2 * Copyright © 2015 Intel Corporation
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
11 * The above copyright notice and this permission notice (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
20 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
25 * This file implements VkQueue, VkFence, and VkSemaphore
30 #include <sys/eventfd.h>
32 #include "anv_private.h"
35 #include "genxml/gen7_pack.h"
38 anv_device_execbuf(struct anv_device
*device
,
39 struct drm_i915_gem_execbuffer2
*execbuf
,
40 struct anv_bo
**execbuf_bos
)
42 int ret
= anv_gem_execbuffer(device
, execbuf
);
44 /* We don't know the real error. */
46 return vk_errorf(VK_ERROR_DEVICE_LOST
, "execbuf2 failed: %m");
49 struct drm_i915_gem_exec_object2
*objects
=
50 (void *)(uintptr_t)execbuf
->buffers_ptr
;
51 for (uint32_t k
= 0; k
< execbuf
->buffer_count
; k
++)
52 execbuf_bos
[k
]->offset
= objects
[k
].offset
;
58 anv_device_submit_simple_batch(struct anv_device
*device
,
59 struct anv_batch
*batch
)
61 struct drm_i915_gem_execbuffer2 execbuf
;
62 struct drm_i915_gem_exec_object2 exec2_objects
[1];
63 struct anv_bo bo
, *exec_bos
[1];
64 VkResult result
= VK_SUCCESS
;
67 /* Kernel driver requires 8 byte aligned batch length */
68 size
= align_u32(batch
->next
- batch
->start
, 8);
69 result
= anv_bo_pool_alloc(&device
->batch_bo_pool
, &bo
, size
);
70 if (result
!= VK_SUCCESS
)
73 memcpy(bo
.map
, batch
->start
, size
);
74 if (!device
->info
.has_llc
)
75 gen_flush_range(bo
.map
, size
);
78 exec2_objects
[0].handle
= bo
.gem_handle
;
79 exec2_objects
[0].relocation_count
= 0;
80 exec2_objects
[0].relocs_ptr
= 0;
81 exec2_objects
[0].alignment
= 0;
82 exec2_objects
[0].offset
= bo
.offset
;
83 exec2_objects
[0].flags
= 0;
84 exec2_objects
[0].rsvd1
= 0;
85 exec2_objects
[0].rsvd2
= 0;
87 execbuf
.buffers_ptr
= (uintptr_t) exec2_objects
;
88 execbuf
.buffer_count
= 1;
89 execbuf
.batch_start_offset
= 0;
90 execbuf
.batch_len
= size
;
91 execbuf
.cliprects_ptr
= 0;
92 execbuf
.num_cliprects
= 0;
97 I915_EXEC_HANDLE_LUT
| I915_EXEC_NO_RELOC
| I915_EXEC_RENDER
;
98 execbuf
.rsvd1
= device
->context_id
;
101 result
= anv_device_execbuf(device
, &execbuf
, exec_bos
);
102 if (result
!= VK_SUCCESS
)
105 result
= anv_device_wait(device
, &bo
, INT64_MAX
);
108 anv_bo_pool_free(&device
->batch_bo_pool
, &bo
);
113 VkResult
anv_QueueSubmit(
115 uint32_t submitCount
,
116 const VkSubmitInfo
* pSubmits
,
119 ANV_FROM_HANDLE(anv_queue
, queue
, _queue
);
120 struct anv_device
*device
= queue
->device
;
122 /* Query for device status prior to submitting. Technically, we don't need
123 * to do this. However, if we have a client that's submitting piles of
124 * garbage, we would rather break as early as possible to keep the GPU
125 * hanging contained. If we don't check here, we'll either be waiting for
126 * the kernel to kick us or we'll have to wait until the client waits on a
127 * fence before we actually know whether or not we've hung.
129 VkResult result
= anv_device_query_status(device
);
130 if (result
!= VK_SUCCESS
)
133 /* We lock around QueueSubmit for three main reasons:
135 * 1) When a block pool is resized, we create a new gem handle with a
136 * different size and, in the case of surface states, possibly a
137 * different center offset but we re-use the same anv_bo struct when
138 * we do so. If this happens in the middle of setting up an execbuf,
139 * we could end up with our list of BOs out of sync with our list of
142 * 2) The algorithm we use for building the list of unique buffers isn't
143 * thread-safe. While the client is supposed to syncronize around
144 * QueueSubmit, this would be extremely difficult to debug if it ever
145 * came up in the wild due to a broken app. It's better to play it
146 * safe and just lock around QueueSubmit.
148 * 3) The anv_cmd_buffer_execbuf function may perform relocations in
149 * userspace. Due to the fact that the surface state buffer is shared
150 * between batches, we can't afford to have that happen from multiple
151 * threads at the same time. Even though the user is supposed to
152 * ensure this doesn't happen, we play it safe as in (2) above.
154 * Since the only other things that ever take the device lock such as block
155 * pool resize only rarely happen, this will almost never be contended so
156 * taking a lock isn't really an expensive operation in this case.
158 pthread_mutex_lock(&device
->mutex
);
160 if (fence
&& submitCount
== 0) {
161 /* If we don't have any command buffers, we need to submit a dummy
162 * batch to give GEM something to wait on. We could, potentially,
163 * come up with something more efficient but this shouldn't be a
166 result
= anv_cmd_buffer_execbuf(device
, NULL
, NULL
, 0, NULL
, 0, fence
);
170 for (uint32_t i
= 0; i
< submitCount
; i
++) {
171 /* Fence for this submit. NULL for all but the last one */
172 VkFence submit_fence
= (i
== submitCount
- 1) ? fence
: NULL
;
174 if (pSubmits
[i
].commandBufferCount
== 0) {
175 /* If we don't have any command buffers, we need to submit a dummy
176 * batch to give GEM something to wait on. We could, potentially,
177 * come up with something more efficient but this shouldn't be a
180 result
= anv_cmd_buffer_execbuf(device
, NULL
,
181 pSubmits
[i
].pWaitSemaphores
,
182 pSubmits
[i
].waitSemaphoreCount
,
183 pSubmits
[i
].pSignalSemaphores
,
184 pSubmits
[i
].signalSemaphoreCount
,
186 if (result
!= VK_SUCCESS
)
192 for (uint32_t j
= 0; j
< pSubmits
[i
].commandBufferCount
; j
++) {
193 ANV_FROM_HANDLE(anv_cmd_buffer
, cmd_buffer
,
194 pSubmits
[i
].pCommandBuffers
[j
]);
195 assert(cmd_buffer
->level
== VK_COMMAND_BUFFER_LEVEL_PRIMARY
);
196 assert(!anv_batch_has_error(&cmd_buffer
->batch
));
198 /* Fence for this execbuf. NULL for all but the last one */
199 VkFence execbuf_fence
=
200 (j
== pSubmits
[i
].commandBufferCount
- 1) ? submit_fence
: NULL
;
202 const VkSemaphore
*in_semaphores
= NULL
, *out_semaphores
= NULL
;
203 uint32_t num_in_semaphores
= 0, num_out_semaphores
= 0;
205 /* Only the first batch gets the in semaphores */
206 in_semaphores
= pSubmits
[i
].pWaitSemaphores
;
207 num_in_semaphores
= pSubmits
[i
].waitSemaphoreCount
;
210 if (j
== pSubmits
[i
].commandBufferCount
- 1) {
211 /* Only the last batch gets the out semaphores */
212 out_semaphores
= pSubmits
[i
].pSignalSemaphores
;
213 num_out_semaphores
= pSubmits
[i
].signalSemaphoreCount
;
216 result
= anv_cmd_buffer_execbuf(device
, cmd_buffer
,
217 in_semaphores
, num_in_semaphores
,
218 out_semaphores
, num_out_semaphores
,
220 if (result
!= VK_SUCCESS
)
225 pthread_cond_broadcast(&device
->queue_submit
);
228 if (result
!= VK_SUCCESS
) {
229 /* In the case that something has gone wrong we may end up with an
230 * inconsistent state from which it may not be trivial to recover.
231 * For example, we might have computed address relocations and
232 * any future attempt to re-submit this job will need to know about
233 * this and avoid computing relocation addresses again.
235 * To avoid this sort of issues, we assume that if something was
236 * wrong during submission we must already be in a really bad situation
237 * anyway (such us being out of memory) and return
238 * VK_ERROR_DEVICE_LOST to ensure that clients do not attempt to
239 * submit the same job again to this device.
241 result
= vk_errorf(VK_ERROR_DEVICE_LOST
, "vkQueueSubmit() failed");
245 pthread_mutex_unlock(&device
->mutex
);
250 VkResult
anv_QueueWaitIdle(
253 ANV_FROM_HANDLE(anv_queue
, queue
, _queue
);
255 return anv_DeviceWaitIdle(anv_device_to_handle(queue
->device
));
258 VkResult
anv_CreateFence(
260 const VkFenceCreateInfo
* pCreateInfo
,
261 const VkAllocationCallbacks
* pAllocator
,
264 ANV_FROM_HANDLE(anv_device
, device
, _device
);
265 struct anv_fence
*fence
;
267 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_FENCE_CREATE_INFO
);
269 fence
= vk_zalloc2(&device
->alloc
, pAllocator
, sizeof(*fence
), 8,
270 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
272 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
274 if (device
->instance
->physicalDevice
.has_syncobj_wait
) {
275 fence
->permanent
.type
= ANV_FENCE_TYPE_SYNCOBJ
;
277 uint32_t create_flags
= 0;
278 if (pCreateInfo
->flags
& VK_FENCE_CREATE_SIGNALED_BIT
)
279 create_flags
|= DRM_SYNCOBJ_CREATE_SIGNALED
;
281 fence
->permanent
.syncobj
= anv_gem_syncobj_create(device
, create_flags
);
282 if (!fence
->permanent
.syncobj
)
283 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
285 fence
->permanent
.type
= ANV_FENCE_TYPE_BO
;
287 VkResult result
= anv_bo_pool_alloc(&device
->batch_bo_pool
,
288 &fence
->permanent
.bo
.bo
, 4096);
289 if (result
!= VK_SUCCESS
)
292 if (pCreateInfo
->flags
& VK_FENCE_CREATE_SIGNALED_BIT
) {
293 fence
->permanent
.bo
.state
= ANV_BO_FENCE_STATE_SIGNALED
;
295 fence
->permanent
.bo
.state
= ANV_BO_FENCE_STATE_RESET
;
299 *pFence
= anv_fence_to_handle(fence
);
305 anv_fence_impl_cleanup(struct anv_device
*device
,
306 struct anv_fence_impl
*impl
)
308 switch (impl
->type
) {
309 case ANV_FENCE_TYPE_NONE
:
310 /* Dummy. Nothing to do */
313 case ANV_FENCE_TYPE_BO
:
314 anv_bo_pool_free(&device
->batch_bo_pool
, &impl
->bo
.bo
);
317 case ANV_FENCE_TYPE_SYNCOBJ
:
318 anv_gem_syncobj_destroy(device
, impl
->syncobj
);
322 unreachable("Invalid fence type");
325 void anv_DestroyFence(
328 const VkAllocationCallbacks
* pAllocator
)
330 ANV_FROM_HANDLE(anv_device
, device
, _device
);
331 ANV_FROM_HANDLE(anv_fence
, fence
, _fence
);
336 anv_fence_impl_cleanup(device
, &fence
->temporary
);
337 anv_fence_impl_cleanup(device
, &fence
->permanent
);
339 vk_free2(&device
->alloc
, pAllocator
, fence
);
342 VkResult
anv_ResetFences(
345 const VkFence
* pFences
)
347 ANV_FROM_HANDLE(anv_device
, device
, _device
);
349 for (uint32_t i
= 0; i
< fenceCount
; i
++) {
350 ANV_FROM_HANDLE(anv_fence
, fence
, pFences
[i
]);
352 /* From the Vulkan 1.0.53 spec:
354 * "If any member of pFences currently has its payload imported with
355 * temporary permanence, that fence’s prior permanent payload is
356 * first restored. The remaining operations described therefore
357 * operate on the restored payload.
359 if (fence
->temporary
.type
!= ANV_FENCE_TYPE_NONE
) {
360 anv_fence_impl_cleanup(device
, &fence
->temporary
);
361 fence
->temporary
.type
= ANV_FENCE_TYPE_NONE
;
364 struct anv_fence_impl
*impl
= &fence
->permanent
;
366 switch (impl
->type
) {
367 case ANV_FENCE_TYPE_BO
:
368 impl
->bo
.state
= ANV_BO_FENCE_STATE_RESET
;
371 case ANV_FENCE_TYPE_SYNCOBJ
:
372 anv_gem_syncobj_reset(device
, impl
->syncobj
);
376 unreachable("Invalid fence type");
383 VkResult
anv_GetFenceStatus(
387 ANV_FROM_HANDLE(anv_device
, device
, _device
);
388 ANV_FROM_HANDLE(anv_fence
, fence
, _fence
);
390 if (unlikely(device
->lost
))
391 return VK_ERROR_DEVICE_LOST
;
393 struct anv_fence_impl
*impl
=
394 fence
->temporary
.type
!= ANV_FENCE_TYPE_NONE
?
395 &fence
->temporary
: &fence
->permanent
;
397 switch (impl
->type
) {
398 case ANV_FENCE_TYPE_BO
:
399 /* BO fences don't support import/export */
400 assert(fence
->temporary
.type
== ANV_FENCE_TYPE_NONE
);
401 switch (impl
->bo
.state
) {
402 case ANV_BO_FENCE_STATE_RESET
:
403 /* If it hasn't even been sent off to the GPU yet, it's not ready */
406 case ANV_BO_FENCE_STATE_SIGNALED
:
407 /* It's been signaled, return success */
410 case ANV_BO_FENCE_STATE_SUBMITTED
: {
411 VkResult result
= anv_device_bo_busy(device
, &impl
->bo
.bo
);
412 if (result
== VK_SUCCESS
) {
413 impl
->bo
.state
= ANV_BO_FENCE_STATE_SIGNALED
;
420 unreachable("Invalid fence status");
423 case ANV_FENCE_TYPE_SYNCOBJ
: {
424 int ret
= anv_gem_syncobj_wait(device
, &impl
->syncobj
, 1, 0, true);
426 if (errno
== ETIME
) {
429 /* We don't know the real error. */
431 return vk_errorf(VK_ERROR_DEVICE_LOST
,
432 "drm_syncobj_wait failed: %m");
440 unreachable("Invalid fence type");
444 #define NSEC_PER_SEC 1000000000
445 #define INT_TYPE_MAX(type) ((1ull << (sizeof(type) * 8 - 1)) - 1)
450 struct timespec current
;
451 clock_gettime(CLOCK_MONOTONIC
, ¤t
);
452 return (uint64_t)current
.tv_sec
* NSEC_PER_SEC
+ current
.tv_nsec
;
456 anv_wait_for_syncobj_fences(struct anv_device
*device
,
458 const VkFence
*pFences
,
462 uint32_t *syncobjs
= vk_zalloc(&device
->alloc
,
463 sizeof(*syncobjs
) * fenceCount
, 8,
464 VK_SYSTEM_ALLOCATION_SCOPE_COMMAND
);
466 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
468 for (uint32_t i
= 0; i
< fenceCount
; i
++) {
469 ANV_FROM_HANDLE(anv_fence
, fence
, pFences
[i
]);
470 assert(fence
->permanent
.type
== ANV_FENCE_TYPE_SYNCOBJ
);
472 struct anv_fence_impl
*impl
=
473 fence
->temporary
.type
!= ANV_FENCE_TYPE_NONE
?
474 &fence
->temporary
: &fence
->permanent
;
476 assert(impl
->type
== ANV_FENCE_TYPE_SYNCOBJ
);
477 syncobjs
[i
] = impl
->syncobj
;
480 int64_t abs_timeout_ns
= 0;
482 uint64_t current_ns
= gettime_ns();
484 /* Add but saturate to INT32_MAX */
485 if (current_ns
+ _timeout
< current_ns
)
486 abs_timeout_ns
= INT64_MAX
;
487 else if (current_ns
+ _timeout
> INT64_MAX
)
488 abs_timeout_ns
= INT64_MAX
;
490 abs_timeout_ns
= current_ns
+ _timeout
;
493 /* The gem_syncobj_wait ioctl may return early due to an inherent
494 * limitation in the way it computes timeouts. Loop until we've actually
495 * passed the timeout.
499 ret
= anv_gem_syncobj_wait(device
, syncobjs
, fenceCount
,
500 abs_timeout_ns
, waitAll
);
501 } while (ret
== -1 && errno
== ETIME
&& gettime_ns() < abs_timeout_ns
);
503 vk_free(&device
->alloc
, syncobjs
);
506 if (errno
== ETIME
) {
509 /* We don't know the real error. */
511 return vk_errorf(VK_ERROR_DEVICE_LOST
,
512 "drm_syncobj_wait failed: %m");
520 anv_wait_for_bo_fences(struct anv_device
*device
,
522 const VkFence
*pFences
,
528 /* DRM_IOCTL_I915_GEM_WAIT uses a signed 64 bit timeout and is supposed
529 * to block indefinitely timeouts <= 0. Unfortunately, this was broken
530 * for a couple of kernel releases. Since there's no way to know
531 * whether or not the kernel we're using is one of the broken ones, the
532 * best we can do is to clamp the timeout to INT64_MAX. This limits the
533 * maximum timeout from 584 years to 292 years - likely not a big deal.
535 int64_t timeout
= MIN2(_timeout
, INT64_MAX
);
537 VkResult result
= VK_SUCCESS
;
538 uint32_t pending_fences
= fenceCount
;
539 while (pending_fences
) {
541 bool signaled_fences
= false;
542 for (uint32_t i
= 0; i
< fenceCount
; i
++) {
543 ANV_FROM_HANDLE(anv_fence
, fence
, pFences
[i
]);
545 /* This function assumes that all fences are BO fences and that they
546 * have no temporary state. Since BO fences will never be exported,
547 * this should be a safe assumption.
549 assert(fence
->permanent
.type
== ANV_FENCE_TYPE_BO
);
550 assert(fence
->temporary
.type
== ANV_FENCE_TYPE_NONE
);
551 struct anv_fence_impl
*impl
= &fence
->permanent
;
553 switch (impl
->bo
.state
) {
554 case ANV_BO_FENCE_STATE_RESET
:
555 /* This fence hasn't been submitted yet, we'll catch it the next
556 * time around. Yes, this may mean we dead-loop but, short of
557 * lots of locking and a condition variable, there's not much that
558 * we can do about that.
563 case ANV_BO_FENCE_STATE_SIGNALED
:
564 /* This fence is not pending. If waitAll isn't set, we can return
565 * early. Otherwise, we have to keep going.
573 case ANV_BO_FENCE_STATE_SUBMITTED
:
574 /* These are the fences we really care about. Go ahead and wait
575 * on it until we hit a timeout.
577 result
= anv_device_wait(device
, &impl
->bo
.bo
, timeout
);
580 impl
->bo
.state
= ANV_BO_FENCE_STATE_SIGNALED
;
581 signaled_fences
= true;
595 if (pending_fences
&& !signaled_fences
) {
596 /* If we've hit this then someone decided to vkWaitForFences before
597 * they've actually submitted any of them to a queue. This is a
598 * fairly pessimal case, so it's ok to lock here and use a standard
599 * pthreads condition variable.
601 pthread_mutex_lock(&device
->mutex
);
603 /* It's possible that some of the fences have changed state since the
604 * last time we checked. Now that we have the lock, check for
605 * pending fences again and don't wait if it's changed.
607 uint32_t now_pending_fences
= 0;
608 for (uint32_t i
= 0; i
< fenceCount
; i
++) {
609 ANV_FROM_HANDLE(anv_fence
, fence
, pFences
[i
]);
610 if (fence
->permanent
.bo
.state
== ANV_BO_FENCE_STATE_RESET
)
611 now_pending_fences
++;
613 assert(now_pending_fences
<= pending_fences
);
615 if (now_pending_fences
== pending_fences
) {
616 struct timespec before
;
617 clock_gettime(CLOCK_MONOTONIC
, &before
);
619 uint32_t abs_nsec
= before
.tv_nsec
+ timeout
% NSEC_PER_SEC
;
620 uint64_t abs_sec
= before
.tv_sec
+ (abs_nsec
/ NSEC_PER_SEC
) +
621 (timeout
/ NSEC_PER_SEC
);
622 abs_nsec
%= NSEC_PER_SEC
;
624 /* Avoid roll-over in tv_sec on 32-bit systems if the user
625 * provided timeout is UINT64_MAX
627 struct timespec abstime
;
628 abstime
.tv_nsec
= abs_nsec
;
629 abstime
.tv_sec
= MIN2(abs_sec
, INT_TYPE_MAX(abstime
.tv_sec
));
631 ret
= pthread_cond_timedwait(&device
->queue_submit
,
632 &device
->mutex
, &abstime
);
633 assert(ret
!= EINVAL
);
635 struct timespec after
;
636 clock_gettime(CLOCK_MONOTONIC
, &after
);
637 uint64_t time_elapsed
=
638 ((uint64_t)after
.tv_sec
* NSEC_PER_SEC
+ after
.tv_nsec
) -
639 ((uint64_t)before
.tv_sec
* NSEC_PER_SEC
+ before
.tv_nsec
);
641 if (time_elapsed
>= timeout
) {
642 pthread_mutex_unlock(&device
->mutex
);
647 timeout
-= time_elapsed
;
650 pthread_mutex_unlock(&device
->mutex
);
655 if (unlikely(device
->lost
))
656 return VK_ERROR_DEVICE_LOST
;
661 VkResult
anv_WaitForFences(
664 const VkFence
* pFences
,
668 ANV_FROM_HANDLE(anv_device
, device
, _device
);
670 if (unlikely(device
->lost
))
671 return VK_ERROR_DEVICE_LOST
;
673 if (device
->instance
->physicalDevice
.has_syncobj_wait
) {
674 return anv_wait_for_syncobj_fences(device
, fenceCount
, pFences
,
677 return anv_wait_for_bo_fences(device
, fenceCount
, pFences
,
682 void anv_GetPhysicalDeviceExternalFencePropertiesKHR(
683 VkPhysicalDevice physicalDevice
,
684 const VkPhysicalDeviceExternalFenceInfoKHR
* pExternalFenceInfo
,
685 VkExternalFencePropertiesKHR
* pExternalFenceProperties
)
687 ANV_FROM_HANDLE(anv_physical_device
, device
, physicalDevice
);
689 switch (pExternalFenceInfo
->handleType
) {
690 case VK_EXTERNAL_FENCE_HANDLE_TYPE_OPAQUE_FD_BIT_KHR
:
691 case VK_EXTERNAL_FENCE_HANDLE_TYPE_SYNC_FD_BIT_KHR
:
692 if (device
->has_syncobj_wait
) {
693 pExternalFenceProperties
->exportFromImportedHandleTypes
=
694 VK_EXTERNAL_FENCE_HANDLE_TYPE_OPAQUE_FD_BIT_KHR
|
695 VK_EXTERNAL_FENCE_HANDLE_TYPE_SYNC_FD_BIT_KHR
;
696 pExternalFenceProperties
->compatibleHandleTypes
=
697 VK_EXTERNAL_FENCE_HANDLE_TYPE_OPAQUE_FD_BIT_KHR
|
698 VK_EXTERNAL_FENCE_HANDLE_TYPE_SYNC_FD_BIT_KHR
;
699 pExternalFenceProperties
->externalFenceFeatures
=
700 VK_EXTERNAL_FENCE_FEATURE_EXPORTABLE_BIT_KHR
|
701 VK_EXTERNAL_FENCE_FEATURE_IMPORTABLE_BIT_KHR
;
710 pExternalFenceProperties
->exportFromImportedHandleTypes
= 0;
711 pExternalFenceProperties
->compatibleHandleTypes
= 0;
712 pExternalFenceProperties
->externalFenceFeatures
= 0;
715 VkResult
anv_ImportFenceFdKHR(
717 const VkImportFenceFdInfoKHR
* pImportFenceFdInfo
)
719 ANV_FROM_HANDLE(anv_device
, device
, _device
);
720 ANV_FROM_HANDLE(anv_fence
, fence
, pImportFenceFdInfo
->fence
);
721 int fd
= pImportFenceFdInfo
->fd
;
723 assert(pImportFenceFdInfo
->sType
==
724 VK_STRUCTURE_TYPE_IMPORT_FENCE_FD_INFO_KHR
);
726 struct anv_fence_impl new_impl
= {
727 .type
= ANV_FENCE_TYPE_NONE
,
730 switch (pImportFenceFdInfo
->handleType
) {
731 case VK_EXTERNAL_FENCE_HANDLE_TYPE_OPAQUE_FD_BIT_KHR
:
732 new_impl
.type
= ANV_FENCE_TYPE_SYNCOBJ
;
734 new_impl
.syncobj
= anv_gem_syncobj_fd_to_handle(device
, fd
);
735 if (!new_impl
.syncobj
)
736 return vk_error(VK_ERROR_INVALID_EXTERNAL_HANDLE_KHR
);
740 case VK_EXTERNAL_FENCE_HANDLE_TYPE_SYNC_FD_BIT_KHR
:
741 /* Sync files are a bit tricky. Because we want to continue using the
742 * syncobj implementation of WaitForFences, we don't use the sync file
743 * directly but instead import it into a syncobj.
745 new_impl
.type
= ANV_FENCE_TYPE_SYNCOBJ
;
747 new_impl
.syncobj
= anv_gem_syncobj_create(device
, 0);
748 if (!new_impl
.syncobj
)
749 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
751 if (anv_gem_syncobj_import_sync_file(device
, new_impl
.syncobj
, fd
)) {
752 anv_gem_syncobj_destroy(device
, new_impl
.syncobj
);
753 return vk_errorf(VK_ERROR_INVALID_EXTERNAL_HANDLE_KHR
,
754 "syncobj sync file import failed: %m");
759 return vk_error(VK_ERROR_INVALID_EXTERNAL_HANDLE_KHR
);
762 /* From the Vulkan 1.0.53 spec:
764 * "Importing a fence payload from a file descriptor transfers
765 * ownership of the file descriptor from the application to the
766 * Vulkan implementation. The application must not perform any
767 * operations on the file descriptor after a successful import."
769 * If the import fails, we leave the file descriptor open.
773 if (pImportFenceFdInfo
->flags
& VK_FENCE_IMPORT_TEMPORARY_BIT_KHR
) {
774 anv_fence_impl_cleanup(device
, &fence
->temporary
);
775 fence
->temporary
= new_impl
;
777 anv_fence_impl_cleanup(device
, &fence
->permanent
);
778 fence
->permanent
= new_impl
;
784 VkResult
anv_GetFenceFdKHR(
786 const VkFenceGetFdInfoKHR
* pGetFdInfo
,
789 ANV_FROM_HANDLE(anv_device
, device
, _device
);
790 ANV_FROM_HANDLE(anv_fence
, fence
, pGetFdInfo
->fence
);
792 assert(pGetFdInfo
->sType
== VK_STRUCTURE_TYPE_FENCE_GET_FD_INFO_KHR
);
794 struct anv_fence_impl
*impl
=
795 fence
->temporary
.type
!= ANV_FENCE_TYPE_NONE
?
796 &fence
->temporary
: &fence
->permanent
;
798 assert(impl
->type
== ANV_FENCE_TYPE_SYNCOBJ
);
799 switch (pGetFdInfo
->handleType
) {
800 case VK_EXTERNAL_FENCE_HANDLE_TYPE_OPAQUE_FD_BIT_KHR
: {
801 int fd
= anv_gem_syncobj_handle_to_fd(device
, impl
->syncobj
);
803 return vk_error(VK_ERROR_TOO_MANY_OBJECTS
);
809 case VK_EXTERNAL_FENCE_HANDLE_TYPE_SYNC_FD_BIT_KHR
: {
810 int fd
= anv_gem_syncobj_export_sync_file(device
, impl
->syncobj
);
812 return vk_error(VK_ERROR_TOO_MANY_OBJECTS
);
819 unreachable("Invalid fence export handle type");
822 /* From the Vulkan 1.0.53 spec:
824 * "Export operations have the same transference as the specified handle
825 * type’s import operations. [...] If the fence was using a
826 * temporarily imported payload, the fence’s prior permanent payload
829 if (impl
== &fence
->temporary
)
830 anv_fence_impl_cleanup(device
, impl
);
835 // Queue semaphore functions
837 VkResult
anv_CreateSemaphore(
839 const VkSemaphoreCreateInfo
* pCreateInfo
,
840 const VkAllocationCallbacks
* pAllocator
,
841 VkSemaphore
* pSemaphore
)
843 ANV_FROM_HANDLE(anv_device
, device
, _device
);
844 struct anv_semaphore
*semaphore
;
846 assert(pCreateInfo
->sType
== VK_STRUCTURE_TYPE_SEMAPHORE_CREATE_INFO
);
848 semaphore
= vk_alloc2(&device
->alloc
, pAllocator
, sizeof(*semaphore
), 8,
849 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
850 if (semaphore
== NULL
)
851 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
853 const VkExportSemaphoreCreateInfoKHR
*export
=
854 vk_find_struct_const(pCreateInfo
->pNext
, EXPORT_SEMAPHORE_CREATE_INFO_KHR
);
855 VkExternalSemaphoreHandleTypeFlagsKHR handleTypes
=
856 export
? export
->handleTypes
: 0;
858 if (handleTypes
== 0) {
859 /* The DRM execbuffer ioctl always execute in-oder so long as you stay
860 * on the same ring. Since we don't expose the blit engine as a DMA
861 * queue, a dummy no-op semaphore is a perfectly valid implementation.
863 semaphore
->permanent
.type
= ANV_SEMAPHORE_TYPE_DUMMY
;
864 } else if (handleTypes
& VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_FD_BIT_KHR
) {
865 assert(handleTypes
== VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_FD_BIT_KHR
);
866 if (device
->instance
->physicalDevice
.has_syncobj
) {
867 semaphore
->permanent
.type
= ANV_SEMAPHORE_TYPE_DRM_SYNCOBJ
;
868 semaphore
->permanent
.syncobj
= anv_gem_syncobj_create(device
, 0);
869 if (!semaphore
->permanent
.syncobj
) {
870 vk_free2(&device
->alloc
, pAllocator
, semaphore
);
871 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
874 semaphore
->permanent
.type
= ANV_SEMAPHORE_TYPE_BO
;
875 VkResult result
= anv_bo_cache_alloc(device
, &device
->bo_cache
,
876 4096, &semaphore
->permanent
.bo
);
877 if (result
!= VK_SUCCESS
) {
878 vk_free2(&device
->alloc
, pAllocator
, semaphore
);
882 /* If we're going to use this as a fence, we need to *not* have the
883 * EXEC_OBJECT_ASYNC bit set.
885 assert(!(semaphore
->permanent
.bo
->flags
& EXEC_OBJECT_ASYNC
));
887 } else if (handleTypes
& VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_SYNC_FD_BIT_KHR
) {
888 assert(handleTypes
== VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_SYNC_FD_BIT_KHR
);
890 semaphore
->permanent
.type
= ANV_SEMAPHORE_TYPE_SYNC_FILE
;
891 semaphore
->permanent
.fd
= -1;
893 assert(!"Unknown handle type");
894 vk_free2(&device
->alloc
, pAllocator
, semaphore
);
895 return vk_error(VK_ERROR_INVALID_EXTERNAL_HANDLE_KHR
);
898 semaphore
->temporary
.type
= ANV_SEMAPHORE_TYPE_NONE
;
900 *pSemaphore
= anv_semaphore_to_handle(semaphore
);
906 anv_semaphore_impl_cleanup(struct anv_device
*device
,
907 struct anv_semaphore_impl
*impl
)
909 switch (impl
->type
) {
910 case ANV_SEMAPHORE_TYPE_NONE
:
911 case ANV_SEMAPHORE_TYPE_DUMMY
:
912 /* Dummy. Nothing to do */
915 case ANV_SEMAPHORE_TYPE_BO
:
916 anv_bo_cache_release(device
, &device
->bo_cache
, impl
->bo
);
919 case ANV_SEMAPHORE_TYPE_SYNC_FILE
:
923 case ANV_SEMAPHORE_TYPE_DRM_SYNCOBJ
:
924 anv_gem_syncobj_destroy(device
, impl
->syncobj
);
928 unreachable("Invalid semaphore type");
932 anv_semaphore_reset_temporary(struct anv_device
*device
,
933 struct anv_semaphore
*semaphore
)
935 if (semaphore
->temporary
.type
== ANV_SEMAPHORE_TYPE_NONE
)
938 anv_semaphore_impl_cleanup(device
, &semaphore
->temporary
);
939 semaphore
->temporary
.type
= ANV_SEMAPHORE_TYPE_NONE
;
942 void anv_DestroySemaphore(
944 VkSemaphore _semaphore
,
945 const VkAllocationCallbacks
* pAllocator
)
947 ANV_FROM_HANDLE(anv_device
, device
, _device
);
948 ANV_FROM_HANDLE(anv_semaphore
, semaphore
, _semaphore
);
950 if (semaphore
== NULL
)
953 anv_semaphore_impl_cleanup(device
, &semaphore
->temporary
);
954 anv_semaphore_impl_cleanup(device
, &semaphore
->permanent
);
956 vk_free2(&device
->alloc
, pAllocator
, semaphore
);
959 void anv_GetPhysicalDeviceExternalSemaphorePropertiesKHR(
960 VkPhysicalDevice physicalDevice
,
961 const VkPhysicalDeviceExternalSemaphoreInfoKHR
* pExternalSemaphoreInfo
,
962 VkExternalSemaphorePropertiesKHR
* pExternalSemaphoreProperties
)
964 ANV_FROM_HANDLE(anv_physical_device
, device
, physicalDevice
);
966 switch (pExternalSemaphoreInfo
->handleType
) {
967 case VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_FD_BIT_KHR
:
968 pExternalSemaphoreProperties
->exportFromImportedHandleTypes
=
969 VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_FD_BIT_KHR
;
970 pExternalSemaphoreProperties
->compatibleHandleTypes
=
971 VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_FD_BIT_KHR
;
972 pExternalSemaphoreProperties
->externalSemaphoreFeatures
=
973 VK_EXTERNAL_SEMAPHORE_FEATURE_EXPORTABLE_BIT_KHR
|
974 VK_EXTERNAL_SEMAPHORE_FEATURE_IMPORTABLE_BIT_KHR
;
977 case VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_SYNC_FD_BIT_KHR
:
978 if (device
->has_exec_fence
) {
979 pExternalSemaphoreProperties
->exportFromImportedHandleTypes
= 0;
980 pExternalSemaphoreProperties
->compatibleHandleTypes
=
981 VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_SYNC_FD_BIT_KHR
;
982 pExternalSemaphoreProperties
->externalSemaphoreFeatures
=
983 VK_EXTERNAL_SEMAPHORE_FEATURE_EXPORTABLE_BIT_KHR
|
984 VK_EXTERNAL_SEMAPHORE_FEATURE_IMPORTABLE_BIT_KHR
;
993 pExternalSemaphoreProperties
->exportFromImportedHandleTypes
= 0;
994 pExternalSemaphoreProperties
->compatibleHandleTypes
= 0;
995 pExternalSemaphoreProperties
->externalSemaphoreFeatures
= 0;
998 VkResult
anv_ImportSemaphoreFdKHR(
1000 const VkImportSemaphoreFdInfoKHR
* pImportSemaphoreFdInfo
)
1002 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1003 ANV_FROM_HANDLE(anv_semaphore
, semaphore
, pImportSemaphoreFdInfo
->semaphore
);
1004 int fd
= pImportSemaphoreFdInfo
->fd
;
1006 struct anv_semaphore_impl new_impl
= {
1007 .type
= ANV_SEMAPHORE_TYPE_NONE
,
1010 switch (pImportSemaphoreFdInfo
->handleType
) {
1011 case VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_FD_BIT_KHR
:
1012 if (device
->instance
->physicalDevice
.has_syncobj
) {
1013 new_impl
.type
= ANV_SEMAPHORE_TYPE_DRM_SYNCOBJ
;
1015 new_impl
.syncobj
= anv_gem_syncobj_fd_to_handle(device
, fd
);
1016 if (!new_impl
.syncobj
)
1017 return vk_error(VK_ERROR_INVALID_EXTERNAL_HANDLE_KHR
);
1019 /* From the Vulkan spec:
1021 * "Importing semaphore state from a file descriptor transfers
1022 * ownership of the file descriptor from the application to the
1023 * Vulkan implementation. The application must not perform any
1024 * operations on the file descriptor after a successful import."
1026 * If the import fails, we leave the file descriptor open.
1028 close(pImportSemaphoreFdInfo
->fd
);
1030 new_impl
.type
= ANV_SEMAPHORE_TYPE_BO
;
1032 VkResult result
= anv_bo_cache_import(device
, &device
->bo_cache
,
1033 fd
, 4096, &new_impl
.bo
);
1034 if (result
!= VK_SUCCESS
)
1037 /* If we're going to use this as a fence, we need to *not* have the
1038 * EXEC_OBJECT_ASYNC bit set.
1040 assert(!(new_impl
.bo
->flags
& EXEC_OBJECT_ASYNC
));
1044 case VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_SYNC_FD_BIT_KHR
:
1045 new_impl
= (struct anv_semaphore_impl
) {
1046 .type
= ANV_SEMAPHORE_TYPE_SYNC_FILE
,
1052 return vk_error(VK_ERROR_INVALID_EXTERNAL_HANDLE_KHR
);
1055 if (pImportSemaphoreFdInfo
->flags
& VK_SEMAPHORE_IMPORT_TEMPORARY_BIT_KHR
) {
1056 anv_semaphore_impl_cleanup(device
, &semaphore
->temporary
);
1057 semaphore
->temporary
= new_impl
;
1059 anv_semaphore_impl_cleanup(device
, &semaphore
->permanent
);
1060 semaphore
->permanent
= new_impl
;
1066 VkResult
anv_GetSemaphoreFdKHR(
1068 const VkSemaphoreGetFdInfoKHR
* pGetFdInfo
,
1071 ANV_FROM_HANDLE(anv_device
, device
, _device
);
1072 ANV_FROM_HANDLE(anv_semaphore
, semaphore
, pGetFdInfo
->semaphore
);
1076 assert(pGetFdInfo
->sType
== VK_STRUCTURE_TYPE_SEMAPHORE_GET_FD_INFO_KHR
);
1078 struct anv_semaphore_impl
*impl
=
1079 semaphore
->temporary
.type
!= ANV_SEMAPHORE_TYPE_NONE
?
1080 &semaphore
->temporary
: &semaphore
->permanent
;
1082 switch (impl
->type
) {
1083 case ANV_SEMAPHORE_TYPE_BO
:
1084 result
= anv_bo_cache_export(device
, &device
->bo_cache
, impl
->bo
, pFd
);
1085 if (result
!= VK_SUCCESS
)
1089 case ANV_SEMAPHORE_TYPE_SYNC_FILE
:
1090 /* There are two reasons why this could happen:
1092 * 1) The user is trying to export without submitting something that
1093 * signals the semaphore. If this is the case, it's their bug so
1094 * what we return here doesn't matter.
1096 * 2) The kernel didn't give us a file descriptor. The most likely
1097 * reason for this is running out of file descriptors.
1100 return vk_error(VK_ERROR_TOO_MANY_OBJECTS
);
1104 /* From the Vulkan 1.0.53 spec:
1106 * "...exporting a semaphore payload to a handle with copy
1107 * transference has the same side effects on the source
1108 * semaphore’s payload as executing a semaphore wait operation."
1110 * In other words, it may still be a SYNC_FD semaphore, but it's now
1111 * considered to have been waited on and no longer has a sync file
1117 case ANV_SEMAPHORE_TYPE_DRM_SYNCOBJ
:
1118 fd
= anv_gem_syncobj_handle_to_fd(device
, impl
->syncobj
);
1120 return vk_error(VK_ERROR_TOO_MANY_OBJECTS
);
1125 return vk_error(VK_ERROR_INVALID_EXTERNAL_HANDLE_KHR
);
1128 /* From the Vulkan 1.0.53 spec:
1130 * "Export operations have the same transference as the specified handle
1131 * type’s import operations. [...] If the semaphore was using a
1132 * temporarily imported payload, the semaphore’s prior permanent payload
1135 if (impl
== &semaphore
->temporary
)
1136 anv_semaphore_impl_cleanup(device
, impl
);