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                 struct drm_gpu_scheduler *sched = entity->rq_list[i]->sched;
 134
 135                 if (!entity->rq_list[i]->sched->ready) {
 136                         DRM_WARN("sched%s is not ready, skipping", sched->name);
 137                         continue;
 138                 }
 139
 140                 num_jobs = atomic_read(&sched->num_jobs);
 141                 if (num_jobs < min_jobs) {
 142                         min_jobs = num_jobs;
 143                         rq = entity->rq_list[i];
 144                 }
 145         }
 146
 147         return rq;
 148 }
 149
 150 /**
 151  * drm_sched_entity_flush - Flush a context entity
 152  *
 153  * @entity: scheduler entity
 154  * @timeout: time to wait in for Q to become empty in jiffies.
 155  *
 156  * Splitting drm_sched_entity_fini() into two functions, The first one does the
 157  * waiting, removes the entity from the runqueue and returns an error when the
 158  * process was killed.
 159  *
 160  * Returns the remaining time in jiffies left from the input timeout
 161  */
 162 long drm_sched_entity_flush(struct drm_sched_entity *entity, long timeout)
 163 {
 164         struct drm_gpu_scheduler *sched;
 165         struct task_struct *last_user;
 166         long ret = timeout;
 167
 168         sched = entity->rq->sched;
 169         /**
 170          * The client will not queue more IBs during this fini, consume existing
 171          * queued IBs or discard them on SIGKILL
 172          */
 173         if (current->flags & PF_EXITING) {
 174                 if (timeout)
 175                         ret = wait_event_timeout(
 176                                         sched->job_scheduled,
 177                                         drm_sched_entity_is_idle(entity),
 178                                         timeout);
 179         } else {
 180                 wait_event_killable(sched->job_scheduled,
 181                                     drm_sched_entity_is_idle(entity));
 182         }
 183
 184         /* For killed process disable any more IBs enqueue right now */
 185         last_user = cmpxchg(&entity->last_user, current->group_leader, NULL);
 186         if ((!last_user || last_user == current->group_leader) &&
 187             (current->flags & PF_EXITING) && (current->exit_code == SIGKILL)) {
 188                 spin_lock(&entity->rq_lock);
 189                 entity->stopped = true;
 190                 drm_sched_rq_remove_entity(entity->rq, entity);
 191                 spin_unlock(&entity->rq_lock);
 192         }
 193
 194         return ret;
 195 }
 196 EXPORT_SYMBOL(drm_sched_entity_flush);
 197
 198 /**
 199  * drm_sched_entity_kill_jobs - helper for drm_sched_entity_kill_jobs
 200  *
 201  * @f: signaled fence
 202  * @cb: our callback structure
 203  *
 204  * Signal the scheduler finished fence when the entity in question is killed.
 205  */
 206 static void drm_sched_entity_kill_jobs_cb(struct dma_fence *f,
 207                                           struct dma_fence_cb *cb)
 208 {
 209         struct drm_sched_job *job = container_of(cb, struct drm_sched_job,
 210                                                  finish_cb);
 211
 212         drm_sched_fence_finished(job->s_fence);
 213         WARN_ON(job->s_fence->parent);
 214         job->sched->ops->free_job(job);
 215 }
 216
 217 /**
 218  * drm_sched_entity_kill_jobs - Make sure all remaining jobs are killed
 219  *
 220  * @entity: entity which is cleaned up
 221  *
 222  * Makes sure that all remaining jobs in an entity are killed before it is
 223  * destroyed.
 224  */
 225 static void drm_sched_entity_kill_jobs(struct drm_sched_entity *entity)
 226 {
 227         struct drm_sched_job *job;
 228         int r;
 229
 230         while ((job = to_drm_sched_job(spsc_queue_pop(&entity->job_queue)))) {
 231                 struct drm_sched_fence *s_fence = job->s_fence;
 232
 233                 drm_sched_fence_scheduled(s_fence);
 234                 dma_fence_set_error(&s_fence->finished, -ESRCH);
 235
 236                 /*
 237                  * When pipe is hanged by older entity, new entity might
 238                  * not even have chance to submit it's first job to HW
 239                  * and so entity->last_scheduled will remain NULL
 240                  */
 241                 if (!entity->last_scheduled) {
 242                         drm_sched_entity_kill_jobs_cb(NULL, &job->finish_cb);
 243                         continue;
 244                 }
 245
 246                 r = dma_fence_add_callback(entity->last_scheduled,
 247                                            &job->finish_cb,
 248                                            drm_sched_entity_kill_jobs_cb);
 249                 if (r == -ENOENT)
 250                         drm_sched_entity_kill_jobs_cb(NULL, &job->finish_cb);
 251                 else if (r)
 252                         DRM_ERROR("fence add callback failed (%d)\n", r);
 253         }
 254 }
 255
 256 /**
 257  * drm_sched_entity_cleanup - Destroy a context entity
 258  *
 259  * @entity: scheduler entity
 260  *
 261  * This should be called after @drm_sched_entity_do_release. It goes over the
 262  * entity and signals all jobs with an error code if the process was killed.
 263  *
 264  */
 265 void drm_sched_entity_fini(struct drm_sched_entity *entity)
 266 {
 267         struct drm_gpu_scheduler *sched;
 268
 269         sched = entity->rq->sched;
 270         drm_sched_rq_remove_entity(entity->rq, entity);
 271
 272         /* Consumption of existing IBs wasn't completed. Forcefully
 273          * remove them here.
 274          */
 275         if (spsc_queue_peek(&entity->job_queue)) {
 276                 /* Park the kernel for a moment to make sure it isn't processing
 277                  * our enity.
 278                  */
 279                 kthread_park(sched->thread);
 280                 kthread_unpark(sched->thread);
 281                 if (entity->dependency) {
 282                         dma_fence_remove_callback(entity->dependency,
 283                                                   &entity->cb);
 284                         dma_fence_put(entity->dependency);
 285                         entity->dependency = NULL;
 286                 }
 287
 288                 drm_sched_entity_kill_jobs(entity);
 289         }
 290
 291         dma_fence_put(entity->last_scheduled);
 292         entity->last_scheduled = NULL;
 293         kfree(entity->rq_list);
 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_rq_priority - helper for drm_sched_entity_set_priority
 340  */
 341 static void drm_sched_entity_set_rq_priority(struct drm_sched_rq **rq,
 342                                              enum drm_sched_priority priority)
 343 {
 344         *rq = &(*rq)->sched->sched_rq[priority];
 345 }
 346
 347 /**
 348  * drm_sched_entity_set_priority - Sets priority of the entity
 349  *
 350  * @entity: scheduler entity
 351  * @priority: scheduler priority
 352  *
 353  * Update the priority of runqueus used for the entity.
 354  */
 355 void drm_sched_entity_set_priority(struct drm_sched_entity *entity,
 356                                    enum drm_sched_priority priority)
 357 {
 358         unsigned int i;
 359
 360         spin_lock(&entity->rq_lock);
 361
 362         for (i = 0; i < entity->num_rq_list; ++i)
 363                 drm_sched_entity_set_rq_priority(&entity->rq_list[i], priority);
 364
 365         drm_sched_rq_remove_entity(entity->rq, entity);
 366         drm_sched_entity_set_rq_priority(&entity->rq, priority);
 367         drm_sched_rq_add_entity(entity->rq, entity);
 368
 369         spin_unlock(&entity->rq_lock);
 370 }
 371 EXPORT_SYMBOL(drm_sched_entity_set_priority);
 372
 373 /**
 374  * drm_sched_entity_add_dependency_cb - add callback for the entities dependency
 375  *
 376  * @entity: entity with dependency
 377  *
 378  * Add a callback to the current dependency of the entity to wake up the
 379  * scheduler when the entity becomes available.
 380  */
 381 static bool drm_sched_entity_add_dependency_cb(struct drm_sched_entity *entity)
 382 {
 383         struct drm_gpu_scheduler *sched = entity->rq->sched;
 384         struct dma_fence *fence = entity->dependency;
 385         struct drm_sched_fence *s_fence;
 386
 387         if (fence->context == entity->fence_context ||
 388             fence->context == entity->fence_context + 1) {
 389                 /*
 390                  * Fence is a scheduled/finished fence from a job
 391                  * which belongs to the same entity, we can ignore
 392                  * fences from ourself
 393                  */
 394                 dma_fence_put(entity->dependency);
 395                 return false;
 396         }
 397
 398         s_fence = to_drm_sched_fence(fence);
 399         if (s_fence && s_fence->sched == sched) {
 400
 401                 /*
 402                  * Fence is from the same scheduler, only need to wait for
 403                  * it to be scheduled
 404                  */
 405                 fence = dma_fence_get(&s_fence->scheduled);
 406                 dma_fence_put(entity->dependency);
 407                 entity->dependency = fence;
 408                 if (!dma_fence_add_callback(fence, &entity->cb,
 409                                             drm_sched_entity_clear_dep))
 410                         return true;
 411
 412                 /* Ignore it when it is already scheduled */
 413                 dma_fence_put(fence);
 414                 return false;
 415         }
 416
 417         if (!dma_fence_add_callback(entity->dependency, &entity->cb,
 418                                     drm_sched_entity_wakeup))
 419                 return true;
 420
 421         dma_fence_put(entity->dependency);
 422         return false;
 423 }
 424
 425 /**
 426  * drm_sched_entity_pop_job - get a ready to be scheduled job from the entity
 427  *
 428  * @entity: entity to get the job from
 429  *
 430  * Process all dependencies and try to get one job from the entities queue.
 431  */
 432 struct drm_sched_job *drm_sched_entity_pop_job(struct drm_sched_entity *entity)
 433 {
 434         struct drm_gpu_scheduler *sched = entity->rq->sched;
 435         struct drm_sched_job *sched_job;
 436
 437         sched_job = to_drm_sched_job(spsc_queue_peek(&entity->job_queue));
 438         if (!sched_job)
 439                 return NULL;
 440
 441         while ((entity->dependency =
 442                         sched->ops->dependency(sched_job, entity))) {
 443                 trace_drm_sched_job_wait_dep(sched_job, entity->dependency);
 444
 445                 if (drm_sched_entity_add_dependency_cb(entity))
 446                         return NULL;
 447         }
 448
 449         /* skip jobs from entity that marked guilty */
 450         if (entity->guilty && atomic_read(entity->guilty))
 451                 dma_fence_set_error(&sched_job->s_fence->finished, -ECANCELED);
 452
 453         dma_fence_put(entity->last_scheduled);
 454         entity->last_scheduled = dma_fence_get(&sched_job->s_fence->finished);
 455
 456         spsc_queue_pop(&entity->job_queue);
 457         return sched_job;
 458 }
 459
 460 /**
 461  * drm_sched_entity_select_rq - select a new rq for the entity
 462  *
 463  * @entity: scheduler entity
 464  *
 465  * Check all prerequisites and select a new rq for the entity for load
 466  * balancing.
 467  */
 468 void drm_sched_entity_select_rq(struct drm_sched_entity *entity)
 469 {
 470         struct dma_fence *fence;
 471         struct drm_sched_rq *rq;
 472
 473         if (spsc_queue_count(&entity->job_queue) || entity->num_rq_list <= 1)
 474                 return;
 475
 476         fence = READ_ONCE(entity->last_scheduled);
 477         if (fence && !dma_fence_is_signaled(fence))
 478                 return;
 479
 480         rq = drm_sched_entity_get_free_sched(entity);
 481         if (rq == entity->rq)
 482                 return;
 483
 484         spin_lock(&entity->rq_lock);
 485         drm_sched_rq_remove_entity(entity->rq, entity);
 486         entity->rq = rq;
 487         spin_unlock(&entity->rq_lock);
 488 }
 489
 490 /**
 491  * drm_sched_entity_push_job - Submit a job to the entity's job queue
 492  *
 493  * @sched_job: job to submit
 494  * @entity: scheduler entity
 495  *
 496  * Note: To guarantee that the order of insertion to queue matches
 497  * the job's fence sequence number this function should be
 498  * called with drm_sched_job_init under common lock.
 499  *
 500  * Returns 0 for success, negative error code otherwise.
 501  */
 502 void drm_sched_entity_push_job(struct drm_sched_job *sched_job,
 503                                struct drm_sched_entity *entity)
 504 {
 505         bool first;
 506
 507         trace_drm_sched_job(sched_job, entity);
 508         atomic_inc(&entity->rq->sched->num_jobs);
 509         WRITE_ONCE(entity->last_user, current->group_leader);
 510         first = spsc_queue_push(&entity->job_queue, &sched_job->queue_node);
 511
 512         /* first job wakes up scheduler */
 513         if (first) {
 514                 /* Add the entity to the run queue */
 515                 spin_lock(&entity->rq_lock);
 516                 if (entity->stopped) {
 517                         spin_unlock(&entity->rq_lock);
 518
 519                         DRM_ERROR("Trying to push to a killed entity\n");
 520                         return;
 521                 }
 522                 drm_sched_rq_add_entity(entity->rq, entity);
 523                 spin_unlock(&entity->rq_lock);
 524                 drm_sched_wakeup(entity->rq->sched);
 525         }
 526 }
 527 EXPORT_SYMBOL(drm_sched_entity_push_job);