drivers/gpu/drm/scheduler/sched_entity.c

   1 /*
   2  * Copyright 2015 Advanced Micro Devices, Inc.
   3  *
   4  * Permission is hereby granted, free of charge, to any person obtaining a
   5  * copy of this software and associated documentation files (the "Software"),
   6  * to deal in the Software without restriction, including without limitation
   7  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
   8  * and/or sell copies of the Software, and to permit persons to whom the
   9  * Software is furnished to do so, subject to the following conditions:
  10  *
  11  * The above copyright notice and this permission notice shall be included in
  12  * all copies or substantial portions of the Software.
  13  *
  14  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  15  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  16  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  17  * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
  18  * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
  19  * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
  20  * OTHER DEALINGS IN THE SOFTWARE.
  21  *
  22  */
  23
  24 #include <linux/kthread.h>
  25 #include <linux/slab.h>
  26 #include <linux/completion.h>
  27
  28 #include <drm/drm_print.h>
  29 #include <drm/gpu_scheduler.h>
  30
  31 #include "gpu_scheduler_trace.h"
  32
  33 #define to_drm_sched_job(sched_job)             \
  34                 container_of((sched_job), struct drm_sched_job, queue_node)
  35
  36 /**
  37  * drm_sched_entity_init - Init a context entity used by scheduler when
  38  * submit to HW ring.
  39  *
  40  * @entity: scheduler entity to init
  41  * @priority: priority of the entity
  42  * @sched_list: the list of drm scheds on which jobs from this
  43  *           entity can be submitted
  44  * @num_sched_list: number of drm sched in sched_list
  45  * @guilty: atomic_t set to 1 when a job on this queue
  46  *          is found to be guilty causing a timeout
  47  *
  48  * Note that the &sched_list must have at least one element to schedule the entity.
  49  *
  50  * For changing @priority later on at runtime see
  51  * drm_sched_entity_set_priority(). For changing the set of schedulers
  52  * @sched_list at runtime see drm_sched_entity_modify_sched().
  53  *
  54  * An entity is cleaned up by callind drm_sched_entity_fini(). See also
  55  * drm_sched_entity_destroy().
  56  *
  57  * Returns 0 on success or a negative error code on failure.
  58  */
  59 int drm_sched_entity_init(struct drm_sched_entity *entity,
  60                           enum drm_sched_priority priority,
  61                           struct drm_gpu_scheduler **sched_list,
  62                           unsigned int num_sched_list,
  63                           atomic_t *guilty)
  64 {
  65         if (!(entity && sched_list && (num_sched_list == 0 || sched_list[0])))
  66                 return -EINVAL;
  67
  68         memset(entity, 0, sizeof(struct drm_sched_entity));
  69         INIT_LIST_HEAD(&entity->list);
  70         entity->rq = NULL;
  71         entity->guilty = guilty;
  72         entity->num_sched_list = num_sched_list;
  73         entity->priority = priority;
  74         entity->sched_list = num_sched_list > 1 ? sched_list : NULL;
  75         RCU_INIT_POINTER(entity->last_scheduled, NULL);
  76         RB_CLEAR_NODE(&entity->rb_tree_node);
  77
  78         if(num_sched_list)
  79                 entity->rq = &sched_list[0]->sched_rq[entity->priority];
  80
  81         init_completion(&entity->entity_idle);
  82
  83         /* We start in an idle state. */
  84         complete_all(&entity->entity_idle);
  85
  86         spin_lock_init(&entity->rq_lock);
  87         spsc_queue_init(&entity->job_queue);
  88
  89         atomic_set(&entity->fence_seq, 0);
  90         entity->fence_context = dma_fence_context_alloc(2);
  91
  92         return 0;
  93 }
  94 EXPORT_SYMBOL(drm_sched_entity_init);
  95
  96 /**
  97  * drm_sched_entity_modify_sched - Modify sched of an entity
  98  * @entity: scheduler entity to init
  99  * @sched_list: the list of new drm scheds which will replace
 100  *               existing entity->sched_list
 101  * @num_sched_list: number of drm sched in sched_list
 102  *
 103  * Note that this must be called under the same common lock for @entity as
 104  * drm_sched_job_arm() and drm_sched_entity_push_job(), or the driver needs to
 105  * guarantee through some other means that this is never called while new jobs
 106  * can be pushed to @entity.
 107  */
 108 void drm_sched_entity_modify_sched(struct drm_sched_entity *entity,
 109                                     struct drm_gpu_scheduler **sched_list,
 110                                     unsigned int num_sched_list)
 111 {
 112         WARN_ON(!num_sched_list || !sched_list);
 113
 114         entity->sched_list = sched_list;
 115         entity->num_sched_list = num_sched_list;
 116 }
 117 EXPORT_SYMBOL(drm_sched_entity_modify_sched);
 118
 119 static bool drm_sched_entity_is_idle(struct drm_sched_entity *entity)
 120 {
 121         rmb(); /* for list_empty to work without lock */
 122
 123         if (list_empty(&entity->list) ||
 124             spsc_queue_count(&entity->job_queue) == 0 ||
 125             entity->stopped)
 126                 return true;
 127
 128         return false;
 129 }
 130
 131 /* Return true if entity could provide a job. */
 132 bool drm_sched_entity_is_ready(struct drm_sched_entity *entity)
 133 {
 134         if (spsc_queue_peek(&entity->job_queue) == NULL)
 135                 return false;
 136
 137         if (READ_ONCE(entity->dependency))
 138                 return false;
 139
 140         return true;
 141 }
 142
 143 /**
 144  * drm_sched_entity_error - return error of last scheduled job
 145  * @entity: scheduler entity to check
 146  *
 147  * Opportunistically return the error of the last scheduled job. Result can
 148  * change any time when new jobs are pushed to the hw.
 149  */
 150 int drm_sched_entity_error(struct drm_sched_entity *entity)
 151 {
 152         struct dma_fence *fence;
 153         int r;
 154
 155         rcu_read_lock();
 156         fence = rcu_dereference(entity->last_scheduled);
 157         r = fence ? fence->error : 0;
 158         rcu_read_unlock();
 159
 160         return r;
 161 }
 162 EXPORT_SYMBOL(drm_sched_entity_error);
 163
 164 static void drm_sched_entity_kill_jobs_work(struct work_struct *wrk)
 165 {
 166         struct drm_sched_job *job = container_of(wrk, typeof(*job), work);
 167
 168         drm_sched_fence_finished(job->s_fence, -ESRCH);
 169         WARN_ON(job->s_fence->parent);
 170         job->sched->ops->free_job(job);
 171 }
 172
 173 /* Signal the scheduler finished fence when the entity in question is killed. */
 174 static void drm_sched_entity_kill_jobs_cb(struct dma_fence *f,
 175                                           struct dma_fence_cb *cb)
 176 {
 177         struct drm_sched_job *job = container_of(cb, struct drm_sched_job,
 178                                                  finish_cb);
 179         int r;
 180
 181         dma_fence_put(f);
 182
 183         /* Wait for all dependencies to avoid data corruptions */
 184         while (!xa_empty(&job->dependencies)) {
 185                 f = xa_erase(&job->dependencies, job->last_dependency++);
 186                 r = dma_fence_add_callback(f, &job->finish_cb,
 187                                            drm_sched_entity_kill_jobs_cb);
 188                 if (!r)
 189                         return;
 190
 191                 dma_fence_put(f);
 192         }
 193
 194         INIT_WORK(&job->work, drm_sched_entity_kill_jobs_work);
 195         schedule_work(&job->work);
 196 }
 197
 198 /* Remove the entity from the scheduler and kill all pending jobs */
 199 static void drm_sched_entity_kill(struct drm_sched_entity *entity)
 200 {
 201         struct drm_sched_job *job;
 202         struct dma_fence *prev;
 203
 204         if (!entity->rq)
 205                 return;
 206
 207         spin_lock(&entity->rq_lock);
 208         entity->stopped = true;
 209         drm_sched_rq_remove_entity(entity->rq, entity);
 210         spin_unlock(&entity->rq_lock);
 211
 212         /* Make sure this entity is not used by the scheduler at the moment */
 213         wait_for_completion(&entity->entity_idle);
 214
 215         /* The entity is guaranteed to not be used by the scheduler */
 216         prev = rcu_dereference_check(entity->last_scheduled, true);
 217         dma_fence_get(prev);
 218         while ((job = to_drm_sched_job(spsc_queue_pop(&entity->job_queue)))) {
 219                 struct drm_sched_fence *s_fence = job->s_fence;
 220
 221                 dma_fence_get(&s_fence->finished);
 222                 if (!prev || dma_fence_add_callback(prev, &job->finish_cb,
 223                                            drm_sched_entity_kill_jobs_cb))
 224                         drm_sched_entity_kill_jobs_cb(NULL, &job->finish_cb);
 225
 226                 prev = &s_fence->finished;
 227         }
 228         dma_fence_put(prev);
 229 }
 230
 231 /**
 232  * drm_sched_entity_flush - Flush a context entity
 233  *
 234  * @entity: scheduler entity
 235  * @timeout: time to wait in for Q to become empty in jiffies.
 236  *
 237  * Splitting drm_sched_entity_fini() into two functions, The first one does the
 238  * waiting, removes the entity from the runqueue and returns an error when the
 239  * process was killed.
 240  *
 241  * Returns the remaining time in jiffies left from the input timeout
 242  */
 243 long drm_sched_entity_flush(struct drm_sched_entity *entity, long timeout)
 244 {
 245         struct drm_gpu_scheduler *sched;
 246         struct task_struct *last_user;
 247         long ret = timeout;
 248
 249         if (!entity->rq)
 250                 return 0;
 251
 252         sched = entity->rq->sched;
 253         /**
 254          * The client will not queue more IBs during this fini, consume existing
 255          * queued IBs or discard them on SIGKILL
 256          */
 257         if (current->flags & PF_EXITING) {
 258                 if (timeout)
 259                         ret = wait_event_timeout(
 260                                         sched->job_scheduled,
 261                                         drm_sched_entity_is_idle(entity),
 262                                         timeout);
 263         } else {
 264                 wait_event_killable(sched->job_scheduled,
 265                                     drm_sched_entity_is_idle(entity));
 266         }
 267
 268         /* For killed process disable any more IBs enqueue right now */
 269         last_user = cmpxchg(&entity->last_user, current->group_leader, NULL);
 270         if ((!last_user || last_user == current->group_leader) &&
 271             (current->flags & PF_EXITING) && (current->exit_code == SIGKILL))
 272                 drm_sched_entity_kill(entity);
 273
 274         return ret;
 275 }
 276 EXPORT_SYMBOL(drm_sched_entity_flush);
 277
 278 /**
 279  * drm_sched_entity_fini - Destroy a context entity
 280  *
 281  * @entity: scheduler entity
 282  *
 283  * Cleanups up @entity which has been initialized by drm_sched_entity_init().
 284  *
 285  * If there are potentially job still in flight or getting newly queued
 286  * drm_sched_entity_flush() must be called first. This function then goes over
 287  * the entity and signals all jobs with an error code if the process was killed.
 288  */
 289 void drm_sched_entity_fini(struct drm_sched_entity *entity)
 290 {
 291         /*
 292          * If consumption of existing IBs wasn't completed. Forcefully remove
 293          * them here. Also makes sure that the scheduler won't touch this entity
 294          * any more.
 295          */
 296         drm_sched_entity_kill(entity);
 297
 298         if (entity->dependency) {
 299                 dma_fence_remove_callback(entity->dependency, &entity->cb);
 300                 dma_fence_put(entity->dependency);
 301                 entity->dependency = NULL;
 302         }
 303
 304         dma_fence_put(rcu_dereference_check(entity->last_scheduled, true));
 305         RCU_INIT_POINTER(entity->last_scheduled, NULL);
 306 }
 307 EXPORT_SYMBOL(drm_sched_entity_fini);
 308
 309 /**
 310  * drm_sched_entity_destroy - Destroy a context entity
 311  * @entity: scheduler entity
 312  *
 313  * Calls drm_sched_entity_flush() and drm_sched_entity_fini() as a
 314  * convenience wrapper.
 315  */
 316 void drm_sched_entity_destroy(struct drm_sched_entity *entity)
 317 {
 318         drm_sched_entity_flush(entity, MAX_WAIT_SCHED_ENTITY_Q_EMPTY);
 319         drm_sched_entity_fini(entity);
 320 }
 321 EXPORT_SYMBOL(drm_sched_entity_destroy);
 322
 323 /* drm_sched_entity_clear_dep - callback to clear the entities dependency */
 324 static void drm_sched_entity_clear_dep(struct dma_fence *f,
 325                                        struct dma_fence_cb *cb)
 326 {
 327         struct drm_sched_entity *entity =
 328                 container_of(cb, struct drm_sched_entity, cb);
 329
 330         entity->dependency = NULL;
 331         dma_fence_put(f);
 332 }
 333
 334 /*
 335  * drm_sched_entity_clear_dep - callback to clear the entities dependency and
 336  * wake up scheduler
 337  */
 338 static void drm_sched_entity_wakeup(struct dma_fence *f,
 339                                     struct dma_fence_cb *cb)
 340 {
 341         struct drm_sched_entity *entity =
 342                 container_of(cb, struct drm_sched_entity, cb);
 343
 344         drm_sched_entity_clear_dep(f, cb);
 345         drm_sched_wakeup_if_can_queue(entity->rq->sched);
 346 }
 347
 348 /**
 349  * drm_sched_entity_set_priority - Sets priority of the entity
 350  *
 351  * @entity: scheduler entity
 352  * @priority: scheduler priority
 353  *
 354  * Update the priority of runqueus used for the entity.
 355  */
 356 void drm_sched_entity_set_priority(struct drm_sched_entity *entity,
 357                                    enum drm_sched_priority priority)
 358 {
 359         spin_lock(&entity->rq_lock);
 360         entity->priority = priority;
 361         spin_unlock(&entity->rq_lock);
 362 }
 363 EXPORT_SYMBOL(drm_sched_entity_set_priority);
 364
 365 /*
 366  * Add a callback to the current dependency of the entity to wake up the
 367  * scheduler when the entity becomes available.
 368  */
 369 static bool drm_sched_entity_add_dependency_cb(struct drm_sched_entity *entity)
 370 {
 371         struct drm_gpu_scheduler *sched = entity->rq->sched;
 372         struct dma_fence *fence = entity->dependency;
 373         struct drm_sched_fence *s_fence;
 374
 375         if (fence->context == entity->fence_context ||
 376             fence->context == entity->fence_context + 1) {
 377                 /*
 378                  * Fence is a scheduled/finished fence from a job
 379                  * which belongs to the same entity, we can ignore
 380                  * fences from ourself
 381                  */
 382                 dma_fence_put(entity->dependency);
 383                 return false;
 384         }
 385
 386         s_fence = to_drm_sched_fence(fence);
 387         if (s_fence && s_fence->sched == sched &&
 388             !test_bit(DRM_SCHED_FENCE_DONT_PIPELINE, &fence->flags)) {
 389
 390                 /*
 391                  * Fence is from the same scheduler, only need to wait for
 392                  * it to be scheduled
 393                  */
 394                 fence = dma_fence_get(&s_fence->scheduled);
 395                 dma_fence_put(entity->dependency);
 396                 entity->dependency = fence;
 397                 if (!dma_fence_add_callback(fence, &entity->cb,
 398                                             drm_sched_entity_clear_dep))
 399                         return true;
 400
 401                 /* Ignore it when it is already scheduled */
 402                 dma_fence_put(fence);
 403                 return false;
 404         }
 405
 406         if (!dma_fence_add_callback(entity->dependency, &entity->cb,
 407                                     drm_sched_entity_wakeup))
 408                 return true;
 409
 410         dma_fence_put(entity->dependency);
 411         return false;
 412 }
 413
 414 static struct dma_fence *
 415 drm_sched_job_dependency(struct drm_sched_job *job,
 416                          struct drm_sched_entity *entity)
 417 {
 418         if (!xa_empty(&job->dependencies))
 419                 return xa_erase(&job->dependencies, job->last_dependency++);
 420
 421         if (job->sched->ops->prepare_job)
 422                 return job->sched->ops->prepare_job(job, entity);
 423
 424         return NULL;
 425 }
 426
 427 struct drm_sched_job *drm_sched_entity_pop_job(struct drm_sched_entity *entity)
 428 {
 429         struct drm_sched_job *sched_job;
 430
 431         sched_job = to_drm_sched_job(spsc_queue_peek(&entity->job_queue));
 432         if (!sched_job)
 433                 return NULL;
 434
 435         while ((entity->dependency =
 436                         drm_sched_job_dependency(sched_job, entity))) {
 437                 trace_drm_sched_job_wait_dep(sched_job, entity->dependency);
 438
 439                 if (drm_sched_entity_add_dependency_cb(entity))
 440                         return NULL;
 441         }
 442
 443         /* skip jobs from entity that marked guilty */
 444         if (entity->guilty && atomic_read(entity->guilty))
 445                 dma_fence_set_error(&sched_job->s_fence->finished, -ECANCELED);
 446
 447         dma_fence_put(rcu_dereference_check(entity->last_scheduled, true));
 448         rcu_assign_pointer(entity->last_scheduled,
 449                            dma_fence_get(&sched_job->s_fence->finished));
 450
 451         /*
 452          * If the queue is empty we allow drm_sched_entity_select_rq() to
 453          * locklessly access ->last_scheduled. This only works if we set the
 454          * pointer before we dequeue and if we a write barrier here.
 455          */
 456         smp_wmb();
 457
 458         spsc_queue_pop(&entity->job_queue);
 459
 460         /*
 461          * Update the entity's location in the min heap according to
 462          * the timestamp of the next job, if any.
 463          */
 464         if (drm_sched_policy == DRM_SCHED_POLICY_FIFO) {
 465                 struct drm_sched_job *next;
 466
 467                 next = to_drm_sched_job(spsc_queue_peek(&entity->job_queue));
 468                 if (next)
 469                         drm_sched_rq_update_fifo(entity, next->submit_ts);
 470         }
 471
 472         /* Jobs and entities might have different lifecycles. Since we're
 473          * removing the job from the entities queue, set the jobs entity pointer
 474          * to NULL to prevent any future access of the entity through this job.
 475          */
 476         sched_job->entity = NULL;
 477
 478         return sched_job;
 479 }
 480
 481 void drm_sched_entity_select_rq(struct drm_sched_entity *entity)
 482 {
 483         struct dma_fence *fence;
 484         struct drm_gpu_scheduler *sched;
 485         struct drm_sched_rq *rq;
 486
 487         /* single possible engine and already selected */
 488         if (!entity->sched_list)
 489                 return;
 490
 491         /* queue non-empty, stay on the same engine */
 492         if (spsc_queue_count(&entity->job_queue))
 493                 return;
 494
 495         /*
 496          * Only when the queue is empty are we guaranteed that the scheduler
 497          * thread cannot change ->last_scheduled. To enforce ordering we need
 498          * a read barrier here. See drm_sched_entity_pop_job() for the other
 499          * side.
 500          */
 501         smp_rmb();
 502
 503         fence = rcu_dereference_check(entity->last_scheduled, true);
 504
 505         /* stay on the same engine if the previous job hasn't finished */
 506         if (fence && !dma_fence_is_signaled(fence))
 507                 return;
 508
 509         spin_lock(&entity->rq_lock);
 510         sched = drm_sched_pick_best(entity->sched_list, entity->num_sched_list);
 511         rq = sched ? &sched->sched_rq[entity->priority] : NULL;
 512         if (rq != entity->rq) {
 513                 drm_sched_rq_remove_entity(entity->rq, entity);
 514                 entity->rq = rq;
 515         }
 516         spin_unlock(&entity->rq_lock);
 517
 518         if (entity->num_sched_list == 1)
 519                 entity->sched_list = NULL;
 520 }
 521
 522 /**
 523  * drm_sched_entity_push_job - Submit a job to the entity's job queue
 524  * @sched_job: job to submit
 525  *
 526  * Note: To guarantee that the order of insertion to queue matches the job's
 527  * fence sequence number this function should be called with drm_sched_job_arm()
 528  * under common lock for the struct drm_sched_entity that was set up for
 529  * @sched_job in drm_sched_job_init().
 530  *
 531  * Returns 0 for success, negative error code otherwise.
 532  */
 533 void drm_sched_entity_push_job(struct drm_sched_job *sched_job)
 534 {
 535         struct drm_sched_entity *entity = sched_job->entity;
 536         bool first;
 537         ktime_t submit_ts;
 538
 539         trace_drm_sched_job(sched_job, entity);
 540         atomic_inc(entity->rq->sched->score);
 541         WRITE_ONCE(entity->last_user, current->group_leader);
 542
 543         /*
 544          * After the sched_job is pushed into the entity queue, it may be
 545          * completed and freed up at any time. We can no longer access it.
 546          * Make sure to set the submit_ts first, to avoid a race.
 547          */
 548         sched_job->submit_ts = submit_ts = ktime_get();
 549         first = spsc_queue_push(&entity->job_queue, &sched_job->queue_node);
 550
 551         /* first job wakes up scheduler */
 552         if (first) {
 553                 /* Add the entity to the run queue */
 554                 spin_lock(&entity->rq_lock);
 555                 if (entity->stopped) {
 556                         spin_unlock(&entity->rq_lock);
 557
 558                         DRM_ERROR("Trying to push to a killed entity\n");
 559                         return;
 560                 }
 561
 562                 drm_sched_rq_add_entity(entity->rq, entity);
 563                 spin_unlock(&entity->rq_lock);
 564
 565                 if (drm_sched_policy == DRM_SCHED_POLICY_FIFO)
 566                         drm_sched_rq_update_fifo(entity, submit_ts);
 567
 568                 drm_sched_wakeup_if_can_queue(entity->rq->sched);
 569         }
 570 }
 571 EXPORT_SYMBOL(drm_sched_entity_push_job);