1 files changed, 223 insertions, 75 deletions

diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_sync.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_sync.c
index a6899180b265..d6ae9974c952 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_sync.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_sync.c
@@ -1,3 +1,4 @@
+// SPDX-License-Identifier: MIT
 /*
  * Copyright 2014 Advanced Micro Devices, Inc.
  * All Rights Reserved.
@@ -28,9 +29,11 @@
  *    Christian König <christian.koenig@amd.com>
  */
 
-#include <drm/drmP.h>
+#include <linux/dma-fence-chain.h>
+
 #include "amdgpu.h"
 #include "amdgpu_trace.h"
+#include "amdgpu_amdkfd.h"
 
 struct amdgpu_sync_entry {
 	struct hlist_node	node;
@@ -49,7 +52,6 @@ static struct kmem_cache *amdgpu_sync_slab;
 void amdgpu_sync_create(struct amdgpu_sync *sync)
 {
 	hash_init(sync->fences);
-	sync->last_vm_update = NULL;
 }
 
 /**
@@ -63,7 +65,7 @@ void amdgpu_sync_create(struct amdgpu_sync *sync)
 static bool amdgpu_sync_same_dev(struct amdgpu_device *adev,
 				 struct dma_fence *f)
 {
-	struct amd_sched_fence *s_fence = to_amd_sched_fence(f);
+	struct drm_sched_fence *s_fence = to_drm_sched_fence(f);
 
 	if (s_fence) {
 		struct amdgpu_ring *ring;
@@ -78,17 +80,26 @@ static bool amdgpu_sync_same_dev(struct amdgpu_device *adev,
 /**
  * amdgpu_sync_get_owner - extract the owner of a fence
  *
- * @fence: fence get the owner from
+ * @f: fence get the owner from
  *
  * Extract who originally created the fence.
  */
 static void *amdgpu_sync_get_owner(struct dma_fence *f)
 {
-	struct amd_sched_fence *s_fence = to_amd_sched_fence(f);
+	struct drm_sched_fence *s_fence;
+	struct amdgpu_amdkfd_fence *kfd_fence;
+
+	if (!f)
+		return AMDGPU_FENCE_OWNER_UNDEFINED;
 
+	s_fence = to_drm_sched_fence(f);
 	if (s_fence)
 		return s_fence->owner;
 
+	kfd_fence = to_amdgpu_amdkfd_fence(f);
+	if (kfd_fence)
+		return AMDGPU_FENCE_OWNER_KFD;
+
 	return AMDGPU_FENCE_OWNER_UNDEFINED;
 }
 
@@ -124,11 +135,16 @@ static bool amdgpu_sync_add_later(struct amdgpu_sync *sync, struct dma_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;
+		if (dma_fence_is_signaled(e->fence)) {
+			dma_fence_put(e->fence);
+			e->fence = dma_fence_get(f);
+			return true;
+		}
 
-		amdgpu_sync_keep_later(&e->fence, f);
-		return true;
+		if (likely(e->fence->context == f->context)) {
+			amdgpu_sync_keep_later(&e->fence, f);
+			return true;
+		}
 	}
 	return false;
 }
@@ -137,25 +153,23 @@ static bool amdgpu_sync_add_later(struct amdgpu_sync *sync, struct dma_fence *f)
  * amdgpu_sync_fence - remember to sync to this fence
  *
  * @sync: sync object to add fence to
- * @fence: fence to sync to
+ * @f: fence to sync to
+ * @flags: memory allocation flags to use when allocating sync entry
  *
+ * Add the fence to the sync object.
  */
-int amdgpu_sync_fence(struct amdgpu_device *adev, struct amdgpu_sync *sync,
-		      struct dma_fence *f)
+int amdgpu_sync_fence(struct amdgpu_sync *sync, struct dma_fence *f,
+		      gfp_t flags)
 {
 	struct amdgpu_sync_entry *e;
 
 	if (!f)
 		return 0;
 
-	if (amdgpu_sync_same_dev(adev, f) &&
-	    amdgpu_sync_get_owner(f) == AMDGPU_FENCE_OWNER_VM)
-		amdgpu_sync_keep_later(&sync->last_vm_update, f);
-
 	if (amdgpu_sync_add_later(sync, f))
 		return 0;
 
-	e = kmem_cache_alloc(amdgpu_sync_slab, GFP_KERNEL);
+	e = kmem_cache_alloc(amdgpu_sync_slab, flags);
 	if (!e)
 		return -ENOMEM;
 
@@ -164,66 +178,132 @@ int amdgpu_sync_fence(struct amdgpu_device *adev, struct amdgpu_sync *sync,
 	return 0;
 }
 
+/* Determine based on the owner and mode if we should sync to a fence or not */
+static bool amdgpu_sync_test_fence(struct amdgpu_device *adev,
+				   enum amdgpu_sync_mode mode,
+				   void *owner, struct dma_fence *f)
+{
+	void *fence_owner = amdgpu_sync_get_owner(f);
+
+	/* Always sync to moves, no matter what */
+	if (fence_owner == AMDGPU_FENCE_OWNER_UNDEFINED)
+		return true;
+
+	/* We only want to trigger KFD eviction fences on
+	 * evict or move jobs. Skip KFD fences otherwise.
+	 */
+	if (fence_owner == AMDGPU_FENCE_OWNER_KFD &&
+	    owner != AMDGPU_FENCE_OWNER_UNDEFINED)
+		return false;
+
+	/* Never sync to VM updates either. */
+	if (fence_owner == AMDGPU_FENCE_OWNER_VM &&
+	    owner != AMDGPU_FENCE_OWNER_UNDEFINED &&
+	    owner != AMDGPU_FENCE_OWNER_KFD)
+		return false;
+
+	/* Ignore fences depending on the sync mode */
+	switch (mode) {
+	case AMDGPU_SYNC_ALWAYS:
+		return true;
+
+	case AMDGPU_SYNC_NE_OWNER:
+		if (amdgpu_sync_same_dev(adev, f) &&
+		    fence_owner == owner)
+			return false;
+		break;
+
+	case AMDGPU_SYNC_EQ_OWNER:
+		if (amdgpu_sync_same_dev(adev, f) &&
+		    fence_owner != owner)
+			return false;
+		break;
+
+	case AMDGPU_SYNC_EXPLICIT:
+		return false;
+	}
+
+	WARN(debug_evictions && fence_owner == AMDGPU_FENCE_OWNER_KFD,
+	     "Adding eviction fence to sync obj");
+	return true;
+}
+
 /**
  * amdgpu_sync_resv - sync to a reservation object
  *
+ * @adev: amdgpu device
  * @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
+ * @mode: how owner affects which fences we sync to
+ * @owner: owner of the planned job submission
  *
  * Sync to the fence
  */
-int amdgpu_sync_resv(struct amdgpu_device *adev,
-		     struct amdgpu_sync *sync,
-		     struct reservation_object *resv,
+int amdgpu_sync_resv(struct amdgpu_device *adev, struct amdgpu_sync *sync,
+		     struct dma_resv *resv, enum amdgpu_sync_mode mode,
 		     void *owner)
 {
-	struct reservation_object_list *flist;
+	struct dma_resv_iter cursor;
 	struct dma_fence *f;
-	void *fence_owner;
-	unsigned i;
-	int r = 0;
+	int r;
 
 	if (resv == NULL)
 		return -EINVAL;
+	/* Implicitly sync only to KERNEL, WRITE and READ */
+	dma_resv_for_each_fence(&cursor, resv, DMA_RESV_USAGE_READ, f) {
+		dma_fence_chain_for_each(f, f) {
+			struct dma_fence *tmp = dma_fence_chain_contained(f);
+
+			if (amdgpu_sync_test_fence(adev, mode, owner, tmp)) {
+				r = amdgpu_sync_fence(sync, f, GFP_KERNEL);
+				dma_fence_put(f);
+				if (r)
+					return r;
+				break;
+			}
+		}
+	}
+	return 0;
+}
 
-	/* always sync to the exclusive fence */
-	f = reservation_object_get_excl(resv);
-	r = amdgpu_sync_fence(adev, sync, f);
+/**
+ * amdgpu_sync_kfd - sync to KFD fences
+ *
+ * @sync: sync object to add KFD fences to
+ * @resv: reservation object with KFD fences
+ *
+ * Extract all KFD fences and add them to the sync object.
+ */
+int amdgpu_sync_kfd(struct amdgpu_sync *sync, struct dma_resv *resv)
+{
+	struct dma_resv_iter cursor;
+	struct dma_fence *f;
+	int r = 0;
 
-	flist = reservation_object_get_list(resv);
-	if (!flist || r)
-		return r;
+	dma_resv_iter_begin(&cursor, resv, DMA_RESV_USAGE_BOOKKEEP);
+	dma_resv_for_each_fence_unlocked(&cursor, f) {
+		void *fence_owner = amdgpu_sync_get_owner(f);
 
-	for (i = 0; i < flist->shared_count; ++i) {
-		f = rcu_dereference_protected(flist->shared[i],
-					      reservation_object_held(resv));
-		if (amdgpu_sync_same_dev(adev, f)) {
-			/* VM updates are only interesting
-			 * for other VM updates and moves.
-			 */
-			fence_owner = amdgpu_sync_get_owner(f);
-			if ((owner != AMDGPU_FENCE_OWNER_UNDEFINED) &&
-			    (fence_owner != AMDGPU_FENCE_OWNER_UNDEFINED) &&
-			    ((owner == AMDGPU_FENCE_OWNER_VM) !=
-			     (fence_owner == AMDGPU_FENCE_OWNER_VM)))
-				continue;
-
-			/* Ignore fence from the same owner as
-			 * long as it isn't undefined.
-			 */
-			if (owner != AMDGPU_FENCE_OWNER_UNDEFINED &&
-			    fence_owner == owner)
-				continue;
-		}
+		if (fence_owner != AMDGPU_FENCE_OWNER_KFD)
+			continue;
 
-		r = amdgpu_sync_fence(adev, sync, f);
+		r = amdgpu_sync_fence(sync, f, GFP_KERNEL);
 		if (r)
 			break;
 	}
+	dma_resv_iter_end(&cursor);
+
 	return r;
 }
 
+/* Free the entry back to the slab */
+static void amdgpu_sync_entry_free(struct amdgpu_sync_entry *e)
+{
+	hash_del(&e->node);
+	dma_fence_put(e->fence);
+	kmem_cache_free(amdgpu_sync_slab, e);
+}
+
 /**
  * amdgpu_sync_peek_fence - get the next fence not signaled yet
  *
@@ -242,8 +322,12 @@ struct dma_fence *amdgpu_sync_peek_fence(struct amdgpu_sync *sync,
 
 	hash_for_each_safe(sync->fences, i, tmp, e, node) {
 		struct dma_fence *f = e->fence;
-		struct amd_sched_fence *s_fence = to_amd_sched_fence(f);
+		struct drm_sched_fence *s_fence = to_drm_sched_fence(f);
 
+		if (dma_fence_is_signaled(f)) {
+			amdgpu_sync_entry_free(e);
+			continue;
+		}
 		if (ring && s_fence) {
 			/* For fences from the same ring it is sufficient
 			 * when they are scheduled.
@@ -256,13 +340,6 @@ struct dma_fence *amdgpu_sync_peek_fence(struct amdgpu_sync *sync,
 			}
 		}
 
-		if (dma_fence_is_signaled(f)) {
-			hash_del(&e->node);
-			dma_fence_put(f);
-			kmem_cache_free(amdgpu_sync_slab, e);
-			continue;
-		}
-
 		return f;
 	}
 
@@ -298,6 +375,86 @@ struct dma_fence *amdgpu_sync_get_fence(struct amdgpu_sync *sync)
 	return NULL;
 }
 
+/**
+ * amdgpu_sync_clone - clone a sync object
+ *
+ * @source: sync object to clone
+ * @clone: pointer to destination sync object
+ *
+ * Adds references to all unsignaled fences in @source to @clone. Also
+ * removes signaled fences from @source while at it.
+ */
+int amdgpu_sync_clone(struct amdgpu_sync *source, struct amdgpu_sync *clone)
+{
+	struct amdgpu_sync_entry *e;
+	struct hlist_node *tmp;
+	struct dma_fence *f;
+	int i, r;
+
+	hash_for_each_safe(source->fences, i, tmp, e, node) {
+		f = e->fence;
+		if (!dma_fence_is_signaled(f)) {
+			r = amdgpu_sync_fence(clone, f, GFP_KERNEL);
+			if (r)
+				return r;
+		} else {
+			amdgpu_sync_entry_free(e);
+		}
+	}
+
+	return 0;
+}
+
+/**
+ * amdgpu_sync_move - move all fences from src to dst
+ *
+ * @src: source of the fences, empty after function
+ * @dst: destination for the fences
+ *
+ * Moves all fences from source to destination. All fences in destination are
+ * freed and source is empty after the function call.
+ */
+void amdgpu_sync_move(struct amdgpu_sync *src, struct amdgpu_sync *dst)
+{
+	unsigned int i;
+
+	amdgpu_sync_free(dst);
+
+	for (i = 0; i < HASH_SIZE(src->fences); ++i)
+		hlist_move_list(&src->fences[i], &dst->fences[i]);
+}
+
+/**
+ * amdgpu_sync_push_to_job - push fences into job
+ * @sync: sync object to get the fences from
+ * @job: job to push the fences into
+ *
+ * Add all unsignaled fences from sync to job.
+ */
+int amdgpu_sync_push_to_job(struct amdgpu_sync *sync, struct amdgpu_job *job)
+{
+	struct amdgpu_sync_entry *e;
+	struct hlist_node *tmp;
+	struct dma_fence *f;
+	int i, r;
+
+	hash_for_each_safe(sync->fences, i, tmp, e, node) {
+		f = e->fence;
+		if (dma_fence_is_signaled(f)) {
+			amdgpu_sync_entry_free(e);
+			continue;
+		}
+
+		dma_fence_get(f);
+		r = drm_sched_job_add_dependency(&job->base, f);
+		if (r) {
+			dma_fence_put(f);
+			return r;
+		}
+	}
+	return 0;
+}
+
 int amdgpu_sync_wait(struct amdgpu_sync *sync, bool intr)
 {
 	struct amdgpu_sync_entry *e;
@@ -309,9 +466,7 @@ int amdgpu_sync_wait(struct amdgpu_sync *sync, bool intr)
 		if (r)
 			return r;
 
-		hash_del(&e->node);
-		dma_fence_put(e->fence);
-		kmem_cache_free(amdgpu_sync_slab, e);
+		amdgpu_sync_entry_free(e);
 	}
 
 	return 0;
@@ -328,15 +483,10 @@ void amdgpu_sync_free(struct amdgpu_sync *sync)
 {
 	struct amdgpu_sync_entry *e;
 	struct hlist_node *tmp;
-	unsigned i;
-
-	hash_for_each_safe(sync->fences, i, tmp, e, node) {
-		hash_del(&e->node);
-		dma_fence_put(e->fence);
-		kmem_cache_free(amdgpu_sync_slab, e);
-	}
+	unsigned int i;
 
-	dma_fence_put(sync->last_vm_update);
+	hash_for_each_safe(sync->fences, i, tmp, e, node)
+		amdgpu_sync_entry_free(e);
 }
 
 /**
@@ -346,9 +496,7 @@ void amdgpu_sync_free(struct amdgpu_sync *sync)
  */
 int amdgpu_sync_init(void)
 {
-	amdgpu_sync_slab = kmem_cache_create(
-		"amdgpu_sync", sizeof(struct amdgpu_sync_entry), 0,
-		SLAB_HWCACHE_ALIGN, NULL);
+	amdgpu_sync_slab = KMEM_CACHE(amdgpu_sync_entry, SLAB_HWCACHE_ALIGN);
 	if (!amdgpu_sync_slab)
 		return -ENOMEM;