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 /* These are often used as an (initial) index 37 * to an array, and as such should start at 0. 38 */ 39 enum drm_sched_priority { 40 DRM_SCHED_PRIORITY_MIN, 41 DRM_SCHED_PRIORITY_NORMAL, 42 DRM_SCHED_PRIORITY_HIGH, 43 DRM_SCHED_PRIORITY_KERNEL, 44 45 DRM_SCHED_PRIORITY_COUNT, 46 DRM_SCHED_PRIORITY_UNSET = -2 47 }; 48 49 /** 50 * struct drm_sched_entity - A wrapper around a job queue (typically 51 * attached to the DRM file_priv). 52 * 53 * @list: used to append this struct to the list of entities in the 54 * runqueue. 55 * @rq: runqueue on which this entity is currently scheduled. 56 * @sched_list: A list of schedulers (drm_gpu_schedulers). 57 * Jobs from this entity can be scheduled on any scheduler 58 * on this list. 59 * @num_sched_list: number of drm_gpu_schedulers in the sched_list. 60 * @priority: priority of the entity 61 * @rq_lock: lock to modify the runqueue to which this entity belongs. 62 * @job_queue: the list of jobs of this entity. 63 * @fence_seq: a linearly increasing seqno incremented with each 64 * new &drm_sched_fence which is part of the entity. 65 * @fence_context: a unique context for all the fences which belong 66 * to this entity. 67 * The &drm_sched_fence.scheduled uses the 68 * fence_context but &drm_sched_fence.finished uses 69 * fence_context + 1. 70 * @dependency: the dependency fence of the job which is on the top 71 * of the job queue. 72 * @cb: callback for the dependency fence above. 73 * @guilty: points to ctx's guilty. 74 * @fini_status: contains the exit status in case the process was signalled. 75 * @last_scheduled: points to the finished fence of the last scheduled job. 76 * @last_user: last group leader pushing a job into the entity. 77 * @stopped: Marks the enity as removed from rq and destined for termination. 78 * @entity_idle: Signals when enityt is not in use 79 * 80 * Entities will emit jobs in order to their corresponding hardware 81 * ring, and the scheduler will alternate between entities based on 82 * scheduling policy. 83 */ 84 struct drm_sched_entity { 85 struct list_head list; 86 struct drm_sched_rq *rq; 87 struct drm_gpu_scheduler **sched_list; 88 unsigned int num_sched_list; 89 enum drm_sched_priority priority; 90 spinlock_t rq_lock; 91 92 struct spsc_queue job_queue; 93 94 atomic_t fence_seq; 95 uint64_t fence_context; 96 97 struct dma_fence *dependency; 98 struct dma_fence_cb cb; 99 atomic_t *guilty; 100 struct dma_fence *last_scheduled; 101 struct task_struct *last_user; 102 bool stopped; 103 struct completion entity_idle; 104 }; 105 106 /** 107 * struct drm_sched_rq - queue of entities to be scheduled. 108 * 109 * @lock: to modify the entities list. 110 * @sched: the scheduler to which this rq belongs to. 111 * @entities: list of the entities to be scheduled. 112 * @current_entity: the entity which is to be scheduled. 113 * 114 * Run queue is a set of entities scheduling command submissions for 115 * one specific ring. It implements the scheduling policy that selects 116 * the next entity to emit commands from. 117 */ 118 struct drm_sched_rq { 119 spinlock_t lock; 120 struct drm_gpu_scheduler *sched; 121 struct list_head entities; 122 struct drm_sched_entity *current_entity; 123 }; 124 125 /** 126 * struct drm_sched_fence - fences corresponding to the scheduling of a job. 127 */ 128 struct drm_sched_fence { 129 /** 130 * @scheduled: this fence is what will be signaled by the scheduler 131 * when the job is scheduled. 132 */ 133 struct dma_fence scheduled; 134 135 /** 136 * @finished: this fence is what will be signaled by the scheduler 137 * when the job is completed. 138 * 139 * When setting up an out fence for the job, you should use 140 * this, since it's available immediately upon 141 * drm_sched_job_init(), and the fence returned by the driver 142 * from run_job() won't be created until the dependencies have 143 * resolved. 144 */ 145 struct dma_fence finished; 146 147 /** 148 * @parent: the fence returned by &drm_sched_backend_ops.run_job 149 * when scheduling the job on hardware. We signal the 150 * &drm_sched_fence.finished fence once parent is signalled. 151 */ 152 struct dma_fence *parent; 153 /** 154 * @sched: the scheduler instance to which the job having this struct 155 * belongs to. 156 */ 157 struct drm_gpu_scheduler *sched; 158 /** 159 * @lock: the lock used by the scheduled and the finished fences. 160 */ 161 spinlock_t lock; 162 /** 163 * @owner: job owner for debugging 164 */ 165 void *owner; 166 }; 167 168 struct drm_sched_fence *to_drm_sched_fence(struct dma_fence *f); 169 170 /** 171 * struct drm_sched_job - A job to be run by an entity. 172 * 173 * @queue_node: used to append this struct to the queue of jobs in an entity. 174 * @sched: the scheduler instance on which this job is scheduled. 175 * @s_fence: contains the fences for the scheduling of job. 176 * @finish_cb: the callback for the finished fence. 177 * @node: used to append this struct to the @drm_gpu_scheduler.ring_mirror_list. 178 * @id: a unique id assigned to each job scheduled on the scheduler. 179 * @karma: increment on every hang caused by this job. If this exceeds the hang 180 * limit of the scheduler then the job is marked guilty and will not 181 * be scheduled further. 182 * @s_priority: the priority of the job. 183 * @entity: the entity to which this job belongs. 184 * @cb: the callback for the parent fence in s_fence. 185 * 186 * A job is created by the driver using drm_sched_job_init(), and 187 * should call drm_sched_entity_push_job() once it wants the scheduler 188 * to schedule the job. 189 */ 190 struct drm_sched_job { 191 struct spsc_node queue_node; 192 struct drm_gpu_scheduler *sched; 193 struct drm_sched_fence *s_fence; 194 struct dma_fence_cb finish_cb; 195 struct list_head node; 196 uint64_t id; 197 atomic_t karma; 198 enum drm_sched_priority s_priority; 199 struct drm_sched_entity *entity; 200 struct dma_fence_cb cb; 201 }; 202 203 static inline bool drm_sched_invalidate_job(struct drm_sched_job *s_job, 204 int threshold) 205 { 206 return (s_job && atomic_inc_return(&s_job->karma) > threshold); 207 } 208 209 /** 210 * struct drm_sched_backend_ops 211 * 212 * Define the backend operations called by the scheduler, 213 * these functions should be implemented in driver side. 214 */ 215 struct drm_sched_backend_ops { 216 /** 217 * @dependency: Called when the scheduler is considering scheduling 218 * this job next, to get another struct dma_fence for this job to 219 * block on. Once it returns NULL, run_job() may be called. 220 */ 221 struct dma_fence *(*dependency)(struct drm_sched_job *sched_job, 222 struct drm_sched_entity *s_entity); 223 224 /** 225 * @run_job: Called to execute the job once all of the dependencies 226 * have been resolved. This may be called multiple times, if 227 * timedout_job() has happened and drm_sched_job_recovery() 228 * decides to try it again. 229 */ 230 struct dma_fence *(*run_job)(struct drm_sched_job *sched_job); 231 232 /** 233 * @timedout_job: Called when a job has taken too long to execute, 234 * to trigger GPU recovery. 235 */ 236 void (*timedout_job)(struct drm_sched_job *sched_job); 237 238 /** 239 * @free_job: Called once the job's finished fence has been signaled 240 * and it's time to clean it up. 241 */ 242 void (*free_job)(struct drm_sched_job *sched_job); 243 }; 244 245 /** 246 * struct drm_gpu_scheduler 247 * 248 * @ops: backend operations provided by the driver. 249 * @hw_submission_limit: the max size of the hardware queue. 250 * @timeout: the time after which a job is removed from the scheduler. 251 * @name: name of the ring for which this scheduler is being used. 252 * @sched_rq: priority wise array of run queues. 253 * @wake_up_worker: the wait queue on which the scheduler sleeps until a job 254 * is ready to be scheduled. 255 * @job_scheduled: once @drm_sched_entity_do_release is called the scheduler 256 * waits on this wait queue until all the scheduled jobs are 257 * finished. 258 * @hw_rq_count: the number of jobs currently in the hardware queue. 259 * @job_id_count: used to assign unique id to the each job. 260 * @work_tdr: schedules a delayed call to @drm_sched_job_timedout after the 261 * timeout interval is over. 262 * @thread: the kthread on which the scheduler which run. 263 * @ring_mirror_list: the list of jobs which are currently in the job queue. 264 * @job_list_lock: lock to protect the ring_mirror_list. 265 * @hang_limit: once the hangs by a job crosses this limit then it is marked 266 * guilty and it will be considered for scheduling further. 267 * @score: score to help loadbalancer pick a idle sched 268 * @ready: marks if the underlying HW is ready to work 269 * @free_guilty: A hit to time out handler to free the guilty job. 270 * 271 * One scheduler is implemented for each hardware ring. 272 */ 273 struct drm_gpu_scheduler { 274 const struct drm_sched_backend_ops *ops; 275 uint32_t hw_submission_limit; 276 long timeout; 277 const char *name; 278 struct drm_sched_rq sched_rq[DRM_SCHED_PRIORITY_COUNT]; 279 wait_queue_head_t wake_up_worker; 280 wait_queue_head_t job_scheduled; 281 atomic_t hw_rq_count; 282 atomic64_t job_id_count; 283 struct delayed_work work_tdr; 284 struct task_struct *thread; 285 struct list_head ring_mirror_list; 286 spinlock_t job_list_lock; 287 int hang_limit; 288 atomic_t score; 289 bool ready; 290 bool free_guilty; 291 }; 292 293 int drm_sched_init(struct drm_gpu_scheduler *sched, 294 const struct drm_sched_backend_ops *ops, 295 uint32_t hw_submission, unsigned hang_limit, long timeout, 296 const char *name); 297 298 void drm_sched_fini(struct drm_gpu_scheduler *sched); 299 int drm_sched_job_init(struct drm_sched_job *job, 300 struct drm_sched_entity *entity, 301 void *owner); 302 void drm_sched_entity_modify_sched(struct drm_sched_entity *entity, 303 struct drm_gpu_scheduler **sched_list, 304 unsigned int num_sched_list); 305 306 void drm_sched_job_cleanup(struct drm_sched_job *job); 307 void drm_sched_wakeup(struct drm_gpu_scheduler *sched); 308 void drm_sched_stop(struct drm_gpu_scheduler *sched, struct drm_sched_job *bad); 309 void drm_sched_start(struct drm_gpu_scheduler *sched, bool full_recovery); 310 void drm_sched_resubmit_jobs(struct drm_gpu_scheduler *sched); 311 void drm_sched_increase_karma(struct drm_sched_job *bad); 312 bool drm_sched_dependency_optimized(struct dma_fence* fence, 313 struct drm_sched_entity *entity); 314 void drm_sched_fault(struct drm_gpu_scheduler *sched); 315 void drm_sched_job_kickout(struct drm_sched_job *s_job); 316 317 void drm_sched_rq_add_entity(struct drm_sched_rq *rq, 318 struct drm_sched_entity *entity); 319 void drm_sched_rq_remove_entity(struct drm_sched_rq *rq, 320 struct drm_sched_entity *entity); 321 322 int drm_sched_entity_init(struct drm_sched_entity *entity, 323 enum drm_sched_priority priority, 324 struct drm_gpu_scheduler **sched_list, 325 unsigned int num_sched_list, 326 atomic_t *guilty); 327 long drm_sched_entity_flush(struct drm_sched_entity *entity, long timeout); 328 void drm_sched_entity_fini(struct drm_sched_entity *entity); 329 void drm_sched_entity_destroy(struct drm_sched_entity *entity); 330 void drm_sched_entity_select_rq(struct drm_sched_entity *entity); 331 struct drm_sched_job *drm_sched_entity_pop_job(struct drm_sched_entity *entity); 332 void drm_sched_entity_push_job(struct drm_sched_job *sched_job, 333 struct drm_sched_entity *entity); 334 void drm_sched_entity_set_priority(struct drm_sched_entity *entity, 335 enum drm_sched_priority priority); 336 bool drm_sched_entity_is_ready(struct drm_sched_entity *entity); 337 338 struct drm_sched_fence *drm_sched_fence_create( 339 struct drm_sched_entity *s_entity, void *owner); 340 void drm_sched_fence_scheduled(struct drm_sched_fence *fence); 341 void drm_sched_fence_finished(struct drm_sched_fence *fence); 342 343 unsigned long drm_sched_suspend_timeout(struct drm_gpu_scheduler *sched); 344 void drm_sched_resume_timeout(struct drm_gpu_scheduler *sched, 345 unsigned long remaining); 346 struct drm_gpu_scheduler * 347 drm_sched_pick_best(struct drm_gpu_scheduler **sched_list, 348 unsigned int num_sched_list); 349 350 #endif 351