xref: /openbmc/linux/drivers/gpu/drm/msm/msm_gpu.c (revision e8254a8e)
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 bool made_progress(struct msm_gpu *gpu, struct msm_ringbuffer *ring)
496 {
497 	if (ring->hangcheck_progress_retries >= DRM_MSM_HANGCHECK_PROGRESS_RETRIES)
498 		return false;
499 
500 	if (!gpu->funcs->progress)
501 		return false;
502 
503 	if (!gpu->funcs->progress(gpu, ring))
504 		return false;
505 
506 	ring->hangcheck_progress_retries++;
507 	return true;
508 }
509 
510 static void hangcheck_handler(struct timer_list *t)
511 {
512 	struct msm_gpu *gpu = from_timer(gpu, t, hangcheck_timer);
513 	struct drm_device *dev = gpu->dev;
514 	struct msm_ringbuffer *ring = gpu->funcs->active_ring(gpu);
515 	uint32_t fence = ring->memptrs->fence;
516 
517 	if (fence != ring->hangcheck_fence) {
518 		/* some progress has been made.. ya! */
519 		ring->hangcheck_fence = fence;
520 		ring->hangcheck_progress_retries = 0;
521 	} else if (fence_before(fence, ring->fctx->last_fence) &&
522 			!made_progress(gpu, ring)) {
523 		/* no progress and not done.. hung! */
524 		ring->hangcheck_fence = fence;
525 		ring->hangcheck_progress_retries = 0;
526 		DRM_DEV_ERROR(dev->dev, "%s: hangcheck detected gpu lockup rb %d!\n",
527 				gpu->name, ring->id);
528 		DRM_DEV_ERROR(dev->dev, "%s:     completed fence: %u\n",
529 				gpu->name, fence);
530 		DRM_DEV_ERROR(dev->dev, "%s:     submitted fence: %u\n",
531 				gpu->name, ring->fctx->last_fence);
532 
533 		kthread_queue_work(gpu->worker, &gpu->recover_work);
534 	}
535 
536 	/* if still more pending work, reset the hangcheck timer: */
537 	if (fence_after(ring->fctx->last_fence, ring->hangcheck_fence))
538 		hangcheck_timer_reset(gpu);
539 
540 	/* workaround for missing irq: */
541 	msm_gpu_retire(gpu);
542 }
543 
544 /*
545  * Performance Counters:
546  */
547 
548 /* called under perf_lock */
549 static int update_hw_cntrs(struct msm_gpu *gpu, uint32_t ncntrs, uint32_t *cntrs)
550 {
551 	uint32_t current_cntrs[ARRAY_SIZE(gpu->last_cntrs)];
552 	int i, n = min(ncntrs, gpu->num_perfcntrs);
553 
554 	/* read current values: */
555 	for (i = 0; i < gpu->num_perfcntrs; i++)
556 		current_cntrs[i] = gpu_read(gpu, gpu->perfcntrs[i].sample_reg);
557 
558 	/* update cntrs: */
559 	for (i = 0; i < n; i++)
560 		cntrs[i] = current_cntrs[i] - gpu->last_cntrs[i];
561 
562 	/* save current values: */
563 	for (i = 0; i < gpu->num_perfcntrs; i++)
564 		gpu->last_cntrs[i] = current_cntrs[i];
565 
566 	return n;
567 }
568 
569 static void update_sw_cntrs(struct msm_gpu *gpu)
570 {
571 	ktime_t time;
572 	uint32_t elapsed;
573 	unsigned long flags;
574 
575 	spin_lock_irqsave(&gpu->perf_lock, flags);
576 	if (!gpu->perfcntr_active)
577 		goto out;
578 
579 	time = ktime_get();
580 	elapsed = ktime_to_us(ktime_sub(time, gpu->last_sample.time));
581 
582 	gpu->totaltime += elapsed;
583 	if (gpu->last_sample.active)
584 		gpu->activetime += elapsed;
585 
586 	gpu->last_sample.active = msm_gpu_active(gpu);
587 	gpu->last_sample.time = time;
588 
589 out:
590 	spin_unlock_irqrestore(&gpu->perf_lock, flags);
591 }
592 
593 void msm_gpu_perfcntr_start(struct msm_gpu *gpu)
594 {
595 	unsigned long flags;
596 
597 	pm_runtime_get_sync(&gpu->pdev->dev);
598 
599 	spin_lock_irqsave(&gpu->perf_lock, flags);
600 	/* we could dynamically enable/disable perfcntr registers too.. */
601 	gpu->last_sample.active = msm_gpu_active(gpu);
602 	gpu->last_sample.time = ktime_get();
603 	gpu->activetime = gpu->totaltime = 0;
604 	gpu->perfcntr_active = true;
605 	update_hw_cntrs(gpu, 0, NULL);
606 	spin_unlock_irqrestore(&gpu->perf_lock, flags);
607 }
608 
609 void msm_gpu_perfcntr_stop(struct msm_gpu *gpu)
610 {
611 	gpu->perfcntr_active = false;
612 	pm_runtime_put_sync(&gpu->pdev->dev);
613 }
614 
615 /* returns -errno or # of cntrs sampled */
616 int msm_gpu_perfcntr_sample(struct msm_gpu *gpu, uint32_t *activetime,
617 		uint32_t *totaltime, uint32_t ncntrs, uint32_t *cntrs)
618 {
619 	unsigned long flags;
620 	int ret;
621 
622 	spin_lock_irqsave(&gpu->perf_lock, flags);
623 
624 	if (!gpu->perfcntr_active) {
625 		ret = -EINVAL;
626 		goto out;
627 	}
628 
629 	*activetime = gpu->activetime;
630 	*totaltime = gpu->totaltime;
631 
632 	gpu->activetime = gpu->totaltime = 0;
633 
634 	ret = update_hw_cntrs(gpu, ncntrs, cntrs);
635 
636 out:
637 	spin_unlock_irqrestore(&gpu->perf_lock, flags);
638 
639 	return ret;
640 }
641 
642 /*
643  * Cmdstream submission/retirement:
644  */
645 
646 static void retire_submit(struct msm_gpu *gpu, struct msm_ringbuffer *ring,
647 		struct msm_gem_submit *submit)
648 {
649 	int index = submit->seqno % MSM_GPU_SUBMIT_STATS_COUNT;
650 	volatile struct msm_gpu_submit_stats *stats;
651 	u64 elapsed, clock = 0, cycles;
652 	unsigned long flags;
653 
654 	stats = &ring->memptrs->stats[index];
655 	/* Convert 19.2Mhz alwayson ticks to nanoseconds for elapsed time */
656 	elapsed = (stats->alwayson_end - stats->alwayson_start) * 10000;
657 	do_div(elapsed, 192);
658 
659 	cycles = stats->cpcycles_end - stats->cpcycles_start;
660 
661 	/* Calculate the clock frequency from the number of CP cycles */
662 	if (elapsed) {
663 		clock = cycles * 1000;
664 		do_div(clock, elapsed);
665 	}
666 
667 	submit->queue->ctx->elapsed_ns += elapsed;
668 	submit->queue->ctx->cycles     += cycles;
669 
670 	trace_msm_gpu_submit_retired(submit, elapsed, clock,
671 		stats->alwayson_start, stats->alwayson_end);
672 
673 	msm_submit_retire(submit);
674 
675 	pm_runtime_mark_last_busy(&gpu->pdev->dev);
676 
677 	spin_lock_irqsave(&ring->submit_lock, flags);
678 	list_del(&submit->node);
679 	spin_unlock_irqrestore(&ring->submit_lock, flags);
680 
681 	/* Update devfreq on transition from active->idle: */
682 	mutex_lock(&gpu->active_lock);
683 	gpu->active_submits--;
684 	WARN_ON(gpu->active_submits < 0);
685 	if (!gpu->active_submits) {
686 		msm_devfreq_idle(gpu);
687 		pm_runtime_put_autosuspend(&gpu->pdev->dev);
688 	}
689 
690 	mutex_unlock(&gpu->active_lock);
691 
692 	msm_gem_submit_put(submit);
693 }
694 
695 static void retire_submits(struct msm_gpu *gpu)
696 {
697 	int i;
698 
699 	/* Retire the commits starting with highest priority */
700 	for (i = 0; i < gpu->nr_rings; i++) {
701 		struct msm_ringbuffer *ring = gpu->rb[i];
702 
703 		while (true) {
704 			struct msm_gem_submit *submit = NULL;
705 			unsigned long flags;
706 
707 			spin_lock_irqsave(&ring->submit_lock, flags);
708 			submit = list_first_entry_or_null(&ring->submits,
709 					struct msm_gem_submit, node);
710 			spin_unlock_irqrestore(&ring->submit_lock, flags);
711 
712 			/*
713 			 * If no submit, we are done.  If submit->fence hasn't
714 			 * been signalled, then later submits are not signalled
715 			 * either, so we are also done.
716 			 */
717 			if (submit && dma_fence_is_signaled(submit->hw_fence)) {
718 				retire_submit(gpu, ring, submit);
719 			} else {
720 				break;
721 			}
722 		}
723 	}
724 
725 	wake_up_all(&gpu->retire_event);
726 }
727 
728 static void retire_worker(struct kthread_work *work)
729 {
730 	struct msm_gpu *gpu = container_of(work, struct msm_gpu, retire_work);
731 
732 	retire_submits(gpu);
733 }
734 
735 /* call from irq handler to schedule work to retire bo's */
736 void msm_gpu_retire(struct msm_gpu *gpu)
737 {
738 	int i;
739 
740 	for (i = 0; i < gpu->nr_rings; i++)
741 		msm_update_fence(gpu->rb[i]->fctx, gpu->rb[i]->memptrs->fence);
742 
743 	kthread_queue_work(gpu->worker, &gpu->retire_work);
744 	update_sw_cntrs(gpu);
745 }
746 
747 /* add bo's to gpu's ring, and kick gpu: */
748 void msm_gpu_submit(struct msm_gpu *gpu, struct msm_gem_submit *submit)
749 {
750 	struct drm_device *dev = gpu->dev;
751 	struct msm_drm_private *priv = dev->dev_private;
752 	struct msm_ringbuffer *ring = submit->ring;
753 	unsigned long flags;
754 
755 	WARN_ON(!mutex_is_locked(&gpu->lock));
756 
757 	pm_runtime_get_sync(&gpu->pdev->dev);
758 
759 	msm_gpu_hw_init(gpu);
760 
761 	submit->seqno = submit->hw_fence->seqno;
762 
763 	msm_rd_dump_submit(priv->rd, submit, NULL);
764 
765 	update_sw_cntrs(gpu);
766 
767 	/*
768 	 * ring->submits holds a ref to the submit, to deal with the case
769 	 * that a submit completes before msm_ioctl_gem_submit() returns.
770 	 */
771 	msm_gem_submit_get(submit);
772 
773 	spin_lock_irqsave(&ring->submit_lock, flags);
774 	list_add_tail(&submit->node, &ring->submits);
775 	spin_unlock_irqrestore(&ring->submit_lock, flags);
776 
777 	/* Update devfreq on transition from idle->active: */
778 	mutex_lock(&gpu->active_lock);
779 	if (!gpu->active_submits) {
780 		pm_runtime_get(&gpu->pdev->dev);
781 		msm_devfreq_active(gpu);
782 	}
783 	gpu->active_submits++;
784 	mutex_unlock(&gpu->active_lock);
785 
786 	gpu->funcs->submit(gpu, submit);
787 	gpu->cur_ctx_seqno = submit->queue->ctx->seqno;
788 
789 	pm_runtime_put(&gpu->pdev->dev);
790 	hangcheck_timer_reset(gpu);
791 }
792 
793 /*
794  * Init/Cleanup:
795  */
796 
797 static irqreturn_t irq_handler(int irq, void *data)
798 {
799 	struct msm_gpu *gpu = data;
800 	return gpu->funcs->irq(gpu);
801 }
802 
803 static int get_clocks(struct platform_device *pdev, struct msm_gpu *gpu)
804 {
805 	int ret = devm_clk_bulk_get_all(&pdev->dev, &gpu->grp_clks);
806 
807 	if (ret < 1) {
808 		gpu->nr_clocks = 0;
809 		return ret;
810 	}
811 
812 	gpu->nr_clocks = ret;
813 
814 	gpu->core_clk = msm_clk_bulk_get_clock(gpu->grp_clks,
815 		gpu->nr_clocks, "core");
816 
817 	gpu->rbbmtimer_clk = msm_clk_bulk_get_clock(gpu->grp_clks,
818 		gpu->nr_clocks, "rbbmtimer");
819 
820 	return 0;
821 }
822 
823 /* Return a new address space for a msm_drm_private instance */
824 struct msm_gem_address_space *
825 msm_gpu_create_private_address_space(struct msm_gpu *gpu, struct task_struct *task)
826 {
827 	struct msm_gem_address_space *aspace = NULL;
828 	if (!gpu)
829 		return NULL;
830 
831 	/*
832 	 * If the target doesn't support private address spaces then return
833 	 * the global one
834 	 */
835 	if (gpu->funcs->create_private_address_space) {
836 		aspace = gpu->funcs->create_private_address_space(gpu);
837 		if (!IS_ERR(aspace))
838 			aspace->pid = get_pid(task_pid(task));
839 	}
840 
841 	if (IS_ERR_OR_NULL(aspace))
842 		aspace = msm_gem_address_space_get(gpu->aspace);
843 
844 	return aspace;
845 }
846 
847 int msm_gpu_init(struct drm_device *drm, struct platform_device *pdev,
848 		struct msm_gpu *gpu, const struct msm_gpu_funcs *funcs,
849 		const char *name, struct msm_gpu_config *config)
850 {
851 	struct msm_drm_private *priv = drm->dev_private;
852 	int i, ret, nr_rings = config->nr_rings;
853 	void *memptrs;
854 	uint64_t memptrs_iova;
855 
856 	if (WARN_ON(gpu->num_perfcntrs > ARRAY_SIZE(gpu->last_cntrs)))
857 		gpu->num_perfcntrs = ARRAY_SIZE(gpu->last_cntrs);
858 
859 	gpu->dev = drm;
860 	gpu->funcs = funcs;
861 	gpu->name = name;
862 
863 	gpu->worker = kthread_create_worker(0, "gpu-worker");
864 	if (IS_ERR(gpu->worker)) {
865 		ret = PTR_ERR(gpu->worker);
866 		gpu->worker = NULL;
867 		goto fail;
868 	}
869 
870 	sched_set_fifo_low(gpu->worker->task);
871 
872 	mutex_init(&gpu->active_lock);
873 	mutex_init(&gpu->lock);
874 	init_waitqueue_head(&gpu->retire_event);
875 	kthread_init_work(&gpu->retire_work, retire_worker);
876 	kthread_init_work(&gpu->recover_work, recover_worker);
877 	kthread_init_work(&gpu->fault_work, fault_worker);
878 
879 	priv->hangcheck_period = DRM_MSM_HANGCHECK_DEFAULT_PERIOD;
880 
881 	/*
882 	 * If progress detection is supported, halve the hangcheck timer
883 	 * duration, as it takes two iterations of the hangcheck handler
884 	 * to detect a hang.
885 	 */
886 	if (funcs->progress)
887 		priv->hangcheck_period /= 2;
888 
889 	timer_setup(&gpu->hangcheck_timer, hangcheck_handler, 0);
890 
891 	spin_lock_init(&gpu->perf_lock);
892 
893 
894 	/* Map registers: */
895 	gpu->mmio = msm_ioremap(pdev, config->ioname);
896 	if (IS_ERR(gpu->mmio)) {
897 		ret = PTR_ERR(gpu->mmio);
898 		goto fail;
899 	}
900 
901 	/* Get Interrupt: */
902 	gpu->irq = platform_get_irq(pdev, 0);
903 	if (gpu->irq < 0) {
904 		ret = gpu->irq;
905 		DRM_DEV_ERROR(drm->dev, "failed to get irq: %d\n", ret);
906 		goto fail;
907 	}
908 
909 	ret = devm_request_irq(&pdev->dev, gpu->irq, irq_handler,
910 			IRQF_TRIGGER_HIGH, "gpu-irq", gpu);
911 	if (ret) {
912 		DRM_DEV_ERROR(drm->dev, "failed to request IRQ%u: %d\n", gpu->irq, ret);
913 		goto fail;
914 	}
915 
916 	ret = get_clocks(pdev, gpu);
917 	if (ret)
918 		goto fail;
919 
920 	gpu->ebi1_clk = msm_clk_get(pdev, "bus");
921 	DBG("ebi1_clk: %p", gpu->ebi1_clk);
922 	if (IS_ERR(gpu->ebi1_clk))
923 		gpu->ebi1_clk = NULL;
924 
925 	/* Acquire regulators: */
926 	gpu->gpu_reg = devm_regulator_get(&pdev->dev, "vdd");
927 	DBG("gpu_reg: %p", gpu->gpu_reg);
928 	if (IS_ERR(gpu->gpu_reg))
929 		gpu->gpu_reg = NULL;
930 
931 	gpu->gpu_cx = devm_regulator_get(&pdev->dev, "vddcx");
932 	DBG("gpu_cx: %p", gpu->gpu_cx);
933 	if (IS_ERR(gpu->gpu_cx))
934 		gpu->gpu_cx = NULL;
935 
936 	gpu->cx_collapse = devm_reset_control_get_optional_exclusive(&pdev->dev,
937 			"cx_collapse");
938 
939 	gpu->pdev = pdev;
940 	platform_set_drvdata(pdev, &gpu->adreno_smmu);
941 
942 	msm_devfreq_init(gpu);
943 
944 
945 	gpu->aspace = gpu->funcs->create_address_space(gpu, pdev);
946 
947 	if (gpu->aspace == NULL)
948 		DRM_DEV_INFO(drm->dev, "%s: no IOMMU, fallback to VRAM carveout!\n", name);
949 	else if (IS_ERR(gpu->aspace)) {
950 		ret = PTR_ERR(gpu->aspace);
951 		goto fail;
952 	}
953 
954 	memptrs = msm_gem_kernel_new(drm,
955 		sizeof(struct msm_rbmemptrs) * nr_rings,
956 		check_apriv(gpu, MSM_BO_WC), gpu->aspace, &gpu->memptrs_bo,
957 		&memptrs_iova);
958 
959 	if (IS_ERR(memptrs)) {
960 		ret = PTR_ERR(memptrs);
961 		DRM_DEV_ERROR(drm->dev, "could not allocate memptrs: %d\n", ret);
962 		goto fail;
963 	}
964 
965 	msm_gem_object_set_name(gpu->memptrs_bo, "memptrs");
966 
967 	if (nr_rings > ARRAY_SIZE(gpu->rb)) {
968 		DRM_DEV_INFO_ONCE(drm->dev, "Only creating %zu ringbuffers\n",
969 			ARRAY_SIZE(gpu->rb));
970 		nr_rings = ARRAY_SIZE(gpu->rb);
971 	}
972 
973 	/* Create ringbuffer(s): */
974 	for (i = 0; i < nr_rings; i++) {
975 		gpu->rb[i] = msm_ringbuffer_new(gpu, i, memptrs, memptrs_iova);
976 
977 		if (IS_ERR(gpu->rb[i])) {
978 			ret = PTR_ERR(gpu->rb[i]);
979 			DRM_DEV_ERROR(drm->dev,
980 				"could not create ringbuffer %d: %d\n", i, ret);
981 			goto fail;
982 		}
983 
984 		memptrs += sizeof(struct msm_rbmemptrs);
985 		memptrs_iova += sizeof(struct msm_rbmemptrs);
986 	}
987 
988 	gpu->nr_rings = nr_rings;
989 
990 	refcount_set(&gpu->sysprof_active, 1);
991 
992 	return 0;
993 
994 fail:
995 	for (i = 0; i < ARRAY_SIZE(gpu->rb); i++)  {
996 		msm_ringbuffer_destroy(gpu->rb[i]);
997 		gpu->rb[i] = NULL;
998 	}
999 
1000 	msm_gem_kernel_put(gpu->memptrs_bo, gpu->aspace);
1001 
1002 	platform_set_drvdata(pdev, NULL);
1003 	return ret;
1004 }
1005 
1006 void msm_gpu_cleanup(struct msm_gpu *gpu)
1007 {
1008 	int i;
1009 
1010 	DBG("%s", gpu->name);
1011 
1012 	for (i = 0; i < ARRAY_SIZE(gpu->rb); i++) {
1013 		msm_ringbuffer_destroy(gpu->rb[i]);
1014 		gpu->rb[i] = NULL;
1015 	}
1016 
1017 	msm_gem_kernel_put(gpu->memptrs_bo, gpu->aspace);
1018 
1019 	if (!IS_ERR_OR_NULL(gpu->aspace)) {
1020 		gpu->aspace->mmu->funcs->detach(gpu->aspace->mmu);
1021 		msm_gem_address_space_put(gpu->aspace);
1022 	}
1023 
1024 	if (gpu->worker) {
1025 		kthread_destroy_worker(gpu->worker);
1026 	}
1027 
1028 	msm_devfreq_cleanup(gpu);
1029 
1030 	platform_set_drvdata(gpu->pdev, NULL);
1031 }
1032