xref: /openbmc/linux/drivers/gpu/drm/msm/msm_gpu.c (revision bbb774d9)
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 "drm/drm_drv.h"
8 
9 #include "msm_gpu.h"
10 #include "msm_gem.h"
11 #include "msm_mmu.h"
12 #include "msm_fence.h"
13 #include "msm_gpu_trace.h"
14 #include "adreno/adreno_gpu.h"
15 
16 #include <generated/utsrelease.h>
17 #include <linux/string_helpers.h>
18 #include <linux/devcoredump.h>
19 #include <linux/sched/task.h>
20 
21 /*
22  * Power Management:
23  */
24 
25 static int enable_pwrrail(struct msm_gpu *gpu)
26 {
27 	struct drm_device *dev = gpu->dev;
28 	int ret = 0;
29 
30 	if (gpu->gpu_reg) {
31 		ret = regulator_enable(gpu->gpu_reg);
32 		if (ret) {
33 			DRM_DEV_ERROR(dev->dev, "failed to enable 'gpu_reg': %d\n", ret);
34 			return ret;
35 		}
36 	}
37 
38 	if (gpu->gpu_cx) {
39 		ret = regulator_enable(gpu->gpu_cx);
40 		if (ret) {
41 			DRM_DEV_ERROR(dev->dev, "failed to enable 'gpu_cx': %d\n", ret);
42 			return ret;
43 		}
44 	}
45 
46 	return 0;
47 }
48 
49 static int disable_pwrrail(struct msm_gpu *gpu)
50 {
51 	if (gpu->gpu_cx)
52 		regulator_disable(gpu->gpu_cx);
53 	if (gpu->gpu_reg)
54 		regulator_disable(gpu->gpu_reg);
55 	return 0;
56 }
57 
58 static int enable_clk(struct msm_gpu *gpu)
59 {
60 	if (gpu->core_clk && gpu->fast_rate)
61 		clk_set_rate(gpu->core_clk, gpu->fast_rate);
62 
63 	/* Set the RBBM timer rate to 19.2Mhz */
64 	if (gpu->rbbmtimer_clk)
65 		clk_set_rate(gpu->rbbmtimer_clk, 19200000);
66 
67 	return clk_bulk_prepare_enable(gpu->nr_clocks, gpu->grp_clks);
68 }
69 
70 static int disable_clk(struct msm_gpu *gpu)
71 {
72 	clk_bulk_disable_unprepare(gpu->nr_clocks, gpu->grp_clks);
73 
74 	/*
75 	 * Set the clock to a deliberately low rate. On older targets the clock
76 	 * speed had to be non zero to avoid problems. On newer targets this
77 	 * will be rounded down to zero anyway so it all works out.
78 	 */
79 	if (gpu->core_clk)
80 		clk_set_rate(gpu->core_clk, 27000000);
81 
82 	if (gpu->rbbmtimer_clk)
83 		clk_set_rate(gpu->rbbmtimer_clk, 0);
84 
85 	return 0;
86 }
87 
88 static int enable_axi(struct msm_gpu *gpu)
89 {
90 	return clk_prepare_enable(gpu->ebi1_clk);
91 }
92 
93 static int disable_axi(struct msm_gpu *gpu)
94 {
95 	clk_disable_unprepare(gpu->ebi1_clk);
96 	return 0;
97 }
98 
99 int msm_gpu_pm_resume(struct msm_gpu *gpu)
100 {
101 	int ret;
102 
103 	DBG("%s", gpu->name);
104 	trace_msm_gpu_resume(0);
105 
106 	ret = enable_pwrrail(gpu);
107 	if (ret)
108 		return ret;
109 
110 	ret = enable_clk(gpu);
111 	if (ret)
112 		return ret;
113 
114 	ret = enable_axi(gpu);
115 	if (ret)
116 		return ret;
117 
118 	msm_devfreq_resume(gpu);
119 
120 	gpu->needs_hw_init = true;
121 
122 	return 0;
123 }
124 
125 int msm_gpu_pm_suspend(struct msm_gpu *gpu)
126 {
127 	int ret;
128 
129 	DBG("%s", gpu->name);
130 	trace_msm_gpu_suspend(0);
131 
132 	msm_devfreq_suspend(gpu);
133 
134 	ret = disable_axi(gpu);
135 	if (ret)
136 		return ret;
137 
138 	ret = disable_clk(gpu);
139 	if (ret)
140 		return ret;
141 
142 	ret = disable_pwrrail(gpu);
143 	if (ret)
144 		return ret;
145 
146 	gpu->suspend_count++;
147 
148 	return 0;
149 }
150 
151 void msm_gpu_show_fdinfo(struct msm_gpu *gpu, struct msm_file_private *ctx,
152 			 struct drm_printer *p)
153 {
154 	drm_printf(p, "drm-driver:\t%s\n", gpu->dev->driver->name);
155 	drm_printf(p, "drm-client-id:\t%u\n", ctx->seqno);
156 	drm_printf(p, "drm-engine-gpu:\t%llu ns\n", ctx->elapsed_ns);
157 	drm_printf(p, "drm-cycles-gpu:\t%llu\n", ctx->cycles);
158 	drm_printf(p, "drm-maxfreq-gpu:\t%u Hz\n", gpu->fast_rate);
159 }
160 
161 int msm_gpu_hw_init(struct msm_gpu *gpu)
162 {
163 	int ret;
164 
165 	WARN_ON(!mutex_is_locked(&gpu->lock));
166 
167 	if (!gpu->needs_hw_init)
168 		return 0;
169 
170 	disable_irq(gpu->irq);
171 	ret = gpu->funcs->hw_init(gpu);
172 	if (!ret)
173 		gpu->needs_hw_init = false;
174 	enable_irq(gpu->irq);
175 
176 	return ret;
177 }
178 
179 #ifdef CONFIG_DEV_COREDUMP
180 static ssize_t msm_gpu_devcoredump_read(char *buffer, loff_t offset,
181 		size_t count, void *data, size_t datalen)
182 {
183 	struct msm_gpu *gpu = data;
184 	struct drm_print_iterator iter;
185 	struct drm_printer p;
186 	struct msm_gpu_state *state;
187 
188 	state = msm_gpu_crashstate_get(gpu);
189 	if (!state)
190 		return 0;
191 
192 	iter.data = buffer;
193 	iter.offset = 0;
194 	iter.start = offset;
195 	iter.remain = count;
196 
197 	p = drm_coredump_printer(&iter);
198 
199 	drm_printf(&p, "---\n");
200 	drm_printf(&p, "kernel: " UTS_RELEASE "\n");
201 	drm_printf(&p, "module: " KBUILD_MODNAME "\n");
202 	drm_printf(&p, "time: %lld.%09ld\n",
203 		state->time.tv_sec, state->time.tv_nsec);
204 	if (state->comm)
205 		drm_printf(&p, "comm: %s\n", state->comm);
206 	if (state->cmd)
207 		drm_printf(&p, "cmdline: %s\n", state->cmd);
208 
209 	gpu->funcs->show(gpu, state, &p);
210 
211 	msm_gpu_crashstate_put(gpu);
212 
213 	return count - iter.remain;
214 }
215 
216 static void msm_gpu_devcoredump_free(void *data)
217 {
218 	struct msm_gpu *gpu = data;
219 
220 	msm_gpu_crashstate_put(gpu);
221 }
222 
223 static void msm_gpu_crashstate_get_bo(struct msm_gpu_state *state,
224 		struct msm_gem_object *obj, u64 iova, bool full)
225 {
226 	struct msm_gpu_state_bo *state_bo = &state->bos[state->nr_bos];
227 
228 	/* Don't record write only objects */
229 	state_bo->size = obj->base.size;
230 	state_bo->iova = iova;
231 
232 	BUILD_BUG_ON(sizeof(state_bo->name) != sizeof(obj->name));
233 
234 	memcpy(state_bo->name, obj->name, sizeof(state_bo->name));
235 
236 	if (full) {
237 		void *ptr;
238 
239 		state_bo->data = kvmalloc(obj->base.size, GFP_KERNEL);
240 		if (!state_bo->data)
241 			goto out;
242 
243 		msm_gem_lock(&obj->base);
244 		ptr = msm_gem_get_vaddr_active(&obj->base);
245 		msm_gem_unlock(&obj->base);
246 		if (IS_ERR(ptr)) {
247 			kvfree(state_bo->data);
248 			state_bo->data = NULL;
249 			goto out;
250 		}
251 
252 		memcpy(state_bo->data, ptr, obj->base.size);
253 		msm_gem_put_vaddr(&obj->base);
254 	}
255 out:
256 	state->nr_bos++;
257 }
258 
259 static void msm_gpu_crashstate_capture(struct msm_gpu *gpu,
260 		struct msm_gem_submit *submit, char *comm, char *cmd)
261 {
262 	struct msm_gpu_state *state;
263 
264 	/* Check if the target supports capturing crash state */
265 	if (!gpu->funcs->gpu_state_get)
266 		return;
267 
268 	/* Only save one crash state at a time */
269 	if (gpu->crashstate)
270 		return;
271 
272 	state = gpu->funcs->gpu_state_get(gpu);
273 	if (IS_ERR_OR_NULL(state))
274 		return;
275 
276 	/* Fill in the additional crash state information */
277 	state->comm = kstrdup(comm, GFP_KERNEL);
278 	state->cmd = kstrdup(cmd, GFP_KERNEL);
279 	state->fault_info = gpu->fault_info;
280 
281 	if (submit) {
282 		int i;
283 
284 		state->bos = kcalloc(submit->nr_bos,
285 			sizeof(struct msm_gpu_state_bo), GFP_KERNEL);
286 
287 		for (i = 0; state->bos && i < submit->nr_bos; i++) {
288 			msm_gpu_crashstate_get_bo(state, submit->bos[i].obj,
289 						  submit->bos[i].iova,
290 						  should_dump(submit, i));
291 		}
292 	}
293 
294 	/* Set the active crash state to be dumped on failure */
295 	gpu->crashstate = state;
296 
297 	/* FIXME: Release the crashstate if this errors out? */
298 	dev_coredumpm(gpu->dev->dev, THIS_MODULE, gpu, 0, GFP_KERNEL,
299 		msm_gpu_devcoredump_read, msm_gpu_devcoredump_free);
300 }
301 #else
302 static void msm_gpu_crashstate_capture(struct msm_gpu *gpu,
303 		struct msm_gem_submit *submit, char *comm, char *cmd)
304 {
305 }
306 #endif
307 
308 /*
309  * Hangcheck detection for locked gpu:
310  */
311 
312 static struct msm_gem_submit *
313 find_submit(struct msm_ringbuffer *ring, uint32_t fence)
314 {
315 	struct msm_gem_submit *submit;
316 	unsigned long flags;
317 
318 	spin_lock_irqsave(&ring->submit_lock, flags);
319 	list_for_each_entry(submit, &ring->submits, node) {
320 		if (submit->seqno == fence) {
321 			spin_unlock_irqrestore(&ring->submit_lock, flags);
322 			return submit;
323 		}
324 	}
325 	spin_unlock_irqrestore(&ring->submit_lock, flags);
326 
327 	return NULL;
328 }
329 
330 static void retire_submits(struct msm_gpu *gpu);
331 
332 static void get_comm_cmdline(struct msm_gem_submit *submit, char **comm, char **cmd)
333 {
334 	struct msm_file_private *ctx = submit->queue->ctx;
335 	struct task_struct *task;
336 
337 	/* Note that kstrdup will return NULL if argument is NULL: */
338 	*comm = kstrdup(ctx->comm, GFP_KERNEL);
339 	*cmd  = kstrdup(ctx->cmdline, GFP_KERNEL);
340 
341 	task = get_pid_task(submit->pid, PIDTYPE_PID);
342 	if (!task)
343 		return;
344 
345 	if (!*comm)
346 		*comm = kstrdup(task->comm, GFP_KERNEL);
347 
348 	if (!*cmd)
349 		*cmd = kstrdup_quotable_cmdline(task, GFP_KERNEL);
350 
351 	put_task_struct(task);
352 }
353 
354 static void recover_worker(struct kthread_work *work)
355 {
356 	struct msm_gpu *gpu = container_of(work, struct msm_gpu, recover_work);
357 	struct drm_device *dev = gpu->dev;
358 	struct msm_drm_private *priv = dev->dev_private;
359 	struct msm_gem_submit *submit;
360 	struct msm_ringbuffer *cur_ring = gpu->funcs->active_ring(gpu);
361 	char *comm = NULL, *cmd = NULL;
362 	int i;
363 
364 	mutex_lock(&gpu->lock);
365 
366 	DRM_DEV_ERROR(dev->dev, "%s: hangcheck recover!\n", gpu->name);
367 
368 	submit = find_submit(cur_ring, cur_ring->memptrs->fence + 1);
369 	if (submit) {
370 		/* Increment the fault counts */
371 		submit->queue->faults++;
372 		if (submit->aspace)
373 			submit->aspace->faults++;
374 
375 		get_comm_cmdline(submit, &comm, &cmd);
376 
377 		if (comm && cmd) {
378 			DRM_DEV_ERROR(dev->dev, "%s: offending task: %s (%s)\n",
379 				gpu->name, comm, cmd);
380 
381 			msm_rd_dump_submit(priv->hangrd, submit,
382 				"offending task: %s (%s)", comm, cmd);
383 		} else {
384 			msm_rd_dump_submit(priv->hangrd, submit, NULL);
385 		}
386 	} else {
387 		/*
388 		 * We couldn't attribute this fault to any particular context,
389 		 * so increment the global fault count instead.
390 		 */
391 		gpu->global_faults++;
392 	}
393 
394 	/* Record the crash state */
395 	pm_runtime_get_sync(&gpu->pdev->dev);
396 	msm_gpu_crashstate_capture(gpu, submit, comm, cmd);
397 	pm_runtime_put_sync(&gpu->pdev->dev);
398 
399 	kfree(cmd);
400 	kfree(comm);
401 
402 	/*
403 	 * Update all the rings with the latest and greatest fence.. this
404 	 * needs to happen after msm_rd_dump_submit() to ensure that the
405 	 * bo's referenced by the offending submit are still around.
406 	 */
407 	for (i = 0; i < gpu->nr_rings; i++) {
408 		struct msm_ringbuffer *ring = gpu->rb[i];
409 
410 		uint32_t fence = ring->memptrs->fence;
411 
412 		/*
413 		 * For the current (faulting?) ring/submit advance the fence by
414 		 * one more to clear the faulting submit
415 		 */
416 		if (ring == cur_ring)
417 			ring->memptrs->fence = ++fence;
418 
419 		msm_update_fence(ring->fctx, fence);
420 	}
421 
422 	if (msm_gpu_active(gpu)) {
423 		/* retire completed submits, plus the one that hung: */
424 		retire_submits(gpu);
425 
426 		pm_runtime_get_sync(&gpu->pdev->dev);
427 		gpu->funcs->recover(gpu);
428 		pm_runtime_put_sync(&gpu->pdev->dev);
429 
430 		/*
431 		 * Replay all remaining submits starting with highest priority
432 		 * ring
433 		 */
434 		for (i = 0; i < gpu->nr_rings; i++) {
435 			struct msm_ringbuffer *ring = gpu->rb[i];
436 			unsigned long flags;
437 
438 			spin_lock_irqsave(&ring->submit_lock, flags);
439 			list_for_each_entry(submit, &ring->submits, node)
440 				gpu->funcs->submit(gpu, submit);
441 			spin_unlock_irqrestore(&ring->submit_lock, flags);
442 		}
443 	}
444 
445 	mutex_unlock(&gpu->lock);
446 
447 	msm_gpu_retire(gpu);
448 }
449 
450 static void fault_worker(struct kthread_work *work)
451 {
452 	struct msm_gpu *gpu = container_of(work, struct msm_gpu, fault_work);
453 	struct msm_gem_submit *submit;
454 	struct msm_ringbuffer *cur_ring = gpu->funcs->active_ring(gpu);
455 	char *comm = NULL, *cmd = NULL;
456 
457 	mutex_lock(&gpu->lock);
458 
459 	submit = find_submit(cur_ring, cur_ring->memptrs->fence + 1);
460 	if (submit && submit->fault_dumped)
461 		goto resume_smmu;
462 
463 	if (submit) {
464 		get_comm_cmdline(submit, &comm, &cmd);
465 
466 		/*
467 		 * When we get GPU iova faults, we can get 1000s of them,
468 		 * but we really only want to log the first one.
469 		 */
470 		submit->fault_dumped = true;
471 	}
472 
473 	/* Record the crash state */
474 	pm_runtime_get_sync(&gpu->pdev->dev);
475 	msm_gpu_crashstate_capture(gpu, submit, comm, cmd);
476 	pm_runtime_put_sync(&gpu->pdev->dev);
477 
478 	kfree(cmd);
479 	kfree(comm);
480 
481 resume_smmu:
482 	memset(&gpu->fault_info, 0, sizeof(gpu->fault_info));
483 	gpu->aspace->mmu->funcs->resume_translation(gpu->aspace->mmu);
484 
485 	mutex_unlock(&gpu->lock);
486 }
487 
488 static void hangcheck_timer_reset(struct msm_gpu *gpu)
489 {
490 	struct msm_drm_private *priv = gpu->dev->dev_private;
491 	mod_timer(&gpu->hangcheck_timer,
492 			round_jiffies_up(jiffies + msecs_to_jiffies(priv->hangcheck_period)));
493 }
494 
495 static void hangcheck_handler(struct timer_list *t)
496 {
497 	struct msm_gpu *gpu = from_timer(gpu, t, hangcheck_timer);
498 	struct drm_device *dev = gpu->dev;
499 	struct msm_ringbuffer *ring = gpu->funcs->active_ring(gpu);
500 	uint32_t fence = ring->memptrs->fence;
501 
502 	if (fence != ring->hangcheck_fence) {
503 		/* some progress has been made.. ya! */
504 		ring->hangcheck_fence = fence;
505 	} else if (fence_before(fence, ring->fctx->last_fence)) {
506 		/* no progress and not done.. hung! */
507 		ring->hangcheck_fence = fence;
508 		DRM_DEV_ERROR(dev->dev, "%s: hangcheck detected gpu lockup rb %d!\n",
509 				gpu->name, ring->id);
510 		DRM_DEV_ERROR(dev->dev, "%s:     completed fence: %u\n",
511 				gpu->name, fence);
512 		DRM_DEV_ERROR(dev->dev, "%s:     submitted fence: %u\n",
513 				gpu->name, ring->fctx->last_fence);
514 
515 		kthread_queue_work(gpu->worker, &gpu->recover_work);
516 	}
517 
518 	/* if still more pending work, reset the hangcheck timer: */
519 	if (fence_after(ring->fctx->last_fence, ring->hangcheck_fence))
520 		hangcheck_timer_reset(gpu);
521 
522 	/* workaround for missing irq: */
523 	msm_gpu_retire(gpu);
524 }
525 
526 /*
527  * Performance Counters:
528  */
529 
530 /* called under perf_lock */
531 static int update_hw_cntrs(struct msm_gpu *gpu, uint32_t ncntrs, uint32_t *cntrs)
532 {
533 	uint32_t current_cntrs[ARRAY_SIZE(gpu->last_cntrs)];
534 	int i, n = min(ncntrs, gpu->num_perfcntrs);
535 
536 	/* read current values: */
537 	for (i = 0; i < gpu->num_perfcntrs; i++)
538 		current_cntrs[i] = gpu_read(gpu, gpu->perfcntrs[i].sample_reg);
539 
540 	/* update cntrs: */
541 	for (i = 0; i < n; i++)
542 		cntrs[i] = current_cntrs[i] - gpu->last_cntrs[i];
543 
544 	/* save current values: */
545 	for (i = 0; i < gpu->num_perfcntrs; i++)
546 		gpu->last_cntrs[i] = current_cntrs[i];
547 
548 	return n;
549 }
550 
551 static void update_sw_cntrs(struct msm_gpu *gpu)
552 {
553 	ktime_t time;
554 	uint32_t elapsed;
555 	unsigned long flags;
556 
557 	spin_lock_irqsave(&gpu->perf_lock, flags);
558 	if (!gpu->perfcntr_active)
559 		goto out;
560 
561 	time = ktime_get();
562 	elapsed = ktime_to_us(ktime_sub(time, gpu->last_sample.time));
563 
564 	gpu->totaltime += elapsed;
565 	if (gpu->last_sample.active)
566 		gpu->activetime += elapsed;
567 
568 	gpu->last_sample.active = msm_gpu_active(gpu);
569 	gpu->last_sample.time = time;
570 
571 out:
572 	spin_unlock_irqrestore(&gpu->perf_lock, flags);
573 }
574 
575 void msm_gpu_perfcntr_start(struct msm_gpu *gpu)
576 {
577 	unsigned long flags;
578 
579 	pm_runtime_get_sync(&gpu->pdev->dev);
580 
581 	spin_lock_irqsave(&gpu->perf_lock, flags);
582 	/* we could dynamically enable/disable perfcntr registers too.. */
583 	gpu->last_sample.active = msm_gpu_active(gpu);
584 	gpu->last_sample.time = ktime_get();
585 	gpu->activetime = gpu->totaltime = 0;
586 	gpu->perfcntr_active = true;
587 	update_hw_cntrs(gpu, 0, NULL);
588 	spin_unlock_irqrestore(&gpu->perf_lock, flags);
589 }
590 
591 void msm_gpu_perfcntr_stop(struct msm_gpu *gpu)
592 {
593 	gpu->perfcntr_active = false;
594 	pm_runtime_put_sync(&gpu->pdev->dev);
595 }
596 
597 /* returns -errno or # of cntrs sampled */
598 int msm_gpu_perfcntr_sample(struct msm_gpu *gpu, uint32_t *activetime,
599 		uint32_t *totaltime, uint32_t ncntrs, uint32_t *cntrs)
600 {
601 	unsigned long flags;
602 	int ret;
603 
604 	spin_lock_irqsave(&gpu->perf_lock, flags);
605 
606 	if (!gpu->perfcntr_active) {
607 		ret = -EINVAL;
608 		goto out;
609 	}
610 
611 	*activetime = gpu->activetime;
612 	*totaltime = gpu->totaltime;
613 
614 	gpu->activetime = gpu->totaltime = 0;
615 
616 	ret = update_hw_cntrs(gpu, ncntrs, cntrs);
617 
618 out:
619 	spin_unlock_irqrestore(&gpu->perf_lock, flags);
620 
621 	return ret;
622 }
623 
624 /*
625  * Cmdstream submission/retirement:
626  */
627 
628 static void retire_submit(struct msm_gpu *gpu, struct msm_ringbuffer *ring,
629 		struct msm_gem_submit *submit)
630 {
631 	int index = submit->seqno % MSM_GPU_SUBMIT_STATS_COUNT;
632 	volatile struct msm_gpu_submit_stats *stats;
633 	u64 elapsed, clock = 0, cycles;
634 	unsigned long flags;
635 
636 	stats = &ring->memptrs->stats[index];
637 	/* Convert 19.2Mhz alwayson ticks to nanoseconds for elapsed time */
638 	elapsed = (stats->alwayson_end - stats->alwayson_start) * 10000;
639 	do_div(elapsed, 192);
640 
641 	cycles = stats->cpcycles_end - stats->cpcycles_start;
642 
643 	/* Calculate the clock frequency from the number of CP cycles */
644 	if (elapsed) {
645 		clock = cycles * 1000;
646 		do_div(clock, elapsed);
647 	}
648 
649 	submit->queue->ctx->elapsed_ns += elapsed;
650 	submit->queue->ctx->cycles     += cycles;
651 
652 	trace_msm_gpu_submit_retired(submit, elapsed, clock,
653 		stats->alwayson_start, stats->alwayson_end);
654 
655 	msm_submit_retire(submit);
656 
657 	pm_runtime_mark_last_busy(&gpu->pdev->dev);
658 
659 	spin_lock_irqsave(&ring->submit_lock, flags);
660 	list_del(&submit->node);
661 	spin_unlock_irqrestore(&ring->submit_lock, flags);
662 
663 	/* Update devfreq on transition from active->idle: */
664 	mutex_lock(&gpu->active_lock);
665 	gpu->active_submits--;
666 	WARN_ON(gpu->active_submits < 0);
667 	if (!gpu->active_submits)
668 		msm_devfreq_idle(gpu);
669 	mutex_unlock(&gpu->active_lock);
670 
671 	pm_runtime_put_autosuspend(&gpu->pdev->dev);
672 
673 	msm_gem_submit_put(submit);
674 }
675 
676 static void retire_submits(struct msm_gpu *gpu)
677 {
678 	int i;
679 
680 	/* Retire the commits starting with highest priority */
681 	for (i = 0; i < gpu->nr_rings; i++) {
682 		struct msm_ringbuffer *ring = gpu->rb[i];
683 
684 		while (true) {
685 			struct msm_gem_submit *submit = NULL;
686 			unsigned long flags;
687 
688 			spin_lock_irqsave(&ring->submit_lock, flags);
689 			submit = list_first_entry_or_null(&ring->submits,
690 					struct msm_gem_submit, node);
691 			spin_unlock_irqrestore(&ring->submit_lock, flags);
692 
693 			/*
694 			 * If no submit, we are done.  If submit->fence hasn't
695 			 * been signalled, then later submits are not signalled
696 			 * either, so we are also done.
697 			 */
698 			if (submit && dma_fence_is_signaled(submit->hw_fence)) {
699 				retire_submit(gpu, ring, submit);
700 			} else {
701 				break;
702 			}
703 		}
704 	}
705 
706 	wake_up_all(&gpu->retire_event);
707 }
708 
709 static void retire_worker(struct kthread_work *work)
710 {
711 	struct msm_gpu *gpu = container_of(work, struct msm_gpu, retire_work);
712 
713 	retire_submits(gpu);
714 }
715 
716 /* call from irq handler to schedule work to retire bo's */
717 void msm_gpu_retire(struct msm_gpu *gpu)
718 {
719 	int i;
720 
721 	for (i = 0; i < gpu->nr_rings; i++)
722 		msm_update_fence(gpu->rb[i]->fctx, gpu->rb[i]->memptrs->fence);
723 
724 	kthread_queue_work(gpu->worker, &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 {
731 	struct drm_device *dev = gpu->dev;
732 	struct msm_drm_private *priv = dev->dev_private;
733 	struct msm_ringbuffer *ring = submit->ring;
734 	unsigned long flags;
735 
736 	WARN_ON(!mutex_is_locked(&gpu->lock));
737 
738 	pm_runtime_get_sync(&gpu->pdev->dev);
739 
740 	msm_gpu_hw_init(gpu);
741 
742 	submit->seqno = submit->hw_fence->seqno;
743 
744 	msm_rd_dump_submit(priv->rd, submit, NULL);
745 
746 	update_sw_cntrs(gpu);
747 
748 	/*
749 	 * ring->submits holds a ref to the submit, to deal with the case
750 	 * that a submit completes before msm_ioctl_gem_submit() returns.
751 	 */
752 	msm_gem_submit_get(submit);
753 
754 	spin_lock_irqsave(&ring->submit_lock, flags);
755 	list_add_tail(&submit->node, &ring->submits);
756 	spin_unlock_irqrestore(&ring->submit_lock, flags);
757 
758 	/* Update devfreq on transition from idle->active: */
759 	mutex_lock(&gpu->active_lock);
760 	if (!gpu->active_submits)
761 		msm_devfreq_active(gpu);
762 	gpu->active_submits++;
763 	mutex_unlock(&gpu->active_lock);
764 
765 	gpu->funcs->submit(gpu, submit);
766 	gpu->cur_ctx_seqno = submit->queue->ctx->seqno;
767 
768 	hangcheck_timer_reset(gpu);
769 }
770 
771 /*
772  * Init/Cleanup:
773  */
774 
775 static irqreturn_t irq_handler(int irq, void *data)
776 {
777 	struct msm_gpu *gpu = data;
778 	return gpu->funcs->irq(gpu);
779 }
780 
781 static int get_clocks(struct platform_device *pdev, struct msm_gpu *gpu)
782 {
783 	int ret = devm_clk_bulk_get_all(&pdev->dev, &gpu->grp_clks);
784 
785 	if (ret < 1) {
786 		gpu->nr_clocks = 0;
787 		return ret;
788 	}
789 
790 	gpu->nr_clocks = ret;
791 
792 	gpu->core_clk = msm_clk_bulk_get_clock(gpu->grp_clks,
793 		gpu->nr_clocks, "core");
794 
795 	gpu->rbbmtimer_clk = msm_clk_bulk_get_clock(gpu->grp_clks,
796 		gpu->nr_clocks, "rbbmtimer");
797 
798 	return 0;
799 }
800 
801 /* Return a new address space for a msm_drm_private instance */
802 struct msm_gem_address_space *
803 msm_gpu_create_private_address_space(struct msm_gpu *gpu, struct task_struct *task)
804 {
805 	struct msm_gem_address_space *aspace = NULL;
806 	if (!gpu)
807 		return NULL;
808 
809 	/*
810 	 * If the target doesn't support private address spaces then return
811 	 * the global one
812 	 */
813 	if (gpu->funcs->create_private_address_space) {
814 		aspace = gpu->funcs->create_private_address_space(gpu);
815 		if (!IS_ERR(aspace))
816 			aspace->pid = get_pid(task_pid(task));
817 	}
818 
819 	if (IS_ERR_OR_NULL(aspace))
820 		aspace = msm_gem_address_space_get(gpu->aspace);
821 
822 	return aspace;
823 }
824 
825 int msm_gpu_init(struct drm_device *drm, struct platform_device *pdev,
826 		struct msm_gpu *gpu, const struct msm_gpu_funcs *funcs,
827 		const char *name, struct msm_gpu_config *config)
828 {
829 	int i, ret, nr_rings = config->nr_rings;
830 	void *memptrs;
831 	uint64_t memptrs_iova;
832 
833 	if (WARN_ON(gpu->num_perfcntrs > ARRAY_SIZE(gpu->last_cntrs)))
834 		gpu->num_perfcntrs = ARRAY_SIZE(gpu->last_cntrs);
835 
836 	gpu->dev = drm;
837 	gpu->funcs = funcs;
838 	gpu->name = name;
839 
840 	gpu->worker = kthread_create_worker(0, "gpu-worker");
841 	if (IS_ERR(gpu->worker)) {
842 		ret = PTR_ERR(gpu->worker);
843 		gpu->worker = NULL;
844 		goto fail;
845 	}
846 
847 	sched_set_fifo_low(gpu->worker->task);
848 
849 	INIT_LIST_HEAD(&gpu->active_list);
850 	mutex_init(&gpu->active_lock);
851 	mutex_init(&gpu->lock);
852 	init_waitqueue_head(&gpu->retire_event);
853 	kthread_init_work(&gpu->retire_work, retire_worker);
854 	kthread_init_work(&gpu->recover_work, recover_worker);
855 	kthread_init_work(&gpu->fault_work, fault_worker);
856 
857 	timer_setup(&gpu->hangcheck_timer, hangcheck_handler, 0);
858 
859 	spin_lock_init(&gpu->perf_lock);
860 
861 
862 	/* Map registers: */
863 	gpu->mmio = msm_ioremap(pdev, config->ioname);
864 	if (IS_ERR(gpu->mmio)) {
865 		ret = PTR_ERR(gpu->mmio);
866 		goto fail;
867 	}
868 
869 	/* Get Interrupt: */
870 	gpu->irq = platform_get_irq(pdev, 0);
871 	if (gpu->irq < 0) {
872 		ret = gpu->irq;
873 		DRM_DEV_ERROR(drm->dev, "failed to get irq: %d\n", ret);
874 		goto fail;
875 	}
876 
877 	ret = devm_request_irq(&pdev->dev, gpu->irq, irq_handler,
878 			IRQF_TRIGGER_HIGH, "gpu-irq", gpu);
879 	if (ret) {
880 		DRM_DEV_ERROR(drm->dev, "failed to request IRQ%u: %d\n", gpu->irq, ret);
881 		goto fail;
882 	}
883 
884 	ret = get_clocks(pdev, gpu);
885 	if (ret)
886 		goto fail;
887 
888 	gpu->ebi1_clk = msm_clk_get(pdev, "bus");
889 	DBG("ebi1_clk: %p", gpu->ebi1_clk);
890 	if (IS_ERR(gpu->ebi1_clk))
891 		gpu->ebi1_clk = NULL;
892 
893 	/* Acquire regulators: */
894 	gpu->gpu_reg = devm_regulator_get(&pdev->dev, "vdd");
895 	DBG("gpu_reg: %p", gpu->gpu_reg);
896 	if (IS_ERR(gpu->gpu_reg))
897 		gpu->gpu_reg = NULL;
898 
899 	gpu->gpu_cx = devm_regulator_get(&pdev->dev, "vddcx");
900 	DBG("gpu_cx: %p", gpu->gpu_cx);
901 	if (IS_ERR(gpu->gpu_cx))
902 		gpu->gpu_cx = NULL;
903 
904 	gpu->pdev = pdev;
905 	platform_set_drvdata(pdev, &gpu->adreno_smmu);
906 
907 	msm_devfreq_init(gpu);
908 
909 
910 	gpu->aspace = gpu->funcs->create_address_space(gpu, pdev);
911 
912 	if (gpu->aspace == NULL)
913 		DRM_DEV_INFO(drm->dev, "%s: no IOMMU, fallback to VRAM carveout!\n", name);
914 	else if (IS_ERR(gpu->aspace)) {
915 		ret = PTR_ERR(gpu->aspace);
916 		goto fail;
917 	}
918 
919 	memptrs = msm_gem_kernel_new(drm,
920 		sizeof(struct msm_rbmemptrs) * nr_rings,
921 		check_apriv(gpu, MSM_BO_WC), gpu->aspace, &gpu->memptrs_bo,
922 		&memptrs_iova);
923 
924 	if (IS_ERR(memptrs)) {
925 		ret = PTR_ERR(memptrs);
926 		DRM_DEV_ERROR(drm->dev, "could not allocate memptrs: %d\n", ret);
927 		goto fail;
928 	}
929 
930 	msm_gem_object_set_name(gpu->memptrs_bo, "memptrs");
931 
932 	if (nr_rings > ARRAY_SIZE(gpu->rb)) {
933 		DRM_DEV_INFO_ONCE(drm->dev, "Only creating %zu ringbuffers\n",
934 			ARRAY_SIZE(gpu->rb));
935 		nr_rings = ARRAY_SIZE(gpu->rb);
936 	}
937 
938 	/* Create ringbuffer(s): */
939 	for (i = 0; i < nr_rings; i++) {
940 		gpu->rb[i] = msm_ringbuffer_new(gpu, i, memptrs, memptrs_iova);
941 
942 		if (IS_ERR(gpu->rb[i])) {
943 			ret = PTR_ERR(gpu->rb[i]);
944 			DRM_DEV_ERROR(drm->dev,
945 				"could not create ringbuffer %d: %d\n", i, ret);
946 			goto fail;
947 		}
948 
949 		memptrs += sizeof(struct msm_rbmemptrs);
950 		memptrs_iova += sizeof(struct msm_rbmemptrs);
951 	}
952 
953 	gpu->nr_rings = nr_rings;
954 
955 	refcount_set(&gpu->sysprof_active, 1);
956 
957 	return 0;
958 
959 fail:
960 	for (i = 0; i < ARRAY_SIZE(gpu->rb); i++)  {
961 		msm_ringbuffer_destroy(gpu->rb[i]);
962 		gpu->rb[i] = NULL;
963 	}
964 
965 	msm_gem_kernel_put(gpu->memptrs_bo, gpu->aspace);
966 
967 	platform_set_drvdata(pdev, NULL);
968 	return ret;
969 }
970 
971 void msm_gpu_cleanup(struct msm_gpu *gpu)
972 {
973 	int i;
974 
975 	DBG("%s", gpu->name);
976 
977 	WARN_ON(!list_empty(&gpu->active_list));
978 
979 	for (i = 0; i < ARRAY_SIZE(gpu->rb); i++) {
980 		msm_ringbuffer_destroy(gpu->rb[i]);
981 		gpu->rb[i] = NULL;
982 	}
983 
984 	msm_gem_kernel_put(gpu->memptrs_bo, gpu->aspace);
985 
986 	if (!IS_ERR_OR_NULL(gpu->aspace)) {
987 		gpu->aspace->mmu->funcs->detach(gpu->aspace->mmu);
988 		msm_gem_address_space_put(gpu->aspace);
989 	}
990 
991 	if (gpu->worker) {
992 		kthread_destroy_worker(gpu->worker);
993 	}
994 
995 	msm_devfreq_cleanup(gpu);
996 }
997