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: the sched_list should have at least one element to schedule
  49  *       the entity
  50  *
  51  * Returns 0 on success or a negative error code on failure.
  52  */
  53 int drm_sched_entity_init(struct drm_sched_entity *entity,
  54                           enum drm_sched_priority priority,
  55                           struct drm_gpu_scheduler **sched_list,
  56                           unsigned int num_sched_list,
  57                           atomic_t *guilty)
  58 {
  59         if (!(entity && sched_list && (num_sched_list == 0 || sched_list[0])))
  60                 return -EINVAL;
  61
  62         memset(entity, 0, sizeof(struct drm_sched_entity));
  63         INIT_LIST_HEAD(&entity->list);
  64         entity->rq = NULL;
  65         entity->guilty = guilty;
  66         entity->num_sched_list = num_sched_list;
  67         entity->priority = priority;
  68         entity->sched_list = num_sched_list > 1 ? sched_list : NULL;
  69         entity->last_scheduled = NULL;
  70
  71         if(num_sched_list)
  72                 entity->rq = &sched_list[0]->sched_rq[entity->priority];
  73
  74         init_completion(&entity->entity_idle);
  75
  76         spin_lock_init(&entity->rq_lock);
  77         spsc_queue_init(&entity->job_queue);
  78
  79         atomic_set(&entity->fence_seq, 0);
  80         entity->fence_context = dma_fence_context_alloc(2);
  81
  82         return 0;
  83 }
  84 EXPORT_SYMBOL(drm_sched_entity_init);
  85
  86 /**
  87  * drm_sched_entity_modify_sched - Modify sched of an entity
  88  * @entity: scheduler entity to init
  89  * @sched_list: the list of new drm scheds which will replace
  90  *               existing entity->sched_list
  91  * @num_sched_list: number of drm sched in sched_list
  92  */
  93 void drm_sched_entity_modify_sched(struct drm_sched_entity *entity,
  94                                     struct drm_gpu_scheduler **sched_list,
  95                                     unsigned int num_sched_list)
  96 {
  97         WARN_ON(!num_sched_list || !sched_list);
  98
  99         entity->sched_list = sched_list;
 100         entity->num_sched_list = num_sched_list;
 101 }
 102 EXPORT_SYMBOL(drm_sched_entity_modify_sched);
 103
 104 /**
 105  * drm_sched_entity_is_idle - Check if entity is idle
 106  *
 107  * @entity: scheduler entity
 108  *
 109  * Returns true if the entity does not have any unscheduled jobs.
 110  */
 111 static bool drm_sched_entity_is_idle(struct drm_sched_entity *entity)
 112 {
 113         rmb(); /* for list_empty to work without lock */
 114
 115         if (list_empty(&entity->list) ||
 116             spsc_queue_count(&entity->job_queue) == 0 ||
 117             entity->stopped)
 118                 return true;
 119
 120         return false;
 121 }
 122
 123 /**
 124  * drm_sched_entity_is_ready - Check if entity is ready
 125  *
 126  * @entity: scheduler entity
 127  *
 128  * Return true if entity could provide a job.
 129  */
 130 bool drm_sched_entity_is_ready(struct drm_sched_entity *entity)
 131 {
 132         if (spsc_queue_peek(&entity->job_queue) == NULL)
 133                 return false;
 134
 135         if (READ_ONCE(entity->dependency))
 136                 return false;
 137
 138         return true;
 139 }
 140
 141 /**
 142  * drm_sched_entity_flush - Flush a context entity
 143  *
 144  * @entity: scheduler entity
 145  * @timeout: time to wait in for Q to become empty in jiffies.
 146  *
 147  * Splitting drm_sched_entity_fini() into two functions, The first one does the
 148  * waiting, removes the entity from the runqueue and returns an error when the
 149  * process was killed.
 150  *
 151  * Returns the remaining time in jiffies left from the input timeout
 152  */
 153 long drm_sched_entity_flush(struct drm_sched_entity *entity, long timeout)
 154 {
 155         struct drm_gpu_scheduler *sched;
 156         struct task_struct *last_user;
 157         long ret = timeout;
 158
 159         if (!entity->rq)
 160                 return 0;
 161
 162         sched = entity->rq->sched;
 163         /**
 164          * The client will not queue more IBs during this fini, consume existing
 165          * queued IBs or discard them on SIGKILL
 166          */
 167         if (current->flags & PF_EXITING) {
 168                 if (timeout)
 169                         ret = wait_event_timeout(
 170                                         sched->job_scheduled,
 171                                         drm_sched_entity_is_idle(entity),
 172                                         timeout);
 173         } else {
 174                 wait_event_killable(sched->job_scheduled,
 175                                     drm_sched_entity_is_idle(entity));
 176         }
 177
 178         /* For killed process disable any more IBs enqueue right now */
 179         last_user = cmpxchg(&entity->last_user, current->group_leader, NULL);
 180         if ((!last_user || last_user == current->group_leader) &&
 181             (current->flags & PF_EXITING) && (current->exit_code == SIGKILL)) {
 182                 spin_lock(&entity->rq_lock);
 183                 entity->stopped = true;
 184                 drm_sched_rq_remove_entity(entity->rq, entity);
 185                 spin_unlock(&entity->rq_lock);
 186         }
 187
 188         return ret;
 189 }
 190 EXPORT_SYMBOL(drm_sched_entity_flush);
 191
 192 /**
 193  * drm_sched_entity_kill_jobs - helper for drm_sched_entity_kill_jobs
 194  *
 195  * @f: signaled fence
 196  * @cb: our callback structure
 197  *
 198  * Signal the scheduler finished fence when the entity in question is killed.
 199  */
 200 static void drm_sched_entity_kill_jobs_cb(struct dma_fence *f,
 201                                           struct dma_fence_cb *cb)
 202 {
 203         struct drm_sched_job *job = container_of(cb, struct drm_sched_job,
 204                                                  finish_cb);
 205
 206         drm_sched_fence_finished(job->s_fence);
 207         WARN_ON(job->s_fence->parent);
 208         job->sched->ops->free_job(job);
 209 }
 210
 211 /**
 212  * drm_sched_entity_kill_jobs - Make sure all remaining jobs are killed
 213  *
 214  * @entity: entity which is cleaned up
 215  *
 216  * Makes sure that all remaining jobs in an entity are killed before it is
 217  * destroyed.
 218  */
 219 static void drm_sched_entity_kill_jobs(struct drm_sched_entity *entity)
 220 {
 221         struct drm_sched_job *job;
 222         struct dma_fence *f;
 223         int r;
 224
 225         while ((job = to_drm_sched_job(spsc_queue_pop(&entity->job_queue)))) {
 226                 struct drm_sched_fence *s_fence = job->s_fence;
 227
 228                 /* Wait for all dependencies to avoid data corruptions */
 229                 while ((f = job->sched->ops->dependency(job, entity)))
 230                         dma_fence_wait(f, false);
 231
 232                 drm_sched_fence_scheduled(s_fence);
 233                 dma_fence_set_error(&s_fence->finished, -ESRCH);
 234
 235                 /*
 236                  * When pipe is hanged by older entity, new entity might
 237                  * not even have chance to submit it's first job to HW
 238                  * and so entity->last_scheduled will remain NULL
 239                  */
 240                 if (!entity->last_scheduled) {
 241                         drm_sched_entity_kill_jobs_cb(NULL, &job->finish_cb);
 242                         continue;
 243                 }
 244
 245                 r = dma_fence_add_callback(entity->last_scheduled,
 246                                            &job->finish_cb,
 247                                            drm_sched_entity_kill_jobs_cb);
 248                 if (r == -ENOENT)
 249                         drm_sched_entity_kill_jobs_cb(NULL, &job->finish_cb);
 250                 else if (r)
 251                         DRM_ERROR("fence add callback failed (%d)\n", r);
 252         }
 253 }
 254
 255 /**
 256  * drm_sched_entity_cleanup - Destroy a context entity
 257  *
 258  * @entity: scheduler entity
 259  *
 260  * This should be called after @drm_sched_entity_do_release. It goes over the
 261  * entity and signals all jobs with an error code if the process was killed.
 262  *
 263  */
 264 void drm_sched_entity_fini(struct drm_sched_entity *entity)
 265 {
 266         struct drm_gpu_scheduler *sched = NULL;
 267
 268         if (entity->rq) {
 269                 sched = entity->rq->sched;
 270                 drm_sched_rq_remove_entity(entity->rq, entity);
 271         }
 272
 273         /* Consumption of existing IBs wasn't completed. Forcefully
 274          * remove them here.
 275          */
 276         if (spsc_queue_count(&entity->job_queue)) {
 277                 if (sched) {
 278                         /*
 279                          * Wait for thread to idle to make sure it isn't processing
 280                          * this entity.
 281                          */
 282                         wait_for_completion(&entity->entity_idle);
 283
 284                 }
 285                 if (entity->dependency) {
 286                         dma_fence_remove_callback(entity->dependency,
 287                                                   &entity->cb);
 288                         dma_fence_put(entity->dependency);
 289                         entity->dependency = NULL;
 290                 }
 291
 292                 drm_sched_entity_kill_jobs(entity);
 293         }
 294
 295         dma_fence_put(entity->last_scheduled);
 296         entity->last_scheduled = NULL;
 297 }
 298 EXPORT_SYMBOL(drm_sched_entity_fini);
 299
 300 /**
 301  * drm_sched_entity_fini - Destroy a context entity
 302  *
 303  * @entity: scheduler entity
 304  *
 305  * Calls drm_sched_entity_do_release() and drm_sched_entity_cleanup()
 306  */
 307 void drm_sched_entity_destroy(struct drm_sched_entity *entity)
 308 {
 309         drm_sched_entity_flush(entity, MAX_WAIT_SCHED_ENTITY_Q_EMPTY);
 310         drm_sched_entity_fini(entity);
 311 }
 312 EXPORT_SYMBOL(drm_sched_entity_destroy);
 313
 314 /**
 315  * drm_sched_entity_clear_dep - callback to clear the entities dependency
 316  */
 317 static void drm_sched_entity_clear_dep(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         entity->dependency = NULL;
 324         dma_fence_put(f);
 325 }
 326
 327 /**
 328  * drm_sched_entity_clear_dep - callback to clear the entities dependency and
 329  * wake up scheduler
 330  */
 331 static void drm_sched_entity_wakeup(struct dma_fence *f,
 332                                     struct dma_fence_cb *cb)
 333 {
 334         struct drm_sched_entity *entity =
 335                 container_of(cb, struct drm_sched_entity, cb);
 336
 337         drm_sched_entity_clear_dep(f, cb);
 338         drm_sched_wakeup(entity->rq->sched);
 339 }
 340
 341 /**
 342  * drm_sched_entity_set_priority - Sets priority of the entity
 343  *
 344  * @entity: scheduler entity
 345  * @priority: scheduler priority
 346  *
 347  * Update the priority of runqueus used for the entity.
 348  */
 349 void drm_sched_entity_set_priority(struct drm_sched_entity *entity,
 350                                    enum drm_sched_priority priority)
 351 {
 352         spin_lock(&entity->rq_lock);
 353         entity->priority = priority;
 354         spin_unlock(&entity->rq_lock);
 355 }
 356 EXPORT_SYMBOL(drm_sched_entity_set_priority);
 357
 358 /**
 359  * drm_sched_entity_add_dependency_cb - add callback for the entities dependency
 360  *
 361  * @entity: entity with dependency
 362  *
 363  * Add a callback to the current dependency of the entity to wake up the
 364  * scheduler when the entity becomes available.
 365  */
 366 static bool drm_sched_entity_add_dependency_cb(struct drm_sched_entity *entity)
 367 {
 368         struct drm_gpu_scheduler *sched = entity->rq->sched;
 369         struct dma_fence *fence = entity->dependency;
 370         struct drm_sched_fence *s_fence;
 371
 372         if (fence->context == entity->fence_context ||
 373             fence->context == entity->fence_context + 1) {
 374                 /*
 375                  * Fence is a scheduled/finished fence from a job
 376                  * which belongs to the same entity, we can ignore
 377                  * fences from ourself
 378                  */
 379                 dma_fence_put(entity->dependency);
 380                 return false;
 381         }
 382
 383         s_fence = to_drm_sched_fence(fence);
 384         if (s_fence && s_fence->sched == sched) {
 385
 386                 /*
 387                  * Fence is from the same scheduler, only need to wait for
 388                  * it to be scheduled
 389                  */
 390                 fence = dma_fence_get(&s_fence->scheduled);
 391                 dma_fence_put(entity->dependency);
 392                 entity->dependency = fence;
 393                 if (!dma_fence_add_callback(fence, &entity->cb,
 394                                             drm_sched_entity_clear_dep))
 395                         return true;
 396
 397                 /* Ignore it when it is already scheduled */
 398                 dma_fence_put(fence);
 399                 return false;
 400         }
 401
 402         if (!dma_fence_add_callback(entity->dependency, &entity->cb,
 403                                     drm_sched_entity_wakeup))
 404                 return true;
 405
 406         dma_fence_put(entity->dependency);
 407         return false;
 408 }
 409
 410 /**
 411  * drm_sched_entity_pop_job - get a ready to be scheduled job from the entity
 412  *
 413  * @entity: entity to get the job from
 414  *
 415  * Process all dependencies and try to get one job from the entities queue.
 416  */
 417 struct drm_sched_job *drm_sched_entity_pop_job(struct drm_sched_entity *entity)
 418 {
 419         struct drm_gpu_scheduler *sched = entity->rq->sched;
 420         struct drm_sched_job *sched_job;
 421
 422         sched_job = to_drm_sched_job(spsc_queue_peek(&entity->job_queue));
 423         if (!sched_job)
 424                 return NULL;
 425
 426         while ((entity->dependency =
 427                         sched->ops->dependency(sched_job, entity))) {
 428                 trace_drm_sched_job_wait_dep(sched_job, entity->dependency);
 429
 430                 if (drm_sched_entity_add_dependency_cb(entity))
 431                         return NULL;
 432         }
 433
 434         /* skip jobs from entity that marked guilty */
 435         if (entity->guilty && atomic_read(entity->guilty))
 436                 dma_fence_set_error(&sched_job->s_fence->finished, -ECANCELED);
 437
 438         dma_fence_put(entity->last_scheduled);
 439         entity->last_scheduled = dma_fence_get(&sched_job->s_fence->finished);
 440
 441         spsc_queue_pop(&entity->job_queue);
 442         return sched_job;
 443 }
 444
 445 /**
 446  * drm_sched_entity_select_rq - select a new rq for the entity
 447  *
 448  * @entity: scheduler entity
 449  *
 450  * Check all prerequisites and select a new rq for the entity for load
 451  * balancing.
 452  */
 453 void drm_sched_entity_select_rq(struct drm_sched_entity *entity)
 454 {
 455         struct dma_fence *fence;
 456         struct drm_gpu_scheduler *sched;
 457         struct drm_sched_rq *rq;
 458
 459         if (spsc_queue_count(&entity->job_queue) || entity->num_sched_list <= 1)
 460                 return;
 461
 462         fence = READ_ONCE(entity->last_scheduled);
 463         if (fence && !dma_fence_is_signaled(fence))
 464                 return;
 465
 466         spin_lock(&entity->rq_lock);
 467         sched = drm_sched_pick_best(entity->sched_list, entity->num_sched_list);
 468         rq = sched ? &sched->sched_rq[entity->priority] : NULL;
 469         if (rq != entity->rq) {
 470                 drm_sched_rq_remove_entity(entity->rq, entity);
 471                 entity->rq = rq;
 472         }
 473
 474         spin_unlock(&entity->rq_lock);
 475 }
 476
 477 /**
 478  * drm_sched_entity_push_job - Submit a job to the entity's job queue
 479  *
 480  * @sched_job: job to submit
 481  * @entity: scheduler entity
 482  *
 483  * Note: To guarantee that the order of insertion to queue matches
 484  * the job's fence sequence number this function should be
 485  * called with drm_sched_job_init under common lock.
 486  *
 487  * Returns 0 for success, negative error code otherwise.
 488  */
 489 void drm_sched_entity_push_job(struct drm_sched_job *sched_job,
 490                                struct drm_sched_entity *entity)
 491 {
 492         bool first;
 493
 494         trace_drm_sched_job(sched_job, entity);
 495         atomic_inc(&entity->rq->sched->score);
 496         WRITE_ONCE(entity->last_user, current->group_leader);
 497         first = spsc_queue_push(&entity->job_queue, &sched_job->queue_node);
 498
 499         /* first job wakes up scheduler */
 500         if (first) {
 501                 /* Add the entity to the run queue */
 502                 spin_lock(&entity->rq_lock);
 503                 if (entity->stopped) {
 504                         spin_unlock(&entity->rq_lock);
 505
 506                         DRM_ERROR("Trying to push to a killed entity\n");
 507                         return;
 508                 }
 509                 drm_sched_rq_add_entity(entity->rq, entity);
 510                 spin_unlock(&entity->rq_lock);
 511                 drm_sched_wakeup(entity->rq->sched);
 512         }
 513 }
 514 EXPORT_SYMBOL(drm_sched_entity_push_job);