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 <drm/gpu_scheduler.h>
  26
  27 #include "gpu_scheduler_trace.h"
  28
  29 #define to_drm_sched_job(sched_job)             \
  30                 container_of((sched_job), struct drm_sched_job, queue_node)
  31
  32 /**
  33  * drm_sched_entity_init - Init a context entity used by scheduler when
  34  * submit to HW ring.
  35  *
  36  * @entity: scheduler entity to init
  37  * @rq_list: the list of run queue on which jobs from this
  38  *           entity can be submitted
  39  * @num_rq_list: number of run queue in rq_list
  40  * @guilty: atomic_t set to 1 when a job on this queue
  41  *          is found to be guilty causing a timeout
  42  *
  43  * Note: the rq_list should have atleast one element to schedule
  44  *       the entity
  45  *
  46  * Returns 0 on success or a negative error code on failure.
  47  */
  48 int drm_sched_entity_init(struct drm_sched_entity *entity,
  49                           struct drm_sched_rq **rq_list,
  50                           unsigned int num_rq_list,
  51                           atomic_t *guilty)
  52 {
  53         int i;
  54
  55         if (!(entity && rq_list && num_rq_list > 0 && rq_list[0]))
  56                 return -EINVAL;
  57
  58         memset(entity, 0, sizeof(struct drm_sched_entity));
  59         INIT_LIST_HEAD(&entity->list);
  60         entity->rq = rq_list[0];
  61         entity->guilty = guilty;
  62         entity->num_rq_list = num_rq_list;
  63         entity->rq_list = kcalloc(num_rq_list, sizeof(struct drm_sched_rq *),
  64                                 GFP_KERNEL);
  65         if (!entity->rq_list)
  66                 return -ENOMEM;
  67
  68         for (i = 0; i < num_rq_list; ++i)
  69                 entity->rq_list[i] = rq_list[i];
  70         entity->last_scheduled = NULL;
  71
  72         spin_lock_init(&entity->rq_lock);
  73         spsc_queue_init(&entity->job_queue);
  74
  75         atomic_set(&entity->fence_seq, 0);
  76         entity->fence_context = dma_fence_context_alloc(2);
  77
  78         return 0;
  79 }
  80 EXPORT_SYMBOL(drm_sched_entity_init);
  81
  82 /**
  83  * drm_sched_entity_is_idle - Check if entity is idle
  84  *
  85  * @entity: scheduler entity
  86  *
  87  * Returns true if the entity does not have any unscheduled jobs.
  88  */
  89 static bool drm_sched_entity_is_idle(struct drm_sched_entity *entity)
  90 {
  91         rmb(); /* for list_empty to work without lock */
  92
  93         if (list_empty(&entity->list) ||
  94             spsc_queue_peek(&entity->job_queue) == NULL)
  95                 return true;
  96
  97         return false;
  98 }
  99
 100 /**
 101  * drm_sched_entity_is_ready - Check if entity is ready
 102  *
 103  * @entity: scheduler entity
 104  *
 105  * Return true if entity could provide a job.
 106  */
 107 bool drm_sched_entity_is_ready(struct drm_sched_entity *entity)
 108 {
 109         if (spsc_queue_peek(&entity->job_queue) == NULL)
 110                 return false;
 111
 112         if (READ_ONCE(entity->dependency))
 113                 return false;
 114
 115         return true;
 116 }
 117
 118 /**
 119  * drm_sched_entity_get_free_sched - Get the rq from rq_list with least load
 120  *
 121  * @entity: scheduler entity
 122  *
 123  * Return the pointer to the rq with least load.
 124  */
 125 static struct drm_sched_rq *
 126 drm_sched_entity_get_free_sched(struct drm_sched_entity *entity)
 127 {
 128         struct drm_sched_rq *rq = NULL;
 129         unsigned int min_jobs = UINT_MAX, num_jobs;
 130         int i;
 131
 132         for (i = 0; i < entity->num_rq_list; ++i) {
 133                 num_jobs = atomic_read(&entity->rq_list[i]->sched->num_jobs);
 134                 if (num_jobs < min_jobs) {
 135                         min_jobs = num_jobs;
 136                         rq = entity->rq_list[i];
 137                 }
 138         }
 139
 140         return rq;
 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         sched = entity->rq->sched;
 162         /**
 163          * The client will not queue more IBs during this fini, consume existing
 164          * queued IBs or discard them on SIGKILL
 165          */
 166         if (current->flags & PF_EXITING) {
 167                 if (timeout)
 168                         ret = wait_event_timeout(
 169                                         sched->job_scheduled,
 170                                         drm_sched_entity_is_idle(entity),
 171                                         timeout);
 172         } else {
 173                 wait_event_killable(sched->job_scheduled,
 174                                     drm_sched_entity_is_idle(entity));
 175         }
 176
 177         /* For killed process disable any more IBs enqueue right now */
 178         last_user = cmpxchg(&entity->last_user, current->group_leader, NULL);
 179         if ((!last_user || last_user == current->group_leader) &&
 180             (current->flags & PF_EXITING) && (current->exit_code == SIGKILL)) {
 181                 spin_lock(&entity->rq_lock);
 182                 entity->stopped = true;
 183                 drm_sched_rq_remove_entity(entity->rq, entity);
 184                 spin_unlock(&entity->rq_lock);
 185         }
 186
 187         return ret;
 188 }
 189 EXPORT_SYMBOL(drm_sched_entity_flush);
 190
 191 /**
 192  * drm_sched_entity_kill_jobs - helper for drm_sched_entity_kill_jobs
 193  *
 194  * @f: signaled fence
 195  * @cb: our callback structure
 196  *
 197  * Signal the scheduler finished fence when the entity in question is killed.
 198  */
 199 static void drm_sched_entity_kill_jobs_cb(struct dma_fence *f,
 200                                           struct dma_fence_cb *cb)
 201 {
 202         struct drm_sched_job *job = container_of(cb, struct drm_sched_job,
 203                                                  finish_cb);
 204
 205         drm_sched_fence_finished(job->s_fence);
 206         WARN_ON(job->s_fence->parent);
 207         dma_fence_put(&job->s_fence->finished);
 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         int r;
 223
 224         while ((job = to_drm_sched_job(spsc_queue_pop(&entity->job_queue)))) {
 225                 struct drm_sched_fence *s_fence = job->s_fence;
 226
 227                 drm_sched_fence_scheduled(s_fence);
 228                 dma_fence_set_error(&s_fence->finished, -ESRCH);
 229
 230                 /*
 231                  * When pipe is hanged by older entity, new entity might
 232                  * not even have chance to submit it's first job to HW
 233                  * and so entity->last_scheduled will remain NULL
 234                  */
 235                 if (!entity->last_scheduled) {
 236                         drm_sched_entity_kill_jobs_cb(NULL, &job->finish_cb);
 237                         continue;
 238                 }
 239
 240                 r = dma_fence_add_callback(entity->last_scheduled,
 241                                            &job->finish_cb,
 242                                            drm_sched_entity_kill_jobs_cb);
 243                 if (r == -ENOENT)
 244                         drm_sched_entity_kill_jobs_cb(NULL, &job->finish_cb);
 245                 else if (r)
 246                         DRM_ERROR("fence add callback failed (%d)\n", r);
 247         }
 248 }
 249
 250 /**
 251  * drm_sched_entity_cleanup - Destroy a context entity
 252  *
 253  * @entity: scheduler entity
 254  *
 255  * This should be called after @drm_sched_entity_do_release. It goes over the
 256  * entity and signals all jobs with an error code if the process was killed.
 257  *
 258  */
 259 void drm_sched_entity_fini(struct drm_sched_entity *entity)
 260 {
 261         struct drm_gpu_scheduler *sched;
 262
 263         sched = entity->rq->sched;
 264         drm_sched_rq_remove_entity(entity->rq, entity);
 265
 266         /* Consumption of existing IBs wasn't completed. Forcefully
 267          * remove them here.
 268          */
 269         if (spsc_queue_peek(&entity->job_queue)) {
 270                 /* Park the kernel for a moment to make sure it isn't processing
 271                  * our enity.
 272                  */
 273                 kthread_park(sched->thread);
 274                 kthread_unpark(sched->thread);
 275                 if (entity->dependency) {
 276                         dma_fence_remove_callback(entity->dependency,
 277                                                   &entity->cb);
 278                         dma_fence_put(entity->dependency);
 279                         entity->dependency = NULL;
 280                 }
 281
 282                 drm_sched_entity_kill_jobs(entity);
 283         }
 284
 285         dma_fence_put(entity->last_scheduled);
 286         entity->last_scheduled = NULL;
 287         kfree(entity->rq_list);
 288 }
 289 EXPORT_SYMBOL(drm_sched_entity_fini);
 290
 291 /**
 292  * drm_sched_entity_fini - Destroy a context entity
 293  *
 294  * @entity: scheduler entity
 295  *
 296  * Calls drm_sched_entity_do_release() and drm_sched_entity_cleanup()
 297  */
 298 void drm_sched_entity_destroy(struct drm_sched_entity *entity)
 299 {
 300         drm_sched_entity_flush(entity, MAX_WAIT_SCHED_ENTITY_Q_EMPTY);
 301         drm_sched_entity_fini(entity);
 302 }
 303 EXPORT_SYMBOL(drm_sched_entity_destroy);
 304
 305 /**
 306  * drm_sched_entity_clear_dep - callback to clear the entities dependency
 307  */
 308 static void drm_sched_entity_clear_dep(struct dma_fence *f,
 309                                        struct dma_fence_cb *cb)
 310 {
 311         struct drm_sched_entity *entity =
 312                 container_of(cb, struct drm_sched_entity, cb);
 313
 314         entity->dependency = NULL;
 315         dma_fence_put(f);
 316 }
 317
 318 /**
 319  * drm_sched_entity_clear_dep - callback to clear the entities dependency and
 320  * wake up scheduler
 321  */
 322 static void drm_sched_entity_wakeup(struct dma_fence *f,
 323                                     struct dma_fence_cb *cb)
 324 {
 325         struct drm_sched_entity *entity =
 326                 container_of(cb, struct drm_sched_entity, cb);
 327
 328         drm_sched_entity_clear_dep(f, cb);
 329         drm_sched_wakeup(entity->rq->sched);
 330 }
 331
 332 /**
 333  * drm_sched_entity_set_rq_priority - helper for drm_sched_entity_set_priority
 334  */
 335 static void drm_sched_entity_set_rq_priority(struct drm_sched_rq **rq,
 336                                              enum drm_sched_priority priority)
 337 {
 338         *rq = &(*rq)->sched->sched_rq[priority];
 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         unsigned int i;
 353
 354         spin_lock(&entity->rq_lock);
 355
 356         for (i = 0; i < entity->num_rq_list; ++i)
 357                 drm_sched_entity_set_rq_priority(&entity->rq_list[i], priority);
 358
 359         drm_sched_rq_remove_entity(entity->rq, entity);
 360         drm_sched_entity_set_rq_priority(&entity->rq, priority);
 361         drm_sched_rq_add_entity(entity->rq, entity);
 362
 363         spin_unlock(&entity->rq_lock);
 364 }
 365 EXPORT_SYMBOL(drm_sched_entity_set_priority);
 366
 367 /**
 368  * drm_sched_entity_add_dependency_cb - add callback for the entities dependency
 369  *
 370  * @entity: entity with dependency
 371  *
 372  * Add a callback to the current dependency of the entity to wake up the
 373  * scheduler when the entity becomes available.
 374  */
 375 static bool drm_sched_entity_add_dependency_cb(struct drm_sched_entity *entity)
 376 {
 377         struct drm_gpu_scheduler *sched = entity->rq->sched;
 378         struct dma_fence *fence = entity->dependency;
 379         struct drm_sched_fence *s_fence;
 380
 381         if (fence->context == entity->fence_context ||
 382             fence->context == entity->fence_context + 1) {
 383                 /*
 384                  * Fence is a scheduled/finished fence from a job
 385                  * which belongs to the same entity, we can ignore
 386                  * fences from ourself
 387                  */
 388                 dma_fence_put(entity->dependency);
 389                 return false;
 390         }
 391
 392         s_fence = to_drm_sched_fence(fence);
 393         if (s_fence && s_fence->sched == sched) {
 394
 395                 /*
 396                  * Fence is from the same scheduler, only need to wait for
 397                  * it to be scheduled
 398                  */
 399                 fence = dma_fence_get(&s_fence->scheduled);
 400                 dma_fence_put(entity->dependency);
 401                 entity->dependency = fence;
 402                 if (!dma_fence_add_callback(fence, &entity->cb,
 403                                             drm_sched_entity_clear_dep))
 404                         return true;
 405
 406                 /* Ignore it when it is already scheduled */
 407                 dma_fence_put(fence);
 408                 return false;
 409         }
 410
 411         if (!dma_fence_add_callback(entity->dependency, &entity->cb,
 412                                     drm_sched_entity_wakeup))
 413                 return true;
 414
 415         dma_fence_put(entity->dependency);
 416         return false;
 417 }
 418
 419 /**
 420  * drm_sched_entity_pop_job - get a ready to be scheduled job from the entity
 421  *
 422  * @entity: entity to get the job from
 423  *
 424  * Process all dependencies and try to get one job from the entities queue.
 425  */
 426 struct drm_sched_job *drm_sched_entity_pop_job(struct drm_sched_entity *entity)
 427 {
 428         struct drm_gpu_scheduler *sched = entity->rq->sched;
 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                         sched->ops->dependency(sched_job, entity))) {
 437
 438                 if (drm_sched_entity_add_dependency_cb(entity)) {
 439
 440                         trace_drm_sched_job_wait_dep(sched_job,
 441                                                      entity->dependency);
 442                         return NULL;
 443                 }
 444         }
 445
 446         /* skip jobs from entity that marked guilty */
 447         if (entity->guilty && atomic_read(entity->guilty))
 448                 dma_fence_set_error(&sched_job->s_fence->finished, -ECANCELED);
 449
 450         dma_fence_put(entity->last_scheduled);
 451         entity->last_scheduled = dma_fence_get(&sched_job->s_fence->finished);
 452
 453         spsc_queue_pop(&entity->job_queue);
 454         return sched_job;
 455 }
 456
 457 /**
 458  * drm_sched_entity_select_rq - select a new rq for the entity
 459  *
 460  * @entity: scheduler entity
 461  *
 462  * Check all prerequisites and select a new rq for the entity for load
 463  * balancing.
 464  */
 465 void drm_sched_entity_select_rq(struct drm_sched_entity *entity)
 466 {
 467         struct dma_fence *fence;
 468         struct drm_sched_rq *rq;
 469
 470         if (spsc_queue_count(&entity->job_queue) || entity->num_rq_list <= 1)
 471                 return;
 472
 473         fence = READ_ONCE(entity->last_scheduled);
 474         if (fence && !dma_fence_is_signaled(fence))
 475                 return;
 476
 477         rq = drm_sched_entity_get_free_sched(entity);
 478         if (rq == entity->rq)
 479                 return;
 480
 481         spin_lock(&entity->rq_lock);
 482         drm_sched_rq_remove_entity(entity->rq, entity);
 483         entity->rq = rq;
 484         spin_unlock(&entity->rq_lock);
 485 }
 486
 487 /**
 488  * drm_sched_entity_push_job - Submit a job to the entity's job queue
 489  *
 490  * @sched_job: job to submit
 491  * @entity: scheduler entity
 492  *
 493  * Note: To guarantee that the order of insertion to queue matches
 494  * the job's fence sequence number this function should be
 495  * called with drm_sched_job_init under common lock.
 496  *
 497  * Returns 0 for success, negative error code otherwise.
 498  */
 499 void drm_sched_entity_push_job(struct drm_sched_job *sched_job,
 500                                struct drm_sched_entity *entity)
 501 {
 502         bool first;
 503
 504         trace_drm_sched_job(sched_job, entity);
 505         atomic_inc(&entity->rq->sched->num_jobs);
 506         WRITE_ONCE(entity->last_user, current->group_leader);
 507         first = spsc_queue_push(&entity->job_queue, &sched_job->queue_node);
 508
 509         /* first job wakes up scheduler */
 510         if (first) {
 511                 /* Add the entity to the run queue */
 512                 spin_lock(&entity->rq_lock);
 513                 if (entity->stopped) {
 514                         spin_unlock(&entity->rq_lock);
 515
 516                         DRM_ERROR("Trying to push to a killed entity\n");
 517                         return;
 518                 }
 519                 drm_sched_rq_add_entity(entity->rq, entity);
 520                 spin_unlock(&entity->rq_lock);
 521                 drm_sched_wakeup(entity->rq->sched);
 522         }
 523 }
 524 EXPORT_SYMBOL(drm_sched_entity_push_job);