xref: /openbmc/linux/drivers/gpu/drm/msm/msm_gpu.c (revision 11a163f2)
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/devfreq.h>
17 #include <linux/devcoredump.h>
18 #include <linux/sched/task.h>
19 
20 /*
21  * Power Management:
22  */
23 
24 static int msm_devfreq_target(struct device *dev, unsigned long *freq,
25 		u32 flags)
26 {
27 	struct msm_gpu *gpu = dev_to_gpu(dev);
28 	struct dev_pm_opp *opp;
29 
30 	opp = devfreq_recommended_opp(dev, freq, flags);
31 
32 	if (IS_ERR(opp))
33 		return PTR_ERR(opp);
34 
35 	trace_msm_gpu_freq_change(dev_pm_opp_get_freq(opp));
36 
37 	if (gpu->funcs->gpu_set_freq)
38 		gpu->funcs->gpu_set_freq(gpu, opp);
39 	else
40 		clk_set_rate(gpu->core_clk, *freq);
41 
42 	dev_pm_opp_put(opp);
43 
44 	return 0;
45 }
46 
47 static int msm_devfreq_get_dev_status(struct device *dev,
48 		struct devfreq_dev_status *status)
49 {
50 	struct msm_gpu *gpu = dev_to_gpu(dev);
51 	ktime_t time;
52 
53 	if (gpu->funcs->gpu_get_freq)
54 		status->current_frequency = gpu->funcs->gpu_get_freq(gpu);
55 	else
56 		status->current_frequency = clk_get_rate(gpu->core_clk);
57 
58 	status->busy_time = gpu->funcs->gpu_busy(gpu);
59 
60 	time = ktime_get();
61 	status->total_time = ktime_us_delta(time, gpu->devfreq.time);
62 	gpu->devfreq.time = time;
63 
64 	return 0;
65 }
66 
67 static int msm_devfreq_get_cur_freq(struct device *dev, unsigned long *freq)
68 {
69 	struct msm_gpu *gpu = dev_to_gpu(dev);
70 
71 	if (gpu->funcs->gpu_get_freq)
72 		*freq = gpu->funcs->gpu_get_freq(gpu);
73 	else
74 		*freq = clk_get_rate(gpu->core_clk);
75 
76 	return 0;
77 }
78 
79 static struct devfreq_dev_profile msm_devfreq_profile = {
80 	.polling_ms = 10,
81 	.target = msm_devfreq_target,
82 	.get_dev_status = msm_devfreq_get_dev_status,
83 	.get_cur_freq = msm_devfreq_get_cur_freq,
84 };
85 
86 static void msm_devfreq_init(struct msm_gpu *gpu)
87 {
88 	/* We need target support to do devfreq */
89 	if (!gpu->funcs->gpu_busy)
90 		return;
91 
92 	msm_devfreq_profile.initial_freq = gpu->fast_rate;
93 
94 	/*
95 	 * Don't set the freq_table or max_state and let devfreq build the table
96 	 * from OPP
97 	 * After a deferred probe, these may have be left to non-zero values,
98 	 * so set them back to zero before creating the devfreq device
99 	 */
100 	msm_devfreq_profile.freq_table = NULL;
101 	msm_devfreq_profile.max_state = 0;
102 
103 	gpu->devfreq.devfreq = devm_devfreq_add_device(&gpu->pdev->dev,
104 			&msm_devfreq_profile, DEVFREQ_GOV_SIMPLE_ONDEMAND,
105 			NULL);
106 
107 	if (IS_ERR(gpu->devfreq.devfreq)) {
108 		DRM_DEV_ERROR(&gpu->pdev->dev, "Couldn't initialize GPU devfreq\n");
109 		gpu->devfreq.devfreq = NULL;
110 	}
111 
112 	devfreq_suspend_device(gpu->devfreq.devfreq);
113 }
114 
115 static int enable_pwrrail(struct msm_gpu *gpu)
116 {
117 	struct drm_device *dev = gpu->dev;
118 	int ret = 0;
119 
120 	if (gpu->gpu_reg) {
121 		ret = regulator_enable(gpu->gpu_reg);
122 		if (ret) {
123 			DRM_DEV_ERROR(dev->dev, "failed to enable 'gpu_reg': %d\n", ret);
124 			return ret;
125 		}
126 	}
127 
128 	if (gpu->gpu_cx) {
129 		ret = regulator_enable(gpu->gpu_cx);
130 		if (ret) {
131 			DRM_DEV_ERROR(dev->dev, "failed to enable 'gpu_cx': %d\n", ret);
132 			return ret;
133 		}
134 	}
135 
136 	return 0;
137 }
138 
139 static int disable_pwrrail(struct msm_gpu *gpu)
140 {
141 	if (gpu->gpu_cx)
142 		regulator_disable(gpu->gpu_cx);
143 	if (gpu->gpu_reg)
144 		regulator_disable(gpu->gpu_reg);
145 	return 0;
146 }
147 
148 static int enable_clk(struct msm_gpu *gpu)
149 {
150 	if (gpu->core_clk && gpu->fast_rate)
151 		clk_set_rate(gpu->core_clk, gpu->fast_rate);
152 
153 	/* Set the RBBM timer rate to 19.2Mhz */
154 	if (gpu->rbbmtimer_clk)
155 		clk_set_rate(gpu->rbbmtimer_clk, 19200000);
156 
157 	return clk_bulk_prepare_enable(gpu->nr_clocks, gpu->grp_clks);
158 }
159 
160 static int disable_clk(struct msm_gpu *gpu)
161 {
162 	clk_bulk_disable_unprepare(gpu->nr_clocks, gpu->grp_clks);
163 
164 	/*
165 	 * Set the clock to a deliberately low rate. On older targets the clock
166 	 * speed had to be non zero to avoid problems. On newer targets this
167 	 * will be rounded down to zero anyway so it all works out.
168 	 */
169 	if (gpu->core_clk)
170 		clk_set_rate(gpu->core_clk, 27000000);
171 
172 	if (gpu->rbbmtimer_clk)
173 		clk_set_rate(gpu->rbbmtimer_clk, 0);
174 
175 	return 0;
176 }
177 
178 static int enable_axi(struct msm_gpu *gpu)
179 {
180 	if (gpu->ebi1_clk)
181 		clk_prepare_enable(gpu->ebi1_clk);
182 	return 0;
183 }
184 
185 static int disable_axi(struct msm_gpu *gpu)
186 {
187 	if (gpu->ebi1_clk)
188 		clk_disable_unprepare(gpu->ebi1_clk);
189 	return 0;
190 }
191 
192 void msm_gpu_resume_devfreq(struct msm_gpu *gpu)
193 {
194 	gpu->devfreq.busy_cycles = 0;
195 	gpu->devfreq.time = ktime_get();
196 
197 	devfreq_resume_device(gpu->devfreq.devfreq);
198 }
199 
200 int msm_gpu_pm_resume(struct msm_gpu *gpu)
201 {
202 	int ret;
203 
204 	DBG("%s", gpu->name);
205 	trace_msm_gpu_resume(0);
206 
207 	ret = enable_pwrrail(gpu);
208 	if (ret)
209 		return ret;
210 
211 	ret = enable_clk(gpu);
212 	if (ret)
213 		return ret;
214 
215 	ret = enable_axi(gpu);
216 	if (ret)
217 		return ret;
218 
219 	msm_gpu_resume_devfreq(gpu);
220 
221 	gpu->needs_hw_init = true;
222 
223 	return 0;
224 }
225 
226 int msm_gpu_pm_suspend(struct msm_gpu *gpu)
227 {
228 	int ret;
229 
230 	DBG("%s", gpu->name);
231 	trace_msm_gpu_suspend(0);
232 
233 	devfreq_suspend_device(gpu->devfreq.devfreq);
234 
235 	ret = disable_axi(gpu);
236 	if (ret)
237 		return ret;
238 
239 	ret = disable_clk(gpu);
240 	if (ret)
241 		return ret;
242 
243 	ret = disable_pwrrail(gpu);
244 	if (ret)
245 		return ret;
246 
247 	return 0;
248 }
249 
250 int msm_gpu_hw_init(struct msm_gpu *gpu)
251 {
252 	int ret;
253 
254 	WARN_ON(!mutex_is_locked(&gpu->dev->struct_mutex));
255 
256 	if (!gpu->needs_hw_init)
257 		return 0;
258 
259 	disable_irq(gpu->irq);
260 	ret = gpu->funcs->hw_init(gpu);
261 	if (!ret)
262 		gpu->needs_hw_init = false;
263 	enable_irq(gpu->irq);
264 
265 	return ret;
266 }
267 
268 #ifdef CONFIG_DEV_COREDUMP
269 static ssize_t msm_gpu_devcoredump_read(char *buffer, loff_t offset,
270 		size_t count, void *data, size_t datalen)
271 {
272 	struct msm_gpu *gpu = data;
273 	struct drm_print_iterator iter;
274 	struct drm_printer p;
275 	struct msm_gpu_state *state;
276 
277 	state = msm_gpu_crashstate_get(gpu);
278 	if (!state)
279 		return 0;
280 
281 	iter.data = buffer;
282 	iter.offset = 0;
283 	iter.start = offset;
284 	iter.remain = count;
285 
286 	p = drm_coredump_printer(&iter);
287 
288 	drm_printf(&p, "---\n");
289 	drm_printf(&p, "kernel: " UTS_RELEASE "\n");
290 	drm_printf(&p, "module: " KBUILD_MODNAME "\n");
291 	drm_printf(&p, "time: %lld.%09ld\n",
292 		state->time.tv_sec, state->time.tv_nsec);
293 	if (state->comm)
294 		drm_printf(&p, "comm: %s\n", state->comm);
295 	if (state->cmd)
296 		drm_printf(&p, "cmdline: %s\n", state->cmd);
297 
298 	gpu->funcs->show(gpu, state, &p);
299 
300 	msm_gpu_crashstate_put(gpu);
301 
302 	return count - iter.remain;
303 }
304 
305 static void msm_gpu_devcoredump_free(void *data)
306 {
307 	struct msm_gpu *gpu = data;
308 
309 	msm_gpu_crashstate_put(gpu);
310 }
311 
312 static void msm_gpu_crashstate_get_bo(struct msm_gpu_state *state,
313 		struct msm_gem_object *obj, u64 iova, u32 flags)
314 {
315 	struct msm_gpu_state_bo *state_bo = &state->bos[state->nr_bos];
316 
317 	/* Don't record write only objects */
318 	state_bo->size = obj->base.size;
319 	state_bo->iova = iova;
320 
321 	/* Only store data for non imported buffer objects marked for read */
322 	if ((flags & MSM_SUBMIT_BO_READ) && !obj->base.import_attach) {
323 		void *ptr;
324 
325 		state_bo->data = kvmalloc(obj->base.size, GFP_KERNEL);
326 		if (!state_bo->data)
327 			goto out;
328 
329 		ptr = msm_gem_get_vaddr_active(&obj->base);
330 		if (IS_ERR(ptr)) {
331 			kvfree(state_bo->data);
332 			state_bo->data = NULL;
333 			goto out;
334 		}
335 
336 		memcpy(state_bo->data, ptr, obj->base.size);
337 		msm_gem_put_vaddr(&obj->base);
338 	}
339 out:
340 	state->nr_bos++;
341 }
342 
343 static void msm_gpu_crashstate_capture(struct msm_gpu *gpu,
344 		struct msm_gem_submit *submit, char *comm, char *cmd)
345 {
346 	struct msm_gpu_state *state;
347 
348 	/* Check if the target supports capturing crash state */
349 	if (!gpu->funcs->gpu_state_get)
350 		return;
351 
352 	/* Only save one crash state at a time */
353 	if (gpu->crashstate)
354 		return;
355 
356 	state = gpu->funcs->gpu_state_get(gpu);
357 	if (IS_ERR_OR_NULL(state))
358 		return;
359 
360 	/* Fill in the additional crash state information */
361 	state->comm = kstrdup(comm, GFP_KERNEL);
362 	state->cmd = kstrdup(cmd, GFP_KERNEL);
363 
364 	if (submit) {
365 		int i, nr = 0;
366 
367 		/* count # of buffers to dump: */
368 		for (i = 0; i < submit->nr_bos; i++)
369 			if (should_dump(submit, i))
370 				nr++;
371 		/* always dump cmd bo's, but don't double count them: */
372 		for (i = 0; i < submit->nr_cmds; i++)
373 			if (!should_dump(submit, submit->cmd[i].idx))
374 				nr++;
375 
376 		state->bos = kcalloc(nr,
377 			sizeof(struct msm_gpu_state_bo), GFP_KERNEL);
378 
379 		for (i = 0; i < submit->nr_bos; i++) {
380 			if (should_dump(submit, i)) {
381 				msm_gpu_crashstate_get_bo(state, submit->bos[i].obj,
382 					submit->bos[i].iova, submit->bos[i].flags);
383 			}
384 		}
385 
386 		for (i = 0; state->bos && i < submit->nr_cmds; i++) {
387 			int idx = submit->cmd[i].idx;
388 
389 			if (!should_dump(submit, submit->cmd[i].idx)) {
390 				msm_gpu_crashstate_get_bo(state, submit->bos[idx].obj,
391 					submit->bos[idx].iova, submit->bos[idx].flags);
392 			}
393 		}
394 	}
395 
396 	/* Set the active crash state to be dumped on failure */
397 	gpu->crashstate = state;
398 
399 	/* FIXME: Release the crashstate if this errors out? */
400 	dev_coredumpm(gpu->dev->dev, THIS_MODULE, gpu, 0, GFP_KERNEL,
401 		msm_gpu_devcoredump_read, msm_gpu_devcoredump_free);
402 }
403 #else
404 static void msm_gpu_crashstate_capture(struct msm_gpu *gpu,
405 		struct msm_gem_submit *submit, char *comm, char *cmd)
406 {
407 }
408 #endif
409 
410 /*
411  * Hangcheck detection for locked gpu:
412  */
413 
414 static void update_fences(struct msm_gpu *gpu, struct msm_ringbuffer *ring,
415 		uint32_t fence)
416 {
417 	struct msm_gem_submit *submit;
418 
419 	list_for_each_entry(submit, &ring->submits, node) {
420 		if (submit->seqno > fence)
421 			break;
422 
423 		msm_update_fence(submit->ring->fctx,
424 			submit->fence->seqno);
425 	}
426 }
427 
428 static struct msm_gem_submit *
429 find_submit(struct msm_ringbuffer *ring, uint32_t fence)
430 {
431 	struct msm_gem_submit *submit;
432 
433 	WARN_ON(!mutex_is_locked(&ring->gpu->dev->struct_mutex));
434 
435 	list_for_each_entry(submit, &ring->submits, node)
436 		if (submit->seqno == fence)
437 			return submit;
438 
439 	return NULL;
440 }
441 
442 static void retire_submits(struct msm_gpu *gpu);
443 
444 static void recover_worker(struct work_struct *work)
445 {
446 	struct msm_gpu *gpu = container_of(work, struct msm_gpu, recover_work);
447 	struct drm_device *dev = gpu->dev;
448 	struct msm_drm_private *priv = dev->dev_private;
449 	struct msm_gem_submit *submit;
450 	struct msm_ringbuffer *cur_ring = gpu->funcs->active_ring(gpu);
451 	char *comm = NULL, *cmd = NULL;
452 	int i;
453 
454 	mutex_lock(&dev->struct_mutex);
455 
456 	DRM_DEV_ERROR(dev->dev, "%s: hangcheck recover!\n", gpu->name);
457 
458 	submit = find_submit(cur_ring, cur_ring->memptrs->fence + 1);
459 	if (submit) {
460 		struct task_struct *task;
461 
462 		/* Increment the fault counts */
463 		gpu->global_faults++;
464 		submit->queue->faults++;
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 		if (comm && cmd) {
474 			DRM_DEV_ERROR(dev->dev, "%s: offending task: %s (%s)\n",
475 				gpu->name, comm, cmd);
476 
477 			msm_rd_dump_submit(priv->hangrd, submit,
478 				"offending task: %s (%s)", comm, cmd);
479 		} else
480 			msm_rd_dump_submit(priv->hangrd, submit, NULL);
481 	}
482 
483 	/* Record the crash state */
484 	pm_runtime_get_sync(&gpu->pdev->dev);
485 	msm_gpu_crashstate_capture(gpu, submit, comm, cmd);
486 	pm_runtime_put_sync(&gpu->pdev->dev);
487 
488 	kfree(cmd);
489 	kfree(comm);
490 
491 	/*
492 	 * Update all the rings with the latest and greatest fence.. this
493 	 * needs to happen after msm_rd_dump_submit() to ensure that the
494 	 * bo's referenced by the offending submit are still around.
495 	 */
496 	for (i = 0; i < gpu->nr_rings; i++) {
497 		struct msm_ringbuffer *ring = gpu->rb[i];
498 
499 		uint32_t fence = ring->memptrs->fence;
500 
501 		/*
502 		 * For the current (faulting?) ring/submit advance the fence by
503 		 * one more to clear the faulting submit
504 		 */
505 		if (ring == cur_ring)
506 			fence++;
507 
508 		update_fences(gpu, ring, fence);
509 	}
510 
511 	if (msm_gpu_active(gpu)) {
512 		/* retire completed submits, plus the one that hung: */
513 		retire_submits(gpu);
514 
515 		pm_runtime_get_sync(&gpu->pdev->dev);
516 		gpu->funcs->recover(gpu);
517 		pm_runtime_put_sync(&gpu->pdev->dev);
518 
519 		/*
520 		 * Replay all remaining submits starting with highest priority
521 		 * ring
522 		 */
523 		for (i = 0; i < gpu->nr_rings; i++) {
524 			struct msm_ringbuffer *ring = gpu->rb[i];
525 
526 			list_for_each_entry(submit, &ring->submits, node)
527 				gpu->funcs->submit(gpu, submit);
528 		}
529 	}
530 
531 	mutex_unlock(&dev->struct_mutex);
532 
533 	msm_gpu_retire(gpu);
534 }
535 
536 static void hangcheck_timer_reset(struct msm_gpu *gpu)
537 {
538 	DBG("%s", gpu->name);
539 	mod_timer(&gpu->hangcheck_timer,
540 			round_jiffies_up(jiffies + DRM_MSM_HANGCHECK_JIFFIES));
541 }
542 
543 static void hangcheck_handler(struct timer_list *t)
544 {
545 	struct msm_gpu *gpu = from_timer(gpu, t, hangcheck_timer);
546 	struct drm_device *dev = gpu->dev;
547 	struct msm_drm_private *priv = dev->dev_private;
548 	struct msm_ringbuffer *ring = gpu->funcs->active_ring(gpu);
549 	uint32_t fence = ring->memptrs->fence;
550 
551 	if (fence != ring->hangcheck_fence) {
552 		/* some progress has been made.. ya! */
553 		ring->hangcheck_fence = fence;
554 	} else if (fence < ring->seqno) {
555 		/* no progress and not done.. hung! */
556 		ring->hangcheck_fence = fence;
557 		DRM_DEV_ERROR(dev->dev, "%s: hangcheck detected gpu lockup rb %d!\n",
558 				gpu->name, ring->id);
559 		DRM_DEV_ERROR(dev->dev, "%s:     completed fence: %u\n",
560 				gpu->name, fence);
561 		DRM_DEV_ERROR(dev->dev, "%s:     submitted fence: %u\n",
562 				gpu->name, ring->seqno);
563 
564 		queue_work(priv->wq, &gpu->recover_work);
565 	}
566 
567 	/* if still more pending work, reset the hangcheck timer: */
568 	if (ring->seqno > ring->hangcheck_fence)
569 		hangcheck_timer_reset(gpu);
570 
571 	/* workaround for missing irq: */
572 	queue_work(priv->wq, &gpu->retire_work);
573 }
574 
575 /*
576  * Performance Counters:
577  */
578 
579 /* called under perf_lock */
580 static int update_hw_cntrs(struct msm_gpu *gpu, uint32_t ncntrs, uint32_t *cntrs)
581 {
582 	uint32_t current_cntrs[ARRAY_SIZE(gpu->last_cntrs)];
583 	int i, n = min(ncntrs, gpu->num_perfcntrs);
584 
585 	/* read current values: */
586 	for (i = 0; i < gpu->num_perfcntrs; i++)
587 		current_cntrs[i] = gpu_read(gpu, gpu->perfcntrs[i].sample_reg);
588 
589 	/* update cntrs: */
590 	for (i = 0; i < n; i++)
591 		cntrs[i] = current_cntrs[i] - gpu->last_cntrs[i];
592 
593 	/* save current values: */
594 	for (i = 0; i < gpu->num_perfcntrs; i++)
595 		gpu->last_cntrs[i] = current_cntrs[i];
596 
597 	return n;
598 }
599 
600 static void update_sw_cntrs(struct msm_gpu *gpu)
601 {
602 	ktime_t time;
603 	uint32_t elapsed;
604 	unsigned long flags;
605 
606 	spin_lock_irqsave(&gpu->perf_lock, flags);
607 	if (!gpu->perfcntr_active)
608 		goto out;
609 
610 	time = ktime_get();
611 	elapsed = ktime_to_us(ktime_sub(time, gpu->last_sample.time));
612 
613 	gpu->totaltime += elapsed;
614 	if (gpu->last_sample.active)
615 		gpu->activetime += elapsed;
616 
617 	gpu->last_sample.active = msm_gpu_active(gpu);
618 	gpu->last_sample.time = time;
619 
620 out:
621 	spin_unlock_irqrestore(&gpu->perf_lock, flags);
622 }
623 
624 void msm_gpu_perfcntr_start(struct msm_gpu *gpu)
625 {
626 	unsigned long flags;
627 
628 	pm_runtime_get_sync(&gpu->pdev->dev);
629 
630 	spin_lock_irqsave(&gpu->perf_lock, flags);
631 	/* we could dynamically enable/disable perfcntr registers too.. */
632 	gpu->last_sample.active = msm_gpu_active(gpu);
633 	gpu->last_sample.time = ktime_get();
634 	gpu->activetime = gpu->totaltime = 0;
635 	gpu->perfcntr_active = true;
636 	update_hw_cntrs(gpu, 0, NULL);
637 	spin_unlock_irqrestore(&gpu->perf_lock, flags);
638 }
639 
640 void msm_gpu_perfcntr_stop(struct msm_gpu *gpu)
641 {
642 	gpu->perfcntr_active = false;
643 	pm_runtime_put_sync(&gpu->pdev->dev);
644 }
645 
646 /* returns -errno or # of cntrs sampled */
647 int msm_gpu_perfcntr_sample(struct msm_gpu *gpu, uint32_t *activetime,
648 		uint32_t *totaltime, uint32_t ncntrs, uint32_t *cntrs)
649 {
650 	unsigned long flags;
651 	int ret;
652 
653 	spin_lock_irqsave(&gpu->perf_lock, flags);
654 
655 	if (!gpu->perfcntr_active) {
656 		ret = -EINVAL;
657 		goto out;
658 	}
659 
660 	*activetime = gpu->activetime;
661 	*totaltime = gpu->totaltime;
662 
663 	gpu->activetime = gpu->totaltime = 0;
664 
665 	ret = update_hw_cntrs(gpu, ncntrs, cntrs);
666 
667 out:
668 	spin_unlock_irqrestore(&gpu->perf_lock, flags);
669 
670 	return ret;
671 }
672 
673 /*
674  * Cmdstream submission/retirement:
675  */
676 
677 static void retire_submit(struct msm_gpu *gpu, struct msm_ringbuffer *ring,
678 		struct msm_gem_submit *submit)
679 {
680 	int index = submit->seqno % MSM_GPU_SUBMIT_STATS_COUNT;
681 	volatile struct msm_gpu_submit_stats *stats;
682 	u64 elapsed, clock = 0;
683 	int i;
684 
685 	stats = &ring->memptrs->stats[index];
686 	/* Convert 19.2Mhz alwayson ticks to nanoseconds for elapsed time */
687 	elapsed = (stats->alwayson_end - stats->alwayson_start) * 10000;
688 	do_div(elapsed, 192);
689 
690 	/* Calculate the clock frequency from the number of CP cycles */
691 	if (elapsed) {
692 		clock = (stats->cpcycles_end - stats->cpcycles_start) * 1000;
693 		do_div(clock, elapsed);
694 	}
695 
696 	trace_msm_gpu_submit_retired(submit, elapsed, clock,
697 		stats->alwayson_start, stats->alwayson_end);
698 
699 	for (i = 0; i < submit->nr_bos; i++) {
700 		struct msm_gem_object *msm_obj = submit->bos[i].obj;
701 
702 		msm_gem_active_put(&msm_obj->base);
703 		msm_gem_unpin_iova(&msm_obj->base, submit->aspace);
704 		drm_gem_object_put_locked(&msm_obj->base);
705 	}
706 
707 	pm_runtime_mark_last_busy(&gpu->pdev->dev);
708 	pm_runtime_put_autosuspend(&gpu->pdev->dev);
709 	msm_gem_submit_free(submit);
710 }
711 
712 static void retire_submits(struct msm_gpu *gpu)
713 {
714 	struct drm_device *dev = gpu->dev;
715 	struct msm_gem_submit *submit, *tmp;
716 	int i;
717 
718 	WARN_ON(!mutex_is_locked(&dev->struct_mutex));
719 
720 	/* Retire the commits starting with highest priority */
721 	for (i = 0; i < gpu->nr_rings; i++) {
722 		struct msm_ringbuffer *ring = gpu->rb[i];
723 
724 		list_for_each_entry_safe(submit, tmp, &ring->submits, node) {
725 			if (dma_fence_is_signaled(submit->fence))
726 				retire_submit(gpu, ring, submit);
727 		}
728 	}
729 }
730 
731 static void retire_worker(struct work_struct *work)
732 {
733 	struct msm_gpu *gpu = container_of(work, struct msm_gpu, retire_work);
734 	struct drm_device *dev = gpu->dev;
735 	int i;
736 
737 	for (i = 0; i < gpu->nr_rings; i++)
738 		update_fences(gpu, gpu->rb[i], gpu->rb[i]->memptrs->fence);
739 
740 	mutex_lock(&dev->struct_mutex);
741 	retire_submits(gpu);
742 	mutex_unlock(&dev->struct_mutex);
743 }
744 
745 /* call from irq handler to schedule work to retire bo's */
746 void msm_gpu_retire(struct msm_gpu *gpu)
747 {
748 	struct msm_drm_private *priv = gpu->dev->dev_private;
749 	queue_work(priv->wq, &gpu->retire_work);
750 	update_sw_cntrs(gpu);
751 }
752 
753 /* add bo's to gpu's ring, and kick gpu: */
754 void msm_gpu_submit(struct msm_gpu *gpu, struct msm_gem_submit *submit)
755 {
756 	struct drm_device *dev = gpu->dev;
757 	struct msm_drm_private *priv = dev->dev_private;
758 	struct msm_ringbuffer *ring = submit->ring;
759 	int i;
760 
761 	WARN_ON(!mutex_is_locked(&dev->struct_mutex));
762 
763 	pm_runtime_get_sync(&gpu->pdev->dev);
764 
765 	msm_gpu_hw_init(gpu);
766 
767 	submit->seqno = ++ring->seqno;
768 
769 	list_add_tail(&submit->node, &ring->submits);
770 
771 	msm_rd_dump_submit(priv->rd, submit, NULL);
772 
773 	update_sw_cntrs(gpu);
774 
775 	for (i = 0; i < submit->nr_bos; i++) {
776 		struct msm_gem_object *msm_obj = submit->bos[i].obj;
777 		struct drm_gem_object *drm_obj = &msm_obj->base;
778 		uint64_t iova;
779 
780 		/* can't happen yet.. but when we add 2d support we'll have
781 		 * to deal w/ cross-ring synchronization:
782 		 */
783 		WARN_ON(is_active(msm_obj) && (msm_obj->gpu != gpu));
784 
785 		/* submit takes a reference to the bo and iova until retired: */
786 		drm_gem_object_get(&msm_obj->base);
787 		msm_gem_get_and_pin_iova(&msm_obj->base, submit->aspace, &iova);
788 
789 		if (submit->bos[i].flags & MSM_SUBMIT_BO_WRITE)
790 			dma_resv_add_excl_fence(drm_obj->resv, submit->fence);
791 		else if (submit->bos[i].flags & MSM_SUBMIT_BO_READ)
792 			dma_resv_add_shared_fence(drm_obj->resv, submit->fence);
793 
794 		msm_gem_active_get(drm_obj, gpu);
795 	}
796 
797 	gpu->funcs->submit(gpu, submit);
798 	priv->lastctx = submit->queue->ctx;
799 
800 	hangcheck_timer_reset(gpu);
801 }
802 
803 /*
804  * Init/Cleanup:
805  */
806 
807 static irqreturn_t irq_handler(int irq, void *data)
808 {
809 	struct msm_gpu *gpu = data;
810 	return gpu->funcs->irq(gpu);
811 }
812 
813 static int get_clocks(struct platform_device *pdev, struct msm_gpu *gpu)
814 {
815 	int ret = devm_clk_bulk_get_all(&pdev->dev, &gpu->grp_clks);
816 
817 	if (ret < 1) {
818 		gpu->nr_clocks = 0;
819 		return ret;
820 	}
821 
822 	gpu->nr_clocks = ret;
823 
824 	gpu->core_clk = msm_clk_bulk_get_clock(gpu->grp_clks,
825 		gpu->nr_clocks, "core");
826 
827 	gpu->rbbmtimer_clk = msm_clk_bulk_get_clock(gpu->grp_clks,
828 		gpu->nr_clocks, "rbbmtimer");
829 
830 	return 0;
831 }
832 
833 /* Return a new address space for a msm_drm_private instance */
834 struct msm_gem_address_space *
835 msm_gpu_create_private_address_space(struct msm_gpu *gpu, struct task_struct *task)
836 {
837 	struct msm_gem_address_space *aspace = NULL;
838 	if (!gpu)
839 		return NULL;
840 
841 	/*
842 	 * If the target doesn't support private address spaces then return
843 	 * the global one
844 	 */
845 	if (gpu->funcs->create_private_address_space) {
846 		aspace = gpu->funcs->create_private_address_space(gpu);
847 		if (!IS_ERR(aspace))
848 			aspace->pid = get_pid(task_pid(task));
849 	}
850 
851 	if (IS_ERR_OR_NULL(aspace))
852 		aspace = msm_gem_address_space_get(gpu->aspace);
853 
854 	return aspace;
855 }
856 
857 int msm_gpu_init(struct drm_device *drm, struct platform_device *pdev,
858 		struct msm_gpu *gpu, const struct msm_gpu_funcs *funcs,
859 		const char *name, struct msm_gpu_config *config)
860 {
861 	int i, ret, nr_rings = config->nr_rings;
862 	void *memptrs;
863 	uint64_t memptrs_iova;
864 
865 	if (WARN_ON(gpu->num_perfcntrs > ARRAY_SIZE(gpu->last_cntrs)))
866 		gpu->num_perfcntrs = ARRAY_SIZE(gpu->last_cntrs);
867 
868 	gpu->dev = drm;
869 	gpu->funcs = funcs;
870 	gpu->name = name;
871 
872 	INIT_LIST_HEAD(&gpu->active_list);
873 	INIT_WORK(&gpu->retire_work, retire_worker);
874 	INIT_WORK(&gpu->recover_work, recover_worker);
875 
876 
877 	timer_setup(&gpu->hangcheck_timer, hangcheck_handler, 0);
878 
879 	spin_lock_init(&gpu->perf_lock);
880 
881 
882 	/* Map registers: */
883 	gpu->mmio = msm_ioremap(pdev, config->ioname, name);
884 	if (IS_ERR(gpu->mmio)) {
885 		ret = PTR_ERR(gpu->mmio);
886 		goto fail;
887 	}
888 
889 	/* Get Interrupt: */
890 	gpu->irq = platform_get_irq(pdev, 0);
891 	if (gpu->irq < 0) {
892 		ret = gpu->irq;
893 		DRM_DEV_ERROR(drm->dev, "failed to get irq: %d\n", ret);
894 		goto fail;
895 	}
896 
897 	ret = devm_request_irq(&pdev->dev, gpu->irq, irq_handler,
898 			IRQF_TRIGGER_HIGH, gpu->name, gpu);
899 	if (ret) {
900 		DRM_DEV_ERROR(drm->dev, "failed to request IRQ%u: %d\n", gpu->irq, ret);
901 		goto fail;
902 	}
903 
904 	ret = get_clocks(pdev, gpu);
905 	if (ret)
906 		goto fail;
907 
908 	gpu->ebi1_clk = msm_clk_get(pdev, "bus");
909 	DBG("ebi1_clk: %p", gpu->ebi1_clk);
910 	if (IS_ERR(gpu->ebi1_clk))
911 		gpu->ebi1_clk = NULL;
912 
913 	/* Acquire regulators: */
914 	gpu->gpu_reg = devm_regulator_get(&pdev->dev, "vdd");
915 	DBG("gpu_reg: %p", gpu->gpu_reg);
916 	if (IS_ERR(gpu->gpu_reg))
917 		gpu->gpu_reg = NULL;
918 
919 	gpu->gpu_cx = devm_regulator_get(&pdev->dev, "vddcx");
920 	DBG("gpu_cx: %p", gpu->gpu_cx);
921 	if (IS_ERR(gpu->gpu_cx))
922 		gpu->gpu_cx = NULL;
923 
924 	gpu->pdev = pdev;
925 	platform_set_drvdata(pdev, &gpu->adreno_smmu);
926 
927 	msm_devfreq_init(gpu);
928 
929 
930 	gpu->aspace = gpu->funcs->create_address_space(gpu, pdev);
931 
932 	if (gpu->aspace == NULL)
933 		DRM_DEV_INFO(drm->dev, "%s: no IOMMU, fallback to VRAM carveout!\n", name);
934 	else if (IS_ERR(gpu->aspace)) {
935 		ret = PTR_ERR(gpu->aspace);
936 		goto fail;
937 	}
938 
939 	memptrs = msm_gem_kernel_new(drm,
940 		sizeof(struct msm_rbmemptrs) * nr_rings,
941 		check_apriv(gpu, MSM_BO_UNCACHED), gpu->aspace, &gpu->memptrs_bo,
942 		&memptrs_iova);
943 
944 	if (IS_ERR(memptrs)) {
945 		ret = PTR_ERR(memptrs);
946 		DRM_DEV_ERROR(drm->dev, "could not allocate memptrs: %d\n", ret);
947 		goto fail;
948 	}
949 
950 	msm_gem_object_set_name(gpu->memptrs_bo, "memptrs");
951 
952 	if (nr_rings > ARRAY_SIZE(gpu->rb)) {
953 		DRM_DEV_INFO_ONCE(drm->dev, "Only creating %zu ringbuffers\n",
954 			ARRAY_SIZE(gpu->rb));
955 		nr_rings = ARRAY_SIZE(gpu->rb);
956 	}
957 
958 	/* Create ringbuffer(s): */
959 	for (i = 0; i < nr_rings; i++) {
960 		gpu->rb[i] = msm_ringbuffer_new(gpu, i, memptrs, memptrs_iova);
961 
962 		if (IS_ERR(gpu->rb[i])) {
963 			ret = PTR_ERR(gpu->rb[i]);
964 			DRM_DEV_ERROR(drm->dev,
965 				"could not create ringbuffer %d: %d\n", i, ret);
966 			goto fail;
967 		}
968 
969 		memptrs += sizeof(struct msm_rbmemptrs);
970 		memptrs_iova += sizeof(struct msm_rbmemptrs);
971 	}
972 
973 	gpu->nr_rings = nr_rings;
974 
975 	return 0;
976 
977 fail:
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, false);
984 
985 	platform_set_drvdata(pdev, NULL);
986 	return ret;
987 }
988 
989 void msm_gpu_cleanup(struct msm_gpu *gpu)
990 {
991 	int i;
992 
993 	DBG("%s", gpu->name);
994 
995 	WARN_ON(!list_empty(&gpu->active_list));
996 
997 	for (i = 0; i < ARRAY_SIZE(gpu->rb); i++) {
998 		msm_ringbuffer_destroy(gpu->rb[i]);
999 		gpu->rb[i] = NULL;
1000 	}
1001 
1002 	msm_gem_kernel_put(gpu->memptrs_bo, gpu->aspace, false);
1003 
1004 	if (!IS_ERR_OR_NULL(gpu->aspace)) {
1005 		gpu->aspace->mmu->funcs->detach(gpu->aspace->mmu);
1006 		msm_gem_address_space_put(gpu->aspace);
1007 	}
1008 }
1009