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