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_is_idle - Check if entity is idle
  88  *
  89  * @entity: scheduler entity
  90  *
  91  * Returns true if the entity does not have any unscheduled jobs.
  92  */
  93 static bool drm_sched_entity_is_idle(struct drm_sched_entity *entity)
  94 {
  95         rmb(); /* for list_empty to work without lock */
  96
  97         if (list_empty(&entity->list) ||
  98             spsc_queue_count(&entity->job_queue) == 0)
  99                 return true;
 100
 101         return false;
 102 }
 103
 104 /**
 105  * drm_sched_entity_is_ready - Check if entity is ready
 106  *
 107  * @entity: scheduler entity
 108  *
 109  * Return true if entity could provide a job.
 110  */
 111 bool drm_sched_entity_is_ready(struct drm_sched_entity *entity)
 112 {
 113         if (spsc_queue_peek(&entity->job_queue) == NULL)
 114                 return false;
 115
 116         if (READ_ONCE(entity->dependency))
 117                 return false;
 118
 119         return true;
 120 }
 121
 122 /**
 123  * drm_sched_entity_get_free_sched - Get the rq from rq_list with least load
 124  *
 125  * @entity: scheduler entity
 126  *
 127  * Return the pointer to the rq with least load.
 128  */
 129 static struct drm_sched_rq *
 130 drm_sched_entity_get_free_sched(struct drm_sched_entity *entity)
 131 {
 132         struct drm_sched_rq *rq = NULL;
 133         unsigned int min_score = UINT_MAX, num_score;
 134         int i;
 135
 136         for (i = 0; i < entity->num_sched_list; ++i) {
 137                 struct drm_gpu_scheduler *sched = entity->sched_list[i];
 138
 139                 if (!entity->sched_list[i]->ready) {
 140                         DRM_WARN("sched%s is not ready, skipping", sched->name);
 141                         continue;
 142                 }
 143
 144                 num_score = atomic_read(&sched->score);
 145                 if (num_score < min_score) {
 146                         min_score = num_score;
 147                         rq = &entity->sched_list[i]->sched_rq[entity->priority];
 148                 }
 149         }
 150
 151         return rq;
 152 }
 153
 154 /**
 155  * drm_sched_entity_flush - Flush a context entity
 156  *
 157  * @entity: scheduler entity
 158  * @timeout: time to wait in for Q to become empty in jiffies.
 159  *
 160  * Splitting drm_sched_entity_fini() into two functions, The first one does the
 161  * waiting, removes the entity from the runqueue and returns an error when the
 162  * process was killed.
 163  *
 164  * Returns the remaining time in jiffies left from the input timeout
 165  */
 166 long drm_sched_entity_flush(struct drm_sched_entity *entity, long timeout)
 167 {
 168         struct drm_gpu_scheduler *sched;
 169         struct task_struct *last_user;
 170         long ret = timeout;
 171
 172         if (!entity->rq)
 173                 return 0;
 174
 175         sched = entity->rq->sched;
 176         /**
 177          * The client will not queue more IBs during this fini, consume existing
 178          * queued IBs or discard them on SIGKILL
 179          */
 180         if (current->flags & PF_EXITING) {
 181                 if (timeout)
 182                         ret = wait_event_timeout(
 183                                         sched->job_scheduled,
 184                                         drm_sched_entity_is_idle(entity),
 185                                         timeout);
 186         } else {
 187                 wait_event_killable(sched->job_scheduled,
 188                                     drm_sched_entity_is_idle(entity));
 189         }
 190
 191         /* For killed process disable any more IBs enqueue right now */
 192         last_user = cmpxchg(&entity->last_user, current->group_leader, NULL);
 193         if ((!last_user || last_user == current->group_leader) &&
 194             (current->flags & PF_EXITING) && (current->exit_code == SIGKILL)) {
 195                 spin_lock(&entity->rq_lock);
 196                 entity->stopped = true;
 197                 drm_sched_rq_remove_entity(entity->rq, entity);
 198                 spin_unlock(&entity->rq_lock);
 199         }
 200
 201         return ret;
 202 }
 203 EXPORT_SYMBOL(drm_sched_entity_flush);
 204
 205 /**
 206  * drm_sched_entity_kill_jobs - helper for drm_sched_entity_kill_jobs
 207  *
 208  * @f: signaled fence
 209  * @cb: our callback structure
 210  *
 211  * Signal the scheduler finished fence when the entity in question is killed.
 212  */
 213 static void drm_sched_entity_kill_jobs_cb(struct dma_fence *f,
 214                                           struct dma_fence_cb *cb)
 215 {
 216         struct drm_sched_job *job = container_of(cb, struct drm_sched_job,
 217                                                  finish_cb);
 218
 219         drm_sched_fence_finished(job->s_fence);
 220         WARN_ON(job->s_fence->parent);
 221         job->sched->ops->free_job(job);
 222 }
 223
 224 /**
 225  * drm_sched_entity_kill_jobs - Make sure all remaining jobs are killed
 226  *
 227  * @entity: entity which is cleaned up
 228  *
 229  * Makes sure that all remaining jobs in an entity are killed before it is
 230  * destroyed.
 231  */
 232 static void drm_sched_entity_kill_jobs(struct drm_sched_entity *entity)
 233 {
 234         struct drm_sched_job *job;
 235         int r;
 236
 237         while ((job = to_drm_sched_job(spsc_queue_pop(&entity->job_queue)))) {
 238                 struct drm_sched_fence *s_fence = job->s_fence;
 239
 240                 drm_sched_fence_scheduled(s_fence);
 241                 dma_fence_set_error(&s_fence->finished, -ESRCH);
 242
 243                 /*
 244                  * When pipe is hanged by older entity, new entity might
 245                  * not even have chance to submit it's first job to HW
 246                  * and so entity->last_scheduled will remain NULL
 247                  */
 248                 if (!entity->last_scheduled) {
 249                         drm_sched_entity_kill_jobs_cb(NULL, &job->finish_cb);
 250                         continue;
 251                 }
 252
 253                 r = dma_fence_add_callback(entity->last_scheduled,
 254                                            &job->finish_cb,
 255                                            drm_sched_entity_kill_jobs_cb);
 256                 if (r == -ENOENT)
 257                         drm_sched_entity_kill_jobs_cb(NULL, &job->finish_cb);
 258                 else if (r)
 259                         DRM_ERROR("fence add callback failed (%d)\n", r);
 260         }
 261 }
 262
 263 /**
 264  * drm_sched_entity_cleanup - Destroy a context entity
 265  *
 266  * @entity: scheduler entity
 267  *
 268  * This should be called after @drm_sched_entity_do_release. It goes over the
 269  * entity and signals all jobs with an error code if the process was killed.
 270  *
 271  */
 272 void drm_sched_entity_fini(struct drm_sched_entity *entity)
 273 {
 274         struct drm_gpu_scheduler *sched = NULL;
 275
 276         if (entity->rq) {
 277                 sched = entity->rq->sched;
 278                 drm_sched_rq_remove_entity(entity->rq, entity);
 279         }
 280
 281         /* Consumption of existing IBs wasn't completed. Forcefully
 282          * remove them here.
 283          */
 284         if (spsc_queue_count(&entity->job_queue)) {
 285                 if (sched) {
 286                         /*
 287                          * Wait for thread to idle to make sure it isn't processing
 288                          * this entity.
 289                          */
 290                         wait_for_completion(&entity->entity_idle);
 291
 292                 }
 293                 if (entity->dependency) {
 294                         dma_fence_remove_callback(entity->dependency,
 295                                                   &entity->cb);
 296                         dma_fence_put(entity->dependency);
 297                         entity->dependency = NULL;
 298                 }
 299
 300                 drm_sched_entity_kill_jobs(entity);
 301         }
 302
 303         dma_fence_put(entity->last_scheduled);
 304         entity->last_scheduled = NULL;
 305 }
 306 EXPORT_SYMBOL(drm_sched_entity_fini);
 307
 308 /**
 309  * drm_sched_entity_fini - Destroy a context entity
 310  *
 311  * @entity: scheduler entity
 312  *
 313  * Calls drm_sched_entity_do_release() and drm_sched_entity_cleanup()
 314  */
 315 void drm_sched_entity_destroy(struct drm_sched_entity *entity)
 316 {
 317         drm_sched_entity_flush(entity, MAX_WAIT_SCHED_ENTITY_Q_EMPTY);
 318         drm_sched_entity_fini(entity);
 319 }
 320 EXPORT_SYMBOL(drm_sched_entity_destroy);
 321
 322 /**
 323  * drm_sched_entity_clear_dep - callback to clear the entities dependency
 324  */
 325 static void drm_sched_entity_clear_dep(struct dma_fence *f,
 326                                        struct dma_fence_cb *cb)
 327 {
 328         struct drm_sched_entity *entity =
 329                 container_of(cb, struct drm_sched_entity, cb);
 330
 331         entity->dependency = NULL;
 332         dma_fence_put(f);
 333 }
 334
 335 /**
 336  * drm_sched_entity_clear_dep - callback to clear the entities dependency and
 337  * wake up scheduler
 338  */
 339 static void drm_sched_entity_wakeup(struct dma_fence *f,
 340                                     struct dma_fence_cb *cb)
 341 {
 342         struct drm_sched_entity *entity =
 343                 container_of(cb, struct drm_sched_entity, cb);
 344
 345         drm_sched_entity_clear_dep(f, cb);
 346         drm_sched_wakeup(entity->rq->sched);
 347 }
 348
 349 /**
 350  * drm_sched_entity_set_priority - Sets priority of the entity
 351  *
 352  * @entity: scheduler entity
 353  * @priority: scheduler priority
 354  *
 355  * Update the priority of runqueus used for the entity.
 356  */
 357 void drm_sched_entity_set_priority(struct drm_sched_entity *entity,
 358                                    enum drm_sched_priority priority)
 359 {
 360         spin_lock(&entity->rq_lock);
 361         entity->priority = priority;
 362         spin_unlock(&entity->rq_lock);
 363 }
 364 EXPORT_SYMBOL(drm_sched_entity_set_priority);
 365
 366 /**
 367  * drm_sched_entity_add_dependency_cb - add callback for the entities dependency
 368  *
 369  * @entity: entity with dependency
 370  *
 371  * Add a callback to the current dependency of the entity to wake up the
 372  * scheduler when the entity becomes available.
 373  */
 374 static bool drm_sched_entity_add_dependency_cb(struct drm_sched_entity *entity)
 375 {
 376         struct drm_gpu_scheduler *sched = entity->rq->sched;
 377         struct dma_fence *fence = entity->dependency;
 378         struct drm_sched_fence *s_fence;
 379
 380         if (fence->context == entity->fence_context ||
 381             fence->context == entity->fence_context + 1) {
 382                 /*
 383                  * Fence is a scheduled/finished fence from a job
 384                  * which belongs to the same entity, we can ignore
 385                  * fences from ourself
 386                  */
 387                 dma_fence_put(entity->dependency);
 388                 return false;
 389         }
 390
 391         s_fence = to_drm_sched_fence(fence);
 392         if (s_fence && s_fence->sched == sched) {
 393
 394                 /*
 395                  * Fence is from the same scheduler, only need to wait for
 396                  * it to be scheduled
 397                  */
 398                 fence = dma_fence_get(&s_fence->scheduled);
 399                 dma_fence_put(entity->dependency);
 400                 entity->dependency = fence;
 401                 if (!dma_fence_add_callback(fence, &entity->cb,
 402                                             drm_sched_entity_clear_dep))
 403                         return true;
 404
 405                 /* Ignore it when it is already scheduled */
 406                 dma_fence_put(fence);
 407                 return false;
 408         }
 409
 410         if (!dma_fence_add_callback(entity->dependency, &entity->cb,
 411                                     drm_sched_entity_wakeup))
 412                 return true;
 413
 414         dma_fence_put(entity->dependency);
 415         return false;
 416 }
 417
 418 /**
 419  * drm_sched_entity_pop_job - get a ready to be scheduled job from the entity
 420  *
 421  * @entity: entity to get the job from
 422  *
 423  * Process all dependencies and try to get one job from the entities queue.
 424  */
 425 struct drm_sched_job *drm_sched_entity_pop_job(struct drm_sched_entity *entity)
 426 {
 427         struct drm_gpu_scheduler *sched = entity->rq->sched;
 428         struct drm_sched_job *sched_job;
 429
 430         sched_job = to_drm_sched_job(spsc_queue_peek(&entity->job_queue));
 431         if (!sched_job)
 432                 return NULL;
 433
 434         while ((entity->dependency =
 435                         sched->ops->dependency(sched_job, entity))) {
 436                 trace_drm_sched_job_wait_dep(sched_job, entity->dependency);
 437
 438                 if (drm_sched_entity_add_dependency_cb(entity))
 439                         return NULL;
 440         }
 441
 442         /* skip jobs from entity that marked guilty */
 443         if (entity->guilty && atomic_read(entity->guilty))
 444                 dma_fence_set_error(&sched_job->s_fence->finished, -ECANCELED);
 445
 446         dma_fence_put(entity->last_scheduled);
 447         entity->last_scheduled = dma_fence_get(&sched_job->s_fence->finished);
 448
 449         spsc_queue_pop(&entity->job_queue);
 450         return sched_job;
 451 }
 452
 453 /**
 454  * drm_sched_entity_select_rq - select a new rq for the entity
 455  *
 456  * @entity: scheduler entity
 457  *
 458  * Check all prerequisites and select a new rq for the entity for load
 459  * balancing.
 460  */
 461 void drm_sched_entity_select_rq(struct drm_sched_entity *entity)
 462 {
 463         struct dma_fence *fence;
 464         struct drm_sched_rq *rq;
 465
 466         if (spsc_queue_count(&entity->job_queue) || entity->num_sched_list <= 1)
 467                 return;
 468
 469         fence = READ_ONCE(entity->last_scheduled);
 470         if (fence && !dma_fence_is_signaled(fence))
 471                 return;
 472
 473         spin_lock(&entity->rq_lock);
 474         rq = drm_sched_entity_get_free_sched(entity);
 475         if (rq != entity->rq) {
 476                 drm_sched_rq_remove_entity(entity->rq, entity);
 477                 entity->rq = rq;
 478         }
 479
 480         spin_unlock(&entity->rq_lock);
 481 }
 482
 483 /**
 484  * drm_sched_entity_push_job - Submit a job to the entity's job queue
 485  *
 486  * @sched_job: job to submit
 487  * @entity: scheduler entity
 488  *
 489  * Note: To guarantee that the order of insertion to queue matches
 490  * the job's fence sequence number this function should be
 491  * called with drm_sched_job_init under common lock.
 492  *
 493  * Returns 0 for success, negative error code otherwise.
 494  */
 495 void drm_sched_entity_push_job(struct drm_sched_job *sched_job,
 496                                struct drm_sched_entity *entity)
 497 {
 498         bool first;
 499
 500         trace_drm_sched_job(sched_job, entity);
 501         atomic_inc(&entity->rq->sched->score);
 502         WRITE_ONCE(entity->last_user, current->group_leader);
 503         first = spsc_queue_push(&entity->job_queue, &sched_job->queue_node);
 504
 505         /* first job wakes up scheduler */
 506         if (first) {
 507                 /* Add the entity to the run queue */
 508                 spin_lock(&entity->rq_lock);
 509                 if (entity->stopped) {
 510                         spin_unlock(&entity->rq_lock);
 511
 512                         DRM_ERROR("Trying to push to a killed entity\n");
 513                         return;
 514                 }
 515                 drm_sched_rq_add_entity(entity->rq, entity);
 516                 spin_unlock(&entity->rq_lock);
 517                 drm_sched_wakeup(entity->rq->sched);
 518         }
 519 }
 520 EXPORT_SYMBOL(drm_sched_entity_push_job);