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 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 trace_amdgpu_runpm_reference_dumps(1, __func__);
185 ptr = &ring->fence_drv.fences[seq & ring->fence_drv.num_fences_mask];
186 if (unlikely(rcu_dereference_protected(*ptr, 1))) {
187 struct dma_fence *old;
188
189 rcu_read_lock();
190 old = dma_fence_get_rcu_safe(ptr);
191 rcu_read_unlock();
192
193 if (old) {
194 r = dma_fence_wait(old, false);
195 dma_fence_put(old);
196 if (r)
197 return r;
198 }
199 }
200
201 to_amdgpu_fence(fence)->start_timestamp = ktime_get();
202
203 /* This function can't be called concurrently anyway, otherwise
204 * emitting the fence would mess up the hardware ring buffer.
205 */
206 rcu_assign_pointer(*ptr, dma_fence_get(fence));
207
208 *f = fence;
209
210 return 0;
211}
212
213/**
214 * amdgpu_fence_emit_polling - emit a fence on the requeste ring
215 *
216 * @ring: ring the fence is associated with
217 * @s: resulting sequence number
218 * @timeout: the timeout for waiting in usecs
219 *
220 * Emits a fence command on the requested ring (all asics).
221 * Used For polling fence.
222 * Returns 0 on success, -ENOMEM on failure.
223 */
224int amdgpu_fence_emit_polling(struct amdgpu_ring *ring, uint32_t *s,
225 uint32_t timeout)
226{
227 uint32_t seq;
228 signed long r;
229
230 if (!s)
231 return -EINVAL;
232
233 seq = ++ring->fence_drv.sync_seq;
234 r = amdgpu_fence_wait_polling(ring,
235 seq - ring->fence_drv.num_fences_mask,
236 timeout);
237 if (r < 1)
238 return -ETIMEDOUT;
239
240 amdgpu_ring_emit_fence(ring, ring->fence_drv.gpu_addr,
241 seq, 0);
242
243 *s = seq;
244
245 return 0;
246}
247
248/**
249 * amdgpu_fence_schedule_fallback - schedule fallback check
250 *
251 * @ring: pointer to struct amdgpu_ring
252 *
253 * Start a timer as fallback to our interrupts.
254 */
255static void amdgpu_fence_schedule_fallback(struct amdgpu_ring *ring)
256{
257 mod_timer(&ring->fence_drv.fallback_timer,
258 jiffies + AMDGPU_FENCE_JIFFIES_TIMEOUT);
259}
260
261/**
262 * amdgpu_fence_process - check for fence activity
263 *
264 * @ring: pointer to struct amdgpu_ring
265 *
266 * Checks the current fence value and calculates the last
267 * signalled fence value. Wakes the fence queue if the
268 * sequence number has increased.
269 *
270 * Returns true if fence was processed
271 */
272bool amdgpu_fence_process(struct amdgpu_ring *ring)
273{
274 struct amdgpu_fence_driver *drv = &ring->fence_drv;
275 struct amdgpu_device *adev = ring->adev;
276 uint32_t seq, last_seq;
277
278 do {
279 last_seq = atomic_read(&ring->fence_drv.last_seq);
280 seq = amdgpu_fence_read(ring);
281
282 } while (atomic_cmpxchg(&drv->last_seq, last_seq, seq) != last_seq);
283
284 if (del_timer(&ring->fence_drv.fallback_timer) &&
285 seq != ring->fence_drv.sync_seq)
286 amdgpu_fence_schedule_fallback(ring);
287
288 if (unlikely(seq == last_seq))
289 return false;
290
291 last_seq &= drv->num_fences_mask;
292 seq &= drv->num_fences_mask;
293
294 do {
295 struct dma_fence *fence, **ptr;
296
297 ++last_seq;
298 last_seq &= drv->num_fences_mask;
299 ptr = &drv->fences[last_seq];
300
301 /* There is always exactly one thread signaling this fence slot */
302 fence = rcu_dereference_protected(*ptr, 1);
303 RCU_INIT_POINTER(*ptr, NULL);
304
305 if (!fence)
306 continue;
307
308 dma_fence_signal(fence);
309 dma_fence_put(fence);
310 pm_runtime_mark_last_busy(adev_to_drm(adev)->dev);
311 pm_runtime_put_autosuspend(adev_to_drm(adev)->dev);
312 trace_amdgpu_runpm_reference_dumps(0, __func__);
313 } while (last_seq != seq);
314
315 return true;
316}
317
318/**
319 * amdgpu_fence_fallback - fallback for hardware interrupts
320 *
321 * @t: timer context used to obtain the pointer to ring structure
322 *
323 * Checks for fence activity.
324 */
325static void amdgpu_fence_fallback(struct timer_list *t)
326{
327 struct amdgpu_ring *ring = from_timer(ring, t,
328 fence_drv.fallback_timer);
329
330 if (amdgpu_fence_process(ring))
331 DRM_WARN("Fence fallback timer expired on ring %s\n", ring->name);
332}
333
334/**
335 * amdgpu_fence_wait_empty - wait for all fences to signal
336 *
337 * @ring: ring index the fence is associated with
338 *
339 * Wait for all fences on the requested ring to signal (all asics).
340 * Returns 0 if the fences have passed, error for all other cases.
341 */
342int amdgpu_fence_wait_empty(struct amdgpu_ring *ring)
343{
344 uint64_t seq = READ_ONCE(ring->fence_drv.sync_seq);
345 struct dma_fence *fence, **ptr;
346 int r;
347
348 if (!seq)
349 return 0;
350
351 ptr = &ring->fence_drv.fences[seq & ring->fence_drv.num_fences_mask];
352 rcu_read_lock();
353 fence = rcu_dereference(*ptr);
354 if (!fence || !dma_fence_get_rcu(fence)) {
355 rcu_read_unlock();
356 return 0;
357 }
358 rcu_read_unlock();
359
360 r = dma_fence_wait(fence, false);
361 dma_fence_put(fence);
362 return r;
363}
364
365/**
366 * amdgpu_fence_wait_polling - busy wait for givn sequence number
367 *
368 * @ring: ring index the fence is associated with
369 * @wait_seq: sequence number to wait
370 * @timeout: the timeout for waiting in usecs
371 *
372 * Wait for all fences on the requested ring to signal (all asics).
373 * Returns left time if no timeout, 0 or minus if timeout.
374 */
375signed long amdgpu_fence_wait_polling(struct amdgpu_ring *ring,
376 uint32_t wait_seq,
377 signed long timeout)
378{
379
380 while ((int32_t)(wait_seq - amdgpu_fence_read(ring)) > 0 && timeout > 0) {
381 udelay(2);
382 timeout -= 2;
383 }
384 return timeout > 0 ? timeout : 0;
385}
386/**
387 * amdgpu_fence_count_emitted - get the count of emitted fences
388 *
389 * @ring: ring the fence is associated with
390 *
391 * Get the number of fences emitted on the requested ring (all asics).
392 * Returns the number of emitted fences on the ring. Used by the
393 * dynpm code to ring track activity.
394 */
395unsigned int amdgpu_fence_count_emitted(struct amdgpu_ring *ring)
396{
397 uint64_t emitted;
398
399 /* We are not protected by ring lock when reading the last sequence
400 * but it's ok to report slightly wrong fence count here.
401 */
402 emitted = 0x100000000ull;
403 emitted -= atomic_read(&ring->fence_drv.last_seq);
404 emitted += READ_ONCE(ring->fence_drv.sync_seq);
405 return lower_32_bits(emitted);
406}
407
408/**
409 * amdgpu_fence_last_unsignaled_time_us - the time fence emitted until now
410 * @ring: ring the fence is associated with
411 *
412 * Find the earliest fence unsignaled until now, calculate the time delta
413 * between the time fence emitted and now.
414 */
415u64 amdgpu_fence_last_unsignaled_time_us(struct amdgpu_ring *ring)
416{
417 struct amdgpu_fence_driver *drv = &ring->fence_drv;
418 struct dma_fence *fence;
419 uint32_t last_seq, sync_seq;
420
421 last_seq = atomic_read(&ring->fence_drv.last_seq);
422 sync_seq = READ_ONCE(ring->fence_drv.sync_seq);
423 if (last_seq == sync_seq)
424 return 0;
425
426 ++last_seq;
427 last_seq &= drv->num_fences_mask;
428 fence = drv->fences[last_seq];
429 if (!fence)
430 return 0;
431
432 return ktime_us_delta(ktime_get(),
433 to_amdgpu_fence(fence)->start_timestamp);
434}
435
436/**
437 * amdgpu_fence_update_start_timestamp - update the timestamp of the fence
438 * @ring: ring the fence is associated with
439 * @seq: the fence seq number to update.
440 * @timestamp: the start timestamp to update.
441 *
442 * The function called at the time the fence and related ib is about to
443 * resubmit to gpu in MCBP scenario. Thus we do not consider race condition
444 * with amdgpu_fence_process to modify the same fence.
445 */
446void amdgpu_fence_update_start_timestamp(struct amdgpu_ring *ring, uint32_t seq, ktime_t timestamp)
447{
448 struct amdgpu_fence_driver *drv = &ring->fence_drv;
449 struct dma_fence *fence;
450
451 seq &= drv->num_fences_mask;
452 fence = drv->fences[seq];
453 if (!fence)
454 return;
455
456 to_amdgpu_fence(fence)->start_timestamp = timestamp;
457}
458
459/**
460 * amdgpu_fence_driver_start_ring - make the fence driver
461 * ready for use on the requested ring.
462 *
463 * @ring: ring to start the fence driver on
464 * @irq_src: interrupt source to use for this ring
465 * @irq_type: interrupt type to use for this ring
466 *
467 * Make the fence driver ready for processing (all asics).
468 * Not all asics have all rings, so each asic will only
469 * start the fence driver on the rings it has.
470 * Returns 0 for success, errors for failure.
471 */
472int amdgpu_fence_driver_start_ring(struct amdgpu_ring *ring,
473 struct amdgpu_irq_src *irq_src,
474 unsigned int irq_type)
475{
476 struct amdgpu_device *adev = ring->adev;
477 uint64_t index;
478
479 if (ring->funcs->type != AMDGPU_RING_TYPE_UVD) {
480 ring->fence_drv.cpu_addr = ring->fence_cpu_addr;
481 ring->fence_drv.gpu_addr = ring->fence_gpu_addr;
482 } else {
483 /* put fence directly behind firmware */
484 index = ALIGN(adev->uvd.fw->size, 8);
485 ring->fence_drv.cpu_addr = adev->uvd.inst[ring->me].cpu_addr + index;
486 ring->fence_drv.gpu_addr = adev->uvd.inst[ring->me].gpu_addr + index;
487 }
488 amdgpu_fence_write(ring, atomic_read(&ring->fence_drv.last_seq));
489
490 ring->fence_drv.irq_src = irq_src;
491 ring->fence_drv.irq_type = irq_type;
492 ring->fence_drv.initialized = true;
493
494 DRM_DEV_DEBUG(adev->dev, "fence driver on ring %s use gpu addr 0x%016llx\n",
495 ring->name, ring->fence_drv.gpu_addr);
496 return 0;
497}
498
499/**
500 * amdgpu_fence_driver_init_ring - init the fence driver
501 * for the requested ring.
502 *
503 * @ring: ring to init the fence driver on
504 *
505 * Init the fence driver for the requested ring (all asics).
506 * Helper function for amdgpu_fence_driver_init().
507 */
508int amdgpu_fence_driver_init_ring(struct amdgpu_ring *ring)
509{
510 struct amdgpu_device *adev = ring->adev;
511
512 if (!adev)
513 return -EINVAL;
514
515 if (!is_power_of_2(ring->num_hw_submission))
516 return -EINVAL;
517
518 ring->fence_drv.cpu_addr = NULL;
519 ring->fence_drv.gpu_addr = 0;
520 ring->fence_drv.sync_seq = 0;
521 atomic_set(&ring->fence_drv.last_seq, 0);
522 ring->fence_drv.initialized = false;
523
524 timer_setup(&ring->fence_drv.fallback_timer, amdgpu_fence_fallback, 0);
525
526 ring->fence_drv.num_fences_mask = ring->num_hw_submission * 2 - 1;
527 spin_lock_init(&ring->fence_drv.lock);
528 ring->fence_drv.fences = kcalloc(ring->num_hw_submission * 2, sizeof(void *),
529 GFP_KERNEL);
530
531 if (!ring->fence_drv.fences)
532 return -ENOMEM;
533
534 return 0;
535}
536
537/**
538 * amdgpu_fence_driver_sw_init - init the fence driver
539 * for all possible rings.
540 *
541 * @adev: amdgpu device pointer
542 *
543 * Init the fence driver for all possible rings (all asics).
544 * Not all asics have all rings, so each asic will only
545 * start the fence driver on the rings it has using
546 * amdgpu_fence_driver_start_ring().
547 * Returns 0 for success.
548 */
549int amdgpu_fence_driver_sw_init(struct amdgpu_device *adev)
550{
551 return 0;
552}
553
554/**
555 * amdgpu_fence_need_ring_interrupt_restore - helper function to check whether
556 * fence driver interrupts need to be restored.
557 *
558 * @ring: ring that to be checked
559 *
560 * Interrupts for rings that belong to GFX IP don't need to be restored
561 * when the target power state is s0ix.
562 *
563 * Return true if need to restore interrupts, false otherwise.
564 */
565static bool amdgpu_fence_need_ring_interrupt_restore(struct amdgpu_ring *ring)
566{
567 struct amdgpu_device *adev = ring->adev;
568 bool is_gfx_power_domain = false;
569
570 switch (ring->funcs->type) {
571 case AMDGPU_RING_TYPE_SDMA:
572 /* SDMA 5.x+ is part of GFX power domain so it's covered by GFXOFF */
573 if (amdgpu_ip_version(adev, SDMA0_HWIP, 0) >=
574 IP_VERSION(5, 0, 0))
575 is_gfx_power_domain = true;
576 break;
577 case AMDGPU_RING_TYPE_GFX:
578 case AMDGPU_RING_TYPE_COMPUTE:
579 case AMDGPU_RING_TYPE_KIQ:
580 case AMDGPU_RING_TYPE_MES:
581 is_gfx_power_domain = true;
582 break;
583 default:
584 break;
585 }
586
587 return !(adev->in_s0ix && is_gfx_power_domain);
588}
589
590/**
591 * amdgpu_fence_driver_hw_fini - tear down the fence driver
592 * for all possible rings.
593 *
594 * @adev: amdgpu device pointer
595 *
596 * Tear down the fence driver for all possible rings (all asics).
597 */
598void amdgpu_fence_driver_hw_fini(struct amdgpu_device *adev)
599{
600 int i, r;
601
602 for (i = 0; i < AMDGPU_MAX_RINGS; i++) {
603 struct amdgpu_ring *ring = adev->rings[i];
604
605 if (!ring || !ring->fence_drv.initialized)
606 continue;
607
608 /* You can't wait for HW to signal if it's gone */
609 if (!drm_dev_is_unplugged(adev_to_drm(adev)))
610 r = amdgpu_fence_wait_empty(ring);
611 else
612 r = -ENODEV;
613 /* no need to trigger GPU reset as we are unloading */
614 if (r)
615 amdgpu_fence_driver_force_completion(ring);
616
617 if (!drm_dev_is_unplugged(adev_to_drm(adev)) &&
618 ring->fence_drv.irq_src &&
619 amdgpu_fence_need_ring_interrupt_restore(ring))
620 amdgpu_irq_put(adev, ring->fence_drv.irq_src,
621 ring->fence_drv.irq_type);
622
623 del_timer_sync(&ring->fence_drv.fallback_timer);
624 }
625}
626
627/* Will either stop and flush handlers for amdgpu interrupt or reanble it */
628void amdgpu_fence_driver_isr_toggle(struct amdgpu_device *adev, bool stop)
629{
630 int i;
631
632 for (i = 0; i < AMDGPU_MAX_RINGS; i++) {
633 struct amdgpu_ring *ring = adev->rings[i];
634
635 if (!ring || !ring->fence_drv.initialized || !ring->fence_drv.irq_src)
636 continue;
637
638 if (stop)
639 disable_irq(adev->irq.irq);
640 else
641 enable_irq(adev->irq.irq);
642 }
643}
644
645void amdgpu_fence_driver_sw_fini(struct amdgpu_device *adev)
646{
647 unsigned int i, j;
648
649 for (i = 0; i < AMDGPU_MAX_RINGS; i++) {
650 struct amdgpu_ring *ring = adev->rings[i];
651
652 if (!ring || !ring->fence_drv.initialized)
653 continue;
654
655 /*
656 * Notice we check for sched.ops since there's some
657 * override on the meaning of sched.ready by amdgpu.
658 * The natural check would be sched.ready, which is
659 * set as drm_sched_init() finishes...
660 */
661 if (ring->sched.ops)
662 drm_sched_fini(&ring->sched);
663
664 for (j = 0; j <= ring->fence_drv.num_fences_mask; ++j)
665 dma_fence_put(ring->fence_drv.fences[j]);
666 kfree(ring->fence_drv.fences);
667 ring->fence_drv.fences = NULL;
668 ring->fence_drv.initialized = false;
669 }
670}
671
672/**
673 * amdgpu_fence_driver_hw_init - enable the fence driver
674 * for all possible rings.
675 *
676 * @adev: amdgpu device pointer
677 *
678 * Enable the fence driver for all possible rings (all asics).
679 * Not all asics have all rings, so each asic will only
680 * start the fence driver on the rings it has using
681 * amdgpu_fence_driver_start_ring().
682 * Returns 0 for success.
683 */
684void amdgpu_fence_driver_hw_init(struct amdgpu_device *adev)
685{
686 int i;
687
688 for (i = 0; i < AMDGPU_MAX_RINGS; i++) {
689 struct amdgpu_ring *ring = adev->rings[i];
690
691 if (!ring || !ring->fence_drv.initialized)
692 continue;
693
694 /* enable the interrupt */
695 if (ring->fence_drv.irq_src &&
696 amdgpu_fence_need_ring_interrupt_restore(ring))
697 amdgpu_irq_get(adev, ring->fence_drv.irq_src,
698 ring->fence_drv.irq_type);
699 }
700}
701
702/**
703 * amdgpu_fence_driver_clear_job_fences - clear job embedded fences of ring
704 *
705 * @ring: fence of the ring to be cleared
706 *
707 */
708void amdgpu_fence_driver_clear_job_fences(struct amdgpu_ring *ring)
709{
710 int i;
711 struct dma_fence *old, **ptr;
712
713 for (i = 0; i <= ring->fence_drv.num_fences_mask; i++) {
714 ptr = &ring->fence_drv.fences[i];
715 old = rcu_dereference_protected(*ptr, 1);
716 if (old && old->ops == &amdgpu_job_fence_ops) {
717 struct amdgpu_job *job;
718
719 /* For non-scheduler bad job, i.e. failed ib test, we need to signal
720 * it right here or we won't be able to track them in fence_drv
721 * and they will remain unsignaled during sa_bo free.
722 */
723 job = container_of(old, struct amdgpu_job, hw_fence);
724 if (!job->base.s_fence && !dma_fence_is_signaled(old))
725 dma_fence_signal(old);
726 RCU_INIT_POINTER(*ptr, NULL);
727 dma_fence_put(old);
728 }
729 }
730}
731
732/**
733 * amdgpu_fence_driver_set_error - set error code on fences
734 * @ring: the ring which contains the fences
735 * @error: the error code to set
736 *
737 * Set an error code to all the fences pending on the ring.
738 */
739void amdgpu_fence_driver_set_error(struct amdgpu_ring *ring, int error)
740{
741 struct amdgpu_fence_driver *drv = &ring->fence_drv;
742 unsigned long flags;
743
744 spin_lock_irqsave(&drv->lock, flags);
745 for (unsigned int i = 0; i <= drv->num_fences_mask; ++i) {
746 struct dma_fence *fence;
747
748 fence = rcu_dereference_protected(drv->fences[i],
749 lockdep_is_held(&drv->lock));
750 if (fence && !dma_fence_is_signaled_locked(fence))
751 dma_fence_set_error(fence, error);
752 }
753 spin_unlock_irqrestore(&drv->lock, flags);
754}
755
756/**
757 * amdgpu_fence_driver_force_completion - force signal latest fence of ring
758 *
759 * @ring: fence of the ring to signal
760 *
761 */
762void amdgpu_fence_driver_force_completion(struct amdgpu_ring *ring)
763{
764 amdgpu_fence_driver_set_error(ring, -ECANCELED);
765 amdgpu_fence_write(ring, ring->fence_drv.sync_seq);
766 amdgpu_fence_process(ring);
767}
768
769/*
770 * Common fence implementation
771 */
772
773static const char *amdgpu_fence_get_driver_name(struct dma_fence *fence)
774{
775 return "amdgpu";
776}
777
778static const char *amdgpu_fence_get_timeline_name(struct dma_fence *f)
779{
780 return (const char *)to_amdgpu_fence(f)->ring->name;
781}
782
783static const char *amdgpu_job_fence_get_timeline_name(struct dma_fence *f)
784{
785 struct amdgpu_job *job = container_of(f, struct amdgpu_job, hw_fence);
786
787 return (const char *)to_amdgpu_ring(job->base.sched)->name;
788}
789
790/**
791 * amdgpu_fence_enable_signaling - enable signalling on fence
792 * @f: fence
793 *
794 * This function is called with fence_queue lock held, and adds a callback
795 * to fence_queue that checks if this fence is signaled, and if so it
796 * signals the fence and removes itself.
797 */
798static bool amdgpu_fence_enable_signaling(struct dma_fence *f)
799{
800 if (!timer_pending(&to_amdgpu_fence(f)->ring->fence_drv.fallback_timer))
801 amdgpu_fence_schedule_fallback(to_amdgpu_fence(f)->ring);
802
803 return true;
804}
805
806/**
807 * amdgpu_job_fence_enable_signaling - enable signalling on job fence
808 * @f: fence
809 *
810 * This is the simliar function with amdgpu_fence_enable_signaling above, it
811 * only handles the job embedded fence.
812 */
813static bool amdgpu_job_fence_enable_signaling(struct dma_fence *f)
814{
815 struct amdgpu_job *job = container_of(f, struct amdgpu_job, hw_fence);
816
817 if (!timer_pending(&to_amdgpu_ring(job->base.sched)->fence_drv.fallback_timer))
818 amdgpu_fence_schedule_fallback(to_amdgpu_ring(job->base.sched));
819
820 return true;
821}
822
823/**
824 * amdgpu_fence_free - free up the fence memory
825 *
826 * @rcu: RCU callback head
827 *
828 * Free up the fence memory after the RCU grace period.
829 */
830static void amdgpu_fence_free(struct rcu_head *rcu)
831{
832 struct dma_fence *f = container_of(rcu, struct dma_fence, rcu);
833
834 /* free fence_slab if it's separated fence*/
835 kmem_cache_free(amdgpu_fence_slab, to_amdgpu_fence(f));
836}
837
838/**
839 * amdgpu_job_fence_free - free up the job with embedded fence
840 *
841 * @rcu: RCU callback head
842 *
843 * Free up the job with embedded fence after the RCU grace period.
844 */
845static void amdgpu_job_fence_free(struct rcu_head *rcu)
846{
847 struct dma_fence *f = container_of(rcu, struct dma_fence, rcu);
848
849 /* free job if fence has a parent job */
850 kfree(container_of(f, struct amdgpu_job, hw_fence));
851}
852
853/**
854 * amdgpu_fence_release - callback that fence can be freed
855 *
856 * @f: fence
857 *
858 * This function is called when the reference count becomes zero.
859 * It just RCU schedules freeing up the fence.
860 */
861static void amdgpu_fence_release(struct dma_fence *f)
862{
863 call_rcu(&f->rcu, amdgpu_fence_free);
864}
865
866/**
867 * amdgpu_job_fence_release - callback that job embedded fence can be freed
868 *
869 * @f: fence
870 *
871 * This is the simliar function with amdgpu_fence_release above, it
872 * only handles the job embedded fence.
873 */
874static void amdgpu_job_fence_release(struct dma_fence *f)
875{
876 call_rcu(&f->rcu, amdgpu_job_fence_free);
877}
878
879static const struct dma_fence_ops amdgpu_fence_ops = {
880 .get_driver_name = amdgpu_fence_get_driver_name,
881 .get_timeline_name = amdgpu_fence_get_timeline_name,
882 .enable_signaling = amdgpu_fence_enable_signaling,
883 .release = amdgpu_fence_release,
884};
885
886static const struct dma_fence_ops amdgpu_job_fence_ops = {
887 .get_driver_name = amdgpu_fence_get_driver_name,
888 .get_timeline_name = amdgpu_job_fence_get_timeline_name,
889 .enable_signaling = amdgpu_job_fence_enable_signaling,
890 .release = amdgpu_job_fence_release,
891};
892
893/*
894 * Fence debugfs
895 */
896#if defined(CONFIG_DEBUG_FS)
897static int amdgpu_debugfs_fence_info_show(struct seq_file *m, void *unused)
898{
899 struct amdgpu_device *adev = m->private;
900 int i;
901
902 for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
903 struct amdgpu_ring *ring = adev->rings[i];
904
905 if (!ring || !ring->fence_drv.initialized)
906 continue;
907
908 amdgpu_fence_process(ring);
909
910 seq_printf(m, "--- ring %d (%s) ---\n", i, ring->name);
911 seq_printf(m, "Last signaled fence 0x%08x\n",
912 atomic_read(&ring->fence_drv.last_seq));
913 seq_printf(m, "Last emitted 0x%08x\n",
914 ring->fence_drv.sync_seq);
915
916 if (ring->funcs->type == AMDGPU_RING_TYPE_GFX ||
917 ring->funcs->type == AMDGPU_RING_TYPE_SDMA) {
918 seq_printf(m, "Last signaled trailing fence 0x%08x\n",
919 le32_to_cpu(*ring->trail_fence_cpu_addr));
920 seq_printf(m, "Last emitted 0x%08x\n",
921 ring->trail_seq);
922 }
923
924 if (ring->funcs->type != AMDGPU_RING_TYPE_GFX)
925 continue;
926
927 /* set in CP_VMID_PREEMPT and preemption occurred */
928 seq_printf(m, "Last preempted 0x%08x\n",
929 le32_to_cpu(*(ring->fence_drv.cpu_addr + 2)));
930 /* set in CP_VMID_RESET and reset occurred */
931 seq_printf(m, "Last reset 0x%08x\n",
932 le32_to_cpu(*(ring->fence_drv.cpu_addr + 4)));
933 /* Both preemption and reset occurred */
934 seq_printf(m, "Last both 0x%08x\n",
935 le32_to_cpu(*(ring->fence_drv.cpu_addr + 6)));
936 }
937 return 0;
938}
939
940/*
941 * amdgpu_debugfs_gpu_recover - manually trigger a gpu reset & recover
942 *
943 * Manually trigger a gpu reset at the next fence wait.
944 */
945static int gpu_recover_get(void *data, u64 *val)
946{
947 struct amdgpu_device *adev = (struct amdgpu_device *)data;
948 struct drm_device *dev = adev_to_drm(adev);
949 int r;
950
951 r = pm_runtime_get_sync(dev->dev);
952 if (r < 0) {
953 pm_runtime_put_autosuspend(dev->dev);
954 return 0;
955 }
956
957 if (amdgpu_reset_domain_schedule(adev->reset_domain, &adev->reset_work))
958 flush_work(&adev->reset_work);
959
960 *val = atomic_read(&adev->reset_domain->reset_res);
961
962 pm_runtime_mark_last_busy(dev->dev);
963 pm_runtime_put_autosuspend(dev->dev);
964
965 return 0;
966}
967
968DEFINE_SHOW_ATTRIBUTE(amdgpu_debugfs_fence_info);
969DEFINE_DEBUGFS_ATTRIBUTE(amdgpu_debugfs_gpu_recover_fops, gpu_recover_get, NULL,
970 "%lld\n");
971
972static void amdgpu_debugfs_reset_work(struct work_struct *work)
973{
974 struct amdgpu_device *adev = container_of(work, struct amdgpu_device,
975 reset_work);
976
977 struct amdgpu_reset_context reset_context;
978
979 memset(&reset_context, 0, sizeof(reset_context));
980
981 reset_context.method = AMD_RESET_METHOD_NONE;
982 reset_context.reset_req_dev = adev;
983 set_bit(AMDGPU_NEED_FULL_RESET, &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
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_create(
65 "amdgpu_fence", sizeof(struct amdgpu_fence), 0,
66 SLAB_HWCACHE_ALIGN, NULL);
67 if (!amdgpu_fence_slab)
68 return -ENOMEM;
69 return 0;
70}
71
72void amdgpu_fence_slab_fini(void)
73{
74 rcu_barrier();
75 kmem_cache_destroy(amdgpu_fence_slab);
76}
77/*
78 * Cast helper
79 */
80static const struct dma_fence_ops amdgpu_fence_ops;
81static const struct dma_fence_ops amdgpu_job_fence_ops;
82static inline struct amdgpu_fence *to_amdgpu_fence(struct dma_fence *f)
83{
84 struct amdgpu_fence *__f = container_of(f, struct amdgpu_fence, base);
85
86 if (__f->base.ops == &amdgpu_fence_ops ||
87 __f->base.ops == &amdgpu_job_fence_ops)
88 return __f;
89
90 return NULL;
91}
92
93/**
94 * amdgpu_fence_write - write a fence value
95 *
96 * @ring: ring the fence is associated with
97 * @seq: sequence number to write
98 *
99 * Writes a fence value to memory (all asics).
100 */
101static void amdgpu_fence_write(struct amdgpu_ring *ring, u32 seq)
102{
103 struct amdgpu_fence_driver *drv = &ring->fence_drv;
104
105 if (drv->cpu_addr)
106 *drv->cpu_addr = cpu_to_le32(seq);
107}
108
109/**
110 * amdgpu_fence_read - read a fence value
111 *
112 * @ring: ring the fence is associated with
113 *
114 * Reads a fence value from memory (all asics).
115 * Returns the value of the fence read from memory.
116 */
117static u32 amdgpu_fence_read(struct amdgpu_ring *ring)
118{
119 struct amdgpu_fence_driver *drv = &ring->fence_drv;
120 u32 seq = 0;
121
122 if (drv->cpu_addr)
123 seq = le32_to_cpu(*drv->cpu_addr);
124 else
125 seq = atomic_read(&drv->last_seq);
126
127 return seq;
128}
129
130/**
131 * amdgpu_fence_emit - emit a fence on the requested ring
132 *
133 * @ring: ring the fence is associated with
134 * @f: resulting fence object
135 * @job: job the fence is embedded in
136 * @flags: flags to pass into the subordinate .emit_fence() call
137 *
138 * Emits a fence command on the requested ring (all asics).
139 * Returns 0 on success, -ENOMEM on failure.
140 */
141int amdgpu_fence_emit(struct amdgpu_ring *ring, struct dma_fence **f, struct amdgpu_job *job,
142 unsigned int flags)
143{
144 struct amdgpu_device *adev = ring->adev;
145 struct dma_fence *fence;
146 struct amdgpu_fence *am_fence;
147 struct dma_fence __rcu **ptr;
148 uint32_t seq;
149 int r;
150
151 if (job == NULL) {
152 /* create a sperate hw fence */
153 am_fence = kmem_cache_alloc(amdgpu_fence_slab, GFP_ATOMIC);
154 if (am_fence == NULL)
155 return -ENOMEM;
156 fence = &am_fence->base;
157 am_fence->ring = ring;
158 } else {
159 /* take use of job-embedded fence */
160 fence = &job->hw_fence;
161 }
162
163 seq = ++ring->fence_drv.sync_seq;
164 if (job && job->job_run_counter) {
165 /* reinit seq for resubmitted jobs */
166 fence->seqno = seq;
167 /* TO be inline with external fence creation and other drivers */
168 dma_fence_get(fence);
169 } else {
170 if (job) {
171 dma_fence_init(fence, &amdgpu_job_fence_ops,
172 &ring->fence_drv.lock,
173 adev->fence_context + ring->idx, seq);
174 /* Against remove in amdgpu_job_{free, free_cb} */
175 dma_fence_get(fence);
176 } else {
177 dma_fence_init(fence, &amdgpu_fence_ops,
178 &ring->fence_drv.lock,
179 adev->fence_context + ring->idx, seq);
180 }
181 }
182
183 amdgpu_ring_emit_fence(ring, ring->fence_drv.gpu_addr,
184 seq, flags | AMDGPU_FENCE_FLAG_INT);
185 pm_runtime_get_noresume(adev_to_drm(adev)->dev);
186 trace_amdgpu_runpm_reference_dumps(1, __func__);
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 trace_amdgpu_runpm_reference_dumps(0, __func__);
315 } while (last_seq != seq);
316
317 return true;
318}
319
320/**
321 * amdgpu_fence_fallback - fallback for hardware interrupts
322 *
323 * @t: timer context used to obtain the pointer to ring structure
324 *
325 * Checks for fence activity.
326 */
327static void amdgpu_fence_fallback(struct timer_list *t)
328{
329 struct amdgpu_ring *ring = from_timer(ring, t,
330 fence_drv.fallback_timer);
331
332 if (amdgpu_fence_process(ring))
333 DRM_WARN("Fence fallback timer expired on ring %s\n", ring->name);
334}
335
336/**
337 * amdgpu_fence_wait_empty - wait for all fences to signal
338 *
339 * @ring: ring index the fence is associated with
340 *
341 * Wait for all fences on the requested ring to signal (all asics).
342 * Returns 0 if the fences have passed, error for all other cases.
343 */
344int amdgpu_fence_wait_empty(struct amdgpu_ring *ring)
345{
346 uint64_t seq = READ_ONCE(ring->fence_drv.sync_seq);
347 struct dma_fence *fence, **ptr;
348 int r;
349
350 if (!seq)
351 return 0;
352
353 ptr = &ring->fence_drv.fences[seq & ring->fence_drv.num_fences_mask];
354 rcu_read_lock();
355 fence = rcu_dereference(*ptr);
356 if (!fence || !dma_fence_get_rcu(fence)) {
357 rcu_read_unlock();
358 return 0;
359 }
360 rcu_read_unlock();
361
362 r = dma_fence_wait(fence, false);
363 dma_fence_put(fence);
364 return r;
365}
366
367/**
368 * amdgpu_fence_wait_polling - busy wait for givn sequence number
369 *
370 * @ring: ring index the fence is associated with
371 * @wait_seq: sequence number to wait
372 * @timeout: the timeout for waiting in usecs
373 *
374 * Wait for all fences on the requested ring to signal (all asics).
375 * Returns left time if no timeout, 0 or minus if timeout.
376 */
377signed long amdgpu_fence_wait_polling(struct amdgpu_ring *ring,
378 uint32_t wait_seq,
379 signed long timeout)
380{
381
382 while ((int32_t)(wait_seq - amdgpu_fence_read(ring)) > 0 && timeout > 0) {
383 udelay(2);
384 timeout -= 2;
385 }
386 return timeout > 0 ? timeout : 0;
387}
388/**
389 * amdgpu_fence_count_emitted - get the count of emitted fences
390 *
391 * @ring: ring the fence is associated with
392 *
393 * Get the number of fences emitted on the requested ring (all asics).
394 * Returns the number of emitted fences on the ring. Used by the
395 * dynpm code to ring track activity.
396 */
397unsigned int amdgpu_fence_count_emitted(struct amdgpu_ring *ring)
398{
399 uint64_t emitted;
400
401 /* We are not protected by ring lock when reading the last sequence
402 * but it's ok to report slightly wrong fence count here.
403 */
404 emitted = 0x100000000ull;
405 emitted -= atomic_read(&ring->fence_drv.last_seq);
406 emitted += READ_ONCE(ring->fence_drv.sync_seq);
407 return lower_32_bits(emitted);
408}
409
410/**
411 * amdgpu_fence_last_unsignaled_time_us - the time fence emitted until now
412 * @ring: ring the fence is associated with
413 *
414 * Find the earliest fence unsignaled until now, calculate the time delta
415 * between the time fence emitted and now.
416 */
417u64 amdgpu_fence_last_unsignaled_time_us(struct amdgpu_ring *ring)
418{
419 struct amdgpu_fence_driver *drv = &ring->fence_drv;
420 struct dma_fence *fence;
421 uint32_t last_seq, sync_seq;
422
423 last_seq = atomic_read(&ring->fence_drv.last_seq);
424 sync_seq = READ_ONCE(ring->fence_drv.sync_seq);
425 if (last_seq == sync_seq)
426 return 0;
427
428 ++last_seq;
429 last_seq &= drv->num_fences_mask;
430 fence = drv->fences[last_seq];
431 if (!fence)
432 return 0;
433
434 return ktime_us_delta(ktime_get(),
435 to_amdgpu_fence(fence)->start_timestamp);
436}
437
438/**
439 * amdgpu_fence_update_start_timestamp - update the timestamp of the fence
440 * @ring: ring the fence is associated with
441 * @seq: the fence seq number to update.
442 * @timestamp: the start timestamp to update.
443 *
444 * The function called at the time the fence and related ib is about to
445 * resubmit to gpu in MCBP scenario. Thus we do not consider race condition
446 * with amdgpu_fence_process to modify the same fence.
447 */
448void amdgpu_fence_update_start_timestamp(struct amdgpu_ring *ring, uint32_t seq, ktime_t timestamp)
449{
450 struct amdgpu_fence_driver *drv = &ring->fence_drv;
451 struct dma_fence *fence;
452
453 seq &= drv->num_fences_mask;
454 fence = drv->fences[seq];
455 if (!fence)
456 return;
457
458 to_amdgpu_fence(fence)->start_timestamp = timestamp;
459}
460
461/**
462 * amdgpu_fence_driver_start_ring - make the fence driver
463 * ready for use on the requested ring.
464 *
465 * @ring: ring to start the fence driver on
466 * @irq_src: interrupt source to use for this ring
467 * @irq_type: interrupt type to use for this ring
468 *
469 * Make the fence driver ready for processing (all asics).
470 * Not all asics have all rings, so each asic will only
471 * start the fence driver on the rings it has.
472 * Returns 0 for success, errors for failure.
473 */
474int amdgpu_fence_driver_start_ring(struct amdgpu_ring *ring,
475 struct amdgpu_irq_src *irq_src,
476 unsigned int irq_type)
477{
478 struct amdgpu_device *adev = ring->adev;
479 uint64_t index;
480
481 if (ring->funcs->type != AMDGPU_RING_TYPE_UVD) {
482 ring->fence_drv.cpu_addr = ring->fence_cpu_addr;
483 ring->fence_drv.gpu_addr = ring->fence_gpu_addr;
484 } else {
485 /* put fence directly behind firmware */
486 index = ALIGN(adev->uvd.fw->size, 8);
487 ring->fence_drv.cpu_addr = adev->uvd.inst[ring->me].cpu_addr + index;
488 ring->fence_drv.gpu_addr = adev->uvd.inst[ring->me].gpu_addr + index;
489 }
490 amdgpu_fence_write(ring, atomic_read(&ring->fence_drv.last_seq));
491
492 ring->fence_drv.irq_src = irq_src;
493 ring->fence_drv.irq_type = irq_type;
494 ring->fence_drv.initialized = true;
495
496 DRM_DEV_DEBUG(adev->dev, "fence driver on ring %s use gpu addr 0x%016llx\n",
497 ring->name, ring->fence_drv.gpu_addr);
498 return 0;
499}
500
501/**
502 * amdgpu_fence_driver_init_ring - init the fence driver
503 * for the requested ring.
504 *
505 * @ring: ring to init the fence driver on
506 *
507 * Init the fence driver for the requested ring (all asics).
508 * Helper function for amdgpu_fence_driver_init().
509 */
510int amdgpu_fence_driver_init_ring(struct amdgpu_ring *ring)
511{
512 struct amdgpu_device *adev = ring->adev;
513
514 if (!adev)
515 return -EINVAL;
516
517 if (!is_power_of_2(ring->num_hw_submission))
518 return -EINVAL;
519
520 ring->fence_drv.cpu_addr = NULL;
521 ring->fence_drv.gpu_addr = 0;
522 ring->fence_drv.sync_seq = 0;
523 atomic_set(&ring->fence_drv.last_seq, 0);
524 ring->fence_drv.initialized = false;
525
526 timer_setup(&ring->fence_drv.fallback_timer, amdgpu_fence_fallback, 0);
527
528 ring->fence_drv.num_fences_mask = ring->num_hw_submission * 2 - 1;
529 spin_lock_init(&ring->fence_drv.lock);
530 ring->fence_drv.fences = kcalloc(ring->num_hw_submission * 2, sizeof(void *),
531 GFP_KERNEL);
532
533 if (!ring->fence_drv.fences)
534 return -ENOMEM;
535
536 return 0;
537}
538
539/**
540 * amdgpu_fence_driver_sw_init - init the fence driver
541 * for all possible rings.
542 *
543 * @adev: amdgpu device pointer
544 *
545 * Init the fence driver for all possible rings (all asics).
546 * Not all asics have all rings, so each asic will only
547 * start the fence driver on the rings it has using
548 * amdgpu_fence_driver_start_ring().
549 * Returns 0 for success.
550 */
551int amdgpu_fence_driver_sw_init(struct amdgpu_device *adev)
552{
553 return 0;
554}
555
556/**
557 * amdgpu_fence_need_ring_interrupt_restore - helper function to check whether
558 * fence driver interrupts need to be restored.
559 *
560 * @ring: ring that to be checked
561 *
562 * Interrupts for rings that belong to GFX IP don't need to be restored
563 * when the target power state is s0ix.
564 *
565 * Return true if need to restore interrupts, false otherwise.
566 */
567static bool amdgpu_fence_need_ring_interrupt_restore(struct amdgpu_ring *ring)
568{
569 struct amdgpu_device *adev = ring->adev;
570 bool is_gfx_power_domain = false;
571
572 switch (ring->funcs->type) {
573 case AMDGPU_RING_TYPE_SDMA:
574 /* SDMA 5.x+ is part of GFX power domain so it's covered by GFXOFF */
575 if (amdgpu_ip_version(adev, SDMA0_HWIP, 0) >=
576 IP_VERSION(5, 0, 0))
577 is_gfx_power_domain = true;
578 break;
579 case AMDGPU_RING_TYPE_GFX:
580 case AMDGPU_RING_TYPE_COMPUTE:
581 case AMDGPU_RING_TYPE_KIQ:
582 case AMDGPU_RING_TYPE_MES:
583 is_gfx_power_domain = true;
584 break;
585 default:
586 break;
587 }
588
589 return !(adev->in_s0ix && is_gfx_power_domain);
590}
591
592/**
593 * amdgpu_fence_driver_hw_fini - tear down the fence driver
594 * for all possible rings.
595 *
596 * @adev: amdgpu device pointer
597 *
598 * Tear down the fence driver for all possible rings (all asics).
599 */
600void amdgpu_fence_driver_hw_fini(struct amdgpu_device *adev)
601{
602 int i, r;
603
604 for (i = 0; i < AMDGPU_MAX_RINGS; i++) {
605 struct amdgpu_ring *ring = adev->rings[i];
606
607 if (!ring || !ring->fence_drv.initialized)
608 continue;
609
610 /* You can't wait for HW to signal if it's gone */
611 if (!drm_dev_is_unplugged(adev_to_drm(adev)))
612 r = amdgpu_fence_wait_empty(ring);
613 else
614 r = -ENODEV;
615 /* no need to trigger GPU reset as we are unloading */
616 if (r)
617 amdgpu_fence_driver_force_completion(ring);
618
619 if (!drm_dev_is_unplugged(adev_to_drm(adev)) &&
620 ring->fence_drv.irq_src &&
621 amdgpu_fence_need_ring_interrupt_restore(ring))
622 amdgpu_irq_put(adev, ring->fence_drv.irq_src,
623 ring->fence_drv.irq_type);
624
625 del_timer_sync(&ring->fence_drv.fallback_timer);
626 }
627}
628
629/* Will either stop and flush handlers for amdgpu interrupt or reanble it */
630void amdgpu_fence_driver_isr_toggle(struct amdgpu_device *adev, bool stop)
631{
632 int i;
633
634 for (i = 0; i < AMDGPU_MAX_RINGS; i++) {
635 struct amdgpu_ring *ring = adev->rings[i];
636
637 if (!ring || !ring->fence_drv.initialized || !ring->fence_drv.irq_src)
638 continue;
639
640 if (stop)
641 disable_irq(adev->irq.irq);
642 else
643 enable_irq(adev->irq.irq);
644 }
645}
646
647void amdgpu_fence_driver_sw_fini(struct amdgpu_device *adev)
648{
649 unsigned int i, j;
650
651 for (i = 0; i < AMDGPU_MAX_RINGS; i++) {
652 struct amdgpu_ring *ring = adev->rings[i];
653
654 if (!ring || !ring->fence_drv.initialized)
655 continue;
656
657 /*
658 * Notice we check for sched.ops since there's some
659 * override on the meaning of sched.ready by amdgpu.
660 * The natural check would be sched.ready, which is
661 * set as drm_sched_init() finishes...
662 */
663 if (ring->sched.ops)
664 drm_sched_fini(&ring->sched);
665
666 for (j = 0; j <= ring->fence_drv.num_fences_mask; ++j)
667 dma_fence_put(ring->fence_drv.fences[j]);
668 kfree(ring->fence_drv.fences);
669 ring->fence_drv.fences = NULL;
670 ring->fence_drv.initialized = false;
671 }
672}
673
674/**
675 * amdgpu_fence_driver_hw_init - enable the fence driver
676 * for all possible rings.
677 *
678 * @adev: amdgpu device pointer
679 *
680 * Enable the fence driver for all possible rings (all asics).
681 * Not all asics have all rings, so each asic will only
682 * start the fence driver on the rings it has using
683 * amdgpu_fence_driver_start_ring().
684 * Returns 0 for success.
685 */
686void amdgpu_fence_driver_hw_init(struct amdgpu_device *adev)
687{
688 int i;
689
690 for (i = 0; i < AMDGPU_MAX_RINGS; i++) {
691 struct amdgpu_ring *ring = adev->rings[i];
692
693 if (!ring || !ring->fence_drv.initialized)
694 continue;
695
696 /* enable the interrupt */
697 if (ring->fence_drv.irq_src &&
698 amdgpu_fence_need_ring_interrupt_restore(ring))
699 amdgpu_irq_get(adev, ring->fence_drv.irq_src,
700 ring->fence_drv.irq_type);
701 }
702}
703
704/**
705 * amdgpu_fence_driver_clear_job_fences - clear job embedded fences of ring
706 *
707 * @ring: fence of the ring to be cleared
708 *
709 */
710void amdgpu_fence_driver_clear_job_fences(struct amdgpu_ring *ring)
711{
712 int i;
713 struct dma_fence *old, **ptr;
714
715 for (i = 0; i <= ring->fence_drv.num_fences_mask; i++) {
716 ptr = &ring->fence_drv.fences[i];
717 old = rcu_dereference_protected(*ptr, 1);
718 if (old && old->ops == &amdgpu_job_fence_ops) {
719 struct amdgpu_job *job;
720
721 /* For non-scheduler bad job, i.e. failed ib test, we need to signal
722 * it right here or we won't be able to track them in fence_drv
723 * and they will remain unsignaled during sa_bo free.
724 */
725 job = container_of(old, struct amdgpu_job, hw_fence);
726 if (!job->base.s_fence && !dma_fence_is_signaled(old))
727 dma_fence_signal(old);
728 RCU_INIT_POINTER(*ptr, NULL);
729 dma_fence_put(old);
730 }
731 }
732}
733
734/**
735 * amdgpu_fence_driver_set_error - set error code on fences
736 * @ring: the ring which contains the fences
737 * @error: the error code to set
738 *
739 * Set an error code to all the fences pending on the ring.
740 */
741void amdgpu_fence_driver_set_error(struct amdgpu_ring *ring, int error)
742{
743 struct amdgpu_fence_driver *drv = &ring->fence_drv;
744 unsigned long flags;
745
746 spin_lock_irqsave(&drv->lock, flags);
747 for (unsigned int i = 0; i <= drv->num_fences_mask; ++i) {
748 struct dma_fence *fence;
749
750 fence = rcu_dereference_protected(drv->fences[i],
751 lockdep_is_held(&drv->lock));
752 if (fence && !dma_fence_is_signaled_locked(fence))
753 dma_fence_set_error(fence, error);
754 }
755 spin_unlock_irqrestore(&drv->lock, flags);
756}
757
758/**
759 * amdgpu_fence_driver_force_completion - force signal latest fence of ring
760 *
761 * @ring: fence of the ring to signal
762 *
763 */
764void amdgpu_fence_driver_force_completion(struct amdgpu_ring *ring)
765{
766 amdgpu_fence_driver_set_error(ring, -ECANCELED);
767 amdgpu_fence_write(ring, ring->fence_drv.sync_seq);
768 amdgpu_fence_process(ring);
769}
770
771/*
772 * Common fence implementation
773 */
774
775static const char *amdgpu_fence_get_driver_name(struct dma_fence *fence)
776{
777 return "amdgpu";
778}
779
780static const char *amdgpu_fence_get_timeline_name(struct dma_fence *f)
781{
782 return (const char *)to_amdgpu_fence(f)->ring->name;
783}
784
785static const char *amdgpu_job_fence_get_timeline_name(struct dma_fence *f)
786{
787 struct amdgpu_job *job = container_of(f, struct amdgpu_job, hw_fence);
788
789 return (const char *)to_amdgpu_ring(job->base.sched)->name;
790}
791
792/**
793 * amdgpu_fence_enable_signaling - enable signalling on fence
794 * @f: fence
795 *
796 * This function is called with fence_queue lock held, and adds a callback
797 * to fence_queue that checks if this fence is signaled, and if so it
798 * signals the fence and removes itself.
799 */
800static bool amdgpu_fence_enable_signaling(struct dma_fence *f)
801{
802 if (!timer_pending(&to_amdgpu_fence(f)->ring->fence_drv.fallback_timer))
803 amdgpu_fence_schedule_fallback(to_amdgpu_fence(f)->ring);
804
805 return true;
806}
807
808/**
809 * amdgpu_job_fence_enable_signaling - enable signalling on job fence
810 * @f: fence
811 *
812 * This is the simliar function with amdgpu_fence_enable_signaling above, it
813 * only handles the job embedded fence.
814 */
815static bool amdgpu_job_fence_enable_signaling(struct dma_fence *f)
816{
817 struct amdgpu_job *job = container_of(f, struct amdgpu_job, hw_fence);
818
819 if (!timer_pending(&to_amdgpu_ring(job->base.sched)->fence_drv.fallback_timer))
820 amdgpu_fence_schedule_fallback(to_amdgpu_ring(job->base.sched));
821
822 return true;
823}
824
825/**
826 * amdgpu_fence_free - free up the fence memory
827 *
828 * @rcu: RCU callback head
829 *
830 * Free up the fence memory after the RCU grace period.
831 */
832static void amdgpu_fence_free(struct rcu_head *rcu)
833{
834 struct dma_fence *f = container_of(rcu, struct dma_fence, rcu);
835
836 /* free fence_slab if it's separated fence*/
837 kmem_cache_free(amdgpu_fence_slab, to_amdgpu_fence(f));
838}
839
840/**
841 * amdgpu_job_fence_free - free up the job with embedded fence
842 *
843 * @rcu: RCU callback head
844 *
845 * Free up the job with embedded fence after the RCU grace period.
846 */
847static void amdgpu_job_fence_free(struct rcu_head *rcu)
848{
849 struct dma_fence *f = container_of(rcu, struct dma_fence, rcu);
850
851 /* free job if fence has a parent job */
852 kfree(container_of(f, struct amdgpu_job, hw_fence));
853}
854
855/**
856 * amdgpu_fence_release - callback that fence can be freed
857 *
858 * @f: fence
859 *
860 * This function is called when the reference count becomes zero.
861 * It just RCU schedules freeing up the fence.
862 */
863static void amdgpu_fence_release(struct dma_fence *f)
864{
865 call_rcu(&f->rcu, amdgpu_fence_free);
866}
867
868/**
869 * amdgpu_job_fence_release - callback that job embedded fence can be freed
870 *
871 * @f: fence
872 *
873 * This is the simliar function with amdgpu_fence_release above, it
874 * only handles the job embedded fence.
875 */
876static void amdgpu_job_fence_release(struct dma_fence *f)
877{
878 call_rcu(&f->rcu, amdgpu_job_fence_free);
879}
880
881static const struct dma_fence_ops amdgpu_fence_ops = {
882 .get_driver_name = amdgpu_fence_get_driver_name,
883 .get_timeline_name = amdgpu_fence_get_timeline_name,
884 .enable_signaling = amdgpu_fence_enable_signaling,
885 .release = amdgpu_fence_release,
886};
887
888static const struct dma_fence_ops amdgpu_job_fence_ops = {
889 .get_driver_name = amdgpu_fence_get_driver_name,
890 .get_timeline_name = amdgpu_job_fence_get_timeline_name,
891 .enable_signaling = amdgpu_job_fence_enable_signaling,
892 .release = amdgpu_job_fence_release,
893};
894
895/*
896 * Fence debugfs
897 */
898#if defined(CONFIG_DEBUG_FS)
899static int amdgpu_debugfs_fence_info_show(struct seq_file *m, void *unused)
900{
901 struct amdgpu_device *adev = m->private;
902 int i;
903
904 for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
905 struct amdgpu_ring *ring = adev->rings[i];
906
907 if (!ring || !ring->fence_drv.initialized)
908 continue;
909
910 amdgpu_fence_process(ring);
911
912 seq_printf(m, "--- ring %d (%s) ---\n", i, ring->name);
913 seq_printf(m, "Last signaled fence 0x%08x\n",
914 atomic_read(&ring->fence_drv.last_seq));
915 seq_printf(m, "Last emitted 0x%08x\n",
916 ring->fence_drv.sync_seq);
917
918 if (ring->funcs->type == AMDGPU_RING_TYPE_GFX ||
919 ring->funcs->type == AMDGPU_RING_TYPE_SDMA) {
920 seq_printf(m, "Last signaled trailing fence 0x%08x\n",
921 le32_to_cpu(*ring->trail_fence_cpu_addr));
922 seq_printf(m, "Last emitted 0x%08x\n",
923 ring->trail_seq);
924 }
925
926 if (ring->funcs->type != AMDGPU_RING_TYPE_GFX)
927 continue;
928
929 /* set in CP_VMID_PREEMPT and preemption occurred */
930 seq_printf(m, "Last preempted 0x%08x\n",
931 le32_to_cpu(*(ring->fence_drv.cpu_addr + 2)));
932 /* set in CP_VMID_RESET and reset occurred */
933 seq_printf(m, "Last reset 0x%08x\n",
934 le32_to_cpu(*(ring->fence_drv.cpu_addr + 4)));
935 /* Both preemption and reset occurred */
936 seq_printf(m, "Last both 0x%08x\n",
937 le32_to_cpu(*(ring->fence_drv.cpu_addr + 6)));
938 }
939 return 0;
940}
941
942/*
943 * amdgpu_debugfs_gpu_recover - manually trigger a gpu reset & recover
944 *
945 * Manually trigger a gpu reset at the next fence wait.
946 */
947static int gpu_recover_get(void *data, u64 *val)
948{
949 struct amdgpu_device *adev = (struct amdgpu_device *)data;
950 struct drm_device *dev = adev_to_drm(adev);
951 int r;
952
953 r = pm_runtime_get_sync(dev->dev);
954 if (r < 0) {
955 pm_runtime_put_autosuspend(dev->dev);
956 return 0;
957 }
958
959 if (amdgpu_reset_domain_schedule(adev->reset_domain, &adev->reset_work))
960 flush_work(&adev->reset_work);
961
962 *val = atomic_read(&adev->reset_domain->reset_res);
963
964 pm_runtime_mark_last_busy(dev->dev);
965 pm_runtime_put_autosuspend(dev->dev);
966
967 return 0;
968}
969
970DEFINE_SHOW_ATTRIBUTE(amdgpu_debugfs_fence_info);
971DEFINE_DEBUGFS_ATTRIBUTE(amdgpu_debugfs_gpu_recover_fops, gpu_recover_get, NULL,
972 "%lld\n");
973
974static void amdgpu_debugfs_reset_work(struct work_struct *work)
975{
976 struct amdgpu_device *adev = container_of(work, struct amdgpu_device,
977 reset_work);
978
979 struct amdgpu_reset_context reset_context;
980
981 memset(&reset_context, 0, sizeof(reset_context));
982
983 reset_context.method = AMD_RESET_METHOD_NONE;
984 reset_context.reset_req_dev = adev;
985 set_bit(AMDGPU_NEED_FULL_RESET, &reset_context.flags);
986
987 amdgpu_device_gpu_recover(adev, NULL, &reset_context);
988}
989
990#endif
991
992void amdgpu_debugfs_fence_init(struct amdgpu_device *adev)
993{
994#if defined(CONFIG_DEBUG_FS)
995 struct drm_minor *minor = adev_to_drm(adev)->primary;
996 struct dentry *root = minor->debugfs_root;
997
998 debugfs_create_file("amdgpu_fence_info", 0444, root, adev,
999 &amdgpu_debugfs_fence_info_fops);
1000
1001 if (!amdgpu_sriov_vf(adev)) {
1002
1003 INIT_WORK(&adev->reset_work, amdgpu_debugfs_reset_work);
1004 debugfs_create_file("amdgpu_gpu_recover", 0444, root, adev,
1005 &amdgpu_debugfs_gpu_recover_fops);
1006 }
1007#endif
1008}
1009