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             entity->stopped)
 121                 return true;
 122
 123         return false;
 124 }
 125
 126 /**
 127  * drm_sched_entity_is_ready - Check if entity is ready
 128  *
 129  * @entity: scheduler entity
 130  *
 131  * Return true if entity could provide a job.
 132  */
 133 bool drm_sched_entity_is_ready(struct drm_sched_entity *entity)
 134 {
 135         if (spsc_queue_peek(&entity->job_queue) == NULL)
 136                 return false;
 137
 138         if (READ_ONCE(entity->dependency))
 139                 return false;
 140
 141         return true;
 142 }
 143
 144 /**
 145  * drm_sched_entity_flush - Flush a context entity
 146  *
 147  * @entity: scheduler entity
 148  * @timeout: time to wait in for Q to become empty in jiffies.
 149  *
 150  * Splitting drm_sched_entity_fini() into two functions, The first one does the
 151  * waiting, removes the entity from the runqueue and returns an error when the
 152  * process was killed.
 153  *
 154  * Returns the remaining time in jiffies left from the input timeout
 155  */
 156 long drm_sched_entity_flush(struct drm_sched_entity *entity, long timeout)
 157 {
 158         struct drm_gpu_scheduler *sched;
 159         struct task_struct *last_user;
 160         long ret = timeout;
 161
 162         if (!entity->rq)
 163                 return 0;
 164
 165         sched = entity->rq->sched;
 166         /**
 167          * The client will not queue more IBs during this fini, consume existing
 168          * queued IBs or discard them on SIGKILL
 169          */
 170         if (current->flags & PF_EXITING) {
 171                 if (timeout)
 172                         ret = wait_event_timeout(
 173                                         sched->job_scheduled,
 174                                         drm_sched_entity_is_idle(entity),
 175                                         timeout);
 176         } else {
 177                 wait_event_killable(sched->job_scheduled,
 178                                     drm_sched_entity_is_idle(entity));
 179         }
 180
 181         /* For killed process disable any more IBs enqueue right now */
 182         last_user = cmpxchg(&entity->last_user, current->group_leader, NULL);
 183         if ((!last_user || last_user == current->group_leader) &&
 184             (current->flags & PF_EXITING) && (current->exit_code == SIGKILL)) {
 185                 spin_lock(&entity->rq_lock);
 186                 entity->stopped = true;
 187                 drm_sched_rq_remove_entity(entity->rq, entity);
 188                 spin_unlock(&entity->rq_lock);
 189         }
 190
 191         return ret;
 192 }
 193 EXPORT_SYMBOL(drm_sched_entity_flush);
 194
 195 /**
 196  * drm_sched_entity_kill_jobs_cb - helper for drm_sched_entity_kill_jobs
 197  *
 198  * @f: signaled fence
 199  * @cb: our callback structure
 200  *
 201  * Signal the scheduler finished fence when the entity in question is killed.
 202  */
 203 static void drm_sched_entity_kill_jobs_cb(struct dma_fence *f,
 204                                           struct dma_fence_cb *cb)
 205 {
 206         struct drm_sched_job *job = container_of(cb, struct drm_sched_job,
 207                                                  finish_cb);
 208
 209         drm_sched_fence_finished(job->s_fence);
 210         WARN_ON(job->s_fence->parent);
 211         job->sched->ops->free_job(job);
 212 }
 213
 214 /**
 215  * drm_sched_entity_kill_jobs - Make sure all remaining jobs are killed
 216  *
 217  * @entity: entity which is cleaned up
 218  *
 219  * Makes sure that all remaining jobs in an entity are killed before it is
 220  * destroyed.
 221  */
 222 static void drm_sched_entity_kill_jobs(struct drm_sched_entity *entity)
 223 {
 224         struct drm_sched_job *job;
 225         struct dma_fence *f;
 226         int r;
 227
 228         while ((job = to_drm_sched_job(spsc_queue_pop(&entity->job_queue)))) {
 229                 struct drm_sched_fence *s_fence = job->s_fence;
 230
 231                 /* Wait for all dependencies to avoid data corruptions */
 232                 while ((f = job->sched->ops->dependency(job, entity)))
 233                         dma_fence_wait(f, false);
 234
 235                 drm_sched_fence_scheduled(s_fence);
 236                 dma_fence_set_error(&s_fence->finished, -ESRCH);
 237
 238                 /*
 239                  * When pipe is hanged by older entity, new entity might
 240                  * not even have chance to submit it's first job to HW
 241                  * and so entity->last_scheduled will remain NULL
 242                  */
 243                 if (!entity->last_scheduled) {
 244                         drm_sched_entity_kill_jobs_cb(NULL, &job->finish_cb);
 245                         continue;
 246                 }
 247
 248                 r = dma_fence_add_callback(entity->last_scheduled,
 249                                            &job->finish_cb,
 250                                            drm_sched_entity_kill_jobs_cb);
 251                 if (r == -ENOENT)
 252                         drm_sched_entity_kill_jobs_cb(NULL, &job->finish_cb);
 253                 else if (r)
 254                         DRM_ERROR("fence add callback failed (%d)\n", r);
 255         }
 256 }
 257
 258 /**
 259  * drm_sched_entity_fini - Destroy a context entity
 260  *
 261  * @entity: scheduler entity
 262  *
 263  * This should be called after @drm_sched_entity_do_release. It goes over the
 264  * entity and signals all jobs with an error code if the process was killed.
 265  *
 266  */
 267 void drm_sched_entity_fini(struct drm_sched_entity *entity)
 268 {
 269         struct drm_gpu_scheduler *sched = NULL;
 270
 271         if (entity->rq) {
 272                 sched = entity->rq->sched;
 273                 drm_sched_rq_remove_entity(entity->rq, entity);
 274         }
 275
 276         /* Consumption of existing IBs wasn't completed. Forcefully
 277          * remove them here.
 278          */
 279         if (spsc_queue_count(&entity->job_queue)) {
 280                 if (sched) {
 281                         /*
 282                          * Wait for thread to idle to make sure it isn't processing
 283                          * this entity.
 284                          */
 285                         wait_for_completion(&entity->entity_idle);
 286
 287                 }
 288                 if (entity->dependency) {
 289                         dma_fence_remove_callback(entity->dependency,
 290                                                   &entity->cb);
 291                         dma_fence_put(entity->dependency);
 292                         entity->dependency = NULL;
 293                 }
 294
 295                 drm_sched_entity_kill_jobs(entity);
 296         }
 297
 298         dma_fence_put(entity->last_scheduled);
 299         entity->last_scheduled = NULL;
 300 }
 301 EXPORT_SYMBOL(drm_sched_entity_fini);
 302
 303 /**
 304  * drm_sched_entity_destroy - Destroy a context entity
 305  *
 306  * @entity: scheduler entity
 307  *
 308  * Calls drm_sched_entity_do_release() and drm_sched_entity_cleanup()
 309  */
 310 void drm_sched_entity_destroy(struct drm_sched_entity *entity)
 311 {
 312         drm_sched_entity_flush(entity, MAX_WAIT_SCHED_ENTITY_Q_EMPTY);
 313         drm_sched_entity_fini(entity);
 314 }
 315 EXPORT_SYMBOL(drm_sched_entity_destroy);
 316
 317 /*
 318  * drm_sched_entity_clear_dep - callback to clear the entities dependency
 319  */
 320 static void drm_sched_entity_clear_dep(struct dma_fence *f,
 321                                        struct dma_fence_cb *cb)
 322 {
 323         struct drm_sched_entity *entity =
 324                 container_of(cb, struct drm_sched_entity, cb);
 325
 326         entity->dependency = NULL;
 327         dma_fence_put(f);
 328 }
 329
 330 /*
 331  * drm_sched_entity_clear_dep - callback to clear the entities dependency and
 332  * wake up scheduler
 333  */
 334 static void drm_sched_entity_wakeup(struct dma_fence *f,
 335                                     struct dma_fence_cb *cb)
 336 {
 337         struct drm_sched_entity *entity =
 338                 container_of(cb, struct drm_sched_entity, cb);
 339
 340         drm_sched_entity_clear_dep(f, cb);
 341         drm_sched_wakeup(entity->rq->sched);
 342 }
 343
 344 /**
 345  * drm_sched_entity_set_priority - Sets priority of the entity
 346  *
 347  * @entity: scheduler entity
 348  * @priority: scheduler priority
 349  *
 350  * Update the priority of runqueus used for the entity.
 351  */
 352 void drm_sched_entity_set_priority(struct drm_sched_entity *entity,
 353                                    enum drm_sched_priority priority)
 354 {
 355         spin_lock(&entity->rq_lock);
 356         entity->priority = priority;
 357         spin_unlock(&entity->rq_lock);
 358 }
 359 EXPORT_SYMBOL(drm_sched_entity_set_priority);
 360
 361 /**
 362  * drm_sched_entity_add_dependency_cb - add callback for the entities dependency
 363  *
 364  * @entity: entity with dependency
 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
 389                 /*
 390                  * Fence is from the same scheduler, only need to wait for
 391                  * it to be scheduled
 392                  */
 393                 fence = dma_fence_get(&s_fence->scheduled);
 394                 dma_fence_put(entity->dependency);
 395                 entity->dependency = fence;
 396                 if (!dma_fence_add_callback(fence, &entity->cb,
 397                                             drm_sched_entity_clear_dep))
 398                         return true;
 399
 400                 /* Ignore it when it is already scheduled */
 401                 dma_fence_put(fence);
 402                 return false;
 403         }
 404
 405         if (!dma_fence_add_callback(entity->dependency, &entity->cb,
 406                                     drm_sched_entity_wakeup))
 407                 return true;
 408
 409         dma_fence_put(entity->dependency);
 410         return false;
 411 }
 412
 413 /**
 414  * drm_sched_entity_pop_job - get a ready to be scheduled job from the entity
 415  *
 416  * @entity: entity to get the job from
 417  *
 418  * Process all dependencies and try to get one job from the entities queue.
 419  */
 420 struct drm_sched_job *drm_sched_entity_pop_job(struct drm_sched_entity *entity)
 421 {
 422         struct drm_gpu_scheduler *sched = entity->rq->sched;
 423         struct drm_sched_job *sched_job;
 424
 425         sched_job = to_drm_sched_job(spsc_queue_peek(&entity->job_queue));
 426         if (!sched_job)
 427                 return NULL;
 428
 429         while ((entity->dependency =
 430                         sched->ops->dependency(sched_job, entity))) {
 431                 trace_drm_sched_job_wait_dep(sched_job, entity->dependency);
 432
 433                 if (drm_sched_entity_add_dependency_cb(entity))
 434                         return NULL;
 435         }
 436
 437         /* skip jobs from entity that marked guilty */
 438         if (entity->guilty && atomic_read(entity->guilty))
 439                 dma_fence_set_error(&sched_job->s_fence->finished, -ECANCELED);
 440
 441         dma_fence_put(entity->last_scheduled);
 442         entity->last_scheduled = dma_fence_get(&sched_job->s_fence->finished);
 443
 444         spsc_queue_pop(&entity->job_queue);
 445         return sched_job;
 446 }
 447
 448 /**
 449  * drm_sched_entity_select_rq - select a new rq for the entity
 450  *
 451  * @entity: scheduler entity
 452  *
 453  * Check all prerequisites and select a new rq for the entity for load
 454  * balancing.
 455  */
 456 void drm_sched_entity_select_rq(struct drm_sched_entity *entity)
 457 {
 458         struct dma_fence *fence;
 459         struct drm_gpu_scheduler *sched;
 460         struct drm_sched_rq *rq;
 461
 462         if (spsc_queue_count(&entity->job_queue) || !entity->sched_list)
 463                 return;
 464
 465         fence = READ_ONCE(entity->last_scheduled);
 466         if (fence && !dma_fence_is_signaled(fence))
 467                 return;
 468
 469         spin_lock(&entity->rq_lock);
 470         sched = drm_sched_pick_best(entity->sched_list, entity->num_sched_list);
 471         rq = sched ? &sched->sched_rq[entity->priority] : NULL;
 472         if (rq != entity->rq) {
 473                 drm_sched_rq_remove_entity(entity->rq, entity);
 474                 entity->rq = rq;
 475         }
 476         spin_unlock(&entity->rq_lock);
 477
 478         if (entity->num_sched_list == 1)
 479                 entity->sched_list = NULL;
 480 }
 481
 482 /**
 483  * drm_sched_entity_push_job - Submit a job to the entity's job queue
 484  *
 485  * @sched_job: job to submit
 486  * @entity: scheduler entity
 487  *
 488  * Note: To guarantee that the order of insertion to queue matches
 489  * the job's fence sequence number this function should be
 490  * called with drm_sched_job_init under common lock.
 491  *
 492  * Returns 0 for success, negative error code otherwise.
 493  */
 494 void drm_sched_entity_push_job(struct drm_sched_job *sched_job,
 495                                struct drm_sched_entity *entity)
 496 {
 497         bool first;
 498
 499         trace_drm_sched_job(sched_job, entity);
 500         atomic_inc(entity->rq->sched->score);
 501         WRITE_ONCE(entity->last_user, current->group_leader);
 502         first = spsc_queue_push(&entity->job_queue, &sched_job->queue_node);
 503
 504         /* first job wakes up scheduler */
 505         if (first) {
 506                 /* Add the entity to the run queue */
 507                 spin_lock(&entity->rq_lock);
 508                 if (entity->stopped) {
 509                         spin_unlock(&entity->rq_lock);
 510
 511                         DRM_ERROR("Trying to push to a killed entity\n");
 512                         return;
 513                 }
 514                 drm_sched_rq_add_entity(entity->rq, entity);
 515                 spin_unlock(&entity->rq_lock);
 516                 drm_sched_wakeup(entity->rq->sched);
 517         }
 518 }
 519 EXPORT_SYMBOL(drm_sched_entity_push_job);