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         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 static void drm_sched_entity_kill_jobs_work(struct work_struct *wrk)
 144 {
 145         struct drm_sched_job *job = container_of(wrk, typeof(*job), work);
 146
 147         drm_sched_fence_finished(job->s_fence);
 148         WARN_ON(job->s_fence->parent);
 149         job->sched->ops->free_job(job);
 150 }
 151
 152 /* Signal the scheduler finished fence when the entity in question is killed. */
 153 static void drm_sched_entity_kill_jobs_cb(struct dma_fence *f,
 154                                           struct dma_fence_cb *cb)
 155 {
 156         struct drm_sched_job *job = container_of(cb, struct drm_sched_job,
 157                                                  finish_cb);
 158         int r;
 159
 160         dma_fence_put(f);
 161
 162         /* Wait for all dependencies to avoid data corruptions */
 163         while (!xa_empty(&job->dependencies)) {
 164                 f = xa_erase(&job->dependencies, job->last_dependency++);
 165                 r = dma_fence_add_callback(f, &job->finish_cb,
 166                                            drm_sched_entity_kill_jobs_cb);
 167                 if (!r)
 168                         return;
 169
 170                 dma_fence_put(f);
 171         }
 172
 173         INIT_WORK(&job->work, drm_sched_entity_kill_jobs_work);
 174         schedule_work(&job->work);
 175 }
 176
 177 /* Remove the entity from the scheduler and kill all pending jobs */
 178 static void drm_sched_entity_kill(struct drm_sched_entity *entity)
 179 {
 180         struct drm_sched_job *job;
 181         struct dma_fence *prev;
 182
 183         if (!entity->rq)
 184                 return;
 185
 186         spin_lock(&entity->rq_lock);
 187         entity->stopped = true;
 188         drm_sched_rq_remove_entity(entity->rq, entity);
 189         spin_unlock(&entity->rq_lock);
 190
 191         /* Make sure this entity is not used by the scheduler at the moment */
 192         wait_for_completion(&entity->entity_idle);
 193
 194         prev = dma_fence_get(entity->last_scheduled);
 195         while ((job = to_drm_sched_job(spsc_queue_pop(&entity->job_queue)))) {
 196                 struct drm_sched_fence *s_fence = job->s_fence;
 197
 198                 dma_fence_set_error(&s_fence->finished, -ESRCH);
 199
 200                 dma_fence_get(&s_fence->finished);
 201                 if (!prev || dma_fence_add_callback(prev, &job->finish_cb,
 202                                            drm_sched_entity_kill_jobs_cb))
 203                         drm_sched_entity_kill_jobs_cb(NULL, &job->finish_cb);
 204
 205                 prev = &s_fence->finished;
 206         }
 207         dma_fence_put(prev);
 208 }
 209
 210 /**
 211  * drm_sched_entity_flush - Flush a context entity
 212  *
 213  * @entity: scheduler entity
 214  * @timeout: time to wait in for Q to become empty in jiffies.
 215  *
 216  * Splitting drm_sched_entity_fini() into two functions, The first one does the
 217  * waiting, removes the entity from the runqueue and returns an error when the
 218  * process was killed.
 219  *
 220  * Returns the remaining time in jiffies left from the input timeout
 221  */
 222 long drm_sched_entity_flush(struct drm_sched_entity *entity, long timeout)
 223 {
 224         struct drm_gpu_scheduler *sched;
 225         struct task_struct *last_user;
 226         long ret = timeout;
 227
 228         if (!entity->rq)
 229                 return 0;
 230
 231         sched = entity->rq->sched;
 232         /**
 233          * The client will not queue more IBs during this fini, consume existing
 234          * queued IBs or discard them on SIGKILL
 235          */
 236         if (current->flags & PF_EXITING) {
 237                 if (timeout)
 238                         ret = wait_event_timeout(
 239                                         sched->job_scheduled,
 240                                         drm_sched_entity_is_idle(entity),
 241                                         timeout);
 242         } else {
 243                 wait_event_killable(sched->job_scheduled,
 244                                     drm_sched_entity_is_idle(entity));
 245         }
 246
 247         /* For killed process disable any more IBs enqueue right now */
 248         last_user = cmpxchg(&entity->last_user, current->group_leader, NULL);
 249         if ((!last_user || last_user == current->group_leader) &&
 250             (current->flags & PF_EXITING) && (current->exit_code == SIGKILL))
 251                 drm_sched_entity_kill(entity);
 252
 253         return ret;
 254 }
 255 EXPORT_SYMBOL(drm_sched_entity_flush);
 256
 257 /**
 258  * drm_sched_entity_fini - Destroy a context entity
 259  *
 260  * @entity: scheduler entity
 261  *
 262  * Cleanups up @entity which has been initialized by drm_sched_entity_init().
 263  *
 264  * If there are potentially job still in flight or getting newly queued
 265  * drm_sched_entity_flush() must be called first. This function then goes over
 266  * the entity and signals all jobs with an error code if the process was killed.
 267  */
 268 void drm_sched_entity_fini(struct drm_sched_entity *entity)
 269 {
 270         /*
 271          * If consumption of existing IBs wasn't completed. Forcefully remove
 272          * them here. Also makes sure that the scheduler won't touch this entity
 273          * any more.
 274          */
 275         drm_sched_entity_kill(entity);
 276
 277         if (entity->dependency) {
 278                 dma_fence_remove_callback(entity->dependency, &entity->cb);
 279                 dma_fence_put(entity->dependency);
 280                 entity->dependency = NULL;
 281         }
 282
 283         dma_fence_put(entity->last_scheduled);
 284         entity->last_scheduled = NULL;
 285 }
 286 EXPORT_SYMBOL(drm_sched_entity_fini);
 287
 288 /**
 289  * drm_sched_entity_destroy - Destroy a context entity
 290  * @entity: scheduler entity
 291  *
 292  * Calls drm_sched_entity_flush() and drm_sched_entity_fini() as a
 293  * convenience wrapper.
 294  */
 295 void drm_sched_entity_destroy(struct drm_sched_entity *entity)
 296 {
 297         drm_sched_entity_flush(entity, MAX_WAIT_SCHED_ENTITY_Q_EMPTY);
 298         drm_sched_entity_fini(entity);
 299 }
 300 EXPORT_SYMBOL(drm_sched_entity_destroy);
 301
 302 /* drm_sched_entity_clear_dep - callback to clear the entities dependency */
 303 static void drm_sched_entity_clear_dep(struct dma_fence *f,
 304                                        struct dma_fence_cb *cb)
 305 {
 306         struct drm_sched_entity *entity =
 307                 container_of(cb, struct drm_sched_entity, cb);
 308
 309         entity->dependency = NULL;
 310         dma_fence_put(f);
 311 }
 312
 313 /*
 314  * drm_sched_entity_clear_dep - callback to clear the entities dependency and
 315  * wake up scheduler
 316  */
 317 static void drm_sched_entity_wakeup(struct dma_fence *f,
 318                                     struct dma_fence_cb *cb)
 319 {
 320         struct drm_sched_entity *entity =
 321                 container_of(cb, struct drm_sched_entity, cb);
 322
 323         drm_sched_entity_clear_dep(f, cb);
 324         drm_sched_wakeup(entity->rq->sched);
 325 }
 326
 327 /**
 328  * drm_sched_entity_set_priority - Sets priority of the entity
 329  *
 330  * @entity: scheduler entity
 331  * @priority: scheduler priority
 332  *
 333  * Update the priority of runqueus used for the entity.
 334  */
 335 void drm_sched_entity_set_priority(struct drm_sched_entity *entity,
 336                                    enum drm_sched_priority priority)
 337 {
 338         spin_lock(&entity->rq_lock);
 339         entity->priority = priority;
 340         spin_unlock(&entity->rq_lock);
 341 }
 342 EXPORT_SYMBOL(drm_sched_entity_set_priority);
 343
 344 /*
 345  * Add a callback to the current dependency of the entity to wake up the
 346  * scheduler when the entity becomes available.
 347  */
 348 static bool drm_sched_entity_add_dependency_cb(struct drm_sched_entity *entity)
 349 {
 350         struct drm_gpu_scheduler *sched = entity->rq->sched;
 351         struct dma_fence *fence = entity->dependency;
 352         struct drm_sched_fence *s_fence;
 353
 354         if (fence->context == entity->fence_context ||
 355             fence->context == entity->fence_context + 1) {
 356                 /*
 357                  * Fence is a scheduled/finished fence from a job
 358                  * which belongs to the same entity, we can ignore
 359                  * fences from ourself
 360                  */
 361                 dma_fence_put(entity->dependency);
 362                 return false;
 363         }
 364
 365         s_fence = to_drm_sched_fence(fence);
 366         if (s_fence && s_fence->sched == sched &&
 367             !test_bit(DRM_SCHED_FENCE_DONT_PIPELINE, &fence->flags)) {
 368
 369                 /*
 370                  * Fence is from the same scheduler, only need to wait for
 371                  * it to be scheduled
 372                  */
 373                 fence = dma_fence_get(&s_fence->scheduled);
 374                 dma_fence_put(entity->dependency);
 375                 entity->dependency = fence;
 376                 if (!dma_fence_add_callback(fence, &entity->cb,
 377                                             drm_sched_entity_clear_dep))
 378                         return true;
 379
 380                 /* Ignore it when it is already scheduled */
 381                 dma_fence_put(fence);
 382                 return false;
 383         }
 384
 385         if (!dma_fence_add_callback(entity->dependency, &entity->cb,
 386                                     drm_sched_entity_wakeup))
 387                 return true;
 388
 389         dma_fence_put(entity->dependency);
 390         return false;
 391 }
 392
 393 static struct dma_fence *
 394 drm_sched_job_dependency(struct drm_sched_job *job,
 395                          struct drm_sched_entity *entity)
 396 {
 397         if (!xa_empty(&job->dependencies))
 398                 return xa_erase(&job->dependencies, job->last_dependency++);
 399
 400         if (job->sched->ops->prepare_job)
 401                 return job->sched->ops->prepare_job(job, entity);
 402
 403         return NULL;
 404 }
 405
 406 struct drm_sched_job *drm_sched_entity_pop_job(struct drm_sched_entity *entity)
 407 {
 408         struct drm_sched_job *sched_job;
 409
 410         sched_job = to_drm_sched_job(spsc_queue_peek(&entity->job_queue));
 411         if (!sched_job)
 412                 return NULL;
 413
 414         while ((entity->dependency =
 415                         drm_sched_job_dependency(sched_job, entity))) {
 416                 trace_drm_sched_job_wait_dep(sched_job, entity->dependency);
 417
 418                 if (drm_sched_entity_add_dependency_cb(entity))
 419                         return NULL;
 420         }
 421
 422         /* skip jobs from entity that marked guilty */
 423         if (entity->guilty && atomic_read(entity->guilty))
 424                 dma_fence_set_error(&sched_job->s_fence->finished, -ECANCELED);
 425
 426         dma_fence_put(entity->last_scheduled);
 427
 428         entity->last_scheduled = dma_fence_get(&sched_job->s_fence->finished);
 429
 430         /*
 431          * If the queue is empty we allow drm_sched_entity_select_rq() to
 432          * locklessly access ->last_scheduled. This only works if we set the
 433          * pointer before we dequeue and if we a write barrier here.
 434          */
 435         smp_wmb();
 436
 437         spsc_queue_pop(&entity->job_queue);
 438
 439         /*
 440          * Update the entity's location in the min heap according to
 441          * the timestamp of the next job, if any.
 442          */
 443         if (drm_sched_policy == DRM_SCHED_POLICY_FIFO) {
 444                 struct drm_sched_job *next;
 445
 446                 next = to_drm_sched_job(spsc_queue_peek(&entity->job_queue));
 447                 if (next)
 448                         drm_sched_rq_update_fifo(entity, next->submit_ts);
 449         }
 450
 451         /* Jobs and entities might have different lifecycles. Since we're
 452          * removing the job from the entities queue, set the jobs entity pointer
 453          * to NULL to prevent any future access of the entity through this job.
 454          */
 455         sched_job->entity = NULL;
 456
 457         return sched_job;
 458 }
 459
 460 void drm_sched_entity_select_rq(struct drm_sched_entity *entity)
 461 {
 462         struct dma_fence *fence;
 463         struct drm_gpu_scheduler *sched;
 464         struct drm_sched_rq *rq;
 465
 466         /* single possible engine and already selected */
 467         if (!entity->sched_list)
 468                 return;
 469
 470         /* queue non-empty, stay on the same engine */
 471         if (spsc_queue_count(&entity->job_queue))
 472                 return;
 473
 474         /*
 475          * Only when the queue is empty are we guaranteed that the scheduler
 476          * thread cannot change ->last_scheduled. To enforce ordering we need
 477          * a read barrier here. See drm_sched_entity_pop_job() for the other
 478          * side.
 479          */
 480         smp_rmb();
 481
 482         fence = entity->last_scheduled;
 483
 484         /* stay on the same engine if the previous job hasn't finished */
 485         if (fence && !dma_fence_is_signaled(fence))
 486                 return;
 487
 488         spin_lock(&entity->rq_lock);
 489         sched = drm_sched_pick_best(entity->sched_list, entity->num_sched_list);
 490         rq = sched ? &sched->sched_rq[entity->priority] : NULL;
 491         if (rq != entity->rq) {
 492                 drm_sched_rq_remove_entity(entity->rq, entity);
 493                 entity->rq = rq;
 494         }
 495         spin_unlock(&entity->rq_lock);
 496
 497         if (entity->num_sched_list == 1)
 498                 entity->sched_list = NULL;
 499 }
 500
 501 /**
 502  * drm_sched_entity_push_job - Submit a job to the entity's job queue
 503  * @sched_job: job to submit
 504  *
 505  * Note: To guarantee that the order of insertion to queue matches the job's
 506  * fence sequence number this function should be called with drm_sched_job_arm()
 507  * under common lock for the struct drm_sched_entity that was set up for
 508  * @sched_job in drm_sched_job_init().
 509  *
 510  * Returns 0 for success, negative error code otherwise.
 511  */
 512 void drm_sched_entity_push_job(struct drm_sched_job *sched_job)
 513 {
 514         struct drm_sched_entity *entity = sched_job->entity;
 515         bool first;
 516         ktime_t submit_ts;
 517
 518         trace_drm_sched_job(sched_job, entity);
 519         atomic_inc(entity->rq->sched->score);
 520         WRITE_ONCE(entity->last_user, current->group_leader);
 521
 522         /*
 523          * After the sched_job is pushed into the entity queue, it may be
 524          * completed and freed up at any time. We can no longer access it.
 525          * Make sure to set the submit_ts first, to avoid a race.
 526          */
 527         sched_job->submit_ts = submit_ts = ktime_get();
 528         first = spsc_queue_push(&entity->job_queue, &sched_job->queue_node);
 529
 530         /* first job wakes up scheduler */
 531         if (first) {
 532                 /* Add the entity to the run queue */
 533                 spin_lock(&entity->rq_lock);
 534                 if (entity->stopped) {
 535                         spin_unlock(&entity->rq_lock);
 536
 537                         DRM_ERROR("Trying to push to a killed entity\n");
 538                         return;
 539                 }
 540
 541                 drm_sched_rq_add_entity(entity->rq, entity);
 542                 spin_unlock(&entity->rq_lock);
 543
 544                 if (drm_sched_policy == DRM_SCHED_POLICY_FIFO)
 545                         drm_sched_rq_update_fifo(entity, submit_ts);
 546
 547                 drm_sched_wakeup(entity->rq->sched);
 548         }
 549 }
 550 EXPORT_SYMBOL(drm_sched_entity_push_job);