drivers/gpu/drm/scheduler/sched_entity.c

   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  */
  59 int 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         spin_lock_init(&entity->rq_lock);
  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 }
  94 EXPORT_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  */
 108 void 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 }
 117 EXPORT_SYMBOL(drm_sched_entity_modify_sched);
 118
 119 static 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. */
 132 bool 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  */
 150 int 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 }
 162 EXPORT_SYMBOL(drm_sched_entity_error);
 163
 164 static 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. */
 174 static 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 */
 215 static 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  */
 259 long drm_sched_entity_flush(struct drm_sched_entity *entity, long timeout)
 260 {
 261         struct drm_gpu_scheduler *sched;
 262         struct task_struct *last_user;
 263         long ret = timeout;
 264
 265         if (!entity->rq)
 266                 return 0;
 267
 268         sched = entity->rq->sched;
 269         /**
 270          * The client will not queue more IBs during this fini, consume existing
 271          * queued IBs or discard them on SIGKILL
 272          */
 273         if (current->flags & PF_EXITING) {
 274                 if (timeout)
 275                         ret = wait_event_timeout(
 276                                         sched->job_scheduled,
 277                                         drm_sched_entity_is_idle(entity),
 278                                         timeout);
 279         } else {
 280                 wait_event_killable(sched->job_scheduled,
 281                                     drm_sched_entity_is_idle(entity));
 282         }
 283
 284         /* For killed process disable any more IBs enqueue right now */
 285         last_user = cmpxchg(&entity->last_user, current->group_leader, NULL);
 286         if ((!last_user || last_user == current->group_leader) &&
 287             (current->flags & PF_EXITING) && (current->exit_code == SIGKILL))
 288                 drm_sched_entity_kill(entity);
 289
 290         return ret;
 291 }
 292 EXPORT_SYMBOL(drm_sched_entity_flush);
 293
 294 /**
 295  * drm_sched_entity_fini - Destroy a context entity
 296  *
 297  * @entity: scheduler entity
 298  *
 299  * Cleanups up @entity which has been initialized by drm_sched_entity_init().
 300  *
 301  * If there are potentially job still in flight or getting newly queued
 302  * drm_sched_entity_flush() must be called first. This function then goes over
 303  * the entity and signals all jobs with an error code if the process was killed.
 304  */
 305 void drm_sched_entity_fini(struct drm_sched_entity *entity)
 306 {
 307         /*
 308          * If consumption of existing IBs wasn't completed. Forcefully remove
 309          * them here. Also makes sure that the scheduler won't touch this entity
 310          * any more.
 311          */
 312         drm_sched_entity_kill(entity);
 313
 314         if (entity->dependency) {
 315                 dma_fence_remove_callback(entity->dependency, &entity->cb);
 316                 dma_fence_put(entity->dependency);
 317                 entity->dependency = NULL;
 318         }
 319
 320         dma_fence_put(rcu_dereference_check(entity->last_scheduled, true));
 321         RCU_INIT_POINTER(entity->last_scheduled, NULL);
 322 }
 323 EXPORT_SYMBOL(drm_sched_entity_fini);
 324
 325 /**
 326  * drm_sched_entity_destroy - Destroy a context entity
 327  * @entity: scheduler entity
 328  *
 329  * Calls drm_sched_entity_flush() and drm_sched_entity_fini() as a
 330  * convenience wrapper.
 331  */
 332 void drm_sched_entity_destroy(struct drm_sched_entity *entity)
 333 {
 334         drm_sched_entity_flush(entity, MAX_WAIT_SCHED_ENTITY_Q_EMPTY);
 335         drm_sched_entity_fini(entity);
 336 }
 337 EXPORT_SYMBOL(drm_sched_entity_destroy);
 338
 339 /* drm_sched_entity_clear_dep - callback to clear the entities dependency */
 340 static void drm_sched_entity_clear_dep(struct dma_fence *f,
 341                                        struct dma_fence_cb *cb)
 342 {
 343         struct drm_sched_entity *entity =
 344                 container_of(cb, struct drm_sched_entity, cb);
 345
 346         entity->dependency = NULL;
 347         dma_fence_put(f);
 348 }
 349
 350 /*
 351  * drm_sched_entity_clear_dep - callback to clear the entities dependency and
 352  * wake up scheduler
 353  */
 354 static void drm_sched_entity_wakeup(struct dma_fence *f,
 355                                     struct dma_fence_cb *cb)
 356 {
 357         struct drm_sched_entity *entity =
 358                 container_of(cb, struct drm_sched_entity, cb);
 359
 360         drm_sched_entity_clear_dep(f, cb);
 361         drm_sched_wakeup_if_can_queue(entity->rq->sched);
 362 }
 363
 364 /**
 365  * drm_sched_entity_set_priority - Sets priority of the entity
 366  *
 367  * @entity: scheduler entity
 368  * @priority: scheduler priority
 369  *
 370  * Update the priority of runqueus used for the entity.
 371  */
 372 void drm_sched_entity_set_priority(struct drm_sched_entity *entity,
 373                                    enum drm_sched_priority priority)
 374 {
 375         spin_lock(&entity->rq_lock);
 376         entity->priority = priority;
 377         spin_unlock(&entity->rq_lock);
 378 }
 379 EXPORT_SYMBOL(drm_sched_entity_set_priority);
 380
 381 /*
 382  * Add a callback to the current dependency of the entity to wake up the
 383  * scheduler when the entity becomes available.
 384  */
 385 static bool drm_sched_entity_add_dependency_cb(struct drm_sched_entity *entity)
 386 {
 387         struct drm_gpu_scheduler *sched = entity->rq->sched;
 388         struct dma_fence *fence = entity->dependency;
 389         struct drm_sched_fence *s_fence;
 390
 391         if (fence->context == entity->fence_context ||
 392             fence->context == entity->fence_context + 1) {
 393                 /*
 394                  * Fence is a scheduled/finished fence from a job
 395                  * which belongs to the same entity, we can ignore
 396                  * fences from ourself
 397                  */
 398                 dma_fence_put(entity->dependency);
 399                 return false;
 400         }
 401
 402         s_fence = to_drm_sched_fence(fence);
 403         if (!fence->error && s_fence && s_fence->sched == sched &&
 404             !test_bit(DRM_SCHED_FENCE_DONT_PIPELINE, &fence->flags)) {
 405
 406                 /*
 407                  * Fence is from the same scheduler, only need to wait for
 408                  * it to be scheduled
 409                  */
 410                 fence = dma_fence_get(&s_fence->scheduled);
 411                 dma_fence_put(entity->dependency);
 412                 entity->dependency = fence;
 413                 if (!dma_fence_add_callback(fence, &entity->cb,
 414                                             drm_sched_entity_clear_dep))
 415                         return true;
 416
 417                 /* Ignore it when it is already scheduled */
 418                 dma_fence_put(fence);
 419                 return false;
 420         }
 421
 422         if (!dma_fence_add_callback(entity->dependency, &entity->cb,
 423                                     drm_sched_entity_wakeup))
 424                 return true;
 425
 426         dma_fence_put(entity->dependency);
 427         return false;
 428 }
 429
 430 static struct dma_fence *
 431 drm_sched_job_dependency(struct drm_sched_job *job,
 432                          struct drm_sched_entity *entity)
 433 {
 434         struct dma_fence *f;
 435
 436         /* We keep the fence around, so we can iterate over all dependencies
 437          * in drm_sched_entity_kill_jobs_cb() to ensure all deps are signaled
 438          * before killing the job.
 439          */
 440         f = xa_load(&job->dependencies, job->last_dependency);
 441         if (f) {
 442                 job->last_dependency++;
 443                 return dma_fence_get(f);
 444         }
 445
 446         if (job->sched->ops->prepare_job)
 447                 return job->sched->ops->prepare_job(job, entity);
 448
 449         return NULL;
 450 }
 451
 452 struct drm_sched_job *drm_sched_entity_pop_job(struct drm_sched_entity *entity)
 453 {
 454         struct drm_sched_job *sched_job;
 455
 456         sched_job = to_drm_sched_job(spsc_queue_peek(&entity->job_queue));
 457         if (!sched_job)
 458                 return NULL;
 459
 460         while ((entity->dependency =
 461                         drm_sched_job_dependency(sched_job, entity))) {
 462                 trace_drm_sched_job_wait_dep(sched_job, entity->dependency);
 463
 464                 if (drm_sched_entity_add_dependency_cb(entity))
 465                         return NULL;
 466         }
 467
 468         /* skip jobs from entity that marked guilty */
 469         if (entity->guilty && atomic_read(entity->guilty))
 470                 dma_fence_set_error(&sched_job->s_fence->finished, -ECANCELED);
 471
 472         dma_fence_put(rcu_dereference_check(entity->last_scheduled, true));
 473         rcu_assign_pointer(entity->last_scheduled,
 474                            dma_fence_get(&sched_job->s_fence->finished));
 475
 476         /*
 477          * If the queue is empty we allow drm_sched_entity_select_rq() to
 478          * locklessly access ->last_scheduled. This only works if we set the
 479          * pointer before we dequeue and if we a write barrier here.
 480          */
 481         smp_wmb();
 482
 483         spsc_queue_pop(&entity->job_queue);
 484
 485         /*
 486          * Update the entity's location in the min heap according to
 487          * the timestamp of the next job, if any.
 488          */
 489         if (drm_sched_policy == DRM_SCHED_POLICY_FIFO) {
 490                 struct drm_sched_job *next;
 491
 492                 next = to_drm_sched_job(spsc_queue_peek(&entity->job_queue));
 493                 if (next)
 494                         drm_sched_rq_update_fifo(entity, next->submit_ts);
 495         }
 496
 497         /* Jobs and entities might have different lifecycles. Since we're
 498          * removing the job from the entities queue, set the jobs entity pointer
 499          * to NULL to prevent any future access of the entity through this job.
 500          */
 501         sched_job->entity = NULL;
 502
 503         return sched_job;
 504 }
 505
 506 void drm_sched_entity_select_rq(struct drm_sched_entity *entity)
 507 {
 508         struct dma_fence *fence;
 509         struct drm_gpu_scheduler *sched;
 510         struct drm_sched_rq *rq;
 511
 512         /* single possible engine and already selected */
 513         if (!entity->sched_list)
 514                 return;
 515
 516         /* queue non-empty, stay on the same engine */
 517         if (spsc_queue_count(&entity->job_queue))
 518                 return;
 519
 520         /*
 521          * Only when the queue is empty are we guaranteed that the scheduler
 522          * thread cannot change ->last_scheduled. To enforce ordering we need
 523          * a read barrier here. See drm_sched_entity_pop_job() for the other
 524          * side.
 525          */
 526         smp_rmb();
 527
 528         fence = rcu_dereference_check(entity->last_scheduled, true);
 529
 530         /* stay on the same engine if the previous job hasn't finished */
 531         if (fence && !dma_fence_is_signaled(fence))
 532                 return;
 533
 534         spin_lock(&entity->rq_lock);
 535         sched = drm_sched_pick_best(entity->sched_list, entity->num_sched_list);
 536         rq = sched ? &sched->sched_rq[entity->priority] : NULL;
 537         if (rq != entity->rq) {
 538                 drm_sched_rq_remove_entity(entity->rq, entity);
 539                 entity->rq = rq;
 540         }
 541         spin_unlock(&entity->rq_lock);
 542
 543         if (entity->num_sched_list == 1)
 544                 entity->sched_list = NULL;
 545 }
 546
 547 /**
 548  * drm_sched_entity_push_job - Submit a job to the entity's job queue
 549  * @sched_job: job to submit
 550  *
 551  * Note: To guarantee that the order of insertion to queue matches the job's
 552  * fence sequence number this function should be called with drm_sched_job_arm()
 553  * under common lock for the struct drm_sched_entity that was set up for
 554  * @sched_job in drm_sched_job_init().
 555  *
 556  * Returns 0 for success, negative error code otherwise.
 557  */
 558 void drm_sched_entity_push_job(struct drm_sched_job *sched_job)
 559 {
 560         struct drm_sched_entity *entity = sched_job->entity;
 561         bool first;
 562         ktime_t submit_ts;
 563
 564         trace_drm_sched_job(sched_job, entity);
 565         atomic_inc(entity->rq->sched->score);
 566         WRITE_ONCE(entity->last_user, current->group_leader);
 567
 568         /*
 569          * After the sched_job is pushed into the entity queue, it may be
 570          * completed and freed up at any time. We can no longer access it.
 571          * Make sure to set the submit_ts first, to avoid a race.
 572          */
 573         sched_job->submit_ts = submit_ts = ktime_get();
 574         first = spsc_queue_push(&entity->job_queue, &sched_job->queue_node);
 575
 576         /* first job wakes up scheduler */
 577         if (first) {
 578                 /* Add the entity to the run queue */
 579                 spin_lock(&entity->rq_lock);
 580                 if (entity->stopped) {
 581                         spin_unlock(&entity->rq_lock);
 582
 583                         DRM_ERROR("Trying to push to a killed entity\n");
 584                         return;
 585                 }
 586
 587                 drm_sched_rq_add_entity(entity->rq, entity);
 588                 spin_unlock(&entity->rq_lock);
 589
 590                 if (drm_sched_policy == DRM_SCHED_POLICY_FIFO)
 591                         drm_sched_rq_update_fifo(entity, submit_ts);
 592
 593                 drm_sched_wakeup_if_can_queue(entity->rq->sched);
 594         }
 595 }
 596 EXPORT_SYMBOL(drm_sched_entity_push_job);