1 // SPDX-License-Identifier: GPL-2.0
2 /* Copyright 2019 Linaro, Ltd, Rob Herring <robh@kernel.org> */
3 /* Copyright 2019 Collabora ltd. */
4 #include <linux/delay.h>
5 #include <linux/interrupt.h>
6 #include <linux/io.h>
7 #include <linux/iopoll.h>
8 #include <linux/platform_device.h>
9 #include <linux/pm_runtime.h>
10 #include <linux/dma-resv.h>
11 #include <drm/gpu_scheduler.h>
12 #include <drm/panfrost_drm.h>
13 
14 #include "panfrost_device.h"
15 #include "panfrost_devfreq.h"
16 #include "panfrost_job.h"
17 #include "panfrost_features.h"
18 #include "panfrost_issues.h"
19 #include "panfrost_gem.h"
20 #include "panfrost_regs.h"
21 #include "panfrost_gpu.h"
22 #include "panfrost_mmu.h"
23 
24 #define JOB_TIMEOUT_MS 500
25 
26 #define job_write(dev, reg, data) writel(data, dev->iomem + (reg))
27 #define job_read(dev, reg) readl(dev->iomem + (reg))
28 
29 struct panfrost_queue_state {
30 	struct drm_gpu_scheduler sched;
31 	u64 fence_context;
32 	u64 emit_seqno;
33 };
34 
35 struct panfrost_job_slot {
36 	struct panfrost_queue_state queue[NUM_JOB_SLOTS];
37 	spinlock_t job_lock;
38 	int irq;
39 };
40 
41 static struct panfrost_job *
42 to_panfrost_job(struct drm_sched_job *sched_job)
43 {
44 	return container_of(sched_job, struct panfrost_job, base);
45 }
46 
47 struct panfrost_fence {
48 	struct dma_fence base;
49 	struct drm_device *dev;
50 	/* panfrost seqno for signaled() test */
51 	u64 seqno;
52 	int queue;
53 };
54 
55 static inline struct panfrost_fence *
56 to_panfrost_fence(struct dma_fence *fence)
57 {
58 	return (struct panfrost_fence *)fence;
59 }
60 
61 static const char *panfrost_fence_get_driver_name(struct dma_fence *fence)
62 {
63 	return "panfrost";
64 }
65 
66 static const char *panfrost_fence_get_timeline_name(struct dma_fence *fence)
67 {
68 	struct panfrost_fence *f = to_panfrost_fence(fence);
69 
70 	switch (f->queue) {
71 	case 0:
72 		return "panfrost-js-0";
73 	case 1:
74 		return "panfrost-js-1";
75 	case 2:
76 		return "panfrost-js-2";
77 	default:
78 		return NULL;
79 	}
80 }
81 
82 static const struct dma_fence_ops panfrost_fence_ops = {
83 	.get_driver_name = panfrost_fence_get_driver_name,
84 	.get_timeline_name = panfrost_fence_get_timeline_name,
85 };
86 
87 static struct dma_fence *panfrost_fence_create(struct panfrost_device *pfdev, int js_num)
88 {
89 	struct panfrost_fence *fence;
90 	struct panfrost_job_slot *js = pfdev->js;
91 
92 	fence = kzalloc(sizeof(*fence), GFP_KERNEL);
93 	if (!fence)
94 		return ERR_PTR(-ENOMEM);
95 
96 	fence->dev = pfdev->ddev;
97 	fence->queue = js_num;
98 	fence->seqno = ++js->queue[js_num].emit_seqno;
99 	dma_fence_init(&fence->base, &panfrost_fence_ops, &js->job_lock,
100 		       js->queue[js_num].fence_context, fence->seqno);
101 
102 	return &fence->base;
103 }
104 
105 int panfrost_job_get_slot(struct panfrost_job *job)
106 {
107 	/* JS0: fragment jobs.
108 	 * JS1: vertex/tiler jobs
109 	 * JS2: compute jobs
110 	 */
111 	if (job->requirements & PANFROST_JD_REQ_FS)
112 		return 0;
113 
114 /* Not exposed to userspace yet */
115 #if 0
116 	if (job->requirements & PANFROST_JD_REQ_ONLY_COMPUTE) {
117 		if ((job->requirements & PANFROST_JD_REQ_CORE_GRP_MASK) &&
118 		    (job->pfdev->features.nr_core_groups == 2))
119 			return 2;
120 		if (panfrost_has_hw_issue(job->pfdev, HW_ISSUE_8987))
121 			return 2;
122 	}
123 #endif
124 	return 1;
125 }
126 
127 static void panfrost_job_write_affinity(struct panfrost_device *pfdev,
128 					u32 requirements,
129 					int js)
130 {
131 	u64 affinity;
132 
133 	/*
134 	 * Use all cores for now.
135 	 * Eventually we may need to support tiler only jobs and h/w with
136 	 * multiple (2) coherent core groups
137 	 */
138 	affinity = pfdev->features.shader_present;
139 
140 	job_write(pfdev, JS_AFFINITY_NEXT_LO(js), lower_32_bits(affinity));
141 	job_write(pfdev, JS_AFFINITY_NEXT_HI(js), upper_32_bits(affinity));
142 }
143 
144 static u32
145 panfrost_get_job_chain_flag(const struct panfrost_job *job)
146 {
147 	struct panfrost_fence *f = to_panfrost_fence(job->done_fence);
148 
149 	if (!panfrost_has_hw_feature(job->pfdev, HW_FEATURE_JOBCHAIN_DISAMBIGUATION))
150 		return 0;
151 
152 	return (f->seqno & 1) ? JS_CONFIG_JOB_CHAIN_FLAG : 0;
153 }
154 
155 static struct panfrost_job *
156 panfrost_dequeue_job(struct panfrost_device *pfdev, int slot)
157 {
158 	struct panfrost_job *job = pfdev->jobs[slot][0];
159 
160 	WARN_ON(!job);
161 	pfdev->jobs[slot][0] = pfdev->jobs[slot][1];
162 	pfdev->jobs[slot][1] = NULL;
163 
164 	return job;
165 }
166 
167 static unsigned int
168 panfrost_enqueue_job(struct panfrost_device *pfdev, int slot,
169 		     struct panfrost_job *job)
170 {
171 	if (WARN_ON(!job))
172 		return 0;
173 
174 	if (!pfdev->jobs[slot][0]) {
175 		pfdev->jobs[slot][0] = job;
176 		return 0;
177 	}
178 
179 	WARN_ON(pfdev->jobs[slot][1]);
180 	pfdev->jobs[slot][1] = job;
181 	WARN_ON(panfrost_get_job_chain_flag(job) ==
182 		panfrost_get_job_chain_flag(pfdev->jobs[slot][0]));
183 	return 1;
184 }
185 
186 static void panfrost_job_hw_submit(struct panfrost_job *job, int js)
187 {
188 	struct panfrost_device *pfdev = job->pfdev;
189 	unsigned int subslot;
190 	u32 cfg;
191 	u64 jc_head = job->jc;
192 	int ret;
193 
194 	panfrost_devfreq_record_busy(&pfdev->pfdevfreq);
195 
196 	ret = pm_runtime_get_sync(pfdev->dev);
197 	if (ret < 0)
198 		return;
199 
200 	if (WARN_ON(job_read(pfdev, JS_COMMAND_NEXT(js)))) {
201 		return;
202 	}
203 
204 	cfg = panfrost_mmu_as_get(pfdev, job->mmu);
205 
206 	job_write(pfdev, JS_HEAD_NEXT_LO(js), lower_32_bits(jc_head));
207 	job_write(pfdev, JS_HEAD_NEXT_HI(js), upper_32_bits(jc_head));
208 
209 	panfrost_job_write_affinity(pfdev, job->requirements, js);
210 
211 	/* start MMU, medium priority, cache clean/flush on end, clean/flush on
212 	 * start */
213 	cfg |= JS_CONFIG_THREAD_PRI(8) |
214 		JS_CONFIG_START_FLUSH_CLEAN_INVALIDATE |
215 		JS_CONFIG_END_FLUSH_CLEAN_INVALIDATE |
216 		panfrost_get_job_chain_flag(job);
217 
218 	if (panfrost_has_hw_feature(pfdev, HW_FEATURE_FLUSH_REDUCTION))
219 		cfg |= JS_CONFIG_ENABLE_FLUSH_REDUCTION;
220 
221 	if (panfrost_has_hw_issue(pfdev, HW_ISSUE_10649))
222 		cfg |= JS_CONFIG_START_MMU;
223 
224 	job_write(pfdev, JS_CONFIG_NEXT(js), cfg);
225 
226 	if (panfrost_has_hw_feature(pfdev, HW_FEATURE_FLUSH_REDUCTION))
227 		job_write(pfdev, JS_FLUSH_ID_NEXT(js), job->flush_id);
228 
229 	/* GO ! */
230 
231 	spin_lock(&pfdev->js->job_lock);
232 	subslot = panfrost_enqueue_job(pfdev, js, job);
233 	/* Don't queue the job if a reset is in progress */
234 	if (!atomic_read(&pfdev->reset.pending)) {
235 		job_write(pfdev, JS_COMMAND_NEXT(js), JS_COMMAND_START);
236 		dev_dbg(pfdev->dev,
237 			"JS: Submitting atom %p to js[%d][%d] with head=0x%llx AS %d",
238 			job, js, subslot, jc_head, cfg & 0xf);
239 	}
240 	spin_unlock(&pfdev->js->job_lock);
241 }
242 
243 static int panfrost_acquire_object_fences(struct drm_gem_object **bos,
244 					  int bo_count,
245 					  struct drm_sched_job *job)
246 {
247 	int i, ret;
248 
249 	for (i = 0; i < bo_count; i++) {
250 		ret = dma_resv_reserve_fences(bos[i]->resv, 1);
251 		if (ret)
252 			return ret;
253 
254 		/* panfrost always uses write mode in its current uapi */
255 		ret = drm_sched_job_add_implicit_dependencies(job, bos[i],
256 							      true);
257 		if (ret)
258 			return ret;
259 	}
260 
261 	return 0;
262 }
263 
264 static void panfrost_attach_object_fences(struct drm_gem_object **bos,
265 					  int bo_count,
266 					  struct dma_fence *fence)
267 {
268 	int i;
269 
270 	for (i = 0; i < bo_count; i++)
271 		dma_resv_add_fence(bos[i]->resv, fence, DMA_RESV_USAGE_WRITE);
272 }
273 
274 int panfrost_job_push(struct panfrost_job *job)
275 {
276 	struct panfrost_device *pfdev = job->pfdev;
277 	struct ww_acquire_ctx acquire_ctx;
278 	int ret = 0;
279 
280 	ret = drm_gem_lock_reservations(job->bos, job->bo_count,
281 					    &acquire_ctx);
282 	if (ret)
283 		return ret;
284 
285 	mutex_lock(&pfdev->sched_lock);
286 	drm_sched_job_arm(&job->base);
287 
288 	job->render_done_fence = dma_fence_get(&job->base.s_fence->finished);
289 
290 	ret = panfrost_acquire_object_fences(job->bos, job->bo_count,
291 					     &job->base);
292 	if (ret) {
293 		mutex_unlock(&pfdev->sched_lock);
294 		goto unlock;
295 	}
296 
297 	kref_get(&job->refcount); /* put by scheduler job completion */
298 
299 	drm_sched_entity_push_job(&job->base);
300 
301 	mutex_unlock(&pfdev->sched_lock);
302 
303 	panfrost_attach_object_fences(job->bos, job->bo_count,
304 				      job->render_done_fence);
305 
306 unlock:
307 	drm_gem_unlock_reservations(job->bos, job->bo_count, &acquire_ctx);
308 
309 	return ret;
310 }
311 
312 static void panfrost_job_cleanup(struct kref *ref)
313 {
314 	struct panfrost_job *job = container_of(ref, struct panfrost_job,
315 						refcount);
316 	unsigned int i;
317 
318 	dma_fence_put(job->done_fence);
319 	dma_fence_put(job->render_done_fence);
320 
321 	if (job->mappings) {
322 		for (i = 0; i < job->bo_count; i++) {
323 			if (!job->mappings[i])
324 				break;
325 
326 			atomic_dec(&job->mappings[i]->obj->gpu_usecount);
327 			panfrost_gem_mapping_put(job->mappings[i]);
328 		}
329 		kvfree(job->mappings);
330 	}
331 
332 	if (job->bos) {
333 		for (i = 0; i < job->bo_count; i++)
334 			drm_gem_object_put(job->bos[i]);
335 
336 		kvfree(job->bos);
337 	}
338 
339 	kfree(job);
340 }
341 
342 void panfrost_job_put(struct panfrost_job *job)
343 {
344 	kref_put(&job->refcount, panfrost_job_cleanup);
345 }
346 
347 static void panfrost_job_free(struct drm_sched_job *sched_job)
348 {
349 	struct panfrost_job *job = to_panfrost_job(sched_job);
350 
351 	drm_sched_job_cleanup(sched_job);
352 
353 	panfrost_job_put(job);
354 }
355 
356 static struct dma_fence *panfrost_job_run(struct drm_sched_job *sched_job)
357 {
358 	struct panfrost_job *job = to_panfrost_job(sched_job);
359 	struct panfrost_device *pfdev = job->pfdev;
360 	int slot = panfrost_job_get_slot(job);
361 	struct dma_fence *fence = NULL;
362 
363 	if (unlikely(job->base.s_fence->finished.error))
364 		return NULL;
365 
366 	/* Nothing to execute: can happen if the job has finished while
367 	 * we were resetting the GPU.
368 	 */
369 	if (!job->jc)
370 		return NULL;
371 
372 	fence = panfrost_fence_create(pfdev, slot);
373 	if (IS_ERR(fence))
374 		return fence;
375 
376 	if (job->done_fence)
377 		dma_fence_put(job->done_fence);
378 	job->done_fence = dma_fence_get(fence);
379 
380 	panfrost_job_hw_submit(job, slot);
381 
382 	return fence;
383 }
384 
385 void panfrost_job_enable_interrupts(struct panfrost_device *pfdev)
386 {
387 	int j;
388 	u32 irq_mask = 0;
389 
390 	for (j = 0; j < NUM_JOB_SLOTS; j++) {
391 		irq_mask |= MK_JS_MASK(j);
392 	}
393 
394 	job_write(pfdev, JOB_INT_CLEAR, irq_mask);
395 	job_write(pfdev, JOB_INT_MASK, irq_mask);
396 }
397 
398 static void panfrost_job_handle_err(struct panfrost_device *pfdev,
399 				    struct panfrost_job *job,
400 				    unsigned int js)
401 {
402 	u32 js_status = job_read(pfdev, JS_STATUS(js));
403 	const char *exception_name = panfrost_exception_name(js_status);
404 	bool signal_fence = true;
405 
406 	if (!panfrost_exception_is_fault(js_status)) {
407 		dev_dbg(pfdev->dev, "js event, js=%d, status=%s, head=0x%x, tail=0x%x",
408 			js, exception_name,
409 			job_read(pfdev, JS_HEAD_LO(js)),
410 			job_read(pfdev, JS_TAIL_LO(js)));
411 	} else {
412 		dev_err(pfdev->dev, "js fault, js=%d, status=%s, head=0x%x, tail=0x%x",
413 			js, exception_name,
414 			job_read(pfdev, JS_HEAD_LO(js)),
415 			job_read(pfdev, JS_TAIL_LO(js)));
416 	}
417 
418 	if (js_status == DRM_PANFROST_EXCEPTION_STOPPED) {
419 		/* Update the job head so we can resume */
420 		job->jc = job_read(pfdev, JS_TAIL_LO(js)) |
421 			  ((u64)job_read(pfdev, JS_TAIL_HI(js)) << 32);
422 
423 		/* The job will be resumed, don't signal the fence */
424 		signal_fence = false;
425 	} else if (js_status == DRM_PANFROST_EXCEPTION_TERMINATED) {
426 		/* Job has been hard-stopped, flag it as canceled */
427 		dma_fence_set_error(job->done_fence, -ECANCELED);
428 		job->jc = 0;
429 	} else if (panfrost_exception_is_fault(js_status)) {
430 		/* We might want to provide finer-grained error code based on
431 		 * the exception type, but unconditionally setting to EINVAL
432 		 * is good enough for now.
433 		 */
434 		dma_fence_set_error(job->done_fence, -EINVAL);
435 		job->jc = 0;
436 	}
437 
438 	panfrost_mmu_as_put(pfdev, job->mmu);
439 	panfrost_devfreq_record_idle(&pfdev->pfdevfreq);
440 
441 	if (signal_fence)
442 		dma_fence_signal_locked(job->done_fence);
443 
444 	pm_runtime_put_autosuspend(pfdev->dev);
445 
446 	if (panfrost_exception_needs_reset(pfdev, js_status)) {
447 		atomic_set(&pfdev->reset.pending, 1);
448 		drm_sched_fault(&pfdev->js->queue[js].sched);
449 	}
450 }
451 
452 static void panfrost_job_handle_done(struct panfrost_device *pfdev,
453 				     struct panfrost_job *job)
454 {
455 	/* Set ->jc to 0 to avoid re-submitting an already finished job (can
456 	 * happen when we receive the DONE interrupt while doing a GPU reset).
457 	 */
458 	job->jc = 0;
459 	panfrost_mmu_as_put(pfdev, job->mmu);
460 	panfrost_devfreq_record_idle(&pfdev->pfdevfreq);
461 
462 	dma_fence_signal_locked(job->done_fence);
463 	pm_runtime_put_autosuspend(pfdev->dev);
464 }
465 
466 static void panfrost_job_handle_irq(struct panfrost_device *pfdev, u32 status)
467 {
468 	struct panfrost_job *done[NUM_JOB_SLOTS][2] = {};
469 	struct panfrost_job *failed[NUM_JOB_SLOTS] = {};
470 	u32 js_state = 0, js_events = 0;
471 	unsigned int i, j;
472 
473 	/* First we collect all failed/done jobs. */
474 	while (status) {
475 		u32 js_state_mask = 0;
476 
477 		for (j = 0; j < NUM_JOB_SLOTS; j++) {
478 			if (status & MK_JS_MASK(j))
479 				js_state_mask |= MK_JS_MASK(j);
480 
481 			if (status & JOB_INT_MASK_DONE(j)) {
482 				if (done[j][0])
483 					done[j][1] = panfrost_dequeue_job(pfdev, j);
484 				else
485 					done[j][0] = panfrost_dequeue_job(pfdev, j);
486 			}
487 
488 			if (status & JOB_INT_MASK_ERR(j)) {
489 				/* Cancel the next submission. Will be submitted
490 				 * after we're done handling this failure if
491 				 * there's no reset pending.
492 				 */
493 				job_write(pfdev, JS_COMMAND_NEXT(j), JS_COMMAND_NOP);
494 				failed[j] = panfrost_dequeue_job(pfdev, j);
495 			}
496 		}
497 
498 		/* JS_STATE is sampled when JOB_INT_CLEAR is written.
499 		 * For each BIT(slot) or BIT(slot + 16) bit written to
500 		 * JOB_INT_CLEAR, the corresponding bits in JS_STATE
501 		 * (BIT(slot) and BIT(slot + 16)) are updated, but this
502 		 * is racy. If we only have one job done at the time we
503 		 * read JOB_INT_RAWSTAT but the second job fails before we
504 		 * clear the status, we end up with a status containing
505 		 * only the DONE bit and consider both jobs as DONE since
506 		 * JS_STATE reports both NEXT and CURRENT as inactive.
507 		 * To prevent that, let's repeat this clear+read steps
508 		 * until status is 0.
509 		 */
510 		job_write(pfdev, JOB_INT_CLEAR, status);
511 		js_state &= ~js_state_mask;
512 		js_state |= job_read(pfdev, JOB_INT_JS_STATE) & js_state_mask;
513 		js_events |= status;
514 		status = job_read(pfdev, JOB_INT_RAWSTAT);
515 	}
516 
517 	/* Then we handle the dequeued jobs. */
518 	for (j = 0; j < NUM_JOB_SLOTS; j++) {
519 		if (!(js_events & MK_JS_MASK(j)))
520 			continue;
521 
522 		if (failed[j]) {
523 			panfrost_job_handle_err(pfdev, failed[j], j);
524 		} else if (pfdev->jobs[j][0] && !(js_state & MK_JS_MASK(j))) {
525 			/* When the current job doesn't fail, the JM dequeues
526 			 * the next job without waiting for an ACK, this means
527 			 * we can have 2 jobs dequeued and only catch the
528 			 * interrupt when the second one is done. If both slots
529 			 * are inactive, but one job remains in pfdev->jobs[j],
530 			 * consider it done. Of course that doesn't apply if a
531 			 * failure happened since we cancelled execution of the
532 			 * job in _NEXT (see above).
533 			 */
534 			if (WARN_ON(!done[j][0]))
535 				done[j][0] = panfrost_dequeue_job(pfdev, j);
536 			else
537 				done[j][1] = panfrost_dequeue_job(pfdev, j);
538 		}
539 
540 		for (i = 0; i < ARRAY_SIZE(done[0]) && done[j][i]; i++)
541 			panfrost_job_handle_done(pfdev, done[j][i]);
542 	}
543 
544 	/* And finally we requeue jobs that were waiting in the second slot
545 	 * and have been stopped if we detected a failure on the first slot.
546 	 */
547 	for (j = 0; j < NUM_JOB_SLOTS; j++) {
548 		if (!(js_events & MK_JS_MASK(j)))
549 			continue;
550 
551 		if (!failed[j] || !pfdev->jobs[j][0])
552 			continue;
553 
554 		if (pfdev->jobs[j][0]->jc == 0) {
555 			/* The job was cancelled, signal the fence now */
556 			struct panfrost_job *canceled = panfrost_dequeue_job(pfdev, j);
557 
558 			dma_fence_set_error(canceled->done_fence, -ECANCELED);
559 			panfrost_job_handle_done(pfdev, canceled);
560 		} else if (!atomic_read(&pfdev->reset.pending)) {
561 			/* Requeue the job we removed if no reset is pending */
562 			job_write(pfdev, JS_COMMAND_NEXT(j), JS_COMMAND_START);
563 		}
564 	}
565 }
566 
567 static void panfrost_job_handle_irqs(struct panfrost_device *pfdev)
568 {
569 	u32 status = job_read(pfdev, JOB_INT_RAWSTAT);
570 
571 	while (status) {
572 		pm_runtime_mark_last_busy(pfdev->dev);
573 
574 		spin_lock(&pfdev->js->job_lock);
575 		panfrost_job_handle_irq(pfdev, status);
576 		spin_unlock(&pfdev->js->job_lock);
577 		status = job_read(pfdev, JOB_INT_RAWSTAT);
578 	}
579 }
580 
581 static u32 panfrost_active_slots(struct panfrost_device *pfdev,
582 				 u32 *js_state_mask, u32 js_state)
583 {
584 	u32 rawstat;
585 
586 	if (!(js_state & *js_state_mask))
587 		return 0;
588 
589 	rawstat = job_read(pfdev, JOB_INT_RAWSTAT);
590 	if (rawstat) {
591 		unsigned int i;
592 
593 		for (i = 0; i < NUM_JOB_SLOTS; i++) {
594 			if (rawstat & MK_JS_MASK(i))
595 				*js_state_mask &= ~MK_JS_MASK(i);
596 		}
597 	}
598 
599 	return js_state & *js_state_mask;
600 }
601 
602 static void
603 panfrost_reset(struct panfrost_device *pfdev,
604 	       struct drm_sched_job *bad)
605 {
606 	u32 js_state, js_state_mask = 0xffffffff;
607 	unsigned int i, j;
608 	bool cookie;
609 	int ret;
610 
611 	if (!atomic_read(&pfdev->reset.pending))
612 		return;
613 
614 	/* Stop the schedulers.
615 	 *
616 	 * FIXME: We temporarily get out of the dma_fence_signalling section
617 	 * because the cleanup path generate lockdep splats when taking locks
618 	 * to release job resources. We should rework the code to follow this
619 	 * pattern:
620 	 *
621 	 *	try_lock
622 	 *	if (locked)
623 	 *		release
624 	 *	else
625 	 *		schedule_work_to_release_later
626 	 */
627 	for (i = 0; i < NUM_JOB_SLOTS; i++)
628 		drm_sched_stop(&pfdev->js->queue[i].sched, bad);
629 
630 	cookie = dma_fence_begin_signalling();
631 
632 	if (bad)
633 		drm_sched_increase_karma(bad);
634 
635 	/* Mask job interrupts and synchronize to make sure we won't be
636 	 * interrupted during our reset.
637 	 */
638 	job_write(pfdev, JOB_INT_MASK, 0);
639 	synchronize_irq(pfdev->js->irq);
640 
641 	for (i = 0; i < NUM_JOB_SLOTS; i++) {
642 		/* Cancel the next job and soft-stop the running job. */
643 		job_write(pfdev, JS_COMMAND_NEXT(i), JS_COMMAND_NOP);
644 		job_write(pfdev, JS_COMMAND(i), JS_COMMAND_SOFT_STOP);
645 	}
646 
647 	/* Wait at most 10ms for soft-stops to complete */
648 	ret = readl_poll_timeout(pfdev->iomem + JOB_INT_JS_STATE, js_state,
649 				 !panfrost_active_slots(pfdev, &js_state_mask, js_state),
650 				 10, 10000);
651 
652 	if (ret)
653 		dev_err(pfdev->dev, "Soft-stop failed\n");
654 
655 	/* Handle the remaining interrupts before we reset. */
656 	panfrost_job_handle_irqs(pfdev);
657 
658 	/* Remaining interrupts have been handled, but we might still have
659 	 * stuck jobs. Let's make sure the PM counters stay balanced by
660 	 * manually calling pm_runtime_put_noidle() and
661 	 * panfrost_devfreq_record_idle() for each stuck job.
662 	 */
663 	spin_lock(&pfdev->js->job_lock);
664 	for (i = 0; i < NUM_JOB_SLOTS; i++) {
665 		for (j = 0; j < ARRAY_SIZE(pfdev->jobs[0]) && pfdev->jobs[i][j]; j++) {
666 			pm_runtime_put_noidle(pfdev->dev);
667 			panfrost_devfreq_record_idle(&pfdev->pfdevfreq);
668 		}
669 	}
670 	memset(pfdev->jobs, 0, sizeof(pfdev->jobs));
671 	spin_unlock(&pfdev->js->job_lock);
672 
673 	/* Proceed with reset now. */
674 	panfrost_device_reset(pfdev);
675 
676 	/* panfrost_device_reset() unmasks job interrupts, but we want to
677 	 * keep them masked a bit longer.
678 	 */
679 	job_write(pfdev, JOB_INT_MASK, 0);
680 
681 	/* GPU has been reset, we can clear the reset pending bit. */
682 	atomic_set(&pfdev->reset.pending, 0);
683 
684 	/* Now resubmit jobs that were previously queued but didn't have a
685 	 * chance to finish.
686 	 * FIXME: We temporarily get out of the DMA fence signalling section
687 	 * while resubmitting jobs because the job submission logic will
688 	 * allocate memory with the GFP_KERNEL flag which can trigger memory
689 	 * reclaim and exposes a lock ordering issue.
690 	 */
691 	dma_fence_end_signalling(cookie);
692 	for (i = 0; i < NUM_JOB_SLOTS; i++)
693 		drm_sched_resubmit_jobs(&pfdev->js->queue[i].sched);
694 	cookie = dma_fence_begin_signalling();
695 
696 	/* Restart the schedulers */
697 	for (i = 0; i < NUM_JOB_SLOTS; i++)
698 		drm_sched_start(&pfdev->js->queue[i].sched, true);
699 
700 	/* Re-enable job interrupts now that everything has been restarted. */
701 	job_write(pfdev, JOB_INT_MASK,
702 		  GENMASK(16 + NUM_JOB_SLOTS - 1, 16) |
703 		  GENMASK(NUM_JOB_SLOTS - 1, 0));
704 
705 	dma_fence_end_signalling(cookie);
706 }
707 
708 static enum drm_gpu_sched_stat panfrost_job_timedout(struct drm_sched_job
709 						     *sched_job)
710 {
711 	struct panfrost_job *job = to_panfrost_job(sched_job);
712 	struct panfrost_device *pfdev = job->pfdev;
713 	int js = panfrost_job_get_slot(job);
714 
715 	/*
716 	 * If the GPU managed to complete this jobs fence, the timeout is
717 	 * spurious. Bail out.
718 	 */
719 	if (dma_fence_is_signaled(job->done_fence))
720 		return DRM_GPU_SCHED_STAT_NOMINAL;
721 
722 	dev_err(pfdev->dev, "gpu sched timeout, js=%d, config=0x%x, status=0x%x, head=0x%x, tail=0x%x, sched_job=%p",
723 		js,
724 		job_read(pfdev, JS_CONFIG(js)),
725 		job_read(pfdev, JS_STATUS(js)),
726 		job_read(pfdev, JS_HEAD_LO(js)),
727 		job_read(pfdev, JS_TAIL_LO(js)),
728 		sched_job);
729 
730 	atomic_set(&pfdev->reset.pending, 1);
731 	panfrost_reset(pfdev, sched_job);
732 
733 	return DRM_GPU_SCHED_STAT_NOMINAL;
734 }
735 
736 static void panfrost_reset_work(struct work_struct *work)
737 {
738 	struct panfrost_device *pfdev;
739 
740 	pfdev = container_of(work, struct panfrost_device, reset.work);
741 	panfrost_reset(pfdev, NULL);
742 }
743 
744 static const struct drm_sched_backend_ops panfrost_sched_ops = {
745 	.run_job = panfrost_job_run,
746 	.timedout_job = panfrost_job_timedout,
747 	.free_job = panfrost_job_free
748 };
749 
750 static irqreturn_t panfrost_job_irq_handler_thread(int irq, void *data)
751 {
752 	struct panfrost_device *pfdev = data;
753 
754 	panfrost_job_handle_irqs(pfdev);
755 	job_write(pfdev, JOB_INT_MASK,
756 		  GENMASK(16 + NUM_JOB_SLOTS - 1, 16) |
757 		  GENMASK(NUM_JOB_SLOTS - 1, 0));
758 	return IRQ_HANDLED;
759 }
760 
761 static irqreturn_t panfrost_job_irq_handler(int irq, void *data)
762 {
763 	struct panfrost_device *pfdev = data;
764 	u32 status = job_read(pfdev, JOB_INT_STAT);
765 
766 	if (!status)
767 		return IRQ_NONE;
768 
769 	job_write(pfdev, JOB_INT_MASK, 0);
770 	return IRQ_WAKE_THREAD;
771 }
772 
773 int panfrost_job_init(struct panfrost_device *pfdev)
774 {
775 	struct panfrost_job_slot *js;
776 	unsigned int nentries = 2;
777 	int ret, j;
778 
779 	/* All GPUs have two entries per queue, but without jobchain
780 	 * disambiguation stopping the right job in the close path is tricky,
781 	 * so let's just advertise one entry in that case.
782 	 */
783 	if (!panfrost_has_hw_feature(pfdev, HW_FEATURE_JOBCHAIN_DISAMBIGUATION))
784 		nentries = 1;
785 
786 	pfdev->js = js = devm_kzalloc(pfdev->dev, sizeof(*js), GFP_KERNEL);
787 	if (!js)
788 		return -ENOMEM;
789 
790 	INIT_WORK(&pfdev->reset.work, panfrost_reset_work);
791 	spin_lock_init(&js->job_lock);
792 
793 	js->irq = platform_get_irq_byname(to_platform_device(pfdev->dev), "job");
794 	if (js->irq <= 0)
795 		return -ENODEV;
796 
797 	ret = devm_request_threaded_irq(pfdev->dev, js->irq,
798 					panfrost_job_irq_handler,
799 					panfrost_job_irq_handler_thread,
800 					IRQF_SHARED, KBUILD_MODNAME "-job",
801 					pfdev);
802 	if (ret) {
803 		dev_err(pfdev->dev, "failed to request job irq");
804 		return ret;
805 	}
806 
807 	pfdev->reset.wq = alloc_ordered_workqueue("panfrost-reset", 0);
808 	if (!pfdev->reset.wq)
809 		return -ENOMEM;
810 
811 	for (j = 0; j < NUM_JOB_SLOTS; j++) {
812 		js->queue[j].fence_context = dma_fence_context_alloc(1);
813 
814 		ret = drm_sched_init(&js->queue[j].sched,
815 				     &panfrost_sched_ops,
816 				     nentries, 0,
817 				     msecs_to_jiffies(JOB_TIMEOUT_MS),
818 				     pfdev->reset.wq,
819 				     NULL, "pan_js", pfdev->dev);
820 		if (ret) {
821 			dev_err(pfdev->dev, "Failed to create scheduler: %d.", ret);
822 			goto err_sched;
823 		}
824 	}
825 
826 	panfrost_job_enable_interrupts(pfdev);
827 
828 	return 0;
829 
830 err_sched:
831 	for (j--; j >= 0; j--)
832 		drm_sched_fini(&js->queue[j].sched);
833 
834 	destroy_workqueue(pfdev->reset.wq);
835 	return ret;
836 }
837 
838 void panfrost_job_fini(struct panfrost_device *pfdev)
839 {
840 	struct panfrost_job_slot *js = pfdev->js;
841 	int j;
842 
843 	job_write(pfdev, JOB_INT_MASK, 0);
844 
845 	for (j = 0; j < NUM_JOB_SLOTS; j++) {
846 		drm_sched_fini(&js->queue[j].sched);
847 	}
848 
849 	cancel_work_sync(&pfdev->reset.work);
850 	destroy_workqueue(pfdev->reset.wq);
851 }
852 
853 int panfrost_job_open(struct panfrost_file_priv *panfrost_priv)
854 {
855 	struct panfrost_device *pfdev = panfrost_priv->pfdev;
856 	struct panfrost_job_slot *js = pfdev->js;
857 	struct drm_gpu_scheduler *sched;
858 	int ret, i;
859 
860 	for (i = 0; i < NUM_JOB_SLOTS; i++) {
861 		sched = &js->queue[i].sched;
862 		ret = drm_sched_entity_init(&panfrost_priv->sched_entity[i],
863 					    DRM_SCHED_PRIORITY_NORMAL, &sched,
864 					    1, NULL);
865 		if (WARN_ON(ret))
866 			return ret;
867 	}
868 	return 0;
869 }
870 
871 void panfrost_job_close(struct panfrost_file_priv *panfrost_priv)
872 {
873 	struct panfrost_device *pfdev = panfrost_priv->pfdev;
874 	int i;
875 
876 	for (i = 0; i < NUM_JOB_SLOTS; i++)
877 		drm_sched_entity_destroy(&panfrost_priv->sched_entity[i]);
878 
879 	/* Kill in-flight jobs */
880 	spin_lock(&pfdev->js->job_lock);
881 	for (i = 0; i < NUM_JOB_SLOTS; i++) {
882 		struct drm_sched_entity *entity = &panfrost_priv->sched_entity[i];
883 		int j;
884 
885 		for (j = ARRAY_SIZE(pfdev->jobs[0]) - 1; j >= 0; j--) {
886 			struct panfrost_job *job = pfdev->jobs[i][j];
887 			u32 cmd;
888 
889 			if (!job || job->base.entity != entity)
890 				continue;
891 
892 			if (j == 1) {
893 				/* Try to cancel the job before it starts */
894 				job_write(pfdev, JS_COMMAND_NEXT(i), JS_COMMAND_NOP);
895 				/* Reset the job head so it doesn't get restarted if
896 				 * the job in the first slot failed.
897 				 */
898 				job->jc = 0;
899 			}
900 
901 			if (panfrost_has_hw_feature(pfdev, HW_FEATURE_JOBCHAIN_DISAMBIGUATION)) {
902 				cmd = panfrost_get_job_chain_flag(job) ?
903 				      JS_COMMAND_HARD_STOP_1 :
904 				      JS_COMMAND_HARD_STOP_0;
905 			} else {
906 				cmd = JS_COMMAND_HARD_STOP;
907 			}
908 
909 			job_write(pfdev, JS_COMMAND(i), cmd);
910 		}
911 	}
912 	spin_unlock(&pfdev->js->job_lock);
913 }
914 
915 int panfrost_job_is_idle(struct panfrost_device *pfdev)
916 {
917 	struct panfrost_job_slot *js = pfdev->js;
918 	int i;
919 
920 	for (i = 0; i < NUM_JOB_SLOTS; i++) {
921 		/* If there are any jobs in the HW queue, we're not idle */
922 		if (atomic_read(&js->queue[i].sched.hw_rq_count))
923 			return false;
924 	}
925 
926 	return true;
927 }
928