xref: /linux/drivers/gpu/drm/msm/msm_gpu.c (revision 132db93572821ec2fdf81e354cc40f558faf7e4f)
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/pm_opp.h>
17 #include <linux/devfreq.h>
18 #include <linux/devcoredump.h>
19 #include <linux/sched/task.h>
20 
21 /*
22  * Power Management:
23  */
24 
25 static int msm_devfreq_target(struct device *dev, unsigned long *freq,
26 		u32 flags)
27 {
28 	struct msm_gpu *gpu = platform_get_drvdata(to_platform_device(dev));
29 	struct dev_pm_opp *opp;
30 
31 	opp = devfreq_recommended_opp(dev, freq, flags);
32 
33 	if (IS_ERR(opp))
34 		return PTR_ERR(opp);
35 
36 	if (gpu->funcs->gpu_set_freq)
37 		gpu->funcs->gpu_set_freq(gpu, (u64)*freq);
38 	else
39 		clk_set_rate(gpu->core_clk, *freq);
40 
41 	dev_pm_opp_put(opp);
42 
43 	return 0;
44 }
45 
46 static int msm_devfreq_get_dev_status(struct device *dev,
47 		struct devfreq_dev_status *status)
48 {
49 	struct msm_gpu *gpu = platform_get_drvdata(to_platform_device(dev));
50 	ktime_t time;
51 
52 	if (gpu->funcs->gpu_get_freq)
53 		status->current_frequency = gpu->funcs->gpu_get_freq(gpu);
54 	else
55 		status->current_frequency = clk_get_rate(gpu->core_clk);
56 
57 	status->busy_time = gpu->funcs->gpu_busy(gpu);
58 
59 	time = ktime_get();
60 	status->total_time = ktime_us_delta(time, gpu->devfreq.time);
61 	gpu->devfreq.time = time;
62 
63 	return 0;
64 }
65 
66 static int msm_devfreq_get_cur_freq(struct device *dev, unsigned long *freq)
67 {
68 	struct msm_gpu *gpu = platform_get_drvdata(to_platform_device(dev));
69 
70 	if (gpu->funcs->gpu_get_freq)
71 		*freq = gpu->funcs->gpu_get_freq(gpu);
72 	else
73 		*freq = clk_get_rate(gpu->core_clk);
74 
75 	return 0;
76 }
77 
78 static struct devfreq_dev_profile msm_devfreq_profile = {
79 	.polling_ms = 10,
80 	.target = msm_devfreq_target,
81 	.get_dev_status = msm_devfreq_get_dev_status,
82 	.get_cur_freq = msm_devfreq_get_cur_freq,
83 };
84 
85 static void msm_devfreq_init(struct msm_gpu *gpu)
86 {
87 	/* We need target support to do devfreq */
88 	if (!gpu->funcs->gpu_busy)
89 		return;
90 
91 	msm_devfreq_profile.initial_freq = gpu->fast_rate;
92 
93 	/*
94 	 * Don't set the freq_table or max_state and let devfreq build the table
95 	 * from OPP
96 	 */
97 
98 	gpu->devfreq.devfreq = devm_devfreq_add_device(&gpu->pdev->dev,
99 			&msm_devfreq_profile, DEVFREQ_GOV_SIMPLE_ONDEMAND,
100 			NULL);
101 
102 	if (IS_ERR(gpu->devfreq.devfreq)) {
103 		DRM_DEV_ERROR(&gpu->pdev->dev, "Couldn't initialize GPU devfreq\n");
104 		gpu->devfreq.devfreq = NULL;
105 	}
106 
107 	devfreq_suspend_device(gpu->devfreq.devfreq);
108 }
109 
110 static int enable_pwrrail(struct msm_gpu *gpu)
111 {
112 	struct drm_device *dev = gpu->dev;
113 	int ret = 0;
114 
115 	if (gpu->gpu_reg) {
116 		ret = regulator_enable(gpu->gpu_reg);
117 		if (ret) {
118 			DRM_DEV_ERROR(dev->dev, "failed to enable 'gpu_reg': %d\n", ret);
119 			return ret;
120 		}
121 	}
122 
123 	if (gpu->gpu_cx) {
124 		ret = regulator_enable(gpu->gpu_cx);
125 		if (ret) {
126 			DRM_DEV_ERROR(dev->dev, "failed to enable 'gpu_cx': %d\n", ret);
127 			return ret;
128 		}
129 	}
130 
131 	return 0;
132 }
133 
134 static int disable_pwrrail(struct msm_gpu *gpu)
135 {
136 	if (gpu->gpu_cx)
137 		regulator_disable(gpu->gpu_cx);
138 	if (gpu->gpu_reg)
139 		regulator_disable(gpu->gpu_reg);
140 	return 0;
141 }
142 
143 static int enable_clk(struct msm_gpu *gpu)
144 {
145 	if (gpu->core_clk && gpu->fast_rate)
146 		clk_set_rate(gpu->core_clk, gpu->fast_rate);
147 
148 	/* Set the RBBM timer rate to 19.2Mhz */
149 	if (gpu->rbbmtimer_clk)
150 		clk_set_rate(gpu->rbbmtimer_clk, 19200000);
151 
152 	return clk_bulk_prepare_enable(gpu->nr_clocks, gpu->grp_clks);
153 }
154 
155 static int disable_clk(struct msm_gpu *gpu)
156 {
157 	clk_bulk_disable_unprepare(gpu->nr_clocks, gpu->grp_clks);
158 
159 	/*
160 	 * Set the clock to a deliberately low rate. On older targets the clock
161 	 * speed had to be non zero to avoid problems. On newer targets this
162 	 * will be rounded down to zero anyway so it all works out.
163 	 */
164 	if (gpu->core_clk)
165 		clk_set_rate(gpu->core_clk, 27000000);
166 
167 	if (gpu->rbbmtimer_clk)
168 		clk_set_rate(gpu->rbbmtimer_clk, 0);
169 
170 	return 0;
171 }
172 
173 static int enable_axi(struct msm_gpu *gpu)
174 {
175 	if (gpu->ebi1_clk)
176 		clk_prepare_enable(gpu->ebi1_clk);
177 	return 0;
178 }
179 
180 static int disable_axi(struct msm_gpu *gpu)
181 {
182 	if (gpu->ebi1_clk)
183 		clk_disable_unprepare(gpu->ebi1_clk);
184 	return 0;
185 }
186 
187 void msm_gpu_resume_devfreq(struct msm_gpu *gpu)
188 {
189 	gpu->devfreq.busy_cycles = 0;
190 	gpu->devfreq.time = ktime_get();
191 
192 	devfreq_resume_device(gpu->devfreq.devfreq);
193 }
194 
195 int msm_gpu_pm_resume(struct msm_gpu *gpu)
196 {
197 	int ret;
198 
199 	DBG("%s", gpu->name);
200 
201 	ret = enable_pwrrail(gpu);
202 	if (ret)
203 		return ret;
204 
205 	ret = enable_clk(gpu);
206 	if (ret)
207 		return ret;
208 
209 	ret = enable_axi(gpu);
210 	if (ret)
211 		return ret;
212 
213 	msm_gpu_resume_devfreq(gpu);
214 
215 	gpu->needs_hw_init = true;
216 
217 	return 0;
218 }
219 
220 int msm_gpu_pm_suspend(struct msm_gpu *gpu)
221 {
222 	int ret;
223 
224 	DBG("%s", gpu->name);
225 
226 	devfreq_suspend_device(gpu->devfreq.devfreq);
227 
228 	ret = disable_axi(gpu);
229 	if (ret)
230 		return ret;
231 
232 	ret = disable_clk(gpu);
233 	if (ret)
234 		return ret;
235 
236 	ret = disable_pwrrail(gpu);
237 	if (ret)
238 		return ret;
239 
240 	return 0;
241 }
242 
243 int msm_gpu_hw_init(struct msm_gpu *gpu)
244 {
245 	int ret;
246 
247 	WARN_ON(!mutex_is_locked(&gpu->dev->struct_mutex));
248 
249 	if (!gpu->needs_hw_init)
250 		return 0;
251 
252 	disable_irq(gpu->irq);
253 	ret = gpu->funcs->hw_init(gpu);
254 	if (!ret)
255 		gpu->needs_hw_init = false;
256 	enable_irq(gpu->irq);
257 
258 	return ret;
259 }
260 
261 #ifdef CONFIG_DEV_COREDUMP
262 static ssize_t msm_gpu_devcoredump_read(char *buffer, loff_t offset,
263 		size_t count, void *data, size_t datalen)
264 {
265 	struct msm_gpu *gpu = data;
266 	struct drm_print_iterator iter;
267 	struct drm_printer p;
268 	struct msm_gpu_state *state;
269 
270 	state = msm_gpu_crashstate_get(gpu);
271 	if (!state)
272 		return 0;
273 
274 	iter.data = buffer;
275 	iter.offset = 0;
276 	iter.start = offset;
277 	iter.remain = count;
278 
279 	p = drm_coredump_printer(&iter);
280 
281 	drm_printf(&p, "---\n");
282 	drm_printf(&p, "kernel: " UTS_RELEASE "\n");
283 	drm_printf(&p, "module: " KBUILD_MODNAME "\n");
284 	drm_printf(&p, "time: %lld.%09ld\n",
285 		state->time.tv_sec, state->time.tv_nsec);
286 	if (state->comm)
287 		drm_printf(&p, "comm: %s\n", state->comm);
288 	if (state->cmd)
289 		drm_printf(&p, "cmdline: %s\n", state->cmd);
290 
291 	gpu->funcs->show(gpu, state, &p);
292 
293 	msm_gpu_crashstate_put(gpu);
294 
295 	return count - iter.remain;
296 }
297 
298 static void msm_gpu_devcoredump_free(void *data)
299 {
300 	struct msm_gpu *gpu = data;
301 
302 	msm_gpu_crashstate_put(gpu);
303 }
304 
305 static void msm_gpu_crashstate_get_bo(struct msm_gpu_state *state,
306 		struct msm_gem_object *obj, u64 iova, u32 flags)
307 {
308 	struct msm_gpu_state_bo *state_bo = &state->bos[state->nr_bos];
309 
310 	/* Don't record write only objects */
311 	state_bo->size = obj->base.size;
312 	state_bo->iova = iova;
313 
314 	/* Only store data for non imported buffer objects marked for read */
315 	if ((flags & MSM_SUBMIT_BO_READ) && !obj->base.import_attach) {
316 		void *ptr;
317 
318 		state_bo->data = kvmalloc(obj->base.size, GFP_KERNEL);
319 		if (!state_bo->data)
320 			goto out;
321 
322 		ptr = msm_gem_get_vaddr_active(&obj->base);
323 		if (IS_ERR(ptr)) {
324 			kvfree(state_bo->data);
325 			state_bo->data = NULL;
326 			goto out;
327 		}
328 
329 		memcpy(state_bo->data, ptr, obj->base.size);
330 		msm_gem_put_vaddr(&obj->base);
331 	}
332 out:
333 	state->nr_bos++;
334 }
335 
336 static void msm_gpu_crashstate_capture(struct msm_gpu *gpu,
337 		struct msm_gem_submit *submit, char *comm, char *cmd)
338 {
339 	struct msm_gpu_state *state;
340 
341 	/* Check if the target supports capturing crash state */
342 	if (!gpu->funcs->gpu_state_get)
343 		return;
344 
345 	/* Only save one crash state at a time */
346 	if (gpu->crashstate)
347 		return;
348 
349 	state = gpu->funcs->gpu_state_get(gpu);
350 	if (IS_ERR_OR_NULL(state))
351 		return;
352 
353 	/* Fill in the additional crash state information */
354 	state->comm = kstrdup(comm, GFP_KERNEL);
355 	state->cmd = kstrdup(cmd, GFP_KERNEL);
356 
357 	if (submit) {
358 		int i, nr = 0;
359 
360 		/* count # of buffers to dump: */
361 		for (i = 0; i < submit->nr_bos; i++)
362 			if (should_dump(submit, i))
363 				nr++;
364 		/* always dump cmd bo's, but don't double count them: */
365 		for (i = 0; i < submit->nr_cmds; i++)
366 			if (!should_dump(submit, submit->cmd[i].idx))
367 				nr++;
368 
369 		state->bos = kcalloc(nr,
370 			sizeof(struct msm_gpu_state_bo), GFP_KERNEL);
371 
372 		for (i = 0; i < submit->nr_bos; i++) {
373 			if (should_dump(submit, i)) {
374 				msm_gpu_crashstate_get_bo(state, submit->bos[i].obj,
375 					submit->bos[i].iova, submit->bos[i].flags);
376 			}
377 		}
378 
379 		for (i = 0; state->bos && i < submit->nr_cmds; i++) {
380 			int idx = submit->cmd[i].idx;
381 
382 			if (!should_dump(submit, submit->cmd[i].idx)) {
383 				msm_gpu_crashstate_get_bo(state, submit->bos[idx].obj,
384 					submit->bos[idx].iova, submit->bos[idx].flags);
385 			}
386 		}
387 	}
388 
389 	/* Set the active crash state to be dumped on failure */
390 	gpu->crashstate = state;
391 
392 	/* FIXME: Release the crashstate if this errors out? */
393 	dev_coredumpm(gpu->dev->dev, THIS_MODULE, gpu, 0, GFP_KERNEL,
394 		msm_gpu_devcoredump_read, msm_gpu_devcoredump_free);
395 }
396 #else
397 static void msm_gpu_crashstate_capture(struct msm_gpu *gpu,
398 		struct msm_gem_submit *submit, char *comm, char *cmd)
399 {
400 }
401 #endif
402 
403 /*
404  * Hangcheck detection for locked gpu:
405  */
406 
407 static void update_fences(struct msm_gpu *gpu, struct msm_ringbuffer *ring,
408 		uint32_t fence)
409 {
410 	struct msm_gem_submit *submit;
411 
412 	list_for_each_entry(submit, &ring->submits, node) {
413 		if (submit->seqno > fence)
414 			break;
415 
416 		msm_update_fence(submit->ring->fctx,
417 			submit->fence->seqno);
418 	}
419 }
420 
421 static struct msm_gem_submit *
422 find_submit(struct msm_ringbuffer *ring, uint32_t fence)
423 {
424 	struct msm_gem_submit *submit;
425 
426 	WARN_ON(!mutex_is_locked(&ring->gpu->dev->struct_mutex));
427 
428 	list_for_each_entry(submit, &ring->submits, node)
429 		if (submit->seqno == fence)
430 			return submit;
431 
432 	return NULL;
433 }
434 
435 static void retire_submits(struct msm_gpu *gpu);
436 
437 static void recover_worker(struct work_struct *work)
438 {
439 	struct msm_gpu *gpu = container_of(work, struct msm_gpu, recover_work);
440 	struct drm_device *dev = gpu->dev;
441 	struct msm_drm_private *priv = dev->dev_private;
442 	struct msm_gem_submit *submit;
443 	struct msm_ringbuffer *cur_ring = gpu->funcs->active_ring(gpu);
444 	char *comm = NULL, *cmd = NULL;
445 	int i;
446 
447 	mutex_lock(&dev->struct_mutex);
448 
449 	DRM_DEV_ERROR(dev->dev, "%s: hangcheck recover!\n", gpu->name);
450 
451 	submit = find_submit(cur_ring, cur_ring->memptrs->fence + 1);
452 	if (submit) {
453 		struct task_struct *task;
454 
455 		/* Increment the fault counts */
456 		gpu->global_faults++;
457 		submit->queue->faults++;
458 
459 		task = get_pid_task(submit->pid, PIDTYPE_PID);
460 		if (task) {
461 			comm = kstrdup(task->comm, GFP_KERNEL);
462 			cmd = kstrdup_quotable_cmdline(task, GFP_KERNEL);
463 			put_task_struct(task);
464 		}
465 
466 		if (comm && cmd) {
467 			DRM_DEV_ERROR(dev->dev, "%s: offending task: %s (%s)\n",
468 				gpu->name, comm, cmd);
469 
470 			msm_rd_dump_submit(priv->hangrd, submit,
471 				"offending task: %s (%s)", comm, cmd);
472 		} else
473 			msm_rd_dump_submit(priv->hangrd, submit, NULL);
474 	}
475 
476 	/* Record the crash state */
477 	pm_runtime_get_sync(&gpu->pdev->dev);
478 	msm_gpu_crashstate_capture(gpu, submit, comm, cmd);
479 	pm_runtime_put_sync(&gpu->pdev->dev);
480 
481 	kfree(cmd);
482 	kfree(comm);
483 
484 	/*
485 	 * Update all the rings with the latest and greatest fence.. this
486 	 * needs to happen after msm_rd_dump_submit() to ensure that the
487 	 * bo's referenced by the offending submit are still around.
488 	 */
489 	for (i = 0; i < gpu->nr_rings; i++) {
490 		struct msm_ringbuffer *ring = gpu->rb[i];
491 
492 		uint32_t fence = ring->memptrs->fence;
493 
494 		/*
495 		 * For the current (faulting?) ring/submit advance the fence by
496 		 * one more to clear the faulting submit
497 		 */
498 		if (ring == cur_ring)
499 			fence++;
500 
501 		update_fences(gpu, ring, fence);
502 	}
503 
504 	if (msm_gpu_active(gpu)) {
505 		/* retire completed submits, plus the one that hung: */
506 		retire_submits(gpu);
507 
508 		pm_runtime_get_sync(&gpu->pdev->dev);
509 		gpu->funcs->recover(gpu);
510 		pm_runtime_put_sync(&gpu->pdev->dev);
511 
512 		/*
513 		 * Replay all remaining submits starting with highest priority
514 		 * ring
515 		 */
516 		for (i = 0; i < gpu->nr_rings; i++) {
517 			struct msm_ringbuffer *ring = gpu->rb[i];
518 
519 			list_for_each_entry(submit, &ring->submits, node)
520 				gpu->funcs->submit(gpu, submit, NULL);
521 		}
522 	}
523 
524 	mutex_unlock(&dev->struct_mutex);
525 
526 	msm_gpu_retire(gpu);
527 }
528 
529 static void hangcheck_timer_reset(struct msm_gpu *gpu)
530 {
531 	DBG("%s", gpu->name);
532 	mod_timer(&gpu->hangcheck_timer,
533 			round_jiffies_up(jiffies + DRM_MSM_HANGCHECK_JIFFIES));
534 }
535 
536 static void hangcheck_handler(struct timer_list *t)
537 {
538 	struct msm_gpu *gpu = from_timer(gpu, t, hangcheck_timer);
539 	struct drm_device *dev = gpu->dev;
540 	struct msm_drm_private *priv = dev->dev_private;
541 	struct msm_ringbuffer *ring = gpu->funcs->active_ring(gpu);
542 	uint32_t fence = ring->memptrs->fence;
543 
544 	if (fence != ring->hangcheck_fence) {
545 		/* some progress has been made.. ya! */
546 		ring->hangcheck_fence = fence;
547 	} else if (fence < ring->seqno) {
548 		/* no progress and not done.. hung! */
549 		ring->hangcheck_fence = fence;
550 		DRM_DEV_ERROR(dev->dev, "%s: hangcheck detected gpu lockup rb %d!\n",
551 				gpu->name, ring->id);
552 		DRM_DEV_ERROR(dev->dev, "%s:     completed fence: %u\n",
553 				gpu->name, fence);
554 		DRM_DEV_ERROR(dev->dev, "%s:     submitted fence: %u\n",
555 				gpu->name, ring->seqno);
556 
557 		queue_work(priv->wq, &gpu->recover_work);
558 	}
559 
560 	/* if still more pending work, reset the hangcheck timer: */
561 	if (ring->seqno > ring->hangcheck_fence)
562 		hangcheck_timer_reset(gpu);
563 
564 	/* workaround for missing irq: */
565 	queue_work(priv->wq, &gpu->retire_work);
566 }
567 
568 /*
569  * Performance Counters:
570  */
571 
572 /* called under perf_lock */
573 static int update_hw_cntrs(struct msm_gpu *gpu, uint32_t ncntrs, uint32_t *cntrs)
574 {
575 	uint32_t current_cntrs[ARRAY_SIZE(gpu->last_cntrs)];
576 	int i, n = min(ncntrs, gpu->num_perfcntrs);
577 
578 	/* read current values: */
579 	for (i = 0; i < gpu->num_perfcntrs; i++)
580 		current_cntrs[i] = gpu_read(gpu, gpu->perfcntrs[i].sample_reg);
581 
582 	/* update cntrs: */
583 	for (i = 0; i < n; i++)
584 		cntrs[i] = current_cntrs[i] - gpu->last_cntrs[i];
585 
586 	/* save current values: */
587 	for (i = 0; i < gpu->num_perfcntrs; i++)
588 		gpu->last_cntrs[i] = current_cntrs[i];
589 
590 	return n;
591 }
592 
593 static void update_sw_cntrs(struct msm_gpu *gpu)
594 {
595 	ktime_t time;
596 	uint32_t elapsed;
597 	unsigned long flags;
598 
599 	spin_lock_irqsave(&gpu->perf_lock, flags);
600 	if (!gpu->perfcntr_active)
601 		goto out;
602 
603 	time = ktime_get();
604 	elapsed = ktime_to_us(ktime_sub(time, gpu->last_sample.time));
605 
606 	gpu->totaltime += elapsed;
607 	if (gpu->last_sample.active)
608 		gpu->activetime += elapsed;
609 
610 	gpu->last_sample.active = msm_gpu_active(gpu);
611 	gpu->last_sample.time = time;
612 
613 out:
614 	spin_unlock_irqrestore(&gpu->perf_lock, flags);
615 }
616 
617 void msm_gpu_perfcntr_start(struct msm_gpu *gpu)
618 {
619 	unsigned long flags;
620 
621 	pm_runtime_get_sync(&gpu->pdev->dev);
622 
623 	spin_lock_irqsave(&gpu->perf_lock, flags);
624 	/* we could dynamically enable/disable perfcntr registers too.. */
625 	gpu->last_sample.active = msm_gpu_active(gpu);
626 	gpu->last_sample.time = ktime_get();
627 	gpu->activetime = gpu->totaltime = 0;
628 	gpu->perfcntr_active = true;
629 	update_hw_cntrs(gpu, 0, NULL);
630 	spin_unlock_irqrestore(&gpu->perf_lock, flags);
631 }
632 
633 void msm_gpu_perfcntr_stop(struct msm_gpu *gpu)
634 {
635 	gpu->perfcntr_active = false;
636 	pm_runtime_put_sync(&gpu->pdev->dev);
637 }
638 
639 /* returns -errno or # of cntrs sampled */
640 int msm_gpu_perfcntr_sample(struct msm_gpu *gpu, uint32_t *activetime,
641 		uint32_t *totaltime, uint32_t ncntrs, uint32_t *cntrs)
642 {
643 	unsigned long flags;
644 	int ret;
645 
646 	spin_lock_irqsave(&gpu->perf_lock, flags);
647 
648 	if (!gpu->perfcntr_active) {
649 		ret = -EINVAL;
650 		goto out;
651 	}
652 
653 	*activetime = gpu->activetime;
654 	*totaltime = gpu->totaltime;
655 
656 	gpu->activetime = gpu->totaltime = 0;
657 
658 	ret = update_hw_cntrs(gpu, ncntrs, cntrs);
659 
660 out:
661 	spin_unlock_irqrestore(&gpu->perf_lock, flags);
662 
663 	return ret;
664 }
665 
666 /*
667  * Cmdstream submission/retirement:
668  */
669 
670 static void retire_submit(struct msm_gpu *gpu, struct msm_ringbuffer *ring,
671 		struct msm_gem_submit *submit)
672 {
673 	int index = submit->seqno % MSM_GPU_SUBMIT_STATS_COUNT;
674 	volatile struct msm_gpu_submit_stats *stats;
675 	u64 elapsed, clock = 0;
676 	int i;
677 
678 	stats = &ring->memptrs->stats[index];
679 	/* Convert 19.2Mhz alwayson ticks to nanoseconds for elapsed time */
680 	elapsed = (stats->alwayson_end - stats->alwayson_start) * 10000;
681 	do_div(elapsed, 192);
682 
683 	/* Calculate the clock frequency from the number of CP cycles */
684 	if (elapsed) {
685 		clock = (stats->cpcycles_end - stats->cpcycles_start) * 1000;
686 		do_div(clock, elapsed);
687 	}
688 
689 	trace_msm_gpu_submit_retired(submit, elapsed, clock,
690 		stats->alwayson_start, stats->alwayson_end);
691 
692 	for (i = 0; i < submit->nr_bos; i++) {
693 		struct msm_gem_object *msm_obj = submit->bos[i].obj;
694 		/* move to inactive: */
695 		msm_gem_move_to_inactive(&msm_obj->base);
696 		msm_gem_unpin_iova(&msm_obj->base, submit->aspace);
697 		drm_gem_object_put(&msm_obj->base);
698 	}
699 
700 	pm_runtime_mark_last_busy(&gpu->pdev->dev);
701 	pm_runtime_put_autosuspend(&gpu->pdev->dev);
702 	msm_gem_submit_free(submit);
703 }
704 
705 static void retire_submits(struct msm_gpu *gpu)
706 {
707 	struct drm_device *dev = gpu->dev;
708 	struct msm_gem_submit *submit, *tmp;
709 	int i;
710 
711 	WARN_ON(!mutex_is_locked(&dev->struct_mutex));
712 
713 	/* Retire the commits starting with highest priority */
714 	for (i = 0; i < gpu->nr_rings; i++) {
715 		struct msm_ringbuffer *ring = gpu->rb[i];
716 
717 		list_for_each_entry_safe(submit, tmp, &ring->submits, node) {
718 			if (dma_fence_is_signaled(submit->fence))
719 				retire_submit(gpu, ring, submit);
720 		}
721 	}
722 }
723 
724 static void retire_worker(struct work_struct *work)
725 {
726 	struct msm_gpu *gpu = container_of(work, struct msm_gpu, retire_work);
727 	struct drm_device *dev = gpu->dev;
728 	int i;
729 
730 	for (i = 0; i < gpu->nr_rings; i++)
731 		update_fences(gpu, gpu->rb[i], gpu->rb[i]->memptrs->fence);
732 
733 	mutex_lock(&dev->struct_mutex);
734 	retire_submits(gpu);
735 	mutex_unlock(&dev->struct_mutex);
736 }
737 
738 /* call from irq handler to schedule work to retire bo's */
739 void msm_gpu_retire(struct msm_gpu *gpu)
740 {
741 	struct msm_drm_private *priv = gpu->dev->dev_private;
742 	queue_work(priv->wq, &gpu->retire_work);
743 	update_sw_cntrs(gpu);
744 }
745 
746 /* add bo's to gpu's ring, and kick gpu: */
747 void msm_gpu_submit(struct msm_gpu *gpu, struct msm_gem_submit *submit,
748 		struct msm_file_private *ctx)
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 	int i;
754 
755 	WARN_ON(!mutex_is_locked(&dev->struct_mutex));
756 
757 	pm_runtime_get_sync(&gpu->pdev->dev);
758 
759 	msm_gpu_hw_init(gpu);
760 
761 	submit->seqno = ++ring->seqno;
762 
763 	list_add_tail(&submit->node, &ring->submits);
764 
765 	msm_rd_dump_submit(priv->rd, submit, NULL);
766 
767 	update_sw_cntrs(gpu);
768 
769 	for (i = 0; i < submit->nr_bos; i++) {
770 		struct msm_gem_object *msm_obj = submit->bos[i].obj;
771 		uint64_t iova;
772 
773 		/* can't happen yet.. but when we add 2d support we'll have
774 		 * to deal w/ cross-ring synchronization:
775 		 */
776 		WARN_ON(is_active(msm_obj) && (msm_obj->gpu != gpu));
777 
778 		/* submit takes a reference to the bo and iova until retired: */
779 		drm_gem_object_get(&msm_obj->base);
780 		msm_gem_get_and_pin_iova(&msm_obj->base, submit->aspace, &iova);
781 
782 		if (submit->bos[i].flags & MSM_SUBMIT_BO_WRITE)
783 			msm_gem_move_to_active(&msm_obj->base, gpu, true, submit->fence);
784 		else if (submit->bos[i].flags & MSM_SUBMIT_BO_READ)
785 			msm_gem_move_to_active(&msm_obj->base, gpu, false, submit->fence);
786 	}
787 
788 	gpu->funcs->submit(gpu, submit, ctx);
789 	priv->lastctx = ctx;
790 
791 	hangcheck_timer_reset(gpu);
792 }
793 
794 /*
795  * Init/Cleanup:
796  */
797 
798 static irqreturn_t irq_handler(int irq, void *data)
799 {
800 	struct msm_gpu *gpu = data;
801 	return gpu->funcs->irq(gpu);
802 }
803 
804 static int get_clocks(struct platform_device *pdev, struct msm_gpu *gpu)
805 {
806 	int ret = devm_clk_bulk_get_all(&pdev->dev, &gpu->grp_clks);
807 
808 	if (ret < 1) {
809 		gpu->nr_clocks = 0;
810 		return ret;
811 	}
812 
813 	gpu->nr_clocks = ret;
814 
815 	gpu->core_clk = msm_clk_bulk_get_clock(gpu->grp_clks,
816 		gpu->nr_clocks, "core");
817 
818 	gpu->rbbmtimer_clk = msm_clk_bulk_get_clock(gpu->grp_clks,
819 		gpu->nr_clocks, "rbbmtimer");
820 
821 	return 0;
822 }
823 
824 int msm_gpu_init(struct drm_device *drm, struct platform_device *pdev,
825 		struct msm_gpu *gpu, const struct msm_gpu_funcs *funcs,
826 		const char *name, struct msm_gpu_config *config)
827 {
828 	int i, ret, nr_rings = config->nr_rings;
829 	void *memptrs;
830 	uint64_t memptrs_iova;
831 
832 	if (WARN_ON(gpu->num_perfcntrs > ARRAY_SIZE(gpu->last_cntrs)))
833 		gpu->num_perfcntrs = ARRAY_SIZE(gpu->last_cntrs);
834 
835 	gpu->dev = drm;
836 	gpu->funcs = funcs;
837 	gpu->name = name;
838 
839 	INIT_LIST_HEAD(&gpu->active_list);
840 	INIT_WORK(&gpu->retire_work, retire_worker);
841 	INIT_WORK(&gpu->recover_work, recover_worker);
842 
843 
844 	timer_setup(&gpu->hangcheck_timer, hangcheck_handler, 0);
845 
846 	spin_lock_init(&gpu->perf_lock);
847 
848 
849 	/* Map registers: */
850 	gpu->mmio = msm_ioremap(pdev, config->ioname, name);
851 	if (IS_ERR(gpu->mmio)) {
852 		ret = PTR_ERR(gpu->mmio);
853 		goto fail;
854 	}
855 
856 	/* Get Interrupt: */
857 	gpu->irq = platform_get_irq(pdev, 0);
858 	if (gpu->irq < 0) {
859 		ret = gpu->irq;
860 		DRM_DEV_ERROR(drm->dev, "failed to get irq: %d\n", ret);
861 		goto fail;
862 	}
863 
864 	ret = devm_request_irq(&pdev->dev, gpu->irq, irq_handler,
865 			IRQF_TRIGGER_HIGH, gpu->name, gpu);
866 	if (ret) {
867 		DRM_DEV_ERROR(drm->dev, "failed to request IRQ%u: %d\n", gpu->irq, ret);
868 		goto fail;
869 	}
870 
871 	ret = get_clocks(pdev, gpu);
872 	if (ret)
873 		goto fail;
874 
875 	gpu->ebi1_clk = msm_clk_get(pdev, "bus");
876 	DBG("ebi1_clk: %p", gpu->ebi1_clk);
877 	if (IS_ERR(gpu->ebi1_clk))
878 		gpu->ebi1_clk = NULL;
879 
880 	/* Acquire regulators: */
881 	gpu->gpu_reg = devm_regulator_get(&pdev->dev, "vdd");
882 	DBG("gpu_reg: %p", gpu->gpu_reg);
883 	if (IS_ERR(gpu->gpu_reg))
884 		gpu->gpu_reg = NULL;
885 
886 	gpu->gpu_cx = devm_regulator_get(&pdev->dev, "vddcx");
887 	DBG("gpu_cx: %p", gpu->gpu_cx);
888 	if (IS_ERR(gpu->gpu_cx))
889 		gpu->gpu_cx = NULL;
890 
891 	gpu->pdev = pdev;
892 	platform_set_drvdata(pdev, gpu);
893 
894 	msm_devfreq_init(gpu);
895 
896 
897 	gpu->aspace = gpu->funcs->create_address_space(gpu, pdev);
898 
899 	if (gpu->aspace == NULL)
900 		DRM_DEV_INFO(drm->dev, "%s: no IOMMU, fallback to VRAM carveout!\n", name);
901 	else if (IS_ERR(gpu->aspace)) {
902 		ret = PTR_ERR(gpu->aspace);
903 		goto fail;
904 	}
905 
906 	memptrs = msm_gem_kernel_new(drm,
907 		sizeof(struct msm_rbmemptrs) * nr_rings,
908 		MSM_BO_UNCACHED, gpu->aspace, &gpu->memptrs_bo,
909 		&memptrs_iova);
910 
911 	if (IS_ERR(memptrs)) {
912 		ret = PTR_ERR(memptrs);
913 		DRM_DEV_ERROR(drm->dev, "could not allocate memptrs: %d\n", ret);
914 		goto fail;
915 	}
916 
917 	msm_gem_object_set_name(gpu->memptrs_bo, "memptrs");
918 
919 	if (nr_rings > ARRAY_SIZE(gpu->rb)) {
920 		DRM_DEV_INFO_ONCE(drm->dev, "Only creating %zu ringbuffers\n",
921 			ARRAY_SIZE(gpu->rb));
922 		nr_rings = ARRAY_SIZE(gpu->rb);
923 	}
924 
925 	/* Create ringbuffer(s): */
926 	for (i = 0; i < nr_rings; i++) {
927 		gpu->rb[i] = msm_ringbuffer_new(gpu, i, memptrs, memptrs_iova);
928 
929 		if (IS_ERR(gpu->rb[i])) {
930 			ret = PTR_ERR(gpu->rb[i]);
931 			DRM_DEV_ERROR(drm->dev,
932 				"could not create ringbuffer %d: %d\n", i, ret);
933 			goto fail;
934 		}
935 
936 		memptrs += sizeof(struct msm_rbmemptrs);
937 		memptrs_iova += sizeof(struct msm_rbmemptrs);
938 	}
939 
940 	gpu->nr_rings = nr_rings;
941 
942 	return 0;
943 
944 fail:
945 	for (i = 0; i < ARRAY_SIZE(gpu->rb); i++)  {
946 		msm_ringbuffer_destroy(gpu->rb[i]);
947 		gpu->rb[i] = NULL;
948 	}
949 
950 	msm_gem_kernel_put(gpu->memptrs_bo, gpu->aspace, false);
951 
952 	platform_set_drvdata(pdev, NULL);
953 	return ret;
954 }
955 
956 void msm_gpu_cleanup(struct msm_gpu *gpu)
957 {
958 	int i;
959 
960 	DBG("%s", gpu->name);
961 
962 	WARN_ON(!list_empty(&gpu->active_list));
963 
964 	for (i = 0; i < ARRAY_SIZE(gpu->rb); i++) {
965 		msm_ringbuffer_destroy(gpu->rb[i]);
966 		gpu->rb[i] = NULL;
967 	}
968 
969 	msm_gem_kernel_put(gpu->memptrs_bo, gpu->aspace, false);
970 
971 	if (!IS_ERR_OR_NULL(gpu->aspace)) {
972 		gpu->aspace->mmu->funcs->detach(gpu->aspace->mmu);
973 		msm_gem_address_space_put(gpu->aspace);
974 	}
975 }
976