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 that the &sched_list must have at least one element to schedule the entity.
  49  *
  50  * For changing @priority later on at runtime see
  51  * drm_sched_entity_set_priority(). For changing the set of schedulers
  52  * @sched_list at runtime see drm_sched_entity_modify_sched().
  53  *
  54  * An entity is cleaned up by callind drm_sched_entity_fini(). See also
  55  * drm_sched_entity_destroy().
  56  *
  57  * Returns 0 on success or a negative error code on failure.
  58  */
  59 int drm_sched_entity_init(struct drm_sched_entity *entity,
  60                           enum drm_sched_priority priority,
  61                           struct drm_gpu_scheduler **sched_list,
  62                           unsigned int num_sched_list,
  63                           atomic_t *guilty)
  64 {
  65         if (!(entity && sched_list && (num_sched_list == 0 || sched_list[0])))
  66                 return -EINVAL;
  67
  68         memset(entity, 0, sizeof(struct drm_sched_entity));
  69         INIT_LIST_HEAD(&entity->list);
  70         entity->rq = NULL;
  71         entity->guilty = guilty;
  72         entity->num_sched_list = num_sched_list;
  73         entity->priority = priority;
  74         entity->sched_list = num_sched_list > 1 ? sched_list : NULL;
  75         entity->last_scheduled = NULL;
  76
  77         if(num_sched_list)
  78                 entity->rq = &sched_list[0]->sched_rq[entity->priority];
  79
  80         init_completion(&entity->entity_idle);
  81
  82         /* We start in an idle state. */
  83         complete(&entity->entity_idle);
  84
  85         spin_lock_init(&entity->rq_lock);
  86         spsc_queue_init(&entity->job_queue);
  87
  88         atomic_set(&entity->fence_seq, 0);
  89         entity->fence_context = dma_fence_context_alloc(2);
  90
  91         return 0;
  92 }
  93 EXPORT_SYMBOL(drm_sched_entity_init);
  94
  95 /**
  96  * drm_sched_entity_modify_sched - Modify sched of an entity
  97  * @entity: scheduler entity to init
  98  * @sched_list: the list of new drm scheds which will replace
  99  *               existing entity->sched_list
 100  * @num_sched_list: number of drm sched in sched_list
 101  *
 102  * Note that this must be called under the same common lock for @entity as
 103  * drm_sched_job_arm() and drm_sched_entity_push_job(), or the driver needs to
 104  * guarantee through some other means that this is never called while new jobs
 105  * can be pushed to @entity.
 106  */
 107 void drm_sched_entity_modify_sched(struct drm_sched_entity *entity,
 108                                     struct drm_gpu_scheduler **sched_list,
 109                                     unsigned int num_sched_list)
 110 {
 111         WARN_ON(!num_sched_list || !sched_list);
 112
 113         entity->sched_list = sched_list;
 114         entity->num_sched_list = num_sched_list;
 115 }
 116 EXPORT_SYMBOL(drm_sched_entity_modify_sched);
 117
 118 static bool drm_sched_entity_is_idle(struct drm_sched_entity *entity)
 119 {
 120         rmb(); /* for list_empty to work without lock */
 121
 122         if (list_empty(&entity->list) ||
 123             spsc_queue_count(&entity->job_queue) == 0 ||
 124             entity->stopped)
 125                 return true;
 126
 127         return false;
 128 }
 129
 130 /* Return true if entity could provide a job. */
 131 bool drm_sched_entity_is_ready(struct drm_sched_entity *entity)
 132 {
 133         if (spsc_queue_peek(&entity->job_queue) == NULL)
 134                 return false;
 135
 136         if (READ_ONCE(entity->dependency))
 137                 return false;
 138
 139         return true;
 140 }
 141
 142 /**
 143  * drm_sched_entity_flush - Flush a context entity
 144  *
 145  * @entity: scheduler entity
 146  * @timeout: time to wait in for Q to become empty in jiffies.
 147  *
 148  * Splitting drm_sched_entity_fini() into two functions, The first one does the
 149  * waiting, removes the entity from the runqueue and returns an error when the
 150  * process was killed.
 151  *
 152  * Returns the remaining time in jiffies left from the input timeout
 153  */
 154 long drm_sched_entity_flush(struct drm_sched_entity *entity, long timeout)
 155 {
 156         struct drm_gpu_scheduler *sched;
 157         struct task_struct *last_user;
 158         long ret = timeout;
 159
 160         if (!entity->rq)
 161                 return 0;
 162
 163         sched = entity->rq->sched;
 164         /**
 165          * The client will not queue more IBs during this fini, consume existing
 166          * queued IBs or discard them on SIGKILL
 167          */
 168         if (current->flags & PF_EXITING) {
 169                 if (timeout)
 170                         ret = wait_event_timeout(
 171                                         sched->job_scheduled,
 172                                         drm_sched_entity_is_idle(entity),
 173                                         timeout);
 174         } else {
 175                 wait_event_killable(sched->job_scheduled,
 176                                     drm_sched_entity_is_idle(entity));
 177         }
 178
 179         /* For killed process disable any more IBs enqueue right now */
 180         last_user = cmpxchg(&entity->last_user, current->group_leader, NULL);
 181         if ((!last_user || last_user == current->group_leader) &&
 182             (current->flags & PF_EXITING) && (current->exit_code == SIGKILL)) {
 183                 spin_lock(&entity->rq_lock);
 184                 entity->stopped = true;
 185                 drm_sched_rq_remove_entity(entity->rq, entity);
 186                 spin_unlock(&entity->rq_lock);
 187         }
 188
 189         return ret;
 190 }
 191 EXPORT_SYMBOL(drm_sched_entity_flush);
 192
 193 static void drm_sched_entity_kill_jobs_irq_work(struct irq_work *wrk)
 194 {
 195         struct drm_sched_job *job = container_of(wrk, typeof(*job), work);
 196
 197         drm_sched_fence_finished(job->s_fence);
 198         WARN_ON(job->s_fence->parent);
 199         job->sched->ops->free_job(job);
 200 }
 201
 202
 203 /* Signal the scheduler finished fence when the entity in question is killed. */
 204 static void drm_sched_entity_kill_jobs_cb(struct dma_fence *f,
 205                                           struct dma_fence_cb *cb)
 206 {
 207         struct drm_sched_job *job = container_of(cb, struct drm_sched_job,
 208                                                  finish_cb);
 209
 210         init_irq_work(&job->work, drm_sched_entity_kill_jobs_irq_work);
 211         irq_work_queue(&job->work);
 212 }
 213
 214 static struct dma_fence *
 215 drm_sched_job_dependency(struct drm_sched_job *job,
 216                          struct drm_sched_entity *entity)
 217 {
 218         if (!xa_empty(&job->dependencies))
 219                 return xa_erase(&job->dependencies, job->last_dependency++);
 220
 221         if (job->sched->ops->dependency)
 222                 return job->sched->ops->dependency(job, entity);
 223
 224         return NULL;
 225 }
 226
 227 static void drm_sched_entity_kill_jobs(struct drm_sched_entity *entity)
 228 {
 229         struct drm_sched_job *job;
 230         struct dma_fence *f;
 231         int r;
 232
 233         while ((job = to_drm_sched_job(spsc_queue_pop(&entity->job_queue)))) {
 234                 struct drm_sched_fence *s_fence = job->s_fence;
 235
 236                 /* Wait for all dependencies to avoid data corruptions */
 237                 while ((f = drm_sched_job_dependency(job, entity)))
 238                         dma_fence_wait(f, false);
 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_fini - Destroy a context entity
 265  *
 266  * @entity: scheduler entity
 267  *
 268  * Cleanups up @entity which has been initialized by drm_sched_entity_init().
 269  *
 270  * If there are potentially job still in flight or getting newly queued
 271  * drm_sched_entity_flush() must be called first. This function then goes over
 272  * the entity and signals all jobs with an error code if the process was killed.
 273  */
 274 void drm_sched_entity_fini(struct drm_sched_entity *entity)
 275 {
 276         struct drm_gpu_scheduler *sched = NULL;
 277
 278         if (entity->rq) {
 279                 sched = entity->rq->sched;
 280                 drm_sched_rq_remove_entity(entity->rq, entity);
 281         }
 282
 283         /* Consumption of existing IBs wasn't completed. Forcefully
 284          * remove them here.
 285          */
 286         if (spsc_queue_count(&entity->job_queue)) {
 287                 if (sched) {
 288                         /*
 289                          * Wait for thread to idle to make sure it isn't processing
 290                          * this entity.
 291                          */
 292                         wait_for_completion(&entity->entity_idle);
 293
 294                 }
 295                 if (entity->dependency) {
 296                         dma_fence_remove_callback(entity->dependency,
 297                                                   &entity->cb);
 298                         dma_fence_put(entity->dependency);
 299                         entity->dependency = NULL;
 300                 }
 301
 302                 drm_sched_entity_kill_jobs(entity);
 303         }
 304
 305         dma_fence_put(entity->last_scheduled);
 306         entity->last_scheduled = NULL;
 307 }
 308 EXPORT_SYMBOL(drm_sched_entity_fini);
 309
 310 /**
 311  * drm_sched_entity_destroy - Destroy a context entity
 312  * @entity: scheduler entity
 313  *
 314  * Calls drm_sched_entity_flush() and drm_sched_entity_fini() as a
 315  * convenience wrapper.
 316  */
 317 void drm_sched_entity_destroy(struct drm_sched_entity *entity)
 318 {
 319         drm_sched_entity_flush(entity, MAX_WAIT_SCHED_ENTITY_Q_EMPTY);
 320         drm_sched_entity_fini(entity);
 321 }
 322 EXPORT_SYMBOL(drm_sched_entity_destroy);
 323
 324 /* drm_sched_entity_clear_dep - callback to clear the entities dependency */
 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  * Add a callback to the current dependency of the entity to wake up the
 368  * scheduler when the entity becomes available.
 369  */
 370 static bool drm_sched_entity_add_dependency_cb(struct drm_sched_entity *entity)
 371 {
 372         struct drm_gpu_scheduler *sched = entity->rq->sched;
 373         struct dma_fence *fence = entity->dependency;
 374         struct drm_sched_fence *s_fence;
 375
 376         if (fence->context == entity->fence_context ||
 377             fence->context == entity->fence_context + 1) {
 378                 /*
 379                  * Fence is a scheduled/finished fence from a job
 380                  * which belongs to the same entity, we can ignore
 381                  * fences from ourself
 382                  */
 383                 dma_fence_put(entity->dependency);
 384                 return false;
 385         }
 386
 387         s_fence = to_drm_sched_fence(fence);
 388         if (s_fence && s_fence->sched == sched) {
 389
 390                 /*
 391                  * Fence is from the same scheduler, only need to wait for
 392                  * it to be scheduled
 393                  */
 394                 fence = dma_fence_get(&s_fence->scheduled);
 395                 dma_fence_put(entity->dependency);
 396                 entity->dependency = fence;
 397                 if (!dma_fence_add_callback(fence, &entity->cb,
 398                                             drm_sched_entity_clear_dep))
 399                         return true;
 400
 401                 /* Ignore it when it is already scheduled */
 402                 dma_fence_put(fence);
 403                 return false;
 404         }
 405
 406         if (!dma_fence_add_callback(entity->dependency, &entity->cb,
 407                                     drm_sched_entity_wakeup))
 408                 return true;
 409
 410         dma_fence_put(entity->dependency);
 411         return false;
 412 }
 413
 414 struct drm_sched_job *drm_sched_entity_pop_job(struct drm_sched_entity *entity)
 415 {
 416         struct drm_sched_job *sched_job;
 417
 418         sched_job = to_drm_sched_job(spsc_queue_peek(&entity->job_queue));
 419         if (!sched_job)
 420                 return NULL;
 421
 422         while ((entity->dependency =
 423                         drm_sched_job_dependency(sched_job, entity))) {
 424                 trace_drm_sched_job_wait_dep(sched_job, entity->dependency);
 425
 426                 if (drm_sched_entity_add_dependency_cb(entity))
 427                         return NULL;
 428         }
 429
 430         /* skip jobs from entity that marked guilty */
 431         if (entity->guilty && atomic_read(entity->guilty))
 432                 dma_fence_set_error(&sched_job->s_fence->finished, -ECANCELED);
 433
 434         dma_fence_put(entity->last_scheduled);
 435
 436         entity->last_scheduled = dma_fence_get(&sched_job->s_fence->finished);
 437
 438         /*
 439          * If the queue is empty we allow drm_sched_entity_select_rq() to
 440          * locklessly access ->last_scheduled. This only works if we set the
 441          * pointer before we dequeue and if we a write barrier here.
 442          */
 443         smp_wmb();
 444
 445         spsc_queue_pop(&entity->job_queue);
 446         return sched_job;
 447 }
 448
 449 void drm_sched_entity_select_rq(struct drm_sched_entity *entity)
 450 {
 451         struct dma_fence *fence;
 452         struct drm_gpu_scheduler *sched;
 453         struct drm_sched_rq *rq;
 454
 455         /* single possible engine and already selected */
 456         if (!entity->sched_list)
 457                 return;
 458
 459         /* queue non-empty, stay on the same engine */
 460         if (spsc_queue_count(&entity->job_queue))
 461                 return;
 462
 463         /*
 464          * Only when the queue is empty are we guaranteed that the scheduler
 465          * thread cannot change ->last_scheduled. To enforce ordering we need
 466          * a read barrier here. See drm_sched_entity_pop_job() for the other
 467          * side.
 468          */
 469         smp_rmb();
 470
 471         fence = entity->last_scheduled;
 472
 473         /* stay on the same engine if the previous job hasn't finished */
 474         if (fence && !dma_fence_is_signaled(fence))
 475                 return;
 476
 477         spin_lock(&entity->rq_lock);
 478         sched = drm_sched_pick_best(entity->sched_list, entity->num_sched_list);
 479         rq = sched ? &sched->sched_rq[entity->priority] : NULL;
 480         if (rq != entity->rq) {
 481                 drm_sched_rq_remove_entity(entity->rq, entity);
 482                 entity->rq = rq;
 483         }
 484         spin_unlock(&entity->rq_lock);
 485
 486         if (entity->num_sched_list == 1)
 487                 entity->sched_list = NULL;
 488 }
 489
 490 /**
 491  * drm_sched_entity_push_job - Submit a job to the entity's job queue
 492  * @sched_job: job to submit
 493  *
 494  * Note: To guarantee that the order of insertion to queue matches the job's
 495  * fence sequence number this function should be called with drm_sched_job_arm()
 496  * under common lock for the struct drm_sched_entity that was set up for
 497  * @sched_job in drm_sched_job_init().
 498  *
 499  * Returns 0 for success, negative error code otherwise.
 500  */
 501 void drm_sched_entity_push_job(struct drm_sched_job *sched_job)
 502 {
 503         struct drm_sched_entity *entity = sched_job->entity;
 504         bool first;
 505
 506         trace_drm_sched_job(sched_job, entity);
 507         atomic_inc(entity->rq->sched->score);
 508         WRITE_ONCE(entity->last_user, current->group_leader);
 509         first = spsc_queue_push(&entity->job_queue, &sched_job->queue_node);
 510
 511         /* first job wakes up scheduler */
 512         if (first) {
 513                 /* Add the entity to the run queue */
 514                 spin_lock(&entity->rq_lock);
 515                 if (entity->stopped) {
 516                         spin_unlock(&entity->rq_lock);
 517
 518                         DRM_ERROR("Trying to push to a killed entity\n");
 519                         return;
 520                 }
 521                 drm_sched_rq_add_entity(entity->rq, entity);
 522                 spin_unlock(&entity->rq_lock);
 523                 drm_sched_wakeup(entity->rq->sched);
 524         }
 525 }
 526 EXPORT_SYMBOL(drm_sched_entity_push_job);