xref: /linux/drivers/gpu/drm/msm/msm_gpu.c (revision 4b132aacb0768ac1e652cf517097ea6f237214b9)
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 		dev_pm_opp_set_rate(&gpu->pdev->dev, 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 		dev_pm_opp_set_rate(&gpu->pdev->dev, 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-engine-gpu:\t%llu ns\n", ctx->elapsed_ns);
155 	drm_printf(p, "drm-cycles-gpu:\t%llu\n", ctx->cycles);
156 	drm_printf(p, "drm-maxfreq-gpu:\t%u Hz\n", gpu->fast_rate);
157 }
158 
159 int msm_gpu_hw_init(struct msm_gpu *gpu)
160 {
161 	int ret;
162 
163 	WARN_ON(!mutex_is_locked(&gpu->lock));
164 
165 	if (!gpu->needs_hw_init)
166 		return 0;
167 
168 	disable_irq(gpu->irq);
169 	ret = gpu->funcs->hw_init(gpu);
170 	if (!ret)
171 		gpu->needs_hw_init = false;
172 	enable_irq(gpu->irq);
173 
174 	return ret;
175 }
176 
177 #ifdef CONFIG_DEV_COREDUMP
178 static ssize_t msm_gpu_devcoredump_read(char *buffer, loff_t offset,
179 		size_t count, void *data, size_t datalen)
180 {
181 	struct msm_gpu *gpu = data;
182 	struct drm_print_iterator iter;
183 	struct drm_printer p;
184 	struct msm_gpu_state *state;
185 
186 	state = msm_gpu_crashstate_get(gpu);
187 	if (!state)
188 		return 0;
189 
190 	iter.data = buffer;
191 	iter.offset = 0;
192 	iter.start = offset;
193 	iter.remain = count;
194 
195 	p = drm_coredump_printer(&iter);
196 
197 	drm_printf(&p, "---\n");
198 	drm_printf(&p, "kernel: " UTS_RELEASE "\n");
199 	drm_printf(&p, "module: " KBUILD_MODNAME "\n");
200 	drm_printf(&p, "time: %lld.%09ld\n",
201 		state->time.tv_sec, state->time.tv_nsec);
202 	if (state->comm)
203 		drm_printf(&p, "comm: %s\n", state->comm);
204 	if (state->cmd)
205 		drm_printf(&p, "cmdline: %s\n", state->cmd);
206 
207 	gpu->funcs->show(gpu, state, &p);
208 
209 	msm_gpu_crashstate_put(gpu);
210 
211 	return count - iter.remain;
212 }
213 
214 static void msm_gpu_devcoredump_free(void *data)
215 {
216 	struct msm_gpu *gpu = data;
217 
218 	msm_gpu_crashstate_put(gpu);
219 }
220 
221 static void msm_gpu_crashstate_get_bo(struct msm_gpu_state *state,
222 		struct drm_gem_object *obj, u64 iova, bool full)
223 {
224 	struct msm_gpu_state_bo *state_bo = &state->bos[state->nr_bos];
225 	struct msm_gem_object *msm_obj = to_msm_bo(obj);
226 
227 	/* Don't record write only objects */
228 	state_bo->size = obj->size;
229 	state_bo->flags = msm_obj->flags;
230 	state_bo->iova = iova;
231 
232 	BUILD_BUG_ON(sizeof(state_bo->name) != sizeof(msm_obj->name));
233 
234 	memcpy(state_bo->name, msm_obj->name, sizeof(state_bo->name));
235 
236 	if (full) {
237 		void *ptr;
238 
239 		state_bo->data = kvmalloc(obj->size, GFP_KERNEL);
240 		if (!state_bo->data)
241 			goto out;
242 
243 		msm_gem_lock(obj);
244 		ptr = msm_gem_get_vaddr_active(obj);
245 		msm_gem_unlock(obj);
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->size);
253 		msm_gem_put_vaddr(obj);
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 	dev_coredumpm(&gpu->pdev->dev, THIS_MODULE, gpu, 0, GFP_KERNEL,
298 		msm_gpu_devcoredump_read, msm_gpu_devcoredump_free);
299 }
300 #else
301 static void msm_gpu_crashstate_capture(struct msm_gpu *gpu,
302 		struct msm_gem_submit *submit, char *comm, char *cmd)
303 {
304 }
305 #endif
306 
307 /*
308  * Hangcheck detection for locked gpu:
309  */
310 
311 static struct msm_gem_submit *
312 find_submit(struct msm_ringbuffer *ring, uint32_t fence)
313 {
314 	struct msm_gem_submit *submit;
315 	unsigned long flags;
316 
317 	spin_lock_irqsave(&ring->submit_lock, flags);
318 	list_for_each_entry(submit, &ring->submits, node) {
319 		if (submit->seqno == fence) {
320 			spin_unlock_irqrestore(&ring->submit_lock, flags);
321 			return submit;
322 		}
323 	}
324 	spin_unlock_irqrestore(&ring->submit_lock, flags);
325 
326 	return NULL;
327 }
328 
329 static void retire_submits(struct msm_gpu *gpu);
330 
331 static void get_comm_cmdline(struct msm_gem_submit *submit, char **comm, char **cmd)
332 {
333 	struct msm_file_private *ctx = submit->queue->ctx;
334 	struct task_struct *task;
335 
336 	WARN_ON(!mutex_is_locked(&submit->gpu->lock));
337 
338 	/* Note that kstrdup will return NULL if argument is NULL: */
339 	*comm = kstrdup(ctx->comm, GFP_KERNEL);
340 	*cmd  = kstrdup(ctx->cmdline, GFP_KERNEL);
341 
342 	task = get_pid_task(submit->pid, PIDTYPE_PID);
343 	if (!task)
344 		return;
345 
346 	if (!*comm)
347 		*comm = kstrdup(task->comm, GFP_KERNEL);
348 
349 	if (!*cmd)
350 		*cmd = kstrdup_quotable_cmdline(task, GFP_KERNEL);
351 
352 	put_task_struct(task);
353 }
354 
355 static void recover_worker(struct kthread_work *work)
356 {
357 	struct msm_gpu *gpu = container_of(work, struct msm_gpu, recover_work);
358 	struct drm_device *dev = gpu->dev;
359 	struct msm_drm_private *priv = dev->dev_private;
360 	struct msm_gem_submit *submit;
361 	struct msm_ringbuffer *cur_ring = gpu->funcs->active_ring(gpu);
362 	char *comm = NULL, *cmd = NULL;
363 	int i;
364 
365 	mutex_lock(&gpu->lock);
366 
367 	DRM_DEV_ERROR(dev->dev, "%s: hangcheck recover!\n", gpu->name);
368 
369 	submit = find_submit(cur_ring, cur_ring->memptrs->fence + 1);
370 
371 	/*
372 	 * If the submit retired while we were waiting for the worker to run,
373 	 * or waiting to acquire the gpu lock, then nothing more to do.
374 	 */
375 	if (!submit)
376 		goto out_unlock;
377 
378 	/* Increment the fault counts */
379 	submit->queue->faults++;
380 	if (submit->aspace)
381 		submit->aspace->faults++;
382 
383 	get_comm_cmdline(submit, &comm, &cmd);
384 
385 	if (comm && cmd) {
386 		DRM_DEV_ERROR(dev->dev, "%s: offending task: %s (%s)\n",
387 			      gpu->name, comm, cmd);
388 
389 		msm_rd_dump_submit(priv->hangrd, submit,
390 				   "offending task: %s (%s)", comm, cmd);
391 	} else {
392 		DRM_DEV_ERROR(dev->dev, "%s: offending task: unknown\n", gpu->name);
393 
394 		msm_rd_dump_submit(priv->hangrd, submit, NULL);
395 	}
396 
397 	/* Record the crash state */
398 	pm_runtime_get_sync(&gpu->pdev->dev);
399 	msm_gpu_crashstate_capture(gpu, submit, comm, cmd);
400 
401 	kfree(cmd);
402 	kfree(comm);
403 
404 	/*
405 	 * Update all the rings with the latest and greatest fence.. this
406 	 * needs to happen after msm_rd_dump_submit() to ensure that the
407 	 * bo's referenced by the offending submit are still around.
408 	 */
409 	for (i = 0; i < gpu->nr_rings; i++) {
410 		struct msm_ringbuffer *ring = gpu->rb[i];
411 
412 		uint32_t fence = ring->memptrs->fence;
413 
414 		/*
415 		 * For the current (faulting?) ring/submit advance the fence by
416 		 * one more to clear the faulting submit
417 		 */
418 		if (ring == cur_ring)
419 			ring->memptrs->fence = ++fence;
420 
421 		msm_update_fence(ring->fctx, fence);
422 	}
423 
424 	if (msm_gpu_active(gpu)) {
425 		/* retire completed submits, plus the one that hung: */
426 		retire_submits(gpu);
427 
428 		gpu->funcs->recover(gpu);
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 	pm_runtime_put(&gpu->pdev->dev);
446 
447 out_unlock:
448 	mutex_unlock(&gpu->lock);
449 
450 	msm_gpu_retire(gpu);
451 }
452 
453 static void fault_worker(struct kthread_work *work)
454 {
455 	struct msm_gpu *gpu = container_of(work, struct msm_gpu, fault_work);
456 	struct msm_gem_submit *submit;
457 	struct msm_ringbuffer *cur_ring = gpu->funcs->active_ring(gpu);
458 	char *comm = NULL, *cmd = NULL;
459 
460 	mutex_lock(&gpu->lock);
461 
462 	submit = find_submit(cur_ring, cur_ring->memptrs->fence + 1);
463 	if (submit && submit->fault_dumped)
464 		goto resume_smmu;
465 
466 	if (submit) {
467 		get_comm_cmdline(submit, &comm, &cmd);
468 
469 		/*
470 		 * When we get GPU iova faults, we can get 1000s of them,
471 		 * but we really only want to log the first one.
472 		 */
473 		submit->fault_dumped = true;
474 	}
475 
476 	/* Record the crash state */
477 	pm_runtime_get_sync(&gpu->pdev->dev);
478 	msm_gpu_crashstate_capture(gpu, submit, comm, cmd);
479 	pm_runtime_put_sync(&gpu->pdev->dev);
480 
481 	kfree(cmd);
482 	kfree(comm);
483 
484 resume_smmu:
485 	memset(&gpu->fault_info, 0, sizeof(gpu->fault_info));
486 	gpu->aspace->mmu->funcs->resume_translation(gpu->aspace->mmu);
487 
488 	mutex_unlock(&gpu->lock);
489 }
490 
491 static void hangcheck_timer_reset(struct msm_gpu *gpu)
492 {
493 	struct msm_drm_private *priv = gpu->dev->dev_private;
494 	mod_timer(&gpu->hangcheck_timer,
495 			round_jiffies_up(jiffies + msecs_to_jiffies(priv->hangcheck_period)));
496 }
497 
498 static bool made_progress(struct msm_gpu *gpu, struct msm_ringbuffer *ring)
499 {
500 	if (ring->hangcheck_progress_retries >= DRM_MSM_HANGCHECK_PROGRESS_RETRIES)
501 		return false;
502 
503 	if (!gpu->funcs->progress)
504 		return false;
505 
506 	if (!gpu->funcs->progress(gpu, ring))
507 		return false;
508 
509 	ring->hangcheck_progress_retries++;
510 	return true;
511 }
512 
513 static void hangcheck_handler(struct timer_list *t)
514 {
515 	struct msm_gpu *gpu = from_timer(gpu, t, hangcheck_timer);
516 	struct drm_device *dev = gpu->dev;
517 	struct msm_ringbuffer *ring = gpu->funcs->active_ring(gpu);
518 	uint32_t fence = ring->memptrs->fence;
519 
520 	if (fence != ring->hangcheck_fence) {
521 		/* some progress has been made.. ya! */
522 		ring->hangcheck_fence = fence;
523 		ring->hangcheck_progress_retries = 0;
524 	} else if (fence_before(fence, ring->fctx->last_fence) &&
525 			!made_progress(gpu, ring)) {
526 		/* no progress and not done.. hung! */
527 		ring->hangcheck_fence = fence;
528 		ring->hangcheck_progress_retries = 0;
529 		DRM_DEV_ERROR(dev->dev, "%s: hangcheck detected gpu lockup rb %d!\n",
530 				gpu->name, ring->id);
531 		DRM_DEV_ERROR(dev->dev, "%s:     completed fence: %u\n",
532 				gpu->name, fence);
533 		DRM_DEV_ERROR(dev->dev, "%s:     submitted fence: %u\n",
534 				gpu->name, ring->fctx->last_fence);
535 
536 		kthread_queue_work(gpu->worker, &gpu->recover_work);
537 	}
538 
539 	/* if still more pending work, reset the hangcheck timer: */
540 	if (fence_after(ring->fctx->last_fence, ring->hangcheck_fence))
541 		hangcheck_timer_reset(gpu);
542 
543 	/* workaround for missing irq: */
544 	msm_gpu_retire(gpu);
545 }
546 
547 /*
548  * Performance Counters:
549  */
550 
551 /* called under perf_lock */
552 static int update_hw_cntrs(struct msm_gpu *gpu, uint32_t ncntrs, uint32_t *cntrs)
553 {
554 	uint32_t current_cntrs[ARRAY_SIZE(gpu->last_cntrs)];
555 	int i, n = min(ncntrs, gpu->num_perfcntrs);
556 
557 	/* read current values: */
558 	for (i = 0; i < gpu->num_perfcntrs; i++)
559 		current_cntrs[i] = gpu_read(gpu, gpu->perfcntrs[i].sample_reg);
560 
561 	/* update cntrs: */
562 	for (i = 0; i < n; i++)
563 		cntrs[i] = current_cntrs[i] - gpu->last_cntrs[i];
564 
565 	/* save current values: */
566 	for (i = 0; i < gpu->num_perfcntrs; i++)
567 		gpu->last_cntrs[i] = current_cntrs[i];
568 
569 	return n;
570 }
571 
572 static void update_sw_cntrs(struct msm_gpu *gpu)
573 {
574 	ktime_t time;
575 	uint32_t elapsed;
576 	unsigned long flags;
577 
578 	spin_lock_irqsave(&gpu->perf_lock, flags);
579 	if (!gpu->perfcntr_active)
580 		goto out;
581 
582 	time = ktime_get();
583 	elapsed = ktime_to_us(ktime_sub(time, gpu->last_sample.time));
584 
585 	gpu->totaltime += elapsed;
586 	if (gpu->last_sample.active)
587 		gpu->activetime += elapsed;
588 
589 	gpu->last_sample.active = msm_gpu_active(gpu);
590 	gpu->last_sample.time = time;
591 
592 out:
593 	spin_unlock_irqrestore(&gpu->perf_lock, flags);
594 }
595 
596 void msm_gpu_perfcntr_start(struct msm_gpu *gpu)
597 {
598 	unsigned long flags;
599 
600 	pm_runtime_get_sync(&gpu->pdev->dev);
601 
602 	spin_lock_irqsave(&gpu->perf_lock, flags);
603 	/* we could dynamically enable/disable perfcntr registers too.. */
604 	gpu->last_sample.active = msm_gpu_active(gpu);
605 	gpu->last_sample.time = ktime_get();
606 	gpu->activetime = gpu->totaltime = 0;
607 	gpu->perfcntr_active = true;
608 	update_hw_cntrs(gpu, 0, NULL);
609 	spin_unlock_irqrestore(&gpu->perf_lock, flags);
610 }
611 
612 void msm_gpu_perfcntr_stop(struct msm_gpu *gpu)
613 {
614 	gpu->perfcntr_active = false;
615 	pm_runtime_put_sync(&gpu->pdev->dev);
616 }
617 
618 /* returns -errno or # of cntrs sampled */
619 int msm_gpu_perfcntr_sample(struct msm_gpu *gpu, uint32_t *activetime,
620 		uint32_t *totaltime, uint32_t ncntrs, uint32_t *cntrs)
621 {
622 	unsigned long flags;
623 	int ret;
624 
625 	spin_lock_irqsave(&gpu->perf_lock, flags);
626 
627 	if (!gpu->perfcntr_active) {
628 		ret = -EINVAL;
629 		goto out;
630 	}
631 
632 	*activetime = gpu->activetime;
633 	*totaltime = gpu->totaltime;
634 
635 	gpu->activetime = gpu->totaltime = 0;
636 
637 	ret = update_hw_cntrs(gpu, ncntrs, cntrs);
638 
639 out:
640 	spin_unlock_irqrestore(&gpu->perf_lock, flags);
641 
642 	return ret;
643 }
644 
645 /*
646  * Cmdstream submission/retirement:
647  */
648 
649 static void retire_submit(struct msm_gpu *gpu, struct msm_ringbuffer *ring,
650 		struct msm_gem_submit *submit)
651 {
652 	int index = submit->seqno % MSM_GPU_SUBMIT_STATS_COUNT;
653 	volatile struct msm_gpu_submit_stats *stats;
654 	u64 elapsed, clock = 0, cycles;
655 	unsigned long flags;
656 
657 	stats = &ring->memptrs->stats[index];
658 	/* Convert 19.2Mhz alwayson ticks to nanoseconds for elapsed time */
659 	elapsed = (stats->alwayson_end - stats->alwayson_start) * 10000;
660 	do_div(elapsed, 192);
661 
662 	cycles = stats->cpcycles_end - stats->cpcycles_start;
663 
664 	/* Calculate the clock frequency from the number of CP cycles */
665 	if (elapsed) {
666 		clock = cycles * 1000;
667 		do_div(clock, elapsed);
668 	}
669 
670 	submit->queue->ctx->elapsed_ns += elapsed;
671 	submit->queue->ctx->cycles     += cycles;
672 
673 	trace_msm_gpu_submit_retired(submit, elapsed, clock,
674 		stats->alwayson_start, stats->alwayson_end);
675 
676 	msm_submit_retire(submit);
677 
678 	pm_runtime_mark_last_busy(&gpu->pdev->dev);
679 
680 	spin_lock_irqsave(&ring->submit_lock, flags);
681 	list_del(&submit->node);
682 	spin_unlock_irqrestore(&ring->submit_lock, flags);
683 
684 	/* Update devfreq on transition from active->idle: */
685 	mutex_lock(&gpu->active_lock);
686 	gpu->active_submits--;
687 	WARN_ON(gpu->active_submits < 0);
688 	if (!gpu->active_submits) {
689 		msm_devfreq_idle(gpu);
690 		pm_runtime_put_autosuspend(&gpu->pdev->dev);
691 	}
692 
693 	mutex_unlock(&gpu->active_lock);
694 
695 	msm_gem_submit_put(submit);
696 }
697 
698 static void retire_submits(struct msm_gpu *gpu)
699 {
700 	int i;
701 
702 	/* Retire the commits starting with highest priority */
703 	for (i = 0; i < gpu->nr_rings; i++) {
704 		struct msm_ringbuffer *ring = gpu->rb[i];
705 
706 		while (true) {
707 			struct msm_gem_submit *submit = NULL;
708 			unsigned long flags;
709 
710 			spin_lock_irqsave(&ring->submit_lock, flags);
711 			submit = list_first_entry_or_null(&ring->submits,
712 					struct msm_gem_submit, node);
713 			spin_unlock_irqrestore(&ring->submit_lock, flags);
714 
715 			/*
716 			 * If no submit, we are done.  If submit->fence hasn't
717 			 * been signalled, then later submits are not signalled
718 			 * either, so we are also done.
719 			 */
720 			if (submit && dma_fence_is_signaled(submit->hw_fence)) {
721 				retire_submit(gpu, ring, submit);
722 			} else {
723 				break;
724 			}
725 		}
726 	}
727 
728 	wake_up_all(&gpu->retire_event);
729 }
730 
731 static void retire_worker(struct kthread_work *work)
732 {
733 	struct msm_gpu *gpu = container_of(work, struct msm_gpu, retire_work);
734 
735 	retire_submits(gpu);
736 }
737 
738 /* call from irq handler to schedule work to retire bo's */
739 void msm_gpu_retire(struct msm_gpu *gpu)
740 {
741 	int i;
742 
743 	for (i = 0; i < gpu->nr_rings; i++)
744 		msm_update_fence(gpu->rb[i]->fctx, gpu->rb[i]->memptrs->fence);
745 
746 	kthread_queue_work(gpu->worker, &gpu->retire_work);
747 	update_sw_cntrs(gpu);
748 }
749 
750 /* add bo's to gpu's ring, and kick gpu: */
751 void msm_gpu_submit(struct msm_gpu *gpu, struct msm_gem_submit *submit)
752 {
753 	struct msm_ringbuffer *ring = submit->ring;
754 	unsigned long flags;
755 
756 	WARN_ON(!mutex_is_locked(&gpu->lock));
757 
758 	pm_runtime_get_sync(&gpu->pdev->dev);
759 
760 	msm_gpu_hw_init(gpu);
761 
762 	submit->seqno = submit->hw_fence->seqno;
763 
764 	update_sw_cntrs(gpu);
765 
766 	/*
767 	 * ring->submits holds a ref to the submit, to deal with the case
768 	 * that a submit completes before msm_ioctl_gem_submit() returns.
769 	 */
770 	msm_gem_submit_get(submit);
771 
772 	spin_lock_irqsave(&ring->submit_lock, flags);
773 	list_add_tail(&submit->node, &ring->submits);
774 	spin_unlock_irqrestore(&ring->submit_lock, flags);
775 
776 	/* Update devfreq on transition from idle->active: */
777 	mutex_lock(&gpu->active_lock);
778 	if (!gpu->active_submits) {
779 		pm_runtime_get(&gpu->pdev->dev);
780 		msm_devfreq_active(gpu);
781 	}
782 	gpu->active_submits++;
783 	mutex_unlock(&gpu->active_lock);
784 
785 	gpu->funcs->submit(gpu, submit);
786 	gpu->cur_ctx_seqno = submit->queue->ctx->seqno;
787 
788 	pm_runtime_put(&gpu->pdev->dev);
789 	hangcheck_timer_reset(gpu);
790 }
791 
792 /*
793  * Init/Cleanup:
794  */
795 
796 static irqreturn_t irq_handler(int irq, void *data)
797 {
798 	struct msm_gpu *gpu = data;
799 	return gpu->funcs->irq(gpu);
800 }
801 
802 static int get_clocks(struct platform_device *pdev, struct msm_gpu *gpu)
803 {
804 	int ret = devm_clk_bulk_get_all(&pdev->dev, &gpu->grp_clks);
805 
806 	if (ret < 1) {
807 		gpu->nr_clocks = 0;
808 		return ret;
809 	}
810 
811 	gpu->nr_clocks = ret;
812 
813 	gpu->core_clk = msm_clk_bulk_get_clock(gpu->grp_clks,
814 		gpu->nr_clocks, "core");
815 
816 	gpu->rbbmtimer_clk = msm_clk_bulk_get_clock(gpu->grp_clks,
817 		gpu->nr_clocks, "rbbmtimer");
818 
819 	return 0;
820 }
821 
822 /* Return a new address space for a msm_drm_private instance */
823 struct msm_gem_address_space *
824 msm_gpu_create_private_address_space(struct msm_gpu *gpu, struct task_struct *task)
825 {
826 	struct msm_gem_address_space *aspace = NULL;
827 	if (!gpu)
828 		return NULL;
829 
830 	/*
831 	 * If the target doesn't support private address spaces then return
832 	 * the global one
833 	 */
834 	if (gpu->funcs->create_private_address_space) {
835 		aspace = gpu->funcs->create_private_address_space(gpu);
836 		if (!IS_ERR(aspace))
837 			aspace->pid = get_pid(task_pid(task));
838 	}
839 
840 	if (IS_ERR_OR_NULL(aspace))
841 		aspace = msm_gem_address_space_get(gpu->aspace);
842 
843 	return aspace;
844 }
845 
846 int msm_gpu_init(struct drm_device *drm, struct platform_device *pdev,
847 		struct msm_gpu *gpu, const struct msm_gpu_funcs *funcs,
848 		const char *name, struct msm_gpu_config *config)
849 {
850 	struct msm_drm_private *priv = drm->dev_private;
851 	int i, ret, nr_rings = config->nr_rings;
852 	void *memptrs;
853 	uint64_t memptrs_iova;
854 
855 	if (WARN_ON(gpu->num_perfcntrs > ARRAY_SIZE(gpu->last_cntrs)))
856 		gpu->num_perfcntrs = ARRAY_SIZE(gpu->last_cntrs);
857 
858 	gpu->dev = drm;
859 	gpu->funcs = funcs;
860 	gpu->name = name;
861 
862 	gpu->worker = kthread_create_worker(0, "gpu-worker");
863 	if (IS_ERR(gpu->worker)) {
864 		ret = PTR_ERR(gpu->worker);
865 		gpu->worker = NULL;
866 		goto fail;
867 	}
868 
869 	sched_set_fifo_low(gpu->worker->task);
870 
871 	mutex_init(&gpu->active_lock);
872 	mutex_init(&gpu->lock);
873 	init_waitqueue_head(&gpu->retire_event);
874 	kthread_init_work(&gpu->retire_work, retire_worker);
875 	kthread_init_work(&gpu->recover_work, recover_worker);
876 	kthread_init_work(&gpu->fault_work, fault_worker);
877 
878 	priv->hangcheck_period = DRM_MSM_HANGCHECK_DEFAULT_PERIOD;
879 
880 	/*
881 	 * If progress detection is supported, halve the hangcheck timer
882 	 * duration, as it takes two iterations of the hangcheck handler
883 	 * to detect a hang.
884 	 */
885 	if (funcs->progress)
886 		priv->hangcheck_period /= 2;
887 
888 	timer_setup(&gpu->hangcheck_timer, hangcheck_handler, 0);
889 
890 	spin_lock_init(&gpu->perf_lock);
891 
892 
893 	/* Map registers: */
894 	gpu->mmio = msm_ioremap(pdev, config->ioname);
895 	if (IS_ERR(gpu->mmio)) {
896 		ret = PTR_ERR(gpu->mmio);
897 		goto fail;
898 	}
899 
900 	/* Get Interrupt: */
901 	gpu->irq = platform_get_irq(pdev, 0);
902 	if (gpu->irq < 0) {
903 		ret = gpu->irq;
904 		goto fail;
905 	}
906 
907 	ret = devm_request_irq(&pdev->dev, gpu->irq, irq_handler,
908 			IRQF_TRIGGER_HIGH, "gpu-irq", gpu);
909 	if (ret) {
910 		DRM_DEV_ERROR(drm->dev, "failed to request IRQ%u: %d\n", gpu->irq, ret);
911 		goto fail;
912 	}
913 
914 	ret = get_clocks(pdev, gpu);
915 	if (ret)
916 		goto fail;
917 
918 	gpu->ebi1_clk = msm_clk_get(pdev, "bus");
919 	DBG("ebi1_clk: %p", gpu->ebi1_clk);
920 	if (IS_ERR(gpu->ebi1_clk))
921 		gpu->ebi1_clk = NULL;
922 
923 	/* Acquire regulators: */
924 	gpu->gpu_reg = devm_regulator_get(&pdev->dev, "vdd");
925 	DBG("gpu_reg: %p", gpu->gpu_reg);
926 	if (IS_ERR(gpu->gpu_reg))
927 		gpu->gpu_reg = NULL;
928 
929 	gpu->gpu_cx = devm_regulator_get(&pdev->dev, "vddcx");
930 	DBG("gpu_cx: %p", gpu->gpu_cx);
931 	if (IS_ERR(gpu->gpu_cx))
932 		gpu->gpu_cx = NULL;
933 
934 	gpu->pdev = pdev;
935 	platform_set_drvdata(pdev, &gpu->adreno_smmu);
936 
937 	msm_devfreq_init(gpu);
938 
939 
940 	gpu->aspace = gpu->funcs->create_address_space(gpu, pdev);
941 
942 	if (gpu->aspace == NULL)
943 		DRM_DEV_INFO(drm->dev, "%s: no IOMMU, fallback to VRAM carveout!\n", name);
944 	else if (IS_ERR(gpu->aspace)) {
945 		ret = PTR_ERR(gpu->aspace);
946 		goto fail;
947 	}
948 
949 	memptrs = msm_gem_kernel_new(drm,
950 		sizeof(struct msm_rbmemptrs) * nr_rings,
951 		check_apriv(gpu, MSM_BO_WC), gpu->aspace, &gpu->memptrs_bo,
952 		&memptrs_iova);
953 
954 	if (IS_ERR(memptrs)) {
955 		ret = PTR_ERR(memptrs);
956 		DRM_DEV_ERROR(drm->dev, "could not allocate memptrs: %d\n", ret);
957 		goto fail;
958 	}
959 
960 	msm_gem_object_set_name(gpu->memptrs_bo, "memptrs");
961 
962 	if (nr_rings > ARRAY_SIZE(gpu->rb)) {
963 		DRM_DEV_INFO_ONCE(drm->dev, "Only creating %zu ringbuffers\n",
964 			ARRAY_SIZE(gpu->rb));
965 		nr_rings = ARRAY_SIZE(gpu->rb);
966 	}
967 
968 	/* Create ringbuffer(s): */
969 	for (i = 0; i < nr_rings; i++) {
970 		gpu->rb[i] = msm_ringbuffer_new(gpu, i, memptrs, memptrs_iova);
971 
972 		if (IS_ERR(gpu->rb[i])) {
973 			ret = PTR_ERR(gpu->rb[i]);
974 			DRM_DEV_ERROR(drm->dev,
975 				"could not create ringbuffer %d: %d\n", i, ret);
976 			goto fail;
977 		}
978 
979 		memptrs += sizeof(struct msm_rbmemptrs);
980 		memptrs_iova += sizeof(struct msm_rbmemptrs);
981 	}
982 
983 	gpu->nr_rings = nr_rings;
984 
985 	refcount_set(&gpu->sysprof_active, 1);
986 
987 	return 0;
988 
989 fail:
990 	for (i = 0; i < ARRAY_SIZE(gpu->rb); i++)  {
991 		msm_ringbuffer_destroy(gpu->rb[i]);
992 		gpu->rb[i] = NULL;
993 	}
994 
995 	msm_gem_kernel_put(gpu->memptrs_bo, gpu->aspace);
996 
997 	platform_set_drvdata(pdev, NULL);
998 	return ret;
999 }
1000 
1001 void msm_gpu_cleanup(struct msm_gpu *gpu)
1002 {
1003 	int i;
1004 
1005 	DBG("%s", gpu->name);
1006 
1007 	for (i = 0; i < ARRAY_SIZE(gpu->rb); i++) {
1008 		msm_ringbuffer_destroy(gpu->rb[i]);
1009 		gpu->rb[i] = NULL;
1010 	}
1011 
1012 	msm_gem_kernel_put(gpu->memptrs_bo, gpu->aspace);
1013 
1014 	if (!IS_ERR_OR_NULL(gpu->aspace)) {
1015 		gpu->aspace->mmu->funcs->detach(gpu->aspace->mmu);
1016 		msm_gem_address_space_put(gpu->aspace);
1017 	}
1018 
1019 	if (gpu->worker) {
1020 		kthread_destroy_worker(gpu->worker);
1021 	}
1022 
1023 	msm_devfreq_cleanup(gpu);
1024 
1025 	platform_set_drvdata(gpu->pdev, NULL);
1026 }
1027