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_drv.h>
40 #include "amdgpu.h"
41 #include "amdgpu_trace.h"
42 #include "amdgpu_reset.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 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 	ktime_t				start_timestamp;
59 };
60 
61 static struct kmem_cache *amdgpu_fence_slab;
62 
63 int amdgpu_fence_slab_init(void)
64 {
65 	amdgpu_fence_slab = kmem_cache_create(
66 		"amdgpu_fence", sizeof(struct amdgpu_fence), 0,
67 		SLAB_HWCACHE_ALIGN, NULL);
68 	if (!amdgpu_fence_slab)
69 		return -ENOMEM;
70 	return 0;
71 }
72 
73 void amdgpu_fence_slab_fini(void)
74 {
75 	rcu_barrier();
76 	kmem_cache_destroy(amdgpu_fence_slab);
77 }
78 /*
79  * Cast helper
80  */
81 static const struct dma_fence_ops amdgpu_fence_ops;
82 static const struct dma_fence_ops amdgpu_job_fence_ops;
83 static inline struct amdgpu_fence *to_amdgpu_fence(struct dma_fence *f)
84 {
85 	struct amdgpu_fence *__f = container_of(f, struct amdgpu_fence, base);
86 
87 	if (__f->base.ops == &amdgpu_fence_ops ||
88 	    __f->base.ops == &amdgpu_job_fence_ops)
89 		return __f;
90 
91 	return NULL;
92 }
93 
94 /**
95  * amdgpu_fence_write - write a fence value
96  *
97  * @ring: ring the fence is associated with
98  * @seq: sequence number to write
99  *
100  * Writes a fence value to memory (all asics).
101  */
102 static void amdgpu_fence_write(struct amdgpu_ring *ring, u32 seq)
103 {
104 	struct amdgpu_fence_driver *drv = &ring->fence_drv;
105 
106 	if (drv->cpu_addr)
107 		*drv->cpu_addr = cpu_to_le32(seq);
108 }
109 
110 /**
111  * amdgpu_fence_read - read a fence value
112  *
113  * @ring: ring the fence is associated with
114  *
115  * Reads a fence value from memory (all asics).
116  * Returns the value of the fence read from memory.
117  */
118 static u32 amdgpu_fence_read(struct amdgpu_ring *ring)
119 {
120 	struct amdgpu_fence_driver *drv = &ring->fence_drv;
121 	u32 seq = 0;
122 
123 	if (drv->cpu_addr)
124 		seq = le32_to_cpu(*drv->cpu_addr);
125 	else
126 		seq = atomic_read(&drv->last_seq);
127 
128 	return seq;
129 }
130 
131 /**
132  * amdgpu_fence_emit - emit a fence on the requested ring
133  *
134  * @ring: ring the fence is associated with
135  * @f: resulting fence object
136  * @job: job the fence is embedded in
137  * @flags: flags to pass into the subordinate .emit_fence() call
138  *
139  * Emits a fence command on the requested ring (all asics).
140  * Returns 0 on success, -ENOMEM on failure.
141  */
142 int amdgpu_fence_emit(struct amdgpu_ring *ring, struct dma_fence **f, struct amdgpu_job *job,
143 		      unsigned flags)
144 {
145 	struct amdgpu_device *adev = ring->adev;
146 	struct dma_fence *fence;
147 	struct amdgpu_fence *am_fence;
148 	struct dma_fence __rcu **ptr;
149 	uint32_t seq;
150 	int r;
151 
152 	if (job == NULL) {
153 		/* create a sperate hw fence */
154 		am_fence = kmem_cache_alloc(amdgpu_fence_slab, GFP_ATOMIC);
155 		if (am_fence == NULL)
156 			return -ENOMEM;
157 		fence = &am_fence->base;
158 		am_fence->ring = ring;
159 	} else {
160 		/* take use of job-embedded fence */
161 		fence = &job->hw_fence;
162 	}
163 
164 	seq = ++ring->fence_drv.sync_seq;
165 	if (job && job->job_run_counter) {
166 		/* reinit seq for resubmitted jobs */
167 		fence->seqno = seq;
168 		/* TO be inline with external fence creation and other drivers */
169 		dma_fence_get(fence);
170 	} else {
171 		if (job) {
172 			dma_fence_init(fence, &amdgpu_job_fence_ops,
173 				       &ring->fence_drv.lock,
174 				       adev->fence_context + ring->idx, seq);
175 			/* Against remove in amdgpu_job_{free, free_cb} */
176 			dma_fence_get(fence);
177 		}
178 		else
179 			dma_fence_init(fence, &amdgpu_fence_ops,
180 				       &ring->fence_drv.lock,
181 				       adev->fence_context + ring->idx, seq);
182 	}
183 
184 	amdgpu_ring_emit_fence(ring, ring->fence_drv.gpu_addr,
185 			       seq, flags | AMDGPU_FENCE_FLAG_INT);
186 	pm_runtime_get_noresume(adev_to_drm(adev)->dev);
187 	ptr = &ring->fence_drv.fences[seq & ring->fence_drv.num_fences_mask];
188 	if (unlikely(rcu_dereference_protected(*ptr, 1))) {
189 		struct dma_fence *old;
190 
191 		rcu_read_lock();
192 		old = dma_fence_get_rcu_safe(ptr);
193 		rcu_read_unlock();
194 
195 		if (old) {
196 			r = dma_fence_wait(old, false);
197 			dma_fence_put(old);
198 			if (r)
199 				return r;
200 		}
201 	}
202 
203 	to_amdgpu_fence(fence)->start_timestamp = ktime_get();
204 
205 	/* This function can't be called concurrently anyway, otherwise
206 	 * emitting the fence would mess up the hardware ring buffer.
207 	 */
208 	rcu_assign_pointer(*ptr, dma_fence_get(fence));
209 
210 	*f = fence;
211 
212 	return 0;
213 }
214 
215 /**
216  * amdgpu_fence_emit_polling - emit a fence on the requeste ring
217  *
218  * @ring: ring the fence is associated with
219  * @s: resulting sequence number
220  * @timeout: the timeout for waiting in usecs
221  *
222  * Emits a fence command on the requested ring (all asics).
223  * Used For polling fence.
224  * Returns 0 on success, -ENOMEM on failure.
225  */
226 int amdgpu_fence_emit_polling(struct amdgpu_ring *ring, uint32_t *s,
227 			      uint32_t timeout)
228 {
229 	uint32_t seq;
230 	signed long r;
231 
232 	if (!s)
233 		return -EINVAL;
234 
235 	seq = ++ring->fence_drv.sync_seq;
236 	r = amdgpu_fence_wait_polling(ring,
237 				      seq - ring->fence_drv.num_fences_mask,
238 				      timeout);
239 	if (r < 1)
240 		return -ETIMEDOUT;
241 
242 	amdgpu_ring_emit_fence(ring, ring->fence_drv.gpu_addr,
243 			       seq, 0);
244 
245 	*s = seq;
246 
247 	return 0;
248 }
249 
250 /**
251  * amdgpu_fence_schedule_fallback - schedule fallback check
252  *
253  * @ring: pointer to struct amdgpu_ring
254  *
255  * Start a timer as fallback to our interrupts.
256  */
257 static void amdgpu_fence_schedule_fallback(struct amdgpu_ring *ring)
258 {
259 	mod_timer(&ring->fence_drv.fallback_timer,
260 		  jiffies + AMDGPU_FENCE_JIFFIES_TIMEOUT);
261 }
262 
263 /**
264  * amdgpu_fence_process - check for fence activity
265  *
266  * @ring: pointer to struct amdgpu_ring
267  *
268  * Checks the current fence value and calculates the last
269  * signalled fence value. Wakes the fence queue if the
270  * sequence number has increased.
271  *
272  * Returns true if fence was processed
273  */
274 bool amdgpu_fence_process(struct amdgpu_ring *ring)
275 {
276 	struct amdgpu_fence_driver *drv = &ring->fence_drv;
277 	struct amdgpu_device *adev = ring->adev;
278 	uint32_t seq, last_seq;
279 
280 	do {
281 		last_seq = atomic_read(&ring->fence_drv.last_seq);
282 		seq = amdgpu_fence_read(ring);
283 
284 	} while (atomic_cmpxchg(&drv->last_seq, last_seq, seq) != last_seq);
285 
286 	if (del_timer(&ring->fence_drv.fallback_timer) &&
287 	    seq != ring->fence_drv.sync_seq)
288 		amdgpu_fence_schedule_fallback(ring);
289 
290 	if (unlikely(seq == last_seq))
291 		return false;
292 
293 	last_seq &= drv->num_fences_mask;
294 	seq &= drv->num_fences_mask;
295 
296 	do {
297 		struct dma_fence *fence, **ptr;
298 
299 		++last_seq;
300 		last_seq &= drv->num_fences_mask;
301 		ptr = &drv->fences[last_seq];
302 
303 		/* There is always exactly one thread signaling this fence slot */
304 		fence = rcu_dereference_protected(*ptr, 1);
305 		RCU_INIT_POINTER(*ptr, NULL);
306 
307 		if (!fence)
308 			continue;
309 
310 		dma_fence_signal(fence);
311 		dma_fence_put(fence);
312 		pm_runtime_mark_last_busy(adev_to_drm(adev)->dev);
313 		pm_runtime_put_autosuspend(adev_to_drm(adev)->dev);
314 	} while (last_seq != seq);
315 
316 	return true;
317 }
318 
319 /**
320  * amdgpu_fence_fallback - fallback for hardware interrupts
321  *
322  * @t: timer context used to obtain the pointer to ring structure
323  *
324  * Checks for fence activity.
325  */
326 static void amdgpu_fence_fallback(struct timer_list *t)
327 {
328 	struct amdgpu_ring *ring = from_timer(ring, t,
329 					      fence_drv.fallback_timer);
330 
331 	if (amdgpu_fence_process(ring))
332 		DRM_WARN("Fence fallback timer expired on ring %s\n", ring->name);
333 }
334 
335 /**
336  * amdgpu_fence_wait_empty - wait for all fences to signal
337  *
338  * @ring: ring index the fence is associated with
339  *
340  * Wait for all fences on the requested ring to signal (all asics).
341  * Returns 0 if the fences have passed, error for all other cases.
342  */
343 int amdgpu_fence_wait_empty(struct amdgpu_ring *ring)
344 {
345 	uint64_t seq = READ_ONCE(ring->fence_drv.sync_seq);
346 	struct dma_fence *fence, **ptr;
347 	int r;
348 
349 	if (!seq)
350 		return 0;
351 
352 	ptr = &ring->fence_drv.fences[seq & ring->fence_drv.num_fences_mask];
353 	rcu_read_lock();
354 	fence = rcu_dereference(*ptr);
355 	if (!fence || !dma_fence_get_rcu(fence)) {
356 		rcu_read_unlock();
357 		return 0;
358 	}
359 	rcu_read_unlock();
360 
361 	r = dma_fence_wait(fence, false);
362 	dma_fence_put(fence);
363 	return r;
364 }
365 
366 /**
367  * amdgpu_fence_wait_polling - busy wait for givn sequence number
368  *
369  * @ring: ring index the fence is associated with
370  * @wait_seq: sequence number to wait
371  * @timeout: the timeout for waiting in usecs
372  *
373  * Wait for all fences on the requested ring to signal (all asics).
374  * Returns left time if no timeout, 0 or minus if timeout.
375  */
376 signed long amdgpu_fence_wait_polling(struct amdgpu_ring *ring,
377 				      uint32_t wait_seq,
378 				      signed long timeout)
379 {
380 	uint32_t seq;
381 
382 	do {
383 		seq = amdgpu_fence_read(ring);
384 		udelay(5);
385 		timeout -= 5;
386 	} while ((int32_t)(wait_seq - seq) > 0 && timeout > 0);
387 
388 	return timeout > 0 ? timeout : 0;
389 }
390 /**
391  * amdgpu_fence_count_emitted - get the count of emitted fences
392  *
393  * @ring: ring the fence is associated with
394  *
395  * Get the number of fences emitted on the requested ring (all asics).
396  * Returns the number of emitted fences on the ring.  Used by the
397  * dynpm code to ring track activity.
398  */
399 unsigned amdgpu_fence_count_emitted(struct amdgpu_ring *ring)
400 {
401 	uint64_t emitted;
402 
403 	/* We are not protected by ring lock when reading the last sequence
404 	 * but it's ok to report slightly wrong fence count here.
405 	 */
406 	emitted = 0x100000000ull;
407 	emitted -= atomic_read(&ring->fence_drv.last_seq);
408 	emitted += READ_ONCE(ring->fence_drv.sync_seq);
409 	return lower_32_bits(emitted);
410 }
411 
412 /**
413  * amdgpu_fence_last_unsignaled_time_us - the time fence emitted until now
414  * @ring: ring the fence is associated with
415  *
416  * Find the earliest fence unsignaled until now, calculate the time delta
417  * between the time fence emitted and now.
418  */
419 u64 amdgpu_fence_last_unsignaled_time_us(struct amdgpu_ring *ring)
420 {
421 	struct amdgpu_fence_driver *drv = &ring->fence_drv;
422 	struct dma_fence *fence;
423 	uint32_t last_seq, sync_seq;
424 
425 	last_seq = atomic_read(&ring->fence_drv.last_seq);
426 	sync_seq = READ_ONCE(ring->fence_drv.sync_seq);
427 	if (last_seq == sync_seq)
428 		return 0;
429 
430 	++last_seq;
431 	last_seq &= drv->num_fences_mask;
432 	fence = drv->fences[last_seq];
433 	if (!fence)
434 		return 0;
435 
436 	return ktime_us_delta(ktime_get(),
437 		to_amdgpu_fence(fence)->start_timestamp);
438 }
439 
440 /**
441  * amdgpu_fence_update_start_timestamp - update the timestamp of the fence
442  * @ring: ring the fence is associated with
443  * @seq: the fence seq number to update.
444  * @timestamp: the start timestamp to update.
445  *
446  * The function called at the time the fence and related ib is about to
447  * resubmit to gpu in MCBP scenario. Thus we do not consider race condition
448  * with amdgpu_fence_process to modify the same fence.
449  */
450 void amdgpu_fence_update_start_timestamp(struct amdgpu_ring *ring, uint32_t seq, ktime_t timestamp)
451 {
452 	struct amdgpu_fence_driver *drv = &ring->fence_drv;
453 	struct dma_fence *fence;
454 
455 	seq &= drv->num_fences_mask;
456 	fence = drv->fences[seq];
457 	if (!fence)
458 		return;
459 
460 	to_amdgpu_fence(fence)->start_timestamp = timestamp;
461 }
462 
463 /**
464  * amdgpu_fence_driver_start_ring - make the fence driver
465  * ready for use on the requested ring.
466  *
467  * @ring: ring to start the fence driver on
468  * @irq_src: interrupt source to use for this ring
469  * @irq_type: interrupt type to use for this ring
470  *
471  * Make the fence driver ready for processing (all asics).
472  * Not all asics have all rings, so each asic will only
473  * start the fence driver on the rings it has.
474  * Returns 0 for success, errors for failure.
475  */
476 int amdgpu_fence_driver_start_ring(struct amdgpu_ring *ring,
477 				   struct amdgpu_irq_src *irq_src,
478 				   unsigned irq_type)
479 {
480 	struct amdgpu_device *adev = ring->adev;
481 	uint64_t index;
482 
483 	if (ring->funcs->type != AMDGPU_RING_TYPE_UVD) {
484 		ring->fence_drv.cpu_addr = ring->fence_cpu_addr;
485 		ring->fence_drv.gpu_addr = ring->fence_gpu_addr;
486 	} else {
487 		/* put fence directly behind firmware */
488 		index = ALIGN(adev->uvd.fw->size, 8);
489 		ring->fence_drv.cpu_addr = adev->uvd.inst[ring->me].cpu_addr + index;
490 		ring->fence_drv.gpu_addr = adev->uvd.inst[ring->me].gpu_addr + index;
491 	}
492 	amdgpu_fence_write(ring, atomic_read(&ring->fence_drv.last_seq));
493 
494 	ring->fence_drv.irq_src = irq_src;
495 	ring->fence_drv.irq_type = irq_type;
496 	ring->fence_drv.initialized = true;
497 
498 	DRM_DEV_DEBUG(adev->dev, "fence driver on ring %s use gpu addr 0x%016llx\n",
499 		      ring->name, ring->fence_drv.gpu_addr);
500 	return 0;
501 }
502 
503 /**
504  * amdgpu_fence_driver_init_ring - init the fence driver
505  * for the requested ring.
506  *
507  * @ring: ring to init the fence driver on
508  *
509  * Init the fence driver for the requested ring (all asics).
510  * Helper function for amdgpu_fence_driver_init().
511  */
512 int amdgpu_fence_driver_init_ring(struct amdgpu_ring *ring)
513 {
514 	struct amdgpu_device *adev = ring->adev;
515 
516 	if (!adev)
517 		return -EINVAL;
518 
519 	if (!is_power_of_2(ring->num_hw_submission))
520 		return -EINVAL;
521 
522 	ring->fence_drv.cpu_addr = NULL;
523 	ring->fence_drv.gpu_addr = 0;
524 	ring->fence_drv.sync_seq = 0;
525 	atomic_set(&ring->fence_drv.last_seq, 0);
526 	ring->fence_drv.initialized = false;
527 
528 	timer_setup(&ring->fence_drv.fallback_timer, amdgpu_fence_fallback, 0);
529 
530 	ring->fence_drv.num_fences_mask = ring->num_hw_submission * 2 - 1;
531 	spin_lock_init(&ring->fence_drv.lock);
532 	ring->fence_drv.fences = kcalloc(ring->num_hw_submission * 2, sizeof(void *),
533 					 GFP_KERNEL);
534 
535 	if (!ring->fence_drv.fences)
536 		return -ENOMEM;
537 
538 	return 0;
539 }
540 
541 /**
542  * amdgpu_fence_driver_sw_init - init the fence driver
543  * for all possible rings.
544  *
545  * @adev: amdgpu device pointer
546  *
547  * Init the fence driver for all possible rings (all asics).
548  * Not all asics have all rings, so each asic will only
549  * start the fence driver on the rings it has using
550  * amdgpu_fence_driver_start_ring().
551  * Returns 0 for success.
552  */
553 int amdgpu_fence_driver_sw_init(struct amdgpu_device *adev)
554 {
555 	return 0;
556 }
557 
558 /**
559  * amdgpu_fence_driver_hw_fini - tear down the fence driver
560  * for all possible rings.
561  *
562  * @adev: amdgpu device pointer
563  *
564  * Tear down the fence driver for all possible rings (all asics).
565  */
566 void amdgpu_fence_driver_hw_fini(struct amdgpu_device *adev)
567 {
568 	int i, r;
569 
570 	for (i = 0; i < AMDGPU_MAX_RINGS; i++) {
571 		struct amdgpu_ring *ring = adev->rings[i];
572 
573 		if (!ring || !ring->fence_drv.initialized)
574 			continue;
575 
576 		/* You can't wait for HW to signal if it's gone */
577 		if (!drm_dev_is_unplugged(adev_to_drm(adev)))
578 			r = amdgpu_fence_wait_empty(ring);
579 		else
580 			r = -ENODEV;
581 		/* no need to trigger GPU reset as we are unloading */
582 		if (r)
583 			amdgpu_fence_driver_force_completion(ring);
584 
585 		if (ring->fence_drv.irq_src)
586 			amdgpu_irq_put(adev, ring->fence_drv.irq_src,
587 				       ring->fence_drv.irq_type);
588 
589 		del_timer_sync(&ring->fence_drv.fallback_timer);
590 	}
591 }
592 
593 /* Will either stop and flush handlers for amdgpu interrupt or reanble it */
594 void amdgpu_fence_driver_isr_toggle(struct amdgpu_device *adev, bool stop)
595 {
596 	int i;
597 
598 	for (i = 0; i < AMDGPU_MAX_RINGS; i++) {
599 		struct amdgpu_ring *ring = adev->rings[i];
600 
601 		if (!ring || !ring->fence_drv.initialized || !ring->fence_drv.irq_src)
602 			continue;
603 
604 		if (stop)
605 			disable_irq(adev->irq.irq);
606 		else
607 			enable_irq(adev->irq.irq);
608 	}
609 }
610 
611 void amdgpu_fence_driver_sw_fini(struct amdgpu_device *adev)
612 {
613 	unsigned int i, j;
614 
615 	for (i = 0; i < AMDGPU_MAX_RINGS; i++) {
616 		struct amdgpu_ring *ring = adev->rings[i];
617 
618 		if (!ring || !ring->fence_drv.initialized)
619 			continue;
620 
621 		/*
622 		 * Notice we check for sched.ops since there's some
623 		 * override on the meaning of sched.ready by amdgpu.
624 		 * The natural check would be sched.ready, which is
625 		 * set as drm_sched_init() finishes...
626 		 */
627 		if (ring->sched.ops)
628 			drm_sched_fini(&ring->sched);
629 
630 		for (j = 0; j <= ring->fence_drv.num_fences_mask; ++j)
631 			dma_fence_put(ring->fence_drv.fences[j]);
632 		kfree(ring->fence_drv.fences);
633 		ring->fence_drv.fences = NULL;
634 		ring->fence_drv.initialized = false;
635 	}
636 }
637 
638 /**
639  * amdgpu_fence_driver_hw_init - enable the fence driver
640  * for all possible rings.
641  *
642  * @adev: amdgpu device pointer
643  *
644  * Enable the fence driver for all possible rings (all asics).
645  * Not all asics have all rings, so each asic will only
646  * start the fence driver on the rings it has using
647  * amdgpu_fence_driver_start_ring().
648  * Returns 0 for success.
649  */
650 void amdgpu_fence_driver_hw_init(struct amdgpu_device *adev)
651 {
652 	int i;
653 
654 	for (i = 0; i < AMDGPU_MAX_RINGS; i++) {
655 		struct amdgpu_ring *ring = adev->rings[i];
656 		if (!ring || !ring->fence_drv.initialized)
657 			continue;
658 
659 		/* enable the interrupt */
660 		if (ring->fence_drv.irq_src)
661 			amdgpu_irq_get(adev, ring->fence_drv.irq_src,
662 				       ring->fence_drv.irq_type);
663 	}
664 }
665 
666 /**
667  * amdgpu_fence_driver_clear_job_fences - clear job embedded fences of ring
668  *
669  * @ring: fence of the ring to be cleared
670  *
671  */
672 void amdgpu_fence_driver_clear_job_fences(struct amdgpu_ring *ring)
673 {
674 	int i;
675 	struct dma_fence *old, **ptr;
676 
677 	for (i = 0; i <= ring->fence_drv.num_fences_mask; i++) {
678 		ptr = &ring->fence_drv.fences[i];
679 		old = rcu_dereference_protected(*ptr, 1);
680 		if (old && old->ops == &amdgpu_job_fence_ops) {
681 			struct amdgpu_job *job;
682 
683 			/* For non-scheduler bad job, i.e. failed ib test, we need to signal
684 			 * it right here or we won't be able to track them in fence_drv
685 			 * and they will remain unsignaled during sa_bo free.
686 			 */
687 			job = container_of(old, struct amdgpu_job, hw_fence);
688 			if (!job->base.s_fence && !dma_fence_is_signaled(old))
689 				dma_fence_signal(old);
690 			RCU_INIT_POINTER(*ptr, NULL);
691 			dma_fence_put(old);
692 		}
693 	}
694 }
695 
696 /**
697  * amdgpu_fence_driver_force_completion - force signal latest fence of ring
698  *
699  * @ring: fence of the ring to signal
700  *
701  */
702 void amdgpu_fence_driver_force_completion(struct amdgpu_ring *ring)
703 {
704 	amdgpu_fence_write(ring, ring->fence_drv.sync_seq);
705 	amdgpu_fence_process(ring);
706 }
707 
708 /*
709  * Common fence implementation
710  */
711 
712 static const char *amdgpu_fence_get_driver_name(struct dma_fence *fence)
713 {
714 	return "amdgpu";
715 }
716 
717 static const char *amdgpu_fence_get_timeline_name(struct dma_fence *f)
718 {
719 	return (const char *)to_amdgpu_fence(f)->ring->name;
720 }
721 
722 static const char *amdgpu_job_fence_get_timeline_name(struct dma_fence *f)
723 {
724 	struct amdgpu_job *job = container_of(f, struct amdgpu_job, hw_fence);
725 
726 	return (const char *)to_amdgpu_ring(job->base.sched)->name;
727 }
728 
729 /**
730  * amdgpu_fence_enable_signaling - enable signalling on fence
731  * @f: fence
732  *
733  * This function is called with fence_queue lock held, and adds a callback
734  * to fence_queue that checks if this fence is signaled, and if so it
735  * signals the fence and removes itself.
736  */
737 static bool amdgpu_fence_enable_signaling(struct dma_fence *f)
738 {
739 	if (!timer_pending(&to_amdgpu_fence(f)->ring->fence_drv.fallback_timer))
740 		amdgpu_fence_schedule_fallback(to_amdgpu_fence(f)->ring);
741 
742 	return true;
743 }
744 
745 /**
746  * amdgpu_job_fence_enable_signaling - enable signalling on job fence
747  * @f: fence
748  *
749  * This is the simliar function with amdgpu_fence_enable_signaling above, it
750  * only handles the job embedded fence.
751  */
752 static bool amdgpu_job_fence_enable_signaling(struct dma_fence *f)
753 {
754 	struct amdgpu_job *job = container_of(f, struct amdgpu_job, hw_fence);
755 
756 	if (!timer_pending(&to_amdgpu_ring(job->base.sched)->fence_drv.fallback_timer))
757 		amdgpu_fence_schedule_fallback(to_amdgpu_ring(job->base.sched));
758 
759 	return true;
760 }
761 
762 /**
763  * amdgpu_fence_free - free up the fence memory
764  *
765  * @rcu: RCU callback head
766  *
767  * Free up the fence memory after the RCU grace period.
768  */
769 static void amdgpu_fence_free(struct rcu_head *rcu)
770 {
771 	struct dma_fence *f = container_of(rcu, struct dma_fence, rcu);
772 
773 	/* free fence_slab if it's separated fence*/
774 	kmem_cache_free(amdgpu_fence_slab, to_amdgpu_fence(f));
775 }
776 
777 /**
778  * amdgpu_job_fence_free - free up the job with embedded fence
779  *
780  * @rcu: RCU callback head
781  *
782  * Free up the job with embedded fence after the RCU grace period.
783  */
784 static void amdgpu_job_fence_free(struct rcu_head *rcu)
785 {
786 	struct dma_fence *f = container_of(rcu, struct dma_fence, rcu);
787 
788 	/* free job if fence has a parent job */
789 	kfree(container_of(f, struct amdgpu_job, hw_fence));
790 }
791 
792 /**
793  * amdgpu_fence_release - callback that fence can be freed
794  *
795  * @f: fence
796  *
797  * This function is called when the reference count becomes zero.
798  * It just RCU schedules freeing up the fence.
799  */
800 static void amdgpu_fence_release(struct dma_fence *f)
801 {
802 	call_rcu(&f->rcu, amdgpu_fence_free);
803 }
804 
805 /**
806  * amdgpu_job_fence_release - callback that job embedded fence can be freed
807  *
808  * @f: fence
809  *
810  * This is the simliar function with amdgpu_fence_release above, it
811  * only handles the job embedded fence.
812  */
813 static void amdgpu_job_fence_release(struct dma_fence *f)
814 {
815 	call_rcu(&f->rcu, amdgpu_job_fence_free);
816 }
817 
818 static const struct dma_fence_ops amdgpu_fence_ops = {
819 	.get_driver_name = amdgpu_fence_get_driver_name,
820 	.get_timeline_name = amdgpu_fence_get_timeline_name,
821 	.enable_signaling = amdgpu_fence_enable_signaling,
822 	.release = amdgpu_fence_release,
823 };
824 
825 static const struct dma_fence_ops amdgpu_job_fence_ops = {
826 	.get_driver_name = amdgpu_fence_get_driver_name,
827 	.get_timeline_name = amdgpu_job_fence_get_timeline_name,
828 	.enable_signaling = amdgpu_job_fence_enable_signaling,
829 	.release = amdgpu_job_fence_release,
830 };
831 
832 /*
833  * Fence debugfs
834  */
835 #if defined(CONFIG_DEBUG_FS)
836 static int amdgpu_debugfs_fence_info_show(struct seq_file *m, void *unused)
837 {
838 	struct amdgpu_device *adev = (struct amdgpu_device *)m->private;
839 	int i;
840 
841 	for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
842 		struct amdgpu_ring *ring = adev->rings[i];
843 		if (!ring || !ring->fence_drv.initialized)
844 			continue;
845 
846 		amdgpu_fence_process(ring);
847 
848 		seq_printf(m, "--- ring %d (%s) ---\n", i, ring->name);
849 		seq_printf(m, "Last signaled fence          0x%08x\n",
850 			   atomic_read(&ring->fence_drv.last_seq));
851 		seq_printf(m, "Last emitted                 0x%08x\n",
852 			   ring->fence_drv.sync_seq);
853 
854 		if (ring->funcs->type == AMDGPU_RING_TYPE_GFX ||
855 		    ring->funcs->type == AMDGPU_RING_TYPE_SDMA) {
856 			seq_printf(m, "Last signaled trailing fence 0x%08x\n",
857 				   le32_to_cpu(*ring->trail_fence_cpu_addr));
858 			seq_printf(m, "Last emitted                 0x%08x\n",
859 				   ring->trail_seq);
860 		}
861 
862 		if (ring->funcs->type != AMDGPU_RING_TYPE_GFX)
863 			continue;
864 
865 		/* set in CP_VMID_PREEMPT and preemption occurred */
866 		seq_printf(m, "Last preempted               0x%08x\n",
867 			   le32_to_cpu(*(ring->fence_drv.cpu_addr + 2)));
868 		/* set in CP_VMID_RESET and reset occurred */
869 		seq_printf(m, "Last reset                   0x%08x\n",
870 			   le32_to_cpu(*(ring->fence_drv.cpu_addr + 4)));
871 		/* Both preemption and reset occurred */
872 		seq_printf(m, "Last both                    0x%08x\n",
873 			   le32_to_cpu(*(ring->fence_drv.cpu_addr + 6)));
874 	}
875 	return 0;
876 }
877 
878 /*
879  * amdgpu_debugfs_gpu_recover - manually trigger a gpu reset & recover
880  *
881  * Manually trigger a gpu reset at the next fence wait.
882  */
883 static int gpu_recover_get(void *data, u64 *val)
884 {
885 	struct amdgpu_device *adev = (struct amdgpu_device *)data;
886 	struct drm_device *dev = adev_to_drm(adev);
887 	int r;
888 
889 	r = pm_runtime_get_sync(dev->dev);
890 	if (r < 0) {
891 		pm_runtime_put_autosuspend(dev->dev);
892 		return 0;
893 	}
894 
895 	if (amdgpu_reset_domain_schedule(adev->reset_domain, &adev->reset_work))
896 		flush_work(&adev->reset_work);
897 
898 	*val = atomic_read(&adev->reset_domain->reset_res);
899 
900 	pm_runtime_mark_last_busy(dev->dev);
901 	pm_runtime_put_autosuspend(dev->dev);
902 
903 	return 0;
904 }
905 
906 DEFINE_SHOW_ATTRIBUTE(amdgpu_debugfs_fence_info);
907 DEFINE_DEBUGFS_ATTRIBUTE(amdgpu_debugfs_gpu_recover_fops, gpu_recover_get, NULL,
908 			 "%lld\n");
909 
910 static void amdgpu_debugfs_reset_work(struct work_struct *work)
911 {
912 	struct amdgpu_device *adev = container_of(work, struct amdgpu_device,
913 						  reset_work);
914 
915 	struct amdgpu_reset_context reset_context;
916 	memset(&reset_context, 0, sizeof(reset_context));
917 
918 	reset_context.method = AMD_RESET_METHOD_NONE;
919 	reset_context.reset_req_dev = adev;
920 	set_bit(AMDGPU_NEED_FULL_RESET, &reset_context.flags);
921 
922 	amdgpu_device_gpu_recover(adev, NULL, &reset_context);
923 }
924 
925 #endif
926 
927 void amdgpu_debugfs_fence_init(struct amdgpu_device *adev)
928 {
929 #if defined(CONFIG_DEBUG_FS)
930 	struct drm_minor *minor = adev_to_drm(adev)->primary;
931 	struct dentry *root = minor->debugfs_root;
932 
933 	debugfs_create_file("amdgpu_fence_info", 0444, root, adev,
934 			    &amdgpu_debugfs_fence_info_fops);
935 
936 	if (!amdgpu_sriov_vf(adev)) {
937 
938 		INIT_WORK(&adev->reset_work, amdgpu_debugfs_reset_work);
939 		debugfs_create_file("amdgpu_gpu_recover", 0444, root, adev,
940 				    &amdgpu_debugfs_gpu_recover_fops);
941 	}
942 #endif
943 }
944 
945