xref: /linux/drivers/gpu/drm/msm/msm_gpu.c (revision b0f84a84fff180718995b1269da2988e5b28be42)
1 /*
2  * Copyright (C) 2013 Red Hat
3  * Author: Rob Clark <robdclark@gmail.com>
4  *
5  * This program is free software; you can redistribute it and/or modify it
6  * under the terms of the GNU General Public License version 2 as published by
7  * the Free Software Foundation.
8  *
9  * This program is distributed in the hope that it will be useful, but WITHOUT
10  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
12  * more details.
13  *
14  * You should have received a copy of the GNU General Public License along with
15  * this program.  If not, see <http://www.gnu.org/licenses/>.
16  */
17 
18 #include "msm_gpu.h"
19 #include "msm_gem.h"
20 #include "msm_mmu.h"
21 #include "msm_fence.h"
22 #include "msm_gpu_trace.h"
23 #include "adreno/adreno_gpu.h"
24 
25 #include <generated/utsrelease.h>
26 #include <linux/string_helpers.h>
27 #include <linux/pm_opp.h>
28 #include <linux/devfreq.h>
29 #include <linux/devcoredump.h>
30 
31 /*
32  * Power Management:
33  */
34 
35 static int msm_devfreq_target(struct device *dev, unsigned long *freq,
36 		u32 flags)
37 {
38 	struct msm_gpu *gpu = platform_get_drvdata(to_platform_device(dev));
39 	struct dev_pm_opp *opp;
40 
41 	opp = devfreq_recommended_opp(dev, freq, flags);
42 
43 	if (IS_ERR(opp))
44 		return PTR_ERR(opp);
45 
46 	if (gpu->funcs->gpu_set_freq)
47 		gpu->funcs->gpu_set_freq(gpu, (u64)*freq);
48 	else
49 		clk_set_rate(gpu->core_clk, *freq);
50 
51 	dev_pm_opp_put(opp);
52 
53 	return 0;
54 }
55 
56 static int msm_devfreq_get_dev_status(struct device *dev,
57 		struct devfreq_dev_status *status)
58 {
59 	struct msm_gpu *gpu = platform_get_drvdata(to_platform_device(dev));
60 	ktime_t time;
61 
62 	if (gpu->funcs->gpu_get_freq)
63 		status->current_frequency = gpu->funcs->gpu_get_freq(gpu);
64 	else
65 		status->current_frequency = clk_get_rate(gpu->core_clk);
66 
67 	status->busy_time = gpu->funcs->gpu_busy(gpu);
68 
69 	time = ktime_get();
70 	status->total_time = ktime_us_delta(time, gpu->devfreq.time);
71 	gpu->devfreq.time = time;
72 
73 	return 0;
74 }
75 
76 static int msm_devfreq_get_cur_freq(struct device *dev, unsigned long *freq)
77 {
78 	struct msm_gpu *gpu = platform_get_drvdata(to_platform_device(dev));
79 
80 	if (gpu->funcs->gpu_get_freq)
81 		*freq = gpu->funcs->gpu_get_freq(gpu);
82 	else
83 		*freq = clk_get_rate(gpu->core_clk);
84 
85 	return 0;
86 }
87 
88 static struct devfreq_dev_profile msm_devfreq_profile = {
89 	.polling_ms = 10,
90 	.target = msm_devfreq_target,
91 	.get_dev_status = msm_devfreq_get_dev_status,
92 	.get_cur_freq = msm_devfreq_get_cur_freq,
93 };
94 
95 static void msm_devfreq_init(struct msm_gpu *gpu)
96 {
97 	/* We need target support to do devfreq */
98 	if (!gpu->funcs->gpu_busy)
99 		return;
100 
101 	msm_devfreq_profile.initial_freq = gpu->fast_rate;
102 
103 	/*
104 	 * Don't set the freq_table or max_state and let devfreq build the table
105 	 * from OPP
106 	 */
107 
108 	gpu->devfreq.devfreq = devm_devfreq_add_device(&gpu->pdev->dev,
109 			&msm_devfreq_profile, "simple_ondemand", NULL);
110 
111 	if (IS_ERR(gpu->devfreq.devfreq)) {
112 		DRM_DEV_ERROR(&gpu->pdev->dev, "Couldn't initialize GPU devfreq\n");
113 		gpu->devfreq.devfreq = NULL;
114 	}
115 
116 	devfreq_suspend_device(gpu->devfreq.devfreq);
117 }
118 
119 static int enable_pwrrail(struct msm_gpu *gpu)
120 {
121 	struct drm_device *dev = gpu->dev;
122 	int ret = 0;
123 
124 	if (gpu->gpu_reg) {
125 		ret = regulator_enable(gpu->gpu_reg);
126 		if (ret) {
127 			DRM_DEV_ERROR(dev->dev, "failed to enable 'gpu_reg': %d\n", ret);
128 			return ret;
129 		}
130 	}
131 
132 	if (gpu->gpu_cx) {
133 		ret = regulator_enable(gpu->gpu_cx);
134 		if (ret) {
135 			DRM_DEV_ERROR(dev->dev, "failed to enable 'gpu_cx': %d\n", ret);
136 			return ret;
137 		}
138 	}
139 
140 	return 0;
141 }
142 
143 static int disable_pwrrail(struct msm_gpu *gpu)
144 {
145 	if (gpu->gpu_cx)
146 		regulator_disable(gpu->gpu_cx);
147 	if (gpu->gpu_reg)
148 		regulator_disable(gpu->gpu_reg);
149 	return 0;
150 }
151 
152 static int enable_clk(struct msm_gpu *gpu)
153 {
154 	if (gpu->core_clk && gpu->fast_rate)
155 		clk_set_rate(gpu->core_clk, gpu->fast_rate);
156 
157 	/* Set the RBBM timer rate to 19.2Mhz */
158 	if (gpu->rbbmtimer_clk)
159 		clk_set_rate(gpu->rbbmtimer_clk, 19200000);
160 
161 	return clk_bulk_prepare_enable(gpu->nr_clocks, gpu->grp_clks);
162 }
163 
164 static int disable_clk(struct msm_gpu *gpu)
165 {
166 	clk_bulk_disable_unprepare(gpu->nr_clocks, gpu->grp_clks);
167 
168 	/*
169 	 * Set the clock to a deliberately low rate. On older targets the clock
170 	 * speed had to be non zero to avoid problems. On newer targets this
171 	 * will be rounded down to zero anyway so it all works out.
172 	 */
173 	if (gpu->core_clk)
174 		clk_set_rate(gpu->core_clk, 27000000);
175 
176 	if (gpu->rbbmtimer_clk)
177 		clk_set_rate(gpu->rbbmtimer_clk, 0);
178 
179 	return 0;
180 }
181 
182 static int enable_axi(struct msm_gpu *gpu)
183 {
184 	if (gpu->ebi1_clk)
185 		clk_prepare_enable(gpu->ebi1_clk);
186 	return 0;
187 }
188 
189 static int disable_axi(struct msm_gpu *gpu)
190 {
191 	if (gpu->ebi1_clk)
192 		clk_disable_unprepare(gpu->ebi1_clk);
193 	return 0;
194 }
195 
196 void msm_gpu_resume_devfreq(struct msm_gpu *gpu)
197 {
198 	gpu->devfreq.busy_cycles = 0;
199 	gpu->devfreq.time = ktime_get();
200 
201 	devfreq_resume_device(gpu->devfreq.devfreq);
202 }
203 
204 int msm_gpu_pm_resume(struct msm_gpu *gpu)
205 {
206 	int ret;
207 
208 	DBG("%s", gpu->name);
209 
210 	ret = enable_pwrrail(gpu);
211 	if (ret)
212 		return ret;
213 
214 	ret = enable_clk(gpu);
215 	if (ret)
216 		return ret;
217 
218 	ret = enable_axi(gpu);
219 	if (ret)
220 		return ret;
221 
222 	msm_gpu_resume_devfreq(gpu);
223 
224 	gpu->needs_hw_init = true;
225 
226 	return 0;
227 }
228 
229 int msm_gpu_pm_suspend(struct msm_gpu *gpu)
230 {
231 	int ret;
232 
233 	DBG("%s", gpu->name);
234 
235 	devfreq_suspend_device(gpu->devfreq.devfreq);
236 
237 	ret = disable_axi(gpu);
238 	if (ret)
239 		return ret;
240 
241 	ret = disable_clk(gpu);
242 	if (ret)
243 		return ret;
244 
245 	ret = disable_pwrrail(gpu);
246 	if (ret)
247 		return ret;
248 
249 	return 0;
250 }
251 
252 int msm_gpu_hw_init(struct msm_gpu *gpu)
253 {
254 	int ret;
255 
256 	WARN_ON(!mutex_is_locked(&gpu->dev->struct_mutex));
257 
258 	if (!gpu->needs_hw_init)
259 		return 0;
260 
261 	disable_irq(gpu->irq);
262 	ret = gpu->funcs->hw_init(gpu);
263 	if (!ret)
264 		gpu->needs_hw_init = false;
265 	enable_irq(gpu->irq);
266 
267 	return ret;
268 }
269 
270 #ifdef CONFIG_DEV_COREDUMP
271 static ssize_t msm_gpu_devcoredump_read(char *buffer, loff_t offset,
272 		size_t count, void *data, size_t datalen)
273 {
274 	struct msm_gpu *gpu = data;
275 	struct drm_print_iterator iter;
276 	struct drm_printer p;
277 	struct msm_gpu_state *state;
278 
279 	state = msm_gpu_crashstate_get(gpu);
280 	if (!state)
281 		return 0;
282 
283 	iter.data = buffer;
284 	iter.offset = 0;
285 	iter.start = offset;
286 	iter.remain = count;
287 
288 	p = drm_coredump_printer(&iter);
289 
290 	drm_printf(&p, "---\n");
291 	drm_printf(&p, "kernel: " UTS_RELEASE "\n");
292 	drm_printf(&p, "module: " KBUILD_MODNAME "\n");
293 	drm_printf(&p, "time: %lld.%09ld\n",
294 		state->time.tv_sec, state->time.tv_nsec);
295 	if (state->comm)
296 		drm_printf(&p, "comm: %s\n", state->comm);
297 	if (state->cmd)
298 		drm_printf(&p, "cmdline: %s\n", state->cmd);
299 
300 	gpu->funcs->show(gpu, state, &p);
301 
302 	msm_gpu_crashstate_put(gpu);
303 
304 	return count - iter.remain;
305 }
306 
307 static void msm_gpu_devcoredump_free(void *data)
308 {
309 	struct msm_gpu *gpu = data;
310 
311 	msm_gpu_crashstate_put(gpu);
312 }
313 
314 static void msm_gpu_crashstate_get_bo(struct msm_gpu_state *state,
315 		struct msm_gem_object *obj, u64 iova, u32 flags)
316 {
317 	struct msm_gpu_state_bo *state_bo = &state->bos[state->nr_bos];
318 
319 	/* Don't record write only objects */
320 	state_bo->size = obj->base.size;
321 	state_bo->iova = iova;
322 
323 	/* Only store data for non imported buffer objects marked for read */
324 	if ((flags & MSM_SUBMIT_BO_READ) && !obj->base.import_attach) {
325 		void *ptr;
326 
327 		state_bo->data = kvmalloc(obj->base.size, GFP_KERNEL);
328 		if (!state_bo->data)
329 			goto out;
330 
331 		ptr = msm_gem_get_vaddr_active(&obj->base);
332 		if (IS_ERR(ptr)) {
333 			kvfree(state_bo->data);
334 			state_bo->data = NULL;
335 			goto out;
336 		}
337 
338 		memcpy(state_bo->data, ptr, obj->base.size);
339 		msm_gem_put_vaddr(&obj->base);
340 	}
341 out:
342 	state->nr_bos++;
343 }
344 
345 static void msm_gpu_crashstate_capture(struct msm_gpu *gpu,
346 		struct msm_gem_submit *submit, char *comm, char *cmd)
347 {
348 	struct msm_gpu_state *state;
349 
350 	/* Check if the target supports capturing crash state */
351 	if (!gpu->funcs->gpu_state_get)
352 		return;
353 
354 	/* Only save one crash state at a time */
355 	if (gpu->crashstate)
356 		return;
357 
358 	state = gpu->funcs->gpu_state_get(gpu);
359 	if (IS_ERR_OR_NULL(state))
360 		return;
361 
362 	/* Fill in the additional crash state information */
363 	state->comm = kstrdup(comm, GFP_KERNEL);
364 	state->cmd = kstrdup(cmd, GFP_KERNEL);
365 
366 	if (submit) {
367 		int i;
368 
369 		state->bos = kcalloc(submit->nr_cmds,
370 			sizeof(struct msm_gpu_state_bo), GFP_KERNEL);
371 
372 		for (i = 0; state->bos && i < submit->nr_cmds; i++) {
373 			int idx = submit->cmd[i].idx;
374 
375 			msm_gpu_crashstate_get_bo(state, submit->bos[idx].obj,
376 				submit->bos[idx].iova, submit->bos[idx].flags);
377 		}
378 	}
379 
380 	/* Set the active crash state to be dumped on failure */
381 	gpu->crashstate = state;
382 
383 	/* FIXME: Release the crashstate if this errors out? */
384 	dev_coredumpm(gpu->dev->dev, THIS_MODULE, gpu, 0, GFP_KERNEL,
385 		msm_gpu_devcoredump_read, msm_gpu_devcoredump_free);
386 }
387 #else
388 static void msm_gpu_crashstate_capture(struct msm_gpu *gpu,
389 		struct msm_gem_submit *submit, char *comm, char *cmd)
390 {
391 }
392 #endif
393 
394 /*
395  * Hangcheck detection for locked gpu:
396  */
397 
398 static void update_fences(struct msm_gpu *gpu, struct msm_ringbuffer *ring,
399 		uint32_t fence)
400 {
401 	struct msm_gem_submit *submit;
402 
403 	list_for_each_entry(submit, &ring->submits, node) {
404 		if (submit->seqno > fence)
405 			break;
406 
407 		msm_update_fence(submit->ring->fctx,
408 			submit->fence->seqno);
409 	}
410 }
411 
412 static struct msm_gem_submit *
413 find_submit(struct msm_ringbuffer *ring, uint32_t fence)
414 {
415 	struct msm_gem_submit *submit;
416 
417 	WARN_ON(!mutex_is_locked(&ring->gpu->dev->struct_mutex));
418 
419 	list_for_each_entry(submit, &ring->submits, node)
420 		if (submit->seqno == fence)
421 			return submit;
422 
423 	return NULL;
424 }
425 
426 static void retire_submits(struct msm_gpu *gpu);
427 
428 static void recover_worker(struct work_struct *work)
429 {
430 	struct msm_gpu *gpu = container_of(work, struct msm_gpu, recover_work);
431 	struct drm_device *dev = gpu->dev;
432 	struct msm_drm_private *priv = dev->dev_private;
433 	struct msm_gem_submit *submit;
434 	struct msm_ringbuffer *cur_ring = gpu->funcs->active_ring(gpu);
435 	char *comm = NULL, *cmd = NULL;
436 	int i;
437 
438 	mutex_lock(&dev->struct_mutex);
439 
440 	DRM_DEV_ERROR(dev->dev, "%s: hangcheck recover!\n", gpu->name);
441 
442 	submit = find_submit(cur_ring, cur_ring->memptrs->fence + 1);
443 	if (submit) {
444 		struct task_struct *task;
445 
446 		/* Increment the fault counts */
447 		gpu->global_faults++;
448 		submit->queue->faults++;
449 
450 		task = get_pid_task(submit->pid, PIDTYPE_PID);
451 		if (task) {
452 			comm = kstrdup(task->comm, GFP_KERNEL);
453 			cmd = kstrdup_quotable_cmdline(task, GFP_KERNEL);
454 			put_task_struct(task);
455 		}
456 
457 		if (comm && cmd) {
458 			DRM_DEV_ERROR(dev->dev, "%s: offending task: %s (%s)\n",
459 				gpu->name, comm, cmd);
460 
461 			msm_rd_dump_submit(priv->hangrd, submit,
462 				"offending task: %s (%s)", comm, cmd);
463 		} else
464 			msm_rd_dump_submit(priv->hangrd, submit, NULL);
465 	}
466 
467 	/* Record the crash state */
468 	pm_runtime_get_sync(&gpu->pdev->dev);
469 	msm_gpu_crashstate_capture(gpu, submit, comm, cmd);
470 	pm_runtime_put_sync(&gpu->pdev->dev);
471 
472 	kfree(cmd);
473 	kfree(comm);
474 
475 	/*
476 	 * Update all the rings with the latest and greatest fence.. this
477 	 * needs to happen after msm_rd_dump_submit() to ensure that the
478 	 * bo's referenced by the offending submit are still around.
479 	 */
480 	for (i = 0; i < gpu->nr_rings; i++) {
481 		struct msm_ringbuffer *ring = gpu->rb[i];
482 
483 		uint32_t fence = ring->memptrs->fence;
484 
485 		/*
486 		 * For the current (faulting?) ring/submit advance the fence by
487 		 * one more to clear the faulting submit
488 		 */
489 		if (ring == cur_ring)
490 			fence++;
491 
492 		update_fences(gpu, ring, fence);
493 	}
494 
495 	if (msm_gpu_active(gpu)) {
496 		/* retire completed submits, plus the one that hung: */
497 		retire_submits(gpu);
498 
499 		pm_runtime_get_sync(&gpu->pdev->dev);
500 		gpu->funcs->recover(gpu);
501 		pm_runtime_put_sync(&gpu->pdev->dev);
502 
503 		/*
504 		 * Replay all remaining submits starting with highest priority
505 		 * ring
506 		 */
507 		for (i = 0; i < gpu->nr_rings; i++) {
508 			struct msm_ringbuffer *ring = gpu->rb[i];
509 
510 			list_for_each_entry(submit, &ring->submits, node)
511 				gpu->funcs->submit(gpu, submit, NULL);
512 		}
513 	}
514 
515 	mutex_unlock(&dev->struct_mutex);
516 
517 	msm_gpu_retire(gpu);
518 }
519 
520 static void hangcheck_timer_reset(struct msm_gpu *gpu)
521 {
522 	DBG("%s", gpu->name);
523 	mod_timer(&gpu->hangcheck_timer,
524 			round_jiffies_up(jiffies + DRM_MSM_HANGCHECK_JIFFIES));
525 }
526 
527 static void hangcheck_handler(struct timer_list *t)
528 {
529 	struct msm_gpu *gpu = from_timer(gpu, t, hangcheck_timer);
530 	struct drm_device *dev = gpu->dev;
531 	struct msm_drm_private *priv = dev->dev_private;
532 	struct msm_ringbuffer *ring = gpu->funcs->active_ring(gpu);
533 	uint32_t fence = ring->memptrs->fence;
534 
535 	if (fence != ring->hangcheck_fence) {
536 		/* some progress has been made.. ya! */
537 		ring->hangcheck_fence = fence;
538 	} else if (fence < ring->seqno) {
539 		/* no progress and not done.. hung! */
540 		ring->hangcheck_fence = fence;
541 		DRM_DEV_ERROR(dev->dev, "%s: hangcheck detected gpu lockup rb %d!\n",
542 				gpu->name, ring->id);
543 		DRM_DEV_ERROR(dev->dev, "%s:     completed fence: %u\n",
544 				gpu->name, fence);
545 		DRM_DEV_ERROR(dev->dev, "%s:     submitted fence: %u\n",
546 				gpu->name, ring->seqno);
547 
548 		queue_work(priv->wq, &gpu->recover_work);
549 	}
550 
551 	/* if still more pending work, reset the hangcheck timer: */
552 	if (ring->seqno > ring->hangcheck_fence)
553 		hangcheck_timer_reset(gpu);
554 
555 	/* workaround for missing irq: */
556 	queue_work(priv->wq, &gpu->retire_work);
557 }
558 
559 /*
560  * Performance Counters:
561  */
562 
563 /* called under perf_lock */
564 static int update_hw_cntrs(struct msm_gpu *gpu, uint32_t ncntrs, uint32_t *cntrs)
565 {
566 	uint32_t current_cntrs[ARRAY_SIZE(gpu->last_cntrs)];
567 	int i, n = min(ncntrs, gpu->num_perfcntrs);
568 
569 	/* read current values: */
570 	for (i = 0; i < gpu->num_perfcntrs; i++)
571 		current_cntrs[i] = gpu_read(gpu, gpu->perfcntrs[i].sample_reg);
572 
573 	/* update cntrs: */
574 	for (i = 0; i < n; i++)
575 		cntrs[i] = current_cntrs[i] - gpu->last_cntrs[i];
576 
577 	/* save current values: */
578 	for (i = 0; i < gpu->num_perfcntrs; i++)
579 		gpu->last_cntrs[i] = current_cntrs[i];
580 
581 	return n;
582 }
583 
584 static void update_sw_cntrs(struct msm_gpu *gpu)
585 {
586 	ktime_t time;
587 	uint32_t elapsed;
588 	unsigned long flags;
589 
590 	spin_lock_irqsave(&gpu->perf_lock, flags);
591 	if (!gpu->perfcntr_active)
592 		goto out;
593 
594 	time = ktime_get();
595 	elapsed = ktime_to_us(ktime_sub(time, gpu->last_sample.time));
596 
597 	gpu->totaltime += elapsed;
598 	if (gpu->last_sample.active)
599 		gpu->activetime += elapsed;
600 
601 	gpu->last_sample.active = msm_gpu_active(gpu);
602 	gpu->last_sample.time = time;
603 
604 out:
605 	spin_unlock_irqrestore(&gpu->perf_lock, flags);
606 }
607 
608 void msm_gpu_perfcntr_start(struct msm_gpu *gpu)
609 {
610 	unsigned long flags;
611 
612 	pm_runtime_get_sync(&gpu->pdev->dev);
613 
614 	spin_lock_irqsave(&gpu->perf_lock, flags);
615 	/* we could dynamically enable/disable perfcntr registers too.. */
616 	gpu->last_sample.active = msm_gpu_active(gpu);
617 	gpu->last_sample.time = ktime_get();
618 	gpu->activetime = gpu->totaltime = 0;
619 	gpu->perfcntr_active = true;
620 	update_hw_cntrs(gpu, 0, NULL);
621 	spin_unlock_irqrestore(&gpu->perf_lock, flags);
622 }
623 
624 void msm_gpu_perfcntr_stop(struct msm_gpu *gpu)
625 {
626 	gpu->perfcntr_active = false;
627 	pm_runtime_put_sync(&gpu->pdev->dev);
628 }
629 
630 /* returns -errno or # of cntrs sampled */
631 int msm_gpu_perfcntr_sample(struct msm_gpu *gpu, uint32_t *activetime,
632 		uint32_t *totaltime, uint32_t ncntrs, uint32_t *cntrs)
633 {
634 	unsigned long flags;
635 	int ret;
636 
637 	spin_lock_irqsave(&gpu->perf_lock, flags);
638 
639 	if (!gpu->perfcntr_active) {
640 		ret = -EINVAL;
641 		goto out;
642 	}
643 
644 	*activetime = gpu->activetime;
645 	*totaltime = gpu->totaltime;
646 
647 	gpu->activetime = gpu->totaltime = 0;
648 
649 	ret = update_hw_cntrs(gpu, ncntrs, cntrs);
650 
651 out:
652 	spin_unlock_irqrestore(&gpu->perf_lock, flags);
653 
654 	return ret;
655 }
656 
657 /*
658  * Cmdstream submission/retirement:
659  */
660 
661 static void retire_submit(struct msm_gpu *gpu, struct msm_ringbuffer *ring,
662 		struct msm_gem_submit *submit)
663 {
664 	int index = submit->seqno % MSM_GPU_SUBMIT_STATS_COUNT;
665 	volatile struct msm_gpu_submit_stats *stats;
666 	u64 elapsed, clock = 0;
667 	int i;
668 
669 	stats = &ring->memptrs->stats[index];
670 	/* Convert 19.2Mhz alwayson ticks to nanoseconds for elapsed time */
671 	elapsed = (stats->alwayson_end - stats->alwayson_start) * 10000;
672 	do_div(elapsed, 192);
673 
674 	/* Calculate the clock frequency from the number of CP cycles */
675 	if (elapsed) {
676 		clock = (stats->cpcycles_end - stats->cpcycles_start) * 1000;
677 		do_div(clock, elapsed);
678 	}
679 
680 	trace_msm_gpu_submit_retired(submit, elapsed, clock,
681 		stats->alwayson_start, stats->alwayson_end);
682 
683 	for (i = 0; i < submit->nr_bos; i++) {
684 		struct msm_gem_object *msm_obj = submit->bos[i].obj;
685 		/* move to inactive: */
686 		msm_gem_move_to_inactive(&msm_obj->base);
687 		msm_gem_unpin_iova(&msm_obj->base, submit->aspace);
688 		drm_gem_object_put(&msm_obj->base);
689 	}
690 
691 	pm_runtime_mark_last_busy(&gpu->pdev->dev);
692 	pm_runtime_put_autosuspend(&gpu->pdev->dev);
693 	msm_gem_submit_free(submit);
694 }
695 
696 static void retire_submits(struct msm_gpu *gpu)
697 {
698 	struct drm_device *dev = gpu->dev;
699 	struct msm_gem_submit *submit, *tmp;
700 	int i;
701 
702 	WARN_ON(!mutex_is_locked(&dev->struct_mutex));
703 
704 	/* Retire the commits starting with highest priority */
705 	for (i = 0; i < gpu->nr_rings; i++) {
706 		struct msm_ringbuffer *ring = gpu->rb[i];
707 
708 		list_for_each_entry_safe(submit, tmp, &ring->submits, node) {
709 			if (dma_fence_is_signaled(submit->fence))
710 				retire_submit(gpu, ring, submit);
711 		}
712 	}
713 }
714 
715 static void retire_worker(struct work_struct *work)
716 {
717 	struct msm_gpu *gpu = container_of(work, struct msm_gpu, retire_work);
718 	struct drm_device *dev = gpu->dev;
719 	int i;
720 
721 	for (i = 0; i < gpu->nr_rings; i++)
722 		update_fences(gpu, gpu->rb[i], gpu->rb[i]->memptrs->fence);
723 
724 	mutex_lock(&dev->struct_mutex);
725 	retire_submits(gpu);
726 	mutex_unlock(&dev->struct_mutex);
727 }
728 
729 /* call from irq handler to schedule work to retire bo's */
730 void msm_gpu_retire(struct msm_gpu *gpu)
731 {
732 	struct msm_drm_private *priv = gpu->dev->dev_private;
733 	queue_work(priv->wq, &gpu->retire_work);
734 	update_sw_cntrs(gpu);
735 }
736 
737 /* add bo's to gpu's ring, and kick gpu: */
738 void msm_gpu_submit(struct msm_gpu *gpu, struct msm_gem_submit *submit,
739 		struct msm_file_private *ctx)
740 {
741 	struct drm_device *dev = gpu->dev;
742 	struct msm_drm_private *priv = dev->dev_private;
743 	struct msm_ringbuffer *ring = submit->ring;
744 	int i;
745 
746 	WARN_ON(!mutex_is_locked(&dev->struct_mutex));
747 
748 	pm_runtime_get_sync(&gpu->pdev->dev);
749 
750 	msm_gpu_hw_init(gpu);
751 
752 	submit->seqno = ++ring->seqno;
753 
754 	list_add_tail(&submit->node, &ring->submits);
755 
756 	msm_rd_dump_submit(priv->rd, submit, NULL);
757 
758 	update_sw_cntrs(gpu);
759 
760 	for (i = 0; i < submit->nr_bos; i++) {
761 		struct msm_gem_object *msm_obj = submit->bos[i].obj;
762 		uint64_t iova;
763 
764 		/* can't happen yet.. but when we add 2d support we'll have
765 		 * to deal w/ cross-ring synchronization:
766 		 */
767 		WARN_ON(is_active(msm_obj) && (msm_obj->gpu != gpu));
768 
769 		/* submit takes a reference to the bo and iova until retired: */
770 		drm_gem_object_get(&msm_obj->base);
771 		msm_gem_get_and_pin_iova(&msm_obj->base, submit->aspace, &iova);
772 
773 		if (submit->bos[i].flags & MSM_SUBMIT_BO_WRITE)
774 			msm_gem_move_to_active(&msm_obj->base, gpu, true, submit->fence);
775 		else if (submit->bos[i].flags & MSM_SUBMIT_BO_READ)
776 			msm_gem_move_to_active(&msm_obj->base, gpu, false, submit->fence);
777 	}
778 
779 	gpu->funcs->submit(gpu, submit, ctx);
780 	priv->lastctx = ctx;
781 
782 	hangcheck_timer_reset(gpu);
783 }
784 
785 /*
786  * Init/Cleanup:
787  */
788 
789 static irqreturn_t irq_handler(int irq, void *data)
790 {
791 	struct msm_gpu *gpu = data;
792 	return gpu->funcs->irq(gpu);
793 }
794 
795 static int get_clocks(struct platform_device *pdev, struct msm_gpu *gpu)
796 {
797 	int ret = msm_clk_bulk_get(&pdev->dev, &gpu->grp_clks);
798 
799 	if (ret < 1) {
800 		gpu->nr_clocks = 0;
801 		return ret;
802 	}
803 
804 	gpu->nr_clocks = ret;
805 
806 	gpu->core_clk = msm_clk_bulk_get_clock(gpu->grp_clks,
807 		gpu->nr_clocks, "core");
808 
809 	gpu->rbbmtimer_clk = msm_clk_bulk_get_clock(gpu->grp_clks,
810 		gpu->nr_clocks, "rbbmtimer");
811 
812 	return 0;
813 }
814 
815 static struct msm_gem_address_space *
816 msm_gpu_create_address_space(struct msm_gpu *gpu, struct platform_device *pdev,
817 		uint64_t va_start, uint64_t va_end)
818 {
819 	struct msm_gem_address_space *aspace;
820 	int ret;
821 
822 	/*
823 	 * Setup IOMMU.. eventually we will (I think) do this once per context
824 	 * and have separate page tables per context.  For now, to keep things
825 	 * simple and to get something working, just use a single address space:
826 	 */
827 	if (!adreno_is_a2xx(to_adreno_gpu(gpu))) {
828 		struct iommu_domain *iommu = iommu_domain_alloc(&platform_bus_type);
829 		if (!iommu)
830 			return NULL;
831 
832 		iommu->geometry.aperture_start = va_start;
833 		iommu->geometry.aperture_end = va_end;
834 
835 		DRM_DEV_INFO(gpu->dev->dev, "%s: using IOMMU\n", gpu->name);
836 
837 		aspace = msm_gem_address_space_create(&pdev->dev, iommu, "gpu");
838 		if (IS_ERR(aspace))
839 			iommu_domain_free(iommu);
840 	} else {
841 		aspace = msm_gem_address_space_create_a2xx(&pdev->dev, gpu, "gpu",
842 			va_start, va_end);
843 	}
844 
845 	if (IS_ERR(aspace)) {
846 		DRM_DEV_ERROR(gpu->dev->dev, "failed to init mmu: %ld\n",
847 			PTR_ERR(aspace));
848 		return ERR_CAST(aspace);
849 	}
850 
851 	ret = aspace->mmu->funcs->attach(aspace->mmu, NULL, 0);
852 	if (ret) {
853 		msm_gem_address_space_put(aspace);
854 		return ERR_PTR(ret);
855 	}
856 
857 	return aspace;
858 }
859 
860 int msm_gpu_init(struct drm_device *drm, struct platform_device *pdev,
861 		struct msm_gpu *gpu, const struct msm_gpu_funcs *funcs,
862 		const char *name, struct msm_gpu_config *config)
863 {
864 	int i, ret, nr_rings = config->nr_rings;
865 	void *memptrs;
866 	uint64_t memptrs_iova;
867 
868 	if (WARN_ON(gpu->num_perfcntrs > ARRAY_SIZE(gpu->last_cntrs)))
869 		gpu->num_perfcntrs = ARRAY_SIZE(gpu->last_cntrs);
870 
871 	gpu->dev = drm;
872 	gpu->funcs = funcs;
873 	gpu->name = name;
874 
875 	INIT_LIST_HEAD(&gpu->active_list);
876 	INIT_WORK(&gpu->retire_work, retire_worker);
877 	INIT_WORK(&gpu->recover_work, recover_worker);
878 
879 
880 	timer_setup(&gpu->hangcheck_timer, hangcheck_handler, 0);
881 
882 	spin_lock_init(&gpu->perf_lock);
883 
884 
885 	/* Map registers: */
886 	gpu->mmio = msm_ioremap(pdev, config->ioname, name);
887 	if (IS_ERR(gpu->mmio)) {
888 		ret = PTR_ERR(gpu->mmio);
889 		goto fail;
890 	}
891 
892 	/* Get Interrupt: */
893 	gpu->irq = platform_get_irq(pdev, 0);
894 	if (gpu->irq < 0) {
895 		ret = gpu->irq;
896 		DRM_DEV_ERROR(drm->dev, "failed to get irq: %d\n", ret);
897 		goto fail;
898 	}
899 
900 	ret = devm_request_irq(&pdev->dev, gpu->irq, irq_handler,
901 			IRQF_TRIGGER_HIGH, gpu->name, gpu);
902 	if (ret) {
903 		DRM_DEV_ERROR(drm->dev, "failed to request IRQ%u: %d\n", gpu->irq, ret);
904 		goto fail;
905 	}
906 
907 	ret = get_clocks(pdev, gpu);
908 	if (ret)
909 		goto fail;
910 
911 	gpu->ebi1_clk = msm_clk_get(pdev, "bus");
912 	DBG("ebi1_clk: %p", gpu->ebi1_clk);
913 	if (IS_ERR(gpu->ebi1_clk))
914 		gpu->ebi1_clk = NULL;
915 
916 	/* Acquire regulators: */
917 	gpu->gpu_reg = devm_regulator_get(&pdev->dev, "vdd");
918 	DBG("gpu_reg: %p", gpu->gpu_reg);
919 	if (IS_ERR(gpu->gpu_reg))
920 		gpu->gpu_reg = NULL;
921 
922 	gpu->gpu_cx = devm_regulator_get(&pdev->dev, "vddcx");
923 	DBG("gpu_cx: %p", gpu->gpu_cx);
924 	if (IS_ERR(gpu->gpu_cx))
925 		gpu->gpu_cx = NULL;
926 
927 	gpu->pdev = pdev;
928 	platform_set_drvdata(pdev, gpu);
929 
930 	msm_devfreq_init(gpu);
931 
932 	gpu->aspace = msm_gpu_create_address_space(gpu, pdev,
933 		config->va_start, config->va_end);
934 
935 	if (gpu->aspace == NULL)
936 		DRM_DEV_INFO(drm->dev, "%s: no IOMMU, fallback to VRAM carveout!\n", name);
937 	else if (IS_ERR(gpu->aspace)) {
938 		ret = PTR_ERR(gpu->aspace);
939 		goto fail;
940 	}
941 
942 	memptrs = msm_gem_kernel_new(drm,
943 		sizeof(struct msm_rbmemptrs) * nr_rings,
944 		MSM_BO_UNCACHED, gpu->aspace, &gpu->memptrs_bo,
945 		&memptrs_iova);
946 
947 	if (IS_ERR(memptrs)) {
948 		ret = PTR_ERR(memptrs);
949 		DRM_DEV_ERROR(drm->dev, "could not allocate memptrs: %d\n", ret);
950 		goto fail;
951 	}
952 
953 	msm_gem_object_set_name(gpu->memptrs_bo, "memptrs");
954 
955 	if (nr_rings > ARRAY_SIZE(gpu->rb)) {
956 		DRM_DEV_INFO_ONCE(drm->dev, "Only creating %zu ringbuffers\n",
957 			ARRAY_SIZE(gpu->rb));
958 		nr_rings = ARRAY_SIZE(gpu->rb);
959 	}
960 
961 	/* Create ringbuffer(s): */
962 	for (i = 0; i < nr_rings; i++) {
963 		gpu->rb[i] = msm_ringbuffer_new(gpu, i, memptrs, memptrs_iova);
964 
965 		if (IS_ERR(gpu->rb[i])) {
966 			ret = PTR_ERR(gpu->rb[i]);
967 			DRM_DEV_ERROR(drm->dev,
968 				"could not create ringbuffer %d: %d\n", i, ret);
969 			goto fail;
970 		}
971 
972 		memptrs += sizeof(struct msm_rbmemptrs);
973 		memptrs_iova += sizeof(struct msm_rbmemptrs);
974 	}
975 
976 	gpu->nr_rings = nr_rings;
977 
978 	return 0;
979 
980 fail:
981 	for (i = 0; i < ARRAY_SIZE(gpu->rb); i++)  {
982 		msm_ringbuffer_destroy(gpu->rb[i]);
983 		gpu->rb[i] = NULL;
984 	}
985 
986 	msm_gem_kernel_put(gpu->memptrs_bo, gpu->aspace, false);
987 
988 	platform_set_drvdata(pdev, NULL);
989 	return ret;
990 }
991 
992 void msm_gpu_cleanup(struct msm_gpu *gpu)
993 {
994 	int i;
995 
996 	DBG("%s", gpu->name);
997 
998 	WARN_ON(!list_empty(&gpu->active_list));
999 
1000 	for (i = 0; i < ARRAY_SIZE(gpu->rb); i++) {
1001 		msm_ringbuffer_destroy(gpu->rb[i]);
1002 		gpu->rb[i] = NULL;
1003 	}
1004 
1005 	msm_gem_kernel_put(gpu->memptrs_bo, gpu->aspace, false);
1006 
1007 	if (!IS_ERR_OR_NULL(gpu->aspace)) {
1008 		gpu->aspace->mmu->funcs->detach(gpu->aspace->mmu,
1009 			NULL, 0);
1010 		msm_gem_address_space_put(gpu->aspace);
1011 	}
1012 }
1013