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
32#include <linux/atomic.h>
33#include <linux/debugfs.h>
34#include <linux/firmware.h>
35#include <linux/kref.h>
36#include <linux/sched/signal.h>
37#include <linux/seq_file.h>
38#include <linux/slab.h>
39#include <linux/wait.h>
40
41#include <drm/drm_device.h>
42#include <drm/drm_file.h>
43
44#include "radeon.h"
45#include "radeon_reg.h"
46#include "radeon_trace.h"
47
48/*
49 * Fences mark an event in the GPUs pipeline and are used
50 * for GPU/CPU synchronization. When the fence is written,
51 * it is expected that all buffers associated with that fence
52 * are no longer in use by the associated ring on the GPU and
53 * that the relevant GPU caches have been flushed. Whether
54 * we use a scratch register or memory location depends on the asic
55 * and whether writeback is enabled.
56 */
57
58/**
59 * radeon_fence_write - write a fence value
60 *
61 * @rdev: radeon_device pointer
62 * @seq: sequence number to write
63 * @ring: ring index the fence is associated with
64 *
65 * Writes a fence value to memory or a scratch register (all asics).
66 */
67static void radeon_fence_write(struct radeon_device *rdev, u32 seq, int ring)
68{
69 struct radeon_fence_driver *drv = &rdev->fence_drv[ring];
70
71 if (likely(rdev->wb.enabled || !drv->scratch_reg)) {
72 if (drv->cpu_addr)
73 *drv->cpu_addr = cpu_to_le32(seq);
74 } else {
75 WREG32(drv->scratch_reg, seq);
76 }
77}
78
79/**
80 * radeon_fence_read - read a fence value
81 *
82 * @rdev: radeon_device pointer
83 * @ring: ring index the fence is associated with
84 *
85 * Reads a fence value from memory or a scratch register (all asics).
86 * Returns the value of the fence read from memory or register.
87 */
88static u32 radeon_fence_read(struct radeon_device *rdev, int ring)
89{
90 struct radeon_fence_driver *drv = &rdev->fence_drv[ring];
91 u32 seq = 0;
92
93 if (likely(rdev->wb.enabled || !drv->scratch_reg)) {
94 if (drv->cpu_addr)
95 seq = le32_to_cpu(*drv->cpu_addr);
96 else
97 seq = lower_32_bits(atomic64_read(&drv->last_seq));
98 } else {
99 seq = RREG32(drv->scratch_reg);
100 }
101 return seq;
102}
103
104/**
105 * radeon_fence_schedule_check - schedule lockup check
106 *
107 * @rdev: radeon_device pointer
108 * @ring: ring index we should work with
109 *
110 * Queues a delayed work item to check for lockups.
111 */
112static void radeon_fence_schedule_check(struct radeon_device *rdev, int ring)
113{
114 /*
115 * Do not reset the timer here with mod_delayed_work,
116 * this can livelock in an interaction with TTM delayed destroy.
117 */
118 queue_delayed_work(system_power_efficient_wq,
119 &rdev->fence_drv[ring].lockup_work,
120 RADEON_FENCE_JIFFIES_TIMEOUT);
121}
122
123/**
124 * radeon_fence_emit - emit a fence on the requested ring
125 *
126 * @rdev: radeon_device pointer
127 * @fence: radeon fence object
128 * @ring: ring index the fence is associated with
129 *
130 * Emits a fence command on the requested ring (all asics).
131 * Returns 0 on success, -ENOMEM on failure.
132 */
133int radeon_fence_emit(struct radeon_device *rdev,
134 struct radeon_fence **fence,
135 int ring)
136{
137 u64 seq;
138
139 /* we are protected by the ring emission mutex */
140 *fence = kmalloc(sizeof(struct radeon_fence), GFP_KERNEL);
141 if ((*fence) == NULL)
142 return -ENOMEM;
143
144 (*fence)->rdev = rdev;
145 (*fence)->seq = seq = ++rdev->fence_drv[ring].sync_seq[ring];
146 (*fence)->ring = ring;
147 (*fence)->is_vm_update = false;
148 dma_fence_init(&(*fence)->base, &radeon_fence_ops,
149 &rdev->fence_queue.lock,
150 rdev->fence_context + ring,
151 seq);
152 radeon_fence_ring_emit(rdev, ring, *fence);
153 trace_radeon_fence_emit(rdev_to_drm(rdev), ring, (*fence)->seq);
154 radeon_fence_schedule_check(rdev, ring);
155 return 0;
156}
157
158/*
159 * radeon_fence_check_signaled - callback from fence_queue
160 *
161 * this function is called with fence_queue lock held, which is also used
162 * for the fence locking itself, so unlocked variants are used for
163 * fence_signal, and remove_wait_queue.
164 */
165static int radeon_fence_check_signaled(wait_queue_entry_t *wait,
166 unsigned int mode, int flags, void *key)
167{
168 struct radeon_fence *fence;
169 u64 seq;
170
171 fence = container_of(wait, struct radeon_fence, fence_wake);
172
173 /*
174 * We cannot use radeon_fence_process here because we're already
175 * in the waitqueue, in a call from wake_up_all.
176 */
177 seq = atomic64_read(&fence->rdev->fence_drv[fence->ring].last_seq);
178 if (seq >= fence->seq) {
179 dma_fence_signal_locked(&fence->base);
180 radeon_irq_kms_sw_irq_put(fence->rdev, fence->ring);
181 __remove_wait_queue(&fence->rdev->fence_queue, &fence->fence_wake);
182 dma_fence_put(&fence->base);
183 }
184 return 0;
185}
186
187/**
188 * radeon_fence_activity - check for fence activity
189 *
190 * @rdev: radeon_device pointer
191 * @ring: ring index the fence is associated with
192 *
193 * Checks the current fence value and calculates the last
194 * signalled fence value. Returns true if activity occured
195 * on the ring, and the fence_queue should be waken up.
196 */
197static bool radeon_fence_activity(struct radeon_device *rdev, int ring)
198{
199 uint64_t seq, last_seq, last_emitted;
200 unsigned int count_loop = 0;
201 bool wake = false;
202
203 /* Note there is a scenario here for an infinite loop but it's
204 * very unlikely to happen. For it to happen, the current polling
205 * process need to be interrupted by another process and another
206 * process needs to update the last_seq btw the atomic read and
207 * xchg of the current process.
208 *
209 * More over for this to go in infinite loop there need to be
210 * continuously new fence signaled ie radeon_fence_read needs
211 * to return a different value each time for both the currently
212 * polling process and the other process that xchg the last_seq
213 * btw atomic read and xchg of the current process. And the
214 * value the other process set as last seq must be higher than
215 * the seq value we just read. Which means that current process
216 * need to be interrupted after radeon_fence_read and before
217 * atomic xchg.
218 *
219 * To be even more safe we count the number of time we loop and
220 * we bail after 10 loop just accepting the fact that we might
221 * have temporarly set the last_seq not to the true real last
222 * seq but to an older one.
223 */
224 last_seq = atomic64_read(&rdev->fence_drv[ring].last_seq);
225 do {
226 last_emitted = rdev->fence_drv[ring].sync_seq[ring];
227 seq = radeon_fence_read(rdev, ring);
228 seq |= last_seq & 0xffffffff00000000LL;
229 if (seq < last_seq) {
230 seq &= 0xffffffff;
231 seq |= last_emitted & 0xffffffff00000000LL;
232 }
233
234 if (seq <= last_seq || seq > last_emitted)
235 break;
236
237 /* If we loop over we don't want to return without
238 * checking if a fence is signaled as it means that the
239 * seq we just read is different from the previous on.
240 */
241 wake = true;
242 last_seq = seq;
243 if ((count_loop++) > 10) {
244 /* We looped over too many time leave with the
245 * fact that we might have set an older fence
246 * seq then the current real last seq as signaled
247 * by the hw.
248 */
249 break;
250 }
251 } while (atomic64_xchg(&rdev->fence_drv[ring].last_seq, seq) > seq);
252
253 if (seq < last_emitted)
254 radeon_fence_schedule_check(rdev, ring);
255
256 return wake;
257}
258
259/**
260 * radeon_fence_check_lockup - check for hardware lockup
261 *
262 * @work: delayed work item
263 *
264 * Checks for fence activity and if there is none probe
265 * the hardware if a lockup occured.
266 */
267static void radeon_fence_check_lockup(struct work_struct *work)
268{
269 struct radeon_fence_driver *fence_drv;
270 struct radeon_device *rdev;
271 int ring;
272
273 fence_drv = container_of(work, struct radeon_fence_driver,
274 lockup_work.work);
275 rdev = fence_drv->rdev;
276 ring = fence_drv - &rdev->fence_drv[0];
277
278 if (!down_read_trylock(&rdev->exclusive_lock)) {
279 /* just reschedule the check if a reset is going on */
280 radeon_fence_schedule_check(rdev, ring);
281 return;
282 }
283
284 if (fence_drv->delayed_irq && rdev->irq.installed) {
285 unsigned long irqflags;
286
287 fence_drv->delayed_irq = false;
288 spin_lock_irqsave(&rdev->irq.lock, irqflags);
289 radeon_irq_set(rdev);
290 spin_unlock_irqrestore(&rdev->irq.lock, irqflags);
291 }
292
293 if (radeon_fence_activity(rdev, ring))
294 wake_up_all(&rdev->fence_queue);
295
296 else if (radeon_ring_is_lockup(rdev, ring, &rdev->ring[ring])) {
297
298 /* good news we believe it's a lockup */
299 dev_warn(rdev->dev, "GPU lockup (current fence id 0x%016llx last fence id 0x%016llx on ring %d)\n",
300 (uint64_t)atomic64_read(&fence_drv->last_seq),
301 fence_drv->sync_seq[ring], ring);
302
303 /* remember that we need an reset */
304 rdev->needs_reset = true;
305 wake_up_all(&rdev->fence_queue);
306 }
307 up_read(&rdev->exclusive_lock);
308}
309
310/**
311 * radeon_fence_process - process a fence
312 *
313 * @rdev: radeon_device pointer
314 * @ring: ring index the fence is associated with
315 *
316 * Checks the current fence value and wakes the fence queue
317 * if the sequence number has increased (all asics).
318 */
319void radeon_fence_process(struct radeon_device *rdev, int ring)
320{
321 if (radeon_fence_activity(rdev, ring))
322 wake_up_all(&rdev->fence_queue);
323}
324
325/**
326 * radeon_fence_seq_signaled - check if a fence sequence number has signaled
327 *
328 * @rdev: radeon device pointer
329 * @seq: sequence number
330 * @ring: ring index the fence is associated with
331 *
332 * Check if the last signaled fence sequnce number is >= the requested
333 * sequence number (all asics).
334 * Returns true if the fence has signaled (current fence value
335 * is >= requested value) or false if it has not (current fence
336 * value is < the requested value. Helper function for
337 * radeon_fence_signaled().
338 */
339static bool radeon_fence_seq_signaled(struct radeon_device *rdev,
340 u64 seq, unsigned int ring)
341{
342 if (atomic64_read(&rdev->fence_drv[ring].last_seq) >= seq)
343 return true;
344
345 /* poll new last sequence at least once */
346 radeon_fence_process(rdev, ring);
347 if (atomic64_read(&rdev->fence_drv[ring].last_seq) >= seq)
348 return true;
349
350 return false;
351}
352
353static bool radeon_fence_is_signaled(struct dma_fence *f)
354{
355 struct radeon_fence *fence = to_radeon_fence(f);
356 struct radeon_device *rdev = fence->rdev;
357 unsigned int ring = fence->ring;
358 u64 seq = fence->seq;
359
360 if (atomic64_read(&rdev->fence_drv[ring].last_seq) >= seq)
361 return true;
362
363 if (down_read_trylock(&rdev->exclusive_lock)) {
364 radeon_fence_process(rdev, ring);
365 up_read(&rdev->exclusive_lock);
366
367 if (atomic64_read(&rdev->fence_drv[ring].last_seq) >= seq)
368 return true;
369 }
370 return false;
371}
372
373/**
374 * radeon_fence_enable_signaling - enable signalling on fence
375 * @f: fence
376 *
377 * This function is called with fence_queue lock held, and adds a callback
378 * to fence_queue that checks if this fence is signaled, and if so it
379 * signals the fence and removes itself.
380 */
381static bool radeon_fence_enable_signaling(struct dma_fence *f)
382{
383 struct radeon_fence *fence = to_radeon_fence(f);
384 struct radeon_device *rdev = fence->rdev;
385
386 if (atomic64_read(&rdev->fence_drv[fence->ring].last_seq) >= fence->seq)
387 return false;
388
389 if (down_read_trylock(&rdev->exclusive_lock)) {
390 radeon_irq_kms_sw_irq_get(rdev, fence->ring);
391
392 if (radeon_fence_activity(rdev, fence->ring))
393 wake_up_all_locked(&rdev->fence_queue);
394
395 /* did fence get signaled after we enabled the sw irq? */
396 if (atomic64_read(&rdev->fence_drv[fence->ring].last_seq) >= fence->seq) {
397 radeon_irq_kms_sw_irq_put(rdev, fence->ring);
398 up_read(&rdev->exclusive_lock);
399 return false;
400 }
401
402 up_read(&rdev->exclusive_lock);
403 } else {
404 /* we're probably in a lockup, lets not fiddle too much */
405 if (radeon_irq_kms_sw_irq_get_delayed(rdev, fence->ring))
406 rdev->fence_drv[fence->ring].delayed_irq = true;
407 radeon_fence_schedule_check(rdev, fence->ring);
408 }
409
410 fence->fence_wake.flags = 0;
411 fence->fence_wake.private = NULL;
412 fence->fence_wake.func = radeon_fence_check_signaled;
413 __add_wait_queue(&rdev->fence_queue, &fence->fence_wake);
414 dma_fence_get(f);
415 return true;
416}
417
418/**
419 * radeon_fence_signaled - check if a fence has signaled
420 *
421 * @fence: radeon fence object
422 *
423 * Check if the requested fence has signaled (all asics).
424 * Returns true if the fence has signaled or false if it has not.
425 */
426bool radeon_fence_signaled(struct radeon_fence *fence)
427{
428 if (!fence)
429 return true;
430
431 if (radeon_fence_seq_signaled(fence->rdev, fence->seq, fence->ring)) {
432 dma_fence_signal(&fence->base);
433 return true;
434 }
435 return false;
436}
437
438/**
439 * radeon_fence_any_seq_signaled - check if any sequence number is signaled
440 *
441 * @rdev: radeon device pointer
442 * @seq: sequence numbers
443 *
444 * Check if the last signaled fence sequnce number is >= the requested
445 * sequence number (all asics).
446 * Returns true if any has signaled (current value is >= requested value)
447 * or false if it has not. Helper function for radeon_fence_wait_seq.
448 */
449static bool radeon_fence_any_seq_signaled(struct radeon_device *rdev, u64 *seq)
450{
451 unsigned int i;
452
453 for (i = 0; i < RADEON_NUM_RINGS; ++i) {
454 if (seq[i] && radeon_fence_seq_signaled(rdev, seq[i], i))
455 return true;
456 }
457 return false;
458}
459
460/**
461 * radeon_fence_wait_seq_timeout - wait for a specific sequence numbers
462 *
463 * @rdev: radeon device pointer
464 * @target_seq: sequence number(s) we want to wait for
465 * @intr: use interruptable sleep
466 * @timeout: maximum time to wait, or MAX_SCHEDULE_TIMEOUT for infinite wait
467 *
468 * Wait for the requested sequence number(s) to be written by any ring
469 * (all asics). Sequnce number array is indexed by ring id.
470 * @intr selects whether to use interruptable (true) or non-interruptable
471 * (false) sleep when waiting for the sequence number. Helper function
472 * for radeon_fence_wait_*().
473 * Returns remaining time if the sequence number has passed, 0 when
474 * the wait timeout, or an error for all other cases.
475 * -EDEADLK is returned when a GPU lockup has been detected.
476 */
477static long radeon_fence_wait_seq_timeout(struct radeon_device *rdev,
478 u64 *target_seq, bool intr,
479 long timeout)
480{
481 long r;
482 int i;
483
484 if (radeon_fence_any_seq_signaled(rdev, target_seq))
485 return timeout;
486
487 /* enable IRQs and tracing */
488 for (i = 0; i < RADEON_NUM_RINGS; ++i) {
489 if (!target_seq[i])
490 continue;
491
492 trace_radeon_fence_wait_begin(rdev_to_drm(rdev), i, target_seq[i]);
493 radeon_irq_kms_sw_irq_get(rdev, i);
494 }
495
496 if (intr) {
497 r = wait_event_interruptible_timeout(rdev->fence_queue, (
498 radeon_fence_any_seq_signaled(rdev, target_seq)
499 || rdev->needs_reset), timeout);
500 } else {
501 r = wait_event_timeout(rdev->fence_queue, (
502 radeon_fence_any_seq_signaled(rdev, target_seq)
503 || rdev->needs_reset), timeout);
504 }
505
506 if (rdev->needs_reset)
507 r = -EDEADLK;
508
509 for (i = 0; i < RADEON_NUM_RINGS; ++i) {
510 if (!target_seq[i])
511 continue;
512
513 radeon_irq_kms_sw_irq_put(rdev, i);
514 trace_radeon_fence_wait_end(rdev_to_drm(rdev), i, target_seq[i]);
515 }
516
517 return r;
518}
519
520/**
521 * radeon_fence_wait_timeout - wait for a fence to signal with timeout
522 *
523 * @fence: radeon fence object
524 * @intr: use interruptible sleep
525 *
526 * Wait for the requested fence to signal (all asics).
527 * @intr selects whether to use interruptable (true) or non-interruptable
528 * (false) sleep when waiting for the fence.
529 * @timeout: maximum time to wait, or MAX_SCHEDULE_TIMEOUT for infinite wait
530 * Returns remaining time if the sequence number has passed, 0 when
531 * the wait timeout, or an error for all other cases.
532 */
533long radeon_fence_wait_timeout(struct radeon_fence *fence, bool intr, long timeout)
534{
535 uint64_t seq[RADEON_NUM_RINGS] = {};
536 long r;
537
538 /*
539 * This function should not be called on !radeon fences.
540 * If this is the case, it would mean this function can
541 * also be called on radeon fences belonging to another card.
542 * exclusive_lock is not held in that case.
543 */
544 if (WARN_ON_ONCE(!to_radeon_fence(&fence->base)))
545 return dma_fence_wait(&fence->base, intr);
546
547 seq[fence->ring] = fence->seq;
548 r = radeon_fence_wait_seq_timeout(fence->rdev, seq, intr, timeout);
549 if (r <= 0)
550 return r;
551
552 dma_fence_signal(&fence->base);
553 return r;
554}
555
556/**
557 * radeon_fence_wait - wait for a fence to signal
558 *
559 * @fence: radeon fence object
560 * @intr: use interruptible sleep
561 *
562 * Wait for the requested fence to signal (all asics).
563 * @intr selects whether to use interruptable (true) or non-interruptable
564 * (false) sleep when waiting for the fence.
565 * Returns 0 if the fence has passed, error for all other cases.
566 */
567int radeon_fence_wait(struct radeon_fence *fence, bool intr)
568{
569 long r = radeon_fence_wait_timeout(fence, intr, MAX_SCHEDULE_TIMEOUT);
570
571 if (r > 0)
572 return 0;
573 else
574 return r;
575}
576
577/**
578 * radeon_fence_wait_any - wait for a fence to signal on any ring
579 *
580 * @rdev: radeon device pointer
581 * @fences: radeon fence object(s)
582 * @intr: use interruptable sleep
583 *
584 * Wait for any requested fence to signal (all asics). Fence
585 * array is indexed by ring id. @intr selects whether to use
586 * interruptable (true) or non-interruptable (false) sleep when
587 * waiting for the fences. Used by the suballocator.
588 * Returns 0 if any fence has passed, error for all other cases.
589 */
590int radeon_fence_wait_any(struct radeon_device *rdev,
591 struct radeon_fence **fences,
592 bool intr)
593{
594 uint64_t seq[RADEON_NUM_RINGS];
595 unsigned int i, num_rings = 0;
596 long r;
597
598 for (i = 0; i < RADEON_NUM_RINGS; ++i) {
599 seq[i] = 0;
600
601 if (!fences[i])
602 continue;
603
604 seq[i] = fences[i]->seq;
605 ++num_rings;
606 }
607
608 /* nothing to wait for ? */
609 if (num_rings == 0)
610 return -ENOENT;
611
612 r = radeon_fence_wait_seq_timeout(rdev, seq, intr, MAX_SCHEDULE_TIMEOUT);
613 if (r < 0)
614 return r;
615
616 return 0;
617}
618
619/**
620 * radeon_fence_wait_next - wait for the next fence to signal
621 *
622 * @rdev: radeon device pointer
623 * @ring: ring index the fence is associated with
624 *
625 * Wait for the next fence on the requested ring to signal (all asics).
626 * Returns 0 if the next fence has passed, error for all other cases.
627 * Caller must hold ring lock.
628 */
629int radeon_fence_wait_next(struct radeon_device *rdev, int ring)
630{
631 uint64_t seq[RADEON_NUM_RINGS] = {};
632 long r;
633
634 seq[ring] = atomic64_read(&rdev->fence_drv[ring].last_seq) + 1ULL;
635 if (seq[ring] >= rdev->fence_drv[ring].sync_seq[ring]) {
636 /* nothing to wait for, last_seq is already
637 * the last emited fence
638 */
639 return -ENOENT;
640 }
641
642 r = radeon_fence_wait_seq_timeout(rdev, seq, false, MAX_SCHEDULE_TIMEOUT);
643 if (r < 0)
644 return r;
645
646 return 0;
647}
648
649/**
650 * radeon_fence_wait_empty - wait for all fences to signal
651 *
652 * @rdev: radeon device pointer
653 * @ring: ring index the fence is associated with
654 *
655 * Wait for all fences on the requested ring to signal (all asics).
656 * Returns 0 if the fences have passed, error for all other cases.
657 * Caller must hold ring lock.
658 */
659int radeon_fence_wait_empty(struct radeon_device *rdev, int ring)
660{
661 uint64_t seq[RADEON_NUM_RINGS] = {};
662 long r;
663
664 seq[ring] = rdev->fence_drv[ring].sync_seq[ring];
665 if (!seq[ring])
666 return 0;
667
668 r = radeon_fence_wait_seq_timeout(rdev, seq, false, MAX_SCHEDULE_TIMEOUT);
669 if (r < 0) {
670 if (r == -EDEADLK)
671 return -EDEADLK;
672
673 dev_err(rdev->dev, "error waiting for ring[%d] to become idle (%ld)\n",
674 ring, r);
675 }
676 return 0;
677}
678
679/**
680 * radeon_fence_ref - take a ref on a fence
681 *
682 * @fence: radeon fence object
683 *
684 * Take a reference on a fence (all asics).
685 * Returns the fence.
686 */
687struct radeon_fence *radeon_fence_ref(struct radeon_fence *fence)
688{
689 dma_fence_get(&fence->base);
690 return fence;
691}
692
693/**
694 * radeon_fence_unref - remove a ref on a fence
695 *
696 * @fence: radeon fence object
697 *
698 * Remove a reference on a fence (all asics).
699 */
700void radeon_fence_unref(struct radeon_fence **fence)
701{
702 struct radeon_fence *tmp = *fence;
703
704 *fence = NULL;
705 if (tmp)
706 dma_fence_put(&tmp->base);
707}
708
709/**
710 * radeon_fence_count_emitted - get the count of emitted fences
711 *
712 * @rdev: radeon device pointer
713 * @ring: ring index the fence is associated with
714 *
715 * Get the number of fences emitted on the requested ring (all asics).
716 * Returns the number of emitted fences on the ring. Used by the
717 * dynpm code to ring track activity.
718 */
719unsigned int radeon_fence_count_emitted(struct radeon_device *rdev, int ring)
720{
721 uint64_t emitted;
722
723 /* We are not protected by ring lock when reading the last sequence
724 * but it's ok to report slightly wrong fence count here.
725 */
726 radeon_fence_process(rdev, ring);
727 emitted = rdev->fence_drv[ring].sync_seq[ring]
728 - atomic64_read(&rdev->fence_drv[ring].last_seq);
729 /* to avoid 32bits warp around */
730 if (emitted > 0x10000000)
731 emitted = 0x10000000;
732
733 return (unsigned int)emitted;
734}
735
736/**
737 * radeon_fence_need_sync - do we need a semaphore
738 *
739 * @fence: radeon fence object
740 * @dst_ring: which ring to check against
741 *
742 * Check if the fence needs to be synced against another ring
743 * (all asics). If so, we need to emit a semaphore.
744 * Returns true if we need to sync with another ring, false if
745 * not.
746 */
747bool radeon_fence_need_sync(struct radeon_fence *fence, int dst_ring)
748{
749 struct radeon_fence_driver *fdrv;
750
751 if (!fence)
752 return false;
753
754 if (fence->ring == dst_ring)
755 return false;
756
757 /* we are protected by the ring mutex */
758 fdrv = &fence->rdev->fence_drv[dst_ring];
759 if (fence->seq <= fdrv->sync_seq[fence->ring])
760 return false;
761
762 return true;
763}
764
765/**
766 * radeon_fence_note_sync - record the sync point
767 *
768 * @fence: radeon fence object
769 * @dst_ring: which ring to check against
770 *
771 * Note the sequence number at which point the fence will
772 * be synced with the requested ring (all asics).
773 */
774void radeon_fence_note_sync(struct radeon_fence *fence, int dst_ring)
775{
776 struct radeon_fence_driver *dst, *src;
777 unsigned int i;
778
779 if (!fence)
780 return;
781
782 if (fence->ring == dst_ring)
783 return;
784
785 /* we are protected by the ring mutex */
786 src = &fence->rdev->fence_drv[fence->ring];
787 dst = &fence->rdev->fence_drv[dst_ring];
788 for (i = 0; i < RADEON_NUM_RINGS; ++i) {
789 if (i == dst_ring)
790 continue;
791
792 dst->sync_seq[i] = max(dst->sync_seq[i], src->sync_seq[i]);
793 }
794}
795
796/**
797 * radeon_fence_driver_start_ring - make the fence driver
798 * ready for use on the requested ring.
799 *
800 * @rdev: radeon device pointer
801 * @ring: ring index to start the fence driver on
802 *
803 * Make the fence driver ready for processing (all asics).
804 * Not all asics have all rings, so each asic will only
805 * start the fence driver on the rings it has.
806 * Returns 0 for success, errors for failure.
807 */
808int radeon_fence_driver_start_ring(struct radeon_device *rdev, int ring)
809{
810 uint64_t index;
811 int r;
812
813 radeon_scratch_free(rdev, rdev->fence_drv[ring].scratch_reg);
814 if (rdev->wb.use_event || !radeon_ring_supports_scratch_reg(rdev, &rdev->ring[ring])) {
815 rdev->fence_drv[ring].scratch_reg = 0;
816 if (ring != R600_RING_TYPE_UVD_INDEX) {
817 index = R600_WB_EVENT_OFFSET + ring * 4;
818 rdev->fence_drv[ring].cpu_addr = &rdev->wb.wb[index/4];
819 rdev->fence_drv[ring].gpu_addr = rdev->wb.gpu_addr +
820 index;
821
822 } else {
823 /* put fence directly behind firmware */
824 index = ALIGN(rdev->uvd_fw->size, 8);
825 rdev->fence_drv[ring].cpu_addr = rdev->uvd.cpu_addr + index;
826 rdev->fence_drv[ring].gpu_addr = rdev->uvd.gpu_addr + index;
827 }
828
829 } else {
830 r = radeon_scratch_get(rdev, &rdev->fence_drv[ring].scratch_reg);
831 if (r) {
832 dev_err(rdev->dev, "fence failed to get scratch register\n");
833 return r;
834 }
835 index = RADEON_WB_SCRATCH_OFFSET +
836 rdev->fence_drv[ring].scratch_reg -
837 rdev->scratch.reg_base;
838 rdev->fence_drv[ring].cpu_addr = &rdev->wb.wb[index/4];
839 rdev->fence_drv[ring].gpu_addr = rdev->wb.gpu_addr + index;
840 }
841 radeon_fence_write(rdev, atomic64_read(&rdev->fence_drv[ring].last_seq), ring);
842 rdev->fence_drv[ring].initialized = true;
843 dev_info(rdev->dev, "fence driver on ring %d use gpu addr 0x%016llx\n",
844 ring, rdev->fence_drv[ring].gpu_addr);
845 return 0;
846}
847
848/**
849 * radeon_fence_driver_init_ring - init the fence driver
850 * for the requested ring.
851 *
852 * @rdev: radeon device pointer
853 * @ring: ring index to start the fence driver on
854 *
855 * Init the fence driver for the requested ring (all asics).
856 * Helper function for radeon_fence_driver_init().
857 */
858static void radeon_fence_driver_init_ring(struct radeon_device *rdev, int ring)
859{
860 int i;
861
862 rdev->fence_drv[ring].scratch_reg = -1;
863 rdev->fence_drv[ring].cpu_addr = NULL;
864 rdev->fence_drv[ring].gpu_addr = 0;
865 for (i = 0; i < RADEON_NUM_RINGS; ++i)
866 rdev->fence_drv[ring].sync_seq[i] = 0;
867 atomic64_set(&rdev->fence_drv[ring].last_seq, 0);
868 rdev->fence_drv[ring].initialized = false;
869 INIT_DELAYED_WORK(&rdev->fence_drv[ring].lockup_work,
870 radeon_fence_check_lockup);
871 rdev->fence_drv[ring].rdev = rdev;
872}
873
874/**
875 * radeon_fence_driver_init - init the fence driver
876 * for all possible rings.
877 *
878 * @rdev: radeon device pointer
879 *
880 * Init the fence driver for all possible rings (all asics).
881 * Not all asics have all rings, so each asic will only
882 * start the fence driver on the rings it has using
883 * radeon_fence_driver_start_ring().
884 */
885void radeon_fence_driver_init(struct radeon_device *rdev)
886{
887 int ring;
888
889 init_waitqueue_head(&rdev->fence_queue);
890 for (ring = 0; ring < RADEON_NUM_RINGS; ring++)
891 radeon_fence_driver_init_ring(rdev, ring);
892
893 radeon_debugfs_fence_init(rdev);
894}
895
896/**
897 * radeon_fence_driver_fini - tear down the fence driver
898 * for all possible rings.
899 *
900 * @rdev: radeon device pointer
901 *
902 * Tear down the fence driver for all possible rings (all asics).
903 */
904void radeon_fence_driver_fini(struct radeon_device *rdev)
905{
906 int ring, r;
907
908 mutex_lock(&rdev->ring_lock);
909 for (ring = 0; ring < RADEON_NUM_RINGS; ring++) {
910 if (!rdev->fence_drv[ring].initialized)
911 continue;
912 r = radeon_fence_wait_empty(rdev, ring);
913 if (r) {
914 /* no need to trigger GPU reset as we are unloading */
915 radeon_fence_driver_force_completion(rdev, ring);
916 }
917 cancel_delayed_work_sync(&rdev->fence_drv[ring].lockup_work);
918 wake_up_all(&rdev->fence_queue);
919 radeon_scratch_free(rdev, rdev->fence_drv[ring].scratch_reg);
920 rdev->fence_drv[ring].initialized = false;
921 }
922 mutex_unlock(&rdev->ring_lock);
923}
924
925/**
926 * radeon_fence_driver_force_completion - force all fence waiter to complete
927 *
928 * @rdev: radeon device pointer
929 * @ring: the ring to complete
930 *
931 * In case of GPU reset failure make sure no process keep waiting on fence
932 * that will never complete.
933 */
934void radeon_fence_driver_force_completion(struct radeon_device *rdev, int ring)
935{
936 if (rdev->fence_drv[ring].initialized) {
937 radeon_fence_write(rdev, rdev->fence_drv[ring].sync_seq[ring], ring);
938 cancel_delayed_work_sync(&rdev->fence_drv[ring].lockup_work);
939 }
940}
941
942
943/*
944 * Fence debugfs
945 */
946#if defined(CONFIG_DEBUG_FS)
947static int radeon_debugfs_fence_info_show(struct seq_file *m, void *data)
948{
949 struct radeon_device *rdev = m->private;
950 int i, j;
951
952 for (i = 0; i < RADEON_NUM_RINGS; ++i) {
953 if (!rdev->fence_drv[i].initialized)
954 continue;
955
956 radeon_fence_process(rdev, i);
957
958 seq_printf(m, "--- ring %d ---\n", i);
959 seq_printf(m, "Last signaled fence 0x%016llx\n",
960 (unsigned long long)atomic64_read(&rdev->fence_drv[i].last_seq));
961 seq_printf(m, "Last emitted 0x%016llx\n",
962 rdev->fence_drv[i].sync_seq[i]);
963
964 for (j = 0; j < RADEON_NUM_RINGS; ++j) {
965 if (i != j && rdev->fence_drv[j].initialized)
966 seq_printf(m, "Last sync to ring %d 0x%016llx\n",
967 j, rdev->fence_drv[i].sync_seq[j]);
968 }
969 }
970 return 0;
971}
972
973/*
974 * radeon_debugfs_gpu_reset - manually trigger a gpu reset
975 *
976 * Manually trigger a gpu reset at the next fence wait.
977 */
978static int radeon_debugfs_gpu_reset(void *data, u64 *val)
979{
980 struct radeon_device *rdev = (struct radeon_device *)data;
981
982 down_read(&rdev->exclusive_lock);
983 *val = rdev->needs_reset;
984 rdev->needs_reset = true;
985 wake_up_all(&rdev->fence_queue);
986 up_read(&rdev->exclusive_lock);
987
988 return 0;
989}
990DEFINE_SHOW_ATTRIBUTE(radeon_debugfs_fence_info);
991DEFINE_DEBUGFS_ATTRIBUTE(radeon_debugfs_gpu_reset_fops,
992 radeon_debugfs_gpu_reset, NULL, "%lld\n");
993#endif
994
995void radeon_debugfs_fence_init(struct radeon_device *rdev)
996{
997#if defined(CONFIG_DEBUG_FS)
998 struct dentry *root = rdev_to_drm(rdev)->primary->debugfs_root;
999
1000 debugfs_create_file("radeon_gpu_reset", 0444, root, rdev,
1001 &radeon_debugfs_gpu_reset_fops);
1002 debugfs_create_file("radeon_fence_info", 0444, root, rdev,
1003 &radeon_debugfs_fence_info_fops);
1004
1005
1006#endif
1007}
1008
1009static const char *radeon_fence_get_driver_name(struct dma_fence *fence)
1010{
1011 return "radeon";
1012}
1013
1014static const char *radeon_fence_get_timeline_name(struct dma_fence *f)
1015{
1016 struct radeon_fence *fence = to_radeon_fence(f);
1017
1018 switch (fence->ring) {
1019 case RADEON_RING_TYPE_GFX_INDEX: return "radeon.gfx";
1020 case CAYMAN_RING_TYPE_CP1_INDEX: return "radeon.cp1";
1021 case CAYMAN_RING_TYPE_CP2_INDEX: return "radeon.cp2";
1022 case R600_RING_TYPE_DMA_INDEX: return "radeon.dma";
1023 case CAYMAN_RING_TYPE_DMA1_INDEX: return "radeon.dma1";
1024 case R600_RING_TYPE_UVD_INDEX: return "radeon.uvd";
1025 case TN_RING_TYPE_VCE1_INDEX: return "radeon.vce1";
1026 case TN_RING_TYPE_VCE2_INDEX: return "radeon.vce2";
1027 default:
1028 WARN_ON_ONCE(1);
1029 return "radeon.unk";
1030 }
1031}
1032
1033static inline bool radeon_test_signaled(struct radeon_fence *fence)
1034{
1035 return test_bit(DMA_FENCE_FLAG_SIGNALED_BIT, &fence->base.flags);
1036}
1037
1038struct radeon_wait_cb {
1039 struct dma_fence_cb base;
1040 struct task_struct *task;
1041};
1042
1043static void
1044radeon_fence_wait_cb(struct dma_fence *fence, struct dma_fence_cb *cb)
1045{
1046 struct radeon_wait_cb *wait =
1047 container_of(cb, struct radeon_wait_cb, base);
1048
1049 wake_up_process(wait->task);
1050}
1051
1052static signed long radeon_fence_default_wait(struct dma_fence *f, bool intr,
1053 signed long t)
1054{
1055 struct radeon_fence *fence = to_radeon_fence(f);
1056 struct radeon_device *rdev = fence->rdev;
1057 struct radeon_wait_cb cb;
1058
1059 cb.task = current;
1060
1061 if (dma_fence_add_callback(f, &cb.base, radeon_fence_wait_cb))
1062 return t;
1063
1064 while (t > 0) {
1065 if (intr)
1066 set_current_state(TASK_INTERRUPTIBLE);
1067 else
1068 set_current_state(TASK_UNINTERRUPTIBLE);
1069
1070 /*
1071 * radeon_test_signaled must be called after
1072 * set_current_state to prevent a race with wake_up_process
1073 */
1074 if (radeon_test_signaled(fence))
1075 break;
1076
1077 if (rdev->needs_reset) {
1078 t = -EDEADLK;
1079 break;
1080 }
1081
1082 t = schedule_timeout(t);
1083
1084 if (t > 0 && intr && signal_pending(current))
1085 t = -ERESTARTSYS;
1086 }
1087
1088 __set_current_state(TASK_RUNNING);
1089 dma_fence_remove_callback(f, &cb.base);
1090
1091 return t;
1092}
1093
1094const struct dma_fence_ops radeon_fence_ops = {
1095 .get_driver_name = radeon_fence_get_driver_name,
1096 .get_timeline_name = radeon_fence_get_timeline_name,
1097 .enable_signaling = radeon_fence_enable_signaling,
1098 .signaled = radeon_fence_is_signaled,
1099 .wait = radeon_fence_default_wait,
1100 .release = NULL,
1101};
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 <drm/drmP.h>
38#include "radeon_reg.h"
39#include "radeon.h"
40#include "radeon_trace.h"
41
42/*
43 * Fences
44 * Fences mark an event in the GPUs pipeline and are used
45 * for GPU/CPU synchronization. When the fence is written,
46 * it is expected that all buffers associated with that fence
47 * are no longer in use by the associated ring on the GPU and
48 * that the the relevant GPU caches have been flushed. Whether
49 * we use a scratch register or memory location depends on the asic
50 * and whether writeback is enabled.
51 */
52
53/**
54 * radeon_fence_write - write a fence value
55 *
56 * @rdev: radeon_device pointer
57 * @seq: sequence number to write
58 * @ring: ring index the fence is associated with
59 *
60 * Writes a fence value to memory or a scratch register (all asics).
61 */
62static void radeon_fence_write(struct radeon_device *rdev, u32 seq, int ring)
63{
64 struct radeon_fence_driver *drv = &rdev->fence_drv[ring];
65 if (likely(rdev->wb.enabled || !drv->scratch_reg)) {
66 if (drv->cpu_addr) {
67 *drv->cpu_addr = cpu_to_le32(seq);
68 }
69 } else {
70 WREG32(drv->scratch_reg, seq);
71 }
72}
73
74/**
75 * radeon_fence_read - read a fence value
76 *
77 * @rdev: radeon_device pointer
78 * @ring: ring index the fence is associated with
79 *
80 * Reads a fence value from memory or a scratch register (all asics).
81 * Returns the value of the fence read from memory or register.
82 */
83static u32 radeon_fence_read(struct radeon_device *rdev, int ring)
84{
85 struct radeon_fence_driver *drv = &rdev->fence_drv[ring];
86 u32 seq = 0;
87
88 if (likely(rdev->wb.enabled || !drv->scratch_reg)) {
89 if (drv->cpu_addr) {
90 seq = le32_to_cpu(*drv->cpu_addr);
91 } else {
92 seq = lower_32_bits(atomic64_read(&drv->last_seq));
93 }
94 } else {
95 seq = RREG32(drv->scratch_reg);
96 }
97 return seq;
98}
99
100/**
101 * radeon_fence_emit - emit a fence on the requested ring
102 *
103 * @rdev: radeon_device pointer
104 * @fence: radeon fence object
105 * @ring: ring index the fence is associated with
106 *
107 * Emits a fence command on the requested ring (all asics).
108 * Returns 0 on success, -ENOMEM on failure.
109 */
110int radeon_fence_emit(struct radeon_device *rdev,
111 struct radeon_fence **fence,
112 int ring)
113{
114 /* we are protected by the ring emission mutex */
115 *fence = kmalloc(sizeof(struct radeon_fence), GFP_KERNEL);
116 if ((*fence) == NULL) {
117 return -ENOMEM;
118 }
119 kref_init(&((*fence)->kref));
120 (*fence)->rdev = rdev;
121 (*fence)->seq = ++rdev->fence_drv[ring].sync_seq[ring];
122 (*fence)->ring = ring;
123 radeon_fence_ring_emit(rdev, ring, *fence);
124 trace_radeon_fence_emit(rdev->ddev, ring, (*fence)->seq);
125 return 0;
126}
127
128/**
129 * radeon_fence_process - process a fence
130 *
131 * @rdev: radeon_device pointer
132 * @ring: ring index the fence is associated with
133 *
134 * Checks the current fence value and wakes the fence queue
135 * if the sequence number has increased (all asics).
136 */
137void radeon_fence_process(struct radeon_device *rdev, int ring)
138{
139 uint64_t seq, last_seq, last_emitted;
140 unsigned count_loop = 0;
141 bool wake = false;
142
143 /* Note there is a scenario here for an infinite loop but it's
144 * very unlikely to happen. For it to happen, the current polling
145 * process need to be interrupted by another process and another
146 * process needs to update the last_seq btw the atomic read and
147 * xchg of the current process.
148 *
149 * More over for this to go in infinite loop there need to be
150 * continuously new fence signaled ie radeon_fence_read needs
151 * to return a different value each time for both the currently
152 * polling process and the other process that xchg the last_seq
153 * btw atomic read and xchg of the current process. And the
154 * value the other process set as last seq must be higher than
155 * the seq value we just read. Which means that current process
156 * need to be interrupted after radeon_fence_read and before
157 * atomic xchg.
158 *
159 * To be even more safe we count the number of time we loop and
160 * we bail after 10 loop just accepting the fact that we might
161 * have temporarly set the last_seq not to the true real last
162 * seq but to an older one.
163 */
164 last_seq = atomic64_read(&rdev->fence_drv[ring].last_seq);
165 do {
166 last_emitted = rdev->fence_drv[ring].sync_seq[ring];
167 seq = radeon_fence_read(rdev, ring);
168 seq |= last_seq & 0xffffffff00000000LL;
169 if (seq < last_seq) {
170 seq &= 0xffffffff;
171 seq |= last_emitted & 0xffffffff00000000LL;
172 }
173
174 if (seq <= last_seq || seq > last_emitted) {
175 break;
176 }
177 /* If we loop over we don't want to return without
178 * checking if a fence is signaled as it means that the
179 * seq we just read is different from the previous on.
180 */
181 wake = true;
182 last_seq = seq;
183 if ((count_loop++) > 10) {
184 /* We looped over too many time leave with the
185 * fact that we might have set an older fence
186 * seq then the current real last seq as signaled
187 * by the hw.
188 */
189 break;
190 }
191 } while (atomic64_xchg(&rdev->fence_drv[ring].last_seq, seq) > seq);
192
193 if (wake)
194 wake_up_all(&rdev->fence_queue);
195}
196
197/**
198 * radeon_fence_destroy - destroy a fence
199 *
200 * @kref: fence kref
201 *
202 * Frees the fence object (all asics).
203 */
204static void radeon_fence_destroy(struct kref *kref)
205{
206 struct radeon_fence *fence;
207
208 fence = container_of(kref, struct radeon_fence, kref);
209 kfree(fence);
210}
211
212/**
213 * radeon_fence_seq_signaled - check if a fence sequence number has signaled
214 *
215 * @rdev: radeon device pointer
216 * @seq: sequence number
217 * @ring: ring index the fence is associated with
218 *
219 * Check if the last signaled fence sequnce number is >= the requested
220 * sequence number (all asics).
221 * Returns true if the fence has signaled (current fence value
222 * is >= requested value) or false if it has not (current fence
223 * value is < the requested value. Helper function for
224 * radeon_fence_signaled().
225 */
226static bool radeon_fence_seq_signaled(struct radeon_device *rdev,
227 u64 seq, unsigned ring)
228{
229 if (atomic64_read(&rdev->fence_drv[ring].last_seq) >= seq) {
230 return true;
231 }
232 /* poll new last sequence at least once */
233 radeon_fence_process(rdev, ring);
234 if (atomic64_read(&rdev->fence_drv[ring].last_seq) >= seq) {
235 return true;
236 }
237 return false;
238}
239
240/**
241 * radeon_fence_signaled - check if a fence has signaled
242 *
243 * @fence: radeon fence object
244 *
245 * Check if the requested fence has signaled (all asics).
246 * Returns true if the fence has signaled or false if it has not.
247 */
248bool radeon_fence_signaled(struct radeon_fence *fence)
249{
250 if (!fence) {
251 return true;
252 }
253 if (fence->seq == RADEON_FENCE_SIGNALED_SEQ) {
254 return true;
255 }
256 if (radeon_fence_seq_signaled(fence->rdev, fence->seq, fence->ring)) {
257 fence->seq = RADEON_FENCE_SIGNALED_SEQ;
258 return true;
259 }
260 return false;
261}
262
263/**
264 * radeon_fence_any_seq_signaled - check if any sequence number is signaled
265 *
266 * @rdev: radeon device pointer
267 * @seq: sequence numbers
268 *
269 * Check if the last signaled fence sequnce number is >= the requested
270 * sequence number (all asics).
271 * Returns true if any has signaled (current value is >= requested value)
272 * or false if it has not. Helper function for radeon_fence_wait_seq.
273 */
274static bool radeon_fence_any_seq_signaled(struct radeon_device *rdev, u64 *seq)
275{
276 unsigned i;
277
278 for (i = 0; i < RADEON_NUM_RINGS; ++i) {
279 if (seq[i] && radeon_fence_seq_signaled(rdev, seq[i], i))
280 return true;
281 }
282 return false;
283}
284
285/**
286 * radeon_fence_wait_seq - wait for a specific sequence numbers
287 *
288 * @rdev: radeon device pointer
289 * @target_seq: sequence number(s) we want to wait for
290 * @intr: use interruptable sleep
291 *
292 * Wait for the requested sequence number(s) to be written by any ring
293 * (all asics). Sequnce number array is indexed by ring id.
294 * @intr selects whether to use interruptable (true) or non-interruptable
295 * (false) sleep when waiting for the sequence number. Helper function
296 * for radeon_fence_wait_*().
297 * Returns 0 if the sequence number has passed, error for all other cases.
298 * -EDEADLK is returned when a GPU lockup has been detected.
299 */
300static int radeon_fence_wait_seq(struct radeon_device *rdev, u64 *target_seq,
301 bool intr)
302{
303 uint64_t last_seq[RADEON_NUM_RINGS];
304 bool signaled;
305 int i, r;
306
307 while (!radeon_fence_any_seq_signaled(rdev, target_seq)) {
308
309 /* Save current sequence values, used to check for GPU lockups */
310 for (i = 0; i < RADEON_NUM_RINGS; ++i) {
311 if (!target_seq[i])
312 continue;
313
314 last_seq[i] = atomic64_read(&rdev->fence_drv[i].last_seq);
315 trace_radeon_fence_wait_begin(rdev->ddev, i, target_seq[i]);
316 radeon_irq_kms_sw_irq_get(rdev, i);
317 }
318
319 if (intr) {
320 r = wait_event_interruptible_timeout(rdev->fence_queue, (
321 (signaled = radeon_fence_any_seq_signaled(rdev, target_seq))
322 || rdev->needs_reset), RADEON_FENCE_JIFFIES_TIMEOUT);
323 } else {
324 r = wait_event_timeout(rdev->fence_queue, (
325 (signaled = radeon_fence_any_seq_signaled(rdev, target_seq))
326 || rdev->needs_reset), RADEON_FENCE_JIFFIES_TIMEOUT);
327 }
328
329 for (i = 0; i < RADEON_NUM_RINGS; ++i) {
330 if (!target_seq[i])
331 continue;
332
333 radeon_irq_kms_sw_irq_put(rdev, i);
334 trace_radeon_fence_wait_end(rdev->ddev, i, target_seq[i]);
335 }
336
337 if (unlikely(r < 0))
338 return r;
339
340 if (unlikely(!signaled)) {
341 if (rdev->needs_reset)
342 return -EDEADLK;
343
344 /* we were interrupted for some reason and fence
345 * isn't signaled yet, resume waiting */
346 if (r)
347 continue;
348
349 for (i = 0; i < RADEON_NUM_RINGS; ++i) {
350 if (!target_seq[i])
351 continue;
352
353 if (last_seq[i] != atomic64_read(&rdev->fence_drv[i].last_seq))
354 break;
355 }
356
357 if (i != RADEON_NUM_RINGS)
358 continue;
359
360 for (i = 0; i < RADEON_NUM_RINGS; ++i) {
361 if (!target_seq[i])
362 continue;
363
364 if (radeon_ring_is_lockup(rdev, i, &rdev->ring[i]))
365 break;
366 }
367
368 if (i < RADEON_NUM_RINGS) {
369 /* good news we believe it's a lockup */
370 dev_warn(rdev->dev, "GPU lockup (waiting for "
371 "0x%016llx last fence id 0x%016llx on"
372 " ring %d)\n",
373 target_seq[i], last_seq[i], i);
374
375 /* remember that we need an reset */
376 rdev->needs_reset = true;
377 wake_up_all(&rdev->fence_queue);
378 return -EDEADLK;
379 }
380 }
381 }
382 return 0;
383}
384
385/**
386 * radeon_fence_wait - wait for a fence to signal
387 *
388 * @fence: radeon fence object
389 * @intr: use interruptable sleep
390 *
391 * Wait for the requested fence to signal (all asics).
392 * @intr selects whether to use interruptable (true) or non-interruptable
393 * (false) sleep when waiting for the fence.
394 * Returns 0 if the fence has passed, error for all other cases.
395 */
396int radeon_fence_wait(struct radeon_fence *fence, bool intr)
397{
398 uint64_t seq[RADEON_NUM_RINGS] = {};
399 int r;
400
401 if (fence == NULL) {
402 WARN(1, "Querying an invalid fence : %p !\n", fence);
403 return -EINVAL;
404 }
405
406 seq[fence->ring] = fence->seq;
407 if (seq[fence->ring] == RADEON_FENCE_SIGNALED_SEQ)
408 return 0;
409
410 r = radeon_fence_wait_seq(fence->rdev, seq, intr);
411 if (r)
412 return r;
413
414 fence->seq = RADEON_FENCE_SIGNALED_SEQ;
415 return 0;
416}
417
418/**
419 * radeon_fence_wait_any - wait for a fence to signal on any ring
420 *
421 * @rdev: radeon device pointer
422 * @fences: radeon fence object(s)
423 * @intr: use interruptable sleep
424 *
425 * Wait for any requested fence to signal (all asics). Fence
426 * array is indexed by ring id. @intr selects whether to use
427 * interruptable (true) or non-interruptable (false) sleep when
428 * waiting for the fences. Used by the suballocator.
429 * Returns 0 if any fence has passed, error for all other cases.
430 */
431int radeon_fence_wait_any(struct radeon_device *rdev,
432 struct radeon_fence **fences,
433 bool intr)
434{
435 uint64_t seq[RADEON_NUM_RINGS];
436 unsigned i, num_rings = 0;
437 int r;
438
439 for (i = 0; i < RADEON_NUM_RINGS; ++i) {
440 seq[i] = 0;
441
442 if (!fences[i]) {
443 continue;
444 }
445
446 seq[i] = fences[i]->seq;
447 ++num_rings;
448
449 /* test if something was allready signaled */
450 if (seq[i] == RADEON_FENCE_SIGNALED_SEQ)
451 return 0;
452 }
453
454 /* nothing to wait for ? */
455 if (num_rings == 0)
456 return -ENOENT;
457
458 r = radeon_fence_wait_seq(rdev, seq, intr);
459 if (r) {
460 return r;
461 }
462 return 0;
463}
464
465/**
466 * radeon_fence_wait_next - wait for the next fence to signal
467 *
468 * @rdev: radeon device pointer
469 * @ring: ring index the fence is associated with
470 *
471 * Wait for the next fence on the requested ring to signal (all asics).
472 * Returns 0 if the next fence has passed, error for all other cases.
473 * Caller must hold ring lock.
474 */
475int radeon_fence_wait_next(struct radeon_device *rdev, int ring)
476{
477 uint64_t seq[RADEON_NUM_RINGS] = {};
478
479 seq[ring] = atomic64_read(&rdev->fence_drv[ring].last_seq) + 1ULL;
480 if (seq[ring] >= rdev->fence_drv[ring].sync_seq[ring]) {
481 /* nothing to wait for, last_seq is
482 already the last emited fence */
483 return -ENOENT;
484 }
485 return radeon_fence_wait_seq(rdev, seq, false);
486}
487
488/**
489 * radeon_fence_wait_empty - wait for all fences to signal
490 *
491 * @rdev: radeon device pointer
492 * @ring: ring index the fence is associated with
493 *
494 * Wait for all fences on the requested ring to signal (all asics).
495 * Returns 0 if the fences have passed, error for all other cases.
496 * Caller must hold ring lock.
497 */
498int radeon_fence_wait_empty(struct radeon_device *rdev, int ring)
499{
500 uint64_t seq[RADEON_NUM_RINGS] = {};
501 int r;
502
503 seq[ring] = rdev->fence_drv[ring].sync_seq[ring];
504 if (!seq[ring])
505 return 0;
506
507 r = radeon_fence_wait_seq(rdev, seq, false);
508 if (r) {
509 if (r == -EDEADLK)
510 return -EDEADLK;
511
512 dev_err(rdev->dev, "error waiting for ring[%d] to become idle (%d)\n",
513 ring, r);
514 }
515 return 0;
516}
517
518/**
519 * radeon_fence_ref - take a ref on a fence
520 *
521 * @fence: radeon fence object
522 *
523 * Take a reference on a fence (all asics).
524 * Returns the fence.
525 */
526struct radeon_fence *radeon_fence_ref(struct radeon_fence *fence)
527{
528 kref_get(&fence->kref);
529 return fence;
530}
531
532/**
533 * radeon_fence_unref - remove a ref on a fence
534 *
535 * @fence: radeon fence object
536 *
537 * Remove a reference on a fence (all asics).
538 */
539void radeon_fence_unref(struct radeon_fence **fence)
540{
541 struct radeon_fence *tmp = *fence;
542
543 *fence = NULL;
544 if (tmp) {
545 kref_put(&tmp->kref, radeon_fence_destroy);
546 }
547}
548
549/**
550 * radeon_fence_count_emitted - get the count of emitted fences
551 *
552 * @rdev: radeon device pointer
553 * @ring: ring index the fence is associated with
554 *
555 * Get the number of fences emitted on the requested ring (all asics).
556 * Returns the number of emitted fences on the ring. Used by the
557 * dynpm code to ring track activity.
558 */
559unsigned radeon_fence_count_emitted(struct radeon_device *rdev, int ring)
560{
561 uint64_t emitted;
562
563 /* We are not protected by ring lock when reading the last sequence
564 * but it's ok to report slightly wrong fence count here.
565 */
566 radeon_fence_process(rdev, ring);
567 emitted = rdev->fence_drv[ring].sync_seq[ring]
568 - atomic64_read(&rdev->fence_drv[ring].last_seq);
569 /* to avoid 32bits warp around */
570 if (emitted > 0x10000000) {
571 emitted = 0x10000000;
572 }
573 return (unsigned)emitted;
574}
575
576/**
577 * radeon_fence_need_sync - do we need a semaphore
578 *
579 * @fence: radeon fence object
580 * @dst_ring: which ring to check against
581 *
582 * Check if the fence needs to be synced against another ring
583 * (all asics). If so, we need to emit a semaphore.
584 * Returns true if we need to sync with another ring, false if
585 * not.
586 */
587bool radeon_fence_need_sync(struct radeon_fence *fence, int dst_ring)
588{
589 struct radeon_fence_driver *fdrv;
590
591 if (!fence) {
592 return false;
593 }
594
595 if (fence->ring == dst_ring) {
596 return false;
597 }
598
599 /* we are protected by the ring mutex */
600 fdrv = &fence->rdev->fence_drv[dst_ring];
601 if (fence->seq <= fdrv->sync_seq[fence->ring]) {
602 return false;
603 }
604
605 return true;
606}
607
608/**
609 * radeon_fence_note_sync - record the sync point
610 *
611 * @fence: radeon fence object
612 * @dst_ring: which ring to check against
613 *
614 * Note the sequence number at which point the fence will
615 * be synced with the requested ring (all asics).
616 */
617void radeon_fence_note_sync(struct radeon_fence *fence, int dst_ring)
618{
619 struct radeon_fence_driver *dst, *src;
620 unsigned i;
621
622 if (!fence) {
623 return;
624 }
625
626 if (fence->ring == dst_ring) {
627 return;
628 }
629
630 /* we are protected by the ring mutex */
631 src = &fence->rdev->fence_drv[fence->ring];
632 dst = &fence->rdev->fence_drv[dst_ring];
633 for (i = 0; i < RADEON_NUM_RINGS; ++i) {
634 if (i == dst_ring) {
635 continue;
636 }
637 dst->sync_seq[i] = max(dst->sync_seq[i], src->sync_seq[i]);
638 }
639}
640
641/**
642 * radeon_fence_driver_start_ring - make the fence driver
643 * ready for use on the requested ring.
644 *
645 * @rdev: radeon device pointer
646 * @ring: ring index to start the fence driver on
647 *
648 * Make the fence driver ready for processing (all asics).
649 * Not all asics have all rings, so each asic will only
650 * start the fence driver on the rings it has.
651 * Returns 0 for success, errors for failure.
652 */
653int radeon_fence_driver_start_ring(struct radeon_device *rdev, int ring)
654{
655 uint64_t index;
656 int r;
657
658 radeon_scratch_free(rdev, rdev->fence_drv[ring].scratch_reg);
659 if (rdev->wb.use_event || !radeon_ring_supports_scratch_reg(rdev, &rdev->ring[ring])) {
660 rdev->fence_drv[ring].scratch_reg = 0;
661 if (ring != R600_RING_TYPE_UVD_INDEX) {
662 index = R600_WB_EVENT_OFFSET + ring * 4;
663 rdev->fence_drv[ring].cpu_addr = &rdev->wb.wb[index/4];
664 rdev->fence_drv[ring].gpu_addr = rdev->wb.gpu_addr +
665 index;
666
667 } else {
668 /* put fence directly behind firmware */
669 index = ALIGN(rdev->uvd_fw->size, 8);
670 rdev->fence_drv[ring].cpu_addr = rdev->uvd.cpu_addr + index;
671 rdev->fence_drv[ring].gpu_addr = rdev->uvd.gpu_addr + index;
672 }
673
674 } else {
675 r = radeon_scratch_get(rdev, &rdev->fence_drv[ring].scratch_reg);
676 if (r) {
677 dev_err(rdev->dev, "fence failed to get scratch register\n");
678 return r;
679 }
680 index = RADEON_WB_SCRATCH_OFFSET +
681 rdev->fence_drv[ring].scratch_reg -
682 rdev->scratch.reg_base;
683 rdev->fence_drv[ring].cpu_addr = &rdev->wb.wb[index/4];
684 rdev->fence_drv[ring].gpu_addr = rdev->wb.gpu_addr + index;
685 }
686 radeon_fence_write(rdev, atomic64_read(&rdev->fence_drv[ring].last_seq), ring);
687 rdev->fence_drv[ring].initialized = true;
688 dev_info(rdev->dev, "fence driver on ring %d use gpu addr 0x%016llx and cpu addr 0x%p\n",
689 ring, rdev->fence_drv[ring].gpu_addr, rdev->fence_drv[ring].cpu_addr);
690 return 0;
691}
692
693/**
694 * radeon_fence_driver_init_ring - init the fence driver
695 * for the requested ring.
696 *
697 * @rdev: radeon device pointer
698 * @ring: ring index to start the fence driver on
699 *
700 * Init the fence driver for the requested ring (all asics).
701 * Helper function for radeon_fence_driver_init().
702 */
703static void radeon_fence_driver_init_ring(struct radeon_device *rdev, int ring)
704{
705 int i;
706
707 rdev->fence_drv[ring].scratch_reg = -1;
708 rdev->fence_drv[ring].cpu_addr = NULL;
709 rdev->fence_drv[ring].gpu_addr = 0;
710 for (i = 0; i < RADEON_NUM_RINGS; ++i)
711 rdev->fence_drv[ring].sync_seq[i] = 0;
712 atomic64_set(&rdev->fence_drv[ring].last_seq, 0);
713 rdev->fence_drv[ring].initialized = false;
714}
715
716/**
717 * radeon_fence_driver_init - init the fence driver
718 * for all possible rings.
719 *
720 * @rdev: radeon device pointer
721 *
722 * Init the fence driver for all possible rings (all asics).
723 * Not all asics have all rings, so each asic will only
724 * start the fence driver on the rings it has using
725 * radeon_fence_driver_start_ring().
726 * Returns 0 for success.
727 */
728int radeon_fence_driver_init(struct radeon_device *rdev)
729{
730 int ring;
731
732 init_waitqueue_head(&rdev->fence_queue);
733 for (ring = 0; ring < RADEON_NUM_RINGS; ring++) {
734 radeon_fence_driver_init_ring(rdev, ring);
735 }
736 if (radeon_debugfs_fence_init(rdev)) {
737 dev_err(rdev->dev, "fence debugfs file creation failed\n");
738 }
739 return 0;
740}
741
742/**
743 * radeon_fence_driver_fini - tear down the fence driver
744 * for all possible rings.
745 *
746 * @rdev: radeon device pointer
747 *
748 * Tear down the fence driver for all possible rings (all asics).
749 */
750void radeon_fence_driver_fini(struct radeon_device *rdev)
751{
752 int ring, r;
753
754 mutex_lock(&rdev->ring_lock);
755 for (ring = 0; ring < RADEON_NUM_RINGS; ring++) {
756 if (!rdev->fence_drv[ring].initialized)
757 continue;
758 r = radeon_fence_wait_empty(rdev, ring);
759 if (r) {
760 /* no need to trigger GPU reset as we are unloading */
761 radeon_fence_driver_force_completion(rdev);
762 }
763 wake_up_all(&rdev->fence_queue);
764 radeon_scratch_free(rdev, rdev->fence_drv[ring].scratch_reg);
765 rdev->fence_drv[ring].initialized = false;
766 }
767 mutex_unlock(&rdev->ring_lock);
768}
769
770/**
771 * radeon_fence_driver_force_completion - force all fence waiter to complete
772 *
773 * @rdev: radeon device pointer
774 *
775 * In case of GPU reset failure make sure no process keep waiting on fence
776 * that will never complete.
777 */
778void radeon_fence_driver_force_completion(struct radeon_device *rdev)
779{
780 int ring;
781
782 for (ring = 0; ring < RADEON_NUM_RINGS; ring++) {
783 if (!rdev->fence_drv[ring].initialized)
784 continue;
785 radeon_fence_write(rdev, rdev->fence_drv[ring].sync_seq[ring], ring);
786 }
787}
788
789
790/*
791 * Fence debugfs
792 */
793#if defined(CONFIG_DEBUG_FS)
794static int radeon_debugfs_fence_info(struct seq_file *m, void *data)
795{
796 struct drm_info_node *node = (struct drm_info_node *)m->private;
797 struct drm_device *dev = node->minor->dev;
798 struct radeon_device *rdev = dev->dev_private;
799 int i, j;
800
801 for (i = 0; i < RADEON_NUM_RINGS; ++i) {
802 if (!rdev->fence_drv[i].initialized)
803 continue;
804
805 radeon_fence_process(rdev, i);
806
807 seq_printf(m, "--- ring %d ---\n", i);
808 seq_printf(m, "Last signaled fence 0x%016llx\n",
809 (unsigned long long)atomic64_read(&rdev->fence_drv[i].last_seq));
810 seq_printf(m, "Last emitted 0x%016llx\n",
811 rdev->fence_drv[i].sync_seq[i]);
812
813 for (j = 0; j < RADEON_NUM_RINGS; ++j) {
814 if (i != j && rdev->fence_drv[j].initialized)
815 seq_printf(m, "Last sync to ring %d 0x%016llx\n",
816 j, rdev->fence_drv[i].sync_seq[j]);
817 }
818 }
819 return 0;
820}
821
822static struct drm_info_list radeon_debugfs_fence_list[] = {
823 {"radeon_fence_info", &radeon_debugfs_fence_info, 0, NULL},
824};
825#endif
826
827int radeon_debugfs_fence_init(struct radeon_device *rdev)
828{
829#if defined(CONFIG_DEBUG_FS)
830 return radeon_debugfs_add_files(rdev, radeon_debugfs_fence_list, 1);
831#else
832 return 0;
833#endif
834}