xref: /openbsd-src/sys/dev/pci/drm/amd/amdgpu/amdgpu_ctx.c (revision c020cf82e0cc147236f01a8dca7052034cf9d30d)
1 /*
2  * Copyright 2015 Advanced Micro Devices, Inc.
3  *
4  * Permission is hereby granted, free of charge, to any person obtaining a
5  * copy of this software and associated documentation files (the "Software"),
6  * to deal in the Software without restriction, including without limitation
7  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8  * and/or sell copies of the Software, and to permit persons to whom the
9  * Software is furnished to do so, subject to the following conditions:
10  *
11  * The above copyright notice and this permission notice shall be included in
12  * all copies or substantial portions of the Software.
13  *
14  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
15  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
16  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
17  * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
18  * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
19  * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
20  * OTHER DEALINGS IN THE SOFTWARE.
21  *
22  * Authors: monk liu <monk.liu@amd.com>
23  */
24 
25 #include <drm/drm_auth.h>
26 #include "amdgpu.h"
27 #include "amdgpu_sched.h"
28 #include "amdgpu_ras.h"
29 
30 #define to_amdgpu_ctx_entity(e)	\
31 	container_of((e), struct amdgpu_ctx_entity, entity)
32 
33 const unsigned int amdgpu_ctx_num_entities[AMDGPU_HW_IP_NUM] = {
34 	[AMDGPU_HW_IP_GFX]	=	1,
35 	[AMDGPU_HW_IP_COMPUTE]	=	4,
36 	[AMDGPU_HW_IP_DMA]	=	2,
37 	[AMDGPU_HW_IP_UVD]	=	1,
38 	[AMDGPU_HW_IP_VCE]	=	1,
39 	[AMDGPU_HW_IP_UVD_ENC]	=	1,
40 	[AMDGPU_HW_IP_VCN_DEC]	=	1,
41 	[AMDGPU_HW_IP_VCN_ENC]	=	1,
42 	[AMDGPU_HW_IP_VCN_JPEG]	=	1,
43 };
44 
45 static int amdgpu_ctx_priority_permit(struct drm_file *filp,
46 				      enum drm_sched_priority priority)
47 {
48 	if (priority < 0 || priority >= DRM_SCHED_PRIORITY_MAX)
49 		return -EINVAL;
50 
51 	/* NORMAL and below are accessible by everyone */
52 	if (priority <= DRM_SCHED_PRIORITY_NORMAL)
53 		return 0;
54 
55 #ifdef notyet
56 	if (capable(CAP_SYS_NICE))
57 		return 0;
58 #endif
59 
60 	if (drm_is_current_master(filp))
61 		return 0;
62 
63 	return -EACCES;
64 }
65 
66 static enum gfx_pipe_priority amdgpu_ctx_sched_prio_to_compute_prio(enum drm_sched_priority prio)
67 {
68 	switch (prio) {
69 	case DRM_SCHED_PRIORITY_HIGH_HW:
70 	case DRM_SCHED_PRIORITY_KERNEL:
71 		return AMDGPU_GFX_PIPE_PRIO_HIGH;
72 	default:
73 		return AMDGPU_GFX_PIPE_PRIO_NORMAL;
74 	}
75 }
76 
77 static int amdgpu_ctx_init_entity(struct amdgpu_ctx *ctx, const u32 hw_ip, const u32 ring)
78 {
79 	struct amdgpu_device *adev = ctx->adev;
80 	struct amdgpu_ctx_entity *entity;
81 	struct drm_gpu_scheduler **scheds = NULL, *sched = NULL;
82 	unsigned num_scheds = 0;
83 	enum gfx_pipe_priority hw_prio;
84 	enum drm_sched_priority priority;
85 	int r;
86 
87 	entity = kcalloc(1, offsetof(typeof(*entity), fences[amdgpu_sched_jobs]),
88 			 GFP_KERNEL);
89 	if (!entity)
90 		return  -ENOMEM;
91 
92 	entity->sequence = 1;
93 	priority = (ctx->override_priority == DRM_SCHED_PRIORITY_UNSET) ?
94 				ctx->init_priority : ctx->override_priority;
95 	switch (hw_ip) {
96 	case AMDGPU_HW_IP_GFX:
97 		sched = &adev->gfx.gfx_ring[0].sched;
98 		scheds = &sched;
99 		num_scheds = 1;
100 		break;
101 	case AMDGPU_HW_IP_COMPUTE:
102 		hw_prio = amdgpu_ctx_sched_prio_to_compute_prio(priority);
103 		scheds = adev->gfx.compute_prio_sched[hw_prio];
104 		num_scheds = adev->gfx.num_compute_sched[hw_prio];
105 		break;
106 	case AMDGPU_HW_IP_DMA:
107 		scheds = adev->sdma.sdma_sched;
108 		num_scheds = adev->sdma.num_sdma_sched;
109 		break;
110 	case AMDGPU_HW_IP_UVD:
111 		sched = &adev->uvd.inst[0].ring.sched;
112 		scheds = &sched;
113 		num_scheds = 1;
114 		break;
115 	case AMDGPU_HW_IP_VCE:
116 		sched = &adev->vce.ring[0].sched;
117 		scheds = &sched;
118 		num_scheds = 1;
119 		break;
120 	case AMDGPU_HW_IP_UVD_ENC:
121 		sched = &adev->uvd.inst[0].ring_enc[0].sched;
122 		scheds = &sched;
123 		num_scheds = 1;
124 		break;
125 	case AMDGPU_HW_IP_VCN_DEC:
126 		sched = drm_sched_pick_best(adev->vcn.vcn_dec_sched,
127 					    adev->vcn.num_vcn_dec_sched);
128 		scheds = &sched;
129 		num_scheds = 1;
130 		break;
131 	case AMDGPU_HW_IP_VCN_ENC:
132 		sched = drm_sched_pick_best(adev->vcn.vcn_enc_sched,
133 					    adev->vcn.num_vcn_enc_sched);
134 		scheds = &sched;
135 		num_scheds = 1;
136 		break;
137 	case AMDGPU_HW_IP_VCN_JPEG:
138 		scheds = adev->jpeg.jpeg_sched;
139 		num_scheds =  adev->jpeg.num_jpeg_sched;
140 		break;
141 	}
142 
143 	r = drm_sched_entity_init(&entity->entity, priority, scheds, num_scheds,
144 				  &ctx->guilty);
145 	if (r)
146 		goto error_free_entity;
147 
148 	ctx->entities[hw_ip][ring] = entity;
149 	return 0;
150 
151 error_free_entity:
152 	kfree(entity);
153 
154 	return r;
155 }
156 
157 static int amdgpu_ctx_init(struct amdgpu_device *adev,
158 			   enum drm_sched_priority priority,
159 			   struct drm_file *filp,
160 			   struct amdgpu_ctx *ctx)
161 {
162 	int r;
163 
164 	r = amdgpu_ctx_priority_permit(filp, priority);
165 	if (r)
166 		return r;
167 
168 	memset(ctx, 0, sizeof(*ctx));
169 
170 	ctx->adev = adev;
171 
172 	kref_init(&ctx->refcount);
173 	mtx_init(&ctx->ring_lock, IPL_TTY);
174 	rw_init(&ctx->lock, "amctxlk");
175 
176 	ctx->reset_counter = atomic_read(&adev->gpu_reset_counter);
177 	ctx->reset_counter_query = ctx->reset_counter;
178 	ctx->vram_lost_counter = atomic_read(&adev->vram_lost_counter);
179 	ctx->init_priority = priority;
180 	ctx->override_priority = DRM_SCHED_PRIORITY_UNSET;
181 
182 	return 0;
183 
184 }
185 
186 static void amdgpu_ctx_fini_entity(struct amdgpu_ctx_entity *entity)
187 {
188 
189 	int i;
190 
191 	if (!entity)
192 		return;
193 
194 	for (i = 0; i < amdgpu_sched_jobs; ++i)
195 		dma_fence_put(entity->fences[i]);
196 
197 	kfree(entity);
198 }
199 
200 static void amdgpu_ctx_fini(struct kref *ref)
201 {
202 	struct amdgpu_ctx *ctx = container_of(ref, struct amdgpu_ctx, refcount);
203 	struct amdgpu_device *adev = ctx->adev;
204 	unsigned i, j;
205 
206 	if (!adev)
207 		return;
208 
209 	for (i = 0; i < AMDGPU_HW_IP_NUM; ++i) {
210 		for (j = 0; j < AMDGPU_MAX_ENTITY_NUM; ++j) {
211 			amdgpu_ctx_fini_entity(ctx->entities[i][j]);
212 			ctx->entities[i][j] = NULL;
213 		}
214 	}
215 
216 	mutex_destroy(&ctx->lock);
217 	kfree(ctx);
218 }
219 
220 int amdgpu_ctx_get_entity(struct amdgpu_ctx *ctx, u32 hw_ip, u32 instance,
221 			  u32 ring, struct drm_sched_entity **entity)
222 {
223 	int r;
224 
225 	if (hw_ip >= AMDGPU_HW_IP_NUM) {
226 		DRM_ERROR("unknown HW IP type: %d\n", hw_ip);
227 		return -EINVAL;
228 	}
229 
230 	/* Right now all IPs have only one instance - multiple rings. */
231 	if (instance != 0) {
232 		DRM_DEBUG("invalid ip instance: %d\n", instance);
233 		return -EINVAL;
234 	}
235 
236 	if (ring >= amdgpu_ctx_num_entities[hw_ip]) {
237 		DRM_DEBUG("invalid ring: %d %d\n", hw_ip, ring);
238 		return -EINVAL;
239 	}
240 
241 	if (ctx->entities[hw_ip][ring] == NULL) {
242 		r = amdgpu_ctx_init_entity(ctx, hw_ip, ring);
243 		if (r)
244 			return r;
245 	}
246 
247 	*entity = &ctx->entities[hw_ip][ring]->entity;
248 	return 0;
249 }
250 
251 static int amdgpu_ctx_alloc(struct amdgpu_device *adev,
252 			    struct amdgpu_fpriv *fpriv,
253 			    struct drm_file *filp,
254 			    enum drm_sched_priority priority,
255 			    uint32_t *id)
256 {
257 	struct amdgpu_ctx_mgr *mgr = &fpriv->ctx_mgr;
258 	struct amdgpu_ctx *ctx;
259 	int r;
260 
261 	ctx = kmalloc(sizeof(*ctx), GFP_KERNEL);
262 	if (!ctx)
263 		return -ENOMEM;
264 
265 	mutex_lock(&mgr->lock);
266 	r = idr_alloc(&mgr->ctx_handles, ctx, 1, AMDGPU_VM_MAX_NUM_CTX, GFP_KERNEL);
267 	if (r < 0) {
268 		mutex_unlock(&mgr->lock);
269 		kfree(ctx);
270 		return r;
271 	}
272 
273 	*id = (uint32_t)r;
274 	r = amdgpu_ctx_init(adev, priority, filp, ctx);
275 	if (r) {
276 		idr_remove(&mgr->ctx_handles, *id);
277 		*id = 0;
278 		kfree(ctx);
279 	}
280 	mutex_unlock(&mgr->lock);
281 	return r;
282 }
283 
284 static void amdgpu_ctx_do_release(struct kref *ref)
285 {
286 	struct amdgpu_ctx *ctx;
287 	u32 i, j;
288 
289 	ctx = container_of(ref, struct amdgpu_ctx, refcount);
290 	for (i = 0; i < AMDGPU_HW_IP_NUM; ++i) {
291 		for (j = 0; j < amdgpu_ctx_num_entities[i]; ++j) {
292 			if (!ctx->entities[i][j])
293 				continue;
294 
295 			drm_sched_entity_destroy(&ctx->entities[i][j]->entity);
296 		}
297 	}
298 
299 	amdgpu_ctx_fini(ref);
300 }
301 
302 static int amdgpu_ctx_free(struct amdgpu_fpriv *fpriv, uint32_t id)
303 {
304 	struct amdgpu_ctx_mgr *mgr = &fpriv->ctx_mgr;
305 	struct amdgpu_ctx *ctx;
306 
307 	mutex_lock(&mgr->lock);
308 	ctx = idr_remove(&mgr->ctx_handles, id);
309 	if (ctx)
310 		kref_put(&ctx->refcount, amdgpu_ctx_do_release);
311 	mutex_unlock(&mgr->lock);
312 	return ctx ? 0 : -EINVAL;
313 }
314 
315 static int amdgpu_ctx_query(struct amdgpu_device *adev,
316 			    struct amdgpu_fpriv *fpriv, uint32_t id,
317 			    union drm_amdgpu_ctx_out *out)
318 {
319 	struct amdgpu_ctx *ctx;
320 	struct amdgpu_ctx_mgr *mgr;
321 	unsigned reset_counter;
322 
323 	if (!fpriv)
324 		return -EINVAL;
325 
326 	mgr = &fpriv->ctx_mgr;
327 	mutex_lock(&mgr->lock);
328 	ctx = idr_find(&mgr->ctx_handles, id);
329 	if (!ctx) {
330 		mutex_unlock(&mgr->lock);
331 		return -EINVAL;
332 	}
333 
334 	/* TODO: these two are always zero */
335 	out->state.flags = 0x0;
336 	out->state.hangs = 0x0;
337 
338 	/* determine if a GPU reset has occured since the last call */
339 	reset_counter = atomic_read(&adev->gpu_reset_counter);
340 	/* TODO: this should ideally return NO, GUILTY, or INNOCENT. */
341 	if (ctx->reset_counter_query == reset_counter)
342 		out->state.reset_status = AMDGPU_CTX_NO_RESET;
343 	else
344 		out->state.reset_status = AMDGPU_CTX_UNKNOWN_RESET;
345 	ctx->reset_counter_query = reset_counter;
346 
347 	mutex_unlock(&mgr->lock);
348 	return 0;
349 }
350 
351 static int amdgpu_ctx_query2(struct amdgpu_device *adev,
352 	struct amdgpu_fpriv *fpriv, uint32_t id,
353 	union drm_amdgpu_ctx_out *out)
354 {
355 	struct amdgpu_ctx *ctx;
356 	struct amdgpu_ctx_mgr *mgr;
357 	unsigned long ras_counter;
358 
359 	if (!fpriv)
360 		return -EINVAL;
361 
362 	mgr = &fpriv->ctx_mgr;
363 	mutex_lock(&mgr->lock);
364 	ctx = idr_find(&mgr->ctx_handles, id);
365 	if (!ctx) {
366 		mutex_unlock(&mgr->lock);
367 		return -EINVAL;
368 	}
369 
370 	out->state.flags = 0x0;
371 	out->state.hangs = 0x0;
372 
373 	if (ctx->reset_counter != atomic_read(&adev->gpu_reset_counter))
374 		out->state.flags |= AMDGPU_CTX_QUERY2_FLAGS_RESET;
375 
376 	if (ctx->vram_lost_counter != atomic_read(&adev->vram_lost_counter))
377 		out->state.flags |= AMDGPU_CTX_QUERY2_FLAGS_VRAMLOST;
378 
379 	if (atomic_read(&ctx->guilty))
380 		out->state.flags |= AMDGPU_CTX_QUERY2_FLAGS_GUILTY;
381 
382 	/*query ue count*/
383 	ras_counter = amdgpu_ras_query_error_count(adev, false);
384 	/*ras counter is monotonic increasing*/
385 	if (ras_counter != ctx->ras_counter_ue) {
386 		out->state.flags |= AMDGPU_CTX_QUERY2_FLAGS_RAS_UE;
387 		ctx->ras_counter_ue = ras_counter;
388 	}
389 
390 	/*query ce count*/
391 	ras_counter = amdgpu_ras_query_error_count(adev, true);
392 	if (ras_counter != ctx->ras_counter_ce) {
393 		out->state.flags |= AMDGPU_CTX_QUERY2_FLAGS_RAS_CE;
394 		ctx->ras_counter_ce = ras_counter;
395 	}
396 
397 	mutex_unlock(&mgr->lock);
398 	return 0;
399 }
400 
401 int amdgpu_ctx_ioctl(struct drm_device *dev, void *data,
402 		     struct drm_file *filp)
403 {
404 	int r;
405 	uint32_t id;
406 	enum drm_sched_priority priority;
407 
408 	union drm_amdgpu_ctx *args = data;
409 	struct amdgpu_device *adev = dev->dev_private;
410 	struct amdgpu_fpriv *fpriv = filp->driver_priv;
411 
412 	r = 0;
413 	id = args->in.ctx_id;
414 	priority = amdgpu_to_sched_priority(args->in.priority);
415 
416 	/* For backwards compatibility reasons, we need to accept
417 	 * ioctls with garbage in the priority field */
418 	if (priority == DRM_SCHED_PRIORITY_INVALID)
419 		priority = DRM_SCHED_PRIORITY_NORMAL;
420 
421 	switch (args->in.op) {
422 	case AMDGPU_CTX_OP_ALLOC_CTX:
423 		r = amdgpu_ctx_alloc(adev, fpriv, filp, priority, &id);
424 		args->out.alloc.ctx_id = id;
425 		break;
426 	case AMDGPU_CTX_OP_FREE_CTX:
427 		r = amdgpu_ctx_free(fpriv, id);
428 		break;
429 	case AMDGPU_CTX_OP_QUERY_STATE:
430 		r = amdgpu_ctx_query(adev, fpriv, id, &args->out);
431 		break;
432 	case AMDGPU_CTX_OP_QUERY_STATE2:
433 		r = amdgpu_ctx_query2(adev, fpriv, id, &args->out);
434 		break;
435 	default:
436 		return -EINVAL;
437 	}
438 
439 	return r;
440 }
441 
442 struct amdgpu_ctx *amdgpu_ctx_get(struct amdgpu_fpriv *fpriv, uint32_t id)
443 {
444 	struct amdgpu_ctx *ctx;
445 	struct amdgpu_ctx_mgr *mgr;
446 
447 	if (!fpriv)
448 		return NULL;
449 
450 	mgr = &fpriv->ctx_mgr;
451 
452 	mutex_lock(&mgr->lock);
453 	ctx = idr_find(&mgr->ctx_handles, id);
454 	if (ctx)
455 		kref_get(&ctx->refcount);
456 	mutex_unlock(&mgr->lock);
457 	return ctx;
458 }
459 
460 int amdgpu_ctx_put(struct amdgpu_ctx *ctx)
461 {
462 	if (ctx == NULL)
463 		return -EINVAL;
464 
465 	kref_put(&ctx->refcount, amdgpu_ctx_do_release);
466 	return 0;
467 }
468 
469 void amdgpu_ctx_add_fence(struct amdgpu_ctx *ctx,
470 			  struct drm_sched_entity *entity,
471 			  struct dma_fence *fence, uint64_t* handle)
472 {
473 	struct amdgpu_ctx_entity *centity = to_amdgpu_ctx_entity(entity);
474 	uint64_t seq = centity->sequence;
475 	struct dma_fence *other = NULL;
476 	unsigned idx = 0;
477 
478 	idx = seq & (amdgpu_sched_jobs - 1);
479 	other = centity->fences[idx];
480 	if (other)
481 		BUG_ON(!dma_fence_is_signaled(other));
482 
483 	dma_fence_get(fence);
484 
485 	spin_lock(&ctx->ring_lock);
486 	centity->fences[idx] = fence;
487 	centity->sequence++;
488 	spin_unlock(&ctx->ring_lock);
489 
490 	dma_fence_put(other);
491 	if (handle)
492 		*handle = seq;
493 }
494 
495 struct dma_fence *amdgpu_ctx_get_fence(struct amdgpu_ctx *ctx,
496 				       struct drm_sched_entity *entity,
497 				       uint64_t seq)
498 {
499 	struct amdgpu_ctx_entity *centity = to_amdgpu_ctx_entity(entity);
500 	struct dma_fence *fence;
501 
502 	spin_lock(&ctx->ring_lock);
503 
504 	if (seq == ~0ull)
505 		seq = centity->sequence - 1;
506 
507 	if (seq >= centity->sequence) {
508 		spin_unlock(&ctx->ring_lock);
509 		return ERR_PTR(-EINVAL);
510 	}
511 
512 
513 	if (seq + amdgpu_sched_jobs < centity->sequence) {
514 		spin_unlock(&ctx->ring_lock);
515 		return NULL;
516 	}
517 
518 	fence = dma_fence_get(centity->fences[seq & (amdgpu_sched_jobs - 1)]);
519 	spin_unlock(&ctx->ring_lock);
520 
521 	return fence;
522 }
523 
524 static void amdgpu_ctx_set_entity_priority(struct amdgpu_ctx *ctx,
525 					    struct amdgpu_ctx_entity *aentity,
526 					    int hw_ip,
527 					    enum drm_sched_priority priority)
528 {
529 	struct amdgpu_device *adev = ctx->adev;
530 	enum gfx_pipe_priority hw_prio;
531 	struct drm_gpu_scheduler **scheds = NULL;
532 	unsigned num_scheds;
533 
534 	/* set sw priority */
535 	drm_sched_entity_set_priority(&aentity->entity, priority);
536 
537 	/* set hw priority */
538 	if (hw_ip == AMDGPU_HW_IP_COMPUTE) {
539 		hw_prio = amdgpu_ctx_sched_prio_to_compute_prio(priority);
540 		scheds = adev->gfx.compute_prio_sched[hw_prio];
541 		num_scheds = adev->gfx.num_compute_sched[hw_prio];
542 		drm_sched_entity_modify_sched(&aentity->entity, scheds,
543 					      num_scheds);
544 	}
545 }
546 
547 void amdgpu_ctx_priority_override(struct amdgpu_ctx *ctx,
548 				  enum drm_sched_priority priority)
549 {
550 	enum drm_sched_priority ctx_prio;
551 	unsigned i, j;
552 
553 	ctx->override_priority = priority;
554 
555 	ctx_prio = (ctx->override_priority == DRM_SCHED_PRIORITY_UNSET) ?
556 			ctx->init_priority : ctx->override_priority;
557 	for (i = 0; i < AMDGPU_HW_IP_NUM; ++i) {
558 		for (j = 0; j < amdgpu_ctx_num_entities[i]; ++j) {
559 			if (!ctx->entities[i][j])
560 				continue;
561 
562 			amdgpu_ctx_set_entity_priority(ctx, ctx->entities[i][j],
563 						       i, ctx_prio);
564 		}
565 	}
566 }
567 
568 int amdgpu_ctx_wait_prev_fence(struct amdgpu_ctx *ctx,
569 			       struct drm_sched_entity *entity)
570 {
571 	struct amdgpu_ctx_entity *centity = to_amdgpu_ctx_entity(entity);
572 	struct dma_fence *other;
573 	unsigned idx;
574 	long r;
575 
576 	spin_lock(&ctx->ring_lock);
577 	idx = centity->sequence & (amdgpu_sched_jobs - 1);
578 	other = dma_fence_get(centity->fences[idx]);
579 	spin_unlock(&ctx->ring_lock);
580 
581 	if (!other)
582 		return 0;
583 
584 	r = dma_fence_wait(other, true);
585 	if (r < 0 && r != -ERESTARTSYS)
586 		DRM_ERROR("Error (%ld) waiting for fence!\n", r);
587 
588 	dma_fence_put(other);
589 	return r;
590 }
591 
592 void amdgpu_ctx_mgr_init(struct amdgpu_ctx_mgr *mgr)
593 {
594 	rw_init(&mgr->lock, "mgrlk");
595 	idr_init(&mgr->ctx_handles);
596 }
597 
598 long amdgpu_ctx_mgr_entity_flush(struct amdgpu_ctx_mgr *mgr, long timeout)
599 {
600 	struct amdgpu_ctx *ctx;
601 	struct idr *idp;
602 	uint32_t id, i, j;
603 
604 	idp = &mgr->ctx_handles;
605 
606 	mutex_lock(&mgr->lock);
607 	idr_for_each_entry(idp, ctx, id) {
608 		for (i = 0; i < AMDGPU_HW_IP_NUM; ++i) {
609 			for (j = 0; j < amdgpu_ctx_num_entities[i]; ++j) {
610 				struct drm_sched_entity *entity;
611 
612 				if (!ctx->entities[i][j])
613 					continue;
614 
615 				entity = &ctx->entities[i][j]->entity;
616 				timeout = drm_sched_entity_flush(entity, timeout);
617 			}
618 		}
619 	}
620 	mutex_unlock(&mgr->lock);
621 	return timeout;
622 }
623 
624 void amdgpu_ctx_mgr_entity_fini(struct amdgpu_ctx_mgr *mgr)
625 {
626 	struct amdgpu_ctx *ctx;
627 	struct idr *idp;
628 	uint32_t id, i, j;
629 
630 	idp = &mgr->ctx_handles;
631 
632 	idr_for_each_entry(idp, ctx, id) {
633 		if (kref_read(&ctx->refcount) != 1) {
634 			DRM_ERROR("ctx %p is still alive\n", ctx);
635 			continue;
636 		}
637 
638 		for (i = 0; i < AMDGPU_HW_IP_NUM; ++i) {
639 			for (j = 0; j < amdgpu_ctx_num_entities[i]; ++j) {
640 				struct drm_sched_entity *entity;
641 
642 				if (!ctx->entities[i][j])
643 					continue;
644 
645 				entity = &ctx->entities[i][j]->entity;
646 				drm_sched_entity_fini(entity);
647 			}
648 		}
649 	}
650 }
651 
652 void amdgpu_ctx_mgr_fini(struct amdgpu_ctx_mgr *mgr)
653 {
654 	struct amdgpu_ctx *ctx;
655 	struct idr *idp;
656 	uint32_t id;
657 
658 	amdgpu_ctx_mgr_entity_fini(mgr);
659 
660 	idp = &mgr->ctx_handles;
661 
662 	idr_for_each_entry(idp, ctx, id) {
663 		if (kref_put(&ctx->refcount, amdgpu_ctx_fini) != 1)
664 			DRM_ERROR("ctx %p is still alive\n", ctx);
665 	}
666 
667 	idr_destroy(&mgr->ctx_handles);
668 	mutex_destroy(&mgr->lock);
669 }
670 
671 
672 static void amdgpu_ctx_init_compute_sched(struct amdgpu_device *adev)
673 {
674 	int num_compute_sched_normal = 0;
675 	int num_compute_sched_high = AMDGPU_MAX_COMPUTE_RINGS - 1;
676 	int i;
677 
678 	/* use one drm sched array, gfx.compute_sched to store both high and
679 	 * normal priority drm compute schedulers */
680 	for (i = 0; i < adev->gfx.num_compute_rings; i++) {
681 		if (!adev->gfx.compute_ring[i].has_high_prio)
682 			adev->gfx.compute_sched[num_compute_sched_normal++] =
683 				&adev->gfx.compute_ring[i].sched;
684 		else
685 			adev->gfx.compute_sched[num_compute_sched_high--] =
686 				&adev->gfx.compute_ring[i].sched;
687 	}
688 
689 	/* compute ring only has two priority for now */
690 	i = AMDGPU_GFX_PIPE_PRIO_NORMAL;
691 	adev->gfx.compute_prio_sched[i] = &adev->gfx.compute_sched[0];
692 	adev->gfx.num_compute_sched[i] = num_compute_sched_normal;
693 
694 	i = AMDGPU_GFX_PIPE_PRIO_HIGH;
695 	if (num_compute_sched_high == (AMDGPU_MAX_COMPUTE_RINGS - 1)) {
696 		/* When compute has no high priority rings then use */
697 		/* normal priority sched array */
698 		adev->gfx.compute_prio_sched[i] = &adev->gfx.compute_sched[0];
699 		adev->gfx.num_compute_sched[i] = num_compute_sched_normal;
700 	} else {
701 		adev->gfx.compute_prio_sched[i] =
702 			&adev->gfx.compute_sched[num_compute_sched_high - 1];
703 		adev->gfx.num_compute_sched[i] =
704 			adev->gfx.num_compute_rings - num_compute_sched_normal;
705 	}
706 }
707 
708 void amdgpu_ctx_init_sched(struct amdgpu_device *adev)
709 {
710 	int i, j;
711 
712 	amdgpu_ctx_init_compute_sched(adev);
713 	for (i = 0; i < adev->gfx.num_gfx_rings; i++) {
714 		adev->gfx.gfx_sched[i] = &adev->gfx.gfx_ring[i].sched;
715 		adev->gfx.num_gfx_sched++;
716 	}
717 
718 	for (i = 0; i < adev->sdma.num_instances; i++) {
719 		adev->sdma.sdma_sched[i] = &adev->sdma.instance[i].ring.sched;
720 		adev->sdma.num_sdma_sched++;
721 	}
722 
723 	for (i = 0; i < adev->vcn.num_vcn_inst; ++i) {
724 		if (adev->vcn.harvest_config & (1 << i))
725 			continue;
726 		adev->vcn.vcn_dec_sched[adev->vcn.num_vcn_dec_sched++] =
727 			&adev->vcn.inst[i].ring_dec.sched;
728 	}
729 
730 	for (i = 0; i < adev->vcn.num_vcn_inst; ++i) {
731 		if (adev->vcn.harvest_config & (1 << i))
732 			continue;
733 		for (j = 0; j < adev->vcn.num_enc_rings; ++j)
734 			adev->vcn.vcn_enc_sched[adev->vcn.num_vcn_enc_sched++] =
735 				&adev->vcn.inst[i].ring_enc[j].sched;
736 	}
737 
738 	for (i = 0; i < adev->jpeg.num_jpeg_inst; ++i) {
739 		if (adev->jpeg.harvest_config & (1 << i))
740 			continue;
741 		adev->jpeg.jpeg_sched[adev->jpeg.num_jpeg_sched++] =
742 			&adev->jpeg.inst[i].ring_dec.sched;
743 	}
744 }
745