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 "amdgpu.h"
27#include "amdgpu_sched.h"
28#include "amdgpu_ras.h"
29
30#define to_amdgpu_ctx_entity(e) \
31 container_of((e), struct amdgpu_ctx_entity, entity)
32
33const unsigned int amdgpu_ctx_num_entities[AMDGPU_HW_IP_NUM] = {
34 [AMDGPU_HW_IP_GFX] = 1,
35 [AMDGPU_HW_IP_COMPUTE] = 4,
36 [AMDGPU_HW_IP_DMA] = 2,
37 [AMDGPU_HW_IP_UVD] = 1,
38 [AMDGPU_HW_IP_VCE] = 1,
39 [AMDGPU_HW_IP_UVD_ENC] = 1,
40 [AMDGPU_HW_IP_VCN_DEC] = 1,
41 [AMDGPU_HW_IP_VCN_ENC] = 1,
42 [AMDGPU_HW_IP_VCN_JPEG] = 1,
43};
44
45static int amdgpu_ctx_total_num_entities(void)
46{
47 unsigned i, num_entities = 0;
48
49 for (i = 0; i < AMDGPU_HW_IP_NUM; ++i)
50 num_entities += amdgpu_ctx_num_entities[i];
51
52 return num_entities;
53}
54
55static int amdgpu_ctx_priority_permit(struct drm_file *filp,
56 enum drm_sched_priority priority)
57{
58 /* NORMAL and below are accessible by everyone */
59 if (priority <= DRM_SCHED_PRIORITY_NORMAL)
60 return 0;
61
62 if (capable(CAP_SYS_NICE))
63 return 0;
64
65 if (drm_is_current_master(filp))
66 return 0;
67
68 return -EACCES;
69}
70
71static int amdgpu_ctx_init(struct amdgpu_device *adev,
72 enum drm_sched_priority priority,
73 struct drm_file *filp,
74 struct amdgpu_ctx *ctx)
75{
76 unsigned num_entities = amdgpu_ctx_total_num_entities();
77 unsigned i, j, k;
78 int r;
79
80 if (priority < 0 || priority >= DRM_SCHED_PRIORITY_MAX)
81 return -EINVAL;
82
83 r = amdgpu_ctx_priority_permit(filp, priority);
84 if (r)
85 return r;
86
87 memset(ctx, 0, sizeof(*ctx));
88 ctx->adev = adev;
89
90 ctx->fences = kcalloc(amdgpu_sched_jobs * num_entities,
91 sizeof(struct dma_fence*), GFP_KERNEL);
92 if (!ctx->fences)
93 return -ENOMEM;
94
95 ctx->entities[0] = kcalloc(num_entities,
96 sizeof(struct amdgpu_ctx_entity),
97 GFP_KERNEL);
98 if (!ctx->entities[0]) {
99 r = -ENOMEM;
100 goto error_free_fences;
101 }
102
103 for (i = 0; i < num_entities; ++i) {
104 struct amdgpu_ctx_entity *entity = &ctx->entities[0][i];
105
106 entity->sequence = 1;
107 entity->fences = &ctx->fences[amdgpu_sched_jobs * i];
108 }
109 for (i = 1; i < AMDGPU_HW_IP_NUM; ++i)
110 ctx->entities[i] = ctx->entities[i - 1] +
111 amdgpu_ctx_num_entities[i - 1];
112
113 kref_init(&ctx->refcount);
114 spin_lock_init(&ctx->ring_lock);
115 mutex_init(&ctx->lock);
116
117 ctx->reset_counter = atomic_read(&adev->gpu_reset_counter);
118 ctx->reset_counter_query = ctx->reset_counter;
119 ctx->vram_lost_counter = atomic_read(&adev->vram_lost_counter);
120 ctx->init_priority = priority;
121 ctx->override_priority = DRM_SCHED_PRIORITY_UNSET;
122
123 for (i = 0; i < AMDGPU_HW_IP_NUM; ++i) {
124 struct amdgpu_ring *rings[AMDGPU_MAX_RINGS];
125 struct drm_sched_rq *rqs[AMDGPU_MAX_RINGS];
126 unsigned num_rings = 0;
127 unsigned num_rqs = 0;
128
129 switch (i) {
130 case AMDGPU_HW_IP_GFX:
131 rings[0] = &adev->gfx.gfx_ring[0];
132 num_rings = 1;
133 break;
134 case AMDGPU_HW_IP_COMPUTE:
135 for (j = 0; j < adev->gfx.num_compute_rings; ++j)
136 rings[j] = &adev->gfx.compute_ring[j];
137 num_rings = adev->gfx.num_compute_rings;
138 break;
139 case AMDGPU_HW_IP_DMA:
140 for (j = 0; j < adev->sdma.num_instances; ++j)
141 rings[j] = &adev->sdma.instance[j].ring;
142 num_rings = adev->sdma.num_instances;
143 break;
144 case AMDGPU_HW_IP_UVD:
145 rings[0] = &adev->uvd.inst[0].ring;
146 num_rings = 1;
147 break;
148 case AMDGPU_HW_IP_VCE:
149 rings[0] = &adev->vce.ring[0];
150 num_rings = 1;
151 break;
152 case AMDGPU_HW_IP_UVD_ENC:
153 rings[0] = &adev->uvd.inst[0].ring_enc[0];
154 num_rings = 1;
155 break;
156 case AMDGPU_HW_IP_VCN_DEC:
157 for (j = 0; j < adev->vcn.num_vcn_inst; ++j) {
158 if (adev->vcn.harvest_config & (1 << j))
159 continue;
160 rings[num_rings++] = &adev->vcn.inst[j].ring_dec;
161 }
162 break;
163 case AMDGPU_HW_IP_VCN_ENC:
164 for (j = 0; j < adev->vcn.num_vcn_inst; ++j) {
165 if (adev->vcn.harvest_config & (1 << j))
166 continue;
167 for (k = 0; k < adev->vcn.num_enc_rings; ++k)
168 rings[num_rings++] = &adev->vcn.inst[j].ring_enc[k];
169 }
170 break;
171 case AMDGPU_HW_IP_VCN_JPEG:
172 for (j = 0; j < adev->vcn.num_vcn_inst; ++j) {
173 if (adev->vcn.harvest_config & (1 << j))
174 continue;
175 rings[num_rings++] = &adev->vcn.inst[j].ring_jpeg;
176 }
177 break;
178 }
179
180 for (j = 0; j < num_rings; ++j) {
181 if (!rings[j]->adev)
182 continue;
183
184 rqs[num_rqs++] = &rings[j]->sched.sched_rq[priority];
185 }
186
187 for (j = 0; j < amdgpu_ctx_num_entities[i]; ++j)
188 r = drm_sched_entity_init(&ctx->entities[i][j].entity,
189 rqs, num_rqs, &ctx->guilty);
190 if (r)
191 goto error_cleanup_entities;
192 }
193
194 return 0;
195
196error_cleanup_entities:
197 for (i = 0; i < num_entities; ++i)
198 drm_sched_entity_destroy(&ctx->entities[0][i].entity);
199 kfree(ctx->entities[0]);
200
201error_free_fences:
202 kfree(ctx->fences);
203 ctx->fences = NULL;
204 return r;
205}
206
207static void amdgpu_ctx_fini(struct kref *ref)
208{
209 struct amdgpu_ctx *ctx = container_of(ref, struct amdgpu_ctx, refcount);
210 unsigned num_entities = amdgpu_ctx_total_num_entities();
211 struct amdgpu_device *adev = ctx->adev;
212 unsigned i, j;
213
214 if (!adev)
215 return;
216
217 for (i = 0; i < num_entities; ++i)
218 for (j = 0; j < amdgpu_sched_jobs; ++j)
219 dma_fence_put(ctx->entities[0][i].fences[j]);
220 kfree(ctx->fences);
221 kfree(ctx->entities[0]);
222
223 mutex_destroy(&ctx->lock);
224
225 kfree(ctx);
226}
227
228int amdgpu_ctx_get_entity(struct amdgpu_ctx *ctx, u32 hw_ip, u32 instance,
229 u32 ring, struct drm_sched_entity **entity)
230{
231 if (hw_ip >= AMDGPU_HW_IP_NUM) {
232 DRM_ERROR("unknown HW IP type: %d\n", hw_ip);
233 return -EINVAL;
234 }
235
236 /* Right now all IPs have only one instance - multiple rings. */
237 if (instance != 0) {
238 DRM_DEBUG("invalid ip instance: %d\n", instance);
239 return -EINVAL;
240 }
241
242 if (ring >= amdgpu_ctx_num_entities[hw_ip]) {
243 DRM_DEBUG("invalid ring: %d %d\n", hw_ip, ring);
244 return -EINVAL;
245 }
246
247 *entity = &ctx->entities[hw_ip][ring].entity;
248 return 0;
249}
250
251static int amdgpu_ctx_alloc(struct amdgpu_device *adev,
252 struct amdgpu_fpriv *fpriv,
253 struct drm_file *filp,
254 enum drm_sched_priority priority,
255 uint32_t *id)
256{
257 struct amdgpu_ctx_mgr *mgr = &fpriv->ctx_mgr;
258 struct amdgpu_ctx *ctx;
259 int r;
260
261 ctx = kmalloc(sizeof(*ctx), GFP_KERNEL);
262 if (!ctx)
263 return -ENOMEM;
264
265 mutex_lock(&mgr->lock);
266 r = idr_alloc(&mgr->ctx_handles, ctx, 1, AMDGPU_VM_MAX_NUM_CTX, GFP_KERNEL);
267 if (r < 0) {
268 mutex_unlock(&mgr->lock);
269 kfree(ctx);
270 return r;
271 }
272
273 *id = (uint32_t)r;
274 r = amdgpu_ctx_init(adev, priority, filp, ctx);
275 if (r) {
276 idr_remove(&mgr->ctx_handles, *id);
277 *id = 0;
278 kfree(ctx);
279 }
280 mutex_unlock(&mgr->lock);
281 return r;
282}
283
284static void amdgpu_ctx_do_release(struct kref *ref)
285{
286 struct amdgpu_ctx *ctx;
287 unsigned num_entities;
288 u32 i;
289
290 ctx = container_of(ref, struct amdgpu_ctx, refcount);
291
292 num_entities = amdgpu_ctx_total_num_entities();
293 for (i = 0; i < num_entities; i++)
294 drm_sched_entity_destroy(&ctx->entities[0][i].entity);
295
296 amdgpu_ctx_fini(ref);
297}
298
299static int amdgpu_ctx_free(struct amdgpu_fpriv *fpriv, uint32_t id)
300{
301 struct amdgpu_ctx_mgr *mgr = &fpriv->ctx_mgr;
302 struct amdgpu_ctx *ctx;
303
304 mutex_lock(&mgr->lock);
305 ctx = idr_remove(&mgr->ctx_handles, id);
306 if (ctx)
307 kref_put(&ctx->refcount, amdgpu_ctx_do_release);
308 mutex_unlock(&mgr->lock);
309 return ctx ? 0 : -EINVAL;
310}
311
312static int amdgpu_ctx_query(struct amdgpu_device *adev,
313 struct amdgpu_fpriv *fpriv, uint32_t id,
314 union drm_amdgpu_ctx_out *out)
315{
316 struct amdgpu_ctx *ctx;
317 struct amdgpu_ctx_mgr *mgr;
318 unsigned reset_counter;
319
320 if (!fpriv)
321 return -EINVAL;
322
323 mgr = &fpriv->ctx_mgr;
324 mutex_lock(&mgr->lock);
325 ctx = idr_find(&mgr->ctx_handles, id);
326 if (!ctx) {
327 mutex_unlock(&mgr->lock);
328 return -EINVAL;
329 }
330
331 /* TODO: these two are always zero */
332 out->state.flags = 0x0;
333 out->state.hangs = 0x0;
334
335 /* determine if a GPU reset has occured since the last call */
336 reset_counter = atomic_read(&adev->gpu_reset_counter);
337 /* TODO: this should ideally return NO, GUILTY, or INNOCENT. */
338 if (ctx->reset_counter_query == reset_counter)
339 out->state.reset_status = AMDGPU_CTX_NO_RESET;
340 else
341 out->state.reset_status = AMDGPU_CTX_UNKNOWN_RESET;
342 ctx->reset_counter_query = reset_counter;
343
344 mutex_unlock(&mgr->lock);
345 return 0;
346}
347
348static int amdgpu_ctx_query2(struct amdgpu_device *adev,
349 struct amdgpu_fpriv *fpriv, uint32_t id,
350 union drm_amdgpu_ctx_out *out)
351{
352 struct amdgpu_ctx *ctx;
353 struct amdgpu_ctx_mgr *mgr;
354 unsigned long ras_counter;
355
356 if (!fpriv)
357 return -EINVAL;
358
359 mgr = &fpriv->ctx_mgr;
360 mutex_lock(&mgr->lock);
361 ctx = idr_find(&mgr->ctx_handles, id);
362 if (!ctx) {
363 mutex_unlock(&mgr->lock);
364 return -EINVAL;
365 }
366
367 out->state.flags = 0x0;
368 out->state.hangs = 0x0;
369
370 if (ctx->reset_counter != atomic_read(&adev->gpu_reset_counter))
371 out->state.flags |= AMDGPU_CTX_QUERY2_FLAGS_RESET;
372
373 if (ctx->vram_lost_counter != atomic_read(&adev->vram_lost_counter))
374 out->state.flags |= AMDGPU_CTX_QUERY2_FLAGS_VRAMLOST;
375
376 if (atomic_read(&ctx->guilty))
377 out->state.flags |= AMDGPU_CTX_QUERY2_FLAGS_GUILTY;
378
379 /*query ue count*/
380 ras_counter = amdgpu_ras_query_error_count(adev, false);
381 /*ras counter is monotonic increasing*/
382 if (ras_counter != ctx->ras_counter_ue) {
383 out->state.flags |= AMDGPU_CTX_QUERY2_FLAGS_RAS_UE;
384 ctx->ras_counter_ue = ras_counter;
385 }
386
387 /*query ce count*/
388 ras_counter = amdgpu_ras_query_error_count(adev, true);
389 if (ras_counter != ctx->ras_counter_ce) {
390 out->state.flags |= AMDGPU_CTX_QUERY2_FLAGS_RAS_CE;
391 ctx->ras_counter_ce = ras_counter;
392 }
393
394 mutex_unlock(&mgr->lock);
395 return 0;
396}
397
398int amdgpu_ctx_ioctl(struct drm_device *dev, void *data,
399 struct drm_file *filp)
400{
401 int r;
402 uint32_t id;
403 enum drm_sched_priority priority;
404
405 union drm_amdgpu_ctx *args = data;
406 struct amdgpu_device *adev = dev->dev_private;
407 struct amdgpu_fpriv *fpriv = filp->driver_priv;
408
409 r = 0;
410 id = args->in.ctx_id;
411 priority = amdgpu_to_sched_priority(args->in.priority);
412
413 /* For backwards compatibility reasons, we need to accept
414 * ioctls with garbage in the priority field */
415 if (priority == DRM_SCHED_PRIORITY_INVALID)
416 priority = DRM_SCHED_PRIORITY_NORMAL;
417
418 switch (args->in.op) {
419 case AMDGPU_CTX_OP_ALLOC_CTX:
420 r = amdgpu_ctx_alloc(adev, fpriv, filp, priority, &id);
421 args->out.alloc.ctx_id = id;
422 break;
423 case AMDGPU_CTX_OP_FREE_CTX:
424 r = amdgpu_ctx_free(fpriv, id);
425 break;
426 case AMDGPU_CTX_OP_QUERY_STATE:
427 r = amdgpu_ctx_query(adev, fpriv, id, &args->out);
428 break;
429 case AMDGPU_CTX_OP_QUERY_STATE2:
430 r = amdgpu_ctx_query2(adev, fpriv, id, &args->out);
431 break;
432 default:
433 return -EINVAL;
434 }
435
436 return r;
437}
438
439struct amdgpu_ctx *amdgpu_ctx_get(struct amdgpu_fpriv *fpriv, uint32_t id)
440{
441 struct amdgpu_ctx *ctx;
442 struct amdgpu_ctx_mgr *mgr;
443
444 if (!fpriv)
445 return NULL;
446
447 mgr = &fpriv->ctx_mgr;
448
449 mutex_lock(&mgr->lock);
450 ctx = idr_find(&mgr->ctx_handles, id);
451 if (ctx)
452 kref_get(&ctx->refcount);
453 mutex_unlock(&mgr->lock);
454 return ctx;
455}
456
457int amdgpu_ctx_put(struct amdgpu_ctx *ctx)
458{
459 if (ctx == NULL)
460 return -EINVAL;
461
462 kref_put(&ctx->refcount, amdgpu_ctx_do_release);
463 return 0;
464}
465
466void amdgpu_ctx_add_fence(struct amdgpu_ctx *ctx,
467 struct drm_sched_entity *entity,
468 struct dma_fence *fence, uint64_t* handle)
469{
470 struct amdgpu_ctx_entity *centity = to_amdgpu_ctx_entity(entity);
471 uint64_t seq = centity->sequence;
472 struct dma_fence *other = NULL;
473 unsigned idx = 0;
474
475 idx = seq & (amdgpu_sched_jobs - 1);
476 other = centity->fences[idx];
477 if (other)
478 BUG_ON(!dma_fence_is_signaled(other));
479
480 dma_fence_get(fence);
481
482 spin_lock(&ctx->ring_lock);
483 centity->fences[idx] = fence;
484 centity->sequence++;
485 spin_unlock(&ctx->ring_lock);
486
487 dma_fence_put(other);
488 if (handle)
489 *handle = seq;
490}
491
492struct dma_fence *amdgpu_ctx_get_fence(struct amdgpu_ctx *ctx,
493 struct drm_sched_entity *entity,
494 uint64_t seq)
495{
496 struct amdgpu_ctx_entity *centity = to_amdgpu_ctx_entity(entity);
497 struct dma_fence *fence;
498
499 spin_lock(&ctx->ring_lock);
500
501 if (seq == ~0ull)
502 seq = centity->sequence - 1;
503
504 if (seq >= centity->sequence) {
505 spin_unlock(&ctx->ring_lock);
506 return ERR_PTR(-EINVAL);
507 }
508
509
510 if (seq + amdgpu_sched_jobs < centity->sequence) {
511 spin_unlock(&ctx->ring_lock);
512 return NULL;
513 }
514
515 fence = dma_fence_get(centity->fences[seq & (amdgpu_sched_jobs - 1)]);
516 spin_unlock(&ctx->ring_lock);
517
518 return fence;
519}
520
521void amdgpu_ctx_priority_override(struct amdgpu_ctx *ctx,
522 enum drm_sched_priority priority)
523{
524 unsigned num_entities = amdgpu_ctx_total_num_entities();
525 enum drm_sched_priority ctx_prio;
526 unsigned i;
527
528 ctx->override_priority = priority;
529
530 ctx_prio = (ctx->override_priority == DRM_SCHED_PRIORITY_UNSET) ?
531 ctx->init_priority : ctx->override_priority;
532
533 for (i = 0; i < num_entities; i++) {
534 struct drm_sched_entity *entity = &ctx->entities[0][i].entity;
535
536 drm_sched_entity_set_priority(entity, ctx_prio);
537 }
538}
539
540int amdgpu_ctx_wait_prev_fence(struct amdgpu_ctx *ctx,
541 struct drm_sched_entity *entity)
542{
543 struct amdgpu_ctx_entity *centity = to_amdgpu_ctx_entity(entity);
544 struct dma_fence *other;
545 unsigned idx;
546 long r;
547
548 spin_lock(&ctx->ring_lock);
549 idx = centity->sequence & (amdgpu_sched_jobs - 1);
550 other = dma_fence_get(centity->fences[idx]);
551 spin_unlock(&ctx->ring_lock);
552
553 if (!other)
554 return 0;
555
556 r = dma_fence_wait(other, true);
557 if (r < 0 && r != -ERESTARTSYS)
558 DRM_ERROR("Error (%ld) waiting for fence!\n", r);
559
560 dma_fence_put(other);
561 return r;
562}
563
564void amdgpu_ctx_mgr_init(struct amdgpu_ctx_mgr *mgr)
565{
566 mutex_init(&mgr->lock);
567 idr_init(&mgr->ctx_handles);
568}
569
570long amdgpu_ctx_mgr_entity_flush(struct amdgpu_ctx_mgr *mgr, long timeout)
571{
572 unsigned num_entities = amdgpu_ctx_total_num_entities();
573 struct amdgpu_ctx *ctx;
574 struct idr *idp;
575 uint32_t id, i;
576
577 idp = &mgr->ctx_handles;
578
579 mutex_lock(&mgr->lock);
580 idr_for_each_entry(idp, ctx, id) {
581 for (i = 0; i < num_entities; i++) {
582 struct drm_sched_entity *entity;
583
584 entity = &ctx->entities[0][i].entity;
585 timeout = drm_sched_entity_flush(entity, timeout);
586 }
587 }
588 mutex_unlock(&mgr->lock);
589 return timeout;
590}
591
592void amdgpu_ctx_mgr_entity_fini(struct amdgpu_ctx_mgr *mgr)
593{
594 unsigned num_entities = amdgpu_ctx_total_num_entities();
595 struct amdgpu_ctx *ctx;
596 struct idr *idp;
597 uint32_t id, i;
598
599 idp = &mgr->ctx_handles;
600
601 idr_for_each_entry(idp, ctx, id) {
602 if (kref_read(&ctx->refcount) != 1) {
603 DRM_ERROR("ctx %p is still alive\n", ctx);
604 continue;
605 }
606
607 for (i = 0; i < num_entities; i++) {
608 mutex_lock(&ctx->adev->lock_reset);
609 drm_sched_entity_fini(&ctx->entities[0][i].entity);
610 mutex_unlock(&ctx->adev->lock_reset);
611 }
612 }
613}
614
615void amdgpu_ctx_mgr_fini(struct amdgpu_ctx_mgr *mgr)
616{
617 struct amdgpu_ctx *ctx;
618 struct idr *idp;
619 uint32_t id;
620
621 amdgpu_ctx_mgr_entity_fini(mgr);
622
623 idp = &mgr->ctx_handles;
624
625 idr_for_each_entry(idp, ctx, id) {
626 if (kref_put(&ctx->refcount, amdgpu_ctx_fini) != 1)
627 DRM_ERROR("ctx %p is still alive\n", ctx);
628 }
629
630 idr_destroy(&mgr->ctx_handles);
631 mutex_destroy(&mgr->lock);
632}
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}