Loading...
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
53struct amdgpu_fence {
54 struct dma_fence base;
55
56 /* RB, DMA, etc. */
57 struct amdgpu_ring *ring;
58 ktime_t start_timestamp;
59};
60
61static struct kmem_cache *amdgpu_fence_slab;
62
63int 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
73void amdgpu_fence_slab_fini(void)
74{
75 rcu_barrier();
76 kmem_cache_destroy(amdgpu_fence_slab);
77}
78/*
79 * Cast helper
80 */
81static const struct dma_fence_ops amdgpu_fence_ops;
82static const struct dma_fence_ops amdgpu_job_fence_ops;
83static 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 */
102static 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 */
118static 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 */
142int 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 */
226int 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 */
257static 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 */
274bool 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 */
326static 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 */
343int 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 */
376signed 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 */
399unsigned 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 */
419u64 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 */
450void 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 */
476int 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 */
512int 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 */
553int 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 */
566void 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 */
594void 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
611void 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 */
650void 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 */
672void 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 RCU_INIT_POINTER(*ptr, NULL);
682 dma_fence_put(old);
683 }
684 }
685}
686
687/**
688 * amdgpu_fence_driver_force_completion - force signal latest fence of ring
689 *
690 * @ring: fence of the ring to signal
691 *
692 */
693void amdgpu_fence_driver_force_completion(struct amdgpu_ring *ring)
694{
695 amdgpu_fence_write(ring, ring->fence_drv.sync_seq);
696 amdgpu_fence_process(ring);
697}
698
699/*
700 * Common fence implementation
701 */
702
703static const char *amdgpu_fence_get_driver_name(struct dma_fence *fence)
704{
705 return "amdgpu";
706}
707
708static const char *amdgpu_fence_get_timeline_name(struct dma_fence *f)
709{
710 return (const char *)to_amdgpu_fence(f)->ring->name;
711}
712
713static const char *amdgpu_job_fence_get_timeline_name(struct dma_fence *f)
714{
715 struct amdgpu_job *job = container_of(f, struct amdgpu_job, hw_fence);
716
717 return (const char *)to_amdgpu_ring(job->base.sched)->name;
718}
719
720/**
721 * amdgpu_fence_enable_signaling - enable signalling on fence
722 * @f: fence
723 *
724 * This function is called with fence_queue lock held, and adds a callback
725 * to fence_queue that checks if this fence is signaled, and if so it
726 * signals the fence and removes itself.
727 */
728static bool amdgpu_fence_enable_signaling(struct dma_fence *f)
729{
730 if (!timer_pending(&to_amdgpu_fence(f)->ring->fence_drv.fallback_timer))
731 amdgpu_fence_schedule_fallback(to_amdgpu_fence(f)->ring);
732
733 return true;
734}
735
736/**
737 * amdgpu_job_fence_enable_signaling - enable signalling on job fence
738 * @f: fence
739 *
740 * This is the simliar function with amdgpu_fence_enable_signaling above, it
741 * only handles the job embedded fence.
742 */
743static bool amdgpu_job_fence_enable_signaling(struct dma_fence *f)
744{
745 struct amdgpu_job *job = container_of(f, struct amdgpu_job, hw_fence);
746
747 if (!timer_pending(&to_amdgpu_ring(job->base.sched)->fence_drv.fallback_timer))
748 amdgpu_fence_schedule_fallback(to_amdgpu_ring(job->base.sched));
749
750 return true;
751}
752
753/**
754 * amdgpu_fence_free - free up the fence memory
755 *
756 * @rcu: RCU callback head
757 *
758 * Free up the fence memory after the RCU grace period.
759 */
760static void amdgpu_fence_free(struct rcu_head *rcu)
761{
762 struct dma_fence *f = container_of(rcu, struct dma_fence, rcu);
763
764 /* free fence_slab if it's separated fence*/
765 kmem_cache_free(amdgpu_fence_slab, to_amdgpu_fence(f));
766}
767
768/**
769 * amdgpu_job_fence_free - free up the job with embedded fence
770 *
771 * @rcu: RCU callback head
772 *
773 * Free up the job with embedded fence after the RCU grace period.
774 */
775static void amdgpu_job_fence_free(struct rcu_head *rcu)
776{
777 struct dma_fence *f = container_of(rcu, struct dma_fence, rcu);
778
779 /* free job if fence has a parent job */
780 kfree(container_of(f, struct amdgpu_job, hw_fence));
781}
782
783/**
784 * amdgpu_fence_release - callback that fence can be freed
785 *
786 * @f: fence
787 *
788 * This function is called when the reference count becomes zero.
789 * It just RCU schedules freeing up the fence.
790 */
791static void amdgpu_fence_release(struct dma_fence *f)
792{
793 call_rcu(&f->rcu, amdgpu_fence_free);
794}
795
796/**
797 * amdgpu_job_fence_release - callback that job embedded fence can be freed
798 *
799 * @f: fence
800 *
801 * This is the simliar function with amdgpu_fence_release above, it
802 * only handles the job embedded fence.
803 */
804static void amdgpu_job_fence_release(struct dma_fence *f)
805{
806 call_rcu(&f->rcu, amdgpu_job_fence_free);
807}
808
809static const struct dma_fence_ops amdgpu_fence_ops = {
810 .get_driver_name = amdgpu_fence_get_driver_name,
811 .get_timeline_name = amdgpu_fence_get_timeline_name,
812 .enable_signaling = amdgpu_fence_enable_signaling,
813 .release = amdgpu_fence_release,
814};
815
816static const struct dma_fence_ops amdgpu_job_fence_ops = {
817 .get_driver_name = amdgpu_fence_get_driver_name,
818 .get_timeline_name = amdgpu_job_fence_get_timeline_name,
819 .enable_signaling = amdgpu_job_fence_enable_signaling,
820 .release = amdgpu_job_fence_release,
821};
822
823/*
824 * Fence debugfs
825 */
826#if defined(CONFIG_DEBUG_FS)
827static int amdgpu_debugfs_fence_info_show(struct seq_file *m, void *unused)
828{
829 struct amdgpu_device *adev = (struct amdgpu_device *)m->private;
830 int i;
831
832 for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
833 struct amdgpu_ring *ring = adev->rings[i];
834 if (!ring || !ring->fence_drv.initialized)
835 continue;
836
837 amdgpu_fence_process(ring);
838
839 seq_printf(m, "--- ring %d (%s) ---\n", i, ring->name);
840 seq_printf(m, "Last signaled fence 0x%08x\n",
841 atomic_read(&ring->fence_drv.last_seq));
842 seq_printf(m, "Last emitted 0x%08x\n",
843 ring->fence_drv.sync_seq);
844
845 if (ring->funcs->type == AMDGPU_RING_TYPE_GFX ||
846 ring->funcs->type == AMDGPU_RING_TYPE_SDMA) {
847 seq_printf(m, "Last signaled trailing fence 0x%08x\n",
848 le32_to_cpu(*ring->trail_fence_cpu_addr));
849 seq_printf(m, "Last emitted 0x%08x\n",
850 ring->trail_seq);
851 }
852
853 if (ring->funcs->type != AMDGPU_RING_TYPE_GFX)
854 continue;
855
856 /* set in CP_VMID_PREEMPT and preemption occurred */
857 seq_printf(m, "Last preempted 0x%08x\n",
858 le32_to_cpu(*(ring->fence_drv.cpu_addr + 2)));
859 /* set in CP_VMID_RESET and reset occurred */
860 seq_printf(m, "Last reset 0x%08x\n",
861 le32_to_cpu(*(ring->fence_drv.cpu_addr + 4)));
862 /* Both preemption and reset occurred */
863 seq_printf(m, "Last both 0x%08x\n",
864 le32_to_cpu(*(ring->fence_drv.cpu_addr + 6)));
865 }
866 return 0;
867}
868
869/*
870 * amdgpu_debugfs_gpu_recover - manually trigger a gpu reset & recover
871 *
872 * Manually trigger a gpu reset at the next fence wait.
873 */
874static int gpu_recover_get(void *data, u64 *val)
875{
876 struct amdgpu_device *adev = (struct amdgpu_device *)data;
877 struct drm_device *dev = adev_to_drm(adev);
878 int r;
879
880 r = pm_runtime_get_sync(dev->dev);
881 if (r < 0) {
882 pm_runtime_put_autosuspend(dev->dev);
883 return 0;
884 }
885
886 if (amdgpu_reset_domain_schedule(adev->reset_domain, &adev->reset_work))
887 flush_work(&adev->reset_work);
888
889 *val = atomic_read(&adev->reset_domain->reset_res);
890
891 pm_runtime_mark_last_busy(dev->dev);
892 pm_runtime_put_autosuspend(dev->dev);
893
894 return 0;
895}
896
897DEFINE_SHOW_ATTRIBUTE(amdgpu_debugfs_fence_info);
898DEFINE_DEBUGFS_ATTRIBUTE(amdgpu_debugfs_gpu_recover_fops, gpu_recover_get, NULL,
899 "%lld\n");
900
901static void amdgpu_debugfs_reset_work(struct work_struct *work)
902{
903 struct amdgpu_device *adev = container_of(work, struct amdgpu_device,
904 reset_work);
905
906 struct amdgpu_reset_context reset_context;
907 memset(&reset_context, 0, sizeof(reset_context));
908
909 reset_context.method = AMD_RESET_METHOD_NONE;
910 reset_context.reset_req_dev = adev;
911 set_bit(AMDGPU_NEED_FULL_RESET, &reset_context.flags);
912
913 amdgpu_device_gpu_recover(adev, NULL, &reset_context);
914}
915
916#endif
917
918void amdgpu_debugfs_fence_init(struct amdgpu_device *adev)
919{
920#if defined(CONFIG_DEBUG_FS)
921 struct drm_minor *minor = adev_to_drm(adev)->primary;
922 struct dentry *root = minor->debugfs_root;
923
924 debugfs_create_file("amdgpu_fence_info", 0444, root, adev,
925 &amdgpu_debugfs_fence_info_fops);
926
927 if (!amdgpu_sriov_vf(adev)) {
928
929 INIT_WORK(&adev->reset_work, amdgpu_debugfs_reset_work);
930 debugfs_create_file("amdgpu_gpu_recover", 0444, root, adev,
931 &amdgpu_debugfs_gpu_recover_fops);
932 }
933#endif
934}
935
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 mark an event in the GPUs pipeline and are used
46 * for GPU/CPU synchronization. When the fence is written,
47 * it is expected that all buffers associated with that fence
48 * are no longer in use by the associated ring on the GPU and
49 * that the relevant GPU caches have been flushed.
50 */
51
52struct amdgpu_fence {
53 struct dma_fence base;
54
55 /* RB, DMA, etc. */
56 struct amdgpu_ring *ring;
57 ktime_t start_timestamp;
58};
59
60static struct kmem_cache *amdgpu_fence_slab;
61
62int amdgpu_fence_slab_init(void)
63{
64 amdgpu_fence_slab = KMEM_CACHE(amdgpu_fence, SLAB_HWCACHE_ALIGN);
65 if (!amdgpu_fence_slab)
66 return -ENOMEM;
67 return 0;
68}
69
70void amdgpu_fence_slab_fini(void)
71{
72 rcu_barrier();
73 kmem_cache_destroy(amdgpu_fence_slab);
74}
75/*
76 * Cast helper
77 */
78static const struct dma_fence_ops amdgpu_fence_ops;
79static const struct dma_fence_ops amdgpu_job_fence_ops;
80static inline struct amdgpu_fence *to_amdgpu_fence(struct dma_fence *f)
81{
82 struct amdgpu_fence *__f = container_of(f, struct amdgpu_fence, base);
83
84 if (__f->base.ops == &amdgpu_fence_ops ||
85 __f->base.ops == &amdgpu_job_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 */
99static 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 */
115static 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 * @job: job the fence is embedded in
134 * @flags: flags to pass into the subordinate .emit_fence() call
135 *
136 * Emits a fence command on the requested ring (all asics).
137 * Returns 0 on success, -ENOMEM on failure.
138 */
139int amdgpu_fence_emit(struct amdgpu_ring *ring, struct dma_fence **f, struct amdgpu_job *job,
140 unsigned int flags)
141{
142 struct amdgpu_device *adev = ring->adev;
143 struct dma_fence *fence;
144 struct amdgpu_fence *am_fence;
145 struct dma_fence __rcu **ptr;
146 uint32_t seq;
147 int r;
148
149 if (job == NULL) {
150 /* create a sperate hw fence */
151 am_fence = kmem_cache_alloc(amdgpu_fence_slab, GFP_ATOMIC);
152 if (am_fence == NULL)
153 return -ENOMEM;
154 fence = &am_fence->base;
155 am_fence->ring = ring;
156 } else {
157 /* take use of job-embedded fence */
158 fence = &job->hw_fence;
159 }
160
161 seq = ++ring->fence_drv.sync_seq;
162 if (job && job->job_run_counter) {
163 /* reinit seq for resubmitted jobs */
164 fence->seqno = seq;
165 /* TO be inline with external fence creation and other drivers */
166 dma_fence_get(fence);
167 } else {
168 if (job) {
169 dma_fence_init(fence, &amdgpu_job_fence_ops,
170 &ring->fence_drv.lock,
171 adev->fence_context + ring->idx, seq);
172 /* Against remove in amdgpu_job_{free, free_cb} */
173 dma_fence_get(fence);
174 } else {
175 dma_fence_init(fence, &amdgpu_fence_ops,
176 &ring->fence_drv.lock,
177 adev->fence_context + ring->idx, seq);
178 }
179 }
180
181 amdgpu_ring_emit_fence(ring, ring->fence_drv.gpu_addr,
182 seq, flags | AMDGPU_FENCE_FLAG_INT);
183 pm_runtime_get_noresume(adev_to_drm(adev)->dev);
184 ptr = &ring->fence_drv.fences[seq & ring->fence_drv.num_fences_mask];
185 if (unlikely(rcu_dereference_protected(*ptr, 1))) {
186 struct dma_fence *old;
187
188 rcu_read_lock();
189 old = dma_fence_get_rcu_safe(ptr);
190 rcu_read_unlock();
191
192 if (old) {
193 r = dma_fence_wait(old, false);
194 dma_fence_put(old);
195 if (r)
196 return r;
197 }
198 }
199
200 to_amdgpu_fence(fence)->start_timestamp = ktime_get();
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 */
223int 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 */
254static 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 */
271bool 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 */
323static 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 */
340int 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 */
373signed long amdgpu_fence_wait_polling(struct amdgpu_ring *ring,
374 uint32_t wait_seq,
375 signed long timeout)
376{
377
378 while ((int32_t)(wait_seq - amdgpu_fence_read(ring)) > 0 && timeout > 0) {
379 udelay(2);
380 timeout -= 2;
381 }
382 return timeout > 0 ? timeout : 0;
383}
384/**
385 * amdgpu_fence_count_emitted - get the count of emitted fences
386 *
387 * @ring: ring the fence is associated with
388 *
389 * Get the number of fences emitted on the requested ring (all asics).
390 * Returns the number of emitted fences on the ring. Used by the
391 * dynpm code to ring track activity.
392 */
393unsigned int amdgpu_fence_count_emitted(struct amdgpu_ring *ring)
394{
395 uint64_t emitted;
396
397 /* We are not protected by ring lock when reading the last sequence
398 * but it's ok to report slightly wrong fence count here.
399 */
400 emitted = 0x100000000ull;
401 emitted -= atomic_read(&ring->fence_drv.last_seq);
402 emitted += READ_ONCE(ring->fence_drv.sync_seq);
403 return lower_32_bits(emitted);
404}
405
406/**
407 * amdgpu_fence_last_unsignaled_time_us - the time fence emitted until now
408 * @ring: ring the fence is associated with
409 *
410 * Find the earliest fence unsignaled until now, calculate the time delta
411 * between the time fence emitted and now.
412 */
413u64 amdgpu_fence_last_unsignaled_time_us(struct amdgpu_ring *ring)
414{
415 struct amdgpu_fence_driver *drv = &ring->fence_drv;
416 struct dma_fence *fence;
417 uint32_t last_seq, sync_seq;
418
419 last_seq = atomic_read(&ring->fence_drv.last_seq);
420 sync_seq = READ_ONCE(ring->fence_drv.sync_seq);
421 if (last_seq == sync_seq)
422 return 0;
423
424 ++last_seq;
425 last_seq &= drv->num_fences_mask;
426 fence = drv->fences[last_seq];
427 if (!fence)
428 return 0;
429
430 return ktime_us_delta(ktime_get(),
431 to_amdgpu_fence(fence)->start_timestamp);
432}
433
434/**
435 * amdgpu_fence_update_start_timestamp - update the timestamp of the fence
436 * @ring: ring the fence is associated with
437 * @seq: the fence seq number to update.
438 * @timestamp: the start timestamp to update.
439 *
440 * The function called at the time the fence and related ib is about to
441 * resubmit to gpu in MCBP scenario. Thus we do not consider race condition
442 * with amdgpu_fence_process to modify the same fence.
443 */
444void amdgpu_fence_update_start_timestamp(struct amdgpu_ring *ring, uint32_t seq, ktime_t timestamp)
445{
446 struct amdgpu_fence_driver *drv = &ring->fence_drv;
447 struct dma_fence *fence;
448
449 seq &= drv->num_fences_mask;
450 fence = drv->fences[seq];
451 if (!fence)
452 return;
453
454 to_amdgpu_fence(fence)->start_timestamp = timestamp;
455}
456
457/**
458 * amdgpu_fence_driver_start_ring - make the fence driver
459 * ready for use on the requested ring.
460 *
461 * @ring: ring to start the fence driver on
462 * @irq_src: interrupt source to use for this ring
463 * @irq_type: interrupt type to use for this ring
464 *
465 * Make the fence driver ready for processing (all asics).
466 * Not all asics have all rings, so each asic will only
467 * start the fence driver on the rings it has.
468 * Returns 0 for success, errors for failure.
469 */
470int amdgpu_fence_driver_start_ring(struct amdgpu_ring *ring,
471 struct amdgpu_irq_src *irq_src,
472 unsigned int irq_type)
473{
474 struct amdgpu_device *adev = ring->adev;
475 uint64_t index;
476
477 if (ring->funcs->type != AMDGPU_RING_TYPE_UVD) {
478 ring->fence_drv.cpu_addr = ring->fence_cpu_addr;
479 ring->fence_drv.gpu_addr = ring->fence_gpu_addr;
480 } else {
481 /* put fence directly behind firmware */
482 index = ALIGN(adev->uvd.fw->size, 8);
483 ring->fence_drv.cpu_addr = adev->uvd.inst[ring->me].cpu_addr + index;
484 ring->fence_drv.gpu_addr = adev->uvd.inst[ring->me].gpu_addr + index;
485 }
486 amdgpu_fence_write(ring, atomic_read(&ring->fence_drv.last_seq));
487
488 ring->fence_drv.irq_src = irq_src;
489 ring->fence_drv.irq_type = irq_type;
490 ring->fence_drv.initialized = true;
491
492 DRM_DEV_DEBUG(adev->dev, "fence driver on ring %s use gpu addr 0x%016llx\n",
493 ring->name, ring->fence_drv.gpu_addr);
494 return 0;
495}
496
497/**
498 * amdgpu_fence_driver_init_ring - init the fence driver
499 * for the requested ring.
500 *
501 * @ring: ring to init the fence driver on
502 *
503 * Init the fence driver for the requested ring (all asics).
504 * Helper function for amdgpu_fence_driver_init().
505 */
506int amdgpu_fence_driver_init_ring(struct amdgpu_ring *ring)
507{
508 struct amdgpu_device *adev = ring->adev;
509
510 if (!adev)
511 return -EINVAL;
512
513 if (!is_power_of_2(ring->num_hw_submission))
514 return -EINVAL;
515
516 ring->fence_drv.cpu_addr = NULL;
517 ring->fence_drv.gpu_addr = 0;
518 ring->fence_drv.sync_seq = 0;
519 atomic_set(&ring->fence_drv.last_seq, 0);
520 ring->fence_drv.initialized = false;
521
522 timer_setup(&ring->fence_drv.fallback_timer, amdgpu_fence_fallback, 0);
523
524 ring->fence_drv.num_fences_mask = ring->num_hw_submission * 2 - 1;
525 spin_lock_init(&ring->fence_drv.lock);
526 ring->fence_drv.fences = kcalloc(ring->num_hw_submission * 2, sizeof(void *),
527 GFP_KERNEL);
528
529 if (!ring->fence_drv.fences)
530 return -ENOMEM;
531
532 return 0;
533}
534
535/**
536 * amdgpu_fence_driver_sw_init - init the fence driver
537 * for all possible rings.
538 *
539 * @adev: amdgpu device pointer
540 *
541 * Init the fence driver for all possible rings (all asics).
542 * Not all asics have all rings, so each asic will only
543 * start the fence driver on the rings it has using
544 * amdgpu_fence_driver_start_ring().
545 * Returns 0 for success.
546 */
547int amdgpu_fence_driver_sw_init(struct amdgpu_device *adev)
548{
549 return 0;
550}
551
552/**
553 * amdgpu_fence_need_ring_interrupt_restore - helper function to check whether
554 * fence driver interrupts need to be restored.
555 *
556 * @ring: ring that to be checked
557 *
558 * Interrupts for rings that belong to GFX IP don't need to be restored
559 * when the target power state is s0ix.
560 *
561 * Return true if need to restore interrupts, false otherwise.
562 */
563static bool amdgpu_fence_need_ring_interrupt_restore(struct amdgpu_ring *ring)
564{
565 struct amdgpu_device *adev = ring->adev;
566 bool is_gfx_power_domain = false;
567
568 switch (ring->funcs->type) {
569 case AMDGPU_RING_TYPE_SDMA:
570 /* SDMA 5.x+ is part of GFX power domain so it's covered by GFXOFF */
571 if (amdgpu_ip_version(adev, SDMA0_HWIP, 0) >=
572 IP_VERSION(5, 0, 0))
573 is_gfx_power_domain = true;
574 break;
575 case AMDGPU_RING_TYPE_GFX:
576 case AMDGPU_RING_TYPE_COMPUTE:
577 case AMDGPU_RING_TYPE_KIQ:
578 case AMDGPU_RING_TYPE_MES:
579 is_gfx_power_domain = true;
580 break;
581 default:
582 break;
583 }
584
585 return !(adev->in_s0ix && is_gfx_power_domain);
586}
587
588/**
589 * amdgpu_fence_driver_hw_fini - tear down the fence driver
590 * for all possible rings.
591 *
592 * @adev: amdgpu device pointer
593 *
594 * Tear down the fence driver for all possible rings (all asics).
595 */
596void amdgpu_fence_driver_hw_fini(struct amdgpu_device *adev)
597{
598 int i, r;
599
600 for (i = 0; i < AMDGPU_MAX_RINGS; i++) {
601 struct amdgpu_ring *ring = adev->rings[i];
602
603 if (!ring || !ring->fence_drv.initialized)
604 continue;
605
606 /* You can't wait for HW to signal if it's gone */
607 if (!drm_dev_is_unplugged(adev_to_drm(adev)))
608 r = amdgpu_fence_wait_empty(ring);
609 else
610 r = -ENODEV;
611 /* no need to trigger GPU reset as we are unloading */
612 if (r)
613 amdgpu_fence_driver_force_completion(ring);
614
615 if (!drm_dev_is_unplugged(adev_to_drm(adev)) &&
616 ring->fence_drv.irq_src &&
617 amdgpu_fence_need_ring_interrupt_restore(ring))
618 amdgpu_irq_put(adev, ring->fence_drv.irq_src,
619 ring->fence_drv.irq_type);
620
621 del_timer_sync(&ring->fence_drv.fallback_timer);
622 }
623}
624
625/* Will either stop and flush handlers for amdgpu interrupt or reanble it */
626void amdgpu_fence_driver_isr_toggle(struct amdgpu_device *adev, bool stop)
627{
628 int i;
629
630 for (i = 0; i < AMDGPU_MAX_RINGS; i++) {
631 struct amdgpu_ring *ring = adev->rings[i];
632
633 if (!ring || !ring->fence_drv.initialized || !ring->fence_drv.irq_src)
634 continue;
635
636 if (stop)
637 disable_irq(adev->irq.irq);
638 else
639 enable_irq(adev->irq.irq);
640 }
641}
642
643void amdgpu_fence_driver_sw_fini(struct amdgpu_device *adev)
644{
645 unsigned int i, j;
646
647 for (i = 0; i < AMDGPU_MAX_RINGS; i++) {
648 struct amdgpu_ring *ring = adev->rings[i];
649
650 if (!ring || !ring->fence_drv.initialized)
651 continue;
652
653 /*
654 * Notice we check for sched.ops since there's some
655 * override on the meaning of sched.ready by amdgpu.
656 * The natural check would be sched.ready, which is
657 * set as drm_sched_init() finishes...
658 */
659 if (ring->sched.ops)
660 drm_sched_fini(&ring->sched);
661
662 for (j = 0; j <= ring->fence_drv.num_fences_mask; ++j)
663 dma_fence_put(ring->fence_drv.fences[j]);
664 kfree(ring->fence_drv.fences);
665 ring->fence_drv.fences = NULL;
666 ring->fence_drv.initialized = false;
667 }
668}
669
670/**
671 * amdgpu_fence_driver_hw_init - enable the fence driver
672 * for all possible rings.
673 *
674 * @adev: amdgpu device pointer
675 *
676 * Enable the fence driver for all possible rings (all asics).
677 * Not all asics have all rings, so each asic will only
678 * start the fence driver on the rings it has using
679 * amdgpu_fence_driver_start_ring().
680 * Returns 0 for success.
681 */
682void amdgpu_fence_driver_hw_init(struct amdgpu_device *adev)
683{
684 int i;
685
686 for (i = 0; i < AMDGPU_MAX_RINGS; i++) {
687 struct amdgpu_ring *ring = adev->rings[i];
688
689 if (!ring || !ring->fence_drv.initialized)
690 continue;
691
692 /* enable the interrupt */
693 if (ring->fence_drv.irq_src &&
694 amdgpu_fence_need_ring_interrupt_restore(ring))
695 amdgpu_irq_get(adev, ring->fence_drv.irq_src,
696 ring->fence_drv.irq_type);
697 }
698}
699
700/**
701 * amdgpu_fence_driver_clear_job_fences - clear job embedded fences of ring
702 *
703 * @ring: fence of the ring to be cleared
704 *
705 */
706void amdgpu_fence_driver_clear_job_fences(struct amdgpu_ring *ring)
707{
708 int i;
709 struct dma_fence *old, **ptr;
710
711 for (i = 0; i <= ring->fence_drv.num_fences_mask; i++) {
712 ptr = &ring->fence_drv.fences[i];
713 old = rcu_dereference_protected(*ptr, 1);
714 if (old && old->ops == &amdgpu_job_fence_ops) {
715 struct amdgpu_job *job;
716
717 /* For non-scheduler bad job, i.e. failed ib test, we need to signal
718 * it right here or we won't be able to track them in fence_drv
719 * and they will remain unsignaled during sa_bo free.
720 */
721 job = container_of(old, struct amdgpu_job, hw_fence);
722 if (!job->base.s_fence && !dma_fence_is_signaled(old))
723 dma_fence_signal(old);
724 RCU_INIT_POINTER(*ptr, NULL);
725 dma_fence_put(old);
726 }
727 }
728}
729
730/**
731 * amdgpu_fence_driver_set_error - set error code on fences
732 * @ring: the ring which contains the fences
733 * @error: the error code to set
734 *
735 * Set an error code to all the fences pending on the ring.
736 */
737void amdgpu_fence_driver_set_error(struct amdgpu_ring *ring, int error)
738{
739 struct amdgpu_fence_driver *drv = &ring->fence_drv;
740 unsigned long flags;
741
742 spin_lock_irqsave(&drv->lock, flags);
743 for (unsigned int i = 0; i <= drv->num_fences_mask; ++i) {
744 struct dma_fence *fence;
745
746 fence = rcu_dereference_protected(drv->fences[i],
747 lockdep_is_held(&drv->lock));
748 if (fence && !dma_fence_is_signaled_locked(fence))
749 dma_fence_set_error(fence, error);
750 }
751 spin_unlock_irqrestore(&drv->lock, flags);
752}
753
754/**
755 * amdgpu_fence_driver_force_completion - force signal latest fence of ring
756 *
757 * @ring: fence of the ring to signal
758 *
759 */
760void amdgpu_fence_driver_force_completion(struct amdgpu_ring *ring)
761{
762 amdgpu_fence_driver_set_error(ring, -ECANCELED);
763 amdgpu_fence_write(ring, ring->fence_drv.sync_seq);
764 amdgpu_fence_process(ring);
765}
766
767/*
768 * Common fence implementation
769 */
770
771static const char *amdgpu_fence_get_driver_name(struct dma_fence *fence)
772{
773 return "amdgpu";
774}
775
776static const char *amdgpu_fence_get_timeline_name(struct dma_fence *f)
777{
778 return (const char *)to_amdgpu_fence(f)->ring->name;
779}
780
781static const char *amdgpu_job_fence_get_timeline_name(struct dma_fence *f)
782{
783 struct amdgpu_job *job = container_of(f, struct amdgpu_job, hw_fence);
784
785 return (const char *)to_amdgpu_ring(job->base.sched)->name;
786}
787
788/**
789 * amdgpu_fence_enable_signaling - enable signalling on fence
790 * @f: fence
791 *
792 * This function is called with fence_queue lock held, and adds a callback
793 * to fence_queue that checks if this fence is signaled, and if so it
794 * signals the fence and removes itself.
795 */
796static bool amdgpu_fence_enable_signaling(struct dma_fence *f)
797{
798 if (!timer_pending(&to_amdgpu_fence(f)->ring->fence_drv.fallback_timer))
799 amdgpu_fence_schedule_fallback(to_amdgpu_fence(f)->ring);
800
801 return true;
802}
803
804/**
805 * amdgpu_job_fence_enable_signaling - enable signalling on job fence
806 * @f: fence
807 *
808 * This is the simliar function with amdgpu_fence_enable_signaling above, it
809 * only handles the job embedded fence.
810 */
811static bool amdgpu_job_fence_enable_signaling(struct dma_fence *f)
812{
813 struct amdgpu_job *job = container_of(f, struct amdgpu_job, hw_fence);
814
815 if (!timer_pending(&to_amdgpu_ring(job->base.sched)->fence_drv.fallback_timer))
816 amdgpu_fence_schedule_fallback(to_amdgpu_ring(job->base.sched));
817
818 return true;
819}
820
821/**
822 * amdgpu_fence_free - free up the fence memory
823 *
824 * @rcu: RCU callback head
825 *
826 * Free up the fence memory after the RCU grace period.
827 */
828static void amdgpu_fence_free(struct rcu_head *rcu)
829{
830 struct dma_fence *f = container_of(rcu, struct dma_fence, rcu);
831
832 /* free fence_slab if it's separated fence*/
833 kmem_cache_free(amdgpu_fence_slab, to_amdgpu_fence(f));
834}
835
836/**
837 * amdgpu_job_fence_free - free up the job with embedded fence
838 *
839 * @rcu: RCU callback head
840 *
841 * Free up the job with embedded fence after the RCU grace period.
842 */
843static void amdgpu_job_fence_free(struct rcu_head *rcu)
844{
845 struct dma_fence *f = container_of(rcu, struct dma_fence, rcu);
846
847 /* free job if fence has a parent job */
848 kfree(container_of(f, struct amdgpu_job, hw_fence));
849}
850
851/**
852 * amdgpu_fence_release - callback that fence can be freed
853 *
854 * @f: fence
855 *
856 * This function is called when the reference count becomes zero.
857 * It just RCU schedules freeing up the fence.
858 */
859static void amdgpu_fence_release(struct dma_fence *f)
860{
861 call_rcu(&f->rcu, amdgpu_fence_free);
862}
863
864/**
865 * amdgpu_job_fence_release - callback that job embedded fence can be freed
866 *
867 * @f: fence
868 *
869 * This is the simliar function with amdgpu_fence_release above, it
870 * only handles the job embedded fence.
871 */
872static void amdgpu_job_fence_release(struct dma_fence *f)
873{
874 call_rcu(&f->rcu, amdgpu_job_fence_free);
875}
876
877static const struct dma_fence_ops amdgpu_fence_ops = {
878 .get_driver_name = amdgpu_fence_get_driver_name,
879 .get_timeline_name = amdgpu_fence_get_timeline_name,
880 .enable_signaling = amdgpu_fence_enable_signaling,
881 .release = amdgpu_fence_release,
882};
883
884static const struct dma_fence_ops amdgpu_job_fence_ops = {
885 .get_driver_name = amdgpu_fence_get_driver_name,
886 .get_timeline_name = amdgpu_job_fence_get_timeline_name,
887 .enable_signaling = amdgpu_job_fence_enable_signaling,
888 .release = amdgpu_job_fence_release,
889};
890
891/*
892 * Fence debugfs
893 */
894#if defined(CONFIG_DEBUG_FS)
895static int amdgpu_debugfs_fence_info_show(struct seq_file *m, void *unused)
896{
897 struct amdgpu_device *adev = m->private;
898 int i;
899
900 for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
901 struct amdgpu_ring *ring = adev->rings[i];
902
903 if (!ring || !ring->fence_drv.initialized)
904 continue;
905
906 amdgpu_fence_process(ring);
907
908 seq_printf(m, "--- ring %d (%s) ---\n", i, ring->name);
909 seq_printf(m, "Last signaled fence 0x%08x\n",
910 atomic_read(&ring->fence_drv.last_seq));
911 seq_printf(m, "Last emitted 0x%08x\n",
912 ring->fence_drv.sync_seq);
913
914 if (ring->funcs->type == AMDGPU_RING_TYPE_GFX ||
915 ring->funcs->type == AMDGPU_RING_TYPE_SDMA) {
916 seq_printf(m, "Last signaled trailing fence 0x%08x\n",
917 le32_to_cpu(*ring->trail_fence_cpu_addr));
918 seq_printf(m, "Last emitted 0x%08x\n",
919 ring->trail_seq);
920 }
921
922 if (ring->funcs->type != AMDGPU_RING_TYPE_GFX)
923 continue;
924
925 /* set in CP_VMID_PREEMPT and preemption occurred */
926 seq_printf(m, "Last preempted 0x%08x\n",
927 le32_to_cpu(*(ring->fence_drv.cpu_addr + 2)));
928 /* set in CP_VMID_RESET and reset occurred */
929 seq_printf(m, "Last reset 0x%08x\n",
930 le32_to_cpu(*(ring->fence_drv.cpu_addr + 4)));
931 /* Both preemption and reset occurred */
932 seq_printf(m, "Last both 0x%08x\n",
933 le32_to_cpu(*(ring->fence_drv.cpu_addr + 6)));
934 }
935 return 0;
936}
937
938/*
939 * amdgpu_debugfs_gpu_recover - manually trigger a gpu reset & recover
940 *
941 * Manually trigger a gpu reset at the next fence wait.
942 */
943static int gpu_recover_get(void *data, u64 *val)
944{
945 struct amdgpu_device *adev = (struct amdgpu_device *)data;
946 struct drm_device *dev = adev_to_drm(adev);
947 int r;
948
949 r = pm_runtime_get_sync(dev->dev);
950 if (r < 0) {
951 pm_runtime_put_autosuspend(dev->dev);
952 return 0;
953 }
954
955 if (amdgpu_reset_domain_schedule(adev->reset_domain, &adev->reset_work))
956 flush_work(&adev->reset_work);
957
958 *val = atomic_read(&adev->reset_domain->reset_res);
959
960 pm_runtime_mark_last_busy(dev->dev);
961 pm_runtime_put_autosuspend(dev->dev);
962
963 return 0;
964}
965
966DEFINE_SHOW_ATTRIBUTE(amdgpu_debugfs_fence_info);
967DEFINE_DEBUGFS_ATTRIBUTE(amdgpu_debugfs_gpu_recover_fops, gpu_recover_get, NULL,
968 "%lld\n");
969
970static void amdgpu_debugfs_reset_work(struct work_struct *work)
971{
972 struct amdgpu_device *adev = container_of(work, struct amdgpu_device,
973 reset_work);
974
975 struct amdgpu_reset_context reset_context;
976
977 memset(&reset_context, 0, sizeof(reset_context));
978
979 reset_context.method = AMD_RESET_METHOD_NONE;
980 reset_context.reset_req_dev = adev;
981 reset_context.src = AMDGPU_RESET_SRC_USER;
982 set_bit(AMDGPU_NEED_FULL_RESET, &reset_context.flags);
983 set_bit(AMDGPU_SKIP_COREDUMP, &reset_context.flags);
984
985 amdgpu_device_gpu_recover(adev, NULL, &reset_context);
986}
987
988#endif
989
990void amdgpu_debugfs_fence_init(struct amdgpu_device *adev)
991{
992#if defined(CONFIG_DEBUG_FS)
993 struct drm_minor *minor = adev_to_drm(adev)->primary;
994 struct dentry *root = minor->debugfs_root;
995
996 debugfs_create_file("amdgpu_fence_info", 0444, root, adev,
997 &amdgpu_debugfs_fence_info_fops);
998
999 if (!amdgpu_sriov_vf(adev)) {
1000
1001 INIT_WORK(&adev->reset_work, amdgpu_debugfs_reset_work);
1002 debugfs_create_file("amdgpu_gpu_recover", 0444, root, adev,
1003 &amdgpu_debugfs_gpu_recover_fops);
1004 }
1005#endif
1006}
1007