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