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
43/*
44 * Fences
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 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};
58
59static struct kmem_cache *amdgpu_fence_slab;
60
61int amdgpu_fence_slab_init(void)
62{
63 amdgpu_fence_slab = kmem_cache_create(
64 "amdgpu_fence", sizeof(struct amdgpu_fence), 0,
65 SLAB_HWCACHE_ALIGN, NULL);
66 if (!amdgpu_fence_slab)
67 return -ENOMEM;
68 return 0;
69}
70
71void amdgpu_fence_slab_fini(void)
72{
73 rcu_barrier();
74 kmem_cache_destroy(amdgpu_fence_slab);
75}
76/*
77 * Cast helper
78 */
79static const struct dma_fence_ops amdgpu_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 return __f;
86
87 return NULL;
88}
89
90/**
91 * amdgpu_fence_write - write a fence value
92 *
93 * @ring: ring the fence is associated with
94 * @seq: sequence number to write
95 *
96 * Writes a fence value to memory (all asics).
97 */
98static void amdgpu_fence_write(struct amdgpu_ring *ring, u32 seq)
99{
100 struct amdgpu_fence_driver *drv = &ring->fence_drv;
101
102 if (drv->cpu_addr)
103 *drv->cpu_addr = cpu_to_le32(seq);
104}
105
106/**
107 * amdgpu_fence_read - read a fence value
108 *
109 * @ring: ring the fence is associated with
110 *
111 * Reads a fence value from memory (all asics).
112 * Returns the value of the fence read from memory.
113 */
114static u32 amdgpu_fence_read(struct amdgpu_ring *ring)
115{
116 struct amdgpu_fence_driver *drv = &ring->fence_drv;
117 u32 seq = 0;
118
119 if (drv->cpu_addr)
120 seq = le32_to_cpu(*drv->cpu_addr);
121 else
122 seq = atomic_read(&drv->last_seq);
123
124 return seq;
125}
126
127/**
128 * amdgpu_fence_emit - emit a fence on the requested ring
129 *
130 * @ring: ring the fence is associated with
131 * @f: resulting fence object
132 * @flags: flags to pass into the subordinate .emit_fence() call
133 *
134 * Emits a fence command on the requested ring (all asics).
135 * Returns 0 on success, -ENOMEM on failure.
136 */
137int amdgpu_fence_emit(struct amdgpu_ring *ring, struct dma_fence **f,
138 unsigned flags)
139{
140 struct amdgpu_device *adev = ring->adev;
141 struct amdgpu_fence *fence;
142 struct dma_fence __rcu **ptr;
143 uint32_t seq;
144 int r;
145
146 fence = kmem_cache_alloc(amdgpu_fence_slab, GFP_KERNEL);
147 if (fence == NULL)
148 return -ENOMEM;
149
150 seq = ++ring->fence_drv.sync_seq;
151 fence->ring = ring;
152 dma_fence_init(&fence->base, &amdgpu_fence_ops,
153 &ring->fence_drv.lock,
154 adev->fence_context + ring->idx,
155 seq);
156 amdgpu_ring_emit_fence(ring, ring->fence_drv.gpu_addr,
157 seq, flags | AMDGPU_FENCE_FLAG_INT);
158 pm_runtime_get_noresume(adev_to_drm(adev)->dev);
159 ptr = &ring->fence_drv.fences[seq & ring->fence_drv.num_fences_mask];
160 if (unlikely(rcu_dereference_protected(*ptr, 1))) {
161 struct dma_fence *old;
162
163 rcu_read_lock();
164 old = dma_fence_get_rcu_safe(ptr);
165 rcu_read_unlock();
166
167 if (old) {
168 r = dma_fence_wait(old, false);
169 dma_fence_put(old);
170 if (r)
171 return r;
172 }
173 }
174
175 /* This function can't be called concurrently anyway, otherwise
176 * emitting the fence would mess up the hardware ring buffer.
177 */
178 rcu_assign_pointer(*ptr, dma_fence_get(&fence->base));
179
180 *f = &fence->base;
181
182 return 0;
183}
184
185/**
186 * amdgpu_fence_emit_polling - emit a fence on the requeste ring
187 *
188 * @ring: ring the fence is associated with
189 * @s: resulting sequence number
190 * @timeout: the timeout for waiting in usecs
191 *
192 * Emits a fence command on the requested ring (all asics).
193 * Used For polling fence.
194 * Returns 0 on success, -ENOMEM on failure.
195 */
196int amdgpu_fence_emit_polling(struct amdgpu_ring *ring, uint32_t *s,
197 uint32_t timeout)
198{
199 uint32_t seq;
200 signed long r;
201
202 if (!s)
203 return -EINVAL;
204
205 seq = ++ring->fence_drv.sync_seq;
206 r = amdgpu_fence_wait_polling(ring,
207 seq - ring->fence_drv.num_fences_mask,
208 timeout);
209 if (r < 1)
210 return -ETIMEDOUT;
211
212 amdgpu_ring_emit_fence(ring, ring->fence_drv.gpu_addr,
213 seq, 0);
214
215 *s = seq;
216
217 return 0;
218}
219
220/**
221 * amdgpu_fence_schedule_fallback - schedule fallback check
222 *
223 * @ring: pointer to struct amdgpu_ring
224 *
225 * Start a timer as fallback to our interrupts.
226 */
227static void amdgpu_fence_schedule_fallback(struct amdgpu_ring *ring)
228{
229 mod_timer(&ring->fence_drv.fallback_timer,
230 jiffies + AMDGPU_FENCE_JIFFIES_TIMEOUT);
231}
232
233/**
234 * amdgpu_fence_process - check for fence activity
235 *
236 * @ring: pointer to struct amdgpu_ring
237 *
238 * Checks the current fence value and calculates the last
239 * signalled fence value. Wakes the fence queue if the
240 * sequence number has increased.
241 *
242 * Returns true if fence was processed
243 */
244bool amdgpu_fence_process(struct amdgpu_ring *ring)
245{
246 struct amdgpu_fence_driver *drv = &ring->fence_drv;
247 struct amdgpu_device *adev = ring->adev;
248 uint32_t seq, last_seq;
249 int r;
250
251 do {
252 last_seq = atomic_read(&ring->fence_drv.last_seq);
253 seq = amdgpu_fence_read(ring);
254
255 } while (atomic_cmpxchg(&drv->last_seq, last_seq, seq) != last_seq);
256
257 if (del_timer(&ring->fence_drv.fallback_timer) &&
258 seq != ring->fence_drv.sync_seq)
259 amdgpu_fence_schedule_fallback(ring);
260
261 if (unlikely(seq == last_seq))
262 return false;
263
264 last_seq &= drv->num_fences_mask;
265 seq &= drv->num_fences_mask;
266
267 do {
268 struct dma_fence *fence, **ptr;
269
270 ++last_seq;
271 last_seq &= drv->num_fences_mask;
272 ptr = &drv->fences[last_seq];
273
274 /* There is always exactly one thread signaling this fence slot */
275 fence = rcu_dereference_protected(*ptr, 1);
276 RCU_INIT_POINTER(*ptr, NULL);
277
278 if (!fence)
279 continue;
280
281 r = dma_fence_signal(fence);
282 if (!r)
283 DMA_FENCE_TRACE(fence, "signaled from irq context\n");
284 else
285 BUG();
286
287 dma_fence_put(fence);
288 pm_runtime_mark_last_busy(adev_to_drm(adev)->dev);
289 pm_runtime_put_autosuspend(adev_to_drm(adev)->dev);
290 } while (last_seq != seq);
291
292 return true;
293}
294
295/**
296 * amdgpu_fence_fallback - fallback for hardware interrupts
297 *
298 * @t: timer context used to obtain the pointer to ring structure
299 *
300 * Checks for fence activity.
301 */
302static void amdgpu_fence_fallback(struct timer_list *t)
303{
304 struct amdgpu_ring *ring = from_timer(ring, t,
305 fence_drv.fallback_timer);
306
307 if (amdgpu_fence_process(ring))
308 DRM_WARN("Fence fallback timer expired on ring %s\n", ring->name);
309}
310
311/**
312 * amdgpu_fence_wait_empty - wait for all fences to signal
313 *
314 * @ring: ring index the fence is associated with
315 *
316 * Wait for all fences on the requested ring to signal (all asics).
317 * Returns 0 if the fences have passed, error for all other cases.
318 */
319int amdgpu_fence_wait_empty(struct amdgpu_ring *ring)
320{
321 uint64_t seq = READ_ONCE(ring->fence_drv.sync_seq);
322 struct dma_fence *fence, **ptr;
323 int r;
324
325 if (!seq)
326 return 0;
327
328 ptr = &ring->fence_drv.fences[seq & ring->fence_drv.num_fences_mask];
329 rcu_read_lock();
330 fence = rcu_dereference(*ptr);
331 if (!fence || !dma_fence_get_rcu(fence)) {
332 rcu_read_unlock();
333 return 0;
334 }
335 rcu_read_unlock();
336
337 r = dma_fence_wait(fence, false);
338 dma_fence_put(fence);
339 return r;
340}
341
342/**
343 * amdgpu_fence_wait_polling - busy wait for givn sequence number
344 *
345 * @ring: ring index the fence is associated with
346 * @wait_seq: sequence number to wait
347 * @timeout: the timeout for waiting in usecs
348 *
349 * Wait for all fences on the requested ring to signal (all asics).
350 * Returns left time if no timeout, 0 or minus if timeout.
351 */
352signed long amdgpu_fence_wait_polling(struct amdgpu_ring *ring,
353 uint32_t wait_seq,
354 signed long timeout)
355{
356 uint32_t seq;
357
358 do {
359 seq = amdgpu_fence_read(ring);
360 udelay(5);
361 timeout -= 5;
362 } while ((int32_t)(wait_seq - seq) > 0 && timeout > 0);
363
364 return timeout > 0 ? timeout : 0;
365}
366/**
367 * amdgpu_fence_count_emitted - get the count of emitted fences
368 *
369 * @ring: ring the fence is associated with
370 *
371 * Get the number of fences emitted on the requested ring (all asics).
372 * Returns the number of emitted fences on the ring. Used by the
373 * dynpm code to ring track activity.
374 */
375unsigned amdgpu_fence_count_emitted(struct amdgpu_ring *ring)
376{
377 uint64_t emitted;
378
379 /* We are not protected by ring lock when reading the last sequence
380 * but it's ok to report slightly wrong fence count here.
381 */
382 amdgpu_fence_process(ring);
383 emitted = 0x100000000ull;
384 emitted -= atomic_read(&ring->fence_drv.last_seq);
385 emitted += READ_ONCE(ring->fence_drv.sync_seq);
386 return lower_32_bits(emitted);
387}
388
389/**
390 * amdgpu_fence_driver_start_ring - make the fence driver
391 * ready for use on the requested ring.
392 *
393 * @ring: ring to start the fence driver on
394 * @irq_src: interrupt source to use for this ring
395 * @irq_type: interrupt type to use for this ring
396 *
397 * Make the fence driver ready for processing (all asics).
398 * Not all asics have all rings, so each asic will only
399 * start the fence driver on the rings it has.
400 * Returns 0 for success, errors for failure.
401 */
402int amdgpu_fence_driver_start_ring(struct amdgpu_ring *ring,
403 struct amdgpu_irq_src *irq_src,
404 unsigned irq_type)
405{
406 struct amdgpu_device *adev = ring->adev;
407 uint64_t index;
408
409 if (ring->funcs->type != AMDGPU_RING_TYPE_UVD) {
410 ring->fence_drv.cpu_addr = &adev->wb.wb[ring->fence_offs];
411 ring->fence_drv.gpu_addr = adev->wb.gpu_addr + (ring->fence_offs * 4);
412 } else {
413 /* put fence directly behind firmware */
414 index = ALIGN(adev->uvd.fw->size, 8);
415 ring->fence_drv.cpu_addr = adev->uvd.inst[ring->me].cpu_addr + index;
416 ring->fence_drv.gpu_addr = adev->uvd.inst[ring->me].gpu_addr + index;
417 }
418 amdgpu_fence_write(ring, atomic_read(&ring->fence_drv.last_seq));
419
420 ring->fence_drv.irq_src = irq_src;
421 ring->fence_drv.irq_type = irq_type;
422 ring->fence_drv.initialized = true;
423
424 DRM_DEV_DEBUG(adev->dev, "fence driver on ring %s use gpu addr 0x%016llx\n",
425 ring->name, ring->fence_drv.gpu_addr);
426 return 0;
427}
428
429/**
430 * amdgpu_fence_driver_init_ring - init the fence driver
431 * for the requested ring.
432 *
433 * @ring: ring to init the fence driver on
434 * @num_hw_submission: number of entries on the hardware queue
435 * @sched_score: optional score atomic shared with other schedulers
436 *
437 * Init the fence driver for the requested ring (all asics).
438 * Helper function for amdgpu_fence_driver_init().
439 */
440int amdgpu_fence_driver_init_ring(struct amdgpu_ring *ring,
441 unsigned num_hw_submission,
442 atomic_t *sched_score)
443{
444 struct amdgpu_device *adev = ring->adev;
445 long timeout;
446 int r;
447
448 if (!adev)
449 return -EINVAL;
450
451 if (!is_power_of_2(num_hw_submission))
452 return -EINVAL;
453
454 ring->fence_drv.cpu_addr = NULL;
455 ring->fence_drv.gpu_addr = 0;
456 ring->fence_drv.sync_seq = 0;
457 atomic_set(&ring->fence_drv.last_seq, 0);
458 ring->fence_drv.initialized = false;
459
460 timer_setup(&ring->fence_drv.fallback_timer, amdgpu_fence_fallback, 0);
461
462 ring->fence_drv.num_fences_mask = num_hw_submission * 2 - 1;
463 spin_lock_init(&ring->fence_drv.lock);
464 ring->fence_drv.fences = kcalloc(num_hw_submission * 2, sizeof(void *),
465 GFP_KERNEL);
466 if (!ring->fence_drv.fences)
467 return -ENOMEM;
468
469 /* No need to setup the GPU scheduler for rings that don't need it */
470 if (ring->no_scheduler)
471 return 0;
472
473 switch (ring->funcs->type) {
474 case AMDGPU_RING_TYPE_GFX:
475 timeout = adev->gfx_timeout;
476 break;
477 case AMDGPU_RING_TYPE_COMPUTE:
478 timeout = adev->compute_timeout;
479 break;
480 case AMDGPU_RING_TYPE_SDMA:
481 timeout = adev->sdma_timeout;
482 break;
483 default:
484 timeout = adev->video_timeout;
485 break;
486 }
487
488 r = drm_sched_init(&ring->sched, &amdgpu_sched_ops,
489 num_hw_submission, amdgpu_job_hang_limit,
490 timeout, sched_score, ring->name);
491 if (r) {
492 DRM_ERROR("Failed to create scheduler on ring %s.\n",
493 ring->name);
494 return r;
495 }
496
497 return 0;
498}
499
500/**
501 * amdgpu_fence_driver_sw_init - init the fence driver
502 * for all possible rings.
503 *
504 * @adev: amdgpu device pointer
505 *
506 * Init the fence driver for all possible rings (all asics).
507 * Not all asics have all rings, so each asic will only
508 * start the fence driver on the rings it has using
509 * amdgpu_fence_driver_start_ring().
510 * Returns 0 for success.
511 */
512int amdgpu_fence_driver_sw_init(struct amdgpu_device *adev)
513{
514 return 0;
515}
516
517/**
518 * amdgpu_fence_driver_hw_fini - tear down the fence driver
519 * for all possible rings.
520 *
521 * @adev: amdgpu device pointer
522 *
523 * Tear down the fence driver for all possible rings (all asics).
524 */
525void amdgpu_fence_driver_hw_fini(struct amdgpu_device *adev)
526{
527 int i, r;
528
529 for (i = 0; i < AMDGPU_MAX_RINGS; i++) {
530 struct amdgpu_ring *ring = adev->rings[i];
531
532 if (!ring || !ring->fence_drv.initialized)
533 continue;
534
535 if (!ring->no_scheduler)
536 drm_sched_stop(&ring->sched, NULL);
537
538 /* You can't wait for HW to signal if it's gone */
539 if (!drm_dev_is_unplugged(&adev->ddev))
540 r = amdgpu_fence_wait_empty(ring);
541 else
542 r = -ENODEV;
543 /* no need to trigger GPU reset as we are unloading */
544 if (r)
545 amdgpu_fence_driver_force_completion(ring);
546
547 if (ring->fence_drv.irq_src)
548 amdgpu_irq_put(adev, ring->fence_drv.irq_src,
549 ring->fence_drv.irq_type);
550
551 del_timer_sync(&ring->fence_drv.fallback_timer);
552 }
553}
554
555void amdgpu_fence_driver_sw_fini(struct amdgpu_device *adev)
556{
557 unsigned int i, j;
558
559 for (i = 0; i < AMDGPU_MAX_RINGS; i++) {
560 struct amdgpu_ring *ring = adev->rings[i];
561
562 if (!ring || !ring->fence_drv.initialized)
563 continue;
564
565 if (!ring->no_scheduler)
566 drm_sched_fini(&ring->sched);
567
568 for (j = 0; j <= ring->fence_drv.num_fences_mask; ++j)
569 dma_fence_put(ring->fence_drv.fences[j]);
570 kfree(ring->fence_drv.fences);
571 ring->fence_drv.fences = NULL;
572 ring->fence_drv.initialized = false;
573 }
574}
575
576/**
577 * amdgpu_fence_driver_hw_init - enable the fence driver
578 * for all possible rings.
579 *
580 * @adev: amdgpu device pointer
581 *
582 * Enable the fence driver for all possible rings (all asics).
583 * Not all asics have all rings, so each asic will only
584 * start the fence driver on the rings it has using
585 * amdgpu_fence_driver_start_ring().
586 * Returns 0 for success.
587 */
588void amdgpu_fence_driver_hw_init(struct amdgpu_device *adev)
589{
590 int i;
591
592 for (i = 0; i < AMDGPU_MAX_RINGS; i++) {
593 struct amdgpu_ring *ring = adev->rings[i];
594 if (!ring || !ring->fence_drv.initialized)
595 continue;
596
597 if (!ring->no_scheduler) {
598 drm_sched_resubmit_jobs(&ring->sched);
599 drm_sched_start(&ring->sched, true);
600 }
601
602 /* enable the interrupt */
603 if (ring->fence_drv.irq_src)
604 amdgpu_irq_get(adev, ring->fence_drv.irq_src,
605 ring->fence_drv.irq_type);
606 }
607}
608
609/**
610 * amdgpu_fence_driver_force_completion - force signal latest fence of ring
611 *
612 * @ring: fence of the ring to signal
613 *
614 */
615void amdgpu_fence_driver_force_completion(struct amdgpu_ring *ring)
616{
617 amdgpu_fence_write(ring, ring->fence_drv.sync_seq);
618 amdgpu_fence_process(ring);
619}
620
621/*
622 * Common fence implementation
623 */
624
625static const char *amdgpu_fence_get_driver_name(struct dma_fence *fence)
626{
627 return "amdgpu";
628}
629
630static const char *amdgpu_fence_get_timeline_name(struct dma_fence *f)
631{
632 struct amdgpu_fence *fence = to_amdgpu_fence(f);
633 return (const char *)fence->ring->name;
634}
635
636/**
637 * amdgpu_fence_enable_signaling - enable signalling on fence
638 * @f: fence
639 *
640 * This function is called with fence_queue lock held, and adds a callback
641 * to fence_queue that checks if this fence is signaled, and if so it
642 * signals the fence and removes itself.
643 */
644static bool amdgpu_fence_enable_signaling(struct dma_fence *f)
645{
646 struct amdgpu_fence *fence = to_amdgpu_fence(f);
647 struct amdgpu_ring *ring = fence->ring;
648
649 if (!timer_pending(&ring->fence_drv.fallback_timer))
650 amdgpu_fence_schedule_fallback(ring);
651
652 DMA_FENCE_TRACE(&fence->base, "armed on ring %i!\n", ring->idx);
653
654 return true;
655}
656
657/**
658 * amdgpu_fence_free - free up the fence memory
659 *
660 * @rcu: RCU callback head
661 *
662 * Free up the fence memory after the RCU grace period.
663 */
664static void amdgpu_fence_free(struct rcu_head *rcu)
665{
666 struct dma_fence *f = container_of(rcu, struct dma_fence, rcu);
667 struct amdgpu_fence *fence = to_amdgpu_fence(f);
668 kmem_cache_free(amdgpu_fence_slab, fence);
669}
670
671/**
672 * amdgpu_fence_release - callback that fence can be freed
673 *
674 * @f: fence
675 *
676 * This function is called when the reference count becomes zero.
677 * It just RCU schedules freeing up the fence.
678 */
679static void amdgpu_fence_release(struct dma_fence *f)
680{
681 call_rcu(&f->rcu, amdgpu_fence_free);
682}
683
684static const struct dma_fence_ops amdgpu_fence_ops = {
685 .get_driver_name = amdgpu_fence_get_driver_name,
686 .get_timeline_name = amdgpu_fence_get_timeline_name,
687 .enable_signaling = amdgpu_fence_enable_signaling,
688 .release = amdgpu_fence_release,
689};
690
691/*
692 * Fence debugfs
693 */
694#if defined(CONFIG_DEBUG_FS)
695static int amdgpu_debugfs_fence_info_show(struct seq_file *m, void *unused)
696{
697 struct amdgpu_device *adev = (struct amdgpu_device *)m->private;
698 int i;
699
700 for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
701 struct amdgpu_ring *ring = adev->rings[i];
702 if (!ring || !ring->fence_drv.initialized)
703 continue;
704
705 amdgpu_fence_process(ring);
706
707 seq_printf(m, "--- ring %d (%s) ---\n", i, ring->name);
708 seq_printf(m, "Last signaled fence 0x%08x\n",
709 atomic_read(&ring->fence_drv.last_seq));
710 seq_printf(m, "Last emitted 0x%08x\n",
711 ring->fence_drv.sync_seq);
712
713 if (ring->funcs->type == AMDGPU_RING_TYPE_GFX ||
714 ring->funcs->type == AMDGPU_RING_TYPE_SDMA) {
715 seq_printf(m, "Last signaled trailing fence 0x%08x\n",
716 le32_to_cpu(*ring->trail_fence_cpu_addr));
717 seq_printf(m, "Last emitted 0x%08x\n",
718 ring->trail_seq);
719 }
720
721 if (ring->funcs->type != AMDGPU_RING_TYPE_GFX)
722 continue;
723
724 /* set in CP_VMID_PREEMPT and preemption occurred */
725 seq_printf(m, "Last preempted 0x%08x\n",
726 le32_to_cpu(*(ring->fence_drv.cpu_addr + 2)));
727 /* set in CP_VMID_RESET and reset occurred */
728 seq_printf(m, "Last reset 0x%08x\n",
729 le32_to_cpu(*(ring->fence_drv.cpu_addr + 4)));
730 /* Both preemption and reset occurred */
731 seq_printf(m, "Last both 0x%08x\n",
732 le32_to_cpu(*(ring->fence_drv.cpu_addr + 6)));
733 }
734 return 0;
735}
736
737/*
738 * amdgpu_debugfs_gpu_recover - manually trigger a gpu reset & recover
739 *
740 * Manually trigger a gpu reset at the next fence wait.
741 */
742static int gpu_recover_get(void *data, u64 *val)
743{
744 struct amdgpu_device *adev = (struct amdgpu_device *)data;
745 struct drm_device *dev = adev_to_drm(adev);
746 int r;
747
748 r = pm_runtime_get_sync(dev->dev);
749 if (r < 0) {
750 pm_runtime_put_autosuspend(dev->dev);
751 return 0;
752 }
753
754 *val = amdgpu_device_gpu_recover(adev, NULL);
755
756 pm_runtime_mark_last_busy(dev->dev);
757 pm_runtime_put_autosuspend(dev->dev);
758
759 return 0;
760}
761
762DEFINE_SHOW_ATTRIBUTE(amdgpu_debugfs_fence_info);
763DEFINE_DEBUGFS_ATTRIBUTE(amdgpu_debugfs_gpu_recover_fops, gpu_recover_get, NULL,
764 "%lld\n");
765
766#endif
767
768void amdgpu_debugfs_fence_init(struct amdgpu_device *adev)
769{
770#if defined(CONFIG_DEBUG_FS)
771 struct drm_minor *minor = adev_to_drm(adev)->primary;
772 struct dentry *root = minor->debugfs_root;
773
774 debugfs_create_file("amdgpu_fence_info", 0444, root, adev,
775 &amdgpu_debugfs_fence_info_fops);
776
777 if (!amdgpu_sriov_vf(adev))
778 debugfs_create_file("amdgpu_gpu_recover", 0444, root, adev,
779 &amdgpu_debugfs_gpu_recover_fops);
780#endif
781}
782
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
38#include <drm/drm_debugfs.h>
39
40#include "amdgpu.h"
41#include "amdgpu_trace.h"
42
43/*
44 * Fences
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 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};
58
59static struct kmem_cache *amdgpu_fence_slab;
60
61int amdgpu_fence_slab_init(void)
62{
63 amdgpu_fence_slab = kmem_cache_create(
64 "amdgpu_fence", sizeof(struct amdgpu_fence), 0,
65 SLAB_HWCACHE_ALIGN, NULL);
66 if (!amdgpu_fence_slab)
67 return -ENOMEM;
68 return 0;
69}
70
71void amdgpu_fence_slab_fini(void)
72{
73 rcu_barrier();
74 kmem_cache_destroy(amdgpu_fence_slab);
75}
76/*
77 * Cast helper
78 */
79static const struct dma_fence_ops amdgpu_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 return __f;
86
87 return NULL;
88}
89
90/**
91 * amdgpu_fence_write - write a fence value
92 *
93 * @ring: ring the fence is associated with
94 * @seq: sequence number to write
95 *
96 * Writes a fence value to memory (all asics).
97 */
98static void amdgpu_fence_write(struct amdgpu_ring *ring, u32 seq)
99{
100 struct amdgpu_fence_driver *drv = &ring->fence_drv;
101
102 if (drv->cpu_addr)
103 *drv->cpu_addr = cpu_to_le32(seq);
104}
105
106/**
107 * amdgpu_fence_read - read a fence value
108 *
109 * @ring: ring the fence is associated with
110 *
111 * Reads a fence value from memory (all asics).
112 * Returns the value of the fence read from memory.
113 */
114static u32 amdgpu_fence_read(struct amdgpu_ring *ring)
115{
116 struct amdgpu_fence_driver *drv = &ring->fence_drv;
117 u32 seq = 0;
118
119 if (drv->cpu_addr)
120 seq = le32_to_cpu(*drv->cpu_addr);
121 else
122 seq = atomic_read(&drv->last_seq);
123
124 return seq;
125}
126
127/**
128 * amdgpu_fence_emit - emit a fence on the requested ring
129 *
130 * @ring: ring the fence is associated with
131 * @f: resulting fence object
132 *
133 * Emits a fence command on the requested ring (all asics).
134 * Returns 0 on success, -ENOMEM on failure.
135 */
136int amdgpu_fence_emit(struct amdgpu_ring *ring, struct dma_fence **f,
137 unsigned flags)
138{
139 struct amdgpu_device *adev = ring->adev;
140 struct amdgpu_fence *fence;
141 struct dma_fence __rcu **ptr;
142 uint32_t seq;
143 int r;
144
145 fence = kmem_cache_alloc(amdgpu_fence_slab, GFP_KERNEL);
146 if (fence == NULL)
147 return -ENOMEM;
148
149 seq = ++ring->fence_drv.sync_seq;
150 fence->ring = ring;
151 dma_fence_init(&fence->base, &amdgpu_fence_ops,
152 &ring->fence_drv.lock,
153 adev->fence_context + ring->idx,
154 seq);
155 amdgpu_ring_emit_fence(ring, ring->fence_drv.gpu_addr,
156 seq, flags | AMDGPU_FENCE_FLAG_INT);
157
158 ptr = &ring->fence_drv.fences[seq & ring->fence_drv.num_fences_mask];
159 if (unlikely(rcu_dereference_protected(*ptr, 1))) {
160 struct dma_fence *old;
161
162 rcu_read_lock();
163 old = dma_fence_get_rcu_safe(ptr);
164 rcu_read_unlock();
165
166 if (old) {
167 r = dma_fence_wait(old, false);
168 dma_fence_put(old);
169 if (r)
170 return r;
171 }
172 }
173
174 /* This function can't be called concurrently anyway, otherwise
175 * emitting the fence would mess up the hardware ring buffer.
176 */
177 rcu_assign_pointer(*ptr, dma_fence_get(&fence->base));
178
179 *f = &fence->base;
180
181 return 0;
182}
183
184/**
185 * amdgpu_fence_emit_polling - emit a fence on the requeste ring
186 *
187 * @ring: ring the fence is associated with
188 * @s: resulting sequence number
189 *
190 * Emits a fence command on the requested ring (all asics).
191 * Used For polling fence.
192 * Returns 0 on success, -ENOMEM on failure.
193 */
194int amdgpu_fence_emit_polling(struct amdgpu_ring *ring, uint32_t *s)
195{
196 uint32_t seq;
197
198 if (!s)
199 return -EINVAL;
200
201 seq = ++ring->fence_drv.sync_seq;
202 amdgpu_ring_emit_fence(ring, ring->fence_drv.gpu_addr,
203 seq, 0);
204
205 *s = seq;
206
207 return 0;
208}
209
210/**
211 * amdgpu_fence_schedule_fallback - schedule fallback check
212 *
213 * @ring: pointer to struct amdgpu_ring
214 *
215 * Start a timer as fallback to our interrupts.
216 */
217static void amdgpu_fence_schedule_fallback(struct amdgpu_ring *ring)
218{
219 mod_timer(&ring->fence_drv.fallback_timer,
220 jiffies + AMDGPU_FENCE_JIFFIES_TIMEOUT);
221}
222
223/**
224 * amdgpu_fence_process - check for fence activity
225 *
226 * @ring: pointer to struct amdgpu_ring
227 *
228 * Checks the current fence value and calculates the last
229 * signalled fence value. Wakes the fence queue if the
230 * sequence number has increased.
231 *
232 * Returns true if fence was processed
233 */
234bool amdgpu_fence_process(struct amdgpu_ring *ring)
235{
236 struct amdgpu_fence_driver *drv = &ring->fence_drv;
237 uint32_t seq, last_seq;
238 int r;
239
240 do {
241 last_seq = atomic_read(&ring->fence_drv.last_seq);
242 seq = amdgpu_fence_read(ring);
243
244 } while (atomic_cmpxchg(&drv->last_seq, last_seq, seq) != last_seq);
245
246 if (del_timer(&ring->fence_drv.fallback_timer) &&
247 seq != ring->fence_drv.sync_seq)
248 amdgpu_fence_schedule_fallback(ring);
249
250 if (unlikely(seq == last_seq))
251 return false;
252
253 last_seq &= drv->num_fences_mask;
254 seq &= drv->num_fences_mask;
255
256 do {
257 struct dma_fence *fence, **ptr;
258
259 ++last_seq;
260 last_seq &= drv->num_fences_mask;
261 ptr = &drv->fences[last_seq];
262
263 /* There is always exactly one thread signaling this fence slot */
264 fence = rcu_dereference_protected(*ptr, 1);
265 RCU_INIT_POINTER(*ptr, NULL);
266
267 if (!fence)
268 continue;
269
270 r = dma_fence_signal(fence);
271 if (!r)
272 DMA_FENCE_TRACE(fence, "signaled from irq context\n");
273 else
274 BUG();
275
276 dma_fence_put(fence);
277 } while (last_seq != seq);
278
279 return true;
280}
281
282/**
283 * amdgpu_fence_fallback - fallback for hardware interrupts
284 *
285 * @work: delayed work item
286 *
287 * Checks for fence activity.
288 */
289static void amdgpu_fence_fallback(struct timer_list *t)
290{
291 struct amdgpu_ring *ring = from_timer(ring, t,
292 fence_drv.fallback_timer);
293
294 if (amdgpu_fence_process(ring))
295 DRM_WARN("Fence fallback timer expired on ring %s\n", ring->name);
296}
297
298/**
299 * amdgpu_fence_wait_empty - wait for all fences to signal
300 *
301 * @adev: amdgpu device pointer
302 * @ring: ring index the fence is associated with
303 *
304 * Wait for all fences on the requested ring to signal (all asics).
305 * Returns 0 if the fences have passed, error for all other cases.
306 */
307int amdgpu_fence_wait_empty(struct amdgpu_ring *ring)
308{
309 uint64_t seq = READ_ONCE(ring->fence_drv.sync_seq);
310 struct dma_fence *fence, **ptr;
311 int r;
312
313 if (!seq)
314 return 0;
315
316 ptr = &ring->fence_drv.fences[seq & ring->fence_drv.num_fences_mask];
317 rcu_read_lock();
318 fence = rcu_dereference(*ptr);
319 if (!fence || !dma_fence_get_rcu(fence)) {
320 rcu_read_unlock();
321 return 0;
322 }
323 rcu_read_unlock();
324
325 r = dma_fence_wait(fence, false);
326 dma_fence_put(fence);
327 return r;
328}
329
330/**
331 * amdgpu_fence_wait_polling - busy wait for givn sequence number
332 *
333 * @ring: ring index the fence is associated with
334 * @wait_seq: sequence number to wait
335 * @timeout: the timeout for waiting in usecs
336 *
337 * Wait for all fences on the requested ring to signal (all asics).
338 * Returns left time if no timeout, 0 or minus if timeout.
339 */
340signed long amdgpu_fence_wait_polling(struct amdgpu_ring *ring,
341 uint32_t wait_seq,
342 signed long timeout)
343{
344 uint32_t seq;
345
346 do {
347 seq = amdgpu_fence_read(ring);
348 udelay(5);
349 timeout -= 5;
350 } while ((int32_t)(wait_seq - seq) > 0 && timeout > 0);
351
352 return timeout > 0 ? timeout : 0;
353}
354/**
355 * amdgpu_fence_count_emitted - get the count of emitted fences
356 *
357 * @ring: ring the fence is associated with
358 *
359 * Get the number of fences emitted on the requested ring (all asics).
360 * Returns the number of emitted fences on the ring. Used by the
361 * dynpm code to ring track activity.
362 */
363unsigned amdgpu_fence_count_emitted(struct amdgpu_ring *ring)
364{
365 uint64_t emitted;
366
367 /* We are not protected by ring lock when reading the last sequence
368 * but it's ok to report slightly wrong fence count here.
369 */
370 amdgpu_fence_process(ring);
371 emitted = 0x100000000ull;
372 emitted -= atomic_read(&ring->fence_drv.last_seq);
373 emitted += READ_ONCE(ring->fence_drv.sync_seq);
374 return lower_32_bits(emitted);
375}
376
377/**
378 * amdgpu_fence_driver_start_ring - make the fence driver
379 * ready for use on the requested ring.
380 *
381 * @ring: ring to start the fence driver on
382 * @irq_src: interrupt source to use for this ring
383 * @irq_type: interrupt type to use for this ring
384 *
385 * Make the fence driver ready for processing (all asics).
386 * Not all asics have all rings, so each asic will only
387 * start the fence driver on the rings it has.
388 * Returns 0 for success, errors for failure.
389 */
390int amdgpu_fence_driver_start_ring(struct amdgpu_ring *ring,
391 struct amdgpu_irq_src *irq_src,
392 unsigned irq_type)
393{
394 struct amdgpu_device *adev = ring->adev;
395 uint64_t index;
396
397 if (ring->funcs->type != AMDGPU_RING_TYPE_UVD) {
398 ring->fence_drv.cpu_addr = &adev->wb.wb[ring->fence_offs];
399 ring->fence_drv.gpu_addr = adev->wb.gpu_addr + (ring->fence_offs * 4);
400 } else {
401 /* put fence directly behind firmware */
402 index = ALIGN(adev->uvd.fw->size, 8);
403 ring->fence_drv.cpu_addr = adev->uvd.inst[ring->me].cpu_addr + index;
404 ring->fence_drv.gpu_addr = adev->uvd.inst[ring->me].gpu_addr + index;
405 }
406 amdgpu_fence_write(ring, atomic_read(&ring->fence_drv.last_seq));
407 amdgpu_irq_get(adev, irq_src, irq_type);
408
409 ring->fence_drv.irq_src = irq_src;
410 ring->fence_drv.irq_type = irq_type;
411 ring->fence_drv.initialized = true;
412
413 DRM_DEV_DEBUG(adev->dev, "fence driver on ring %s use gpu addr "
414 "0x%016llx, cpu addr 0x%p\n", ring->name,
415 ring->fence_drv.gpu_addr, ring->fence_drv.cpu_addr);
416 return 0;
417}
418
419/**
420 * amdgpu_fence_driver_init_ring - init the fence driver
421 * for the requested ring.
422 *
423 * @ring: ring to init the fence driver on
424 * @num_hw_submission: number of entries on the hardware queue
425 *
426 * Init the fence driver for the requested ring (all asics).
427 * Helper function for amdgpu_fence_driver_init().
428 */
429int amdgpu_fence_driver_init_ring(struct amdgpu_ring *ring,
430 unsigned num_hw_submission)
431{
432 struct amdgpu_device *adev = ring->adev;
433 long timeout;
434 int r;
435
436 if (!adev)
437 return -EINVAL;
438
439 /* Check that num_hw_submission is a power of two */
440 if ((num_hw_submission & (num_hw_submission - 1)) != 0)
441 return -EINVAL;
442
443 ring->fence_drv.cpu_addr = NULL;
444 ring->fence_drv.gpu_addr = 0;
445 ring->fence_drv.sync_seq = 0;
446 atomic_set(&ring->fence_drv.last_seq, 0);
447 ring->fence_drv.initialized = false;
448
449 timer_setup(&ring->fence_drv.fallback_timer, amdgpu_fence_fallback, 0);
450
451 ring->fence_drv.num_fences_mask = num_hw_submission * 2 - 1;
452 spin_lock_init(&ring->fence_drv.lock);
453 ring->fence_drv.fences = kcalloc(num_hw_submission * 2, sizeof(void *),
454 GFP_KERNEL);
455 if (!ring->fence_drv.fences)
456 return -ENOMEM;
457
458 /* No need to setup the GPU scheduler for KIQ ring */
459 if (ring->funcs->type != AMDGPU_RING_TYPE_KIQ) {
460 switch (ring->funcs->type) {
461 case AMDGPU_RING_TYPE_GFX:
462 timeout = adev->gfx_timeout;
463 break;
464 case AMDGPU_RING_TYPE_COMPUTE:
465 /*
466 * For non-sriov case, no timeout enforce
467 * on compute ring by default. Unless user
468 * specifies a timeout for compute ring.
469 *
470 * For sriov case, always use the timeout
471 * as gfx ring
472 */
473 if (!amdgpu_sriov_vf(ring->adev))
474 timeout = adev->compute_timeout;
475 else
476 timeout = adev->gfx_timeout;
477 break;
478 case AMDGPU_RING_TYPE_SDMA:
479 timeout = adev->sdma_timeout;
480 break;
481 default:
482 timeout = adev->video_timeout;
483 break;
484 }
485
486 r = drm_sched_init(&ring->sched, &amdgpu_sched_ops,
487 num_hw_submission, amdgpu_job_hang_limit,
488 timeout, ring->name);
489 if (r) {
490 DRM_ERROR("Failed to create scheduler on ring %s.\n",
491 ring->name);
492 return r;
493 }
494 }
495
496 return 0;
497}
498
499/**
500 * amdgpu_fence_driver_init - init the fence driver
501 * for all possible rings.
502 *
503 * @adev: amdgpu device pointer
504 *
505 * Init the fence driver for all possible rings (all asics).
506 * Not all asics have all rings, so each asic will only
507 * start the fence driver on the rings it has using
508 * amdgpu_fence_driver_start_ring().
509 * Returns 0 for success.
510 */
511int amdgpu_fence_driver_init(struct amdgpu_device *adev)
512{
513 if (amdgpu_debugfs_fence_init(adev))
514 dev_err(adev->dev, "fence debugfs file creation failed\n");
515
516 return 0;
517}
518
519/**
520 * amdgpu_fence_driver_fini - tear down the fence driver
521 * for all possible rings.
522 *
523 * @adev: amdgpu device pointer
524 *
525 * Tear down the fence driver for all possible rings (all asics).
526 */
527void amdgpu_fence_driver_fini(struct amdgpu_device *adev)
528{
529 unsigned i, j;
530 int r;
531
532 for (i = 0; i < AMDGPU_MAX_RINGS; i++) {
533 struct amdgpu_ring *ring = adev->rings[i];
534
535 if (!ring || !ring->fence_drv.initialized)
536 continue;
537 r = amdgpu_fence_wait_empty(ring);
538 if (r) {
539 /* no need to trigger GPU reset as we are unloading */
540 amdgpu_fence_driver_force_completion(ring);
541 }
542 amdgpu_irq_put(adev, ring->fence_drv.irq_src,
543 ring->fence_drv.irq_type);
544 drm_sched_fini(&ring->sched);
545 del_timer_sync(&ring->fence_drv.fallback_timer);
546 for (j = 0; j <= ring->fence_drv.num_fences_mask; ++j)
547 dma_fence_put(ring->fence_drv.fences[j]);
548 kfree(ring->fence_drv.fences);
549 ring->fence_drv.fences = NULL;
550 ring->fence_drv.initialized = false;
551 }
552}
553
554/**
555 * amdgpu_fence_driver_suspend - suspend the fence driver
556 * for all possible rings.
557 *
558 * @adev: amdgpu device pointer
559 *
560 * Suspend the fence driver for all possible rings (all asics).
561 */
562void amdgpu_fence_driver_suspend(struct amdgpu_device *adev)
563{
564 int i, r;
565
566 for (i = 0; i < AMDGPU_MAX_RINGS; i++) {
567 struct amdgpu_ring *ring = adev->rings[i];
568 if (!ring || !ring->fence_drv.initialized)
569 continue;
570
571 /* wait for gpu to finish processing current batch */
572 r = amdgpu_fence_wait_empty(ring);
573 if (r) {
574 /* delay GPU reset to resume */
575 amdgpu_fence_driver_force_completion(ring);
576 }
577
578 /* disable the interrupt */
579 amdgpu_irq_put(adev, ring->fence_drv.irq_src,
580 ring->fence_drv.irq_type);
581 }
582}
583
584/**
585 * amdgpu_fence_driver_resume - resume the fence driver
586 * for all possible rings.
587 *
588 * @adev: amdgpu device pointer
589 *
590 * Resume the fence driver for all possible rings (all asics).
591 * Not all asics have all rings, so each asic will only
592 * start the fence driver on the rings it has using
593 * amdgpu_fence_driver_start_ring().
594 * Returns 0 for success.
595 */
596void amdgpu_fence_driver_resume(struct amdgpu_device *adev)
597{
598 int i;
599
600 for (i = 0; i < AMDGPU_MAX_RINGS; i++) {
601 struct amdgpu_ring *ring = adev->rings[i];
602 if (!ring || !ring->fence_drv.initialized)
603 continue;
604
605 /* enable the interrupt */
606 amdgpu_irq_get(adev, ring->fence_drv.irq_src,
607 ring->fence_drv.irq_type);
608 }
609}
610
611/**
612 * amdgpu_fence_driver_force_completion - force signal latest fence of ring
613 *
614 * @ring: fence of the ring to signal
615 *
616 */
617void amdgpu_fence_driver_force_completion(struct amdgpu_ring *ring)
618{
619 amdgpu_fence_write(ring, ring->fence_drv.sync_seq);
620 amdgpu_fence_process(ring);
621}
622
623/*
624 * Common fence implementation
625 */
626
627static const char *amdgpu_fence_get_driver_name(struct dma_fence *fence)
628{
629 return "amdgpu";
630}
631
632static const char *amdgpu_fence_get_timeline_name(struct dma_fence *f)
633{
634 struct amdgpu_fence *fence = to_amdgpu_fence(f);
635 return (const char *)fence->ring->name;
636}
637
638/**
639 * amdgpu_fence_enable_signaling - enable signalling on fence
640 * @fence: fence
641 *
642 * This function is called with fence_queue lock held, and adds a callback
643 * to fence_queue that checks if this fence is signaled, and if so it
644 * signals the fence and removes itself.
645 */
646static bool amdgpu_fence_enable_signaling(struct dma_fence *f)
647{
648 struct amdgpu_fence *fence = to_amdgpu_fence(f);
649 struct amdgpu_ring *ring = fence->ring;
650
651 if (!timer_pending(&ring->fence_drv.fallback_timer))
652 amdgpu_fence_schedule_fallback(ring);
653
654 DMA_FENCE_TRACE(&fence->base, "armed on ring %i!\n", ring->idx);
655
656 return true;
657}
658
659/**
660 * amdgpu_fence_free - free up the fence memory
661 *
662 * @rcu: RCU callback head
663 *
664 * Free up the fence memory after the RCU grace period.
665 */
666static void amdgpu_fence_free(struct rcu_head *rcu)
667{
668 struct dma_fence *f = container_of(rcu, struct dma_fence, rcu);
669 struct amdgpu_fence *fence = to_amdgpu_fence(f);
670 kmem_cache_free(amdgpu_fence_slab, fence);
671}
672
673/**
674 * amdgpu_fence_release - callback that fence can be freed
675 *
676 * @fence: fence
677 *
678 * This function is called when the reference count becomes zero.
679 * It just RCU schedules freeing up the fence.
680 */
681static void amdgpu_fence_release(struct dma_fence *f)
682{
683 call_rcu(&f->rcu, amdgpu_fence_free);
684}
685
686static const struct dma_fence_ops amdgpu_fence_ops = {
687 .get_driver_name = amdgpu_fence_get_driver_name,
688 .get_timeline_name = amdgpu_fence_get_timeline_name,
689 .enable_signaling = amdgpu_fence_enable_signaling,
690 .release = amdgpu_fence_release,
691};
692
693/*
694 * Fence debugfs
695 */
696#if defined(CONFIG_DEBUG_FS)
697static int amdgpu_debugfs_fence_info(struct seq_file *m, void *data)
698{
699 struct drm_info_node *node = (struct drm_info_node *)m->private;
700 struct drm_device *dev = node->minor->dev;
701 struct amdgpu_device *adev = dev->dev_private;
702 int i;
703
704 for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
705 struct amdgpu_ring *ring = adev->rings[i];
706 if (!ring || !ring->fence_drv.initialized)
707 continue;
708
709 amdgpu_fence_process(ring);
710
711 seq_printf(m, "--- ring %d (%s) ---\n", i, ring->name);
712 seq_printf(m, "Last signaled fence 0x%08x\n",
713 atomic_read(&ring->fence_drv.last_seq));
714 seq_printf(m, "Last emitted 0x%08x\n",
715 ring->fence_drv.sync_seq);
716
717 if (ring->funcs->type == AMDGPU_RING_TYPE_GFX ||
718 ring->funcs->type == AMDGPU_RING_TYPE_SDMA) {
719 seq_printf(m, "Last signaled trailing fence 0x%08x\n",
720 le32_to_cpu(*ring->trail_fence_cpu_addr));
721 seq_printf(m, "Last emitted 0x%08x\n",
722 ring->trail_seq);
723 }
724
725 if (ring->funcs->type != AMDGPU_RING_TYPE_GFX)
726 continue;
727
728 /* set in CP_VMID_PREEMPT and preemption occurred */
729 seq_printf(m, "Last preempted 0x%08x\n",
730 le32_to_cpu(*(ring->fence_drv.cpu_addr + 2)));
731 /* set in CP_VMID_RESET and reset occurred */
732 seq_printf(m, "Last reset 0x%08x\n",
733 le32_to_cpu(*(ring->fence_drv.cpu_addr + 4)));
734 /* Both preemption and reset occurred */
735 seq_printf(m, "Last both 0x%08x\n",
736 le32_to_cpu(*(ring->fence_drv.cpu_addr + 6)));
737 }
738 return 0;
739}
740
741/**
742 * amdgpu_debugfs_gpu_recover - manually trigger a gpu reset & recover
743 *
744 * Manually trigger a gpu reset at the next fence wait.
745 */
746static int amdgpu_debugfs_gpu_recover(struct seq_file *m, void *data)
747{
748 struct drm_info_node *node = (struct drm_info_node *) m->private;
749 struct drm_device *dev = node->minor->dev;
750 struct amdgpu_device *adev = dev->dev_private;
751
752 seq_printf(m, "gpu recover\n");
753 amdgpu_device_gpu_recover(adev, NULL);
754
755 return 0;
756}
757
758static const struct drm_info_list amdgpu_debugfs_fence_list[] = {
759 {"amdgpu_fence_info", &amdgpu_debugfs_fence_info, 0, NULL},
760 {"amdgpu_gpu_recover", &amdgpu_debugfs_gpu_recover, 0, NULL}
761};
762
763static const struct drm_info_list amdgpu_debugfs_fence_list_sriov[] = {
764 {"amdgpu_fence_info", &amdgpu_debugfs_fence_info, 0, NULL},
765};
766#endif
767
768int amdgpu_debugfs_fence_init(struct amdgpu_device *adev)
769{
770#if defined(CONFIG_DEBUG_FS)
771 if (amdgpu_sriov_vf(adev))
772 return amdgpu_debugfs_add_files(adev, amdgpu_debugfs_fence_list_sriov, 1);
773 return amdgpu_debugfs_add_files(adev, amdgpu_debugfs_fence_list, 2);
774#else
775 return 0;
776#endif
777}
778