Loading...
Note: File does not exist in v3.1.
1// SPDX-License-Identifier: GPL-2.0+
2/*
3 * Copyright (C) 2014-2018 Broadcom
4 * Copyright (C) 2023 Raspberry Pi
5 */
6
7#include <drm/drm_syncobj.h>
8
9#include "v3d_drv.h"
10#include "v3d_regs.h"
11#include "v3d_trace.h"
12
13/* Takes the reservation lock on all the BOs being referenced, so that
14 * at queue submit time we can update the reservations.
15 *
16 * We don't lock the RCL the tile alloc/state BOs, or overflow memory
17 * (all of which are on exec->unref_list). They're entirely private
18 * to v3d, so we don't attach dma-buf fences to them.
19 */
20static int
21v3d_lock_bo_reservations(struct v3d_job *job,
22 struct ww_acquire_ctx *acquire_ctx)
23{
24 int i, ret;
25
26 ret = drm_gem_lock_reservations(job->bo, job->bo_count, acquire_ctx);
27 if (ret)
28 return ret;
29
30 for (i = 0; i < job->bo_count; i++) {
31 ret = dma_resv_reserve_fences(job->bo[i]->resv, 1);
32 if (ret)
33 goto fail;
34
35 ret = drm_sched_job_add_implicit_dependencies(&job->base,
36 job->bo[i], true);
37 if (ret)
38 goto fail;
39 }
40
41 return 0;
42
43fail:
44 drm_gem_unlock_reservations(job->bo, job->bo_count, acquire_ctx);
45 return ret;
46}
47
48/**
49 * v3d_lookup_bos() - Sets up job->bo[] with the GEM objects
50 * referenced by the job.
51 * @dev: DRM device
52 * @file_priv: DRM file for this fd
53 * @job: V3D job being set up
54 * @bo_handles: GEM handles
55 * @bo_count: Number of GEM handles passed in
56 *
57 * The command validator needs to reference BOs by their index within
58 * the submitted job's BO list. This does the validation of the job's
59 * BO list and reference counting for the lifetime of the job.
60 *
61 * Note that this function doesn't need to unreference the BOs on
62 * failure, because that will happen at v3d_exec_cleanup() time.
63 */
64static int
65v3d_lookup_bos(struct drm_device *dev,
66 struct drm_file *file_priv,
67 struct v3d_job *job,
68 u64 bo_handles,
69 u32 bo_count)
70{
71 job->bo_count = bo_count;
72
73 if (!job->bo_count) {
74 /* See comment on bo_index for why we have to check
75 * this.
76 */
77 DRM_DEBUG("Rendering requires BOs\n");
78 return -EINVAL;
79 }
80
81 return drm_gem_objects_lookup(file_priv,
82 (void __user *)(uintptr_t)bo_handles,
83 job->bo_count, &job->bo);
84}
85
86static void
87v3d_job_free(struct kref *ref)
88{
89 struct v3d_job *job = container_of(ref, struct v3d_job, refcount);
90 int i;
91
92 if (job->bo) {
93 for (i = 0; i < job->bo_count; i++)
94 drm_gem_object_put(job->bo[i]);
95 kvfree(job->bo);
96 }
97
98 dma_fence_put(job->irq_fence);
99 dma_fence_put(job->done_fence);
100
101 if (job->perfmon)
102 v3d_perfmon_put(job->perfmon);
103
104 kfree(job);
105}
106
107static void
108v3d_render_job_free(struct kref *ref)
109{
110 struct v3d_render_job *job = container_of(ref, struct v3d_render_job,
111 base.refcount);
112 struct v3d_bo *bo, *save;
113
114 list_for_each_entry_safe(bo, save, &job->unref_list, unref_head) {
115 drm_gem_object_put(&bo->base.base);
116 }
117
118 v3d_job_free(ref);
119}
120
121void v3d_job_cleanup(struct v3d_job *job)
122{
123 if (!job)
124 return;
125
126 drm_sched_job_cleanup(&job->base);
127 v3d_job_put(job);
128}
129
130void v3d_job_put(struct v3d_job *job)
131{
132 if (!job)
133 return;
134
135 kref_put(&job->refcount, job->free);
136}
137
138static int
139v3d_job_allocate(void **container, size_t size)
140{
141 *container = kcalloc(1, size, GFP_KERNEL);
142 if (!*container) {
143 DRM_ERROR("Cannot allocate memory for V3D job.\n");
144 return -ENOMEM;
145 }
146
147 return 0;
148}
149
150static void
151v3d_job_deallocate(void **container)
152{
153 kfree(*container);
154 *container = NULL;
155}
156
157static int
158v3d_job_init(struct v3d_dev *v3d, struct drm_file *file_priv,
159 struct v3d_job *job, void (*free)(struct kref *ref),
160 u32 in_sync, struct v3d_submit_ext *se, enum v3d_queue queue)
161{
162 struct v3d_file_priv *v3d_priv = file_priv->driver_priv;
163 bool has_multisync = se && (se->flags & DRM_V3D_EXT_ID_MULTI_SYNC);
164 int ret, i;
165
166 job->v3d = v3d;
167 job->free = free;
168 job->file = file_priv;
169
170 ret = drm_sched_job_init(&job->base, &v3d_priv->sched_entity[queue],
171 1, v3d_priv);
172 if (ret)
173 return ret;
174
175 if (has_multisync) {
176 if (se->in_sync_count && se->wait_stage == queue) {
177 struct drm_v3d_sem __user *handle = u64_to_user_ptr(se->in_syncs);
178
179 for (i = 0; i < se->in_sync_count; i++) {
180 struct drm_v3d_sem in;
181
182 if (copy_from_user(&in, handle++, sizeof(in))) {
183 ret = -EFAULT;
184 DRM_DEBUG("Failed to copy wait dep handle.\n");
185 goto fail_deps;
186 }
187 ret = drm_sched_job_add_syncobj_dependency(&job->base, file_priv, in.handle, 0);
188
189 // TODO: Investigate why this was filtered out for the IOCTL.
190 if (ret && ret != -ENOENT)
191 goto fail_deps;
192 }
193 }
194 } else {
195 ret = drm_sched_job_add_syncobj_dependency(&job->base, file_priv, in_sync, 0);
196
197 // TODO: Investigate why this was filtered out for the IOCTL.
198 if (ret && ret != -ENOENT)
199 goto fail_deps;
200 }
201
202 kref_init(&job->refcount);
203
204 return 0;
205
206fail_deps:
207 drm_sched_job_cleanup(&job->base);
208 return ret;
209}
210
211static void
212v3d_push_job(struct v3d_job *job)
213{
214 drm_sched_job_arm(&job->base);
215
216 job->done_fence = dma_fence_get(&job->base.s_fence->finished);
217
218 /* put by scheduler job completion */
219 kref_get(&job->refcount);
220
221 drm_sched_entity_push_job(&job->base);
222}
223
224static void
225v3d_attach_fences_and_unlock_reservation(struct drm_file *file_priv,
226 struct v3d_job *job,
227 struct ww_acquire_ctx *acquire_ctx,
228 u32 out_sync,
229 struct v3d_submit_ext *se,
230 struct dma_fence *done_fence)
231{
232 struct drm_syncobj *sync_out;
233 bool has_multisync = se && (se->flags & DRM_V3D_EXT_ID_MULTI_SYNC);
234 int i;
235
236 for (i = 0; i < job->bo_count; i++) {
237 /* XXX: Use shared fences for read-only objects. */
238 dma_resv_add_fence(job->bo[i]->resv, job->done_fence,
239 DMA_RESV_USAGE_WRITE);
240 }
241
242 drm_gem_unlock_reservations(job->bo, job->bo_count, acquire_ctx);
243
244 /* Update the return sync object for the job */
245 /* If it only supports a single signal semaphore*/
246 if (!has_multisync) {
247 sync_out = drm_syncobj_find(file_priv, out_sync);
248 if (sync_out) {
249 drm_syncobj_replace_fence(sync_out, done_fence);
250 drm_syncobj_put(sync_out);
251 }
252 return;
253 }
254
255 /* If multiple semaphores extension is supported */
256 if (se->out_sync_count) {
257 for (i = 0; i < se->out_sync_count; i++) {
258 drm_syncobj_replace_fence(se->out_syncs[i].syncobj,
259 done_fence);
260 drm_syncobj_put(se->out_syncs[i].syncobj);
261 }
262 kvfree(se->out_syncs);
263 }
264}
265
266static int
267v3d_setup_csd_jobs_and_bos(struct drm_file *file_priv,
268 struct v3d_dev *v3d,
269 struct drm_v3d_submit_csd *args,
270 struct v3d_csd_job **job,
271 struct v3d_job **clean_job,
272 struct v3d_submit_ext *se,
273 struct ww_acquire_ctx *acquire_ctx)
274{
275 int ret;
276
277 ret = v3d_job_allocate((void *)job, sizeof(**job));
278 if (ret)
279 return ret;
280
281 ret = v3d_job_init(v3d, file_priv, &(*job)->base,
282 v3d_job_free, args->in_sync, se, V3D_CSD);
283 if (ret) {
284 v3d_job_deallocate((void *)job);
285 return ret;
286 }
287
288 ret = v3d_job_allocate((void *)clean_job, sizeof(**clean_job));
289 if (ret)
290 return ret;
291
292 ret = v3d_job_init(v3d, file_priv, *clean_job,
293 v3d_job_free, 0, NULL, V3D_CACHE_CLEAN);
294 if (ret) {
295 v3d_job_deallocate((void *)clean_job);
296 return ret;
297 }
298
299 (*job)->args = *args;
300
301 ret = v3d_lookup_bos(&v3d->drm, file_priv, *clean_job,
302 args->bo_handles, args->bo_handle_count);
303 if (ret)
304 return ret;
305
306 return v3d_lock_bo_reservations(*clean_job, acquire_ctx);
307}
308
309static void
310v3d_put_multisync_post_deps(struct v3d_submit_ext *se)
311{
312 unsigned int i;
313
314 if (!(se && se->out_sync_count))
315 return;
316
317 for (i = 0; i < se->out_sync_count; i++)
318 drm_syncobj_put(se->out_syncs[i].syncobj);
319 kvfree(se->out_syncs);
320}
321
322static int
323v3d_get_multisync_post_deps(struct drm_file *file_priv,
324 struct v3d_submit_ext *se,
325 u32 count, u64 handles)
326{
327 struct drm_v3d_sem __user *post_deps;
328 int i, ret;
329
330 if (!count)
331 return 0;
332
333 se->out_syncs = (struct v3d_submit_outsync *)
334 kvmalloc_array(count,
335 sizeof(struct v3d_submit_outsync),
336 GFP_KERNEL);
337 if (!se->out_syncs)
338 return -ENOMEM;
339
340 post_deps = u64_to_user_ptr(handles);
341
342 for (i = 0; i < count; i++) {
343 struct drm_v3d_sem out;
344
345 if (copy_from_user(&out, post_deps++, sizeof(out))) {
346 ret = -EFAULT;
347 DRM_DEBUG("Failed to copy post dep handles\n");
348 goto fail;
349 }
350
351 se->out_syncs[i].syncobj = drm_syncobj_find(file_priv,
352 out.handle);
353 if (!se->out_syncs[i].syncobj) {
354 ret = -EINVAL;
355 goto fail;
356 }
357 }
358 se->out_sync_count = count;
359
360 return 0;
361
362fail:
363 for (i--; i >= 0; i--)
364 drm_syncobj_put(se->out_syncs[i].syncobj);
365 kvfree(se->out_syncs);
366
367 return ret;
368}
369
370/* Get data for multiple binary semaphores synchronization. Parse syncobj
371 * to be signaled when job completes (out_sync).
372 */
373static int
374v3d_get_multisync_submit_deps(struct drm_file *file_priv,
375 struct drm_v3d_extension __user *ext,
376 struct v3d_submit_ext *se)
377{
378 struct drm_v3d_multi_sync multisync;
379 int ret;
380
381 if (se->in_sync_count || se->out_sync_count) {
382 DRM_DEBUG("Two multisync extensions were added to the same job.");
383 return -EINVAL;
384 }
385
386 if (copy_from_user(&multisync, ext, sizeof(multisync)))
387 return -EFAULT;
388
389 if (multisync.pad)
390 return -EINVAL;
391
392 ret = v3d_get_multisync_post_deps(file_priv, se, multisync.out_sync_count,
393 multisync.out_syncs);
394 if (ret)
395 return ret;
396
397 se->in_sync_count = multisync.in_sync_count;
398 se->in_syncs = multisync.in_syncs;
399 se->flags |= DRM_V3D_EXT_ID_MULTI_SYNC;
400 se->wait_stage = multisync.wait_stage;
401
402 return 0;
403}
404
405/* Get data for the indirect CSD job submission. */
406static int
407v3d_get_cpu_indirect_csd_params(struct drm_file *file_priv,
408 struct drm_v3d_extension __user *ext,
409 struct v3d_cpu_job *job)
410{
411 struct v3d_file_priv *v3d_priv = file_priv->driver_priv;
412 struct v3d_dev *v3d = v3d_priv->v3d;
413 struct drm_v3d_indirect_csd indirect_csd;
414 struct v3d_indirect_csd_info *info = &job->indirect_csd;
415
416 if (!job) {
417 DRM_DEBUG("CPU job extension was attached to a GPU job.\n");
418 return -EINVAL;
419 }
420
421 if (job->job_type) {
422 DRM_DEBUG("Two CPU job extensions were added to the same CPU job.\n");
423 return -EINVAL;
424 }
425
426 if (copy_from_user(&indirect_csd, ext, sizeof(indirect_csd)))
427 return -EFAULT;
428
429 if (!v3d_has_csd(v3d)) {
430 DRM_DEBUG("Attempting CSD submit on non-CSD hardware.\n");
431 return -EINVAL;
432 }
433
434 job->job_type = V3D_CPU_JOB_TYPE_INDIRECT_CSD;
435 info->offset = indirect_csd.offset;
436 info->wg_size = indirect_csd.wg_size;
437 memcpy(&info->wg_uniform_offsets, &indirect_csd.wg_uniform_offsets,
438 sizeof(indirect_csd.wg_uniform_offsets));
439
440 info->indirect = drm_gem_object_lookup(file_priv, indirect_csd.indirect);
441
442 return v3d_setup_csd_jobs_and_bos(file_priv, v3d, &indirect_csd.submit,
443 &info->job, &info->clean_job,
444 NULL, &info->acquire_ctx);
445}
446
447/* Get data for the query timestamp job submission. */
448static int
449v3d_get_cpu_timestamp_query_params(struct drm_file *file_priv,
450 struct drm_v3d_extension __user *ext,
451 struct v3d_cpu_job *job)
452{
453 u32 __user *offsets, *syncs;
454 struct drm_v3d_timestamp_query timestamp;
455
456 if (!job) {
457 DRM_DEBUG("CPU job extension was attached to a GPU job.\n");
458 return -EINVAL;
459 }
460
461 if (job->job_type) {
462 DRM_DEBUG("Two CPU job extensions were added to the same CPU job.\n");
463 return -EINVAL;
464 }
465
466 if (copy_from_user(×tamp, ext, sizeof(timestamp)))
467 return -EFAULT;
468
469 if (timestamp.pad)
470 return -EINVAL;
471
472 job->job_type = V3D_CPU_JOB_TYPE_TIMESTAMP_QUERY;
473
474 job->timestamp_query.queries = kvmalloc_array(timestamp.count,
475 sizeof(struct v3d_timestamp_query),
476 GFP_KERNEL);
477 if (!job->timestamp_query.queries)
478 return -ENOMEM;
479
480 offsets = u64_to_user_ptr(timestamp.offsets);
481 syncs = u64_to_user_ptr(timestamp.syncs);
482
483 for (int i = 0; i < timestamp.count; i++) {
484 u32 offset, sync;
485
486 if (copy_from_user(&offset, offsets++, sizeof(offset))) {
487 kvfree(job->timestamp_query.queries);
488 return -EFAULT;
489 }
490
491 job->timestamp_query.queries[i].offset = offset;
492
493 if (copy_from_user(&sync, syncs++, sizeof(sync))) {
494 kvfree(job->timestamp_query.queries);
495 return -EFAULT;
496 }
497
498 job->timestamp_query.queries[i].syncobj = drm_syncobj_find(file_priv, sync);
499 }
500 job->timestamp_query.count = timestamp.count;
501
502 return 0;
503}
504
505static int
506v3d_get_cpu_reset_timestamp_params(struct drm_file *file_priv,
507 struct drm_v3d_extension __user *ext,
508 struct v3d_cpu_job *job)
509{
510 u32 __user *syncs;
511 struct drm_v3d_reset_timestamp_query reset;
512
513 if (!job) {
514 DRM_DEBUG("CPU job extension was attached to a GPU job.\n");
515 return -EINVAL;
516 }
517
518 if (job->job_type) {
519 DRM_DEBUG("Two CPU job extensions were added to the same CPU job.\n");
520 return -EINVAL;
521 }
522
523 if (copy_from_user(&reset, ext, sizeof(reset)))
524 return -EFAULT;
525
526 job->job_type = V3D_CPU_JOB_TYPE_RESET_TIMESTAMP_QUERY;
527
528 job->timestamp_query.queries = kvmalloc_array(reset.count,
529 sizeof(struct v3d_timestamp_query),
530 GFP_KERNEL);
531 if (!job->timestamp_query.queries)
532 return -ENOMEM;
533
534 syncs = u64_to_user_ptr(reset.syncs);
535
536 for (int i = 0; i < reset.count; i++) {
537 u32 sync;
538
539 job->timestamp_query.queries[i].offset = reset.offset + 8 * i;
540
541 if (copy_from_user(&sync, syncs++, sizeof(sync))) {
542 kvfree(job->timestamp_query.queries);
543 return -EFAULT;
544 }
545
546 job->timestamp_query.queries[i].syncobj = drm_syncobj_find(file_priv, sync);
547 }
548 job->timestamp_query.count = reset.count;
549
550 return 0;
551}
552
553/* Get data for the copy timestamp query results job submission. */
554static int
555v3d_get_cpu_copy_query_results_params(struct drm_file *file_priv,
556 struct drm_v3d_extension __user *ext,
557 struct v3d_cpu_job *job)
558{
559 u32 __user *offsets, *syncs;
560 struct drm_v3d_copy_timestamp_query copy;
561 int i;
562
563 if (!job) {
564 DRM_DEBUG("CPU job extension was attached to a GPU job.\n");
565 return -EINVAL;
566 }
567
568 if (job->job_type) {
569 DRM_DEBUG("Two CPU job extensions were added to the same CPU job.\n");
570 return -EINVAL;
571 }
572
573 if (copy_from_user(©, ext, sizeof(copy)))
574 return -EFAULT;
575
576 if (copy.pad)
577 return -EINVAL;
578
579 job->job_type = V3D_CPU_JOB_TYPE_COPY_TIMESTAMP_QUERY;
580
581 job->timestamp_query.queries = kvmalloc_array(copy.count,
582 sizeof(struct v3d_timestamp_query),
583 GFP_KERNEL);
584 if (!job->timestamp_query.queries)
585 return -ENOMEM;
586
587 offsets = u64_to_user_ptr(copy.offsets);
588 syncs = u64_to_user_ptr(copy.syncs);
589
590 for (i = 0; i < copy.count; i++) {
591 u32 offset, sync;
592
593 if (copy_from_user(&offset, offsets++, sizeof(offset))) {
594 kvfree(job->timestamp_query.queries);
595 return -EFAULT;
596 }
597
598 job->timestamp_query.queries[i].offset = offset;
599
600 if (copy_from_user(&sync, syncs++, sizeof(sync))) {
601 kvfree(job->timestamp_query.queries);
602 return -EFAULT;
603 }
604
605 job->timestamp_query.queries[i].syncobj = drm_syncobj_find(file_priv, sync);
606 }
607 job->timestamp_query.count = copy.count;
608
609 job->copy.do_64bit = copy.do_64bit;
610 job->copy.do_partial = copy.do_partial;
611 job->copy.availability_bit = copy.availability_bit;
612 job->copy.offset = copy.offset;
613 job->copy.stride = copy.stride;
614
615 return 0;
616}
617
618static int
619v3d_get_cpu_reset_performance_params(struct drm_file *file_priv,
620 struct drm_v3d_extension __user *ext,
621 struct v3d_cpu_job *job)
622{
623 u32 __user *syncs;
624 u64 __user *kperfmon_ids;
625 struct drm_v3d_reset_performance_query reset;
626
627 if (!job) {
628 DRM_DEBUG("CPU job extension was attached to a GPU job.\n");
629 return -EINVAL;
630 }
631
632 if (job->job_type) {
633 DRM_DEBUG("Two CPU job extensions were added to the same CPU job.\n");
634 return -EINVAL;
635 }
636
637 if (copy_from_user(&reset, ext, sizeof(reset)))
638 return -EFAULT;
639
640 job->job_type = V3D_CPU_JOB_TYPE_RESET_PERFORMANCE_QUERY;
641
642 job->performance_query.queries = kvmalloc_array(reset.count,
643 sizeof(struct v3d_performance_query),
644 GFP_KERNEL);
645 if (!job->performance_query.queries)
646 return -ENOMEM;
647
648 syncs = u64_to_user_ptr(reset.syncs);
649 kperfmon_ids = u64_to_user_ptr(reset.kperfmon_ids);
650
651 for (int i = 0; i < reset.count; i++) {
652 u32 sync;
653 u64 ids;
654 u32 __user *ids_pointer;
655 u32 id;
656
657 if (copy_from_user(&sync, syncs++, sizeof(sync))) {
658 kvfree(job->performance_query.queries);
659 return -EFAULT;
660 }
661
662 job->performance_query.queries[i].syncobj = drm_syncobj_find(file_priv, sync);
663
664 if (copy_from_user(&ids, kperfmon_ids++, sizeof(ids))) {
665 kvfree(job->performance_query.queries);
666 return -EFAULT;
667 }
668
669 ids_pointer = u64_to_user_ptr(ids);
670
671 for (int j = 0; j < reset.nperfmons; j++) {
672 if (copy_from_user(&id, ids_pointer++, sizeof(id))) {
673 kvfree(job->performance_query.queries);
674 return -EFAULT;
675 }
676
677 job->performance_query.queries[i].kperfmon_ids[j] = id;
678 }
679 }
680 job->performance_query.count = reset.count;
681 job->performance_query.nperfmons = reset.nperfmons;
682
683 return 0;
684}
685
686static int
687v3d_get_cpu_copy_performance_query_params(struct drm_file *file_priv,
688 struct drm_v3d_extension __user *ext,
689 struct v3d_cpu_job *job)
690{
691 u32 __user *syncs;
692 u64 __user *kperfmon_ids;
693 struct drm_v3d_copy_performance_query copy;
694
695 if (!job) {
696 DRM_DEBUG("CPU job extension was attached to a GPU job.\n");
697 return -EINVAL;
698 }
699
700 if (job->job_type) {
701 DRM_DEBUG("Two CPU job extensions were added to the same CPU job.\n");
702 return -EINVAL;
703 }
704
705 if (copy_from_user(©, ext, sizeof(copy)))
706 return -EFAULT;
707
708 if (copy.pad)
709 return -EINVAL;
710
711 job->job_type = V3D_CPU_JOB_TYPE_COPY_PERFORMANCE_QUERY;
712
713 job->performance_query.queries = kvmalloc_array(copy.count,
714 sizeof(struct v3d_performance_query),
715 GFP_KERNEL);
716 if (!job->performance_query.queries)
717 return -ENOMEM;
718
719 syncs = u64_to_user_ptr(copy.syncs);
720 kperfmon_ids = u64_to_user_ptr(copy.kperfmon_ids);
721
722 for (int i = 0; i < copy.count; i++) {
723 u32 sync;
724 u64 ids;
725 u32 __user *ids_pointer;
726 u32 id;
727
728 if (copy_from_user(&sync, syncs++, sizeof(sync))) {
729 kvfree(job->performance_query.queries);
730 return -EFAULT;
731 }
732
733 job->performance_query.queries[i].syncobj = drm_syncobj_find(file_priv, sync);
734
735 if (copy_from_user(&ids, kperfmon_ids++, sizeof(ids))) {
736 kvfree(job->performance_query.queries);
737 return -EFAULT;
738 }
739
740 ids_pointer = u64_to_user_ptr(ids);
741
742 for (int j = 0; j < copy.nperfmons; j++) {
743 if (copy_from_user(&id, ids_pointer++, sizeof(id))) {
744 kvfree(job->performance_query.queries);
745 return -EFAULT;
746 }
747
748 job->performance_query.queries[i].kperfmon_ids[j] = id;
749 }
750 }
751 job->performance_query.count = copy.count;
752 job->performance_query.nperfmons = copy.nperfmons;
753 job->performance_query.ncounters = copy.ncounters;
754
755 job->copy.do_64bit = copy.do_64bit;
756 job->copy.do_partial = copy.do_partial;
757 job->copy.availability_bit = copy.availability_bit;
758 job->copy.offset = copy.offset;
759 job->copy.stride = copy.stride;
760
761 return 0;
762}
763
764/* Whenever userspace sets ioctl extensions, v3d_get_extensions parses data
765 * according to the extension id (name).
766 */
767static int
768v3d_get_extensions(struct drm_file *file_priv,
769 u64 ext_handles,
770 struct v3d_submit_ext *se,
771 struct v3d_cpu_job *job)
772{
773 struct drm_v3d_extension __user *user_ext;
774 int ret;
775
776 user_ext = u64_to_user_ptr(ext_handles);
777 while (user_ext) {
778 struct drm_v3d_extension ext;
779
780 if (copy_from_user(&ext, user_ext, sizeof(ext))) {
781 DRM_DEBUG("Failed to copy submit extension\n");
782 return -EFAULT;
783 }
784
785 switch (ext.id) {
786 case DRM_V3D_EXT_ID_MULTI_SYNC:
787 ret = v3d_get_multisync_submit_deps(file_priv, user_ext, se);
788 break;
789 case DRM_V3D_EXT_ID_CPU_INDIRECT_CSD:
790 ret = v3d_get_cpu_indirect_csd_params(file_priv, user_ext, job);
791 break;
792 case DRM_V3D_EXT_ID_CPU_TIMESTAMP_QUERY:
793 ret = v3d_get_cpu_timestamp_query_params(file_priv, user_ext, job);
794 break;
795 case DRM_V3D_EXT_ID_CPU_RESET_TIMESTAMP_QUERY:
796 ret = v3d_get_cpu_reset_timestamp_params(file_priv, user_ext, job);
797 break;
798 case DRM_V3D_EXT_ID_CPU_COPY_TIMESTAMP_QUERY:
799 ret = v3d_get_cpu_copy_query_results_params(file_priv, user_ext, job);
800 break;
801 case DRM_V3D_EXT_ID_CPU_RESET_PERFORMANCE_QUERY:
802 ret = v3d_get_cpu_reset_performance_params(file_priv, user_ext, job);
803 break;
804 case DRM_V3D_EXT_ID_CPU_COPY_PERFORMANCE_QUERY:
805 ret = v3d_get_cpu_copy_performance_query_params(file_priv, user_ext, job);
806 break;
807 default:
808 DRM_DEBUG_DRIVER("Unknown extension id: %d\n", ext.id);
809 return -EINVAL;
810 }
811
812 if (ret)
813 return ret;
814
815 user_ext = u64_to_user_ptr(ext.next);
816 }
817
818 return 0;
819}
820
821/**
822 * v3d_submit_cl_ioctl() - Submits a job (frame) to the V3D.
823 * @dev: DRM device
824 * @data: ioctl argument
825 * @file_priv: DRM file for this fd
826 *
827 * This is the main entrypoint for userspace to submit a 3D frame to
828 * the GPU. Userspace provides the binner command list (if
829 * applicable), and the kernel sets up the render command list to draw
830 * to the framebuffer described in the ioctl, using the command lists
831 * that the 3D engine's binner will produce.
832 */
833int
834v3d_submit_cl_ioctl(struct drm_device *dev, void *data,
835 struct drm_file *file_priv)
836{
837 struct v3d_dev *v3d = to_v3d_dev(dev);
838 struct v3d_file_priv *v3d_priv = file_priv->driver_priv;
839 struct drm_v3d_submit_cl *args = data;
840 struct v3d_submit_ext se = {0};
841 struct v3d_bin_job *bin = NULL;
842 struct v3d_render_job *render = NULL;
843 struct v3d_job *clean_job = NULL;
844 struct v3d_job *last_job;
845 struct ww_acquire_ctx acquire_ctx;
846 int ret = 0;
847
848 trace_v3d_submit_cl_ioctl(&v3d->drm, args->rcl_start, args->rcl_end);
849
850 if (args->pad)
851 return -EINVAL;
852
853 if (args->flags &&
854 args->flags & ~(DRM_V3D_SUBMIT_CL_FLUSH_CACHE |
855 DRM_V3D_SUBMIT_EXTENSION)) {
856 DRM_INFO("invalid flags: %d\n", args->flags);
857 return -EINVAL;
858 }
859
860 if (args->flags & DRM_V3D_SUBMIT_EXTENSION) {
861 ret = v3d_get_extensions(file_priv, args->extensions, &se, NULL);
862 if (ret) {
863 DRM_DEBUG("Failed to get extensions.\n");
864 return ret;
865 }
866 }
867
868 ret = v3d_job_allocate((void *)&render, sizeof(*render));
869 if (ret)
870 return ret;
871
872 ret = v3d_job_init(v3d, file_priv, &render->base,
873 v3d_render_job_free, args->in_sync_rcl, &se, V3D_RENDER);
874 if (ret) {
875 v3d_job_deallocate((void *)&render);
876 goto fail;
877 }
878
879 render->start = args->rcl_start;
880 render->end = args->rcl_end;
881 INIT_LIST_HEAD(&render->unref_list);
882
883 if (args->bcl_start != args->bcl_end) {
884 ret = v3d_job_allocate((void *)&bin, sizeof(*bin));
885 if (ret)
886 goto fail;
887
888 ret = v3d_job_init(v3d, file_priv, &bin->base,
889 v3d_job_free, args->in_sync_bcl, &se, V3D_BIN);
890 if (ret) {
891 v3d_job_deallocate((void *)&bin);
892 goto fail;
893 }
894
895 bin->start = args->bcl_start;
896 bin->end = args->bcl_end;
897 bin->qma = args->qma;
898 bin->qms = args->qms;
899 bin->qts = args->qts;
900 bin->render = render;
901 }
902
903 if (args->flags & DRM_V3D_SUBMIT_CL_FLUSH_CACHE) {
904 ret = v3d_job_allocate((void *)&clean_job, sizeof(*clean_job));
905 if (ret)
906 goto fail;
907
908 ret = v3d_job_init(v3d, file_priv, clean_job,
909 v3d_job_free, 0, NULL, V3D_CACHE_CLEAN);
910 if (ret) {
911 v3d_job_deallocate((void *)&clean_job);
912 goto fail;
913 }
914
915 last_job = clean_job;
916 } else {
917 last_job = &render->base;
918 }
919
920 ret = v3d_lookup_bos(dev, file_priv, last_job,
921 args->bo_handles, args->bo_handle_count);
922 if (ret)
923 goto fail;
924
925 ret = v3d_lock_bo_reservations(last_job, &acquire_ctx);
926 if (ret)
927 goto fail;
928
929 if (args->perfmon_id) {
930 render->base.perfmon = v3d_perfmon_find(v3d_priv,
931 args->perfmon_id);
932
933 if (!render->base.perfmon) {
934 ret = -ENOENT;
935 goto fail_perfmon;
936 }
937 }
938
939 mutex_lock(&v3d->sched_lock);
940 if (bin) {
941 bin->base.perfmon = render->base.perfmon;
942 v3d_perfmon_get(bin->base.perfmon);
943 v3d_push_job(&bin->base);
944
945 ret = drm_sched_job_add_dependency(&render->base.base,
946 dma_fence_get(bin->base.done_fence));
947 if (ret)
948 goto fail_unreserve;
949 }
950
951 v3d_push_job(&render->base);
952
953 if (clean_job) {
954 struct dma_fence *render_fence =
955 dma_fence_get(render->base.done_fence);
956 ret = drm_sched_job_add_dependency(&clean_job->base,
957 render_fence);
958 if (ret)
959 goto fail_unreserve;
960 clean_job->perfmon = render->base.perfmon;
961 v3d_perfmon_get(clean_job->perfmon);
962 v3d_push_job(clean_job);
963 }
964
965 mutex_unlock(&v3d->sched_lock);
966
967 v3d_attach_fences_and_unlock_reservation(file_priv,
968 last_job,
969 &acquire_ctx,
970 args->out_sync,
971 &se,
972 last_job->done_fence);
973
974 v3d_job_put(&bin->base);
975 v3d_job_put(&render->base);
976 v3d_job_put(clean_job);
977
978 return 0;
979
980fail_unreserve:
981 mutex_unlock(&v3d->sched_lock);
982fail_perfmon:
983 drm_gem_unlock_reservations(last_job->bo,
984 last_job->bo_count, &acquire_ctx);
985fail:
986 v3d_job_cleanup((void *)bin);
987 v3d_job_cleanup((void *)render);
988 v3d_job_cleanup(clean_job);
989 v3d_put_multisync_post_deps(&se);
990
991 return ret;
992}
993
994/**
995 * v3d_submit_tfu_ioctl() - Submits a TFU (texture formatting) job to the V3D.
996 * @dev: DRM device
997 * @data: ioctl argument
998 * @file_priv: DRM file for this fd
999 *
1000 * Userspace provides the register setup for the TFU, which we don't
1001 * need to validate since the TFU is behind the MMU.
1002 */
1003int
1004v3d_submit_tfu_ioctl(struct drm_device *dev, void *data,
1005 struct drm_file *file_priv)
1006{
1007 struct v3d_dev *v3d = to_v3d_dev(dev);
1008 struct drm_v3d_submit_tfu *args = data;
1009 struct v3d_submit_ext se = {0};
1010 struct v3d_tfu_job *job = NULL;
1011 struct ww_acquire_ctx acquire_ctx;
1012 int ret = 0;
1013
1014 trace_v3d_submit_tfu_ioctl(&v3d->drm, args->iia);
1015
1016 if (args->flags && !(args->flags & DRM_V3D_SUBMIT_EXTENSION)) {
1017 DRM_DEBUG("invalid flags: %d\n", args->flags);
1018 return -EINVAL;
1019 }
1020
1021 if (args->flags & DRM_V3D_SUBMIT_EXTENSION) {
1022 ret = v3d_get_extensions(file_priv, args->extensions, &se, NULL);
1023 if (ret) {
1024 DRM_DEBUG("Failed to get extensions.\n");
1025 return ret;
1026 }
1027 }
1028
1029 ret = v3d_job_allocate((void *)&job, sizeof(*job));
1030 if (ret)
1031 return ret;
1032
1033 ret = v3d_job_init(v3d, file_priv, &job->base,
1034 v3d_job_free, args->in_sync, &se, V3D_TFU);
1035 if (ret) {
1036 v3d_job_deallocate((void *)&job);
1037 goto fail;
1038 }
1039
1040 job->base.bo = kcalloc(ARRAY_SIZE(args->bo_handles),
1041 sizeof(*job->base.bo), GFP_KERNEL);
1042 if (!job->base.bo) {
1043 ret = -ENOMEM;
1044 goto fail;
1045 }
1046
1047 job->args = *args;
1048
1049 for (job->base.bo_count = 0;
1050 job->base.bo_count < ARRAY_SIZE(args->bo_handles);
1051 job->base.bo_count++) {
1052 struct drm_gem_object *bo;
1053
1054 if (!args->bo_handles[job->base.bo_count])
1055 break;
1056
1057 bo = drm_gem_object_lookup(file_priv, args->bo_handles[job->base.bo_count]);
1058 if (!bo) {
1059 DRM_DEBUG("Failed to look up GEM BO %d: %d\n",
1060 job->base.bo_count,
1061 args->bo_handles[job->base.bo_count]);
1062 ret = -ENOENT;
1063 goto fail;
1064 }
1065 job->base.bo[job->base.bo_count] = bo;
1066 }
1067
1068 ret = v3d_lock_bo_reservations(&job->base, &acquire_ctx);
1069 if (ret)
1070 goto fail;
1071
1072 mutex_lock(&v3d->sched_lock);
1073 v3d_push_job(&job->base);
1074 mutex_unlock(&v3d->sched_lock);
1075
1076 v3d_attach_fences_and_unlock_reservation(file_priv,
1077 &job->base, &acquire_ctx,
1078 args->out_sync,
1079 &se,
1080 job->base.done_fence);
1081
1082 v3d_job_put(&job->base);
1083
1084 return 0;
1085
1086fail:
1087 v3d_job_cleanup((void *)job);
1088 v3d_put_multisync_post_deps(&se);
1089
1090 return ret;
1091}
1092
1093/**
1094 * v3d_submit_csd_ioctl() - Submits a CSD (compute shader) job to the V3D.
1095 * @dev: DRM device
1096 * @data: ioctl argument
1097 * @file_priv: DRM file for this fd
1098 *
1099 * Userspace provides the register setup for the CSD, which we don't
1100 * need to validate since the CSD is behind the MMU.
1101 */
1102int
1103v3d_submit_csd_ioctl(struct drm_device *dev, void *data,
1104 struct drm_file *file_priv)
1105{
1106 struct v3d_dev *v3d = to_v3d_dev(dev);
1107 struct v3d_file_priv *v3d_priv = file_priv->driver_priv;
1108 struct drm_v3d_submit_csd *args = data;
1109 struct v3d_submit_ext se = {0};
1110 struct v3d_csd_job *job = NULL;
1111 struct v3d_job *clean_job = NULL;
1112 struct ww_acquire_ctx acquire_ctx;
1113 int ret;
1114
1115 trace_v3d_submit_csd_ioctl(&v3d->drm, args->cfg[5], args->cfg[6]);
1116
1117 if (args->pad)
1118 return -EINVAL;
1119
1120 if (!v3d_has_csd(v3d)) {
1121 DRM_DEBUG("Attempting CSD submit on non-CSD hardware\n");
1122 return -EINVAL;
1123 }
1124
1125 if (args->flags && !(args->flags & DRM_V3D_SUBMIT_EXTENSION)) {
1126 DRM_INFO("invalid flags: %d\n", args->flags);
1127 return -EINVAL;
1128 }
1129
1130 if (args->flags & DRM_V3D_SUBMIT_EXTENSION) {
1131 ret = v3d_get_extensions(file_priv, args->extensions, &se, NULL);
1132 if (ret) {
1133 DRM_DEBUG("Failed to get extensions.\n");
1134 return ret;
1135 }
1136 }
1137
1138 ret = v3d_setup_csd_jobs_and_bos(file_priv, v3d, args,
1139 &job, &clean_job, &se,
1140 &acquire_ctx);
1141 if (ret)
1142 goto fail;
1143
1144 if (args->perfmon_id) {
1145 job->base.perfmon = v3d_perfmon_find(v3d_priv,
1146 args->perfmon_id);
1147 if (!job->base.perfmon) {
1148 ret = -ENOENT;
1149 goto fail_perfmon;
1150 }
1151 }
1152
1153 mutex_lock(&v3d->sched_lock);
1154 v3d_push_job(&job->base);
1155
1156 ret = drm_sched_job_add_dependency(&clean_job->base,
1157 dma_fence_get(job->base.done_fence));
1158 if (ret)
1159 goto fail_unreserve;
1160
1161 v3d_push_job(clean_job);
1162 mutex_unlock(&v3d->sched_lock);
1163
1164 v3d_attach_fences_and_unlock_reservation(file_priv,
1165 clean_job,
1166 &acquire_ctx,
1167 args->out_sync,
1168 &se,
1169 clean_job->done_fence);
1170
1171 v3d_job_put(&job->base);
1172 v3d_job_put(clean_job);
1173
1174 return 0;
1175
1176fail_unreserve:
1177 mutex_unlock(&v3d->sched_lock);
1178fail_perfmon:
1179 drm_gem_unlock_reservations(clean_job->bo, clean_job->bo_count,
1180 &acquire_ctx);
1181fail:
1182 v3d_job_cleanup((void *)job);
1183 v3d_job_cleanup(clean_job);
1184 v3d_put_multisync_post_deps(&se);
1185
1186 return ret;
1187}
1188
1189static const unsigned int cpu_job_bo_handle_count[] = {
1190 [V3D_CPU_JOB_TYPE_INDIRECT_CSD] = 1,
1191 [V3D_CPU_JOB_TYPE_TIMESTAMP_QUERY] = 1,
1192 [V3D_CPU_JOB_TYPE_RESET_TIMESTAMP_QUERY] = 1,
1193 [V3D_CPU_JOB_TYPE_COPY_TIMESTAMP_QUERY] = 2,
1194 [V3D_CPU_JOB_TYPE_RESET_PERFORMANCE_QUERY] = 0,
1195 [V3D_CPU_JOB_TYPE_COPY_PERFORMANCE_QUERY] = 1,
1196};
1197
1198/**
1199 * v3d_submit_cpu_ioctl() - Submits a CPU job to the V3D.
1200 * @dev: DRM device
1201 * @data: ioctl argument
1202 * @file_priv: DRM file for this fd
1203 *
1204 * Userspace specifies the CPU job type and data required to perform its
1205 * operations through the drm_v3d_extension struct.
1206 */
1207int
1208v3d_submit_cpu_ioctl(struct drm_device *dev, void *data,
1209 struct drm_file *file_priv)
1210{
1211 struct v3d_dev *v3d = to_v3d_dev(dev);
1212 struct drm_v3d_submit_cpu *args = data;
1213 struct v3d_submit_ext se = {0};
1214 struct v3d_submit_ext *out_se = NULL;
1215 struct v3d_cpu_job *cpu_job = NULL;
1216 struct v3d_csd_job *csd_job = NULL;
1217 struct v3d_job *clean_job = NULL;
1218 struct ww_acquire_ctx acquire_ctx;
1219 int ret;
1220
1221 if (args->flags && !(args->flags & DRM_V3D_SUBMIT_EXTENSION)) {
1222 DRM_INFO("Invalid flags: %d\n", args->flags);
1223 return -EINVAL;
1224 }
1225
1226 ret = v3d_job_allocate((void *)&cpu_job, sizeof(*cpu_job));
1227 if (ret)
1228 return ret;
1229
1230 if (args->flags & DRM_V3D_SUBMIT_EXTENSION) {
1231 ret = v3d_get_extensions(file_priv, args->extensions, &se, cpu_job);
1232 if (ret) {
1233 DRM_DEBUG("Failed to get extensions.\n");
1234 goto fail;
1235 }
1236 }
1237
1238 /* Every CPU job must have a CPU job user extension */
1239 if (!cpu_job->job_type) {
1240 DRM_DEBUG("CPU job must have a CPU job user extension.\n");
1241 ret = -EINVAL;
1242 goto fail;
1243 }
1244
1245 if (args->bo_handle_count != cpu_job_bo_handle_count[cpu_job->job_type]) {
1246 DRM_DEBUG("This CPU job was not submitted with the proper number of BOs.\n");
1247 ret = -EINVAL;
1248 goto fail;
1249 }
1250
1251 trace_v3d_submit_cpu_ioctl(&v3d->drm, cpu_job->job_type);
1252
1253 ret = v3d_job_init(v3d, file_priv, &cpu_job->base,
1254 v3d_job_free, 0, &se, V3D_CPU);
1255 if (ret) {
1256 v3d_job_deallocate((void *)&cpu_job);
1257 goto fail;
1258 }
1259
1260 clean_job = cpu_job->indirect_csd.clean_job;
1261 csd_job = cpu_job->indirect_csd.job;
1262
1263 if (args->bo_handle_count) {
1264 ret = v3d_lookup_bos(dev, file_priv, &cpu_job->base,
1265 args->bo_handles, args->bo_handle_count);
1266 if (ret)
1267 goto fail;
1268
1269 ret = v3d_lock_bo_reservations(&cpu_job->base, &acquire_ctx);
1270 if (ret)
1271 goto fail;
1272 }
1273
1274 mutex_lock(&v3d->sched_lock);
1275 v3d_push_job(&cpu_job->base);
1276
1277 switch (cpu_job->job_type) {
1278 case V3D_CPU_JOB_TYPE_INDIRECT_CSD:
1279 ret = drm_sched_job_add_dependency(&csd_job->base.base,
1280 dma_fence_get(cpu_job->base.done_fence));
1281 if (ret)
1282 goto fail_unreserve;
1283
1284 v3d_push_job(&csd_job->base);
1285
1286 ret = drm_sched_job_add_dependency(&clean_job->base,
1287 dma_fence_get(csd_job->base.done_fence));
1288 if (ret)
1289 goto fail_unreserve;
1290
1291 v3d_push_job(clean_job);
1292
1293 break;
1294 default:
1295 break;
1296 }
1297 mutex_unlock(&v3d->sched_lock);
1298
1299 out_se = (cpu_job->job_type == V3D_CPU_JOB_TYPE_INDIRECT_CSD) ? NULL : &se;
1300
1301 v3d_attach_fences_and_unlock_reservation(file_priv,
1302 &cpu_job->base,
1303 &acquire_ctx, 0,
1304 out_se, cpu_job->base.done_fence);
1305
1306 switch (cpu_job->job_type) {
1307 case V3D_CPU_JOB_TYPE_INDIRECT_CSD:
1308 v3d_attach_fences_and_unlock_reservation(file_priv,
1309 clean_job,
1310 &cpu_job->indirect_csd.acquire_ctx,
1311 0, &se, clean_job->done_fence);
1312 break;
1313 default:
1314 break;
1315 }
1316
1317 v3d_job_put(&cpu_job->base);
1318 v3d_job_put(&csd_job->base);
1319 v3d_job_put(clean_job);
1320
1321 return 0;
1322
1323fail_unreserve:
1324 mutex_unlock(&v3d->sched_lock);
1325
1326 drm_gem_unlock_reservations(cpu_job->base.bo, cpu_job->base.bo_count,
1327 &acquire_ctx);
1328
1329 drm_gem_unlock_reservations(clean_job->bo, clean_job->bo_count,
1330 &cpu_job->indirect_csd.acquire_ctx);
1331
1332fail:
1333 v3d_job_cleanup((void *)cpu_job);
1334 v3d_job_cleanup((void *)csd_job);
1335 v3d_job_cleanup(clean_job);
1336 v3d_put_multisync_post_deps(&se);
1337 kvfree(cpu_job->timestamp_query.queries);
1338 kvfree(cpu_job->performance_query.queries);
1339
1340 return ret;
1341}