r600g: userspace fence to avoid kernel call for testing bo busy status

Signed-off-by: Jerome Glisse <[email protected]>

6 files changed, 103 insertions, 47 deletions

diff --git a/src/gallium/drivers/r600/r600.h b/src/gallium/drivers/r600/r600.h
index 630177d6add..24e25cec0db 100644
--- a/src/gallium/drivers/r600/r600.h
+++ b/src/gallium/drivers/r600/r600.h
@@ -233,6 +233,10 @@ struct r600_context {
 	u32			*pm4;
 	struct list_head	query_list;
 	unsigned		num_query_running;
+	unsigned		fence;
+	struct list_head	fenced_bo;
+	unsigned		*cfence;
+	struct r600_bo		*fence_bo;
 };
 
 struct r600_draw {
diff --git a/src/gallium/winsys/r600/drm/evergreen_hw_context.c b/src/gallium/winsys/r600/drm/evergreen_hw_context.c
index 1355b079450..2093a2d09c1 100644
--- a/src/gallium/winsys/r600/drm/evergreen_hw_context.c
+++ b/src/gallium/winsys/r600/drm/evergreen_hw_context.c
@@ -613,6 +613,13 @@ int evergreen_context_init(struct r600_context *ctx, struct radeon *radeon)
 		r = -ENOMEM;
 		goto out_err;
 	}
+	/* save 16dwords space for fence mecanism */
+	ctx->pm4_ndwords -= 16;
+
+	r = r600_context_init_fence(ctx);
+	if (r) {
+		goto out_err;
+	}
 
 	/* init dirty list */
 	LIST_INITHEAD(&ctx->dirty);
diff --git a/src/gallium/winsys/r600/drm/r600_hw_context.c b/src/gallium/winsys/r600/drm/r600_hw_context.c
index b379499f060..7d81d734571 100644
--- a/src/gallium/winsys/r600/drm/r600_hw_context.c
+++ b/src/gallium/winsys/r600/drm/r600_hw_context.c
@@ -41,6 +41,44 @@
 
 #define GROUP_FORCE_NEW_BLOCK	0
 
+int r600_context_init_fence(struct r600_context *ctx)
+{
+	ctx->fence = 1;
+	ctx->fence_bo = r600_bo(ctx->radeon, 4096, 0, 0);
+	if (ctx->fence_bo == NULL) {
+		return -ENOMEM;
+	}
+	ctx->cfence = r600_bo_map(ctx->radeon, ctx->fence_bo, PB_USAGE_UNSYNCHRONIZED, NULL);
+	*ctx->cfence = 0;
+	LIST_INITHEAD(&ctx->fenced_bo);
+	return 0;
+}
+
+static void INLINE r600_context_update_fenced_list(struct r600_context *ctx)
+{
+	for (int i = 0; i < ctx->creloc; i++) {
+		if (!LIST_IS_EMPTY(&ctx->bo[i]->fencedlist))
+			LIST_DELINIT(&ctx->bo[i]->fencedlist);
+		LIST_ADDTAIL(&ctx->bo[i]->fencedlist, &ctx->fenced_bo);
+		ctx->bo[i]->fence = ctx->fence;
+		ctx->bo[i]->ctx = ctx;
+	}
+}
+
+static void INLINE r600_context_fence_wraparound(struct r600_context *ctx, unsigned fence)
+{
+	struct radeon_bo *bo, *tmp;
+
+	LIST_FOR_EACH_ENTRY_SAFE(bo, tmp, &ctx->fenced_bo, fencedlist) {
+		if (bo->fence <= *ctx->cfence) {
+			LIST_DELINIT(&bo->fencedlist);
+			bo->fence = 0;
+		} else {
+			bo->fence = fence;
+		}
+	}
+}
+
 int r600_context_add_block(struct r600_context *ctx, const struct r600_reg *reg, unsigned nreg)
 {
 	struct r600_block *block;
@@ -572,6 +610,9 @@ void r600_context_fini(struct r600_context *ctx)
 	}
 	free(ctx->reloc);
 	free(ctx->pm4);
+	if (ctx->fence_bo) {
+		r600_bo_reference(ctx->radeon, &ctx->fence_bo, NULL);
+	}
 	memset(ctx, 0, sizeof(struct r600_context));
 }
 
@@ -691,6 +732,13 @@ int r600_context_init(struct r600_context *ctx, struct radeon *radeon)
 		r = -ENOMEM;
 		goto out_err;
 	}
+	/* save 16dwords space for fence mecanism */
+	ctx->pm4_ndwords -= 16;
+
+	r = r600_context_init_fence(ctx);
+	if (r) {
+		goto out_err;
+	}
 
 	/* init dirty list */
 	LIST_INITHEAD(&ctx->dirty);
@@ -1019,6 +1067,7 @@ void r600_context_flush(struct r600_context *ctx)
 	struct drm_radeon_cs drmib;
 	struct drm_radeon_cs_chunk chunks[2];
 	uint64_t chunk_array[2];
+	unsigned fence;
 	int r;
 
 	if (!ctx->pm4_cdwords)
@@ -1028,6 +1077,18 @@ void r600_context_flush(struct r600_context *ctx)
 	r600_context_queries_suspend(ctx);
 
 	radeon_bo_pbmgr_flush_maps(ctx->radeon->kman);
+
+	/* emit fence */
+	ctx->pm4[ctx->pm4_cdwords++] = PKT3(PKT3_EVENT_WRITE_EOP, 4);
+	ctx->pm4[ctx->pm4_cdwords++] = EVENT_TYPE_CACHE_FLUSH_AND_INV_TS_EVENT | (5 << 8);
+	ctx->pm4[ctx->pm4_cdwords++] = 0;
+	ctx->pm4[ctx->pm4_cdwords++] = (1 << 29) | (0 << 24);
+	ctx->pm4[ctx->pm4_cdwords++] = ctx->fence;
+	ctx->pm4[ctx->pm4_cdwords++] = 0;
+	ctx->pm4[ctx->pm4_cdwords++] = PKT3(PKT3_NOP, 0);
+	ctx->pm4[ctx->pm4_cdwords++] = 0;
+	r600_context_bo_reloc(ctx, &ctx->pm4[ctx->pm4_cdwords - 1], ctx->fence_bo);
+
 #if 1
 	/* emit cs */
 	drmib.num_chunks = 2;
@@ -1043,8 +1104,18 @@ void r600_context_flush(struct r600_context *ctx)
 	r = drmCommandWriteRead(ctx->radeon->fd, DRM_RADEON_CS, &drmib,
 				sizeof(struct drm_radeon_cs));
 #endif
+
+	r600_context_update_fenced_list(ctx);
+
+	fence = ctx->fence + 1;
+	if (fence < ctx->fence) {
+		/* wrap around */
+		fence = 1;
+		r600_context_fence_wraparound(ctx, fence);
+	}
+	ctx->fence = fence;
+
 	/* restart */
-	radeon_bo_fencelist(ctx->radeon, ctx->bo, ctx->creloc);
 	for (int i = 0; i < ctx->creloc; i++) {
 		ctx->bo[i]->reloc = NULL;
 		ctx->bo[i]->last_flush = 0;
diff --git a/src/gallium/winsys/r600/drm/r600_priv.h b/src/gallium/winsys/r600/drm/r600_priv.h
index ea2cf347785..a693a5b5ab8 100644
--- a/src/gallium/winsys/r600/drm/r600_priv.h
+++ b/src/gallium/winsys/r600/drm/r600_priv.h
@@ -64,9 +64,9 @@ struct radeon_bo {
 	unsigned			map_count;
 	void				*data;
 	struct list_head		fencedlist;
+	unsigned			fence;
+	struct r600_context		*ctx;
 	boolean				shared;
-	int64_t				last_busy;
-	boolean				set_busy;
 	struct r600_reloc		*reloc;
 	unsigned			reloc_id;
 	unsigned			last_flush;
@@ -103,6 +103,7 @@ struct pb_buffer *radeon_bo_pb_create_buffer_from_handle(struct pb_manager *_mgr
 							 uint32_t handle);
 
 /* r600_hw_context.c */
+int r600_context_init_fence(struct r600_context *ctx);
 void r600_context_bo_reloc(struct r600_context *ctx, u32 *pm4, struct r600_bo *rbo);
 void r600_context_bo_flush(struct r600_context *ctx, unsigned flush_flags,
 				unsigned flush_mask, struct r600_bo *rbo);
diff --git a/src/gallium/winsys/r600/drm/r600d.h b/src/gallium/winsys/r600/drm/r600d.h
index ccc9ffaf8e3..d91f7737af3 100644
--- a/src/gallium/winsys/r600/drm/r600d.h
+++ b/src/gallium/winsys/r600/drm/r600d.h
@@ -91,6 +91,7 @@
 #define PKT3_SET_CTL_CONST                     0x6F
 #define PKT3_SURFACE_BASE_UPDATE               0x73
 
+#define EVENT_TYPE_CACHE_FLUSH_AND_INV_TS_EVENT 0x14
 #define EVENT_TYPE_ZPASS_DONE                  0x15
 #define EVENT_TYPE_CACHE_FLUSH_AND_INV_EVENT   0x16
 
diff --git a/src/gallium/winsys/r600/drm/radeon_bo.c b/src/gallium/winsys/r600/drm/radeon_bo.c
index 42a23f0ab8f..836d5d77e09 100644
--- a/src/gallium/winsys/r600/drm/radeon_bo.c
+++ b/src/gallium/winsys/r600/drm/radeon_bo.c
@@ -128,7 +128,6 @@ struct radeon_bo *radeon_bo(struct radeon *radeon, unsigned handle,
 	return bo;
 }
 
-
 static void radeon_bo_destroy(struct radeon *radeon, struct radeon_bo *bo)
 {
 	struct drm_gem_close args;
@@ -158,8 +157,14 @@ int radeon_bo_wait(struct radeon *radeon, struct radeon_bo *bo)
 	struct drm_radeon_gem_wait_idle args;
 	int ret;
 
-	if (LIST_IS_EMPTY(&bo->fencedlist) && !bo->shared)
+	if (!bo->fence && !bo->shared)
+		return 0;
+
+	if (bo->fence <= *bo->ctx->cfence) {
+		LIST_DELINIT(&bo->fencedlist);
+		bo->fence = 0;
 		return 0;
+	}
 
 	/* Zero out args to make valgrind happy */
 	memset(&args, 0, sizeof(args));
@@ -171,22 +176,20 @@ int radeon_bo_wait(struct radeon *radeon, struct radeon_bo *bo)
 	return ret;
 }
 
-#define BO_BUSY_BACKOFF 10000
-
 int radeon_bo_busy(struct radeon *radeon, struct radeon_bo *bo, uint32_t *domain)
 {
 	struct drm_radeon_gem_busy args;
 	int ret;
-	int64_t now;
-
-	now = os_time_get();
-	if (LIST_IS_EMPTY(&bo->fencedlist) && !bo->shared)
-		return 0;
-
-	if (bo->set_busy && (now - bo->last_busy < BO_BUSY_BACKOFF))
-		return -EBUSY;
 
-	bo->set_busy = FALSE;
+	if (!bo->shared) {
+		if (!bo->fence)
+			return 0;
+		if (bo->fence <= *bo->ctx->cfence) {
+			LIST_DELINIT(&bo->fencedlist);
+			bo->fence = 0;
+			return 0;
+		}
+	}
 
 	memset(&args, 0, sizeof(args));
 	args.handle = bo->handle;
@@ -195,37 +198,6 @@ int radeon_bo_busy(struct radeon *radeon, struct radeon_bo *bo, uint32_t *domain
 	ret = drmCommandWriteRead(radeon->fd, DRM_RADEON_GEM_BUSY,
 			&args, sizeof(args));
 
-	if (ret == 0) {
-		struct radeon_bo *entry, *tent;
-		LIST_FOR_EACH_ENTRY_SAFE(entry, tent, &bo->fencedlist, fencedlist) {
-			LIST_DELINIT(&entry->fencedlist);
-		}
-	} else {
-		bo->set_busy = TRUE;
-		bo->last_busy = now;
-	}
 	*domain = args.domain;
 	return ret;
 }
-
-int radeon_bo_fencelist(struct radeon *radeon, struct radeon_bo **bolist,
-			uint32_t num_bo)
-{
-	struct radeon_bo *first;
-	int i;
-
-	first = bolist[0];
-
-	if (!LIST_IS_EMPTY(&first->fencedlist))
-		LIST_DELINIT(&first->fencedlist);
-
-	LIST_INITHEAD(&first->fencedlist);
-
-	for (i = 1; i < num_bo; i++) {
-		if (!LIST_IS_EMPTY(&bolist[i]->fencedlist))
-			LIST_DELINIT(&bolist[i]->fencedlist);
-
-		LIST_ADDTAIL(&bolist[i]->fencedlist, &first->fencedlist);
-	}
-	return 0;
-}