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  * @rq_list: the list of run queue on which jobs from this
  42  *           entity can be submitted
  43  * @num_rq_list: number of run queue in rq_list
  44  * @guilty: atomic_t set to 1 when a job on this queue
  45  *          is found to be guilty causing a timeout
  46  *
  47  * Note: the rq_list should have atleast one element to schedule
  48  *       the entity
  49  *
  50  * Returns 0 on success or a negative error code on failure.
  51  */
  52 int drm_sched_entity_init(struct drm_sched_entity *entity,
  53                           struct drm_sched_rq **rq_list,
  54                           unsigned int num_rq_list,
  55                           atomic_t *guilty)
  56 {
  57         int i;
  58
  59         if (!(entity && rq_list && (num_rq_list == 0 || rq_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_rq_list = num_rq_list;
  67         entity->rq_list = kcalloc(num_rq_list, sizeof(struct drm_sched_rq *),
  68                                 GFP_KERNEL);
  69         if (!entity->rq_list)
  70                 return -ENOMEM;
  71
  72         init_completion(&entity->entity_idle);
  73
  74         for (i = 0; i < num_rq_list; ++i)
  75                 entity->rq_list[i] = rq_list[i];
  76
  77         if (num_rq_list)
  78                 entity->rq = rq_list[0];
  79
  80         entity->last_scheduled = NULL;
  81
  82         spin_lock_init(&entity->rq_lock);
  83         spsc_queue_init(&entity->job_queue);
  84
  85         atomic_set(&entity->fence_seq, 0);
  86         entity->fence_context = dma_fence_context_alloc(2);
  87
  88         return 0;
  89 }
  90 EXPORT_SYMBOL(drm_sched_entity_init);
  91
  92 /**
  93  * drm_sched_entity_is_idle - Check if entity is idle
  94  *
  95  * @entity: scheduler entity
  96  *
  97  * Returns true if the entity does not have any unscheduled jobs.
  98  */
  99 static bool drm_sched_entity_is_idle(struct drm_sched_entity *entity)
 100 {
 101         rmb(); /* for list_empty to work without lock */
 102
 103         if (list_empty(&entity->list) ||
 104             spsc_queue_count(&entity->job_queue) == 0)
 105                 return true;
 106
 107         return false;
 108 }
 109
 110 /**
 111  * drm_sched_entity_is_ready - Check if entity is ready
 112  *
 113  * @entity: scheduler entity
 114  *
 115  * Return true if entity could provide a job.
 116  */
 117 bool drm_sched_entity_is_ready(struct drm_sched_entity *entity)
 118 {
 119         if (spsc_queue_peek(&entity->job_queue) == NULL)
 120                 return false;
 121
 122         if (READ_ONCE(entity->dependency))
 123                 return false;
 124
 125         return true;
 126 }
 127
 128 /**
 129  * drm_sched_entity_get_free_sched - Get the rq from rq_list with least load
 130  *
 131  * @entity: scheduler entity
 132  *
 133  * Return the pointer to the rq with least load.
 134  */
 135 static struct drm_sched_rq *
 136 drm_sched_entity_get_free_sched(struct drm_sched_entity *entity)
 137 {
 138         struct drm_sched_rq *rq = NULL;
 139         unsigned int min_jobs = UINT_MAX, num_jobs;
 140         int i;
 141
 142         for (i = 0; i < entity->num_rq_list; ++i) {
 143                 struct drm_gpu_scheduler *sched = entity->rq_list[i]->sched;
 144
 145                 if (!entity->rq_list[i]->sched->ready) {
 146                         DRM_WARN("sched%s is not ready, skipping", sched->name);
 147                         continue;
 148                 }
 149
 150                 num_jobs = atomic_read(&sched->num_jobs);
 151                 if (num_jobs < min_jobs) {
 152                         min_jobs = num_jobs;
 153                         rq = entity->rq_list[i];
 154                 }
 155         }
 156
 157         return rq;
 158 }
 159
 160 /**
 161  * drm_sched_entity_flush - Flush a context entity
 162  *
 163  * @entity: scheduler entity
 164  * @timeout: time to wait in for Q to become empty in jiffies.
 165  *
 166  * Splitting drm_sched_entity_fini() into two functions, The first one does the
 167  * waiting, removes the entity from the runqueue and returns an error when the
 168  * process was killed.
 169  *
 170  * Returns the remaining time in jiffies left from the input timeout
 171  */
 172 long drm_sched_entity_flush(struct drm_sched_entity *entity, long timeout)
 173 {
 174         struct drm_gpu_scheduler *sched;
 175         struct task_struct *last_user;
 176         long ret = timeout;
 177
 178         if (!entity->rq)
 179                 return 0;
 180
 181         sched = entity->rq->sched;
 182         /**
 183          * The client will not queue more IBs during this fini, consume existing
 184          * queued IBs or discard them on SIGKILL
 185          */
 186         if (current->flags & PF_EXITING) {
 187                 if (timeout)
 188                         ret = wait_event_timeout(
 189                                         sched->job_scheduled,
 190                                         drm_sched_entity_is_idle(entity),
 191                                         timeout);
 192         } else {
 193                 wait_event_killable(sched->job_scheduled,
 194                                     drm_sched_entity_is_idle(entity));
 195         }
 196
 197         /* For killed process disable any more IBs enqueue right now */
 198         last_user = cmpxchg(&entity->last_user, current->group_leader, NULL);
 199         if ((!last_user || last_user == current->group_leader) &&
 200             (current->flags & PF_EXITING) && (current->exit_code == SIGKILL)) {
 201                 spin_lock(&entity->rq_lock);
 202                 entity->stopped = true;
 203                 drm_sched_rq_remove_entity(entity->rq, entity);
 204                 spin_unlock(&entity->rq_lock);
 205         }
 206
 207         return ret;
 208 }
 209 EXPORT_SYMBOL(drm_sched_entity_flush);
 210
 211 /**
 212  * drm_sched_entity_kill_jobs - helper for drm_sched_entity_kill_jobs
 213  *
 214  * @f: signaled fence
 215  * @cb: our callback structure
 216  *
 217  * Signal the scheduler finished fence when the entity in question is killed.
 218  */
 219 static void drm_sched_entity_kill_jobs_cb(struct dma_fence *f,
 220                                           struct dma_fence_cb *cb)
 221 {
 222         struct drm_sched_job *job = container_of(cb, struct drm_sched_job,
 223                                                  finish_cb);
 224
 225         drm_sched_fence_finished(job->s_fence);
 226         WARN_ON(job->s_fence->parent);
 227         job->sched->ops->free_job(job);
 228 }
 229
 230 /**
 231  * drm_sched_entity_kill_jobs - Make sure all remaining jobs are killed
 232  *
 233  * @entity: entity which is cleaned up
 234  *
 235  * Makes sure that all remaining jobs in an entity are killed before it is
 236  * destroyed.
 237  */
 238 static void drm_sched_entity_kill_jobs(struct drm_sched_entity *entity)
 239 {
 240         struct drm_sched_job *job;
 241         int r;
 242
 243         while ((job = to_drm_sched_job(spsc_queue_pop(&entity->job_queue)))) {
 244                 struct drm_sched_fence *s_fence = job->s_fence;
 245
 246                 drm_sched_fence_scheduled(s_fence);
 247                 dma_fence_set_error(&s_fence->finished, -ESRCH);
 248
 249                 /*
 250                  * When pipe is hanged by older entity, new entity might
 251                  * not even have chance to submit it's first job to HW
 252                  * and so entity->last_scheduled will remain NULL
 253                  */
 254                 if (!entity->last_scheduled) {
 255                         drm_sched_entity_kill_jobs_cb(NULL, &job->finish_cb);
 256                         continue;
 257                 }
 258
 259                 r = dma_fence_add_callback(entity->last_scheduled,
 260                                            &job->finish_cb,
 261                                            drm_sched_entity_kill_jobs_cb);
 262                 if (r == -ENOENT)
 263                         drm_sched_entity_kill_jobs_cb(NULL, &job->finish_cb);
 264                 else if (r)
 265                         DRM_ERROR("fence add callback failed (%d)\n", r);
 266         }
 267 }
 268
 269 /**
 270  * drm_sched_entity_cleanup - Destroy a context entity
 271  *
 272  * @entity: scheduler entity
 273  *
 274  * This should be called after @drm_sched_entity_do_release. It goes over the
 275  * entity and signals all jobs with an error code if the process was killed.
 276  *
 277  */
 278 void drm_sched_entity_fini(struct drm_sched_entity *entity)
 279 {
 280         struct drm_gpu_scheduler *sched = NULL;
 281
 282         if (entity->rq) {
 283                 sched = entity->rq->sched;
 284                 drm_sched_rq_remove_entity(entity->rq, entity);
 285         }
 286
 287         /* Consumption of existing IBs wasn't completed. Forcefully
 288          * remove them here.
 289          */
 290         if (spsc_queue_count(&entity->job_queue)) {
 291                 if (sched) {
 292                         /*
 293                          * Wait for thread to idle to make sure it isn't processing
 294                          * this entity.
 295                          */
 296                         wait_for_completion(&entity->entity_idle);
 297
 298                 }
 299                 if (entity->dependency) {
 300                         dma_fence_remove_callback(entity->dependency,
 301                                                   &entity->cb);
 302                         dma_fence_put(entity->dependency);
 303                         entity->dependency = NULL;
 304                 }
 305
 306                 drm_sched_entity_kill_jobs(entity);
 307         }
 308
 309         dma_fence_put(entity->last_scheduled);
 310         entity->last_scheduled = NULL;
 311         kfree(entity->rq_list);
 312 }
 313 EXPORT_SYMBOL(drm_sched_entity_fini);
 314
 315 /**
 316  * drm_sched_entity_fini - Destroy a context entity
 317  *
 318  * @entity: scheduler entity
 319  *
 320  * Calls drm_sched_entity_do_release() and drm_sched_entity_cleanup()
 321  */
 322 void drm_sched_entity_destroy(struct drm_sched_entity *entity)
 323 {
 324         drm_sched_entity_flush(entity, MAX_WAIT_SCHED_ENTITY_Q_EMPTY);
 325         drm_sched_entity_fini(entity);
 326 }
 327 EXPORT_SYMBOL(drm_sched_entity_destroy);
 328
 329 /**
 330  * drm_sched_entity_clear_dep - callback to clear the entities dependency
 331  */
 332 static void drm_sched_entity_clear_dep(struct dma_fence *f,
 333                                        struct dma_fence_cb *cb)
 334 {
 335         struct drm_sched_entity *entity =
 336                 container_of(cb, struct drm_sched_entity, cb);
 337
 338         entity->dependency = NULL;
 339         dma_fence_put(f);
 340 }
 341
 342 /**
 343  * drm_sched_entity_clear_dep - callback to clear the entities dependency and
 344  * wake up scheduler
 345  */
 346 static void drm_sched_entity_wakeup(struct dma_fence *f,
 347                                     struct dma_fence_cb *cb)
 348 {
 349         struct drm_sched_entity *entity =
 350                 container_of(cb, struct drm_sched_entity, cb);
 351
 352         drm_sched_entity_clear_dep(f, cb);
 353         drm_sched_wakeup(entity->rq->sched);
 354 }
 355
 356 /**
 357  * drm_sched_entity_set_rq_priority - helper for drm_sched_entity_set_priority
 358  */
 359 static void drm_sched_entity_set_rq_priority(struct drm_sched_rq **rq,
 360                                              enum drm_sched_priority priority)
 361 {
 362         *rq = &(*rq)->sched->sched_rq[priority];
 363 }
 364
 365 /**
 366  * drm_sched_entity_set_priority - Sets priority of the entity
 367  *
 368  * @entity: scheduler entity
 369  * @priority: scheduler priority
 370  *
 371  * Update the priority of runqueus used for the entity.
 372  */
 373 void drm_sched_entity_set_priority(struct drm_sched_entity *entity,
 374                                    enum drm_sched_priority priority)
 375 {
 376         unsigned int i;
 377
 378         spin_lock(&entity->rq_lock);
 379
 380         for (i = 0; i < entity->num_rq_list; ++i)
 381                 drm_sched_entity_set_rq_priority(&entity->rq_list[i], priority);
 382
 383         if (entity->rq) {
 384                 drm_sched_rq_remove_entity(entity->rq, entity);
 385                 drm_sched_entity_set_rq_priority(&entity->rq, priority);
 386                 drm_sched_rq_add_entity(entity->rq, entity);
 387         }
 388
 389         spin_unlock(&entity->rq_lock);
 390 }
 391 EXPORT_SYMBOL(drm_sched_entity_set_priority);
 392
 393 /**
 394  * drm_sched_entity_add_dependency_cb - add callback for the entities dependency
 395  *
 396  * @entity: entity with dependency
 397  *
 398  * Add a callback to the current dependency of the entity to wake up the
 399  * scheduler when the entity becomes available.
 400  */
 401 static bool drm_sched_entity_add_dependency_cb(struct drm_sched_entity *entity)
 402 {
 403         struct drm_gpu_scheduler *sched = entity->rq->sched;
 404         struct dma_fence *fence = entity->dependency;
 405         struct drm_sched_fence *s_fence;
 406
 407         if (fence->context == entity->fence_context ||
 408             fence->context == entity->fence_context + 1) {
 409                 /*
 410                  * Fence is a scheduled/finished fence from a job
 411                  * which belongs to the same entity, we can ignore
 412                  * fences from ourself
 413                  */
 414                 dma_fence_put(entity->dependency);
 415                 return false;
 416         }
 417
 418         s_fence = to_drm_sched_fence(fence);
 419         if (s_fence && s_fence->sched == sched) {
 420
 421                 /*
 422                  * Fence is from the same scheduler, only need to wait for
 423                  * it to be scheduled
 424                  */
 425                 fence = dma_fence_get(&s_fence->scheduled);
 426                 dma_fence_put(entity->dependency);
 427                 entity->dependency = fence;
 428                 if (!dma_fence_add_callback(fence, &entity->cb,
 429                                             drm_sched_entity_clear_dep))
 430                         return true;
 431
 432                 /* Ignore it when it is already scheduled */
 433                 dma_fence_put(fence);
 434                 return false;
 435         }
 436
 437         if (!dma_fence_add_callback(entity->dependency, &entity->cb,
 438                                     drm_sched_entity_wakeup))
 439                 return true;
 440
 441         dma_fence_put(entity->dependency);
 442         return false;
 443 }
 444
 445 /**
 446  * drm_sched_entity_pop_job - get a ready to be scheduled job from the entity
 447  *
 448  * @entity: entity to get the job from
 449  *
 450  * Process all dependencies and try to get one job from the entities queue.
 451  */
 452 struct drm_sched_job *drm_sched_entity_pop_job(struct drm_sched_entity *entity)
 453 {
 454         struct drm_gpu_scheduler *sched = entity->rq->sched;
 455         struct drm_sched_job *sched_job;
 456
 457         sched_job = to_drm_sched_job(spsc_queue_peek(&entity->job_queue));
 458         if (!sched_job)
 459                 return NULL;
 460
 461         while ((entity->dependency =
 462                         sched->ops->dependency(sched_job, entity))) {
 463                 trace_drm_sched_job_wait_dep(sched_job, entity->dependency);
 464
 465                 if (drm_sched_entity_add_dependency_cb(entity))
 466                         return NULL;
 467         }
 468
 469         /* skip jobs from entity that marked guilty */
 470         if (entity->guilty && atomic_read(entity->guilty))
 471                 dma_fence_set_error(&sched_job->s_fence->finished, -ECANCELED);
 472
 473         dma_fence_put(entity->last_scheduled);
 474         entity->last_scheduled = dma_fence_get(&sched_job->s_fence->finished);
 475
 476         spsc_queue_pop(&entity->job_queue);
 477         return sched_job;
 478 }
 479
 480 /**
 481  * drm_sched_entity_select_rq - select a new rq for the entity
 482  *
 483  * @entity: scheduler entity
 484  *
 485  * Check all prerequisites and select a new rq for the entity for load
 486  * balancing.
 487  */
 488 void drm_sched_entity_select_rq(struct drm_sched_entity *entity)
 489 {
 490         struct dma_fence *fence;
 491         struct drm_sched_rq *rq;
 492
 493         if (spsc_queue_count(&entity->job_queue) || entity->num_rq_list <= 1)
 494                 return;
 495
 496         fence = READ_ONCE(entity->last_scheduled);
 497         if (fence && !dma_fence_is_signaled(fence))
 498                 return;
 499
 500         rq = drm_sched_entity_get_free_sched(entity);
 501         if (rq == entity->rq)
 502                 return;
 503
 504         spin_lock(&entity->rq_lock);
 505         drm_sched_rq_remove_entity(entity->rq, entity);
 506         entity->rq = rq;
 507         spin_unlock(&entity->rq_lock);
 508 }
 509
 510 /**
 511  * drm_sched_entity_push_job - Submit a job to the entity's job queue
 512  *
 513  * @sched_job: job to submit
 514  * @entity: scheduler entity
 515  *
 516  * Note: To guarantee that the order of insertion to queue matches
 517  * the job's fence sequence number this function should be
 518  * called with drm_sched_job_init under common lock.
 519  *
 520  * Returns 0 for success, negative error code otherwise.
 521  */
 522 void drm_sched_entity_push_job(struct drm_sched_job *sched_job,
 523                                struct drm_sched_entity *entity)
 524 {
 525         bool first;
 526
 527         trace_drm_sched_job(sched_job, entity);
 528         atomic_inc(&entity->rq->sched->num_jobs);
 529         WRITE_ONCE(entity->last_user, current->group_leader);
 530         first = spsc_queue_push(&entity->job_queue, &sched_job->queue_node);
 531
 532         /* first job wakes up scheduler */
 533         if (first) {
 534                 /* Add the entity to the run queue */
 535                 spin_lock(&entity->rq_lock);
 536                 if (entity->stopped) {
 537                         spin_unlock(&entity->rq_lock);
 538
 539                         DRM_ERROR("Trying to push to a killed entity\n");
 540                         return;
 541                 }
 542                 drm_sched_rq_add_entity(entity->rq, entity);
 543                 spin_unlock(&entity->rq_lock);
 544                 drm_sched_wakeup(entity->rq->sched);
 545         }
 546 }
 547 EXPORT_SYMBOL(drm_sched_entity_push_job);