drm/amdgpu: optionally do a writeback but don't invalidate TC for IB fences

There is a new IB flag that enables this new behavior.
Full invalidation is unnecessary for RELEASE_MEM and doesn't make sense
when draw calls from two adjacent gfx IBs run in parallel. This will be
the new default for Mesa.

v2: bump the version

Signed-off-by: Marek Olšák <marek.olsak@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c

@@ -75,9 +75,10 @@
  * - 3.23.0 - Add query for VRAM lost counter
  * - 3.24.0 - Add high priority compute support for gfx9
  * - 3.25.0 - Add support for sensor query info (stable pstate sclk/mclk).
+ * - 3.26.0 - GFX9: Process AMDGPU_IB_FLAG_TC_WB_NOT_INVALIDATE.
  */
 #define KMS_DRIVER_MAJOR	3
-#define KMS_DRIVER_MINOR	25
+#define KMS_DRIVER_MINOR	26
 #define KMS_DRIVER_PATCHLEVEL	0
 
 int amdgpu_vram_limit = 0;

drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c

@@ -131,7 +131,8 @@ static u32 amdgpu_fence_read(struct amdgpu_ring *ring)
  * Emits a fence command on the requested ring (all asics).
  * Returns 0 on success, -ENOMEM on failure.
  */
-int amdgpu_fence_emit(struct amdgpu_ring *ring, struct dma_fence **f)
+int amdgpu_fence_emit(struct amdgpu_ring *ring, struct dma_fence **f,
+		      unsigned flags)
 {
 	struct amdgpu_device *adev = ring->adev;
 	struct amdgpu_fence *fence;
@@ -149,7 +150,7 @@ int amdgpu_fence_emit(struct amdgpu_ring *ring, struct dma_fence **f)
 		       adev->fence_context + ring->idx,
 		       seq);
 	amdgpu_ring_emit_fence(ring, ring->fence_drv.gpu_addr,
-			       seq, AMDGPU_FENCE_FLAG_INT);
+			       seq, flags | AMDGPU_FENCE_FLAG_INT);
 
 	ptr = &ring->fence_drv.fences[seq & ring->fence_drv.num_fences_mask];
 	/* This function can't be called concurrently anyway, otherwise

drivers/gpu/drm/amd/amdgpu/amdgpu_ib.c

@@ -127,6 +127,7 @@ int amdgpu_ib_schedule(struct amdgpu_ring *ring, unsigned num_ibs,
 	struct amdgpu_vm *vm;
 	uint64_t fence_ctx;
 	uint32_t status = 0, alloc_size;
+	unsigned fence_flags = 0;
 
 	unsigned i;
 	int r = 0;
@@ -227,7 +228,10 @@ int amdgpu_ib_schedule(struct amdgpu_ring *ring, unsigned num_ibs,
 #endif
 		amdgpu_asic_invalidate_hdp(adev, ring);
 
-	r = amdgpu_fence_emit(ring, f);
+	if (ib->flags & AMDGPU_IB_FLAG_TC_WB_NOT_INVALIDATE)
+		fence_flags |= AMDGPU_FENCE_FLAG_TC_WB_ONLY;
+
+	r = amdgpu_fence_emit(ring, f, fence_flags);
 	if (r) {
 		dev_err(adev->dev, "failed to emit fence (%d)\n", r);
 		if (job && job->vmid)
@@ -242,7 +246,7 @@ int amdgpu_ib_schedule(struct amdgpu_ring *ring, unsigned num_ibs,
 	/* wrap the last IB with fence */
 	if (job && job->uf_addr) {
 		amdgpu_ring_emit_fence(ring, job->uf_addr, job->uf_sequence,
-				       AMDGPU_FENCE_FLAG_64BIT);
+				       fence_flags | AMDGPU_FENCE_FLAG_64BIT);
 	}
 
 	if (patch_offset != ~0 && ring->funcs->patch_cond_exec)

drivers/gpu/drm/amd/amdgpu/amdgpu_ring.h

@@ -42,6 +42,7 @@
 
 #define AMDGPU_FENCE_FLAG_64BIT         (1 << 0)
 #define AMDGPU_FENCE_FLAG_INT           (1 << 1)
+#define AMDGPU_FENCE_FLAG_TC_WB_ONLY    (1 << 2)
 
 enum amdgpu_ring_type {
 	AMDGPU_RING_TYPE_GFX,
@@ -90,7 +91,8 @@ int amdgpu_fence_driver_start_ring(struct amdgpu_ring *ring,
 				   unsigned irq_type);
 void amdgpu_fence_driver_suspend(struct amdgpu_device *adev);
 void amdgpu_fence_driver_resume(struct amdgpu_device *adev);
-int amdgpu_fence_emit(struct amdgpu_ring *ring, struct dma_fence **fence);
+int amdgpu_fence_emit(struct amdgpu_ring *ring, struct dma_fence **fence,
+		      unsigned flags);
 int amdgpu_fence_emit_polling(struct amdgpu_ring *ring, uint32_t *s);
 void amdgpu_fence_process(struct amdgpu_ring *ring);
 int amdgpu_fence_wait_empty(struct amdgpu_ring *ring);

drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c

@@ -633,7 +633,7 @@ int amdgpu_vm_flush(struct amdgpu_ring *ring, struct amdgpu_job *job, bool need_
 		amdgpu_gmc_emit_pasid_mapping(ring, job->vmid, job->pasid);
 
 	if (vm_flush_needed || pasid_mapping_needed) {
-		r = amdgpu_fence_emit(ring, &fence);
+		r = amdgpu_fence_emit(ring, &fence, 0);
 		if (r)
 			return r;
 	}

drivers/gpu/drm/amd/amdgpu/gfx_v9_0.c

@@ -3775,13 +3775,16 @@ static void gfx_v9_0_ring_emit_fence(struct amdgpu_ring *ring, u64 addr,
 {
 	bool write64bit = flags & AMDGPU_FENCE_FLAG_64BIT;
 	bool int_sel = flags & AMDGPU_FENCE_FLAG_INT;
+	bool writeback = flags & AMDGPU_FENCE_FLAG_TC_WB_ONLY;
 
 	/* RELEASE_MEM - flush caches, send int */
 	amdgpu_ring_write(ring, PACKET3(PACKET3_RELEASE_MEM, 6));
-	amdgpu_ring_write(ring, (EOP_TCL1_ACTION_EN |
+	amdgpu_ring_write(ring, ((writeback ? (EOP_TC_WB_ACTION_EN |
-				 EOP_TC_ACTION_EN |
+					       EOP_TC_NC_ACTION_EN) :
-				 EOP_TC_WB_ACTION_EN |
+					      (EOP_TCL1_ACTION_EN |
-				 EOP_TC_MD_ACTION_EN |
+					       EOP_TC_ACTION_EN |
+					       EOP_TC_WB_ACTION_EN |
+					       EOP_TC_MD_ACTION_EN)) |
 				 EVENT_TYPE(CACHE_FLUSH_AND_INV_TS_EVENT) |
 				 EVENT_INDEX(5)));
 	amdgpu_ring_write(ring, DATA_SEL(write64bit ? 2 : 1) | INT_SEL(int_sel ? 2 : 0));

drivers/gpu/drm/amd/amdgpu/soc15d.h

@@ -159,6 +159,7 @@
 #define		EOP_TC_WB_ACTION_EN                     (1 << 15) /* L2 */
 #define		EOP_TCL1_ACTION_EN                      (1 << 16)
 #define		EOP_TC_ACTION_EN                        (1 << 17) /* L2 */
+#define		EOP_TC_NC_ACTION_EN			(1 << 19)
 #define		EOP_TC_MD_ACTION_EN			(1 << 21) /* L2 metadata */
 
 #define		DATA_SEL(x)                             ((x) << 29)

include/uapi/drm/amdgpu_drm.h

@@ -526,6 +526,10 @@ union drm_amdgpu_cs {
 /* Preempt flag, IB should set Pre_enb bit if PREEMPT flag detected */
 #define AMDGPU_IB_FLAG_PREEMPT (1<<2)
 
+/* The IB fence should do the L2 writeback but not invalidate any shader
+ * caches (L2/vL1/sL1/I$). */
+#define AMDGPU_IB_FLAG_TC_WB_NOT_INVALIDATE (1 << 3)
+
 struct drm_amdgpu_cs_chunk_ib {
 	__u32 _pad;
 	/** AMDGPU_IB_FLAG_* */