1 files changed, 166 insertions, 167 deletions

diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_sync.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_sync.c
index 181ce39ef5e5..34a92808bbd4 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_sync.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_sync.c
@@ -37,6 +37,8 @@ struct amdgpu_sync_entry {
 	struct fence		*fence;
 };
 
+static struct kmem_cache *amdgpu_sync_slab;
+
 /**
  * amdgpu_sync_create - zero init sync object
  *
@@ -46,26 +48,22 @@ struct amdgpu_sync_entry {
  */
 void amdgpu_sync_create(struct amdgpu_sync *sync)
 {
-	unsigned i;
-
-	for (i = 0; i < AMDGPU_NUM_SYNCS; ++i)
-		sync->semaphores[i] = NULL;
-
-	for (i = 0; i < AMDGPU_MAX_RINGS; ++i)
-		sync->sync_to[i] = NULL;
-
 	hash_init(sync->fences);
 	sync->last_vm_update = NULL;
 }
 
+/**
+ * amdgpu_sync_same_dev - test if fence belong to us
+ *
+ * @adev: amdgpu device to use for the test
+ * @f: fence to test
+ *
+ * Test if the fence was issued by us.
+ */
 static bool amdgpu_sync_same_dev(struct amdgpu_device *adev, struct fence *f)
 {
-	struct amdgpu_fence *a_fence = to_amdgpu_fence(f);
 	struct amd_sched_fence *s_fence = to_amd_sched_fence(f);
 
-	if (a_fence)
-		return a_fence->ring->adev == adev;
-
 	if (s_fence) {
 		struct amdgpu_ring *ring;
 
@@ -76,17 +74,31 @@ static bool amdgpu_sync_same_dev(struct amdgpu_device *adev, struct fence *f)
 	return false;
 }
 
-static bool amdgpu_sync_test_owner(struct fence *f, void *owner)
+/**
+ * amdgpu_sync_get_owner - extract the owner of a fence
+ *
+ * @fence: fence get the owner from
+ *
+ * Extract who originally created the fence.
+ */
+static void *amdgpu_sync_get_owner(struct fence *f)
 {
-	struct amdgpu_fence *a_fence = to_amdgpu_fence(f);
 	struct amd_sched_fence *s_fence = to_amd_sched_fence(f);
+
 	if (s_fence)
-		return s_fence->owner == owner;
-	if (a_fence)
-		return a_fence->owner == owner;
-	return false;
+		return s_fence->owner;
+
+	return AMDGPU_FENCE_OWNER_UNDEFINED;
 }
 
+/**
+ * amdgpu_sync_keep_later - Keep the later fence
+ *
+ * @keep: existing fence to test
+ * @fence: new fence
+ *
+ * Either keep the existing fence or the new one, depending which one is later.
+ */
 static void amdgpu_sync_keep_later(struct fence **keep, struct fence *fence)
 {
 	if (*keep && fence_is_later(*keep, fence))
@@ -97,6 +109,29 @@ static void amdgpu_sync_keep_later(struct fence **keep, struct fence *fence)
 }
 
 /**
+ * amdgpu_sync_add_later - add the fence to the hash
+ *
+ * @sync: sync object to add the fence to
+ * @f: fence to add
+ *
+ * Tries to add the fence to an existing hash entry. Returns true when an entry
+ * was found, false otherwise.
+ */
+static bool amdgpu_sync_add_later(struct amdgpu_sync *sync, struct fence *f)
+{
+	struct amdgpu_sync_entry *e;
+
+	hash_for_each_possible(sync->fences, e, node, f->context) {
+		if (unlikely(e->fence->context != f->context))
+			continue;
+
+		amdgpu_sync_keep_later(&e->fence, f);
+		return true;
+	}
+	return false;
+}
+
+/**
  * amdgpu_sync_fence - remember to sync to this fence
  *
  * @sync: sync object to add fence to
@@ -107,59 +142,34 @@ int amdgpu_sync_fence(struct amdgpu_device *adev, struct amdgpu_sync *sync,
 		      struct fence *f)
 {
 	struct amdgpu_sync_entry *e;
-	struct amdgpu_fence *fence;
 
 	if (!f)
 		return 0;
 
 	if (amdgpu_sync_same_dev(adev, f) &&
-	    amdgpu_sync_test_owner(f, AMDGPU_FENCE_OWNER_VM))
+	    amdgpu_sync_get_owner(f) == AMDGPU_FENCE_OWNER_VM)
 		amdgpu_sync_keep_later(&sync->last_vm_update, f);
 
-	fence = to_amdgpu_fence(f);
-	if (!fence || fence->ring->adev != adev) {
-		hash_for_each_possible(sync->fences, e, node, f->context) {
-			if (unlikely(e->fence->context != f->context))
-				continue;
-
-			amdgpu_sync_keep_later(&e->fence, f);
-			return 0;
-		}
-
-		e = kmalloc(sizeof(struct amdgpu_sync_entry), GFP_KERNEL);
-		if (!e)
-			return -ENOMEM;
-
-		hash_add(sync->fences, &e->node, f->context);
-		e->fence = fence_get(f);
+	if (amdgpu_sync_add_later(sync, f))
 		return 0;
-	}
 
-	amdgpu_sync_keep_later(&sync->sync_to[fence->ring->idx], f);
+	e = kmem_cache_alloc(amdgpu_sync_slab, GFP_KERNEL);
+	if (!e)
+		return -ENOMEM;
 
+	hash_add(sync->fences, &e->node, f->context);
+	e->fence = fence_get(f);
 	return 0;
 }
 
-static void *amdgpu_sync_get_owner(struct fence *f)
-{
-	struct amdgpu_fence *a_fence = to_amdgpu_fence(f);
-	struct amd_sched_fence *s_fence = to_amd_sched_fence(f);
-
-	if (s_fence)
-		return s_fence->owner;
-	else if (a_fence)
-		return a_fence->owner;
-	return AMDGPU_FENCE_OWNER_UNDEFINED;
-}
-
 /**
- * amdgpu_sync_resv - use the semaphores to sync to a reservation object
+ * amdgpu_sync_resv - sync to a reservation object
  *
  * @sync: sync object to add fences from reservation object to
  * @resv: reservation object with embedded fence
  * @shared: true if we should only sync to the exclusive fence
  *
- * Sync to the fence using the semaphore objects
+ * Sync to the fence
  */
 int amdgpu_sync_resv(struct amdgpu_device *adev,
 		     struct amdgpu_sync *sync,
@@ -212,144 +222,117 @@ int amdgpu_sync_resv(struct amdgpu_device *adev,
 	return r;
 }
 
-struct fence *amdgpu_sync_get_fence(struct amdgpu_sync *sync)
+/**
+ * amdgpu_sync_is_idle - test if all fences are signaled
+ *
+ * @sync: the sync object
+ *
+ * Returns true if all fences in the sync object are signaled.
+ */
+bool amdgpu_sync_is_idle(struct amdgpu_sync *sync)
 {
 	struct amdgpu_sync_entry *e;
 	struct hlist_node *tmp;
-	struct fence *f;
 	int i;
 
 	hash_for_each_safe(sync->fences, i, tmp, e, node) {
+		struct fence *f = e->fence;
 
-		f = e->fence;
-
-		hash_del(&e->node);
-		kfree(e);
-
-		if (!fence_is_signaled(f))
-			return f;
+		if (fence_is_signaled(f)) {
+			hash_del(&e->node);
+			fence_put(f);
+			kmem_cache_free(amdgpu_sync_slab, e);
+			continue;
+		}
 
-		fence_put(f);
+		return false;
 	}
-	return NULL;
+
+	return true;
 }
 
-int amdgpu_sync_wait(struct amdgpu_sync *sync)
+/**
+ * amdgpu_sync_cycle_fences - move fences from one sync object into another
+ *
+ * @dst: the destination sync object
+ * @src: the source sync object
+ * @fence: fence to add to source
+ *
+ * Remove all fences from source and put them into destination and add
+ * fence as new one into source.
+ */
+int amdgpu_sync_cycle_fences(struct amdgpu_sync *dst, struct amdgpu_sync *src,
+			     struct fence *fence)
 {
-	struct amdgpu_sync_entry *e;
+	struct amdgpu_sync_entry *e, *newone;
 	struct hlist_node *tmp;
-	int i, r;
+	int i;
 
-	hash_for_each_safe(sync->fences, i, tmp, e, node) {
-		r = fence_wait(e->fence, false);
-		if (r)
-			return r;
+	/* Allocate the new entry before moving the old ones */
+	newone = kmem_cache_alloc(amdgpu_sync_slab, GFP_KERNEL);
+	if (!newone)
+		return -ENOMEM;
 
-		hash_del(&e->node);
-		fence_put(e->fence);
-		kfree(e);
-	}
+	hash_for_each_safe(src->fences, i, tmp, e, node) {
+		struct fence *f = e->fence;
 
-	if (amdgpu_enable_semaphores)
-		return 0;
+		hash_del(&e->node);
+		if (fence_is_signaled(f)) {
+			fence_put(f);
+			kmem_cache_free(amdgpu_sync_slab, e);
+			continue;
+		}
 
-	for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
-		struct fence *fence = sync->sync_to[i];
-		if (!fence)
+		if (amdgpu_sync_add_later(dst, f)) {
+			kmem_cache_free(amdgpu_sync_slab, e);
 			continue;
+		}
 
-		r = fence_wait(fence, false);
-		if (r)
-			return r;
+		hash_add(dst->fences, &e->node, f->context);
 	}
 
+	hash_add(src->fences, &newone->node, fence->context);
+	newone->fence = fence_get(fence);
+
 	return 0;
 }
 
-/**
- * amdgpu_sync_rings - sync ring to all registered fences
- *
- * @sync: sync object to use
- * @ring: ring that needs sync
- *
- * Ensure that all registered fences are signaled before letting
- * the ring continue. The caller must hold the ring lock.
- */
-int amdgpu_sync_rings(struct amdgpu_sync *sync,
-		      struct amdgpu_ring *ring)
+struct fence *amdgpu_sync_get_fence(struct amdgpu_sync *sync)
 {
-	struct amdgpu_device *adev = ring->adev;
-	unsigned count = 0;
-	int i, r;
-
-	for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
-		struct amdgpu_ring *other = adev->rings[i];
-		struct amdgpu_semaphore *semaphore;
-		struct amdgpu_fence *fence;
-
-		if (!sync->sync_to[i])
-			continue;
+	struct amdgpu_sync_entry *e;
+	struct hlist_node *tmp;
+	struct fence *f;
+	int i;
 
-		fence = to_amdgpu_fence(sync->sync_to[i]);
+	hash_for_each_safe(sync->fences, i, tmp, e, node) {
 
-		/* check if we really need to sync */
-		if (!amdgpu_enable_scheduler &&
-		    !amdgpu_fence_need_sync(fence, ring))
-			continue;
+		f = e->fence;
 
-		/* prevent GPU deadlocks */
-		if (!other->ready) {
-			dev_err(adev->dev, "Syncing to a disabled ring!");
-			return -EINVAL;
-		}
+		hash_del(&e->node);
+		kmem_cache_free(amdgpu_sync_slab, e);
 
-		if (amdgpu_enable_scheduler || !amdgpu_enable_semaphores) {
-			r = fence_wait(sync->sync_to[i], true);
-			if (r)
-				return r;
-			continue;
-		}
+		if (!fence_is_signaled(f))
+			return f;
 
-		if (count >= AMDGPU_NUM_SYNCS) {
-			/* not enough room, wait manually */
-			r = fence_wait(&fence->base, false);
-			if (r)
-				return r;
-			continue;
-		}
-		r = amdgpu_semaphore_create(adev, &semaphore);
-		if (r)
-			return r;
+		fence_put(f);
+	}
+	return NULL;
+}
 
-		sync->semaphores[count++] = semaphore;
+int amdgpu_sync_wait(struct amdgpu_sync *sync)
+{
+	struct amdgpu_sync_entry *e;
+	struct hlist_node *tmp;
+	int i, r;
 
-		/* allocate enough space for sync command */
-		r = amdgpu_ring_alloc(other, 16);
+	hash_for_each_safe(sync->fences, i, tmp, e, node) {
+		r = fence_wait(e->fence, false);
 		if (r)
 			return r;
 
-		/* emit the signal semaphore */
-		if (!amdgpu_semaphore_emit_signal(other, semaphore)) {
-			/* signaling wasn't successful wait manually */
-			amdgpu_ring_undo(other);
-			r = fence_wait(&fence->base, false);
-			if (r)
-				return r;
-			continue;
-		}
-
-		/* we assume caller has already allocated space on waiters ring */
-		if (!amdgpu_semaphore_emit_wait(ring, semaphore)) {
-			/* waiting wasn't successful wait manually */
-			amdgpu_ring_undo(other);
-			r = fence_wait(&fence->base, false);
-			if (r)
-				return r;
-			continue;
-		}
-
-		amdgpu_ring_commit(other);
-		amdgpu_fence_note_sync(fence, ring);
+		hash_del(&e->node);
+		fence_put(e->fence);
+		kmem_cache_free(amdgpu_sync_slab, e);
 	}
 
 	return 0;
@@ -358,15 +341,11 @@ int amdgpu_sync_rings(struct amdgpu_sync *sync,
 /**
  * amdgpu_sync_free - free the sync object
  *
- * @adev: amdgpu_device pointer
  * @sync: sync object to use
- * @fence: fence to use for the free
  *
- * Free the sync object by freeing all semaphores in it.
+ * Free the sync object.
  */
-void amdgpu_sync_free(struct amdgpu_device *adev,
-		      struct amdgpu_sync *sync,
-		      struct fence *fence)
+void amdgpu_sync_free(struct amdgpu_sync *sync)
 {
 	struct amdgpu_sync_entry *e;
 	struct hlist_node *tmp;
@@ -375,14 +354,34 @@ void amdgpu_sync_free(struct amdgpu_device *adev,
 	hash_for_each_safe(sync->fences, i, tmp, e, node) {
 		hash_del(&e->node);
 		fence_put(e->fence);
-		kfree(e);
+		kmem_cache_free(amdgpu_sync_slab, e);
 	}
 
-	for (i = 0; i < AMDGPU_NUM_SYNCS; ++i)
-		amdgpu_semaphore_free(adev, &sync->semaphores[i], fence);
+	fence_put(sync->last_vm_update);
+}
 
-	for (i = 0; i < AMDGPU_MAX_RINGS; ++i)
-		fence_put(sync->sync_to[i]);
+/**
+ * amdgpu_sync_init - init sync object subsystem
+ *
+ * Allocate the slab allocator.
+ */
+int amdgpu_sync_init(void)
+{
+	amdgpu_sync_slab = kmem_cache_create(
+		"amdgpu_sync", sizeof(struct amdgpu_sync_entry), 0,
+		SLAB_HWCACHE_ALIGN, NULL);
+	if (!amdgpu_sync_slab)
+		return -ENOMEM;
 
-	fence_put(sync->last_vm_update);
+	return 0;
+}
+
+/**
+ * amdgpu_sync_fini - fini sync object subsystem
+ *
+ * Free the slab allocator.
+ */
+void amdgpu_sync_fini(void)
+{
+	kmem_cache_destroy(amdgpu_sync_slab);
 }