xref: /openbmc/linux/drivers/gpu/drm/msm/msm_gpu.c (revision ca637c0e)
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 
26 static 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 
50 static 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 
59 static 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 
71 static 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 
89 static int enable_axi(struct msm_gpu *gpu)
90 {
91 	return clk_prepare_enable(gpu->ebi1_clk);
92 }
93 
94 static int disable_axi(struct msm_gpu *gpu)
95 {
96 	clk_disable_unprepare(gpu->ebi1_clk);
97 	return 0;
98 }
99 
100 int 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 
126 int 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 
152 void 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 
162 int 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
181 static 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 
217 static void msm_gpu_devcoredump_free(void *data)
218 {
219 	struct msm_gpu *gpu = data;
220 
221 	msm_gpu_crashstate_put(gpu);
222 }
223 
224 static 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 	}
256 out:
257 	state->nr_bos++;
258 }
259 
260 static 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
303 static 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 
313 static struct msm_gem_submit *
314 find_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 
331 static void retire_submits(struct msm_gpu *gpu);
332 
333 static 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 	/* 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 
355 static 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 	if (submit) {
371 		/* Increment the fault counts */
372 		submit->queue->faults++;
373 		if (submit->aspace)
374 			submit->aspace->faults++;
375 
376 		get_comm_cmdline(submit, &comm, &cmd);
377 
378 		if (comm && cmd) {
379 			DRM_DEV_ERROR(dev->dev, "%s: offending task: %s (%s)\n",
380 				gpu->name, comm, cmd);
381 
382 			msm_rd_dump_submit(priv->hangrd, submit,
383 				"offending task: %s (%s)", comm, cmd);
384 		} else {
385 			msm_rd_dump_submit(priv->hangrd, submit, NULL);
386 		}
387 	} else {
388 		/*
389 		 * We couldn't attribute this fault to any particular context,
390 		 * so increment the global fault count instead.
391 		 */
392 		gpu->global_faults++;
393 	}
394 
395 	/* Record the crash state */
396 	pm_runtime_get_sync(&gpu->pdev->dev);
397 	msm_gpu_crashstate_capture(gpu, submit, comm, cmd);
398 
399 	kfree(cmd);
400 	kfree(comm);
401 
402 	/*
403 	 * Update all the rings with the latest and greatest fence.. this
404 	 * needs to happen after msm_rd_dump_submit() to ensure that the
405 	 * bo's referenced by the offending submit are still around.
406 	 */
407 	for (i = 0; i < gpu->nr_rings; i++) {
408 		struct msm_ringbuffer *ring = gpu->rb[i];
409 
410 		uint32_t fence = ring->memptrs->fence;
411 
412 		/*
413 		 * For the current (faulting?) ring/submit advance the fence by
414 		 * one more to clear the faulting submit
415 		 */
416 		if (ring == cur_ring)
417 			ring->memptrs->fence = ++fence;
418 
419 		msm_update_fence(ring->fctx, fence);
420 	}
421 
422 	if (msm_gpu_active(gpu)) {
423 		/* retire completed submits, plus the one that hung: */
424 		retire_submits(gpu);
425 
426 		gpu->funcs->recover(gpu);
427 
428 		/*
429 		 * Replay all remaining submits starting with highest priority
430 		 * ring
431 		 */
432 		for (i = 0; i < gpu->nr_rings; i++) {
433 			struct msm_ringbuffer *ring = gpu->rb[i];
434 			unsigned long flags;
435 
436 			spin_lock_irqsave(&ring->submit_lock, flags);
437 			list_for_each_entry(submit, &ring->submits, node)
438 				gpu->funcs->submit(gpu, submit);
439 			spin_unlock_irqrestore(&ring->submit_lock, flags);
440 		}
441 	}
442 
443 	pm_runtime_put(&gpu->pdev->dev);
444 
445 	mutex_unlock(&gpu->lock);
446 
447 	msm_gpu_retire(gpu);
448 }
449 
450 static void fault_worker(struct kthread_work *work)
451 {
452 	struct msm_gpu *gpu = container_of(work, struct msm_gpu, fault_work);
453 	struct msm_gem_submit *submit;
454 	struct msm_ringbuffer *cur_ring = gpu->funcs->active_ring(gpu);
455 	char *comm = NULL, *cmd = NULL;
456 
457 	mutex_lock(&gpu->lock);
458 
459 	submit = find_submit(cur_ring, cur_ring->memptrs->fence + 1);
460 	if (submit && submit->fault_dumped)
461 		goto resume_smmu;
462 
463 	if (submit) {
464 		get_comm_cmdline(submit, &comm, &cmd);
465 
466 		/*
467 		 * When we get GPU iova faults, we can get 1000s of them,
468 		 * but we really only want to log the first one.
469 		 */
470 		submit->fault_dumped = true;
471 	}
472 
473 	/* Record the crash state */
474 	pm_runtime_get_sync(&gpu->pdev->dev);
475 	msm_gpu_crashstate_capture(gpu, submit, comm, cmd);
476 	pm_runtime_put_sync(&gpu->pdev->dev);
477 
478 	kfree(cmd);
479 	kfree(comm);
480 
481 resume_smmu:
482 	memset(&gpu->fault_info, 0, sizeof(gpu->fault_info));
483 	gpu->aspace->mmu->funcs->resume_translation(gpu->aspace->mmu);
484 
485 	mutex_unlock(&gpu->lock);
486 }
487 
488 static void hangcheck_timer_reset(struct msm_gpu *gpu)
489 {
490 	struct msm_drm_private *priv = gpu->dev->dev_private;
491 	mod_timer(&gpu->hangcheck_timer,
492 			round_jiffies_up(jiffies + msecs_to_jiffies(priv->hangcheck_period)));
493 }
494 
495 static void hangcheck_handler(struct timer_list *t)
496 {
497 	struct msm_gpu *gpu = from_timer(gpu, t, hangcheck_timer);
498 	struct drm_device *dev = gpu->dev;
499 	struct msm_ringbuffer *ring = gpu->funcs->active_ring(gpu);
500 	uint32_t fence = ring->memptrs->fence;
501 
502 	if (fence != ring->hangcheck_fence) {
503 		/* some progress has been made.. ya! */
504 		ring->hangcheck_fence = fence;
505 	} else if (fence_before(fence, ring->fctx->last_fence)) {
506 		/* no progress and not done.. hung! */
507 		ring->hangcheck_fence = fence;
508 		DRM_DEV_ERROR(dev->dev, "%s: hangcheck detected gpu lockup rb %d!\n",
509 				gpu->name, ring->id);
510 		DRM_DEV_ERROR(dev->dev, "%s:     completed fence: %u\n",
511 				gpu->name, fence);
512 		DRM_DEV_ERROR(dev->dev, "%s:     submitted fence: %u\n",
513 				gpu->name, ring->fctx->last_fence);
514 
515 		kthread_queue_work(gpu->worker, &gpu->recover_work);
516 	}
517 
518 	/* if still more pending work, reset the hangcheck timer: */
519 	if (fence_after(ring->fctx->last_fence, ring->hangcheck_fence))
520 		hangcheck_timer_reset(gpu);
521 
522 	/* workaround for missing irq: */
523 	msm_gpu_retire(gpu);
524 }
525 
526 /*
527  * Performance Counters:
528  */
529 
530 /* called under perf_lock */
531 static int update_hw_cntrs(struct msm_gpu *gpu, uint32_t ncntrs, uint32_t *cntrs)
532 {
533 	uint32_t current_cntrs[ARRAY_SIZE(gpu->last_cntrs)];
534 	int i, n = min(ncntrs, gpu->num_perfcntrs);
535 
536 	/* read current values: */
537 	for (i = 0; i < gpu->num_perfcntrs; i++)
538 		current_cntrs[i] = gpu_read(gpu, gpu->perfcntrs[i].sample_reg);
539 
540 	/* update cntrs: */
541 	for (i = 0; i < n; i++)
542 		cntrs[i] = current_cntrs[i] - gpu->last_cntrs[i];
543 
544 	/* save current values: */
545 	for (i = 0; i < gpu->num_perfcntrs; i++)
546 		gpu->last_cntrs[i] = current_cntrs[i];
547 
548 	return n;
549 }
550 
551 static void update_sw_cntrs(struct msm_gpu *gpu)
552 {
553 	ktime_t time;
554 	uint32_t elapsed;
555 	unsigned long flags;
556 
557 	spin_lock_irqsave(&gpu->perf_lock, flags);
558 	if (!gpu->perfcntr_active)
559 		goto out;
560 
561 	time = ktime_get();
562 	elapsed = ktime_to_us(ktime_sub(time, gpu->last_sample.time));
563 
564 	gpu->totaltime += elapsed;
565 	if (gpu->last_sample.active)
566 		gpu->activetime += elapsed;
567 
568 	gpu->last_sample.active = msm_gpu_active(gpu);
569 	gpu->last_sample.time = time;
570 
571 out:
572 	spin_unlock_irqrestore(&gpu->perf_lock, flags);
573 }
574 
575 void msm_gpu_perfcntr_start(struct msm_gpu *gpu)
576 {
577 	unsigned long flags;
578 
579 	pm_runtime_get_sync(&gpu->pdev->dev);
580 
581 	spin_lock_irqsave(&gpu->perf_lock, flags);
582 	/* we could dynamically enable/disable perfcntr registers too.. */
583 	gpu->last_sample.active = msm_gpu_active(gpu);
584 	gpu->last_sample.time = ktime_get();
585 	gpu->activetime = gpu->totaltime = 0;
586 	gpu->perfcntr_active = true;
587 	update_hw_cntrs(gpu, 0, NULL);
588 	spin_unlock_irqrestore(&gpu->perf_lock, flags);
589 }
590 
591 void msm_gpu_perfcntr_stop(struct msm_gpu *gpu)
592 {
593 	gpu->perfcntr_active = false;
594 	pm_runtime_put_sync(&gpu->pdev->dev);
595 }
596 
597 /* returns -errno or # of cntrs sampled */
598 int msm_gpu_perfcntr_sample(struct msm_gpu *gpu, uint32_t *activetime,
599 		uint32_t *totaltime, uint32_t ncntrs, uint32_t *cntrs)
600 {
601 	unsigned long flags;
602 	int ret;
603 
604 	spin_lock_irqsave(&gpu->perf_lock, flags);
605 
606 	if (!gpu->perfcntr_active) {
607 		ret = -EINVAL;
608 		goto out;
609 	}
610 
611 	*activetime = gpu->activetime;
612 	*totaltime = gpu->totaltime;
613 
614 	gpu->activetime = gpu->totaltime = 0;
615 
616 	ret = update_hw_cntrs(gpu, ncntrs, cntrs);
617 
618 out:
619 	spin_unlock_irqrestore(&gpu->perf_lock, flags);
620 
621 	return ret;
622 }
623 
624 /*
625  * Cmdstream submission/retirement:
626  */
627 
628 static void retire_submit(struct msm_gpu *gpu, struct msm_ringbuffer *ring,
629 		struct msm_gem_submit *submit)
630 {
631 	int index = submit->seqno % MSM_GPU_SUBMIT_STATS_COUNT;
632 	volatile struct msm_gpu_submit_stats *stats;
633 	u64 elapsed, clock = 0, cycles;
634 	unsigned long flags;
635 
636 	stats = &ring->memptrs->stats[index];
637 	/* Convert 19.2Mhz alwayson ticks to nanoseconds for elapsed time */
638 	elapsed = (stats->alwayson_end - stats->alwayson_start) * 10000;
639 	do_div(elapsed, 192);
640 
641 	cycles = stats->cpcycles_end - stats->cpcycles_start;
642 
643 	/* Calculate the clock frequency from the number of CP cycles */
644 	if (elapsed) {
645 		clock = cycles * 1000;
646 		do_div(clock, elapsed);
647 	}
648 
649 	submit->queue->ctx->elapsed_ns += elapsed;
650 	submit->queue->ctx->cycles     += cycles;
651 
652 	trace_msm_gpu_submit_retired(submit, elapsed, clock,
653 		stats->alwayson_start, stats->alwayson_end);
654 
655 	msm_submit_retire(submit);
656 
657 	pm_runtime_mark_last_busy(&gpu->pdev->dev);
658 
659 	spin_lock_irqsave(&ring->submit_lock, flags);
660 	list_del(&submit->node);
661 	spin_unlock_irqrestore(&ring->submit_lock, flags);
662 
663 	/* Update devfreq on transition from active->idle: */
664 	mutex_lock(&gpu->active_lock);
665 	gpu->active_submits--;
666 	WARN_ON(gpu->active_submits < 0);
667 	if (!gpu->active_submits) {
668 		msm_devfreq_idle(gpu);
669 		pm_runtime_put_autosuspend(&gpu->pdev->dev);
670 	}
671 
672 	mutex_unlock(&gpu->active_lock);
673 
674 	msm_gem_submit_put(submit);
675 }
676 
677 static void retire_submits(struct msm_gpu *gpu)
678 {
679 	int i;
680 
681 	/* Retire the commits starting with highest priority */
682 	for (i = 0; i < gpu->nr_rings; i++) {
683 		struct msm_ringbuffer *ring = gpu->rb[i];
684 
685 		while (true) {
686 			struct msm_gem_submit *submit = NULL;
687 			unsigned long flags;
688 
689 			spin_lock_irqsave(&ring->submit_lock, flags);
690 			submit = list_first_entry_or_null(&ring->submits,
691 					struct msm_gem_submit, node);
692 			spin_unlock_irqrestore(&ring->submit_lock, flags);
693 
694 			/*
695 			 * If no submit, we are done.  If submit->fence hasn't
696 			 * been signalled, then later submits are not signalled
697 			 * either, so we are also done.
698 			 */
699 			if (submit && dma_fence_is_signaled(submit->hw_fence)) {
700 				retire_submit(gpu, ring, submit);
701 			} else {
702 				break;
703 			}
704 		}
705 	}
706 
707 	wake_up_all(&gpu->retire_event);
708 }
709 
710 static void retire_worker(struct kthread_work *work)
711 {
712 	struct msm_gpu *gpu = container_of(work, struct msm_gpu, retire_work);
713 
714 	retire_submits(gpu);
715 }
716 
717 /* call from irq handler to schedule work to retire bo's */
718 void msm_gpu_retire(struct msm_gpu *gpu)
719 {
720 	int i;
721 
722 	for (i = 0; i < gpu->nr_rings; i++)
723 		msm_update_fence(gpu->rb[i]->fctx, gpu->rb[i]->memptrs->fence);
724 
725 	kthread_queue_work(gpu->worker, &gpu->retire_work);
726 	update_sw_cntrs(gpu);
727 }
728 
729 /* add bo's to gpu's ring, and kick gpu: */
730 void msm_gpu_submit(struct msm_gpu *gpu, struct msm_gem_submit *submit)
731 {
732 	struct drm_device *dev = gpu->dev;
733 	struct msm_drm_private *priv = dev->dev_private;
734 	struct msm_ringbuffer *ring = submit->ring;
735 	unsigned long flags;
736 
737 	WARN_ON(!mutex_is_locked(&gpu->lock));
738 
739 	pm_runtime_get_sync(&gpu->pdev->dev);
740 
741 	msm_gpu_hw_init(gpu);
742 
743 	submit->seqno = submit->hw_fence->seqno;
744 
745 	msm_rd_dump_submit(priv->rd, submit, NULL);
746 
747 	update_sw_cntrs(gpu);
748 
749 	/*
750 	 * ring->submits holds a ref to the submit, to deal with the case
751 	 * that a submit completes before msm_ioctl_gem_submit() returns.
752 	 */
753 	msm_gem_submit_get(submit);
754 
755 	spin_lock_irqsave(&ring->submit_lock, flags);
756 	list_add_tail(&submit->node, &ring->submits);
757 	spin_unlock_irqrestore(&ring->submit_lock, flags);
758 
759 	/* Update devfreq on transition from idle->active: */
760 	mutex_lock(&gpu->active_lock);
761 	if (!gpu->active_submits) {
762 		pm_runtime_get(&gpu->pdev->dev);
763 		msm_devfreq_active(gpu);
764 	}
765 	gpu->active_submits++;
766 	mutex_unlock(&gpu->active_lock);
767 
768 	gpu->funcs->submit(gpu, submit);
769 	gpu->cur_ctx_seqno = submit->queue->ctx->seqno;
770 
771 	pm_runtime_put(&gpu->pdev->dev);
772 	hangcheck_timer_reset(gpu);
773 }
774 
775 /*
776  * Init/Cleanup:
777  */
778 
779 static irqreturn_t irq_handler(int irq, void *data)
780 {
781 	struct msm_gpu *gpu = data;
782 	return gpu->funcs->irq(gpu);
783 }
784 
785 static int get_clocks(struct platform_device *pdev, struct msm_gpu *gpu)
786 {
787 	int ret = devm_clk_bulk_get_all(&pdev->dev, &gpu->grp_clks);
788 
789 	if (ret < 1) {
790 		gpu->nr_clocks = 0;
791 		return ret;
792 	}
793 
794 	gpu->nr_clocks = ret;
795 
796 	gpu->core_clk = msm_clk_bulk_get_clock(gpu->grp_clks,
797 		gpu->nr_clocks, "core");
798 
799 	gpu->rbbmtimer_clk = msm_clk_bulk_get_clock(gpu->grp_clks,
800 		gpu->nr_clocks, "rbbmtimer");
801 
802 	return 0;
803 }
804 
805 /* Return a new address space for a msm_drm_private instance */
806 struct msm_gem_address_space *
807 msm_gpu_create_private_address_space(struct msm_gpu *gpu, struct task_struct *task)
808 {
809 	struct msm_gem_address_space *aspace = NULL;
810 	if (!gpu)
811 		return NULL;
812 
813 	/*
814 	 * If the target doesn't support private address spaces then return
815 	 * the global one
816 	 */
817 	if (gpu->funcs->create_private_address_space) {
818 		aspace = gpu->funcs->create_private_address_space(gpu);
819 		if (!IS_ERR(aspace))
820 			aspace->pid = get_pid(task_pid(task));
821 	}
822 
823 	if (IS_ERR_OR_NULL(aspace))
824 		aspace = msm_gem_address_space_get(gpu->aspace);
825 
826 	return aspace;
827 }
828 
829 int msm_gpu_init(struct drm_device *drm, struct platform_device *pdev,
830 		struct msm_gpu *gpu, const struct msm_gpu_funcs *funcs,
831 		const char *name, struct msm_gpu_config *config)
832 {
833 	int i, ret, nr_rings = config->nr_rings;
834 	void *memptrs;
835 	uint64_t memptrs_iova;
836 
837 	if (WARN_ON(gpu->num_perfcntrs > ARRAY_SIZE(gpu->last_cntrs)))
838 		gpu->num_perfcntrs = ARRAY_SIZE(gpu->last_cntrs);
839 
840 	gpu->dev = drm;
841 	gpu->funcs = funcs;
842 	gpu->name = name;
843 
844 	gpu->worker = kthread_create_worker(0, "gpu-worker");
845 	if (IS_ERR(gpu->worker)) {
846 		ret = PTR_ERR(gpu->worker);
847 		gpu->worker = NULL;
848 		goto fail;
849 	}
850 
851 	sched_set_fifo_low(gpu->worker->task);
852 
853 	mutex_init(&gpu->active_lock);
854 	mutex_init(&gpu->lock);
855 	init_waitqueue_head(&gpu->retire_event);
856 	kthread_init_work(&gpu->retire_work, retire_worker);
857 	kthread_init_work(&gpu->recover_work, recover_worker);
858 	kthread_init_work(&gpu->fault_work, fault_worker);
859 
860 	timer_setup(&gpu->hangcheck_timer, hangcheck_handler, 0);
861 
862 	spin_lock_init(&gpu->perf_lock);
863 
864 
865 	/* Map registers: */
866 	gpu->mmio = msm_ioremap(pdev, config->ioname);
867 	if (IS_ERR(gpu->mmio)) {
868 		ret = PTR_ERR(gpu->mmio);
869 		goto fail;
870 	}
871 
872 	/* Get Interrupt: */
873 	gpu->irq = platform_get_irq(pdev, 0);
874 	if (gpu->irq < 0) {
875 		ret = gpu->irq;
876 		DRM_DEV_ERROR(drm->dev, "failed to get irq: %d\n", ret);
877 		goto fail;
878 	}
879 
880 	ret = devm_request_irq(&pdev->dev, gpu->irq, irq_handler,
881 			IRQF_TRIGGER_HIGH, "gpu-irq", gpu);
882 	if (ret) {
883 		DRM_DEV_ERROR(drm->dev, "failed to request IRQ%u: %d\n", gpu->irq, ret);
884 		goto fail;
885 	}
886 
887 	ret = get_clocks(pdev, gpu);
888 	if (ret)
889 		goto fail;
890 
891 	gpu->ebi1_clk = msm_clk_get(pdev, "bus");
892 	DBG("ebi1_clk: %p", gpu->ebi1_clk);
893 	if (IS_ERR(gpu->ebi1_clk))
894 		gpu->ebi1_clk = NULL;
895 
896 	/* Acquire regulators: */
897 	gpu->gpu_reg = devm_regulator_get(&pdev->dev, "vdd");
898 	DBG("gpu_reg: %p", gpu->gpu_reg);
899 	if (IS_ERR(gpu->gpu_reg))
900 		gpu->gpu_reg = NULL;
901 
902 	gpu->gpu_cx = devm_regulator_get(&pdev->dev, "vddcx");
903 	DBG("gpu_cx: %p", gpu->gpu_cx);
904 	if (IS_ERR(gpu->gpu_cx))
905 		gpu->gpu_cx = NULL;
906 
907 	gpu->cx_collapse = devm_reset_control_get_optional_exclusive(&pdev->dev,
908 			"cx_collapse");
909 
910 	gpu->pdev = pdev;
911 	platform_set_drvdata(pdev, &gpu->adreno_smmu);
912 
913 	msm_devfreq_init(gpu);
914 
915 
916 	gpu->aspace = gpu->funcs->create_address_space(gpu, pdev);
917 
918 	if (gpu->aspace == NULL)
919 		DRM_DEV_INFO(drm->dev, "%s: no IOMMU, fallback to VRAM carveout!\n", name);
920 	else if (IS_ERR(gpu->aspace)) {
921 		ret = PTR_ERR(gpu->aspace);
922 		goto fail;
923 	}
924 
925 	memptrs = msm_gem_kernel_new(drm,
926 		sizeof(struct msm_rbmemptrs) * nr_rings,
927 		check_apriv(gpu, MSM_BO_WC), gpu->aspace, &gpu->memptrs_bo,
928 		&memptrs_iova);
929 
930 	if (IS_ERR(memptrs)) {
931 		ret = PTR_ERR(memptrs);
932 		DRM_DEV_ERROR(drm->dev, "could not allocate memptrs: %d\n", ret);
933 		goto fail;
934 	}
935 
936 	msm_gem_object_set_name(gpu->memptrs_bo, "memptrs");
937 
938 	if (nr_rings > ARRAY_SIZE(gpu->rb)) {
939 		DRM_DEV_INFO_ONCE(drm->dev, "Only creating %zu ringbuffers\n",
940 			ARRAY_SIZE(gpu->rb));
941 		nr_rings = ARRAY_SIZE(gpu->rb);
942 	}
943 
944 	/* Create ringbuffer(s): */
945 	for (i = 0; i < nr_rings; i++) {
946 		gpu->rb[i] = msm_ringbuffer_new(gpu, i, memptrs, memptrs_iova);
947 
948 		if (IS_ERR(gpu->rb[i])) {
949 			ret = PTR_ERR(gpu->rb[i]);
950 			DRM_DEV_ERROR(drm->dev,
951 				"could not create ringbuffer %d: %d\n", i, ret);
952 			goto fail;
953 		}
954 
955 		memptrs += sizeof(struct msm_rbmemptrs);
956 		memptrs_iova += sizeof(struct msm_rbmemptrs);
957 	}
958 
959 	gpu->nr_rings = nr_rings;
960 
961 	refcount_set(&gpu->sysprof_active, 1);
962 
963 	return 0;
964 
965 fail:
966 	for (i = 0; i < ARRAY_SIZE(gpu->rb); i++)  {
967 		msm_ringbuffer_destroy(gpu->rb[i]);
968 		gpu->rb[i] = NULL;
969 	}
970 
971 	msm_gem_kernel_put(gpu->memptrs_bo, gpu->aspace);
972 
973 	platform_set_drvdata(pdev, NULL);
974 	return ret;
975 }
976 
977 void msm_gpu_cleanup(struct msm_gpu *gpu)
978 {
979 	int i;
980 
981 	DBG("%s", gpu->name);
982 
983 	for (i = 0; i < ARRAY_SIZE(gpu->rb); i++) {
984 		msm_ringbuffer_destroy(gpu->rb[i]);
985 		gpu->rb[i] = NULL;
986 	}
987 
988 	msm_gem_kernel_put(gpu->memptrs_bo, gpu->aspace);
989 
990 	if (!IS_ERR_OR_NULL(gpu->aspace)) {
991 		gpu->aspace->mmu->funcs->detach(gpu->aspace->mmu);
992 		msm_gem_address_space_put(gpu->aspace);
993 	}
994 
995 	if (gpu->worker) {
996 		kthread_destroy_worker(gpu->worker);
997 	}
998 
999 	msm_devfreq_cleanup(gpu);
1000 
1001 	platform_set_drvdata(gpu->pdev, NULL);
1002 }
1003