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 #ifndef _DRM_GPU_SCHEDULER_H_ 25 #define _DRM_GPU_SCHEDULER_H_ 26 27 #include <drm/spsc_queue.h> 28 #include <linux/dma-fence.h> 29 #include <linux/completion.h> 30 #include <linux/xarray.h> 31 #include <linux/workqueue.h> 32 33 #define MAX_WAIT_SCHED_ENTITY_Q_EMPTY msecs_to_jiffies(1000) 34 35 /** 36 * DRM_SCHED_FENCE_DONT_PIPELINE - Prevent dependency pipelining 37 * 38 * Setting this flag on a scheduler fence prevents pipelining of jobs depending 39 * on this fence. In other words we always insert a full CPU round trip before 40 * dependent jobs are pushed to the hw queue. 41 */ 42 #define DRM_SCHED_FENCE_DONT_PIPELINE DMA_FENCE_FLAG_USER_BITS 43 44 /** 45 * DRM_SCHED_FENCE_FLAG_HAS_DEADLINE_BIT - A fence deadline hint has been set 46 * 47 * Because we could have a deadline hint can be set before the backing hw 48 * fence is created, we need to keep track of whether a deadline has already 49 * been set. 50 */ 51 #define DRM_SCHED_FENCE_FLAG_HAS_DEADLINE_BIT (DMA_FENCE_FLAG_USER_BITS + 1) 52 53 enum dma_resv_usage; 54 struct dma_resv; 55 struct drm_gem_object; 56 57 struct drm_gpu_scheduler; 58 struct drm_sched_rq; 59 60 struct drm_file; 61 62 /* These are often used as an (initial) index 63 * to an array, and as such should start at 0. 64 */ 65 enum drm_sched_priority { 66 DRM_SCHED_PRIORITY_KERNEL, 67 DRM_SCHED_PRIORITY_HIGH, 68 DRM_SCHED_PRIORITY_NORMAL, 69 DRM_SCHED_PRIORITY_LOW, 70 71 DRM_SCHED_PRIORITY_COUNT 72 }; 73 74 /* Used to choose between FIFO and RR job-scheduling */ 75 extern int drm_sched_policy; 76 77 #define DRM_SCHED_POLICY_RR 0 78 #define DRM_SCHED_POLICY_FIFO 1 79 80 /** 81 * struct drm_sched_entity - A wrapper around a job queue (typically 82 * attached to the DRM file_priv). 83 * 84 * Entities will emit jobs in order to their corresponding hardware 85 * ring, and the scheduler will alternate between entities based on 86 * scheduling policy. 87 */ 88 struct drm_sched_entity { 89 /** 90 * @list: 91 * 92 * Used to append this struct to the list of entities in the runqueue 93 * @rq under &drm_sched_rq.entities. 94 * 95 * Protected by &drm_sched_rq.lock of @rq. 96 */ 97 struct list_head list; 98 99 /** 100 * @lock: 101 * 102 * Lock protecting the run-queue (@rq) to which this entity belongs, 103 * @priority and the list of schedulers (@sched_list, @num_sched_list). 104 */ 105 spinlock_t lock; 106 107 /** 108 * @rq: 109 * 110 * Runqueue on which this entity is currently scheduled. 111 * 112 * FIXME: Locking is very unclear for this. Writers are protected by 113 * @lock, but readers are generally lockless and seem to just race with 114 * not even a READ_ONCE. 115 */ 116 struct drm_sched_rq *rq; 117 118 /** 119 * @sched_list: 120 * 121 * A list of schedulers (struct drm_gpu_scheduler). Jobs from this entity can 122 * be scheduled on any scheduler on this list. 123 * 124 * This can be modified by calling drm_sched_entity_modify_sched(). 125 * Locking is entirely up to the driver, see the above function for more 126 * details. 127 * 128 * This will be set to NULL if &num_sched_list equals 1 and @rq has been 129 * set already. 130 * 131 * FIXME: This means priority changes through 132 * drm_sched_entity_set_priority() will be lost henceforth in this case. 133 */ 134 struct drm_gpu_scheduler **sched_list; 135 136 /** 137 * @num_sched_list: 138 * 139 * Number of drm_gpu_schedulers in the @sched_list. 140 */ 141 unsigned int num_sched_list; 142 143 /** 144 * @priority: 145 * 146 * Priority of the entity. This can be modified by calling 147 * drm_sched_entity_set_priority(). Protected by @lock. 148 */ 149 enum drm_sched_priority priority; 150 151 /** 152 * @job_queue: the list of jobs of this entity. 153 */ 154 struct spsc_queue job_queue; 155 156 /** 157 * @fence_seq: 158 * 159 * A linearly increasing seqno incremented with each new 160 * &drm_sched_fence which is part of the entity. 161 * 162 * FIXME: Callers of drm_sched_job_arm() need to ensure correct locking, 163 * this doesn't need to be atomic. 164 */ 165 atomic_t fence_seq; 166 167 /** 168 * @fence_context: 169 * 170 * A unique context for all the fences which belong to this entity. The 171 * &drm_sched_fence.scheduled uses the fence_context but 172 * &drm_sched_fence.finished uses fence_context + 1. 173 */ 174 uint64_t fence_context; 175 176 /** 177 * @dependency: 178 * 179 * The dependency fence of the job which is on the top of the job queue. 180 */ 181 struct dma_fence *dependency; 182 183 /** 184 * @cb: 185 * 186 * Callback for the dependency fence above. 187 */ 188 struct dma_fence_cb cb; 189 190 /** 191 * @guilty: 192 * 193 * Points to entities' guilty. 194 */ 195 atomic_t *guilty; 196 197 /** 198 * @last_scheduled: 199 * 200 * Points to the finished fence of the last scheduled job. Only written 201 * by the scheduler thread, can be accessed locklessly from 202 * drm_sched_job_arm() if the queue is empty. 203 */ 204 struct dma_fence __rcu *last_scheduled; 205 206 /** 207 * @last_user: last group leader pushing a job into the entity. 208 */ 209 struct task_struct *last_user; 210 211 /** 212 * @stopped: 213 * 214 * Marks the enity as removed from rq and destined for 215 * termination. This is set by calling drm_sched_entity_flush() and by 216 * drm_sched_fini(). 217 */ 218 bool stopped; 219 220 /** 221 * @entity_idle: 222 * 223 * Signals when entity is not in use, used to sequence entity cleanup in 224 * drm_sched_entity_fini(). 225 */ 226 struct completion entity_idle; 227 228 /** 229 * @oldest_job_waiting: 230 * 231 * Marks earliest job waiting in SW queue 232 */ 233 ktime_t oldest_job_waiting; 234 235 /** 236 * @rb_tree_node: 237 * 238 * The node used to insert this entity into time based priority queue 239 */ 240 struct rb_node rb_tree_node; 241 242 }; 243 244 /** 245 * struct drm_sched_rq - queue of entities to be scheduled. 246 * 247 * @sched: the scheduler to which this rq belongs to. 248 * @lock: protects @entities, @rb_tree_root and @current_entity. 249 * @current_entity: the entity which is to be scheduled. 250 * @entities: list of the entities to be scheduled. 251 * @rb_tree_root: root of time based priority queue of entities for FIFO scheduling 252 * 253 * Run queue is a set of entities scheduling command submissions for 254 * one specific ring. It implements the scheduling policy that selects 255 * the next entity to emit commands from. 256 */ 257 struct drm_sched_rq { 258 struct drm_gpu_scheduler *sched; 259 260 spinlock_t lock; 261 /* Following members are protected by the @lock: */ 262 struct drm_sched_entity *current_entity; 263 struct list_head entities; 264 struct rb_root_cached rb_tree_root; 265 }; 266 267 /** 268 * struct drm_sched_fence - fences corresponding to the scheduling of a job. 269 */ 270 struct drm_sched_fence { 271 /** 272 * @scheduled: this fence is what will be signaled by the scheduler 273 * when the job is scheduled. 274 */ 275 struct dma_fence scheduled; 276 277 /** 278 * @finished: this fence is what will be signaled by the scheduler 279 * when the job is completed. 280 * 281 * When setting up an out fence for the job, you should use 282 * this, since it's available immediately upon 283 * drm_sched_job_init(), and the fence returned by the driver 284 * from run_job() won't be created until the dependencies have 285 * resolved. 286 */ 287 struct dma_fence finished; 288 289 /** 290 * @deadline: deadline set on &drm_sched_fence.finished which 291 * potentially needs to be propagated to &drm_sched_fence.parent 292 */ 293 ktime_t deadline; 294 295 /** 296 * @parent: the fence returned by &drm_sched_backend_ops.run_job 297 * when scheduling the job on hardware. We signal the 298 * &drm_sched_fence.finished fence once parent is signalled. 299 */ 300 struct dma_fence *parent; 301 /** 302 * @sched: the scheduler instance to which the job having this struct 303 * belongs to. 304 */ 305 struct drm_gpu_scheduler *sched; 306 /** 307 * @lock: the lock used by the scheduled and the finished fences. 308 */ 309 spinlock_t lock; 310 /** 311 * @owner: job owner for debugging 312 */ 313 void *owner; 314 }; 315 316 struct drm_sched_fence *to_drm_sched_fence(struct dma_fence *f); 317 318 /** 319 * struct drm_sched_job - A job to be run by an entity. 320 * 321 * @queue_node: used to append this struct to the queue of jobs in an entity. 322 * @list: a job participates in a "pending" and "done" lists. 323 * @sched: the scheduler instance on which this job is scheduled. 324 * @s_fence: contains the fences for the scheduling of job. 325 * @finish_cb: the callback for the finished fence. 326 * @credits: the number of credits this job contributes to the scheduler 327 * @work: Helper to reschedule job kill to different context. 328 * @id: a unique id assigned to each job scheduled on the scheduler. 329 * @karma: increment on every hang caused by this job. If this exceeds the hang 330 * limit of the scheduler then the job is marked guilty and will not 331 * be scheduled further. 332 * @s_priority: the priority of the job. 333 * @entity: the entity to which this job belongs. 334 * @cb: the callback for the parent fence in s_fence. 335 * 336 * A job is created by the driver using drm_sched_job_init(), and 337 * should call drm_sched_entity_push_job() once it wants the scheduler 338 * to schedule the job. 339 */ 340 struct drm_sched_job { 341 struct spsc_node queue_node; 342 struct list_head list; 343 struct drm_gpu_scheduler *sched; 344 struct drm_sched_fence *s_fence; 345 346 u32 credits; 347 348 /* 349 * work is used only after finish_cb has been used and will not be 350 * accessed anymore. 351 */ 352 union { 353 struct dma_fence_cb finish_cb; 354 struct work_struct work; 355 }; 356 357 uint64_t id; 358 atomic_t karma; 359 enum drm_sched_priority s_priority; 360 struct drm_sched_entity *entity; 361 struct dma_fence_cb cb; 362 /** 363 * @dependencies: 364 * 365 * Contains the dependencies as struct dma_fence for this job, see 366 * drm_sched_job_add_dependency() and 367 * drm_sched_job_add_implicit_dependencies(). 368 */ 369 struct xarray dependencies; 370 371 /** @last_dependency: tracks @dependencies as they signal */ 372 unsigned long last_dependency; 373 374 /** 375 * @submit_ts: 376 * 377 * When the job was pushed into the entity queue. 378 */ 379 ktime_t submit_ts; 380 }; 381 382 static inline bool drm_sched_invalidate_job(struct drm_sched_job *s_job, 383 int threshold) 384 { 385 return s_job && atomic_inc_return(&s_job->karma) > threshold; 386 } 387 388 enum drm_gpu_sched_stat { 389 DRM_GPU_SCHED_STAT_NONE, /* Reserve 0 */ 390 DRM_GPU_SCHED_STAT_NOMINAL, 391 DRM_GPU_SCHED_STAT_ENODEV, 392 }; 393 394 /** 395 * struct drm_sched_backend_ops - Define the backend operations 396 * called by the scheduler 397 * 398 * These functions should be implemented in the driver side. 399 */ 400 struct drm_sched_backend_ops { 401 /** 402 * @prepare_job: 403 * 404 * Called when the scheduler is considering scheduling this job next, to 405 * get another struct dma_fence for this job to block on. Once it 406 * returns NULL, run_job() may be called. 407 * 408 * Can be NULL if no additional preparation to the dependencies are 409 * necessary. Skipped when jobs are killed instead of run. 410 */ 411 struct dma_fence *(*prepare_job)(struct drm_sched_job *sched_job, 412 struct drm_sched_entity *s_entity); 413 414 /** 415 * @run_job: Called to execute the job once all of the dependencies 416 * have been resolved. This may be called multiple times, if 417 * timedout_job() has happened and drm_sched_job_recovery() 418 * decides to try it again. 419 */ 420 struct dma_fence *(*run_job)(struct drm_sched_job *sched_job); 421 422 /** 423 * @timedout_job: Called when a job has taken too long to execute, 424 * to trigger GPU recovery. 425 * 426 * This method is called in a workqueue context. 427 * 428 * Drivers typically issue a reset to recover from GPU hangs, and this 429 * procedure usually follows the following workflow: 430 * 431 * 1. Stop the scheduler using drm_sched_stop(). This will park the 432 * scheduler thread and cancel the timeout work, guaranteeing that 433 * nothing is queued while we reset the hardware queue 434 * 2. Try to gracefully stop non-faulty jobs (optional) 435 * 3. Issue a GPU reset (driver-specific) 436 * 4. Re-submit jobs using drm_sched_resubmit_jobs() 437 * 5. Restart the scheduler using drm_sched_start(). At that point, new 438 * jobs can be queued, and the scheduler thread is unblocked 439 * 440 * Note that some GPUs have distinct hardware queues but need to reset 441 * the GPU globally, which requires extra synchronization between the 442 * timeout handler of the different &drm_gpu_scheduler. One way to 443 * achieve this synchronization is to create an ordered workqueue 444 * (using alloc_ordered_workqueue()) at the driver level, and pass this 445 * queue to drm_sched_init(), to guarantee that timeout handlers are 446 * executed sequentially. The above workflow needs to be slightly 447 * adjusted in that case: 448 * 449 * 1. Stop all schedulers impacted by the reset using drm_sched_stop() 450 * 2. Try to gracefully stop non-faulty jobs on all queues impacted by 451 * the reset (optional) 452 * 3. Issue a GPU reset on all faulty queues (driver-specific) 453 * 4. Re-submit jobs on all schedulers impacted by the reset using 454 * drm_sched_resubmit_jobs() 455 * 5. Restart all schedulers that were stopped in step #1 using 456 * drm_sched_start() 457 * 458 * Return DRM_GPU_SCHED_STAT_NOMINAL, when all is normal, 459 * and the underlying driver has started or completed recovery. 460 * 461 * Return DRM_GPU_SCHED_STAT_ENODEV, if the device is no longer 462 * available, i.e. has been unplugged. 463 */ 464 enum drm_gpu_sched_stat (*timedout_job)(struct drm_sched_job *sched_job); 465 466 /** 467 * @free_job: Called once the job's finished fence has been signaled 468 * and it's time to clean it up. 469 */ 470 void (*free_job)(struct drm_sched_job *sched_job); 471 472 /** 473 * @update_job_credits: Called when the scheduler is considering this 474 * job for execution. 475 * 476 * This callback returns the number of credits the job would take if 477 * pushed to the hardware. Drivers may use this to dynamically update 478 * the job's credit count. For instance, deduct the number of credits 479 * for already signalled native fences. 480 * 481 * This callback is optional. 482 */ 483 u32 (*update_job_credits)(struct drm_sched_job *sched_job); 484 }; 485 486 /** 487 * struct drm_gpu_scheduler - scheduler instance-specific data 488 * 489 * @ops: backend operations provided by the driver. 490 * @credit_limit: the credit limit of this scheduler 491 * @credit_count: the current credit count of this scheduler 492 * @timeout: the time after which a job is removed from the scheduler. 493 * @name: name of the ring for which this scheduler is being used. 494 * @num_rqs: Number of run-queues. This is at most DRM_SCHED_PRIORITY_COUNT, 495 * as there's usually one run-queue per priority, but could be less. 496 * @sched_rq: An allocated array of run-queues of size @num_rqs; 497 * @job_scheduled: once @drm_sched_entity_do_release is called the scheduler 498 * waits on this wait queue until all the scheduled jobs are 499 * finished. 500 * @job_id_count: used to assign unique id to the each job. 501 * @submit_wq: workqueue used to queue @work_run_job and @work_free_job 502 * @timeout_wq: workqueue used to queue @work_tdr 503 * @work_run_job: work which calls run_job op of each scheduler. 504 * @work_free_job: work which calls free_job op of each scheduler. 505 * @work_tdr: schedules a delayed call to @drm_sched_job_timedout after the 506 * timeout interval is over. 507 * @pending_list: the list of jobs which are currently in the job queue. 508 * @job_list_lock: lock to protect the pending_list. 509 * @hang_limit: once the hangs by a job crosses this limit then it is marked 510 * guilty and it will no longer be considered for scheduling. 511 * @score: score to help loadbalancer pick a idle sched 512 * @_score: score used when the driver doesn't provide one 513 * @ready: marks if the underlying HW is ready to work 514 * @free_guilty: A hit to time out handler to free the guilty job. 515 * @pause_submit: pause queuing of @work_run_job on @submit_wq 516 * @own_submit_wq: scheduler owns allocation of @submit_wq 517 * @dev: system &struct device 518 * 519 * One scheduler is implemented for each hardware ring. 520 */ 521 struct drm_gpu_scheduler { 522 const struct drm_sched_backend_ops *ops; 523 u32 credit_limit; 524 atomic_t credit_count; 525 long timeout; 526 const char *name; 527 u32 num_rqs; 528 struct drm_sched_rq **sched_rq; 529 wait_queue_head_t job_scheduled; 530 atomic64_t job_id_count; 531 struct workqueue_struct *submit_wq; 532 struct workqueue_struct *timeout_wq; 533 struct work_struct work_run_job; 534 struct work_struct work_free_job; 535 struct delayed_work work_tdr; 536 struct list_head pending_list; 537 spinlock_t job_list_lock; 538 int hang_limit; 539 atomic_t *score; 540 atomic_t _score; 541 bool ready; 542 bool free_guilty; 543 bool pause_submit; 544 bool own_submit_wq; 545 struct device *dev; 546 }; 547 548 int drm_sched_init(struct drm_gpu_scheduler *sched, 549 const struct drm_sched_backend_ops *ops, 550 struct workqueue_struct *submit_wq, 551 u32 num_rqs, u32 credit_limit, unsigned int hang_limit, 552 long timeout, struct workqueue_struct *timeout_wq, 553 atomic_t *score, const char *name, struct device *dev); 554 555 void drm_sched_fini(struct drm_gpu_scheduler *sched); 556 int drm_sched_job_init(struct drm_sched_job *job, 557 struct drm_sched_entity *entity, 558 u32 credits, void *owner); 559 void drm_sched_job_arm(struct drm_sched_job *job); 560 int drm_sched_job_add_dependency(struct drm_sched_job *job, 561 struct dma_fence *fence); 562 int drm_sched_job_add_syncobj_dependency(struct drm_sched_job *job, 563 struct drm_file *file, 564 u32 handle, 565 u32 point); 566 int drm_sched_job_add_resv_dependencies(struct drm_sched_job *job, 567 struct dma_resv *resv, 568 enum dma_resv_usage usage); 569 int drm_sched_job_add_implicit_dependencies(struct drm_sched_job *job, 570 struct drm_gem_object *obj, 571 bool write); 572 573 574 void drm_sched_entity_modify_sched(struct drm_sched_entity *entity, 575 struct drm_gpu_scheduler **sched_list, 576 unsigned int num_sched_list); 577 578 void drm_sched_tdr_queue_imm(struct drm_gpu_scheduler *sched); 579 void drm_sched_job_cleanup(struct drm_sched_job *job); 580 void drm_sched_wakeup(struct drm_gpu_scheduler *sched); 581 bool drm_sched_wqueue_ready(struct drm_gpu_scheduler *sched); 582 void drm_sched_wqueue_stop(struct drm_gpu_scheduler *sched); 583 void drm_sched_wqueue_start(struct drm_gpu_scheduler *sched); 584 void drm_sched_stop(struct drm_gpu_scheduler *sched, struct drm_sched_job *bad); 585 void drm_sched_start(struct drm_gpu_scheduler *sched, int errno); 586 void drm_sched_resubmit_jobs(struct drm_gpu_scheduler *sched); 587 void drm_sched_increase_karma(struct drm_sched_job *bad); 588 void drm_sched_reset_karma(struct drm_sched_job *bad); 589 void drm_sched_increase_karma_ext(struct drm_sched_job *bad, int type); 590 bool drm_sched_dependency_optimized(struct dma_fence* fence, 591 struct drm_sched_entity *entity); 592 void drm_sched_fault(struct drm_gpu_scheduler *sched); 593 594 void drm_sched_rq_add_entity(struct drm_sched_rq *rq, 595 struct drm_sched_entity *entity); 596 void drm_sched_rq_remove_entity(struct drm_sched_rq *rq, 597 struct drm_sched_entity *entity); 598 599 void drm_sched_rq_update_fifo_locked(struct drm_sched_entity *entity, 600 struct drm_sched_rq *rq, ktime_t ts); 601 602 int drm_sched_entity_init(struct drm_sched_entity *entity, 603 enum drm_sched_priority priority, 604 struct drm_gpu_scheduler **sched_list, 605 unsigned int num_sched_list, 606 atomic_t *guilty); 607 long drm_sched_entity_flush(struct drm_sched_entity *entity, long timeout); 608 void drm_sched_entity_fini(struct drm_sched_entity *entity); 609 void drm_sched_entity_destroy(struct drm_sched_entity *entity); 610 void drm_sched_entity_select_rq(struct drm_sched_entity *entity); 611 struct drm_sched_job *drm_sched_entity_pop_job(struct drm_sched_entity *entity); 612 void drm_sched_entity_push_job(struct drm_sched_job *sched_job); 613 void drm_sched_entity_set_priority(struct drm_sched_entity *entity, 614 enum drm_sched_priority priority); 615 bool drm_sched_entity_is_ready(struct drm_sched_entity *entity); 616 int drm_sched_entity_error(struct drm_sched_entity *entity); 617 618 struct drm_sched_fence *drm_sched_fence_alloc( 619 struct drm_sched_entity *s_entity, void *owner); 620 void drm_sched_fence_init(struct drm_sched_fence *fence, 621 struct drm_sched_entity *entity); 622 void drm_sched_fence_free(struct drm_sched_fence *fence); 623 624 void drm_sched_fence_scheduled(struct drm_sched_fence *fence, 625 struct dma_fence *parent); 626 void drm_sched_fence_finished(struct drm_sched_fence *fence, int result); 627 628 unsigned long drm_sched_suspend_timeout(struct drm_gpu_scheduler *sched); 629 void drm_sched_resume_timeout(struct drm_gpu_scheduler *sched, 630 unsigned long remaining); 631 struct drm_gpu_scheduler * 632 drm_sched_pick_best(struct drm_gpu_scheduler **sched_list, 633 unsigned int num_sched_list); 634 635 #endif 636