struct drm_i915_gem_relocation_entry *entry;
int index;
- const uint32_t domain =
- target_bo->is_winsys_bo ? I915_GEM_DOMAIN_RENDER : 0;
-
VkResult result = anv_reloc_list_grow(list, alloc, 1);
if (result != VK_SUCCESS)
return result;
entry->delta = delta;
entry->offset = offset;
entry->presumed_offset = target_bo->offset;
- entry->read_domains = domain;
- entry->write_domain = domain;
+ entry->read_domains = 0;
+ entry->write_domain = 0;
VG(VALGRIND_CHECK_MEM_IS_DEFINED(entry, sizeof(*entry)));
return VK_SUCCESS;
* Functions related to anv_batch_bo
*-----------------------------------------------------------------------*/
-static inline struct anv_batch_bo *
+static struct anv_batch_bo *
anv_cmd_buffer_current_batch_bo(struct anv_cmd_buffer *cmd_buffer)
{
return LIST_ENTRY(struct anv_batch_bo, cmd_buffer->batch_bos.prev, link);
struct anv_address
anv_cmd_buffer_surface_base_address(struct anv_cmd_buffer *cmd_buffer)
{
+ struct anv_state *bt_block = u_vector_head(&cmd_buffer->bt_block_states);
return (struct anv_address) {
- .bo = &cmd_buffer->device->surface_state_block_pool.bo,
- .offset = *(int32_t *)u_vector_head(&cmd_buffer->bt_blocks),
+ .bo = &cmd_buffer->device->surface_state_pool.block_pool.bo,
+ .offset = bt_block->offset,
};
}
anv_cmd_buffer_alloc_binding_table(struct anv_cmd_buffer *cmd_buffer,
uint32_t entries, uint32_t *state_offset)
{
- struct anv_block_pool *block_pool =
- &cmd_buffer->device->surface_state_block_pool;
- int32_t *bt_block = u_vector_head(&cmd_buffer->bt_blocks);
+ struct anv_state_pool *state_pool = &cmd_buffer->device->surface_state_pool;
+ struct anv_state *bt_block = u_vector_head(&cmd_buffer->bt_block_states);
struct anv_state state;
state.alloc_size = align_u32(entries * 4, 32);
- if (cmd_buffer->bt_next + state.alloc_size > block_pool->block_size)
+ if (cmd_buffer->bt_next + state.alloc_size > state_pool->block_size)
return (struct anv_state) { 0 };
state.offset = cmd_buffer->bt_next;
- state.map = block_pool->map + *bt_block + state.offset;
+ state.map = state_pool->block_pool.map + bt_block->offset + state.offset;
cmd_buffer->bt_next += state.alloc_size;
- assert(*bt_block < 0);
- *state_offset = -(*bt_block);
+ assert(bt_block->offset < 0);
+ *state_offset = -bt_block->offset;
return state;
}
VkResult
anv_cmd_buffer_new_binding_table_block(struct anv_cmd_buffer *cmd_buffer)
{
- struct anv_block_pool *block_pool =
- &cmd_buffer->device->surface_state_block_pool;
+ struct anv_state_pool *state_pool = &cmd_buffer->device->surface_state_pool;
- int32_t *offset = u_vector_add(&cmd_buffer->bt_blocks);
- if (offset == NULL) {
+ struct anv_state *bt_block = u_vector_add(&cmd_buffer->bt_block_states);
+ if (bt_block == NULL) {
anv_batch_set_error(&cmd_buffer->batch, VK_ERROR_OUT_OF_HOST_MEMORY);
return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
}
- *offset = anv_block_pool_alloc_back(block_pool);
+ *bt_block = anv_state_pool_alloc_back(state_pool);
cmd_buffer->bt_next = 0;
return VK_SUCCESS;
*(struct anv_batch_bo **)u_vector_add(&cmd_buffer->seen_bbos) = batch_bo;
- success = u_vector_init(&cmd_buffer->bt_blocks, sizeof(int32_t),
- 8 * sizeof(int32_t));
+ /* u_vector requires power-of-two size elements */
+ unsigned pow2_state_size = util_next_power_of_two(sizeof(struct anv_state));
+ success = u_vector_init(&cmd_buffer->bt_block_states,
+ pow2_state_size, 8 * pow2_state_size);
if (!success)
goto fail_seen_bbos;
return VK_SUCCESS;
fail_bt_blocks:
- u_vector_finish(&cmd_buffer->bt_blocks);
+ u_vector_finish(&cmd_buffer->bt_block_states);
fail_seen_bbos:
u_vector_finish(&cmd_buffer->seen_bbos);
fail_batch_bo:
void
anv_cmd_buffer_fini_batch_bo_chain(struct anv_cmd_buffer *cmd_buffer)
{
- int32_t *bt_block;
- u_vector_foreach(bt_block, &cmd_buffer->bt_blocks) {
- anv_block_pool_free(&cmd_buffer->device->surface_state_block_pool,
- *bt_block);
- }
- u_vector_finish(&cmd_buffer->bt_blocks);
+ struct anv_state *bt_block;
+ u_vector_foreach(bt_block, &cmd_buffer->bt_block_states)
+ anv_state_pool_free(&cmd_buffer->device->surface_state_pool, *bt_block);
+ u_vector_finish(&cmd_buffer->bt_block_states);
anv_reloc_list_finish(&cmd_buffer->surface_relocs, &cmd_buffer->pool->alloc);
&cmd_buffer->batch,
GEN8_MI_BATCH_BUFFER_START_length * 4);
- while (u_vector_length(&cmd_buffer->bt_blocks) > 1) {
- int32_t *bt_block = u_vector_remove(&cmd_buffer->bt_blocks);
- anv_block_pool_free(&cmd_buffer->device->surface_state_block_pool,
- *bt_block);
+ while (u_vector_length(&cmd_buffer->bt_block_states) > 1) {
+ struct anv_state *bt_block = u_vector_remove(&cmd_buffer->bt_block_states);
+ anv_state_pool_free(&cmd_buffer->device->surface_state_pool, *bt_block);
}
- assert(u_vector_length(&cmd_buffer->bt_blocks) == 1);
+ assert(u_vector_length(&cmd_buffer->bt_block_states) == 1);
cmd_buffer->bt_next = 0;
cmd_buffer->surface_relocs.num_relocs = 0;
}
}
-static inline VkResult
+static VkResult
anv_cmd_buffer_add_seen_bbos(struct anv_cmd_buffer *cmd_buffer,
struct list_head *list)
{
/* Allocated length of the 'objects' and 'bos' arrays */
uint32_t array_length;
+
+ uint32_t fence_count;
+ uint32_t fence_array_length;
+ struct drm_i915_gem_exec_fence * fences;
+ struct anv_syncobj ** syncobjs;
};
static void
{
vk_free(alloc, exec->objects);
vk_free(alloc, exec->bos);
+ vk_free(alloc, exec->fences);
+ vk_free(alloc, exec->syncobjs);
}
static VkResult
anv_execbuf_add_bo(struct anv_execbuf *exec,
struct anv_bo *bo,
struct anv_reloc_list *relocs,
+ uint32_t extra_flags,
const VkAllocationCallbacks *alloc)
{
struct drm_i915_gem_exec_object2 *obj = NULL;
obj->relocs_ptr = 0;
obj->alignment = 0;
obj->offset = bo->offset;
- obj->flags = bo->is_winsys_bo ? EXEC_OBJECT_WRITE : 0;
+ obj->flags = bo->flags | extra_flags;
obj->rsvd1 = 0;
obj->rsvd2 = 0;
}
obj->relocs_ptr = (uintptr_t) relocs->relocs;
for (size_t i = 0; i < relocs->num_relocs; i++) {
+ VkResult result;
+
/* A quick sanity check on relocations */
assert(relocs->relocs[i].offset < bo->size);
- anv_execbuf_add_bo(exec, relocs->reloc_bos[i], NULL, alloc);
+ result = anv_execbuf_add_bo(exec, relocs->reloc_bos[i], NULL,
+ extra_flags, alloc);
+
+ if (result != VK_SUCCESS)
+ return result;
}
}
return VK_SUCCESS;
}
+static VkResult
+anv_execbuf_add_syncobj(struct anv_execbuf *exec,
+ uint32_t handle, uint32_t flags,
+ const VkAllocationCallbacks *alloc)
+{
+ assert(flags != 0);
+
+ if (exec->fence_count >= exec->fence_array_length) {
+ uint32_t new_len = MAX2(exec->fence_array_length * 2, 64);
+
+ exec->fences = vk_realloc(alloc, exec->fences,
+ new_len * sizeof(*exec->fences),
+ 8, VK_SYSTEM_ALLOCATION_SCOPE_COMMAND);
+ if (exec->fences == NULL)
+ return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
+
+ exec->fence_array_length = new_len;
+ }
+
+ exec->fences[exec->fence_count] = (struct drm_i915_gem_exec_fence) {
+ .handle = handle,
+ .flags = flags,
+ };
+
+ exec->fence_count++;
+
+ return VK_SUCCESS;
+}
+
static void
anv_cmd_buffer_process_relocs(struct anv_cmd_buffer *cmd_buffer,
struct anv_reloc_list *list)
{
unsigned reloc_size = 0;
if (device->info.gen >= 8) {
- /* From the Broadwell PRM Vol. 2a, MI_LOAD_REGISTER_MEM::MemoryAddress:
- *
- * "This field specifies the address of the memory location where the
- * register value specified in the DWord above will read from. The
- * address specifies the DWord location of the data. Range =
- * GraphicsVirtualAddress[63:2] for a DWord register GraphicsAddress
- * [63:48] are ignored by the HW and assumed to be in correct
- * canonical form [63:48] == [47]."
- */
- const int shift = 63 - 47;
reloc_size = sizeof(uint64_t);
- *(uint64_t *)p = (((int64_t)v) << shift) >> shift;
+ *(uint64_t *)p = gen_canonical_address(v);
} else {
reloc_size = sizeof(uint32_t);
*(uint32_t *)p = v;
}
if (flush && !device->info.has_llc)
- anv_flush_range(p, reloc_size);
+ gen_flush_range(p, reloc_size);
}
static void
-adjust_relocations_from_state_pool(struct anv_block_pool *pool,
+adjust_relocations_from_state_pool(struct anv_state_pool *pool,
struct anv_reloc_list *relocs,
uint32_t last_pool_center_bo_offset)
{
- assert(last_pool_center_bo_offset <= pool->center_bo_offset);
- uint32_t delta = pool->center_bo_offset - last_pool_center_bo_offset;
+ assert(last_pool_center_bo_offset <= pool->block_pool.center_bo_offset);
+ uint32_t delta = pool->block_pool.center_bo_offset - last_pool_center_bo_offset;
for (size_t i = 0; i < relocs->num_relocs; i++) {
/* All of the relocations from this block pool to other BO's should
}
static void
-adjust_relocations_to_state_pool(struct anv_block_pool *pool,
+adjust_relocations_to_state_pool(struct anv_state_pool *pool,
struct anv_bo *from_bo,
struct anv_reloc_list *relocs,
uint32_t last_pool_center_bo_offset)
{
- assert(last_pool_center_bo_offset <= pool->center_bo_offset);
- uint32_t delta = pool->center_bo_offset - last_pool_center_bo_offset;
+ assert(last_pool_center_bo_offset <= pool->block_pool.center_bo_offset);
+ uint32_t delta = pool->block_pool.center_bo_offset - last_pool_center_bo_offset;
/* When we initially emit relocations into a block pool, we don't
* actually know what the final center_bo_offset will be so we just emit
* relocations that point to the pool bo with the correct offset.
*/
for (size_t i = 0; i < relocs->num_relocs; i++) {
- if (relocs->reloc_bos[i] == &pool->bo) {
+ if (relocs->reloc_bos[i] == &pool->block_pool.bo) {
/* Adjust the delta value in the relocation to correctly
* correspond to the new delta. Initially, this value may have
* been negative (if treated as unsigned), but we trust in
* use by the GPU at the moment.
*/
assert(relocs->relocs[i].offset < from_bo->size);
- write_reloc(pool->device, from_bo->map + relocs->relocs[i].offset,
+ write_reloc(pool->block_pool.device,
+ from_bo->map + relocs->relocs[i].offset,
relocs->relocs[i].presumed_offset +
relocs->relocs[i].delta, false);
}
* given time. The only option is to always relocate them.
*/
anv_reloc_list_apply(cmd_buffer->device, &cmd_buffer->surface_relocs,
- &cmd_buffer->device->surface_state_block_pool.bo,
+ &cmd_buffer->device->surface_state_pool.block_pool.bo,
true /* always relocate surface states */);
/* Since we own all of the batch buffers, we know what values are stored
return true;
}
-VkResult
-anv_cmd_buffer_execbuf(struct anv_device *device,
- struct anv_cmd_buffer *cmd_buffer)
+static VkResult
+setup_execbuf_for_cmd_buffer(struct anv_execbuf *execbuf,
+ struct anv_cmd_buffer *cmd_buffer)
{
struct anv_batch *batch = &cmd_buffer->batch;
- struct anv_block_pool *ss_pool =
- &cmd_buffer->device->surface_state_block_pool;
-
- struct anv_execbuf execbuf;
- anv_execbuf_init(&execbuf);
+ struct anv_state_pool *ss_pool =
+ &cmd_buffer->device->surface_state_pool;
adjust_relocations_from_state_pool(ss_pool, &cmd_buffer->surface_relocs,
cmd_buffer->last_ss_pool_center);
- VkResult result =
- anv_execbuf_add_bo(&execbuf, &ss_pool->bo, &cmd_buffer->surface_relocs,
- &cmd_buffer->pool->alloc);
+ VkResult result = anv_execbuf_add_bo(execbuf, &ss_pool->block_pool.bo,
+ &cmd_buffer->surface_relocs, 0,
+ &cmd_buffer->device->alloc);
if (result != VK_SUCCESS)
return result;
adjust_relocations_to_state_pool(ss_pool, &(*bbo)->bo, &(*bbo)->relocs,
cmd_buffer->last_ss_pool_center);
- anv_execbuf_add_bo(&execbuf, &(*bbo)->bo, &(*bbo)->relocs,
- &cmd_buffer->pool->alloc);
+ result = anv_execbuf_add_bo(execbuf, &(*bbo)->bo, &(*bbo)->relocs, 0,
+ &cmd_buffer->device->alloc);
+ if (result != VK_SUCCESS)
+ return result;
}
/* Now that we've adjusted all of the surface state relocations, we need to
* record the surface state pool center so future executions of the command
* buffer can adjust correctly.
*/
- cmd_buffer->last_ss_pool_center = ss_pool->center_bo_offset;
+ cmd_buffer->last_ss_pool_center = ss_pool->block_pool.center_bo_offset;
struct anv_batch_bo *first_batch_bo =
list_first_entry(&cmd_buffer->batch_bos, struct anv_batch_bo, link);
* corresponding to the first batch_bo in the chain with the last
* element in the list.
*/
- if (first_batch_bo->bo.index != execbuf.bo_count - 1) {
+ if (first_batch_bo->bo.index != execbuf->bo_count - 1) {
uint32_t idx = first_batch_bo->bo.index;
- uint32_t last_idx = execbuf.bo_count - 1;
+ uint32_t last_idx = execbuf->bo_count - 1;
- struct drm_i915_gem_exec_object2 tmp_obj = execbuf.objects[idx];
- assert(execbuf.bos[idx] == &first_batch_bo->bo);
+ struct drm_i915_gem_exec_object2 tmp_obj = execbuf->objects[idx];
+ assert(execbuf->bos[idx] == &first_batch_bo->bo);
- execbuf.objects[idx] = execbuf.objects[last_idx];
- execbuf.bos[idx] = execbuf.bos[last_idx];
- execbuf.bos[idx]->index = idx;
+ execbuf->objects[idx] = execbuf->objects[last_idx];
+ execbuf->bos[idx] = execbuf->bos[last_idx];
+ execbuf->bos[idx]->index = idx;
- execbuf.objects[last_idx] = tmp_obj;
- execbuf.bos[last_idx] = &first_batch_bo->bo;
+ execbuf->objects[last_idx] = tmp_obj;
+ execbuf->bos[last_idx] = &first_batch_bo->bo;
first_batch_bo->bo.index = last_idx;
}
}
}
- execbuf.execbuf = (struct drm_i915_gem_execbuffer2) {
- .buffers_ptr = (uintptr_t) execbuf.objects,
- .buffer_count = execbuf.bo_count,
+ execbuf->execbuf = (struct drm_i915_gem_execbuffer2) {
+ .buffers_ptr = (uintptr_t) execbuf->objects,
+ .buffer_count = execbuf->bo_count,
.batch_start_offset = 0,
.batch_len = batch->next - batch->start,
.cliprects_ptr = 0,
.num_cliprects = 0,
.DR1 = 0,
.DR4 = 0,
- .flags = I915_EXEC_HANDLE_LUT | I915_EXEC_RENDER |
- I915_EXEC_CONSTANTS_REL_GENERAL,
+ .flags = I915_EXEC_HANDLE_LUT | I915_EXEC_RENDER,
.rsvd1 = cmd_buffer->device->context_id,
.rsvd2 = 0,
};
- if (relocate_cmd_buffer(cmd_buffer, &execbuf)) {
+ if (relocate_cmd_buffer(cmd_buffer, execbuf)) {
/* If we were able to successfully relocate everything, tell the kernel
* that it can skip doing relocations. The requirement for using
* NO_RELOC is:
* the RENDER_SURFACE_STATE matches presumed_offset, so it should be
* safe for the kernel to relocate them as needed.
*/
- execbuf.execbuf.flags |= I915_EXEC_NO_RELOC;
+ execbuf->execbuf.flags |= I915_EXEC_NO_RELOC;
} else {
/* In the case where we fall back to doing kernel relocations, we need
* to ensure that the relocation list is valid. All relocations on the
cmd_buffer->surface_relocs.relocs[i].presumed_offset = -1;
}
+ return VK_SUCCESS;
+}
+
+static VkResult
+setup_empty_execbuf(struct anv_execbuf *execbuf, struct anv_device *device)
+{
+ VkResult result = anv_execbuf_add_bo(execbuf, &device->trivial_batch_bo,
+ NULL, 0, &device->alloc);
+ if (result != VK_SUCCESS)
+ return result;
+
+ execbuf->execbuf = (struct drm_i915_gem_execbuffer2) {
+ .buffers_ptr = (uintptr_t) execbuf->objects,
+ .buffer_count = execbuf->bo_count,
+ .batch_start_offset = 0,
+ .batch_len = 8, /* GEN7_MI_BATCH_BUFFER_END and NOOP */
+ .flags = I915_EXEC_HANDLE_LUT | I915_EXEC_RENDER,
+ .rsvd1 = device->context_id,
+ .rsvd2 = 0,
+ };
+
+ return VK_SUCCESS;
+}
+
+VkResult
+anv_cmd_buffer_execbuf(struct anv_device *device,
+ struct anv_cmd_buffer *cmd_buffer,
+ const VkSemaphore *in_semaphores,
+ uint32_t num_in_semaphores,
+ const VkSemaphore *out_semaphores,
+ uint32_t num_out_semaphores,
+ VkFence _fence)
+{
+ ANV_FROM_HANDLE(anv_fence, fence, _fence);
+
+ struct anv_execbuf execbuf;
+ anv_execbuf_init(&execbuf);
+
+ int in_fence = -1;
+ VkResult result = VK_SUCCESS;
+ for (uint32_t i = 0; i < num_in_semaphores; i++) {
+ ANV_FROM_HANDLE(anv_semaphore, semaphore, in_semaphores[i]);
+ struct anv_semaphore_impl *impl =
+ semaphore->temporary.type != ANV_SEMAPHORE_TYPE_NONE ?
+ &semaphore->temporary : &semaphore->permanent;
+
+ switch (impl->type) {
+ case ANV_SEMAPHORE_TYPE_BO:
+ result = anv_execbuf_add_bo(&execbuf, impl->bo, NULL,
+ 0, &device->alloc);
+ if (result != VK_SUCCESS)
+ return result;
+ break;
+
+ case ANV_SEMAPHORE_TYPE_SYNC_FILE:
+ if (in_fence == -1) {
+ in_fence = impl->fd;
+ } else {
+ int merge = anv_gem_sync_file_merge(device, in_fence, impl->fd);
+ if (merge == -1)
+ return vk_error(VK_ERROR_INVALID_EXTERNAL_HANDLE);
+
+ close(impl->fd);
+ close(in_fence);
+ in_fence = merge;
+ }
+
+ impl->fd = -1;
+ break;
+
+ case ANV_SEMAPHORE_TYPE_DRM_SYNCOBJ:
+ result = anv_execbuf_add_syncobj(&execbuf, impl->syncobj,
+ I915_EXEC_FENCE_WAIT,
+ &device->alloc);
+ if (result != VK_SUCCESS)
+ return result;
+ break;
+
+ default:
+ break;
+ }
+ }
+
+ bool need_out_fence = false;
+ for (uint32_t i = 0; i < num_out_semaphores; i++) {
+ ANV_FROM_HANDLE(anv_semaphore, semaphore, out_semaphores[i]);
+
+ /* Under most circumstances, out fences won't be temporary. However,
+ * the spec does allow it for opaque_fd. From the Vulkan 1.0.53 spec:
+ *
+ * "If the import is temporary, the implementation must restore the
+ * semaphore to its prior permanent state after submitting the next
+ * semaphore wait operation."
+ *
+ * The spec says nothing whatsoever about signal operations on
+ * temporarily imported semaphores so it appears they are allowed.
+ * There are also CTS tests that require this to work.
+ */
+ struct anv_semaphore_impl *impl =
+ semaphore->temporary.type != ANV_SEMAPHORE_TYPE_NONE ?
+ &semaphore->temporary : &semaphore->permanent;
+
+ switch (impl->type) {
+ case ANV_SEMAPHORE_TYPE_BO:
+ result = anv_execbuf_add_bo(&execbuf, impl->bo, NULL,
+ EXEC_OBJECT_WRITE, &device->alloc);
+ if (result != VK_SUCCESS)
+ return result;
+ break;
+
+ case ANV_SEMAPHORE_TYPE_SYNC_FILE:
+ need_out_fence = true;
+ break;
+
+ case ANV_SEMAPHORE_TYPE_DRM_SYNCOBJ:
+ result = anv_execbuf_add_syncobj(&execbuf, impl->syncobj,
+ I915_EXEC_FENCE_SIGNAL,
+ &device->alloc);
+ if (result != VK_SUCCESS)
+ return result;
+ break;
+
+ default:
+ break;
+ }
+ }
+
+ if (fence) {
+ /* Under most circumstances, out fences won't be temporary. However,
+ * the spec does allow it for opaque_fd. From the Vulkan 1.0.53 spec:
+ *
+ * "If the import is temporary, the implementation must restore the
+ * semaphore to its prior permanent state after submitting the next
+ * semaphore wait operation."
+ *
+ * The spec says nothing whatsoever about signal operations on
+ * temporarily imported semaphores so it appears they are allowed.
+ * There are also CTS tests that require this to work.
+ */
+ struct anv_fence_impl *impl =
+ fence->temporary.type != ANV_FENCE_TYPE_NONE ?
+ &fence->temporary : &fence->permanent;
+
+ switch (impl->type) {
+ case ANV_FENCE_TYPE_BO:
+ result = anv_execbuf_add_bo(&execbuf, &impl->bo.bo, NULL,
+ EXEC_OBJECT_WRITE, &device->alloc);
+ if (result != VK_SUCCESS)
+ return result;
+ break;
+
+ case ANV_FENCE_TYPE_SYNCOBJ:
+ result = anv_execbuf_add_syncobj(&execbuf, impl->syncobj,
+ I915_EXEC_FENCE_SIGNAL,
+ &device->alloc);
+ if (result != VK_SUCCESS)
+ return result;
+ break;
+
+ default:
+ unreachable("Invalid fence type");
+ }
+ }
+
+ if (cmd_buffer)
+ result = setup_execbuf_for_cmd_buffer(&execbuf, cmd_buffer);
+ else
+ result = setup_empty_execbuf(&execbuf, device);
+
+ if (result != VK_SUCCESS)
+ return result;
+
+ if (execbuf.fence_count > 0) {
+ assert(device->instance->physicalDevice.has_syncobj);
+ execbuf.execbuf.flags |= I915_EXEC_FENCE_ARRAY;
+ execbuf.execbuf.num_cliprects = execbuf.fence_count;
+ execbuf.execbuf.cliprects_ptr = (uintptr_t) execbuf.fences;
+ }
+
+ if (in_fence != -1) {
+ execbuf.execbuf.flags |= I915_EXEC_FENCE_IN;
+ execbuf.execbuf.rsvd2 |= (uint32_t)in_fence;
+ }
+
+ if (need_out_fence)
+ execbuf.execbuf.flags |= I915_EXEC_FENCE_OUT;
+
result = anv_device_execbuf(device, &execbuf.execbuf, execbuf.bos);
- anv_execbuf_finish(&execbuf, &cmd_buffer->pool->alloc);
+ /* Execbuf does not consume the in_fence. It's our job to close it. */
+ if (in_fence != -1)
+ close(in_fence);
+
+ for (uint32_t i = 0; i < num_in_semaphores; i++) {
+ ANV_FROM_HANDLE(anv_semaphore, semaphore, in_semaphores[i]);
+ /* From the Vulkan 1.0.53 spec:
+ *
+ * "If the import is temporary, the implementation must restore the
+ * semaphore to its prior permanent state after submitting the next
+ * semaphore wait operation."
+ *
+ * This has to happen after the execbuf in case we close any syncobjs in
+ * the process.
+ */
+ anv_semaphore_reset_temporary(device, semaphore);
+ }
+
+ if (fence && fence->permanent.type == ANV_FENCE_TYPE_BO) {
+ /* BO fences can't be shared, so they can't be temporary. */
+ assert(fence->temporary.type == ANV_FENCE_TYPE_NONE);
+
+ /* Once the execbuf has returned, we need to set the fence state to
+ * SUBMITTED. We can't do this before calling execbuf because
+ * anv_GetFenceStatus does take the global device lock before checking
+ * fence->state.
+ *
+ * We set the fence state to SUBMITTED regardless of whether or not the
+ * execbuf succeeds because we need to ensure that vkWaitForFences() and
+ * vkGetFenceStatus() return a valid result (VK_ERROR_DEVICE_LOST or
+ * VK_SUCCESS) in a finite amount of time even if execbuf fails.
+ */
+ fence->permanent.bo.state = ANV_BO_FENCE_STATE_SUBMITTED;
+ }
+
+ if (result == VK_SUCCESS && need_out_fence) {
+ int out_fence = execbuf.execbuf.rsvd2 >> 32;
+ for (uint32_t i = 0; i < num_out_semaphores; i++) {
+ ANV_FROM_HANDLE(anv_semaphore, semaphore, out_semaphores[i]);
+ /* Out fences can't have temporary state because that would imply
+ * that we imported a sync file and are trying to signal it.
+ */
+ assert(semaphore->temporary.type == ANV_SEMAPHORE_TYPE_NONE);
+ struct anv_semaphore_impl *impl = &semaphore->permanent;
+
+ if (impl->type == ANV_SEMAPHORE_TYPE_SYNC_FILE) {
+ assert(impl->fd == -1);
+ impl->fd = dup(out_fence);
+ }
+ }
+ close(out_fence);
+ }
+
+ anv_execbuf_finish(&execbuf, &device->alloc);
return result;
}