1 /* 2 * Completely Fair Scheduling (CFS) Class (SCHED_NORMAL/SCHED_BATCH) 3 * 4 * Copyright (C) 2007 Red Hat, Inc., Ingo Molnar <mingo@redhat.com> 5 * 6 * Interactivity improvements by Mike Galbraith 7 * (C) 2007 Mike Galbraith <efault@gmx.de> 8 * 9 * Various enhancements by Dmitry Adamushko. 10 * (C) 2007 Dmitry Adamushko <dmitry.adamushko@gmail.com> 11 * 12 * Group scheduling enhancements by Srivatsa Vaddagiri 13 * Copyright IBM Corporation, 2007 14 * Author: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com> 15 * 16 * Scaled math optimizations by Thomas Gleixner 17 * Copyright (C) 2007, Thomas Gleixner <tglx@linutronix.de> 18 * 19 * Adaptive scheduling granularity, math enhancements by Peter Zijlstra 20 * Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com> 21 */ 22 23 #include <linux/latencytop.h> 24 #include <linux/sched.h> 25 #include <linux/cpumask.h> 26 #include <linux/slab.h> 27 #include <linux/profile.h> 28 #include <linux/interrupt.h> 29 30 #include <trace/events/sched.h> 31 32 #include "sched.h" 33 34 /* 35 * Targeted preemption latency for CPU-bound tasks: 36 * (default: 6ms * (1 + ilog(ncpus)), units: nanoseconds) 37 * 38 * NOTE: this latency value is not the same as the concept of 39 * 'timeslice length' - timeslices in CFS are of variable length 40 * and have no persistent notion like in traditional, time-slice 41 * based scheduling concepts. 42 * 43 * (to see the precise effective timeslice length of your workload, 44 * run vmstat and monitor the context-switches (cs) field) 45 */ 46 unsigned int sysctl_sched_latency = 6000000ULL; 47 unsigned int normalized_sysctl_sched_latency = 6000000ULL; 48 49 /* 50 * The initial- and re-scaling of tunables is configurable 51 * (default SCHED_TUNABLESCALING_LOG = *(1+ilog(ncpus)) 52 * 53 * Options are: 54 * SCHED_TUNABLESCALING_NONE - unscaled, always *1 55 * SCHED_TUNABLESCALING_LOG - scaled logarithmical, *1+ilog(ncpus) 56 * SCHED_TUNABLESCALING_LINEAR - scaled linear, *ncpus 57 */ 58 enum sched_tunable_scaling sysctl_sched_tunable_scaling 59 = SCHED_TUNABLESCALING_LOG; 60 61 /* 62 * Minimal preemption granularity for CPU-bound tasks: 63 * (default: 0.75 msec * (1 + ilog(ncpus)), units: nanoseconds) 64 */ 65 unsigned int sysctl_sched_min_granularity = 750000ULL; 66 unsigned int normalized_sysctl_sched_min_granularity = 750000ULL; 67 68 /* 69 * is kept at sysctl_sched_latency / sysctl_sched_min_granularity 70 */ 71 static unsigned int sched_nr_latency = 8; 72 73 /* 74 * After fork, child runs first. If set to 0 (default) then 75 * parent will (try to) run first. 76 */ 77 unsigned int sysctl_sched_child_runs_first __read_mostly; 78 79 /* 80 * SCHED_OTHER wake-up granularity. 81 * (default: 1 msec * (1 + ilog(ncpus)), units: nanoseconds) 82 * 83 * This option delays the preemption effects of decoupled workloads 84 * and reduces their over-scheduling. Synchronous workloads will still 85 * have immediate wakeup/sleep latencies. 86 */ 87 unsigned int sysctl_sched_wakeup_granularity = 1000000UL; 88 unsigned int normalized_sysctl_sched_wakeup_granularity = 1000000UL; 89 90 const_debug unsigned int sysctl_sched_migration_cost = 500000UL; 91 92 /* 93 * The exponential sliding window over which load is averaged for shares 94 * distribution. 95 * (default: 10msec) 96 */ 97 unsigned int __read_mostly sysctl_sched_shares_window = 10000000UL; 98 99 #ifdef CONFIG_CFS_BANDWIDTH 100 /* 101 * Amount of runtime to allocate from global (tg) to local (per-cfs_rq) pool 102 * each time a cfs_rq requests quota. 103 * 104 * Note: in the case that the slice exceeds the runtime remaining (either due 105 * to consumption or the quota being specified to be smaller than the slice) 106 * we will always only issue the remaining available time. 107 * 108 * default: 5 msec, units: microseconds 109 */ 110 unsigned int sysctl_sched_cfs_bandwidth_slice = 5000UL; 111 #endif 112 113 /* 114 * Increase the granularity value when there are more CPUs, 115 * because with more CPUs the 'effective latency' as visible 116 * to users decreases. But the relationship is not linear, 117 * so pick a second-best guess by going with the log2 of the 118 * number of CPUs. 119 * 120 * This idea comes from the SD scheduler of Con Kolivas: 121 */ 122 static int get_update_sysctl_factor(void) 123 { 124 unsigned int cpus = min_t(int, num_online_cpus(), 8); 125 unsigned int factor; 126 127 switch (sysctl_sched_tunable_scaling) { 128 case SCHED_TUNABLESCALING_NONE: 129 factor = 1; 130 break; 131 case SCHED_TUNABLESCALING_LINEAR: 132 factor = cpus; 133 break; 134 case SCHED_TUNABLESCALING_LOG: 135 default: 136 factor = 1 + ilog2(cpus); 137 break; 138 } 139 140 return factor; 141 } 142 143 static void update_sysctl(void) 144 { 145 unsigned int factor = get_update_sysctl_factor(); 146 147 #define SET_SYSCTL(name) \ 148 (sysctl_##name = (factor) * normalized_sysctl_##name) 149 SET_SYSCTL(sched_min_granularity); 150 SET_SYSCTL(sched_latency); 151 SET_SYSCTL(sched_wakeup_granularity); 152 #undef SET_SYSCTL 153 } 154 155 void sched_init_granularity(void) 156 { 157 update_sysctl(); 158 } 159 160 #if BITS_PER_LONG == 32 161 # define WMULT_CONST (~0UL) 162 #else 163 # define WMULT_CONST (1UL << 32) 164 #endif 165 166 #define WMULT_SHIFT 32 167 168 /* 169 * Shift right and round: 170 */ 171 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y)) 172 173 /* 174 * delta *= weight / lw 175 */ 176 static unsigned long 177 calc_delta_mine(unsigned long delta_exec, unsigned long weight, 178 struct load_weight *lw) 179 { 180 u64 tmp; 181 182 /* 183 * weight can be less than 2^SCHED_LOAD_RESOLUTION for task group sched 184 * entities since MIN_SHARES = 2. Treat weight as 1 if less than 185 * 2^SCHED_LOAD_RESOLUTION. 186 */ 187 if (likely(weight > (1UL << SCHED_LOAD_RESOLUTION))) 188 tmp = (u64)delta_exec * scale_load_down(weight); 189 else 190 tmp = (u64)delta_exec; 191 192 if (!lw->inv_weight) { 193 unsigned long w = scale_load_down(lw->weight); 194 195 if (BITS_PER_LONG > 32 && unlikely(w >= WMULT_CONST)) 196 lw->inv_weight = 1; 197 else if (unlikely(!w)) 198 lw->inv_weight = WMULT_CONST; 199 else 200 lw->inv_weight = WMULT_CONST / w; 201 } 202 203 /* 204 * Check whether we'd overflow the 64-bit multiplication: 205 */ 206 if (unlikely(tmp > WMULT_CONST)) 207 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight, 208 WMULT_SHIFT/2); 209 else 210 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT); 211 212 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX); 213 } 214 215 216 const struct sched_class fair_sched_class; 217 218 /************************************************************** 219 * CFS operations on generic schedulable entities: 220 */ 221 222 #ifdef CONFIG_FAIR_GROUP_SCHED 223 224 /* cpu runqueue to which this cfs_rq is attached */ 225 static inline struct rq *rq_of(struct cfs_rq *cfs_rq) 226 { 227 return cfs_rq->rq; 228 } 229 230 /* An entity is a task if it doesn't "own" a runqueue */ 231 #define entity_is_task(se) (!se->my_q) 232 233 static inline struct task_struct *task_of(struct sched_entity *se) 234 { 235 #ifdef CONFIG_SCHED_DEBUG 236 WARN_ON_ONCE(!entity_is_task(se)); 237 #endif 238 return container_of(se, struct task_struct, se); 239 } 240 241 /* Walk up scheduling entities hierarchy */ 242 #define for_each_sched_entity(se) \ 243 for (; se; se = se->parent) 244 245 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p) 246 { 247 return p->se.cfs_rq; 248 } 249 250 /* runqueue on which this entity is (to be) queued */ 251 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se) 252 { 253 return se->cfs_rq; 254 } 255 256 /* runqueue "owned" by this group */ 257 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp) 258 { 259 return grp->my_q; 260 } 261 262 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq) 263 { 264 if (!cfs_rq->on_list) { 265 /* 266 * Ensure we either appear before our parent (if already 267 * enqueued) or force our parent to appear after us when it is 268 * enqueued. The fact that we always enqueue bottom-up 269 * reduces this to two cases. 270 */ 271 if (cfs_rq->tg->parent && 272 cfs_rq->tg->parent->cfs_rq[cpu_of(rq_of(cfs_rq))]->on_list) { 273 list_add_rcu(&cfs_rq->leaf_cfs_rq_list, 274 &rq_of(cfs_rq)->leaf_cfs_rq_list); 275 } else { 276 list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list, 277 &rq_of(cfs_rq)->leaf_cfs_rq_list); 278 } 279 280 cfs_rq->on_list = 1; 281 } 282 } 283 284 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq) 285 { 286 if (cfs_rq->on_list) { 287 list_del_rcu(&cfs_rq->leaf_cfs_rq_list); 288 cfs_rq->on_list = 0; 289 } 290 } 291 292 /* Iterate thr' all leaf cfs_rq's on a runqueue */ 293 #define for_each_leaf_cfs_rq(rq, cfs_rq) \ 294 list_for_each_entry_rcu(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list) 295 296 /* Do the two (enqueued) entities belong to the same group ? */ 297 static inline int 298 is_same_group(struct sched_entity *se, struct sched_entity *pse) 299 { 300 if (se->cfs_rq == pse->cfs_rq) 301 return 1; 302 303 return 0; 304 } 305 306 static inline struct sched_entity *parent_entity(struct sched_entity *se) 307 { 308 return se->parent; 309 } 310 311 /* return depth at which a sched entity is present in the hierarchy */ 312 static inline int depth_se(struct sched_entity *se) 313 { 314 int depth = 0; 315 316 for_each_sched_entity(se) 317 depth++; 318 319 return depth; 320 } 321 322 static void 323 find_matching_se(struct sched_entity **se, struct sched_entity **pse) 324 { 325 int se_depth, pse_depth; 326 327 /* 328 * preemption test can be made between sibling entities who are in the 329 * same cfs_rq i.e who have a common parent. Walk up the hierarchy of 330 * both tasks until we find their ancestors who are siblings of common 331 * parent. 332 */ 333 334 /* First walk up until both entities are at same depth */ 335 se_depth = depth_se(*se); 336 pse_depth = depth_se(*pse); 337 338 while (se_depth > pse_depth) { 339 se_depth--; 340 *se = parent_entity(*se); 341 } 342 343 while (pse_depth > se_depth) { 344 pse_depth--; 345 *pse = parent_entity(*pse); 346 } 347 348 while (!is_same_group(*se, *pse)) { 349 *se = parent_entity(*se); 350 *pse = parent_entity(*pse); 351 } 352 } 353 354 #else /* !CONFIG_FAIR_GROUP_SCHED */ 355 356 static inline struct task_struct *task_of(struct sched_entity *se) 357 { 358 return container_of(se, struct task_struct, se); 359 } 360 361 static inline struct rq *rq_of(struct cfs_rq *cfs_rq) 362 { 363 return container_of(cfs_rq, struct rq, cfs); 364 } 365 366 #define entity_is_task(se) 1 367 368 #define for_each_sched_entity(se) \ 369 for (; se; se = NULL) 370 371 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p) 372 { 373 return &task_rq(p)->cfs; 374 } 375 376 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se) 377 { 378 struct task_struct *p = task_of(se); 379 struct rq *rq = task_rq(p); 380 381 return &rq->cfs; 382 } 383 384 /* runqueue "owned" by this group */ 385 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp) 386 { 387 return NULL; 388 } 389 390 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq) 391 { 392 } 393 394 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq) 395 { 396 } 397 398 #define for_each_leaf_cfs_rq(rq, cfs_rq) \ 399 for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL) 400 401 static inline int 402 is_same_group(struct sched_entity *se, struct sched_entity *pse) 403 { 404 return 1; 405 } 406 407 static inline struct sched_entity *parent_entity(struct sched_entity *se) 408 { 409 return NULL; 410 } 411 412 static inline void 413 find_matching_se(struct sched_entity **se, struct sched_entity **pse) 414 { 415 } 416 417 #endif /* CONFIG_FAIR_GROUP_SCHED */ 418 419 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, 420 unsigned long delta_exec); 421 422 /************************************************************** 423 * Scheduling class tree data structure manipulation methods: 424 */ 425 426 static inline u64 max_vruntime(u64 min_vruntime, u64 vruntime) 427 { 428 s64 delta = (s64)(vruntime - min_vruntime); 429 if (delta > 0) 430 min_vruntime = vruntime; 431 432 return min_vruntime; 433 } 434 435 static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime) 436 { 437 s64 delta = (s64)(vruntime - min_vruntime); 438 if (delta < 0) 439 min_vruntime = vruntime; 440 441 return min_vruntime; 442 } 443 444 static inline int entity_before(struct sched_entity *a, 445 struct sched_entity *b) 446 { 447 return (s64)(a->vruntime - b->vruntime) < 0; 448 } 449 450 static void update_min_vruntime(struct cfs_rq *cfs_rq) 451 { 452 u64 vruntime = cfs_rq->min_vruntime; 453 454 if (cfs_rq->curr) 455 vruntime = cfs_rq->curr->vruntime; 456 457 if (cfs_rq->rb_leftmost) { 458 struct sched_entity *se = rb_entry(cfs_rq->rb_leftmost, 459 struct sched_entity, 460 run_node); 461 462 if (!cfs_rq->curr) 463 vruntime = se->vruntime; 464 else 465 vruntime = min_vruntime(vruntime, se->vruntime); 466 } 467 468 cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime); 469 #ifndef CONFIG_64BIT 470 smp_wmb(); 471 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime; 472 #endif 473 } 474 475 /* 476 * Enqueue an entity into the rb-tree: 477 */ 478 static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se) 479 { 480 struct rb_node **link = &cfs_rq->tasks_timeline.rb_node; 481 struct rb_node *parent = NULL; 482 struct sched_entity *entry; 483 int leftmost = 1; 484 485 /* 486 * Find the right place in the rbtree: 487 */ 488 while (*link) { 489 parent = *link; 490 entry = rb_entry(parent, struct sched_entity, run_node); 491 /* 492 * We dont care about collisions. Nodes with 493 * the same key stay together. 494 */ 495 if (entity_before(se, entry)) { 496 link = &parent->rb_left; 497 } else { 498 link = &parent->rb_right; 499 leftmost = 0; 500 } 501 } 502 503 /* 504 * Maintain a cache of leftmost tree entries (it is frequently 505 * used): 506 */ 507 if (leftmost) 508 cfs_rq->rb_leftmost = &se->run_node; 509 510 rb_link_node(&se->run_node, parent, link); 511 rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline); 512 } 513 514 static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se) 515 { 516 if (cfs_rq->rb_leftmost == &se->run_node) { 517 struct rb_node *next_node; 518 519 next_node = rb_next(&se->run_node); 520 cfs_rq->rb_leftmost = next_node; 521 } 522 523 rb_erase(&se->run_node, &cfs_rq->tasks_timeline); 524 } 525 526 struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq) 527 { 528 struct rb_node *left = cfs_rq->rb_leftmost; 529 530 if (!left) 531 return NULL; 532 533 return rb_entry(left, struct sched_entity, run_node); 534 } 535 536 static struct sched_entity *__pick_next_entity(struct sched_entity *se) 537 { 538 struct rb_node *next = rb_next(&se->run_node); 539 540 if (!next) 541 return NULL; 542 543 return rb_entry(next, struct sched_entity, run_node); 544 } 545 546 #ifdef CONFIG_SCHED_DEBUG 547 struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq) 548 { 549 struct rb_node *last = rb_last(&cfs_rq->tasks_timeline); 550 551 if (!last) 552 return NULL; 553 554 return rb_entry(last, struct sched_entity, run_node); 555 } 556 557 /************************************************************** 558 * Scheduling class statistics methods: 559 */ 560 561 int sched_proc_update_handler(struct ctl_table *table, int write, 562 void __user *buffer, size_t *lenp, 563 loff_t *ppos) 564 { 565 int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos); 566 int factor = get_update_sysctl_factor(); 567 568 if (ret || !write) 569 return ret; 570 571 sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency, 572 sysctl_sched_min_granularity); 573 574 #define WRT_SYSCTL(name) \ 575 (normalized_sysctl_##name = sysctl_##name / (factor)) 576 WRT_SYSCTL(sched_min_granularity); 577 WRT_SYSCTL(sched_latency); 578 WRT_SYSCTL(sched_wakeup_granularity); 579 #undef WRT_SYSCTL 580 581 return 0; 582 } 583 #endif 584 585 /* 586 * delta /= w 587 */ 588 static inline unsigned long 589 calc_delta_fair(unsigned long delta, struct sched_entity *se) 590 { 591 if (unlikely(se->load.weight != NICE_0_LOAD)) 592 delta = calc_delta_mine(delta, NICE_0_LOAD, &se->load); 593 594 return delta; 595 } 596 597 /* 598 * The idea is to set a period in which each task runs once. 599 * 600 * When there are too many tasks (sysctl_sched_nr_latency) we have to stretch 601 * this period because otherwise the slices get too small. 602 * 603 * p = (nr <= nl) ? l : l*nr/nl 604 */ 605 static u64 __sched_period(unsigned long nr_running) 606 { 607 u64 period = sysctl_sched_latency; 608 unsigned long nr_latency = sched_nr_latency; 609 610 if (unlikely(nr_running > nr_latency)) { 611 period = sysctl_sched_min_granularity; 612 period *= nr_running; 613 } 614 615 return period; 616 } 617 618 /* 619 * We calculate the wall-time slice from the period by taking a part 620 * proportional to the weight. 621 * 622 * s = p*P[w/rw] 623 */ 624 static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se) 625 { 626 u64 slice = __sched_period(cfs_rq->nr_running + !se->on_rq); 627 628 for_each_sched_entity(se) { 629 struct load_weight *load; 630 struct load_weight lw; 631 632 cfs_rq = cfs_rq_of(se); 633 load = &cfs_rq->load; 634 635 if (unlikely(!se->on_rq)) { 636 lw = cfs_rq->load; 637 638 update_load_add(&lw, se->load.weight); 639 load = &lw; 640 } 641 slice = calc_delta_mine(slice, se->load.weight, load); 642 } 643 return slice; 644 } 645 646 /* 647 * We calculate the vruntime slice of a to be inserted task 648 * 649 * vs = s/w 650 */ 651 static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se) 652 { 653 return calc_delta_fair(sched_slice(cfs_rq, se), se); 654 } 655 656 static void update_cfs_load(struct cfs_rq *cfs_rq, int global_update); 657 static void update_cfs_shares(struct cfs_rq *cfs_rq); 658 659 /* 660 * Update the current task's runtime statistics. Skip current tasks that 661 * are not in our scheduling class. 662 */ 663 static inline void 664 __update_curr(struct cfs_rq *cfs_rq, struct sched_entity *curr, 665 unsigned long delta_exec) 666 { 667 unsigned long delta_exec_weighted; 668 669 schedstat_set(curr->statistics.exec_max, 670 max((u64)delta_exec, curr->statistics.exec_max)); 671 672 curr->sum_exec_runtime += delta_exec; 673 schedstat_add(cfs_rq, exec_clock, delta_exec); 674 delta_exec_weighted = calc_delta_fair(delta_exec, curr); 675 676 curr->vruntime += delta_exec_weighted; 677 update_min_vruntime(cfs_rq); 678 679 #if defined CONFIG_SMP && defined CONFIG_FAIR_GROUP_SCHED 680 cfs_rq->load_unacc_exec_time += delta_exec; 681 #endif 682 } 683 684 static void update_curr(struct cfs_rq *cfs_rq) 685 { 686 struct sched_entity *curr = cfs_rq->curr; 687 u64 now = rq_of(cfs_rq)->clock_task; 688 unsigned long delta_exec; 689 690 if (unlikely(!curr)) 691 return; 692 693 /* 694 * Get the amount of time the current task was running 695 * since the last time we changed load (this cannot 696 * overflow on 32 bits): 697 */ 698 delta_exec = (unsigned long)(now - curr->exec_start); 699 if (!delta_exec) 700 return; 701 702 __update_curr(cfs_rq, curr, delta_exec); 703 curr->exec_start = now; 704 705 if (entity_is_task(curr)) { 706 struct task_struct *curtask = task_of(curr); 707 708 trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime); 709 cpuacct_charge(curtask, delta_exec); 710 account_group_exec_runtime(curtask, delta_exec); 711 } 712 713 account_cfs_rq_runtime(cfs_rq, delta_exec); 714 } 715 716 static inline void 717 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se) 718 { 719 schedstat_set(se->statistics.wait_start, rq_of(cfs_rq)->clock); 720 } 721 722 /* 723 * Task is being enqueued - update stats: 724 */ 725 static void update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se) 726 { 727 /* 728 * Are we enqueueing a waiting task? (for current tasks 729 * a dequeue/enqueue event is a NOP) 730 */ 731 if (se != cfs_rq->curr) 732 update_stats_wait_start(cfs_rq, se); 733 } 734 735 static void 736 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se) 737 { 738 schedstat_set(se->statistics.wait_max, max(se->statistics.wait_max, 739 rq_of(cfs_rq)->clock - se->statistics.wait_start)); 740 schedstat_set(se->statistics.wait_count, se->statistics.wait_count + 1); 741 schedstat_set(se->statistics.wait_sum, se->statistics.wait_sum + 742 rq_of(cfs_rq)->clock - se->statistics.wait_start); 743 #ifdef CONFIG_SCHEDSTATS 744 if (entity_is_task(se)) { 745 trace_sched_stat_wait(task_of(se), 746 rq_of(cfs_rq)->clock - se->statistics.wait_start); 747 } 748 #endif 749 schedstat_set(se->statistics.wait_start, 0); 750 } 751 752 static inline void 753 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se) 754 { 755 /* 756 * Mark the end of the wait period if dequeueing a 757 * waiting task: 758 */ 759 if (se != cfs_rq->curr) 760 update_stats_wait_end(cfs_rq, se); 761 } 762 763 /* 764 * We are picking a new current task - update its stats: 765 */ 766 static inline void 767 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se) 768 { 769 /* 770 * We are starting a new run period: 771 */ 772 se->exec_start = rq_of(cfs_rq)->clock_task; 773 } 774 775 /************************************************** 776 * Scheduling class queueing methods: 777 */ 778 779 #if defined CONFIG_SMP && defined CONFIG_FAIR_GROUP_SCHED 780 static void 781 add_cfs_task_weight(struct cfs_rq *cfs_rq, unsigned long weight) 782 { 783 cfs_rq->task_weight += weight; 784 } 785 #else 786 static inline void 787 add_cfs_task_weight(struct cfs_rq *cfs_rq, unsigned long weight) 788 { 789 } 790 #endif 791 792 static void 793 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se) 794 { 795 update_load_add(&cfs_rq->load, se->load.weight); 796 if (!parent_entity(se)) 797 update_load_add(&rq_of(cfs_rq)->load, se->load.weight); 798 if (entity_is_task(se)) { 799 add_cfs_task_weight(cfs_rq, se->load.weight); 800 list_add(&se->group_node, &cfs_rq->tasks); 801 } 802 cfs_rq->nr_running++; 803 } 804 805 static void 806 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se) 807 { 808 update_load_sub(&cfs_rq->load, se->load.weight); 809 if (!parent_entity(se)) 810 update_load_sub(&rq_of(cfs_rq)->load, se->load.weight); 811 if (entity_is_task(se)) { 812 add_cfs_task_weight(cfs_rq, -se->load.weight); 813 list_del_init(&se->group_node); 814 } 815 cfs_rq->nr_running--; 816 } 817 818 #ifdef CONFIG_FAIR_GROUP_SCHED 819 /* we need this in update_cfs_load and load-balance functions below */ 820 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq); 821 # ifdef CONFIG_SMP 822 static void update_cfs_rq_load_contribution(struct cfs_rq *cfs_rq, 823 int global_update) 824 { 825 struct task_group *tg = cfs_rq->tg; 826 long load_avg; 827 828 load_avg = div64_u64(cfs_rq->load_avg, cfs_rq->load_period+1); 829 load_avg -= cfs_rq->load_contribution; 830 831 if (global_update || abs(load_avg) > cfs_rq->load_contribution / 8) { 832 atomic_add(load_avg, &tg->load_weight); 833 cfs_rq->load_contribution += load_avg; 834 } 835 } 836 837 static void update_cfs_load(struct cfs_rq *cfs_rq, int global_update) 838 { 839 u64 period = sysctl_sched_shares_window; 840 u64 now, delta; 841 unsigned long load = cfs_rq->load.weight; 842 843 if (cfs_rq->tg == &root_task_group || throttled_hierarchy(cfs_rq)) 844 return; 845 846 now = rq_of(cfs_rq)->clock_task; 847 delta = now - cfs_rq->load_stamp; 848 849 /* truncate load history at 4 idle periods */ 850 if (cfs_rq->load_stamp > cfs_rq->load_last && 851 now - cfs_rq->load_last > 4 * period) { 852 cfs_rq->load_period = 0; 853 cfs_rq->load_avg = 0; 854 delta = period - 1; 855 } 856 857 cfs_rq->load_stamp = now; 858 cfs_rq->load_unacc_exec_time = 0; 859 cfs_rq->load_period += delta; 860 if (load) { 861 cfs_rq->load_last = now; 862 cfs_rq->load_avg += delta * load; 863 } 864 865 /* consider updating load contribution on each fold or truncate */ 866 if (global_update || cfs_rq->load_period > period 867 || !cfs_rq->load_period) 868 update_cfs_rq_load_contribution(cfs_rq, global_update); 869 870 while (cfs_rq->load_period > period) { 871 /* 872 * Inline assembly required to prevent the compiler 873 * optimising this loop into a divmod call. 874 * See __iter_div_u64_rem() for another example of this. 875 */ 876 asm("" : "+rm" (cfs_rq->load_period)); 877 cfs_rq->load_period /= 2; 878 cfs_rq->load_avg /= 2; 879 } 880 881 if (!cfs_rq->curr && !cfs_rq->nr_running && !cfs_rq->load_avg) 882 list_del_leaf_cfs_rq(cfs_rq); 883 } 884 885 static inline long calc_tg_weight(struct task_group *tg, struct cfs_rq *cfs_rq) 886 { 887 long tg_weight; 888 889 /* 890 * Use this CPU's actual weight instead of the last load_contribution 891 * to gain a more accurate current total weight. See 892 * update_cfs_rq_load_contribution(). 893 */ 894 tg_weight = atomic_read(&tg->load_weight); 895 tg_weight -= cfs_rq->load_contribution; 896 tg_weight += cfs_rq->load.weight; 897 898 return tg_weight; 899 } 900 901 static long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg) 902 { 903 long tg_weight, load, shares; 904 905 tg_weight = calc_tg_weight(tg, cfs_rq); 906 load = cfs_rq->load.weight; 907 908 shares = (tg->shares * load); 909 if (tg_weight) 910 shares /= tg_weight; 911 912 if (shares < MIN_SHARES) 913 shares = MIN_SHARES; 914 if (shares > tg->shares) 915 shares = tg->shares; 916 917 return shares; 918 } 919 920 static void update_entity_shares_tick(struct cfs_rq *cfs_rq) 921 { 922 if (cfs_rq->load_unacc_exec_time > sysctl_sched_shares_window) { 923 update_cfs_load(cfs_rq, 0); 924 update_cfs_shares(cfs_rq); 925 } 926 } 927 # else /* CONFIG_SMP */ 928 static void update_cfs_load(struct cfs_rq *cfs_rq, int global_update) 929 { 930 } 931 932 static inline long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg) 933 { 934 return tg->shares; 935 } 936 937 static inline void update_entity_shares_tick(struct cfs_rq *cfs_rq) 938 { 939 } 940 # endif /* CONFIG_SMP */ 941 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, 942 unsigned long weight) 943 { 944 if (se->on_rq) { 945 /* commit outstanding execution time */ 946 if (cfs_rq->curr == se) 947 update_curr(cfs_rq); 948 account_entity_dequeue(cfs_rq, se); 949 } 950 951 update_load_set(&se->load, weight); 952 953 if (se->on_rq) 954 account_entity_enqueue(cfs_rq, se); 955 } 956 957 static void update_cfs_shares(struct cfs_rq *cfs_rq) 958 { 959 struct task_group *tg; 960 struct sched_entity *se; 961 long shares; 962 963 tg = cfs_rq->tg; 964 se = tg->se[cpu_of(rq_of(cfs_rq))]; 965 if (!se || throttled_hierarchy(cfs_rq)) 966 return; 967 #ifndef CONFIG_SMP 968 if (likely(se->load.weight == tg->shares)) 969 return; 970 #endif 971 shares = calc_cfs_shares(cfs_rq, tg); 972 973 reweight_entity(cfs_rq_of(se), se, shares); 974 } 975 #else /* CONFIG_FAIR_GROUP_SCHED */ 976 static void update_cfs_load(struct cfs_rq *cfs_rq, int global_update) 977 { 978 } 979 980 static inline void update_cfs_shares(struct cfs_rq *cfs_rq) 981 { 982 } 983 984 static inline void update_entity_shares_tick(struct cfs_rq *cfs_rq) 985 { 986 } 987 #endif /* CONFIG_FAIR_GROUP_SCHED */ 988 989 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se) 990 { 991 #ifdef CONFIG_SCHEDSTATS 992 struct task_struct *tsk = NULL; 993 994 if (entity_is_task(se)) 995 tsk = task_of(se); 996 997 if (se->statistics.sleep_start) { 998 u64 delta = rq_of(cfs_rq)->clock - se->statistics.sleep_start; 999 1000 if ((s64)delta < 0) 1001 delta = 0; 1002 1003 if (unlikely(delta > se->statistics.sleep_max)) 1004 se->statistics.sleep_max = delta; 1005 1006 se->statistics.sum_sleep_runtime += delta; 1007 1008 if (tsk) { 1009 account_scheduler_latency(tsk, delta >> 10, 1); 1010 trace_sched_stat_sleep(tsk, delta); 1011 } 1012 } 1013 if (se->statistics.block_start) { 1014 u64 delta = rq_of(cfs_rq)->clock - se->statistics.block_start; 1015 1016 if ((s64)delta < 0) 1017 delta = 0; 1018 1019 if (unlikely(delta > se->statistics.block_max)) 1020 se->statistics.block_max = delta; 1021 1022 se->statistics.sum_sleep_runtime += delta; 1023 1024 if (tsk) { 1025 if (tsk->in_iowait) { 1026 se->statistics.iowait_sum += delta; 1027 se->statistics.iowait_count++; 1028 trace_sched_stat_iowait(tsk, delta); 1029 } 1030 1031 trace_sched_stat_blocked(tsk, delta); 1032 1033 /* 1034 * Blocking time is in units of nanosecs, so shift by 1035 * 20 to get a milliseconds-range estimation of the 1036 * amount of time that the task spent sleeping: 1037 */ 1038 if (unlikely(prof_on == SLEEP_PROFILING)) { 1039 profile_hits(SLEEP_PROFILING, 1040 (void *)get_wchan(tsk), 1041 delta >> 20); 1042 } 1043 account_scheduler_latency(tsk, delta >> 10, 0); 1044 } 1045 } 1046 #endif 1047 } 1048 1049 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se) 1050 { 1051 #ifdef CONFIG_SCHED_DEBUG 1052 s64 d = se->vruntime - cfs_rq->min_vruntime; 1053 1054 if (d < 0) 1055 d = -d; 1056 1057 if (d > 3*sysctl_sched_latency) 1058 schedstat_inc(cfs_rq, nr_spread_over); 1059 #endif 1060 } 1061 1062 static void 1063 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial) 1064 { 1065 u64 vruntime = cfs_rq->min_vruntime; 1066 1067 /* 1068 * The 'current' period is already promised to the current tasks, 1069 * however the extra weight of the new task will slow them down a 1070 * little, place the new task so that it fits in the slot that 1071 * stays open at the end. 1072 */ 1073 if (initial && sched_feat(START_DEBIT)) 1074 vruntime += sched_vslice(cfs_rq, se); 1075 1076 /* sleeps up to a single latency don't count. */ 1077 if (!initial) { 1078 unsigned long thresh = sysctl_sched_latency; 1079 1080 /* 1081 * Halve their sleep time's effect, to allow 1082 * for a gentler effect of sleepers: 1083 */ 1084 if (sched_feat(GENTLE_FAIR_SLEEPERS)) 1085 thresh >>= 1; 1086 1087 vruntime -= thresh; 1088 } 1089 1090 /* ensure we never gain time by being placed backwards. */ 1091 vruntime = max_vruntime(se->vruntime, vruntime); 1092 1093 se->vruntime = vruntime; 1094 } 1095 1096 static void check_enqueue_throttle(struct cfs_rq *cfs_rq); 1097 1098 static void 1099 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags) 1100 { 1101 /* 1102 * Update the normalized vruntime before updating min_vruntime 1103 * through callig update_curr(). 1104 */ 1105 if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_WAKING)) 1106 se->vruntime += cfs_rq->min_vruntime; 1107 1108 /* 1109 * Update run-time statistics of the 'current'. 1110 */ 1111 update_curr(cfs_rq); 1112 update_cfs_load(cfs_rq, 0); 1113 account_entity_enqueue(cfs_rq, se); 1114 update_cfs_shares(cfs_rq); 1115 1116 if (flags & ENQUEUE_WAKEUP) { 1117 place_entity(cfs_rq, se, 0); 1118 enqueue_sleeper(cfs_rq, se); 1119 } 1120 1121 update_stats_enqueue(cfs_rq, se); 1122 check_spread(cfs_rq, se); 1123 if (se != cfs_rq->curr) 1124 __enqueue_entity(cfs_rq, se); 1125 se->on_rq = 1; 1126 1127 if (cfs_rq->nr_running == 1) { 1128 list_add_leaf_cfs_rq(cfs_rq); 1129 check_enqueue_throttle(cfs_rq); 1130 } 1131 } 1132 1133 static void __clear_buddies_last(struct sched_entity *se) 1134 { 1135 for_each_sched_entity(se) { 1136 struct cfs_rq *cfs_rq = cfs_rq_of(se); 1137 if (cfs_rq->last == se) 1138 cfs_rq->last = NULL; 1139 else 1140 break; 1141 } 1142 } 1143 1144 static void __clear_buddies_next(struct sched_entity *se) 1145 { 1146 for_each_sched_entity(se) { 1147 struct cfs_rq *cfs_rq = cfs_rq_of(se); 1148 if (cfs_rq->next == se) 1149 cfs_rq->next = NULL; 1150 else 1151 break; 1152 } 1153 } 1154 1155 static void __clear_buddies_skip(struct sched_entity *se) 1156 { 1157 for_each_sched_entity(se) { 1158 struct cfs_rq *cfs_rq = cfs_rq_of(se); 1159 if (cfs_rq->skip == se) 1160 cfs_rq->skip = NULL; 1161 else 1162 break; 1163 } 1164 } 1165 1166 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se) 1167 { 1168 if (cfs_rq->last == se) 1169 __clear_buddies_last(se); 1170 1171 if (cfs_rq->next == se) 1172 __clear_buddies_next(se); 1173 1174 if (cfs_rq->skip == se) 1175 __clear_buddies_skip(se); 1176 } 1177 1178 static void return_cfs_rq_runtime(struct cfs_rq *cfs_rq); 1179 1180 static void 1181 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags) 1182 { 1183 /* 1184 * Update run-time statistics of the 'current'. 1185 */ 1186 update_curr(cfs_rq); 1187 1188 update_stats_dequeue(cfs_rq, se); 1189 if (flags & DEQUEUE_SLEEP) { 1190 #ifdef CONFIG_SCHEDSTATS 1191 if (entity_is_task(se)) { 1192 struct task_struct *tsk = task_of(se); 1193 1194 if (tsk->state & TASK_INTERRUPTIBLE) 1195 se->statistics.sleep_start = rq_of(cfs_rq)->clock; 1196 if (tsk->state & TASK_UNINTERRUPTIBLE) 1197 se->statistics.block_start = rq_of(cfs_rq)->clock; 1198 } 1199 #endif 1200 } 1201 1202 clear_buddies(cfs_rq, se); 1203 1204 if (se != cfs_rq->curr) 1205 __dequeue_entity(cfs_rq, se); 1206 se->on_rq = 0; 1207 update_cfs_load(cfs_rq, 0); 1208 account_entity_dequeue(cfs_rq, se); 1209 1210 /* 1211 * Normalize the entity after updating the min_vruntime because the 1212 * update can refer to the ->curr item and we need to reflect this 1213 * movement in our normalized position. 1214 */ 1215 if (!(flags & DEQUEUE_SLEEP)) 1216 se->vruntime -= cfs_rq->min_vruntime; 1217 1218 /* return excess runtime on last dequeue */ 1219 return_cfs_rq_runtime(cfs_rq); 1220 1221 update_min_vruntime(cfs_rq); 1222 update_cfs_shares(cfs_rq); 1223 } 1224 1225 /* 1226 * Preempt the current task with a newly woken task if needed: 1227 */ 1228 static void 1229 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr) 1230 { 1231 unsigned long ideal_runtime, delta_exec; 1232 struct sched_entity *se; 1233 s64 delta; 1234 1235 ideal_runtime = sched_slice(cfs_rq, curr); 1236 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime; 1237 if (delta_exec > ideal_runtime) { 1238 resched_task(rq_of(cfs_rq)->curr); 1239 /* 1240 * The current task ran long enough, ensure it doesn't get 1241 * re-elected due to buddy favours. 1242 */ 1243 clear_buddies(cfs_rq, curr); 1244 return; 1245 } 1246 1247 /* 1248 * Ensure that a task that missed wakeup preemption by a 1249 * narrow margin doesn't have to wait for a full slice. 1250 * This also mitigates buddy induced latencies under load. 1251 */ 1252 if (delta_exec < sysctl_sched_min_granularity) 1253 return; 1254 1255 se = __pick_first_entity(cfs_rq); 1256 delta = curr->vruntime - se->vruntime; 1257 1258 if (delta < 0) 1259 return; 1260 1261 if (delta > ideal_runtime) 1262 resched_task(rq_of(cfs_rq)->curr); 1263 } 1264 1265 static void 1266 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se) 1267 { 1268 /* 'current' is not kept within the tree. */ 1269 if (se->on_rq) { 1270 /* 1271 * Any task has to be enqueued before it get to execute on 1272 * a CPU. So account for the time it spent waiting on the 1273 * runqueue. 1274 */ 1275 update_stats_wait_end(cfs_rq, se); 1276 __dequeue_entity(cfs_rq, se); 1277 } 1278 1279 update_stats_curr_start(cfs_rq, se); 1280 cfs_rq->curr = se; 1281 #ifdef CONFIG_SCHEDSTATS 1282 /* 1283 * Track our maximum slice length, if the CPU's load is at 1284 * least twice that of our own weight (i.e. dont track it 1285 * when there are only lesser-weight tasks around): 1286 */ 1287 if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) { 1288 se->statistics.slice_max = max(se->statistics.slice_max, 1289 se->sum_exec_runtime - se->prev_sum_exec_runtime); 1290 } 1291 #endif 1292 se->prev_sum_exec_runtime = se->sum_exec_runtime; 1293 } 1294 1295 static int 1296 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se); 1297 1298 /* 1299 * Pick the next process, keeping these things in mind, in this order: 1300 * 1) keep things fair between processes/task groups 1301 * 2) pick the "next" process, since someone really wants that to run 1302 * 3) pick the "last" process, for cache locality 1303 * 4) do not run the "skip" process, if something else is available 1304 */ 1305 static struct sched_entity *pick_next_entity(struct cfs_rq *cfs_rq) 1306 { 1307 struct sched_entity *se = __pick_first_entity(cfs_rq); 1308 struct sched_entity *left = se; 1309 1310 /* 1311 * Avoid running the skip buddy, if running something else can 1312 * be done without getting too unfair. 1313 */ 1314 if (cfs_rq->skip == se) { 1315 struct sched_entity *second = __pick_next_entity(se); 1316 if (second && wakeup_preempt_entity(second, left) < 1) 1317 se = second; 1318 } 1319 1320 /* 1321 * Prefer last buddy, try to return the CPU to a preempted task. 1322 */ 1323 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1) 1324 se = cfs_rq->last; 1325 1326 /* 1327 * Someone really wants this to run. If it's not unfair, run it. 1328 */ 1329 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1) 1330 se = cfs_rq->next; 1331 1332 clear_buddies(cfs_rq, se); 1333 1334 return se; 1335 } 1336 1337 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq); 1338 1339 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev) 1340 { 1341 /* 1342 * If still on the runqueue then deactivate_task() 1343 * was not called and update_curr() has to be done: 1344 */ 1345 if (prev->on_rq) 1346 update_curr(cfs_rq); 1347 1348 /* throttle cfs_rqs exceeding runtime */ 1349 check_cfs_rq_runtime(cfs_rq); 1350 1351 check_spread(cfs_rq, prev); 1352 if (prev->on_rq) { 1353 update_stats_wait_start(cfs_rq, prev); 1354 /* Put 'current' back into the tree. */ 1355 __enqueue_entity(cfs_rq, prev); 1356 } 1357 cfs_rq->curr = NULL; 1358 } 1359 1360 static void 1361 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued) 1362 { 1363 /* 1364 * Update run-time statistics of the 'current'. 1365 */ 1366 update_curr(cfs_rq); 1367 1368 /* 1369 * Update share accounting for long-running entities. 1370 */ 1371 update_entity_shares_tick(cfs_rq); 1372 1373 #ifdef CONFIG_SCHED_HRTICK 1374 /* 1375 * queued ticks are scheduled to match the slice, so don't bother 1376 * validating it and just reschedule. 1377 */ 1378 if (queued) { 1379 resched_task(rq_of(cfs_rq)->curr); 1380 return; 1381 } 1382 /* 1383 * don't let the period tick interfere with the hrtick preemption 1384 */ 1385 if (!sched_feat(DOUBLE_TICK) && 1386 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer)) 1387 return; 1388 #endif 1389 1390 if (cfs_rq->nr_running > 1) 1391 check_preempt_tick(cfs_rq, curr); 1392 } 1393 1394 1395 /************************************************** 1396 * CFS bandwidth control machinery 1397 */ 1398 1399 #ifdef CONFIG_CFS_BANDWIDTH 1400 1401 #ifdef HAVE_JUMP_LABEL 1402 static struct jump_label_key __cfs_bandwidth_used; 1403 1404 static inline bool cfs_bandwidth_used(void) 1405 { 1406 return static_branch(&__cfs_bandwidth_used); 1407 } 1408 1409 void account_cfs_bandwidth_used(int enabled, int was_enabled) 1410 { 1411 /* only need to count groups transitioning between enabled/!enabled */ 1412 if (enabled && !was_enabled) 1413 jump_label_inc(&__cfs_bandwidth_used); 1414 else if (!enabled && was_enabled) 1415 jump_label_dec(&__cfs_bandwidth_used); 1416 } 1417 #else /* HAVE_JUMP_LABEL */ 1418 static bool cfs_bandwidth_used(void) 1419 { 1420 return true; 1421 } 1422 1423 void account_cfs_bandwidth_used(int enabled, int was_enabled) {} 1424 #endif /* HAVE_JUMP_LABEL */ 1425 1426 /* 1427 * default period for cfs group bandwidth. 1428 * default: 0.1s, units: nanoseconds 1429 */ 1430 static inline u64 default_cfs_period(void) 1431 { 1432 return 100000000ULL; 1433 } 1434 1435 static inline u64 sched_cfs_bandwidth_slice(void) 1436 { 1437 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC; 1438 } 1439 1440 /* 1441 * Replenish runtime according to assigned quota and update expiration time. 1442 * We use sched_clock_cpu directly instead of rq->clock to avoid adding 1443 * additional synchronization around rq->lock. 1444 * 1445 * requires cfs_b->lock 1446 */ 1447 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b) 1448 { 1449 u64 now; 1450 1451 if (cfs_b->quota == RUNTIME_INF) 1452 return; 1453 1454 now = sched_clock_cpu(smp_processor_id()); 1455 cfs_b->runtime = cfs_b->quota; 1456 cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period); 1457 } 1458 1459 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg) 1460 { 1461 return &tg->cfs_bandwidth; 1462 } 1463 1464 /* returns 0 on failure to allocate runtime */ 1465 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq) 1466 { 1467 struct task_group *tg = cfs_rq->tg; 1468 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg); 1469 u64 amount = 0, min_amount, expires; 1470 1471 /* note: this is a positive sum as runtime_remaining <= 0 */ 1472 min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining; 1473 1474 raw_spin_lock(&cfs_b->lock); 1475 if (cfs_b->quota == RUNTIME_INF) 1476 amount = min_amount; 1477 else { 1478 /* 1479 * If the bandwidth pool has become inactive, then at least one 1480 * period must have elapsed since the last consumption. 1481 * Refresh the global state and ensure bandwidth timer becomes 1482 * active. 1483 */ 1484 if (!cfs_b->timer_active) { 1485 __refill_cfs_bandwidth_runtime(cfs_b); 1486 __start_cfs_bandwidth(cfs_b); 1487 } 1488 1489 if (cfs_b->runtime > 0) { 1490 amount = min(cfs_b->runtime, min_amount); 1491 cfs_b->runtime -= amount; 1492 cfs_b->idle = 0; 1493 } 1494 } 1495 expires = cfs_b->runtime_expires; 1496 raw_spin_unlock(&cfs_b->lock); 1497 1498 cfs_rq->runtime_remaining += amount; 1499 /* 1500 * we may have advanced our local expiration to account for allowed 1501 * spread between our sched_clock and the one on which runtime was 1502 * issued. 1503 */ 1504 if ((s64)(expires - cfs_rq->runtime_expires) > 0) 1505 cfs_rq->runtime_expires = expires; 1506 1507 return cfs_rq->runtime_remaining > 0; 1508 } 1509 1510 /* 1511 * Note: This depends on the synchronization provided by sched_clock and the 1512 * fact that rq->clock snapshots this value. 1513 */ 1514 static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq) 1515 { 1516 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg); 1517 struct rq *rq = rq_of(cfs_rq); 1518 1519 /* if the deadline is ahead of our clock, nothing to do */ 1520 if (likely((s64)(rq->clock - cfs_rq->runtime_expires) < 0)) 1521 return; 1522 1523 if (cfs_rq->runtime_remaining < 0) 1524 return; 1525 1526 /* 1527 * If the local deadline has passed we have to consider the 1528 * possibility that our sched_clock is 'fast' and the global deadline 1529 * has not truly expired. 1530 * 1531 * Fortunately we can check determine whether this the case by checking 1532 * whether the global deadline has advanced. 1533 */ 1534 1535 if ((s64)(cfs_rq->runtime_expires - cfs_b->runtime_expires) >= 0) { 1536 /* extend local deadline, drift is bounded above by 2 ticks */ 1537 cfs_rq->runtime_expires += TICK_NSEC; 1538 } else { 1539 /* global deadline is ahead, expiration has passed */ 1540 cfs_rq->runtime_remaining = 0; 1541 } 1542 } 1543 1544 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, 1545 unsigned long delta_exec) 1546 { 1547 /* dock delta_exec before expiring quota (as it could span periods) */ 1548 cfs_rq->runtime_remaining -= delta_exec; 1549 expire_cfs_rq_runtime(cfs_rq); 1550 1551 if (likely(cfs_rq->runtime_remaining > 0)) 1552 return; 1553 1554 /* 1555 * if we're unable to extend our runtime we resched so that the active 1556 * hierarchy can be throttled 1557 */ 1558 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr)) 1559 resched_task(rq_of(cfs_rq)->curr); 1560 } 1561 1562 static __always_inline void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, 1563 unsigned long delta_exec) 1564 { 1565 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled) 1566 return; 1567 1568 __account_cfs_rq_runtime(cfs_rq, delta_exec); 1569 } 1570 1571 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq) 1572 { 1573 return cfs_bandwidth_used() && cfs_rq->throttled; 1574 } 1575 1576 /* check whether cfs_rq, or any parent, is throttled */ 1577 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq) 1578 { 1579 return cfs_bandwidth_used() && cfs_rq->throttle_count; 1580 } 1581 1582 /* 1583 * Ensure that neither of the group entities corresponding to src_cpu or 1584 * dest_cpu are members of a throttled hierarchy when performing group 1585 * load-balance operations. 1586 */ 1587 static inline int throttled_lb_pair(struct task_group *tg, 1588 int src_cpu, int dest_cpu) 1589 { 1590 struct cfs_rq *src_cfs_rq, *dest_cfs_rq; 1591 1592 src_cfs_rq = tg->cfs_rq[src_cpu]; 1593 dest_cfs_rq = tg->cfs_rq[dest_cpu]; 1594 1595 return throttled_hierarchy(src_cfs_rq) || 1596 throttled_hierarchy(dest_cfs_rq); 1597 } 1598 1599 /* updated child weight may affect parent so we have to do this bottom up */ 1600 static int tg_unthrottle_up(struct task_group *tg, void *data) 1601 { 1602 struct rq *rq = data; 1603 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)]; 1604 1605 cfs_rq->throttle_count--; 1606 #ifdef CONFIG_SMP 1607 if (!cfs_rq->throttle_count) { 1608 u64 delta = rq->clock_task - cfs_rq->load_stamp; 1609 1610 /* leaving throttled state, advance shares averaging windows */ 1611 cfs_rq->load_stamp += delta; 1612 cfs_rq->load_last += delta; 1613 1614 /* update entity weight now that we are on_rq again */ 1615 update_cfs_shares(cfs_rq); 1616 } 1617 #endif 1618 1619 return 0; 1620 } 1621 1622 static int tg_throttle_down(struct task_group *tg, void *data) 1623 { 1624 struct rq *rq = data; 1625 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)]; 1626 1627 /* group is entering throttled state, record last load */ 1628 if (!cfs_rq->throttle_count) 1629 update_cfs_load(cfs_rq, 0); 1630 cfs_rq->throttle_count++; 1631 1632 return 0; 1633 } 1634 1635 static void throttle_cfs_rq(struct cfs_rq *cfs_rq) 1636 { 1637 struct rq *rq = rq_of(cfs_rq); 1638 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg); 1639 struct sched_entity *se; 1640 long task_delta, dequeue = 1; 1641 1642 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))]; 1643 1644 /* account load preceding throttle */ 1645 rcu_read_lock(); 1646 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq); 1647 rcu_read_unlock(); 1648 1649 task_delta = cfs_rq->h_nr_running; 1650 for_each_sched_entity(se) { 1651 struct cfs_rq *qcfs_rq = cfs_rq_of(se); 1652 /* throttled entity or throttle-on-deactivate */ 1653 if (!se->on_rq) 1654 break; 1655 1656 if (dequeue) 1657 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP); 1658 qcfs_rq->h_nr_running -= task_delta; 1659 1660 if (qcfs_rq->load.weight) 1661 dequeue = 0; 1662 } 1663 1664 if (!se) 1665 rq->nr_running -= task_delta; 1666 1667 cfs_rq->throttled = 1; 1668 cfs_rq->throttled_timestamp = rq->clock; 1669 raw_spin_lock(&cfs_b->lock); 1670 list_add_tail_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq); 1671 raw_spin_unlock(&cfs_b->lock); 1672 } 1673 1674 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq) 1675 { 1676 struct rq *rq = rq_of(cfs_rq); 1677 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg); 1678 struct sched_entity *se; 1679 int enqueue = 1; 1680 long task_delta; 1681 1682 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))]; 1683 1684 cfs_rq->throttled = 0; 1685 raw_spin_lock(&cfs_b->lock); 1686 cfs_b->throttled_time += rq->clock - cfs_rq->throttled_timestamp; 1687 list_del_rcu(&cfs_rq->throttled_list); 1688 raw_spin_unlock(&cfs_b->lock); 1689 cfs_rq->throttled_timestamp = 0; 1690 1691 update_rq_clock(rq); 1692 /* update hierarchical throttle state */ 1693 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq); 1694 1695 if (!cfs_rq->load.weight) 1696 return; 1697 1698 task_delta = cfs_rq->h_nr_running; 1699 for_each_sched_entity(se) { 1700 if (se->on_rq) 1701 enqueue = 0; 1702 1703 cfs_rq = cfs_rq_of(se); 1704 if (enqueue) 1705 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP); 1706 cfs_rq->h_nr_running += task_delta; 1707 1708 if (cfs_rq_throttled(cfs_rq)) 1709 break; 1710 } 1711 1712 if (!se) 1713 rq->nr_running += task_delta; 1714 1715 /* determine whether we need to wake up potentially idle cpu */ 1716 if (rq->curr == rq->idle && rq->cfs.nr_running) 1717 resched_task(rq->curr); 1718 } 1719 1720 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b, 1721 u64 remaining, u64 expires) 1722 { 1723 struct cfs_rq *cfs_rq; 1724 u64 runtime = remaining; 1725 1726 rcu_read_lock(); 1727 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq, 1728 throttled_list) { 1729 struct rq *rq = rq_of(cfs_rq); 1730 1731 raw_spin_lock(&rq->lock); 1732 if (!cfs_rq_throttled(cfs_rq)) 1733 goto next; 1734 1735 runtime = -cfs_rq->runtime_remaining + 1; 1736 if (runtime > remaining) 1737 runtime = remaining; 1738 remaining -= runtime; 1739 1740 cfs_rq->runtime_remaining += runtime; 1741 cfs_rq->runtime_expires = expires; 1742 1743 /* we check whether we're throttled above */ 1744 if (cfs_rq->runtime_remaining > 0) 1745 unthrottle_cfs_rq(cfs_rq); 1746 1747 next: 1748 raw_spin_unlock(&rq->lock); 1749 1750 if (!remaining) 1751 break; 1752 } 1753 rcu_read_unlock(); 1754 1755 return remaining; 1756 } 1757 1758 /* 1759 * Responsible for refilling a task_group's bandwidth and unthrottling its 1760 * cfs_rqs as appropriate. If there has been no activity within the last 1761 * period the timer is deactivated until scheduling resumes; cfs_b->idle is 1762 * used to track this state. 1763 */ 1764 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun) 1765 { 1766 u64 runtime, runtime_expires; 1767 int idle = 1, throttled; 1768 1769 raw_spin_lock(&cfs_b->lock); 1770 /* no need to continue the timer with no bandwidth constraint */ 1771 if (cfs_b->quota == RUNTIME_INF) 1772 goto out_unlock; 1773 1774 throttled = !list_empty(&cfs_b->throttled_cfs_rq); 1775 /* idle depends on !throttled (for the case of a large deficit) */ 1776 idle = cfs_b->idle && !throttled; 1777 cfs_b->nr_periods += overrun; 1778 1779 /* if we're going inactive then everything else can be deferred */ 1780 if (idle) 1781 goto out_unlock; 1782 1783 __refill_cfs_bandwidth_runtime(cfs_b); 1784 1785 if (!throttled) { 1786 /* mark as potentially idle for the upcoming period */ 1787 cfs_b->idle = 1; 1788 goto out_unlock; 1789 } 1790 1791 /* account preceding periods in which throttling occurred */ 1792 cfs_b->nr_throttled += overrun; 1793 1794 /* 1795 * There are throttled entities so we must first use the new bandwidth 1796 * to unthrottle them before making it generally available. This 1797 * ensures that all existing debts will be paid before a new cfs_rq is 1798 * allowed to run. 1799 */ 1800 runtime = cfs_b->runtime; 1801 runtime_expires = cfs_b->runtime_expires; 1802 cfs_b->runtime = 0; 1803 1804 /* 1805 * This check is repeated as we are holding onto the new bandwidth 1806 * while we unthrottle. This can potentially race with an unthrottled 1807 * group trying to acquire new bandwidth from the global pool. 1808 */ 1809 while (throttled && runtime > 0) { 1810 raw_spin_unlock(&cfs_b->lock); 1811 /* we can't nest cfs_b->lock while distributing bandwidth */ 1812 runtime = distribute_cfs_runtime(cfs_b, runtime, 1813 runtime_expires); 1814 raw_spin_lock(&cfs_b->lock); 1815 1816 throttled = !list_empty(&cfs_b->throttled_cfs_rq); 1817 } 1818 1819 /* return (any) remaining runtime */ 1820 cfs_b->runtime = runtime; 1821 /* 1822 * While we are ensured activity in the period following an 1823 * unthrottle, this also covers the case in which the new bandwidth is 1824 * insufficient to cover the existing bandwidth deficit. (Forcing the 1825 * timer to remain active while there are any throttled entities.) 1826 */ 1827 cfs_b->idle = 0; 1828 out_unlock: 1829 if (idle) 1830 cfs_b->timer_active = 0; 1831 raw_spin_unlock(&cfs_b->lock); 1832 1833 return idle; 1834 } 1835 1836 /* a cfs_rq won't donate quota below this amount */ 1837 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC; 1838 /* minimum remaining period time to redistribute slack quota */ 1839 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC; 1840 /* how long we wait to gather additional slack before distributing */ 1841 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC; 1842 1843 /* are we near the end of the current quota period? */ 1844 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire) 1845 { 1846 struct hrtimer *refresh_timer = &cfs_b->period_timer; 1847 u64 remaining; 1848 1849 /* if the call-back is running a quota refresh is already occurring */ 1850 if (hrtimer_callback_running(refresh_timer)) 1851 return 1; 1852 1853 /* is a quota refresh about to occur? */ 1854 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer)); 1855 if (remaining < min_expire) 1856 return 1; 1857 1858 return 0; 1859 } 1860 1861 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b) 1862 { 1863 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration; 1864 1865 /* if there's a quota refresh soon don't bother with slack */ 1866 if (runtime_refresh_within(cfs_b, min_left)) 1867 return; 1868 1869 start_bandwidth_timer(&cfs_b->slack_timer, 1870 ns_to_ktime(cfs_bandwidth_slack_period)); 1871 } 1872 1873 /* we know any runtime found here is valid as update_curr() precedes return */ 1874 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq) 1875 { 1876 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg); 1877 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime; 1878 1879 if (slack_runtime <= 0) 1880 return; 1881 1882 raw_spin_lock(&cfs_b->lock); 1883 if (cfs_b->quota != RUNTIME_INF && 1884 cfs_rq->runtime_expires == cfs_b->runtime_expires) { 1885 cfs_b->runtime += slack_runtime; 1886 1887 /* we are under rq->lock, defer unthrottling using a timer */ 1888 if (cfs_b->runtime > sched_cfs_bandwidth_slice() && 1889 !list_empty(&cfs_b->throttled_cfs_rq)) 1890 start_cfs_slack_bandwidth(cfs_b); 1891 } 1892 raw_spin_unlock(&cfs_b->lock); 1893 1894 /* even if it's not valid for return we don't want to try again */ 1895 cfs_rq->runtime_remaining -= slack_runtime; 1896 } 1897 1898 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) 1899 { 1900 if (!cfs_bandwidth_used()) 1901 return; 1902 1903 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running) 1904 return; 1905 1906 __return_cfs_rq_runtime(cfs_rq); 1907 } 1908 1909 /* 1910 * This is done with a timer (instead of inline with bandwidth return) since 1911 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs. 1912 */ 1913 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b) 1914 { 1915 u64 runtime = 0, slice = sched_cfs_bandwidth_slice(); 1916 u64 expires; 1917 1918 /* confirm we're still not at a refresh boundary */ 1919 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) 1920 return; 1921 1922 raw_spin_lock(&cfs_b->lock); 1923 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice) { 1924 runtime = cfs_b->runtime; 1925 cfs_b->runtime = 0; 1926 } 1927 expires = cfs_b->runtime_expires; 1928 raw_spin_unlock(&cfs_b->lock); 1929 1930 if (!runtime) 1931 return; 1932 1933 runtime = distribute_cfs_runtime(cfs_b, runtime, expires); 1934 1935 raw_spin_lock(&cfs_b->lock); 1936 if (expires == cfs_b->runtime_expires) 1937 cfs_b->runtime = runtime; 1938 raw_spin_unlock(&cfs_b->lock); 1939 } 1940 1941 /* 1942 * When a group wakes up we want to make sure that its quota is not already 1943 * expired/exceeded, otherwise it may be allowed to steal additional ticks of 1944 * runtime as update_curr() throttling can not not trigger until it's on-rq. 1945 */ 1946 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) 1947 { 1948 if (!cfs_bandwidth_used()) 1949 return; 1950 1951 /* an active group must be handled by the update_curr()->put() path */ 1952 if (!cfs_rq->runtime_enabled || cfs_rq->curr) 1953 return; 1954 1955 /* ensure the group is not already throttled */ 1956 if (cfs_rq_throttled(cfs_rq)) 1957 return; 1958 1959 /* update runtime allocation */ 1960 account_cfs_rq_runtime(cfs_rq, 0); 1961 if (cfs_rq->runtime_remaining <= 0) 1962 throttle_cfs_rq(cfs_rq); 1963 } 1964 1965 /* conditionally throttle active cfs_rq's from put_prev_entity() */ 1966 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq) 1967 { 1968 if (!cfs_bandwidth_used()) 1969 return; 1970 1971 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0)) 1972 return; 1973 1974 /* 1975 * it's possible for a throttled entity to be forced into a running 1976 * state (e.g. set_curr_task), in this case we're finished. 1977 */ 1978 if (cfs_rq_throttled(cfs_rq)) 1979 return; 1980 1981 throttle_cfs_rq(cfs_rq); 1982 } 1983 1984 static inline u64 default_cfs_period(void); 1985 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun); 1986 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b); 1987 1988 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer) 1989 { 1990 struct cfs_bandwidth *cfs_b = 1991 container_of(timer, struct cfs_bandwidth, slack_timer); 1992 do_sched_cfs_slack_timer(cfs_b); 1993 1994 return HRTIMER_NORESTART; 1995 } 1996 1997 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer) 1998 { 1999 struct cfs_bandwidth *cfs_b = 2000 container_of(timer, struct cfs_bandwidth, period_timer); 2001 ktime_t now; 2002 int overrun; 2003 int idle = 0; 2004 2005 for (;;) { 2006 now = hrtimer_cb_get_time(timer); 2007 overrun = hrtimer_forward(timer, now, cfs_b->period); 2008 2009 if (!overrun) 2010 break; 2011 2012 idle = do_sched_cfs_period_timer(cfs_b, overrun); 2013 } 2014 2015 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART; 2016 } 2017 2018 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) 2019 { 2020 raw_spin_lock_init(&cfs_b->lock); 2021 cfs_b->runtime = 0; 2022 cfs_b->quota = RUNTIME_INF; 2023 cfs_b->period = ns_to_ktime(default_cfs_period()); 2024 2025 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq); 2026 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); 2027 cfs_b->period_timer.function = sched_cfs_period_timer; 2028 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); 2029 cfs_b->slack_timer.function = sched_cfs_slack_timer; 2030 } 2031 2032 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) 2033 { 2034 cfs_rq->runtime_enabled = 0; 2035 INIT_LIST_HEAD(&cfs_rq->throttled_list); 2036 } 2037 2038 /* requires cfs_b->lock, may release to reprogram timer */ 2039 void __start_cfs_bandwidth(struct cfs_bandwidth *cfs_b) 2040 { 2041 /* 2042 * The timer may be active because we're trying to set a new bandwidth 2043 * period or because we're racing with the tear-down path 2044 * (timer_active==0 becomes visible before the hrtimer call-back 2045 * terminates). In either case we ensure that it's re-programmed 2046 */ 2047 while (unlikely(hrtimer_active(&cfs_b->period_timer))) { 2048 raw_spin_unlock(&cfs_b->lock); 2049 /* ensure cfs_b->lock is available while we wait */ 2050 hrtimer_cancel(&cfs_b->period_timer); 2051 2052 raw_spin_lock(&cfs_b->lock); 2053 /* if someone else restarted the timer then we're done */ 2054 if (cfs_b->timer_active) 2055 return; 2056 } 2057 2058 cfs_b->timer_active = 1; 2059 start_bandwidth_timer(&cfs_b->period_timer, cfs_b->period); 2060 } 2061 2062 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) 2063 { 2064 hrtimer_cancel(&cfs_b->period_timer); 2065 hrtimer_cancel(&cfs_b->slack_timer); 2066 } 2067 2068 void unthrottle_offline_cfs_rqs(struct rq *rq) 2069 { 2070 struct cfs_rq *cfs_rq; 2071 2072 for_each_leaf_cfs_rq(rq, cfs_rq) { 2073 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg); 2074 2075 if (!cfs_rq->runtime_enabled) 2076 continue; 2077 2078 /* 2079 * clock_task is not advancing so we just need to make sure 2080 * there's some valid quota amount 2081 */ 2082 cfs_rq->runtime_remaining = cfs_b->quota; 2083 if (cfs_rq_throttled(cfs_rq)) 2084 unthrottle_cfs_rq(cfs_rq); 2085 } 2086 } 2087 2088 #else /* CONFIG_CFS_BANDWIDTH */ 2089 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, 2090 unsigned long delta_exec) {} 2091 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq) {} 2092 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {} 2093 static void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {} 2094 2095 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq) 2096 { 2097 return 0; 2098 } 2099 2100 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq) 2101 { 2102 return 0; 2103 } 2104 2105 static inline int throttled_lb_pair(struct task_group *tg, 2106 int src_cpu, int dest_cpu) 2107 { 2108 return 0; 2109 } 2110 2111 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {} 2112 2113 #ifdef CONFIG_FAIR_GROUP_SCHED 2114 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {} 2115 #endif 2116 2117 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg) 2118 { 2119 return NULL; 2120 } 2121 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {} 2122 void unthrottle_offline_cfs_rqs(struct rq *rq) {} 2123 2124 #endif /* CONFIG_CFS_BANDWIDTH */ 2125 2126 /************************************************** 2127 * CFS operations on tasks: 2128 */ 2129 2130 #ifdef CONFIG_SCHED_HRTICK 2131 static void hrtick_start_fair(struct rq *rq, struct task_struct *p) 2132 { 2133 struct sched_entity *se = &p->se; 2134 struct cfs_rq *cfs_rq = cfs_rq_of(se); 2135 2136 WARN_ON(task_rq(p) != rq); 2137 2138 if (cfs_rq->nr_running > 1) { 2139 u64 slice = sched_slice(cfs_rq, se); 2140 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime; 2141 s64 delta = slice - ran; 2142 2143 if (delta < 0) { 2144 if (rq->curr == p) 2145 resched_task(p); 2146 return; 2147 } 2148 2149 /* 2150 * Don't schedule slices shorter than 10000ns, that just 2151 * doesn't make sense. Rely on vruntime for fairness. 2152 */ 2153 if (rq->curr != p) 2154 delta = max_t(s64, 10000LL, delta); 2155 2156 hrtick_start(rq, delta); 2157 } 2158 } 2159 2160 /* 2161 * called from enqueue/dequeue and updates the hrtick when the 2162 * current task is from our class and nr_running is low enough 2163 * to matter. 2164 */ 2165 static void hrtick_update(struct rq *rq) 2166 { 2167 struct task_struct *curr = rq->curr; 2168 2169 if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class) 2170 return; 2171 2172 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency) 2173 hrtick_start_fair(rq, curr); 2174 } 2175 #else /* !CONFIG_SCHED_HRTICK */ 2176 static inline void 2177 hrtick_start_fair(struct rq *rq, struct task_struct *p) 2178 { 2179 } 2180 2181 static inline void hrtick_update(struct rq *rq) 2182 { 2183 } 2184 #endif 2185 2186 /* 2187 * The enqueue_task method is called before nr_running is 2188 * increased. Here we update the fair scheduling stats and 2189 * then put the task into the rbtree: 2190 */ 2191 static void 2192 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags) 2193 { 2194 struct cfs_rq *cfs_rq; 2195 struct sched_entity *se = &p->se; 2196 2197 for_each_sched_entity(se) { 2198 if (se->on_rq) 2199 break; 2200 cfs_rq = cfs_rq_of(se); 2201 enqueue_entity(cfs_rq, se, flags); 2202 2203 /* 2204 * end evaluation on encountering a throttled cfs_rq 2205 * 2206 * note: in the case of encountering a throttled cfs_rq we will 2207 * post the final h_nr_running increment below. 2208 */ 2209 if (cfs_rq_throttled(cfs_rq)) 2210 break; 2211 cfs_rq->h_nr_running++; 2212 2213 flags = ENQUEUE_WAKEUP; 2214 } 2215 2216 for_each_sched_entity(se) { 2217 cfs_rq = cfs_rq_of(se); 2218 cfs_rq->h_nr_running++; 2219 2220 if (cfs_rq_throttled(cfs_rq)) 2221 break; 2222 2223 update_cfs_load(cfs_rq, 0); 2224 update_cfs_shares(cfs_rq); 2225 } 2226 2227 if (!se) 2228 inc_nr_running(rq); 2229 hrtick_update(rq); 2230 } 2231 2232 static void set_next_buddy(struct sched_entity *se); 2233 2234 /* 2235 * The dequeue_task method is called before nr_running is 2236 * decreased. We remove the task from the rbtree and 2237 * update the fair scheduling stats: 2238 */ 2239 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags) 2240 { 2241 struct cfs_rq *cfs_rq; 2242 struct sched_entity *se = &p->se; 2243 int task_sleep = flags & DEQUEUE_SLEEP; 2244 2245 for_each_sched_entity(se) { 2246 cfs_rq = cfs_rq_of(se); 2247 dequeue_entity(cfs_rq, se, flags); 2248 2249 /* 2250 * end evaluation on encountering a throttled cfs_rq 2251 * 2252 * note: in the case of encountering a throttled cfs_rq we will 2253 * post the final h_nr_running decrement below. 2254 */ 2255 if (cfs_rq_throttled(cfs_rq)) 2256 break; 2257 cfs_rq->h_nr_running--; 2258 2259 /* Don't dequeue parent if it has other entities besides us */ 2260 if (cfs_rq->load.weight) { 2261 /* 2262 * Bias pick_next to pick a task from this cfs_rq, as 2263 * p is sleeping when it is within its sched_slice. 2264 */ 2265 if (task_sleep && parent_entity(se)) 2266 set_next_buddy(parent_entity(se)); 2267 2268 /* avoid re-evaluating load for this entity */ 2269 se = parent_entity(se); 2270 break; 2271 } 2272 flags |= DEQUEUE_SLEEP; 2273 } 2274 2275 for_each_sched_entity(se) { 2276 cfs_rq = cfs_rq_of(se); 2277 cfs_rq->h_nr_running--; 2278 2279 if (cfs_rq_throttled(cfs_rq)) 2280 break; 2281 2282 update_cfs_load(cfs_rq, 0); 2283 update_cfs_shares(cfs_rq); 2284 } 2285 2286 if (!se) 2287 dec_nr_running(rq); 2288 hrtick_update(rq); 2289 } 2290 2291 #ifdef CONFIG_SMP 2292 /* Used instead of source_load when we know the type == 0 */ 2293 static unsigned long weighted_cpuload(const int cpu) 2294 { 2295 return cpu_rq(cpu)->load.weight; 2296 } 2297 2298 /* 2299 * Return a low guess at the load of a migration-source cpu weighted 2300 * according to the scheduling class and "nice" value. 2301 * 2302 * We want to under-estimate the load of migration sources, to 2303 * balance conservatively. 2304 */ 2305 static unsigned long source_load(int cpu, int type) 2306 { 2307 struct rq *rq = cpu_rq(cpu); 2308 unsigned long total = weighted_cpuload(cpu); 2309 2310 if (type == 0 || !sched_feat(LB_BIAS)) 2311 return total; 2312 2313 return min(rq->cpu_load[type-1], total); 2314 } 2315 2316 /* 2317 * Return a high guess at the load of a migration-target cpu weighted 2318 * according to the scheduling class and "nice" value. 2319 */ 2320 static unsigned long target_load(int cpu, int type) 2321 { 2322 struct rq *rq = cpu_rq(cpu); 2323 unsigned long total = weighted_cpuload(cpu); 2324 2325 if (type == 0 || !sched_feat(LB_BIAS)) 2326 return total; 2327 2328 return max(rq->cpu_load[type-1], total); 2329 } 2330 2331 static unsigned long power_of(int cpu) 2332 { 2333 return cpu_rq(cpu)->cpu_power; 2334 } 2335 2336 static unsigned long cpu_avg_load_per_task(int cpu) 2337 { 2338 struct rq *rq = cpu_rq(cpu); 2339 unsigned long nr_running = ACCESS_ONCE(rq->nr_running); 2340 2341 if (nr_running) 2342 return rq->load.weight / nr_running; 2343 2344 return 0; 2345 } 2346 2347 2348 static void task_waking_fair(struct task_struct *p) 2349 { 2350 struct sched_entity *se = &p->se; 2351 struct cfs_rq *cfs_rq = cfs_rq_of(se); 2352 u64 min_vruntime; 2353 2354 #ifndef CONFIG_64BIT 2355 u64 min_vruntime_copy; 2356 2357 do { 2358 min_vruntime_copy = cfs_rq->min_vruntime_copy; 2359 smp_rmb(); 2360 min_vruntime = cfs_rq->min_vruntime; 2361 } while (min_vruntime != min_vruntime_copy); 2362 #else 2363 min_vruntime = cfs_rq->min_vruntime; 2364 #endif 2365 2366 se->vruntime -= min_vruntime; 2367 } 2368 2369 #ifdef CONFIG_FAIR_GROUP_SCHED 2370 /* 2371 * effective_load() calculates the load change as seen from the root_task_group 2372 * 2373 * Adding load to a group doesn't make a group heavier, but can cause movement 2374 * of group shares between cpus. Assuming the shares were perfectly aligned one 2375 * can calculate the shift in shares. 2376 * 2377 * Calculate the effective load difference if @wl is added (subtracted) to @tg 2378 * on this @cpu and results in a total addition (subtraction) of @wg to the 2379 * total group weight. 2380 * 2381 * Given a runqueue weight distribution (rw_i) we can compute a shares 2382 * distribution (s_i) using: 2383 * 2384 * s_i = rw_i / \Sum rw_j (1) 2385 * 2386 * Suppose we have 4 CPUs and our @tg is a direct child of the root group and 2387 * has 7 equal weight tasks, distributed as below (rw_i), with the resulting 2388 * shares distribution (s_i): 2389 * 2390 * rw_i = { 2, 4, 1, 0 } 2391 * s_i = { 2/7, 4/7, 1/7, 0 } 2392 * 2393 * As per wake_affine() we're interested in the load of two CPUs (the CPU the 2394 * task used to run on and the CPU the waker is running on), we need to 2395 * compute the effect of waking a task on either CPU and, in case of a sync 2396 * wakeup, compute the effect of the current task going to sleep. 2397 * 2398 * So for a change of @wl to the local @cpu with an overall group weight change 2399 * of @wl we can compute the new shares distribution (s'_i) using: 2400 * 2401 * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2) 2402 * 2403 * Suppose we're interested in CPUs 0 and 1, and want to compute the load 2404 * differences in waking a task to CPU 0. The additional task changes the 2405 * weight and shares distributions like: 2406 * 2407 * rw'_i = { 3, 4, 1, 0 } 2408 * s'_i = { 3/8, 4/8, 1/8, 0 } 2409 * 2410 * We can then compute the difference in effective weight by using: 2411 * 2412 * dw_i = S * (s'_i - s_i) (3) 2413 * 2414 * Where 'S' is the group weight as seen by its parent. 2415 * 2416 * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7) 2417 * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 - 2418 * 4/7) times the weight of the group. 2419 */ 2420 static long effective_load(struct task_group *tg, int cpu, long wl, long wg) 2421 { 2422 struct sched_entity *se = tg->se[cpu]; 2423 2424 if (!tg->parent) /* the trivial, non-cgroup case */ 2425 return wl; 2426 2427 for_each_sched_entity(se) { 2428 long w, W; 2429 2430 tg = se->my_q->tg; 2431 2432 /* 2433 * W = @wg + \Sum rw_j 2434 */ 2435 W = wg + calc_tg_weight(tg, se->my_q); 2436 2437 /* 2438 * w = rw_i + @wl 2439 */ 2440 w = se->my_q->load.weight + wl; 2441 2442 /* 2443 * wl = S * s'_i; see (2) 2444 */ 2445 if (W > 0 && w < W) 2446 wl = (w * tg->shares) / W; 2447 else 2448 wl = tg->shares; 2449 2450 /* 2451 * Per the above, wl is the new se->load.weight value; since 2452 * those are clipped to [MIN_SHARES, ...) do so now. See 2453 * calc_cfs_shares(). 2454 */ 2455 if (wl < MIN_SHARES) 2456 wl = MIN_SHARES; 2457 2458 /* 2459 * wl = dw_i = S * (s'_i - s_i); see (3) 2460 */ 2461 wl -= se->load.weight; 2462 2463 /* 2464 * Recursively apply this logic to all parent groups to compute 2465 * the final effective load change on the root group. Since 2466 * only the @tg group gets extra weight, all parent groups can 2467 * only redistribute existing shares. @wl is the shift in shares 2468 * resulting from this level per the above. 2469 */ 2470 wg = 0; 2471 } 2472 2473 return wl; 2474 } 2475 #else 2476 2477 static inline unsigned long effective_load(struct task_group *tg, int cpu, 2478 unsigned long wl, unsigned long wg) 2479 { 2480 return wl; 2481 } 2482 2483 #endif 2484 2485 static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync) 2486 { 2487 s64 this_load, load; 2488 int idx, this_cpu, prev_cpu; 2489 unsigned long tl_per_task; 2490 struct task_group *tg; 2491 unsigned long weight; 2492 int balanced; 2493 2494 idx = sd->wake_idx; 2495 this_cpu = smp_processor_id(); 2496 prev_cpu = task_cpu(p); 2497 load = source_load(prev_cpu, idx); 2498 this_load = target_load(this_cpu, idx); 2499 2500 /* 2501 * If sync wakeup then subtract the (maximum possible) 2502 * effect of the currently running task from the load 2503 * of the current CPU: 2504 */ 2505 if (sync) { 2506 tg = task_group(current); 2507 weight = current->se.load.weight; 2508 2509 this_load += effective_load(tg, this_cpu, -weight, -weight); 2510 load += effective_load(tg, prev_cpu, 0, -weight); 2511 } 2512 2513 tg = task_group(p); 2514 weight = p->se.load.weight; 2515 2516 /* 2517 * In low-load situations, where prev_cpu is idle and this_cpu is idle 2518 * due to the sync cause above having dropped this_load to 0, we'll 2519 * always have an imbalance, but there's really nothing you can do 2520 * about that, so that's good too. 2521 * 2522 * Otherwise check if either cpus are near enough in load to allow this 2523 * task to be woken on this_cpu. 2524 */ 2525 if (this_load > 0) { 2526 s64 this_eff_load, prev_eff_load; 2527 2528 this_eff_load = 100; 2529 this_eff_load *= power_of(prev_cpu); 2530 this_eff_load *= this_load + 2531 effective_load(tg, this_cpu, weight, weight); 2532 2533 prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2; 2534 prev_eff_load *= power_of(this_cpu); 2535 prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight); 2536 2537 balanced = this_eff_load <= prev_eff_load; 2538 } else 2539 balanced = true; 2540 2541 /* 2542 * If the currently running task will sleep within 2543 * a reasonable amount of time then attract this newly 2544 * woken task: 2545 */ 2546 if (sync && balanced) 2547 return 1; 2548 2549 schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts); 2550 tl_per_task = cpu_avg_load_per_task(this_cpu); 2551 2552 if (balanced || 2553 (this_load <= load && 2554 this_load + target_load(prev_cpu, idx) <= tl_per_task)) { 2555 /* 2556 * This domain has SD_WAKE_AFFINE and 2557 * p is cache cold in this domain, and 2558 * there is no bad imbalance. 2559 */ 2560 schedstat_inc(sd, ttwu_move_affine); 2561 schedstat_inc(p, se.statistics.nr_wakeups_affine); 2562 2563 return 1; 2564 } 2565 return 0; 2566 } 2567 2568 /* 2569 * find_idlest_group finds and returns the least busy CPU group within the 2570 * domain. 2571 */ 2572 static struct sched_group * 2573 find_idlest_group(struct sched_domain *sd, struct task_struct *p, 2574 int this_cpu, int load_idx) 2575 { 2576 struct sched_group *idlest = NULL, *group = sd->groups; 2577 unsigned long min_load = ULONG_MAX, this_load = 0; 2578 int imbalance = 100 + (sd->imbalance_pct-100)/2; 2579 2580 do { 2581 unsigned long load, avg_load; 2582 int local_group; 2583 int i; 2584 2585 /* Skip over this group if it has no CPUs allowed */ 2586 if (!cpumask_intersects(sched_group_cpus(group), 2587 tsk_cpus_allowed(p))) 2588 continue; 2589 2590 local_group = cpumask_test_cpu(this_cpu, 2591 sched_group_cpus(group)); 2592 2593 /* Tally up the load of all CPUs in the group */ 2594 avg_load = 0; 2595 2596 for_each_cpu(i, sched_group_cpus(group)) { 2597 /* Bias balancing toward cpus of our domain */ 2598 if (local_group) 2599 load = source_load(i, load_idx); 2600 else 2601 load = target_load(i, load_idx); 2602 2603 avg_load += load; 2604 } 2605 2606 /* Adjust by relative CPU power of the group */ 2607 avg_load = (avg_load * SCHED_POWER_SCALE) / group->sgp->power; 2608 2609 if (local_group) { 2610 this_load = avg_load; 2611 } else if (avg_load < min_load) { 2612 min_load = avg_load; 2613 idlest = group; 2614 } 2615 } while (group = group->next, group != sd->groups); 2616 2617 if (!idlest || 100*this_load < imbalance*min_load) 2618 return NULL; 2619 return idlest; 2620 } 2621 2622 /* 2623 * find_idlest_cpu - find the idlest cpu among the cpus in group. 2624 */ 2625 static int 2626 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu) 2627 { 2628 unsigned long load, min_load = ULONG_MAX; 2629 int idlest = -1; 2630 int i; 2631 2632 /* Traverse only the allowed CPUs */ 2633 for_each_cpu_and(i, sched_group_cpus(group), tsk_cpus_allowed(p)) { 2634 load = weighted_cpuload(i); 2635 2636 if (load < min_load || (load == min_load && i == this_cpu)) { 2637 min_load = load; 2638 idlest = i; 2639 } 2640 } 2641 2642 return idlest; 2643 } 2644 2645 /* 2646 * Try and locate an idle CPU in the sched_domain. 2647 */ 2648 static int select_idle_sibling(struct task_struct *p, int target) 2649 { 2650 int cpu = smp_processor_id(); 2651 int prev_cpu = task_cpu(p); 2652 struct sched_domain *sd; 2653 struct sched_group *sg; 2654 int i; 2655 2656 /* 2657 * If the task is going to be woken-up on this cpu and if it is 2658 * already idle, then it is the right target. 2659 */ 2660 if (target == cpu && idle_cpu(cpu)) 2661 return cpu; 2662 2663 /* 2664 * If the task is going to be woken-up on the cpu where it previously 2665 * ran and if it is currently idle, then it the right target. 2666 */ 2667 if (target == prev_cpu && idle_cpu(prev_cpu)) 2668 return prev_cpu; 2669 2670 /* 2671 * Otherwise, iterate the domains and find an elegible idle cpu. 2672 */ 2673 rcu_read_lock(); 2674 2675 sd = rcu_dereference(per_cpu(sd_llc, target)); 2676 for_each_lower_domain(sd) { 2677 sg = sd->groups; 2678 do { 2679 if (!cpumask_intersects(sched_group_cpus(sg), 2680 tsk_cpus_allowed(p))) 2681 goto next; 2682 2683 for_each_cpu(i, sched_group_cpus(sg)) { 2684 if (!idle_cpu(i)) 2685 goto next; 2686 } 2687 2688 target = cpumask_first_and(sched_group_cpus(sg), 2689 tsk_cpus_allowed(p)); 2690 goto done; 2691 next: 2692 sg = sg->next; 2693 } while (sg != sd->groups); 2694 } 2695 done: 2696 rcu_read_unlock(); 2697 2698 return target; 2699 } 2700 2701 /* 2702 * sched_balance_self: balance the current task (running on cpu) in domains 2703 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and 2704 * SD_BALANCE_EXEC. 2705 * 2706 * Balance, ie. select the least loaded group. 2707 * 2708 * Returns the target CPU number, or the same CPU if no balancing is needed. 2709 * 2710 * preempt must be disabled. 2711 */ 2712 static int 2713 select_task_rq_fair(struct task_struct *p, int sd_flag, int wake_flags) 2714 { 2715 struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL; 2716 int cpu = smp_processor_id(); 2717 int prev_cpu = task_cpu(p); 2718 int new_cpu = cpu; 2719 int want_affine = 0; 2720 int want_sd = 1; 2721 int sync = wake_flags & WF_SYNC; 2722 2723 if (p->rt.nr_cpus_allowed == 1) 2724 return prev_cpu; 2725 2726 if (sd_flag & SD_BALANCE_WAKE) { 2727 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) 2728 want_affine = 1; 2729 new_cpu = prev_cpu; 2730 } 2731 2732 rcu_read_lock(); 2733 for_each_domain(cpu, tmp) { 2734 if (!(tmp->flags & SD_LOAD_BALANCE)) 2735 continue; 2736 2737 /* 2738 * If power savings logic is enabled for a domain, see if we 2739 * are not overloaded, if so, don't balance wider. 2740 */ 2741 if (tmp->flags & (SD_POWERSAVINGS_BALANCE|SD_PREFER_LOCAL)) { 2742 unsigned long power = 0; 2743 unsigned long nr_running = 0; 2744 unsigned long capacity; 2745 int i; 2746 2747 for_each_cpu(i, sched_domain_span(tmp)) { 2748 power += power_of(i); 2749 nr_running += cpu_rq(i)->cfs.nr_running; 2750 } 2751 2752 capacity = DIV_ROUND_CLOSEST(power, SCHED_POWER_SCALE); 2753 2754 if (tmp->flags & SD_POWERSAVINGS_BALANCE) 2755 nr_running /= 2; 2756 2757 if (nr_running < capacity) 2758 want_sd = 0; 2759 } 2760 2761 /* 2762 * If both cpu and prev_cpu are part of this domain, 2763 * cpu is a valid SD_WAKE_AFFINE target. 2764 */ 2765 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) && 2766 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) { 2767 affine_sd = tmp; 2768 want_affine = 0; 2769 } 2770 2771 if (!want_sd && !want_affine) 2772 break; 2773 2774 if (!(tmp->flags & sd_flag)) 2775 continue; 2776 2777 if (want_sd) 2778 sd = tmp; 2779 } 2780 2781 if (affine_sd) { 2782 if (cpu == prev_cpu || wake_affine(affine_sd, p, sync)) 2783 prev_cpu = cpu; 2784 2785 new_cpu = select_idle_sibling(p, prev_cpu); 2786 goto unlock; 2787 } 2788 2789 while (sd) { 2790 int load_idx = sd->forkexec_idx; 2791 struct sched_group *group; 2792 int weight; 2793 2794 if (!(sd->flags & sd_flag)) { 2795 sd = sd->child; 2796 continue; 2797 } 2798 2799 if (sd_flag & SD_BALANCE_WAKE) 2800 load_idx = sd->wake_idx; 2801 2802 group = find_idlest_group(sd, p, cpu, load_idx); 2803 if (!group) { 2804 sd = sd->child; 2805 continue; 2806 } 2807 2808 new_cpu = find_idlest_cpu(group, p, cpu); 2809 if (new_cpu == -1 || new_cpu == cpu) { 2810 /* Now try balancing at a lower domain level of cpu */ 2811 sd = sd->child; 2812 continue; 2813 } 2814 2815 /* Now try balancing at a lower domain level of new_cpu */ 2816 cpu = new_cpu; 2817 weight = sd->span_weight; 2818 sd = NULL; 2819 for_each_domain(cpu, tmp) { 2820 if (weight <= tmp->span_weight) 2821 break; 2822 if (tmp->flags & sd_flag) 2823 sd = tmp; 2824 } 2825 /* while loop will break here if sd == NULL */ 2826 } 2827 unlock: 2828 rcu_read_unlock(); 2829 2830 return new_cpu; 2831 } 2832 #endif /* CONFIG_SMP */ 2833 2834 static unsigned long 2835 wakeup_gran(struct sched_entity *curr, struct sched_entity *se) 2836 { 2837 unsigned long gran = sysctl_sched_wakeup_granularity; 2838 2839 /* 2840 * Since its curr running now, convert the gran from real-time 2841 * to virtual-time in his units. 2842 * 2843 * By using 'se' instead of 'curr' we penalize light tasks, so 2844 * they get preempted easier. That is, if 'se' < 'curr' then 2845 * the resulting gran will be larger, therefore penalizing the 2846 * lighter, if otoh 'se' > 'curr' then the resulting gran will 2847 * be smaller, again penalizing the lighter task. 2848 * 2849 * This is especially important for buddies when the leftmost 2850 * task is higher priority than the buddy. 2851 */ 2852 return calc_delta_fair(gran, se); 2853 } 2854 2855 /* 2856 * Should 'se' preempt 'curr'. 2857 * 2858 * |s1 2859 * |s2 2860 * |s3 2861 * g 2862 * |<--->|c 2863 * 2864 * w(c, s1) = -1 2865 * w(c, s2) = 0 2866 * w(c, s3) = 1 2867 * 2868 */ 2869 static int 2870 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se) 2871 { 2872 s64 gran, vdiff = curr->vruntime - se->vruntime; 2873 2874 if (vdiff <= 0) 2875 return -1; 2876 2877 gran = wakeup_gran(curr, se); 2878 if (vdiff > gran) 2879 return 1; 2880 2881 return 0; 2882 } 2883 2884 static void set_last_buddy(struct sched_entity *se) 2885 { 2886 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE)) 2887 return; 2888 2889 for_each_sched_entity(se) 2890 cfs_rq_of(se)->last = se; 2891 } 2892 2893 static void set_next_buddy(struct sched_entity *se) 2894 { 2895 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE)) 2896 return; 2897 2898 for_each_sched_entity(se) 2899 cfs_rq_of(se)->next = se; 2900 } 2901 2902 static void set_skip_buddy(struct sched_entity *se) 2903 { 2904 for_each_sched_entity(se) 2905 cfs_rq_of(se)->skip = se; 2906 } 2907 2908 /* 2909 * Preempt the current task with a newly woken task if needed: 2910 */ 2911 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags) 2912 { 2913 struct task_struct *curr = rq->curr; 2914 struct sched_entity *se = &curr->se, *pse = &p->se; 2915 struct cfs_rq *cfs_rq = task_cfs_rq(curr); 2916 int scale = cfs_rq->nr_running >= sched_nr_latency; 2917 int next_buddy_marked = 0; 2918 2919 if (unlikely(se == pse)) 2920 return; 2921 2922 /* 2923 * This is possible from callers such as pull_task(), in which we 2924 * unconditionally check_prempt_curr() after an enqueue (which may have 2925 * lead to a throttle). This both saves work and prevents false 2926 * next-buddy nomination below. 2927 */ 2928 if (unlikely(throttled_hierarchy(cfs_rq_of(pse)))) 2929 return; 2930 2931 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) { 2932 set_next_buddy(pse); 2933 next_buddy_marked = 1; 2934 } 2935 2936 /* 2937 * We can come here with TIF_NEED_RESCHED already set from new task 2938 * wake up path. 2939 * 2940 * Note: this also catches the edge-case of curr being in a throttled 2941 * group (e.g. via set_curr_task), since update_curr() (in the 2942 * enqueue of curr) will have resulted in resched being set. This 2943 * prevents us from potentially nominating it as a false LAST_BUDDY 2944 * below. 2945 */ 2946 if (test_tsk_need_resched(curr)) 2947 return; 2948 2949 /* Idle tasks are by definition preempted by non-idle tasks. */ 2950 if (unlikely(curr->policy == SCHED_IDLE) && 2951 likely(p->policy != SCHED_IDLE)) 2952 goto preempt; 2953 2954 /* 2955 * Batch and idle tasks do not preempt non-idle tasks (their preemption 2956 * is driven by the tick): 2957 */ 2958 if (unlikely(p->policy != SCHED_NORMAL)) 2959 return; 2960 2961 find_matching_se(&se, &pse); 2962 update_curr(cfs_rq_of(se)); 2963 BUG_ON(!pse); 2964 if (wakeup_preempt_entity(se, pse) == 1) { 2965 /* 2966 * Bias pick_next to pick the sched entity that is 2967 * triggering this preemption. 2968 */ 2969 if (!next_buddy_marked) 2970 set_next_buddy(pse); 2971 goto preempt; 2972 } 2973 2974 return; 2975 2976 preempt: 2977 resched_task(curr); 2978 /* 2979 * Only set the backward buddy when the current task is still 2980 * on the rq. This can happen when a wakeup gets interleaved 2981 * with schedule on the ->pre_schedule() or idle_balance() 2982 * point, either of which can * drop the rq lock. 2983 * 2984 * Also, during early boot the idle thread is in the fair class, 2985 * for obvious reasons its a bad idea to schedule back to it. 2986 */ 2987 if (unlikely(!se->on_rq || curr == rq->idle)) 2988 return; 2989 2990 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se)) 2991 set_last_buddy(se); 2992 } 2993 2994 static struct task_struct *pick_next_task_fair(struct rq *rq) 2995 { 2996 struct task_struct *p; 2997 struct cfs_rq *cfs_rq = &rq->cfs; 2998 struct sched_entity *se; 2999 3000 if (!cfs_rq->nr_running) 3001 return NULL; 3002 3003 do { 3004 se = pick_next_entity(cfs_rq); 3005 set_next_entity(cfs_rq, se); 3006 cfs_rq = group_cfs_rq(se); 3007 } while (cfs_rq); 3008 3009 p = task_of(se); 3010 if (hrtick_enabled(rq)) 3011 hrtick_start_fair(rq, p); 3012 3013 return p; 3014 } 3015 3016 /* 3017 * Account for a descheduled task: 3018 */ 3019 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev) 3020 { 3021 struct sched_entity *se = &prev->se; 3022 struct cfs_rq *cfs_rq; 3023 3024 for_each_sched_entity(se) { 3025 cfs_rq = cfs_rq_of(se); 3026 put_prev_entity(cfs_rq, se); 3027 } 3028 } 3029 3030 /* 3031 * sched_yield() is very simple 3032 * 3033 * The magic of dealing with the ->skip buddy is in pick_next_entity. 3034 */ 3035 static void yield_task_fair(struct rq *rq) 3036 { 3037 struct task_struct *curr = rq->curr; 3038 struct cfs_rq *cfs_rq = task_cfs_rq(curr); 3039 struct sched_entity *se = &curr->se; 3040 3041 /* 3042 * Are we the only task in the tree? 3043 */ 3044 if (unlikely(rq->nr_running == 1)) 3045 return; 3046 3047 clear_buddies(cfs_rq, se); 3048 3049 if (curr->policy != SCHED_BATCH) { 3050 update_rq_clock(rq); 3051 /* 3052 * Update run-time statistics of the 'current'. 3053 */ 3054 update_curr(cfs_rq); 3055 /* 3056 * Tell update_rq_clock() that we've just updated, 3057 * so we don't do microscopic update in schedule() 3058 * and double the fastpath cost. 3059 */ 3060 rq->skip_clock_update = 1; 3061 } 3062 3063 set_skip_buddy(se); 3064 } 3065 3066 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt) 3067 { 3068 struct sched_entity *se = &p->se; 3069 3070 /* throttled hierarchies are not runnable */ 3071 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se))) 3072 return false; 3073 3074 /* Tell the scheduler that we'd really like pse to run next. */ 3075 set_next_buddy(se); 3076 3077 yield_task_fair(rq); 3078 3079 return true; 3080 } 3081 3082 #ifdef CONFIG_SMP 3083 /************************************************** 3084 * Fair scheduling class load-balancing methods: 3085 */ 3086 3087 /* 3088 * pull_task - move a task from a remote runqueue to the local runqueue. 3089 * Both runqueues must be locked. 3090 */ 3091 static void pull_task(struct rq *src_rq, struct task_struct *p, 3092 struct rq *this_rq, int this_cpu) 3093 { 3094 deactivate_task(src_rq, p, 0); 3095 set_task_cpu(p, this_cpu); 3096 activate_task(this_rq, p, 0); 3097 check_preempt_curr(this_rq, p, 0); 3098 } 3099 3100 /* 3101 * Is this task likely cache-hot: 3102 */ 3103 static int 3104 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd) 3105 { 3106 s64 delta; 3107 3108 if (p->sched_class != &fair_sched_class) 3109 return 0; 3110 3111 if (unlikely(p->policy == SCHED_IDLE)) 3112 return 0; 3113 3114 /* 3115 * Buddy candidates are cache hot: 3116 */ 3117 if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running && 3118 (&p->se == cfs_rq_of(&p->se)->next || 3119 &p->se == cfs_rq_of(&p->se)->last)) 3120 return 1; 3121 3122 if (sysctl_sched_migration_cost == -1) 3123 return 1; 3124 if (sysctl_sched_migration_cost == 0) 3125 return 0; 3126 3127 delta = now - p->se.exec_start; 3128 3129 return delta < (s64)sysctl_sched_migration_cost; 3130 } 3131 3132 #define LBF_ALL_PINNED 0x01 3133 #define LBF_NEED_BREAK 0x02 /* clears into HAD_BREAK */ 3134 #define LBF_HAD_BREAK 0x04 3135 #define LBF_HAD_BREAKS 0x0C /* count HAD_BREAKs overflows into ABORT */ 3136 #define LBF_ABORT 0x10 3137 3138 /* 3139 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu? 3140 */ 3141 static 3142 int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu, 3143 struct sched_domain *sd, enum cpu_idle_type idle, 3144 int *lb_flags) 3145 { 3146 int tsk_cache_hot = 0; 3147 /* 3148 * We do not migrate tasks that are: 3149 * 1) running (obviously), or 3150 * 2) cannot be migrated to this CPU due to cpus_allowed, or 3151 * 3) are cache-hot on their current CPU. 3152 */ 3153 if (!cpumask_test_cpu(this_cpu, tsk_cpus_allowed(p))) { 3154 schedstat_inc(p, se.statistics.nr_failed_migrations_affine); 3155 return 0; 3156 } 3157 *lb_flags &= ~LBF_ALL_PINNED; 3158 3159 if (task_running(rq, p)) { 3160 schedstat_inc(p, se.statistics.nr_failed_migrations_running); 3161 return 0; 3162 } 3163 3164 /* 3165 * Aggressive migration if: 3166 * 1) task is cache cold, or 3167 * 2) too many balance attempts have failed. 3168 */ 3169 3170 tsk_cache_hot = task_hot(p, rq->clock_task, sd); 3171 if (!tsk_cache_hot || 3172 sd->nr_balance_failed > sd->cache_nice_tries) { 3173 #ifdef CONFIG_SCHEDSTATS 3174 if (tsk_cache_hot) { 3175 schedstat_inc(sd, lb_hot_gained[idle]); 3176 schedstat_inc(p, se.statistics.nr_forced_migrations); 3177 } 3178 #endif 3179 return 1; 3180 } 3181 3182 if (tsk_cache_hot) { 3183 schedstat_inc(p, se.statistics.nr_failed_migrations_hot); 3184 return 0; 3185 } 3186 return 1; 3187 } 3188 3189 /* 3190 * move_one_task tries to move exactly one task from busiest to this_rq, as 3191 * part of active balancing operations within "domain". 3192 * Returns 1 if successful and 0 otherwise. 3193 * 3194 * Called with both runqueues locked. 3195 */ 3196 static int 3197 move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest, 3198 struct sched_domain *sd, enum cpu_idle_type idle) 3199 { 3200 struct task_struct *p, *n; 3201 struct cfs_rq *cfs_rq; 3202 int pinned = 0; 3203 3204 for_each_leaf_cfs_rq(busiest, cfs_rq) { 3205 list_for_each_entry_safe(p, n, &cfs_rq->tasks, se.group_node) { 3206 if (throttled_lb_pair(task_group(p), 3207 busiest->cpu, this_cpu)) 3208 break; 3209 3210 if (!can_migrate_task(p, busiest, this_cpu, 3211 sd, idle, &pinned)) 3212 continue; 3213 3214 pull_task(busiest, p, this_rq, this_cpu); 3215 /* 3216 * Right now, this is only the second place pull_task() 3217 * is called, so we can safely collect pull_task() 3218 * stats here rather than inside pull_task(). 3219 */ 3220 schedstat_inc(sd, lb_gained[idle]); 3221 return 1; 3222 } 3223 } 3224 3225 return 0; 3226 } 3227 3228 static unsigned long 3229 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest, 3230 unsigned long max_load_move, struct sched_domain *sd, 3231 enum cpu_idle_type idle, int *lb_flags, 3232 struct cfs_rq *busiest_cfs_rq) 3233 { 3234 int loops = 0, pulled = 0; 3235 long rem_load_move = max_load_move; 3236 struct task_struct *p, *n; 3237 3238 if (max_load_move == 0) 3239 goto out; 3240 3241 list_for_each_entry_safe(p, n, &busiest_cfs_rq->tasks, se.group_node) { 3242 if (loops++ > sysctl_sched_nr_migrate) { 3243 *lb_flags |= LBF_NEED_BREAK; 3244 break; 3245 } 3246 3247 if ((p->se.load.weight >> 1) > rem_load_move || 3248 !can_migrate_task(p, busiest, this_cpu, sd, idle, 3249 lb_flags)) 3250 continue; 3251 3252 pull_task(busiest, p, this_rq, this_cpu); 3253 pulled++; 3254 rem_load_move -= p->se.load.weight; 3255 3256 #ifdef CONFIG_PREEMPT 3257 /* 3258 * NEWIDLE balancing is a source of latency, so preemptible 3259 * kernels will stop after the first task is pulled to minimize 3260 * the critical section. 3261 */ 3262 if (idle == CPU_NEWLY_IDLE) { 3263 *lb_flags |= LBF_ABORT; 3264 break; 3265 } 3266 #endif 3267 3268 /* 3269 * We only want to steal up to the prescribed amount of 3270 * weighted load. 3271 */ 3272 if (rem_load_move <= 0) 3273 break; 3274 } 3275 out: 3276 /* 3277 * Right now, this is one of only two places pull_task() is called, 3278 * so we can safely collect pull_task() stats here rather than 3279 * inside pull_task(). 3280 */ 3281 schedstat_add(sd, lb_gained[idle], pulled); 3282 3283 return max_load_move - rem_load_move; 3284 } 3285 3286 #ifdef CONFIG_FAIR_GROUP_SCHED 3287 /* 3288 * update tg->load_weight by folding this cpu's load_avg 3289 */ 3290 static int update_shares_cpu(struct task_group *tg, int cpu) 3291 { 3292 struct cfs_rq *cfs_rq; 3293 unsigned long flags; 3294 struct rq *rq; 3295 3296 if (!tg->se[cpu]) 3297 return 0; 3298 3299 rq = cpu_rq(cpu); 3300 cfs_rq = tg->cfs_rq[cpu]; 3301 3302 raw_spin_lock_irqsave(&rq->lock, flags); 3303 3304 update_rq_clock(rq); 3305 update_cfs_load(cfs_rq, 1); 3306 3307 /* 3308 * We need to update shares after updating tg->load_weight in 3309 * order to adjust the weight of groups with long running tasks. 3310 */ 3311 update_cfs_shares(cfs_rq); 3312 3313 raw_spin_unlock_irqrestore(&rq->lock, flags); 3314 3315 return 0; 3316 } 3317 3318 static void update_shares(int cpu) 3319 { 3320 struct cfs_rq *cfs_rq; 3321 struct rq *rq = cpu_rq(cpu); 3322 3323 rcu_read_lock(); 3324 /* 3325 * Iterates the task_group tree in a bottom up fashion, see 3326 * list_add_leaf_cfs_rq() for details. 3327 */ 3328 for_each_leaf_cfs_rq(rq, cfs_rq) { 3329 /* throttled entities do not contribute to load */ 3330 if (throttled_hierarchy(cfs_rq)) 3331 continue; 3332 3333 update_shares_cpu(cfs_rq->tg, cpu); 3334 } 3335 rcu_read_unlock(); 3336 } 3337 3338 /* 3339 * Compute the cpu's hierarchical load factor for each task group. 3340 * This needs to be done in a top-down fashion because the load of a child 3341 * group is a fraction of its parents load. 3342 */ 3343 static int tg_load_down(struct task_group *tg, void *data) 3344 { 3345 unsigned long load; 3346 long cpu = (long)data; 3347 3348 if (!tg->parent) { 3349 load = cpu_rq(cpu)->load.weight; 3350 } else { 3351 load = tg->parent->cfs_rq[cpu]->h_load; 3352 load *= tg->se[cpu]->load.weight; 3353 load /= tg->parent->cfs_rq[cpu]->load.weight + 1; 3354 } 3355 3356 tg->cfs_rq[cpu]->h_load = load; 3357 3358 return 0; 3359 } 3360 3361 static void update_h_load(long cpu) 3362 { 3363 walk_tg_tree(tg_load_down, tg_nop, (void *)cpu); 3364 } 3365 3366 static unsigned long 3367 load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest, 3368 unsigned long max_load_move, 3369 struct sched_domain *sd, enum cpu_idle_type idle, 3370 int *lb_flags) 3371 { 3372 long rem_load_move = max_load_move; 3373 struct cfs_rq *busiest_cfs_rq; 3374 3375 rcu_read_lock(); 3376 update_h_load(cpu_of(busiest)); 3377 3378 for_each_leaf_cfs_rq(busiest, busiest_cfs_rq) { 3379 unsigned long busiest_h_load = busiest_cfs_rq->h_load; 3380 unsigned long busiest_weight = busiest_cfs_rq->load.weight; 3381 u64 rem_load, moved_load; 3382 3383 if (*lb_flags & (LBF_NEED_BREAK|LBF_ABORT)) 3384 break; 3385 3386 /* 3387 * empty group or part of a throttled hierarchy 3388 */ 3389 if (!busiest_cfs_rq->task_weight || 3390 throttled_lb_pair(busiest_cfs_rq->tg, cpu_of(busiest), this_cpu)) 3391 continue; 3392 3393 rem_load = (u64)rem_load_move * busiest_weight; 3394 rem_load = div_u64(rem_load, busiest_h_load + 1); 3395 3396 moved_load = balance_tasks(this_rq, this_cpu, busiest, 3397 rem_load, sd, idle, lb_flags, 3398 busiest_cfs_rq); 3399 3400 if (!moved_load) 3401 continue; 3402 3403 moved_load *= busiest_h_load; 3404 moved_load = div_u64(moved_load, busiest_weight + 1); 3405 3406 rem_load_move -= moved_load; 3407 if (rem_load_move < 0) 3408 break; 3409 } 3410 rcu_read_unlock(); 3411 3412 return max_load_move - rem_load_move; 3413 } 3414 #else 3415 static inline void update_shares(int cpu) 3416 { 3417 } 3418 3419 static unsigned long 3420 load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest, 3421 unsigned long max_load_move, 3422 struct sched_domain *sd, enum cpu_idle_type idle, 3423 int *lb_flags) 3424 { 3425 return balance_tasks(this_rq, this_cpu, busiest, 3426 max_load_move, sd, idle, lb_flags, 3427 &busiest->cfs); 3428 } 3429 #endif 3430 3431 /* 3432 * move_tasks tries to move up to max_load_move weighted load from busiest to 3433 * this_rq, as part of a balancing operation within domain "sd". 3434 * Returns 1 if successful and 0 otherwise. 3435 * 3436 * Called with both runqueues locked. 3437 */ 3438 static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest, 3439 unsigned long max_load_move, 3440 struct sched_domain *sd, enum cpu_idle_type idle, 3441 int *lb_flags) 3442 { 3443 unsigned long total_load_moved = 0, load_moved; 3444 3445 do { 3446 load_moved = load_balance_fair(this_rq, this_cpu, busiest, 3447 max_load_move - total_load_moved, 3448 sd, idle, lb_flags); 3449 3450 total_load_moved += load_moved; 3451 3452 if (*lb_flags & (LBF_NEED_BREAK|LBF_ABORT)) 3453 break; 3454 3455 #ifdef CONFIG_PREEMPT 3456 /* 3457 * NEWIDLE balancing is a source of latency, so preemptible 3458 * kernels will stop after the first task is pulled to minimize 3459 * the critical section. 3460 */ 3461 if (idle == CPU_NEWLY_IDLE && this_rq->nr_running) { 3462 *lb_flags |= LBF_ABORT; 3463 break; 3464 } 3465 #endif 3466 } while (load_moved && max_load_move > total_load_moved); 3467 3468 return total_load_moved > 0; 3469 } 3470 3471 /********** Helpers for find_busiest_group ************************/ 3472 /* 3473 * sd_lb_stats - Structure to store the statistics of a sched_domain 3474 * during load balancing. 3475 */ 3476 struct sd_lb_stats { 3477 struct sched_group *busiest; /* Busiest group in this sd */ 3478 struct sched_group *this; /* Local group in this sd */ 3479 unsigned long total_load; /* Total load of all groups in sd */ 3480 unsigned long total_pwr; /* Total power of all groups in sd */ 3481 unsigned long avg_load; /* Average load across all groups in sd */ 3482 3483 /** Statistics of this group */ 3484 unsigned long this_load; 3485 unsigned long this_load_per_task; 3486 unsigned long this_nr_running; 3487 unsigned long this_has_capacity; 3488 unsigned int this_idle_cpus; 3489 3490 /* Statistics of the busiest group */ 3491 unsigned int busiest_idle_cpus; 3492 unsigned long max_load; 3493 unsigned long busiest_load_per_task; 3494 unsigned long busiest_nr_running; 3495 unsigned long busiest_group_capacity; 3496 unsigned long busiest_has_capacity; 3497 unsigned int busiest_group_weight; 3498 3499 int group_imb; /* Is there imbalance in this sd */ 3500 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT) 3501 int power_savings_balance; /* Is powersave balance needed for this sd */ 3502 struct sched_group *group_min; /* Least loaded group in sd */ 3503 struct sched_group *group_leader; /* Group which relieves group_min */ 3504 unsigned long min_load_per_task; /* load_per_task in group_min */ 3505 unsigned long leader_nr_running; /* Nr running of group_leader */ 3506 unsigned long min_nr_running; /* Nr running of group_min */ 3507 #endif 3508 }; 3509 3510 /* 3511 * sg_lb_stats - stats of a sched_group required for load_balancing 3512 */ 3513 struct sg_lb_stats { 3514 unsigned long avg_load; /*Avg load across the CPUs of the group */ 3515 unsigned long group_load; /* Total load over the CPUs of the group */ 3516 unsigned long sum_nr_running; /* Nr tasks running in the group */ 3517 unsigned long sum_weighted_load; /* Weighted load of group's tasks */ 3518 unsigned long group_capacity; 3519 unsigned long idle_cpus; 3520 unsigned long group_weight; 3521 int group_imb; /* Is there an imbalance in the group ? */ 3522 int group_has_capacity; /* Is there extra capacity in the group? */ 3523 }; 3524 3525 /** 3526 * get_sd_load_idx - Obtain the load index for a given sched domain. 3527 * @sd: The sched_domain whose load_idx is to be obtained. 3528 * @idle: The Idle status of the CPU for whose sd load_icx is obtained. 3529 */ 3530 static inline int get_sd_load_idx(struct sched_domain *sd, 3531 enum cpu_idle_type idle) 3532 { 3533 int load_idx; 3534 3535 switch (idle) { 3536 case CPU_NOT_IDLE: 3537 load_idx = sd->busy_idx; 3538 break; 3539 3540 case CPU_NEWLY_IDLE: 3541 load_idx = sd->newidle_idx; 3542 break; 3543 default: 3544 load_idx = sd->idle_idx; 3545 break; 3546 } 3547 3548 return load_idx; 3549 } 3550 3551 3552 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT) 3553 /** 3554 * init_sd_power_savings_stats - Initialize power savings statistics for 3555 * the given sched_domain, during load balancing. 3556 * 3557 * @sd: Sched domain whose power-savings statistics are to be initialized. 3558 * @sds: Variable containing the statistics for sd. 3559 * @idle: Idle status of the CPU at which we're performing load-balancing. 3560 */ 3561 static inline void init_sd_power_savings_stats(struct sched_domain *sd, 3562 struct sd_lb_stats *sds, enum cpu_idle_type idle) 3563 { 3564 /* 3565 * Busy processors will not participate in power savings 3566 * balance. 3567 */ 3568 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE)) 3569 sds->power_savings_balance = 0; 3570 else { 3571 sds->power_savings_balance = 1; 3572 sds->min_nr_running = ULONG_MAX; 3573 sds->leader_nr_running = 0; 3574 } 3575 } 3576 3577 /** 3578 * update_sd_power_savings_stats - Update the power saving stats for a 3579 * sched_domain while performing load balancing. 3580 * 3581 * @group: sched_group belonging to the sched_domain under consideration. 3582 * @sds: Variable containing the statistics of the sched_domain 3583 * @local_group: Does group contain the CPU for which we're performing 3584 * load balancing ? 3585 * @sgs: Variable containing the statistics of the group. 3586 */ 3587 static inline void update_sd_power_savings_stats(struct sched_group *group, 3588 struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs) 3589 { 3590 3591 if (!sds->power_savings_balance) 3592 return; 3593 3594 /* 3595 * If the local group is idle or completely loaded 3596 * no need to do power savings balance at this domain 3597 */ 3598 if (local_group && (sds->this_nr_running >= sgs->group_capacity || 3599 !sds->this_nr_running)) 3600 sds->power_savings_balance = 0; 3601 3602 /* 3603 * If a group is already running at full capacity or idle, 3604 * don't include that group in power savings calculations 3605 */ 3606 if (!sds->power_savings_balance || 3607 sgs->sum_nr_running >= sgs->group_capacity || 3608 !sgs->sum_nr_running) 3609 return; 3610 3611 /* 3612 * Calculate the group which has the least non-idle load. 3613 * This is the group from where we need to pick up the load 3614 * for saving power 3615 */ 3616 if ((sgs->sum_nr_running < sds->min_nr_running) || 3617 (sgs->sum_nr_running == sds->min_nr_running && 3618 group_first_cpu(group) > group_first_cpu(sds->group_min))) { 3619 sds->group_min = group; 3620 sds->min_nr_running = sgs->sum_nr_running; 3621 sds->min_load_per_task = sgs->sum_weighted_load / 3622 sgs->sum_nr_running; 3623 } 3624 3625 /* 3626 * Calculate the group which is almost near its 3627 * capacity but still has some space to pick up some load 3628 * from other group and save more power 3629 */ 3630 if (sgs->sum_nr_running + 1 > sgs->group_capacity) 3631 return; 3632 3633 if (sgs->sum_nr_running > sds->leader_nr_running || 3634 (sgs->sum_nr_running == sds->leader_nr_running && 3635 group_first_cpu(group) < group_first_cpu(sds->group_leader))) { 3636 sds->group_leader = group; 3637 sds->leader_nr_running = sgs->sum_nr_running; 3638 } 3639 } 3640 3641 /** 3642 * check_power_save_busiest_group - see if there is potential for some power-savings balance 3643 * @sds: Variable containing the statistics of the sched_domain 3644 * under consideration. 3645 * @this_cpu: Cpu at which we're currently performing load-balancing. 3646 * @imbalance: Variable to store the imbalance. 3647 * 3648 * Description: 3649 * Check if we have potential to perform some power-savings balance. 3650 * If yes, set the busiest group to be the least loaded group in the 3651 * sched_domain, so that it's CPUs can be put to idle. 3652 * 3653 * Returns 1 if there is potential to perform power-savings balance. 3654 * Else returns 0. 3655 */ 3656 static inline int check_power_save_busiest_group(struct sd_lb_stats *sds, 3657 int this_cpu, unsigned long *imbalance) 3658 { 3659 if (!sds->power_savings_balance) 3660 return 0; 3661 3662 if (sds->this != sds->group_leader || 3663 sds->group_leader == sds->group_min) 3664 return 0; 3665 3666 *imbalance = sds->min_load_per_task; 3667 sds->busiest = sds->group_min; 3668 3669 return 1; 3670 3671 } 3672 #else /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */ 3673 static inline void init_sd_power_savings_stats(struct sched_domain *sd, 3674 struct sd_lb_stats *sds, enum cpu_idle_type idle) 3675 { 3676 return; 3677 } 3678 3679 static inline void update_sd_power_savings_stats(struct sched_group *group, 3680 struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs) 3681 { 3682 return; 3683 } 3684 3685 static inline int check_power_save_busiest_group(struct sd_lb_stats *sds, 3686 int this_cpu, unsigned long *imbalance) 3687 { 3688 return 0; 3689 } 3690 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */ 3691 3692 3693 unsigned long default_scale_freq_power(struct sched_domain *sd, int cpu) 3694 { 3695 return SCHED_POWER_SCALE; 3696 } 3697 3698 unsigned long __weak arch_scale_freq_power(struct sched_domain *sd, int cpu) 3699 { 3700 return default_scale_freq_power(sd, cpu); 3701 } 3702 3703 unsigned long default_scale_smt_power(struct sched_domain *sd, int cpu) 3704 { 3705 unsigned long weight = sd->span_weight; 3706 unsigned long smt_gain = sd->smt_gain; 3707 3708 smt_gain /= weight; 3709 3710 return smt_gain; 3711 } 3712 3713 unsigned long __weak arch_scale_smt_power(struct sched_domain *sd, int cpu) 3714 { 3715 return default_scale_smt_power(sd, cpu); 3716 } 3717 3718 unsigned long scale_rt_power(int cpu) 3719 { 3720 struct rq *rq = cpu_rq(cpu); 3721 u64 total, available; 3722 3723 total = sched_avg_period() + (rq->clock - rq->age_stamp); 3724 3725 if (unlikely(total < rq->rt_avg)) { 3726 /* Ensures that power won't end up being negative */ 3727 available = 0; 3728 } else { 3729 available = total - rq->rt_avg; 3730 } 3731 3732 if (unlikely((s64)total < SCHED_POWER_SCALE)) 3733 total = SCHED_POWER_SCALE; 3734 3735 total >>= SCHED_POWER_SHIFT; 3736 3737 return div_u64(available, total); 3738 } 3739 3740 static void update_cpu_power(struct sched_domain *sd, int cpu) 3741 { 3742 unsigned long weight = sd->span_weight; 3743 unsigned long power = SCHED_POWER_SCALE; 3744 struct sched_group *sdg = sd->groups; 3745 3746 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) { 3747 if (sched_feat(ARCH_POWER)) 3748 power *= arch_scale_smt_power(sd, cpu); 3749 else 3750 power *= default_scale_smt_power(sd, cpu); 3751 3752 power >>= SCHED_POWER_SHIFT; 3753 } 3754 3755 sdg->sgp->power_orig = power; 3756 3757 if (sched_feat(ARCH_POWER)) 3758 power *= arch_scale_freq_power(sd, cpu); 3759 else 3760 power *= default_scale_freq_power(sd, cpu); 3761 3762 power >>= SCHED_POWER_SHIFT; 3763 3764 power *= scale_rt_power(cpu); 3765 power >>= SCHED_POWER_SHIFT; 3766 3767 if (!power) 3768 power = 1; 3769 3770 cpu_rq(cpu)->cpu_power = power; 3771 sdg->sgp->power = power; 3772 } 3773 3774 void update_group_power(struct sched_domain *sd, int cpu) 3775 { 3776 struct sched_domain *child = sd->child; 3777 struct sched_group *group, *sdg = sd->groups; 3778 unsigned long power; 3779 3780 if (!child) { 3781 update_cpu_power(sd, cpu); 3782 return; 3783 } 3784 3785 power = 0; 3786 3787 group = child->groups; 3788 do { 3789 power += group->sgp->power; 3790 group = group->next; 3791 } while (group != child->groups); 3792 3793 sdg->sgp->power = power; 3794 } 3795 3796 /* 3797 * Try and fix up capacity for tiny siblings, this is needed when 3798 * things like SD_ASYM_PACKING need f_b_g to select another sibling 3799 * which on its own isn't powerful enough. 3800 * 3801 * See update_sd_pick_busiest() and check_asym_packing(). 3802 */ 3803 static inline int 3804 fix_small_capacity(struct sched_domain *sd, struct sched_group *group) 3805 { 3806 /* 3807 * Only siblings can have significantly less than SCHED_POWER_SCALE 3808 */ 3809 if (!(sd->flags & SD_SHARE_CPUPOWER)) 3810 return 0; 3811 3812 /* 3813 * If ~90% of the cpu_power is still there, we're good. 3814 */ 3815 if (group->sgp->power * 32 > group->sgp->power_orig * 29) 3816 return 1; 3817 3818 return 0; 3819 } 3820 3821 /** 3822 * update_sg_lb_stats - Update sched_group's statistics for load balancing. 3823 * @sd: The sched_domain whose statistics are to be updated. 3824 * @group: sched_group whose statistics are to be updated. 3825 * @this_cpu: Cpu for which load balance is currently performed. 3826 * @idle: Idle status of this_cpu 3827 * @load_idx: Load index of sched_domain of this_cpu for load calc. 3828 * @local_group: Does group contain this_cpu. 3829 * @cpus: Set of cpus considered for load balancing. 3830 * @balance: Should we balance. 3831 * @sgs: variable to hold the statistics for this group. 3832 */ 3833 static inline void update_sg_lb_stats(struct sched_domain *sd, 3834 struct sched_group *group, int this_cpu, 3835 enum cpu_idle_type idle, int load_idx, 3836 int local_group, const struct cpumask *cpus, 3837 int *balance, struct sg_lb_stats *sgs) 3838 { 3839 unsigned long load, max_cpu_load, min_cpu_load, max_nr_running; 3840 int i; 3841 unsigned int balance_cpu = -1, first_idle_cpu = 0; 3842 unsigned long avg_load_per_task = 0; 3843 3844 if (local_group) 3845 balance_cpu = group_first_cpu(group); 3846 3847 /* Tally up the load of all CPUs in the group */ 3848 max_cpu_load = 0; 3849 min_cpu_load = ~0UL; 3850 max_nr_running = 0; 3851 3852 for_each_cpu_and(i, sched_group_cpus(group), cpus) { 3853 struct rq *rq = cpu_rq(i); 3854 3855 /* Bias balancing toward cpus of our domain */ 3856 if (local_group) { 3857 if (idle_cpu(i) && !first_idle_cpu) { 3858 first_idle_cpu = 1; 3859 balance_cpu = i; 3860 } 3861 3862 load = target_load(i, load_idx); 3863 } else { 3864 load = source_load(i, load_idx); 3865 if (load > max_cpu_load) { 3866 max_cpu_load = load; 3867 max_nr_running = rq->nr_running; 3868 } 3869 if (min_cpu_load > load) 3870 min_cpu_load = load; 3871 } 3872 3873 sgs->group_load += load; 3874 sgs->sum_nr_running += rq->nr_running; 3875 sgs->sum_weighted_load += weighted_cpuload(i); 3876 if (idle_cpu(i)) 3877 sgs->idle_cpus++; 3878 } 3879 3880 /* 3881 * First idle cpu or the first cpu(busiest) in this sched group 3882 * is eligible for doing load balancing at this and above 3883 * domains. In the newly idle case, we will allow all the cpu's 3884 * to do the newly idle load balance. 3885 */ 3886 if (idle != CPU_NEWLY_IDLE && local_group) { 3887 if (balance_cpu != this_cpu) { 3888 *balance = 0; 3889 return; 3890 } 3891 update_group_power(sd, this_cpu); 3892 } 3893 3894 /* Adjust by relative CPU power of the group */ 3895 sgs->avg_load = (sgs->group_load*SCHED_POWER_SCALE) / group->sgp->power; 3896 3897 /* 3898 * Consider the group unbalanced when the imbalance is larger 3899 * than the average weight of a task. 3900 * 3901 * APZ: with cgroup the avg task weight can vary wildly and 3902 * might not be a suitable number - should we keep a 3903 * normalized nr_running number somewhere that negates 3904 * the hierarchy? 3905 */ 3906 if (sgs->sum_nr_running) 3907 avg_load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running; 3908 3909 if ((max_cpu_load - min_cpu_load) >= avg_load_per_task && max_nr_running > 1) 3910 sgs->group_imb = 1; 3911 3912 sgs->group_capacity = DIV_ROUND_CLOSEST(group->sgp->power, 3913 SCHED_POWER_SCALE); 3914 if (!sgs->group_capacity) 3915 sgs->group_capacity = fix_small_capacity(sd, group); 3916 sgs->group_weight = group->group_weight; 3917 3918 if (sgs->group_capacity > sgs->sum_nr_running) 3919 sgs->group_has_capacity = 1; 3920 } 3921 3922 /** 3923 * update_sd_pick_busiest - return 1 on busiest group 3924 * @sd: sched_domain whose statistics are to be checked 3925 * @sds: sched_domain statistics 3926 * @sg: sched_group candidate to be checked for being the busiest 3927 * @sgs: sched_group statistics 3928 * @this_cpu: the current cpu 3929 * 3930 * Determine if @sg is a busier group than the previously selected 3931 * busiest group. 3932 */ 3933 static bool update_sd_pick_busiest(struct sched_domain *sd, 3934 struct sd_lb_stats *sds, 3935 struct sched_group *sg, 3936 struct sg_lb_stats *sgs, 3937 int this_cpu) 3938 { 3939 if (sgs->avg_load <= sds->max_load) 3940 return false; 3941 3942 if (sgs->sum_nr_running > sgs->group_capacity) 3943 return true; 3944 3945 if (sgs->group_imb) 3946 return true; 3947 3948 /* 3949 * ASYM_PACKING needs to move all the work to the lowest 3950 * numbered CPUs in the group, therefore mark all groups 3951 * higher than ourself as busy. 3952 */ 3953 if ((sd->flags & SD_ASYM_PACKING) && sgs->sum_nr_running && 3954 this_cpu < group_first_cpu(sg)) { 3955 if (!sds->busiest) 3956 return true; 3957 3958 if (group_first_cpu(sds->busiest) > group_first_cpu(sg)) 3959 return true; 3960 } 3961 3962 return false; 3963 } 3964 3965 /** 3966 * update_sd_lb_stats - Update sched_domain's statistics for load balancing. 3967 * @sd: sched_domain whose statistics are to be updated. 3968 * @this_cpu: Cpu for which load balance is currently performed. 3969 * @idle: Idle status of this_cpu 3970 * @cpus: Set of cpus considered for load balancing. 3971 * @balance: Should we balance. 3972 * @sds: variable to hold the statistics for this sched_domain. 3973 */ 3974 static inline void update_sd_lb_stats(struct sched_domain *sd, int this_cpu, 3975 enum cpu_idle_type idle, const struct cpumask *cpus, 3976 int *balance, struct sd_lb_stats *sds) 3977 { 3978 struct sched_domain *child = sd->child; 3979 struct sched_group *sg = sd->groups; 3980 struct sg_lb_stats sgs; 3981 int load_idx, prefer_sibling = 0; 3982 3983 if (child && child->flags & SD_PREFER_SIBLING) 3984 prefer_sibling = 1; 3985 3986 init_sd_power_savings_stats(sd, sds, idle); 3987 load_idx = get_sd_load_idx(sd, idle); 3988 3989 do { 3990 int local_group; 3991 3992 local_group = cpumask_test_cpu(this_cpu, sched_group_cpus(sg)); 3993 memset(&sgs, 0, sizeof(sgs)); 3994 update_sg_lb_stats(sd, sg, this_cpu, idle, load_idx, 3995 local_group, cpus, balance, &sgs); 3996 3997 if (local_group && !(*balance)) 3998 return; 3999 4000 sds->total_load += sgs.group_load; 4001 sds->total_pwr += sg->sgp->power; 4002 4003 /* 4004 * In case the child domain prefers tasks go to siblings 4005 * first, lower the sg capacity to one so that we'll try 4006 * and move all the excess tasks away. We lower the capacity 4007 * of a group only if the local group has the capacity to fit 4008 * these excess tasks, i.e. nr_running < group_capacity. The 4009 * extra check prevents the case where you always pull from the 4010 * heaviest group when it is already under-utilized (possible 4011 * with a large weight task outweighs the tasks on the system). 4012 */ 4013 if (prefer_sibling && !local_group && sds->this_has_capacity) 4014 sgs.group_capacity = min(sgs.group_capacity, 1UL); 4015 4016 if (local_group) { 4017 sds->this_load = sgs.avg_load; 4018 sds->this = sg; 4019 sds->this_nr_running = sgs.sum_nr_running; 4020 sds->this_load_per_task = sgs.sum_weighted_load; 4021 sds->this_has_capacity = sgs.group_has_capacity; 4022 sds->this_idle_cpus = sgs.idle_cpus; 4023 } else if (update_sd_pick_busiest(sd, sds, sg, &sgs, this_cpu)) { 4024 sds->max_load = sgs.avg_load; 4025 sds->busiest = sg; 4026 sds->busiest_nr_running = sgs.sum_nr_running; 4027 sds->busiest_idle_cpus = sgs.idle_cpus; 4028 sds->busiest_group_capacity = sgs.group_capacity; 4029 sds->busiest_load_per_task = sgs.sum_weighted_load; 4030 sds->busiest_has_capacity = sgs.group_has_capacity; 4031 sds->busiest_group_weight = sgs.group_weight; 4032 sds->group_imb = sgs.group_imb; 4033 } 4034 4035 update_sd_power_savings_stats(sg, sds, local_group, &sgs); 4036 sg = sg->next; 4037 } while (sg != sd->groups); 4038 } 4039 4040 /** 4041 * check_asym_packing - Check to see if the group is packed into the 4042 * sched doman. 4043 * 4044 * This is primarily intended to used at the sibling level. Some 4045 * cores like POWER7 prefer to use lower numbered SMT threads. In the 4046 * case of POWER7, it can move to lower SMT modes only when higher 4047 * threads are idle. When in lower SMT modes, the threads will 4048 * perform better since they share less core resources. Hence when we 4049 * have idle threads, we want them to be the higher ones. 4050 * 4051 * This packing function is run on idle threads. It checks to see if 4052 * the busiest CPU in this domain (core in the P7 case) has a higher 4053 * CPU number than the packing function is being run on. Here we are 4054 * assuming lower CPU number will be equivalent to lower a SMT thread 4055 * number. 4056 * 4057 * Returns 1 when packing is required and a task should be moved to 4058 * this CPU. The amount of the imbalance is returned in *imbalance. 4059 * 4060 * @sd: The sched_domain whose packing is to be checked. 4061 * @sds: Statistics of the sched_domain which is to be packed 4062 * @this_cpu: The cpu at whose sched_domain we're performing load-balance. 4063 * @imbalance: returns amount of imbalanced due to packing. 4064 */ 4065 static int check_asym_packing(struct sched_domain *sd, 4066 struct sd_lb_stats *sds, 4067 int this_cpu, unsigned long *imbalance) 4068 { 4069 int busiest_cpu; 4070 4071 if (!(sd->flags & SD_ASYM_PACKING)) 4072 return 0; 4073 4074 if (!sds->busiest) 4075 return 0; 4076 4077 busiest_cpu = group_first_cpu(sds->busiest); 4078 if (this_cpu > busiest_cpu) 4079 return 0; 4080 4081 *imbalance = DIV_ROUND_CLOSEST(sds->max_load * sds->busiest->sgp->power, 4082 SCHED_POWER_SCALE); 4083 return 1; 4084 } 4085 4086 /** 4087 * fix_small_imbalance - Calculate the minor imbalance that exists 4088 * amongst the groups of a sched_domain, during 4089 * load balancing. 4090 * @sds: Statistics of the sched_domain whose imbalance is to be calculated. 4091 * @this_cpu: The cpu at whose sched_domain we're performing load-balance. 4092 * @imbalance: Variable to store the imbalance. 4093 */ 4094 static inline void fix_small_imbalance(struct sd_lb_stats *sds, 4095 int this_cpu, unsigned long *imbalance) 4096 { 4097 unsigned long tmp, pwr_now = 0, pwr_move = 0; 4098 unsigned int imbn = 2; 4099 unsigned long scaled_busy_load_per_task; 4100 4101 if (sds->this_nr_running) { 4102 sds->this_load_per_task /= sds->this_nr_running; 4103 if (sds->busiest_load_per_task > 4104 sds->this_load_per_task) 4105 imbn = 1; 4106 } else 4107 sds->this_load_per_task = 4108 cpu_avg_load_per_task(this_cpu); 4109 4110 scaled_busy_load_per_task = sds->busiest_load_per_task 4111 * SCHED_POWER_SCALE; 4112 scaled_busy_load_per_task /= sds->busiest->sgp->power; 4113 4114 if (sds->max_load - sds->this_load + scaled_busy_load_per_task >= 4115 (scaled_busy_load_per_task * imbn)) { 4116 *imbalance = sds->busiest_load_per_task; 4117 return; 4118 } 4119 4120 /* 4121 * OK, we don't have enough imbalance to justify moving tasks, 4122 * however we may be able to increase total CPU power used by 4123 * moving them. 4124 */ 4125 4126 pwr_now += sds->busiest->sgp->power * 4127 min(sds->busiest_load_per_task, sds->max_load); 4128 pwr_now += sds->this->sgp->power * 4129 min(sds->this_load_per_task, sds->this_load); 4130 pwr_now /= SCHED_POWER_SCALE; 4131 4132 /* Amount of load we'd subtract */ 4133 tmp = (sds->busiest_load_per_task * SCHED_POWER_SCALE) / 4134 sds->busiest->sgp->power; 4135 if (sds->max_load > tmp) 4136 pwr_move += sds->busiest->sgp->power * 4137 min(sds->busiest_load_per_task, sds->max_load - tmp); 4138 4139 /* Amount of load we'd add */ 4140 if (sds->max_load * sds->busiest->sgp->power < 4141 sds->busiest_load_per_task * SCHED_POWER_SCALE) 4142 tmp = (sds->max_load * sds->busiest->sgp->power) / 4143 sds->this->sgp->power; 4144 else 4145 tmp = (sds->busiest_load_per_task * SCHED_POWER_SCALE) / 4146 sds->this->sgp->power; 4147 pwr_move += sds->this->sgp->power * 4148 min(sds->this_load_per_task, sds->this_load + tmp); 4149 pwr_move /= SCHED_POWER_SCALE; 4150 4151 /* Move if we gain throughput */ 4152 if (pwr_move > pwr_now) 4153 *imbalance = sds->busiest_load_per_task; 4154 } 4155 4156 /** 4157 * calculate_imbalance - Calculate the amount of imbalance present within the 4158 * groups of a given sched_domain during load balance. 4159 * @sds: statistics of the sched_domain whose imbalance is to be calculated. 4160 * @this_cpu: Cpu for which currently load balance is being performed. 4161 * @imbalance: The variable to store the imbalance. 4162 */ 4163 static inline void calculate_imbalance(struct sd_lb_stats *sds, int this_cpu, 4164 unsigned long *imbalance) 4165 { 4166 unsigned long max_pull, load_above_capacity = ~0UL; 4167 4168 sds->busiest_load_per_task /= sds->busiest_nr_running; 4169 if (sds->group_imb) { 4170 sds->busiest_load_per_task = 4171 min(sds->busiest_load_per_task, sds->avg_load); 4172 } 4173 4174 /* 4175 * In the presence of smp nice balancing, certain scenarios can have 4176 * max load less than avg load(as we skip the groups at or below 4177 * its cpu_power, while calculating max_load..) 4178 */ 4179 if (sds->max_load < sds->avg_load) { 4180 *imbalance = 0; 4181 return fix_small_imbalance(sds, this_cpu, imbalance); 4182 } 4183 4184 if (!sds->group_imb) { 4185 /* 4186 * Don't want to pull so many tasks that a group would go idle. 4187 */ 4188 load_above_capacity = (sds->busiest_nr_running - 4189 sds->busiest_group_capacity); 4190 4191 load_above_capacity *= (SCHED_LOAD_SCALE * SCHED_POWER_SCALE); 4192 4193 load_above_capacity /= sds->busiest->sgp->power; 4194 } 4195 4196 /* 4197 * We're trying to get all the cpus to the average_load, so we don't 4198 * want to push ourselves above the average load, nor do we wish to 4199 * reduce the max loaded cpu below the average load. At the same time, 4200 * we also don't want to reduce the group load below the group capacity 4201 * (so that we can implement power-savings policies etc). Thus we look 4202 * for the minimum possible imbalance. 4203 * Be careful of negative numbers as they'll appear as very large values 4204 * with unsigned longs. 4205 */ 4206 max_pull = min(sds->max_load - sds->avg_load, load_above_capacity); 4207 4208 /* How much load to actually move to equalise the imbalance */ 4209 *imbalance = min(max_pull * sds->busiest->sgp->power, 4210 (sds->avg_load - sds->this_load) * sds->this->sgp->power) 4211 / SCHED_POWER_SCALE; 4212 4213 /* 4214 * if *imbalance is less than the average load per runnable task 4215 * there is no guarantee that any tasks will be moved so we'll have 4216 * a think about bumping its value to force at least one task to be 4217 * moved 4218 */ 4219 if (*imbalance < sds->busiest_load_per_task) 4220 return fix_small_imbalance(sds, this_cpu, imbalance); 4221 4222 } 4223 4224 /******* find_busiest_group() helpers end here *********************/ 4225 4226 /** 4227 * find_busiest_group - Returns the busiest group within the sched_domain 4228 * if there is an imbalance. If there isn't an imbalance, and 4229 * the user has opted for power-savings, it returns a group whose 4230 * CPUs can be put to idle by rebalancing those tasks elsewhere, if 4231 * such a group exists. 4232 * 4233 * Also calculates the amount of weighted load which should be moved 4234 * to restore balance. 4235 * 4236 * @sd: The sched_domain whose busiest group is to be returned. 4237 * @this_cpu: The cpu for which load balancing is currently being performed. 4238 * @imbalance: Variable which stores amount of weighted load which should 4239 * be moved to restore balance/put a group to idle. 4240 * @idle: The idle status of this_cpu. 4241 * @cpus: The set of CPUs under consideration for load-balancing. 4242 * @balance: Pointer to a variable indicating if this_cpu 4243 * is the appropriate cpu to perform load balancing at this_level. 4244 * 4245 * Returns: - the busiest group if imbalance exists. 4246 * - If no imbalance and user has opted for power-savings balance, 4247 * return the least loaded group whose CPUs can be 4248 * put to idle by rebalancing its tasks onto our group. 4249 */ 4250 static struct sched_group * 4251 find_busiest_group(struct sched_domain *sd, int this_cpu, 4252 unsigned long *imbalance, enum cpu_idle_type idle, 4253 const struct cpumask *cpus, int *balance) 4254 { 4255 struct sd_lb_stats sds; 4256 4257 memset(&sds, 0, sizeof(sds)); 4258 4259 /* 4260 * Compute the various statistics relavent for load balancing at 4261 * this level. 4262 */ 4263 update_sd_lb_stats(sd, this_cpu, idle, cpus, balance, &sds); 4264 4265 /* 4266 * this_cpu is not the appropriate cpu to perform load balancing at 4267 * this level. 4268 */ 4269 if (!(*balance)) 4270 goto ret; 4271 4272 if ((idle == CPU_IDLE || idle == CPU_NEWLY_IDLE) && 4273 check_asym_packing(sd, &sds, this_cpu, imbalance)) 4274 return sds.busiest; 4275 4276 /* There is no busy sibling group to pull tasks from */ 4277 if (!sds.busiest || sds.busiest_nr_running == 0) 4278 goto out_balanced; 4279 4280 sds.avg_load = (SCHED_POWER_SCALE * sds.total_load) / sds.total_pwr; 4281 4282 /* 4283 * If the busiest group is imbalanced the below checks don't 4284 * work because they assumes all things are equal, which typically 4285 * isn't true due to cpus_allowed constraints and the like. 4286 */ 4287 if (sds.group_imb) 4288 goto force_balance; 4289 4290 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */ 4291 if (idle == CPU_NEWLY_IDLE && sds.this_has_capacity && 4292 !sds.busiest_has_capacity) 4293 goto force_balance; 4294 4295 /* 4296 * If the local group is more busy than the selected busiest group 4297 * don't try and pull any tasks. 4298 */ 4299 if (sds.this_load >= sds.max_load) 4300 goto out_balanced; 4301 4302 /* 4303 * Don't pull any tasks if this group is already above the domain 4304 * average load. 4305 */ 4306 if (sds.this_load >= sds.avg_load) 4307 goto out_balanced; 4308 4309 if (idle == CPU_IDLE) { 4310 /* 4311 * This cpu is idle. If the busiest group load doesn't 4312 * have more tasks than the number of available cpu's and 4313 * there is no imbalance between this and busiest group 4314 * wrt to idle cpu's, it is balanced. 4315 */ 4316 if ((sds.this_idle_cpus <= sds.busiest_idle_cpus + 1) && 4317 sds.busiest_nr_running <= sds.busiest_group_weight) 4318 goto out_balanced; 4319 } else { 4320 /* 4321 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use 4322 * imbalance_pct to be conservative. 4323 */ 4324 if (100 * sds.max_load <= sd->imbalance_pct * sds.this_load) 4325 goto out_balanced; 4326 } 4327 4328 force_balance: 4329 /* Looks like there is an imbalance. Compute it */ 4330 calculate_imbalance(&sds, this_cpu, imbalance); 4331 return sds.busiest; 4332 4333 out_balanced: 4334 /* 4335 * There is no obvious imbalance. But check if we can do some balancing 4336 * to save power. 4337 */ 4338 if (check_power_save_busiest_group(&sds, this_cpu, imbalance)) 4339 return sds.busiest; 4340 ret: 4341 *imbalance = 0; 4342 return NULL; 4343 } 4344 4345 /* 4346 * find_busiest_queue - find the busiest runqueue among the cpus in group. 4347 */ 4348 static struct rq * 4349 find_busiest_queue(struct sched_domain *sd, struct sched_group *group, 4350 enum cpu_idle_type idle, unsigned long imbalance, 4351 const struct cpumask *cpus) 4352 { 4353 struct rq *busiest = NULL, *rq; 4354 unsigned long max_load = 0; 4355 int i; 4356 4357 for_each_cpu(i, sched_group_cpus(group)) { 4358 unsigned long power = power_of(i); 4359 unsigned long capacity = DIV_ROUND_CLOSEST(power, 4360 SCHED_POWER_SCALE); 4361 unsigned long wl; 4362 4363 if (!capacity) 4364 capacity = fix_small_capacity(sd, group); 4365 4366 if (!cpumask_test_cpu(i, cpus)) 4367 continue; 4368 4369 rq = cpu_rq(i); 4370 wl = weighted_cpuload(i); 4371 4372 /* 4373 * When comparing with imbalance, use weighted_cpuload() 4374 * which is not scaled with the cpu power. 4375 */ 4376 if (capacity && rq->nr_running == 1 && wl > imbalance) 4377 continue; 4378 4379 /* 4380 * For the load comparisons with the other cpu's, consider 4381 * the weighted_cpuload() scaled with the cpu power, so that 4382 * the load can be moved away from the cpu that is potentially 4383 * running at a lower capacity. 4384 */ 4385 wl = (wl * SCHED_POWER_SCALE) / power; 4386 4387 if (wl > max_load) { 4388 max_load = wl; 4389 busiest = rq; 4390 } 4391 } 4392 4393 return busiest; 4394 } 4395 4396 /* 4397 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but 4398 * so long as it is large enough. 4399 */ 4400 #define MAX_PINNED_INTERVAL 512 4401 4402 /* Working cpumask for load_balance and load_balance_newidle. */ 4403 DEFINE_PER_CPU(cpumask_var_t, load_balance_tmpmask); 4404 4405 static int need_active_balance(struct sched_domain *sd, int idle, 4406 int busiest_cpu, int this_cpu) 4407 { 4408 if (idle == CPU_NEWLY_IDLE) { 4409 4410 /* 4411 * ASYM_PACKING needs to force migrate tasks from busy but 4412 * higher numbered CPUs in order to pack all tasks in the 4413 * lowest numbered CPUs. 4414 */ 4415 if ((sd->flags & SD_ASYM_PACKING) && busiest_cpu > this_cpu) 4416 return 1; 4417 4418 /* 4419 * The only task running in a non-idle cpu can be moved to this 4420 * cpu in an attempt to completely freeup the other CPU 4421 * package. 4422 * 4423 * The package power saving logic comes from 4424 * find_busiest_group(). If there are no imbalance, then 4425 * f_b_g() will return NULL. However when sched_mc={1,2} then 4426 * f_b_g() will select a group from which a running task may be 4427 * pulled to this cpu in order to make the other package idle. 4428 * If there is no opportunity to make a package idle and if 4429 * there are no imbalance, then f_b_g() will return NULL and no 4430 * action will be taken in load_balance_newidle(). 4431 * 4432 * Under normal task pull operation due to imbalance, there 4433 * will be more than one task in the source run queue and 4434 * move_tasks() will succeed. ld_moved will be true and this 4435 * active balance code will not be triggered. 4436 */ 4437 if (sched_mc_power_savings < POWERSAVINGS_BALANCE_WAKEUP) 4438 return 0; 4439 } 4440 4441 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2); 4442 } 4443 4444 static int active_load_balance_cpu_stop(void *data); 4445 4446 /* 4447 * Check this_cpu to ensure it is balanced within domain. Attempt to move 4448 * tasks if there is an imbalance. 4449 */ 4450 static int load_balance(int this_cpu, struct rq *this_rq, 4451 struct sched_domain *sd, enum cpu_idle_type idle, 4452 int *balance) 4453 { 4454 int ld_moved, lb_flags = 0, active_balance = 0; 4455 struct sched_group *group; 4456 unsigned long imbalance; 4457 struct rq *busiest; 4458 unsigned long flags; 4459 struct cpumask *cpus = __get_cpu_var(load_balance_tmpmask); 4460 4461 cpumask_copy(cpus, cpu_active_mask); 4462 4463 schedstat_inc(sd, lb_count[idle]); 4464 4465 redo: 4466 group = find_busiest_group(sd, this_cpu, &imbalance, idle, 4467 cpus, balance); 4468 4469 if (*balance == 0) 4470 goto out_balanced; 4471 4472 if (!group) { 4473 schedstat_inc(sd, lb_nobusyg[idle]); 4474 goto out_balanced; 4475 } 4476 4477 busiest = find_busiest_queue(sd, group, idle, imbalance, cpus); 4478 if (!busiest) { 4479 schedstat_inc(sd, lb_nobusyq[idle]); 4480 goto out_balanced; 4481 } 4482 4483 BUG_ON(busiest == this_rq); 4484 4485 schedstat_add(sd, lb_imbalance[idle], imbalance); 4486 4487 ld_moved = 0; 4488 if (busiest->nr_running > 1) { 4489 /* 4490 * Attempt to move tasks. If find_busiest_group has found 4491 * an imbalance but busiest->nr_running <= 1, the group is 4492 * still unbalanced. ld_moved simply stays zero, so it is 4493 * correctly treated as an imbalance. 4494 */ 4495 lb_flags |= LBF_ALL_PINNED; 4496 local_irq_save(flags); 4497 double_rq_lock(this_rq, busiest); 4498 ld_moved = move_tasks(this_rq, this_cpu, busiest, 4499 imbalance, sd, idle, &lb_flags); 4500 double_rq_unlock(this_rq, busiest); 4501 local_irq_restore(flags); 4502 4503 /* 4504 * some other cpu did the load balance for us. 4505 */ 4506 if (ld_moved && this_cpu != smp_processor_id()) 4507 resched_cpu(this_cpu); 4508 4509 if (lb_flags & LBF_ABORT) 4510 goto out_balanced; 4511 4512 if (lb_flags & LBF_NEED_BREAK) { 4513 lb_flags += LBF_HAD_BREAK - LBF_NEED_BREAK; 4514 if (lb_flags & LBF_ABORT) 4515 goto out_balanced; 4516 goto redo; 4517 } 4518 4519 /* All tasks on this runqueue were pinned by CPU affinity */ 4520 if (unlikely(lb_flags & LBF_ALL_PINNED)) { 4521 cpumask_clear_cpu(cpu_of(busiest), cpus); 4522 if (!cpumask_empty(cpus)) 4523 goto redo; 4524 goto out_balanced; 4525 } 4526 } 4527 4528 if (!ld_moved) { 4529 schedstat_inc(sd, lb_failed[idle]); 4530 /* 4531 * Increment the failure counter only on periodic balance. 4532 * We do not want newidle balance, which can be very 4533 * frequent, pollute the failure counter causing 4534 * excessive cache_hot migrations and active balances. 4535 */ 4536 if (idle != CPU_NEWLY_IDLE) 4537 sd->nr_balance_failed++; 4538 4539 if (need_active_balance(sd, idle, cpu_of(busiest), this_cpu)) { 4540 raw_spin_lock_irqsave(&busiest->lock, flags); 4541 4542 /* don't kick the active_load_balance_cpu_stop, 4543 * if the curr task on busiest cpu can't be 4544 * moved to this_cpu 4545 */ 4546 if (!cpumask_test_cpu(this_cpu, 4547 tsk_cpus_allowed(busiest->curr))) { 4548 raw_spin_unlock_irqrestore(&busiest->lock, 4549 flags); 4550 lb_flags |= LBF_ALL_PINNED; 4551 goto out_one_pinned; 4552 } 4553 4554 /* 4555 * ->active_balance synchronizes accesses to 4556 * ->active_balance_work. Once set, it's cleared 4557 * only after active load balance is finished. 4558 */ 4559 if (!busiest->active_balance) { 4560 busiest->active_balance = 1; 4561 busiest->push_cpu = this_cpu; 4562 active_balance = 1; 4563 } 4564 raw_spin_unlock_irqrestore(&busiest->lock, flags); 4565 4566 if (active_balance) 4567 stop_one_cpu_nowait(cpu_of(busiest), 4568 active_load_balance_cpu_stop, busiest, 4569 &busiest->active_balance_work); 4570 4571 /* 4572 * We've kicked active balancing, reset the failure 4573 * counter. 4574 */ 4575 sd->nr_balance_failed = sd->cache_nice_tries+1; 4576 } 4577 } else 4578 sd->nr_balance_failed = 0; 4579 4580 if (likely(!active_balance)) { 4581 /* We were unbalanced, so reset the balancing interval */ 4582 sd->balance_interval = sd->min_interval; 4583 } else { 4584 /* 4585 * If we've begun active balancing, start to back off. This 4586 * case may not be covered by the all_pinned logic if there 4587 * is only 1 task on the busy runqueue (because we don't call 4588 * move_tasks). 4589 */ 4590 if (sd->balance_interval < sd->max_interval) 4591 sd->balance_interval *= 2; 4592 } 4593 4594 goto out; 4595 4596 out_balanced: 4597 schedstat_inc(sd, lb_balanced[idle]); 4598 4599 sd->nr_balance_failed = 0; 4600 4601 out_one_pinned: 4602 /* tune up the balancing interval */ 4603 if (((lb_flags & LBF_ALL_PINNED) && 4604 sd->balance_interval < MAX_PINNED_INTERVAL) || 4605 (sd->balance_interval < sd->max_interval)) 4606 sd->balance_interval *= 2; 4607 4608 ld_moved = 0; 4609 out: 4610 return ld_moved; 4611 } 4612 4613 /* 4614 * idle_balance is called by schedule() if this_cpu is about to become 4615 * idle. Attempts to pull tasks from other CPUs. 4616 */ 4617 void idle_balance(int this_cpu, struct rq *this_rq) 4618 { 4619 struct sched_domain *sd; 4620 int pulled_task = 0; 4621 unsigned long next_balance = jiffies + HZ; 4622 4623 this_rq->idle_stamp = this_rq->clock; 4624 4625 if (this_rq->avg_idle < sysctl_sched_migration_cost) 4626 return; 4627 4628 /* 4629 * Drop the rq->lock, but keep IRQ/preempt disabled. 4630 */ 4631 raw_spin_unlock(&this_rq->lock); 4632 4633 update_shares(this_cpu); 4634 rcu_read_lock(); 4635 for_each_domain(this_cpu, sd) { 4636 unsigned long interval; 4637 int balance = 1; 4638 4639 if (!(sd->flags & SD_LOAD_BALANCE)) 4640 continue; 4641 4642 if (sd->flags & SD_BALANCE_NEWIDLE) { 4643 /* If we've pulled tasks over stop searching: */ 4644 pulled_task = load_balance(this_cpu, this_rq, 4645 sd, CPU_NEWLY_IDLE, &balance); 4646 } 4647 4648 interval = msecs_to_jiffies(sd->balance_interval); 4649 if (time_after(next_balance, sd->last_balance + interval)) 4650 next_balance = sd->last_balance + interval; 4651 if (pulled_task) { 4652 this_rq->idle_stamp = 0; 4653 break; 4654 } 4655 } 4656 rcu_read_unlock(); 4657 4658 raw_spin_lock(&this_rq->lock); 4659 4660 if (pulled_task || time_after(jiffies, this_rq->next_balance)) { 4661 /* 4662 * We are going idle. next_balance may be set based on 4663 * a busy processor. So reset next_balance. 4664 */ 4665 this_rq->next_balance = next_balance; 4666 } 4667 } 4668 4669 /* 4670 * active_load_balance_cpu_stop is run by cpu stopper. It pushes 4671 * running tasks off the busiest CPU onto idle CPUs. It requires at 4672 * least 1 task to be running on each physical CPU where possible, and 4673 * avoids physical / logical imbalances. 4674 */ 4675 static int active_load_balance_cpu_stop(void *data) 4676 { 4677 struct rq *busiest_rq = data; 4678 int busiest_cpu = cpu_of(busiest_rq); 4679 int target_cpu = busiest_rq->push_cpu; 4680 struct rq *target_rq = cpu_rq(target_cpu); 4681 struct sched_domain *sd; 4682 4683 raw_spin_lock_irq(&busiest_rq->lock); 4684 4685 /* make sure the requested cpu hasn't gone down in the meantime */ 4686 if (unlikely(busiest_cpu != smp_processor_id() || 4687 !busiest_rq->active_balance)) 4688 goto out_unlock; 4689 4690 /* Is there any task to move? */ 4691 if (busiest_rq->nr_running <= 1) 4692 goto out_unlock; 4693 4694 /* 4695 * This condition is "impossible", if it occurs 4696 * we need to fix it. Originally reported by 4697 * Bjorn Helgaas on a 128-cpu setup. 4698 */ 4699 BUG_ON(busiest_rq == target_rq); 4700 4701 /* move a task from busiest_rq to target_rq */ 4702 double_lock_balance(busiest_rq, target_rq); 4703 4704 /* Search for an sd spanning us and the target CPU. */ 4705 rcu_read_lock(); 4706 for_each_domain(target_cpu, sd) { 4707 if ((sd->flags & SD_LOAD_BALANCE) && 4708 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd))) 4709 break; 4710 } 4711 4712 if (likely(sd)) { 4713 schedstat_inc(sd, alb_count); 4714 4715 if (move_one_task(target_rq, target_cpu, busiest_rq, 4716 sd, CPU_IDLE)) 4717 schedstat_inc(sd, alb_pushed); 4718 else 4719 schedstat_inc(sd, alb_failed); 4720 } 4721 rcu_read_unlock(); 4722 double_unlock_balance(busiest_rq, target_rq); 4723 out_unlock: 4724 busiest_rq->active_balance = 0; 4725 raw_spin_unlock_irq(&busiest_rq->lock); 4726 return 0; 4727 } 4728 4729 #ifdef CONFIG_NO_HZ 4730 /* 4731 * idle load balancing details 4732 * - When one of the busy CPUs notice that there may be an idle rebalancing 4733 * needed, they will kick the idle load balancer, which then does idle 4734 * load balancing for all the idle CPUs. 4735 */ 4736 static struct { 4737 cpumask_var_t idle_cpus_mask; 4738 atomic_t nr_cpus; 4739 unsigned long next_balance; /* in jiffy units */ 4740 } nohz ____cacheline_aligned; 4741 4742 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT) 4743 /** 4744 * lowest_flag_domain - Return lowest sched_domain containing flag. 4745 * @cpu: The cpu whose lowest level of sched domain is to 4746 * be returned. 4747 * @flag: The flag to check for the lowest sched_domain 4748 * for the given cpu. 4749 * 4750 * Returns the lowest sched_domain of a cpu which contains the given flag. 4751 */ 4752 static inline struct sched_domain *lowest_flag_domain(int cpu, int flag) 4753 { 4754 struct sched_domain *sd; 4755 4756 for_each_domain(cpu, sd) 4757 if (sd->flags & flag) 4758 break; 4759 4760 return sd; 4761 } 4762 4763 /** 4764 * for_each_flag_domain - Iterates over sched_domains containing the flag. 4765 * @cpu: The cpu whose domains we're iterating over. 4766 * @sd: variable holding the value of the power_savings_sd 4767 * for cpu. 4768 * @flag: The flag to filter the sched_domains to be iterated. 4769 * 4770 * Iterates over all the scheduler domains for a given cpu that has the 'flag' 4771 * set, starting from the lowest sched_domain to the highest. 4772 */ 4773 #define for_each_flag_domain(cpu, sd, flag) \ 4774 for (sd = lowest_flag_domain(cpu, flag); \ 4775 (sd && (sd->flags & flag)); sd = sd->parent) 4776 4777 /** 4778 * find_new_ilb - Finds the optimum idle load balancer for nomination. 4779 * @cpu: The cpu which is nominating a new idle_load_balancer. 4780 * 4781 * Returns: Returns the id of the idle load balancer if it exists, 4782 * Else, returns >= nr_cpu_ids. 4783 * 4784 * This algorithm picks the idle load balancer such that it belongs to a 4785 * semi-idle powersavings sched_domain. The idea is to try and avoid 4786 * completely idle packages/cores just for the purpose of idle load balancing 4787 * when there are other idle cpu's which are better suited for that job. 4788 */ 4789 static int find_new_ilb(int cpu) 4790 { 4791 int ilb = cpumask_first(nohz.idle_cpus_mask); 4792 struct sched_group *ilbg; 4793 struct sched_domain *sd; 4794 4795 /* 4796 * Have idle load balancer selection from semi-idle packages only 4797 * when power-aware load balancing is enabled 4798 */ 4799 if (!(sched_smt_power_savings || sched_mc_power_savings)) 4800 goto out_done; 4801 4802 /* 4803 * Optimize for the case when we have no idle CPUs or only one 4804 * idle CPU. Don't walk the sched_domain hierarchy in such cases 4805 */ 4806 if (cpumask_weight(nohz.idle_cpus_mask) < 2) 4807 goto out_done; 4808 4809 rcu_read_lock(); 4810 for_each_flag_domain(cpu, sd, SD_POWERSAVINGS_BALANCE) { 4811 ilbg = sd->groups; 4812 4813 do { 4814 if (ilbg->group_weight != 4815 atomic_read(&ilbg->sgp->nr_busy_cpus)) { 4816 ilb = cpumask_first_and(nohz.idle_cpus_mask, 4817 sched_group_cpus(ilbg)); 4818 goto unlock; 4819 } 4820 4821 ilbg = ilbg->next; 4822 4823 } while (ilbg != sd->groups); 4824 } 4825 unlock: 4826 rcu_read_unlock(); 4827 4828 out_done: 4829 if (ilb < nr_cpu_ids && idle_cpu(ilb)) 4830 return ilb; 4831 4832 return nr_cpu_ids; 4833 } 4834 #else /* (CONFIG_SCHED_MC || CONFIG_SCHED_SMT) */ 4835 static inline int find_new_ilb(int call_cpu) 4836 { 4837 return nr_cpu_ids; 4838 } 4839 #endif 4840 4841 /* 4842 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the 4843 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle 4844 * CPU (if there is one). 4845 */ 4846 static void nohz_balancer_kick(int cpu) 4847 { 4848 int ilb_cpu; 4849 4850 nohz.next_balance++; 4851 4852 ilb_cpu = find_new_ilb(cpu); 4853 4854 if (ilb_cpu >= nr_cpu_ids) 4855 return; 4856 4857 if (test_and_set_bit(NOHZ_BALANCE_KICK, nohz_flags(ilb_cpu))) 4858 return; 4859 /* 4860 * Use smp_send_reschedule() instead of resched_cpu(). 4861 * This way we generate a sched IPI on the target cpu which 4862 * is idle. And the softirq performing nohz idle load balance 4863 * will be run before returning from the IPI. 4864 */ 4865 smp_send_reschedule(ilb_cpu); 4866 return; 4867 } 4868 4869 static inline void clear_nohz_tick_stopped(int cpu) 4870 { 4871 if (unlikely(test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))) { 4872 cpumask_clear_cpu(cpu, nohz.idle_cpus_mask); 4873 atomic_dec(&nohz.nr_cpus); 4874 clear_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)); 4875 } 4876 } 4877 4878 static inline void set_cpu_sd_state_busy(void) 4879 { 4880 struct sched_domain *sd; 4881 int cpu = smp_processor_id(); 4882 4883 if (!test_bit(NOHZ_IDLE, nohz_flags(cpu))) 4884 return; 4885 clear_bit(NOHZ_IDLE, nohz_flags(cpu)); 4886 4887 rcu_read_lock(); 4888 for_each_domain(cpu, sd) 4889 atomic_inc(&sd->groups->sgp->nr_busy_cpus); 4890 rcu_read_unlock(); 4891 } 4892 4893 void set_cpu_sd_state_idle(void) 4894 { 4895 struct sched_domain *sd; 4896 int cpu = smp_processor_id(); 4897 4898 if (test_bit(NOHZ_IDLE, nohz_flags(cpu))) 4899 return; 4900 set_bit(NOHZ_IDLE, nohz_flags(cpu)); 4901 4902 rcu_read_lock(); 4903 for_each_domain(cpu, sd) 4904 atomic_dec(&sd->groups->sgp->nr_busy_cpus); 4905 rcu_read_unlock(); 4906 } 4907 4908 /* 4909 * This routine will record that this cpu is going idle with tick stopped. 4910 * This info will be used in performing idle load balancing in the future. 4911 */ 4912 void select_nohz_load_balancer(int stop_tick) 4913 { 4914 int cpu = smp_processor_id(); 4915 4916 /* 4917 * If this cpu is going down, then nothing needs to be done. 4918 */ 4919 if (!cpu_active(cpu)) 4920 return; 4921 4922 if (stop_tick) { 4923 if (test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu))) 4924 return; 4925 4926 cpumask_set_cpu(cpu, nohz.idle_cpus_mask); 4927 atomic_inc(&nohz.nr_cpus); 4928 set_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)); 4929 } 4930 return; 4931 } 4932 4933 static int __cpuinit sched_ilb_notifier(struct notifier_block *nfb, 4934 unsigned long action, void *hcpu) 4935 { 4936 switch (action & ~CPU_TASKS_FROZEN) { 4937 case CPU_DYING: 4938 clear_nohz_tick_stopped(smp_processor_id()); 4939 return NOTIFY_OK; 4940 default: 4941 return NOTIFY_DONE; 4942 } 4943 } 4944 #endif 4945 4946 static DEFINE_SPINLOCK(balancing); 4947 4948 static unsigned long __read_mostly max_load_balance_interval = HZ/10; 4949 4950 /* 4951 * Scale the max load_balance interval with the number of CPUs in the system. 4952 * This trades load-balance latency on larger machines for less cross talk. 4953 */ 4954 void update_max_interval(void) 4955 { 4956 max_load_balance_interval = HZ*num_online_cpus()/10; 4957 } 4958 4959 /* 4960 * It checks each scheduling domain to see if it is due to be balanced, 4961 * and initiates a balancing operation if so. 4962 * 4963 * Balancing parameters are set up in arch_init_sched_domains. 4964 */ 4965 static void rebalance_domains(int cpu, enum cpu_idle_type idle) 4966 { 4967 int balance = 1; 4968 struct rq *rq = cpu_rq(cpu); 4969 unsigned long interval; 4970 struct sched_domain *sd; 4971 /* Earliest time when we have to do rebalance again */ 4972 unsigned long next_balance = jiffies + 60*HZ; 4973 int update_next_balance = 0; 4974 int need_serialize; 4975 4976 update_shares(cpu); 4977 4978 rcu_read_lock(); 4979 for_each_domain(cpu, sd) { 4980 if (!(sd->flags & SD_LOAD_BALANCE)) 4981 continue; 4982 4983 interval = sd->balance_interval; 4984 if (idle != CPU_IDLE) 4985 interval *= sd->busy_factor; 4986 4987 /* scale ms to jiffies */ 4988 interval = msecs_to_jiffies(interval); 4989 interval = clamp(interval, 1UL, max_load_balance_interval); 4990 4991 need_serialize = sd->flags & SD_SERIALIZE; 4992 4993 if (need_serialize) { 4994 if (!spin_trylock(&balancing)) 4995 goto out; 4996 } 4997 4998 if (time_after_eq(jiffies, sd->last_balance + interval)) { 4999 if (load_balance(cpu, rq, sd, idle, &balance)) { 5000 /* 5001 * We've pulled tasks over so either we're no 5002 * longer idle. 5003 */ 5004 idle = CPU_NOT_IDLE; 5005 } 5006 sd->last_balance = jiffies; 5007 } 5008 if (need_serialize) 5009 spin_unlock(&balancing); 5010 out: 5011 if (time_after(next_balance, sd->last_balance + interval)) { 5012 next_balance = sd->last_balance + interval; 5013 update_next_balance = 1; 5014 } 5015 5016 /* 5017 * Stop the load balance at this level. There is another 5018 * CPU in our sched group which is doing load balancing more 5019 * actively. 5020 */ 5021 if (!balance) 5022 break; 5023 } 5024 rcu_read_unlock(); 5025 5026 /* 5027 * next_balance will be updated only when there is a need. 5028 * When the cpu is attached to null domain for ex, it will not be 5029 * updated. 5030 */ 5031 if (likely(update_next_balance)) 5032 rq->next_balance = next_balance; 5033 } 5034 5035 #ifdef CONFIG_NO_HZ 5036 /* 5037 * In CONFIG_NO_HZ case, the idle balance kickee will do the 5038 * rebalancing for all the cpus for whom scheduler ticks are stopped. 5039 */ 5040 static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle) 5041 { 5042 struct rq *this_rq = cpu_rq(this_cpu); 5043 struct rq *rq; 5044 int balance_cpu; 5045 5046 if (idle != CPU_IDLE || 5047 !test_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu))) 5048 goto end; 5049 5050 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) { 5051 if (balance_cpu == this_cpu || !idle_cpu(balance_cpu)) 5052 continue; 5053 5054 /* 5055 * If this cpu gets work to do, stop the load balancing 5056 * work being done for other cpus. Next load 5057 * balancing owner will pick it up. 5058 */ 5059 if (need_resched()) 5060 break; 5061 5062 raw_spin_lock_irq(&this_rq->lock); 5063 update_rq_clock(this_rq); 5064 update_cpu_load(this_rq); 5065 raw_spin_unlock_irq(&this_rq->lock); 5066 5067 rebalance_domains(balance_cpu, CPU_IDLE); 5068 5069 rq = cpu_rq(balance_cpu); 5070 if (time_after(this_rq->next_balance, rq->next_balance)) 5071 this_rq->next_balance = rq->next_balance; 5072 } 5073 nohz.next_balance = this_rq->next_balance; 5074 end: 5075 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu)); 5076 } 5077 5078 /* 5079 * Current heuristic for kicking the idle load balancer in the presence 5080 * of an idle cpu is the system. 5081 * - This rq has more than one task. 5082 * - At any scheduler domain level, this cpu's scheduler group has multiple 5083 * busy cpu's exceeding the group's power. 5084 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler 5085 * domain span are idle. 5086 */ 5087 static inline int nohz_kick_needed(struct rq *rq, int cpu) 5088 { 5089 unsigned long now = jiffies; 5090 struct sched_domain *sd; 5091 5092 if (unlikely(idle_cpu(cpu))) 5093 return 0; 5094 5095 /* 5096 * We may be recently in ticked or tickless idle mode. At the first 5097 * busy tick after returning from idle, we will update the busy stats. 5098 */ 5099 set_cpu_sd_state_busy(); 5100 clear_nohz_tick_stopped(cpu); 5101 5102 /* 5103 * None are in tickless mode and hence no need for NOHZ idle load 5104 * balancing. 5105 */ 5106 if (likely(!atomic_read(&nohz.nr_cpus))) 5107 return 0; 5108 5109 if (time_before(now, nohz.next_balance)) 5110 return 0; 5111 5112 if (rq->nr_running >= 2) 5113 goto need_kick; 5114 5115 rcu_read_lock(); 5116 for_each_domain(cpu, sd) { 5117 struct sched_group *sg = sd->groups; 5118 struct sched_group_power *sgp = sg->sgp; 5119 int nr_busy = atomic_read(&sgp->nr_busy_cpus); 5120 5121 if (sd->flags & SD_SHARE_PKG_RESOURCES && nr_busy > 1) 5122 goto need_kick_unlock; 5123 5124 if (sd->flags & SD_ASYM_PACKING && nr_busy != sg->group_weight 5125 && (cpumask_first_and(nohz.idle_cpus_mask, 5126 sched_domain_span(sd)) < cpu)) 5127 goto need_kick_unlock; 5128 5129 if (!(sd->flags & (SD_SHARE_PKG_RESOURCES | SD_ASYM_PACKING))) 5130 break; 5131 } 5132 rcu_read_unlock(); 5133 return 0; 5134 5135 need_kick_unlock: 5136 rcu_read_unlock(); 5137 need_kick: 5138 return 1; 5139 } 5140 #else 5141 static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle) { } 5142 #endif 5143 5144 /* 5145 * run_rebalance_domains is triggered when needed from the scheduler tick. 5146 * Also triggered for nohz idle balancing (with nohz_balancing_kick set). 5147 */ 5148 static void run_rebalance_domains(struct softirq_action *h) 5149 { 5150 int this_cpu = smp_processor_id(); 5151 struct rq *this_rq = cpu_rq(this_cpu); 5152 enum cpu_idle_type idle = this_rq->idle_balance ? 5153 CPU_IDLE : CPU_NOT_IDLE; 5154 5155 rebalance_domains(this_cpu, idle); 5156 5157 /* 5158 * If this cpu has a pending nohz_balance_kick, then do the 5159 * balancing on behalf of the other idle cpus whose ticks are 5160 * stopped. 5161 */ 5162 nohz_idle_balance(this_cpu, idle); 5163 } 5164 5165 static inline int on_null_domain(int cpu) 5166 { 5167 return !rcu_dereference_sched(cpu_rq(cpu)->sd); 5168 } 5169 5170 /* 5171 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing. 5172 */ 5173 void trigger_load_balance(struct rq *rq, int cpu) 5174 { 5175 /* Don't need to rebalance while attached to NULL domain */ 5176 if (time_after_eq(jiffies, rq->next_balance) && 5177 likely(!on_null_domain(cpu))) 5178 raise_softirq(SCHED_SOFTIRQ); 5179 #ifdef CONFIG_NO_HZ 5180 if (nohz_kick_needed(rq, cpu) && likely(!on_null_domain(cpu))) 5181 nohz_balancer_kick(cpu); 5182 #endif 5183 } 5184 5185 static void rq_online_fair(struct rq *rq) 5186 { 5187 update_sysctl(); 5188 } 5189 5190 static void rq_offline_fair(struct rq *rq) 5191 { 5192 update_sysctl(); 5193 } 5194 5195 #endif /* CONFIG_SMP */ 5196 5197 /* 5198 * scheduler tick hitting a task of our scheduling class: 5199 */ 5200 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued) 5201 { 5202 struct cfs_rq *cfs_rq; 5203 struct sched_entity *se = &curr->se; 5204 5205 for_each_sched_entity(se) { 5206 cfs_rq = cfs_rq_of(se); 5207 entity_tick(cfs_rq, se, queued); 5208 } 5209 } 5210 5211 /* 5212 * called on fork with the child task as argument from the parent's context 5213 * - child not yet on the tasklist 5214 * - preemption disabled 5215 */ 5216 static void task_fork_fair(struct task_struct *p) 5217 { 5218 struct cfs_rq *cfs_rq; 5219 struct sched_entity *se = &p->se, *curr; 5220 int this_cpu = smp_processor_id(); 5221 struct rq *rq = this_rq(); 5222 unsigned long flags; 5223 5224 raw_spin_lock_irqsave(&rq->lock, flags); 5225 5226 update_rq_clock(rq); 5227 5228 cfs_rq = task_cfs_rq(current); 5229 curr = cfs_rq->curr; 5230 5231 if (unlikely(task_cpu(p) != this_cpu)) { 5232 rcu_read_lock(); 5233 __set_task_cpu(p, this_cpu); 5234 rcu_read_unlock(); 5235 } 5236 5237 update_curr(cfs_rq); 5238 5239 if (curr) 5240 se->vruntime = curr->vruntime; 5241 place_entity(cfs_rq, se, 1); 5242 5243 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) { 5244 /* 5245 * Upon rescheduling, sched_class::put_prev_task() will place 5246 * 'current' within the tree based on its new key value. 5247 */ 5248 swap(curr->vruntime, se->vruntime); 5249 resched_task(rq->curr); 5250 } 5251 5252 se->vruntime -= cfs_rq->min_vruntime; 5253 5254 raw_spin_unlock_irqrestore(&rq->lock, flags); 5255 } 5256 5257 /* 5258 * Priority of the task has changed. Check to see if we preempt 5259 * the current task. 5260 */ 5261 static void 5262 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio) 5263 { 5264 if (!p->se.on_rq) 5265 return; 5266 5267 /* 5268 * Reschedule if we are currently running on this runqueue and 5269 * our priority decreased, or if we are not currently running on 5270 * this runqueue and our priority is higher than the current's 5271 */ 5272 if (rq->curr == p) { 5273 if (p->prio > oldprio) 5274 resched_task(rq->curr); 5275 } else 5276 check_preempt_curr(rq, p, 0); 5277 } 5278 5279 static void switched_from_fair(struct rq *rq, struct task_struct *p) 5280 { 5281 struct sched_entity *se = &p->se; 5282 struct cfs_rq *cfs_rq = cfs_rq_of(se); 5283 5284 /* 5285 * Ensure the task's vruntime is normalized, so that when its 5286 * switched back to the fair class the enqueue_entity(.flags=0) will 5287 * do the right thing. 5288 * 5289 * If it was on_rq, then the dequeue_entity(.flags=0) will already 5290 * have normalized the vruntime, if it was !on_rq, then only when 5291 * the task is sleeping will it still have non-normalized vruntime. 5292 */ 5293 if (!se->on_rq && p->state != TASK_RUNNING) { 5294 /* 5295 * Fix up our vruntime so that the current sleep doesn't 5296 * cause 'unlimited' sleep bonus. 5297 */ 5298 place_entity(cfs_rq, se, 0); 5299 se->vruntime -= cfs_rq->min_vruntime; 5300 } 5301 } 5302 5303 /* 5304 * We switched to the sched_fair class. 5305 */ 5306 static void switched_to_fair(struct rq *rq, struct task_struct *p) 5307 { 5308 if (!p->se.on_rq) 5309 return; 5310 5311 /* 5312 * We were most likely switched from sched_rt, so 5313 * kick off the schedule if running, otherwise just see 5314 * if we can still preempt the current task. 5315 */ 5316 if (rq->curr == p) 5317 resched_task(rq->curr); 5318 else 5319 check_preempt_curr(rq, p, 0); 5320 } 5321 5322 /* Account for a task changing its policy or group. 5323 * 5324 * This routine is mostly called to set cfs_rq->curr field when a task 5325 * migrates between groups/classes. 5326 */ 5327 static void set_curr_task_fair(struct rq *rq) 5328 { 5329 struct sched_entity *se = &rq->curr->se; 5330 5331 for_each_sched_entity(se) { 5332 struct cfs_rq *cfs_rq = cfs_rq_of(se); 5333 5334 set_next_entity(cfs_rq, se); 5335 /* ensure bandwidth has been allocated on our new cfs_rq */ 5336 account_cfs_rq_runtime(cfs_rq, 0); 5337 } 5338 } 5339 5340 void init_cfs_rq(struct cfs_rq *cfs_rq) 5341 { 5342 cfs_rq->tasks_timeline = RB_ROOT; 5343 INIT_LIST_HEAD(&cfs_rq->tasks); 5344 cfs_rq->min_vruntime = (u64)(-(1LL << 20)); 5345 #ifndef CONFIG_64BIT 5346 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime; 5347 #endif 5348 } 5349 5350 #ifdef CONFIG_FAIR_GROUP_SCHED 5351 static void task_move_group_fair(struct task_struct *p, int on_rq) 5352 { 5353 /* 5354 * If the task was not on the rq at the time of this cgroup movement 5355 * it must have been asleep, sleeping tasks keep their ->vruntime 5356 * absolute on their old rq until wakeup (needed for the fair sleeper 5357 * bonus in place_entity()). 5358 * 5359 * If it was on the rq, we've just 'preempted' it, which does convert 5360 * ->vruntime to a relative base. 5361 * 5362 * Make sure both cases convert their relative position when migrating 5363 * to another cgroup's rq. This does somewhat interfere with the 5364 * fair sleeper stuff for the first placement, but who cares. 5365 */ 5366 /* 5367 * When !on_rq, vruntime of the task has usually NOT been normalized. 5368 * But there are some cases where it has already been normalized: 5369 * 5370 * - Moving a forked child which is waiting for being woken up by 5371 * wake_up_new_task(). 5372 * - Moving a task which has been woken up by try_to_wake_up() and 5373 * waiting for actually being woken up by sched_ttwu_pending(). 5374 * 5375 * To prevent boost or penalty in the new cfs_rq caused by delta 5376 * min_vruntime between the two cfs_rqs, we skip vruntime adjustment. 5377 */ 5378 if (!on_rq && (!p->se.sum_exec_runtime || p->state == TASK_WAKING)) 5379 on_rq = 1; 5380 5381 if (!on_rq) 5382 p->se.vruntime -= cfs_rq_of(&p->se)->min_vruntime; 5383 set_task_rq(p, task_cpu(p)); 5384 if (!on_rq) 5385 p->se.vruntime += cfs_rq_of(&p->se)->min_vruntime; 5386 } 5387 5388 void free_fair_sched_group(struct task_group *tg) 5389 { 5390 int i; 5391 5392 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg)); 5393 5394 for_each_possible_cpu(i) { 5395 if (tg->cfs_rq) 5396 kfree(tg->cfs_rq[i]); 5397 if (tg->se) 5398 kfree(tg->se[i]); 5399 } 5400 5401 kfree(tg->cfs_rq); 5402 kfree(tg->se); 5403 } 5404 5405 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent) 5406 { 5407 struct cfs_rq *cfs_rq; 5408 struct sched_entity *se; 5409 int i; 5410 5411 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL); 5412 if (!tg->cfs_rq) 5413 goto err; 5414 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL); 5415 if (!tg->se) 5416 goto err; 5417 5418 tg->shares = NICE_0_LOAD; 5419 5420 init_cfs_bandwidth(tg_cfs_bandwidth(tg)); 5421 5422 for_each_possible_cpu(i) { 5423 cfs_rq = kzalloc_node(sizeof(struct cfs_rq), 5424 GFP_KERNEL, cpu_to_node(i)); 5425 if (!cfs_rq) 5426 goto err; 5427 5428 se = kzalloc_node(sizeof(struct sched_entity), 5429 GFP_KERNEL, cpu_to_node(i)); 5430 if (!se) 5431 goto err_free_rq; 5432 5433 init_cfs_rq(cfs_rq); 5434 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]); 5435 } 5436 5437 return 1; 5438 5439 err_free_rq: 5440 kfree(cfs_rq); 5441 err: 5442 return 0; 5443 } 5444 5445 void unregister_fair_sched_group(struct task_group *tg, int cpu) 5446 { 5447 struct rq *rq = cpu_rq(cpu); 5448 unsigned long flags; 5449 5450 /* 5451 * Only empty task groups can be destroyed; so we can speculatively 5452 * check on_list without danger of it being re-added. 5453 */ 5454 if (!tg->cfs_rq[cpu]->on_list) 5455 return; 5456 5457 raw_spin_lock_irqsave(&rq->lock, flags); 5458 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]); 5459 raw_spin_unlock_irqrestore(&rq->lock, flags); 5460 } 5461 5462 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq, 5463 struct sched_entity *se, int cpu, 5464 struct sched_entity *parent) 5465 { 5466 struct rq *rq = cpu_rq(cpu); 5467 5468 cfs_rq->tg = tg; 5469 cfs_rq->rq = rq; 5470 #ifdef CONFIG_SMP 5471 /* allow initial update_cfs_load() to truncate */ 5472 cfs_rq->load_stamp = 1; 5473 #endif 5474 init_cfs_rq_runtime(cfs_rq); 5475 5476 tg->cfs_rq[cpu] = cfs_rq; 5477 tg->se[cpu] = se; 5478 5479 /* se could be NULL for root_task_group */ 5480 if (!se) 5481 return; 5482 5483 if (!parent) 5484 se->cfs_rq = &rq->cfs; 5485 else 5486 se->cfs_rq = parent->my_q; 5487 5488 se->my_q = cfs_rq; 5489 update_load_set(&se->load, 0); 5490 se->parent = parent; 5491 } 5492 5493 static DEFINE_MUTEX(shares_mutex); 5494 5495 int sched_group_set_shares(struct task_group *tg, unsigned long shares) 5496 { 5497 int i; 5498 unsigned long flags; 5499 5500 /* 5501 * We can't change the weight of the root cgroup. 5502 */ 5503 if (!tg->se[0]) 5504 return -EINVAL; 5505 5506 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES)); 5507 5508 mutex_lock(&shares_mutex); 5509 if (tg->shares == shares) 5510 goto done; 5511 5512 tg->shares = shares; 5513 for_each_possible_cpu(i) { 5514 struct rq *rq = cpu_rq(i); 5515 struct sched_entity *se; 5516 5517 se = tg->se[i]; 5518 /* Propagate contribution to hierarchy */ 5519 raw_spin_lock_irqsave(&rq->lock, flags); 5520 for_each_sched_entity(se) 5521 update_cfs_shares(group_cfs_rq(se)); 5522 raw_spin_unlock_irqrestore(&rq->lock, flags); 5523 } 5524 5525 done: 5526 mutex_unlock(&shares_mutex); 5527 return 0; 5528 } 5529 #else /* CONFIG_FAIR_GROUP_SCHED */ 5530 5531 void free_fair_sched_group(struct task_group *tg) { } 5532 5533 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent) 5534 { 5535 return 1; 5536 } 5537 5538 void unregister_fair_sched_group(struct task_group *tg, int cpu) { } 5539 5540 #endif /* CONFIG_FAIR_GROUP_SCHED */ 5541 5542 5543 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task) 5544 { 5545 struct sched_entity *se = &task->se; 5546 unsigned int rr_interval = 0; 5547 5548 /* 5549 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise 5550 * idle runqueue: 5551 */ 5552 if (rq->cfs.load.weight) 5553 rr_interval = NS_TO_JIFFIES(sched_slice(&rq->cfs, se)); 5554 5555 return rr_interval; 5556 } 5557 5558 /* 5559 * All the scheduling class methods: 5560 */ 5561 const struct sched_class fair_sched_class = { 5562 .next = &idle_sched_class, 5563 .enqueue_task = enqueue_task_fair, 5564 .dequeue_task = dequeue_task_fair, 5565 .yield_task = yield_task_fair, 5566 .yield_to_task = yield_to_task_fair, 5567 5568 .check_preempt_curr = check_preempt_wakeup, 5569 5570 .pick_next_task = pick_next_task_fair, 5571 .put_prev_task = put_prev_task_fair, 5572 5573 #ifdef CONFIG_SMP 5574 .select_task_rq = select_task_rq_fair, 5575 5576 .rq_online = rq_online_fair, 5577 .rq_offline = rq_offline_fair, 5578 5579 .task_waking = task_waking_fair, 5580 #endif 5581 5582 .set_curr_task = set_curr_task_fair, 5583 .task_tick = task_tick_fair, 5584 .task_fork = task_fork_fair, 5585 5586 .prio_changed = prio_changed_fair, 5587 .switched_from = switched_from_fair, 5588 .switched_to = switched_to_fair, 5589 5590 .get_rr_interval = get_rr_interval_fair, 5591 5592 #ifdef CONFIG_FAIR_GROUP_SCHED 5593 .task_move_group = task_move_group_fair, 5594 #endif 5595 }; 5596 5597 #ifdef CONFIG_SCHED_DEBUG 5598 void print_cfs_stats(struct seq_file *m, int cpu) 5599 { 5600 struct cfs_rq *cfs_rq; 5601 5602 rcu_read_lock(); 5603 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq) 5604 print_cfs_rq(m, cpu, cfs_rq); 5605 rcu_read_unlock(); 5606 } 5607 #endif 5608 5609 __init void init_sched_fair_class(void) 5610 { 5611 #ifdef CONFIG_SMP 5612 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains); 5613 5614 #ifdef CONFIG_NO_HZ 5615 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT); 5616 cpu_notifier(sched_ilb_notifier, 0); 5617 #endif 5618 #endif /* SMP */ 5619 5620 } 5621