sched_entity.c revision 1.8
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 */ 23 24#include <linux/kthread.h> 25#include <linux/slab.h> 26#include <linux/completion.h> 27 28#include <drm/drm_print.h> 29#include <drm/gpu_scheduler.h> 30 31#include "gpu_scheduler_trace.h" 32 33#define to_drm_sched_job(sched_job) \ 34 container_of((sched_job), struct drm_sched_job, queue_node) 35 36/** 37 * drm_sched_entity_init - Init a context entity used by scheduler when 38 * submit to HW ring. 39 * 40 * @entity: scheduler entity to init 41 * @priority: priority of the entity 42 * @sched_list: the list of drm scheds on which jobs from this 43 * entity can be submitted 44 * @num_sched_list: number of drm sched in sched_list 45 * @guilty: atomic_t set to 1 when a job on this queue 46 * is found to be guilty causing a timeout 47 * 48 * Note that the &sched_list must have at least one element to schedule the entity. 49 * 50 * For changing @priority later on at runtime see 51 * drm_sched_entity_set_priority(). For changing the set of schedulers 52 * @sched_list at runtime see drm_sched_entity_modify_sched(). 53 * 54 * An entity is cleaned up by callind drm_sched_entity_fini(). See also 55 * drm_sched_entity_destroy(). 56 * 57 * Returns 0 on success or a negative error code on failure. 58 */ 59int drm_sched_entity_init(struct drm_sched_entity *entity, 60 enum drm_sched_priority priority, 61 struct drm_gpu_scheduler **sched_list, 62 unsigned int num_sched_list, 63 atomic_t *guilty) 64{ 65 if (!(entity && sched_list && (num_sched_list == 0 || sched_list[0]))) 66 return -EINVAL; 67 68 memset(entity, 0, sizeof(struct drm_sched_entity)); 69 INIT_LIST_HEAD(&entity->list); 70 entity->rq = NULL; 71 entity->guilty = guilty; 72 entity->num_sched_list = num_sched_list; 73 entity->priority = priority; 74 entity->sched_list = num_sched_list > 1 ? sched_list : NULL; 75 RCU_INIT_POINTER(entity->last_scheduled, NULL); 76 RB_CLEAR_NODE(&entity->rb_tree_node); 77 78 if(num_sched_list) 79 entity->rq = &sched_list[0]->sched_rq[entity->priority]; 80 81 init_completion(&entity->entity_idle); 82 83 /* We start in an idle state. */ 84 complete_all(&entity->entity_idle); 85 86 mtx_init(&entity->rq_lock, IPL_NONE); 87 spsc_queue_init(&entity->job_queue); 88 89 atomic_set(&entity->fence_seq, 0); 90 entity->fence_context = dma_fence_context_alloc(2); 91 92 return 0; 93} 94EXPORT_SYMBOL(drm_sched_entity_init); 95 96/** 97 * drm_sched_entity_modify_sched - Modify sched of an entity 98 * @entity: scheduler entity to init 99 * @sched_list: the list of new drm scheds which will replace 100 * existing entity->sched_list 101 * @num_sched_list: number of drm sched in sched_list 102 * 103 * Note that this must be called under the same common lock for @entity as 104 * drm_sched_job_arm() and drm_sched_entity_push_job(), or the driver needs to 105 * guarantee through some other means that this is never called while new jobs 106 * can be pushed to @entity. 107 */ 108void drm_sched_entity_modify_sched(struct drm_sched_entity *entity, 109 struct drm_gpu_scheduler **sched_list, 110 unsigned int num_sched_list) 111{ 112 WARN_ON(!num_sched_list || !sched_list); 113 114 entity->sched_list = sched_list; 115 entity->num_sched_list = num_sched_list; 116} 117EXPORT_SYMBOL(drm_sched_entity_modify_sched); 118 119static bool drm_sched_entity_is_idle(struct drm_sched_entity *entity) 120{ 121 rmb(); /* for list_empty to work without lock */ 122 123 if (list_empty(&entity->list) || 124 spsc_queue_count(&entity->job_queue) == 0 || 125 entity->stopped) 126 return true; 127 128 return false; 129} 130 131/* Return true if entity could provide a job. */ 132bool drm_sched_entity_is_ready(struct drm_sched_entity *entity) 133{ 134 if (spsc_queue_peek(&entity->job_queue) == NULL) 135 return false; 136 137 if (READ_ONCE(entity->dependency)) 138 return false; 139 140 return true; 141} 142 143/** 144 * drm_sched_entity_error - return error of last scheduled job 145 * @entity: scheduler entity to check 146 * 147 * Opportunistically return the error of the last scheduled job. Result can 148 * change any time when new jobs are pushed to the hw. 149 */ 150int drm_sched_entity_error(struct drm_sched_entity *entity) 151{ 152 struct dma_fence *fence; 153 int r; 154 155 rcu_read_lock(); 156 fence = rcu_dereference(entity->last_scheduled); 157 r = fence ? fence->error : 0; 158 rcu_read_unlock(); 159 160 return r; 161} 162EXPORT_SYMBOL(drm_sched_entity_error); 163 164static void drm_sched_entity_kill_jobs_work(struct work_struct *wrk) 165{ 166 struct drm_sched_job *job = container_of(wrk, typeof(*job), work); 167 168 drm_sched_fence_finished(job->s_fence, -ESRCH); 169 WARN_ON(job->s_fence->parent); 170 job->sched->ops->free_job(job); 171} 172 173/* Signal the scheduler finished fence when the entity in question is killed. */ 174static void drm_sched_entity_kill_jobs_cb(struct dma_fence *f, 175 struct dma_fence_cb *cb) 176{ 177 struct drm_sched_job *job = container_of(cb, struct drm_sched_job, 178 finish_cb); 179 unsigned long index; 180 181 dma_fence_put(f); 182 183 /* Wait for all dependencies to avoid data corruptions */ 184 xa_for_each(&job->dependencies, index, f) { 185 struct drm_sched_fence *s_fence = to_drm_sched_fence(f); 186 187 if (s_fence && f == &s_fence->scheduled) { 188 /* The dependencies array had a reference on the scheduled 189 * fence, and the finished fence refcount might have 190 * dropped to zero. Use dma_fence_get_rcu() so we get 191 * a NULL fence in that case. 192 */ 193 f = dma_fence_get_rcu(&s_fence->finished); 194 195 /* Now that we have a reference on the finished fence, 196 * we can release the reference the dependencies array 197 * had on the scheduled fence. 198 */ 199 dma_fence_put(&s_fence->scheduled); 200 } 201 202 xa_erase(&job->dependencies, index); 203 if (f && !dma_fence_add_callback(f, &job->finish_cb, 204 drm_sched_entity_kill_jobs_cb)) 205 return; 206 207 dma_fence_put(f); 208 } 209 210 INIT_WORK(&job->work, drm_sched_entity_kill_jobs_work); 211 schedule_work(&job->work); 212} 213 214/* Remove the entity from the scheduler and kill all pending jobs */ 215static void drm_sched_entity_kill(struct drm_sched_entity *entity) 216{ 217 struct drm_sched_job *job; 218 struct dma_fence *prev; 219 220 if (!entity->rq) 221 return; 222 223 spin_lock(&entity->rq_lock); 224 entity->stopped = true; 225 drm_sched_rq_remove_entity(entity->rq, entity); 226 spin_unlock(&entity->rq_lock); 227 228 /* Make sure this entity is not used by the scheduler at the moment */ 229 wait_for_completion(&entity->entity_idle); 230 231 /* The entity is guaranteed to not be used by the scheduler */ 232 prev = rcu_dereference_check(entity->last_scheduled, true); 233 dma_fence_get(prev); 234 while ((job = to_drm_sched_job(spsc_queue_pop(&entity->job_queue)))) { 235 struct drm_sched_fence *s_fence = job->s_fence; 236 237 dma_fence_get(&s_fence->finished); 238 if (!prev || dma_fence_add_callback(prev, &job->finish_cb, 239 drm_sched_entity_kill_jobs_cb)) 240 drm_sched_entity_kill_jobs_cb(NULL, &job->finish_cb); 241 242 prev = &s_fence->finished; 243 } 244 dma_fence_put(prev); 245} 246 247/** 248 * drm_sched_entity_flush - Flush a context entity 249 * 250 * @entity: scheduler entity 251 * @timeout: time to wait in for Q to become empty in jiffies. 252 * 253 * Splitting drm_sched_entity_fini() into two functions, The first one does the 254 * waiting, removes the entity from the runqueue and returns an error when the 255 * process was killed. 256 * 257 * Returns the remaining time in jiffies left from the input timeout 258 */ 259long drm_sched_entity_flush(struct drm_sched_entity *entity, long timeout) 260{ 261 struct drm_gpu_scheduler *sched; 262#ifdef __linux__ 263 struct task_struct *last_user; 264#else 265 struct process *last_user, *curpr; 266#endif 267 long ret = timeout; 268 269 if (!entity->rq) 270 return 0; 271 272 sched = entity->rq->sched; 273 /** 274 * The client will not queue more IBs during this fini, consume existing 275 * queued IBs or discard them on SIGKILL 276 */ 277#ifdef __linux__ 278 if (current->flags & PF_EXITING) { 279#else 280 curpr = curproc->p_p; 281 if (curpr->ps_flags & PS_EXITING) { 282#endif 283 if (timeout) 284 ret = wait_event_timeout( 285 sched->job_scheduled, 286 drm_sched_entity_is_idle(entity), 287 timeout); 288 } else { 289 wait_event_killable(sched->job_scheduled, 290 drm_sched_entity_is_idle(entity)); 291 } 292 293 /* For killed process disable any more IBs enqueue right now */ 294#ifdef __linux__ 295 last_user = cmpxchg(&entity->last_user, current->group_leader, NULL); 296 if ((!last_user || last_user == current->group_leader) && 297 (current->flags & PF_EXITING) && (current->exit_code == SIGKILL)) 298#else 299 last_user = cmpxchg(&entity->last_user, curpr, NULL); 300 if ((!last_user || last_user == curproc->p_p) && 301 (curpr->ps_flags & PS_EXITING) && 302 (curpr->ps_xsig == SIGKILL)) 303#endif 304 drm_sched_entity_kill(entity); 305 306 return ret; 307} 308EXPORT_SYMBOL(drm_sched_entity_flush); 309 310/** 311 * drm_sched_entity_fini - Destroy a context entity 312 * 313 * @entity: scheduler entity 314 * 315 * Cleanups up @entity which has been initialized by drm_sched_entity_init(). 316 * 317 * If there are potentially job still in flight or getting newly queued 318 * drm_sched_entity_flush() must be called first. This function then goes over 319 * the entity and signals all jobs with an error code if the process was killed. 320 */ 321void drm_sched_entity_fini(struct drm_sched_entity *entity) 322{ 323 /* 324 * If consumption of existing IBs wasn't completed. Forcefully remove 325 * them here. Also makes sure that the scheduler won't touch this entity 326 * any more. 327 */ 328 drm_sched_entity_kill(entity); 329 330 if (entity->dependency) { 331 dma_fence_remove_callback(entity->dependency, &entity->cb); 332 dma_fence_put(entity->dependency); 333 entity->dependency = NULL; 334 } 335 336 dma_fence_put(rcu_dereference_check(entity->last_scheduled, true)); 337 RCU_INIT_POINTER(entity->last_scheduled, NULL); 338} 339EXPORT_SYMBOL(drm_sched_entity_fini); 340 341/** 342 * drm_sched_entity_destroy - Destroy a context entity 343 * @entity: scheduler entity 344 * 345 * Calls drm_sched_entity_flush() and drm_sched_entity_fini() as a 346 * convenience wrapper. 347 */ 348void drm_sched_entity_destroy(struct drm_sched_entity *entity) 349{ 350 drm_sched_entity_flush(entity, MAX_WAIT_SCHED_ENTITY_Q_EMPTY); 351 drm_sched_entity_fini(entity); 352} 353EXPORT_SYMBOL(drm_sched_entity_destroy); 354 355/* drm_sched_entity_clear_dep - callback to clear the entities dependency */ 356static void drm_sched_entity_clear_dep(struct dma_fence *f, 357 struct dma_fence_cb *cb) 358{ 359 struct drm_sched_entity *entity = 360 container_of(cb, struct drm_sched_entity, cb); 361 362 entity->dependency = NULL; 363 dma_fence_put(f); 364} 365 366/* 367 * drm_sched_entity_clear_dep - callback to clear the entities dependency and 368 * wake up scheduler 369 */ 370static void drm_sched_entity_wakeup(struct dma_fence *f, 371 struct dma_fence_cb *cb) 372{ 373 struct drm_sched_entity *entity = 374 container_of(cb, struct drm_sched_entity, cb); 375 376 drm_sched_entity_clear_dep(f, cb); 377 drm_sched_wakeup_if_can_queue(entity->rq->sched); 378} 379 380/** 381 * drm_sched_entity_set_priority - Sets priority of the entity 382 * 383 * @entity: scheduler entity 384 * @priority: scheduler priority 385 * 386 * Update the priority of runqueus used for the entity. 387 */ 388void drm_sched_entity_set_priority(struct drm_sched_entity *entity, 389 enum drm_sched_priority priority) 390{ 391 spin_lock(&entity->rq_lock); 392 entity->priority = priority; 393 spin_unlock(&entity->rq_lock); 394} 395EXPORT_SYMBOL(drm_sched_entity_set_priority); 396 397/* 398 * Add a callback to the current dependency of the entity to wake up the 399 * scheduler when the entity becomes available. 400 */ 401static bool drm_sched_entity_add_dependency_cb(struct drm_sched_entity *entity) 402{ 403 struct drm_gpu_scheduler *sched = entity->rq->sched; 404 struct dma_fence *fence = entity->dependency; 405 struct drm_sched_fence *s_fence; 406 407 if (fence->context == entity->fence_context || 408 fence->context == entity->fence_context + 1) { 409 /* 410 * Fence is a scheduled/finished fence from a job 411 * which belongs to the same entity, we can ignore 412 * fences from ourself 413 */ 414 dma_fence_put(entity->dependency); 415 return false; 416 } 417 418 s_fence = to_drm_sched_fence(fence); 419 if (!fence->error && s_fence && s_fence->sched == sched && 420 !test_bit(DRM_SCHED_FENCE_DONT_PIPELINE, &fence->flags)) { 421 422 /* 423 * Fence is from the same scheduler, only need to wait for 424 * it to be scheduled 425 */ 426 fence = dma_fence_get(&s_fence->scheduled); 427 dma_fence_put(entity->dependency); 428 entity->dependency = fence; 429 if (!dma_fence_add_callback(fence, &entity->cb, 430 drm_sched_entity_clear_dep)) 431 return true; 432 433 /* Ignore it when it is already scheduled */ 434 dma_fence_put(fence); 435 return false; 436 } 437 438 if (!dma_fence_add_callback(entity->dependency, &entity->cb, 439 drm_sched_entity_wakeup)) 440 return true; 441 442 dma_fence_put(entity->dependency); 443 return false; 444} 445 446static struct dma_fence * 447drm_sched_job_dependency(struct drm_sched_job *job, 448 struct drm_sched_entity *entity) 449{ 450 struct dma_fence *f; 451 452 /* We keep the fence around, so we can iterate over all dependencies 453 * in drm_sched_entity_kill_jobs_cb() to ensure all deps are signaled 454 * before killing the job. 455 */ 456 f = xa_load(&job->dependencies, job->last_dependency); 457 if (f) { 458 job->last_dependency++; 459 return dma_fence_get(f); 460 } 461 462 if (job->sched->ops->prepare_job) 463 return job->sched->ops->prepare_job(job, entity); 464 465 return NULL; 466} 467 468struct drm_sched_job *drm_sched_entity_pop_job(struct drm_sched_entity *entity) 469{ 470 struct drm_sched_job *sched_job; 471 472 sched_job = to_drm_sched_job(spsc_queue_peek(&entity->job_queue)); 473 if (!sched_job) 474 return NULL; 475 476 while ((entity->dependency = 477 drm_sched_job_dependency(sched_job, entity))) { 478 trace_drm_sched_job_wait_dep(sched_job, entity->dependency); 479 480 if (drm_sched_entity_add_dependency_cb(entity)) 481 return NULL; 482 } 483 484 /* skip jobs from entity that marked guilty */ 485 if (entity->guilty && atomic_read(entity->guilty)) 486 dma_fence_set_error(&sched_job->s_fence->finished, -ECANCELED); 487 488 dma_fence_put(rcu_dereference_check(entity->last_scheduled, true)); 489 rcu_assign_pointer(entity->last_scheduled, 490 dma_fence_get(&sched_job->s_fence->finished)); 491 492 /* 493 * If the queue is empty we allow drm_sched_entity_select_rq() to 494 * locklessly access ->last_scheduled. This only works if we set the 495 * pointer before we dequeue and if we a write barrier here. 496 */ 497 smp_wmb(); 498 499 spsc_queue_pop(&entity->job_queue); 500 501 /* 502 * Update the entity's location in the min heap according to 503 * the timestamp of the next job, if any. 504 */ 505 if (drm_sched_policy == DRM_SCHED_POLICY_FIFO) { 506 struct drm_sched_job *next; 507 508 next = to_drm_sched_job(spsc_queue_peek(&entity->job_queue)); 509 if (next) 510 drm_sched_rq_update_fifo(entity, next->submit_ts); 511 } 512 513 /* Jobs and entities might have different lifecycles. Since we're 514 * removing the job from the entities queue, set the jobs entity pointer 515 * to NULL to prevent any future access of the entity through this job. 516 */ 517 sched_job->entity = NULL; 518 519 return sched_job; 520} 521 522void drm_sched_entity_select_rq(struct drm_sched_entity *entity) 523{ 524 struct dma_fence *fence; 525 struct drm_gpu_scheduler *sched; 526 struct drm_sched_rq *rq; 527 528 /* single possible engine and already selected */ 529 if (!entity->sched_list) 530 return; 531 532 /* queue non-empty, stay on the same engine */ 533 if (spsc_queue_count(&entity->job_queue)) 534 return; 535 536 /* 537 * Only when the queue is empty are we guaranteed that the scheduler 538 * thread cannot change ->last_scheduled. To enforce ordering we need 539 * a read barrier here. See drm_sched_entity_pop_job() for the other 540 * side. 541 */ 542 smp_rmb(); 543 544 fence = rcu_dereference_check(entity->last_scheduled, true); 545 546 /* stay on the same engine if the previous job hasn't finished */ 547 if (fence && !dma_fence_is_signaled(fence)) 548 return; 549 550 spin_lock(&entity->rq_lock); 551 sched = drm_sched_pick_best(entity->sched_list, entity->num_sched_list); 552 rq = sched ? &sched->sched_rq[entity->priority] : NULL; 553 if (rq != entity->rq) { 554 drm_sched_rq_remove_entity(entity->rq, entity); 555 entity->rq = rq; 556 } 557 spin_unlock(&entity->rq_lock); 558 559 if (entity->num_sched_list == 1) 560 entity->sched_list = NULL; 561} 562 563/** 564 * drm_sched_entity_push_job - Submit a job to the entity's job queue 565 * @sched_job: job to submit 566 * 567 * Note: To guarantee that the order of insertion to queue matches the job's 568 * fence sequence number this function should be called with drm_sched_job_arm() 569 * under common lock for the struct drm_sched_entity that was set up for 570 * @sched_job in drm_sched_job_init(). 571 * 572 * Returns 0 for success, negative error code otherwise. 573 */ 574void drm_sched_entity_push_job(struct drm_sched_job *sched_job) 575{ 576 struct drm_sched_entity *entity = sched_job->entity; 577 bool first; 578 ktime_t submit_ts; 579 580 trace_drm_sched_job(sched_job, entity); 581 atomic_inc(entity->rq->sched->score); 582#ifdef __linux__ 583 WRITE_ONCE(entity->last_user, current->group_leader); 584#else 585 WRITE_ONCE(entity->last_user, curproc->p_p); 586#endif 587 588 /* 589 * After the sched_job is pushed into the entity queue, it may be 590 * completed and freed up at any time. We can no longer access it. 591 * Make sure to set the submit_ts first, to avoid a race. 592 */ 593 sched_job->submit_ts = submit_ts = ktime_get(); 594 first = spsc_queue_push(&entity->job_queue, &sched_job->queue_node); 595 596 /* first job wakes up scheduler */ 597 if (first) { 598 /* Add the entity to the run queue */ 599 spin_lock(&entity->rq_lock); 600 if (entity->stopped) { 601 spin_unlock(&entity->rq_lock); 602 603 DRM_ERROR("Trying to push to a killed entity\n"); 604 return; 605 } 606 607 drm_sched_rq_add_entity(entity->rq, entity); 608 spin_unlock(&entity->rq_lock); 609 610 if (drm_sched_policy == DRM_SCHED_POLICY_FIFO) 611 drm_sched_rq_update_fifo(entity, submit_ts); 612 613 drm_sched_wakeup_if_can_queue(entity->rq->sched); 614 } 615} 616EXPORT_SYMBOL(drm_sched_entity_push_job); 617