Loading...
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * Copyright (C) 2013 Red Hat
4 * Author: Rob Clark <robdclark@gmail.com>
5 */
6
7#include "drm/drm_drv.h"
8
9#include "msm_gpu.h"
10#include "msm_gem.h"
11#include "msm_mmu.h"
12#include "msm_fence.h"
13#include "msm_gpu_trace.h"
14//#include "adreno/adreno_gpu.h"
15
16#include <generated/utsrelease.h>
17#include <linux/string_helpers.h>
18#include <linux/devcoredump.h>
19#include <linux/sched/task.h>
20
21/*
22 * Power Management:
23 */
24
25static int enable_pwrrail(struct msm_gpu *gpu)
26{
27 struct drm_device *dev = gpu->dev;
28 int ret = 0;
29
30 if (gpu->gpu_reg) {
31 ret = regulator_enable(gpu->gpu_reg);
32 if (ret) {
33 DRM_DEV_ERROR(dev->dev, "failed to enable 'gpu_reg': %d\n", ret);
34 return ret;
35 }
36 }
37
38 if (gpu->gpu_cx) {
39 ret = regulator_enable(gpu->gpu_cx);
40 if (ret) {
41 DRM_DEV_ERROR(dev->dev, "failed to enable 'gpu_cx': %d\n", ret);
42 return ret;
43 }
44 }
45
46 return 0;
47}
48
49static int disable_pwrrail(struct msm_gpu *gpu)
50{
51 if (gpu->gpu_cx)
52 regulator_disable(gpu->gpu_cx);
53 if (gpu->gpu_reg)
54 regulator_disable(gpu->gpu_reg);
55 return 0;
56}
57
58static int enable_clk(struct msm_gpu *gpu)
59{
60 if (gpu->core_clk && gpu->fast_rate)
61 dev_pm_opp_set_rate(&gpu->pdev->dev, gpu->fast_rate);
62
63 /* Set the RBBM timer rate to 19.2Mhz */
64 if (gpu->rbbmtimer_clk)
65 clk_set_rate(gpu->rbbmtimer_clk, 19200000);
66
67 return clk_bulk_prepare_enable(gpu->nr_clocks, gpu->grp_clks);
68}
69
70static int disable_clk(struct msm_gpu *gpu)
71{
72 clk_bulk_disable_unprepare(gpu->nr_clocks, gpu->grp_clks);
73
74 /*
75 * Set the clock to a deliberately low rate. On older targets the clock
76 * speed had to be non zero to avoid problems. On newer targets this
77 * will be rounded down to zero anyway so it all works out.
78 */
79 if (gpu->core_clk)
80 dev_pm_opp_set_rate(&gpu->pdev->dev, 27000000);
81
82 if (gpu->rbbmtimer_clk)
83 clk_set_rate(gpu->rbbmtimer_clk, 0);
84
85 return 0;
86}
87
88static int enable_axi(struct msm_gpu *gpu)
89{
90 return clk_prepare_enable(gpu->ebi1_clk);
91}
92
93static int disable_axi(struct msm_gpu *gpu)
94{
95 clk_disable_unprepare(gpu->ebi1_clk);
96 return 0;
97}
98
99int msm_gpu_pm_resume(struct msm_gpu *gpu)
100{
101 int ret;
102
103 DBG("%s", gpu->name);
104 trace_msm_gpu_resume(0);
105
106 ret = enable_pwrrail(gpu);
107 if (ret)
108 return ret;
109
110 ret = enable_clk(gpu);
111 if (ret)
112 return ret;
113
114 ret = enable_axi(gpu);
115 if (ret)
116 return ret;
117
118 msm_devfreq_resume(gpu);
119
120 gpu->needs_hw_init = true;
121
122 return 0;
123}
124
125int msm_gpu_pm_suspend(struct msm_gpu *gpu)
126{
127 int ret;
128
129 DBG("%s", gpu->name);
130 trace_msm_gpu_suspend(0);
131
132 msm_devfreq_suspend(gpu);
133
134 ret = disable_axi(gpu);
135 if (ret)
136 return ret;
137
138 ret = disable_clk(gpu);
139 if (ret)
140 return ret;
141
142 ret = disable_pwrrail(gpu);
143 if (ret)
144 return ret;
145
146 gpu->suspend_count++;
147
148 return 0;
149}
150
151void msm_gpu_show_fdinfo(struct msm_gpu *gpu, struct msm_file_private *ctx,
152 struct drm_printer *p)
153{
154 drm_printf(p, "drm-engine-gpu:\t%llu ns\n", ctx->elapsed_ns);
155 drm_printf(p, "drm-cycles-gpu:\t%llu\n", ctx->cycles);
156 drm_printf(p, "drm-maxfreq-gpu:\t%u Hz\n", gpu->fast_rate);
157}
158
159int msm_gpu_hw_init(struct msm_gpu *gpu)
160{
161 int ret;
162
163 WARN_ON(!mutex_is_locked(&gpu->lock));
164
165 if (!gpu->needs_hw_init)
166 return 0;
167
168 disable_irq(gpu->irq);
169 ret = gpu->funcs->hw_init(gpu);
170 if (!ret)
171 gpu->needs_hw_init = false;
172 enable_irq(gpu->irq);
173
174 return ret;
175}
176
177#ifdef CONFIG_DEV_COREDUMP
178static ssize_t msm_gpu_devcoredump_read(char *buffer, loff_t offset,
179 size_t count, void *data, size_t datalen)
180{
181 struct msm_gpu *gpu = data;
182 struct drm_print_iterator iter;
183 struct drm_printer p;
184 struct msm_gpu_state *state;
185
186 state = msm_gpu_crashstate_get(gpu);
187 if (!state)
188 return 0;
189
190 iter.data = buffer;
191 iter.offset = 0;
192 iter.start = offset;
193 iter.remain = count;
194
195 p = drm_coredump_printer(&iter);
196
197 drm_printf(&p, "---\n");
198 drm_printf(&p, "kernel: " UTS_RELEASE "\n");
199 drm_printf(&p, "module: " KBUILD_MODNAME "\n");
200 drm_printf(&p, "time: %lld.%09ld\n",
201 state->time.tv_sec, state->time.tv_nsec);
202 if (state->comm)
203 drm_printf(&p, "comm: %s\n", state->comm);
204 if (state->cmd)
205 drm_printf(&p, "cmdline: %s\n", state->cmd);
206
207 gpu->funcs->show(gpu, state, &p);
208
209 msm_gpu_crashstate_put(gpu);
210
211 return count - iter.remain;
212}
213
214static void msm_gpu_devcoredump_free(void *data)
215{
216 struct msm_gpu *gpu = data;
217
218 msm_gpu_crashstate_put(gpu);
219}
220
221static void msm_gpu_crashstate_get_bo(struct msm_gpu_state *state,
222 struct drm_gem_object *obj, u64 iova, bool full)
223{
224 struct msm_gpu_state_bo *state_bo = &state->bos[state->nr_bos];
225 struct msm_gem_object *msm_obj = to_msm_bo(obj);
226
227 /* Don't record write only objects */
228 state_bo->size = obj->size;
229 state_bo->flags = msm_obj->flags;
230 state_bo->iova = iova;
231
232 BUILD_BUG_ON(sizeof(state_bo->name) != sizeof(msm_obj->name));
233
234 memcpy(state_bo->name, msm_obj->name, sizeof(state_bo->name));
235
236 if (full) {
237 void *ptr;
238
239 state_bo->data = kvmalloc(obj->size, GFP_KERNEL);
240 if (!state_bo->data)
241 goto out;
242
243 msm_gem_lock(obj);
244 ptr = msm_gem_get_vaddr_active(obj);
245 msm_gem_unlock(obj);
246 if (IS_ERR(ptr)) {
247 kvfree(state_bo->data);
248 state_bo->data = NULL;
249 goto out;
250 }
251
252 memcpy(state_bo->data, ptr, obj->size);
253 msm_gem_put_vaddr(obj);
254 }
255out:
256 state->nr_bos++;
257}
258
259static void msm_gpu_crashstate_capture(struct msm_gpu *gpu,
260 struct msm_gem_submit *submit, char *comm, char *cmd)
261{
262 struct msm_gpu_state *state;
263
264 /* Check if the target supports capturing crash state */
265 if (!gpu->funcs->gpu_state_get)
266 return;
267
268 /* Only save one crash state at a time */
269 if (gpu->crashstate)
270 return;
271
272 state = gpu->funcs->gpu_state_get(gpu);
273 if (IS_ERR_OR_NULL(state))
274 return;
275
276 /* Fill in the additional crash state information */
277 state->comm = kstrdup(comm, GFP_KERNEL);
278 state->cmd = kstrdup(cmd, GFP_KERNEL);
279 state->fault_info = gpu->fault_info;
280
281 if (submit) {
282 int i;
283
284 state->bos = kcalloc(submit->nr_bos,
285 sizeof(struct msm_gpu_state_bo), GFP_KERNEL);
286
287 for (i = 0; state->bos && i < submit->nr_bos; i++) {
288 msm_gpu_crashstate_get_bo(state, submit->bos[i].obj,
289 submit->bos[i].iova,
290 should_dump(submit, i));
291 }
292 }
293
294 /* Set the active crash state to be dumped on failure */
295 gpu->crashstate = state;
296
297 dev_coredumpm(&gpu->pdev->dev, THIS_MODULE, gpu, 0, GFP_KERNEL,
298 msm_gpu_devcoredump_read, msm_gpu_devcoredump_free);
299}
300#else
301static void msm_gpu_crashstate_capture(struct msm_gpu *gpu,
302 struct msm_gem_submit *submit, char *comm, char *cmd)
303{
304}
305#endif
306
307/*
308 * Hangcheck detection for locked gpu:
309 */
310
311static struct msm_gem_submit *
312find_submit(struct msm_ringbuffer *ring, uint32_t fence)
313{
314 struct msm_gem_submit *submit;
315 unsigned long flags;
316
317 spin_lock_irqsave(&ring->submit_lock, flags);
318 list_for_each_entry(submit, &ring->submits, node) {
319 if (submit->seqno == fence) {
320 spin_unlock_irqrestore(&ring->submit_lock, flags);
321 return submit;
322 }
323 }
324 spin_unlock_irqrestore(&ring->submit_lock, flags);
325
326 return NULL;
327}
328
329static void retire_submits(struct msm_gpu *gpu);
330
331static void get_comm_cmdline(struct msm_gem_submit *submit, char **comm, char **cmd)
332{
333 struct msm_file_private *ctx = submit->queue->ctx;
334 struct task_struct *task;
335
336 WARN_ON(!mutex_is_locked(&submit->gpu->lock));
337
338 /* Note that kstrdup will return NULL if argument is NULL: */
339 *comm = kstrdup(ctx->comm, GFP_KERNEL);
340 *cmd = kstrdup(ctx->cmdline, GFP_KERNEL);
341
342 task = get_pid_task(submit->pid, PIDTYPE_PID);
343 if (!task)
344 return;
345
346 if (!*comm)
347 *comm = kstrdup(task->comm, GFP_KERNEL);
348
349 if (!*cmd)
350 *cmd = kstrdup_quotable_cmdline(task, GFP_KERNEL);
351
352 put_task_struct(task);
353}
354
355static void recover_worker(struct kthread_work *work)
356{
357 struct msm_gpu *gpu = container_of(work, struct msm_gpu, recover_work);
358 struct drm_device *dev = gpu->dev;
359 struct msm_drm_private *priv = dev->dev_private;
360 struct msm_gem_submit *submit;
361 struct msm_ringbuffer *cur_ring = gpu->funcs->active_ring(gpu);
362 char *comm = NULL, *cmd = NULL;
363 int i;
364
365 mutex_lock(&gpu->lock);
366
367 DRM_DEV_ERROR(dev->dev, "%s: hangcheck recover!\n", gpu->name);
368
369 submit = find_submit(cur_ring, cur_ring->memptrs->fence + 1);
370
371 /*
372 * If the submit retired while we were waiting for the worker to run,
373 * or waiting to acquire the gpu lock, then nothing more to do.
374 */
375 if (!submit)
376 goto out_unlock;
377
378 /* Increment the fault counts */
379 submit->queue->faults++;
380 if (submit->aspace)
381 submit->aspace->faults++;
382
383 get_comm_cmdline(submit, &comm, &cmd);
384
385 if (comm && cmd) {
386 DRM_DEV_ERROR(dev->dev, "%s: offending task: %s (%s)\n",
387 gpu->name, comm, cmd);
388
389 msm_rd_dump_submit(priv->hangrd, submit,
390 "offending task: %s (%s)", comm, cmd);
391 } else {
392 DRM_DEV_ERROR(dev->dev, "%s: offending task: unknown\n", gpu->name);
393
394 msm_rd_dump_submit(priv->hangrd, submit, NULL);
395 }
396
397 /* Record the crash state */
398 pm_runtime_get_sync(&gpu->pdev->dev);
399 msm_gpu_crashstate_capture(gpu, submit, comm, cmd);
400
401 kfree(cmd);
402 kfree(comm);
403
404 /*
405 * Update all the rings with the latest and greatest fence.. this
406 * needs to happen after msm_rd_dump_submit() to ensure that the
407 * bo's referenced by the offending submit are still around.
408 */
409 for (i = 0; i < gpu->nr_rings; i++) {
410 struct msm_ringbuffer *ring = gpu->rb[i];
411
412 uint32_t fence = ring->memptrs->fence;
413
414 /*
415 * For the current (faulting?) ring/submit advance the fence by
416 * one more to clear the faulting submit
417 */
418 if (ring == cur_ring)
419 ring->memptrs->fence = ++fence;
420
421 msm_update_fence(ring->fctx, fence);
422 }
423
424 if (msm_gpu_active(gpu)) {
425 /* retire completed submits, plus the one that hung: */
426 retire_submits(gpu);
427
428 gpu->funcs->recover(gpu);
429
430 /*
431 * Replay all remaining submits starting with highest priority
432 * ring
433 */
434 for (i = 0; i < gpu->nr_rings; i++) {
435 struct msm_ringbuffer *ring = gpu->rb[i];
436 unsigned long flags;
437
438 spin_lock_irqsave(&ring->submit_lock, flags);
439 list_for_each_entry(submit, &ring->submits, node)
440 gpu->funcs->submit(gpu, submit);
441 spin_unlock_irqrestore(&ring->submit_lock, flags);
442 }
443 }
444
445 pm_runtime_put(&gpu->pdev->dev);
446
447out_unlock:
448 mutex_unlock(&gpu->lock);
449
450 msm_gpu_retire(gpu);
451}
452
453static void fault_worker(struct kthread_work *work)
454{
455 struct msm_gpu *gpu = container_of(work, struct msm_gpu, fault_work);
456 struct msm_gem_submit *submit;
457 struct msm_ringbuffer *cur_ring = gpu->funcs->active_ring(gpu);
458 char *comm = NULL, *cmd = NULL;
459
460 mutex_lock(&gpu->lock);
461
462 submit = find_submit(cur_ring, cur_ring->memptrs->fence + 1);
463 if (submit && submit->fault_dumped)
464 goto resume_smmu;
465
466 if (submit) {
467 get_comm_cmdline(submit, &comm, &cmd);
468
469 /*
470 * When we get GPU iova faults, we can get 1000s of them,
471 * but we really only want to log the first one.
472 */
473 submit->fault_dumped = true;
474 }
475
476 /* Record the crash state */
477 pm_runtime_get_sync(&gpu->pdev->dev);
478 msm_gpu_crashstate_capture(gpu, submit, comm, cmd);
479 pm_runtime_put_sync(&gpu->pdev->dev);
480
481 kfree(cmd);
482 kfree(comm);
483
484resume_smmu:
485 memset(&gpu->fault_info, 0, sizeof(gpu->fault_info));
486 gpu->aspace->mmu->funcs->resume_translation(gpu->aspace->mmu);
487
488 mutex_unlock(&gpu->lock);
489}
490
491static void hangcheck_timer_reset(struct msm_gpu *gpu)
492{
493 struct msm_drm_private *priv = gpu->dev->dev_private;
494 mod_timer(&gpu->hangcheck_timer,
495 round_jiffies_up(jiffies + msecs_to_jiffies(priv->hangcheck_period)));
496}
497
498static bool made_progress(struct msm_gpu *gpu, struct msm_ringbuffer *ring)
499{
500 if (ring->hangcheck_progress_retries >= DRM_MSM_HANGCHECK_PROGRESS_RETRIES)
501 return false;
502
503 if (!gpu->funcs->progress)
504 return false;
505
506 if (!gpu->funcs->progress(gpu, ring))
507 return false;
508
509 ring->hangcheck_progress_retries++;
510 return true;
511}
512
513static void hangcheck_handler(struct timer_list *t)
514{
515 struct msm_gpu *gpu = from_timer(gpu, t, hangcheck_timer);
516 struct drm_device *dev = gpu->dev;
517 struct msm_ringbuffer *ring = gpu->funcs->active_ring(gpu);
518 uint32_t fence = ring->memptrs->fence;
519
520 if (fence != ring->hangcheck_fence) {
521 /* some progress has been made.. ya! */
522 ring->hangcheck_fence = fence;
523 ring->hangcheck_progress_retries = 0;
524 } else if (fence_before(fence, ring->fctx->last_fence) &&
525 !made_progress(gpu, ring)) {
526 /* no progress and not done.. hung! */
527 ring->hangcheck_fence = fence;
528 ring->hangcheck_progress_retries = 0;
529 DRM_DEV_ERROR(dev->dev, "%s: hangcheck detected gpu lockup rb %d!\n",
530 gpu->name, ring->id);
531 DRM_DEV_ERROR(dev->dev, "%s: completed fence: %u\n",
532 gpu->name, fence);
533 DRM_DEV_ERROR(dev->dev, "%s: submitted fence: %u\n",
534 gpu->name, ring->fctx->last_fence);
535
536 kthread_queue_work(gpu->worker, &gpu->recover_work);
537 }
538
539 /* if still more pending work, reset the hangcheck timer: */
540 if (fence_after(ring->fctx->last_fence, ring->hangcheck_fence))
541 hangcheck_timer_reset(gpu);
542
543 /* workaround for missing irq: */
544 msm_gpu_retire(gpu);
545}
546
547/*
548 * Performance Counters:
549 */
550
551/* called under perf_lock */
552static int update_hw_cntrs(struct msm_gpu *gpu, uint32_t ncntrs, uint32_t *cntrs)
553{
554 uint32_t current_cntrs[ARRAY_SIZE(gpu->last_cntrs)];
555 int i, n = min(ncntrs, gpu->num_perfcntrs);
556
557 /* read current values: */
558 for (i = 0; i < gpu->num_perfcntrs; i++)
559 current_cntrs[i] = gpu_read(gpu, gpu->perfcntrs[i].sample_reg);
560
561 /* update cntrs: */
562 for (i = 0; i < n; i++)
563 cntrs[i] = current_cntrs[i] - gpu->last_cntrs[i];
564
565 /* save current values: */
566 for (i = 0; i < gpu->num_perfcntrs; i++)
567 gpu->last_cntrs[i] = current_cntrs[i];
568
569 return n;
570}
571
572static void update_sw_cntrs(struct msm_gpu *gpu)
573{
574 ktime_t time;
575 uint32_t elapsed;
576 unsigned long flags;
577
578 spin_lock_irqsave(&gpu->perf_lock, flags);
579 if (!gpu->perfcntr_active)
580 goto out;
581
582 time = ktime_get();
583 elapsed = ktime_to_us(ktime_sub(time, gpu->last_sample.time));
584
585 gpu->totaltime += elapsed;
586 if (gpu->last_sample.active)
587 gpu->activetime += elapsed;
588
589 gpu->last_sample.active = msm_gpu_active(gpu);
590 gpu->last_sample.time = time;
591
592out:
593 spin_unlock_irqrestore(&gpu->perf_lock, flags);
594}
595
596void msm_gpu_perfcntr_start(struct msm_gpu *gpu)
597{
598 unsigned long flags;
599
600 pm_runtime_get_sync(&gpu->pdev->dev);
601
602 spin_lock_irqsave(&gpu->perf_lock, flags);
603 /* we could dynamically enable/disable perfcntr registers too.. */
604 gpu->last_sample.active = msm_gpu_active(gpu);
605 gpu->last_sample.time = ktime_get();
606 gpu->activetime = gpu->totaltime = 0;
607 gpu->perfcntr_active = true;
608 update_hw_cntrs(gpu, 0, NULL);
609 spin_unlock_irqrestore(&gpu->perf_lock, flags);
610}
611
612void msm_gpu_perfcntr_stop(struct msm_gpu *gpu)
613{
614 gpu->perfcntr_active = false;
615 pm_runtime_put_sync(&gpu->pdev->dev);
616}
617
618/* returns -errno or # of cntrs sampled */
619int msm_gpu_perfcntr_sample(struct msm_gpu *gpu, uint32_t *activetime,
620 uint32_t *totaltime, uint32_t ncntrs, uint32_t *cntrs)
621{
622 unsigned long flags;
623 int ret;
624
625 spin_lock_irqsave(&gpu->perf_lock, flags);
626
627 if (!gpu->perfcntr_active) {
628 ret = -EINVAL;
629 goto out;
630 }
631
632 *activetime = gpu->activetime;
633 *totaltime = gpu->totaltime;
634
635 gpu->activetime = gpu->totaltime = 0;
636
637 ret = update_hw_cntrs(gpu, ncntrs, cntrs);
638
639out:
640 spin_unlock_irqrestore(&gpu->perf_lock, flags);
641
642 return ret;
643}
644
645/*
646 * Cmdstream submission/retirement:
647 */
648
649static void retire_submit(struct msm_gpu *gpu, struct msm_ringbuffer *ring,
650 struct msm_gem_submit *submit)
651{
652 int index = submit->seqno % MSM_GPU_SUBMIT_STATS_COUNT;
653 volatile struct msm_gpu_submit_stats *stats;
654 u64 elapsed, clock = 0, cycles;
655 unsigned long flags;
656
657 stats = &ring->memptrs->stats[index];
658 /* Convert 19.2Mhz alwayson ticks to nanoseconds for elapsed time */
659 elapsed = (stats->alwayson_end - stats->alwayson_start) * 10000;
660 do_div(elapsed, 192);
661
662 cycles = stats->cpcycles_end - stats->cpcycles_start;
663
664 /* Calculate the clock frequency from the number of CP cycles */
665 if (elapsed) {
666 clock = cycles * 1000;
667 do_div(clock, elapsed);
668 }
669
670 submit->queue->ctx->elapsed_ns += elapsed;
671 submit->queue->ctx->cycles += cycles;
672
673 trace_msm_gpu_submit_retired(submit, elapsed, clock,
674 stats->alwayson_start, stats->alwayson_end);
675
676 msm_submit_retire(submit);
677
678 pm_runtime_mark_last_busy(&gpu->pdev->dev);
679
680 spin_lock_irqsave(&ring->submit_lock, flags);
681 list_del(&submit->node);
682 spin_unlock_irqrestore(&ring->submit_lock, flags);
683
684 /* Update devfreq on transition from active->idle: */
685 mutex_lock(&gpu->active_lock);
686 gpu->active_submits--;
687 WARN_ON(gpu->active_submits < 0);
688 if (!gpu->active_submits) {
689 msm_devfreq_idle(gpu);
690 pm_runtime_put_autosuspend(&gpu->pdev->dev);
691 }
692
693 mutex_unlock(&gpu->active_lock);
694
695 msm_gem_submit_put(submit);
696}
697
698static void retire_submits(struct msm_gpu *gpu)
699{
700 int i;
701
702 /* Retire the commits starting with highest priority */
703 for (i = 0; i < gpu->nr_rings; i++) {
704 struct msm_ringbuffer *ring = gpu->rb[i];
705
706 while (true) {
707 struct msm_gem_submit *submit = NULL;
708 unsigned long flags;
709
710 spin_lock_irqsave(&ring->submit_lock, flags);
711 submit = list_first_entry_or_null(&ring->submits,
712 struct msm_gem_submit, node);
713 spin_unlock_irqrestore(&ring->submit_lock, flags);
714
715 /*
716 * If no submit, we are done. If submit->fence hasn't
717 * been signalled, then later submits are not signalled
718 * either, so we are also done.
719 */
720 if (submit && dma_fence_is_signaled(submit->hw_fence)) {
721 retire_submit(gpu, ring, submit);
722 } else {
723 break;
724 }
725 }
726 }
727
728 wake_up_all(&gpu->retire_event);
729}
730
731static void retire_worker(struct kthread_work *work)
732{
733 struct msm_gpu *gpu = container_of(work, struct msm_gpu, retire_work);
734
735 retire_submits(gpu);
736}
737
738/* call from irq handler to schedule work to retire bo's */
739void msm_gpu_retire(struct msm_gpu *gpu)
740{
741 int i;
742
743 for (i = 0; i < gpu->nr_rings; i++)
744 msm_update_fence(gpu->rb[i]->fctx, gpu->rb[i]->memptrs->fence);
745
746 kthread_queue_work(gpu->worker, &gpu->retire_work);
747 update_sw_cntrs(gpu);
748}
749
750/* add bo's to gpu's ring, and kick gpu: */
751void msm_gpu_submit(struct msm_gpu *gpu, struct msm_gem_submit *submit)
752{
753 struct msm_ringbuffer *ring = submit->ring;
754 unsigned long flags;
755
756 WARN_ON(!mutex_is_locked(&gpu->lock));
757
758 pm_runtime_get_sync(&gpu->pdev->dev);
759
760 msm_gpu_hw_init(gpu);
761
762 submit->seqno = submit->hw_fence->seqno;
763
764 update_sw_cntrs(gpu);
765
766 /*
767 * ring->submits holds a ref to the submit, to deal with the case
768 * that a submit completes before msm_ioctl_gem_submit() returns.
769 */
770 msm_gem_submit_get(submit);
771
772 spin_lock_irqsave(&ring->submit_lock, flags);
773 list_add_tail(&submit->node, &ring->submits);
774 spin_unlock_irqrestore(&ring->submit_lock, flags);
775
776 /* Update devfreq on transition from idle->active: */
777 mutex_lock(&gpu->active_lock);
778 if (!gpu->active_submits) {
779 pm_runtime_get(&gpu->pdev->dev);
780 msm_devfreq_active(gpu);
781 }
782 gpu->active_submits++;
783 mutex_unlock(&gpu->active_lock);
784
785 gpu->funcs->submit(gpu, submit);
786 submit->ring->cur_ctx_seqno = submit->queue->ctx->seqno;
787
788 pm_runtime_put(&gpu->pdev->dev);
789 hangcheck_timer_reset(gpu);
790}
791
792/*
793 * Init/Cleanup:
794 */
795
796static irqreturn_t irq_handler(int irq, void *data)
797{
798 struct msm_gpu *gpu = data;
799 return gpu->funcs->irq(gpu);
800}
801
802static int get_clocks(struct platform_device *pdev, struct msm_gpu *gpu)
803{
804 int ret = devm_clk_bulk_get_all(&pdev->dev, &gpu->grp_clks);
805
806 if (ret < 1) {
807 gpu->nr_clocks = 0;
808 return ret;
809 }
810
811 gpu->nr_clocks = ret;
812
813 gpu->core_clk = msm_clk_bulk_get_clock(gpu->grp_clks,
814 gpu->nr_clocks, "core");
815
816 gpu->rbbmtimer_clk = msm_clk_bulk_get_clock(gpu->grp_clks,
817 gpu->nr_clocks, "rbbmtimer");
818
819 return 0;
820}
821
822/* Return a new address space for a msm_drm_private instance */
823struct msm_gem_address_space *
824msm_gpu_create_private_address_space(struct msm_gpu *gpu, struct task_struct *task)
825{
826 struct msm_gem_address_space *aspace = NULL;
827 if (!gpu)
828 return NULL;
829
830 /*
831 * If the target doesn't support private address spaces then return
832 * the global one
833 */
834 if (gpu->funcs->create_private_address_space) {
835 aspace = gpu->funcs->create_private_address_space(gpu);
836 if (!IS_ERR(aspace))
837 aspace->pid = get_pid(task_pid(task));
838 }
839
840 if (IS_ERR_OR_NULL(aspace))
841 aspace = msm_gem_address_space_get(gpu->aspace);
842
843 return aspace;
844}
845
846int msm_gpu_init(struct drm_device *drm, struct platform_device *pdev,
847 struct msm_gpu *gpu, const struct msm_gpu_funcs *funcs,
848 const char *name, struct msm_gpu_config *config)
849{
850 struct msm_drm_private *priv = drm->dev_private;
851 int i, ret, nr_rings = config->nr_rings;
852 void *memptrs;
853 uint64_t memptrs_iova;
854
855 if (WARN_ON(gpu->num_perfcntrs > ARRAY_SIZE(gpu->last_cntrs)))
856 gpu->num_perfcntrs = ARRAY_SIZE(gpu->last_cntrs);
857
858 gpu->dev = drm;
859 gpu->funcs = funcs;
860 gpu->name = name;
861
862 gpu->worker = kthread_create_worker(0, "gpu-worker");
863 if (IS_ERR(gpu->worker)) {
864 ret = PTR_ERR(gpu->worker);
865 gpu->worker = NULL;
866 goto fail;
867 }
868
869 sched_set_fifo_low(gpu->worker->task);
870
871 mutex_init(&gpu->active_lock);
872 mutex_init(&gpu->lock);
873 init_waitqueue_head(&gpu->retire_event);
874 kthread_init_work(&gpu->retire_work, retire_worker);
875 kthread_init_work(&gpu->recover_work, recover_worker);
876 kthread_init_work(&gpu->fault_work, fault_worker);
877
878 priv->hangcheck_period = DRM_MSM_HANGCHECK_DEFAULT_PERIOD;
879
880 /*
881 * If progress detection is supported, halve the hangcheck timer
882 * duration, as it takes two iterations of the hangcheck handler
883 * to detect a hang.
884 */
885 if (funcs->progress)
886 priv->hangcheck_period /= 2;
887
888 timer_setup(&gpu->hangcheck_timer, hangcheck_handler, 0);
889
890 spin_lock_init(&gpu->perf_lock);
891
892
893 /* Map registers: */
894 gpu->mmio = msm_ioremap(pdev, config->ioname);
895 if (IS_ERR(gpu->mmio)) {
896 ret = PTR_ERR(gpu->mmio);
897 goto fail;
898 }
899
900 /* Get Interrupt: */
901 gpu->irq = platform_get_irq(pdev, 0);
902 if (gpu->irq < 0) {
903 ret = gpu->irq;
904 goto fail;
905 }
906
907 ret = devm_request_irq(&pdev->dev, gpu->irq, irq_handler,
908 IRQF_TRIGGER_HIGH, "gpu-irq", gpu);
909 if (ret) {
910 DRM_DEV_ERROR(drm->dev, "failed to request IRQ%u: %d\n", gpu->irq, ret);
911 goto fail;
912 }
913
914 ret = get_clocks(pdev, gpu);
915 if (ret)
916 goto fail;
917
918 gpu->ebi1_clk = msm_clk_get(pdev, "bus");
919 DBG("ebi1_clk: %p", gpu->ebi1_clk);
920 if (IS_ERR(gpu->ebi1_clk))
921 gpu->ebi1_clk = NULL;
922
923 /* Acquire regulators: */
924 gpu->gpu_reg = devm_regulator_get(&pdev->dev, "vdd");
925 DBG("gpu_reg: %p", gpu->gpu_reg);
926 if (IS_ERR(gpu->gpu_reg))
927 gpu->gpu_reg = NULL;
928
929 gpu->gpu_cx = devm_regulator_get(&pdev->dev, "vddcx");
930 DBG("gpu_cx: %p", gpu->gpu_cx);
931 if (IS_ERR(gpu->gpu_cx))
932 gpu->gpu_cx = NULL;
933
934 platform_set_drvdata(pdev, &gpu->adreno_smmu);
935
936 msm_devfreq_init(gpu);
937
938
939 gpu->aspace = gpu->funcs->create_address_space(gpu, pdev);
940
941 if (gpu->aspace == NULL)
942 DRM_DEV_INFO(drm->dev, "%s: no IOMMU, fallback to VRAM carveout!\n", name);
943 else if (IS_ERR(gpu->aspace)) {
944 ret = PTR_ERR(gpu->aspace);
945 goto fail;
946 }
947
948 memptrs = msm_gem_kernel_new(drm,
949 sizeof(struct msm_rbmemptrs) * nr_rings,
950 check_apriv(gpu, MSM_BO_WC), gpu->aspace, &gpu->memptrs_bo,
951 &memptrs_iova);
952
953 if (IS_ERR(memptrs)) {
954 ret = PTR_ERR(memptrs);
955 DRM_DEV_ERROR(drm->dev, "could not allocate memptrs: %d\n", ret);
956 goto fail;
957 }
958
959 msm_gem_object_set_name(gpu->memptrs_bo, "memptrs");
960
961 if (nr_rings > ARRAY_SIZE(gpu->rb)) {
962 DRM_DEV_INFO_ONCE(drm->dev, "Only creating %zu ringbuffers\n",
963 ARRAY_SIZE(gpu->rb));
964 nr_rings = ARRAY_SIZE(gpu->rb);
965 }
966
967 /* Create ringbuffer(s): */
968 for (i = 0; i < nr_rings; i++) {
969 gpu->rb[i] = msm_ringbuffer_new(gpu, i, memptrs, memptrs_iova);
970
971 if (IS_ERR(gpu->rb[i])) {
972 ret = PTR_ERR(gpu->rb[i]);
973 DRM_DEV_ERROR(drm->dev,
974 "could not create ringbuffer %d: %d\n", i, ret);
975 goto fail;
976 }
977
978 memptrs += sizeof(struct msm_rbmemptrs);
979 memptrs_iova += sizeof(struct msm_rbmemptrs);
980 }
981
982 gpu->nr_rings = nr_rings;
983
984 refcount_set(&gpu->sysprof_active, 1);
985
986 return 0;
987
988fail:
989 for (i = 0; i < ARRAY_SIZE(gpu->rb); i++) {
990 msm_ringbuffer_destroy(gpu->rb[i]);
991 gpu->rb[i] = NULL;
992 }
993
994 msm_gem_kernel_put(gpu->memptrs_bo, gpu->aspace);
995
996 platform_set_drvdata(pdev, NULL);
997 return ret;
998}
999
1000void msm_gpu_cleanup(struct msm_gpu *gpu)
1001{
1002 int i;
1003
1004 DBG("%s", gpu->name);
1005
1006 for (i = 0; i < ARRAY_SIZE(gpu->rb); i++) {
1007 msm_ringbuffer_destroy(gpu->rb[i]);
1008 gpu->rb[i] = NULL;
1009 }
1010
1011 msm_gem_kernel_put(gpu->memptrs_bo, gpu->aspace);
1012
1013 if (!IS_ERR_OR_NULL(gpu->aspace)) {
1014 gpu->aspace->mmu->funcs->detach(gpu->aspace->mmu);
1015 msm_gem_address_space_put(gpu->aspace);
1016 }
1017
1018 if (gpu->worker) {
1019 kthread_destroy_worker(gpu->worker);
1020 }
1021
1022 msm_devfreq_cleanup(gpu);
1023
1024 platform_set_drvdata(gpu->pdev, NULL);
1025}
1/*
2 * Copyright (C) 2013 Red Hat
3 * Author: Rob Clark <robdclark@gmail.com>
4 *
5 * This program is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 as published by
7 * the Free Software Foundation.
8 *
9 * This program is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for
12 * more details.
13 *
14 * You should have received a copy of the GNU General Public License along with
15 * this program. If not, see <http://www.gnu.org/licenses/>.
16 */
17
18#include "msm_gpu.h"
19#include "msm_gem.h"
20#include "msm_mmu.h"
21#include "msm_fence.h"
22
23#include <linux/string_helpers.h>
24#include <linux/pm_opp.h>
25#include <linux/devfreq.h>
26
27
28/*
29 * Power Management:
30 */
31
32static int msm_devfreq_target(struct device *dev, unsigned long *freq,
33 u32 flags)
34{
35 struct msm_gpu *gpu = platform_get_drvdata(to_platform_device(dev));
36 struct dev_pm_opp *opp;
37
38 opp = devfreq_recommended_opp(dev, freq, flags);
39
40 if (IS_ERR(opp))
41 return PTR_ERR(opp);
42
43 clk_set_rate(gpu->core_clk, *freq);
44 dev_pm_opp_put(opp);
45
46 return 0;
47}
48
49static int msm_devfreq_get_dev_status(struct device *dev,
50 struct devfreq_dev_status *status)
51{
52 struct msm_gpu *gpu = platform_get_drvdata(to_platform_device(dev));
53 u64 cycles;
54 u32 freq = ((u32) status->current_frequency) / 1000000;
55 ktime_t time;
56
57 status->current_frequency = (unsigned long) clk_get_rate(gpu->core_clk);
58 gpu->funcs->gpu_busy(gpu, &cycles);
59
60 status->busy_time = ((u32) (cycles - gpu->devfreq.busy_cycles)) / freq;
61
62 gpu->devfreq.busy_cycles = cycles;
63
64 time = ktime_get();
65 status->total_time = ktime_us_delta(time, gpu->devfreq.time);
66 gpu->devfreq.time = time;
67
68 return 0;
69}
70
71static int msm_devfreq_get_cur_freq(struct device *dev, unsigned long *freq)
72{
73 struct msm_gpu *gpu = platform_get_drvdata(to_platform_device(dev));
74
75 *freq = (unsigned long) clk_get_rate(gpu->core_clk);
76
77 return 0;
78}
79
80static struct devfreq_dev_profile msm_devfreq_profile = {
81 .polling_ms = 10,
82 .target = msm_devfreq_target,
83 .get_dev_status = msm_devfreq_get_dev_status,
84 .get_cur_freq = msm_devfreq_get_cur_freq,
85};
86
87static void msm_devfreq_init(struct msm_gpu *gpu)
88{
89 /* We need target support to do devfreq */
90 if (!gpu->funcs->gpu_busy)
91 return;
92
93 msm_devfreq_profile.initial_freq = gpu->fast_rate;
94
95 /*
96 * Don't set the freq_table or max_state and let devfreq build the table
97 * from OPP
98 */
99
100 gpu->devfreq.devfreq = devm_devfreq_add_device(&gpu->pdev->dev,
101 &msm_devfreq_profile, "simple_ondemand", NULL);
102
103 if (IS_ERR(gpu->devfreq.devfreq)) {
104 dev_err(&gpu->pdev->dev, "Couldn't initialize GPU devfreq\n");
105 gpu->devfreq.devfreq = NULL;
106 }
107}
108
109static int enable_pwrrail(struct msm_gpu *gpu)
110{
111 struct drm_device *dev = gpu->dev;
112 int ret = 0;
113
114 if (gpu->gpu_reg) {
115 ret = regulator_enable(gpu->gpu_reg);
116 if (ret) {
117 dev_err(dev->dev, "failed to enable 'gpu_reg': %d\n", ret);
118 return ret;
119 }
120 }
121
122 if (gpu->gpu_cx) {
123 ret = regulator_enable(gpu->gpu_cx);
124 if (ret) {
125 dev_err(dev->dev, "failed to enable 'gpu_cx': %d\n", ret);
126 return ret;
127 }
128 }
129
130 return 0;
131}
132
133static int disable_pwrrail(struct msm_gpu *gpu)
134{
135 if (gpu->gpu_cx)
136 regulator_disable(gpu->gpu_cx);
137 if (gpu->gpu_reg)
138 regulator_disable(gpu->gpu_reg);
139 return 0;
140}
141
142static int enable_clk(struct msm_gpu *gpu)
143{
144 int i;
145
146 if (gpu->core_clk && gpu->fast_rate)
147 clk_set_rate(gpu->core_clk, gpu->fast_rate);
148
149 /* Set the RBBM timer rate to 19.2Mhz */
150 if (gpu->rbbmtimer_clk)
151 clk_set_rate(gpu->rbbmtimer_clk, 19200000);
152
153 for (i = gpu->nr_clocks - 1; i >= 0; i--)
154 if (gpu->grp_clks[i])
155 clk_prepare(gpu->grp_clks[i]);
156
157 for (i = gpu->nr_clocks - 1; i >= 0; i--)
158 if (gpu->grp_clks[i])
159 clk_enable(gpu->grp_clks[i]);
160
161 return 0;
162}
163
164static int disable_clk(struct msm_gpu *gpu)
165{
166 int i;
167
168 for (i = gpu->nr_clocks - 1; i >= 0; i--)
169 if (gpu->grp_clks[i])
170 clk_disable(gpu->grp_clks[i]);
171
172 for (i = gpu->nr_clocks - 1; i >= 0; i--)
173 if (gpu->grp_clks[i])
174 clk_unprepare(gpu->grp_clks[i]);
175
176 /*
177 * Set the clock to a deliberately low rate. On older targets the clock
178 * speed had to be non zero to avoid problems. On newer targets this
179 * will be rounded down to zero anyway so it all works out.
180 */
181 if (gpu->core_clk)
182 clk_set_rate(gpu->core_clk, 27000000);
183
184 if (gpu->rbbmtimer_clk)
185 clk_set_rate(gpu->rbbmtimer_clk, 0);
186
187 return 0;
188}
189
190static int enable_axi(struct msm_gpu *gpu)
191{
192 if (gpu->ebi1_clk)
193 clk_prepare_enable(gpu->ebi1_clk);
194 return 0;
195}
196
197static int disable_axi(struct msm_gpu *gpu)
198{
199 if (gpu->ebi1_clk)
200 clk_disable_unprepare(gpu->ebi1_clk);
201 return 0;
202}
203
204int msm_gpu_pm_resume(struct msm_gpu *gpu)
205{
206 int ret;
207
208 DBG("%s", gpu->name);
209
210 ret = enable_pwrrail(gpu);
211 if (ret)
212 return ret;
213
214 ret = enable_clk(gpu);
215 if (ret)
216 return ret;
217
218 ret = enable_axi(gpu);
219 if (ret)
220 return ret;
221
222 if (gpu->devfreq.devfreq) {
223 gpu->devfreq.busy_cycles = 0;
224 gpu->devfreq.time = ktime_get();
225
226 devfreq_resume_device(gpu->devfreq.devfreq);
227 }
228
229 gpu->needs_hw_init = true;
230
231 return 0;
232}
233
234int msm_gpu_pm_suspend(struct msm_gpu *gpu)
235{
236 int ret;
237
238 DBG("%s", gpu->name);
239
240 if (gpu->devfreq.devfreq)
241 devfreq_suspend_device(gpu->devfreq.devfreq);
242
243 ret = disable_axi(gpu);
244 if (ret)
245 return ret;
246
247 ret = disable_clk(gpu);
248 if (ret)
249 return ret;
250
251 ret = disable_pwrrail(gpu);
252 if (ret)
253 return ret;
254
255 return 0;
256}
257
258int msm_gpu_hw_init(struct msm_gpu *gpu)
259{
260 int ret;
261
262 WARN_ON(!mutex_is_locked(&gpu->dev->struct_mutex));
263
264 if (!gpu->needs_hw_init)
265 return 0;
266
267 disable_irq(gpu->irq);
268 ret = gpu->funcs->hw_init(gpu);
269 if (!ret)
270 gpu->needs_hw_init = false;
271 enable_irq(gpu->irq);
272
273 return ret;
274}
275
276/*
277 * Hangcheck detection for locked gpu:
278 */
279
280static void update_fences(struct msm_gpu *gpu, struct msm_ringbuffer *ring,
281 uint32_t fence)
282{
283 struct msm_gem_submit *submit;
284
285 list_for_each_entry(submit, &ring->submits, node) {
286 if (submit->seqno > fence)
287 break;
288
289 msm_update_fence(submit->ring->fctx,
290 submit->fence->seqno);
291 }
292}
293
294static struct msm_gem_submit *
295find_submit(struct msm_ringbuffer *ring, uint32_t fence)
296{
297 struct msm_gem_submit *submit;
298
299 WARN_ON(!mutex_is_locked(&ring->gpu->dev->struct_mutex));
300
301 list_for_each_entry(submit, &ring->submits, node)
302 if (submit->seqno == fence)
303 return submit;
304
305 return NULL;
306}
307
308static void retire_submits(struct msm_gpu *gpu);
309
310static void recover_worker(struct work_struct *work)
311{
312 struct msm_gpu *gpu = container_of(work, struct msm_gpu, recover_work);
313 struct drm_device *dev = gpu->dev;
314 struct msm_drm_private *priv = dev->dev_private;
315 struct msm_gem_submit *submit;
316 struct msm_ringbuffer *cur_ring = gpu->funcs->active_ring(gpu);
317 int i;
318
319 mutex_lock(&dev->struct_mutex);
320
321 dev_err(dev->dev, "%s: hangcheck recover!\n", gpu->name);
322
323 submit = find_submit(cur_ring, cur_ring->memptrs->fence + 1);
324 if (submit) {
325 struct task_struct *task;
326
327 rcu_read_lock();
328 task = pid_task(submit->pid, PIDTYPE_PID);
329 if (task) {
330 char *cmd;
331
332 /*
333 * So slightly annoying, in other paths like
334 * mmap'ing gem buffers, mmap_sem is acquired
335 * before struct_mutex, which means we can't
336 * hold struct_mutex across the call to
337 * get_cmdline(). But submits are retired
338 * from the same in-order workqueue, so we can
339 * safely drop the lock here without worrying
340 * about the submit going away.
341 */
342 mutex_unlock(&dev->struct_mutex);
343 cmd = kstrdup_quotable_cmdline(task, GFP_KERNEL);
344 mutex_lock(&dev->struct_mutex);
345
346 dev_err(dev->dev, "%s: offending task: %s (%s)\n",
347 gpu->name, task->comm, cmd);
348
349 msm_rd_dump_submit(priv->hangrd, submit,
350 "offending task: %s (%s)", task->comm, cmd);
351
352 kfree(cmd);
353 } else {
354 msm_rd_dump_submit(priv->hangrd, submit, NULL);
355 }
356 rcu_read_unlock();
357 }
358
359
360 /*
361 * Update all the rings with the latest and greatest fence.. this
362 * needs to happen after msm_rd_dump_submit() to ensure that the
363 * bo's referenced by the offending submit are still around.
364 */
365 for (i = 0; i < gpu->nr_rings; i++) {
366 struct msm_ringbuffer *ring = gpu->rb[i];
367
368 uint32_t fence = ring->memptrs->fence;
369
370 /*
371 * For the current (faulting?) ring/submit advance the fence by
372 * one more to clear the faulting submit
373 */
374 if (ring == cur_ring)
375 fence++;
376
377 update_fences(gpu, ring, fence);
378 }
379
380 if (msm_gpu_active(gpu)) {
381 /* retire completed submits, plus the one that hung: */
382 retire_submits(gpu);
383
384 pm_runtime_get_sync(&gpu->pdev->dev);
385 gpu->funcs->recover(gpu);
386 pm_runtime_put_sync(&gpu->pdev->dev);
387
388 /*
389 * Replay all remaining submits starting with highest priority
390 * ring
391 */
392 for (i = 0; i < gpu->nr_rings; i++) {
393 struct msm_ringbuffer *ring = gpu->rb[i];
394
395 list_for_each_entry(submit, &ring->submits, node)
396 gpu->funcs->submit(gpu, submit, NULL);
397 }
398 }
399
400 mutex_unlock(&dev->struct_mutex);
401
402 msm_gpu_retire(gpu);
403}
404
405static void hangcheck_timer_reset(struct msm_gpu *gpu)
406{
407 DBG("%s", gpu->name);
408 mod_timer(&gpu->hangcheck_timer,
409 round_jiffies_up(jiffies + DRM_MSM_HANGCHECK_JIFFIES));
410}
411
412static void hangcheck_handler(struct timer_list *t)
413{
414 struct msm_gpu *gpu = from_timer(gpu, t, hangcheck_timer);
415 struct drm_device *dev = gpu->dev;
416 struct msm_drm_private *priv = dev->dev_private;
417 struct msm_ringbuffer *ring = gpu->funcs->active_ring(gpu);
418 uint32_t fence = ring->memptrs->fence;
419
420 if (fence != ring->hangcheck_fence) {
421 /* some progress has been made.. ya! */
422 ring->hangcheck_fence = fence;
423 } else if (fence < ring->seqno) {
424 /* no progress and not done.. hung! */
425 ring->hangcheck_fence = fence;
426 dev_err(dev->dev, "%s: hangcheck detected gpu lockup rb %d!\n",
427 gpu->name, ring->id);
428 dev_err(dev->dev, "%s: completed fence: %u\n",
429 gpu->name, fence);
430 dev_err(dev->dev, "%s: submitted fence: %u\n",
431 gpu->name, ring->seqno);
432
433 queue_work(priv->wq, &gpu->recover_work);
434 }
435
436 /* if still more pending work, reset the hangcheck timer: */
437 if (ring->seqno > ring->hangcheck_fence)
438 hangcheck_timer_reset(gpu);
439
440 /* workaround for missing irq: */
441 queue_work(priv->wq, &gpu->retire_work);
442}
443
444/*
445 * Performance Counters:
446 */
447
448/* called under perf_lock */
449static int update_hw_cntrs(struct msm_gpu *gpu, uint32_t ncntrs, uint32_t *cntrs)
450{
451 uint32_t current_cntrs[ARRAY_SIZE(gpu->last_cntrs)];
452 int i, n = min(ncntrs, gpu->num_perfcntrs);
453
454 /* read current values: */
455 for (i = 0; i < gpu->num_perfcntrs; i++)
456 current_cntrs[i] = gpu_read(gpu, gpu->perfcntrs[i].sample_reg);
457
458 /* update cntrs: */
459 for (i = 0; i < n; i++)
460 cntrs[i] = current_cntrs[i] - gpu->last_cntrs[i];
461
462 /* save current values: */
463 for (i = 0; i < gpu->num_perfcntrs; i++)
464 gpu->last_cntrs[i] = current_cntrs[i];
465
466 return n;
467}
468
469static void update_sw_cntrs(struct msm_gpu *gpu)
470{
471 ktime_t time;
472 uint32_t elapsed;
473 unsigned long flags;
474
475 spin_lock_irqsave(&gpu->perf_lock, flags);
476 if (!gpu->perfcntr_active)
477 goto out;
478
479 time = ktime_get();
480 elapsed = ktime_to_us(ktime_sub(time, gpu->last_sample.time));
481
482 gpu->totaltime += elapsed;
483 if (gpu->last_sample.active)
484 gpu->activetime += elapsed;
485
486 gpu->last_sample.active = msm_gpu_active(gpu);
487 gpu->last_sample.time = time;
488
489out:
490 spin_unlock_irqrestore(&gpu->perf_lock, flags);
491}
492
493void msm_gpu_perfcntr_start(struct msm_gpu *gpu)
494{
495 unsigned long flags;
496
497 pm_runtime_get_sync(&gpu->pdev->dev);
498
499 spin_lock_irqsave(&gpu->perf_lock, flags);
500 /* we could dynamically enable/disable perfcntr registers too.. */
501 gpu->last_sample.active = msm_gpu_active(gpu);
502 gpu->last_sample.time = ktime_get();
503 gpu->activetime = gpu->totaltime = 0;
504 gpu->perfcntr_active = true;
505 update_hw_cntrs(gpu, 0, NULL);
506 spin_unlock_irqrestore(&gpu->perf_lock, flags);
507}
508
509void msm_gpu_perfcntr_stop(struct msm_gpu *gpu)
510{
511 gpu->perfcntr_active = false;
512 pm_runtime_put_sync(&gpu->pdev->dev);
513}
514
515/* returns -errno or # of cntrs sampled */
516int msm_gpu_perfcntr_sample(struct msm_gpu *gpu, uint32_t *activetime,
517 uint32_t *totaltime, uint32_t ncntrs, uint32_t *cntrs)
518{
519 unsigned long flags;
520 int ret;
521
522 spin_lock_irqsave(&gpu->perf_lock, flags);
523
524 if (!gpu->perfcntr_active) {
525 ret = -EINVAL;
526 goto out;
527 }
528
529 *activetime = gpu->activetime;
530 *totaltime = gpu->totaltime;
531
532 gpu->activetime = gpu->totaltime = 0;
533
534 ret = update_hw_cntrs(gpu, ncntrs, cntrs);
535
536out:
537 spin_unlock_irqrestore(&gpu->perf_lock, flags);
538
539 return ret;
540}
541
542/*
543 * Cmdstream submission/retirement:
544 */
545
546static void retire_submit(struct msm_gpu *gpu, struct msm_gem_submit *submit)
547{
548 int i;
549
550 for (i = 0; i < submit->nr_bos; i++) {
551 struct msm_gem_object *msm_obj = submit->bos[i].obj;
552 /* move to inactive: */
553 msm_gem_move_to_inactive(&msm_obj->base);
554 msm_gem_put_iova(&msm_obj->base, gpu->aspace);
555 drm_gem_object_put(&msm_obj->base);
556 }
557
558 pm_runtime_mark_last_busy(&gpu->pdev->dev);
559 pm_runtime_put_autosuspend(&gpu->pdev->dev);
560 msm_gem_submit_free(submit);
561}
562
563static void retire_submits(struct msm_gpu *gpu)
564{
565 struct drm_device *dev = gpu->dev;
566 struct msm_gem_submit *submit, *tmp;
567 int i;
568
569 WARN_ON(!mutex_is_locked(&dev->struct_mutex));
570
571 /* Retire the commits starting with highest priority */
572 for (i = 0; i < gpu->nr_rings; i++) {
573 struct msm_ringbuffer *ring = gpu->rb[i];
574
575 list_for_each_entry_safe(submit, tmp, &ring->submits, node) {
576 if (dma_fence_is_signaled(submit->fence))
577 retire_submit(gpu, submit);
578 }
579 }
580}
581
582static void retire_worker(struct work_struct *work)
583{
584 struct msm_gpu *gpu = container_of(work, struct msm_gpu, retire_work);
585 struct drm_device *dev = gpu->dev;
586 int i;
587
588 for (i = 0; i < gpu->nr_rings; i++)
589 update_fences(gpu, gpu->rb[i], gpu->rb[i]->memptrs->fence);
590
591 mutex_lock(&dev->struct_mutex);
592 retire_submits(gpu);
593 mutex_unlock(&dev->struct_mutex);
594}
595
596/* call from irq handler to schedule work to retire bo's */
597void msm_gpu_retire(struct msm_gpu *gpu)
598{
599 struct msm_drm_private *priv = gpu->dev->dev_private;
600 queue_work(priv->wq, &gpu->retire_work);
601 update_sw_cntrs(gpu);
602}
603
604/* add bo's to gpu's ring, and kick gpu: */
605void msm_gpu_submit(struct msm_gpu *gpu, struct msm_gem_submit *submit,
606 struct msm_file_private *ctx)
607{
608 struct drm_device *dev = gpu->dev;
609 struct msm_drm_private *priv = dev->dev_private;
610 struct msm_ringbuffer *ring = submit->ring;
611 int i;
612
613 WARN_ON(!mutex_is_locked(&dev->struct_mutex));
614
615 pm_runtime_get_sync(&gpu->pdev->dev);
616
617 msm_gpu_hw_init(gpu);
618
619 submit->seqno = ++ring->seqno;
620
621 list_add_tail(&submit->node, &ring->submits);
622
623 msm_rd_dump_submit(priv->rd, submit, NULL);
624
625 update_sw_cntrs(gpu);
626
627 for (i = 0; i < submit->nr_bos; i++) {
628 struct msm_gem_object *msm_obj = submit->bos[i].obj;
629 uint64_t iova;
630
631 /* can't happen yet.. but when we add 2d support we'll have
632 * to deal w/ cross-ring synchronization:
633 */
634 WARN_ON(is_active(msm_obj) && (msm_obj->gpu != gpu));
635
636 /* submit takes a reference to the bo and iova until retired: */
637 drm_gem_object_get(&msm_obj->base);
638 msm_gem_get_iova(&msm_obj->base,
639 submit->gpu->aspace, &iova);
640
641 if (submit->bos[i].flags & MSM_SUBMIT_BO_WRITE)
642 msm_gem_move_to_active(&msm_obj->base, gpu, true, submit->fence);
643 else if (submit->bos[i].flags & MSM_SUBMIT_BO_READ)
644 msm_gem_move_to_active(&msm_obj->base, gpu, false, submit->fence);
645 }
646
647 gpu->funcs->submit(gpu, submit, ctx);
648 priv->lastctx = ctx;
649
650 hangcheck_timer_reset(gpu);
651}
652
653/*
654 * Init/Cleanup:
655 */
656
657static irqreturn_t irq_handler(int irq, void *data)
658{
659 struct msm_gpu *gpu = data;
660 return gpu->funcs->irq(gpu);
661}
662
663static struct clk *get_clock(struct device *dev, const char *name)
664{
665 struct clk *clk = devm_clk_get(dev, name);
666
667 return IS_ERR(clk) ? NULL : clk;
668}
669
670static int get_clocks(struct platform_device *pdev, struct msm_gpu *gpu)
671{
672 struct device *dev = &pdev->dev;
673 struct property *prop;
674 const char *name;
675 int i = 0;
676
677 gpu->nr_clocks = of_property_count_strings(dev->of_node, "clock-names");
678 if (gpu->nr_clocks < 1) {
679 gpu->nr_clocks = 0;
680 return 0;
681 }
682
683 gpu->grp_clks = devm_kcalloc(dev, sizeof(struct clk *), gpu->nr_clocks,
684 GFP_KERNEL);
685 if (!gpu->grp_clks) {
686 gpu->nr_clocks = 0;
687 return -ENOMEM;
688 }
689
690 of_property_for_each_string(dev->of_node, "clock-names", prop, name) {
691 gpu->grp_clks[i] = get_clock(dev, name);
692
693 /* Remember the key clocks that we need to control later */
694 if (!strcmp(name, "core") || !strcmp(name, "core_clk"))
695 gpu->core_clk = gpu->grp_clks[i];
696 else if (!strcmp(name, "rbbmtimer") || !strcmp(name, "rbbmtimer_clk"))
697 gpu->rbbmtimer_clk = gpu->grp_clks[i];
698
699 ++i;
700 }
701
702 return 0;
703}
704
705static struct msm_gem_address_space *
706msm_gpu_create_address_space(struct msm_gpu *gpu, struct platform_device *pdev,
707 uint64_t va_start, uint64_t va_end)
708{
709 struct iommu_domain *iommu;
710 struct msm_gem_address_space *aspace;
711 int ret;
712
713 /*
714 * Setup IOMMU.. eventually we will (I think) do this once per context
715 * and have separate page tables per context. For now, to keep things
716 * simple and to get something working, just use a single address space:
717 */
718 iommu = iommu_domain_alloc(&platform_bus_type);
719 if (!iommu)
720 return NULL;
721
722 iommu->geometry.aperture_start = va_start;
723 iommu->geometry.aperture_end = va_end;
724
725 dev_info(gpu->dev->dev, "%s: using IOMMU\n", gpu->name);
726
727 aspace = msm_gem_address_space_create(&pdev->dev, iommu, "gpu");
728 if (IS_ERR(aspace)) {
729 dev_err(gpu->dev->dev, "failed to init iommu: %ld\n",
730 PTR_ERR(aspace));
731 iommu_domain_free(iommu);
732 return ERR_CAST(aspace);
733 }
734
735 ret = aspace->mmu->funcs->attach(aspace->mmu, NULL, 0);
736 if (ret) {
737 msm_gem_address_space_put(aspace);
738 return ERR_PTR(ret);
739 }
740
741 return aspace;
742}
743
744int msm_gpu_init(struct drm_device *drm, struct platform_device *pdev,
745 struct msm_gpu *gpu, const struct msm_gpu_funcs *funcs,
746 const char *name, struct msm_gpu_config *config)
747{
748 int i, ret, nr_rings = config->nr_rings;
749 void *memptrs;
750 uint64_t memptrs_iova;
751
752 if (WARN_ON(gpu->num_perfcntrs > ARRAY_SIZE(gpu->last_cntrs)))
753 gpu->num_perfcntrs = ARRAY_SIZE(gpu->last_cntrs);
754
755 gpu->dev = drm;
756 gpu->funcs = funcs;
757 gpu->name = name;
758
759 INIT_LIST_HEAD(&gpu->active_list);
760 INIT_WORK(&gpu->retire_work, retire_worker);
761 INIT_WORK(&gpu->recover_work, recover_worker);
762
763
764 timer_setup(&gpu->hangcheck_timer, hangcheck_handler, 0);
765
766 spin_lock_init(&gpu->perf_lock);
767
768
769 /* Map registers: */
770 gpu->mmio = msm_ioremap(pdev, config->ioname, name);
771 if (IS_ERR(gpu->mmio)) {
772 ret = PTR_ERR(gpu->mmio);
773 goto fail;
774 }
775
776 /* Get Interrupt: */
777 gpu->irq = platform_get_irq_byname(pdev, config->irqname);
778 if (gpu->irq < 0) {
779 ret = gpu->irq;
780 dev_err(drm->dev, "failed to get irq: %d\n", ret);
781 goto fail;
782 }
783
784 ret = devm_request_irq(&pdev->dev, gpu->irq, irq_handler,
785 IRQF_TRIGGER_HIGH, gpu->name, gpu);
786 if (ret) {
787 dev_err(drm->dev, "failed to request IRQ%u: %d\n", gpu->irq, ret);
788 goto fail;
789 }
790
791 ret = get_clocks(pdev, gpu);
792 if (ret)
793 goto fail;
794
795 gpu->ebi1_clk = msm_clk_get(pdev, "bus");
796 DBG("ebi1_clk: %p", gpu->ebi1_clk);
797 if (IS_ERR(gpu->ebi1_clk))
798 gpu->ebi1_clk = NULL;
799
800 /* Acquire regulators: */
801 gpu->gpu_reg = devm_regulator_get(&pdev->dev, "vdd");
802 DBG("gpu_reg: %p", gpu->gpu_reg);
803 if (IS_ERR(gpu->gpu_reg))
804 gpu->gpu_reg = NULL;
805
806 gpu->gpu_cx = devm_regulator_get(&pdev->dev, "vddcx");
807 DBG("gpu_cx: %p", gpu->gpu_cx);
808 if (IS_ERR(gpu->gpu_cx))
809 gpu->gpu_cx = NULL;
810
811 gpu->pdev = pdev;
812 platform_set_drvdata(pdev, gpu);
813
814 msm_devfreq_init(gpu);
815
816 gpu->aspace = msm_gpu_create_address_space(gpu, pdev,
817 config->va_start, config->va_end);
818
819 if (gpu->aspace == NULL)
820 dev_info(drm->dev, "%s: no IOMMU, fallback to VRAM carveout!\n", name);
821 else if (IS_ERR(gpu->aspace)) {
822 ret = PTR_ERR(gpu->aspace);
823 goto fail;
824 }
825
826 memptrs = msm_gem_kernel_new(drm, sizeof(*gpu->memptrs_bo),
827 MSM_BO_UNCACHED, gpu->aspace, &gpu->memptrs_bo,
828 &memptrs_iova);
829
830 if (IS_ERR(memptrs)) {
831 ret = PTR_ERR(memptrs);
832 dev_err(drm->dev, "could not allocate memptrs: %d\n", ret);
833 goto fail;
834 }
835
836 if (nr_rings > ARRAY_SIZE(gpu->rb)) {
837 DRM_DEV_INFO_ONCE(drm->dev, "Only creating %zu ringbuffers\n",
838 ARRAY_SIZE(gpu->rb));
839 nr_rings = ARRAY_SIZE(gpu->rb);
840 }
841
842 /* Create ringbuffer(s): */
843 for (i = 0; i < nr_rings; i++) {
844 gpu->rb[i] = msm_ringbuffer_new(gpu, i, memptrs, memptrs_iova);
845
846 if (IS_ERR(gpu->rb[i])) {
847 ret = PTR_ERR(gpu->rb[i]);
848 dev_err(drm->dev,
849 "could not create ringbuffer %d: %d\n", i, ret);
850 goto fail;
851 }
852
853 memptrs += sizeof(struct msm_rbmemptrs);
854 memptrs_iova += sizeof(struct msm_rbmemptrs);
855 }
856
857 gpu->nr_rings = nr_rings;
858
859 return 0;
860
861fail:
862 for (i = 0; i < ARRAY_SIZE(gpu->rb); i++) {
863 msm_ringbuffer_destroy(gpu->rb[i]);
864 gpu->rb[i] = NULL;
865 }
866
867 if (gpu->memptrs_bo) {
868 msm_gem_put_vaddr(gpu->memptrs_bo);
869 msm_gem_put_iova(gpu->memptrs_bo, gpu->aspace);
870 drm_gem_object_put_unlocked(gpu->memptrs_bo);
871 }
872
873 platform_set_drvdata(pdev, NULL);
874 return ret;
875}
876
877void msm_gpu_cleanup(struct msm_gpu *gpu)
878{
879 int i;
880
881 DBG("%s", gpu->name);
882
883 WARN_ON(!list_empty(&gpu->active_list));
884
885 for (i = 0; i < ARRAY_SIZE(gpu->rb); i++) {
886 msm_ringbuffer_destroy(gpu->rb[i]);
887 gpu->rb[i] = NULL;
888 }
889
890 if (gpu->memptrs_bo) {
891 msm_gem_put_vaddr(gpu->memptrs_bo);
892 msm_gem_put_iova(gpu->memptrs_bo, gpu->aspace);
893 drm_gem_object_put_unlocked(gpu->memptrs_bo);
894 }
895
896 if (!IS_ERR_OR_NULL(gpu->aspace)) {
897 gpu->aspace->mmu->funcs->detach(gpu->aspace->mmu,
898 NULL, 0);
899 msm_gem_address_space_put(gpu->aspace);
900 }
901}