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 (!sched_list[0]->sched_rq) {
  79                 /* Warn drivers not to do this and to fix their DRM
  80                  * calling order.
  81                  */
  82                 pr_warn("%s: called with uninitialized scheduler\n", __func__);
  83         } else if (num_sched_list) {
  84                 /* The "priority" of an entity cannot exceed the number
  85                  * of run-queues of a scheduler.
  86                  */
  87                 if (entity->priority >= sched_list[0]->num_rqs)
  88                         entity->priority = max_t(u32, sched_list[0]->num_rqs,
  89                                                  DRM_SCHED_PRIORITY_MIN);
  90                 entity->rq = sched_list[0]->sched_rq[entity->priority];
  91         }
  92
  93         init_completion(&entity->entity_idle);
  94
  95         /* We start in an idle state. */
  96         complete_all(&entity->entity_idle);
  97
  98         spin_lock_init(&entity->rq_lock);
  99         spsc_queue_init(&entity->job_queue);
 100
 101         atomic_set(&entity->fence_seq, 0);
 102         entity->fence_context = dma_fence_context_alloc(2);
 103
 104         return 0;
 105 }
 106 EXPORT_SYMBOL(drm_sched_entity_init);
 107
 108 /**
 109  * drm_sched_entity_modify_sched - Modify sched of an entity
 110  * @entity: scheduler entity to init
 111  * @sched_list: the list of new drm scheds which will replace
 112  *               existing entity->sched_list
 113  * @num_sched_list: number of drm sched in sched_list
 114  *
 115  * Note that this must be called under the same common lock for @entity as
 116  * drm_sched_job_arm() and drm_sched_entity_push_job(), or the driver needs to
 117  * guarantee through some other means that this is never called while new jobs
 118  * can be pushed to @entity.
 119  */
 120 void drm_sched_entity_modify_sched(struct drm_sched_entity *entity,
 121                                     struct drm_gpu_scheduler **sched_list,
 122                                     unsigned int num_sched_list)
 123 {
 124         WARN_ON(!num_sched_list || !sched_list);
 125
 126         entity->sched_list = sched_list;
 127         entity->num_sched_list = num_sched_list;
 128 }
 129 EXPORT_SYMBOL(drm_sched_entity_modify_sched);
 130
 131 static bool drm_sched_entity_is_idle(struct drm_sched_entity *entity)
 132 {
 133         rmb(); /* for list_empty to work without lock */
 134
 135         if (list_empty(&entity->list) ||
 136             spsc_queue_count(&entity->job_queue) == 0 ||
 137             entity->stopped)
 138                 return true;
 139
 140         return false;
 141 }
 142
 143 /* Return true if entity could provide a job. */
 144 bool drm_sched_entity_is_ready(struct drm_sched_entity *entity)
 145 {
 146         if (spsc_queue_peek(&entity->job_queue) == NULL)
 147                 return false;
 148
 149         if (READ_ONCE(entity->dependency))
 150                 return false;
 151
 152         return true;
 153 }
 154
 155 /**
 156  * drm_sched_entity_error - return error of last scheduled job
 157  * @entity: scheduler entity to check
 158  *
 159  * Opportunistically return the error of the last scheduled job. Result can
 160  * change any time when new jobs are pushed to the hw.
 161  */
 162 int drm_sched_entity_error(struct drm_sched_entity *entity)
 163 {
 164         struct dma_fence *fence;
 165         int r;
 166
 167         rcu_read_lock();
 168         fence = rcu_dereference(entity->last_scheduled);
 169         r = fence ? fence->error : 0;
 170         rcu_read_unlock();
 171
 172         return r;
 173 }
 174 EXPORT_SYMBOL(drm_sched_entity_error);
 175
 176 static void drm_sched_entity_kill_jobs_work(struct work_struct *wrk)
 177 {
 178         struct drm_sched_job *job = container_of(wrk, typeof(*job), work);
 179
 180         drm_sched_fence_finished(job->s_fence, -ESRCH);
 181         WARN_ON(job->s_fence->parent);
 182         job->sched->ops->free_job(job);
 183 }
 184
 185 /* Signal the scheduler finished fence when the entity in question is killed. */
 186 static void drm_sched_entity_kill_jobs_cb(struct dma_fence *f,
 187                                           struct dma_fence_cb *cb)
 188 {
 189         struct drm_sched_job *job = container_of(cb, struct drm_sched_job,
 190                                                  finish_cb);
 191         unsigned long index;
 192
 193         dma_fence_put(f);
 194
 195         /* Wait for all dependencies to avoid data corruptions */
 196         xa_for_each(&job->dependencies, index, f) {
 197                 struct drm_sched_fence *s_fence = to_drm_sched_fence(f);
 198
 199                 if (s_fence && f == &s_fence->scheduled) {
 200                         /* The dependencies array had a reference on the scheduled
 201                          * fence, and the finished fence refcount might have
 202                          * dropped to zero. Use dma_fence_get_rcu() so we get
 203                          * a NULL fence in that case.
 204                          */
 205                         f = dma_fence_get_rcu(&s_fence->finished);
 206
 207                         /* Now that we have a reference on the finished fence,
 208                          * we can release the reference the dependencies array
 209                          * had on the scheduled fence.
 210                          */
 211                         dma_fence_put(&s_fence->scheduled);
 212                 }
 213
 214                 xa_erase(&job->dependencies, index);
 215                 if (f && !dma_fence_add_callback(f, &job->finish_cb,
 216                                                  drm_sched_entity_kill_jobs_cb))
 217                         return;
 218
 219                 dma_fence_put(f);
 220         }
 221
 222         INIT_WORK(&job->work, drm_sched_entity_kill_jobs_work);
 223         schedule_work(&job->work);
 224 }
 225
 226 /* Remove the entity from the scheduler and kill all pending jobs */
 227 static void drm_sched_entity_kill(struct drm_sched_entity *entity)
 228 {
 229         struct drm_sched_job *job;
 230         struct dma_fence *prev;
 231
 232         if (!entity->rq)
 233                 return;
 234
 235         spin_lock(&entity->rq_lock);
 236         entity->stopped = true;
 237         drm_sched_rq_remove_entity(entity->rq, entity);
 238         spin_unlock(&entity->rq_lock);
 239
 240         /* Make sure this entity is not used by the scheduler at the moment */
 241         wait_for_completion(&entity->entity_idle);
 242
 243         /* The entity is guaranteed to not be used by the scheduler */
 244         prev = rcu_dereference_check(entity->last_scheduled, true);
 245         dma_fence_get(prev);
 246         while ((job = to_drm_sched_job(spsc_queue_pop(&entity->job_queue)))) {
 247                 struct drm_sched_fence *s_fence = job->s_fence;
 248
 249                 dma_fence_get(&s_fence->finished);
 250                 if (!prev || dma_fence_add_callback(prev, &job->finish_cb,
 251                                            drm_sched_entity_kill_jobs_cb))
 252                         drm_sched_entity_kill_jobs_cb(NULL, &job->finish_cb);
 253
 254                 prev = &s_fence->finished;
 255         }
 256         dma_fence_put(prev);
 257 }
 258
 259 /**
 260  * drm_sched_entity_flush - Flush a context entity
 261  *
 262  * @entity: scheduler entity
 263  * @timeout: time to wait in for Q to become empty in jiffies.
 264  *
 265  * Splitting drm_sched_entity_fini() into two functions, The first one does the
 266  * waiting, removes the entity from the runqueue and returns an error when the
 267  * process was killed.
 268  *
 269  * Returns the remaining time in jiffies left from the input timeout
 270  */
 271 long drm_sched_entity_flush(struct drm_sched_entity *entity, long timeout)
 272 {
 273         struct drm_gpu_scheduler *sched;
 274         struct task_struct *last_user;
 275         long ret = timeout;
 276
 277         if (!entity->rq)
 278                 return 0;
 279
 280         sched = entity->rq->sched;
 281         /**
 282          * The client will not queue more IBs during this fini, consume existing
 283          * queued IBs or discard them on SIGKILL
 284          */
 285         if (current->flags & PF_EXITING) {
 286                 if (timeout)
 287                         ret = wait_event_timeout(
 288                                         sched->job_scheduled,
 289                                         drm_sched_entity_is_idle(entity),
 290                                         timeout);
 291         } else {
 292                 wait_event_killable(sched->job_scheduled,
 293                                     drm_sched_entity_is_idle(entity));
 294         }
 295
 296         /* For killed process disable any more IBs enqueue right now */
 297         last_user = cmpxchg(&entity->last_user, current->group_leader, NULL);
 298         if ((!last_user || last_user == current->group_leader) &&
 299             (current->flags & PF_EXITING) && (current->exit_code == SIGKILL))
 300                 drm_sched_entity_kill(entity);
 301
 302         return ret;
 303 }
 304 EXPORT_SYMBOL(drm_sched_entity_flush);
 305
 306 /**
 307  * drm_sched_entity_fini - Destroy a context entity
 308  *
 309  * @entity: scheduler entity
 310  *
 311  * Cleanups up @entity which has been initialized by drm_sched_entity_init().
 312  *
 313  * If there are potentially job still in flight or getting newly queued
 314  * drm_sched_entity_flush() must be called first. This function then goes over
 315  * the entity and signals all jobs with an error code if the process was killed.
 316  */
 317 void drm_sched_entity_fini(struct drm_sched_entity *entity)
 318 {
 319         /*
 320          * If consumption of existing IBs wasn't completed. Forcefully remove
 321          * them here. Also makes sure that the scheduler won't touch this entity
 322          * any more.
 323          */
 324         drm_sched_entity_kill(entity);
 325
 326         if (entity->dependency) {
 327                 dma_fence_remove_callback(entity->dependency, &entity->cb);
 328                 dma_fence_put(entity->dependency);
 329                 entity->dependency = NULL;
 330         }
 331
 332         dma_fence_put(rcu_dereference_check(entity->last_scheduled, true));
 333         RCU_INIT_POINTER(entity->last_scheduled, NULL);
 334 }
 335 EXPORT_SYMBOL(drm_sched_entity_fini);
 336
 337 /**
 338  * drm_sched_entity_destroy - Destroy a context entity
 339  * @entity: scheduler entity
 340  *
 341  * Calls drm_sched_entity_flush() and drm_sched_entity_fini() as a
 342  * convenience wrapper.
 343  */
 344 void drm_sched_entity_destroy(struct drm_sched_entity *entity)
 345 {
 346         drm_sched_entity_flush(entity, MAX_WAIT_SCHED_ENTITY_Q_EMPTY);
 347         drm_sched_entity_fini(entity);
 348 }
 349 EXPORT_SYMBOL(drm_sched_entity_destroy);
 350
 351 /* drm_sched_entity_clear_dep - callback to clear the entities dependency */
 352 static void drm_sched_entity_clear_dep(struct dma_fence *f,
 353                                        struct dma_fence_cb *cb)
 354 {
 355         struct drm_sched_entity *entity =
 356                 container_of(cb, struct drm_sched_entity, cb);
 357
 358         entity->dependency = NULL;
 359         dma_fence_put(f);
 360 }
 361
 362 /*
 363  * drm_sched_entity_clear_dep - callback to clear the entities dependency and
 364  * wake up scheduler
 365  */
 366 static void drm_sched_entity_wakeup(struct dma_fence *f,
 367                                     struct dma_fence_cb *cb)
 368 {
 369         struct drm_sched_entity *entity =
 370                 container_of(cb, struct drm_sched_entity, cb);
 371
 372         drm_sched_entity_clear_dep(f, cb);
 373         drm_sched_wakeup_if_can_queue(entity->rq->sched);
 374 }
 375
 376 /**
 377  * drm_sched_entity_set_priority - Sets priority of the entity
 378  *
 379  * @entity: scheduler entity
 380  * @priority: scheduler priority
 381  *
 382  * Update the priority of runqueus used for the entity.
 383  */
 384 void drm_sched_entity_set_priority(struct drm_sched_entity *entity,
 385                                    enum drm_sched_priority priority)
 386 {
 387         spin_lock(&entity->rq_lock);
 388         entity->priority = priority;
 389         spin_unlock(&entity->rq_lock);
 390 }
 391 EXPORT_SYMBOL(drm_sched_entity_set_priority);
 392
 393 /*
 394  * Add a callback to the current dependency of the entity to wake up the
 395  * scheduler when the entity becomes available.
 396  */
 397 static bool drm_sched_entity_add_dependency_cb(struct drm_sched_entity *entity)
 398 {
 399         struct drm_gpu_scheduler *sched = entity->rq->sched;
 400         struct dma_fence *fence = entity->dependency;
 401         struct drm_sched_fence *s_fence;
 402
 403         if (fence->context == entity->fence_context ||
 404             fence->context == entity->fence_context + 1) {
 405                 /*
 406                  * Fence is a scheduled/finished fence from a job
 407                  * which belongs to the same entity, we can ignore
 408                  * fences from ourself
 409                  */
 410                 dma_fence_put(entity->dependency);
 411                 return false;
 412         }
 413
 414         s_fence = to_drm_sched_fence(fence);
 415         if (!fence->error && s_fence && s_fence->sched == sched &&
 416             !test_bit(DRM_SCHED_FENCE_DONT_PIPELINE, &fence->flags)) {
 417
 418                 /*
 419                  * Fence is from the same scheduler, only need to wait for
 420                  * it to be scheduled
 421                  */
 422                 fence = dma_fence_get(&s_fence->scheduled);
 423                 dma_fence_put(entity->dependency);
 424                 entity->dependency = fence;
 425                 if (!dma_fence_add_callback(fence, &entity->cb,
 426                                             drm_sched_entity_clear_dep))
 427                         return true;
 428
 429                 /* Ignore it when it is already scheduled */
 430                 dma_fence_put(fence);
 431                 return false;
 432         }
 433
 434         if (!dma_fence_add_callback(entity->dependency, &entity->cb,
 435                                     drm_sched_entity_wakeup))
 436                 return true;
 437
 438         dma_fence_put(entity->dependency);
 439         return false;
 440 }
 441
 442 static struct dma_fence *
 443 drm_sched_job_dependency(struct drm_sched_job *job,
 444                          struct drm_sched_entity *entity)
 445 {
 446         struct dma_fence *f;
 447
 448         /* We keep the fence around, so we can iterate over all dependencies
 449          * in drm_sched_entity_kill_jobs_cb() to ensure all deps are signaled
 450          * before killing the job.
 451          */
 452         f = xa_load(&job->dependencies, job->last_dependency);
 453         if (f) {
 454                 job->last_dependency++;
 455                 return dma_fence_get(f);
 456         }
 457
 458         if (job->sched->ops->prepare_job)
 459                 return job->sched->ops->prepare_job(job, entity);
 460
 461         return NULL;
 462 }
 463
 464 struct drm_sched_job *drm_sched_entity_pop_job(struct drm_sched_entity *entity)
 465 {
 466         struct drm_sched_job *sched_job;
 467
 468         sched_job = to_drm_sched_job(spsc_queue_peek(&entity->job_queue));
 469         if (!sched_job)
 470                 return NULL;
 471
 472         while ((entity->dependency =
 473                         drm_sched_job_dependency(sched_job, entity))) {
 474                 trace_drm_sched_job_wait_dep(sched_job, entity->dependency);
 475
 476                 if (drm_sched_entity_add_dependency_cb(entity))
 477                         return NULL;
 478         }
 479
 480         /* skip jobs from entity that marked guilty */
 481         if (entity->guilty && atomic_read(entity->guilty))
 482                 dma_fence_set_error(&sched_job->s_fence->finished, -ECANCELED);
 483
 484         dma_fence_put(rcu_dereference_check(entity->last_scheduled, true));
 485         rcu_assign_pointer(entity->last_scheduled,
 486                            dma_fence_get(&sched_job->s_fence->finished));
 487
 488         /*
 489          * If the queue is empty we allow drm_sched_entity_select_rq() to
 490          * locklessly access ->last_scheduled. This only works if we set the
 491          * pointer before we dequeue and if we a write barrier here.
 492          */
 493         smp_wmb();
 494
 495         spsc_queue_pop(&entity->job_queue);
 496
 497         /*
 498          * Update the entity's location in the min heap according to
 499          * the timestamp of the next job, if any.
 500          */
 501         if (drm_sched_policy == DRM_SCHED_POLICY_FIFO) {
 502                 struct drm_sched_job *next;
 503
 504                 next = to_drm_sched_job(spsc_queue_peek(&entity->job_queue));
 505                 if (next)
 506                         drm_sched_rq_update_fifo(entity, next->submit_ts);
 507         }
 508
 509         /* Jobs and entities might have different lifecycles. Since we're
 510          * removing the job from the entities queue, set the jobs entity pointer
 511          * to NULL to prevent any future access of the entity through this job.
 512          */
 513         sched_job->entity = NULL;
 514
 515         return sched_job;
 516 }
 517
 518 void drm_sched_entity_select_rq(struct drm_sched_entity *entity)
 519 {
 520         struct dma_fence *fence;
 521         struct drm_gpu_scheduler *sched;
 522         struct drm_sched_rq *rq;
 523
 524         /* single possible engine and already selected */
 525         if (!entity->sched_list)
 526                 return;
 527
 528         /* queue non-empty, stay on the same engine */
 529         if (spsc_queue_count(&entity->job_queue))
 530                 return;
 531
 532         /*
 533          * Only when the queue is empty are we guaranteed that the scheduler
 534          * thread cannot change ->last_scheduled. To enforce ordering we need
 535          * a read barrier here. See drm_sched_entity_pop_job() for the other
 536          * side.
 537          */
 538         smp_rmb();
 539
 540         fence = rcu_dereference_check(entity->last_scheduled, true);
 541
 542         /* stay on the same engine if the previous job hasn't finished */
 543         if (fence && !dma_fence_is_signaled(fence))
 544                 return;
 545
 546         spin_lock(&entity->rq_lock);
 547         sched = drm_sched_pick_best(entity->sched_list, entity->num_sched_list);
 548         rq = sched ? sched->sched_rq[entity->priority] : NULL;
 549         if (rq != entity->rq) {
 550                 drm_sched_rq_remove_entity(entity->rq, entity);
 551                 entity->rq = rq;
 552         }
 553         spin_unlock(&entity->rq_lock);
 554
 555         if (entity->num_sched_list == 1)
 556                 entity->sched_list = NULL;
 557 }
 558
 559 /**
 560  * drm_sched_entity_push_job - Submit a job to the entity's job queue
 561  * @sched_job: job to submit
 562  *
 563  * Note: To guarantee that the order of insertion to queue matches the job's
 564  * fence sequence number this function should be called with drm_sched_job_arm()
 565  * under common lock for the struct drm_sched_entity that was set up for
 566  * @sched_job in drm_sched_job_init().
 567  *
 568  * Returns 0 for success, negative error code otherwise.
 569  */
 570 void drm_sched_entity_push_job(struct drm_sched_job *sched_job)
 571 {
 572         struct drm_sched_entity *entity = sched_job->entity;
 573         bool first;
 574         ktime_t submit_ts;
 575
 576         trace_drm_sched_job(sched_job, entity);
 577         atomic_inc(entity->rq->sched->score);
 578         WRITE_ONCE(entity->last_user, current->group_leader);
 579
 580         /*
 581          * After the sched_job is pushed into the entity queue, it may be
 582          * completed and freed up at any time. We can no longer access it.
 583          * Make sure to set the submit_ts first, to avoid a race.
 584          */
 585         sched_job->submit_ts = submit_ts = ktime_get();
 586         first = spsc_queue_push(&entity->job_queue, &sched_job->queue_node);
 587
 588         /* first job wakes up scheduler */
 589         if (first) {
 590                 /* Add the entity to the run queue */
 591                 spin_lock(&entity->rq_lock);
 592                 if (entity->stopped) {
 593                         spin_unlock(&entity->rq_lock);
 594
 595                         DRM_ERROR("Trying to push to a killed entity\n");
 596                         return;
 597                 }
 598
 599                 drm_sched_rq_add_entity(entity->rq, entity);
 600                 spin_unlock(&entity->rq_lock);
 601
 602                 if (drm_sched_policy == DRM_SCHED_POLICY_FIFO)
 603                         drm_sched_rq_update_fifo(entity, submit_ts);
 604
 605                 drm_sched_wakeup_if_can_queue(entity->rq->sched);
 606         }
 607 }
 608 EXPORT_SYMBOL(drm_sched_entity_push_job);