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 #ifdef __linux__ 100 struct task_struct *last_user; 101 #else 102 struct process *last_user; 103 #endif 104 bool stopped; 105 struct completion entity_idle; 106 }; 107 108 /** 109 * struct drm_sched_rq - queue of entities to be scheduled. 110 * 111 * @lock: to modify the entities list. 112 * @sched: the scheduler to which this rq belongs to. 113 * @entities: list of the entities to be scheduled. 114 * @current_entity: the entity which is to be scheduled. 115 * 116 * Run queue is a set of entities scheduling command submissions for 117 * one specific ring. It implements the scheduling policy that selects 118 * the next entity to emit commands from. 119 */ 120 struct drm_sched_rq { 121 spinlock_t lock; 122 struct drm_gpu_scheduler *sched; 123 struct list_head entities; 124 struct drm_sched_entity *current_entity; 125 }; 126 127 /** 128 * struct drm_sched_fence - fences corresponding to the scheduling of a job. 129 */ 130 struct drm_sched_fence { 131 /** 132 * @scheduled: this fence is what will be signaled by the scheduler 133 * when the job is scheduled. 134 */ 135 struct dma_fence scheduled; 136 137 /** 138 * @finished: this fence is what will be signaled by the scheduler 139 * when the job is completed. 140 * 141 * When setting up an out fence for the job, you should use 142 * this, since it's available immediately upon 143 * drm_sched_job_init(), and the fence returned by the driver 144 * from run_job() won't be created until the dependencies have 145 * resolved. 146 */ 147 struct dma_fence finished; 148 149 /** 150 * @parent: the fence returned by &drm_sched_backend_ops.run_job 151 * when scheduling the job on hardware. We signal the 152 * &drm_sched_fence.finished fence once parent is signalled. 153 */ 154 struct dma_fence *parent; 155 /** 156 * @sched: the scheduler instance to which the job having this struct 157 * belongs to. 158 */ 159 struct drm_gpu_scheduler *sched; 160 /** 161 * @lock: the lock used by the scheduled and the finished fences. 162 */ 163 spinlock_t lock; 164 /** 165 * @owner: job owner for debugging 166 */ 167 void *owner; 168 }; 169 170 struct drm_sched_fence *to_drm_sched_fence(struct dma_fence *f); 171 172 /** 173 * struct drm_sched_job - A job to be run by an entity. 174 * 175 * @queue_node: used to append this struct to the queue of jobs in an entity. 176 * @sched: the scheduler instance on which this job is scheduled. 177 * @s_fence: contains the fences for the scheduling of job. 178 * @finish_cb: the callback for the finished fence. 179 * @node: used to append this struct to the @drm_gpu_scheduler.ring_mirror_list. 180 * @id: a unique id assigned to each job scheduled on the scheduler. 181 * @karma: increment on every hang caused by this job. If this exceeds the hang 182 * limit of the scheduler then the job is marked guilty and will not 183 * be scheduled further. 184 * @s_priority: the priority of the job. 185 * @entity: the entity to which this job belongs. 186 * @cb: the callback for the parent fence in s_fence. 187 * 188 * A job is created by the driver using drm_sched_job_init(), and 189 * should call drm_sched_entity_push_job() once it wants the scheduler 190 * to schedule the job. 191 */ 192 struct drm_sched_job { 193 struct spsc_node queue_node; 194 struct drm_gpu_scheduler *sched; 195 struct drm_sched_fence *s_fence; 196 struct dma_fence_cb finish_cb; 197 struct list_head node; 198 uint64_t id; 199 atomic_t karma; 200 enum drm_sched_priority s_priority; 201 struct drm_sched_entity *entity; 202 struct dma_fence_cb cb; 203 }; 204 205 static inline bool drm_sched_invalidate_job(struct drm_sched_job *s_job, 206 int threshold) 207 { 208 return (s_job && atomic_inc_return(&s_job->karma) > threshold); 209 } 210 211 /** 212 * struct drm_sched_backend_ops 213 * 214 * Define the backend operations called by the scheduler, 215 * these functions should be implemented in driver side. 216 */ 217 struct drm_sched_backend_ops { 218 /** 219 * @dependency: Called when the scheduler is considering scheduling 220 * this job next, to get another struct dma_fence for this job to 221 * block on. Once it returns NULL, run_job() may be called. 222 */ 223 struct dma_fence *(*dependency)(struct drm_sched_job *sched_job, 224 struct drm_sched_entity *s_entity); 225 226 /** 227 * @run_job: Called to execute the job once all of the dependencies 228 * have been resolved. This may be called multiple times, if 229 * timedout_job() has happened and drm_sched_job_recovery() 230 * decides to try it again. 231 */ 232 struct dma_fence *(*run_job)(struct drm_sched_job *sched_job); 233 234 /** 235 * @timedout_job: Called when a job has taken too long to execute, 236 * to trigger GPU recovery. 237 */ 238 void (*timedout_job)(struct drm_sched_job *sched_job); 239 240 /** 241 * @free_job: Called once the job's finished fence has been signaled 242 * and it's time to clean it up. 243 */ 244 void (*free_job)(struct drm_sched_job *sched_job); 245 }; 246 247 /** 248 * struct drm_gpu_scheduler 249 * 250 * @ops: backend operations provided by the driver. 251 * @hw_submission_limit: the max size of the hardware queue. 252 * @timeout: the time after which a job is removed from the scheduler. 253 * @name: name of the ring for which this scheduler is being used. 254 * @sched_rq: priority wise array of run queues. 255 * @wake_up_worker: the wait queue on which the scheduler sleeps until a job 256 * is ready to be scheduled. 257 * @job_scheduled: once @drm_sched_entity_do_release is called the scheduler 258 * waits on this wait queue until all the scheduled jobs are 259 * finished. 260 * @hw_rq_count: the number of jobs currently in the hardware queue. 261 * @job_id_count: used to assign unique id to the each job. 262 * @work_tdr: schedules a delayed call to @drm_sched_job_timedout after the 263 * timeout interval is over. 264 * @thread: the kthread on which the scheduler which run. 265 * @ring_mirror_list: the list of jobs which are currently in the job queue. 266 * @job_list_lock: lock to protect the ring_mirror_list. 267 * @hang_limit: once the hangs by a job crosses this limit then it is marked 268 * guilty and it will be considered for scheduling further. 269 * @num_jobs: the number of jobs in queue in the scheduler 270 * @ready: marks if the underlying HW is ready to work 271 * @free_guilty: A hit to time out handler to free the guilty job. 272 * 273 * One scheduler is implemented for each hardware ring. 274 */ 275 struct drm_gpu_scheduler { 276 const struct drm_sched_backend_ops *ops; 277 uint32_t hw_submission_limit; 278 long timeout; 279 const char *name; 280 struct drm_sched_rq sched_rq[DRM_SCHED_PRIORITY_MAX]; 281 wait_queue_head_t wake_up_worker; 282 wait_queue_head_t job_scheduled; 283 atomic_t hw_rq_count; 284 atomic64_t job_id_count; 285 struct delayed_work work_tdr; 286 #ifdef __linux__ 287 struct task_struct *thread; 288 #else 289 struct proc *thread; 290 #endif 291 struct list_head ring_mirror_list; 292 spinlock_t job_list_lock; 293 int hang_limit; 294 atomic_t num_jobs; 295 bool ready; 296 bool free_guilty; 297 }; 298 299 int drm_sched_init(struct drm_gpu_scheduler *sched, 300 const struct drm_sched_backend_ops *ops, 301 uint32_t hw_submission, unsigned hang_limit, long timeout, 302 const char *name); 303 304 void drm_sched_fini(struct drm_gpu_scheduler *sched); 305 int drm_sched_job_init(struct drm_sched_job *job, 306 struct drm_sched_entity *entity, 307 void *owner); 308 void drm_sched_entity_modify_sched(struct drm_sched_entity *entity, 309 struct drm_gpu_scheduler **sched_list, 310 unsigned int num_sched_list); 311 312 void drm_sched_job_cleanup(struct drm_sched_job *job); 313 void drm_sched_wakeup(struct drm_gpu_scheduler *sched); 314 void drm_sched_stop(struct drm_gpu_scheduler *sched, struct drm_sched_job *bad); 315 void drm_sched_start(struct drm_gpu_scheduler *sched, bool full_recovery); 316 void drm_sched_resubmit_jobs(struct drm_gpu_scheduler *sched); 317 void drm_sched_increase_karma(struct drm_sched_job *bad); 318 bool drm_sched_dependency_optimized(struct dma_fence* fence, 319 struct drm_sched_entity *entity); 320 void drm_sched_fault(struct drm_gpu_scheduler *sched); 321 void drm_sched_job_kickout(struct drm_sched_job *s_job); 322 323 void drm_sched_rq_add_entity(struct drm_sched_rq *rq, 324 struct drm_sched_entity *entity); 325 void drm_sched_rq_remove_entity(struct drm_sched_rq *rq, 326 struct drm_sched_entity *entity); 327 328 int drm_sched_entity_init(struct drm_sched_entity *entity, 329 enum drm_sched_priority priority, 330 struct drm_gpu_scheduler **sched_list, 331 unsigned int num_sched_list, 332 atomic_t *guilty); 333 long drm_sched_entity_flush(struct drm_sched_entity *entity, long timeout); 334 void drm_sched_entity_fini(struct drm_sched_entity *entity); 335 void drm_sched_entity_destroy(struct drm_sched_entity *entity); 336 void drm_sched_entity_select_rq(struct drm_sched_entity *entity); 337 struct drm_sched_job *drm_sched_entity_pop_job(struct drm_sched_entity *entity); 338 void drm_sched_entity_push_job(struct drm_sched_job *sched_job, 339 struct drm_sched_entity *entity); 340 void drm_sched_entity_set_priority(struct drm_sched_entity *entity, 341 enum drm_sched_priority priority); 342 bool drm_sched_entity_is_ready(struct drm_sched_entity *entity); 343 344 struct drm_sched_fence *drm_sched_fence_create( 345 struct drm_sched_entity *s_entity, void *owner); 346 void drm_sched_fence_scheduled(struct drm_sched_fence *fence); 347 void drm_sched_fence_finished(struct drm_sched_fence *fence); 348 349 unsigned long drm_sched_suspend_timeout(struct drm_gpu_scheduler *sched); 350 void drm_sched_resume_timeout(struct drm_gpu_scheduler *sched, 351 unsigned long remaining); 352 struct drm_gpu_scheduler * 353 drm_sched_pick_best(struct drm_gpu_scheduler **sched_list, 354 unsigned int num_sched_list); 355 356 #endif 357