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#include <linux/sched.h> 25#include <linux/cpumask.h> 26#include <linux/cpuidle.h> 27#include <linux/slab.h> 28#include <linux/profile.h> 29#include <linux/interrupt.h> 30#include <linux/mempolicy.h> 31#include <linux/migrate.h> 32#include <linux/task_work.h> 33 34#include <trace/events/sched.h> 35 36#include "sched.h" 37 38/* 39 * Targeted preemption latency for CPU-bound tasks: 40 * (default: 6ms * (1 + ilog(ncpus)), units: nanoseconds) 41 * 42 * NOTE: this latency value is not the same as the concept of 43 * 'timeslice length' - timeslices in CFS are of variable length 44 * and have no persistent notion like in traditional, time-slice 45 * based scheduling concepts. 46 * 47 * (to see the precise effective timeslice length of your workload, 48 * run vmstat and monitor the context-switches (cs) field) 49 */ 50unsigned int sysctl_sched_latency = 6000000ULL; 51unsigned int normalized_sysctl_sched_latency = 6000000ULL; 52 53/* 54 * The initial- and re-scaling of tunables is configurable 55 * (default SCHED_TUNABLESCALING_LOG = *(1+ilog(ncpus)) 56 * 57 * Options are: 58 * SCHED_TUNABLESCALING_NONE - unscaled, always *1 59 * SCHED_TUNABLESCALING_LOG - scaled logarithmical, *1+ilog(ncpus) 60 * SCHED_TUNABLESCALING_LINEAR - scaled linear, *ncpus 61 */ 62enum sched_tunable_scaling sysctl_sched_tunable_scaling 63 = SCHED_TUNABLESCALING_LOG; 64 65/* 66 * Minimal preemption granularity for CPU-bound tasks: 67 * (default: 0.75 msec * (1 + ilog(ncpus)), units: nanoseconds) 68 */ 69unsigned int sysctl_sched_min_granularity = 750000ULL; 70unsigned int normalized_sysctl_sched_min_granularity = 750000ULL; 71 72/* 73 * is kept at sysctl_sched_latency / sysctl_sched_min_granularity 74 */ 75static unsigned int sched_nr_latency = 8; 76 77/* 78 * After fork, child runs first. If set to 0 (default) then 79 * parent will (try to) run first. 80 */ 81unsigned int sysctl_sched_child_runs_first __read_mostly; 82 83/* 84 * SCHED_OTHER wake-up granularity. 85 * (default: 1 msec * (1 + ilog(ncpus)), units: nanoseconds) 86 * 87 * This option delays the preemption effects of decoupled workloads 88 * and reduces their over-scheduling. Synchronous workloads will still 89 * have immediate wakeup/sleep latencies. 90 */ 91unsigned int sysctl_sched_wakeup_granularity = 1000000UL; 92unsigned int normalized_sysctl_sched_wakeup_granularity = 1000000UL; 93 94const_debug unsigned int sysctl_sched_migration_cost = 500000UL; 95 96/* 97 * The exponential sliding window over which load is averaged for shares 98 * distribution. 99 * (default: 10msec) 100 */ 101unsigned int __read_mostly sysctl_sched_shares_window = 10000000UL; 102 103#ifdef CONFIG_CFS_BANDWIDTH 104/* 105 * Amount of runtime to allocate from global (tg) to local (per-cfs_rq) pool 106 * each time a cfs_rq requests quota. 107 * 108 * Note: in the case that the slice exceeds the runtime remaining (either due 109 * to consumption or the quota being specified to be smaller than the slice) 110 * we will always only issue the remaining available time. 111 * 112 * default: 5 msec, units: microseconds 113 */ 114unsigned int sysctl_sched_cfs_bandwidth_slice = 5000UL; 115#endif 116 117static inline void update_load_add(struct load_weight *lw, unsigned long inc) 118{ 119 lw->weight += inc; 120 lw->inv_weight = 0; 121} 122 123static inline void update_load_sub(struct load_weight *lw, unsigned long dec) 124{ 125 lw->weight -= dec; 126 lw->inv_weight = 0; 127} 128 129static inline void update_load_set(struct load_weight *lw, unsigned long w) 130{ 131 lw->weight = w; 132 lw->inv_weight = 0; 133} 134 135/* 136 * Increase the granularity value when there are more CPUs, 137 * because with more CPUs the 'effective latency' as visible 138 * to users decreases. But the relationship is not linear, 139 * so pick a second-best guess by going with the log2 of the 140 * number of CPUs. 141 * 142 * This idea comes from the SD scheduler of Con Kolivas: 143 */ 144static int get_update_sysctl_factor(void) 145{ 146 unsigned int cpus = min_t(int, num_online_cpus(), 8); 147 unsigned int factor; 148 149 switch (sysctl_sched_tunable_scaling) { 150 case SCHED_TUNABLESCALING_NONE: 151 factor = 1; 152 break; 153 case SCHED_TUNABLESCALING_LINEAR: 154 factor = cpus; 155 break; 156 case SCHED_TUNABLESCALING_LOG: 157 default: 158 factor = 1 + ilog2(cpus); 159 break; 160 } 161 162 return factor; 163} 164 165static void update_sysctl(void) 166{ 167 unsigned int factor = get_update_sysctl_factor(); 168 169#define SET_SYSCTL(name) \ 170 (sysctl_##name = (factor) * normalized_sysctl_##name) 171 SET_SYSCTL(sched_min_granularity); 172 SET_SYSCTL(sched_latency); 173 SET_SYSCTL(sched_wakeup_granularity); 174#undef SET_SYSCTL 175} 176 177void sched_init_granularity(void) 178{ 179 update_sysctl(); 180} 181 182#define WMULT_CONST (~0U) 183#define WMULT_SHIFT 32 184 185static void __update_inv_weight(struct load_weight *lw) 186{ 187 unsigned long w; 188 189 if (likely(lw->inv_weight)) 190 return; 191 192 w = scale_load_down(lw->weight); 193 194 if (BITS_PER_LONG > 32 && unlikely(w >= WMULT_CONST)) 195 lw->inv_weight = 1; 196 else if (unlikely(!w)) 197 lw->inv_weight = WMULT_CONST; 198 else 199 lw->inv_weight = WMULT_CONST / w; 200} 201 202/* 203 * delta_exec * weight / lw.weight 204 * OR 205 * (delta_exec * (weight * lw->inv_weight)) >> WMULT_SHIFT 206 * 207 * Either weight := NICE_0_LOAD and lw \e prio_to_wmult[], in which case 208 * we're guaranteed shift stays positive because inv_weight is guaranteed to 209 * fit 32 bits, and NICE_0_LOAD gives another 10 bits; therefore shift >= 22. 210 * 211 * Or, weight =< lw.weight (because lw.weight is the runqueue weight), thus 212 * weight/lw.weight <= 1, and therefore our shift will also be positive. 213 */ 214static u64 __calc_delta(u64 delta_exec, unsigned long weight, struct load_weight *lw) 215{ 216 u64 fact = scale_load_down(weight); 217 int shift = WMULT_SHIFT; 218 219 __update_inv_weight(lw); 220 221 if (unlikely(fact >> 32)) { 222 while (fact >> 32) { 223 fact >>= 1; 224 shift--; 225 } 226 } 227 228 /* hint to use a 32x32->64 mul */ 229 fact = (u64)(u32)fact * lw->inv_weight; 230 231 while (fact >> 32) { 232 fact >>= 1; 233 shift--; 234 } 235 236 return mul_u64_u32_shr(delta_exec, fact, shift); 237} 238 239 240const struct sched_class fair_sched_class; 241 242/************************************************************** 243 * CFS operations on generic schedulable entities: 244 */ 245 246#ifdef CONFIG_FAIR_GROUP_SCHED 247 248/* cpu runqueue to which this cfs_rq is attached */ 249static inline struct rq *rq_of(struct cfs_rq *cfs_rq) 250{ 251 return cfs_rq->rq; 252} 253 254/* An entity is a task if it doesn't "own" a runqueue */ 255#define entity_is_task(se) (!se->my_q) 256 257static inline struct task_struct *task_of(struct sched_entity *se) 258{ 259#ifdef CONFIG_SCHED_DEBUG 260 WARN_ON_ONCE(!entity_is_task(se)); 261#endif 262 return container_of(se, struct task_struct, se); 263} 264 265/* Walk up scheduling entities hierarchy */ 266#define for_each_sched_entity(se) \ 267 for (; se; se = se->parent) 268 269static inline struct cfs_rq *task_cfs_rq(struct task_struct *p) 270{ 271 return p->se.cfs_rq; 272} 273 274/* runqueue on which this entity is (to be) queued */ 275static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se) 276{ 277 return se->cfs_rq; 278} 279 280/* runqueue "owned" by this group */ 281static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp) 282{ 283 return grp->my_q; 284} 285 286static void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq, 287 int force_update); 288 289static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq) 290{ 291 if (!cfs_rq->on_list) { 292 /* 293 * Ensure we either appear before our parent (if already 294 * enqueued) or force our parent to appear after us when it is 295 * enqueued. The fact that we always enqueue bottom-up 296 * reduces this to two cases. 297 */ 298 if (cfs_rq->tg->parent && 299 cfs_rq->tg->parent->cfs_rq[cpu_of(rq_of(cfs_rq))]->on_list) { 300 list_add_rcu(&cfs_rq->leaf_cfs_rq_list, 301 &rq_of(cfs_rq)->leaf_cfs_rq_list); 302 } else { 303 list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list, 304 &rq_of(cfs_rq)->leaf_cfs_rq_list); 305 } 306 307 cfs_rq->on_list = 1; 308 /* We should have no load, but we need to update last_decay. */ 309 update_cfs_rq_blocked_load(cfs_rq, 0); 310 } 311} 312 313static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq) 314{ 315 if (cfs_rq->on_list) { 316 list_del_rcu(&cfs_rq->leaf_cfs_rq_list); 317 cfs_rq->on_list = 0; 318 } 319} 320 321/* Iterate thr' all leaf cfs_rq's on a runqueue */ 322#define for_each_leaf_cfs_rq(rq, cfs_rq) \ 323 list_for_each_entry_rcu(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list) 324 325/* Do the two (enqueued) entities belong to the same group ? */ 326static inline struct cfs_rq * 327is_same_group(struct sched_entity *se, struct sched_entity *pse) 328{ 329 if (se->cfs_rq == pse->cfs_rq) 330 return se->cfs_rq; 331 332 return NULL; 333} 334 335static inline struct sched_entity *parent_entity(struct sched_entity *se) 336{ 337 return se->parent; 338} 339 340static void 341find_matching_se(struct sched_entity **se, struct sched_entity **pse) 342{ 343 int se_depth, pse_depth; 344 345 /* 346 * preemption test can be made between sibling entities who are in the 347 * same cfs_rq i.e who have a common parent. Walk up the hierarchy of 348 * both tasks until we find their ancestors who are siblings of common 349 * parent. 350 */ 351 352 /* First walk up until both entities are at same depth */ 353 se_depth = (*se)->depth; 354 pse_depth = (*pse)->depth; 355 356 while (se_depth > pse_depth) { 357 se_depth--; 358 *se = parent_entity(*se); 359 } 360 361 while (pse_depth > se_depth) { 362 pse_depth--; 363 *pse = parent_entity(*pse); 364 } 365 366 while (!is_same_group(*se, *pse)) { 367 *se = parent_entity(*se); 368 *pse = parent_entity(*pse); 369 } 370} 371 372#else /* !CONFIG_FAIR_GROUP_SCHED */ 373 374static inline struct task_struct *task_of(struct sched_entity *se) 375{ 376 return container_of(se, struct task_struct, se); 377} 378 379static inline struct rq *rq_of(struct cfs_rq *cfs_rq) 380{ 381 return container_of(cfs_rq, struct rq, cfs); 382} 383 384#define entity_is_task(se) 1 385 386#define for_each_sched_entity(se) \ 387 for (; se; se = NULL) 388 389static inline struct cfs_rq *task_cfs_rq(struct task_struct *p) 390{ 391 return &task_rq(p)->cfs; 392} 393 394static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se) 395{ 396 struct task_struct *p = task_of(se); 397 struct rq *rq = task_rq(p); 398 399 return &rq->cfs; 400} 401 402/* runqueue "owned" by this group */ 403static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp) 404{ 405 return NULL; 406} 407 408static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq) 409{ 410} 411 412static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq) 413{ 414} 415 416#define for_each_leaf_cfs_rq(rq, cfs_rq) \ 417 for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL) 418 419static inline struct sched_entity *parent_entity(struct sched_entity *se) 420{ 421 return NULL; 422} 423 424static inline void 425find_matching_se(struct sched_entity **se, struct sched_entity **pse) 426{ 427} 428 429#endif /* CONFIG_FAIR_GROUP_SCHED */ 430 431static __always_inline 432void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec); 433 434/************************************************************** 435 * Scheduling class tree data structure manipulation methods: 436 */ 437 438static inline u64 max_vruntime(u64 max_vruntime, u64 vruntime) 439{ 440 s64 delta = (s64)(vruntime - max_vruntime); 441 if (delta > 0) 442 max_vruntime = vruntime; 443 444 return max_vruntime; 445} 446 447static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime) 448{ 449 s64 delta = (s64)(vruntime - min_vruntime); 450 if (delta < 0) 451 min_vruntime = vruntime; 452 453 return min_vruntime; 454} 455 456static inline int entity_before(struct sched_entity *a, 457 struct sched_entity *b) 458{ 459 return (s64)(a->vruntime - b->vruntime) < 0; 460} 461 462static void update_min_vruntime(struct cfs_rq *cfs_rq) 463{ 464 u64 vruntime = cfs_rq->min_vruntime; 465 466 if (cfs_rq->curr) 467 vruntime = cfs_rq->curr->vruntime; 468 469 if (cfs_rq->rb_leftmost) { 470 struct sched_entity *se = rb_entry(cfs_rq->rb_leftmost, 471 struct sched_entity, 472 run_node); 473 474 if (!cfs_rq->curr) 475 vruntime = se->vruntime; 476 else 477 vruntime = min_vruntime(vruntime, se->vruntime); 478 } 479 480 /* ensure we never gain time by being placed backwards. */ 481 cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime); 482#ifndef CONFIG_64BIT 483 smp_wmb(); 484 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime; 485#endif 486} 487 488/* 489 * Enqueue an entity into the rb-tree: 490 */ 491static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se) 492{ 493 struct rb_node **link = &cfs_rq->tasks_timeline.rb_node; 494 struct rb_node *parent = NULL; 495 struct sched_entity *entry; 496 int leftmost = 1; 497 498 /* 499 * Find the right place in the rbtree: 500 */ 501 while (*link) { 502 parent = *link; 503 entry = rb_entry(parent, struct sched_entity, run_node); 504 /* 505 * We dont care about collisions. Nodes with 506 * the same key stay together. 507 */ 508 if (entity_before(se, entry)) { 509 link = &parent->rb_left; 510 } else { 511 link = &parent->rb_right; 512 leftmost = 0; 513 } 514 } 515 516 /* 517 * Maintain a cache of leftmost tree entries (it is frequently 518 * used): 519 */ 520 if (leftmost) 521 cfs_rq->rb_leftmost = &se->run_node; 522 523 rb_link_node(&se->run_node, parent, link); 524 rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline); 525} 526 527static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se) 528{ 529 if (cfs_rq->rb_leftmost == &se->run_node) { 530 struct rb_node *next_node; 531 532 next_node = rb_next(&se->run_node); 533 cfs_rq->rb_leftmost = next_node; 534 } 535 536 rb_erase(&se->run_node, &cfs_rq->tasks_timeline); 537} 538 539struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq) 540{ 541 struct rb_node *left = cfs_rq->rb_leftmost; 542 543 if (!left) 544 return NULL; 545 546 return rb_entry(left, struct sched_entity, run_node); 547} 548 549static struct sched_entity *__pick_next_entity(struct sched_entity *se) 550{ 551 struct rb_node *next = rb_next(&se->run_node); 552 553 if (!next) 554 return NULL; 555 556 return rb_entry(next, struct sched_entity, run_node); 557} 558 559#ifdef CONFIG_SCHED_DEBUG 560struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq) 561{ 562 struct rb_node *last = rb_last(&cfs_rq->tasks_timeline); 563 564 if (!last) 565 return NULL; 566 567 return rb_entry(last, struct sched_entity, run_node); 568} 569 570/************************************************************** 571 * Scheduling class statistics methods: 572 */ 573 574int sched_proc_update_handler(struct ctl_table *table, int write, 575 void __user *buffer, size_t *lenp, 576 loff_t *ppos) 577{ 578 int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos); 579 int factor = get_update_sysctl_factor(); 580 581 if (ret || !write) 582 return ret; 583 584 sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency, 585 sysctl_sched_min_granularity); 586 587#define WRT_SYSCTL(name) \ 588 (normalized_sysctl_##name = sysctl_##name / (factor)) 589 WRT_SYSCTL(sched_min_granularity); 590 WRT_SYSCTL(sched_latency); 591 WRT_SYSCTL(sched_wakeup_granularity); 592#undef WRT_SYSCTL 593 594 return 0; 595} 596#endif 597 598/* 599 * delta /= w 600 */ 601static inline u64 calc_delta_fair(u64 delta, struct sched_entity *se) 602{ 603 if (unlikely(se->load.weight != NICE_0_LOAD)) 604 delta = __calc_delta(delta, NICE_0_LOAD, &se->load); 605 606 return delta; 607} 608 609/* 610 * The idea is to set a period in which each task runs once. 611 * 612 * When there are too many tasks (sched_nr_latency) we have to stretch 613 * this period because otherwise the slices get too small. 614 * 615 * p = (nr <= nl) ? l : l*nr/nl 616 */ 617static u64 __sched_period(unsigned long nr_running) 618{ 619 u64 period = sysctl_sched_latency; 620 unsigned long nr_latency = sched_nr_latency; 621 622 if (unlikely(nr_running > nr_latency)) { 623 period = sysctl_sched_min_granularity; 624 period *= nr_running; 625 } 626 627 return period; 628} 629 630/* 631 * We calculate the wall-time slice from the period by taking a part 632 * proportional to the weight. 633 * 634 * s = p*P[w/rw] 635 */ 636static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se) 637{ 638 u64 slice = __sched_period(cfs_rq->nr_running + !se->on_rq); 639 640 for_each_sched_entity(se) { 641 struct load_weight *load; 642 struct load_weight lw; 643 644 cfs_rq = cfs_rq_of(se); 645 load = &cfs_rq->load; 646 647 if (unlikely(!se->on_rq)) { 648 lw = cfs_rq->load; 649 650 update_load_add(&lw, se->load.weight); 651 load = &lw; 652 } 653 slice = __calc_delta(slice, se->load.weight, load); 654 } 655 return slice; 656} 657 658/* 659 * We calculate the vruntime slice of a to-be-inserted task. 660 * 661 * vs = s/w 662 */ 663static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se) 664{ 665 return calc_delta_fair(sched_slice(cfs_rq, se), se); 666} 667 668#ifdef CONFIG_SMP 669static int select_idle_sibling(struct task_struct *p, int cpu); 670static unsigned long task_h_load(struct task_struct *p); 671 672static inline void __update_task_entity_contrib(struct sched_entity *se); 673 674/* Give new task start runnable values to heavy its load in infant time */ 675void init_task_runnable_average(struct task_struct *p) 676{ 677 u32 slice; 678 679 p->se.avg.decay_count = 0; 680 slice = sched_slice(task_cfs_rq(p), &p->se) >> 10; 681 p->se.avg.runnable_avg_sum = slice; 682 p->se.avg.runnable_avg_period = slice; 683 __update_task_entity_contrib(&p->se); 684} 685#else 686void init_task_runnable_average(struct task_struct *p) 687{ 688} 689#endif 690 691/* 692 * Update the current task's runtime statistics. 693 */ 694static void update_curr(struct cfs_rq *cfs_rq) 695{ 696 struct sched_entity *curr = cfs_rq->curr; 697 u64 now = rq_clock_task(rq_of(cfs_rq)); 698 u64 delta_exec; 699 700 if (unlikely(!curr)) 701 return; 702 703 delta_exec = now - curr->exec_start; 704 if (unlikely((s64)delta_exec <= 0)) 705 return; 706 707 curr->exec_start = now; 708 709 schedstat_set(curr->statistics.exec_max, 710 max(delta_exec, curr->statistics.exec_max)); 711 712 curr->sum_exec_runtime += delta_exec; 713 schedstat_add(cfs_rq, exec_clock, delta_exec); 714 715 curr->vruntime += calc_delta_fair(delta_exec, curr); 716 update_min_vruntime(cfs_rq); 717 718 if (entity_is_task(curr)) { 719 struct task_struct *curtask = task_of(curr); 720 721 trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime); 722 cpuacct_charge(curtask, delta_exec); 723 account_group_exec_runtime(curtask, delta_exec); 724 } 725 726 account_cfs_rq_runtime(cfs_rq, delta_exec); 727} 728 729static void update_curr_fair(struct rq *rq) 730{ 731 update_curr(cfs_rq_of(&rq->curr->se)); 732} 733 734static inline void 735update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se) 736{ 737 schedstat_set(se->statistics.wait_start, rq_clock(rq_of(cfs_rq))); 738} 739 740/* 741 * Task is being enqueued - update stats: 742 */ 743static void update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se) 744{ 745 /* 746 * Are we enqueueing a waiting task? (for current tasks 747 * a dequeue/enqueue event is a NOP) 748 */ 749 if (se != cfs_rq->curr) 750 update_stats_wait_start(cfs_rq, se); 751} 752 753static void 754update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se) 755{ 756 schedstat_set(se->statistics.wait_max, max(se->statistics.wait_max, 757 rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start)); 758 schedstat_set(se->statistics.wait_count, se->statistics.wait_count + 1); 759 schedstat_set(se->statistics.wait_sum, se->statistics.wait_sum + 760 rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start); 761#ifdef CONFIG_SCHEDSTATS 762 if (entity_is_task(se)) { 763 trace_sched_stat_wait(task_of(se), 764 rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start); 765 } 766#endif 767 schedstat_set(se->statistics.wait_start, 0); 768} 769 770static inline void 771update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se) 772{ 773 /* 774 * Mark the end of the wait period if dequeueing a 775 * waiting task: 776 */ 777 if (se != cfs_rq->curr) 778 update_stats_wait_end(cfs_rq, se); 779} 780 781/* 782 * We are picking a new current task - update its stats: 783 */ 784static inline void 785update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se) 786{ 787 /* 788 * We are starting a new run period: 789 */ 790 se->exec_start = rq_clock_task(rq_of(cfs_rq)); 791} 792 793/************************************************** 794 * Scheduling class queueing methods: 795 */ 796 797#ifdef CONFIG_NUMA_BALANCING 798/* 799 * Approximate time to scan a full NUMA task in ms. The task scan period is 800 * calculated based on the tasks virtual memory size and 801 * numa_balancing_scan_size. 802 */ 803unsigned int sysctl_numa_balancing_scan_period_min = 1000; 804unsigned int sysctl_numa_balancing_scan_period_max = 60000; 805 806/* Portion of address space to scan in MB */ 807unsigned int sysctl_numa_balancing_scan_size = 256; 808 809/* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */ 810unsigned int sysctl_numa_balancing_scan_delay = 1000; 811 812static unsigned int task_nr_scan_windows(struct task_struct *p) 813{ 814 unsigned long rss = 0; 815 unsigned long nr_scan_pages; 816 817 /* 818 * Calculations based on RSS as non-present and empty pages are skipped 819 * by the PTE scanner and NUMA hinting faults should be trapped based 820 * on resident pages 821 */ 822 nr_scan_pages = sysctl_numa_balancing_scan_size << (20 - PAGE_SHIFT); 823 rss = get_mm_rss(p->mm); 824 if (!rss) 825 rss = nr_scan_pages; 826 827 rss = round_up(rss, nr_scan_pages); 828 return rss / nr_scan_pages; 829} 830 831/* For sanitys sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */ 832#define MAX_SCAN_WINDOW 2560 833 834static unsigned int task_scan_min(struct task_struct *p) 835{ 836 unsigned int scan_size = ACCESS_ONCE(sysctl_numa_balancing_scan_size); 837 unsigned int scan, floor; 838 unsigned int windows = 1; 839 840 if (scan_size < MAX_SCAN_WINDOW) 841 windows = MAX_SCAN_WINDOW / scan_size; 842 floor = 1000 / windows; 843 844 scan = sysctl_numa_balancing_scan_period_min / task_nr_scan_windows(p); 845 return max_t(unsigned int, floor, scan); 846} 847 848static unsigned int task_scan_max(struct task_struct *p) 849{ 850 unsigned int smin = task_scan_min(p); 851 unsigned int smax; 852 853 /* Watch for min being lower than max due to floor calculations */ 854 smax = sysctl_numa_balancing_scan_period_max / task_nr_scan_windows(p); 855 return max(smin, smax); 856} 857 858static void account_numa_enqueue(struct rq *rq, struct task_struct *p) 859{ 860 rq->nr_numa_running += (p->numa_preferred_nid != -1); 861 rq->nr_preferred_running += (p->numa_preferred_nid == task_node(p)); 862} 863 864static void account_numa_dequeue(struct rq *rq, struct task_struct *p) 865{ 866 rq->nr_numa_running -= (p->numa_preferred_nid != -1); 867 rq->nr_preferred_running -= (p->numa_preferred_nid == task_node(p)); 868} 869 870struct numa_group { 871 atomic_t refcount; 872 873 spinlock_t lock; /* nr_tasks, tasks */ 874 int nr_tasks; 875 pid_t gid; 876 struct list_head task_list; 877 878 struct rcu_head rcu; 879 nodemask_t active_nodes; 880 unsigned long total_faults; 881 /* 882 * Faults_cpu is used to decide whether memory should move 883 * towards the CPU. As a consequence, these stats are weighted 884 * more by CPU use than by memory faults. 885 */ 886 unsigned long *faults_cpu; 887 unsigned long faults[0]; 888}; 889 890/* Shared or private faults. */ 891#define NR_NUMA_HINT_FAULT_TYPES 2 892 893/* Memory and CPU locality */ 894#define NR_NUMA_HINT_FAULT_STATS (NR_NUMA_HINT_FAULT_TYPES * 2) 895 896/* Averaged statistics, and temporary buffers. */ 897#define NR_NUMA_HINT_FAULT_BUCKETS (NR_NUMA_HINT_FAULT_STATS * 2) 898 899pid_t task_numa_group_id(struct task_struct *p) 900{ 901 return p->numa_group ? p->numa_group->gid : 0; 902} 903 904static inline int task_faults_idx(int nid, int priv) 905{ 906 return NR_NUMA_HINT_FAULT_TYPES * nid + priv; 907} 908 909static inline unsigned long task_faults(struct task_struct *p, int nid) 910{ 911 if (!p->numa_faults_memory) 912 return 0; 913 914 return p->numa_faults_memory[task_faults_idx(nid, 0)] + 915 p->numa_faults_memory[task_faults_idx(nid, 1)]; 916} 917 918static inline unsigned long group_faults(struct task_struct *p, int nid) 919{ 920 if (!p->numa_group) 921 return 0; 922 923 return p->numa_group->faults[task_faults_idx(nid, 0)] + 924 p->numa_group->faults[task_faults_idx(nid, 1)]; 925} 926 927static inline unsigned long group_faults_cpu(struct numa_group *group, int nid) 928{ 929 return group->faults_cpu[task_faults_idx(nid, 0)] + 930 group->faults_cpu[task_faults_idx(nid, 1)]; 931} 932 933/* 934 * These return the fraction of accesses done by a particular task, or 935 * task group, on a particular numa node. The group weight is given a 936 * larger multiplier, in order to group tasks together that are almost 937 * evenly spread out between numa nodes. 938 */ 939static inline unsigned long task_weight(struct task_struct *p, int nid) 940{ 941 unsigned long total_faults; 942 943 if (!p->numa_faults_memory) 944 return 0; 945 946 total_faults = p->total_numa_faults; 947 948 if (!total_faults) 949 return 0; 950 951 return 1000 * task_faults(p, nid) / total_faults; 952} 953 954static inline unsigned long group_weight(struct task_struct *p, int nid) 955{ 956 if (!p->numa_group || !p->numa_group->total_faults) 957 return 0; 958 959 return 1000 * group_faults(p, nid) / p->numa_group->total_faults; 960} 961 962bool should_numa_migrate_memory(struct task_struct *p, struct page * page, 963 int src_nid, int dst_cpu) 964{ 965 struct numa_group *ng = p->numa_group; 966 int dst_nid = cpu_to_node(dst_cpu); 967 int last_cpupid, this_cpupid; 968 969 this_cpupid = cpu_pid_to_cpupid(dst_cpu, current->pid); 970 971 /* 972 * Multi-stage node selection is used in conjunction with a periodic 973 * migration fault to build a temporal task<->page relation. By using 974 * a two-stage filter we remove short/unlikely relations. 975 * 976 * Using P(p) ~ n_p / n_t as per frequentist probability, we can equate 977 * a task's usage of a particular page (n_p) per total usage of this 978 * page (n_t) (in a given time-span) to a probability. 979 * 980 * Our periodic faults will sample this probability and getting the 981 * same result twice in a row, given these samples are fully 982 * independent, is then given by P(n)^2, provided our sample period 983 * is sufficiently short compared to the usage pattern. 984 * 985 * This quadric squishes small probabilities, making it less likely we 986 * act on an unlikely task<->page relation. 987 */ 988 last_cpupid = page_cpupid_xchg_last(page, this_cpupid); 989 if (!cpupid_pid_unset(last_cpupid) && 990 cpupid_to_nid(last_cpupid) != dst_nid) 991 return false; 992 993 /* Always allow migrate on private faults */ 994 if (cpupid_match_pid(p, last_cpupid)) 995 return true; 996 997 /* A shared fault, but p->numa_group has not been set up yet. */ 998 if (!ng) 999 return true; 1000 1001 /* 1002 * Do not migrate if the destination is not a node that 1003 * is actively used by this numa group. 1004 */ 1005 if (!node_isset(dst_nid, ng->active_nodes)) 1006 return false; 1007 1008 /* 1009 * Source is a node that is not actively used by this 1010 * numa group, while the destination is. Migrate. 1011 */ 1012 if (!node_isset(src_nid, ng->active_nodes)) 1013 return true; 1014 1015 /* 1016 * Both source and destination are nodes in active 1017 * use by this numa group. Maximize memory bandwidth 1018 * by migrating from more heavily used groups, to less 1019 * heavily used ones, spreading the load around. 1020 * Use a 1/4 hysteresis to avoid spurious page movement. 1021 */ 1022 return group_faults(p, dst_nid) < (group_faults(p, src_nid) * 3 / 4); 1023} 1024 1025static unsigned long weighted_cpuload(const int cpu); 1026static unsigned long source_load(int cpu, int type); 1027static unsigned long target_load(int cpu, int type); 1028static unsigned long capacity_of(int cpu); 1029static long effective_load(struct task_group *tg, int cpu, long wl, long wg); 1030 1031/* Cached statistics for all CPUs within a node */ 1032struct numa_stats { 1033 unsigned long nr_running; 1034 unsigned long load; 1035 1036 /* Total compute capacity of CPUs on a node */ 1037 unsigned long compute_capacity; 1038 1039 /* Approximate capacity in terms of runnable tasks on a node */ 1040 unsigned long task_capacity; 1041 int has_free_capacity; 1042}; 1043 1044/* 1045 * XXX borrowed from update_sg_lb_stats 1046 */ 1047static void update_numa_stats(struct numa_stats *ns, int nid) 1048{ 1049 int smt, cpu, cpus = 0; 1050 unsigned long capacity; 1051 1052 memset(ns, 0, sizeof(*ns)); 1053 for_each_cpu(cpu, cpumask_of_node(nid)) { 1054 struct rq *rq = cpu_rq(cpu); 1055 1056 ns->nr_running += rq->nr_running; 1057 ns->load += weighted_cpuload(cpu); 1058 ns->compute_capacity += capacity_of(cpu); 1059 1060 cpus++; 1061 } 1062 1063 /* 1064 * If we raced with hotplug and there are no CPUs left in our mask 1065 * the @ns structure is NULL'ed and task_numa_compare() will 1066 * not find this node attractive. 1067 * 1068 * We'll either bail at !has_free_capacity, or we'll detect a huge 1069 * imbalance and bail there. 1070 */ 1071 if (!cpus) 1072 return; 1073 1074 /* smt := ceil(cpus / capacity), assumes: 1 < smt_power < 2 */ 1075 smt = DIV_ROUND_UP(SCHED_CAPACITY_SCALE * cpus, ns->compute_capacity); 1076 capacity = cpus / smt; /* cores */ 1077 1078 ns->task_capacity = min_t(unsigned, capacity, 1079 DIV_ROUND_CLOSEST(ns->compute_capacity, SCHED_CAPACITY_SCALE)); 1080 ns->has_free_capacity = (ns->nr_running < ns->task_capacity); 1081} 1082 1083struct task_numa_env { 1084 struct task_struct *p; 1085 1086 int src_cpu, src_nid; 1087 int dst_cpu, dst_nid; 1088 1089 struct numa_stats src_stats, dst_stats; 1090 1091 int imbalance_pct; 1092 1093 struct task_struct *best_task; 1094 long best_imp; 1095 int best_cpu; 1096}; 1097 1098static void task_numa_assign(struct task_numa_env *env, 1099 struct task_struct *p, long imp) 1100{ 1101 if (env->best_task) 1102 put_task_struct(env->best_task); 1103 if (p) 1104 get_task_struct(p); 1105 1106 env->best_task = p; 1107 env->best_imp = imp; 1108 env->best_cpu = env->dst_cpu; 1109} 1110 1111static bool load_too_imbalanced(long src_load, long dst_load, 1112 struct task_numa_env *env) 1113{ 1114 long imb, old_imb; 1115 long orig_src_load, orig_dst_load; 1116 long src_capacity, dst_capacity; 1117 1118 /* 1119 * The load is corrected for the CPU capacity available on each node. 1120 * 1121 * src_load dst_load 1122 * ------------ vs --------- 1123 * src_capacity dst_capacity 1124 */ 1125 src_capacity = env->src_stats.compute_capacity; 1126 dst_capacity = env->dst_stats.compute_capacity; 1127 1128 /* We care about the slope of the imbalance, not the direction. */ 1129 if (dst_load < src_load) 1130 swap(dst_load, src_load); 1131 1132 /* Is the difference below the threshold? */ 1133 imb = dst_load * src_capacity * 100 - 1134 src_load * dst_capacity * env->imbalance_pct; 1135 if (imb <= 0) 1136 return false; 1137 1138 /* 1139 * The imbalance is above the allowed threshold. 1140 * Compare it with the old imbalance. 1141 */ 1142 orig_src_load = env->src_stats.load; 1143 orig_dst_load = env->dst_stats.load; 1144 1145 if (orig_dst_load < orig_src_load) 1146 swap(orig_dst_load, orig_src_load); 1147 1148 old_imb = orig_dst_load * src_capacity * 100 - 1149 orig_src_load * dst_capacity * env->imbalance_pct; 1150 1151 /* Would this change make things worse? */ 1152 return (imb > old_imb); 1153} 1154 1155/* 1156 * This checks if the overall compute and NUMA accesses of the system would 1157 * be improved if the source tasks was migrated to the target dst_cpu taking 1158 * into account that it might be best if task running on the dst_cpu should 1159 * be exchanged with the source task 1160 */ 1161static void task_numa_compare(struct task_numa_env *env, 1162 long taskimp, long groupimp) 1163{ 1164 struct rq *src_rq = cpu_rq(env->src_cpu); 1165 struct rq *dst_rq = cpu_rq(env->dst_cpu); 1166 struct task_struct *cur; 1167 long src_load, dst_load; 1168 long load; 1169 long imp = env->p->numa_group ? groupimp : taskimp; 1170 long moveimp = imp; 1171 1172 rcu_read_lock(); 1173 1174 raw_spin_lock_irq(&dst_rq->lock); 1175 cur = dst_rq->curr; 1176 /* 1177 * No need to move the exiting task, and this ensures that ->curr 1178 * wasn't reaped and thus get_task_struct() in task_numa_assign() 1179 * is safe under RCU read lock. 1180 * Note that rcu_read_lock() itself can't protect from the final 1181 * put_task_struct() after the last schedule(). 1182 */ 1183 if ((cur->flags & PF_EXITING) || is_idle_task(cur)) 1184 cur = NULL; 1185 raw_spin_unlock_irq(&dst_rq->lock); 1186 1187 /* 1188 * Because we have preemption enabled we can get migrated around and 1189 * end try selecting ourselves (current == env->p) as a swap candidate. 1190 */ 1191 if (cur == env->p) 1192 goto unlock; 1193 1194 /* 1195 * "imp" is the fault differential for the source task between the 1196 * source and destination node. Calculate the total differential for 1197 * the source task and potential destination task. The more negative 1198 * the value is, the more rmeote accesses that would be expected to 1199 * be incurred if the tasks were swapped. 1200 */ 1201 if (cur) { 1202 /* Skip this swap candidate if cannot move to the source cpu */ 1203 if (!cpumask_test_cpu(env->src_cpu, tsk_cpus_allowed(cur))) 1204 goto unlock; 1205 1206 /* 1207 * If dst and source tasks are in the same NUMA group, or not 1208 * in any group then look only at task weights. 1209 */ 1210 if (cur->numa_group == env->p->numa_group) { 1211 imp = taskimp + task_weight(cur, env->src_nid) - 1212 task_weight(cur, env->dst_nid); 1213 /* 1214 * Add some hysteresis to prevent swapping the 1215 * tasks within a group over tiny differences. 1216 */ 1217 if (cur->numa_group) 1218 imp -= imp/16; 1219 } else { 1220 /* 1221 * Compare the group weights. If a task is all by 1222 * itself (not part of a group), use the task weight 1223 * instead. 1224 */ 1225 if (cur->numa_group) 1226 imp += group_weight(cur, env->src_nid) - 1227 group_weight(cur, env->dst_nid); 1228 else 1229 imp += task_weight(cur, env->src_nid) - 1230 task_weight(cur, env->dst_nid); 1231 } 1232 } 1233 1234 if (imp <= env->best_imp && moveimp <= env->best_imp) 1235 goto unlock; 1236 1237 if (!cur) { 1238 /* Is there capacity at our destination? */ 1239 if (env->src_stats.nr_running <= env->src_stats.task_capacity && 1240 !env->dst_stats.has_free_capacity) 1241 goto unlock; 1242 1243 goto balance; 1244 } 1245 1246 /* Balance doesn't matter much if we're running a task per cpu */ 1247 if (imp > env->best_imp && src_rq->nr_running == 1 && 1248 dst_rq->nr_running == 1) 1249 goto assign; 1250 1251 /* 1252 * In the overloaded case, try and keep the load balanced. 1253 */ 1254balance: 1255 load = task_h_load(env->p); 1256 dst_load = env->dst_stats.load + load; 1257 src_load = env->src_stats.load - load; 1258 1259 if (moveimp > imp && moveimp > env->best_imp) { 1260 /* 1261 * If the improvement from just moving env->p direction is 1262 * better than swapping tasks around, check if a move is 1263 * possible. Store a slightly smaller score than moveimp, 1264 * so an actually idle CPU will win. 1265 */ 1266 if (!load_too_imbalanced(src_load, dst_load, env)) { 1267 imp = moveimp - 1; 1268 cur = NULL; 1269 goto assign; 1270 } 1271 } 1272 1273 if (imp <= env->best_imp) 1274 goto unlock; 1275 1276 if (cur) { 1277 load = task_h_load(cur); 1278 dst_load -= load; 1279 src_load += load; 1280 } 1281 1282 if (load_too_imbalanced(src_load, dst_load, env)) 1283 goto unlock; 1284 1285 /* 1286 * One idle CPU per node is evaluated for a task numa move. 1287 * Call select_idle_sibling to maybe find a better one. 1288 */ 1289 if (!cur) 1290 env->dst_cpu = select_idle_sibling(env->p, env->dst_cpu); 1291 1292assign: 1293 task_numa_assign(env, cur, imp); 1294unlock: 1295 rcu_read_unlock(); 1296} 1297 1298static void task_numa_find_cpu(struct task_numa_env *env, 1299 long taskimp, long groupimp) 1300{ 1301 int cpu; 1302 1303 for_each_cpu(cpu, cpumask_of_node(env->dst_nid)) { 1304 /* Skip this CPU if the source task cannot migrate */ 1305 if (!cpumask_test_cpu(cpu, tsk_cpus_allowed(env->p))) 1306 continue; 1307 1308 env->dst_cpu = cpu; 1309 task_numa_compare(env, taskimp, groupimp); 1310 } 1311} 1312 1313static int task_numa_migrate(struct task_struct *p) 1314{ 1315 struct task_numa_env env = { 1316 .p = p, 1317 1318 .src_cpu = task_cpu(p), 1319 .src_nid = task_node(p), 1320 1321 .imbalance_pct = 112, 1322 1323 .best_task = NULL, 1324 .best_imp = 0, 1325 .best_cpu = -1 1326 }; 1327 struct sched_domain *sd; 1328 unsigned long taskweight, groupweight; 1329 int nid, ret; 1330 long taskimp, groupimp; 1331 1332 /* 1333 * Pick the lowest SD_NUMA domain, as that would have the smallest 1334 * imbalance and would be the first to start moving tasks about. 1335 * 1336 * And we want to avoid any moving of tasks about, as that would create 1337 * random movement of tasks -- counter the numa conditions we're trying 1338 * to satisfy here. 1339 */ 1340 rcu_read_lock(); 1341 sd = rcu_dereference(per_cpu(sd_numa, env.src_cpu)); 1342 if (sd) 1343 env.imbalance_pct = 100 + (sd->imbalance_pct - 100) / 2; 1344 rcu_read_unlock(); 1345 1346 /* 1347 * Cpusets can break the scheduler domain tree into smaller 1348 * balance domains, some of which do not cross NUMA boundaries. 1349 * Tasks that are "trapped" in such domains cannot be migrated 1350 * elsewhere, so there is no point in (re)trying. 1351 */ 1352 if (unlikely(!sd)) { 1353 p->numa_preferred_nid = task_node(p); 1354 return -EINVAL; 1355 } 1356 1357 taskweight = task_weight(p, env.src_nid); 1358 groupweight = group_weight(p, env.src_nid); 1359 update_numa_stats(&env.src_stats, env.src_nid); 1360 env.dst_nid = p->numa_preferred_nid; 1361 taskimp = task_weight(p, env.dst_nid) - taskweight; 1362 groupimp = group_weight(p, env.dst_nid) - groupweight; 1363 update_numa_stats(&env.dst_stats, env.dst_nid); 1364 1365 /* Try to find a spot on the preferred nid. */ 1366 task_numa_find_cpu(&env, taskimp, groupimp); 1367 1368 /* No space available on the preferred nid. Look elsewhere. */ 1369 if (env.best_cpu == -1) { 1370 for_each_online_node(nid) { 1371 if (nid == env.src_nid || nid == p->numa_preferred_nid) 1372 continue; 1373 1374 /* Only consider nodes where both task and groups benefit */ 1375 taskimp = task_weight(p, nid) - taskweight; 1376 groupimp = group_weight(p, nid) - groupweight; 1377 if (taskimp < 0 && groupimp < 0) 1378 continue; 1379 1380 env.dst_nid = nid; 1381 update_numa_stats(&env.dst_stats, env.dst_nid); 1382 task_numa_find_cpu(&env, taskimp, groupimp); 1383 } 1384 } 1385 1386 /* 1387 * If the task is part of a workload that spans multiple NUMA nodes, 1388 * and is migrating into one of the workload's active nodes, remember 1389 * this node as the task's preferred numa node, so the workload can 1390 * settle down. 1391 * A task that migrated to a second choice node will be better off 1392 * trying for a better one later. Do not set the preferred node here. 1393 */ 1394 if (p->numa_group) { 1395 if (env.best_cpu == -1) 1396 nid = env.src_nid; 1397 else 1398 nid = env.dst_nid; 1399 1400 if (node_isset(nid, p->numa_group->active_nodes)) 1401 sched_setnuma(p, env.dst_nid); 1402 } 1403 1404 /* No better CPU than the current one was found. */ 1405 if (env.best_cpu == -1) 1406 return -EAGAIN; 1407 1408 /* 1409 * Reset the scan period if the task is being rescheduled on an 1410 * alternative node to recheck if the tasks is now properly placed. 1411 */ 1412 p->numa_scan_period = task_scan_min(p); 1413 1414 if (env.best_task == NULL) { 1415 ret = migrate_task_to(p, env.best_cpu); 1416 if (ret != 0) 1417 trace_sched_stick_numa(p, env.src_cpu, env.best_cpu); 1418 return ret; 1419 } 1420 1421 ret = migrate_swap(p, env.best_task); 1422 if (ret != 0) 1423 trace_sched_stick_numa(p, env.src_cpu, task_cpu(env.best_task)); 1424 put_task_struct(env.best_task); 1425 return ret; 1426} 1427 1428/* Attempt to migrate a task to a CPU on the preferred node. */ 1429static void numa_migrate_preferred(struct task_struct *p) 1430{ 1431 unsigned long interval = HZ; 1432 1433 /* This task has no NUMA fault statistics yet */ 1434 if (unlikely(p->numa_preferred_nid == -1 || !p->numa_faults_memory)) 1435 return; 1436 1437 /* Periodically retry migrating the task to the preferred node */ 1438 interval = min(interval, msecs_to_jiffies(p->numa_scan_period) / 16); 1439 p->numa_migrate_retry = jiffies + interval; 1440 1441 /* Success if task is already running on preferred CPU */ 1442 if (task_node(p) == p->numa_preferred_nid) 1443 return; 1444 1445 /* Otherwise, try migrate to a CPU on the preferred node */ 1446 task_numa_migrate(p); 1447} 1448 1449/* 1450 * Find the nodes on which the workload is actively running. We do this by 1451 * tracking the nodes from which NUMA hinting faults are triggered. This can 1452 * be different from the set of nodes where the workload's memory is currently 1453 * located. 1454 * 1455 * The bitmask is used to make smarter decisions on when to do NUMA page 1456 * migrations, To prevent flip-flopping, and excessive page migrations, nodes 1457 * are added when they cause over 6/16 of the maximum number of faults, but 1458 * only removed when they drop below 3/16. 1459 */ 1460static void update_numa_active_node_mask(struct numa_group *numa_group) 1461{ 1462 unsigned long faults, max_faults = 0; 1463 int nid; 1464 1465 for_each_online_node(nid) { 1466 faults = group_faults_cpu(numa_group, nid); 1467 if (faults > max_faults) 1468 max_faults = faults; 1469 } 1470 1471 for_each_online_node(nid) { 1472 faults = group_faults_cpu(numa_group, nid); 1473 if (!node_isset(nid, numa_group->active_nodes)) { 1474 if (faults > max_faults * 6 / 16) 1475 node_set(nid, numa_group->active_nodes); 1476 } else if (faults < max_faults * 3 / 16) 1477 node_clear(nid, numa_group->active_nodes); 1478 } 1479} 1480 1481/* 1482 * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS 1483 * increments. The more local the fault statistics are, the higher the scan 1484 * period will be for the next scan window. If local/(local+remote) ratio is 1485 * below NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS) 1486 * the scan period will decrease. Aim for 70% local accesses. 1487 */ 1488#define NUMA_PERIOD_SLOTS 10 1489#define NUMA_PERIOD_THRESHOLD 7 1490 1491/* 1492 * Increase the scan period (slow down scanning) if the majority of 1493 * our memory is already on our local node, or if the majority of 1494 * the page accesses are shared with other processes. 1495 * Otherwise, decrease the scan period. 1496 */ 1497static void update_task_scan_period(struct task_struct *p, 1498 unsigned long shared, unsigned long private) 1499{ 1500 unsigned int period_slot; 1501 int ratio; 1502 int diff; 1503 1504 unsigned long remote = p->numa_faults_locality[0]; 1505 unsigned long local = p->numa_faults_locality[1]; 1506 1507 /* 1508 * If there were no record hinting faults then either the task is 1509 * completely idle or all activity is areas that are not of interest 1510 * to automatic numa balancing. Scan slower 1511 */ 1512 if (local + shared == 0) { 1513 p->numa_scan_period = min(p->numa_scan_period_max, 1514 p->numa_scan_period << 1); 1515 1516 p->mm->numa_next_scan = jiffies + 1517 msecs_to_jiffies(p->numa_scan_period); 1518 1519 return; 1520 } 1521 1522 /* 1523 * Prepare to scale scan period relative to the current period. 1524 * == NUMA_PERIOD_THRESHOLD scan period stays the same 1525 * < NUMA_PERIOD_THRESHOLD scan period decreases (scan faster) 1526 * >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower) 1527 */ 1528 period_slot = DIV_ROUND_UP(p->numa_scan_period, NUMA_PERIOD_SLOTS); 1529 ratio = (local * NUMA_PERIOD_SLOTS) / (local + remote); 1530 if (ratio >= NUMA_PERIOD_THRESHOLD) { 1531 int slot = ratio - NUMA_PERIOD_THRESHOLD; 1532 if (!slot) 1533 slot = 1; 1534 diff = slot * period_slot; 1535 } else { 1536 diff = -(NUMA_PERIOD_THRESHOLD - ratio) * period_slot; 1537 1538 /* 1539 * Scale scan rate increases based on sharing. There is an 1540 * inverse relationship between the degree of sharing and 1541 * the adjustment made to the scanning period. Broadly 1542 * speaking the intent is that there is little point 1543 * scanning faster if shared accesses dominate as it may 1544 * simply bounce migrations uselessly 1545 */ 1546 ratio = DIV_ROUND_UP(private * NUMA_PERIOD_SLOTS, (private + shared + 1)); 1547 diff = (diff * ratio) / NUMA_PERIOD_SLOTS; 1548 } 1549 1550 p->numa_scan_period = clamp(p->numa_scan_period + diff, 1551 task_scan_min(p), task_scan_max(p)); 1552 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality)); 1553} 1554 1555/* 1556 * Get the fraction of time the task has been running since the last 1557 * NUMA placement cycle. The scheduler keeps similar statistics, but 1558 * decays those on a 32ms period, which is orders of magnitude off 1559 * from the dozens-of-seconds NUMA balancing period. Use the scheduler 1560 * stats only if the task is so new there are no NUMA statistics yet. 1561 */ 1562static u64 numa_get_avg_runtime(struct task_struct *p, u64 *period) 1563{ 1564 u64 runtime, delta, now; 1565 /* Use the start of this time slice to avoid calculations. */ 1566 now = p->se.exec_start; 1567 runtime = p->se.sum_exec_runtime; 1568 1569 if (p->last_task_numa_placement) { 1570 delta = runtime - p->last_sum_exec_runtime; 1571 *period = now - p->last_task_numa_placement; 1572 } else { 1573 delta = p->se.avg.runnable_avg_sum; 1574 *period = p->se.avg.runnable_avg_period; 1575 } 1576 1577 p->last_sum_exec_runtime = runtime; 1578 p->last_task_numa_placement = now; 1579 1580 return delta; 1581} 1582 1583static void task_numa_placement(struct task_struct *p) 1584{ 1585 int seq, nid, max_nid = -1, max_group_nid = -1; 1586 unsigned long max_faults = 0, max_group_faults = 0; 1587 unsigned long fault_types[2] = { 0, 0 }; 1588 unsigned long total_faults; 1589 u64 runtime, period; 1590 spinlock_t *group_lock = NULL; 1591 1592 seq = ACCESS_ONCE(p->mm->numa_scan_seq); 1593 if (p->numa_scan_seq == seq) 1594 return; 1595 p->numa_scan_seq = seq; 1596 p->numa_scan_period_max = task_scan_max(p); 1597 1598 total_faults = p->numa_faults_locality[0] + 1599 p->numa_faults_locality[1]; 1600 runtime = numa_get_avg_runtime(p, &period); 1601 1602 /* If the task is part of a group prevent parallel updates to group stats */ 1603 if (p->numa_group) { 1604 group_lock = &p->numa_group->lock; 1605 spin_lock_irq(group_lock); 1606 } 1607 1608 /* Find the node with the highest number of faults */ 1609 for_each_online_node(nid) { 1610 unsigned long faults = 0, group_faults = 0; 1611 int priv, i; 1612 1613 for (priv = 0; priv < NR_NUMA_HINT_FAULT_TYPES; priv++) { 1614 long diff, f_diff, f_weight; 1615 1616 i = task_faults_idx(nid, priv); 1617 1618 /* Decay existing window, copy faults since last scan */ 1619 diff = p->numa_faults_buffer_memory[i] - p->numa_faults_memory[i] / 2; 1620 fault_types[priv] += p->numa_faults_buffer_memory[i]; 1621 p->numa_faults_buffer_memory[i] = 0; 1622 1623 /* 1624 * Normalize the faults_from, so all tasks in a group 1625 * count according to CPU use, instead of by the raw 1626 * number of faults. Tasks with little runtime have 1627 * little over-all impact on throughput, and thus their 1628 * faults are less important. 1629 */ 1630 f_weight = div64_u64(runtime << 16, period + 1); 1631 f_weight = (f_weight * p->numa_faults_buffer_cpu[i]) / 1632 (total_faults + 1); 1633 f_diff = f_weight - p->numa_faults_cpu[i] / 2; 1634 p->numa_faults_buffer_cpu[i] = 0; 1635 1636 p->numa_faults_memory[i] += diff; 1637 p->numa_faults_cpu[i] += f_diff; 1638 faults += p->numa_faults_memory[i]; 1639 p->total_numa_faults += diff; 1640 if (p->numa_group) { 1641 /* safe because we can only change our own group */ 1642 p->numa_group->faults[i] += diff; 1643 p->numa_group->faults_cpu[i] += f_diff; 1644 p->numa_group->total_faults += diff; 1645 group_faults += p->numa_group->faults[i]; 1646 } 1647 } 1648 1649 if (faults > max_faults) { 1650 max_faults = faults; 1651 max_nid = nid; 1652 } 1653 1654 if (group_faults > max_group_faults) { 1655 max_group_faults = group_faults; 1656 max_group_nid = nid; 1657 } 1658 } 1659 1660 update_task_scan_period(p, fault_types[0], fault_types[1]); 1661 1662 if (p->numa_group) { 1663 update_numa_active_node_mask(p->numa_group); 1664 spin_unlock_irq(group_lock); 1665 max_nid = max_group_nid; 1666 } 1667 1668 if (max_faults) { 1669 /* Set the new preferred node */ 1670 if (max_nid != p->numa_preferred_nid) 1671 sched_setnuma(p, max_nid); 1672 1673 if (task_node(p) != p->numa_preferred_nid) 1674 numa_migrate_preferred(p); 1675 } 1676} 1677 1678static inline int get_numa_group(struct numa_group *grp) 1679{ 1680 return atomic_inc_not_zero(&grp->refcount); 1681} 1682 1683static inline void put_numa_group(struct numa_group *grp) 1684{ 1685 if (atomic_dec_and_test(&grp->refcount)) 1686 kfree_rcu(grp, rcu); 1687} 1688 1689static void task_numa_group(struct task_struct *p, int cpupid, int flags, 1690 int *priv) 1691{ 1692 struct numa_group *grp, *my_grp; 1693 struct task_struct *tsk; 1694 bool join = false; 1695 int cpu = cpupid_to_cpu(cpupid); 1696 int i; 1697 1698 if (unlikely(!p->numa_group)) { 1699 unsigned int size = sizeof(struct numa_group) + 1700 4*nr_node_ids*sizeof(unsigned long); 1701 1702 grp = kzalloc(size, GFP_KERNEL | __GFP_NOWARN); 1703 if (!grp) 1704 return; 1705 1706 atomic_set(&grp->refcount, 1); 1707 spin_lock_init(&grp->lock); 1708 INIT_LIST_HEAD(&grp->task_list); 1709 grp->gid = p->pid; 1710 /* Second half of the array tracks nids where faults happen */ 1711 grp->faults_cpu = grp->faults + NR_NUMA_HINT_FAULT_TYPES * 1712 nr_node_ids; 1713 1714 node_set(task_node(current), grp->active_nodes); 1715 1716 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) 1717 grp->faults[i] = p->numa_faults_memory[i]; 1718 1719 grp->total_faults = p->total_numa_faults; 1720 1721 list_add(&p->numa_entry, &grp->task_list); 1722 grp->nr_tasks++; 1723 rcu_assign_pointer(p->numa_group, grp); 1724 } 1725 1726 rcu_read_lock(); 1727 tsk = ACCESS_ONCE(cpu_rq(cpu)->curr); 1728 1729 if (!cpupid_match_pid(tsk, cpupid)) 1730 goto no_join; 1731 1732 grp = rcu_dereference(tsk->numa_group); 1733 if (!grp) 1734 goto no_join; 1735 1736 my_grp = p->numa_group; 1737 if (grp == my_grp) 1738 goto no_join; 1739 1740 /* 1741 * Only join the other group if its bigger; if we're the bigger group, 1742 * the other task will join us. 1743 */ 1744 if (my_grp->nr_tasks > grp->nr_tasks) 1745 goto no_join; 1746 1747 /* 1748 * Tie-break on the grp address. 1749 */ 1750 if (my_grp->nr_tasks == grp->nr_tasks && my_grp > grp) 1751 goto no_join; 1752 1753 /* Always join threads in the same process. */ 1754 if (tsk->mm == current->mm) 1755 join = true; 1756 1757 /* Simple filter to avoid false positives due to PID collisions */ 1758 if (flags & TNF_SHARED) 1759 join = true; 1760 1761 /* Update priv based on whether false sharing was detected */ 1762 *priv = !join; 1763 1764 if (join && !get_numa_group(grp)) 1765 goto no_join; 1766 1767 rcu_read_unlock(); 1768 1769 if (!join) 1770 return; 1771 1772 BUG_ON(irqs_disabled()); 1773 double_lock_irq(&my_grp->lock, &grp->lock); 1774 1775 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) { 1776 my_grp->faults[i] -= p->numa_faults_memory[i]; 1777 grp->faults[i] += p->numa_faults_memory[i]; 1778 } 1779 my_grp->total_faults -= p->total_numa_faults; 1780 grp->total_faults += p->total_numa_faults; 1781 1782 list_move(&p->numa_entry, &grp->task_list); 1783 my_grp->nr_tasks--; 1784 grp->nr_tasks++; 1785 1786 spin_unlock(&my_grp->lock); 1787 spin_unlock_irq(&grp->lock); 1788 1789 rcu_assign_pointer(p->numa_group, grp); 1790 1791 put_numa_group(my_grp); 1792 return; 1793 1794no_join: 1795 rcu_read_unlock(); 1796 return; 1797} 1798 1799void task_numa_free(struct task_struct *p) 1800{ 1801 struct numa_group *grp = p->numa_group; 1802 void *numa_faults = p->numa_faults_memory; 1803 unsigned long flags; 1804 int i; 1805 1806 if (grp) { 1807 spin_lock_irqsave(&grp->lock, flags); 1808 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) 1809 grp->faults[i] -= p->numa_faults_memory[i]; 1810 grp->total_faults -= p->total_numa_faults; 1811 1812 list_del(&p->numa_entry); 1813 grp->nr_tasks--; 1814 spin_unlock_irqrestore(&grp->lock, flags); 1815 RCU_INIT_POINTER(p->numa_group, NULL); 1816 put_numa_group(grp); 1817 } 1818 1819 p->numa_faults_memory = NULL; 1820 p->numa_faults_buffer_memory = NULL; 1821 p->numa_faults_cpu= NULL; 1822 p->numa_faults_buffer_cpu = NULL; 1823 kfree(numa_faults); 1824} 1825 1826/* 1827 * Got a PROT_NONE fault for a page on @node. 1828 */ 1829void task_numa_fault(int last_cpupid, int mem_node, int pages, int flags) 1830{ 1831 struct task_struct *p = current; 1832 bool migrated = flags & TNF_MIGRATED; 1833 int cpu_node = task_node(current); 1834 int local = !!(flags & TNF_FAULT_LOCAL); 1835 int priv; 1836 1837 if (!numabalancing_enabled) 1838 return; 1839 1840 /* for example, ksmd faulting in a user's mm */ 1841 if (!p->mm) 1842 return; 1843 1844 /* Allocate buffer to track faults on a per-node basis */ 1845 if (unlikely(!p->numa_faults_memory)) { 1846 int size = sizeof(*p->numa_faults_memory) * 1847 NR_NUMA_HINT_FAULT_BUCKETS * nr_node_ids; 1848 1849 p->numa_faults_memory = kzalloc(size, GFP_KERNEL|__GFP_NOWARN); 1850 if (!p->numa_faults_memory) 1851 return; 1852 1853 BUG_ON(p->numa_faults_buffer_memory); 1854 /* 1855 * The averaged statistics, shared & private, memory & cpu, 1856 * occupy the first half of the array. The second half of the 1857 * array is for current counters, which are averaged into the 1858 * first set by task_numa_placement. 1859 */ 1860 p->numa_faults_cpu = p->numa_faults_memory + (2 * nr_node_ids); 1861 p->numa_faults_buffer_memory = p->numa_faults_memory + (4 * nr_node_ids); 1862 p->numa_faults_buffer_cpu = p->numa_faults_memory + (6 * nr_node_ids); 1863 p->total_numa_faults = 0; 1864 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality)); 1865 } 1866 1867 /* 1868 * First accesses are treated as private, otherwise consider accesses 1869 * to be private if the accessing pid has not changed 1870 */ 1871 if (unlikely(last_cpupid == (-1 & LAST_CPUPID_MASK))) { 1872 priv = 1; 1873 } else { 1874 priv = cpupid_match_pid(p, last_cpupid); 1875 if (!priv && !(flags & TNF_NO_GROUP)) 1876 task_numa_group(p, last_cpupid, flags, &priv); 1877 } 1878 1879 /* 1880 * If a workload spans multiple NUMA nodes, a shared fault that 1881 * occurs wholly within the set of nodes that the workload is 1882 * actively using should be counted as local. This allows the 1883 * scan rate to slow down when a workload has settled down. 1884 */ 1885 if (!priv && !local && p->numa_group && 1886 node_isset(cpu_node, p->numa_group->active_nodes) && 1887 node_isset(mem_node, p->numa_group->active_nodes)) 1888 local = 1; 1889 1890 task_numa_placement(p); 1891 1892 /* 1893 * Retry task to preferred node migration periodically, in case it 1894 * case it previously failed, or the scheduler moved us. 1895 */ 1896 if (time_after(jiffies, p->numa_migrate_retry)) 1897 numa_migrate_preferred(p); 1898 1899 if (migrated) 1900 p->numa_pages_migrated += pages; 1901 1902 p->numa_faults_buffer_memory[task_faults_idx(mem_node, priv)] += pages; 1903 p->numa_faults_buffer_cpu[task_faults_idx(cpu_node, priv)] += pages; 1904 p->numa_faults_locality[local] += pages; 1905} 1906 1907static void reset_ptenuma_scan(struct task_struct *p) 1908{ 1909 ACCESS_ONCE(p->mm->numa_scan_seq)++; 1910 p->mm->numa_scan_offset = 0; 1911} 1912 1913/* 1914 * The expensive part of numa migration is done from task_work context. 1915 * Triggered from task_tick_numa(). 1916 */ 1917void task_numa_work(struct callback_head *work) 1918{ 1919 unsigned long migrate, next_scan, now = jiffies; 1920 struct task_struct *p = current; 1921 struct mm_struct *mm = p->mm; 1922 struct vm_area_struct *vma; 1923 unsigned long start, end; 1924 unsigned long nr_pte_updates = 0; 1925 long pages; 1926 1927 WARN_ON_ONCE(p != container_of(work, struct task_struct, numa_work)); 1928 1929 work->next = work; /* protect against double add */ 1930 /* 1931 * Who cares about NUMA placement when they're dying. 1932 * 1933 * NOTE: make sure not to dereference p->mm before this check, 1934 * exit_task_work() happens _after_ exit_mm() so we could be called 1935 * without p->mm even though we still had it when we enqueued this 1936 * work. 1937 */ 1938 if (p->flags & PF_EXITING) 1939 return; 1940 1941 if (!mm->numa_next_scan) { 1942 mm->numa_next_scan = now + 1943 msecs_to_jiffies(sysctl_numa_balancing_scan_delay); 1944 } 1945 1946 /* 1947 * Enforce maximal scan/migration frequency.. 1948 */ 1949 migrate = mm->numa_next_scan; 1950 if (time_before(now, migrate)) 1951 return; 1952 1953 if (p->numa_scan_period == 0) { 1954 p->numa_scan_period_max = task_scan_max(p); 1955 p->numa_scan_period = task_scan_min(p); 1956 } 1957 1958 next_scan = now + msecs_to_jiffies(p->numa_scan_period); 1959 if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate) 1960 return; 1961 1962 /* 1963 * Delay this task enough that another task of this mm will likely win 1964 * the next time around. 1965 */ 1966 p->node_stamp += 2 * TICK_NSEC; 1967 1968 start = mm->numa_scan_offset; 1969 pages = sysctl_numa_balancing_scan_size; 1970 pages <<= 20 - PAGE_SHIFT; /* MB in pages */ 1971 if (!pages) 1972 return; 1973 1974 down_read(&mm->mmap_sem); 1975 vma = find_vma(mm, start); 1976 if (!vma) { 1977 reset_ptenuma_scan(p); 1978 start = 0; 1979 vma = mm->mmap; 1980 } 1981 for (; vma; vma = vma->vm_next) { 1982 if (!vma_migratable(vma) || !vma_policy_mof(vma)) 1983 continue; 1984 1985 /* 1986 * Shared library pages mapped by multiple processes are not 1987 * migrated as it is expected they are cache replicated. Avoid 1988 * hinting faults in read-only file-backed mappings or the vdso 1989 * as migrating the pages will be of marginal benefit. 1990 */ 1991 if (!vma->vm_mm || 1992 (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ))) 1993 continue; 1994 1995 /* 1996 * Skip inaccessible VMAs to avoid any confusion between 1997 * PROT_NONE and NUMA hinting ptes 1998 */ 1999 if (!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE))) 2000 continue; 2001 2002 do { 2003 start = max(start, vma->vm_start); 2004 end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE); 2005 end = min(end, vma->vm_end); 2006 nr_pte_updates += change_prot_numa(vma, start, end); 2007 2008 /* 2009 * Scan sysctl_numa_balancing_scan_size but ensure that 2010 * at least one PTE is updated so that unused virtual 2011 * address space is quickly skipped. 2012 */ 2013 if (nr_pte_updates) 2014 pages -= (end - start) >> PAGE_SHIFT; 2015 2016 start = end; 2017 if (pages <= 0) 2018 goto out; 2019 2020 cond_resched(); 2021 } while (end != vma->vm_end); 2022 } 2023 2024out: 2025 /* 2026 * It is possible to reach the end of the VMA list but the last few 2027 * VMAs are not guaranteed to the vma_migratable. If they are not, we 2028 * would find the !migratable VMA on the next scan but not reset the 2029 * scanner to the start so check it now. 2030 */ 2031 if (vma) 2032 mm->numa_scan_offset = start; 2033 else 2034 reset_ptenuma_scan(p); 2035 up_read(&mm->mmap_sem); 2036} 2037 2038/* 2039 * Drive the periodic memory faults.. 2040 */ 2041void task_tick_numa(struct rq *rq, struct task_struct *curr) 2042{ 2043 struct callback_head *work = &curr->numa_work; 2044 u64 period, now; 2045 2046 /* 2047 * We don't care about NUMA placement if we don't have memory. 2048 */ 2049 if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work) 2050 return; 2051 2052 /* 2053 * Using runtime rather than walltime has the dual advantage that 2054 * we (mostly) drive the selection from busy threads and that the 2055 * task needs to have done some actual work before we bother with 2056 * NUMA placement. 2057 */ 2058 now = curr->se.sum_exec_runtime; 2059 period = (u64)curr->numa_scan_period * NSEC_PER_MSEC; 2060 2061 if (now - curr->node_stamp > period) { 2062 if (!curr->node_stamp) 2063 curr->numa_scan_period = task_scan_min(curr); 2064 curr->node_stamp += period; 2065 2066 if (!time_before(jiffies, curr->mm->numa_next_scan)) { 2067 init_task_work(work, task_numa_work); /* TODO: move this into sched_fork() */ 2068 task_work_add(curr, work, true); 2069 } 2070 } 2071} 2072#else 2073static void task_tick_numa(struct rq *rq, struct task_struct *curr) 2074{ 2075} 2076 2077static inline void account_numa_enqueue(struct rq *rq, struct task_struct *p) 2078{ 2079} 2080 2081static inline void account_numa_dequeue(struct rq *rq, struct task_struct *p) 2082{ 2083} 2084#endif /* CONFIG_NUMA_BALANCING */ 2085 2086static void 2087account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se) 2088{ 2089 update_load_add(&cfs_rq->load, se->load.weight); 2090 if (!parent_entity(se)) 2091 update_load_add(&rq_of(cfs_rq)->load, se->load.weight); 2092#ifdef CONFIG_SMP 2093 if (entity_is_task(se)) { 2094 struct rq *rq = rq_of(cfs_rq); 2095 2096 account_numa_enqueue(rq, task_of(se)); 2097 list_add(&se->group_node, &rq->cfs_tasks); 2098 } 2099#endif 2100 cfs_rq->nr_running++; 2101} 2102 2103static void 2104account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se) 2105{ 2106 update_load_sub(&cfs_rq->load, se->load.weight); 2107 if (!parent_entity(se)) 2108 update_load_sub(&rq_of(cfs_rq)->load, se->load.weight); 2109 if (entity_is_task(se)) { 2110 account_numa_dequeue(rq_of(cfs_rq), task_of(se)); 2111 list_del_init(&se->group_node); 2112 } 2113 cfs_rq->nr_running--; 2114} 2115 2116#ifdef CONFIG_FAIR_GROUP_SCHED 2117# ifdef CONFIG_SMP 2118static inline long calc_tg_weight(struct task_group *tg, struct cfs_rq *cfs_rq) 2119{ 2120 long tg_weight; 2121 2122 /* 2123 * Use this CPU's actual weight instead of the last load_contribution 2124 * to gain a more accurate current total weight. See 2125 * update_cfs_rq_load_contribution(). 2126 */ 2127 tg_weight = atomic_long_read(&tg->load_avg); 2128 tg_weight -= cfs_rq->tg_load_contrib; 2129 tg_weight += cfs_rq->load.weight; 2130 2131 return tg_weight; 2132} 2133 2134static long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg) 2135{ 2136 long tg_weight, load, shares; 2137 2138 tg_weight = calc_tg_weight(tg, cfs_rq); 2139 load = cfs_rq->load.weight; 2140 2141 shares = (tg->shares * load); 2142 if (tg_weight) 2143 shares /= tg_weight; 2144 2145 if (shares < MIN_SHARES) 2146 shares = MIN_SHARES; 2147 if (shares > tg->shares) 2148 shares = tg->shares; 2149 2150 return shares; 2151} 2152# else /* CONFIG_SMP */ 2153static inline long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg) 2154{ 2155 return tg->shares; 2156} 2157# endif /* CONFIG_SMP */ 2158static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, 2159 unsigned long weight) 2160{ 2161 if (se->on_rq) { 2162 /* commit outstanding execution time */ 2163 if (cfs_rq->curr == se) 2164 update_curr(cfs_rq); 2165 account_entity_dequeue(cfs_rq, se); 2166 } 2167 2168 update_load_set(&se->load, weight); 2169 2170 if (se->on_rq) 2171 account_entity_enqueue(cfs_rq, se); 2172} 2173 2174static inline int throttled_hierarchy(struct cfs_rq *cfs_rq); 2175 2176static void update_cfs_shares(struct cfs_rq *cfs_rq) 2177{ 2178 struct task_group *tg; 2179 struct sched_entity *se; 2180 long shares; 2181 2182 tg = cfs_rq->tg; 2183 se = tg->se[cpu_of(rq_of(cfs_rq))]; 2184 if (!se || throttled_hierarchy(cfs_rq)) 2185 return; 2186#ifndef CONFIG_SMP 2187 if (likely(se->load.weight == tg->shares)) 2188 return; 2189#endif 2190 shares = calc_cfs_shares(cfs_rq, tg); 2191 2192 reweight_entity(cfs_rq_of(se), se, shares); 2193} 2194#else /* CONFIG_FAIR_GROUP_SCHED */ 2195static inline void update_cfs_shares(struct cfs_rq *cfs_rq) 2196{ 2197} 2198#endif /* CONFIG_FAIR_GROUP_SCHED */ 2199 2200#ifdef CONFIG_SMP 2201/* 2202 * We choose a half-life close to 1 scheduling period. 2203 * Note: The tables below are dependent on this value. 2204 */ 2205#define LOAD_AVG_PERIOD 32 2206#define LOAD_AVG_MAX 47742 /* maximum possible load avg */ 2207#define LOAD_AVG_MAX_N 345 /* number of full periods to produce LOAD_MAX_AVG */ 2208 2209/* Precomputed fixed inverse multiplies for multiplication by y^n */ 2210static const u32 runnable_avg_yN_inv[] = { 2211 0xffffffff, 0xfa83b2da, 0xf5257d14, 0xefe4b99a, 0xeac0c6e6, 0xe5b906e6, 2212 0xe0ccdeeb, 0xdbfbb796, 0xd744fcc9, 0xd2a81d91, 0xce248c14, 0xc9b9bd85, 2213 0xc5672a10, 0xc12c4cc9, 0xbd08a39e, 0xb8fbaf46, 0xb504f333, 0xb123f581, 2214 0xad583ee9, 0xa9a15ab4, 0xa5fed6a9, 0xa2704302, 0x9ef5325f, 0x9b8d39b9, 2215 0x9837f050, 0x94f4efa8, 0x91c3d373, 0x8ea4398a, 0x8b95c1e3, 0x88980e80, 2216 0x85aac367, 0x82cd8698, 2217}; 2218 2219/* 2220 * Precomputed \Sum y^k { 1<=k<=n }. These are floor(true_value) to prevent 2221 * over-estimates when re-combining. 2222 */ 2223static const u32 runnable_avg_yN_sum[] = { 2224 0, 1002, 1982, 2941, 3880, 4798, 5697, 6576, 7437, 8279, 9103, 2225 9909,10698,11470,12226,12966,13690,14398,15091,15769,16433,17082, 2226 17718,18340,18949,19545,20128,20698,21256,21802,22336,22859,23371, 2227}; 2228 2229/* 2230 * Approximate: 2231 * val * y^n, where y^32 ~= 0.5 (~1 scheduling period) 2232 */ 2233static __always_inline u64 decay_load(u64 val, u64 n) 2234{ 2235 unsigned int local_n; 2236 2237 if (!n) 2238 return val; 2239 else if (unlikely(n > LOAD_AVG_PERIOD * 63)) 2240 return 0; 2241 2242 /* after bounds checking we can collapse to 32-bit */ 2243 local_n = n; 2244 2245 /* 2246 * As y^PERIOD = 1/2, we can combine 2247 * y^n = 1/2^(n/PERIOD) * y^(n%PERIOD) 2248 * With a look-up table which covers y^n (n<PERIOD) 2249 * 2250 * To achieve constant time decay_load. 2251 */ 2252 if (unlikely(local_n >= LOAD_AVG_PERIOD)) { 2253 val >>= local_n / LOAD_AVG_PERIOD; 2254 local_n %= LOAD_AVG_PERIOD; 2255 } 2256 2257 val *= runnable_avg_yN_inv[local_n]; 2258 /* We don't use SRR here since we always want to round down. */ 2259 return val >> 32; 2260} 2261 2262/* 2263 * For updates fully spanning n periods, the contribution to runnable 2264 * average will be: \Sum 1024*y^n 2265 * 2266 * We can compute this reasonably efficiently by combining: 2267 * y^PERIOD = 1/2 with precomputed \Sum 1024*y^n {for n <PERIOD} 2268 */ 2269static u32 __compute_runnable_contrib(u64 n) 2270{ 2271 u32 contrib = 0; 2272 2273 if (likely(n <= LOAD_AVG_PERIOD)) 2274 return runnable_avg_yN_sum[n]; 2275 else if (unlikely(n >= LOAD_AVG_MAX_N)) 2276 return LOAD_AVG_MAX; 2277 2278 /* Compute \Sum k^n combining precomputed values for k^i, \Sum k^j */ 2279 do { 2280 contrib /= 2; /* y^LOAD_AVG_PERIOD = 1/2 */ 2281 contrib += runnable_avg_yN_sum[LOAD_AVG_PERIOD]; 2282 2283 n -= LOAD_AVG_PERIOD; 2284 } while (n > LOAD_AVG_PERIOD); 2285 2286 contrib = decay_load(contrib, n); 2287 return contrib + runnable_avg_yN_sum[n]; 2288} 2289 2290/* 2291 * We can represent the historical contribution to runnable average as the 2292 * coefficients of a geometric series. To do this we sub-divide our runnable 2293 * history into segments of approximately 1ms (1024us); label the segment that 2294 * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g. 2295 * 2296 * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ... 2297 * p0 p1 p2 2298 * (now) (~1ms ago) (~2ms ago) 2299 * 2300 * Let u_i denote the fraction of p_i that the entity was runnable. 2301 * 2302 * We then designate the fractions u_i as our co-efficients, yielding the 2303 * following representation of historical load: 2304 * u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ... 2305 * 2306 * We choose y based on the with of a reasonably scheduling period, fixing: 2307 * y^32 = 0.5 2308 * 2309 * This means that the contribution to load ~32ms ago (u_32) will be weighted 2310 * approximately half as much as the contribution to load within the last ms 2311 * (u_0). 2312 * 2313 * When a period "rolls over" and we have new u_0`, multiplying the previous 2314 * sum again by y is sufficient to update: 2315 * load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... ) 2316 * = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}] 2317 */ 2318static __always_inline int __update_entity_runnable_avg(u64 now, 2319 struct sched_avg *sa, 2320 int runnable) 2321{ 2322 u64 delta, periods; 2323 u32 runnable_contrib; 2324 int delta_w, decayed = 0; 2325 2326 delta = now - sa->last_runnable_update; 2327 /* 2328 * This should only happen when time goes backwards, which it 2329 * unfortunately does during sched clock init when we swap over to TSC. 2330 */ 2331 if ((s64)delta < 0) { 2332 sa->last_runnable_update = now; 2333 return 0; 2334 } 2335 2336 /* 2337 * Use 1024ns as the unit of measurement since it's a reasonable 2338 * approximation of 1us and fast to compute. 2339 */ 2340 delta >>= 10; 2341 if (!delta) 2342 return 0; 2343 sa->last_runnable_update = now; 2344 2345 /* delta_w is the amount already accumulated against our next period */ 2346 delta_w = sa->runnable_avg_period % 1024; 2347 if (delta + delta_w >= 1024) { 2348 /* period roll-over */ 2349 decayed = 1; 2350 2351 /* 2352 * Now that we know we're crossing a period boundary, figure 2353 * out how much from delta we need to complete the current 2354 * period and accrue it. 2355 */ 2356 delta_w = 1024 - delta_w; 2357 if (runnable) 2358 sa->runnable_avg_sum += delta_w; 2359 sa->runnable_avg_period += delta_w; 2360 2361 delta -= delta_w; 2362 2363 /* Figure out how many additional periods this update spans */ 2364 periods = delta / 1024; 2365 delta %= 1024; 2366 2367 sa->runnable_avg_sum = decay_load(sa->runnable_avg_sum, 2368 periods + 1); 2369 sa->runnable_avg_period = decay_load(sa->runnable_avg_period, 2370 periods + 1); 2371 2372 /* Efficiently calculate \sum (1..n_period) 1024*y^i */ 2373 runnable_contrib = __compute_runnable_contrib(periods); 2374 if (runnable) 2375 sa->runnable_avg_sum += runnable_contrib; 2376 sa->runnable_avg_period += runnable_contrib; 2377 } 2378 2379 /* Remainder of delta accrued against u_0` */ 2380 if (runnable) 2381 sa->runnable_avg_sum += delta; 2382 sa->runnable_avg_period += delta; 2383 2384 return decayed; 2385} 2386 2387/* Synchronize an entity's decay with its parenting cfs_rq.*/ 2388static inline u64 __synchronize_entity_decay(struct sched_entity *se) 2389{ 2390 struct cfs_rq *cfs_rq = cfs_rq_of(se); 2391 u64 decays = atomic64_read(&cfs_rq->decay_counter); 2392 2393 decays -= se->avg.decay_count; 2394 if (!decays) 2395 return 0; 2396 2397 se->avg.load_avg_contrib = decay_load(se->avg.load_avg_contrib, decays); 2398 se->avg.decay_count = 0; 2399 2400 return decays; 2401} 2402 2403#ifdef CONFIG_FAIR_GROUP_SCHED 2404static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq *cfs_rq, 2405 int force_update) 2406{ 2407 struct task_group *tg = cfs_rq->tg; 2408 long tg_contrib; 2409 2410 tg_contrib = cfs_rq->runnable_load_avg + cfs_rq->blocked_load_avg; 2411 tg_contrib -= cfs_rq->tg_load_contrib; 2412 2413 if (!tg_contrib) 2414 return; 2415 2416 if (force_update || abs(tg_contrib) > cfs_rq->tg_load_contrib / 8) { 2417 atomic_long_add(tg_contrib, &tg->load_avg); 2418 cfs_rq->tg_load_contrib += tg_contrib; 2419 } 2420} 2421 2422/* 2423 * Aggregate cfs_rq runnable averages into an equivalent task_group 2424 * representation for computing load contributions. 2425 */ 2426static inline void __update_tg_runnable_avg(struct sched_avg *sa, 2427 struct cfs_rq *cfs_rq) 2428{ 2429 struct task_group *tg = cfs_rq->tg; 2430 long contrib; 2431 2432 /* The fraction of a cpu used by this cfs_rq */ 2433 contrib = div_u64((u64)sa->runnable_avg_sum << NICE_0_SHIFT, 2434 sa->runnable_avg_period + 1); 2435 contrib -= cfs_rq->tg_runnable_contrib; 2436 2437 if (abs(contrib) > cfs_rq->tg_runnable_contrib / 64) { 2438 atomic_add(contrib, &tg->runnable_avg); 2439 cfs_rq->tg_runnable_contrib += contrib; 2440 } 2441} 2442 2443static inline void __update_group_entity_contrib(struct sched_entity *se) 2444{ 2445 struct cfs_rq *cfs_rq = group_cfs_rq(se); 2446 struct task_group *tg = cfs_rq->tg; 2447 int runnable_avg; 2448 2449 u64 contrib; 2450 2451 contrib = cfs_rq->tg_load_contrib * tg->shares; 2452 se->avg.load_avg_contrib = div_u64(contrib, 2453 atomic_long_read(&tg->load_avg) + 1); 2454 2455 /* 2456 * For group entities we need to compute a correction term in the case 2457 * that they are consuming <1 cpu so that we would contribute the same 2458 * load as a task of equal weight. 2459 * 2460 * Explicitly co-ordinating this measurement would be expensive, but 2461 * fortunately the sum of each cpus contribution forms a usable 2462 * lower-bound on the true value. 2463 * 2464 * Consider the aggregate of 2 contributions. Either they are disjoint 2465 * (and the sum represents true value) or they are disjoint and we are 2466 * understating by the aggregate of their overlap. 2467 * 2468 * Extending this to N cpus, for a given overlap, the maximum amount we 2469 * understand is then n_i(n_i+1)/2 * w_i where n_i is the number of 2470 * cpus that overlap for this interval and w_i is the interval width. 2471 * 2472 * On a small machine; the first term is well-bounded which bounds the 2473 * total error since w_i is a subset of the period. Whereas on a 2474 * larger machine, while this first term can be larger, if w_i is the 2475 * of consequential size guaranteed to see n_i*w_i quickly converge to 2476 * our upper bound of 1-cpu. 2477 */ 2478 runnable_avg = atomic_read(&tg->runnable_avg); 2479 if (runnable_avg < NICE_0_LOAD) { 2480 se->avg.load_avg_contrib *= runnable_avg; 2481 se->avg.load_avg_contrib >>= NICE_0_SHIFT; 2482 } 2483} 2484 2485static inline void update_rq_runnable_avg(struct rq *rq, int runnable) 2486{ 2487 __update_entity_runnable_avg(rq_clock_task(rq), &rq->avg, runnable); 2488 __update_tg_runnable_avg(&rq->avg, &rq->cfs); 2489} 2490#else /* CONFIG_FAIR_GROUP_SCHED */ 2491static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq *cfs_rq, 2492 int force_update) {} 2493static inline void __update_tg_runnable_avg(struct sched_avg *sa, 2494 struct cfs_rq *cfs_rq) {} 2495static inline void __update_group_entity_contrib(struct sched_entity *se) {} 2496static inline void update_rq_runnable_avg(struct rq *rq, int runnable) {} 2497#endif /* CONFIG_FAIR_GROUP_SCHED */ 2498 2499static inline void __update_task_entity_contrib(struct sched_entity *se) 2500{ 2501 u32 contrib; 2502 2503 /* avoid overflowing a 32-bit type w/ SCHED_LOAD_SCALE */ 2504 contrib = se->avg.runnable_avg_sum * scale_load_down(se->load.weight); 2505 contrib /= (se->avg.runnable_avg_period + 1); 2506 se->avg.load_avg_contrib = scale_load(contrib); 2507} 2508 2509/* Compute the current contribution to load_avg by se, return any delta */ 2510static long __update_entity_load_avg_contrib(struct sched_entity *se) 2511{ 2512 long old_contrib = se->avg.load_avg_contrib; 2513 2514 if (entity_is_task(se)) { 2515 __update_task_entity_contrib(se); 2516 } else { 2517 __update_tg_runnable_avg(&se->avg, group_cfs_rq(se)); 2518 __update_group_entity_contrib(se); 2519 } 2520 2521 return se->avg.load_avg_contrib - old_contrib; 2522} 2523 2524static inline void subtract_blocked_load_contrib(struct cfs_rq *cfs_rq, 2525 long load_contrib) 2526{ 2527 if (likely(load_contrib < cfs_rq->blocked_load_avg)) 2528 cfs_rq->blocked_load_avg -= load_contrib; 2529 else 2530 cfs_rq->blocked_load_avg = 0; 2531} 2532 2533static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq); 2534 2535/* Update a sched_entity's runnable average */ 2536static inline void update_entity_load_avg(struct sched_entity *se, 2537 int update_cfs_rq) 2538{ 2539 struct cfs_rq *cfs_rq = cfs_rq_of(se); 2540 long contrib_delta; 2541 u64 now; 2542 2543 /* 2544 * For a group entity we need to use their owned cfs_rq_clock_task() in 2545 * case they are the parent of a throttled hierarchy. 2546 */ 2547 if (entity_is_task(se)) 2548 now = cfs_rq_clock_task(cfs_rq); 2549 else 2550 now = cfs_rq_clock_task(group_cfs_rq(se)); 2551 2552 if (!__update_entity_runnable_avg(now, &se->avg, se->on_rq)) 2553 return; 2554 2555 contrib_delta = __update_entity_load_avg_contrib(se); 2556 2557 if (!update_cfs_rq) 2558 return; 2559 2560 if (se->on_rq) 2561 cfs_rq->runnable_load_avg += contrib_delta; 2562 else 2563 subtract_blocked_load_contrib(cfs_rq, -contrib_delta); 2564} 2565 2566/* 2567 * Decay the load contributed by all blocked children and account this so that 2568 * their contribution may appropriately discounted when they wake up. 2569 */ 2570static void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq, int force_update) 2571{ 2572 u64 now = cfs_rq_clock_task(cfs_rq) >> 20; 2573 u64 decays; 2574 2575 decays = now - cfs_rq->last_decay; 2576 if (!decays && !force_update) 2577 return; 2578 2579 if (atomic_long_read(&cfs_rq->removed_load)) { 2580 unsigned long removed_load; 2581 removed_load = atomic_long_xchg(&cfs_rq->removed_load, 0); 2582 subtract_blocked_load_contrib(cfs_rq, removed_load); 2583 } 2584 2585 if (decays) { 2586 cfs_rq->blocked_load_avg = decay_load(cfs_rq->blocked_load_avg, 2587 decays); 2588 atomic64_add(decays, &cfs_rq->decay_counter); 2589 cfs_rq->last_decay = now; 2590 } 2591 2592 __update_cfs_rq_tg_load_contrib(cfs_rq, force_update); 2593} 2594 2595/* Add the load generated by se into cfs_rq's child load-average */ 2596static inline void enqueue_entity_load_avg(struct cfs_rq *cfs_rq, 2597 struct sched_entity *se, 2598 int wakeup) 2599{ 2600 /* 2601 * We track migrations using entity decay_count <= 0, on a wake-up 2602 * migration we use a negative decay count to track the remote decays 2603 * accumulated while sleeping. 2604 * 2605 * Newly forked tasks are enqueued with se->avg.decay_count == 0, they 2606 * are seen by enqueue_entity_load_avg() as a migration with an already 2607 * constructed load_avg_contrib. 2608 */ 2609 if (unlikely(se->avg.decay_count <= 0)) { 2610 se->avg.last_runnable_update = rq_clock_task(rq_of(cfs_rq)); 2611 if (se->avg.decay_count) { 2612 /* 2613 * In a wake-up migration we have to approximate the 2614 * time sleeping. This is because we can't synchronize 2615 * clock_task between the two cpus, and it is not 2616 * guaranteed to be read-safe. Instead, we can 2617 * approximate this using our carried decays, which are 2618 * explicitly atomically readable. 2619 */ 2620 se->avg.last_runnable_update -= (-se->avg.decay_count) 2621 << 20; 2622 update_entity_load_avg(se, 0); 2623 /* Indicate that we're now synchronized and on-rq */ 2624 se->avg.decay_count = 0; 2625 } 2626 wakeup = 0; 2627 } else { 2628 __synchronize_entity_decay(se); 2629 } 2630 2631 /* migrated tasks did not contribute to our blocked load */ 2632 if (wakeup) { 2633 subtract_blocked_load_contrib(cfs_rq, se->avg.load_avg_contrib); 2634 update_entity_load_avg(se, 0); 2635 } 2636 2637 cfs_rq->runnable_load_avg += se->avg.load_avg_contrib; 2638 /* we force update consideration on load-balancer moves */ 2639 update_cfs_rq_blocked_load(cfs_rq, !wakeup); 2640} 2641 2642/* 2643 * Remove se's load from this cfs_rq child load-average, if the entity is 2644 * transitioning to a blocked state we track its projected decay using 2645 * blocked_load_avg. 2646 */ 2647static inline void dequeue_entity_load_avg(struct cfs_rq *cfs_rq, 2648 struct sched_entity *se, 2649 int sleep) 2650{ 2651 update_entity_load_avg(se, 1); 2652 /* we force update consideration on load-balancer moves */ 2653 update_cfs_rq_blocked_load(cfs_rq, !sleep); 2654 2655 cfs_rq->runnable_load_avg -= se->avg.load_avg_contrib; 2656 if (sleep) { 2657 cfs_rq->blocked_load_avg += se->avg.load_avg_contrib; 2658 se->avg.decay_count = atomic64_read(&cfs_rq->decay_counter); 2659 } /* migrations, e.g. sleep=0 leave decay_count == 0 */ 2660} 2661 2662/* 2663 * Update the rq's load with the elapsed running time before entering 2664 * idle. if the last scheduled task is not a CFS task, idle_enter will 2665 * be the only way to update the runnable statistic. 2666 */ 2667void idle_enter_fair(struct rq *this_rq) 2668{ 2669 update_rq_runnable_avg(this_rq, 1); 2670} 2671 2672/* 2673 * Update the rq's load with the elapsed idle time before a task is 2674 * scheduled. if the newly scheduled task is not a CFS task, idle_exit will 2675 * be the only way to update the runnable statistic. 2676 */ 2677void idle_exit_fair(struct rq *this_rq) 2678{ 2679 update_rq_runnable_avg(this_rq, 0); 2680} 2681 2682static int idle_balance(struct rq *this_rq); 2683 2684#else /* CONFIG_SMP */ 2685 2686static inline void update_entity_load_avg(struct sched_entity *se, 2687 int update_cfs_rq) {} 2688static inline void update_rq_runnable_avg(struct rq *rq, int runnable) {} 2689static inline void enqueue_entity_load_avg(struct cfs_rq *cfs_rq, 2690 struct sched_entity *se, 2691 int wakeup) {} 2692static inline void dequeue_entity_load_avg(struct cfs_rq *cfs_rq, 2693 struct sched_entity *se, 2694 int sleep) {} 2695static inline void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq, 2696 int force_update) {} 2697 2698static inline int idle_balance(struct rq *rq) 2699{ 2700 return 0; 2701} 2702 2703#endif /* CONFIG_SMP */ 2704 2705static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se) 2706{ 2707#ifdef CONFIG_SCHEDSTATS 2708 struct task_struct *tsk = NULL; 2709 2710 if (entity_is_task(se)) 2711 tsk = task_of(se); 2712 2713 if (se->statistics.sleep_start) { 2714 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.sleep_start; 2715 2716 if ((s64)delta < 0) 2717 delta = 0; 2718 2719 if (unlikely(delta > se->statistics.sleep_max)) 2720 se->statistics.sleep_max = delta; 2721 2722 se->statistics.sleep_start = 0; 2723 se->statistics.sum_sleep_runtime += delta; 2724 2725 if (tsk) { 2726 account_scheduler_latency(tsk, delta >> 10, 1); 2727 trace_sched_stat_sleep(tsk, delta); 2728 } 2729 } 2730 if (se->statistics.block_start) { 2731 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.block_start; 2732 2733 if ((s64)delta < 0) 2734 delta = 0; 2735 2736 if (unlikely(delta > se->statistics.block_max)) 2737 se->statistics.block_max = delta; 2738 2739 se->statistics.block_start = 0; 2740 se->statistics.sum_sleep_runtime += delta; 2741 2742 if (tsk) { 2743 if (tsk->in_iowait) { 2744 se->statistics.iowait_sum += delta; 2745 se->statistics.iowait_count++; 2746 trace_sched_stat_iowait(tsk, delta); 2747 } 2748 2749 trace_sched_stat_blocked(tsk, delta); 2750 2751 /* 2752 * Blocking time is in units of nanosecs, so shift by 2753 * 20 to get a milliseconds-range estimation of the 2754 * amount of time that the task spent sleeping: 2755 */ 2756 if (unlikely(prof_on == SLEEP_PROFILING)) { 2757 profile_hits(SLEEP_PROFILING, 2758 (void *)get_wchan(tsk), 2759 delta >> 20); 2760 } 2761 account_scheduler_latency(tsk, delta >> 10, 0); 2762 } 2763 } 2764#endif 2765} 2766 2767static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se) 2768{ 2769#ifdef CONFIG_SCHED_DEBUG 2770 s64 d = se->vruntime - cfs_rq->min_vruntime; 2771 2772 if (d < 0) 2773 d = -d; 2774 2775 if (d > 3*sysctl_sched_latency) 2776 schedstat_inc(cfs_rq, nr_spread_over); 2777#endif 2778} 2779 2780static void 2781place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial) 2782{ 2783 u64 vruntime = cfs_rq->min_vruntime; 2784 2785 /* 2786 * The 'current' period is already promised to the current tasks, 2787 * however the extra weight of the new task will slow them down a 2788 * little, place the new task so that it fits in the slot that 2789 * stays open at the end. 2790 */ 2791 if (initial && sched_feat(START_DEBIT)) 2792 vruntime += sched_vslice(cfs_rq, se); 2793 2794 /* sleeps up to a single latency don't count. */ 2795 if (!initial) { 2796 unsigned long thresh = sysctl_sched_latency; 2797 2798 /* 2799 * Halve their sleep time's effect, to allow 2800 * for a gentler effect of sleepers: 2801 */ 2802 if (sched_feat(GENTLE_FAIR_SLEEPERS)) 2803 thresh >>= 1; 2804 2805 vruntime -= thresh; 2806 } 2807 2808 /* ensure we never gain time by being placed backwards. */ 2809 se->vruntime = max_vruntime(se->vruntime, vruntime); 2810} 2811 2812static void check_enqueue_throttle(struct cfs_rq *cfs_rq); 2813 2814static void 2815enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags) 2816{ 2817 /* 2818 * Update the normalized vruntime before updating min_vruntime 2819 * through calling update_curr(). 2820 */ 2821 if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_WAKING)) 2822 se->vruntime += cfs_rq->min_vruntime; 2823 2824 /* 2825 * Update run-time statistics of the 'current'. 2826 */ 2827 update_curr(cfs_rq); 2828 enqueue_entity_load_avg(cfs_rq, se, flags & ENQUEUE_WAKEUP); 2829 account_entity_enqueue(cfs_rq, se); 2830 update_cfs_shares(cfs_rq); 2831 2832 if (flags & ENQUEUE_WAKEUP) { 2833 place_entity(cfs_rq, se, 0); 2834 enqueue_sleeper(cfs_rq, se); 2835 } 2836 2837 update_stats_enqueue(cfs_rq, se); 2838 check_spread(cfs_rq, se); 2839 if (se != cfs_rq->curr) 2840 __enqueue_entity(cfs_rq, se); 2841 se->on_rq = 1; 2842 2843 if (cfs_rq->nr_running == 1) { 2844 list_add_leaf_cfs_rq(cfs_rq); 2845 check_enqueue_throttle(cfs_rq); 2846 } 2847} 2848 2849static void __clear_buddies_last(struct sched_entity *se) 2850{ 2851 for_each_sched_entity(se) { 2852 struct cfs_rq *cfs_rq = cfs_rq_of(se); 2853 if (cfs_rq->last != se) 2854 break; 2855 2856 cfs_rq->last = NULL; 2857 } 2858} 2859 2860static void __clear_buddies_next(struct sched_entity *se) 2861{ 2862 for_each_sched_entity(se) { 2863 struct cfs_rq *cfs_rq = cfs_rq_of(se); 2864 if (cfs_rq->next != se) 2865 break; 2866 2867 cfs_rq->next = NULL; 2868 } 2869} 2870 2871static void __clear_buddies_skip(struct sched_entity *se) 2872{ 2873 for_each_sched_entity(se) { 2874 struct cfs_rq *cfs_rq = cfs_rq_of(se); 2875 if (cfs_rq->skip != se) 2876 break; 2877 2878 cfs_rq->skip = NULL; 2879 } 2880} 2881 2882static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se) 2883{ 2884 if (cfs_rq->last == se) 2885 __clear_buddies_last(se); 2886 2887 if (cfs_rq->next == se) 2888 __clear_buddies_next(se); 2889 2890 if (cfs_rq->skip == se) 2891 __clear_buddies_skip(se); 2892} 2893 2894static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq); 2895 2896static void 2897dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags) 2898{ 2899 /* 2900 * Update run-time statistics of the 'current'. 2901 */ 2902 update_curr(cfs_rq); 2903 dequeue_entity_load_avg(cfs_rq, se, flags & DEQUEUE_SLEEP); 2904 2905 update_stats_dequeue(cfs_rq, se); 2906 if (flags & DEQUEUE_SLEEP) { 2907#ifdef CONFIG_SCHEDSTATS 2908 if (entity_is_task(se)) { 2909 struct task_struct *tsk = task_of(se); 2910 2911 if (tsk->state & TASK_INTERRUPTIBLE) 2912 se->statistics.sleep_start = rq_clock(rq_of(cfs_rq)); 2913 if (tsk->state & TASK_UNINTERRUPTIBLE) 2914 se->statistics.block_start = rq_clock(rq_of(cfs_rq)); 2915 } 2916#endif 2917 } 2918 2919 clear_buddies(cfs_rq, se); 2920 2921 if (se != cfs_rq->curr) 2922 __dequeue_entity(cfs_rq, se); 2923 se->on_rq = 0; 2924 account_entity_dequeue(cfs_rq, se); 2925 2926 /* 2927 * Normalize the entity after updating the min_vruntime because the 2928 * update can refer to the ->curr item and we need to reflect this 2929 * movement in our normalized position. 2930 */ 2931 if (!(flags & DEQUEUE_SLEEP)) 2932 se->vruntime -= cfs_rq->min_vruntime; 2933 2934 /* return excess runtime on last dequeue */ 2935 return_cfs_rq_runtime(cfs_rq); 2936 2937 update_min_vruntime(cfs_rq); 2938 update_cfs_shares(cfs_rq); 2939} 2940 2941/* 2942 * Preempt the current task with a newly woken task if needed: 2943 */ 2944static void 2945check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr) 2946{ 2947 unsigned long ideal_runtime, delta_exec; 2948 struct sched_entity *se; 2949 s64 delta; 2950 2951 ideal_runtime = sched_slice(cfs_rq, curr); 2952 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime; 2953 if (delta_exec > ideal_runtime) { 2954 resched_curr(rq_of(cfs_rq)); 2955 /* 2956 * The current task ran long enough, ensure it doesn't get 2957 * re-elected due to buddy favours. 2958 */ 2959 clear_buddies(cfs_rq, curr); 2960 return; 2961 } 2962 2963 /* 2964 * Ensure that a task that missed wakeup preemption by a 2965 * narrow margin doesn't have to wait for a full slice. 2966 * This also mitigates buddy induced latencies under load. 2967 */ 2968 if (delta_exec < sysctl_sched_min_granularity) 2969 return; 2970 2971 se = __pick_first_entity(cfs_rq); 2972 delta = curr->vruntime - se->vruntime; 2973 2974 if (delta < 0) 2975 return; 2976 2977 if (delta > ideal_runtime) 2978 resched_curr(rq_of(cfs_rq)); 2979} 2980 2981static void 2982set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se) 2983{ 2984 /* 'current' is not kept within the tree. */ 2985 if (se->on_rq) { 2986 /* 2987 * Any task has to be enqueued before it get to execute on 2988 * a CPU. So account for the time it spent waiting on the 2989 * runqueue. 2990 */ 2991 update_stats_wait_end(cfs_rq, se); 2992 __dequeue_entity(cfs_rq, se); 2993 } 2994 2995 update_stats_curr_start(cfs_rq, se); 2996 cfs_rq->curr = se; 2997#ifdef CONFIG_SCHEDSTATS 2998 /* 2999 * Track our maximum slice length, if the CPU's load is at 3000 * least twice that of our own weight (i.e. dont track it 3001 * when there are only lesser-weight tasks around): 3002 */ 3003 if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) { 3004 se->statistics.slice_max = max(se->statistics.slice_max, 3005 se->sum_exec_runtime - se->prev_sum_exec_runtime); 3006 } 3007#endif 3008 se->prev_sum_exec_runtime = se->sum_exec_runtime; 3009} 3010 3011static int 3012wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se); 3013 3014/* 3015 * Pick the next process, keeping these things in mind, in this order: 3016 * 1) keep things fair between processes/task groups 3017 * 2) pick the "next" process, since someone really wants that to run 3018 * 3) pick the "last" process, for cache locality 3019 * 4) do not run the "skip" process, if something else is available 3020 */ 3021static struct sched_entity * 3022pick_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *curr) 3023{ 3024 struct sched_entity *left = __pick_first_entity(cfs_rq); 3025 struct sched_entity *se; 3026 3027 /* 3028 * If curr is set we have to see if its left of the leftmost entity 3029 * still in the tree, provided there was anything in the tree at all. 3030 */ 3031 if (!left || (curr && entity_before(curr, left))) 3032 left = curr; 3033 3034 se = left; /* ideally we run the leftmost entity */ 3035 3036 /* 3037 * Avoid running the skip buddy, if running something else can 3038 * be done without getting too unfair. 3039 */ 3040 if (cfs_rq->skip == se) { 3041 struct sched_entity *second; 3042 3043 if (se == curr) { 3044 second = __pick_first_entity(cfs_rq); 3045 } else { 3046 second = __pick_next_entity(se); 3047 if (!second || (curr && entity_before(curr, second))) 3048 second = curr; 3049 } 3050 3051 if (second && wakeup_preempt_entity(second, left) < 1) 3052 se = second; 3053 } 3054 3055 /* 3056 * Prefer last buddy, try to return the CPU to a preempted task. 3057 */ 3058 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1) 3059 se = cfs_rq->last; 3060 3061 /* 3062 * Someone really wants this to run. If it's not unfair, run it. 3063 */ 3064 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1) 3065 se = cfs_rq->next; 3066 3067 clear_buddies(cfs_rq, se); 3068 3069 return se; 3070} 3071 3072static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq); 3073 3074static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev) 3075{ 3076 /* 3077 * If still on the runqueue then deactivate_task() 3078 * was not called and update_curr() has to be done: 3079 */ 3080 if (prev->on_rq) 3081 update_curr(cfs_rq); 3082 3083 /* throttle cfs_rqs exceeding runtime */ 3084 check_cfs_rq_runtime(cfs_rq); 3085 3086 check_spread(cfs_rq, prev); 3087 if (prev->on_rq) { 3088 update_stats_wait_start(cfs_rq, prev); 3089 /* Put 'current' back into the tree. */ 3090 __enqueue_entity(cfs_rq, prev); 3091 /* in !on_rq case, update occurred at dequeue */ 3092 update_entity_load_avg(prev, 1); 3093 } 3094 cfs_rq->curr = NULL; 3095} 3096 3097static void 3098entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued) 3099{ 3100 /* 3101 * Update run-time statistics of the 'current'. 3102 */ 3103 update_curr(cfs_rq); 3104 3105 /* 3106 * Ensure that runnable average is periodically updated. 3107 */ 3108 update_entity_load_avg(curr, 1); 3109 update_cfs_rq_blocked_load(cfs_rq, 1); 3110 update_cfs_shares(cfs_rq); 3111 3112#ifdef CONFIG_SCHED_HRTICK 3113 /* 3114 * queued ticks are scheduled to match the slice, so don't bother 3115 * validating it and just reschedule. 3116 */ 3117 if (queued) { 3118 resched_curr(rq_of(cfs_rq)); 3119 return; 3120 } 3121 /* 3122 * don't let the period tick interfere with the hrtick preemption 3123 */ 3124 if (!sched_feat(DOUBLE_TICK) && 3125 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer)) 3126 return; 3127#endif 3128 3129 if (cfs_rq->nr_running > 1) 3130 check_preempt_tick(cfs_rq, curr); 3131} 3132 3133 3134/************************************************** 3135 * CFS bandwidth control machinery 3136 */ 3137 3138#ifdef CONFIG_CFS_BANDWIDTH 3139 3140#ifdef HAVE_JUMP_LABEL 3141static struct static_key __cfs_bandwidth_used; 3142 3143static inline bool cfs_bandwidth_used(void) 3144{ 3145 return static_key_false(&__cfs_bandwidth_used); 3146} 3147 3148void cfs_bandwidth_usage_inc(void) 3149{ 3150 static_key_slow_inc(&__cfs_bandwidth_used); 3151} 3152 3153void cfs_bandwidth_usage_dec(void) 3154{ 3155 static_key_slow_dec(&__cfs_bandwidth_used); 3156} 3157#else /* HAVE_JUMP_LABEL */ 3158static bool cfs_bandwidth_used(void) 3159{ 3160 return true; 3161} 3162 3163void cfs_bandwidth_usage_inc(void) {} 3164void cfs_bandwidth_usage_dec(void) {} 3165#endif /* HAVE_JUMP_LABEL */ 3166 3167/* 3168 * default period for cfs group bandwidth. 3169 * default: 0.1s, units: nanoseconds 3170 */ 3171static inline u64 default_cfs_period(void) 3172{ 3173 return 100000000ULL; 3174} 3175 3176static inline u64 sched_cfs_bandwidth_slice(void) 3177{ 3178 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC; 3179} 3180 3181/* 3182 * Replenish runtime according to assigned quota and update expiration time. 3183 * We use sched_clock_cpu directly instead of rq->clock to avoid adding 3184 * additional synchronization around rq->lock. 3185 * 3186 * requires cfs_b->lock 3187 */ 3188void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b) 3189{ 3190 u64 now; 3191 3192 if (cfs_b->quota == RUNTIME_INF) 3193 return; 3194 3195 now = sched_clock_cpu(smp_processor_id()); 3196 cfs_b->runtime = cfs_b->quota; 3197 cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period); 3198} 3199 3200static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg) 3201{ 3202 return &tg->cfs_bandwidth; 3203} 3204 3205/* rq->task_clock normalized against any time this cfs_rq has spent throttled */ 3206static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq) 3207{ 3208 if (unlikely(cfs_rq->throttle_count)) 3209 return cfs_rq->throttled_clock_task; 3210 3211 return rq_clock_task(rq_of(cfs_rq)) - cfs_rq->throttled_clock_task_time; 3212} 3213 3214/* returns 0 on failure to allocate runtime */ 3215static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq) 3216{ 3217 struct task_group *tg = cfs_rq->tg; 3218 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg); 3219 u64 amount = 0, min_amount, expires; 3220 3221 /* note: this is a positive sum as runtime_remaining <= 0 */ 3222 min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining; 3223 3224 raw_spin_lock(&cfs_b->lock); 3225 if (cfs_b->quota == RUNTIME_INF) 3226 amount = min_amount; 3227 else { 3228 /* 3229 * If the bandwidth pool has become inactive, then at least one 3230 * period must have elapsed since the last consumption. 3231 * Refresh the global state and ensure bandwidth timer becomes 3232 * active. 3233 */ 3234 if (!cfs_b->timer_active) { 3235 __refill_cfs_bandwidth_runtime(cfs_b); 3236 __start_cfs_bandwidth(cfs_b, false); 3237 } 3238 3239 if (cfs_b->runtime > 0) { 3240 amount = min(cfs_b->runtime, min_amount); 3241 cfs_b->runtime -= amount; 3242 cfs_b->idle = 0; 3243 } 3244 } 3245 expires = cfs_b->runtime_expires; 3246 raw_spin_unlock(&cfs_b->lock); 3247 3248 cfs_rq->runtime_remaining += amount; 3249 /* 3250 * we may have advanced our local expiration to account for allowed 3251 * spread between our sched_clock and the one on which runtime was 3252 * issued. 3253 */ 3254 if ((s64)(expires - cfs_rq->runtime_expires) > 0) 3255 cfs_rq->runtime_expires = expires; 3256 3257 return cfs_rq->runtime_remaining > 0; 3258} 3259 3260/* 3261 * Note: This depends on the synchronization provided by sched_clock and the 3262 * fact that rq->clock snapshots this value. 3263 */ 3264static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq) 3265{ 3266 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg); 3267 3268 /* if the deadline is ahead of our clock, nothing to do */ 3269 if (likely((s64)(rq_clock(rq_of(cfs_rq)) - cfs_rq->runtime_expires) < 0)) 3270 return; 3271 3272 if (cfs_rq->runtime_remaining < 0) 3273 return; 3274 3275 /* 3276 * If the local deadline has passed we have to consider the 3277 * possibility that our sched_clock is 'fast' and the global deadline 3278 * has not truly expired. 3279 * 3280 * Fortunately we can check determine whether this the case by checking 3281 * whether the global deadline has advanced. It is valid to compare 3282 * cfs_b->runtime_expires without any locks since we only care about 3283 * exact equality, so a partial write will still work. 3284 */ 3285 3286 if (cfs_rq->runtime_expires != cfs_b->runtime_expires) { 3287 /* extend local deadline, drift is bounded above by 2 ticks */ 3288 cfs_rq->runtime_expires += TICK_NSEC; 3289 } else { 3290 /* global deadline is ahead, expiration has passed */ 3291 cfs_rq->runtime_remaining = 0; 3292 } 3293} 3294 3295static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) 3296{ 3297 /* dock delta_exec before expiring quota (as it could span periods) */ 3298 cfs_rq->runtime_remaining -= delta_exec; 3299 expire_cfs_rq_runtime(cfs_rq); 3300 3301 if (likely(cfs_rq->runtime_remaining > 0)) 3302 return; 3303 3304 /* 3305 * if we're unable to extend our runtime we resched so that the active 3306 * hierarchy can be throttled 3307 */ 3308 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr)) 3309 resched_curr(rq_of(cfs_rq)); 3310} 3311 3312static __always_inline 3313void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) 3314{ 3315 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled) 3316 return; 3317 3318 __account_cfs_rq_runtime(cfs_rq, delta_exec); 3319} 3320 3321static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq) 3322{ 3323 return cfs_bandwidth_used() && cfs_rq->throttled; 3324} 3325 3326/* check whether cfs_rq, or any parent, is throttled */ 3327static inline int throttled_hierarchy(struct cfs_rq *cfs_rq) 3328{ 3329 return cfs_bandwidth_used() && cfs_rq->throttle_count; 3330} 3331 3332/* 3333 * Ensure that neither of the group entities corresponding to src_cpu or 3334 * dest_cpu are members of a throttled hierarchy when performing group 3335 * load-balance operations. 3336 */ 3337static inline int throttled_lb_pair(struct task_group *tg, 3338 int src_cpu, int dest_cpu) 3339{ 3340 struct cfs_rq *src_cfs_rq, *dest_cfs_rq; 3341 3342 src_cfs_rq = tg->cfs_rq[src_cpu]; 3343 dest_cfs_rq = tg->cfs_rq[dest_cpu]; 3344 3345 return throttled_hierarchy(src_cfs_rq) || 3346 throttled_hierarchy(dest_cfs_rq); 3347} 3348 3349/* updated child weight may affect parent so we have to do this bottom up */ 3350static int tg_unthrottle_up(struct task_group *tg, void *data) 3351{ 3352 struct rq *rq = data; 3353 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)]; 3354 3355 cfs_rq->throttle_count--; 3356#ifdef CONFIG_SMP 3357 if (!cfs_rq->throttle_count) { 3358 /* adjust cfs_rq_clock_task() */ 3359 cfs_rq->throttled_clock_task_time += rq_clock_task(rq) - 3360 cfs_rq->throttled_clock_task; 3361 } 3362#endif 3363 3364 return 0; 3365} 3366 3367static int tg_throttle_down(struct task_group *tg, void *data) 3368{ 3369 struct rq *rq = data; 3370 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)]; 3371 3372 /* group is entering throttled state, stop time */ 3373 if (!cfs_rq->throttle_count) 3374 cfs_rq->throttled_clock_task = rq_clock_task(rq); 3375 cfs_rq->throttle_count++; 3376 3377 return 0; 3378} 3379 3380static void throttle_cfs_rq(struct cfs_rq *cfs_rq) 3381{ 3382 struct rq *rq = rq_of(cfs_rq); 3383 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg); 3384 struct sched_entity *se; 3385 long task_delta, dequeue = 1; 3386 3387 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))]; 3388 3389 /* freeze hierarchy runnable averages while throttled */ 3390 rcu_read_lock(); 3391 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq); 3392 rcu_read_unlock(); 3393 3394 task_delta = cfs_rq->h_nr_running; 3395 for_each_sched_entity(se) { 3396 struct cfs_rq *qcfs_rq = cfs_rq_of(se); 3397 /* throttled entity or throttle-on-deactivate */ 3398 if (!se->on_rq) 3399 break; 3400 3401 if (dequeue) 3402 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP); 3403 qcfs_rq->h_nr_running -= task_delta; 3404 3405 if (qcfs_rq->load.weight) 3406 dequeue = 0; 3407 } 3408 3409 if (!se) 3410 sub_nr_running(rq, task_delta); 3411 3412 cfs_rq->throttled = 1; 3413 cfs_rq->throttled_clock = rq_clock(rq); 3414 raw_spin_lock(&cfs_b->lock); 3415 /* 3416 * Add to the _head_ of the list, so that an already-started 3417 * distribute_cfs_runtime will not see us 3418 */ 3419 list_add_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq); 3420 if (!cfs_b->timer_active) 3421 __start_cfs_bandwidth(cfs_b, false); 3422 raw_spin_unlock(&cfs_b->lock); 3423} 3424 3425void unthrottle_cfs_rq(struct cfs_rq *cfs_rq) 3426{ 3427 struct rq *rq = rq_of(cfs_rq); 3428 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg); 3429 struct sched_entity *se; 3430 int enqueue = 1; 3431 long task_delta; 3432 3433 se = cfs_rq->tg->se[cpu_of(rq)]; 3434 3435 cfs_rq->throttled = 0; 3436 3437 update_rq_clock(rq); 3438 3439 raw_spin_lock(&cfs_b->lock); 3440 cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock; 3441 list_del_rcu(&cfs_rq->throttled_list); 3442 raw_spin_unlock(&cfs_b->lock); 3443 3444 /* update hierarchical throttle state */ 3445 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq); 3446 3447 if (!cfs_rq->load.weight) 3448 return; 3449 3450 task_delta = cfs_rq->h_nr_running; 3451 for_each_sched_entity(se) { 3452 if (se->on_rq) 3453 enqueue = 0; 3454 3455 cfs_rq = cfs_rq_of(se); 3456 if (enqueue) 3457 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP); 3458 cfs_rq->h_nr_running += task_delta; 3459 3460 if (cfs_rq_throttled(cfs_rq)) 3461 break; 3462 } 3463 3464 if (!se) 3465 add_nr_running(rq, task_delta); 3466 3467 /* determine whether we need to wake up potentially idle cpu */ 3468 if (rq->curr == rq->idle && rq->cfs.nr_running) 3469 resched_curr(rq); 3470} 3471 3472static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b, 3473 u64 remaining, u64 expires) 3474{ 3475 struct cfs_rq *cfs_rq; 3476 u64 runtime; 3477 u64 starting_runtime = remaining; 3478 3479 rcu_read_lock(); 3480 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq, 3481 throttled_list) { 3482 struct rq *rq = rq_of(cfs_rq); 3483 3484 raw_spin_lock(&rq->lock); 3485 if (!cfs_rq_throttled(cfs_rq)) 3486 goto next; 3487 3488 runtime = -cfs_rq->runtime_remaining + 1; 3489 if (runtime > remaining) 3490 runtime = remaining; 3491 remaining -= runtime; 3492 3493 cfs_rq->runtime_remaining += runtime; 3494 cfs_rq->runtime_expires = expires; 3495 3496 /* we check whether we're throttled above */ 3497 if (cfs_rq->runtime_remaining > 0) 3498 unthrottle_cfs_rq(cfs_rq); 3499 3500next: 3501 raw_spin_unlock(&rq->lock); 3502 3503 if (!remaining) 3504 break; 3505 } 3506 rcu_read_unlock(); 3507 3508 return starting_runtime - remaining; 3509} 3510 3511/* 3512 * Responsible for refilling a task_group's bandwidth and unthrottling its 3513 * cfs_rqs as appropriate. If there has been no activity within the last 3514 * period the timer is deactivated until scheduling resumes; cfs_b->idle is 3515 * used to track this state. 3516 */ 3517static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun) 3518{ 3519 u64 runtime, runtime_expires; 3520 int throttled; 3521 3522 /* no need to continue the timer with no bandwidth constraint */ 3523 if (cfs_b->quota == RUNTIME_INF) 3524 goto out_deactivate; 3525 3526 throttled = !list_empty(&cfs_b->throttled_cfs_rq); 3527 cfs_b->nr_periods += overrun; 3528 3529 /* 3530 * idle depends on !throttled (for the case of a large deficit), and if 3531 * we're going inactive then everything else can be deferred 3532 */ 3533 if (cfs_b->idle && !throttled) 3534 goto out_deactivate; 3535 3536 /* 3537 * if we have relooped after returning idle once, we need to update our 3538 * status as actually running, so that other cpus doing 3539 * __start_cfs_bandwidth will stop trying to cancel us. 3540 */ 3541 cfs_b->timer_active = 1; 3542 3543 __refill_cfs_bandwidth_runtime(cfs_b); 3544 3545 if (!throttled) { 3546 /* mark as potentially idle for the upcoming period */ 3547 cfs_b->idle = 1; 3548 return 0; 3549 } 3550 3551 /* account preceding periods in which throttling occurred */ 3552 cfs_b->nr_throttled += overrun; 3553 3554 runtime_expires = cfs_b->runtime_expires; 3555 3556 /* 3557 * This check is repeated as we are holding onto the new bandwidth while 3558 * we unthrottle. This can potentially race with an unthrottled group 3559 * trying to acquire new bandwidth from the global pool. This can result 3560 * in us over-using our runtime if it is all used during this loop, but 3561 * only by limited amounts in that extreme case. 3562 */ 3563 while (throttled && cfs_b->runtime > 0) { 3564 runtime = cfs_b->runtime; 3565 raw_spin_unlock(&cfs_b->lock); 3566 /* we can't nest cfs_b->lock while distributing bandwidth */ 3567 runtime = distribute_cfs_runtime(cfs_b, runtime, 3568 runtime_expires); 3569 raw_spin_lock(&cfs_b->lock); 3570 3571 throttled = !list_empty(&cfs_b->throttled_cfs_rq); 3572 3573 cfs_b->runtime -= min(runtime, cfs_b->runtime); 3574 } 3575 3576 /* 3577 * While we are ensured activity in the period following an 3578 * unthrottle, this also covers the case in which the new bandwidth is 3579 * insufficient to cover the existing bandwidth deficit. (Forcing the 3580 * timer to remain active while there are any throttled entities.) 3581 */ 3582 cfs_b->idle = 0; 3583 3584 return 0; 3585 3586out_deactivate: 3587 cfs_b->timer_active = 0; 3588 return 1; 3589} 3590 3591/* a cfs_rq won't donate quota below this amount */ 3592static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC; 3593/* minimum remaining period time to redistribute slack quota */ 3594static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC; 3595/* how long we wait to gather additional slack before distributing */ 3596static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC; 3597 3598/* 3599 * Are we near the end of the current quota period? 3600 * 3601 * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the 3602 * hrtimer base being cleared by __hrtimer_start_range_ns. In the case of 3603 * migrate_hrtimers, base is never cleared, so we are fine. 3604 */ 3605static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire) 3606{ 3607 struct hrtimer *refresh_timer = &cfs_b->period_timer; 3608 u64 remaining; 3609 3610 /* if the call-back is running a quota refresh is already occurring */ 3611 if (hrtimer_callback_running(refresh_timer)) 3612 return 1; 3613 3614 /* is a quota refresh about to occur? */ 3615 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer)); 3616 if (remaining < min_expire) 3617 return 1; 3618 3619 return 0; 3620} 3621 3622static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b) 3623{ 3624 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration; 3625 3626 /* if there's a quota refresh soon don't bother with slack */ 3627 if (runtime_refresh_within(cfs_b, min_left)) 3628 return; 3629 3630 start_bandwidth_timer(&cfs_b->slack_timer, 3631 ns_to_ktime(cfs_bandwidth_slack_period)); 3632} 3633 3634/* we know any runtime found here is valid as update_curr() precedes return */ 3635static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq) 3636{ 3637 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg); 3638 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime; 3639 3640 if (slack_runtime <= 0) 3641 return; 3642 3643 raw_spin_lock(&cfs_b->lock); 3644 if (cfs_b->quota != RUNTIME_INF && 3645 cfs_rq->runtime_expires == cfs_b->runtime_expires) { 3646 cfs_b->runtime += slack_runtime; 3647 3648 /* we are under rq->lock, defer unthrottling using a timer */ 3649 if (cfs_b->runtime > sched_cfs_bandwidth_slice() && 3650 !list_empty(&cfs_b->throttled_cfs_rq)) 3651 start_cfs_slack_bandwidth(cfs_b); 3652 } 3653 raw_spin_unlock(&cfs_b->lock); 3654 3655 /* even if it's not valid for return we don't want to try again */ 3656 cfs_rq->runtime_remaining -= slack_runtime; 3657} 3658 3659static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) 3660{ 3661 if (!cfs_bandwidth_used()) 3662 return; 3663 3664 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running) 3665 return; 3666 3667 __return_cfs_rq_runtime(cfs_rq); 3668} 3669 3670/* 3671 * This is done with a timer (instead of inline with bandwidth return) since 3672 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs. 3673 */ 3674static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b) 3675{ 3676 u64 runtime = 0, slice = sched_cfs_bandwidth_slice(); 3677 u64 expires; 3678 3679 /* confirm we're still not at a refresh boundary */ 3680 raw_spin_lock(&cfs_b->lock); 3681 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) { 3682 raw_spin_unlock(&cfs_b->lock); 3683 return; 3684 } 3685 3686 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice) 3687 runtime = cfs_b->runtime; 3688 3689 expires = cfs_b->runtime_expires; 3690 raw_spin_unlock(&cfs_b->lock); 3691 3692 if (!runtime) 3693 return; 3694 3695 runtime = distribute_cfs_runtime(cfs_b, runtime, expires); 3696 3697 raw_spin_lock(&cfs_b->lock); 3698 if (expires == cfs_b->runtime_expires) 3699 cfs_b->runtime -= min(runtime, cfs_b->runtime); 3700 raw_spin_unlock(&cfs_b->lock); 3701} 3702 3703/* 3704 * When a group wakes up we want to make sure that its quota is not already 3705 * expired/exceeded, otherwise it may be allowed to steal additional ticks of 3706 * runtime as update_curr() throttling can not not trigger until it's on-rq. 3707 */ 3708static void check_enqueue_throttle(struct cfs_rq *cfs_rq) 3709{ 3710 if (!cfs_bandwidth_used()) 3711 return; 3712 3713 /* an active group must be handled by the update_curr()->put() path */ 3714 if (!cfs_rq->runtime_enabled || cfs_rq->curr) 3715 return; 3716 3717 /* ensure the group is not already throttled */ 3718 if (cfs_rq_throttled(cfs_rq)) 3719 return; 3720 3721 /* update runtime allocation */ 3722 account_cfs_rq_runtime(cfs_rq, 0); 3723 if (cfs_rq->runtime_remaining <= 0) 3724 throttle_cfs_rq(cfs_rq); 3725} 3726 3727/* conditionally throttle active cfs_rq's from put_prev_entity() */ 3728static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) 3729{ 3730 if (!cfs_bandwidth_used()) 3731 return false; 3732 3733 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0)) 3734 return false; 3735 3736 /* 3737 * it's possible for a throttled entity to be forced into a running 3738 * state (e.g. set_curr_task), in this case we're finished. 3739 */ 3740 if (cfs_rq_throttled(cfs_rq)) 3741 return true; 3742 3743 throttle_cfs_rq(cfs_rq); 3744 return true; 3745} 3746 3747static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer) 3748{ 3749 struct cfs_bandwidth *cfs_b = 3750 container_of(timer, struct cfs_bandwidth, slack_timer); 3751 do_sched_cfs_slack_timer(cfs_b); 3752 3753 return HRTIMER_NORESTART; 3754} 3755 3756static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer) 3757{ 3758 struct cfs_bandwidth *cfs_b = 3759 container_of(timer, struct cfs_bandwidth, period_timer); 3760 ktime_t now; 3761 int overrun; 3762 int idle = 0; 3763 3764 raw_spin_lock(&cfs_b->lock); 3765 for (;;) { 3766 now = hrtimer_cb_get_time(timer); 3767 overrun = hrtimer_forward(timer, now, cfs_b->period); 3768 3769 if (!overrun) 3770 break; 3771 3772 idle = do_sched_cfs_period_timer(cfs_b, overrun); 3773 } 3774 raw_spin_unlock(&cfs_b->lock); 3775 3776 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART; 3777} 3778 3779void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) 3780{ 3781 raw_spin_lock_init(&cfs_b->lock); 3782 cfs_b->runtime = 0; 3783 cfs_b->quota = RUNTIME_INF; 3784 cfs_b->period = ns_to_ktime(default_cfs_period()); 3785 3786 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq); 3787 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); 3788 cfs_b->period_timer.function = sched_cfs_period_timer; 3789 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); 3790 cfs_b->slack_timer.function = sched_cfs_slack_timer; 3791} 3792 3793static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) 3794{ 3795 cfs_rq->runtime_enabled = 0; 3796 INIT_LIST_HEAD(&cfs_rq->throttled_list); 3797} 3798 3799/* requires cfs_b->lock, may release to reprogram timer */ 3800void __start_cfs_bandwidth(struct cfs_bandwidth *cfs_b, bool force) 3801{ 3802 /* 3803 * The timer may be active because we're trying to set a new bandwidth 3804 * period or because we're racing with the tear-down path 3805 * (timer_active==0 becomes visible before the hrtimer call-back 3806 * terminates). In either case we ensure that it's re-programmed 3807 */ 3808 while (unlikely(hrtimer_active(&cfs_b->period_timer)) && 3809 hrtimer_try_to_cancel(&cfs_b->period_timer) < 0) { 3810 /* bounce the lock to allow do_sched_cfs_period_timer to run */ 3811 raw_spin_unlock(&cfs_b->lock); 3812 cpu_relax(); 3813 raw_spin_lock(&cfs_b->lock); 3814 /* if someone else restarted the timer then we're done */ 3815 if (!force && cfs_b->timer_active) 3816 return; 3817 } 3818 3819 cfs_b->timer_active = 1; 3820 start_bandwidth_timer(&cfs_b->period_timer, cfs_b->period); 3821} 3822 3823static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) 3824{ 3825 hrtimer_cancel(&cfs_b->period_timer); 3826 hrtimer_cancel(&cfs_b->slack_timer); 3827} 3828 3829static void __maybe_unused update_runtime_enabled(struct rq *rq) 3830{ 3831 struct cfs_rq *cfs_rq; 3832 3833 for_each_leaf_cfs_rq(rq, cfs_rq) { 3834 struct cfs_bandwidth *cfs_b = &cfs_rq->tg->cfs_bandwidth; 3835 3836 raw_spin_lock(&cfs_b->lock); 3837 cfs_rq->runtime_enabled = cfs_b->quota != RUNTIME_INF; 3838 raw_spin_unlock(&cfs_b->lock); 3839 } 3840} 3841 3842static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq) 3843{ 3844 struct cfs_rq *cfs_rq; 3845 3846 for_each_leaf_cfs_rq(rq, cfs_rq) { 3847 if (!cfs_rq->runtime_enabled) 3848 continue; 3849 3850 /* 3851 * clock_task is not advancing so we just need to make sure 3852 * there's some valid quota amount 3853 */ 3854 cfs_rq->runtime_remaining = 1; 3855 /* 3856 * Offline rq is schedulable till cpu is completely disabled 3857 * in take_cpu_down(), so we prevent new cfs throttling here. 3858 */ 3859 cfs_rq->runtime_enabled = 0; 3860 3861 if (cfs_rq_throttled(cfs_rq)) 3862 unthrottle_cfs_rq(cfs_rq); 3863 } 3864} 3865 3866#else /* CONFIG_CFS_BANDWIDTH */ 3867static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq) 3868{ 3869 return rq_clock_task(rq_of(cfs_rq)); 3870} 3871 3872static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) {} 3873static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) { return false; } 3874static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {} 3875static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {} 3876 3877static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq) 3878{ 3879 return 0; 3880} 3881 3882static inline int throttled_hierarchy(struct cfs_rq *cfs_rq) 3883{ 3884 return 0; 3885} 3886 3887static inline int throttled_lb_pair(struct task_group *tg, 3888 int src_cpu, int dest_cpu) 3889{ 3890 return 0; 3891} 3892 3893void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {} 3894 3895#ifdef CONFIG_FAIR_GROUP_SCHED 3896static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {} 3897#endif 3898 3899static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg) 3900{ 3901 return NULL; 3902} 3903static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {} 3904static inline void update_runtime_enabled(struct rq *rq) {} 3905static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {} 3906 3907#endif /* CONFIG_CFS_BANDWIDTH */ 3908 3909/************************************************** 3910 * CFS operations on tasks: 3911 */ 3912 3913#ifdef CONFIG_SCHED_HRTICK 3914static void hrtick_start_fair(struct rq *rq, struct task_struct *p) 3915{ 3916 struct sched_entity *se = &p->se; 3917 struct cfs_rq *cfs_rq = cfs_rq_of(se); 3918 3919 WARN_ON(task_rq(p) != rq); 3920 3921 if (cfs_rq->nr_running > 1) { 3922 u64 slice = sched_slice(cfs_rq, se); 3923 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime; 3924 s64 delta = slice - ran; 3925 3926 if (delta < 0) { 3927 if (rq->curr == p) 3928 resched_curr(rq); 3929 return; 3930 } 3931 hrtick_start(rq, delta); 3932 } 3933} 3934 3935/* 3936 * called from enqueue/dequeue and updates the hrtick when the 3937 * current task is from our class and nr_running is low enough 3938 * to matter. 3939 */ 3940static void hrtick_update(struct rq *rq) 3941{ 3942 struct task_struct *curr = rq->curr; 3943 3944 if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class) 3945 return; 3946 3947 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency) 3948 hrtick_start_fair(rq, curr); 3949} 3950#else /* !CONFIG_SCHED_HRTICK */ 3951static inline void 3952hrtick_start_fair(struct rq *rq, struct task_struct *p) 3953{ 3954} 3955 3956static inline void hrtick_update(struct rq *rq) 3957{ 3958} 3959#endif 3960 3961/* 3962 * The enqueue_task method is called before nr_running is 3963 * increased. Here we update the fair scheduling stats and 3964 * then put the task into the rbtree: 3965 */ 3966static void 3967enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags) 3968{ 3969 struct cfs_rq *cfs_rq; 3970 struct sched_entity *se = &p->se; 3971 3972 for_each_sched_entity(se) { 3973 if (se->on_rq) 3974 break; 3975 cfs_rq = cfs_rq_of(se); 3976 enqueue_entity(cfs_rq, se, flags); 3977 3978 /* 3979 * end evaluation on encountering a throttled cfs_rq 3980 * 3981 * note: in the case of encountering a throttled cfs_rq we will 3982 * post the final h_nr_running increment below. 3983 */ 3984 if (cfs_rq_throttled(cfs_rq)) 3985 break; 3986 cfs_rq->h_nr_running++; 3987 3988 flags = ENQUEUE_WAKEUP; 3989 } 3990 3991 for_each_sched_entity(se) { 3992 cfs_rq = cfs_rq_of(se); 3993 cfs_rq->h_nr_running++; 3994 3995 if (cfs_rq_throttled(cfs_rq)) 3996 break; 3997 3998 update_cfs_shares(cfs_rq); 3999 update_entity_load_avg(se, 1); 4000 } 4001 4002 if (!se) { 4003 update_rq_runnable_avg(rq, rq->nr_running); 4004 add_nr_running(rq, 1); 4005 } 4006 hrtick_update(rq); 4007} 4008 4009static void set_next_buddy(struct sched_entity *se); 4010 4011/* 4012 * The dequeue_task method is called before nr_running is 4013 * decreased. We remove the task from the rbtree and 4014 * update the fair scheduling stats: 4015 */ 4016static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags) 4017{ 4018 struct cfs_rq *cfs_rq; 4019 struct sched_entity *se = &p->se; 4020 int task_sleep = flags & DEQUEUE_SLEEP; 4021 4022 for_each_sched_entity(se) { 4023 cfs_rq = cfs_rq_of(se); 4024 dequeue_entity(cfs_rq, se, flags); 4025 4026 /* 4027 * end evaluation on encountering a throttled cfs_rq 4028 * 4029 * note: in the case of encountering a throttled cfs_rq we will 4030 * post the final h_nr_running decrement below. 4031 */ 4032 if (cfs_rq_throttled(cfs_rq)) 4033 break; 4034 cfs_rq->h_nr_running--; 4035 4036 /* Don't dequeue parent if it has other entities besides us */ 4037 if (cfs_rq->load.weight) { 4038 /* 4039 * Bias pick_next to pick a task from this cfs_rq, as 4040 * p is sleeping when it is within its sched_slice. 4041 */ 4042 if (task_sleep && parent_entity(se)) 4043 set_next_buddy(parent_entity(se)); 4044 4045 /* avoid re-evaluating load for this entity */ 4046 se = parent_entity(se); 4047 break; 4048 } 4049 flags |= DEQUEUE_SLEEP; 4050 } 4051 4052 for_each_sched_entity(se) { 4053 cfs_rq = cfs_rq_of(se); 4054 cfs_rq->h_nr_running--; 4055 4056 if (cfs_rq_throttled(cfs_rq)) 4057 break; 4058 4059 update_cfs_shares(cfs_rq); 4060 update_entity_load_avg(se, 1); 4061 } 4062 4063 if (!se) { 4064 sub_nr_running(rq, 1); 4065 update_rq_runnable_avg(rq, 1); 4066 } 4067 hrtick_update(rq); 4068} 4069 4070#ifdef CONFIG_SMP 4071/* Used instead of source_load when we know the type == 0 */ 4072static unsigned long weighted_cpuload(const int cpu) 4073{ 4074 return cpu_rq(cpu)->cfs.runnable_load_avg; 4075} 4076 4077/* 4078 * Return a low guess at the load of a migration-source cpu weighted 4079 * according to the scheduling class and "nice" value. 4080 * 4081 * We want to under-estimate the load of migration sources, to 4082 * balance conservatively. 4083 */ 4084static unsigned long source_load(int cpu, int type) 4085{ 4086 struct rq *rq = cpu_rq(cpu); 4087 unsigned long total = weighted_cpuload(cpu); 4088 4089 if (type == 0 || !sched_feat(LB_BIAS)) 4090 return total; 4091 4092 return min(rq->cpu_load[type-1], total); 4093} 4094 4095/* 4096 * Return a high guess at the load of a migration-target cpu weighted 4097 * according to the scheduling class and "nice" value. 4098 */ 4099static unsigned long target_load(int cpu, int type) 4100{ 4101 struct rq *rq = cpu_rq(cpu); 4102 unsigned long total = weighted_cpuload(cpu); 4103 4104 if (type == 0 || !sched_feat(LB_BIAS)) 4105 return total; 4106 4107 return max(rq->cpu_load[type-1], total); 4108} 4109 4110static unsigned long capacity_of(int cpu) 4111{ 4112 return cpu_rq(cpu)->cpu_capacity; 4113} 4114 4115static unsigned long cpu_avg_load_per_task(int cpu) 4116{ 4117 struct rq *rq = cpu_rq(cpu); 4118 unsigned long nr_running = ACCESS_ONCE(rq->cfs.h_nr_running); 4119 unsigned long load_avg = rq->cfs.runnable_load_avg; 4120 4121 if (nr_running) 4122 return load_avg / nr_running; 4123 4124 return 0; 4125} 4126 4127static void record_wakee(struct task_struct *p) 4128{ 4129 /* 4130 * Rough decay (wiping) for cost saving, don't worry 4131 * about the boundary, really active task won't care 4132 * about the loss. 4133 */ 4134 if (time_after(jiffies, current->wakee_flip_decay_ts + HZ)) { 4135 current->wakee_flips >>= 1; 4136 current->wakee_flip_decay_ts = jiffies; 4137 } 4138 4139 if (current->last_wakee != p) { 4140 current->last_wakee = p; 4141 current->wakee_flips++; 4142 } 4143} 4144 4145static void task_waking_fair(struct task_struct *p) 4146{ 4147 struct sched_entity *se = &p->se; 4148 struct cfs_rq *cfs_rq = cfs_rq_of(se); 4149 u64 min_vruntime; 4150 4151#ifndef CONFIG_64BIT 4152 u64 min_vruntime_copy; 4153 4154 do { 4155 min_vruntime_copy = cfs_rq->min_vruntime_copy; 4156 smp_rmb(); 4157 min_vruntime = cfs_rq->min_vruntime; 4158 } while (min_vruntime != min_vruntime_copy); 4159#else 4160 min_vruntime = cfs_rq->min_vruntime; 4161#endif 4162 4163 se->vruntime -= min_vruntime; 4164 record_wakee(p); 4165} 4166 4167#ifdef CONFIG_FAIR_GROUP_SCHED 4168/* 4169 * effective_load() calculates the load change as seen from the root_task_group 4170 * 4171 * Adding load to a group doesn't make a group heavier, but can cause movement 4172 * of group shares between cpus. Assuming the shares were perfectly aligned one 4173 * can calculate the shift in shares. 4174 * 4175 * Calculate the effective load difference if @wl is added (subtracted) to @tg 4176 * on this @cpu and results in a total addition (subtraction) of @wg to the 4177 * total group weight. 4178 * 4179 * Given a runqueue weight distribution (rw_i) we can compute a shares 4180 * distribution (s_i) using: 4181 * 4182 * s_i = rw_i / \Sum rw_j (1) 4183 * 4184 * Suppose we have 4 CPUs and our @tg is a direct child of the root group and 4185 * has 7 equal weight tasks, distributed as below (rw_i), with the resulting 4186 * shares distribution (s_i): 4187 * 4188 * rw_i = { 2, 4, 1, 0 } 4189 * s_i = { 2/7, 4/7, 1/7, 0 } 4190 * 4191 * As per wake_affine() we're interested in the load of two CPUs (the CPU the 4192 * task used to run on and the CPU the waker is running on), we need to 4193 * compute the effect of waking a task on either CPU and, in case of a sync 4194 * wakeup, compute the effect of the current task going to sleep. 4195 * 4196 * So for a change of @wl to the local @cpu with an overall group weight change 4197 * of @wl we can compute the new shares distribution (s'_i) using: 4198 * 4199 * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2) 4200 * 4201 * Suppose we're interested in CPUs 0 and 1, and want to compute the load 4202 * differences in waking a task to CPU 0. The additional task changes the 4203 * weight and shares distributions like: 4204 * 4205 * rw'_i = { 3, 4, 1, 0 } 4206 * s'_i = { 3/8, 4/8, 1/8, 0 } 4207 * 4208 * We can then compute the difference in effective weight by using: 4209 * 4210 * dw_i = S * (s'_i - s_i) (3) 4211 * 4212 * Where 'S' is the group weight as seen by its parent. 4213 * 4214 * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7) 4215 * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 - 4216 * 4/7) times the weight of the group. 4217 */ 4218static long effective_load(struct task_group *tg, int cpu, long wl, long wg) 4219{ 4220 struct sched_entity *se = tg->se[cpu]; 4221 4222 if (!tg->parent) /* the trivial, non-cgroup case */ 4223 return wl; 4224 4225 for_each_sched_entity(se) { 4226 long w, W; 4227 4228 tg = se->my_q->tg; 4229 4230 /* 4231 * W = @wg + \Sum rw_j 4232 */ 4233 W = wg + calc_tg_weight(tg, se->my_q); 4234 4235 /* 4236 * w = rw_i + @wl 4237 */ 4238 w = se->my_q->load.weight + wl; 4239 4240 /* 4241 * wl = S * s'_i; see (2) 4242 */ 4243 if (W > 0 && w < W) 4244 wl = (w * tg->shares) / W; 4245 else 4246 wl = tg->shares; 4247 4248 /* 4249 * Per the above, wl is the new se->load.weight value; since 4250 * those are clipped to [MIN_SHARES, ...) do so now. See 4251 * calc_cfs_shares(). 4252 */ 4253 if (wl < MIN_SHARES) 4254 wl = MIN_SHARES; 4255 4256 /* 4257 * wl = dw_i = S * (s'_i - s_i); see (3) 4258 */ 4259 wl -= se->load.weight; 4260 4261 /* 4262 * Recursively apply this logic to all parent groups to compute 4263 * the final effective load change on the root group. Since 4264 * only the @tg group gets extra weight, all parent groups can 4265 * only redistribute existing shares. @wl is the shift in shares 4266 * resulting from this level per the above. 4267 */ 4268 wg = 0; 4269 } 4270 4271 return wl; 4272} 4273#else 4274 4275static long effective_load(struct task_group *tg, int cpu, long wl, long wg) 4276{ 4277 return wl; 4278} 4279 4280#endif 4281 4282static int wake_wide(struct task_struct *p) 4283{ 4284 int factor = this_cpu_read(sd_llc_size); 4285 4286 /* 4287 * Yeah, it's the switching-frequency, could means many wakee or 4288 * rapidly switch, use factor here will just help to automatically 4289 * adjust the loose-degree, so bigger node will lead to more pull. 4290 */ 4291 if (p->wakee_flips > factor) { 4292 /* 4293 * wakee is somewhat hot, it needs certain amount of cpu 4294 * resource, so if waker is far more hot, prefer to leave 4295 * it alone. 4296 */ 4297 if (current->wakee_flips > (factor * p->wakee_flips)) 4298 return 1; 4299 } 4300 4301 return 0; 4302} 4303 4304static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync) 4305{ 4306 s64 this_load, load; 4307 s64 this_eff_load, prev_eff_load; 4308 int idx, this_cpu, prev_cpu; 4309 struct task_group *tg; 4310 unsigned long weight; 4311 int balanced; 4312 4313 /* 4314 * If we wake multiple tasks be careful to not bounce 4315 * ourselves around too much. 4316 */ 4317 if (wake_wide(p)) 4318 return 0; 4319 4320 idx = sd->wake_idx; 4321 this_cpu = smp_processor_id(); 4322 prev_cpu = task_cpu(p); 4323 load = source_load(prev_cpu, idx); 4324 this_load = target_load(this_cpu, idx); 4325 4326 /* 4327 * If sync wakeup then subtract the (maximum possible) 4328 * effect of the currently running task from the load 4329 * of the current CPU: 4330 */ 4331 if (sync) { 4332 tg = task_group(current); 4333 weight = current->se.load.weight; 4334 4335 this_load += effective_load(tg, this_cpu, -weight, -weight); 4336 load += effective_load(tg, prev_cpu, 0, -weight); 4337 } 4338 4339 tg = task_group(p); 4340 weight = p->se.load.weight; 4341 4342 /* 4343 * In low-load situations, where prev_cpu is idle and this_cpu is idle 4344 * due to the sync cause above having dropped this_load to 0, we'll 4345 * always have an imbalance, but there's really nothing you can do 4346 * about that, so that's good too. 4347 * 4348 * Otherwise check if either cpus are near enough in load to allow this 4349 * task to be woken on this_cpu. 4350 */ 4351 this_eff_load = 100; 4352 this_eff_load *= capacity_of(prev_cpu); 4353 4354 prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2; 4355 prev_eff_load *= capacity_of(this_cpu); 4356 4357 if (this_load > 0) { 4358 this_eff_load *= this_load + 4359 effective_load(tg, this_cpu, weight, weight); 4360 4361 prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight); 4362 } 4363 4364 balanced = this_eff_load <= prev_eff_load; 4365 4366 schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts); 4367 4368 if (!balanced) 4369 return 0; 4370 4371 schedstat_inc(sd, ttwu_move_affine); 4372 schedstat_inc(p, se.statistics.nr_wakeups_affine); 4373 4374 return 1; 4375} 4376 4377/* 4378 * find_idlest_group finds and returns the least busy CPU group within the 4379 * domain. 4380 */ 4381static struct sched_group * 4382find_idlest_group(struct sched_domain *sd, struct task_struct *p, 4383 int this_cpu, int sd_flag) 4384{ 4385 struct sched_group *idlest = NULL, *group = sd->groups; 4386 unsigned long min_load = ULONG_MAX, this_load = 0; 4387 int load_idx = sd->forkexec_idx; 4388 int imbalance = 100 + (sd->imbalance_pct-100)/2; 4389 4390 if (sd_flag & SD_BALANCE_WAKE) 4391 load_idx = sd->wake_idx; 4392 4393 do { 4394 unsigned long load, avg_load; 4395 int local_group; 4396 int i; 4397 4398 /* Skip over this group if it has no CPUs allowed */ 4399 if (!cpumask_intersects(sched_group_cpus(group), 4400 tsk_cpus_allowed(p))) 4401 continue; 4402 4403 local_group = cpumask_test_cpu(this_cpu, 4404 sched_group_cpus(group)); 4405 4406 /* Tally up the load of all CPUs in the group */ 4407 avg_load = 0; 4408 4409 for_each_cpu(i, sched_group_cpus(group)) { 4410 /* Bias balancing toward cpus of our domain */ 4411 if (local_group) 4412 load = source_load(i, load_idx); 4413 else 4414 load = target_load(i, load_idx); 4415 4416 avg_load += load; 4417 } 4418 4419 /* Adjust by relative CPU capacity of the group */ 4420 avg_load = (avg_load * SCHED_CAPACITY_SCALE) / group->sgc->capacity; 4421 4422 if (local_group) { 4423 this_load = avg_load; 4424 } else if (avg_load < min_load) { 4425 min_load = avg_load; 4426 idlest = group; 4427 } 4428 } while (group = group->next, group != sd->groups); 4429 4430 if (!idlest || 100*this_load < imbalance*min_load) 4431 return NULL; 4432 return idlest; 4433} 4434 4435/* 4436 * find_idlest_cpu - find the idlest cpu among the cpus in group. 4437 */ 4438static int 4439find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu) 4440{ 4441 unsigned long load, min_load = ULONG_MAX; 4442 unsigned int min_exit_latency = UINT_MAX; 4443 u64 latest_idle_timestamp = 0; 4444 int least_loaded_cpu = this_cpu; 4445 int shallowest_idle_cpu = -1; 4446 int i; 4447 4448 /* Traverse only the allowed CPUs */ 4449 for_each_cpu_and(i, sched_group_cpus(group), tsk_cpus_allowed(p)) { 4450 if (idle_cpu(i)) { 4451 struct rq *rq = cpu_rq(i); 4452 struct cpuidle_state *idle = idle_get_state(rq); 4453 if (idle && idle->exit_latency < min_exit_latency) { 4454 /* 4455 * We give priority to a CPU whose idle state 4456 * has the smallest exit latency irrespective 4457 * of any idle timestamp. 4458 */ 4459 min_exit_latency = idle->exit_latency; 4460 latest_idle_timestamp = rq->idle_stamp; 4461 shallowest_idle_cpu = i; 4462 } else if ((!idle || idle->exit_latency == min_exit_latency) && 4463 rq->idle_stamp > latest_idle_timestamp) { 4464 /* 4465 * If equal or no active idle state, then 4466 * the most recently idled CPU might have 4467 * a warmer cache. 4468 */ 4469 latest_idle_timestamp = rq->idle_stamp; 4470 shallowest_idle_cpu = i; 4471 } 4472 } else { 4473 load = weighted_cpuload(i); 4474 if (load < min_load || (load == min_load && i == this_cpu)) { 4475 min_load = load; 4476 least_loaded_cpu = i; 4477 } 4478 } 4479 } 4480 4481 return shallowest_idle_cpu != -1 ? shallowest_idle_cpu : least_loaded_cpu; 4482} 4483 4484/* 4485 * Try and locate an idle CPU in the sched_domain. 4486 */ 4487static int select_idle_sibling(struct task_struct *p, int target) 4488{ 4489 struct sched_domain *sd; 4490 struct sched_group *sg; 4491 int i = task_cpu(p); 4492 4493 if (idle_cpu(target)) 4494 return target; 4495 4496 /* 4497 * If the prevous cpu is cache affine and idle, don't be stupid. 4498 */ 4499 if (i != target && cpus_share_cache(i, target) && idle_cpu(i)) 4500 return i; 4501 4502 /* 4503 * Otherwise, iterate the domains and find an elegible idle cpu. 4504 */ 4505 sd = rcu_dereference(per_cpu(sd_llc, target)); 4506 for_each_lower_domain(sd) { 4507 sg = sd->groups; 4508 do { 4509 if (!cpumask_intersects(sched_group_cpus(sg), 4510 tsk_cpus_allowed(p))) 4511 goto next; 4512 4513 for_each_cpu(i, sched_group_cpus(sg)) { 4514 if (i == target || !idle_cpu(i)) 4515 goto next; 4516 } 4517 4518 target = cpumask_first_and(sched_group_cpus(sg), 4519 tsk_cpus_allowed(p)); 4520 goto done; 4521next: 4522 sg = sg->next; 4523 } while (sg != sd->groups); 4524 } 4525done: 4526 return target; 4527} 4528 4529/* 4530 * select_task_rq_fair: Select target runqueue for the waking task in domains 4531 * that have the 'sd_flag' flag set. In practice, this is SD_BALANCE_WAKE, 4532 * SD_BALANCE_FORK, or SD_BALANCE_EXEC. 4533 * 4534 * Balances load by selecting the idlest cpu in the idlest group, or under 4535 * certain conditions an idle sibling cpu if the domain has SD_WAKE_AFFINE set. 4536 * 4537 * Returns the target cpu number. 4538 * 4539 * preempt must be disabled. 4540 */ 4541static int 4542select_task_rq_fair(struct task_struct *p, int prev_cpu, int sd_flag, int wake_flags) 4543{ 4544 struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL; 4545 int cpu = smp_processor_id(); 4546 int new_cpu = cpu; 4547 int want_affine = 0; 4548 int sync = wake_flags & WF_SYNC; 4549 4550 if (p->nr_cpus_allowed == 1) 4551 return prev_cpu; 4552 4553 if (sd_flag & SD_BALANCE_WAKE) 4554 want_affine = cpumask_test_cpu(cpu, tsk_cpus_allowed(p)); 4555 4556 rcu_read_lock(); 4557 for_each_domain(cpu, tmp) { 4558 if (!(tmp->flags & SD_LOAD_BALANCE)) 4559 continue; 4560 4561 /* 4562 * If both cpu and prev_cpu are part of this domain, 4563 * cpu is a valid SD_WAKE_AFFINE target. 4564 */ 4565 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) && 4566 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) { 4567 affine_sd = tmp; 4568 break; 4569 } 4570 4571 if (tmp->flags & sd_flag) 4572 sd = tmp; 4573 } 4574 4575 if (affine_sd && cpu != prev_cpu && wake_affine(affine_sd, p, sync)) 4576 prev_cpu = cpu; 4577 4578 if (sd_flag & SD_BALANCE_WAKE) { 4579 new_cpu = select_idle_sibling(p, prev_cpu); 4580 goto unlock; 4581 } 4582 4583 while (sd) { 4584 struct sched_group *group; 4585 int weight; 4586 4587 if (!(sd->flags & sd_flag)) { 4588 sd = sd->child; 4589 continue; 4590 } 4591 4592 group = find_idlest_group(sd, p, cpu, sd_flag); 4593 if (!group) { 4594 sd = sd->child; 4595 continue; 4596 } 4597 4598 new_cpu = find_idlest_cpu(group, p, cpu); 4599 if (new_cpu == -1 || new_cpu == cpu) { 4600 /* Now try balancing at a lower domain level of cpu */ 4601 sd = sd->child; 4602 continue; 4603 } 4604 4605 /* Now try balancing at a lower domain level of new_cpu */ 4606 cpu = new_cpu; 4607 weight = sd->span_weight; 4608 sd = NULL; 4609 for_each_domain(cpu, tmp) { 4610 if (weight <= tmp->span_weight) 4611 break; 4612 if (tmp->flags & sd_flag) 4613 sd = tmp; 4614 } 4615 /* while loop will break here if sd == NULL */ 4616 } 4617unlock: 4618 rcu_read_unlock(); 4619 4620 return new_cpu; 4621} 4622 4623/* 4624 * Called immediately before a task is migrated to a new cpu; task_cpu(p) and 4625 * cfs_rq_of(p) references at time of call are still valid and identify the 4626 * previous cpu. However, the caller only guarantees p->pi_lock is held; no 4627 * other assumptions, including the state of rq->lock, should be made. 4628 */ 4629static void 4630migrate_task_rq_fair(struct task_struct *p, int next_cpu) 4631{ 4632 struct sched_entity *se = &p->se; 4633 struct cfs_rq *cfs_rq = cfs_rq_of(se); 4634 4635 /* 4636 * Load tracking: accumulate removed load so that it can be processed 4637 * when we next update owning cfs_rq under rq->lock. Tasks contribute 4638 * to blocked load iff they have a positive decay-count. It can never 4639 * be negative here since on-rq tasks have decay-count == 0. 4640 */ 4641 if (se->avg.decay_count) { 4642 se->avg.decay_count = -__synchronize_entity_decay(se); 4643 atomic_long_add(se->avg.load_avg_contrib, 4644 &cfs_rq->removed_load); 4645 } 4646 4647 /* We have migrated, no longer consider this task hot */ 4648 se->exec_start = 0; 4649} 4650#endif /* CONFIG_SMP */ 4651 4652static unsigned long 4653wakeup_gran(struct sched_entity *curr, struct sched_entity *se) 4654{ 4655 unsigned long gran = sysctl_sched_wakeup_granularity; 4656 4657 /* 4658 * Since its curr running now, convert the gran from real-time 4659 * to virtual-time in his units. 4660 * 4661 * By using 'se' instead of 'curr' we penalize light tasks, so 4662 * they get preempted easier. That is, if 'se' < 'curr' then 4663 * the resulting gran will be larger, therefore penalizing the 4664 * lighter, if otoh 'se' > 'curr' then the resulting gran will 4665 * be smaller, again penalizing the lighter task. 4666 * 4667 * This is especially important for buddies when the leftmost 4668 * task is higher priority than the buddy. 4669 */ 4670 return calc_delta_fair(gran, se); 4671} 4672 4673/* 4674 * Should 'se' preempt 'curr'. 4675 * 4676 * |s1 4677 * |s2 4678 * |s3 4679 * g 4680 * |<--->|c 4681 * 4682 * w(c, s1) = -1 4683 * w(c, s2) = 0 4684 * w(c, s3) = 1 4685 * 4686 */ 4687static int 4688wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se) 4689{ 4690 s64 gran, vdiff = curr->vruntime - se->vruntime; 4691 4692 if (vdiff <= 0) 4693 return -1; 4694 4695 gran = wakeup_gran(curr, se); 4696 if (vdiff > gran) 4697 return 1; 4698 4699 return 0; 4700} 4701 4702static void set_last_buddy(struct sched_entity *se) 4703{ 4704 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE)) 4705 return; 4706 4707 for_each_sched_entity(se) 4708 cfs_rq_of(se)->last = se; 4709} 4710 4711static void set_next_buddy(struct sched_entity *se) 4712{ 4713 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE)) 4714 return; 4715 4716 for_each_sched_entity(se) 4717 cfs_rq_of(se)->next = se; 4718} 4719 4720static void set_skip_buddy(struct sched_entity *se) 4721{ 4722 for_each_sched_entity(se) 4723 cfs_rq_of(se)->skip = se; 4724} 4725 4726/* 4727 * Preempt the current task with a newly woken task if needed: 4728 */ 4729static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags) 4730{ 4731 struct task_struct *curr = rq->curr; 4732 struct sched_entity *se = &curr->se, *pse = &p->se; 4733 struct cfs_rq *cfs_rq = task_cfs_rq(curr); 4734 int scale = cfs_rq->nr_running >= sched_nr_latency; 4735 int next_buddy_marked = 0; 4736 4737 if (unlikely(se == pse)) 4738 return; 4739 4740 /* 4741 * This is possible from callers such as attach_tasks(), in which we 4742 * unconditionally check_prempt_curr() after an enqueue (which may have 4743 * lead to a throttle). This both saves work and prevents false 4744 * next-buddy nomination below. 4745 */ 4746 if (unlikely(throttled_hierarchy(cfs_rq_of(pse)))) 4747 return; 4748 4749 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) { 4750 set_next_buddy(pse); 4751 next_buddy_marked = 1; 4752 } 4753 4754 /* 4755 * We can come here with TIF_NEED_RESCHED already set from new task 4756 * wake up path. 4757 * 4758 * Note: this also catches the edge-case of curr being in a throttled 4759 * group (e.g. via set_curr_task), since update_curr() (in the 4760 * enqueue of curr) will have resulted in resched being set. This 4761 * prevents us from potentially nominating it as a false LAST_BUDDY 4762 * below. 4763 */ 4764 if (test_tsk_need_resched(curr)) 4765 return; 4766 4767 /* Idle tasks are by definition preempted by non-idle tasks. */ 4768 if (unlikely(curr->policy == SCHED_IDLE) && 4769 likely(p->policy != SCHED_IDLE)) 4770 goto preempt; 4771 4772 /* 4773 * Batch and idle tasks do not preempt non-idle tasks (their preemption 4774 * is driven by the tick): 4775 */ 4776 if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION)) 4777 return; 4778 4779 find_matching_se(&se, &pse); 4780 update_curr(cfs_rq_of(se)); 4781 BUG_ON(!pse); 4782 if (wakeup_preempt_entity(se, pse) == 1) { 4783 /* 4784 * Bias pick_next to pick the sched entity that is 4785 * triggering this preemption. 4786 */ 4787 if (!next_buddy_marked) 4788 set_next_buddy(pse); 4789 goto preempt; 4790 } 4791 4792 return; 4793 4794preempt: 4795 resched_curr(rq); 4796 /* 4797 * Only set the backward buddy when the current task is still 4798 * on the rq. This can happen when a wakeup gets interleaved 4799 * with schedule on the ->pre_schedule() or idle_balance() 4800 * point, either of which can * drop the rq lock. 4801 * 4802 * Also, during early boot the idle thread is in the fair class, 4803 * for obvious reasons its a bad idea to schedule back to it. 4804 */ 4805 if (unlikely(!se->on_rq || curr == rq->idle)) 4806 return; 4807 4808 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se)) 4809 set_last_buddy(se); 4810} 4811 4812static struct task_struct * 4813pick_next_task_fair(struct rq *rq, struct task_struct *prev) 4814{ 4815 struct cfs_rq *cfs_rq = &rq->cfs; 4816 struct sched_entity *se; 4817 struct task_struct *p; 4818 int new_tasks; 4819 4820again: 4821#ifdef CONFIG_FAIR_GROUP_SCHED 4822 if (!cfs_rq->nr_running) 4823 goto idle; 4824 4825 if (prev->sched_class != &fair_sched_class) 4826 goto simple; 4827 4828 /* 4829 * Because of the set_next_buddy() in dequeue_task_fair() it is rather 4830 * likely that a next task is from the same cgroup as the current. 4831 * 4832 * Therefore attempt to avoid putting and setting the entire cgroup 4833 * hierarchy, only change the part that actually changes. 4834 */ 4835 4836 do { 4837 struct sched_entity *curr = cfs_rq->curr; 4838 4839 /* 4840 * Since we got here without doing put_prev_entity() we also 4841 * have to consider cfs_rq->curr. If it is still a runnable 4842 * entity, update_curr() will update its vruntime, otherwise 4843 * forget we've ever seen it. 4844 */ 4845 if (curr && curr->on_rq) 4846 update_curr(cfs_rq); 4847 else 4848 curr = NULL; 4849 4850 /* 4851 * This call to check_cfs_rq_runtime() will do the throttle and 4852 * dequeue its entity in the parent(s). Therefore the 'simple' 4853 * nr_running test will indeed be correct. 4854 */ 4855 if (unlikely(check_cfs_rq_runtime(cfs_rq))) 4856 goto simple; 4857 4858 se = pick_next_entity(cfs_rq, curr); 4859 cfs_rq = group_cfs_rq(se); 4860 } while (cfs_rq); 4861 4862 p = task_of(se); 4863 4864 /* 4865 * Since we haven't yet done put_prev_entity and if the selected task 4866 * is a different task than we started out with, try and touch the 4867 * least amount of cfs_rqs. 4868 */ 4869 if (prev != p) { 4870 struct sched_entity *pse = &prev->se; 4871 4872 while (!(cfs_rq = is_same_group(se, pse))) { 4873 int se_depth = se->depth; 4874 int pse_depth = pse->depth; 4875 4876 if (se_depth <= pse_depth) { 4877 put_prev_entity(cfs_rq_of(pse), pse); 4878 pse = parent_entity(pse); 4879 } 4880 if (se_depth >= pse_depth) { 4881 set_next_entity(cfs_rq_of(se), se); 4882 se = parent_entity(se); 4883 } 4884 } 4885 4886 put_prev_entity(cfs_rq, pse); 4887 set_next_entity(cfs_rq, se); 4888 } 4889 4890 if (hrtick_enabled(rq)) 4891 hrtick_start_fair(rq, p); 4892 4893 return p; 4894simple: 4895 cfs_rq = &rq->cfs; 4896#endif 4897 4898 if (!cfs_rq->nr_running) 4899 goto idle; 4900 4901 put_prev_task(rq, prev); 4902 4903 do { 4904 se = pick_next_entity(cfs_rq, NULL); 4905 set_next_entity(cfs_rq, se); 4906 cfs_rq = group_cfs_rq(se); 4907 } while (cfs_rq); 4908 4909 p = task_of(se); 4910 4911 if (hrtick_enabled(rq)) 4912 hrtick_start_fair(rq, p); 4913 4914 return p; 4915 4916idle: 4917 new_tasks = idle_balance(rq); 4918 /* 4919 * Because idle_balance() releases (and re-acquires) rq->lock, it is 4920 * possible for any higher priority task to appear. In that case we 4921 * must re-start the pick_next_entity() loop. 4922 */ 4923 if (new_tasks < 0) 4924 return RETRY_TASK; 4925 4926 if (new_tasks > 0) 4927 goto again; 4928 4929 return NULL; 4930} 4931 4932/* 4933 * Account for a descheduled task: 4934 */ 4935static void put_prev_task_fair(struct rq *rq, struct task_struct *prev) 4936{ 4937 struct sched_entity *se = &prev->se; 4938 struct cfs_rq *cfs_rq; 4939 4940 for_each_sched_entity(se) { 4941 cfs_rq = cfs_rq_of(se); 4942 put_prev_entity(cfs_rq, se); 4943 } 4944} 4945 4946/* 4947 * sched_yield() is very simple 4948 * 4949 * The magic of dealing with the ->skip buddy is in pick_next_entity. 4950 */ 4951static void yield_task_fair(struct rq *rq) 4952{ 4953 struct task_struct *curr = rq->curr; 4954 struct cfs_rq *cfs_rq = task_cfs_rq(curr); 4955 struct sched_entity *se = &curr->se; 4956 4957 /* 4958 * Are we the only task in the tree? 4959 */ 4960 if (unlikely(rq->nr_running == 1)) 4961 return; 4962 4963 clear_buddies(cfs_rq, se); 4964 4965 if (curr->policy != SCHED_BATCH) { 4966 update_rq_clock(rq); 4967 /* 4968 * Update run-time statistics of the 'current'. 4969 */ 4970 update_curr(cfs_rq); 4971 /* 4972 * Tell update_rq_clock() that we've just updated, 4973 * so we don't do microscopic update in schedule() 4974 * and double the fastpath cost. 4975 */ 4976 rq->skip_clock_update = 1; 4977 } 4978 4979 set_skip_buddy(se); 4980} 4981 4982static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt) 4983{ 4984 struct sched_entity *se = &p->se; 4985 4986 /* throttled hierarchies are not runnable */ 4987 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se))) 4988 return false; 4989 4990 /* Tell the scheduler that we'd really like pse to run next. */ 4991 set_next_buddy(se); 4992 4993 yield_task_fair(rq); 4994 4995 return true; 4996} 4997 4998#ifdef CONFIG_SMP 4999/************************************************** 5000 * Fair scheduling class load-balancing methods. 5001 * 5002 * BASICS 5003 * 5004 * The purpose of load-balancing is to achieve the same basic fairness the 5005 * per-cpu scheduler provides, namely provide a proportional amount of compute 5006 * time to each task. This is expressed in the following equation: 5007 * 5008 * W_i,n/P_i == W_j,n/P_j for all i,j (1) 5009 * 5010 * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight 5011 * W_i,0 is defined as: 5012 * 5013 * W_i,0 = \Sum_j w_i,j (2) 5014 * 5015 * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight 5016 * is derived from the nice value as per prio_to_weight[]. 5017 * 5018 * The weight average is an exponential decay average of the instantaneous 5019 * weight: 5020 * 5021 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3) 5022 * 5023 * C_i is the compute capacity of cpu i, typically it is the 5024 * fraction of 'recent' time available for SCHED_OTHER task execution. But it 5025 * can also include other factors [XXX]. 5026 * 5027 * To achieve this balance we define a measure of imbalance which follows 5028 * directly from (1): 5029 * 5030 * imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j } (4) 5031 * 5032 * We them move tasks around to minimize the imbalance. In the continuous 5033 * function space it is obvious this converges, in the discrete case we get 5034 * a few fun cases generally called infeasible weight scenarios. 5035 * 5036 * [XXX expand on: 5037 * - infeasible weights; 5038 * - local vs global optima in the discrete case. ] 5039 * 5040 * 5041 * SCHED DOMAINS 5042 * 5043 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2) 5044 * for all i,j solution, we create a tree of cpus that follows the hardware 5045 * topology where each level pairs two lower groups (or better). This results 5046 * in O(log n) layers. Furthermore we reduce the number of cpus going up the 5047 * tree to only the first of the previous level and we decrease the frequency 5048 * of load-balance at each level inv. proportional to the number of cpus in 5049 * the groups. 5050 * 5051 * This yields: 5052 * 5053 * log_2 n 1 n 5054 * \Sum { --- * --- * 2^i } = O(n) (5) 5055 * i = 0 2^i 2^i 5056 * `- size of each group 5057 * | | `- number of cpus doing load-balance 5058 * | `- freq 5059 * `- sum over all levels 5060 * 5061 * Coupled with a limit on how many tasks we can migrate every balance pass, 5062 * this makes (5) the runtime complexity of the balancer. 5063 * 5064 * An important property here is that each CPU is still (indirectly) connected 5065 * to every other cpu in at most O(log n) steps: 5066 * 5067 * The adjacency matrix of the resulting graph is given by: 5068 * 5069 * log_2 n 5070 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6) 5071 * k = 0 5072 * 5073 * And you'll find that: 5074 * 5075 * A^(log_2 n)_i,j != 0 for all i,j (7) 5076 * 5077 * Showing there's indeed a path between every cpu in at most O(log n) steps. 5078 * The task movement gives a factor of O(m), giving a convergence complexity 5079 * of: 5080 * 5081 * O(nm log n), n := nr_cpus, m := nr_tasks (8) 5082 * 5083 * 5084 * WORK CONSERVING 5085 * 5086 * In order to avoid CPUs going idle while there's still work to do, new idle 5087 * balancing is more aggressive and has the newly idle cpu iterate up the domain 5088 * tree itself instead of relying on other CPUs to bring it work. 5089 * 5090 * This adds some complexity to both (5) and (8) but it reduces the total idle 5091 * time. 5092 * 5093 * [XXX more?] 5094 * 5095 * 5096 * CGROUPS 5097 * 5098 * Cgroups make a horror show out of (2), instead of a simple sum we get: 5099 * 5100 * s_k,i 5101 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9) 5102 * S_k 5103 * 5104 * Where 5105 * 5106 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10) 5107 * 5108 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i. 5109 * 5110 * The big problem is S_k, its a global sum needed to compute a local (W_i) 5111 * property. 5112 * 5113 * [XXX write more on how we solve this.. _after_ merging pjt's patches that 5114 * rewrite all of this once again.] 5115 */ 5116 5117static unsigned long __read_mostly max_load_balance_interval = HZ/10; 5118 5119enum fbq_type { regular, remote, all }; 5120 5121#define LBF_ALL_PINNED 0x01 5122#define LBF_NEED_BREAK 0x02 5123#define LBF_DST_PINNED 0x04 5124#define LBF_SOME_PINNED 0x08 5125 5126struct lb_env { 5127 struct sched_domain *sd; 5128 5129 struct rq *src_rq; 5130 int src_cpu; 5131 5132 int dst_cpu; 5133 struct rq *dst_rq; 5134 5135 struct cpumask *dst_grpmask; 5136 int new_dst_cpu; 5137 enum cpu_idle_type idle; 5138 long imbalance; 5139 /* The set of CPUs under consideration for load-balancing */ 5140 struct cpumask *cpus; 5141 5142 unsigned int flags; 5143 5144 unsigned int loop; 5145 unsigned int loop_break; 5146 unsigned int loop_max; 5147 5148 enum fbq_type fbq_type; 5149 struct list_head tasks; 5150}; 5151 5152/* 5153 * Is this task likely cache-hot: 5154 */ 5155static int task_hot(struct task_struct *p, struct lb_env *env) 5156{ 5157 s64 delta; 5158 5159 lockdep_assert_held(&env->src_rq->lock); 5160 5161 if (p->sched_class != &fair_sched_class) 5162 return 0; 5163 5164 if (unlikely(p->policy == SCHED_IDLE)) 5165 return 0; 5166 5167 /* 5168 * Buddy candidates are cache hot: 5169 */ 5170 if (sched_feat(CACHE_HOT_BUDDY) && env->dst_rq->nr_running && 5171 (&p->se == cfs_rq_of(&p->se)->next || 5172 &p->se == cfs_rq_of(&p->se)->last)) 5173 return 1; 5174 5175 if (sysctl_sched_migration_cost == -1) 5176 return 1; 5177 if (sysctl_sched_migration_cost == 0) 5178 return 0; 5179 5180 delta = rq_clock_task(env->src_rq) - p->se.exec_start; 5181 5182 return delta < (s64)sysctl_sched_migration_cost; 5183} 5184 5185#ifdef CONFIG_NUMA_BALANCING 5186/* Returns true if the destination node has incurred more faults */ 5187static bool migrate_improves_locality(struct task_struct *p, struct lb_env *env) 5188{ 5189 struct numa_group *numa_group = rcu_dereference(p->numa_group); 5190 int src_nid, dst_nid; 5191 5192 if (!sched_feat(NUMA_FAVOUR_HIGHER) || !p->numa_faults_memory || 5193 !(env->sd->flags & SD_NUMA)) { 5194 return false; 5195 } 5196 5197 src_nid = cpu_to_node(env->src_cpu); 5198 dst_nid = cpu_to_node(env->dst_cpu); 5199 5200 if (src_nid == dst_nid) 5201 return false; 5202 5203 if (numa_group) { 5204 /* Task is already in the group's interleave set. */ 5205 if (node_isset(src_nid, numa_group->active_nodes)) 5206 return false; 5207 5208 /* Task is moving into the group's interleave set. */ 5209 if (node_isset(dst_nid, numa_group->active_nodes)) 5210 return true; 5211 5212 return group_faults(p, dst_nid) > group_faults(p, src_nid); 5213 } 5214 5215 /* Encourage migration to the preferred node. */ 5216 if (dst_nid == p->numa_preferred_nid) 5217 return true; 5218 5219 return task_faults(p, dst_nid) > task_faults(p, src_nid); 5220} 5221 5222 5223static bool migrate_degrades_locality(struct task_struct *p, struct lb_env *env) 5224{ 5225 struct numa_group *numa_group = rcu_dereference(p->numa_group); 5226 int src_nid, dst_nid; 5227 5228 if (!sched_feat(NUMA) || !sched_feat(NUMA_RESIST_LOWER)) 5229 return false; 5230 5231 if (!p->numa_faults_memory || !(env->sd->flags & SD_NUMA)) 5232 return false; 5233 5234 src_nid = cpu_to_node(env->src_cpu); 5235 dst_nid = cpu_to_node(env->dst_cpu); 5236 5237 if (src_nid == dst_nid) 5238 return false; 5239 5240 if (numa_group) { 5241 /* Task is moving within/into the group's interleave set. */ 5242 if (node_isset(dst_nid, numa_group->active_nodes)) 5243 return false; 5244 5245 /* Task is moving out of the group's interleave set. */ 5246 if (node_isset(src_nid, numa_group->active_nodes)) 5247 return true; 5248 5249 return group_faults(p, dst_nid) < group_faults(p, src_nid); 5250 } 5251 5252 /* Migrating away from the preferred node is always bad. */ 5253 if (src_nid == p->numa_preferred_nid) 5254 return true; 5255 5256 return task_faults(p, dst_nid) < task_faults(p, src_nid); 5257} 5258 5259#else 5260static inline bool migrate_improves_locality(struct task_struct *p, 5261 struct lb_env *env) 5262{ 5263 return false; 5264} 5265 5266static inline bool migrate_degrades_locality(struct task_struct *p, 5267 struct lb_env *env) 5268{ 5269 return false; 5270} 5271#endif 5272 5273/* 5274 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu? 5275 */ 5276static 5277int can_migrate_task(struct task_struct *p, struct lb_env *env) 5278{ 5279 int tsk_cache_hot = 0; 5280 5281 lockdep_assert_held(&env->src_rq->lock); 5282 5283 /* 5284 * We do not migrate tasks that are: 5285 * 1) throttled_lb_pair, or 5286 * 2) cannot be migrated to this CPU due to cpus_allowed, or 5287 * 3) running (obviously), or 5288 * 4) are cache-hot on their current CPU. 5289 */ 5290 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu)) 5291 return 0; 5292 5293 if (!cpumask_test_cpu(env->dst_cpu, tsk_cpus_allowed(p))) { 5294 int cpu; 5295 5296 schedstat_inc(p, se.statistics.nr_failed_migrations_affine); 5297 5298 env->flags |= LBF_SOME_PINNED; 5299 5300 /* 5301 * Remember if this task can be migrated to any other cpu in 5302 * our sched_group. We may want to revisit it if we couldn't 5303 * meet load balance goals by pulling other tasks on src_cpu. 5304 * 5305 * Also avoid computing new_dst_cpu if we have already computed 5306 * one in current iteration. 5307 */ 5308 if (!env->dst_grpmask || (env->flags & LBF_DST_PINNED)) 5309 return 0; 5310 5311 /* Prevent to re-select dst_cpu via env's cpus */ 5312 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) { 5313 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) { 5314 env->flags |= LBF_DST_PINNED; 5315 env->new_dst_cpu = cpu; 5316 break; 5317 } 5318 } 5319 5320 return 0; 5321 } 5322 5323 /* Record that we found atleast one task that could run on dst_cpu */ 5324 env->flags &= ~LBF_ALL_PINNED; 5325 5326 if (task_running(env->src_rq, p)) { 5327 schedstat_inc(p, se.statistics.nr_failed_migrations_running); 5328 return 0; 5329 } 5330 5331 /* 5332 * Aggressive migration if: 5333 * 1) destination numa is preferred 5334 * 2) task is cache cold, or 5335 * 3) too many balance attempts have failed. 5336 */ 5337 tsk_cache_hot = task_hot(p, env); 5338 if (!tsk_cache_hot) 5339 tsk_cache_hot = migrate_degrades_locality(p, env); 5340 5341 if (migrate_improves_locality(p, env) || !tsk_cache_hot || 5342 env->sd->nr_balance_failed > env->sd->cache_nice_tries) { 5343 if (tsk_cache_hot) { 5344 schedstat_inc(env->sd, lb_hot_gained[env->idle]); 5345 schedstat_inc(p, se.statistics.nr_forced_migrations); 5346 } 5347 return 1; 5348 } 5349 5350 schedstat_inc(p, se.statistics.nr_failed_migrations_hot); 5351 return 0; 5352} 5353 5354/* 5355 * detach_task() -- detach the task for the migration specified in env 5356 */ 5357static void detach_task(struct task_struct *p, struct lb_env *env) 5358{ 5359 lockdep_assert_held(&env->src_rq->lock); 5360 5361 deactivate_task(env->src_rq, p, 0); 5362 p->on_rq = TASK_ON_RQ_MIGRATING; 5363 set_task_cpu(p, env->dst_cpu); 5364} 5365 5366/* 5367 * detach_one_task() -- tries to dequeue exactly one task from env->src_rq, as 5368 * part of active balancing operations within "domain". 5369 * 5370 * Returns a task if successful and NULL otherwise. 5371 */ 5372static struct task_struct *detach_one_task(struct lb_env *env) 5373{ 5374 struct task_struct *p, *n; 5375 5376 lockdep_assert_held(&env->src_rq->lock); 5377 5378 list_for_each_entry_safe(p, n, &env->src_rq->cfs_tasks, se.group_node) { 5379 if (!can_migrate_task(p, env)) 5380 continue; 5381 5382 detach_task(p, env); 5383 5384 /* 5385 * Right now, this is only the second place where 5386 * lb_gained[env->idle] is updated (other is detach_tasks) 5387 * so we can safely collect stats here rather than 5388 * inside detach_tasks(). 5389 */ 5390 schedstat_inc(env->sd, lb_gained[env->idle]); 5391 return p; 5392 } 5393 return NULL; 5394} 5395 5396static const unsigned int sched_nr_migrate_break = 32; 5397 5398/* 5399 * detach_tasks() -- tries to detach up to imbalance weighted load from 5400 * busiest_rq, as part of a balancing operation within domain "sd". 5401 * 5402 * Returns number of detached tasks if successful and 0 otherwise. 5403 */ 5404static int detach_tasks(struct lb_env *env) 5405{ 5406 struct list_head *tasks = &env->src_rq->cfs_tasks; 5407 struct task_struct *p; 5408 unsigned long load; 5409 int detached = 0; 5410 5411 lockdep_assert_held(&env->src_rq->lock); 5412 5413 if (env->imbalance <= 0) 5414 return 0; 5415 5416 while (!list_empty(tasks)) { 5417 p = list_first_entry(tasks, struct task_struct, se.group_node); 5418 5419 env->loop++; 5420 /* We've more or less seen every task there is, call it quits */ 5421 if (env->loop > env->loop_max) 5422 break; 5423 5424 /* take a breather every nr_migrate tasks */ 5425 if (env->loop > env->loop_break) { 5426 env->loop_break += sched_nr_migrate_break; 5427 env->flags |= LBF_NEED_BREAK; 5428 break; 5429 } 5430 5431 if (!can_migrate_task(p, env)) 5432 goto next; 5433 5434 load = task_h_load(p); 5435 5436 if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed) 5437 goto next; 5438 5439 if ((load / 2) > env->imbalance) 5440 goto next; 5441 5442 detach_task(p, env); 5443 list_add(&p->se.group_node, &env->tasks); 5444 5445 detached++; 5446 env->imbalance -= load; 5447 5448#ifdef CONFIG_PREEMPT 5449 /* 5450 * NEWIDLE balancing is a source of latency, so preemptible 5451 * kernels will stop after the first task is detached to minimize 5452 * the critical section. 5453 */ 5454 if (env->idle == CPU_NEWLY_IDLE) 5455 break; 5456#endif 5457 5458 /* 5459 * We only want to steal up to the prescribed amount of 5460 * weighted load. 5461 */ 5462 if (env->imbalance <= 0) 5463 break; 5464 5465 continue; 5466next: 5467 list_move_tail(&p->se.group_node, tasks); 5468 } 5469 5470 /* 5471 * Right now, this is one of only two places we collect this stat 5472 * so we can safely collect detach_one_task() stats here rather 5473 * than inside detach_one_task(). 5474 */ 5475 schedstat_add(env->sd, lb_gained[env->idle], detached); 5476 5477 return detached; 5478} 5479 5480/* 5481 * attach_task() -- attach the task detached by detach_task() to its new rq. 5482 */ 5483static void attach_task(struct rq *rq, struct task_struct *p) 5484{ 5485 lockdep_assert_held(&rq->lock); 5486 5487 BUG_ON(task_rq(p) != rq); 5488 p->on_rq = TASK_ON_RQ_QUEUED; 5489 activate_task(rq, p, 0); 5490 check_preempt_curr(rq, p, 0); 5491} 5492 5493/* 5494 * attach_one_task() -- attaches the task returned from detach_one_task() to 5495 * its new rq. 5496 */ 5497static void attach_one_task(struct rq *rq, struct task_struct *p) 5498{ 5499 raw_spin_lock(&rq->lock); 5500 attach_task(rq, p); 5501 raw_spin_unlock(&rq->lock); 5502} 5503 5504/* 5505 * attach_tasks() -- attaches all tasks detached by detach_tasks() to their 5506 * new rq. 5507 */ 5508static void attach_tasks(struct lb_env *env) 5509{ 5510 struct list_head *tasks = &env->tasks; 5511 struct task_struct *p; 5512 5513 raw_spin_lock(&env->dst_rq->lock); 5514 5515 while (!list_empty(tasks)) { 5516 p = list_first_entry(tasks, struct task_struct, se.group_node); 5517 list_del_init(&p->se.group_node); 5518 5519 attach_task(env->dst_rq, p); 5520 } 5521 5522 raw_spin_unlock(&env->dst_rq->lock); 5523} 5524 5525#ifdef CONFIG_FAIR_GROUP_SCHED 5526/* 5527 * update tg->load_weight by folding this cpu's load_avg 5528 */ 5529static void __update_blocked_averages_cpu(struct task_group *tg, int cpu) 5530{ 5531 struct sched_entity *se = tg->se[cpu]; 5532 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu]; 5533 5534 /* throttled entities do not contribute to load */ 5535 if (throttled_hierarchy(cfs_rq)) 5536 return; 5537 5538 update_cfs_rq_blocked_load(cfs_rq, 1); 5539 5540 if (se) { 5541 update_entity_load_avg(se, 1); 5542 /* 5543 * We pivot on our runnable average having decayed to zero for 5544 * list removal. This generally implies that all our children 5545 * have also been removed (modulo rounding error or bandwidth 5546 * control); however, such cases are rare and we can fix these 5547 * at enqueue. 5548 * 5549 * TODO: fix up out-of-order children on enqueue. 5550 */ 5551 if (!se->avg.runnable_avg_sum && !cfs_rq->nr_running) 5552 list_del_leaf_cfs_rq(cfs_rq); 5553 } else { 5554 struct rq *rq = rq_of(cfs_rq); 5555 update_rq_runnable_avg(rq, rq->nr_running); 5556 } 5557} 5558 5559static void update_blocked_averages(int cpu) 5560{ 5561 struct rq *rq = cpu_rq(cpu); 5562 struct cfs_rq *cfs_rq; 5563 unsigned long flags; 5564 5565 raw_spin_lock_irqsave(&rq->lock, flags); 5566 update_rq_clock(rq); 5567 /* 5568 * Iterates the task_group tree in a bottom up fashion, see 5569 * list_add_leaf_cfs_rq() for details. 5570 */ 5571 for_each_leaf_cfs_rq(rq, cfs_rq) { 5572 /* 5573 * Note: We may want to consider periodically releasing 5574 * rq->lock about these updates so that creating many task 5575 * groups does not result in continually extending hold time. 5576 */ 5577 __update_blocked_averages_cpu(cfs_rq->tg, rq->cpu); 5578 } 5579 5580 raw_spin_unlock_irqrestore(&rq->lock, flags); 5581} 5582 5583/* 5584 * Compute the hierarchical load factor for cfs_rq and all its ascendants. 5585 * This needs to be done in a top-down fashion because the load of a child 5586 * group is a fraction of its parents load. 5587 */ 5588static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq) 5589{ 5590 struct rq *rq = rq_of(cfs_rq); 5591 struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)]; 5592 unsigned long now = jiffies; 5593 unsigned long load; 5594 5595 if (cfs_rq->last_h_load_update == now) 5596 return; 5597 5598 cfs_rq->h_load_next = NULL; 5599 for_each_sched_entity(se) { 5600 cfs_rq = cfs_rq_of(se); 5601 cfs_rq->h_load_next = se; 5602 if (cfs_rq->last_h_load_update == now) 5603 break; 5604 } 5605 5606 if (!se) { 5607 cfs_rq->h_load = cfs_rq->runnable_load_avg; 5608 cfs_rq->last_h_load_update = now; 5609 } 5610 5611 while ((se = cfs_rq->h_load_next) != NULL) { 5612 load = cfs_rq->h_load; 5613 load = div64_ul(load * se->avg.load_avg_contrib, 5614 cfs_rq->runnable_load_avg + 1); 5615 cfs_rq = group_cfs_rq(se); 5616 cfs_rq->h_load = load; 5617 cfs_rq->last_h_load_update = now; 5618 } 5619} 5620 5621static unsigned long task_h_load(struct task_struct *p) 5622{ 5623 struct cfs_rq *cfs_rq = task_cfs_rq(p); 5624 5625 update_cfs_rq_h_load(cfs_rq); 5626 return div64_ul(p->se.avg.load_avg_contrib * cfs_rq->h_load, 5627 cfs_rq->runnable_load_avg + 1); 5628} 5629#else 5630static inline void update_blocked_averages(int cpu) 5631{ 5632} 5633 5634static unsigned long task_h_load(struct task_struct *p) 5635{ 5636 return p->se.avg.load_avg_contrib; 5637} 5638#endif 5639 5640/********** Helpers for find_busiest_group ************************/ 5641 5642enum group_type { 5643 group_other = 0, 5644 group_imbalanced, 5645 group_overloaded, 5646}; 5647 5648/* 5649 * sg_lb_stats - stats of a sched_group required for load_balancing 5650 */ 5651struct sg_lb_stats { 5652 unsigned long avg_load; /*Avg load across the CPUs of the group */ 5653 unsigned long group_load; /* Total load over the CPUs of the group */ 5654 unsigned long sum_weighted_load; /* Weighted load of group's tasks */ 5655 unsigned long load_per_task; 5656 unsigned long group_capacity; 5657 unsigned int sum_nr_running; /* Nr tasks running in the group */ 5658 unsigned int group_capacity_factor; 5659 unsigned int idle_cpus; 5660 unsigned int group_weight; 5661 enum group_type group_type; 5662 int group_has_free_capacity; 5663#ifdef CONFIG_NUMA_BALANCING 5664 unsigned int nr_numa_running; 5665 unsigned int nr_preferred_running; 5666#endif 5667}; 5668 5669/* 5670 * sd_lb_stats - Structure to store the statistics of a sched_domain 5671 * during load balancing. 5672 */ 5673struct sd_lb_stats { 5674 struct sched_group *busiest; /* Busiest group in this sd */ 5675 struct sched_group *local; /* Local group in this sd */ 5676 unsigned long total_load; /* Total load of all groups in sd */ 5677 unsigned long total_capacity; /* Total capacity of all groups in sd */ 5678 unsigned long avg_load; /* Average load across all groups in sd */ 5679 5680 struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */ 5681 struct sg_lb_stats local_stat; /* Statistics of the local group */ 5682}; 5683 5684static inline void init_sd_lb_stats(struct sd_lb_stats *sds) 5685{ 5686 /* 5687 * Skimp on the clearing to avoid duplicate work. We can avoid clearing 5688 * local_stat because update_sg_lb_stats() does a full clear/assignment. 5689 * We must however clear busiest_stat::avg_load because 5690 * update_sd_pick_busiest() reads this before assignment. 5691 */ 5692 *sds = (struct sd_lb_stats){ 5693 .busiest = NULL, 5694 .local = NULL, 5695 .total_load = 0UL, 5696 .total_capacity = 0UL, 5697 .busiest_stat = { 5698 .avg_load = 0UL, 5699 .sum_nr_running = 0, 5700 .group_type = group_other, 5701 }, 5702 }; 5703} 5704 5705/** 5706 * get_sd_load_idx - Obtain the load index for a given sched domain. 5707 * @sd: The sched_domain whose load_idx is to be obtained. 5708 * @idle: The idle status of the CPU for whose sd load_idx is obtained. 5709 * 5710 * Return: The load index. 5711 */ 5712static inline int get_sd_load_idx(struct sched_domain *sd, 5713 enum cpu_idle_type idle) 5714{ 5715 int load_idx; 5716 5717 switch (idle) { 5718 case CPU_NOT_IDLE: 5719 load_idx = sd->busy_idx; 5720 break; 5721 5722 case CPU_NEWLY_IDLE: 5723 load_idx = sd->newidle_idx; 5724 break; 5725 default: 5726 load_idx = sd->idle_idx; 5727 break; 5728 } 5729 5730 return load_idx; 5731} 5732 5733static unsigned long default_scale_capacity(struct sched_domain *sd, int cpu) 5734{ 5735 return SCHED_CAPACITY_SCALE; 5736} 5737 5738unsigned long __weak arch_scale_freq_capacity(struct sched_domain *sd, int cpu) 5739{ 5740 return default_scale_capacity(sd, cpu); 5741} 5742 5743static unsigned long default_scale_cpu_capacity(struct sched_domain *sd, int cpu) 5744{ 5745 if ((sd->flags & SD_SHARE_CPUCAPACITY) && (sd->span_weight > 1)) 5746 return sd->smt_gain / sd->span_weight; 5747 5748 return SCHED_CAPACITY_SCALE; 5749} 5750 5751unsigned long __weak arch_scale_cpu_capacity(struct sched_domain *sd, int cpu) 5752{ 5753 return default_scale_cpu_capacity(sd, cpu); 5754} 5755 5756static unsigned long scale_rt_capacity(int cpu) 5757{ 5758 struct rq *rq = cpu_rq(cpu); 5759 u64 total, available, age_stamp, avg; 5760 s64 delta; 5761 5762 /* 5763 * Since we're reading these variables without serialization make sure 5764 * we read them once before doing sanity checks on them. 5765 */ 5766 age_stamp = ACCESS_ONCE(rq->age_stamp); 5767 avg = ACCESS_ONCE(rq->rt_avg); 5768 5769 delta = rq_clock(rq) - age_stamp; 5770 if (unlikely(delta < 0)) 5771 delta = 0; 5772 5773 total = sched_avg_period() + delta; 5774 5775 if (unlikely(total < avg)) { 5776 /* Ensures that capacity won't end up being negative */ 5777 available = 0; 5778 } else { 5779 available = total - avg; 5780 } 5781 5782 if (unlikely((s64)total < SCHED_CAPACITY_SCALE)) 5783 total = SCHED_CAPACITY_SCALE; 5784 5785 total >>= SCHED_CAPACITY_SHIFT; 5786 5787 return div_u64(available, total); 5788} 5789 5790static void update_cpu_capacity(struct sched_domain *sd, int cpu) 5791{ 5792 unsigned long capacity = SCHED_CAPACITY_SCALE; 5793 struct sched_group *sdg = sd->groups; 5794 5795 if (sched_feat(ARCH_CAPACITY)) 5796 capacity *= arch_scale_cpu_capacity(sd, cpu); 5797 else 5798 capacity *= default_scale_cpu_capacity(sd, cpu); 5799 5800 capacity >>= SCHED_CAPACITY_SHIFT; 5801 5802 sdg->sgc->capacity_orig = capacity; 5803 5804 if (sched_feat(ARCH_CAPACITY)) 5805 capacity *= arch_scale_freq_capacity(sd, cpu); 5806 else 5807 capacity *= default_scale_capacity(sd, cpu); 5808 5809 capacity >>= SCHED_CAPACITY_SHIFT; 5810 5811 capacity *= scale_rt_capacity(cpu); 5812 capacity >>= SCHED_CAPACITY_SHIFT; 5813 5814 if (!capacity) 5815 capacity = 1; 5816 5817 cpu_rq(cpu)->cpu_capacity = capacity; 5818 sdg->sgc->capacity = capacity; 5819} 5820 5821void update_group_capacity(struct sched_domain *sd, int cpu) 5822{ 5823 struct sched_domain *child = sd->child; 5824 struct sched_group *group, *sdg = sd->groups; 5825 unsigned long capacity, capacity_orig; 5826 unsigned long interval; 5827 5828 interval = msecs_to_jiffies(sd->balance_interval); 5829 interval = clamp(interval, 1UL, max_load_balance_interval); 5830 sdg->sgc->next_update = jiffies + interval; 5831 5832 if (!child) { 5833 update_cpu_capacity(sd, cpu); 5834 return; 5835 } 5836 5837 capacity_orig = capacity = 0; 5838 5839 if (child->flags & SD_OVERLAP) { 5840 /* 5841 * SD_OVERLAP domains cannot assume that child groups 5842 * span the current group. 5843 */ 5844 5845 for_each_cpu(cpu, sched_group_cpus(sdg)) { 5846 struct sched_group_capacity *sgc; 5847 struct rq *rq = cpu_rq(cpu); 5848 5849 /* 5850 * build_sched_domains() -> init_sched_groups_capacity() 5851 * gets here before we've attached the domains to the 5852 * runqueues. 5853 * 5854 * Use capacity_of(), which is set irrespective of domains 5855 * in update_cpu_capacity(). 5856 * 5857 * This avoids capacity/capacity_orig from being 0 and 5858 * causing divide-by-zero issues on boot. 5859 * 5860 * Runtime updates will correct capacity_orig. 5861 */ 5862 if (unlikely(!rq->sd)) { 5863 capacity_orig += capacity_of(cpu); 5864 capacity += capacity_of(cpu); 5865 continue; 5866 } 5867 5868 sgc = rq->sd->groups->sgc; 5869 capacity_orig += sgc->capacity_orig; 5870 capacity += sgc->capacity; 5871 } 5872 } else { 5873 /* 5874 * !SD_OVERLAP domains can assume that child groups 5875 * span the current group. 5876 */ 5877 5878 group = child->groups; 5879 do { 5880 capacity_orig += group->sgc->capacity_orig; 5881 capacity += group->sgc->capacity; 5882 group = group->next; 5883 } while (group != child->groups); 5884 } 5885 5886 sdg->sgc->capacity_orig = capacity_orig; 5887 sdg->sgc->capacity = capacity; 5888} 5889 5890/* 5891 * Try and fix up capacity for tiny siblings, this is needed when 5892 * things like SD_ASYM_PACKING need f_b_g to select another sibling 5893 * which on its own isn't powerful enough. 5894 * 5895 * See update_sd_pick_busiest() and check_asym_packing(). 5896 */ 5897static inline int 5898fix_small_capacity(struct sched_domain *sd, struct sched_group *group) 5899{ 5900 /* 5901 * Only siblings can have significantly less than SCHED_CAPACITY_SCALE 5902 */ 5903 if (!(sd->flags & SD_SHARE_CPUCAPACITY)) 5904 return 0; 5905 5906 /* 5907 * If ~90% of the cpu_capacity is still there, we're good. 5908 */ 5909 if (group->sgc->capacity * 32 > group->sgc->capacity_orig * 29) 5910 return 1; 5911 5912 return 0; 5913} 5914 5915/* 5916 * Group imbalance indicates (and tries to solve) the problem where balancing 5917 * groups is inadequate due to tsk_cpus_allowed() constraints. 5918 * 5919 * Imagine a situation of two groups of 4 cpus each and 4 tasks each with a 5920 * cpumask covering 1 cpu of the first group and 3 cpus of the second group. 5921 * Something like: 5922 * 5923 * { 0 1 2 3 } { 4 5 6 7 } 5924 * * * * * 5925 * 5926 * If we were to balance group-wise we'd place two tasks in the first group and 5927 * two tasks in the second group. Clearly this is undesired as it will overload 5928 * cpu 3 and leave one of the cpus in the second group unused. 5929 * 5930 * The current solution to this issue is detecting the skew in the first group 5931 * by noticing the lower domain failed to reach balance and had difficulty 5932 * moving tasks due to affinity constraints. 5933 * 5934 * When this is so detected; this group becomes a candidate for busiest; see 5935 * update_sd_pick_busiest(). And calculate_imbalance() and 5936 * find_busiest_group() avoid some of the usual balance conditions to allow it 5937 * to create an effective group imbalance. 5938 * 5939 * This is a somewhat tricky proposition since the next run might not find the 5940 * group imbalance and decide the groups need to be balanced again. A most 5941 * subtle and fragile situation. 5942 */ 5943 5944static inline int sg_imbalanced(struct sched_group *group) 5945{ 5946 return group->sgc->imbalance; 5947} 5948 5949/* 5950 * Compute the group capacity factor. 5951 * 5952 * Avoid the issue where N*frac(smt_capacity) >= 1 creates 'phantom' cores by 5953 * first dividing out the smt factor and computing the actual number of cores 5954 * and limit unit capacity with that. 5955 */ 5956static inline int sg_capacity_factor(struct lb_env *env, struct sched_group *group) 5957{ 5958 unsigned int capacity_factor, smt, cpus; 5959 unsigned int capacity, capacity_orig; 5960 5961 capacity = group->sgc->capacity; 5962 capacity_orig = group->sgc->capacity_orig; 5963 cpus = group->group_weight; 5964 5965 /* smt := ceil(cpus / capacity), assumes: 1 < smt_capacity < 2 */ 5966 smt = DIV_ROUND_UP(SCHED_CAPACITY_SCALE * cpus, capacity_orig); 5967 capacity_factor = cpus / smt; /* cores */ 5968 5969 capacity_factor = min_t(unsigned, 5970 capacity_factor, DIV_ROUND_CLOSEST(capacity, SCHED_CAPACITY_SCALE)); 5971 if (!capacity_factor) 5972 capacity_factor = fix_small_capacity(env->sd, group); 5973 5974 return capacity_factor; 5975} 5976 5977static enum group_type 5978group_classify(struct sched_group *group, struct sg_lb_stats *sgs) 5979{ 5980 if (sgs->sum_nr_running > sgs->group_capacity_factor) 5981 return group_overloaded; 5982 5983 if (sg_imbalanced(group)) 5984 return group_imbalanced; 5985 5986 return group_other; 5987} 5988 5989/** 5990 * update_sg_lb_stats - Update sched_group's statistics for load balancing. 5991 * @env: The load balancing environment. 5992 * @group: sched_group whose statistics are to be updated. 5993 * @load_idx: Load index of sched_domain of this_cpu for load calc. 5994 * @local_group: Does group contain this_cpu. 5995 * @sgs: variable to hold the statistics for this group. 5996 * @overload: Indicate more than one runnable task for any CPU. 5997 */ 5998static inline void update_sg_lb_stats(struct lb_env *env, 5999 struct sched_group *group, int load_idx, 6000 int local_group, struct sg_lb_stats *sgs, 6001 bool *overload) 6002{ 6003 unsigned long load; 6004 int i; 6005 6006 memset(sgs, 0, sizeof(*sgs)); 6007 6008 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) { 6009 struct rq *rq = cpu_rq(i); 6010 6011 /* Bias balancing toward cpus of our domain */ 6012 if (local_group) 6013 load = target_load(i, load_idx); 6014 else 6015 load = source_load(i, load_idx); 6016 6017 sgs->group_load += load; 6018 sgs->sum_nr_running += rq->cfs.h_nr_running; 6019 6020 if (rq->nr_running > 1) 6021 *overload = true; 6022 6023#ifdef CONFIG_NUMA_BALANCING 6024 sgs->nr_numa_running += rq->nr_numa_running; 6025 sgs->nr_preferred_running += rq->nr_preferred_running; 6026#endif 6027 sgs->sum_weighted_load += weighted_cpuload(i); 6028 if (idle_cpu(i)) 6029 sgs->idle_cpus++; 6030 } 6031 6032 /* Adjust by relative CPU capacity of the group */ 6033 sgs->group_capacity = group->sgc->capacity; 6034 sgs->avg_load = (sgs->group_load*SCHED_CAPACITY_SCALE) / sgs->group_capacity; 6035 6036 if (sgs->sum_nr_running) 6037 sgs->load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running; 6038 6039 sgs->group_weight = group->group_weight; 6040 sgs->group_capacity_factor = sg_capacity_factor(env, group); 6041 sgs->group_type = group_classify(group, sgs); 6042 6043 if (sgs->group_capacity_factor > sgs->sum_nr_running) 6044 sgs->group_has_free_capacity = 1; 6045} 6046 6047/** 6048 * update_sd_pick_busiest - return 1 on busiest group 6049 * @env: The load balancing environment. 6050 * @sds: sched_domain statistics 6051 * @sg: sched_group candidate to be checked for being the busiest 6052 * @sgs: sched_group statistics 6053 * 6054 * Determine if @sg is a busier group than the previously selected 6055 * busiest group. 6056 * 6057 * Return: %true if @sg is a busier group than the previously selected 6058 * busiest group. %false otherwise. 6059 */ 6060static bool update_sd_pick_busiest(struct lb_env *env, 6061 struct sd_lb_stats *sds, 6062 struct sched_group *sg, 6063 struct sg_lb_stats *sgs) 6064{ 6065 struct sg_lb_stats *busiest = &sds->busiest_stat; 6066 6067 if (sgs->group_type > busiest->group_type) 6068 return true; 6069 6070 if (sgs->group_type < busiest->group_type) 6071 return false; 6072 6073 if (sgs->avg_load <= busiest->avg_load) 6074 return false; 6075 6076 /* This is the busiest node in its class. */ 6077 if (!(env->sd->flags & SD_ASYM_PACKING)) 6078 return true; 6079 6080 /* 6081 * ASYM_PACKING needs to move all the work to the lowest 6082 * numbered CPUs in the group, therefore mark all groups 6083 * higher than ourself as busy. 6084 */ 6085 if (sgs->sum_nr_running && env->dst_cpu < group_first_cpu(sg)) { 6086 if (!sds->busiest) 6087 return true; 6088 6089 if (group_first_cpu(sds->busiest) > group_first_cpu(sg)) 6090 return true; 6091 } 6092 6093 return false; 6094} 6095 6096#ifdef CONFIG_NUMA_BALANCING 6097static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs) 6098{ 6099 if (sgs->sum_nr_running > sgs->nr_numa_running) 6100 return regular; 6101 if (sgs->sum_nr_running > sgs->nr_preferred_running) 6102 return remote; 6103 return all; 6104} 6105 6106static inline enum fbq_type fbq_classify_rq(struct rq *rq) 6107{ 6108 if (rq->nr_running > rq->nr_numa_running) 6109 return regular; 6110 if (rq->nr_running > rq->nr_preferred_running) 6111 return remote; 6112 return all; 6113} 6114#else 6115static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs) 6116{ 6117 return all; 6118} 6119 6120static inline enum fbq_type fbq_classify_rq(struct rq *rq) 6121{ 6122 return regular; 6123} 6124#endif /* CONFIG_NUMA_BALANCING */ 6125 6126/** 6127 * update_sd_lb_stats - Update sched_domain's statistics for load balancing. 6128 * @env: The load balancing environment. 6129 * @sds: variable to hold the statistics for this sched_domain. 6130 */ 6131static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds) 6132{ 6133 struct sched_domain *child = env->sd->child; 6134 struct sched_group *sg = env->sd->groups; 6135 struct sg_lb_stats tmp_sgs; 6136 int load_idx, prefer_sibling = 0; 6137 bool overload = false; 6138 6139 if (child && child->flags & SD_PREFER_SIBLING) 6140 prefer_sibling = 1; 6141 6142 load_idx = get_sd_load_idx(env->sd, env->idle); 6143 6144 do { 6145 struct sg_lb_stats *sgs = &tmp_sgs; 6146 int local_group; 6147 6148 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_cpus(sg)); 6149 if (local_group) { 6150 sds->local = sg; 6151 sgs = &sds->local_stat; 6152 6153 if (env->idle != CPU_NEWLY_IDLE || 6154 time_after_eq(jiffies, sg->sgc->next_update)) 6155 update_group_capacity(env->sd, env->dst_cpu); 6156 } 6157 6158 update_sg_lb_stats(env, sg, load_idx, local_group, sgs, 6159 &overload); 6160 6161 if (local_group) 6162 goto next_group; 6163 6164 /* 6165 * In case the child domain prefers tasks go to siblings 6166 * first, lower the sg capacity factor to one so that we'll try 6167 * and move all the excess tasks away. We lower the capacity 6168 * of a group only if the local group has the capacity to fit 6169 * these excess tasks, i.e. nr_running < group_capacity_factor. The 6170 * extra check prevents the case where you always pull from the 6171 * heaviest group when it is already under-utilized (possible 6172 * with a large weight task outweighs the tasks on the system). 6173 */ 6174 if (prefer_sibling && sds->local && 6175 sds->local_stat.group_has_free_capacity) 6176 sgs->group_capacity_factor = min(sgs->group_capacity_factor, 1U); 6177 6178 if (update_sd_pick_busiest(env, sds, sg, sgs)) { 6179 sds->busiest = sg; 6180 sds->busiest_stat = *sgs; 6181 } 6182 6183next_group: 6184 /* Now, start updating sd_lb_stats */ 6185 sds->total_load += sgs->group_load; 6186 sds->total_capacity += sgs->group_capacity; 6187 6188 sg = sg->next; 6189 } while (sg != env->sd->groups); 6190 6191 if (env->sd->flags & SD_NUMA) 6192 env->fbq_type = fbq_classify_group(&sds->busiest_stat); 6193 6194 if (!env->sd->parent) { 6195 /* update overload indicator if we are at root domain */ 6196 if (env->dst_rq->rd->overload != overload) 6197 env->dst_rq->rd->overload = overload; 6198 } 6199 6200} 6201 6202/** 6203 * check_asym_packing - Check to see if the group is packed into the 6204 * sched doman. 6205 * 6206 * This is primarily intended to used at the sibling level. Some 6207 * cores like POWER7 prefer to use lower numbered SMT threads. In the 6208 * case of POWER7, it can move to lower SMT modes only when higher 6209 * threads are idle. When in lower SMT modes, the threads will 6210 * perform better since they share less core resources. Hence when we 6211 * have idle threads, we want them to be the higher ones. 6212 * 6213 * This packing function is run on idle threads. It checks to see if 6214 * the busiest CPU in this domain (core in the P7 case) has a higher 6215 * CPU number than the packing function is being run on. Here we are 6216 * assuming lower CPU number will be equivalent to lower a SMT thread 6217 * number. 6218 * 6219 * Return: 1 when packing is required and a task should be moved to 6220 * this CPU. The amount of the imbalance is returned in *imbalance. 6221 * 6222 * @env: The load balancing environment. 6223 * @sds: Statistics of the sched_domain which is to be packed 6224 */ 6225static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds) 6226{ 6227 int busiest_cpu; 6228 6229 if (!(env->sd->flags & SD_ASYM_PACKING)) 6230 return 0; 6231 6232 if (!sds->busiest) 6233 return 0; 6234 6235 busiest_cpu = group_first_cpu(sds->busiest); 6236 if (env->dst_cpu > busiest_cpu) 6237 return 0; 6238 6239 env->imbalance = DIV_ROUND_CLOSEST( 6240 sds->busiest_stat.avg_load * sds->busiest_stat.group_capacity, 6241 SCHED_CAPACITY_SCALE); 6242 6243 return 1; 6244} 6245 6246/** 6247 * fix_small_imbalance - Calculate the minor imbalance that exists 6248 * amongst the groups of a sched_domain, during 6249 * load balancing. 6250 * @env: The load balancing environment. 6251 * @sds: Statistics of the sched_domain whose imbalance is to be calculated. 6252 */ 6253static inline 6254void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds) 6255{ 6256 unsigned long tmp, capa_now = 0, capa_move = 0; 6257 unsigned int imbn = 2; 6258 unsigned long scaled_busy_load_per_task; 6259 struct sg_lb_stats *local, *busiest; 6260 6261 local = &sds->local_stat; 6262 busiest = &sds->busiest_stat; 6263 6264 if (!local->sum_nr_running) 6265 local->load_per_task = cpu_avg_load_per_task(env->dst_cpu); 6266 else if (busiest->load_per_task > local->load_per_task) 6267 imbn = 1; 6268 6269 scaled_busy_load_per_task = 6270 (busiest->load_per_task * SCHED_CAPACITY_SCALE) / 6271 busiest->group_capacity; 6272 6273 if (busiest->avg_load + scaled_busy_load_per_task >= 6274 local->avg_load + (scaled_busy_load_per_task * imbn)) { 6275 env->imbalance = busiest->load_per_task; 6276 return; 6277 } 6278 6279 /* 6280 * OK, we don't have enough imbalance to justify moving tasks, 6281 * however we may be able to increase total CPU capacity used by 6282 * moving them. 6283 */ 6284 6285 capa_now += busiest->group_capacity * 6286 min(busiest->load_per_task, busiest->avg_load); 6287 capa_now += local->group_capacity * 6288 min(local->load_per_task, local->avg_load); 6289 capa_now /= SCHED_CAPACITY_SCALE; 6290 6291 /* Amount of load we'd subtract */ 6292 if (busiest->avg_load > scaled_busy_load_per_task) { 6293 capa_move += busiest->group_capacity * 6294 min(busiest->load_per_task, 6295 busiest->avg_load - scaled_busy_load_per_task); 6296 } 6297 6298 /* Amount of load we'd add */ 6299 if (busiest->avg_load * busiest->group_capacity < 6300 busiest->load_per_task * SCHED_CAPACITY_SCALE) { 6301 tmp = (busiest->avg_load * busiest->group_capacity) / 6302 local->group_capacity; 6303 } else { 6304 tmp = (busiest->load_per_task * SCHED_CAPACITY_SCALE) / 6305 local->group_capacity; 6306 } 6307 capa_move += local->group_capacity * 6308 min(local->load_per_task, local->avg_load + tmp); 6309 capa_move /= SCHED_CAPACITY_SCALE; 6310 6311 /* Move if we gain throughput */ 6312 if (capa_move > capa_now) 6313 env->imbalance = busiest->load_per_task; 6314} 6315 6316/** 6317 * calculate_imbalance - Calculate the amount of imbalance present within the 6318 * groups of a given sched_domain during load balance. 6319 * @env: load balance environment 6320 * @sds: statistics of the sched_domain whose imbalance is to be calculated. 6321 */ 6322static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds) 6323{ 6324 unsigned long max_pull, load_above_capacity = ~0UL; 6325 struct sg_lb_stats *local, *busiest; 6326 6327 local = &sds->local_stat; 6328 busiest = &sds->busiest_stat; 6329 6330 if (busiest->group_type == group_imbalanced) { 6331 /* 6332 * In the group_imb case we cannot rely on group-wide averages 6333 * to ensure cpu-load equilibrium, look at wider averages. XXX 6334 */ 6335 busiest->load_per_task = 6336 min(busiest->load_per_task, sds->avg_load); 6337 } 6338 6339 /* 6340 * In the presence of smp nice balancing, certain scenarios can have 6341 * max load less than avg load(as we skip the groups at or below 6342 * its cpu_capacity, while calculating max_load..) 6343 */ 6344 if (busiest->avg_load <= sds->avg_load || 6345 local->avg_load >= sds->avg_load) { 6346 env->imbalance = 0; 6347 return fix_small_imbalance(env, sds); 6348 } 6349 6350 /* 6351 * If there aren't any idle cpus, avoid creating some. 6352 */ 6353 if (busiest->group_type == group_overloaded && 6354 local->group_type == group_overloaded) { 6355 load_above_capacity = 6356 (busiest->sum_nr_running - busiest->group_capacity_factor); 6357 6358 load_above_capacity *= (SCHED_LOAD_SCALE * SCHED_CAPACITY_SCALE); 6359 load_above_capacity /= busiest->group_capacity; 6360 } 6361 6362 /* 6363 * We're trying to get all the cpus to the average_load, so we don't 6364 * want to push ourselves above the average load, nor do we wish to 6365 * reduce the max loaded cpu below the average load. At the same time, 6366 * we also don't want to reduce the group load below the group capacity 6367 * (so that we can implement power-savings policies etc). Thus we look 6368 * for the minimum possible imbalance. 6369 */ 6370 max_pull = min(busiest->avg_load - sds->avg_load, load_above_capacity); 6371 6372 /* How much load to actually move to equalise the imbalance */ 6373 env->imbalance = min( 6374 max_pull * busiest->group_capacity, 6375 (sds->avg_load - local->avg_load) * local->group_capacity 6376 ) / SCHED_CAPACITY_SCALE; 6377 6378 /* 6379 * if *imbalance is less than the average load per runnable task 6380 * there is no guarantee that any tasks will be moved so we'll have 6381 * a think about bumping its value to force at least one task to be 6382 * moved 6383 */ 6384 if (env->imbalance < busiest->load_per_task) 6385 return fix_small_imbalance(env, sds); 6386} 6387 6388/******* find_busiest_group() helpers end here *********************/ 6389 6390/** 6391 * find_busiest_group - Returns the busiest group within the sched_domain 6392 * if there is an imbalance. If there isn't an imbalance, and 6393 * the user has opted for power-savings, it returns a group whose 6394 * CPUs can be put to idle by rebalancing those tasks elsewhere, if 6395 * such a group exists. 6396 * 6397 * Also calculates the amount of weighted load which should be moved 6398 * to restore balance. 6399 * 6400 * @env: The load balancing environment. 6401 * 6402 * Return: - The busiest group if imbalance exists. 6403 * - If no imbalance and user has opted for power-savings balance, 6404 * return the least loaded group whose CPUs can be 6405 * put to idle by rebalancing its tasks onto our group. 6406 */ 6407static struct sched_group *find_busiest_group(struct lb_env *env) 6408{ 6409 struct sg_lb_stats *local, *busiest; 6410 struct sd_lb_stats sds; 6411 6412 init_sd_lb_stats(&sds); 6413 6414 /* 6415 * Compute the various statistics relavent for load balancing at 6416 * this level. 6417 */ 6418 update_sd_lb_stats(env, &sds); 6419 local = &sds.local_stat; 6420 busiest = &sds.busiest_stat; 6421 6422 if ((env->idle == CPU_IDLE || env->idle == CPU_NEWLY_IDLE) && 6423 check_asym_packing(env, &sds)) 6424 return sds.busiest; 6425 6426 /* There is no busy sibling group to pull tasks from */ 6427 if (!sds.busiest || busiest->sum_nr_running == 0) 6428 goto out_balanced; 6429 6430 sds.avg_load = (SCHED_CAPACITY_SCALE * sds.total_load) 6431 / sds.total_capacity; 6432 6433 /* 6434 * If the busiest group is imbalanced the below checks don't 6435 * work because they assume all things are equal, which typically 6436 * isn't true due to cpus_allowed constraints and the like. 6437 */ 6438 if (busiest->group_type == group_imbalanced) 6439 goto force_balance; 6440 6441 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */ 6442 if (env->idle == CPU_NEWLY_IDLE && local->group_has_free_capacity && 6443 !busiest->group_has_free_capacity) 6444 goto force_balance; 6445 6446 /* 6447 * If the local group is busier than the selected busiest group 6448 * don't try and pull any tasks. 6449 */ 6450 if (local->avg_load >= busiest->avg_load) 6451 goto out_balanced; 6452 6453 /* 6454 * Don't pull any tasks if this group is already above the domain 6455 * average load. 6456 */ 6457 if (local->avg_load >= sds.avg_load) 6458 goto out_balanced; 6459 6460 if (env->idle == CPU_IDLE) { 6461 /* 6462 * This cpu is idle. If the busiest group is not overloaded 6463 * and there is no imbalance between this and busiest group 6464 * wrt idle cpus, it is balanced. The imbalance becomes 6465 * significant if the diff is greater than 1 otherwise we 6466 * might end up to just move the imbalance on another group 6467 */ 6468 if ((busiest->group_type != group_overloaded) && 6469 (local->idle_cpus <= (busiest->idle_cpus + 1))) 6470 goto out_balanced; 6471 } else { 6472 /* 6473 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use 6474 * imbalance_pct to be conservative. 6475 */ 6476 if (100 * busiest->avg_load <= 6477 env->sd->imbalance_pct * local->avg_load) 6478 goto out_balanced; 6479 } 6480 6481force_balance: 6482 /* Looks like there is an imbalance. Compute it */ 6483 calculate_imbalance(env, &sds); 6484 return sds.busiest; 6485 6486out_balanced: 6487 env->imbalance = 0; 6488 return NULL; 6489} 6490 6491/* 6492 * find_busiest_queue - find the busiest runqueue among the cpus in group. 6493 */ 6494static struct rq *find_busiest_queue(struct lb_env *env, 6495 struct sched_group *group) 6496{ 6497 struct rq *busiest = NULL, *rq; 6498 unsigned long busiest_load = 0, busiest_capacity = 1; 6499 int i; 6500 6501 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) { 6502 unsigned long capacity, capacity_factor, wl; 6503 enum fbq_type rt; 6504 6505 rq = cpu_rq(i); 6506 rt = fbq_classify_rq(rq); 6507 6508 /* 6509 * We classify groups/runqueues into three groups: 6510 * - regular: there are !numa tasks 6511 * - remote: there are numa tasks that run on the 'wrong' node 6512 * - all: there is no distinction 6513 * 6514 * In order to avoid migrating ideally placed numa tasks, 6515 * ignore those when there's better options. 6516 * 6517 * If we ignore the actual busiest queue to migrate another 6518 * task, the next balance pass can still reduce the busiest 6519 * queue by moving tasks around inside the node. 6520 * 6521 * If we cannot move enough load due to this classification 6522 * the next pass will adjust the group classification and 6523 * allow migration of more tasks. 6524 * 6525 * Both cases only affect the total convergence complexity. 6526 */ 6527 if (rt > env->fbq_type) 6528 continue; 6529 6530 capacity = capacity_of(i); 6531 capacity_factor = DIV_ROUND_CLOSEST(capacity, SCHED_CAPACITY_SCALE); 6532 if (!capacity_factor) 6533 capacity_factor = fix_small_capacity(env->sd, group); 6534 6535 wl = weighted_cpuload(i); 6536 6537 /* 6538 * When comparing with imbalance, use weighted_cpuload() 6539 * which is not scaled with the cpu capacity. 6540 */ 6541 if (capacity_factor && rq->nr_running == 1 && wl > env->imbalance) 6542 continue; 6543 6544 /* 6545 * For the load comparisons with the other cpu's, consider 6546 * the weighted_cpuload() scaled with the cpu capacity, so 6547 * that the load can be moved away from the cpu that is 6548 * potentially running at a lower capacity. 6549 * 6550 * Thus we're looking for max(wl_i / capacity_i), crosswise 6551 * multiplication to rid ourselves of the division works out 6552 * to: wl_i * capacity_j > wl_j * capacity_i; where j is 6553 * our previous maximum. 6554 */ 6555 if (wl * busiest_capacity > busiest_load * capacity) { 6556 busiest_load = wl; 6557 busiest_capacity = capacity; 6558 busiest = rq; 6559 } 6560 } 6561 6562 return busiest; 6563} 6564 6565/* 6566 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but 6567 * so long as it is large enough. 6568 */ 6569#define MAX_PINNED_INTERVAL 512 6570 6571/* Working cpumask for load_balance and load_balance_newidle. */ 6572DEFINE_PER_CPU(cpumask_var_t, load_balance_mask); 6573 6574static int need_active_balance(struct lb_env *env) 6575{ 6576 struct sched_domain *sd = env->sd; 6577 6578 if (env->idle == CPU_NEWLY_IDLE) { 6579 6580 /* 6581 * ASYM_PACKING needs to force migrate tasks from busy but 6582 * higher numbered CPUs in order to pack all tasks in the 6583 * lowest numbered CPUs. 6584 */ 6585 if ((sd->flags & SD_ASYM_PACKING) && env->src_cpu > env->dst_cpu) 6586 return 1; 6587 } 6588 6589 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2); 6590} 6591 6592static int active_load_balance_cpu_stop(void *data); 6593 6594static int should_we_balance(struct lb_env *env) 6595{ 6596 struct sched_group *sg = env->sd->groups; 6597 struct cpumask *sg_cpus, *sg_mask; 6598 int cpu, balance_cpu = -1; 6599 6600 /* 6601 * In the newly idle case, we will allow all the cpu's 6602 * to do the newly idle load balance. 6603 */ 6604 if (env->idle == CPU_NEWLY_IDLE) 6605 return 1; 6606 6607 sg_cpus = sched_group_cpus(sg); 6608 sg_mask = sched_group_mask(sg); 6609 /* Try to find first idle cpu */ 6610 for_each_cpu_and(cpu, sg_cpus, env->cpus) { 6611 if (!cpumask_test_cpu(cpu, sg_mask) || !idle_cpu(cpu)) 6612 continue; 6613 6614 balance_cpu = cpu; 6615 break; 6616 } 6617 6618 if (balance_cpu == -1) 6619 balance_cpu = group_balance_cpu(sg); 6620 6621 /* 6622 * First idle cpu or the first cpu(busiest) in this sched group 6623 * is eligible for doing load balancing at this and above domains. 6624 */ 6625 return balance_cpu == env->dst_cpu; 6626} 6627 6628/* 6629 * Check this_cpu to ensure it is balanced within domain. Attempt to move 6630 * tasks if there is an imbalance. 6631 */ 6632static int load_balance(int this_cpu, struct rq *this_rq, 6633 struct sched_domain *sd, enum cpu_idle_type idle, 6634 int *continue_balancing) 6635{ 6636 int ld_moved, cur_ld_moved, active_balance = 0; 6637 struct sched_domain *sd_parent = sd->parent; 6638 struct sched_group *group; 6639 struct rq *busiest; 6640 unsigned long flags; 6641 struct cpumask *cpus = this_cpu_cpumask_var_ptr(load_balance_mask); 6642 6643 struct lb_env env = { 6644 .sd = sd, 6645 .dst_cpu = this_cpu, 6646 .dst_rq = this_rq, 6647 .dst_grpmask = sched_group_cpus(sd->groups), 6648 .idle = idle, 6649 .loop_break = sched_nr_migrate_break, 6650 .cpus = cpus, 6651 .fbq_type = all, 6652 .tasks = LIST_HEAD_INIT(env.tasks), 6653 }; 6654 6655 /* 6656 * For NEWLY_IDLE load_balancing, we don't need to consider 6657 * other cpus in our group 6658 */ 6659 if (idle == CPU_NEWLY_IDLE) 6660 env.dst_grpmask = NULL; 6661 6662 cpumask_copy(cpus, cpu_active_mask); 6663 6664 schedstat_inc(sd, lb_count[idle]); 6665 6666redo: 6667 if (!should_we_balance(&env)) { 6668 *continue_balancing = 0; 6669 goto out_balanced; 6670 } 6671 6672 group = find_busiest_group(&env); 6673 if (!group) { 6674 schedstat_inc(sd, lb_nobusyg[idle]); 6675 goto out_balanced; 6676 } 6677 6678 busiest = find_busiest_queue(&env, group); 6679 if (!busiest) { 6680 schedstat_inc(sd, lb_nobusyq[idle]); 6681 goto out_balanced; 6682 } 6683 6684 BUG_ON(busiest == env.dst_rq); 6685 6686 schedstat_add(sd, lb_imbalance[idle], env.imbalance); 6687 6688 ld_moved = 0; 6689 if (busiest->nr_running > 1) { 6690 /* 6691 * Attempt to move tasks. If find_busiest_group has found 6692 * an imbalance but busiest->nr_running <= 1, the group is 6693 * still unbalanced. ld_moved simply stays zero, so it is 6694 * correctly treated as an imbalance. 6695 */ 6696 env.flags |= LBF_ALL_PINNED; 6697 env.src_cpu = busiest->cpu; 6698 env.src_rq = busiest; 6699 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running); 6700 6701more_balance: 6702 raw_spin_lock_irqsave(&busiest->lock, flags); 6703 6704 /* 6705 * cur_ld_moved - load moved in current iteration 6706 * ld_moved - cumulative load moved across iterations 6707 */ 6708 cur_ld_moved = detach_tasks(&env); 6709 6710 /* 6711 * We've detached some tasks from busiest_rq. Every 6712 * task is masked "TASK_ON_RQ_MIGRATING", so we can safely 6713 * unlock busiest->lock, and we are able to be sure 6714 * that nobody can manipulate the tasks in parallel. 6715 * See task_rq_lock() family for the details. 6716 */ 6717 6718 raw_spin_unlock(&busiest->lock); 6719 6720 if (cur_ld_moved) { 6721 attach_tasks(&env); 6722 ld_moved += cur_ld_moved; 6723 } 6724 6725 local_irq_restore(flags); 6726 6727 if (env.flags & LBF_NEED_BREAK) { 6728 env.flags &= ~LBF_NEED_BREAK; 6729 goto more_balance; 6730 } 6731 6732 /* 6733 * Revisit (affine) tasks on src_cpu that couldn't be moved to 6734 * us and move them to an alternate dst_cpu in our sched_group 6735 * where they can run. The upper limit on how many times we 6736 * iterate on same src_cpu is dependent on number of cpus in our 6737 * sched_group. 6738 * 6739 * This changes load balance semantics a bit on who can move 6740 * load to a given_cpu. In addition to the given_cpu itself 6741 * (or a ilb_cpu acting on its behalf where given_cpu is 6742 * nohz-idle), we now have balance_cpu in a position to move 6743 * load to given_cpu. In rare situations, this may cause 6744 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding 6745 * _independently_ and at _same_ time to move some load to 6746 * given_cpu) causing exceess load to be moved to given_cpu. 6747 * This however should not happen so much in practice and 6748 * moreover subsequent load balance cycles should correct the 6749 * excess load moved. 6750 */ 6751 if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) { 6752 6753 /* Prevent to re-select dst_cpu via env's cpus */ 6754 cpumask_clear_cpu(env.dst_cpu, env.cpus); 6755 6756 env.dst_rq = cpu_rq(env.new_dst_cpu); 6757 env.dst_cpu = env.new_dst_cpu; 6758 env.flags &= ~LBF_DST_PINNED; 6759 env.loop = 0; 6760 env.loop_break = sched_nr_migrate_break; 6761 6762 /* 6763 * Go back to "more_balance" rather than "redo" since we 6764 * need to continue with same src_cpu. 6765 */ 6766 goto more_balance; 6767 } 6768 6769 /* 6770 * We failed to reach balance because of affinity. 6771 */ 6772 if (sd_parent) { 6773 int *group_imbalance = &sd_parent->groups->sgc->imbalance; 6774 6775 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0) 6776 *group_imbalance = 1; 6777 } 6778 6779 /* All tasks on this runqueue were pinned by CPU affinity */ 6780 if (unlikely(env.flags & LBF_ALL_PINNED)) { 6781 cpumask_clear_cpu(cpu_of(busiest), cpus); 6782 if (!cpumask_empty(cpus)) { 6783 env.loop = 0; 6784 env.loop_break = sched_nr_migrate_break; 6785 goto redo; 6786 } 6787 goto out_all_pinned; 6788 } 6789 } 6790 6791 if (!ld_moved) { 6792 schedstat_inc(sd, lb_failed[idle]); 6793 /* 6794 * Increment the failure counter only on periodic balance. 6795 * We do not want newidle balance, which can be very 6796 * frequent, pollute the failure counter causing 6797 * excessive cache_hot migrations and active balances. 6798 */ 6799 if (idle != CPU_NEWLY_IDLE) 6800 sd->nr_balance_failed++; 6801 6802 if (need_active_balance(&env)) { 6803 raw_spin_lock_irqsave(&busiest->lock, flags); 6804 6805 /* don't kick the active_load_balance_cpu_stop, 6806 * if the curr task on busiest cpu can't be 6807 * moved to this_cpu 6808 */ 6809 if (!cpumask_test_cpu(this_cpu, 6810 tsk_cpus_allowed(busiest->curr))) { 6811 raw_spin_unlock_irqrestore(&busiest->lock, 6812 flags); 6813 env.flags |= LBF_ALL_PINNED; 6814 goto out_one_pinned; 6815 } 6816 6817 /* 6818 * ->active_balance synchronizes accesses to 6819 * ->active_balance_work. Once set, it's cleared 6820 * only after active load balance is finished. 6821 */ 6822 if (!busiest->active_balance) { 6823 busiest->active_balance = 1; 6824 busiest->push_cpu = this_cpu; 6825 active_balance = 1; 6826 } 6827 raw_spin_unlock_irqrestore(&busiest->lock, flags); 6828 6829 if (active_balance) { 6830 stop_one_cpu_nowait(cpu_of(busiest), 6831 active_load_balance_cpu_stop, busiest, 6832 &busiest->active_balance_work); 6833 } 6834 6835 /* 6836 * We've kicked active balancing, reset the failure 6837 * counter. 6838 */ 6839 sd->nr_balance_failed = sd->cache_nice_tries+1; 6840 } 6841 } else 6842 sd->nr_balance_failed = 0; 6843 6844 if (likely(!active_balance)) { 6845 /* We were unbalanced, so reset the balancing interval */ 6846 sd->balance_interval = sd->min_interval; 6847 } else { 6848 /* 6849 * If we've begun active balancing, start to back off. This 6850 * case may not be covered by the all_pinned logic if there 6851 * is only 1 task on the busy runqueue (because we don't call 6852 * detach_tasks). 6853 */ 6854 if (sd->balance_interval < sd->max_interval) 6855 sd->balance_interval *= 2; 6856 } 6857 6858 goto out; 6859 6860out_balanced: 6861 /* 6862 * We reach balance although we may have faced some affinity 6863 * constraints. Clear the imbalance flag if it was set. 6864 */ 6865 if (sd_parent) { 6866 int *group_imbalance = &sd_parent->groups->sgc->imbalance; 6867 6868 if (*group_imbalance) 6869 *group_imbalance = 0; 6870 } 6871 6872out_all_pinned: 6873 /* 6874 * We reach balance because all tasks are pinned at this level so 6875 * we can't migrate them. Let the imbalance flag set so parent level 6876 * can try to migrate them. 6877 */ 6878 schedstat_inc(sd, lb_balanced[idle]); 6879 6880 sd->nr_balance_failed = 0; 6881 6882out_one_pinned: 6883 /* tune up the balancing interval */ 6884 if (((env.flags & LBF_ALL_PINNED) && 6885 sd->balance_interval < MAX_PINNED_INTERVAL) || 6886 (sd->balance_interval < sd->max_interval)) 6887 sd->balance_interval *= 2; 6888 6889 ld_moved = 0; 6890out: 6891 return ld_moved; 6892} 6893 6894static inline unsigned long 6895get_sd_balance_interval(struct sched_domain *sd, int cpu_busy) 6896{ 6897 unsigned long interval = sd->balance_interval; 6898 6899 if (cpu_busy) 6900 interval *= sd->busy_factor; 6901 6902 /* scale ms to jiffies */ 6903 interval = msecs_to_jiffies(interval); 6904 interval = clamp(interval, 1UL, max_load_balance_interval); 6905 6906 return interval; 6907} 6908 6909static inline void 6910update_next_balance(struct sched_domain *sd, int cpu_busy, unsigned long *next_balance) 6911{ 6912 unsigned long interval, next; 6913 6914 interval = get_sd_balance_interval(sd, cpu_busy); 6915 next = sd->last_balance + interval; 6916 6917 if (time_after(*next_balance, next)) 6918 *next_balance = next; 6919} 6920 6921/* 6922 * idle_balance is called by schedule() if this_cpu is about to become 6923 * idle. Attempts to pull tasks from other CPUs. 6924 */ 6925static int idle_balance(struct rq *this_rq) 6926{ 6927 unsigned long next_balance = jiffies + HZ; 6928 int this_cpu = this_rq->cpu; 6929 struct sched_domain *sd; 6930 int pulled_task = 0; 6931 u64 curr_cost = 0; 6932 6933 idle_enter_fair(this_rq); 6934 6935 /* 6936 * We must set idle_stamp _before_ calling idle_balance(), such that we 6937 * measure the duration of idle_balance() as idle time. 6938 */ 6939 this_rq->idle_stamp = rq_clock(this_rq); 6940 6941 if (this_rq->avg_idle < sysctl_sched_migration_cost || 6942 !this_rq->rd->overload) { 6943 rcu_read_lock(); 6944 sd = rcu_dereference_check_sched_domain(this_rq->sd); 6945 if (sd) 6946 update_next_balance(sd, 0, &next_balance); 6947 rcu_read_unlock(); 6948 6949 goto out; 6950 } 6951 6952 /* 6953 * Drop the rq->lock, but keep IRQ/preempt disabled. 6954 */ 6955 raw_spin_unlock(&this_rq->lock); 6956 6957 update_blocked_averages(this_cpu); 6958 rcu_read_lock(); 6959 for_each_domain(this_cpu, sd) { 6960 int continue_balancing = 1; 6961 u64 t0, domain_cost; 6962 6963 if (!(sd->flags & SD_LOAD_BALANCE)) 6964 continue; 6965 6966 if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost) { 6967 update_next_balance(sd, 0, &next_balance); 6968 break; 6969 } 6970 6971 if (sd->flags & SD_BALANCE_NEWIDLE) { 6972 t0 = sched_clock_cpu(this_cpu); 6973 6974 pulled_task = load_balance(this_cpu, this_rq, 6975 sd, CPU_NEWLY_IDLE, 6976 &continue_balancing); 6977 6978 domain_cost = sched_clock_cpu(this_cpu) - t0; 6979 if (domain_cost > sd->max_newidle_lb_cost) 6980 sd->max_newidle_lb_cost = domain_cost; 6981 6982 curr_cost += domain_cost; 6983 } 6984 6985 update_next_balance(sd, 0, &next_balance); 6986 6987 /* 6988 * Stop searching for tasks to pull if there are 6989 * now runnable tasks on this rq. 6990 */ 6991 if (pulled_task || this_rq->nr_running > 0) 6992 break; 6993 } 6994 rcu_read_unlock(); 6995 6996 raw_spin_lock(&this_rq->lock); 6997 6998 if (curr_cost > this_rq->max_idle_balance_cost) 6999 this_rq->max_idle_balance_cost = curr_cost; 7000 7001 /* 7002 * While browsing the domains, we released the rq lock, a task could 7003 * have been enqueued in the meantime. Since we're not going idle, 7004 * pretend we pulled a task. 7005 */ 7006 if (this_rq->cfs.h_nr_running && !pulled_task) 7007 pulled_task = 1; 7008 7009out: 7010 /* Move the next balance forward */ 7011 if (time_after(this_rq->next_balance, next_balance)) 7012 this_rq->next_balance = next_balance; 7013 7014 /* Is there a task of a high priority class? */ 7015 if (this_rq->nr_running != this_rq->cfs.h_nr_running) 7016 pulled_task = -1; 7017 7018 if (pulled_task) { 7019 idle_exit_fair(this_rq); 7020 this_rq->idle_stamp = 0; 7021 } 7022 7023 return pulled_task; 7024} 7025 7026/* 7027 * active_load_balance_cpu_stop is run by cpu stopper. It pushes 7028 * running tasks off the busiest CPU onto idle CPUs. It requires at 7029 * least 1 task to be running on each physical CPU where possible, and 7030 * avoids physical / logical imbalances. 7031 */ 7032static int active_load_balance_cpu_stop(void *data) 7033{ 7034 struct rq *busiest_rq = data; 7035 int busiest_cpu = cpu_of(busiest_rq); 7036 int target_cpu = busiest_rq->push_cpu; 7037 struct rq *target_rq = cpu_rq(target_cpu); 7038 struct sched_domain *sd; 7039 struct task_struct *p = NULL; 7040 7041 raw_spin_lock_irq(&busiest_rq->lock); 7042 7043 /* make sure the requested cpu hasn't gone down in the meantime */ 7044 if (unlikely(busiest_cpu != smp_processor_id() || 7045 !busiest_rq->active_balance)) 7046 goto out_unlock; 7047 7048 /* Is there any task to move? */ 7049 if (busiest_rq->nr_running <= 1) 7050 goto out_unlock; 7051 7052 /* 7053 * This condition is "impossible", if it occurs 7054 * we need to fix it. Originally reported by 7055 * Bjorn Helgaas on a 128-cpu setup. 7056 */ 7057 BUG_ON(busiest_rq == target_rq); 7058 7059 /* Search for an sd spanning us and the target CPU. */ 7060 rcu_read_lock(); 7061 for_each_domain(target_cpu, sd) { 7062 if ((sd->flags & SD_LOAD_BALANCE) && 7063 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd))) 7064 break; 7065 } 7066 7067 if (likely(sd)) { 7068 struct lb_env env = { 7069 .sd = sd, 7070 .dst_cpu = target_cpu, 7071 .dst_rq = target_rq, 7072 .src_cpu = busiest_rq->cpu, 7073 .src_rq = busiest_rq, 7074 .idle = CPU_IDLE, 7075 }; 7076 7077 schedstat_inc(sd, alb_count); 7078 7079 p = detach_one_task(&env); 7080 if (p) 7081 schedstat_inc(sd, alb_pushed); 7082 else 7083 schedstat_inc(sd, alb_failed); 7084 } 7085 rcu_read_unlock(); 7086out_unlock: 7087 busiest_rq->active_balance = 0; 7088 raw_spin_unlock(&busiest_rq->lock); 7089 7090 if (p) 7091 attach_one_task(target_rq, p); 7092 7093 local_irq_enable(); 7094 7095 return 0; 7096} 7097 7098static inline int on_null_domain(struct rq *rq) 7099{ 7100 return unlikely(!rcu_dereference_sched(rq->sd)); 7101} 7102 7103#ifdef CONFIG_NO_HZ_COMMON 7104/* 7105 * idle load balancing details 7106 * - When one of the busy CPUs notice that there may be an idle rebalancing 7107 * needed, they will kick the idle load balancer, which then does idle 7108 * load balancing for all the idle CPUs. 7109 */ 7110static struct { 7111 cpumask_var_t idle_cpus_mask; 7112 atomic_t nr_cpus; 7113 unsigned long next_balance; /* in jiffy units */ 7114} nohz ____cacheline_aligned; 7115 7116static inline int find_new_ilb(void) 7117{ 7118 int ilb = cpumask_first(nohz.idle_cpus_mask); 7119 7120 if (ilb < nr_cpu_ids && idle_cpu(ilb)) 7121 return ilb; 7122 7123 return nr_cpu_ids; 7124} 7125 7126/* 7127 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the 7128 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle 7129 * CPU (if there is one). 7130 */ 7131static void nohz_balancer_kick(void) 7132{ 7133 int ilb_cpu; 7134 7135 nohz.next_balance++; 7136 7137 ilb_cpu = find_new_ilb(); 7138 7139 if (ilb_cpu >= nr_cpu_ids) 7140 return; 7141 7142 if (test_and_set_bit(NOHZ_BALANCE_KICK, nohz_flags(ilb_cpu))) 7143 return; 7144 /* 7145 * Use smp_send_reschedule() instead of resched_cpu(). 7146 * This way we generate a sched IPI on the target cpu which 7147 * is idle. And the softirq performing nohz idle load balance 7148 * will be run before returning from the IPI. 7149 */ 7150 smp_send_reschedule(ilb_cpu); 7151 return; 7152} 7153 7154static inline void nohz_balance_exit_idle(int cpu) 7155{ 7156 if (unlikely(test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))) { 7157 /* 7158 * Completely isolated CPUs don't ever set, so we must test. 7159 */ 7160 if (likely(cpumask_test_cpu(cpu, nohz.idle_cpus_mask))) { 7161 cpumask_clear_cpu(cpu, nohz.idle_cpus_mask); 7162 atomic_dec(&nohz.nr_cpus); 7163 } 7164 clear_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)); 7165 } 7166} 7167 7168static inline void set_cpu_sd_state_busy(void) 7169{ 7170 struct sched_domain *sd; 7171 int cpu = smp_processor_id(); 7172 7173 rcu_read_lock(); 7174 sd = rcu_dereference(per_cpu(sd_busy, cpu)); 7175 7176 if (!sd || !sd->nohz_idle) 7177 goto unlock; 7178 sd->nohz_idle = 0; 7179 7180 atomic_inc(&sd->groups->sgc->nr_busy_cpus); 7181unlock: 7182 rcu_read_unlock(); 7183} 7184 7185void set_cpu_sd_state_idle(void) 7186{ 7187 struct sched_domain *sd; 7188 int cpu = smp_processor_id(); 7189 7190 rcu_read_lock(); 7191 sd = rcu_dereference(per_cpu(sd_busy, cpu)); 7192 7193 if (!sd || sd->nohz_idle) 7194 goto unlock; 7195 sd->nohz_idle = 1; 7196 7197 atomic_dec(&sd->groups->sgc->nr_busy_cpus); 7198unlock: 7199 rcu_read_unlock(); 7200} 7201 7202/* 7203 * This routine will record that the cpu is going idle with tick stopped. 7204 * This info will be used in performing idle load balancing in the future. 7205 */ 7206void nohz_balance_enter_idle(int cpu) 7207{ 7208 /* 7209 * If this cpu is going down, then nothing needs to be done. 7210 */ 7211 if (!cpu_active(cpu)) 7212 return; 7213 7214 if (test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu))) 7215 return; 7216 7217 /* 7218 * If we're a completely isolated CPU, we don't play. 7219 */ 7220 if (on_null_domain(cpu_rq(cpu))) 7221 return; 7222 7223 cpumask_set_cpu(cpu, nohz.idle_cpus_mask); 7224 atomic_inc(&nohz.nr_cpus); 7225 set_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)); 7226} 7227 7228static int sched_ilb_notifier(struct notifier_block *nfb, 7229 unsigned long action, void *hcpu) 7230{ 7231 switch (action & ~CPU_TASKS_FROZEN) { 7232 case CPU_DYING: 7233 nohz_balance_exit_idle(smp_processor_id()); 7234 return NOTIFY_OK; 7235 default: 7236 return NOTIFY_DONE; 7237 } 7238} 7239#endif 7240 7241static DEFINE_SPINLOCK(balancing); 7242 7243/* 7244 * Scale the max load_balance interval with the number of CPUs in the system. 7245 * This trades load-balance latency on larger machines for less cross talk. 7246 */ 7247void update_max_interval(void) 7248{ 7249 max_load_balance_interval = HZ*num_online_cpus()/10; 7250} 7251 7252/* 7253 * It checks each scheduling domain to see if it is due to be balanced, 7254 * and initiates a balancing operation if so. 7255 * 7256 * Balancing parameters are set up in init_sched_domains. 7257 */ 7258static void rebalance_domains(struct rq *rq, enum cpu_idle_type idle) 7259{ 7260 int continue_balancing = 1; 7261 int cpu = rq->cpu; 7262 unsigned long interval; 7263 struct sched_domain *sd; 7264 /* Earliest time when we have to do rebalance again */ 7265 unsigned long next_balance = jiffies + 60*HZ; 7266 int update_next_balance = 0; 7267 int need_serialize, need_decay = 0; 7268 u64 max_cost = 0; 7269 7270 update_blocked_averages(cpu); 7271 7272 rcu_read_lock(); 7273 for_each_domain(cpu, sd) { 7274 /* 7275 * Decay the newidle max times here because this is a regular 7276 * visit to all the domains. Decay ~1% per second. 7277 */ 7278 if (time_after(jiffies, sd->next_decay_max_lb_cost)) { 7279 sd->max_newidle_lb_cost = 7280 (sd->max_newidle_lb_cost * 253) / 256; 7281 sd->next_decay_max_lb_cost = jiffies + HZ; 7282 need_decay = 1; 7283 } 7284 max_cost += sd->max_newidle_lb_cost; 7285 7286 if (!(sd->flags & SD_LOAD_BALANCE)) 7287 continue; 7288 7289 /* 7290 * Stop the load balance at this level. There is another 7291 * CPU in our sched group which is doing load balancing more 7292 * actively. 7293 */ 7294 if (!continue_balancing) { 7295 if (need_decay) 7296 continue; 7297 break; 7298 } 7299 7300 interval = get_sd_balance_interval(sd, idle != CPU_IDLE); 7301 7302 need_serialize = sd->flags & SD_SERIALIZE; 7303 if (need_serialize) { 7304 if (!spin_trylock(&balancing)) 7305 goto out; 7306 } 7307 7308 if (time_after_eq(jiffies, sd->last_balance + interval)) { 7309 if (load_balance(cpu, rq, sd, idle, &continue_balancing)) { 7310 /* 7311 * The LBF_DST_PINNED logic could have changed 7312 * env->dst_cpu, so we can't know our idle 7313 * state even if we migrated tasks. Update it. 7314 */ 7315 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE; 7316 } 7317 sd->last_balance = jiffies; 7318 interval = get_sd_balance_interval(sd, idle != CPU_IDLE); 7319 } 7320 if (need_serialize) 7321 spin_unlock(&balancing); 7322out: 7323 if (time_after(next_balance, sd->last_balance + interval)) { 7324 next_balance = sd->last_balance + interval; 7325 update_next_balance = 1; 7326 } 7327 } 7328 if (need_decay) { 7329 /* 7330 * Ensure the rq-wide value also decays but keep it at a 7331 * reasonable floor to avoid funnies with rq->avg_idle. 7332 */ 7333 rq->max_idle_balance_cost = 7334 max((u64)sysctl_sched_migration_cost, max_cost); 7335 } 7336 rcu_read_unlock(); 7337 7338 /* 7339 * next_balance will be updated only when there is a need. 7340 * When the cpu is attached to null domain for ex, it will not be 7341 * updated. 7342 */ 7343 if (likely(update_next_balance)) 7344 rq->next_balance = next_balance; 7345} 7346 7347#ifdef CONFIG_NO_HZ_COMMON 7348/* 7349 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the 7350 * rebalancing for all the cpus for whom scheduler ticks are stopped. 7351 */ 7352static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle) 7353{ 7354 int this_cpu = this_rq->cpu; 7355 struct rq *rq; 7356 int balance_cpu; 7357 7358 if (idle != CPU_IDLE || 7359 !test_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu))) 7360 goto end; 7361 7362 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) { 7363 if (balance_cpu == this_cpu || !idle_cpu(balance_cpu)) 7364 continue; 7365 7366 /* 7367 * If this cpu gets work to do, stop the load balancing 7368 * work being done for other cpus. Next load 7369 * balancing owner will pick it up. 7370 */ 7371 if (need_resched()) 7372 break; 7373 7374 rq = cpu_rq(balance_cpu); 7375 7376 /* 7377 * If time for next balance is due, 7378 * do the balance. 7379 */ 7380 if (time_after_eq(jiffies, rq->next_balance)) { 7381 raw_spin_lock_irq(&rq->lock); 7382 update_rq_clock(rq); 7383 update_idle_cpu_load(rq); 7384 raw_spin_unlock_irq(&rq->lock); 7385 rebalance_domains(rq, CPU_IDLE); 7386 } 7387 7388 if (time_after(this_rq->next_balance, rq->next_balance)) 7389 this_rq->next_balance = rq->next_balance; 7390 } 7391 nohz.next_balance = this_rq->next_balance; 7392end: 7393 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu)); 7394} 7395 7396/* 7397 * Current heuristic for kicking the idle load balancer in the presence 7398 * of an idle cpu is the system. 7399 * - This rq has more than one task. 7400 * - At any scheduler domain level, this cpu's scheduler group has multiple 7401 * busy cpu's exceeding the group's capacity. 7402 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler 7403 * domain span are idle. 7404 */ 7405static inline int nohz_kick_needed(struct rq *rq) 7406{ 7407 unsigned long now = jiffies; 7408 struct sched_domain *sd; 7409 struct sched_group_capacity *sgc; 7410 int nr_busy, cpu = rq->cpu; 7411 7412 if (unlikely(rq->idle_balance)) 7413 return 0; 7414 7415 /* 7416 * We may be recently in ticked or tickless idle mode. At the first 7417 * busy tick after returning from idle, we will update the busy stats. 7418 */ 7419 set_cpu_sd_state_busy(); 7420 nohz_balance_exit_idle(cpu); 7421 7422 /* 7423 * None are in tickless mode and hence no need for NOHZ idle load 7424 * balancing. 7425 */ 7426 if (likely(!atomic_read(&nohz.nr_cpus))) 7427 return 0; 7428 7429 if (time_before(now, nohz.next_balance)) 7430 return 0; 7431 7432 if (rq->nr_running >= 2) 7433 goto need_kick; 7434 7435 rcu_read_lock(); 7436 sd = rcu_dereference(per_cpu(sd_busy, cpu)); 7437 7438 if (sd) { 7439 sgc = sd->groups->sgc; 7440 nr_busy = atomic_read(&sgc->nr_busy_cpus); 7441 7442 if (nr_busy > 1) 7443 goto need_kick_unlock; 7444 } 7445 7446 sd = rcu_dereference(per_cpu(sd_asym, cpu)); 7447 7448 if (sd && (cpumask_first_and(nohz.idle_cpus_mask, 7449 sched_domain_span(sd)) < cpu)) 7450 goto need_kick_unlock; 7451 7452 rcu_read_unlock(); 7453 return 0; 7454 7455need_kick_unlock: 7456 rcu_read_unlock(); 7457need_kick: 7458 return 1; 7459} 7460#else 7461static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle) { } 7462#endif 7463 7464/* 7465 * run_rebalance_domains is triggered when needed from the scheduler tick. 7466 * Also triggered for nohz idle balancing (with nohz_balancing_kick set). 7467 */ 7468static void run_rebalance_domains(struct softirq_action *h) 7469{ 7470 struct rq *this_rq = this_rq(); 7471 enum cpu_idle_type idle = this_rq->idle_balance ? 7472 CPU_IDLE : CPU_NOT_IDLE; 7473 7474 rebalance_domains(this_rq, idle); 7475 7476 /* 7477 * If this cpu has a pending nohz_balance_kick, then do the 7478 * balancing on behalf of the other idle cpus whose ticks are 7479 * stopped. 7480 */ 7481 nohz_idle_balance(this_rq, idle); 7482} 7483 7484/* 7485 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing. 7486 */ 7487void trigger_load_balance(struct rq *rq) 7488{ 7489 /* Don't need to rebalance while attached to NULL domain */ 7490 if (unlikely(on_null_domain(rq))) 7491 return; 7492 7493 if (time_after_eq(jiffies, rq->next_balance)) 7494 raise_softirq(SCHED_SOFTIRQ); 7495#ifdef CONFIG_NO_HZ_COMMON 7496 if (nohz_kick_needed(rq)) 7497 nohz_balancer_kick(); 7498#endif 7499} 7500 7501static void rq_online_fair(struct rq *rq) 7502{ 7503 update_sysctl(); 7504 7505 update_runtime_enabled(rq); 7506} 7507 7508static void rq_offline_fair(struct rq *rq) 7509{ 7510 update_sysctl(); 7511 7512 /* Ensure any throttled groups are reachable by pick_next_task */ 7513 unthrottle_offline_cfs_rqs(rq); 7514} 7515 7516#endif /* CONFIG_SMP */ 7517 7518/* 7519 * scheduler tick hitting a task of our scheduling class: 7520 */ 7521static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued) 7522{ 7523 struct cfs_rq *cfs_rq; 7524 struct sched_entity *se = &curr->se; 7525 7526 for_each_sched_entity(se) { 7527 cfs_rq = cfs_rq_of(se); 7528 entity_tick(cfs_rq, se, queued); 7529 } 7530 7531 if (numabalancing_enabled) 7532 task_tick_numa(rq, curr); 7533 7534 update_rq_runnable_avg(rq, 1); 7535} 7536 7537/* 7538 * called on fork with the child task as argument from the parent's context 7539 * - child not yet on the tasklist 7540 * - preemption disabled 7541 */ 7542static void task_fork_fair(struct task_struct *p) 7543{ 7544 struct cfs_rq *cfs_rq; 7545 struct sched_entity *se = &p->se, *curr; 7546 int this_cpu = smp_processor_id(); 7547 struct rq *rq = this_rq(); 7548 unsigned long flags; 7549 7550 raw_spin_lock_irqsave(&rq->lock, flags); 7551 7552 update_rq_clock(rq); 7553 7554 cfs_rq = task_cfs_rq(current); 7555 curr = cfs_rq->curr; 7556 7557 /* 7558 * Not only the cpu but also the task_group of the parent might have 7559 * been changed after parent->se.parent,cfs_rq were copied to 7560 * child->se.parent,cfs_rq. So call __set_task_cpu() to make those 7561 * of child point to valid ones. 7562 */ 7563 rcu_read_lock(); 7564 __set_task_cpu(p, this_cpu); 7565 rcu_read_unlock(); 7566 7567 update_curr(cfs_rq); 7568 7569 if (curr) 7570 se->vruntime = curr->vruntime; 7571 place_entity(cfs_rq, se, 1); 7572 7573 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) { 7574 /* 7575 * Upon rescheduling, sched_class::put_prev_task() will place 7576 * 'current' within the tree based on its new key value. 7577 */ 7578 swap(curr->vruntime, se->vruntime); 7579 resched_curr(rq); 7580 } 7581 7582 se->vruntime -= cfs_rq->min_vruntime; 7583 7584 raw_spin_unlock_irqrestore(&rq->lock, flags); 7585} 7586 7587/* 7588 * Priority of the task has changed. Check to see if we preempt 7589 * the current task. 7590 */ 7591static void 7592prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio) 7593{ 7594 if (!task_on_rq_queued(p)) 7595 return; 7596 7597 /* 7598 * Reschedule if we are currently running on this runqueue and 7599 * our priority decreased, or if we are not currently running on 7600 * this runqueue and our priority is higher than the current's 7601 */ 7602 if (rq->curr == p) { 7603 if (p->prio > oldprio) 7604 resched_curr(rq); 7605 } else 7606 check_preempt_curr(rq, p, 0); 7607} 7608 7609static void switched_from_fair(struct rq *rq, struct task_struct *p) 7610{ 7611 struct sched_entity *se = &p->se; 7612 struct cfs_rq *cfs_rq = cfs_rq_of(se); 7613 7614 /* 7615 * Ensure the task's vruntime is normalized, so that when it's 7616 * switched back to the fair class the enqueue_entity(.flags=0) will 7617 * do the right thing. 7618 * 7619 * If it's queued, then the dequeue_entity(.flags=0) will already 7620 * have normalized the vruntime, if it's !queued, then only when 7621 * the task is sleeping will it still have non-normalized vruntime. 7622 */ 7623 if (!task_on_rq_queued(p) && p->state != TASK_RUNNING) { 7624 /* 7625 * Fix up our vruntime so that the current sleep doesn't 7626 * cause 'unlimited' sleep bonus. 7627 */ 7628 place_entity(cfs_rq, se, 0); 7629 se->vruntime -= cfs_rq->min_vruntime; 7630 } 7631 7632#ifdef CONFIG_SMP 7633 /* 7634 * Remove our load from contribution when we leave sched_fair 7635 * and ensure we don't carry in an old decay_count if we 7636 * switch back. 7637 */ 7638 if (se->avg.decay_count) { 7639 __synchronize_entity_decay(se); 7640 subtract_blocked_load_contrib(cfs_rq, se->avg.load_avg_contrib); 7641 } 7642#endif 7643} 7644 7645/* 7646 * We switched to the sched_fair class. 7647 */ 7648static void switched_to_fair(struct rq *rq, struct task_struct *p) 7649{ 7650#ifdef CONFIG_FAIR_GROUP_SCHED 7651 struct sched_entity *se = &p->se; 7652 /* 7653 * Since the real-depth could have been changed (only FAIR 7654 * class maintain depth value), reset depth properly. 7655 */ 7656 se->depth = se->parent ? se->parent->depth + 1 : 0; 7657#endif 7658 if (!task_on_rq_queued(p)) 7659 return; 7660 7661 /* 7662 * We were most likely switched from sched_rt, so 7663 * kick off the schedule if running, otherwise just see 7664 * if we can still preempt the current task. 7665 */ 7666 if (rq->curr == p) 7667 resched_curr(rq); 7668 else 7669 check_preempt_curr(rq, p, 0); 7670} 7671 7672/* Account for a task changing its policy or group. 7673 * 7674 * This routine is mostly called to set cfs_rq->curr field when a task 7675 * migrates between groups/classes. 7676 */ 7677static void set_curr_task_fair(struct rq *rq) 7678{ 7679 struct sched_entity *se = &rq->curr->se; 7680 7681 for_each_sched_entity(se) { 7682 struct cfs_rq *cfs_rq = cfs_rq_of(se); 7683 7684 set_next_entity(cfs_rq, se); 7685 /* ensure bandwidth has been allocated on our new cfs_rq */ 7686 account_cfs_rq_runtime(cfs_rq, 0); 7687 } 7688} 7689 7690void init_cfs_rq(struct cfs_rq *cfs_rq) 7691{ 7692 cfs_rq->tasks_timeline = RB_ROOT; 7693 cfs_rq->min_vruntime = (u64)(-(1LL << 20)); 7694#ifndef CONFIG_64BIT 7695 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime; 7696#endif 7697#ifdef CONFIG_SMP 7698 atomic64_set(&cfs_rq->decay_counter, 1); 7699 atomic_long_set(&cfs_rq->removed_load, 0); 7700#endif 7701} 7702 7703#ifdef CONFIG_FAIR_GROUP_SCHED 7704static void task_move_group_fair(struct task_struct *p, int queued) 7705{ 7706 struct sched_entity *se = &p->se; 7707 struct cfs_rq *cfs_rq; 7708 7709 /* 7710 * If the task was not on the rq at the time of this cgroup movement 7711 * it must have been asleep, sleeping tasks keep their ->vruntime 7712 * absolute on their old rq until wakeup (needed for the fair sleeper 7713 * bonus in place_entity()). 7714 * 7715 * If it was on the rq, we've just 'preempted' it, which does convert 7716 * ->vruntime to a relative base. 7717 * 7718 * Make sure both cases convert their relative position when migrating 7719 * to another cgroup's rq. This does somewhat interfere with the 7720 * fair sleeper stuff for the first placement, but who cares. 7721 */ 7722 /* 7723 * When !queued, vruntime of the task has usually NOT been normalized. 7724 * But there are some cases where it has already been normalized: 7725 * 7726 * - Moving a forked child which is waiting for being woken up by 7727 * wake_up_new_task(). 7728 * - Moving a task which has been woken up by try_to_wake_up() and 7729 * waiting for actually being woken up by sched_ttwu_pending(). 7730 * 7731 * To prevent boost or penalty in the new cfs_rq caused by delta 7732 * min_vruntime between the two cfs_rqs, we skip vruntime adjustment. 7733 */ 7734 if (!queued && (!se->sum_exec_runtime || p->state == TASK_WAKING)) 7735 queued = 1; 7736 7737 if (!queued) 7738 se->vruntime -= cfs_rq_of(se)->min_vruntime; 7739 set_task_rq(p, task_cpu(p)); 7740 se->depth = se->parent ? se->parent->depth + 1 : 0; 7741 if (!queued) { 7742 cfs_rq = cfs_rq_of(se); 7743 se->vruntime += cfs_rq->min_vruntime; 7744#ifdef CONFIG_SMP 7745 /* 7746 * migrate_task_rq_fair() will have removed our previous 7747 * contribution, but we must synchronize for ongoing future 7748 * decay. 7749 */ 7750 se->avg.decay_count = atomic64_read(&cfs_rq->decay_counter); 7751 cfs_rq->blocked_load_avg += se->avg.load_avg_contrib; 7752#endif 7753 } 7754} 7755 7756void free_fair_sched_group(struct task_group *tg) 7757{ 7758 int i; 7759 7760 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg)); 7761 7762 for_each_possible_cpu(i) { 7763 if (tg->cfs_rq) 7764 kfree(tg->cfs_rq[i]); 7765 if (tg->se) 7766 kfree(tg->se[i]); 7767 } 7768 7769 kfree(tg->cfs_rq); 7770 kfree(tg->se); 7771} 7772 7773int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent) 7774{ 7775 struct cfs_rq *cfs_rq; 7776 struct sched_entity *se; 7777 int i; 7778 7779 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL); 7780 if (!tg->cfs_rq) 7781 goto err; 7782 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL); 7783 if (!tg->se) 7784 goto err; 7785 7786 tg->shares = NICE_0_LOAD; 7787 7788 init_cfs_bandwidth(tg_cfs_bandwidth(tg)); 7789 7790 for_each_possible_cpu(i) { 7791 cfs_rq = kzalloc_node(sizeof(struct cfs_rq), 7792 GFP_KERNEL, cpu_to_node(i)); 7793 if (!cfs_rq) 7794 goto err; 7795 7796 se = kzalloc_node(sizeof(struct sched_entity), 7797 GFP_KERNEL, cpu_to_node(i)); 7798 if (!se) 7799 goto err_free_rq; 7800 7801 init_cfs_rq(cfs_rq); 7802 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]); 7803 } 7804 7805 return 1; 7806 7807err_free_rq: 7808 kfree(cfs_rq); 7809err: 7810 return 0; 7811} 7812 7813void unregister_fair_sched_group(struct task_group *tg, int cpu) 7814{ 7815 struct rq *rq = cpu_rq(cpu); 7816 unsigned long flags; 7817 7818 /* 7819 * Only empty task groups can be destroyed; so we can speculatively 7820 * check on_list without danger of it being re-added. 7821 */ 7822 if (!tg->cfs_rq[cpu]->on_list) 7823 return; 7824 7825 raw_spin_lock_irqsave(&rq->lock, flags); 7826 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]); 7827 raw_spin_unlock_irqrestore(&rq->lock, flags); 7828} 7829 7830void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq, 7831 struct sched_entity *se, int cpu, 7832 struct sched_entity *parent) 7833{ 7834 struct rq *rq = cpu_rq(cpu); 7835 7836 cfs_rq->tg = tg; 7837 cfs_rq->rq = rq; 7838 init_cfs_rq_runtime(cfs_rq); 7839 7840 tg->cfs_rq[cpu] = cfs_rq; 7841 tg->se[cpu] = se; 7842 7843 /* se could be NULL for root_task_group */ 7844 if (!se) 7845 return; 7846 7847 if (!parent) { 7848 se->cfs_rq = &rq->cfs; 7849 se->depth = 0; 7850 } else { 7851 se->cfs_rq = parent->my_q; 7852 se->depth = parent->depth + 1; 7853 } 7854 7855 se->my_q = cfs_rq; 7856 /* guarantee group entities always have weight */ 7857 update_load_set(&se->load, NICE_0_LOAD); 7858 se->parent = parent; 7859} 7860 7861static DEFINE_MUTEX(shares_mutex); 7862 7863int sched_group_set_shares(struct task_group *tg, unsigned long shares) 7864{ 7865 int i; 7866 unsigned long flags; 7867 7868 /* 7869 * We can't change the weight of the root cgroup. 7870 */ 7871 if (!tg->se[0]) 7872 return -EINVAL; 7873 7874 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES)); 7875 7876 mutex_lock(&shares_mutex); 7877 if (tg->shares == shares) 7878 goto done; 7879 7880 tg->shares = shares; 7881 for_each_possible_cpu(i) { 7882 struct rq *rq = cpu_rq(i); 7883 struct sched_entity *se; 7884 7885 se = tg->se[i]; 7886 /* Propagate contribution to hierarchy */ 7887 raw_spin_lock_irqsave(&rq->lock, flags); 7888 7889 /* Possible calls to update_curr() need rq clock */ 7890 update_rq_clock(rq); 7891 for_each_sched_entity(se) 7892 update_cfs_shares(group_cfs_rq(se)); 7893 raw_spin_unlock_irqrestore(&rq->lock, flags); 7894 } 7895 7896done: 7897 mutex_unlock(&shares_mutex); 7898 return 0; 7899} 7900#else /* CONFIG_FAIR_GROUP_SCHED */ 7901 7902void free_fair_sched_group(struct task_group *tg) { } 7903 7904int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent) 7905{ 7906 return 1; 7907} 7908 7909void unregister_fair_sched_group(struct task_group *tg, int cpu) { } 7910 7911#endif /* CONFIG_FAIR_GROUP_SCHED */ 7912 7913 7914static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task) 7915{ 7916 struct sched_entity *se = &task->se; 7917 unsigned int rr_interval = 0; 7918 7919 /* 7920 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise 7921 * idle runqueue: 7922 */ 7923 if (rq->cfs.load.weight) 7924 rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se)); 7925 7926 return rr_interval; 7927} 7928 7929/* 7930 * All the scheduling class methods: 7931 */ 7932const struct sched_class fair_sched_class = { 7933 .next = &idle_sched_class, 7934 .enqueue_task = enqueue_task_fair, 7935 .dequeue_task = dequeue_task_fair, 7936 .yield_task = yield_task_fair, 7937 .yield_to_task = yield_to_task_fair, 7938 7939 .check_preempt_curr = check_preempt_wakeup, 7940 7941 .pick_next_task = pick_next_task_fair, 7942 .put_prev_task = put_prev_task_fair, 7943 7944#ifdef CONFIG_SMP 7945 .select_task_rq = select_task_rq_fair, 7946 .migrate_task_rq = migrate_task_rq_fair, 7947 7948 .rq_online = rq_online_fair, 7949 .rq_offline = rq_offline_fair, 7950 7951 .task_waking = task_waking_fair, 7952#endif 7953 7954 .set_curr_task = set_curr_task_fair, 7955 .task_tick = task_tick_fair, 7956 .task_fork = task_fork_fair, 7957 7958 .prio_changed = prio_changed_fair, 7959 .switched_from = switched_from_fair, 7960 .switched_to = switched_to_fair, 7961 7962 .get_rr_interval = get_rr_interval_fair, 7963 7964 .update_curr = update_curr_fair, 7965 7966#ifdef CONFIG_FAIR_GROUP_SCHED 7967 .task_move_group = task_move_group_fair, 7968#endif 7969}; 7970 7971#ifdef CONFIG_SCHED_DEBUG 7972void print_cfs_stats(struct seq_file *m, int cpu) 7973{ 7974 struct cfs_rq *cfs_rq; 7975 7976 rcu_read_lock(); 7977 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq) 7978 print_cfs_rq(m, cpu, cfs_rq); 7979 rcu_read_unlock(); 7980} 7981#endif 7982 7983__init void init_sched_fair_class(void) 7984{ 7985#ifdef CONFIG_SMP 7986 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains); 7987 7988#ifdef CONFIG_NO_HZ_COMMON 7989 nohz.next_balance = jiffies; 7990 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT); 7991 cpu_notifier(sched_ilb_notifier, 0); 7992#endif 7993#endif /* SMP */ 7994 7995} 7996