xref: /linux/drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c (revision be471fe332f7f14aa6828010b220d7e6902b91a0)
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 "amdgpu.h"
40 #include "amdgpu_trace.h"
41 
42 /*
43  * Fences
44  * Fences mark an event in the GPUs pipeline and are used
45  * for GPU/CPU synchronization.  When the fence is written,
46  * it is expected that all buffers associated with that fence
47  * are no longer in use by the associated ring on the GPU and
48  * that the the relevant GPU caches have been flushed.
49  */
50 
51 struct amdgpu_fence {
52 	struct dma_fence base;
53 
54 	/* RB, DMA, etc. */
55 	struct amdgpu_ring		*ring;
56 };
57 
58 static struct kmem_cache *amdgpu_fence_slab;
59 
60 int amdgpu_fence_slab_init(void)
61 {
62 	amdgpu_fence_slab = kmem_cache_create(
63 		"amdgpu_fence", sizeof(struct amdgpu_fence), 0,
64 		SLAB_HWCACHE_ALIGN, NULL);
65 	if (!amdgpu_fence_slab)
66 		return -ENOMEM;
67 	return 0;
68 }
69 
70 void amdgpu_fence_slab_fini(void)
71 {
72 	rcu_barrier();
73 	kmem_cache_destroy(amdgpu_fence_slab);
74 }
75 /*
76  * Cast helper
77  */
78 static const struct dma_fence_ops amdgpu_fence_ops;
79 static inline struct amdgpu_fence *to_amdgpu_fence(struct dma_fence *f)
80 {
81 	struct amdgpu_fence *__f = container_of(f, struct amdgpu_fence, base);
82 
83 	if (__f->base.ops == &amdgpu_fence_ops)
84 		return __f;
85 
86 	return NULL;
87 }
88 
89 /**
90  * amdgpu_fence_write - write a fence value
91  *
92  * @ring: ring the fence is associated with
93  * @seq: sequence number to write
94  *
95  * Writes a fence value to memory (all asics).
96  */
97 static void amdgpu_fence_write(struct amdgpu_ring *ring, u32 seq)
98 {
99 	struct amdgpu_fence_driver *drv = &ring->fence_drv;
100 
101 	if (drv->cpu_addr)
102 		*drv->cpu_addr = cpu_to_le32(seq);
103 }
104 
105 /**
106  * amdgpu_fence_read - read a fence value
107  *
108  * @ring: ring the fence is associated with
109  *
110  * Reads a fence value from memory (all asics).
111  * Returns the value of the fence read from memory.
112  */
113 static u32 amdgpu_fence_read(struct amdgpu_ring *ring)
114 {
115 	struct amdgpu_fence_driver *drv = &ring->fence_drv;
116 	u32 seq = 0;
117 
118 	if (drv->cpu_addr)
119 		seq = le32_to_cpu(*drv->cpu_addr);
120 	else
121 		seq = atomic_read(&drv->last_seq);
122 
123 	return seq;
124 }
125 
126 /**
127  * amdgpu_fence_emit - emit a fence on the requested ring
128  *
129  * @ring: ring the fence is associated with
130  * @f: resulting fence object
131  * @flags: flags to pass into the subordinate .emit_fence() call
132  *
133  * Emits a fence command on the requested ring (all asics).
134  * Returns 0 on success, -ENOMEM on failure.
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 	pm_runtime_get_noresume(adev_to_drm(adev)->dev);
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 	/* This function can't be called concurrently anyway, otherwise
175 	 * emitting the fence would mess up the hardware ring buffer.
176 	 */
177 	rcu_assign_pointer(*ptr, dma_fence_get(&fence->base));
178 
179 	*f = &fence->base;
180 
181 	return 0;
182 }
183 
184 /**
185  * amdgpu_fence_emit_polling - emit a fence on the requeste ring
186  *
187  * @ring: ring the fence is associated with
188  * @s: resulting sequence number
189  * @timeout: the timeout for waiting in usecs
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  */
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  */
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  */
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  * @t: timer context used to obtain the pointer to ring structure
298  *
299  * Checks for fence activity.
300  */
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  * @ring: ring index the fence is associated with
314  *
315  * Wait for all fences on the requested ring to signal (all asics).
316  * Returns 0 if the fences have passed, error for all other cases.
317  */
318 int amdgpu_fence_wait_empty(struct amdgpu_ring *ring)
319 {
320 	uint64_t seq = READ_ONCE(ring->fence_drv.sync_seq);
321 	struct dma_fence *fence, **ptr;
322 	int r;
323 
324 	if (!seq)
325 		return 0;
326 
327 	ptr = &ring->fence_drv.fences[seq & ring->fence_drv.num_fences_mask];
328 	rcu_read_lock();
329 	fence = rcu_dereference(*ptr);
330 	if (!fence || !dma_fence_get_rcu(fence)) {
331 		rcu_read_unlock();
332 		return 0;
333 	}
334 	rcu_read_unlock();
335 
336 	r = dma_fence_wait(fence, false);
337 	dma_fence_put(fence);
338 	return r;
339 }
340 
341 /**
342  * amdgpu_fence_wait_polling - busy wait for givn sequence number
343  *
344  * @ring: ring index the fence is associated with
345  * @wait_seq: sequence number to wait
346  * @timeout: the timeout for waiting in usecs
347  *
348  * Wait for all fences on the requested ring to signal (all asics).
349  * Returns left time if no timeout, 0 or minus if timeout.
350  */
351 signed long amdgpu_fence_wait_polling(struct amdgpu_ring *ring,
352 				      uint32_t wait_seq,
353 				      signed long timeout)
354 {
355 	uint32_t seq;
356 
357 	do {
358 		seq = amdgpu_fence_read(ring);
359 		udelay(5);
360 		timeout -= 5;
361 	} while ((int32_t)(wait_seq - seq) > 0 && timeout > 0);
362 
363 	return timeout > 0 ? timeout : 0;
364 }
365 /**
366  * amdgpu_fence_count_emitted - get the count of emitted fences
367  *
368  * @ring: ring the fence is associated with
369  *
370  * Get the number of fences emitted on the requested ring (all asics).
371  * Returns the number of emitted fences on the ring.  Used by the
372  * dynpm code to ring track activity.
373  */
374 unsigned amdgpu_fence_count_emitted(struct amdgpu_ring *ring)
375 {
376 	uint64_t emitted;
377 
378 	/* We are not protected by ring lock when reading the last sequence
379 	 * but it's ok to report slightly wrong fence count here.
380 	 */
381 	amdgpu_fence_process(ring);
382 	emitted = 0x100000000ull;
383 	emitted -= atomic_read(&ring->fence_drv.last_seq);
384 	emitted += READ_ONCE(ring->fence_drv.sync_seq);
385 	return lower_32_bits(emitted);
386 }
387 
388 /**
389  * amdgpu_fence_driver_start_ring - make the fence driver
390  * ready for use on the requested ring.
391  *
392  * @ring: ring to start the fence driver on
393  * @irq_src: interrupt source to use for this ring
394  * @irq_type: interrupt type to use for this ring
395  *
396  * Make the fence driver ready for processing (all asics).
397  * Not all asics have all rings, so each asic will only
398  * start the fence driver on the rings it has.
399  * Returns 0 for success, errors for failure.
400  */
401 int amdgpu_fence_driver_start_ring(struct amdgpu_ring *ring,
402 				   struct amdgpu_irq_src *irq_src,
403 				   unsigned irq_type)
404 {
405 	struct amdgpu_device *adev = ring->adev;
406 	uint64_t index;
407 
408 	if (ring->funcs->type != AMDGPU_RING_TYPE_UVD) {
409 		ring->fence_drv.cpu_addr = &adev->wb.wb[ring->fence_offs];
410 		ring->fence_drv.gpu_addr = adev->wb.gpu_addr + (ring->fence_offs * 4);
411 	} else {
412 		/* put fence directly behind firmware */
413 		index = ALIGN(adev->uvd.fw->size, 8);
414 		ring->fence_drv.cpu_addr = adev->uvd.inst[ring->me].cpu_addr + index;
415 		ring->fence_drv.gpu_addr = adev->uvd.inst[ring->me].gpu_addr + index;
416 	}
417 	amdgpu_fence_write(ring, atomic_read(&ring->fence_drv.last_seq));
418 
419 	if (irq_src)
420 		amdgpu_irq_get(adev, irq_src, irq_type);
421 
422 	ring->fence_drv.irq_src = irq_src;
423 	ring->fence_drv.irq_type = irq_type;
424 	ring->fence_drv.initialized = true;
425 
426 	DRM_DEV_DEBUG(adev->dev, "fence driver on ring %s use gpu addr 0x%016llx\n",
427 		      ring->name, ring->fence_drv.gpu_addr);
428 	return 0;
429 }
430 
431 /**
432  * amdgpu_fence_driver_init_ring - init the fence driver
433  * for the requested ring.
434  *
435  * @ring: ring to init the fence driver on
436  * @num_hw_submission: number of entries on the hardware queue
437  *
438  * Init the fence driver for the requested ring (all asics).
439  * Helper function for amdgpu_fence_driver_init().
440  */
441 int amdgpu_fence_driver_init_ring(struct amdgpu_ring *ring,
442 				  unsigned num_hw_submission,
443 				  atomic_t *sched_score)
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 		return 0;
473 
474 	switch (ring->funcs->type) {
475 	case AMDGPU_RING_TYPE_GFX:
476 		timeout = adev->gfx_timeout;
477 		break;
478 	case AMDGPU_RING_TYPE_COMPUTE:
479 		timeout = adev->compute_timeout;
480 		break;
481 	case AMDGPU_RING_TYPE_SDMA:
482 		timeout = adev->sdma_timeout;
483 		break;
484 	default:
485 		timeout = adev->video_timeout;
486 		break;
487 	}
488 
489 	r = drm_sched_init(&ring->sched, &amdgpu_sched_ops,
490 			   num_hw_submission, amdgpu_job_hang_limit,
491 			   timeout, sched_score, ring->name);
492 	if (r) {
493 		DRM_ERROR("Failed to create scheduler on ring %s.\n",
494 			  ring->name);
495 		return r;
496 	}
497 
498 	return 0;
499 }
500 
501 /**
502  * amdgpu_fence_driver_init - init the fence driver
503  * for all possible rings.
504  *
505  * @adev: amdgpu device pointer
506  *
507  * Init the fence driver for all possible rings (all asics).
508  * Not all asics have all rings, so each asic will only
509  * start the fence driver on the rings it has using
510  * amdgpu_fence_driver_start_ring().
511  * Returns 0 for success.
512  */
513 int amdgpu_fence_driver_init(struct amdgpu_device *adev)
514 {
515 	return 0;
516 }
517 
518 /**
519  * amdgpu_fence_driver_fini - tear down the fence driver
520  * for all possible rings.
521  *
522  * @adev: amdgpu device pointer
523  *
524  * Tear down the fence driver for all possible rings (all asics).
525  */
526 void amdgpu_fence_driver_fini(struct amdgpu_device *adev)
527 {
528 	unsigned i, j;
529 	int r;
530 
531 	for (i = 0; i < AMDGPU_MAX_RINGS; i++) {
532 		struct amdgpu_ring *ring = adev->rings[i];
533 
534 		if (!ring || !ring->fence_drv.initialized)
535 			continue;
536 		if (!ring->no_scheduler)
537 			drm_sched_fini(&ring->sched);
538 		r = amdgpu_fence_wait_empty(ring);
539 		if (r) {
540 			/* no need to trigger GPU reset as we are unloading */
541 			amdgpu_fence_driver_force_completion(ring);
542 		}
543 		if (ring->fence_drv.irq_src)
544 			amdgpu_irq_put(adev, ring->fence_drv.irq_src,
545 				       ring->fence_drv.irq_type);
546 
547 		del_timer_sync(&ring->fence_drv.fallback_timer);
548 		for (j = 0; j <= ring->fence_drv.num_fences_mask; ++j)
549 			dma_fence_put(ring->fence_drv.fences[j]);
550 		kfree(ring->fence_drv.fences);
551 		ring->fence_drv.fences = NULL;
552 		ring->fence_drv.initialized = false;
553 	}
554 }
555 
556 /**
557  * amdgpu_fence_driver_suspend - suspend the fence driver
558  * for all possible rings.
559  *
560  * @adev: amdgpu device pointer
561  *
562  * Suspend the fence driver for all possible rings (all asics).
563  */
564 void amdgpu_fence_driver_suspend(struct amdgpu_device *adev)
565 {
566 	int i, r;
567 
568 	for (i = 0; i < AMDGPU_MAX_RINGS; i++) {
569 		struct amdgpu_ring *ring = adev->rings[i];
570 		if (!ring || !ring->fence_drv.initialized)
571 			continue;
572 
573 		/* wait for gpu to finish processing current batch */
574 		r = amdgpu_fence_wait_empty(ring);
575 		if (r) {
576 			/* delay GPU reset to resume */
577 			amdgpu_fence_driver_force_completion(ring);
578 		}
579 
580 		/* disable the interrupt */
581 		if (ring->fence_drv.irq_src)
582 			amdgpu_irq_put(adev, ring->fence_drv.irq_src,
583 				       ring->fence_drv.irq_type);
584 	}
585 }
586 
587 /**
588  * amdgpu_fence_driver_resume - resume the fence driver
589  * for all possible rings.
590  *
591  * @adev: amdgpu device pointer
592  *
593  * Resume the fence driver for all possible rings (all asics).
594  * Not all asics have all rings, so each asic will only
595  * start the fence driver on the rings it has using
596  * amdgpu_fence_driver_start_ring().
597  * Returns 0 for success.
598  */
599 void amdgpu_fence_driver_resume(struct amdgpu_device *adev)
600 {
601 	int i;
602 
603 	for (i = 0; i < AMDGPU_MAX_RINGS; i++) {
604 		struct amdgpu_ring *ring = adev->rings[i];
605 		if (!ring || !ring->fence_drv.initialized)
606 			continue;
607 
608 		/* enable the interrupt */
609 		if (ring->fence_drv.irq_src)
610 			amdgpu_irq_get(adev, ring->fence_drv.irq_src,
611 				       ring->fence_drv.irq_type);
612 	}
613 }
614 
615 /**
616  * amdgpu_fence_driver_force_completion - force signal latest fence of ring
617  *
618  * @ring: fence of the ring to signal
619  *
620  */
621 void amdgpu_fence_driver_force_completion(struct amdgpu_ring *ring)
622 {
623 	amdgpu_fence_write(ring, ring->fence_drv.sync_seq);
624 	amdgpu_fence_process(ring);
625 }
626 
627 /*
628  * Common fence implementation
629  */
630 
631 static const char *amdgpu_fence_get_driver_name(struct dma_fence *fence)
632 {
633 	return "amdgpu";
634 }
635 
636 static const char *amdgpu_fence_get_timeline_name(struct dma_fence *f)
637 {
638 	struct amdgpu_fence *fence = to_amdgpu_fence(f);
639 	return (const char *)fence->ring->name;
640 }
641 
642 /**
643  * amdgpu_fence_enable_signaling - enable signalling on fence
644  * @f: fence
645  *
646  * This function is called with fence_queue lock held, and adds a callback
647  * to fence_queue that checks if this fence is signaled, and if so it
648  * signals the fence and removes itself.
649  */
650 static bool amdgpu_fence_enable_signaling(struct dma_fence *f)
651 {
652 	struct amdgpu_fence *fence = to_amdgpu_fence(f);
653 	struct amdgpu_ring *ring = fence->ring;
654 
655 	if (!timer_pending(&ring->fence_drv.fallback_timer))
656 		amdgpu_fence_schedule_fallback(ring);
657 
658 	DMA_FENCE_TRACE(&fence->base, "armed on ring %i!\n", ring->idx);
659 
660 	return true;
661 }
662 
663 /**
664  * amdgpu_fence_free - free up the fence memory
665  *
666  * @rcu: RCU callback head
667  *
668  * Free up the fence memory after the RCU grace period.
669  */
670 static void amdgpu_fence_free(struct rcu_head *rcu)
671 {
672 	struct dma_fence *f = container_of(rcu, struct dma_fence, rcu);
673 	struct amdgpu_fence *fence = to_amdgpu_fence(f);
674 	kmem_cache_free(amdgpu_fence_slab, fence);
675 }
676 
677 /**
678  * amdgpu_fence_release - callback that fence can be freed
679  *
680  * @f: fence
681  *
682  * This function is called when the reference count becomes zero.
683  * It just RCU schedules freeing up the fence.
684  */
685 static void amdgpu_fence_release(struct dma_fence *f)
686 {
687 	call_rcu(&f->rcu, amdgpu_fence_free);
688 }
689 
690 static const struct dma_fence_ops amdgpu_fence_ops = {
691 	.get_driver_name = amdgpu_fence_get_driver_name,
692 	.get_timeline_name = amdgpu_fence_get_timeline_name,
693 	.enable_signaling = amdgpu_fence_enable_signaling,
694 	.release = amdgpu_fence_release,
695 };
696 
697 /*
698  * Fence debugfs
699  */
700 #if defined(CONFIG_DEBUG_FS)
701 static int amdgpu_debugfs_fence_info_show(struct seq_file *m, void *unused)
702 {
703 	struct amdgpu_device *adev = (struct amdgpu_device *)m->private;
704 	int i;
705 
706 	for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
707 		struct amdgpu_ring *ring = adev->rings[i];
708 		if (!ring || !ring->fence_drv.initialized)
709 			continue;
710 
711 		amdgpu_fence_process(ring);
712 
713 		seq_printf(m, "--- ring %d (%s) ---\n", i, ring->name);
714 		seq_printf(m, "Last signaled fence          0x%08x\n",
715 			   atomic_read(&ring->fence_drv.last_seq));
716 		seq_printf(m, "Last emitted                 0x%08x\n",
717 			   ring->fence_drv.sync_seq);
718 
719 		if (ring->funcs->type == AMDGPU_RING_TYPE_GFX ||
720 		    ring->funcs->type == AMDGPU_RING_TYPE_SDMA) {
721 			seq_printf(m, "Last signaled trailing fence 0x%08x\n",
722 				   le32_to_cpu(*ring->trail_fence_cpu_addr));
723 			seq_printf(m, "Last emitted                 0x%08x\n",
724 				   ring->trail_seq);
725 		}
726 
727 		if (ring->funcs->type != AMDGPU_RING_TYPE_GFX)
728 			continue;
729 
730 		/* set in CP_VMID_PREEMPT and preemption occurred */
731 		seq_printf(m, "Last preempted               0x%08x\n",
732 			   le32_to_cpu(*(ring->fence_drv.cpu_addr + 2)));
733 		/* set in CP_VMID_RESET and reset occurred */
734 		seq_printf(m, "Last reset                   0x%08x\n",
735 			   le32_to_cpu(*(ring->fence_drv.cpu_addr + 4)));
736 		/* Both preemption and reset occurred */
737 		seq_printf(m, "Last both                    0x%08x\n",
738 			   le32_to_cpu(*(ring->fence_drv.cpu_addr + 6)));
739 	}
740 	return 0;
741 }
742 
743 /*
744  * amdgpu_debugfs_gpu_recover - manually trigger a gpu reset & recover
745  *
746  * Manually trigger a gpu reset at the next fence wait.
747  */
748 static int gpu_recover_get(void *data, u64 *val)
749 {
750 	struct amdgpu_device *adev = (struct amdgpu_device *)data;
751 	struct drm_device *dev = adev_to_drm(adev);
752 	int r;
753 
754 	r = pm_runtime_get_sync(dev->dev);
755 	if (r < 0) {
756 		pm_runtime_put_autosuspend(dev->dev);
757 		return 0;
758 	}
759 
760 	*val = amdgpu_device_gpu_recover(adev, NULL);
761 
762 	pm_runtime_mark_last_busy(dev->dev);
763 	pm_runtime_put_autosuspend(dev->dev);
764 
765 	return 0;
766 }
767 
768 DEFINE_SHOW_ATTRIBUTE(amdgpu_debugfs_fence_info);
769 DEFINE_DEBUGFS_ATTRIBUTE(amdgpu_debugfs_gpu_recover_fops, gpu_recover_get, NULL,
770 			 "%lld\n");
771 
772 #endif
773 
774 void amdgpu_debugfs_fence_init(struct amdgpu_device *adev)
775 {
776 #if defined(CONFIG_DEBUG_FS)
777 	struct drm_minor *minor = adev_to_drm(adev)->primary;
778 	struct dentry *root = minor->debugfs_root;
779 
780 	debugfs_create_file("amdgpu_fence_info", 0444, root, adev,
781 			    &amdgpu_debugfs_fence_info_fops);
782 
783 	if (!amdgpu_sriov_vf(adev))
784 		debugfs_create_file("amdgpu_gpu_recover", 0444, root, adev,
785 				    &amdgpu_debugfs_gpu_recover_fops);
786 #endif
787 }
788 
789