xref: /openbmc/linux/drivers/gpu/drm/msm/msm_gpu.c (revision dbf563ee)
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 = platform_get_drvdata(to_platform_device(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 	if (gpu->funcs->gpu_set_freq)
36 		gpu->funcs->gpu_set_freq(gpu, opp);
37 	else
38 		clk_set_rate(gpu->core_clk, *freq);
39 
40 	dev_pm_opp_put(opp);
41 
42 	return 0;
43 }
44 
45 static int msm_devfreq_get_dev_status(struct device *dev,
46 		struct devfreq_dev_status *status)
47 {
48 	struct msm_gpu *gpu = platform_get_drvdata(to_platform_device(dev));
49 	ktime_t time;
50 
51 	if (gpu->funcs->gpu_get_freq)
52 		status->current_frequency = gpu->funcs->gpu_get_freq(gpu);
53 	else
54 		status->current_frequency = clk_get_rate(gpu->core_clk);
55 
56 	status->busy_time = gpu->funcs->gpu_busy(gpu);
57 
58 	time = ktime_get();
59 	status->total_time = ktime_us_delta(time, gpu->devfreq.time);
60 	gpu->devfreq.time = time;
61 
62 	return 0;
63 }
64 
65 static int msm_devfreq_get_cur_freq(struct device *dev, unsigned long *freq)
66 {
67 	struct msm_gpu *gpu = platform_get_drvdata(to_platform_device(dev));
68 
69 	if (gpu->funcs->gpu_get_freq)
70 		*freq = gpu->funcs->gpu_get_freq(gpu);
71 	else
72 		*freq = clk_get_rate(gpu->core_clk);
73 
74 	return 0;
75 }
76 
77 static struct devfreq_dev_profile msm_devfreq_profile = {
78 	.polling_ms = 10,
79 	.target = msm_devfreq_target,
80 	.get_dev_status = msm_devfreq_get_dev_status,
81 	.get_cur_freq = msm_devfreq_get_cur_freq,
82 };
83 
84 static void msm_devfreq_init(struct msm_gpu *gpu)
85 {
86 	/* We need target support to do devfreq */
87 	if (!gpu->funcs->gpu_busy)
88 		return;
89 
90 	msm_devfreq_profile.initial_freq = gpu->fast_rate;
91 
92 	/*
93 	 * Don't set the freq_table or max_state and let devfreq build the table
94 	 * from OPP
95 	 * After a deferred probe, these may have be left to non-zero values,
96 	 * so set them back to zero before creating the devfreq device
97 	 */
98 	msm_devfreq_profile.freq_table = NULL;
99 	msm_devfreq_profile.max_state = 0;
100 
101 	gpu->devfreq.devfreq = devm_devfreq_add_device(&gpu->pdev->dev,
102 			&msm_devfreq_profile, DEVFREQ_GOV_SIMPLE_ONDEMAND,
103 			NULL);
104 
105 	if (IS_ERR(gpu->devfreq.devfreq)) {
106 		DRM_DEV_ERROR(&gpu->pdev->dev, "Couldn't initialize GPU devfreq\n");
107 		gpu->devfreq.devfreq = NULL;
108 	}
109 
110 	devfreq_suspend_device(gpu->devfreq.devfreq);
111 }
112 
113 static int enable_pwrrail(struct msm_gpu *gpu)
114 {
115 	struct drm_device *dev = gpu->dev;
116 	int ret = 0;
117 
118 	if (gpu->gpu_reg) {
119 		ret = regulator_enable(gpu->gpu_reg);
120 		if (ret) {
121 			DRM_DEV_ERROR(dev->dev, "failed to enable 'gpu_reg': %d\n", ret);
122 			return ret;
123 		}
124 	}
125 
126 	if (gpu->gpu_cx) {
127 		ret = regulator_enable(gpu->gpu_cx);
128 		if (ret) {
129 			DRM_DEV_ERROR(dev->dev, "failed to enable 'gpu_cx': %d\n", ret);
130 			return ret;
131 		}
132 	}
133 
134 	return 0;
135 }
136 
137 static int disable_pwrrail(struct msm_gpu *gpu)
138 {
139 	if (gpu->gpu_cx)
140 		regulator_disable(gpu->gpu_cx);
141 	if (gpu->gpu_reg)
142 		regulator_disable(gpu->gpu_reg);
143 	return 0;
144 }
145 
146 static int enable_clk(struct msm_gpu *gpu)
147 {
148 	if (gpu->core_clk && gpu->fast_rate)
149 		clk_set_rate(gpu->core_clk, gpu->fast_rate);
150 
151 	/* Set the RBBM timer rate to 19.2Mhz */
152 	if (gpu->rbbmtimer_clk)
153 		clk_set_rate(gpu->rbbmtimer_clk, 19200000);
154 
155 	return clk_bulk_prepare_enable(gpu->nr_clocks, gpu->grp_clks);
156 }
157 
158 static int disable_clk(struct msm_gpu *gpu)
159 {
160 	clk_bulk_disable_unprepare(gpu->nr_clocks, gpu->grp_clks);
161 
162 	/*
163 	 * Set the clock to a deliberately low rate. On older targets the clock
164 	 * speed had to be non zero to avoid problems. On newer targets this
165 	 * will be rounded down to zero anyway so it all works out.
166 	 */
167 	if (gpu->core_clk)
168 		clk_set_rate(gpu->core_clk, 27000000);
169 
170 	if (gpu->rbbmtimer_clk)
171 		clk_set_rate(gpu->rbbmtimer_clk, 0);
172 
173 	return 0;
174 }
175 
176 static int enable_axi(struct msm_gpu *gpu)
177 {
178 	if (gpu->ebi1_clk)
179 		clk_prepare_enable(gpu->ebi1_clk);
180 	return 0;
181 }
182 
183 static int disable_axi(struct msm_gpu *gpu)
184 {
185 	if (gpu->ebi1_clk)
186 		clk_disable_unprepare(gpu->ebi1_clk);
187 	return 0;
188 }
189 
190 void msm_gpu_resume_devfreq(struct msm_gpu *gpu)
191 {
192 	gpu->devfreq.busy_cycles = 0;
193 	gpu->devfreq.time = ktime_get();
194 
195 	devfreq_resume_device(gpu->devfreq.devfreq);
196 }
197 
198 int msm_gpu_pm_resume(struct msm_gpu *gpu)
199 {
200 	int ret;
201 
202 	DBG("%s", gpu->name);
203 
204 	ret = enable_pwrrail(gpu);
205 	if (ret)
206 		return ret;
207 
208 	ret = enable_clk(gpu);
209 	if (ret)
210 		return ret;
211 
212 	ret = enable_axi(gpu);
213 	if (ret)
214 		return ret;
215 
216 	msm_gpu_resume_devfreq(gpu);
217 
218 	gpu->needs_hw_init = true;
219 
220 	return 0;
221 }
222 
223 int msm_gpu_pm_suspend(struct msm_gpu *gpu)
224 {
225 	int ret;
226 
227 	DBG("%s", gpu->name);
228 
229 	devfreq_suspend_device(gpu->devfreq.devfreq);
230 
231 	ret = disable_axi(gpu);
232 	if (ret)
233 		return ret;
234 
235 	ret = disable_clk(gpu);
236 	if (ret)
237 		return ret;
238 
239 	ret = disable_pwrrail(gpu);
240 	if (ret)
241 		return ret;
242 
243 	return 0;
244 }
245 
246 int msm_gpu_hw_init(struct msm_gpu *gpu)
247 {
248 	int ret;
249 
250 	WARN_ON(!mutex_is_locked(&gpu->dev->struct_mutex));
251 
252 	if (!gpu->needs_hw_init)
253 		return 0;
254 
255 	disable_irq(gpu->irq);
256 	ret = gpu->funcs->hw_init(gpu);
257 	if (!ret)
258 		gpu->needs_hw_init = false;
259 	enable_irq(gpu->irq);
260 
261 	return ret;
262 }
263 
264 #ifdef CONFIG_DEV_COREDUMP
265 static ssize_t msm_gpu_devcoredump_read(char *buffer, loff_t offset,
266 		size_t count, void *data, size_t datalen)
267 {
268 	struct msm_gpu *gpu = data;
269 	struct drm_print_iterator iter;
270 	struct drm_printer p;
271 	struct msm_gpu_state *state;
272 
273 	state = msm_gpu_crashstate_get(gpu);
274 	if (!state)
275 		return 0;
276 
277 	iter.data = buffer;
278 	iter.offset = 0;
279 	iter.start = offset;
280 	iter.remain = count;
281 
282 	p = drm_coredump_printer(&iter);
283 
284 	drm_printf(&p, "---\n");
285 	drm_printf(&p, "kernel: " UTS_RELEASE "\n");
286 	drm_printf(&p, "module: " KBUILD_MODNAME "\n");
287 	drm_printf(&p, "time: %lld.%09ld\n",
288 		state->time.tv_sec, state->time.tv_nsec);
289 	if (state->comm)
290 		drm_printf(&p, "comm: %s\n", state->comm);
291 	if (state->cmd)
292 		drm_printf(&p, "cmdline: %s\n", state->cmd);
293 
294 	gpu->funcs->show(gpu, state, &p);
295 
296 	msm_gpu_crashstate_put(gpu);
297 
298 	return count - iter.remain;
299 }
300 
301 static void msm_gpu_devcoredump_free(void *data)
302 {
303 	struct msm_gpu *gpu = data;
304 
305 	msm_gpu_crashstate_put(gpu);
306 }
307 
308 static void msm_gpu_crashstate_get_bo(struct msm_gpu_state *state,
309 		struct msm_gem_object *obj, u64 iova, u32 flags)
310 {
311 	struct msm_gpu_state_bo *state_bo = &state->bos[state->nr_bos];
312 
313 	/* Don't record write only objects */
314 	state_bo->size = obj->base.size;
315 	state_bo->iova = iova;
316 
317 	/* Only store data for non imported buffer objects marked for read */
318 	if ((flags & MSM_SUBMIT_BO_READ) && !obj->base.import_attach) {
319 		void *ptr;
320 
321 		state_bo->data = kvmalloc(obj->base.size, GFP_KERNEL);
322 		if (!state_bo->data)
323 			goto out;
324 
325 		ptr = msm_gem_get_vaddr_active(&obj->base);
326 		if (IS_ERR(ptr)) {
327 			kvfree(state_bo->data);
328 			state_bo->data = NULL;
329 			goto out;
330 		}
331 
332 		memcpy(state_bo->data, ptr, obj->base.size);
333 		msm_gem_put_vaddr(&obj->base);
334 	}
335 out:
336 	state->nr_bos++;
337 }
338 
339 static void msm_gpu_crashstate_capture(struct msm_gpu *gpu,
340 		struct msm_gem_submit *submit, char *comm, char *cmd)
341 {
342 	struct msm_gpu_state *state;
343 
344 	/* Check if the target supports capturing crash state */
345 	if (!gpu->funcs->gpu_state_get)
346 		return;
347 
348 	/* Only save one crash state at a time */
349 	if (gpu->crashstate)
350 		return;
351 
352 	state = gpu->funcs->gpu_state_get(gpu);
353 	if (IS_ERR_OR_NULL(state))
354 		return;
355 
356 	/* Fill in the additional crash state information */
357 	state->comm = kstrdup(comm, GFP_KERNEL);
358 	state->cmd = kstrdup(cmd, GFP_KERNEL);
359 
360 	if (submit) {
361 		int i, nr = 0;
362 
363 		/* count # of buffers to dump: */
364 		for (i = 0; i < submit->nr_bos; i++)
365 			if (should_dump(submit, i))
366 				nr++;
367 		/* always dump cmd bo's, but don't double count them: */
368 		for (i = 0; i < submit->nr_cmds; i++)
369 			if (!should_dump(submit, submit->cmd[i].idx))
370 				nr++;
371 
372 		state->bos = kcalloc(nr,
373 			sizeof(struct msm_gpu_state_bo), GFP_KERNEL);
374 
375 		for (i = 0; i < submit->nr_bos; i++) {
376 			if (should_dump(submit, i)) {
377 				msm_gpu_crashstate_get_bo(state, submit->bos[i].obj,
378 					submit->bos[i].iova, submit->bos[i].flags);
379 			}
380 		}
381 
382 		for (i = 0; state->bos && i < submit->nr_cmds; i++) {
383 			int idx = submit->cmd[i].idx;
384 
385 			if (!should_dump(submit, submit->cmd[i].idx)) {
386 				msm_gpu_crashstate_get_bo(state, submit->bos[idx].obj,
387 					submit->bos[idx].iova, submit->bos[idx].flags);
388 			}
389 		}
390 	}
391 
392 	/* Set the active crash state to be dumped on failure */
393 	gpu->crashstate = state;
394 
395 	/* FIXME: Release the crashstate if this errors out? */
396 	dev_coredumpm(gpu->dev->dev, THIS_MODULE, gpu, 0, GFP_KERNEL,
397 		msm_gpu_devcoredump_read, msm_gpu_devcoredump_free);
398 }
399 #else
400 static void msm_gpu_crashstate_capture(struct msm_gpu *gpu,
401 		struct msm_gem_submit *submit, char *comm, char *cmd)
402 {
403 }
404 #endif
405 
406 /*
407  * Hangcheck detection for locked gpu:
408  */
409 
410 static void update_fences(struct msm_gpu *gpu, struct msm_ringbuffer *ring,
411 		uint32_t fence)
412 {
413 	struct msm_gem_submit *submit;
414 
415 	list_for_each_entry(submit, &ring->submits, node) {
416 		if (submit->seqno > fence)
417 			break;
418 
419 		msm_update_fence(submit->ring->fctx,
420 			submit->fence->seqno);
421 	}
422 }
423 
424 static struct msm_gem_submit *
425 find_submit(struct msm_ringbuffer *ring, uint32_t fence)
426 {
427 	struct msm_gem_submit *submit;
428 
429 	WARN_ON(!mutex_is_locked(&ring->gpu->dev->struct_mutex));
430 
431 	list_for_each_entry(submit, &ring->submits, node)
432 		if (submit->seqno == fence)
433 			return submit;
434 
435 	return NULL;
436 }
437 
438 static void retire_submits(struct msm_gpu *gpu);
439 
440 static void recover_worker(struct work_struct *work)
441 {
442 	struct msm_gpu *gpu = container_of(work, struct msm_gpu, recover_work);
443 	struct drm_device *dev = gpu->dev;
444 	struct msm_drm_private *priv = dev->dev_private;
445 	struct msm_gem_submit *submit;
446 	struct msm_ringbuffer *cur_ring = gpu->funcs->active_ring(gpu);
447 	char *comm = NULL, *cmd = NULL;
448 	int i;
449 
450 	mutex_lock(&dev->struct_mutex);
451 
452 	DRM_DEV_ERROR(dev->dev, "%s: hangcheck recover!\n", gpu->name);
453 
454 	submit = find_submit(cur_ring, cur_ring->memptrs->fence + 1);
455 	if (submit) {
456 		struct task_struct *task;
457 
458 		/* Increment the fault counts */
459 		gpu->global_faults++;
460 		submit->queue->faults++;
461 
462 		task = get_pid_task(submit->pid, PIDTYPE_PID);
463 		if (task) {
464 			comm = kstrdup(task->comm, GFP_KERNEL);
465 			cmd = kstrdup_quotable_cmdline(task, GFP_KERNEL);
466 			put_task_struct(task);
467 		}
468 
469 		if (comm && cmd) {
470 			DRM_DEV_ERROR(dev->dev, "%s: offending task: %s (%s)\n",
471 				gpu->name, comm, cmd);
472 
473 			msm_rd_dump_submit(priv->hangrd, submit,
474 				"offending task: %s (%s)", comm, cmd);
475 		} else
476 			msm_rd_dump_submit(priv->hangrd, submit, NULL);
477 	}
478 
479 	/* Record the crash state */
480 	pm_runtime_get_sync(&gpu->pdev->dev);
481 	msm_gpu_crashstate_capture(gpu, submit, comm, cmd);
482 	pm_runtime_put_sync(&gpu->pdev->dev);
483 
484 	kfree(cmd);
485 	kfree(comm);
486 
487 	/*
488 	 * Update all the rings with the latest and greatest fence.. this
489 	 * needs to happen after msm_rd_dump_submit() to ensure that the
490 	 * bo's referenced by the offending submit are still around.
491 	 */
492 	for (i = 0; i < gpu->nr_rings; i++) {
493 		struct msm_ringbuffer *ring = gpu->rb[i];
494 
495 		uint32_t fence = ring->memptrs->fence;
496 
497 		/*
498 		 * For the current (faulting?) ring/submit advance the fence by
499 		 * one more to clear the faulting submit
500 		 */
501 		if (ring == cur_ring)
502 			fence++;
503 
504 		update_fences(gpu, ring, fence);
505 	}
506 
507 	if (msm_gpu_active(gpu)) {
508 		/* retire completed submits, plus the one that hung: */
509 		retire_submits(gpu);
510 
511 		pm_runtime_get_sync(&gpu->pdev->dev);
512 		gpu->funcs->recover(gpu);
513 		pm_runtime_put_sync(&gpu->pdev->dev);
514 
515 		/*
516 		 * Replay all remaining submits starting with highest priority
517 		 * ring
518 		 */
519 		for (i = 0; i < gpu->nr_rings; i++) {
520 			struct msm_ringbuffer *ring = gpu->rb[i];
521 
522 			list_for_each_entry(submit, &ring->submits, node)
523 				gpu->funcs->submit(gpu, submit, NULL);
524 		}
525 	}
526 
527 	mutex_unlock(&dev->struct_mutex);
528 
529 	msm_gpu_retire(gpu);
530 }
531 
532 static void hangcheck_timer_reset(struct msm_gpu *gpu)
533 {
534 	DBG("%s", gpu->name);
535 	mod_timer(&gpu->hangcheck_timer,
536 			round_jiffies_up(jiffies + DRM_MSM_HANGCHECK_JIFFIES));
537 }
538 
539 static void hangcheck_handler(struct timer_list *t)
540 {
541 	struct msm_gpu *gpu = from_timer(gpu, t, hangcheck_timer);
542 	struct drm_device *dev = gpu->dev;
543 	struct msm_drm_private *priv = dev->dev_private;
544 	struct msm_ringbuffer *ring = gpu->funcs->active_ring(gpu);
545 	uint32_t fence = ring->memptrs->fence;
546 
547 	if (fence != ring->hangcheck_fence) {
548 		/* some progress has been made.. ya! */
549 		ring->hangcheck_fence = fence;
550 	} else if (fence < ring->seqno) {
551 		/* no progress and not done.. hung! */
552 		ring->hangcheck_fence = fence;
553 		DRM_DEV_ERROR(dev->dev, "%s: hangcheck detected gpu lockup rb %d!\n",
554 				gpu->name, ring->id);
555 		DRM_DEV_ERROR(dev->dev, "%s:     completed fence: %u\n",
556 				gpu->name, fence);
557 		DRM_DEV_ERROR(dev->dev, "%s:     submitted fence: %u\n",
558 				gpu->name, ring->seqno);
559 
560 		queue_work(priv->wq, &gpu->recover_work);
561 	}
562 
563 	/* if still more pending work, reset the hangcheck timer: */
564 	if (ring->seqno > ring->hangcheck_fence)
565 		hangcheck_timer_reset(gpu);
566 
567 	/* workaround for missing irq: */
568 	queue_work(priv->wq, &gpu->retire_work);
569 }
570 
571 /*
572  * Performance Counters:
573  */
574 
575 /* called under perf_lock */
576 static int update_hw_cntrs(struct msm_gpu *gpu, uint32_t ncntrs, uint32_t *cntrs)
577 {
578 	uint32_t current_cntrs[ARRAY_SIZE(gpu->last_cntrs)];
579 	int i, n = min(ncntrs, gpu->num_perfcntrs);
580 
581 	/* read current values: */
582 	for (i = 0; i < gpu->num_perfcntrs; i++)
583 		current_cntrs[i] = gpu_read(gpu, gpu->perfcntrs[i].sample_reg);
584 
585 	/* update cntrs: */
586 	for (i = 0; i < n; i++)
587 		cntrs[i] = current_cntrs[i] - gpu->last_cntrs[i];
588 
589 	/* save current values: */
590 	for (i = 0; i < gpu->num_perfcntrs; i++)
591 		gpu->last_cntrs[i] = current_cntrs[i];
592 
593 	return n;
594 }
595 
596 static void update_sw_cntrs(struct msm_gpu *gpu)
597 {
598 	ktime_t time;
599 	uint32_t elapsed;
600 	unsigned long flags;
601 
602 	spin_lock_irqsave(&gpu->perf_lock, flags);
603 	if (!gpu->perfcntr_active)
604 		goto out;
605 
606 	time = ktime_get();
607 	elapsed = ktime_to_us(ktime_sub(time, gpu->last_sample.time));
608 
609 	gpu->totaltime += elapsed;
610 	if (gpu->last_sample.active)
611 		gpu->activetime += elapsed;
612 
613 	gpu->last_sample.active = msm_gpu_active(gpu);
614 	gpu->last_sample.time = time;
615 
616 out:
617 	spin_unlock_irqrestore(&gpu->perf_lock, flags);
618 }
619 
620 void msm_gpu_perfcntr_start(struct msm_gpu *gpu)
621 {
622 	unsigned long flags;
623 
624 	pm_runtime_get_sync(&gpu->pdev->dev);
625 
626 	spin_lock_irqsave(&gpu->perf_lock, flags);
627 	/* we could dynamically enable/disable perfcntr registers too.. */
628 	gpu->last_sample.active = msm_gpu_active(gpu);
629 	gpu->last_sample.time = ktime_get();
630 	gpu->activetime = gpu->totaltime = 0;
631 	gpu->perfcntr_active = true;
632 	update_hw_cntrs(gpu, 0, NULL);
633 	spin_unlock_irqrestore(&gpu->perf_lock, flags);
634 }
635 
636 void msm_gpu_perfcntr_stop(struct msm_gpu *gpu)
637 {
638 	gpu->perfcntr_active = false;
639 	pm_runtime_put_sync(&gpu->pdev->dev);
640 }
641 
642 /* returns -errno or # of cntrs sampled */
643 int msm_gpu_perfcntr_sample(struct msm_gpu *gpu, uint32_t *activetime,
644 		uint32_t *totaltime, uint32_t ncntrs, uint32_t *cntrs)
645 {
646 	unsigned long flags;
647 	int ret;
648 
649 	spin_lock_irqsave(&gpu->perf_lock, flags);
650 
651 	if (!gpu->perfcntr_active) {
652 		ret = -EINVAL;
653 		goto out;
654 	}
655 
656 	*activetime = gpu->activetime;
657 	*totaltime = gpu->totaltime;
658 
659 	gpu->activetime = gpu->totaltime = 0;
660 
661 	ret = update_hw_cntrs(gpu, ncntrs, cntrs);
662 
663 out:
664 	spin_unlock_irqrestore(&gpu->perf_lock, flags);
665 
666 	return ret;
667 }
668 
669 /*
670  * Cmdstream submission/retirement:
671  */
672 
673 static void retire_submit(struct msm_gpu *gpu, struct msm_ringbuffer *ring,
674 		struct msm_gem_submit *submit)
675 {
676 	int index = submit->seqno % MSM_GPU_SUBMIT_STATS_COUNT;
677 	volatile struct msm_gpu_submit_stats *stats;
678 	u64 elapsed, clock = 0;
679 	int i;
680 
681 	stats = &ring->memptrs->stats[index];
682 	/* Convert 19.2Mhz alwayson ticks to nanoseconds for elapsed time */
683 	elapsed = (stats->alwayson_end - stats->alwayson_start) * 10000;
684 	do_div(elapsed, 192);
685 
686 	/* Calculate the clock frequency from the number of CP cycles */
687 	if (elapsed) {
688 		clock = (stats->cpcycles_end - stats->cpcycles_start) * 1000;
689 		do_div(clock, elapsed);
690 	}
691 
692 	trace_msm_gpu_submit_retired(submit, elapsed, clock,
693 		stats->alwayson_start, stats->alwayson_end);
694 
695 	for (i = 0; i < submit->nr_bos; i++) {
696 		struct msm_gem_object *msm_obj = submit->bos[i].obj;
697 		/* move to inactive: */
698 		msm_gem_move_to_inactive(&msm_obj->base);
699 		msm_gem_unpin_iova(&msm_obj->base, submit->aspace);
700 		drm_gem_object_put_locked(&msm_obj->base);
701 	}
702 
703 	pm_runtime_mark_last_busy(&gpu->pdev->dev);
704 	pm_runtime_put_autosuspend(&gpu->pdev->dev);
705 	msm_gem_submit_free(submit);
706 }
707 
708 static void retire_submits(struct msm_gpu *gpu)
709 {
710 	struct drm_device *dev = gpu->dev;
711 	struct msm_gem_submit *submit, *tmp;
712 	int i;
713 
714 	WARN_ON(!mutex_is_locked(&dev->struct_mutex));
715 
716 	/* Retire the commits starting with highest priority */
717 	for (i = 0; i < gpu->nr_rings; i++) {
718 		struct msm_ringbuffer *ring = gpu->rb[i];
719 
720 		list_for_each_entry_safe(submit, tmp, &ring->submits, node) {
721 			if (dma_fence_is_signaled(submit->fence))
722 				retire_submit(gpu, ring, submit);
723 		}
724 	}
725 }
726 
727 static void retire_worker(struct work_struct *work)
728 {
729 	struct msm_gpu *gpu = container_of(work, struct msm_gpu, retire_work);
730 	struct drm_device *dev = gpu->dev;
731 	int i;
732 
733 	for (i = 0; i < gpu->nr_rings; i++)
734 		update_fences(gpu, gpu->rb[i], gpu->rb[i]->memptrs->fence);
735 
736 	mutex_lock(&dev->struct_mutex);
737 	retire_submits(gpu);
738 	mutex_unlock(&dev->struct_mutex);
739 }
740 
741 /* call from irq handler to schedule work to retire bo's */
742 void msm_gpu_retire(struct msm_gpu *gpu)
743 {
744 	struct msm_drm_private *priv = gpu->dev->dev_private;
745 	queue_work(priv->wq, &gpu->retire_work);
746 	update_sw_cntrs(gpu);
747 }
748 
749 /* add bo's to gpu's ring, and kick gpu: */
750 void msm_gpu_submit(struct msm_gpu *gpu, struct msm_gem_submit *submit,
751 		struct msm_file_private *ctx)
752 {
753 	struct drm_device *dev = gpu->dev;
754 	struct msm_drm_private *priv = dev->dev_private;
755 	struct msm_ringbuffer *ring = submit->ring;
756 	int i;
757 
758 	WARN_ON(!mutex_is_locked(&dev->struct_mutex));
759 
760 	pm_runtime_get_sync(&gpu->pdev->dev);
761 
762 	msm_gpu_hw_init(gpu);
763 
764 	submit->seqno = ++ring->seqno;
765 
766 	list_add_tail(&submit->node, &ring->submits);
767 
768 	msm_rd_dump_submit(priv->rd, submit, NULL);
769 
770 	update_sw_cntrs(gpu);
771 
772 	for (i = 0; i < submit->nr_bos; i++) {
773 		struct msm_gem_object *msm_obj = submit->bos[i].obj;
774 		uint64_t iova;
775 
776 		/* can't happen yet.. but when we add 2d support we'll have
777 		 * to deal w/ cross-ring synchronization:
778 		 */
779 		WARN_ON(is_active(msm_obj) && (msm_obj->gpu != gpu));
780 
781 		/* submit takes a reference to the bo and iova until retired: */
782 		drm_gem_object_get(&msm_obj->base);
783 		msm_gem_get_and_pin_iova(&msm_obj->base, submit->aspace, &iova);
784 
785 		if (submit->bos[i].flags & MSM_SUBMIT_BO_WRITE)
786 			msm_gem_move_to_active(&msm_obj->base, gpu, true, submit->fence);
787 		else if (submit->bos[i].flags & MSM_SUBMIT_BO_READ)
788 			msm_gem_move_to_active(&msm_obj->base, gpu, false, submit->fence);
789 	}
790 
791 	gpu->funcs->submit(gpu, submit, ctx);
792 	priv->lastctx = ctx;
793 
794 	hangcheck_timer_reset(gpu);
795 }
796 
797 /*
798  * Init/Cleanup:
799  */
800 
801 static irqreturn_t irq_handler(int irq, void *data)
802 {
803 	struct msm_gpu *gpu = data;
804 	return gpu->funcs->irq(gpu);
805 }
806 
807 static int get_clocks(struct platform_device *pdev, struct msm_gpu *gpu)
808 {
809 	int ret = devm_clk_bulk_get_all(&pdev->dev, &gpu->grp_clks);
810 
811 	if (ret < 1) {
812 		gpu->nr_clocks = 0;
813 		return ret;
814 	}
815 
816 	gpu->nr_clocks = ret;
817 
818 	gpu->core_clk = msm_clk_bulk_get_clock(gpu->grp_clks,
819 		gpu->nr_clocks, "core");
820 
821 	gpu->rbbmtimer_clk = msm_clk_bulk_get_clock(gpu->grp_clks,
822 		gpu->nr_clocks, "rbbmtimer");
823 
824 	return 0;
825 }
826 
827 int msm_gpu_init(struct drm_device *drm, struct platform_device *pdev,
828 		struct msm_gpu *gpu, const struct msm_gpu_funcs *funcs,
829 		const char *name, struct msm_gpu_config *config)
830 {
831 	int i, ret, nr_rings = config->nr_rings;
832 	void *memptrs;
833 	uint64_t memptrs_iova;
834 
835 	if (WARN_ON(gpu->num_perfcntrs > ARRAY_SIZE(gpu->last_cntrs)))
836 		gpu->num_perfcntrs = ARRAY_SIZE(gpu->last_cntrs);
837 
838 	gpu->dev = drm;
839 	gpu->funcs = funcs;
840 	gpu->name = name;
841 
842 	INIT_LIST_HEAD(&gpu->active_list);
843 	INIT_WORK(&gpu->retire_work, retire_worker);
844 	INIT_WORK(&gpu->recover_work, recover_worker);
845 
846 
847 	timer_setup(&gpu->hangcheck_timer, hangcheck_handler, 0);
848 
849 	spin_lock_init(&gpu->perf_lock);
850 
851 
852 	/* Map registers: */
853 	gpu->mmio = msm_ioremap(pdev, config->ioname, name);
854 	if (IS_ERR(gpu->mmio)) {
855 		ret = PTR_ERR(gpu->mmio);
856 		goto fail;
857 	}
858 
859 	/* Get Interrupt: */
860 	gpu->irq = platform_get_irq(pdev, 0);
861 	if (gpu->irq < 0) {
862 		ret = gpu->irq;
863 		DRM_DEV_ERROR(drm->dev, "failed to get irq: %d\n", ret);
864 		goto fail;
865 	}
866 
867 	ret = devm_request_irq(&pdev->dev, gpu->irq, irq_handler,
868 			IRQF_TRIGGER_HIGH, gpu->name, gpu);
869 	if (ret) {
870 		DRM_DEV_ERROR(drm->dev, "failed to request IRQ%u: %d\n", gpu->irq, ret);
871 		goto fail;
872 	}
873 
874 	ret = get_clocks(pdev, gpu);
875 	if (ret)
876 		goto fail;
877 
878 	gpu->ebi1_clk = msm_clk_get(pdev, "bus");
879 	DBG("ebi1_clk: %p", gpu->ebi1_clk);
880 	if (IS_ERR(gpu->ebi1_clk))
881 		gpu->ebi1_clk = NULL;
882 
883 	/* Acquire regulators: */
884 	gpu->gpu_reg = devm_regulator_get(&pdev->dev, "vdd");
885 	DBG("gpu_reg: %p", gpu->gpu_reg);
886 	if (IS_ERR(gpu->gpu_reg))
887 		gpu->gpu_reg = NULL;
888 
889 	gpu->gpu_cx = devm_regulator_get(&pdev->dev, "vddcx");
890 	DBG("gpu_cx: %p", gpu->gpu_cx);
891 	if (IS_ERR(gpu->gpu_cx))
892 		gpu->gpu_cx = NULL;
893 
894 	gpu->pdev = pdev;
895 	platform_set_drvdata(pdev, gpu);
896 
897 	msm_devfreq_init(gpu);
898 
899 
900 	gpu->aspace = gpu->funcs->create_address_space(gpu, pdev);
901 
902 	if (gpu->aspace == NULL)
903 		DRM_DEV_INFO(drm->dev, "%s: no IOMMU, fallback to VRAM carveout!\n", name);
904 	else if (IS_ERR(gpu->aspace)) {
905 		ret = PTR_ERR(gpu->aspace);
906 		goto fail;
907 	}
908 
909 	memptrs = msm_gem_kernel_new(drm,
910 		sizeof(struct msm_rbmemptrs) * nr_rings,
911 		check_apriv(gpu, MSM_BO_UNCACHED), gpu->aspace, &gpu->memptrs_bo,
912 		&memptrs_iova);
913 
914 	if (IS_ERR(memptrs)) {
915 		ret = PTR_ERR(memptrs);
916 		DRM_DEV_ERROR(drm->dev, "could not allocate memptrs: %d\n", ret);
917 		goto fail;
918 	}
919 
920 	msm_gem_object_set_name(gpu->memptrs_bo, "memptrs");
921 
922 	if (nr_rings > ARRAY_SIZE(gpu->rb)) {
923 		DRM_DEV_INFO_ONCE(drm->dev, "Only creating %zu ringbuffers\n",
924 			ARRAY_SIZE(gpu->rb));
925 		nr_rings = ARRAY_SIZE(gpu->rb);
926 	}
927 
928 	/* Create ringbuffer(s): */
929 	for (i = 0; i < nr_rings; i++) {
930 		gpu->rb[i] = msm_ringbuffer_new(gpu, i, memptrs, memptrs_iova);
931 
932 		if (IS_ERR(gpu->rb[i])) {
933 			ret = PTR_ERR(gpu->rb[i]);
934 			DRM_DEV_ERROR(drm->dev,
935 				"could not create ringbuffer %d: %d\n", i, ret);
936 			goto fail;
937 		}
938 
939 		memptrs += sizeof(struct msm_rbmemptrs);
940 		memptrs_iova += sizeof(struct msm_rbmemptrs);
941 	}
942 
943 	gpu->nr_rings = nr_rings;
944 
945 	return 0;
946 
947 fail:
948 	for (i = 0; i < ARRAY_SIZE(gpu->rb); i++)  {
949 		msm_ringbuffer_destroy(gpu->rb[i]);
950 		gpu->rb[i] = NULL;
951 	}
952 
953 	msm_gem_kernel_put(gpu->memptrs_bo, gpu->aspace, false);
954 
955 	platform_set_drvdata(pdev, NULL);
956 	return ret;
957 }
958 
959 void msm_gpu_cleanup(struct msm_gpu *gpu)
960 {
961 	int i;
962 
963 	DBG("%s", gpu->name);
964 
965 	WARN_ON(!list_empty(&gpu->active_list));
966 
967 	for (i = 0; i < ARRAY_SIZE(gpu->rb); i++) {
968 		msm_ringbuffer_destroy(gpu->rb[i]);
969 		gpu->rb[i] = NULL;
970 	}
971 
972 	msm_gem_kernel_put(gpu->memptrs_bo, gpu->aspace, false);
973 
974 	if (!IS_ERR_OR_NULL(gpu->aspace)) {
975 		gpu->aspace->mmu->funcs->detach(gpu->aspace->mmu);
976 		msm_gem_address_space_put(gpu->aspace);
977 	}
978 }
979