include/drm/gpu_scheduler.h

   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
  31 #define MAX_WAIT_SCHED_ENTITY_Q_EMPTY msecs_to_jiffies(1000)
  32
  33 struct drm_gpu_scheduler;
  34 struct drm_sched_rq;
  35
  36 enum drm_sched_priority {
  37         DRM_SCHED_PRIORITY_MIN,
  38         DRM_SCHED_PRIORITY_LOW = DRM_SCHED_PRIORITY_MIN,
  39         DRM_SCHED_PRIORITY_NORMAL,
  40         DRM_SCHED_PRIORITY_HIGH_SW,
  41         DRM_SCHED_PRIORITY_HIGH_HW,
  42         DRM_SCHED_PRIORITY_KERNEL,
  43         DRM_SCHED_PRIORITY_MAX,
  44         DRM_SCHED_PRIORITY_INVALID = -1,
  45         DRM_SCHED_PRIORITY_UNSET = -2
  46 };
  47
  48 /**
  49  * struct drm_sched_entity - A wrapper around a job queue (typically
  50  * attached to the DRM file_priv).
  51  *
  52  * @list: used to append this struct to the list of entities in the
  53  *        runqueue.
  54  * @rq: runqueue on which this entity is currently scheduled.
  55  * @sched_list: A list of schedulers (drm_gpu_schedulers).
  56  *              Jobs from this entity can be scheduled on any scheduler
  57  *              on this list.
  58  * @num_sched_list: number of drm_gpu_schedulers in the sched_list.
  59  * @rq_lock: lock to modify the runqueue to which this entity belongs.
  60  * @job_queue: the list of jobs of this entity.
  61  * @fence_seq: a linearly increasing seqno incremented with each
  62  *             new &drm_sched_fence which is part of the entity.
  63  * @fence_context: a unique context for all the fences which belong
  64  *                 to this entity.
  65  *                 The &drm_sched_fence.scheduled uses the
  66  *                 fence_context but &drm_sched_fence.finished uses
  67  *                 fence_context + 1.
  68  * @dependency: the dependency fence of the job which is on the top
  69  *              of the job queue.
  70  * @cb: callback for the dependency fence above.
  71  * @guilty: points to ctx's guilty.
  72  * @fini_status: contains the exit status in case the process was signalled.
  73  * @last_scheduled: points to the finished fence of the last scheduled job.
  74  * @last_user: last group leader pushing a job into the entity.
  75  * @stopped: Marks the enity as removed from rq and destined for termination.
  76  * @entity_idle: Signals when enityt is not in use
  77  *
  78  * Entities will emit jobs in order to their corresponding hardware
  79  * ring, and the scheduler will alternate between entities based on
  80  * scheduling policy.
  81  */
  82 struct drm_sched_entity {
  83         struct list_head                list;
  84         struct drm_sched_rq             *rq;
  85         struct drm_gpu_scheduler        **sched_list;
  86         unsigned int                    num_sched_list;
  87         enum drm_sched_priority         priority;
  88         spinlock_t                      rq_lock;
  89
  90         struct spsc_queue               job_queue;
  91
  92         atomic_t                        fence_seq;
  93         uint64_t                        fence_context;
  94
  95         struct dma_fence                *dependency;
  96         struct dma_fence_cb             cb;
  97         atomic_t                        *guilty;
  98         struct dma_fence                *last_scheduled;
  99         struct task_struct              *last_user;
 100         bool                            stopped;
 101         struct completion               entity_idle;
 102 };
 103
 104 /**
 105  * struct drm_sched_rq - queue of entities to be scheduled.
 106  *
 107  * @lock: to modify the entities list.
 108  * @sched: the scheduler to which this rq belongs to.
 109  * @entities: list of the entities to be scheduled.
 110  * @current_entity: the entity which is to be scheduled.
 111  *
 112  * Run queue is a set of entities scheduling command submissions for
 113  * one specific ring. It implements the scheduling policy that selects
 114  * the next entity to emit commands from.
 115  */
 116 struct drm_sched_rq {
 117         spinlock_t                      lock;
 118         struct drm_gpu_scheduler        *sched;
 119         struct list_head                entities;
 120         struct drm_sched_entity         *current_entity;
 121 };
 122
 123 /**
 124  * struct drm_sched_fence - fences corresponding to the scheduling of a job.
 125  */
 126 struct drm_sched_fence {
 127         /**
 128          * @scheduled: this fence is what will be signaled by the scheduler
 129          * when the job is scheduled.
 130          */
 131         struct dma_fence                scheduled;
 132
 133         /**
 134          * @finished: this fence is what will be signaled by the scheduler
 135          * when the job is completed.
 136          *
 137          * When setting up an out fence for the job, you should use
 138          * this, since it's available immediately upon
 139          * drm_sched_job_init(), and the fence returned by the driver
 140          * from run_job() won't be created until the dependencies have
 141          * resolved.
 142          */
 143         struct dma_fence                finished;
 144
 145         /**
 146          * @parent: the fence returned by &drm_sched_backend_ops.run_job
 147          * when scheduling the job on hardware. We signal the
 148          * &drm_sched_fence.finished fence once parent is signalled.
 149          */
 150         struct dma_fence                *parent;
 151         /**
 152          * @sched: the scheduler instance to which the job having this struct
 153          * belongs to.
 154          */
 155         struct drm_gpu_scheduler        *sched;
 156         /**
 157          * @lock: the lock used by the scheduled and the finished fences.
 158          */
 159         spinlock_t                      lock;
 160         /**
 161          * @owner: job owner for debugging
 162          */
 163         void                            *owner;
 164 };
 165
 166 struct drm_sched_fence *to_drm_sched_fence(struct dma_fence *f);
 167
 168 /**
 169  * struct drm_sched_job - A job to be run by an entity.
 170  *
 171  * @queue_node: used to append this struct to the queue of jobs in an entity.
 172  * @sched: the scheduler instance on which this job is scheduled.
 173  * @s_fence: contains the fences for the scheduling of job.
 174  * @finish_cb: the callback for the finished fence.
 175  * @node: used to append this struct to the @drm_gpu_scheduler.ring_mirror_list.
 176  * @id: a unique id assigned to each job scheduled on the scheduler.
 177  * @karma: increment on every hang caused by this job. If this exceeds the hang
 178  *         limit of the scheduler then the job is marked guilty and will not
 179  *         be scheduled further.
 180  * @s_priority: the priority of the job.
 181  * @entity: the entity to which this job belongs.
 182  * @cb: the callback for the parent fence in s_fence.
 183  *
 184  * A job is created by the driver using drm_sched_job_init(), and
 185  * should call drm_sched_entity_push_job() once it wants the scheduler
 186  * to schedule the job.
 187  */
 188 struct drm_sched_job {
 189         struct spsc_node                queue_node;
 190         struct drm_gpu_scheduler        *sched;
 191         struct drm_sched_fence          *s_fence;
 192         struct dma_fence_cb             finish_cb;
 193         struct list_head                node;
 194         uint64_t                        id;
 195         atomic_t                        karma;
 196         enum drm_sched_priority         s_priority;
 197         struct drm_sched_entity  *entity;
 198         struct dma_fence_cb             cb;
 199 };
 200
 201 static inline bool drm_sched_invalidate_job(struct drm_sched_job *s_job,
 202                                             int threshold)
 203 {
 204         return (s_job && atomic_inc_return(&s_job->karma) > threshold);
 205 }
 206
 207 /**
 208  * struct drm_sched_backend_ops
 209  *
 210  * Define the backend operations called by the scheduler,
 211  * these functions should be implemented in driver side.
 212  */
 213 struct drm_sched_backend_ops {
 214         /**
 215          * @dependency: Called when the scheduler is considering scheduling
 216          * this job next, to get another struct dma_fence for this job to
 217          * block on.  Once it returns NULL, run_job() may be called.
 218          */
 219         struct dma_fence *(*dependency)(struct drm_sched_job *sched_job,
 220                                         struct drm_sched_entity *s_entity);
 221
 222         /**
 223          * @run_job: Called to execute the job once all of the dependencies
 224          * have been resolved.  This may be called multiple times, if
 225          * timedout_job() has happened and drm_sched_job_recovery()
 226          * decides to try it again.
 227          */
 228         struct dma_fence *(*run_job)(struct drm_sched_job *sched_job);
 229
 230         /**
 231          * @timedout_job: Called when a job has taken too long to execute,
 232          * to trigger GPU recovery.
 233          */
 234         void (*timedout_job)(struct drm_sched_job *sched_job);
 235
 236         /**
 237          * @free_job: Called once the job's finished fence has been signaled
 238          * and it's time to clean it up.
 239          */
 240         void (*free_job)(struct drm_sched_job *sched_job);
 241 };
 242
 243 /**
 244  * struct drm_gpu_scheduler
 245  *
 246  * @ops: backend operations provided by the driver.
 247  * @hw_submission_limit: the max size of the hardware queue.
 248  * @timeout: the time after which a job is removed from the scheduler.
 249  * @name: name of the ring for which this scheduler is being used.
 250  * @sched_rq: priority wise array of run queues.
 251  * @wake_up_worker: the wait queue on which the scheduler sleeps until a job
 252  *                  is ready to be scheduled.
 253  * @job_scheduled: once @drm_sched_entity_do_release is called the scheduler
 254  *                 waits on this wait queue until all the scheduled jobs are
 255  *                 finished.
 256  * @hw_rq_count: the number of jobs currently in the hardware queue.
 257  * @job_id_count: used to assign unique id to the each job.
 258  * @work_tdr: schedules a delayed call to @drm_sched_job_timedout after the
 259  *            timeout interval is over.
 260  * @thread: the kthread on which the scheduler which run.
 261  * @ring_mirror_list: the list of jobs which are currently in the job queue.
 262  * @job_list_lock: lock to protect the ring_mirror_list.
 263  * @hang_limit: once the hangs by a job crosses this limit then it is marked
 264  *              guilty and it will be considered for scheduling further.
 265  * @num_jobs: the number of jobs in queue in the scheduler
 266  * @ready: marks if the underlying HW is ready to work
 267  * @free_guilty: A hit to time out handler to free the guilty job.
 268  *
 269  * One scheduler is implemented for each hardware ring.
 270  */
 271 struct drm_gpu_scheduler {
 272         const struct drm_sched_backend_ops      *ops;
 273         uint32_t                        hw_submission_limit;
 274         long                            timeout;
 275         const char                      *name;
 276         struct drm_sched_rq             sched_rq[DRM_SCHED_PRIORITY_MAX];
 277         wait_queue_head_t               wake_up_worker;
 278         wait_queue_head_t               job_scheduled;
 279         atomic_t                        hw_rq_count;
 280         atomic64_t                      job_id_count;
 281         struct delayed_work             work_tdr;
 282         struct task_struct              *thread;
 283         struct list_head                ring_mirror_list;
 284         spinlock_t                      job_list_lock;
 285         int                             hang_limit;
 286         atomic_t                        num_jobs;
 287         bool                    ready;
 288         bool                            free_guilty;
 289 };
 290
 291 int drm_sched_init(struct drm_gpu_scheduler *sched,
 292                    const struct drm_sched_backend_ops *ops,
 293                    uint32_t hw_submission, unsigned hang_limit, long timeout,
 294                    const char *name);
 295
 296 void drm_sched_fini(struct drm_gpu_scheduler *sched);
 297 int drm_sched_job_init(struct drm_sched_job *job,
 298                        struct drm_sched_entity *entity,
 299                        void *owner);
 300 void drm_sched_entity_modify_sched(struct drm_sched_entity *entity,
 301                                     struct drm_gpu_scheduler **sched_list,
 302                                    unsigned int num_sched_list);
 303
 304 void drm_sched_job_cleanup(struct drm_sched_job *job);
 305 void drm_sched_wakeup(struct drm_gpu_scheduler *sched);
 306 void drm_sched_stop(struct drm_gpu_scheduler *sched, struct drm_sched_job *bad);
 307 void drm_sched_start(struct drm_gpu_scheduler *sched, bool full_recovery);
 308 void drm_sched_resubmit_jobs(struct drm_gpu_scheduler *sched);
 309 void drm_sched_increase_karma(struct drm_sched_job *bad);
 310 bool drm_sched_dependency_optimized(struct dma_fence* fence,
 311                                     struct drm_sched_entity *entity);
 312 void drm_sched_fault(struct drm_gpu_scheduler *sched);
 313 void drm_sched_job_kickout(struct drm_sched_job *s_job);
 314
 315 void drm_sched_rq_add_entity(struct drm_sched_rq *rq,
 316                              struct drm_sched_entity *entity);
 317 void drm_sched_rq_remove_entity(struct drm_sched_rq *rq,
 318                                 struct drm_sched_entity *entity);
 319
 320 int drm_sched_entity_init(struct drm_sched_entity *entity,
 321                           enum drm_sched_priority priority,
 322                           struct drm_gpu_scheduler **sched_list,
 323                           unsigned int num_sched_list,
 324                           atomic_t *guilty);
 325 long drm_sched_entity_flush(struct drm_sched_entity *entity, long timeout);
 326 void drm_sched_entity_fini(struct drm_sched_entity *entity);
 327 void drm_sched_entity_destroy(struct drm_sched_entity *entity);
 328 void drm_sched_entity_select_rq(struct drm_sched_entity *entity);
 329 struct drm_sched_job *drm_sched_entity_pop_job(struct drm_sched_entity *entity);
 330 void drm_sched_entity_push_job(struct drm_sched_job *sched_job,
 331                                struct drm_sched_entity *entity);
 332 void drm_sched_entity_set_priority(struct drm_sched_entity *entity,
 333                                    enum drm_sched_priority priority);
 334 bool drm_sched_entity_is_ready(struct drm_sched_entity *entity);
 335
 336 struct drm_sched_fence *drm_sched_fence_create(
 337         struct drm_sched_entity *s_entity, void *owner);
 338 void drm_sched_fence_scheduled(struct drm_sched_fence *fence);
 339 void drm_sched_fence_finished(struct drm_sched_fence *fence);
 340
 341 unsigned long drm_sched_suspend_timeout(struct drm_gpu_scheduler *sched);
 342 void drm_sched_resume_timeout(struct drm_gpu_scheduler *sched,
 343                                 unsigned long remaining);
 344 struct drm_gpu_scheduler *
 345 drm_sched_pick_best(struct drm_gpu_scheduler **sched_list,
 346                      unsigned int num_sched_list);
 347
 348 #endif