627 lines · c
1/*2 * Copyright 2015 Advanced Micro Devices, Inc.3 *4 * Permission is hereby granted, free of charge, to any person obtaining a5 * copy of this software and associated documentation files (the "Software"),6 * to deal in the Software without restriction, including without limitation7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,8 * and/or sell copies of the Software, and to permit persons to whom the9 * Software is furnished to do so, subject to the following conditions:10 *11 * The above copyright notice and this permission notice shall be included in12 * all copies or substantial portions of the Software.13 *14 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR15 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,16 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL17 * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR18 * 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 OR20 * 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 when38 * submit to HW ring.39 *40 * @entity: scheduler entity to init41 * @priority: priority of the entity42 * @sched_list: the list of drm scheds on which jobs from this43 * entity can be submitted44 * @num_sched_list: number of drm sched in sched_list45 * @guilty: atomic_t set to 1 when a job on this queue46 * is found to be guilty causing a timeout47 *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 see51 * drm_sched_entity_set_priority(). For changing the set of schedulers52 * @sched_list at runtime see drm_sched_entity_modify_sched().53 *54 * An entity is cleaned up by callind drm_sched_entity_fini(). See also55 * drm_sched_entity_destroy().56 *57 * Returns 0 on success or a negative error code on failure.58 */59int 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 /*75 * It's perfectly valid to initialize an entity without having a valid76 * scheduler attached. It's just not valid to use the scheduler before it77 * is initialized itself.78 */79 entity->sched_list = num_sched_list > 1 ? sched_list : NULL;80 RCU_INIT_POINTER(entity->last_scheduled, NULL);81 RB_CLEAR_NODE(&entity->rb_tree_node);82 83 if (num_sched_list && !sched_list[0]->sched_rq) {84 /* Since every entry covered by num_sched_list85 * should be non-NULL and therefore we warn drivers86 * not to do this and to fix their DRM calling order.87 */88 pr_warn("%s: called with uninitialized scheduler\n", __func__);89 } else if (num_sched_list) {90 /* The "priority" of an entity cannot exceed the number of run-queues of a91 * scheduler. Protect against num_rqs being 0, by converting to signed. Choose92 * the lowest priority available.93 */94 if (entity->priority >= sched_list[0]->num_rqs) {95 drm_err(sched_list[0], "entity with out-of-bounds priority:%u num_rqs:%u\n",96 entity->priority, sched_list[0]->num_rqs);97 entity->priority = max_t(s32, (s32) sched_list[0]->num_rqs - 1,98 (s32) DRM_SCHED_PRIORITY_KERNEL);99 }100 entity->rq = sched_list[0]->sched_rq[entity->priority];101 }102 103 init_completion(&entity->entity_idle);104 105 /* We start in an idle state. */106 complete_all(&entity->entity_idle);107 108 spin_lock_init(&entity->rq_lock);109 spsc_queue_init(&entity->job_queue);110 111 atomic_set(&entity->fence_seq, 0);112 entity->fence_context = dma_fence_context_alloc(2);113 114 return 0;115}116EXPORT_SYMBOL(drm_sched_entity_init);117 118/**119 * drm_sched_entity_modify_sched - Modify sched of an entity120 * @entity: scheduler entity to init121 * @sched_list: the list of new drm scheds which will replace122 * existing entity->sched_list123 * @num_sched_list: number of drm sched in sched_list124 *125 * Note that this must be called under the same common lock for @entity as126 * drm_sched_job_arm() and drm_sched_entity_push_job(), or the driver needs to127 * guarantee through some other means that this is never called while new jobs128 * can be pushed to @entity.129 */130void drm_sched_entity_modify_sched(struct drm_sched_entity *entity,131 struct drm_gpu_scheduler **sched_list,132 unsigned int num_sched_list)133{134 WARN_ON(!num_sched_list || !sched_list);135 136 spin_lock(&entity->rq_lock);137 entity->sched_list = sched_list;138 entity->num_sched_list = num_sched_list;139 spin_unlock(&entity->rq_lock);140}141EXPORT_SYMBOL(drm_sched_entity_modify_sched);142 143static bool drm_sched_entity_is_idle(struct drm_sched_entity *entity)144{145 rmb(); /* for list_empty to work without lock */146 147 if (list_empty(&entity->list) ||148 spsc_queue_count(&entity->job_queue) == 0 ||149 entity->stopped)150 return true;151 152 return false;153}154 155/* Return true if entity could provide a job. */156bool drm_sched_entity_is_ready(struct drm_sched_entity *entity)157{158 if (spsc_queue_peek(&entity->job_queue) == NULL)159 return false;160 161 if (READ_ONCE(entity->dependency))162 return false;163 164 return true;165}166 167/**168 * drm_sched_entity_error - return error of last scheduled job169 * @entity: scheduler entity to check170 *171 * Opportunistically return the error of the last scheduled job. Result can172 * change any time when new jobs are pushed to the hw.173 */174int drm_sched_entity_error(struct drm_sched_entity *entity)175{176 struct dma_fence *fence;177 int r;178 179 rcu_read_lock();180 fence = rcu_dereference(entity->last_scheduled);181 r = fence ? fence->error : 0;182 rcu_read_unlock();183 184 return r;185}186EXPORT_SYMBOL(drm_sched_entity_error);187 188static void drm_sched_entity_kill_jobs_work(struct work_struct *wrk)189{190 struct drm_sched_job *job = container_of(wrk, typeof(*job), work);191 192 drm_sched_fence_finished(job->s_fence, -ESRCH);193 WARN_ON(job->s_fence->parent);194 job->sched->ops->free_job(job);195}196 197/* Signal the scheduler finished fence when the entity in question is killed. */198static void drm_sched_entity_kill_jobs_cb(struct dma_fence *f,199 struct dma_fence_cb *cb)200{201 struct drm_sched_job *job = container_of(cb, struct drm_sched_job,202 finish_cb);203 unsigned long index;204 205 dma_fence_put(f);206 207 /* Wait for all dependencies to avoid data corruptions */208 xa_for_each(&job->dependencies, index, f) {209 struct drm_sched_fence *s_fence = to_drm_sched_fence(f);210 211 if (s_fence && f == &s_fence->scheduled) {212 /* The dependencies array had a reference on the scheduled213 * fence, and the finished fence refcount might have214 * dropped to zero. Use dma_fence_get_rcu() so we get215 * a NULL fence in that case.216 */217 f = dma_fence_get_rcu(&s_fence->finished);218 219 /* Now that we have a reference on the finished fence,220 * we can release the reference the dependencies array221 * had on the scheduled fence.222 */223 dma_fence_put(&s_fence->scheduled);224 }225 226 xa_erase(&job->dependencies, index);227 if (f && !dma_fence_add_callback(f, &job->finish_cb,228 drm_sched_entity_kill_jobs_cb))229 return;230 231 dma_fence_put(f);232 }233 234 INIT_WORK(&job->work, drm_sched_entity_kill_jobs_work);235 schedule_work(&job->work);236}237 238/* Remove the entity from the scheduler and kill all pending jobs */239static void drm_sched_entity_kill(struct drm_sched_entity *entity)240{241 struct drm_sched_job *job;242 struct dma_fence *prev;243 244 if (!entity->rq)245 return;246 247 spin_lock(&entity->rq_lock);248 entity->stopped = true;249 drm_sched_rq_remove_entity(entity->rq, entity);250 spin_unlock(&entity->rq_lock);251 252 /* Make sure this entity is not used by the scheduler at the moment */253 wait_for_completion(&entity->entity_idle);254 255 /* The entity is guaranteed to not be used by the scheduler */256 prev = rcu_dereference_check(entity->last_scheduled, true);257 dma_fence_get(prev);258 while ((job = to_drm_sched_job(spsc_queue_pop(&entity->job_queue)))) {259 struct drm_sched_fence *s_fence = job->s_fence;260 261 dma_fence_get(&s_fence->finished);262 if (!prev || dma_fence_add_callback(prev, &job->finish_cb,263 drm_sched_entity_kill_jobs_cb))264 drm_sched_entity_kill_jobs_cb(NULL, &job->finish_cb);265 266 prev = &s_fence->finished;267 }268 dma_fence_put(prev);269}270 271/**272 * drm_sched_entity_flush - Flush a context entity273 *274 * @entity: scheduler entity275 * @timeout: time to wait in for Q to become empty in jiffies.276 *277 * Splitting drm_sched_entity_fini() into two functions, The first one does the278 * waiting, removes the entity from the runqueue and returns an error when the279 * process was killed.280 *281 * Returns the remaining time in jiffies left from the input timeout282 */283long drm_sched_entity_flush(struct drm_sched_entity *entity, long timeout)284{285 struct drm_gpu_scheduler *sched;286 struct task_struct *last_user;287 long ret = timeout;288 289 if (!entity->rq)290 return 0;291 292 sched = entity->rq->sched;293 /**294 * The client will not queue more IBs during this fini, consume existing295 * queued IBs or discard them on SIGKILL296 */297 if (current->flags & PF_EXITING) {298 if (timeout)299 ret = wait_event_timeout(300 sched->job_scheduled,301 drm_sched_entity_is_idle(entity),302 timeout);303 } else {304 wait_event_killable(sched->job_scheduled,305 drm_sched_entity_is_idle(entity));306 }307 308 /* For killed process disable any more IBs enqueue right now */309 last_user = cmpxchg(&entity->last_user, current->group_leader, NULL);310 if ((!last_user || last_user == current->group_leader) &&311 (current->flags & PF_EXITING) && (current->exit_code == SIGKILL))312 drm_sched_entity_kill(entity);313 314 return ret;315}316EXPORT_SYMBOL(drm_sched_entity_flush);317 318/**319 * drm_sched_entity_fini - Destroy a context entity320 *321 * @entity: scheduler entity322 *323 * Cleanups up @entity which has been initialized by drm_sched_entity_init().324 *325 * If there are potentially job still in flight or getting newly queued326 * drm_sched_entity_flush() must be called first. This function then goes over327 * the entity and signals all jobs with an error code if the process was killed.328 */329void drm_sched_entity_fini(struct drm_sched_entity *entity)330{331 /*332 * If consumption of existing IBs wasn't completed. Forcefully remove333 * them here. Also makes sure that the scheduler won't touch this entity334 * any more.335 */336 drm_sched_entity_kill(entity);337 338 if (entity->dependency) {339 dma_fence_remove_callback(entity->dependency, &entity->cb);340 dma_fence_put(entity->dependency);341 entity->dependency = NULL;342 }343 344 dma_fence_put(rcu_dereference_check(entity->last_scheduled, true));345 RCU_INIT_POINTER(entity->last_scheduled, NULL);346}347EXPORT_SYMBOL(drm_sched_entity_fini);348 349/**350 * drm_sched_entity_destroy - Destroy a context entity351 * @entity: scheduler entity352 *353 * Calls drm_sched_entity_flush() and drm_sched_entity_fini() as a354 * convenience wrapper.355 */356void drm_sched_entity_destroy(struct drm_sched_entity *entity)357{358 drm_sched_entity_flush(entity, MAX_WAIT_SCHED_ENTITY_Q_EMPTY);359 drm_sched_entity_fini(entity);360}361EXPORT_SYMBOL(drm_sched_entity_destroy);362 363/* drm_sched_entity_clear_dep - callback to clear the entities dependency */364static void drm_sched_entity_clear_dep(struct dma_fence *f,365 struct dma_fence_cb *cb)366{367 struct drm_sched_entity *entity =368 container_of(cb, struct drm_sched_entity, cb);369 370 entity->dependency = NULL;371 dma_fence_put(f);372}373 374/*375 * drm_sched_entity_clear_dep - callback to clear the entities dependency and376 * wake up scheduler377 */378static void drm_sched_entity_wakeup(struct dma_fence *f,379 struct dma_fence_cb *cb)380{381 struct drm_sched_entity *entity =382 container_of(cb, struct drm_sched_entity, cb);383 384 drm_sched_entity_clear_dep(f, cb);385 drm_sched_wakeup(entity->rq->sched);386}387 388/**389 * drm_sched_entity_set_priority - Sets priority of the entity390 *391 * @entity: scheduler entity392 * @priority: scheduler priority393 *394 * Update the priority of runqueus used for the entity.395 */396void drm_sched_entity_set_priority(struct drm_sched_entity *entity,397 enum drm_sched_priority priority)398{399 spin_lock(&entity->rq_lock);400 entity->priority = priority;401 spin_unlock(&entity->rq_lock);402}403EXPORT_SYMBOL(drm_sched_entity_set_priority);404 405/*406 * Add a callback to the current dependency of the entity to wake up the407 * scheduler when the entity becomes available.408 */409static bool drm_sched_entity_add_dependency_cb(struct drm_sched_entity *entity)410{411 struct drm_gpu_scheduler *sched = entity->rq->sched;412 struct dma_fence *fence = entity->dependency;413 struct drm_sched_fence *s_fence;414 415 if (fence->context == entity->fence_context ||416 fence->context == entity->fence_context + 1) {417 /*418 * Fence is a scheduled/finished fence from a job419 * which belongs to the same entity, we can ignore420 * fences from ourself421 */422 dma_fence_put(entity->dependency);423 return false;424 }425 426 s_fence = to_drm_sched_fence(fence);427 if (!fence->error && s_fence && s_fence->sched == sched &&428 !test_bit(DRM_SCHED_FENCE_DONT_PIPELINE, &fence->flags)) {429 430 /*431 * Fence is from the same scheduler, only need to wait for432 * it to be scheduled433 */434 fence = dma_fence_get(&s_fence->scheduled);435 dma_fence_put(entity->dependency);436 entity->dependency = fence;437 if (!dma_fence_add_callback(fence, &entity->cb,438 drm_sched_entity_clear_dep))439 return true;440 441 /* Ignore it when it is already scheduled */442 dma_fence_put(fence);443 return false;444 }445 446 if (!dma_fence_add_callback(entity->dependency, &entity->cb,447 drm_sched_entity_wakeup))448 return true;449 450 dma_fence_put(entity->dependency);451 return false;452}453 454static struct dma_fence *455drm_sched_job_dependency(struct drm_sched_job *job,456 struct drm_sched_entity *entity)457{458 struct dma_fence *f;459 460 /* We keep the fence around, so we can iterate over all dependencies461 * in drm_sched_entity_kill_jobs_cb() to ensure all deps are signaled462 * before killing the job.463 */464 f = xa_load(&job->dependencies, job->last_dependency);465 if (f) {466 job->last_dependency++;467 return dma_fence_get(f);468 }469 470 if (job->sched->ops->prepare_job)471 return job->sched->ops->prepare_job(job, entity);472 473 return NULL;474}475 476struct drm_sched_job *drm_sched_entity_pop_job(struct drm_sched_entity *entity)477{478 struct drm_sched_job *sched_job;479 480 sched_job = to_drm_sched_job(spsc_queue_peek(&entity->job_queue));481 if (!sched_job)482 return NULL;483 484 while ((entity->dependency =485 drm_sched_job_dependency(sched_job, entity))) {486 trace_drm_sched_job_wait_dep(sched_job, entity->dependency);487 488 if (drm_sched_entity_add_dependency_cb(entity))489 return NULL;490 }491 492 /* skip jobs from entity that marked guilty */493 if (entity->guilty && atomic_read(entity->guilty))494 dma_fence_set_error(&sched_job->s_fence->finished, -ECANCELED);495 496 dma_fence_put(rcu_dereference_check(entity->last_scheduled, true));497 rcu_assign_pointer(entity->last_scheduled,498 dma_fence_get(&sched_job->s_fence->finished));499 500 /*501 * If the queue is empty we allow drm_sched_entity_select_rq() to502 * locklessly access ->last_scheduled. This only works if we set the503 * pointer before we dequeue and if we a write barrier here.504 */505 smp_wmb();506 507 spsc_queue_pop(&entity->job_queue);508 509 /*510 * Update the entity's location in the min heap according to511 * the timestamp of the next job, if any.512 */513 if (drm_sched_policy == DRM_SCHED_POLICY_FIFO) {514 struct drm_sched_job *next;515 516 next = to_drm_sched_job(spsc_queue_peek(&entity->job_queue));517 if (next)518 drm_sched_rq_update_fifo(entity, next->submit_ts);519 }520 521 /* Jobs and entities might have different lifecycles. Since we're522 * removing the job from the entities queue, set the jobs entity pointer523 * to NULL to prevent any future access of the entity through this job.524 */525 sched_job->entity = NULL;526 527 return sched_job;528}529 530void drm_sched_entity_select_rq(struct drm_sched_entity *entity)531{532 struct dma_fence *fence;533 struct drm_gpu_scheduler *sched;534 struct drm_sched_rq *rq;535 536 /* single possible engine and already selected */537 if (!entity->sched_list)538 return;539 540 /* queue non-empty, stay on the same engine */541 if (spsc_queue_count(&entity->job_queue))542 return;543 544 /*545 * Only when the queue is empty are we guaranteed that the scheduler546 * thread cannot change ->last_scheduled. To enforce ordering we need547 * a read barrier here. See drm_sched_entity_pop_job() for the other548 * side.549 */550 smp_rmb();551 552 fence = rcu_dereference_check(entity->last_scheduled, true);553 554 /* stay on the same engine if the previous job hasn't finished */555 if (fence && !dma_fence_is_signaled(fence))556 return;557 558 spin_lock(&entity->rq_lock);559 sched = drm_sched_pick_best(entity->sched_list, entity->num_sched_list);560 rq = sched ? sched->sched_rq[entity->priority] : NULL;561 if (rq != entity->rq) {562 drm_sched_rq_remove_entity(entity->rq, entity);563 entity->rq = rq;564 }565 spin_unlock(&entity->rq_lock);566 567 if (entity->num_sched_list == 1)568 entity->sched_list = NULL;569}570 571/**572 * drm_sched_entity_push_job - Submit a job to the entity's job queue573 * @sched_job: job to submit574 *575 * Note: To guarantee that the order of insertion to queue matches the job's576 * fence sequence number this function should be called with drm_sched_job_arm()577 * under common lock for the struct drm_sched_entity that was set up for578 * @sched_job in drm_sched_job_init().579 *580 * Returns 0 for success, negative error code otherwise.581 */582void drm_sched_entity_push_job(struct drm_sched_job *sched_job)583{584 struct drm_sched_entity *entity = sched_job->entity;585 bool first;586 ktime_t submit_ts;587 588 trace_drm_sched_job(sched_job, entity);589 atomic_inc(entity->rq->sched->score);590 WRITE_ONCE(entity->last_user, current->group_leader);591 592 /*593 * After the sched_job is pushed into the entity queue, it may be594 * completed and freed up at any time. We can no longer access it.595 * Make sure to set the submit_ts first, to avoid a race.596 */597 sched_job->submit_ts = submit_ts = ktime_get();598 first = spsc_queue_push(&entity->job_queue, &sched_job->queue_node);599 600 /* first job wakes up scheduler */601 if (first) {602 struct drm_gpu_scheduler *sched;603 struct drm_sched_rq *rq;604 605 /* Add the entity to the run queue */606 spin_lock(&entity->rq_lock);607 if (entity->stopped) {608 spin_unlock(&entity->rq_lock);609 610 DRM_ERROR("Trying to push to a killed entity\n");611 return;612 }613 614 rq = entity->rq;615 sched = rq->sched;616 617 drm_sched_rq_add_entity(rq, entity);618 spin_unlock(&entity->rq_lock);619 620 if (drm_sched_policy == DRM_SCHED_POLICY_FIFO)621 drm_sched_rq_update_fifo(entity, submit_ts);622 623 drm_sched_wakeup(sched);624 }625}626EXPORT_SYMBOL(drm_sched_entity_push_job);627