| 1 | /* |
| 2 | * Completely Fair Scheduling (CFS) Class (SCHED_NORMAL/SCHED_BATCH) |
| 3 | * |
| 4 | * Copyright (C) 2007 Red Hat, Inc., Ingo Molnar <mingo@redhat.com> |
| 5 | * |
| 6 | * Interactivity improvements by Mike Galbraith |
| 7 | * (C) 2007 Mike Galbraith <efault@gmx.de> |
| 8 | * |
| 9 | * Various enhancements by Dmitry Adamushko. |
| 10 | * (C) 2007 Dmitry Adamushko <dmitry.adamushko@gmail.com> |
| 11 | * |
| 12 | * Group scheduling enhancements by Srivatsa Vaddagiri |
| 13 | * Copyright IBM Corporation, 2007 |
| 14 | * Author: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com> |
| 15 | * |
| 16 | * Scaled math optimizations by Thomas Gleixner |
| 17 | * Copyright (C) 2007, Thomas Gleixner <tglx@linutronix.de> |
| 18 | * |
| 19 | * Adaptive scheduling granularity, math enhancements by Peter Zijlstra |
| 20 | * Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com> |
| 21 | */ |
| 22 | |
| 23 | #include <linux/latencytop.h> |
| 24 | |
| 25 | /* |
| 26 | * Targeted preemption latency for CPU-bound tasks: |
| 27 | * (default: 20ms * (1 + ilog(ncpus)), units: nanoseconds) |
| 28 | * |
| 29 | * NOTE: this latency value is not the same as the concept of |
| 30 | * 'timeslice length' - timeslices in CFS are of variable length |
| 31 | * and have no persistent notion like in traditional, time-slice |
| 32 | * based scheduling concepts. |
| 33 | * |
| 34 | * (to see the precise effective timeslice length of your workload, |
| 35 | * run vmstat and monitor the context-switches (cs) field) |
| 36 | */ |
| 37 | unsigned int sysctl_sched_latency = 20000000ULL; |
| 38 | |
| 39 | /* |
| 40 | * Minimal preemption granularity for CPU-bound tasks: |
| 41 | * (default: 4 msec * (1 + ilog(ncpus)), units: nanoseconds) |
| 42 | */ |
| 43 | unsigned int sysctl_sched_min_granularity = 4000000ULL; |
| 44 | |
| 45 | /* |
| 46 | * is kept at sysctl_sched_latency / sysctl_sched_min_granularity |
| 47 | */ |
| 48 | static unsigned int sched_nr_latency = 5; |
| 49 | |
| 50 | /* |
| 51 | * After fork, child runs first. (default) If set to 0 then |
| 52 | * parent will (try to) run first. |
| 53 | */ |
| 54 | const_debug unsigned int sysctl_sched_child_runs_first = 1; |
| 55 | |
| 56 | /* |
| 57 | * sys_sched_yield() compat mode |
| 58 | * |
| 59 | * This option switches the agressive yield implementation of the |
| 60 | * old scheduler back on. |
| 61 | */ |
| 62 | unsigned int __read_mostly sysctl_sched_compat_yield; |
| 63 | |
| 64 | /* |
| 65 | * SCHED_OTHER wake-up granularity. |
| 66 | * (default: 10 msec * (1 + ilog(ncpus)), units: nanoseconds) |
| 67 | * |
| 68 | * This option delays the preemption effects of decoupled workloads |
| 69 | * and reduces their over-scheduling. Synchronous workloads will still |
| 70 | * have immediate wakeup/sleep latencies. |
| 71 | */ |
| 72 | unsigned int sysctl_sched_wakeup_granularity = 10000000UL; |
| 73 | |
| 74 | const_debug unsigned int sysctl_sched_migration_cost = 500000UL; |
| 75 | |
| 76 | /************************************************************** |
| 77 | * CFS operations on generic schedulable entities: |
| 78 | */ |
| 79 | |
| 80 | #ifdef CONFIG_FAIR_GROUP_SCHED |
| 81 | |
| 82 | /* cpu runqueue to which this cfs_rq is attached */ |
| 83 | static inline struct rq *rq_of(struct cfs_rq *cfs_rq) |
| 84 | { |
| 85 | return cfs_rq->rq; |
| 86 | } |
| 87 | |
| 88 | /* An entity is a task if it doesn't "own" a runqueue */ |
| 89 | #define entity_is_task(se) (!se->my_q) |
| 90 | |
| 91 | #else /* CONFIG_FAIR_GROUP_SCHED */ |
| 92 | |
| 93 | static inline struct rq *rq_of(struct cfs_rq *cfs_rq) |
| 94 | { |
| 95 | return container_of(cfs_rq, struct rq, cfs); |
| 96 | } |
| 97 | |
| 98 | #define entity_is_task(se) 1 |
| 99 | |
| 100 | #endif /* CONFIG_FAIR_GROUP_SCHED */ |
| 101 | |
| 102 | static inline struct task_struct *task_of(struct sched_entity *se) |
| 103 | { |
| 104 | return container_of(se, struct task_struct, se); |
| 105 | } |
| 106 | |
| 107 | |
| 108 | /************************************************************** |
| 109 | * Scheduling class tree data structure manipulation methods: |
| 110 | */ |
| 111 | |
| 112 | static inline u64 max_vruntime(u64 min_vruntime, u64 vruntime) |
| 113 | { |
| 114 | s64 delta = (s64)(vruntime - min_vruntime); |
| 115 | if (delta > 0) |
| 116 | min_vruntime = vruntime; |
| 117 | |
| 118 | return min_vruntime; |
| 119 | } |
| 120 | |
| 121 | static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime) |
| 122 | { |
| 123 | s64 delta = (s64)(vruntime - min_vruntime); |
| 124 | if (delta < 0) |
| 125 | min_vruntime = vruntime; |
| 126 | |
| 127 | return min_vruntime; |
| 128 | } |
| 129 | |
| 130 | static inline s64 entity_key(struct cfs_rq *cfs_rq, struct sched_entity *se) |
| 131 | { |
| 132 | return se->vruntime - cfs_rq->min_vruntime; |
| 133 | } |
| 134 | |
| 135 | /* |
| 136 | * Enqueue an entity into the rb-tree: |
| 137 | */ |
| 138 | static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se) |
| 139 | { |
| 140 | struct rb_node **link = &cfs_rq->tasks_timeline.rb_node; |
| 141 | struct rb_node *parent = NULL; |
| 142 | struct sched_entity *entry; |
| 143 | s64 key = entity_key(cfs_rq, se); |
| 144 | int leftmost = 1; |
| 145 | |
| 146 | /* |
| 147 | * Find the right place in the rbtree: |
| 148 | */ |
| 149 | while (*link) { |
| 150 | parent = *link; |
| 151 | entry = rb_entry(parent, struct sched_entity, run_node); |
| 152 | /* |
| 153 | * We dont care about collisions. Nodes with |
| 154 | * the same key stay together. |
| 155 | */ |
| 156 | if (key < entity_key(cfs_rq, entry)) { |
| 157 | link = &parent->rb_left; |
| 158 | } else { |
| 159 | link = &parent->rb_right; |
| 160 | leftmost = 0; |
| 161 | } |
| 162 | } |
| 163 | |
| 164 | /* |
| 165 | * Maintain a cache of leftmost tree entries (it is frequently |
| 166 | * used): |
| 167 | */ |
| 168 | if (leftmost) { |
| 169 | cfs_rq->rb_leftmost = &se->run_node; |
| 170 | /* |
| 171 | * maintain cfs_rq->min_vruntime to be a monotonic increasing |
| 172 | * value tracking the leftmost vruntime in the tree. |
| 173 | */ |
| 174 | cfs_rq->min_vruntime = |
| 175 | max_vruntime(cfs_rq->min_vruntime, se->vruntime); |
| 176 | } |
| 177 | |
| 178 | rb_link_node(&se->run_node, parent, link); |
| 179 | rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline); |
| 180 | } |
| 181 | |
| 182 | static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se) |
| 183 | { |
| 184 | if (cfs_rq->rb_leftmost == &se->run_node) { |
| 185 | struct rb_node *next_node; |
| 186 | struct sched_entity *next; |
| 187 | |
| 188 | next_node = rb_next(&se->run_node); |
| 189 | cfs_rq->rb_leftmost = next_node; |
| 190 | |
| 191 | if (next_node) { |
| 192 | next = rb_entry(next_node, |
| 193 | struct sched_entity, run_node); |
| 194 | cfs_rq->min_vruntime = |
| 195 | max_vruntime(cfs_rq->min_vruntime, |
| 196 | next->vruntime); |
| 197 | } |
| 198 | } |
| 199 | |
| 200 | if (cfs_rq->next == se) |
| 201 | cfs_rq->next = NULL; |
| 202 | |
| 203 | rb_erase(&se->run_node, &cfs_rq->tasks_timeline); |
| 204 | } |
| 205 | |
| 206 | static inline struct rb_node *first_fair(struct cfs_rq *cfs_rq) |
| 207 | { |
| 208 | return cfs_rq->rb_leftmost; |
| 209 | } |
| 210 | |
| 211 | static struct sched_entity *__pick_next_entity(struct cfs_rq *cfs_rq) |
| 212 | { |
| 213 | return rb_entry(first_fair(cfs_rq), struct sched_entity, run_node); |
| 214 | } |
| 215 | |
| 216 | static inline struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq) |
| 217 | { |
| 218 | struct rb_node *last = rb_last(&cfs_rq->tasks_timeline); |
| 219 | |
| 220 | if (!last) |
| 221 | return NULL; |
| 222 | |
| 223 | return rb_entry(last, struct sched_entity, run_node); |
| 224 | } |
| 225 | |
| 226 | /************************************************************** |
| 227 | * Scheduling class statistics methods: |
| 228 | */ |
| 229 | |
| 230 | #ifdef CONFIG_SCHED_DEBUG |
| 231 | int sched_nr_latency_handler(struct ctl_table *table, int write, |
| 232 | struct file *filp, void __user *buffer, size_t *lenp, |
| 233 | loff_t *ppos) |
| 234 | { |
| 235 | int ret = proc_dointvec_minmax(table, write, filp, buffer, lenp, ppos); |
| 236 | |
| 237 | if (ret || !write) |
| 238 | return ret; |
| 239 | |
| 240 | sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency, |
| 241 | sysctl_sched_min_granularity); |
| 242 | |
| 243 | return 0; |
| 244 | } |
| 245 | #endif |
| 246 | |
| 247 | /* |
| 248 | * The idea is to set a period in which each task runs once. |
| 249 | * |
| 250 | * When there are too many tasks (sysctl_sched_nr_latency) we have to stretch |
| 251 | * this period because otherwise the slices get too small. |
| 252 | * |
| 253 | * p = (nr <= nl) ? l : l*nr/nl |
| 254 | */ |
| 255 | static u64 __sched_period(unsigned long nr_running) |
| 256 | { |
| 257 | u64 period = sysctl_sched_latency; |
| 258 | unsigned long nr_latency = sched_nr_latency; |
| 259 | |
| 260 | if (unlikely(nr_running > nr_latency)) { |
| 261 | period = sysctl_sched_min_granularity; |
| 262 | period *= nr_running; |
| 263 | } |
| 264 | |
| 265 | return period; |
| 266 | } |
| 267 | |
| 268 | /* |
| 269 | * We calculate the wall-time slice from the period by taking a part |
| 270 | * proportional to the weight. |
| 271 | * |
| 272 | * s = p*w/rw |
| 273 | */ |
| 274 | static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se) |
| 275 | { |
| 276 | return calc_delta_mine(__sched_period(cfs_rq->nr_running), |
| 277 | se->load.weight, &cfs_rq->load); |
| 278 | } |
| 279 | |
| 280 | /* |
| 281 | * We calculate the vruntime slice. |
| 282 | * |
| 283 | * vs = s/w = p/rw |
| 284 | */ |
| 285 | static u64 __sched_vslice(unsigned long rq_weight, unsigned long nr_running) |
| 286 | { |
| 287 | u64 vslice = __sched_period(nr_running); |
| 288 | |
| 289 | vslice *= NICE_0_LOAD; |
| 290 | do_div(vslice, rq_weight); |
| 291 | |
| 292 | return vslice; |
| 293 | } |
| 294 | |
| 295 | static u64 sched_vslice_add(struct cfs_rq *cfs_rq, struct sched_entity *se) |
| 296 | { |
| 297 | return __sched_vslice(cfs_rq->load.weight + se->load.weight, |
| 298 | cfs_rq->nr_running + 1); |
| 299 | } |
| 300 | |
| 301 | /* |
| 302 | * Update the current task's runtime statistics. Skip current tasks that |
| 303 | * are not in our scheduling class. |
| 304 | */ |
| 305 | static inline void |
| 306 | __update_curr(struct cfs_rq *cfs_rq, struct sched_entity *curr, |
| 307 | unsigned long delta_exec) |
| 308 | { |
| 309 | unsigned long delta_exec_weighted; |
| 310 | |
| 311 | schedstat_set(curr->exec_max, max((u64)delta_exec, curr->exec_max)); |
| 312 | |
| 313 | curr->sum_exec_runtime += delta_exec; |
| 314 | schedstat_add(cfs_rq, exec_clock, delta_exec); |
| 315 | delta_exec_weighted = delta_exec; |
| 316 | if (unlikely(curr->load.weight != NICE_0_LOAD)) { |
| 317 | delta_exec_weighted = calc_delta_fair(delta_exec_weighted, |
| 318 | &curr->load); |
| 319 | } |
| 320 | curr->vruntime += delta_exec_weighted; |
| 321 | } |
| 322 | |
| 323 | static void update_curr(struct cfs_rq *cfs_rq) |
| 324 | { |
| 325 | struct sched_entity *curr = cfs_rq->curr; |
| 326 | u64 now = rq_of(cfs_rq)->clock; |
| 327 | unsigned long delta_exec; |
| 328 | |
| 329 | if (unlikely(!curr)) |
| 330 | return; |
| 331 | |
| 332 | /* |
| 333 | * Get the amount of time the current task was running |
| 334 | * since the last time we changed load (this cannot |
| 335 | * overflow on 32 bits): |
| 336 | */ |
| 337 | delta_exec = (unsigned long)(now - curr->exec_start); |
| 338 | |
| 339 | __update_curr(cfs_rq, curr, delta_exec); |
| 340 | curr->exec_start = now; |
| 341 | |
| 342 | if (entity_is_task(curr)) { |
| 343 | struct task_struct *curtask = task_of(curr); |
| 344 | |
| 345 | cpuacct_charge(curtask, delta_exec); |
| 346 | } |
| 347 | } |
| 348 | |
| 349 | static inline void |
| 350 | update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se) |
| 351 | { |
| 352 | schedstat_set(se->wait_start, rq_of(cfs_rq)->clock); |
| 353 | } |
| 354 | |
| 355 | /* |
| 356 | * Task is being enqueued - update stats: |
| 357 | */ |
| 358 | static void update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se) |
| 359 | { |
| 360 | /* |
| 361 | * Are we enqueueing a waiting task? (for current tasks |
| 362 | * a dequeue/enqueue event is a NOP) |
| 363 | */ |
| 364 | if (se != cfs_rq->curr) |
| 365 | update_stats_wait_start(cfs_rq, se); |
| 366 | } |
| 367 | |
| 368 | static void |
| 369 | update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se) |
| 370 | { |
| 371 | schedstat_set(se->wait_max, max(se->wait_max, |
| 372 | rq_of(cfs_rq)->clock - se->wait_start)); |
| 373 | schedstat_set(se->wait_count, se->wait_count + 1); |
| 374 | schedstat_set(se->wait_sum, se->wait_sum + |
| 375 | rq_of(cfs_rq)->clock - se->wait_start); |
| 376 | schedstat_set(se->wait_start, 0); |
| 377 | } |
| 378 | |
| 379 | static inline void |
| 380 | update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se) |
| 381 | { |
| 382 | /* |
| 383 | * Mark the end of the wait period if dequeueing a |
| 384 | * waiting task: |
| 385 | */ |
| 386 | if (se != cfs_rq->curr) |
| 387 | update_stats_wait_end(cfs_rq, se); |
| 388 | } |
| 389 | |
| 390 | /* |
| 391 | * We are picking a new current task - update its stats: |
| 392 | */ |
| 393 | static inline void |
| 394 | update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se) |
| 395 | { |
| 396 | /* |
| 397 | * We are starting a new run period: |
| 398 | */ |
| 399 | se->exec_start = rq_of(cfs_rq)->clock; |
| 400 | } |
| 401 | |
| 402 | /************************************************** |
| 403 | * Scheduling class queueing methods: |
| 404 | */ |
| 405 | |
| 406 | static void |
| 407 | account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se) |
| 408 | { |
| 409 | update_load_add(&cfs_rq->load, se->load.weight); |
| 410 | cfs_rq->nr_running++; |
| 411 | se->on_rq = 1; |
| 412 | } |
| 413 | |
| 414 | static void |
| 415 | account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se) |
| 416 | { |
| 417 | update_load_sub(&cfs_rq->load, se->load.weight); |
| 418 | cfs_rq->nr_running--; |
| 419 | se->on_rq = 0; |
| 420 | } |
| 421 | |
| 422 | static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se) |
| 423 | { |
| 424 | #ifdef CONFIG_SCHEDSTATS |
| 425 | if (se->sleep_start) { |
| 426 | u64 delta = rq_of(cfs_rq)->clock - se->sleep_start; |
| 427 | struct task_struct *tsk = task_of(se); |
| 428 | |
| 429 | if ((s64)delta < 0) |
| 430 | delta = 0; |
| 431 | |
| 432 | if (unlikely(delta > se->sleep_max)) |
| 433 | se->sleep_max = delta; |
| 434 | |
| 435 | se->sleep_start = 0; |
| 436 | se->sum_sleep_runtime += delta; |
| 437 | |
| 438 | account_scheduler_latency(tsk, delta >> 10, 1); |
| 439 | } |
| 440 | if (se->block_start) { |
| 441 | u64 delta = rq_of(cfs_rq)->clock - se->block_start; |
| 442 | struct task_struct *tsk = task_of(se); |
| 443 | |
| 444 | if ((s64)delta < 0) |
| 445 | delta = 0; |
| 446 | |
| 447 | if (unlikely(delta > se->block_max)) |
| 448 | se->block_max = delta; |
| 449 | |
| 450 | se->block_start = 0; |
| 451 | se->sum_sleep_runtime += delta; |
| 452 | |
| 453 | /* |
| 454 | * Blocking time is in units of nanosecs, so shift by 20 to |
| 455 | * get a milliseconds-range estimation of the amount of |
| 456 | * time that the task spent sleeping: |
| 457 | */ |
| 458 | if (unlikely(prof_on == SLEEP_PROFILING)) { |
| 459 | |
| 460 | profile_hits(SLEEP_PROFILING, (void *)get_wchan(tsk), |
| 461 | delta >> 20); |
| 462 | } |
| 463 | account_scheduler_latency(tsk, delta >> 10, 0); |
| 464 | } |
| 465 | #endif |
| 466 | } |
| 467 | |
| 468 | static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se) |
| 469 | { |
| 470 | #ifdef CONFIG_SCHED_DEBUG |
| 471 | s64 d = se->vruntime - cfs_rq->min_vruntime; |
| 472 | |
| 473 | if (d < 0) |
| 474 | d = -d; |
| 475 | |
| 476 | if (d > 3*sysctl_sched_latency) |
| 477 | schedstat_inc(cfs_rq, nr_spread_over); |
| 478 | #endif |
| 479 | } |
| 480 | |
| 481 | static void |
| 482 | place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial) |
| 483 | { |
| 484 | u64 vruntime; |
| 485 | |
| 486 | if (first_fair(cfs_rq)) { |
| 487 | vruntime = min_vruntime(cfs_rq->min_vruntime, |
| 488 | __pick_next_entity(cfs_rq)->vruntime); |
| 489 | } else |
| 490 | vruntime = cfs_rq->min_vruntime; |
| 491 | |
| 492 | /* |
| 493 | * The 'current' period is already promised to the current tasks, |
| 494 | * however the extra weight of the new task will slow them down a |
| 495 | * little, place the new task so that it fits in the slot that |
| 496 | * stays open at the end. |
| 497 | */ |
| 498 | if (initial && sched_feat(START_DEBIT)) |
| 499 | vruntime += sched_vslice_add(cfs_rq, se); |
| 500 | |
| 501 | if (!initial) { |
| 502 | /* sleeps upto a single latency don't count. */ |
| 503 | if (sched_feat(NEW_FAIR_SLEEPERS)) { |
| 504 | vruntime -= calc_delta_fair(sysctl_sched_latency, |
| 505 | &cfs_rq->load); |
| 506 | } |
| 507 | |
| 508 | /* ensure we never gain time by being placed backwards. */ |
| 509 | vruntime = max_vruntime(se->vruntime, vruntime); |
| 510 | } |
| 511 | |
| 512 | se->vruntime = vruntime; |
| 513 | } |
| 514 | |
| 515 | static void |
| 516 | enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int wakeup) |
| 517 | { |
| 518 | /* |
| 519 | * Update run-time statistics of the 'current'. |
| 520 | */ |
| 521 | update_curr(cfs_rq); |
| 522 | |
| 523 | if (wakeup) { |
| 524 | place_entity(cfs_rq, se, 0); |
| 525 | enqueue_sleeper(cfs_rq, se); |
| 526 | } |
| 527 | |
| 528 | update_stats_enqueue(cfs_rq, se); |
| 529 | check_spread(cfs_rq, se); |
| 530 | if (se != cfs_rq->curr) |
| 531 | __enqueue_entity(cfs_rq, se); |
| 532 | account_entity_enqueue(cfs_rq, se); |
| 533 | } |
| 534 | |
| 535 | static void update_avg(u64 *avg, u64 sample) |
| 536 | { |
| 537 | s64 diff = sample - *avg; |
| 538 | *avg += diff >> 3; |
| 539 | } |
| 540 | |
| 541 | static void update_avg_stats(struct cfs_rq *cfs_rq, struct sched_entity *se) |
| 542 | { |
| 543 | if (!se->last_wakeup) |
| 544 | return; |
| 545 | |
| 546 | update_avg(&se->avg_overlap, se->sum_exec_runtime - se->last_wakeup); |
| 547 | se->last_wakeup = 0; |
| 548 | } |
| 549 | |
| 550 | static void |
| 551 | dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int sleep) |
| 552 | { |
| 553 | /* |
| 554 | * Update run-time statistics of the 'current'. |
| 555 | */ |
| 556 | update_curr(cfs_rq); |
| 557 | |
| 558 | update_stats_dequeue(cfs_rq, se); |
| 559 | if (sleep) { |
| 560 | update_avg_stats(cfs_rq, se); |
| 561 | #ifdef CONFIG_SCHEDSTATS |
| 562 | if (entity_is_task(se)) { |
| 563 | struct task_struct *tsk = task_of(se); |
| 564 | |
| 565 | if (tsk->state & TASK_INTERRUPTIBLE) |
| 566 | se->sleep_start = rq_of(cfs_rq)->clock; |
| 567 | if (tsk->state & TASK_UNINTERRUPTIBLE) |
| 568 | se->block_start = rq_of(cfs_rq)->clock; |
| 569 | } |
| 570 | #endif |
| 571 | } |
| 572 | |
| 573 | if (se != cfs_rq->curr) |
| 574 | __dequeue_entity(cfs_rq, se); |
| 575 | account_entity_dequeue(cfs_rq, se); |
| 576 | } |
| 577 | |
| 578 | /* |
| 579 | * Preempt the current task with a newly woken task if needed: |
| 580 | */ |
| 581 | static void |
| 582 | check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr) |
| 583 | { |
| 584 | unsigned long ideal_runtime, delta_exec; |
| 585 | |
| 586 | ideal_runtime = sched_slice(cfs_rq, curr); |
| 587 | delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime; |
| 588 | if (delta_exec > ideal_runtime) |
| 589 | resched_task(rq_of(cfs_rq)->curr); |
| 590 | } |
| 591 | |
| 592 | static void |
| 593 | set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se) |
| 594 | { |
| 595 | /* 'current' is not kept within the tree. */ |
| 596 | if (se->on_rq) { |
| 597 | /* |
| 598 | * Any task has to be enqueued before it get to execute on |
| 599 | * a CPU. So account for the time it spent waiting on the |
| 600 | * runqueue. |
| 601 | */ |
| 602 | update_stats_wait_end(cfs_rq, se); |
| 603 | __dequeue_entity(cfs_rq, se); |
| 604 | } |
| 605 | |
| 606 | update_stats_curr_start(cfs_rq, se); |
| 607 | cfs_rq->curr = se; |
| 608 | #ifdef CONFIG_SCHEDSTATS |
| 609 | /* |
| 610 | * Track our maximum slice length, if the CPU's load is at |
| 611 | * least twice that of our own weight (i.e. dont track it |
| 612 | * when there are only lesser-weight tasks around): |
| 613 | */ |
| 614 | if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) { |
| 615 | se->slice_max = max(se->slice_max, |
| 616 | se->sum_exec_runtime - se->prev_sum_exec_runtime); |
| 617 | } |
| 618 | #endif |
| 619 | se->prev_sum_exec_runtime = se->sum_exec_runtime; |
| 620 | } |
| 621 | |
| 622 | static int |
| 623 | wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se); |
| 624 | |
| 625 | static struct sched_entity * |
| 626 | pick_next(struct cfs_rq *cfs_rq, struct sched_entity *se) |
| 627 | { |
| 628 | if (!cfs_rq->next) |
| 629 | return se; |
| 630 | |
| 631 | if (wakeup_preempt_entity(cfs_rq->next, se) != 0) |
| 632 | return se; |
| 633 | |
| 634 | return cfs_rq->next; |
| 635 | } |
| 636 | |
| 637 | static struct sched_entity *pick_next_entity(struct cfs_rq *cfs_rq) |
| 638 | { |
| 639 | struct sched_entity *se = NULL; |
| 640 | |
| 641 | if (first_fair(cfs_rq)) { |
| 642 | se = __pick_next_entity(cfs_rq); |
| 643 | se = pick_next(cfs_rq, se); |
| 644 | set_next_entity(cfs_rq, se); |
| 645 | } |
| 646 | |
| 647 | return se; |
| 648 | } |
| 649 | |
| 650 | static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev) |
| 651 | { |
| 652 | /* |
| 653 | * If still on the runqueue then deactivate_task() |
| 654 | * was not called and update_curr() has to be done: |
| 655 | */ |
| 656 | if (prev->on_rq) |
| 657 | update_curr(cfs_rq); |
| 658 | |
| 659 | check_spread(cfs_rq, prev); |
| 660 | if (prev->on_rq) { |
| 661 | update_stats_wait_start(cfs_rq, prev); |
| 662 | /* Put 'current' back into the tree. */ |
| 663 | __enqueue_entity(cfs_rq, prev); |
| 664 | } |
| 665 | cfs_rq->curr = NULL; |
| 666 | } |
| 667 | |
| 668 | static void |
| 669 | entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued) |
| 670 | { |
| 671 | /* |
| 672 | * Update run-time statistics of the 'current'. |
| 673 | */ |
| 674 | update_curr(cfs_rq); |
| 675 | |
| 676 | #ifdef CONFIG_SCHED_HRTICK |
| 677 | /* |
| 678 | * queued ticks are scheduled to match the slice, so don't bother |
| 679 | * validating it and just reschedule. |
| 680 | */ |
| 681 | if (queued) |
| 682 | return resched_task(rq_of(cfs_rq)->curr); |
| 683 | /* |
| 684 | * don't let the period tick interfere with the hrtick preemption |
| 685 | */ |
| 686 | if (!sched_feat(DOUBLE_TICK) && |
| 687 | hrtimer_active(&rq_of(cfs_rq)->hrtick_timer)) |
| 688 | return; |
| 689 | #endif |
| 690 | |
| 691 | if (cfs_rq->nr_running > 1 || !sched_feat(WAKEUP_PREEMPT)) |
| 692 | check_preempt_tick(cfs_rq, curr); |
| 693 | } |
| 694 | |
| 695 | /************************************************** |
| 696 | * CFS operations on tasks: |
| 697 | */ |
| 698 | |
| 699 | #ifdef CONFIG_FAIR_GROUP_SCHED |
| 700 | |
| 701 | /* Walk up scheduling entities hierarchy */ |
| 702 | #define for_each_sched_entity(se) \ |
| 703 | for (; se; se = se->parent) |
| 704 | |
| 705 | static inline struct cfs_rq *task_cfs_rq(struct task_struct *p) |
| 706 | { |
| 707 | return p->se.cfs_rq; |
| 708 | } |
| 709 | |
| 710 | /* runqueue on which this entity is (to be) queued */ |
| 711 | static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se) |
| 712 | { |
| 713 | return se->cfs_rq; |
| 714 | } |
| 715 | |
| 716 | /* runqueue "owned" by this group */ |
| 717 | static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp) |
| 718 | { |
| 719 | return grp->my_q; |
| 720 | } |
| 721 | |
| 722 | /* Given a group's cfs_rq on one cpu, return its corresponding cfs_rq on |
| 723 | * another cpu ('this_cpu') |
| 724 | */ |
| 725 | static inline struct cfs_rq *cpu_cfs_rq(struct cfs_rq *cfs_rq, int this_cpu) |
| 726 | { |
| 727 | return cfs_rq->tg->cfs_rq[this_cpu]; |
| 728 | } |
| 729 | |
| 730 | /* Iterate thr' all leaf cfs_rq's on a runqueue */ |
| 731 | #define for_each_leaf_cfs_rq(rq, cfs_rq) \ |
| 732 | list_for_each_entry_rcu(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list) |
| 733 | |
| 734 | /* Do the two (enqueued) entities belong to the same group ? */ |
| 735 | static inline int |
| 736 | is_same_group(struct sched_entity *se, struct sched_entity *pse) |
| 737 | { |
| 738 | if (se->cfs_rq == pse->cfs_rq) |
| 739 | return 1; |
| 740 | |
| 741 | return 0; |
| 742 | } |
| 743 | |
| 744 | static inline struct sched_entity *parent_entity(struct sched_entity *se) |
| 745 | { |
| 746 | return se->parent; |
| 747 | } |
| 748 | |
| 749 | #else /* CONFIG_FAIR_GROUP_SCHED */ |
| 750 | |
| 751 | #define for_each_sched_entity(se) \ |
| 752 | for (; se; se = NULL) |
| 753 | |
| 754 | static inline struct cfs_rq *task_cfs_rq(struct task_struct *p) |
| 755 | { |
| 756 | return &task_rq(p)->cfs; |
| 757 | } |
| 758 | |
| 759 | static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se) |
| 760 | { |
| 761 | struct task_struct *p = task_of(se); |
| 762 | struct rq *rq = task_rq(p); |
| 763 | |
| 764 | return &rq->cfs; |
| 765 | } |
| 766 | |
| 767 | /* runqueue "owned" by this group */ |
| 768 | static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp) |
| 769 | { |
| 770 | return NULL; |
| 771 | } |
| 772 | |
| 773 | static inline struct cfs_rq *cpu_cfs_rq(struct cfs_rq *cfs_rq, int this_cpu) |
| 774 | { |
| 775 | return &cpu_rq(this_cpu)->cfs; |
| 776 | } |
| 777 | |
| 778 | #define for_each_leaf_cfs_rq(rq, cfs_rq) \ |
| 779 | for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL) |
| 780 | |
| 781 | static inline int |
| 782 | is_same_group(struct sched_entity *se, struct sched_entity *pse) |
| 783 | { |
| 784 | return 1; |
| 785 | } |
| 786 | |
| 787 | static inline struct sched_entity *parent_entity(struct sched_entity *se) |
| 788 | { |
| 789 | return NULL; |
| 790 | } |
| 791 | |
| 792 | #endif /* CONFIG_FAIR_GROUP_SCHED */ |
| 793 | |
| 794 | #ifdef CONFIG_SCHED_HRTICK |
| 795 | static void hrtick_start_fair(struct rq *rq, struct task_struct *p) |
| 796 | { |
| 797 | int requeue = rq->curr == p; |
| 798 | struct sched_entity *se = &p->se; |
| 799 | struct cfs_rq *cfs_rq = cfs_rq_of(se); |
| 800 | |
| 801 | WARN_ON(task_rq(p) != rq); |
| 802 | |
| 803 | if (hrtick_enabled(rq) && cfs_rq->nr_running > 1) { |
| 804 | u64 slice = sched_slice(cfs_rq, se); |
| 805 | u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime; |
| 806 | s64 delta = slice - ran; |
| 807 | |
| 808 | if (delta < 0) { |
| 809 | if (rq->curr == p) |
| 810 | resched_task(p); |
| 811 | return; |
| 812 | } |
| 813 | |
| 814 | /* |
| 815 | * Don't schedule slices shorter than 10000ns, that just |
| 816 | * doesn't make sense. Rely on vruntime for fairness. |
| 817 | */ |
| 818 | if (!requeue) |
| 819 | delta = max(10000LL, delta); |
| 820 | |
| 821 | hrtick_start(rq, delta, requeue); |
| 822 | } |
| 823 | } |
| 824 | #else |
| 825 | static inline void |
| 826 | hrtick_start_fair(struct rq *rq, struct task_struct *p) |
| 827 | { |
| 828 | } |
| 829 | #endif |
| 830 | |
| 831 | /* |
| 832 | * The enqueue_task method is called before nr_running is |
| 833 | * increased. Here we update the fair scheduling stats and |
| 834 | * then put the task into the rbtree: |
| 835 | */ |
| 836 | static void enqueue_task_fair(struct rq *rq, struct task_struct *p, int wakeup) |
| 837 | { |
| 838 | struct cfs_rq *cfs_rq; |
| 839 | struct sched_entity *se = &p->se; |
| 840 | |
| 841 | for_each_sched_entity(se) { |
| 842 | if (se->on_rq) |
| 843 | break; |
| 844 | cfs_rq = cfs_rq_of(se); |
| 845 | enqueue_entity(cfs_rq, se, wakeup); |
| 846 | wakeup = 1; |
| 847 | } |
| 848 | |
| 849 | hrtick_start_fair(rq, rq->curr); |
| 850 | } |
| 851 | |
| 852 | /* |
| 853 | * The dequeue_task method is called before nr_running is |
| 854 | * decreased. We remove the task from the rbtree and |
| 855 | * update the fair scheduling stats: |
| 856 | */ |
| 857 | static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int sleep) |
| 858 | { |
| 859 | struct cfs_rq *cfs_rq; |
| 860 | struct sched_entity *se = &p->se; |
| 861 | |
| 862 | for_each_sched_entity(se) { |
| 863 | cfs_rq = cfs_rq_of(se); |
| 864 | dequeue_entity(cfs_rq, se, sleep); |
| 865 | /* Don't dequeue parent if it has other entities besides us */ |
| 866 | if (cfs_rq->load.weight) |
| 867 | break; |
| 868 | sleep = 1; |
| 869 | } |
| 870 | |
| 871 | hrtick_start_fair(rq, rq->curr); |
| 872 | } |
| 873 | |
| 874 | /* |
| 875 | * sched_yield() support is very simple - we dequeue and enqueue. |
| 876 | * |
| 877 | * If compat_yield is turned on then we requeue to the end of the tree. |
| 878 | */ |
| 879 | static void yield_task_fair(struct rq *rq) |
| 880 | { |
| 881 | struct task_struct *curr = rq->curr; |
| 882 | struct cfs_rq *cfs_rq = task_cfs_rq(curr); |
| 883 | struct sched_entity *rightmost, *se = &curr->se; |
| 884 | |
| 885 | /* |
| 886 | * Are we the only task in the tree? |
| 887 | */ |
| 888 | if (unlikely(cfs_rq->nr_running == 1)) |
| 889 | return; |
| 890 | |
| 891 | if (likely(!sysctl_sched_compat_yield) && curr->policy != SCHED_BATCH) { |
| 892 | __update_rq_clock(rq); |
| 893 | /* |
| 894 | * Update run-time statistics of the 'current'. |
| 895 | */ |
| 896 | update_curr(cfs_rq); |
| 897 | |
| 898 | return; |
| 899 | } |
| 900 | /* |
| 901 | * Find the rightmost entry in the rbtree: |
| 902 | */ |
| 903 | rightmost = __pick_last_entity(cfs_rq); |
| 904 | /* |
| 905 | * Already in the rightmost position? |
| 906 | */ |
| 907 | if (unlikely(rightmost->vruntime < se->vruntime)) |
| 908 | return; |
| 909 | |
| 910 | /* |
| 911 | * Minimally necessary key value to be last in the tree: |
| 912 | * Upon rescheduling, sched_class::put_prev_task() will place |
| 913 | * 'current' within the tree based on its new key value. |
| 914 | */ |
| 915 | se->vruntime = rightmost->vruntime + 1; |
| 916 | } |
| 917 | |
| 918 | /* |
| 919 | * wake_idle() will wake a task on an idle cpu if task->cpu is |
| 920 | * not idle and an idle cpu is available. The span of cpus to |
| 921 | * search starts with cpus closest then further out as needed, |
| 922 | * so we always favor a closer, idle cpu. |
| 923 | * |
| 924 | * Returns the CPU we should wake onto. |
| 925 | */ |
| 926 | #if defined(ARCH_HAS_SCHED_WAKE_IDLE) |
| 927 | static int wake_idle(int cpu, struct task_struct *p) |
| 928 | { |
| 929 | cpumask_t tmp; |
| 930 | struct sched_domain *sd; |
| 931 | int i; |
| 932 | |
| 933 | /* |
| 934 | * If it is idle, then it is the best cpu to run this task. |
| 935 | * |
| 936 | * This cpu is also the best, if it has more than one task already. |
| 937 | * Siblings must be also busy(in most cases) as they didn't already |
| 938 | * pickup the extra load from this cpu and hence we need not check |
| 939 | * sibling runqueue info. This will avoid the checks and cache miss |
| 940 | * penalities associated with that. |
| 941 | */ |
| 942 | if (idle_cpu(cpu) || cpu_rq(cpu)->nr_running > 1) |
| 943 | return cpu; |
| 944 | |
| 945 | for_each_domain(cpu, sd) { |
| 946 | if (sd->flags & SD_WAKE_IDLE) { |
| 947 | cpus_and(tmp, sd->span, p->cpus_allowed); |
| 948 | for_each_cpu_mask(i, tmp) { |
| 949 | if (idle_cpu(i)) { |
| 950 | if (i != task_cpu(p)) { |
| 951 | schedstat_inc(p, |
| 952 | se.nr_wakeups_idle); |
| 953 | } |
| 954 | return i; |
| 955 | } |
| 956 | } |
| 957 | } else { |
| 958 | break; |
| 959 | } |
| 960 | } |
| 961 | return cpu; |
| 962 | } |
| 963 | #else |
| 964 | static inline int wake_idle(int cpu, struct task_struct *p) |
| 965 | { |
| 966 | return cpu; |
| 967 | } |
| 968 | #endif |
| 969 | |
| 970 | #ifdef CONFIG_SMP |
| 971 | |
| 972 | static const struct sched_class fair_sched_class; |
| 973 | |
| 974 | static int |
| 975 | wake_affine(struct rq *rq, struct sched_domain *this_sd, struct rq *this_rq, |
| 976 | struct task_struct *p, int prev_cpu, int this_cpu, int sync, |
| 977 | int idx, unsigned long load, unsigned long this_load, |
| 978 | unsigned int imbalance) |
| 979 | { |
| 980 | struct task_struct *curr = this_rq->curr; |
| 981 | unsigned long tl = this_load; |
| 982 | unsigned long tl_per_task; |
| 983 | |
| 984 | if (!(this_sd->flags & SD_WAKE_AFFINE)) |
| 985 | return 0; |
| 986 | |
| 987 | /* |
| 988 | * If the currently running task will sleep within |
| 989 | * a reasonable amount of time then attract this newly |
| 990 | * woken task: |
| 991 | */ |
| 992 | if (sync && curr->sched_class == &fair_sched_class) { |
| 993 | if (curr->se.avg_overlap < sysctl_sched_migration_cost && |
| 994 | p->se.avg_overlap < sysctl_sched_migration_cost) |
| 995 | return 1; |
| 996 | } |
| 997 | |
| 998 | schedstat_inc(p, se.nr_wakeups_affine_attempts); |
| 999 | tl_per_task = cpu_avg_load_per_task(this_cpu); |
| 1000 | |
| 1001 | /* |
| 1002 | * If sync wakeup then subtract the (maximum possible) |
| 1003 | * effect of the currently running task from the load |
| 1004 | * of the current CPU: |
| 1005 | */ |
| 1006 | if (sync) |
| 1007 | tl -= current->se.load.weight; |
| 1008 | |
| 1009 | if ((tl <= load && tl + target_load(prev_cpu, idx) <= tl_per_task) || |
| 1010 | 100*(tl + p->se.load.weight) <= imbalance*load) { |
| 1011 | /* |
| 1012 | * This domain has SD_WAKE_AFFINE and |
| 1013 | * p is cache cold in this domain, and |
| 1014 | * there is no bad imbalance. |
| 1015 | */ |
| 1016 | schedstat_inc(this_sd, ttwu_move_affine); |
| 1017 | schedstat_inc(p, se.nr_wakeups_affine); |
| 1018 | |
| 1019 | return 1; |
| 1020 | } |
| 1021 | return 0; |
| 1022 | } |
| 1023 | |
| 1024 | static int select_task_rq_fair(struct task_struct *p, int sync) |
| 1025 | { |
| 1026 | struct sched_domain *sd, *this_sd = NULL; |
| 1027 | int prev_cpu, this_cpu, new_cpu; |
| 1028 | unsigned long load, this_load; |
| 1029 | struct rq *rq, *this_rq; |
| 1030 | unsigned int imbalance; |
| 1031 | int idx; |
| 1032 | |
| 1033 | prev_cpu = task_cpu(p); |
| 1034 | rq = task_rq(p); |
| 1035 | this_cpu = smp_processor_id(); |
| 1036 | this_rq = cpu_rq(this_cpu); |
| 1037 | new_cpu = prev_cpu; |
| 1038 | |
| 1039 | /* |
| 1040 | * 'this_sd' is the first domain that both |
| 1041 | * this_cpu and prev_cpu are present in: |
| 1042 | */ |
| 1043 | for_each_domain(this_cpu, sd) { |
| 1044 | if (cpu_isset(prev_cpu, sd->span)) { |
| 1045 | this_sd = sd; |
| 1046 | break; |
| 1047 | } |
| 1048 | } |
| 1049 | |
| 1050 | if (unlikely(!cpu_isset(this_cpu, p->cpus_allowed))) |
| 1051 | goto out; |
| 1052 | |
| 1053 | /* |
| 1054 | * Check for affine wakeup and passive balancing possibilities. |
| 1055 | */ |
| 1056 | if (!this_sd) |
| 1057 | goto out; |
| 1058 | |
| 1059 | idx = this_sd->wake_idx; |
| 1060 | |
| 1061 | imbalance = 100 + (this_sd->imbalance_pct - 100) / 2; |
| 1062 | |
| 1063 | load = source_load(prev_cpu, idx); |
| 1064 | this_load = target_load(this_cpu, idx); |
| 1065 | |
| 1066 | if (wake_affine(rq, this_sd, this_rq, p, prev_cpu, this_cpu, sync, idx, |
| 1067 | load, this_load, imbalance)) |
| 1068 | return this_cpu; |
| 1069 | |
| 1070 | if (prev_cpu == this_cpu) |
| 1071 | goto out; |
| 1072 | |
| 1073 | /* |
| 1074 | * Start passive balancing when half the imbalance_pct |
| 1075 | * limit is reached. |
| 1076 | */ |
| 1077 | if (this_sd->flags & SD_WAKE_BALANCE) { |
| 1078 | if (imbalance*this_load <= 100*load) { |
| 1079 | schedstat_inc(this_sd, ttwu_move_balance); |
| 1080 | schedstat_inc(p, se.nr_wakeups_passive); |
| 1081 | return this_cpu; |
| 1082 | } |
| 1083 | } |
| 1084 | |
| 1085 | out: |
| 1086 | return wake_idle(new_cpu, p); |
| 1087 | } |
| 1088 | #endif /* CONFIG_SMP */ |
| 1089 | |
| 1090 | static unsigned long wakeup_gran(struct sched_entity *se) |
| 1091 | { |
| 1092 | unsigned long gran = sysctl_sched_wakeup_granularity; |
| 1093 | |
| 1094 | /* |
| 1095 | * More easily preempt - nice tasks, while not making |
| 1096 | * it harder for + nice tasks. |
| 1097 | */ |
| 1098 | if (unlikely(se->load.weight > NICE_0_LOAD)) |
| 1099 | gran = calc_delta_fair(gran, &se->load); |
| 1100 | |
| 1101 | return gran; |
| 1102 | } |
| 1103 | |
| 1104 | /* |
| 1105 | * Should 'se' preempt 'curr'. |
| 1106 | * |
| 1107 | * |s1 |
| 1108 | * |s2 |
| 1109 | * |s3 |
| 1110 | * g |
| 1111 | * |<--->|c |
| 1112 | * |
| 1113 | * w(c, s1) = -1 |
| 1114 | * w(c, s2) = 0 |
| 1115 | * w(c, s3) = 1 |
| 1116 | * |
| 1117 | */ |
| 1118 | static int |
| 1119 | wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se) |
| 1120 | { |
| 1121 | s64 gran, vdiff = curr->vruntime - se->vruntime; |
| 1122 | |
| 1123 | if (vdiff < 0) |
| 1124 | return -1; |
| 1125 | |
| 1126 | gran = wakeup_gran(curr); |
| 1127 | if (vdiff > gran) |
| 1128 | return 1; |
| 1129 | |
| 1130 | return 0; |
| 1131 | } |
| 1132 | |
| 1133 | /* |
| 1134 | * Preempt the current task with a newly woken task if needed: |
| 1135 | */ |
| 1136 | static void check_preempt_wakeup(struct rq *rq, struct task_struct *p) |
| 1137 | { |
| 1138 | struct task_struct *curr = rq->curr; |
| 1139 | struct cfs_rq *cfs_rq = task_cfs_rq(curr); |
| 1140 | struct sched_entity *se = &curr->se, *pse = &p->se; |
| 1141 | |
| 1142 | if (unlikely(rt_prio(p->prio))) { |
| 1143 | update_rq_clock(rq); |
| 1144 | update_curr(cfs_rq); |
| 1145 | resched_task(curr); |
| 1146 | return; |
| 1147 | } |
| 1148 | |
| 1149 | se->last_wakeup = se->sum_exec_runtime; |
| 1150 | if (unlikely(se == pse)) |
| 1151 | return; |
| 1152 | |
| 1153 | cfs_rq_of(pse)->next = pse; |
| 1154 | |
| 1155 | /* |
| 1156 | * Batch tasks do not preempt (their preemption is driven by |
| 1157 | * the tick): |
| 1158 | */ |
| 1159 | if (unlikely(p->policy == SCHED_BATCH)) |
| 1160 | return; |
| 1161 | |
| 1162 | if (!sched_feat(WAKEUP_PREEMPT)) |
| 1163 | return; |
| 1164 | |
| 1165 | while (!is_same_group(se, pse)) { |
| 1166 | se = parent_entity(se); |
| 1167 | pse = parent_entity(pse); |
| 1168 | } |
| 1169 | |
| 1170 | if (wakeup_preempt_entity(se, pse) == 1) |
| 1171 | resched_task(curr); |
| 1172 | } |
| 1173 | |
| 1174 | static struct task_struct *pick_next_task_fair(struct rq *rq) |
| 1175 | { |
| 1176 | struct task_struct *p; |
| 1177 | struct cfs_rq *cfs_rq = &rq->cfs; |
| 1178 | struct sched_entity *se; |
| 1179 | |
| 1180 | if (unlikely(!cfs_rq->nr_running)) |
| 1181 | return NULL; |
| 1182 | |
| 1183 | do { |
| 1184 | se = pick_next_entity(cfs_rq); |
| 1185 | cfs_rq = group_cfs_rq(se); |
| 1186 | } while (cfs_rq); |
| 1187 | |
| 1188 | p = task_of(se); |
| 1189 | hrtick_start_fair(rq, p); |
| 1190 | |
| 1191 | return p; |
| 1192 | } |
| 1193 | |
| 1194 | /* |
| 1195 | * Account for a descheduled task: |
| 1196 | */ |
| 1197 | static void put_prev_task_fair(struct rq *rq, struct task_struct *prev) |
| 1198 | { |
| 1199 | struct sched_entity *se = &prev->se; |
| 1200 | struct cfs_rq *cfs_rq; |
| 1201 | |
| 1202 | for_each_sched_entity(se) { |
| 1203 | cfs_rq = cfs_rq_of(se); |
| 1204 | put_prev_entity(cfs_rq, se); |
| 1205 | } |
| 1206 | } |
| 1207 | |
| 1208 | #ifdef CONFIG_SMP |
| 1209 | /************************************************** |
| 1210 | * Fair scheduling class load-balancing methods: |
| 1211 | */ |
| 1212 | |
| 1213 | /* |
| 1214 | * Load-balancing iterator. Note: while the runqueue stays locked |
| 1215 | * during the whole iteration, the current task might be |
| 1216 | * dequeued so the iterator has to be dequeue-safe. Here we |
| 1217 | * achieve that by always pre-iterating before returning |
| 1218 | * the current task: |
| 1219 | */ |
| 1220 | static struct task_struct * |
| 1221 | __load_balance_iterator(struct cfs_rq *cfs_rq, struct rb_node *curr) |
| 1222 | { |
| 1223 | struct task_struct *p; |
| 1224 | |
| 1225 | if (!curr) |
| 1226 | return NULL; |
| 1227 | |
| 1228 | p = rb_entry(curr, struct task_struct, se.run_node); |
| 1229 | cfs_rq->rb_load_balance_curr = rb_next(curr); |
| 1230 | |
| 1231 | return p; |
| 1232 | } |
| 1233 | |
| 1234 | static struct task_struct *load_balance_start_fair(void *arg) |
| 1235 | { |
| 1236 | struct cfs_rq *cfs_rq = arg; |
| 1237 | |
| 1238 | return __load_balance_iterator(cfs_rq, first_fair(cfs_rq)); |
| 1239 | } |
| 1240 | |
| 1241 | static struct task_struct *load_balance_next_fair(void *arg) |
| 1242 | { |
| 1243 | struct cfs_rq *cfs_rq = arg; |
| 1244 | |
| 1245 | return __load_balance_iterator(cfs_rq, cfs_rq->rb_load_balance_curr); |
| 1246 | } |
| 1247 | |
| 1248 | #ifdef CONFIG_FAIR_GROUP_SCHED |
| 1249 | static int cfs_rq_best_prio(struct cfs_rq *cfs_rq) |
| 1250 | { |
| 1251 | struct sched_entity *curr; |
| 1252 | struct task_struct *p; |
| 1253 | |
| 1254 | if (!cfs_rq->nr_running || !first_fair(cfs_rq)) |
| 1255 | return MAX_PRIO; |
| 1256 | |
| 1257 | curr = cfs_rq->curr; |
| 1258 | if (!curr) |
| 1259 | curr = __pick_next_entity(cfs_rq); |
| 1260 | |
| 1261 | p = task_of(curr); |
| 1262 | |
| 1263 | return p->prio; |
| 1264 | } |
| 1265 | #endif |
| 1266 | |
| 1267 | static unsigned long |
| 1268 | load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest, |
| 1269 | unsigned long max_load_move, |
| 1270 | struct sched_domain *sd, enum cpu_idle_type idle, |
| 1271 | int *all_pinned, int *this_best_prio) |
| 1272 | { |
| 1273 | struct cfs_rq *busy_cfs_rq; |
| 1274 | long rem_load_move = max_load_move; |
| 1275 | struct rq_iterator cfs_rq_iterator; |
| 1276 | |
| 1277 | cfs_rq_iterator.start = load_balance_start_fair; |
| 1278 | cfs_rq_iterator.next = load_balance_next_fair; |
| 1279 | |
| 1280 | for_each_leaf_cfs_rq(busiest, busy_cfs_rq) { |
| 1281 | #ifdef CONFIG_FAIR_GROUP_SCHED |
| 1282 | struct cfs_rq *this_cfs_rq; |
| 1283 | long imbalance; |
| 1284 | unsigned long maxload; |
| 1285 | |
| 1286 | this_cfs_rq = cpu_cfs_rq(busy_cfs_rq, this_cpu); |
| 1287 | |
| 1288 | imbalance = busy_cfs_rq->load.weight - this_cfs_rq->load.weight; |
| 1289 | /* Don't pull if this_cfs_rq has more load than busy_cfs_rq */ |
| 1290 | if (imbalance <= 0) |
| 1291 | continue; |
| 1292 | |
| 1293 | /* Don't pull more than imbalance/2 */ |
| 1294 | imbalance /= 2; |
| 1295 | maxload = min(rem_load_move, imbalance); |
| 1296 | |
| 1297 | *this_best_prio = cfs_rq_best_prio(this_cfs_rq); |
| 1298 | #else |
| 1299 | # define maxload rem_load_move |
| 1300 | #endif |
| 1301 | /* |
| 1302 | * pass busy_cfs_rq argument into |
| 1303 | * load_balance_[start|next]_fair iterators |
| 1304 | */ |
| 1305 | cfs_rq_iterator.arg = busy_cfs_rq; |
| 1306 | rem_load_move -= balance_tasks(this_rq, this_cpu, busiest, |
| 1307 | maxload, sd, idle, all_pinned, |
| 1308 | this_best_prio, |
| 1309 | &cfs_rq_iterator); |
| 1310 | |
| 1311 | if (rem_load_move <= 0) |
| 1312 | break; |
| 1313 | } |
| 1314 | |
| 1315 | return max_load_move - rem_load_move; |
| 1316 | } |
| 1317 | |
| 1318 | static int |
| 1319 | move_one_task_fair(struct rq *this_rq, int this_cpu, struct rq *busiest, |
| 1320 | struct sched_domain *sd, enum cpu_idle_type idle) |
| 1321 | { |
| 1322 | struct cfs_rq *busy_cfs_rq; |
| 1323 | struct rq_iterator cfs_rq_iterator; |
| 1324 | |
| 1325 | cfs_rq_iterator.start = load_balance_start_fair; |
| 1326 | cfs_rq_iterator.next = load_balance_next_fair; |
| 1327 | |
| 1328 | for_each_leaf_cfs_rq(busiest, busy_cfs_rq) { |
| 1329 | /* |
| 1330 | * pass busy_cfs_rq argument into |
| 1331 | * load_balance_[start|next]_fair iterators |
| 1332 | */ |
| 1333 | cfs_rq_iterator.arg = busy_cfs_rq; |
| 1334 | if (iter_move_one_task(this_rq, this_cpu, busiest, sd, idle, |
| 1335 | &cfs_rq_iterator)) |
| 1336 | return 1; |
| 1337 | } |
| 1338 | |
| 1339 | return 0; |
| 1340 | } |
| 1341 | #endif |
| 1342 | |
| 1343 | /* |
| 1344 | * scheduler tick hitting a task of our scheduling class: |
| 1345 | */ |
| 1346 | static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued) |
| 1347 | { |
| 1348 | struct cfs_rq *cfs_rq; |
| 1349 | struct sched_entity *se = &curr->se; |
| 1350 | |
| 1351 | for_each_sched_entity(se) { |
| 1352 | cfs_rq = cfs_rq_of(se); |
| 1353 | entity_tick(cfs_rq, se, queued); |
| 1354 | } |
| 1355 | } |
| 1356 | |
| 1357 | #define swap(a, b) do { typeof(a) tmp = (a); (a) = (b); (b) = tmp; } while (0) |
| 1358 | |
| 1359 | /* |
| 1360 | * Share the fairness runtime between parent and child, thus the |
| 1361 | * total amount of pressure for CPU stays equal - new tasks |
| 1362 | * get a chance to run but frequent forkers are not allowed to |
| 1363 | * monopolize the CPU. Note: the parent runqueue is locked, |
| 1364 | * the child is not running yet. |
| 1365 | */ |
| 1366 | static void task_new_fair(struct rq *rq, struct task_struct *p) |
| 1367 | { |
| 1368 | struct cfs_rq *cfs_rq = task_cfs_rq(p); |
| 1369 | struct sched_entity *se = &p->se, *curr = cfs_rq->curr; |
| 1370 | int this_cpu = smp_processor_id(); |
| 1371 | |
| 1372 | sched_info_queued(p); |
| 1373 | |
| 1374 | update_curr(cfs_rq); |
| 1375 | place_entity(cfs_rq, se, 1); |
| 1376 | |
| 1377 | /* 'curr' will be NULL if the child belongs to a different group */ |
| 1378 | if (sysctl_sched_child_runs_first && this_cpu == task_cpu(p) && |
| 1379 | curr && curr->vruntime < se->vruntime) { |
| 1380 | /* |
| 1381 | * Upon rescheduling, sched_class::put_prev_task() will place |
| 1382 | * 'current' within the tree based on its new key value. |
| 1383 | */ |
| 1384 | swap(curr->vruntime, se->vruntime); |
| 1385 | } |
| 1386 | |
| 1387 | enqueue_task_fair(rq, p, 0); |
| 1388 | resched_task(rq->curr); |
| 1389 | } |
| 1390 | |
| 1391 | /* |
| 1392 | * Priority of the task has changed. Check to see if we preempt |
| 1393 | * the current task. |
| 1394 | */ |
| 1395 | static void prio_changed_fair(struct rq *rq, struct task_struct *p, |
| 1396 | int oldprio, int running) |
| 1397 | { |
| 1398 | /* |
| 1399 | * Reschedule if we are currently running on this runqueue and |
| 1400 | * our priority decreased, or if we are not currently running on |
| 1401 | * this runqueue and our priority is higher than the current's |
| 1402 | */ |
| 1403 | if (running) { |
| 1404 | if (p->prio > oldprio) |
| 1405 | resched_task(rq->curr); |
| 1406 | } else |
| 1407 | check_preempt_curr(rq, p); |
| 1408 | } |
| 1409 | |
| 1410 | /* |
| 1411 | * We switched to the sched_fair class. |
| 1412 | */ |
| 1413 | static void switched_to_fair(struct rq *rq, struct task_struct *p, |
| 1414 | int running) |
| 1415 | { |
| 1416 | /* |
| 1417 | * We were most likely switched from sched_rt, so |
| 1418 | * kick off the schedule if running, otherwise just see |
| 1419 | * if we can still preempt the current task. |
| 1420 | */ |
| 1421 | if (running) |
| 1422 | resched_task(rq->curr); |
| 1423 | else |
| 1424 | check_preempt_curr(rq, p); |
| 1425 | } |
| 1426 | |
| 1427 | /* Account for a task changing its policy or group. |
| 1428 | * |
| 1429 | * This routine is mostly called to set cfs_rq->curr field when a task |
| 1430 | * migrates between groups/classes. |
| 1431 | */ |
| 1432 | static void set_curr_task_fair(struct rq *rq) |
| 1433 | { |
| 1434 | struct sched_entity *se = &rq->curr->se; |
| 1435 | |
| 1436 | for_each_sched_entity(se) |
| 1437 | set_next_entity(cfs_rq_of(se), se); |
| 1438 | } |
| 1439 | |
| 1440 | #ifdef CONFIG_FAIR_GROUP_SCHED |
| 1441 | static void moved_group_fair(struct task_struct *p) |
| 1442 | { |
| 1443 | struct cfs_rq *cfs_rq = task_cfs_rq(p); |
| 1444 | |
| 1445 | update_curr(cfs_rq); |
| 1446 | place_entity(cfs_rq, &p->se, 1); |
| 1447 | } |
| 1448 | #endif |
| 1449 | |
| 1450 | /* |
| 1451 | * All the scheduling class methods: |
| 1452 | */ |
| 1453 | static const struct sched_class fair_sched_class = { |
| 1454 | .next = &idle_sched_class, |
| 1455 | .enqueue_task = enqueue_task_fair, |
| 1456 | .dequeue_task = dequeue_task_fair, |
| 1457 | .yield_task = yield_task_fair, |
| 1458 | #ifdef CONFIG_SMP |
| 1459 | .select_task_rq = select_task_rq_fair, |
| 1460 | #endif /* CONFIG_SMP */ |
| 1461 | |
| 1462 | .check_preempt_curr = check_preempt_wakeup, |
| 1463 | |
| 1464 | .pick_next_task = pick_next_task_fair, |
| 1465 | .put_prev_task = put_prev_task_fair, |
| 1466 | |
| 1467 | #ifdef CONFIG_SMP |
| 1468 | .load_balance = load_balance_fair, |
| 1469 | .move_one_task = move_one_task_fair, |
| 1470 | #endif |
| 1471 | |
| 1472 | .set_curr_task = set_curr_task_fair, |
| 1473 | .task_tick = task_tick_fair, |
| 1474 | .task_new = task_new_fair, |
| 1475 | |
| 1476 | .prio_changed = prio_changed_fair, |
| 1477 | .switched_to = switched_to_fair, |
| 1478 | |
| 1479 | #ifdef CONFIG_FAIR_GROUP_SCHED |
| 1480 | .moved_group = moved_group_fair, |
| 1481 | #endif |
| 1482 | }; |
| 1483 | |
| 1484 | #ifdef CONFIG_SCHED_DEBUG |
| 1485 | static void print_cfs_stats(struct seq_file *m, int cpu) |
| 1486 | { |
| 1487 | struct cfs_rq *cfs_rq; |
| 1488 | |
| 1489 | #ifdef CONFIG_FAIR_GROUP_SCHED |
| 1490 | print_cfs_rq(m, cpu, &cpu_rq(cpu)->cfs); |
| 1491 | #endif |
| 1492 | rcu_read_lock(); |
| 1493 | for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq) |
| 1494 | print_cfs_rq(m, cpu, cfs_rq); |
| 1495 | rcu_read_unlock(); |
| 1496 | } |
| 1497 | #endif |