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 */ 50 unsigned int sysctl_sched_latency = 6000000ULL; 51 unsigned 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 */ 62 enum 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 */ 69 unsigned int sysctl_sched_min_granularity = 750000ULL; 70 unsigned int normalized_sysctl_sched_min_granularity = 750000ULL; 71 72 /* 73 * is kept at sysctl_sched_latency / sysctl_sched_min_granularity 74 */ 75 static 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 */ 81 unsigned 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 */ 91 unsigned int sysctl_sched_wakeup_granularity = 1000000UL; 92 unsigned int normalized_sysctl_sched_wakeup_granularity = 1000000UL; 93 94 const_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 */ 101 unsigned 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 */ 114 unsigned int sysctl_sched_cfs_bandwidth_slice = 5000UL; 115 #endif 116 117 static inline void update_load_add(struct load_weight *lw, unsigned long inc) 118 { 119 lw->weight += inc; 120 lw->inv_weight = 0; 121 } 122 123 static inline void update_load_sub(struct load_weight *lw, unsigned long dec) 124 { 125 lw->weight -= dec; 126 lw->inv_weight = 0; 127 } 128 129 static 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 */ 144 static 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 } 164 165 static 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 } 176 177 void sched_init_granularity(void) 178 { 179 update_sysctl(); 180 } 181 182 #define WMULT_CONST (~0U) 183 #define WMULT_SHIFT 32 184 185 static 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 */ 214 static 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 } 238 239 240 const 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 */ 249 static 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) 256 257 static 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) 268 269 static 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 */ 275 static 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 */ 281 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp) 282 { 283 return grp->my_q; 284 } 285 286 static void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq, 287 int force_update); 288 289 static 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 } 312 313 static 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 ? */ 326 static inline struct cfs_rq * 327 is_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 } 334 335 static inline struct sched_entity *parent_entity(struct sched_entity *se) 336 { 337 return se->parent; 338 } 339 340 static void 341 find_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 */ 373 374 static inline struct task_struct *task_of(struct sched_entity *se) 375 { 376 return container_of(se, struct task_struct, se); 377 } 378 379 static 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) 388 389 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p) 390 { 391 return &task_rq(p)->cfs; 392 } 393 394 static 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 */ 403 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp) 404 { 405 return NULL; 406 } 407 408 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq) 409 { 410 } 411 412 static 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) 418 419 static inline struct sched_entity *parent_entity(struct sched_entity *se) 420 { 421 return NULL; 422 } 423 424 static inline void 425 find_matching_se(struct sched_entity **se, struct sched_entity **pse) 426 { 427 } 428 429 #endif /* CONFIG_FAIR_GROUP_SCHED */ 430 431 static __always_inline 432 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec); 433 434 /************************************************************** 435 * Scheduling class tree data structure manipulation methods: 436 */ 437 438 static 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 } 446 447 static 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 } 455 456 static inline int entity_before(struct sched_entity *a, 457 struct sched_entity *b) 458 { 459 return (s64)(a->vruntime - b->vruntime) < 0; 460 } 461 462 static 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 */ 491 static 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 } 526 527 static 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 } 538 539 struct 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 } 548 549 static 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 560 struct 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 */ 573 574 int 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 */ 601 static 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 */ 617 static 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 */ 636 static 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 */ 663 static 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 669 static int select_idle_sibling(struct task_struct *p, int cpu); 670 static unsigned long task_h_load(struct task_struct *p); 671 672 static 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 */ 675 void init_task_runnable_average(struct task_struct *p) 676 { 677 u32 slice; 678 679 slice = sched_slice(task_cfs_rq(p), &p->se) >> 10; 680 p->se.avg.runnable_avg_sum = slice; 681 p->se.avg.runnable_avg_period = slice; 682 __update_task_entity_contrib(&p->se); 683 } 684 #else 685 void init_task_runnable_average(struct task_struct *p) 686 { 687 } 688 #endif 689 690 /* 691 * Update the current task's runtime statistics. 692 */ 693 static void update_curr(struct cfs_rq *cfs_rq) 694 { 695 struct sched_entity *curr = cfs_rq->curr; 696 u64 now = rq_clock_task(rq_of(cfs_rq)); 697 u64 delta_exec; 698 699 if (unlikely(!curr)) 700 return; 701 702 delta_exec = now - curr->exec_start; 703 if (unlikely((s64)delta_exec <= 0)) 704 return; 705 706 curr->exec_start = now; 707 708 schedstat_set(curr->statistics.exec_max, 709 max(delta_exec, curr->statistics.exec_max)); 710 711 curr->sum_exec_runtime += delta_exec; 712 schedstat_add(cfs_rq, exec_clock, delta_exec); 713 714 curr->vruntime += calc_delta_fair(delta_exec, curr); 715 update_min_vruntime(cfs_rq); 716 717 if (entity_is_task(curr)) { 718 struct task_struct *curtask = task_of(curr); 719 720 trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime); 721 cpuacct_charge(curtask, delta_exec); 722 account_group_exec_runtime(curtask, delta_exec); 723 } 724 725 account_cfs_rq_runtime(cfs_rq, delta_exec); 726 } 727 728 static void update_curr_fair(struct rq *rq) 729 { 730 update_curr(cfs_rq_of(&rq->curr->se)); 731 } 732 733 static inline void 734 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se) 735 { 736 schedstat_set(se->statistics.wait_start, rq_clock(rq_of(cfs_rq))); 737 } 738 739 /* 740 * Task is being enqueued - update stats: 741 */ 742 static void update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se) 743 { 744 /* 745 * Are we enqueueing a waiting task? (for current tasks 746 * a dequeue/enqueue event is a NOP) 747 */ 748 if (se != cfs_rq->curr) 749 update_stats_wait_start(cfs_rq, se); 750 } 751 752 static void 753 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se) 754 { 755 schedstat_set(se->statistics.wait_max, max(se->statistics.wait_max, 756 rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start)); 757 schedstat_set(se->statistics.wait_count, se->statistics.wait_count + 1); 758 schedstat_set(se->statistics.wait_sum, se->statistics.wait_sum + 759 rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start); 760 #ifdef CONFIG_SCHEDSTATS 761 if (entity_is_task(se)) { 762 trace_sched_stat_wait(task_of(se), 763 rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start); 764 } 765 #endif 766 schedstat_set(se->statistics.wait_start, 0); 767 } 768 769 static inline void 770 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se) 771 { 772 /* 773 * Mark the end of the wait period if dequeueing a 774 * waiting task: 775 */ 776 if (se != cfs_rq->curr) 777 update_stats_wait_end(cfs_rq, se); 778 } 779 780 /* 781 * We are picking a new current task - update its stats: 782 */ 783 static inline void 784 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se) 785 { 786 /* 787 * We are starting a new run period: 788 */ 789 se->exec_start = rq_clock_task(rq_of(cfs_rq)); 790 } 791 792 /************************************************** 793 * Scheduling class queueing methods: 794 */ 795 796 #ifdef CONFIG_NUMA_BALANCING 797 /* 798 * Approximate time to scan a full NUMA task in ms. The task scan period is 799 * calculated based on the tasks virtual memory size and 800 * numa_balancing_scan_size. 801 */ 802 unsigned int sysctl_numa_balancing_scan_period_min = 1000; 803 unsigned int sysctl_numa_balancing_scan_period_max = 60000; 804 805 /* Portion of address space to scan in MB */ 806 unsigned int sysctl_numa_balancing_scan_size = 256; 807 808 /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */ 809 unsigned int sysctl_numa_balancing_scan_delay = 1000; 810 811 static unsigned int task_nr_scan_windows(struct task_struct *p) 812 { 813 unsigned long rss = 0; 814 unsigned long nr_scan_pages; 815 816 /* 817 * Calculations based on RSS as non-present and empty pages are skipped 818 * by the PTE scanner and NUMA hinting faults should be trapped based 819 * on resident pages 820 */ 821 nr_scan_pages = sysctl_numa_balancing_scan_size << (20 - PAGE_SHIFT); 822 rss = get_mm_rss(p->mm); 823 if (!rss) 824 rss = nr_scan_pages; 825 826 rss = round_up(rss, nr_scan_pages); 827 return rss / nr_scan_pages; 828 } 829 830 /* For sanitys sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */ 831 #define MAX_SCAN_WINDOW 2560 832 833 static unsigned int task_scan_min(struct task_struct *p) 834 { 835 unsigned int scan_size = ACCESS_ONCE(sysctl_numa_balancing_scan_size); 836 unsigned int scan, floor; 837 unsigned int windows = 1; 838 839 if (scan_size < MAX_SCAN_WINDOW) 840 windows = MAX_SCAN_WINDOW / scan_size; 841 floor = 1000 / windows; 842 843 scan = sysctl_numa_balancing_scan_period_min / task_nr_scan_windows(p); 844 return max_t(unsigned int, floor, scan); 845 } 846 847 static unsigned int task_scan_max(struct task_struct *p) 848 { 849 unsigned int smin = task_scan_min(p); 850 unsigned int smax; 851 852 /* Watch for min being lower than max due to floor calculations */ 853 smax = sysctl_numa_balancing_scan_period_max / task_nr_scan_windows(p); 854 return max(smin, smax); 855 } 856 857 static void account_numa_enqueue(struct rq *rq, struct task_struct *p) 858 { 859 rq->nr_numa_running += (p->numa_preferred_nid != -1); 860 rq->nr_preferred_running += (p->numa_preferred_nid == task_node(p)); 861 } 862 863 static void account_numa_dequeue(struct rq *rq, struct task_struct *p) 864 { 865 rq->nr_numa_running -= (p->numa_preferred_nid != -1); 866 rq->nr_preferred_running -= (p->numa_preferred_nid == task_node(p)); 867 } 868 869 struct numa_group { 870 atomic_t refcount; 871 872 spinlock_t lock; /* nr_tasks, tasks */ 873 int nr_tasks; 874 pid_t gid; 875 876 struct rcu_head rcu; 877 nodemask_t active_nodes; 878 unsigned long total_faults; 879 /* 880 * Faults_cpu is used to decide whether memory should move 881 * towards the CPU. As a consequence, these stats are weighted 882 * more by CPU use than by memory faults. 883 */ 884 unsigned long *faults_cpu; 885 unsigned long faults[0]; 886 }; 887 888 /* Shared or private faults. */ 889 #define NR_NUMA_HINT_FAULT_TYPES 2 890 891 /* Memory and CPU locality */ 892 #define NR_NUMA_HINT_FAULT_STATS (NR_NUMA_HINT_FAULT_TYPES * 2) 893 894 /* Averaged statistics, and temporary buffers. */ 895 #define NR_NUMA_HINT_FAULT_BUCKETS (NR_NUMA_HINT_FAULT_STATS * 2) 896 897 pid_t task_numa_group_id(struct task_struct *p) 898 { 899 return p->numa_group ? p->numa_group->gid : 0; 900 } 901 902 /* 903 * The averaged statistics, shared & private, memory & cpu, 904 * occupy the first half of the array. The second half of the 905 * array is for current counters, which are averaged into the 906 * first set by task_numa_placement. 907 */ 908 static inline int task_faults_idx(enum numa_faults_stats s, int nid, int priv) 909 { 910 return NR_NUMA_HINT_FAULT_TYPES * (s * nr_node_ids + nid) + priv; 911 } 912 913 static inline unsigned long task_faults(struct task_struct *p, int nid) 914 { 915 if (!p->numa_faults) 916 return 0; 917 918 return p->numa_faults[task_faults_idx(NUMA_MEM, nid, 0)] + 919 p->numa_faults[task_faults_idx(NUMA_MEM, nid, 1)]; 920 } 921 922 static inline unsigned long group_faults(struct task_struct *p, int nid) 923 { 924 if (!p->numa_group) 925 return 0; 926 927 return p->numa_group->faults[task_faults_idx(NUMA_MEM, nid, 0)] + 928 p->numa_group->faults[task_faults_idx(NUMA_MEM, nid, 1)]; 929 } 930 931 static inline unsigned long group_faults_cpu(struct numa_group *group, int nid) 932 { 933 return group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 0)] + 934 group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 1)]; 935 } 936 937 /* Handle placement on systems where not all nodes are directly connected. */ 938 static unsigned long score_nearby_nodes(struct task_struct *p, int nid, 939 int maxdist, bool task) 940 { 941 unsigned long score = 0; 942 int node; 943 944 /* 945 * All nodes are directly connected, and the same distance 946 * from each other. No need for fancy placement algorithms. 947 */ 948 if (sched_numa_topology_type == NUMA_DIRECT) 949 return 0; 950 951 /* 952 * This code is called for each node, introducing N^2 complexity, 953 * which should be ok given the number of nodes rarely exceeds 8. 954 */ 955 for_each_online_node(node) { 956 unsigned long faults; 957 int dist = node_distance(nid, node); 958 959 /* 960 * The furthest away nodes in the system are not interesting 961 * for placement; nid was already counted. 962 */ 963 if (dist == sched_max_numa_distance || node == nid) 964 continue; 965 966 /* 967 * On systems with a backplane NUMA topology, compare groups 968 * of nodes, and move tasks towards the group with the most 969 * memory accesses. When comparing two nodes at distance 970 * "hoplimit", only nodes closer by than "hoplimit" are part 971 * of each group. Skip other nodes. 972 */ 973 if (sched_numa_topology_type == NUMA_BACKPLANE && 974 dist > maxdist) 975 continue; 976 977 /* Add up the faults from nearby nodes. */ 978 if (task) 979 faults = task_faults(p, node); 980 else 981 faults = group_faults(p, node); 982 983 /* 984 * On systems with a glueless mesh NUMA topology, there are 985 * no fixed "groups of nodes". Instead, nodes that are not 986 * directly connected bounce traffic through intermediate 987 * nodes; a numa_group can occupy any set of nodes. 988 * The further away a node is, the less the faults count. 989 * This seems to result in good task placement. 990 */ 991 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) { 992 faults *= (sched_max_numa_distance - dist); 993 faults /= (sched_max_numa_distance - LOCAL_DISTANCE); 994 } 995 996 score += faults; 997 } 998 999 return score; 1000 } 1001 1002 /* 1003 * These return the fraction of accesses done by a particular task, or 1004 * task group, on a particular numa node. The group weight is given a 1005 * larger multiplier, in order to group tasks together that are almost 1006 * evenly spread out between numa nodes. 1007 */ 1008 static inline unsigned long task_weight(struct task_struct *p, int nid, 1009 int dist) 1010 { 1011 unsigned long faults, total_faults; 1012 1013 if (!p->numa_faults) 1014 return 0; 1015 1016 total_faults = p->total_numa_faults; 1017 1018 if (!total_faults) 1019 return 0; 1020 1021 faults = task_faults(p, nid); 1022 faults += score_nearby_nodes(p, nid, dist, true); 1023 1024 return 1000 * faults / total_faults; 1025 } 1026 1027 static inline unsigned long group_weight(struct task_struct *p, int nid, 1028 int dist) 1029 { 1030 unsigned long faults, total_faults; 1031 1032 if (!p->numa_group) 1033 return 0; 1034 1035 total_faults = p->numa_group->total_faults; 1036 1037 if (!total_faults) 1038 return 0; 1039 1040 faults = group_faults(p, nid); 1041 faults += score_nearby_nodes(p, nid, dist, false); 1042 1043 return 1000 * faults / total_faults; 1044 } 1045 1046 bool should_numa_migrate_memory(struct task_struct *p, struct page * page, 1047 int src_nid, int dst_cpu) 1048 { 1049 struct numa_group *ng = p->numa_group; 1050 int dst_nid = cpu_to_node(dst_cpu); 1051 int last_cpupid, this_cpupid; 1052 1053 this_cpupid = cpu_pid_to_cpupid(dst_cpu, current->pid); 1054 1055 /* 1056 * Multi-stage node selection is used in conjunction with a periodic 1057 * migration fault to build a temporal task<->page relation. By using 1058 * a two-stage filter we remove short/unlikely relations. 1059 * 1060 * Using P(p) ~ n_p / n_t as per frequentist probability, we can equate 1061 * a task's usage of a particular page (n_p) per total usage of this 1062 * page (n_t) (in a given time-span) to a probability. 1063 * 1064 * Our periodic faults will sample this probability and getting the 1065 * same result twice in a row, given these samples are fully 1066 * independent, is then given by P(n)^2, provided our sample period 1067 * is sufficiently short compared to the usage pattern. 1068 * 1069 * This quadric squishes small probabilities, making it less likely we 1070 * act on an unlikely task<->page relation. 1071 */ 1072 last_cpupid = page_cpupid_xchg_last(page, this_cpupid); 1073 if (!cpupid_pid_unset(last_cpupid) && 1074 cpupid_to_nid(last_cpupid) != dst_nid) 1075 return false; 1076 1077 /* Always allow migrate on private faults */ 1078 if (cpupid_match_pid(p, last_cpupid)) 1079 return true; 1080 1081 /* A shared fault, but p->numa_group has not been set up yet. */ 1082 if (!ng) 1083 return true; 1084 1085 /* 1086 * Do not migrate if the destination is not a node that 1087 * is actively used by this numa group. 1088 */ 1089 if (!node_isset(dst_nid, ng->active_nodes)) 1090 return false; 1091 1092 /* 1093 * Source is a node that is not actively used by this 1094 * numa group, while the destination is. Migrate. 1095 */ 1096 if (!node_isset(src_nid, ng->active_nodes)) 1097 return true; 1098 1099 /* 1100 * Both source and destination are nodes in active 1101 * use by this numa group. Maximize memory bandwidth 1102 * by migrating from more heavily used groups, to less 1103 * heavily used ones, spreading the load around. 1104 * Use a 1/4 hysteresis to avoid spurious page movement. 1105 */ 1106 return group_faults(p, dst_nid) < (group_faults(p, src_nid) * 3 / 4); 1107 } 1108 1109 static unsigned long weighted_cpuload(const int cpu); 1110 static unsigned long source_load(int cpu, int type); 1111 static unsigned long target_load(int cpu, int type); 1112 static unsigned long capacity_of(int cpu); 1113 static long effective_load(struct task_group *tg, int cpu, long wl, long wg); 1114 1115 /* Cached statistics for all CPUs within a node */ 1116 struct numa_stats { 1117 unsigned long nr_running; 1118 unsigned long load; 1119 1120 /* Total compute capacity of CPUs on a node */ 1121 unsigned long compute_capacity; 1122 1123 /* Approximate capacity in terms of runnable tasks on a node */ 1124 unsigned long task_capacity; 1125 int has_free_capacity; 1126 }; 1127 1128 /* 1129 * XXX borrowed from update_sg_lb_stats 1130 */ 1131 static void update_numa_stats(struct numa_stats *ns, int nid) 1132 { 1133 int smt, cpu, cpus = 0; 1134 unsigned long capacity; 1135 1136 memset(ns, 0, sizeof(*ns)); 1137 for_each_cpu(cpu, cpumask_of_node(nid)) { 1138 struct rq *rq = cpu_rq(cpu); 1139 1140 ns->nr_running += rq->nr_running; 1141 ns->load += weighted_cpuload(cpu); 1142 ns->compute_capacity += capacity_of(cpu); 1143 1144 cpus++; 1145 } 1146 1147 /* 1148 * If we raced with hotplug and there are no CPUs left in our mask 1149 * the @ns structure is NULL'ed and task_numa_compare() will 1150 * not find this node attractive. 1151 * 1152 * We'll either bail at !has_free_capacity, or we'll detect a huge 1153 * imbalance and bail there. 1154 */ 1155 if (!cpus) 1156 return; 1157 1158 /* smt := ceil(cpus / capacity), assumes: 1 < smt_power < 2 */ 1159 smt = DIV_ROUND_UP(SCHED_CAPACITY_SCALE * cpus, ns->compute_capacity); 1160 capacity = cpus / smt; /* cores */ 1161 1162 ns->task_capacity = min_t(unsigned, capacity, 1163 DIV_ROUND_CLOSEST(ns->compute_capacity, SCHED_CAPACITY_SCALE)); 1164 ns->has_free_capacity = (ns->nr_running < ns->task_capacity); 1165 } 1166 1167 struct task_numa_env { 1168 struct task_struct *p; 1169 1170 int src_cpu, src_nid; 1171 int dst_cpu, dst_nid; 1172 1173 struct numa_stats src_stats, dst_stats; 1174 1175 int imbalance_pct; 1176 int dist; 1177 1178 struct task_struct *best_task; 1179 long best_imp; 1180 int best_cpu; 1181 }; 1182 1183 static void task_numa_assign(struct task_numa_env *env, 1184 struct task_struct *p, long imp) 1185 { 1186 if (env->best_task) 1187 put_task_struct(env->best_task); 1188 if (p) 1189 get_task_struct(p); 1190 1191 env->best_task = p; 1192 env->best_imp = imp; 1193 env->best_cpu = env->dst_cpu; 1194 } 1195 1196 static bool load_too_imbalanced(long src_load, long dst_load, 1197 struct task_numa_env *env) 1198 { 1199 long imb, old_imb; 1200 long orig_src_load, orig_dst_load; 1201 long src_capacity, dst_capacity; 1202 1203 /* 1204 * The load is corrected for the CPU capacity available on each node. 1205 * 1206 * src_load dst_load 1207 * ------------ vs --------- 1208 * src_capacity dst_capacity 1209 */ 1210 src_capacity = env->src_stats.compute_capacity; 1211 dst_capacity = env->dst_stats.compute_capacity; 1212 1213 /* We care about the slope of the imbalance, not the direction. */ 1214 if (dst_load < src_load) 1215 swap(dst_load, src_load); 1216 1217 /* Is the difference below the threshold? */ 1218 imb = dst_load * src_capacity * 100 - 1219 src_load * dst_capacity * env->imbalance_pct; 1220 if (imb <= 0) 1221 return false; 1222 1223 /* 1224 * The imbalance is above the allowed threshold. 1225 * Compare it with the old imbalance. 1226 */ 1227 orig_src_load = env->src_stats.load; 1228 orig_dst_load = env->dst_stats.load; 1229 1230 if (orig_dst_load < orig_src_load) 1231 swap(orig_dst_load, orig_src_load); 1232 1233 old_imb = orig_dst_load * src_capacity * 100 - 1234 orig_src_load * dst_capacity * env->imbalance_pct; 1235 1236 /* Would this change make things worse? */ 1237 return (imb > old_imb); 1238 } 1239 1240 /* 1241 * This checks if the overall compute and NUMA accesses of the system would 1242 * be improved if the source tasks was migrated to the target dst_cpu taking 1243 * into account that it might be best if task running on the dst_cpu should 1244 * be exchanged with the source task 1245 */ 1246 static void task_numa_compare(struct task_numa_env *env, 1247 long taskimp, long groupimp) 1248 { 1249 struct rq *src_rq = cpu_rq(env->src_cpu); 1250 struct rq *dst_rq = cpu_rq(env->dst_cpu); 1251 struct task_struct *cur; 1252 long src_load, dst_load; 1253 long load; 1254 long imp = env->p->numa_group ? groupimp : taskimp; 1255 long moveimp = imp; 1256 int dist = env->dist; 1257 1258 rcu_read_lock(); 1259 1260 raw_spin_lock_irq(&dst_rq->lock); 1261 cur = dst_rq->curr; 1262 /* 1263 * No need to move the exiting task, and this ensures that ->curr 1264 * wasn't reaped and thus get_task_struct() in task_numa_assign() 1265 * is safe under RCU read lock. 1266 * Note that rcu_read_lock() itself can't protect from the final 1267 * put_task_struct() after the last schedule(). 1268 */ 1269 if ((cur->flags & PF_EXITING) || is_idle_task(cur)) 1270 cur = NULL; 1271 raw_spin_unlock_irq(&dst_rq->lock); 1272 1273 /* 1274 * Because we have preemption enabled we can get migrated around and 1275 * end try selecting ourselves (current == env->p) as a swap candidate. 1276 */ 1277 if (cur == env->p) 1278 goto unlock; 1279 1280 /* 1281 * "imp" is the fault differential for the source task between the 1282 * source and destination node. Calculate the total differential for 1283 * the source task and potential destination task. The more negative 1284 * the value is, the more rmeote accesses that would be expected to 1285 * be incurred if the tasks were swapped. 1286 */ 1287 if (cur) { 1288 /* Skip this swap candidate if cannot move to the source cpu */ 1289 if (!cpumask_test_cpu(env->src_cpu, tsk_cpus_allowed(cur))) 1290 goto unlock; 1291 1292 /* 1293 * If dst and source tasks are in the same NUMA group, or not 1294 * in any group then look only at task weights. 1295 */ 1296 if (cur->numa_group == env->p->numa_group) { 1297 imp = taskimp + task_weight(cur, env->src_nid, dist) - 1298 task_weight(cur, env->dst_nid, dist); 1299 /* 1300 * Add some hysteresis to prevent swapping the 1301 * tasks within a group over tiny differences. 1302 */ 1303 if (cur->numa_group) 1304 imp -= imp/16; 1305 } else { 1306 /* 1307 * Compare the group weights. If a task is all by 1308 * itself (not part of a group), use the task weight 1309 * instead. 1310 */ 1311 if (cur->numa_group) 1312 imp += group_weight(cur, env->src_nid, dist) - 1313 group_weight(cur, env->dst_nid, dist); 1314 else 1315 imp += task_weight(cur, env->src_nid, dist) - 1316 task_weight(cur, env->dst_nid, dist); 1317 } 1318 } 1319 1320 if (imp <= env->best_imp && moveimp <= env->best_imp) 1321 goto unlock; 1322 1323 if (!cur) { 1324 /* Is there capacity at our destination? */ 1325 if (env->src_stats.nr_running <= env->src_stats.task_capacity && 1326 !env->dst_stats.has_free_capacity) 1327 goto unlock; 1328 1329 goto balance; 1330 } 1331 1332 /* Balance doesn't matter much if we're running a task per cpu */ 1333 if (imp > env->best_imp && src_rq->nr_running == 1 && 1334 dst_rq->nr_running == 1) 1335 goto assign; 1336 1337 /* 1338 * In the overloaded case, try and keep the load balanced. 1339 */ 1340 balance: 1341 load = task_h_load(env->p); 1342 dst_load = env->dst_stats.load + load; 1343 src_load = env->src_stats.load - load; 1344 1345 if (moveimp > imp && moveimp > env->best_imp) { 1346 /* 1347 * If the improvement from just moving env->p direction is 1348 * better than swapping tasks around, check if a move is 1349 * possible. Store a slightly smaller score than moveimp, 1350 * so an actually idle CPU will win. 1351 */ 1352 if (!load_too_imbalanced(src_load, dst_load, env)) { 1353 imp = moveimp - 1; 1354 cur = NULL; 1355 goto assign; 1356 } 1357 } 1358 1359 if (imp <= env->best_imp) 1360 goto unlock; 1361 1362 if (cur) { 1363 load = task_h_load(cur); 1364 dst_load -= load; 1365 src_load += load; 1366 } 1367 1368 if (load_too_imbalanced(src_load, dst_load, env)) 1369 goto unlock; 1370 1371 /* 1372 * One idle CPU per node is evaluated for a task numa move. 1373 * Call select_idle_sibling to maybe find a better one. 1374 */ 1375 if (!cur) 1376 env->dst_cpu = select_idle_sibling(env->p, env->dst_cpu); 1377 1378 assign: 1379 task_numa_assign(env, cur, imp); 1380 unlock: 1381 rcu_read_unlock(); 1382 } 1383 1384 static void task_numa_find_cpu(struct task_numa_env *env, 1385 long taskimp, long groupimp) 1386 { 1387 int cpu; 1388 1389 for_each_cpu(cpu, cpumask_of_node(env->dst_nid)) { 1390 /* Skip this CPU if the source task cannot migrate */ 1391 if (!cpumask_test_cpu(cpu, tsk_cpus_allowed(env->p))) 1392 continue; 1393 1394 env->dst_cpu = cpu; 1395 task_numa_compare(env, taskimp, groupimp); 1396 } 1397 } 1398 1399 static int task_numa_migrate(struct task_struct *p) 1400 { 1401 struct task_numa_env env = { 1402 .p = p, 1403 1404 .src_cpu = task_cpu(p), 1405 .src_nid = task_node(p), 1406 1407 .imbalance_pct = 112, 1408 1409 .best_task = NULL, 1410 .best_imp = 0, 1411 .best_cpu = -1 1412 }; 1413 struct sched_domain *sd; 1414 unsigned long taskweight, groupweight; 1415 int nid, ret, dist; 1416 long taskimp, groupimp; 1417 1418 /* 1419 * Pick the lowest SD_NUMA domain, as that would have the smallest 1420 * imbalance and would be the first to start moving tasks about. 1421 * 1422 * And we want to avoid any moving of tasks about, as that would create 1423 * random movement of tasks -- counter the numa conditions we're trying 1424 * to satisfy here. 1425 */ 1426 rcu_read_lock(); 1427 sd = rcu_dereference(per_cpu(sd_numa, env.src_cpu)); 1428 if (sd) 1429 env.imbalance_pct = 100 + (sd->imbalance_pct - 100) / 2; 1430 rcu_read_unlock(); 1431 1432 /* 1433 * Cpusets can break the scheduler domain tree into smaller 1434 * balance domains, some of which do not cross NUMA boundaries. 1435 * Tasks that are "trapped" in such domains cannot be migrated 1436 * elsewhere, so there is no point in (re)trying. 1437 */ 1438 if (unlikely(!sd)) { 1439 p->numa_preferred_nid = task_node(p); 1440 return -EINVAL; 1441 } 1442 1443 env.dst_nid = p->numa_preferred_nid; 1444 dist = env.dist = node_distance(env.src_nid, env.dst_nid); 1445 taskweight = task_weight(p, env.src_nid, dist); 1446 groupweight = group_weight(p, env.src_nid, dist); 1447 update_numa_stats(&env.src_stats, env.src_nid); 1448 taskimp = task_weight(p, env.dst_nid, dist) - taskweight; 1449 groupimp = group_weight(p, env.dst_nid, dist) - groupweight; 1450 update_numa_stats(&env.dst_stats, env.dst_nid); 1451 1452 /* Try to find a spot on the preferred nid. */ 1453 task_numa_find_cpu(&env, taskimp, groupimp); 1454 1455 /* 1456 * Look at other nodes in these cases: 1457 * - there is no space available on the preferred_nid 1458 * - the task is part of a numa_group that is interleaved across 1459 * multiple NUMA nodes; in order to better consolidate the group, 1460 * we need to check other locations. 1461 */ 1462 if (env.best_cpu == -1 || (p->numa_group && 1463 nodes_weight(p->numa_group->active_nodes) > 1)) { 1464 for_each_online_node(nid) { 1465 if (nid == env.src_nid || nid == p->numa_preferred_nid) 1466 continue; 1467 1468 dist = node_distance(env.src_nid, env.dst_nid); 1469 if (sched_numa_topology_type == NUMA_BACKPLANE && 1470 dist != env.dist) { 1471 taskweight = task_weight(p, env.src_nid, dist); 1472 groupweight = group_weight(p, env.src_nid, dist); 1473 } 1474 1475 /* Only consider nodes where both task and groups benefit */ 1476 taskimp = task_weight(p, nid, dist) - taskweight; 1477 groupimp = group_weight(p, nid, dist) - groupweight; 1478 if (taskimp < 0 && groupimp < 0) 1479 continue; 1480 1481 env.dist = dist; 1482 env.dst_nid = nid; 1483 update_numa_stats(&env.dst_stats, env.dst_nid); 1484 task_numa_find_cpu(&env, taskimp, groupimp); 1485 } 1486 } 1487 1488 /* 1489 * If the task is part of a workload that spans multiple NUMA nodes, 1490 * and is migrating into one of the workload's active nodes, remember 1491 * this node as the task's preferred numa node, so the workload can 1492 * settle down. 1493 * A task that migrated to a second choice node will be better off 1494 * trying for a better one later. Do not set the preferred node here. 1495 */ 1496 if (p->numa_group) { 1497 if (env.best_cpu == -1) 1498 nid = env.src_nid; 1499 else 1500 nid = env.dst_nid; 1501 1502 if (node_isset(nid, p->numa_group->active_nodes)) 1503 sched_setnuma(p, env.dst_nid); 1504 } 1505 1506 /* No better CPU than the current one was found. */ 1507 if (env.best_cpu == -1) 1508 return -EAGAIN; 1509 1510 /* 1511 * Reset the scan period if the task is being rescheduled on an 1512 * alternative node to recheck if the tasks is now properly placed. 1513 */ 1514 p->numa_scan_period = task_scan_min(p); 1515 1516 if (env.best_task == NULL) { 1517 ret = migrate_task_to(p, env.best_cpu); 1518 if (ret != 0) 1519 trace_sched_stick_numa(p, env.src_cpu, env.best_cpu); 1520 return ret; 1521 } 1522 1523 ret = migrate_swap(p, env.best_task); 1524 if (ret != 0) 1525 trace_sched_stick_numa(p, env.src_cpu, task_cpu(env.best_task)); 1526 put_task_struct(env.best_task); 1527 return ret; 1528 } 1529 1530 /* Attempt to migrate a task to a CPU on the preferred node. */ 1531 static void numa_migrate_preferred(struct task_struct *p) 1532 { 1533 unsigned long interval = HZ; 1534 1535 /* This task has no NUMA fault statistics yet */ 1536 if (unlikely(p->numa_preferred_nid == -1 || !p->numa_faults)) 1537 return; 1538 1539 /* Periodically retry migrating the task to the preferred node */ 1540 interval = min(interval, msecs_to_jiffies(p->numa_scan_period) / 16); 1541 p->numa_migrate_retry = jiffies + interval; 1542 1543 /* Success if task is already running on preferred CPU */ 1544 if (task_node(p) == p->numa_preferred_nid) 1545 return; 1546 1547 /* Otherwise, try migrate to a CPU on the preferred node */ 1548 task_numa_migrate(p); 1549 } 1550 1551 /* 1552 * Find the nodes on which the workload is actively running. We do this by 1553 * tracking the nodes from which NUMA hinting faults are triggered. This can 1554 * be different from the set of nodes where the workload's memory is currently 1555 * located. 1556 * 1557 * The bitmask is used to make smarter decisions on when to do NUMA page 1558 * migrations, To prevent flip-flopping, and excessive page migrations, nodes 1559 * are added when they cause over 6/16 of the maximum number of faults, but 1560 * only removed when they drop below 3/16. 1561 */ 1562 static void update_numa_active_node_mask(struct numa_group *numa_group) 1563 { 1564 unsigned long faults, max_faults = 0; 1565 int nid; 1566 1567 for_each_online_node(nid) { 1568 faults = group_faults_cpu(numa_group, nid); 1569 if (faults > max_faults) 1570 max_faults = faults; 1571 } 1572 1573 for_each_online_node(nid) { 1574 faults = group_faults_cpu(numa_group, nid); 1575 if (!node_isset(nid, numa_group->active_nodes)) { 1576 if (faults > max_faults * 6 / 16) 1577 node_set(nid, numa_group->active_nodes); 1578 } else if (faults < max_faults * 3 / 16) 1579 node_clear(nid, numa_group->active_nodes); 1580 } 1581 } 1582 1583 /* 1584 * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS 1585 * increments. The more local the fault statistics are, the higher the scan 1586 * period will be for the next scan window. If local/(local+remote) ratio is 1587 * below NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS) 1588 * the scan period will decrease. Aim for 70% local accesses. 1589 */ 1590 #define NUMA_PERIOD_SLOTS 10 1591 #define NUMA_PERIOD_THRESHOLD 7 1592 1593 /* 1594 * Increase the scan period (slow down scanning) if the majority of 1595 * our memory is already on our local node, or if the majority of 1596 * the page accesses are shared with other processes. 1597 * Otherwise, decrease the scan period. 1598 */ 1599 static void update_task_scan_period(struct task_struct *p, 1600 unsigned long shared, unsigned long private) 1601 { 1602 unsigned int period_slot; 1603 int ratio; 1604 int diff; 1605 1606 unsigned long remote = p->numa_faults_locality[0]; 1607 unsigned long local = p->numa_faults_locality[1]; 1608 1609 /* 1610 * If there were no record hinting faults then either the task is 1611 * completely idle or all activity is areas that are not of interest 1612 * to automatic numa balancing. Scan slower 1613 */ 1614 if (local + shared == 0) { 1615 p->numa_scan_period = min(p->numa_scan_period_max, 1616 p->numa_scan_period << 1); 1617 1618 p->mm->numa_next_scan = jiffies + 1619 msecs_to_jiffies(p->numa_scan_period); 1620 1621 return; 1622 } 1623 1624 /* 1625 * Prepare to scale scan period relative to the current period. 1626 * == NUMA_PERIOD_THRESHOLD scan period stays the same 1627 * < NUMA_PERIOD_THRESHOLD scan period decreases (scan faster) 1628 * >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower) 1629 */ 1630 period_slot = DIV_ROUND_UP(p->numa_scan_period, NUMA_PERIOD_SLOTS); 1631 ratio = (local * NUMA_PERIOD_SLOTS) / (local + remote); 1632 if (ratio >= NUMA_PERIOD_THRESHOLD) { 1633 int slot = ratio - NUMA_PERIOD_THRESHOLD; 1634 if (!slot) 1635 slot = 1; 1636 diff = slot * period_slot; 1637 } else { 1638 diff = -(NUMA_PERIOD_THRESHOLD - ratio) * period_slot; 1639 1640 /* 1641 * Scale scan rate increases based on sharing. There is an 1642 * inverse relationship between the degree of sharing and 1643 * the adjustment made to the scanning period. Broadly 1644 * speaking the intent is that there is little point 1645 * scanning faster if shared accesses dominate as it may 1646 * simply bounce migrations uselessly 1647 */ 1648 ratio = DIV_ROUND_UP(private * NUMA_PERIOD_SLOTS, (private + shared + 1)); 1649 diff = (diff * ratio) / NUMA_PERIOD_SLOTS; 1650 } 1651 1652 p->numa_scan_period = clamp(p->numa_scan_period + diff, 1653 task_scan_min(p), task_scan_max(p)); 1654 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality)); 1655 } 1656 1657 /* 1658 * Get the fraction of time the task has been running since the last 1659 * NUMA placement cycle. The scheduler keeps similar statistics, but 1660 * decays those on a 32ms period, which is orders of magnitude off 1661 * from the dozens-of-seconds NUMA balancing period. Use the scheduler 1662 * stats only if the task is so new there are no NUMA statistics yet. 1663 */ 1664 static u64 numa_get_avg_runtime(struct task_struct *p, u64 *period) 1665 { 1666 u64 runtime, delta, now; 1667 /* Use the start of this time slice to avoid calculations. */ 1668 now = p->se.exec_start; 1669 runtime = p->se.sum_exec_runtime; 1670 1671 if (p->last_task_numa_placement) { 1672 delta = runtime - p->last_sum_exec_runtime; 1673 *period = now - p->last_task_numa_placement; 1674 } else { 1675 delta = p->se.avg.runnable_avg_sum; 1676 *period = p->se.avg.runnable_avg_period; 1677 } 1678 1679 p->last_sum_exec_runtime = runtime; 1680 p->last_task_numa_placement = now; 1681 1682 return delta; 1683 } 1684 1685 /* 1686 * Determine the preferred nid for a task in a numa_group. This needs to 1687 * be done in a way that produces consistent results with group_weight, 1688 * otherwise workloads might not converge. 1689 */ 1690 static int preferred_group_nid(struct task_struct *p, int nid) 1691 { 1692 nodemask_t nodes; 1693 int dist; 1694 1695 /* Direct connections between all NUMA nodes. */ 1696 if (sched_numa_topology_type == NUMA_DIRECT) 1697 return nid; 1698 1699 /* 1700 * On a system with glueless mesh NUMA topology, group_weight 1701 * scores nodes according to the number of NUMA hinting faults on 1702 * both the node itself, and on nearby nodes. 1703 */ 1704 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) { 1705 unsigned long score, max_score = 0; 1706 int node, max_node = nid; 1707 1708 dist = sched_max_numa_distance; 1709 1710 for_each_online_node(node) { 1711 score = group_weight(p, node, dist); 1712 if (score > max_score) { 1713 max_score = score; 1714 max_node = node; 1715 } 1716 } 1717 return max_node; 1718 } 1719 1720 /* 1721 * Finding the preferred nid in a system with NUMA backplane 1722 * interconnect topology is more involved. The goal is to locate 1723 * tasks from numa_groups near each other in the system, and 1724 * untangle workloads from different sides of the system. This requires 1725 * searching down the hierarchy of node groups, recursively searching 1726 * inside the highest scoring group of nodes. The nodemask tricks 1727 * keep the complexity of the search down. 1728 */ 1729 nodes = node_online_map; 1730 for (dist = sched_max_numa_distance; dist > LOCAL_DISTANCE; dist--) { 1731 unsigned long max_faults = 0; 1732 nodemask_t max_group = NODE_MASK_NONE; 1733 int a, b; 1734 1735 /* Are there nodes at this distance from each other? */ 1736 if (!find_numa_distance(dist)) 1737 continue; 1738 1739 for_each_node_mask(a, nodes) { 1740 unsigned long faults = 0; 1741 nodemask_t this_group; 1742 nodes_clear(this_group); 1743 1744 /* Sum group's NUMA faults; includes a==b case. */ 1745 for_each_node_mask(b, nodes) { 1746 if (node_distance(a, b) < dist) { 1747 faults += group_faults(p, b); 1748 node_set(b, this_group); 1749 node_clear(b, nodes); 1750 } 1751 } 1752 1753 /* Remember the top group. */ 1754 if (faults > max_faults) { 1755 max_faults = faults; 1756 max_group = this_group; 1757 /* 1758 * subtle: at the smallest distance there is 1759 * just one node left in each "group", the 1760 * winner is the preferred nid. 1761 */ 1762 nid = a; 1763 } 1764 } 1765 /* Next round, evaluate the nodes within max_group. */ 1766 nodes = max_group; 1767 } 1768 return nid; 1769 } 1770 1771 static void task_numa_placement(struct task_struct *p) 1772 { 1773 int seq, nid, max_nid = -1, max_group_nid = -1; 1774 unsigned long max_faults = 0, max_group_faults = 0; 1775 unsigned long fault_types[2] = { 0, 0 }; 1776 unsigned long total_faults; 1777 u64 runtime, period; 1778 spinlock_t *group_lock = NULL; 1779 1780 seq = ACCESS_ONCE(p->mm->numa_scan_seq); 1781 if (p->numa_scan_seq == seq) 1782 return; 1783 p->numa_scan_seq = seq; 1784 p->numa_scan_period_max = task_scan_max(p); 1785 1786 total_faults = p->numa_faults_locality[0] + 1787 p->numa_faults_locality[1]; 1788 runtime = numa_get_avg_runtime(p, &period); 1789 1790 /* If the task is part of a group prevent parallel updates to group stats */ 1791 if (p->numa_group) { 1792 group_lock = &p->numa_group->lock; 1793 spin_lock_irq(group_lock); 1794 } 1795 1796 /* Find the node with the highest number of faults */ 1797 for_each_online_node(nid) { 1798 /* Keep track of the offsets in numa_faults array */ 1799 int mem_idx, membuf_idx, cpu_idx, cpubuf_idx; 1800 unsigned long faults = 0, group_faults = 0; 1801 int priv; 1802 1803 for (priv = 0; priv < NR_NUMA_HINT_FAULT_TYPES; priv++) { 1804 long diff, f_diff, f_weight; 1805 1806 mem_idx = task_faults_idx(NUMA_MEM, nid, priv); 1807 membuf_idx = task_faults_idx(NUMA_MEMBUF, nid, priv); 1808 cpu_idx = task_faults_idx(NUMA_CPU, nid, priv); 1809 cpubuf_idx = task_faults_idx(NUMA_CPUBUF, nid, priv); 1810 1811 /* Decay existing window, copy faults since last scan */ 1812 diff = p->numa_faults[membuf_idx] - p->numa_faults[mem_idx] / 2; 1813 fault_types[priv] += p->numa_faults[membuf_idx]; 1814 p->numa_faults[membuf_idx] = 0; 1815 1816 /* 1817 * Normalize the faults_from, so all tasks in a group 1818 * count according to CPU use, instead of by the raw 1819 * number of faults. Tasks with little runtime have 1820 * little over-all impact on throughput, and thus their 1821 * faults are less important. 1822 */ 1823 f_weight = div64_u64(runtime << 16, period + 1); 1824 f_weight = (f_weight * p->numa_faults[cpubuf_idx]) / 1825 (total_faults + 1); 1826 f_diff = f_weight - p->numa_faults[cpu_idx] / 2; 1827 p->numa_faults[cpubuf_idx] = 0; 1828 1829 p->numa_faults[mem_idx] += diff; 1830 p->numa_faults[cpu_idx] += f_diff; 1831 faults += p->numa_faults[mem_idx]; 1832 p->total_numa_faults += diff; 1833 if (p->numa_group) { 1834 /* 1835 * safe because we can only change our own group 1836 * 1837 * mem_idx represents the offset for a given 1838 * nid and priv in a specific region because it 1839 * is at the beginning of the numa_faults array. 1840 */ 1841 p->numa_group->faults[mem_idx] += diff; 1842 p->numa_group->faults_cpu[mem_idx] += f_diff; 1843 p->numa_group->total_faults += diff; 1844 group_faults += p->numa_group->faults[mem_idx]; 1845 } 1846 } 1847 1848 if (faults > max_faults) { 1849 max_faults = faults; 1850 max_nid = nid; 1851 } 1852 1853 if (group_faults > max_group_faults) { 1854 max_group_faults = group_faults; 1855 max_group_nid = nid; 1856 } 1857 } 1858 1859 update_task_scan_period(p, fault_types[0], fault_types[1]); 1860 1861 if (p->numa_group) { 1862 update_numa_active_node_mask(p->numa_group); 1863 spin_unlock_irq(group_lock); 1864 max_nid = preferred_group_nid(p, max_group_nid); 1865 } 1866 1867 if (max_faults) { 1868 /* Set the new preferred node */ 1869 if (max_nid != p->numa_preferred_nid) 1870 sched_setnuma(p, max_nid); 1871 1872 if (task_node(p) != p->numa_preferred_nid) 1873 numa_migrate_preferred(p); 1874 } 1875 } 1876 1877 static inline int get_numa_group(struct numa_group *grp) 1878 { 1879 return atomic_inc_not_zero(&grp->refcount); 1880 } 1881 1882 static inline void put_numa_group(struct numa_group *grp) 1883 { 1884 if (atomic_dec_and_test(&grp->refcount)) 1885 kfree_rcu(grp, rcu); 1886 } 1887 1888 static void task_numa_group(struct task_struct *p, int cpupid, int flags, 1889 int *priv) 1890 { 1891 struct numa_group *grp, *my_grp; 1892 struct task_struct *tsk; 1893 bool join = false; 1894 int cpu = cpupid_to_cpu(cpupid); 1895 int i; 1896 1897 if (unlikely(!p->numa_group)) { 1898 unsigned int size = sizeof(struct numa_group) + 1899 4*nr_node_ids*sizeof(unsigned long); 1900 1901 grp = kzalloc(size, GFP_KERNEL | __GFP_NOWARN); 1902 if (!grp) 1903 return; 1904 1905 atomic_set(&grp->refcount, 1); 1906 spin_lock_init(&grp->lock); 1907 grp->gid = p->pid; 1908 /* Second half of the array tracks nids where faults happen */ 1909 grp->faults_cpu = grp->faults + NR_NUMA_HINT_FAULT_TYPES * 1910 nr_node_ids; 1911 1912 node_set(task_node(current), grp->active_nodes); 1913 1914 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) 1915 grp->faults[i] = p->numa_faults[i]; 1916 1917 grp->total_faults = p->total_numa_faults; 1918 1919 grp->nr_tasks++; 1920 rcu_assign_pointer(p->numa_group, grp); 1921 } 1922 1923 rcu_read_lock(); 1924 tsk = ACCESS_ONCE(cpu_rq(cpu)->curr); 1925 1926 if (!cpupid_match_pid(tsk, cpupid)) 1927 goto no_join; 1928 1929 grp = rcu_dereference(tsk->numa_group); 1930 if (!grp) 1931 goto no_join; 1932 1933 my_grp = p->numa_group; 1934 if (grp == my_grp) 1935 goto no_join; 1936 1937 /* 1938 * Only join the other group if its bigger; if we're the bigger group, 1939 * the other task will join us. 1940 */ 1941 if (my_grp->nr_tasks > grp->nr_tasks) 1942 goto no_join; 1943 1944 /* 1945 * Tie-break on the grp address. 1946 */ 1947 if (my_grp->nr_tasks == grp->nr_tasks && my_grp > grp) 1948 goto no_join; 1949 1950 /* Always join threads in the same process. */ 1951 if (tsk->mm == current->mm) 1952 join = true; 1953 1954 /* Simple filter to avoid false positives due to PID collisions */ 1955 if (flags & TNF_SHARED) 1956 join = true; 1957 1958 /* Update priv based on whether false sharing was detected */ 1959 *priv = !join; 1960 1961 if (join && !get_numa_group(grp)) 1962 goto no_join; 1963 1964 rcu_read_unlock(); 1965 1966 if (!join) 1967 return; 1968 1969 BUG_ON(irqs_disabled()); 1970 double_lock_irq(&my_grp->lock, &grp->lock); 1971 1972 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) { 1973 my_grp->faults[i] -= p->numa_faults[i]; 1974 grp->faults[i] += p->numa_faults[i]; 1975 } 1976 my_grp->total_faults -= p->total_numa_faults; 1977 grp->total_faults += p->total_numa_faults; 1978 1979 my_grp->nr_tasks--; 1980 grp->nr_tasks++; 1981 1982 spin_unlock(&my_grp->lock); 1983 spin_unlock_irq(&grp->lock); 1984 1985 rcu_assign_pointer(p->numa_group, grp); 1986 1987 put_numa_group(my_grp); 1988 return; 1989 1990 no_join: 1991 rcu_read_unlock(); 1992 return; 1993 } 1994 1995 void task_numa_free(struct task_struct *p) 1996 { 1997 struct numa_group *grp = p->numa_group; 1998 void *numa_faults = p->numa_faults; 1999 unsigned long flags; 2000 int i; 2001 2002 if (grp) { 2003 spin_lock_irqsave(&grp->lock, flags); 2004 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) 2005 grp->faults[i] -= p->numa_faults[i]; 2006 grp->total_faults -= p->total_numa_faults; 2007 2008 grp->nr_tasks--; 2009 spin_unlock_irqrestore(&grp->lock, flags); 2010 RCU_INIT_POINTER(p->numa_group, NULL); 2011 put_numa_group(grp); 2012 } 2013 2014 p->numa_faults = NULL; 2015 kfree(numa_faults); 2016 } 2017 2018 /* 2019 * Got a PROT_NONE fault for a page on @node. 2020 */ 2021 void task_numa_fault(int last_cpupid, int mem_node, int pages, int flags) 2022 { 2023 struct task_struct *p = current; 2024 bool migrated = flags & TNF_MIGRATED; 2025 int cpu_node = task_node(current); 2026 int local = !!(flags & TNF_FAULT_LOCAL); 2027 int priv; 2028 2029 if (!numabalancing_enabled) 2030 return; 2031 2032 /* for example, ksmd faulting in a user's mm */ 2033 if (!p->mm) 2034 return; 2035 2036 /* Allocate buffer to track faults on a per-node basis */ 2037 if (unlikely(!p->numa_faults)) { 2038 int size = sizeof(*p->numa_faults) * 2039 NR_NUMA_HINT_FAULT_BUCKETS * nr_node_ids; 2040 2041 p->numa_faults = kzalloc(size, GFP_KERNEL|__GFP_NOWARN); 2042 if (!p->numa_faults) 2043 return; 2044 2045 p->total_numa_faults = 0; 2046 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality)); 2047 } 2048 2049 /* 2050 * First accesses are treated as private, otherwise consider accesses 2051 * to be private if the accessing pid has not changed 2052 */ 2053 if (unlikely(last_cpupid == (-1 & LAST_CPUPID_MASK))) { 2054 priv = 1; 2055 } else { 2056 priv = cpupid_match_pid(p, last_cpupid); 2057 if (!priv && !(flags & TNF_NO_GROUP)) 2058 task_numa_group(p, last_cpupid, flags, &priv); 2059 } 2060 2061 /* 2062 * If a workload spans multiple NUMA nodes, a shared fault that 2063 * occurs wholly within the set of nodes that the workload is 2064 * actively using should be counted as local. This allows the 2065 * scan rate to slow down when a workload has settled down. 2066 */ 2067 if (!priv && !local && p->numa_group && 2068 node_isset(cpu_node, p->numa_group->active_nodes) && 2069 node_isset(mem_node, p->numa_group->active_nodes)) 2070 local = 1; 2071 2072 task_numa_placement(p); 2073 2074 /* 2075 * Retry task to preferred node migration periodically, in case it 2076 * case it previously failed, or the scheduler moved us. 2077 */ 2078 if (time_after(jiffies, p->numa_migrate_retry)) 2079 numa_migrate_preferred(p); 2080 2081 if (migrated) 2082 p->numa_pages_migrated += pages; 2083 2084 p->numa_faults[task_faults_idx(NUMA_MEMBUF, mem_node, priv)] += pages; 2085 p->numa_faults[task_faults_idx(NUMA_CPUBUF, cpu_node, priv)] += pages; 2086 p->numa_faults_locality[local] += pages; 2087 } 2088 2089 static void reset_ptenuma_scan(struct task_struct *p) 2090 { 2091 ACCESS_ONCE(p->mm->numa_scan_seq)++; 2092 p->mm->numa_scan_offset = 0; 2093 } 2094 2095 /* 2096 * The expensive part of numa migration is done from task_work context. 2097 * Triggered from task_tick_numa(). 2098 */ 2099 void task_numa_work(struct callback_head *work) 2100 { 2101 unsigned long migrate, next_scan, now = jiffies; 2102 struct task_struct *p = current; 2103 struct mm_struct *mm = p->mm; 2104 struct vm_area_struct *vma; 2105 unsigned long start, end; 2106 unsigned long nr_pte_updates = 0; 2107 long pages; 2108 2109 WARN_ON_ONCE(p != container_of(work, struct task_struct, numa_work)); 2110 2111 work->next = work; /* protect against double add */ 2112 /* 2113 * Who cares about NUMA placement when they're dying. 2114 * 2115 * NOTE: make sure not to dereference p->mm before this check, 2116 * exit_task_work() happens _after_ exit_mm() so we could be called 2117 * without p->mm even though we still had it when we enqueued this 2118 * work. 2119 */ 2120 if (p->flags & PF_EXITING) 2121 return; 2122 2123 if (!mm->numa_next_scan) { 2124 mm->numa_next_scan = now + 2125 msecs_to_jiffies(sysctl_numa_balancing_scan_delay); 2126 } 2127 2128 /* 2129 * Enforce maximal scan/migration frequency.. 2130 */ 2131 migrate = mm->numa_next_scan; 2132 if (time_before(now, migrate)) 2133 return; 2134 2135 if (p->numa_scan_period == 0) { 2136 p->numa_scan_period_max = task_scan_max(p); 2137 p->numa_scan_period = task_scan_min(p); 2138 } 2139 2140 next_scan = now + msecs_to_jiffies(p->numa_scan_period); 2141 if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate) 2142 return; 2143 2144 /* 2145 * Delay this task enough that another task of this mm will likely win 2146 * the next time around. 2147 */ 2148 p->node_stamp += 2 * TICK_NSEC; 2149 2150 start = mm->numa_scan_offset; 2151 pages = sysctl_numa_balancing_scan_size; 2152 pages <<= 20 - PAGE_SHIFT; /* MB in pages */ 2153 if (!pages) 2154 return; 2155 2156 down_read(&mm->mmap_sem); 2157 vma = find_vma(mm, start); 2158 if (!vma) { 2159 reset_ptenuma_scan(p); 2160 start = 0; 2161 vma = mm->mmap; 2162 } 2163 for (; vma; vma = vma->vm_next) { 2164 if (!vma_migratable(vma) || !vma_policy_mof(vma)) 2165 continue; 2166 2167 /* 2168 * Shared library pages mapped by multiple processes are not 2169 * migrated as it is expected they are cache replicated. Avoid 2170 * hinting faults in read-only file-backed mappings or the vdso 2171 * as migrating the pages will be of marginal benefit. 2172 */ 2173 if (!vma->vm_mm || 2174 (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ))) 2175 continue; 2176 2177 /* 2178 * Skip inaccessible VMAs to avoid any confusion between 2179 * PROT_NONE and NUMA hinting ptes 2180 */ 2181 if (!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE))) 2182 continue; 2183 2184 do { 2185 start = max(start, vma->vm_start); 2186 end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE); 2187 end = min(end, vma->vm_end); 2188 nr_pte_updates += change_prot_numa(vma, start, end); 2189 2190 /* 2191 * Scan sysctl_numa_balancing_scan_size but ensure that 2192 * at least one PTE is updated so that unused virtual 2193 * address space is quickly skipped. 2194 */ 2195 if (nr_pte_updates) 2196 pages -= (end - start) >> PAGE_SHIFT; 2197 2198 start = end; 2199 if (pages <= 0) 2200 goto out; 2201 2202 cond_resched(); 2203 } while (end != vma->vm_end); 2204 } 2205 2206 out: 2207 /* 2208 * It is possible to reach the end of the VMA list but the last few 2209 * VMAs are not guaranteed to the vma_migratable. If they are not, we 2210 * would find the !migratable VMA on the next scan but not reset the 2211 * scanner to the start so check it now. 2212 */ 2213 if (vma) 2214 mm->numa_scan_offset = start; 2215 else 2216 reset_ptenuma_scan(p); 2217 up_read(&mm->mmap_sem); 2218 } 2219 2220 /* 2221 * Drive the periodic memory faults.. 2222 */ 2223 void task_tick_numa(struct rq *rq, struct task_struct *curr) 2224 { 2225 struct callback_head *work = &curr->numa_work; 2226 u64 period, now; 2227 2228 /* 2229 * We don't care about NUMA placement if we don't have memory. 2230 */ 2231 if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work) 2232 return; 2233 2234 /* 2235 * Using runtime rather than walltime has the dual advantage that 2236 * we (mostly) drive the selection from busy threads and that the 2237 * task needs to have done some actual work before we bother with 2238 * NUMA placement. 2239 */ 2240 now = curr->se.sum_exec_runtime; 2241 period = (u64)curr->numa_scan_period * NSEC_PER_MSEC; 2242 2243 if (now - curr->node_stamp > period) { 2244 if (!curr->node_stamp) 2245 curr->numa_scan_period = task_scan_min(curr); 2246 curr->node_stamp += period; 2247 2248 if (!time_before(jiffies, curr->mm->numa_next_scan)) { 2249 init_task_work(work, task_numa_work); /* TODO: move this into sched_fork() */ 2250 task_work_add(curr, work, true); 2251 } 2252 } 2253 } 2254 #else 2255 static void task_tick_numa(struct rq *rq, struct task_struct *curr) 2256 { 2257 } 2258 2259 static inline void account_numa_enqueue(struct rq *rq, struct task_struct *p) 2260 { 2261 } 2262 2263 static inline void account_numa_dequeue(struct rq *rq, struct task_struct *p) 2264 { 2265 } 2266 #endif /* CONFIG_NUMA_BALANCING */ 2267 2268 static void 2269 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se) 2270 { 2271 update_load_add(&cfs_rq->load, se->load.weight); 2272 if (!parent_entity(se)) 2273 update_load_add(&rq_of(cfs_rq)->load, se->load.weight); 2274 #ifdef CONFIG_SMP 2275 if (entity_is_task(se)) { 2276 struct rq *rq = rq_of(cfs_rq); 2277 2278 account_numa_enqueue(rq, task_of(se)); 2279 list_add(&se->group_node, &rq->cfs_tasks); 2280 } 2281 #endif 2282 cfs_rq->nr_running++; 2283 } 2284 2285 static void 2286 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se) 2287 { 2288 update_load_sub(&cfs_rq->load, se->load.weight); 2289 if (!parent_entity(se)) 2290 update_load_sub(&rq_of(cfs_rq)->load, se->load.weight); 2291 if (entity_is_task(se)) { 2292 account_numa_dequeue(rq_of(cfs_rq), task_of(se)); 2293 list_del_init(&se->group_node); 2294 } 2295 cfs_rq->nr_running--; 2296 } 2297 2298 #ifdef CONFIG_FAIR_GROUP_SCHED 2299 # ifdef CONFIG_SMP 2300 static inline long calc_tg_weight(struct task_group *tg, struct cfs_rq *cfs_rq) 2301 { 2302 long tg_weight; 2303 2304 /* 2305 * Use this CPU's actual weight instead of the last load_contribution 2306 * to gain a more accurate current total weight. See 2307 * update_cfs_rq_load_contribution(). 2308 */ 2309 tg_weight = atomic_long_read(&tg->load_avg); 2310 tg_weight -= cfs_rq->tg_load_contrib; 2311 tg_weight += cfs_rq->load.weight; 2312 2313 return tg_weight; 2314 } 2315 2316 static long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg) 2317 { 2318 long tg_weight, load, shares; 2319 2320 tg_weight = calc_tg_weight(tg, cfs_rq); 2321 load = cfs_rq->load.weight; 2322 2323 shares = (tg->shares * load); 2324 if (tg_weight) 2325 shares /= tg_weight; 2326 2327 if (shares < MIN_SHARES) 2328 shares = MIN_SHARES; 2329 if (shares > tg->shares) 2330 shares = tg->shares; 2331 2332 return shares; 2333 } 2334 # else /* CONFIG_SMP */ 2335 static inline long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg) 2336 { 2337 return tg->shares; 2338 } 2339 # endif /* CONFIG_SMP */ 2340 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, 2341 unsigned long weight) 2342 { 2343 if (se->on_rq) { 2344 /* commit outstanding execution time */ 2345 if (cfs_rq->curr == se) 2346 update_curr(cfs_rq); 2347 account_entity_dequeue(cfs_rq, se); 2348 } 2349 2350 update_load_set(&se->load, weight); 2351 2352 if (se->on_rq) 2353 account_entity_enqueue(cfs_rq, se); 2354 } 2355 2356 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq); 2357 2358 static void update_cfs_shares(struct cfs_rq *cfs_rq) 2359 { 2360 struct task_group *tg; 2361 struct sched_entity *se; 2362 long shares; 2363 2364 tg = cfs_rq->tg; 2365 se = tg->se[cpu_of(rq_of(cfs_rq))]; 2366 if (!se || throttled_hierarchy(cfs_rq)) 2367 return; 2368 #ifndef CONFIG_SMP 2369 if (likely(se->load.weight == tg->shares)) 2370 return; 2371 #endif 2372 shares = calc_cfs_shares(cfs_rq, tg); 2373 2374 reweight_entity(cfs_rq_of(se), se, shares); 2375 } 2376 #else /* CONFIG_FAIR_GROUP_SCHED */ 2377 static inline void update_cfs_shares(struct cfs_rq *cfs_rq) 2378 { 2379 } 2380 #endif /* CONFIG_FAIR_GROUP_SCHED */ 2381 2382 #ifdef CONFIG_SMP 2383 /* 2384 * We choose a half-life close to 1 scheduling period. 2385 * Note: The tables below are dependent on this value. 2386 */ 2387 #define LOAD_AVG_PERIOD 32 2388 #define LOAD_AVG_MAX 47742 /* maximum possible load avg */ 2389 #define LOAD_AVG_MAX_N 345 /* number of full periods to produce LOAD_MAX_AVG */ 2390 2391 /* Precomputed fixed inverse multiplies for multiplication by y^n */ 2392 static const u32 runnable_avg_yN_inv[] = { 2393 0xffffffff, 0xfa83b2da, 0xf5257d14, 0xefe4b99a, 0xeac0c6e6, 0xe5b906e6, 2394 0xe0ccdeeb, 0xdbfbb796, 0xd744fcc9, 0xd2a81d91, 0xce248c14, 0xc9b9bd85, 2395 0xc5672a10, 0xc12c4cc9, 0xbd08a39e, 0xb8fbaf46, 0xb504f333, 0xb123f581, 2396 0xad583ee9, 0xa9a15ab4, 0xa5fed6a9, 0xa2704302, 0x9ef5325f, 0x9b8d39b9, 2397 0x9837f050, 0x94f4efa8, 0x91c3d373, 0x8ea4398a, 0x8b95c1e3, 0x88980e80, 2398 0x85aac367, 0x82cd8698, 2399 }; 2400 2401 /* 2402 * Precomputed \Sum y^k { 1<=k<=n }. These are floor(true_value) to prevent 2403 * over-estimates when re-combining. 2404 */ 2405 static const u32 runnable_avg_yN_sum[] = { 2406 0, 1002, 1982, 2941, 3880, 4798, 5697, 6576, 7437, 8279, 9103, 2407 9909,10698,11470,12226,12966,13690,14398,15091,15769,16433,17082, 2408 17718,18340,18949,19545,20128,20698,21256,21802,22336,22859,23371, 2409 }; 2410 2411 /* 2412 * Approximate: 2413 * val * y^n, where y^32 ~= 0.5 (~1 scheduling period) 2414 */ 2415 static __always_inline u64 decay_load(u64 val, u64 n) 2416 { 2417 unsigned int local_n; 2418 2419 if (!n) 2420 return val; 2421 else if (unlikely(n > LOAD_AVG_PERIOD * 63)) 2422 return 0; 2423 2424 /* after bounds checking we can collapse to 32-bit */ 2425 local_n = n; 2426 2427 /* 2428 * As y^PERIOD = 1/2, we can combine 2429 * y^n = 1/2^(n/PERIOD) * y^(n%PERIOD) 2430 * With a look-up table which covers y^n (n<PERIOD) 2431 * 2432 * To achieve constant time decay_load. 2433 */ 2434 if (unlikely(local_n >= LOAD_AVG_PERIOD)) { 2435 val >>= local_n / LOAD_AVG_PERIOD; 2436 local_n %= LOAD_AVG_PERIOD; 2437 } 2438 2439 val *= runnable_avg_yN_inv[local_n]; 2440 /* We don't use SRR here since we always want to round down. */ 2441 return val >> 32; 2442 } 2443 2444 /* 2445 * For updates fully spanning n periods, the contribution to runnable 2446 * average will be: \Sum 1024*y^n 2447 * 2448 * We can compute this reasonably efficiently by combining: 2449 * y^PERIOD = 1/2 with precomputed \Sum 1024*y^n {for n <PERIOD} 2450 */ 2451 static u32 __compute_runnable_contrib(u64 n) 2452 { 2453 u32 contrib = 0; 2454 2455 if (likely(n <= LOAD_AVG_PERIOD)) 2456 return runnable_avg_yN_sum[n]; 2457 else if (unlikely(n >= LOAD_AVG_MAX_N)) 2458 return LOAD_AVG_MAX; 2459 2460 /* Compute \Sum k^n combining precomputed values for k^i, \Sum k^j */ 2461 do { 2462 contrib /= 2; /* y^LOAD_AVG_PERIOD = 1/2 */ 2463 contrib += runnable_avg_yN_sum[LOAD_AVG_PERIOD]; 2464 2465 n -= LOAD_AVG_PERIOD; 2466 } while (n > LOAD_AVG_PERIOD); 2467 2468 contrib = decay_load(contrib, n); 2469 return contrib + runnable_avg_yN_sum[n]; 2470 } 2471 2472 /* 2473 * We can represent the historical contribution to runnable average as the 2474 * coefficients of a geometric series. To do this we sub-divide our runnable 2475 * history into segments of approximately 1ms (1024us); label the segment that 2476 * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g. 2477 * 2478 * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ... 2479 * p0 p1 p2 2480 * (now) (~1ms ago) (~2ms ago) 2481 * 2482 * Let u_i denote the fraction of p_i that the entity was runnable. 2483 * 2484 * We then designate the fractions u_i as our co-efficients, yielding the 2485 * following representation of historical load: 2486 * u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ... 2487 * 2488 * We choose y based on the with of a reasonably scheduling period, fixing: 2489 * y^32 = 0.5 2490 * 2491 * This means that the contribution to load ~32ms ago (u_32) will be weighted 2492 * approximately half as much as the contribution to load within the last ms 2493 * (u_0). 2494 * 2495 * When a period "rolls over" and we have new u_0`, multiplying the previous 2496 * sum again by y is sufficient to update: 2497 * load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... ) 2498 * = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}] 2499 */ 2500 static __always_inline int __update_entity_runnable_avg(u64 now, 2501 struct sched_avg *sa, 2502 int runnable) 2503 { 2504 u64 delta, periods; 2505 u32 runnable_contrib; 2506 int delta_w, decayed = 0; 2507 2508 delta = now - sa->last_runnable_update; 2509 /* 2510 * This should only happen when time goes backwards, which it 2511 * unfortunately does during sched clock init when we swap over to TSC. 2512 */ 2513 if ((s64)delta < 0) { 2514 sa->last_runnable_update = now; 2515 return 0; 2516 } 2517 2518 /* 2519 * Use 1024ns as the unit of measurement since it's a reasonable 2520 * approximation of 1us and fast to compute. 2521 */ 2522 delta >>= 10; 2523 if (!delta) 2524 return 0; 2525 sa->last_runnable_update = now; 2526 2527 /* delta_w is the amount already accumulated against our next period */ 2528 delta_w = sa->runnable_avg_period % 1024; 2529 if (delta + delta_w >= 1024) { 2530 /* period roll-over */ 2531 decayed = 1; 2532 2533 /* 2534 * Now that we know we're crossing a period boundary, figure 2535 * out how much from delta we need to complete the current 2536 * period and accrue it. 2537 */ 2538 delta_w = 1024 - delta_w; 2539 if (runnable) 2540 sa->runnable_avg_sum += delta_w; 2541 sa->runnable_avg_period += delta_w; 2542 2543 delta -= delta_w; 2544 2545 /* Figure out how many additional periods this update spans */ 2546 periods = delta / 1024; 2547 delta %= 1024; 2548 2549 sa->runnable_avg_sum = decay_load(sa->runnable_avg_sum, 2550 periods + 1); 2551 sa->runnable_avg_period = decay_load(sa->runnable_avg_period, 2552 periods + 1); 2553 2554 /* Efficiently calculate \sum (1..n_period) 1024*y^i */ 2555 runnable_contrib = __compute_runnable_contrib(periods); 2556 if (runnable) 2557 sa->runnable_avg_sum += runnable_contrib; 2558 sa->runnable_avg_period += runnable_contrib; 2559 } 2560 2561 /* Remainder of delta accrued against u_0` */ 2562 if (runnable) 2563 sa->runnable_avg_sum += delta; 2564 sa->runnable_avg_period += delta; 2565 2566 return decayed; 2567 } 2568 2569 /* Synchronize an entity's decay with its parenting cfs_rq.*/ 2570 static inline u64 __synchronize_entity_decay(struct sched_entity *se) 2571 { 2572 struct cfs_rq *cfs_rq = cfs_rq_of(se); 2573 u64 decays = atomic64_read(&cfs_rq->decay_counter); 2574 2575 decays -= se->avg.decay_count; 2576 se->avg.decay_count = 0; 2577 if (!decays) 2578 return 0; 2579 2580 se->avg.load_avg_contrib = decay_load(se->avg.load_avg_contrib, decays); 2581 2582 return decays; 2583 } 2584 2585 #ifdef CONFIG_FAIR_GROUP_SCHED 2586 static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq *cfs_rq, 2587 int force_update) 2588 { 2589 struct task_group *tg = cfs_rq->tg; 2590 long tg_contrib; 2591 2592 tg_contrib = cfs_rq->runnable_load_avg + cfs_rq->blocked_load_avg; 2593 tg_contrib -= cfs_rq->tg_load_contrib; 2594 2595 if (!tg_contrib) 2596 return; 2597 2598 if (force_update || abs(tg_contrib) > cfs_rq->tg_load_contrib / 8) { 2599 atomic_long_add(tg_contrib, &tg->load_avg); 2600 cfs_rq->tg_load_contrib += tg_contrib; 2601 } 2602 } 2603 2604 /* 2605 * Aggregate cfs_rq runnable averages into an equivalent task_group 2606 * representation for computing load contributions. 2607 */ 2608 static inline void __update_tg_runnable_avg(struct sched_avg *sa, 2609 struct cfs_rq *cfs_rq) 2610 { 2611 struct task_group *tg = cfs_rq->tg; 2612 long contrib; 2613 2614 /* The fraction of a cpu used by this cfs_rq */ 2615 contrib = div_u64((u64)sa->runnable_avg_sum << NICE_0_SHIFT, 2616 sa->runnable_avg_period + 1); 2617 contrib -= cfs_rq->tg_runnable_contrib; 2618 2619 if (abs(contrib) > cfs_rq->tg_runnable_contrib / 64) { 2620 atomic_add(contrib, &tg->runnable_avg); 2621 cfs_rq->tg_runnable_contrib += contrib; 2622 } 2623 } 2624 2625 static inline void __update_group_entity_contrib(struct sched_entity *se) 2626 { 2627 struct cfs_rq *cfs_rq = group_cfs_rq(se); 2628 struct task_group *tg = cfs_rq->tg; 2629 int runnable_avg; 2630 2631 u64 contrib; 2632 2633 contrib = cfs_rq->tg_load_contrib * tg->shares; 2634 se->avg.load_avg_contrib = div_u64(contrib, 2635 atomic_long_read(&tg->load_avg) + 1); 2636 2637 /* 2638 * For group entities we need to compute a correction term in the case 2639 * that they are consuming <1 cpu so that we would contribute the same 2640 * load as a task of equal weight. 2641 * 2642 * Explicitly co-ordinating this measurement would be expensive, but 2643 * fortunately the sum of each cpus contribution forms a usable 2644 * lower-bound on the true value. 2645 * 2646 * Consider the aggregate of 2 contributions. Either they are disjoint 2647 * (and the sum represents true value) or they are disjoint and we are 2648 * understating by the aggregate of their overlap. 2649 * 2650 * Extending this to N cpus, for a given overlap, the maximum amount we 2651 * understand is then n_i(n_i+1)/2 * w_i where n_i is the number of 2652 * cpus that overlap for this interval and w_i is the interval width. 2653 * 2654 * On a small machine; the first term is well-bounded which bounds the 2655 * total error since w_i is a subset of the period. Whereas on a 2656 * larger machine, while this first term can be larger, if w_i is the 2657 * of consequential size guaranteed to see n_i*w_i quickly converge to 2658 * our upper bound of 1-cpu. 2659 */ 2660 runnable_avg = atomic_read(&tg->runnable_avg); 2661 if (runnable_avg < NICE_0_LOAD) { 2662 se->avg.load_avg_contrib *= runnable_avg; 2663 se->avg.load_avg_contrib >>= NICE_0_SHIFT; 2664 } 2665 } 2666 2667 static inline void update_rq_runnable_avg(struct rq *rq, int runnable) 2668 { 2669 __update_entity_runnable_avg(rq_clock_task(rq), &rq->avg, runnable); 2670 __update_tg_runnable_avg(&rq->avg, &rq->cfs); 2671 } 2672 #else /* CONFIG_FAIR_GROUP_SCHED */ 2673 static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq *cfs_rq, 2674 int force_update) {} 2675 static inline void __update_tg_runnable_avg(struct sched_avg *sa, 2676 struct cfs_rq *cfs_rq) {} 2677 static inline void __update_group_entity_contrib(struct sched_entity *se) {} 2678 static inline void update_rq_runnable_avg(struct rq *rq, int runnable) {} 2679 #endif /* CONFIG_FAIR_GROUP_SCHED */ 2680 2681 static inline void __update_task_entity_contrib(struct sched_entity *se) 2682 { 2683 u32 contrib; 2684 2685 /* avoid overflowing a 32-bit type w/ SCHED_LOAD_SCALE */ 2686 contrib = se->avg.runnable_avg_sum * scale_load_down(se->load.weight); 2687 contrib /= (se->avg.runnable_avg_period + 1); 2688 se->avg.load_avg_contrib = scale_load(contrib); 2689 } 2690 2691 /* Compute the current contribution to load_avg by se, return any delta */ 2692 static long __update_entity_load_avg_contrib(struct sched_entity *se) 2693 { 2694 long old_contrib = se->avg.load_avg_contrib; 2695 2696 if (entity_is_task(se)) { 2697 __update_task_entity_contrib(se); 2698 } else { 2699 __update_tg_runnable_avg(&se->avg, group_cfs_rq(se)); 2700 __update_group_entity_contrib(se); 2701 } 2702 2703 return se->avg.load_avg_contrib - old_contrib; 2704 } 2705 2706 static inline void subtract_blocked_load_contrib(struct cfs_rq *cfs_rq, 2707 long load_contrib) 2708 { 2709 if (likely(load_contrib < cfs_rq->blocked_load_avg)) 2710 cfs_rq->blocked_load_avg -= load_contrib; 2711 else 2712 cfs_rq->blocked_load_avg = 0; 2713 } 2714 2715 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq); 2716 2717 /* Update a sched_entity's runnable average */ 2718 static inline void update_entity_load_avg(struct sched_entity *se, 2719 int update_cfs_rq) 2720 { 2721 struct cfs_rq *cfs_rq = cfs_rq_of(se); 2722 long contrib_delta; 2723 u64 now; 2724 2725 /* 2726 * For a group entity we need to use their owned cfs_rq_clock_task() in 2727 * case they are the parent of a throttled hierarchy. 2728 */ 2729 if (entity_is_task(se)) 2730 now = cfs_rq_clock_task(cfs_rq); 2731 else 2732 now = cfs_rq_clock_task(group_cfs_rq(se)); 2733 2734 if (!__update_entity_runnable_avg(now, &se->avg, se->on_rq)) 2735 return; 2736 2737 contrib_delta = __update_entity_load_avg_contrib(se); 2738 2739 if (!update_cfs_rq) 2740 return; 2741 2742 if (se->on_rq) 2743 cfs_rq->runnable_load_avg += contrib_delta; 2744 else 2745 subtract_blocked_load_contrib(cfs_rq, -contrib_delta); 2746 } 2747 2748 /* 2749 * Decay the load contributed by all blocked children and account this so that 2750 * their contribution may appropriately discounted when they wake up. 2751 */ 2752 static void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq, int force_update) 2753 { 2754 u64 now = cfs_rq_clock_task(cfs_rq) >> 20; 2755 u64 decays; 2756 2757 decays = now - cfs_rq->last_decay; 2758 if (!decays && !force_update) 2759 return; 2760 2761 if (atomic_long_read(&cfs_rq->removed_load)) { 2762 unsigned long removed_load; 2763 removed_load = atomic_long_xchg(&cfs_rq->removed_load, 0); 2764 subtract_blocked_load_contrib(cfs_rq, removed_load); 2765 } 2766 2767 if (decays) { 2768 cfs_rq->blocked_load_avg = decay_load(cfs_rq->blocked_load_avg, 2769 decays); 2770 atomic64_add(decays, &cfs_rq->decay_counter); 2771 cfs_rq->last_decay = now; 2772 } 2773 2774 __update_cfs_rq_tg_load_contrib(cfs_rq, force_update); 2775 } 2776 2777 /* Add the load generated by se into cfs_rq's child load-average */ 2778 static inline void enqueue_entity_load_avg(struct cfs_rq *cfs_rq, 2779 struct sched_entity *se, 2780 int wakeup) 2781 { 2782 /* 2783 * We track migrations using entity decay_count <= 0, on a wake-up 2784 * migration we use a negative decay count to track the remote decays 2785 * accumulated while sleeping. 2786 * 2787 * Newly forked tasks are enqueued with se->avg.decay_count == 0, they 2788 * are seen by enqueue_entity_load_avg() as a migration with an already 2789 * constructed load_avg_contrib. 2790 */ 2791 if (unlikely(se->avg.decay_count <= 0)) { 2792 se->avg.last_runnable_update = rq_clock_task(rq_of(cfs_rq)); 2793 if (se->avg.decay_count) { 2794 /* 2795 * In a wake-up migration we have to approximate the 2796 * time sleeping. This is because we can't synchronize 2797 * clock_task between the two cpus, and it is not 2798 * guaranteed to be read-safe. Instead, we can 2799 * approximate this using our carried decays, which are 2800 * explicitly atomically readable. 2801 */ 2802 se->avg.last_runnable_update -= (-se->avg.decay_count) 2803 << 20; 2804 update_entity_load_avg(se, 0); 2805 /* Indicate that we're now synchronized and on-rq */ 2806 se->avg.decay_count = 0; 2807 } 2808 wakeup = 0; 2809 } else { 2810 __synchronize_entity_decay(se); 2811 } 2812 2813 /* migrated tasks did not contribute to our blocked load */ 2814 if (wakeup) { 2815 subtract_blocked_load_contrib(cfs_rq, se->avg.load_avg_contrib); 2816 update_entity_load_avg(se, 0); 2817 } 2818 2819 cfs_rq->runnable_load_avg += se->avg.load_avg_contrib; 2820 /* we force update consideration on load-balancer moves */ 2821 update_cfs_rq_blocked_load(cfs_rq, !wakeup); 2822 } 2823 2824 /* 2825 * Remove se's load from this cfs_rq child load-average, if the entity is 2826 * transitioning to a blocked state we track its projected decay using 2827 * blocked_load_avg. 2828 */ 2829 static inline void dequeue_entity_load_avg(struct cfs_rq *cfs_rq, 2830 struct sched_entity *se, 2831 int sleep) 2832 { 2833 update_entity_load_avg(se, 1); 2834 /* we force update consideration on load-balancer moves */ 2835 update_cfs_rq_blocked_load(cfs_rq, !sleep); 2836 2837 cfs_rq->runnable_load_avg -= se->avg.load_avg_contrib; 2838 if (sleep) { 2839 cfs_rq->blocked_load_avg += se->avg.load_avg_contrib; 2840 se->avg.decay_count = atomic64_read(&cfs_rq->decay_counter); 2841 } /* migrations, e.g. sleep=0 leave decay_count == 0 */ 2842 } 2843 2844 /* 2845 * Update the rq's load with the elapsed running time before entering 2846 * idle. if the last scheduled task is not a CFS task, idle_enter will 2847 * be the only way to update the runnable statistic. 2848 */ 2849 void idle_enter_fair(struct rq *this_rq) 2850 { 2851 update_rq_runnable_avg(this_rq, 1); 2852 } 2853 2854 /* 2855 * Update the rq's load with the elapsed idle time before a task is 2856 * scheduled. if the newly scheduled task is not a CFS task, idle_exit will 2857 * be the only way to update the runnable statistic. 2858 */ 2859 void idle_exit_fair(struct rq *this_rq) 2860 { 2861 update_rq_runnable_avg(this_rq, 0); 2862 } 2863 2864 static int idle_balance(struct rq *this_rq); 2865 2866 #else /* CONFIG_SMP */ 2867 2868 static inline void update_entity_load_avg(struct sched_entity *se, 2869 int update_cfs_rq) {} 2870 static inline void update_rq_runnable_avg(struct rq *rq, int runnable) {} 2871 static inline void enqueue_entity_load_avg(struct cfs_rq *cfs_rq, 2872 struct sched_entity *se, 2873 int wakeup) {} 2874 static inline void dequeue_entity_load_avg(struct cfs_rq *cfs_rq, 2875 struct sched_entity *se, 2876 int sleep) {} 2877 static inline void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq, 2878 int force_update) {} 2879 2880 static inline int idle_balance(struct rq *rq) 2881 { 2882 return 0; 2883 } 2884 2885 #endif /* CONFIG_SMP */ 2886 2887 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se) 2888 { 2889 #ifdef CONFIG_SCHEDSTATS 2890 struct task_struct *tsk = NULL; 2891 2892 if (entity_is_task(se)) 2893 tsk = task_of(se); 2894 2895 if (se->statistics.sleep_start) { 2896 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.sleep_start; 2897 2898 if ((s64)delta < 0) 2899 delta = 0; 2900 2901 if (unlikely(delta > se->statistics.sleep_max)) 2902 se->statistics.sleep_max = delta; 2903 2904 se->statistics.sleep_start = 0; 2905 se->statistics.sum_sleep_runtime += delta; 2906 2907 if (tsk) { 2908 account_scheduler_latency(tsk, delta >> 10, 1); 2909 trace_sched_stat_sleep(tsk, delta); 2910 } 2911 } 2912 if (se->statistics.block_start) { 2913 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.block_start; 2914 2915 if ((s64)delta < 0) 2916 delta = 0; 2917 2918 if (unlikely(delta > se->statistics.block_max)) 2919 se->statistics.block_max = delta; 2920 2921 se->statistics.block_start = 0; 2922 se->statistics.sum_sleep_runtime += delta; 2923 2924 if (tsk) { 2925 if (tsk->in_iowait) { 2926 se->statistics.iowait_sum += delta; 2927 se->statistics.iowait_count++; 2928 trace_sched_stat_iowait(tsk, delta); 2929 } 2930 2931 trace_sched_stat_blocked(tsk, delta); 2932 2933 /* 2934 * Blocking time is in units of nanosecs, so shift by 2935 * 20 to get a milliseconds-range estimation of the 2936 * amount of time that the task spent sleeping: 2937 */ 2938 if (unlikely(prof_on == SLEEP_PROFILING)) { 2939 profile_hits(SLEEP_PROFILING, 2940 (void *)get_wchan(tsk), 2941 delta >> 20); 2942 } 2943 account_scheduler_latency(tsk, delta >> 10, 0); 2944 } 2945 } 2946 #endif 2947 } 2948 2949 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se) 2950 { 2951 #ifdef CONFIG_SCHED_DEBUG 2952 s64 d = se->vruntime - cfs_rq->min_vruntime; 2953 2954 if (d < 0) 2955 d = -d; 2956 2957 if (d > 3*sysctl_sched_latency) 2958 schedstat_inc(cfs_rq, nr_spread_over); 2959 #endif 2960 } 2961 2962 static void 2963 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial) 2964 { 2965 u64 vruntime = cfs_rq->min_vruntime; 2966 2967 /* 2968 * The 'current' period is already promised to the current tasks, 2969 * however the extra weight of the new task will slow them down a 2970 * little, place the new task so that it fits in the slot that 2971 * stays open at the end. 2972 */ 2973 if (initial && sched_feat(START_DEBIT)) 2974 vruntime += sched_vslice(cfs_rq, se); 2975 2976 /* sleeps up to a single latency don't count. */ 2977 if (!initial) { 2978 unsigned long thresh = sysctl_sched_latency; 2979 2980 /* 2981 * Halve their sleep time's effect, to allow 2982 * for a gentler effect of sleepers: 2983 */ 2984 if (sched_feat(GENTLE_FAIR_SLEEPERS)) 2985 thresh >>= 1; 2986 2987 vruntime -= thresh; 2988 } 2989 2990 /* ensure we never gain time by being placed backwards. */ 2991 se->vruntime = max_vruntime(se->vruntime, vruntime); 2992 } 2993 2994 static void check_enqueue_throttle(struct cfs_rq *cfs_rq); 2995 2996 static void 2997 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags) 2998 { 2999 /* 3000 * Update the normalized vruntime before updating min_vruntime 3001 * through calling update_curr(). 3002 */ 3003 if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_WAKING)) 3004 se->vruntime += cfs_rq->min_vruntime; 3005 3006 /* 3007 * Update run-time statistics of the 'current'. 3008 */ 3009 update_curr(cfs_rq); 3010 enqueue_entity_load_avg(cfs_rq, se, flags & ENQUEUE_WAKEUP); 3011 account_entity_enqueue(cfs_rq, se); 3012 update_cfs_shares(cfs_rq); 3013 3014 if (flags & ENQUEUE_WAKEUP) { 3015 place_entity(cfs_rq, se, 0); 3016 enqueue_sleeper(cfs_rq, se); 3017 } 3018 3019 update_stats_enqueue(cfs_rq, se); 3020 check_spread(cfs_rq, se); 3021 if (se != cfs_rq->curr) 3022 __enqueue_entity(cfs_rq, se); 3023 se->on_rq = 1; 3024 3025 if (cfs_rq->nr_running == 1) { 3026 list_add_leaf_cfs_rq(cfs_rq); 3027 check_enqueue_throttle(cfs_rq); 3028 } 3029 } 3030 3031 static void __clear_buddies_last(struct sched_entity *se) 3032 { 3033 for_each_sched_entity(se) { 3034 struct cfs_rq *cfs_rq = cfs_rq_of(se); 3035 if (cfs_rq->last != se) 3036 break; 3037 3038 cfs_rq->last = NULL; 3039 } 3040 } 3041 3042 static void __clear_buddies_next(struct sched_entity *se) 3043 { 3044 for_each_sched_entity(se) { 3045 struct cfs_rq *cfs_rq = cfs_rq_of(se); 3046 if (cfs_rq->next != se) 3047 break; 3048 3049 cfs_rq->next = NULL; 3050 } 3051 } 3052 3053 static void __clear_buddies_skip(struct sched_entity *se) 3054 { 3055 for_each_sched_entity(se) { 3056 struct cfs_rq *cfs_rq = cfs_rq_of(se); 3057 if (cfs_rq->skip != se) 3058 break; 3059 3060 cfs_rq->skip = NULL; 3061 } 3062 } 3063 3064 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se) 3065 { 3066 if (cfs_rq->last == se) 3067 __clear_buddies_last(se); 3068 3069 if (cfs_rq->next == se) 3070 __clear_buddies_next(se); 3071 3072 if (cfs_rq->skip == se) 3073 __clear_buddies_skip(se); 3074 } 3075 3076 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq); 3077 3078 static void 3079 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags) 3080 { 3081 /* 3082 * Update run-time statistics of the 'current'. 3083 */ 3084 update_curr(cfs_rq); 3085 dequeue_entity_load_avg(cfs_rq, se, flags & DEQUEUE_SLEEP); 3086 3087 update_stats_dequeue(cfs_rq, se); 3088 if (flags & DEQUEUE_SLEEP) { 3089 #ifdef CONFIG_SCHEDSTATS 3090 if (entity_is_task(se)) { 3091 struct task_struct *tsk = task_of(se); 3092 3093 if (tsk->state & TASK_INTERRUPTIBLE) 3094 se->statistics.sleep_start = rq_clock(rq_of(cfs_rq)); 3095 if (tsk->state & TASK_UNINTERRUPTIBLE) 3096 se->statistics.block_start = rq_clock(rq_of(cfs_rq)); 3097 } 3098 #endif 3099 } 3100 3101 clear_buddies(cfs_rq, se); 3102 3103 if (se != cfs_rq->curr) 3104 __dequeue_entity(cfs_rq, se); 3105 se->on_rq = 0; 3106 account_entity_dequeue(cfs_rq, se); 3107 3108 /* 3109 * Normalize the entity after updating the min_vruntime because the 3110 * update can refer to the ->curr item and we need to reflect this 3111 * movement in our normalized position. 3112 */ 3113 if (!(flags & DEQUEUE_SLEEP)) 3114 se->vruntime -= cfs_rq->min_vruntime; 3115 3116 /* return excess runtime on last dequeue */ 3117 return_cfs_rq_runtime(cfs_rq); 3118 3119 update_min_vruntime(cfs_rq); 3120 update_cfs_shares(cfs_rq); 3121 } 3122 3123 /* 3124 * Preempt the current task with a newly woken task if needed: 3125 */ 3126 static void 3127 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr) 3128 { 3129 unsigned long ideal_runtime, delta_exec; 3130 struct sched_entity *se; 3131 s64 delta; 3132 3133 ideal_runtime = sched_slice(cfs_rq, curr); 3134 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime; 3135 if (delta_exec > ideal_runtime) { 3136 resched_curr(rq_of(cfs_rq)); 3137 /* 3138 * The current task ran long enough, ensure it doesn't get 3139 * re-elected due to buddy favours. 3140 */ 3141 clear_buddies(cfs_rq, curr); 3142 return; 3143 } 3144 3145 /* 3146 * Ensure that a task that missed wakeup preemption by a 3147 * narrow margin doesn't have to wait for a full slice. 3148 * This also mitigates buddy induced latencies under load. 3149 */ 3150 if (delta_exec < sysctl_sched_min_granularity) 3151 return; 3152 3153 se = __pick_first_entity(cfs_rq); 3154 delta = curr->vruntime - se->vruntime; 3155 3156 if (delta < 0) 3157 return; 3158 3159 if (delta > ideal_runtime) 3160 resched_curr(rq_of(cfs_rq)); 3161 } 3162 3163 static void 3164 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se) 3165 { 3166 /* 'current' is not kept within the tree. */ 3167 if (se->on_rq) { 3168 /* 3169 * Any task has to be enqueued before it get to execute on 3170 * a CPU. So account for the time it spent waiting on the 3171 * runqueue. 3172 */ 3173 update_stats_wait_end(cfs_rq, se); 3174 __dequeue_entity(cfs_rq, se); 3175 } 3176 3177 update_stats_curr_start(cfs_rq, se); 3178 cfs_rq->curr = se; 3179 #ifdef CONFIG_SCHEDSTATS 3180 /* 3181 * Track our maximum slice length, if the CPU's load is at 3182 * least twice that of our own weight (i.e. dont track it 3183 * when there are only lesser-weight tasks around): 3184 */ 3185 if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) { 3186 se->statistics.slice_max = max(se->statistics.slice_max, 3187 se->sum_exec_runtime - se->prev_sum_exec_runtime); 3188 } 3189 #endif 3190 se->prev_sum_exec_runtime = se->sum_exec_runtime; 3191 } 3192 3193 static int 3194 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se); 3195 3196 /* 3197 * Pick the next process, keeping these things in mind, in this order: 3198 * 1) keep things fair between processes/task groups 3199 * 2) pick the "next" process, since someone really wants that to run 3200 * 3) pick the "last" process, for cache locality 3201 * 4) do not run the "skip" process, if something else is available 3202 */ 3203 static struct sched_entity * 3204 pick_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *curr) 3205 { 3206 struct sched_entity *left = __pick_first_entity(cfs_rq); 3207 struct sched_entity *se; 3208 3209 /* 3210 * If curr is set we have to see if its left of the leftmost entity 3211 * still in the tree, provided there was anything in the tree at all. 3212 */ 3213 if (!left || (curr && entity_before(curr, left))) 3214 left = curr; 3215 3216 se = left; /* ideally we run the leftmost entity */ 3217 3218 /* 3219 * Avoid running the skip buddy, if running something else can 3220 * be done without getting too unfair. 3221 */ 3222 if (cfs_rq->skip == se) { 3223 struct sched_entity *second; 3224 3225 if (se == curr) { 3226 second = __pick_first_entity(cfs_rq); 3227 } else { 3228 second = __pick_next_entity(se); 3229 if (!second || (curr && entity_before(curr, second))) 3230 second = curr; 3231 } 3232 3233 if (second && wakeup_preempt_entity(second, left) < 1) 3234 se = second; 3235 } 3236 3237 /* 3238 * Prefer last buddy, try to return the CPU to a preempted task. 3239 */ 3240 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1) 3241 se = cfs_rq->last; 3242 3243 /* 3244 * Someone really wants this to run. If it's not unfair, run it. 3245 */ 3246 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1) 3247 se = cfs_rq->next; 3248 3249 clear_buddies(cfs_rq, se); 3250 3251 return se; 3252 } 3253 3254 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq); 3255 3256 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev) 3257 { 3258 /* 3259 * If still on the runqueue then deactivate_task() 3260 * was not called and update_curr() has to be done: 3261 */ 3262 if (prev->on_rq) 3263 update_curr(cfs_rq); 3264 3265 /* throttle cfs_rqs exceeding runtime */ 3266 check_cfs_rq_runtime(cfs_rq); 3267 3268 check_spread(cfs_rq, prev); 3269 if (prev->on_rq) { 3270 update_stats_wait_start(cfs_rq, prev); 3271 /* Put 'current' back into the tree. */ 3272 __enqueue_entity(cfs_rq, prev); 3273 /* in !on_rq case, update occurred at dequeue */ 3274 update_entity_load_avg(prev, 1); 3275 } 3276 cfs_rq->curr = NULL; 3277 } 3278 3279 static void 3280 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued) 3281 { 3282 /* 3283 * Update run-time statistics of the 'current'. 3284 */ 3285 update_curr(cfs_rq); 3286 3287 /* 3288 * Ensure that runnable average is periodically updated. 3289 */ 3290 update_entity_load_avg(curr, 1); 3291 update_cfs_rq_blocked_load(cfs_rq, 1); 3292 update_cfs_shares(cfs_rq); 3293 3294 #ifdef CONFIG_SCHED_HRTICK 3295 /* 3296 * queued ticks are scheduled to match the slice, so don't bother 3297 * validating it and just reschedule. 3298 */ 3299 if (queued) { 3300 resched_curr(rq_of(cfs_rq)); 3301 return; 3302 } 3303 /* 3304 * don't let the period tick interfere with the hrtick preemption 3305 */ 3306 if (!sched_feat(DOUBLE_TICK) && 3307 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer)) 3308 return; 3309 #endif 3310 3311 if (cfs_rq->nr_running > 1) 3312 check_preempt_tick(cfs_rq, curr); 3313 } 3314 3315 3316 /************************************************** 3317 * CFS bandwidth control machinery 3318 */ 3319 3320 #ifdef CONFIG_CFS_BANDWIDTH 3321 3322 #ifdef HAVE_JUMP_LABEL 3323 static struct static_key __cfs_bandwidth_used; 3324 3325 static inline bool cfs_bandwidth_used(void) 3326 { 3327 return static_key_false(&__cfs_bandwidth_used); 3328 } 3329 3330 void cfs_bandwidth_usage_inc(void) 3331 { 3332 static_key_slow_inc(&__cfs_bandwidth_used); 3333 } 3334 3335 void cfs_bandwidth_usage_dec(void) 3336 { 3337 static_key_slow_dec(&__cfs_bandwidth_used); 3338 } 3339 #else /* HAVE_JUMP_LABEL */ 3340 static bool cfs_bandwidth_used(void) 3341 { 3342 return true; 3343 } 3344 3345 void cfs_bandwidth_usage_inc(void) {} 3346 void cfs_bandwidth_usage_dec(void) {} 3347 #endif /* HAVE_JUMP_LABEL */ 3348 3349 /* 3350 * default period for cfs group bandwidth. 3351 * default: 0.1s, units: nanoseconds 3352 */ 3353 static inline u64 default_cfs_period(void) 3354 { 3355 return 100000000ULL; 3356 } 3357 3358 static inline u64 sched_cfs_bandwidth_slice(void) 3359 { 3360 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC; 3361 } 3362 3363 /* 3364 * Replenish runtime according to assigned quota and update expiration time. 3365 * We use sched_clock_cpu directly instead of rq->clock to avoid adding 3366 * additional synchronization around rq->lock. 3367 * 3368 * requires cfs_b->lock 3369 */ 3370 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b) 3371 { 3372 u64 now; 3373 3374 if (cfs_b->quota == RUNTIME_INF) 3375 return; 3376 3377 now = sched_clock_cpu(smp_processor_id()); 3378 cfs_b->runtime = cfs_b->quota; 3379 cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period); 3380 } 3381 3382 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg) 3383 { 3384 return &tg->cfs_bandwidth; 3385 } 3386 3387 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */ 3388 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq) 3389 { 3390 if (unlikely(cfs_rq->throttle_count)) 3391 return cfs_rq->throttled_clock_task; 3392 3393 return rq_clock_task(rq_of(cfs_rq)) - cfs_rq->throttled_clock_task_time; 3394 } 3395 3396 /* returns 0 on failure to allocate runtime */ 3397 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq) 3398 { 3399 struct task_group *tg = cfs_rq->tg; 3400 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg); 3401 u64 amount = 0, min_amount, expires; 3402 3403 /* note: this is a positive sum as runtime_remaining <= 0 */ 3404 min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining; 3405 3406 raw_spin_lock(&cfs_b->lock); 3407 if (cfs_b->quota == RUNTIME_INF) 3408 amount = min_amount; 3409 else { 3410 /* 3411 * If the bandwidth pool has become inactive, then at least one 3412 * period must have elapsed since the last consumption. 3413 * Refresh the global state and ensure bandwidth timer becomes 3414 * active. 3415 */ 3416 if (!cfs_b->timer_active) { 3417 __refill_cfs_bandwidth_runtime(cfs_b); 3418 __start_cfs_bandwidth(cfs_b, false); 3419 } 3420 3421 if (cfs_b->runtime > 0) { 3422 amount = min(cfs_b->runtime, min_amount); 3423 cfs_b->runtime -= amount; 3424 cfs_b->idle = 0; 3425 } 3426 } 3427 expires = cfs_b->runtime_expires; 3428 raw_spin_unlock(&cfs_b->lock); 3429 3430 cfs_rq->runtime_remaining += amount; 3431 /* 3432 * we may have advanced our local expiration to account for allowed 3433 * spread between our sched_clock and the one on which runtime was 3434 * issued. 3435 */ 3436 if ((s64)(expires - cfs_rq->runtime_expires) > 0) 3437 cfs_rq->runtime_expires = expires; 3438 3439 return cfs_rq->runtime_remaining > 0; 3440 } 3441 3442 /* 3443 * Note: This depends on the synchronization provided by sched_clock and the 3444 * fact that rq->clock snapshots this value. 3445 */ 3446 static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq) 3447 { 3448 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg); 3449 3450 /* if the deadline is ahead of our clock, nothing to do */ 3451 if (likely((s64)(rq_clock(rq_of(cfs_rq)) - cfs_rq->runtime_expires) < 0)) 3452 return; 3453 3454 if (cfs_rq->runtime_remaining < 0) 3455 return; 3456 3457 /* 3458 * If the local deadline has passed we have to consider the 3459 * possibility that our sched_clock is 'fast' and the global deadline 3460 * has not truly expired. 3461 * 3462 * Fortunately we can check determine whether this the case by checking 3463 * whether the global deadline has advanced. It is valid to compare 3464 * cfs_b->runtime_expires without any locks since we only care about 3465 * exact equality, so a partial write will still work. 3466 */ 3467 3468 if (cfs_rq->runtime_expires != cfs_b->runtime_expires) { 3469 /* extend local deadline, drift is bounded above by 2 ticks */ 3470 cfs_rq->runtime_expires += TICK_NSEC; 3471 } else { 3472 /* global deadline is ahead, expiration has passed */ 3473 cfs_rq->runtime_remaining = 0; 3474 } 3475 } 3476 3477 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) 3478 { 3479 /* dock delta_exec before expiring quota (as it could span periods) */ 3480 cfs_rq->runtime_remaining -= delta_exec; 3481 expire_cfs_rq_runtime(cfs_rq); 3482 3483 if (likely(cfs_rq->runtime_remaining > 0)) 3484 return; 3485 3486 /* 3487 * if we're unable to extend our runtime we resched so that the active 3488 * hierarchy can be throttled 3489 */ 3490 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr)) 3491 resched_curr(rq_of(cfs_rq)); 3492 } 3493 3494 static __always_inline 3495 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) 3496 { 3497 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled) 3498 return; 3499 3500 __account_cfs_rq_runtime(cfs_rq, delta_exec); 3501 } 3502 3503 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq) 3504 { 3505 return cfs_bandwidth_used() && cfs_rq->throttled; 3506 } 3507 3508 /* check whether cfs_rq, or any parent, is throttled */ 3509 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq) 3510 { 3511 return cfs_bandwidth_used() && cfs_rq->throttle_count; 3512 } 3513 3514 /* 3515 * Ensure that neither of the group entities corresponding to src_cpu or 3516 * dest_cpu are members of a throttled hierarchy when performing group 3517 * load-balance operations. 3518 */ 3519 static inline int throttled_lb_pair(struct task_group *tg, 3520 int src_cpu, int dest_cpu) 3521 { 3522 struct cfs_rq *src_cfs_rq, *dest_cfs_rq; 3523 3524 src_cfs_rq = tg->cfs_rq[src_cpu]; 3525 dest_cfs_rq = tg->cfs_rq[dest_cpu]; 3526 3527 return throttled_hierarchy(src_cfs_rq) || 3528 throttled_hierarchy(dest_cfs_rq); 3529 } 3530 3531 /* updated child weight may affect parent so we have to do this bottom up */ 3532 static int tg_unthrottle_up(struct task_group *tg, void *data) 3533 { 3534 struct rq *rq = data; 3535 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)]; 3536 3537 cfs_rq->throttle_count--; 3538 #ifdef CONFIG_SMP 3539 if (!cfs_rq->throttle_count) { 3540 /* adjust cfs_rq_clock_task() */ 3541 cfs_rq->throttled_clock_task_time += rq_clock_task(rq) - 3542 cfs_rq->throttled_clock_task; 3543 } 3544 #endif 3545 3546 return 0; 3547 } 3548 3549 static int tg_throttle_down(struct task_group *tg, void *data) 3550 { 3551 struct rq *rq = data; 3552 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)]; 3553 3554 /* group is entering throttled state, stop time */ 3555 if (!cfs_rq->throttle_count) 3556 cfs_rq->throttled_clock_task = rq_clock_task(rq); 3557 cfs_rq->throttle_count++; 3558 3559 return 0; 3560 } 3561 3562 static void throttle_cfs_rq(struct cfs_rq *cfs_rq) 3563 { 3564 struct rq *rq = rq_of(cfs_rq); 3565 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg); 3566 struct sched_entity *se; 3567 long task_delta, dequeue = 1; 3568 3569 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))]; 3570 3571 /* freeze hierarchy runnable averages while throttled */ 3572 rcu_read_lock(); 3573 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq); 3574 rcu_read_unlock(); 3575 3576 task_delta = cfs_rq->h_nr_running; 3577 for_each_sched_entity(se) { 3578 struct cfs_rq *qcfs_rq = cfs_rq_of(se); 3579 /* throttled entity or throttle-on-deactivate */ 3580 if (!se->on_rq) 3581 break; 3582 3583 if (dequeue) 3584 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP); 3585 qcfs_rq->h_nr_running -= task_delta; 3586 3587 if (qcfs_rq->load.weight) 3588 dequeue = 0; 3589 } 3590 3591 if (!se) 3592 sub_nr_running(rq, task_delta); 3593 3594 cfs_rq->throttled = 1; 3595 cfs_rq->throttled_clock = rq_clock(rq); 3596 raw_spin_lock(&cfs_b->lock); 3597 /* 3598 * Add to the _head_ of the list, so that an already-started 3599 * distribute_cfs_runtime will not see us 3600 */ 3601 list_add_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq); 3602 if (!cfs_b->timer_active) 3603 __start_cfs_bandwidth(cfs_b, false); 3604 raw_spin_unlock(&cfs_b->lock); 3605 } 3606 3607 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq) 3608 { 3609 struct rq *rq = rq_of(cfs_rq); 3610 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg); 3611 struct sched_entity *se; 3612 int enqueue = 1; 3613 long task_delta; 3614 3615 se = cfs_rq->tg->se[cpu_of(rq)]; 3616 3617 cfs_rq->throttled = 0; 3618 3619 update_rq_clock(rq); 3620 3621 raw_spin_lock(&cfs_b->lock); 3622 cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock; 3623 list_del_rcu(&cfs_rq->throttled_list); 3624 raw_spin_unlock(&cfs_b->lock); 3625 3626 /* update hierarchical throttle state */ 3627 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq); 3628 3629 if (!cfs_rq->load.weight) 3630 return; 3631 3632 task_delta = cfs_rq->h_nr_running; 3633 for_each_sched_entity(se) { 3634 if (se->on_rq) 3635 enqueue = 0; 3636 3637 cfs_rq = cfs_rq_of(se); 3638 if (enqueue) 3639 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP); 3640 cfs_rq->h_nr_running += task_delta; 3641 3642 if (cfs_rq_throttled(cfs_rq)) 3643 break; 3644 } 3645 3646 if (!se) 3647 add_nr_running(rq, task_delta); 3648 3649 /* determine whether we need to wake up potentially idle cpu */ 3650 if (rq->curr == rq->idle && rq->cfs.nr_running) 3651 resched_curr(rq); 3652 } 3653 3654 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b, 3655 u64 remaining, u64 expires) 3656 { 3657 struct cfs_rq *cfs_rq; 3658 u64 runtime; 3659 u64 starting_runtime = remaining; 3660 3661 rcu_read_lock(); 3662 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq, 3663 throttled_list) { 3664 struct rq *rq = rq_of(cfs_rq); 3665 3666 raw_spin_lock(&rq->lock); 3667 if (!cfs_rq_throttled(cfs_rq)) 3668 goto next; 3669 3670 runtime = -cfs_rq->runtime_remaining + 1; 3671 if (runtime > remaining) 3672 runtime = remaining; 3673 remaining -= runtime; 3674 3675 cfs_rq->runtime_remaining += runtime; 3676 cfs_rq->runtime_expires = expires; 3677 3678 /* we check whether we're throttled above */ 3679 if (cfs_rq->runtime_remaining > 0) 3680 unthrottle_cfs_rq(cfs_rq); 3681 3682 next: 3683 raw_spin_unlock(&rq->lock); 3684 3685 if (!remaining) 3686 break; 3687 } 3688 rcu_read_unlock(); 3689 3690 return starting_runtime - remaining; 3691 } 3692 3693 /* 3694 * Responsible for refilling a task_group's bandwidth and unthrottling its 3695 * cfs_rqs as appropriate. If there has been no activity within the last 3696 * period the timer is deactivated until scheduling resumes; cfs_b->idle is 3697 * used to track this state. 3698 */ 3699 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun) 3700 { 3701 u64 runtime, runtime_expires; 3702 int throttled; 3703 3704 /* no need to continue the timer with no bandwidth constraint */ 3705 if (cfs_b->quota == RUNTIME_INF) 3706 goto out_deactivate; 3707 3708 throttled = !list_empty(&cfs_b->throttled_cfs_rq); 3709 cfs_b->nr_periods += overrun; 3710 3711 /* 3712 * idle depends on !throttled (for the case of a large deficit), and if 3713 * we're going inactive then everything else can be deferred 3714 */ 3715 if (cfs_b->idle && !throttled) 3716 goto out_deactivate; 3717 3718 /* 3719 * if we have relooped after returning idle once, we need to update our 3720 * status as actually running, so that other cpus doing 3721 * __start_cfs_bandwidth will stop trying to cancel us. 3722 */ 3723 cfs_b->timer_active = 1; 3724 3725 __refill_cfs_bandwidth_runtime(cfs_b); 3726 3727 if (!throttled) { 3728 /* mark as potentially idle for the upcoming period */ 3729 cfs_b->idle = 1; 3730 return 0; 3731 } 3732 3733 /* account preceding periods in which throttling occurred */ 3734 cfs_b->nr_throttled += overrun; 3735 3736 runtime_expires = cfs_b->runtime_expires; 3737 3738 /* 3739 * This check is repeated as we are holding onto the new bandwidth while 3740 * we unthrottle. This can potentially race with an unthrottled group 3741 * trying to acquire new bandwidth from the global pool. This can result 3742 * in us over-using our runtime if it is all used during this loop, but 3743 * only by limited amounts in that extreme case. 3744 */ 3745 while (throttled && cfs_b->runtime > 0) { 3746 runtime = cfs_b->runtime; 3747 raw_spin_unlock(&cfs_b->lock); 3748 /* we can't nest cfs_b->lock while distributing bandwidth */ 3749 runtime = distribute_cfs_runtime(cfs_b, runtime, 3750 runtime_expires); 3751 raw_spin_lock(&cfs_b->lock); 3752 3753 throttled = !list_empty(&cfs_b->throttled_cfs_rq); 3754 3755 cfs_b->runtime -= min(runtime, cfs_b->runtime); 3756 } 3757 3758 /* 3759 * While we are ensured activity in the period following an 3760 * unthrottle, this also covers the case in which the new bandwidth is 3761 * insufficient to cover the existing bandwidth deficit. (Forcing the 3762 * timer to remain active while there are any throttled entities.) 3763 */ 3764 cfs_b->idle = 0; 3765 3766 return 0; 3767 3768 out_deactivate: 3769 cfs_b->timer_active = 0; 3770 return 1; 3771 } 3772 3773 /* a cfs_rq won't donate quota below this amount */ 3774 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC; 3775 /* minimum remaining period time to redistribute slack quota */ 3776 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC; 3777 /* how long we wait to gather additional slack before distributing */ 3778 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC; 3779 3780 /* 3781 * Are we near the end of the current quota period? 3782 * 3783 * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the 3784 * hrtimer base being cleared by __hrtimer_start_range_ns. In the case of 3785 * migrate_hrtimers, base is never cleared, so we are fine. 3786 */ 3787 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire) 3788 { 3789 struct hrtimer *refresh_timer = &cfs_b->period_timer; 3790 u64 remaining; 3791 3792 /* if the call-back is running a quota refresh is already occurring */ 3793 if (hrtimer_callback_running(refresh_timer)) 3794 return 1; 3795 3796 /* is a quota refresh about to occur? */ 3797 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer)); 3798 if (remaining < min_expire) 3799 return 1; 3800 3801 return 0; 3802 } 3803 3804 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b) 3805 { 3806 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration; 3807 3808 /* if there's a quota refresh soon don't bother with slack */ 3809 if (runtime_refresh_within(cfs_b, min_left)) 3810 return; 3811 3812 start_bandwidth_timer(&cfs_b->slack_timer, 3813 ns_to_ktime(cfs_bandwidth_slack_period)); 3814 } 3815 3816 /* we know any runtime found here is valid as update_curr() precedes return */ 3817 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq) 3818 { 3819 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg); 3820 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime; 3821 3822 if (slack_runtime <= 0) 3823 return; 3824 3825 raw_spin_lock(&cfs_b->lock); 3826 if (cfs_b->quota != RUNTIME_INF && 3827 cfs_rq->runtime_expires == cfs_b->runtime_expires) { 3828 cfs_b->runtime += slack_runtime; 3829 3830 /* we are under rq->lock, defer unthrottling using a timer */ 3831 if (cfs_b->runtime > sched_cfs_bandwidth_slice() && 3832 !list_empty(&cfs_b->throttled_cfs_rq)) 3833 start_cfs_slack_bandwidth(cfs_b); 3834 } 3835 raw_spin_unlock(&cfs_b->lock); 3836 3837 /* even if it's not valid for return we don't want to try again */ 3838 cfs_rq->runtime_remaining -= slack_runtime; 3839 } 3840 3841 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) 3842 { 3843 if (!cfs_bandwidth_used()) 3844 return; 3845 3846 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running) 3847 return; 3848 3849 __return_cfs_rq_runtime(cfs_rq); 3850 } 3851 3852 /* 3853 * This is done with a timer (instead of inline with bandwidth return) since 3854 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs. 3855 */ 3856 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b) 3857 { 3858 u64 runtime = 0, slice = sched_cfs_bandwidth_slice(); 3859 u64 expires; 3860 3861 /* confirm we're still not at a refresh boundary */ 3862 raw_spin_lock(&cfs_b->lock); 3863 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) { 3864 raw_spin_unlock(&cfs_b->lock); 3865 return; 3866 } 3867 3868 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice) 3869 runtime = cfs_b->runtime; 3870 3871 expires = cfs_b->runtime_expires; 3872 raw_spin_unlock(&cfs_b->lock); 3873 3874 if (!runtime) 3875 return; 3876 3877 runtime = distribute_cfs_runtime(cfs_b, runtime, expires); 3878 3879 raw_spin_lock(&cfs_b->lock); 3880 if (expires == cfs_b->runtime_expires) 3881 cfs_b->runtime -= min(runtime, cfs_b->runtime); 3882 raw_spin_unlock(&cfs_b->lock); 3883 } 3884 3885 /* 3886 * When a group wakes up we want to make sure that its quota is not already 3887 * expired/exceeded, otherwise it may be allowed to steal additional ticks of 3888 * runtime as update_curr() throttling can not not trigger until it's on-rq. 3889 */ 3890 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) 3891 { 3892 if (!cfs_bandwidth_used()) 3893 return; 3894 3895 /* an active group must be handled by the update_curr()->put() path */ 3896 if (!cfs_rq->runtime_enabled || cfs_rq->curr) 3897 return; 3898 3899 /* ensure the group is not already throttled */ 3900 if (cfs_rq_throttled(cfs_rq)) 3901 return; 3902 3903 /* update runtime allocation */ 3904 account_cfs_rq_runtime(cfs_rq, 0); 3905 if (cfs_rq->runtime_remaining <= 0) 3906 throttle_cfs_rq(cfs_rq); 3907 } 3908 3909 /* conditionally throttle active cfs_rq's from put_prev_entity() */ 3910 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) 3911 { 3912 if (!cfs_bandwidth_used()) 3913 return false; 3914 3915 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0)) 3916 return false; 3917 3918 /* 3919 * it's possible for a throttled entity to be forced into a running 3920 * state (e.g. set_curr_task), in this case we're finished. 3921 */ 3922 if (cfs_rq_throttled(cfs_rq)) 3923 return true; 3924 3925 throttle_cfs_rq(cfs_rq); 3926 return true; 3927 } 3928 3929 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer) 3930 { 3931 struct cfs_bandwidth *cfs_b = 3932 container_of(timer, struct cfs_bandwidth, slack_timer); 3933 do_sched_cfs_slack_timer(cfs_b); 3934 3935 return HRTIMER_NORESTART; 3936 } 3937 3938 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer) 3939 { 3940 struct cfs_bandwidth *cfs_b = 3941 container_of(timer, struct cfs_bandwidth, period_timer); 3942 ktime_t now; 3943 int overrun; 3944 int idle = 0; 3945 3946 raw_spin_lock(&cfs_b->lock); 3947 for (;;) { 3948 now = hrtimer_cb_get_time(timer); 3949 overrun = hrtimer_forward(timer, now, cfs_b->period); 3950 3951 if (!overrun) 3952 break; 3953 3954 idle = do_sched_cfs_period_timer(cfs_b, overrun); 3955 } 3956 raw_spin_unlock(&cfs_b->lock); 3957 3958 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART; 3959 } 3960 3961 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) 3962 { 3963 raw_spin_lock_init(&cfs_b->lock); 3964 cfs_b->runtime = 0; 3965 cfs_b->quota = RUNTIME_INF; 3966 cfs_b->period = ns_to_ktime(default_cfs_period()); 3967 3968 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq); 3969 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); 3970 cfs_b->period_timer.function = sched_cfs_period_timer; 3971 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); 3972 cfs_b->slack_timer.function = sched_cfs_slack_timer; 3973 } 3974 3975 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) 3976 { 3977 cfs_rq->runtime_enabled = 0; 3978 INIT_LIST_HEAD(&cfs_rq->throttled_list); 3979 } 3980 3981 /* requires cfs_b->lock, may release to reprogram timer */ 3982 void __start_cfs_bandwidth(struct cfs_bandwidth *cfs_b, bool force) 3983 { 3984 /* 3985 * The timer may be active because we're trying to set a new bandwidth 3986 * period or because we're racing with the tear-down path 3987 * (timer_active==0 becomes visible before the hrtimer call-back 3988 * terminates). In either case we ensure that it's re-programmed 3989 */ 3990 while (unlikely(hrtimer_active(&cfs_b->period_timer)) && 3991 hrtimer_try_to_cancel(&cfs_b->period_timer) < 0) { 3992 /* bounce the lock to allow do_sched_cfs_period_timer to run */ 3993 raw_spin_unlock(&cfs_b->lock); 3994 cpu_relax(); 3995 raw_spin_lock(&cfs_b->lock); 3996 /* if someone else restarted the timer then we're done */ 3997 if (!force && cfs_b->timer_active) 3998 return; 3999 } 4000 4001 cfs_b->timer_active = 1; 4002 start_bandwidth_timer(&cfs_b->period_timer, cfs_b->period); 4003 } 4004 4005 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) 4006 { 4007 /* init_cfs_bandwidth() was not called */ 4008 if (!cfs_b->throttled_cfs_rq.next) 4009 return; 4010 4011 hrtimer_cancel(&cfs_b->period_timer); 4012 hrtimer_cancel(&cfs_b->slack_timer); 4013 } 4014 4015 static void __maybe_unused update_runtime_enabled(struct rq *rq) 4016 { 4017 struct cfs_rq *cfs_rq; 4018 4019 for_each_leaf_cfs_rq(rq, cfs_rq) { 4020 struct cfs_bandwidth *cfs_b = &cfs_rq->tg->cfs_bandwidth; 4021 4022 raw_spin_lock(&cfs_b->lock); 4023 cfs_rq->runtime_enabled = cfs_b->quota != RUNTIME_INF; 4024 raw_spin_unlock(&cfs_b->lock); 4025 } 4026 } 4027 4028 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq) 4029 { 4030 struct cfs_rq *cfs_rq; 4031 4032 for_each_leaf_cfs_rq(rq, cfs_rq) { 4033 if (!cfs_rq->runtime_enabled) 4034 continue; 4035 4036 /* 4037 * clock_task is not advancing so we just need to make sure 4038 * there's some valid quota amount 4039 */ 4040 cfs_rq->runtime_remaining = 1; 4041 /* 4042 * Offline rq is schedulable till cpu is completely disabled 4043 * in take_cpu_down(), so we prevent new cfs throttling here. 4044 */ 4045 cfs_rq->runtime_enabled = 0; 4046 4047 if (cfs_rq_throttled(cfs_rq)) 4048 unthrottle_cfs_rq(cfs_rq); 4049 } 4050 } 4051 4052 #else /* CONFIG_CFS_BANDWIDTH */ 4053 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq) 4054 { 4055 return rq_clock_task(rq_of(cfs_rq)); 4056 } 4057 4058 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) {} 4059 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) { return false; } 4060 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {} 4061 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {} 4062 4063 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq) 4064 { 4065 return 0; 4066 } 4067 4068 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq) 4069 { 4070 return 0; 4071 } 4072 4073 static inline int throttled_lb_pair(struct task_group *tg, 4074 int src_cpu, int dest_cpu) 4075 { 4076 return 0; 4077 } 4078 4079 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {} 4080 4081 #ifdef CONFIG_FAIR_GROUP_SCHED 4082 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {} 4083 #endif 4084 4085 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg) 4086 { 4087 return NULL; 4088 } 4089 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {} 4090 static inline void update_runtime_enabled(struct rq *rq) {} 4091 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {} 4092 4093 #endif /* CONFIG_CFS_BANDWIDTH */ 4094 4095 /************************************************** 4096 * CFS operations on tasks: 4097 */ 4098 4099 #ifdef CONFIG_SCHED_HRTICK 4100 static void hrtick_start_fair(struct rq *rq, struct task_struct *p) 4101 { 4102 struct sched_entity *se = &p->se; 4103 struct cfs_rq *cfs_rq = cfs_rq_of(se); 4104 4105 WARN_ON(task_rq(p) != rq); 4106 4107 if (cfs_rq->nr_running > 1) { 4108 u64 slice = sched_slice(cfs_rq, se); 4109 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime; 4110 s64 delta = slice - ran; 4111 4112 if (delta < 0) { 4113 if (rq->curr == p) 4114 resched_curr(rq); 4115 return; 4116 } 4117 hrtick_start(rq, delta); 4118 } 4119 } 4120 4121 /* 4122 * called from enqueue/dequeue and updates the hrtick when the 4123 * current task is from our class and nr_running is low enough 4124 * to matter. 4125 */ 4126 static void hrtick_update(struct rq *rq) 4127 { 4128 struct task_struct *curr = rq->curr; 4129 4130 if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class) 4131 return; 4132 4133 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency) 4134 hrtick_start_fair(rq, curr); 4135 } 4136 #else /* !CONFIG_SCHED_HRTICK */ 4137 static inline void 4138 hrtick_start_fair(struct rq *rq, struct task_struct *p) 4139 { 4140 } 4141 4142 static inline void hrtick_update(struct rq *rq) 4143 { 4144 } 4145 #endif 4146 4147 /* 4148 * The enqueue_task method is called before nr_running is 4149 * increased. Here we update the fair scheduling stats and 4150 * then put the task into the rbtree: 4151 */ 4152 static void 4153 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags) 4154 { 4155 struct cfs_rq *cfs_rq; 4156 struct sched_entity *se = &p->se; 4157 4158 for_each_sched_entity(se) { 4159 if (se->on_rq) 4160 break; 4161 cfs_rq = cfs_rq_of(se); 4162 enqueue_entity(cfs_rq, se, flags); 4163 4164 /* 4165 * end evaluation on encountering a throttled cfs_rq 4166 * 4167 * note: in the case of encountering a throttled cfs_rq we will 4168 * post the final h_nr_running increment below. 4169 */ 4170 if (cfs_rq_throttled(cfs_rq)) 4171 break; 4172 cfs_rq->h_nr_running++; 4173 4174 flags = ENQUEUE_WAKEUP; 4175 } 4176 4177 for_each_sched_entity(se) { 4178 cfs_rq = cfs_rq_of(se); 4179 cfs_rq->h_nr_running++; 4180 4181 if (cfs_rq_throttled(cfs_rq)) 4182 break; 4183 4184 update_cfs_shares(cfs_rq); 4185 update_entity_load_avg(se, 1); 4186 } 4187 4188 if (!se) { 4189 update_rq_runnable_avg(rq, rq->nr_running); 4190 add_nr_running(rq, 1); 4191 } 4192 hrtick_update(rq); 4193 } 4194 4195 static void set_next_buddy(struct sched_entity *se); 4196 4197 /* 4198 * The dequeue_task method is called before nr_running is 4199 * decreased. We remove the task from the rbtree and 4200 * update the fair scheduling stats: 4201 */ 4202 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags) 4203 { 4204 struct cfs_rq *cfs_rq; 4205 struct sched_entity *se = &p->se; 4206 int task_sleep = flags & DEQUEUE_SLEEP; 4207 4208 for_each_sched_entity(se) { 4209 cfs_rq = cfs_rq_of(se); 4210 dequeue_entity(cfs_rq, se, flags); 4211 4212 /* 4213 * end evaluation on encountering a throttled cfs_rq 4214 * 4215 * note: in the case of encountering a throttled cfs_rq we will 4216 * post the final h_nr_running decrement below. 4217 */ 4218 if (cfs_rq_throttled(cfs_rq)) 4219 break; 4220 cfs_rq->h_nr_running--; 4221 4222 /* Don't dequeue parent if it has other entities besides us */ 4223 if (cfs_rq->load.weight) { 4224 /* 4225 * Bias pick_next to pick a task from this cfs_rq, as 4226 * p is sleeping when it is within its sched_slice. 4227 */ 4228 if (task_sleep && parent_entity(se)) 4229 set_next_buddy(parent_entity(se)); 4230 4231 /* avoid re-evaluating load for this entity */ 4232 se = parent_entity(se); 4233 break; 4234 } 4235 flags |= DEQUEUE_SLEEP; 4236 } 4237 4238 for_each_sched_entity(se) { 4239 cfs_rq = cfs_rq_of(se); 4240 cfs_rq->h_nr_running--; 4241 4242 if (cfs_rq_throttled(cfs_rq)) 4243 break; 4244 4245 update_cfs_shares(cfs_rq); 4246 update_entity_load_avg(se, 1); 4247 } 4248 4249 if (!se) { 4250 sub_nr_running(rq, 1); 4251 update_rq_runnable_avg(rq, 1); 4252 } 4253 hrtick_update(rq); 4254 } 4255 4256 #ifdef CONFIG_SMP 4257 /* Used instead of source_load when we know the type == 0 */ 4258 static unsigned long weighted_cpuload(const int cpu) 4259 { 4260 return cpu_rq(cpu)->cfs.runnable_load_avg; 4261 } 4262 4263 /* 4264 * Return a low guess at the load of a migration-source cpu weighted 4265 * according to the scheduling class and "nice" value. 4266 * 4267 * We want to under-estimate the load of migration sources, to 4268 * balance conservatively. 4269 */ 4270 static unsigned long source_load(int cpu, int type) 4271 { 4272 struct rq *rq = cpu_rq(cpu); 4273 unsigned long total = weighted_cpuload(cpu); 4274 4275 if (type == 0 || !sched_feat(LB_BIAS)) 4276 return total; 4277 4278 return min(rq->cpu_load[type-1], total); 4279 } 4280 4281 /* 4282 * Return a high guess at the load of a migration-target cpu weighted 4283 * according to the scheduling class and "nice" value. 4284 */ 4285 static unsigned long target_load(int cpu, int type) 4286 { 4287 struct rq *rq = cpu_rq(cpu); 4288 unsigned long total = weighted_cpuload(cpu); 4289 4290 if (type == 0 || !sched_feat(LB_BIAS)) 4291 return total; 4292 4293 return max(rq->cpu_load[type-1], total); 4294 } 4295 4296 static unsigned long capacity_of(int cpu) 4297 { 4298 return cpu_rq(cpu)->cpu_capacity; 4299 } 4300 4301 static unsigned long cpu_avg_load_per_task(int cpu) 4302 { 4303 struct rq *rq = cpu_rq(cpu); 4304 unsigned long nr_running = ACCESS_ONCE(rq->cfs.h_nr_running); 4305 unsigned long load_avg = rq->cfs.runnable_load_avg; 4306 4307 if (nr_running) 4308 return load_avg / nr_running; 4309 4310 return 0; 4311 } 4312 4313 static void record_wakee(struct task_struct *p) 4314 { 4315 /* 4316 * Rough decay (wiping) for cost saving, don't worry 4317 * about the boundary, really active task won't care 4318 * about the loss. 4319 */ 4320 if (time_after(jiffies, current->wakee_flip_decay_ts + HZ)) { 4321 current->wakee_flips >>= 1; 4322 current->wakee_flip_decay_ts = jiffies; 4323 } 4324 4325 if (current->last_wakee != p) { 4326 current->last_wakee = p; 4327 current->wakee_flips++; 4328 } 4329 } 4330 4331 static void task_waking_fair(struct task_struct *p) 4332 { 4333 struct sched_entity *se = &p->se; 4334 struct cfs_rq *cfs_rq = cfs_rq_of(se); 4335 u64 min_vruntime; 4336 4337 #ifndef CONFIG_64BIT 4338 u64 min_vruntime_copy; 4339 4340 do { 4341 min_vruntime_copy = cfs_rq->min_vruntime_copy; 4342 smp_rmb(); 4343 min_vruntime = cfs_rq->min_vruntime; 4344 } while (min_vruntime != min_vruntime_copy); 4345 #else 4346 min_vruntime = cfs_rq->min_vruntime; 4347 #endif 4348 4349 se->vruntime -= min_vruntime; 4350 record_wakee(p); 4351 } 4352 4353 #ifdef CONFIG_FAIR_GROUP_SCHED 4354 /* 4355 * effective_load() calculates the load change as seen from the root_task_group 4356 * 4357 * Adding load to a group doesn't make a group heavier, but can cause movement 4358 * of group shares between cpus. Assuming the shares were perfectly aligned one 4359 * can calculate the shift in shares. 4360 * 4361 * Calculate the effective load difference if @wl is added (subtracted) to @tg 4362 * on this @cpu and results in a total addition (subtraction) of @wg to the 4363 * total group weight. 4364 * 4365 * Given a runqueue weight distribution (rw_i) we can compute a shares 4366 * distribution (s_i) using: 4367 * 4368 * s_i = rw_i / \Sum rw_j (1) 4369 * 4370 * Suppose we have 4 CPUs and our @tg is a direct child of the root group and 4371 * has 7 equal weight tasks, distributed as below (rw_i), with the resulting 4372 * shares distribution (s_i): 4373 * 4374 * rw_i = { 2, 4, 1, 0 } 4375 * s_i = { 2/7, 4/7, 1/7, 0 } 4376 * 4377 * As per wake_affine() we're interested in the load of two CPUs (the CPU the 4378 * task used to run on and the CPU the waker is running on), we need to 4379 * compute the effect of waking a task on either CPU and, in case of a sync 4380 * wakeup, compute the effect of the current task going to sleep. 4381 * 4382 * So for a change of @wl to the local @cpu with an overall group weight change 4383 * of @wl we can compute the new shares distribution (s'_i) using: 4384 * 4385 * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2) 4386 * 4387 * Suppose we're interested in CPUs 0 and 1, and want to compute the load 4388 * differences in waking a task to CPU 0. The additional task changes the 4389 * weight and shares distributions like: 4390 * 4391 * rw'_i = { 3, 4, 1, 0 } 4392 * s'_i = { 3/8, 4/8, 1/8, 0 } 4393 * 4394 * We can then compute the difference in effective weight by using: 4395 * 4396 * dw_i = S * (s'_i - s_i) (3) 4397 * 4398 * Where 'S' is the group weight as seen by its parent. 4399 * 4400 * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7) 4401 * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 - 4402 * 4/7) times the weight of the group. 4403 */ 4404 static long effective_load(struct task_group *tg, int cpu, long wl, long wg) 4405 { 4406 struct sched_entity *se = tg->se[cpu]; 4407 4408 if (!tg->parent) /* the trivial, non-cgroup case */ 4409 return wl; 4410 4411 for_each_sched_entity(se) { 4412 long w, W; 4413 4414 tg = se->my_q->tg; 4415 4416 /* 4417 * W = @wg + \Sum rw_j 4418 */ 4419 W = wg + calc_tg_weight(tg, se->my_q); 4420 4421 /* 4422 * w = rw_i + @wl 4423 */ 4424 w = se->my_q->load.weight + wl; 4425 4426 /* 4427 * wl = S * s'_i; see (2) 4428 */ 4429 if (W > 0 && w < W) 4430 wl = (w * (long)tg->shares) / W; 4431 else 4432 wl = tg->shares; 4433 4434 /* 4435 * Per the above, wl is the new se->load.weight value; since 4436 * those are clipped to [MIN_SHARES, ...) do so now. See 4437 * calc_cfs_shares(). 4438 */ 4439 if (wl < MIN_SHARES) 4440 wl = MIN_SHARES; 4441 4442 /* 4443 * wl = dw_i = S * (s'_i - s_i); see (3) 4444 */ 4445 wl -= se->load.weight; 4446 4447 /* 4448 * Recursively apply this logic to all parent groups to compute 4449 * the final effective load change on the root group. Since 4450 * only the @tg group gets extra weight, all parent groups can 4451 * only redistribute existing shares. @wl is the shift in shares 4452 * resulting from this level per the above. 4453 */ 4454 wg = 0; 4455 } 4456 4457 return wl; 4458 } 4459 #else 4460 4461 static long effective_load(struct task_group *tg, int cpu, long wl, long wg) 4462 { 4463 return wl; 4464 } 4465 4466 #endif 4467 4468 static int wake_wide(struct task_struct *p) 4469 { 4470 int factor = this_cpu_read(sd_llc_size); 4471 4472 /* 4473 * Yeah, it's the switching-frequency, could means many wakee or 4474 * rapidly switch, use factor here will just help to automatically 4475 * adjust the loose-degree, so bigger node will lead to more pull. 4476 */ 4477 if (p->wakee_flips > factor) { 4478 /* 4479 * wakee is somewhat hot, it needs certain amount of cpu 4480 * resource, so if waker is far more hot, prefer to leave 4481 * it alone. 4482 */ 4483 if (current->wakee_flips > (factor * p->wakee_flips)) 4484 return 1; 4485 } 4486 4487 return 0; 4488 } 4489 4490 static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync) 4491 { 4492 s64 this_load, load; 4493 s64 this_eff_load, prev_eff_load; 4494 int idx, this_cpu, prev_cpu; 4495 struct task_group *tg; 4496 unsigned long weight; 4497 int balanced; 4498 4499 /* 4500 * If we wake multiple tasks be careful to not bounce 4501 * ourselves around too much. 4502 */ 4503 if (wake_wide(p)) 4504 return 0; 4505 4506 idx = sd->wake_idx; 4507 this_cpu = smp_processor_id(); 4508 prev_cpu = task_cpu(p); 4509 load = source_load(prev_cpu, idx); 4510 this_load = target_load(this_cpu, idx); 4511 4512 /* 4513 * If sync wakeup then subtract the (maximum possible) 4514 * effect of the currently running task from the load 4515 * of the current CPU: 4516 */ 4517 if (sync) { 4518 tg = task_group(current); 4519 weight = current->se.load.weight; 4520 4521 this_load += effective_load(tg, this_cpu, -weight, -weight); 4522 load += effective_load(tg, prev_cpu, 0, -weight); 4523 } 4524 4525 tg = task_group(p); 4526 weight = p->se.load.weight; 4527 4528 /* 4529 * In low-load situations, where prev_cpu is idle and this_cpu is idle 4530 * due to the sync cause above having dropped this_load to 0, we'll 4531 * always have an imbalance, but there's really nothing you can do 4532 * about that, so that's good too. 4533 * 4534 * Otherwise check if either cpus are near enough in load to allow this 4535 * task to be woken on this_cpu. 4536 */ 4537 this_eff_load = 100; 4538 this_eff_load *= capacity_of(prev_cpu); 4539 4540 prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2; 4541 prev_eff_load *= capacity_of(this_cpu); 4542 4543 if (this_load > 0) { 4544 this_eff_load *= this_load + 4545 effective_load(tg, this_cpu, weight, weight); 4546 4547 prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight); 4548 } 4549 4550 balanced = this_eff_load <= prev_eff_load; 4551 4552 schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts); 4553 4554 if (!balanced) 4555 return 0; 4556 4557 schedstat_inc(sd, ttwu_move_affine); 4558 schedstat_inc(p, se.statistics.nr_wakeups_affine); 4559 4560 return 1; 4561 } 4562 4563 /* 4564 * find_idlest_group finds and returns the least busy CPU group within the 4565 * domain. 4566 */ 4567 static struct sched_group * 4568 find_idlest_group(struct sched_domain *sd, struct task_struct *p, 4569 int this_cpu, int sd_flag) 4570 { 4571 struct sched_group *idlest = NULL, *group = sd->groups; 4572 unsigned long min_load = ULONG_MAX, this_load = 0; 4573 int load_idx = sd->forkexec_idx; 4574 int imbalance = 100 + (sd->imbalance_pct-100)/2; 4575 4576 if (sd_flag & SD_BALANCE_WAKE) 4577 load_idx = sd->wake_idx; 4578 4579 do { 4580 unsigned long load, avg_load; 4581 int local_group; 4582 int i; 4583 4584 /* Skip over this group if it has no CPUs allowed */ 4585 if (!cpumask_intersects(sched_group_cpus(group), 4586 tsk_cpus_allowed(p))) 4587 continue; 4588 4589 local_group = cpumask_test_cpu(this_cpu, 4590 sched_group_cpus(group)); 4591 4592 /* Tally up the load of all CPUs in the group */ 4593 avg_load = 0; 4594 4595 for_each_cpu(i, sched_group_cpus(group)) { 4596 /* Bias balancing toward cpus of our domain */ 4597 if (local_group) 4598 load = source_load(i, load_idx); 4599 else 4600 load = target_load(i, load_idx); 4601 4602 avg_load += load; 4603 } 4604 4605 /* Adjust by relative CPU capacity of the group */ 4606 avg_load = (avg_load * SCHED_CAPACITY_SCALE) / group->sgc->capacity; 4607 4608 if (local_group) { 4609 this_load = avg_load; 4610 } else if (avg_load < min_load) { 4611 min_load = avg_load; 4612 idlest = group; 4613 } 4614 } while (group = group->next, group != sd->groups); 4615 4616 if (!idlest || 100*this_load < imbalance*min_load) 4617 return NULL; 4618 return idlest; 4619 } 4620 4621 /* 4622 * find_idlest_cpu - find the idlest cpu among the cpus in group. 4623 */ 4624 static int 4625 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu) 4626 { 4627 unsigned long load, min_load = ULONG_MAX; 4628 unsigned int min_exit_latency = UINT_MAX; 4629 u64 latest_idle_timestamp = 0; 4630 int least_loaded_cpu = this_cpu; 4631 int shallowest_idle_cpu = -1; 4632 int i; 4633 4634 /* Traverse only the allowed CPUs */ 4635 for_each_cpu_and(i, sched_group_cpus(group), tsk_cpus_allowed(p)) { 4636 if (idle_cpu(i)) { 4637 struct rq *rq = cpu_rq(i); 4638 struct cpuidle_state *idle = idle_get_state(rq); 4639 if (idle && idle->exit_latency < min_exit_latency) { 4640 /* 4641 * We give priority to a CPU whose idle state 4642 * has the smallest exit latency irrespective 4643 * of any idle timestamp. 4644 */ 4645 min_exit_latency = idle->exit_latency; 4646 latest_idle_timestamp = rq->idle_stamp; 4647 shallowest_idle_cpu = i; 4648 } else if ((!idle || idle->exit_latency == min_exit_latency) && 4649 rq->idle_stamp > latest_idle_timestamp) { 4650 /* 4651 * If equal or no active idle state, then 4652 * the most recently idled CPU might have 4653 * a warmer cache. 4654 */ 4655 latest_idle_timestamp = rq->idle_stamp; 4656 shallowest_idle_cpu = i; 4657 } 4658 } else if (shallowest_idle_cpu == -1) { 4659 load = weighted_cpuload(i); 4660 if (load < min_load || (load == min_load && i == this_cpu)) { 4661 min_load = load; 4662 least_loaded_cpu = i; 4663 } 4664 } 4665 } 4666 4667 return shallowest_idle_cpu != -1 ? shallowest_idle_cpu : least_loaded_cpu; 4668 } 4669 4670 /* 4671 * Try and locate an idle CPU in the sched_domain. 4672 */ 4673 static int select_idle_sibling(struct task_struct *p, int target) 4674 { 4675 struct sched_domain *sd; 4676 struct sched_group *sg; 4677 int i = task_cpu(p); 4678 4679 if (idle_cpu(target)) 4680 return target; 4681 4682 /* 4683 * If the prevous cpu is cache affine and idle, don't be stupid. 4684 */ 4685 if (i != target && cpus_share_cache(i, target) && idle_cpu(i)) 4686 return i; 4687 4688 /* 4689 * Otherwise, iterate the domains and find an elegible idle cpu. 4690 */ 4691 sd = rcu_dereference(per_cpu(sd_llc, target)); 4692 for_each_lower_domain(sd) { 4693 sg = sd->groups; 4694 do { 4695 if (!cpumask_intersects(sched_group_cpus(sg), 4696 tsk_cpus_allowed(p))) 4697 goto next; 4698 4699 for_each_cpu(i, sched_group_cpus(sg)) { 4700 if (i == target || !idle_cpu(i)) 4701 goto next; 4702 } 4703 4704 target = cpumask_first_and(sched_group_cpus(sg), 4705 tsk_cpus_allowed(p)); 4706 goto done; 4707 next: 4708 sg = sg->next; 4709 } while (sg != sd->groups); 4710 } 4711 done: 4712 return target; 4713 } 4714 4715 /* 4716 * select_task_rq_fair: Select target runqueue for the waking task in domains 4717 * that have the 'sd_flag' flag set. In practice, this is SD_BALANCE_WAKE, 4718 * SD_BALANCE_FORK, or SD_BALANCE_EXEC. 4719 * 4720 * Balances load by selecting the idlest cpu in the idlest group, or under 4721 * certain conditions an idle sibling cpu if the domain has SD_WAKE_AFFINE set. 4722 * 4723 * Returns the target cpu number. 4724 * 4725 * preempt must be disabled. 4726 */ 4727 static int 4728 select_task_rq_fair(struct task_struct *p, int prev_cpu, int sd_flag, int wake_flags) 4729 { 4730 struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL; 4731 int cpu = smp_processor_id(); 4732 int new_cpu = cpu; 4733 int want_affine = 0; 4734 int sync = wake_flags & WF_SYNC; 4735 4736 if (sd_flag & SD_BALANCE_WAKE) 4737 want_affine = cpumask_test_cpu(cpu, tsk_cpus_allowed(p)); 4738 4739 rcu_read_lock(); 4740 for_each_domain(cpu, tmp) { 4741 if (!(tmp->flags & SD_LOAD_BALANCE)) 4742 continue; 4743 4744 /* 4745 * If both cpu and prev_cpu are part of this domain, 4746 * cpu is a valid SD_WAKE_AFFINE target. 4747 */ 4748 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) && 4749 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) { 4750 affine_sd = tmp; 4751 break; 4752 } 4753 4754 if (tmp->flags & sd_flag) 4755 sd = tmp; 4756 } 4757 4758 if (affine_sd && cpu != prev_cpu && wake_affine(affine_sd, p, sync)) 4759 prev_cpu = cpu; 4760 4761 if (sd_flag & SD_BALANCE_WAKE) { 4762 new_cpu = select_idle_sibling(p, prev_cpu); 4763 goto unlock; 4764 } 4765 4766 while (sd) { 4767 struct sched_group *group; 4768 int weight; 4769 4770 if (!(sd->flags & sd_flag)) { 4771 sd = sd->child; 4772 continue; 4773 } 4774 4775 group = find_idlest_group(sd, p, cpu, sd_flag); 4776 if (!group) { 4777 sd = sd->child; 4778 continue; 4779 } 4780 4781 new_cpu = find_idlest_cpu(group, p, cpu); 4782 if (new_cpu == -1 || new_cpu == cpu) { 4783 /* Now try balancing at a lower domain level of cpu */ 4784 sd = sd->child; 4785 continue; 4786 } 4787 4788 /* Now try balancing at a lower domain level of new_cpu */ 4789 cpu = new_cpu; 4790 weight = sd->span_weight; 4791 sd = NULL; 4792 for_each_domain(cpu, tmp) { 4793 if (weight <= tmp->span_weight) 4794 break; 4795 if (tmp->flags & sd_flag) 4796 sd = tmp; 4797 } 4798 /* while loop will break here if sd == NULL */ 4799 } 4800 unlock: 4801 rcu_read_unlock(); 4802 4803 return new_cpu; 4804 } 4805 4806 /* 4807 * Called immediately before a task is migrated to a new cpu; task_cpu(p) and 4808 * cfs_rq_of(p) references at time of call are still valid and identify the 4809 * previous cpu. However, the caller only guarantees p->pi_lock is held; no 4810 * other assumptions, including the state of rq->lock, should be made. 4811 */ 4812 static void 4813 migrate_task_rq_fair(struct task_struct *p, int next_cpu) 4814 { 4815 struct sched_entity *se = &p->se; 4816 struct cfs_rq *cfs_rq = cfs_rq_of(se); 4817 4818 /* 4819 * Load tracking: accumulate removed load so that it can be processed 4820 * when we next update owning cfs_rq under rq->lock. Tasks contribute 4821 * to blocked load iff they have a positive decay-count. It can never 4822 * be negative here since on-rq tasks have decay-count == 0. 4823 */ 4824 if (se->avg.decay_count) { 4825 se->avg.decay_count = -__synchronize_entity_decay(se); 4826 atomic_long_add(se->avg.load_avg_contrib, 4827 &cfs_rq->removed_load); 4828 } 4829 4830 /* We have migrated, no longer consider this task hot */ 4831 se->exec_start = 0; 4832 } 4833 #endif /* CONFIG_SMP */ 4834 4835 static unsigned long 4836 wakeup_gran(struct sched_entity *curr, struct sched_entity *se) 4837 { 4838 unsigned long gran = sysctl_sched_wakeup_granularity; 4839 4840 /* 4841 * Since its curr running now, convert the gran from real-time 4842 * to virtual-time in his units. 4843 * 4844 * By using 'se' instead of 'curr' we penalize light tasks, so 4845 * they get preempted easier. That is, if 'se' < 'curr' then 4846 * the resulting gran will be larger, therefore penalizing the 4847 * lighter, if otoh 'se' > 'curr' then the resulting gran will 4848 * be smaller, again penalizing the lighter task. 4849 * 4850 * This is especially important for buddies when the leftmost 4851 * task is higher priority than the buddy. 4852 */ 4853 return calc_delta_fair(gran, se); 4854 } 4855 4856 /* 4857 * Should 'se' preempt 'curr'. 4858 * 4859 * |s1 4860 * |s2 4861 * |s3 4862 * g 4863 * |<--->|c 4864 * 4865 * w(c, s1) = -1 4866 * w(c, s2) = 0 4867 * w(c, s3) = 1 4868 * 4869 */ 4870 static int 4871 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se) 4872 { 4873 s64 gran, vdiff = curr->vruntime - se->vruntime; 4874 4875 if (vdiff <= 0) 4876 return -1; 4877 4878 gran = wakeup_gran(curr, se); 4879 if (vdiff > gran) 4880 return 1; 4881 4882 return 0; 4883 } 4884 4885 static void set_last_buddy(struct sched_entity *se) 4886 { 4887 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE)) 4888 return; 4889 4890 for_each_sched_entity(se) 4891 cfs_rq_of(se)->last = se; 4892 } 4893 4894 static void set_next_buddy(struct sched_entity *se) 4895 { 4896 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE)) 4897 return; 4898 4899 for_each_sched_entity(se) 4900 cfs_rq_of(se)->next = se; 4901 } 4902 4903 static void set_skip_buddy(struct sched_entity *se) 4904 { 4905 for_each_sched_entity(se) 4906 cfs_rq_of(se)->skip = se; 4907 } 4908 4909 /* 4910 * Preempt the current task with a newly woken task if needed: 4911 */ 4912 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags) 4913 { 4914 struct task_struct *curr = rq->curr; 4915 struct sched_entity *se = &curr->se, *pse = &p->se; 4916 struct cfs_rq *cfs_rq = task_cfs_rq(curr); 4917 int scale = cfs_rq->nr_running >= sched_nr_latency; 4918 int next_buddy_marked = 0; 4919 4920 if (unlikely(se == pse)) 4921 return; 4922 4923 /* 4924 * This is possible from callers such as attach_tasks(), in which we 4925 * unconditionally check_prempt_curr() after an enqueue (which may have 4926 * lead to a throttle). This both saves work and prevents false 4927 * next-buddy nomination below. 4928 */ 4929 if (unlikely(throttled_hierarchy(cfs_rq_of(pse)))) 4930 return; 4931 4932 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) { 4933 set_next_buddy(pse); 4934 next_buddy_marked = 1; 4935 } 4936 4937 /* 4938 * We can come here with TIF_NEED_RESCHED already set from new task 4939 * wake up path. 4940 * 4941 * Note: this also catches the edge-case of curr being in a throttled 4942 * group (e.g. via set_curr_task), since update_curr() (in the 4943 * enqueue of curr) will have resulted in resched being set. This 4944 * prevents us from potentially nominating it as a false LAST_BUDDY 4945 * below. 4946 */ 4947 if (test_tsk_need_resched(curr)) 4948 return; 4949 4950 /* Idle tasks are by definition preempted by non-idle tasks. */ 4951 if (unlikely(curr->policy == SCHED_IDLE) && 4952 likely(p->policy != SCHED_IDLE)) 4953 goto preempt; 4954 4955 /* 4956 * Batch and idle tasks do not preempt non-idle tasks (their preemption 4957 * is driven by the tick): 4958 */ 4959 if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION)) 4960 return; 4961 4962 find_matching_se(&se, &pse); 4963 update_curr(cfs_rq_of(se)); 4964 BUG_ON(!pse); 4965 if (wakeup_preempt_entity(se, pse) == 1) { 4966 /* 4967 * Bias pick_next to pick the sched entity that is 4968 * triggering this preemption. 4969 */ 4970 if (!next_buddy_marked) 4971 set_next_buddy(pse); 4972 goto preempt; 4973 } 4974 4975 return; 4976 4977 preempt: 4978 resched_curr(rq); 4979 /* 4980 * Only set the backward buddy when the current task is still 4981 * on the rq. This can happen when a wakeup gets interleaved 4982 * with schedule on the ->pre_schedule() or idle_balance() 4983 * point, either of which can * drop the rq lock. 4984 * 4985 * Also, during early boot the idle thread is in the fair class, 4986 * for obvious reasons its a bad idea to schedule back to it. 4987 */ 4988 if (unlikely(!se->on_rq || curr == rq->idle)) 4989 return; 4990 4991 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se)) 4992 set_last_buddy(se); 4993 } 4994 4995 static struct task_struct * 4996 pick_next_task_fair(struct rq *rq, struct task_struct *prev) 4997 { 4998 struct cfs_rq *cfs_rq = &rq->cfs; 4999 struct sched_entity *se; 5000 struct task_struct *p; 5001 int new_tasks; 5002 5003 again: 5004 #ifdef CONFIG_FAIR_GROUP_SCHED 5005 if (!cfs_rq->nr_running) 5006 goto idle; 5007 5008 if (prev->sched_class != &fair_sched_class) 5009 goto simple; 5010 5011 /* 5012 * Because of the set_next_buddy() in dequeue_task_fair() it is rather 5013 * likely that a next task is from the same cgroup as the current. 5014 * 5015 * Therefore attempt to avoid putting and setting the entire cgroup 5016 * hierarchy, only change the part that actually changes. 5017 */ 5018 5019 do { 5020 struct sched_entity *curr = cfs_rq->curr; 5021 5022 /* 5023 * Since we got here without doing put_prev_entity() we also 5024 * have to consider cfs_rq->curr. If it is still a runnable 5025 * entity, update_curr() will update its vruntime, otherwise 5026 * forget we've ever seen it. 5027 */ 5028 if (curr && curr->on_rq) 5029 update_curr(cfs_rq); 5030 else 5031 curr = NULL; 5032 5033 /* 5034 * This call to check_cfs_rq_runtime() will do the throttle and 5035 * dequeue its entity in the parent(s). Therefore the 'simple' 5036 * nr_running test will indeed be correct. 5037 */ 5038 if (unlikely(check_cfs_rq_runtime(cfs_rq))) 5039 goto simple; 5040 5041 se = pick_next_entity(cfs_rq, curr); 5042 cfs_rq = group_cfs_rq(se); 5043 } while (cfs_rq); 5044 5045 p = task_of(se); 5046 5047 /* 5048 * Since we haven't yet done put_prev_entity and if the selected task 5049 * is a different task than we started out with, try and touch the 5050 * least amount of cfs_rqs. 5051 */ 5052 if (prev != p) { 5053 struct sched_entity *pse = &prev->se; 5054 5055 while (!(cfs_rq = is_same_group(se, pse))) { 5056 int se_depth = se->depth; 5057 int pse_depth = pse->depth; 5058 5059 if (se_depth <= pse_depth) { 5060 put_prev_entity(cfs_rq_of(pse), pse); 5061 pse = parent_entity(pse); 5062 } 5063 if (se_depth >= pse_depth) { 5064 set_next_entity(cfs_rq_of(se), se); 5065 se = parent_entity(se); 5066 } 5067 } 5068 5069 put_prev_entity(cfs_rq, pse); 5070 set_next_entity(cfs_rq, se); 5071 } 5072 5073 if (hrtick_enabled(rq)) 5074 hrtick_start_fair(rq, p); 5075 5076 return p; 5077 simple: 5078 cfs_rq = &rq->cfs; 5079 #endif 5080 5081 if (!cfs_rq->nr_running) 5082 goto idle; 5083 5084 put_prev_task(rq, prev); 5085 5086 do { 5087 se = pick_next_entity(cfs_rq, NULL); 5088 set_next_entity(cfs_rq, se); 5089 cfs_rq = group_cfs_rq(se); 5090 } while (cfs_rq); 5091 5092 p = task_of(se); 5093 5094 if (hrtick_enabled(rq)) 5095 hrtick_start_fair(rq, p); 5096 5097 return p; 5098 5099 idle: 5100 new_tasks = idle_balance(rq); 5101 /* 5102 * Because idle_balance() releases (and re-acquires) rq->lock, it is 5103 * possible for any higher priority task to appear. In that case we 5104 * must re-start the pick_next_entity() loop. 5105 */ 5106 if (new_tasks < 0) 5107 return RETRY_TASK; 5108 5109 if (new_tasks > 0) 5110 goto again; 5111 5112 return NULL; 5113 } 5114 5115 /* 5116 * Account for a descheduled task: 5117 */ 5118 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev) 5119 { 5120 struct sched_entity *se = &prev->se; 5121 struct cfs_rq *cfs_rq; 5122 5123 for_each_sched_entity(se) { 5124 cfs_rq = cfs_rq_of(se); 5125 put_prev_entity(cfs_rq, se); 5126 } 5127 } 5128 5129 /* 5130 * sched_yield() is very simple 5131 * 5132 * The magic of dealing with the ->skip buddy is in pick_next_entity. 5133 */ 5134 static void yield_task_fair(struct rq *rq) 5135 { 5136 struct task_struct *curr = rq->curr; 5137 struct cfs_rq *cfs_rq = task_cfs_rq(curr); 5138 struct sched_entity *se = &curr->se; 5139 5140 /* 5141 * Are we the only task in the tree? 5142 */ 5143 if (unlikely(rq->nr_running == 1)) 5144 return; 5145 5146 clear_buddies(cfs_rq, se); 5147 5148 if (curr->policy != SCHED_BATCH) { 5149 update_rq_clock(rq); 5150 /* 5151 * Update run-time statistics of the 'current'. 5152 */ 5153 update_curr(cfs_rq); 5154 /* 5155 * Tell update_rq_clock() that we've just updated, 5156 * so we don't do microscopic update in schedule() 5157 * and double the fastpath cost. 5158 */ 5159 rq_clock_skip_update(rq, true); 5160 } 5161 5162 set_skip_buddy(se); 5163 } 5164 5165 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt) 5166 { 5167 struct sched_entity *se = &p->se; 5168 5169 /* throttled hierarchies are not runnable */ 5170 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se))) 5171 return false; 5172 5173 /* Tell the scheduler that we'd really like pse to run next. */ 5174 set_next_buddy(se); 5175 5176 yield_task_fair(rq); 5177 5178 return true; 5179 } 5180 5181 #ifdef CONFIG_SMP 5182 /************************************************** 5183 * Fair scheduling class load-balancing methods. 5184 * 5185 * BASICS 5186 * 5187 * The purpose of load-balancing is to achieve the same basic fairness the 5188 * per-cpu scheduler provides, namely provide a proportional amount of compute 5189 * time to each task. This is expressed in the following equation: 5190 * 5191 * W_i,n/P_i == W_j,n/P_j for all i,j (1) 5192 * 5193 * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight 5194 * W_i,0 is defined as: 5195 * 5196 * W_i,0 = \Sum_j w_i,j (2) 5197 * 5198 * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight 5199 * is derived from the nice value as per prio_to_weight[]. 5200 * 5201 * The weight average is an exponential decay average of the instantaneous 5202 * weight: 5203 * 5204 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3) 5205 * 5206 * C_i is the compute capacity of cpu i, typically it is the 5207 * fraction of 'recent' time available for SCHED_OTHER task execution. But it 5208 * can also include other factors [XXX]. 5209 * 5210 * To achieve this balance we define a measure of imbalance which follows 5211 * directly from (1): 5212 * 5213 * imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j } (4) 5214 * 5215 * We them move tasks around to minimize the imbalance. In the continuous 5216 * function space it is obvious this converges, in the discrete case we get 5217 * a few fun cases generally called infeasible weight scenarios. 5218 * 5219 * [XXX expand on: 5220 * - infeasible weights; 5221 * - local vs global optima in the discrete case. ] 5222 * 5223 * 5224 * SCHED DOMAINS 5225 * 5226 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2) 5227 * for all i,j solution, we create a tree of cpus that follows the hardware 5228 * topology where each level pairs two lower groups (or better). This results 5229 * in O(log n) layers. Furthermore we reduce the number of cpus going up the 5230 * tree to only the first of the previous level and we decrease the frequency 5231 * of load-balance at each level inv. proportional to the number of cpus in 5232 * the groups. 5233 * 5234 * This yields: 5235 * 5236 * log_2 n 1 n 5237 * \Sum { --- * --- * 2^i } = O(n) (5) 5238 * i = 0 2^i 2^i 5239 * `- size of each group 5240 * | | `- number of cpus doing load-balance 5241 * | `- freq 5242 * `- sum over all levels 5243 * 5244 * Coupled with a limit on how many tasks we can migrate every balance pass, 5245 * this makes (5) the runtime complexity of the balancer. 5246 * 5247 * An important property here is that each CPU is still (indirectly) connected 5248 * to every other cpu in at most O(log n) steps: 5249 * 5250 * The adjacency matrix of the resulting graph is given by: 5251 * 5252 * log_2 n 5253 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6) 5254 * k = 0 5255 * 5256 * And you'll find that: 5257 * 5258 * A^(log_2 n)_i,j != 0 for all i,j (7) 5259 * 5260 * Showing there's indeed a path between every cpu in at most O(log n) steps. 5261 * The task movement gives a factor of O(m), giving a convergence complexity 5262 * of: 5263 * 5264 * O(nm log n), n := nr_cpus, m := nr_tasks (8) 5265 * 5266 * 5267 * WORK CONSERVING 5268 * 5269 * In order to avoid CPUs going idle while there's still work to do, new idle 5270 * balancing is more aggressive and has the newly idle cpu iterate up the domain 5271 * tree itself instead of relying on other CPUs to bring it work. 5272 * 5273 * This adds some complexity to both (5) and (8) but it reduces the total idle 5274 * time. 5275 * 5276 * [XXX more?] 5277 * 5278 * 5279 * CGROUPS 5280 * 5281 * Cgroups make a horror show out of (2), instead of a simple sum we get: 5282 * 5283 * s_k,i 5284 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9) 5285 * S_k 5286 * 5287 * Where 5288 * 5289 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10) 5290 * 5291 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i. 5292 * 5293 * The big problem is S_k, its a global sum needed to compute a local (W_i) 5294 * property. 5295 * 5296 * [XXX write more on how we solve this.. _after_ merging pjt's patches that 5297 * rewrite all of this once again.] 5298 */ 5299 5300 static unsigned long __read_mostly max_load_balance_interval = HZ/10; 5301 5302 enum fbq_type { regular, remote, all }; 5303 5304 #define LBF_ALL_PINNED 0x01 5305 #define LBF_NEED_BREAK 0x02 5306 #define LBF_DST_PINNED 0x04 5307 #define LBF_SOME_PINNED 0x08 5308 5309 struct lb_env { 5310 struct sched_domain *sd; 5311 5312 struct rq *src_rq; 5313 int src_cpu; 5314 5315 int dst_cpu; 5316 struct rq *dst_rq; 5317 5318 struct cpumask *dst_grpmask; 5319 int new_dst_cpu; 5320 enum cpu_idle_type idle; 5321 long imbalance; 5322 /* The set of CPUs under consideration for load-balancing */ 5323 struct cpumask *cpus; 5324 5325 unsigned int flags; 5326 5327 unsigned int loop; 5328 unsigned int loop_break; 5329 unsigned int loop_max; 5330 5331 enum fbq_type fbq_type; 5332 struct list_head tasks; 5333 }; 5334 5335 /* 5336 * Is this task likely cache-hot: 5337 */ 5338 static int task_hot(struct task_struct *p, struct lb_env *env) 5339 { 5340 s64 delta; 5341 5342 lockdep_assert_held(&env->src_rq->lock); 5343 5344 if (p->sched_class != &fair_sched_class) 5345 return 0; 5346 5347 if (unlikely(p->policy == SCHED_IDLE)) 5348 return 0; 5349 5350 /* 5351 * Buddy candidates are cache hot: 5352 */ 5353 if (sched_feat(CACHE_HOT_BUDDY) && env->dst_rq->nr_running && 5354 (&p->se == cfs_rq_of(&p->se)->next || 5355 &p->se == cfs_rq_of(&p->se)->last)) 5356 return 1; 5357 5358 if (sysctl_sched_migration_cost == -1) 5359 return 1; 5360 if (sysctl_sched_migration_cost == 0) 5361 return 0; 5362 5363 delta = rq_clock_task(env->src_rq) - p->se.exec_start; 5364 5365 return delta < (s64)sysctl_sched_migration_cost; 5366 } 5367 5368 #ifdef CONFIG_NUMA_BALANCING 5369 /* Returns true if the destination node has incurred more faults */ 5370 static bool migrate_improves_locality(struct task_struct *p, struct lb_env *env) 5371 { 5372 struct numa_group *numa_group = rcu_dereference(p->numa_group); 5373 int src_nid, dst_nid; 5374 5375 if (!sched_feat(NUMA_FAVOUR_HIGHER) || !p->numa_faults || 5376 !(env->sd->flags & SD_NUMA)) { 5377 return false; 5378 } 5379 5380 src_nid = cpu_to_node(env->src_cpu); 5381 dst_nid = cpu_to_node(env->dst_cpu); 5382 5383 if (src_nid == dst_nid) 5384 return false; 5385 5386 if (numa_group) { 5387 /* Task is already in the group's interleave set. */ 5388 if (node_isset(src_nid, numa_group->active_nodes)) 5389 return false; 5390 5391 /* Task is moving into the group's interleave set. */ 5392 if (node_isset(dst_nid, numa_group->active_nodes)) 5393 return true; 5394 5395 return group_faults(p, dst_nid) > group_faults(p, src_nid); 5396 } 5397 5398 /* Encourage migration to the preferred node. */ 5399 if (dst_nid == p->numa_preferred_nid) 5400 return true; 5401 5402 return task_faults(p, dst_nid) > task_faults(p, src_nid); 5403 } 5404 5405 5406 static bool migrate_degrades_locality(struct task_struct *p, struct lb_env *env) 5407 { 5408 struct numa_group *numa_group = rcu_dereference(p->numa_group); 5409 int src_nid, dst_nid; 5410 5411 if (!sched_feat(NUMA) || !sched_feat(NUMA_RESIST_LOWER)) 5412 return false; 5413 5414 if (!p->numa_faults || !(env->sd->flags & SD_NUMA)) 5415 return false; 5416 5417 src_nid = cpu_to_node(env->src_cpu); 5418 dst_nid = cpu_to_node(env->dst_cpu); 5419 5420 if (src_nid == dst_nid) 5421 return false; 5422 5423 if (numa_group) { 5424 /* Task is moving within/into the group's interleave set. */ 5425 if (node_isset(dst_nid, numa_group->active_nodes)) 5426 return false; 5427 5428 /* Task is moving out of the group's interleave set. */ 5429 if (node_isset(src_nid, numa_group->active_nodes)) 5430 return true; 5431 5432 return group_faults(p, dst_nid) < group_faults(p, src_nid); 5433 } 5434 5435 /* Migrating away from the preferred node is always bad. */ 5436 if (src_nid == p->numa_preferred_nid) 5437 return true; 5438 5439 return task_faults(p, dst_nid) < task_faults(p, src_nid); 5440 } 5441 5442 #else 5443 static inline bool migrate_improves_locality(struct task_struct *p, 5444 struct lb_env *env) 5445 { 5446 return false; 5447 } 5448 5449 static inline bool migrate_degrades_locality(struct task_struct *p, 5450 struct lb_env *env) 5451 { 5452 return false; 5453 } 5454 #endif 5455 5456 /* 5457 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu? 5458 */ 5459 static 5460 int can_migrate_task(struct task_struct *p, struct lb_env *env) 5461 { 5462 int tsk_cache_hot = 0; 5463 5464 lockdep_assert_held(&env->src_rq->lock); 5465 5466 /* 5467 * We do not migrate tasks that are: 5468 * 1) throttled_lb_pair, or 5469 * 2) cannot be migrated to this CPU due to cpus_allowed, or 5470 * 3) running (obviously), or 5471 * 4) are cache-hot on their current CPU. 5472 */ 5473 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu)) 5474 return 0; 5475 5476 if (!cpumask_test_cpu(env->dst_cpu, tsk_cpus_allowed(p))) { 5477 int cpu; 5478 5479 schedstat_inc(p, se.statistics.nr_failed_migrations_affine); 5480 5481 env->flags |= LBF_SOME_PINNED; 5482 5483 /* 5484 * Remember if this task can be migrated to any other cpu in 5485 * our sched_group. We may want to revisit it if we couldn't 5486 * meet load balance goals by pulling other tasks on src_cpu. 5487 * 5488 * Also avoid computing new_dst_cpu if we have already computed 5489 * one in current iteration. 5490 */ 5491 if (!env->dst_grpmask || (env->flags & LBF_DST_PINNED)) 5492 return 0; 5493 5494 /* Prevent to re-select dst_cpu via env's cpus */ 5495 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) { 5496 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) { 5497 env->flags |= LBF_DST_PINNED; 5498 env->new_dst_cpu = cpu; 5499 break; 5500 } 5501 } 5502 5503 return 0; 5504 } 5505 5506 /* Record that we found atleast one task that could run on dst_cpu */ 5507 env->flags &= ~LBF_ALL_PINNED; 5508 5509 if (task_running(env->src_rq, p)) { 5510 schedstat_inc(p, se.statistics.nr_failed_migrations_running); 5511 return 0; 5512 } 5513 5514 /* 5515 * Aggressive migration if: 5516 * 1) destination numa is preferred 5517 * 2) task is cache cold, or 5518 * 3) too many balance attempts have failed. 5519 */ 5520 tsk_cache_hot = task_hot(p, env); 5521 if (!tsk_cache_hot) 5522 tsk_cache_hot = migrate_degrades_locality(p, env); 5523 5524 if (migrate_improves_locality(p, env) || !tsk_cache_hot || 5525 env->sd->nr_balance_failed > env->sd->cache_nice_tries) { 5526 if (tsk_cache_hot) { 5527 schedstat_inc(env->sd, lb_hot_gained[env->idle]); 5528 schedstat_inc(p, se.statistics.nr_forced_migrations); 5529 } 5530 return 1; 5531 } 5532 5533 schedstat_inc(p, se.statistics.nr_failed_migrations_hot); 5534 return 0; 5535 } 5536 5537 /* 5538 * detach_task() -- detach the task for the migration specified in env 5539 */ 5540 static void detach_task(struct task_struct *p, struct lb_env *env) 5541 { 5542 lockdep_assert_held(&env->src_rq->lock); 5543 5544 deactivate_task(env->src_rq, p, 0); 5545 p->on_rq = TASK_ON_RQ_MIGRATING; 5546 set_task_cpu(p, env->dst_cpu); 5547 } 5548 5549 /* 5550 * detach_one_task() -- tries to dequeue exactly one task from env->src_rq, as 5551 * part of active balancing operations within "domain". 5552 * 5553 * Returns a task if successful and NULL otherwise. 5554 */ 5555 static struct task_struct *detach_one_task(struct lb_env *env) 5556 { 5557 struct task_struct *p, *n; 5558 5559 lockdep_assert_held(&env->src_rq->lock); 5560 5561 list_for_each_entry_safe(p, n, &env->src_rq->cfs_tasks, se.group_node) { 5562 if (!can_migrate_task(p, env)) 5563 continue; 5564 5565 detach_task(p, env); 5566 5567 /* 5568 * Right now, this is only the second place where 5569 * lb_gained[env->idle] is updated (other is detach_tasks) 5570 * so we can safely collect stats here rather than 5571 * inside detach_tasks(). 5572 */ 5573 schedstat_inc(env->sd, lb_gained[env->idle]); 5574 return p; 5575 } 5576 return NULL; 5577 } 5578 5579 static const unsigned int sched_nr_migrate_break = 32; 5580 5581 /* 5582 * detach_tasks() -- tries to detach up to imbalance weighted load from 5583 * busiest_rq, as part of a balancing operation within domain "sd". 5584 * 5585 * Returns number of detached tasks if successful and 0 otherwise. 5586 */ 5587 static int detach_tasks(struct lb_env *env) 5588 { 5589 struct list_head *tasks = &env->src_rq->cfs_tasks; 5590 struct task_struct *p; 5591 unsigned long load; 5592 int detached = 0; 5593 5594 lockdep_assert_held(&env->src_rq->lock); 5595 5596 if (env->imbalance <= 0) 5597 return 0; 5598 5599 while (!list_empty(tasks)) { 5600 p = list_first_entry(tasks, struct task_struct, se.group_node); 5601 5602 env->loop++; 5603 /* We've more or less seen every task there is, call it quits */ 5604 if (env->loop > env->loop_max) 5605 break; 5606 5607 /* take a breather every nr_migrate tasks */ 5608 if (env->loop > env->loop_break) { 5609 env->loop_break += sched_nr_migrate_break; 5610 env->flags |= LBF_NEED_BREAK; 5611 break; 5612 } 5613 5614 if (!can_migrate_task(p, env)) 5615 goto next; 5616 5617 load = task_h_load(p); 5618 5619 if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed) 5620 goto next; 5621 5622 if ((load / 2) > env->imbalance) 5623 goto next; 5624 5625 detach_task(p, env); 5626 list_add(&p->se.group_node, &env->tasks); 5627 5628 detached++; 5629 env->imbalance -= load; 5630 5631 #ifdef CONFIG_PREEMPT 5632 /* 5633 * NEWIDLE balancing is a source of latency, so preemptible 5634 * kernels will stop after the first task is detached to minimize 5635 * the critical section. 5636 */ 5637 if (env->idle == CPU_NEWLY_IDLE) 5638 break; 5639 #endif 5640 5641 /* 5642 * We only want to steal up to the prescribed amount of 5643 * weighted load. 5644 */ 5645 if (env->imbalance <= 0) 5646 break; 5647 5648 continue; 5649 next: 5650 list_move_tail(&p->se.group_node, tasks); 5651 } 5652 5653 /* 5654 * Right now, this is one of only two places we collect this stat 5655 * so we can safely collect detach_one_task() stats here rather 5656 * than inside detach_one_task(). 5657 */ 5658 schedstat_add(env->sd, lb_gained[env->idle], detached); 5659 5660 return detached; 5661 } 5662 5663 /* 5664 * attach_task() -- attach the task detached by detach_task() to its new rq. 5665 */ 5666 static void attach_task(struct rq *rq, struct task_struct *p) 5667 { 5668 lockdep_assert_held(&rq->lock); 5669 5670 BUG_ON(task_rq(p) != rq); 5671 p->on_rq = TASK_ON_RQ_QUEUED; 5672 activate_task(rq, p, 0); 5673 check_preempt_curr(rq, p, 0); 5674 } 5675 5676 /* 5677 * attach_one_task() -- attaches the task returned from detach_one_task() to 5678 * its new rq. 5679 */ 5680 static void attach_one_task(struct rq *rq, struct task_struct *p) 5681 { 5682 raw_spin_lock(&rq->lock); 5683 attach_task(rq, p); 5684 raw_spin_unlock(&rq->lock); 5685 } 5686 5687 /* 5688 * attach_tasks() -- attaches all tasks detached by detach_tasks() to their 5689 * new rq. 5690 */ 5691 static void attach_tasks(struct lb_env *env) 5692 { 5693 struct list_head *tasks = &env->tasks; 5694 struct task_struct *p; 5695 5696 raw_spin_lock(&env->dst_rq->lock); 5697 5698 while (!list_empty(tasks)) { 5699 p = list_first_entry(tasks, struct task_struct, se.group_node); 5700 list_del_init(&p->se.group_node); 5701 5702 attach_task(env->dst_rq, p); 5703 } 5704 5705 raw_spin_unlock(&env->dst_rq->lock); 5706 } 5707 5708 #ifdef CONFIG_FAIR_GROUP_SCHED 5709 /* 5710 * update tg->load_weight by folding this cpu's load_avg 5711 */ 5712 static void __update_blocked_averages_cpu(struct task_group *tg, int cpu) 5713 { 5714 struct sched_entity *se = tg->se[cpu]; 5715 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu]; 5716 5717 /* throttled entities do not contribute to load */ 5718 if (throttled_hierarchy(cfs_rq)) 5719 return; 5720 5721 update_cfs_rq_blocked_load(cfs_rq, 1); 5722 5723 if (se) { 5724 update_entity_load_avg(se, 1); 5725 /* 5726 * We pivot on our runnable average having decayed to zero for 5727 * list removal. This generally implies that all our children 5728 * have also been removed (modulo rounding error or bandwidth 5729 * control); however, such cases are rare and we can fix these 5730 * at enqueue. 5731 * 5732 * TODO: fix up out-of-order children on enqueue. 5733 */ 5734 if (!se->avg.runnable_avg_sum && !cfs_rq->nr_running) 5735 list_del_leaf_cfs_rq(cfs_rq); 5736 } else { 5737 struct rq *rq = rq_of(cfs_rq); 5738 update_rq_runnable_avg(rq, rq->nr_running); 5739 } 5740 } 5741 5742 static void update_blocked_averages(int cpu) 5743 { 5744 struct rq *rq = cpu_rq(cpu); 5745 struct cfs_rq *cfs_rq; 5746 unsigned long flags; 5747 5748 raw_spin_lock_irqsave(&rq->lock, flags); 5749 update_rq_clock(rq); 5750 /* 5751 * Iterates the task_group tree in a bottom up fashion, see 5752 * list_add_leaf_cfs_rq() for details. 5753 */ 5754 for_each_leaf_cfs_rq(rq, cfs_rq) { 5755 /* 5756 * Note: We may want to consider periodically releasing 5757 * rq->lock about these updates so that creating many task 5758 * groups does not result in continually extending hold time. 5759 */ 5760 __update_blocked_averages_cpu(cfs_rq->tg, rq->cpu); 5761 } 5762 5763 raw_spin_unlock_irqrestore(&rq->lock, flags); 5764 } 5765 5766 /* 5767 * Compute the hierarchical load factor for cfs_rq and all its ascendants. 5768 * This needs to be done in a top-down fashion because the load of a child 5769 * group is a fraction of its parents load. 5770 */ 5771 static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq) 5772 { 5773 struct rq *rq = rq_of(cfs_rq); 5774 struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)]; 5775 unsigned long now = jiffies; 5776 unsigned long load; 5777 5778 if (cfs_rq->last_h_load_update == now) 5779 return; 5780 5781 cfs_rq->h_load_next = NULL; 5782 for_each_sched_entity(se) { 5783 cfs_rq = cfs_rq_of(se); 5784 cfs_rq->h_load_next = se; 5785 if (cfs_rq->last_h_load_update == now) 5786 break; 5787 } 5788 5789 if (!se) { 5790 cfs_rq->h_load = cfs_rq->runnable_load_avg; 5791 cfs_rq->last_h_load_update = now; 5792 } 5793 5794 while ((se = cfs_rq->h_load_next) != NULL) { 5795 load = cfs_rq->h_load; 5796 load = div64_ul(load * se->avg.load_avg_contrib, 5797 cfs_rq->runnable_load_avg + 1); 5798 cfs_rq = group_cfs_rq(se); 5799 cfs_rq->h_load = load; 5800 cfs_rq->last_h_load_update = now; 5801 } 5802 } 5803 5804 static unsigned long task_h_load(struct task_struct *p) 5805 { 5806 struct cfs_rq *cfs_rq = task_cfs_rq(p); 5807 5808 update_cfs_rq_h_load(cfs_rq); 5809 return div64_ul(p->se.avg.load_avg_contrib * cfs_rq->h_load, 5810 cfs_rq->runnable_load_avg + 1); 5811 } 5812 #else 5813 static inline void update_blocked_averages(int cpu) 5814 { 5815 } 5816 5817 static unsigned long task_h_load(struct task_struct *p) 5818 { 5819 return p->se.avg.load_avg_contrib; 5820 } 5821 #endif 5822 5823 /********** Helpers for find_busiest_group ************************/ 5824 5825 enum group_type { 5826 group_other = 0, 5827 group_imbalanced, 5828 group_overloaded, 5829 }; 5830 5831 /* 5832 * sg_lb_stats - stats of a sched_group required for load_balancing 5833 */ 5834 struct sg_lb_stats { 5835 unsigned long avg_load; /*Avg load across the CPUs of the group */ 5836 unsigned long group_load; /* Total load over the CPUs of the group */ 5837 unsigned long sum_weighted_load; /* Weighted load of group's tasks */ 5838 unsigned long load_per_task; 5839 unsigned long group_capacity; 5840 unsigned int sum_nr_running; /* Nr tasks running in the group */ 5841 unsigned int group_capacity_factor; 5842 unsigned int idle_cpus; 5843 unsigned int group_weight; 5844 enum group_type group_type; 5845 int group_has_free_capacity; 5846 #ifdef CONFIG_NUMA_BALANCING 5847 unsigned int nr_numa_running; 5848 unsigned int nr_preferred_running; 5849 #endif 5850 }; 5851 5852 /* 5853 * sd_lb_stats - Structure to store the statistics of a sched_domain 5854 * during load balancing. 5855 */ 5856 struct sd_lb_stats { 5857 struct sched_group *busiest; /* Busiest group in this sd */ 5858 struct sched_group *local; /* Local group in this sd */ 5859 unsigned long total_load; /* Total load of all groups in sd */ 5860 unsigned long total_capacity; /* Total capacity of all groups in sd */ 5861 unsigned long avg_load; /* Average load across all groups in sd */ 5862 5863 struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */ 5864 struct sg_lb_stats local_stat; /* Statistics of the local group */ 5865 }; 5866 5867 static inline void init_sd_lb_stats(struct sd_lb_stats *sds) 5868 { 5869 /* 5870 * Skimp on the clearing to avoid duplicate work. We can avoid clearing 5871 * local_stat because update_sg_lb_stats() does a full clear/assignment. 5872 * We must however clear busiest_stat::avg_load because 5873 * update_sd_pick_busiest() reads this before assignment. 5874 */ 5875 *sds = (struct sd_lb_stats){ 5876 .busiest = NULL, 5877 .local = NULL, 5878 .total_load = 0UL, 5879 .total_capacity = 0UL, 5880 .busiest_stat = { 5881 .avg_load = 0UL, 5882 .sum_nr_running = 0, 5883 .group_type = group_other, 5884 }, 5885 }; 5886 } 5887 5888 /** 5889 * get_sd_load_idx - Obtain the load index for a given sched domain. 5890 * @sd: The sched_domain whose load_idx is to be obtained. 5891 * @idle: The idle status of the CPU for whose sd load_idx is obtained. 5892 * 5893 * Return: The load index. 5894 */ 5895 static inline int get_sd_load_idx(struct sched_domain *sd, 5896 enum cpu_idle_type idle) 5897 { 5898 int load_idx; 5899 5900 switch (idle) { 5901 case CPU_NOT_IDLE: 5902 load_idx = sd->busy_idx; 5903 break; 5904 5905 case CPU_NEWLY_IDLE: 5906 load_idx = sd->newidle_idx; 5907 break; 5908 default: 5909 load_idx = sd->idle_idx; 5910 break; 5911 } 5912 5913 return load_idx; 5914 } 5915 5916 static unsigned long default_scale_capacity(struct sched_domain *sd, int cpu) 5917 { 5918 return SCHED_CAPACITY_SCALE; 5919 } 5920 5921 unsigned long __weak arch_scale_freq_capacity(struct sched_domain *sd, int cpu) 5922 { 5923 return default_scale_capacity(sd, cpu); 5924 } 5925 5926 static unsigned long default_scale_cpu_capacity(struct sched_domain *sd, int cpu) 5927 { 5928 if ((sd->flags & SD_SHARE_CPUCAPACITY) && (sd->span_weight > 1)) 5929 return sd->smt_gain / sd->span_weight; 5930 5931 return SCHED_CAPACITY_SCALE; 5932 } 5933 5934 unsigned long __weak arch_scale_cpu_capacity(struct sched_domain *sd, int cpu) 5935 { 5936 return default_scale_cpu_capacity(sd, cpu); 5937 } 5938 5939 static unsigned long scale_rt_capacity(int cpu) 5940 { 5941 struct rq *rq = cpu_rq(cpu); 5942 u64 total, available, age_stamp, avg; 5943 s64 delta; 5944 5945 /* 5946 * Since we're reading these variables without serialization make sure 5947 * we read them once before doing sanity checks on them. 5948 */ 5949 age_stamp = ACCESS_ONCE(rq->age_stamp); 5950 avg = ACCESS_ONCE(rq->rt_avg); 5951 delta = __rq_clock_broken(rq) - age_stamp; 5952 5953 if (unlikely(delta < 0)) 5954 delta = 0; 5955 5956 total = sched_avg_period() + delta; 5957 5958 if (unlikely(total < avg)) { 5959 /* Ensures that capacity won't end up being negative */ 5960 available = 0; 5961 } else { 5962 available = total - avg; 5963 } 5964 5965 if (unlikely((s64)total < SCHED_CAPACITY_SCALE)) 5966 total = SCHED_CAPACITY_SCALE; 5967 5968 total >>= SCHED_CAPACITY_SHIFT; 5969 5970 return div_u64(available, total); 5971 } 5972 5973 static void update_cpu_capacity(struct sched_domain *sd, int cpu) 5974 { 5975 unsigned long capacity = SCHED_CAPACITY_SCALE; 5976 struct sched_group *sdg = sd->groups; 5977 5978 if (sched_feat(ARCH_CAPACITY)) 5979 capacity *= arch_scale_cpu_capacity(sd, cpu); 5980 else 5981 capacity *= default_scale_cpu_capacity(sd, cpu); 5982 5983 capacity >>= SCHED_CAPACITY_SHIFT; 5984 5985 sdg->sgc->capacity_orig = capacity; 5986 5987 if (sched_feat(ARCH_CAPACITY)) 5988 capacity *= arch_scale_freq_capacity(sd, cpu); 5989 else 5990 capacity *= default_scale_capacity(sd, cpu); 5991 5992 capacity >>= SCHED_CAPACITY_SHIFT; 5993 5994 capacity *= scale_rt_capacity(cpu); 5995 capacity >>= SCHED_CAPACITY_SHIFT; 5996 5997 if (!capacity) 5998 capacity = 1; 5999 6000 cpu_rq(cpu)->cpu_capacity = capacity; 6001 sdg->sgc->capacity = capacity; 6002 } 6003 6004 void update_group_capacity(struct sched_domain *sd, int cpu) 6005 { 6006 struct sched_domain *child = sd->child; 6007 struct sched_group *group, *sdg = sd->groups; 6008 unsigned long capacity, capacity_orig; 6009 unsigned long interval; 6010 6011 interval = msecs_to_jiffies(sd->balance_interval); 6012 interval = clamp(interval, 1UL, max_load_balance_interval); 6013 sdg->sgc->next_update = jiffies + interval; 6014 6015 if (!child) { 6016 update_cpu_capacity(sd, cpu); 6017 return; 6018 } 6019 6020 capacity_orig = capacity = 0; 6021 6022 if (child->flags & SD_OVERLAP) { 6023 /* 6024 * SD_OVERLAP domains cannot assume that child groups 6025 * span the current group. 6026 */ 6027 6028 for_each_cpu(cpu, sched_group_cpus(sdg)) { 6029 struct sched_group_capacity *sgc; 6030 struct rq *rq = cpu_rq(cpu); 6031 6032 /* 6033 * build_sched_domains() -> init_sched_groups_capacity() 6034 * gets here before we've attached the domains to the 6035 * runqueues. 6036 * 6037 * Use capacity_of(), which is set irrespective of domains 6038 * in update_cpu_capacity(). 6039 * 6040 * This avoids capacity/capacity_orig from being 0 and 6041 * causing divide-by-zero issues on boot. 6042 * 6043 * Runtime updates will correct capacity_orig. 6044 */ 6045 if (unlikely(!rq->sd)) { 6046 capacity_orig += capacity_of(cpu); 6047 capacity += capacity_of(cpu); 6048 continue; 6049 } 6050 6051 sgc = rq->sd->groups->sgc; 6052 capacity_orig += sgc->capacity_orig; 6053 capacity += sgc->capacity; 6054 } 6055 } else { 6056 /* 6057 * !SD_OVERLAP domains can assume that child groups 6058 * span the current group. 6059 */ 6060 6061 group = child->groups; 6062 do { 6063 capacity_orig += group->sgc->capacity_orig; 6064 capacity += group->sgc->capacity; 6065 group = group->next; 6066 } while (group != child->groups); 6067 } 6068 6069 sdg->sgc->capacity_orig = capacity_orig; 6070 sdg->sgc->capacity = capacity; 6071 } 6072 6073 /* 6074 * Try and fix up capacity for tiny siblings, this is needed when 6075 * things like SD_ASYM_PACKING need f_b_g to select another sibling 6076 * which on its own isn't powerful enough. 6077 * 6078 * See update_sd_pick_busiest() and check_asym_packing(). 6079 */ 6080 static inline int 6081 fix_small_capacity(struct sched_domain *sd, struct sched_group *group) 6082 { 6083 /* 6084 * Only siblings can have significantly less than SCHED_CAPACITY_SCALE 6085 */ 6086 if (!(sd->flags & SD_SHARE_CPUCAPACITY)) 6087 return 0; 6088 6089 /* 6090 * If ~90% of the cpu_capacity is still there, we're good. 6091 */ 6092 if (group->sgc->capacity * 32 > group->sgc->capacity_orig * 29) 6093 return 1; 6094 6095 return 0; 6096 } 6097 6098 /* 6099 * Group imbalance indicates (and tries to solve) the problem where balancing 6100 * groups is inadequate due to tsk_cpus_allowed() constraints. 6101 * 6102 * Imagine a situation of two groups of 4 cpus each and 4 tasks each with a 6103 * cpumask covering 1 cpu of the first group and 3 cpus of the second group. 6104 * Something like: 6105 * 6106 * { 0 1 2 3 } { 4 5 6 7 } 6107 * * * * * 6108 * 6109 * If we were to balance group-wise we'd place two tasks in the first group and 6110 * two tasks in the second group. Clearly this is undesired as it will overload 6111 * cpu 3 and leave one of the cpus in the second group unused. 6112 * 6113 * The current solution to this issue is detecting the skew in the first group 6114 * by noticing the lower domain failed to reach balance and had difficulty 6115 * moving tasks due to affinity constraints. 6116 * 6117 * When this is so detected; this group becomes a candidate for busiest; see 6118 * update_sd_pick_busiest(). And calculate_imbalance() and 6119 * find_busiest_group() avoid some of the usual balance conditions to allow it 6120 * to create an effective group imbalance. 6121 * 6122 * This is a somewhat tricky proposition since the next run might not find the 6123 * group imbalance and decide the groups need to be balanced again. A most 6124 * subtle and fragile situation. 6125 */ 6126 6127 static inline int sg_imbalanced(struct sched_group *group) 6128 { 6129 return group->sgc->imbalance; 6130 } 6131 6132 /* 6133 * Compute the group capacity factor. 6134 * 6135 * Avoid the issue where N*frac(smt_capacity) >= 1 creates 'phantom' cores by 6136 * first dividing out the smt factor and computing the actual number of cores 6137 * and limit unit capacity with that. 6138 */ 6139 static inline int sg_capacity_factor(struct lb_env *env, struct sched_group *group) 6140 { 6141 unsigned int capacity_factor, smt, cpus; 6142 unsigned int capacity, capacity_orig; 6143 6144 capacity = group->sgc->capacity; 6145 capacity_orig = group->sgc->capacity_orig; 6146 cpus = group->group_weight; 6147 6148 /* smt := ceil(cpus / capacity), assumes: 1 < smt_capacity < 2 */ 6149 smt = DIV_ROUND_UP(SCHED_CAPACITY_SCALE * cpus, capacity_orig); 6150 capacity_factor = cpus / smt; /* cores */ 6151 6152 capacity_factor = min_t(unsigned, 6153 capacity_factor, DIV_ROUND_CLOSEST(capacity, SCHED_CAPACITY_SCALE)); 6154 if (!capacity_factor) 6155 capacity_factor = fix_small_capacity(env->sd, group); 6156 6157 return capacity_factor; 6158 } 6159 6160 static enum group_type 6161 group_classify(struct sched_group *group, struct sg_lb_stats *sgs) 6162 { 6163 if (sgs->sum_nr_running > sgs->group_capacity_factor) 6164 return group_overloaded; 6165 6166 if (sg_imbalanced(group)) 6167 return group_imbalanced; 6168 6169 return group_other; 6170 } 6171 6172 /** 6173 * update_sg_lb_stats - Update sched_group's statistics for load balancing. 6174 * @env: The load balancing environment. 6175 * @group: sched_group whose statistics are to be updated. 6176 * @load_idx: Load index of sched_domain of this_cpu for load calc. 6177 * @local_group: Does group contain this_cpu. 6178 * @sgs: variable to hold the statistics for this group. 6179 * @overload: Indicate more than one runnable task for any CPU. 6180 */ 6181 static inline void update_sg_lb_stats(struct lb_env *env, 6182 struct sched_group *group, int load_idx, 6183 int local_group, struct sg_lb_stats *sgs, 6184 bool *overload) 6185 { 6186 unsigned long load; 6187 int i; 6188 6189 memset(sgs, 0, sizeof(*sgs)); 6190 6191 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) { 6192 struct rq *rq = cpu_rq(i); 6193 6194 /* Bias balancing toward cpus of our domain */ 6195 if (local_group) 6196 load = target_load(i, load_idx); 6197 else 6198 load = source_load(i, load_idx); 6199 6200 sgs->group_load += load; 6201 sgs->sum_nr_running += rq->cfs.h_nr_running; 6202 6203 if (rq->nr_running > 1) 6204 *overload = true; 6205 6206 #ifdef CONFIG_NUMA_BALANCING 6207 sgs->nr_numa_running += rq->nr_numa_running; 6208 sgs->nr_preferred_running += rq->nr_preferred_running; 6209 #endif 6210 sgs->sum_weighted_load += weighted_cpuload(i); 6211 if (idle_cpu(i)) 6212 sgs->idle_cpus++; 6213 } 6214 6215 /* Adjust by relative CPU capacity of the group */ 6216 sgs->group_capacity = group->sgc->capacity; 6217 sgs->avg_load = (sgs->group_load*SCHED_CAPACITY_SCALE) / sgs->group_capacity; 6218 6219 if (sgs->sum_nr_running) 6220 sgs->load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running; 6221 6222 sgs->group_weight = group->group_weight; 6223 sgs->group_capacity_factor = sg_capacity_factor(env, group); 6224 sgs->group_type = group_classify(group, sgs); 6225 6226 if (sgs->group_capacity_factor > sgs->sum_nr_running) 6227 sgs->group_has_free_capacity = 1; 6228 } 6229 6230 /** 6231 * update_sd_pick_busiest - return 1 on busiest group 6232 * @env: The load balancing environment. 6233 * @sds: sched_domain statistics 6234 * @sg: sched_group candidate to be checked for being the busiest 6235 * @sgs: sched_group statistics 6236 * 6237 * Determine if @sg is a busier group than the previously selected 6238 * busiest group. 6239 * 6240 * Return: %true if @sg is a busier group than the previously selected 6241 * busiest group. %false otherwise. 6242 */ 6243 static bool update_sd_pick_busiest(struct lb_env *env, 6244 struct sd_lb_stats *sds, 6245 struct sched_group *sg, 6246 struct sg_lb_stats *sgs) 6247 { 6248 struct sg_lb_stats *busiest = &sds->busiest_stat; 6249 6250 if (sgs->group_type > busiest->group_type) 6251 return true; 6252 6253 if (sgs->group_type < busiest->group_type) 6254 return false; 6255 6256 if (sgs->avg_load <= busiest->avg_load) 6257 return false; 6258 6259 /* This is the busiest node in its class. */ 6260 if (!(env->sd->flags & SD_ASYM_PACKING)) 6261 return true; 6262 6263 /* 6264 * ASYM_PACKING needs to move all the work to the lowest 6265 * numbered CPUs in the group, therefore mark all groups 6266 * higher than ourself as busy. 6267 */ 6268 if (sgs->sum_nr_running && env->dst_cpu < group_first_cpu(sg)) { 6269 if (!sds->busiest) 6270 return true; 6271 6272 if (group_first_cpu(sds->busiest) > group_first_cpu(sg)) 6273 return true; 6274 } 6275 6276 return false; 6277 } 6278 6279 #ifdef CONFIG_NUMA_BALANCING 6280 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs) 6281 { 6282 if (sgs->sum_nr_running > sgs->nr_numa_running) 6283 return regular; 6284 if (sgs->sum_nr_running > sgs->nr_preferred_running) 6285 return remote; 6286 return all; 6287 } 6288 6289 static inline enum fbq_type fbq_classify_rq(struct rq *rq) 6290 { 6291 if (rq->nr_running > rq->nr_numa_running) 6292 return regular; 6293 if (rq->nr_running > rq->nr_preferred_running) 6294 return remote; 6295 return all; 6296 } 6297 #else 6298 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs) 6299 { 6300 return all; 6301 } 6302 6303 static inline enum fbq_type fbq_classify_rq(struct rq *rq) 6304 { 6305 return regular; 6306 } 6307 #endif /* CONFIG_NUMA_BALANCING */ 6308 6309 /** 6310 * update_sd_lb_stats - Update sched_domain's statistics for load balancing. 6311 * @env: The load balancing environment. 6312 * @sds: variable to hold the statistics for this sched_domain. 6313 */ 6314 static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds) 6315 { 6316 struct sched_domain *child = env->sd->child; 6317 struct sched_group *sg = env->sd->groups; 6318 struct sg_lb_stats tmp_sgs; 6319 int load_idx, prefer_sibling = 0; 6320 bool overload = false; 6321 6322 if (child && child->flags & SD_PREFER_SIBLING) 6323 prefer_sibling = 1; 6324 6325 load_idx = get_sd_load_idx(env->sd, env->idle); 6326 6327 do { 6328 struct sg_lb_stats *sgs = &tmp_sgs; 6329 int local_group; 6330 6331 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_cpus(sg)); 6332 if (local_group) { 6333 sds->local = sg; 6334 sgs = &sds->local_stat; 6335 6336 if (env->idle != CPU_NEWLY_IDLE || 6337 time_after_eq(jiffies, sg->sgc->next_update)) 6338 update_group_capacity(env->sd, env->dst_cpu); 6339 } 6340 6341 update_sg_lb_stats(env, sg, load_idx, local_group, sgs, 6342 &overload); 6343 6344 if (local_group) 6345 goto next_group; 6346 6347 /* 6348 * In case the child domain prefers tasks go to siblings 6349 * first, lower the sg capacity factor to one so that we'll try 6350 * and move all the excess tasks away. We lower the capacity 6351 * of a group only if the local group has the capacity to fit 6352 * these excess tasks, i.e. nr_running < group_capacity_factor. The 6353 * extra check prevents the case where you always pull from the 6354 * heaviest group when it is already under-utilized (possible 6355 * with a large weight task outweighs the tasks on the system). 6356 */ 6357 if (prefer_sibling && sds->local && 6358 sds->local_stat.group_has_free_capacity) { 6359 sgs->group_capacity_factor = min(sgs->group_capacity_factor, 1U); 6360 sgs->group_type = group_classify(sg, sgs); 6361 } 6362 6363 if (update_sd_pick_busiest(env, sds, sg, sgs)) { 6364 sds->busiest = sg; 6365 sds->busiest_stat = *sgs; 6366 } 6367 6368 next_group: 6369 /* Now, start updating sd_lb_stats */ 6370 sds->total_load += sgs->group_load; 6371 sds->total_capacity += sgs->group_capacity; 6372 6373 sg = sg->next; 6374 } while (sg != env->sd->groups); 6375 6376 if (env->sd->flags & SD_NUMA) 6377 env->fbq_type = fbq_classify_group(&sds->busiest_stat); 6378 6379 if (!env->sd->parent) { 6380 /* update overload indicator if we are at root domain */ 6381 if (env->dst_rq->rd->overload != overload) 6382 env->dst_rq->rd->overload = overload; 6383 } 6384 6385 } 6386 6387 /** 6388 * check_asym_packing - Check to see if the group is packed into the 6389 * sched doman. 6390 * 6391 * This is primarily intended to used at the sibling level. Some 6392 * cores like POWER7 prefer to use lower numbered SMT threads. In the 6393 * case of POWER7, it can move to lower SMT modes only when higher 6394 * threads are idle. When in lower SMT modes, the threads will 6395 * perform better since they share less core resources. Hence when we 6396 * have idle threads, we want them to be the higher ones. 6397 * 6398 * This packing function is run on idle threads. It checks to see if 6399 * the busiest CPU in this domain (core in the P7 case) has a higher 6400 * CPU number than the packing function is being run on. Here we are 6401 * assuming lower CPU number will be equivalent to lower a SMT thread 6402 * number. 6403 * 6404 * Return: 1 when packing is required and a task should be moved to 6405 * this CPU. The amount of the imbalance is returned in *imbalance. 6406 * 6407 * @env: The load balancing environment. 6408 * @sds: Statistics of the sched_domain which is to be packed 6409 */ 6410 static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds) 6411 { 6412 int busiest_cpu; 6413 6414 if (!(env->sd->flags & SD_ASYM_PACKING)) 6415 return 0; 6416 6417 if (!sds->busiest) 6418 return 0; 6419 6420 busiest_cpu = group_first_cpu(sds->busiest); 6421 if (env->dst_cpu > busiest_cpu) 6422 return 0; 6423 6424 env->imbalance = DIV_ROUND_CLOSEST( 6425 sds->busiest_stat.avg_load * sds->busiest_stat.group_capacity, 6426 SCHED_CAPACITY_SCALE); 6427 6428 return 1; 6429 } 6430 6431 /** 6432 * fix_small_imbalance - Calculate the minor imbalance that exists 6433 * amongst the groups of a sched_domain, during 6434 * load balancing. 6435 * @env: The load balancing environment. 6436 * @sds: Statistics of the sched_domain whose imbalance is to be calculated. 6437 */ 6438 static inline 6439 void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds) 6440 { 6441 unsigned long tmp, capa_now = 0, capa_move = 0; 6442 unsigned int imbn = 2; 6443 unsigned long scaled_busy_load_per_task; 6444 struct sg_lb_stats *local, *busiest; 6445 6446 local = &sds->local_stat; 6447 busiest = &sds->busiest_stat; 6448 6449 if (!local->sum_nr_running) 6450 local->load_per_task = cpu_avg_load_per_task(env->dst_cpu); 6451 else if (busiest->load_per_task > local->load_per_task) 6452 imbn = 1; 6453 6454 scaled_busy_load_per_task = 6455 (busiest->load_per_task * SCHED_CAPACITY_SCALE) / 6456 busiest->group_capacity; 6457 6458 if (busiest->avg_load + scaled_busy_load_per_task >= 6459 local->avg_load + (scaled_busy_load_per_task * imbn)) { 6460 env->imbalance = busiest->load_per_task; 6461 return; 6462 } 6463 6464 /* 6465 * OK, we don't have enough imbalance to justify moving tasks, 6466 * however we may be able to increase total CPU capacity used by 6467 * moving them. 6468 */ 6469 6470 capa_now += busiest->group_capacity * 6471 min(busiest->load_per_task, busiest->avg_load); 6472 capa_now += local->group_capacity * 6473 min(local->load_per_task, local->avg_load); 6474 capa_now /= SCHED_CAPACITY_SCALE; 6475 6476 /* Amount of load we'd subtract */ 6477 if (busiest->avg_load > scaled_busy_load_per_task) { 6478 capa_move += busiest->group_capacity * 6479 min(busiest->load_per_task, 6480 busiest->avg_load - scaled_busy_load_per_task); 6481 } 6482 6483 /* Amount of load we'd add */ 6484 if (busiest->avg_load * busiest->group_capacity < 6485 busiest->load_per_task * SCHED_CAPACITY_SCALE) { 6486 tmp = (busiest->avg_load * busiest->group_capacity) / 6487 local->group_capacity; 6488 } else { 6489 tmp = (busiest->load_per_task * SCHED_CAPACITY_SCALE) / 6490 local->group_capacity; 6491 } 6492 capa_move += local->group_capacity * 6493 min(local->load_per_task, local->avg_load + tmp); 6494 capa_move /= SCHED_CAPACITY_SCALE; 6495 6496 /* Move if we gain throughput */ 6497 if (capa_move > capa_now) 6498 env->imbalance = busiest->load_per_task; 6499 } 6500 6501 /** 6502 * calculate_imbalance - Calculate the amount of imbalance present within the 6503 * groups of a given sched_domain during load balance. 6504 * @env: load balance environment 6505 * @sds: statistics of the sched_domain whose imbalance is to be calculated. 6506 */ 6507 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds) 6508 { 6509 unsigned long max_pull, load_above_capacity = ~0UL; 6510 struct sg_lb_stats *local, *busiest; 6511 6512 local = &sds->local_stat; 6513 busiest = &sds->busiest_stat; 6514 6515 if (busiest->group_type == group_imbalanced) { 6516 /* 6517 * In the group_imb case we cannot rely on group-wide averages 6518 * to ensure cpu-load equilibrium, look at wider averages. XXX 6519 */ 6520 busiest->load_per_task = 6521 min(busiest->load_per_task, sds->avg_load); 6522 } 6523 6524 /* 6525 * In the presence of smp nice balancing, certain scenarios can have 6526 * max load less than avg load(as we skip the groups at or below 6527 * its cpu_capacity, while calculating max_load..) 6528 */ 6529 if (busiest->avg_load <= sds->avg_load || 6530 local->avg_load >= sds->avg_load) { 6531 env->imbalance = 0; 6532 return fix_small_imbalance(env, sds); 6533 } 6534 6535 /* 6536 * If there aren't any idle cpus, avoid creating some. 6537 */ 6538 if (busiest->group_type == group_overloaded && 6539 local->group_type == group_overloaded) { 6540 load_above_capacity = 6541 (busiest->sum_nr_running - busiest->group_capacity_factor); 6542 6543 load_above_capacity *= (SCHED_LOAD_SCALE * SCHED_CAPACITY_SCALE); 6544 load_above_capacity /= busiest->group_capacity; 6545 } 6546 6547 /* 6548 * We're trying to get all the cpus to the average_load, so we don't 6549 * want to push ourselves above the average load, nor do we wish to 6550 * reduce the max loaded cpu below the average load. At the same time, 6551 * we also don't want to reduce the group load below the group capacity 6552 * (so that we can implement power-savings policies etc). Thus we look 6553 * for the minimum possible imbalance. 6554 */ 6555 max_pull = min(busiest->avg_load - sds->avg_load, load_above_capacity); 6556 6557 /* How much load to actually move to equalise the imbalance */ 6558 env->imbalance = min( 6559 max_pull * busiest->group_capacity, 6560 (sds->avg_load - local->avg_load) * local->group_capacity 6561 ) / SCHED_CAPACITY_SCALE; 6562 6563 /* 6564 * if *imbalance is less than the average load per runnable task 6565 * there is no guarantee that any tasks will be moved so we'll have 6566 * a think about bumping its value to force at least one task to be 6567 * moved 6568 */ 6569 if (env->imbalance < busiest->load_per_task) 6570 return fix_small_imbalance(env, sds); 6571 } 6572 6573 /******* find_busiest_group() helpers end here *********************/ 6574 6575 /** 6576 * find_busiest_group - Returns the busiest group within the sched_domain 6577 * if there is an imbalance. If there isn't an imbalance, and 6578 * the user has opted for power-savings, it returns a group whose 6579 * CPUs can be put to idle by rebalancing those tasks elsewhere, if 6580 * such a group exists. 6581 * 6582 * Also calculates the amount of weighted load which should be moved 6583 * to restore balance. 6584 * 6585 * @env: The load balancing environment. 6586 * 6587 * Return: - The busiest group if imbalance exists. 6588 * - If no imbalance and user has opted for power-savings balance, 6589 * return the least loaded group whose CPUs can be 6590 * put to idle by rebalancing its tasks onto our group. 6591 */ 6592 static struct sched_group *find_busiest_group(struct lb_env *env) 6593 { 6594 struct sg_lb_stats *local, *busiest; 6595 struct sd_lb_stats sds; 6596 6597 init_sd_lb_stats(&sds); 6598 6599 /* 6600 * Compute the various statistics relavent for load balancing at 6601 * this level. 6602 */ 6603 update_sd_lb_stats(env, &sds); 6604 local = &sds.local_stat; 6605 busiest = &sds.busiest_stat; 6606 6607 if ((env->idle == CPU_IDLE || env->idle == CPU_NEWLY_IDLE) && 6608 check_asym_packing(env, &sds)) 6609 return sds.busiest; 6610 6611 /* There is no busy sibling group to pull tasks from */ 6612 if (!sds.busiest || busiest->sum_nr_running == 0) 6613 goto out_balanced; 6614 6615 sds.avg_load = (SCHED_CAPACITY_SCALE * sds.total_load) 6616 / sds.total_capacity; 6617 6618 /* 6619 * If the busiest group is imbalanced the below checks don't 6620 * work because they assume all things are equal, which typically 6621 * isn't true due to cpus_allowed constraints and the like. 6622 */ 6623 if (busiest->group_type == group_imbalanced) 6624 goto force_balance; 6625 6626 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */ 6627 if (env->idle == CPU_NEWLY_IDLE && local->group_has_free_capacity && 6628 !busiest->group_has_free_capacity) 6629 goto force_balance; 6630 6631 /* 6632 * If the local group is busier than the selected busiest group 6633 * don't try and pull any tasks. 6634 */ 6635 if (local->avg_load >= busiest->avg_load) 6636 goto out_balanced; 6637 6638 /* 6639 * Don't pull any tasks if this group is already above the domain 6640 * average load. 6641 */ 6642 if (local->avg_load >= sds.avg_load) 6643 goto out_balanced; 6644 6645 if (env->idle == CPU_IDLE) { 6646 /* 6647 * This cpu is idle. If the busiest group is not overloaded 6648 * and there is no imbalance between this and busiest group 6649 * wrt idle cpus, it is balanced. The imbalance becomes 6650 * significant if the diff is greater than 1 otherwise we 6651 * might end up to just move the imbalance on another group 6652 */ 6653 if ((busiest->group_type != group_overloaded) && 6654 (local->idle_cpus <= (busiest->idle_cpus + 1))) 6655 goto out_balanced; 6656 } else { 6657 /* 6658 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use 6659 * imbalance_pct to be conservative. 6660 */ 6661 if (100 * busiest->avg_load <= 6662 env->sd->imbalance_pct * local->avg_load) 6663 goto out_balanced; 6664 } 6665 6666 force_balance: 6667 /* Looks like there is an imbalance. Compute it */ 6668 calculate_imbalance(env, &sds); 6669 return sds.busiest; 6670 6671 out_balanced: 6672 env->imbalance = 0; 6673 return NULL; 6674 } 6675 6676 /* 6677 * find_busiest_queue - find the busiest runqueue among the cpus in group. 6678 */ 6679 static struct rq *find_busiest_queue(struct lb_env *env, 6680 struct sched_group *group) 6681 { 6682 struct rq *busiest = NULL, *rq; 6683 unsigned long busiest_load = 0, busiest_capacity = 1; 6684 int i; 6685 6686 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) { 6687 unsigned long capacity, capacity_factor, wl; 6688 enum fbq_type rt; 6689 6690 rq = cpu_rq(i); 6691 rt = fbq_classify_rq(rq); 6692 6693 /* 6694 * We classify groups/runqueues into three groups: 6695 * - regular: there are !numa tasks 6696 * - remote: there are numa tasks that run on the 'wrong' node 6697 * - all: there is no distinction 6698 * 6699 * In order to avoid migrating ideally placed numa tasks, 6700 * ignore those when there's better options. 6701 * 6702 * If we ignore the actual busiest queue to migrate another 6703 * task, the next balance pass can still reduce the busiest 6704 * queue by moving tasks around inside the node. 6705 * 6706 * If we cannot move enough load due to this classification 6707 * the next pass will adjust the group classification and 6708 * allow migration of more tasks. 6709 * 6710 * Both cases only affect the total convergence complexity. 6711 */ 6712 if (rt > env->fbq_type) 6713 continue; 6714 6715 capacity = capacity_of(i); 6716 capacity_factor = DIV_ROUND_CLOSEST(capacity, SCHED_CAPACITY_SCALE); 6717 if (!capacity_factor) 6718 capacity_factor = fix_small_capacity(env->sd, group); 6719 6720 wl = weighted_cpuload(i); 6721 6722 /* 6723 * When comparing with imbalance, use weighted_cpuload() 6724 * which is not scaled with the cpu capacity. 6725 */ 6726 if (capacity_factor && rq->nr_running == 1 && wl > env->imbalance) 6727 continue; 6728 6729 /* 6730 * For the load comparisons with the other cpu's, consider 6731 * the weighted_cpuload() scaled with the cpu capacity, so 6732 * that the load can be moved away from the cpu that is 6733 * potentially running at a lower capacity. 6734 * 6735 * Thus we're looking for max(wl_i / capacity_i), crosswise 6736 * multiplication to rid ourselves of the division works out 6737 * to: wl_i * capacity_j > wl_j * capacity_i; where j is 6738 * our previous maximum. 6739 */ 6740 if (wl * busiest_capacity > busiest_load * capacity) { 6741 busiest_load = wl; 6742 busiest_capacity = capacity; 6743 busiest = rq; 6744 } 6745 } 6746 6747 return busiest; 6748 } 6749 6750 /* 6751 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but 6752 * so long as it is large enough. 6753 */ 6754 #define MAX_PINNED_INTERVAL 512 6755 6756 /* Working cpumask for load_balance and load_balance_newidle. */ 6757 DEFINE_PER_CPU(cpumask_var_t, load_balance_mask); 6758 6759 static int need_active_balance(struct lb_env *env) 6760 { 6761 struct sched_domain *sd = env->sd; 6762 6763 if (env->idle == CPU_NEWLY_IDLE) { 6764 6765 /* 6766 * ASYM_PACKING needs to force migrate tasks from busy but 6767 * higher numbered CPUs in order to pack all tasks in the 6768 * lowest numbered CPUs. 6769 */ 6770 if ((sd->flags & SD_ASYM_PACKING) && env->src_cpu > env->dst_cpu) 6771 return 1; 6772 } 6773 6774 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2); 6775 } 6776 6777 static int active_load_balance_cpu_stop(void *data); 6778 6779 static int should_we_balance(struct lb_env *env) 6780 { 6781 struct sched_group *sg = env->sd->groups; 6782 struct cpumask *sg_cpus, *sg_mask; 6783 int cpu, balance_cpu = -1; 6784 6785 /* 6786 * In the newly idle case, we will allow all the cpu's 6787 * to do the newly idle load balance. 6788 */ 6789 if (env->idle == CPU_NEWLY_IDLE) 6790 return 1; 6791 6792 sg_cpus = sched_group_cpus(sg); 6793 sg_mask = sched_group_mask(sg); 6794 /* Try to find first idle cpu */ 6795 for_each_cpu_and(cpu, sg_cpus, env->cpus) { 6796 if (!cpumask_test_cpu(cpu, sg_mask) || !idle_cpu(cpu)) 6797 continue; 6798 6799 balance_cpu = cpu; 6800 break; 6801 } 6802 6803 if (balance_cpu == -1) 6804 balance_cpu = group_balance_cpu(sg); 6805 6806 /* 6807 * First idle cpu or the first cpu(busiest) in this sched group 6808 * is eligible for doing load balancing at this and above domains. 6809 */ 6810 return balance_cpu == env->dst_cpu; 6811 } 6812 6813 /* 6814 * Check this_cpu to ensure it is balanced within domain. Attempt to move 6815 * tasks if there is an imbalance. 6816 */ 6817 static int load_balance(int this_cpu, struct rq *this_rq, 6818 struct sched_domain *sd, enum cpu_idle_type idle, 6819 int *continue_balancing) 6820 { 6821 int ld_moved, cur_ld_moved, active_balance = 0; 6822 struct sched_domain *sd_parent = sd->parent; 6823 struct sched_group *group; 6824 struct rq *busiest; 6825 unsigned long flags; 6826 struct cpumask *cpus = this_cpu_cpumask_var_ptr(load_balance_mask); 6827 6828 struct lb_env env = { 6829 .sd = sd, 6830 .dst_cpu = this_cpu, 6831 .dst_rq = this_rq, 6832 .dst_grpmask = sched_group_cpus(sd->groups), 6833 .idle = idle, 6834 .loop_break = sched_nr_migrate_break, 6835 .cpus = cpus, 6836 .fbq_type = all, 6837 .tasks = LIST_HEAD_INIT(env.tasks), 6838 }; 6839 6840 /* 6841 * For NEWLY_IDLE load_balancing, we don't need to consider 6842 * other cpus in our group 6843 */ 6844 if (idle == CPU_NEWLY_IDLE) 6845 env.dst_grpmask = NULL; 6846 6847 cpumask_copy(cpus, cpu_active_mask); 6848 6849 schedstat_inc(sd, lb_count[idle]); 6850 6851 redo: 6852 if (!should_we_balance(&env)) { 6853 *continue_balancing = 0; 6854 goto out_balanced; 6855 } 6856 6857 group = find_busiest_group(&env); 6858 if (!group) { 6859 schedstat_inc(sd, lb_nobusyg[idle]); 6860 goto out_balanced; 6861 } 6862 6863 busiest = find_busiest_queue(&env, group); 6864 if (!busiest) { 6865 schedstat_inc(sd, lb_nobusyq[idle]); 6866 goto out_balanced; 6867 } 6868 6869 BUG_ON(busiest == env.dst_rq); 6870 6871 schedstat_add(sd, lb_imbalance[idle], env.imbalance); 6872 6873 ld_moved = 0; 6874 if (busiest->nr_running > 1) { 6875 /* 6876 * Attempt to move tasks. If find_busiest_group has found 6877 * an imbalance but busiest->nr_running <= 1, the group is 6878 * still unbalanced. ld_moved simply stays zero, so it is 6879 * correctly treated as an imbalance. 6880 */ 6881 env.flags |= LBF_ALL_PINNED; 6882 env.src_cpu = busiest->cpu; 6883 env.src_rq = busiest; 6884 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running); 6885 6886 more_balance: 6887 raw_spin_lock_irqsave(&busiest->lock, flags); 6888 6889 /* 6890 * cur_ld_moved - load moved in current iteration 6891 * ld_moved - cumulative load moved across iterations 6892 */ 6893 cur_ld_moved = detach_tasks(&env); 6894 6895 /* 6896 * We've detached some tasks from busiest_rq. Every 6897 * task is masked "TASK_ON_RQ_MIGRATING", so we can safely 6898 * unlock busiest->lock, and we are able to be sure 6899 * that nobody can manipulate the tasks in parallel. 6900 * See task_rq_lock() family for the details. 6901 */ 6902 6903 raw_spin_unlock(&busiest->lock); 6904 6905 if (cur_ld_moved) { 6906 attach_tasks(&env); 6907 ld_moved += cur_ld_moved; 6908 } 6909 6910 local_irq_restore(flags); 6911 6912 if (env.flags & LBF_NEED_BREAK) { 6913 env.flags &= ~LBF_NEED_BREAK; 6914 goto more_balance; 6915 } 6916 6917 /* 6918 * Revisit (affine) tasks on src_cpu that couldn't be moved to 6919 * us and move them to an alternate dst_cpu in our sched_group 6920 * where they can run. The upper limit on how many times we 6921 * iterate on same src_cpu is dependent on number of cpus in our 6922 * sched_group. 6923 * 6924 * This changes load balance semantics a bit on who can move 6925 * load to a given_cpu. In addition to the given_cpu itself 6926 * (or a ilb_cpu acting on its behalf where given_cpu is 6927 * nohz-idle), we now have balance_cpu in a position to move 6928 * load to given_cpu. In rare situations, this may cause 6929 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding 6930 * _independently_ and at _same_ time to move some load to 6931 * given_cpu) causing exceess load to be moved to given_cpu. 6932 * This however should not happen so much in practice and 6933 * moreover subsequent load balance cycles should correct the 6934 * excess load moved. 6935 */ 6936 if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) { 6937 6938 /* Prevent to re-select dst_cpu via env's cpus */ 6939 cpumask_clear_cpu(env.dst_cpu, env.cpus); 6940 6941 env.dst_rq = cpu_rq(env.new_dst_cpu); 6942 env.dst_cpu = env.new_dst_cpu; 6943 env.flags &= ~LBF_DST_PINNED; 6944 env.loop = 0; 6945 env.loop_break = sched_nr_migrate_break; 6946 6947 /* 6948 * Go back to "more_balance" rather than "redo" since we 6949 * need to continue with same src_cpu. 6950 */ 6951 goto more_balance; 6952 } 6953 6954 /* 6955 * We failed to reach balance because of affinity. 6956 */ 6957 if (sd_parent) { 6958 int *group_imbalance = &sd_parent->groups->sgc->imbalance; 6959 6960 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0) 6961 *group_imbalance = 1; 6962 } 6963 6964 /* All tasks on this runqueue were pinned by CPU affinity */ 6965 if (unlikely(env.flags & LBF_ALL_PINNED)) { 6966 cpumask_clear_cpu(cpu_of(busiest), cpus); 6967 if (!cpumask_empty(cpus)) { 6968 env.loop = 0; 6969 env.loop_break = sched_nr_migrate_break; 6970 goto redo; 6971 } 6972 goto out_all_pinned; 6973 } 6974 } 6975 6976 if (!ld_moved) { 6977 schedstat_inc(sd, lb_failed[idle]); 6978 /* 6979 * Increment the failure counter only on periodic balance. 6980 * We do not want newidle balance, which can be very 6981 * frequent, pollute the failure counter causing 6982 * excessive cache_hot migrations and active balances. 6983 */ 6984 if (idle != CPU_NEWLY_IDLE) 6985 sd->nr_balance_failed++; 6986 6987 if (need_active_balance(&env)) { 6988 raw_spin_lock_irqsave(&busiest->lock, flags); 6989 6990 /* don't kick the active_load_balance_cpu_stop, 6991 * if the curr task on busiest cpu can't be 6992 * moved to this_cpu 6993 */ 6994 if (!cpumask_test_cpu(this_cpu, 6995 tsk_cpus_allowed(busiest->curr))) { 6996 raw_spin_unlock_irqrestore(&busiest->lock, 6997 flags); 6998 env.flags |= LBF_ALL_PINNED; 6999 goto out_one_pinned; 7000 } 7001 7002 /* 7003 * ->active_balance synchronizes accesses to 7004 * ->active_balance_work. Once set, it's cleared 7005 * only after active load balance is finished. 7006 */ 7007 if (!busiest->active_balance) { 7008 busiest->active_balance = 1; 7009 busiest->push_cpu = this_cpu; 7010 active_balance = 1; 7011 } 7012 raw_spin_unlock_irqrestore(&busiest->lock, flags); 7013 7014 if (active_balance) { 7015 stop_one_cpu_nowait(cpu_of(busiest), 7016 active_load_balance_cpu_stop, busiest, 7017 &busiest->active_balance_work); 7018 } 7019 7020 /* 7021 * We've kicked active balancing, reset the failure 7022 * counter. 7023 */ 7024 sd->nr_balance_failed = sd->cache_nice_tries+1; 7025 } 7026 } else 7027 sd->nr_balance_failed = 0; 7028 7029 if (likely(!active_balance)) { 7030 /* We were unbalanced, so reset the balancing interval */ 7031 sd->balance_interval = sd->min_interval; 7032 } else { 7033 /* 7034 * If we've begun active balancing, start to back off. This 7035 * case may not be covered by the all_pinned logic if there 7036 * is only 1 task on the busy runqueue (because we don't call 7037 * detach_tasks). 7038 */ 7039 if (sd->balance_interval < sd->max_interval) 7040 sd->balance_interval *= 2; 7041 } 7042 7043 goto out; 7044 7045 out_balanced: 7046 /* 7047 * We reach balance although we may have faced some affinity 7048 * constraints. Clear the imbalance flag if it was set. 7049 */ 7050 if (sd_parent) { 7051 int *group_imbalance = &sd_parent->groups->sgc->imbalance; 7052 7053 if (*group_imbalance) 7054 *group_imbalance = 0; 7055 } 7056 7057 out_all_pinned: 7058 /* 7059 * We reach balance because all tasks are pinned at this level so 7060 * we can't migrate them. Let the imbalance flag set so parent level 7061 * can try to migrate them. 7062 */ 7063 schedstat_inc(sd, lb_balanced[idle]); 7064 7065 sd->nr_balance_failed = 0; 7066 7067 out_one_pinned: 7068 /* tune up the balancing interval */ 7069 if (((env.flags & LBF_ALL_PINNED) && 7070 sd->balance_interval < MAX_PINNED_INTERVAL) || 7071 (sd->balance_interval < sd->max_interval)) 7072 sd->balance_interval *= 2; 7073 7074 ld_moved = 0; 7075 out: 7076 return ld_moved; 7077 } 7078 7079 static inline unsigned long 7080 get_sd_balance_interval(struct sched_domain *sd, int cpu_busy) 7081 { 7082 unsigned long interval = sd->balance_interval; 7083 7084 if (cpu_busy) 7085 interval *= sd->busy_factor; 7086 7087 /* scale ms to jiffies */ 7088 interval = msecs_to_jiffies(interval); 7089 interval = clamp(interval, 1UL, max_load_balance_interval); 7090 7091 return interval; 7092 } 7093 7094 static inline void 7095 update_next_balance(struct sched_domain *sd, int cpu_busy, unsigned long *next_balance) 7096 { 7097 unsigned long interval, next; 7098 7099 interval = get_sd_balance_interval(sd, cpu_busy); 7100 next = sd->last_balance + interval; 7101 7102 if (time_after(*next_balance, next)) 7103 *next_balance = next; 7104 } 7105 7106 /* 7107 * idle_balance is called by schedule() if this_cpu is about to become 7108 * idle. Attempts to pull tasks from other CPUs. 7109 */ 7110 static int idle_balance(struct rq *this_rq) 7111 { 7112 unsigned long next_balance = jiffies + HZ; 7113 int this_cpu = this_rq->cpu; 7114 struct sched_domain *sd; 7115 int pulled_task = 0; 7116 u64 curr_cost = 0; 7117 7118 idle_enter_fair(this_rq); 7119 7120 /* 7121 * We must set idle_stamp _before_ calling idle_balance(), such that we 7122 * measure the duration of idle_balance() as idle time. 7123 */ 7124 this_rq->idle_stamp = rq_clock(this_rq); 7125 7126 if (this_rq->avg_idle < sysctl_sched_migration_cost || 7127 !this_rq->rd->overload) { 7128 rcu_read_lock(); 7129 sd = rcu_dereference_check_sched_domain(this_rq->sd); 7130 if (sd) 7131 update_next_balance(sd, 0, &next_balance); 7132 rcu_read_unlock(); 7133 7134 goto out; 7135 } 7136 7137 /* 7138 * Drop the rq->lock, but keep IRQ/preempt disabled. 7139 */ 7140 raw_spin_unlock(&this_rq->lock); 7141 7142 update_blocked_averages(this_cpu); 7143 rcu_read_lock(); 7144 for_each_domain(this_cpu, sd) { 7145 int continue_balancing = 1; 7146 u64 t0, domain_cost; 7147 7148 if (!(sd->flags & SD_LOAD_BALANCE)) 7149 continue; 7150 7151 if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost) { 7152 update_next_balance(sd, 0, &next_balance); 7153 break; 7154 } 7155 7156 if (sd->flags & SD_BALANCE_NEWIDLE) { 7157 t0 = sched_clock_cpu(this_cpu); 7158 7159 pulled_task = load_balance(this_cpu, this_rq, 7160 sd, CPU_NEWLY_IDLE, 7161 &continue_balancing); 7162 7163 domain_cost = sched_clock_cpu(this_cpu) - t0; 7164 if (domain_cost > sd->max_newidle_lb_cost) 7165 sd->max_newidle_lb_cost = domain_cost; 7166 7167 curr_cost += domain_cost; 7168 } 7169 7170 update_next_balance(sd, 0, &next_balance); 7171 7172 /* 7173 * Stop searching for tasks to pull if there are 7174 * now runnable tasks on this rq. 7175 */ 7176 if (pulled_task || this_rq->nr_running > 0) 7177 break; 7178 } 7179 rcu_read_unlock(); 7180 7181 raw_spin_lock(&this_rq->lock); 7182 7183 if (curr_cost > this_rq->max_idle_balance_cost) 7184 this_rq->max_idle_balance_cost = curr_cost; 7185 7186 /* 7187 * While browsing the domains, we released the rq lock, a task could 7188 * have been enqueued in the meantime. Since we're not going idle, 7189 * pretend we pulled a task. 7190 */ 7191 if (this_rq->cfs.h_nr_running && !pulled_task) 7192 pulled_task = 1; 7193 7194 out: 7195 /* Move the next balance forward */ 7196 if (time_after(this_rq->next_balance, next_balance)) 7197 this_rq->next_balance = next_balance; 7198 7199 /* Is there a task of a high priority class? */ 7200 if (this_rq->nr_running != this_rq->cfs.h_nr_running) 7201 pulled_task = -1; 7202 7203 if (pulled_task) { 7204 idle_exit_fair(this_rq); 7205 this_rq->idle_stamp = 0; 7206 } 7207 7208 return pulled_task; 7209 } 7210 7211 /* 7212 * active_load_balance_cpu_stop is run by cpu stopper. It pushes 7213 * running tasks off the busiest CPU onto idle CPUs. It requires at 7214 * least 1 task to be running on each physical CPU where possible, and 7215 * avoids physical / logical imbalances. 7216 */ 7217 static int active_load_balance_cpu_stop(void *data) 7218 { 7219 struct rq *busiest_rq = data; 7220 int busiest_cpu = cpu_of(busiest_rq); 7221 int target_cpu = busiest_rq->push_cpu; 7222 struct rq *target_rq = cpu_rq(target_cpu); 7223 struct sched_domain *sd; 7224 struct task_struct *p = NULL; 7225 7226 raw_spin_lock_irq(&busiest_rq->lock); 7227 7228 /* make sure the requested cpu hasn't gone down in the meantime */ 7229 if (unlikely(busiest_cpu != smp_processor_id() || 7230 !busiest_rq->active_balance)) 7231 goto out_unlock; 7232 7233 /* Is there any task to move? */ 7234 if (busiest_rq->nr_running <= 1) 7235 goto out_unlock; 7236 7237 /* 7238 * This condition is "impossible", if it occurs 7239 * we need to fix it. Originally reported by 7240 * Bjorn Helgaas on a 128-cpu setup. 7241 */ 7242 BUG_ON(busiest_rq == target_rq); 7243 7244 /* Search for an sd spanning us and the target CPU. */ 7245 rcu_read_lock(); 7246 for_each_domain(target_cpu, sd) { 7247 if ((sd->flags & SD_LOAD_BALANCE) && 7248 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd))) 7249 break; 7250 } 7251 7252 if (likely(sd)) { 7253 struct lb_env env = { 7254 .sd = sd, 7255 .dst_cpu = target_cpu, 7256 .dst_rq = target_rq, 7257 .src_cpu = busiest_rq->cpu, 7258 .src_rq = busiest_rq, 7259 .idle = CPU_IDLE, 7260 }; 7261 7262 schedstat_inc(sd, alb_count); 7263 7264 p = detach_one_task(&env); 7265 if (p) 7266 schedstat_inc(sd, alb_pushed); 7267 else 7268 schedstat_inc(sd, alb_failed); 7269 } 7270 rcu_read_unlock(); 7271 out_unlock: 7272 busiest_rq->active_balance = 0; 7273 raw_spin_unlock(&busiest_rq->lock); 7274 7275 if (p) 7276 attach_one_task(target_rq, p); 7277 7278 local_irq_enable(); 7279 7280 return 0; 7281 } 7282 7283 static inline int on_null_domain(struct rq *rq) 7284 { 7285 return unlikely(!rcu_dereference_sched(rq->sd)); 7286 } 7287 7288 #ifdef CONFIG_NO_HZ_COMMON 7289 /* 7290 * idle load balancing details 7291 * - When one of the busy CPUs notice that there may be an idle rebalancing 7292 * needed, they will kick the idle load balancer, which then does idle 7293 * load balancing for all the idle CPUs. 7294 */ 7295 static struct { 7296 cpumask_var_t idle_cpus_mask; 7297 atomic_t nr_cpus; 7298 unsigned long next_balance; /* in jiffy units */ 7299 } nohz ____cacheline_aligned; 7300 7301 static inline int find_new_ilb(void) 7302 { 7303 int ilb = cpumask_first(nohz.idle_cpus_mask); 7304 7305 if (ilb < nr_cpu_ids && idle_cpu(ilb)) 7306 return ilb; 7307 7308 return nr_cpu_ids; 7309 } 7310 7311 /* 7312 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the 7313 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle 7314 * CPU (if there is one). 7315 */ 7316 static void nohz_balancer_kick(void) 7317 { 7318 int ilb_cpu; 7319 7320 nohz.next_balance++; 7321 7322 ilb_cpu = find_new_ilb(); 7323 7324 if (ilb_cpu >= nr_cpu_ids) 7325 return; 7326 7327 if (test_and_set_bit(NOHZ_BALANCE_KICK, nohz_flags(ilb_cpu))) 7328 return; 7329 /* 7330 * Use smp_send_reschedule() instead of resched_cpu(). 7331 * This way we generate a sched IPI on the target cpu which 7332 * is idle. And the softirq performing nohz idle load balance 7333 * will be run before returning from the IPI. 7334 */ 7335 smp_send_reschedule(ilb_cpu); 7336 return; 7337 } 7338 7339 static inline void nohz_balance_exit_idle(int cpu) 7340 { 7341 if (unlikely(test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))) { 7342 /* 7343 * Completely isolated CPUs don't ever set, so we must test. 7344 */ 7345 if (likely(cpumask_test_cpu(cpu, nohz.idle_cpus_mask))) { 7346 cpumask_clear_cpu(cpu, nohz.idle_cpus_mask); 7347 atomic_dec(&nohz.nr_cpus); 7348 } 7349 clear_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)); 7350 } 7351 } 7352 7353 static inline void set_cpu_sd_state_busy(void) 7354 { 7355 struct sched_domain *sd; 7356 int cpu = smp_processor_id(); 7357 7358 rcu_read_lock(); 7359 sd = rcu_dereference(per_cpu(sd_busy, cpu)); 7360 7361 if (!sd || !sd->nohz_idle) 7362 goto unlock; 7363 sd->nohz_idle = 0; 7364 7365 atomic_inc(&sd->groups->sgc->nr_busy_cpus); 7366 unlock: 7367 rcu_read_unlock(); 7368 } 7369 7370 void set_cpu_sd_state_idle(void) 7371 { 7372 struct sched_domain *sd; 7373 int cpu = smp_processor_id(); 7374 7375 rcu_read_lock(); 7376 sd = rcu_dereference(per_cpu(sd_busy, cpu)); 7377 7378 if (!sd || sd->nohz_idle) 7379 goto unlock; 7380 sd->nohz_idle = 1; 7381 7382 atomic_dec(&sd->groups->sgc->nr_busy_cpus); 7383 unlock: 7384 rcu_read_unlock(); 7385 } 7386 7387 /* 7388 * This routine will record that the cpu is going idle with tick stopped. 7389 * This info will be used in performing idle load balancing in the future. 7390 */ 7391 void nohz_balance_enter_idle(int cpu) 7392 { 7393 /* 7394 * If this cpu is going down, then nothing needs to be done. 7395 */ 7396 if (!cpu_active(cpu)) 7397 return; 7398 7399 if (test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu))) 7400 return; 7401 7402 /* 7403 * If we're a completely isolated CPU, we don't play. 7404 */ 7405 if (on_null_domain(cpu_rq(cpu))) 7406 return; 7407 7408 cpumask_set_cpu(cpu, nohz.idle_cpus_mask); 7409 atomic_inc(&nohz.nr_cpus); 7410 set_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)); 7411 } 7412 7413 static int sched_ilb_notifier(struct notifier_block *nfb, 7414 unsigned long action, void *hcpu) 7415 { 7416 switch (action & ~CPU_TASKS_FROZEN) { 7417 case CPU_DYING: 7418 nohz_balance_exit_idle(smp_processor_id()); 7419 return NOTIFY_OK; 7420 default: 7421 return NOTIFY_DONE; 7422 } 7423 } 7424 #endif 7425 7426 static DEFINE_SPINLOCK(balancing); 7427 7428 /* 7429 * Scale the max load_balance interval with the number of CPUs in the system. 7430 * This trades load-balance latency on larger machines for less cross talk. 7431 */ 7432 void update_max_interval(void) 7433 { 7434 max_load_balance_interval = HZ*num_online_cpus()/10; 7435 } 7436 7437 /* 7438 * It checks each scheduling domain to see if it is due to be balanced, 7439 * and initiates a balancing operation if so. 7440 * 7441 * Balancing parameters are set up in init_sched_domains. 7442 */ 7443 static void rebalance_domains(struct rq *rq, enum cpu_idle_type idle) 7444 { 7445 int continue_balancing = 1; 7446 int cpu = rq->cpu; 7447 unsigned long interval; 7448 struct sched_domain *sd; 7449 /* Earliest time when we have to do rebalance again */ 7450 unsigned long next_balance = jiffies + 60*HZ; 7451 int update_next_balance = 0; 7452 int need_serialize, need_decay = 0; 7453 u64 max_cost = 0; 7454 7455 update_blocked_averages(cpu); 7456 7457 rcu_read_lock(); 7458 for_each_domain(cpu, sd) { 7459 /* 7460 * Decay the newidle max times here because this is a regular 7461 * visit to all the domains. Decay ~1% per second. 7462 */ 7463 if (time_after(jiffies, sd->next_decay_max_lb_cost)) { 7464 sd->max_newidle_lb_cost = 7465 (sd->max_newidle_lb_cost * 253) / 256; 7466 sd->next_decay_max_lb_cost = jiffies + HZ; 7467 need_decay = 1; 7468 } 7469 max_cost += sd->max_newidle_lb_cost; 7470 7471 if (!(sd->flags & SD_LOAD_BALANCE)) 7472 continue; 7473 7474 /* 7475 * Stop the load balance at this level. There is another 7476 * CPU in our sched group which is doing load balancing more 7477 * actively. 7478 */ 7479 if (!continue_balancing) { 7480 if (need_decay) 7481 continue; 7482 break; 7483 } 7484 7485 interval = get_sd_balance_interval(sd, idle != CPU_IDLE); 7486 7487 need_serialize = sd->flags & SD_SERIALIZE; 7488 if (need_serialize) { 7489 if (!spin_trylock(&balancing)) 7490 goto out; 7491 } 7492 7493 if (time_after_eq(jiffies, sd->last_balance + interval)) { 7494 if (load_balance(cpu, rq, sd, idle, &continue_balancing)) { 7495 /* 7496 * The LBF_DST_PINNED logic could have changed 7497 * env->dst_cpu, so we can't know our idle 7498 * state even if we migrated tasks. Update it. 7499 */ 7500 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE; 7501 } 7502 sd->last_balance = jiffies; 7503 interval = get_sd_balance_interval(sd, idle != CPU_IDLE); 7504 } 7505 if (need_serialize) 7506 spin_unlock(&balancing); 7507 out: 7508 if (time_after(next_balance, sd->last_balance + interval)) { 7509 next_balance = sd->last_balance + interval; 7510 update_next_balance = 1; 7511 } 7512 } 7513 if (need_decay) { 7514 /* 7515 * Ensure the rq-wide value also decays but keep it at a 7516 * reasonable floor to avoid funnies with rq->avg_idle. 7517 */ 7518 rq->max_idle_balance_cost = 7519 max((u64)sysctl_sched_migration_cost, max_cost); 7520 } 7521 rcu_read_unlock(); 7522 7523 /* 7524 * next_balance will be updated only when there is a need. 7525 * When the cpu is attached to null domain for ex, it will not be 7526 * updated. 7527 */ 7528 if (likely(update_next_balance)) 7529 rq->next_balance = next_balance; 7530 } 7531 7532 #ifdef CONFIG_NO_HZ_COMMON 7533 /* 7534 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the 7535 * rebalancing for all the cpus for whom scheduler ticks are stopped. 7536 */ 7537 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle) 7538 { 7539 int this_cpu = this_rq->cpu; 7540 struct rq *rq; 7541 int balance_cpu; 7542 7543 if (idle != CPU_IDLE || 7544 !test_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu))) 7545 goto end; 7546 7547 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) { 7548 if (balance_cpu == this_cpu || !idle_cpu(balance_cpu)) 7549 continue; 7550 7551 /* 7552 * If this cpu gets work to do, stop the load balancing 7553 * work being done for other cpus. Next load 7554 * balancing owner will pick it up. 7555 */ 7556 if (need_resched()) 7557 break; 7558 7559 rq = cpu_rq(balance_cpu); 7560 7561 /* 7562 * If time for next balance is due, 7563 * do the balance. 7564 */ 7565 if (time_after_eq(jiffies, rq->next_balance)) { 7566 raw_spin_lock_irq(&rq->lock); 7567 update_rq_clock(rq); 7568 update_idle_cpu_load(rq); 7569 raw_spin_unlock_irq(&rq->lock); 7570 rebalance_domains(rq, CPU_IDLE); 7571 } 7572 7573 if (time_after(this_rq->next_balance, rq->next_balance)) 7574 this_rq->next_balance = rq->next_balance; 7575 } 7576 nohz.next_balance = this_rq->next_balance; 7577 end: 7578 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu)); 7579 } 7580 7581 /* 7582 * Current heuristic for kicking the idle load balancer in the presence 7583 * of an idle cpu is the system. 7584 * - This rq has more than one task. 7585 * - At any scheduler domain level, this cpu's scheduler group has multiple 7586 * busy cpu's exceeding the group's capacity. 7587 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler 7588 * domain span are idle. 7589 */ 7590 static inline int nohz_kick_needed(struct rq *rq) 7591 { 7592 unsigned long now = jiffies; 7593 struct sched_domain *sd; 7594 struct sched_group_capacity *sgc; 7595 int nr_busy, cpu = rq->cpu; 7596 7597 if (unlikely(rq->idle_balance)) 7598 return 0; 7599 7600 /* 7601 * We may be recently in ticked or tickless idle mode. At the first 7602 * busy tick after returning from idle, we will update the busy stats. 7603 */ 7604 set_cpu_sd_state_busy(); 7605 nohz_balance_exit_idle(cpu); 7606 7607 /* 7608 * None are in tickless mode and hence no need for NOHZ idle load 7609 * balancing. 7610 */ 7611 if (likely(!atomic_read(&nohz.nr_cpus))) 7612 return 0; 7613 7614 if (time_before(now, nohz.next_balance)) 7615 return 0; 7616 7617 if (rq->nr_running >= 2) 7618 goto need_kick; 7619 7620 rcu_read_lock(); 7621 sd = rcu_dereference(per_cpu(sd_busy, cpu)); 7622 7623 if (sd) { 7624 sgc = sd->groups->sgc; 7625 nr_busy = atomic_read(&sgc->nr_busy_cpus); 7626 7627 if (nr_busy > 1) 7628 goto need_kick_unlock; 7629 } 7630 7631 sd = rcu_dereference(per_cpu(sd_asym, cpu)); 7632 7633 if (sd && (cpumask_first_and(nohz.idle_cpus_mask, 7634 sched_domain_span(sd)) < cpu)) 7635 goto need_kick_unlock; 7636 7637 rcu_read_unlock(); 7638 return 0; 7639 7640 need_kick_unlock: 7641 rcu_read_unlock(); 7642 need_kick: 7643 return 1; 7644 } 7645 #else 7646 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle) { } 7647 #endif 7648 7649 /* 7650 * run_rebalance_domains is triggered when needed from the scheduler tick. 7651 * Also triggered for nohz idle balancing (with nohz_balancing_kick set). 7652 */ 7653 static void run_rebalance_domains(struct softirq_action *h) 7654 { 7655 struct rq *this_rq = this_rq(); 7656 enum cpu_idle_type idle = this_rq->idle_balance ? 7657 CPU_IDLE : CPU_NOT_IDLE; 7658 7659 rebalance_domains(this_rq, idle); 7660 7661 /* 7662 * If this cpu has a pending nohz_balance_kick, then do the 7663 * balancing on behalf of the other idle cpus whose ticks are 7664 * stopped. 7665 */ 7666 nohz_idle_balance(this_rq, idle); 7667 } 7668 7669 /* 7670 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing. 7671 */ 7672 void trigger_load_balance(struct rq *rq) 7673 { 7674 /* Don't need to rebalance while attached to NULL domain */ 7675 if (unlikely(on_null_domain(rq))) 7676 return; 7677 7678 if (time_after_eq(jiffies, rq->next_balance)) 7679 raise_softirq(SCHED_SOFTIRQ); 7680 #ifdef CONFIG_NO_HZ_COMMON 7681 if (nohz_kick_needed(rq)) 7682 nohz_balancer_kick(); 7683 #endif 7684 } 7685 7686 static void rq_online_fair(struct rq *rq) 7687 { 7688 update_sysctl(); 7689 7690 update_runtime_enabled(rq); 7691 } 7692 7693 static void rq_offline_fair(struct rq *rq) 7694 { 7695 update_sysctl(); 7696 7697 /* Ensure any throttled groups are reachable by pick_next_task */ 7698 unthrottle_offline_cfs_rqs(rq); 7699 } 7700 7701 #endif /* CONFIG_SMP */ 7702 7703 /* 7704 * scheduler tick hitting a task of our scheduling class: 7705 */ 7706 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued) 7707 { 7708 struct cfs_rq *cfs_rq; 7709 struct sched_entity *se = &curr->se; 7710 7711 for_each_sched_entity(se) { 7712 cfs_rq = cfs_rq_of(se); 7713 entity_tick(cfs_rq, se, queued); 7714 } 7715 7716 if (numabalancing_enabled) 7717 task_tick_numa(rq, curr); 7718 7719 update_rq_runnable_avg(rq, 1); 7720 } 7721 7722 /* 7723 * called on fork with the child task as argument from the parent's context 7724 * - child not yet on the tasklist 7725 * - preemption disabled 7726 */ 7727 static void task_fork_fair(struct task_struct *p) 7728 { 7729 struct cfs_rq *cfs_rq; 7730 struct sched_entity *se = &p->se, *curr; 7731 int this_cpu = smp_processor_id(); 7732 struct rq *rq = this_rq(); 7733 unsigned long flags; 7734 7735 raw_spin_lock_irqsave(&rq->lock, flags); 7736 7737 update_rq_clock(rq); 7738 7739 cfs_rq = task_cfs_rq(current); 7740 curr = cfs_rq->curr; 7741 7742 /* 7743 * Not only the cpu but also the task_group of the parent might have 7744 * been changed after parent->se.parent,cfs_rq were copied to 7745 * child->se.parent,cfs_rq. So call __set_task_cpu() to make those 7746 * of child point to valid ones. 7747 */ 7748 rcu_read_lock(); 7749 __set_task_cpu(p, this_cpu); 7750 rcu_read_unlock(); 7751 7752 update_curr(cfs_rq); 7753 7754 if (curr) 7755 se->vruntime = curr->vruntime; 7756 place_entity(cfs_rq, se, 1); 7757 7758 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) { 7759 /* 7760 * Upon rescheduling, sched_class::put_prev_task() will place 7761 * 'current' within the tree based on its new key value. 7762 */ 7763 swap(curr->vruntime, se->vruntime); 7764 resched_curr(rq); 7765 } 7766 7767 se->vruntime -= cfs_rq->min_vruntime; 7768 7769 raw_spin_unlock_irqrestore(&rq->lock, flags); 7770 } 7771 7772 /* 7773 * Priority of the task has changed. Check to see if we preempt 7774 * the current task. 7775 */ 7776 static void 7777 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio) 7778 { 7779 if (!task_on_rq_queued(p)) 7780 return; 7781 7782 /* 7783 * Reschedule if we are currently running on this runqueue and 7784 * our priority decreased, or if we are not currently running on 7785 * this runqueue and our priority is higher than the current's 7786 */ 7787 if (rq->curr == p) { 7788 if (p->prio > oldprio) 7789 resched_curr(rq); 7790 } else 7791 check_preempt_curr(rq, p, 0); 7792 } 7793 7794 static void switched_from_fair(struct rq *rq, struct task_struct *p) 7795 { 7796 struct sched_entity *se = &p->se; 7797 struct cfs_rq *cfs_rq = cfs_rq_of(se); 7798 7799 /* 7800 * Ensure the task's vruntime is normalized, so that when it's 7801 * switched back to the fair class the enqueue_entity(.flags=0) will 7802 * do the right thing. 7803 * 7804 * If it's queued, then the dequeue_entity(.flags=0) will already 7805 * have normalized the vruntime, if it's !queued, then only when 7806 * the task is sleeping will it still have non-normalized vruntime. 7807 */ 7808 if (!task_on_rq_queued(p) && p->state != TASK_RUNNING) { 7809 /* 7810 * Fix up our vruntime so that the current sleep doesn't 7811 * cause 'unlimited' sleep bonus. 7812 */ 7813 place_entity(cfs_rq, se, 0); 7814 se->vruntime -= cfs_rq->min_vruntime; 7815 } 7816 7817 #ifdef CONFIG_SMP 7818 /* 7819 * Remove our load from contribution when we leave sched_fair 7820 * and ensure we don't carry in an old decay_count if we 7821 * switch back. 7822 */ 7823 if (se->avg.decay_count) { 7824 __synchronize_entity_decay(se); 7825 subtract_blocked_load_contrib(cfs_rq, se->avg.load_avg_contrib); 7826 } 7827 #endif 7828 } 7829 7830 /* 7831 * We switched to the sched_fair class. 7832 */ 7833 static void switched_to_fair(struct rq *rq, struct task_struct *p) 7834 { 7835 #ifdef CONFIG_FAIR_GROUP_SCHED 7836 struct sched_entity *se = &p->se; 7837 /* 7838 * Since the real-depth could have been changed (only FAIR 7839 * class maintain depth value), reset depth properly. 7840 */ 7841 se->depth = se->parent ? se->parent->depth + 1 : 0; 7842 #endif 7843 if (!task_on_rq_queued(p)) 7844 return; 7845 7846 /* 7847 * We were most likely switched from sched_rt, so 7848 * kick off the schedule if running, otherwise just see 7849 * if we can still preempt the current task. 7850 */ 7851 if (rq->curr == p) 7852 resched_curr(rq); 7853 else 7854 check_preempt_curr(rq, p, 0); 7855 } 7856 7857 /* Account for a task changing its policy or group. 7858 * 7859 * This routine is mostly called to set cfs_rq->curr field when a task 7860 * migrates between groups/classes. 7861 */ 7862 static void set_curr_task_fair(struct rq *rq) 7863 { 7864 struct sched_entity *se = &rq->curr->se; 7865 7866 for_each_sched_entity(se) { 7867 struct cfs_rq *cfs_rq = cfs_rq_of(se); 7868 7869 set_next_entity(cfs_rq, se); 7870 /* ensure bandwidth has been allocated on our new cfs_rq */ 7871 account_cfs_rq_runtime(cfs_rq, 0); 7872 } 7873 } 7874 7875 void init_cfs_rq(struct cfs_rq *cfs_rq) 7876 { 7877 cfs_rq->tasks_timeline = RB_ROOT; 7878 cfs_rq->min_vruntime = (u64)(-(1LL << 20)); 7879 #ifndef CONFIG_64BIT 7880 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime; 7881 #endif 7882 #ifdef CONFIG_SMP 7883 atomic64_set(&cfs_rq->decay_counter, 1); 7884 atomic_long_set(&cfs_rq->removed_load, 0); 7885 #endif 7886 } 7887 7888 #ifdef CONFIG_FAIR_GROUP_SCHED 7889 static void task_move_group_fair(struct task_struct *p, int queued) 7890 { 7891 struct sched_entity *se = &p->se; 7892 struct cfs_rq *cfs_rq; 7893 7894 /* 7895 * If the task was not on the rq at the time of this cgroup movement 7896 * it must have been asleep, sleeping tasks keep their ->vruntime 7897 * absolute on their old rq until wakeup (needed for the fair sleeper 7898 * bonus in place_entity()). 7899 * 7900 * If it was on the rq, we've just 'preempted' it, which does convert 7901 * ->vruntime to a relative base. 7902 * 7903 * Make sure both cases convert their relative position when migrating 7904 * to another cgroup's rq. This does somewhat interfere with the 7905 * fair sleeper stuff for the first placement, but who cares. 7906 */ 7907 /* 7908 * When !queued, vruntime of the task has usually NOT been normalized. 7909 * But there are some cases where it has already been normalized: 7910 * 7911 * - Moving a forked child which is waiting for being woken up by 7912 * wake_up_new_task(). 7913 * - Moving a task which has been woken up by try_to_wake_up() and 7914 * waiting for actually being woken up by sched_ttwu_pending(). 7915 * 7916 * To prevent boost or penalty in the new cfs_rq caused by delta 7917 * min_vruntime between the two cfs_rqs, we skip vruntime adjustment. 7918 */ 7919 if (!queued && (!se->sum_exec_runtime || p->state == TASK_WAKING)) 7920 queued = 1; 7921 7922 if (!queued) 7923 se->vruntime -= cfs_rq_of(se)->min_vruntime; 7924 set_task_rq(p, task_cpu(p)); 7925 se->depth = se->parent ? se->parent->depth + 1 : 0; 7926 if (!queued) { 7927 cfs_rq = cfs_rq_of(se); 7928 se->vruntime += cfs_rq->min_vruntime; 7929 #ifdef CONFIG_SMP 7930 /* 7931 * migrate_task_rq_fair() will have removed our previous 7932 * contribution, but we must synchronize for ongoing future 7933 * decay. 7934 */ 7935 se->avg.decay_count = atomic64_read(&cfs_rq->decay_counter); 7936 cfs_rq->blocked_load_avg += se->avg.load_avg_contrib; 7937 #endif 7938 } 7939 } 7940 7941 void free_fair_sched_group(struct task_group *tg) 7942 { 7943 int i; 7944 7945 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg)); 7946 7947 for_each_possible_cpu(i) { 7948 if (tg->cfs_rq) 7949 kfree(tg->cfs_rq[i]); 7950 if (tg->se) 7951 kfree(tg->se[i]); 7952 } 7953 7954 kfree(tg->cfs_rq); 7955 kfree(tg->se); 7956 } 7957 7958 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent) 7959 { 7960 struct cfs_rq *cfs_rq; 7961 struct sched_entity *se; 7962 int i; 7963 7964 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL); 7965 if (!tg->cfs_rq) 7966 goto err; 7967 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL); 7968 if (!tg->se) 7969 goto err; 7970 7971 tg->shares = NICE_0_LOAD; 7972 7973 init_cfs_bandwidth(tg_cfs_bandwidth(tg)); 7974 7975 for_each_possible_cpu(i) { 7976 cfs_rq = kzalloc_node(sizeof(struct cfs_rq), 7977 GFP_KERNEL, cpu_to_node(i)); 7978 if (!cfs_rq) 7979 goto err; 7980 7981 se = kzalloc_node(sizeof(struct sched_entity), 7982 GFP_KERNEL, cpu_to_node(i)); 7983 if (!se) 7984 goto err_free_rq; 7985 7986 init_cfs_rq(cfs_rq); 7987 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]); 7988 } 7989 7990 return 1; 7991 7992 err_free_rq: 7993 kfree(cfs_rq); 7994 err: 7995 return 0; 7996 } 7997 7998 void unregister_fair_sched_group(struct task_group *tg, int cpu) 7999 { 8000 struct rq *rq = cpu_rq(cpu); 8001 unsigned long flags; 8002 8003 /* 8004 * Only empty task groups can be destroyed; so we can speculatively 8005 * check on_list without danger of it being re-added. 8006 */ 8007 if (!tg->cfs_rq[cpu]->on_list) 8008 return; 8009 8010 raw_spin_lock_irqsave(&rq->lock, flags); 8011 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]); 8012 raw_spin_unlock_irqrestore(&rq->lock, flags); 8013 } 8014 8015 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq, 8016 struct sched_entity *se, int cpu, 8017 struct sched_entity *parent) 8018 { 8019 struct rq *rq = cpu_rq(cpu); 8020 8021 cfs_rq->tg = tg; 8022 cfs_rq->rq = rq; 8023 init_cfs_rq_runtime(cfs_rq); 8024 8025 tg->cfs_rq[cpu] = cfs_rq; 8026 tg->se[cpu] = se; 8027 8028 /* se could be NULL for root_task_group */ 8029 if (!se) 8030 return; 8031 8032 if (!parent) { 8033 se->cfs_rq = &rq->cfs; 8034 se->depth = 0; 8035 } else { 8036 se->cfs_rq = parent->my_q; 8037 se->depth = parent->depth + 1; 8038 } 8039 8040 se->my_q = cfs_rq; 8041 /* guarantee group entities always have weight */ 8042 update_load_set(&se->load, NICE_0_LOAD); 8043 se->parent = parent; 8044 } 8045 8046 static DEFINE_MUTEX(shares_mutex); 8047 8048 int sched_group_set_shares(struct task_group *tg, unsigned long shares) 8049 { 8050 int i; 8051 unsigned long flags; 8052 8053 /* 8054 * We can't change the weight of the root cgroup. 8055 */ 8056 if (!tg->se[0]) 8057 return -EINVAL; 8058 8059 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES)); 8060 8061 mutex_lock(&shares_mutex); 8062 if (tg->shares == shares) 8063 goto done; 8064 8065 tg->shares = shares; 8066 for_each_possible_cpu(i) { 8067 struct rq *rq = cpu_rq(i); 8068 struct sched_entity *se; 8069 8070 se = tg->se[i]; 8071 /* Propagate contribution to hierarchy */ 8072 raw_spin_lock_irqsave(&rq->lock, flags); 8073 8074 /* Possible calls to update_curr() need rq clock */ 8075 update_rq_clock(rq); 8076 for_each_sched_entity(se) 8077 update_cfs_shares(group_cfs_rq(se)); 8078 raw_spin_unlock_irqrestore(&rq->lock, flags); 8079 } 8080 8081 done: 8082 mutex_unlock(&shares_mutex); 8083 return 0; 8084 } 8085 #else /* CONFIG_FAIR_GROUP_SCHED */ 8086 8087 void free_fair_sched_group(struct task_group *tg) { } 8088 8089 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent) 8090 { 8091 return 1; 8092 } 8093 8094 void unregister_fair_sched_group(struct task_group *tg, int cpu) { } 8095 8096 #endif /* CONFIG_FAIR_GROUP_SCHED */ 8097 8098 8099 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task) 8100 { 8101 struct sched_entity *se = &task->se; 8102 unsigned int rr_interval = 0; 8103 8104 /* 8105 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise 8106 * idle runqueue: 8107 */ 8108 if (rq->cfs.load.weight) 8109 rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se)); 8110 8111 return rr_interval; 8112 } 8113 8114 /* 8115 * All the scheduling class methods: 8116 */ 8117 const struct sched_class fair_sched_class = { 8118 .next = &idle_sched_class, 8119 .enqueue_task = enqueue_task_fair, 8120 .dequeue_task = dequeue_task_fair, 8121 .yield_task = yield_task_fair, 8122 .yield_to_task = yield_to_task_fair, 8123 8124 .check_preempt_curr = check_preempt_wakeup, 8125 8126 .pick_next_task = pick_next_task_fair, 8127 .put_prev_task = put_prev_task_fair, 8128 8129 #ifdef CONFIG_SMP 8130 .select_task_rq = select_task_rq_fair, 8131 .migrate_task_rq = migrate_task_rq_fair, 8132 8133 .rq_online = rq_online_fair, 8134 .rq_offline = rq_offline_fair, 8135 8136 .task_waking = task_waking_fair, 8137 #endif 8138 8139 .set_curr_task = set_curr_task_fair, 8140 .task_tick = task_tick_fair, 8141 .task_fork = task_fork_fair, 8142 8143 .prio_changed = prio_changed_fair, 8144 .switched_from = switched_from_fair, 8145 .switched_to = switched_to_fair, 8146 8147 .get_rr_interval = get_rr_interval_fair, 8148 8149 .update_curr = update_curr_fair, 8150 8151 #ifdef CONFIG_FAIR_GROUP_SCHED 8152 .task_move_group = task_move_group_fair, 8153 #endif 8154 }; 8155 8156 #ifdef CONFIG_SCHED_DEBUG 8157 void print_cfs_stats(struct seq_file *m, int cpu) 8158 { 8159 struct cfs_rq *cfs_rq; 8160 8161 rcu_read_lock(); 8162 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq) 8163 print_cfs_rq(m, cpu, cfs_rq); 8164 rcu_read_unlock(); 8165 } 8166 #endif 8167 8168 __init void init_sched_fair_class(void) 8169 { 8170 #ifdef CONFIG_SMP 8171 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains); 8172 8173 #ifdef CONFIG_NO_HZ_COMMON 8174 nohz.next_balance = jiffies; 8175 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT); 8176 cpu_notifier(sched_ilb_notifier, 0); 8177 #endif 8178 #endif /* SMP */ 8179 8180 } 8181