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