radv: add support for local bos. (v3)

This uses the new kernel interfaces for reduced cs overhead,
We only set the local flag for memory allocations that don't have
 a dedicated allocation and ones that aren't imports.

v2: add to all the internal buffer creation paths.
v3: missed some command submission paths, handle 0/empty bo lists.

Reviewed-by: Bas Nieuwenhuizen <[email protected]>
Signed-off-by: Dave Airlie <[email protected]>

11 files changed, 51 insertions, 22 deletions

diff --git a/src/amd/vulkan/radv_cmd_buffer.c b/src/amd/vulkan/radv_cmd_buffer.c
index 08a05277fa5..1c276168f83 100644
--- a/src/amd/vulkan/radv_cmd_buffer.c
+++ b/src/amd/vulkan/radv_cmd_buffer.c
@@ -313,7 +313,8 @@ radv_cmd_buffer_resize_upload_buf(struct radv_cmd_buffer *cmd_buffer,
 	bo = device->ws->buffer_create(device->ws,
 				       new_size, 4096,
 				       RADEON_DOMAIN_GTT,
-				       RADEON_FLAG_CPU_ACCESS);
+				       RADEON_FLAG_CPU_ACCESS|
+				       RADEON_FLAG_NO_INTERPROCESS_SHARING);
 
 	if (!bo) {
 		cmd_buffer->record_result = VK_ERROR_OUT_OF_DEVICE_MEMORY;
diff --git a/src/amd/vulkan/radv_debug.c b/src/amd/vulkan/radv_debug.c
index b69c05b64f3..cdf8e7a1145 100644
--- a/src/amd/vulkan/radv_debug.c
+++ b/src/amd/vulkan/radv_debug.c
@@ -61,7 +61,8 @@ radv_init_trace(struct radv_device *device)
 
 	device->trace_bo = ws->buffer_create(ws, TRACE_BO_SIZE, 8,
 					     RADEON_DOMAIN_VRAM,
-					     RADEON_FLAG_CPU_ACCESS);
+					     RADEON_FLAG_CPU_ACCESS|
+					     RADEON_FLAG_NO_INTERPROCESS_SHARING);
 	if (!device->trace_bo)
 		return false;
 
diff --git a/src/amd/vulkan/radv_descriptor_set.c b/src/amd/vulkan/radv_descriptor_set.c
index c6b736bb689..167944f4e2f 100644
--- a/src/amd/vulkan/radv_descriptor_set.c
+++ b/src/amd/vulkan/radv_descriptor_set.c
@@ -431,7 +431,7 @@ VkResult radv_CreateDescriptorPool(
 
 	if (bo_size) {
 		pool->bo = device->ws->buffer_create(device->ws, bo_size,
-							32, RADEON_DOMAIN_VRAM, 0);
+							32, RADEON_DOMAIN_VRAM, RADEON_FLAG_NO_INTERPROCESS_SHARING);
 		pool->mapped_ptr = (uint8_t*)device->ws->buffer_map(pool->bo);
 	}
 	pool->size = bo_size;
diff --git a/src/amd/vulkan/radv_device.c b/src/amd/vulkan/radv_device.c
index 19ff8fec647..d25e9c97ba8 100644
--- a/src/amd/vulkan/radv_device.c
+++ b/src/amd/vulkan/radv_device.c
@@ -1394,6 +1394,7 @@ radv_get_preamble_cs(struct radv_queue *queue,
 	unsigned tess_factor_ring_size = 0, tess_offchip_ring_size = 0;
 	unsigned max_offchip_buffers;
 	unsigned hs_offchip_param = 0;
+	uint32_t ring_bo_flags = RADEON_FLAG_NO_CPU_ACCESS | RADEON_FLAG_NO_INTERPROCESS_SHARING;
 	if (!queue->has_tess_rings) {
 		if (needs_tess_rings)
 			add_tess_rings = true;
@@ -1427,7 +1428,7 @@ radv_get_preamble_cs(struct radv_queue *queue,
 		                                              scratch_size,
 		                                              4096,
 		                                              RADEON_DOMAIN_VRAM,
-		                                              RADEON_FLAG_NO_CPU_ACCESS);
+		                                              ring_bo_flags);
 		if (!scratch_bo)
 			goto fail;
 	} else
@@ -1438,7 +1439,7 @@ radv_get_preamble_cs(struct radv_queue *queue,
 		                                                      compute_scratch_size,
 		                                                      4096,
 		                                                      RADEON_DOMAIN_VRAM,
-		                                                      RADEON_FLAG_NO_CPU_ACCESS);
+		                                                      ring_bo_flags);
 		if (!compute_scratch_bo)
 			goto fail;
 
@@ -1450,7 +1451,7 @@ radv_get_preamble_cs(struct radv_queue *queue,
 								esgs_ring_size,
 								4096,
 								RADEON_DOMAIN_VRAM,
-								RADEON_FLAG_NO_CPU_ACCESS);
+								ring_bo_flags);
 		if (!esgs_ring_bo)
 			goto fail;
 	} else {
@@ -1463,7 +1464,7 @@ radv_get_preamble_cs(struct radv_queue *queue,
 								gsvs_ring_size,
 								4096,
 								RADEON_DOMAIN_VRAM,
-								RADEON_FLAG_NO_CPU_ACCESS);
+								ring_bo_flags);
 		if (!gsvs_ring_bo)
 			goto fail;
 	} else {
@@ -1476,14 +1477,14 @@ radv_get_preamble_cs(struct radv_queue *queue,
 								       tess_factor_ring_size,
 								       256,
 								       RADEON_DOMAIN_VRAM,
-								       RADEON_FLAG_NO_CPU_ACCESS);
+								       ring_bo_flags);
 		if (!tess_factor_ring_bo)
 			goto fail;
 		tess_offchip_ring_bo = queue->device->ws->buffer_create(queue->device->ws,
 								       tess_offchip_ring_size,
 								       256,
 								       RADEON_DOMAIN_VRAM,
-								       RADEON_FLAG_NO_CPU_ACCESS);
+									ring_bo_flags);
 		if (!tess_offchip_ring_bo)
 			goto fail;
 	} else {
@@ -1510,7 +1511,7 @@ radv_get_preamble_cs(struct radv_queue *queue,
 		                                                 size,
 		                                                 4096,
 		                                                 RADEON_DOMAIN_VRAM,
-		                                                 RADEON_FLAG_CPU_ACCESS);
+		                                                 RADEON_FLAG_CPU_ACCESS|RADEON_FLAG_NO_INTERPROCESS_SHARING);
 		if (!descriptor_bo)
 			goto fail;
 	} else
@@ -2119,6 +2120,9 @@ VkResult radv_alloc_memory(VkDevice                        _device,
 	if (mem_flags & RADV_MEM_IMPLICIT_SYNC)
 		flags |= RADEON_FLAG_IMPLICIT_SYNC;
 
+	if (!dedicate_info && !import_info)
+		flags |= RADEON_FLAG_NO_INTERPROCESS_SHARING;
+
 	mem->bo = device->ws->buffer_create(device->ws, alloc_size, device->physical_device->rad_info.max_alignment,
 					       domain, flags);
 
@@ -2682,7 +2686,7 @@ VkResult radv_CreateEvent(
 
 	event->bo = device->ws->buffer_create(device->ws, 8, 8,
 					      RADEON_DOMAIN_GTT,
-					      RADEON_FLAG_VA_UNCACHED | RADEON_FLAG_CPU_ACCESS);
+					      RADEON_FLAG_VA_UNCACHED | RADEON_FLAG_CPU_ACCESS | RADEON_FLAG_NO_INTERPROCESS_SHARING);
 	if (!event->bo) {
 		vk_free2(&device->alloc, pAllocator, event);
 		return VK_ERROR_OUT_OF_DEVICE_MEMORY;
diff --git a/src/amd/vulkan/radv_query.c b/src/amd/vulkan/radv_query.c
index 06045d6b41b..06bf14ab680 100644
--- a/src/amd/vulkan/radv_query.c
+++ b/src/amd/vulkan/radv_query.c
@@ -780,7 +780,7 @@ VkResult radv_CreateQueryPool(
 		size += 4 * pCreateInfo->queryCount;
 
 	pool->bo = device->ws->buffer_create(device->ws, size,
-					     64, RADEON_DOMAIN_GTT, 0);
+					     64, RADEON_DOMAIN_GTT, RADEON_FLAG_NO_INTERPROCESS_SHARING);
 
 	if (!pool->bo) {
 		vk_free2(&device->alloc, pAllocator, pool);
diff --git a/src/amd/vulkan/radv_radeon_winsys.h b/src/amd/vulkan/radv_radeon_winsys.h
index cf5a9e8f069..395c8499b3d 100644
--- a/src/amd/vulkan/radv_radeon_winsys.h
+++ b/src/amd/vulkan/radv_radeon_winsys.h
@@ -54,6 +54,7 @@ enum radeon_bo_flag { /* bitfield */
 	RADEON_FLAG_VIRTUAL =       (1 << 3),
 	RADEON_FLAG_VA_UNCACHED =   (1 << 4),
 	RADEON_FLAG_IMPLICIT_SYNC = (1 << 5),
+	RADEON_FLAG_NO_INTERPROCESS_SHARING = (1 << 6),
 };
 
 enum radeon_bo_usage { /* bitfield */
diff --git a/src/amd/vulkan/radv_shader.c b/src/amd/vulkan/radv_shader.c
index 59039170687..7f10798fdf4 100644
--- a/src/amd/vulkan/radv_shader.c
+++ b/src/amd/vulkan/radv_shader.c
@@ -325,7 +325,7 @@ radv_alloc_shader_memory(struct radv_device *device,
 
 	slab->size = 256 * 1024;
 	slab->bo = device->ws->buffer_create(device->ws, slab->size, 256,
-	                                     RADEON_DOMAIN_VRAM, 0);
+	                                     RADEON_DOMAIN_VRAM, RADEON_FLAG_NO_INTERPROCESS_SHARING);
 	slab->ptr = (char*)device->ws->buffer_map(slab->bo);
 	list_inithead(&slab->shaders);
 
diff --git a/src/amd/vulkan/si_cmd_buffer.c b/src/amd/vulkan/si_cmd_buffer.c
index 20144d39ea3..89ee399817d 100644
--- a/src/amd/vulkan/si_cmd_buffer.c
+++ b/src/amd/vulkan/si_cmd_buffer.c
@@ -571,7 +571,8 @@ cik_create_gfx_config(struct radv_device *device)
 	device->gfx_init = device->ws->buffer_create(device->ws,
 						     cs->cdw * 4, 4096,
 						     RADEON_DOMAIN_GTT,
-						     RADEON_FLAG_CPU_ACCESS);
+						     RADEON_FLAG_CPU_ACCESS|
+						     RADEON_FLAG_NO_INTERPROCESS_SHARING);
 	if (!device->gfx_init)
 		goto fail;
 
diff --git a/src/amd/vulkan/winsys/amdgpu/radv_amdgpu_bo.c b/src/amd/vulkan/winsys/amdgpu/radv_amdgpu_bo.c
index 15099b318e7..dac549a20ad 100644
--- a/src/amd/vulkan/winsys/amdgpu/radv_amdgpu_bo.c
+++ b/src/amd/vulkan/winsys/amdgpu/radv_amdgpu_bo.c
@@ -332,6 +332,10 @@ radv_amdgpu_winsys_bo_create(struct radeon_winsys *_ws,
 		request.flags |= AMDGPU_GEM_CREATE_CPU_GTT_USWC;
 	if (!(flags & RADEON_FLAG_IMPLICIT_SYNC) && ws->info.drm_minor >= 22)
 		request.flags |= AMDGPU_GEM_CREATE_EXPLICIT_SYNC;
+	if (flags & RADEON_FLAG_NO_INTERPROCESS_SHARING && ws->info.drm_minor >= 20) {
+		bo->is_local = true;
+		request.flags |= AMDGPU_GEM_CREATE_VM_ALWAYS_VALID;
+	}
 
 	/* this won't do anything on pre 4.9 kernels */
 	if (ws->zero_all_vram_allocs && (initial_domain & RADEON_DOMAIN_VRAM))
diff --git a/src/amd/vulkan/winsys/amdgpu/radv_amdgpu_bo.h b/src/amd/vulkan/winsys/amdgpu/radv_amdgpu_bo.h
index f32e4308386..f9aac9451c0 100644
--- a/src/amd/vulkan/winsys/amdgpu/radv_amdgpu_bo.h
+++ b/src/amd/vulkan/winsys/amdgpu/radv_amdgpu_bo.h
@@ -45,6 +45,7 @@ struct radv_amdgpu_winsys_bo {
 	uint64_t size;
 	struct radv_amdgpu_winsys *ws;
 	bool is_virtual;
+	bool is_local;
 	int ref_count;
 
 	union {
diff --git a/src/amd/vulkan/winsys/amdgpu/radv_amdgpu_cs.c b/src/amd/vulkan/winsys/amdgpu/radv_amdgpu_cs.c
index 46e5b767033..939c221e0c8 100644
--- a/src/amd/vulkan/winsys/amdgpu/radv_amdgpu_cs.c
+++ b/src/amd/vulkan/winsys/amdgpu/radv_amdgpu_cs.c
@@ -202,7 +202,8 @@ radv_amdgpu_cs_create(struct radeon_winsys *ws,
 	if (cs->ws->use_ib_bos) {
 		cs->ib_buffer = ws->buffer_create(ws, ib_size, 0,
 						RADEON_DOMAIN_GTT,
-						RADEON_FLAG_CPU_ACCESS);
+						RADEON_FLAG_CPU_ACCESS|
+						  RADEON_FLAG_NO_INTERPROCESS_SHARING);
 		if (!cs->ib_buffer) {
 			free(cs);
 			return NULL;
@@ -287,7 +288,8 @@ static void radv_amdgpu_cs_grow(struct radeon_winsys_cs *_cs, size_t min_size)
 
 	cs->ib_buffer = cs->ws->base.buffer_create(&cs->ws->base, ib_size, 0,
 						   RADEON_DOMAIN_GTT,
-						   RADEON_FLAG_CPU_ACCESS);
+						   RADEON_FLAG_CPU_ACCESS|
+						   RADEON_FLAG_NO_INTERPROCESS_SHARING);
 
 	if (!cs->ib_buffer) {
 		cs->base.cdw = 0;
@@ -471,6 +473,9 @@ static void radv_amdgpu_cs_add_buffer(struct radeon_winsys_cs *_cs,
 		return;
 	}
 
+	if (bo->is_local)
+		return;
+
 	radv_amdgpu_cs_add_buffer_internal(cs, bo->bo, priority);
 }
 
@@ -541,6 +546,10 @@ static int radv_amdgpu_create_bo_list(struct radv_amdgpu_winsys *ws,
 	} else if (count == 1 && !extra_bo && !extra_cs &&
 	           !radv_amdgpu_cs(cs_array[0])->num_virtual_buffers) {
 		struct radv_amdgpu_cs *cs = (struct radv_amdgpu_cs*)cs_array[0];
+		if (cs->num_buffers == 0) {
+			*bo_list = 0;
+			return 0;
+		}
 		r = amdgpu_bo_list_create(ws->dev, cs->num_buffers, cs->handles,
 					  cs->priorities, bo_list);
 	} else {
@@ -556,7 +565,10 @@ static int radv_amdgpu_create_bo_list(struct radv_amdgpu_winsys *ws,
 		if (extra_cs) {
 			total_buffer_count += ((struct radv_amdgpu_cs*)extra_cs)->num_buffers;
 		}
-
+		if (total_buffer_count == 0) {
+			*bo_list = 0;
+			return 0;
+		}
 		amdgpu_bo_handle *handles = malloc(sizeof(amdgpu_bo_handle) * total_buffer_count);
 		uint8_t *priorities = malloc(sizeof(uint8_t) * total_buffer_count);
 		if (!handles || !priorities) {
@@ -721,7 +733,8 @@ static int radv_amdgpu_winsys_cs_submit_chained(struct radeon_winsys_ctx *_ctx,
 					"see dmesg for more information.\n");
 	}
 
-	amdgpu_bo_list_destroy(bo_list);
+	if (bo_list)
+		amdgpu_bo_list_destroy(bo_list);
 
 	if (fence)
 		radv_amdgpu_request_to_fence(ctx, fence, &request);
@@ -795,7 +808,8 @@ static int radv_amdgpu_winsys_cs_submit_fallback(struct radeon_winsys_ctx *_ctx,
 						"see dmesg for more information.\n");
 		}
 
-		amdgpu_bo_list_destroy(bo_list);
+		if (bo_list)
+			amdgpu_bo_list_destroy(bo_list);
 
 		if (r)
 			return r;
@@ -856,7 +870,7 @@ static int radv_amdgpu_winsys_cs_submit_sysmem(struct radeon_winsys_ctx *_ctx,
 		}
 		assert(cnt);
 
-		bo = ws->buffer_create(ws, 4 * size, 4096, RADEON_DOMAIN_GTT, RADEON_FLAG_CPU_ACCESS);
+		bo = ws->buffer_create(ws, 4 * size, 4096, RADEON_DOMAIN_GTT, RADEON_FLAG_CPU_ACCESS|RADEON_FLAG_NO_INTERPROCESS_SHARING);
 		ptr = ws->buffer_map(bo);
 
 		if (preamble_cs) {
@@ -905,7 +919,8 @@ static int radv_amdgpu_winsys_cs_submit_sysmem(struct radeon_winsys_ctx *_ctx,
 						"see dmesg for more information.\n");
 		}
 
-		amdgpu_bo_list_destroy(bo_list);
+		if (bo_list)
+			amdgpu_bo_list_destroy(bo_list);
 
 		ws->buffer_destroy(bo);
 		if (r)
@@ -1038,7 +1053,8 @@ static struct radeon_winsys_ctx *radv_amdgpu_ctx_create(struct radeon_winsys *_w
 	assert(AMDGPU_HW_IP_NUM * MAX_RINGS_PER_TYPE * sizeof(uint64_t) <= 4096);
 	ctx->fence_bo = ws->base.buffer_create(&ws->base, 4096, 8,
 	                                      RADEON_DOMAIN_GTT,
-	                                      RADEON_FLAG_CPU_ACCESS);
+	                                      RADEON_FLAG_CPU_ACCESS|
+					       RADEON_FLAG_NO_INTERPROCESS_SHARING);
 	if (ctx->fence_bo)
 		ctx->fence_map = (uint64_t*)ws->base.buffer_map(ctx->fence_bo);
 	if (ctx->fence_map)