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 #include <linux/nospec.h>
30 
31 #define to_amdgpu_ctx_entity(e)	\
32 	container_of((e), struct amdgpu_ctx_entity, entity)
33 
34 const unsigned int amdgpu_ctx_num_entities[AMDGPU_HW_IP_NUM] = {
35 	[AMDGPU_HW_IP_GFX]	=	1,
36 	[AMDGPU_HW_IP_COMPUTE]	=	4,
37 	[AMDGPU_HW_IP_DMA]	=	2,
38 	[AMDGPU_HW_IP_UVD]	=	1,
39 	[AMDGPU_HW_IP_VCE]	=	1,
40 	[AMDGPU_HW_IP_UVD_ENC]	=	1,
41 	[AMDGPU_HW_IP_VCN_DEC]	=	1,
42 	[AMDGPU_HW_IP_VCN_ENC]	=	1,
43 	[AMDGPU_HW_IP_VCN_JPEG]	=	1,
44 };
45 
46 static int amdgpu_ctx_priority_permit(struct drm_file *filp,
47 				      enum drm_sched_priority priority)
48 {
49 	if (priority < 0 || priority >= DRM_SCHED_PRIORITY_COUNT)
50 		return -EINVAL;
51 
52 	/* NORMAL and below are accessible by everyone */
53 	if (priority <= DRM_SCHED_PRIORITY_NORMAL)
54 		return 0;
55 
56 	if (capable(CAP_SYS_NICE))
57 		return 0;
58 
59 	if (drm_is_current_master(filp))
60 		return 0;
61 
62 	return -EACCES;
63 }
64 
65 static enum gfx_pipe_priority amdgpu_ctx_sched_prio_to_compute_prio(enum drm_sched_priority prio)
66 {
67 	switch (prio) {
68 	case DRM_SCHED_PRIORITY_HIGH:
69 	case DRM_SCHED_PRIORITY_KERNEL:
70 		return AMDGPU_GFX_PIPE_PRIO_HIGH;
71 	default:
72 		return AMDGPU_GFX_PIPE_PRIO_NORMAL;
73 	}
74 }
75 
76 static unsigned int amdgpu_ctx_prio_sched_to_hw(struct amdgpu_device *adev,
77 						 enum drm_sched_priority prio,
78 						 u32 hw_ip)
79 {
80 	unsigned int hw_prio;
81 
82 	hw_prio = (hw_ip == AMDGPU_HW_IP_COMPUTE) ?
83 			amdgpu_ctx_sched_prio_to_compute_prio(prio) :
84 			AMDGPU_RING_PRIO_DEFAULT;
85 	hw_ip = array_index_nospec(hw_ip, AMDGPU_HW_IP_NUM);
86 	if (adev->gpu_sched[hw_ip][hw_prio].num_scheds == 0)
87 		hw_prio = AMDGPU_RING_PRIO_DEFAULT;
88 
89 	return hw_prio;
90 }
91 
92 static int amdgpu_ctx_init_entity(struct amdgpu_ctx *ctx, u32 hw_ip,
93 				   const u32 ring)
94 {
95 	struct amdgpu_device *adev = ctx->adev;
96 	struct amdgpu_ctx_entity *entity;
97 	struct drm_gpu_scheduler **scheds = NULL, *sched = NULL;
98 	unsigned num_scheds = 0;
99 	unsigned int hw_prio;
100 	enum drm_sched_priority priority;
101 	int r;
102 
103 	entity = kzalloc(struct_size(entity, fences, amdgpu_sched_jobs),
104 			 GFP_KERNEL);
105 	if (!entity)
106 		return  -ENOMEM;
107 
108 	entity->sequence = 1;
109 	priority = (ctx->override_priority == DRM_SCHED_PRIORITY_UNSET) ?
110 				ctx->init_priority : ctx->override_priority;
111 	hw_prio = amdgpu_ctx_prio_sched_to_hw(adev, priority, hw_ip);
112 
113 	hw_ip = array_index_nospec(hw_ip, AMDGPU_HW_IP_NUM);
114 	scheds = adev->gpu_sched[hw_ip][hw_prio].sched;
115 	num_scheds = adev->gpu_sched[hw_ip][hw_prio].num_scheds;
116 
117 	/* disable load balance if the hw engine retains context among dependent jobs */
118 	if (hw_ip == AMDGPU_HW_IP_VCN_ENC ||
119 	    hw_ip == AMDGPU_HW_IP_VCN_DEC ||
120 	    hw_ip == AMDGPU_HW_IP_UVD_ENC ||
121 	    hw_ip == AMDGPU_HW_IP_UVD) {
122 		sched = drm_sched_pick_best(scheds, num_scheds);
123 		scheds = &sched;
124 		num_scheds = 1;
125 	}
126 
127 	r = drm_sched_entity_init(&entity->entity, priority, scheds, num_scheds,
128 				  &ctx->guilty);
129 	if (r)
130 		goto error_free_entity;
131 
132 	ctx->entities[hw_ip][ring] = entity;
133 	return 0;
134 
135 error_free_entity:
136 	kfree(entity);
137 
138 	return r;
139 }
140 
141 static int amdgpu_ctx_init(struct amdgpu_device *adev,
142 			   enum drm_sched_priority priority,
143 			   struct drm_file *filp,
144 			   struct amdgpu_ctx *ctx)
145 {
146 	int r;
147 
148 	r = amdgpu_ctx_priority_permit(filp, priority);
149 	if (r)
150 		return r;
151 
152 	memset(ctx, 0, sizeof(*ctx));
153 
154 	ctx->adev = adev;
155 
156 	kref_init(&ctx->refcount);
157 	spin_lock_init(&ctx->ring_lock);
158 	mutex_init(&ctx->lock);
159 
160 	ctx->reset_counter = atomic_read(&adev->gpu_reset_counter);
161 	ctx->reset_counter_query = ctx->reset_counter;
162 	ctx->vram_lost_counter = atomic_read(&adev->vram_lost_counter);
163 	ctx->init_priority = priority;
164 	ctx->override_priority = DRM_SCHED_PRIORITY_UNSET;
165 
166 	return 0;
167 }
168 
169 static void amdgpu_ctx_fini_entity(struct amdgpu_ctx_entity *entity)
170 {
171 
172 	int i;
173 
174 	if (!entity)
175 		return;
176 
177 	for (i = 0; i < amdgpu_sched_jobs; ++i)
178 		dma_fence_put(entity->fences[i]);
179 
180 	kfree(entity);
181 }
182 
183 static void amdgpu_ctx_fini(struct kref *ref)
184 {
185 	struct amdgpu_ctx *ctx = container_of(ref, struct amdgpu_ctx, refcount);
186 	struct amdgpu_device *adev = ctx->adev;
187 	unsigned i, j;
188 
189 	if (!adev)
190 		return;
191 
192 	for (i = 0; i < AMDGPU_HW_IP_NUM; ++i) {
193 		for (j = 0; j < AMDGPU_MAX_ENTITY_NUM; ++j) {
194 			amdgpu_ctx_fini_entity(ctx->entities[i][j]);
195 			ctx->entities[i][j] = NULL;
196 		}
197 	}
198 
199 	mutex_destroy(&ctx->lock);
200 	kfree(ctx);
201 }
202 
203 int amdgpu_ctx_get_entity(struct amdgpu_ctx *ctx, u32 hw_ip, u32 instance,
204 			  u32 ring, struct drm_sched_entity **entity)
205 {
206 	int r;
207 
208 	if (hw_ip >= AMDGPU_HW_IP_NUM) {
209 		DRM_ERROR("unknown HW IP type: %d\n", hw_ip);
210 		return -EINVAL;
211 	}
212 
213 	/* Right now all IPs have only one instance - multiple rings. */
214 	if (instance != 0) {
215 		DRM_DEBUG("invalid ip instance: %d\n", instance);
216 		return -EINVAL;
217 	}
218 
219 	if (ring >= amdgpu_ctx_num_entities[hw_ip]) {
220 		DRM_DEBUG("invalid ring: %d %d\n", hw_ip, ring);
221 		return -EINVAL;
222 	}
223 
224 	if (ctx->entities[hw_ip][ring] == NULL) {
225 		r = amdgpu_ctx_init_entity(ctx, hw_ip, ring);
226 		if (r)
227 			return r;
228 	}
229 
230 	*entity = &ctx->entities[hw_ip][ring]->entity;
231 	return 0;
232 }
233 
234 static int amdgpu_ctx_alloc(struct amdgpu_device *adev,
235 			    struct amdgpu_fpriv *fpriv,
236 			    struct drm_file *filp,
237 			    enum drm_sched_priority priority,
238 			    uint32_t *id)
239 {
240 	struct amdgpu_ctx_mgr *mgr = &fpriv->ctx_mgr;
241 	struct amdgpu_ctx *ctx;
242 	int r;
243 
244 	ctx = kmalloc(sizeof(*ctx), GFP_KERNEL);
245 	if (!ctx)
246 		return -ENOMEM;
247 
248 	mutex_lock(&mgr->lock);
249 	r = idr_alloc(&mgr->ctx_handles, ctx, 1, AMDGPU_VM_MAX_NUM_CTX, GFP_KERNEL);
250 	if (r < 0) {
251 		mutex_unlock(&mgr->lock);
252 		kfree(ctx);
253 		return r;
254 	}
255 
256 	*id = (uint32_t)r;
257 	r = amdgpu_ctx_init(adev, priority, filp, ctx);
258 	if (r) {
259 		idr_remove(&mgr->ctx_handles, *id);
260 		*id = 0;
261 		kfree(ctx);
262 	}
263 	mutex_unlock(&mgr->lock);
264 	return r;
265 }
266 
267 static void amdgpu_ctx_do_release(struct kref *ref)
268 {
269 	struct amdgpu_ctx *ctx;
270 	u32 i, j;
271 
272 	ctx = container_of(ref, struct amdgpu_ctx, refcount);
273 	for (i = 0; i < AMDGPU_HW_IP_NUM; ++i) {
274 		for (j = 0; j < amdgpu_ctx_num_entities[i]; ++j) {
275 			if (!ctx->entities[i][j])
276 				continue;
277 
278 			drm_sched_entity_destroy(&ctx->entities[i][j]->entity);
279 		}
280 	}
281 
282 	amdgpu_ctx_fini(ref);
283 }
284 
285 static int amdgpu_ctx_free(struct amdgpu_fpriv *fpriv, uint32_t id)
286 {
287 	struct amdgpu_ctx_mgr *mgr = &fpriv->ctx_mgr;
288 	struct amdgpu_ctx *ctx;
289 
290 	mutex_lock(&mgr->lock);
291 	ctx = idr_remove(&mgr->ctx_handles, id);
292 	if (ctx)
293 		kref_put(&ctx->refcount, amdgpu_ctx_do_release);
294 	mutex_unlock(&mgr->lock);
295 	return ctx ? 0 : -EINVAL;
296 }
297 
298 static int amdgpu_ctx_query(struct amdgpu_device *adev,
299 			    struct amdgpu_fpriv *fpriv, uint32_t id,
300 			    union drm_amdgpu_ctx_out *out)
301 {
302 	struct amdgpu_ctx *ctx;
303 	struct amdgpu_ctx_mgr *mgr;
304 	unsigned reset_counter;
305 
306 	if (!fpriv)
307 		return -EINVAL;
308 
309 	mgr = &fpriv->ctx_mgr;
310 	mutex_lock(&mgr->lock);
311 	ctx = idr_find(&mgr->ctx_handles, id);
312 	if (!ctx) {
313 		mutex_unlock(&mgr->lock);
314 		return -EINVAL;
315 	}
316 
317 	/* TODO: these two are always zero */
318 	out->state.flags = 0x0;
319 	out->state.hangs = 0x0;
320 
321 	/* determine if a GPU reset has occured since the last call */
322 	reset_counter = atomic_read(&adev->gpu_reset_counter);
323 	/* TODO: this should ideally return NO, GUILTY, or INNOCENT. */
324 	if (ctx->reset_counter_query == reset_counter)
325 		out->state.reset_status = AMDGPU_CTX_NO_RESET;
326 	else
327 		out->state.reset_status = AMDGPU_CTX_UNKNOWN_RESET;
328 	ctx->reset_counter_query = reset_counter;
329 
330 	mutex_unlock(&mgr->lock);
331 	return 0;
332 }
333 
334 #define AMDGPU_RAS_COUNTE_DELAY_MS 3000
335 
336 static int amdgpu_ctx_query2(struct amdgpu_device *adev,
337 			     struct amdgpu_fpriv *fpriv, uint32_t id,
338 			     union drm_amdgpu_ctx_out *out)
339 {
340 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
341 	struct amdgpu_ctx *ctx;
342 	struct amdgpu_ctx_mgr *mgr;
343 
344 	if (!fpriv)
345 		return -EINVAL;
346 
347 	mgr = &fpriv->ctx_mgr;
348 	mutex_lock(&mgr->lock);
349 	ctx = idr_find(&mgr->ctx_handles, id);
350 	if (!ctx) {
351 		mutex_unlock(&mgr->lock);
352 		return -EINVAL;
353 	}
354 
355 	out->state.flags = 0x0;
356 	out->state.hangs = 0x0;
357 
358 	if (ctx->reset_counter != atomic_read(&adev->gpu_reset_counter))
359 		out->state.flags |= AMDGPU_CTX_QUERY2_FLAGS_RESET;
360 
361 	if (ctx->vram_lost_counter != atomic_read(&adev->vram_lost_counter))
362 		out->state.flags |= AMDGPU_CTX_QUERY2_FLAGS_VRAMLOST;
363 
364 	if (atomic_read(&ctx->guilty))
365 		out->state.flags |= AMDGPU_CTX_QUERY2_FLAGS_GUILTY;
366 
367 	if (adev->ras_enabled && con) {
368 		/* Return the cached values in O(1),
369 		 * and schedule delayed work to cache
370 		 * new vaues.
371 		 */
372 		int ce_count, ue_count;
373 
374 		ce_count = atomic_read(&con->ras_ce_count);
375 		ue_count = atomic_read(&con->ras_ue_count);
376 
377 		if (ce_count != ctx->ras_counter_ce) {
378 			ctx->ras_counter_ce = ce_count;
379 			out->state.flags |= AMDGPU_CTX_QUERY2_FLAGS_RAS_CE;
380 		}
381 
382 		if (ue_count != ctx->ras_counter_ue) {
383 			ctx->ras_counter_ue = ue_count;
384 			out->state.flags |= AMDGPU_CTX_QUERY2_FLAGS_RAS_UE;
385 		}
386 
387 		schedule_delayed_work(&con->ras_counte_delay_work,
388 				      msecs_to_jiffies(AMDGPU_RAS_COUNTE_DELAY_MS));
389 	}
390 
391 	mutex_unlock(&mgr->lock);
392 	return 0;
393 }
394 
395 int amdgpu_ctx_ioctl(struct drm_device *dev, void *data,
396 		     struct drm_file *filp)
397 {
398 	int r;
399 	uint32_t id;
400 	enum drm_sched_priority priority;
401 
402 	union drm_amdgpu_ctx *args = data;
403 	struct amdgpu_device *adev = drm_to_adev(dev);
404 	struct amdgpu_fpriv *fpriv = filp->driver_priv;
405 
406 	id = args->in.ctx_id;
407 	r = amdgpu_to_sched_priority(args->in.priority, &priority);
408 
409 	/* For backwards compatibility reasons, we need to accept
410 	 * ioctls with garbage in the priority field */
411 	if (r == -EINVAL)
412 		priority = DRM_SCHED_PRIORITY_NORMAL;
413 
414 	switch (args->in.op) {
415 	case AMDGPU_CTX_OP_ALLOC_CTX:
416 		r = amdgpu_ctx_alloc(adev, fpriv, filp, priority, &id);
417 		args->out.alloc.ctx_id = id;
418 		break;
419 	case AMDGPU_CTX_OP_FREE_CTX:
420 		r = amdgpu_ctx_free(fpriv, id);
421 		break;
422 	case AMDGPU_CTX_OP_QUERY_STATE:
423 		r = amdgpu_ctx_query(adev, fpriv, id, &args->out);
424 		break;
425 	case AMDGPU_CTX_OP_QUERY_STATE2:
426 		r = amdgpu_ctx_query2(adev, fpriv, id, &args->out);
427 		break;
428 	default:
429 		return -EINVAL;
430 	}
431 
432 	return r;
433 }
434 
435 struct amdgpu_ctx *amdgpu_ctx_get(struct amdgpu_fpriv *fpriv, uint32_t id)
436 {
437 	struct amdgpu_ctx *ctx;
438 	struct amdgpu_ctx_mgr *mgr;
439 
440 	if (!fpriv)
441 		return NULL;
442 
443 	mgr = &fpriv->ctx_mgr;
444 
445 	mutex_lock(&mgr->lock);
446 	ctx = idr_find(&mgr->ctx_handles, id);
447 	if (ctx)
448 		kref_get(&ctx->refcount);
449 	mutex_unlock(&mgr->lock);
450 	return ctx;
451 }
452 
453 int amdgpu_ctx_put(struct amdgpu_ctx *ctx)
454 {
455 	if (ctx == NULL)
456 		return -EINVAL;
457 
458 	kref_put(&ctx->refcount, amdgpu_ctx_do_release);
459 	return 0;
460 }
461 
462 void amdgpu_ctx_add_fence(struct amdgpu_ctx *ctx,
463 			  struct drm_sched_entity *entity,
464 			  struct dma_fence *fence, uint64_t *handle)
465 {
466 	struct amdgpu_ctx_entity *centity = to_amdgpu_ctx_entity(entity);
467 	uint64_t seq = centity->sequence;
468 	struct dma_fence *other = NULL;
469 	unsigned idx = 0;
470 
471 	idx = seq & (amdgpu_sched_jobs - 1);
472 	other = centity->fences[idx];
473 	if (other)
474 		BUG_ON(!dma_fence_is_signaled(other));
475 
476 	dma_fence_get(fence);
477 
478 	spin_lock(&ctx->ring_lock);
479 	centity->fences[idx] = fence;
480 	centity->sequence++;
481 	spin_unlock(&ctx->ring_lock);
482 
483 	dma_fence_put(other);
484 	if (handle)
485 		*handle = seq;
486 }
487 
488 struct dma_fence *amdgpu_ctx_get_fence(struct amdgpu_ctx *ctx,
489 				       struct drm_sched_entity *entity,
490 				       uint64_t seq)
491 {
492 	struct amdgpu_ctx_entity *centity = to_amdgpu_ctx_entity(entity);
493 	struct dma_fence *fence;
494 
495 	spin_lock(&ctx->ring_lock);
496 
497 	if (seq == ~0ull)
498 		seq = centity->sequence - 1;
499 
500 	if (seq >= centity->sequence) {
501 		spin_unlock(&ctx->ring_lock);
502 		return ERR_PTR(-EINVAL);
503 	}
504 
505 
506 	if (seq + amdgpu_sched_jobs < centity->sequence) {
507 		spin_unlock(&ctx->ring_lock);
508 		return NULL;
509 	}
510 
511 	fence = dma_fence_get(centity->fences[seq & (amdgpu_sched_jobs - 1)]);
512 	spin_unlock(&ctx->ring_lock);
513 
514 	return fence;
515 }
516 
517 static void amdgpu_ctx_set_entity_priority(struct amdgpu_ctx *ctx,
518 					    struct amdgpu_ctx_entity *aentity,
519 					    int hw_ip,
520 					    enum drm_sched_priority priority)
521 {
522 	struct amdgpu_device *adev = ctx->adev;
523 	unsigned int hw_prio;
524 	struct drm_gpu_scheduler **scheds = NULL;
525 	unsigned num_scheds;
526 
527 	/* set sw priority */
528 	drm_sched_entity_set_priority(&aentity->entity, priority);
529 
530 	/* set hw priority */
531 	if (hw_ip == AMDGPU_HW_IP_COMPUTE) {
532 		hw_prio = amdgpu_ctx_prio_sched_to_hw(adev, priority,
533 						      AMDGPU_HW_IP_COMPUTE);
534 		hw_prio = array_index_nospec(hw_prio, AMDGPU_RING_PRIO_MAX);
535 		scheds = adev->gpu_sched[hw_ip][hw_prio].sched;
536 		num_scheds = adev->gpu_sched[hw_ip][hw_prio].num_scheds;
537 		drm_sched_entity_modify_sched(&aentity->entity, scheds,
538 					      num_scheds);
539 	}
540 }
541 
542 void amdgpu_ctx_priority_override(struct amdgpu_ctx *ctx,
543 				  enum drm_sched_priority priority)
544 {
545 	enum drm_sched_priority ctx_prio;
546 	unsigned i, j;
547 
548 	ctx->override_priority = priority;
549 
550 	ctx_prio = (ctx->override_priority == DRM_SCHED_PRIORITY_UNSET) ?
551 			ctx->init_priority : ctx->override_priority;
552 	for (i = 0; i < AMDGPU_HW_IP_NUM; ++i) {
553 		for (j = 0; j < amdgpu_ctx_num_entities[i]; ++j) {
554 			if (!ctx->entities[i][j])
555 				continue;
556 
557 			amdgpu_ctx_set_entity_priority(ctx, ctx->entities[i][j],
558 						       i, ctx_prio);
559 		}
560 	}
561 }
562 
563 int amdgpu_ctx_wait_prev_fence(struct amdgpu_ctx *ctx,
564 			       struct drm_sched_entity *entity)
565 {
566 	struct amdgpu_ctx_entity *centity = to_amdgpu_ctx_entity(entity);
567 	struct dma_fence *other;
568 	unsigned idx;
569 	long r;
570 
571 	spin_lock(&ctx->ring_lock);
572 	idx = centity->sequence & (amdgpu_sched_jobs - 1);
573 	other = dma_fence_get(centity->fences[idx]);
574 	spin_unlock(&ctx->ring_lock);
575 
576 	if (!other)
577 		return 0;
578 
579 	r = dma_fence_wait(other, true);
580 	if (r < 0 && r != -ERESTARTSYS)
581 		DRM_ERROR("Error (%ld) waiting for fence!\n", r);
582 
583 	dma_fence_put(other);
584 	return r;
585 }
586 
587 void amdgpu_ctx_mgr_init(struct amdgpu_ctx_mgr *mgr)
588 {
589 	mutex_init(&mgr->lock);
590 	idr_init(&mgr->ctx_handles);
591 }
592 
593 long amdgpu_ctx_mgr_entity_flush(struct amdgpu_ctx_mgr *mgr, long timeout)
594 {
595 	struct amdgpu_ctx *ctx;
596 	struct idr *idp;
597 	uint32_t id, i, j;
598 
599 	idp = &mgr->ctx_handles;
600 
601 	mutex_lock(&mgr->lock);
602 	idr_for_each_entry(idp, ctx, id) {
603 		for (i = 0; i < AMDGPU_HW_IP_NUM; ++i) {
604 			for (j = 0; j < amdgpu_ctx_num_entities[i]; ++j) {
605 				struct drm_sched_entity *entity;
606 
607 				if (!ctx->entities[i][j])
608 					continue;
609 
610 				entity = &ctx->entities[i][j]->entity;
611 				timeout = drm_sched_entity_flush(entity, timeout);
612 			}
613 		}
614 	}
615 	mutex_unlock(&mgr->lock);
616 	return timeout;
617 }
618 
619 void amdgpu_ctx_mgr_entity_fini(struct amdgpu_ctx_mgr *mgr)
620 {
621 	struct amdgpu_ctx *ctx;
622 	struct idr *idp;
623 	uint32_t id, i, j;
624 
625 	idp = &mgr->ctx_handles;
626 
627 	idr_for_each_entry(idp, ctx, id) {
628 		if (kref_read(&ctx->refcount) != 1) {
629 			DRM_ERROR("ctx %p is still alive\n", ctx);
630 			continue;
631 		}
632 
633 		for (i = 0; i < AMDGPU_HW_IP_NUM; ++i) {
634 			for (j = 0; j < amdgpu_ctx_num_entities[i]; ++j) {
635 				struct drm_sched_entity *entity;
636 
637 				if (!ctx->entities[i][j])
638 					continue;
639 
640 				entity = &ctx->entities[i][j]->entity;
641 				drm_sched_entity_fini(entity);
642 			}
643 		}
644 	}
645 }
646 
647 void amdgpu_ctx_mgr_fini(struct amdgpu_ctx_mgr *mgr)
648 {
649 	struct amdgpu_ctx *ctx;
650 	struct idr *idp;
651 	uint32_t id;
652 
653 	amdgpu_ctx_mgr_entity_fini(mgr);
654 
655 	idp = &mgr->ctx_handles;
656 
657 	idr_for_each_entry(idp, ctx, id) {
658 		if (kref_put(&ctx->refcount, amdgpu_ctx_fini) != 1)
659 			DRM_ERROR("ctx %p is still alive\n", ctx);
660 	}
661 
662 	idr_destroy(&mgr->ctx_handles);
663 	mutex_destroy(&mgr->lock);
664 }
665 
666 static void amdgpu_ctx_fence_time(struct amdgpu_ctx *ctx,
667 		struct amdgpu_ctx_entity *centity, ktime_t *total, ktime_t *max)
668 {
669 	ktime_t now, t1;
670 	uint32_t i;
671 
672 	*total = *max = 0;
673 
674 	now = ktime_get();
675 	for (i = 0; i < amdgpu_sched_jobs; i++) {
676 		struct dma_fence *fence;
677 		struct drm_sched_fence *s_fence;
678 
679 		spin_lock(&ctx->ring_lock);
680 		fence = dma_fence_get(centity->fences[i]);
681 		spin_unlock(&ctx->ring_lock);
682 		if (!fence)
683 			continue;
684 		s_fence = to_drm_sched_fence(fence);
685 		if (!dma_fence_is_signaled(&s_fence->scheduled)) {
686 			dma_fence_put(fence);
687 			continue;
688 		}
689 		t1 = s_fence->scheduled.timestamp;
690 		if (!ktime_before(t1, now)) {
691 			dma_fence_put(fence);
692 			continue;
693 		}
694 		if (dma_fence_is_signaled(&s_fence->finished) &&
695 			s_fence->finished.timestamp < now)
696 			*total += ktime_sub(s_fence->finished.timestamp, t1);
697 		else
698 			*total += ktime_sub(now, t1);
699 		t1 = ktime_sub(now, t1);
700 		dma_fence_put(fence);
701 		*max = max(t1, *max);
702 	}
703 }
704 
705 ktime_t amdgpu_ctx_mgr_fence_usage(struct amdgpu_ctx_mgr *mgr, uint32_t hwip,
706 		uint32_t idx, uint64_t *elapsed)
707 {
708 	struct idr *idp;
709 	struct amdgpu_ctx *ctx;
710 	uint32_t id;
711 	struct amdgpu_ctx_entity *centity;
712 	ktime_t total = 0, max = 0;
713 
714 	if (idx >= AMDGPU_MAX_ENTITY_NUM)
715 		return 0;
716 	idp = &mgr->ctx_handles;
717 	mutex_lock(&mgr->lock);
718 	idr_for_each_entry(idp, ctx, id) {
719 		ktime_t ttotal, tmax;
720 
721 		if (!ctx->entities[hwip][idx])
722 			continue;
723 
724 		centity = ctx->entities[hwip][idx];
725 		amdgpu_ctx_fence_time(ctx, centity, &ttotal, &tmax);
726 
727 		/* Harmonic mean approximation diverges for very small
728 		 * values. If ratio < 0.01% ignore
729 		 */
730 		if (AMDGPU_CTX_FENCE_USAGE_MIN_RATIO(tmax, ttotal))
731 			continue;
732 
733 		total = ktime_add(total, ttotal);
734 		max = ktime_after(tmax, max) ? tmax : max;
735 	}
736 
737 	mutex_unlock(&mgr->lock);
738 	if (elapsed)
739 		*elapsed = max;
740 
741 	return total;
742 }
743