xref: /linux/drivers/gpu/drm/msm/msm_gpu.c (revision 4b660dbd9ee2059850fd30e0df420ca7a38a1856)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * Copyright (C) 2013 Red Hat
4  * Author: Rob Clark <robdclark@gmail.com>
5  */
6 
7 #include "drm/drm_drv.h"
8 
9 #include "msm_gpu.h"
10 #include "msm_gem.h"
11 #include "msm_mmu.h"
12 #include "msm_fence.h"
13 #include "msm_gpu_trace.h"
14 #include "adreno/adreno_gpu.h"
15 
16 #include <generated/utsrelease.h>
17 #include <linux/string_helpers.h>
18 #include <linux/devcoredump.h>
19 #include <linux/sched/task.h>
20 
21 /*
22  * Power Management:
23  */
24 
25 static int enable_pwrrail(struct msm_gpu *gpu)
26 {
27 	struct drm_device *dev = gpu->dev;
28 	int ret = 0;
29 
30 	if (gpu->gpu_reg) {
31 		ret = regulator_enable(gpu->gpu_reg);
32 		if (ret) {
33 			DRM_DEV_ERROR(dev->dev, "failed to enable 'gpu_reg': %d\n", ret);
34 			return ret;
35 		}
36 	}
37 
38 	if (gpu->gpu_cx) {
39 		ret = regulator_enable(gpu->gpu_cx);
40 		if (ret) {
41 			DRM_DEV_ERROR(dev->dev, "failed to enable 'gpu_cx': %d\n", ret);
42 			return ret;
43 		}
44 	}
45 
46 	return 0;
47 }
48 
49 static int disable_pwrrail(struct msm_gpu *gpu)
50 {
51 	if (gpu->gpu_cx)
52 		regulator_disable(gpu->gpu_cx);
53 	if (gpu->gpu_reg)
54 		regulator_disable(gpu->gpu_reg);
55 	return 0;
56 }
57 
58 static int enable_clk(struct msm_gpu *gpu)
59 {
60 	if (gpu->core_clk && gpu->fast_rate)
61 		dev_pm_opp_set_rate(&gpu->pdev->dev, gpu->fast_rate);
62 
63 	/* Set the RBBM timer rate to 19.2Mhz */
64 	if (gpu->rbbmtimer_clk)
65 		clk_set_rate(gpu->rbbmtimer_clk, 19200000);
66 
67 	return clk_bulk_prepare_enable(gpu->nr_clocks, gpu->grp_clks);
68 }
69 
70 static int disable_clk(struct msm_gpu *gpu)
71 {
72 	clk_bulk_disable_unprepare(gpu->nr_clocks, gpu->grp_clks);
73 
74 	/*
75 	 * Set the clock to a deliberately low rate. On older targets the clock
76 	 * speed had to be non zero to avoid problems. On newer targets this
77 	 * will be rounded down to zero anyway so it all works out.
78 	 */
79 	if (gpu->core_clk)
80 		dev_pm_opp_set_rate(&gpu->pdev->dev, 27000000);
81 
82 	if (gpu->rbbmtimer_clk)
83 		clk_set_rate(gpu->rbbmtimer_clk, 0);
84 
85 	return 0;
86 }
87 
88 static int enable_axi(struct msm_gpu *gpu)
89 {
90 	return clk_prepare_enable(gpu->ebi1_clk);
91 }
92 
93 static int disable_axi(struct msm_gpu *gpu)
94 {
95 	clk_disable_unprepare(gpu->ebi1_clk);
96 	return 0;
97 }
98 
99 int msm_gpu_pm_resume(struct msm_gpu *gpu)
100 {
101 	int ret;
102 
103 	DBG("%s", gpu->name);
104 	trace_msm_gpu_resume(0);
105 
106 	ret = enable_pwrrail(gpu);
107 	if (ret)
108 		return ret;
109 
110 	ret = enable_clk(gpu);
111 	if (ret)
112 		return ret;
113 
114 	ret = enable_axi(gpu);
115 	if (ret)
116 		return ret;
117 
118 	msm_devfreq_resume(gpu);
119 
120 	gpu->needs_hw_init = true;
121 
122 	return 0;
123 }
124 
125 int msm_gpu_pm_suspend(struct msm_gpu *gpu)
126 {
127 	int ret;
128 
129 	DBG("%s", gpu->name);
130 	trace_msm_gpu_suspend(0);
131 
132 	msm_devfreq_suspend(gpu);
133 
134 	ret = disable_axi(gpu);
135 	if (ret)
136 		return ret;
137 
138 	ret = disable_clk(gpu);
139 	if (ret)
140 		return ret;
141 
142 	ret = disable_pwrrail(gpu);
143 	if (ret)
144 		return ret;
145 
146 	gpu->suspend_count++;
147 
148 	return 0;
149 }
150 
151 void msm_gpu_show_fdinfo(struct msm_gpu *gpu, struct msm_file_private *ctx,
152 			 struct drm_printer *p)
153 {
154 	drm_printf(p, "drm-engine-gpu:\t%llu ns\n", ctx->elapsed_ns);
155 	drm_printf(p, "drm-cycles-gpu:\t%llu\n", ctx->cycles);
156 	drm_printf(p, "drm-maxfreq-gpu:\t%u Hz\n", gpu->fast_rate);
157 }
158 
159 int msm_gpu_hw_init(struct msm_gpu *gpu)
160 {
161 	int ret;
162 
163 	WARN_ON(!mutex_is_locked(&gpu->lock));
164 
165 	if (!gpu->needs_hw_init)
166 		return 0;
167 
168 	disable_irq(gpu->irq);
169 	ret = gpu->funcs->hw_init(gpu);
170 	if (!ret)
171 		gpu->needs_hw_init = false;
172 	enable_irq(gpu->irq);
173 
174 	return ret;
175 }
176 
177 #ifdef CONFIG_DEV_COREDUMP
178 static ssize_t msm_gpu_devcoredump_read(char *buffer, loff_t offset,
179 		size_t count, void *data, size_t datalen)
180 {
181 	struct msm_gpu *gpu = data;
182 	struct drm_print_iterator iter;
183 	struct drm_printer p;
184 	struct msm_gpu_state *state;
185 
186 	state = msm_gpu_crashstate_get(gpu);
187 	if (!state)
188 		return 0;
189 
190 	iter.data = buffer;
191 	iter.offset = 0;
192 	iter.start = offset;
193 	iter.remain = count;
194 
195 	p = drm_coredump_printer(&iter);
196 
197 	drm_printf(&p, "---\n");
198 	drm_printf(&p, "kernel: " UTS_RELEASE "\n");
199 	drm_printf(&p, "module: " KBUILD_MODNAME "\n");
200 	drm_printf(&p, "time: %lld.%09ld\n",
201 		state->time.tv_sec, state->time.tv_nsec);
202 	if (state->comm)
203 		drm_printf(&p, "comm: %s\n", state->comm);
204 	if (state->cmd)
205 		drm_printf(&p, "cmdline: %s\n", state->cmd);
206 
207 	gpu->funcs->show(gpu, state, &p);
208 
209 	msm_gpu_crashstate_put(gpu);
210 
211 	return count - iter.remain;
212 }
213 
214 static void msm_gpu_devcoredump_free(void *data)
215 {
216 	struct msm_gpu *gpu = data;
217 
218 	msm_gpu_crashstate_put(gpu);
219 }
220 
221 static void msm_gpu_crashstate_get_bo(struct msm_gpu_state *state,
222 		struct drm_gem_object *obj, u64 iova, bool full)
223 {
224 	struct msm_gpu_state_bo *state_bo = &state->bos[state->nr_bos];
225 
226 	/* Don't record write only objects */
227 	state_bo->size = obj->size;
228 	state_bo->iova = iova;
229 
230 	BUILD_BUG_ON(sizeof(state_bo->name) != sizeof(to_msm_bo(obj)->name));
231 
232 	memcpy(state_bo->name, to_msm_bo(obj)->name, sizeof(state_bo->name));
233 
234 	if (full) {
235 		void *ptr;
236 
237 		state_bo->data = kvmalloc(obj->size, GFP_KERNEL);
238 		if (!state_bo->data)
239 			goto out;
240 
241 		msm_gem_lock(obj);
242 		ptr = msm_gem_get_vaddr_active(obj);
243 		msm_gem_unlock(obj);
244 		if (IS_ERR(ptr)) {
245 			kvfree(state_bo->data);
246 			state_bo->data = NULL;
247 			goto out;
248 		}
249 
250 		memcpy(state_bo->data, ptr, obj->size);
251 		msm_gem_put_vaddr(obj);
252 	}
253 out:
254 	state->nr_bos++;
255 }
256 
257 static void msm_gpu_crashstate_capture(struct msm_gpu *gpu,
258 		struct msm_gem_submit *submit, char *comm, char *cmd)
259 {
260 	struct msm_gpu_state *state;
261 
262 	/* Check if the target supports capturing crash state */
263 	if (!gpu->funcs->gpu_state_get)
264 		return;
265 
266 	/* Only save one crash state at a time */
267 	if (gpu->crashstate)
268 		return;
269 
270 	state = gpu->funcs->gpu_state_get(gpu);
271 	if (IS_ERR_OR_NULL(state))
272 		return;
273 
274 	/* Fill in the additional crash state information */
275 	state->comm = kstrdup(comm, GFP_KERNEL);
276 	state->cmd = kstrdup(cmd, GFP_KERNEL);
277 	state->fault_info = gpu->fault_info;
278 
279 	if (submit) {
280 		int i;
281 
282 		state->bos = kcalloc(submit->nr_bos,
283 			sizeof(struct msm_gpu_state_bo), GFP_KERNEL);
284 
285 		for (i = 0; state->bos && i < submit->nr_bos; i++) {
286 			msm_gpu_crashstate_get_bo(state, submit->bos[i].obj,
287 						  submit->bos[i].iova,
288 						  should_dump(submit, i));
289 		}
290 	}
291 
292 	/* Set the active crash state to be dumped on failure */
293 	gpu->crashstate = state;
294 
295 	dev_coredumpm(&gpu->pdev->dev, THIS_MODULE, gpu, 0, GFP_KERNEL,
296 		msm_gpu_devcoredump_read, msm_gpu_devcoredump_free);
297 }
298 #else
299 static void msm_gpu_crashstate_capture(struct msm_gpu *gpu,
300 		struct msm_gem_submit *submit, char *comm, char *cmd)
301 {
302 }
303 #endif
304 
305 /*
306  * Hangcheck detection for locked gpu:
307  */
308 
309 static struct msm_gem_submit *
310 find_submit(struct msm_ringbuffer *ring, uint32_t fence)
311 {
312 	struct msm_gem_submit *submit;
313 	unsigned long flags;
314 
315 	spin_lock_irqsave(&ring->submit_lock, flags);
316 	list_for_each_entry(submit, &ring->submits, node) {
317 		if (submit->seqno == fence) {
318 			spin_unlock_irqrestore(&ring->submit_lock, flags);
319 			return submit;
320 		}
321 	}
322 	spin_unlock_irqrestore(&ring->submit_lock, flags);
323 
324 	return NULL;
325 }
326 
327 static void retire_submits(struct msm_gpu *gpu);
328 
329 static void get_comm_cmdline(struct msm_gem_submit *submit, char **comm, char **cmd)
330 {
331 	struct msm_file_private *ctx = submit->queue->ctx;
332 	struct task_struct *task;
333 
334 	WARN_ON(!mutex_is_locked(&submit->gpu->lock));
335 
336 	/* Note that kstrdup will return NULL if argument is NULL: */
337 	*comm = kstrdup(ctx->comm, GFP_KERNEL);
338 	*cmd  = kstrdup(ctx->cmdline, GFP_KERNEL);
339 
340 	task = get_pid_task(submit->pid, PIDTYPE_PID);
341 	if (!task)
342 		return;
343 
344 	if (!*comm)
345 		*comm = kstrdup(task->comm, GFP_KERNEL);
346 
347 	if (!*cmd)
348 		*cmd = kstrdup_quotable_cmdline(task, GFP_KERNEL);
349 
350 	put_task_struct(task);
351 }
352 
353 static void recover_worker(struct kthread_work *work)
354 {
355 	struct msm_gpu *gpu = container_of(work, struct msm_gpu, recover_work);
356 	struct drm_device *dev = gpu->dev;
357 	struct msm_drm_private *priv = dev->dev_private;
358 	struct msm_gem_submit *submit;
359 	struct msm_ringbuffer *cur_ring = gpu->funcs->active_ring(gpu);
360 	char *comm = NULL, *cmd = NULL;
361 	int i;
362 
363 	mutex_lock(&gpu->lock);
364 
365 	DRM_DEV_ERROR(dev->dev, "%s: hangcheck recover!\n", gpu->name);
366 
367 	submit = find_submit(cur_ring, cur_ring->memptrs->fence + 1);
368 
369 	/*
370 	 * If the submit retired while we were waiting for the worker to run,
371 	 * or waiting to acquire the gpu lock, then nothing more to do.
372 	 */
373 	if (!submit)
374 		goto out_unlock;
375 
376 	/* Increment the fault counts */
377 	submit->queue->faults++;
378 	if (submit->aspace)
379 		submit->aspace->faults++;
380 
381 	get_comm_cmdline(submit, &comm, &cmd);
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 		DRM_DEV_ERROR(dev->dev, "%s: offending task: unknown\n", gpu->name);
391 
392 		msm_rd_dump_submit(priv->hangrd, submit, NULL);
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 out_unlock:
446 	mutex_unlock(&gpu->lock);
447 
448 	msm_gpu_retire(gpu);
449 }
450 
451 static void fault_worker(struct kthread_work *work)
452 {
453 	struct msm_gpu *gpu = container_of(work, struct msm_gpu, fault_work);
454 	struct msm_gem_submit *submit;
455 	struct msm_ringbuffer *cur_ring = gpu->funcs->active_ring(gpu);
456 	char *comm = NULL, *cmd = NULL;
457 
458 	mutex_lock(&gpu->lock);
459 
460 	submit = find_submit(cur_ring, cur_ring->memptrs->fence + 1);
461 	if (submit && submit->fault_dumped)
462 		goto resume_smmu;
463 
464 	if (submit) {
465 		get_comm_cmdline(submit, &comm, &cmd);
466 
467 		/*
468 		 * When we get GPU iova faults, we can get 1000s of them,
469 		 * but we really only want to log the first one.
470 		 */
471 		submit->fault_dumped = true;
472 	}
473 
474 	/* Record the crash state */
475 	pm_runtime_get_sync(&gpu->pdev->dev);
476 	msm_gpu_crashstate_capture(gpu, submit, comm, cmd);
477 	pm_runtime_put_sync(&gpu->pdev->dev);
478 
479 	kfree(cmd);
480 	kfree(comm);
481 
482 resume_smmu:
483 	memset(&gpu->fault_info, 0, sizeof(gpu->fault_info));
484 	gpu->aspace->mmu->funcs->resume_translation(gpu->aspace->mmu);
485 
486 	mutex_unlock(&gpu->lock);
487 }
488 
489 static void hangcheck_timer_reset(struct msm_gpu *gpu)
490 {
491 	struct msm_drm_private *priv = gpu->dev->dev_private;
492 	mod_timer(&gpu->hangcheck_timer,
493 			round_jiffies_up(jiffies + msecs_to_jiffies(priv->hangcheck_period)));
494 }
495 
496 static bool made_progress(struct msm_gpu *gpu, struct msm_ringbuffer *ring)
497 {
498 	if (ring->hangcheck_progress_retries >= DRM_MSM_HANGCHECK_PROGRESS_RETRIES)
499 		return false;
500 
501 	if (!gpu->funcs->progress)
502 		return false;
503 
504 	if (!gpu->funcs->progress(gpu, ring))
505 		return false;
506 
507 	ring->hangcheck_progress_retries++;
508 	return true;
509 }
510 
511 static void hangcheck_handler(struct timer_list *t)
512 {
513 	struct msm_gpu *gpu = from_timer(gpu, t, hangcheck_timer);
514 	struct drm_device *dev = gpu->dev;
515 	struct msm_ringbuffer *ring = gpu->funcs->active_ring(gpu);
516 	uint32_t fence = ring->memptrs->fence;
517 
518 	if (fence != ring->hangcheck_fence) {
519 		/* some progress has been made.. ya! */
520 		ring->hangcheck_fence = fence;
521 		ring->hangcheck_progress_retries = 0;
522 	} else if (fence_before(fence, ring->fctx->last_fence) &&
523 			!made_progress(gpu, ring)) {
524 		/* no progress and not done.. hung! */
525 		ring->hangcheck_fence = fence;
526 		ring->hangcheck_progress_retries = 0;
527 		DRM_DEV_ERROR(dev->dev, "%s: hangcheck detected gpu lockup rb %d!\n",
528 				gpu->name, ring->id);
529 		DRM_DEV_ERROR(dev->dev, "%s:     completed fence: %u\n",
530 				gpu->name, fence);
531 		DRM_DEV_ERROR(dev->dev, "%s:     submitted fence: %u\n",
532 				gpu->name, ring->fctx->last_fence);
533 
534 		kthread_queue_work(gpu->worker, &gpu->recover_work);
535 	}
536 
537 	/* if still more pending work, reset the hangcheck timer: */
538 	if (fence_after(ring->fctx->last_fence, ring->hangcheck_fence))
539 		hangcheck_timer_reset(gpu);
540 
541 	/* workaround for missing irq: */
542 	msm_gpu_retire(gpu);
543 }
544 
545 /*
546  * Performance Counters:
547  */
548 
549 /* called under perf_lock */
550 static int update_hw_cntrs(struct msm_gpu *gpu, uint32_t ncntrs, uint32_t *cntrs)
551 {
552 	uint32_t current_cntrs[ARRAY_SIZE(gpu->last_cntrs)];
553 	int i, n = min(ncntrs, gpu->num_perfcntrs);
554 
555 	/* read current values: */
556 	for (i = 0; i < gpu->num_perfcntrs; i++)
557 		current_cntrs[i] = gpu_read(gpu, gpu->perfcntrs[i].sample_reg);
558 
559 	/* update cntrs: */
560 	for (i = 0; i < n; i++)
561 		cntrs[i] = current_cntrs[i] - gpu->last_cntrs[i];
562 
563 	/* save current values: */
564 	for (i = 0; i < gpu->num_perfcntrs; i++)
565 		gpu->last_cntrs[i] = current_cntrs[i];
566 
567 	return n;
568 }
569 
570 static void update_sw_cntrs(struct msm_gpu *gpu)
571 {
572 	ktime_t time;
573 	uint32_t elapsed;
574 	unsigned long flags;
575 
576 	spin_lock_irqsave(&gpu->perf_lock, flags);
577 	if (!gpu->perfcntr_active)
578 		goto out;
579 
580 	time = ktime_get();
581 	elapsed = ktime_to_us(ktime_sub(time, gpu->last_sample.time));
582 
583 	gpu->totaltime += elapsed;
584 	if (gpu->last_sample.active)
585 		gpu->activetime += elapsed;
586 
587 	gpu->last_sample.active = msm_gpu_active(gpu);
588 	gpu->last_sample.time = time;
589 
590 out:
591 	spin_unlock_irqrestore(&gpu->perf_lock, flags);
592 }
593 
594 void msm_gpu_perfcntr_start(struct msm_gpu *gpu)
595 {
596 	unsigned long flags;
597 
598 	pm_runtime_get_sync(&gpu->pdev->dev);
599 
600 	spin_lock_irqsave(&gpu->perf_lock, flags);
601 	/* we could dynamically enable/disable perfcntr registers too.. */
602 	gpu->last_sample.active = msm_gpu_active(gpu);
603 	gpu->last_sample.time = ktime_get();
604 	gpu->activetime = gpu->totaltime = 0;
605 	gpu->perfcntr_active = true;
606 	update_hw_cntrs(gpu, 0, NULL);
607 	spin_unlock_irqrestore(&gpu->perf_lock, flags);
608 }
609 
610 void msm_gpu_perfcntr_stop(struct msm_gpu *gpu)
611 {
612 	gpu->perfcntr_active = false;
613 	pm_runtime_put_sync(&gpu->pdev->dev);
614 }
615 
616 /* returns -errno or # of cntrs sampled */
617 int msm_gpu_perfcntr_sample(struct msm_gpu *gpu, uint32_t *activetime,
618 		uint32_t *totaltime, uint32_t ncntrs, uint32_t *cntrs)
619 {
620 	unsigned long flags;
621 	int ret;
622 
623 	spin_lock_irqsave(&gpu->perf_lock, flags);
624 
625 	if (!gpu->perfcntr_active) {
626 		ret = -EINVAL;
627 		goto out;
628 	}
629 
630 	*activetime = gpu->activetime;
631 	*totaltime = gpu->totaltime;
632 
633 	gpu->activetime = gpu->totaltime = 0;
634 
635 	ret = update_hw_cntrs(gpu, ncntrs, cntrs);
636 
637 out:
638 	spin_unlock_irqrestore(&gpu->perf_lock, flags);
639 
640 	return ret;
641 }
642 
643 /*
644  * Cmdstream submission/retirement:
645  */
646 
647 static void retire_submit(struct msm_gpu *gpu, struct msm_ringbuffer *ring,
648 		struct msm_gem_submit *submit)
649 {
650 	int index = submit->seqno % MSM_GPU_SUBMIT_STATS_COUNT;
651 	volatile struct msm_gpu_submit_stats *stats;
652 	u64 elapsed, clock = 0, cycles;
653 	unsigned long flags;
654 
655 	stats = &ring->memptrs->stats[index];
656 	/* Convert 19.2Mhz alwayson ticks to nanoseconds for elapsed time */
657 	elapsed = (stats->alwayson_end - stats->alwayson_start) * 10000;
658 	do_div(elapsed, 192);
659 
660 	cycles = stats->cpcycles_end - stats->cpcycles_start;
661 
662 	/* Calculate the clock frequency from the number of CP cycles */
663 	if (elapsed) {
664 		clock = cycles * 1000;
665 		do_div(clock, elapsed);
666 	}
667 
668 	submit->queue->ctx->elapsed_ns += elapsed;
669 	submit->queue->ctx->cycles     += cycles;
670 
671 	trace_msm_gpu_submit_retired(submit, elapsed, clock,
672 		stats->alwayson_start, stats->alwayson_end);
673 
674 	msm_submit_retire(submit);
675 
676 	pm_runtime_mark_last_busy(&gpu->pdev->dev);
677 
678 	spin_lock_irqsave(&ring->submit_lock, flags);
679 	list_del(&submit->node);
680 	spin_unlock_irqrestore(&ring->submit_lock, flags);
681 
682 	/* Update devfreq on transition from active->idle: */
683 	mutex_lock(&gpu->active_lock);
684 	gpu->active_submits--;
685 	WARN_ON(gpu->active_submits < 0);
686 	if (!gpu->active_submits) {
687 		msm_devfreq_idle(gpu);
688 		pm_runtime_put_autosuspend(&gpu->pdev->dev);
689 	}
690 
691 	mutex_unlock(&gpu->active_lock);
692 
693 	msm_gem_submit_put(submit);
694 }
695 
696 static void retire_submits(struct msm_gpu *gpu)
697 {
698 	int i;
699 
700 	/* Retire the commits starting with highest priority */
701 	for (i = 0; i < gpu->nr_rings; i++) {
702 		struct msm_ringbuffer *ring = gpu->rb[i];
703 
704 		while (true) {
705 			struct msm_gem_submit *submit = NULL;
706 			unsigned long flags;
707 
708 			spin_lock_irqsave(&ring->submit_lock, flags);
709 			submit = list_first_entry_or_null(&ring->submits,
710 					struct msm_gem_submit, node);
711 			spin_unlock_irqrestore(&ring->submit_lock, flags);
712 
713 			/*
714 			 * If no submit, we are done.  If submit->fence hasn't
715 			 * been signalled, then later submits are not signalled
716 			 * either, so we are also done.
717 			 */
718 			if (submit && dma_fence_is_signaled(submit->hw_fence)) {
719 				retire_submit(gpu, ring, submit);
720 			} else {
721 				break;
722 			}
723 		}
724 	}
725 
726 	wake_up_all(&gpu->retire_event);
727 }
728 
729 static void retire_worker(struct kthread_work *work)
730 {
731 	struct msm_gpu *gpu = container_of(work, struct msm_gpu, retire_work);
732 
733 	retire_submits(gpu);
734 }
735 
736 /* call from irq handler to schedule work to retire bo's */
737 void msm_gpu_retire(struct msm_gpu *gpu)
738 {
739 	int i;
740 
741 	for (i = 0; i < gpu->nr_rings; i++)
742 		msm_update_fence(gpu->rb[i]->fctx, gpu->rb[i]->memptrs->fence);
743 
744 	kthread_queue_work(gpu->worker, &gpu->retire_work);
745 	update_sw_cntrs(gpu);
746 }
747 
748 /* add bo's to gpu's ring, and kick gpu: */
749 void msm_gpu_submit(struct msm_gpu *gpu, struct msm_gem_submit *submit)
750 {
751 	struct msm_ringbuffer *ring = submit->ring;
752 	unsigned long flags;
753 
754 	WARN_ON(!mutex_is_locked(&gpu->lock));
755 
756 	pm_runtime_get_sync(&gpu->pdev->dev);
757 
758 	msm_gpu_hw_init(gpu);
759 
760 	submit->seqno = submit->hw_fence->seqno;
761 
762 	update_sw_cntrs(gpu);
763 
764 	/*
765 	 * ring->submits holds a ref to the submit, to deal with the case
766 	 * that a submit completes before msm_ioctl_gem_submit() returns.
767 	 */
768 	msm_gem_submit_get(submit);
769 
770 	spin_lock_irqsave(&ring->submit_lock, flags);
771 	list_add_tail(&submit->node, &ring->submits);
772 	spin_unlock_irqrestore(&ring->submit_lock, flags);
773 
774 	/* Update devfreq on transition from idle->active: */
775 	mutex_lock(&gpu->active_lock);
776 	if (!gpu->active_submits) {
777 		pm_runtime_get(&gpu->pdev->dev);
778 		msm_devfreq_active(gpu);
779 	}
780 	gpu->active_submits++;
781 	mutex_unlock(&gpu->active_lock);
782 
783 	gpu->funcs->submit(gpu, submit);
784 	gpu->cur_ctx_seqno = submit->queue->ctx->seqno;
785 
786 	pm_runtime_put(&gpu->pdev->dev);
787 	hangcheck_timer_reset(gpu);
788 }
789 
790 /*
791  * Init/Cleanup:
792  */
793 
794 static irqreturn_t irq_handler(int irq, void *data)
795 {
796 	struct msm_gpu *gpu = data;
797 	return gpu->funcs->irq(gpu);
798 }
799 
800 static int get_clocks(struct platform_device *pdev, struct msm_gpu *gpu)
801 {
802 	int ret = devm_clk_bulk_get_all(&pdev->dev, &gpu->grp_clks);
803 
804 	if (ret < 1) {
805 		gpu->nr_clocks = 0;
806 		return ret;
807 	}
808 
809 	gpu->nr_clocks = ret;
810 
811 	gpu->core_clk = msm_clk_bulk_get_clock(gpu->grp_clks,
812 		gpu->nr_clocks, "core");
813 
814 	gpu->rbbmtimer_clk = msm_clk_bulk_get_clock(gpu->grp_clks,
815 		gpu->nr_clocks, "rbbmtimer");
816 
817 	return 0;
818 }
819 
820 /* Return a new address space for a msm_drm_private instance */
821 struct msm_gem_address_space *
822 msm_gpu_create_private_address_space(struct msm_gpu *gpu, struct task_struct *task)
823 {
824 	struct msm_gem_address_space *aspace = NULL;
825 	if (!gpu)
826 		return NULL;
827 
828 	/*
829 	 * If the target doesn't support private address spaces then return
830 	 * the global one
831 	 */
832 	if (gpu->funcs->create_private_address_space) {
833 		aspace = gpu->funcs->create_private_address_space(gpu);
834 		if (!IS_ERR(aspace))
835 			aspace->pid = get_pid(task_pid(task));
836 	}
837 
838 	if (IS_ERR_OR_NULL(aspace))
839 		aspace = msm_gem_address_space_get(gpu->aspace);
840 
841 	return aspace;
842 }
843 
844 int msm_gpu_init(struct drm_device *drm, struct platform_device *pdev,
845 		struct msm_gpu *gpu, const struct msm_gpu_funcs *funcs,
846 		const char *name, struct msm_gpu_config *config)
847 {
848 	struct msm_drm_private *priv = drm->dev_private;
849 	int i, ret, nr_rings = config->nr_rings;
850 	void *memptrs;
851 	uint64_t memptrs_iova;
852 
853 	if (WARN_ON(gpu->num_perfcntrs > ARRAY_SIZE(gpu->last_cntrs)))
854 		gpu->num_perfcntrs = ARRAY_SIZE(gpu->last_cntrs);
855 
856 	gpu->dev = drm;
857 	gpu->funcs = funcs;
858 	gpu->name = name;
859 
860 	gpu->worker = kthread_create_worker(0, "gpu-worker");
861 	if (IS_ERR(gpu->worker)) {
862 		ret = PTR_ERR(gpu->worker);
863 		gpu->worker = NULL;
864 		goto fail;
865 	}
866 
867 	sched_set_fifo_low(gpu->worker->task);
868 
869 	mutex_init(&gpu->active_lock);
870 	mutex_init(&gpu->lock);
871 	init_waitqueue_head(&gpu->retire_event);
872 	kthread_init_work(&gpu->retire_work, retire_worker);
873 	kthread_init_work(&gpu->recover_work, recover_worker);
874 	kthread_init_work(&gpu->fault_work, fault_worker);
875 
876 	priv->hangcheck_period = DRM_MSM_HANGCHECK_DEFAULT_PERIOD;
877 
878 	/*
879 	 * If progress detection is supported, halve the hangcheck timer
880 	 * duration, as it takes two iterations of the hangcheck handler
881 	 * to detect a hang.
882 	 */
883 	if (funcs->progress)
884 		priv->hangcheck_period /= 2;
885 
886 	timer_setup(&gpu->hangcheck_timer, hangcheck_handler, 0);
887 
888 	spin_lock_init(&gpu->perf_lock);
889 
890 
891 	/* Map registers: */
892 	gpu->mmio = msm_ioremap(pdev, config->ioname);
893 	if (IS_ERR(gpu->mmio)) {
894 		ret = PTR_ERR(gpu->mmio);
895 		goto fail;
896 	}
897 
898 	/* Get Interrupt: */
899 	gpu->irq = platform_get_irq(pdev, 0);
900 	if (gpu->irq < 0) {
901 		ret = gpu->irq;
902 		goto fail;
903 	}
904 
905 	ret = devm_request_irq(&pdev->dev, gpu->irq, irq_handler,
906 			IRQF_TRIGGER_HIGH, "gpu-irq", gpu);
907 	if (ret) {
908 		DRM_DEV_ERROR(drm->dev, "failed to request IRQ%u: %d\n", gpu->irq, ret);
909 		goto fail;
910 	}
911 
912 	ret = get_clocks(pdev, gpu);
913 	if (ret)
914 		goto fail;
915 
916 	gpu->ebi1_clk = msm_clk_get(pdev, "bus");
917 	DBG("ebi1_clk: %p", gpu->ebi1_clk);
918 	if (IS_ERR(gpu->ebi1_clk))
919 		gpu->ebi1_clk = NULL;
920 
921 	/* Acquire regulators: */
922 	gpu->gpu_reg = devm_regulator_get(&pdev->dev, "vdd");
923 	DBG("gpu_reg: %p", gpu->gpu_reg);
924 	if (IS_ERR(gpu->gpu_reg))
925 		gpu->gpu_reg = NULL;
926 
927 	gpu->gpu_cx = devm_regulator_get(&pdev->dev, "vddcx");
928 	DBG("gpu_cx: %p", gpu->gpu_cx);
929 	if (IS_ERR(gpu->gpu_cx))
930 		gpu->gpu_cx = NULL;
931 
932 	gpu->pdev = pdev;
933 	platform_set_drvdata(pdev, &gpu->adreno_smmu);
934 
935 	msm_devfreq_init(gpu);
936 
937 
938 	gpu->aspace = gpu->funcs->create_address_space(gpu, pdev);
939 
940 	if (gpu->aspace == NULL)
941 		DRM_DEV_INFO(drm->dev, "%s: no IOMMU, fallback to VRAM carveout!\n", name);
942 	else if (IS_ERR(gpu->aspace)) {
943 		ret = PTR_ERR(gpu->aspace);
944 		goto fail;
945 	}
946 
947 	memptrs = msm_gem_kernel_new(drm,
948 		sizeof(struct msm_rbmemptrs) * nr_rings,
949 		check_apriv(gpu, MSM_BO_WC), gpu->aspace, &gpu->memptrs_bo,
950 		&memptrs_iova);
951 
952 	if (IS_ERR(memptrs)) {
953 		ret = PTR_ERR(memptrs);
954 		DRM_DEV_ERROR(drm->dev, "could not allocate memptrs: %d\n", ret);
955 		goto fail;
956 	}
957 
958 	msm_gem_object_set_name(gpu->memptrs_bo, "memptrs");
959 
960 	if (nr_rings > ARRAY_SIZE(gpu->rb)) {
961 		DRM_DEV_INFO_ONCE(drm->dev, "Only creating %zu ringbuffers\n",
962 			ARRAY_SIZE(gpu->rb));
963 		nr_rings = ARRAY_SIZE(gpu->rb);
964 	}
965 
966 	/* Create ringbuffer(s): */
967 	for (i = 0; i < nr_rings; i++) {
968 		gpu->rb[i] = msm_ringbuffer_new(gpu, i, memptrs, memptrs_iova);
969 
970 		if (IS_ERR(gpu->rb[i])) {
971 			ret = PTR_ERR(gpu->rb[i]);
972 			DRM_DEV_ERROR(drm->dev,
973 				"could not create ringbuffer %d: %d\n", i, ret);
974 			goto fail;
975 		}
976 
977 		memptrs += sizeof(struct msm_rbmemptrs);
978 		memptrs_iova += sizeof(struct msm_rbmemptrs);
979 	}
980 
981 	gpu->nr_rings = nr_rings;
982 
983 	refcount_set(&gpu->sysprof_active, 1);
984 
985 	return 0;
986 
987 fail:
988 	for (i = 0; i < ARRAY_SIZE(gpu->rb); i++)  {
989 		msm_ringbuffer_destroy(gpu->rb[i]);
990 		gpu->rb[i] = NULL;
991 	}
992 
993 	msm_gem_kernel_put(gpu->memptrs_bo, gpu->aspace);
994 
995 	platform_set_drvdata(pdev, NULL);
996 	return ret;
997 }
998 
999 void msm_gpu_cleanup(struct msm_gpu *gpu)
1000 {
1001 	int i;
1002 
1003 	DBG("%s", gpu->name);
1004 
1005 	for (i = 0; i < ARRAY_SIZE(gpu->rb); i++) {
1006 		msm_ringbuffer_destroy(gpu->rb[i]);
1007 		gpu->rb[i] = NULL;
1008 	}
1009 
1010 	msm_gem_kernel_put(gpu->memptrs_bo, gpu->aspace);
1011 
1012 	if (!IS_ERR_OR_NULL(gpu->aspace)) {
1013 		gpu->aspace->mmu->funcs->detach(gpu->aspace->mmu);
1014 		msm_gem_address_space_put(gpu->aspace);
1015 	}
1016 
1017 	if (gpu->worker) {
1018 		kthread_destroy_worker(gpu->worker);
1019 	}
1020 
1021 	msm_devfreq_cleanup(gpu);
1022 
1023 	platform_set_drvdata(gpu->pdev, NULL);
1024 }
1025