drm/amdgpu: add user fence context map v2

This is a prerequisite for the GPU scheduler to make the order
of submission independent from the order of execution.

v2: properly implement the locking

Signed-off-by: Christian König <christian.koenig@amd.com>
Reviewed-by: Jammy Zhou <Jammy.Zhou@amd.com>
Reviewed-by: Chunming Zhou <david1.zhou@amd.com>

4 files changed, 110 insertions, 30 deletions

diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu.h b/drivers/gpu/drm/amd/amdgpu/amdgpu.h
index 70e783a..0220d98 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu.h
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu.h
@@ -415,6 +415,8 @@ struct amdgpu_user_fence {
 	struct amdgpu_bo 	*bo;
 	/* write-back address offset to bo start */
 	uint32_t                offset;
+	/* resulting sequence number */
+	uint64_t                sequence;
 };
 
 int amdgpu_fence_driver_init(struct amdgpu_device *adev);
@@ -985,9 +987,18 @@ struct amdgpu_vm_manager {
  * context related structures
  */
 
+#define AMDGPU_CTX_MAX_CS_PENDING	16
+
+struct amdgpu_ctx_ring {
+	uint64_t	sequence;
+	struct fence	*fences[AMDGPU_CTX_MAX_CS_PENDING];
+};
+
 struct amdgpu_ctx {
 	struct kref		refcount;
 	unsigned		reset_counter;
+	spinlock_t		ring_lock;
+	struct amdgpu_ctx_ring	rings[AMDGPU_MAX_RINGS];
 };
 
 struct amdgpu_ctx_mgr {
@@ -1007,6 +1018,11 @@ void amdgpu_ctx_fini(struct amdgpu_fpriv *fpriv);
 struct amdgpu_ctx *amdgpu_ctx_get(struct amdgpu_fpriv *fpriv, uint32_t id);
 int amdgpu_ctx_put(struct amdgpu_ctx *ctx);
 
+uint64_t amdgpu_ctx_add_fence(struct amdgpu_ctx *ctx, struct amdgpu_ring *ring,
+			      struct fence *fence);
+struct fence *amdgpu_ctx_get_fence(struct amdgpu_ctx *ctx,
+				   struct amdgpu_ring *ring, uint64_t seq);
+
 int amdgpu_ctx_ioctl(struct drm_device *dev, void *data,
 		     struct drm_file *filp);
 
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c
index 53e6a10f..cef8360 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c
@@ -698,9 +698,9 @@ static int amdgpu_cs_dependencies(struct amdgpu_device *adev,
 			sizeof(struct drm_amdgpu_cs_chunk_dep);
 
 		for (j = 0; j < num_deps; ++j) {
-			struct amdgpu_fence *fence;
 			struct amdgpu_ring *ring;
 			struct amdgpu_ctx *ctx;
+			struct fence *fence;
 
 			r = amdgpu_cs_get_ring(adev, deps[j].ip_type,
 					       deps[j].ip_instance,
@@ -712,20 +712,20 @@ static int amdgpu_cs_dependencies(struct amdgpu_device *adev,
 			if (ctx == NULL)
 				return -EINVAL;
 
-			r = amdgpu_fence_recreate(ring, p->filp,
-						  deps[j].handle,
-						  &fence);
-			if (r) {
+			fence = amdgpu_ctx_get_fence(ctx, ring,
+						     deps[j].handle);
+			if (IS_ERR(fence)) {
+				r = PTR_ERR(fence);
 				amdgpu_ctx_put(ctx);
 				return r;
-			}
-
-			r = amdgpu_sync_fence(adev, &ib->sync, &fence->base);
-			amdgpu_fence_unref(&fence);
-			amdgpu_ctx_put(ctx);
 
-			if (r)
-				return r;
+			} else if (fence) {
+				r = amdgpu_sync_fence(adev, &ib->sync, fence);
+				fence_put(fence);
+				amdgpu_ctx_put(ctx);
+				if (r)
+					return r;
+			}
 		}
 	}
 
@@ -773,8 +773,11 @@ int amdgpu_cs_ioctl(struct drm_device *dev, void *data, struct drm_file *filp)
 		r = amdgpu_cs_ib_fill(adev, &parser);
 	}
 
-	if (!r)
+	if (!r) {
 		r = amdgpu_cs_dependencies(adev, &parser);
+		if (r)
+			DRM_ERROR("Failed in the dependencies handling %d!\n", r);
+	}
 
 	if (r) {
 		amdgpu_cs_parser_fini(&parser, r, reserved_buffers);
@@ -791,7 +794,7 @@ int amdgpu_cs_ioctl(struct drm_device *dev, void *data, struct drm_file *filp)
 		goto out;
 	}
 
-	cs->out.handle = parser.ibs[parser.num_ibs - 1].fence->seq;
+	cs->out.handle = parser.uf.sequence;
 out:
 	amdgpu_cs_parser_fini(&parser, r, true);
 	up_read(&adev->exclusive_lock);
@@ -814,30 +817,31 @@ int amdgpu_cs_wait_ioctl(struct drm_device *dev, void *data,
 	union drm_amdgpu_wait_cs *wait = data;
 	struct amdgpu_device *adev = dev->dev_private;
 	unsigned long timeout = amdgpu_gem_timeout(wait->in.timeout);
-	struct amdgpu_fence *fence = NULL;
 	struct amdgpu_ring *ring = NULL;
 	struct amdgpu_ctx *ctx;
+	struct fence *fence;
 	long r;
 
+	r = amdgpu_cs_get_ring(adev, wait->in.ip_type, wait->in.ip_instance,
+			       wait->in.ring, &ring);
+	if (r)
+		return r;
+
 	ctx = amdgpu_ctx_get(filp->driver_priv, wait->in.ctx_id);
 	if (ctx == NULL)
 		return -EINVAL;
 
-	r = amdgpu_cs_get_ring(adev, wait->in.ip_type, wait->in.ip_instance,
-			       wait->in.ring, &ring);
-	if (r) {
-		amdgpu_ctx_put(ctx);
-		return r;
-	}
+	fence = amdgpu_ctx_get_fence(ctx, ring, wait->in.handle);
+	if (IS_ERR(fence))
+		r = PTR_ERR(fence);
 
-	r = amdgpu_fence_recreate(ring, filp, wait->in.handle, &fence);
-	if (r) {
-		amdgpu_ctx_put(ctx);
-		return r;
-	}
+	else if (fence) {
+		r = fence_wait_timeout(fence, true, timeout);
+		fence_put(fence);
+
+	} else
+		r = 1;
 
-	r = fence_wait_timeout(&fence->base, true, timeout);
-	amdgpu_fence_unref(&fence);
 	amdgpu_ctx_put(ctx);
 	if (r < 0)
 		return r;
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ctx.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_ctx.c
index e63cfb7..c23bfd8 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ctx.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ctx.c
@@ -28,17 +28,22 @@
 static void amdgpu_ctx_do_release(struct kref *ref)
 {
 	struct amdgpu_ctx *ctx;
+	unsigned i, j;
 
 	ctx = container_of(ref, struct amdgpu_ctx, refcount);
+
+	for (i = 0; i < AMDGPU_MAX_RINGS; ++i)
+		for (j = 0; j < AMDGPU_CTX_MAX_CS_PENDING; ++j)
+			fence_put(ctx->rings[i].fences[j]);
 	kfree(ctx);
 }
 
 int amdgpu_ctx_alloc(struct amdgpu_device *adev, struct amdgpu_fpriv *fpriv,
 		     uint32_t *id)
 {
-	int r;
 	struct amdgpu_ctx *ctx;
 	struct amdgpu_ctx_mgr *mgr = &fpriv->ctx_mgr;
+	int i, r;
 
 	ctx = kmalloc(sizeof(*ctx), GFP_KERNEL);
 	if (!ctx)
@@ -55,6 +60,9 @@ int amdgpu_ctx_alloc(struct amdgpu_device *adev, struct amdgpu_fpriv *fpriv,
 
 	memset(ctx, 0, sizeof(*ctx));
 	kref_init(&ctx->refcount);
+	spin_lock_init(&ctx->ring_lock);
+	for (i = 0; i < AMDGPU_MAX_RINGS; ++i)
+		ctx->rings[i].sequence = 1;
 	mutex_unlock(&mgr->lock);
 
 	return 0;
@@ -177,3 +185,53 @@ int amdgpu_ctx_put(struct amdgpu_ctx *ctx)
 	kref_put(&ctx->refcount, amdgpu_ctx_do_release);
 	return 0;
 }
+
+uint64_t amdgpu_ctx_add_fence(struct amdgpu_ctx *ctx, struct amdgpu_ring *ring,
+			      struct fence *fence)
+{
+	struct amdgpu_ctx_ring *cring = & ctx->rings[ring->idx];
+	uint64_t seq = cring->sequence;
+	unsigned idx = seq % AMDGPU_CTX_MAX_CS_PENDING;
+	struct fence *other = cring->fences[idx];
+
+	if (other) {
+		signed long r;
+		r = fence_wait_timeout(other, false, MAX_SCHEDULE_TIMEOUT);
+		if (r < 0)
+			DRM_ERROR("Error (%ld) waiting for fence!\n", r);
+	}
+
+	fence_get(fence);
+
+	spin_lock(&ctx->ring_lock);
+	cring->fences[idx] = fence;
+	cring->sequence++;
+	spin_unlock(&ctx->ring_lock);
+
+	fence_put(other);
+
+	return seq;
+}
+
+struct fence *amdgpu_ctx_get_fence(struct amdgpu_ctx *ctx,
+				   struct amdgpu_ring *ring, uint64_t seq)
+{
+	struct amdgpu_ctx_ring *cring = & ctx->rings[ring->idx];
+	struct fence *fence;
+
+	spin_lock(&ctx->ring_lock);
+	if (seq >= cring->sequence) {
+		spin_unlock(&ctx->ring_lock);
+		return ERR_PTR(-EINVAL);
+	}
+
+	if (seq < cring->sequence - AMDGPU_CTX_MAX_CS_PENDING) {
+		spin_unlock(&ctx->ring_lock);
+		return NULL;
+	}
+
+	fence = fence_get(cring->fences[seq % AMDGPU_CTX_MAX_CS_PENDING]);
+	spin_unlock(&ctx->ring_lock);
+
+	return fence;
+}
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ib.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_ib.c
index 2722815..95d5334 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ib.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ib.c
@@ -219,8 +219,10 @@ int amdgpu_ib_schedule(struct amdgpu_device *adev, unsigned num_ibs,
 	/* wrap the last IB with fence */
 	if (ib->user) {
 		uint64_t addr = amdgpu_bo_gpu_offset(ib->user->bo);
+		ib->user->sequence = amdgpu_ctx_add_fence(ib->ctx, ring,
+							  &ib->fence->base);
 		addr += ib->user->offset;
-		amdgpu_ring_emit_fence(ring, addr, ib->fence->seq,
+		amdgpu_ring_emit_fence(ring, addr, ib->user->sequence,
 				       AMDGPU_FENCE_FLAG_64BIT);
 	}