xref: /openbmc/linux/drivers/gpu/drm/amd/amdgpu/amdgpu_ctx.c (revision d43c76c8208c1c5e83bcf79c3f08974b231188f3)
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 <drm/drm_drv.h>
27 #include "amdgpu.h"
28 #include "amdgpu_sched.h"
29 #include "amdgpu_ras.h"
30 #include <linux/nospec.h>
31 
32 #define to_amdgpu_ctx_entity(e)	\
33 	container_of((e), struct amdgpu_ctx_entity, entity)
34 
35 const unsigned int amdgpu_ctx_num_entities[AMDGPU_HW_IP_NUM] = {
36 	[AMDGPU_HW_IP_GFX]	=	1,
37 	[AMDGPU_HW_IP_COMPUTE]	=	4,
38 	[AMDGPU_HW_IP_DMA]	=	2,
39 	[AMDGPU_HW_IP_UVD]	=	1,
40 	[AMDGPU_HW_IP_VCE]	=	1,
41 	[AMDGPU_HW_IP_UVD_ENC]	=	1,
42 	[AMDGPU_HW_IP_VCN_DEC]	=	1,
43 	[AMDGPU_HW_IP_VCN_ENC]	=	1,
44 	[AMDGPU_HW_IP_VCN_JPEG]	=	1,
45 };
46 
47 bool amdgpu_ctx_priority_is_valid(int32_t ctx_prio)
48 {
49 	switch (ctx_prio) {
50 	case AMDGPU_CTX_PRIORITY_VERY_LOW:
51 	case AMDGPU_CTX_PRIORITY_LOW:
52 	case AMDGPU_CTX_PRIORITY_NORMAL:
53 	case AMDGPU_CTX_PRIORITY_HIGH:
54 	case AMDGPU_CTX_PRIORITY_VERY_HIGH:
55 		return true;
56 	default:
57 	case AMDGPU_CTX_PRIORITY_UNSET:
58 		return false;
59 	}
60 }
61 
62 static enum drm_sched_priority
63 amdgpu_ctx_to_drm_sched_prio(int32_t ctx_prio)
64 {
65 	switch (ctx_prio) {
66 	case AMDGPU_CTX_PRIORITY_UNSET:
67 		pr_warn_once("AMD-->DRM context priority value UNSET-->NORMAL");
68 		return DRM_SCHED_PRIORITY_NORMAL;
69 
70 	case AMDGPU_CTX_PRIORITY_VERY_LOW:
71 		return DRM_SCHED_PRIORITY_MIN;
72 
73 	case AMDGPU_CTX_PRIORITY_LOW:
74 		return DRM_SCHED_PRIORITY_MIN;
75 
76 	case AMDGPU_CTX_PRIORITY_NORMAL:
77 		return DRM_SCHED_PRIORITY_NORMAL;
78 
79 	case AMDGPU_CTX_PRIORITY_HIGH:
80 		return DRM_SCHED_PRIORITY_HIGH;
81 
82 	case AMDGPU_CTX_PRIORITY_VERY_HIGH:
83 		return DRM_SCHED_PRIORITY_HIGH;
84 
85 	/* This should not happen as we sanitized userspace provided priority
86 	 * already, WARN if this happens.
87 	 */
88 	default:
89 		WARN(1, "Invalid context priority %d\n", ctx_prio);
90 		return DRM_SCHED_PRIORITY_NORMAL;
91 	}
92 
93 }
94 
95 static int amdgpu_ctx_priority_permit(struct drm_file *filp,
96 				      int32_t priority)
97 {
98 	if (!amdgpu_ctx_priority_is_valid(priority))
99 		return -EINVAL;
100 
101 	/* NORMAL and below are accessible by everyone */
102 	if (priority <= AMDGPU_CTX_PRIORITY_NORMAL)
103 		return 0;
104 
105 	if (capable(CAP_SYS_NICE))
106 		return 0;
107 
108 	if (drm_is_current_master(filp))
109 		return 0;
110 
111 	return -EACCES;
112 }
113 
114 static enum amdgpu_gfx_pipe_priority amdgpu_ctx_prio_to_gfx_pipe_prio(int32_t prio)
115 {
116 	switch (prio) {
117 	case AMDGPU_CTX_PRIORITY_HIGH:
118 	case AMDGPU_CTX_PRIORITY_VERY_HIGH:
119 		return AMDGPU_GFX_PIPE_PRIO_HIGH;
120 	default:
121 		return AMDGPU_GFX_PIPE_PRIO_NORMAL;
122 	}
123 }
124 
125 static enum amdgpu_ring_priority_level amdgpu_ctx_sched_prio_to_ring_prio(int32_t prio)
126 {
127 	switch (prio) {
128 	case AMDGPU_CTX_PRIORITY_HIGH:
129 		return AMDGPU_RING_PRIO_1;
130 	case AMDGPU_CTX_PRIORITY_VERY_HIGH:
131 		return AMDGPU_RING_PRIO_2;
132 	default:
133 		return AMDGPU_RING_PRIO_0;
134 	}
135 }
136 
137 static unsigned int amdgpu_ctx_get_hw_prio(struct amdgpu_ctx *ctx, u32 hw_ip)
138 {
139 	struct amdgpu_device *adev = ctx->mgr->adev;
140 	unsigned int hw_prio;
141 	int32_t ctx_prio;
142 
143 	ctx_prio = (ctx->override_priority == AMDGPU_CTX_PRIORITY_UNSET) ?
144 			ctx->init_priority : ctx->override_priority;
145 
146 	switch (hw_ip) {
147 	case AMDGPU_HW_IP_GFX:
148 	case AMDGPU_HW_IP_COMPUTE:
149 		hw_prio = amdgpu_ctx_prio_to_gfx_pipe_prio(ctx_prio);
150 		break;
151 	case AMDGPU_HW_IP_VCE:
152 	case AMDGPU_HW_IP_VCN_ENC:
153 		hw_prio = amdgpu_ctx_sched_prio_to_ring_prio(ctx_prio);
154 		break;
155 	default:
156 		hw_prio = AMDGPU_RING_PRIO_DEFAULT;
157 		break;
158 	}
159 
160 	hw_ip = array_index_nospec(hw_ip, AMDGPU_HW_IP_NUM);
161 	if (adev->gpu_sched[hw_ip][hw_prio].num_scheds == 0)
162 		hw_prio = AMDGPU_RING_PRIO_DEFAULT;
163 
164 	return hw_prio;
165 }
166 
167 /* Calculate the time spend on the hw */
168 static ktime_t amdgpu_ctx_fence_time(struct dma_fence *fence)
169 {
170 	struct drm_sched_fence *s_fence;
171 
172 	if (!fence)
173 		return ns_to_ktime(0);
174 
175 	/* When the fence is not even scheduled it can't have spend time */
176 	s_fence = to_drm_sched_fence(fence);
177 	if (!test_bit(DMA_FENCE_FLAG_TIMESTAMP_BIT, &s_fence->scheduled.flags))
178 		return ns_to_ktime(0);
179 
180 	/* When it is still running account how much already spend */
181 	if (!test_bit(DMA_FENCE_FLAG_TIMESTAMP_BIT, &s_fence->finished.flags))
182 		return ktime_sub(ktime_get(), s_fence->scheduled.timestamp);
183 
184 	return ktime_sub(s_fence->finished.timestamp,
185 			 s_fence->scheduled.timestamp);
186 }
187 
188 static ktime_t amdgpu_ctx_entity_time(struct amdgpu_ctx *ctx,
189 				      struct amdgpu_ctx_entity *centity)
190 {
191 	ktime_t res = ns_to_ktime(0);
192 	uint32_t i;
193 
194 	spin_lock(&ctx->ring_lock);
195 	for (i = 0; i < amdgpu_sched_jobs; i++) {
196 		res = ktime_add(res, amdgpu_ctx_fence_time(centity->fences[i]));
197 	}
198 	spin_unlock(&ctx->ring_lock);
199 	return res;
200 }
201 
202 static int amdgpu_ctx_init_entity(struct amdgpu_ctx *ctx, u32 hw_ip,
203 				  const u32 ring)
204 {
205 	struct drm_gpu_scheduler **scheds = NULL, *sched = NULL;
206 	struct amdgpu_device *adev = ctx->mgr->adev;
207 	struct amdgpu_ctx_entity *entity;
208 	enum drm_sched_priority drm_prio;
209 	unsigned int hw_prio, num_scheds;
210 	int32_t ctx_prio;
211 	int r;
212 
213 	entity = kzalloc(struct_size(entity, fences, amdgpu_sched_jobs),
214 			 GFP_KERNEL);
215 	if (!entity)
216 		return  -ENOMEM;
217 
218 	ctx_prio = (ctx->override_priority == AMDGPU_CTX_PRIORITY_UNSET) ?
219 			ctx->init_priority : ctx->override_priority;
220 	entity->hw_ip = hw_ip;
221 	entity->sequence = 1;
222 	hw_prio = amdgpu_ctx_get_hw_prio(ctx, hw_ip);
223 	drm_prio = amdgpu_ctx_to_drm_sched_prio(ctx_prio);
224 
225 	hw_ip = array_index_nospec(hw_ip, AMDGPU_HW_IP_NUM);
226 
227 	if (!(adev)->xcp_mgr) {
228 		scheds = adev->gpu_sched[hw_ip][hw_prio].sched;
229 		num_scheds = adev->gpu_sched[hw_ip][hw_prio].num_scheds;
230 	} else {
231 		struct amdgpu_fpriv *fpriv;
232 
233 		fpriv = container_of(ctx->ctx_mgr, struct amdgpu_fpriv, ctx_mgr);
234 		r = amdgpu_xcp_select_scheds(adev, hw_ip, hw_prio, fpriv,
235 						&num_scheds, &scheds);
236 		if (r)
237 			goto cleanup_entity;
238 	}
239 
240 	/* disable load balance if the hw engine retains context among dependent jobs */
241 	if (hw_ip == AMDGPU_HW_IP_VCN_ENC ||
242 	    hw_ip == AMDGPU_HW_IP_VCN_DEC ||
243 	    hw_ip == AMDGPU_HW_IP_UVD_ENC ||
244 	    hw_ip == AMDGPU_HW_IP_UVD) {
245 		sched = drm_sched_pick_best(scheds, num_scheds);
246 		scheds = &sched;
247 		num_scheds = 1;
248 	}
249 
250 	r = drm_sched_entity_init(&entity->entity, drm_prio, scheds, num_scheds,
251 				  &ctx->guilty);
252 	if (r)
253 		goto error_free_entity;
254 
255 	/* It's not an error if we fail to install the new entity */
256 	if (cmpxchg(&ctx->entities[hw_ip][ring], NULL, entity))
257 		goto cleanup_entity;
258 
259 	return 0;
260 
261 cleanup_entity:
262 	drm_sched_entity_fini(&entity->entity);
263 
264 error_free_entity:
265 	kfree(entity);
266 
267 	return r;
268 }
269 
270 static ktime_t amdgpu_ctx_fini_entity(struct amdgpu_device *adev,
271 				  struct amdgpu_ctx_entity *entity)
272 {
273 	ktime_t res = ns_to_ktime(0);
274 	int i;
275 
276 	if (!entity)
277 		return res;
278 
279 	for (i = 0; i < amdgpu_sched_jobs; ++i) {
280 		res = ktime_add(res, amdgpu_ctx_fence_time(entity->fences[i]));
281 		dma_fence_put(entity->fences[i]);
282 	}
283 
284 	amdgpu_xcp_release_sched(adev, entity);
285 
286 	kfree(entity);
287 	return res;
288 }
289 
290 static int amdgpu_ctx_get_stable_pstate(struct amdgpu_ctx *ctx,
291 					u32 *stable_pstate)
292 {
293 	struct amdgpu_device *adev = ctx->mgr->adev;
294 	enum amd_dpm_forced_level current_level;
295 
296 	current_level = amdgpu_dpm_get_performance_level(adev);
297 
298 	switch (current_level) {
299 	case AMD_DPM_FORCED_LEVEL_PROFILE_STANDARD:
300 		*stable_pstate = AMDGPU_CTX_STABLE_PSTATE_STANDARD;
301 		break;
302 	case AMD_DPM_FORCED_LEVEL_PROFILE_MIN_SCLK:
303 		*stable_pstate = AMDGPU_CTX_STABLE_PSTATE_MIN_SCLK;
304 		break;
305 	case AMD_DPM_FORCED_LEVEL_PROFILE_MIN_MCLK:
306 		*stable_pstate = AMDGPU_CTX_STABLE_PSTATE_MIN_MCLK;
307 		break;
308 	case AMD_DPM_FORCED_LEVEL_PROFILE_PEAK:
309 		*stable_pstate = AMDGPU_CTX_STABLE_PSTATE_PEAK;
310 		break;
311 	default:
312 		*stable_pstate = AMDGPU_CTX_STABLE_PSTATE_NONE;
313 		break;
314 	}
315 	return 0;
316 }
317 
318 static int amdgpu_ctx_init(struct amdgpu_ctx_mgr *mgr, int32_t priority,
319 			   struct drm_file *filp, struct amdgpu_ctx *ctx)
320 {
321 	struct amdgpu_fpriv *fpriv = filp->driver_priv;
322 	u32 current_stable_pstate;
323 	int r;
324 
325 	r = amdgpu_ctx_priority_permit(filp, priority);
326 	if (r)
327 		return r;
328 
329 	memset(ctx, 0, sizeof(*ctx));
330 
331 	kref_init(&ctx->refcount);
332 	ctx->mgr = mgr;
333 	spin_lock_init(&ctx->ring_lock);
334 
335 	ctx->reset_counter = atomic_read(&mgr->adev->gpu_reset_counter);
336 	ctx->reset_counter_query = ctx->reset_counter;
337 	ctx->generation = amdgpu_vm_generation(mgr->adev, &fpriv->vm);
338 	ctx->init_priority = priority;
339 	ctx->override_priority = AMDGPU_CTX_PRIORITY_UNSET;
340 
341 	r = amdgpu_ctx_get_stable_pstate(ctx, &current_stable_pstate);
342 	if (r)
343 		return r;
344 
345 	if (mgr->adev->pm.stable_pstate_ctx)
346 		ctx->stable_pstate = mgr->adev->pm.stable_pstate_ctx->stable_pstate;
347 	else
348 		ctx->stable_pstate = current_stable_pstate;
349 
350 	ctx->ctx_mgr = &(fpriv->ctx_mgr);
351 	return 0;
352 }
353 
354 static int amdgpu_ctx_set_stable_pstate(struct amdgpu_ctx *ctx,
355 					u32 stable_pstate)
356 {
357 	struct amdgpu_device *adev = ctx->mgr->adev;
358 	enum amd_dpm_forced_level level;
359 	u32 current_stable_pstate;
360 	int r;
361 
362 	mutex_lock(&adev->pm.stable_pstate_ctx_lock);
363 	if (adev->pm.stable_pstate_ctx && adev->pm.stable_pstate_ctx != ctx) {
364 		r = -EBUSY;
365 		goto done;
366 	}
367 
368 	r = amdgpu_ctx_get_stable_pstate(ctx, &current_stable_pstate);
369 	if (r || (stable_pstate == current_stable_pstate))
370 		goto done;
371 
372 	switch (stable_pstate) {
373 	case AMDGPU_CTX_STABLE_PSTATE_NONE:
374 		level = AMD_DPM_FORCED_LEVEL_AUTO;
375 		break;
376 	case AMDGPU_CTX_STABLE_PSTATE_STANDARD:
377 		level = AMD_DPM_FORCED_LEVEL_PROFILE_STANDARD;
378 		break;
379 	case AMDGPU_CTX_STABLE_PSTATE_MIN_SCLK:
380 		level = AMD_DPM_FORCED_LEVEL_PROFILE_MIN_SCLK;
381 		break;
382 	case AMDGPU_CTX_STABLE_PSTATE_MIN_MCLK:
383 		level = AMD_DPM_FORCED_LEVEL_PROFILE_MIN_MCLK;
384 		break;
385 	case AMDGPU_CTX_STABLE_PSTATE_PEAK:
386 		level = AMD_DPM_FORCED_LEVEL_PROFILE_PEAK;
387 		break;
388 	default:
389 		r = -EINVAL;
390 		goto done;
391 	}
392 
393 	r = amdgpu_dpm_force_performance_level(adev, level);
394 
395 	if (level == AMD_DPM_FORCED_LEVEL_AUTO)
396 		adev->pm.stable_pstate_ctx = NULL;
397 	else
398 		adev->pm.stable_pstate_ctx = ctx;
399 done:
400 	mutex_unlock(&adev->pm.stable_pstate_ctx_lock);
401 
402 	return r;
403 }
404 
405 static void amdgpu_ctx_fini(struct kref *ref)
406 {
407 	struct amdgpu_ctx *ctx = container_of(ref, struct amdgpu_ctx, refcount);
408 	struct amdgpu_ctx_mgr *mgr = ctx->mgr;
409 	struct amdgpu_device *adev = mgr->adev;
410 	unsigned i, j, idx;
411 
412 	if (!adev)
413 		return;
414 
415 	for (i = 0; i < AMDGPU_HW_IP_NUM; ++i) {
416 		for (j = 0; j < AMDGPU_MAX_ENTITY_NUM; ++j) {
417 			ktime_t spend;
418 
419 			spend = amdgpu_ctx_fini_entity(adev, ctx->entities[i][j]);
420 			atomic64_add(ktime_to_ns(spend), &mgr->time_spend[i]);
421 		}
422 	}
423 
424 	if (drm_dev_enter(adev_to_drm(adev), &idx)) {
425 		amdgpu_ctx_set_stable_pstate(ctx, ctx->stable_pstate);
426 		drm_dev_exit(idx);
427 	}
428 
429 	kfree(ctx);
430 }
431 
432 int amdgpu_ctx_get_entity(struct amdgpu_ctx *ctx, u32 hw_ip, u32 instance,
433 			  u32 ring, struct drm_sched_entity **entity)
434 {
435 	int r;
436 	struct drm_sched_entity *ctx_entity;
437 
438 	if (hw_ip >= AMDGPU_HW_IP_NUM) {
439 		DRM_ERROR("unknown HW IP type: %d\n", hw_ip);
440 		return -EINVAL;
441 	}
442 
443 	/* Right now all IPs have only one instance - multiple rings. */
444 	if (instance != 0) {
445 		DRM_DEBUG("invalid ip instance: %d\n", instance);
446 		return -EINVAL;
447 	}
448 
449 	if (ring >= amdgpu_ctx_num_entities[hw_ip]) {
450 		DRM_DEBUG("invalid ring: %d %d\n", hw_ip, ring);
451 		return -EINVAL;
452 	}
453 
454 	if (ctx->entities[hw_ip][ring] == NULL) {
455 		r = amdgpu_ctx_init_entity(ctx, hw_ip, ring);
456 		if (r)
457 			return r;
458 	}
459 
460 	ctx_entity = &ctx->entities[hw_ip][ring]->entity;
461 	r = drm_sched_entity_error(ctx_entity);
462 	if (r) {
463 		DRM_DEBUG("error entity %p\n", ctx_entity);
464 		return r;
465 	}
466 
467 	*entity = ctx_entity;
468 	return 0;
469 }
470 
471 static int amdgpu_ctx_alloc(struct amdgpu_device *adev,
472 			    struct amdgpu_fpriv *fpriv,
473 			    struct drm_file *filp,
474 			    int32_t priority,
475 			    uint32_t *id)
476 {
477 	struct amdgpu_ctx_mgr *mgr = &fpriv->ctx_mgr;
478 	struct amdgpu_ctx *ctx;
479 	int r;
480 
481 	ctx = kmalloc(sizeof(*ctx), GFP_KERNEL);
482 	if (!ctx)
483 		return -ENOMEM;
484 
485 	mutex_lock(&mgr->lock);
486 	r = idr_alloc(&mgr->ctx_handles, ctx, 1, AMDGPU_VM_MAX_NUM_CTX, GFP_KERNEL);
487 	if (r < 0) {
488 		mutex_unlock(&mgr->lock);
489 		kfree(ctx);
490 		return r;
491 	}
492 
493 	*id = (uint32_t)r;
494 	r = amdgpu_ctx_init(mgr, priority, filp, ctx);
495 	if (r) {
496 		idr_remove(&mgr->ctx_handles, *id);
497 		*id = 0;
498 		kfree(ctx);
499 	}
500 	mutex_unlock(&mgr->lock);
501 	return r;
502 }
503 
504 static void amdgpu_ctx_do_release(struct kref *ref)
505 {
506 	struct amdgpu_ctx *ctx;
507 	u32 i, j;
508 
509 	ctx = container_of(ref, struct amdgpu_ctx, refcount);
510 	for (i = 0; i < AMDGPU_HW_IP_NUM; ++i) {
511 		for (j = 0; j < amdgpu_ctx_num_entities[i]; ++j) {
512 			if (!ctx->entities[i][j])
513 				continue;
514 
515 			drm_sched_entity_destroy(&ctx->entities[i][j]->entity);
516 		}
517 	}
518 
519 	amdgpu_ctx_fini(ref);
520 }
521 
522 static int amdgpu_ctx_free(struct amdgpu_fpriv *fpriv, uint32_t id)
523 {
524 	struct amdgpu_ctx_mgr *mgr = &fpriv->ctx_mgr;
525 	struct amdgpu_ctx *ctx;
526 
527 	mutex_lock(&mgr->lock);
528 	ctx = idr_remove(&mgr->ctx_handles, id);
529 	if (ctx)
530 		kref_put(&ctx->refcount, amdgpu_ctx_do_release);
531 	mutex_unlock(&mgr->lock);
532 	return ctx ? 0 : -EINVAL;
533 }
534 
535 static int amdgpu_ctx_query(struct amdgpu_device *adev,
536 			    struct amdgpu_fpriv *fpriv, uint32_t id,
537 			    union drm_amdgpu_ctx_out *out)
538 {
539 	struct amdgpu_ctx *ctx;
540 	struct amdgpu_ctx_mgr *mgr;
541 	unsigned reset_counter;
542 
543 	if (!fpriv)
544 		return -EINVAL;
545 
546 	mgr = &fpriv->ctx_mgr;
547 	mutex_lock(&mgr->lock);
548 	ctx = idr_find(&mgr->ctx_handles, id);
549 	if (!ctx) {
550 		mutex_unlock(&mgr->lock);
551 		return -EINVAL;
552 	}
553 
554 	/* TODO: these two are always zero */
555 	out->state.flags = 0x0;
556 	out->state.hangs = 0x0;
557 
558 	/* determine if a GPU reset has occured since the last call */
559 	reset_counter = atomic_read(&adev->gpu_reset_counter);
560 	/* TODO: this should ideally return NO, GUILTY, or INNOCENT. */
561 	if (ctx->reset_counter_query == reset_counter)
562 		out->state.reset_status = AMDGPU_CTX_NO_RESET;
563 	else
564 		out->state.reset_status = AMDGPU_CTX_UNKNOWN_RESET;
565 	ctx->reset_counter_query = reset_counter;
566 
567 	mutex_unlock(&mgr->lock);
568 	return 0;
569 }
570 
571 #define AMDGPU_RAS_COUNTE_DELAY_MS 3000
572 
573 static int amdgpu_ctx_query2(struct amdgpu_device *adev,
574 			     struct amdgpu_fpriv *fpriv, uint32_t id,
575 			     union drm_amdgpu_ctx_out *out)
576 {
577 	struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
578 	struct amdgpu_ctx *ctx;
579 	struct amdgpu_ctx_mgr *mgr;
580 
581 	if (!fpriv)
582 		return -EINVAL;
583 
584 	mgr = &fpriv->ctx_mgr;
585 	mutex_lock(&mgr->lock);
586 	ctx = idr_find(&mgr->ctx_handles, id);
587 	if (!ctx) {
588 		mutex_unlock(&mgr->lock);
589 		return -EINVAL;
590 	}
591 
592 	out->state.flags = 0x0;
593 	out->state.hangs = 0x0;
594 
595 	if (ctx->reset_counter != atomic_read(&adev->gpu_reset_counter))
596 		out->state.flags |= AMDGPU_CTX_QUERY2_FLAGS_RESET;
597 
598 	if (ctx->generation != amdgpu_vm_generation(adev, &fpriv->vm))
599 		out->state.flags |= AMDGPU_CTX_QUERY2_FLAGS_VRAMLOST;
600 
601 	if (atomic_read(&ctx->guilty))
602 		out->state.flags |= AMDGPU_CTX_QUERY2_FLAGS_GUILTY;
603 
604 	if (amdgpu_in_reset(adev))
605 		out->state.flags |= AMDGPU_CTX_QUERY2_FLAGS_RESET_IN_PROGRESS;
606 
607 	if (adev->ras_enabled && con) {
608 		/* Return the cached values in O(1),
609 		 * and schedule delayed work to cache
610 		 * new vaues.
611 		 */
612 		int ce_count, ue_count;
613 
614 		ce_count = atomic_read(&con->ras_ce_count);
615 		ue_count = atomic_read(&con->ras_ue_count);
616 
617 		if (ce_count != ctx->ras_counter_ce) {
618 			ctx->ras_counter_ce = ce_count;
619 			out->state.flags |= AMDGPU_CTX_QUERY2_FLAGS_RAS_CE;
620 		}
621 
622 		if (ue_count != ctx->ras_counter_ue) {
623 			ctx->ras_counter_ue = ue_count;
624 			out->state.flags |= AMDGPU_CTX_QUERY2_FLAGS_RAS_UE;
625 		}
626 
627 		schedule_delayed_work(&con->ras_counte_delay_work,
628 				      msecs_to_jiffies(AMDGPU_RAS_COUNTE_DELAY_MS));
629 	}
630 
631 	mutex_unlock(&mgr->lock);
632 	return 0;
633 }
634 
635 
636 
637 static int amdgpu_ctx_stable_pstate(struct amdgpu_device *adev,
638 				    struct amdgpu_fpriv *fpriv, uint32_t id,
639 				    bool set, u32 *stable_pstate)
640 {
641 	struct amdgpu_ctx *ctx;
642 	struct amdgpu_ctx_mgr *mgr;
643 	int r;
644 
645 	if (!fpriv)
646 		return -EINVAL;
647 
648 	mgr = &fpriv->ctx_mgr;
649 	mutex_lock(&mgr->lock);
650 	ctx = idr_find(&mgr->ctx_handles, id);
651 	if (!ctx) {
652 		mutex_unlock(&mgr->lock);
653 		return -EINVAL;
654 	}
655 
656 	if (set)
657 		r = amdgpu_ctx_set_stable_pstate(ctx, *stable_pstate);
658 	else
659 		r = amdgpu_ctx_get_stable_pstate(ctx, stable_pstate);
660 
661 	mutex_unlock(&mgr->lock);
662 	return r;
663 }
664 
665 int amdgpu_ctx_ioctl(struct drm_device *dev, void *data,
666 		     struct drm_file *filp)
667 {
668 	int r;
669 	uint32_t id, stable_pstate;
670 	int32_t priority;
671 
672 	union drm_amdgpu_ctx *args = data;
673 	struct amdgpu_device *adev = drm_to_adev(dev);
674 	struct amdgpu_fpriv *fpriv = filp->driver_priv;
675 
676 	id = args->in.ctx_id;
677 	priority = args->in.priority;
678 
679 	/* For backwards compatibility reasons, we need to accept
680 	 * ioctls with garbage in the priority field */
681 	if (!amdgpu_ctx_priority_is_valid(priority))
682 		priority = AMDGPU_CTX_PRIORITY_NORMAL;
683 
684 	switch (args->in.op) {
685 	case AMDGPU_CTX_OP_ALLOC_CTX:
686 		r = amdgpu_ctx_alloc(adev, fpriv, filp, priority, &id);
687 		args->out.alloc.ctx_id = id;
688 		break;
689 	case AMDGPU_CTX_OP_FREE_CTX:
690 		r = amdgpu_ctx_free(fpriv, id);
691 		break;
692 	case AMDGPU_CTX_OP_QUERY_STATE:
693 		r = amdgpu_ctx_query(adev, fpriv, id, &args->out);
694 		break;
695 	case AMDGPU_CTX_OP_QUERY_STATE2:
696 		r = amdgpu_ctx_query2(adev, fpriv, id, &args->out);
697 		break;
698 	case AMDGPU_CTX_OP_GET_STABLE_PSTATE:
699 		if (args->in.flags)
700 			return -EINVAL;
701 		r = amdgpu_ctx_stable_pstate(adev, fpriv, id, false, &stable_pstate);
702 		if (!r)
703 			args->out.pstate.flags = stable_pstate;
704 		break;
705 	case AMDGPU_CTX_OP_SET_STABLE_PSTATE:
706 		if (args->in.flags & ~AMDGPU_CTX_STABLE_PSTATE_FLAGS_MASK)
707 			return -EINVAL;
708 		stable_pstate = args->in.flags & AMDGPU_CTX_STABLE_PSTATE_FLAGS_MASK;
709 		if (stable_pstate > AMDGPU_CTX_STABLE_PSTATE_PEAK)
710 			return -EINVAL;
711 		r = amdgpu_ctx_stable_pstate(adev, fpriv, id, true, &stable_pstate);
712 		break;
713 	default:
714 		return -EINVAL;
715 	}
716 
717 	return r;
718 }
719 
720 struct amdgpu_ctx *amdgpu_ctx_get(struct amdgpu_fpriv *fpriv, uint32_t id)
721 {
722 	struct amdgpu_ctx *ctx;
723 	struct amdgpu_ctx_mgr *mgr;
724 
725 	if (!fpriv)
726 		return NULL;
727 
728 	mgr = &fpriv->ctx_mgr;
729 
730 	mutex_lock(&mgr->lock);
731 	ctx = idr_find(&mgr->ctx_handles, id);
732 	if (ctx)
733 		kref_get(&ctx->refcount);
734 	mutex_unlock(&mgr->lock);
735 	return ctx;
736 }
737 
738 int amdgpu_ctx_put(struct amdgpu_ctx *ctx)
739 {
740 	if (ctx == NULL)
741 		return -EINVAL;
742 
743 	kref_put(&ctx->refcount, amdgpu_ctx_do_release);
744 	return 0;
745 }
746 
747 uint64_t amdgpu_ctx_add_fence(struct amdgpu_ctx *ctx,
748 			      struct drm_sched_entity *entity,
749 			      struct dma_fence *fence)
750 {
751 	struct amdgpu_ctx_entity *centity = to_amdgpu_ctx_entity(entity);
752 	uint64_t seq = centity->sequence;
753 	struct dma_fence *other = NULL;
754 	unsigned idx = 0;
755 
756 	idx = seq & (amdgpu_sched_jobs - 1);
757 	other = centity->fences[idx];
758 	WARN_ON(other && !dma_fence_is_signaled(other));
759 
760 	dma_fence_get(fence);
761 
762 	spin_lock(&ctx->ring_lock);
763 	centity->fences[idx] = fence;
764 	centity->sequence++;
765 	spin_unlock(&ctx->ring_lock);
766 
767 	atomic64_add(ktime_to_ns(amdgpu_ctx_fence_time(other)),
768 		     &ctx->mgr->time_spend[centity->hw_ip]);
769 
770 	dma_fence_put(other);
771 	return seq;
772 }
773 
774 struct dma_fence *amdgpu_ctx_get_fence(struct amdgpu_ctx *ctx,
775 				       struct drm_sched_entity *entity,
776 				       uint64_t seq)
777 {
778 	struct amdgpu_ctx_entity *centity = to_amdgpu_ctx_entity(entity);
779 	struct dma_fence *fence;
780 
781 	spin_lock(&ctx->ring_lock);
782 
783 	if (seq == ~0ull)
784 		seq = centity->sequence - 1;
785 
786 	if (seq >= centity->sequence) {
787 		spin_unlock(&ctx->ring_lock);
788 		return ERR_PTR(-EINVAL);
789 	}
790 
791 
792 	if (seq + amdgpu_sched_jobs < centity->sequence) {
793 		spin_unlock(&ctx->ring_lock);
794 		return NULL;
795 	}
796 
797 	fence = dma_fence_get(centity->fences[seq & (amdgpu_sched_jobs - 1)]);
798 	spin_unlock(&ctx->ring_lock);
799 
800 	return fence;
801 }
802 
803 static void amdgpu_ctx_set_entity_priority(struct amdgpu_ctx *ctx,
804 					   struct amdgpu_ctx_entity *aentity,
805 					   int hw_ip,
806 					   int32_t priority)
807 {
808 	struct amdgpu_device *adev = ctx->mgr->adev;
809 	unsigned int hw_prio;
810 	struct drm_gpu_scheduler **scheds = NULL;
811 	unsigned num_scheds;
812 
813 	/* set sw priority */
814 	drm_sched_entity_set_priority(&aentity->entity,
815 				      amdgpu_ctx_to_drm_sched_prio(priority));
816 
817 	/* set hw priority */
818 	if (hw_ip == AMDGPU_HW_IP_COMPUTE || hw_ip == AMDGPU_HW_IP_GFX) {
819 		hw_prio = amdgpu_ctx_get_hw_prio(ctx, hw_ip);
820 		hw_prio = array_index_nospec(hw_prio, AMDGPU_RING_PRIO_MAX);
821 		scheds = adev->gpu_sched[hw_ip][hw_prio].sched;
822 		num_scheds = adev->gpu_sched[hw_ip][hw_prio].num_scheds;
823 		drm_sched_entity_modify_sched(&aentity->entity, scheds,
824 					      num_scheds);
825 	}
826 }
827 
828 void amdgpu_ctx_priority_override(struct amdgpu_ctx *ctx,
829 				  int32_t priority)
830 {
831 	int32_t ctx_prio;
832 	unsigned i, j;
833 
834 	ctx->override_priority = priority;
835 
836 	ctx_prio = (ctx->override_priority == AMDGPU_CTX_PRIORITY_UNSET) ?
837 			ctx->init_priority : ctx->override_priority;
838 	for (i = 0; i < AMDGPU_HW_IP_NUM; ++i) {
839 		for (j = 0; j < amdgpu_ctx_num_entities[i]; ++j) {
840 			if (!ctx->entities[i][j])
841 				continue;
842 
843 			amdgpu_ctx_set_entity_priority(ctx, ctx->entities[i][j],
844 						       i, ctx_prio);
845 		}
846 	}
847 }
848 
849 int amdgpu_ctx_wait_prev_fence(struct amdgpu_ctx *ctx,
850 			       struct drm_sched_entity *entity)
851 {
852 	struct amdgpu_ctx_entity *centity = to_amdgpu_ctx_entity(entity);
853 	struct dma_fence *other;
854 	unsigned idx;
855 	long r;
856 
857 	spin_lock(&ctx->ring_lock);
858 	idx = centity->sequence & (amdgpu_sched_jobs - 1);
859 	other = dma_fence_get(centity->fences[idx]);
860 	spin_unlock(&ctx->ring_lock);
861 
862 	if (!other)
863 		return 0;
864 
865 	r = dma_fence_wait(other, true);
866 	if (r < 0 && r != -ERESTARTSYS)
867 		DRM_ERROR("Error (%ld) waiting for fence!\n", r);
868 
869 	dma_fence_put(other);
870 	return r;
871 }
872 
873 void amdgpu_ctx_mgr_init(struct amdgpu_ctx_mgr *mgr,
874 			 struct amdgpu_device *adev)
875 {
876 	unsigned int i;
877 
878 	mgr->adev = adev;
879 	mutex_init(&mgr->lock);
880 	idr_init_base(&mgr->ctx_handles, 1);
881 
882 	for (i = 0; i < AMDGPU_HW_IP_NUM; ++i)
883 		atomic64_set(&mgr->time_spend[i], 0);
884 }
885 
886 long amdgpu_ctx_mgr_entity_flush(struct amdgpu_ctx_mgr *mgr, long timeout)
887 {
888 	struct amdgpu_ctx *ctx;
889 	struct idr *idp;
890 	uint32_t id, i, j;
891 
892 	idp = &mgr->ctx_handles;
893 
894 	mutex_lock(&mgr->lock);
895 	idr_for_each_entry(idp, ctx, id) {
896 		for (i = 0; i < AMDGPU_HW_IP_NUM; ++i) {
897 			for (j = 0; j < amdgpu_ctx_num_entities[i]; ++j) {
898 				struct drm_sched_entity *entity;
899 
900 				if (!ctx->entities[i][j])
901 					continue;
902 
903 				entity = &ctx->entities[i][j]->entity;
904 				timeout = drm_sched_entity_flush(entity, timeout);
905 			}
906 		}
907 	}
908 	mutex_unlock(&mgr->lock);
909 	return timeout;
910 }
911 
912 void amdgpu_ctx_mgr_entity_fini(struct amdgpu_ctx_mgr *mgr)
913 {
914 	struct amdgpu_ctx *ctx;
915 	struct idr *idp;
916 	uint32_t id, i, j;
917 
918 	idp = &mgr->ctx_handles;
919 
920 	idr_for_each_entry(idp, ctx, id) {
921 		if (kref_read(&ctx->refcount) != 1) {
922 			DRM_ERROR("ctx %p is still alive\n", ctx);
923 			continue;
924 		}
925 
926 		for (i = 0; i < AMDGPU_HW_IP_NUM; ++i) {
927 			for (j = 0; j < amdgpu_ctx_num_entities[i]; ++j) {
928 				struct drm_sched_entity *entity;
929 
930 				if (!ctx->entities[i][j])
931 					continue;
932 
933 				entity = &ctx->entities[i][j]->entity;
934 				drm_sched_entity_fini(entity);
935 			}
936 		}
937 	}
938 }
939 
940 void amdgpu_ctx_mgr_fini(struct amdgpu_ctx_mgr *mgr)
941 {
942 	struct amdgpu_ctx *ctx;
943 	struct idr *idp;
944 	uint32_t id;
945 
946 	amdgpu_ctx_mgr_entity_fini(mgr);
947 
948 	idp = &mgr->ctx_handles;
949 
950 	idr_for_each_entry(idp, ctx, id) {
951 		if (kref_put(&ctx->refcount, amdgpu_ctx_fini) != 1)
952 			DRM_ERROR("ctx %p is still alive\n", ctx);
953 	}
954 
955 	idr_destroy(&mgr->ctx_handles);
956 	mutex_destroy(&mgr->lock);
957 }
958 
959 void amdgpu_ctx_mgr_usage(struct amdgpu_ctx_mgr *mgr,
960 			  ktime_t usage[AMDGPU_HW_IP_NUM])
961 {
962 	struct amdgpu_ctx *ctx;
963 	unsigned int hw_ip, i;
964 	uint32_t id;
965 
966 	/*
967 	 * This is a little bit racy because it can be that a ctx or a fence are
968 	 * destroyed just in the moment we try to account them. But that is ok
969 	 * since exactly that case is explicitely allowed by the interface.
970 	 */
971 	mutex_lock(&mgr->lock);
972 	for (hw_ip = 0; hw_ip < AMDGPU_HW_IP_NUM; ++hw_ip) {
973 		uint64_t ns = atomic64_read(&mgr->time_spend[hw_ip]);
974 
975 		usage[hw_ip] = ns_to_ktime(ns);
976 	}
977 
978 	idr_for_each_entry(&mgr->ctx_handles, ctx, id) {
979 		for (hw_ip = 0; hw_ip < AMDGPU_HW_IP_NUM; ++hw_ip) {
980 			for (i = 0; i < amdgpu_ctx_num_entities[hw_ip]; ++i) {
981 				struct amdgpu_ctx_entity *centity;
982 				ktime_t spend;
983 
984 				centity = ctx->entities[hw_ip][i];
985 				if (!centity)
986 					continue;
987 				spend = amdgpu_ctx_entity_time(ctx, centity);
988 				usage[hw_ip] = ktime_add(usage[hw_ip], spend);
989 			}
990 		}
991 	}
992 	mutex_unlock(&mgr->lock);
993 }
994