Loading...
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}
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/reset.h>
20#include <linux/sched/task.h>
21
22/*
23 * Power Management:
24 */
25
26static int enable_pwrrail(struct msm_gpu *gpu)
27{
28 struct drm_device *dev = gpu->dev;
29 int ret = 0;
30
31 if (gpu->gpu_reg) {
32 ret = regulator_enable(gpu->gpu_reg);
33 if (ret) {
34 DRM_DEV_ERROR(dev->dev, "failed to enable 'gpu_reg': %d\n", ret);
35 return ret;
36 }
37 }
38
39 if (gpu->gpu_cx) {
40 ret = regulator_enable(gpu->gpu_cx);
41 if (ret) {
42 DRM_DEV_ERROR(dev->dev, "failed to enable 'gpu_cx': %d\n", ret);
43 return ret;
44 }
45 }
46
47 return 0;
48}
49
50static int disable_pwrrail(struct msm_gpu *gpu)
51{
52 if (gpu->gpu_cx)
53 regulator_disable(gpu->gpu_cx);
54 if (gpu->gpu_reg)
55 regulator_disable(gpu->gpu_reg);
56 return 0;
57}
58
59static int enable_clk(struct msm_gpu *gpu)
60{
61 if (gpu->core_clk && gpu->fast_rate)
62 clk_set_rate(gpu->core_clk, gpu->fast_rate);
63
64 /* Set the RBBM timer rate to 19.2Mhz */
65 if (gpu->rbbmtimer_clk)
66 clk_set_rate(gpu->rbbmtimer_clk, 19200000);
67
68 return clk_bulk_prepare_enable(gpu->nr_clocks, gpu->grp_clks);
69}
70
71static int disable_clk(struct msm_gpu *gpu)
72{
73 clk_bulk_disable_unprepare(gpu->nr_clocks, gpu->grp_clks);
74
75 /*
76 * Set the clock to a deliberately low rate. On older targets the clock
77 * speed had to be non zero to avoid problems. On newer targets this
78 * will be rounded down to zero anyway so it all works out.
79 */
80 if (gpu->core_clk)
81 clk_set_rate(gpu->core_clk, 27000000);
82
83 if (gpu->rbbmtimer_clk)
84 clk_set_rate(gpu->rbbmtimer_clk, 0);
85
86 return 0;
87}
88
89static int enable_axi(struct msm_gpu *gpu)
90{
91 return clk_prepare_enable(gpu->ebi1_clk);
92}
93
94static int disable_axi(struct msm_gpu *gpu)
95{
96 clk_disable_unprepare(gpu->ebi1_clk);
97 return 0;
98}
99
100int msm_gpu_pm_resume(struct msm_gpu *gpu)
101{
102 int ret;
103
104 DBG("%s", gpu->name);
105 trace_msm_gpu_resume(0);
106
107 ret = enable_pwrrail(gpu);
108 if (ret)
109 return ret;
110
111 ret = enable_clk(gpu);
112 if (ret)
113 return ret;
114
115 ret = enable_axi(gpu);
116 if (ret)
117 return ret;
118
119 msm_devfreq_resume(gpu);
120
121 gpu->needs_hw_init = true;
122
123 return 0;
124}
125
126int msm_gpu_pm_suspend(struct msm_gpu *gpu)
127{
128 int ret;
129
130 DBG("%s", gpu->name);
131 trace_msm_gpu_suspend(0);
132
133 msm_devfreq_suspend(gpu);
134
135 ret = disable_axi(gpu);
136 if (ret)
137 return ret;
138
139 ret = disable_clk(gpu);
140 if (ret)
141 return ret;
142
143 ret = disable_pwrrail(gpu);
144 if (ret)
145 return ret;
146
147 gpu->suspend_count++;
148
149 return 0;
150}
151
152void msm_gpu_show_fdinfo(struct msm_gpu *gpu, struct msm_file_private *ctx,
153 struct drm_printer *p)
154{
155 drm_printf(p, "drm-driver:\t%s\n", gpu->dev->driver->name);
156 drm_printf(p, "drm-client-id:\t%u\n", ctx->seqno);
157 drm_printf(p, "drm-engine-gpu:\t%llu ns\n", ctx->elapsed_ns);
158 drm_printf(p, "drm-cycles-gpu:\t%llu\n", ctx->cycles);
159 drm_printf(p, "drm-maxfreq-gpu:\t%u Hz\n", gpu->fast_rate);
160}
161
162int msm_gpu_hw_init(struct msm_gpu *gpu)
163{
164 int ret;
165
166 WARN_ON(!mutex_is_locked(&gpu->lock));
167
168 if (!gpu->needs_hw_init)
169 return 0;
170
171 disable_irq(gpu->irq);
172 ret = gpu->funcs->hw_init(gpu);
173 if (!ret)
174 gpu->needs_hw_init = false;
175 enable_irq(gpu->irq);
176
177 return ret;
178}
179
180#ifdef CONFIG_DEV_COREDUMP
181static ssize_t msm_gpu_devcoredump_read(char *buffer, loff_t offset,
182 size_t count, void *data, size_t datalen)
183{
184 struct msm_gpu *gpu = data;
185 struct drm_print_iterator iter;
186 struct drm_printer p;
187 struct msm_gpu_state *state;
188
189 state = msm_gpu_crashstate_get(gpu);
190 if (!state)
191 return 0;
192
193 iter.data = buffer;
194 iter.offset = 0;
195 iter.start = offset;
196 iter.remain = count;
197
198 p = drm_coredump_printer(&iter);
199
200 drm_printf(&p, "---\n");
201 drm_printf(&p, "kernel: " UTS_RELEASE "\n");
202 drm_printf(&p, "module: " KBUILD_MODNAME "\n");
203 drm_printf(&p, "time: %lld.%09ld\n",
204 state->time.tv_sec, state->time.tv_nsec);
205 if (state->comm)
206 drm_printf(&p, "comm: %s\n", state->comm);
207 if (state->cmd)
208 drm_printf(&p, "cmdline: %s\n", state->cmd);
209
210 gpu->funcs->show(gpu, state, &p);
211
212 msm_gpu_crashstate_put(gpu);
213
214 return count - iter.remain;
215}
216
217static void msm_gpu_devcoredump_free(void *data)
218{
219 struct msm_gpu *gpu = data;
220
221 msm_gpu_crashstate_put(gpu);
222}
223
224static void msm_gpu_crashstate_get_bo(struct msm_gpu_state *state,
225 struct msm_gem_object *obj, u64 iova, bool full)
226{
227 struct msm_gpu_state_bo *state_bo = &state->bos[state->nr_bos];
228
229 /* Don't record write only objects */
230 state_bo->size = obj->base.size;
231 state_bo->iova = iova;
232
233 BUILD_BUG_ON(sizeof(state_bo->name) != sizeof(obj->name));
234
235 memcpy(state_bo->name, obj->name, sizeof(state_bo->name));
236
237 if (full) {
238 void *ptr;
239
240 state_bo->data = kvmalloc(obj->base.size, GFP_KERNEL);
241 if (!state_bo->data)
242 goto out;
243
244 msm_gem_lock(&obj->base);
245 ptr = msm_gem_get_vaddr_active(&obj->base);
246 msm_gem_unlock(&obj->base);
247 if (IS_ERR(ptr)) {
248 kvfree(state_bo->data);
249 state_bo->data = NULL;
250 goto out;
251 }
252
253 memcpy(state_bo->data, ptr, obj->base.size);
254 msm_gem_put_vaddr(&obj->base);
255 }
256out:
257 state->nr_bos++;
258}
259
260static void msm_gpu_crashstate_capture(struct msm_gpu *gpu,
261 struct msm_gem_submit *submit, char *comm, char *cmd)
262{
263 struct msm_gpu_state *state;
264
265 /* Check if the target supports capturing crash state */
266 if (!gpu->funcs->gpu_state_get)
267 return;
268
269 /* Only save one crash state at a time */
270 if (gpu->crashstate)
271 return;
272
273 state = gpu->funcs->gpu_state_get(gpu);
274 if (IS_ERR_OR_NULL(state))
275 return;
276
277 /* Fill in the additional crash state information */
278 state->comm = kstrdup(comm, GFP_KERNEL);
279 state->cmd = kstrdup(cmd, GFP_KERNEL);
280 state->fault_info = gpu->fault_info;
281
282 if (submit) {
283 int i;
284
285 state->bos = kcalloc(submit->nr_bos,
286 sizeof(struct msm_gpu_state_bo), GFP_KERNEL);
287
288 for (i = 0; state->bos && i < submit->nr_bos; i++) {
289 msm_gpu_crashstate_get_bo(state, submit->bos[i].obj,
290 submit->bos[i].iova,
291 should_dump(submit, i));
292 }
293 }
294
295 /* Set the active crash state to be dumped on failure */
296 gpu->crashstate = state;
297
298 /* FIXME: Release the crashstate if this errors out? */
299 dev_coredumpm(gpu->dev->dev, THIS_MODULE, gpu, 0, GFP_KERNEL,
300 msm_gpu_devcoredump_read, msm_gpu_devcoredump_free);
301}
302#else
303static void msm_gpu_crashstate_capture(struct msm_gpu *gpu,
304 struct msm_gem_submit *submit, char *comm, char *cmd)
305{
306}
307#endif
308
309/*
310 * Hangcheck detection for locked gpu:
311 */
312
313static struct msm_gem_submit *
314find_submit(struct msm_ringbuffer *ring, uint32_t fence)
315{
316 struct msm_gem_submit *submit;
317 unsigned long flags;
318
319 spin_lock_irqsave(&ring->submit_lock, flags);
320 list_for_each_entry(submit, &ring->submits, node) {
321 if (submit->seqno == fence) {
322 spin_unlock_irqrestore(&ring->submit_lock, flags);
323 return submit;
324 }
325 }
326 spin_unlock_irqrestore(&ring->submit_lock, flags);
327
328 return NULL;
329}
330
331static void retire_submits(struct msm_gpu *gpu);
332
333static void get_comm_cmdline(struct msm_gem_submit *submit, char **comm, char **cmd)
334{
335 struct msm_file_private *ctx = submit->queue->ctx;
336 struct task_struct *task;
337
338 WARN_ON(!mutex_is_locked(&submit->gpu->lock));
339
340 /* Note that kstrdup will return NULL if argument is NULL: */
341 *comm = kstrdup(ctx->comm, GFP_KERNEL);
342 *cmd = kstrdup(ctx->cmdline, GFP_KERNEL);
343
344 task = get_pid_task(submit->pid, PIDTYPE_PID);
345 if (!task)
346 return;
347
348 if (!*comm)
349 *comm = kstrdup(task->comm, GFP_KERNEL);
350
351 if (!*cmd)
352 *cmd = kstrdup_quotable_cmdline(task, GFP_KERNEL);
353
354 put_task_struct(task);
355}
356
357static void recover_worker(struct kthread_work *work)
358{
359 struct msm_gpu *gpu = container_of(work, struct msm_gpu, recover_work);
360 struct drm_device *dev = gpu->dev;
361 struct msm_drm_private *priv = dev->dev_private;
362 struct msm_gem_submit *submit;
363 struct msm_ringbuffer *cur_ring = gpu->funcs->active_ring(gpu);
364 char *comm = NULL, *cmd = NULL;
365 int i;
366
367 mutex_lock(&gpu->lock);
368
369 DRM_DEV_ERROR(dev->dev, "%s: hangcheck recover!\n", gpu->name);
370
371 submit = find_submit(cur_ring, cur_ring->memptrs->fence + 1);
372 if (submit) {
373 /* Increment the fault counts */
374 submit->queue->faults++;
375 if (submit->aspace)
376 submit->aspace->faults++;
377
378 get_comm_cmdline(submit, &comm, &cmd);
379
380 if (comm && cmd) {
381 DRM_DEV_ERROR(dev->dev, "%s: offending task: %s (%s)\n",
382 gpu->name, comm, cmd);
383
384 msm_rd_dump_submit(priv->hangrd, submit,
385 "offending task: %s (%s)", comm, cmd);
386 } else {
387 msm_rd_dump_submit(priv->hangrd, submit, NULL);
388 }
389 } else {
390 /*
391 * We couldn't attribute this fault to any particular context,
392 * so increment the global fault count instead.
393 */
394 gpu->global_faults++;
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
447 mutex_unlock(&gpu->lock);
448
449 msm_gpu_retire(gpu);
450}
451
452static void fault_worker(struct kthread_work *work)
453{
454 struct msm_gpu *gpu = container_of(work, struct msm_gpu, fault_work);
455 struct msm_gem_submit *submit;
456 struct msm_ringbuffer *cur_ring = gpu->funcs->active_ring(gpu);
457 char *comm = NULL, *cmd = NULL;
458
459 mutex_lock(&gpu->lock);
460
461 submit = find_submit(cur_ring, cur_ring->memptrs->fence + 1);
462 if (submit && submit->fault_dumped)
463 goto resume_smmu;
464
465 if (submit) {
466 get_comm_cmdline(submit, &comm, &cmd);
467
468 /*
469 * When we get GPU iova faults, we can get 1000s of them,
470 * but we really only want to log the first one.
471 */
472 submit->fault_dumped = true;
473 }
474
475 /* Record the crash state */
476 pm_runtime_get_sync(&gpu->pdev->dev);
477 msm_gpu_crashstate_capture(gpu, submit, comm, cmd);
478 pm_runtime_put_sync(&gpu->pdev->dev);
479
480 kfree(cmd);
481 kfree(comm);
482
483resume_smmu:
484 memset(&gpu->fault_info, 0, sizeof(gpu->fault_info));
485 gpu->aspace->mmu->funcs->resume_translation(gpu->aspace->mmu);
486
487 mutex_unlock(&gpu->lock);
488}
489
490static void hangcheck_timer_reset(struct msm_gpu *gpu)
491{
492 struct msm_drm_private *priv = gpu->dev->dev_private;
493 mod_timer(&gpu->hangcheck_timer,
494 round_jiffies_up(jiffies + msecs_to_jiffies(priv->hangcheck_period)));
495}
496
497static bool made_progress(struct msm_gpu *gpu, struct msm_ringbuffer *ring)
498{
499 if (ring->hangcheck_progress_retries >= DRM_MSM_HANGCHECK_PROGRESS_RETRIES)
500 return false;
501
502 if (!gpu->funcs->progress)
503 return false;
504
505 if (!gpu->funcs->progress(gpu, ring))
506 return false;
507
508 ring->hangcheck_progress_retries++;
509 return true;
510}
511
512static void hangcheck_handler(struct timer_list *t)
513{
514 struct msm_gpu *gpu = from_timer(gpu, t, hangcheck_timer);
515 struct drm_device *dev = gpu->dev;
516 struct msm_ringbuffer *ring = gpu->funcs->active_ring(gpu);
517 uint32_t fence = ring->memptrs->fence;
518
519 if (fence != ring->hangcheck_fence) {
520 /* some progress has been made.. ya! */
521 ring->hangcheck_fence = fence;
522 ring->hangcheck_progress_retries = 0;
523 } else if (fence_before(fence, ring->fctx->last_fence) &&
524 !made_progress(gpu, ring)) {
525 /* no progress and not done.. hung! */
526 ring->hangcheck_fence = fence;
527 ring->hangcheck_progress_retries = 0;
528 DRM_DEV_ERROR(dev->dev, "%s: hangcheck detected gpu lockup rb %d!\n",
529 gpu->name, ring->id);
530 DRM_DEV_ERROR(dev->dev, "%s: completed fence: %u\n",
531 gpu->name, fence);
532 DRM_DEV_ERROR(dev->dev, "%s: submitted fence: %u\n",
533 gpu->name, ring->fctx->last_fence);
534
535 kthread_queue_work(gpu->worker, &gpu->recover_work);
536 }
537
538 /* if still more pending work, reset the hangcheck timer: */
539 if (fence_after(ring->fctx->last_fence, ring->hangcheck_fence))
540 hangcheck_timer_reset(gpu);
541
542 /* workaround for missing irq: */
543 msm_gpu_retire(gpu);
544}
545
546/*
547 * Performance Counters:
548 */
549
550/* called under perf_lock */
551static int update_hw_cntrs(struct msm_gpu *gpu, uint32_t ncntrs, uint32_t *cntrs)
552{
553 uint32_t current_cntrs[ARRAY_SIZE(gpu->last_cntrs)];
554 int i, n = min(ncntrs, gpu->num_perfcntrs);
555
556 /* read current values: */
557 for (i = 0; i < gpu->num_perfcntrs; i++)
558 current_cntrs[i] = gpu_read(gpu, gpu->perfcntrs[i].sample_reg);
559
560 /* update cntrs: */
561 for (i = 0; i < n; i++)
562 cntrs[i] = current_cntrs[i] - gpu->last_cntrs[i];
563
564 /* save current values: */
565 for (i = 0; i < gpu->num_perfcntrs; i++)
566 gpu->last_cntrs[i] = current_cntrs[i];
567
568 return n;
569}
570
571static void update_sw_cntrs(struct msm_gpu *gpu)
572{
573 ktime_t time;
574 uint32_t elapsed;
575 unsigned long flags;
576
577 spin_lock_irqsave(&gpu->perf_lock, flags);
578 if (!gpu->perfcntr_active)
579 goto out;
580
581 time = ktime_get();
582 elapsed = ktime_to_us(ktime_sub(time, gpu->last_sample.time));
583
584 gpu->totaltime += elapsed;
585 if (gpu->last_sample.active)
586 gpu->activetime += elapsed;
587
588 gpu->last_sample.active = msm_gpu_active(gpu);
589 gpu->last_sample.time = time;
590
591out:
592 spin_unlock_irqrestore(&gpu->perf_lock, flags);
593}
594
595void msm_gpu_perfcntr_start(struct msm_gpu *gpu)
596{
597 unsigned long flags;
598
599 pm_runtime_get_sync(&gpu->pdev->dev);
600
601 spin_lock_irqsave(&gpu->perf_lock, flags);
602 /* we could dynamically enable/disable perfcntr registers too.. */
603 gpu->last_sample.active = msm_gpu_active(gpu);
604 gpu->last_sample.time = ktime_get();
605 gpu->activetime = gpu->totaltime = 0;
606 gpu->perfcntr_active = true;
607 update_hw_cntrs(gpu, 0, NULL);
608 spin_unlock_irqrestore(&gpu->perf_lock, flags);
609}
610
611void msm_gpu_perfcntr_stop(struct msm_gpu *gpu)
612{
613 gpu->perfcntr_active = false;
614 pm_runtime_put_sync(&gpu->pdev->dev);
615}
616
617/* returns -errno or # of cntrs sampled */
618int msm_gpu_perfcntr_sample(struct msm_gpu *gpu, uint32_t *activetime,
619 uint32_t *totaltime, uint32_t ncntrs, uint32_t *cntrs)
620{
621 unsigned long flags;
622 int ret;
623
624 spin_lock_irqsave(&gpu->perf_lock, flags);
625
626 if (!gpu->perfcntr_active) {
627 ret = -EINVAL;
628 goto out;
629 }
630
631 *activetime = gpu->activetime;
632 *totaltime = gpu->totaltime;
633
634 gpu->activetime = gpu->totaltime = 0;
635
636 ret = update_hw_cntrs(gpu, ncntrs, cntrs);
637
638out:
639 spin_unlock_irqrestore(&gpu->perf_lock, flags);
640
641 return ret;
642}
643
644/*
645 * Cmdstream submission/retirement:
646 */
647
648static void retire_submit(struct msm_gpu *gpu, struct msm_ringbuffer *ring,
649 struct msm_gem_submit *submit)
650{
651 int index = submit->seqno % MSM_GPU_SUBMIT_STATS_COUNT;
652 volatile struct msm_gpu_submit_stats *stats;
653 u64 elapsed, clock = 0, cycles;
654 unsigned long flags;
655
656 stats = &ring->memptrs->stats[index];
657 /* Convert 19.2Mhz alwayson ticks to nanoseconds for elapsed time */
658 elapsed = (stats->alwayson_end - stats->alwayson_start) * 10000;
659 do_div(elapsed, 192);
660
661 cycles = stats->cpcycles_end - stats->cpcycles_start;
662
663 /* Calculate the clock frequency from the number of CP cycles */
664 if (elapsed) {
665 clock = cycles * 1000;
666 do_div(clock, elapsed);
667 }
668
669 submit->queue->ctx->elapsed_ns += elapsed;
670 submit->queue->ctx->cycles += cycles;
671
672 trace_msm_gpu_submit_retired(submit, elapsed, clock,
673 stats->alwayson_start, stats->alwayson_end);
674
675 msm_submit_retire(submit);
676
677 pm_runtime_mark_last_busy(&gpu->pdev->dev);
678
679 spin_lock_irqsave(&ring->submit_lock, flags);
680 list_del(&submit->node);
681 spin_unlock_irqrestore(&ring->submit_lock, flags);
682
683 /* Update devfreq on transition from active->idle: */
684 mutex_lock(&gpu->active_lock);
685 gpu->active_submits--;
686 WARN_ON(gpu->active_submits < 0);
687 if (!gpu->active_submits) {
688 msm_devfreq_idle(gpu);
689 pm_runtime_put_autosuspend(&gpu->pdev->dev);
690 }
691
692 mutex_unlock(&gpu->active_lock);
693
694 msm_gem_submit_put(submit);
695}
696
697static void retire_submits(struct msm_gpu *gpu)
698{
699 int i;
700
701 /* Retire the commits starting with highest priority */
702 for (i = 0; i < gpu->nr_rings; i++) {
703 struct msm_ringbuffer *ring = gpu->rb[i];
704
705 while (true) {
706 struct msm_gem_submit *submit = NULL;
707 unsigned long flags;
708
709 spin_lock_irqsave(&ring->submit_lock, flags);
710 submit = list_first_entry_or_null(&ring->submits,
711 struct msm_gem_submit, node);
712 spin_unlock_irqrestore(&ring->submit_lock, flags);
713
714 /*
715 * If no submit, we are done. If submit->fence hasn't
716 * been signalled, then later submits are not signalled
717 * either, so we are also done.
718 */
719 if (submit && dma_fence_is_signaled(submit->hw_fence)) {
720 retire_submit(gpu, ring, submit);
721 } else {
722 break;
723 }
724 }
725 }
726
727 wake_up_all(&gpu->retire_event);
728}
729
730static void retire_worker(struct kthread_work *work)
731{
732 struct msm_gpu *gpu = container_of(work, struct msm_gpu, retire_work);
733
734 retire_submits(gpu);
735}
736
737/* call from irq handler to schedule work to retire bo's */
738void msm_gpu_retire(struct msm_gpu *gpu)
739{
740 int i;
741
742 for (i = 0; i < gpu->nr_rings; i++)
743 msm_update_fence(gpu->rb[i]->fctx, gpu->rb[i]->memptrs->fence);
744
745 kthread_queue_work(gpu->worker, &gpu->retire_work);
746 update_sw_cntrs(gpu);
747}
748
749/* add bo's to gpu's ring, and kick gpu: */
750void msm_gpu_submit(struct msm_gpu *gpu, struct msm_gem_submit *submit)
751{
752 struct drm_device *dev = gpu->dev;
753 struct msm_drm_private *priv = dev->dev_private;
754 struct msm_ringbuffer *ring = submit->ring;
755 unsigned long flags;
756
757 WARN_ON(!mutex_is_locked(&gpu->lock));
758
759 pm_runtime_get_sync(&gpu->pdev->dev);
760
761 msm_gpu_hw_init(gpu);
762
763 submit->seqno = submit->hw_fence->seqno;
764
765 msm_rd_dump_submit(priv->rd, submit, NULL);
766
767 update_sw_cntrs(gpu);
768
769 /*
770 * ring->submits holds a ref to the submit, to deal with the case
771 * that a submit completes before msm_ioctl_gem_submit() returns.
772 */
773 msm_gem_submit_get(submit);
774
775 spin_lock_irqsave(&ring->submit_lock, flags);
776 list_add_tail(&submit->node, &ring->submits);
777 spin_unlock_irqrestore(&ring->submit_lock, flags);
778
779 /* Update devfreq on transition from idle->active: */
780 mutex_lock(&gpu->active_lock);
781 if (!gpu->active_submits) {
782 pm_runtime_get(&gpu->pdev->dev);
783 msm_devfreq_active(gpu);
784 }
785 gpu->active_submits++;
786 mutex_unlock(&gpu->active_lock);
787
788 gpu->funcs->submit(gpu, submit);
789 gpu->cur_ctx_seqno = submit->queue->ctx->seqno;
790
791 pm_runtime_put(&gpu->pdev->dev);
792 hangcheck_timer_reset(gpu);
793}
794
795/*
796 * Init/Cleanup:
797 */
798
799static irqreturn_t irq_handler(int irq, void *data)
800{
801 struct msm_gpu *gpu = data;
802 return gpu->funcs->irq(gpu);
803}
804
805static int get_clocks(struct platform_device *pdev, struct msm_gpu *gpu)
806{
807 int ret = devm_clk_bulk_get_all(&pdev->dev, &gpu->grp_clks);
808
809 if (ret < 1) {
810 gpu->nr_clocks = 0;
811 return ret;
812 }
813
814 gpu->nr_clocks = ret;
815
816 gpu->core_clk = msm_clk_bulk_get_clock(gpu->grp_clks,
817 gpu->nr_clocks, "core");
818
819 gpu->rbbmtimer_clk = msm_clk_bulk_get_clock(gpu->grp_clks,
820 gpu->nr_clocks, "rbbmtimer");
821
822 return 0;
823}
824
825/* Return a new address space for a msm_drm_private instance */
826struct msm_gem_address_space *
827msm_gpu_create_private_address_space(struct msm_gpu *gpu, struct task_struct *task)
828{
829 struct msm_gem_address_space *aspace = NULL;
830 if (!gpu)
831 return NULL;
832
833 /*
834 * If the target doesn't support private address spaces then return
835 * the global one
836 */
837 if (gpu->funcs->create_private_address_space) {
838 aspace = gpu->funcs->create_private_address_space(gpu);
839 if (!IS_ERR(aspace))
840 aspace->pid = get_pid(task_pid(task));
841 }
842
843 if (IS_ERR_OR_NULL(aspace))
844 aspace = msm_gem_address_space_get(gpu->aspace);
845
846 return aspace;
847}
848
849int msm_gpu_init(struct drm_device *drm, struct platform_device *pdev,
850 struct msm_gpu *gpu, const struct msm_gpu_funcs *funcs,
851 const char *name, struct msm_gpu_config *config)
852{
853 struct msm_drm_private *priv = drm->dev_private;
854 int i, ret, nr_rings = config->nr_rings;
855 void *memptrs;
856 uint64_t memptrs_iova;
857
858 if (WARN_ON(gpu->num_perfcntrs > ARRAY_SIZE(gpu->last_cntrs)))
859 gpu->num_perfcntrs = ARRAY_SIZE(gpu->last_cntrs);
860
861 gpu->dev = drm;
862 gpu->funcs = funcs;
863 gpu->name = name;
864
865 gpu->worker = kthread_create_worker(0, "gpu-worker");
866 if (IS_ERR(gpu->worker)) {
867 ret = PTR_ERR(gpu->worker);
868 gpu->worker = NULL;
869 goto fail;
870 }
871
872 sched_set_fifo_low(gpu->worker->task);
873
874 mutex_init(&gpu->active_lock);
875 mutex_init(&gpu->lock);
876 init_waitqueue_head(&gpu->retire_event);
877 kthread_init_work(&gpu->retire_work, retire_worker);
878 kthread_init_work(&gpu->recover_work, recover_worker);
879 kthread_init_work(&gpu->fault_work, fault_worker);
880
881 priv->hangcheck_period = DRM_MSM_HANGCHECK_DEFAULT_PERIOD;
882
883 /*
884 * If progress detection is supported, halve the hangcheck timer
885 * duration, as it takes two iterations of the hangcheck handler
886 * to detect a hang.
887 */
888 if (funcs->progress)
889 priv->hangcheck_period /= 2;
890
891 timer_setup(&gpu->hangcheck_timer, hangcheck_handler, 0);
892
893 spin_lock_init(&gpu->perf_lock);
894
895
896 /* Map registers: */
897 gpu->mmio = msm_ioremap(pdev, config->ioname);
898 if (IS_ERR(gpu->mmio)) {
899 ret = PTR_ERR(gpu->mmio);
900 goto fail;
901 }
902
903 /* Get Interrupt: */
904 gpu->irq = platform_get_irq(pdev, 0);
905 if (gpu->irq < 0) {
906 ret = gpu->irq;
907 DRM_DEV_ERROR(drm->dev, "failed to get irq: %d\n", ret);
908 goto fail;
909 }
910
911 ret = devm_request_irq(&pdev->dev, gpu->irq, irq_handler,
912 IRQF_TRIGGER_HIGH, "gpu-irq", gpu);
913 if (ret) {
914 DRM_DEV_ERROR(drm->dev, "failed to request IRQ%u: %d\n", gpu->irq, ret);
915 goto fail;
916 }
917
918 ret = get_clocks(pdev, gpu);
919 if (ret)
920 goto fail;
921
922 gpu->ebi1_clk = msm_clk_get(pdev, "bus");
923 DBG("ebi1_clk: %p", gpu->ebi1_clk);
924 if (IS_ERR(gpu->ebi1_clk))
925 gpu->ebi1_clk = NULL;
926
927 /* Acquire regulators: */
928 gpu->gpu_reg = devm_regulator_get(&pdev->dev, "vdd");
929 DBG("gpu_reg: %p", gpu->gpu_reg);
930 if (IS_ERR(gpu->gpu_reg))
931 gpu->gpu_reg = NULL;
932
933 gpu->gpu_cx = devm_regulator_get(&pdev->dev, "vddcx");
934 DBG("gpu_cx: %p", gpu->gpu_cx);
935 if (IS_ERR(gpu->gpu_cx))
936 gpu->gpu_cx = NULL;
937
938 gpu->cx_collapse = devm_reset_control_get_optional_exclusive(&pdev->dev,
939 "cx_collapse");
940
941 gpu->pdev = pdev;
942 platform_set_drvdata(pdev, &gpu->adreno_smmu);
943
944 msm_devfreq_init(gpu);
945
946
947 gpu->aspace = gpu->funcs->create_address_space(gpu, pdev);
948
949 if (gpu->aspace == NULL)
950 DRM_DEV_INFO(drm->dev, "%s: no IOMMU, fallback to VRAM carveout!\n", name);
951 else if (IS_ERR(gpu->aspace)) {
952 ret = PTR_ERR(gpu->aspace);
953 goto fail;
954 }
955
956 memptrs = msm_gem_kernel_new(drm,
957 sizeof(struct msm_rbmemptrs) * nr_rings,
958 check_apriv(gpu, MSM_BO_WC), gpu->aspace, &gpu->memptrs_bo,
959 &memptrs_iova);
960
961 if (IS_ERR(memptrs)) {
962 ret = PTR_ERR(memptrs);
963 DRM_DEV_ERROR(drm->dev, "could not allocate memptrs: %d\n", ret);
964 goto fail;
965 }
966
967 msm_gem_object_set_name(gpu->memptrs_bo, "memptrs");
968
969 if (nr_rings > ARRAY_SIZE(gpu->rb)) {
970 DRM_DEV_INFO_ONCE(drm->dev, "Only creating %zu ringbuffers\n",
971 ARRAY_SIZE(gpu->rb));
972 nr_rings = ARRAY_SIZE(gpu->rb);
973 }
974
975 /* Create ringbuffer(s): */
976 for (i = 0; i < nr_rings; i++) {
977 gpu->rb[i] = msm_ringbuffer_new(gpu, i, memptrs, memptrs_iova);
978
979 if (IS_ERR(gpu->rb[i])) {
980 ret = PTR_ERR(gpu->rb[i]);
981 DRM_DEV_ERROR(drm->dev,
982 "could not create ringbuffer %d: %d\n", i, ret);
983 goto fail;
984 }
985
986 memptrs += sizeof(struct msm_rbmemptrs);
987 memptrs_iova += sizeof(struct msm_rbmemptrs);
988 }
989
990 gpu->nr_rings = nr_rings;
991
992 refcount_set(&gpu->sysprof_active, 1);
993
994 return 0;
995
996fail:
997 for (i = 0; i < ARRAY_SIZE(gpu->rb); i++) {
998 msm_ringbuffer_destroy(gpu->rb[i]);
999 gpu->rb[i] = NULL;
1000 }
1001
1002 msm_gem_kernel_put(gpu->memptrs_bo, gpu->aspace);
1003
1004 platform_set_drvdata(pdev, NULL);
1005 return ret;
1006}
1007
1008void msm_gpu_cleanup(struct msm_gpu *gpu)
1009{
1010 int i;
1011
1012 DBG("%s", gpu->name);
1013
1014 for (i = 0; i < ARRAY_SIZE(gpu->rb); i++) {
1015 msm_ringbuffer_destroy(gpu->rb[i]);
1016 gpu->rb[i] = NULL;
1017 }
1018
1019 msm_gem_kernel_put(gpu->memptrs_bo, gpu->aspace);
1020
1021 if (!IS_ERR_OR_NULL(gpu->aspace)) {
1022 gpu->aspace->mmu->funcs->detach(gpu->aspace->mmu);
1023 msm_gem_address_space_put(gpu->aspace);
1024 }
1025
1026 if (gpu->worker) {
1027 kthread_destroy_worker(gpu->worker);
1028 }
1029
1030 msm_devfreq_cleanup(gpu);
1031
1032 platform_set_drvdata(gpu->pdev, NULL);
1033}