drm/radeon: allow semaphore emission to fail

To workaround bugs and/or certain limits it's sometimes
useful to fall back to waiting on fences.

Signed-off-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Cc: stable@vger.kernel.org

17 files changed, 191 insertions, 147 deletions

diff --git a/drivers/gpu/drm/radeon/cik.c b/drivers/gpu/drm/radeon/cik.c
index ae92aa0..811fc1b 100644
--- a/drivers/gpu/drm/radeon/cik.c
+++ b/drivers/gpu/drm/radeon/cik.c
@@ -3556,7 +3556,7 @@ void cik_fence_compute_ring_emit(struct radeon_device *rdev,
 	radeon_ring_write(ring, 0);
 }
 
-void cik_semaphore_ring_emit(struct radeon_device *rdev,
+bool cik_semaphore_ring_emit(struct radeon_device *rdev,
 			     struct radeon_ring *ring,
 			     struct radeon_semaphore *semaphore,
 			     bool emit_wait)
@@ -3567,6 +3567,8 @@ void cik_semaphore_ring_emit(struct radeon_device *rdev,
 	radeon_ring_write(ring, PACKET3(PACKET3_MEM_SEMAPHORE, 1));
 	radeon_ring_write(ring, addr & 0xffffffff);
 	radeon_ring_write(ring, (upper_32_bits(addr) & 0xffff) | sel);
+
+	return true;
 }
 
 /**
@@ -3609,13 +3611,8 @@ int cik_copy_cpdma(struct radeon_device *rdev,
 		return r;
 	}
 
-	if (radeon_fence_need_sync(*fence, ring->idx)) {
-		radeon_semaphore_sync_rings(rdev, sem, (*fence)->ring,
-					    ring->idx);
-		radeon_fence_note_sync(*fence, ring->idx);
-	} else {
-		radeon_semaphore_free(rdev, &sem, NULL);
-	}
+	radeon_semaphore_sync_to(sem, *fence);
+	radeon_semaphore_sync_rings(rdev, sem, ring->idx);
 
 	for (i = 0; i < num_loops; i++) {
 		cur_size_in_bytes = size_in_bytes;
diff --git a/drivers/gpu/drm/radeon/cik_sdma.c b/drivers/gpu/drm/radeon/cik_sdma.c
index 9c9529d..0300727 100644
--- a/drivers/gpu/drm/radeon/cik_sdma.c
+++ b/drivers/gpu/drm/radeon/cik_sdma.c
@@ -130,7 +130,7 @@ void cik_sdma_fence_ring_emit(struct radeon_device *rdev,
  * Add a DMA semaphore packet to the ring wait on or signal
  * other rings (CIK).
  */
-void cik_sdma_semaphore_ring_emit(struct radeon_device *rdev,
+bool cik_sdma_semaphore_ring_emit(struct radeon_device *rdev,
 				  struct radeon_ring *ring,
 				  struct radeon_semaphore *semaphore,
 				  bool emit_wait)
@@ -141,6 +141,8 @@ void cik_sdma_semaphore_ring_emit(struct radeon_device *rdev,
 	radeon_ring_write(ring, SDMA_PACKET(SDMA_OPCODE_SEMAPHORE, 0, extra_bits));
 	radeon_ring_write(ring, addr & 0xfffffff8);
 	radeon_ring_write(ring, upper_32_bits(addr) & 0xffffffff);
+
+	return true;
 }
 
 /**
@@ -443,13 +445,8 @@ int cik_copy_dma(struct radeon_device *rdev,
 		return r;
 	}
 
-	if (radeon_fence_need_sync(*fence, ring->idx)) {
-		radeon_semaphore_sync_rings(rdev, sem, (*fence)->ring,
-					    ring->idx);
-		radeon_fence_note_sync(*fence, ring->idx);
-	} else {
-		radeon_semaphore_free(rdev, &sem, NULL);
-	}
+	radeon_semaphore_sync_to(sem, *fence);
+	radeon_semaphore_sync_rings(rdev, sem, ring->idx);
 
 	for (i = 0; i < num_loops; i++) {
 		cur_size_in_bytes = size_in_bytes;
diff --git a/drivers/gpu/drm/radeon/evergreen_dma.c b/drivers/gpu/drm/radeon/evergreen_dma.c
index 6a0656d..a37b544 100644
--- a/drivers/gpu/drm/radeon/evergreen_dma.c
+++ b/drivers/gpu/drm/radeon/evergreen_dma.c
@@ -131,13 +131,8 @@ int evergreen_copy_dma(struct radeon_device *rdev,
 		return r;
 	}
 
-	if (radeon_fence_need_sync(*fence, ring->idx)) {
-		radeon_semaphore_sync_rings(rdev, sem, (*fence)->ring,
-					    ring->idx);
-		radeon_fence_note_sync(*fence, ring->idx);
-	} else {
-		radeon_semaphore_free(rdev, &sem, NULL);
-	}
+	radeon_semaphore_sync_to(sem, *fence);
+	radeon_semaphore_sync_rings(rdev, sem, ring->idx);
 
 	for (i = 0; i < num_loops; i++) {
 		cur_size_in_dw = size_in_dw;
diff --git a/drivers/gpu/drm/radeon/r100.c b/drivers/gpu/drm/radeon/r100.c
index 784983d..10abc4d 100644
--- a/drivers/gpu/drm/radeon/r100.c
+++ b/drivers/gpu/drm/radeon/r100.c
@@ -869,13 +869,14 @@ void r100_fence_ring_emit(struct radeon_device *rdev,
 	radeon_ring_write(ring, RADEON_SW_INT_FIRE);
 }
 
-void r100_semaphore_ring_emit(struct radeon_device *rdev,
+bool r100_semaphore_ring_emit(struct radeon_device *rdev,
 			      struct radeon_ring *ring,
 			      struct radeon_semaphore *semaphore,
 			      bool emit_wait)
 {
 	/* Unused on older asics, since we don't have semaphores or multiple rings */
 	BUG();
+	return false;
 }
 
 int r100_copy_blit(struct radeon_device *rdev,
diff --git a/drivers/gpu/drm/radeon/r600.c b/drivers/gpu/drm/radeon/r600.c
index 4e609e8..9ad0673 100644
--- a/drivers/gpu/drm/radeon/r600.c
+++ b/drivers/gpu/drm/radeon/r600.c
@@ -2650,7 +2650,7 @@ void r600_fence_ring_emit(struct radeon_device *rdev,
 	}
 }
 
-void r600_semaphore_ring_emit(struct radeon_device *rdev,
+bool r600_semaphore_ring_emit(struct radeon_device *rdev,
 			      struct radeon_ring *ring,
 			      struct radeon_semaphore *semaphore,
 			      bool emit_wait)
@@ -2664,6 +2664,8 @@ void r600_semaphore_ring_emit(struct radeon_device *rdev,
 	radeon_ring_write(ring, PACKET3(PACKET3_MEM_SEMAPHORE, 1));
 	radeon_ring_write(ring, addr & 0xffffffff);
 	radeon_ring_write(ring, (upper_32_bits(addr) & 0xff) | sel);
+
+	return true;
 }
 
 /**
@@ -2706,13 +2708,8 @@ int r600_copy_cpdma(struct radeon_device *rdev,
 		return r;
 	}
 
-	if (radeon_fence_need_sync(*fence, ring->idx)) {
-		radeon_semaphore_sync_rings(rdev, sem, (*fence)->ring,
-					    ring->idx);
-		radeon_fence_note_sync(*fence, ring->idx);
-	} else {
-		radeon_semaphore_free(rdev, &sem, NULL);
-	}
+	radeon_semaphore_sync_to(sem, *fence);
+	radeon_semaphore_sync_rings(rdev, sem, ring->idx);
 
 	radeon_ring_write(ring, PACKET3(PACKET3_SET_CONFIG_REG, 1));
 	radeon_ring_write(ring, (WAIT_UNTIL - PACKET3_SET_CONFIG_REG_OFFSET) >> 2);
diff --git a/drivers/gpu/drm/radeon/r600_dma.c b/drivers/gpu/drm/radeon/r600_dma.c
index 3b31745..7844d15 100644
--- a/drivers/gpu/drm/radeon/r600_dma.c
+++ b/drivers/gpu/drm/radeon/r600_dma.c
@@ -311,7 +311,7 @@ void r600_dma_fence_ring_emit(struct radeon_device *rdev,
  * Add a DMA semaphore packet to the ring wait on or signal
  * other rings (r6xx-SI).
  */
-void r600_dma_semaphore_ring_emit(struct radeon_device *rdev,
+bool r600_dma_semaphore_ring_emit(struct radeon_device *rdev,
 				  struct radeon_ring *ring,
 				  struct radeon_semaphore *semaphore,
 				  bool emit_wait)
@@ -322,6 +322,8 @@ void r600_dma_semaphore_ring_emit(struct radeon_device *rdev,
 	radeon_ring_write(ring, DMA_PACKET(DMA_PACKET_SEMAPHORE, 0, s, 0));
 	radeon_ring_write(ring, addr & 0xfffffffc);
 	radeon_ring_write(ring, upper_32_bits(addr) & 0xff);
+
+	return true;
 }
 
 /**
@@ -462,13 +464,8 @@ int r600_copy_dma(struct radeon_device *rdev,
 		return r;
 	}
 
-	if (radeon_fence_need_sync(*fence, ring->idx)) {
-		radeon_semaphore_sync_rings(rdev, sem, (*fence)->ring,
-					    ring->idx);
-		radeon_fence_note_sync(*fence, ring->idx);
-	} else {
-		radeon_semaphore_free(rdev, &sem, NULL);
-	}
+	radeon_semaphore_sync_to(sem, *fence);
+	radeon_semaphore_sync_rings(rdev, sem, ring->idx);
 
 	for (i = 0; i < num_loops; i++) {
 		cur_size_in_dw = size_in_dw;
diff --git a/drivers/gpu/drm/radeon/radeon.h b/drivers/gpu/drm/radeon/radeon.h
index b9ee992..2fe2f63 100644
--- a/drivers/gpu/drm/radeon/radeon.h
+++ b/drivers/gpu/drm/radeon/radeon.h
@@ -348,6 +348,7 @@ int radeon_fence_emit(struct radeon_device *rdev, struct radeon_fence **fence, i
 void radeon_fence_process(struct radeon_device *rdev, int ring);
 bool radeon_fence_signaled(struct radeon_fence *fence);
 int radeon_fence_wait(struct radeon_fence *fence, bool interruptible);
+int radeon_fence_wait_locked(struct radeon_fence *fence);
 int radeon_fence_wait_next_locked(struct radeon_device *rdev, int ring);
 int radeon_fence_wait_empty_locked(struct radeon_device *rdev, int ring);
 int radeon_fence_wait_any(struct radeon_device *rdev,
@@ -548,17 +549,20 @@ struct radeon_semaphore {
 	struct radeon_sa_bo		*sa_bo;
 	signed				waiters;
 	uint64_t			gpu_addr;
+	struct radeon_fence		*sync_to[RADEON_NUM_RINGS];
 };
 
 int radeon_semaphore_create(struct radeon_device *rdev,
 			    struct radeon_semaphore **semaphore);
-void radeon_semaphore_emit_signal(struct radeon_device *rdev, int ring,
+bool radeon_semaphore_emit_signal(struct radeon_device *rdev, int ring,
 				  struct radeon_semaphore *semaphore);
-void radeon_semaphore_emit_wait(struct radeon_device *rdev, int ring,
+bool radeon_semaphore_emit_wait(struct radeon_device *rdev, int ring,
 				struct radeon_semaphore *semaphore);
+void radeon_semaphore_sync_to(struct radeon_semaphore *semaphore,
+			      struct radeon_fence *fence);
 int radeon_semaphore_sync_rings(struct radeon_device *rdev,
 				struct radeon_semaphore *semaphore,
-				int signaler, int waiter);
+				int waiting_ring);
 void radeon_semaphore_free(struct radeon_device *rdev,
 			   struct radeon_semaphore **semaphore,
 			   struct radeon_fence *fence);
@@ -765,7 +769,6 @@ struct radeon_ib {
 	struct radeon_fence		*fence;
 	struct radeon_vm		*vm;
 	bool				is_const_ib;
-	struct radeon_fence		*sync_to[RADEON_NUM_RINGS];
 	struct radeon_semaphore		*semaphore;
 };
 
@@ -921,7 +924,6 @@ int radeon_ib_get(struct radeon_device *rdev, int ring,
 		  struct radeon_ib *ib, struct radeon_vm *vm,
 		  unsigned size);
 void radeon_ib_free(struct radeon_device *rdev, struct radeon_ib *ib);
-void radeon_ib_sync_to(struct radeon_ib *ib, struct radeon_fence *fence);
 int radeon_ib_schedule(struct radeon_device *rdev, struct radeon_ib *ib,
 		       struct radeon_ib *const_ib);
 int radeon_ib_pool_init(struct radeon_device *rdev);
@@ -1638,7 +1640,7 @@ struct radeon_asic_ring {
 	/* command emmit functions */
 	void (*ib_execute)(struct radeon_device *rdev, struct radeon_ib *ib);
 	void (*emit_fence)(struct radeon_device *rdev, struct radeon_fence *fence);
-	void (*emit_semaphore)(struct radeon_device *rdev, struct radeon_ring *cp,
+	bool (*emit_semaphore)(struct radeon_device *rdev, struct radeon_ring *cp,
 			       struct radeon_semaphore *semaphore, bool emit_wait);
 	void (*vm_flush)(struct radeon_device *rdev, int ridx, struct radeon_vm *vm);
 
diff --git a/drivers/gpu/drm/radeon/radeon_asic.h b/drivers/gpu/drm/radeon/radeon_asic.h
index f2833ee..c9fd97b 100644
--- a/drivers/gpu/drm/radeon/radeon_asic.h
+++ b/drivers/gpu/drm/radeon/radeon_asic.h
@@ -80,7 +80,7 @@ int r100_irq_set(struct radeon_device *rdev);
 int r100_irq_process(struct radeon_device *rdev);
 void r100_fence_ring_emit(struct radeon_device *rdev,
 			  struct radeon_fence *fence);
-void r100_semaphore_ring_emit(struct radeon_device *rdev,
+bool r100_semaphore_ring_emit(struct radeon_device *rdev,
 			      struct radeon_ring *cp,
 			      struct radeon_semaphore *semaphore,
 			      bool emit_wait);
@@ -313,13 +313,13 @@ int r600_cs_parse(struct radeon_cs_parser *p);
 int r600_dma_cs_parse(struct radeon_cs_parser *p);
 void r600_fence_ring_emit(struct radeon_device *rdev,
 			  struct radeon_fence *fence);
-void r600_semaphore_ring_emit(struct radeon_device *rdev,
+bool r600_semaphore_ring_emit(struct radeon_device *rdev,
 			      struct radeon_ring *cp,
 			      struct radeon_semaphore *semaphore,
 			      bool emit_wait);
 void r600_dma_fence_ring_emit(struct radeon_device *rdev,
 			      struct radeon_fence *fence);
-void r600_dma_semaphore_ring_emit(struct radeon_device *rdev,
+bool r600_dma_semaphore_ring_emit(struct radeon_device *rdev,
 				  struct radeon_ring *ring,
 				  struct radeon_semaphore *semaphore,
 				  bool emit_wait);
@@ -566,10 +566,6 @@ int sumo_dpm_force_performance_level(struct radeon_device *rdev,
  */
 void cayman_fence_ring_emit(struct radeon_device *rdev,
 			    struct radeon_fence *fence);
-void cayman_uvd_semaphore_emit(struct radeon_device *rdev,
-			       struct radeon_ring *ring,
-			       struct radeon_semaphore *semaphore,
-			       bool emit_wait);
 void cayman_pcie_gart_tlb_flush(struct radeon_device *rdev);
 int cayman_init(struct radeon_device *rdev);
 void cayman_fini(struct radeon_device *rdev);
@@ -697,7 +693,7 @@ void cik_pciep_wreg(struct radeon_device *rdev, uint32_t reg, uint32_t v);
 int cik_set_uvd_clocks(struct radeon_device *rdev, u32 vclk, u32 dclk);
 void cik_sdma_fence_ring_emit(struct radeon_device *rdev,
 			      struct radeon_fence *fence);
-void cik_sdma_semaphore_ring_emit(struct radeon_device *rdev,
+bool cik_sdma_semaphore_ring_emit(struct radeon_device *rdev,
 				  struct radeon_ring *ring,
 				  struct radeon_semaphore *semaphore,
 				  bool emit_wait);
@@ -717,7 +713,7 @@ void cik_fence_gfx_ring_emit(struct radeon_device *rdev,
 			     struct radeon_fence *fence);
 void cik_fence_compute_ring_emit(struct radeon_device *rdev,
 				 struct radeon_fence *fence);
-void cik_semaphore_ring_emit(struct radeon_device *rdev,
+bool cik_semaphore_ring_emit(struct radeon_device *rdev,
 			     struct radeon_ring *cp,
 			     struct radeon_semaphore *semaphore,
 			     bool emit_wait);
@@ -807,7 +803,7 @@ void uvd_v1_0_stop(struct radeon_device *rdev);
 
 int uvd_v1_0_ring_test(struct radeon_device *rdev, struct radeon_ring *ring);
 int uvd_v1_0_ib_test(struct radeon_device *rdev, struct radeon_ring *ring);
-void uvd_v1_0_semaphore_emit(struct radeon_device *rdev,
+bool uvd_v1_0_semaphore_emit(struct radeon_device *rdev,
 			     struct radeon_ring *ring,
 			     struct radeon_semaphore *semaphore,
 			     bool emit_wait);
@@ -819,7 +815,7 @@ void uvd_v2_2_fence_emit(struct radeon_device *rdev,
 			 struct radeon_fence *fence);
 
 /* uvd v3.1 */
-void uvd_v3_1_semaphore_emit(struct radeon_device *rdev,
+bool uvd_v3_1_semaphore_emit(struct radeon_device *rdev,
 			     struct radeon_ring *ring,
 			     struct radeon_semaphore *semaphore,
 			     bool emit_wait);
diff --git a/drivers/gpu/drm/radeon/radeon_cs.c b/drivers/gpu/drm/radeon/radeon_cs.c
index 26ca223..f41594b 100644
--- a/drivers/gpu/drm/radeon/radeon_cs.c
+++ b/drivers/gpu/drm/radeon/radeon_cs.c
@@ -159,7 +159,8 @@ static void radeon_cs_sync_rings(struct radeon_cs_parser *p)
 		if (!p->relocs[i].robj)
 			continue;
 
-		radeon_ib_sync_to(&p->ib, p->relocs[i].robj->tbo.sync_obj);
+		radeon_semaphore_sync_to(p->ib.semaphore,
+					 p->relocs[i].robj->tbo.sync_obj);
 	}
 }
 
@@ -411,9 +412,9 @@ static int radeon_cs_ib_vm_chunk(struct radeon_device *rdev,
 		goto out;
 	}
 	radeon_cs_sync_rings(parser);
-	radeon_ib_sync_to(&parser->ib, vm->fence);
-	radeon_ib_sync_to(&parser->ib, radeon_vm_grab_id(
-		rdev, vm, parser->ring));
+	radeon_semaphore_sync_to(parser->ib.semaphore, vm->fence);
+	radeon_semaphore_sync_to(parser->ib.semaphore,
+				 radeon_vm_grab_id(rdev, vm, parser->ring));
 
 	if ((rdev->family >= CHIP_TAHITI) &&
 	    (parser->chunk_const_ib_idx != -1)) {
diff --git a/drivers/gpu/drm/radeon/radeon_fence.c b/drivers/gpu/drm/radeon/radeon_fence.c
index 281d14c..d3a86e4 100644
--- a/drivers/gpu/drm/radeon/radeon_fence.c
+++ b/drivers/gpu/drm/radeon/radeon_fence.c
@@ -472,6 +472,36 @@ int radeon_fence_wait_any(struct radeon_device *rdev,
 }
 
 /**
+ * radeon_fence_wait_locked - wait for a fence to signal
+ *
+ * @fence: radeon fence object
+ *
+ * Wait for the requested fence to signal (all asics).
+ * Returns 0 if the fence has passed, error for all other cases.
+ */
+int radeon_fence_wait_locked(struct radeon_fence *fence)
+{
+	uint64_t seq[RADEON_NUM_RINGS] = {};
+	int 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)
+		return 0;
+
+	r = radeon_fence_wait_seq(fence->rdev, seq, false, false);
+	if (r)
+		return r;
+
+	fence->seq = RADEON_FENCE_SIGNALED_SEQ;
+	return 0;
+}
+
+/**
  * radeon_fence_wait_next_locked - wait for the next fence to signal
  *
  * @rdev: radeon device pointer
diff --git a/drivers/gpu/drm/radeon/radeon_gart.c b/drivers/gpu/drm/radeon/radeon_gart.c
index 8a83b89..cd7489b 100644
--- a/drivers/gpu/drm/radeon/radeon_gart.c
+++ b/drivers/gpu/drm/radeon/radeon_gart.c
@@ -651,7 +651,7 @@ retry:
 	radeon_asic_vm_set_page(rdev, &ib, vm->pd_gpu_addr,
 				0, pd_entries, 0, 0);
 
-	radeon_ib_sync_to(&ib, vm->fence);
+	radeon_semaphore_sync_to(ib.semaphore, vm->fence);
 	r = radeon_ib_schedule(rdev, &ib, NULL);
 	if (r) {
 		radeon_ib_free(rdev, &ib);
@@ -1220,7 +1220,7 @@ int radeon_vm_bo_update_pte(struct radeon_device *rdev,
 	radeon_vm_update_ptes(rdev, vm, &ib, bo_va->soffset, bo_va->eoffset,
 			      addr, radeon_vm_page_flags(bo_va->flags));
 
-	radeon_ib_sync_to(&ib, vm->fence);
+	radeon_semaphore_sync_to(ib.semaphore, vm->fence);
 	r = radeon_ib_schedule(rdev, &ib, NULL);
 	if (r) {
 		radeon_ib_free(rdev, &ib);
diff --git a/drivers/gpu/drm/radeon/radeon_ring.c b/drivers/gpu/drm/radeon/radeon_ring.c
index 18254e1..9214403 100644
--- a/drivers/gpu/drm/radeon/radeon_ring.c
+++ b/drivers/gpu/drm/radeon/radeon_ring.c
@@ -61,7 +61,7 @@ int radeon_ib_get(struct radeon_device *rdev, int ring,
 		  struct radeon_ib *ib, struct radeon_vm *vm,
 		  unsigned size)
 {
-	int i, r;
+	int r;
 
 	r = radeon_sa_bo_new(rdev, &rdev->ring_tmp_bo, &ib->sa_bo, size, 256, true);
 	if (r) {
@@ -87,8 +87,6 @@ int radeon_ib_get(struct radeon_device *rdev, int ring,
 		ib->gpu_addr = radeon_sa_bo_gpu_addr(ib->sa_bo);
 	}
 	ib->is_const_ib = false;
-	for (i = 0; i < RADEON_NUM_RINGS; ++i)
-		ib->sync_to[i] = NULL;
 
 	return 0;
 }
@@ -109,25 +107,6 @@ void radeon_ib_free(struct radeon_device *rdev, struct radeon_ib *ib)
 }
 
 /**
- * radeon_ib_sync_to - sync to fence before executing the IB
- *
- * @ib: IB object to add fence to
- * @fence: fence to sync to
- *
- * Sync to the fence before executing the IB
- */
-void radeon_ib_sync_to(struct radeon_ib *ib, struct radeon_fence *fence)
-{
-	struct radeon_fence *other;
-
-	if (!fence)
-		return;
-
-	other = ib->sync_to[fence->ring];
-	ib->sync_to[fence->ring] = radeon_fence_later(fence, other);
-}
-
-/**
  * radeon_ib_schedule - schedule an IB (Indirect Buffer) on the ring
  *
  * @rdev: radeon_device pointer
@@ -151,8 +130,7 @@ int radeon_ib_schedule(struct radeon_device *rdev, struct radeon_ib *ib,
 		       struct radeon_ib *const_ib)
 {
 	struct radeon_ring *ring = &rdev->ring[ib->ring];
-	bool need_sync = false;
-	int i, r = 0;
+	int r = 0;
 
 	if (!ib->length_dw || !ring->ready) {
 		/* TODO: Nothings in the ib we should report. */
@@ -166,19 +144,15 @@ int radeon_ib_schedule(struct radeon_device *rdev, struct radeon_ib *ib,
 		dev_err(rdev->dev, "scheduling IB failed (%d).\n", r);
 		return r;
 	}
-	for (i = 0; i < RADEON_NUM_RINGS; ++i) {
-		struct radeon_fence *fence = ib->sync_to[i];
-		if (radeon_fence_need_sync(fence, ib->ring)) {
-			need_sync = true;
-			radeon_semaphore_sync_rings(rdev, ib->semaphore,
-						    fence->ring, ib->ring);
-			radeon_fence_note_sync(fence, ib->ring);
-		}
-	}
-	/* immediately free semaphore when we don't need to sync */
-	if (!need_sync) {
-		radeon_semaphore_free(rdev, &ib->semaphore, NULL);
+
+	/* sync with other rings */
+	r = radeon_semaphore_sync_rings(rdev, ib->semaphore, ib->ring);
+	if (r) {
+		dev_err(rdev->dev, "failed to sync rings (%d)\n", r);
+		radeon_ring_unlock_undo(rdev, ring);
+		return r;
 	}
+
 	/* if we can't remember our last VM flush then flush now! */
 	/* XXX figure out why we have to flush for every IB */
 	if (ib->vm /*&& !ib->vm->last_flush*/) {
diff --git a/drivers/gpu/drm/radeon/radeon_semaphore.c b/drivers/gpu/drm/radeon/radeon_semaphore.c
index 97d73bf..2b42aa1 100644
--- a/drivers/gpu/drm/radeon/radeon_semaphore.c
+++ b/drivers/gpu/drm/radeon/radeon_semaphore.c
@@ -34,7 +34,7 @@
 int radeon_semaphore_create(struct radeon_device *rdev,
 			    struct radeon_semaphore **semaphore)
 {
-	int r;
+	int i, r;
 
 	*semaphore = kmalloc(sizeof(struct radeon_semaphore), GFP_KERNEL);
 	if (*semaphore == NULL) {
@@ -50,58 +50,121 @@ int radeon_semaphore_create(struct radeon_device *rdev,
 	(*semaphore)->waiters = 0;
 	(*semaphore)->gpu_addr = radeon_sa_bo_gpu_addr((*semaphore)->sa_bo);
 	*((uint64_t*)radeon_sa_bo_cpu_addr((*semaphore)->sa_bo)) = 0;
+
+	for (i = 0; i < RADEON_NUM_RINGS; ++i)
+		(*semaphore)->sync_to[i] = NULL;
+
 	return 0;
 }
 
-void radeon_semaphore_emit_signal(struct radeon_device *rdev, int ring,
+bool radeon_semaphore_emit_signal(struct radeon_device *rdev, int ridx,
 			          struct radeon_semaphore *semaphore)
 {
-	trace_radeon_semaphore_signale(ring, semaphore);
+	struct radeon_ring *ring = &rdev->ring[ridx];
+
+	trace_radeon_semaphore_signale(ridx, semaphore);
+
+	if (radeon_semaphore_ring_emit(rdev, ridx, ring, semaphore, false)) {
+		--semaphore->waiters;
 
-	--semaphore->waiters;
-	radeon_semaphore_ring_emit(rdev, ring, &rdev->ring[ring], semaphore, false);
+		/* for debugging lockup only, used by sysfs debug files */
+		ring->last_semaphore_signal_addr = semaphore->gpu_addr;
+		return true;
+	}
+	return false;
 }
 
-void radeon_semaphore_emit_wait(struct radeon_device *rdev, int ring,
+bool radeon_semaphore_emit_wait(struct radeon_device *rdev, int ridx,
 			        struct radeon_semaphore *semaphore)
 {
-	trace_radeon_semaphore_wait(ring, semaphore);
+	struct radeon_ring *ring = &rdev->ring[ridx];
+
+	trace_radeon_semaphore_wait(ridx, semaphore);
+
+	if (radeon_semaphore_ring_emit(rdev, ridx, ring, semaphore, true)) {
+		++semaphore->waiters;
 
-	++semaphore->waiters;
-	radeon_semaphore_ring_emit(rdev, ring, &rdev->ring[ring], semaphore, true);
+		/* for debugging lockup only, used by sysfs debug files */
+		ring->last_semaphore_wait_addr = semaphore->gpu_addr;
+		return true;
+	}
+	return false;
 }
 
-/* caller must hold ring lock */
+/**
+ * radeon_semaphore_sync_to - use the semaphore to sync to a fence
+ *
+ * @semaphore: semaphore object to add fence to
+ * @fence: fence to sync to
+ *
+ * Sync to the fence using this semaphore object
+ */
+void radeon_semaphore_sync_to(struct radeon_semaphore *semaphore,
+			      struct radeon_fence *fence)
+{
+        struct radeon_fence *other;
+
+        if (!fence)
+                return;
+
+        other = semaphore->sync_to[fence->ring];
+        semaphore->sync_to[fence->ring] = radeon_fence_later(fence, other);
+}
+
+/**
+ * radeon_semaphore_sync_rings - sync ring to all registered fences
+ *
+ * @rdev: radeon_device pointer
+ * @semaphore: semaphore object to use for sync
+ * @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 radeon_semaphore_sync_rings(struct radeon_device *rdev,
 				struct radeon_semaphore *semaphore,
-				int signaler, int waiter)
+				int ring)
 {
-	int r;
+	int i, r;
 
-	/* no need to signal and wait on the same ring */
-	if (signaler == waiter) {
-		return 0;
-	}
+        for (i = 0; i < RADEON_NUM_RINGS; ++i) {
+		struct radeon_fence *fence = semaphore->sync_to[i];
 
-	/* prevent GPU deadlocks */
-	if (!rdev->ring[signaler].ready) {
-		dev_err(rdev->dev, "Trying to sync to a disabled ring!");
-		return -EINVAL;
-	}
+		/* check if we really need to sync */
+                if (!radeon_fence_need_sync(fence, ring))
+			continue;
 
-	r = radeon_ring_alloc(rdev, &rdev->ring[signaler], 8);
-	if (r) {
-		return r;
-	}
-	radeon_semaphore_emit_signal(rdev, signaler, semaphore);
-	radeon_ring_commit(rdev, &rdev->ring[signaler]);
+		/* prevent GPU deadlocks */
+		if (!rdev->ring[i].ready) {
+			dev_err(rdev->dev, "Syncing to a disabled ring!");
+			return -EINVAL;
+		}
 
-	/* we assume caller has already allocated space on waiters ring */
-	radeon_semaphore_emit_wait(rdev, waiter, semaphore);
+		/* allocate enough space for sync command */
+		r = radeon_ring_alloc(rdev, &rdev->ring[i], 16);
+		if (r) {
+			return r;
+		}
 
-	/* for debugging lockup only, used by sysfs debug files */
-	rdev->ring[signaler].last_semaphore_signal_addr = semaphore->gpu_addr;
-	rdev->ring[waiter].last_semaphore_wait_addr = semaphore->gpu_addr;
+		/* emit the signal semaphore */
+		if (!radeon_semaphore_emit_signal(rdev, i, semaphore)) {
+			/* signaling wasn't successful wait manually */
+			radeon_ring_undo(&rdev->ring[i]);
+			radeon_fence_wait_locked(fence);
+			continue;
+		}
+
+		/* we assume caller has already allocated space on waiters ring */
+		if (!radeon_semaphore_emit_wait(rdev, ring, semaphore)) {
+			/* waiting wasn't successful wait manually */
+			radeon_ring_undo(&rdev->ring[i]);
+			radeon_fence_wait_locked(fence);
+			continue;
+		}
+
+		radeon_ring_commit(rdev, &rdev->ring[i]);
+		radeon_fence_note_sync(fence, ring);
+	}
 
 	return 0;
 }
diff --git a/drivers/gpu/drm/radeon/rv770_dma.c b/drivers/gpu/drm/radeon/rv770_dma.c
index f9b02e3..aca8cbe 100644
--- a/drivers/gpu/drm/radeon/rv770_dma.c
+++ b/drivers/gpu/drm/radeon/rv770_dma.c
@@ -66,13 +66,8 @@ int rv770_copy_dma(struct radeon_device *rdev,
 		return r;
 	}
 
-	if (radeon_fence_need_sync(*fence, ring->idx)) {
-		radeon_semaphore_sync_rings(rdev, sem, (*fence)->ring,
-					    ring->idx);
-		radeon_fence_note_sync(*fence, ring->idx);
-	} else {
-		radeon_semaphore_free(rdev, &sem, NULL);
-	}
+	radeon_semaphore_sync_to(sem, *fence);
+	radeon_semaphore_sync_rings(rdev, sem, ring->idx);
 
 	for (i = 0; i < num_loops; i++) {
 		cur_size_in_dw = size_in_dw;
diff --git a/drivers/gpu/drm/radeon/si_dma.c b/drivers/gpu/drm/radeon/si_dma.c
index 8e8f461..59be2cf 100644
--- a/drivers/gpu/drm/radeon/si_dma.c
+++ b/drivers/gpu/drm/radeon/si_dma.c
@@ -195,13 +195,8 @@ int si_copy_dma(struct radeon_device *rdev,
 		return r;
 	}
 
-	if (radeon_fence_need_sync(*fence, ring->idx)) {
-		radeon_semaphore_sync_rings(rdev, sem, (*fence)->ring,
-					    ring->idx);
-		radeon_fence_note_sync(*fence, ring->idx);
-	} else {
-		radeon_semaphore_free(rdev, &sem, NULL);
-	}
+	radeon_semaphore_sync_to(sem, *fence);
+	radeon_semaphore_sync_rings(rdev, sem, ring->idx);
 
 	for (i = 0; i < num_loops; i++) {
 		cur_size_in_bytes = size_in_bytes;
diff --git a/drivers/gpu/drm/radeon/uvd_v1_0.c b/drivers/gpu/drm/radeon/uvd_v1_0.c
index 7266805..d4a68af 100644
--- a/drivers/gpu/drm/radeon/uvd_v1_0.c
+++ b/drivers/gpu/drm/radeon/uvd_v1_0.c
@@ -357,7 +357,7 @@ int uvd_v1_0_ring_test(struct radeon_device *rdev, struct radeon_ring *ring)
  *
  * Emit a semaphore command (either wait or signal) to the UVD ring.
  */
-void uvd_v1_0_semaphore_emit(struct radeon_device *rdev,
+bool uvd_v1_0_semaphore_emit(struct radeon_device *rdev,
 			     struct radeon_ring *ring,
 			     struct radeon_semaphore *semaphore,
 			     bool emit_wait)
@@ -372,6 +372,8 @@ void uvd_v1_0_semaphore_emit(struct radeon_device *rdev,
 
 	radeon_ring_write(ring, PACKET0(UVD_SEMA_CMD, 0));
 	radeon_ring_write(ring, emit_wait ? 1 : 0);
+
+	return true;
 }
 
 /**
diff --git a/drivers/gpu/drm/radeon/uvd_v3_1.c b/drivers/gpu/drm/radeon/uvd_v3_1.c
index 5b6fa1f..d722db2 100644
--- a/drivers/gpu/drm/radeon/uvd_v3_1.c
+++ b/drivers/gpu/drm/radeon/uvd_v3_1.c
@@ -37,7 +37,7 @@
  *
  * Emit a semaphore command (either wait or signal) to the UVD ring.
  */
-void uvd_v3_1_semaphore_emit(struct radeon_device *rdev,
+bool uvd_v3_1_semaphore_emit(struct radeon_device *rdev,
 			     struct radeon_ring *ring,
 			     struct radeon_semaphore *semaphore,
 			     bool emit_wait)
@@ -52,4 +52,6 @@ void uvd_v3_1_semaphore_emit(struct radeon_device *rdev,
 
 	radeon_ring_write(ring, PACKET0(UVD_SEMA_CMD, 0));
 	radeon_ring_write(ring, 0x80 | (emit_wait ? 1 : 0));
+
+	return true;
 }