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         /*
  75          * It's perfectly valid to initialize an entity without having a valid
  76          * scheduler attached. It's just not valid to use the scheduler before it
  77          * is initialized itself.
  78          */
  79         entity->sched_list = num_sched_list > 1 ? sched_list : NULL;
  80         RCU_INIT_POINTER(entity->last_scheduled, NULL);
  81         RB_CLEAR_NODE(&entity->rb_tree_node);
  82
  83         if (num_sched_list && !sched_list[0]->sched_rq) {
  84                 /* Since every entry covered by num_sched_list
  85                  * should be non-NULL and therefore we warn drivers
  86                  * not to do this and to fix their DRM calling order.
  87                  */
  88                 pr_warn("%s: called with uninitialized scheduler\n", __func__);
  89         } else if (num_sched_list) {
  90                 /* The "priority" of an entity cannot exceed the number of run-queues of a
  91                  * scheduler. Protect against num_rqs being 0, by converting to signed. Choose
  92                  * the lowest priority available.
  93                  */
  94                 if (entity->priority >= sched_list[0]->num_rqs) {
  95                         drm_err(sched_list[0], "entity with out-of-bounds priority:%u num_rqs:%u\n",
  96                                 entity->priority, sched_list[0]->num_rqs);
  97                         entity->priority = max_t(s32, (s32) sched_list[0]->num_rqs - 1,
  98                                                  (s32) DRM_SCHED_PRIORITY_KERNEL);
  99                 }
 100                 entity->rq = sched_list[0]->sched_rq[entity->priority];
 101         }
 102
 103         init_completion(&entity->entity_idle);
 104
 105         /* We start in an idle state. */
 106         complete_all(&entity->entity_idle);
 107
 108         spin_lock_init(&entity->rq_lock);
 109         spsc_queue_init(&entity->job_queue);
 110
 111         atomic_set(&entity->fence_seq, 0);
 112         entity->fence_context = dma_fence_context_alloc(2);
 113
 114         return 0;
 115 }
 116 EXPORT_SYMBOL(drm_sched_entity_init);
 117
 118 /**
 119  * drm_sched_entity_modify_sched - Modify sched of an entity
 120  * @entity: scheduler entity to init
 121  * @sched_list: the list of new drm scheds which will replace
 122  *               existing entity->sched_list
 123  * @num_sched_list: number of drm sched in sched_list
 124  *
 125  * Note that this must be called under the same common lock for @entity as
 126  * drm_sched_job_arm() and drm_sched_entity_push_job(), or the driver needs to
 127  * guarantee through some other means that this is never called while new jobs
 128  * can be pushed to @entity.
 129  */
 130 void drm_sched_entity_modify_sched(struct drm_sched_entity *entity,
 131                                     struct drm_gpu_scheduler **sched_list,
 132                                     unsigned int num_sched_list)
 133 {
 134         WARN_ON(!num_sched_list || !sched_list);
 135
 136         entity->sched_list = sched_list;
 137         entity->num_sched_list = num_sched_list;
 138 }
 139 EXPORT_SYMBOL(drm_sched_entity_modify_sched);
 140
 141 static bool drm_sched_entity_is_idle(struct drm_sched_entity *entity)
 142 {
 143         rmb(); /* for list_empty to work without lock */
 144
 145         if (list_empty(&entity->list) ||
 146             spsc_queue_count(&entity->job_queue) == 0 ||
 147             entity->stopped)
 148                 return true;
 149
 150         return false;
 151 }
 152
 153 /* Return true if entity could provide a job. */
 154 bool drm_sched_entity_is_ready(struct drm_sched_entity *entity)
 155 {
 156         if (spsc_queue_peek(&entity->job_queue) == NULL)
 157                 return false;
 158
 159         if (READ_ONCE(entity->dependency))
 160                 return false;
 161
 162         return true;
 163 }
 164
 165 /**
 166  * drm_sched_entity_error - return error of last scheduled job
 167  * @entity: scheduler entity to check
 168  *
 169  * Opportunistically return the error of the last scheduled job. Result can
 170  * change any time when new jobs are pushed to the hw.
 171  */
 172 int drm_sched_entity_error(struct drm_sched_entity *entity)
 173 {
 174         struct dma_fence *fence;
 175         int r;
 176
 177         rcu_read_lock();
 178         fence = rcu_dereference(entity->last_scheduled);
 179         r = fence ? fence->error : 0;
 180         rcu_read_unlock();
 181
 182         return r;
 183 }
 184 EXPORT_SYMBOL(drm_sched_entity_error);
 185
 186 static void drm_sched_entity_kill_jobs_work(struct work_struct *wrk)
 187 {
 188         struct drm_sched_job *job = container_of(wrk, typeof(*job), work);
 189
 190         drm_sched_fence_finished(job->s_fence, -ESRCH);
 191         WARN_ON(job->s_fence->parent);
 192         job->sched->ops->free_job(job);
 193 }
 194
 195 /* Signal the scheduler finished fence when the entity in question is killed. */
 196 static void drm_sched_entity_kill_jobs_cb(struct dma_fence *f,
 197                                           struct dma_fence_cb *cb)
 198 {
 199         struct drm_sched_job *job = container_of(cb, struct drm_sched_job,
 200                                                  finish_cb);
 201         unsigned long index;
 202
 203         dma_fence_put(f);
 204
 205         /* Wait for all dependencies to avoid data corruptions */
 206         xa_for_each(&job->dependencies, index, f) {
 207                 struct drm_sched_fence *s_fence = to_drm_sched_fence(f);
 208
 209                 if (s_fence && f == &s_fence->scheduled) {
 210                         /* The dependencies array had a reference on the scheduled
 211                          * fence, and the finished fence refcount might have
 212                          * dropped to zero. Use dma_fence_get_rcu() so we get
 213                          * a NULL fence in that case.
 214                          */
 215                         f = dma_fence_get_rcu(&s_fence->finished);
 216
 217                         /* Now that we have a reference on the finished fence,
 218                          * we can release the reference the dependencies array
 219                          * had on the scheduled fence.
 220                          */
 221                         dma_fence_put(&s_fence->scheduled);
 222                 }
 223
 224                 xa_erase(&job->dependencies, index);
 225                 if (f && !dma_fence_add_callback(f, &job->finish_cb,
 226                                                  drm_sched_entity_kill_jobs_cb))
 227                         return;
 228
 229                 dma_fence_put(f);
 230         }
 231
 232         INIT_WORK(&job->work, drm_sched_entity_kill_jobs_work);
 233         schedule_work(&job->work);
 234 }
 235
 236 /* Remove the entity from the scheduler and kill all pending jobs */
 237 static void drm_sched_entity_kill(struct drm_sched_entity *entity)
 238 {
 239         struct drm_sched_job *job;
 240         struct dma_fence *prev;
 241
 242         if (!entity->rq)
 243                 return;
 244
 245         spin_lock(&entity->rq_lock);
 246         entity->stopped = true;
 247         drm_sched_rq_remove_entity(entity->rq, entity);
 248         spin_unlock(&entity->rq_lock);
 249
 250         /* Make sure this entity is not used by the scheduler at the moment */
 251         wait_for_completion(&entity->entity_idle);
 252
 253         /* The entity is guaranteed to not be used by the scheduler */
 254         prev = rcu_dereference_check(entity->last_scheduled, true);
 255         dma_fence_get(prev);
 256         while ((job = to_drm_sched_job(spsc_queue_pop(&entity->job_queue)))) {
 257                 struct drm_sched_fence *s_fence = job->s_fence;
 258
 259                 dma_fence_get(&s_fence->finished);
 260                 if (!prev || dma_fence_add_callback(prev, &job->finish_cb,
 261                                            drm_sched_entity_kill_jobs_cb))
 262                         drm_sched_entity_kill_jobs_cb(NULL, &job->finish_cb);
 263
 264                 prev = &s_fence->finished;
 265         }
 266         dma_fence_put(prev);
 267 }
 268
 269 /**
 270  * drm_sched_entity_flush - Flush a context entity
 271  *
 272  * @entity: scheduler entity
 273  * @timeout: time to wait in for Q to become empty in jiffies.
 274  *
 275  * Splitting drm_sched_entity_fini() into two functions, The first one does the
 276  * waiting, removes the entity from the runqueue and returns an error when the
 277  * process was killed.
 278  *
 279  * Returns the remaining time in jiffies left from the input timeout
 280  */
 281 long drm_sched_entity_flush(struct drm_sched_entity *entity, long timeout)
 282 {
 283         struct drm_gpu_scheduler *sched;
 284         struct task_struct *last_user;
 285         long ret = timeout;
 286
 287         if (!entity->rq)
 288                 return 0;
 289
 290         sched = entity->rq->sched;
 291         /**
 292          * The client will not queue more IBs during this fini, consume existing
 293          * queued IBs or discard them on SIGKILL
 294          */
 295         if (current->flags & PF_EXITING) {
 296                 if (timeout)
 297                         ret = wait_event_timeout(
 298                                         sched->job_scheduled,
 299                                         drm_sched_entity_is_idle(entity),
 300                                         timeout);
 301         } else {
 302                 wait_event_killable(sched->job_scheduled,
 303                                     drm_sched_entity_is_idle(entity));
 304         }
 305
 306         /* For killed process disable any more IBs enqueue right now */
 307         last_user = cmpxchg(&entity->last_user, current->group_leader, NULL);
 308         if ((!last_user || last_user == current->group_leader) &&
 309             (current->flags & PF_EXITING) && (current->exit_code == SIGKILL))
 310                 drm_sched_entity_kill(entity);
 311
 312         return ret;
 313 }
 314 EXPORT_SYMBOL(drm_sched_entity_flush);
 315
 316 /**
 317  * drm_sched_entity_fini - Destroy a context entity
 318  *
 319  * @entity: scheduler entity
 320  *
 321  * Cleanups up @entity which has been initialized by drm_sched_entity_init().
 322  *
 323  * If there are potentially job still in flight or getting newly queued
 324  * drm_sched_entity_flush() must be called first. This function then goes over
 325  * the entity and signals all jobs with an error code if the process was killed.
 326  */
 327 void drm_sched_entity_fini(struct drm_sched_entity *entity)
 328 {
 329         /*
 330          * If consumption of existing IBs wasn't completed. Forcefully remove
 331          * them here. Also makes sure that the scheduler won't touch this entity
 332          * any more.
 333          */
 334         drm_sched_entity_kill(entity);
 335
 336         if (entity->dependency) {
 337                 dma_fence_remove_callback(entity->dependency, &entity->cb);
 338                 dma_fence_put(entity->dependency);
 339                 entity->dependency = NULL;
 340         }
 341
 342         dma_fence_put(rcu_dereference_check(entity->last_scheduled, true));
 343         RCU_INIT_POINTER(entity->last_scheduled, NULL);
 344 }
 345 EXPORT_SYMBOL(drm_sched_entity_fini);
 346
 347 /**
 348  * drm_sched_entity_destroy - Destroy a context entity
 349  * @entity: scheduler entity
 350  *
 351  * Calls drm_sched_entity_flush() and drm_sched_entity_fini() as a
 352  * convenience wrapper.
 353  */
 354 void drm_sched_entity_destroy(struct drm_sched_entity *entity)
 355 {
 356         drm_sched_entity_flush(entity, MAX_WAIT_SCHED_ENTITY_Q_EMPTY);
 357         drm_sched_entity_fini(entity);
 358 }
 359 EXPORT_SYMBOL(drm_sched_entity_destroy);
 360
 361 /* drm_sched_entity_clear_dep - callback to clear the entities dependency */
 362 static void drm_sched_entity_clear_dep(struct dma_fence *f,
 363                                        struct dma_fence_cb *cb)
 364 {
 365         struct drm_sched_entity *entity =
 366                 container_of(cb, struct drm_sched_entity, cb);
 367
 368         entity->dependency = NULL;
 369         dma_fence_put(f);
 370 }
 371
 372 /*
 373  * drm_sched_entity_clear_dep - callback to clear the entities dependency and
 374  * wake up scheduler
 375  */
 376 static void drm_sched_entity_wakeup(struct dma_fence *f,
 377                                     struct dma_fence_cb *cb)
 378 {
 379         struct drm_sched_entity *entity =
 380                 container_of(cb, struct drm_sched_entity, cb);
 381
 382         drm_sched_entity_clear_dep(f, cb);
 383         drm_sched_wakeup(entity->rq->sched, entity);
 384 }
 385
 386 /**
 387  * drm_sched_entity_set_priority - Sets priority of the entity
 388  *
 389  * @entity: scheduler entity
 390  * @priority: scheduler priority
 391  *
 392  * Update the priority of runqueus used for the entity.
 393  */
 394 void drm_sched_entity_set_priority(struct drm_sched_entity *entity,
 395                                    enum drm_sched_priority priority)
 396 {
 397         spin_lock(&entity->rq_lock);
 398         entity->priority = priority;
 399         spin_unlock(&entity->rq_lock);
 400 }
 401 EXPORT_SYMBOL(drm_sched_entity_set_priority);
 402
 403 /*
 404  * Add a callback to the current dependency of the entity to wake up the
 405  * scheduler when the entity becomes available.
 406  */
 407 static bool drm_sched_entity_add_dependency_cb(struct drm_sched_entity *entity)
 408 {
 409         struct drm_gpu_scheduler *sched = entity->rq->sched;
 410         struct dma_fence *fence = entity->dependency;
 411         struct drm_sched_fence *s_fence;
 412
 413         if (fence->context == entity->fence_context ||
 414             fence->context == entity->fence_context + 1) {
 415                 /*
 416                  * Fence is a scheduled/finished fence from a job
 417                  * which belongs to the same entity, we can ignore
 418                  * fences from ourself
 419                  */
 420                 dma_fence_put(entity->dependency);
 421                 return false;
 422         }
 423
 424         s_fence = to_drm_sched_fence(fence);
 425         if (!fence->error && s_fence && s_fence->sched == sched &&
 426             !test_bit(DRM_SCHED_FENCE_DONT_PIPELINE, &fence->flags)) {
 427
 428                 /*
 429                  * Fence is from the same scheduler, only need to wait for
 430                  * it to be scheduled
 431                  */
 432                 fence = dma_fence_get(&s_fence->scheduled);
 433                 dma_fence_put(entity->dependency);
 434                 entity->dependency = fence;
 435                 if (!dma_fence_add_callback(fence, &entity->cb,
 436                                             drm_sched_entity_clear_dep))
 437                         return true;
 438
 439                 /* Ignore it when it is already scheduled */
 440                 dma_fence_put(fence);
 441                 return false;
 442         }
 443
 444         if (!dma_fence_add_callback(entity->dependency, &entity->cb,
 445                                     drm_sched_entity_wakeup))
 446                 return true;
 447
 448         dma_fence_put(entity->dependency);
 449         return false;
 450 }
 451
 452 static struct dma_fence *
 453 drm_sched_job_dependency(struct drm_sched_job *job,
 454                          struct drm_sched_entity *entity)
 455 {
 456         struct dma_fence *f;
 457
 458         /* We keep the fence around, so we can iterate over all dependencies
 459          * in drm_sched_entity_kill_jobs_cb() to ensure all deps are signaled
 460          * before killing the job.
 461          */
 462         f = xa_load(&job->dependencies, job->last_dependency);
 463         if (f) {
 464                 job->last_dependency++;
 465                 return dma_fence_get(f);
 466         }
 467
 468         if (job->sched->ops->prepare_job)
 469                 return job->sched->ops->prepare_job(job, entity);
 470
 471         return NULL;
 472 }
 473
 474 struct drm_sched_job *drm_sched_entity_pop_job(struct drm_sched_entity *entity)
 475 {
 476         struct drm_sched_job *sched_job;
 477
 478         sched_job = to_drm_sched_job(spsc_queue_peek(&entity->job_queue));
 479         if (!sched_job)
 480                 return NULL;
 481
 482         while ((entity->dependency =
 483                         drm_sched_job_dependency(sched_job, entity))) {
 484                 trace_drm_sched_job_wait_dep(sched_job, entity->dependency);
 485
 486                 if (drm_sched_entity_add_dependency_cb(entity))
 487                         return NULL;
 488         }
 489
 490         /* skip jobs from entity that marked guilty */
 491         if (entity->guilty && atomic_read(entity->guilty))
 492                 dma_fence_set_error(&sched_job->s_fence->finished, -ECANCELED);
 493
 494         dma_fence_put(rcu_dereference_check(entity->last_scheduled, true));
 495         rcu_assign_pointer(entity->last_scheduled,
 496                            dma_fence_get(&sched_job->s_fence->finished));
 497
 498         /*
 499          * If the queue is empty we allow drm_sched_entity_select_rq() to
 500          * locklessly access ->last_scheduled. This only works if we set the
 501          * pointer before we dequeue and if we a write barrier here.
 502          */
 503         smp_wmb();
 504
 505         spsc_queue_pop(&entity->job_queue);
 506
 507         /*
 508          * Update the entity's location in the min heap according to
 509          * the timestamp of the next job, if any.
 510          */
 511         if (drm_sched_policy == DRM_SCHED_POLICY_FIFO) {
 512                 struct drm_sched_job *next;
 513
 514                 next = to_drm_sched_job(spsc_queue_peek(&entity->job_queue));
 515                 if (next)
 516                         drm_sched_rq_update_fifo(entity, next->submit_ts);
 517         }
 518
 519         /* Jobs and entities might have different lifecycles. Since we're
 520          * removing the job from the entities queue, set the jobs entity pointer
 521          * to NULL to prevent any future access of the entity through this job.
 522          */
 523         sched_job->entity = NULL;
 524
 525         return sched_job;
 526 }
 527
 528 void drm_sched_entity_select_rq(struct drm_sched_entity *entity)
 529 {
 530         struct dma_fence *fence;
 531         struct drm_gpu_scheduler *sched;
 532         struct drm_sched_rq *rq;
 533
 534         /* single possible engine and already selected */
 535         if (!entity->sched_list)
 536                 return;
 537
 538         /* queue non-empty, stay on the same engine */
 539         if (spsc_queue_count(&entity->job_queue))
 540                 return;
 541
 542         /*
 543          * Only when the queue is empty are we guaranteed that the scheduler
 544          * thread cannot change ->last_scheduled. To enforce ordering we need
 545          * a read barrier here. See drm_sched_entity_pop_job() for the other
 546          * side.
 547          */
 548         smp_rmb();
 549
 550         fence = rcu_dereference_check(entity->last_scheduled, true);
 551
 552         /* stay on the same engine if the previous job hasn't finished */
 553         if (fence && !dma_fence_is_signaled(fence))
 554                 return;
 555
 556         spin_lock(&entity->rq_lock);
 557         sched = drm_sched_pick_best(entity->sched_list, entity->num_sched_list);
 558         rq = sched ? sched->sched_rq[entity->priority] : NULL;
 559         if (rq != entity->rq) {
 560                 drm_sched_rq_remove_entity(entity->rq, entity);
 561                 entity->rq = rq;
 562         }
 563         spin_unlock(&entity->rq_lock);
 564
 565         if (entity->num_sched_list == 1)
 566                 entity->sched_list = NULL;
 567 }
 568
 569 /**
 570  * drm_sched_entity_push_job - Submit a job to the entity's job queue
 571  * @sched_job: job to submit
 572  *
 573  * Note: To guarantee that the order of insertion to queue matches the job's
 574  * fence sequence number this function should be called with drm_sched_job_arm()
 575  * under common lock for the struct drm_sched_entity that was set up for
 576  * @sched_job in drm_sched_job_init().
 577  *
 578  * Returns 0 for success, negative error code otherwise.
 579  */
 580 void drm_sched_entity_push_job(struct drm_sched_job *sched_job)
 581 {
 582         struct drm_sched_entity *entity = sched_job->entity;
 583         bool first;
 584         ktime_t submit_ts;
 585
 586         trace_drm_sched_job(sched_job, entity);
 587         atomic_inc(entity->rq->sched->score);
 588         WRITE_ONCE(entity->last_user, current->group_leader);
 589
 590         /*
 591          * After the sched_job is pushed into the entity queue, it may be
 592          * completed and freed up at any time. We can no longer access it.
 593          * Make sure to set the submit_ts first, to avoid a race.
 594          */
 595         sched_job->submit_ts = submit_ts = ktime_get();
 596         first = spsc_queue_push(&entity->job_queue, &sched_job->queue_node);
 597
 598         /* first job wakes up scheduler */
 599         if (first) {
 600                 /* Add the entity to the run queue */
 601                 spin_lock(&entity->rq_lock);
 602                 if (entity->stopped) {
 603                         spin_unlock(&entity->rq_lock);
 604
 605                         DRM_ERROR("Trying to push to a killed entity\n");
 606                         return;
 607                 }
 608
 609                 drm_sched_rq_add_entity(entity->rq, entity);
 610                 spin_unlock(&entity->rq_lock);
 611
 612                 if (drm_sched_policy == DRM_SCHED_POLICY_FIFO)
 613                         drm_sched_rq_update_fifo(entity, submit_ts);
 614
 615                 drm_sched_wakeup(entity->rq->sched, entity);
 616         }
 617 }
 618 EXPORT_SYMBOL(drm_sched_entity_push_job);