1 /*
2 * Copyright 2009 Jerome Glisse.
3 * All Rights Reserved.
4 *
5 * Permission is hereby granted, free of charge, to any person obtaining a
6 * copy of this software and associated documentation files (the
7 * "Software"), to deal in the Software without restriction, including
8 * without limitation the rights to use, copy, modify, merge, publish,
9 * distribute, sub license, and/or sell copies of the Software, and to
10 * permit persons to whom the Software is furnished to do so, subject to
11 * the following conditions:
12 *
13 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
14 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
15 * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
16 * THE COPYRIGHT HOLDERS, AUTHORS AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM,
17 * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
18 * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
19 * USE OR OTHER DEALINGS IN THE SOFTWARE.
20 *
21 * The above copyright notice and this permission notice (including the
22 * next paragraph) shall be included in all copies or substantial portions
23 * of the Software.
24 *
25 */
26 /*
27 * Authors:
28 * Jerome Glisse <glisse@freedesktop.org>
29 * Dave Airlie
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 #include <linux/pm_runtime.h>
38
39 #include <drm/drm_debugfs.h>
40
41 #include "amdgpu.h"
42 #include "amdgpu_trace.h"
43
44 /*
45 * Fences
46 * Fences mark an event in the GPUs pipeline and are used
47 * for GPU/CPU synchronization. When the fence is written,
48 * it is expected that all buffers associated with that fence
49 * are no longer in use by the associated ring on the GPU and
50 * that the the relevant GPU caches have been flushed.
51 */
52
53 struct amdgpu_fence {
54 struct dma_fence base;
55
56 /* RB, DMA, etc. */
57 struct amdgpu_ring *ring;
58 };
59
60 static struct kmem_cache *amdgpu_fence_slab;
61
amdgpu_fence_slab_init(void)62 int amdgpu_fence_slab_init(void)
63 {
64 amdgpu_fence_slab = kmem_cache_create(
65 "amdgpu_fence", sizeof(struct amdgpu_fence), 0,
66 SLAB_HWCACHE_ALIGN, NULL);
67 if (!amdgpu_fence_slab)
68 return -ENOMEM;
69 return 0;
70 }
71
amdgpu_fence_slab_fini(void)72 void amdgpu_fence_slab_fini(void)
73 {
74 rcu_barrier();
75 kmem_cache_destroy(amdgpu_fence_slab);
76 }
77 /*
78 * Cast helper
79 */
80 static const struct dma_fence_ops amdgpu_fence_ops;
to_amdgpu_fence(struct dma_fence * f)81 static inline struct amdgpu_fence *to_amdgpu_fence(struct dma_fence *f)
82 {
83 struct amdgpu_fence *__f = container_of(f, struct amdgpu_fence, base);
84
85 if (__f->base.ops == &amdgpu_fence_ops)
86 return __f;
87
88 return NULL;
89 }
90
91 /**
92 * amdgpu_fence_write - write a fence value
93 *
94 * @ring: ring the fence is associated with
95 * @seq: sequence number to write
96 *
97 * Writes a fence value to memory (all asics).
98 */
amdgpu_fence_write(struct amdgpu_ring * ring,u32 seq)99 static void amdgpu_fence_write(struct amdgpu_ring *ring, u32 seq)
100 {
101 struct amdgpu_fence_driver *drv = &ring->fence_drv;
102
103 if (drv->cpu_addr)
104 *drv->cpu_addr = cpu_to_le32(seq);
105 }
106
107 /**
108 * amdgpu_fence_read - read a fence value
109 *
110 * @ring: ring the fence is associated with
111 *
112 * Reads a fence value from memory (all asics).
113 * Returns the value of the fence read from memory.
114 */
amdgpu_fence_read(struct amdgpu_ring * ring)115 static u32 amdgpu_fence_read(struct amdgpu_ring *ring)
116 {
117 struct amdgpu_fence_driver *drv = &ring->fence_drv;
118 u32 seq = 0;
119
120 if (drv->cpu_addr)
121 seq = le32_to_cpu(*drv->cpu_addr);
122 else
123 seq = atomic_read(&drv->last_seq);
124
125 return seq;
126 }
127
128 /**
129 * amdgpu_fence_emit - emit a fence on the requested ring
130 *
131 * @ring: ring the fence is associated with
132 * @f: resulting fence object
133 *
134 * Emits a fence command on the requested ring (all asics).
135 * Returns 0 on success, -ENOMEM on failure.
136 */
amdgpu_fence_emit(struct amdgpu_ring * ring,struct dma_fence ** f,unsigned flags)137 int amdgpu_fence_emit(struct amdgpu_ring *ring, struct dma_fence **f,
138 unsigned flags)
139 {
140 struct amdgpu_device *adev = ring->adev;
141 struct amdgpu_fence *fence;
142 struct dma_fence __rcu **ptr;
143 uint32_t seq;
144 int r;
145
146 fence = kmem_cache_alloc(amdgpu_fence_slab, GFP_KERNEL);
147 if (fence == NULL)
148 return -ENOMEM;
149
150 seq = ++ring->fence_drv.sync_seq;
151 fence->ring = ring;
152 dma_fence_init(&fence->base, &amdgpu_fence_ops,
153 &ring->fence_drv.lock,
154 adev->fence_context + ring->idx,
155 seq);
156 amdgpu_ring_emit_fence(ring, ring->fence_drv.gpu_addr,
157 seq, flags | AMDGPU_FENCE_FLAG_INT);
158 pm_runtime_get_noresume(adev_to_drm(adev)->dev);
159 ptr = &ring->fence_drv.fences[seq & ring->fence_drv.num_fences_mask];
160 if (unlikely(rcu_dereference_protected(*ptr, 1))) {
161 struct dma_fence *old;
162
163 rcu_read_lock();
164 old = dma_fence_get_rcu_safe(ptr);
165 rcu_read_unlock();
166
167 if (old) {
168 r = dma_fence_wait(old, false);
169 dma_fence_put(old);
170 if (r)
171 return r;
172 }
173 }
174
175 /* This function can't be called concurrently anyway, otherwise
176 * emitting the fence would mess up the hardware ring buffer.
177 */
178 rcu_assign_pointer(*ptr, dma_fence_get(&fence->base));
179
180 *f = &fence->base;
181
182 return 0;
183 }
184
185 /**
186 * amdgpu_fence_emit_polling - emit a fence on the requeste ring
187 *
188 * @ring: ring the fence is associated with
189 * @s: resulting sequence number
190 *
191 * Emits a fence command on the requested ring (all asics).
192 * Used For polling fence.
193 * Returns 0 on success, -ENOMEM on failure.
194 */
amdgpu_fence_emit_polling(struct amdgpu_ring * ring,uint32_t * s,uint32_t timeout)195 int amdgpu_fence_emit_polling(struct amdgpu_ring *ring, uint32_t *s,
196 uint32_t timeout)
197 {
198 uint32_t seq;
199 signed long r;
200
201 if (!s)
202 return -EINVAL;
203
204 seq = ++ring->fence_drv.sync_seq;
205 r = amdgpu_fence_wait_polling(ring,
206 seq - ring->fence_drv.num_fences_mask,
207 timeout);
208 if (r < 1)
209 return -ETIMEDOUT;
210
211 amdgpu_ring_emit_fence(ring, ring->fence_drv.gpu_addr,
212 seq, 0);
213
214 *s = seq;
215
216 return 0;
217 }
218
219 /**
220 * amdgpu_fence_schedule_fallback - schedule fallback check
221 *
222 * @ring: pointer to struct amdgpu_ring
223 *
224 * Start a timer as fallback to our interrupts.
225 */
amdgpu_fence_schedule_fallback(struct amdgpu_ring * ring)226 static void amdgpu_fence_schedule_fallback(struct amdgpu_ring *ring)
227 {
228 mod_timer(&ring->fence_drv.fallback_timer,
229 jiffies + AMDGPU_FENCE_JIFFIES_TIMEOUT);
230 }
231
232 /**
233 * amdgpu_fence_process - check for fence activity
234 *
235 * @ring: pointer to struct amdgpu_ring
236 *
237 * Checks the current fence value and calculates the last
238 * signalled fence value. Wakes the fence queue if the
239 * sequence number has increased.
240 *
241 * Returns true if fence was processed
242 */
amdgpu_fence_process(struct amdgpu_ring * ring)243 bool amdgpu_fence_process(struct amdgpu_ring *ring)
244 {
245 struct amdgpu_fence_driver *drv = &ring->fence_drv;
246 struct amdgpu_device *adev = ring->adev;
247 uint32_t seq, last_seq;
248 int r;
249
250 do {
251 last_seq = atomic_read(&ring->fence_drv.last_seq);
252 seq = amdgpu_fence_read(ring);
253
254 } while (atomic_cmpxchg(&drv->last_seq, last_seq, seq) != last_seq);
255
256 if (del_timer(&ring->fence_drv.fallback_timer) &&
257 seq != ring->fence_drv.sync_seq)
258 amdgpu_fence_schedule_fallback(ring);
259
260 if (unlikely(seq == last_seq))
261 return false;
262
263 last_seq &= drv->num_fences_mask;
264 seq &= drv->num_fences_mask;
265
266 do {
267 struct dma_fence *fence, **ptr;
268
269 ++last_seq;
270 last_seq &= drv->num_fences_mask;
271 ptr = &drv->fences[last_seq];
272
273 /* There is always exactly one thread signaling this fence slot */
274 fence = rcu_dereference_protected(*ptr, 1);
275 RCU_INIT_POINTER(*ptr, NULL);
276
277 if (!fence)
278 continue;
279
280 r = dma_fence_signal(fence);
281 if (!r)
282 DMA_FENCE_TRACE(fence, "signaled from irq context\n");
283 else
284 BUG();
285
286 dma_fence_put(fence);
287 pm_runtime_mark_last_busy(adev_to_drm(adev)->dev);
288 pm_runtime_put_autosuspend(adev_to_drm(adev)->dev);
289 } while (last_seq != seq);
290
291 return true;
292 }
293
294 /**
295 * amdgpu_fence_fallback - fallback for hardware interrupts
296 *
297 * @work: delayed work item
298 *
299 * Checks for fence activity.
300 */
amdgpu_fence_fallback(struct timer_list * t)301 static void amdgpu_fence_fallback(struct timer_list *t)
302 {
303 struct amdgpu_ring *ring = from_timer(ring, t,
304 fence_drv.fallback_timer);
305
306 if (amdgpu_fence_process(ring))
307 DRM_WARN("Fence fallback timer expired on ring %s\n", ring->name);
308 }
309
310 /**
311 * amdgpu_fence_wait_empty - wait for all fences to signal
312 *
313 * @adev: amdgpu device pointer
314 * @ring: ring index the fence is associated with
315 *
316 * Wait for all fences on the requested ring to signal (all asics).
317 * Returns 0 if the fences have passed, error for all other cases.
318 */
amdgpu_fence_wait_empty(struct amdgpu_ring * ring)319 int amdgpu_fence_wait_empty(struct amdgpu_ring *ring)
320 {
321 uint64_t seq = READ_ONCE(ring->fence_drv.sync_seq);
322 struct dma_fence *fence, **ptr;
323 int r;
324
325 if (!seq)
326 return 0;
327
328 ptr = &ring->fence_drv.fences[seq & ring->fence_drv.num_fences_mask];
329 rcu_read_lock();
330 fence = rcu_dereference(*ptr);
331 if (!fence || !dma_fence_get_rcu(fence)) {
332 rcu_read_unlock();
333 return 0;
334 }
335 rcu_read_unlock();
336
337 r = dma_fence_wait(fence, false);
338 dma_fence_put(fence);
339 return r;
340 }
341
342 /**
343 * amdgpu_fence_wait_polling - busy wait for givn sequence number
344 *
345 * @ring: ring index the fence is associated with
346 * @wait_seq: sequence number to wait
347 * @timeout: the timeout for waiting in usecs
348 *
349 * Wait for all fences on the requested ring to signal (all asics).
350 * Returns left time if no timeout, 0 or minus if timeout.
351 */
amdgpu_fence_wait_polling(struct amdgpu_ring * ring,uint32_t wait_seq,signed long timeout)352 signed long amdgpu_fence_wait_polling(struct amdgpu_ring *ring,
353 uint32_t wait_seq,
354 signed long timeout)
355 {
356 uint32_t seq;
357
358 do {
359 seq = amdgpu_fence_read(ring);
360 udelay(5);
361 timeout -= 5;
362 } while ((int32_t)(wait_seq - seq) > 0 && timeout > 0);
363
364 return timeout > 0 ? timeout : 0;
365 }
366 /**
367 * amdgpu_fence_count_emitted - get the count of emitted fences
368 *
369 * @ring: ring the fence is associated with
370 *
371 * Get the number of fences emitted on the requested ring (all asics).
372 * Returns the number of emitted fences on the ring. Used by the
373 * dynpm code to ring track activity.
374 */
amdgpu_fence_count_emitted(struct amdgpu_ring * ring)375 unsigned amdgpu_fence_count_emitted(struct amdgpu_ring *ring)
376 {
377 uint64_t emitted;
378
379 /* We are not protected by ring lock when reading the last sequence
380 * but it's ok to report slightly wrong fence count here.
381 */
382 amdgpu_fence_process(ring);
383 emitted = 0x100000000ull;
384 emitted -= atomic_read(&ring->fence_drv.last_seq);
385 emitted += READ_ONCE(ring->fence_drv.sync_seq);
386 return lower_32_bits(emitted);
387 }
388
389 /**
390 * amdgpu_fence_driver_start_ring - make the fence driver
391 * ready for use on the requested ring.
392 *
393 * @ring: ring to start the fence driver on
394 * @irq_src: interrupt source to use for this ring
395 * @irq_type: interrupt type to use for this ring
396 *
397 * Make the fence driver ready for processing (all asics).
398 * Not all asics have all rings, so each asic will only
399 * start the fence driver on the rings it has.
400 * Returns 0 for success, errors for failure.
401 */
amdgpu_fence_driver_start_ring(struct amdgpu_ring * ring,struct amdgpu_irq_src * irq_src,unsigned irq_type)402 int amdgpu_fence_driver_start_ring(struct amdgpu_ring *ring,
403 struct amdgpu_irq_src *irq_src,
404 unsigned irq_type)
405 {
406 struct amdgpu_device *adev = ring->adev;
407 uint64_t index;
408
409 if (ring->funcs->type != AMDGPU_RING_TYPE_UVD) {
410 ring->fence_drv.cpu_addr = &adev->wb.wb[ring->fence_offs];
411 ring->fence_drv.gpu_addr = adev->wb.gpu_addr + (ring->fence_offs * 4);
412 } else {
413 /* put fence directly behind firmware */
414 index = ALIGN(adev->uvd.fw->size, 8);
415 ring->fence_drv.cpu_addr = adev->uvd.inst[ring->me].cpu_addr + index;
416 ring->fence_drv.gpu_addr = adev->uvd.inst[ring->me].gpu_addr + index;
417 }
418 amdgpu_fence_write(ring, atomic_read(&ring->fence_drv.last_seq));
419
420 if (irq_src)
421 amdgpu_irq_get(adev, irq_src, irq_type);
422
423 ring->fence_drv.irq_src = irq_src;
424 ring->fence_drv.irq_type = irq_type;
425 ring->fence_drv.initialized = true;
426
427 DRM_DEV_DEBUG(adev->dev, "fence driver on ring %s use gpu addr 0x%016llx\n",
428 ring->name, ring->fence_drv.gpu_addr);
429 return 0;
430 }
431
432 /**
433 * amdgpu_fence_driver_init_ring - init the fence driver
434 * for the requested ring.
435 *
436 * @ring: ring to init the fence driver on
437 * @num_hw_submission: number of entries on the hardware queue
438 *
439 * Init the fence driver for the requested ring (all asics).
440 * Helper function for amdgpu_fence_driver_init().
441 */
amdgpu_fence_driver_init_ring(struct amdgpu_ring * ring,unsigned num_hw_submission)442 int amdgpu_fence_driver_init_ring(struct amdgpu_ring *ring,
443 unsigned num_hw_submission)
444 {
445 struct amdgpu_device *adev = ring->adev;
446 long timeout;
447 int r;
448
449 if (!adev)
450 return -EINVAL;
451
452 if (!is_power_of_2(num_hw_submission))
453 return -EINVAL;
454
455 ring->fence_drv.cpu_addr = NULL;
456 ring->fence_drv.gpu_addr = 0;
457 ring->fence_drv.sync_seq = 0;
458 atomic_set(&ring->fence_drv.last_seq, 0);
459 ring->fence_drv.initialized = false;
460
461 timer_setup(&ring->fence_drv.fallback_timer, amdgpu_fence_fallback, 0);
462
463 ring->fence_drv.num_fences_mask = num_hw_submission * 2 - 1;
464 spin_lock_init(&ring->fence_drv.lock);
465 ring->fence_drv.fences = kcalloc(num_hw_submission * 2, sizeof(void *),
466 GFP_KERNEL);
467 if (!ring->fence_drv.fences)
468 return -ENOMEM;
469
470 /* No need to setup the GPU scheduler for rings that don't need it */
471 if (!ring->no_scheduler) {
472 switch (ring->funcs->type) {
473 case AMDGPU_RING_TYPE_GFX:
474 timeout = adev->gfx_timeout;
475 break;
476 case AMDGPU_RING_TYPE_COMPUTE:
477 timeout = adev->compute_timeout;
478 break;
479 case AMDGPU_RING_TYPE_SDMA:
480 timeout = adev->sdma_timeout;
481 break;
482 default:
483 timeout = adev->video_timeout;
484 break;
485 }
486
487 r = drm_sched_init(&ring->sched, &amdgpu_sched_ops,
488 num_hw_submission, amdgpu_job_hang_limit,
489 timeout, ring->name);
490 if (r) {
491 DRM_ERROR("Failed to create scheduler on ring %s.\n",
492 ring->name);
493 return r;
494 }
495 }
496
497 return 0;
498 }
499
500 /**
501 * amdgpu_fence_driver_init - init the fence driver
502 * for all possible rings.
503 *
504 * @adev: amdgpu device pointer
505 *
506 * Init the fence driver for all possible rings (all asics).
507 * Not all asics have all rings, so each asic will only
508 * start the fence driver on the rings it has using
509 * amdgpu_fence_driver_start_ring().
510 * Returns 0 for success.
511 */
amdgpu_fence_driver_init(struct amdgpu_device * adev)512 int amdgpu_fence_driver_init(struct amdgpu_device *adev)
513 {
514 return 0;
515 }
516
517 /**
518 * amdgpu_fence_driver_fini - tear down the fence driver
519 * for all possible rings.
520 *
521 * @adev: amdgpu device pointer
522 *
523 * Tear down the fence driver for all possible rings (all asics).
524 */
amdgpu_fence_driver_fini(struct amdgpu_device * adev)525 void amdgpu_fence_driver_fini(struct amdgpu_device *adev)
526 {
527 unsigned i, j;
528 int r;
529
530 for (i = 0; i < AMDGPU_MAX_RINGS; i++) {
531 struct amdgpu_ring *ring = adev->rings[i];
532
533 if (!ring || !ring->fence_drv.initialized)
534 continue;
535 if (!ring->no_scheduler)
536 drm_sched_fini(&ring->sched);
537 r = amdgpu_fence_wait_empty(ring);
538 if (r) {
539 /* no need to trigger GPU reset as we are unloading */
540 amdgpu_fence_driver_force_completion(ring);
541 }
542 if (ring->fence_drv.irq_src)
543 amdgpu_irq_put(adev, ring->fence_drv.irq_src,
544 ring->fence_drv.irq_type);
545
546 del_timer_sync(&ring->fence_drv.fallback_timer);
547 for (j = 0; j <= ring->fence_drv.num_fences_mask; ++j)
548 dma_fence_put(ring->fence_drv.fences[j]);
549 kfree(ring->fence_drv.fences);
550 ring->fence_drv.fences = NULL;
551 ring->fence_drv.initialized = false;
552 }
553 }
554
555 /**
556 * amdgpu_fence_driver_suspend - suspend the fence driver
557 * for all possible rings.
558 *
559 * @adev: amdgpu device pointer
560 *
561 * Suspend the fence driver for all possible rings (all asics).
562 */
amdgpu_fence_driver_suspend(struct amdgpu_device * adev)563 void amdgpu_fence_driver_suspend(struct amdgpu_device *adev)
564 {
565 int i, r;
566
567 for (i = 0; i < AMDGPU_MAX_RINGS; i++) {
568 struct amdgpu_ring *ring = adev->rings[i];
569 if (!ring || !ring->fence_drv.initialized)
570 continue;
571
572 /* wait for gpu to finish processing current batch */
573 r = amdgpu_fence_wait_empty(ring);
574 if (r) {
575 /* delay GPU reset to resume */
576 amdgpu_fence_driver_force_completion(ring);
577 }
578
579 /* disable the interrupt */
580 if (ring->fence_drv.irq_src)
581 amdgpu_irq_put(adev, ring->fence_drv.irq_src,
582 ring->fence_drv.irq_type);
583 }
584 }
585
586 /**
587 * amdgpu_fence_driver_resume - resume the fence driver
588 * for all possible rings.
589 *
590 * @adev: amdgpu device pointer
591 *
592 * Resume the fence driver for all possible rings (all asics).
593 * Not all asics have all rings, so each asic will only
594 * start the fence driver on the rings it has using
595 * amdgpu_fence_driver_start_ring().
596 * Returns 0 for success.
597 */
amdgpu_fence_driver_resume(struct amdgpu_device * adev)598 void amdgpu_fence_driver_resume(struct amdgpu_device *adev)
599 {
600 int i;
601
602 for (i = 0; i < AMDGPU_MAX_RINGS; i++) {
603 struct amdgpu_ring *ring = adev->rings[i];
604 if (!ring || !ring->fence_drv.initialized)
605 continue;
606
607 /* enable the interrupt */
608 if (ring->fence_drv.irq_src)
609 amdgpu_irq_get(adev, ring->fence_drv.irq_src,
610 ring->fence_drv.irq_type);
611 }
612 }
613
614 /**
615 * amdgpu_fence_driver_force_completion - force signal latest fence of ring
616 *
617 * @ring: fence of the ring to signal
618 *
619 */
amdgpu_fence_driver_force_completion(struct amdgpu_ring * ring)620 void amdgpu_fence_driver_force_completion(struct amdgpu_ring *ring)
621 {
622 amdgpu_fence_write(ring, ring->fence_drv.sync_seq);
623 amdgpu_fence_process(ring);
624 }
625
626 /*
627 * Common fence implementation
628 */
629
amdgpu_fence_get_driver_name(struct dma_fence * fence)630 static const char *amdgpu_fence_get_driver_name(struct dma_fence *fence)
631 {
632 return "amdgpu";
633 }
634
amdgpu_fence_get_timeline_name(struct dma_fence * f)635 static const char *amdgpu_fence_get_timeline_name(struct dma_fence *f)
636 {
637 struct amdgpu_fence *fence = to_amdgpu_fence(f);
638 return (const char *)fence->ring->name;
639 }
640
641 /**
642 * amdgpu_fence_enable_signaling - enable signalling on fence
643 * @fence: fence
644 *
645 * This function is called with fence_queue lock held, and adds a callback
646 * to fence_queue that checks if this fence is signaled, and if so it
647 * signals the fence and removes itself.
648 */
amdgpu_fence_enable_signaling(struct dma_fence * f)649 static bool amdgpu_fence_enable_signaling(struct dma_fence *f)
650 {
651 struct amdgpu_fence *fence = to_amdgpu_fence(f);
652 struct amdgpu_ring *ring = fence->ring;
653
654 if (!timer_pending(&ring->fence_drv.fallback_timer))
655 amdgpu_fence_schedule_fallback(ring);
656
657 DMA_FENCE_TRACE(&fence->base, "armed on ring %i!\n", ring->idx);
658
659 return true;
660 }
661
662 /**
663 * amdgpu_fence_free - free up the fence memory
664 *
665 * @rcu: RCU callback head
666 *
667 * Free up the fence memory after the RCU grace period.
668 */
amdgpu_fence_free(struct rcu_head * rcu)669 static void amdgpu_fence_free(struct rcu_head *rcu)
670 {
671 struct dma_fence *f = container_of(rcu, struct dma_fence, rcu);
672 struct amdgpu_fence *fence = to_amdgpu_fence(f);
673 kmem_cache_free(amdgpu_fence_slab, fence);
674 }
675
676 /**
677 * amdgpu_fence_release - callback that fence can be freed
678 *
679 * @fence: fence
680 *
681 * This function is called when the reference count becomes zero.
682 * It just RCU schedules freeing up the fence.
683 */
amdgpu_fence_release(struct dma_fence * f)684 static void amdgpu_fence_release(struct dma_fence *f)
685 {
686 call_rcu(&f->rcu, amdgpu_fence_free);
687 }
688
689 static const struct dma_fence_ops amdgpu_fence_ops = {
690 .get_driver_name = amdgpu_fence_get_driver_name,
691 .get_timeline_name = amdgpu_fence_get_timeline_name,
692 .enable_signaling = amdgpu_fence_enable_signaling,
693 .release = amdgpu_fence_release,
694 };
695
696 /*
697 * Fence debugfs
698 */
699 #if defined(CONFIG_DEBUG_FS)
amdgpu_debugfs_fence_info(struct seq_file * m,void * data)700 static int amdgpu_debugfs_fence_info(struct seq_file *m, void *data)
701 {
702 struct drm_info_node *node = (struct drm_info_node *)m->private;
703 struct drm_device *dev = node->minor->dev;
704 struct amdgpu_device *adev = drm_to_adev(dev);
705 int i;
706
707 for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
708 struct amdgpu_ring *ring = adev->rings[i];
709 if (!ring || !ring->fence_drv.initialized)
710 continue;
711
712 amdgpu_fence_process(ring);
713
714 seq_printf(m, "--- ring %d (%s) ---\n", i, ring->name);
715 seq_printf(m, "Last signaled fence 0x%08x\n",
716 atomic_read(&ring->fence_drv.last_seq));
717 seq_printf(m, "Last emitted 0x%08x\n",
718 ring->fence_drv.sync_seq);
719
720 if (ring->funcs->type == AMDGPU_RING_TYPE_GFX ||
721 ring->funcs->type == AMDGPU_RING_TYPE_SDMA) {
722 seq_printf(m, "Last signaled trailing fence 0x%08x\n",
723 le32_to_cpu(*ring->trail_fence_cpu_addr));
724 seq_printf(m, "Last emitted 0x%08x\n",
725 ring->trail_seq);
726 }
727
728 if (ring->funcs->type != AMDGPU_RING_TYPE_GFX)
729 continue;
730
731 /* set in CP_VMID_PREEMPT and preemption occurred */
732 seq_printf(m, "Last preempted 0x%08x\n",
733 le32_to_cpu(*(ring->fence_drv.cpu_addr + 2)));
734 /* set in CP_VMID_RESET and reset occurred */
735 seq_printf(m, "Last reset 0x%08x\n",
736 le32_to_cpu(*(ring->fence_drv.cpu_addr + 4)));
737 /* Both preemption and reset occurred */
738 seq_printf(m, "Last both 0x%08x\n",
739 le32_to_cpu(*(ring->fence_drv.cpu_addr + 6)));
740 }
741 return 0;
742 }
743
744 /**
745 * amdgpu_debugfs_gpu_recover - manually trigger a gpu reset & recover
746 *
747 * Manually trigger a gpu reset at the next fence wait.
748 */
amdgpu_debugfs_gpu_recover(struct seq_file * m,void * data)749 static int amdgpu_debugfs_gpu_recover(struct seq_file *m, void *data)
750 {
751 struct drm_info_node *node = (struct drm_info_node *) m->private;
752 struct drm_device *dev = node->minor->dev;
753 struct amdgpu_device *adev = drm_to_adev(dev);
754 int r;
755
756 r = pm_runtime_get_sync(dev->dev);
757 if (r < 0) {
758 pm_runtime_put_autosuspend(dev->dev);
759 return 0;
760 }
761
762 seq_printf(m, "gpu recover\n");
763 amdgpu_device_gpu_recover(adev, NULL);
764
765 pm_runtime_mark_last_busy(dev->dev);
766 pm_runtime_put_autosuspend(dev->dev);
767
768 return 0;
769 }
770
771 static const struct drm_info_list amdgpu_debugfs_fence_list[] = {
772 {"amdgpu_fence_info", &amdgpu_debugfs_fence_info, 0, NULL},
773 {"amdgpu_gpu_recover", &amdgpu_debugfs_gpu_recover, 0, NULL}
774 };
775
776 static const struct drm_info_list amdgpu_debugfs_fence_list_sriov[] = {
777 {"amdgpu_fence_info", &amdgpu_debugfs_fence_info, 0, NULL},
778 };
779 #endif
780
amdgpu_debugfs_fence_init(struct amdgpu_device * adev)781 int amdgpu_debugfs_fence_init(struct amdgpu_device *adev)
782 {
783 #if defined(CONFIG_DEBUG_FS)
784 if (amdgpu_sriov_vf(adev))
785 return amdgpu_debugfs_add_files(adev, amdgpu_debugfs_fence_list_sriov,
786 ARRAY_SIZE(amdgpu_debugfs_fence_list_sriov));
787 return amdgpu_debugfs_add_files(adev, amdgpu_debugfs_fence_list,
788 ARRAY_SIZE(amdgpu_debugfs_fence_list));
789 #else
790 return 0;
791 #endif
792 }
793
794