This source file includes following definitions.
- vc4_queue_hangcheck
- vc4_free_hang_state
- vc4_get_hang_state_ioctl
- vc4_save_hang_state
- vc4_reset
- vc4_reset_work
- vc4_hangcheck_elapsed
- submit_cl
- vc4_wait_for_seqno
- vc4_flush_caches
- vc4_flush_texture_caches
- vc4_submit_next_bin_job
- vc4_submit_next_render_job
- vc4_move_job_to_render
- vc4_update_bo_seqnos
- vc4_unlock_bo_reservations
- vc4_lock_bo_reservations
- vc4_queue_submit
- vc4_cl_lookup_bos
- vc4_get_bcl
- vc4_complete_exec
- vc4_job_handle_completed
- vc4_seqno_cb_work
- vc4_queue_seqno_cb
- vc4_job_done_work
- vc4_wait_for_seqno_ioctl_helper
- vc4_wait_seqno_ioctl
- vc4_wait_bo_ioctl
- vc4_submit_cl_ioctl
- vc4_gem_init
- vc4_gem_destroy
- vc4_gem_madvise_ioctl
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24 #include <linux/module.h>
25 #include <linux/platform_device.h>
26 #include <linux/pm_runtime.h>
27 #include <linux/device.h>
28 #include <linux/io.h>
29 #include <linux/sched/signal.h>
30 #include <linux/dma-fence-array.h>
31
32 #include <drm/drm_syncobj.h>
33
34 #include "uapi/drm/vc4_drm.h"
35 #include "vc4_drv.h"
36 #include "vc4_regs.h"
37 #include "vc4_trace.h"
38
39 static void
40 vc4_queue_hangcheck(struct drm_device *dev)
41 {
42 struct vc4_dev *vc4 = to_vc4_dev(dev);
43
44 mod_timer(&vc4->hangcheck.timer,
45 round_jiffies_up(jiffies + msecs_to_jiffies(100)));
46 }
47
48 struct vc4_hang_state {
49 struct drm_vc4_get_hang_state user_state;
50
51 u32 bo_count;
52 struct drm_gem_object **bo;
53 };
54
55 static void
56 vc4_free_hang_state(struct drm_device *dev, struct vc4_hang_state *state)
57 {
58 unsigned int i;
59
60 for (i = 0; i < state->user_state.bo_count; i++)
61 drm_gem_object_put_unlocked(state->bo[i]);
62
63 kfree(state);
64 }
65
66 int
67 vc4_get_hang_state_ioctl(struct drm_device *dev, void *data,
68 struct drm_file *file_priv)
69 {
70 struct drm_vc4_get_hang_state *get_state = data;
71 struct drm_vc4_get_hang_state_bo *bo_state;
72 struct vc4_hang_state *kernel_state;
73 struct drm_vc4_get_hang_state *state;
74 struct vc4_dev *vc4 = to_vc4_dev(dev);
75 unsigned long irqflags;
76 u32 i;
77 int ret = 0;
78
79 if (!vc4->v3d) {
80 DRM_DEBUG("VC4_GET_HANG_STATE with no VC4 V3D probed\n");
81 return -ENODEV;
82 }
83
84 spin_lock_irqsave(&vc4->job_lock, irqflags);
85 kernel_state = vc4->hang_state;
86 if (!kernel_state) {
87 spin_unlock_irqrestore(&vc4->job_lock, irqflags);
88 return -ENOENT;
89 }
90 state = &kernel_state->user_state;
91
92
93
94
95 if (get_state->bo_count < state->bo_count) {
96 get_state->bo_count = state->bo_count;
97 spin_unlock_irqrestore(&vc4->job_lock, irqflags);
98 return 0;
99 }
100
101 vc4->hang_state = NULL;
102 spin_unlock_irqrestore(&vc4->job_lock, irqflags);
103
104
105 state->bo = get_state->bo;
106 memcpy(get_state, state, sizeof(*state));
107
108 bo_state = kcalloc(state->bo_count, sizeof(*bo_state), GFP_KERNEL);
109 if (!bo_state) {
110 ret = -ENOMEM;
111 goto err_free;
112 }
113
114 for (i = 0; i < state->bo_count; i++) {
115 struct vc4_bo *vc4_bo = to_vc4_bo(kernel_state->bo[i]);
116 u32 handle;
117
118 ret = drm_gem_handle_create(file_priv, kernel_state->bo[i],
119 &handle);
120
121 if (ret) {
122 state->bo_count = i;
123 goto err_delete_handle;
124 }
125 bo_state[i].handle = handle;
126 bo_state[i].paddr = vc4_bo->base.paddr;
127 bo_state[i].size = vc4_bo->base.base.size;
128 }
129
130 if (copy_to_user(u64_to_user_ptr(get_state->bo),
131 bo_state,
132 state->bo_count * sizeof(*bo_state)))
133 ret = -EFAULT;
134
135 err_delete_handle:
136 if (ret) {
137 for (i = 0; i < state->bo_count; i++)
138 drm_gem_handle_delete(file_priv, bo_state[i].handle);
139 }
140
141 err_free:
142 vc4_free_hang_state(dev, kernel_state);
143 kfree(bo_state);
144
145 return ret;
146 }
147
148 static void
149 vc4_save_hang_state(struct drm_device *dev)
150 {
151 struct vc4_dev *vc4 = to_vc4_dev(dev);
152 struct drm_vc4_get_hang_state *state;
153 struct vc4_hang_state *kernel_state;
154 struct vc4_exec_info *exec[2];
155 struct vc4_bo *bo;
156 unsigned long irqflags;
157 unsigned int i, j, k, unref_list_count;
158
159 kernel_state = kcalloc(1, sizeof(*kernel_state), GFP_KERNEL);
160 if (!kernel_state)
161 return;
162
163 state = &kernel_state->user_state;
164
165 spin_lock_irqsave(&vc4->job_lock, irqflags);
166 exec[0] = vc4_first_bin_job(vc4);
167 exec[1] = vc4_first_render_job(vc4);
168 if (!exec[0] && !exec[1]) {
169 spin_unlock_irqrestore(&vc4->job_lock, irqflags);
170 return;
171 }
172
173
174 state->bo_count = 0;
175 for (i = 0; i < 2; i++) {
176 if (!exec[i])
177 continue;
178
179 unref_list_count = 0;
180 list_for_each_entry(bo, &exec[i]->unref_list, unref_head)
181 unref_list_count++;
182 state->bo_count += exec[i]->bo_count + unref_list_count;
183 }
184
185 kernel_state->bo = kcalloc(state->bo_count,
186 sizeof(*kernel_state->bo), GFP_ATOMIC);
187
188 if (!kernel_state->bo) {
189 spin_unlock_irqrestore(&vc4->job_lock, irqflags);
190 return;
191 }
192
193 k = 0;
194 for (i = 0; i < 2; i++) {
195 if (!exec[i])
196 continue;
197
198 for (j = 0; j < exec[i]->bo_count; j++) {
199 bo = to_vc4_bo(&exec[i]->bo[j]->base);
200
201
202
203
204
205 WARN_ON(!refcount_read(&bo->usecnt));
206 refcount_inc(&bo->usecnt);
207 drm_gem_object_get(&exec[i]->bo[j]->base);
208 kernel_state->bo[k++] = &exec[i]->bo[j]->base;
209 }
210
211 list_for_each_entry(bo, &exec[i]->unref_list, unref_head) {
212
213
214
215 drm_gem_object_get(&bo->base.base);
216 kernel_state->bo[k++] = &bo->base.base;
217 }
218 }
219
220 WARN_ON_ONCE(k != state->bo_count);
221
222 if (exec[0])
223 state->start_bin = exec[0]->ct0ca;
224 if (exec[1])
225 state->start_render = exec[1]->ct1ca;
226
227 spin_unlock_irqrestore(&vc4->job_lock, irqflags);
228
229 state->ct0ca = V3D_READ(V3D_CTNCA(0));
230 state->ct0ea = V3D_READ(V3D_CTNEA(0));
231
232 state->ct1ca = V3D_READ(V3D_CTNCA(1));
233 state->ct1ea = V3D_READ(V3D_CTNEA(1));
234
235 state->ct0cs = V3D_READ(V3D_CTNCS(0));
236 state->ct1cs = V3D_READ(V3D_CTNCS(1));
237
238 state->ct0ra0 = V3D_READ(V3D_CT00RA0);
239 state->ct1ra0 = V3D_READ(V3D_CT01RA0);
240
241 state->bpca = V3D_READ(V3D_BPCA);
242 state->bpcs = V3D_READ(V3D_BPCS);
243 state->bpoa = V3D_READ(V3D_BPOA);
244 state->bpos = V3D_READ(V3D_BPOS);
245
246 state->vpmbase = V3D_READ(V3D_VPMBASE);
247
248 state->dbge = V3D_READ(V3D_DBGE);
249 state->fdbgo = V3D_READ(V3D_FDBGO);
250 state->fdbgb = V3D_READ(V3D_FDBGB);
251 state->fdbgr = V3D_READ(V3D_FDBGR);
252 state->fdbgs = V3D_READ(V3D_FDBGS);
253 state->errstat = V3D_READ(V3D_ERRSTAT);
254
255
256
257
258
259
260
261
262 for (i = 0; i < kernel_state->user_state.bo_count; i++) {
263 struct vc4_bo *bo = to_vc4_bo(kernel_state->bo[i]);
264
265 if (bo->madv == __VC4_MADV_NOTSUPP)
266 continue;
267
268 mutex_lock(&bo->madv_lock);
269 if (!WARN_ON(bo->madv == __VC4_MADV_PURGED))
270 bo->madv = VC4_MADV_WILLNEED;
271 refcount_dec(&bo->usecnt);
272 mutex_unlock(&bo->madv_lock);
273 }
274
275 spin_lock_irqsave(&vc4->job_lock, irqflags);
276 if (vc4->hang_state) {
277 spin_unlock_irqrestore(&vc4->job_lock, irqflags);
278 vc4_free_hang_state(dev, kernel_state);
279 } else {
280 vc4->hang_state = kernel_state;
281 spin_unlock_irqrestore(&vc4->job_lock, irqflags);
282 }
283 }
284
285 static void
286 vc4_reset(struct drm_device *dev)
287 {
288 struct vc4_dev *vc4 = to_vc4_dev(dev);
289
290 DRM_INFO("Resetting GPU.\n");
291
292 mutex_lock(&vc4->power_lock);
293 if (vc4->power_refcount) {
294
295
296
297 pm_runtime_put_sync_suspend(&vc4->v3d->pdev->dev);
298 pm_runtime_get_sync(&vc4->v3d->pdev->dev);
299 }
300 mutex_unlock(&vc4->power_lock);
301
302 vc4_irq_reset(dev);
303
304
305
306
307
308 vc4_queue_hangcheck(dev);
309 }
310
311 static void
312 vc4_reset_work(struct work_struct *work)
313 {
314 struct vc4_dev *vc4 =
315 container_of(work, struct vc4_dev, hangcheck.reset_work);
316
317 vc4_save_hang_state(vc4->dev);
318
319 vc4_reset(vc4->dev);
320 }
321
322 static void
323 vc4_hangcheck_elapsed(struct timer_list *t)
324 {
325 struct vc4_dev *vc4 = from_timer(vc4, t, hangcheck.timer);
326 struct drm_device *dev = vc4->dev;
327 uint32_t ct0ca, ct1ca;
328 unsigned long irqflags;
329 struct vc4_exec_info *bin_exec, *render_exec;
330
331 spin_lock_irqsave(&vc4->job_lock, irqflags);
332
333 bin_exec = vc4_first_bin_job(vc4);
334 render_exec = vc4_first_render_job(vc4);
335
336
337 if (!bin_exec && !render_exec) {
338 spin_unlock_irqrestore(&vc4->job_lock, irqflags);
339 return;
340 }
341
342 ct0ca = V3D_READ(V3D_CTNCA(0));
343 ct1ca = V3D_READ(V3D_CTNCA(1));
344
345
346
347
348 if ((bin_exec && ct0ca != bin_exec->last_ct0ca) ||
349 (render_exec && ct1ca != render_exec->last_ct1ca)) {
350 if (bin_exec)
351 bin_exec->last_ct0ca = ct0ca;
352 if (render_exec)
353 render_exec->last_ct1ca = ct1ca;
354 spin_unlock_irqrestore(&vc4->job_lock, irqflags);
355 vc4_queue_hangcheck(dev);
356 return;
357 }
358
359 spin_unlock_irqrestore(&vc4->job_lock, irqflags);
360
361
362
363
364
365 schedule_work(&vc4->hangcheck.reset_work);
366 }
367
368 static void
369 submit_cl(struct drm_device *dev, uint32_t thread, uint32_t start, uint32_t end)
370 {
371 struct vc4_dev *vc4 = to_vc4_dev(dev);
372
373
374
375
376 V3D_WRITE(V3D_CTNCA(thread), start);
377 V3D_WRITE(V3D_CTNEA(thread), end);
378 }
379
380 int
381 vc4_wait_for_seqno(struct drm_device *dev, uint64_t seqno, uint64_t timeout_ns,
382 bool interruptible)
383 {
384 struct vc4_dev *vc4 = to_vc4_dev(dev);
385 int ret = 0;
386 unsigned long timeout_expire;
387 DEFINE_WAIT(wait);
388
389 if (vc4->finished_seqno >= seqno)
390 return 0;
391
392 if (timeout_ns == 0)
393 return -ETIME;
394
395 timeout_expire = jiffies + nsecs_to_jiffies(timeout_ns);
396
397 trace_vc4_wait_for_seqno_begin(dev, seqno, timeout_ns);
398 for (;;) {
399 prepare_to_wait(&vc4->job_wait_queue, &wait,
400 interruptible ? TASK_INTERRUPTIBLE :
401 TASK_UNINTERRUPTIBLE);
402
403 if (interruptible && signal_pending(current)) {
404 ret = -ERESTARTSYS;
405 break;
406 }
407
408 if (vc4->finished_seqno >= seqno)
409 break;
410
411 if (timeout_ns != ~0ull) {
412 if (time_after_eq(jiffies, timeout_expire)) {
413 ret = -ETIME;
414 break;
415 }
416 schedule_timeout(timeout_expire - jiffies);
417 } else {
418 schedule();
419 }
420 }
421
422 finish_wait(&vc4->job_wait_queue, &wait);
423 trace_vc4_wait_for_seqno_end(dev, seqno);
424
425 return ret;
426 }
427
428 static void
429 vc4_flush_caches(struct drm_device *dev)
430 {
431 struct vc4_dev *vc4 = to_vc4_dev(dev);
432
433
434
435
436
437 V3D_WRITE(V3D_L2CACTL,
438 V3D_L2CACTL_L2CCLR);
439
440 V3D_WRITE(V3D_SLCACTL,
441 VC4_SET_FIELD(0xf, V3D_SLCACTL_T1CC) |
442 VC4_SET_FIELD(0xf, V3D_SLCACTL_T0CC) |
443 VC4_SET_FIELD(0xf, V3D_SLCACTL_UCC) |
444 VC4_SET_FIELD(0xf, V3D_SLCACTL_ICC));
445 }
446
447 static void
448 vc4_flush_texture_caches(struct drm_device *dev)
449 {
450 struct vc4_dev *vc4 = to_vc4_dev(dev);
451
452 V3D_WRITE(V3D_L2CACTL,
453 V3D_L2CACTL_L2CCLR);
454
455 V3D_WRITE(V3D_SLCACTL,
456 VC4_SET_FIELD(0xf, V3D_SLCACTL_T1CC) |
457 VC4_SET_FIELD(0xf, V3D_SLCACTL_T0CC));
458 }
459
460
461
462
463
464
465 void
466 vc4_submit_next_bin_job(struct drm_device *dev)
467 {
468 struct vc4_dev *vc4 = to_vc4_dev(dev);
469 struct vc4_exec_info *exec;
470
471 again:
472 exec = vc4_first_bin_job(vc4);
473 if (!exec)
474 return;
475
476 vc4_flush_caches(dev);
477
478
479
480
481 if (exec->perfmon && vc4->active_perfmon != exec->perfmon)
482 vc4_perfmon_start(vc4, exec->perfmon);
483
484
485
486
487 if (exec->ct0ca != exec->ct0ea) {
488 submit_cl(dev, 0, exec->ct0ca, exec->ct0ea);
489 } else {
490 struct vc4_exec_info *next;
491
492 vc4_move_job_to_render(dev, exec);
493 next = vc4_first_bin_job(vc4);
494
495
496
497
498
499
500 if (next && next->perfmon == exec->perfmon)
501 goto again;
502 }
503 }
504
505 void
506 vc4_submit_next_render_job(struct drm_device *dev)
507 {
508 struct vc4_dev *vc4 = to_vc4_dev(dev);
509 struct vc4_exec_info *exec = vc4_first_render_job(vc4);
510
511 if (!exec)
512 return;
513
514
515
516
517
518
519
520 vc4_flush_texture_caches(dev);
521
522 submit_cl(dev, 1, exec->ct1ca, exec->ct1ea);
523 }
524
525 void
526 vc4_move_job_to_render(struct drm_device *dev, struct vc4_exec_info *exec)
527 {
528 struct vc4_dev *vc4 = to_vc4_dev(dev);
529 bool was_empty = list_empty(&vc4->render_job_list);
530
531 list_move_tail(&exec->head, &vc4->render_job_list);
532 if (was_empty)
533 vc4_submit_next_render_job(dev);
534 }
535
536 static void
537 vc4_update_bo_seqnos(struct vc4_exec_info *exec, uint64_t seqno)
538 {
539 struct vc4_bo *bo;
540 unsigned i;
541
542 for (i = 0; i < exec->bo_count; i++) {
543 bo = to_vc4_bo(&exec->bo[i]->base);
544 bo->seqno = seqno;
545
546 dma_resv_add_shared_fence(bo->base.base.resv, exec->fence);
547 }
548
549 list_for_each_entry(bo, &exec->unref_list, unref_head) {
550 bo->seqno = seqno;
551 }
552
553 for (i = 0; i < exec->rcl_write_bo_count; i++) {
554 bo = to_vc4_bo(&exec->rcl_write_bo[i]->base);
555 bo->write_seqno = seqno;
556
557 dma_resv_add_excl_fence(bo->base.base.resv, exec->fence);
558 }
559 }
560
561 static void
562 vc4_unlock_bo_reservations(struct drm_device *dev,
563 struct vc4_exec_info *exec,
564 struct ww_acquire_ctx *acquire_ctx)
565 {
566 int i;
567
568 for (i = 0; i < exec->bo_count; i++) {
569 struct drm_gem_object *bo = &exec->bo[i]->base;
570
571 ww_mutex_unlock(&bo->resv->lock);
572 }
573
574 ww_acquire_fini(acquire_ctx);
575 }
576
577
578
579
580
581
582
583
584 static int
585 vc4_lock_bo_reservations(struct drm_device *dev,
586 struct vc4_exec_info *exec,
587 struct ww_acquire_ctx *acquire_ctx)
588 {
589 int contended_lock = -1;
590 int i, ret;
591 struct drm_gem_object *bo;
592
593 ww_acquire_init(acquire_ctx, &reservation_ww_class);
594
595 retry:
596 if (contended_lock != -1) {
597 bo = &exec->bo[contended_lock]->base;
598 ret = ww_mutex_lock_slow_interruptible(&bo->resv->lock,
599 acquire_ctx);
600 if (ret) {
601 ww_acquire_done(acquire_ctx);
602 return ret;
603 }
604 }
605
606 for (i = 0; i < exec->bo_count; i++) {
607 if (i == contended_lock)
608 continue;
609
610 bo = &exec->bo[i]->base;
611
612 ret = ww_mutex_lock_interruptible(&bo->resv->lock, acquire_ctx);
613 if (ret) {
614 int j;
615
616 for (j = 0; j < i; j++) {
617 bo = &exec->bo[j]->base;
618 ww_mutex_unlock(&bo->resv->lock);
619 }
620
621 if (contended_lock != -1 && contended_lock >= i) {
622 bo = &exec->bo[contended_lock]->base;
623
624 ww_mutex_unlock(&bo->resv->lock);
625 }
626
627 if (ret == -EDEADLK) {
628 contended_lock = i;
629 goto retry;
630 }
631
632 ww_acquire_done(acquire_ctx);
633 return ret;
634 }
635 }
636
637 ww_acquire_done(acquire_ctx);
638
639
640
641
642 for (i = 0; i < exec->bo_count; i++) {
643 bo = &exec->bo[i]->base;
644
645 ret = dma_resv_reserve_shared(bo->resv, 1);
646 if (ret) {
647 vc4_unlock_bo_reservations(dev, exec, acquire_ctx);
648 return ret;
649 }
650 }
651
652 return 0;
653 }
654
655
656
657
658
659
660
661
662
663
664 static int
665 vc4_queue_submit(struct drm_device *dev, struct vc4_exec_info *exec,
666 struct ww_acquire_ctx *acquire_ctx,
667 struct drm_syncobj *out_sync)
668 {
669 struct vc4_dev *vc4 = to_vc4_dev(dev);
670 struct vc4_exec_info *renderjob;
671 uint64_t seqno;
672 unsigned long irqflags;
673 struct vc4_fence *fence;
674
675 fence = kzalloc(sizeof(*fence), GFP_KERNEL);
676 if (!fence)
677 return -ENOMEM;
678 fence->dev = dev;
679
680 spin_lock_irqsave(&vc4->job_lock, irqflags);
681
682 seqno = ++vc4->emit_seqno;
683 exec->seqno = seqno;
684
685 dma_fence_init(&fence->base, &vc4_fence_ops, &vc4->job_lock,
686 vc4->dma_fence_context, exec->seqno);
687 fence->seqno = exec->seqno;
688 exec->fence = &fence->base;
689
690 if (out_sync)
691 drm_syncobj_replace_fence(out_sync, exec->fence);
692
693 vc4_update_bo_seqnos(exec, seqno);
694
695 vc4_unlock_bo_reservations(dev, exec, acquire_ctx);
696
697 list_add_tail(&exec->head, &vc4->bin_job_list);
698
699
700
701
702
703
704 renderjob = vc4_first_render_job(vc4);
705 if (vc4_first_bin_job(vc4) == exec &&
706 (!renderjob || renderjob->perfmon == exec->perfmon)) {
707 vc4_submit_next_bin_job(dev);
708 vc4_queue_hangcheck(dev);
709 }
710
711 spin_unlock_irqrestore(&vc4->job_lock, irqflags);
712
713 return 0;
714 }
715
716
717
718
719
720
721
722
723
724
725
726
727 static int
728 vc4_cl_lookup_bos(struct drm_device *dev,
729 struct drm_file *file_priv,
730 struct vc4_exec_info *exec)
731 {
732 struct drm_vc4_submit_cl *args = exec->args;
733 uint32_t *handles;
734 int ret = 0;
735 int i;
736
737 exec->bo_count = args->bo_handle_count;
738
739 if (!exec->bo_count) {
740
741
742
743 DRM_DEBUG("Rendering requires BOs to validate\n");
744 return -EINVAL;
745 }
746
747 exec->bo = kvmalloc_array(exec->bo_count,
748 sizeof(struct drm_gem_cma_object *),
749 GFP_KERNEL | __GFP_ZERO);
750 if (!exec->bo) {
751 DRM_ERROR("Failed to allocate validated BO pointers\n");
752 return -ENOMEM;
753 }
754
755 handles = kvmalloc_array(exec->bo_count, sizeof(uint32_t), GFP_KERNEL);
756 if (!handles) {
757 ret = -ENOMEM;
758 DRM_ERROR("Failed to allocate incoming GEM handles\n");
759 goto fail;
760 }
761
762 if (copy_from_user(handles, u64_to_user_ptr(args->bo_handles),
763 exec->bo_count * sizeof(uint32_t))) {
764 ret = -EFAULT;
765 DRM_ERROR("Failed to copy in GEM handles\n");
766 goto fail;
767 }
768
769 spin_lock(&file_priv->table_lock);
770 for (i = 0; i < exec->bo_count; i++) {
771 struct drm_gem_object *bo = idr_find(&file_priv->object_idr,
772 handles[i]);
773 if (!bo) {
774 DRM_DEBUG("Failed to look up GEM BO %d: %d\n",
775 i, handles[i]);
776 ret = -EINVAL;
777 break;
778 }
779
780 drm_gem_object_get(bo);
781 exec->bo[i] = (struct drm_gem_cma_object *)bo;
782 }
783 spin_unlock(&file_priv->table_lock);
784
785 if (ret)
786 goto fail_put_bo;
787
788 for (i = 0; i < exec->bo_count; i++) {
789 ret = vc4_bo_inc_usecnt(to_vc4_bo(&exec->bo[i]->base));
790 if (ret)
791 goto fail_dec_usecnt;
792 }
793
794 kvfree(handles);
795 return 0;
796
797 fail_dec_usecnt:
798
799
800
801
802
803
804
805
806 for (i-- ; i >= 0; i--)
807 vc4_bo_dec_usecnt(to_vc4_bo(&exec->bo[i]->base));
808
809 fail_put_bo:
810
811 for (i = 0; i < exec->bo_count && exec->bo[i]; i++)
812 drm_gem_object_put_unlocked(&exec->bo[i]->base);
813
814 fail:
815 kvfree(handles);
816 kvfree(exec->bo);
817 exec->bo = NULL;
818 return ret;
819 }
820
821 static int
822 vc4_get_bcl(struct drm_device *dev, struct vc4_exec_info *exec)
823 {
824 struct drm_vc4_submit_cl *args = exec->args;
825 struct vc4_dev *vc4 = to_vc4_dev(dev);
826 void *temp = NULL;
827 void *bin;
828 int ret = 0;
829 uint32_t bin_offset = 0;
830 uint32_t shader_rec_offset = roundup(bin_offset + args->bin_cl_size,
831 16);
832 uint32_t uniforms_offset = shader_rec_offset + args->shader_rec_size;
833 uint32_t exec_size = uniforms_offset + args->uniforms_size;
834 uint32_t temp_size = exec_size + (sizeof(struct vc4_shader_state) *
835 args->shader_rec_count);
836 struct vc4_bo *bo;
837
838 if (shader_rec_offset < args->bin_cl_size ||
839 uniforms_offset < shader_rec_offset ||
840 exec_size < uniforms_offset ||
841 args->shader_rec_count >= (UINT_MAX /
842 sizeof(struct vc4_shader_state)) ||
843 temp_size < exec_size) {
844 DRM_DEBUG("overflow in exec arguments\n");
845 ret = -EINVAL;
846 goto fail;
847 }
848
849
850
851
852
853
854
855
856 temp = kvmalloc_array(temp_size, 1, GFP_KERNEL);
857 if (!temp) {
858 DRM_ERROR("Failed to allocate storage for copying "
859 "in bin/render CLs.\n");
860 ret = -ENOMEM;
861 goto fail;
862 }
863 bin = temp + bin_offset;
864 exec->shader_rec_u = temp + shader_rec_offset;
865 exec->uniforms_u = temp + uniforms_offset;
866 exec->shader_state = temp + exec_size;
867 exec->shader_state_size = args->shader_rec_count;
868
869 if (copy_from_user(bin,
870 u64_to_user_ptr(args->bin_cl),
871 args->bin_cl_size)) {
872 ret = -EFAULT;
873 goto fail;
874 }
875
876 if (copy_from_user(exec->shader_rec_u,
877 u64_to_user_ptr(args->shader_rec),
878 args->shader_rec_size)) {
879 ret = -EFAULT;
880 goto fail;
881 }
882
883 if (copy_from_user(exec->uniforms_u,
884 u64_to_user_ptr(args->uniforms),
885 args->uniforms_size)) {
886 ret = -EFAULT;
887 goto fail;
888 }
889
890 bo = vc4_bo_create(dev, exec_size, true, VC4_BO_TYPE_BCL);
891 if (IS_ERR(bo)) {
892 DRM_ERROR("Couldn't allocate BO for binning\n");
893 ret = PTR_ERR(bo);
894 goto fail;
895 }
896 exec->exec_bo = &bo->base;
897
898 list_add_tail(&to_vc4_bo(&exec->exec_bo->base)->unref_head,
899 &exec->unref_list);
900
901 exec->ct0ca = exec->exec_bo->paddr + bin_offset;
902
903 exec->bin_u = bin;
904
905 exec->shader_rec_v = exec->exec_bo->vaddr + shader_rec_offset;
906 exec->shader_rec_p = exec->exec_bo->paddr + shader_rec_offset;
907 exec->shader_rec_size = args->shader_rec_size;
908
909 exec->uniforms_v = exec->exec_bo->vaddr + uniforms_offset;
910 exec->uniforms_p = exec->exec_bo->paddr + uniforms_offset;
911 exec->uniforms_size = args->uniforms_size;
912
913 ret = vc4_validate_bin_cl(dev,
914 exec->exec_bo->vaddr + bin_offset,
915 bin,
916 exec);
917 if (ret)
918 goto fail;
919
920 ret = vc4_validate_shader_recs(dev, exec);
921 if (ret)
922 goto fail;
923
924 if (exec->found_tile_binning_mode_config_packet) {
925 ret = vc4_v3d_bin_bo_get(vc4, &exec->bin_bo_used);
926 if (ret)
927 goto fail;
928 }
929
930
931
932
933
934 ret = vc4_wait_for_seqno(dev, exec->bin_dep_seqno, ~0ull, true);
935
936 fail:
937 kvfree(temp);
938 return ret;
939 }
940
941 static void
942 vc4_complete_exec(struct drm_device *dev, struct vc4_exec_info *exec)
943 {
944 struct vc4_dev *vc4 = to_vc4_dev(dev);
945 unsigned long irqflags;
946 unsigned i;
947
948
949
950
951 if (exec->fence) {
952 dma_fence_signal(exec->fence);
953 dma_fence_put(exec->fence);
954 }
955
956 if (exec->bo) {
957 for (i = 0; i < exec->bo_count; i++) {
958 struct vc4_bo *bo = to_vc4_bo(&exec->bo[i]->base);
959
960 vc4_bo_dec_usecnt(bo);
961 drm_gem_object_put_unlocked(&exec->bo[i]->base);
962 }
963 kvfree(exec->bo);
964 }
965
966 while (!list_empty(&exec->unref_list)) {
967 struct vc4_bo *bo = list_first_entry(&exec->unref_list,
968 struct vc4_bo, unref_head);
969 list_del(&bo->unref_head);
970 drm_gem_object_put_unlocked(&bo->base.base);
971 }
972
973
974 spin_lock_irqsave(&vc4->job_lock, irqflags);
975 vc4->bin_alloc_used &= ~exec->bin_slots;
976 spin_unlock_irqrestore(&vc4->job_lock, irqflags);
977
978
979 if (exec->bin_bo_used)
980 vc4_v3d_bin_bo_put(vc4);
981
982
983 vc4_perfmon_put(exec->perfmon);
984
985 vc4_v3d_pm_put(vc4);
986
987 kfree(exec);
988 }
989
990 void
991 vc4_job_handle_completed(struct vc4_dev *vc4)
992 {
993 unsigned long irqflags;
994 struct vc4_seqno_cb *cb, *cb_temp;
995
996 spin_lock_irqsave(&vc4->job_lock, irqflags);
997 while (!list_empty(&vc4->job_done_list)) {
998 struct vc4_exec_info *exec =
999 list_first_entry(&vc4->job_done_list,
1000 struct vc4_exec_info, head);
1001 list_del(&exec->head);
1002
1003 spin_unlock_irqrestore(&vc4->job_lock, irqflags);
1004 vc4_complete_exec(vc4->dev, exec);
1005 spin_lock_irqsave(&vc4->job_lock, irqflags);
1006 }
1007
1008 list_for_each_entry_safe(cb, cb_temp, &vc4->seqno_cb_list, work.entry) {
1009 if (cb->seqno <= vc4->finished_seqno) {
1010 list_del_init(&cb->work.entry);
1011 schedule_work(&cb->work);
1012 }
1013 }
1014
1015 spin_unlock_irqrestore(&vc4->job_lock, irqflags);
1016 }
1017
1018 static void vc4_seqno_cb_work(struct work_struct *work)
1019 {
1020 struct vc4_seqno_cb *cb = container_of(work, struct vc4_seqno_cb, work);
1021
1022 cb->func(cb);
1023 }
1024
1025 int vc4_queue_seqno_cb(struct drm_device *dev,
1026 struct vc4_seqno_cb *cb, uint64_t seqno,
1027 void (*func)(struct vc4_seqno_cb *cb))
1028 {
1029 struct vc4_dev *vc4 = to_vc4_dev(dev);
1030 int ret = 0;
1031 unsigned long irqflags;
1032
1033 cb->func = func;
1034 INIT_WORK(&cb->work, vc4_seqno_cb_work);
1035
1036 spin_lock_irqsave(&vc4->job_lock, irqflags);
1037 if (seqno > vc4->finished_seqno) {
1038 cb->seqno = seqno;
1039 list_add_tail(&cb->work.entry, &vc4->seqno_cb_list);
1040 } else {
1041 schedule_work(&cb->work);
1042 }
1043 spin_unlock_irqrestore(&vc4->job_lock, irqflags);
1044
1045 return ret;
1046 }
1047
1048
1049
1050
1051
1052 static void
1053 vc4_job_done_work(struct work_struct *work)
1054 {
1055 struct vc4_dev *vc4 =
1056 container_of(work, struct vc4_dev, job_done_work);
1057
1058 vc4_job_handle_completed(vc4);
1059 }
1060
1061 static int
1062 vc4_wait_for_seqno_ioctl_helper(struct drm_device *dev,
1063 uint64_t seqno,
1064 uint64_t *timeout_ns)
1065 {
1066 unsigned long start = jiffies;
1067 int ret = vc4_wait_for_seqno(dev, seqno, *timeout_ns, true);
1068
1069 if ((ret == -EINTR || ret == -ERESTARTSYS) && *timeout_ns != ~0ull) {
1070 uint64_t delta = jiffies_to_nsecs(jiffies - start);
1071
1072 if (*timeout_ns >= delta)
1073 *timeout_ns -= delta;
1074 }
1075
1076 return ret;
1077 }
1078
1079 int
1080 vc4_wait_seqno_ioctl(struct drm_device *dev, void *data,
1081 struct drm_file *file_priv)
1082 {
1083 struct drm_vc4_wait_seqno *args = data;
1084
1085 return vc4_wait_for_seqno_ioctl_helper(dev, args->seqno,
1086 &args->timeout_ns);
1087 }
1088
1089 int
1090 vc4_wait_bo_ioctl(struct drm_device *dev, void *data,
1091 struct drm_file *file_priv)
1092 {
1093 int ret;
1094 struct drm_vc4_wait_bo *args = data;
1095 struct drm_gem_object *gem_obj;
1096 struct vc4_bo *bo;
1097
1098 if (args->pad != 0)
1099 return -EINVAL;
1100
1101 gem_obj = drm_gem_object_lookup(file_priv, args->handle);
1102 if (!gem_obj) {
1103 DRM_DEBUG("Failed to look up GEM BO %d\n", args->handle);
1104 return -EINVAL;
1105 }
1106 bo = to_vc4_bo(gem_obj);
1107
1108 ret = vc4_wait_for_seqno_ioctl_helper(dev, bo->seqno,
1109 &args->timeout_ns);
1110
1111 drm_gem_object_put_unlocked(gem_obj);
1112 return ret;
1113 }
1114
1115
1116
1117
1118
1119
1120
1121
1122
1123
1124
1125
1126
1127 int
1128 vc4_submit_cl_ioctl(struct drm_device *dev, void *data,
1129 struct drm_file *file_priv)
1130 {
1131 struct vc4_dev *vc4 = to_vc4_dev(dev);
1132 struct vc4_file *vc4file = file_priv->driver_priv;
1133 struct drm_vc4_submit_cl *args = data;
1134 struct drm_syncobj *out_sync = NULL;
1135 struct vc4_exec_info *exec;
1136 struct ww_acquire_ctx acquire_ctx;
1137 struct dma_fence *in_fence;
1138 int ret = 0;
1139
1140 if (!vc4->v3d) {
1141 DRM_DEBUG("VC4_SUBMIT_CL with no VC4 V3D probed\n");
1142 return -ENODEV;
1143 }
1144
1145 if ((args->flags & ~(VC4_SUBMIT_CL_USE_CLEAR_COLOR |
1146 VC4_SUBMIT_CL_FIXED_RCL_ORDER |
1147 VC4_SUBMIT_CL_RCL_ORDER_INCREASING_X |
1148 VC4_SUBMIT_CL_RCL_ORDER_INCREASING_Y)) != 0) {
1149 DRM_DEBUG("Unknown flags: 0x%02x\n", args->flags);
1150 return -EINVAL;
1151 }
1152
1153 if (args->pad2 != 0) {
1154 DRM_DEBUG("Invalid pad: 0x%08x\n", args->pad2);
1155 return -EINVAL;
1156 }
1157
1158 exec = kcalloc(1, sizeof(*exec), GFP_KERNEL);
1159 if (!exec) {
1160 DRM_ERROR("malloc failure on exec struct\n");
1161 return -ENOMEM;
1162 }
1163
1164 ret = vc4_v3d_pm_get(vc4);
1165 if (ret) {
1166 kfree(exec);
1167 return ret;
1168 }
1169
1170 exec->args = args;
1171 INIT_LIST_HEAD(&exec->unref_list);
1172
1173 ret = vc4_cl_lookup_bos(dev, file_priv, exec);
1174 if (ret)
1175 goto fail;
1176
1177 if (args->perfmonid) {
1178 exec->perfmon = vc4_perfmon_find(vc4file,
1179 args->perfmonid);
1180 if (!exec->perfmon) {
1181 ret = -ENOENT;
1182 goto fail;
1183 }
1184 }
1185
1186 if (args->in_sync) {
1187 ret = drm_syncobj_find_fence(file_priv, args->in_sync,
1188 0, 0, &in_fence);
1189 if (ret)
1190 goto fail;
1191
1192
1193
1194
1195
1196
1197 if (!dma_fence_match_context(in_fence,
1198 vc4->dma_fence_context)) {
1199 ret = dma_fence_wait(in_fence, true);
1200 if (ret) {
1201 dma_fence_put(in_fence);
1202 goto fail;
1203 }
1204 }
1205
1206 dma_fence_put(in_fence);
1207 }
1208
1209 if (exec->args->bin_cl_size != 0) {
1210 ret = vc4_get_bcl(dev, exec);
1211 if (ret)
1212 goto fail;
1213 } else {
1214 exec->ct0ca = 0;
1215 exec->ct0ea = 0;
1216 }
1217
1218 ret = vc4_get_rcl(dev, exec);
1219 if (ret)
1220 goto fail;
1221
1222 ret = vc4_lock_bo_reservations(dev, exec, &acquire_ctx);
1223 if (ret)
1224 goto fail;
1225
1226 if (args->out_sync) {
1227 out_sync = drm_syncobj_find(file_priv, args->out_sync);
1228 if (!out_sync) {
1229 ret = -EINVAL;
1230 goto fail;
1231 }
1232
1233
1234
1235
1236
1237
1238 }
1239
1240
1241
1242
1243 exec->args = NULL;
1244
1245 ret = vc4_queue_submit(dev, exec, &acquire_ctx, out_sync);
1246
1247
1248
1249
1250 if (out_sync)
1251 drm_syncobj_put(out_sync);
1252
1253 if (ret)
1254 goto fail;
1255
1256
1257 args->seqno = vc4->emit_seqno;
1258
1259 return 0;
1260
1261 fail:
1262 vc4_complete_exec(vc4->dev, exec);
1263
1264 return ret;
1265 }
1266
1267 void
1268 vc4_gem_init(struct drm_device *dev)
1269 {
1270 struct vc4_dev *vc4 = to_vc4_dev(dev);
1271
1272 vc4->dma_fence_context = dma_fence_context_alloc(1);
1273
1274 INIT_LIST_HEAD(&vc4->bin_job_list);
1275 INIT_LIST_HEAD(&vc4->render_job_list);
1276 INIT_LIST_HEAD(&vc4->job_done_list);
1277 INIT_LIST_HEAD(&vc4->seqno_cb_list);
1278 spin_lock_init(&vc4->job_lock);
1279
1280 INIT_WORK(&vc4->hangcheck.reset_work, vc4_reset_work);
1281 timer_setup(&vc4->hangcheck.timer, vc4_hangcheck_elapsed, 0);
1282
1283 INIT_WORK(&vc4->job_done_work, vc4_job_done_work);
1284
1285 mutex_init(&vc4->power_lock);
1286
1287 INIT_LIST_HEAD(&vc4->purgeable.list);
1288 mutex_init(&vc4->purgeable.lock);
1289 }
1290
1291 void
1292 vc4_gem_destroy(struct drm_device *dev)
1293 {
1294 struct vc4_dev *vc4 = to_vc4_dev(dev);
1295
1296
1297
1298
1299 WARN_ON(vc4->emit_seqno != vc4->finished_seqno);
1300
1301
1302
1303
1304 if (vc4->bin_bo) {
1305 drm_gem_object_put_unlocked(&vc4->bin_bo->base.base);
1306 vc4->bin_bo = NULL;
1307 }
1308
1309 if (vc4->hang_state)
1310 vc4_free_hang_state(dev, vc4->hang_state);
1311 }
1312
1313 int vc4_gem_madvise_ioctl(struct drm_device *dev, void *data,
1314 struct drm_file *file_priv)
1315 {
1316 struct drm_vc4_gem_madvise *args = data;
1317 struct drm_gem_object *gem_obj;
1318 struct vc4_bo *bo;
1319 int ret;
1320
1321 switch (args->madv) {
1322 case VC4_MADV_DONTNEED:
1323 case VC4_MADV_WILLNEED:
1324 break;
1325 default:
1326 return -EINVAL;
1327 }
1328
1329 if (args->pad != 0)
1330 return -EINVAL;
1331
1332 gem_obj = drm_gem_object_lookup(file_priv, args->handle);
1333 if (!gem_obj) {
1334 DRM_DEBUG("Failed to look up GEM BO %d\n", args->handle);
1335 return -ENOENT;
1336 }
1337
1338 bo = to_vc4_bo(gem_obj);
1339
1340
1341 if (bo->madv == __VC4_MADV_NOTSUPP) {
1342 DRM_DEBUG("madvise not supported on this BO\n");
1343 ret = -EINVAL;
1344 goto out_put_gem;
1345 }
1346
1347
1348
1349
1350 if (gem_obj->import_attach) {
1351 DRM_DEBUG("madvise not supported on imported BOs\n");
1352 ret = -EINVAL;
1353 goto out_put_gem;
1354 }
1355
1356 mutex_lock(&bo->madv_lock);
1357
1358 if (args->madv == VC4_MADV_DONTNEED && bo->madv == VC4_MADV_WILLNEED &&
1359 !refcount_read(&bo->usecnt)) {
1360
1361
1362
1363
1364 vc4_bo_add_to_purgeable_pool(bo);
1365 } else if (args->madv == VC4_MADV_WILLNEED &&
1366 bo->madv == VC4_MADV_DONTNEED &&
1367 !refcount_read(&bo->usecnt)) {
1368
1369
1370
1371 vc4_bo_remove_from_purgeable_pool(bo);
1372 }
1373
1374
1375 args->retained = bo->madv != __VC4_MADV_PURGED;
1376
1377
1378 if (bo->madv != __VC4_MADV_PURGED)
1379 bo->madv = args->madv;
1380
1381 mutex_unlock(&bo->madv_lock);
1382
1383 ret = 0;
1384
1385 out_put_gem:
1386 drm_gem_object_put_unlocked(gem_obj);
1387
1388 return ret;
1389 }