xref: /openbmc/linux/drivers/gpu/drm/msm/msm_gpu.c (revision dc6a81c3)
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;
359 
360 		state->bos = kcalloc(submit->nr_cmds,
361 			sizeof(struct msm_gpu_state_bo), GFP_KERNEL);
362 
363 		for (i = 0; state->bos && i < submit->nr_cmds; i++) {
364 			int idx = submit->cmd[i].idx;
365 
366 			msm_gpu_crashstate_get_bo(state, submit->bos[idx].obj,
367 				submit->bos[idx].iova, submit->bos[idx].flags);
368 		}
369 	}
370 
371 	/* Set the active crash state to be dumped on failure */
372 	gpu->crashstate = state;
373 
374 	/* FIXME: Release the crashstate if this errors out? */
375 	dev_coredumpm(gpu->dev->dev, THIS_MODULE, gpu, 0, GFP_KERNEL,
376 		msm_gpu_devcoredump_read, msm_gpu_devcoredump_free);
377 }
378 #else
379 static void msm_gpu_crashstate_capture(struct msm_gpu *gpu,
380 		struct msm_gem_submit *submit, char *comm, char *cmd)
381 {
382 }
383 #endif
384 
385 /*
386  * Hangcheck detection for locked gpu:
387  */
388 
389 static void update_fences(struct msm_gpu *gpu, struct msm_ringbuffer *ring,
390 		uint32_t fence)
391 {
392 	struct msm_gem_submit *submit;
393 
394 	list_for_each_entry(submit, &ring->submits, node) {
395 		if (submit->seqno > fence)
396 			break;
397 
398 		msm_update_fence(submit->ring->fctx,
399 			submit->fence->seqno);
400 	}
401 }
402 
403 static struct msm_gem_submit *
404 find_submit(struct msm_ringbuffer *ring, uint32_t fence)
405 {
406 	struct msm_gem_submit *submit;
407 
408 	WARN_ON(!mutex_is_locked(&ring->gpu->dev->struct_mutex));
409 
410 	list_for_each_entry(submit, &ring->submits, node)
411 		if (submit->seqno == fence)
412 			return submit;
413 
414 	return NULL;
415 }
416 
417 static void retire_submits(struct msm_gpu *gpu);
418 
419 static void recover_worker(struct work_struct *work)
420 {
421 	struct msm_gpu *gpu = container_of(work, struct msm_gpu, recover_work);
422 	struct drm_device *dev = gpu->dev;
423 	struct msm_drm_private *priv = dev->dev_private;
424 	struct msm_gem_submit *submit;
425 	struct msm_ringbuffer *cur_ring = gpu->funcs->active_ring(gpu);
426 	char *comm = NULL, *cmd = NULL;
427 	int i;
428 
429 	mutex_lock(&dev->struct_mutex);
430 
431 	DRM_DEV_ERROR(dev->dev, "%s: hangcheck recover!\n", gpu->name);
432 
433 	submit = find_submit(cur_ring, cur_ring->memptrs->fence + 1);
434 	if (submit) {
435 		struct task_struct *task;
436 
437 		/* Increment the fault counts */
438 		gpu->global_faults++;
439 		submit->queue->faults++;
440 
441 		task = get_pid_task(submit->pid, PIDTYPE_PID);
442 		if (task) {
443 			comm = kstrdup(task->comm, GFP_KERNEL);
444 			cmd = kstrdup_quotable_cmdline(task, GFP_KERNEL);
445 			put_task_struct(task);
446 		}
447 
448 		if (comm && cmd) {
449 			DRM_DEV_ERROR(dev->dev, "%s: offending task: %s (%s)\n",
450 				gpu->name, comm, cmd);
451 
452 			msm_rd_dump_submit(priv->hangrd, submit,
453 				"offending task: %s (%s)", comm, cmd);
454 		} else
455 			msm_rd_dump_submit(priv->hangrd, submit, NULL);
456 	}
457 
458 	/* Record the crash state */
459 	pm_runtime_get_sync(&gpu->pdev->dev);
460 	msm_gpu_crashstate_capture(gpu, submit, comm, cmd);
461 	pm_runtime_put_sync(&gpu->pdev->dev);
462 
463 	kfree(cmd);
464 	kfree(comm);
465 
466 	/*
467 	 * Update all the rings with the latest and greatest fence.. this
468 	 * needs to happen after msm_rd_dump_submit() to ensure that the
469 	 * bo's referenced by the offending submit are still around.
470 	 */
471 	for (i = 0; i < gpu->nr_rings; i++) {
472 		struct msm_ringbuffer *ring = gpu->rb[i];
473 
474 		uint32_t fence = ring->memptrs->fence;
475 
476 		/*
477 		 * For the current (faulting?) ring/submit advance the fence by
478 		 * one more to clear the faulting submit
479 		 */
480 		if (ring == cur_ring)
481 			fence++;
482 
483 		update_fences(gpu, ring, fence);
484 	}
485 
486 	if (msm_gpu_active(gpu)) {
487 		/* retire completed submits, plus the one that hung: */
488 		retire_submits(gpu);
489 
490 		pm_runtime_get_sync(&gpu->pdev->dev);
491 		gpu->funcs->recover(gpu);
492 		pm_runtime_put_sync(&gpu->pdev->dev);
493 
494 		/*
495 		 * Replay all remaining submits starting with highest priority
496 		 * ring
497 		 */
498 		for (i = 0; i < gpu->nr_rings; i++) {
499 			struct msm_ringbuffer *ring = gpu->rb[i];
500 
501 			list_for_each_entry(submit, &ring->submits, node)
502 				gpu->funcs->submit(gpu, submit, NULL);
503 		}
504 	}
505 
506 	mutex_unlock(&dev->struct_mutex);
507 
508 	msm_gpu_retire(gpu);
509 }
510 
511 static void hangcheck_timer_reset(struct msm_gpu *gpu)
512 {
513 	DBG("%s", gpu->name);
514 	mod_timer(&gpu->hangcheck_timer,
515 			round_jiffies_up(jiffies + DRM_MSM_HANGCHECK_JIFFIES));
516 }
517 
518 static void hangcheck_handler(struct timer_list *t)
519 {
520 	struct msm_gpu *gpu = from_timer(gpu, t, hangcheck_timer);
521 	struct drm_device *dev = gpu->dev;
522 	struct msm_drm_private *priv = dev->dev_private;
523 	struct msm_ringbuffer *ring = gpu->funcs->active_ring(gpu);
524 	uint32_t fence = ring->memptrs->fence;
525 
526 	if (fence != ring->hangcheck_fence) {
527 		/* some progress has been made.. ya! */
528 		ring->hangcheck_fence = fence;
529 	} else if (fence < ring->seqno) {
530 		/* no progress and not done.. hung! */
531 		ring->hangcheck_fence = fence;
532 		DRM_DEV_ERROR(dev->dev, "%s: hangcheck detected gpu lockup rb %d!\n",
533 				gpu->name, ring->id);
534 		DRM_DEV_ERROR(dev->dev, "%s:     completed fence: %u\n",
535 				gpu->name, fence);
536 		DRM_DEV_ERROR(dev->dev, "%s:     submitted fence: %u\n",
537 				gpu->name, ring->seqno);
538 
539 		queue_work(priv->wq, &gpu->recover_work);
540 	}
541 
542 	/* if still more pending work, reset the hangcheck timer: */
543 	if (ring->seqno > ring->hangcheck_fence)
544 		hangcheck_timer_reset(gpu);
545 
546 	/* workaround for missing irq: */
547 	queue_work(priv->wq, &gpu->retire_work);
548 }
549 
550 /*
551  * Performance Counters:
552  */
553 
554 /* called under perf_lock */
555 static int update_hw_cntrs(struct msm_gpu *gpu, uint32_t ncntrs, uint32_t *cntrs)
556 {
557 	uint32_t current_cntrs[ARRAY_SIZE(gpu->last_cntrs)];
558 	int i, n = min(ncntrs, gpu->num_perfcntrs);
559 
560 	/* read current values: */
561 	for (i = 0; i < gpu->num_perfcntrs; i++)
562 		current_cntrs[i] = gpu_read(gpu, gpu->perfcntrs[i].sample_reg);
563 
564 	/* update cntrs: */
565 	for (i = 0; i < n; i++)
566 		cntrs[i] = current_cntrs[i] - gpu->last_cntrs[i];
567 
568 	/* save current values: */
569 	for (i = 0; i < gpu->num_perfcntrs; i++)
570 		gpu->last_cntrs[i] = current_cntrs[i];
571 
572 	return n;
573 }
574 
575 static void update_sw_cntrs(struct msm_gpu *gpu)
576 {
577 	ktime_t time;
578 	uint32_t elapsed;
579 	unsigned long flags;
580 
581 	spin_lock_irqsave(&gpu->perf_lock, flags);
582 	if (!gpu->perfcntr_active)
583 		goto out;
584 
585 	time = ktime_get();
586 	elapsed = ktime_to_us(ktime_sub(time, gpu->last_sample.time));
587 
588 	gpu->totaltime += elapsed;
589 	if (gpu->last_sample.active)
590 		gpu->activetime += elapsed;
591 
592 	gpu->last_sample.active = msm_gpu_active(gpu);
593 	gpu->last_sample.time = time;
594 
595 out:
596 	spin_unlock_irqrestore(&gpu->perf_lock, flags);
597 }
598 
599 void msm_gpu_perfcntr_start(struct msm_gpu *gpu)
600 {
601 	unsigned long flags;
602 
603 	pm_runtime_get_sync(&gpu->pdev->dev);
604 
605 	spin_lock_irqsave(&gpu->perf_lock, flags);
606 	/* we could dynamically enable/disable perfcntr registers too.. */
607 	gpu->last_sample.active = msm_gpu_active(gpu);
608 	gpu->last_sample.time = ktime_get();
609 	gpu->activetime = gpu->totaltime = 0;
610 	gpu->perfcntr_active = true;
611 	update_hw_cntrs(gpu, 0, NULL);
612 	spin_unlock_irqrestore(&gpu->perf_lock, flags);
613 }
614 
615 void msm_gpu_perfcntr_stop(struct msm_gpu *gpu)
616 {
617 	gpu->perfcntr_active = false;
618 	pm_runtime_put_sync(&gpu->pdev->dev);
619 }
620 
621 /* returns -errno or # of cntrs sampled */
622 int msm_gpu_perfcntr_sample(struct msm_gpu *gpu, uint32_t *activetime,
623 		uint32_t *totaltime, uint32_t ncntrs, uint32_t *cntrs)
624 {
625 	unsigned long flags;
626 	int ret;
627 
628 	spin_lock_irqsave(&gpu->perf_lock, flags);
629 
630 	if (!gpu->perfcntr_active) {
631 		ret = -EINVAL;
632 		goto out;
633 	}
634 
635 	*activetime = gpu->activetime;
636 	*totaltime = gpu->totaltime;
637 
638 	gpu->activetime = gpu->totaltime = 0;
639 
640 	ret = update_hw_cntrs(gpu, ncntrs, cntrs);
641 
642 out:
643 	spin_unlock_irqrestore(&gpu->perf_lock, flags);
644 
645 	return ret;
646 }
647 
648 /*
649  * Cmdstream submission/retirement:
650  */
651 
652 static void retire_submit(struct msm_gpu *gpu, struct msm_ringbuffer *ring,
653 		struct msm_gem_submit *submit)
654 {
655 	int index = submit->seqno % MSM_GPU_SUBMIT_STATS_COUNT;
656 	volatile struct msm_gpu_submit_stats *stats;
657 	u64 elapsed, clock = 0;
658 	int i;
659 
660 	stats = &ring->memptrs->stats[index];
661 	/* Convert 19.2Mhz alwayson ticks to nanoseconds for elapsed time */
662 	elapsed = (stats->alwayson_end - stats->alwayson_start) * 10000;
663 	do_div(elapsed, 192);
664 
665 	/* Calculate the clock frequency from the number of CP cycles */
666 	if (elapsed) {
667 		clock = (stats->cpcycles_end - stats->cpcycles_start) * 1000;
668 		do_div(clock, elapsed);
669 	}
670 
671 	trace_msm_gpu_submit_retired(submit, elapsed, clock,
672 		stats->alwayson_start, stats->alwayson_end);
673 
674 	for (i = 0; i < submit->nr_bos; i++) {
675 		struct msm_gem_object *msm_obj = submit->bos[i].obj;
676 		/* move to inactive: */
677 		msm_gem_move_to_inactive(&msm_obj->base);
678 		msm_gem_unpin_iova(&msm_obj->base, submit->aspace);
679 		drm_gem_object_put(&msm_obj->base);
680 	}
681 
682 	pm_runtime_mark_last_busy(&gpu->pdev->dev);
683 	pm_runtime_put_autosuspend(&gpu->pdev->dev);
684 	msm_gem_submit_free(submit);
685 }
686 
687 static void retire_submits(struct msm_gpu *gpu)
688 {
689 	struct drm_device *dev = gpu->dev;
690 	struct msm_gem_submit *submit, *tmp;
691 	int i;
692 
693 	WARN_ON(!mutex_is_locked(&dev->struct_mutex));
694 
695 	/* Retire the commits starting with highest priority */
696 	for (i = 0; i < gpu->nr_rings; i++) {
697 		struct msm_ringbuffer *ring = gpu->rb[i];
698 
699 		list_for_each_entry_safe(submit, tmp, &ring->submits, node) {
700 			if (dma_fence_is_signaled(submit->fence))
701 				retire_submit(gpu, ring, submit);
702 		}
703 	}
704 }
705 
706 static void retire_worker(struct work_struct *work)
707 {
708 	struct msm_gpu *gpu = container_of(work, struct msm_gpu, retire_work);
709 	struct drm_device *dev = gpu->dev;
710 	int i;
711 
712 	for (i = 0; i < gpu->nr_rings; i++)
713 		update_fences(gpu, gpu->rb[i], gpu->rb[i]->memptrs->fence);
714 
715 	mutex_lock(&dev->struct_mutex);
716 	retire_submits(gpu);
717 	mutex_unlock(&dev->struct_mutex);
718 }
719 
720 /* call from irq handler to schedule work to retire bo's */
721 void msm_gpu_retire(struct msm_gpu *gpu)
722 {
723 	struct msm_drm_private *priv = gpu->dev->dev_private;
724 	queue_work(priv->wq, &gpu->retire_work);
725 	update_sw_cntrs(gpu);
726 }
727 
728 /* add bo's to gpu's ring, and kick gpu: */
729 void msm_gpu_submit(struct msm_gpu *gpu, struct msm_gem_submit *submit,
730 		struct msm_file_private *ctx)
731 {
732 	struct drm_device *dev = gpu->dev;
733 	struct msm_drm_private *priv = dev->dev_private;
734 	struct msm_ringbuffer *ring = submit->ring;
735 	int i;
736 
737 	WARN_ON(!mutex_is_locked(&dev->struct_mutex));
738 
739 	pm_runtime_get_sync(&gpu->pdev->dev);
740 
741 	msm_gpu_hw_init(gpu);
742 
743 	submit->seqno = ++ring->seqno;
744 
745 	list_add_tail(&submit->node, &ring->submits);
746 
747 	msm_rd_dump_submit(priv->rd, submit, NULL);
748 
749 	update_sw_cntrs(gpu);
750 
751 	for (i = 0; i < submit->nr_bos; i++) {
752 		struct msm_gem_object *msm_obj = submit->bos[i].obj;
753 		uint64_t iova;
754 
755 		/* can't happen yet.. but when we add 2d support we'll have
756 		 * to deal w/ cross-ring synchronization:
757 		 */
758 		WARN_ON(is_active(msm_obj) && (msm_obj->gpu != gpu));
759 
760 		/* submit takes a reference to the bo and iova until retired: */
761 		drm_gem_object_get(&msm_obj->base);
762 		msm_gem_get_and_pin_iova(&msm_obj->base, submit->aspace, &iova);
763 
764 		if (submit->bos[i].flags & MSM_SUBMIT_BO_WRITE)
765 			msm_gem_move_to_active(&msm_obj->base, gpu, true, submit->fence);
766 		else if (submit->bos[i].flags & MSM_SUBMIT_BO_READ)
767 			msm_gem_move_to_active(&msm_obj->base, gpu, false, submit->fence);
768 	}
769 
770 	gpu->funcs->submit(gpu, submit, ctx);
771 	priv->lastctx = ctx;
772 
773 	hangcheck_timer_reset(gpu);
774 }
775 
776 /*
777  * Init/Cleanup:
778  */
779 
780 static irqreturn_t irq_handler(int irq, void *data)
781 {
782 	struct msm_gpu *gpu = data;
783 	return gpu->funcs->irq(gpu);
784 }
785 
786 static int get_clocks(struct platform_device *pdev, struct msm_gpu *gpu)
787 {
788 	int ret = devm_clk_bulk_get_all(&pdev->dev, &gpu->grp_clks);
789 
790 	if (ret < 1) {
791 		gpu->nr_clocks = 0;
792 		return ret;
793 	}
794 
795 	gpu->nr_clocks = ret;
796 
797 	gpu->core_clk = msm_clk_bulk_get_clock(gpu->grp_clks,
798 		gpu->nr_clocks, "core");
799 
800 	gpu->rbbmtimer_clk = msm_clk_bulk_get_clock(gpu->grp_clks,
801 		gpu->nr_clocks, "rbbmtimer");
802 
803 	return 0;
804 }
805 
806 static struct msm_gem_address_space *
807 msm_gpu_create_address_space(struct msm_gpu *gpu, struct platform_device *pdev,
808 		uint64_t va_start, uint64_t va_end)
809 {
810 	struct msm_gem_address_space *aspace;
811 	int ret;
812 
813 	/*
814 	 * Setup IOMMU.. eventually we will (I think) do this once per context
815 	 * and have separate page tables per context.  For now, to keep things
816 	 * simple and to get something working, just use a single address space:
817 	 */
818 	if (!adreno_is_a2xx(to_adreno_gpu(gpu))) {
819 		struct iommu_domain *iommu = iommu_domain_alloc(&platform_bus_type);
820 		if (!iommu)
821 			return NULL;
822 
823 		iommu->geometry.aperture_start = va_start;
824 		iommu->geometry.aperture_end = va_end;
825 
826 		DRM_DEV_INFO(gpu->dev->dev, "%s: using IOMMU\n", gpu->name);
827 
828 		aspace = msm_gem_address_space_create(&pdev->dev, iommu, "gpu");
829 		if (IS_ERR(aspace))
830 			iommu_domain_free(iommu);
831 	} else {
832 		aspace = msm_gem_address_space_create_a2xx(&pdev->dev, gpu, "gpu",
833 			va_start, va_end);
834 	}
835 
836 	if (IS_ERR(aspace)) {
837 		DRM_DEV_ERROR(gpu->dev->dev, "failed to init mmu: %ld\n",
838 			PTR_ERR(aspace));
839 		return ERR_CAST(aspace);
840 	}
841 
842 	ret = aspace->mmu->funcs->attach(aspace->mmu);
843 	if (ret) {
844 		msm_gem_address_space_put(aspace);
845 		return ERR_PTR(ret);
846 	}
847 
848 	return aspace;
849 }
850 
851 int msm_gpu_init(struct drm_device *drm, struct platform_device *pdev,
852 		struct msm_gpu *gpu, const struct msm_gpu_funcs *funcs,
853 		const char *name, struct msm_gpu_config *config)
854 {
855 	int i, ret, nr_rings = config->nr_rings;
856 	void *memptrs;
857 	uint64_t memptrs_iova;
858 
859 	if (WARN_ON(gpu->num_perfcntrs > ARRAY_SIZE(gpu->last_cntrs)))
860 		gpu->num_perfcntrs = ARRAY_SIZE(gpu->last_cntrs);
861 
862 	gpu->dev = drm;
863 	gpu->funcs = funcs;
864 	gpu->name = name;
865 
866 	INIT_LIST_HEAD(&gpu->active_list);
867 	INIT_WORK(&gpu->retire_work, retire_worker);
868 	INIT_WORK(&gpu->recover_work, recover_worker);
869 
870 
871 	timer_setup(&gpu->hangcheck_timer, hangcheck_handler, 0);
872 
873 	spin_lock_init(&gpu->perf_lock);
874 
875 
876 	/* Map registers: */
877 	gpu->mmio = msm_ioremap(pdev, config->ioname, name);
878 	if (IS_ERR(gpu->mmio)) {
879 		ret = PTR_ERR(gpu->mmio);
880 		goto fail;
881 	}
882 
883 	/* Get Interrupt: */
884 	gpu->irq = platform_get_irq(pdev, 0);
885 	if (gpu->irq < 0) {
886 		ret = gpu->irq;
887 		DRM_DEV_ERROR(drm->dev, "failed to get irq: %d\n", ret);
888 		goto fail;
889 	}
890 
891 	ret = devm_request_irq(&pdev->dev, gpu->irq, irq_handler,
892 			IRQF_TRIGGER_HIGH, gpu->name, gpu);
893 	if (ret) {
894 		DRM_DEV_ERROR(drm->dev, "failed to request IRQ%u: %d\n", gpu->irq, ret);
895 		goto fail;
896 	}
897 
898 	ret = get_clocks(pdev, gpu);
899 	if (ret)
900 		goto fail;
901 
902 	gpu->ebi1_clk = msm_clk_get(pdev, "bus");
903 	DBG("ebi1_clk: %p", gpu->ebi1_clk);
904 	if (IS_ERR(gpu->ebi1_clk))
905 		gpu->ebi1_clk = NULL;
906 
907 	/* Acquire regulators: */
908 	gpu->gpu_reg = devm_regulator_get(&pdev->dev, "vdd");
909 	DBG("gpu_reg: %p", gpu->gpu_reg);
910 	if (IS_ERR(gpu->gpu_reg))
911 		gpu->gpu_reg = NULL;
912 
913 	gpu->gpu_cx = devm_regulator_get(&pdev->dev, "vddcx");
914 	DBG("gpu_cx: %p", gpu->gpu_cx);
915 	if (IS_ERR(gpu->gpu_cx))
916 		gpu->gpu_cx = NULL;
917 
918 	gpu->pdev = pdev;
919 	platform_set_drvdata(pdev, gpu);
920 
921 	msm_devfreq_init(gpu);
922 
923 	gpu->aspace = msm_gpu_create_address_space(gpu, pdev,
924 		config->va_start, config->va_end);
925 
926 	if (gpu->aspace == NULL)
927 		DRM_DEV_INFO(drm->dev, "%s: no IOMMU, fallback to VRAM carveout!\n", name);
928 	else if (IS_ERR(gpu->aspace)) {
929 		ret = PTR_ERR(gpu->aspace);
930 		goto fail;
931 	}
932 
933 	memptrs = msm_gem_kernel_new(drm,
934 		sizeof(struct msm_rbmemptrs) * nr_rings,
935 		MSM_BO_UNCACHED, gpu->aspace, &gpu->memptrs_bo,
936 		&memptrs_iova);
937 
938 	if (IS_ERR(memptrs)) {
939 		ret = PTR_ERR(memptrs);
940 		DRM_DEV_ERROR(drm->dev, "could not allocate memptrs: %d\n", ret);
941 		goto fail;
942 	}
943 
944 	msm_gem_object_set_name(gpu->memptrs_bo, "memptrs");
945 
946 	if (nr_rings > ARRAY_SIZE(gpu->rb)) {
947 		DRM_DEV_INFO_ONCE(drm->dev, "Only creating %zu ringbuffers\n",
948 			ARRAY_SIZE(gpu->rb));
949 		nr_rings = ARRAY_SIZE(gpu->rb);
950 	}
951 
952 	/* Create ringbuffer(s): */
953 	for (i = 0; i < nr_rings; i++) {
954 		gpu->rb[i] = msm_ringbuffer_new(gpu, i, memptrs, memptrs_iova);
955 
956 		if (IS_ERR(gpu->rb[i])) {
957 			ret = PTR_ERR(gpu->rb[i]);
958 			DRM_DEV_ERROR(drm->dev,
959 				"could not create ringbuffer %d: %d\n", i, ret);
960 			goto fail;
961 		}
962 
963 		memptrs += sizeof(struct msm_rbmemptrs);
964 		memptrs_iova += sizeof(struct msm_rbmemptrs);
965 	}
966 
967 	gpu->nr_rings = nr_rings;
968 
969 	return 0;
970 
971 fail:
972 	for (i = 0; i < ARRAY_SIZE(gpu->rb); i++)  {
973 		msm_ringbuffer_destroy(gpu->rb[i]);
974 		gpu->rb[i] = NULL;
975 	}
976 
977 	msm_gem_kernel_put(gpu->memptrs_bo, gpu->aspace, false);
978 
979 	platform_set_drvdata(pdev, NULL);
980 	return ret;
981 }
982 
983 void msm_gpu_cleanup(struct msm_gpu *gpu)
984 {
985 	int i;
986 
987 	DBG("%s", gpu->name);
988 
989 	WARN_ON(!list_empty(&gpu->active_list));
990 
991 	for (i = 0; i < ARRAY_SIZE(gpu->rb); i++) {
992 		msm_ringbuffer_destroy(gpu->rb[i]);
993 		gpu->rb[i] = NULL;
994 	}
995 
996 	msm_gem_kernel_put(gpu->memptrs_bo, gpu->aspace, false);
997 
998 	if (!IS_ERR_OR_NULL(gpu->aspace)) {
999 		gpu->aspace->mmu->funcs->detach(gpu->aspace->mmu);
1000 		msm_gem_address_space_put(gpu->aspace);
1001 	}
1002 }
1003