This source file includes following definitions.
- amdgpu_fence_slab_init
- amdgpu_fence_slab_fini
- to_amdgpu_fence
- amdgpu_fence_write
- amdgpu_fence_read
- amdgpu_fence_emit
- amdgpu_fence_emit_polling
- amdgpu_fence_schedule_fallback
- amdgpu_fence_process
- amdgpu_fence_fallback
- amdgpu_fence_wait_empty
- amdgpu_fence_wait_polling
- amdgpu_fence_count_emitted
- amdgpu_fence_driver_start_ring
- amdgpu_fence_driver_init_ring
- amdgpu_fence_driver_init
- amdgpu_fence_driver_fini
- amdgpu_fence_driver_suspend
- amdgpu_fence_driver_resume
- amdgpu_fence_driver_force_completion
- amdgpu_fence_get_driver_name
- amdgpu_fence_get_timeline_name
- amdgpu_fence_enable_signaling
- amdgpu_fence_free
- amdgpu_fence_release
- amdgpu_debugfs_fence_info
- amdgpu_debugfs_gpu_recover
- amdgpu_debugfs_fence_init
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31 #include <linux/seq_file.h>
32 #include <linux/atomic.h>
33 #include <linux/wait.h>
34 #include <linux/kref.h>
35 #include <linux/slab.h>
36 #include <linux/firmware.h>
37
38 #include <drm/drm_debugfs.h>
39
40 #include "amdgpu.h"
41 #include "amdgpu_trace.h"
42
43
44
45
46
47
48
49
50
51
52 struct amdgpu_fence {
53 struct dma_fence base;
54
55
56 struct amdgpu_ring *ring;
57 };
58
59 static struct kmem_cache *amdgpu_fence_slab;
60
61 int amdgpu_fence_slab_init(void)
62 {
63 amdgpu_fence_slab = kmem_cache_create(
64 "amdgpu_fence", sizeof(struct amdgpu_fence), 0,
65 SLAB_HWCACHE_ALIGN, NULL);
66 if (!amdgpu_fence_slab)
67 return -ENOMEM;
68 return 0;
69 }
70
71 void amdgpu_fence_slab_fini(void)
72 {
73 rcu_barrier();
74 kmem_cache_destroy(amdgpu_fence_slab);
75 }
76
77
78
79 static const struct dma_fence_ops amdgpu_fence_ops;
80 static inline struct amdgpu_fence *to_amdgpu_fence(struct dma_fence *f)
81 {
82 struct amdgpu_fence *__f = container_of(f, struct amdgpu_fence, base);
83
84 if (__f->base.ops == &amdgpu_fence_ops)
85 return __f;
86
87 return NULL;
88 }
89
90
91
92
93
94
95
96
97
98 static void amdgpu_fence_write(struct amdgpu_ring *ring, u32 seq)
99 {
100 struct amdgpu_fence_driver *drv = &ring->fence_drv;
101
102 if (drv->cpu_addr)
103 *drv->cpu_addr = cpu_to_le32(seq);
104 }
105
106
107
108
109
110
111
112
113
114 static u32 amdgpu_fence_read(struct amdgpu_ring *ring)
115 {
116 struct amdgpu_fence_driver *drv = &ring->fence_drv;
117 u32 seq = 0;
118
119 if (drv->cpu_addr)
120 seq = le32_to_cpu(*drv->cpu_addr);
121 else
122 seq = atomic_read(&drv->last_seq);
123
124 return seq;
125 }
126
127
128
129
130
131
132
133
134
135
136 int amdgpu_fence_emit(struct amdgpu_ring *ring, struct dma_fence **f,
137 unsigned flags)
138 {
139 struct amdgpu_device *adev = ring->adev;
140 struct amdgpu_fence *fence;
141 struct dma_fence __rcu **ptr;
142 uint32_t seq;
143 int r;
144
145 fence = kmem_cache_alloc(amdgpu_fence_slab, GFP_KERNEL);
146 if (fence == NULL)
147 return -ENOMEM;
148
149 seq = ++ring->fence_drv.sync_seq;
150 fence->ring = ring;
151 dma_fence_init(&fence->base, &amdgpu_fence_ops,
152 &ring->fence_drv.lock,
153 adev->fence_context + ring->idx,
154 seq);
155 amdgpu_ring_emit_fence(ring, ring->fence_drv.gpu_addr,
156 seq, flags | AMDGPU_FENCE_FLAG_INT);
157
158 ptr = &ring->fence_drv.fences[seq & ring->fence_drv.num_fences_mask];
159 if (unlikely(rcu_dereference_protected(*ptr, 1))) {
160 struct dma_fence *old;
161
162 rcu_read_lock();
163 old = dma_fence_get_rcu_safe(ptr);
164 rcu_read_unlock();
165
166 if (old) {
167 r = dma_fence_wait(old, false);
168 dma_fence_put(old);
169 if (r)
170 return r;
171 }
172 }
173
174
175
176
177 rcu_assign_pointer(*ptr, dma_fence_get(&fence->base));
178
179 *f = &fence->base;
180
181 return 0;
182 }
183
184
185
186
187
188
189
190
191
192
193
194 int amdgpu_fence_emit_polling(struct amdgpu_ring *ring, uint32_t *s)
195 {
196 uint32_t seq;
197
198 if (!s)
199 return -EINVAL;
200
201 seq = ++ring->fence_drv.sync_seq;
202 amdgpu_ring_emit_fence(ring, ring->fence_drv.gpu_addr,
203 seq, 0);
204
205 *s = seq;
206
207 return 0;
208 }
209
210
211
212
213
214
215
216
217 static void amdgpu_fence_schedule_fallback(struct amdgpu_ring *ring)
218 {
219 mod_timer(&ring->fence_drv.fallback_timer,
220 jiffies + AMDGPU_FENCE_JIFFIES_TIMEOUT);
221 }
222
223
224
225
226
227
228
229
230
231
232
233
234 bool amdgpu_fence_process(struct amdgpu_ring *ring)
235 {
236 struct amdgpu_fence_driver *drv = &ring->fence_drv;
237 uint32_t seq, last_seq;
238 int r;
239
240 do {
241 last_seq = atomic_read(&ring->fence_drv.last_seq);
242 seq = amdgpu_fence_read(ring);
243
244 } while (atomic_cmpxchg(&drv->last_seq, last_seq, seq) != last_seq);
245
246 if (del_timer(&ring->fence_drv.fallback_timer) &&
247 seq != ring->fence_drv.sync_seq)
248 amdgpu_fence_schedule_fallback(ring);
249
250 if (unlikely(seq == last_seq))
251 return false;
252
253 last_seq &= drv->num_fences_mask;
254 seq &= drv->num_fences_mask;
255
256 do {
257 struct dma_fence *fence, **ptr;
258
259 ++last_seq;
260 last_seq &= drv->num_fences_mask;
261 ptr = &drv->fences[last_seq];
262
263
264 fence = rcu_dereference_protected(*ptr, 1);
265 RCU_INIT_POINTER(*ptr, NULL);
266
267 if (!fence)
268 continue;
269
270 r = dma_fence_signal(fence);
271 if (!r)
272 DMA_FENCE_TRACE(fence, "signaled from irq context\n");
273 else
274 BUG();
275
276 dma_fence_put(fence);
277 } while (last_seq != seq);
278
279 return true;
280 }
281
282
283
284
285
286
287
288
289 static void amdgpu_fence_fallback(struct timer_list *t)
290 {
291 struct amdgpu_ring *ring = from_timer(ring, t,
292 fence_drv.fallback_timer);
293
294 if (amdgpu_fence_process(ring))
295 DRM_WARN("Fence fallback timer expired on ring %s\n", ring->name);
296 }
297
298
299
300
301
302
303
304
305
306
307 int amdgpu_fence_wait_empty(struct amdgpu_ring *ring)
308 {
309 uint64_t seq = READ_ONCE(ring->fence_drv.sync_seq);
310 struct dma_fence *fence, **ptr;
311 int r;
312
313 if (!seq)
314 return 0;
315
316 ptr = &ring->fence_drv.fences[seq & ring->fence_drv.num_fences_mask];
317 rcu_read_lock();
318 fence = rcu_dereference(*ptr);
319 if (!fence || !dma_fence_get_rcu(fence)) {
320 rcu_read_unlock();
321 return 0;
322 }
323 rcu_read_unlock();
324
325 r = dma_fence_wait(fence, false);
326 dma_fence_put(fence);
327 return r;
328 }
329
330
331
332
333
334
335
336
337
338
339
340 signed long amdgpu_fence_wait_polling(struct amdgpu_ring *ring,
341 uint32_t wait_seq,
342 signed long timeout)
343 {
344 uint32_t seq;
345
346 do {
347 seq = amdgpu_fence_read(ring);
348 udelay(5);
349 timeout -= 5;
350 } while ((int32_t)(wait_seq - seq) > 0 && timeout > 0);
351
352 return timeout > 0 ? timeout : 0;
353 }
354
355
356
357
358
359
360
361
362
363 unsigned amdgpu_fence_count_emitted(struct amdgpu_ring *ring)
364 {
365 uint64_t emitted;
366
367
368
369
370 amdgpu_fence_process(ring);
371 emitted = 0x100000000ull;
372 emitted -= atomic_read(&ring->fence_drv.last_seq);
373 emitted += READ_ONCE(ring->fence_drv.sync_seq);
374 return lower_32_bits(emitted);
375 }
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390 int amdgpu_fence_driver_start_ring(struct amdgpu_ring *ring,
391 struct amdgpu_irq_src *irq_src,
392 unsigned irq_type)
393 {
394 struct amdgpu_device *adev = ring->adev;
395 uint64_t index;
396
397 if (ring->funcs->type != AMDGPU_RING_TYPE_UVD) {
398 ring->fence_drv.cpu_addr = &adev->wb.wb[ring->fence_offs];
399 ring->fence_drv.gpu_addr = adev->wb.gpu_addr + (ring->fence_offs * 4);
400 } else {
401
402 index = ALIGN(adev->uvd.fw->size, 8);
403 ring->fence_drv.cpu_addr = adev->uvd.inst[ring->me].cpu_addr + index;
404 ring->fence_drv.gpu_addr = adev->uvd.inst[ring->me].gpu_addr + index;
405 }
406 amdgpu_fence_write(ring, atomic_read(&ring->fence_drv.last_seq));
407 amdgpu_irq_get(adev, irq_src, irq_type);
408
409 ring->fence_drv.irq_src = irq_src;
410 ring->fence_drv.irq_type = irq_type;
411 ring->fence_drv.initialized = true;
412
413 DRM_DEV_DEBUG(adev->dev, "fence driver on ring %s use gpu addr "
414 "0x%016llx, cpu addr 0x%p\n", ring->name,
415 ring->fence_drv.gpu_addr, ring->fence_drv.cpu_addr);
416 return 0;
417 }
418
419
420
421
422
423
424
425
426
427
428
429 int amdgpu_fence_driver_init_ring(struct amdgpu_ring *ring,
430 unsigned num_hw_submission)
431 {
432 struct amdgpu_device *adev = ring->adev;
433 long timeout;
434 int r;
435
436 if (!adev)
437 return -EINVAL;
438
439
440 if ((num_hw_submission & (num_hw_submission - 1)) != 0)
441 return -EINVAL;
442
443 ring->fence_drv.cpu_addr = NULL;
444 ring->fence_drv.gpu_addr = 0;
445 ring->fence_drv.sync_seq = 0;
446 atomic_set(&ring->fence_drv.last_seq, 0);
447 ring->fence_drv.initialized = false;
448
449 timer_setup(&ring->fence_drv.fallback_timer, amdgpu_fence_fallback, 0);
450
451 ring->fence_drv.num_fences_mask = num_hw_submission * 2 - 1;
452 spin_lock_init(&ring->fence_drv.lock);
453 ring->fence_drv.fences = kcalloc(num_hw_submission * 2, sizeof(void *),
454 GFP_KERNEL);
455 if (!ring->fence_drv.fences)
456 return -ENOMEM;
457
458
459 if (ring->funcs->type != AMDGPU_RING_TYPE_KIQ) {
460 switch (ring->funcs->type) {
461 case AMDGPU_RING_TYPE_GFX:
462 timeout = adev->gfx_timeout;
463 break;
464 case AMDGPU_RING_TYPE_COMPUTE:
465
466
467
468
469
470
471
472
473 if (!amdgpu_sriov_vf(ring->adev))
474 timeout = adev->compute_timeout;
475 else
476 timeout = adev->gfx_timeout;
477 break;
478 case AMDGPU_RING_TYPE_SDMA:
479 timeout = adev->sdma_timeout;
480 break;
481 default:
482 timeout = adev->video_timeout;
483 break;
484 }
485
486 r = drm_sched_init(&ring->sched, &amdgpu_sched_ops,
487 num_hw_submission, amdgpu_job_hang_limit,
488 timeout, ring->name);
489 if (r) {
490 DRM_ERROR("Failed to create scheduler on ring %s.\n",
491 ring->name);
492 return r;
493 }
494 }
495
496 return 0;
497 }
498
499
500
501
502
503
504
505
506
507
508
509
510
511 int amdgpu_fence_driver_init(struct amdgpu_device *adev)
512 {
513 if (amdgpu_debugfs_fence_init(adev))
514 dev_err(adev->dev, "fence debugfs file creation failed\n");
515
516 return 0;
517 }
518
519
520
521
522
523
524
525
526
527 void amdgpu_fence_driver_fini(struct amdgpu_device *adev)
528 {
529 unsigned i, j;
530 int r;
531
532 for (i = 0; i < AMDGPU_MAX_RINGS; i++) {
533 struct amdgpu_ring *ring = adev->rings[i];
534
535 if (!ring || !ring->fence_drv.initialized)
536 continue;
537 r = amdgpu_fence_wait_empty(ring);
538 if (r) {
539
540 amdgpu_fence_driver_force_completion(ring);
541 }
542 amdgpu_irq_put(adev, ring->fence_drv.irq_src,
543 ring->fence_drv.irq_type);
544 drm_sched_fini(&ring->sched);
545 del_timer_sync(&ring->fence_drv.fallback_timer);
546 for (j = 0; j <= ring->fence_drv.num_fences_mask; ++j)
547 dma_fence_put(ring->fence_drv.fences[j]);
548 kfree(ring->fence_drv.fences);
549 ring->fence_drv.fences = NULL;
550 ring->fence_drv.initialized = false;
551 }
552 }
553
554
555
556
557
558
559
560
561
562 void amdgpu_fence_driver_suspend(struct amdgpu_device *adev)
563 {
564 int i, r;
565
566 for (i = 0; i < AMDGPU_MAX_RINGS; i++) {
567 struct amdgpu_ring *ring = adev->rings[i];
568 if (!ring || !ring->fence_drv.initialized)
569 continue;
570
571
572 r = amdgpu_fence_wait_empty(ring);
573 if (r) {
574
575 amdgpu_fence_driver_force_completion(ring);
576 }
577
578
579 amdgpu_irq_put(adev, ring->fence_drv.irq_src,
580 ring->fence_drv.irq_type);
581 }
582 }
583
584
585
586
587
588
589
590
591
592
593
594
595
596 void amdgpu_fence_driver_resume(struct amdgpu_device *adev)
597 {
598 int i;
599
600 for (i = 0; i < AMDGPU_MAX_RINGS; i++) {
601 struct amdgpu_ring *ring = adev->rings[i];
602 if (!ring || !ring->fence_drv.initialized)
603 continue;
604
605
606 amdgpu_irq_get(adev, ring->fence_drv.irq_src,
607 ring->fence_drv.irq_type);
608 }
609 }
610
611
612
613
614
615
616
617 void amdgpu_fence_driver_force_completion(struct amdgpu_ring *ring)
618 {
619 amdgpu_fence_write(ring, ring->fence_drv.sync_seq);
620 amdgpu_fence_process(ring);
621 }
622
623
624
625
626
627 static const char *amdgpu_fence_get_driver_name(struct dma_fence *fence)
628 {
629 return "amdgpu";
630 }
631
632 static const char *amdgpu_fence_get_timeline_name(struct dma_fence *f)
633 {
634 struct amdgpu_fence *fence = to_amdgpu_fence(f);
635 return (const char *)fence->ring->name;
636 }
637
638
639
640
641
642
643
644
645
646 static bool amdgpu_fence_enable_signaling(struct dma_fence *f)
647 {
648 struct amdgpu_fence *fence = to_amdgpu_fence(f);
649 struct amdgpu_ring *ring = fence->ring;
650
651 if (!timer_pending(&ring->fence_drv.fallback_timer))
652 amdgpu_fence_schedule_fallback(ring);
653
654 DMA_FENCE_TRACE(&fence->base, "armed on ring %i!\n", ring->idx);
655
656 return true;
657 }
658
659
660
661
662
663
664
665
666 static void amdgpu_fence_free(struct rcu_head *rcu)
667 {
668 struct dma_fence *f = container_of(rcu, struct dma_fence, rcu);
669 struct amdgpu_fence *fence = to_amdgpu_fence(f);
670 kmem_cache_free(amdgpu_fence_slab, fence);
671 }
672
673
674
675
676
677
678
679
680
681 static void amdgpu_fence_release(struct dma_fence *f)
682 {
683 call_rcu(&f->rcu, amdgpu_fence_free);
684 }
685
686 static const struct dma_fence_ops amdgpu_fence_ops = {
687 .get_driver_name = amdgpu_fence_get_driver_name,
688 .get_timeline_name = amdgpu_fence_get_timeline_name,
689 .enable_signaling = amdgpu_fence_enable_signaling,
690 .release = amdgpu_fence_release,
691 };
692
693
694
695
696 #if defined(CONFIG_DEBUG_FS)
697 static int amdgpu_debugfs_fence_info(struct seq_file *m, void *data)
698 {
699 struct drm_info_node *node = (struct drm_info_node *)m->private;
700 struct drm_device *dev = node->minor->dev;
701 struct amdgpu_device *adev = dev->dev_private;
702 int i;
703
704 for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
705 struct amdgpu_ring *ring = adev->rings[i];
706 if (!ring || !ring->fence_drv.initialized)
707 continue;
708
709 amdgpu_fence_process(ring);
710
711 seq_printf(m, "--- ring %d (%s) ---\n", i, ring->name);
712 seq_printf(m, "Last signaled fence 0x%08x\n",
713 atomic_read(&ring->fence_drv.last_seq));
714 seq_printf(m, "Last emitted 0x%08x\n",
715 ring->fence_drv.sync_seq);
716
717 if (ring->funcs->type == AMDGPU_RING_TYPE_GFX ||
718 ring->funcs->type == AMDGPU_RING_TYPE_SDMA) {
719 seq_printf(m, "Last signaled trailing fence 0x%08x\n",
720 le32_to_cpu(*ring->trail_fence_cpu_addr));
721 seq_printf(m, "Last emitted 0x%08x\n",
722 ring->trail_seq);
723 }
724
725 if (ring->funcs->type != AMDGPU_RING_TYPE_GFX)
726 continue;
727
728
729 seq_printf(m, "Last preempted 0x%08x\n",
730 le32_to_cpu(*(ring->fence_drv.cpu_addr + 2)));
731
732 seq_printf(m, "Last reset 0x%08x\n",
733 le32_to_cpu(*(ring->fence_drv.cpu_addr + 4)));
734
735 seq_printf(m, "Last both 0x%08x\n",
736 le32_to_cpu(*(ring->fence_drv.cpu_addr + 6)));
737 }
738 return 0;
739 }
740
741
742
743
744
745
746 static int amdgpu_debugfs_gpu_recover(struct seq_file *m, void *data)
747 {
748 struct drm_info_node *node = (struct drm_info_node *) m->private;
749 struct drm_device *dev = node->minor->dev;
750 struct amdgpu_device *adev = dev->dev_private;
751
752 seq_printf(m, "gpu recover\n");
753 amdgpu_device_gpu_recover(adev, NULL);
754
755 return 0;
756 }
757
758 static const struct drm_info_list amdgpu_debugfs_fence_list[] = {
759 {"amdgpu_fence_info", &amdgpu_debugfs_fence_info, 0, NULL},
760 {"amdgpu_gpu_recover", &amdgpu_debugfs_gpu_recover, 0, NULL}
761 };
762
763 static const struct drm_info_list amdgpu_debugfs_fence_list_sriov[] = {
764 {"amdgpu_fence_info", &amdgpu_debugfs_fence_info, 0, NULL},
765 };
766 #endif
767
768 int amdgpu_debugfs_fence_init(struct amdgpu_device *adev)
769 {
770 #if defined(CONFIG_DEBUG_FS)
771 if (amdgpu_sriov_vf(adev))
772 return amdgpu_debugfs_add_files(adev, amdgpu_debugfs_fence_list_sriov, 1);
773 return amdgpu_debugfs_add_files(adev, amdgpu_debugfs_fence_list, 2);
774 #else
775 return 0;
776 #endif
777 }
778