xref: /openbmc/linux/drivers/gpu/drm/lima/lima_sched.c (revision fb574682)
1 // SPDX-License-Identifier: GPL-2.0 OR MIT
2 /* Copyright 2017-2019 Qiang Yu <yuq825@gmail.com> */
3 
4 #include <linux/kthread.h>
5 #include <linux/slab.h>
6 #include <linux/vmalloc.h>
7 #include <linux/pm_runtime.h>
8 
9 #include "lima_devfreq.h"
10 #include "lima_drv.h"
11 #include "lima_sched.h"
12 #include "lima_vm.h"
13 #include "lima_mmu.h"
14 #include "lima_l2_cache.h"
15 #include "lima_gem.h"
16 #include "lima_trace.h"
17 
18 struct lima_fence {
19 	struct dma_fence base;
20 	struct lima_sched_pipe *pipe;
21 };
22 
23 static struct kmem_cache *lima_fence_slab;
24 static int lima_fence_slab_refcnt;
25 
26 int lima_sched_slab_init(void)
27 {
28 	if (!lima_fence_slab) {
29 		lima_fence_slab = kmem_cache_create(
30 			"lima_fence", sizeof(struct lima_fence), 0,
31 			SLAB_HWCACHE_ALIGN, NULL);
32 		if (!lima_fence_slab)
33 			return -ENOMEM;
34 	}
35 
36 	lima_fence_slab_refcnt++;
37 	return 0;
38 }
39 
40 void lima_sched_slab_fini(void)
41 {
42 	if (!--lima_fence_slab_refcnt) {
43 		kmem_cache_destroy(lima_fence_slab);
44 		lima_fence_slab = NULL;
45 	}
46 }
47 
48 static inline struct lima_fence *to_lima_fence(struct dma_fence *fence)
49 {
50 	return container_of(fence, struct lima_fence, base);
51 }
52 
53 static const char *lima_fence_get_driver_name(struct dma_fence *fence)
54 {
55 	return "lima";
56 }
57 
58 static const char *lima_fence_get_timeline_name(struct dma_fence *fence)
59 {
60 	struct lima_fence *f = to_lima_fence(fence);
61 
62 	return f->pipe->base.name;
63 }
64 
65 static void lima_fence_release_rcu(struct rcu_head *rcu)
66 {
67 	struct dma_fence *f = container_of(rcu, struct dma_fence, rcu);
68 	struct lima_fence *fence = to_lima_fence(f);
69 
70 	kmem_cache_free(lima_fence_slab, fence);
71 }
72 
73 static void lima_fence_release(struct dma_fence *fence)
74 {
75 	struct lima_fence *f = to_lima_fence(fence);
76 
77 	call_rcu(&f->base.rcu, lima_fence_release_rcu);
78 }
79 
80 static const struct dma_fence_ops lima_fence_ops = {
81 	.get_driver_name = lima_fence_get_driver_name,
82 	.get_timeline_name = lima_fence_get_timeline_name,
83 	.release = lima_fence_release,
84 };
85 
86 static struct lima_fence *lima_fence_create(struct lima_sched_pipe *pipe)
87 {
88 	struct lima_fence *fence;
89 
90 	fence = kmem_cache_zalloc(lima_fence_slab, GFP_KERNEL);
91 	if (!fence)
92 		return NULL;
93 
94 	fence->pipe = pipe;
95 	dma_fence_init(&fence->base, &lima_fence_ops, &pipe->fence_lock,
96 		       pipe->fence_context, ++pipe->fence_seqno);
97 
98 	return fence;
99 }
100 
101 static inline struct lima_sched_task *to_lima_task(struct drm_sched_job *job)
102 {
103 	return container_of(job, struct lima_sched_task, base);
104 }
105 
106 static inline struct lima_sched_pipe *to_lima_pipe(struct drm_gpu_scheduler *sched)
107 {
108 	return container_of(sched, struct lima_sched_pipe, base);
109 }
110 
111 int lima_sched_task_init(struct lima_sched_task *task,
112 			 struct lima_sched_context *context,
113 			 struct lima_bo **bos, int num_bos,
114 			 struct lima_vm *vm)
115 {
116 	int err, i;
117 
118 	task->bos = kmemdup(bos, sizeof(*bos) * num_bos, GFP_KERNEL);
119 	if (!task->bos)
120 		return -ENOMEM;
121 
122 	for (i = 0; i < num_bos; i++)
123 		drm_gem_object_get(&bos[i]->base.base);
124 
125 	err = drm_sched_job_init(&task->base, &context->base, vm);
126 	if (err) {
127 		kfree(task->bos);
128 		return err;
129 	}
130 
131 	task->num_bos = num_bos;
132 	task->vm = lima_vm_get(vm);
133 
134 	xa_init_flags(&task->deps, XA_FLAGS_ALLOC);
135 
136 	return 0;
137 }
138 
139 void lima_sched_task_fini(struct lima_sched_task *task)
140 {
141 	struct dma_fence *fence;
142 	unsigned long index;
143 	int i;
144 
145 	drm_sched_job_cleanup(&task->base);
146 
147 	xa_for_each(&task->deps, index, fence) {
148 		dma_fence_put(fence);
149 	}
150 	xa_destroy(&task->deps);
151 
152 	if (task->bos) {
153 		for (i = 0; i < task->num_bos; i++)
154 			drm_gem_object_put_unlocked(&task->bos[i]->base.base);
155 		kfree(task->bos);
156 	}
157 
158 	lima_vm_put(task->vm);
159 }
160 
161 int lima_sched_context_init(struct lima_sched_pipe *pipe,
162 			    struct lima_sched_context *context,
163 			    atomic_t *guilty)
164 {
165 	struct drm_gpu_scheduler *sched = &pipe->base;
166 
167 	return drm_sched_entity_init(&context->base, DRM_SCHED_PRIORITY_NORMAL,
168 				     &sched, 1, guilty);
169 }
170 
171 void lima_sched_context_fini(struct lima_sched_pipe *pipe,
172 			     struct lima_sched_context *context)
173 {
174 	drm_sched_entity_fini(&context->base);
175 }
176 
177 struct dma_fence *lima_sched_context_queue_task(struct lima_sched_context *context,
178 						struct lima_sched_task *task)
179 {
180 	struct dma_fence *fence = dma_fence_get(&task->base.s_fence->finished);
181 
182 	trace_lima_task_submit(task);
183 	drm_sched_entity_push_job(&task->base, &context->base);
184 	return fence;
185 }
186 
187 static struct dma_fence *lima_sched_dependency(struct drm_sched_job *job,
188 					       struct drm_sched_entity *entity)
189 {
190 	struct lima_sched_task *task = to_lima_task(job);
191 
192 	if (!xa_empty(&task->deps))
193 		return xa_erase(&task->deps, task->last_dep++);
194 
195 	return NULL;
196 }
197 
198 static int lima_pm_busy(struct lima_device *ldev)
199 {
200 	int ret;
201 
202 	/* resume GPU if it has been suspended by runtime PM */
203 	ret = pm_runtime_get_sync(ldev->dev);
204 	if (ret < 0)
205 		return ret;
206 
207 	lima_devfreq_record_busy(&ldev->devfreq);
208 	return 0;
209 }
210 
211 static void lima_pm_idle(struct lima_device *ldev)
212 {
213 	lima_devfreq_record_idle(&ldev->devfreq);
214 
215 	/* GPU can do auto runtime suspend */
216 	pm_runtime_mark_last_busy(ldev->dev);
217 	pm_runtime_put_autosuspend(ldev->dev);
218 }
219 
220 static struct dma_fence *lima_sched_run_job(struct drm_sched_job *job)
221 {
222 	struct lima_sched_task *task = to_lima_task(job);
223 	struct lima_sched_pipe *pipe = to_lima_pipe(job->sched);
224 	struct lima_device *ldev = pipe->ldev;
225 	struct lima_fence *fence;
226 	struct dma_fence *ret;
227 	int i, err;
228 
229 	/* after GPU reset */
230 	if (job->s_fence->finished.error < 0)
231 		return NULL;
232 
233 	fence = lima_fence_create(pipe);
234 	if (!fence)
235 		return NULL;
236 
237 	err = lima_pm_busy(ldev);
238 	if (err < 0) {
239 		dma_fence_put(&fence->base);
240 		return NULL;
241 	}
242 
243 	task->fence = &fence->base;
244 
245 	/* for caller usage of the fence, otherwise irq handler
246 	 * may consume the fence before caller use it
247 	 */
248 	ret = dma_fence_get(task->fence);
249 
250 	pipe->current_task = task;
251 
252 	/* this is needed for MMU to work correctly, otherwise GP/PP
253 	 * will hang or page fault for unknown reason after running for
254 	 * a while.
255 	 *
256 	 * Need to investigate:
257 	 * 1. is it related to TLB
258 	 * 2. how much performance will be affected by L2 cache flush
259 	 * 3. can we reduce the calling of this function because all
260 	 *    GP/PP use the same L2 cache on mali400
261 	 *
262 	 * TODO:
263 	 * 1. move this to task fini to save some wait time?
264 	 * 2. when GP/PP use different l2 cache, need PP wait GP l2
265 	 *    cache flush?
266 	 */
267 	for (i = 0; i < pipe->num_l2_cache; i++)
268 		lima_l2_cache_flush(pipe->l2_cache[i]);
269 
270 	lima_vm_put(pipe->current_vm);
271 	pipe->current_vm = lima_vm_get(task->vm);
272 
273 	if (pipe->bcast_mmu)
274 		lima_mmu_switch_vm(pipe->bcast_mmu, pipe->current_vm);
275 	else {
276 		for (i = 0; i < pipe->num_mmu; i++)
277 			lima_mmu_switch_vm(pipe->mmu[i], pipe->current_vm);
278 	}
279 
280 	trace_lima_task_run(task);
281 
282 	pipe->error = false;
283 	pipe->task_run(pipe, task);
284 
285 	return task->fence;
286 }
287 
288 static void lima_sched_build_error_task_list(struct lima_sched_task *task)
289 {
290 	struct lima_sched_error_task *et;
291 	struct lima_sched_pipe *pipe = to_lima_pipe(task->base.sched);
292 	struct lima_ip *ip = pipe->processor[0];
293 	int pipe_id = ip->id == lima_ip_gp ? lima_pipe_gp : lima_pipe_pp;
294 	struct lima_device *dev = ip->dev;
295 	struct lima_sched_context *sched_ctx =
296 		container_of(task->base.entity,
297 			     struct lima_sched_context, base);
298 	struct lima_ctx *ctx =
299 		container_of(sched_ctx, struct lima_ctx, context[pipe_id]);
300 	struct lima_dump_task *dt;
301 	struct lima_dump_chunk *chunk;
302 	struct lima_dump_chunk_pid *pid_chunk;
303 	struct lima_dump_chunk_buffer *buffer_chunk;
304 	u32 size, task_size, mem_size;
305 	int i;
306 
307 	mutex_lock(&dev->error_task_list_lock);
308 
309 	if (dev->dump.num_tasks >= lima_max_error_tasks) {
310 		dev_info(dev->dev, "fail to save task state from %s pid %d: "
311 			 "error task list is full\n", ctx->pname, ctx->pid);
312 		goto out;
313 	}
314 
315 	/* frame chunk */
316 	size = sizeof(struct lima_dump_chunk) + pipe->frame_size;
317 	/* process name chunk */
318 	size += sizeof(struct lima_dump_chunk) + sizeof(ctx->pname);
319 	/* pid chunk */
320 	size += sizeof(struct lima_dump_chunk);
321 	/* buffer chunks */
322 	for (i = 0; i < task->num_bos; i++) {
323 		struct lima_bo *bo = task->bos[i];
324 
325 		size += sizeof(struct lima_dump_chunk);
326 		size += bo->heap_size ? bo->heap_size : lima_bo_size(bo);
327 	}
328 
329 	task_size = size + sizeof(struct lima_dump_task);
330 	mem_size = task_size + sizeof(*et);
331 	et = kvmalloc(mem_size, GFP_KERNEL);
332 	if (!et) {
333 		dev_err(dev->dev, "fail to alloc task dump buffer of size %x\n",
334 			mem_size);
335 		goto out;
336 	}
337 
338 	et->data = et + 1;
339 	et->size = task_size;
340 
341 	dt = et->data;
342 	memset(dt, 0, sizeof(*dt));
343 	dt->id = pipe_id;
344 	dt->size = size;
345 
346 	chunk = (struct lima_dump_chunk *)(dt + 1);
347 	memset(chunk, 0, sizeof(*chunk));
348 	chunk->id = LIMA_DUMP_CHUNK_FRAME;
349 	chunk->size = pipe->frame_size;
350 	memcpy(chunk + 1, task->frame, pipe->frame_size);
351 	dt->num_chunks++;
352 
353 	chunk = (void *)(chunk + 1) + chunk->size;
354 	memset(chunk, 0, sizeof(*chunk));
355 	chunk->id = LIMA_DUMP_CHUNK_PROCESS_NAME;
356 	chunk->size = sizeof(ctx->pname);
357 	memcpy(chunk + 1, ctx->pname, sizeof(ctx->pname));
358 	dt->num_chunks++;
359 
360 	pid_chunk = (void *)(chunk + 1) + chunk->size;
361 	memset(pid_chunk, 0, sizeof(*pid_chunk));
362 	pid_chunk->id = LIMA_DUMP_CHUNK_PROCESS_ID;
363 	pid_chunk->pid = ctx->pid;
364 	dt->num_chunks++;
365 
366 	buffer_chunk = (void *)(pid_chunk + 1) + pid_chunk->size;
367 	for (i = 0; i < task->num_bos; i++) {
368 		struct lima_bo *bo = task->bos[i];
369 		void *data;
370 
371 		memset(buffer_chunk, 0, sizeof(*buffer_chunk));
372 		buffer_chunk->id = LIMA_DUMP_CHUNK_BUFFER;
373 		buffer_chunk->va = lima_vm_get_va(task->vm, bo);
374 
375 		if (bo->heap_size) {
376 			buffer_chunk->size = bo->heap_size;
377 
378 			data = vmap(bo->base.pages, bo->heap_size >> PAGE_SHIFT,
379 				    VM_MAP, pgprot_writecombine(PAGE_KERNEL));
380 			if (!data) {
381 				kvfree(et);
382 				goto out;
383 			}
384 
385 			memcpy(buffer_chunk + 1, data, buffer_chunk->size);
386 
387 			vunmap(data);
388 		} else {
389 			buffer_chunk->size = lima_bo_size(bo);
390 
391 			data = drm_gem_shmem_vmap(&bo->base.base);
392 			if (IS_ERR_OR_NULL(data)) {
393 				kvfree(et);
394 				goto out;
395 			}
396 
397 			memcpy(buffer_chunk + 1, data, buffer_chunk->size);
398 
399 			drm_gem_shmem_vunmap(&bo->base.base, data);
400 		}
401 
402 		buffer_chunk = (void *)(buffer_chunk + 1) + buffer_chunk->size;
403 		dt->num_chunks++;
404 	}
405 
406 	list_add(&et->list, &dev->error_task_list);
407 	dev->dump.size += et->size;
408 	dev->dump.num_tasks++;
409 
410 	dev_info(dev->dev, "save error task state success\n");
411 
412 out:
413 	mutex_unlock(&dev->error_task_list_lock);
414 }
415 
416 static void lima_sched_timedout_job(struct drm_sched_job *job)
417 {
418 	struct lima_sched_pipe *pipe = to_lima_pipe(job->sched);
419 	struct lima_sched_task *task = to_lima_task(job);
420 	struct lima_device *ldev = pipe->ldev;
421 
422 	if (!pipe->error)
423 		DRM_ERROR("lima job timeout\n");
424 
425 	drm_sched_stop(&pipe->base, &task->base);
426 
427 	drm_sched_increase_karma(&task->base);
428 
429 	lima_sched_build_error_task_list(task);
430 
431 	pipe->task_error(pipe);
432 
433 	if (pipe->bcast_mmu)
434 		lima_mmu_page_fault_resume(pipe->bcast_mmu);
435 	else {
436 		int i;
437 
438 		for (i = 0; i < pipe->num_mmu; i++)
439 			lima_mmu_page_fault_resume(pipe->mmu[i]);
440 	}
441 
442 	lima_vm_put(pipe->current_vm);
443 	pipe->current_vm = NULL;
444 	pipe->current_task = NULL;
445 
446 	lima_pm_idle(ldev);
447 
448 	drm_sched_resubmit_jobs(&pipe->base);
449 	drm_sched_start(&pipe->base, true);
450 }
451 
452 static void lima_sched_free_job(struct drm_sched_job *job)
453 {
454 	struct lima_sched_task *task = to_lima_task(job);
455 	struct lima_sched_pipe *pipe = to_lima_pipe(job->sched);
456 	struct lima_vm *vm = task->vm;
457 	struct lima_bo **bos = task->bos;
458 	int i;
459 
460 	dma_fence_put(task->fence);
461 
462 	for (i = 0; i < task->num_bos; i++)
463 		lima_vm_bo_del(vm, bos[i]);
464 
465 	lima_sched_task_fini(task);
466 	kmem_cache_free(pipe->task_slab, task);
467 }
468 
469 static const struct drm_sched_backend_ops lima_sched_ops = {
470 	.dependency = lima_sched_dependency,
471 	.run_job = lima_sched_run_job,
472 	.timedout_job = lima_sched_timedout_job,
473 	.free_job = lima_sched_free_job,
474 };
475 
476 static void lima_sched_recover_work(struct work_struct *work)
477 {
478 	struct lima_sched_pipe *pipe =
479 		container_of(work, struct lima_sched_pipe, recover_work);
480 	int i;
481 
482 	for (i = 0; i < pipe->num_l2_cache; i++)
483 		lima_l2_cache_flush(pipe->l2_cache[i]);
484 
485 	if (pipe->bcast_mmu) {
486 		lima_mmu_flush_tlb(pipe->bcast_mmu);
487 	} else {
488 		for (i = 0; i < pipe->num_mmu; i++)
489 			lima_mmu_flush_tlb(pipe->mmu[i]);
490 	}
491 
492 	if (pipe->task_recover(pipe))
493 		drm_sched_fault(&pipe->base);
494 }
495 
496 int lima_sched_pipe_init(struct lima_sched_pipe *pipe, const char *name)
497 {
498 	unsigned int timeout = lima_sched_timeout_ms > 0 ?
499 			       lima_sched_timeout_ms : 500;
500 
501 	pipe->fence_context = dma_fence_context_alloc(1);
502 	spin_lock_init(&pipe->fence_lock);
503 
504 	INIT_WORK(&pipe->recover_work, lima_sched_recover_work);
505 
506 	return drm_sched_init(&pipe->base, &lima_sched_ops, 1, 0,
507 			      msecs_to_jiffies(timeout), name);
508 }
509 
510 void lima_sched_pipe_fini(struct lima_sched_pipe *pipe)
511 {
512 	drm_sched_fini(&pipe->base);
513 }
514 
515 void lima_sched_pipe_task_done(struct lima_sched_pipe *pipe)
516 {
517 	struct lima_sched_task *task = pipe->current_task;
518 	struct lima_device *ldev = pipe->ldev;
519 
520 	if (pipe->error) {
521 		if (task && task->recoverable)
522 			schedule_work(&pipe->recover_work);
523 		else
524 			drm_sched_fault(&pipe->base);
525 	} else {
526 		pipe->task_fini(pipe);
527 		dma_fence_signal(task->fence);
528 
529 		lima_pm_idle(ldev);
530 	}
531 }
532