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 21 */ 22 23 #include <linux/sched.h> 24 #include <linux/latencytop.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 unsigned int get_update_sysctl_factor(void) 145 { 146 unsigned int cpus = min_t(unsigned 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 sched_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 inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq) 287 { 288 if (!cfs_rq->on_list) { 289 /* 290 * Ensure we either appear before our parent (if already 291 * enqueued) or force our parent to appear after us when it is 292 * enqueued. The fact that we always enqueue bottom-up 293 * reduces this to two cases. 294 */ 295 if (cfs_rq->tg->parent && 296 cfs_rq->tg->parent->cfs_rq[cpu_of(rq_of(cfs_rq))]->on_list) { 297 list_add_rcu(&cfs_rq->leaf_cfs_rq_list, 298 &rq_of(cfs_rq)->leaf_cfs_rq_list); 299 } else { 300 list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list, 301 &rq_of(cfs_rq)->leaf_cfs_rq_list); 302 } 303 304 cfs_rq->on_list = 1; 305 } 306 } 307 308 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq) 309 { 310 if (cfs_rq->on_list) { 311 list_del_rcu(&cfs_rq->leaf_cfs_rq_list); 312 cfs_rq->on_list = 0; 313 } 314 } 315 316 /* Iterate thr' all leaf cfs_rq's on a runqueue */ 317 #define for_each_leaf_cfs_rq(rq, cfs_rq) \ 318 list_for_each_entry_rcu(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list) 319 320 /* Do the two (enqueued) entities belong to the same group ? */ 321 static inline struct cfs_rq * 322 is_same_group(struct sched_entity *se, struct sched_entity *pse) 323 { 324 if (se->cfs_rq == pse->cfs_rq) 325 return se->cfs_rq; 326 327 return NULL; 328 } 329 330 static inline struct sched_entity *parent_entity(struct sched_entity *se) 331 { 332 return se->parent; 333 } 334 335 static void 336 find_matching_se(struct sched_entity **se, struct sched_entity **pse) 337 { 338 int se_depth, pse_depth; 339 340 /* 341 * preemption test can be made between sibling entities who are in the 342 * same cfs_rq i.e who have a common parent. Walk up the hierarchy of 343 * both tasks until we find their ancestors who are siblings of common 344 * parent. 345 */ 346 347 /* First walk up until both entities are at same depth */ 348 se_depth = (*se)->depth; 349 pse_depth = (*pse)->depth; 350 351 while (se_depth > pse_depth) { 352 se_depth--; 353 *se = parent_entity(*se); 354 } 355 356 while (pse_depth > se_depth) { 357 pse_depth--; 358 *pse = parent_entity(*pse); 359 } 360 361 while (!is_same_group(*se, *pse)) { 362 *se = parent_entity(*se); 363 *pse = parent_entity(*pse); 364 } 365 } 366 367 #else /* !CONFIG_FAIR_GROUP_SCHED */ 368 369 static inline struct task_struct *task_of(struct sched_entity *se) 370 { 371 return container_of(se, struct task_struct, se); 372 } 373 374 static inline struct rq *rq_of(struct cfs_rq *cfs_rq) 375 { 376 return container_of(cfs_rq, struct rq, cfs); 377 } 378 379 #define entity_is_task(se) 1 380 381 #define for_each_sched_entity(se) \ 382 for (; se; se = NULL) 383 384 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p) 385 { 386 return &task_rq(p)->cfs; 387 } 388 389 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se) 390 { 391 struct task_struct *p = task_of(se); 392 struct rq *rq = task_rq(p); 393 394 return &rq->cfs; 395 } 396 397 /* runqueue "owned" by this group */ 398 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp) 399 { 400 return NULL; 401 } 402 403 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq) 404 { 405 } 406 407 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq) 408 { 409 } 410 411 #define for_each_leaf_cfs_rq(rq, cfs_rq) \ 412 for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL) 413 414 static inline struct sched_entity *parent_entity(struct sched_entity *se) 415 { 416 return NULL; 417 } 418 419 static inline void 420 find_matching_se(struct sched_entity **se, struct sched_entity **pse) 421 { 422 } 423 424 #endif /* CONFIG_FAIR_GROUP_SCHED */ 425 426 static __always_inline 427 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec); 428 429 /************************************************************** 430 * Scheduling class tree data structure manipulation methods: 431 */ 432 433 static inline u64 max_vruntime(u64 max_vruntime, u64 vruntime) 434 { 435 s64 delta = (s64)(vruntime - max_vruntime); 436 if (delta > 0) 437 max_vruntime = vruntime; 438 439 return max_vruntime; 440 } 441 442 static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime) 443 { 444 s64 delta = (s64)(vruntime - min_vruntime); 445 if (delta < 0) 446 min_vruntime = vruntime; 447 448 return min_vruntime; 449 } 450 451 static inline int entity_before(struct sched_entity *a, 452 struct sched_entity *b) 453 { 454 return (s64)(a->vruntime - b->vruntime) < 0; 455 } 456 457 static void update_min_vruntime(struct cfs_rq *cfs_rq) 458 { 459 struct sched_entity *curr = cfs_rq->curr; 460 461 u64 vruntime = cfs_rq->min_vruntime; 462 463 if (curr) { 464 if (curr->on_rq) 465 vruntime = curr->vruntime; 466 else 467 curr = NULL; 468 } 469 470 if (cfs_rq->rb_leftmost) { 471 struct sched_entity *se = rb_entry(cfs_rq->rb_leftmost, 472 struct sched_entity, 473 run_node); 474 475 if (!curr) 476 vruntime = se->vruntime; 477 else 478 vruntime = min_vruntime(vruntime, se->vruntime); 479 } 480 481 /* ensure we never gain time by being placed backwards. */ 482 cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime); 483 #ifndef CONFIG_64BIT 484 smp_wmb(); 485 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime; 486 #endif 487 } 488 489 /* 490 * Enqueue an entity into the rb-tree: 491 */ 492 static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se) 493 { 494 struct rb_node **link = &cfs_rq->tasks_timeline.rb_node; 495 struct rb_node *parent = NULL; 496 struct sched_entity *entry; 497 int leftmost = 1; 498 499 /* 500 * Find the right place in the rbtree: 501 */ 502 while (*link) { 503 parent = *link; 504 entry = rb_entry(parent, struct sched_entity, run_node); 505 /* 506 * We dont care about collisions. Nodes with 507 * the same key stay together. 508 */ 509 if (entity_before(se, entry)) { 510 link = &parent->rb_left; 511 } else { 512 link = &parent->rb_right; 513 leftmost = 0; 514 } 515 } 516 517 /* 518 * Maintain a cache of leftmost tree entries (it is frequently 519 * used): 520 */ 521 if (leftmost) 522 cfs_rq->rb_leftmost = &se->run_node; 523 524 rb_link_node(&se->run_node, parent, link); 525 rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline); 526 } 527 528 static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se) 529 { 530 if (cfs_rq->rb_leftmost == &se->run_node) { 531 struct rb_node *next_node; 532 533 next_node = rb_next(&se->run_node); 534 cfs_rq->rb_leftmost = next_node; 535 } 536 537 rb_erase(&se->run_node, &cfs_rq->tasks_timeline); 538 } 539 540 struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq) 541 { 542 struct rb_node *left = cfs_rq->rb_leftmost; 543 544 if (!left) 545 return NULL; 546 547 return rb_entry(left, struct sched_entity, run_node); 548 } 549 550 static struct sched_entity *__pick_next_entity(struct sched_entity *se) 551 { 552 struct rb_node *next = rb_next(&se->run_node); 553 554 if (!next) 555 return NULL; 556 557 return rb_entry(next, struct sched_entity, run_node); 558 } 559 560 #ifdef CONFIG_SCHED_DEBUG 561 struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq) 562 { 563 struct rb_node *last = rb_last(&cfs_rq->tasks_timeline); 564 565 if (!last) 566 return NULL; 567 568 return rb_entry(last, struct sched_entity, run_node); 569 } 570 571 /************************************************************** 572 * Scheduling class statistics methods: 573 */ 574 575 int sched_proc_update_handler(struct ctl_table *table, int write, 576 void __user *buffer, size_t *lenp, 577 loff_t *ppos) 578 { 579 int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos); 580 unsigned int factor = get_update_sysctl_factor(); 581 582 if (ret || !write) 583 return ret; 584 585 sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency, 586 sysctl_sched_min_granularity); 587 588 #define WRT_SYSCTL(name) \ 589 (normalized_sysctl_##name = sysctl_##name / (factor)) 590 WRT_SYSCTL(sched_min_granularity); 591 WRT_SYSCTL(sched_latency); 592 WRT_SYSCTL(sched_wakeup_granularity); 593 #undef WRT_SYSCTL 594 595 return 0; 596 } 597 #endif 598 599 /* 600 * delta /= w 601 */ 602 static inline u64 calc_delta_fair(u64 delta, struct sched_entity *se) 603 { 604 if (unlikely(se->load.weight != NICE_0_LOAD)) 605 delta = __calc_delta(delta, NICE_0_LOAD, &se->load); 606 607 return delta; 608 } 609 610 /* 611 * The idea is to set a period in which each task runs once. 612 * 613 * When there are too many tasks (sched_nr_latency) we have to stretch 614 * this period because otherwise the slices get too small. 615 * 616 * p = (nr <= nl) ? l : l*nr/nl 617 */ 618 static u64 __sched_period(unsigned long nr_running) 619 { 620 if (unlikely(nr_running > sched_nr_latency)) 621 return nr_running * sysctl_sched_min_granularity; 622 else 623 return sysctl_sched_latency; 624 } 625 626 /* 627 * We calculate the wall-time slice from the period by taking a part 628 * proportional to the weight. 629 * 630 * s = p*P[w/rw] 631 */ 632 static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se) 633 { 634 u64 slice = __sched_period(cfs_rq->nr_running + !se->on_rq); 635 636 for_each_sched_entity(se) { 637 struct load_weight *load; 638 struct load_weight lw; 639 640 cfs_rq = cfs_rq_of(se); 641 load = &cfs_rq->load; 642 643 if (unlikely(!se->on_rq)) { 644 lw = cfs_rq->load; 645 646 update_load_add(&lw, se->load.weight); 647 load = &lw; 648 } 649 slice = __calc_delta(slice, se->load.weight, load); 650 } 651 return slice; 652 } 653 654 /* 655 * We calculate the vruntime slice of a to-be-inserted task. 656 * 657 * vs = s/w 658 */ 659 static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se) 660 { 661 return calc_delta_fair(sched_slice(cfs_rq, se), se); 662 } 663 664 #ifdef CONFIG_SMP 665 static int select_idle_sibling(struct task_struct *p, int cpu); 666 static unsigned long task_h_load(struct task_struct *p); 667 668 /* 669 * We choose a half-life close to 1 scheduling period. 670 * Note: The tables runnable_avg_yN_inv and runnable_avg_yN_sum are 671 * dependent on this value. 672 */ 673 #define LOAD_AVG_PERIOD 32 674 #define LOAD_AVG_MAX 47742 /* maximum possible load avg */ 675 #define LOAD_AVG_MAX_N 345 /* number of full periods to produce LOAD_AVG_MAX */ 676 677 /* Give new sched_entity start runnable values to heavy its load in infant time */ 678 void init_entity_runnable_average(struct sched_entity *se) 679 { 680 struct sched_avg *sa = &se->avg; 681 682 sa->last_update_time = 0; 683 /* 684 * sched_avg's period_contrib should be strictly less then 1024, so 685 * we give it 1023 to make sure it is almost a period (1024us), and 686 * will definitely be update (after enqueue). 687 */ 688 sa->period_contrib = 1023; 689 /* 690 * Tasks are intialized with full load to be seen as heavy tasks until 691 * they get a chance to stabilize to their real load level. 692 * Group entities are intialized with zero load to reflect the fact that 693 * nothing has been attached to the task group yet. 694 */ 695 if (entity_is_task(se)) 696 sa->load_avg = scale_load_down(se->load.weight); 697 sa->load_sum = sa->load_avg * LOAD_AVG_MAX; 698 /* 699 * At this point, util_avg won't be used in select_task_rq_fair anyway 700 */ 701 sa->util_avg = 0; 702 sa->util_sum = 0; 703 /* when this task enqueue'ed, it will contribute to its cfs_rq's load_avg */ 704 } 705 706 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq); 707 static int update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq, bool update_freq); 708 static void update_tg_load_avg(struct cfs_rq *cfs_rq, int force); 709 static void attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se); 710 711 /* 712 * With new tasks being created, their initial util_avgs are extrapolated 713 * based on the cfs_rq's current util_avg: 714 * 715 * util_avg = cfs_rq->util_avg / (cfs_rq->load_avg + 1) * se.load.weight 716 * 717 * However, in many cases, the above util_avg does not give a desired 718 * value. Moreover, the sum of the util_avgs may be divergent, such 719 * as when the series is a harmonic series. 720 * 721 * To solve this problem, we also cap the util_avg of successive tasks to 722 * only 1/2 of the left utilization budget: 723 * 724 * util_avg_cap = (1024 - cfs_rq->avg.util_avg) / 2^n 725 * 726 * where n denotes the nth task. 727 * 728 * For example, a simplest series from the beginning would be like: 729 * 730 * task util_avg: 512, 256, 128, 64, 32, 16, 8, ... 731 * cfs_rq util_avg: 512, 768, 896, 960, 992, 1008, 1016, ... 732 * 733 * Finally, that extrapolated util_avg is clamped to the cap (util_avg_cap) 734 * if util_avg > util_avg_cap. 735 */ 736 void post_init_entity_util_avg(struct sched_entity *se) 737 { 738 struct cfs_rq *cfs_rq = cfs_rq_of(se); 739 struct sched_avg *sa = &se->avg; 740 long cap = (long)(SCHED_CAPACITY_SCALE - cfs_rq->avg.util_avg) / 2; 741 u64 now = cfs_rq_clock_task(cfs_rq); 742 int tg_update; 743 744 if (cap > 0) { 745 if (cfs_rq->avg.util_avg != 0) { 746 sa->util_avg = cfs_rq->avg.util_avg * se->load.weight; 747 sa->util_avg /= (cfs_rq->avg.load_avg + 1); 748 749 if (sa->util_avg > cap) 750 sa->util_avg = cap; 751 } else { 752 sa->util_avg = cap; 753 } 754 sa->util_sum = sa->util_avg * LOAD_AVG_MAX; 755 } 756 757 if (entity_is_task(se)) { 758 struct task_struct *p = task_of(se); 759 if (p->sched_class != &fair_sched_class) { 760 /* 761 * For !fair tasks do: 762 * 763 update_cfs_rq_load_avg(now, cfs_rq, false); 764 attach_entity_load_avg(cfs_rq, se); 765 switched_from_fair(rq, p); 766 * 767 * such that the next switched_to_fair() has the 768 * expected state. 769 */ 770 se->avg.last_update_time = now; 771 return; 772 } 773 } 774 775 tg_update = update_cfs_rq_load_avg(now, cfs_rq, false); 776 attach_entity_load_avg(cfs_rq, se); 777 if (tg_update) 778 update_tg_load_avg(cfs_rq, false); 779 } 780 781 #else /* !CONFIG_SMP */ 782 void init_entity_runnable_average(struct sched_entity *se) 783 { 784 } 785 void post_init_entity_util_avg(struct sched_entity *se) 786 { 787 } 788 static void update_tg_load_avg(struct cfs_rq *cfs_rq, int force) 789 { 790 } 791 #endif /* CONFIG_SMP */ 792 793 /* 794 * Update the current task's runtime statistics. 795 */ 796 static void update_curr(struct cfs_rq *cfs_rq) 797 { 798 struct sched_entity *curr = cfs_rq->curr; 799 u64 now = rq_clock_task(rq_of(cfs_rq)); 800 u64 delta_exec; 801 802 if (unlikely(!curr)) 803 return; 804 805 delta_exec = now - curr->exec_start; 806 if (unlikely((s64)delta_exec <= 0)) 807 return; 808 809 curr->exec_start = now; 810 811 schedstat_set(curr->statistics.exec_max, 812 max(delta_exec, curr->statistics.exec_max)); 813 814 curr->sum_exec_runtime += delta_exec; 815 schedstat_add(cfs_rq, exec_clock, delta_exec); 816 817 curr->vruntime += calc_delta_fair(delta_exec, curr); 818 update_min_vruntime(cfs_rq); 819 820 if (entity_is_task(curr)) { 821 struct task_struct *curtask = task_of(curr); 822 823 trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime); 824 cpuacct_charge(curtask, delta_exec); 825 account_group_exec_runtime(curtask, delta_exec); 826 } 827 828 account_cfs_rq_runtime(cfs_rq, delta_exec); 829 } 830 831 static void update_curr_fair(struct rq *rq) 832 { 833 update_curr(cfs_rq_of(&rq->curr->se)); 834 } 835 836 #ifdef CONFIG_SCHEDSTATS 837 static inline void 838 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se) 839 { 840 u64 wait_start = rq_clock(rq_of(cfs_rq)); 841 842 if (entity_is_task(se) && task_on_rq_migrating(task_of(se)) && 843 likely(wait_start > se->statistics.wait_start)) 844 wait_start -= se->statistics.wait_start; 845 846 se->statistics.wait_start = wait_start; 847 } 848 849 static void 850 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se) 851 { 852 struct task_struct *p; 853 u64 delta; 854 855 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start; 856 857 if (entity_is_task(se)) { 858 p = task_of(se); 859 if (task_on_rq_migrating(p)) { 860 /* 861 * Preserve migrating task's wait time so wait_start 862 * time stamp can be adjusted to accumulate wait time 863 * prior to migration. 864 */ 865 se->statistics.wait_start = delta; 866 return; 867 } 868 trace_sched_stat_wait(p, delta); 869 } 870 871 se->statistics.wait_max = max(se->statistics.wait_max, delta); 872 se->statistics.wait_count++; 873 se->statistics.wait_sum += delta; 874 se->statistics.wait_start = 0; 875 } 876 877 /* 878 * Task is being enqueued - update stats: 879 */ 880 static inline void 881 update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se) 882 { 883 /* 884 * Are we enqueueing a waiting task? (for current tasks 885 * a dequeue/enqueue event is a NOP) 886 */ 887 if (se != cfs_rq->curr) 888 update_stats_wait_start(cfs_rq, se); 889 } 890 891 static inline void 892 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags) 893 { 894 /* 895 * Mark the end of the wait period if dequeueing a 896 * waiting task: 897 */ 898 if (se != cfs_rq->curr) 899 update_stats_wait_end(cfs_rq, se); 900 901 if (flags & DEQUEUE_SLEEP) { 902 if (entity_is_task(se)) { 903 struct task_struct *tsk = task_of(se); 904 905 if (tsk->state & TASK_INTERRUPTIBLE) 906 se->statistics.sleep_start = rq_clock(rq_of(cfs_rq)); 907 if (tsk->state & TASK_UNINTERRUPTIBLE) 908 se->statistics.block_start = rq_clock(rq_of(cfs_rq)); 909 } 910 } 911 912 } 913 #else 914 static inline void 915 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se) 916 { 917 } 918 919 static inline void 920 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se) 921 { 922 } 923 924 static inline void 925 update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se) 926 { 927 } 928 929 static inline void 930 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags) 931 { 932 } 933 #endif 934 935 /* 936 * We are picking a new current task - update its stats: 937 */ 938 static inline void 939 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se) 940 { 941 /* 942 * We are starting a new run period: 943 */ 944 se->exec_start = rq_clock_task(rq_of(cfs_rq)); 945 } 946 947 /************************************************** 948 * Scheduling class queueing methods: 949 */ 950 951 #ifdef CONFIG_NUMA_BALANCING 952 /* 953 * Approximate time to scan a full NUMA task in ms. The task scan period is 954 * calculated based on the tasks virtual memory size and 955 * numa_balancing_scan_size. 956 */ 957 unsigned int sysctl_numa_balancing_scan_period_min = 1000; 958 unsigned int sysctl_numa_balancing_scan_period_max = 60000; 959 960 /* Portion of address space to scan in MB */ 961 unsigned int sysctl_numa_balancing_scan_size = 256; 962 963 /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */ 964 unsigned int sysctl_numa_balancing_scan_delay = 1000; 965 966 static unsigned int task_nr_scan_windows(struct task_struct *p) 967 { 968 unsigned long rss = 0; 969 unsigned long nr_scan_pages; 970 971 /* 972 * Calculations based on RSS as non-present and empty pages are skipped 973 * by the PTE scanner and NUMA hinting faults should be trapped based 974 * on resident pages 975 */ 976 nr_scan_pages = sysctl_numa_balancing_scan_size << (20 - PAGE_SHIFT); 977 rss = get_mm_rss(p->mm); 978 if (!rss) 979 rss = nr_scan_pages; 980 981 rss = round_up(rss, nr_scan_pages); 982 return rss / nr_scan_pages; 983 } 984 985 /* For sanitys sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */ 986 #define MAX_SCAN_WINDOW 2560 987 988 static unsigned int task_scan_min(struct task_struct *p) 989 { 990 unsigned int scan_size = READ_ONCE(sysctl_numa_balancing_scan_size); 991 unsigned int scan, floor; 992 unsigned int windows = 1; 993 994 if (scan_size < MAX_SCAN_WINDOW) 995 windows = MAX_SCAN_WINDOW / scan_size; 996 floor = 1000 / windows; 997 998 scan = sysctl_numa_balancing_scan_period_min / task_nr_scan_windows(p); 999 return max_t(unsigned int, floor, scan); 1000 } 1001 1002 static unsigned int task_scan_max(struct task_struct *p) 1003 { 1004 unsigned int smin = task_scan_min(p); 1005 unsigned int smax; 1006 1007 /* Watch for min being lower than max due to floor calculations */ 1008 smax = sysctl_numa_balancing_scan_period_max / task_nr_scan_windows(p); 1009 return max(smin, smax); 1010 } 1011 1012 static void account_numa_enqueue(struct rq *rq, struct task_struct *p) 1013 { 1014 rq->nr_numa_running += (p->numa_preferred_nid != -1); 1015 rq->nr_preferred_running += (p->numa_preferred_nid == task_node(p)); 1016 } 1017 1018 static void account_numa_dequeue(struct rq *rq, struct task_struct *p) 1019 { 1020 rq->nr_numa_running -= (p->numa_preferred_nid != -1); 1021 rq->nr_preferred_running -= (p->numa_preferred_nid == task_node(p)); 1022 } 1023 1024 struct numa_group { 1025 atomic_t refcount; 1026 1027 spinlock_t lock; /* nr_tasks, tasks */ 1028 int nr_tasks; 1029 pid_t gid; 1030 int active_nodes; 1031 1032 struct rcu_head rcu; 1033 unsigned long total_faults; 1034 unsigned long max_faults_cpu; 1035 /* 1036 * Faults_cpu is used to decide whether memory should move 1037 * towards the CPU. As a consequence, these stats are weighted 1038 * more by CPU use than by memory faults. 1039 */ 1040 unsigned long *faults_cpu; 1041 unsigned long faults[0]; 1042 }; 1043 1044 /* Shared or private faults. */ 1045 #define NR_NUMA_HINT_FAULT_TYPES 2 1046 1047 /* Memory and CPU locality */ 1048 #define NR_NUMA_HINT_FAULT_STATS (NR_NUMA_HINT_FAULT_TYPES * 2) 1049 1050 /* Averaged statistics, and temporary buffers. */ 1051 #define NR_NUMA_HINT_FAULT_BUCKETS (NR_NUMA_HINT_FAULT_STATS * 2) 1052 1053 pid_t task_numa_group_id(struct task_struct *p) 1054 { 1055 return p->numa_group ? p->numa_group->gid : 0; 1056 } 1057 1058 /* 1059 * The averaged statistics, shared & private, memory & cpu, 1060 * occupy the first half of the array. The second half of the 1061 * array is for current counters, which are averaged into the 1062 * first set by task_numa_placement. 1063 */ 1064 static inline int task_faults_idx(enum numa_faults_stats s, int nid, int priv) 1065 { 1066 return NR_NUMA_HINT_FAULT_TYPES * (s * nr_node_ids + nid) + priv; 1067 } 1068 1069 static inline unsigned long task_faults(struct task_struct *p, int nid) 1070 { 1071 if (!p->numa_faults) 1072 return 0; 1073 1074 return p->numa_faults[task_faults_idx(NUMA_MEM, nid, 0)] + 1075 p->numa_faults[task_faults_idx(NUMA_MEM, nid, 1)]; 1076 } 1077 1078 static inline unsigned long group_faults(struct task_struct *p, int nid) 1079 { 1080 if (!p->numa_group) 1081 return 0; 1082 1083 return p->numa_group->faults[task_faults_idx(NUMA_MEM, nid, 0)] + 1084 p->numa_group->faults[task_faults_idx(NUMA_MEM, nid, 1)]; 1085 } 1086 1087 static inline unsigned long group_faults_cpu(struct numa_group *group, int nid) 1088 { 1089 return group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 0)] + 1090 group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 1)]; 1091 } 1092 1093 /* 1094 * A node triggering more than 1/3 as many NUMA faults as the maximum is 1095 * considered part of a numa group's pseudo-interleaving set. Migrations 1096 * between these nodes are slowed down, to allow things to settle down. 1097 */ 1098 #define ACTIVE_NODE_FRACTION 3 1099 1100 static bool numa_is_active_node(int nid, struct numa_group *ng) 1101 { 1102 return group_faults_cpu(ng, nid) * ACTIVE_NODE_FRACTION > ng->max_faults_cpu; 1103 } 1104 1105 /* Handle placement on systems where not all nodes are directly connected. */ 1106 static unsigned long score_nearby_nodes(struct task_struct *p, int nid, 1107 int maxdist, bool task) 1108 { 1109 unsigned long score = 0; 1110 int node; 1111 1112 /* 1113 * All nodes are directly connected, and the same distance 1114 * from each other. No need for fancy placement algorithms. 1115 */ 1116 if (sched_numa_topology_type == NUMA_DIRECT) 1117 return 0; 1118 1119 /* 1120 * This code is called for each node, introducing N^2 complexity, 1121 * which should be ok given the number of nodes rarely exceeds 8. 1122 */ 1123 for_each_online_node(node) { 1124 unsigned long faults; 1125 int dist = node_distance(nid, node); 1126 1127 /* 1128 * The furthest away nodes in the system are not interesting 1129 * for placement; nid was already counted. 1130 */ 1131 if (dist == sched_max_numa_distance || node == nid) 1132 continue; 1133 1134 /* 1135 * On systems with a backplane NUMA topology, compare groups 1136 * of nodes, and move tasks towards the group with the most 1137 * memory accesses. When comparing two nodes at distance 1138 * "hoplimit", only nodes closer by than "hoplimit" are part 1139 * of each group. Skip other nodes. 1140 */ 1141 if (sched_numa_topology_type == NUMA_BACKPLANE && 1142 dist > maxdist) 1143 continue; 1144 1145 /* Add up the faults from nearby nodes. */ 1146 if (task) 1147 faults = task_faults(p, node); 1148 else 1149 faults = group_faults(p, node); 1150 1151 /* 1152 * On systems with a glueless mesh NUMA topology, there are 1153 * no fixed "groups of nodes". Instead, nodes that are not 1154 * directly connected bounce traffic through intermediate 1155 * nodes; a numa_group can occupy any set of nodes. 1156 * The further away a node is, the less the faults count. 1157 * This seems to result in good task placement. 1158 */ 1159 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) { 1160 faults *= (sched_max_numa_distance - dist); 1161 faults /= (sched_max_numa_distance - LOCAL_DISTANCE); 1162 } 1163 1164 score += faults; 1165 } 1166 1167 return score; 1168 } 1169 1170 /* 1171 * These return the fraction of accesses done by a particular task, or 1172 * task group, on a particular numa node. The group weight is given a 1173 * larger multiplier, in order to group tasks together that are almost 1174 * evenly spread out between numa nodes. 1175 */ 1176 static inline unsigned long task_weight(struct task_struct *p, int nid, 1177 int dist) 1178 { 1179 unsigned long faults, total_faults; 1180 1181 if (!p->numa_faults) 1182 return 0; 1183 1184 total_faults = p->total_numa_faults; 1185 1186 if (!total_faults) 1187 return 0; 1188 1189 faults = task_faults(p, nid); 1190 faults += score_nearby_nodes(p, nid, dist, true); 1191 1192 return 1000 * faults / total_faults; 1193 } 1194 1195 static inline unsigned long group_weight(struct task_struct *p, int nid, 1196 int dist) 1197 { 1198 unsigned long faults, total_faults; 1199 1200 if (!p->numa_group) 1201 return 0; 1202 1203 total_faults = p->numa_group->total_faults; 1204 1205 if (!total_faults) 1206 return 0; 1207 1208 faults = group_faults(p, nid); 1209 faults += score_nearby_nodes(p, nid, dist, false); 1210 1211 return 1000 * faults / total_faults; 1212 } 1213 1214 bool should_numa_migrate_memory(struct task_struct *p, struct page * page, 1215 int src_nid, int dst_cpu) 1216 { 1217 struct numa_group *ng = p->numa_group; 1218 int dst_nid = cpu_to_node(dst_cpu); 1219 int last_cpupid, this_cpupid; 1220 1221 this_cpupid = cpu_pid_to_cpupid(dst_cpu, current->pid); 1222 1223 /* 1224 * Multi-stage node selection is used in conjunction with a periodic 1225 * migration fault to build a temporal task<->page relation. By using 1226 * a two-stage filter we remove short/unlikely relations. 1227 * 1228 * Using P(p) ~ n_p / n_t as per frequentist probability, we can equate 1229 * a task's usage of a particular page (n_p) per total usage of this 1230 * page (n_t) (in a given time-span) to a probability. 1231 * 1232 * Our periodic faults will sample this probability and getting the 1233 * same result twice in a row, given these samples are fully 1234 * independent, is then given by P(n)^2, provided our sample period 1235 * is sufficiently short compared to the usage pattern. 1236 * 1237 * This quadric squishes small probabilities, making it less likely we 1238 * act on an unlikely task<->page relation. 1239 */ 1240 last_cpupid = page_cpupid_xchg_last(page, this_cpupid); 1241 if (!cpupid_pid_unset(last_cpupid) && 1242 cpupid_to_nid(last_cpupid) != dst_nid) 1243 return false; 1244 1245 /* Always allow migrate on private faults */ 1246 if (cpupid_match_pid(p, last_cpupid)) 1247 return true; 1248 1249 /* A shared fault, but p->numa_group has not been set up yet. */ 1250 if (!ng) 1251 return true; 1252 1253 /* 1254 * Destination node is much more heavily used than the source 1255 * node? Allow migration. 1256 */ 1257 if (group_faults_cpu(ng, dst_nid) > group_faults_cpu(ng, src_nid) * 1258 ACTIVE_NODE_FRACTION) 1259 return true; 1260 1261 /* 1262 * Distribute memory according to CPU & memory use on each node, 1263 * with 3/4 hysteresis to avoid unnecessary memory migrations: 1264 * 1265 * faults_cpu(dst) 3 faults_cpu(src) 1266 * --------------- * - > --------------- 1267 * faults_mem(dst) 4 faults_mem(src) 1268 */ 1269 return group_faults_cpu(ng, dst_nid) * group_faults(p, src_nid) * 3 > 1270 group_faults_cpu(ng, src_nid) * group_faults(p, dst_nid) * 4; 1271 } 1272 1273 static unsigned long weighted_cpuload(const int cpu); 1274 static unsigned long source_load(int cpu, int type); 1275 static unsigned long target_load(int cpu, int type); 1276 static unsigned long capacity_of(int cpu); 1277 static long effective_load(struct task_group *tg, int cpu, long wl, long wg); 1278 1279 /* Cached statistics for all CPUs within a node */ 1280 struct numa_stats { 1281 unsigned long nr_running; 1282 unsigned long load; 1283 1284 /* Total compute capacity of CPUs on a node */ 1285 unsigned long compute_capacity; 1286 1287 /* Approximate capacity in terms of runnable tasks on a node */ 1288 unsigned long task_capacity; 1289 int has_free_capacity; 1290 }; 1291 1292 /* 1293 * XXX borrowed from update_sg_lb_stats 1294 */ 1295 static void update_numa_stats(struct numa_stats *ns, int nid) 1296 { 1297 int smt, cpu, cpus = 0; 1298 unsigned long capacity; 1299 1300 memset(ns, 0, sizeof(*ns)); 1301 for_each_cpu(cpu, cpumask_of_node(nid)) { 1302 struct rq *rq = cpu_rq(cpu); 1303 1304 ns->nr_running += rq->nr_running; 1305 ns->load += weighted_cpuload(cpu); 1306 ns->compute_capacity += capacity_of(cpu); 1307 1308 cpus++; 1309 } 1310 1311 /* 1312 * If we raced with hotplug and there are no CPUs left in our mask 1313 * the @ns structure is NULL'ed and task_numa_compare() will 1314 * not find this node attractive. 1315 * 1316 * We'll either bail at !has_free_capacity, or we'll detect a huge 1317 * imbalance and bail there. 1318 */ 1319 if (!cpus) 1320 return; 1321 1322 /* smt := ceil(cpus / capacity), assumes: 1 < smt_power < 2 */ 1323 smt = DIV_ROUND_UP(SCHED_CAPACITY_SCALE * cpus, ns->compute_capacity); 1324 capacity = cpus / smt; /* cores */ 1325 1326 ns->task_capacity = min_t(unsigned, capacity, 1327 DIV_ROUND_CLOSEST(ns->compute_capacity, SCHED_CAPACITY_SCALE)); 1328 ns->has_free_capacity = (ns->nr_running < ns->task_capacity); 1329 } 1330 1331 struct task_numa_env { 1332 struct task_struct *p; 1333 1334 int src_cpu, src_nid; 1335 int dst_cpu, dst_nid; 1336 1337 struct numa_stats src_stats, dst_stats; 1338 1339 int imbalance_pct; 1340 int dist; 1341 1342 struct task_struct *best_task; 1343 long best_imp; 1344 int best_cpu; 1345 }; 1346 1347 static void task_numa_assign(struct task_numa_env *env, 1348 struct task_struct *p, long imp) 1349 { 1350 if (env->best_task) 1351 put_task_struct(env->best_task); 1352 if (p) 1353 get_task_struct(p); 1354 1355 env->best_task = p; 1356 env->best_imp = imp; 1357 env->best_cpu = env->dst_cpu; 1358 } 1359 1360 static bool load_too_imbalanced(long src_load, long dst_load, 1361 struct task_numa_env *env) 1362 { 1363 long imb, old_imb; 1364 long orig_src_load, orig_dst_load; 1365 long src_capacity, dst_capacity; 1366 1367 /* 1368 * The load is corrected for the CPU capacity available on each node. 1369 * 1370 * src_load dst_load 1371 * ------------ vs --------- 1372 * src_capacity dst_capacity 1373 */ 1374 src_capacity = env->src_stats.compute_capacity; 1375 dst_capacity = env->dst_stats.compute_capacity; 1376 1377 /* We care about the slope of the imbalance, not the direction. */ 1378 if (dst_load < src_load) 1379 swap(dst_load, src_load); 1380 1381 /* Is the difference below the threshold? */ 1382 imb = dst_load * src_capacity * 100 - 1383 src_load * dst_capacity * env->imbalance_pct; 1384 if (imb <= 0) 1385 return false; 1386 1387 /* 1388 * The imbalance is above the allowed threshold. 1389 * Compare it with the old imbalance. 1390 */ 1391 orig_src_load = env->src_stats.load; 1392 orig_dst_load = env->dst_stats.load; 1393 1394 if (orig_dst_load < orig_src_load) 1395 swap(orig_dst_load, orig_src_load); 1396 1397 old_imb = orig_dst_load * src_capacity * 100 - 1398 orig_src_load * dst_capacity * env->imbalance_pct; 1399 1400 /* Would this change make things worse? */ 1401 return (imb > old_imb); 1402 } 1403 1404 /* 1405 * This checks if the overall compute and NUMA accesses of the system would 1406 * be improved if the source tasks was migrated to the target dst_cpu taking 1407 * into account that it might be best if task running on the dst_cpu should 1408 * be exchanged with the source task 1409 */ 1410 static void task_numa_compare(struct task_numa_env *env, 1411 long taskimp, long groupimp) 1412 { 1413 struct rq *src_rq = cpu_rq(env->src_cpu); 1414 struct rq *dst_rq = cpu_rq(env->dst_cpu); 1415 struct task_struct *cur; 1416 long src_load, dst_load; 1417 long load; 1418 long imp = env->p->numa_group ? groupimp : taskimp; 1419 long moveimp = imp; 1420 int dist = env->dist; 1421 1422 rcu_read_lock(); 1423 cur = task_rcu_dereference(&dst_rq->curr); 1424 if (cur && ((cur->flags & PF_EXITING) || is_idle_task(cur))) 1425 cur = NULL; 1426 1427 /* 1428 * Because we have preemption enabled we can get migrated around and 1429 * end try selecting ourselves (current == env->p) as a swap candidate. 1430 */ 1431 if (cur == env->p) 1432 goto unlock; 1433 1434 /* 1435 * "imp" is the fault differential for the source task between the 1436 * source and destination node. Calculate the total differential for 1437 * the source task and potential destination task. The more negative 1438 * the value is, the more rmeote accesses that would be expected to 1439 * be incurred if the tasks were swapped. 1440 */ 1441 if (cur) { 1442 /* Skip this swap candidate if cannot move to the source cpu */ 1443 if (!cpumask_test_cpu(env->src_cpu, tsk_cpus_allowed(cur))) 1444 goto unlock; 1445 1446 /* 1447 * If dst and source tasks are in the same NUMA group, or not 1448 * in any group then look only at task weights. 1449 */ 1450 if (cur->numa_group == env->p->numa_group) { 1451 imp = taskimp + task_weight(cur, env->src_nid, dist) - 1452 task_weight(cur, env->dst_nid, dist); 1453 /* 1454 * Add some hysteresis to prevent swapping the 1455 * tasks within a group over tiny differences. 1456 */ 1457 if (cur->numa_group) 1458 imp -= imp/16; 1459 } else { 1460 /* 1461 * Compare the group weights. If a task is all by 1462 * itself (not part of a group), use the task weight 1463 * instead. 1464 */ 1465 if (cur->numa_group) 1466 imp += group_weight(cur, env->src_nid, dist) - 1467 group_weight(cur, env->dst_nid, dist); 1468 else 1469 imp += task_weight(cur, env->src_nid, dist) - 1470 task_weight(cur, env->dst_nid, dist); 1471 } 1472 } 1473 1474 if (imp <= env->best_imp && moveimp <= env->best_imp) 1475 goto unlock; 1476 1477 if (!cur) { 1478 /* Is there capacity at our destination? */ 1479 if (env->src_stats.nr_running <= env->src_stats.task_capacity && 1480 !env->dst_stats.has_free_capacity) 1481 goto unlock; 1482 1483 goto balance; 1484 } 1485 1486 /* Balance doesn't matter much if we're running a task per cpu */ 1487 if (imp > env->best_imp && src_rq->nr_running == 1 && 1488 dst_rq->nr_running == 1) 1489 goto assign; 1490 1491 /* 1492 * In the overloaded case, try and keep the load balanced. 1493 */ 1494 balance: 1495 load = task_h_load(env->p); 1496 dst_load = env->dst_stats.load + load; 1497 src_load = env->src_stats.load - load; 1498 1499 if (moveimp > imp && moveimp > env->best_imp) { 1500 /* 1501 * If the improvement from just moving env->p direction is 1502 * better than swapping tasks around, check if a move is 1503 * possible. Store a slightly smaller score than moveimp, 1504 * so an actually idle CPU will win. 1505 */ 1506 if (!load_too_imbalanced(src_load, dst_load, env)) { 1507 imp = moveimp - 1; 1508 cur = NULL; 1509 goto assign; 1510 } 1511 } 1512 1513 if (imp <= env->best_imp) 1514 goto unlock; 1515 1516 if (cur) { 1517 load = task_h_load(cur); 1518 dst_load -= load; 1519 src_load += load; 1520 } 1521 1522 if (load_too_imbalanced(src_load, dst_load, env)) 1523 goto unlock; 1524 1525 /* 1526 * One idle CPU per node is evaluated for a task numa move. 1527 * Call select_idle_sibling to maybe find a better one. 1528 */ 1529 if (!cur) 1530 env->dst_cpu = select_idle_sibling(env->p, env->dst_cpu); 1531 1532 assign: 1533 task_numa_assign(env, cur, imp); 1534 unlock: 1535 rcu_read_unlock(); 1536 } 1537 1538 static void task_numa_find_cpu(struct task_numa_env *env, 1539 long taskimp, long groupimp) 1540 { 1541 int cpu; 1542 1543 for_each_cpu(cpu, cpumask_of_node(env->dst_nid)) { 1544 /* Skip this CPU if the source task cannot migrate */ 1545 if (!cpumask_test_cpu(cpu, tsk_cpus_allowed(env->p))) 1546 continue; 1547 1548 env->dst_cpu = cpu; 1549 task_numa_compare(env, taskimp, groupimp); 1550 } 1551 } 1552 1553 /* Only move tasks to a NUMA node less busy than the current node. */ 1554 static bool numa_has_capacity(struct task_numa_env *env) 1555 { 1556 struct numa_stats *src = &env->src_stats; 1557 struct numa_stats *dst = &env->dst_stats; 1558 1559 if (src->has_free_capacity && !dst->has_free_capacity) 1560 return false; 1561 1562 /* 1563 * Only consider a task move if the source has a higher load 1564 * than the destination, corrected for CPU capacity on each node. 1565 * 1566 * src->load dst->load 1567 * --------------------- vs --------------------- 1568 * src->compute_capacity dst->compute_capacity 1569 */ 1570 if (src->load * dst->compute_capacity * env->imbalance_pct > 1571 1572 dst->load * src->compute_capacity * 100) 1573 return true; 1574 1575 return false; 1576 } 1577 1578 static int task_numa_migrate(struct task_struct *p) 1579 { 1580 struct task_numa_env env = { 1581 .p = p, 1582 1583 .src_cpu = task_cpu(p), 1584 .src_nid = task_node(p), 1585 1586 .imbalance_pct = 112, 1587 1588 .best_task = NULL, 1589 .best_imp = 0, 1590 .best_cpu = -1, 1591 }; 1592 struct sched_domain *sd; 1593 unsigned long taskweight, groupweight; 1594 int nid, ret, dist; 1595 long taskimp, groupimp; 1596 1597 /* 1598 * Pick the lowest SD_NUMA domain, as that would have the smallest 1599 * imbalance and would be the first to start moving tasks about. 1600 * 1601 * And we want to avoid any moving of tasks about, as that would create 1602 * random movement of tasks -- counter the numa conditions we're trying 1603 * to satisfy here. 1604 */ 1605 rcu_read_lock(); 1606 sd = rcu_dereference(per_cpu(sd_numa, env.src_cpu)); 1607 if (sd) 1608 env.imbalance_pct = 100 + (sd->imbalance_pct - 100) / 2; 1609 rcu_read_unlock(); 1610 1611 /* 1612 * Cpusets can break the scheduler domain tree into smaller 1613 * balance domains, some of which do not cross NUMA boundaries. 1614 * Tasks that are "trapped" in such domains cannot be migrated 1615 * elsewhere, so there is no point in (re)trying. 1616 */ 1617 if (unlikely(!sd)) { 1618 p->numa_preferred_nid = task_node(p); 1619 return -EINVAL; 1620 } 1621 1622 env.dst_nid = p->numa_preferred_nid; 1623 dist = env.dist = node_distance(env.src_nid, env.dst_nid); 1624 taskweight = task_weight(p, env.src_nid, dist); 1625 groupweight = group_weight(p, env.src_nid, dist); 1626 update_numa_stats(&env.src_stats, env.src_nid); 1627 taskimp = task_weight(p, env.dst_nid, dist) - taskweight; 1628 groupimp = group_weight(p, env.dst_nid, dist) - groupweight; 1629 update_numa_stats(&env.dst_stats, env.dst_nid); 1630 1631 /* Try to find a spot on the preferred nid. */ 1632 if (numa_has_capacity(&env)) 1633 task_numa_find_cpu(&env, taskimp, groupimp); 1634 1635 /* 1636 * Look at other nodes in these cases: 1637 * - there is no space available on the preferred_nid 1638 * - the task is part of a numa_group that is interleaved across 1639 * multiple NUMA nodes; in order to better consolidate the group, 1640 * we need to check other locations. 1641 */ 1642 if (env.best_cpu == -1 || (p->numa_group && p->numa_group->active_nodes > 1)) { 1643 for_each_online_node(nid) { 1644 if (nid == env.src_nid || nid == p->numa_preferred_nid) 1645 continue; 1646 1647 dist = node_distance(env.src_nid, env.dst_nid); 1648 if (sched_numa_topology_type == NUMA_BACKPLANE && 1649 dist != env.dist) { 1650 taskweight = task_weight(p, env.src_nid, dist); 1651 groupweight = group_weight(p, env.src_nid, dist); 1652 } 1653 1654 /* Only consider nodes where both task and groups benefit */ 1655 taskimp = task_weight(p, nid, dist) - taskweight; 1656 groupimp = group_weight(p, nid, dist) - groupweight; 1657 if (taskimp < 0 && groupimp < 0) 1658 continue; 1659 1660 env.dist = dist; 1661 env.dst_nid = nid; 1662 update_numa_stats(&env.dst_stats, env.dst_nid); 1663 if (numa_has_capacity(&env)) 1664 task_numa_find_cpu(&env, taskimp, groupimp); 1665 } 1666 } 1667 1668 /* 1669 * If the task is part of a workload that spans multiple NUMA nodes, 1670 * and is migrating into one of the workload's active nodes, remember 1671 * this node as the task's preferred numa node, so the workload can 1672 * settle down. 1673 * A task that migrated to a second choice node will be better off 1674 * trying for a better one later. Do not set the preferred node here. 1675 */ 1676 if (p->numa_group) { 1677 struct numa_group *ng = p->numa_group; 1678 1679 if (env.best_cpu == -1) 1680 nid = env.src_nid; 1681 else 1682 nid = env.dst_nid; 1683 1684 if (ng->active_nodes > 1 && numa_is_active_node(env.dst_nid, ng)) 1685 sched_setnuma(p, env.dst_nid); 1686 } 1687 1688 /* No better CPU than the current one was found. */ 1689 if (env.best_cpu == -1) 1690 return -EAGAIN; 1691 1692 /* 1693 * Reset the scan period if the task is being rescheduled on an 1694 * alternative node to recheck if the tasks is now properly placed. 1695 */ 1696 p->numa_scan_period = task_scan_min(p); 1697 1698 if (env.best_task == NULL) { 1699 ret = migrate_task_to(p, env.best_cpu); 1700 if (ret != 0) 1701 trace_sched_stick_numa(p, env.src_cpu, env.best_cpu); 1702 return ret; 1703 } 1704 1705 ret = migrate_swap(p, env.best_task); 1706 if (ret != 0) 1707 trace_sched_stick_numa(p, env.src_cpu, task_cpu(env.best_task)); 1708 put_task_struct(env.best_task); 1709 return ret; 1710 } 1711 1712 /* Attempt to migrate a task to a CPU on the preferred node. */ 1713 static void numa_migrate_preferred(struct task_struct *p) 1714 { 1715 unsigned long interval = HZ; 1716 1717 /* This task has no NUMA fault statistics yet */ 1718 if (unlikely(p->numa_preferred_nid == -1 || !p->numa_faults)) 1719 return; 1720 1721 /* Periodically retry migrating the task to the preferred node */ 1722 interval = min(interval, msecs_to_jiffies(p->numa_scan_period) / 16); 1723 p->numa_migrate_retry = jiffies + interval; 1724 1725 /* Success if task is already running on preferred CPU */ 1726 if (task_node(p) == p->numa_preferred_nid) 1727 return; 1728 1729 /* Otherwise, try migrate to a CPU on the preferred node */ 1730 task_numa_migrate(p); 1731 } 1732 1733 /* 1734 * Find out how many nodes on the workload is actively running on. Do this by 1735 * tracking the nodes from which NUMA hinting faults are triggered. This can 1736 * be different from the set of nodes where the workload's memory is currently 1737 * located. 1738 */ 1739 static void numa_group_count_active_nodes(struct numa_group *numa_group) 1740 { 1741 unsigned long faults, max_faults = 0; 1742 int nid, active_nodes = 0; 1743 1744 for_each_online_node(nid) { 1745 faults = group_faults_cpu(numa_group, nid); 1746 if (faults > max_faults) 1747 max_faults = faults; 1748 } 1749 1750 for_each_online_node(nid) { 1751 faults = group_faults_cpu(numa_group, nid); 1752 if (faults * ACTIVE_NODE_FRACTION > max_faults) 1753 active_nodes++; 1754 } 1755 1756 numa_group->max_faults_cpu = max_faults; 1757 numa_group->active_nodes = active_nodes; 1758 } 1759 1760 /* 1761 * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS 1762 * increments. The more local the fault statistics are, the higher the scan 1763 * period will be for the next scan window. If local/(local+remote) ratio is 1764 * below NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS) 1765 * the scan period will decrease. Aim for 70% local accesses. 1766 */ 1767 #define NUMA_PERIOD_SLOTS 10 1768 #define NUMA_PERIOD_THRESHOLD 7 1769 1770 /* 1771 * Increase the scan period (slow down scanning) if the majority of 1772 * our memory is already on our local node, or if the majority of 1773 * the page accesses are shared with other processes. 1774 * Otherwise, decrease the scan period. 1775 */ 1776 static void update_task_scan_period(struct task_struct *p, 1777 unsigned long shared, unsigned long private) 1778 { 1779 unsigned int period_slot; 1780 int ratio; 1781 int diff; 1782 1783 unsigned long remote = p->numa_faults_locality[0]; 1784 unsigned long local = p->numa_faults_locality[1]; 1785 1786 /* 1787 * If there were no record hinting faults then either the task is 1788 * completely idle or all activity is areas that are not of interest 1789 * to automatic numa balancing. Related to that, if there were failed 1790 * migration then it implies we are migrating too quickly or the local 1791 * node is overloaded. In either case, scan slower 1792 */ 1793 if (local + shared == 0 || p->numa_faults_locality[2]) { 1794 p->numa_scan_period = min(p->numa_scan_period_max, 1795 p->numa_scan_period << 1); 1796 1797 p->mm->numa_next_scan = jiffies + 1798 msecs_to_jiffies(p->numa_scan_period); 1799 1800 return; 1801 } 1802 1803 /* 1804 * Prepare to scale scan period relative to the current period. 1805 * == NUMA_PERIOD_THRESHOLD scan period stays the same 1806 * < NUMA_PERIOD_THRESHOLD scan period decreases (scan faster) 1807 * >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower) 1808 */ 1809 period_slot = DIV_ROUND_UP(p->numa_scan_period, NUMA_PERIOD_SLOTS); 1810 ratio = (local * NUMA_PERIOD_SLOTS) / (local + remote); 1811 if (ratio >= NUMA_PERIOD_THRESHOLD) { 1812 int slot = ratio - NUMA_PERIOD_THRESHOLD; 1813 if (!slot) 1814 slot = 1; 1815 diff = slot * period_slot; 1816 } else { 1817 diff = -(NUMA_PERIOD_THRESHOLD - ratio) * period_slot; 1818 1819 /* 1820 * Scale scan rate increases based on sharing. There is an 1821 * inverse relationship between the degree of sharing and 1822 * the adjustment made to the scanning period. Broadly 1823 * speaking the intent is that there is little point 1824 * scanning faster if shared accesses dominate as it may 1825 * simply bounce migrations uselessly 1826 */ 1827 ratio = DIV_ROUND_UP(private * NUMA_PERIOD_SLOTS, (private + shared + 1)); 1828 diff = (diff * ratio) / NUMA_PERIOD_SLOTS; 1829 } 1830 1831 p->numa_scan_period = clamp(p->numa_scan_period + diff, 1832 task_scan_min(p), task_scan_max(p)); 1833 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality)); 1834 } 1835 1836 /* 1837 * Get the fraction of time the task has been running since the last 1838 * NUMA placement cycle. The scheduler keeps similar statistics, but 1839 * decays those on a 32ms period, which is orders of magnitude off 1840 * from the dozens-of-seconds NUMA balancing period. Use the scheduler 1841 * stats only if the task is so new there are no NUMA statistics yet. 1842 */ 1843 static u64 numa_get_avg_runtime(struct task_struct *p, u64 *period) 1844 { 1845 u64 runtime, delta, now; 1846 /* Use the start of this time slice to avoid calculations. */ 1847 now = p->se.exec_start; 1848 runtime = p->se.sum_exec_runtime; 1849 1850 if (p->last_task_numa_placement) { 1851 delta = runtime - p->last_sum_exec_runtime; 1852 *period = now - p->last_task_numa_placement; 1853 } else { 1854 delta = p->se.avg.load_sum / p->se.load.weight; 1855 *period = LOAD_AVG_MAX; 1856 } 1857 1858 p->last_sum_exec_runtime = runtime; 1859 p->last_task_numa_placement = now; 1860 1861 return delta; 1862 } 1863 1864 /* 1865 * Determine the preferred nid for a task in a numa_group. This needs to 1866 * be done in a way that produces consistent results with group_weight, 1867 * otherwise workloads might not converge. 1868 */ 1869 static int preferred_group_nid(struct task_struct *p, int nid) 1870 { 1871 nodemask_t nodes; 1872 int dist; 1873 1874 /* Direct connections between all NUMA nodes. */ 1875 if (sched_numa_topology_type == NUMA_DIRECT) 1876 return nid; 1877 1878 /* 1879 * On a system with glueless mesh NUMA topology, group_weight 1880 * scores nodes according to the number of NUMA hinting faults on 1881 * both the node itself, and on nearby nodes. 1882 */ 1883 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) { 1884 unsigned long score, max_score = 0; 1885 int node, max_node = nid; 1886 1887 dist = sched_max_numa_distance; 1888 1889 for_each_online_node(node) { 1890 score = group_weight(p, node, dist); 1891 if (score > max_score) { 1892 max_score = score; 1893 max_node = node; 1894 } 1895 } 1896 return max_node; 1897 } 1898 1899 /* 1900 * Finding the preferred nid in a system with NUMA backplane 1901 * interconnect topology is more involved. The goal is to locate 1902 * tasks from numa_groups near each other in the system, and 1903 * untangle workloads from different sides of the system. This requires 1904 * searching down the hierarchy of node groups, recursively searching 1905 * inside the highest scoring group of nodes. The nodemask tricks 1906 * keep the complexity of the search down. 1907 */ 1908 nodes = node_online_map; 1909 for (dist = sched_max_numa_distance; dist > LOCAL_DISTANCE; dist--) { 1910 unsigned long max_faults = 0; 1911 nodemask_t max_group = NODE_MASK_NONE; 1912 int a, b; 1913 1914 /* Are there nodes at this distance from each other? */ 1915 if (!find_numa_distance(dist)) 1916 continue; 1917 1918 for_each_node_mask(a, nodes) { 1919 unsigned long faults = 0; 1920 nodemask_t this_group; 1921 nodes_clear(this_group); 1922 1923 /* Sum group's NUMA faults; includes a==b case. */ 1924 for_each_node_mask(b, nodes) { 1925 if (node_distance(a, b) < dist) { 1926 faults += group_faults(p, b); 1927 node_set(b, this_group); 1928 node_clear(b, nodes); 1929 } 1930 } 1931 1932 /* Remember the top group. */ 1933 if (faults > max_faults) { 1934 max_faults = faults; 1935 max_group = this_group; 1936 /* 1937 * subtle: at the smallest distance there is 1938 * just one node left in each "group", the 1939 * winner is the preferred nid. 1940 */ 1941 nid = a; 1942 } 1943 } 1944 /* Next round, evaluate the nodes within max_group. */ 1945 if (!max_faults) 1946 break; 1947 nodes = max_group; 1948 } 1949 return nid; 1950 } 1951 1952 static void task_numa_placement(struct task_struct *p) 1953 { 1954 int seq, nid, max_nid = -1, max_group_nid = -1; 1955 unsigned long max_faults = 0, max_group_faults = 0; 1956 unsigned long fault_types[2] = { 0, 0 }; 1957 unsigned long total_faults; 1958 u64 runtime, period; 1959 spinlock_t *group_lock = NULL; 1960 1961 /* 1962 * The p->mm->numa_scan_seq field gets updated without 1963 * exclusive access. Use READ_ONCE() here to ensure 1964 * that the field is read in a single access: 1965 */ 1966 seq = READ_ONCE(p->mm->numa_scan_seq); 1967 if (p->numa_scan_seq == seq) 1968 return; 1969 p->numa_scan_seq = seq; 1970 p->numa_scan_period_max = task_scan_max(p); 1971 1972 total_faults = p->numa_faults_locality[0] + 1973 p->numa_faults_locality[1]; 1974 runtime = numa_get_avg_runtime(p, &period); 1975 1976 /* If the task is part of a group prevent parallel updates to group stats */ 1977 if (p->numa_group) { 1978 group_lock = &p->numa_group->lock; 1979 spin_lock_irq(group_lock); 1980 } 1981 1982 /* Find the node with the highest number of faults */ 1983 for_each_online_node(nid) { 1984 /* Keep track of the offsets in numa_faults array */ 1985 int mem_idx, membuf_idx, cpu_idx, cpubuf_idx; 1986 unsigned long faults = 0, group_faults = 0; 1987 int priv; 1988 1989 for (priv = 0; priv < NR_NUMA_HINT_FAULT_TYPES; priv++) { 1990 long diff, f_diff, f_weight; 1991 1992 mem_idx = task_faults_idx(NUMA_MEM, nid, priv); 1993 membuf_idx = task_faults_idx(NUMA_MEMBUF, nid, priv); 1994 cpu_idx = task_faults_idx(NUMA_CPU, nid, priv); 1995 cpubuf_idx = task_faults_idx(NUMA_CPUBUF, nid, priv); 1996 1997 /* Decay existing window, copy faults since last scan */ 1998 diff = p->numa_faults[membuf_idx] - p->numa_faults[mem_idx] / 2; 1999 fault_types[priv] += p->numa_faults[membuf_idx]; 2000 p->numa_faults[membuf_idx] = 0; 2001 2002 /* 2003 * Normalize the faults_from, so all tasks in a group 2004 * count according to CPU use, instead of by the raw 2005 * number of faults. Tasks with little runtime have 2006 * little over-all impact on throughput, and thus their 2007 * faults are less important. 2008 */ 2009 f_weight = div64_u64(runtime << 16, period + 1); 2010 f_weight = (f_weight * p->numa_faults[cpubuf_idx]) / 2011 (total_faults + 1); 2012 f_diff = f_weight - p->numa_faults[cpu_idx] / 2; 2013 p->numa_faults[cpubuf_idx] = 0; 2014 2015 p->numa_faults[mem_idx] += diff; 2016 p->numa_faults[cpu_idx] += f_diff; 2017 faults += p->numa_faults[mem_idx]; 2018 p->total_numa_faults += diff; 2019 if (p->numa_group) { 2020 /* 2021 * safe because we can only change our own group 2022 * 2023 * mem_idx represents the offset for a given 2024 * nid and priv in a specific region because it 2025 * is at the beginning of the numa_faults array. 2026 */ 2027 p->numa_group->faults[mem_idx] += diff; 2028 p->numa_group->faults_cpu[mem_idx] += f_diff; 2029 p->numa_group->total_faults += diff; 2030 group_faults += p->numa_group->faults[mem_idx]; 2031 } 2032 } 2033 2034 if (faults > max_faults) { 2035 max_faults = faults; 2036 max_nid = nid; 2037 } 2038 2039 if (group_faults > max_group_faults) { 2040 max_group_faults = group_faults; 2041 max_group_nid = nid; 2042 } 2043 } 2044 2045 update_task_scan_period(p, fault_types[0], fault_types[1]); 2046 2047 if (p->numa_group) { 2048 numa_group_count_active_nodes(p->numa_group); 2049 spin_unlock_irq(group_lock); 2050 max_nid = preferred_group_nid(p, max_group_nid); 2051 } 2052 2053 if (max_faults) { 2054 /* Set the new preferred node */ 2055 if (max_nid != p->numa_preferred_nid) 2056 sched_setnuma(p, max_nid); 2057 2058 if (task_node(p) != p->numa_preferred_nid) 2059 numa_migrate_preferred(p); 2060 } 2061 } 2062 2063 static inline int get_numa_group(struct numa_group *grp) 2064 { 2065 return atomic_inc_not_zero(&grp->refcount); 2066 } 2067 2068 static inline void put_numa_group(struct numa_group *grp) 2069 { 2070 if (atomic_dec_and_test(&grp->refcount)) 2071 kfree_rcu(grp, rcu); 2072 } 2073 2074 static void task_numa_group(struct task_struct *p, int cpupid, int flags, 2075 int *priv) 2076 { 2077 struct numa_group *grp, *my_grp; 2078 struct task_struct *tsk; 2079 bool join = false; 2080 int cpu = cpupid_to_cpu(cpupid); 2081 int i; 2082 2083 if (unlikely(!p->numa_group)) { 2084 unsigned int size = sizeof(struct numa_group) + 2085 4*nr_node_ids*sizeof(unsigned long); 2086 2087 grp = kzalloc(size, GFP_KERNEL | __GFP_NOWARN); 2088 if (!grp) 2089 return; 2090 2091 atomic_set(&grp->refcount, 1); 2092 grp->active_nodes = 1; 2093 grp->max_faults_cpu = 0; 2094 spin_lock_init(&grp->lock); 2095 grp->gid = p->pid; 2096 /* Second half of the array tracks nids where faults happen */ 2097 grp->faults_cpu = grp->faults + NR_NUMA_HINT_FAULT_TYPES * 2098 nr_node_ids; 2099 2100 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) 2101 grp->faults[i] = p->numa_faults[i]; 2102 2103 grp->total_faults = p->total_numa_faults; 2104 2105 grp->nr_tasks++; 2106 rcu_assign_pointer(p->numa_group, grp); 2107 } 2108 2109 rcu_read_lock(); 2110 tsk = READ_ONCE(cpu_rq(cpu)->curr); 2111 2112 if (!cpupid_match_pid(tsk, cpupid)) 2113 goto no_join; 2114 2115 grp = rcu_dereference(tsk->numa_group); 2116 if (!grp) 2117 goto no_join; 2118 2119 my_grp = p->numa_group; 2120 if (grp == my_grp) 2121 goto no_join; 2122 2123 /* 2124 * Only join the other group if its bigger; if we're the bigger group, 2125 * the other task will join us. 2126 */ 2127 if (my_grp->nr_tasks > grp->nr_tasks) 2128 goto no_join; 2129 2130 /* 2131 * Tie-break on the grp address. 2132 */ 2133 if (my_grp->nr_tasks == grp->nr_tasks && my_grp > grp) 2134 goto no_join; 2135 2136 /* Always join threads in the same process. */ 2137 if (tsk->mm == current->mm) 2138 join = true; 2139 2140 /* Simple filter to avoid false positives due to PID collisions */ 2141 if (flags & TNF_SHARED) 2142 join = true; 2143 2144 /* Update priv based on whether false sharing was detected */ 2145 *priv = !join; 2146 2147 if (join && !get_numa_group(grp)) 2148 goto no_join; 2149 2150 rcu_read_unlock(); 2151 2152 if (!join) 2153 return; 2154 2155 BUG_ON(irqs_disabled()); 2156 double_lock_irq(&my_grp->lock, &grp->lock); 2157 2158 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) { 2159 my_grp->faults[i] -= p->numa_faults[i]; 2160 grp->faults[i] += p->numa_faults[i]; 2161 } 2162 my_grp->total_faults -= p->total_numa_faults; 2163 grp->total_faults += p->total_numa_faults; 2164 2165 my_grp->nr_tasks--; 2166 grp->nr_tasks++; 2167 2168 spin_unlock(&my_grp->lock); 2169 spin_unlock_irq(&grp->lock); 2170 2171 rcu_assign_pointer(p->numa_group, grp); 2172 2173 put_numa_group(my_grp); 2174 return; 2175 2176 no_join: 2177 rcu_read_unlock(); 2178 return; 2179 } 2180 2181 void task_numa_free(struct task_struct *p) 2182 { 2183 struct numa_group *grp = p->numa_group; 2184 void *numa_faults = p->numa_faults; 2185 unsigned long flags; 2186 int i; 2187 2188 if (grp) { 2189 spin_lock_irqsave(&grp->lock, flags); 2190 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) 2191 grp->faults[i] -= p->numa_faults[i]; 2192 grp->total_faults -= p->total_numa_faults; 2193 2194 grp->nr_tasks--; 2195 spin_unlock_irqrestore(&grp->lock, flags); 2196 RCU_INIT_POINTER(p->numa_group, NULL); 2197 put_numa_group(grp); 2198 } 2199 2200 p->numa_faults = NULL; 2201 kfree(numa_faults); 2202 } 2203 2204 /* 2205 * Got a PROT_NONE fault for a page on @node. 2206 */ 2207 void task_numa_fault(int last_cpupid, int mem_node, int pages, int flags) 2208 { 2209 struct task_struct *p = current; 2210 bool migrated = flags & TNF_MIGRATED; 2211 int cpu_node = task_node(current); 2212 int local = !!(flags & TNF_FAULT_LOCAL); 2213 struct numa_group *ng; 2214 int priv; 2215 2216 if (!static_branch_likely(&sched_numa_balancing)) 2217 return; 2218 2219 /* for example, ksmd faulting in a user's mm */ 2220 if (!p->mm) 2221 return; 2222 2223 /* Allocate buffer to track faults on a per-node basis */ 2224 if (unlikely(!p->numa_faults)) { 2225 int size = sizeof(*p->numa_faults) * 2226 NR_NUMA_HINT_FAULT_BUCKETS * nr_node_ids; 2227 2228 p->numa_faults = kzalloc(size, GFP_KERNEL|__GFP_NOWARN); 2229 if (!p->numa_faults) 2230 return; 2231 2232 p->total_numa_faults = 0; 2233 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality)); 2234 } 2235 2236 /* 2237 * First accesses are treated as private, otherwise consider accesses 2238 * to be private if the accessing pid has not changed 2239 */ 2240 if (unlikely(last_cpupid == (-1 & LAST_CPUPID_MASK))) { 2241 priv = 1; 2242 } else { 2243 priv = cpupid_match_pid(p, last_cpupid); 2244 if (!priv && !(flags & TNF_NO_GROUP)) 2245 task_numa_group(p, last_cpupid, flags, &priv); 2246 } 2247 2248 /* 2249 * If a workload spans multiple NUMA nodes, a shared fault that 2250 * occurs wholly within the set of nodes that the workload is 2251 * actively using should be counted as local. This allows the 2252 * scan rate to slow down when a workload has settled down. 2253 */ 2254 ng = p->numa_group; 2255 if (!priv && !local && ng && ng->active_nodes > 1 && 2256 numa_is_active_node(cpu_node, ng) && 2257 numa_is_active_node(mem_node, ng)) 2258 local = 1; 2259 2260 task_numa_placement(p); 2261 2262 /* 2263 * Retry task to preferred node migration periodically, in case it 2264 * case it previously failed, or the scheduler moved us. 2265 */ 2266 if (time_after(jiffies, p->numa_migrate_retry)) 2267 numa_migrate_preferred(p); 2268 2269 if (migrated) 2270 p->numa_pages_migrated += pages; 2271 if (flags & TNF_MIGRATE_FAIL) 2272 p->numa_faults_locality[2] += pages; 2273 2274 p->numa_faults[task_faults_idx(NUMA_MEMBUF, mem_node, priv)] += pages; 2275 p->numa_faults[task_faults_idx(NUMA_CPUBUF, cpu_node, priv)] += pages; 2276 p->numa_faults_locality[local] += pages; 2277 } 2278 2279 static void reset_ptenuma_scan(struct task_struct *p) 2280 { 2281 /* 2282 * We only did a read acquisition of the mmap sem, so 2283 * p->mm->numa_scan_seq is written to without exclusive access 2284 * and the update is not guaranteed to be atomic. That's not 2285 * much of an issue though, since this is just used for 2286 * statistical sampling. Use READ_ONCE/WRITE_ONCE, which are not 2287 * expensive, to avoid any form of compiler optimizations: 2288 */ 2289 WRITE_ONCE(p->mm->numa_scan_seq, READ_ONCE(p->mm->numa_scan_seq) + 1); 2290 p->mm->numa_scan_offset = 0; 2291 } 2292 2293 /* 2294 * The expensive part of numa migration is done from task_work context. 2295 * Triggered from task_tick_numa(). 2296 */ 2297 void task_numa_work(struct callback_head *work) 2298 { 2299 unsigned long migrate, next_scan, now = jiffies; 2300 struct task_struct *p = current; 2301 struct mm_struct *mm = p->mm; 2302 u64 runtime = p->se.sum_exec_runtime; 2303 struct vm_area_struct *vma; 2304 unsigned long start, end; 2305 unsigned long nr_pte_updates = 0; 2306 long pages, virtpages; 2307 2308 WARN_ON_ONCE(p != container_of(work, struct task_struct, numa_work)); 2309 2310 work->next = work; /* protect against double add */ 2311 /* 2312 * Who cares about NUMA placement when they're dying. 2313 * 2314 * NOTE: make sure not to dereference p->mm before this check, 2315 * exit_task_work() happens _after_ exit_mm() so we could be called 2316 * without p->mm even though we still had it when we enqueued this 2317 * work. 2318 */ 2319 if (p->flags & PF_EXITING) 2320 return; 2321 2322 if (!mm->numa_next_scan) { 2323 mm->numa_next_scan = now + 2324 msecs_to_jiffies(sysctl_numa_balancing_scan_delay); 2325 } 2326 2327 /* 2328 * Enforce maximal scan/migration frequency.. 2329 */ 2330 migrate = mm->numa_next_scan; 2331 if (time_before(now, migrate)) 2332 return; 2333 2334 if (p->numa_scan_period == 0) { 2335 p->numa_scan_period_max = task_scan_max(p); 2336 p->numa_scan_period = task_scan_min(p); 2337 } 2338 2339 next_scan = now + msecs_to_jiffies(p->numa_scan_period); 2340 if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate) 2341 return; 2342 2343 /* 2344 * Delay this task enough that another task of this mm will likely win 2345 * the next time around. 2346 */ 2347 p->node_stamp += 2 * TICK_NSEC; 2348 2349 start = mm->numa_scan_offset; 2350 pages = sysctl_numa_balancing_scan_size; 2351 pages <<= 20 - PAGE_SHIFT; /* MB in pages */ 2352 virtpages = pages * 8; /* Scan up to this much virtual space */ 2353 if (!pages) 2354 return; 2355 2356 2357 down_read(&mm->mmap_sem); 2358 vma = find_vma(mm, start); 2359 if (!vma) { 2360 reset_ptenuma_scan(p); 2361 start = 0; 2362 vma = mm->mmap; 2363 } 2364 for (; vma; vma = vma->vm_next) { 2365 if (!vma_migratable(vma) || !vma_policy_mof(vma) || 2366 is_vm_hugetlb_page(vma) || (vma->vm_flags & VM_MIXEDMAP)) { 2367 continue; 2368 } 2369 2370 /* 2371 * Shared library pages mapped by multiple processes are not 2372 * migrated as it is expected they are cache replicated. Avoid 2373 * hinting faults in read-only file-backed mappings or the vdso 2374 * as migrating the pages will be of marginal benefit. 2375 */ 2376 if (!vma->vm_mm || 2377 (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ))) 2378 continue; 2379 2380 /* 2381 * Skip inaccessible VMAs to avoid any confusion between 2382 * PROT_NONE and NUMA hinting ptes 2383 */ 2384 if (!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE))) 2385 continue; 2386 2387 do { 2388 start = max(start, vma->vm_start); 2389 end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE); 2390 end = min(end, vma->vm_end); 2391 nr_pte_updates = change_prot_numa(vma, start, end); 2392 2393 /* 2394 * Try to scan sysctl_numa_balancing_size worth of 2395 * hpages that have at least one present PTE that 2396 * is not already pte-numa. If the VMA contains 2397 * areas that are unused or already full of prot_numa 2398 * PTEs, scan up to virtpages, to skip through those 2399 * areas faster. 2400 */ 2401 if (nr_pte_updates) 2402 pages -= (end - start) >> PAGE_SHIFT; 2403 virtpages -= (end - start) >> PAGE_SHIFT; 2404 2405 start = end; 2406 if (pages <= 0 || virtpages <= 0) 2407 goto out; 2408 2409 cond_resched(); 2410 } while (end != vma->vm_end); 2411 } 2412 2413 out: 2414 /* 2415 * It is possible to reach the end of the VMA list but the last few 2416 * VMAs are not guaranteed to the vma_migratable. If they are not, we 2417 * would find the !migratable VMA on the next scan but not reset the 2418 * scanner to the start so check it now. 2419 */ 2420 if (vma) 2421 mm->numa_scan_offset = start; 2422 else 2423 reset_ptenuma_scan(p); 2424 up_read(&mm->mmap_sem); 2425 2426 /* 2427 * Make sure tasks use at least 32x as much time to run other code 2428 * than they used here, to limit NUMA PTE scanning overhead to 3% max. 2429 * Usually update_task_scan_period slows down scanning enough; on an 2430 * overloaded system we need to limit overhead on a per task basis. 2431 */ 2432 if (unlikely(p->se.sum_exec_runtime != runtime)) { 2433 u64 diff = p->se.sum_exec_runtime - runtime; 2434 p->node_stamp += 32 * diff; 2435 } 2436 } 2437 2438 /* 2439 * Drive the periodic memory faults.. 2440 */ 2441 void task_tick_numa(struct rq *rq, struct task_struct *curr) 2442 { 2443 struct callback_head *work = &curr->numa_work; 2444 u64 period, now; 2445 2446 /* 2447 * We don't care about NUMA placement if we don't have memory. 2448 */ 2449 if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work) 2450 return; 2451 2452 /* 2453 * Using runtime rather than walltime has the dual advantage that 2454 * we (mostly) drive the selection from busy threads and that the 2455 * task needs to have done some actual work before we bother with 2456 * NUMA placement. 2457 */ 2458 now = curr->se.sum_exec_runtime; 2459 period = (u64)curr->numa_scan_period * NSEC_PER_MSEC; 2460 2461 if (now > curr->node_stamp + period) { 2462 if (!curr->node_stamp) 2463 curr->numa_scan_period = task_scan_min(curr); 2464 curr->node_stamp += period; 2465 2466 if (!time_before(jiffies, curr->mm->numa_next_scan)) { 2467 init_task_work(work, task_numa_work); /* TODO: move this into sched_fork() */ 2468 task_work_add(curr, work, true); 2469 } 2470 } 2471 } 2472 #else 2473 static void task_tick_numa(struct rq *rq, struct task_struct *curr) 2474 { 2475 } 2476 2477 static inline void account_numa_enqueue(struct rq *rq, struct task_struct *p) 2478 { 2479 } 2480 2481 static inline void account_numa_dequeue(struct rq *rq, struct task_struct *p) 2482 { 2483 } 2484 #endif /* CONFIG_NUMA_BALANCING */ 2485 2486 static void 2487 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se) 2488 { 2489 update_load_add(&cfs_rq->load, se->load.weight); 2490 if (!parent_entity(se)) 2491 update_load_add(&rq_of(cfs_rq)->load, se->load.weight); 2492 #ifdef CONFIG_SMP 2493 if (entity_is_task(se)) { 2494 struct rq *rq = rq_of(cfs_rq); 2495 2496 account_numa_enqueue(rq, task_of(se)); 2497 list_add(&se->group_node, &rq->cfs_tasks); 2498 } 2499 #endif 2500 cfs_rq->nr_running++; 2501 } 2502 2503 static void 2504 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se) 2505 { 2506 update_load_sub(&cfs_rq->load, se->load.weight); 2507 if (!parent_entity(se)) 2508 update_load_sub(&rq_of(cfs_rq)->load, se->load.weight); 2509 #ifdef CONFIG_SMP 2510 if (entity_is_task(se)) { 2511 account_numa_dequeue(rq_of(cfs_rq), task_of(se)); 2512 list_del_init(&se->group_node); 2513 } 2514 #endif 2515 cfs_rq->nr_running--; 2516 } 2517 2518 #ifdef CONFIG_FAIR_GROUP_SCHED 2519 # ifdef CONFIG_SMP 2520 static long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg) 2521 { 2522 long tg_weight, load, shares; 2523 2524 /* 2525 * This really should be: cfs_rq->avg.load_avg, but instead we use 2526 * cfs_rq->load.weight, which is its upper bound. This helps ramp up 2527 * the shares for small weight interactive tasks. 2528 */ 2529 load = scale_load_down(cfs_rq->load.weight); 2530 2531 tg_weight = atomic_long_read(&tg->load_avg); 2532 2533 /* Ensure tg_weight >= load */ 2534 tg_weight -= cfs_rq->tg_load_avg_contrib; 2535 tg_weight += load; 2536 2537 shares = (tg->shares * load); 2538 if (tg_weight) 2539 shares /= tg_weight; 2540 2541 if (shares < MIN_SHARES) 2542 shares = MIN_SHARES; 2543 if (shares > tg->shares) 2544 shares = tg->shares; 2545 2546 return shares; 2547 } 2548 # else /* CONFIG_SMP */ 2549 static inline long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg) 2550 { 2551 return tg->shares; 2552 } 2553 # endif /* CONFIG_SMP */ 2554 2555 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, 2556 unsigned long weight) 2557 { 2558 if (se->on_rq) { 2559 /* commit outstanding execution time */ 2560 if (cfs_rq->curr == se) 2561 update_curr(cfs_rq); 2562 account_entity_dequeue(cfs_rq, se); 2563 } 2564 2565 update_load_set(&se->load, weight); 2566 2567 if (se->on_rq) 2568 account_entity_enqueue(cfs_rq, se); 2569 } 2570 2571 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq); 2572 2573 static void update_cfs_shares(struct cfs_rq *cfs_rq) 2574 { 2575 struct task_group *tg; 2576 struct sched_entity *se; 2577 long shares; 2578 2579 tg = cfs_rq->tg; 2580 se = tg->se[cpu_of(rq_of(cfs_rq))]; 2581 if (!se || throttled_hierarchy(cfs_rq)) 2582 return; 2583 #ifndef CONFIG_SMP 2584 if (likely(se->load.weight == tg->shares)) 2585 return; 2586 #endif 2587 shares = calc_cfs_shares(cfs_rq, tg); 2588 2589 reweight_entity(cfs_rq_of(se), se, shares); 2590 } 2591 #else /* CONFIG_FAIR_GROUP_SCHED */ 2592 static inline void update_cfs_shares(struct cfs_rq *cfs_rq) 2593 { 2594 } 2595 #endif /* CONFIG_FAIR_GROUP_SCHED */ 2596 2597 #ifdef CONFIG_SMP 2598 /* Precomputed fixed inverse multiplies for multiplication by y^n */ 2599 static const u32 runnable_avg_yN_inv[] = { 2600 0xffffffff, 0xfa83b2da, 0xf5257d14, 0xefe4b99a, 0xeac0c6e6, 0xe5b906e6, 2601 0xe0ccdeeb, 0xdbfbb796, 0xd744fcc9, 0xd2a81d91, 0xce248c14, 0xc9b9bd85, 2602 0xc5672a10, 0xc12c4cc9, 0xbd08a39e, 0xb8fbaf46, 0xb504f333, 0xb123f581, 2603 0xad583ee9, 0xa9a15ab4, 0xa5fed6a9, 0xa2704302, 0x9ef5325f, 0x9b8d39b9, 2604 0x9837f050, 0x94f4efa8, 0x91c3d373, 0x8ea4398a, 0x8b95c1e3, 0x88980e80, 2605 0x85aac367, 0x82cd8698, 2606 }; 2607 2608 /* 2609 * Precomputed \Sum y^k { 1<=k<=n }. These are floor(true_value) to prevent 2610 * over-estimates when re-combining. 2611 */ 2612 static const u32 runnable_avg_yN_sum[] = { 2613 0, 1002, 1982, 2941, 3880, 4798, 5697, 6576, 7437, 8279, 9103, 2614 9909,10698,11470,12226,12966,13690,14398,15091,15769,16433,17082, 2615 17718,18340,18949,19545,20128,20698,21256,21802,22336,22859,23371, 2616 }; 2617 2618 /* 2619 * Precomputed \Sum y^k { 1<=k<=n, where n%32=0). Values are rolled down to 2620 * lower integers. See Documentation/scheduler/sched-avg.txt how these 2621 * were generated: 2622 */ 2623 static const u32 __accumulated_sum_N32[] = { 2624 0, 23371, 35056, 40899, 43820, 45281, 2625 46011, 46376, 46559, 46650, 46696, 46719, 2626 }; 2627 2628 /* 2629 * Approximate: 2630 * val * y^n, where y^32 ~= 0.5 (~1 scheduling period) 2631 */ 2632 static __always_inline u64 decay_load(u64 val, u64 n) 2633 { 2634 unsigned int local_n; 2635 2636 if (!n) 2637 return val; 2638 else if (unlikely(n > LOAD_AVG_PERIOD * 63)) 2639 return 0; 2640 2641 /* after bounds checking we can collapse to 32-bit */ 2642 local_n = n; 2643 2644 /* 2645 * As y^PERIOD = 1/2, we can combine 2646 * y^n = 1/2^(n/PERIOD) * y^(n%PERIOD) 2647 * With a look-up table which covers y^n (n<PERIOD) 2648 * 2649 * To achieve constant time decay_load. 2650 */ 2651 if (unlikely(local_n >= LOAD_AVG_PERIOD)) { 2652 val >>= local_n / LOAD_AVG_PERIOD; 2653 local_n %= LOAD_AVG_PERIOD; 2654 } 2655 2656 val = mul_u64_u32_shr(val, runnable_avg_yN_inv[local_n], 32); 2657 return val; 2658 } 2659 2660 /* 2661 * For updates fully spanning n periods, the contribution to runnable 2662 * average will be: \Sum 1024*y^n 2663 * 2664 * We can compute this reasonably efficiently by combining: 2665 * y^PERIOD = 1/2 with precomputed \Sum 1024*y^n {for n <PERIOD} 2666 */ 2667 static u32 __compute_runnable_contrib(u64 n) 2668 { 2669 u32 contrib = 0; 2670 2671 if (likely(n <= LOAD_AVG_PERIOD)) 2672 return runnable_avg_yN_sum[n]; 2673 else if (unlikely(n >= LOAD_AVG_MAX_N)) 2674 return LOAD_AVG_MAX; 2675 2676 /* Since n < LOAD_AVG_MAX_N, n/LOAD_AVG_PERIOD < 11 */ 2677 contrib = __accumulated_sum_N32[n/LOAD_AVG_PERIOD]; 2678 n %= LOAD_AVG_PERIOD; 2679 contrib = decay_load(contrib, n); 2680 return contrib + runnable_avg_yN_sum[n]; 2681 } 2682 2683 #define cap_scale(v, s) ((v)*(s) >> SCHED_CAPACITY_SHIFT) 2684 2685 /* 2686 * We can represent the historical contribution to runnable average as the 2687 * coefficients of a geometric series. To do this we sub-divide our runnable 2688 * history into segments of approximately 1ms (1024us); label the segment that 2689 * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g. 2690 * 2691 * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ... 2692 * p0 p1 p2 2693 * (now) (~1ms ago) (~2ms ago) 2694 * 2695 * Let u_i denote the fraction of p_i that the entity was runnable. 2696 * 2697 * We then designate the fractions u_i as our co-efficients, yielding the 2698 * following representation of historical load: 2699 * u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ... 2700 * 2701 * We choose y based on the with of a reasonably scheduling period, fixing: 2702 * y^32 = 0.5 2703 * 2704 * This means that the contribution to load ~32ms ago (u_32) will be weighted 2705 * approximately half as much as the contribution to load within the last ms 2706 * (u_0). 2707 * 2708 * When a period "rolls over" and we have new u_0`, multiplying the previous 2709 * sum again by y is sufficient to update: 2710 * load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... ) 2711 * = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}] 2712 */ 2713 static __always_inline int 2714 __update_load_avg(u64 now, int cpu, struct sched_avg *sa, 2715 unsigned long weight, int running, struct cfs_rq *cfs_rq) 2716 { 2717 u64 delta, scaled_delta, periods; 2718 u32 contrib; 2719 unsigned int delta_w, scaled_delta_w, decayed = 0; 2720 unsigned long scale_freq, scale_cpu; 2721 2722 delta = now - sa->last_update_time; 2723 /* 2724 * This should only happen when time goes backwards, which it 2725 * unfortunately does during sched clock init when we swap over to TSC. 2726 */ 2727 if ((s64)delta < 0) { 2728 sa->last_update_time = now; 2729 return 0; 2730 } 2731 2732 /* 2733 * Use 1024ns as the unit of measurement since it's a reasonable 2734 * approximation of 1us and fast to compute. 2735 */ 2736 delta >>= 10; 2737 if (!delta) 2738 return 0; 2739 sa->last_update_time = now; 2740 2741 scale_freq = arch_scale_freq_capacity(NULL, cpu); 2742 scale_cpu = arch_scale_cpu_capacity(NULL, cpu); 2743 2744 /* delta_w is the amount already accumulated against our next period */ 2745 delta_w = sa->period_contrib; 2746 if (delta + delta_w >= 1024) { 2747 decayed = 1; 2748 2749 /* how much left for next period will start over, we don't know yet */ 2750 sa->period_contrib = 0; 2751 2752 /* 2753 * Now that we know we're crossing a period boundary, figure 2754 * out how much from delta we need to complete the current 2755 * period and accrue it. 2756 */ 2757 delta_w = 1024 - delta_w; 2758 scaled_delta_w = cap_scale(delta_w, scale_freq); 2759 if (weight) { 2760 sa->load_sum += weight * scaled_delta_w; 2761 if (cfs_rq) { 2762 cfs_rq->runnable_load_sum += 2763 weight * scaled_delta_w; 2764 } 2765 } 2766 if (running) 2767 sa->util_sum += scaled_delta_w * scale_cpu; 2768 2769 delta -= delta_w; 2770 2771 /* Figure out how many additional periods this update spans */ 2772 periods = delta / 1024; 2773 delta %= 1024; 2774 2775 sa->load_sum = decay_load(sa->load_sum, periods + 1); 2776 if (cfs_rq) { 2777 cfs_rq->runnable_load_sum = 2778 decay_load(cfs_rq->runnable_load_sum, periods + 1); 2779 } 2780 sa->util_sum = decay_load((u64)(sa->util_sum), periods + 1); 2781 2782 /* Efficiently calculate \sum (1..n_period) 1024*y^i */ 2783 contrib = __compute_runnable_contrib(periods); 2784 contrib = cap_scale(contrib, scale_freq); 2785 if (weight) { 2786 sa->load_sum += weight * contrib; 2787 if (cfs_rq) 2788 cfs_rq->runnable_load_sum += weight * contrib; 2789 } 2790 if (running) 2791 sa->util_sum += contrib * scale_cpu; 2792 } 2793 2794 /* Remainder of delta accrued against u_0` */ 2795 scaled_delta = cap_scale(delta, scale_freq); 2796 if (weight) { 2797 sa->load_sum += weight * scaled_delta; 2798 if (cfs_rq) 2799 cfs_rq->runnable_load_sum += weight * scaled_delta; 2800 } 2801 if (running) 2802 sa->util_sum += scaled_delta * scale_cpu; 2803 2804 sa->period_contrib += delta; 2805 2806 if (decayed) { 2807 sa->load_avg = div_u64(sa->load_sum, LOAD_AVG_MAX); 2808 if (cfs_rq) { 2809 cfs_rq->runnable_load_avg = 2810 div_u64(cfs_rq->runnable_load_sum, LOAD_AVG_MAX); 2811 } 2812 sa->util_avg = sa->util_sum / LOAD_AVG_MAX; 2813 } 2814 2815 return decayed; 2816 } 2817 2818 #ifdef CONFIG_FAIR_GROUP_SCHED 2819 /* 2820 * Updating tg's load_avg is necessary before update_cfs_share (which is done) 2821 * and effective_load (which is not done because it is too costly). 2822 */ 2823 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force) 2824 { 2825 long delta = cfs_rq->avg.load_avg - cfs_rq->tg_load_avg_contrib; 2826 2827 /* 2828 * No need to update load_avg for root_task_group as it is not used. 2829 */ 2830 if (cfs_rq->tg == &root_task_group) 2831 return; 2832 2833 if (force || abs(delta) > cfs_rq->tg_load_avg_contrib / 64) { 2834 atomic_long_add(delta, &cfs_rq->tg->load_avg); 2835 cfs_rq->tg_load_avg_contrib = cfs_rq->avg.load_avg; 2836 } 2837 } 2838 2839 /* 2840 * Called within set_task_rq() right before setting a task's cpu. The 2841 * caller only guarantees p->pi_lock is held; no other assumptions, 2842 * including the state of rq->lock, should be made. 2843 */ 2844 void set_task_rq_fair(struct sched_entity *se, 2845 struct cfs_rq *prev, struct cfs_rq *next) 2846 { 2847 if (!sched_feat(ATTACH_AGE_LOAD)) 2848 return; 2849 2850 /* 2851 * We are supposed to update the task to "current" time, then its up to 2852 * date and ready to go to new CPU/cfs_rq. But we have difficulty in 2853 * getting what current time is, so simply throw away the out-of-date 2854 * time. This will result in the wakee task is less decayed, but giving 2855 * the wakee more load sounds not bad. 2856 */ 2857 if (se->avg.last_update_time && prev) { 2858 u64 p_last_update_time; 2859 u64 n_last_update_time; 2860 2861 #ifndef CONFIG_64BIT 2862 u64 p_last_update_time_copy; 2863 u64 n_last_update_time_copy; 2864 2865 do { 2866 p_last_update_time_copy = prev->load_last_update_time_copy; 2867 n_last_update_time_copy = next->load_last_update_time_copy; 2868 2869 smp_rmb(); 2870 2871 p_last_update_time = prev->avg.last_update_time; 2872 n_last_update_time = next->avg.last_update_time; 2873 2874 } while (p_last_update_time != p_last_update_time_copy || 2875 n_last_update_time != n_last_update_time_copy); 2876 #else 2877 p_last_update_time = prev->avg.last_update_time; 2878 n_last_update_time = next->avg.last_update_time; 2879 #endif 2880 __update_load_avg(p_last_update_time, cpu_of(rq_of(prev)), 2881 &se->avg, 0, 0, NULL); 2882 se->avg.last_update_time = n_last_update_time; 2883 } 2884 } 2885 #else /* CONFIG_FAIR_GROUP_SCHED */ 2886 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force) {} 2887 #endif /* CONFIG_FAIR_GROUP_SCHED */ 2888 2889 static inline void cfs_rq_util_change(struct cfs_rq *cfs_rq) 2890 { 2891 struct rq *rq = rq_of(cfs_rq); 2892 int cpu = cpu_of(rq); 2893 2894 if (cpu == smp_processor_id() && &rq->cfs == cfs_rq) { 2895 unsigned long max = rq->cpu_capacity_orig; 2896 2897 /* 2898 * There are a few boundary cases this might miss but it should 2899 * get called often enough that that should (hopefully) not be 2900 * a real problem -- added to that it only calls on the local 2901 * CPU, so if we enqueue remotely we'll miss an update, but 2902 * the next tick/schedule should update. 2903 * 2904 * It will not get called when we go idle, because the idle 2905 * thread is a different class (!fair), nor will the utilization 2906 * number include things like RT tasks. 2907 * 2908 * As is, the util number is not freq-invariant (we'd have to 2909 * implement arch_scale_freq_capacity() for that). 2910 * 2911 * See cpu_util(). 2912 */ 2913 cpufreq_update_util(rq_clock(rq), 2914 min(cfs_rq->avg.util_avg, max), max); 2915 } 2916 } 2917 2918 /* 2919 * Unsigned subtract and clamp on underflow. 2920 * 2921 * Explicitly do a load-store to ensure the intermediate value never hits 2922 * memory. This allows lockless observations without ever seeing the negative 2923 * values. 2924 */ 2925 #define sub_positive(_ptr, _val) do { \ 2926 typeof(_ptr) ptr = (_ptr); \ 2927 typeof(*ptr) val = (_val); \ 2928 typeof(*ptr) res, var = READ_ONCE(*ptr); \ 2929 res = var - val; \ 2930 if (res > var) \ 2931 res = 0; \ 2932 WRITE_ONCE(*ptr, res); \ 2933 } while (0) 2934 2935 /** 2936 * update_cfs_rq_load_avg - update the cfs_rq's load/util averages 2937 * @now: current time, as per cfs_rq_clock_task() 2938 * @cfs_rq: cfs_rq to update 2939 * @update_freq: should we call cfs_rq_util_change() or will the call do so 2940 * 2941 * The cfs_rq avg is the direct sum of all its entities (blocked and runnable) 2942 * avg. The immediate corollary is that all (fair) tasks must be attached, see 2943 * post_init_entity_util_avg(). 2944 * 2945 * cfs_rq->avg is used for task_h_load() and update_cfs_share() for example. 2946 * 2947 * Returns true if the load decayed or we removed utilization. It is expected 2948 * that one calls update_tg_load_avg() on this condition, but after you've 2949 * modified the cfs_rq avg (attach/detach), such that we propagate the new 2950 * avg up. 2951 */ 2952 static inline int 2953 update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq, bool update_freq) 2954 { 2955 struct sched_avg *sa = &cfs_rq->avg; 2956 int decayed, removed_load = 0, removed_util = 0; 2957 2958 if (atomic_long_read(&cfs_rq->removed_load_avg)) { 2959 s64 r = atomic_long_xchg(&cfs_rq->removed_load_avg, 0); 2960 sub_positive(&sa->load_avg, r); 2961 sub_positive(&sa->load_sum, r * LOAD_AVG_MAX); 2962 removed_load = 1; 2963 } 2964 2965 if (atomic_long_read(&cfs_rq->removed_util_avg)) { 2966 long r = atomic_long_xchg(&cfs_rq->removed_util_avg, 0); 2967 sub_positive(&sa->util_avg, r); 2968 sub_positive(&sa->util_sum, r * LOAD_AVG_MAX); 2969 removed_util = 1; 2970 } 2971 2972 decayed = __update_load_avg(now, cpu_of(rq_of(cfs_rq)), sa, 2973 scale_load_down(cfs_rq->load.weight), cfs_rq->curr != NULL, cfs_rq); 2974 2975 #ifndef CONFIG_64BIT 2976 smp_wmb(); 2977 cfs_rq->load_last_update_time_copy = sa->last_update_time; 2978 #endif 2979 2980 if (update_freq && (decayed || removed_util)) 2981 cfs_rq_util_change(cfs_rq); 2982 2983 return decayed || removed_load; 2984 } 2985 2986 /* Update task and its cfs_rq load average */ 2987 static inline void update_load_avg(struct sched_entity *se, int update_tg) 2988 { 2989 struct cfs_rq *cfs_rq = cfs_rq_of(se); 2990 u64 now = cfs_rq_clock_task(cfs_rq); 2991 struct rq *rq = rq_of(cfs_rq); 2992 int cpu = cpu_of(rq); 2993 2994 /* 2995 * Track task load average for carrying it to new CPU after migrated, and 2996 * track group sched_entity load average for task_h_load calc in migration 2997 */ 2998 __update_load_avg(now, cpu, &se->avg, 2999 se->on_rq * scale_load_down(se->load.weight), 3000 cfs_rq->curr == se, NULL); 3001 3002 if (update_cfs_rq_load_avg(now, cfs_rq, true) && update_tg) 3003 update_tg_load_avg(cfs_rq, 0); 3004 } 3005 3006 /** 3007 * attach_entity_load_avg - attach this entity to its cfs_rq load avg 3008 * @cfs_rq: cfs_rq to attach to 3009 * @se: sched_entity to attach 3010 * 3011 * Must call update_cfs_rq_load_avg() before this, since we rely on 3012 * cfs_rq->avg.last_update_time being current. 3013 */ 3014 static void attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) 3015 { 3016 if (!sched_feat(ATTACH_AGE_LOAD)) 3017 goto skip_aging; 3018 3019 /* 3020 * If we got migrated (either between CPUs or between cgroups) we'll 3021 * have aged the average right before clearing @last_update_time. 3022 * 3023 * Or we're fresh through post_init_entity_util_avg(). 3024 */ 3025 if (se->avg.last_update_time) { 3026 __update_load_avg(cfs_rq->avg.last_update_time, cpu_of(rq_of(cfs_rq)), 3027 &se->avg, 0, 0, NULL); 3028 3029 /* 3030 * XXX: we could have just aged the entire load away if we've been 3031 * absent from the fair class for too long. 3032 */ 3033 } 3034 3035 skip_aging: 3036 se->avg.last_update_time = cfs_rq->avg.last_update_time; 3037 cfs_rq->avg.load_avg += se->avg.load_avg; 3038 cfs_rq->avg.load_sum += se->avg.load_sum; 3039 cfs_rq->avg.util_avg += se->avg.util_avg; 3040 cfs_rq->avg.util_sum += se->avg.util_sum; 3041 3042 cfs_rq_util_change(cfs_rq); 3043 } 3044 3045 /** 3046 * detach_entity_load_avg - detach this entity from its cfs_rq load avg 3047 * @cfs_rq: cfs_rq to detach from 3048 * @se: sched_entity to detach 3049 * 3050 * Must call update_cfs_rq_load_avg() before this, since we rely on 3051 * cfs_rq->avg.last_update_time being current. 3052 */ 3053 static void detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) 3054 { 3055 __update_load_avg(cfs_rq->avg.last_update_time, cpu_of(rq_of(cfs_rq)), 3056 &se->avg, se->on_rq * scale_load_down(se->load.weight), 3057 cfs_rq->curr == se, NULL); 3058 3059 sub_positive(&cfs_rq->avg.load_avg, se->avg.load_avg); 3060 sub_positive(&cfs_rq->avg.load_sum, se->avg.load_sum); 3061 sub_positive(&cfs_rq->avg.util_avg, se->avg.util_avg); 3062 sub_positive(&cfs_rq->avg.util_sum, se->avg.util_sum); 3063 3064 cfs_rq_util_change(cfs_rq); 3065 } 3066 3067 /* Add the load generated by se into cfs_rq's load average */ 3068 static inline void 3069 enqueue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) 3070 { 3071 struct sched_avg *sa = &se->avg; 3072 u64 now = cfs_rq_clock_task(cfs_rq); 3073 int migrated, decayed; 3074 3075 migrated = !sa->last_update_time; 3076 if (!migrated) { 3077 __update_load_avg(now, cpu_of(rq_of(cfs_rq)), sa, 3078 se->on_rq * scale_load_down(se->load.weight), 3079 cfs_rq->curr == se, NULL); 3080 } 3081 3082 decayed = update_cfs_rq_load_avg(now, cfs_rq, !migrated); 3083 3084 cfs_rq->runnable_load_avg += sa->load_avg; 3085 cfs_rq->runnable_load_sum += sa->load_sum; 3086 3087 if (migrated) 3088 attach_entity_load_avg(cfs_rq, se); 3089 3090 if (decayed || migrated) 3091 update_tg_load_avg(cfs_rq, 0); 3092 } 3093 3094 /* Remove the runnable load generated by se from cfs_rq's runnable load average */ 3095 static inline void 3096 dequeue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) 3097 { 3098 update_load_avg(se, 1); 3099 3100 cfs_rq->runnable_load_avg = 3101 max_t(long, cfs_rq->runnable_load_avg - se->avg.load_avg, 0); 3102 cfs_rq->runnable_load_sum = 3103 max_t(s64, cfs_rq->runnable_load_sum - se->avg.load_sum, 0); 3104 } 3105 3106 #ifndef CONFIG_64BIT 3107 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq) 3108 { 3109 u64 last_update_time_copy; 3110 u64 last_update_time; 3111 3112 do { 3113 last_update_time_copy = cfs_rq->load_last_update_time_copy; 3114 smp_rmb(); 3115 last_update_time = cfs_rq->avg.last_update_time; 3116 } while (last_update_time != last_update_time_copy); 3117 3118 return last_update_time; 3119 } 3120 #else 3121 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq) 3122 { 3123 return cfs_rq->avg.last_update_time; 3124 } 3125 #endif 3126 3127 /* 3128 * Task first catches up with cfs_rq, and then subtract 3129 * itself from the cfs_rq (task must be off the queue now). 3130 */ 3131 void remove_entity_load_avg(struct sched_entity *se) 3132 { 3133 struct cfs_rq *cfs_rq = cfs_rq_of(se); 3134 u64 last_update_time; 3135 3136 /* 3137 * tasks cannot exit without having gone through wake_up_new_task() -> 3138 * post_init_entity_util_avg() which will have added things to the 3139 * cfs_rq, so we can remove unconditionally. 3140 * 3141 * Similarly for groups, they will have passed through 3142 * post_init_entity_util_avg() before unregister_sched_fair_group() 3143 * calls this. 3144 */ 3145 3146 last_update_time = cfs_rq_last_update_time(cfs_rq); 3147 3148 __update_load_avg(last_update_time, cpu_of(rq_of(cfs_rq)), &se->avg, 0, 0, NULL); 3149 atomic_long_add(se->avg.load_avg, &cfs_rq->removed_load_avg); 3150 atomic_long_add(se->avg.util_avg, &cfs_rq->removed_util_avg); 3151 } 3152 3153 static inline unsigned long cfs_rq_runnable_load_avg(struct cfs_rq *cfs_rq) 3154 { 3155 return cfs_rq->runnable_load_avg; 3156 } 3157 3158 static inline unsigned long cfs_rq_load_avg(struct cfs_rq *cfs_rq) 3159 { 3160 return cfs_rq->avg.load_avg; 3161 } 3162 3163 static int idle_balance(struct rq *this_rq); 3164 3165 #else /* CONFIG_SMP */ 3166 3167 static inline int 3168 update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq, bool update_freq) 3169 { 3170 return 0; 3171 } 3172 3173 static inline void update_load_avg(struct sched_entity *se, int not_used) 3174 { 3175 struct cfs_rq *cfs_rq = cfs_rq_of(se); 3176 struct rq *rq = rq_of(cfs_rq); 3177 3178 cpufreq_trigger_update(rq_clock(rq)); 3179 } 3180 3181 static inline void 3182 enqueue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {} 3183 static inline void 3184 dequeue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {} 3185 static inline void remove_entity_load_avg(struct sched_entity *se) {} 3186 3187 static inline void 3188 attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {} 3189 static inline void 3190 detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {} 3191 3192 static inline int idle_balance(struct rq *rq) 3193 { 3194 return 0; 3195 } 3196 3197 #endif /* CONFIG_SMP */ 3198 3199 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se) 3200 { 3201 #ifdef CONFIG_SCHEDSTATS 3202 struct task_struct *tsk = NULL; 3203 3204 if (entity_is_task(se)) 3205 tsk = task_of(se); 3206 3207 if (se->statistics.sleep_start) { 3208 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.sleep_start; 3209 3210 if ((s64)delta < 0) 3211 delta = 0; 3212 3213 if (unlikely(delta > se->statistics.sleep_max)) 3214 se->statistics.sleep_max = delta; 3215 3216 se->statistics.sleep_start = 0; 3217 se->statistics.sum_sleep_runtime += delta; 3218 3219 if (tsk) { 3220 account_scheduler_latency(tsk, delta >> 10, 1); 3221 trace_sched_stat_sleep(tsk, delta); 3222 } 3223 } 3224 if (se->statistics.block_start) { 3225 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.block_start; 3226 3227 if ((s64)delta < 0) 3228 delta = 0; 3229 3230 if (unlikely(delta > se->statistics.block_max)) 3231 se->statistics.block_max = delta; 3232 3233 se->statistics.block_start = 0; 3234 se->statistics.sum_sleep_runtime += delta; 3235 3236 if (tsk) { 3237 if (tsk->in_iowait) { 3238 se->statistics.iowait_sum += delta; 3239 se->statistics.iowait_count++; 3240 trace_sched_stat_iowait(tsk, delta); 3241 } 3242 3243 trace_sched_stat_blocked(tsk, delta); 3244 3245 /* 3246 * Blocking time is in units of nanosecs, so shift by 3247 * 20 to get a milliseconds-range estimation of the 3248 * amount of time that the task spent sleeping: 3249 */ 3250 if (unlikely(prof_on == SLEEP_PROFILING)) { 3251 profile_hits(SLEEP_PROFILING, 3252 (void *)get_wchan(tsk), 3253 delta >> 20); 3254 } 3255 account_scheduler_latency(tsk, delta >> 10, 0); 3256 } 3257 } 3258 #endif 3259 } 3260 3261 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se) 3262 { 3263 #ifdef CONFIG_SCHED_DEBUG 3264 s64 d = se->vruntime - cfs_rq->min_vruntime; 3265 3266 if (d < 0) 3267 d = -d; 3268 3269 if (d > 3*sysctl_sched_latency) 3270 schedstat_inc(cfs_rq, nr_spread_over); 3271 #endif 3272 } 3273 3274 static void 3275 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial) 3276 { 3277 u64 vruntime = cfs_rq->min_vruntime; 3278 3279 /* 3280 * The 'current' period is already promised to the current tasks, 3281 * however the extra weight of the new task will slow them down a 3282 * little, place the new task so that it fits in the slot that 3283 * stays open at the end. 3284 */ 3285 if (initial && sched_feat(START_DEBIT)) 3286 vruntime += sched_vslice(cfs_rq, se); 3287 3288 /* sleeps up to a single latency don't count. */ 3289 if (!initial) { 3290 unsigned long thresh = sysctl_sched_latency; 3291 3292 /* 3293 * Halve their sleep time's effect, to allow 3294 * for a gentler effect of sleepers: 3295 */ 3296 if (sched_feat(GENTLE_FAIR_SLEEPERS)) 3297 thresh >>= 1; 3298 3299 vruntime -= thresh; 3300 } 3301 3302 /* ensure we never gain time by being placed backwards. */ 3303 se->vruntime = max_vruntime(se->vruntime, vruntime); 3304 } 3305 3306 static void check_enqueue_throttle(struct cfs_rq *cfs_rq); 3307 3308 static inline void check_schedstat_required(void) 3309 { 3310 #ifdef CONFIG_SCHEDSTATS 3311 if (schedstat_enabled()) 3312 return; 3313 3314 /* Force schedstat enabled if a dependent tracepoint is active */ 3315 if (trace_sched_stat_wait_enabled() || 3316 trace_sched_stat_sleep_enabled() || 3317 trace_sched_stat_iowait_enabled() || 3318 trace_sched_stat_blocked_enabled() || 3319 trace_sched_stat_runtime_enabled()) { 3320 printk_deferred_once("Scheduler tracepoints stat_sleep, stat_iowait, " 3321 "stat_blocked and stat_runtime require the " 3322 "kernel parameter schedstats=enabled or " 3323 "kernel.sched_schedstats=1\n"); 3324 } 3325 #endif 3326 } 3327 3328 3329 /* 3330 * MIGRATION 3331 * 3332 * dequeue 3333 * update_curr() 3334 * update_min_vruntime() 3335 * vruntime -= min_vruntime 3336 * 3337 * enqueue 3338 * update_curr() 3339 * update_min_vruntime() 3340 * vruntime += min_vruntime 3341 * 3342 * this way the vruntime transition between RQs is done when both 3343 * min_vruntime are up-to-date. 3344 * 3345 * WAKEUP (remote) 3346 * 3347 * ->migrate_task_rq_fair() (p->state == TASK_WAKING) 3348 * vruntime -= min_vruntime 3349 * 3350 * enqueue 3351 * update_curr() 3352 * update_min_vruntime() 3353 * vruntime += min_vruntime 3354 * 3355 * this way we don't have the most up-to-date min_vruntime on the originating 3356 * CPU and an up-to-date min_vruntime on the destination CPU. 3357 */ 3358 3359 static void 3360 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags) 3361 { 3362 bool renorm = !(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_MIGRATED); 3363 bool curr = cfs_rq->curr == se; 3364 3365 /* 3366 * If we're the current task, we must renormalise before calling 3367 * update_curr(). 3368 */ 3369 if (renorm && curr) 3370 se->vruntime += cfs_rq->min_vruntime; 3371 3372 update_curr(cfs_rq); 3373 3374 /* 3375 * Otherwise, renormalise after, such that we're placed at the current 3376 * moment in time, instead of some random moment in the past. Being 3377 * placed in the past could significantly boost this task to the 3378 * fairness detriment of existing tasks. 3379 */ 3380 if (renorm && !curr) 3381 se->vruntime += cfs_rq->min_vruntime; 3382 3383 enqueue_entity_load_avg(cfs_rq, se); 3384 account_entity_enqueue(cfs_rq, se); 3385 update_cfs_shares(cfs_rq); 3386 3387 if (flags & ENQUEUE_WAKEUP) { 3388 place_entity(cfs_rq, se, 0); 3389 if (schedstat_enabled()) 3390 enqueue_sleeper(cfs_rq, se); 3391 } 3392 3393 check_schedstat_required(); 3394 if (schedstat_enabled()) { 3395 update_stats_enqueue(cfs_rq, se); 3396 check_spread(cfs_rq, se); 3397 } 3398 if (!curr) 3399 __enqueue_entity(cfs_rq, se); 3400 se->on_rq = 1; 3401 3402 if (cfs_rq->nr_running == 1) { 3403 list_add_leaf_cfs_rq(cfs_rq); 3404 check_enqueue_throttle(cfs_rq); 3405 } 3406 } 3407 3408 static void __clear_buddies_last(struct sched_entity *se) 3409 { 3410 for_each_sched_entity(se) { 3411 struct cfs_rq *cfs_rq = cfs_rq_of(se); 3412 if (cfs_rq->last != se) 3413 break; 3414 3415 cfs_rq->last = NULL; 3416 } 3417 } 3418 3419 static void __clear_buddies_next(struct sched_entity *se) 3420 { 3421 for_each_sched_entity(se) { 3422 struct cfs_rq *cfs_rq = cfs_rq_of(se); 3423 if (cfs_rq->next != se) 3424 break; 3425 3426 cfs_rq->next = NULL; 3427 } 3428 } 3429 3430 static void __clear_buddies_skip(struct sched_entity *se) 3431 { 3432 for_each_sched_entity(se) { 3433 struct cfs_rq *cfs_rq = cfs_rq_of(se); 3434 if (cfs_rq->skip != se) 3435 break; 3436 3437 cfs_rq->skip = NULL; 3438 } 3439 } 3440 3441 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se) 3442 { 3443 if (cfs_rq->last == se) 3444 __clear_buddies_last(se); 3445 3446 if (cfs_rq->next == se) 3447 __clear_buddies_next(se); 3448 3449 if (cfs_rq->skip == se) 3450 __clear_buddies_skip(se); 3451 } 3452 3453 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq); 3454 3455 static void 3456 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags) 3457 { 3458 /* 3459 * Update run-time statistics of the 'current'. 3460 */ 3461 update_curr(cfs_rq); 3462 dequeue_entity_load_avg(cfs_rq, se); 3463 3464 if (schedstat_enabled()) 3465 update_stats_dequeue(cfs_rq, se, flags); 3466 3467 clear_buddies(cfs_rq, se); 3468 3469 if (se != cfs_rq->curr) 3470 __dequeue_entity(cfs_rq, se); 3471 se->on_rq = 0; 3472 account_entity_dequeue(cfs_rq, se); 3473 3474 /* 3475 * Normalize after update_curr(); which will also have moved 3476 * min_vruntime if @se is the one holding it back. But before doing 3477 * update_min_vruntime() again, which will discount @se's position and 3478 * can move min_vruntime forward still more. 3479 */ 3480 if (!(flags & DEQUEUE_SLEEP)) 3481 se->vruntime -= cfs_rq->min_vruntime; 3482 3483 /* return excess runtime on last dequeue */ 3484 return_cfs_rq_runtime(cfs_rq); 3485 3486 update_cfs_shares(cfs_rq); 3487 3488 /* 3489 * Now advance min_vruntime if @se was the entity holding it back, 3490 * except when: DEQUEUE_SAVE && !DEQUEUE_MOVE, in this case we'll be 3491 * put back on, and if we advance min_vruntime, we'll be placed back 3492 * further than we started -- ie. we'll be penalized. 3493 */ 3494 if ((flags & (DEQUEUE_SAVE | DEQUEUE_MOVE)) == DEQUEUE_SAVE) 3495 update_min_vruntime(cfs_rq); 3496 } 3497 3498 /* 3499 * Preempt the current task with a newly woken task if needed: 3500 */ 3501 static void 3502 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr) 3503 { 3504 unsigned long ideal_runtime, delta_exec; 3505 struct sched_entity *se; 3506 s64 delta; 3507 3508 ideal_runtime = sched_slice(cfs_rq, curr); 3509 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime; 3510 if (delta_exec > ideal_runtime) { 3511 resched_curr(rq_of(cfs_rq)); 3512 /* 3513 * The current task ran long enough, ensure it doesn't get 3514 * re-elected due to buddy favours. 3515 */ 3516 clear_buddies(cfs_rq, curr); 3517 return; 3518 } 3519 3520 /* 3521 * Ensure that a task that missed wakeup preemption by a 3522 * narrow margin doesn't have to wait for a full slice. 3523 * This also mitigates buddy induced latencies under load. 3524 */ 3525 if (delta_exec < sysctl_sched_min_granularity) 3526 return; 3527 3528 se = __pick_first_entity(cfs_rq); 3529 delta = curr->vruntime - se->vruntime; 3530 3531 if (delta < 0) 3532 return; 3533 3534 if (delta > ideal_runtime) 3535 resched_curr(rq_of(cfs_rq)); 3536 } 3537 3538 static void 3539 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se) 3540 { 3541 /* 'current' is not kept within the tree. */ 3542 if (se->on_rq) { 3543 /* 3544 * Any task has to be enqueued before it get to execute on 3545 * a CPU. So account for the time it spent waiting on the 3546 * runqueue. 3547 */ 3548 if (schedstat_enabled()) 3549 update_stats_wait_end(cfs_rq, se); 3550 __dequeue_entity(cfs_rq, se); 3551 update_load_avg(se, 1); 3552 } 3553 3554 update_stats_curr_start(cfs_rq, se); 3555 cfs_rq->curr = se; 3556 #ifdef CONFIG_SCHEDSTATS 3557 /* 3558 * Track our maximum slice length, if the CPU's load is at 3559 * least twice that of our own weight (i.e. dont track it 3560 * when there are only lesser-weight tasks around): 3561 */ 3562 if (schedstat_enabled() && rq_of(cfs_rq)->load.weight >= 2*se->load.weight) { 3563 se->statistics.slice_max = max(se->statistics.slice_max, 3564 se->sum_exec_runtime - se->prev_sum_exec_runtime); 3565 } 3566 #endif 3567 se->prev_sum_exec_runtime = se->sum_exec_runtime; 3568 } 3569 3570 static int 3571 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se); 3572 3573 /* 3574 * Pick the next process, keeping these things in mind, in this order: 3575 * 1) keep things fair between processes/task groups 3576 * 2) pick the "next" process, since someone really wants that to run 3577 * 3) pick the "last" process, for cache locality 3578 * 4) do not run the "skip" process, if something else is available 3579 */ 3580 static struct sched_entity * 3581 pick_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *curr) 3582 { 3583 struct sched_entity *left = __pick_first_entity(cfs_rq); 3584 struct sched_entity *se; 3585 3586 /* 3587 * If curr is set we have to see if its left of the leftmost entity 3588 * still in the tree, provided there was anything in the tree at all. 3589 */ 3590 if (!left || (curr && entity_before(curr, left))) 3591 left = curr; 3592 3593 se = left; /* ideally we run the leftmost entity */ 3594 3595 /* 3596 * Avoid running the skip buddy, if running something else can 3597 * be done without getting too unfair. 3598 */ 3599 if (cfs_rq->skip == se) { 3600 struct sched_entity *second; 3601 3602 if (se == curr) { 3603 second = __pick_first_entity(cfs_rq); 3604 } else { 3605 second = __pick_next_entity(se); 3606 if (!second || (curr && entity_before(curr, second))) 3607 second = curr; 3608 } 3609 3610 if (second && wakeup_preempt_entity(second, left) < 1) 3611 se = second; 3612 } 3613 3614 /* 3615 * Prefer last buddy, try to return the CPU to a preempted task. 3616 */ 3617 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1) 3618 se = cfs_rq->last; 3619 3620 /* 3621 * Someone really wants this to run. If it's not unfair, run it. 3622 */ 3623 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1) 3624 se = cfs_rq->next; 3625 3626 clear_buddies(cfs_rq, se); 3627 3628 return se; 3629 } 3630 3631 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq); 3632 3633 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev) 3634 { 3635 /* 3636 * If still on the runqueue then deactivate_task() 3637 * was not called and update_curr() has to be done: 3638 */ 3639 if (prev->on_rq) 3640 update_curr(cfs_rq); 3641 3642 /* throttle cfs_rqs exceeding runtime */ 3643 check_cfs_rq_runtime(cfs_rq); 3644 3645 if (schedstat_enabled()) { 3646 check_spread(cfs_rq, prev); 3647 if (prev->on_rq) 3648 update_stats_wait_start(cfs_rq, prev); 3649 } 3650 3651 if (prev->on_rq) { 3652 /* Put 'current' back into the tree. */ 3653 __enqueue_entity(cfs_rq, prev); 3654 /* in !on_rq case, update occurred at dequeue */ 3655 update_load_avg(prev, 0); 3656 } 3657 cfs_rq->curr = NULL; 3658 } 3659 3660 static void 3661 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued) 3662 { 3663 /* 3664 * Update run-time statistics of the 'current'. 3665 */ 3666 update_curr(cfs_rq); 3667 3668 /* 3669 * Ensure that runnable average is periodically updated. 3670 */ 3671 update_load_avg(curr, 1); 3672 update_cfs_shares(cfs_rq); 3673 3674 #ifdef CONFIG_SCHED_HRTICK 3675 /* 3676 * queued ticks are scheduled to match the slice, so don't bother 3677 * validating it and just reschedule. 3678 */ 3679 if (queued) { 3680 resched_curr(rq_of(cfs_rq)); 3681 return; 3682 } 3683 /* 3684 * don't let the period tick interfere with the hrtick preemption 3685 */ 3686 if (!sched_feat(DOUBLE_TICK) && 3687 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer)) 3688 return; 3689 #endif 3690 3691 if (cfs_rq->nr_running > 1) 3692 check_preempt_tick(cfs_rq, curr); 3693 } 3694 3695 3696 /************************************************** 3697 * CFS bandwidth control machinery 3698 */ 3699 3700 #ifdef CONFIG_CFS_BANDWIDTH 3701 3702 #ifdef HAVE_JUMP_LABEL 3703 static struct static_key __cfs_bandwidth_used; 3704 3705 static inline bool cfs_bandwidth_used(void) 3706 { 3707 return static_key_false(&__cfs_bandwidth_used); 3708 } 3709 3710 void cfs_bandwidth_usage_inc(void) 3711 { 3712 static_key_slow_inc(&__cfs_bandwidth_used); 3713 } 3714 3715 void cfs_bandwidth_usage_dec(void) 3716 { 3717 static_key_slow_dec(&__cfs_bandwidth_used); 3718 } 3719 #else /* HAVE_JUMP_LABEL */ 3720 static bool cfs_bandwidth_used(void) 3721 { 3722 return true; 3723 } 3724 3725 void cfs_bandwidth_usage_inc(void) {} 3726 void cfs_bandwidth_usage_dec(void) {} 3727 #endif /* HAVE_JUMP_LABEL */ 3728 3729 /* 3730 * default period for cfs group bandwidth. 3731 * default: 0.1s, units: nanoseconds 3732 */ 3733 static inline u64 default_cfs_period(void) 3734 { 3735 return 100000000ULL; 3736 } 3737 3738 static inline u64 sched_cfs_bandwidth_slice(void) 3739 { 3740 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC; 3741 } 3742 3743 /* 3744 * Replenish runtime according to assigned quota and update expiration time. 3745 * We use sched_clock_cpu directly instead of rq->clock to avoid adding 3746 * additional synchronization around rq->lock. 3747 * 3748 * requires cfs_b->lock 3749 */ 3750 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b) 3751 { 3752 u64 now; 3753 3754 if (cfs_b->quota == RUNTIME_INF) 3755 return; 3756 3757 now = sched_clock_cpu(smp_processor_id()); 3758 cfs_b->runtime = cfs_b->quota; 3759 cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period); 3760 } 3761 3762 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg) 3763 { 3764 return &tg->cfs_bandwidth; 3765 } 3766 3767 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */ 3768 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq) 3769 { 3770 if (unlikely(cfs_rq->throttle_count)) 3771 return cfs_rq->throttled_clock_task - cfs_rq->throttled_clock_task_time; 3772 3773 return rq_clock_task(rq_of(cfs_rq)) - cfs_rq->throttled_clock_task_time; 3774 } 3775 3776 /* returns 0 on failure to allocate runtime */ 3777 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq) 3778 { 3779 struct task_group *tg = cfs_rq->tg; 3780 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg); 3781 u64 amount = 0, min_amount, expires; 3782 3783 /* note: this is a positive sum as runtime_remaining <= 0 */ 3784 min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining; 3785 3786 raw_spin_lock(&cfs_b->lock); 3787 if (cfs_b->quota == RUNTIME_INF) 3788 amount = min_amount; 3789 else { 3790 start_cfs_bandwidth(cfs_b); 3791 3792 if (cfs_b->runtime > 0) { 3793 amount = min(cfs_b->runtime, min_amount); 3794 cfs_b->runtime -= amount; 3795 cfs_b->idle = 0; 3796 } 3797 } 3798 expires = cfs_b->runtime_expires; 3799 raw_spin_unlock(&cfs_b->lock); 3800 3801 cfs_rq->runtime_remaining += amount; 3802 /* 3803 * we may have advanced our local expiration to account for allowed 3804 * spread between our sched_clock and the one on which runtime was 3805 * issued. 3806 */ 3807 if ((s64)(expires - cfs_rq->runtime_expires) > 0) 3808 cfs_rq->runtime_expires = expires; 3809 3810 return cfs_rq->runtime_remaining > 0; 3811 } 3812 3813 /* 3814 * Note: This depends on the synchronization provided by sched_clock and the 3815 * fact that rq->clock snapshots this value. 3816 */ 3817 static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq) 3818 { 3819 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg); 3820 3821 /* if the deadline is ahead of our clock, nothing to do */ 3822 if (likely((s64)(rq_clock(rq_of(cfs_rq)) - cfs_rq->runtime_expires) < 0)) 3823 return; 3824 3825 if (cfs_rq->runtime_remaining < 0) 3826 return; 3827 3828 /* 3829 * If the local deadline has passed we have to consider the 3830 * possibility that our sched_clock is 'fast' and the global deadline 3831 * has not truly expired. 3832 * 3833 * Fortunately we can check determine whether this the case by checking 3834 * whether the global deadline has advanced. It is valid to compare 3835 * cfs_b->runtime_expires without any locks since we only care about 3836 * exact equality, so a partial write will still work. 3837 */ 3838 3839 if (cfs_rq->runtime_expires != cfs_b->runtime_expires) { 3840 /* extend local deadline, drift is bounded above by 2 ticks */ 3841 cfs_rq->runtime_expires += TICK_NSEC; 3842 } else { 3843 /* global deadline is ahead, expiration has passed */ 3844 cfs_rq->runtime_remaining = 0; 3845 } 3846 } 3847 3848 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) 3849 { 3850 /* dock delta_exec before expiring quota (as it could span periods) */ 3851 cfs_rq->runtime_remaining -= delta_exec; 3852 expire_cfs_rq_runtime(cfs_rq); 3853 3854 if (likely(cfs_rq->runtime_remaining > 0)) 3855 return; 3856 3857 /* 3858 * if we're unable to extend our runtime we resched so that the active 3859 * hierarchy can be throttled 3860 */ 3861 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr)) 3862 resched_curr(rq_of(cfs_rq)); 3863 } 3864 3865 static __always_inline 3866 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) 3867 { 3868 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled) 3869 return; 3870 3871 __account_cfs_rq_runtime(cfs_rq, delta_exec); 3872 } 3873 3874 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq) 3875 { 3876 return cfs_bandwidth_used() && cfs_rq->throttled; 3877 } 3878 3879 /* check whether cfs_rq, or any parent, is throttled */ 3880 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq) 3881 { 3882 return cfs_bandwidth_used() && cfs_rq->throttle_count; 3883 } 3884 3885 /* 3886 * Ensure that neither of the group entities corresponding to src_cpu or 3887 * dest_cpu are members of a throttled hierarchy when performing group 3888 * load-balance operations. 3889 */ 3890 static inline int throttled_lb_pair(struct task_group *tg, 3891 int src_cpu, int dest_cpu) 3892 { 3893 struct cfs_rq *src_cfs_rq, *dest_cfs_rq; 3894 3895 src_cfs_rq = tg->cfs_rq[src_cpu]; 3896 dest_cfs_rq = tg->cfs_rq[dest_cpu]; 3897 3898 return throttled_hierarchy(src_cfs_rq) || 3899 throttled_hierarchy(dest_cfs_rq); 3900 } 3901 3902 /* updated child weight may affect parent so we have to do this bottom up */ 3903 static int tg_unthrottle_up(struct task_group *tg, void *data) 3904 { 3905 struct rq *rq = data; 3906 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)]; 3907 3908 cfs_rq->throttle_count--; 3909 if (!cfs_rq->throttle_count) { 3910 /* adjust cfs_rq_clock_task() */ 3911 cfs_rq->throttled_clock_task_time += rq_clock_task(rq) - 3912 cfs_rq->throttled_clock_task; 3913 } 3914 3915 return 0; 3916 } 3917 3918 static int tg_throttle_down(struct task_group *tg, void *data) 3919 { 3920 struct rq *rq = data; 3921 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)]; 3922 3923 /* group is entering throttled state, stop time */ 3924 if (!cfs_rq->throttle_count) 3925 cfs_rq->throttled_clock_task = rq_clock_task(rq); 3926 cfs_rq->throttle_count++; 3927 3928 return 0; 3929 } 3930 3931 static void throttle_cfs_rq(struct cfs_rq *cfs_rq) 3932 { 3933 struct rq *rq = rq_of(cfs_rq); 3934 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg); 3935 struct sched_entity *se; 3936 long task_delta, dequeue = 1; 3937 bool empty; 3938 3939 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))]; 3940 3941 /* freeze hierarchy runnable averages while throttled */ 3942 rcu_read_lock(); 3943 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq); 3944 rcu_read_unlock(); 3945 3946 task_delta = cfs_rq->h_nr_running; 3947 for_each_sched_entity(se) { 3948 struct cfs_rq *qcfs_rq = cfs_rq_of(se); 3949 /* throttled entity or throttle-on-deactivate */ 3950 if (!se->on_rq) 3951 break; 3952 3953 if (dequeue) 3954 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP); 3955 qcfs_rq->h_nr_running -= task_delta; 3956 3957 if (qcfs_rq->load.weight) 3958 dequeue = 0; 3959 } 3960 3961 if (!se) 3962 sub_nr_running(rq, task_delta); 3963 3964 cfs_rq->throttled = 1; 3965 cfs_rq->throttled_clock = rq_clock(rq); 3966 raw_spin_lock(&cfs_b->lock); 3967 empty = list_empty(&cfs_b->throttled_cfs_rq); 3968 3969 /* 3970 * Add to the _head_ of the list, so that an already-started 3971 * distribute_cfs_runtime will not see us 3972 */ 3973 list_add_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq); 3974 3975 /* 3976 * If we're the first throttled task, make sure the bandwidth 3977 * timer is running. 3978 */ 3979 if (empty) 3980 start_cfs_bandwidth(cfs_b); 3981 3982 raw_spin_unlock(&cfs_b->lock); 3983 } 3984 3985 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq) 3986 { 3987 struct rq *rq = rq_of(cfs_rq); 3988 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg); 3989 struct sched_entity *se; 3990 int enqueue = 1; 3991 long task_delta; 3992 3993 se = cfs_rq->tg->se[cpu_of(rq)]; 3994 3995 cfs_rq->throttled = 0; 3996 3997 update_rq_clock(rq); 3998 3999 raw_spin_lock(&cfs_b->lock); 4000 cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock; 4001 list_del_rcu(&cfs_rq->throttled_list); 4002 raw_spin_unlock(&cfs_b->lock); 4003 4004 /* update hierarchical throttle state */ 4005 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq); 4006 4007 if (!cfs_rq->load.weight) 4008 return; 4009 4010 task_delta = cfs_rq->h_nr_running; 4011 for_each_sched_entity(se) { 4012 if (se->on_rq) 4013 enqueue = 0; 4014 4015 cfs_rq = cfs_rq_of(se); 4016 if (enqueue) 4017 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP); 4018 cfs_rq->h_nr_running += task_delta; 4019 4020 if (cfs_rq_throttled(cfs_rq)) 4021 break; 4022 } 4023 4024 if (!se) 4025 add_nr_running(rq, task_delta); 4026 4027 /* determine whether we need to wake up potentially idle cpu */ 4028 if (rq->curr == rq->idle && rq->cfs.nr_running) 4029 resched_curr(rq); 4030 } 4031 4032 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b, 4033 u64 remaining, u64 expires) 4034 { 4035 struct cfs_rq *cfs_rq; 4036 u64 runtime; 4037 u64 starting_runtime = remaining; 4038 4039 rcu_read_lock(); 4040 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq, 4041 throttled_list) { 4042 struct rq *rq = rq_of(cfs_rq); 4043 4044 raw_spin_lock(&rq->lock); 4045 if (!cfs_rq_throttled(cfs_rq)) 4046 goto next; 4047 4048 runtime = -cfs_rq->runtime_remaining + 1; 4049 if (runtime > remaining) 4050 runtime = remaining; 4051 remaining -= runtime; 4052 4053 cfs_rq->runtime_remaining += runtime; 4054 cfs_rq->runtime_expires = expires; 4055 4056 /* we check whether we're throttled above */ 4057 if (cfs_rq->runtime_remaining > 0) 4058 unthrottle_cfs_rq(cfs_rq); 4059 4060 next: 4061 raw_spin_unlock(&rq->lock); 4062 4063 if (!remaining) 4064 break; 4065 } 4066 rcu_read_unlock(); 4067 4068 return starting_runtime - remaining; 4069 } 4070 4071 /* 4072 * Responsible for refilling a task_group's bandwidth and unthrottling its 4073 * cfs_rqs as appropriate. If there has been no activity within the last 4074 * period the timer is deactivated until scheduling resumes; cfs_b->idle is 4075 * used to track this state. 4076 */ 4077 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun) 4078 { 4079 u64 runtime, runtime_expires; 4080 int throttled; 4081 4082 /* no need to continue the timer with no bandwidth constraint */ 4083 if (cfs_b->quota == RUNTIME_INF) 4084 goto out_deactivate; 4085 4086 throttled = !list_empty(&cfs_b->throttled_cfs_rq); 4087 cfs_b->nr_periods += overrun; 4088 4089 /* 4090 * idle depends on !throttled (for the case of a large deficit), and if 4091 * we're going inactive then everything else can be deferred 4092 */ 4093 if (cfs_b->idle && !throttled) 4094 goto out_deactivate; 4095 4096 __refill_cfs_bandwidth_runtime(cfs_b); 4097 4098 if (!throttled) { 4099 /* mark as potentially idle for the upcoming period */ 4100 cfs_b->idle = 1; 4101 return 0; 4102 } 4103 4104 /* account preceding periods in which throttling occurred */ 4105 cfs_b->nr_throttled += overrun; 4106 4107 runtime_expires = cfs_b->runtime_expires; 4108 4109 /* 4110 * This check is repeated as we are holding onto the new bandwidth while 4111 * we unthrottle. This can potentially race with an unthrottled group 4112 * trying to acquire new bandwidth from the global pool. This can result 4113 * in us over-using our runtime if it is all used during this loop, but 4114 * only by limited amounts in that extreme case. 4115 */ 4116 while (throttled && cfs_b->runtime > 0) { 4117 runtime = cfs_b->runtime; 4118 raw_spin_unlock(&cfs_b->lock); 4119 /* we can't nest cfs_b->lock while distributing bandwidth */ 4120 runtime = distribute_cfs_runtime(cfs_b, runtime, 4121 runtime_expires); 4122 raw_spin_lock(&cfs_b->lock); 4123 4124 throttled = !list_empty(&cfs_b->throttled_cfs_rq); 4125 4126 cfs_b->runtime -= min(runtime, cfs_b->runtime); 4127 } 4128 4129 /* 4130 * While we are ensured activity in the period following an 4131 * unthrottle, this also covers the case in which the new bandwidth is 4132 * insufficient to cover the existing bandwidth deficit. (Forcing the 4133 * timer to remain active while there are any throttled entities.) 4134 */ 4135 cfs_b->idle = 0; 4136 4137 return 0; 4138 4139 out_deactivate: 4140 return 1; 4141 } 4142 4143 /* a cfs_rq won't donate quota below this amount */ 4144 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC; 4145 /* minimum remaining period time to redistribute slack quota */ 4146 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC; 4147 /* how long we wait to gather additional slack before distributing */ 4148 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC; 4149 4150 /* 4151 * Are we near the end of the current quota period? 4152 * 4153 * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the 4154 * hrtimer base being cleared by hrtimer_start. In the case of 4155 * migrate_hrtimers, base is never cleared, so we are fine. 4156 */ 4157 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire) 4158 { 4159 struct hrtimer *refresh_timer = &cfs_b->period_timer; 4160 u64 remaining; 4161 4162 /* if the call-back is running a quota refresh is already occurring */ 4163 if (hrtimer_callback_running(refresh_timer)) 4164 return 1; 4165 4166 /* is a quota refresh about to occur? */ 4167 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer)); 4168 if (remaining < min_expire) 4169 return 1; 4170 4171 return 0; 4172 } 4173 4174 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b) 4175 { 4176 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration; 4177 4178 /* if there's a quota refresh soon don't bother with slack */ 4179 if (runtime_refresh_within(cfs_b, min_left)) 4180 return; 4181 4182 hrtimer_start(&cfs_b->slack_timer, 4183 ns_to_ktime(cfs_bandwidth_slack_period), 4184 HRTIMER_MODE_REL); 4185 } 4186 4187 /* we know any runtime found here is valid as update_curr() precedes return */ 4188 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq) 4189 { 4190 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg); 4191 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime; 4192 4193 if (slack_runtime <= 0) 4194 return; 4195 4196 raw_spin_lock(&cfs_b->lock); 4197 if (cfs_b->quota != RUNTIME_INF && 4198 cfs_rq->runtime_expires == cfs_b->runtime_expires) { 4199 cfs_b->runtime += slack_runtime; 4200 4201 /* we are under rq->lock, defer unthrottling using a timer */ 4202 if (cfs_b->runtime > sched_cfs_bandwidth_slice() && 4203 !list_empty(&cfs_b->throttled_cfs_rq)) 4204 start_cfs_slack_bandwidth(cfs_b); 4205 } 4206 raw_spin_unlock(&cfs_b->lock); 4207 4208 /* even if it's not valid for return we don't want to try again */ 4209 cfs_rq->runtime_remaining -= slack_runtime; 4210 } 4211 4212 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) 4213 { 4214 if (!cfs_bandwidth_used()) 4215 return; 4216 4217 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running) 4218 return; 4219 4220 __return_cfs_rq_runtime(cfs_rq); 4221 } 4222 4223 /* 4224 * This is done with a timer (instead of inline with bandwidth return) since 4225 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs. 4226 */ 4227 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b) 4228 { 4229 u64 runtime = 0, slice = sched_cfs_bandwidth_slice(); 4230 u64 expires; 4231 4232 /* confirm we're still not at a refresh boundary */ 4233 raw_spin_lock(&cfs_b->lock); 4234 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) { 4235 raw_spin_unlock(&cfs_b->lock); 4236 return; 4237 } 4238 4239 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice) 4240 runtime = cfs_b->runtime; 4241 4242 expires = cfs_b->runtime_expires; 4243 raw_spin_unlock(&cfs_b->lock); 4244 4245 if (!runtime) 4246 return; 4247 4248 runtime = distribute_cfs_runtime(cfs_b, runtime, expires); 4249 4250 raw_spin_lock(&cfs_b->lock); 4251 if (expires == cfs_b->runtime_expires) 4252 cfs_b->runtime -= min(runtime, cfs_b->runtime); 4253 raw_spin_unlock(&cfs_b->lock); 4254 } 4255 4256 /* 4257 * When a group wakes up we want to make sure that its quota is not already 4258 * expired/exceeded, otherwise it may be allowed to steal additional ticks of 4259 * runtime as update_curr() throttling can not not trigger until it's on-rq. 4260 */ 4261 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) 4262 { 4263 if (!cfs_bandwidth_used()) 4264 return; 4265 4266 /* an active group must be handled by the update_curr()->put() path */ 4267 if (!cfs_rq->runtime_enabled || cfs_rq->curr) 4268 return; 4269 4270 /* ensure the group is not already throttled */ 4271 if (cfs_rq_throttled(cfs_rq)) 4272 return; 4273 4274 /* update runtime allocation */ 4275 account_cfs_rq_runtime(cfs_rq, 0); 4276 if (cfs_rq->runtime_remaining <= 0) 4277 throttle_cfs_rq(cfs_rq); 4278 } 4279 4280 static void sync_throttle(struct task_group *tg, int cpu) 4281 { 4282 struct cfs_rq *pcfs_rq, *cfs_rq; 4283 4284 if (!cfs_bandwidth_used()) 4285 return; 4286 4287 if (!tg->parent) 4288 return; 4289 4290 cfs_rq = tg->cfs_rq[cpu]; 4291 pcfs_rq = tg->parent->cfs_rq[cpu]; 4292 4293 cfs_rq->throttle_count = pcfs_rq->throttle_count; 4294 cfs_rq->throttled_clock_task = rq_clock_task(cpu_rq(cpu)); 4295 } 4296 4297 /* conditionally throttle active cfs_rq's from put_prev_entity() */ 4298 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) 4299 { 4300 if (!cfs_bandwidth_used()) 4301 return false; 4302 4303 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0)) 4304 return false; 4305 4306 /* 4307 * it's possible for a throttled entity to be forced into a running 4308 * state (e.g. set_curr_task), in this case we're finished. 4309 */ 4310 if (cfs_rq_throttled(cfs_rq)) 4311 return true; 4312 4313 throttle_cfs_rq(cfs_rq); 4314 return true; 4315 } 4316 4317 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer) 4318 { 4319 struct cfs_bandwidth *cfs_b = 4320 container_of(timer, struct cfs_bandwidth, slack_timer); 4321 4322 do_sched_cfs_slack_timer(cfs_b); 4323 4324 return HRTIMER_NORESTART; 4325 } 4326 4327 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer) 4328 { 4329 struct cfs_bandwidth *cfs_b = 4330 container_of(timer, struct cfs_bandwidth, period_timer); 4331 int overrun; 4332 int idle = 0; 4333 4334 raw_spin_lock(&cfs_b->lock); 4335 for (;;) { 4336 overrun = hrtimer_forward_now(timer, cfs_b->period); 4337 if (!overrun) 4338 break; 4339 4340 idle = do_sched_cfs_period_timer(cfs_b, overrun); 4341 } 4342 if (idle) 4343 cfs_b->period_active = 0; 4344 raw_spin_unlock(&cfs_b->lock); 4345 4346 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART; 4347 } 4348 4349 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) 4350 { 4351 raw_spin_lock_init(&cfs_b->lock); 4352 cfs_b->runtime = 0; 4353 cfs_b->quota = RUNTIME_INF; 4354 cfs_b->period = ns_to_ktime(default_cfs_period()); 4355 4356 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq); 4357 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_PINNED); 4358 cfs_b->period_timer.function = sched_cfs_period_timer; 4359 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); 4360 cfs_b->slack_timer.function = sched_cfs_slack_timer; 4361 } 4362 4363 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) 4364 { 4365 cfs_rq->runtime_enabled = 0; 4366 INIT_LIST_HEAD(&cfs_rq->throttled_list); 4367 } 4368 4369 void start_cfs_bandwidth(struct cfs_bandwidth *cfs_b) 4370 { 4371 lockdep_assert_held(&cfs_b->lock); 4372 4373 if (!cfs_b->period_active) { 4374 cfs_b->period_active = 1; 4375 hrtimer_forward_now(&cfs_b->period_timer, cfs_b->period); 4376 hrtimer_start_expires(&cfs_b->period_timer, HRTIMER_MODE_ABS_PINNED); 4377 } 4378 } 4379 4380 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) 4381 { 4382 /* init_cfs_bandwidth() was not called */ 4383 if (!cfs_b->throttled_cfs_rq.next) 4384 return; 4385 4386 hrtimer_cancel(&cfs_b->period_timer); 4387 hrtimer_cancel(&cfs_b->slack_timer); 4388 } 4389 4390 static void __maybe_unused update_runtime_enabled(struct rq *rq) 4391 { 4392 struct cfs_rq *cfs_rq; 4393 4394 for_each_leaf_cfs_rq(rq, cfs_rq) { 4395 struct cfs_bandwidth *cfs_b = &cfs_rq->tg->cfs_bandwidth; 4396 4397 raw_spin_lock(&cfs_b->lock); 4398 cfs_rq->runtime_enabled = cfs_b->quota != RUNTIME_INF; 4399 raw_spin_unlock(&cfs_b->lock); 4400 } 4401 } 4402 4403 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq) 4404 { 4405 struct cfs_rq *cfs_rq; 4406 4407 for_each_leaf_cfs_rq(rq, cfs_rq) { 4408 if (!cfs_rq->runtime_enabled) 4409 continue; 4410 4411 /* 4412 * clock_task is not advancing so we just need to make sure 4413 * there's some valid quota amount 4414 */ 4415 cfs_rq->runtime_remaining = 1; 4416 /* 4417 * Offline rq is schedulable till cpu is completely disabled 4418 * in take_cpu_down(), so we prevent new cfs throttling here. 4419 */ 4420 cfs_rq->runtime_enabled = 0; 4421 4422 if (cfs_rq_throttled(cfs_rq)) 4423 unthrottle_cfs_rq(cfs_rq); 4424 } 4425 } 4426 4427 #else /* CONFIG_CFS_BANDWIDTH */ 4428 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq) 4429 { 4430 return rq_clock_task(rq_of(cfs_rq)); 4431 } 4432 4433 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) {} 4434 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) { return false; } 4435 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {} 4436 static inline void sync_throttle(struct task_group *tg, int cpu) {} 4437 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {} 4438 4439 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq) 4440 { 4441 return 0; 4442 } 4443 4444 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq) 4445 { 4446 return 0; 4447 } 4448 4449 static inline int throttled_lb_pair(struct task_group *tg, 4450 int src_cpu, int dest_cpu) 4451 { 4452 return 0; 4453 } 4454 4455 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {} 4456 4457 #ifdef CONFIG_FAIR_GROUP_SCHED 4458 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {} 4459 #endif 4460 4461 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg) 4462 { 4463 return NULL; 4464 } 4465 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {} 4466 static inline void update_runtime_enabled(struct rq *rq) {} 4467 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {} 4468 4469 #endif /* CONFIG_CFS_BANDWIDTH */ 4470 4471 /************************************************** 4472 * CFS operations on tasks: 4473 */ 4474 4475 #ifdef CONFIG_SCHED_HRTICK 4476 static void hrtick_start_fair(struct rq *rq, struct task_struct *p) 4477 { 4478 struct sched_entity *se = &p->se; 4479 struct cfs_rq *cfs_rq = cfs_rq_of(se); 4480 4481 WARN_ON(task_rq(p) != rq); 4482 4483 if (cfs_rq->nr_running > 1) { 4484 u64 slice = sched_slice(cfs_rq, se); 4485 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime; 4486 s64 delta = slice - ran; 4487 4488 if (delta < 0) { 4489 if (rq->curr == p) 4490 resched_curr(rq); 4491 return; 4492 } 4493 hrtick_start(rq, delta); 4494 } 4495 } 4496 4497 /* 4498 * called from enqueue/dequeue and updates the hrtick when the 4499 * current task is from our class and nr_running is low enough 4500 * to matter. 4501 */ 4502 static void hrtick_update(struct rq *rq) 4503 { 4504 struct task_struct *curr = rq->curr; 4505 4506 if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class) 4507 return; 4508 4509 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency) 4510 hrtick_start_fair(rq, curr); 4511 } 4512 #else /* !CONFIG_SCHED_HRTICK */ 4513 static inline void 4514 hrtick_start_fair(struct rq *rq, struct task_struct *p) 4515 { 4516 } 4517 4518 static inline void hrtick_update(struct rq *rq) 4519 { 4520 } 4521 #endif 4522 4523 /* 4524 * The enqueue_task method is called before nr_running is 4525 * increased. Here we update the fair scheduling stats and 4526 * then put the task into the rbtree: 4527 */ 4528 static void 4529 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags) 4530 { 4531 struct cfs_rq *cfs_rq; 4532 struct sched_entity *se = &p->se; 4533 4534 for_each_sched_entity(se) { 4535 if (se->on_rq) 4536 break; 4537 cfs_rq = cfs_rq_of(se); 4538 enqueue_entity(cfs_rq, se, flags); 4539 4540 /* 4541 * end evaluation on encountering a throttled cfs_rq 4542 * 4543 * note: in the case of encountering a throttled cfs_rq we will 4544 * post the final h_nr_running increment below. 4545 */ 4546 if (cfs_rq_throttled(cfs_rq)) 4547 break; 4548 cfs_rq->h_nr_running++; 4549 4550 flags = ENQUEUE_WAKEUP; 4551 } 4552 4553 for_each_sched_entity(se) { 4554 cfs_rq = cfs_rq_of(se); 4555 cfs_rq->h_nr_running++; 4556 4557 if (cfs_rq_throttled(cfs_rq)) 4558 break; 4559 4560 update_load_avg(se, 1); 4561 update_cfs_shares(cfs_rq); 4562 } 4563 4564 if (!se) 4565 add_nr_running(rq, 1); 4566 4567 hrtick_update(rq); 4568 } 4569 4570 static void set_next_buddy(struct sched_entity *se); 4571 4572 /* 4573 * The dequeue_task method is called before nr_running is 4574 * decreased. We remove the task from the rbtree and 4575 * update the fair scheduling stats: 4576 */ 4577 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags) 4578 { 4579 struct cfs_rq *cfs_rq; 4580 struct sched_entity *se = &p->se; 4581 int task_sleep = flags & DEQUEUE_SLEEP; 4582 4583 for_each_sched_entity(se) { 4584 cfs_rq = cfs_rq_of(se); 4585 dequeue_entity(cfs_rq, se, flags); 4586 4587 /* 4588 * end evaluation on encountering a throttled cfs_rq 4589 * 4590 * note: in the case of encountering a throttled cfs_rq we will 4591 * post the final h_nr_running decrement below. 4592 */ 4593 if (cfs_rq_throttled(cfs_rq)) 4594 break; 4595 cfs_rq->h_nr_running--; 4596 4597 /* Don't dequeue parent if it has other entities besides us */ 4598 if (cfs_rq->load.weight) { 4599 /* Avoid re-evaluating load for this entity: */ 4600 se = parent_entity(se); 4601 /* 4602 * Bias pick_next to pick a task from this cfs_rq, as 4603 * p is sleeping when it is within its sched_slice. 4604 */ 4605 if (task_sleep && se && !throttled_hierarchy(cfs_rq)) 4606 set_next_buddy(se); 4607 break; 4608 } 4609 flags |= DEQUEUE_SLEEP; 4610 } 4611 4612 for_each_sched_entity(se) { 4613 cfs_rq = cfs_rq_of(se); 4614 cfs_rq->h_nr_running--; 4615 4616 if (cfs_rq_throttled(cfs_rq)) 4617 break; 4618 4619 update_load_avg(se, 1); 4620 update_cfs_shares(cfs_rq); 4621 } 4622 4623 if (!se) 4624 sub_nr_running(rq, 1); 4625 4626 hrtick_update(rq); 4627 } 4628 4629 #ifdef CONFIG_SMP 4630 #ifdef CONFIG_NO_HZ_COMMON 4631 /* 4632 * per rq 'load' arrray crap; XXX kill this. 4633 */ 4634 4635 /* 4636 * The exact cpuload calculated at every tick would be: 4637 * 4638 * load' = (1 - 1/2^i) * load + (1/2^i) * cur_load 4639 * 4640 * If a cpu misses updates for n ticks (as it was idle) and update gets 4641 * called on the n+1-th tick when cpu may be busy, then we have: 4642 * 4643 * load_n = (1 - 1/2^i)^n * load_0 4644 * load_n+1 = (1 - 1/2^i) * load_n + (1/2^i) * cur_load 4645 * 4646 * decay_load_missed() below does efficient calculation of 4647 * 4648 * load' = (1 - 1/2^i)^n * load 4649 * 4650 * Because x^(n+m) := x^n * x^m we can decompose any x^n in power-of-2 factors. 4651 * This allows us to precompute the above in said factors, thereby allowing the 4652 * reduction of an arbitrary n in O(log_2 n) steps. (See also 4653 * fixed_power_int()) 4654 * 4655 * The calculation is approximated on a 128 point scale. 4656 */ 4657 #define DEGRADE_SHIFT 7 4658 4659 static const u8 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128}; 4660 static const u8 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = { 4661 { 0, 0, 0, 0, 0, 0, 0, 0 }, 4662 { 64, 32, 8, 0, 0, 0, 0, 0 }, 4663 { 96, 72, 40, 12, 1, 0, 0, 0 }, 4664 { 112, 98, 75, 43, 15, 1, 0, 0 }, 4665 { 120, 112, 98, 76, 45, 16, 2, 0 } 4666 }; 4667 4668 /* 4669 * Update cpu_load for any missed ticks, due to tickless idle. The backlog 4670 * would be when CPU is idle and so we just decay the old load without 4671 * adding any new load. 4672 */ 4673 static unsigned long 4674 decay_load_missed(unsigned long load, unsigned long missed_updates, int idx) 4675 { 4676 int j = 0; 4677 4678 if (!missed_updates) 4679 return load; 4680 4681 if (missed_updates >= degrade_zero_ticks[idx]) 4682 return 0; 4683 4684 if (idx == 1) 4685 return load >> missed_updates; 4686 4687 while (missed_updates) { 4688 if (missed_updates % 2) 4689 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT; 4690 4691 missed_updates >>= 1; 4692 j++; 4693 } 4694 return load; 4695 } 4696 #endif /* CONFIG_NO_HZ_COMMON */ 4697 4698 /** 4699 * __cpu_load_update - update the rq->cpu_load[] statistics 4700 * @this_rq: The rq to update statistics for 4701 * @this_load: The current load 4702 * @pending_updates: The number of missed updates 4703 * 4704 * Update rq->cpu_load[] statistics. This function is usually called every 4705 * scheduler tick (TICK_NSEC). 4706 * 4707 * This function computes a decaying average: 4708 * 4709 * load[i]' = (1 - 1/2^i) * load[i] + (1/2^i) * load 4710 * 4711 * Because of NOHZ it might not get called on every tick which gives need for 4712 * the @pending_updates argument. 4713 * 4714 * load[i]_n = (1 - 1/2^i) * load[i]_n-1 + (1/2^i) * load_n-1 4715 * = A * load[i]_n-1 + B ; A := (1 - 1/2^i), B := (1/2^i) * load 4716 * = A * (A * load[i]_n-2 + B) + B 4717 * = A * (A * (A * load[i]_n-3 + B) + B) + B 4718 * = A^3 * load[i]_n-3 + (A^2 + A + 1) * B 4719 * = A^n * load[i]_0 + (A^(n-1) + A^(n-2) + ... + 1) * B 4720 * = A^n * load[i]_0 + ((1 - A^n) / (1 - A)) * B 4721 * = (1 - 1/2^i)^n * (load[i]_0 - load) + load 4722 * 4723 * In the above we've assumed load_n := load, which is true for NOHZ_FULL as 4724 * any change in load would have resulted in the tick being turned back on. 4725 * 4726 * For regular NOHZ, this reduces to: 4727 * 4728 * load[i]_n = (1 - 1/2^i)^n * load[i]_0 4729 * 4730 * see decay_load_misses(). For NOHZ_FULL we get to subtract and add the extra 4731 * term. 4732 */ 4733 static void cpu_load_update(struct rq *this_rq, unsigned long this_load, 4734 unsigned long pending_updates) 4735 { 4736 unsigned long __maybe_unused tickless_load = this_rq->cpu_load[0]; 4737 int i, scale; 4738 4739 this_rq->nr_load_updates++; 4740 4741 /* Update our load: */ 4742 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */ 4743 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) { 4744 unsigned long old_load, new_load; 4745 4746 /* scale is effectively 1 << i now, and >> i divides by scale */ 4747 4748 old_load = this_rq->cpu_load[i]; 4749 #ifdef CONFIG_NO_HZ_COMMON 4750 old_load = decay_load_missed(old_load, pending_updates - 1, i); 4751 if (tickless_load) { 4752 old_load -= decay_load_missed(tickless_load, pending_updates - 1, i); 4753 /* 4754 * old_load can never be a negative value because a 4755 * decayed tickless_load cannot be greater than the 4756 * original tickless_load. 4757 */ 4758 old_load += tickless_load; 4759 } 4760 #endif 4761 new_load = this_load; 4762 /* 4763 * Round up the averaging division if load is increasing. This 4764 * prevents us from getting stuck on 9 if the load is 10, for 4765 * example. 4766 */ 4767 if (new_load > old_load) 4768 new_load += scale - 1; 4769 4770 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i; 4771 } 4772 4773 sched_avg_update(this_rq); 4774 } 4775 4776 /* Used instead of source_load when we know the type == 0 */ 4777 static unsigned long weighted_cpuload(const int cpu) 4778 { 4779 return cfs_rq_runnable_load_avg(&cpu_rq(cpu)->cfs); 4780 } 4781 4782 #ifdef CONFIG_NO_HZ_COMMON 4783 /* 4784 * There is no sane way to deal with nohz on smp when using jiffies because the 4785 * cpu doing the jiffies update might drift wrt the cpu doing the jiffy reading 4786 * causing off-by-one errors in observed deltas; {0,2} instead of {1,1}. 4787 * 4788 * Therefore we need to avoid the delta approach from the regular tick when 4789 * possible since that would seriously skew the load calculation. This is why we 4790 * use cpu_load_update_periodic() for CPUs out of nohz. However we'll rely on 4791 * jiffies deltas for updates happening while in nohz mode (idle ticks, idle 4792 * loop exit, nohz_idle_balance, nohz full exit...) 4793 * 4794 * This means we might still be one tick off for nohz periods. 4795 */ 4796 4797 static void cpu_load_update_nohz(struct rq *this_rq, 4798 unsigned long curr_jiffies, 4799 unsigned long load) 4800 { 4801 unsigned long pending_updates; 4802 4803 pending_updates = curr_jiffies - this_rq->last_load_update_tick; 4804 if (pending_updates) { 4805 this_rq->last_load_update_tick = curr_jiffies; 4806 /* 4807 * In the regular NOHZ case, we were idle, this means load 0. 4808 * In the NOHZ_FULL case, we were non-idle, we should consider 4809 * its weighted load. 4810 */ 4811 cpu_load_update(this_rq, load, pending_updates); 4812 } 4813 } 4814 4815 /* 4816 * Called from nohz_idle_balance() to update the load ratings before doing the 4817 * idle balance. 4818 */ 4819 static void cpu_load_update_idle(struct rq *this_rq) 4820 { 4821 /* 4822 * bail if there's load or we're actually up-to-date. 4823 */ 4824 if (weighted_cpuload(cpu_of(this_rq))) 4825 return; 4826 4827 cpu_load_update_nohz(this_rq, READ_ONCE(jiffies), 0); 4828 } 4829 4830 /* 4831 * Record CPU load on nohz entry so we know the tickless load to account 4832 * on nohz exit. cpu_load[0] happens then to be updated more frequently 4833 * than other cpu_load[idx] but it should be fine as cpu_load readers 4834 * shouldn't rely into synchronized cpu_load[*] updates. 4835 */ 4836 void cpu_load_update_nohz_start(void) 4837 { 4838 struct rq *this_rq = this_rq(); 4839 4840 /* 4841 * This is all lockless but should be fine. If weighted_cpuload changes 4842 * concurrently we'll exit nohz. And cpu_load write can race with 4843 * cpu_load_update_idle() but both updater would be writing the same. 4844 */ 4845 this_rq->cpu_load[0] = weighted_cpuload(cpu_of(this_rq)); 4846 } 4847 4848 /* 4849 * Account the tickless load in the end of a nohz frame. 4850 */ 4851 void cpu_load_update_nohz_stop(void) 4852 { 4853 unsigned long curr_jiffies = READ_ONCE(jiffies); 4854 struct rq *this_rq = this_rq(); 4855 unsigned long load; 4856 4857 if (curr_jiffies == this_rq->last_load_update_tick) 4858 return; 4859 4860 load = weighted_cpuload(cpu_of(this_rq)); 4861 raw_spin_lock(&this_rq->lock); 4862 update_rq_clock(this_rq); 4863 cpu_load_update_nohz(this_rq, curr_jiffies, load); 4864 raw_spin_unlock(&this_rq->lock); 4865 } 4866 #else /* !CONFIG_NO_HZ_COMMON */ 4867 static inline void cpu_load_update_nohz(struct rq *this_rq, 4868 unsigned long curr_jiffies, 4869 unsigned long load) { } 4870 #endif /* CONFIG_NO_HZ_COMMON */ 4871 4872 static void cpu_load_update_periodic(struct rq *this_rq, unsigned long load) 4873 { 4874 #ifdef CONFIG_NO_HZ_COMMON 4875 /* See the mess around cpu_load_update_nohz(). */ 4876 this_rq->last_load_update_tick = READ_ONCE(jiffies); 4877 #endif 4878 cpu_load_update(this_rq, load, 1); 4879 } 4880 4881 /* 4882 * Called from scheduler_tick() 4883 */ 4884 void cpu_load_update_active(struct rq *this_rq) 4885 { 4886 unsigned long load = weighted_cpuload(cpu_of(this_rq)); 4887 4888 if (tick_nohz_tick_stopped()) 4889 cpu_load_update_nohz(this_rq, READ_ONCE(jiffies), load); 4890 else 4891 cpu_load_update_periodic(this_rq, load); 4892 } 4893 4894 /* 4895 * Return a low guess at the load of a migration-source cpu weighted 4896 * according to the scheduling class and "nice" value. 4897 * 4898 * We want to under-estimate the load of migration sources, to 4899 * balance conservatively. 4900 */ 4901 static unsigned long source_load(int cpu, int type) 4902 { 4903 struct rq *rq = cpu_rq(cpu); 4904 unsigned long total = weighted_cpuload(cpu); 4905 4906 if (type == 0 || !sched_feat(LB_BIAS)) 4907 return total; 4908 4909 return min(rq->cpu_load[type-1], total); 4910 } 4911 4912 /* 4913 * Return a high guess at the load of a migration-target cpu weighted 4914 * according to the scheduling class and "nice" value. 4915 */ 4916 static unsigned long target_load(int cpu, int type) 4917 { 4918 struct rq *rq = cpu_rq(cpu); 4919 unsigned long total = weighted_cpuload(cpu); 4920 4921 if (type == 0 || !sched_feat(LB_BIAS)) 4922 return total; 4923 4924 return max(rq->cpu_load[type-1], total); 4925 } 4926 4927 static unsigned long capacity_of(int cpu) 4928 { 4929 return cpu_rq(cpu)->cpu_capacity; 4930 } 4931 4932 static unsigned long capacity_orig_of(int cpu) 4933 { 4934 return cpu_rq(cpu)->cpu_capacity_orig; 4935 } 4936 4937 static unsigned long cpu_avg_load_per_task(int cpu) 4938 { 4939 struct rq *rq = cpu_rq(cpu); 4940 unsigned long nr_running = READ_ONCE(rq->cfs.h_nr_running); 4941 unsigned long load_avg = weighted_cpuload(cpu); 4942 4943 if (nr_running) 4944 return load_avg / nr_running; 4945 4946 return 0; 4947 } 4948 4949 #ifdef CONFIG_FAIR_GROUP_SCHED 4950 /* 4951 * effective_load() calculates the load change as seen from the root_task_group 4952 * 4953 * Adding load to a group doesn't make a group heavier, but can cause movement 4954 * of group shares between cpus. Assuming the shares were perfectly aligned one 4955 * can calculate the shift in shares. 4956 * 4957 * Calculate the effective load difference if @wl is added (subtracted) to @tg 4958 * on this @cpu and results in a total addition (subtraction) of @wg to the 4959 * total group weight. 4960 * 4961 * Given a runqueue weight distribution (rw_i) we can compute a shares 4962 * distribution (s_i) using: 4963 * 4964 * s_i = rw_i / \Sum rw_j (1) 4965 * 4966 * Suppose we have 4 CPUs and our @tg is a direct child of the root group and 4967 * has 7 equal weight tasks, distributed as below (rw_i), with the resulting 4968 * shares distribution (s_i): 4969 * 4970 * rw_i = { 2, 4, 1, 0 } 4971 * s_i = { 2/7, 4/7, 1/7, 0 } 4972 * 4973 * As per wake_affine() we're interested in the load of two CPUs (the CPU the 4974 * task used to run on and the CPU the waker is running on), we need to 4975 * compute the effect of waking a task on either CPU and, in case of a sync 4976 * wakeup, compute the effect of the current task going to sleep. 4977 * 4978 * So for a change of @wl to the local @cpu with an overall group weight change 4979 * of @wl we can compute the new shares distribution (s'_i) using: 4980 * 4981 * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2) 4982 * 4983 * Suppose we're interested in CPUs 0 and 1, and want to compute the load 4984 * differences in waking a task to CPU 0. The additional task changes the 4985 * weight and shares distributions like: 4986 * 4987 * rw'_i = { 3, 4, 1, 0 } 4988 * s'_i = { 3/8, 4/8, 1/8, 0 } 4989 * 4990 * We can then compute the difference in effective weight by using: 4991 * 4992 * dw_i = S * (s'_i - s_i) (3) 4993 * 4994 * Where 'S' is the group weight as seen by its parent. 4995 * 4996 * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7) 4997 * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 - 4998 * 4/7) times the weight of the group. 4999 */ 5000 static long effective_load(struct task_group *tg, int cpu, long wl, long wg) 5001 { 5002 struct sched_entity *se = tg->se[cpu]; 5003 5004 if (!tg->parent) /* the trivial, non-cgroup case */ 5005 return wl; 5006 5007 for_each_sched_entity(se) { 5008 struct cfs_rq *cfs_rq = se->my_q; 5009 long W, w = cfs_rq_load_avg(cfs_rq); 5010 5011 tg = cfs_rq->tg; 5012 5013 /* 5014 * W = @wg + \Sum rw_j 5015 */ 5016 W = wg + atomic_long_read(&tg->load_avg); 5017 5018 /* Ensure \Sum rw_j >= rw_i */ 5019 W -= cfs_rq->tg_load_avg_contrib; 5020 W += w; 5021 5022 /* 5023 * w = rw_i + @wl 5024 */ 5025 w += wl; 5026 5027 /* 5028 * wl = S * s'_i; see (2) 5029 */ 5030 if (W > 0 && w < W) 5031 wl = (w * (long)tg->shares) / W; 5032 else 5033 wl = tg->shares; 5034 5035 /* 5036 * Per the above, wl is the new se->load.weight value; since 5037 * those are clipped to [MIN_SHARES, ...) do so now. See 5038 * calc_cfs_shares(). 5039 */ 5040 if (wl < MIN_SHARES) 5041 wl = MIN_SHARES; 5042 5043 /* 5044 * wl = dw_i = S * (s'_i - s_i); see (3) 5045 */ 5046 wl -= se->avg.load_avg; 5047 5048 /* 5049 * Recursively apply this logic to all parent groups to compute 5050 * the final effective load change on the root group. Since 5051 * only the @tg group gets extra weight, all parent groups can 5052 * only redistribute existing shares. @wl is the shift in shares 5053 * resulting from this level per the above. 5054 */ 5055 wg = 0; 5056 } 5057 5058 return wl; 5059 } 5060 #else 5061 5062 static long effective_load(struct task_group *tg, int cpu, long wl, long wg) 5063 { 5064 return wl; 5065 } 5066 5067 #endif 5068 5069 static void record_wakee(struct task_struct *p) 5070 { 5071 /* 5072 * Only decay a single time; tasks that have less then 1 wakeup per 5073 * jiffy will not have built up many flips. 5074 */ 5075 if (time_after(jiffies, current->wakee_flip_decay_ts + HZ)) { 5076 current->wakee_flips >>= 1; 5077 current->wakee_flip_decay_ts = jiffies; 5078 } 5079 5080 if (current->last_wakee != p) { 5081 current->last_wakee = p; 5082 current->wakee_flips++; 5083 } 5084 } 5085 5086 /* 5087 * Detect M:N waker/wakee relationships via a switching-frequency heuristic. 5088 * 5089 * A waker of many should wake a different task than the one last awakened 5090 * at a frequency roughly N times higher than one of its wakees. 5091 * 5092 * In order to determine whether we should let the load spread vs consolidating 5093 * to shared cache, we look for a minimum 'flip' frequency of llc_size in one 5094 * partner, and a factor of lls_size higher frequency in the other. 5095 * 5096 * With both conditions met, we can be relatively sure that the relationship is 5097 * non-monogamous, with partner count exceeding socket size. 5098 * 5099 * Waker/wakee being client/server, worker/dispatcher, interrupt source or 5100 * whatever is irrelevant, spread criteria is apparent partner count exceeds 5101 * socket size. 5102 */ 5103 static int wake_wide(struct task_struct *p) 5104 { 5105 unsigned int master = current->wakee_flips; 5106 unsigned int slave = p->wakee_flips; 5107 int factor = this_cpu_read(sd_llc_size); 5108 5109 if (master < slave) 5110 swap(master, slave); 5111 if (slave < factor || master < slave * factor) 5112 return 0; 5113 return 1; 5114 } 5115 5116 static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync) 5117 { 5118 s64 this_load, load; 5119 s64 this_eff_load, prev_eff_load; 5120 int idx, this_cpu, prev_cpu; 5121 struct task_group *tg; 5122 unsigned long weight; 5123 int balanced; 5124 5125 idx = sd->wake_idx; 5126 this_cpu = smp_processor_id(); 5127 prev_cpu = task_cpu(p); 5128 load = source_load(prev_cpu, idx); 5129 this_load = target_load(this_cpu, idx); 5130 5131 /* 5132 * If sync wakeup then subtract the (maximum possible) 5133 * effect of the currently running task from the load 5134 * of the current CPU: 5135 */ 5136 if (sync) { 5137 tg = task_group(current); 5138 weight = current->se.avg.load_avg; 5139 5140 this_load += effective_load(tg, this_cpu, -weight, -weight); 5141 load += effective_load(tg, prev_cpu, 0, -weight); 5142 } 5143 5144 tg = task_group(p); 5145 weight = p->se.avg.load_avg; 5146 5147 /* 5148 * In low-load situations, where prev_cpu is idle and this_cpu is idle 5149 * due to the sync cause above having dropped this_load to 0, we'll 5150 * always have an imbalance, but there's really nothing you can do 5151 * about that, so that's good too. 5152 * 5153 * Otherwise check if either cpus are near enough in load to allow this 5154 * task to be woken on this_cpu. 5155 */ 5156 this_eff_load = 100; 5157 this_eff_load *= capacity_of(prev_cpu); 5158 5159 prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2; 5160 prev_eff_load *= capacity_of(this_cpu); 5161 5162 if (this_load > 0) { 5163 this_eff_load *= this_load + 5164 effective_load(tg, this_cpu, weight, weight); 5165 5166 prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight); 5167 } 5168 5169 balanced = this_eff_load <= prev_eff_load; 5170 5171 schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts); 5172 5173 if (!balanced) 5174 return 0; 5175 5176 schedstat_inc(sd, ttwu_move_affine); 5177 schedstat_inc(p, se.statistics.nr_wakeups_affine); 5178 5179 return 1; 5180 } 5181 5182 /* 5183 * find_idlest_group finds and returns the least busy CPU group within the 5184 * domain. 5185 */ 5186 static struct sched_group * 5187 find_idlest_group(struct sched_domain *sd, struct task_struct *p, 5188 int this_cpu, int sd_flag) 5189 { 5190 struct sched_group *idlest = NULL, *group = sd->groups; 5191 unsigned long min_load = ULONG_MAX, this_load = 0; 5192 int load_idx = sd->forkexec_idx; 5193 int imbalance = 100 + (sd->imbalance_pct-100)/2; 5194 5195 if (sd_flag & SD_BALANCE_WAKE) 5196 load_idx = sd->wake_idx; 5197 5198 do { 5199 unsigned long load, avg_load; 5200 int local_group; 5201 int i; 5202 5203 /* Skip over this group if it has no CPUs allowed */ 5204 if (!cpumask_intersects(sched_group_cpus(group), 5205 tsk_cpus_allowed(p))) 5206 continue; 5207 5208 local_group = cpumask_test_cpu(this_cpu, 5209 sched_group_cpus(group)); 5210 5211 /* Tally up the load of all CPUs in the group */ 5212 avg_load = 0; 5213 5214 for_each_cpu(i, sched_group_cpus(group)) { 5215 /* Bias balancing toward cpus of our domain */ 5216 if (local_group) 5217 load = source_load(i, load_idx); 5218 else 5219 load = target_load(i, load_idx); 5220 5221 avg_load += load; 5222 } 5223 5224 /* Adjust by relative CPU capacity of the group */ 5225 avg_load = (avg_load * SCHED_CAPACITY_SCALE) / group->sgc->capacity; 5226 5227 if (local_group) { 5228 this_load = avg_load; 5229 } else if (avg_load < min_load) { 5230 min_load = avg_load; 5231 idlest = group; 5232 } 5233 } while (group = group->next, group != sd->groups); 5234 5235 if (!idlest || 100*this_load < imbalance*min_load) 5236 return NULL; 5237 return idlest; 5238 } 5239 5240 /* 5241 * find_idlest_cpu - find the idlest cpu among the cpus in group. 5242 */ 5243 static int 5244 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu) 5245 { 5246 unsigned long load, min_load = ULONG_MAX; 5247 unsigned int min_exit_latency = UINT_MAX; 5248 u64 latest_idle_timestamp = 0; 5249 int least_loaded_cpu = this_cpu; 5250 int shallowest_idle_cpu = -1; 5251 int i; 5252 5253 /* Traverse only the allowed CPUs */ 5254 for_each_cpu_and(i, sched_group_cpus(group), tsk_cpus_allowed(p)) { 5255 if (idle_cpu(i)) { 5256 struct rq *rq = cpu_rq(i); 5257 struct cpuidle_state *idle = idle_get_state(rq); 5258 if (idle && idle->exit_latency < min_exit_latency) { 5259 /* 5260 * We give priority to a CPU whose idle state 5261 * has the smallest exit latency irrespective 5262 * of any idle timestamp. 5263 */ 5264 min_exit_latency = idle->exit_latency; 5265 latest_idle_timestamp = rq->idle_stamp; 5266 shallowest_idle_cpu = i; 5267 } else if ((!idle || idle->exit_latency == min_exit_latency) && 5268 rq->idle_stamp > latest_idle_timestamp) { 5269 /* 5270 * If equal or no active idle state, then 5271 * the most recently idled CPU might have 5272 * a warmer cache. 5273 */ 5274 latest_idle_timestamp = rq->idle_stamp; 5275 shallowest_idle_cpu = i; 5276 } 5277 } else if (shallowest_idle_cpu == -1) { 5278 load = weighted_cpuload(i); 5279 if (load < min_load || (load == min_load && i == this_cpu)) { 5280 min_load = load; 5281 least_loaded_cpu = i; 5282 } 5283 } 5284 } 5285 5286 return shallowest_idle_cpu != -1 ? shallowest_idle_cpu : least_loaded_cpu; 5287 } 5288 5289 /* 5290 * Try and locate an idle CPU in the sched_domain. 5291 */ 5292 static int select_idle_sibling(struct task_struct *p, int target) 5293 { 5294 struct sched_domain *sd; 5295 struct sched_group *sg; 5296 int i = task_cpu(p); 5297 5298 if (idle_cpu(target)) 5299 return target; 5300 5301 /* 5302 * If the prevous cpu is cache affine and idle, don't be stupid. 5303 */ 5304 if (i != target && cpus_share_cache(i, target) && idle_cpu(i)) 5305 return i; 5306 5307 /* 5308 * Otherwise, iterate the domains and find an eligible idle cpu. 5309 * 5310 * A completely idle sched group at higher domains is more 5311 * desirable than an idle group at a lower level, because lower 5312 * domains have smaller groups and usually share hardware 5313 * resources which causes tasks to contend on them, e.g. x86 5314 * hyperthread siblings in the lowest domain (SMT) can contend 5315 * on the shared cpu pipeline. 5316 * 5317 * However, while we prefer idle groups at higher domains 5318 * finding an idle cpu at the lowest domain is still better than 5319 * returning 'target', which we've already established, isn't 5320 * idle. 5321 */ 5322 sd = rcu_dereference(per_cpu(sd_llc, target)); 5323 for_each_lower_domain(sd) { 5324 sg = sd->groups; 5325 do { 5326 if (!cpumask_intersects(sched_group_cpus(sg), 5327 tsk_cpus_allowed(p))) 5328 goto next; 5329 5330 /* Ensure the entire group is idle */ 5331 for_each_cpu(i, sched_group_cpus(sg)) { 5332 if (i == target || !idle_cpu(i)) 5333 goto next; 5334 } 5335 5336 /* 5337 * It doesn't matter which cpu we pick, the 5338 * whole group is idle. 5339 */ 5340 target = cpumask_first_and(sched_group_cpus(sg), 5341 tsk_cpus_allowed(p)); 5342 goto done; 5343 next: 5344 sg = sg->next; 5345 } while (sg != sd->groups); 5346 } 5347 done: 5348 return target; 5349 } 5350 5351 /* 5352 * cpu_util returns the amount of capacity of a CPU that is used by CFS 5353 * tasks. The unit of the return value must be the one of capacity so we can 5354 * compare the utilization with the capacity of the CPU that is available for 5355 * CFS task (ie cpu_capacity). 5356 * 5357 * cfs_rq.avg.util_avg is the sum of running time of runnable tasks plus the 5358 * recent utilization of currently non-runnable tasks on a CPU. It represents 5359 * the amount of utilization of a CPU in the range [0..capacity_orig] where 5360 * capacity_orig is the cpu_capacity available at the highest frequency 5361 * (arch_scale_freq_capacity()). 5362 * The utilization of a CPU converges towards a sum equal to or less than the 5363 * current capacity (capacity_curr <= capacity_orig) of the CPU because it is 5364 * the running time on this CPU scaled by capacity_curr. 5365 * 5366 * Nevertheless, cfs_rq.avg.util_avg can be higher than capacity_curr or even 5367 * higher than capacity_orig because of unfortunate rounding in 5368 * cfs.avg.util_avg or just after migrating tasks and new task wakeups until 5369 * the average stabilizes with the new running time. We need to check that the 5370 * utilization stays within the range of [0..capacity_orig] and cap it if 5371 * necessary. Without utilization capping, a group could be seen as overloaded 5372 * (CPU0 utilization at 121% + CPU1 utilization at 80%) whereas CPU1 has 20% of 5373 * available capacity. We allow utilization to overshoot capacity_curr (but not 5374 * capacity_orig) as it useful for predicting the capacity required after task 5375 * migrations (scheduler-driven DVFS). 5376 */ 5377 static int cpu_util(int cpu) 5378 { 5379 unsigned long util = cpu_rq(cpu)->cfs.avg.util_avg; 5380 unsigned long capacity = capacity_orig_of(cpu); 5381 5382 return (util >= capacity) ? capacity : util; 5383 } 5384 5385 /* 5386 * select_task_rq_fair: Select target runqueue for the waking task in domains 5387 * that have the 'sd_flag' flag set. In practice, this is SD_BALANCE_WAKE, 5388 * SD_BALANCE_FORK, or SD_BALANCE_EXEC. 5389 * 5390 * Balances load by selecting the idlest cpu in the idlest group, or under 5391 * certain conditions an idle sibling cpu if the domain has SD_WAKE_AFFINE set. 5392 * 5393 * Returns the target cpu number. 5394 * 5395 * preempt must be disabled. 5396 */ 5397 static int 5398 select_task_rq_fair(struct task_struct *p, int prev_cpu, int sd_flag, int wake_flags) 5399 { 5400 struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL; 5401 int cpu = smp_processor_id(); 5402 int new_cpu = prev_cpu; 5403 int want_affine = 0; 5404 int sync = wake_flags & WF_SYNC; 5405 5406 if (sd_flag & SD_BALANCE_WAKE) { 5407 record_wakee(p); 5408 want_affine = !wake_wide(p) && cpumask_test_cpu(cpu, tsk_cpus_allowed(p)); 5409 } 5410 5411 rcu_read_lock(); 5412 for_each_domain(cpu, tmp) { 5413 if (!(tmp->flags & SD_LOAD_BALANCE)) 5414 break; 5415 5416 /* 5417 * If both cpu and prev_cpu are part of this domain, 5418 * cpu is a valid SD_WAKE_AFFINE target. 5419 */ 5420 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) && 5421 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) { 5422 affine_sd = tmp; 5423 break; 5424 } 5425 5426 if (tmp->flags & sd_flag) 5427 sd = tmp; 5428 else if (!want_affine) 5429 break; 5430 } 5431 5432 if (affine_sd) { 5433 sd = NULL; /* Prefer wake_affine over balance flags */ 5434 if (cpu != prev_cpu && wake_affine(affine_sd, p, sync)) 5435 new_cpu = cpu; 5436 } 5437 5438 if (!sd) { 5439 if (sd_flag & SD_BALANCE_WAKE) /* XXX always ? */ 5440 new_cpu = select_idle_sibling(p, new_cpu); 5441 5442 } else while (sd) { 5443 struct sched_group *group; 5444 int weight; 5445 5446 if (!(sd->flags & sd_flag)) { 5447 sd = sd->child; 5448 continue; 5449 } 5450 5451 group = find_idlest_group(sd, p, cpu, sd_flag); 5452 if (!group) { 5453 sd = sd->child; 5454 continue; 5455 } 5456 5457 new_cpu = find_idlest_cpu(group, p, cpu); 5458 if (new_cpu == -1 || new_cpu == cpu) { 5459 /* Now try balancing at a lower domain level of cpu */ 5460 sd = sd->child; 5461 continue; 5462 } 5463 5464 /* Now try balancing at a lower domain level of new_cpu */ 5465 cpu = new_cpu; 5466 weight = sd->span_weight; 5467 sd = NULL; 5468 for_each_domain(cpu, tmp) { 5469 if (weight <= tmp->span_weight) 5470 break; 5471 if (tmp->flags & sd_flag) 5472 sd = tmp; 5473 } 5474 /* while loop will break here if sd == NULL */ 5475 } 5476 rcu_read_unlock(); 5477 5478 return new_cpu; 5479 } 5480 5481 /* 5482 * Called immediately before a task is migrated to a new cpu; task_cpu(p) and 5483 * cfs_rq_of(p) references at time of call are still valid and identify the 5484 * previous cpu. The caller guarantees p->pi_lock or task_rq(p)->lock is held. 5485 */ 5486 static void migrate_task_rq_fair(struct task_struct *p) 5487 { 5488 /* 5489 * As blocked tasks retain absolute vruntime the migration needs to 5490 * deal with this by subtracting the old and adding the new 5491 * min_vruntime -- the latter is done by enqueue_entity() when placing 5492 * the task on the new runqueue. 5493 */ 5494 if (p->state == TASK_WAKING) { 5495 struct sched_entity *se = &p->se; 5496 struct cfs_rq *cfs_rq = cfs_rq_of(se); 5497 u64 min_vruntime; 5498 5499 #ifndef CONFIG_64BIT 5500 u64 min_vruntime_copy; 5501 5502 do { 5503 min_vruntime_copy = cfs_rq->min_vruntime_copy; 5504 smp_rmb(); 5505 min_vruntime = cfs_rq->min_vruntime; 5506 } while (min_vruntime != min_vruntime_copy); 5507 #else 5508 min_vruntime = cfs_rq->min_vruntime; 5509 #endif 5510 5511 se->vruntime -= min_vruntime; 5512 } 5513 5514 /* 5515 * We are supposed to update the task to "current" time, then its up to date 5516 * and ready to go to new CPU/cfs_rq. But we have difficulty in getting 5517 * what current time is, so simply throw away the out-of-date time. This 5518 * will result in the wakee task is less decayed, but giving the wakee more 5519 * load sounds not bad. 5520 */ 5521 remove_entity_load_avg(&p->se); 5522 5523 /* Tell new CPU we are migrated */ 5524 p->se.avg.last_update_time = 0; 5525 5526 /* We have migrated, no longer consider this task hot */ 5527 p->se.exec_start = 0; 5528 } 5529 5530 static void task_dead_fair(struct task_struct *p) 5531 { 5532 remove_entity_load_avg(&p->se); 5533 } 5534 #endif /* CONFIG_SMP */ 5535 5536 static unsigned long 5537 wakeup_gran(struct sched_entity *curr, struct sched_entity *se) 5538 { 5539 unsigned long gran = sysctl_sched_wakeup_granularity; 5540 5541 /* 5542 * Since its curr running now, convert the gran from real-time 5543 * to virtual-time in his units. 5544 * 5545 * By using 'se' instead of 'curr' we penalize light tasks, so 5546 * they get preempted easier. That is, if 'se' < 'curr' then 5547 * the resulting gran will be larger, therefore penalizing the 5548 * lighter, if otoh 'se' > 'curr' then the resulting gran will 5549 * be smaller, again penalizing the lighter task. 5550 * 5551 * This is especially important for buddies when the leftmost 5552 * task is higher priority than the buddy. 5553 */ 5554 return calc_delta_fair(gran, se); 5555 } 5556 5557 /* 5558 * Should 'se' preempt 'curr'. 5559 * 5560 * |s1 5561 * |s2 5562 * |s3 5563 * g 5564 * |<--->|c 5565 * 5566 * w(c, s1) = -1 5567 * w(c, s2) = 0 5568 * w(c, s3) = 1 5569 * 5570 */ 5571 static int 5572 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se) 5573 { 5574 s64 gran, vdiff = curr->vruntime - se->vruntime; 5575 5576 if (vdiff <= 0) 5577 return -1; 5578 5579 gran = wakeup_gran(curr, se); 5580 if (vdiff > gran) 5581 return 1; 5582 5583 return 0; 5584 } 5585 5586 static void set_last_buddy(struct sched_entity *se) 5587 { 5588 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE)) 5589 return; 5590 5591 for_each_sched_entity(se) 5592 cfs_rq_of(se)->last = se; 5593 } 5594 5595 static void set_next_buddy(struct sched_entity *se) 5596 { 5597 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE)) 5598 return; 5599 5600 for_each_sched_entity(se) 5601 cfs_rq_of(se)->next = se; 5602 } 5603 5604 static void set_skip_buddy(struct sched_entity *se) 5605 { 5606 for_each_sched_entity(se) 5607 cfs_rq_of(se)->skip = se; 5608 } 5609 5610 /* 5611 * Preempt the current task with a newly woken task if needed: 5612 */ 5613 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags) 5614 { 5615 struct task_struct *curr = rq->curr; 5616 struct sched_entity *se = &curr->se, *pse = &p->se; 5617 struct cfs_rq *cfs_rq = task_cfs_rq(curr); 5618 int scale = cfs_rq->nr_running >= sched_nr_latency; 5619 int next_buddy_marked = 0; 5620 5621 if (unlikely(se == pse)) 5622 return; 5623 5624 /* 5625 * This is possible from callers such as attach_tasks(), in which we 5626 * unconditionally check_prempt_curr() after an enqueue (which may have 5627 * lead to a throttle). This both saves work and prevents false 5628 * next-buddy nomination below. 5629 */ 5630 if (unlikely(throttled_hierarchy(cfs_rq_of(pse)))) 5631 return; 5632 5633 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) { 5634 set_next_buddy(pse); 5635 next_buddy_marked = 1; 5636 } 5637 5638 /* 5639 * We can come here with TIF_NEED_RESCHED already set from new task 5640 * wake up path. 5641 * 5642 * Note: this also catches the edge-case of curr being in a throttled 5643 * group (e.g. via set_curr_task), since update_curr() (in the 5644 * enqueue of curr) will have resulted in resched being set. This 5645 * prevents us from potentially nominating it as a false LAST_BUDDY 5646 * below. 5647 */ 5648 if (test_tsk_need_resched(curr)) 5649 return; 5650 5651 /* Idle tasks are by definition preempted by non-idle tasks. */ 5652 if (unlikely(curr->policy == SCHED_IDLE) && 5653 likely(p->policy != SCHED_IDLE)) 5654 goto preempt; 5655 5656 /* 5657 * Batch and idle tasks do not preempt non-idle tasks (their preemption 5658 * is driven by the tick): 5659 */ 5660 if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION)) 5661 return; 5662 5663 find_matching_se(&se, &pse); 5664 update_curr(cfs_rq_of(se)); 5665 BUG_ON(!pse); 5666 if (wakeup_preempt_entity(se, pse) == 1) { 5667 /* 5668 * Bias pick_next to pick the sched entity that is 5669 * triggering this preemption. 5670 */ 5671 if (!next_buddy_marked) 5672 set_next_buddy(pse); 5673 goto preempt; 5674 } 5675 5676 return; 5677 5678 preempt: 5679 resched_curr(rq); 5680 /* 5681 * Only set the backward buddy when the current task is still 5682 * on the rq. This can happen when a wakeup gets interleaved 5683 * with schedule on the ->pre_schedule() or idle_balance() 5684 * point, either of which can * drop the rq lock. 5685 * 5686 * Also, during early boot the idle thread is in the fair class, 5687 * for obvious reasons its a bad idea to schedule back to it. 5688 */ 5689 if (unlikely(!se->on_rq || curr == rq->idle)) 5690 return; 5691 5692 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se)) 5693 set_last_buddy(se); 5694 } 5695 5696 static struct task_struct * 5697 pick_next_task_fair(struct rq *rq, struct task_struct *prev, struct pin_cookie cookie) 5698 { 5699 struct cfs_rq *cfs_rq = &rq->cfs; 5700 struct sched_entity *se; 5701 struct task_struct *p; 5702 int new_tasks; 5703 5704 again: 5705 #ifdef CONFIG_FAIR_GROUP_SCHED 5706 if (!cfs_rq->nr_running) 5707 goto idle; 5708 5709 if (prev->sched_class != &fair_sched_class) 5710 goto simple; 5711 5712 /* 5713 * Because of the set_next_buddy() in dequeue_task_fair() it is rather 5714 * likely that a next task is from the same cgroup as the current. 5715 * 5716 * Therefore attempt to avoid putting and setting the entire cgroup 5717 * hierarchy, only change the part that actually changes. 5718 */ 5719 5720 do { 5721 struct sched_entity *curr = cfs_rq->curr; 5722 5723 /* 5724 * Since we got here without doing put_prev_entity() we also 5725 * have to consider cfs_rq->curr. If it is still a runnable 5726 * entity, update_curr() will update its vruntime, otherwise 5727 * forget we've ever seen it. 5728 */ 5729 if (curr) { 5730 if (curr->on_rq) 5731 update_curr(cfs_rq); 5732 else 5733 curr = NULL; 5734 5735 /* 5736 * This call to check_cfs_rq_runtime() will do the 5737 * throttle and dequeue its entity in the parent(s). 5738 * Therefore the 'simple' nr_running test will indeed 5739 * be correct. 5740 */ 5741 if (unlikely(check_cfs_rq_runtime(cfs_rq))) 5742 goto simple; 5743 } 5744 5745 se = pick_next_entity(cfs_rq, curr); 5746 cfs_rq = group_cfs_rq(se); 5747 } while (cfs_rq); 5748 5749 p = task_of(se); 5750 5751 /* 5752 * Since we haven't yet done put_prev_entity and if the selected task 5753 * is a different task than we started out with, try and touch the 5754 * least amount of cfs_rqs. 5755 */ 5756 if (prev != p) { 5757 struct sched_entity *pse = &prev->se; 5758 5759 while (!(cfs_rq = is_same_group(se, pse))) { 5760 int se_depth = se->depth; 5761 int pse_depth = pse->depth; 5762 5763 if (se_depth <= pse_depth) { 5764 put_prev_entity(cfs_rq_of(pse), pse); 5765 pse = parent_entity(pse); 5766 } 5767 if (se_depth >= pse_depth) { 5768 set_next_entity(cfs_rq_of(se), se); 5769 se = parent_entity(se); 5770 } 5771 } 5772 5773 put_prev_entity(cfs_rq, pse); 5774 set_next_entity(cfs_rq, se); 5775 } 5776 5777 if (hrtick_enabled(rq)) 5778 hrtick_start_fair(rq, p); 5779 5780 return p; 5781 simple: 5782 cfs_rq = &rq->cfs; 5783 #endif 5784 5785 if (!cfs_rq->nr_running) 5786 goto idle; 5787 5788 put_prev_task(rq, prev); 5789 5790 do { 5791 se = pick_next_entity(cfs_rq, NULL); 5792 set_next_entity(cfs_rq, se); 5793 cfs_rq = group_cfs_rq(se); 5794 } while (cfs_rq); 5795 5796 p = task_of(se); 5797 5798 if (hrtick_enabled(rq)) 5799 hrtick_start_fair(rq, p); 5800 5801 return p; 5802 5803 idle: 5804 /* 5805 * This is OK, because current is on_cpu, which avoids it being picked 5806 * for load-balance and preemption/IRQs are still disabled avoiding 5807 * further scheduler activity on it and we're being very careful to 5808 * re-start the picking loop. 5809 */ 5810 lockdep_unpin_lock(&rq->lock, cookie); 5811 new_tasks = idle_balance(rq); 5812 lockdep_repin_lock(&rq->lock, cookie); 5813 /* 5814 * Because idle_balance() releases (and re-acquires) rq->lock, it is 5815 * possible for any higher priority task to appear. In that case we 5816 * must re-start the pick_next_entity() loop. 5817 */ 5818 if (new_tasks < 0) 5819 return RETRY_TASK; 5820 5821 if (new_tasks > 0) 5822 goto again; 5823 5824 return NULL; 5825 } 5826 5827 /* 5828 * Account for a descheduled task: 5829 */ 5830 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev) 5831 { 5832 struct sched_entity *se = &prev->se; 5833 struct cfs_rq *cfs_rq; 5834 5835 for_each_sched_entity(se) { 5836 cfs_rq = cfs_rq_of(se); 5837 put_prev_entity(cfs_rq, se); 5838 } 5839 } 5840 5841 /* 5842 * sched_yield() is very simple 5843 * 5844 * The magic of dealing with the ->skip buddy is in pick_next_entity. 5845 */ 5846 static void yield_task_fair(struct rq *rq) 5847 { 5848 struct task_struct *curr = rq->curr; 5849 struct cfs_rq *cfs_rq = task_cfs_rq(curr); 5850 struct sched_entity *se = &curr->se; 5851 5852 /* 5853 * Are we the only task in the tree? 5854 */ 5855 if (unlikely(rq->nr_running == 1)) 5856 return; 5857 5858 clear_buddies(cfs_rq, se); 5859 5860 if (curr->policy != SCHED_BATCH) { 5861 update_rq_clock(rq); 5862 /* 5863 * Update run-time statistics of the 'current'. 5864 */ 5865 update_curr(cfs_rq); 5866 /* 5867 * Tell update_rq_clock() that we've just updated, 5868 * so we don't do microscopic update in schedule() 5869 * and double the fastpath cost. 5870 */ 5871 rq_clock_skip_update(rq, true); 5872 } 5873 5874 set_skip_buddy(se); 5875 } 5876 5877 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt) 5878 { 5879 struct sched_entity *se = &p->se; 5880 5881 /* throttled hierarchies are not runnable */ 5882 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se))) 5883 return false; 5884 5885 /* Tell the scheduler that we'd really like pse to run next. */ 5886 set_next_buddy(se); 5887 5888 yield_task_fair(rq); 5889 5890 return true; 5891 } 5892 5893 #ifdef CONFIG_SMP 5894 /************************************************** 5895 * Fair scheduling class load-balancing methods. 5896 * 5897 * BASICS 5898 * 5899 * The purpose of load-balancing is to achieve the same basic fairness the 5900 * per-cpu scheduler provides, namely provide a proportional amount of compute 5901 * time to each task. This is expressed in the following equation: 5902 * 5903 * W_i,n/P_i == W_j,n/P_j for all i,j (1) 5904 * 5905 * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight 5906 * W_i,0 is defined as: 5907 * 5908 * W_i,0 = \Sum_j w_i,j (2) 5909 * 5910 * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight 5911 * is derived from the nice value as per sched_prio_to_weight[]. 5912 * 5913 * The weight average is an exponential decay average of the instantaneous 5914 * weight: 5915 * 5916 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3) 5917 * 5918 * C_i is the compute capacity of cpu i, typically it is the 5919 * fraction of 'recent' time available for SCHED_OTHER task execution. But it 5920 * can also include other factors [XXX]. 5921 * 5922 * To achieve this balance we define a measure of imbalance which follows 5923 * directly from (1): 5924 * 5925 * imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j } (4) 5926 * 5927 * We them move tasks around to minimize the imbalance. In the continuous 5928 * function space it is obvious this converges, in the discrete case we get 5929 * a few fun cases generally called infeasible weight scenarios. 5930 * 5931 * [XXX expand on: 5932 * - infeasible weights; 5933 * - local vs global optima in the discrete case. ] 5934 * 5935 * 5936 * SCHED DOMAINS 5937 * 5938 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2) 5939 * for all i,j solution, we create a tree of cpus that follows the hardware 5940 * topology where each level pairs two lower groups (or better). This results 5941 * in O(log n) layers. Furthermore we reduce the number of cpus going up the 5942 * tree to only the first of the previous level and we decrease the frequency 5943 * of load-balance at each level inv. proportional to the number of cpus in 5944 * the groups. 5945 * 5946 * This yields: 5947 * 5948 * log_2 n 1 n 5949 * \Sum { --- * --- * 2^i } = O(n) (5) 5950 * i = 0 2^i 2^i 5951 * `- size of each group 5952 * | | `- number of cpus doing load-balance 5953 * | `- freq 5954 * `- sum over all levels 5955 * 5956 * Coupled with a limit on how many tasks we can migrate every balance pass, 5957 * this makes (5) the runtime complexity of the balancer. 5958 * 5959 * An important property here is that each CPU is still (indirectly) connected 5960 * to every other cpu in at most O(log n) steps: 5961 * 5962 * The adjacency matrix of the resulting graph is given by: 5963 * 5964 * log_2 n 5965 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6) 5966 * k = 0 5967 * 5968 * And you'll find that: 5969 * 5970 * A^(log_2 n)_i,j != 0 for all i,j (7) 5971 * 5972 * Showing there's indeed a path between every cpu in at most O(log n) steps. 5973 * The task movement gives a factor of O(m), giving a convergence complexity 5974 * of: 5975 * 5976 * O(nm log n), n := nr_cpus, m := nr_tasks (8) 5977 * 5978 * 5979 * WORK CONSERVING 5980 * 5981 * In order to avoid CPUs going idle while there's still work to do, new idle 5982 * balancing is more aggressive and has the newly idle cpu iterate up the domain 5983 * tree itself instead of relying on other CPUs to bring it work. 5984 * 5985 * This adds some complexity to both (5) and (8) but it reduces the total idle 5986 * time. 5987 * 5988 * [XXX more?] 5989 * 5990 * 5991 * CGROUPS 5992 * 5993 * Cgroups make a horror show out of (2), instead of a simple sum we get: 5994 * 5995 * s_k,i 5996 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9) 5997 * S_k 5998 * 5999 * Where 6000 * 6001 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10) 6002 * 6003 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i. 6004 * 6005 * The big problem is S_k, its a global sum needed to compute a local (W_i) 6006 * property. 6007 * 6008 * [XXX write more on how we solve this.. _after_ merging pjt's patches that 6009 * rewrite all of this once again.] 6010 */ 6011 6012 static unsigned long __read_mostly max_load_balance_interval = HZ/10; 6013 6014 enum fbq_type { regular, remote, all }; 6015 6016 #define LBF_ALL_PINNED 0x01 6017 #define LBF_NEED_BREAK 0x02 6018 #define LBF_DST_PINNED 0x04 6019 #define LBF_SOME_PINNED 0x08 6020 6021 struct lb_env { 6022 struct sched_domain *sd; 6023 6024 struct rq *src_rq; 6025 int src_cpu; 6026 6027 int dst_cpu; 6028 struct rq *dst_rq; 6029 6030 struct cpumask *dst_grpmask; 6031 int new_dst_cpu; 6032 enum cpu_idle_type idle; 6033 long imbalance; 6034 /* The set of CPUs under consideration for load-balancing */ 6035 struct cpumask *cpus; 6036 6037 unsigned int flags; 6038 6039 unsigned int loop; 6040 unsigned int loop_break; 6041 unsigned int loop_max; 6042 6043 enum fbq_type fbq_type; 6044 struct list_head tasks; 6045 }; 6046 6047 /* 6048 * Is this task likely cache-hot: 6049 */ 6050 static int task_hot(struct task_struct *p, struct lb_env *env) 6051 { 6052 s64 delta; 6053 6054 lockdep_assert_held(&env->src_rq->lock); 6055 6056 if (p->sched_class != &fair_sched_class) 6057 return 0; 6058 6059 if (unlikely(p->policy == SCHED_IDLE)) 6060 return 0; 6061 6062 /* 6063 * Buddy candidates are cache hot: 6064 */ 6065 if (sched_feat(CACHE_HOT_BUDDY) && env->dst_rq->nr_running && 6066 (&p->se == cfs_rq_of(&p->se)->next || 6067 &p->se == cfs_rq_of(&p->se)->last)) 6068 return 1; 6069 6070 if (sysctl_sched_migration_cost == -1) 6071 return 1; 6072 if (sysctl_sched_migration_cost == 0) 6073 return 0; 6074 6075 delta = rq_clock_task(env->src_rq) - p->se.exec_start; 6076 6077 return delta < (s64)sysctl_sched_migration_cost; 6078 } 6079 6080 #ifdef CONFIG_NUMA_BALANCING 6081 /* 6082 * Returns 1, if task migration degrades locality 6083 * Returns 0, if task migration improves locality i.e migration preferred. 6084 * Returns -1, if task migration is not affected by locality. 6085 */ 6086 static int migrate_degrades_locality(struct task_struct *p, struct lb_env *env) 6087 { 6088 struct numa_group *numa_group = rcu_dereference(p->numa_group); 6089 unsigned long src_faults, dst_faults; 6090 int src_nid, dst_nid; 6091 6092 if (!static_branch_likely(&sched_numa_balancing)) 6093 return -1; 6094 6095 if (!p->numa_faults || !(env->sd->flags & SD_NUMA)) 6096 return -1; 6097 6098 src_nid = cpu_to_node(env->src_cpu); 6099 dst_nid = cpu_to_node(env->dst_cpu); 6100 6101 if (src_nid == dst_nid) 6102 return -1; 6103 6104 /* Migrating away from the preferred node is always bad. */ 6105 if (src_nid == p->numa_preferred_nid) { 6106 if (env->src_rq->nr_running > env->src_rq->nr_preferred_running) 6107 return 1; 6108 else 6109 return -1; 6110 } 6111 6112 /* Encourage migration to the preferred node. */ 6113 if (dst_nid == p->numa_preferred_nid) 6114 return 0; 6115 6116 if (numa_group) { 6117 src_faults = group_faults(p, src_nid); 6118 dst_faults = group_faults(p, dst_nid); 6119 } else { 6120 src_faults = task_faults(p, src_nid); 6121 dst_faults = task_faults(p, dst_nid); 6122 } 6123 6124 return dst_faults < src_faults; 6125 } 6126 6127 #else 6128 static inline int migrate_degrades_locality(struct task_struct *p, 6129 struct lb_env *env) 6130 { 6131 return -1; 6132 } 6133 #endif 6134 6135 /* 6136 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu? 6137 */ 6138 static 6139 int can_migrate_task(struct task_struct *p, struct lb_env *env) 6140 { 6141 int tsk_cache_hot; 6142 6143 lockdep_assert_held(&env->src_rq->lock); 6144 6145 /* 6146 * We do not migrate tasks that are: 6147 * 1) throttled_lb_pair, or 6148 * 2) cannot be migrated to this CPU due to cpus_allowed, or 6149 * 3) running (obviously), or 6150 * 4) are cache-hot on their current CPU. 6151 */ 6152 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu)) 6153 return 0; 6154 6155 if (!cpumask_test_cpu(env->dst_cpu, tsk_cpus_allowed(p))) { 6156 int cpu; 6157 6158 schedstat_inc(p, se.statistics.nr_failed_migrations_affine); 6159 6160 env->flags |= LBF_SOME_PINNED; 6161 6162 /* 6163 * Remember if this task can be migrated to any other cpu in 6164 * our sched_group. We may want to revisit it if we couldn't 6165 * meet load balance goals by pulling other tasks on src_cpu. 6166 * 6167 * Also avoid computing new_dst_cpu if we have already computed 6168 * one in current iteration. 6169 */ 6170 if (!env->dst_grpmask || (env->flags & LBF_DST_PINNED)) 6171 return 0; 6172 6173 /* Prevent to re-select dst_cpu via env's cpus */ 6174 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) { 6175 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) { 6176 env->flags |= LBF_DST_PINNED; 6177 env->new_dst_cpu = cpu; 6178 break; 6179 } 6180 } 6181 6182 return 0; 6183 } 6184 6185 /* Record that we found atleast one task that could run on dst_cpu */ 6186 env->flags &= ~LBF_ALL_PINNED; 6187 6188 if (task_running(env->src_rq, p)) { 6189 schedstat_inc(p, se.statistics.nr_failed_migrations_running); 6190 return 0; 6191 } 6192 6193 /* 6194 * Aggressive migration if: 6195 * 1) destination numa is preferred 6196 * 2) task is cache cold, or 6197 * 3) too many balance attempts have failed. 6198 */ 6199 tsk_cache_hot = migrate_degrades_locality(p, env); 6200 if (tsk_cache_hot == -1) 6201 tsk_cache_hot = task_hot(p, env); 6202 6203 if (tsk_cache_hot <= 0 || 6204 env->sd->nr_balance_failed > env->sd->cache_nice_tries) { 6205 if (tsk_cache_hot == 1) { 6206 schedstat_inc(env->sd, lb_hot_gained[env->idle]); 6207 schedstat_inc(p, se.statistics.nr_forced_migrations); 6208 } 6209 return 1; 6210 } 6211 6212 schedstat_inc(p, se.statistics.nr_failed_migrations_hot); 6213 return 0; 6214 } 6215 6216 /* 6217 * detach_task() -- detach the task for the migration specified in env 6218 */ 6219 static void detach_task(struct task_struct *p, struct lb_env *env) 6220 { 6221 lockdep_assert_held(&env->src_rq->lock); 6222 6223 p->on_rq = TASK_ON_RQ_MIGRATING; 6224 deactivate_task(env->src_rq, p, 0); 6225 set_task_cpu(p, env->dst_cpu); 6226 } 6227 6228 /* 6229 * detach_one_task() -- tries to dequeue exactly one task from env->src_rq, as 6230 * part of active balancing operations within "domain". 6231 * 6232 * Returns a task if successful and NULL otherwise. 6233 */ 6234 static struct task_struct *detach_one_task(struct lb_env *env) 6235 { 6236 struct task_struct *p, *n; 6237 6238 lockdep_assert_held(&env->src_rq->lock); 6239 6240 list_for_each_entry_safe(p, n, &env->src_rq->cfs_tasks, se.group_node) { 6241 if (!can_migrate_task(p, env)) 6242 continue; 6243 6244 detach_task(p, env); 6245 6246 /* 6247 * Right now, this is only the second place where 6248 * lb_gained[env->idle] is updated (other is detach_tasks) 6249 * so we can safely collect stats here rather than 6250 * inside detach_tasks(). 6251 */ 6252 schedstat_inc(env->sd, lb_gained[env->idle]); 6253 return p; 6254 } 6255 return NULL; 6256 } 6257 6258 static const unsigned int sched_nr_migrate_break = 32; 6259 6260 /* 6261 * detach_tasks() -- tries to detach up to imbalance weighted load from 6262 * busiest_rq, as part of a balancing operation within domain "sd". 6263 * 6264 * Returns number of detached tasks if successful and 0 otherwise. 6265 */ 6266 static int detach_tasks(struct lb_env *env) 6267 { 6268 struct list_head *tasks = &env->src_rq->cfs_tasks; 6269 struct task_struct *p; 6270 unsigned long load; 6271 int detached = 0; 6272 6273 lockdep_assert_held(&env->src_rq->lock); 6274 6275 if (env->imbalance <= 0) 6276 return 0; 6277 6278 while (!list_empty(tasks)) { 6279 /* 6280 * We don't want to steal all, otherwise we may be treated likewise, 6281 * which could at worst lead to a livelock crash. 6282 */ 6283 if (env->idle != CPU_NOT_IDLE && env->src_rq->nr_running <= 1) 6284 break; 6285 6286 p = list_first_entry(tasks, struct task_struct, se.group_node); 6287 6288 env->loop++; 6289 /* We've more or less seen every task there is, call it quits */ 6290 if (env->loop > env->loop_max) 6291 break; 6292 6293 /* take a breather every nr_migrate tasks */ 6294 if (env->loop > env->loop_break) { 6295 env->loop_break += sched_nr_migrate_break; 6296 env->flags |= LBF_NEED_BREAK; 6297 break; 6298 } 6299 6300 if (!can_migrate_task(p, env)) 6301 goto next; 6302 6303 load = task_h_load(p); 6304 6305 if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed) 6306 goto next; 6307 6308 if ((load / 2) > env->imbalance) 6309 goto next; 6310 6311 detach_task(p, env); 6312 list_add(&p->se.group_node, &env->tasks); 6313 6314 detached++; 6315 env->imbalance -= load; 6316 6317 #ifdef CONFIG_PREEMPT 6318 /* 6319 * NEWIDLE balancing is a source of latency, so preemptible 6320 * kernels will stop after the first task is detached to minimize 6321 * the critical section. 6322 */ 6323 if (env->idle == CPU_NEWLY_IDLE) 6324 break; 6325 #endif 6326 6327 /* 6328 * We only want to steal up to the prescribed amount of 6329 * weighted load. 6330 */ 6331 if (env->imbalance <= 0) 6332 break; 6333 6334 continue; 6335 next: 6336 list_move_tail(&p->se.group_node, tasks); 6337 } 6338 6339 /* 6340 * Right now, this is one of only two places we collect this stat 6341 * so we can safely collect detach_one_task() stats here rather 6342 * than inside detach_one_task(). 6343 */ 6344 schedstat_add(env->sd, lb_gained[env->idle], detached); 6345 6346 return detached; 6347 } 6348 6349 /* 6350 * attach_task() -- attach the task detached by detach_task() to its new rq. 6351 */ 6352 static void attach_task(struct rq *rq, struct task_struct *p) 6353 { 6354 lockdep_assert_held(&rq->lock); 6355 6356 BUG_ON(task_rq(p) != rq); 6357 activate_task(rq, p, 0); 6358 p->on_rq = TASK_ON_RQ_QUEUED; 6359 check_preempt_curr(rq, p, 0); 6360 } 6361 6362 /* 6363 * attach_one_task() -- attaches the task returned from detach_one_task() to 6364 * its new rq. 6365 */ 6366 static void attach_one_task(struct rq *rq, struct task_struct *p) 6367 { 6368 raw_spin_lock(&rq->lock); 6369 attach_task(rq, p); 6370 raw_spin_unlock(&rq->lock); 6371 } 6372 6373 /* 6374 * attach_tasks() -- attaches all tasks detached by detach_tasks() to their 6375 * new rq. 6376 */ 6377 static void attach_tasks(struct lb_env *env) 6378 { 6379 struct list_head *tasks = &env->tasks; 6380 struct task_struct *p; 6381 6382 raw_spin_lock(&env->dst_rq->lock); 6383 6384 while (!list_empty(tasks)) { 6385 p = list_first_entry(tasks, struct task_struct, se.group_node); 6386 list_del_init(&p->se.group_node); 6387 6388 attach_task(env->dst_rq, p); 6389 } 6390 6391 raw_spin_unlock(&env->dst_rq->lock); 6392 } 6393 6394 #ifdef CONFIG_FAIR_GROUP_SCHED 6395 static void update_blocked_averages(int cpu) 6396 { 6397 struct rq *rq = cpu_rq(cpu); 6398 struct cfs_rq *cfs_rq; 6399 unsigned long flags; 6400 6401 raw_spin_lock_irqsave(&rq->lock, flags); 6402 update_rq_clock(rq); 6403 6404 /* 6405 * Iterates the task_group tree in a bottom up fashion, see 6406 * list_add_leaf_cfs_rq() for details. 6407 */ 6408 for_each_leaf_cfs_rq(rq, cfs_rq) { 6409 /* throttled entities do not contribute to load */ 6410 if (throttled_hierarchy(cfs_rq)) 6411 continue; 6412 6413 if (update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq, true)) 6414 update_tg_load_avg(cfs_rq, 0); 6415 } 6416 raw_spin_unlock_irqrestore(&rq->lock, flags); 6417 } 6418 6419 /* 6420 * Compute the hierarchical load factor for cfs_rq and all its ascendants. 6421 * This needs to be done in a top-down fashion because the load of a child 6422 * group is a fraction of its parents load. 6423 */ 6424 static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq) 6425 { 6426 struct rq *rq = rq_of(cfs_rq); 6427 struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)]; 6428 unsigned long now = jiffies; 6429 unsigned long load; 6430 6431 if (cfs_rq->last_h_load_update == now) 6432 return; 6433 6434 cfs_rq->h_load_next = NULL; 6435 for_each_sched_entity(se) { 6436 cfs_rq = cfs_rq_of(se); 6437 cfs_rq->h_load_next = se; 6438 if (cfs_rq->last_h_load_update == now) 6439 break; 6440 } 6441 6442 if (!se) { 6443 cfs_rq->h_load = cfs_rq_load_avg(cfs_rq); 6444 cfs_rq->last_h_load_update = now; 6445 } 6446 6447 while ((se = cfs_rq->h_load_next) != NULL) { 6448 load = cfs_rq->h_load; 6449 load = div64_ul(load * se->avg.load_avg, 6450 cfs_rq_load_avg(cfs_rq) + 1); 6451 cfs_rq = group_cfs_rq(se); 6452 cfs_rq->h_load = load; 6453 cfs_rq->last_h_load_update = now; 6454 } 6455 } 6456 6457 static unsigned long task_h_load(struct task_struct *p) 6458 { 6459 struct cfs_rq *cfs_rq = task_cfs_rq(p); 6460 6461 update_cfs_rq_h_load(cfs_rq); 6462 return div64_ul(p->se.avg.load_avg * cfs_rq->h_load, 6463 cfs_rq_load_avg(cfs_rq) + 1); 6464 } 6465 #else 6466 static inline void update_blocked_averages(int cpu) 6467 { 6468 struct rq *rq = cpu_rq(cpu); 6469 struct cfs_rq *cfs_rq = &rq->cfs; 6470 unsigned long flags; 6471 6472 raw_spin_lock_irqsave(&rq->lock, flags); 6473 update_rq_clock(rq); 6474 update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq, true); 6475 raw_spin_unlock_irqrestore(&rq->lock, flags); 6476 } 6477 6478 static unsigned long task_h_load(struct task_struct *p) 6479 { 6480 return p->se.avg.load_avg; 6481 } 6482 #endif 6483 6484 /********** Helpers for find_busiest_group ************************/ 6485 6486 enum group_type { 6487 group_other = 0, 6488 group_imbalanced, 6489 group_overloaded, 6490 }; 6491 6492 /* 6493 * sg_lb_stats - stats of a sched_group required for load_balancing 6494 */ 6495 struct sg_lb_stats { 6496 unsigned long avg_load; /*Avg load across the CPUs of the group */ 6497 unsigned long group_load; /* Total load over the CPUs of the group */ 6498 unsigned long sum_weighted_load; /* Weighted load of group's tasks */ 6499 unsigned long load_per_task; 6500 unsigned long group_capacity; 6501 unsigned long group_util; /* Total utilization of the group */ 6502 unsigned int sum_nr_running; /* Nr tasks running in the group */ 6503 unsigned int idle_cpus; 6504 unsigned int group_weight; 6505 enum group_type group_type; 6506 int group_no_capacity; 6507 #ifdef CONFIG_NUMA_BALANCING 6508 unsigned int nr_numa_running; 6509 unsigned int nr_preferred_running; 6510 #endif 6511 }; 6512 6513 /* 6514 * sd_lb_stats - Structure to store the statistics of a sched_domain 6515 * during load balancing. 6516 */ 6517 struct sd_lb_stats { 6518 struct sched_group *busiest; /* Busiest group in this sd */ 6519 struct sched_group *local; /* Local group in this sd */ 6520 unsigned long total_load; /* Total load of all groups in sd */ 6521 unsigned long total_capacity; /* Total capacity of all groups in sd */ 6522 unsigned long avg_load; /* Average load across all groups in sd */ 6523 6524 struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */ 6525 struct sg_lb_stats local_stat; /* Statistics of the local group */ 6526 }; 6527 6528 static inline void init_sd_lb_stats(struct sd_lb_stats *sds) 6529 { 6530 /* 6531 * Skimp on the clearing to avoid duplicate work. We can avoid clearing 6532 * local_stat because update_sg_lb_stats() does a full clear/assignment. 6533 * We must however clear busiest_stat::avg_load because 6534 * update_sd_pick_busiest() reads this before assignment. 6535 */ 6536 *sds = (struct sd_lb_stats){ 6537 .busiest = NULL, 6538 .local = NULL, 6539 .total_load = 0UL, 6540 .total_capacity = 0UL, 6541 .busiest_stat = { 6542 .avg_load = 0UL, 6543 .sum_nr_running = 0, 6544 .group_type = group_other, 6545 }, 6546 }; 6547 } 6548 6549 /** 6550 * get_sd_load_idx - Obtain the load index for a given sched domain. 6551 * @sd: The sched_domain whose load_idx is to be obtained. 6552 * @idle: The idle status of the CPU for whose sd load_idx is obtained. 6553 * 6554 * Return: The load index. 6555 */ 6556 static inline int get_sd_load_idx(struct sched_domain *sd, 6557 enum cpu_idle_type idle) 6558 { 6559 int load_idx; 6560 6561 switch (idle) { 6562 case CPU_NOT_IDLE: 6563 load_idx = sd->busy_idx; 6564 break; 6565 6566 case CPU_NEWLY_IDLE: 6567 load_idx = sd->newidle_idx; 6568 break; 6569 default: 6570 load_idx = sd->idle_idx; 6571 break; 6572 } 6573 6574 return load_idx; 6575 } 6576 6577 static unsigned long scale_rt_capacity(int cpu) 6578 { 6579 struct rq *rq = cpu_rq(cpu); 6580 u64 total, used, age_stamp, avg; 6581 s64 delta; 6582 6583 /* 6584 * Since we're reading these variables without serialization make sure 6585 * we read them once before doing sanity checks on them. 6586 */ 6587 age_stamp = READ_ONCE(rq->age_stamp); 6588 avg = READ_ONCE(rq->rt_avg); 6589 delta = __rq_clock_broken(rq) - age_stamp; 6590 6591 if (unlikely(delta < 0)) 6592 delta = 0; 6593 6594 total = sched_avg_period() + delta; 6595 6596 used = div_u64(avg, total); 6597 6598 if (likely(used < SCHED_CAPACITY_SCALE)) 6599 return SCHED_CAPACITY_SCALE - used; 6600 6601 return 1; 6602 } 6603 6604 static void update_cpu_capacity(struct sched_domain *sd, int cpu) 6605 { 6606 unsigned long capacity = arch_scale_cpu_capacity(sd, cpu); 6607 struct sched_group *sdg = sd->groups; 6608 6609 cpu_rq(cpu)->cpu_capacity_orig = capacity; 6610 6611 capacity *= scale_rt_capacity(cpu); 6612 capacity >>= SCHED_CAPACITY_SHIFT; 6613 6614 if (!capacity) 6615 capacity = 1; 6616 6617 cpu_rq(cpu)->cpu_capacity = capacity; 6618 sdg->sgc->capacity = capacity; 6619 } 6620 6621 void update_group_capacity(struct sched_domain *sd, int cpu) 6622 { 6623 struct sched_domain *child = sd->child; 6624 struct sched_group *group, *sdg = sd->groups; 6625 unsigned long capacity; 6626 unsigned long interval; 6627 6628 interval = msecs_to_jiffies(sd->balance_interval); 6629 interval = clamp(interval, 1UL, max_load_balance_interval); 6630 sdg->sgc->next_update = jiffies + interval; 6631 6632 if (!child) { 6633 update_cpu_capacity(sd, cpu); 6634 return; 6635 } 6636 6637 capacity = 0; 6638 6639 if (child->flags & SD_OVERLAP) { 6640 /* 6641 * SD_OVERLAP domains cannot assume that child groups 6642 * span the current group. 6643 */ 6644 6645 for_each_cpu(cpu, sched_group_cpus(sdg)) { 6646 struct sched_group_capacity *sgc; 6647 struct rq *rq = cpu_rq(cpu); 6648 6649 /* 6650 * build_sched_domains() -> init_sched_groups_capacity() 6651 * gets here before we've attached the domains to the 6652 * runqueues. 6653 * 6654 * Use capacity_of(), which is set irrespective of domains 6655 * in update_cpu_capacity(). 6656 * 6657 * This avoids capacity from being 0 and 6658 * causing divide-by-zero issues on boot. 6659 */ 6660 if (unlikely(!rq->sd)) { 6661 capacity += capacity_of(cpu); 6662 continue; 6663 } 6664 6665 sgc = rq->sd->groups->sgc; 6666 capacity += sgc->capacity; 6667 } 6668 } else { 6669 /* 6670 * !SD_OVERLAP domains can assume that child groups 6671 * span the current group. 6672 */ 6673 6674 group = child->groups; 6675 do { 6676 capacity += group->sgc->capacity; 6677 group = group->next; 6678 } while (group != child->groups); 6679 } 6680 6681 sdg->sgc->capacity = capacity; 6682 } 6683 6684 /* 6685 * Check whether the capacity of the rq has been noticeably reduced by side 6686 * activity. The imbalance_pct is used for the threshold. 6687 * Return true is the capacity is reduced 6688 */ 6689 static inline int 6690 check_cpu_capacity(struct rq *rq, struct sched_domain *sd) 6691 { 6692 return ((rq->cpu_capacity * sd->imbalance_pct) < 6693 (rq->cpu_capacity_orig * 100)); 6694 } 6695 6696 /* 6697 * Group imbalance indicates (and tries to solve) the problem where balancing 6698 * groups is inadequate due to tsk_cpus_allowed() constraints. 6699 * 6700 * Imagine a situation of two groups of 4 cpus each and 4 tasks each with a 6701 * cpumask covering 1 cpu of the first group and 3 cpus of the second group. 6702 * Something like: 6703 * 6704 * { 0 1 2 3 } { 4 5 6 7 } 6705 * * * * * 6706 * 6707 * If we were to balance group-wise we'd place two tasks in the first group and 6708 * two tasks in the second group. Clearly this is undesired as it will overload 6709 * cpu 3 and leave one of the cpus in the second group unused. 6710 * 6711 * The current solution to this issue is detecting the skew in the first group 6712 * by noticing the lower domain failed to reach balance and had difficulty 6713 * moving tasks due to affinity constraints. 6714 * 6715 * When this is so detected; this group becomes a candidate for busiest; see 6716 * update_sd_pick_busiest(). And calculate_imbalance() and 6717 * find_busiest_group() avoid some of the usual balance conditions to allow it 6718 * to create an effective group imbalance. 6719 * 6720 * This is a somewhat tricky proposition since the next run might not find the 6721 * group imbalance and decide the groups need to be balanced again. A most 6722 * subtle and fragile situation. 6723 */ 6724 6725 static inline int sg_imbalanced(struct sched_group *group) 6726 { 6727 return group->sgc->imbalance; 6728 } 6729 6730 /* 6731 * group_has_capacity returns true if the group has spare capacity that could 6732 * be used by some tasks. 6733 * We consider that a group has spare capacity if the * number of task is 6734 * smaller than the number of CPUs or if the utilization is lower than the 6735 * available capacity for CFS tasks. 6736 * For the latter, we use a threshold to stabilize the state, to take into 6737 * account the variance of the tasks' load and to return true if the available 6738 * capacity in meaningful for the load balancer. 6739 * As an example, an available capacity of 1% can appear but it doesn't make 6740 * any benefit for the load balance. 6741 */ 6742 static inline bool 6743 group_has_capacity(struct lb_env *env, struct sg_lb_stats *sgs) 6744 { 6745 if (sgs->sum_nr_running < sgs->group_weight) 6746 return true; 6747 6748 if ((sgs->group_capacity * 100) > 6749 (sgs->group_util * env->sd->imbalance_pct)) 6750 return true; 6751 6752 return false; 6753 } 6754 6755 /* 6756 * group_is_overloaded returns true if the group has more tasks than it can 6757 * handle. 6758 * group_is_overloaded is not equals to !group_has_capacity because a group 6759 * with the exact right number of tasks, has no more spare capacity but is not 6760 * overloaded so both group_has_capacity and group_is_overloaded return 6761 * false. 6762 */ 6763 static inline bool 6764 group_is_overloaded(struct lb_env *env, struct sg_lb_stats *sgs) 6765 { 6766 if (sgs->sum_nr_running <= sgs->group_weight) 6767 return false; 6768 6769 if ((sgs->group_capacity * 100) < 6770 (sgs->group_util * env->sd->imbalance_pct)) 6771 return true; 6772 6773 return false; 6774 } 6775 6776 static inline enum 6777 group_type group_classify(struct sched_group *group, 6778 struct sg_lb_stats *sgs) 6779 { 6780 if (sgs->group_no_capacity) 6781 return group_overloaded; 6782 6783 if (sg_imbalanced(group)) 6784 return group_imbalanced; 6785 6786 return group_other; 6787 } 6788 6789 /** 6790 * update_sg_lb_stats - Update sched_group's statistics for load balancing. 6791 * @env: The load balancing environment. 6792 * @group: sched_group whose statistics are to be updated. 6793 * @load_idx: Load index of sched_domain of this_cpu for load calc. 6794 * @local_group: Does group contain this_cpu. 6795 * @sgs: variable to hold the statistics for this group. 6796 * @overload: Indicate more than one runnable task for any CPU. 6797 */ 6798 static inline void update_sg_lb_stats(struct lb_env *env, 6799 struct sched_group *group, int load_idx, 6800 int local_group, struct sg_lb_stats *sgs, 6801 bool *overload) 6802 { 6803 unsigned long load; 6804 int i, nr_running; 6805 6806 memset(sgs, 0, sizeof(*sgs)); 6807 6808 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) { 6809 struct rq *rq = cpu_rq(i); 6810 6811 /* Bias balancing toward cpus of our domain */ 6812 if (local_group) 6813 load = target_load(i, load_idx); 6814 else 6815 load = source_load(i, load_idx); 6816 6817 sgs->group_load += load; 6818 sgs->group_util += cpu_util(i); 6819 sgs->sum_nr_running += rq->cfs.h_nr_running; 6820 6821 nr_running = rq->nr_running; 6822 if (nr_running > 1) 6823 *overload = true; 6824 6825 #ifdef CONFIG_NUMA_BALANCING 6826 sgs->nr_numa_running += rq->nr_numa_running; 6827 sgs->nr_preferred_running += rq->nr_preferred_running; 6828 #endif 6829 sgs->sum_weighted_load += weighted_cpuload(i); 6830 /* 6831 * No need to call idle_cpu() if nr_running is not 0 6832 */ 6833 if (!nr_running && idle_cpu(i)) 6834 sgs->idle_cpus++; 6835 } 6836 6837 /* Adjust by relative CPU capacity of the group */ 6838 sgs->group_capacity = group->sgc->capacity; 6839 sgs->avg_load = (sgs->group_load*SCHED_CAPACITY_SCALE) / sgs->group_capacity; 6840 6841 if (sgs->sum_nr_running) 6842 sgs->load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running; 6843 6844 sgs->group_weight = group->group_weight; 6845 6846 sgs->group_no_capacity = group_is_overloaded(env, sgs); 6847 sgs->group_type = group_classify(group, sgs); 6848 } 6849 6850 /** 6851 * update_sd_pick_busiest - return 1 on busiest group 6852 * @env: The load balancing environment. 6853 * @sds: sched_domain statistics 6854 * @sg: sched_group candidate to be checked for being the busiest 6855 * @sgs: sched_group statistics 6856 * 6857 * Determine if @sg is a busier group than the previously selected 6858 * busiest group. 6859 * 6860 * Return: %true if @sg is a busier group than the previously selected 6861 * busiest group. %false otherwise. 6862 */ 6863 static bool update_sd_pick_busiest(struct lb_env *env, 6864 struct sd_lb_stats *sds, 6865 struct sched_group *sg, 6866 struct sg_lb_stats *sgs) 6867 { 6868 struct sg_lb_stats *busiest = &sds->busiest_stat; 6869 6870 if (sgs->group_type > busiest->group_type) 6871 return true; 6872 6873 if (sgs->group_type < busiest->group_type) 6874 return false; 6875 6876 if (sgs->avg_load <= busiest->avg_load) 6877 return false; 6878 6879 /* This is the busiest node in its class. */ 6880 if (!(env->sd->flags & SD_ASYM_PACKING)) 6881 return true; 6882 6883 /* No ASYM_PACKING if target cpu is already busy */ 6884 if (env->idle == CPU_NOT_IDLE) 6885 return true; 6886 /* 6887 * ASYM_PACKING needs to move all the work to the lowest 6888 * numbered CPUs in the group, therefore mark all groups 6889 * higher than ourself as busy. 6890 */ 6891 if (sgs->sum_nr_running && env->dst_cpu < group_first_cpu(sg)) { 6892 if (!sds->busiest) 6893 return true; 6894 6895 /* Prefer to move from highest possible cpu's work */ 6896 if (group_first_cpu(sds->busiest) < group_first_cpu(sg)) 6897 return true; 6898 } 6899 6900 return false; 6901 } 6902 6903 #ifdef CONFIG_NUMA_BALANCING 6904 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs) 6905 { 6906 if (sgs->sum_nr_running > sgs->nr_numa_running) 6907 return regular; 6908 if (sgs->sum_nr_running > sgs->nr_preferred_running) 6909 return remote; 6910 return all; 6911 } 6912 6913 static inline enum fbq_type fbq_classify_rq(struct rq *rq) 6914 { 6915 if (rq->nr_running > rq->nr_numa_running) 6916 return regular; 6917 if (rq->nr_running > rq->nr_preferred_running) 6918 return remote; 6919 return all; 6920 } 6921 #else 6922 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs) 6923 { 6924 return all; 6925 } 6926 6927 static inline enum fbq_type fbq_classify_rq(struct rq *rq) 6928 { 6929 return regular; 6930 } 6931 #endif /* CONFIG_NUMA_BALANCING */ 6932 6933 /** 6934 * update_sd_lb_stats - Update sched_domain's statistics for load balancing. 6935 * @env: The load balancing environment. 6936 * @sds: variable to hold the statistics for this sched_domain. 6937 */ 6938 static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds) 6939 { 6940 struct sched_domain *child = env->sd->child; 6941 struct sched_group *sg = env->sd->groups; 6942 struct sg_lb_stats tmp_sgs; 6943 int load_idx, prefer_sibling = 0; 6944 bool overload = false; 6945 6946 if (child && child->flags & SD_PREFER_SIBLING) 6947 prefer_sibling = 1; 6948 6949 load_idx = get_sd_load_idx(env->sd, env->idle); 6950 6951 do { 6952 struct sg_lb_stats *sgs = &tmp_sgs; 6953 int local_group; 6954 6955 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_cpus(sg)); 6956 if (local_group) { 6957 sds->local = sg; 6958 sgs = &sds->local_stat; 6959 6960 if (env->idle != CPU_NEWLY_IDLE || 6961 time_after_eq(jiffies, sg->sgc->next_update)) 6962 update_group_capacity(env->sd, env->dst_cpu); 6963 } 6964 6965 update_sg_lb_stats(env, sg, load_idx, local_group, sgs, 6966 &overload); 6967 6968 if (local_group) 6969 goto next_group; 6970 6971 /* 6972 * In case the child domain prefers tasks go to siblings 6973 * first, lower the sg capacity so that we'll try 6974 * and move all the excess tasks away. We lower the capacity 6975 * of a group only if the local group has the capacity to fit 6976 * these excess tasks. The extra check prevents the case where 6977 * you always pull from the heaviest group when it is already 6978 * under-utilized (possible with a large weight task outweighs 6979 * the tasks on the system). 6980 */ 6981 if (prefer_sibling && sds->local && 6982 group_has_capacity(env, &sds->local_stat) && 6983 (sgs->sum_nr_running > 1)) { 6984 sgs->group_no_capacity = 1; 6985 sgs->group_type = group_classify(sg, sgs); 6986 } 6987 6988 if (update_sd_pick_busiest(env, sds, sg, sgs)) { 6989 sds->busiest = sg; 6990 sds->busiest_stat = *sgs; 6991 } 6992 6993 next_group: 6994 /* Now, start updating sd_lb_stats */ 6995 sds->total_load += sgs->group_load; 6996 sds->total_capacity += sgs->group_capacity; 6997 6998 sg = sg->next; 6999 } while (sg != env->sd->groups); 7000 7001 if (env->sd->flags & SD_NUMA) 7002 env->fbq_type = fbq_classify_group(&sds->busiest_stat); 7003 7004 if (!env->sd->parent) { 7005 /* update overload indicator if we are at root domain */ 7006 if (env->dst_rq->rd->overload != overload) 7007 env->dst_rq->rd->overload = overload; 7008 } 7009 7010 } 7011 7012 /** 7013 * check_asym_packing - Check to see if the group is packed into the 7014 * sched doman. 7015 * 7016 * This is primarily intended to used at the sibling level. Some 7017 * cores like POWER7 prefer to use lower numbered SMT threads. In the 7018 * case of POWER7, it can move to lower SMT modes only when higher 7019 * threads are idle. When in lower SMT modes, the threads will 7020 * perform better since they share less core resources. Hence when we 7021 * have idle threads, we want them to be the higher ones. 7022 * 7023 * This packing function is run on idle threads. It checks to see if 7024 * the busiest CPU in this domain (core in the P7 case) has a higher 7025 * CPU number than the packing function is being run on. Here we are 7026 * assuming lower CPU number will be equivalent to lower a SMT thread 7027 * number. 7028 * 7029 * Return: 1 when packing is required and a task should be moved to 7030 * this CPU. The amount of the imbalance is returned in *imbalance. 7031 * 7032 * @env: The load balancing environment. 7033 * @sds: Statistics of the sched_domain which is to be packed 7034 */ 7035 static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds) 7036 { 7037 int busiest_cpu; 7038 7039 if (!(env->sd->flags & SD_ASYM_PACKING)) 7040 return 0; 7041 7042 if (env->idle == CPU_NOT_IDLE) 7043 return 0; 7044 7045 if (!sds->busiest) 7046 return 0; 7047 7048 busiest_cpu = group_first_cpu(sds->busiest); 7049 if (env->dst_cpu > busiest_cpu) 7050 return 0; 7051 7052 env->imbalance = DIV_ROUND_CLOSEST( 7053 sds->busiest_stat.avg_load * sds->busiest_stat.group_capacity, 7054 SCHED_CAPACITY_SCALE); 7055 7056 return 1; 7057 } 7058 7059 /** 7060 * fix_small_imbalance - Calculate the minor imbalance that exists 7061 * amongst the groups of a sched_domain, during 7062 * load balancing. 7063 * @env: The load balancing environment. 7064 * @sds: Statistics of the sched_domain whose imbalance is to be calculated. 7065 */ 7066 static inline 7067 void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds) 7068 { 7069 unsigned long tmp, capa_now = 0, capa_move = 0; 7070 unsigned int imbn = 2; 7071 unsigned long scaled_busy_load_per_task; 7072 struct sg_lb_stats *local, *busiest; 7073 7074 local = &sds->local_stat; 7075 busiest = &sds->busiest_stat; 7076 7077 if (!local->sum_nr_running) 7078 local->load_per_task = cpu_avg_load_per_task(env->dst_cpu); 7079 else if (busiest->load_per_task > local->load_per_task) 7080 imbn = 1; 7081 7082 scaled_busy_load_per_task = 7083 (busiest->load_per_task * SCHED_CAPACITY_SCALE) / 7084 busiest->group_capacity; 7085 7086 if (busiest->avg_load + scaled_busy_load_per_task >= 7087 local->avg_load + (scaled_busy_load_per_task * imbn)) { 7088 env->imbalance = busiest->load_per_task; 7089 return; 7090 } 7091 7092 /* 7093 * OK, we don't have enough imbalance to justify moving tasks, 7094 * however we may be able to increase total CPU capacity used by 7095 * moving them. 7096 */ 7097 7098 capa_now += busiest->group_capacity * 7099 min(busiest->load_per_task, busiest->avg_load); 7100 capa_now += local->group_capacity * 7101