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         /* We start in an idle state. */
  77         complete(&entity->entity_idle);
  78
  79         spin_lock_init(&entity->rq_lock);
  80         spsc_queue_init(&entity->job_queue);
  81
  82         atomic_set(&entity->fence_seq, 0);
  83         entity->fence_context = dma_fence_context_alloc(2);
  84
  85         return 0;
  86 }
  87 EXPORT_SYMBOL(drm_sched_entity_init);
  88
  89 /**
  90  * drm_sched_entity_modify_sched - Modify sched of an entity
  91  * @entity: scheduler entity to init
  92  * @sched_list: the list of new drm scheds which will replace
  93  *               existing entity->sched_list
  94  * @num_sched_list: number of drm sched in sched_list
  95  */
  96 void drm_sched_entity_modify_sched(struct drm_sched_entity *entity,
  97                                     struct drm_gpu_scheduler **sched_list,
  98                                     unsigned int num_sched_list)
  99 {
 100         WARN_ON(!num_sched_list || !sched_list);
 101
 102         entity->sched_list = sched_list;
 103         entity->num_sched_list = num_sched_list;
 104 }
 105 EXPORT_SYMBOL(drm_sched_entity_modify_sched);
 106
 107 /**
 108  * drm_sched_entity_is_idle - Check if entity is idle
 109  *
 110  * @entity: scheduler entity
 111  *
 112  * Returns true if the entity does not have any unscheduled jobs.
 113  */
 114 static bool drm_sched_entity_is_idle(struct drm_sched_entity *entity)
 115 {
 116         rmb(); /* for list_empty to work without lock */
 117
 118         if (list_empty(&entity->list) ||
 119             spsc_queue_count(&entity->job_queue) == 0)
 120                 return true;
 121
 122         return false;
 123 }
 124
 125 /**
 126  * drm_sched_entity_is_ready - Check if entity is ready
 127  *
 128  * @entity: scheduler entity
 129  *
 130  * Return true if entity could provide a job.
 131  */
 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_flush - Flush a context entity
 145  *
 146  * @entity: scheduler entity
 147  * @timeout: time to wait in for Q to become empty in jiffies.
 148  *
 149  * Splitting drm_sched_entity_fini() into two functions, The first one does the
 150  * waiting, removes the entity from the runqueue and returns an error when the
 151  * process was killed.
 152  *
 153  * Returns the remaining time in jiffies left from the input timeout
 154  */
 155 long drm_sched_entity_flush(struct drm_sched_entity *entity, long timeout)
 156 {
 157         struct drm_gpu_scheduler *sched;
 158         struct task_struct *last_user;
 159         long ret = timeout;
 160
 161         if (!entity->rq)
 162                 return 0;
 163
 164         sched = entity->rq->sched;
 165         /**
 166          * The client will not queue more IBs during this fini, consume existing
 167          * queued IBs or discard them on SIGKILL
 168          */
 169         if (current->flags & PF_EXITING) {
 170                 if (timeout)
 171                         ret = wait_event_timeout(
 172                                         sched->job_scheduled,
 173                                         drm_sched_entity_is_idle(entity),
 174                                         timeout);
 175         } else {
 176                 wait_event_killable(sched->job_scheduled,
 177                                     drm_sched_entity_is_idle(entity));
 178         }
 179
 180         /* For killed process disable any more IBs enqueue right now */
 181         last_user = cmpxchg(&entity->last_user, current->group_leader, NULL);
 182         if ((!last_user || last_user == current->group_leader) &&
 183             (current->flags & PF_EXITING) && (current->exit_code == SIGKILL)) {
 184                 spin_lock(&entity->rq_lock);
 185                 entity->stopped = true;
 186                 drm_sched_rq_remove_entity(entity->rq, entity);
 187                 spin_unlock(&entity->rq_lock);
 188         }
 189
 190         return ret;
 191 }
 192 EXPORT_SYMBOL(drm_sched_entity_flush);
 193
 194 /**
 195  * drm_sched_entity_kill_jobs - helper for drm_sched_entity_kill_jobs
 196  *
 197  * @f: signaled fence
 198  * @cb: our callback structure
 199  *
 200  * Signal the scheduler finished fence when the entity in question is killed.
 201  */
 202 static void drm_sched_entity_kill_jobs_cb(struct dma_fence *f,
 203                                           struct dma_fence_cb *cb)
 204 {
 205         struct drm_sched_job *job = container_of(cb, struct drm_sched_job,
 206                                                  finish_cb);
 207
 208         drm_sched_fence_finished(job->s_fence);
 209         WARN_ON(job->s_fence->parent);
 210         job->sched->ops->free_job(job);
 211 }
 212
 213 /**
 214  * drm_sched_entity_kill_jobs - Make sure all remaining jobs are killed
 215  *
 216  * @entity: entity which is cleaned up
 217  *
 218  * Makes sure that all remaining jobs in an entity are killed before it is
 219  * destroyed.
 220  */
 221 static void drm_sched_entity_kill_jobs(struct drm_sched_entity *entity)
 222 {
 223         struct drm_sched_job *job;
 224         int r;
 225
 226         while ((job = to_drm_sched_job(spsc_queue_pop(&entity->job_queue)))) {
 227                 struct drm_sched_fence *s_fence = job->s_fence;
 228
 229                 drm_sched_fence_scheduled(s_fence);
 230                 dma_fence_set_error(&s_fence->finished, -ESRCH);
 231
 232                 /*
 233                  * When pipe is hanged by older entity, new entity might
 234                  * not even have chance to submit it's first job to HW
 235                  * and so entity->last_scheduled will remain NULL
 236                  */
 237                 if (!entity->last_scheduled) {
 238                         drm_sched_entity_kill_jobs_cb(NULL, &job->finish_cb);
 239                         continue;
 240                 }
 241
 242                 r = dma_fence_add_callback(entity->last_scheduled,
 243                                            &job->finish_cb,
 244                                            drm_sched_entity_kill_jobs_cb);
 245                 if (r == -ENOENT)
 246                         drm_sched_entity_kill_jobs_cb(NULL, &job->finish_cb);
 247                 else if (r)
 248                         DRM_ERROR("fence add callback failed (%d)\n", r);
 249         }
 250 }
 251
 252 /**
 253  * drm_sched_entity_cleanup - Destroy a context entity
 254  *
 255  * @entity: scheduler entity
 256  *
 257  * This should be called after @drm_sched_entity_do_release. It goes over the
 258  * entity and signals all jobs with an error code if the process was killed.
 259  *
 260  */
 261 void drm_sched_entity_fini(struct drm_sched_entity *entity)
 262 {
 263         struct drm_gpu_scheduler *sched = NULL;
 264
 265         if (entity->rq) {
 266                 sched = entity->rq->sched;
 267                 drm_sched_rq_remove_entity(entity->rq, entity);
 268         }
 269
 270         /* Consumption of existing IBs wasn't completed. Forcefully
 271          * remove them here.
 272          */
 273         if (spsc_queue_count(&entity->job_queue)) {
 274                 if (sched) {
 275                         /*
 276                          * Wait for thread to idle to make sure it isn't processing
 277                          * this entity.
 278                          */
 279                         wait_for_completion(&entity->entity_idle);
 280
 281                 }
 282                 if (entity->dependency) {
 283                         dma_fence_remove_callback(entity->dependency,
 284                                                   &entity->cb);
 285                         dma_fence_put(entity->dependency);
 286                         entity->dependency = NULL;
 287                 }
 288
 289                 drm_sched_entity_kill_jobs(entity);
 290         }
 291
 292         dma_fence_put(entity->last_scheduled);
 293         entity->last_scheduled = NULL;
 294 }
 295 EXPORT_SYMBOL(drm_sched_entity_fini);
 296
 297 /**
 298  * drm_sched_entity_fini - Destroy a context entity
 299  *
 300  * @entity: scheduler entity
 301  *
 302  * Calls drm_sched_entity_do_release() and drm_sched_entity_cleanup()
 303  */
 304 void drm_sched_entity_destroy(struct drm_sched_entity *entity)
 305 {
 306         drm_sched_entity_flush(entity, MAX_WAIT_SCHED_ENTITY_Q_EMPTY);
 307         drm_sched_entity_fini(entity);
 308 }
 309 EXPORT_SYMBOL(drm_sched_entity_destroy);
 310
 311 /*
 312  * drm_sched_entity_clear_dep - callback to clear the entities dependency
 313  */
 314 static void drm_sched_entity_clear_dep(struct dma_fence *f,
 315                                        struct dma_fence_cb *cb)
 316 {
 317         struct drm_sched_entity *entity =
 318                 container_of(cb, struct drm_sched_entity, cb);
 319
 320         entity->dependency = NULL;
 321         dma_fence_put(f);
 322 }
 323
 324 /*
 325  * drm_sched_entity_clear_dep - callback to clear the entities dependency and
 326  * wake up scheduler
 327  */
 328 static void drm_sched_entity_wakeup(struct dma_fence *f,
 329                                     struct dma_fence_cb *cb)
 330 {
 331         struct drm_sched_entity *entity =
 332                 container_of(cb, struct drm_sched_entity, cb);
 333
 334         drm_sched_entity_clear_dep(f, cb);
 335         drm_sched_wakeup(entity->rq->sched);
 336 }
 337
 338 /**
 339  * drm_sched_entity_set_priority - Sets priority of the entity
 340  *
 341  * @entity: scheduler entity
 342  * @priority: scheduler priority
 343  *
 344  * Update the priority of runqueus used for the entity.
 345  */
 346 void drm_sched_entity_set_priority(struct drm_sched_entity *entity,
 347                                    enum drm_sched_priority priority)
 348 {
 349         spin_lock(&entity->rq_lock);
 350         entity->priority = priority;
 351         spin_unlock(&entity->rq_lock);
 352 }
 353 EXPORT_SYMBOL(drm_sched_entity_set_priority);
 354
 355 /**
 356  * drm_sched_entity_add_dependency_cb - add callback for the entities dependency
 357  *
 358  * @entity: entity with dependency
 359  *
 360  * Add a callback to the current dependency of the entity to wake up the
 361  * scheduler when the entity becomes available.
 362  */
 363 static bool drm_sched_entity_add_dependency_cb(struct drm_sched_entity *entity)
 364 {
 365         struct drm_gpu_scheduler *sched = entity->rq->sched;
 366         struct dma_fence *fence = entity->dependency;
 367         struct drm_sched_fence *s_fence;
 368
 369         if (fence->context == entity->fence_context ||
 370             fence->context == entity->fence_context + 1) {
 371                 /*
 372                  * Fence is a scheduled/finished fence from a job
 373                  * which belongs to the same entity, we can ignore
 374                  * fences from ourself
 375                  */
 376                 dma_fence_put(entity->dependency);
 377                 return false;
 378         }
 379
 380         s_fence = to_drm_sched_fence(fence);
 381         if (s_fence && s_fence->sched == sched) {
 382
 383                 /*
 384                  * Fence is from the same scheduler, only need to wait for
 385                  * it to be scheduled
 386                  */
 387                 fence = dma_fence_get(&s_fence->scheduled);
 388                 dma_fence_put(entity->dependency);
 389                 entity->dependency = fence;
 390                 if (!dma_fence_add_callback(fence, &entity->cb,
 391                                             drm_sched_entity_clear_dep))
 392                         return true;
 393
 394                 /* Ignore it when it is already scheduled */
 395                 dma_fence_put(fence);
 396                 return false;
 397         }
 398
 399         if (!dma_fence_add_callback(entity->dependency, &entity->cb,
 400                                     drm_sched_entity_wakeup))
 401                 return true;
 402
 403         dma_fence_put(entity->dependency);
 404         return false;
 405 }
 406
 407 /**
 408  * drm_sched_entity_pop_job - get a ready to be scheduled job from the entity
 409  *
 410  * @entity: entity to get the job from
 411  *
 412  * Process all dependencies and try to get one job from the entities queue.
 413  */
 414 struct drm_sched_job *drm_sched_entity_pop_job(struct drm_sched_entity *entity)
 415 {
 416         struct drm_gpu_scheduler *sched = entity->rq->sched;
 417         struct drm_sched_job *sched_job;
 418
 419         sched_job = to_drm_sched_job(spsc_queue_peek(&entity->job_queue));
 420         if (!sched_job)
 421                 return NULL;
 422
 423         while ((entity->dependency =
 424                         sched->ops->dependency(sched_job, entity))) {
 425                 trace_drm_sched_job_wait_dep(sched_job, entity->dependency);
 426
 427                 if (drm_sched_entity_add_dependency_cb(entity))
 428                         return NULL;
 429         }
 430
 431         /* skip jobs from entity that marked guilty */
 432         if (entity->guilty && atomic_read(entity->guilty))
 433                 dma_fence_set_error(&sched_job->s_fence->finished, -ECANCELED);
 434
 435         dma_fence_put(entity->last_scheduled);
 436         entity->last_scheduled = dma_fence_get(&sched_job->s_fence->finished);
 437
 438         spsc_queue_pop(&entity->job_queue);
 439         return sched_job;
 440 }
 441
 442 /**
 443  * drm_sched_entity_select_rq - select a new rq for the entity
 444  *
 445  * @entity: scheduler entity
 446  *
 447  * Check all prerequisites and select a new rq for the entity for load
 448  * balancing.
 449  */
 450 void drm_sched_entity_select_rq(struct drm_sched_entity *entity)
 451 {
 452         struct dma_fence *fence;
 453         struct drm_gpu_scheduler *sched;
 454         struct drm_sched_rq *rq;
 455
 456         if (spsc_queue_count(&entity->job_queue) || !entity->sched_list)
 457                 return;
 458
 459         fence = READ_ONCE(entity->last_scheduled);
 460         if (fence && !dma_fence_is_signaled(fence))
 461                 return;
 462
 463         spin_lock(&entity->rq_lock);
 464         sched = drm_sched_pick_best(entity->sched_list, entity->num_sched_list);
 465         rq = sched ? &sched->sched_rq[entity->priority] : NULL;
 466         if (rq != entity->rq) {
 467                 drm_sched_rq_remove_entity(entity->rq, entity);
 468                 entity->rq = rq;
 469         }
 470         spin_unlock(&entity->rq_lock);
 471
 472         if (entity->num_sched_list == 1)
 473                 entity->sched_list = NULL;
 474 }
 475
 476 /**
 477  * drm_sched_entity_push_job - Submit a job to the entity's job queue
 478  *
 479  * @sched_job: job to submit
 480  * @entity: scheduler entity
 481  *
 482  * Note: To guarantee that the order of insertion to queue matches
 483  * the job's fence sequence number this function should be
 484  * called with drm_sched_job_init under common lock.
 485  *
 486  * Returns 0 for success, negative error code otherwise.
 487  */
 488 void drm_sched_entity_push_job(struct drm_sched_job *sched_job,
 489                                struct drm_sched_entity *entity)
 490 {
 491         bool first;
 492
 493         trace_drm_sched_job(sched_job, entity);
 494         atomic_inc(entity->rq->sched->score);
 495         WRITE_ONCE(entity->last_user, current->group_leader);
 496         first = spsc_queue_push(&entity->job_queue, &sched_job->queue_node);
 497
 498         /* first job wakes up scheduler */
 499         if (first) {
 500                 /* Add the entity to the run queue */
 501                 spin_lock(&entity->rq_lock);
 502                 if (entity->stopped) {
 503                         spin_unlock(&entity->rq_lock);
 504
 505                         DRM_ERROR("Trying to push to a killed entity\n");
 506                         return;
 507                 }
 508                 drm_sched_rq_add_entity(entity->rq, entity);
 509                 spin_unlock(&entity->rq_lock);
 510                 drm_sched_wakeup(entity->rq->sched);
 511         }
 512 }
 513 EXPORT_SYMBOL(drm_sched_entity_push_job);