xref: /linux/drivers/gpu/drm/msm/msm_gpu.c (revision a4eb44a6435d6d8f9e642407a4a06f65eb90ca04)
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 "msm_gpu.h"
8 #include "msm_gem.h"
9 #include "msm_mmu.h"
10 #include "msm_fence.h"
11 #include "msm_gpu_trace.h"
12 #include "adreno/adreno_gpu.h"
13 
14 #include <generated/utsrelease.h>
15 #include <linux/string_helpers.h>
16 #include <linux/devcoredump.h>
17 #include <linux/sched/task.h>
18 
19 /*
20  * Power Management:
21  */
22 
23 static int enable_pwrrail(struct msm_gpu *gpu)
24 {
25 	struct drm_device *dev = gpu->dev;
26 	int ret = 0;
27 
28 	if (gpu->gpu_reg) {
29 		ret = regulator_enable(gpu->gpu_reg);
30 		if (ret) {
31 			DRM_DEV_ERROR(dev->dev, "failed to enable 'gpu_reg': %d\n", ret);
32 			return ret;
33 		}
34 	}
35 
36 	if (gpu->gpu_cx) {
37 		ret = regulator_enable(gpu->gpu_cx);
38 		if (ret) {
39 			DRM_DEV_ERROR(dev->dev, "failed to enable 'gpu_cx': %d\n", ret);
40 			return ret;
41 		}
42 	}
43 
44 	return 0;
45 }
46 
47 static int disable_pwrrail(struct msm_gpu *gpu)
48 {
49 	if (gpu->gpu_cx)
50 		regulator_disable(gpu->gpu_cx);
51 	if (gpu->gpu_reg)
52 		regulator_disable(gpu->gpu_reg);
53 	return 0;
54 }
55 
56 static int enable_clk(struct msm_gpu *gpu)
57 {
58 	if (gpu->core_clk && gpu->fast_rate)
59 		clk_set_rate(gpu->core_clk, gpu->fast_rate);
60 
61 	/* Set the RBBM timer rate to 19.2Mhz */
62 	if (gpu->rbbmtimer_clk)
63 		clk_set_rate(gpu->rbbmtimer_clk, 19200000);
64 
65 	return clk_bulk_prepare_enable(gpu->nr_clocks, gpu->grp_clks);
66 }
67 
68 static int disable_clk(struct msm_gpu *gpu)
69 {
70 	clk_bulk_disable_unprepare(gpu->nr_clocks, gpu->grp_clks);
71 
72 	/*
73 	 * Set the clock to a deliberately low rate. On older targets the clock
74 	 * speed had to be non zero to avoid problems. On newer targets this
75 	 * will be rounded down to zero anyway so it all works out.
76 	 */
77 	if (gpu->core_clk)
78 		clk_set_rate(gpu->core_clk, 27000000);
79 
80 	if (gpu->rbbmtimer_clk)
81 		clk_set_rate(gpu->rbbmtimer_clk, 0);
82 
83 	return 0;
84 }
85 
86 static int enable_axi(struct msm_gpu *gpu)
87 {
88 	return clk_prepare_enable(gpu->ebi1_clk);
89 }
90 
91 static int disable_axi(struct msm_gpu *gpu)
92 {
93 	clk_disable_unprepare(gpu->ebi1_clk);
94 	return 0;
95 }
96 
97 int msm_gpu_pm_resume(struct msm_gpu *gpu)
98 {
99 	int ret;
100 
101 	DBG("%s", gpu->name);
102 	trace_msm_gpu_resume(0);
103 
104 	ret = enable_pwrrail(gpu);
105 	if (ret)
106 		return ret;
107 
108 	ret = enable_clk(gpu);
109 	if (ret)
110 		return ret;
111 
112 	ret = enable_axi(gpu);
113 	if (ret)
114 		return ret;
115 
116 	msm_devfreq_resume(gpu);
117 
118 	gpu->needs_hw_init = true;
119 
120 	return 0;
121 }
122 
123 int msm_gpu_pm_suspend(struct msm_gpu *gpu)
124 {
125 	int ret;
126 
127 	DBG("%s", gpu->name);
128 	trace_msm_gpu_suspend(0);
129 
130 	msm_devfreq_suspend(gpu);
131 
132 	ret = disable_axi(gpu);
133 	if (ret)
134 		return ret;
135 
136 	ret = disable_clk(gpu);
137 	if (ret)
138 		return ret;
139 
140 	ret = disable_pwrrail(gpu);
141 	if (ret)
142 		return ret;
143 
144 	gpu->suspend_count++;
145 
146 	return 0;
147 }
148 
149 int msm_gpu_hw_init(struct msm_gpu *gpu)
150 {
151 	int ret;
152 
153 	WARN_ON(!mutex_is_locked(&gpu->lock));
154 
155 	if (!gpu->needs_hw_init)
156 		return 0;
157 
158 	disable_irq(gpu->irq);
159 	ret = gpu->funcs->hw_init(gpu);
160 	if (!ret)
161 		gpu->needs_hw_init = false;
162 	enable_irq(gpu->irq);
163 
164 	return ret;
165 }
166 
167 static void update_fences(struct msm_gpu *gpu, struct msm_ringbuffer *ring,
168 		uint32_t fence)
169 {
170 	struct msm_gem_submit *submit;
171 	unsigned long flags;
172 
173 	spin_lock_irqsave(&ring->submit_lock, flags);
174 	list_for_each_entry(submit, &ring->submits, node) {
175 		if (fence_after(submit->seqno, fence))
176 			break;
177 
178 		msm_update_fence(submit->ring->fctx,
179 			submit->hw_fence->seqno);
180 		dma_fence_signal(submit->hw_fence);
181 	}
182 	spin_unlock_irqrestore(&ring->submit_lock, flags);
183 }
184 
185 #ifdef CONFIG_DEV_COREDUMP
186 static ssize_t msm_gpu_devcoredump_read(char *buffer, loff_t offset,
187 		size_t count, void *data, size_t datalen)
188 {
189 	struct msm_gpu *gpu = data;
190 	struct drm_print_iterator iter;
191 	struct drm_printer p;
192 	struct msm_gpu_state *state;
193 
194 	state = msm_gpu_crashstate_get(gpu);
195 	if (!state)
196 		return 0;
197 
198 	iter.data = buffer;
199 	iter.offset = 0;
200 	iter.start = offset;
201 	iter.remain = count;
202 
203 	p = drm_coredump_printer(&iter);
204 
205 	drm_printf(&p, "---\n");
206 	drm_printf(&p, "kernel: " UTS_RELEASE "\n");
207 	drm_printf(&p, "module: " KBUILD_MODNAME "\n");
208 	drm_printf(&p, "time: %lld.%09ld\n",
209 		state->time.tv_sec, state->time.tv_nsec);
210 	if (state->comm)
211 		drm_printf(&p, "comm: %s\n", state->comm);
212 	if (state->cmd)
213 		drm_printf(&p, "cmdline: %s\n", state->cmd);
214 
215 	gpu->funcs->show(gpu, state, &p);
216 
217 	msm_gpu_crashstate_put(gpu);
218 
219 	return count - iter.remain;
220 }
221 
222 static void msm_gpu_devcoredump_free(void *data)
223 {
224 	struct msm_gpu *gpu = data;
225 
226 	msm_gpu_crashstate_put(gpu);
227 }
228 
229 static void msm_gpu_crashstate_get_bo(struct msm_gpu_state *state,
230 		struct msm_gem_object *obj, u64 iova, u32 flags)
231 {
232 	struct msm_gpu_state_bo *state_bo = &state->bos[state->nr_bos];
233 
234 	/* Don't record write only objects */
235 	state_bo->size = obj->base.size;
236 	state_bo->iova = iova;
237 
238 	/* Only store data for non imported buffer objects marked for read */
239 	if ((flags & MSM_SUBMIT_BO_READ) && !obj->base.import_attach) {
240 		void *ptr;
241 
242 		state_bo->data = kvmalloc(obj->base.size, GFP_KERNEL);
243 		if (!state_bo->data)
244 			goto out;
245 
246 		msm_gem_lock(&obj->base);
247 		ptr = msm_gem_get_vaddr_active(&obj->base);
248 		msm_gem_unlock(&obj->base);
249 		if (IS_ERR(ptr)) {
250 			kvfree(state_bo->data);
251 			state_bo->data = NULL;
252 			goto out;
253 		}
254 
255 		memcpy(state_bo->data, ptr, obj->base.size);
256 		msm_gem_put_vaddr(&obj->base);
257 	}
258 out:
259 	state->nr_bos++;
260 }
261 
262 static void msm_gpu_crashstate_capture(struct msm_gpu *gpu,
263 		struct msm_gem_submit *submit, char *comm, char *cmd)
264 {
265 	struct msm_gpu_state *state;
266 
267 	/* Check if the target supports capturing crash state */
268 	if (!gpu->funcs->gpu_state_get)
269 		return;
270 
271 	/* Only save one crash state at a time */
272 	if (gpu->crashstate)
273 		return;
274 
275 	state = gpu->funcs->gpu_state_get(gpu);
276 	if (IS_ERR_OR_NULL(state))
277 		return;
278 
279 	/* Fill in the additional crash state information */
280 	state->comm = kstrdup(comm, GFP_KERNEL);
281 	state->cmd = kstrdup(cmd, GFP_KERNEL);
282 	state->fault_info = gpu->fault_info;
283 
284 	if (submit) {
285 		int i, nr = 0;
286 
287 		/* count # of buffers to dump: */
288 		for (i = 0; i < submit->nr_bos; i++)
289 			if (should_dump(submit, i))
290 				nr++;
291 		/* always dump cmd bo's, but don't double count them: */
292 		for (i = 0; i < submit->nr_cmds; i++)
293 			if (!should_dump(submit, submit->cmd[i].idx))
294 				nr++;
295 
296 		state->bos = kcalloc(nr,
297 			sizeof(struct msm_gpu_state_bo), GFP_KERNEL);
298 
299 		for (i = 0; state->bos && i < submit->nr_bos; i++) {
300 			if (should_dump(submit, i)) {
301 				msm_gpu_crashstate_get_bo(state, submit->bos[i].obj,
302 					submit->bos[i].iova, submit->bos[i].flags);
303 			}
304 		}
305 
306 		for (i = 0; state->bos && i < submit->nr_cmds; i++) {
307 			int idx = submit->cmd[i].idx;
308 
309 			if (!should_dump(submit, submit->cmd[i].idx)) {
310 				msm_gpu_crashstate_get_bo(state, submit->bos[idx].obj,
311 					submit->bos[idx].iova, submit->bos[idx].flags);
312 			}
313 		}
314 	}
315 
316 	/* Set the active crash state to be dumped on failure */
317 	gpu->crashstate = state;
318 
319 	/* FIXME: Release the crashstate if this errors out? */
320 	dev_coredumpm(gpu->dev->dev, THIS_MODULE, gpu, 0, GFP_KERNEL,
321 		msm_gpu_devcoredump_read, msm_gpu_devcoredump_free);
322 }
323 #else
324 static void msm_gpu_crashstate_capture(struct msm_gpu *gpu,
325 		struct msm_gem_submit *submit, char *comm, char *cmd)
326 {
327 }
328 #endif
329 
330 /*
331  * Hangcheck detection for locked gpu:
332  */
333 
334 static struct msm_gem_submit *
335 find_submit(struct msm_ringbuffer *ring, uint32_t fence)
336 {
337 	struct msm_gem_submit *submit;
338 	unsigned long flags;
339 
340 	spin_lock_irqsave(&ring->submit_lock, flags);
341 	list_for_each_entry(submit, &ring->submits, node) {
342 		if (submit->seqno == fence) {
343 			spin_unlock_irqrestore(&ring->submit_lock, flags);
344 			return submit;
345 		}
346 	}
347 	spin_unlock_irqrestore(&ring->submit_lock, flags);
348 
349 	return NULL;
350 }
351 
352 static void retire_submits(struct msm_gpu *gpu);
353 
354 static void recover_worker(struct kthread_work *work)
355 {
356 	struct msm_gpu *gpu = container_of(work, struct msm_gpu, recover_work);
357 	struct drm_device *dev = gpu->dev;
358 	struct msm_drm_private *priv = dev->dev_private;
359 	struct msm_gem_submit *submit;
360 	struct msm_ringbuffer *cur_ring = gpu->funcs->active_ring(gpu);
361 	char *comm = NULL, *cmd = NULL;
362 	int i;
363 
364 	mutex_lock(&gpu->lock);
365 
366 	DRM_DEV_ERROR(dev->dev, "%s: hangcheck recover!\n", gpu->name);
367 
368 	submit = find_submit(cur_ring, cur_ring->memptrs->fence + 1);
369 	if (submit) {
370 		struct task_struct *task;
371 
372 		/* Increment the fault counts */
373 		gpu->global_faults++;
374 		submit->queue->faults++;
375 
376 		task = get_pid_task(submit->pid, PIDTYPE_PID);
377 		if (task) {
378 			comm = kstrdup(task->comm, GFP_KERNEL);
379 			cmd = kstrdup_quotable_cmdline(task, GFP_KERNEL);
380 			put_task_struct(task);
381 		}
382 
383 		if (comm && cmd) {
384 			DRM_DEV_ERROR(dev->dev, "%s: offending task: %s (%s)\n",
385 				gpu->name, comm, cmd);
386 
387 			msm_rd_dump_submit(priv->hangrd, submit,
388 				"offending task: %s (%s)", comm, cmd);
389 		} else {
390 			msm_rd_dump_submit(priv->hangrd, submit, NULL);
391 		}
392 	}
393 
394 	/* Record the crash state */
395 	pm_runtime_get_sync(&gpu->pdev->dev);
396 	msm_gpu_crashstate_capture(gpu, submit, comm, cmd);
397 	pm_runtime_put_sync(&gpu->pdev->dev);
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 			fence++;
418 
419 		update_fences(gpu, ring, fence);
420 	}
421 
422 	if (msm_gpu_active(gpu)) {
423 		/* retire completed submits, plus the one that hung: */
424 		retire_submits(gpu);
425 
426 		pm_runtime_get_sync(&gpu->pdev->dev);
427 		gpu->funcs->recover(gpu);
428 		pm_runtime_put_sync(&gpu->pdev->dev);
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 	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 		struct task_struct *task;
465 
466 		task = get_pid_task(submit->pid, PIDTYPE_PID);
467 		if (task) {
468 			comm = kstrdup(task->comm, GFP_KERNEL);
469 			cmd = kstrdup_quotable_cmdline(task, GFP_KERNEL);
470 			put_task_struct(task);
471 		}
472 
473 		/*
474 		 * When we get GPU iova faults, we can get 1000s of them,
475 		 * but we really only want to log the first one.
476 		 */
477 		submit->fault_dumped = true;
478 	}
479 
480 	/* Record the crash state */
481 	pm_runtime_get_sync(&gpu->pdev->dev);
482 	msm_gpu_crashstate_capture(gpu, submit, comm, cmd);
483 	pm_runtime_put_sync(&gpu->pdev->dev);
484 
485 	kfree(cmd);
486 	kfree(comm);
487 
488 resume_smmu:
489 	memset(&gpu->fault_info, 0, sizeof(gpu->fault_info));
490 	gpu->aspace->mmu->funcs->resume_translation(gpu->aspace->mmu);
491 
492 	mutex_unlock(&gpu->lock);
493 }
494 
495 static void hangcheck_timer_reset(struct msm_gpu *gpu)
496 {
497 	struct msm_drm_private *priv = gpu->dev->dev_private;
498 	mod_timer(&gpu->hangcheck_timer,
499 			round_jiffies_up(jiffies + msecs_to_jiffies(priv->hangcheck_period)));
500 }
501 
502 static void hangcheck_handler(struct timer_list *t)
503 {
504 	struct msm_gpu *gpu = from_timer(gpu, t, hangcheck_timer);
505 	struct drm_device *dev = gpu->dev;
506 	struct msm_ringbuffer *ring = gpu->funcs->active_ring(gpu);
507 	uint32_t fence = ring->memptrs->fence;
508 
509 	if (fence != ring->hangcheck_fence) {
510 		/* some progress has been made.. ya! */
511 		ring->hangcheck_fence = fence;
512 	} else if (fence_before(fence, ring->seqno)) {
513 		/* no progress and not done.. hung! */
514 		ring->hangcheck_fence = fence;
515 		DRM_DEV_ERROR(dev->dev, "%s: hangcheck detected gpu lockup rb %d!\n",
516 				gpu->name, ring->id);
517 		DRM_DEV_ERROR(dev->dev, "%s:     completed fence: %u\n",
518 				gpu->name, fence);
519 		DRM_DEV_ERROR(dev->dev, "%s:     submitted fence: %u\n",
520 				gpu->name, ring->seqno);
521 
522 		kthread_queue_work(gpu->worker, &gpu->recover_work);
523 	}
524 
525 	/* if still more pending work, reset the hangcheck timer: */
526 	if (fence_after(ring->seqno, ring->hangcheck_fence))
527 		hangcheck_timer_reset(gpu);
528 
529 	/* workaround for missing irq: */
530 	msm_gpu_retire(gpu);
531 }
532 
533 /*
534  * Performance Counters:
535  */
536 
537 /* called under perf_lock */
538 static int update_hw_cntrs(struct msm_gpu *gpu, uint32_t ncntrs, uint32_t *cntrs)
539 {
540 	uint32_t current_cntrs[ARRAY_SIZE(gpu->last_cntrs)];
541 	int i, n = min(ncntrs, gpu->num_perfcntrs);
542 
543 	/* read current values: */
544 	for (i = 0; i < gpu->num_perfcntrs; i++)
545 		current_cntrs[i] = gpu_read(gpu, gpu->perfcntrs[i].sample_reg);
546 
547 	/* update cntrs: */
548 	for (i = 0; i < n; i++)
549 		cntrs[i] = current_cntrs[i] - gpu->last_cntrs[i];
550 
551 	/* save current values: */
552 	for (i = 0; i < gpu->num_perfcntrs; i++)
553 		gpu->last_cntrs[i] = current_cntrs[i];
554 
555 	return n;
556 }
557 
558 static void update_sw_cntrs(struct msm_gpu *gpu)
559 {
560 	ktime_t time;
561 	uint32_t elapsed;
562 	unsigned long flags;
563 
564 	spin_lock_irqsave(&gpu->perf_lock, flags);
565 	if (!gpu->perfcntr_active)
566 		goto out;
567 
568 	time = ktime_get();
569 	elapsed = ktime_to_us(ktime_sub(time, gpu->last_sample.time));
570 
571 	gpu->totaltime += elapsed;
572 	if (gpu->last_sample.active)
573 		gpu->activetime += elapsed;
574 
575 	gpu->last_sample.active = msm_gpu_active(gpu);
576 	gpu->last_sample.time = time;
577 
578 out:
579 	spin_unlock_irqrestore(&gpu->perf_lock, flags);
580 }
581 
582 void msm_gpu_perfcntr_start(struct msm_gpu *gpu)
583 {
584 	unsigned long flags;
585 
586 	pm_runtime_get_sync(&gpu->pdev->dev);
587 
588 	spin_lock_irqsave(&gpu->perf_lock, flags);
589 	/* we could dynamically enable/disable perfcntr registers too.. */
590 	gpu->last_sample.active = msm_gpu_active(gpu);
591 	gpu->last_sample.time = ktime_get();
592 	gpu->activetime = gpu->totaltime = 0;
593 	gpu->perfcntr_active = true;
594 	update_hw_cntrs(gpu, 0, NULL);
595 	spin_unlock_irqrestore(&gpu->perf_lock, flags);
596 }
597 
598 void msm_gpu_perfcntr_stop(struct msm_gpu *gpu)
599 {
600 	gpu->perfcntr_active = false;
601 	pm_runtime_put_sync(&gpu->pdev->dev);
602 }
603 
604 /* returns -errno or # of cntrs sampled */
605 int msm_gpu_perfcntr_sample(struct msm_gpu *gpu, uint32_t *activetime,
606 		uint32_t *totaltime, uint32_t ncntrs, uint32_t *cntrs)
607 {
608 	unsigned long flags;
609 	int ret;
610 
611 	spin_lock_irqsave(&gpu->perf_lock, flags);
612 
613 	if (!gpu->perfcntr_active) {
614 		ret = -EINVAL;
615 		goto out;
616 	}
617 
618 	*activetime = gpu->activetime;
619 	*totaltime = gpu->totaltime;
620 
621 	gpu->activetime = gpu->totaltime = 0;
622 
623 	ret = update_hw_cntrs(gpu, ncntrs, cntrs);
624 
625 out:
626 	spin_unlock_irqrestore(&gpu->perf_lock, flags);
627 
628 	return ret;
629 }
630 
631 /*
632  * Cmdstream submission/retirement:
633  */
634 
635 static void retire_submit(struct msm_gpu *gpu, struct msm_ringbuffer *ring,
636 		struct msm_gem_submit *submit)
637 {
638 	int index = submit->seqno % MSM_GPU_SUBMIT_STATS_COUNT;
639 	volatile struct msm_gpu_submit_stats *stats;
640 	u64 elapsed, clock = 0;
641 	unsigned long flags;
642 
643 	stats = &ring->memptrs->stats[index];
644 	/* Convert 19.2Mhz alwayson ticks to nanoseconds for elapsed time */
645 	elapsed = (stats->alwayson_end - stats->alwayson_start) * 10000;
646 	do_div(elapsed, 192);
647 
648 	/* Calculate the clock frequency from the number of CP cycles */
649 	if (elapsed) {
650 		clock = (stats->cpcycles_end - stats->cpcycles_start) * 1000;
651 		do_div(clock, elapsed);
652 	}
653 
654 	trace_msm_gpu_submit_retired(submit, elapsed, clock,
655 		stats->alwayson_start, stats->alwayson_end);
656 
657 	msm_submit_retire(submit);
658 
659 	pm_runtime_mark_last_busy(&gpu->pdev->dev);
660 	pm_runtime_put_autosuspend(&gpu->pdev->dev);
661 
662 	spin_lock_irqsave(&ring->submit_lock, flags);
663 	list_del(&submit->node);
664 	spin_unlock_irqrestore(&ring->submit_lock, flags);
665 
666 	/* Update devfreq on transition from active->idle: */
667 	mutex_lock(&gpu->active_lock);
668 	gpu->active_submits--;
669 	WARN_ON(gpu->active_submits < 0);
670 	if (!gpu->active_submits)
671 		msm_devfreq_idle(gpu);
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 		update_fences(gpu, gpu->rb[i], 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 = ++ring->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 		msm_devfreq_active(gpu);
763 	gpu->active_submits++;
764 	mutex_unlock(&gpu->active_lock);
765 
766 	gpu->funcs->submit(gpu, submit);
767 	gpu->cur_ctx_seqno = submit->queue->ctx->seqno;
768 
769 	hangcheck_timer_reset(gpu);
770 }
771 
772 /*
773  * Init/Cleanup:
774  */
775 
776 static irqreturn_t irq_handler(int irq, void *data)
777 {
778 	struct msm_gpu *gpu = data;
779 	return gpu->funcs->irq(gpu);
780 }
781 
782 static int get_clocks(struct platform_device *pdev, struct msm_gpu *gpu)
783 {
784 	int ret = devm_clk_bulk_get_all(&pdev->dev, &gpu->grp_clks);
785 
786 	if (ret < 1) {
787 		gpu->nr_clocks = 0;
788 		return ret;
789 	}
790 
791 	gpu->nr_clocks = ret;
792 
793 	gpu->core_clk = msm_clk_bulk_get_clock(gpu->grp_clks,
794 		gpu->nr_clocks, "core");
795 
796 	gpu->rbbmtimer_clk = msm_clk_bulk_get_clock(gpu->grp_clks,
797 		gpu->nr_clocks, "rbbmtimer");
798 
799 	return 0;
800 }
801 
802 /* Return a new address space for a msm_drm_private instance */
803 struct msm_gem_address_space *
804 msm_gpu_create_private_address_space(struct msm_gpu *gpu, struct task_struct *task)
805 {
806 	struct msm_gem_address_space *aspace = NULL;
807 	if (!gpu)
808 		return NULL;
809 
810 	/*
811 	 * If the target doesn't support private address spaces then return
812 	 * the global one
813 	 */
814 	if (gpu->funcs->create_private_address_space) {
815 		aspace = gpu->funcs->create_private_address_space(gpu);
816 		if (!IS_ERR(aspace))
817 			aspace->pid = get_pid(task_pid(task));
818 	}
819 
820 	if (IS_ERR_OR_NULL(aspace))
821 		aspace = msm_gem_address_space_get(gpu->aspace);
822 
823 	return aspace;
824 }
825 
826 int msm_gpu_init(struct drm_device *drm, struct platform_device *pdev,
827 		struct msm_gpu *gpu, const struct msm_gpu_funcs *funcs,
828 		const char *name, struct msm_gpu_config *config)
829 {
830 	int i, ret, nr_rings = config->nr_rings;
831 	void *memptrs;
832 	uint64_t memptrs_iova;
833 
834 	if (WARN_ON(gpu->num_perfcntrs > ARRAY_SIZE(gpu->last_cntrs)))
835 		gpu->num_perfcntrs = ARRAY_SIZE(gpu->last_cntrs);
836 
837 	gpu->dev = drm;
838 	gpu->funcs = funcs;
839 	gpu->name = name;
840 
841 	gpu->worker = kthread_create_worker(0, "%s-worker", gpu->name);
842 	if (IS_ERR(gpu->worker)) {
843 		ret = PTR_ERR(gpu->worker);
844 		gpu->worker = NULL;
845 		goto fail;
846 	}
847 
848 	sched_set_fifo_low(gpu->worker->task);
849 
850 	INIT_LIST_HEAD(&gpu->active_list);
851 	mutex_init(&gpu->active_lock);
852 	mutex_init(&gpu->lock);
853 	init_waitqueue_head(&gpu->retire_event);
854 	kthread_init_work(&gpu->retire_work, retire_worker);
855 	kthread_init_work(&gpu->recover_work, recover_worker);
856 	kthread_init_work(&gpu->fault_work, fault_worker);
857 
858 	timer_setup(&gpu->hangcheck_timer, hangcheck_handler, 0);
859 
860 	spin_lock_init(&gpu->perf_lock);
861 
862 
863 	/* Map registers: */
864 	gpu->mmio = msm_ioremap(pdev, config->ioname, name);
865 	if (IS_ERR(gpu->mmio)) {
866 		ret = PTR_ERR(gpu->mmio);
867 		goto fail;
868 	}
869 
870 	/* Get Interrupt: */
871 	gpu->irq = platform_get_irq(pdev, 0);
872 	if (gpu->irq < 0) {
873 		ret = gpu->irq;
874 		DRM_DEV_ERROR(drm->dev, "failed to get irq: %d\n", ret);
875 		goto fail;
876 	}
877 
878 	ret = devm_request_irq(&pdev->dev, gpu->irq, irq_handler,
879 			IRQF_TRIGGER_HIGH, gpu->name, gpu);
880 	if (ret) {
881 		DRM_DEV_ERROR(drm->dev, "failed to request IRQ%u: %d\n", gpu->irq, ret);
882 		goto fail;
883 	}
884 
885 	ret = get_clocks(pdev, gpu);
886 	if (ret)
887 		goto fail;
888 
889 	gpu->ebi1_clk = msm_clk_get(pdev, "bus");
890 	DBG("ebi1_clk: %p", gpu->ebi1_clk);
891 	if (IS_ERR(gpu->ebi1_clk))
892 		gpu->ebi1_clk = NULL;
893 
894 	/* Acquire regulators: */
895 	gpu->gpu_reg = devm_regulator_get(&pdev->dev, "vdd");
896 	DBG("gpu_reg: %p", gpu->gpu_reg);
897 	if (IS_ERR(gpu->gpu_reg))
898 		gpu->gpu_reg = NULL;
899 
900 	gpu->gpu_cx = devm_regulator_get(&pdev->dev, "vddcx");
901 	DBG("gpu_cx: %p", gpu->gpu_cx);
902 	if (IS_ERR(gpu->gpu_cx))
903 		gpu->gpu_cx = NULL;
904 
905 	gpu->pdev = pdev;
906 	platform_set_drvdata(pdev, &gpu->adreno_smmu);
907 
908 	msm_devfreq_init(gpu);
909 
910 
911 	gpu->aspace = gpu->funcs->create_address_space(gpu, pdev);
912 
913 	if (gpu->aspace == NULL)
914 		DRM_DEV_INFO(drm->dev, "%s: no IOMMU, fallback to VRAM carveout!\n", name);
915 	else if (IS_ERR(gpu->aspace)) {
916 		ret = PTR_ERR(gpu->aspace);
917 		goto fail;
918 	}
919 
920 	memptrs = msm_gem_kernel_new(drm,
921 		sizeof(struct msm_rbmemptrs) * nr_rings,
922 		check_apriv(gpu, MSM_BO_UNCACHED), gpu->aspace, &gpu->memptrs_bo,
923 		&memptrs_iova);
924 
925 	if (IS_ERR(memptrs)) {
926 		ret = PTR_ERR(memptrs);
927 		DRM_DEV_ERROR(drm->dev, "could not allocate memptrs: %d\n", ret);
928 		goto fail;
929 	}
930 
931 	msm_gem_object_set_name(gpu->memptrs_bo, "memptrs");
932 
933 	if (nr_rings > ARRAY_SIZE(gpu->rb)) {
934 		DRM_DEV_INFO_ONCE(drm->dev, "Only creating %zu ringbuffers\n",
935 			ARRAY_SIZE(gpu->rb));
936 		nr_rings = ARRAY_SIZE(gpu->rb);
937 	}
938 
939 	/* Create ringbuffer(s): */
940 	for (i = 0; i < nr_rings; i++) {
941 		gpu->rb[i] = msm_ringbuffer_new(gpu, i, memptrs, memptrs_iova);
942 
943 		if (IS_ERR(gpu->rb[i])) {
944 			ret = PTR_ERR(gpu->rb[i]);
945 			DRM_DEV_ERROR(drm->dev,
946 				"could not create ringbuffer %d: %d\n", i, ret);
947 			goto fail;
948 		}
949 
950 		memptrs += sizeof(struct msm_rbmemptrs);
951 		memptrs_iova += sizeof(struct msm_rbmemptrs);
952 	}
953 
954 	gpu->nr_rings = nr_rings;
955 
956 	return 0;
957 
958 fail:
959 	for (i = 0; i < ARRAY_SIZE(gpu->rb); i++)  {
960 		msm_ringbuffer_destroy(gpu->rb[i]);
961 		gpu->rb[i] = NULL;
962 	}
963 
964 	msm_gem_kernel_put(gpu->memptrs_bo, gpu->aspace);
965 
966 	platform_set_drvdata(pdev, NULL);
967 	return ret;
968 }
969 
970 void msm_gpu_cleanup(struct msm_gpu *gpu)
971 {
972 	int i;
973 
974 	DBG("%s", gpu->name);
975 
976 	WARN_ON(!list_empty(&gpu->active_list));
977 
978 	for (i = 0; i < ARRAY_SIZE(gpu->rb); i++) {
979 		msm_ringbuffer_destroy(gpu->rb[i]);
980 		gpu->rb[i] = NULL;
981 	}
982 
983 	msm_gem_kernel_put(gpu->memptrs_bo, gpu->aspace);
984 
985 	if (!IS_ERR_OR_NULL(gpu->aspace)) {
986 		gpu->aspace->mmu->funcs->detach(gpu->aspace->mmu);
987 		msm_gem_address_space_put(gpu->aspace);
988 	}
989 
990 	if (gpu->worker) {
991 		kthread_destroy_worker(gpu->worker);
992 	}
993 
994 	msm_devfreq_cleanup(gpu);
995 }
996