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
52struct amdgpu_fence {
53 struct dma_fence base;
54
55
56 struct amdgpu_ring *ring;
57};
58
59static struct kmem_cache *amdgpu_fence_slab;
60
61int 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
71void amdgpu_fence_slab_fini(void)
72{
73 rcu_barrier();
74 kmem_cache_destroy(amdgpu_fence_slab);
75}
76
77
78
79static const struct dma_fence_ops amdgpu_fence_ops;
80static 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
98static 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
114static 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
136int 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
194int 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
217static 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
234bool 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
289static 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
307int 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
340signed 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
363unsigned 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
390int 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
429int 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 timeout = adev->compute_timeout;
466 break;
467 case AMDGPU_RING_TYPE_SDMA:
468 timeout = adev->sdma_timeout;
469 break;
470 default:
471 timeout = adev->video_timeout;
472 break;
473 }
474
475 r = drm_sched_init(&ring->sched, &amdgpu_sched_ops,
476 num_hw_submission, amdgpu_job_hang_limit,
477 timeout, ring->name);
478 if (r) {
479 DRM_ERROR("Failed to create scheduler on ring %s.\n",
480 ring->name);
481 return r;
482 }
483 }
484
485 return 0;
486}
487
488
489
490
491
492
493
494
495
496
497
498
499
500int amdgpu_fence_driver_init(struct amdgpu_device *adev)
501{
502 if (amdgpu_debugfs_fence_init(adev))
503 dev_err(adev->dev, "fence debugfs file creation failed\n");
504
505 return 0;
506}
507
508
509
510
511
512
513
514
515
516void amdgpu_fence_driver_fini(struct amdgpu_device *adev)
517{
518 unsigned i, j;
519 int r;
520
521 for (i = 0; i < AMDGPU_MAX_RINGS; i++) {
522 struct amdgpu_ring *ring = adev->rings[i];
523
524 if (!ring || !ring->fence_drv.initialized)
525 continue;
526 r = amdgpu_fence_wait_empty(ring);
527 if (r) {
528
529 amdgpu_fence_driver_force_completion(ring);
530 }
531 amdgpu_irq_put(adev, ring->fence_drv.irq_src,
532 ring->fence_drv.irq_type);
533 drm_sched_fini(&ring->sched);
534 del_timer_sync(&ring->fence_drv.fallback_timer);
535 for (j = 0; j <= ring->fence_drv.num_fences_mask; ++j)
536 dma_fence_put(ring->fence_drv.fences[j]);
537 kfree(ring->fence_drv.fences);
538 ring->fence_drv.fences = NULL;
539 ring->fence_drv.initialized = false;
540 }
541}
542
543
544
545
546
547
548
549
550
551void amdgpu_fence_driver_suspend(struct amdgpu_device *adev)
552{
553 int i, r;
554
555 for (i = 0; i < AMDGPU_MAX_RINGS; i++) {
556 struct amdgpu_ring *ring = adev->rings[i];
557 if (!ring || !ring->fence_drv.initialized)
558 continue;
559
560
561 r = amdgpu_fence_wait_empty(ring);
562 if (r) {
563
564 amdgpu_fence_driver_force_completion(ring);
565 }
566
567
568 amdgpu_irq_put(adev, ring->fence_drv.irq_src,
569 ring->fence_drv.irq_type);
570 }
571}
572
573
574
575
576
577
578
579
580
581
582
583
584
585void amdgpu_fence_driver_resume(struct amdgpu_device *adev)
586{
587 int i;
588
589 for (i = 0; i < AMDGPU_MAX_RINGS; i++) {
590 struct amdgpu_ring *ring = adev->rings[i];
591 if (!ring || !ring->fence_drv.initialized)
592 continue;
593
594
595 amdgpu_irq_get(adev, ring->fence_drv.irq_src,
596 ring->fence_drv.irq_type);
597 }
598}
599
600
601
602
603
604
605
606void amdgpu_fence_driver_force_completion(struct amdgpu_ring *ring)
607{
608 amdgpu_fence_write(ring, ring->fence_drv.sync_seq);
609 amdgpu_fence_process(ring);
610}
611
612
613
614
615
616static const char *amdgpu_fence_get_driver_name(struct dma_fence *fence)
617{
618 return "amdgpu";
619}
620
621static const char *amdgpu_fence_get_timeline_name(struct dma_fence *f)
622{
623 struct amdgpu_fence *fence = to_amdgpu_fence(f);
624 return (const char *)fence->ring->name;
625}
626
627
628
629
630
631
632
633
634
635static bool amdgpu_fence_enable_signaling(struct dma_fence *f)
636{
637 struct amdgpu_fence *fence = to_amdgpu_fence(f);
638 struct amdgpu_ring *ring = fence->ring;
639
640 if (!timer_pending(&ring->fence_drv.fallback_timer))
641 amdgpu_fence_schedule_fallback(ring);
642
643 DMA_FENCE_TRACE(&fence->base, "armed on ring %i!\n", ring->idx);
644
645 return true;
646}
647
648
649
650
651
652
653
654
655static void amdgpu_fence_free(struct rcu_head *rcu)
656{
657 struct dma_fence *f = container_of(rcu, struct dma_fence, rcu);
658 struct amdgpu_fence *fence = to_amdgpu_fence(f);
659 kmem_cache_free(amdgpu_fence_slab, fence);
660}
661
662
663
664
665
666
667
668
669
670static void amdgpu_fence_release(struct dma_fence *f)
671{
672 call_rcu(&f->rcu, amdgpu_fence_free);
673}
674
675static const struct dma_fence_ops amdgpu_fence_ops = {
676 .get_driver_name = amdgpu_fence_get_driver_name,
677 .get_timeline_name = amdgpu_fence_get_timeline_name,
678 .enable_signaling = amdgpu_fence_enable_signaling,
679 .release = amdgpu_fence_release,
680};
681
682
683
684
685#if defined(CONFIG_DEBUG_FS)
686static int amdgpu_debugfs_fence_info(struct seq_file *m, void *data)
687{
688 struct drm_info_node *node = (struct drm_info_node *)m->private;
689 struct drm_device *dev = node->minor->dev;
690 struct amdgpu_device *adev = dev->dev_private;
691 int i;
692
693 for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
694 struct amdgpu_ring *ring = adev->rings[i];
695 if (!ring || !ring->fence_drv.initialized)
696 continue;
697
698 amdgpu_fence_process(ring);
699
700 seq_printf(m, "--- ring %d (%s) ---\n", i, ring->name);
701 seq_printf(m, "Last signaled fence 0x%08x\n",
702 atomic_read(&ring->fence_drv.last_seq));
703 seq_printf(m, "Last emitted 0x%08x\n",
704 ring->fence_drv.sync_seq);
705
706 if (ring->funcs->type == AMDGPU_RING_TYPE_GFX ||
707 ring->funcs->type == AMDGPU_RING_TYPE_SDMA) {
708 seq_printf(m, "Last signaled trailing fence 0x%08x\n",
709 le32_to_cpu(*ring->trail_fence_cpu_addr));
710 seq_printf(m, "Last emitted 0x%08x\n",
711 ring->trail_seq);
712 }
713
714 if (ring->funcs->type != AMDGPU_RING_TYPE_GFX)
715 continue;
716
717
718 seq_printf(m, "Last preempted 0x%08x\n",
719 le32_to_cpu(*(ring->fence_drv.cpu_addr + 2)));
720
721 seq_printf(m, "Last reset 0x%08x\n",
722 le32_to_cpu(*(ring->fence_drv.cpu_addr + 4)));
723
724 seq_printf(m, "Last both 0x%08x\n",
725 le32_to_cpu(*(ring->fence_drv.cpu_addr + 6)));
726 }
727 return 0;
728}
729
730
731
732
733
734
735static int amdgpu_debugfs_gpu_recover(struct seq_file *m, void *data)
736{
737 struct drm_info_node *node = (struct drm_info_node *) m->private;
738 struct drm_device *dev = node->minor->dev;
739 struct amdgpu_device *adev = dev->dev_private;
740
741 seq_printf(m, "gpu recover\n");
742 amdgpu_device_gpu_recover(adev, NULL);
743
744 return 0;
745}
746
747static const struct drm_info_list amdgpu_debugfs_fence_list[] = {
748 {"amdgpu_fence_info", &amdgpu_debugfs_fence_info, 0, NULL},
749 {"amdgpu_gpu_recover", &amdgpu_debugfs_gpu_recover, 0, NULL}
750};
751
752static const struct drm_info_list amdgpu_debugfs_fence_list_sriov[] = {
753 {"amdgpu_fence_info", &amdgpu_debugfs_fence_info, 0, NULL},
754};
755#endif
756
757int amdgpu_debugfs_fence_init(struct amdgpu_device *adev)
758{
759#if defined(CONFIG_DEBUG_FS)
760 if (amdgpu_sriov_vf(adev))
761 return amdgpu_debugfs_add_files(adev, amdgpu_debugfs_fence_list_sriov, 1);
762 return amdgpu_debugfs_add_files(adev, amdgpu_debugfs_fence_list, 2);
763#else
764 return 0;
765#endif
766}
767
768