device *rdev, int ring)
 {
 	uint64_t seq[RADEON_NUM_RINGS] = {};
-	int r;
+	long r;
 
 	seq[ring] = rdev->fence_drv[ring].sync_seq[ring];
 	if (!seq[ring])
 		return 0;
 
-	r = radeon_fence_wait_seq(rdev, seq, false);
-	if (r) {
+	r = radeon_fence_wait_seq_timeout(rdev, seq, false, MAX_SCHEDULE_TIMEOUT);
+	if (r < 0) {
 		if (r == -EDEADLK)
 			return -EDEADLK;
 
-		dev_err(rdev->dev, "error waiting for ring[%d] to become idle (%d)\n",
+		dev_err(rdev->dev, "error waiting for ring[%d] to become idle (%ld)\n",
 			ring, r);
 	}
 	return 0;

[PATCH 11/19] drm/radeon: use common fence implementation for fences, v2

[Maarten Lankhorst]

Changes since v1:
- Kill the sw interrupt dance, add and use
  radeon_irq_kms_sw_irq_get_delayed instead.
- Change custom wait function, lockdep complained about it.
  Holding exclusive_lock in the wait function might cause deadlocks.
  Instead do all the processing in .enable_signaling, and wait
  on the global fence_queue to pick up gpu resets.
- Process all fences in radeon_gpu_reset after reset to close a race
  with the trylock in enable_signaling.

Signed-off-by: Maarten Lankhorst <maarten.lankhorst@canonical.com>
---
 drivers/gpu/drm/radeon/radeon.h         |   18 +-
 drivers/gpu/drm/radeon/radeon_device.c  |   22 ++-
 drivers/gpu/drm/radeon/radeon_fence.c   |  263 +++++++++++++++++++++++++++----
 drivers/gpu/drm/radeon/radeon_irq_kms.c |   43 +++++
 4 files changed, 293 insertions(+), 53 deletions(-)

diff --git a/drivers/gpu/drm/radeon/radeon.h b/drivers/gpu/drm/radeon/radeon.h
index b01d88fc10cb..599bb28e4104 100644
--- a/drivers/gpu/drm/radeon/radeon.h
+++ b/drivers/gpu/drm/radeon/radeon.h
@@ -64,6 +64,7 @@
 #include <linux/wait.h>
 #include <linux/list.h>
 #include <linux/kref.h>
+#include <linux/fence.h>
 
 #include <ttm/ttm_bo_api.h>
 #include <ttm/ttm_bo_driver.h>
@@ -116,9 +117,6 @@ extern int radeon_deep_color;
 #define RADEONFB_CONN_LIMIT			4
 #define RADEON_BIOS_NUM_SCRATCH			8
 
-/* fence seq are set to this number when signaled */
-#define RADEON_FENCE_SIGNALED_SEQ		0LL
-
 /* internal ring indices */
 /* r1xx+ has gfx CP ring */
 #define RADEON_RING_TYPE_GFX_INDEX		0
@@ -352,12 +350,15 @@ struct radeon_fence_driver {
 };
 
 struct radeon_fence {
+	struct fence base;
+
 	struct radeon_device		*rdev;
-	struct kref			kref;
 	/* protected by radeon_fence.lock */
 	uint64_t			seq;
 	/* RB, DMA, etc. */
 	unsigned			ring;
+
+	wait_queue_t fence_wake;
 };
 
 int radeon_fence_driver_start_ring(struct radeon_device *rdev, int ring);
@@ -769,6 +770,7 @@ struct radeon_irq {
 int radeon_irq_kms_init(struct radeon_device *rdev);
 void radeon_irq_kms_fini(struct radeon_device *rdev);
 void radeon_irq_kms_sw_irq_get(struct radeon_device *rdev, int ring);
+void radeon_irq_kms_sw_irq_get_delayed(struct radeon_device *rdev, int ring);
 void radeon_irq_kms_sw_irq_put(struct radeon_device *rdev, int ring);
 void radeon_irq_kms_pflip_irq_get(struct radeon_device *rdev, int crtc);
 void radeon_irq_kms_pflip_irq_put(struct radeon_device *rdev, int crtc);
@@ -2276,6 +2278,7 @@ struct radeon_device {
 	struct radeon_mman		mman;
 	struct radeon_fence_driver	fence_drv[RADEON_NUM_RINGS];
 	wait_queue_head_t		fence_queue;
+	unsigned			fence_context;
 	struct mutex			ring_lock;
 	struct radeon_ring		ring[RADEON_NUM_RINGS];
 	bool				ib_pool_ready;
@@ -2314,6 +2317,8 @@ struct radeon_device {
 	struct work_struct hotplug_work;
 	struct work_struct audio_work;
 	struct work_struct reset_work;
+	struct work_struct delayed_irq_work;
+
 	int num_crtc; /* number of crtcs */
 	struct mutex dc_hw_i2c_mutex; /* display controller hw i2c mutex */
 	bool has_uvd;
@@ -2366,11 +2371,6 @@ u32 cik_mm_rdoorbell(struct radeon_device *rdev, u32 index);
 void cik_mm_wdoorbell(struct radeon_device *rdev, u32 index, u32 v);
 
 /*
- * Cast helper
- */
-#define to_radeon_fence(p) ((struct radeon_fence *)(p))
-
-/*
  * Registers read & write functions.
  */
 #define RREG8(reg) readb((rdev->rmmio) + (reg))
diff --git a/drivers/gpu/drm/radeon/radeon_device.c b/drivers/gpu/drm/radeon/radeon_device.c
index 21efd32d07ee..dbf3a8bc91d2 100644
--- a/drivers/gpu/drm/radeon/radeon_device.c
+++ b/drivers/gpu/drm/radeon/radeon_device.c
@@ -1213,6 +1213,7 @@ int radeon_device_init(struct radeon_device *rdev,
 	for (i = 0; i < RADEON_NUM_RINGS; i++) {
 		rdev->ring[i].idx = i;
 	}
+	rdev->fence_context = fence_context_alloc(RADEON_NUM_RINGS);
 
 	DRM_INFO("initializing kernel modesetting (%s 0x%04X:0x%04X 0x%04X:0x%04X).\n",
 		radeon_family_name[rdev->family], pdev->vendor, pdev->device,
@@ -1622,15 +1623,13 @@ int radeon_gpu_reset(struct radeon_device *rdev)
 
 	bool saved = false;
 
-	int i, r;
+	int i, r = 0;
 	int resched;
 
 	down_write(&rdev->exclusive_lock);
 
-	if (!rdev->needs_reset) {
-		up_write(&rdev->exclusive_lock);
-		return 0;
-	}
+	if (!rdev->needs_reset)
+		goto out;
 
 	rdev->needs_reset = false;
 
@@ -1697,7 +1696,18 @@ retry:
 		dev_info(rdev->dev, "GPU reset failed\n");
 	}
 
-	up_write(&rdev->exclusive_lock);
+out:
+	/*
+	 * force all waiters to recheck, some may have been
+	 * added while the exclusive_lock was unavailable
+	 */
+	downgrade_write(&rdev->exclusive_lock);
+	for (i = 0; i < RADEON_NUM_RINGS; ++i) {
+		if (rdev->ring[i].ready)
+			radeon_fence_process(rdev, i);
+	}
+	wake_up_all(&rdev->fence_queue);
+	up_read(&rdev->exclusive_lock);
 	return r;
 }
 
diff --git a/drivers/gpu/drm/radeon/radeon_fence.c b/drivers/gpu/drm/radeon/radeon_fence.c
index f79e0b34582b..8ca56ca70503 100644
--- a/drivers/gpu/drm/radeon/radeon_fence.c
+++ b/drivers/gpu/drm/radeon/radeon_fence.c
@@ -39,6 +39,15 @@
 #include "radeon.h"
 #include "radeon_trace.h"
 
+static const struct fence_ops radeon_fence_ops;
+
+#define to_radeon_fence(p) \
+	({								\
+		struct radeon_fence *__f;				\
+		__f = container_of((p), struct radeon_fence, base);	\
+		__f->base.ops == &radeon_fence_ops ? __f : NULL;	\
+	})
+
 /*
  * Fences
  * Fences mark an event in the GPUs pipeline and are used
@@ -111,20 +120,82 @@ int radeon_fence_emit(struct radeon_device *rdev,
 		      struct radeon_fence **fence,
 		      int ring)
 {
+	u64 seq = ++rdev->fence_drv[ring].sync_seq[ring];
+
 	/* we are protected by the ring emission mutex */
 	*fence = kmalloc(sizeof(struct radeon_fence), GFP_KERNEL);
 	if ((*fence) == NULL) {
 		return -ENOMEM;
 	}
-	kref_init(&((*fence)->kref));
-	(*fence)->rdev = rdev;
-	(*fence)->seq = ++rdev->fence_drv[ring].sync_seq[ring];
 	(*fence)->ring = ring;
+	fence_init(&(*fence)->base, &radeon_fence_ops,
+		   &rdev->fence_queue.lock, rdev->fence_context + ring, seq);
+	(*fence)->rdev = rdev;
+	(*fence)->seq = seq;
 	radeon_fence_ring_emit(rdev, ring, *fence);
 	trace_radeon_fence_emit(rdev->ddev, ring, (*fence)->seq);
 	return 0;
 }
 
+/**
+ * __radeon_fence_read_seq - Returns the current fence value without updating
+ *
+ * @rdev: radeon_device pointer
+ * @ring: ring index to return the seqno of
+ */
+static uint64_t __radeon_fence_read_seq(struct radeon_device *rdev, int ring)
+{
+	uint64_t last_seq = atomic64_read(&rdev->fence_drv[ring].last_seq);
+	uint64_t last_emitted = rdev->fence_drv[ring].sync_seq[ring];
+	uint64_t seq = radeon_fence_read(rdev, ring);
+
+	seq = radeon_fence_read(rdev, ring);
+	seq |= last_seq & 0xffffffff00000000LL;
+	if (seq < last_seq) {
+		seq &= 0xffffffff;
+		seq |= last_emitted & 0xffffffff00000000LL;
+	}
+	return seq;
+}
+
+/**
+ * radeon_fence_check_signaled - callback from fence_queue
+ *
+ * this function is called with fence_queue lock held, which is also used
+ * for the fence locking itself, so unlocked variants are used for
+ * fence_signal, and remove_wait_queue.
+ */
+static int radeon_fence_check_signaled(wait_queue_t *wait, unsigned mode, int flags, void *key)
+{
+	struct radeon_fence *fence;
+	u64 seq;
+
+	fence = container_of(wait, struct radeon_fence, fence_wake);
+
+	/*
+	 * We cannot use radeon_fence_process here because we're already
+	 * in the waitqueue, in a call from wake_up_all.
+	 *
+	 * Because the current ring might not have been updated, call
+	 * __radeon_fence_read_seq here to check without updating.
+	 */
+	seq = __radeon_fence_read_seq(fence->rdev, fence->ring);
+	if (seq >= fence->seq) {
+		int ret = fence_signal_locked(&fence->base);
+
+		if (!ret)
+			FENCE_TRACE(&fence->base, "signaled from irq context\n");
+		else
+			FENCE_TRACE(&fence->base, "was already signaled\n");
+
+		radeon_irq_kms_sw_irq_put(fence->rdev, fence->ring);
+		__remove_wait_queue(&fence->rdev->fence_queue, &fence->fence_wake);
+		fence_put(&fence->base);
+	} else
+		FENCE_TRACE(&fence->base, "pending\n");
+	return 0;
+}
+
 static bool __radeon_fence_process(struct radeon_device *rdev, int ring)
 {
 	uint64_t seq, last_seq, last_emitted;
@@ -230,21 +301,6 @@ void radeon_fence_process(struct radeon_device *rdev, int ring)
 }
 
 /**
- * radeon_fence_destroy - destroy a fence
- *
- * @kref: fence kref
- *
- * Frees the fence object (all asics).
- */
-static void radeon_fence_destroy(struct kref *kref)
-{
-	struct radeon_fence *fence;
-
-	fence = container_of(kref, struct radeon_fence, kref);
-	kfree(fence);
-}
-
-/**
  * radeon_fence_seq_signaled - check if a fence sequence number has signaled
  *
  * @rdev: radeon device pointer
@@ -272,6 +328,73 @@ static bool radeon_fence_seq_signaled(struct radeon_device *rdev,
 	return false;
 }
 
+static bool __radeon_fence_signaled(struct fence *f)
+{
+	struct radeon_fence *fence = to_radeon_fence(f);
+	struct radeon_device *rdev = fence->rdev;
+	unsigned ring = fence->ring;
+	u64 seq = fence->seq;
+
+	if (atomic64_read(&rdev->fence_drv[ring].last_seq) >= seq) {
+		return true;
+	}
+
+	if (down_read_trylock(&rdev->exclusive_lock)) {
+		radeon_fence_process(rdev, ring);
+		up_read(&rdev->exclusive_lock);
+
+		if (atomic64_read(&rdev->fence_drv[ring].last_seq) >= seq) {
+			return true;
+		}
+	}
+	return false;
+}
+
+/**
+ * radeon_fence_enable_signaling - enable signalling on fence
+ * @fence: fence
+ *
+ * This function is called with fence_queue lock held, and adds a callback
+ * to fence_queue that checks if this fence is signaled, and if so it
+ * signals the fence and removes itself.
+ */
+static bool radeon_fence_enable_signaling(struct fence *f)
+{
+	struct radeon_fence *fence = to_radeon_fence(f);
+	struct radeon_device *rdev = fence->rdev;
+
+	if (atomic64_read(&rdev->fence_drv[fence->ring].last_seq) >= fence->seq ||
+	    !rdev->ddev->irq_enabled)
+		return false;
+
+	if (down_read_trylock(&rdev->exclusive_lock)) {
+		radeon_irq_kms_sw_irq_get(rdev, fence->ring);
+
+		if (__radeon_fence_process(rdev, fence->ring))
+			wake_up_all_locked(&rdev->fence_queue);
+
+		/* did fence get signaled after we enabled the sw irq? */
+		if (atomic64_read(&rdev->fence_drv[fence->ring].last_seq) >= fence->seq) {
+			radeon_irq_kms_sw_irq_put(rdev, fence->ring);
+			up_read(&rdev->exclusive_lock);
+			return false;
+		}
+
+		up_read(&rdev->exclusive_lock);
+	} else
+		/* we're probably in a lockup, lets not fiddle too much */
+		radeon_irq_kms_sw_irq_get_delayed(rdev, fence->ring);
+
+	fence->fence_wake.flags = 0;
+	fence->fence_wake.private = NULL;
+	fence->fence_wake.func = radeon_fence_check_signaled;
+	__add_wait_queue(&rdev->fence_queue, &fence->fence_wake);
+	fence_get(f);
+
+	FENCE_TRACE(&fence->base, "armed on ring %i!\n", fence->ring);
+	return true;
+}
+
 /**
  * radeon_fence_signaled - check if a fence has signaled
  *
@@ -285,11 +408,13 @@ bool radeon_fence_signaled(struct radeon_fence *fence)
 	if (!fence) {
 		return true;
 	}
-	if (fence->seq == RADEON_FENCE_SIGNALED_SEQ) {
-		return true;
-	}
+
 	if (radeon_fence_seq_signaled(fence->rdev, fence->seq, fence->ring)) {
-		fence->seq = RADEON_FENCE_SIGNALED_SEQ;
+		int ret;
+
+		ret = fence_signal(&fence->base);
+		if (!ret)
+			FENCE_TRACE(&fence->base, "signaled from radeon_fence_signaled\n");
 		return true;
 	}
 	return false;
@@ -400,21 +525,18 @@ int radeon_fence_wait(struct radeon_fence *fence, bool intr)
 	uint64_t seq[RADEON_NUM_RINGS] = {};
 	long r;
 
-	if (fence == NULL) {
-		WARN(1, "Querying an invalid fence : %p !\n", fence);
-		return -EINVAL;
-	}
-
-	seq[fence->ring] = fence->seq;
-	if (seq[fence->ring] == RADEON_FENCE_SIGNALED_SEQ)
+	if (test_bit(FENCE_FLAG_SIGNALED_BIT, &fence->base.flags))
 		return 0;
 
+	seq[fence->ring] = fence->seq;
 	r = radeon_fence_wait_seq_timeout(fence->rdev, seq, intr, MAX_SCHEDULE_TIMEOUT);
 	if (r < 0) {
 		return r;
 	}
 
-	fence->seq = RADEON_FENCE_SIGNALED_SEQ;
+	r = fence_signal(&fence->base);
+	if (!r)
+		FENCE_TRACE(&fence->base, "signaled from fence_wait\n");
 	return 0;
 }
 
@@ -446,12 +568,13 @@ int radeon_fence_wait_any(struct radeon_device *rdev,
 			continue;
 		}
 
+		if (test_bit(FENCE_FLAG_SIGNALED_BIT, &fences[i]->base.flags)) {
+			/* already signaled */
+			return 0;
+		}
+
 		seq[i] = fences[i]->seq;
 		++num_rings;
-
-		/* test if something was allready signaled */
-		if (seq[i] == RADEON_FENCE_SIGNALED_SEQ)
-			return 0;
 	}
 
 	/* nothing to wait for ? */
@@ -532,7 +655,7 @@ int radeon_fence_wait_empty(struct radeon_device *rdev, int ring)
  */
 struct radeon_fence *radeon_fence_ref(struct radeon_fence *fence)
 {
-	kref_get(&fence->kref);
+	fence_get(&fence->base);
 	return fence;
 }
 
@@ -548,9 +671,8 @@ void radeon_fence_unref(struct radeon_fence **fence)
 	struct radeon_fence *tmp = *fence;
 
 	*fence = NULL;
-	if (tmp) {
-		kref_put(&tmp->kref, radeon_fence_destroy);
-	}
+	if (tmp)
+		fence_put(&tmp->base);
 }
 
 /**
@@ -855,6 +977,7 @@ static int radeon_debugfs_gpu_reset(struct seq_file *m, void *data)
 	down_read(&rdev->exclusive_lock);
 	seq_printf(m, "%d\n", rdev->needs_reset);
 	rdev->needs_reset = true;
+	wake_up_all(&rdev->fence_queue);
 	up_read(&rdev->exclusive_lock);
 
 	return 0;
@@ -874,3 +997,67 @@ int radeon_debugfs_fence_init(struct radeon_device *rdev)
 	return 0;
 #endif
 }
+
+static const char *radeon_fence_get_driver_name(struct fence *fence)
+{
+	return "radeon";
+}
+
+static const char *radeon_fence_get_timeline_name(struct fence *f)
+{
+	struct radeon_fence *fence = to_radeon_fence(f);
+	switch (fence->ring) {
+	case RADEON_RING_TYPE_GFX_INDEX: return "radeon.gfx";
+	case CAYMAN_RING_TYPE_CP1_INDEX: return "radeon.cp1";
+	case CAYMAN_RING_TYPE_CP2_INDEX: return "radeon.cp2";
+	case R600_RING_TYPE_DMA_INDEX: return "radeon.dma";
+	case CAYMAN_RING_TYPE_DMA1_INDEX: return "radeon.dma1";
+	case R600_RING_TYPE_UVD_INDEX: return "radeon.uvd";
+	case TN_RING_TYPE_VCE1_INDEX: return "radeon.vce1";
+	case TN_RING_TYPE_VCE2_INDEX: return "radeon.vce2";
+	default: WARN_ON_ONCE(1); return "radeon.unk";
+	}
+}
+
+static signed long __radeon_fence_default_wait(struct fence *f, bool intr,
+					       signed long timeout)
+{
+	struct radeon_fence *fence = to_radeon_fence(f);
+	struct radeon_device *rdev = fence->rdev;
+	bool signaled;
+
+	fence_enable_sw_signaling(&fence->base);
+
+	/*
+	 * This function has to return -EDEADLK, but cannot hold
+	 * exclusive_lock during the wait because some callers
+	 * may already hold it. This means checking needs_reset without
+	 * lock, and not fiddling with any gpu internals.
+	 *
+	 * The callback installed with fence_enable_sw_signaling will
+	 * run before our wait_event_*timeout call, so we will see
+	 * both the signaled fence and the changes to needs_reset.
+	 */
+
+	if (intr)
+		timeout = wait_event_interruptible_timeout(rdev->fence_queue,
+							   ((signaled = (test_bit(FENCE_FLAG_SIGNALED_BIT, &fence->base.flags))) || rdev->needs_reset),
+							   timeout);
+	else
+		timeout = wait_event_timeout(rdev->fence_queue,
+					     ((signaled = (test_bit(FENCE_FLAG_SIGNALED_BIT, &fence->base.flags))) || rdev->needs_reset),
+					     timeout);
+
+	if (timeout > 0 && !signaled)
+		return -EDEADLK;
+	return timeout;
+}
+
+static const struct fence_ops radeon_fence_ops = {
+	.get_driver_name = radeon_fence_get_driver_name,
+	.get_timeline_name = radeon_fence_get_timeline_name,
+	.enable_signaling = radeon_fence_enable_signaling,
+	.signaled = __radeon_fence_signaled,
+	.wait = __radeon_fence_default_wait,
+	.release = NULL,
+};
diff --git a/drivers/gpu/drm/radeon/radeon_irq_kms.c b/drivers/gpu/drm/radeon/radeon_irq_kms.c
index 16807afab362..41310cec61a6 100644
--- a/drivers/gpu/drm/radeon/radeon_irq_kms.c
+++ b/drivers/gpu/drm/radeon/radeon_irq_kms.c
@@ -254,6 +254,25 @@ static bool radeon_msi_ok(struct radeon_device *rdev)
 	return true;
 }
 
+static void radeon_delayed_irq_set(struct work_struct *work)
+{
+	struct radeon_device *rdev;
+	unsigned long irqflags;
+
+	rdev = container_of(work, struct radeon_device, delayed_irq_work);
+
+	down_read(&rdev->exclusive_lock);
+
+	spin_lock_irqsave(&rdev->irq.lock, irqflags);
+	radeon_irq_set(rdev);
+	spin_unlock_irqrestore(&rdev->irq.lock, irqflags);
+
+	/* force all waiters to recheck */
+	wake_up_all(&rdev->fence_queue);
+
+	up_read(&rdev->exclusive_lock);
+}
+
 /**
  * radeon_irq_kms_init - init driver interrupt info
  *
@@ -285,6 +304,7 @@ int radeon_irq_kms_init(struct radeon_device *rdev)
 	INIT_WORK(&rdev->hotplug_work, radeon_hotplug_work_func);
 	INIT_WORK(&rdev->audio_work, r600_audio_update_hdmi);
 	INIT_WORK(&rdev->reset_work, radeon_irq_reset_work_func);
+	INIT_WORK(&rdev->delayed_irq_work, radeon_delayed_irq_set);
 
 	rdev->irq.installed = true;
 	r = drm_irq_install(rdev->ddev, rdev->ddev->pdev->irq);
@@ -313,6 +333,7 @@ void radeon_irq_kms_fini(struct radeon_device *rdev)
 		rdev->irq.installed = false;
 		if (rdev->msi_enabled)
 			pci_disable_msi(rdev->pdev);
+		cancel_work_sync(&rdev->delayed_irq_work);
 		flush_work(&rdev->hotplug_work);
 	}
 }
@@ -342,6 +363,28 @@ void radeon_irq_kms_sw_irq_get(struct radeon_device *rdev, int ring)
 }
 
 /**
+ * radeon_irq_kms_sw_irq_get_delayed - enable software interrupt
+ *
+ * @rdev: radeon device pointer
+ * @ring: ring whose interrupt you want to enable
+ *
+ * Enables the software interrupt for a specific ring (all asics).
+ * The software interrupt is generally used to signal a fence on
+ * a particular ring.
+ *
+ * This function can be used without exclusive_lock grabbed,
+ * and will schedule later work for irqs to be set when needed.
+ */
+void radeon_irq_kms_sw_irq_get_delayed(struct radeon_device *rdev, int ring)
+{
+	if (!rdev->ddev->irq_enabled)
+		return;
+
+	if (atomic_inc_return(&rdev->irq.ring_int[ring]) == 1)
+		schedule_work(&rdev->delayed_irq_work);
+}
+
+/**
  * radeon_irq_kms_sw_irq_put - disable software interrupt
  *
  * @rdev: radeon device pointer

[PATCH 12/19] drm/qxl: rework to new fence interface

[Maarten Lankhorst]

Final driver! \o/

This is not a proper dma_fence because the hardware may never signal
anything, so don't use dma-buf with qxl, ever.

Signed-off-by: Maarten Lankhorst <maarten.lankhorst@canonical.com>
---
 drivers/gpu/drm/qxl/Makefile      |    2 
 drivers/gpu/drm/qxl/qxl_cmd.c     |    5 -
 drivers/gpu/drm/qxl/qxl_debugfs.c |   12 ++-
 drivers/gpu/drm/qxl/qxl_drv.h     |   22 ++---
 drivers/gpu/drm/qxl/qxl_fence.c   |   87 -------------------
 drivers/gpu/drm/qxl/qxl_kms.c     |    2 
 drivers/gpu/drm/qxl/qxl_object.c  |    2 
 drivers/gpu/drm/qxl/qxl_release.c |  166 ++++++++++++++++++++++++++++++++-----
 drivers/gpu/drm/qxl/qxl_ttm.c     |   97 ++++++++++++----------
 9 files changed, 220 insertions(+), 175 deletions(-)
 delete mode 100644 drivers/gpu/drm/qxl/qxl_fence.c

diff --git a/drivers/gpu/drm/qxl/Makefile b/drivers/gpu/drm/qxl/Makefile
index ea046ba691d2..ac0d74852e11 100644
--- a/drivers/gpu/drm/qxl/Makefile
+++ b/drivers/gpu/drm/qxl/Makefile
@@ -4,6 +4,6 @@
 
 ccflags-y := -Iinclude/drm
 
-qxl-y := qxl_drv.o qxl_kms.o qxl_display.o qxl_ttm.o qxl_fb.o qxl_object.o qxl_gem.o qxl_cmd.o qxl_image.o qxl_draw.o qxl_debugfs.o qxl_irq.o qxl_dumb.o qxl_ioctl.o qxl_fence.o qxl_release.o
+qxl-y := qxl_drv.o qxl_kms.o qxl_display.o qxl_ttm.o qxl_fb.o qxl_object.o qxl_gem.o qxl_cmd.o qxl_image.o qxl_draw.o qxl_debugfs.o qxl_irq.o qxl_dumb.o qxl_ioctl.o qxl_release.o
 
 obj-$(CONFIG_DRM_QXL)+= qxl.o
diff --git a/drivers/gpu/drm/qxl/qxl_cmd.c b/drivers/gpu/drm/qxl/qxl_cmd.c
index 45fad7b45486..97823644d347 100644
--- a/drivers/gpu/drm/qxl/qxl_cmd.c
+++ b/drivers/gpu/drm/qxl/qxl_cmd.c
@@ -620,11 +620,6 @@ static int qxl_reap_surf(struct qxl_device *qdev, struct qxl_bo *surf, bool stal
 	if (ret == -EBUSY)
 		return -EBUSY;
 
-	if (surf->fence.num_active_releases > 0 && stall == false) {
-		qxl_bo_unreserve(surf);
-		return -EBUSY;
-	}
-
 	if (stall)
 		mutex_unlock(&qdev->surf_evict_mutex);
 
diff --git a/drivers/gpu/drm/qxl/qxl_debugfs.c b/drivers/gpu/drm/qxl/qxl_debugfs.c
index c3c2bbdc6674..0d144e0646d6 100644
--- a/drivers/gpu/drm/qxl/qxl_debugfs.c
+++ b/drivers/gpu/drm/qxl/qxl_debugfs.c
@@ -57,11 +57,21 @@ qxl_debugfs_buffers_info(struct seq_file *m, void *data)
 	struct qxl_device *qdev = node->minor->dev->dev_private;
 	struct qxl_bo *bo;
 
+	spin_lock(&qdev->release_lock);
 	list_for_each_entry(bo, &qdev->gem.objects, list) {
+		struct reservation_object_list *fobj;
+		int rel;
+
+		rcu_read_lock();
+		fobj = rcu_dereference(bo->tbo.resv->fence);
+		rel = fobj ? fobj->shared_count : 0;
+		rcu_read_unlock();
+
 		seq_printf(m, "size %ld, pc %d, sync obj %p, num releases %d\n",
 			   (unsigned long)bo->gem_base.size, bo->pin_count,
-			   bo->tbo.sync_obj, bo->fence.num_active_releases);
+			   bo->tbo.sync_obj, rel);
 	}
+	spin_unlock(&qdev->release_lock);
 	return 0;
 }
 
diff --git a/drivers/gpu/drm/qxl/qxl_drv.h b/drivers/gpu/drm/qxl/qxl_drv.h
index 36ed40ba773f..d547cbdebeb4 100644
--- a/drivers/gpu/drm/qxl/qxl_drv.h
+++ b/drivers/gpu/drm/qxl/qxl_drv.h
@@ -31,6 +31,7 @@
  * Definitions taken from spice-protocol, plus kernel driver specific bits.
  */
 
+#include <linux/fence.h>
 #include <linux/workqueue.h>
 #include <linux/firmware.h>
 #include <linux/platform_device.h>
@@ -95,13 +96,6 @@ enum {
 	QXL_INTERRUPT_IO_CMD |\
 	QXL_INTERRUPT_CLIENT_MONITORS_CONFIG)
 
-struct qxl_fence {
-	struct qxl_device *qdev;
-	uint32_t num_active_releases;
-	uint32_t *release_ids;
-	struct radix_tree_root tree;
-};
-
 struct qxl_bo {
 	/* Protected by gem.mutex */
 	struct list_head		list;
@@ -113,13 +107,13 @@ struct qxl_bo {
 	unsigned			pin_count;
 	void				*kptr;
 	int                             type;
+
 	/* Constant after initialization */
 	struct drm_gem_object		gem_base;
 	bool is_primary; /* is this now a primary surface */
 	bool hw_surf_alloc;
 	struct qxl_surface surf;
 	uint32_t surface_id;
-	struct qxl_fence fence; /* per bo fence  - list of releases */
 	struct qxl_release *surf_create;
 };
 #define gem_to_qxl_bo(gobj) container_of((gobj), struct qxl_bo, gem_base)
@@ -191,6 +185,8 @@ enum {
  * spice-protocol/qxl_dev.h */
 #define QXL_MAX_RES 96
 struct qxl_release {
+	struct fence base;
+
 	int id;
 	int type;
 	uint32_t release_offset;
@@ -284,7 +280,11 @@ struct qxl_device {
 	uint8_t		slot_gen_bits;
 	uint64_t	va_slot_mask;
 
+	/* XXX: when rcu becomes available, release_lock can be killed */
+	spinlock_t	release_lock;
+	spinlock_t	fence_lock;
 	struct idr	release_idr;
+	uint32_t	release_seqno;
 	spinlock_t release_idr_lock;
 	struct mutex	async_io_mutex;
 	unsigned int last_sent_io_cmd;
@@ -561,10 +561,4 @@ qxl_surface_lookup(struct drm_device *dev, int surface_id);
 void qxl_surface_evict(struct qxl_device *qdev, struct qxl_bo *surf, bool freeing);
 int qxl_update_surface(struct qxl_device *qdev, struct qxl_bo *surf);
 
-/* qxl_fence.c */
-void qxl_fence_add_release_locked(struct qxl_fence *qfence, uint32_t rel_id);
-int qxl_fence_remove_release(struct qxl_fence *qfence, uint32_t rel_id);
-int qxl_fence_init(struct qxl_device *qdev, struct qxl_fence *qfence);
-void qxl_fence_fini(struct qxl_fence *qfence);
-
 #endif
diff --git a/drivers/gpu/drm/qxl/qxl_fence.c b/drivers/gpu/drm/qxl/qxl_fence.c
deleted file mode 100644
index c7248418117d..000000000000
--- a/drivers/gpu/drm/qxl/qxl_fence.c
+++ /dev/null
@@ -1,87 +0,0 @@
-/*
- * Copyright 2013 Red Hat Inc.
- *
- * Permission is hereby granted, free of charge, to any person obtaining a
- * copy of this software and associated documentation files (the "Software"),
- * to deal in the Software without restriction, including without limitation
- * the rights to use, copy, modify, merge, publish, distribute, sublicense,
- * and/or sell copies of the Software, and to permit persons to whom the
- * Software is furnished to do so, subject to the following conditions:
- *
- * The above copyright notice and this permission notice shall be included in
- * all copies or substantial portions of the Software.
- *
- * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
- * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
- * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
- * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
- * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
- * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
- * OTHER DEALINGS IN THE SOFTWARE.
- *
- * Authors: Dave Airlie
- *          Alon Levy
- */
-
-
-#include "qxl_drv.h"
-
-/* QXL fencing-
-
-   When we submit operations to the GPU we pass a release reference to the GPU
-   with them, the release reference is then added to the release ring when
-   the GPU is finished with that particular operation and has removed it from
-   its tree.
-
-   So we have can have multiple outstanding non linear fences per object.
-
-   From a TTM POV we only care if the object has any outstanding releases on
-   it.
-
-   we wait until all outstanding releases are processeed.
-
-   sync object is just a list of release ids that represent that fence on
-   that buffer.
-
-   we just add new releases onto the sync object attached to the object.
-
-   This currently uses a radix tree to store the list of release ids.
-
-   For some reason every so often qxl hw fails to release, things go wrong.
-*/
-/* must be called with the fence lock held */
-void qxl_fence_add_release_locked(struct qxl_fence *qfence, uint32_t rel_id)
-{
-	radix_tree_insert(&qfence->tree, rel_id, qfence);
-	qfence->num_active_releases++;
-}
-
-int qxl_fence_remove_release(struct qxl_fence *qfence, uint32_t rel_id)
-{
-	void *ret;
-	int retval = 0;
-
-	ret = radix_tree_delete(&qfence->tree, rel_id);
-	if (ret == qfence)
-		qfence->num_active_releases--;
-	else {
-		DRM_DEBUG("didn't find fence in radix tree for %d\n", rel_id);
-		retval = -ENOENT;
-	}
-	return retval;
-}
-
-
-int qxl_fence_init(struct qxl_device *qdev, struct qxl_fence *qfence)
-{
-	qfence->qdev = qdev;
-	qfence->num_active_releases = 0;
-	INIT_RADIX_TREE(&qfence->tree, GFP_ATOMIC);
-	return 0;
-}
-
-void qxl_fence_fini(struct qxl_fence *qfence)
-{
-	kfree(qfence->release_ids);
-	qfence->num_active_releases = 0;
-}
diff --git a/drivers/gpu/drm/qxl/qxl_kms.c b/drivers/gpu/drm/qxl/qxl_kms.c
index fd88eb4a3f79..a9e7c30e92c5 100644
--- a/drivers/gpu/drm/qxl/qxl_kms.c
+++ b/drivers/gpu/drm/qxl/qxl_kms.c
@@ -223,6 +223,8 @@ static int qxl_device_init(struct qxl_device *qdev,
 
 	idr_init(&qdev->release_idr);
 	spin_lock_init(&qdev->release_idr_lock);
+	spin_lock_init(&qdev->release_lock);
+	spin_lock_init(&qdev->fence_lock);
 
 	idr_init(&qdev->surf_id_idr);
 	spin_lock_init(&qdev->surf_id_idr_lock);
diff --git a/drivers/gpu/drm/qxl/qxl_object.c b/drivers/gpu/drm/qxl/qxl_object.c
index b95f144f0b49..9981962451d7 100644
--- a/drivers/gpu/drm/qxl/qxl_object.c
+++ b/drivers/gpu/drm/qxl/qxl_object.c
@@ -36,7 +36,6 @@ static void qxl_ttm_bo_destroy(struct ttm_buffer_object *tbo)
 	qdev = (struct qxl_device *)bo->gem_base.dev->dev_private;
 
 	qxl_surface_evict(qdev, bo, false);
-	qxl_fence_fini(&bo->fence);
 	mutex_lock(&qdev->gem.mutex);
 	list_del_init(&bo->list);
 	mutex_unlock(&qdev->gem.mutex);
@@ -99,7 +98,6 @@ int qxl_bo_create(struct qxl_device *qdev,
 	bo->type = domain;
 	bo->pin_count = pinned ? 1 : 0;
 	bo->surface_id = 0;
-	qxl_fence_init(qdev, &bo->fence);
 	INIT_LIST_HEAD(&bo->list);
 
 	if (surf)
diff --git a/drivers/gpu/drm/qxl/qxl_release.c b/drivers/gpu/drm/qxl/qxl_release.c
index 4045ba873ab8..9731d2540a40 100644
--- a/drivers/gpu/drm/qxl/qxl_release.c
+++ b/drivers/gpu/drm/qxl/qxl_release.c
@@ -21,6 +21,7 @@
  */
 #include "qxl_drv.h"
 #include "qxl_object.h"
+#include <trace/events/fence.h>
 
 /*
  * drawable cmd cache - allocate a bunch of VRAM pages, suballocate
@@ -39,6 +40,88 @@
 static const int release_size_per_bo[] = { RELEASE_SIZE, SURFACE_RELEASE_SIZE, RELEASE_SIZE };
 static const int releases_per_bo[] = { RELEASES_PER_BO, SURFACE_RELEASES_PER_BO, RELEASES_PER_BO };
 
+static const char *qxl_get_driver_name(struct fence *fence)
+{
+	return "qxl";
+}
+
+static const char *qxl_get_timeline_name(struct fence *fence)
+{
+	return "release";
+}
+
+static bool qxl_nop_signaling(struct fence *fence)
+{
+	/* fences are always automatically signaled, so just pretend we did this.. */
+	return true;
+}
+
+static long qxl_fence_wait(struct fence *fence, bool intr, signed long timeout)
+{
+	struct qxl_device *qdev;
+	struct qxl_release *release;
+	int count = 0, sc = 0;
+	bool have_drawable_releases;
+	unsigned long cur, end = jiffies + timeout;
+
+	qdev = container_of(fence->lock, struct qxl_device, release_lock);
+	release = container_of(fence, struct qxl_release, base);
+	have_drawable_releases = release->type == QXL_RELEASE_DRAWABLE;
+
+retry:
+	sc++;
+
+	if (fence_is_signaled_locked(fence))
+		goto signaled;
+
+	qxl_io_notify_oom(qdev);
+
+	for (count = 0; count < 11; count++) {
+		if (!qxl_queue_garbage_collect(qdev, true))
+			break;
+
+		if (fence_is_signaled_locked(fence))
+			goto signaled;
+	}
+
+	if (fence_is_signaled_locked(fence))
+		goto signaled;
+
+	if (have_drawable_releases || sc < 4) {
+		if (sc > 2)
+			/* back off */
+			usleep_range(500, 1000);
+
+		if (time_after(jiffies, end))
+			return 0;
+
+		if (have_drawable_releases && sc > 300) {
+			FENCE_WARN(fence, "failed to wait on release %d "
+					  "after spincount %d\n",
+					  fence->context & ~0xf0000000, sc);
+			goto signaled;
+		}
+		goto retry;
+	}
+	/*
+	 * yeah, original sync_obj_wait gave up after 3 spins when
+	 * have_drawable_releases is not set.
+	 */
+
+signaled:
+	cur = jiffies;
+	if (time_after(cur, end))
+		return 0;
+	return end - cur;
+}
+
+static const struct fence_ops qxl_fence_ops = {
+	.get_driver_name = qxl_get_driver_name,
+	.get_timeline_name = qxl_get_timeline_name,
+	.enable_signaling = qxl_nop_signaling,
+	.wait = qxl_fence_wait,
+};
+
 static uint64_t
 qxl_release_alloc(struct qxl_device *qdev, int type,
 		  struct qxl_release **ret)
@@ -46,13 +129,13 @@ qxl_release_alloc(struct qxl_device *qdev, int type,
 	struct qxl_release *release;
 	int handle;
 	size_t size = sizeof(*release);
-	int idr_ret;
 
 	release = kmalloc(size, GFP_KERNEL);
 	if (!release) {
 		DRM_ERROR("Out of memory\n");
 		return 0;
 	}
+	release->base.ops = NULL;
 	release->type = type;
 	release->release_offset = 0;
 	release->surface_release_id = 0;
@@ -60,44 +143,59 @@ qxl_release_alloc(struct qxl_device *qdev, int type,
 
 	idr_preload(GFP_KERNEL);
 	spin_lock(&qdev->release_idr_lock);
-	idr_ret = idr_alloc(&qdev->release_idr, release, 1, 0, GFP_NOWAIT);
+	handle = idr_alloc(&qdev->release_idr, release, 1, 0, GFP_NOWAIT);
+	release->base.seqno = ++qdev->release_seqno;
 	spin_unlock(&qdev->release_idr_lock);
 	idr_preload_end();
-	handle = idr_ret;
-	if (idr_ret < 0)
-		goto release_fail;
+	if (handle < 0) {
+		kfree(release);
+		*ret = NULL;
+		return handle;
+	}
 	*ret = release;
 	QXL_INFO(qdev, "allocated release %lld\n", handle);
 	release->id = handle;
-release_fail:
-
 	return handle;
 }
 
+static void
+qxl_release_free_list(struct qxl_release *release)
+{
+	while (!list_empty(&release->bos)) {
+		struct ttm_validate_buffer *entry;
+
+		entry = container_of(release->bos.next,
+				     struct ttm_validate_buffer, head);
+
+		list_del(&entry->head);
+		kfree(entry);
+	}
+}
+
 void
 qxl_release_free(struct qxl_device *qdev,
 		 struct qxl_release *release)
 {
-	struct qxl_bo_list *entry, *tmp;
 	QXL_INFO(qdev, "release %d, type %d\n", release->id,
 		 release->type);
 
 	if (release->surface_release_id)
 		qxl_surface_id_dealloc(qdev, release->surface_release_id);
 
-	list_for_each_entry_safe(entry, tmp, &release->bos, tv.head) {
-		struct qxl_bo *bo = to_qxl_bo(entry->tv.bo);
-		QXL_INFO(qdev, "release %llx\n",
-			drm_vma_node_offset_addr(&entry->tv.bo->vma_node)
-						- DRM_FILE_OFFSET);
-		qxl_fence_remove_release(&bo->fence, release->id);
-		qxl_bo_unref(&bo);
-		kfree(entry);
-	}
 	spin_lock(&qdev->release_idr_lock);
 	idr_remove(&qdev->release_idr, release->id);
 	spin_unlock(&qdev->release_idr_lock);
-	kfree(release);
+
+	if (release->base.ops) {
+		WARN_ON(list_empty(&release->bos));
+		qxl_release_free_list(release);
+
+		fence_signal(&release->base);
+		fence_put(&release->base);
+	} else {
+		qxl_release_free_list(release);
+		kfree(release);
+	}
 }
 
 static int qxl_release_bo_alloc(struct qxl_device *qdev,
@@ -142,6 +240,10 @@ static int qxl_release_validate_bo(struct qxl_bo *bo)
 			return ret;
 	}
 
+	ret = reservation_object_reserve_shared(bo->tbo.resv);
+	if (ret)
+		return ret;
+
 	/* allocate a surface for reserved + validated buffers */
 	ret = qxl_bo_check_id(bo->gem_base.dev->dev_private, bo);
 	if (ret)
@@ -199,6 +301,8 @@ int qxl_alloc_surface_release_reserved(struct qxl_device *qdev,
 
 		/* stash the release after the create command */
 		idr_ret = qxl_release_alloc(qdev, QXL_RELEASE_SURFACE_CMD, release);
+		if (idr_ret < 0)
+			return idr_ret;
 		bo = qxl_bo_ref(to_qxl_bo(entry->tv.bo));
 
 		(*release)->release_offset = create_rel->release_offset + 64;
@@ -239,6 +343,11 @@ int qxl_alloc_release_reserved(struct qxl_device *qdev, unsigned long size,
 	}
 
 	idr_ret = qxl_release_alloc(qdev, type, release);
+	if (idr_ret < 0) {
+		if (rbo)
+			*rbo = NULL;
+		return idr_ret;
+	}
 
 	mutex_lock(&qdev->release_mutex);
 	if (qdev->current_release_bo_offset[cur_idx] + 1 >= releases_per_bo[cur_idx]) {
@@ -319,12 +428,13 @@ void qxl_release_unmap(struct qxl_device *qdev,
 
 void qxl_release_fence_buffer_objects(struct qxl_release *release)
 {
-	struct ttm_validate_buffer *entry;
 	struct ttm_buffer_object *bo;
 	struct ttm_bo_global *glob;
 	struct ttm_bo_device *bdev;
 	struct ttm_bo_driver *driver;
 	struct qxl_bo *qbo;
+	struct ttm_validate_buffer *entry;
+	struct qxl_device *qdev;
 
 	/* if only one object on the release its the release itself
 	   since these objects are pinned no need to reserve */
@@ -333,23 +443,35 @@ void qxl_release_fence_buffer_objects(struct qxl_release *release)
 
 	bo = list_first_entry(&release->bos, struct ttm_validate_buffer, head)->bo;
 	bdev = bo->bdev;
+	qdev = container_of(bdev, struct qxl_device, mman.bdev);
+
+	/*
+	 * Since we never really allocated a context and we don't want to conflict,
+	 * set the highest bits. This will break if we really allow exporting of dma-bufs.
+	 */
+	fence_init(&release->base, &qxl_fence_ops, &qdev->release_lock,
+		   release->id | 0xf0000000, release->base.seqno);
+	trace_fence_emit(&release->base);
+
 	driver = bdev->driver;
 	glob = bo->glob;
 
 	spin_lock(&glob->lru_lock);
+	/* acquire release_lock to protect bo->resv->fence and its contents */
+	spin_lock(&qdev->release_lock);
 
 	list_for_each_entry(entry, &release->bos, head) {
 		bo = entry->bo;
 		qbo = to_qxl_bo(bo);
 
 		if (!entry->bo->sync_obj)
-			entry->bo->sync_obj = &qbo->fence;
-
-		qxl_fence_add_release_locked(&qbo->fence, release->id);
+			entry->bo->sync_obj = qbo;
 
+		reservation_object_add_shared_fence(bo->resv, &release->base);
 		ttm_bo_add_to_lru(bo);
 		__ttm_bo_unreserve(bo);
 	}
+	spin_unlock(&qdev->release_lock);
 	spin_unlock(&glob->lru_lock);
 	ww_acquire_fini(&release->ticket);
 }
diff --git a/drivers/gpu/drm/qxl/qxl_ttm.c b/drivers/gpu/drm/qxl/qxl_ttm.c
index 71a1baeac14e..6230251fa5b0 100644
--- a/drivers/gpu/drm/qxl/qxl_ttm.c
+++ b/drivers/gpu/drm/qxl/qxl_ttm.c
@@ -355,67 +355,67 @@ static int qxl_bo_move(struct ttm_buffer_object *bo,
 	return ttm_bo_move_memcpy(bo, evict, no_wait_gpu, new_mem);
 }
 
+static bool qxl_sync_obj_signaled(void *sync_obj);
 
 static int qxl_sync_obj_wait(void *sync_obj,
 			     bool lazy, bool interruptible)
 {
-	struct qxl_fence *qfence = (struct qxl_fence *)sync_obj;
-	int count = 0, sc = 0;
-	struct qxl_bo *bo = container_of(qfence, struct qxl_bo, fence);
-
-	if (qfence->num_active_releases == 0)
-		return 0;
+	struct qxl_bo *bo = (struct qxl_bo *)sync_obj;
+	struct qxl_device *qdev = bo->gem_base.dev->dev_private;
+	struct reservation_object_list *fobj;
+	int count = 0, sc = 0, num_release = 0;
+	bool have_drawable_releases;
 
 retry:
 	if (sc == 0) {
 		if (bo->type == QXL_GEM_DOMAIN_SURFACE)
-			qxl_update_surface(qfence->qdev, bo);
+			qxl_update_surface(qdev, bo);
 	} else if (sc >= 1) {
-		qxl_io_notify_oom(qfence->qdev);
+		qxl_io_notify_oom(qdev);
 	}
 
 	sc++;
 
 	for (count = 0; count < 10; count++) {
-		bool ret;
-		ret = qxl_queue_garbage_collect(qfence->qdev, true);
-		if (ret == false)
-			break;
-
-		if (qfence->num_active_releases == 0)
+		if (qxl_sync_obj_signaled(sync_obj))
 			return 0;
+
+		if (!qxl_queue_garbage_collect(qdev, true))
+			break;
 	}
 
-	if (qfence->num_active_releases) {
-		bool have_drawable_releases = false;
-		void **slot;
-		struct radix_tree_iter iter;
-		int release_id;
+	have_drawable_releases = false;
+	num_release = 0;
 
-		radix_tree_for_each_slot(slot, &qfence->tree, &iter, 0) {
-			struct qxl_release *release;
+	spin_lock(&qdev->release_lock);
+	fobj = bo->tbo.resv->fence;
+	for (count = 0; fobj && count < fobj->shared_count; count++) {
+		struct qxl_release *release;
 
-			release_id = iter.index;
-			release = qxl_release_from_id_locked(qfence->qdev, release_id);
-			if (release == NULL)
-				continue;
+		release = container_of(fobj->shared[count],
+				       struct qxl_release, base);
 
-			if (release->type == QXL_RELEASE_DRAWABLE)
-				have_drawable_releases = true;
-		}
+		if (fence_is_signaled(&release->base))
+			continue;
+
+		num_release++;
+
+		if (release->type == QXL_RELEASE_DRAWABLE)
+			have_drawable_releases = true;
+	}
+	spin_unlock(&qdev->release_lock);
+
+	qxl_queue_garbage_collect(qdev, true);
 
-		qxl_queue_garbage_collect(qfence->qdev, true);
-
-		if (have_drawable_releases || sc < 4) {
-			if (sc > 2)
-				/* back off */
-				usleep_range(500, 1000);
-			if (have_drawable_releases && sc > 300) {
-				WARN(1, "sync obj %d still has outstanding releases %d %d %d %ld %d\n", sc, bo->surface_id, bo->is_primary, bo->pin_count, (unsigned long)bo->gem_base.size, qfence->num_active_releases);
-				return -EBUSY;
-			}
-			goto retry;
+	if (have_drawable_releases || sc < 4) {
+		if (sc > 2)
+			/* back off */
+			usleep_range(500, 1000);
+		if (have_drawable_releases && sc > 300) {
+			WARN(1, "sync obj %d still has outstanding releases %d %d %d %ld %d\n", sc, bo->surface_id, bo->is_primary, bo->pin_count, (unsigned long)bo->gem_base.size, num_release);
+			return -EBUSY;
 		}
+		goto retry;
 	}
 	return 0;
 }
@@ -437,8 +437,21 @@ static void *qxl_sync_obj_ref(void *sync_obj)
 
 static bool qxl_sync_obj_signaled(void *sync_obj)
 {
-	struct qxl_fence *qfence = (struct qxl_fence *)sync_obj;
-	return (qfence->num_active_releases == 0);
+	struct qxl_bo *qbo = (struct qxl_bo *)sync_obj;
+	struct qxl_device *qdev = qbo->gem_base.dev->dev_private;
+	struct reservation_object_list *fobj;
+	bool ret = true;
+	unsigned i;
+
+	spin_lock(&qdev->release_lock);
+	fobj = qbo->tbo.resv->fence;
+	for (i = 0; fobj && i < fobj->shared_count; ++i) {
+		ret = fence_is_signaled(fobj->shared[i]);
+		if (!ret)
+			break;
+	}
+	spin_unlock(&qdev->release_lock);
+	return ret;
 }
 
 static void qxl_bo_move_notify(struct ttm_buffer_object *bo,
@@ -475,8 +488,6 @@ static struct ttm_bo_driver qxl_bo_driver = {
 	.move_notify = &qxl_bo_move_notify,
 };
 
-
-
 int qxl_ttm_init(struct qxl_device *qdev)
 {
 	int r;

[PATCH 13/19] drm/vmwgfx: get rid of different types of fence_flags entirely

[Maarten Lankhorst]

Only one type was ever used. This is needed to simplify the fence
support in the next commit.

Signed-off-by: Maarten Lankhorst <maarten.lankhorst@canonical.com>
---
 drivers/gpu/drm/vmwgfx/vmwgfx_buffer.c  |    5 +--
 drivers/gpu/drm/vmwgfx/vmwgfx_drv.h     |    1 -
 drivers/gpu/drm/vmwgfx/vmwgfx_execbuf.c |   14 ++-------
 drivers/gpu/drm/vmwgfx/vmwgfx_fence.c   |   50 ++++++++++++-------------------
 drivers/gpu/drm/vmwgfx/vmwgfx_fence.h   |    8 +----
 5 files changed, 26 insertions(+), 52 deletions(-)

diff --git a/drivers/gpu/drm/vmwgfx/vmwgfx_buffer.c b/drivers/gpu/drm/vmwgfx/vmwgfx_buffer.c
index 4a36bb1dc525..f15718cc631d 100644
--- a/drivers/gpu/drm/vmwgfx/vmwgfx_buffer.c
+++ b/drivers/gpu/drm/vmwgfx/vmwgfx_buffer.c
@@ -792,15 +792,12 @@ static int vmw_sync_obj_flush(void *sync_obj)
 
 static bool vmw_sync_obj_signaled(void *sync_obj)
 {
-	return	vmw_fence_obj_signaled((struct vmw_fence_obj *) sync_obj,
-				       DRM_VMW_FENCE_FLAG_EXEC);
-
+	return vmw_fence_obj_signaled((struct vmw_fence_obj *) sync_obj);
 }
 
 static int vmw_sync_obj_wait(void *sync_obj, bool lazy, bool interruptible)
 {
 	return vmw_fence_obj_wait((struct vmw_fence_obj *) sync_obj,
-				  DRM_VMW_FENCE_FLAG_EXEC,
 				  lazy, interruptible,
 				  VMW_FENCE_WAIT_TIMEOUT);
 }
diff --git a/drivers/gpu/drm/vmwgfx/vmwgfx_drv.h b/drivers/gpu/drm/vmwgfx/vmwgfx_drv.h
index c1811750cc8d..0b1e4b0c919e 100644
--- a/drivers/gpu/drm/vmwgfx/vmwgfx_drv.h
+++ b/drivers/gpu/drm/vmwgfx/vmwgfx_drv.h
@@ -342,7 +342,6 @@ struct vmw_sw_context{
 	uint32_t *cmd_bounce;
 	uint32_t cmd_bounce_size;
 	struct list_head resource_list;
-	uint32_t fence_flags;
 	struct ttm_buffer_object *cur_query_bo;
 	struct list_head res_relocations;
 	uint32_t *buf_start;
diff --git a/drivers/gpu/drm/vmwgfx/vmwgfx_execbuf.c b/drivers/gpu/drm/vmwgfx/vmwgfx_execbuf.c
index b19b2b980cb4..214fcfd13cb2 100644
--- a/drivers/gpu/drm/vmwgfx/vmwgfx_execbuf.c
+++ b/drivers/gpu/drm/vmwgfx/vmwgfx_execbuf.c
@@ -350,8 +350,6 @@ static int vmw_bo_to_validate_list(struct vmw_sw_context *sw_context,
 		vval_buf->validate_as_mob = validate_as_mob;
 	}
 
-	sw_context->fence_flags |= DRM_VMW_FENCE_FLAG_EXEC;
-
 	if (p_val_node)
 		*p_val_node = val_node;
 
@@ -2337,13 +2335,9 @@ int vmw_execbuf_fence_commands(struct drm_file *file_priv,
 
 	if (p_handle != NULL)
 		ret = vmw_user_fence_create(file_priv, dev_priv->fman,
-					    sequence,
-					    DRM_VMW_FENCE_FLAG_EXEC,
-					    p_fence, p_handle);
+					    sequence, p_fence, p_handle);
 	else
-		ret = vmw_fence_create(dev_priv->fman, sequence,
-				       DRM_VMW_FENCE_FLAG_EXEC,
-				       p_fence);
+		ret = vmw_fence_create(dev_priv->fman, sequence, p_fence);
 
 	if (unlikely(ret != 0 && !synced)) {
 		(void) vmw_fallback_wait(dev_priv, false, false,
@@ -2416,8 +2410,7 @@ vmw_execbuf_copy_fence_user(struct vmw_private *dev_priv,
 		ttm_ref_object_base_unref(vmw_fp->tfile,
 					  fence_handle, TTM_REF_USAGE);
 		DRM_ERROR("Fence copy error. Syncing.\n");
-		(void) vmw_fence_obj_wait(fence, fence->signal_mask,
-					  false, false,
+		(void) vmw_fence_obj_wait(fence, false, false,
 					  VMW_FENCE_WAIT_TIMEOUT);
 	}
 }
@@ -2469,7 +2462,6 @@ int vmw_execbuf_process(struct drm_file *file_priv,
 	sw_context->fp = vmw_fpriv(file_priv);
 	sw_context->cur_reloc = 0;
 	sw_context->cur_val_buf = 0;
-	sw_context->fence_flags = 0;
 	INIT_LIST_HEAD(&sw_context->resource_list);
 	sw_context->cur_query_bo = dev_priv->pinned_bo;
 	sw_context->last_query_ctx = NULL;
diff --git a/drivers/gpu/drm/vmwgfx/vmwgfx_fence.c b/drivers/gpu/drm/vmwgfx/vmwgfx_fence.c
index 436b013b4231..05b9eea8e875 100644
--- a/drivers/gpu/drm/vmwgfx/vmwgfx_fence.c
+++ b/drivers/gpu/drm/vmwgfx/vmwgfx_fence.c
@@ -207,9 +207,7 @@ void vmw_fence_manager_takedown(struct vmw_fence_manager *fman)
 }
 
 static int vmw_fence_obj_init(struct vmw_fence_manager *fman,
-			      struct vmw_fence_obj *fence,
-			      u32 seqno,
-			      uint32_t mask,
+			      struct vmw_fence_obj *fence, u32 seqno,
 			      void (*destroy) (struct vmw_fence_obj *fence))
 {
 	unsigned long irq_flags;
@@ -220,7 +218,6 @@ static int vmw_fence_obj_init(struct vmw_fence_manager *fman,
 	INIT_LIST_HEAD(&fence->seq_passed_actions);
 	fence->fman = fman;
 	fence->signaled = 0;
-	fence->signal_mask = mask;
 	kref_init(&fence->kref);
 	fence->destroy = destroy;
 	init_waitqueue_head(&fence->queue);
@@ -356,7 +353,7 @@ static bool vmw_fence_goal_check_locked(struct vmw_fence_obj *fence)
 	u32 goal_seqno;
 	__le32 __iomem *fifo_mem;
 
-	if (fence->signaled & DRM_VMW_FENCE_FLAG_EXEC)
+	if (fence->signaled)
 		return false;
 
 	fifo_mem = fence->fman->dev_priv->mmio_virt;
@@ -386,7 +383,7 @@ rerun:
 	list_for_each_entry_safe(fence, next_fence, &fman->fence_list, head) {
 		if (seqno - fence->seqno < VMW_FENCE_WRAP) {
 			list_del_init(&fence->head);
-			fence->signaled |= DRM_VMW_FENCE_FLAG_EXEC;
+			fence->signaled = 1;
 			INIT_LIST_HEAD(&action_list);
 			list_splice_init(&fence->seq_passed_actions,
 					 &action_list);
@@ -417,8 +414,7 @@ rerun:
 	}
 }
 
-bool vmw_fence_obj_signaled(struct vmw_fence_obj *fence,
-			    uint32_t flags)
+bool vmw_fence_obj_signaled(struct vmw_fence_obj *fence)
 {
 	struct vmw_fence_manager *fman = fence->fman;
 	unsigned long irq_flags;
@@ -428,28 +424,25 @@ bool vmw_fence_obj_signaled(struct vmw_fence_obj *fence,
 	signaled = fence->signaled;
 	spin_unlock_irqrestore(&fman->lock, irq_flags);
 
-	flags &= fence->signal_mask;
-	if ((signaled & flags) == flags)
+	if (signaled)
 		return 1;
 
-	if ((signaled & DRM_VMW_FENCE_FLAG_EXEC) == 0)
-		vmw_fences_update(fman);
+	vmw_fences_update(fman);
 
 	spin_lock_irqsave(&fman->lock, irq_flags);
 	signaled = fence->signaled;
 	spin_unlock_irqrestore(&fman->lock, irq_flags);
 
-	return ((signaled & flags) == flags);
+	return signaled;
 }
 
-int vmw_fence_obj_wait(struct vmw_fence_obj *fence,
-		       uint32_t flags, bool lazy,
+int vmw_fence_obj_wait(struct vmw_fence_obj *fence, bool lazy,
 		       bool interruptible, unsigned long timeout)
 {
 	struct vmw_private *dev_priv = fence->fman->dev_priv;
 	long ret;
 
-	if (likely(vmw_fence_obj_signaled(fence, flags)))
+	if (likely(vmw_fence_obj_signaled(fence)))
 		return 0;
 
 	vmw_fifo_ping_host(dev_priv, SVGA_SYNC_GENERIC);
@@ -458,12 +451,12 @@ int vmw_fence_obj_wait(struct vmw_fence_obj *fence,
 	if (interruptible)
 		ret = wait_event_interruptible_timeout
 			(fence->queue,
-			 vmw_fence_obj_signaled(fence, flags),
+			 vmw_fence_obj_signaled(fence),
 			 timeout);
 	else
 		ret = wait_event_timeout
 			(fence->queue,
-			 vmw_fence_obj_signaled(fence, flags),
+			 vmw_fence_obj_signaled(fence),
 			 timeout);
 
 	vmw_seqno_waiter_remove(dev_priv);
@@ -497,7 +490,6 @@ static void vmw_fence_destroy(struct vmw_fence_obj *fence)
 
 int vmw_fence_create(struct vmw_fence_manager *fman,
 		     uint32_t seqno,
-		     uint32_t mask,
 		     struct vmw_fence_obj **p_fence)
 {
 	struct ttm_mem_global *mem_glob = vmw_mem_glob(fman->dev_priv);
@@ -515,7 +507,7 @@ int vmw_fence_create(struct vmw_fence_manager *fman,
 		goto out_no_object;
 	}
 
-	ret = vmw_fence_obj_init(fman, fence, seqno, mask,
+	ret = vmw_fence_obj_init(fman, fence, seqno,
 				 vmw_fence_destroy);
 	if (unlikely(ret != 0))
 		goto out_err_init;
@@ -559,7 +551,6 @@ static void vmw_user_fence_base_release(struct ttm_base_object **p_base)
 int vmw_user_fence_create(struct drm_file *file_priv,
 			  struct vmw_fence_manager *fman,
 			  uint32_t seqno,
-			  uint32_t mask,
 			  struct vmw_fence_obj **p_fence,
 			  uint32_t *p_handle)
 {
@@ -586,7 +577,7 @@ int vmw_user_fence_create(struct drm_file *file_priv,
 	}
 
 	ret = vmw_fence_obj_init(fman, &ufence->fence, seqno,
-				 mask, vmw_user_fence_destroy);
+				 vmw_user_fence_destroy);
 	if (unlikely(ret != 0)) {
 		kfree(ufence);
 		goto out_no_object;
@@ -647,13 +638,12 @@ void vmw_fence_fifo_down(struct vmw_fence_manager *fman)
 		kref_get(&fence->kref);
 		spin_unlock_irq(&fman->lock);
 
-		ret = vmw_fence_obj_wait(fence, fence->signal_mask,
-					 false, false,
+		ret = vmw_fence_obj_wait(fence, false, false,
 					 VMW_FENCE_WAIT_TIMEOUT);
 
 		if (unlikely(ret != 0)) {
 			list_del_init(&fence->head);
-			fence->signaled |= DRM_VMW_FENCE_FLAG_EXEC;
+			fence->signaled = 1;
 			INIT_LIST_HEAD(&action_list);
 			list_splice_init(&fence->seq_passed_actions,
 					 &action_list);
@@ -716,14 +706,14 @@ int vmw_fence_obj_wait_ioctl(struct drm_device *dev, void *data,
 
 	timeout = jiffies;
 	if (time_after_eq(timeout, (unsigned long)arg->kernel_cookie)) {
-		ret = ((vmw_fence_obj_signaled(fence, arg->flags)) ?
+		ret = ((vmw_fence_obj_signaled(fence)) ?
 		       0 : -EBUSY);
 		goto out;
 	}
 
 	timeout = (unsigned long)arg->kernel_cookie - timeout;
 
-	ret = vmw_fence_obj_wait(fence, arg->flags, arg->lazy, true, timeout);
+	ret = vmw_fence_obj_wait(fence, arg->lazy, true, timeout);
 
 out:
 	ttm_base_object_unref(&base);
@@ -760,10 +750,10 @@ int vmw_fence_obj_signaled_ioctl(struct drm_device *dev, void *data,
 	fence = &(container_of(base, struct vmw_user_fence, base)->fence);
 	fman = fence->fman;
 
-	arg->signaled = vmw_fence_obj_signaled(fence, arg->flags);
+	arg->signaled = vmw_fence_obj_signaled(fence);
 	spin_lock_irq(&fman->lock);
 
-	arg->signaled_flags = fence->signaled;
+	arg->signaled_flags = arg->flags;
 	arg->passed_seqno = dev_priv->last_read_seqno;
 	spin_unlock_irq(&fman->lock);
 
@@ -908,7 +898,7 @@ static void vmw_fence_obj_add_action(struct vmw_fence_obj *fence,
 	spin_lock_irqsave(&fman->lock, irq_flags);
 
 	fman->pending_actions[action->type]++;
-	if (fence->signaled & DRM_VMW_FENCE_FLAG_EXEC) {
+	if (fence->signaled) {
 		struct list_head action_list;
 
 		INIT_LIST_HEAD(&action_list);
diff --git a/drivers/gpu/drm/vmwgfx/vmwgfx_fence.h b/drivers/gpu/drm/vmwgfx/vmwgfx_fence.h
index faf2e7873860..8c18d32bd1c3 100644
--- a/drivers/gpu/drm/vmwgfx/vmwgfx_fence.h
+++ b/drivers/gpu/drm/vmwgfx/vmwgfx_fence.h
@@ -56,7 +56,6 @@ struct vmw_fence_obj {
 	struct vmw_fence_manager *fman;
 	struct list_head head;
 	uint32_t signaled;
-	uint32_t signal_mask;
 	struct list_head seq_passed_actions;
 	void (*destroy)(struct vmw_fence_obj *fence);
 	wait_queue_head_t queue;
@@ -74,10 +73,9 @@ vmw_fence_obj_reference(struct vmw_fence_obj *fence);
 
 extern void vmw_fences_update(struct vmw_fence_manager *fman);
 
-extern bool vmw_fence_obj_signaled(struct vmw_fence_obj *fence,
-				   uint32_t flags);
+extern bool vmw_fence_obj_signaled(struct vmw_fence_obj *fence);
 
-extern int vmw_fence_obj_wait(struct vmw_fence_obj *fence, uint32_t flags,
+extern int vmw_fence_obj_wait(struct vmw_fence_obj *fence,
 			      bool lazy,
 			      bool interruptible, unsigned long timeout);
 
@@ -85,13 +83,11 @@ extern void vmw_fence_obj_flush(struct vmw_fence_obj *fence);
 
 extern int vmw_fence_create(struct vmw_fence_manager *fman,
 			    uint32_t seqno,
-			    uint32_t mask,
 			    struct vmw_fence_obj **p_fence);
 
 extern int vmw_user_fence_create(struct drm_file *file_priv,
 				 struct vmw_fence_manager *fman,
 				 uint32_t sequence,
-				 uint32_t mask,
 				 struct vmw_fence_obj **p_fence,
 				 uint32_t *p_handle);
 

[PATCH 14/19] drm/vmwgfx: rework to new fence interface

[Maarten Lankhorst]

Signed-off-by: Maarten Lankhorst <maarten.lankhorst@canonical.com>
---
 drivers/gpu/drm/vmwgfx/vmwgfx_execbuf.c  |    2 
 drivers/gpu/drm/vmwgfx/vmwgfx_fence.c    |  299 ++++++++++++++++++------------
 drivers/gpu/drm/vmwgfx/vmwgfx_fence.h    |   29 ++-
 drivers/gpu/drm/vmwgfx/vmwgfx_resource.c |    9 -
 4 files changed, 200 insertions(+), 139 deletions(-)

diff --git a/drivers/gpu/drm/vmwgfx/vmwgfx_execbuf.c b/drivers/gpu/drm/vmwgfx/vmwgfx_execbuf.c
index 214fcfd13cb2..0ceaddc8e4f7 100644
--- a/drivers/gpu/drm/vmwgfx/vmwgfx_execbuf.c
+++ b/drivers/gpu/drm/vmwgfx/vmwgfx_execbuf.c
@@ -2389,7 +2389,7 @@ vmw_execbuf_copy_fence_user(struct vmw_private *dev_priv,
 		BUG_ON(fence == NULL);
 
 		fence_rep.handle = fence_handle;
-		fence_rep.seqno = fence->seqno;
+		fence_rep.seqno = fence->base.seqno;
 		vmw_update_seqno(dev_priv, &dev_priv->fifo);
 		fence_rep.passed_seqno = dev_priv->last_read_seqno;
 	}
diff --git a/drivers/gpu/drm/vmwgfx/vmwgfx_fence.c b/drivers/gpu/drm/vmwgfx/vmwgfx_fence.c
index 05b9eea8e875..77f416b7552c 100644
--- a/drivers/gpu/drm/vmwgfx/vmwgfx_fence.c
+++ b/drivers/gpu/drm/vmwgfx/vmwgfx_fence.c
@@ -46,6 +46,7 @@ struct vmw_fence_manager {
 	bool goal_irq_on; /* Protected by @goal_irq_mutex */
 	bool seqno_valid; /* Protected by @lock, and may not be set to true
 			     without the @goal_irq_mutex held. */
+	unsigned ctx;
 };
 
 struct vmw_user_fence {
@@ -80,6 +81,12 @@ struct vmw_event_fence_action {
 	uint32_t *tv_usec;
 };
 
+static struct vmw_fence_manager *
+fman_from_fence(struct vmw_fence_obj *fence)
+{
+	return container_of(fence->base.lock, struct vmw_fence_manager, lock);
+}
+
 /**
  * Note on fencing subsystem usage of irqs:
  * Typically the vmw_fences_update function is called
@@ -102,25 +109,130 @@ struct vmw_event_fence_action {
  * objects with actions attached to them.
  */
 
-static void vmw_fence_obj_destroy_locked(struct kref *kref)
+static void vmw_fence_obj_destroy(struct fence *f)
 {
 	struct vmw_fence_obj *fence =
-		container_of(kref, struct vmw_fence_obj, kref);
+		container_of(f, struct vmw_fence_obj, base);
 
-	struct vmw_fence_manager *fman = fence->fman;
-	unsigned int num_fences;
+	struct vmw_fence_manager *fman = fman_from_fence(fence);
+	unsigned long irq_flags;
 
+	spin_lock_irqsave(&fman->lock, irq_flags);
 	list_del_init(&fence->head);
-	num_fences = --fman->num_fence_objects;
-	spin_unlock_irq(&fman->lock);
-	if (fence->destroy)
-		fence->destroy(fence);
-	else
-		kfree(fence);
+	--fman->num_fence_objects;
+	spin_unlock_irqrestore(&fman->lock, irq_flags);
+	fence->destroy(fence);
+}
 
-	spin_lock_irq(&fman->lock);
+static const char *vmw_fence_get_driver_name(struct fence *f)
+{
+	return "vmwgfx";
+}
+
+static const char *vmw_fence_get_timeline_name(struct fence *f)
+{
+	return "svga";
+}
+
+static bool vmw_fence_enable_signaling(struct fence *f)
+{
+	struct vmw_fence_obj *fence =
+		container_of(f, struct vmw_fence_obj, base);
+
+	struct vmw_fence_manager *fman = fman_from_fence(fence);
+
+	__le32 __iomem *fifo_mem = fman->dev_priv->mmio_virt;
+	u32 seqno = ioread32(fifo_mem + SVGA_FIFO_FENCE);
+	if (seqno - fence->base.seqno < VMW_FENCE_WRAP)
+		return false;
+
+	vmw_fifo_ping_host(fman->dev_priv, SVGA_SYNC_GENERIC);
+
+	return true;
+}
+
+struct vmwgfx_wait_cb {
+	struct fence_cb base;
+	struct task_struct *task;
+};
+
+static void
+vmwgfx_wait_cb(struct fence *fence, struct fence_cb *cb)
+{
+	struct vmwgfx_wait_cb *wait =
+		container_of(cb, struct vmwgfx_wait_cb, base);
+
+	wake_up_process(wait->task);
 }
 
+static void __vmw_fences_update(struct vmw_fence_manager *fman);
+
+static long vmw_fence_wait(struct fence *f, bool intr, signed long timeout)
+{
+	struct vmw_fence_obj *fence =
+		container_of(f, struct vmw_fence_obj, base);
+
+	struct vmw_fence_manager *fman = fman_from_fence(fence);
+	struct vmw_private *dev_priv = fman->dev_priv;
+	struct vmwgfx_wait_cb cb;
+	long ret = timeout;
+	unsigned long irq_flags;
+
+	if (likely(vmw_fence_obj_signaled(fence)))
+		return timeout;
+
+	vmw_fifo_ping_host(dev_priv, SVGA_SYNC_GENERIC);
+	vmw_seqno_waiter_add(dev_priv);
+
+	spin_lock_irqsave(f->lock, irq_flags);
+
+	if (intr && signal_pending(current)) {
+		ret = -ERESTARTSYS;
+		goto out;
+	}
+
+	cb.base.func = vmwgfx_wait_cb;
+	cb.task = current;
+	list_add(&cb.base.node, &f->cb_list);
+
+	while (ret > 0) {
+		__vmw_fences_update(fman);
+		if (test_bit(FENCE_FLAG_SIGNALED_BIT, &f->flags))
+			break;
+
+		if (intr)
+			__set_current_state(TASK_INTERRUPTIBLE);
+		else
+			__set_current_state(TASK_UNINTERRUPTIBLE);
+		spin_unlock_irqrestore(f->lock, irq_flags);
+
+		ret = schedule_timeout(ret);
+
+		spin_lock_irqsave(f->lock, irq_flags);
+		if (ret > 0 && intr && signal_pending(current))
+			ret = -ERESTARTSYS;
+	}
+
+	if (!list_empty(&cb.base.node))
+		list_del(&cb.base.node);
+	__set_current_state(TASK_RUNNING);
+
+out:
+	spin_unlock_irqrestore(f->lock, irq_flags);
+
+	vmw_seqno_waiter_remove(dev_priv);
+
+	return ret;
+}
+
+static struct fence_ops vmw_fence_ops = {
+	.get_driver_name = vmw_fence_get_driver_name,
+	.get_timeline_name = vmw_fence_get_timeline_name,
+	.enable_signaling = vmw_fence_enable_signaling,
+	.wait = vmw_fence_wait,
+	.release = vmw_fence_obj_destroy,
+};
+
 
 /**
  * Execute signal actions on fences recently signaled.
@@ -186,6 +298,7 @@ struct vmw_fence_manager *vmw_fence_manager_init(struct vmw_private *dev_priv)
 	fman->event_fence_action_size =
 		ttm_round_pot(sizeof(struct vmw_event_fence_action));
 	mutex_init(&fman->goal_irq_mutex);
+	fman->ctx = fence_context_alloc(1);
 
 	return fman;
 }
@@ -211,16 +324,12 @@ static int vmw_fence_obj_init(struct vmw_fence_manager *fman,
 			      void (*destroy) (struct vmw_fence_obj *fence))
 {
 	unsigned long irq_flags;
-	unsigned int num_fences;
 	int ret = 0;
 
-	fence->seqno = seqno;
+	fence_init(&fence->base, &vmw_fence_ops, &fman->lock,
+		   fman->ctx, seqno);
 	INIT_LIST_HEAD(&fence->seq_passed_actions);
-	fence->fman = fman;
-	fence->signaled = 0;
-	kref_init(&fence->kref);
 	fence->destroy = destroy;
-	init_waitqueue_head(&fence->queue);
 
 	spin_lock_irqsave(&fman->lock, irq_flags);
 	if (unlikely(fman->fifo_down)) {
@@ -228,7 +337,7 @@ static int vmw_fence_obj_init(struct vmw_fence_manager *fman,
 		goto out_unlock;
 	}
 	list_add_tail(&fence->head, &fman->fence_list);
-	num_fences = ++fman->num_fence_objects;
+	++fman->num_fence_objects;
 
 out_unlock:
 	spin_unlock_irqrestore(&fman->lock, irq_flags);
@@ -236,38 +345,6 @@ out_unlock:
 
 }
 
-struct vmw_fence_obj *vmw_fence_obj_reference(struct vmw_fence_obj *fence)
-{
-	if (unlikely(fence == NULL))
-		return NULL;
-
-	kref_get(&fence->kref);
-	return fence;
-}
-
-/**
- * vmw_fence_obj_unreference
- *
- * Note that this function may not be entered with disabled irqs since
- * it may re-enable them in the destroy function.
- *
- */
-void vmw_fence_obj_unreference(struct vmw_fence_obj **fence_p)
-{
-	struct vmw_fence_obj *fence = *fence_p;
-	struct vmw_fence_manager *fman;
-
-	if (unlikely(fence == NULL))
-		return;
-
-	fman = fence->fman;
-	*fence_p = NULL;
-	spin_lock_irq(&fman->lock);
-	BUG_ON(atomic_read(&fence->kref.refcount) == 0);
-	kref_put(&fence->kref, vmw_fence_obj_destroy_locked);
-	spin_unlock_irq(&fman->lock);
-}
-
 static void vmw_fences_perform_actions(struct vmw_fence_manager *fman,
 				struct list_head *list)
 {
@@ -323,7 +400,7 @@ static bool vmw_fence_goal_new_locked(struct vmw_fence_manager *fman,
 	list_for_each_entry(fence, &fman->fence_list, head) {
 		if (!list_empty(&fence->seq_passed_actions)) {
 			fman->seqno_valid = true;
-			iowrite32(fence->seqno,
+			iowrite32(fence->base.seqno,
 				  fifo_mem + SVGA_FIFO_FENCE_GOAL);
 			break;
 		}
@@ -350,27 +427,27 @@ static bool vmw_fence_goal_new_locked(struct vmw_fence_manager *fman,
  */
 static bool vmw_fence_goal_check_locked(struct vmw_fence_obj *fence)
 {
+	struct vmw_fence_manager *fman = fman_from_fence(fence);
 	u32 goal_seqno;
 	__le32 __iomem *fifo_mem;
 
-	if (fence->signaled)
+	if (fence_is_signaled_locked(&fence->base))
 		return false;
 
-	fifo_mem = fence->fman->dev_priv->mmio_virt;
+	fifo_mem = fman->dev_priv->mmio_virt;
 	goal_seqno = ioread32(fifo_mem + SVGA_FIFO_FENCE_GOAL);
-	if (likely(fence->fman->seqno_valid &&
-		   goal_seqno - fence->seqno < VMW_FENCE_WRAP))
+	if (likely(fman->seqno_valid &&
+		   goal_seqno - fence->base.seqno < VMW_FENCE_WRAP))
 		return false;
 
-	iowrite32(fence->seqno, fifo_mem + SVGA_FIFO_FENCE_GOAL);
-	fence->fman->seqno_valid = true;
+	iowrite32(fence->base.seqno, fifo_mem + SVGA_FIFO_FENCE_GOAL);
+	fman->seqno_valid = true;
 
 	return true;
 }
 
-void vmw_fences_update(struct vmw_fence_manager *fman)
+static void __vmw_fences_update(struct vmw_fence_manager *fman)
 {
-	unsigned long flags;
 	struct vmw_fence_obj *fence, *next_fence;
 	struct list_head action_list;
 	bool needs_rerun;
@@ -379,32 +456,25 @@ void vmw_fences_update(struct vmw_fence_manager *fman)
 
 	seqno = ioread32(fifo_mem + SVGA_FIFO_FENCE);
 rerun:
-	spin_lock_irqsave(&fman->lock, flags);
 	list_for_each_entry_safe(fence, next_fence, &fman->fence_list, head) {
-		if (seqno - fence->seqno < VMW_FENCE_WRAP) {
+		if (seqno - fence->base.seqno < VMW_FENCE_WRAP) {
 			list_del_init(&fence->head);
-			fence->signaled = 1;
+			fence_signal_locked(&fence->base);
 			INIT_LIST_HEAD(&action_list);
 			list_splice_init(&fence->seq_passed_actions,
 					 &action_list);
 			vmw_fences_perform_actions(fman, &action_list);
-			wake_up_all(&fence->queue);
 		} else
 			break;
 	}
 
-	needs_rerun = vmw_fence_goal_new_locked(fman, seqno);
-
-	if (!list_empty(&fman->cleanup_list))
-		(void) schedule_work(&fman->work);
-	spin_unlock_irqrestore(&fman->lock, flags);
-
 	/*
 	 * Rerun if the fence goal seqno was updated, and the
 	 * hardware might have raced with that update, so that
 	 * we missed a fence_goal irq.
 	 */
 
+	needs_rerun = vmw_fence_goal_new_locked(fman, seqno);
 	if (unlikely(needs_rerun)) {
 		new_seqno = ioread32(fifo_mem + SVGA_FIFO_FENCE);
 		if (new_seqno != seqno) {
@@ -412,75 +482,58 @@ rerun:
 			goto rerun;
 		}
 	}
+
+	if (!list_empty(&fman->cleanup_list))
+		(void) schedule_work(&fman->work);
 }
 
-bool vmw_fence_obj_signaled(struct vmw_fence_obj *fence)
+void vmw_fences_update(struct vmw_fence_manager *fman)
 {
-	struct vmw_fence_manager *fman = fence->fman;
 	unsigned long irq_flags;
-	uint32_t signaled;
 
 	spin_lock_irqsave(&fman->lock, irq_flags);
-	signaled = fence->signaled;
+	__vmw_fences_update(fman);
 	spin_unlock_irqrestore(&fman->lock, irq_flags);
+}
+
+bool vmw_fence_obj_signaled(struct vmw_fence_obj *fence)
+{
+	struct vmw_fence_manager *fman = fman_from_fence(fence);
 
-	if (signaled)
+	if (test_bit(FENCE_FLAG_SIGNALED_BIT, &fence->base.flags))
 		return 1;
 
 	vmw_fences_update(fman);
 
-	spin_lock_irqsave(&fman->lock, irq_flags);
-	signaled = fence->signaled;
-	spin_unlock_irqrestore(&fman->lock, irq_flags);
-
-	return signaled;
+	return fence_is_signaled(&fence->base);
 }
 
 int vmw_fence_obj_wait(struct vmw_fence_obj *fence, bool lazy,
 		       bool interruptible, unsigned long timeout)
 {
-	struct vmw_private *dev_priv = fence->fman->dev_priv;
-	long ret;
+	long ret = fence_wait_timeout(&fence->base, interruptible, timeout);
 
-	if (likely(vmw_fence_obj_signaled(fence)))
+	if (likely(ret > 0))
 		return 0;
-
-	vmw_fifo_ping_host(dev_priv, SVGA_SYNC_GENERIC);
-	vmw_seqno_waiter_add(dev_priv);
-
-	if (interruptible)
-		ret = wait_event_interruptible_timeout
-			(fence->queue,
-			 vmw_fence_obj_signaled(fence),
-			 timeout);
+	else if (ret == 0)
+		return -EBUSY;
 	else
-		ret = wait_event_timeout
-			(fence->queue,
-			 vmw_fence_obj_signaled(fence),
-			 timeout);
-
-	vmw_seqno_waiter_remove(dev_priv);
-
-	if (unlikely(ret == 0))
-		ret = -EBUSY;
-	else if (likely(ret > 0))
-		ret = 0;
-
-	return ret;
+		return ret;
 }
 
 void vmw_fence_obj_flush(struct vmw_fence_obj *fence)
 {
-	struct vmw_private *dev_priv = fence->fman->dev_priv;
+	struct vmw_private *dev_priv = fman_from_fence(fence)->dev_priv;
 
 	vmw_fifo_ping_host(dev_priv, SVGA_SYNC_GENERIC);
 }
 
 static void vmw_fence_destroy(struct vmw_fence_obj *fence)
 {
-	struct vmw_fence_manager *fman = fence->fman;
+	struct vmw_fence_manager *fman = fman_from_fence(fence);
+
+	fence_free(&fence->base);
 
-	kfree(fence);
 	/*
 	 * Free kernel space accounting.
 	 */
@@ -527,7 +580,7 @@ static void vmw_user_fence_destroy(struct vmw_fence_obj *fence)
 {
 	struct vmw_user_fence *ufence =
 		container_of(fence, struct vmw_user_fence, fence);
-	struct vmw_fence_manager *fman = fence->fman;
+	struct vmw_fence_manager *fman = fman_from_fence(fence);
 
 	ttm_base_object_kfree(ufence, base);
 	/*
@@ -620,7 +673,6 @@ out_no_object:
 
 void vmw_fence_fifo_down(struct vmw_fence_manager *fman)
 {
-	unsigned long irq_flags;
 	struct list_head action_list;
 	int ret;
 
@@ -629,13 +681,13 @@ void vmw_fence_fifo_down(struct vmw_fence_manager *fman)
 	 * restart when we've released the fman->lock.
 	 */
 
-	spin_lock_irqsave(&fman->lock, irq_flags);
+	spin_lock_irq(&fman->lock);
 	fman->fifo_down = true;
 	while (!list_empty(&fman->fence_list)) {
 		struct vmw_fence_obj *fence =
 			list_entry(fman->fence_list.prev, struct vmw_fence_obj,
 				   head);
-		kref_get(&fence->kref);
+		fence_get(&fence->base);
 		spin_unlock_irq(&fman->lock);
 
 		ret = vmw_fence_obj_wait(fence, false, false,
@@ -643,20 +695,18 @@ void vmw_fence_fifo_down(struct vmw_fence_manager *fman)
 
 		if (unlikely(ret != 0)) {
 			list_del_init(&fence->head);
-			fence->signaled = 1;
+			fence_signal(&fence->base);
 			INIT_LIST_HEAD(&action_list);
 			list_splice_init(&fence->seq_passed_actions,
 					 &action_list);
 			vmw_fences_perform_actions(fman, &action_list);
-			wake_up_all(&fence->queue);
 		}
 
-		spin_lock_irq(&fman->lock);
-
 		BUG_ON(!list_empty(&fence->head));
-		kref_put(&fence->kref, vmw_fence_obj_destroy_locked);
+		fence_put(&fence->base);
+		spin_lock_irq(&fman->lock);
 	}
-	spin_unlock_irqrestore(&fman->lock, irq_flags);
+	spin_unlock_irq(&fman->lock);
 }
 
 void vmw_fence_fifo_up(struct vmw_fence_manager *fman)
@@ -748,12 +798,12 @@ int vmw_fence_obj_signaled_ioctl(struct drm_device *dev, void *data,
 	}
 
 	fence = &(container_of(base, struct vmw_user_fence, base)->fence);
-	fman = fence->fman;
+	fman = fman_from_fence(fence);
 
 	arg->signaled = vmw_fence_obj_signaled(fence);
-	spin_lock_irq(&fman->lock);
 
 	arg->signaled_flags = arg->flags;
+	spin_lock_irq(&fman->lock);
 	arg->passed_seqno = dev_priv->last_read_seqno;
 	spin_unlock_irq(&fman->lock);
 
@@ -866,7 +916,7 @@ static void vmw_event_fence_action_cleanup(struct vmw_fence_action *action)
 {
 	struct vmw_event_fence_action *eaction =
 		container_of(action, struct vmw_event_fence_action, action);
-	struct vmw_fence_manager *fman = eaction->fence->fman;
+	struct vmw_fence_manager *fman = fman_from_fence(eaction->fence);
 	unsigned long irq_flags;
 
 	spin_lock_irqsave(&fman->lock, irq_flags);
@@ -890,7 +940,7 @@ static void vmw_event_fence_action_cleanup(struct vmw_fence_action *action)
 static void vmw_fence_obj_add_action(struct vmw_fence_obj *fence,
 			      struct vmw_fence_action *action)
 {
-	struct vmw_fence_manager *fman = fence->fman;
+	struct vmw_fence_manager *fman = fman_from_fence(fence);
 	unsigned long irq_flags;
 	bool run_update = false;
 
@@ -898,7 +948,7 @@ static void vmw_fence_obj_add_action(struct vmw_fence_obj *fence,
 	spin_lock_irqsave(&fman->lock, irq_flags);
 
 	fman->pending_actions[action->type]++;
-	if (fence->signaled) {
+	if (fence_is_signaled_locked(&fence->base)) {
 		struct list_head action_list;
 
 		INIT_LIST_HEAD(&action_list);
@@ -950,7 +1000,7 @@ int vmw_event_fence_action_queue(struct drm_file *file_priv,
 				 bool interruptible)
 {
 	struct vmw_event_fence_action *eaction;
-	struct vmw_fence_manager *fman = fence->fman;
+	struct vmw_fence_manager *fman = fman_from_fence(fence);
 	struct vmw_fpriv *vmw_fp = vmw_fpriv(file_priv);
 	unsigned long irq_flags;
 
@@ -990,7 +1040,8 @@ static int vmw_event_fence_action_create(struct drm_file *file_priv,
 				  bool interruptible)
 {
 	struct vmw_event_fence_pending *event;
-	struct drm_device *dev = fence->fman->dev_priv->dev;
+	struct vmw_fence_manager *fman = fman_from_fence(fence);
+	struct drm_device *dev = fman->dev_priv->dev;
 	unsigned long irq_flags;
 	int ret;
 
diff --git a/drivers/gpu/drm/vmwgfx/vmwgfx_fence.h b/drivers/gpu/drm/vmwgfx/vmwgfx_fence.h
index 8c18d32bd1c3..26a4add39208 100644
--- a/drivers/gpu/drm/vmwgfx/vmwgfx_fence.h
+++ b/drivers/gpu/drm/vmwgfx/vmwgfx_fence.h
@@ -27,6 +27,8 @@
 
 #ifndef _VMWGFX_FENCE_H_
 
+#include <linux/fence.h>
+
 #define VMW_FENCE_WAIT_TIMEOUT (5*HZ)
 
 struct vmw_private;
@@ -50,15 +52,11 @@ struct vmw_fence_action {
 };
 
 struct vmw_fence_obj {
-	struct kref kref;
-	u32 seqno;
+	struct fence base;
 
-	struct vmw_fence_manager *fman;
 	struct list_head head;
-	uint32_t signaled;
 	struct list_head seq_passed_actions;
 	void (*destroy)(struct vmw_fence_obj *fence);
-	wait_queue_head_t queue;
 };
 
 extern struct vmw_fence_manager *
@@ -66,10 +64,23 @@ vmw_fence_manager_init(struct vmw_private *dev_priv);
 
 extern void vmw_fence_manager_takedown(struct vmw_fence_manager *fman);
 
-extern void vmw_fence_obj_unreference(struct vmw_fence_obj **fence_p);
-
-extern struct vmw_fence_obj *
-vmw_fence_obj_reference(struct vmw_fence_obj *fence);
+static inline void
+vmw_fence_obj_unreference(struct vmw_fence_obj **fence_p)
+{
+	struct vmw_fence_obj *fence = *fence_p;
+
+	*fence_p = NULL;
+	if (fence)
+		fence_put(&fence->base);
+}
+
+static inline struct vmw_fence_obj *
+vmw_fence_obj_reference(struct vmw_fence_obj *fence)
+{
+	if (fence)
+		fence_get(&fence->base);
+	return fence;
+}
 
 extern void vmw_fences_update(struct vmw_fence_manager *fman);
 
diff --git a/drivers/gpu/drm/vmwgfx/vmwgfx_resource.c b/drivers/gpu/drm/vmwgfx/vmwgfx_resource.c
index 936777e469bb..294a0a3e48cb 100644
--- a/drivers/gpu/drm/vmwgfx/vmwgfx_resource.c
+++ b/drivers/gpu/drm/vmwgfx/vmwgfx_resource.c
@@ -1419,21 +1419,20 @@ void vmw_fence_single_bo(struct ttm_buffer_object *bo,
 			 struct vmw_fence_obj *fence)
 {
 	struct ttm_bo_device *bdev = bo->bdev;
-	struct ttm_bo_driver *driver = bdev->driver;
 	struct vmw_fence_obj *old_fence_obj;
 	struct vmw_private *dev_priv =
 		container_of(bdev, struct vmw_private, bdev);
 
-	if (fence == NULL)
+	if (fence == NULL) {
 		vmw_execbuf_fence_commands(NULL, dev_priv, &fence, NULL);
-	else
-		driver->sync_obj_ref(fence);
+	} else
+		vmw_fence_obj_reference(fence);
 
+	reservation_object_add_excl_fence(bo->resv, &fence->base);
 
 	old_fence_obj = bo->sync_obj;
 	bo->sync_obj = fence;
 
-
 	if (old_fence_obj)
 		vmw_fence_obj_unreference(&old_fence_obj);
 }

[PATCH 15/19] drm/ttm: flip the switch, and convert to dma_fence

[Maarten Lankhorst]

---
 drivers/gpu/drm/nouveau/nouveau_bo.c     |   48 +-------
 drivers/gpu/drm/nouveau/nouveau_fence.c  |   24 +---
 drivers/gpu/drm/nouveau/nouveau_fence.h  |    2 
 drivers/gpu/drm/nouveau/nouveau_gem.c    |   16 ++-
 drivers/gpu/drm/qxl/qxl_debugfs.c        |    6 +
 drivers/gpu/drm/qxl/qxl_drv.h            |    2 
 drivers/gpu/drm/qxl/qxl_kms.c            |    1 
 drivers/gpu/drm/qxl/qxl_object.h         |    4 -
 drivers/gpu/drm/qxl/qxl_release.c        |    3 -
 drivers/gpu/drm/qxl/qxl_ttm.c            |  104 ------------------
 drivers/gpu/drm/radeon/radeon_cs.c       |   10 +-
 drivers/gpu/drm/radeon/radeon_display.c  |    3 -
 drivers/gpu/drm/radeon/radeon_object.c   |    4 -
 drivers/gpu/drm/radeon/radeon_ttm.c      |   34 ------
 drivers/gpu/drm/radeon/radeon_uvd.c      |    8 +
 drivers/gpu/drm/radeon/radeon_vm.c       |   14 ++
 drivers/gpu/drm/ttm/ttm_bo.c             |  171 +++++++++++++++++++++---------
 drivers/gpu/drm/ttm/ttm_bo_util.c        |   23 +---
 drivers/gpu/drm/ttm/ttm_execbuf_util.c   |   10 --
 drivers/gpu/drm/vmwgfx/vmwgfx_buffer.c   |   40 -------
 drivers/gpu/drm/vmwgfx/vmwgfx_resource.c |   14 +-
 include/drm/ttm/ttm_bo_api.h             |    2 
 include/drm/ttm/ttm_bo_driver.h          |   26 -----
 include/drm/ttm/ttm_execbuf_util.h       |   10 +-
 24 files changed, 193 insertions(+), 386 deletions(-)

diff --git a/drivers/gpu/drm/nouveau/nouveau_bo.c b/drivers/gpu/drm/nouveau/nouveau_bo.c
index 6e71d3dc2d87..3d8a8e18caea 100644
--- a/drivers/gpu/drm/nouveau/nouveau_bo.c
+++ b/drivers/gpu/drm/nouveau/nouveau_bo.c
@@ -92,13 +92,13 @@ nv10_bo_get_tile_region(struct drm_device *dev, int i)
 
 static void
 nv10_bo_put_tile_region(struct drm_device *dev, struct nouveau_drm_tile *tile,
-			struct nouveau_fence *fence)
+			struct fence *fence)
 {
 	struct nouveau_drm *drm = nouveau_drm(dev);
 
 	if (tile) {
 		spin_lock(&drm->tile.lock);
-		tile->fence = nouveau_fence_ref(fence);
+		tile->fence = nouveau_fence_ref((struct nouveau_fence *)fence);
 		tile->used = false;
 		spin_unlock(&drm->tile.lock);
 	}
@@ -965,7 +965,8 @@ nouveau_bo_move_m2mf(struct ttm_buffer_object *bo, int evict, bool intr,
 		if (ret == 0) {
 			ret = nouveau_fence_new(chan, false, &fence);
 			if (ret == 0) {
-				ret = ttm_bo_move_accel_cleanup(bo, fence,
+				ret = ttm_bo_move_accel_cleanup(bo,
+								&fence->base,
 								evict,
 								no_wait_gpu,
 								new_mem);
@@ -1151,8 +1152,9 @@ nouveau_bo_vm_cleanup(struct ttm_buffer_object *bo,
 {
 	struct nouveau_drm *drm = nouveau_bdev(bo->bdev);
 	struct drm_device *dev = drm->dev;
+	struct fence *fence = reservation_object_get_excl(bo->resv);
 
-	nv10_bo_put_tile_region(dev, *old_tile, bo->sync_obj);
+	nv10_bo_put_tile_region(dev, *old_tile, fence);
 	*old_tile = new_tile;
 }
 
@@ -1423,47 +1425,14 @@ nouveau_ttm_tt_unpopulate(struct ttm_tt *ttm)
 	ttm_pool_unpopulate(ttm);
 }
 
-static void
-nouveau_bo_fence_unref(void **sync_obj)
-{
-	nouveau_fence_unref((struct nouveau_fence **)sync_obj);
-}
-
 void
 nouveau_bo_fence(struct nouveau_bo *nvbo, struct nouveau_fence *fence)
 {
 	struct reservation_object *resv = nvbo->bo.resv;
 
-	nouveau_bo_fence_unref(&nvbo->bo.sync_obj);
-	nvbo->bo.sync_obj = nouveau_fence_ref(fence);
-
 	reservation_object_add_excl_fence(resv, &fence->base);
 }
 
-static void *
-nouveau_bo_fence_ref(void *sync_obj)
-{
-	return nouveau_fence_ref(sync_obj);
-}
-
-static bool
-nouveau_bo_fence_signalled(void *sync_obj)
-{
-	return nouveau_fence_done(sync_obj);
-}
-
-static int
-nouveau_bo_fence_wait(void *sync_obj, bool lazy, bool intr)
-{
-	return nouveau_fence_wait(sync_obj, lazy, intr);
-}
-
-static int
-nouveau_bo_fence_flush(void *sync_obj)
-{
-	return 0;
-}
-
 struct ttm_bo_driver nouveau_bo_driver = {
 	.ttm_tt_create = &nouveau_ttm_tt_create,
 	.ttm_tt_populate = &nouveau_ttm_tt_populate,
@@ -1474,11 +1443,6 @@ struct ttm_bo_driver nouveau_bo_driver = {
 	.move_notify = nouveau_bo_move_ntfy,
 	.move = nouveau_bo_move,
 	.verify_access = nouveau_bo_verify_access,
-	.sync_obj_signaled = nouveau_bo_fence_signalled,
-	.sync_obj_wait = nouveau_bo_fence_wait,
-	.sync_obj_flush = nouveau_bo_fence_flush,
-	.sync_obj_unref = nouveau_bo_fence_unref,
-	.sync_obj_ref = nouveau_bo_fence_ref,
 	.fault_reserve_notify = &nouveau_ttm_fault_reserve_notify,
 	.io_mem_reserve = &nouveau_ttm_io_mem_reserve,
 	.io_mem_free = &nouveau_ttm_io_mem_free,
diff --git a/drivers/gpu/drm/nouveau/nouveau_fence.c b/drivers/gpu/drm/nouveau/nouveau_fence.c
index d24f8ce4341a..9f92ad37637d 100644
--- a/drivers/gpu/drm/nouveau/nouveau_fence.c
+++ b/drivers/gpu/drm/nouveau/nouveau_fence.c
@@ -139,17 +139,18 @@ static bool nouveau_fence_is_signaled(struct fence *f)
 }
 
 void
-nouveau_fence_work(struct nouveau_fence *fence,
+nouveau_fence_work(struct fence *fence,
 		   void (*func)(void *), void *data)
 {
 	struct nouveau_fence_work *work;
 
-	if (fence_is_signaled(&fence->base))
+	if (fence_is_signaled(fence))
 		goto err;
 
 	work = kmalloc(sizeof(*work), GFP_KERNEL);
 	if (!work) {
-		WARN_ON(nouveau_fence_wait(fence, false, false));
+		WARN_ON(nouveau_fence_wait((struct nouveau_fence *)fence,
+					   false, false));
 		goto err;
 	}
 
@@ -157,7 +158,7 @@ nouveau_fence_work(struct nouveau_fence *fence,
 	work->func = func;
 	work->data = data;
 
-	if (fence_add_callback(&fence->base, &work->cb, nouveau_fence_work_cb) < 0)
+	if (fence_add_callback(fence, &work->cb, nouveau_fence_work_cb) < 0)
 		goto err_free;
 	return;
 
@@ -322,14 +323,9 @@ nouveau_fence_sync(struct nouveau_bo *nvbo, struct nouveau_channel *chan)
 	struct reservation_object_list *fobj;
 	int ret = 0, i;
 
-	fence = nvbo->bo.sync_obj;
-	if (fence && fence_is_signaled(fence)) {
-		nouveau_fence_unref((struct nouveau_fence **)
-				    &nvbo->bo.sync_obj);
-		fence = NULL;
-	}
+	fence = reservation_object_get_excl(resv);
 
-	if (fence) {
+	if (fence && !fence_is_signaled(fence)) {
 		struct nouveau_fence *f = container_of(fence,
 						       struct nouveau_fence,
 						       base);
@@ -345,12 +341,8 @@ nouveau_fence_sync(struct nouveau_bo *nvbo, struct nouveau_channel *chan)
 	if (ret)
 		return ret;
 
-	fence = reservation_object_get_excl(resv);
-	if (fence && !nouveau_local_fence(fence, chan->drm))
-		ret = fence_wait(fence, true);
-
 	fobj = reservation_object_get_list(resv);
-	if (!fobj || ret)
+	if (!fobj)
 		return ret;
 
 	for (i = 0; i < fobj->shared_count && !ret; ++i) {
diff --git a/drivers/gpu/drm/nouveau/nouveau_fence.h b/drivers/gpu/drm/nouveau/nouveau_fence.h
index 1989ec22e66e..41abc8a44e3c 100644
--- a/drivers/gpu/drm/nouveau/nouveau_fence.h
+++ b/drivers/gpu/drm/nouveau/nouveau_fence.h
@@ -26,7 +26,7 @@ void nouveau_fence_unref(struct nouveau_fence **);
 
 int  nouveau_fence_emit(struct nouveau_fence *, struct nouveau_channel *);
 bool nouveau_fence_done(struct nouveau_fence *);
-void nouveau_fence_work(struct nouveau_fence *, void (*)(void *), void *);
+void nouveau_fence_work(struct fence *, void (*)(void *), void *);
 int  nouveau_fence_wait(struct nouveau_fence *, bool lazy, bool intr);
 int  nouveau_fence_sync(struct nouveau_bo *, struct nouveau_channel *);
 
diff --git a/drivers/gpu/drm/nouveau/nouveau_gem.c b/drivers/gpu/drm/nouveau/nouveau_gem.c
index 2bd62f4ecf6a..0a1fee7eb211 100644
--- a/drivers/gpu/drm/nouveau/nouveau_gem.c
+++ b/drivers/gpu/drm/nouveau/nouveau_gem.c
@@ -100,13 +100,12 @@ static void
 nouveau_gem_object_unmap(struct nouveau_bo *nvbo, struct nouveau_vma *vma)
 {
 	const bool mapped = nvbo->bo.mem.mem_type != TTM_PL_SYSTEM;
-	struct nouveau_fence *fence = NULL;
+	struct fence *fence = NULL;
 
 	list_del(&vma->head);
 
-	if (mapped) {
-		fence = nouveau_fence_ref(nvbo->bo.sync_obj);
-	}
+	if (mapped)
+		fence = reservation_object_get_excl(nvbo->bo.resv);
 
 	if (fence) {
 		nouveau_fence_work(fence, nouveau_gem_object_delete, vma);
@@ -116,7 +115,6 @@ nouveau_gem_object_unmap(struct nouveau_bo *nvbo, struct nouveau_vma *vma)
 		nouveau_vm_put(vma);
 		kfree(vma);
 	}
-	nouveau_fence_unref(&fence);
 }
 
 void
@@ -876,8 +874,12 @@ nouveau_gem_ioctl_cpu_prep(struct drm_device *dev, void *data,
 	ret = ttm_bo_reserve(&nvbo->bo, true, false, false, 0);
 	if (!ret) {
 		ret = ttm_bo_wait(&nvbo->bo, true, true, true);
-		if (!no_wait && ret)
-			fence = nouveau_fence_ref(nvbo->bo.sync_obj);
+		if (!no_wait && ret) {
+			struct fence *excl;
+
+			excl = reservation_object_get_excl(nvbo->bo.resv);
+			fence = nouveau_fence_ref((struct nouveau_fence *)excl);
+		}
 
 		ttm_bo_unreserve(&nvbo->bo);
 	}
diff --git a/drivers/gpu/drm/qxl/qxl_debugfs.c b/drivers/gpu/drm/qxl/qxl_debugfs.c
index 0d144e0646d6..a4a63fd84803 100644
--- a/drivers/gpu/drm/qxl/qxl_debugfs.c
+++ b/drivers/gpu/drm/qxl/qxl_debugfs.c
@@ -67,9 +67,9 @@ qxl_debugfs_buffers_info(struct seq_file *m, void *data)
 		rel = fobj ? fobj->shared_count : 0;
 		rcu_read_unlock();
 
-		seq_printf(m, "size %ld, pc %d, sync obj %p, num releases %d\n",
-			   (unsigned long)bo->gem_base.size, bo->pin_count,
-			   bo->tbo.sync_obj, rel);
+		seq_printf(m, "size %ld, pc %d, num releases %d\n",
+			   (unsigned long)bo->gem_base.size,
+			   bo->pin_count, rel);
 	}
 	spin_unlock(&qdev->release_lock);
 	return 0;
diff --git a/drivers/gpu/drm/qxl/qxl_drv.h b/drivers/gpu/drm/qxl/qxl_drv.h
index d547cbdebeb4..74e2117ee0e6 100644
--- a/drivers/gpu/drm/qxl/qxl_drv.h
+++ b/drivers/gpu/drm/qxl/qxl_drv.h
@@ -280,9 +280,7 @@ struct qxl_device {
 	uint8_t		slot_gen_bits;
 	uint64_t	va_slot_mask;
 
-	/* XXX: when rcu becomes available, release_lock can be killed */
 	spinlock_t	release_lock;
-	spinlock_t	fence_lock;
 	struct idr	release_idr;
 	uint32_t	release_seqno;
 	spinlock_t release_idr_lock;
diff --git a/drivers/gpu/drm/qxl/qxl_kms.c b/drivers/gpu/drm/qxl/qxl_kms.c
index a9e7c30e92c5..7234561e09d9 100644
--- a/drivers/gpu/drm/qxl/qxl_kms.c
+++ b/drivers/gpu/drm/qxl/qxl_kms.c
@@ -224,7 +224,6 @@ static int qxl_device_init(struct qxl_device *qdev,
 	idr_init(&qdev->release_idr);
 	spin_lock_init(&qdev->release_idr_lock);
 	spin_lock_init(&qdev->release_lock);
-	spin_lock_init(&qdev->fence_lock);
 
 	idr_init(&qdev->surf_id_idr);
 	spin_lock_init(&qdev->surf_id_idr_lock);
diff --git a/drivers/gpu/drm/qxl/qxl_object.h b/drivers/gpu/drm/qxl/qxl_object.h
index 1edaf5768086..37af1bc0dd00 100644
--- a/drivers/gpu/drm/qxl/qxl_object.h
+++ b/drivers/gpu/drm/qxl/qxl_object.h
@@ -78,8 +78,8 @@ static inline int qxl_bo_wait(struct qxl_bo *bo, u32 *mem_type,
 	}
 	if (mem_type)
 		*mem_type = bo->tbo.mem.mem_type;
-	if (bo->tbo.sync_obj)
-		r = ttm_bo_wait(&bo->tbo, true, true, no_wait);
+
+	r = ttm_bo_wait(&bo->tbo, true, true, no_wait);
 	ttm_bo_unreserve(&bo->tbo);
 	return r;
 }
diff --git a/drivers/gpu/drm/qxl/qxl_release.c b/drivers/gpu/drm/qxl/qxl_release.c
index 9731d2540a40..15158c5a5b3a 100644
--- a/drivers/gpu/drm/qxl/qxl_release.c
+++ b/drivers/gpu/drm/qxl/qxl_release.c
@@ -464,9 +464,6 @@ void qxl_release_fence_buffer_objects(struct qxl_release *release)
 		bo = entry->bo;
 		qbo = to_qxl_bo(bo);
 
-		if (!entry->bo->sync_obj)
-			entry->bo->sync_obj = qbo;
-
 		reservation_object_add_shared_fence(bo->resv, &release->base);
 		ttm_bo_add_to_lru(bo);
 		__ttm_bo_unreserve(bo);
diff --git a/drivers/gpu/drm/qxl/qxl_ttm.c b/drivers/gpu/drm/qxl/qxl_ttm.c
index 6230251fa5b0..99b7ee110a98 100644
--- a/drivers/gpu/drm/qxl/qxl_ttm.c
+++ b/drivers/gpu/drm/qxl/qxl_ttm.c
@@ -355,105 +355,6 @@ static int qxl_bo_move(struct ttm_buffer_object *bo,
 	return ttm_bo_move_memcpy(bo, evict, no_wait_gpu, new_mem);
 }
 
-static bool qxl_sync_obj_signaled(void *sync_obj);
-
-static int qxl_sync_obj_wait(void *sync_obj,
-			     bool lazy, bool interruptible)
-{
-	struct qxl_bo *bo = (struct qxl_bo *)sync_obj;
-	struct qxl_device *qdev = bo->gem_base.dev->dev_private;
-	struct reservation_object_list *fobj;
-	int count = 0, sc = 0, num_release = 0;
-	bool have_drawable_releases;
-
-retry:
-	if (sc == 0) {
-		if (bo->type == QXL_GEM_DOMAIN_SURFACE)
-			qxl_update_surface(qdev, bo);
-	} else if (sc >= 1) {
-		qxl_io_notify_oom(qdev);
-	}
-
-	sc++;
-
-	for (count = 0; count < 10; count++) {
-		if (qxl_sync_obj_signaled(sync_obj))
-			return 0;
-
-		if (!qxl_queue_garbage_collect(qdev, true))
-			break;
-	}
-
-	have_drawable_releases = false;
-	num_release = 0;
-
-	spin_lock(&qdev->release_lock);
-	fobj = bo->tbo.resv->fence;
-	for (count = 0; fobj && count < fobj->shared_count; count++) {
-		struct qxl_release *release;
-
-		release = container_of(fobj->shared[count],
-				       struct qxl_release, base);
-
-		if (fence_is_signaled(&release->base))
-			continue;
-
-		num_release++;
-
-		if (release->type == QXL_RELEASE_DRAWABLE)
-			have_drawable_releases = true;
-	}
-	spin_unlock(&qdev->release_lock);
-
-	qxl_queue_garbage_collect(qdev, true);
-
-	if (have_drawable_releases || sc < 4) {
-		if (sc > 2)
-			/* back off */
-			usleep_range(500, 1000);
-		if (have_drawable_releases && sc > 300) {
-			WARN(1, "sync obj %d still has outstanding releases %d %d %d %ld %d\n", sc, bo->surface_id, bo->is_primary, bo->pin_count, (unsigned long)bo->gem_base.size, num_release);
-			return -EBUSY;
-		}
-		goto retry;
-	}
-	return 0;
-}
-
-static int qxl_sync_obj_flush(void *sync_obj)
-{
-	return 0;
-}
-
-static void qxl_sync_obj_unref(void **sync_obj)
-{
-	*sync_obj = NULL;
-}
-
-static void *qxl_sync_obj_ref(void *sync_obj)
-{
-	return sync_obj;
-}
-
-static bool qxl_sync_obj_signaled(void *sync_obj)
-{
-	struct qxl_bo *qbo = (struct qxl_bo *)sync_obj;
-	struct qxl_device *qdev = qbo->gem_base.dev->dev_private;
-	struct reservation_object_list *fobj;
-	bool ret = true;
-	unsigned i;
-
-	spin_lock(&qdev->release_lock);
-	fobj = qbo->tbo.resv->fence;
-	for (i = 0; fobj && i < fobj->shared_count; ++i) {
-		ret = fence_is_signaled(fobj->shared[i]);
-		if (!ret)
-			break;
-	}
-	spin_unlock(&qdev->release_lock);
-	return ret;
-}
-
 static void qxl_bo_move_notify(struct ttm_buffer_object *bo,
 			       struct ttm_mem_reg *new_mem)
 {
@@ -480,11 +381,6 @@ static struct ttm_bo_driver qxl_bo_driver = {
 	.verify_access = &qxl_verify_access,
 	.io_mem_reserve = &qxl_ttm_io_mem_reserve,
 	.io_mem_free = &qxl_ttm_io_mem_free,
-	.sync_obj_signaled = &qxl_sync_obj_signaled,
-	.sync_obj_wait = &qxl_sync_obj_wait,
-	.sync_obj_flush = &qxl_sync_obj_flush,
-	.sync_obj_unref = &qxl_sync_obj_unref,
-	.sync_obj_ref = &qxl_sync_obj_ref,
 	.move_notify = &qxl_bo_move_notify,
 };
 
diff --git a/drivers/gpu/drm/radeon/radeon_cs.c b/drivers/gpu/drm/radeon/radeon_cs.c
index ae763f60c8a0..687b4fdc988e 100644
--- a/drivers/gpu/drm/radeon/radeon_cs.c
+++ b/drivers/gpu/drm/radeon/radeon_cs.c
@@ -228,11 +228,17 @@ static void radeon_cs_sync_rings(struct radeon_cs_parser *p)
 	int i;
 
 	for (i = 0; i < p->nrelocs; i++) {
+		struct reservation_object *resv;
+		struct fence *fence;
+
 		if (!p->relocs[i].robj)
 			continue;
 
+		resv = p->relocs[i].robj->tbo.resv;
+		fence = reservation_object_get_excl(resv);
+
 		radeon_semaphore_sync_to(p->ib.semaphore,
-					 p->relocs[i].robj->tbo.sync_obj);
+					 (struct radeon_fence *)fence);
 	}
 }
 
@@ -402,7 +408,7 @@ static void radeon_cs_parser_fini(struct radeon_cs_parser *parser, int error, bo
 
 		ttm_eu_fence_buffer_objects(&parser->ticket,
 					    &parser->validated,
-					    parser->ib.fence);
+					    &parser->ib.fence->base);
 	} else if (backoff) {
 		ttm_eu_backoff_reservation(&parser->ticket,
 					   &parser->validated);
diff --git a/drivers/gpu/drm/radeon/radeon_display.c b/drivers/gpu/drm/radeon/radeon_display.c
index 3402540aa641..45bdc829bb14 100644
--- a/drivers/gpu/drm/radeon/radeon_display.c
+++ b/drivers/gpu/drm/radeon/radeon_display.c
@@ -481,8 +481,7 @@ static int radeon_crtc_page_flip(struct drm_crtc *crtc,
 		goto cleanup;
 	}
 
-	if (new_rbo->tbo.sync_obj)
-		work->fence = radeon_fence_ref(new_rbo->tbo.sync_obj);
+	work->fence = fence_get(reservation_object_get_excl(new_rbo->tbo.resv));
 
 	radeon_bo_get_tiling_flags(new_rbo, &tiling_flags, NULL);
 	radeon_bo_unreserve(new_rbo);
diff --git a/drivers/gpu/drm/radeon/radeon_object.c b/drivers/gpu/drm/radeon/radeon_object.c
index 8538aebb6580..53104f80d382 100644
--- a/drivers/gpu/drm/radeon/radeon_object.c
+++ b/drivers/gpu/drm/radeon/radeon_object.c
@@ -736,8 +736,8 @@ int radeon_bo_wait(struct radeon_bo *bo, u32 *mem_type, bool no_wait)
 		return r;
 	if (mem_type)
 		*mem_type = bo->tbo.mem.mem_type;
-	if (bo->tbo.sync_obj)
-		r = ttm_bo_wait(&bo->tbo, true, true, no_wait);
+
+	r = ttm_bo_wait(&bo->tbo, true, true, no_wait);
 	ttm_bo_unreserve(&bo->tbo);
 	return r;
 }
diff --git a/drivers/gpu/drm/radeon/radeon_ttm.c b/drivers/gpu/drm/radeon/radeon_ttm.c
index c8a8a5144ec1..715e29f984c1 100644
--- a/drivers/gpu/drm/radeon/radeon_ttm.c
+++ b/drivers/gpu/drm/radeon/radeon_ttm.c
@@ -265,12 +265,12 @@ static int radeon_move_blit(struct ttm_buffer_object *bo,
 	BUILD_BUG_ON((PAGE_SIZE % RADEON_GPU_PAGE_SIZE) != 0);
 
 	/* sync other rings */
-	fence = bo->sync_obj;
+	fence = (struct radeon_fence *)reservation_object_get_excl(bo->resv);
 	r = radeon_copy(rdev, old_start, new_start,
 			new_mem->num_pages * (PAGE_SIZE / RADEON_GPU_PAGE_SIZE), /* GPU pages */
 			&fence);
 	/* FIXME: handle copy error */
-	r = ttm_bo_move_accel_cleanup(bo, (void *)fence,
+	r = ttm_bo_move_accel_cleanup(bo, &fence->base,
 				      evict, no_wait_gpu, new_mem);
 	radeon_fence_unref(&fence);
 	return r;
@@ -483,31 +483,6 @@ static void radeon_ttm_io_mem_free(struct ttm_bo_device *bdev, struct ttm_mem_re
 {
 }
 
-static int radeon_sync_obj_wait(void *sync_obj, bool lazy, bool interruptible)
-{
-	return radeon_fence_wait((struct radeon_fence *)sync_obj, interruptible);
-}
-
-static int radeon_sync_obj_flush(void *sync_obj)
-{
-	return 0;
-}
-
-static void radeon_sync_obj_unref(void **sync_obj)
-{
-	radeon_fence_unref((struct radeon_fence **)sync_obj);
-}
-
-static void *radeon_sync_obj_ref(void *sync_obj)
-{
-	return radeon_fence_ref((struct radeon_fence *)sync_obj);
-}
-
-static bool radeon_sync_obj_signaled(void *sync_obj)
-{
-	return radeon_fence_signaled((struct radeon_fence *)sync_obj);
-}
-
 /*
  * TTM backend functions.
  */
@@ -685,11 +660,6 @@ static struct ttm_bo_driver radeon_bo_driver = {
 	.evict_flags = &radeon_evict_flags,
 	.move = &radeon_bo_move,
 	.verify_access = &radeon_verify_access,
-	.sync_obj_signaled = &radeon_sync_obj_signaled,
-	.sync_obj_wait = &radeon_sync_obj_wait,
-	.sync_obj_flush = &radeon_sync_obj_flush,
-	.sync_obj_unref = &radeon_sync_obj_unref,
-	.sync_obj_ref = &radeon_sync_obj_ref,
 	.move_notify = &radeon_bo_move_notify,
 	.fault_reserve_notify = &radeon_bo_fault_reserve_notify,
 	.io_mem_reserve = &radeon_ttm_io_mem_reserve,
diff --git a/drivers/gpu/drm/radeon/radeon_uvd.c b/drivers/gpu/drm/radeon/radeon_uvd.c
index 67b2a367df40..b35655e2e35e 100644
--- a/drivers/gpu/drm/radeon/radeon_uvd.c
+++ b/drivers/gpu/drm/radeon/radeon_uvd.c
@@ -356,6 +356,7 @@ static int radeon_uvd_cs_msg(struct radeon_cs_parser *p, struct radeon_bo *bo,
 {
 	int32_t *msg, msg_type, handle;
 	unsigned img_size = 0;
+	struct fence *f;
 	void *ptr;
 
 	int i, r;
@@ -365,8 +366,9 @@ static int radeon_uvd_cs_msg(struct radeon_cs_parser *p, struct radeon_bo *bo,
 		return -EINVAL;
 	}
 
-	if (bo->tbo.sync_obj) {
-		r = radeon_fence_wait(bo->tbo.sync_obj, false);
+	f = reservation_object_get_excl(bo->tbo.resv);
+	if (f) {
+		r = radeon_fence_wait((struct radeon_fence *)f, false);
 		if (r) {
 			DRM_ERROR("Failed waiting for UVD message (%d)!\n", r);
 			return r;
@@ -649,7 +651,7 @@ static int radeon_uvd_send_msg(struct radeon_device *rdev,
 	r = radeon_ib_schedule(rdev, &ib, NULL);
 	if (r)
 		goto err;
-	ttm_eu_fence_buffer_objects(&ticket, &head, ib.fence);
+	ttm_eu_fence_buffer_objects(&ticket, &head, &ib.fence->base);
 
 	if (fence)
 		*fence = radeon_fence_ref(ib.fence);
diff --git a/drivers/gpu/drm/radeon/radeon_vm.c b/drivers/gpu/drm/radeon/radeon_vm.c
index 0a0d0291af0d..e353eacb63b1 100644
--- a/drivers/gpu/drm/radeon/radeon_vm.c
+++ b/drivers/gpu/drm/radeon/radeon_vm.c
@@ -389,7 +389,7 @@ static int radeon_vm_clear_bo(struct radeon_device *rdev,
 	if (r)
                 goto error;
 
-	ttm_eu_fence_buffer_objects(&ticket, &head, ib.fence);
+	ttm_eu_fence_buffer_objects(&ticket, &head, &ib.fence->base);
 	radeon_ib_free(rdev, &ib);
 
 	return 0;
@@ -658,7 +658,12 @@ int radeon_vm_update_page_directory(struct radeon_device *rdev,
 					incr, R600_PTE_VALID);
 
 	if (ib.length_dw != 0) {
-		radeon_semaphore_sync_to(ib.semaphore, pd->tbo.sync_obj);
+		struct fence *fence;
+
+		fence = reservation_object_get_excl(pd->tbo.resv);
+		radeon_semaphore_sync_to(ib.semaphore,
+					 (struct radeon_fence *)fence);
+
 		radeon_semaphore_sync_to(ib.semaphore, vm->last_id_use);
 		r = radeon_ib_schedule(rdev, &ib, NULL);
 		if (r) {
@@ -786,8 +791,11 @@ static void radeon_vm_update_ptes(struct radeon_device *rdev,
 		struct radeon_bo *pt = vm->page_tables[pt_idx].bo;
 		unsigned nptes;
 		uint64_t pte;
+		struct fence *fence;
 
-		radeon_semaphore_sync_to(ib->semaphore, pt->tbo.sync_obj);
+		fence = reservation_object_get_excl(pt->tbo.resv);
+		radeon_semaphore_sync_to(ib->semaphore,
+					 (struct radeon_fence *)fence);
 
 		if ((addr & ~mask) == (end & ~mask))
 			nptes = end - addr;
diff --git a/drivers/gpu/drm/ttm/ttm_bo.c b/drivers/gpu/drm/ttm/ttm_bo.c
index c441ac6d546b..9d36cffad3b7 100644
--- a/drivers/gpu/drm/ttm/ttm_bo.c
+++ b/drivers/gpu/drm/ttm/ttm_bo.c
@@ -40,6 +40,7 @@
 #include <linux/file.h>
 #include <linux/module.h>
 #include <linux/atomic.h>
+#include <linux/reservation.h>
 
 #define TTM_ASSERT_LOCKED(param)
 #define TTM_DEBUG(fmt, arg...)
@@ -141,7 +142,6 @@ static void ttm_bo_release_list(struct kref *list_kref)
 	BUG_ON(atomic_read(&bo->list_kref.refcount));
 	BUG_ON(atomic_read(&bo->kref.refcount));
 	BUG_ON(atomic_read(&bo->cpu_writers));
-	BUG_ON(bo->sync_obj != NULL);
 	BUG_ON(bo->mem.mm_node != NULL);
 	BUG_ON(!list_empty(&bo->lru));
 	BUG_ON(!list_empty(&bo->ddestroy));
@@ -402,12 +402,30 @@ static void ttm_bo_cleanup_memtype_use(struct ttm_buffer_object *bo)
 	ww_mutex_unlock (&bo->resv->lock);
 }
 
+static void ttm_bo_flush_all_fences(struct ttm_buffer_object *bo)
+{
+	struct reservation_object_list *fobj;
+	struct fence *fence;
+	int i;
+
+	fobj = reservation_object_get_list(bo->resv);
+	fence = reservation_object_get_excl(bo->resv);
+	if (fence && !fence->ops->signaled)
+		fence_enable_sw_signaling(fence);
+
+	for (i = 0; fobj && i < fobj->shared_count; ++i) {
+		fence = rcu_dereference_protected(fobj->shared[i],
+					reservation_object_held(bo->resv));
+
+		if (!fence->ops->signaled)
+			fence_enable_sw_signaling(fence);
+	}
+}
+
 static void ttm_bo_cleanup_refs_or_queue(struct ttm_buffer_object *bo)
 {
 	struct ttm_bo_device *bdev = bo->bdev;
 	struct ttm_bo_global *glob = bo->glob;
-	struct ttm_bo_driver *driver = bdev->driver;
-	void *sync_obj = NULL;
 	int put_count;
 	int ret;
 
@@ -415,9 +433,7 @@ static void ttm_bo_cleanup_refs_or_queue(struct ttm_buffer_object *bo)
 	ret = __ttm_bo_reserve(bo, false, true, false, NULL);
 
 	if (!ret) {
-		(void) ttm_bo_wait(bo, false, false, true);
-
-		if (!bo->sync_obj) {
+		if (!ttm_bo_wait(bo, false, false, true)) {
 			put_count = ttm_bo_del_from_lru(bo);
 
 			spin_unlock(&glob->lru_lock);
@@ -426,8 +442,8 @@ static void ttm_bo_cleanup_refs_or_queue(struct ttm_buffer_object *bo)
 			ttm_bo_list_ref_sub(bo, put_count, true);
 
 			return;
-		}
-		sync_obj = driver->sync_obj_ref(bo->sync_obj);
+		} else
+			ttm_bo_flush_all_fences(bo);
 
 		/*
 		 * Make NO_EVICT bos immediately available to
@@ -446,14 +462,70 @@ static void ttm_bo_cleanup_refs_or_queue(struct ttm_buffer_object *bo)
 	list_add_tail(&bo->ddestroy, &bdev->ddestroy);
 	spin_unlock(&glob->lru_lock);
 
-	if (sync_obj) {
-		driver->sync_obj_flush(sync_obj);
-		driver->sync_obj_unref(&sync_obj);
-	}
 	schedule_delayed_work(&bdev->wq,
 			      ((HZ / 100) < 1) ? 1 : HZ / 100);
 }
 
+static int ttm_bo_unreserve_and_wait(struct ttm_buffer_object *bo,
+				     bool interruptible)
+{
+	struct ttm_bo_global *glob = bo->glob;
+	struct reservation_object_list *fobj;
+	struct fence *excl = NULL;
+	struct fence **shared = NULL;
+	u32 shared_count = 0, i;
+	int ret = 0;
+
+	fobj = reservation_object_get_list(bo->resv);
+	if (fobj && fobj->shared_count) {
+		shared = kmalloc(sizeof(*shared) * fobj->shared_count,
+				 GFP_KERNEL);
+
+		if (!shared) {
+			ret = -ENOMEM;
+			__ttm_bo_unreserve(bo);
+			spin_unlock(&glob->lru_lock);
+			return ret;
+		}
+
+		for (i = 0; i < fobj->shared_count; ++i) {
+			if (!fence_is_signaled(fobj->shared[i])) {
+				fence_get(fobj->shared[i]);
+				shared[shared_count++] = fobj->shared[i];
+			}
+		}
+		if (!shared_count) {
+			kfree(shared);
+			shared = NULL;
+		}
+	}
+
+	excl = reservation_object_get_excl(bo->resv);
+	if (excl && !fence_is_signaled(excl))
+		fence_get(excl);
+	else
+		excl = NULL;
+
+	__ttm_bo_unreserve(bo);
+	spin_unlock(&glob->lru_lock);
+
+	if (excl) {
+		ret = fence_wait(excl, interruptible);
+		fence_put(excl);
+	}
+
+	if (shared_count > 0) {
+		for (i = 0; i < shared_count; ++i) {
+			if (!ret)
+				ret = fence_wait(shared[i], interruptible);
+			fence_put(shared[i]);
+		}
+		kfree(shared);
+	}
+
+	return ret;
+}
+
 /**
  * function ttm_bo_cleanup_refs_and_unlock
  * If bo idle, remove from delayed- and lru lists, and unref.
@@ -470,8 +542,6 @@ static int ttm_bo_cleanup_refs_and_unlock(struct ttm_buffer_object *bo,
 					  bool interruptible,
 					  bool no_wait_gpu)
 {
-	struct ttm_bo_device *bdev = bo->bdev;
-	struct ttm_bo_driver *driver = bdev->driver;
 	struct ttm_bo_global *glob = bo->glob;
 	int put_count;
 	int ret;
@@ -479,20 +549,7 @@ static int ttm_bo_cleanup_refs_and_unlock(struct ttm_buffer_object *bo,
 	ret = ttm_bo_wait(bo, false, false, true);
 
 	if (ret && !no_wait_gpu) {
-		void *sync_obj;
-
-		/*
-		 * Take a reference to the fence and unreserve,
-		 * at this point the buffer should be dead, so
-		 * no new sync objects can be attached.
-		 */
-		sync_obj = driver->sync_obj_ref(bo->sync_obj);
-
-		__ttm_bo_unreserve(bo);
-		spin_unlock(&glob->lru_lock);
-
-		ret = driver->sync_obj_wait(sync_obj, false, interruptible);
-		driver->sync_obj_unref(&sync_obj);
+		ret = ttm_bo_unreserve_and_wait(bo, interruptible);
 		if (ret)
 			return ret;
 
@@ -1513,41 +1570,51 @@ void ttm_bo_unmap_virtual(struct ttm_buffer_object *bo)
 
 EXPORT_SYMBOL(ttm_bo_unmap_virtual);
 
-
 int ttm_bo_wait(struct ttm_buffer_object *bo,
 		bool lazy, bool interruptible, bool no_wait)
 {
-	struct ttm_bo_driver *driver = bo->bdev->driver;
-	void *sync_obj;
-	int ret = 0;
-
-	lockdep_assert_held(&bo->resv->lock.base);
+	struct reservation_object_list *fobj;
+	struct reservation_object *resv;
+	struct fence *excl;
+	long timeout = 15 * HZ;
+	int i;
 
-	if (likely(bo->sync_obj == NULL))
-		return 0;
+	resv = bo->resv;
+	fobj = reservation_object_get_list(resv);
+	excl = reservation_object_get_excl(resv);
+	if (excl) {
+		if (!fence_is_signaled(excl)) {
+			if (no_wait)
+				return -EBUSY;
 
-	if (bo->sync_obj) {
-		if (driver->sync_obj_signaled(bo->sync_obj)) {
-			driver->sync_obj_unref(&bo->sync_obj);
-			clear_bit(TTM_BO_PRIV_FLAG_MOVING, &bo->priv_flags);
-			return 0;
+			timeout = fence_wait_timeout(excl,
+						     interruptible, timeout);
 		}
+	}
 
-		if (no_wait)
-			return -EBUSY;
+	for (i = 0; fobj && timeout > 0 && i < fobj->shared_count; ++i) {
+		struct fence *fence;
+		fence = rcu_dereference_protected(fobj->shared[i],
+						reservation_object_held(resv));
 
-		sync_obj = driver->sync_obj_ref(bo->sync_obj);
-		ret = driver->sync_obj_wait(sync_obj,
-					    lazy, interruptible);
+		if (!fence_is_signaled(fence)) {
+			if (no_wait)
+				return -EBUSY;
 
-		if (likely(ret == 0)) {
-			clear_bit(TTM_BO_PRIV_FLAG_MOVING,
-				  &bo->priv_flags);
-			driver->sync_obj_unref(&bo->sync_obj);
+			timeout = fence_wait_timeout(fence,
+						     interruptible, timeout);
 		}
-		driver->sync_obj_unref(&sync_obj);
 	}
-	return ret;
+
+	if (timeout < 0)
+		return timeout;
+
+	if (timeout == 0)
+		return -EBUSY;
+
+	reservation_object_add_excl_fence(resv, NULL);
+	clear_bit(TTM_BO_PRIV_FLAG_MOVING, &bo->priv_flags);
+	return 0;
 }
 EXPORT_SYMBOL(ttm_bo_wait);
 
diff --git a/drivers/gpu/drm/ttm/ttm_bo_util.c b/drivers/gpu/drm/ttm/ttm_bo_util.c
index 495aebf0f9c3..824af90cbe31 100644
--- a/drivers/gpu/drm/ttm/ttm_bo_util.c
+++ b/drivers/gpu/drm/ttm/ttm_bo_util.c
@@ -37,6 +37,7 @@
 #include <linux/slab.h>
 #include <linux/vmalloc.h>
 #include <linux/module.h>
+#include <linux/reservation.h>
 
 void ttm_bo_free_old_node(struct ttm_buffer_object *bo)
 {
@@ -444,8 +445,6 @@ static int ttm_buffer_object_transfer(struct ttm_buffer_object *bo,
 				      struct ttm_buffer_object **new_obj)
 {
 	struct ttm_buffer_object *fbo;
-	struct ttm_bo_device *bdev = bo->bdev;
-	struct ttm_bo_driver *driver = bdev->driver;
 	int ret;
 
 	fbo = kmalloc(sizeof(*fbo), GFP_KERNEL);
@@ -466,10 +465,6 @@ static int ttm_buffer_object_transfer(struct ttm_buffer_object *bo,
 	drm_vma_node_reset(&fbo->vma_node);
 	atomic_set(&fbo->cpu_writers, 0);
 
-	if (bo->sync_obj)
-		fbo->sync_obj = driver->sync_obj_ref(bo->sync_obj);
-	else
-		fbo->sync_obj = NULL;
 	kref_init(&fbo->list_kref);
 	kref_init(&fbo->kref);
 	fbo->destroy = &ttm_transfered_destroy;
@@ -642,28 +637,20 @@ void ttm_bo_kunmap(struct ttm_bo_kmap_obj *map)
 EXPORT_SYMBOL(ttm_bo_kunmap);
 
 int ttm_bo_move_accel_cleanup(struct ttm_buffer_object *bo,
-			      void *sync_obj,
+			      struct fence *fence,
 			      bool evict,
 			      bool no_wait_gpu,
 			      struct ttm_mem_reg *new_mem)
 {
 	struct ttm_bo_device *bdev = bo->bdev;
-	struct ttm_bo_driver *driver = bdev->driver;
 	struct ttm_mem_type_manager *man = &bdev->man[new_mem->mem_type];
 	struct ttm_mem_reg *old_mem = &bo->mem;
 	int ret;
 	struct ttm_buffer_object *ghost_obj;
-	void *tmp_obj = NULL;
 
-	if (bo->sync_obj) {
-		tmp_obj = bo->sync_obj;
-		bo->sync_obj = NULL;
-	}
-	bo->sync_obj = driver->sync_obj_ref(sync_obj);
+	reservation_object_add_excl_fence(bo->resv, fence);
 	if (evict) {
 		ret = ttm_bo_wait(bo, false, false, false);
-		if (tmp_obj)
-			driver->sync_obj_unref(&tmp_obj);
 		if (ret)
 			return ret;
 
@@ -684,13 +671,13 @@ int ttm_bo_move_accel_cleanup(struct ttm_buffer_object *bo,
 		 */
 
 		set_bit(TTM_BO_PRIV_FLAG_MOVING, &bo->priv_flags);
-		if (tmp_obj)
-			driver->sync_obj_unref(&tmp_obj);
 
 		ret = ttm_buffer_object_transfer(bo, &ghost_obj);
 		if (ret)
 			return ret;
 
+		reservation_object_add_excl_fence(ghost_obj->resv, fence);
+
 		/**
 		 * If we're not moving to fixed memory, the TTM object
 		 * needs to stay alive. Otherwhise hang it on the ghost
diff --git a/drivers/gpu/drm/ttm/ttm_execbuf_util.c b/drivers/gpu/drm/ttm/ttm_execbuf_util.c
index 108730e9147b..adafc0f8ec06 100644
--- a/drivers/gpu/drm/ttm/ttm_execbuf_util.c
+++ b/drivers/gpu/drm/ttm/ttm_execbuf_util.c
@@ -163,7 +163,7 @@ int ttm_eu_reserve_buffers(struct ww_acquire_ctx *ticket,
 EXPORT_SYMBOL(ttm_eu_reserve_buffers);
 
 void ttm_eu_fence_buffer_objects(struct ww_acquire_ctx *ticket,
-				 struct list_head *list, void *sync_obj)
+				 struct list_head *list, struct fence *fence)
 {
 	struct ttm_validate_buffer *entry;
 	struct ttm_buffer_object *bo;
@@ -183,18 +183,12 @@ void ttm_eu_fence_buffer_objects(struct ww_acquire_ctx *ticket,
 
 	list_for_each_entry(entry, list, head) {
 		bo = entry->bo;
-		entry->old_sync_obj = bo->sync_obj;
-		bo->sync_obj = driver->sync_obj_ref(sync_obj);
+		reservation_object_add_excl_fence(bo->resv, fence);
 		ttm_bo_add_to_lru(bo);
 		__ttm_bo_unreserve(bo);
 	}
 	spin_unlock(&glob->lru_lock);
 	if (ticket)
 		ww_acquire_fini(ticket);
-
-	list_for_each_entry(entry, list, head) {
-		if (entry->old_sync_obj)
-			driver->sync_obj_unref(&entry->old_sync_obj);
-	}
 }
 EXPORT_SYMBOL(ttm_eu_fence_buffer_objects);
diff --git a/drivers/gpu/drm/vmwgfx/vmwgfx_buffer.c b/drivers/gpu/drm/vmwgfx/vmwgfx_buffer.c
index f15718cc631d..656c88485e14 100644
--- a/drivers/gpu/drm/vmwgfx/vmwgfx_buffer.c
+++ b/drivers/gpu/drm/vmwgfx/vmwgfx_buffer.c
@@ -768,41 +768,6 @@ static int vmw_ttm_fault_reserve_notify(struct ttm_buffer_object *bo)
 }
 
 /**
- * FIXME: We're using the old vmware polling method to sync.
- * Do this with fences instead.
- */
-
-static void *vmw_sync_obj_ref(void *sync_obj)
-{
-
-	return (void *)
-		vmw_fence_obj_reference((struct vmw_fence_obj *) sync_obj);
-}
-
-static void vmw_sync_obj_unref(void **sync_obj)
-{
-	vmw_fence_obj_unreference((struct vmw_fence_obj **) sync_obj);
-}
-
-static int vmw_sync_obj_flush(void *sync_obj)
-{
-	vmw_fence_obj_flush((struct vmw_fence_obj *) sync_obj);
-	return 0;
-}
-
-static bool vmw_sync_obj_signaled(void *sync_obj)
-{
-	return vmw_fence_obj_signaled((struct vmw_fence_obj *) sync_obj);
-}
-
-static int vmw_sync_obj_wait(void *sync_obj, bool lazy, bool interruptible)
-{
-	return vmw_fence_obj_wait((struct vmw_fence_obj *) sync_obj,
-				  lazy, interruptible,
-				  VMW_FENCE_WAIT_TIMEOUT);
-}
-
-/**
  * vmw_move_notify - TTM move_notify_callback
  *
  * @bo:             The TTM buffer object about to move.
@@ -839,11 +804,6 @@ struct ttm_bo_driver vmw_bo_driver = {
 	.evict_flags = vmw_evict_flags,
 	.move = NULL,
 	.verify_access = vmw_verify_access,
-	.sync_obj_signaled = vmw_sync_obj_signaled,
-	.sync_obj_wait = vmw_sync_obj_wait,
-	.sync_obj_flush = vmw_sync_obj_flush,
-	.sync_obj_unref = vmw_sync_obj_unref,
-	.sync_obj_ref = vmw_sync_obj_ref,
 	.move_notify = vmw_move_notify,
 	.swap_notify = vmw_swap_notify,
 	.fault_reserve_notify = &vmw_ttm_fault_reserve_notify,
diff --git a/drivers/gpu/drm/vmwgfx/vmwgfx_resource.c b/drivers/gpu/drm/vmwgfx/vmwgfx_resource.c
index 294a0a3e48cb..f0a689ba9685 100644
--- a/drivers/gpu/drm/vmwgfx/vmwgfx_resource.c
+++ b/drivers/gpu/drm/vmwgfx/vmwgfx_resource.c
@@ -1419,22 +1419,16 @@ void vmw_fence_single_bo(struct ttm_buffer_object *bo,
 			 struct vmw_fence_obj *fence)
 {
 	struct ttm_bo_device *bdev = bo->bdev;
-	struct vmw_fence_obj *old_fence_obj;
+
 	struct vmw_private *dev_priv =
 		container_of(bdev, struct vmw_private, bdev);
 
 	if (fence == NULL) {
 		vmw_execbuf_fence_commands(NULL, dev_priv, &fence, NULL);
+		reservation_object_add_excl_fence(bo->resv, &fence->base);
+		fence_put(&fence->base);
 	} else
-		vmw_fence_obj_reference(fence);
-
-	reservation_object_add_excl_fence(bo->resv, &fence->base);
-
-	old_fence_obj = bo->sync_obj;
-	bo->sync_obj = fence;
-
-	if (old_fence_obj)
-		vmw_fence_obj_unreference(&old_fence_obj);
+		reservation_object_add_excl_fence(bo->resv, &fence->base);
 }
 
 /**
diff --git a/include/drm/ttm/ttm_bo_api.h b/include/drm/ttm/ttm_bo_api.h
index 67df9d7c06cc..3b630f3153d0 100644
--- a/include/drm/ttm/ttm_bo_api.h
+++ b/include/drm/ttm/ttm_bo_api.h
@@ -163,7 +163,6 @@ struct ttm_tt;
  * @lru: List head for the lru list.
  * @ddestroy: List head for the delayed destroy list.
  * @swap: List head for swap LRU list.
- * @sync_obj: Pointer to a synchronization object.
  * @priv_flags: Flags describing buffer object internal state.
  * @vma_node: Address space manager node.
  * @offset: The current GPU offset, which can have different meanings
@@ -230,7 +229,6 @@ struct ttm_buffer_object {
 	 * Members protected by a bo reservation.
 	 */
 
-	void *sync_obj;
 	unsigned long priv_flags;
 
 	struct drm_vma_offset_node vma_node;
diff --git a/include/drm/ttm/ttm_bo_driver.h b/include/drm/ttm/ttm_bo_driver.h
index ff725df48595..9792cd047b8c 100644
--- a/include/drm/ttm/ttm_bo_driver.h
+++ b/include/drm/ttm/ttm_bo_driver.h
@@ -311,11 +311,6 @@ struct ttm_mem_type_manager {
  * @move: Callback for a driver to hook in accelerated functions to
  * move a buffer.
  * If set to NULL, a potentially slow memcpy() move is used.
- * @sync_obj_signaled: See ttm_fence_api.h
- * @sync_obj_wait: See ttm_fence_api.h
- * @sync_obj_flush: See ttm_fence_api.h
- * @sync_obj_unref: See ttm_fence_api.h
- * @sync_obj_ref: See ttm_fence_api.h
  */
 
 struct ttm_bo_driver {
@@ -417,23 +412,6 @@ struct ttm_bo_driver {
 	int (*verify_access) (struct ttm_buffer_object *bo,
 			      struct file *filp);
 
-	/**
-	 * In case a driver writer dislikes the TTM fence objects,
-	 * the driver writer can replace those with sync objects of
-	 * his / her own. If it turns out that no driver writer is
-	 * using these. I suggest we remove these hooks and plug in
-	 * fences directly. The bo driver needs the following functionality:
-	 * See the corresponding functions in the fence object API
-	 * documentation.
-	 */
-
-	bool (*sync_obj_signaled) (void *sync_obj);
-	int (*sync_obj_wait) (void *sync_obj,
-			      bool lazy, bool interruptible);
-	int (*sync_obj_flush) (void *sync_obj);
-	void (*sync_obj_unref) (void **sync_obj);
-	void *(*sync_obj_ref) (void *sync_obj);
-
 	/* hook to notify driver about a driver move so it
 	 * can do tiling things */
 	void (*move_notify)(struct ttm_buffer_object *bo,
@@ -1021,7 +999,7 @@ extern void ttm_bo_free_old_node(struct ttm_buffer_object *bo);
  * ttm_bo_move_accel_cleanup.
  *
  * @bo: A pointer to a struct ttm_buffer_object.
- * @sync_obj: A sync object that signals when moving is complete.
+ * @fence: A fence object that signals when moving is complete.
  * @evict: This is an evict move. Don't return until the buffer is idle.
  * @no_wait_gpu: Return immediately if the GPU is busy.
  * @new_mem: struct ttm_mem_reg indicating where to move.
@@ -1035,7 +1013,7 @@ extern void ttm_bo_free_old_node(struct ttm_buffer_object *bo);
  */
 
 extern int ttm_bo_move_accel_cleanup(struct ttm_buffer_object *bo,
-				     void *sync_obj,
+				     struct fence *fence,
 				     bool evict, bool no_wait_gpu,
 				     struct ttm_mem_reg *new_mem);
 /**
diff --git a/include/drm/ttm/ttm_execbuf_util.h b/include/drm/ttm/ttm_execbuf_util.h
index 8490cb8ee0d8..ff11a424f752 100644
--- a/include/drm/ttm/ttm_execbuf_util.h
+++ b/include/drm/ttm/ttm_execbuf_util.h
@@ -39,16 +39,11 @@
  *
  * @head:           list head for thread-private list.
  * @bo:             refcounted buffer object pointer.
- * @reserved:       Indicates whether @bo has been reserved for validation.
- * @removed:        Indicates whether @bo has been removed from lru lists.
- * @put_count:      Number of outstanding references on bo::list_kref.
- * @old_sync_obj:   Pointer to a sync object about to be unreferenced
  */
 
 struct ttm_validate_buffer {
 	struct list_head head;
 	struct ttm_buffer_object *bo;
-	void *old_sync_obj;
 };
 
 /**
@@ -100,7 +95,7 @@ extern int ttm_eu_reserve_buffers(struct ww_acquire_ctx *ticket,
  *
  * @ticket:      ww_acquire_ctx from reserve call
  * @list:        thread private list of ttm_validate_buffer structs.
- * @sync_obj:    The new sync object for the buffers.
+ * @fence:       The new exclusive fence for the buffers.
  *
  * This function should be called when command submission is complete, and
  * it will add a new sync object to bos pointed to by entries on @list.
@@ -109,6 +104,7 @@ extern int ttm_eu_reserve_buffers(struct ww_acquire_ctx *ticket,
  */
 
 extern void ttm_eu_fence_buffer_objects(struct ww_acquire_ctx *ticket,
-					struct list_head *list, void *sync_obj);
+					struct list_head *list,
+					struct fence *fence);
 
 #endif

[PATCH 16/19] drm/nouveau: use rcu in nouveau_gem_ioctl_cpu_prep

[Maarten Lankhorst]

With the conversion to the reservation api this should be safe.

Signed-off-by: Maarten Lankhorst <maarten.lankhorst@canonical.com>
---
 drivers/gpu/drm/nouveau/nouveau_gem.c |   28 ++++++++++++----------------
 1 file changed, 12 insertions(+), 16 deletions(-)

diff --git a/drivers/gpu/drm/nouveau/nouveau_gem.c b/drivers/gpu/drm/nouveau/nouveau_gem.c
index 0a1fee7eb211..b24a6fe6c601 100644
--- a/drivers/gpu/drm/nouveau/nouveau_gem.c
+++ b/drivers/gpu/drm/nouveau/nouveau_gem.c
@@ -863,33 +863,29 @@ nouveau_gem_ioctl_cpu_prep(struct drm_device *dev, void *data,
 	struct drm_gem_object *gem;
 	struct nouveau_bo *nvbo;
 	bool no_wait = !!(req->flags & NOUVEAU_GEM_CPU_PREP_NOWAIT);
+	bool write = !!(req->flags & NOUVEAU_GEM_CPU_PREP_WRITE);
 	int ret;
-	struct nouveau_fence *fence = NULL;
 
 	gem = drm_gem_object_lookup(dev, file_priv, req->handle);
 	if (!gem)
 		return -ENOENT;
 	nvbo = nouveau_gem_object(gem);
 
-	ret = ttm_bo_reserve(&nvbo->bo, true, false, false, 0);
-	if (!ret) {
-		ret = ttm_bo_wait(&nvbo->bo, true, true, true);
-		if (!no_wait && ret) {
-			struct fence *excl;
-
-			excl = reservation_object_get_excl(nvbo->bo.resv);
-			fence = nouveau_fence_ref((struct nouveau_fence *)excl);
-		}
+	if (no_wait)
+		ret = reservation_object_test_signaled_rcu(nvbo->bo.resv, write) ? 0 : -EBUSY;
+	else {
+		long lret;
 
-		ttm_bo_unreserve(&nvbo->bo);
+		lret = reservation_object_wait_timeout_rcu(nvbo->bo.resv, write, true, 30 * HZ);
+		if (!lret)
+			ret = -EBUSY;
+		else if (lret > 0)
+			ret = 0;
+		else
+			ret = lret;
 	}
 	drm_gem_object_unreference_unlocked(gem);
 
-	if (fence) {
-		ret = nouveau_fence_wait(fence, true, no_wait);
-		nouveau_fence_unref(&fence);
-	}
-
 	return ret;
 }
 

[PATCH 17/19] drm/radeon: use rcu waits in some ioctls

[Maarten Lankhorst]

Signed-off-by: Maarten Lankhorst <maarten.lankhorst@canonical.com>
---
 drivers/gpu/drm/radeon/radeon_gem.c |   19 ++++++++++++++-----
 1 file changed, 14 insertions(+), 5 deletions(-)

diff --git a/drivers/gpu/drm/radeon/radeon_gem.c b/drivers/gpu/drm/radeon/radeon_gem.c
index d09650c1d720..7ba883843668 100644
--- a/drivers/gpu/drm/radeon/radeon_gem.c
+++ b/drivers/gpu/drm/radeon/radeon_gem.c
@@ -107,9 +107,12 @@ static int radeon_gem_set_domain(struct drm_gem_object *gobj,
 	}
 	if (domain == RADEON_GEM_DOMAIN_CPU) {
 		/* Asking for cpu access wait for object idle */
-		r = radeon_bo_wait(robj, NULL, false);
-		if (r) {
-			printk(KERN_ERR "Failed to wait for object !\n");
+		r = reservation_object_wait_timeout_rcu(robj->tbo.resv, true, true, 30 * HZ);
+		if (!r)
+			r = -EBUSY;
+
+		if (r < 0 && r != -EINTR) {
+			printk(KERN_ERR "Failed to wait for object: %i\n", r);
 			return r;
 		}
 	}
@@ -357,14 +360,20 @@ int radeon_gem_wait_idle_ioctl(struct drm_device *dev, void *data,
 	struct drm_radeon_gem_wait_idle *args = data;
 	struct drm_gem_object *gobj;
 	struct radeon_bo *robj;
-	int r;
+	int r = 0;
+	long ret;
 
 	gobj = drm_gem_object_lookup(dev, filp, args->handle);
 	if (gobj == NULL) {
 		return -ENOENT;
 	}
 	robj = gem_to_radeon_bo(gobj);
-	r = radeon_bo_wait(robj, NULL, false);
+	ret = reservation_object_wait_timeout_rcu(robj->tbo.resv, true, true, 30 * HZ);
+	if (ret == 0)
+		r = -EBUSY;
+	else if (ret < 0)
+		r = ret;
+
 	/* callback hw specific functions if any */
 	if (rdev->asic->ioctl_wait_idle)
 		robj->rdev->asic->ioctl_wait_idle(rdev, robj);

[PATCH 18/19] drm/vmwgfx: use rcu in vmw_user_dmabuf_synccpu_grab

[Maarten Lankhorst]

Signed-off-by: Maarten Lankhorst <maarten.lankhorst@canonical.com>
---
 drivers/gpu/drm/vmwgfx/vmwgfx_resource.c |   17 ++++++++++-------
 1 file changed, 10 insertions(+), 7 deletions(-)

diff --git a/drivers/gpu/drm/vmwgfx/vmwgfx_resource.c b/drivers/gpu/drm/vmwgfx/vmwgfx_resource.c
index f0a689ba9685..a43930162eb6 100644
--- a/drivers/gpu/drm/vmwgfx/vmwgfx_resource.c
+++ b/drivers/gpu/drm/vmwgfx/vmwgfx_resource.c
@@ -567,13 +567,16 @@ static int vmw_user_dmabuf_synccpu_grab(struct vmw_user_dma_buffer *user_bo,
 	int ret;
 
 	if (flags & drm_vmw_synccpu_allow_cs) {
-		ret = ttm_bo_reserve(bo, true, !!(flags & drm_vmw_synccpu_dontblock), false, 0);
-		if (!ret) {
-			ret = ttm_bo_wait(bo, false, true,
-					  !!(flags & drm_vmw_synccpu_dontblock));
-			ttm_bo_unreserve(bo);
-		}
-		return ret;
+		long lret;
+		if (flags & drm_vmw_synccpu_dontblock)
+			return reservation_object_test_signaled_rcu(bo->resv, true) ? 0 : -EBUSY;
+
+		lret = reservation_object_wait_timeout_rcu(bo->resv, true, true, MAX_SCHEDULE_TIMEOUT);
+		if (!lret)
+			return -EBUSY;
+		else if (lret < 0)
+			return lret;
+		return 0;
 	}
 
 	ret = ttm_bo_synccpu_write_grab

[PATCH 19/19] drm/ttm: use rcu in core ttm

[Maarten Lankhorst]

Signed-off-by: Maarten Lankhorst <maarten.lankhorst@canonical.com>
---
 drivers/gpu/drm/ttm/ttm_bo.c |   76 +++++++-----------------------------------
 1 file changed, 13 insertions(+), 63 deletions(-)

diff --git a/drivers/gpu/drm/ttm/ttm_bo.c b/drivers/gpu/drm/ttm/ttm_bo.c
index 9d36cffad3b7..d119ae4419d4 100644
--- a/drivers/gpu/drm/ttm/ttm_bo.c
+++ b/drivers/gpu/drm/ttm/ttm_bo.c
@@ -466,66 +466,6 @@ static void ttm_bo_cleanup_refs_or_queue(struct ttm_buffer_object *bo)
 			      ((HZ / 100) < 1) ? 1 : HZ / 100);
 }
 
-static int ttm_bo_unreserve_and_wait(struct ttm_buffer_object *bo,
-				     bool interruptible)
-{
-	struct ttm_bo_global *glob = bo->glob;
-	struct reservation_object_list *fobj;
-	struct fence *excl = NULL;
-	struct fence **shared = NULL;
-	u32 shared_count = 0, i;
-	int ret = 0;
-
-	fobj = reservation_object_get_list(bo->resv);
-	if (fobj && fobj->shared_count) {
-		shared = kmalloc(sizeof(*shared) * fobj->shared_count,
-				 GFP_KERNEL);
-
-		if (!shared) {
-			ret = -ENOMEM;
-			__ttm_bo_unreserve(bo);
-			spin_unlock(&glob->lru_lock);
-			return ret;
-		}
-
-		for (i = 0; i < fobj->shared_count; ++i) {
-			if (!fence_is_signaled(fobj->shared[i])) {
-				fence_get(fobj->shared[i]);
-				shared[shared_count++] = fobj->shared[i];
-			}
-		}
-		if (!shared_count) {
-			kfree(shared);
-			shared = NULL;
-		}
-	}
-
-	excl = reservation_object_get_excl(bo->resv);
-	if (excl && !fence_is_signaled(excl))
-		fence_get(excl);
-	else
-		excl = NULL;
-
-	__ttm_bo_unreserve(bo);
-	spin_unlock(&glob->lru_lock);
-
-	if (excl) {
-		ret = fence_wait(excl, interruptible);
-		fence_put(excl);
-	}
-
-	if (shared_count > 0) {
-		for (i = 0; i < shared_count; ++i) {
-			if (!ret)
-				ret = fence_wait(shared[i], interruptible);
-			fence_put(shared[i]);
-		}
-		kfree(shared);
-	}
-
-	return ret;
-}
-
 /**
  * function ttm_bo_cleanup_refs_and_unlock
  * If bo idle, remove from delayed- and lru lists, and unref.
@@ -549,9 +489,19 @@ static int ttm_bo_cleanup_refs_and_unlock(struct ttm_buffer_object *bo,
 	ret = ttm_bo_wait(bo, false, false, true);
 
 	if (ret && !no_wait_gpu) {
-		ret = ttm_bo_unreserve_and_wait(bo, interruptible);
-		if (ret)
-			return ret;
+		long lret;
+		ww_mutex_unlock(&bo->resv->lock);
+		spin_unlock(&glob->lru_lock);
+
+		lret = reservation_object_wait_timeout_rcu(bo->resv,
+							   true,
+							   interruptible,
+							   30 * HZ);
+
+		if (lret < 0)
+			return lret;
+		else if (lret == 0)
+			return -EBUSY;
 
 		spin_lock(&glob->lru_lock);
 		ret = __ttm_bo_reserve(bo, false, true, false, NULL);