xref: /openbmc/linux/drivers/gpu/drm/msm/msm_gpu.c (revision 249592bf)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * Copyright (C) 2013 Red Hat
4  * Author: Rob Clark <robdclark@gmail.com>
5  */
6 
7 #include "msm_gpu.h"
8 #include "msm_gem.h"
9 #include "msm_mmu.h"
10 #include "msm_fence.h"
11 #include "msm_gpu_trace.h"
12 #include "adreno/adreno_gpu.h"
13 
14 #include <generated/utsrelease.h>
15 #include <linux/string_helpers.h>
16 #include <linux/devfreq.h>
17 #include <linux/devfreq_cooling.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 = dev_to_gpu(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 	trace_msm_gpu_freq_change(dev_pm_opp_get_freq(opp));
37 
38 	if (gpu->funcs->gpu_set_freq)
39 		gpu->funcs->gpu_set_freq(gpu, opp);
40 	else
41 		clk_set_rate(gpu->core_clk, *freq);
42 
43 	dev_pm_opp_put(opp);
44 
45 	return 0;
46 }
47 
48 static int msm_devfreq_get_dev_status(struct device *dev,
49 		struct devfreq_dev_status *status)
50 {
51 	struct msm_gpu *gpu = dev_to_gpu(dev);
52 	ktime_t time;
53 
54 	if (gpu->funcs->gpu_get_freq)
55 		status->current_frequency = gpu->funcs->gpu_get_freq(gpu);
56 	else
57 		status->current_frequency = clk_get_rate(gpu->core_clk);
58 
59 	status->busy_time = gpu->funcs->gpu_busy(gpu);
60 
61 	time = ktime_get();
62 	status->total_time = ktime_us_delta(time, gpu->devfreq.time);
63 	gpu->devfreq.time = time;
64 
65 	return 0;
66 }
67 
68 static int msm_devfreq_get_cur_freq(struct device *dev, unsigned long *freq)
69 {
70 	struct msm_gpu *gpu = dev_to_gpu(dev);
71 
72 	if (gpu->funcs->gpu_get_freq)
73 		*freq = gpu->funcs->gpu_get_freq(gpu);
74 	else
75 		*freq = clk_get_rate(gpu->core_clk);
76 
77 	return 0;
78 }
79 
80 static struct devfreq_dev_profile msm_devfreq_profile = {
81 	.polling_ms = 10,
82 	.target = msm_devfreq_target,
83 	.get_dev_status = msm_devfreq_get_dev_status,
84 	.get_cur_freq = msm_devfreq_get_cur_freq,
85 };
86 
87 static void msm_devfreq_init(struct msm_gpu *gpu)
88 {
89 	/* We need target support to do devfreq */
90 	if (!gpu->funcs->gpu_busy)
91 		return;
92 
93 	msm_devfreq_profile.initial_freq = gpu->fast_rate;
94 
95 	/*
96 	 * Don't set the freq_table or max_state and let devfreq build the table
97 	 * from OPP
98 	 * After a deferred probe, these may have be left to non-zero values,
99 	 * so set them back to zero before creating the devfreq device
100 	 */
101 	msm_devfreq_profile.freq_table = NULL;
102 	msm_devfreq_profile.max_state = 0;
103 
104 	gpu->devfreq.devfreq = devm_devfreq_add_device(&gpu->pdev->dev,
105 			&msm_devfreq_profile, DEVFREQ_GOV_SIMPLE_ONDEMAND,
106 			NULL);
107 
108 	if (IS_ERR(gpu->devfreq.devfreq)) {
109 		DRM_DEV_ERROR(&gpu->pdev->dev, "Couldn't initialize GPU devfreq\n");
110 		gpu->devfreq.devfreq = NULL;
111 		return;
112 	}
113 
114 	devfreq_suspend_device(gpu->devfreq.devfreq);
115 
116 	gpu->cooling = of_devfreq_cooling_register(gpu->pdev->dev.of_node,
117 			gpu->devfreq.devfreq);
118 	if (IS_ERR(gpu->cooling)) {
119 		DRM_DEV_ERROR(&gpu->pdev->dev,
120 				"Couldn't register GPU cooling device\n");
121 		gpu->cooling = NULL;
122 	}
123 }
124 
125 static int enable_pwrrail(struct msm_gpu *gpu)
126 {
127 	struct drm_device *dev = gpu->dev;
128 	int ret = 0;
129 
130 	if (gpu->gpu_reg) {
131 		ret = regulator_enable(gpu->gpu_reg);
132 		if (ret) {
133 			DRM_DEV_ERROR(dev->dev, "failed to enable 'gpu_reg': %d\n", ret);
134 			return ret;
135 		}
136 	}
137 
138 	if (gpu->gpu_cx) {
139 		ret = regulator_enable(gpu->gpu_cx);
140 		if (ret) {
141 			DRM_DEV_ERROR(dev->dev, "failed to enable 'gpu_cx': %d\n", ret);
142 			return ret;
143 		}
144 	}
145 
146 	return 0;
147 }
148 
149 static int disable_pwrrail(struct msm_gpu *gpu)
150 {
151 	if (gpu->gpu_cx)
152 		regulator_disable(gpu->gpu_cx);
153 	if (gpu->gpu_reg)
154 		regulator_disable(gpu->gpu_reg);
155 	return 0;
156 }
157 
158 static int enable_clk(struct msm_gpu *gpu)
159 {
160 	if (gpu->core_clk && gpu->fast_rate)
161 		clk_set_rate(gpu->core_clk, gpu->fast_rate);
162 
163 	/* Set the RBBM timer rate to 19.2Mhz */
164 	if (gpu->rbbmtimer_clk)
165 		clk_set_rate(gpu->rbbmtimer_clk, 19200000);
166 
167 	return clk_bulk_prepare_enable(gpu->nr_clocks, gpu->grp_clks);
168 }
169 
170 static int disable_clk(struct msm_gpu *gpu)
171 {
172 	clk_bulk_disable_unprepare(gpu->nr_clocks, gpu->grp_clks);
173 
174 	/*
175 	 * Set the clock to a deliberately low rate. On older targets the clock
176 	 * speed had to be non zero to avoid problems. On newer targets this
177 	 * will be rounded down to zero anyway so it all works out.
178 	 */
179 	if (gpu->core_clk)
180 		clk_set_rate(gpu->core_clk, 27000000);
181 
182 	if (gpu->rbbmtimer_clk)
183 		clk_set_rate(gpu->rbbmtimer_clk, 0);
184 
185 	return 0;
186 }
187 
188 static int enable_axi(struct msm_gpu *gpu)
189 {
190 	return clk_prepare_enable(gpu->ebi1_clk);
191 }
192 
193 static int disable_axi(struct msm_gpu *gpu)
194 {
195 	clk_disable_unprepare(gpu->ebi1_clk);
196 	return 0;
197 }
198 
199 void msm_gpu_resume_devfreq(struct msm_gpu *gpu)
200 {
201 	gpu->devfreq.busy_cycles = 0;
202 	gpu->devfreq.time = ktime_get();
203 
204 	devfreq_resume_device(gpu->devfreq.devfreq);
205 }
206 
207 int msm_gpu_pm_resume(struct msm_gpu *gpu)
208 {
209 	int ret;
210 
211 	DBG("%s", gpu->name);
212 	trace_msm_gpu_resume(0);
213 
214 	ret = enable_pwrrail(gpu);
215 	if (ret)
216 		return ret;
217 
218 	ret = enable_clk(gpu);
219 	if (ret)
220 		return ret;
221 
222 	ret = enable_axi(gpu);
223 	if (ret)
224 		return ret;
225 
226 	msm_gpu_resume_devfreq(gpu);
227 
228 	gpu->needs_hw_init = true;
229 
230 	return 0;
231 }
232 
233 int msm_gpu_pm_suspend(struct msm_gpu *gpu)
234 {
235 	int ret;
236 
237 	DBG("%s", gpu->name);
238 	trace_msm_gpu_suspend(0);
239 
240 	devfreq_suspend_device(gpu->devfreq.devfreq);
241 
242 	ret = disable_axi(gpu);
243 	if (ret)
244 		return ret;
245 
246 	ret = disable_clk(gpu);
247 	if (ret)
248 		return ret;
249 
250 	ret = disable_pwrrail(gpu);
251 	if (ret)
252 		return ret;
253 
254 	gpu->suspend_count++;
255 
256 	return 0;
257 }
258 
259 int msm_gpu_hw_init(struct msm_gpu *gpu)
260 {
261 	int ret;
262 
263 	WARN_ON(!mutex_is_locked(&gpu->dev->struct_mutex));
264 
265 	if (!gpu->needs_hw_init)
266 		return 0;
267 
268 	disable_irq(gpu->irq);
269 	ret = gpu->funcs->hw_init(gpu);
270 	if (!ret)
271 		gpu->needs_hw_init = false;
272 	enable_irq(gpu->irq);
273 
274 	return ret;
275 }
276 
277 static void update_fences(struct msm_gpu *gpu, struct msm_ringbuffer *ring,
278 		uint32_t fence)
279 {
280 	struct msm_gem_submit *submit;
281 
282 	spin_lock(&ring->submit_lock);
283 	list_for_each_entry(submit, &ring->submits, node) {
284 		if (submit->seqno > fence)
285 			break;
286 
287 		msm_update_fence(submit->ring->fctx,
288 			submit->fence->seqno);
289 	}
290 	spin_unlock(&ring->submit_lock);
291 }
292 
293 #ifdef CONFIG_DEV_COREDUMP
294 static ssize_t msm_gpu_devcoredump_read(char *buffer, loff_t offset,
295 		size_t count, void *data, size_t datalen)
296 {
297 	struct msm_gpu *gpu = data;
298 	struct drm_print_iterator iter;
299 	struct drm_printer p;
300 	struct msm_gpu_state *state;
301 
302 	state = msm_gpu_crashstate_get(gpu);
303 	if (!state)
304 		return 0;
305 
306 	iter.data = buffer;
307 	iter.offset = 0;
308 	iter.start = offset;
309 	iter.remain = count;
310 
311 	p = drm_coredump_printer(&iter);
312 
313 	drm_printf(&p, "---\n");
314 	drm_printf(&p, "kernel: " UTS_RELEASE "\n");
315 	drm_printf(&p, "module: " KBUILD_MODNAME "\n");
316 	drm_printf(&p, "time: %lld.%09ld\n",
317 		state->time.tv_sec, state->time.tv_nsec);
318 	if (state->comm)
319 		drm_printf(&p, "comm: %s\n", state->comm);
320 	if (state->cmd)
321 		drm_printf(&p, "cmdline: %s\n", state->cmd);
322 
323 	gpu->funcs->show(gpu, state, &p);
324 
325 	msm_gpu_crashstate_put(gpu);
326 
327 	return count - iter.remain;
328 }
329 
330 static void msm_gpu_devcoredump_free(void *data)
331 {
332 	struct msm_gpu *gpu = data;
333 
334 	msm_gpu_crashstate_put(gpu);
335 }
336 
337 static void msm_gpu_crashstate_get_bo(struct msm_gpu_state *state,
338 		struct msm_gem_object *obj, u64 iova, u32 flags)
339 {
340 	struct msm_gpu_state_bo *state_bo = &state->bos[state->nr_bos];
341 
342 	/* Don't record write only objects */
343 	state_bo->size = obj->base.size;
344 	state_bo->iova = iova;
345 
346 	/* Only store data for non imported buffer objects marked for read */
347 	if ((flags & MSM_SUBMIT_BO_READ) && !obj->base.import_attach) {
348 		void *ptr;
349 
350 		state_bo->data = kvmalloc(obj->base.size, GFP_KERNEL);
351 		if (!state_bo->data)
352 			goto out;
353 
354 		msm_gem_lock(&obj->base);
355 		ptr = msm_gem_get_vaddr_active(&obj->base);
356 		msm_gem_unlock(&obj->base);
357 		if (IS_ERR(ptr)) {
358 			kvfree(state_bo->data);
359 			state_bo->data = NULL;
360 			goto out;
361 		}
362 
363 		memcpy(state_bo->data, ptr, obj->base.size);
364 		msm_gem_put_vaddr(&obj->base);
365 	}
366 out:
367 	state->nr_bos++;
368 }
369 
370 static void msm_gpu_crashstate_capture(struct msm_gpu *gpu,
371 		struct msm_gem_submit *submit, char *comm, char *cmd)
372 {
373 	struct msm_gpu_state *state;
374 
375 	/* Check if the target supports capturing crash state */
376 	if (!gpu->funcs->gpu_state_get)
377 		return;
378 
379 	/* Only save one crash state at a time */
380 	if (gpu->crashstate)
381 		return;
382 
383 	state = gpu->funcs->gpu_state_get(gpu);
384 	if (IS_ERR_OR_NULL(state))
385 		return;
386 
387 	/* Fill in the additional crash state information */
388 	state->comm = kstrdup(comm, GFP_KERNEL);
389 	state->cmd = kstrdup(cmd, GFP_KERNEL);
390 
391 	if (submit) {
392 		int i, nr = 0;
393 
394 		/* count # of buffers to dump: */
395 		for (i = 0; i < submit->nr_bos; i++)
396 			if (should_dump(submit, i))
397 				nr++;
398 		/* always dump cmd bo's, but don't double count them: */
399 		for (i = 0; i < submit->nr_cmds; i++)
400 			if (!should_dump(submit, submit->cmd[i].idx))
401 				nr++;
402 
403 		state->bos = kcalloc(nr,
404 			sizeof(struct msm_gpu_state_bo), GFP_KERNEL);
405 
406 		for (i = 0; i < submit->nr_bos; i++) {
407 			if (should_dump(submit, i)) {
408 				msm_gpu_crashstate_get_bo(state, submit->bos[i].obj,
409 					submit->bos[i].iova, submit->bos[i].flags);
410 			}
411 		}
412 
413 		for (i = 0; state->bos && i < submit->nr_cmds; i++) {
414 			int idx = submit->cmd[i].idx;
415 
416 			if (!should_dump(submit, submit->cmd[i].idx)) {
417 				msm_gpu_crashstate_get_bo(state, submit->bos[idx].obj,
418 					submit->bos[idx].iova, submit->bos[idx].flags);
419 			}
420 		}
421 	}
422 
423 	/* Set the active crash state to be dumped on failure */
424 	gpu->crashstate = state;
425 
426 	/* FIXME: Release the crashstate if this errors out? */
427 	dev_coredumpm(gpu->dev->dev, THIS_MODULE, gpu, 0, GFP_KERNEL,
428 		msm_gpu_devcoredump_read, msm_gpu_devcoredump_free);
429 }
430 #else
431 static void msm_gpu_crashstate_capture(struct msm_gpu *gpu,
432 		struct msm_gem_submit *submit, char *comm, char *cmd)
433 {
434 }
435 #endif
436 
437 /*
438  * Hangcheck detection for locked gpu:
439  */
440 
441 static struct msm_gem_submit *
442 find_submit(struct msm_ringbuffer *ring, uint32_t fence)
443 {
444 	struct msm_gem_submit *submit;
445 
446 	spin_lock(&ring->submit_lock);
447 	list_for_each_entry(submit, &ring->submits, node) {
448 		if (submit->seqno == fence) {
449 			spin_unlock(&ring->submit_lock);
450 			return submit;
451 		}
452 	}
453 	spin_unlock(&ring->submit_lock);
454 
455 	return NULL;
456 }
457 
458 static void retire_submits(struct msm_gpu *gpu);
459 
460 static void recover_worker(struct kthread_work *work)
461 {
462 	struct msm_gpu *gpu = container_of(work, struct msm_gpu, recover_work);
463 	struct drm_device *dev = gpu->dev;
464 	struct msm_drm_private *priv = dev->dev_private;
465 	struct msm_gem_submit *submit;
466 	struct msm_ringbuffer *cur_ring = gpu->funcs->active_ring(gpu);
467 	char *comm = NULL, *cmd = NULL;
468 	int i;
469 
470 	mutex_lock(&dev->struct_mutex);
471 
472 	DRM_DEV_ERROR(dev->dev, "%s: hangcheck recover!\n", gpu->name);
473 
474 	submit = find_submit(cur_ring, cur_ring->memptrs->fence + 1);
475 	if (submit) {
476 		struct task_struct *task;
477 
478 		/* Increment the fault counts */
479 		gpu->global_faults++;
480 		submit->queue->faults++;
481 
482 		task = get_pid_task(submit->pid, PIDTYPE_PID);
483 		if (task) {
484 			comm = kstrdup(task->comm, GFP_KERNEL);
485 			cmd = kstrdup_quotable_cmdline(task, GFP_KERNEL);
486 			put_task_struct(task);
487 		}
488 
489 		/* msm_rd_dump_submit() needs bo locked to dump: */
490 		for (i = 0; i < submit->nr_bos; i++)
491 			msm_gem_lock(&submit->bos[i].obj->base);
492 
493 		if (comm && cmd) {
494 			DRM_DEV_ERROR(dev->dev, "%s: offending task: %s (%s)\n",
495 				gpu->name, comm, cmd);
496 
497 			msm_rd_dump_submit(priv->hangrd, submit,
498 				"offending task: %s (%s)", comm, cmd);
499 		} else {
500 			msm_rd_dump_submit(priv->hangrd, submit, NULL);
501 		}
502 
503 		for (i = 0; i < submit->nr_bos; i++)
504 			msm_gem_unlock(&submit->bos[i].obj->base);
505 	}
506 
507 	/* Record the crash state */
508 	pm_runtime_get_sync(&gpu->pdev->dev);
509 	msm_gpu_crashstate_capture(gpu, submit, comm, cmd);
510 	pm_runtime_put_sync(&gpu->pdev->dev);
511 
512 	kfree(cmd);
513 	kfree(comm);
514 
515 	/*
516 	 * Update all the rings with the latest and greatest fence.. this
517 	 * needs to happen after msm_rd_dump_submit() to ensure that the
518 	 * bo's referenced by the offending submit are still around.
519 	 */
520 	for (i = 0; i < gpu->nr_rings; i++) {
521 		struct msm_ringbuffer *ring = gpu->rb[i];
522 
523 		uint32_t fence = ring->memptrs->fence;
524 
525 		/*
526 		 * For the current (faulting?) ring/submit advance the fence by
527 		 * one more to clear the faulting submit
528 		 */
529 		if (ring == cur_ring)
530 			fence++;
531 
532 		update_fences(gpu, ring, fence);
533 	}
534 
535 	if (msm_gpu_active(gpu)) {
536 		/* retire completed submits, plus the one that hung: */
537 		retire_submits(gpu);
538 
539 		pm_runtime_get_sync(&gpu->pdev->dev);
540 		gpu->funcs->recover(gpu);
541 		pm_runtime_put_sync(&gpu->pdev->dev);
542 
543 		/*
544 		 * Replay all remaining submits starting with highest priority
545 		 * ring
546 		 */
547 		for (i = 0; i < gpu->nr_rings; i++) {
548 			struct msm_ringbuffer *ring = gpu->rb[i];
549 
550 			spin_lock(&ring->submit_lock);
551 			list_for_each_entry(submit, &ring->submits, node)
552 				gpu->funcs->submit(gpu, submit);
553 			spin_unlock(&ring->submit_lock);
554 		}
555 	}
556 
557 	mutex_unlock(&dev->struct_mutex);
558 
559 	msm_gpu_retire(gpu);
560 }
561 
562 static void hangcheck_timer_reset(struct msm_gpu *gpu)
563 {
564 	mod_timer(&gpu->hangcheck_timer,
565 			round_jiffies_up(jiffies + DRM_MSM_HANGCHECK_JIFFIES));
566 }
567 
568 static void hangcheck_handler(struct timer_list *t)
569 {
570 	struct msm_gpu *gpu = from_timer(gpu, t, hangcheck_timer);
571 	struct drm_device *dev = gpu->dev;
572 	struct msm_ringbuffer *ring = gpu->funcs->active_ring(gpu);
573 	uint32_t fence = ring->memptrs->fence;
574 
575 	if (fence != ring->hangcheck_fence) {
576 		/* some progress has been made.. ya! */
577 		ring->hangcheck_fence = fence;
578 	} else if (fence < ring->seqno) {
579 		/* no progress and not done.. hung! */
580 		ring->hangcheck_fence = fence;
581 		DRM_DEV_ERROR(dev->dev, "%s: hangcheck detected gpu lockup rb %d!\n",
582 				gpu->name, ring->id);
583 		DRM_DEV_ERROR(dev->dev, "%s:     completed fence: %u\n",
584 				gpu->name, fence);
585 		DRM_DEV_ERROR(dev->dev, "%s:     submitted fence: %u\n",
586 				gpu->name, ring->seqno);
587 
588 		kthread_queue_work(gpu->worker, &gpu->recover_work);
589 	}
590 
591 	/* if still more pending work, reset the hangcheck timer: */
592 	if (ring->seqno > ring->hangcheck_fence)
593 		hangcheck_timer_reset(gpu);
594 
595 	/* workaround for missing irq: */
596 	kthread_queue_work(gpu->worker, &gpu->retire_work);
597 }
598 
599 /*
600  * Performance Counters:
601  */
602 
603 /* called under perf_lock */
604 static int update_hw_cntrs(struct msm_gpu *gpu, uint32_t ncntrs, uint32_t *cntrs)
605 {
606 	uint32_t current_cntrs[ARRAY_SIZE(gpu->last_cntrs)];
607 	int i, n = min(ncntrs, gpu->num_perfcntrs);
608 
609 	/* read current values: */
610 	for (i = 0; i < gpu->num_perfcntrs; i++)
611 		current_cntrs[i] = gpu_read(gpu, gpu->perfcntrs[i].sample_reg);
612 
613 	/* update cntrs: */
614 	for (i = 0; i < n; i++)
615 		cntrs[i] = current_cntrs[i] - gpu->last_cntrs[i];
616 
617 	/* save current values: */
618 	for (i = 0; i < gpu->num_perfcntrs; i++)
619 		gpu->last_cntrs[i] = current_cntrs[i];
620 
621 	return n;
622 }
623 
624 static void update_sw_cntrs(struct msm_gpu *gpu)
625 {
626 	ktime_t time;
627 	uint32_t elapsed;
628 	unsigned long flags;
629 
630 	spin_lock_irqsave(&gpu->perf_lock, flags);
631 	if (!gpu->perfcntr_active)
632 		goto out;
633 
634 	time = ktime_get();
635 	elapsed = ktime_to_us(ktime_sub(time, gpu->last_sample.time));
636 
637 	gpu->totaltime += elapsed;
638 	if (gpu->last_sample.active)
639 		gpu->activetime += elapsed;
640 
641 	gpu->last_sample.active = msm_gpu_active(gpu);
642 	gpu->last_sample.time = time;
643 
644 out:
645 	spin_unlock_irqrestore(&gpu->perf_lock, flags);
646 }
647 
648 void msm_gpu_perfcntr_start(struct msm_gpu *gpu)
649 {
650 	unsigned long flags;
651 
652 	pm_runtime_get_sync(&gpu->pdev->dev);
653 
654 	spin_lock_irqsave(&gpu->perf_lock, flags);
655 	/* we could dynamically enable/disable perfcntr registers too.. */
656 	gpu->last_sample.active = msm_gpu_active(gpu);
657 	gpu->last_sample.time = ktime_get();
658 	gpu->activetime = gpu->totaltime = 0;
659 	gpu->perfcntr_active = true;
660 	update_hw_cntrs(gpu, 0, NULL);
661 	spin_unlock_irqrestore(&gpu->perf_lock, flags);
662 }
663 
664 void msm_gpu_perfcntr_stop(struct msm_gpu *gpu)
665 {
666 	gpu->perfcntr_active = false;
667 	pm_runtime_put_sync(&gpu->pdev->dev);
668 }
669 
670 /* returns -errno or # of cntrs sampled */
671 int msm_gpu_perfcntr_sample(struct msm_gpu *gpu, uint32_t *activetime,
672 		uint32_t *totaltime, uint32_t ncntrs, uint32_t *cntrs)
673 {
674 	unsigned long flags;
675 	int ret;
676 
677 	spin_lock_irqsave(&gpu->perf_lock, flags);
678 
679 	if (!gpu->perfcntr_active) {
680 		ret = -EINVAL;
681 		goto out;
682 	}
683 
684 	*activetime = gpu->activetime;
685 	*totaltime = gpu->totaltime;
686 
687 	gpu->activetime = gpu->totaltime = 0;
688 
689 	ret = update_hw_cntrs(gpu, ncntrs, cntrs);
690 
691 out:
692 	spin_unlock_irqrestore(&gpu->perf_lock, flags);
693 
694 	return ret;
695 }
696 
697 /*
698  * Cmdstream submission/retirement:
699  */
700 
701 static void retire_submit(struct msm_gpu *gpu, struct msm_ringbuffer *ring,
702 		struct msm_gem_submit *submit)
703 {
704 	int index = submit->seqno % MSM_GPU_SUBMIT_STATS_COUNT;
705 	volatile struct msm_gpu_submit_stats *stats;
706 	u64 elapsed, clock = 0;
707 	int i;
708 
709 	stats = &ring->memptrs->stats[index];
710 	/* Convert 19.2Mhz alwayson ticks to nanoseconds for elapsed time */
711 	elapsed = (stats->alwayson_end - stats->alwayson_start) * 10000;
712 	do_div(elapsed, 192);
713 
714 	/* Calculate the clock frequency from the number of CP cycles */
715 	if (elapsed) {
716 		clock = (stats->cpcycles_end - stats->cpcycles_start) * 1000;
717 		do_div(clock, elapsed);
718 	}
719 
720 	trace_msm_gpu_submit_retired(submit, elapsed, clock,
721 		stats->alwayson_start, stats->alwayson_end);
722 
723 	for (i = 0; i < submit->nr_bos; i++) {
724 		struct drm_gem_object *obj = &submit->bos[i].obj->base;
725 
726 		msm_gem_lock(obj);
727 		msm_gem_active_put(obj);
728 		msm_gem_unpin_iova_locked(obj, submit->aspace);
729 		msm_gem_unlock(obj);
730 		drm_gem_object_put(obj);
731 	}
732 
733 	pm_runtime_mark_last_busy(&gpu->pdev->dev);
734 	pm_runtime_put_autosuspend(&gpu->pdev->dev);
735 
736 	spin_lock(&ring->submit_lock);
737 	list_del(&submit->node);
738 	spin_unlock(&ring->submit_lock);
739 
740 	msm_gem_submit_put(submit);
741 }
742 
743 static void retire_submits(struct msm_gpu *gpu)
744 {
745 	int i;
746 
747 	/* Retire the commits starting with highest priority */
748 	for (i = 0; i < gpu->nr_rings; i++) {
749 		struct msm_ringbuffer *ring = gpu->rb[i];
750 
751 		while (true) {
752 			struct msm_gem_submit *submit = NULL;
753 
754 			spin_lock(&ring->submit_lock);
755 			submit = list_first_entry_or_null(&ring->submits,
756 					struct msm_gem_submit, node);
757 			spin_unlock(&ring->submit_lock);
758 
759 			/*
760 			 * If no submit, we are done.  If submit->fence hasn't
761 			 * been signalled, then later submits are not signalled
762 			 * either, so we are also done.
763 			 */
764 			if (submit && dma_fence_is_signaled(submit->fence)) {
765 				retire_submit(gpu, ring, submit);
766 			} else {
767 				break;
768 			}
769 		}
770 	}
771 }
772 
773 static void retire_worker(struct kthread_work *work)
774 {
775 	struct msm_gpu *gpu = container_of(work, struct msm_gpu, retire_work);
776 	int i;
777 
778 	for (i = 0; i < gpu->nr_rings; i++)
779 		update_fences(gpu, gpu->rb[i], gpu->rb[i]->memptrs->fence);
780 
781 	retire_submits(gpu);
782 }
783 
784 /* call from irq handler to schedule work to retire bo's */
785 void msm_gpu_retire(struct msm_gpu *gpu)
786 {
787 	kthread_queue_work(gpu->worker, &gpu->retire_work);
788 	update_sw_cntrs(gpu);
789 }
790 
791 /* add bo's to gpu's ring, and kick gpu: */
792 void msm_gpu_submit(struct msm_gpu *gpu, struct msm_gem_submit *submit)
793 {
794 	struct drm_device *dev = gpu->dev;
795 	struct msm_drm_private *priv = dev->dev_private;
796 	struct msm_ringbuffer *ring = submit->ring;
797 	int i;
798 
799 	WARN_ON(!mutex_is_locked(&dev->struct_mutex));
800 
801 	pm_runtime_get_sync(&gpu->pdev->dev);
802 
803 	msm_gpu_hw_init(gpu);
804 
805 	submit->seqno = ++ring->seqno;
806 
807 	msm_rd_dump_submit(priv->rd, submit, NULL);
808 
809 	update_sw_cntrs(gpu);
810 
811 	for (i = 0; i < submit->nr_bos; i++) {
812 		struct msm_gem_object *msm_obj = submit->bos[i].obj;
813 		struct drm_gem_object *drm_obj = &msm_obj->base;
814 		uint64_t iova;
815 
816 		/* submit takes a reference to the bo and iova until retired: */
817 		drm_gem_object_get(&msm_obj->base);
818 		msm_gem_get_and_pin_iova_locked(&msm_obj->base, submit->aspace, &iova);
819 
820 		if (submit->bos[i].flags & MSM_SUBMIT_BO_WRITE)
821 			dma_resv_add_excl_fence(drm_obj->resv, submit->fence);
822 		else if (submit->bos[i].flags & MSM_SUBMIT_BO_READ)
823 			dma_resv_add_shared_fence(drm_obj->resv, submit->fence);
824 
825 		msm_gem_active_get(drm_obj, gpu);
826 	}
827 
828 	/*
829 	 * ring->submits holds a ref to the submit, to deal with the case
830 	 * that a submit completes before msm_ioctl_gem_submit() returns.
831 	 */
832 	msm_gem_submit_get(submit);
833 
834 	spin_lock(&ring->submit_lock);
835 	list_add_tail(&submit->node, &ring->submits);
836 	spin_unlock(&ring->submit_lock);
837 
838 	gpu->funcs->submit(gpu, submit);
839 	priv->lastctx = submit->queue->ctx;
840 
841 	hangcheck_timer_reset(gpu);
842 }
843 
844 /*
845  * Init/Cleanup:
846  */
847 
848 static irqreturn_t irq_handler(int irq, void *data)
849 {
850 	struct msm_gpu *gpu = data;
851 	return gpu->funcs->irq(gpu);
852 }
853 
854 static int get_clocks(struct platform_device *pdev, struct msm_gpu *gpu)
855 {
856 	int ret = devm_clk_bulk_get_all(&pdev->dev, &gpu->grp_clks);
857 
858 	if (ret < 1) {
859 		gpu->nr_clocks = 0;
860 		return ret;
861 	}
862 
863 	gpu->nr_clocks = ret;
864 
865 	gpu->core_clk = msm_clk_bulk_get_clock(gpu->grp_clks,
866 		gpu->nr_clocks, "core");
867 
868 	gpu->rbbmtimer_clk = msm_clk_bulk_get_clock(gpu->grp_clks,
869 		gpu->nr_clocks, "rbbmtimer");
870 
871 	return 0;
872 }
873 
874 /* Return a new address space for a msm_drm_private instance */
875 struct msm_gem_address_space *
876 msm_gpu_create_private_address_space(struct msm_gpu *gpu, struct task_struct *task)
877 {
878 	struct msm_gem_address_space *aspace = NULL;
879 	if (!gpu)
880 		return NULL;
881 
882 	/*
883 	 * If the target doesn't support private address spaces then return
884 	 * the global one
885 	 */
886 	if (gpu->funcs->create_private_address_space) {
887 		aspace = gpu->funcs->create_private_address_space(gpu);
888 		if (!IS_ERR(aspace))
889 			aspace->pid = get_pid(task_pid(task));
890 	}
891 
892 	if (IS_ERR_OR_NULL(aspace))
893 		aspace = msm_gem_address_space_get(gpu->aspace);
894 
895 	return aspace;
896 }
897 
898 int msm_gpu_init(struct drm_device *drm, struct platform_device *pdev,
899 		struct msm_gpu *gpu, const struct msm_gpu_funcs *funcs,
900 		const char *name, struct msm_gpu_config *config)
901 {
902 	int i, ret, nr_rings = config->nr_rings;
903 	void *memptrs;
904 	uint64_t memptrs_iova;
905 
906 	if (WARN_ON(gpu->num_perfcntrs > ARRAY_SIZE(gpu->last_cntrs)))
907 		gpu->num_perfcntrs = ARRAY_SIZE(gpu->last_cntrs);
908 
909 	gpu->dev = drm;
910 	gpu->funcs = funcs;
911 	gpu->name = name;
912 
913 	gpu->worker = kthread_create_worker(0, "%s-worker", gpu->name);
914 	if (IS_ERR(gpu->worker)) {
915 		ret = PTR_ERR(gpu->worker);
916 		gpu->worker = NULL;
917 		goto fail;
918 	}
919 
920 	sched_set_fifo_low(gpu->worker->task);
921 
922 	INIT_LIST_HEAD(&gpu->active_list);
923 	kthread_init_work(&gpu->retire_work, retire_worker);
924 	kthread_init_work(&gpu->recover_work, recover_worker);
925 
926 	timer_setup(&gpu->hangcheck_timer, hangcheck_handler, 0);
927 
928 	spin_lock_init(&gpu->perf_lock);
929 
930 
931 	/* Map registers: */
932 	gpu->mmio = msm_ioremap(pdev, config->ioname, name);
933 	if (IS_ERR(gpu->mmio)) {
934 		ret = PTR_ERR(gpu->mmio);
935 		goto fail;
936 	}
937 
938 	/* Get Interrupt: */
939 	gpu->irq = platform_get_irq(pdev, 0);
940 	if (gpu->irq < 0) {
941 		ret = gpu->irq;
942 		DRM_DEV_ERROR(drm->dev, "failed to get irq: %d\n", ret);
943 		goto fail;
944 	}
945 
946 	ret = devm_request_irq(&pdev->dev, gpu->irq, irq_handler,
947 			IRQF_TRIGGER_HIGH, gpu->name, gpu);
948 	if (ret) {
949 		DRM_DEV_ERROR(drm->dev, "failed to request IRQ%u: %d\n", gpu->irq, ret);
950 		goto fail;
951 	}
952 
953 	ret = get_clocks(pdev, gpu);
954 	if (ret)
955 		goto fail;
956 
957 	gpu->ebi1_clk = msm_clk_get(pdev, "bus");
958 	DBG("ebi1_clk: %p", gpu->ebi1_clk);
959 	if (IS_ERR(gpu->ebi1_clk))
960 		gpu->ebi1_clk = NULL;
961 
962 	/* Acquire regulators: */
963 	gpu->gpu_reg = devm_regulator_get(&pdev->dev, "vdd");
964 	DBG("gpu_reg: %p", gpu->gpu_reg);
965 	if (IS_ERR(gpu->gpu_reg))
966 		gpu->gpu_reg = NULL;
967 
968 	gpu->gpu_cx = devm_regulator_get(&pdev->dev, "vddcx");
969 	DBG("gpu_cx: %p", gpu->gpu_cx);
970 	if (IS_ERR(gpu->gpu_cx))
971 		gpu->gpu_cx = NULL;
972 
973 	gpu->pdev = pdev;
974 	platform_set_drvdata(pdev, &gpu->adreno_smmu);
975 
976 	msm_devfreq_init(gpu);
977 
978 
979 	gpu->aspace = gpu->funcs->create_address_space(gpu, pdev);
980 
981 	if (gpu->aspace == NULL)
982 		DRM_DEV_INFO(drm->dev, "%s: no IOMMU, fallback to VRAM carveout!\n", name);
983 	else if (IS_ERR(gpu->aspace)) {
984 		ret = PTR_ERR(gpu->aspace);
985 		goto fail;
986 	}
987 
988 	memptrs = msm_gem_kernel_new(drm,
989 		sizeof(struct msm_rbmemptrs) * nr_rings,
990 		check_apriv(gpu, MSM_BO_UNCACHED), gpu->aspace, &gpu->memptrs_bo,
991 		&memptrs_iova);
992 
993 	if (IS_ERR(memptrs)) {
994 		ret = PTR_ERR(memptrs);
995 		DRM_DEV_ERROR(drm->dev, "could not allocate memptrs: %d\n", ret);
996 		goto fail;
997 	}
998 
999 	msm_gem_object_set_name(gpu->memptrs_bo, "memptrs");
1000 
1001 	if (nr_rings > ARRAY_SIZE(gpu->rb)) {
1002 		DRM_DEV_INFO_ONCE(drm->dev, "Only creating %zu ringbuffers\n",
1003 			ARRAY_SIZE(gpu->rb));
1004 		nr_rings = ARRAY_SIZE(gpu->rb);
1005 	}
1006 
1007 	/* Create ringbuffer(s): */
1008 	for (i = 0; i < nr_rings; i++) {
1009 		gpu->rb[i] = msm_ringbuffer_new(gpu, i, memptrs, memptrs_iova);
1010 
1011 		if (IS_ERR(gpu->rb[i])) {
1012 			ret = PTR_ERR(gpu->rb[i]);
1013 			DRM_DEV_ERROR(drm->dev,
1014 				"could not create ringbuffer %d: %d\n", i, ret);
1015 			goto fail;
1016 		}
1017 
1018 		memptrs += sizeof(struct msm_rbmemptrs);
1019 		memptrs_iova += sizeof(struct msm_rbmemptrs);
1020 	}
1021 
1022 	gpu->nr_rings = nr_rings;
1023 
1024 	return 0;
1025 
1026 fail:
1027 	for (i = 0; i < ARRAY_SIZE(gpu->rb); i++)  {
1028 		msm_ringbuffer_destroy(gpu->rb[i]);
1029 		gpu->rb[i] = NULL;
1030 	}
1031 
1032 	msm_gem_kernel_put(gpu->memptrs_bo, gpu->aspace, false);
1033 
1034 	platform_set_drvdata(pdev, NULL);
1035 	return ret;
1036 }
1037 
1038 void msm_gpu_cleanup(struct msm_gpu *gpu)
1039 {
1040 	int i;
1041 
1042 	DBG("%s", gpu->name);
1043 
1044 	WARN_ON(!list_empty(&gpu->active_list));
1045 
1046 	for (i = 0; i < ARRAY_SIZE(gpu->rb); i++) {
1047 		msm_ringbuffer_destroy(gpu->rb[i]);
1048 		gpu->rb[i] = NULL;
1049 	}
1050 
1051 	msm_gem_kernel_put(gpu->memptrs_bo, gpu->aspace, false);
1052 
1053 	if (!IS_ERR_OR_NULL(gpu->aspace)) {
1054 		gpu->aspace->mmu->funcs->detach(gpu->aspace->mmu);
1055 		msm_gem_address_space_put(gpu->aspace);
1056 	}
1057 
1058 	if (gpu->worker) {
1059 		kthread_destroy_worker(gpu->worker);
1060 	}
1061 
1062 	devfreq_cooling_unregister(gpu->cooling);
1063 }
1064