1 // SPDX-License-Identifier: GPL-2.0 2 /* 3 * Completely Fair Scheduling (CFS) Class (SCHED_NORMAL/SCHED_BATCH) 4 * 5 * Copyright (C) 2007 Red Hat, Inc., Ingo Molnar <mingo@redhat.com> 6 * 7 * Interactivity improvements by Mike Galbraith 8 * (C) 2007 Mike Galbraith <efault@gmx.de> 9 * 10 * Various enhancements by Dmitry Adamushko. 11 * (C) 2007 Dmitry Adamushko <dmitry.adamushko@gmail.com> 12 * 13 * Group scheduling enhancements by Srivatsa Vaddagiri 14 * Copyright IBM Corporation, 2007 15 * Author: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com> 16 * 17 * Scaled math optimizations by Thomas Gleixner 18 * Copyright (C) 2007, Thomas Gleixner <tglx@linutronix.de> 19 * 20 * Adaptive scheduling granularity, math enhancements by Peter Zijlstra 21 * Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra 22 */ 23 #include "sched.h" 24 25 /* 26 * Targeted preemption latency for CPU-bound tasks: 27 * 28 * NOTE: this latency value is not the same as the concept of 29 * 'timeslice length' - timeslices in CFS are of variable length 30 * and have no persistent notion like in traditional, time-slice 31 * based scheduling concepts. 32 * 33 * (to see the precise effective timeslice length of your workload, 34 * run vmstat and monitor the context-switches (cs) field) 35 * 36 * (default: 6ms * (1 + ilog(ncpus)), units: nanoseconds) 37 */ 38 unsigned int sysctl_sched_latency = 6000000ULL; 39 static unsigned int normalized_sysctl_sched_latency = 6000000ULL; 40 41 /* 42 * The initial- and re-scaling of tunables is configurable 43 * 44 * Options are: 45 * 46 * SCHED_TUNABLESCALING_NONE - unscaled, always *1 47 * SCHED_TUNABLESCALING_LOG - scaled logarithmical, *1+ilog(ncpus) 48 * SCHED_TUNABLESCALING_LINEAR - scaled linear, *ncpus 49 * 50 * (default SCHED_TUNABLESCALING_LOG = *(1+ilog(ncpus)) 51 */ 52 unsigned int sysctl_sched_tunable_scaling = SCHED_TUNABLESCALING_LOG; 53 54 /* 55 * Minimal preemption granularity for CPU-bound tasks: 56 * 57 * (default: 0.75 msec * (1 + ilog(ncpus)), units: nanoseconds) 58 */ 59 unsigned int sysctl_sched_min_granularity = 750000ULL; 60 static unsigned int normalized_sysctl_sched_min_granularity = 750000ULL; 61 62 /* 63 * This value is kept at sysctl_sched_latency/sysctl_sched_min_granularity 64 */ 65 static unsigned int sched_nr_latency = 8; 66 67 /* 68 * After fork, child runs first. If set to 0 (default) then 69 * parent will (try to) run first. 70 */ 71 unsigned int sysctl_sched_child_runs_first __read_mostly; 72 73 /* 74 * SCHED_OTHER wake-up granularity. 75 * 76 * This option delays the preemption effects of decoupled workloads 77 * and reduces their over-scheduling. Synchronous workloads will still 78 * have immediate wakeup/sleep latencies. 79 * 80 * (default: 1 msec * (1 + ilog(ncpus)), units: nanoseconds) 81 */ 82 unsigned int sysctl_sched_wakeup_granularity = 1000000UL; 83 static unsigned int normalized_sysctl_sched_wakeup_granularity = 1000000UL; 84 85 const_debug unsigned int sysctl_sched_migration_cost = 500000UL; 86 87 int sched_thermal_decay_shift; 88 static int __init setup_sched_thermal_decay_shift(char *str) 89 { 90 int _shift = 0; 91 92 if (kstrtoint(str, 0, &_shift)) 93 pr_warn("Unable to set scheduler thermal pressure decay shift parameter\n"); 94 95 sched_thermal_decay_shift = clamp(_shift, 0, 10); 96 return 1; 97 } 98 __setup("sched_thermal_decay_shift=", setup_sched_thermal_decay_shift); 99 100 #ifdef CONFIG_SMP 101 /* 102 * For asym packing, by default the lower numbered CPU has higher priority. 103 */ 104 int __weak arch_asym_cpu_priority(int cpu) 105 { 106 return -cpu; 107 } 108 109 /* 110 * The margin used when comparing utilization with CPU capacity. 111 * 112 * (default: ~20%) 113 */ 114 #define fits_capacity(cap, max) ((cap) * 1280 < (max) * 1024) 115 116 /* 117 * The margin used when comparing CPU capacities. 118 * is 'cap1' noticeably greater than 'cap2' 119 * 120 * (default: ~5%) 121 */ 122 #define capacity_greater(cap1, cap2) ((cap1) * 1024 > (cap2) * 1078) 123 #endif 124 125 #ifdef CONFIG_CFS_BANDWIDTH 126 /* 127 * Amount of runtime to allocate from global (tg) to local (per-cfs_rq) pool 128 * each time a cfs_rq requests quota. 129 * 130 * Note: in the case that the slice exceeds the runtime remaining (either due 131 * to consumption or the quota being specified to be smaller than the slice) 132 * we will always only issue the remaining available time. 133 * 134 * (default: 5 msec, units: microseconds) 135 */ 136 unsigned int sysctl_sched_cfs_bandwidth_slice = 5000UL; 137 #endif 138 139 static inline void update_load_add(struct load_weight *lw, unsigned long inc) 140 { 141 lw->weight += inc; 142 lw->inv_weight = 0; 143 } 144 145 static inline void update_load_sub(struct load_weight *lw, unsigned long dec) 146 { 147 lw->weight -= dec; 148 lw->inv_weight = 0; 149 } 150 151 static inline void update_load_set(struct load_weight *lw, unsigned long w) 152 { 153 lw->weight = w; 154 lw->inv_weight = 0; 155 } 156 157 /* 158 * Increase the granularity value when there are more CPUs, 159 * because with more CPUs the 'effective latency' as visible 160 * to users decreases. But the relationship is not linear, 161 * so pick a second-best guess by going with the log2 of the 162 * number of CPUs. 163 * 164 * This idea comes from the SD scheduler of Con Kolivas: 165 */ 166 static unsigned int get_update_sysctl_factor(void) 167 { 168 unsigned int cpus = min_t(unsigned int, num_online_cpus(), 8); 169 unsigned int factor; 170 171 switch (sysctl_sched_tunable_scaling) { 172 case SCHED_TUNABLESCALING_NONE: 173 factor = 1; 174 break; 175 case SCHED_TUNABLESCALING_LINEAR: 176 factor = cpus; 177 break; 178 case SCHED_TUNABLESCALING_LOG: 179 default: 180 factor = 1 + ilog2(cpus); 181 break; 182 } 183 184 return factor; 185 } 186 187 static void update_sysctl(void) 188 { 189 unsigned int factor = get_update_sysctl_factor(); 190 191 #define SET_SYSCTL(name) \ 192 (sysctl_##name = (factor) * normalized_sysctl_##name) 193 SET_SYSCTL(sched_min_granularity); 194 SET_SYSCTL(sched_latency); 195 SET_SYSCTL(sched_wakeup_granularity); 196 #undef SET_SYSCTL 197 } 198 199 void __init sched_init_granularity(void) 200 { 201 update_sysctl(); 202 } 203 204 #define WMULT_CONST (~0U) 205 #define WMULT_SHIFT 32 206 207 static void __update_inv_weight(struct load_weight *lw) 208 { 209 unsigned long w; 210 211 if (likely(lw->inv_weight)) 212 return; 213 214 w = scale_load_down(lw->weight); 215 216 if (BITS_PER_LONG > 32 && unlikely(w >= WMULT_CONST)) 217 lw->inv_weight = 1; 218 else if (unlikely(!w)) 219 lw->inv_weight = WMULT_CONST; 220 else 221 lw->inv_weight = WMULT_CONST / w; 222 } 223 224 /* 225 * delta_exec * weight / lw.weight 226 * OR 227 * (delta_exec * (weight * lw->inv_weight)) >> WMULT_SHIFT 228 * 229 * Either weight := NICE_0_LOAD and lw \e sched_prio_to_wmult[], in which case 230 * we're guaranteed shift stays positive because inv_weight is guaranteed to 231 * fit 32 bits, and NICE_0_LOAD gives another 10 bits; therefore shift >= 22. 232 * 233 * Or, weight =< lw.weight (because lw.weight is the runqueue weight), thus 234 * weight/lw.weight <= 1, and therefore our shift will also be positive. 235 */ 236 static u64 __calc_delta(u64 delta_exec, unsigned long weight, struct load_weight *lw) 237 { 238 u64 fact = scale_load_down(weight); 239 u32 fact_hi = (u32)(fact >> 32); 240 int shift = WMULT_SHIFT; 241 int fs; 242 243 __update_inv_weight(lw); 244 245 if (unlikely(fact_hi)) { 246 fs = fls(fact_hi); 247 shift -= fs; 248 fact >>= fs; 249 } 250 251 fact = mul_u32_u32(fact, lw->inv_weight); 252 253 fact_hi = (u32)(fact >> 32); 254 if (fact_hi) { 255 fs = fls(fact_hi); 256 shift -= fs; 257 fact >>= fs; 258 } 259 260 return mul_u64_u32_shr(delta_exec, fact, shift); 261 } 262 263 264 const struct sched_class fair_sched_class; 265 266 /************************************************************** 267 * CFS operations on generic schedulable entities: 268 */ 269 270 #ifdef CONFIG_FAIR_GROUP_SCHED 271 static inline struct task_struct *task_of(struct sched_entity *se) 272 { 273 SCHED_WARN_ON(!entity_is_task(se)); 274 return container_of(se, struct task_struct, se); 275 } 276 277 /* Walk up scheduling entities hierarchy */ 278 #define for_each_sched_entity(se) \ 279 for (; se; se = se->parent) 280 281 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p) 282 { 283 return p->se.cfs_rq; 284 } 285 286 /* runqueue on which this entity is (to be) queued */ 287 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se) 288 { 289 return se->cfs_rq; 290 } 291 292 /* runqueue "owned" by this group */ 293 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp) 294 { 295 return grp->my_q; 296 } 297 298 static inline void cfs_rq_tg_path(struct cfs_rq *cfs_rq, char *path, int len) 299 { 300 if (!path) 301 return; 302 303 if (cfs_rq && task_group_is_autogroup(cfs_rq->tg)) 304 autogroup_path(cfs_rq->tg, path, len); 305 else if (cfs_rq && cfs_rq->tg->css.cgroup) 306 cgroup_path(cfs_rq->tg->css.cgroup, path, len); 307 else 308 strlcpy(path, "(null)", len); 309 } 310 311 static inline bool list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq) 312 { 313 struct rq *rq = rq_of(cfs_rq); 314 int cpu = cpu_of(rq); 315 316 if (cfs_rq->on_list) 317 return rq->tmp_alone_branch == &rq->leaf_cfs_rq_list; 318 319 cfs_rq->on_list = 1; 320 321 /* 322 * Ensure we either appear before our parent (if already 323 * enqueued) or force our parent to appear after us when it is 324 * enqueued. The fact that we always enqueue bottom-up 325 * reduces this to two cases and a special case for the root 326 * cfs_rq. Furthermore, it also means that we will always reset 327 * tmp_alone_branch either when the branch is connected 328 * to a tree or when we reach the top of the tree 329 */ 330 if (cfs_rq->tg->parent && 331 cfs_rq->tg->parent->cfs_rq[cpu]->on_list) { 332 /* 333 * If parent is already on the list, we add the child 334 * just before. Thanks to circular linked property of 335 * the list, this means to put the child at the tail 336 * of the list that starts by parent. 337 */ 338 list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list, 339 &(cfs_rq->tg->parent->cfs_rq[cpu]->leaf_cfs_rq_list)); 340 /* 341 * The branch is now connected to its tree so we can 342 * reset tmp_alone_branch to the beginning of the 343 * list. 344 */ 345 rq->tmp_alone_branch = &rq->leaf_cfs_rq_list; 346 return true; 347 } 348 349 if (!cfs_rq->tg->parent) { 350 /* 351 * cfs rq without parent should be put 352 * at the tail of the list. 353 */ 354 list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list, 355 &rq->leaf_cfs_rq_list); 356 /* 357 * We have reach the top of a tree so we can reset 358 * tmp_alone_branch to the beginning of the list. 359 */ 360 rq->tmp_alone_branch = &rq->leaf_cfs_rq_list; 361 return true; 362 } 363 364 /* 365 * The parent has not already been added so we want to 366 * make sure that it will be put after us. 367 * tmp_alone_branch points to the begin of the branch 368 * where we will add parent. 369 */ 370 list_add_rcu(&cfs_rq->leaf_cfs_rq_list, rq->tmp_alone_branch); 371 /* 372 * update tmp_alone_branch to points to the new begin 373 * of the branch 374 */ 375 rq->tmp_alone_branch = &cfs_rq->leaf_cfs_rq_list; 376 return false; 377 } 378 379 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq) 380 { 381 if (cfs_rq->on_list) { 382 struct rq *rq = rq_of(cfs_rq); 383 384 /* 385 * With cfs_rq being unthrottled/throttled during an enqueue, 386 * it can happen the tmp_alone_branch points the a leaf that 387 * we finally want to del. In this case, tmp_alone_branch moves 388 * to the prev element but it will point to rq->leaf_cfs_rq_list 389 * at the end of the enqueue. 390 */ 391 if (rq->tmp_alone_branch == &cfs_rq->leaf_cfs_rq_list) 392 rq->tmp_alone_branch = cfs_rq->leaf_cfs_rq_list.prev; 393 394 list_del_rcu(&cfs_rq->leaf_cfs_rq_list); 395 cfs_rq->on_list = 0; 396 } 397 } 398 399 static inline void assert_list_leaf_cfs_rq(struct rq *rq) 400 { 401 SCHED_WARN_ON(rq->tmp_alone_branch != &rq->leaf_cfs_rq_list); 402 } 403 404 /* Iterate thr' all leaf cfs_rq's on a runqueue */ 405 #define for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos) \ 406 list_for_each_entry_safe(cfs_rq, pos, &rq->leaf_cfs_rq_list, \ 407 leaf_cfs_rq_list) 408 409 /* Do the two (enqueued) entities belong to the same group ? */ 410 static inline struct cfs_rq * 411 is_same_group(struct sched_entity *se, struct sched_entity *pse) 412 { 413 if (se->cfs_rq == pse->cfs_rq) 414 return se->cfs_rq; 415 416 return NULL; 417 } 418 419 static inline struct sched_entity *parent_entity(struct sched_entity *se) 420 { 421 return se->parent; 422 } 423 424 static void 425 find_matching_se(struct sched_entity **se, struct sched_entity **pse) 426 { 427 int se_depth, pse_depth; 428 429 /* 430 * preemption test can be made between sibling entities who are in the 431 * same cfs_rq i.e who have a common parent. Walk up the hierarchy of 432 * both tasks until we find their ancestors who are siblings of common 433 * parent. 434 */ 435 436 /* First walk up until both entities are at same depth */ 437 se_depth = (*se)->depth; 438 pse_depth = (*pse)->depth; 439 440 while (se_depth > pse_depth) { 441 se_depth--; 442 *se = parent_entity(*se); 443 } 444 445 while (pse_depth > se_depth) { 446 pse_depth--; 447 *pse = parent_entity(*pse); 448 } 449 450 while (!is_same_group(*se, *pse)) { 451 *se = parent_entity(*se); 452 *pse = parent_entity(*pse); 453 } 454 } 455 456 #else /* !CONFIG_FAIR_GROUP_SCHED */ 457 458 static inline struct task_struct *task_of(struct sched_entity *se) 459 { 460 return container_of(se, struct task_struct, se); 461 } 462 463 #define for_each_sched_entity(se) \ 464 for (; se; se = NULL) 465 466 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p) 467 { 468 return &task_rq(p)->cfs; 469 } 470 471 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se) 472 { 473 struct task_struct *p = task_of(se); 474 struct rq *rq = task_rq(p); 475 476 return &rq->cfs; 477 } 478 479 /* runqueue "owned" by this group */ 480 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp) 481 { 482 return NULL; 483 } 484 485 static inline void cfs_rq_tg_path(struct cfs_rq *cfs_rq, char *path, int len) 486 { 487 if (path) 488 strlcpy(path, "(null)", len); 489 } 490 491 static inline bool list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq) 492 { 493 return true; 494 } 495 496 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq) 497 { 498 } 499 500 static inline void assert_list_leaf_cfs_rq(struct rq *rq) 501 { 502 } 503 504 #define for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos) \ 505 for (cfs_rq = &rq->cfs, pos = NULL; cfs_rq; cfs_rq = pos) 506 507 static inline struct sched_entity *parent_entity(struct sched_entity *se) 508 { 509 return NULL; 510 } 511 512 static inline void 513 find_matching_se(struct sched_entity **se, struct sched_entity **pse) 514 { 515 } 516 517 #endif /* CONFIG_FAIR_GROUP_SCHED */ 518 519 static __always_inline 520 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec); 521 522 /************************************************************** 523 * Scheduling class tree data structure manipulation methods: 524 */ 525 526 static inline u64 max_vruntime(u64 max_vruntime, u64 vruntime) 527 { 528 s64 delta = (s64)(vruntime - max_vruntime); 529 if (delta > 0) 530 max_vruntime = vruntime; 531 532 return max_vruntime; 533 } 534 535 static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime) 536 { 537 s64 delta = (s64)(vruntime - min_vruntime); 538 if (delta < 0) 539 min_vruntime = vruntime; 540 541 return min_vruntime; 542 } 543 544 static inline bool entity_before(struct sched_entity *a, 545 struct sched_entity *b) 546 { 547 return (s64)(a->vruntime - b->vruntime) < 0; 548 } 549 550 #define __node_2_se(node) \ 551 rb_entry((node), struct sched_entity, run_node) 552 553 static void update_min_vruntime(struct cfs_rq *cfs_rq) 554 { 555 struct sched_entity *curr = cfs_rq->curr; 556 struct rb_node *leftmost = rb_first_cached(&cfs_rq->tasks_timeline); 557 558 u64 vruntime = cfs_rq->min_vruntime; 559 560 if (curr) { 561 if (curr->on_rq) 562 vruntime = curr->vruntime; 563 else 564 curr = NULL; 565 } 566 567 if (leftmost) { /* non-empty tree */ 568 struct sched_entity *se = __node_2_se(leftmost); 569 570 if (!curr) 571 vruntime = se->vruntime; 572 else 573 vruntime = min_vruntime(vruntime, se->vruntime); 574 } 575 576 /* ensure we never gain time by being placed backwards. */ 577 cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime); 578 #ifndef CONFIG_64BIT 579 smp_wmb(); 580 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime; 581 #endif 582 } 583 584 static inline bool __entity_less(struct rb_node *a, const struct rb_node *b) 585 { 586 return entity_before(__node_2_se(a), __node_2_se(b)); 587 } 588 589 /* 590 * Enqueue an entity into the rb-tree: 591 */ 592 static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se) 593 { 594 rb_add_cached(&se->run_node, &cfs_rq->tasks_timeline, __entity_less); 595 } 596 597 static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se) 598 { 599 rb_erase_cached(&se->run_node, &cfs_rq->tasks_timeline); 600 } 601 602 struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq) 603 { 604 struct rb_node *left = rb_first_cached(&cfs_rq->tasks_timeline); 605 606 if (!left) 607 return NULL; 608 609 return __node_2_se(left); 610 } 611 612 static struct sched_entity *__pick_next_entity(struct sched_entity *se) 613 { 614 struct rb_node *next = rb_next(&se->run_node); 615 616 if (!next) 617 return NULL; 618 619 return __node_2_se(next); 620 } 621 622 #ifdef CONFIG_SCHED_DEBUG 623 struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq) 624 { 625 struct rb_node *last = rb_last(&cfs_rq->tasks_timeline.rb_root); 626 627 if (!last) 628 return NULL; 629 630 return __node_2_se(last); 631 } 632 633 /************************************************************** 634 * Scheduling class statistics methods: 635 */ 636 637 int sched_update_scaling(void) 638 { 639 unsigned int factor = get_update_sysctl_factor(); 640 641 sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency, 642 sysctl_sched_min_granularity); 643 644 #define WRT_SYSCTL(name) \ 645 (normalized_sysctl_##name = sysctl_##name / (factor)) 646 WRT_SYSCTL(sched_min_granularity); 647 WRT_SYSCTL(sched_latency); 648 WRT_SYSCTL(sched_wakeup_granularity); 649 #undef WRT_SYSCTL 650 651 return 0; 652 } 653 #endif 654 655 /* 656 * delta /= w 657 */ 658 static inline u64 calc_delta_fair(u64 delta, struct sched_entity *se) 659 { 660 if (unlikely(se->load.weight != NICE_0_LOAD)) 661 delta = __calc_delta(delta, NICE_0_LOAD, &se->load); 662 663 return delta; 664 } 665 666 /* 667 * The idea is to set a period in which each task runs once. 668 * 669 * When there are too many tasks (sched_nr_latency) we have to stretch 670 * this period because otherwise the slices get too small. 671 * 672 * p = (nr <= nl) ? l : l*nr/nl 673 */ 674 static u64 __sched_period(unsigned long nr_running) 675 { 676 if (unlikely(nr_running > sched_nr_latency)) 677 return nr_running * sysctl_sched_min_granularity; 678 else 679 return sysctl_sched_latency; 680 } 681 682 /* 683 * We calculate the wall-time slice from the period by taking a part 684 * proportional to the weight. 685 * 686 * s = p*P[w/rw] 687 */ 688 static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se) 689 { 690 unsigned int nr_running = cfs_rq->nr_running; 691 u64 slice; 692 693 if (sched_feat(ALT_PERIOD)) 694 nr_running = rq_of(cfs_rq)->cfs.h_nr_running; 695 696 slice = __sched_period(nr_running + !se->on_rq); 697 698 for_each_sched_entity(se) { 699 struct load_weight *load; 700 struct load_weight lw; 701 702 cfs_rq = cfs_rq_of(se); 703 load = &cfs_rq->load; 704 705 if (unlikely(!se->on_rq)) { 706 lw = cfs_rq->load; 707 708 update_load_add(&lw, se->load.weight); 709 load = &lw; 710 } 711 slice = __calc_delta(slice, se->load.weight, load); 712 } 713 714 if (sched_feat(BASE_SLICE)) 715 slice = max(slice, (u64)sysctl_sched_min_granularity); 716 717 return slice; 718 } 719 720 /* 721 * We calculate the vruntime slice of a to-be-inserted task. 722 * 723 * vs = s/w 724 */ 725 static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se) 726 { 727 return calc_delta_fair(sched_slice(cfs_rq, se), se); 728 } 729 730 #include "pelt.h" 731 #ifdef CONFIG_SMP 732 733 static int select_idle_sibling(struct task_struct *p, int prev_cpu, int cpu); 734 static unsigned long task_h_load(struct task_struct *p); 735 static unsigned long capacity_of(int cpu); 736 737 /* Give new sched_entity start runnable values to heavy its load in infant time */ 738 void init_entity_runnable_average(struct sched_entity *se) 739 { 740 struct sched_avg *sa = &se->avg; 741 742 memset(sa, 0, sizeof(*sa)); 743 744 /* 745 * Tasks are initialized with full load to be seen as heavy tasks until 746 * they get a chance to stabilize to their real load level. 747 * Group entities are initialized with zero load to reflect the fact that 748 * nothing has been attached to the task group yet. 749 */ 750 if (entity_is_task(se)) 751 sa->load_avg = scale_load_down(se->load.weight); 752 753 /* when this task enqueue'ed, it will contribute to its cfs_rq's load_avg */ 754 } 755 756 static void attach_entity_cfs_rq(struct sched_entity *se); 757 758 /* 759 * With new tasks being created, their initial util_avgs are extrapolated 760 * based on the cfs_rq's current util_avg: 761 * 762 * util_avg = cfs_rq->util_avg / (cfs_rq->load_avg + 1) * se.load.weight 763 * 764 * However, in many cases, the above util_avg does not give a desired 765 * value. Moreover, the sum of the util_avgs may be divergent, such 766 * as when the series is a harmonic series. 767 * 768 * To solve this problem, we also cap the util_avg of successive tasks to 769 * only 1/2 of the left utilization budget: 770 * 771 * util_avg_cap = (cpu_scale - cfs_rq->avg.util_avg) / 2^n 772 * 773 * where n denotes the nth task and cpu_scale the CPU capacity. 774 * 775 * For example, for a CPU with 1024 of capacity, a simplest series from 776 * the beginning would be like: 777 * 778 * task util_avg: 512, 256, 128, 64, 32, 16, 8, ... 779 * cfs_rq util_avg: 512, 768, 896, 960, 992, 1008, 1016, ... 780 * 781 * Finally, that extrapolated util_avg is clamped to the cap (util_avg_cap) 782 * if util_avg > util_avg_cap. 783 */ 784 void post_init_entity_util_avg(struct task_struct *p) 785 { 786 struct sched_entity *se = &p->se; 787 struct cfs_rq *cfs_rq = cfs_rq_of(se); 788 struct sched_avg *sa = &se->avg; 789 long cpu_scale = arch_scale_cpu_capacity(cpu_of(rq_of(cfs_rq))); 790 long cap = (long)(cpu_scale - cfs_rq->avg.util_avg) / 2; 791 792 if (cap > 0) { 793 if (cfs_rq->avg.util_avg != 0) { 794 sa->util_avg = cfs_rq->avg.util_avg * se->load.weight; 795 sa->util_avg /= (cfs_rq->avg.load_avg + 1); 796 797 if (sa->util_avg > cap) 798 sa->util_avg = cap; 799 } else { 800 sa->util_avg = cap; 801 } 802 } 803 804 sa->runnable_avg = sa->util_avg; 805 806 if (p->sched_class != &fair_sched_class) { 807 /* 808 * For !fair tasks do: 809 * 810 update_cfs_rq_load_avg(now, cfs_rq); 811 attach_entity_load_avg(cfs_rq, se); 812 switched_from_fair(rq, p); 813 * 814 * such that the next switched_to_fair() has the 815 * expected state. 816 */ 817 se->avg.last_update_time = cfs_rq_clock_pelt(cfs_rq); 818 return; 819 } 820 821 attach_entity_cfs_rq(se); 822 } 823 824 #else /* !CONFIG_SMP */ 825 void init_entity_runnable_average(struct sched_entity *se) 826 { 827 } 828 void post_init_entity_util_avg(struct task_struct *p) 829 { 830 } 831 static void update_tg_load_avg(struct cfs_rq *cfs_rq) 832 { 833 } 834 #endif /* CONFIG_SMP */ 835 836 /* 837 * Update the current task's runtime statistics. 838 */ 839 static void update_curr(struct cfs_rq *cfs_rq) 840 { 841 struct sched_entity *curr = cfs_rq->curr; 842 u64 now = rq_clock_task(rq_of(cfs_rq)); 843 u64 delta_exec; 844 845 if (unlikely(!curr)) 846 return; 847 848 delta_exec = now - curr->exec_start; 849 if (unlikely((s64)delta_exec <= 0)) 850 return; 851 852 curr->exec_start = now; 853 854 schedstat_set(curr->statistics.exec_max, 855 max(delta_exec, curr->statistics.exec_max)); 856 857 curr->sum_exec_runtime += delta_exec; 858 schedstat_add(cfs_rq->exec_clock, delta_exec); 859 860 curr->vruntime += calc_delta_fair(delta_exec, curr); 861 update_min_vruntime(cfs_rq); 862 863 if (entity_is_task(curr)) { 864 struct task_struct *curtask = task_of(curr); 865 866 trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime); 867 cgroup_account_cputime(curtask, delta_exec); 868 account_group_exec_runtime(curtask, delta_exec); 869 } 870 871 account_cfs_rq_runtime(cfs_rq, delta_exec); 872 } 873 874 static void update_curr_fair(struct rq *rq) 875 { 876 update_curr(cfs_rq_of(&rq->curr->se)); 877 } 878 879 static inline void 880 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se) 881 { 882 u64 wait_start, prev_wait_start; 883 884 if (!schedstat_enabled()) 885 return; 886 887 wait_start = rq_clock(rq_of(cfs_rq)); 888 prev_wait_start = schedstat_val(se->statistics.wait_start); 889 890 if (entity_is_task(se) && task_on_rq_migrating(task_of(se)) && 891 likely(wait_start > prev_wait_start)) 892 wait_start -= prev_wait_start; 893 894 __schedstat_set(se->statistics.wait_start, wait_start); 895 } 896 897 static inline void 898 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se) 899 { 900 struct task_struct *p; 901 u64 delta; 902 903 if (!schedstat_enabled()) 904 return; 905 906 /* 907 * When the sched_schedstat changes from 0 to 1, some sched se 908 * maybe already in the runqueue, the se->statistics.wait_start 909 * will be 0.So it will let the delta wrong. We need to avoid this 910 * scenario. 911 */ 912 if (unlikely(!schedstat_val(se->statistics.wait_start))) 913 return; 914 915 delta = rq_clock(rq_of(cfs_rq)) - schedstat_val(se->statistics.wait_start); 916 917 if (entity_is_task(se)) { 918 p = task_of(se); 919 if (task_on_rq_migrating(p)) { 920 /* 921 * Preserve migrating task's wait time so wait_start 922 * time stamp can be adjusted to accumulate wait time 923 * prior to migration. 924 */ 925 __schedstat_set(se->statistics.wait_start, delta); 926 return; 927 } 928 trace_sched_stat_wait(p, delta); 929 } 930 931 __schedstat_set(se->statistics.wait_max, 932 max(schedstat_val(se->statistics.wait_max), delta)); 933 __schedstat_inc(se->statistics.wait_count); 934 __schedstat_add(se->statistics.wait_sum, delta); 935 __schedstat_set(se->statistics.wait_start, 0); 936 } 937 938 static inline void 939 update_stats_enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se) 940 { 941 struct task_struct *tsk = NULL; 942 u64 sleep_start, block_start; 943 944 if (!schedstat_enabled()) 945 return; 946 947 sleep_start = schedstat_val(se->statistics.sleep_start); 948 block_start = schedstat_val(se->statistics.block_start); 949 950 if (entity_is_task(se)) 951 tsk = task_of(se); 952 953 if (sleep_start) { 954 u64 delta = rq_clock(rq_of(cfs_rq)) - sleep_start; 955 956 if ((s64)delta < 0) 957 delta = 0; 958 959 if (unlikely(delta > schedstat_val(se->statistics.sleep_max))) 960 __schedstat_set(se->statistics.sleep_max, delta); 961 962 __schedstat_set(se->statistics.sleep_start, 0); 963 __schedstat_add(se->statistics.sum_sleep_runtime, delta); 964 965 if (tsk) { 966 account_scheduler_latency(tsk, delta >> 10, 1); 967 trace_sched_stat_sleep(tsk, delta); 968 } 969 } 970 if (block_start) { 971 u64 delta = rq_clock(rq_of(cfs_rq)) - block_start; 972 973 if ((s64)delta < 0) 974 delta = 0; 975 976 if (unlikely(delta > schedstat_val(se->statistics.block_max))) 977 __schedstat_set(se->statistics.block_max, delta); 978 979 __schedstat_set(se->statistics.block_start, 0); 980 __schedstat_add(se->statistics.sum_sleep_runtime, delta); 981 982 if (tsk) { 983 if (tsk->in_iowait) { 984 __schedstat_add(se->statistics.iowait_sum, delta); 985 __schedstat_inc(se->statistics.iowait_count); 986 trace_sched_stat_iowait(tsk, delta); 987 } 988 989 trace_sched_stat_blocked(tsk, delta); 990 991 /* 992 * Blocking time is in units of nanosecs, so shift by 993 * 20 to get a milliseconds-range estimation of the 994 * amount of time that the task spent sleeping: 995 */ 996 if (unlikely(prof_on == SLEEP_PROFILING)) { 997 profile_hits(SLEEP_PROFILING, 998 (void *)get_wchan(tsk), 999 delta >> 20); 1000 } 1001 account_scheduler_latency(tsk, delta >> 10, 0); 1002 } 1003 } 1004 } 1005 1006 /* 1007 * Task is being enqueued - update stats: 1008 */ 1009 static inline void 1010 update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags) 1011 { 1012 if (!schedstat_enabled()) 1013 return; 1014 1015 /* 1016 * Are we enqueueing a waiting task? (for current tasks 1017 * a dequeue/enqueue event is a NOP) 1018 */ 1019 if (se != cfs_rq->curr) 1020 update_stats_wait_start(cfs_rq, se); 1021 1022 if (flags & ENQUEUE_WAKEUP) 1023 update_stats_enqueue_sleeper(cfs_rq, se); 1024 } 1025 1026 static inline void 1027 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags) 1028 { 1029 1030 if (!schedstat_enabled()) 1031 return; 1032 1033 /* 1034 * Mark the end of the wait period if dequeueing a 1035 * waiting task: 1036 */ 1037 if (se != cfs_rq->curr) 1038 update_stats_wait_end(cfs_rq, se); 1039 1040 if ((flags & DEQUEUE_SLEEP) && entity_is_task(se)) { 1041 struct task_struct *tsk = task_of(se); 1042 1043 if (tsk->state & TASK_INTERRUPTIBLE) 1044 __schedstat_set(se->statistics.sleep_start, 1045 rq_clock(rq_of(cfs_rq))); 1046 if (tsk->state & TASK_UNINTERRUPTIBLE) 1047 __schedstat_set(se->statistics.block_start, 1048 rq_clock(rq_of(cfs_rq))); 1049 } 1050 } 1051 1052 /* 1053 * We are picking a new current task - update its stats: 1054 */ 1055 static inline void 1056 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se) 1057 { 1058 /* 1059 * We are starting a new run period: 1060 */ 1061 se->exec_start = rq_clock_task(rq_of(cfs_rq)); 1062 } 1063 1064 /************************************************** 1065 * Scheduling class queueing methods: 1066 */ 1067 1068 #ifdef CONFIG_NUMA_BALANCING 1069 /* 1070 * Approximate time to scan a full NUMA task in ms. The task scan period is 1071 * calculated based on the tasks virtual memory size and 1072 * numa_balancing_scan_size. 1073 */ 1074 unsigned int sysctl_numa_balancing_scan_period_min = 1000; 1075 unsigned int sysctl_numa_balancing_scan_period_max = 60000; 1076 1077 /* Portion of address space to scan in MB */ 1078 unsigned int sysctl_numa_balancing_scan_size = 256; 1079 1080 /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */ 1081 unsigned int sysctl_numa_balancing_scan_delay = 1000; 1082 1083 struct numa_group { 1084 refcount_t refcount; 1085 1086 spinlock_t lock; /* nr_tasks, tasks */ 1087 int nr_tasks; 1088 pid_t gid; 1089 int active_nodes; 1090 1091 struct rcu_head rcu; 1092 unsigned long total_faults; 1093 unsigned long max_faults_cpu; 1094 /* 1095 * Faults_cpu is used to decide whether memory should move 1096 * towards the CPU. As a consequence, these stats are weighted 1097 * more by CPU use than by memory faults. 1098 */ 1099 unsigned long *faults_cpu; 1100 unsigned long faults[]; 1101 }; 1102 1103 /* 1104 * For functions that can be called in multiple contexts that permit reading 1105 * ->numa_group (see struct task_struct for locking rules). 1106 */ 1107 static struct numa_group *deref_task_numa_group(struct task_struct *p) 1108 { 1109 return rcu_dereference_check(p->numa_group, p == current || 1110 (lockdep_is_held(&task_rq(p)->lock) && !READ_ONCE(p->on_cpu))); 1111 } 1112 1113 static struct numa_group *deref_curr_numa_group(struct task_struct *p) 1114 { 1115 return rcu_dereference_protected(p->numa_group, p == current); 1116 } 1117 1118 static inline unsigned long group_faults_priv(struct numa_group *ng); 1119 static inline unsigned long group_faults_shared(struct numa_group *ng); 1120 1121 static unsigned int task_nr_scan_windows(struct task_struct *p) 1122 { 1123 unsigned long rss = 0; 1124 unsigned long nr_scan_pages; 1125 1126 /* 1127 * Calculations based on RSS as non-present and empty pages are skipped 1128 * by the PTE scanner and NUMA hinting faults should be trapped based 1129 * on resident pages 1130 */ 1131 nr_scan_pages = sysctl_numa_balancing_scan_size << (20 - PAGE_SHIFT); 1132 rss = get_mm_rss(p->mm); 1133 if (!rss) 1134 rss = nr_scan_pages; 1135 1136 rss = round_up(rss, nr_scan_pages); 1137 return rss / nr_scan_pages; 1138 } 1139 1140 /* For sanity's sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */ 1141 #define MAX_SCAN_WINDOW 2560 1142 1143 static unsigned int task_scan_min(struct task_struct *p) 1144 { 1145 unsigned int scan_size = READ_ONCE(sysctl_numa_balancing_scan_size); 1146 unsigned int scan, floor; 1147 unsigned int windows = 1; 1148 1149 if (scan_size < MAX_SCAN_WINDOW) 1150 windows = MAX_SCAN_WINDOW / scan_size; 1151 floor = 1000 / windows; 1152 1153 scan = sysctl_numa_balancing_scan_period_min / task_nr_scan_windows(p); 1154 return max_t(unsigned int, floor, scan); 1155 } 1156 1157 static unsigned int task_scan_start(struct task_struct *p) 1158 { 1159 unsigned long smin = task_scan_min(p); 1160 unsigned long period = smin; 1161 struct numa_group *ng; 1162 1163 /* Scale the maximum scan period with the amount of shared memory. */ 1164 rcu_read_lock(); 1165 ng = rcu_dereference(p->numa_group); 1166 if (ng) { 1167 unsigned long shared = group_faults_shared(ng); 1168 unsigned long private = group_faults_priv(ng); 1169 1170 period *= refcount_read(&ng->refcount); 1171 period *= shared + 1; 1172 period /= private + shared + 1; 1173 } 1174 rcu_read_unlock(); 1175 1176 return max(smin, period); 1177 } 1178 1179 static unsigned int task_scan_max(struct task_struct *p) 1180 { 1181 unsigned long smin = task_scan_min(p); 1182 unsigned long smax; 1183 struct numa_group *ng; 1184 1185 /* Watch for min being lower than max due to floor calculations */ 1186 smax = sysctl_numa_balancing_scan_period_max / task_nr_scan_windows(p); 1187 1188 /* Scale the maximum scan period with the amount of shared memory. */ 1189 ng = deref_curr_numa_group(p); 1190 if (ng) { 1191 unsigned long shared = group_faults_shared(ng); 1192 unsigned long private = group_faults_priv(ng); 1193 unsigned long period = smax; 1194 1195 period *= refcount_read(&ng->refcount); 1196 period *= shared + 1; 1197 period /= private + shared + 1; 1198 1199 smax = max(smax, period); 1200 } 1201 1202 return max(smin, smax); 1203 } 1204 1205 static void account_numa_enqueue(struct rq *rq, struct task_struct *p) 1206 { 1207 rq->nr_numa_running += (p->numa_preferred_nid != NUMA_NO_NODE); 1208 rq->nr_preferred_running += (p->numa_preferred_nid == task_node(p)); 1209 } 1210 1211 static void account_numa_dequeue(struct rq *rq, struct task_struct *p) 1212 { 1213 rq->nr_numa_running -= (p->numa_preferred_nid != NUMA_NO_NODE); 1214 rq->nr_preferred_running -= (p->numa_preferred_nid == task_node(p)); 1215 } 1216 1217 /* Shared or private faults. */ 1218 #define NR_NUMA_HINT_FAULT_TYPES 2 1219 1220 /* Memory and CPU locality */ 1221 #define NR_NUMA_HINT_FAULT_STATS (NR_NUMA_HINT_FAULT_TYPES * 2) 1222 1223 /* Averaged statistics, and temporary buffers. */ 1224 #define NR_NUMA_HINT_FAULT_BUCKETS (NR_NUMA_HINT_FAULT_STATS * 2) 1225 1226 pid_t task_numa_group_id(struct task_struct *p) 1227 { 1228 struct numa_group *ng; 1229 pid_t gid = 0; 1230 1231 rcu_read_lock(); 1232 ng = rcu_dereference(p->numa_group); 1233 if (ng) 1234 gid = ng->gid; 1235 rcu_read_unlock(); 1236 1237 return gid; 1238 } 1239 1240 /* 1241 * The averaged statistics, shared & private, memory & CPU, 1242 * occupy the first half of the array. The second half of the 1243 * array is for current counters, which are averaged into the 1244 * first set by task_numa_placement. 1245 */ 1246 static inline int task_faults_idx(enum numa_faults_stats s, int nid, int priv) 1247 { 1248 return NR_NUMA_HINT_FAULT_TYPES * (s * nr_node_ids + nid) + priv; 1249 } 1250 1251 static inline unsigned long task_faults(struct task_struct *p, int nid) 1252 { 1253 if (!p->numa_faults) 1254 return 0; 1255 1256 return p->numa_faults[task_faults_idx(NUMA_MEM, nid, 0)] + 1257 p->numa_faults[task_faults_idx(NUMA_MEM, nid, 1)]; 1258 } 1259 1260 static inline unsigned long group_faults(struct task_struct *p, int nid) 1261 { 1262 struct numa_group *ng = deref_task_numa_group(p); 1263 1264 if (!ng) 1265 return 0; 1266 1267 return ng->faults[task_faults_idx(NUMA_MEM, nid, 0)] + 1268 ng->faults[task_faults_idx(NUMA_MEM, nid, 1)]; 1269 } 1270 1271 static inline unsigned long group_faults_cpu(struct numa_group *group, int nid) 1272 { 1273 return group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 0)] + 1274 group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 1)]; 1275 } 1276 1277 static inline unsigned long group_faults_priv(struct numa_group *ng) 1278 { 1279 unsigned long faults = 0; 1280 int node; 1281 1282 for_each_online_node(node) { 1283 faults += ng->faults[task_faults_idx(NUMA_MEM, node, 1)]; 1284 } 1285 1286 return faults; 1287 } 1288 1289 static inline unsigned long group_faults_shared(struct numa_group *ng) 1290 { 1291 unsigned long faults = 0; 1292 int node; 1293 1294 for_each_online_node(node) { 1295 faults += ng->faults[task_faults_idx(NUMA_MEM, node, 0)]; 1296 } 1297 1298 return faults; 1299 } 1300 1301 /* 1302 * A node triggering more than 1/3 as many NUMA faults as the maximum is 1303 * considered part of a numa group's pseudo-interleaving set. Migrations 1304 * between these nodes are slowed down, to allow things to settle down. 1305 */ 1306 #define ACTIVE_NODE_FRACTION 3 1307 1308 static bool numa_is_active_node(int nid, struct numa_group *ng) 1309 { 1310 return group_faults_cpu(ng, nid) * ACTIVE_NODE_FRACTION > ng->max_faults_cpu; 1311 } 1312 1313 /* Handle placement on systems where not all nodes are directly connected. */ 1314 static unsigned long score_nearby_nodes(struct task_struct *p, int nid, 1315 int maxdist, bool task) 1316 { 1317 unsigned long score = 0; 1318 int node; 1319 1320 /* 1321 * All nodes are directly connected, and the same distance 1322 * from each other. No need for fancy placement algorithms. 1323 */ 1324 if (sched_numa_topology_type == NUMA_DIRECT) 1325 return 0; 1326 1327 /* 1328 * This code is called for each node, introducing N^2 complexity, 1329 * which should be ok given the number of nodes rarely exceeds 8. 1330 */ 1331 for_each_online_node(node) { 1332 unsigned long faults; 1333 int dist = node_distance(nid, node); 1334 1335 /* 1336 * The furthest away nodes in the system are not interesting 1337 * for placement; nid was already counted. 1338 */ 1339 if (dist == sched_max_numa_distance || node == nid) 1340 continue; 1341 1342 /* 1343 * On systems with a backplane NUMA topology, compare groups 1344 * of nodes, and move tasks towards the group with the most 1345 * memory accesses. When comparing two nodes at distance 1346 * "hoplimit", only nodes closer by than "hoplimit" are part 1347 * of each group. Skip other nodes. 1348 */ 1349 if (sched_numa_topology_type == NUMA_BACKPLANE && 1350 dist >= maxdist) 1351 continue; 1352 1353 /* Add up the faults from nearby nodes. */ 1354 if (task) 1355 faults = task_faults(p, node); 1356 else 1357 faults = group_faults(p, node); 1358 1359 /* 1360 * On systems with a glueless mesh NUMA topology, there are 1361 * no fixed "groups of nodes". Instead, nodes that are not 1362 * directly connected bounce traffic through intermediate 1363 * nodes; a numa_group can occupy any set of nodes. 1364 * The further away a node is, the less the faults count. 1365 * This seems to result in good task placement. 1366 */ 1367 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) { 1368 faults *= (sched_max_numa_distance - dist); 1369 faults /= (sched_max_numa_distance - LOCAL_DISTANCE); 1370 } 1371 1372 score += faults; 1373 } 1374 1375 return score; 1376 } 1377 1378 /* 1379 * These return the fraction of accesses done by a particular task, or 1380 * task group, on a particular numa node. The group weight is given a 1381 * larger multiplier, in order to group tasks together that are almost 1382 * evenly spread out between numa nodes. 1383 */ 1384 static inline unsigned long task_weight(struct task_struct *p, int nid, 1385 int dist) 1386 { 1387 unsigned long faults, total_faults; 1388 1389 if (!p->numa_faults) 1390 return 0; 1391 1392 total_faults = p->total_numa_faults; 1393 1394 if (!total_faults) 1395 return 0; 1396 1397 faults = task_faults(p, nid); 1398 faults += score_nearby_nodes(p, nid, dist, true); 1399 1400 return 1000 * faults / total_faults; 1401 } 1402 1403 static inline unsigned long group_weight(struct task_struct *p, int nid, 1404 int dist) 1405 { 1406 struct numa_group *ng = deref_task_numa_group(p); 1407 unsigned long faults, total_faults; 1408 1409 if (!ng) 1410 return 0; 1411 1412 total_faults = ng->total_faults; 1413 1414 if (!total_faults) 1415 return 0; 1416 1417 faults = group_faults(p, nid); 1418 faults += score_nearby_nodes(p, nid, dist, false); 1419 1420 return 1000 * faults / total_faults; 1421 } 1422 1423 bool should_numa_migrate_memory(struct task_struct *p, struct page * page, 1424 int src_nid, int dst_cpu) 1425 { 1426 struct numa_group *ng = deref_curr_numa_group(p); 1427 int dst_nid = cpu_to_node(dst_cpu); 1428 int last_cpupid, this_cpupid; 1429 1430 this_cpupid = cpu_pid_to_cpupid(dst_cpu, current->pid); 1431 last_cpupid = page_cpupid_xchg_last(page, this_cpupid); 1432 1433 /* 1434 * Allow first faults or private faults to migrate immediately early in 1435 * the lifetime of a task. The magic number 4 is based on waiting for 1436 * two full passes of the "multi-stage node selection" test that is 1437 * executed below. 1438 */ 1439 if ((p->numa_preferred_nid == NUMA_NO_NODE || p->numa_scan_seq <= 4) && 1440 (cpupid_pid_unset(last_cpupid) || cpupid_match_pid(p, last_cpupid))) 1441 return true; 1442 1443 /* 1444 * Multi-stage node selection is used in conjunction with a periodic 1445 * migration fault to build a temporal task<->page relation. By using 1446 * a two-stage filter we remove short/unlikely relations. 1447 * 1448 * Using P(p) ~ n_p / n_t as per frequentist probability, we can equate 1449 * a task's usage of a particular page (n_p) per total usage of this 1450 * page (n_t) (in a given time-span) to a probability. 1451 * 1452 * Our periodic faults will sample this probability and getting the 1453 * same result twice in a row, given these samples are fully 1454 * independent, is then given by P(n)^2, provided our sample period 1455 * is sufficiently short compared to the usage pattern. 1456 * 1457 * This quadric squishes small probabilities, making it less likely we 1458 * act on an unlikely task<->page relation. 1459 */ 1460 if (!cpupid_pid_unset(last_cpupid) && 1461 cpupid_to_nid(last_cpupid) != dst_nid) 1462 return false; 1463 1464 /* Always allow migrate on private faults */ 1465 if (cpupid_match_pid(p, last_cpupid)) 1466 return true; 1467 1468 /* A shared fault, but p->numa_group has not been set up yet. */ 1469 if (!ng) 1470 return true; 1471 1472 /* 1473 * Destination node is much more heavily used than the source 1474 * node? Allow migration. 1475 */ 1476 if (group_faults_cpu(ng, dst_nid) > group_faults_cpu(ng, src_nid) * 1477 ACTIVE_NODE_FRACTION) 1478 return true; 1479 1480 /* 1481 * Distribute memory according to CPU & memory use on each node, 1482 * with 3/4 hysteresis to avoid unnecessary memory migrations: 1483 * 1484 * faults_cpu(dst) 3 faults_cpu(src) 1485 * --------------- * - > --------------- 1486 * faults_mem(dst) 4 faults_mem(src) 1487 */ 1488 return group_faults_cpu(ng, dst_nid) * group_faults(p, src_nid) * 3 > 1489 group_faults_cpu(ng, src_nid) * group_faults(p, dst_nid) * 4; 1490 } 1491 1492 /* 1493 * 'numa_type' describes the node at the moment of load balancing. 1494 */ 1495 enum numa_type { 1496 /* The node has spare capacity that can be used to run more tasks. */ 1497 node_has_spare = 0, 1498 /* 1499 * The node is fully used and the tasks don't compete for more CPU 1500 * cycles. Nevertheless, some tasks might wait before running. 1501 */ 1502 node_fully_busy, 1503 /* 1504 * The node is overloaded and can't provide expected CPU cycles to all 1505 * tasks. 1506 */ 1507 node_overloaded 1508 }; 1509 1510 /* Cached statistics for all CPUs within a node */ 1511 struct numa_stats { 1512 unsigned long load; 1513 unsigned long runnable; 1514 unsigned long util; 1515 /* Total compute capacity of CPUs on a node */ 1516 unsigned long compute_capacity; 1517 unsigned int nr_running; 1518 unsigned int weight; 1519 enum numa_type node_type; 1520 int idle_cpu; 1521 }; 1522 1523 static inline bool is_core_idle(int cpu) 1524 { 1525 #ifdef CONFIG_SCHED_SMT 1526 int sibling; 1527 1528 for_each_cpu(sibling, cpu_smt_mask(cpu)) { 1529 if (cpu == sibling) 1530 continue; 1531 1532 if (!idle_cpu(cpu)) 1533 return false; 1534 } 1535 #endif 1536 1537 return true; 1538 } 1539 1540 struct task_numa_env { 1541 struct task_struct *p; 1542 1543 int src_cpu, src_nid; 1544 int dst_cpu, dst_nid; 1545 1546 struct numa_stats src_stats, dst_stats; 1547 1548 int imbalance_pct; 1549 int dist; 1550 1551 struct task_struct *best_task; 1552 long best_imp; 1553 int best_cpu; 1554 }; 1555 1556 static unsigned long cpu_load(struct rq *rq); 1557 static unsigned long cpu_runnable(struct rq *rq); 1558 static unsigned long cpu_util(int cpu); 1559 static inline long adjust_numa_imbalance(int imbalance, 1560 int dst_running, int dst_weight); 1561 1562 static inline enum 1563 numa_type numa_classify(unsigned int imbalance_pct, 1564 struct numa_stats *ns) 1565 { 1566 if ((ns->nr_running > ns->weight) && 1567 (((ns->compute_capacity * 100) < (ns->util * imbalance_pct)) || 1568 ((ns->compute_capacity * imbalance_pct) < (ns->runnable * 100)))) 1569 return node_overloaded; 1570 1571 if ((ns->nr_running < ns->weight) || 1572 (((ns->compute_capacity * 100) > (ns->util * imbalance_pct)) && 1573 ((ns->compute_capacity * imbalance_pct) > (ns->runnable * 100)))) 1574 return node_has_spare; 1575 1576 return node_fully_busy; 1577 } 1578 1579 #ifdef CONFIG_SCHED_SMT 1580 /* Forward declarations of select_idle_sibling helpers */ 1581 static inline bool test_idle_cores(int cpu, bool def); 1582 static inline int numa_idle_core(int idle_core, int cpu) 1583 { 1584 if (!static_branch_likely(&sched_smt_present) || 1585 idle_core >= 0 || !test_idle_cores(cpu, false)) 1586 return idle_core; 1587 1588 /* 1589 * Prefer cores instead of packing HT siblings 1590 * and triggering future load balancing. 1591 */ 1592 if (is_core_idle(cpu)) 1593 idle_core = cpu; 1594 1595 return idle_core; 1596 } 1597 #else 1598 static inline int numa_idle_core(int idle_core, int cpu) 1599 { 1600 return idle_core; 1601 } 1602 #endif 1603 1604 /* 1605 * Gather all necessary information to make NUMA balancing placement 1606 * decisions that are compatible with standard load balancer. This 1607 * borrows code and logic from update_sg_lb_stats but sharing a 1608 * common implementation is impractical. 1609 */ 1610 static void update_numa_stats(struct task_numa_env *env, 1611 struct numa_stats *ns, int nid, 1612 bool find_idle) 1613 { 1614 int cpu, idle_core = -1; 1615 1616 memset(ns, 0, sizeof(*ns)); 1617 ns->idle_cpu = -1; 1618 1619 rcu_read_lock(); 1620 for_each_cpu(cpu, cpumask_of_node(nid)) { 1621 struct rq *rq = cpu_rq(cpu); 1622 1623 ns->load += cpu_load(rq); 1624 ns->runnable += cpu_runnable(rq); 1625 ns->util += cpu_util(cpu); 1626 ns->nr_running += rq->cfs.h_nr_running; 1627 ns->compute_capacity += capacity_of(cpu); 1628 1629 if (find_idle && !rq->nr_running && idle_cpu(cpu)) { 1630 if (READ_ONCE(rq->numa_migrate_on) || 1631 !cpumask_test_cpu(cpu, env->p->cpus_ptr)) 1632 continue; 1633 1634 if (ns->idle_cpu == -1) 1635 ns->idle_cpu = cpu; 1636 1637 idle_core = numa_idle_core(idle_core, cpu); 1638 } 1639 } 1640 rcu_read_unlock(); 1641 1642 ns->weight = cpumask_weight(cpumask_of_node(nid)); 1643 1644 ns->node_type = numa_classify(env->imbalance_pct, ns); 1645 1646 if (idle_core >= 0) 1647 ns->idle_cpu = idle_core; 1648 } 1649 1650 static void task_numa_assign(struct task_numa_env *env, 1651 struct task_struct *p, long imp) 1652 { 1653 struct rq *rq = cpu_rq(env->dst_cpu); 1654 1655 /* Check if run-queue part of active NUMA balance. */ 1656 if (env->best_cpu != env->dst_cpu && xchg(&rq->numa_migrate_on, 1)) { 1657 int cpu; 1658 int start = env->dst_cpu; 1659 1660 /* Find alternative idle CPU. */ 1661 for_each_cpu_wrap(cpu, cpumask_of_node(env->dst_nid), start) { 1662 if (cpu == env->best_cpu || !idle_cpu(cpu) || 1663 !cpumask_test_cpu(cpu, env->p->cpus_ptr)) { 1664 continue; 1665 } 1666 1667 env->dst_cpu = cpu; 1668 rq = cpu_rq(env->dst_cpu); 1669 if (!xchg(&rq->numa_migrate_on, 1)) 1670 goto assign; 1671 } 1672 1673 /* Failed to find an alternative idle CPU */ 1674 return; 1675 } 1676 1677 assign: 1678 /* 1679 * Clear previous best_cpu/rq numa-migrate flag, since task now 1680 * found a better CPU to move/swap. 1681 */ 1682 if (env->best_cpu != -1 && env->best_cpu != env->dst_cpu) { 1683 rq = cpu_rq(env->best_cpu); 1684 WRITE_ONCE(rq->numa_migrate_on, 0); 1685 } 1686 1687 if (env->best_task) 1688 put_task_struct(env->best_task); 1689 if (p) 1690 get_task_struct(p); 1691 1692 env->best_task = p; 1693 env->best_imp = imp; 1694 env->best_cpu = env->dst_cpu; 1695 } 1696 1697 static bool load_too_imbalanced(long src_load, long dst_load, 1698 struct task_numa_env *env) 1699 { 1700 long imb, old_imb; 1701 long orig_src_load, orig_dst_load; 1702 long src_capacity, dst_capacity; 1703 1704 /* 1705 * The load is corrected for the CPU capacity available on each node. 1706 * 1707 * src_load dst_load 1708 * ------------ vs --------- 1709 * src_capacity dst_capacity 1710 */ 1711 src_capacity = env->src_stats.compute_capacity; 1712 dst_capacity = env->dst_stats.compute_capacity; 1713 1714 imb = abs(dst_load * src_capacity - src_load * dst_capacity); 1715 1716 orig_src_load = env->src_stats.load; 1717 orig_dst_load = env->dst_stats.load; 1718 1719 old_imb = abs(orig_dst_load * src_capacity - orig_src_load * dst_capacity); 1720 1721 /* Would this change make things worse? */ 1722 return (imb > old_imb); 1723 } 1724 1725 /* 1726 * Maximum NUMA importance can be 1998 (2*999); 1727 * SMALLIMP @ 30 would be close to 1998/64. 1728 * Used to deter task migration. 1729 */ 1730 #define SMALLIMP 30 1731 1732 /* 1733 * This checks if the overall compute and NUMA accesses of the system would 1734 * be improved if the source tasks was migrated to the target dst_cpu taking 1735 * into account that it might be best if task running on the dst_cpu should 1736 * be exchanged with the source task 1737 */ 1738 static bool task_numa_compare(struct task_numa_env *env, 1739 long taskimp, long groupimp, bool maymove) 1740 { 1741 struct numa_group *cur_ng, *p_ng = deref_curr_numa_group(env->p); 1742 struct rq *dst_rq = cpu_rq(env->dst_cpu); 1743 long imp = p_ng ? groupimp : taskimp; 1744 struct task_struct *cur; 1745 long src_load, dst_load; 1746 int dist = env->dist; 1747 long moveimp = imp; 1748 long load; 1749 bool stopsearch = false; 1750 1751 if (READ_ONCE(dst_rq->numa_migrate_on)) 1752 return false; 1753 1754 rcu_read_lock(); 1755 cur = rcu_dereference(dst_rq->curr); 1756 if (cur && ((cur->flags & PF_EXITING) || is_idle_task(cur))) 1757 cur = NULL; 1758 1759 /* 1760 * Because we have preemption enabled we can get migrated around and 1761 * end try selecting ourselves (current == env->p) as a swap candidate. 1762 */ 1763 if (cur == env->p) { 1764 stopsearch = true; 1765 goto unlock; 1766 } 1767 1768 if (!cur) { 1769 if (maymove && moveimp >= env->best_imp) 1770 goto assign; 1771 else 1772 goto unlock; 1773 } 1774 1775 /* Skip this swap candidate if cannot move to the source cpu. */ 1776 if (!cpumask_test_cpu(env->src_cpu, cur->cpus_ptr)) 1777 goto unlock; 1778 1779 /* 1780 * Skip this swap candidate if it is not moving to its preferred 1781 * node and the best task is. 1782 */ 1783 if (env->best_task && 1784 env->best_task->numa_preferred_nid == env->src_nid && 1785 cur->numa_preferred_nid != env->src_nid) { 1786 goto unlock; 1787 } 1788 1789 /* 1790 * "imp" is the fault differential for the source task between the 1791 * source and destination node. Calculate the total differential for 1792 * the source task and potential destination task. The more negative 1793 * the value is, the more remote accesses that would be expected to 1794 * be incurred if the tasks were swapped. 1795 * 1796 * If dst and source tasks are in the same NUMA group, or not 1797 * in any group then look only at task weights. 1798 */ 1799 cur_ng = rcu_dereference(cur->numa_group); 1800 if (cur_ng == p_ng) { 1801 imp = taskimp + task_weight(cur, env->src_nid, dist) - 1802 task_weight(cur, env->dst_nid, dist); 1803 /* 1804 * Add some hysteresis to prevent swapping the 1805 * tasks within a group over tiny differences. 1806 */ 1807 if (cur_ng) 1808 imp -= imp / 16; 1809 } else { 1810 /* 1811 * Compare the group weights. If a task is all by itself 1812 * (not part of a group), use the task weight instead. 1813 */ 1814 if (cur_ng && p_ng) 1815 imp += group_weight(cur, env->src_nid, dist) - 1816 group_weight(cur, env->dst_nid, dist); 1817 else 1818 imp += task_weight(cur, env->src_nid, dist) - 1819 task_weight(cur, env->dst_nid, dist); 1820 } 1821 1822 /* Discourage picking a task already on its preferred node */ 1823 if (cur->numa_preferred_nid == env->dst_nid) 1824 imp -= imp / 16; 1825 1826 /* 1827 * Encourage picking a task that moves to its preferred node. 1828 * This potentially makes imp larger than it's maximum of 1829 * 1998 (see SMALLIMP and task_weight for why) but in this 1830 * case, it does not matter. 1831 */ 1832 if (cur->numa_preferred_nid == env->src_nid) 1833 imp += imp / 8; 1834 1835 if (maymove && moveimp > imp && moveimp > env->best_imp) { 1836 imp = moveimp; 1837 cur = NULL; 1838 goto assign; 1839 } 1840 1841 /* 1842 * Prefer swapping with a task moving to its preferred node over a 1843 * task that is not. 1844 */ 1845 if (env->best_task && cur->numa_preferred_nid == env->src_nid && 1846 env->best_task->numa_preferred_nid != env->src_nid) { 1847 goto assign; 1848 } 1849 1850 /* 1851 * If the NUMA importance is less than SMALLIMP, 1852 * task migration might only result in ping pong 1853 * of tasks and also hurt performance due to cache 1854 * misses. 1855 */ 1856 if (imp < SMALLIMP || imp <= env->best_imp + SMALLIMP / 2) 1857 goto unlock; 1858 1859 /* 1860 * In the overloaded case, try and keep the load balanced. 1861 */ 1862 load = task_h_load(env->p) - task_h_load(cur); 1863 if (!load) 1864 goto assign; 1865 1866 dst_load = env->dst_stats.load + load; 1867 src_load = env->src_stats.load - load; 1868 1869 if (load_too_imbalanced(src_load, dst_load, env)) 1870 goto unlock; 1871 1872 assign: 1873 /* Evaluate an idle CPU for a task numa move. */ 1874 if (!cur) { 1875 int cpu = env->dst_stats.idle_cpu; 1876 1877 /* Nothing cached so current CPU went idle since the search. */ 1878 if (cpu < 0) 1879 cpu = env->dst_cpu; 1880 1881 /* 1882 * If the CPU is no longer truly idle and the previous best CPU 1883 * is, keep using it. 1884 */ 1885 if (!idle_cpu(cpu) && env->best_cpu >= 0 && 1886 idle_cpu(env->best_cpu)) { 1887 cpu = env->best_cpu; 1888 } 1889 1890 env->dst_cpu = cpu; 1891 } 1892 1893 task_numa_assign(env, cur, imp); 1894 1895 /* 1896 * If a move to idle is allowed because there is capacity or load 1897 * balance improves then stop the search. While a better swap 1898 * candidate may exist, a search is not free. 1899 */ 1900 if (maymove && !cur && env->best_cpu >= 0 && idle_cpu(env->best_cpu)) 1901 stopsearch = true; 1902 1903 /* 1904 * If a swap candidate must be identified and the current best task 1905 * moves its preferred node then stop the search. 1906 */ 1907 if (!maymove && env->best_task && 1908 env->best_task->numa_preferred_nid == env->src_nid) { 1909 stopsearch = true; 1910 } 1911 unlock: 1912 rcu_read_unlock(); 1913 1914 return stopsearch; 1915 } 1916 1917 static void task_numa_find_cpu(struct task_numa_env *env, 1918 long taskimp, long groupimp) 1919 { 1920 bool maymove = false; 1921 int cpu; 1922 1923 /* 1924 * If dst node has spare capacity, then check if there is an 1925 * imbalance that would be overruled by the load balancer. 1926 */ 1927 if (env->dst_stats.node_type == node_has_spare) { 1928 unsigned int imbalance; 1929 int src_running, dst_running; 1930 1931 /* 1932 * Would movement cause an imbalance? Note that if src has 1933 * more running tasks that the imbalance is ignored as the 1934 * move improves the imbalance from the perspective of the 1935 * CPU load balancer. 1936 * */ 1937 src_running = env->src_stats.nr_running - 1; 1938 dst_running = env->dst_stats.nr_running + 1; 1939 imbalance = max(0, dst_running - src_running); 1940 imbalance = adjust_numa_imbalance(imbalance, dst_running, 1941 env->dst_stats.weight); 1942 1943 /* Use idle CPU if there is no imbalance */ 1944 if (!imbalance) { 1945 maymove = true; 1946 if (env->dst_stats.idle_cpu >= 0) { 1947 env->dst_cpu = env->dst_stats.idle_cpu; 1948 task_numa_assign(env, NULL, 0); 1949 return; 1950 } 1951 } 1952 } else { 1953 long src_load, dst_load, load; 1954 /* 1955 * If the improvement from just moving env->p direction is better 1956 * than swapping tasks around, check if a move is possible. 1957 */ 1958 load = task_h_load(env->p); 1959 dst_load = env->dst_stats.load + load; 1960 src_load = env->src_stats.load - load; 1961 maymove = !load_too_imbalanced(src_load, dst_load, env); 1962 } 1963 1964 for_each_cpu(cpu, cpumask_of_node(env->dst_nid)) { 1965 /* Skip this CPU if the source task cannot migrate */ 1966 if (!cpumask_test_cpu(cpu, env->p->cpus_ptr)) 1967 continue; 1968 1969 env->dst_cpu = cpu; 1970 if (task_numa_compare(env, taskimp, groupimp, maymove)) 1971 break; 1972 } 1973 } 1974 1975 static int task_numa_migrate(struct task_struct *p) 1976 { 1977 struct task_numa_env env = { 1978 .p = p, 1979 1980 .src_cpu = task_cpu(p), 1981 .src_nid = task_node(p), 1982 1983 .imbalance_pct = 112, 1984 1985 .best_task = NULL, 1986 .best_imp = 0, 1987 .best_cpu = -1, 1988 }; 1989 unsigned long taskweight, groupweight; 1990 struct sched_domain *sd; 1991 long taskimp, groupimp; 1992 struct numa_group *ng; 1993 struct rq *best_rq; 1994 int nid, ret, dist; 1995 1996 /* 1997 * Pick the lowest SD_NUMA domain, as that would have the smallest 1998 * imbalance and would be the first to start moving tasks about. 1999 * 2000 * And we want to avoid any moving of tasks about, as that would create 2001 * random movement of tasks -- counter the numa conditions we're trying 2002 * to satisfy here. 2003 */ 2004 rcu_read_lock(); 2005 sd = rcu_dereference(per_cpu(sd_numa, env.src_cpu)); 2006 if (sd) 2007 env.imbalance_pct = 100 + (sd->imbalance_pct - 100) / 2; 2008 rcu_read_unlock(); 2009 2010 /* 2011 * Cpusets can break the scheduler domain tree into smaller 2012 * balance domains, some of which do not cross NUMA boundaries. 2013 * Tasks that are "trapped" in such domains cannot be migrated 2014 * elsewhere, so there is no point in (re)trying. 2015 */ 2016 if (unlikely(!sd)) { 2017 sched_setnuma(p, task_node(p)); 2018 return -EINVAL; 2019 } 2020 2021 env.dst_nid = p->numa_preferred_nid; 2022 dist = env.dist = node_distance(env.src_nid, env.dst_nid); 2023 taskweight = task_weight(p, env.src_nid, dist); 2024 groupweight = group_weight(p, env.src_nid, dist); 2025 update_numa_stats(&env, &env.src_stats, env.src_nid, false); 2026 taskimp = task_weight(p, env.dst_nid, dist) - taskweight; 2027 groupimp = group_weight(p, env.dst_nid, dist) - groupweight; 2028 update_numa_stats(&env, &env.dst_stats, env.dst_nid, true); 2029 2030 /* Try to find a spot on the preferred nid. */ 2031 task_numa_find_cpu(&env, taskimp, groupimp); 2032 2033 /* 2034 * Look at other nodes in these cases: 2035 * - there is no space available on the preferred_nid 2036 * - the task is part of a numa_group that is interleaved across 2037 * multiple NUMA nodes; in order to better consolidate the group, 2038 * we need to check other locations. 2039 */ 2040 ng = deref_curr_numa_group(p); 2041 if (env.best_cpu == -1 || (ng && ng->active_nodes > 1)) { 2042 for_each_online_node(nid) { 2043 if (nid == env.src_nid || nid == p->numa_preferred_nid) 2044 continue; 2045 2046 dist = node_distance(env.src_nid, env.dst_nid); 2047 if (sched_numa_topology_type == NUMA_BACKPLANE && 2048 dist != env.dist) { 2049 taskweight = task_weight(p, env.src_nid, dist); 2050 groupweight = group_weight(p, env.src_nid, dist); 2051 } 2052 2053 /* Only consider nodes where both task and groups benefit */ 2054 taskimp = task_weight(p, nid, dist) - taskweight; 2055 groupimp = group_weight(p, nid, dist) - groupweight; 2056 if (taskimp < 0 && groupimp < 0) 2057 continue; 2058 2059 env.dist = dist; 2060 env.dst_nid = nid; 2061 update_numa_stats(&env, &env.dst_stats, env.dst_nid, true); 2062 task_numa_find_cpu(&env, taskimp, groupimp); 2063 } 2064 } 2065 2066 /* 2067 * If the task is part of a workload that spans multiple NUMA nodes, 2068 * and is migrating into one of the workload's active nodes, remember 2069 * this node as the task's preferred numa node, so the workload can 2070 * settle down. 2071 * A task that migrated to a second choice node will be better off 2072 * trying for a better one later. Do not set the preferred node here. 2073 */ 2074 if (ng) { 2075 if (env.best_cpu == -1) 2076 nid = env.src_nid; 2077 else 2078 nid = cpu_to_node(env.best_cpu); 2079 2080 if (nid != p->numa_preferred_nid) 2081 sched_setnuma(p, nid); 2082 } 2083 2084 /* No better CPU than the current one was found. */ 2085 if (env.best_cpu == -1) { 2086 trace_sched_stick_numa(p, env.src_cpu, NULL, -1); 2087 return -EAGAIN; 2088 } 2089 2090 best_rq = cpu_rq(env.best_cpu); 2091 if (env.best_task == NULL) { 2092 ret = migrate_task_to(p, env.best_cpu); 2093 WRITE_ONCE(best_rq->numa_migrate_on, 0); 2094 if (ret != 0) 2095 trace_sched_stick_numa(p, env.src_cpu, NULL, env.best_cpu); 2096 return ret; 2097 } 2098 2099 ret = migrate_swap(p, env.best_task, env.best_cpu, env.src_cpu); 2100 WRITE_ONCE(best_rq->numa_migrate_on, 0); 2101 2102 if (ret != 0) 2103 trace_sched_stick_numa(p, env.src_cpu, env.best_task, env.best_cpu); 2104 put_task_struct(env.best_task); 2105 return ret; 2106 } 2107 2108 /* Attempt to migrate a task to a CPU on the preferred node. */ 2109 static void numa_migrate_preferred(struct task_struct *p) 2110 { 2111 unsigned long interval = HZ; 2112 2113 /* This task has no NUMA fault statistics yet */ 2114 if (unlikely(p->numa_preferred_nid == NUMA_NO_NODE || !p->numa_faults)) 2115 return; 2116 2117 /* Periodically retry migrating the task to the preferred node */ 2118 interval = min(interval, msecs_to_jiffies(p->numa_scan_period) / 16); 2119 p->numa_migrate_retry = jiffies + interval; 2120 2121 /* Success if task is already running on preferred CPU */ 2122 if (task_node(p) == p->numa_preferred_nid) 2123 return; 2124 2125 /* Otherwise, try migrate to a CPU on the preferred node */ 2126 task_numa_migrate(p); 2127 } 2128 2129 /* 2130 * Find out how many nodes on the workload is actively running on. Do this by 2131 * tracking the nodes from which NUMA hinting faults are triggered. This can 2132 * be different from the set of nodes where the workload's memory is currently 2133 * located. 2134 */ 2135 static void numa_group_count_active_nodes(struct numa_group *numa_group) 2136 { 2137 unsigned long faults, max_faults = 0; 2138 int nid, active_nodes = 0; 2139 2140 for_each_online_node(nid) { 2141 faults = group_faults_cpu(numa_group, nid); 2142 if (faults > max_faults) 2143 max_faults = faults; 2144 } 2145 2146 for_each_online_node(nid) { 2147 faults = group_faults_cpu(numa_group, nid); 2148 if (faults * ACTIVE_NODE_FRACTION > max_faults) 2149 active_nodes++; 2150 } 2151 2152 numa_group->max_faults_cpu = max_faults; 2153 numa_group->active_nodes = active_nodes; 2154 } 2155 2156 /* 2157 * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS 2158 * increments. The more local the fault statistics are, the higher the scan 2159 * period will be for the next scan window. If local/(local+remote) ratio is 2160 * below NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS) 2161 * the scan period will decrease. Aim for 70% local accesses. 2162 */ 2163 #define NUMA_PERIOD_SLOTS 10 2164 #define NUMA_PERIOD_THRESHOLD 7 2165 2166 /* 2167 * Increase the scan period (slow down scanning) if the majority of 2168 * our memory is already on our local node, or if the majority of 2169 * the page accesses are shared with other processes. 2170 * Otherwise, decrease the scan period. 2171 */ 2172 static void update_task_scan_period(struct task_struct *p, 2173 unsigned long shared, unsigned long private) 2174 { 2175 unsigned int period_slot; 2176 int lr_ratio, ps_ratio; 2177 int diff; 2178 2179 unsigned long remote = p->numa_faults_locality[0]; 2180 unsigned long local = p->numa_faults_locality[1]; 2181 2182 /* 2183 * If there were no record hinting faults then either the task is 2184 * completely idle or all activity is areas that are not of interest 2185 * to automatic numa balancing. Related to that, if there were failed 2186 * migration then it implies we are migrating too quickly or the local 2187 * node is overloaded. In either case, scan slower 2188 */ 2189 if (local + shared == 0 || p->numa_faults_locality[2]) { 2190 p->numa_scan_period = min(p->numa_scan_period_max, 2191 p->numa_scan_period << 1); 2192 2193 p->mm->numa_next_scan = jiffies + 2194 msecs_to_jiffies(p->numa_scan_period); 2195 2196 return; 2197 } 2198 2199 /* 2200 * Prepare to scale scan period relative to the current period. 2201 * == NUMA_PERIOD_THRESHOLD scan period stays the same 2202 * < NUMA_PERIOD_THRESHOLD scan period decreases (scan faster) 2203 * >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower) 2204 */ 2205 period_slot = DIV_ROUND_UP(p->numa_scan_period, NUMA_PERIOD_SLOTS); 2206 lr_ratio = (local * NUMA_PERIOD_SLOTS) / (local + remote); 2207 ps_ratio = (private * NUMA_PERIOD_SLOTS) / (private + shared); 2208 2209 if (ps_ratio >= NUMA_PERIOD_THRESHOLD) { 2210 /* 2211 * Most memory accesses are local. There is no need to 2212 * do fast NUMA scanning, since memory is already local. 2213 */ 2214 int slot = ps_ratio - NUMA_PERIOD_THRESHOLD; 2215 if (!slot) 2216 slot = 1; 2217 diff = slot * period_slot; 2218 } else if (lr_ratio >= NUMA_PERIOD_THRESHOLD) { 2219 /* 2220 * Most memory accesses are shared with other tasks. 2221 * There is no point in continuing fast NUMA scanning, 2222 * since other tasks may just move the memory elsewhere. 2223 */ 2224 int slot = lr_ratio - NUMA_PERIOD_THRESHOLD; 2225 if (!slot) 2226 slot = 1; 2227 diff = slot * period_slot; 2228 } else { 2229 /* 2230 * Private memory faults exceed (SLOTS-THRESHOLD)/SLOTS, 2231 * yet they are not on the local NUMA node. Speed up 2232 * NUMA scanning to get the memory moved over. 2233 */ 2234 int ratio = max(lr_ratio, ps_ratio); 2235 diff = -(NUMA_PERIOD_THRESHOLD - ratio) * period_slot; 2236 } 2237 2238 p->numa_scan_period = clamp(p->numa_scan_period + diff, 2239 task_scan_min(p), task_scan_max(p)); 2240 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality)); 2241 } 2242 2243 /* 2244 * Get the fraction of time the task has been running since the last 2245 * NUMA placement cycle. The scheduler keeps similar statistics, but 2246 * decays those on a 32ms period, which is orders of magnitude off 2247 * from the dozens-of-seconds NUMA balancing period. Use the scheduler 2248 * stats only if the task is so new there are no NUMA statistics yet. 2249 */ 2250 static u64 numa_get_avg_runtime(struct task_struct *p, u64 *period) 2251 { 2252 u64 runtime, delta, now; 2253 /* Use the start of this time slice to avoid calculations. */ 2254 now = p->se.exec_start; 2255 runtime = p->se.sum_exec_runtime; 2256 2257 if (p->last_task_numa_placement) { 2258 delta = runtime - p->last_sum_exec_runtime; 2259 *period = now - p->last_task_numa_placement; 2260 2261 /* Avoid time going backwards, prevent potential divide error: */ 2262 if (unlikely((s64)*period < 0)) 2263 *period = 0; 2264 } else { 2265 delta = p->se.avg.load_sum; 2266 *period = LOAD_AVG_MAX; 2267 } 2268 2269 p->last_sum_exec_runtime = runtime; 2270 p->last_task_numa_placement = now; 2271 2272 return delta; 2273 } 2274 2275 /* 2276 * Determine the preferred nid for a task in a numa_group. This needs to 2277 * be done in a way that produces consistent results with group_weight, 2278 * otherwise workloads might not converge. 2279 */ 2280 static int preferred_group_nid(struct task_struct *p, int nid) 2281 { 2282 nodemask_t nodes; 2283 int dist; 2284 2285 /* Direct connections between all NUMA nodes. */ 2286 if (sched_numa_topology_type == NUMA_DIRECT) 2287 return nid; 2288 2289 /* 2290 * On a system with glueless mesh NUMA topology, group_weight 2291 * scores nodes according to the number of NUMA hinting faults on 2292 * both the node itself, and on nearby nodes. 2293 */ 2294 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) { 2295 unsigned long score, max_score = 0; 2296 int node, max_node = nid; 2297 2298 dist = sched_max_numa_distance; 2299 2300 for_each_online_node(node) { 2301 score = group_weight(p, node, dist); 2302 if (score > max_score) { 2303 max_score = score; 2304 max_node = node; 2305 } 2306 } 2307 return max_node; 2308 } 2309 2310 /* 2311 * Finding the preferred nid in a system with NUMA backplane 2312 * interconnect topology is more involved. The goal is to locate 2313 * tasks from numa_groups near each other in the system, and 2314 * untangle workloads from different sides of the system. This requires 2315 * searching down the hierarchy of node groups, recursively searching 2316 * inside the highest scoring group of nodes. The nodemask tricks 2317 * keep the complexity of the search down. 2318 */ 2319 nodes = node_online_map; 2320 for (dist = sched_max_numa_distance; dist > LOCAL_DISTANCE; dist--) { 2321 unsigned long max_faults = 0; 2322 nodemask_t max_group = NODE_MASK_NONE; 2323 int a, b; 2324 2325 /* Are there nodes at this distance from each other? */ 2326 if (!find_numa_distance(dist)) 2327 continue; 2328 2329 for_each_node_mask(a, nodes) { 2330 unsigned long faults = 0; 2331 nodemask_t this_group; 2332 nodes_clear(this_group); 2333 2334 /* Sum group's NUMA faults; includes a==b case. */ 2335 for_each_node_mask(b, nodes) { 2336 if (node_distance(a, b) < dist) { 2337 faults += group_faults(p, b); 2338 node_set(b, this_group); 2339 node_clear(b, nodes); 2340 } 2341 } 2342 2343 /* Remember the top group. */ 2344 if (faults > max_faults) { 2345 max_faults = faults; 2346 max_group = this_group; 2347 /* 2348 * subtle: at the smallest distance there is 2349 * just one node left in each "group", the 2350 * winner is the preferred nid. 2351 */ 2352 nid = a; 2353 } 2354 } 2355 /* Next round, evaluate the nodes within max_group. */ 2356 if (!max_faults) 2357 break; 2358 nodes = max_group; 2359 } 2360 return nid; 2361 } 2362 2363 static void task_numa_placement(struct task_struct *p) 2364 { 2365 int seq, nid, max_nid = NUMA_NO_NODE; 2366 unsigned long max_faults = 0; 2367 unsigned long fault_types[2] = { 0, 0 }; 2368 unsigned long total_faults; 2369 u64 runtime, period; 2370 spinlock_t *group_lock = NULL; 2371 struct numa_group *ng; 2372 2373 /* 2374 * The p->mm->numa_scan_seq field gets updated without 2375 * exclusive access. Use READ_ONCE() here to ensure 2376 * that the field is read in a single access: 2377 */ 2378 seq = READ_ONCE(p->mm->numa_scan_seq); 2379 if (p->numa_scan_seq == seq) 2380 return; 2381 p->numa_scan_seq = seq; 2382 p->numa_scan_period_max = task_scan_max(p); 2383 2384 total_faults = p->numa_faults_locality[0] + 2385 p->numa_faults_locality[1]; 2386 runtime = numa_get_avg_runtime(p, &period); 2387 2388 /* If the task is part of a group prevent parallel updates to group stats */ 2389 ng = deref_curr_numa_group(p); 2390 if (ng) { 2391 group_lock = &ng->lock; 2392 spin_lock_irq(group_lock); 2393 } 2394 2395 /* Find the node with the highest number of faults */ 2396 for_each_online_node(nid) { 2397 /* Keep track of the offsets in numa_faults array */ 2398 int mem_idx, membuf_idx, cpu_idx, cpubuf_idx; 2399 unsigned long faults = 0, group_faults = 0; 2400 int priv; 2401 2402 for (priv = 0; priv < NR_NUMA_HINT_FAULT_TYPES; priv++) { 2403 long diff, f_diff, f_weight; 2404 2405 mem_idx = task_faults_idx(NUMA_MEM, nid, priv); 2406 membuf_idx = task_faults_idx(NUMA_MEMBUF, nid, priv); 2407 cpu_idx = task_faults_idx(NUMA_CPU, nid, priv); 2408 cpubuf_idx = task_faults_idx(NUMA_CPUBUF, nid, priv); 2409 2410 /* Decay existing window, copy faults since last scan */ 2411 diff = p->numa_faults[membuf_idx] - p->numa_faults[mem_idx] / 2; 2412 fault_types[priv] += p->numa_faults[membuf_idx]; 2413 p->numa_faults[membuf_idx] = 0; 2414 2415 /* 2416 * Normalize the faults_from, so all tasks in a group 2417 * count according to CPU use, instead of by the raw 2418 * number of faults. Tasks with little runtime have 2419 * little over-all impact on throughput, and thus their 2420 * faults are less important. 2421 */ 2422 f_weight = div64_u64(runtime << 16, period + 1); 2423 f_weight = (f_weight * p->numa_faults[cpubuf_idx]) / 2424 (total_faults + 1); 2425 f_diff = f_weight - p->numa_faults[cpu_idx] / 2; 2426 p->numa_faults[cpubuf_idx] = 0; 2427 2428 p->numa_faults[mem_idx] += diff; 2429 p->numa_faults[cpu_idx] += f_diff; 2430 faults += p->numa_faults[mem_idx]; 2431 p->total_numa_faults += diff; 2432 if (ng) { 2433 /* 2434 * safe because we can only change our own group 2435 * 2436 * mem_idx represents the offset for a given 2437 * nid and priv in a specific region because it 2438 * is at the beginning of the numa_faults array. 2439 */ 2440 ng->faults[mem_idx] += diff; 2441 ng->faults_cpu[mem_idx] += f_diff; 2442 ng->total_faults += diff; 2443 group_faults += ng->faults[mem_idx]; 2444 } 2445 } 2446 2447 if (!ng) { 2448 if (faults > max_faults) { 2449 max_faults = faults; 2450 max_nid = nid; 2451 } 2452 } else if (group_faults > max_faults) { 2453 max_faults = group_faults; 2454 max_nid = nid; 2455 } 2456 } 2457 2458 if (ng) { 2459 numa_group_count_active_nodes(ng); 2460 spin_unlock_irq(group_lock); 2461 max_nid = preferred_group_nid(p, max_nid); 2462 } 2463 2464 if (max_faults) { 2465 /* Set the new preferred node */ 2466 if (max_nid != p->numa_preferred_nid) 2467 sched_setnuma(p, max_nid); 2468 } 2469 2470 update_task_scan_period(p, fault_types[0], fault_types[1]); 2471 } 2472 2473 static inline int get_numa_group(struct numa_group *grp) 2474 { 2475 return refcount_inc_not_zero(&grp->refcount); 2476 } 2477 2478 static inline void put_numa_group(struct numa_group *grp) 2479 { 2480 if (refcount_dec_and_test(&grp->refcount)) 2481 kfree_rcu(grp, rcu); 2482 } 2483 2484 static void task_numa_group(struct task_struct *p, int cpupid, int flags, 2485 int *priv) 2486 { 2487 struct numa_group *grp, *my_grp; 2488 struct task_struct *tsk; 2489 bool join = false; 2490 int cpu = cpupid_to_cpu(cpupid); 2491 int i; 2492 2493 if (unlikely(!deref_curr_numa_group(p))) { 2494 unsigned int size = sizeof(struct numa_group) + 2495 4*nr_node_ids*sizeof(unsigned long); 2496 2497 grp = kzalloc(size, GFP_KERNEL | __GFP_NOWARN); 2498 if (!grp) 2499 return; 2500 2501 refcount_set(&grp->refcount, 1); 2502 grp->active_nodes = 1; 2503 grp->max_faults_cpu = 0; 2504 spin_lock_init(&grp->lock); 2505 grp->gid = p->pid; 2506 /* Second half of the array tracks nids where faults happen */ 2507 grp->faults_cpu = grp->faults + NR_NUMA_HINT_FAULT_TYPES * 2508 nr_node_ids; 2509 2510 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) 2511 grp->faults[i] = p->numa_faults[i]; 2512 2513 grp->total_faults = p->total_numa_faults; 2514 2515 grp->nr_tasks++; 2516 rcu_assign_pointer(p->numa_group, grp); 2517 } 2518 2519 rcu_read_lock(); 2520 tsk = READ_ONCE(cpu_rq(cpu)->curr); 2521 2522 if (!cpupid_match_pid(tsk, cpupid)) 2523 goto no_join; 2524 2525 grp = rcu_dereference(tsk->numa_group); 2526 if (!grp) 2527 goto no_join; 2528 2529 my_grp = deref_curr_numa_group(p); 2530 if (grp == my_grp) 2531 goto no_join; 2532 2533 /* 2534 * Only join the other group if its bigger; if we're the bigger group, 2535 * the other task will join us. 2536 */ 2537 if (my_grp->nr_tasks > grp->nr_tasks) 2538 goto no_join; 2539 2540 /* 2541 * Tie-break on the grp address. 2542 */ 2543 if (my_grp->nr_tasks == grp->nr_tasks && my_grp > grp) 2544 goto no_join; 2545 2546 /* Always join threads in the same process. */ 2547 if (tsk->mm == current->mm) 2548 join = true; 2549 2550 /* Simple filter to avoid false positives due to PID collisions */ 2551 if (flags & TNF_SHARED) 2552 join = true; 2553 2554 /* Update priv based on whether false sharing was detected */ 2555 *priv = !join; 2556 2557 if (join && !get_numa_group(grp)) 2558 goto no_join; 2559 2560 rcu_read_unlock(); 2561 2562 if (!join) 2563 return; 2564 2565 BUG_ON(irqs_disabled()); 2566 double_lock_irq(&my_grp->lock, &grp->lock); 2567 2568 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) { 2569 my_grp->faults[i] -= p->numa_faults[i]; 2570 grp->faults[i] += p->numa_faults[i]; 2571 } 2572 my_grp->total_faults -= p->total_numa_faults; 2573 grp->total_faults += p->total_numa_faults; 2574 2575 my_grp->nr_tasks--; 2576 grp->nr_tasks++; 2577 2578 spin_unlock(&my_grp->lock); 2579 spin_unlock_irq(&grp->lock); 2580 2581 rcu_assign_pointer(p->numa_group, grp); 2582 2583 put_numa_group(my_grp); 2584 return; 2585 2586 no_join: 2587 rcu_read_unlock(); 2588 return; 2589 } 2590 2591 /* 2592 * Get rid of NUMA statistics associated with a task (either current or dead). 2593 * If @final is set, the task is dead and has reached refcount zero, so we can 2594 * safely free all relevant data structures. Otherwise, there might be 2595 * concurrent reads from places like load balancing and procfs, and we should 2596 * reset the data back to default state without freeing ->numa_faults. 2597 */ 2598 void task_numa_free(struct task_struct *p, bool final) 2599 { 2600 /* safe: p either is current or is being freed by current */ 2601 struct numa_group *grp = rcu_dereference_raw(p->numa_group); 2602 unsigned long *numa_faults = p->numa_faults; 2603 unsigned long flags; 2604 int i; 2605 2606 if (!numa_faults) 2607 return; 2608 2609 if (grp) { 2610 spin_lock_irqsave(&grp->lock, flags); 2611 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) 2612 grp->faults[i] -= p->numa_faults[i]; 2613 grp->total_faults -= p->total_numa_faults; 2614 2615 grp->nr_tasks--; 2616 spin_unlock_irqrestore(&grp->lock, flags); 2617 RCU_INIT_POINTER(p->numa_group, NULL); 2618 put_numa_group(grp); 2619 } 2620 2621 if (final) { 2622 p->numa_faults = NULL; 2623 kfree(numa_faults); 2624 } else { 2625 p->total_numa_faults = 0; 2626 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) 2627 numa_faults[i] = 0; 2628 } 2629 } 2630 2631 /* 2632 * Got a PROT_NONE fault for a page on @node. 2633 */ 2634 void task_numa_fault(int last_cpupid, int mem_node, int pages, int flags) 2635 { 2636 struct task_struct *p = current; 2637 bool migrated = flags & TNF_MIGRATED; 2638 int cpu_node = task_node(current); 2639 int local = !!(flags & TNF_FAULT_LOCAL); 2640 struct numa_group *ng; 2641 int priv; 2642 2643 if (!static_branch_likely(&sched_numa_balancing)) 2644 return; 2645 2646 /* for example, ksmd faulting in a user's mm */ 2647 if (!p->mm) 2648 return; 2649 2650 /* Allocate buffer to track faults on a per-node basis */ 2651 if (unlikely(!p->numa_faults)) { 2652 int size = sizeof(*p->numa_faults) * 2653 NR_NUMA_HINT_FAULT_BUCKETS * nr_node_ids; 2654 2655 p->numa_faults = kzalloc(size, GFP_KERNEL|__GFP_NOWARN); 2656 if (!p->numa_faults) 2657 return; 2658 2659 p->total_numa_faults = 0; 2660 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality)); 2661 } 2662 2663 /* 2664 * First accesses are treated as private, otherwise consider accesses 2665 * to be private if the accessing pid has not changed 2666 */ 2667 if (unlikely(last_cpupid == (-1 & LAST_CPUPID_MASK))) { 2668 priv = 1; 2669 } else { 2670 priv = cpupid_match_pid(p, last_cpupid); 2671 if (!priv && !(flags & TNF_NO_GROUP)) 2672 task_numa_group(p, last_cpupid, flags, &priv); 2673 } 2674 2675 /* 2676 * If a workload spans multiple NUMA nodes, a shared fault that 2677 * occurs wholly within the set of nodes that the workload is 2678 * actively using should be counted as local. This allows the 2679 * scan rate to slow down when a workload has settled down. 2680 */ 2681 ng = deref_curr_numa_group(p); 2682 if (!priv && !local && ng && ng->active_nodes > 1 && 2683 numa_is_active_node(cpu_node, ng) && 2684 numa_is_active_node(mem_node, ng)) 2685 local = 1; 2686 2687 /* 2688 * Retry to migrate task to preferred node periodically, in case it 2689 * previously failed, or the scheduler moved us. 2690 */ 2691 if (time_after(jiffies, p->numa_migrate_retry)) { 2692 task_numa_placement(p); 2693 numa_migrate_preferred(p); 2694 } 2695 2696 if (migrated) 2697 p->numa_pages_migrated += pages; 2698 if (flags & TNF_MIGRATE_FAIL) 2699 p->numa_faults_locality[2] += pages; 2700 2701 p->numa_faults[task_faults_idx(NUMA_MEMBUF, mem_node, priv)] += pages; 2702 p->numa_faults[task_faults_idx(NUMA_CPUBUF, cpu_node, priv)] += pages; 2703 p->numa_faults_locality[local] += pages; 2704 } 2705 2706 static void reset_ptenuma_scan(struct task_struct *p) 2707 { 2708 /* 2709 * We only did a read acquisition of the mmap sem, so 2710 * p->mm->numa_scan_seq is written to without exclusive access 2711 * and the update is not guaranteed to be atomic. That's not 2712 * much of an issue though, since this is just used for 2713 * statistical sampling. Use READ_ONCE/WRITE_ONCE, which are not 2714 * expensive, to avoid any form of compiler optimizations: 2715 */ 2716 WRITE_ONCE(p->mm->numa_scan_seq, READ_ONCE(p->mm->numa_scan_seq) + 1); 2717 p->mm->numa_scan_offset = 0; 2718 } 2719 2720 /* 2721 * The expensive part of numa migration is done from task_work context. 2722 * Triggered from task_tick_numa(). 2723 */ 2724 static void task_numa_work(struct callback_head *work) 2725 { 2726 unsigned long migrate, next_scan, now = jiffies; 2727 struct task_struct *p = current; 2728 struct mm_struct *mm = p->mm; 2729 u64 runtime = p->se.sum_exec_runtime; 2730 struct vm_area_struct *vma; 2731 unsigned long start, end; 2732 unsigned long nr_pte_updates = 0; 2733 long pages, virtpages; 2734 2735 SCHED_WARN_ON(p != container_of(work, struct task_struct, numa_work)); 2736 2737 work->next = work; 2738 /* 2739 * Who cares about NUMA placement when they're dying. 2740 * 2741 * NOTE: make sure not to dereference p->mm before this check, 2742 * exit_task_work() happens _after_ exit_mm() so we could be called 2743 * without p->mm even though we still had it when we enqueued this 2744 * work. 2745 */ 2746 if (p->flags & PF_EXITING) 2747 return; 2748 2749 if (!mm->numa_next_scan) { 2750 mm->numa_next_scan = now + 2751 msecs_to_jiffies(sysctl_numa_balancing_scan_delay); 2752 } 2753 2754 /* 2755 * Enforce maximal scan/migration frequency.. 2756 */ 2757 migrate = mm->numa_next_scan; 2758 if (time_before(now, migrate)) 2759 return; 2760 2761 if (p->numa_scan_period == 0) { 2762 p->numa_scan_period_max = task_scan_max(p); 2763 p->numa_scan_period = task_scan_start(p); 2764 } 2765 2766 next_scan = now + msecs_to_jiffies(p->numa_scan_period); 2767 if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate) 2768 return; 2769 2770 /* 2771 * Delay this task enough that another task of this mm will likely win 2772 * the next time around. 2773 */ 2774 p->node_stamp += 2 * TICK_NSEC; 2775 2776 start = mm->numa_scan_offset; 2777 pages = sysctl_numa_balancing_scan_size; 2778 pages <<= 20 - PAGE_SHIFT; /* MB in pages */ 2779 virtpages = pages * 8; /* Scan up to this much virtual space */ 2780 if (!pages) 2781 return; 2782 2783 2784 if (!mmap_read_trylock(mm)) 2785 return; 2786 vma = find_vma(mm, start); 2787 if (!vma) { 2788 reset_ptenuma_scan(p); 2789 start = 0; 2790 vma = mm->mmap; 2791 } 2792 for (; vma; vma = vma->vm_next) { 2793 if (!vma_migratable(vma) || !vma_policy_mof(vma) || 2794 is_vm_hugetlb_page(vma) || (vma->vm_flags & VM_MIXEDMAP)) { 2795 continue; 2796 } 2797 2798 /* 2799 * Shared library pages mapped by multiple processes are not 2800 * migrated as it is expected they are cache replicated. Avoid 2801 * hinting faults in read-only file-backed mappings or the vdso 2802 * as migrating the pages will be of marginal benefit. 2803 */ 2804 if (!vma->vm_mm || 2805 (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ))) 2806 continue; 2807 2808 /* 2809 * Skip inaccessible VMAs to avoid any confusion between 2810 * PROT_NONE and NUMA hinting ptes 2811 */ 2812 if (!vma_is_accessible(vma)) 2813 continue; 2814 2815 do { 2816 start = max(start, vma->vm_start); 2817 end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE); 2818 end = min(end, vma->vm_end); 2819 nr_pte_updates = change_prot_numa(vma, start, end); 2820 2821 /* 2822 * Try to scan sysctl_numa_balancing_size worth of 2823 * hpages that have at least one present PTE that 2824 * is not already pte-numa. If the VMA contains 2825 * areas that are unused or already full of prot_numa 2826 * PTEs, scan up to virtpages, to skip through those 2827 * areas faster. 2828 */ 2829 if (nr_pte_updates) 2830 pages -= (end - start) >> PAGE_SHIFT; 2831 virtpages -= (end - start) >> PAGE_SHIFT; 2832 2833 start = end; 2834 if (pages <= 0 || virtpages <= 0) 2835 goto out; 2836 2837 cond_resched(); 2838 } while (end != vma->vm_end); 2839 } 2840 2841 out: 2842 /* 2843 * It is possible to reach the end of the VMA list but the last few 2844 * VMAs are not guaranteed to the vma_migratable. If they are not, we 2845 * would find the !migratable VMA on the next scan but not reset the 2846 * scanner to the start so check it now. 2847 */ 2848 if (vma) 2849 mm->numa_scan_offset = start; 2850 else 2851 reset_ptenuma_scan(p); 2852 mmap_read_unlock(mm); 2853 2854 /* 2855 * Make sure tasks use at least 32x as much time to run other code 2856 * than they used here, to limit NUMA PTE scanning overhead to 3% max. 2857 * Usually update_task_scan_period slows down scanning enough; on an 2858 * overloaded system we need to limit overhead on a per task basis. 2859 */ 2860 if (unlikely(p->se.sum_exec_runtime != runtime)) { 2861 u64 diff = p->se.sum_exec_runtime - runtime; 2862 p->node_stamp += 32 * diff; 2863 } 2864 } 2865 2866 void init_numa_balancing(unsigned long clone_flags, struct task_struct *p) 2867 { 2868 int mm_users = 0; 2869 struct mm_struct *mm = p->mm; 2870 2871 if (mm) { 2872 mm_users = atomic_read(&mm->mm_users); 2873 if (mm_users == 1) { 2874 mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay); 2875 mm->numa_scan_seq = 0; 2876 } 2877 } 2878 p->node_stamp = 0; 2879 p->numa_scan_seq = mm ? mm->numa_scan_seq : 0; 2880 p->numa_scan_period = sysctl_numa_balancing_scan_delay; 2881 /* Protect against double add, see task_tick_numa and task_numa_work */ 2882 p->numa_work.next = &p->numa_work; 2883 p->numa_faults = NULL; 2884 RCU_INIT_POINTER(p->numa_group, NULL); 2885 p->last_task_numa_placement = 0; 2886 p->last_sum_exec_runtime = 0; 2887 2888 init_task_work(&p->numa_work, task_numa_work); 2889 2890 /* New address space, reset the preferred nid */ 2891 if (!(clone_flags & CLONE_VM)) { 2892 p->numa_preferred_nid = NUMA_NO_NODE; 2893 return; 2894 } 2895 2896 /* 2897 * New thread, keep existing numa_preferred_nid which should be copied 2898 * already by arch_dup_task_struct but stagger when scans start. 2899 */ 2900 if (mm) { 2901 unsigned int delay; 2902 2903 delay = min_t(unsigned int, task_scan_max(current), 2904 current->numa_scan_period * mm_users * NSEC_PER_MSEC); 2905 delay += 2 * TICK_NSEC; 2906 p->node_stamp = delay; 2907 } 2908 } 2909 2910 /* 2911 * Drive the periodic memory faults.. 2912 */ 2913 static void task_tick_numa(struct rq *rq, struct task_struct *curr) 2914 { 2915 struct callback_head *work = &curr->numa_work; 2916 u64 period, now; 2917 2918 /* 2919 * We don't care about NUMA placement if we don't have memory. 2920 */ 2921 if ((curr->flags & (PF_EXITING | PF_KTHREAD)) || work->next != work) 2922 return; 2923 2924 /* 2925 * Using runtime rather than walltime has the dual advantage that 2926 * we (mostly) drive the selection from busy threads and that the 2927 * task needs to have done some actual work before we bother with 2928 * NUMA placement. 2929 */ 2930 now = curr->se.sum_exec_runtime; 2931 period = (u64)curr->numa_scan_period * NSEC_PER_MSEC; 2932 2933 if (now > curr->node_stamp + period) { 2934 if (!curr->node_stamp) 2935 curr->numa_scan_period = task_scan_start(curr); 2936 curr->node_stamp += period; 2937 2938 if (!time_before(jiffies, curr->mm->numa_next_scan)) 2939 task_work_add(curr, work, TWA_RESUME); 2940 } 2941 } 2942 2943 static void update_scan_period(struct task_struct *p, int new_cpu) 2944 { 2945 int src_nid = cpu_to_node(task_cpu(p)); 2946 int dst_nid = cpu_to_node(new_cpu); 2947 2948 if (!static_branch_likely(&sched_numa_balancing)) 2949 return; 2950 2951 if (!p->mm || !p->numa_faults || (p->flags & PF_EXITING)) 2952 return; 2953 2954 if (src_nid == dst_nid) 2955 return; 2956 2957 /* 2958 * Allow resets if faults have been trapped before one scan 2959 * has completed. This is most likely due to a new task that 2960 * is pulled cross-node due to wakeups or load balancing. 2961 */ 2962 if (p->numa_scan_seq) { 2963 /* 2964 * Avoid scan adjustments if moving to the preferred 2965 * node or if the task was not previously running on 2966 * the preferred node. 2967 */ 2968 if (dst_nid == p->numa_preferred_nid || 2969 (p->numa_preferred_nid != NUMA_NO_NODE && 2970 src_nid != p->numa_preferred_nid)) 2971 return; 2972 } 2973 2974 p->numa_scan_period = task_scan_start(p); 2975 } 2976 2977 #else 2978 static void task_tick_numa(struct rq *rq, struct task_struct *curr) 2979 { 2980 } 2981 2982 static inline void account_numa_enqueue(struct rq *rq, struct task_struct *p) 2983 { 2984 } 2985 2986 static inline void account_numa_dequeue(struct rq *rq, struct task_struct *p) 2987 { 2988 } 2989 2990 static inline void update_scan_period(struct task_struct *p, int new_cpu) 2991 { 2992 } 2993 2994 #endif /* CONFIG_NUMA_BALANCING */ 2995 2996 static void 2997 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se) 2998 { 2999 update_load_add(&cfs_rq->load, se->load.weight); 3000 #ifdef CONFIG_SMP 3001 if (entity_is_task(se)) { 3002 struct rq *rq = rq_of(cfs_rq); 3003 3004 account_numa_enqueue(rq, task_of(se)); 3005 list_add(&se->group_node, &rq->cfs_tasks); 3006 } 3007 #endif 3008 cfs_rq->nr_running++; 3009 } 3010 3011 static void 3012 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se) 3013 { 3014 update_load_sub(&cfs_rq->load, se->load.weight); 3015 #ifdef CONFIG_SMP 3016 if (entity_is_task(se)) { 3017 account_numa_dequeue(rq_of(cfs_rq), task_of(se)); 3018 list_del_init(&se->group_node); 3019 } 3020 #endif 3021 cfs_rq->nr_running--; 3022 } 3023 3024 /* 3025 * Signed add and clamp on underflow. 3026 * 3027 * Explicitly do a load-store to ensure the intermediate value never hits 3028 * memory. This allows lockless observations without ever seeing the negative 3029 * values. 3030 */ 3031 #define add_positive(_ptr, _val) do { \ 3032 typeof(_ptr) ptr = (_ptr); \ 3033 typeof(_val) val = (_val); \ 3034 typeof(*ptr) res, var = READ_ONCE(*ptr); \ 3035 \ 3036 res = var + val; \ 3037 \ 3038 if (val < 0 && res > var) \ 3039 res = 0; \ 3040 \ 3041 WRITE_ONCE(*ptr, res); \ 3042 } while (0) 3043 3044 /* 3045 * Unsigned subtract and clamp on underflow. 3046 * 3047 * Explicitly do a load-store to ensure the intermediate value never hits 3048 * memory. This allows lockless observations without ever seeing the negative 3049 * values. 3050 */ 3051 #define sub_positive(_ptr, _val) do { \ 3052 typeof(_ptr) ptr = (_ptr); \ 3053 typeof(*ptr) val = (_val); \ 3054 typeof(*ptr) res, var = READ_ONCE(*ptr); \ 3055 res = var - val; \ 3056 if (res > var) \ 3057 res = 0; \ 3058 WRITE_ONCE(*ptr, res); \ 3059 } while (0) 3060 3061 /* 3062 * Remove and clamp on negative, from a local variable. 3063 * 3064 * A variant of sub_positive(), which does not use explicit load-store 3065 * and is thus optimized for local variable updates. 3066 */ 3067 #define lsub_positive(_ptr, _val) do { \ 3068 typeof(_ptr) ptr = (_ptr); \ 3069 *ptr -= min_t(typeof(*ptr), *ptr, _val); \ 3070 } while (0) 3071 3072 #ifdef CONFIG_SMP 3073 static inline void 3074 enqueue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) 3075 { 3076 cfs_rq->avg.load_avg += se->avg.load_avg; 3077 cfs_rq->avg.load_sum += se_weight(se) * se->avg.load_sum; 3078 } 3079 3080 static inline void 3081 dequeue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) 3082 { 3083 sub_positive(&cfs_rq->avg.load_avg, se->avg.load_avg); 3084 sub_positive(&cfs_rq->avg.load_sum, se_weight(se) * se->avg.load_sum); 3085 } 3086 #else 3087 static inline void 3088 enqueue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { } 3089 static inline void 3090 dequeue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { } 3091 #endif 3092 3093 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, 3094 unsigned long weight) 3095 { 3096 if (se->on_rq) { 3097 /* commit outstanding execution time */ 3098 if (cfs_rq->curr == se) 3099 update_curr(cfs_rq); 3100 update_load_sub(&cfs_rq->load, se->load.weight); 3101 } 3102 dequeue_load_avg(cfs_rq, se); 3103 3104 update_load_set(&se->load, weight); 3105 3106 #ifdef CONFIG_SMP 3107 do { 3108 u32 divider = get_pelt_divider(&se->avg); 3109 3110 se->avg.load_avg = div_u64(se_weight(se) * se->avg.load_sum, divider); 3111 } while (0); 3112 #endif 3113 3114 enqueue_load_avg(cfs_rq, se); 3115 if (se->on_rq) 3116 update_load_add(&cfs_rq->load, se->load.weight); 3117 3118 } 3119 3120 void reweight_task(struct task_struct *p, int prio) 3121 { 3122 struct sched_entity *se = &p->se; 3123 struct cfs_rq *cfs_rq = cfs_rq_of(se); 3124 struct load_weight *load = &se->load; 3125 unsigned long weight = scale_load(sched_prio_to_weight[prio]); 3126 3127 reweight_entity(cfs_rq, se, weight); 3128 load->inv_weight = sched_prio_to_wmult[prio]; 3129 } 3130 3131 #ifdef CONFIG_FAIR_GROUP_SCHED 3132 #ifdef CONFIG_SMP 3133 /* 3134 * All this does is approximate the hierarchical proportion which includes that 3135 * global sum we all love to hate. 3136 * 3137 * That is, the weight of a group entity, is the proportional share of the 3138 * group weight based on the group runqueue weights. That is: 3139 * 3140 * tg->weight * grq->load.weight 3141 * ge->load.weight = ----------------------------- (1) 3142 * \Sum grq->load.weight 3143 * 3144 * Now, because computing that sum is prohibitively expensive to compute (been 3145 * there, done that) we approximate it with this average stuff. The average 3146 * moves slower and therefore the approximation is cheaper and more stable. 3147 * 3148 * So instead of the above, we substitute: 3149 * 3150 * grq->load.weight -> grq->avg.load_avg (2) 3151 * 3152 * which yields the following: 3153 * 3154 * tg->weight * grq->avg.load_avg 3155 * ge->load.weight = ------------------------------ (3) 3156 * tg->load_avg 3157 * 3158 * Where: tg->load_avg ~= \Sum grq->avg.load_avg 3159 * 3160 * That is shares_avg, and it is right (given the approximation (2)). 3161 * 3162 * The problem with it is that because the average is slow -- it was designed 3163 * to be exactly that of course -- this leads to transients in boundary 3164 * conditions. In specific, the case where the group was idle and we start the 3165 * one task. It takes time for our CPU's grq->avg.load_avg to build up, 3166 * yielding bad latency etc.. 3167 * 3168 * Now, in that special case (1) reduces to: 3169 * 3170 * tg->weight * grq->load.weight 3171 * ge->load.weight = ----------------------------- = tg->weight (4) 3172 * grp->load.weight 3173 * 3174 * That is, the sum collapses because all other CPUs are idle; the UP scenario. 3175 * 3176 * So what we do is modify our approximation (3) to approach (4) in the (near) 3177 * UP case, like: 3178 * 3179 * ge->load.weight = 3180 * 3181 * tg->weight * grq->load.weight 3182 * --------------------------------------------------- (5) 3183 * tg->load_avg - grq->avg.load_avg + grq->load.weight 3184 * 3185 * But because grq->load.weight can drop to 0, resulting in a divide by zero, 3186 * we need to use grq->avg.load_avg as its lower bound, which then gives: 3187 * 3188 * 3189 * tg->weight * grq->load.weight 3190 * ge->load.weight = ----------------------------- (6) 3191 * tg_load_avg' 3192 * 3193 * Where: 3194 * 3195 * tg_load_avg' = tg->load_avg - grq->avg.load_avg + 3196 * max(grq->load.weight, grq->avg.load_avg) 3197 * 3198 * And that is shares_weight and is icky. In the (near) UP case it approaches 3199 * (4) while in the normal case it approaches (3). It consistently 3200 * overestimates the ge->load.weight and therefore: 3201 * 3202 * \Sum ge->load.weight >= tg->weight 3203 * 3204 * hence icky! 3205 */ 3206 static long calc_group_shares(struct cfs_rq *cfs_rq) 3207 { 3208 long tg_weight, tg_shares, load, shares; 3209 struct task_group *tg = cfs_rq->tg; 3210 3211 tg_shares = READ_ONCE(tg->shares); 3212 3213 load = max(scale_load_down(cfs_rq->load.weight), cfs_rq->avg.load_avg); 3214 3215 tg_weight = atomic_long_read(&tg->load_avg); 3216 3217 /* Ensure tg_weight >= load */ 3218 tg_weight -= cfs_rq->tg_load_avg_contrib; 3219 tg_weight += load; 3220 3221 shares = (tg_shares * load); 3222 if (tg_weight) 3223 shares /= tg_weight; 3224 3225 /* 3226 * MIN_SHARES has to be unscaled here to support per-CPU partitioning 3227 * of a group with small tg->shares value. It is a floor value which is 3228 * assigned as a minimum load.weight to the sched_entity representing 3229 * the group on a CPU. 3230 * 3231 * E.g. on 64-bit for a group with tg->shares of scale_load(15)=15*1024 3232 * on an 8-core system with 8 tasks each runnable on one CPU shares has 3233 * to be 15*1024*1/8=1920 instead of scale_load(MIN_SHARES)=2*1024. In 3234 * case no task is runnable on a CPU MIN_SHARES=2 should be returned 3235 * instead of 0. 3236 */ 3237 return clamp_t(long, shares, MIN_SHARES, tg_shares); 3238 } 3239 #endif /* CONFIG_SMP */ 3240 3241 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq); 3242 3243 /* 3244 * Recomputes the group entity based on the current state of its group 3245 * runqueue. 3246 */ 3247 static void update_cfs_group(struct sched_entity *se) 3248 { 3249 struct cfs_rq *gcfs_rq = group_cfs_rq(se); 3250 long shares; 3251 3252 if (!gcfs_rq) 3253 return; 3254 3255 if (throttled_hierarchy(gcfs_rq)) 3256 return; 3257 3258 #ifndef CONFIG_SMP 3259 shares = READ_ONCE(gcfs_rq->tg->shares); 3260 3261 if (likely(se->load.weight == shares)) 3262 return; 3263 #else 3264 shares = calc_group_shares(gcfs_rq); 3265 #endif 3266 3267 reweight_entity(cfs_rq_of(se), se, shares); 3268 } 3269 3270 #else /* CONFIG_FAIR_GROUP_SCHED */ 3271 static inline void update_cfs_group(struct sched_entity *se) 3272 { 3273 } 3274 #endif /* CONFIG_FAIR_GROUP_SCHED */ 3275 3276 static inline void cfs_rq_util_change(struct cfs_rq *cfs_rq, int flags) 3277 { 3278 struct rq *rq = rq_of(cfs_rq); 3279 3280 if (&rq->cfs == cfs_rq) { 3281 /* 3282 * There are a few boundary cases this might miss but it should 3283 * get called often enough that that should (hopefully) not be 3284 * a real problem. 3285 * 3286 * It will not get called when we go idle, because the idle 3287 * thread is a different class (!fair), nor will the utilization 3288 * number include things like RT tasks. 3289 * 3290 * As is, the util number is not freq-invariant (we'd have to 3291 * implement arch_scale_freq_capacity() for that). 3292 * 3293 * See cpu_util(). 3294 */ 3295 cpufreq_update_util(rq, flags); 3296 } 3297 } 3298 3299 #ifdef CONFIG_SMP 3300 #ifdef CONFIG_FAIR_GROUP_SCHED 3301 /** 3302 * update_tg_load_avg - update the tg's load avg 3303 * @cfs_rq: the cfs_rq whose avg changed 3304 * 3305 * This function 'ensures': tg->load_avg := \Sum tg->cfs_rq[]->avg.load. 3306 * However, because tg->load_avg is a global value there are performance 3307 * considerations. 3308 * 3309 * In order to avoid having to look at the other cfs_rq's, we use a 3310 * differential update where we store the last value we propagated. This in 3311 * turn allows skipping updates if the differential is 'small'. 3312 * 3313 * Updating tg's load_avg is necessary before update_cfs_share(). 3314 */ 3315 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq) 3316 { 3317 long delta = cfs_rq->avg.load_avg - cfs_rq->tg_load_avg_contrib; 3318 3319 /* 3320 * No need to update load_avg for root_task_group as it is not used. 3321 */ 3322 if (cfs_rq->tg == &root_task_group) 3323 return; 3324 3325 if (abs(delta) > cfs_rq->tg_load_avg_contrib / 64) { 3326 atomic_long_add(delta, &cfs_rq->tg->load_avg); 3327 cfs_rq->tg_load_avg_contrib = cfs_rq->avg.load_avg; 3328 } 3329 } 3330 3331 /* 3332 * Called within set_task_rq() right before setting a task's CPU. The 3333 * caller only guarantees p->pi_lock is held; no other assumptions, 3334 * including the state of rq->lock, should be made. 3335 */ 3336 void set_task_rq_fair(struct sched_entity *se, 3337 struct cfs_rq *prev, struct cfs_rq *next) 3338 { 3339 u64 p_last_update_time; 3340 u64 n_last_update_time; 3341 3342 if (!sched_feat(ATTACH_AGE_LOAD)) 3343 return; 3344 3345 /* 3346 * We are supposed to update the task to "current" time, then its up to 3347 * date and ready to go to new CPU/cfs_rq. But we have difficulty in 3348 * getting what current time is, so simply throw away the out-of-date 3349 * time. This will result in the wakee task is less decayed, but giving 3350 * the wakee more load sounds not bad. 3351 */ 3352 if (!(se->avg.last_update_time && prev)) 3353 return; 3354 3355 #ifndef CONFIG_64BIT 3356 { 3357 u64 p_last_update_time_copy; 3358 u64 n_last_update_time_copy; 3359 3360 do { 3361 p_last_update_time_copy = prev->load_last_update_time_copy; 3362 n_last_update_time_copy = next->load_last_update_time_copy; 3363 3364 smp_rmb(); 3365 3366 p_last_update_time = prev->avg.last_update_time; 3367 n_last_update_time = next->avg.last_update_time; 3368 3369 } while (p_last_update_time != p_last_update_time_copy || 3370 n_last_update_time != n_last_update_time_copy); 3371 } 3372 #else 3373 p_last_update_time = prev->avg.last_update_time; 3374 n_last_update_time = next->avg.last_update_time; 3375 #endif 3376 __update_load_avg_blocked_se(p_last_update_time, se); 3377 se->avg.last_update_time = n_last_update_time; 3378 } 3379 3380 3381 /* 3382 * When on migration a sched_entity joins/leaves the PELT hierarchy, we need to 3383 * propagate its contribution. The key to this propagation is the invariant 3384 * that for each group: 3385 * 3386 * ge->avg == grq->avg (1) 3387 * 3388 * _IFF_ we look at the pure running and runnable sums. Because they 3389 * represent the very same entity, just at different points in the hierarchy. 3390 * 3391 * Per the above update_tg_cfs_util() and update_tg_cfs_runnable() are trivial 3392 * and simply copies the running/runnable sum over (but still wrong, because 3393 * the group entity and group rq do not have their PELT windows aligned). 3394 * 3395 * However, update_tg_cfs_load() is more complex. So we have: 3396 * 3397 * ge->avg.load_avg = ge->load.weight * ge->avg.runnable_avg (2) 3398 * 3399 * And since, like util, the runnable part should be directly transferable, 3400 * the following would _appear_ to be the straight forward approach: 3401 * 3402 * grq->avg.load_avg = grq->load.weight * grq->avg.runnable_avg (3) 3403 * 3404 * And per (1) we have: 3405 * 3406 * ge->avg.runnable_avg == grq->avg.runnable_avg 3407 * 3408 * Which gives: 3409 * 3410 * ge->load.weight * grq->avg.load_avg 3411 * ge->avg.load_avg = ----------------------------------- (4) 3412 * grq->load.weight 3413 * 3414 * Except that is wrong! 3415 * 3416 * Because while for entities historical weight is not important and we 3417 * really only care about our future and therefore can consider a pure 3418 * runnable sum, runqueues can NOT do this. 3419 * 3420 * We specifically want runqueues to have a load_avg that includes 3421 * historical weights. Those represent the blocked load, the load we expect 3422 * to (shortly) return to us. This only works by keeping the weights as 3423 * integral part of the sum. We therefore cannot decompose as per (3). 3424 * 3425 * Another reason this doesn't work is that runnable isn't a 0-sum entity. 3426 * Imagine a rq with 2 tasks that each are runnable 2/3 of the time. Then the 3427 * rq itself is runnable anywhere between 2/3 and 1 depending on how the 3428 * runnable section of these tasks overlap (or not). If they were to perfectly 3429 * align the rq as a whole would be runnable 2/3 of the time. If however we 3430 * always have at least 1 runnable task, the rq as a whole is always runnable. 3431 * 3432 * So we'll have to approximate.. :/ 3433 * 3434 * Given the constraint: 3435 * 3436 * ge->avg.running_sum <= ge->avg.runnable_sum <= LOAD_AVG_MAX 3437 * 3438 * We can construct a rule that adds runnable to a rq by assuming minimal 3439 * overlap. 3440 * 3441 * On removal, we'll assume each task is equally runnable; which yields: 3442 * 3443 * grq->avg.runnable_sum = grq->avg.load_sum / grq->load.weight 3444 * 3445 * XXX: only do this for the part of runnable > running ? 3446 * 3447 */ 3448 3449 static inline void 3450 update_tg_cfs_util(struct cfs_rq *cfs_rq, struct sched_entity *se, struct cfs_rq *gcfs_rq) 3451 { 3452 long delta = gcfs_rq->avg.util_avg - se->avg.util_avg; 3453 u32 divider; 3454 3455 /* Nothing to update */ 3456 if (!delta) 3457 return; 3458 3459 /* 3460 * cfs_rq->avg.period_contrib can be used for both cfs_rq and se. 3461 * See ___update_load_avg() for details. 3462 */ 3463 divider = get_pelt_divider(&cfs_rq->avg); 3464 3465 /* Set new sched_entity's utilization */ 3466 se->avg.util_avg = gcfs_rq->avg.util_avg; 3467 se->avg.util_sum = se->avg.util_avg * divider; 3468 3469 /* Update parent cfs_rq utilization */ 3470 add_positive(&cfs_rq->avg.util_avg, delta); 3471 cfs_rq->avg.util_sum = cfs_rq->avg.util_avg * divider; 3472 } 3473 3474 static inline void 3475 update_tg_cfs_runnable(struct cfs_rq *cfs_rq, struct sched_entity *se, struct cfs_rq *gcfs_rq) 3476 { 3477 long delta = gcfs_rq->avg.runnable_avg - se->avg.runnable_avg; 3478 u32 divider; 3479 3480 /* Nothing to update */ 3481 if (!delta) 3482 return; 3483 3484 /* 3485 * cfs_rq->avg.period_contrib can be used for both cfs_rq and se. 3486 * See ___update_load_avg() for details. 3487 */ 3488 divider = get_pelt_divider(&cfs_rq->avg); 3489 3490 /* Set new sched_entity's runnable */ 3491 se->avg.runnable_avg = gcfs_rq->avg.runnable_avg; 3492 se->avg.runnable_sum = se->avg.runnable_avg * divider; 3493 3494 /* Update parent cfs_rq runnable */ 3495 add_positive(&cfs_rq->avg.runnable_avg, delta); 3496 cfs_rq->avg.runnable_sum = cfs_rq->avg.runnable_avg * divider; 3497 } 3498 3499 static inline void 3500 update_tg_cfs_load(struct cfs_rq *cfs_rq, struct sched_entity *se, struct cfs_rq *gcfs_rq) 3501 { 3502 long delta, running_sum, runnable_sum = gcfs_rq->prop_runnable_sum; 3503 unsigned long load_avg; 3504 u64 load_sum = 0; 3505 u32 divider; 3506 3507 if (!runnable_sum) 3508 return; 3509 3510 gcfs_rq->prop_runnable_sum = 0; 3511 3512 /* 3513 * cfs_rq->avg.period_contrib can be used for both cfs_rq and se. 3514 * See ___update_load_avg() for details. 3515 */ 3516 divider = get_pelt_divider(&cfs_rq->avg); 3517 3518 if (runnable_sum >= 0) { 3519 /* 3520 * Add runnable; clip at LOAD_AVG_MAX. Reflects that until 3521 * the CPU is saturated running == runnable. 3522 */ 3523 runnable_sum += se->avg.load_sum; 3524 runnable_sum = min_t(long, runnable_sum, divider); 3525 } else { 3526 /* 3527 * Estimate the new unweighted runnable_sum of the gcfs_rq by 3528 * assuming all tasks are equally runnable. 3529 */ 3530 if (scale_load_down(gcfs_rq->load.weight)) { 3531 load_sum = div_s64(gcfs_rq->avg.load_sum, 3532 scale_load_down(gcfs_rq->load.weight)); 3533 } 3534 3535 /* But make sure to not inflate se's runnable */ 3536 runnable_sum = min(se->avg.load_sum, load_sum); 3537 } 3538 3539 /* 3540 * runnable_sum can't be lower than running_sum 3541 * Rescale running sum to be in the same range as runnable sum 3542 * running_sum is in [0 : LOAD_AVG_MAX << SCHED_CAPACITY_SHIFT] 3543 * runnable_sum is in [0 : LOAD_AVG_MAX] 3544 */ 3545 running_sum = se->avg.util_sum >> SCHED_CAPACITY_SHIFT; 3546 runnable_sum = max(runnable_sum, running_sum); 3547 3548 load_sum = (s64)se_weight(se) * runnable_sum; 3549 load_avg = div_s64(load_sum, divider); 3550 3551 delta = load_avg - se->avg.load_avg; 3552 3553 se->avg.load_sum = runnable_sum; 3554 se->avg.load_avg = load_avg; 3555 3556 add_positive(&cfs_rq->avg.load_avg, delta); 3557 cfs_rq->avg.load_sum = cfs_rq->avg.load_avg * divider; 3558 } 3559 3560 static inline void add_tg_cfs_propagate(struct cfs_rq *cfs_rq, long runnable_sum) 3561 { 3562 cfs_rq->propagate = 1; 3563 cfs_rq->prop_runnable_sum += runnable_sum; 3564 } 3565 3566 /* Update task and its cfs_rq load average */ 3567 static inline int propagate_entity_load_avg(struct sched_entity *se) 3568 { 3569 struct cfs_rq *cfs_rq, *gcfs_rq; 3570 3571 if (entity_is_task(se)) 3572 return 0; 3573 3574 gcfs_rq = group_cfs_rq(se); 3575 if (!gcfs_rq->propagate) 3576 return 0; 3577 3578 gcfs_rq->propagate = 0; 3579 3580 cfs_rq = cfs_rq_of(se); 3581 3582 add_tg_cfs_propagate(cfs_rq, gcfs_rq->prop_runnable_sum); 3583 3584 update_tg_cfs_util(cfs_rq, se, gcfs_rq); 3585 update_tg_cfs_runnable(cfs_rq, se, gcfs_rq); 3586 update_tg_cfs_load(cfs_rq, se, gcfs_rq); 3587 3588 trace_pelt_cfs_tp(cfs_rq); 3589 trace_pelt_se_tp(se); 3590 3591 return 1; 3592 } 3593 3594 /* 3595 * Check if we need to update the load and the utilization of a blocked 3596 * group_entity: 3597 */ 3598 static inline bool skip_blocked_update(struct sched_entity *se) 3599 { 3600 struct cfs_rq *gcfs_rq = group_cfs_rq(se); 3601 3602 /* 3603 * If sched_entity still have not zero load or utilization, we have to 3604 * decay it: 3605 */ 3606 if (se->avg.load_avg || se->avg.util_avg) 3607 return false; 3608 3609 /* 3610 * If there is a pending propagation, we have to update the load and 3611 * the utilization of the sched_entity: 3612 */ 3613 if (gcfs_rq->propagate) 3614 return false; 3615 3616 /* 3617 * Otherwise, the load and the utilization of the sched_entity is 3618 * already zero and there is no pending propagation, so it will be a 3619 * waste of time to try to decay it: 3620 */ 3621 return true; 3622 } 3623 3624 #else /* CONFIG_FAIR_GROUP_SCHED */ 3625 3626 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq) {} 3627 3628 static inline int propagate_entity_load_avg(struct sched_entity *se) 3629 { 3630 return 0; 3631 } 3632 3633 static inline void add_tg_cfs_propagate(struct cfs_rq *cfs_rq, long runnable_sum) {} 3634 3635 #endif /* CONFIG_FAIR_GROUP_SCHED */ 3636 3637 /** 3638 * update_cfs_rq_load_avg - update the cfs_rq's load/util averages 3639 * @now: current time, as per cfs_rq_clock_pelt() 3640 * @cfs_rq: cfs_rq to update 3641 * 3642 * The cfs_rq avg is the direct sum of all its entities (blocked and runnable) 3643 * avg. The immediate corollary is that all (fair) tasks must be attached, see 3644 * post_init_entity_util_avg(). 3645 * 3646 * cfs_rq->avg is used for task_h_load() and update_cfs_share() for example. 3647 * 3648 * Returns true if the load decayed or we removed load. 3649 * 3650 * Since both these conditions indicate a changed cfs_rq->avg.load we should 3651 * call update_tg_load_avg() when this function returns true. 3652 */ 3653 static inline int 3654 update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq) 3655 { 3656 unsigned long removed_load = 0, removed_util = 0, removed_runnable = 0; 3657 struct sched_avg *sa = &cfs_rq->avg; 3658 int decayed = 0; 3659 3660 if (cfs_rq->removed.nr) { 3661 unsigned long r; 3662 u32 divider = get_pelt_divider(&cfs_rq->avg); 3663 3664 raw_spin_lock(&cfs_rq->removed.lock); 3665 swap(cfs_rq->removed.util_avg, removed_util); 3666 swap(cfs_rq->removed.load_avg, removed_load); 3667 swap(cfs_rq->removed.runnable_avg, removed_runnable); 3668 cfs_rq->removed.nr = 0; 3669 raw_spin_unlock(&cfs_rq->removed.lock); 3670 3671 r = removed_load; 3672 sub_positive(&sa->load_avg, r); 3673 sub_positive(&sa->load_sum, r * divider); 3674 3675 r = removed_util; 3676 sub_positive(&sa->util_avg, r); 3677 sub_positive(&sa->util_sum, r * divider); 3678 3679 r = removed_runnable; 3680 sub_positive(&sa->runnable_avg, r); 3681 sub_positive(&sa->runnable_sum, r * divider); 3682 3683 /* 3684 * removed_runnable is the unweighted version of removed_load so we 3685 * can use it to estimate removed_load_sum. 3686 */ 3687 add_tg_cfs_propagate(cfs_rq, 3688 -(long)(removed_runnable * divider) >> SCHED_CAPACITY_SHIFT); 3689 3690 decayed = 1; 3691 } 3692 3693 decayed |= __update_load_avg_cfs_rq(now, cfs_rq); 3694 3695 #ifndef CONFIG_64BIT 3696 smp_wmb(); 3697 cfs_rq->load_last_update_time_copy = sa->last_update_time; 3698 #endif 3699 3700 return decayed; 3701 } 3702 3703 /** 3704 * attach_entity_load_avg - attach this entity to its cfs_rq load avg 3705 * @cfs_rq: cfs_rq to attach to 3706 * @se: sched_entity to attach 3707 * 3708 * Must call update_cfs_rq_load_avg() before this, since we rely on 3709 * cfs_rq->avg.last_update_time being current. 3710 */ 3711 static void attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) 3712 { 3713 /* 3714 * cfs_rq->avg.period_contrib can be used for both cfs_rq and se. 3715 * See ___update_load_avg() for details. 3716 */ 3717 u32 divider = get_pelt_divider(&cfs_rq->avg); 3718 3719 /* 3720 * When we attach the @se to the @cfs_rq, we must align the decay 3721 * window because without that, really weird and wonderful things can 3722 * happen. 3723 * 3724 * XXX illustrate 3725 */ 3726 se->avg.last_update_time = cfs_rq->avg.last_update_time; 3727 se->avg.period_contrib = cfs_rq->avg.period_contrib; 3728 3729 /* 3730 * Hell(o) Nasty stuff.. we need to recompute _sum based on the new 3731 * period_contrib. This isn't strictly correct, but since we're 3732 * entirely outside of the PELT hierarchy, nobody cares if we truncate 3733 * _sum a little. 3734 */ 3735 se->avg.util_sum = se->avg.util_avg * divider; 3736 3737 se->avg.runnable_sum = se->avg.runnable_avg * divider; 3738 3739 se->avg.load_sum = divider; 3740 if (se_weight(se)) { 3741 se->avg.load_sum = 3742 div_u64(se->avg.load_avg * se->avg.load_sum, se_weight(se)); 3743 } 3744 3745 enqueue_load_avg(cfs_rq, se); 3746 cfs_rq->avg.util_avg += se->avg.util_avg; 3747 cfs_rq->avg.util_sum += se->avg.util_sum; 3748 cfs_rq->avg.runnable_avg += se->avg.runnable_avg; 3749 cfs_rq->avg.runnable_sum += se->avg.runnable_sum; 3750 3751 add_tg_cfs_propagate(cfs_rq, se->avg.load_sum); 3752 3753 cfs_rq_util_change(cfs_rq, 0); 3754 3755 trace_pelt_cfs_tp(cfs_rq); 3756 } 3757 3758 /** 3759 * detach_entity_load_avg - detach this entity from its cfs_rq load avg 3760 * @cfs_rq: cfs_rq to detach from 3761 * @se: sched_entity to detach 3762 * 3763 * Must call update_cfs_rq_load_avg() before this, since we rely on 3764 * cfs_rq->avg.last_update_time being current. 3765 */ 3766 static void detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) 3767 { 3768 /* 3769 * cfs_rq->avg.period_contrib can be used for both cfs_rq and se. 3770 * See ___update_load_avg() for details. 3771 */ 3772 u32 divider = get_pelt_divider(&cfs_rq->avg); 3773 3774 dequeue_load_avg(cfs_rq, se); 3775 sub_positive(&cfs_rq->avg.util_avg, se->avg.util_avg); 3776 cfs_rq->avg.util_sum = cfs_rq->avg.util_avg * divider; 3777 sub_positive(&cfs_rq->avg.runnable_avg, se->avg.runnable_avg); 3778 cfs_rq->avg.runnable_sum = cfs_rq->avg.runnable_avg * divider; 3779 3780 add_tg_cfs_propagate(cfs_rq, -se->avg.load_sum); 3781 3782 cfs_rq_util_change(cfs_rq, 0); 3783 3784 trace_pelt_cfs_tp(cfs_rq); 3785 } 3786 3787 /* 3788 * Optional action to be done while updating the load average 3789 */ 3790 #define UPDATE_TG 0x1 3791 #define SKIP_AGE_LOAD 0x2 3792 #define DO_ATTACH 0x4 3793 3794 /* Update task and its cfs_rq load average */ 3795 static inline void update_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags) 3796 { 3797 u64 now = cfs_rq_clock_pelt(cfs_rq); 3798 int decayed; 3799 3800 /* 3801 * Track task load average for carrying it to new CPU after migrated, and 3802 * track group sched_entity load average for task_h_load calc in migration 3803 */ 3804 if (se->avg.last_update_time && !(flags & SKIP_AGE_LOAD)) 3805 __update_load_avg_se(now, cfs_rq, se); 3806 3807 decayed = update_cfs_rq_load_avg(now, cfs_rq); 3808 decayed |= propagate_entity_load_avg(se); 3809 3810 if (!se->avg.last_update_time && (flags & DO_ATTACH)) { 3811 3812 /* 3813 * DO_ATTACH means we're here from enqueue_entity(). 3814 * !last_update_time means we've passed through 3815 * migrate_task_rq_fair() indicating we migrated. 3816 * 3817 * IOW we're enqueueing a task on a new CPU. 3818 */ 3819 attach_entity_load_avg(cfs_rq, se); 3820 update_tg_load_avg(cfs_rq); 3821 3822 } else if (decayed) { 3823 cfs_rq_util_change(cfs_rq, 0); 3824 3825 if (flags & UPDATE_TG) 3826 update_tg_load_avg(cfs_rq); 3827 } 3828 } 3829 3830 #ifndef CONFIG_64BIT 3831 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq) 3832 { 3833 u64 last_update_time_copy; 3834 u64 last_update_time; 3835 3836 do { 3837 last_update_time_copy = cfs_rq->load_last_update_time_copy; 3838 smp_rmb(); 3839 last_update_time = cfs_rq->avg.last_update_time; 3840 } while (last_update_time != last_update_time_copy); 3841 3842 return last_update_time; 3843 } 3844 #else 3845 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq) 3846 { 3847 return cfs_rq->avg.last_update_time; 3848 } 3849 #endif 3850 3851 /* 3852 * Synchronize entity load avg of dequeued entity without locking 3853 * the previous rq. 3854 */ 3855 static void sync_entity_load_avg(struct sched_entity *se) 3856 { 3857 struct cfs_rq *cfs_rq = cfs_rq_of(se); 3858 u64 last_update_time; 3859 3860 last_update_time = cfs_rq_last_update_time(cfs_rq); 3861 __update_load_avg_blocked_se(last_update_time, se); 3862 } 3863 3864 /* 3865 * Task first catches up with cfs_rq, and then subtract 3866 * itself from the cfs_rq (task must be off the queue now). 3867 */ 3868 static void remove_entity_load_avg(struct sched_entity *se) 3869 { 3870 struct cfs_rq *cfs_rq = cfs_rq_of(se); 3871 unsigned long flags; 3872 3873 /* 3874 * tasks cannot exit without having gone through wake_up_new_task() -> 3875 * post_init_entity_util_avg() which will have added things to the 3876 * cfs_rq, so we can remove unconditionally. 3877 */ 3878 3879 sync_entity_load_avg(se); 3880 3881 raw_spin_lock_irqsave(&cfs_rq->removed.lock, flags); 3882 ++cfs_rq->removed.nr; 3883 cfs_rq->removed.util_avg += se->avg.util_avg; 3884 cfs_rq->removed.load_avg += se->avg.load_avg; 3885 cfs_rq->removed.runnable_avg += se->avg.runnable_avg; 3886 raw_spin_unlock_irqrestore(&cfs_rq->removed.lock, flags); 3887 } 3888 3889 static inline unsigned long cfs_rq_runnable_avg(struct cfs_rq *cfs_rq) 3890 { 3891 return cfs_rq->avg.runnable_avg; 3892 } 3893 3894 static inline unsigned long cfs_rq_load_avg(struct cfs_rq *cfs_rq) 3895 { 3896 return cfs_rq->avg.load_avg; 3897 } 3898 3899 static int newidle_balance(struct rq *this_rq, struct rq_flags *rf); 3900 3901 static inline unsigned long task_util(struct task_struct *p) 3902 { 3903 return READ_ONCE(p->se.avg.util_avg); 3904 } 3905 3906 static inline unsigned long _task_util_est(struct task_struct *p) 3907 { 3908 struct util_est ue = READ_ONCE(p->se.avg.util_est); 3909 3910 return max(ue.ewma, (ue.enqueued & ~UTIL_AVG_UNCHANGED)); 3911 } 3912 3913 static inline unsigned long task_util_est(struct task_struct *p) 3914 { 3915 return max(task_util(p), _task_util_est(p)); 3916 } 3917 3918 #ifdef CONFIG_UCLAMP_TASK 3919 static inline unsigned long uclamp_task_util(struct task_struct *p) 3920 { 3921 return clamp(task_util_est(p), 3922 uclamp_eff_value(p, UCLAMP_MIN), 3923 uclamp_eff_value(p, UCLAMP_MAX)); 3924 } 3925 #else 3926 static inline unsigned long uclamp_task_util(struct task_struct *p) 3927 { 3928 return task_util_est(p); 3929 } 3930 #endif 3931 3932 static inline void util_est_enqueue(struct cfs_rq *cfs_rq, 3933 struct task_struct *p) 3934 { 3935 unsigned int enqueued; 3936 3937 if (!sched_feat(UTIL_EST)) 3938 return; 3939 3940 /* Update root cfs_rq's estimated utilization */ 3941 enqueued = cfs_rq->avg.util_est.enqueued; 3942 enqueued += _task_util_est(p); 3943 WRITE_ONCE(cfs_rq->avg.util_est.enqueued, enqueued); 3944 3945 trace_sched_util_est_cfs_tp(cfs_rq); 3946 } 3947 3948 static inline void util_est_dequeue(struct cfs_rq *cfs_rq, 3949 struct task_struct *p) 3950 { 3951 unsigned int enqueued; 3952 3953 if (!sched_feat(UTIL_EST)) 3954 return; 3955 3956 /* Update root cfs_rq's estimated utilization */ 3957 enqueued = cfs_rq->avg.util_est.enqueued; 3958 enqueued -= min_t(unsigned int, enqueued, _task_util_est(p)); 3959 WRITE_ONCE(cfs_rq->avg.util_est.enqueued, enqueued); 3960 3961 trace_sched_util_est_cfs_tp(cfs_rq); 3962 } 3963 3964 #define UTIL_EST_MARGIN (SCHED_CAPACITY_SCALE / 100) 3965 3966 /* 3967 * Check if a (signed) value is within a specified (unsigned) margin, 3968 * based on the observation that: 3969 * 3970 * abs(x) < y := (unsigned)(x + y - 1) < (2 * y - 1) 3971 * 3972 * NOTE: this only works when value + margin < INT_MAX. 3973 */ 3974 static inline bool within_margin(int value, int margin) 3975 { 3976 return ((unsigned int)(value + margin - 1) < (2 * margin - 1)); 3977 } 3978 3979 static inline void util_est_update(struct cfs_rq *cfs_rq, 3980 struct task_struct *p, 3981 bool task_sleep) 3982 { 3983 long last_ewma_diff, last_enqueued_diff; 3984 struct util_est ue; 3985 3986 if (!sched_feat(UTIL_EST)) 3987 return; 3988 3989 /* 3990 * Skip update of task's estimated utilization when the task has not 3991 * yet completed an activation, e.g. being migrated. 3992 */ 3993 if (!task_sleep) 3994 return; 3995 3996 /* 3997 * If the PELT values haven't changed since enqueue time, 3998 * skip the util_est update. 3999 */ 4000 ue = p->se.avg.util_est; 4001 if (ue.enqueued & UTIL_AVG_UNCHANGED) 4002 return; 4003 4004 last_enqueued_diff = ue.enqueued; 4005 4006 /* 4007 * Reset EWMA on utilization increases, the moving average is used only 4008 * to smooth utilization decreases. 4009 */ 4010 ue.enqueued = task_util(p); 4011 if (sched_feat(UTIL_EST_FASTUP)) { 4012 if (ue.ewma < ue.enqueued) { 4013 ue.ewma = ue.enqueued; 4014 goto done; 4015 } 4016 } 4017 4018 /* 4019 * Skip update of task's estimated utilization when its members are 4020 * already ~1% close to its last activation value. 4021 */ 4022 last_ewma_diff = ue.enqueued - ue.ewma; 4023 last_enqueued_diff -= ue.enqueued; 4024 if (within_margin(last_ewma_diff, UTIL_EST_MARGIN)) { 4025 if (!within_margin(last_enqueued_diff, UTIL_EST_MARGIN)) 4026 goto done; 4027 4028 return; 4029 } 4030 4031 /* 4032 * To avoid overestimation of actual task utilization, skip updates if 4033 * we cannot grant there is idle time in this CPU. 4034 */ 4035 if (task_util(p) > capacity_orig_of(cpu_of(rq_of(cfs_rq)))) 4036 return; 4037 4038 /* 4039 * Update Task's estimated utilization 4040 * 4041 * When *p completes an activation we can consolidate another sample 4042 * of the task size. This is done by storing the current PELT value 4043 * as ue.enqueued and by using this value to update the Exponential 4044 * Weighted Moving Average (EWMA): 4045 * 4046 * ewma(t) = w * task_util(p) + (1-w) * ewma(t-1) 4047 * = w * task_util(p) + ewma(t-1) - w * ewma(t-1) 4048 * = w * (task_util(p) - ewma(t-1)) + ewma(t-1) 4049 * = w * ( last_ewma_diff ) + ewma(t-1) 4050 * = w * (last_ewma_diff + ewma(t-1) / w) 4051 * 4052 * Where 'w' is the weight of new samples, which is configured to be 4053 * 0.25, thus making w=1/4 ( >>= UTIL_EST_WEIGHT_SHIFT) 4054 */ 4055 ue.ewma <<= UTIL_EST_WEIGHT_SHIFT; 4056 ue.ewma += last_ewma_diff; 4057 ue.ewma >>= UTIL_EST_WEIGHT_SHIFT; 4058 done: 4059 ue.enqueued |= UTIL_AVG_UNCHANGED; 4060 WRITE_ONCE(p->se.avg.util_est, ue); 4061 4062 trace_sched_util_est_se_tp(&p->se); 4063 } 4064 4065 static inline int task_fits_capacity(struct task_struct *p, long capacity) 4066 { 4067 return fits_capacity(uclamp_task_util(p), capacity); 4068 } 4069 4070 static inline void update_misfit_status(struct task_struct *p, struct rq *rq) 4071 { 4072 if (!static_branch_unlikely(&sched_asym_cpucapacity)) 4073 return; 4074 4075 if (!p || p->nr_cpus_allowed == 1) { 4076 rq->misfit_task_load = 0; 4077 return; 4078 } 4079 4080 if (task_fits_capacity(p, capacity_of(cpu_of(rq)))) { 4081 rq->misfit_task_load = 0; 4082 return; 4083 } 4084 4085 /* 4086 * Make sure that misfit_task_load will not be null even if 4087 * task_h_load() returns 0. 4088 */ 4089 rq->misfit_task_load = max_t(unsigned long, task_h_load(p), 1); 4090 } 4091 4092 #else /* CONFIG_SMP */ 4093 4094 #define UPDATE_TG 0x0 4095 #define SKIP_AGE_LOAD 0x0 4096 #define DO_ATTACH 0x0 4097 4098 static inline void update_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int not_used1) 4099 { 4100 cfs_rq_util_change(cfs_rq, 0); 4101 } 4102 4103 static inline void remove_entity_load_avg(struct sched_entity *se) {} 4104 4105 static inline void 4106 attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {} 4107 static inline void 4108 detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {} 4109 4110 static inline int newidle_balance(struct rq *rq, struct rq_flags *rf) 4111 { 4112 return 0; 4113 } 4114 4115 static inline void 4116 util_est_enqueue(struct cfs_rq *cfs_rq, struct task_struct *p) {} 4117 4118 static inline void 4119 util_est_dequeue(struct cfs_rq *cfs_rq, struct task_struct *p) {} 4120 4121 static inline void 4122 util_est_update(struct cfs_rq *cfs_rq, struct task_struct *p, 4123 bool task_sleep) {} 4124 static inline void update_misfit_status(struct task_struct *p, struct rq *rq) {} 4125 4126 #endif /* CONFIG_SMP */ 4127 4128 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se) 4129 { 4130 #ifdef CONFIG_SCHED_DEBUG 4131 s64 d = se->vruntime - cfs_rq->min_vruntime; 4132 4133 if (d < 0) 4134 d = -d; 4135 4136 if (d > 3*sysctl_sched_latency) 4137 schedstat_inc(cfs_rq->nr_spread_over); 4138 #endif 4139 } 4140 4141 static void 4142 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial) 4143 { 4144 u64 vruntime = cfs_rq->min_vruntime; 4145 4146 /* 4147 * The 'current' period is already promised to the current tasks, 4148 * however the extra weight of the new task will slow them down a 4149 * little, place the new task so that it fits in the slot that 4150 * stays open at the end. 4151 */ 4152 if (initial && sched_feat(START_DEBIT)) 4153 vruntime += sched_vslice(cfs_rq, se); 4154 4155 /* sleeps up to a single latency don't count. */ 4156 if (!initial) { 4157 unsigned long thresh = sysctl_sched_latency; 4158 4159 /* 4160 * Halve their sleep time's effect, to allow 4161 * for a gentler effect of sleepers: 4162 */ 4163 if (sched_feat(GENTLE_FAIR_SLEEPERS)) 4164 thresh >>= 1; 4165 4166 vruntime -= thresh; 4167 } 4168 4169 /* ensure we never gain time by being placed backwards. */ 4170 se->vruntime = max_vruntime(se->vruntime, vruntime); 4171 } 4172 4173 static void check_enqueue_throttle(struct cfs_rq *cfs_rq); 4174 4175 static inline void check_schedstat_required(void) 4176 { 4177 #ifdef CONFIG_SCHEDSTATS 4178 if (schedstat_enabled()) 4179 return; 4180 4181 /* Force schedstat enabled if a dependent tracepoint is active */ 4182 if (trace_sched_stat_wait_enabled() || 4183 trace_sched_stat_sleep_enabled() || 4184 trace_sched_stat_iowait_enabled() || 4185 trace_sched_stat_blocked_enabled() || 4186 trace_sched_stat_runtime_enabled()) { 4187 printk_deferred_once("Scheduler tracepoints stat_sleep, stat_iowait, " 4188 "stat_blocked and stat_runtime require the " 4189 "kernel parameter schedstats=enable or " 4190 "kernel.sched_schedstats=1\n"); 4191 } 4192 #endif 4193 } 4194 4195 static inline bool cfs_bandwidth_used(void); 4196 4197 /* 4198 * MIGRATION 4199 * 4200 * dequeue 4201 * update_curr() 4202 * update_min_vruntime() 4203 * vruntime -= min_vruntime 4204 * 4205 * enqueue 4206 * update_curr() 4207 * update_min_vruntime() 4208 * vruntime += min_vruntime 4209 * 4210 * this way the vruntime transition between RQs is done when both 4211 * min_vruntime are up-to-date. 4212 * 4213 * WAKEUP (remote) 4214 * 4215 * ->migrate_task_rq_fair() (p->state == TASK_WAKING) 4216 * vruntime -= min_vruntime 4217 * 4218 * enqueue 4219 * update_curr() 4220 * update_min_vruntime() 4221 * vruntime += min_vruntime 4222 * 4223 * this way we don't have the most up-to-date min_vruntime on the originating 4224 * CPU and an up-to-date min_vruntime on the destination CPU. 4225 */ 4226 4227 static void 4228 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags) 4229 { 4230 bool renorm = !(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_MIGRATED); 4231 bool curr = cfs_rq->curr == se; 4232 4233 /* 4234 * If we're the current task, we must renormalise before calling 4235 * update_curr(). 4236 */ 4237 if (renorm && curr) 4238 se->vruntime += cfs_rq->min_vruntime; 4239 4240 update_curr(cfs_rq); 4241 4242 /* 4243 * Otherwise, renormalise after, such that we're placed at the current 4244 * moment in time, instead of some random moment in the past. Being 4245 * placed in the past could significantly boost this task to the 4246 * fairness detriment of existing tasks. 4247 */ 4248 if (renorm && !curr) 4249 se->vruntime += cfs_rq->min_vruntime; 4250 4251 /* 4252 * When enqueuing a sched_entity, we must: 4253 * - Update loads to have both entity and cfs_rq synced with now. 4254 * - Add its load to cfs_rq->runnable_avg 4255 * - For group_entity, update its weight to reflect the new share of 4256 * its group cfs_rq 4257 * - Add its new weight to cfs_rq->load.weight 4258 */ 4259 update_load_avg(cfs_rq, se, UPDATE_TG | DO_ATTACH); 4260 se_update_runnable(se); 4261 update_cfs_group(se); 4262 account_entity_enqueue(cfs_rq, se); 4263 4264 if (flags & ENQUEUE_WAKEUP) 4265 place_entity(cfs_rq, se, 0); 4266 4267 check_schedstat_required(); 4268 update_stats_enqueue(cfs_rq, se, flags); 4269 check_spread(cfs_rq, se); 4270 if (!curr) 4271 __enqueue_entity(cfs_rq, se); 4272 se->on_rq = 1; 4273 4274 /* 4275 * When bandwidth control is enabled, cfs might have been removed 4276 * because of a parent been throttled but cfs->nr_running > 1. Try to 4277 * add it unconditionally. 4278 */ 4279 if (cfs_rq->nr_running == 1 || cfs_bandwidth_used()) 4280 list_add_leaf_cfs_rq(cfs_rq); 4281 4282 if (cfs_rq->nr_running == 1) 4283 check_enqueue_throttle(cfs_rq); 4284 } 4285 4286 static void __clear_buddies_last(struct sched_entity *se) 4287 { 4288 for_each_sched_entity(se) { 4289 struct cfs_rq *cfs_rq = cfs_rq_of(se); 4290 if (cfs_rq->last != se) 4291 break; 4292 4293 cfs_rq->last = NULL; 4294 } 4295 } 4296 4297 static void __clear_buddies_next(struct sched_entity *se) 4298 { 4299 for_each_sched_entity(se) { 4300 struct cfs_rq *cfs_rq = cfs_rq_of(se); 4301 if (cfs_rq->next != se) 4302 break; 4303 4304 cfs_rq->next = NULL; 4305 } 4306 } 4307 4308 static void __clear_buddies_skip(struct sched_entity *se) 4309 { 4310 for_each_sched_entity(se) { 4311 struct cfs_rq *cfs_rq = cfs_rq_of(se); 4312 if (cfs_rq->skip != se) 4313 break; 4314 4315 cfs_rq->skip = NULL; 4316 } 4317 } 4318 4319 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se) 4320 { 4321 if (cfs_rq->last == se) 4322 __clear_buddies_last(se); 4323 4324 if (cfs_rq->next == se) 4325 __clear_buddies_next(se); 4326 4327 if (cfs_rq->skip == se) 4328 __clear_buddies_skip(se); 4329 } 4330 4331 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq); 4332 4333 static void 4334 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags) 4335 { 4336 /* 4337 * Update run-time statistics of the 'current'. 4338 */ 4339 update_curr(cfs_rq); 4340 4341 /* 4342 * When dequeuing a sched_entity, we must: 4343 * - Update loads to have both entity and cfs_rq synced with now. 4344 * - Subtract its load from the cfs_rq->runnable_avg. 4345 * - Subtract its previous weight from cfs_rq->load.weight. 4346 * - For group entity, update its weight to reflect the new share 4347 * of its group cfs_rq. 4348 */ 4349 update_load_avg(cfs_rq, se, UPDATE_TG); 4350 se_update_runnable(se); 4351 4352 update_stats_dequeue(cfs_rq, se, flags); 4353 4354 clear_buddies(cfs_rq, se); 4355 4356 if (se != cfs_rq->curr) 4357 __dequeue_entity(cfs_rq, se); 4358 se->on_rq = 0; 4359 account_entity_dequeue(cfs_rq, se); 4360 4361 /* 4362 * Normalize after update_curr(); which will also have moved 4363 * min_vruntime if @se is the one holding it back. But before doing 4364 * update_min_vruntime() again, which will discount @se's position and 4365 * can move min_vruntime forward still more. 4366 */ 4367 if (!(flags & DEQUEUE_SLEEP)) 4368 se->vruntime -= cfs_rq->min_vruntime; 4369 4370 /* return excess runtime on last dequeue */ 4371 return_cfs_rq_runtime(cfs_rq); 4372 4373 update_cfs_group(se); 4374 4375 /* 4376 * Now advance min_vruntime if @se was the entity holding it back, 4377 * except when: DEQUEUE_SAVE && !DEQUEUE_MOVE, in this case we'll be 4378 * put back on, and if we advance min_vruntime, we'll be placed back 4379 * further than we started -- ie. we'll be penalized. 4380 */ 4381 if ((flags & (DEQUEUE_SAVE | DEQUEUE_MOVE)) != DEQUEUE_SAVE) 4382 update_min_vruntime(cfs_rq); 4383 } 4384 4385 /* 4386 * Preempt the current task with a newly woken task if needed: 4387 */ 4388 static void 4389 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr) 4390 { 4391 unsigned long ideal_runtime, delta_exec; 4392 struct sched_entity *se; 4393 s64 delta; 4394 4395 ideal_runtime = sched_slice(cfs_rq, curr); 4396 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime; 4397 if (delta_exec > ideal_runtime) { 4398 resched_curr(rq_of(cfs_rq)); 4399 /* 4400 * The current task ran long enough, ensure it doesn't get 4401 * re-elected due to buddy favours. 4402 */ 4403 clear_buddies(cfs_rq, curr); 4404 return; 4405 } 4406 4407 /* 4408 * Ensure that a task that missed wakeup preemption by a 4409 * narrow margin doesn't have to wait for a full slice. 4410 * This also mitigates buddy induced latencies under load. 4411 */ 4412 if (delta_exec < sysctl_sched_min_granularity) 4413 return; 4414 4415 se = __pick_first_entity(cfs_rq); 4416 delta = curr->vruntime - se->vruntime; 4417 4418 if (delta < 0) 4419 return; 4420 4421 if (delta > ideal_runtime) 4422 resched_curr(rq_of(cfs_rq)); 4423 } 4424 4425 static void 4426 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se) 4427 { 4428 /* 'current' is not kept within the tree. */ 4429 if (se->on_rq) { 4430 /* 4431 * Any task has to be enqueued before it get to execute on 4432 * a CPU. So account for the time it spent waiting on the 4433 * runqueue. 4434 */ 4435 update_stats_wait_end(cfs_rq, se); 4436 __dequeue_entity(cfs_rq, se); 4437 update_load_avg(cfs_rq, se, UPDATE_TG); 4438 } 4439 4440 update_stats_curr_start(cfs_rq, se); 4441 cfs_rq->curr = se; 4442 4443 /* 4444 * Track our maximum slice length, if the CPU's load is at 4445 * least twice that of our own weight (i.e. dont track it 4446 * when there are only lesser-weight tasks around): 4447 */ 4448 if (schedstat_enabled() && 4449 rq_of(cfs_rq)->cfs.load.weight >= 2*se->load.weight) { 4450 schedstat_set(se->statistics.slice_max, 4451 max((u64)schedstat_val(se->statistics.slice_max), 4452 se->sum_exec_runtime - se->prev_sum_exec_runtime)); 4453 } 4454 4455 se->prev_sum_exec_runtime = se->sum_exec_runtime; 4456 } 4457 4458 static int 4459 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se); 4460 4461 /* 4462 * Pick the next process, keeping these things in mind, in this order: 4463 * 1) keep things fair between processes/task groups 4464 * 2) pick the "next" process, since someone really wants that to run 4465 * 3) pick the "last" process, for cache locality 4466 * 4) do not run the "skip" process, if something else is available 4467 */ 4468 static struct sched_entity * 4469 pick_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *curr) 4470 { 4471 struct sched_entity *left = __pick_first_entity(cfs_rq); 4472 struct sched_entity *se; 4473 4474 /* 4475 * If curr is set we have to see if its left of the leftmost entity 4476 * still in the tree, provided there was anything in the tree at all. 4477 */ 4478 if (!left || (curr && entity_before(curr, left))) 4479 left = curr; 4480 4481 se = left; /* ideally we run the leftmost entity */ 4482 4483 /* 4484 * Avoid running the skip buddy, if running something else can 4485 * be done without getting too unfair. 4486 */ 4487 if (cfs_rq->skip == se) { 4488 struct sched_entity *second; 4489 4490 if (se == curr) { 4491 second = __pick_first_entity(cfs_rq); 4492 } else { 4493 second = __pick_next_entity(se); 4494 if (!second || (curr && entity_before(curr, second))) 4495 second = curr; 4496 } 4497 4498 if (second && wakeup_preempt_entity(second, left) < 1) 4499 se = second; 4500 } 4501 4502 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1) { 4503 /* 4504 * Someone really wants this to run. If it's not unfair, run it. 4505 */ 4506 se = cfs_rq->next; 4507 } else if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1) { 4508 /* 4509 * Prefer last buddy, try to return the CPU to a preempted task. 4510 */ 4511 se = cfs_rq->last; 4512 } 4513 4514 clear_buddies(cfs_rq, se); 4515 4516 return se; 4517 } 4518 4519 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq); 4520 4521 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev) 4522 { 4523 /* 4524 * If still on the runqueue then deactivate_task() 4525 * was not called and update_curr() has to be done: 4526 */ 4527 if (prev->on_rq) 4528 update_curr(cfs_rq); 4529 4530 /* throttle cfs_rqs exceeding runtime */ 4531 check_cfs_rq_runtime(cfs_rq); 4532 4533 check_spread(cfs_rq, prev); 4534 4535 if (prev->on_rq) { 4536 update_stats_wait_start(cfs_rq, prev); 4537 /* Put 'current' back into the tree. */ 4538 __enqueue_entity(cfs_rq, prev); 4539 /* in !on_rq case, update occurred at dequeue */ 4540 update_load_avg(cfs_rq, prev, 0); 4541 } 4542 cfs_rq->curr = NULL; 4543 } 4544 4545 static void 4546 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued) 4547 { 4548 /* 4549 * Update run-time statistics of the 'current'. 4550 */ 4551 update_curr(cfs_rq); 4552 4553 /* 4554 * Ensure that runnable average is periodically updated. 4555 */ 4556 update_load_avg(cfs_rq, curr, UPDATE_TG); 4557 update_cfs_group(curr); 4558 4559 #ifdef CONFIG_SCHED_HRTICK 4560 /* 4561 * queued ticks are scheduled to match the slice, so don't bother 4562 * validating it and just reschedule. 4563 */ 4564 if (queued) { 4565 resched_curr(rq_of(cfs_rq)); 4566 return; 4567 } 4568 /* 4569 * don't let the period tick interfere with the hrtick preemption 4570 */ 4571 if (!sched_feat(DOUBLE_TICK) && 4572 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer)) 4573 return; 4574 #endif 4575 4576 if (cfs_rq->nr_running > 1) 4577 check_preempt_tick(cfs_rq, curr); 4578 } 4579 4580 4581 /************************************************** 4582 * CFS bandwidth control machinery 4583 */ 4584 4585 #ifdef CONFIG_CFS_BANDWIDTH 4586 4587 #ifdef CONFIG_JUMP_LABEL 4588 static struct static_key __cfs_bandwidth_used; 4589 4590 static inline bool cfs_bandwidth_used(void) 4591 { 4592 return static_key_false(&__cfs_bandwidth_used); 4593 } 4594 4595 void cfs_bandwidth_usage_inc(void) 4596 { 4597 static_key_slow_inc_cpuslocked(&__cfs_bandwidth_used); 4598 } 4599 4600 void cfs_bandwidth_usage_dec(void) 4601 { 4602 static_key_slow_dec_cpuslocked(&__cfs_bandwidth_used); 4603 } 4604 #else /* CONFIG_JUMP_LABEL */ 4605 static bool cfs_bandwidth_used(void) 4606 { 4607 return true; 4608 } 4609 4610 void cfs_bandwidth_usage_inc(void) {} 4611 void cfs_bandwidth_usage_dec(void) {} 4612 #endif /* CONFIG_JUMP_LABEL */ 4613 4614 /* 4615 * default period for cfs group bandwidth. 4616 * default: 0.1s, units: nanoseconds 4617 */ 4618 static inline u64 default_cfs_period(void) 4619 { 4620 return 100000000ULL; 4621 } 4622 4623 static inline u64 sched_cfs_bandwidth_slice(void) 4624 { 4625 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC; 4626 } 4627 4628 /* 4629 * Replenish runtime according to assigned quota. We use sched_clock_cpu 4630 * directly instead of rq->clock to avoid adding additional synchronization 4631 * around rq->lock. 4632 * 4633 * requires cfs_b->lock 4634 */ 4635 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b) 4636 { 4637 if (cfs_b->quota != RUNTIME_INF) 4638 cfs_b->runtime = cfs_b->quota; 4639 } 4640 4641 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg) 4642 { 4643 return &tg->cfs_bandwidth; 4644 } 4645 4646 /* returns 0 on failure to allocate runtime */ 4647 static int __assign_cfs_rq_runtime(struct cfs_bandwidth *cfs_b, 4648 struct cfs_rq *cfs_rq, u64 target_runtime) 4649 { 4650 u64 min_amount, amount = 0; 4651 4652 lockdep_assert_held(&cfs_b->lock); 4653 4654 /* note: this is a positive sum as runtime_remaining <= 0 */ 4655 min_amount = target_runtime - cfs_rq->runtime_remaining; 4656 4657 if (cfs_b->quota == RUNTIME_INF) 4658 amount = min_amount; 4659 else { 4660 start_cfs_bandwidth(cfs_b); 4661 4662 if (cfs_b->runtime > 0) { 4663 amount = min(cfs_b->runtime, min_amount); 4664 cfs_b->runtime -= amount; 4665 cfs_b->idle = 0; 4666 } 4667 } 4668 4669 cfs_rq->runtime_remaining += amount; 4670 4671 return cfs_rq->runtime_remaining > 0; 4672 } 4673 4674 /* returns 0 on failure to allocate runtime */ 4675 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq) 4676 { 4677 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg); 4678 int ret; 4679 4680 raw_spin_lock(&cfs_b->lock); 4681 ret = __assign_cfs_rq_runtime(cfs_b, cfs_rq, sched_cfs_bandwidth_slice()); 4682 raw_spin_unlock(&cfs_b->lock); 4683 4684 return ret; 4685 } 4686 4687 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) 4688 { 4689 /* dock delta_exec before expiring quota (as it could span periods) */ 4690 cfs_rq->runtime_remaining -= delta_exec; 4691 4692 if (likely(cfs_rq->runtime_remaining > 0)) 4693 return; 4694 4695 if (cfs_rq->throttled) 4696 return; 4697 /* 4698 * if we're unable to extend our runtime we resched so that the active 4699 * hierarchy can be throttled 4700 */ 4701 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr)) 4702 resched_curr(rq_of(cfs_rq)); 4703 } 4704 4705 static __always_inline 4706 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) 4707 { 4708 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled) 4709 return; 4710 4711 __account_cfs_rq_runtime(cfs_rq, delta_exec); 4712 } 4713 4714 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq) 4715 { 4716 return cfs_bandwidth_used() && cfs_rq->throttled; 4717 } 4718 4719 /* check whether cfs_rq, or any parent, is throttled */ 4720 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq) 4721 { 4722 return cfs_bandwidth_used() && cfs_rq->throttle_count; 4723 } 4724 4725 /* 4726 * Ensure that neither of the group entities corresponding to src_cpu or 4727 * dest_cpu are members of a throttled hierarchy when performing group 4728 * load-balance operations. 4729 */ 4730 static inline int throttled_lb_pair(struct task_group *tg, 4731 int src_cpu, int dest_cpu) 4732 { 4733 struct cfs_rq *src_cfs_rq, *dest_cfs_rq; 4734 4735 src_cfs_rq = tg->cfs_rq[src_cpu]; 4736 dest_cfs_rq = tg->cfs_rq[dest_cpu]; 4737 4738 return throttled_hierarchy(src_cfs_rq) || 4739 throttled_hierarchy(dest_cfs_rq); 4740 } 4741 4742 static int tg_unthrottle_up(struct task_group *tg, void *data) 4743 { 4744 struct rq *rq = data; 4745 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)]; 4746 4747 cfs_rq->throttle_count--; 4748 if (!cfs_rq->throttle_count) { 4749 cfs_rq->throttled_clock_task_time += rq_clock_task(rq) - 4750 cfs_rq->throttled_clock_task; 4751 4752 /* Add cfs_rq with already running entity in the list */ 4753 if (cfs_rq->nr_running >= 1) 4754 list_add_leaf_cfs_rq(cfs_rq); 4755 } 4756 4757 return 0; 4758 } 4759 4760 static int tg_throttle_down(struct task_group *tg, void *data) 4761 { 4762 struct rq *rq = data; 4763 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)]; 4764 4765 /* group is entering throttled state, stop time */ 4766 if (!cfs_rq->throttle_count) { 4767 cfs_rq->throttled_clock_task = rq_clock_task(rq); 4768 list_del_leaf_cfs_rq(cfs_rq); 4769 } 4770 cfs_rq->throttle_count++; 4771 4772 return 0; 4773 } 4774 4775 static bool throttle_cfs_rq(struct cfs_rq *cfs_rq) 4776 { 4777 struct rq *rq = rq_of(cfs_rq); 4778 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg); 4779 struct sched_entity *se; 4780 long task_delta, idle_task_delta, dequeue = 1; 4781 4782 raw_spin_lock(&cfs_b->lock); 4783 /* This will start the period timer if necessary */ 4784 if (__assign_cfs_rq_runtime(cfs_b, cfs_rq, 1)) { 4785 /* 4786 * We have raced with bandwidth becoming available, and if we 4787 * actually throttled the timer might not unthrottle us for an 4788 * entire period. We additionally needed to make sure that any 4789 * subsequent check_cfs_rq_runtime calls agree not to throttle 4790 * us, as we may commit to do cfs put_prev+pick_next, so we ask 4791 * for 1ns of runtime rather than just check cfs_b. 4792 */ 4793 dequeue = 0; 4794 } else { 4795 list_add_tail_rcu(&cfs_rq->throttled_list, 4796 &cfs_b->throttled_cfs_rq); 4797 } 4798 raw_spin_unlock(&cfs_b->lock); 4799 4800 if (!dequeue) 4801 return false; /* Throttle no longer required. */ 4802 4803 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))]; 4804 4805 /* freeze hierarchy runnable averages while throttled */ 4806 rcu_read_lock(); 4807 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq); 4808 rcu_read_unlock(); 4809 4810 task_delta = cfs_rq->h_nr_running; 4811 idle_task_delta = cfs_rq->idle_h_nr_running; 4812 for_each_sched_entity(se) { 4813 struct cfs_rq *qcfs_rq = cfs_rq_of(se); 4814 /* throttled entity or throttle-on-deactivate */ 4815 if (!se->on_rq) 4816 goto done; 4817 4818 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP); 4819 4820 qcfs_rq->h_nr_running -= task_delta; 4821 qcfs_rq->idle_h_nr_running -= idle_task_delta; 4822 4823 if (qcfs_rq->load.weight) { 4824 /* Avoid re-evaluating load for this entity: */ 4825 se = parent_entity(se); 4826 break; 4827 } 4828 } 4829 4830 for_each_sched_entity(se) { 4831 struct cfs_rq *qcfs_rq = cfs_rq_of(se); 4832 /* throttled entity or throttle-on-deactivate */ 4833 if (!se->on_rq) 4834 goto done; 4835 4836 update_load_avg(qcfs_rq, se, 0); 4837 se_update_runnable(se); 4838 4839 qcfs_rq->h_nr_running -= task_delta; 4840 qcfs_rq->idle_h_nr_running -= idle_task_delta; 4841 } 4842 4843 /* At this point se is NULL and we are at root level*/ 4844 sub_nr_running(rq, task_delta); 4845 4846 done: 4847 /* 4848 * Note: distribution will already see us throttled via the 4849 * throttled-list. rq->lock protects completion. 4850 */ 4851 cfs_rq->throttled = 1; 4852 cfs_rq->throttled_clock = rq_clock(rq); 4853 return true; 4854 } 4855 4856 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq) 4857 { 4858 struct rq *rq = rq_of(cfs_rq); 4859 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg); 4860 struct sched_entity *se; 4861 long task_delta, idle_task_delta; 4862 4863 se = cfs_rq->tg->se[cpu_of(rq)]; 4864 4865 cfs_rq->throttled = 0; 4866 4867 update_rq_clock(rq); 4868 4869 raw_spin_lock(&cfs_b->lock); 4870 cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock; 4871 list_del_rcu(&cfs_rq->throttled_list); 4872 raw_spin_unlock(&cfs_b->lock); 4873 4874 /* update hierarchical throttle state */ 4875 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq); 4876 4877 if (!cfs_rq->load.weight) 4878 return; 4879 4880 task_delta = cfs_rq->h_nr_running; 4881 idle_task_delta = cfs_rq->idle_h_nr_running; 4882 for_each_sched_entity(se) { 4883 if (se->on_rq) 4884 break; 4885 cfs_rq = cfs_rq_of(se); 4886 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP); 4887 4888 cfs_rq->h_nr_running += task_delta; 4889 cfs_rq->idle_h_nr_running += idle_task_delta; 4890 4891 /* end evaluation on encountering a throttled cfs_rq */ 4892 if (cfs_rq_throttled(cfs_rq)) 4893 goto unthrottle_throttle; 4894 } 4895 4896 for_each_sched_entity(se) { 4897 cfs_rq = cfs_rq_of(se); 4898 4899 update_load_avg(cfs_rq, se, UPDATE_TG); 4900 se_update_runnable(se); 4901 4902 cfs_rq->h_nr_running += task_delta; 4903 cfs_rq->idle_h_nr_running += idle_task_delta; 4904 4905 4906 /* end evaluation on encountering a throttled cfs_rq */ 4907 if (cfs_rq_throttled(cfs_rq)) 4908 goto unthrottle_throttle; 4909 4910 /* 4911 * One parent has been throttled and cfs_rq removed from the 4912 * list. Add it back to not break the leaf list. 4913 */ 4914 if (throttled_hierarchy(cfs_rq)) 4915 list_add_leaf_cfs_rq(cfs_rq); 4916 } 4917 4918 /* At this point se is NULL and we are at root level*/ 4919 add_nr_running(rq, task_delta); 4920 4921 unthrottle_throttle: 4922 /* 4923 * The cfs_rq_throttled() breaks in the above iteration can result in 4924 * incomplete leaf list maintenance, resulting in triggering the 4925 * assertion below. 4926 */ 4927 for_each_sched_entity(se) { 4928 cfs_rq = cfs_rq_of(se); 4929 4930 if (list_add_leaf_cfs_rq(cfs_rq)) 4931 break; 4932 } 4933 4934 assert_list_leaf_cfs_rq(rq); 4935 4936 /* Determine whether we need to wake up potentially idle CPU: */ 4937 if (rq->curr == rq->idle && rq->cfs.nr_running) 4938 resched_curr(rq); 4939 } 4940 4941 static void distribute_cfs_runtime(struct cfs_bandwidth *cfs_b) 4942 { 4943 struct cfs_rq *cfs_rq; 4944 u64 runtime, remaining = 1; 4945 4946 rcu_read_lock(); 4947 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq, 4948 throttled_list) { 4949 struct rq *rq = rq_of(cfs_rq); 4950 struct rq_flags rf; 4951 4952 rq_lock_irqsave(rq, &rf); 4953 if (!cfs_rq_throttled(cfs_rq)) 4954 goto next; 4955 4956 /* By the above check, this should never be true */ 4957 SCHED_WARN_ON(cfs_rq->runtime_remaining > 0); 4958 4959 raw_spin_lock(&cfs_b->lock); 4960 runtime = -cfs_rq->runtime_remaining + 1; 4961 if (runtime > cfs_b->runtime) 4962 runtime = cfs_b->runtime; 4963 cfs_b->runtime -= runtime; 4964 remaining = cfs_b->runtime; 4965 raw_spin_unlock(&cfs_b->lock); 4966 4967 cfs_rq->runtime_remaining += runtime; 4968 4969 /* we check whether we're throttled above */ 4970 if (cfs_rq->runtime_remaining > 0) 4971 unthrottle_cfs_rq(cfs_rq); 4972 4973 next: 4974 rq_unlock_irqrestore(rq, &rf); 4975 4976 if (!remaining) 4977 break; 4978 } 4979 rcu_read_unlock(); 4980 } 4981 4982 /* 4983 * Responsible for refilling a task_group's bandwidth and unthrottling its 4984 * cfs_rqs as appropriate. If there has been no activity within the last 4985 * period the timer is deactivated until scheduling resumes; cfs_b->idle is 4986 * used to track this state. 4987 */ 4988 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun, unsigned long flags) 4989 { 4990 int throttled; 4991 4992 /* no need to continue the timer with no bandwidth constraint */ 4993 if (cfs_b->quota == RUNTIME_INF) 4994 goto out_deactivate; 4995 4996 throttled = !list_empty(&cfs_b->throttled_cfs_rq); 4997 cfs_b->nr_periods += overrun; 4998 4999 /* 5000 * idle depends on !throttled (for the case of a large deficit), and if 5001 * we're going inactive then everything else can be deferred 5002 */ 5003 if (cfs_b->idle && !throttled) 5004 goto out_deactivate; 5005 5006 __refill_cfs_bandwidth_runtime(cfs_b); 5007 5008 if (!throttled) { 5009 /* mark as potentially idle for the upcoming period */ 5010 cfs_b->idle = 1; 5011 return 0; 5012 } 5013 5014 /* account preceding periods in which throttling occurred */ 5015 cfs_b->nr_throttled += overrun; 5016 5017 /* 5018 * This check is repeated as we release cfs_b->lock while we unthrottle. 5019 */ 5020 while (throttled && cfs_b->runtime > 0) { 5021 raw_spin_unlock_irqrestore(&cfs_b->lock, flags); 5022 /* we can't nest cfs_b->lock while distributing bandwidth */ 5023 distribute_cfs_runtime(cfs_b); 5024 raw_spin_lock_irqsave(&cfs_b->lock, flags); 5025 5026 throttled = !list_empty(&cfs_b->throttled_cfs_rq); 5027 } 5028 5029 /* 5030 * While we are ensured activity in the period following an 5031 * unthrottle, this also covers the case in which the new bandwidth is 5032 * insufficient to cover the existing bandwidth deficit. (Forcing the 5033 * timer to remain active while there are any throttled entities.) 5034 */ 5035 cfs_b->idle = 0; 5036 5037 return 0; 5038 5039 out_deactivate: 5040 return 1; 5041 } 5042 5043 /* a cfs_rq won't donate quota below this amount */ 5044 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC; 5045 /* minimum remaining period time to redistribute slack quota */ 5046 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC; 5047 /* how long we wait to gather additional slack before distributing */ 5048 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC; 5049 5050 /* 5051 * Are we near the end of the current quota period? 5052 * 5053 * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the 5054 * hrtimer base being cleared by hrtimer_start. In the case of 5055 * migrate_hrtimers, base is never cleared, so we are fine. 5056 */ 5057 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire) 5058 { 5059 struct hrtimer *refresh_timer = &cfs_b->period_timer; 5060 u64 remaining; 5061 5062 /* if the call-back is running a quota refresh is already occurring */ 5063 if (hrtimer_callback_running(refresh_timer)) 5064 return 1; 5065 5066 /* is a quota refresh about to occur? */ 5067 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer)); 5068 if (remaining < min_expire) 5069 return 1; 5070 5071 return 0; 5072 } 5073 5074 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b) 5075 { 5076 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration; 5077 5078 /* if there's a quota refresh soon don't bother with slack */ 5079 if (runtime_refresh_within(cfs_b, min_left)) 5080 return; 5081 5082 /* don't push forwards an existing deferred unthrottle */ 5083 if (cfs_b->slack_started) 5084 return; 5085 cfs_b->slack_started = true; 5086 5087 hrtimer_start(&cfs_b->slack_timer, 5088 ns_to_ktime(cfs_bandwidth_slack_period), 5089 HRTIMER_MODE_REL); 5090 } 5091 5092 /* we know any runtime found here is valid as update_curr() precedes return */ 5093 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq) 5094 { 5095 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg); 5096 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime; 5097 5098 if (slack_runtime <= 0) 5099 return; 5100 5101 raw_spin_lock(&cfs_b->lock); 5102 if (cfs_b->quota != RUNTIME_INF) { 5103 cfs_b->runtime += slack_runtime; 5104 5105 /* we are under rq->lock, defer unthrottling using a timer */ 5106 if (cfs_b->runtime > sched_cfs_bandwidth_slice() && 5107 !list_empty(&cfs_b->throttled_cfs_rq)) 5108 start_cfs_slack_bandwidth(cfs_b); 5109 } 5110 raw_spin_unlock(&cfs_b->lock); 5111 5112 /* even if it's not valid for return we don't want to try again */ 5113 cfs_rq->runtime_remaining -= slack_runtime; 5114 } 5115 5116 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) 5117 { 5118 if (!cfs_bandwidth_used()) 5119 return; 5120 5121 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running) 5122 return; 5123 5124 __return_cfs_rq_runtime(cfs_rq); 5125 } 5126 5127 /* 5128 * This is done with a timer (instead of inline with bandwidth return) since 5129 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs. 5130 */ 5131 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b) 5132 { 5133 u64 runtime = 0, slice = sched_cfs_bandwidth_slice(); 5134 unsigned long flags; 5135 5136 /* confirm we're still not at a refresh boundary */ 5137 raw_spin_lock_irqsave(&cfs_b->lock, flags); 5138 cfs_b->slack_started = false; 5139 5140 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) { 5141 raw_spin_unlock_irqrestore(&cfs_b->lock, flags); 5142 return; 5143 } 5144 5145 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice) 5146 runtime = cfs_b->runtime; 5147 5148 raw_spin_unlock_irqrestore(&cfs_b->lock, flags); 5149 5150 if (!runtime) 5151 return; 5152 5153 distribute_cfs_runtime(cfs_b); 5154 } 5155 5156 /* 5157 * When a group wakes up we want to make sure that its quota is not already 5158 * expired/exceeded, otherwise it may be allowed to steal additional ticks of 5159 * runtime as update_curr() throttling can not trigger until it's on-rq. 5160 */ 5161 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) 5162 { 5163 if (!cfs_bandwidth_used()) 5164 return; 5165 5166 /* an active group must be handled by the update_curr()->put() path */ 5167 if (!cfs_rq->runtime_enabled || cfs_rq->curr) 5168 return; 5169 5170 /* ensure the group is not already throttled */ 5171 if (cfs_rq_throttled(cfs_rq)) 5172 return; 5173 5174 /* update runtime allocation */ 5175 account_cfs_rq_runtime(cfs_rq, 0); 5176 if (cfs_rq->runtime_remaining <= 0) 5177 throttle_cfs_rq(cfs_rq); 5178 } 5179 5180 static void sync_throttle(struct task_group *tg, int cpu) 5181 { 5182 struct cfs_rq *pcfs_rq, *cfs_rq; 5183 5184 if (!cfs_bandwidth_used()) 5185 return; 5186 5187 if (!tg->parent) 5188 return; 5189 5190 cfs_rq = tg->cfs_rq[cpu]; 5191 pcfs_rq = tg->parent->cfs_rq[cpu]; 5192 5193 cfs_rq->throttle_count = pcfs_rq->throttle_count; 5194 cfs_rq->throttled_clock_task = rq_clock_task(cpu_rq(cpu)); 5195 } 5196 5197 /* conditionally throttle active cfs_rq's from put_prev_entity() */ 5198 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) 5199 { 5200 if (!cfs_bandwidth_used()) 5201 return false; 5202 5203 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0)) 5204 return false; 5205 5206 /* 5207 * it's possible for a throttled entity to be forced into a running 5208 * state (e.g. set_curr_task), in this case we're finished. 5209 */ 5210 if (cfs_rq_throttled(cfs_rq)) 5211 return true; 5212 5213 return throttle_cfs_rq(cfs_rq); 5214 } 5215 5216 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer) 5217 { 5218 struct cfs_bandwidth *cfs_b = 5219 container_of(timer, struct cfs_bandwidth, slack_timer); 5220 5221 do_sched_cfs_slack_timer(cfs_b); 5222 5223 return HRTIMER_NORESTART; 5224 } 5225 5226 extern const u64 max_cfs_quota_period; 5227 5228 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer) 5229 { 5230 struct cfs_bandwidth *cfs_b = 5231 container_of(timer, struct cfs_bandwidth, period_timer); 5232 unsigned long flags; 5233 int overrun; 5234 int idle = 0; 5235 int count = 0; 5236 5237 raw_spin_lock_irqsave(&cfs_b->lock, flags); 5238 for (;;) { 5239 overrun = hrtimer_forward_now(timer, cfs_b->period); 5240 if (!overrun) 5241 break; 5242 5243 idle = do_sched_cfs_period_timer(cfs_b, overrun, flags); 5244 5245 if (++count > 3) { 5246 u64 new, old = ktime_to_ns(cfs_b->period); 5247 5248 /* 5249 * Grow period by a factor of 2 to avoid losing precision. 5250 * Precision loss in the quota/period ratio can cause __cfs_schedulable 5251 * to fail. 5252 */ 5253 new = old * 2; 5254 if (new < max_cfs_quota_period) { 5255 cfs_b->period = ns_to_ktime(new); 5256 cfs_b->quota *= 2; 5257 5258 pr_warn_ratelimited( 5259 "cfs_period_timer[cpu%d]: period too short, scaling up (new cfs_period_us = %lld, cfs_quota_us = %lld)\n", 5260 smp_processor_id(), 5261 div_u64(new, NSEC_PER_USEC), 5262 div_u64(cfs_b->quota, NSEC_PER_USEC)); 5263 } else { 5264 pr_warn_ratelimited( 5265 "cfs_period_timer[cpu%d]: period too short, but cannot scale up without losing precision (cfs_period_us = %lld, cfs_quota_us = %lld)\n", 5266 smp_processor_id(), 5267 div_u64(old, NSEC_PER_USEC), 5268 div_u64(cfs_b->quota, NSEC_PER_USEC)); 5269 } 5270 5271 /* reset count so we don't come right back in here */ 5272 count = 0; 5273 } 5274 } 5275 if (idle) 5276 cfs_b->period_active = 0; 5277 raw_spin_unlock_irqrestore(&cfs_b->lock, flags); 5278 5279 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART; 5280 } 5281 5282 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) 5283 { 5284 raw_spin_lock_init(&cfs_b->lock); 5285 cfs_b->runtime = 0; 5286 cfs_b->quota = RUNTIME_INF; 5287 cfs_b->period = ns_to_ktime(default_cfs_period()); 5288 5289 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq); 5290 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_PINNED); 5291 cfs_b->period_timer.function = sched_cfs_period_timer; 5292 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); 5293 cfs_b->slack_timer.function = sched_cfs_slack_timer; 5294 cfs_b->slack_started = false; 5295 } 5296 5297 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) 5298 { 5299 cfs_rq->runtime_enabled = 0; 5300 INIT_LIST_HEAD(&cfs_rq->throttled_list); 5301 } 5302 5303 void start_cfs_bandwidth(struct cfs_bandwidth *cfs_b) 5304 { 5305 lockdep_assert_held(&cfs_b->lock); 5306 5307 if (cfs_b->period_active) 5308 return; 5309 5310 cfs_b->period_active = 1; 5311 hrtimer_forward_now(&cfs_b->period_timer, cfs_b->period); 5312 hrtimer_start_expires(&cfs_b->period_timer, HRTIMER_MODE_ABS_PINNED); 5313 } 5314 5315 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) 5316 { 5317 /* init_cfs_bandwidth() was not called */ 5318 if (!cfs_b->throttled_cfs_rq.next) 5319 return; 5320 5321 hrtimer_cancel(&cfs_b->period_timer); 5322 hrtimer_cancel(&cfs_b->slack_timer); 5323 } 5324 5325 /* 5326 * Both these CPU hotplug callbacks race against unregister_fair_sched_group() 5327 * 5328 * The race is harmless, since modifying bandwidth settings of unhooked group 5329 * bits doesn't do much. 5330 */ 5331 5332 /* cpu online callback */ 5333 static void __maybe_unused update_runtime_enabled(struct rq *rq) 5334 { 5335 struct task_group *tg; 5336 5337 lockdep_assert_held(&rq->lock); 5338 5339 rcu_read_lock(); 5340 list_for_each_entry_rcu(tg, &task_groups, list) { 5341 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth; 5342 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)]; 5343 5344 raw_spin_lock(&cfs_b->lock); 5345 cfs_rq->runtime_enabled = cfs_b->quota != RUNTIME_INF; 5346 raw_spin_unlock(&cfs_b->lock); 5347 } 5348 rcu_read_unlock(); 5349 } 5350 5351 /* cpu offline callback */ 5352 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq) 5353 { 5354 struct task_group *tg; 5355 5356 lockdep_assert_held(&rq->lock); 5357 5358 rcu_read_lock(); 5359 list_for_each_entry_rcu(tg, &task_groups, list) { 5360 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)]; 5361 5362 if (!cfs_rq->runtime_enabled) 5363 continue; 5364 5365 /* 5366 * clock_task is not advancing so we just need to make sure 5367 * there's some valid quota amount 5368 */ 5369 cfs_rq->runtime_remaining = 1; 5370 /* 5371 * Offline rq is schedulable till CPU is completely disabled 5372 * in take_cpu_down(), so we prevent new cfs throttling here. 5373 */ 5374 cfs_rq->runtime_enabled = 0; 5375 5376 if (cfs_rq_throttled(cfs_rq)) 5377 unthrottle_cfs_rq(cfs_rq); 5378 } 5379 rcu_read_unlock(); 5380 } 5381 5382 #else /* CONFIG_CFS_BANDWIDTH */ 5383 5384 static inline bool cfs_bandwidth_used(void) 5385 { 5386 return false; 5387 } 5388 5389 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) {} 5390 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) { return false; } 5391 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {} 5392 static inline void sync_throttle(struct task_group *tg, int cpu) {} 5393 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {} 5394 5395 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq) 5396 { 5397 return 0; 5398 } 5399 5400 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq) 5401 { 5402 return 0; 5403 } 5404 5405 static inline int throttled_lb_pair(struct task_group *tg, 5406 int src_cpu, int dest_cpu) 5407 { 5408 return 0; 5409 } 5410 5411 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {} 5412 5413 #ifdef CONFIG_FAIR_GROUP_SCHED 5414 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {} 5415 #endif 5416 5417 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg) 5418 { 5419 return NULL; 5420 } 5421 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {} 5422 static inline void update_runtime_enabled(struct rq *rq) {} 5423 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {} 5424 5425 #endif /* CONFIG_CFS_BANDWIDTH */ 5426 5427 /************************************************** 5428 * CFS operations on tasks: 5429 */ 5430 5431 #ifdef CONFIG_SCHED_HRTICK 5432 static void hrtick_start_fair(struct rq *rq, struct task_struct *p) 5433 { 5434 struct sched_entity *se = &p->se; 5435 struct cfs_rq *cfs_rq = cfs_rq_of(se); 5436 5437 SCHED_WARN_ON(task_rq(p) != rq); 5438 5439 if (rq->cfs.h_nr_running > 1) { 5440 u64 slice = sched_slice(cfs_rq, se); 5441 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime; 5442 s64 delta = slice - ran; 5443 5444 if (delta < 0) { 5445 if (task_current(rq, p)) 5446 resched_curr(rq); 5447 return; 5448 } 5449 hrtick_start(rq, delta); 5450 } 5451 } 5452 5453 /* 5454 * called from enqueue/dequeue and updates the hrtick when the 5455 * current task is from our class and nr_running is low enough 5456 * to matter. 5457 */ 5458 static void hrtick_update(struct rq *rq) 5459 { 5460 struct task_struct *curr = rq->curr; 5461 5462 if (!hrtick_enabled_fair(rq) || curr->sched_class != &fair_sched_class) 5463 return; 5464 5465 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency) 5466 hrtick_start_fair(rq, curr); 5467 } 5468 #else /* !CONFIG_SCHED_HRTICK */ 5469 static inline void 5470 hrtick_start_fair(struct rq *rq, struct task_struct *p) 5471 { 5472 } 5473 5474 static inline void hrtick_update(struct rq *rq) 5475 { 5476 } 5477 #endif 5478 5479 #ifdef CONFIG_SMP 5480 static inline unsigned long cpu_util(int cpu); 5481 5482 static inline bool cpu_overutilized(int cpu) 5483 { 5484 return !fits_capacity(cpu_util(cpu), capacity_of(cpu)); 5485 } 5486 5487 static inline void update_overutilized_status(struct rq *rq) 5488 { 5489 if (!READ_ONCE(rq->rd->overutilized) && cpu_overutilized(rq->cpu)) { 5490 WRITE_ONCE(rq->rd->overutilized, SG_OVERUTILIZED); 5491 trace_sched_overutilized_tp(rq->rd, SG_OVERUTILIZED); 5492 } 5493 } 5494 #else 5495 static inline void update_overutilized_status(struct rq *rq) { } 5496 #endif 5497 5498 /* Runqueue only has SCHED_IDLE tasks enqueued */ 5499 static int sched_idle_rq(struct rq *rq) 5500 { 5501 return unlikely(rq->nr_running == rq->cfs.idle_h_nr_running && 5502 rq->nr_running); 5503 } 5504 5505 #ifdef CONFIG_SMP 5506 static int sched_idle_cpu(int cpu) 5507 { 5508 return sched_idle_rq(cpu_rq(cpu)); 5509 } 5510 #endif 5511 5512 /* 5513 * The enqueue_task method is called before nr_running is 5514 * increased. Here we update the fair scheduling stats and 5515 * then put the task into the rbtree: 5516 */ 5517 static void 5518 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags) 5519 { 5520 struct cfs_rq *cfs_rq; 5521 struct sched_entity *se = &p->se; 5522 int idle_h_nr_running = task_has_idle_policy(p); 5523 int task_new = !(flags & ENQUEUE_WAKEUP); 5524 5525 /* 5526 * The code below (indirectly) updates schedutil which looks at 5527 * the cfs_rq utilization to select a frequency. 5528 * Let's add the task's estimated utilization to the cfs_rq's 5529 * estimated utilization, before we update schedutil. 5530 */ 5531 util_est_enqueue(&rq->cfs, p); 5532 5533 /* 5534 * If in_iowait is set, the code below may not trigger any cpufreq 5535 * utilization updates, so do it here explicitly with the IOWAIT flag 5536 * passed. 5537 */ 5538 if (p->in_iowait) 5539 cpufreq_update_util(rq, SCHED_CPUFREQ_IOWAIT); 5540 5541 for_each_sched_entity(se) { 5542 if (se->on_rq) 5543 break; 5544 cfs_rq = cfs_rq_of(se); 5545 enqueue_entity(cfs_rq, se, flags); 5546 5547 cfs_rq->h_nr_running++; 5548 cfs_rq->idle_h_nr_running += idle_h_nr_running; 5549 5550 /* end evaluation on encountering a throttled cfs_rq */ 5551 if (cfs_rq_throttled(cfs_rq)) 5552 goto enqueue_throttle; 5553 5554 flags = ENQUEUE_WAKEUP; 5555 } 5556 5557 for_each_sched_entity(se) { 5558 cfs_rq = cfs_rq_of(se); 5559 5560 update_load_avg(cfs_rq, se, UPDATE_TG); 5561 se_update_runnable(se); 5562 update_cfs_group(se); 5563 5564 cfs_rq->h_nr_running++; 5565 cfs_rq->idle_h_nr_running += idle_h_nr_running; 5566 5567 /* end evaluation on encountering a throttled cfs_rq */ 5568 if (cfs_rq_throttled(cfs_rq)) 5569 goto enqueue_throttle; 5570 5571 /* 5572 * One parent has been throttled and cfs_rq removed from the 5573 * list. Add it back to not break the leaf list. 5574 */ 5575 if (throttled_hierarchy(cfs_rq)) 5576 list_add_leaf_cfs_rq(cfs_rq); 5577 } 5578 5579 /* At this point se is NULL and we are at root level*/ 5580 add_nr_running(rq, 1); 5581 5582 /* 5583 * Since new tasks are assigned an initial util_avg equal to 5584 * half of the spare capacity of their CPU, tiny tasks have the 5585 * ability to cross the overutilized threshold, which will 5586 * result in the load balancer ruining all the task placement 5587 * done by EAS. As a way to mitigate that effect, do not account 5588 * for the first enqueue operation of new tasks during the 5589 * overutilized flag detection. 5590 * 5591 * A better way of solving this problem would be to wait for 5592 * the PELT signals of tasks to converge before taking them 5593 * into account, but that is not straightforward to implement, 5594 * and the following generally works well enough in practice. 5595 */ 5596 if (!task_new) 5597 update_overutilized_status(rq); 5598 5599 enqueue_throttle: 5600 if (cfs_bandwidth_used()) { 5601 /* 5602 * When bandwidth control is enabled; the cfs_rq_throttled() 5603 * breaks in the above iteration can result in incomplete 5604 * leaf list maintenance, resulting in triggering the assertion 5605 * below. 5606 */ 5607 for_each_sched_entity(se) { 5608 cfs_rq = cfs_rq_of(se); 5609 5610 if (list_add_leaf_cfs_rq(cfs_rq)) 5611 break; 5612 } 5613 } 5614 5615 assert_list_leaf_cfs_rq(rq); 5616 5617 hrtick_update(rq); 5618 } 5619 5620 static void set_next_buddy(struct sched_entity *se); 5621 5622 /* 5623 * The dequeue_task method is called before nr_running is 5624 * decreased. We remove the task from the rbtree and 5625 * update the fair scheduling stats: 5626 */ 5627 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags) 5628 { 5629 struct cfs_rq *cfs_rq; 5630 struct sched_entity *se = &p->se; 5631 int task_sleep = flags & DEQUEUE_SLEEP; 5632 int idle_h_nr_running = task_has_idle_policy(p); 5633 bool was_sched_idle = sched_idle_rq(rq); 5634 5635 util_est_dequeue(&rq->cfs, p); 5636 5637 for_each_sched_entity(se) { 5638 cfs_rq = cfs_rq_of(se); 5639 dequeue_entity(cfs_rq, se, flags); 5640 5641 cfs_rq->h_nr_running--; 5642 cfs_rq->idle_h_nr_running -= idle_h_nr_running; 5643 5644 /* end evaluation on encountering a throttled cfs_rq */ 5645 if (cfs_rq_throttled(cfs_rq)) 5646 goto dequeue_throttle; 5647 5648 /* Don't dequeue parent if it has other entities besides us */ 5649 if (cfs_rq->load.weight) { 5650 /* Avoid re-evaluating load for this entity: */ 5651 se = parent_entity(se); 5652 /* 5653 * Bias pick_next to pick a task from this cfs_rq, as 5654 * p is sleeping when it is within its sched_slice. 5655 */ 5656 if (task_sleep && se && !throttled_hierarchy(cfs_rq)) 5657 set_next_buddy(se); 5658 break; 5659 } 5660 flags |= DEQUEUE_SLEEP; 5661 } 5662 5663 for_each_sched_entity(se) { 5664 cfs_rq = cfs_rq_of(se); 5665 5666 update_load_avg(cfs_rq, se, UPDATE_TG); 5667 se_update_runnable(se); 5668 update_cfs_group(se); 5669 5670 cfs_rq->h_nr_running--; 5671 cfs_rq->idle_h_nr_running -= idle_h_nr_running; 5672 5673 /* end evaluation on encountering a throttled cfs_rq */ 5674 if (cfs_rq_throttled(cfs_rq)) 5675 goto dequeue_throttle; 5676 5677 } 5678 5679 /* At this point se is NULL and we are at root level*/ 5680 sub_nr_running(rq, 1); 5681 5682 /* balance early to pull high priority tasks */ 5683 if (unlikely(!was_sched_idle && sched_idle_rq(rq))) 5684 rq->next_balance = jiffies; 5685 5686 dequeue_throttle: 5687 util_est_update(&rq->cfs, p, task_sleep); 5688 hrtick_update(rq); 5689 } 5690 5691 #ifdef CONFIG_SMP 5692 5693 /* Working cpumask for: load_balance, load_balance_newidle. */ 5694 DEFINE_PER_CPU(cpumask_var_t, load_balance_mask); 5695 DEFINE_PER_CPU(cpumask_var_t, select_idle_mask); 5696 5697 #ifdef CONFIG_NO_HZ_COMMON 5698 5699 static struct { 5700 cpumask_var_t idle_cpus_mask; 5701 atomic_t nr_cpus; 5702 int has_blocked; /* Idle CPUS has blocked load */ 5703 unsigned long next_balance; /* in jiffy units */ 5704 unsigned long next_blocked; /* Next update of blocked load in jiffies */ 5705 } nohz ____cacheline_aligned; 5706 5707 #endif /* CONFIG_NO_HZ_COMMON */ 5708 5709 static unsigned long cpu_load(struct rq *rq) 5710 { 5711 return cfs_rq_load_avg(&rq->cfs); 5712 } 5713 5714 /* 5715 * cpu_load_without - compute CPU load without any contributions from *p 5716 * @cpu: the CPU which load is requested 5717 * @p: the task which load should be discounted 5718 * 5719 * The load of a CPU is defined by the load of tasks currently enqueued on that 5720 * CPU as well as tasks which are currently sleeping after an execution on that 5721 * CPU. 5722 * 5723 * This method returns the load of the specified CPU by discounting the load of 5724 * the specified task, whenever the task is currently contributing to the CPU 5725 * load. 5726 */ 5727 static unsigned long cpu_load_without(struct rq *rq, struct task_struct *p) 5728 { 5729 struct cfs_rq *cfs_rq; 5730 unsigned int load; 5731 5732 /* Task has no contribution or is new */ 5733 if (cpu_of(rq) != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time)) 5734 return cpu_load(rq); 5735 5736 cfs_rq = &rq->cfs; 5737 load = READ_ONCE(cfs_rq->avg.load_avg); 5738 5739 /* Discount task's util from CPU's util */ 5740 lsub_positive(&load, task_h_load(p)); 5741 5742 return load; 5743 } 5744 5745 static unsigned long cpu_runnable(struct rq *rq) 5746 { 5747 return cfs_rq_runnable_avg(&rq->cfs); 5748 } 5749 5750 static unsigned long cpu_runnable_without(struct rq *rq, struct task_struct *p) 5751 { 5752 struct cfs_rq *cfs_rq; 5753 unsigned int runnable; 5754 5755 /* Task has no contribution or is new */ 5756 if (cpu_of(rq) != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time)) 5757 return cpu_runnable(rq); 5758 5759 cfs_rq = &rq->cfs; 5760 runnable = READ_ONCE(cfs_rq->avg.runnable_avg); 5761 5762 /* Discount task's runnable from CPU's runnable */ 5763 lsub_positive(&runnable, p->se.avg.runnable_avg); 5764 5765 return runnable; 5766 } 5767 5768 static unsigned long capacity_of(int cpu) 5769 { 5770 return cpu_rq(cpu)->cpu_capacity; 5771 } 5772 5773 static void record_wakee(struct task_struct *p) 5774 { 5775 /* 5776 * Only decay a single time; tasks that have less then 1 wakeup per 5777 * jiffy will not have built up many flips. 5778 */ 5779 if (time_after(jiffies, current->wakee_flip_decay_ts + HZ)) { 5780 current->wakee_flips >>= 1; 5781 current->wakee_flip_decay_ts = jiffies; 5782 } 5783 5784 if (current->last_wakee != p) { 5785 current->last_wakee = p; 5786 current->wakee_flips++; 5787 } 5788 } 5789 5790 /* 5791 * Detect M:N waker/wakee relationships via a switching-frequency heuristic. 5792 * 5793 * A waker of many should wake a different task than the one last awakened 5794 * at a frequency roughly N times higher than one of its wakees. 5795 * 5796 * In order to determine whether we should let the load spread vs consolidating 5797 * to shared cache, we look for a minimum 'flip' frequency of llc_size in one 5798 * partner, and a factor of lls_size higher frequency in the other. 5799 * 5800 * With both conditions met, we can be relatively sure that the relationship is 5801 * non-monogamous, with partner count exceeding socket size. 5802 * 5803 * Waker/wakee being client/server, worker/dispatcher, interrupt source or 5804 * whatever is irrelevant, spread criteria is apparent partner count exceeds 5805 * socket size. 5806 */ 5807 static int wake_wide(struct task_struct *p) 5808 { 5809 unsigned int master = current->wakee_flips; 5810 unsigned int slave = p->wakee_flips; 5811 int factor = __this_cpu_read(sd_llc_size); 5812 5813 if (master < slave) 5814 swap(master, slave); 5815 if (slave < factor || master < slave * factor) 5816 return 0; 5817 return 1; 5818 } 5819 5820 /* 5821 * The purpose of wake_affine() is to quickly determine on which CPU we can run 5822 * soonest. For the purpose of speed we only consider the waking and previous 5823 * CPU. 5824 * 5825 * wake_affine_idle() - only considers 'now', it check if the waking CPU is 5826 * cache-affine and is (or will be) idle. 5827 * 5828 * wake_affine_weight() - considers the weight to reflect the average 5829 * scheduling latency of the CPUs. This seems to work 5830 * for the overloaded case. 5831 */ 5832 static int 5833 wake_affine_idle(int this_cpu, int prev_cpu, int sync) 5834 { 5835 /* 5836 * If this_cpu is idle, it implies the wakeup is from interrupt 5837 * context. Only allow the move if cache is shared. Otherwise an 5838 * interrupt intensive workload could force all tasks onto one 5839 * node depending on the IO topology or IRQ affinity settings. 5840 * 5841 * If the prev_cpu is idle and cache affine then avoid a migration. 5842 * There is no guarantee that the cache hot data from an interrupt 5843 * is more important than cache hot data on the prev_cpu and from 5844 * a cpufreq perspective, it's better to have higher utilisation 5845 * on one CPU. 5846 */ 5847 if (available_idle_cpu(this_cpu) && cpus_share_cache(this_cpu, prev_cpu)) 5848 return available_idle_cpu(prev_cpu) ? prev_cpu : this_cpu; 5849 5850 if (sync && cpu_rq(this_cpu)->nr_running == 1) 5851 return this_cpu; 5852 5853 if (available_idle_cpu(prev_cpu)) 5854 return prev_cpu; 5855 5856 return nr_cpumask_bits; 5857 } 5858 5859 static int 5860 wake_affine_weight(struct sched_domain *sd, struct task_struct *p, 5861 int this_cpu, int prev_cpu, int sync) 5862 { 5863 s64 this_eff_load, prev_eff_load; 5864 unsigned long task_load; 5865 5866 this_eff_load = cpu_load(cpu_rq(this_cpu)); 5867 5868 if (sync) { 5869 unsigned long current_load = task_h_load(current); 5870 5871 if (current_load > this_eff_load) 5872 return this_cpu; 5873 5874 this_eff_load -= current_load; 5875 } 5876 5877 task_load = task_h_load(p); 5878 5879 this_eff_load += task_load; 5880 if (sched_feat(WA_BIAS)) 5881 this_eff_load *= 100; 5882 this_eff_load *= capacity_of(prev_cpu); 5883 5884 prev_eff_load = cpu_load(cpu_rq(prev_cpu)); 5885 prev_eff_load -= task_load; 5886 if (sched_feat(WA_BIAS)) 5887 prev_eff_load *= 100 + (sd->imbalance_pct - 100) / 2; 5888 prev_eff_load *= capacity_of(this_cpu); 5889 5890 /* 5891 * If sync, adjust the weight of prev_eff_load such that if 5892 * prev_eff == this_eff that select_idle_sibling() will consider 5893 * stacking the wakee on top of the waker if no other CPU is 5894 * idle. 5895 */ 5896 if (sync) 5897 prev_eff_load += 1; 5898 5899 return this_eff_load < prev_eff_load ? this_cpu : nr_cpumask_bits; 5900 } 5901 5902 static int wake_affine(struct sched_domain *sd, struct task_struct *p, 5903 int this_cpu, int prev_cpu, int sync) 5904 { 5905 int target = nr_cpumask_bits; 5906 5907 if (sched_feat(WA_IDLE)) 5908 target = wake_affine_idle(this_cpu, prev_cpu, sync); 5909 5910 if (sched_feat(WA_WEIGHT) && target == nr_cpumask_bits) 5911 target = wake_affine_weight(sd, p, this_cpu, prev_cpu, sync); 5912 5913 schedstat_inc(p->se.statistics.nr_wakeups_affine_attempts); 5914 if (target == nr_cpumask_bits) 5915 return prev_cpu; 5916 5917 schedstat_inc(sd->ttwu_move_affine); 5918 schedstat_inc(p->se.statistics.nr_wakeups_affine); 5919 return target; 5920 } 5921 5922 static struct sched_group * 5923 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu); 5924 5925 /* 5926 * find_idlest_group_cpu - find the idlest CPU among the CPUs in the group. 5927 */ 5928 static int 5929 find_idlest_group_cpu(struct sched_group *group, struct task_struct *p, int this_cpu) 5930 { 5931 unsigned long load, min_load = ULONG_MAX; 5932 unsigned int min_exit_latency = UINT_MAX; 5933 u64 latest_idle_timestamp = 0; 5934 int least_loaded_cpu = this_cpu; 5935 int shallowest_idle_cpu = -1; 5936 int i; 5937 5938 /* Check if we have any choice: */ 5939 if (group->group_weight == 1) 5940 return cpumask_first(sched_group_span(group)); 5941 5942 /* Traverse only the allowed CPUs */ 5943 for_each_cpu_and(i, sched_group_span(group), p->cpus_ptr) { 5944 if (sched_idle_cpu(i)) 5945 return i; 5946 5947 if (available_idle_cpu(i)) { 5948 struct rq *rq = cpu_rq(i); 5949 struct cpuidle_state *idle = idle_get_state(rq); 5950 if (idle && idle->exit_latency < min_exit_latency) { 5951 /* 5952 * We give priority to a CPU whose idle state 5953 * has the smallest exit latency irrespective 5954 * of any idle timestamp. 5955 */ 5956 min_exit_latency = idle->exit_latency; 5957 latest_idle_timestamp = rq->idle_stamp; 5958 shallowest_idle_cpu = i; 5959 } else if ((!idle || idle->exit_latency == min_exit_latency) && 5960 rq->idle_stamp > latest_idle_timestamp) { 5961 /* 5962 * If equal or no active idle state, then 5963 * the most recently idled CPU might have 5964 * a warmer cache. 5965 */ 5966 latest_idle_timestamp = rq->idle_stamp; 5967 shallowest_idle_cpu = i; 5968 } 5969 } else if (shallowest_idle_cpu == -1) { 5970 load = cpu_load(cpu_rq(i)); 5971 if (load < min_load) { 5972 min_load = load; 5973 least_loaded_cpu = i; 5974 } 5975 } 5976 } 5977 5978 return shallowest_idle_cpu != -1 ? shallowest_idle_cpu : least_loaded_cpu; 5979 } 5980 5981 static inline int find_idlest_cpu(struct sched_domain *sd, struct task_struct *p, 5982 int cpu, int prev_cpu, int sd_flag) 5983 { 5984 int new_cpu = cpu; 5985 5986 if (!cpumask_intersects(sched_domain_span(sd), p->cpus_ptr)) 5987 return prev_cpu; 5988 5989 /* 5990 * We need task's util for cpu_util_without, sync it up to 5991 * prev_cpu's last_update_time. 5992 */ 5993 if (!(sd_flag & SD_BALANCE_FORK)) 5994 sync_entity_load_avg(&p->se); 5995 5996 while (sd) { 5997 struct sched_group *group; 5998 struct sched_domain *tmp; 5999 int weight; 6000 6001 if (!(sd->flags & sd_flag)) { 6002 sd = sd->child; 6003 continue; 6004 } 6005 6006 group = find_idlest_group(sd, p, cpu); 6007 if (!group) { 6008 sd = sd->child; 6009 continue; 6010 } 6011 6012 new_cpu = find_idlest_group_cpu(group, p, cpu); 6013 if (new_cpu == cpu) { 6014 /* Now try balancing at a lower domain level of 'cpu': */ 6015 sd = sd->child; 6016 continue; 6017 } 6018 6019 /* Now try balancing at a lower domain level of 'new_cpu': */ 6020 cpu = new_cpu; 6021 weight = sd->span_weight; 6022 sd = NULL; 6023 for_each_domain(cpu, tmp) { 6024 if (weight <= tmp->span_weight) 6025 break; 6026 if (tmp->flags & sd_flag) 6027 sd = tmp; 6028 } 6029 } 6030 6031 return new_cpu; 6032 } 6033 6034 static inline int __select_idle_cpu(int cpu) 6035 { 6036 if (available_idle_cpu(cpu) || sched_idle_cpu(cpu)) 6037 return cpu; 6038 6039 return -1; 6040 } 6041 6042 #ifdef CONFIG_SCHED_SMT 6043 DEFINE_STATIC_KEY_FALSE(sched_smt_present); 6044 EXPORT_SYMBOL_GPL(sched_smt_present); 6045 6046 static inline void set_idle_cores(int cpu, int val) 6047 { 6048 struct sched_domain_shared *sds; 6049 6050 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu)); 6051 if (sds) 6052 WRITE_ONCE(sds->has_idle_cores, val); 6053 } 6054 6055 static inline bool test_idle_cores(int cpu, bool def) 6056 { 6057 struct sched_domain_shared *sds; 6058 6059 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu)); 6060 if (sds) 6061 return READ_ONCE(sds->has_idle_cores); 6062 6063 return def; 6064 } 6065 6066 /* 6067 * Scans the local SMT mask to see if the entire core is idle, and records this 6068 * information in sd_llc_shared->has_idle_cores. 6069 * 6070 * Since SMT siblings share all cache levels, inspecting this limited remote 6071 * state should be fairly cheap. 6072 */ 6073 void __update_idle_core(struct rq *rq) 6074 { 6075 int core = cpu_of(rq); 6076 int cpu; 6077 6078 rcu_read_lock(); 6079 if (test_idle_cores(core, true)) 6080 goto unlock; 6081 6082 for_each_cpu(cpu, cpu_smt_mask(core)) { 6083 if (cpu == core) 6084 continue; 6085 6086 if (!available_idle_cpu(cpu)) 6087 goto unlock; 6088 } 6089 6090 set_idle_cores(core, 1); 6091 unlock: 6092 rcu_read_unlock(); 6093 } 6094 6095 /* 6096 * Scan the entire LLC domain for idle cores; this dynamically switches off if 6097 * there are no idle cores left in the system; tracked through 6098 * sd_llc->shared->has_idle_cores and enabled through update_idle_core() above. 6099 */ 6100 static int select_idle_core(struct task_struct *p, int core, struct cpumask *cpus, int *idle_cpu) 6101 { 6102 bool idle = true; 6103 int cpu; 6104 6105 if (!static_branch_likely(&sched_smt_present)) 6106 return __select_idle_cpu(core); 6107 6108 for_each_cpu(cpu, cpu_smt_mask(core)) { 6109 if (!available_idle_cpu(cpu)) { 6110 idle = false; 6111 if (*idle_cpu == -1) { 6112 if (sched_idle_cpu(cpu) && cpumask_test_cpu(cpu, p->cpus_ptr)) { 6113 *idle_cpu = cpu; 6114 break; 6115 } 6116 continue; 6117 } 6118 break; 6119 } 6120 if (*idle_cpu == -1 && cpumask_test_cpu(cpu, p->cpus_ptr)) 6121 *idle_cpu = cpu; 6122 } 6123 6124 if (idle) 6125 return core; 6126 6127 cpumask_andnot(cpus, cpus, cpu_smt_mask(core)); 6128 return -1; 6129 } 6130 6131 /* 6132 * Scan the local SMT mask for idle CPUs. 6133 */ 6134 static int select_idle_smt(struct task_struct *p, struct sched_domain *sd, int target) 6135 { 6136 int cpu; 6137 6138 for_each_cpu(cpu, cpu_smt_mask(target)) { 6139 if (!cpumask_test_cpu(cpu, p->cpus_ptr) || 6140 !cpumask_test_cpu(cpu, sched_domain_span(sd))) 6141 continue; 6142 if (available_idle_cpu(cpu) || sched_idle_cpu(cpu)) 6143 return cpu; 6144 } 6145 6146 return -1; 6147 } 6148 6149 #else /* CONFIG_SCHED_SMT */ 6150 6151 static inline void set_idle_cores(int cpu, int val) 6152 { 6153 } 6154 6155 static inline bool test_idle_cores(int cpu, bool def) 6156 { 6157 return def; 6158 } 6159 6160 static inline int select_idle_core(struct task_struct *p, int core, struct cpumask *cpus, int *idle_cpu) 6161 { 6162 return __select_idle_cpu(core); 6163 } 6164 6165 static inline int select_idle_smt(struct task_struct *p, struct sched_domain *sd, int target) 6166 { 6167 return -1; 6168 } 6169 6170 #endif /* CONFIG_SCHED_SMT */ 6171 6172 /* 6173 * Scan the LLC domain for idle CPUs; this is dynamically regulated by 6174 * comparing the average scan cost (tracked in sd->avg_scan_cost) against the 6175 * average idle time for this rq (as found in rq->avg_idle). 6176 */ 6177 static int select_idle_cpu(struct task_struct *p, struct sched_domain *sd, bool has_idle_core, int target) 6178 { 6179 struct cpumask *cpus = this_cpu_cpumask_var_ptr(select_idle_mask); 6180 int i, cpu, idle_cpu = -1, nr = INT_MAX; 6181 int this = smp_processor_id(); 6182 struct sched_domain *this_sd; 6183 u64 time; 6184 6185 this_sd = rcu_dereference(*this_cpu_ptr(&sd_llc)); 6186 if (!this_sd) 6187 return -1; 6188 6189 cpumask_and(cpus, sched_domain_span(sd), p->cpus_ptr); 6190 6191 if (sched_feat(SIS_PROP) && !has_idle_core) { 6192 u64 avg_cost, avg_idle, span_avg; 6193 6194 /* 6195 * Due to large variance we need a large fuzz factor; 6196 * hackbench in particularly is sensitive here. 6197 */ 6198 avg_idle = this_rq()->avg_idle / 512; 6199 avg_cost = this_sd->avg_scan_cost + 1; 6200 6201 span_avg = sd->span_weight * avg_idle; 6202 if (span_avg > 4*avg_cost) 6203 nr = div_u64(span_avg, avg_cost); 6204 else 6205 nr = 4; 6206 6207 time = cpu_clock(this); 6208 } 6209 6210 for_each_cpu_wrap(cpu, cpus, target) { 6211 if (has_idle_core) { 6212 i = select_idle_core(p, cpu, cpus, &idle_cpu); 6213 if ((unsigned int)i < nr_cpumask_bits) 6214 return i; 6215 6216 } else { 6217 if (!--nr) 6218 return -1; 6219 idle_cpu = __select_idle_cpu(cpu); 6220 if ((unsigned int)idle_cpu < nr_cpumask_bits) 6221 break; 6222 } 6223 } 6224 6225 if (has_idle_core) 6226 set_idle_cores(target, false); 6227 6228 if (sched_feat(SIS_PROP) && !has_idle_core) { 6229 time = cpu_clock(this) - time; 6230 update_avg(&this_sd->avg_scan_cost, time); 6231 } 6232 6233 return idle_cpu; 6234 } 6235 6236 /* 6237 * Scan the asym_capacity domain for idle CPUs; pick the first idle one on which 6238 * the task fits. If no CPU is big enough, but there are idle ones, try to 6239 * maximize capacity. 6240 */ 6241 static int 6242 select_idle_capacity(struct task_struct *p, struct sched_domain *sd, int target) 6243 { 6244 unsigned long task_util, best_cap = 0; 6245 int cpu, best_cpu = -1; 6246 struct cpumask *cpus; 6247 6248 cpus = this_cpu_cpumask_var_ptr(select_idle_mask); 6249 cpumask_and(cpus, sched_domain_span(sd), p->cpus_ptr); 6250 6251 task_util = uclamp_task_util(p); 6252 6253 for_each_cpu_wrap(cpu, cpus, target) { 6254 unsigned long cpu_cap = capacity_of(cpu); 6255 6256 if (!available_idle_cpu(cpu) && !sched_idle_cpu(cpu)) 6257 continue; 6258 if (fits_capacity(task_util, cpu_cap)) 6259 return cpu; 6260 6261 if (cpu_cap > best_cap) { 6262 best_cap = cpu_cap; 6263 best_cpu = cpu; 6264 } 6265 } 6266 6267 return best_cpu; 6268 } 6269 6270 static inline bool asym_fits_capacity(int task_util, int cpu) 6271 { 6272 if (static_branch_unlikely(&sched_asym_cpucapacity)) 6273 return fits_capacity(task_util, capacity_of(cpu)); 6274 6275 return true; 6276 } 6277 6278 /* 6279 * Try and locate an idle core/thread in the LLC cache domain. 6280 */ 6281 static int select_idle_sibling(struct task_struct *p, int prev, int target) 6282 { 6283 bool has_idle_core = false; 6284 struct sched_domain *sd; 6285 unsigned long task_util; 6286 int i, recent_used_cpu; 6287 6288 /* 6289 * On asymmetric system, update task utilization because we will check 6290 * that the task fits with cpu's capacity. 6291 */ 6292 if (static_branch_unlikely(&sched_asym_cpucapacity)) { 6293 sync_entity_load_avg(&p->se); 6294 task_util = uclamp_task_util(p); 6295 } 6296 6297 if ((available_idle_cpu(target) || sched_idle_cpu(target)) && 6298 asym_fits_capacity(task_util, target)) 6299 return target; 6300 6301 /* 6302 * If the previous CPU is cache affine and idle, don't be stupid: 6303 */ 6304 if (prev != target && cpus_share_cache(prev, target) && 6305 (available_idle_cpu(prev) || sched_idle_cpu(prev)) && 6306 asym_fits_capacity(task_util, prev)) 6307 return prev; 6308 6309 /* 6310 * Allow a per-cpu kthread to stack with the wakee if the 6311 * kworker thread and the tasks previous CPUs are the same. 6312 * The assumption is that the wakee queued work for the 6313 * per-cpu kthread that is now complete and the wakeup is 6314 * essentially a sync wakeup. An obvious example of this 6315 * pattern is IO completions. 6316 */ 6317 if (is_per_cpu_kthread(current) && 6318 prev == smp_processor_id() && 6319 this_rq()->nr_running <= 1) { 6320 return prev; 6321 } 6322 6323 /* Check a recently used CPU as a potential idle candidate: */ 6324 recent_used_cpu = p->recent_used_cpu; 6325 if (recent_used_cpu != prev && 6326 recent_used_cpu != target && 6327 cpus_share_cache(recent_used_cpu, target) && 6328 (available_idle_cpu(recent_used_cpu) || sched_idle_cpu(recent_used_cpu)) && 6329 cpumask_test_cpu(p->recent_used_cpu, p->cpus_ptr) && 6330 asym_fits_capacity(task_util, recent_used_cpu)) { 6331 /* 6332 * Replace recent_used_cpu with prev as it is a potential 6333 * candidate for the next wake: 6334 */ 6335 p->recent_used_cpu = prev; 6336 return recent_used_cpu; 6337 } 6338 6339 /* 6340 * For asymmetric CPU capacity systems, our domain of interest is 6341 * sd_asym_cpucapacity rather than sd_llc. 6342 */ 6343 if (static_branch_unlikely(&sched_asym_cpucapacity)) { 6344 sd = rcu_dereference(per_cpu(sd_asym_cpucapacity, target)); 6345 /* 6346 * On an asymmetric CPU capacity system where an exclusive 6347 * cpuset defines a symmetric island (i.e. one unique 6348 * capacity_orig value through the cpuset), the key will be set 6349 * but the CPUs within that cpuset will not have a domain with 6350 * SD_ASYM_CPUCAPACITY. These should follow the usual symmetric 6351 * capacity path. 6352 */ 6353 if (sd) { 6354 i = select_idle_capacity(p, sd, target); 6355 return ((unsigned)i < nr_cpumask_bits) ? i : target; 6356 } 6357 } 6358 6359 sd = rcu_dereference(per_cpu(sd_llc, target)); 6360 if (!sd) 6361 return target; 6362 6363 if (sched_smt_active()) { 6364 has_idle_core = test_idle_cores(target, false); 6365 6366 if (!has_idle_core && cpus_share_cache(prev, target)) { 6367 i = select_idle_smt(p, sd, prev); 6368 if ((unsigned int)i < nr_cpumask_bits) 6369 return i; 6370 } 6371 } 6372 6373 i = select_idle_cpu(p, sd, has_idle_core, target); 6374 if ((unsigned)i < nr_cpumask_bits) 6375 return i; 6376 6377 return target; 6378 } 6379 6380 /** 6381 * cpu_util - Estimates the amount of capacity of a CPU used by CFS tasks. 6382 * @cpu: the CPU to get the utilization of 6383 * 6384 * The unit of the return value must be the one of capacity so we can compare 6385 * the utilization with the capacity of the CPU that is available for CFS task 6386 * (ie cpu_capacity). 6387 * 6388 * cfs_rq.avg.util_avg is the sum of running time of runnable tasks plus the 6389 * recent utilization of currently non-runnable tasks on a CPU. It represents 6390 * the amount of utilization of a CPU in the range [0..capacity_orig] where 6391 * capacity_orig is the cpu_capacity available at the highest frequency 6392 * (arch_scale_freq_capacity()). 6393 * The utilization of a CPU converges towards a sum equal to or less than the 6394 * current capacity (capacity_curr <= capacity_orig) of the CPU because it is 6395 * the running time on this CPU scaled by capacity_curr. 6396 * 6397 * The estimated utilization of a CPU is defined to be the maximum between its 6398 * cfs_rq.avg.util_avg and the sum of the estimated utilization of the tasks 6399 * currently RUNNABLE on that CPU. 6400 * This allows to properly represent the expected utilization of a CPU which 6401 * has just got a big task running since a long sleep period. At the same time 6402 * however it preserves the benefits of the "blocked utilization" in 6403 * describing the potential for other tasks waking up on the same CPU. 6404 * 6405 * Nevertheless, cfs_rq.avg.util_avg can be higher than capacity_curr or even 6406 * higher than capacity_orig because of unfortunate rounding in 6407 * cfs.avg.util_avg or just after migrating tasks and new task wakeups until 6408 * the average stabilizes with the new running time. We need to check that the 6409 * utilization stays within the range of [0..capacity_orig] and cap it if 6410 * necessary. Without utilization capping, a group could be seen as overloaded 6411 * (CPU0 utilization at 121% + CPU1 utilization at 80%) whereas CPU1 has 20% of 6412 * available capacity. We allow utilization to overshoot capacity_curr (but not 6413 * capacity_orig) as it useful for predicting the capacity required after task 6414 * migrations (scheduler-driven DVFS). 6415 * 6416 * Return: the (estimated) utilization for the specified CPU 6417 */ 6418 static inline unsigned long cpu_util(int cpu) 6419 { 6420 struct cfs_rq *cfs_rq; 6421 unsigned int util; 6422 6423 cfs_rq = &cpu_rq(cpu)->cfs; 6424 util = READ_ONCE(cfs_rq->avg.util_avg); 6425 6426 if (sched_feat(UTIL_EST)) 6427 util = max(util, READ_ONCE(cfs_rq->avg.util_est.enqueued)); 6428 6429 return min_t(unsigned long, util, capacity_orig_of(cpu)); 6430 } 6431 6432 /* 6433 * cpu_util_without: compute cpu utilization without any contributions from *p 6434 * @cpu: the CPU which utilization is requested 6435 * @p: the task which utilization should be discounted 6436 * 6437 * The utilization of a CPU is defined by the utilization of tasks currently 6438 * enqueued on that CPU as well as tasks which are currently sleeping after an 6439 * execution on that CPU. 6440 * 6441 * This method returns the utilization of the specified CPU by discounting the 6442 * utilization of the specified task, whenever the task is currently 6443 * contributing to the CPU utilization. 6444 */ 6445 static unsigned long cpu_util_without(int cpu, struct task_struct *p) 6446 { 6447 struct cfs_rq *cfs_rq; 6448 unsigned int util; 6449 6450 /* Task has no contribution or is new */ 6451 if (cpu != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time)) 6452 return cpu_util(cpu); 6453 6454 cfs_rq = &cpu_rq(cpu)->cfs; 6455 util = READ_ONCE(cfs_rq->avg.util_avg); 6456 6457 /* Discount task's util from CPU's util */ 6458 lsub_positive(&util, task_util(p)); 6459 6460 /* 6461 * Covered cases: 6462 * 6463 * a) if *p is the only task sleeping on this CPU, then: 6464 * cpu_util (== task_util) > util_est (== 0) 6465 * and thus we return: 6466 * cpu_util_without = (cpu_util - task_util) = 0 6467 * 6468 * b) if other tasks are SLEEPING on this CPU, which is now exiting 6469 * IDLE, then: 6470 * cpu_util >= task_util 6471 * cpu_util > util_est (== 0) 6472 * and thus we discount *p's blocked utilization to return: 6473 * cpu_util_without = (cpu_util - task_util) >= 0 6474 * 6475 * c) if other tasks are RUNNABLE on that CPU and 6476 * util_est > cpu_util 6477 * then we use util_est since it returns a more restrictive 6478 * estimation of the spare capacity on that CPU, by just 6479 * considering the expected utilization of tasks already 6480 * runnable on that CPU. 6481 * 6482 * Cases a) and b) are covered by the above code, while case c) is 6483 * covered by the following code when estimated utilization is 6484 * enabled. 6485 */ 6486 if (sched_feat(UTIL_EST)) { 6487 unsigned int estimated = 6488 READ_ONCE(cfs_rq->avg.util_est.enqueued); 6489 6490 /* 6491 * Despite the following checks we still have a small window 6492 * for a possible race, when an execl's select_task_rq_fair() 6493 * races with LB's detach_task(): 6494 * 6495 * detach_task() 6496 * p->on_rq = TASK_ON_RQ_MIGRATING; 6497 * ---------------------------------- A 6498 * deactivate_task() \ 6499 * dequeue_task() + RaceTime 6500 * util_est_dequeue() / 6501 * ---------------------------------- B 6502 * 6503 * The additional check on "current == p" it's required to 6504 * properly fix the execl regression and it helps in further 6505 * reducing the chances for the above race. 6506 */ 6507 if (unlikely(task_on_rq_queued(p) || current == p)) 6508 lsub_positive(&estimated, _task_util_est(p)); 6509 6510 util = max(util, estimated); 6511 } 6512 6513 /* 6514 * Utilization (estimated) can exceed the CPU capacity, thus let's 6515 * clamp to the maximum CPU capacity to ensure consistency with 6516 * the cpu_util call. 6517 */ 6518 return min_t(unsigned long, util, capacity_orig_of(cpu)); 6519 } 6520 6521 /* 6522 * Predicts what cpu_util(@cpu) would return if @p was migrated (and enqueued) 6523 * to @dst_cpu. 6524 */ 6525 static unsigned long cpu_util_next(int cpu, struct task_struct *p, int dst_cpu) 6526 { 6527 struct cfs_rq *cfs_rq = &cpu_rq(cpu)->cfs; 6528 unsigned long util_est, util = READ_ONCE(cfs_rq->avg.util_avg); 6529 6530 /* 6531 * If @p migrates from @cpu to another, remove its contribution. Or, 6532 * if @p migrates from another CPU to @cpu, add its contribution. In 6533 * the other cases, @cpu is not impacted by the migration, so the 6534 * util_avg should already be correct. 6535 */ 6536 if (task_cpu(p) == cpu && dst_cpu != cpu) 6537 lsub_positive(&util, task_util(p)); 6538 else if (task_cpu(p) != cpu && dst_cpu == cpu) 6539 util += task_util(p); 6540 6541 if (sched_feat(UTIL_EST)) { 6542 util_est = READ_ONCE(cfs_rq->avg.util_est.enqueued); 6543 6544 /* 6545 * During wake-up, the task isn't enqueued yet and doesn't 6546 * appear in the cfs_rq->avg.util_est.enqueued of any rq, 6547 * so just add it (if needed) to "simulate" what will be 6548 * cpu_util() after the task has been enqueued. 6549 */ 6550 if (dst_cpu == cpu) 6551 util_est += _task_util_est(p); 6552 6553 util = max(util, util_est); 6554 } 6555 6556 return min(util, capacity_orig_of(cpu)); 6557 } 6558 6559 /* 6560 * compute_energy(): Estimates the energy that @pd would consume if @p was 6561 * migrated to @dst_cpu. compute_energy() predicts what will be the utilization 6562 * landscape of @pd's CPUs after the task migration, and uses the Energy Model 6563 * to compute what would be the energy if we decided to actually migrate that 6564 * task. 6565 */ 6566 static long 6567 compute_energy(struct task_struct *p, int dst_cpu, struct perf_domain *pd) 6568 { 6569 struct cpumask *pd_mask = perf_domain_span(pd); 6570 unsigned long cpu_cap = arch_scale_cpu_capacity(cpumask_first(pd_mask)); 6571 unsigned long max_util = 0, sum_util = 0; 6572 int cpu; 6573 6574 /* 6575 * The capacity state of CPUs of the current rd can be driven by CPUs 6576 * of another rd if they belong to the same pd. So, account for the 6577 * utilization of these CPUs too by masking pd with cpu_online_mask 6578 * instead of the rd span. 6579 * 6580 * If an entire pd is outside of the current rd, it will not appear in 6581 * its pd list and will not be accounted by compute_energy(). 6582 */ 6583 for_each_cpu_and(cpu, pd_mask, cpu_online_mask) { 6584 unsigned long util_freq = cpu_util_next(cpu, p, dst_cpu); 6585 unsigned long cpu_util, util_running = util_freq; 6586 struct task_struct *tsk = NULL; 6587 6588 /* 6589 * When @p is placed on @cpu: 6590 * 6591 * util_running = max(cpu_util, cpu_util_est) + 6592 * max(task_util, _task_util_est) 6593 * 6594 * while cpu_util_next is: max(cpu_util + task_util, 6595 * cpu_util_est + _task_util_est) 6596 */ 6597 if (cpu == dst_cpu) { 6598 tsk = p; 6599 util_running = 6600 cpu_util_next(cpu, p, -1) + task_util_est(p); 6601 } 6602 6603 /* 6604 * Busy time computation: utilization clamping is not 6605 * required since the ratio (sum_util / cpu_capacity) 6606 * is already enough to scale the EM reported power 6607 * consumption at the (eventually clamped) cpu_capacity. 6608 */ 6609 sum_util += effective_cpu_util(cpu, util_running, cpu_cap, 6610 ENERGY_UTIL, NULL); 6611 6612 /* 6613 * Performance domain frequency: utilization clamping 6614 * must be considered since it affects the selection 6615 * of the performance domain frequency. 6616 * NOTE: in case RT tasks are running, by default the 6617 * FREQUENCY_UTIL's utilization can be max OPP. 6618 */ 6619 cpu_util = effective_cpu_util(cpu, util_freq, cpu_cap, 6620 FREQUENCY_UTIL, tsk); 6621 max_util = max(max_util, cpu_util); 6622 } 6623 6624 return em_cpu_energy(pd->em_pd, max_util, sum_util); 6625 } 6626 6627 /* 6628 * find_energy_efficient_cpu(): Find most energy-efficient target CPU for the 6629 * waking task. find_energy_efficient_cpu() looks for the CPU with maximum 6630 * spare capacity in each performance domain and uses it as a potential 6631 * candidate to execute the task. Then, it uses the Energy Model to figure 6632 * out which of the CPU candidates is the most energy-efficient. 6633 * 6634 * The rationale for this heuristic is as follows. In a performance domain, 6635 * all the most energy efficient CPU candidates (according to the Energy 6636 * Model) are those for which we'll request a low frequency. When there are 6637 * several CPUs for which the frequency request will be the same, we don't 6638 * have enough data to break the tie between them, because the Energy Model 6639 * only includes active power costs. With this model, if we assume that 6640 * frequency requests follow utilization (e.g. using schedutil), the CPU with 6641 * the maximum spare capacity in a performance domain is guaranteed to be among 6642 * the best candidates of the performance domain. 6643 * 6644 * In practice, it could be preferable from an energy standpoint to pack 6645 * small tasks on a CPU in order to let other CPUs go in deeper idle states, 6646 * but that could also hurt our chances to go cluster idle, and we have no 6647 * ways to tell with the current Energy Model if this is actually a good 6648 * idea or not. So, find_energy_efficient_cpu() basically favors 6649 * cluster-packing, and spreading inside a cluster. That should at least be 6650 * a good thing for latency, and this is consistent with the idea that most 6651 * of the energy savings of EAS come from the asymmetry of the system, and 6652 * not so much from breaking the tie between identical CPUs. That's also the 6653 * reason why EAS is enabled in the topology code only for systems where 6654 * SD_ASYM_CPUCAPACITY is set. 6655 * 6656 * NOTE: Forkees are not accepted in the energy-aware wake-up path because 6657 * they don't have any useful utilization data yet and it's not possible to 6658 * forecast their impact on energy consumption. Consequently, they will be 6659 * placed by find_idlest_cpu() on the least loaded CPU, which might turn out 6660 * to be energy-inefficient in some use-cases. The alternative would be to 6661 * bias new tasks towards specific types of CPUs first, or to try to infer 6662 * their util_avg from the parent task, but those heuristics could hurt 6663 * other use-cases too. So, until someone finds a better way to solve this, 6664 * let's keep things simple by re-using the existing slow path. 6665 */ 6666 static int find_energy_efficient_cpu(struct task_struct *p, int prev_cpu) 6667 { 6668 unsigned long prev_delta = ULONG_MAX, best_delta = ULONG_MAX; 6669 struct root_domain *rd = cpu_rq(smp_processor_id())->rd; 6670 unsigned long cpu_cap, util, base_energy = 0; 6671 int cpu, best_energy_cpu = prev_cpu; 6672 struct sched_domain *sd; 6673 struct perf_domain *pd; 6674 6675 rcu_read_lock(); 6676 pd = rcu_dereference(rd->pd); 6677 if (!pd || READ_ONCE(rd->overutilized)) 6678 goto fail; 6679 6680 /* 6681 * Energy-aware wake-up happens on the lowest sched_domain starting 6682 * from sd_asym_cpucapacity spanning over this_cpu and prev_cpu. 6683 */ 6684 sd = rcu_dereference(*this_cpu_ptr(&sd_asym_cpucapacity)); 6685 while (sd && !cpumask_test_cpu(prev_cpu, sched_domain_span(sd))) 6686 sd = sd->parent; 6687 if (!sd) 6688 goto fail; 6689 6690 sync_entity_load_avg(&p->se); 6691 if (!task_util_est(p)) 6692 goto unlock; 6693 6694 for (; pd; pd = pd->next) { 6695 unsigned long cur_delta, spare_cap, max_spare_cap = 0; 6696 unsigned long base_energy_pd; 6697 int max_spare_cap_cpu = -1; 6698 6699 /* Compute the 'base' energy of the pd, without @p */ 6700 base_energy_pd = compute_energy(p, -1, pd); 6701 base_energy += base_energy_pd; 6702 6703 for_each_cpu_and(cpu, perf_domain_span(pd), sched_domain_span(sd)) { 6704 if (!cpumask_test_cpu(cpu, p->cpus_ptr)) 6705 continue; 6706 6707 util = cpu_util_next(cpu, p, cpu); 6708 cpu_cap = capacity_of(cpu); 6709 spare_cap = cpu_cap; 6710 lsub_positive(&spare_cap, util); 6711 6712 /* 6713 * Skip CPUs that cannot satisfy the capacity request. 6714 * IOW, placing the task there would make the CPU 6715 * overutilized. Take uclamp into account to see how 6716 * much capacity we can get out of the CPU; this is 6717 * aligned with sched_cpu_util(). 6718 */ 6719 util = uclamp_rq_util_with(cpu_rq(cpu), util, p); 6720 if (!fits_capacity(util, cpu_cap)) 6721 continue; 6722 6723 /* Always use prev_cpu as a candidate. */ 6724 if (cpu == prev_cpu) { 6725 prev_delta = compute_energy(p, prev_cpu, pd); 6726 prev_delta -= base_energy_pd; 6727 best_delta = min(best_delta, prev_delta); 6728 } 6729 6730 /* 6731 * Find the CPU with the maximum spare capacity in 6732 * the performance domain 6733 */ 6734 if (spare_cap > max_spare_cap) { 6735 max_spare_cap = spare_cap; 6736 max_spare_cap_cpu = cpu; 6737 } 6738 } 6739 6740 /* Evaluate the energy impact of using this CPU. */ 6741 if (max_spare_cap_cpu >= 0 && max_spare_cap_cpu != prev_cpu) { 6742 cur_delta = compute_energy(p, max_spare_cap_cpu, pd); 6743 cur_delta -= base_energy_pd; 6744 if (cur_delta < best_delta) { 6745 best_delta = cur_delta; 6746 best_energy_cpu = max_spare_cap_cpu; 6747 } 6748 } 6749 } 6750 unlock: 6751 rcu_read_unlock(); 6752 6753 /* 6754 * Pick the best CPU if prev_cpu cannot be used, or if it saves at 6755 * least 6% of the energy used by prev_cpu. 6756 */ 6757 if (prev_delta == ULONG_MAX) 6758 return best_energy_cpu; 6759 6760 if ((prev_delta - best_delta) > ((prev_delta + base_energy) >> 4)) 6761 return best_energy_cpu; 6762 6763 return prev_cpu; 6764 6765 fail: 6766 rcu_read_unlock(); 6767 6768 return -1; 6769 } 6770 6771 /* 6772 * select_task_rq_fair: Select target runqueue for the waking task in domains 6773 * that have the relevant SD flag set. In practice, this is SD_BALANCE_WAKE, 6774 * SD_BALANCE_FORK, or SD_BALANCE_EXEC. 6775 * 6776 * Balances load by selecting the idlest CPU in the idlest group, or under 6777 * certain conditions an idle sibling CPU if the domain has SD_WAKE_AFFINE set. 6778 * 6779 * Returns the target CPU number. 6780 * 6781 * preempt must be disabled. 6782 */ 6783 static int 6784 select_task_rq_fair(struct task_struct *p, int prev_cpu, int wake_flags) 6785 { 6786 int sync = (wake_flags & WF_SYNC) && !(current->flags & PF_EXITING); 6787 struct sched_domain *tmp, *sd = NULL; 6788 int cpu = smp_processor_id(); 6789 int new_cpu = prev_cpu; 6790 int want_affine = 0; 6791 /* SD_flags and WF_flags share the first nibble */ 6792 int sd_flag = wake_flags & 0xF; 6793 6794 if (wake_flags & WF_TTWU) { 6795 record_wakee(p); 6796 6797 if (sched_energy_enabled()) { 6798 new_cpu = find_energy_efficient_cpu(p, prev_cpu); 6799 if (new_cpu >= 0) 6800 return new_cpu; 6801 new_cpu = prev_cpu; 6802 } 6803 6804 want_affine = !wake_wide(p) && cpumask_test_cpu(cpu, p->cpus_ptr); 6805 } 6806 6807 rcu_read_lock(); 6808 for_each_domain(cpu, tmp) { 6809 /* 6810 * If both 'cpu' and 'prev_cpu' are part of this domain, 6811 * cpu is a valid SD_WAKE_AFFINE target. 6812 */ 6813 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) && 6814 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) { 6815 if (cpu != prev_cpu) 6816 new_cpu = wake_affine(tmp, p, cpu, prev_cpu, sync); 6817 6818 sd = NULL; /* Prefer wake_affine over balance flags */ 6819 break; 6820 } 6821 6822 if (tmp->flags & sd_flag) 6823 sd = tmp; 6824 else if (!want_affine) 6825 break; 6826 } 6827 6828 if (unlikely(sd)) { 6829 /* Slow path */ 6830 new_cpu = find_idlest_cpu(sd, p, cpu, prev_cpu, sd_flag); 6831 } else if (wake_flags & WF_TTWU) { /* XXX always ? */ 6832 /* Fast path */ 6833 new_cpu = select_idle_sibling(p, prev_cpu, new_cpu); 6834 6835 if (want_affine) 6836 current->recent_used_cpu = cpu; 6837 } 6838 rcu_read_unlock(); 6839 6840 return new_cpu; 6841 } 6842 6843 static void detach_entity_cfs_rq(struct sched_entity *se); 6844 6845 /* 6846 * Called immediately before a task is migrated to a new CPU; task_cpu(p) and 6847 * cfs_rq_of(p) references at time of call are still valid and identify the 6848 * previous CPU. The caller guarantees p->pi_lock or task_rq(p)->lock is held. 6849 */ 6850 static void migrate_task_rq_fair(struct task_struct *p, int new_cpu) 6851 { 6852 /* 6853 * As blocked tasks retain absolute vruntime the migration needs to 6854 * deal with this by subtracting the old and adding the new 6855 * min_vruntime -- the latter is done by enqueue_entity() when placing 6856 * the task on the new runqueue. 6857 */ 6858 if (p->state == TASK_WAKING) { 6859 struct sched_entity *se = &p->se; 6860 struct cfs_rq *cfs_rq = cfs_rq_of(se); 6861 u64 min_vruntime; 6862 6863 #ifndef CONFIG_64BIT 6864 u64 min_vruntime_copy; 6865 6866 do { 6867 min_vruntime_copy = cfs_rq->min_vruntime_copy; 6868 smp_rmb(); 6869 min_vruntime = cfs_rq->min_vruntime; 6870 } while (min_vruntime != min_vruntime_copy); 6871 #else 6872 min_vruntime = cfs_rq->min_vruntime; 6873 #endif 6874 6875 se->vruntime -= min_vruntime; 6876 } 6877 6878 if (p->on_rq == TASK_ON_RQ_MIGRATING) { 6879 /* 6880 * In case of TASK_ON_RQ_MIGRATING we in fact hold the 'old' 6881 * rq->lock and can modify state directly. 6882 */ 6883 lockdep_assert_held(&task_rq(p)->lock); 6884 detach_entity_cfs_rq(&p->se); 6885 6886 } else { 6887 /* 6888 * We are supposed to update the task to "current" time, then 6889 * its up to date and ready to go to new CPU/cfs_rq. But we 6890 * have difficulty in getting what current time is, so simply 6891 * throw away the out-of-date time. This will result in the 6892 * wakee task is less decayed, but giving the wakee more load 6893 * sounds not bad. 6894 */ 6895 remove_entity_load_avg(&p->se); 6896 } 6897 6898 /* Tell new CPU we are migrated */ 6899 p->se.avg.last_update_time = 0; 6900 6901 /* We have migrated, no longer consider this task hot */ 6902 p->se.exec_start = 0; 6903 6904 update_scan_period(p, new_cpu); 6905 } 6906 6907 static void task_dead_fair(struct task_struct *p) 6908 { 6909 remove_entity_load_avg(&p->se); 6910 } 6911 6912 static int 6913 balance_fair(struct rq *rq, struct task_struct *prev, struct rq_flags *rf) 6914 { 6915 if (rq->nr_running) 6916 return 1; 6917 6918 return newidle_balance(rq, rf) != 0; 6919 } 6920 #endif /* CONFIG_SMP */ 6921 6922 static unsigned long wakeup_gran(struct sched_entity *se) 6923 { 6924 unsigned long gran = sysctl_sched_wakeup_granularity; 6925 6926 /* 6927 * Since its curr running now, convert the gran from real-time 6928 * to virtual-time in his units. 6929 * 6930 * By using 'se' instead of 'curr' we penalize light tasks, so 6931 * they get preempted easier. That is, if 'se' < 'curr' then 6932 * the resulting gran will be larger, therefore penalizing the 6933 * lighter, if otoh 'se' > 'curr' then the resulting gran will 6934 * be smaller, again penalizing the lighter task. 6935 * 6936 * This is especially important for buddies when the leftmost 6937 * task is higher priority than the buddy. 6938 */ 6939 return calc_delta_fair(gran, se); 6940 } 6941 6942 /* 6943 * Should 'se' preempt 'curr'. 6944 * 6945 * |s1 6946 * |s2 6947 * |s3 6948 * g 6949 * |<--->|c 6950 * 6951 * w(c, s1) = -1 6952 * w(c, s2) = 0 6953 * w(c, s3) = 1 6954 * 6955 */ 6956 static int 6957 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se) 6958 { 6959 s64 gran, vdiff = curr->vruntime - se->vruntime; 6960 6961 if (vdiff <= 0) 6962 return -1; 6963 6964 gran = wakeup_gran(se); 6965 if (vdiff > gran) 6966 return 1; 6967 6968 return 0; 6969 } 6970 6971 static void set_last_buddy(struct sched_entity *se) 6972 { 6973 if (entity_is_task(se) && unlikely(task_has_idle_policy(task_of(se)))) 6974 return; 6975 6976 for_each_sched_entity(se) { 6977 if (SCHED_WARN_ON(!se->on_rq)) 6978 return; 6979 cfs_rq_of(se)->last = se; 6980 } 6981 } 6982 6983 static void set_next_buddy(struct sched_entity *se) 6984 { 6985 if (entity_is_task(se) && unlikely(task_has_idle_policy(task_of(se)))) 6986 return; 6987 6988 for_each_sched_entity(se) { 6989 if (SCHED_WARN_ON(!se->on_rq)) 6990 return; 6991 cfs_rq_of(se)->next = se; 6992 } 6993 } 6994 6995 static void set_skip_buddy(struct sched_entity *se) 6996 { 6997 for_each_sched_entity(se) 6998 cfs_rq_of(se)->skip = se; 6999 } 7000 7001 /* 7002 * Preempt the current task with a newly woken task if needed: 7003 */ 7004 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags) 7005 { 7006 struct task_struct *curr = rq->curr; 7007 struct sched_entity *se = &curr->se, *pse = &p->se; 7008 struct cfs_rq *cfs_rq = task_cfs_rq(curr); 7009 int scale = cfs_rq->nr_running >= sched_nr_latency; 7010 int next_buddy_marked = 0; 7011 7012 if (unlikely(se == pse)) 7013 return; 7014 7015 /* 7016 * This is possible from callers such as attach_tasks(), in which we 7017 * unconditionally check_preempt_curr() after an enqueue (which may have 7018 * lead to a throttle). This both saves work and prevents false 7019 * next-buddy nomination below. 7020 */ 7021 if (unlikely(throttled_hierarchy(cfs_rq_of(pse)))) 7022 return; 7023 7024 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) { 7025 set_next_buddy(pse); 7026 next_buddy_marked = 1; 7027 } 7028 7029 /* 7030 * We can come here with TIF_NEED_RESCHED already set from new task 7031 * wake up path. 7032 * 7033 * Note: this also catches the edge-case of curr being in a throttled 7034 * group (e.g. via set_curr_task), since update_curr() (in the 7035 * enqueue of curr) will have resulted in resched being set. This 7036 * prevents us from potentially nominating it as a false LAST_BUDDY 7037 * below. 7038 */ 7039 if (test_tsk_need_resched(curr)) 7040 return; 7041 7042 /* Idle tasks are by definition preempted by non-idle tasks. */ 7043 if (unlikely(task_has_idle_policy(curr)) && 7044 likely(!task_has_idle_policy(p))) 7045 goto preempt; 7046 7047 /* 7048 * Batch and idle tasks do not preempt non-idle tasks (their preemption 7049 * is driven by the tick): 7050 */ 7051 if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION)) 7052 return; 7053 7054 find_matching_se(&se, &pse); 7055 update_curr(cfs_rq_of(se)); 7056 BUG_ON(!pse); 7057 if (wakeup_preempt_entity(se, pse) == 1) { 7058 /* 7059 * Bias pick_next to pick the sched entity that is 7060 * triggering this preemption. 7061 */ 7062 if (!next_buddy_marked) 7063 set_next_buddy(pse); 7064 goto preempt; 7065 } 7066 7067 return; 7068 7069 preempt: 7070 resched_curr(rq); 7071 /* 7072 * Only set the backward buddy when the current task is still 7073 * on the rq. This can happen when a wakeup gets interleaved 7074 * with schedule on the ->pre_schedule() or idle_balance() 7075 * point, either of which can * drop the rq lock. 7076 * 7077 * Also, during early boot the idle thread is in the fair class, 7078 * for obvious reasons its a bad idea to schedule back to it. 7079 */ 7080 if (unlikely(!se->on_rq || curr == rq->idle)) 7081 return; 7082 7083 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se)) 7084 set_last_buddy(se); 7085 } 7086 7087 struct task_struct * 7088 pick_next_task_fair(struct rq *rq, struct task_struct *prev, struct rq_flags *rf) 7089 { 7090 struct cfs_rq *cfs_rq = &rq->cfs; 7091 struct sched_entity *se; 7092 struct task_struct *p; 7093 int new_tasks; 7094 7095 again: 7096 if (!sched_fair_runnable(rq)) 7097 goto idle; 7098 7099 #ifdef CONFIG_FAIR_GROUP_SCHED 7100 if (!prev || prev->sched_class != &fair_sched_class) 7101 goto simple; 7102 7103 /* 7104 * Because of the set_next_buddy() in dequeue_task_fair() it is rather 7105 * likely that a next task is from the same cgroup as the current. 7106 * 7107 * Therefore attempt to avoid putting and setting the entire cgroup 7108 * hierarchy, only change the part that actually changes. 7109 */ 7110 7111 do { 7112 struct sched_entity *curr = cfs_rq->curr; 7113 7114 /* 7115 * Since we got here without doing put_prev_entity() we also 7116 * have to consider cfs_rq->curr. If it is still a runnable 7117 * entity, update_curr() will update its vruntime, otherwise 7118 * forget we've ever seen it. 7119 */ 7120 if (curr) { 7121 if (curr->on_rq) 7122 update_curr(cfs_rq); 7123 else 7124 curr = NULL; 7125 7126 /* 7127 * This call to check_cfs_rq_runtime() will do the 7128 * throttle and dequeue its entity in the parent(s). 7129 * Therefore the nr_running test will indeed 7130 * be correct. 7131 */ 7132 if (unlikely(check_cfs_rq_runtime(cfs_rq))) { 7133 cfs_rq = &rq->cfs; 7134 7135 if (!cfs_rq->nr_running) 7136 goto idle; 7137 7138 goto simple; 7139 } 7140 } 7141 7142 se = pick_next_entity(cfs_rq, curr); 7143 cfs_rq = group_cfs_rq(se); 7144 } while (cfs_rq); 7145 7146 p = task_of(se); 7147 7148 /* 7149 * Since we haven't yet done put_prev_entity and if the selected task 7150 * is a different task than we started out with, try and touch the 7151 * least amount of cfs_rqs. 7152 */ 7153 if (prev != p) { 7154 struct sched_entity *pse = &prev->se; 7155 7156 while (!(cfs_rq = is_same_group(se, pse))) { 7157 int se_depth = se->depth; 7158 int pse_depth = pse->depth; 7159 7160 if (se_depth <= pse_depth) { 7161 put_prev_entity(cfs_rq_of(pse), pse); 7162 pse = parent_entity(pse); 7163 } 7164 if (se_depth >= pse_depth) { 7165 set_next_entity(cfs_rq_of(se), se); 7166 se = parent_entity(se); 7167 } 7168 } 7169 7170 put_prev_entity(cfs_rq, pse); 7171 set_next_entity(cfs_rq, se); 7172 } 7173 7174 goto done; 7175 simple: 7176 #endif 7177 if (prev) 7178 put_prev_task(rq, prev); 7179 7180 do { 7181 se = pick_next_entity(cfs_rq, NULL); 7182 set_next_entity(cfs_rq, se); 7183 cfs_rq = group_cfs_rq(se); 7184 } while (cfs_rq); 7185 7186 p = task_of(se); 7187 7188 done: __maybe_unused; 7189 #ifdef CONFIG_SMP 7190 /* 7191 * Move the next running task to the front of 7192 * the list, so our cfs_tasks list becomes MRU 7193 * one. 7194 */ 7195 list_move(&p->se.group_node, &rq->cfs_tasks); 7196 #endif 7197 7198 if (hrtick_enabled_fair(rq)) 7199 hrtick_start_fair(rq, p); 7200 7201 update_misfit_status(p, rq); 7202 7203 return p; 7204 7205 idle: 7206 if (!rf) 7207 return NULL; 7208 7209 new_tasks = newidle_balance(rq, rf); 7210 7211 /* 7212 * Because newidle_balance() releases (and re-acquires) rq->lock, it is 7213 * possible for any higher priority task to appear. In that case we 7214 * must re-start the pick_next_entity() loop. 7215 */ 7216 if (new_tasks < 0) 7217 return RETRY_TASK; 7218 7219 if (new_tasks > 0) 7220 goto again; 7221 7222 /* 7223 * rq is about to be idle, check if we need to update the 7224 * lost_idle_time of clock_pelt 7225 */ 7226 update_idle_rq_clock_pelt(rq); 7227 7228 return NULL; 7229 } 7230 7231 static struct task_struct *__pick_next_task_fair(struct rq *rq) 7232 { 7233 return pick_next_task_fair(rq, NULL, NULL); 7234 } 7235 7236 /* 7237 * Account for a descheduled task: 7238 */ 7239 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev) 7240 { 7241 struct sched_entity *se = &prev->se; 7242 struct cfs_rq *cfs_rq; 7243 7244 for_each_sched_entity(se) { 7245 cfs_rq = cfs_rq_of(se); 7246 put_prev_entity(cfs_rq, se); 7247 } 7248 } 7249 7250 /* 7251 * sched_yield() is very simple 7252 * 7253 * The magic of dealing with the ->skip buddy is in pick_next_entity. 7254 */ 7255 static void yield_task_fair(struct rq *rq) 7256 { 7257 struct task_struct *curr = rq->curr; 7258 struct cfs_rq *cfs_rq = task_cfs_rq(curr); 7259 struct sched_entity *se = &curr->se; 7260 7261 /* 7262 * Are we the only task in the tree? 7263 */ 7264 if (unlikely(rq->nr_running == 1)) 7265 return; 7266 7267 clear_buddies(cfs_rq, se); 7268 7269 if (curr->policy != SCHED_BATCH) { 7270 update_rq_clock(rq); 7271 /* 7272 * Update run-time statistics of the 'current'. 7273 */ 7274 update_curr(cfs_rq); 7275 /* 7276 * Tell update_rq_clock() that we've just updated, 7277 * so we don't do microscopic update in schedule() 7278 * and double the fastpath cost. 7279 */ 7280 rq_clock_skip_update(rq); 7281 } 7282 7283 set_skip_buddy(se); 7284 } 7285 7286 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p) 7287 { 7288 struct sched_entity *se = &p->se; 7289 7290 /* throttled hierarchies are not runnable */ 7291 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se))) 7292 return false; 7293 7294 /* Tell the scheduler that we'd really like pse to run next. */ 7295 set_next_buddy(se); 7296 7297 yield_task_fair(rq); 7298 7299 return true; 7300 } 7301 7302 #ifdef CONFIG_SMP 7303 /************************************************** 7304 * Fair scheduling class load-balancing methods. 7305 * 7306 * BASICS 7307 * 7308 * The purpose of load-balancing is to achieve the same basic fairness the 7309 * per-CPU scheduler provides, namely provide a proportional amount of compute 7310 * time to each task. This is expressed in the following equation: 7311 * 7312 * W_i,n/P_i == W_j,n/P_j for all i,j (1) 7313 * 7314 * Where W_i,n is the n-th weight average for CPU i. The instantaneous weight 7315 * W_i,0 is defined as: 7316 * 7317 * W_i,0 = \Sum_j w_i,j (2) 7318 * 7319 * Where w_i,j is the weight of the j-th runnable task on CPU i. This weight 7320 * is derived from the nice value as per sched_prio_to_weight[]. 7321 * 7322 * The weight average is an exponential decay average of the instantaneous 7323 * weight: 7324 * 7325 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3) 7326 * 7327 * C_i is the compute capacity of CPU i, typically it is the 7328 * fraction of 'recent' time available for SCHED_OTHER task execution. But it 7329 * can also include other factors [XXX]. 7330 * 7331 * To achieve this balance we define a measure of imbalance which follows 7332 * directly from (1): 7333 * 7334 * imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j } (4) 7335 * 7336 * We them move tasks around to minimize the imbalance. In the continuous 7337 * function space it is obvious this converges, in the discrete case we get 7338 * a few fun cases generally called infeasible weight scenarios. 7339 * 7340 * [XXX expand on: 7341 * - infeasible weights; 7342 * - local vs global optima in the discrete case. ] 7343 * 7344 * 7345 * SCHED DOMAINS 7346 * 7347 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2) 7348 * for all i,j solution, we create a tree of CPUs that follows the hardware 7349 * topology where each level pairs two lower groups (or better). This results 7350 * in O(log n) layers. Furthermore we reduce the number of CPUs going up the 7351 * tree to only the first of the previous level and we decrease the frequency 7352 * of load-balance at each level inv. proportional to the number of CPUs in 7353 * the groups. 7354 * 7355 * This yields: 7356 * 7357 * log_2 n 1 n 7358 * \Sum { --- * --- * 2^i } = O(n) (5) 7359 * i = 0 2^i 2^i 7360 * `- size of each group 7361 * | | `- number of CPUs doing load-balance 7362 * | `- freq 7363 * `- sum over all levels 7364 * 7365 * Coupled with a limit on how many tasks we can migrate every balance pass, 7366 * this makes (5) the runtime complexity of the balancer. 7367 * 7368 * An important property here is that each CPU is still (indirectly) connected 7369 * to every other CPU in at most O(log n) steps: 7370 * 7371 * The adjacency matrix of the resulting graph is given by: 7372 * 7373 * log_2 n 7374 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6) 7375 * k = 0 7376 * 7377 * And you'll find that: 7378 * 7379 * A^(log_2 n)_i,j != 0 for all i,j (7) 7380 * 7381 * Showing there's indeed a path between every CPU in at most O(log n) steps. 7382 * The task movement gives a factor of O(m), giving a convergence complexity 7383 * of: 7384 * 7385 * O(nm log n), n := nr_cpus, m := nr_tasks (8) 7386 * 7387 * 7388 * WORK CONSERVING 7389 * 7390 * In order to avoid CPUs going idle while there's still work to do, new idle 7391 * balancing is more aggressive and has the newly idle CPU iterate up the domain 7392 * tree itself instead of relying on other CPUs to bring it work. 7393 * 7394 * This adds some complexity to both (5) and (8) but it reduces the total idle 7395 * time. 7396 * 7397 * [XXX more?] 7398 * 7399 * 7400 * CGROUPS 7401 * 7402 * Cgroups make a horror show out of (2), instead of a simple sum we get: 7403 * 7404 * s_k,i 7405 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9) 7406 * S_k 7407 * 7408 * Where 7409 * 7410 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10) 7411 * 7412 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on CPU i. 7413 * 7414 * The big problem is S_k, its a global sum needed to compute a local (W_i) 7415 * property. 7416 * 7417 * [XXX write more on how we solve this.. _after_ merging pjt's patches that 7418 * rewrite all of this once again.] 7419 */ 7420 7421 static unsigned long __read_mostly max_load_balance_interval = HZ/10; 7422 7423 enum fbq_type { regular, remote, all }; 7424 7425 /* 7426 * 'group_type' describes the group of CPUs at the moment of load balancing. 7427 * 7428 * The enum is ordered by pulling priority, with the group with lowest priority 7429 * first so the group_type can simply be compared when selecting the busiest 7430 * group. See update_sd_pick_busiest(). 7431 */ 7432 enum group_type { 7433 /* The group has spare capacity that can be used to run more tasks. */ 7434 group_has_spare = 0, 7435 /* 7436 * The group is fully used and the tasks don't compete for more CPU 7437 * cycles. Nevertheless, some tasks might wait before running. 7438 */ 7439 group_fully_busy, 7440 /* 7441 * SD_ASYM_CPUCAPACITY only: One task doesn't fit with CPU's capacity 7442 * and must be migrated to a more powerful CPU. 7443 */ 7444 group_misfit_task, 7445 /* 7446 * SD_ASYM_PACKING only: One local CPU with higher capacity is available, 7447 * and the task should be migrated to it instead of running on the 7448 * current CPU. 7449 */ 7450 group_asym_packing, 7451 /* 7452 * The tasks' affinity constraints previously prevented the scheduler 7453 * from balancing the load across the system. 7454 */ 7455 group_imbalanced, 7456 /* 7457 * The CPU is overloaded and can't provide expected CPU cycles to all 7458 * tasks. 7459 */ 7460 group_overloaded 7461 }; 7462 7463 enum migration_type { 7464 migrate_load = 0, 7465 migrate_util, 7466 migrate_task, 7467 migrate_misfit 7468 }; 7469 7470 #define LBF_ALL_PINNED 0x01 7471 #define LBF_NEED_BREAK 0x02 7472 #define LBF_DST_PINNED 0x04 7473 #define LBF_SOME_PINNED 0x08 7474 #define LBF_ACTIVE_LB 0x10 7475 7476 struct lb_env { 7477 struct sched_domain *sd; 7478 7479 struct rq *src_rq; 7480 int src_cpu; 7481 7482 int dst_cpu; 7483 struct rq *dst_rq; 7484 7485 struct cpumask *dst_grpmask; 7486 int new_dst_cpu; 7487 enum cpu_idle_type idle; 7488 long imbalance; 7489 /* The set of CPUs under consideration for load-balancing */ 7490 struct cpumask *cpus; 7491 7492 unsigned int flags; 7493 7494 unsigned int loop; 7495 unsigned int loop_break; 7496 unsigned int loop_max; 7497 7498 enum fbq_type fbq_type; 7499 enum migration_type migration_type; 7500 struct list_head tasks; 7501 }; 7502 7503 /* 7504 * Is this task likely cache-hot: 7505 */ 7506 static int task_hot(struct task_struct *p, struct lb_env *env) 7507 { 7508 s64 delta; 7509 7510 lockdep_assert_held(&env->src_rq->lock); 7511 7512 if (p->sched_class != &fair_sched_class) 7513 return 0; 7514 7515 if (unlikely(task_has_idle_policy(p))) 7516 return 0; 7517 7518 /* SMT siblings share cache */ 7519 if (env->sd->flags & SD_SHARE_CPUCAPACITY) 7520 return 0; 7521 7522 /* 7523 * Buddy candidates are cache hot: 7524 */ 7525 if (sched_feat(CACHE_HOT_BUDDY) && env->dst_rq->nr_running && 7526 (&p->se == cfs_rq_of(&p->se)->next || 7527 &p->se == cfs_rq_of(&p->se)->last)) 7528 return 1; 7529 7530 if (sysctl_sched_migration_cost == -1) 7531 return 1; 7532 if (sysctl_sched_migration_cost == 0) 7533 return 0; 7534 7535 delta = rq_clock_task(env->src_rq) - p->se.exec_start; 7536 7537 return delta < (s64)sysctl_sched_migration_cost; 7538 } 7539 7540 #ifdef CONFIG_NUMA_BALANCING 7541 /* 7542 * Returns 1, if task migration degrades locality 7543 * Returns 0, if task migration improves locality i.e migration preferred. 7544 * Returns -1, if task migration is not affected by locality. 7545 */ 7546 static int migrate_degrades_locality(struct task_struct *p, struct lb_env *env) 7547 { 7548 struct numa_group *numa_group = rcu_dereference(p->numa_group); 7549 unsigned long src_weight, dst_weight; 7550 int src_nid, dst_nid, dist; 7551 7552 if (!static_branch_likely(&sched_numa_balancing)) 7553 return -1; 7554 7555 if (!p->numa_faults || !(env->sd->flags & SD_NUMA)) 7556 return -1; 7557 7558 src_nid = cpu_to_node(env->src_cpu); 7559 dst_nid = cpu_to_node(env->dst_cpu); 7560 7561 if (src_nid == dst_nid) 7562 return -1; 7563 7564 /* Migrating away from the preferred node is always bad. */ 7565 if (src_nid == p->numa_preferred_nid) { 7566 if (env->src_rq->nr_running > env->src_rq->nr_preferred_running) 7567 return 1; 7568 else 7569 return -1; 7570 } 7571 7572 /* Encourage migration to the preferred node. */ 7573 if (dst_nid == p->numa_preferred_nid) 7574 return 0; 7575 7576 /* Leaving a core idle is often worse than degrading locality. */ 7577 if (env->idle == CPU_IDLE) 7578 return -1; 7579 7580 dist = node_distance(src_nid, dst_nid); 7581 if (numa_group) { 7582 src_weight = group_weight(p, src_nid, dist); 7583 dst_weight = group_weight(p, dst_nid, dist); 7584 } else { 7585 src_weight = task_weight(p, src_nid, dist); 7586 dst_weight = task_weight(p, dst_nid, dist); 7587 } 7588 7589 return dst_weight < src_weight; 7590 } 7591 7592 #else 7593 static inline int migrate_degrades_locality(struct task_struct *p, 7594 struct lb_env *env) 7595 { 7596 return -1; 7597 } 7598 #endif 7599 7600 /* 7601 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu? 7602 */ 7603 static 7604 int can_migrate_task(struct task_struct *p, struct lb_env *env) 7605 { 7606 int tsk_cache_hot; 7607 7608 lockdep_assert_held(&env->src_rq->lock); 7609 7610 /* 7611 * We do not migrate tasks that are: 7612 * 1) throttled_lb_pair, or 7613 * 2) cannot be migrated to this CPU due to cpus_ptr, or 7614 * 3) running (obviously), or 7615 * 4) are cache-hot on their current CPU. 7616 */ 7617 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu)) 7618 return 0; 7619 7620 /* Disregard pcpu kthreads; they are where they need to be. */ 7621 if (kthread_is_per_cpu(p)) 7622 return 0; 7623 7624 if (!cpumask_test_cpu(env->dst_cpu, p->cpus_ptr)) { 7625 int cpu; 7626 7627 schedstat_inc(p->se.statistics.nr_failed_migrations_affine); 7628 7629 env->flags |= LBF_SOME_PINNED; 7630 7631 /* 7632 * Remember if this task can be migrated to any other CPU in 7633 * our sched_group. We may want to revisit it if we couldn't 7634 * meet load balance goals by pulling other tasks on src_cpu. 7635 * 7636 * Avoid computing new_dst_cpu 7637 * - for NEWLY_IDLE 7638 * - if we have already computed one in current iteration 7639 * - if it's an active balance 7640 */ 7641 if (env->idle == CPU_NEWLY_IDLE || 7642 env->flags & (LBF_DST_PINNED | LBF_ACTIVE_LB)) 7643 return 0; 7644 7645 /* Prevent to re-select dst_cpu via env's CPUs: */ 7646 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) { 7647 if (cpumask_test_cpu(cpu, p->cpus_ptr)) { 7648 env->flags |= LBF_DST_PINNED; 7649 env->new_dst_cpu = cpu; 7650 break; 7651 } 7652 } 7653 7654 return 0; 7655 } 7656 7657 /* Record that we found at least one task that could run on dst_cpu */ 7658 env->flags &= ~LBF_ALL_PINNED; 7659 7660 if (task_running(env->src_rq, p)) { 7661 schedstat_inc(p->se.statistics.nr_failed_migrations_running); 7662 return 0; 7663 } 7664 7665 /* 7666 * Aggressive migration if: 7667 * 1) active balance 7668 * 2) destination numa is preferred 7669 * 3) task is cache cold, or 7670 * 4) too many balance attempts have failed. 7671 */ 7672 if (env->flags & LBF_ACTIVE_LB) 7673 return 1; 7674 7675 tsk_cache_hot = migrate_degrades_locality(p, env); 7676 if (tsk_cache_hot == -1) 7677 tsk_cache_hot = task_hot(p, env); 7678 7679 if (tsk_cache_hot <= 0 || 7680 env->sd->nr_balance_failed > env->sd->cache_nice_tries) { 7681 if (tsk_cache_hot == 1) { 7682 schedstat_inc(env->sd->lb_hot_gained[env->idle]); 7683 schedstat_inc(p->se.statistics.nr_forced_migrations); 7684 } 7685 return 1; 7686 } 7687 7688 schedstat_inc(p->se.statistics.nr_failed_migrations_hot); 7689 return 0; 7690 } 7691 7692 /* 7693 * detach_task() -- detach the task for the migration specified in env 7694 */ 7695 static void detach_task(struct task_struct *p, struct lb_env *env) 7696 { 7697 lockdep_assert_held(&env->src_rq->lock); 7698 7699 deactivate_task(env->src_rq, p, DEQUEUE_NOCLOCK); 7700 set_task_cpu(p, env->dst_cpu); 7701 } 7702 7703 /* 7704 * detach_one_task() -- tries to dequeue exactly one task from env->src_rq, as 7705 * part of active balancing operations within "domain". 7706 * 7707 * Returns a task if successful and NULL otherwise. 7708 */ 7709 static struct task_struct *detach_one_task(struct lb_env *env) 7710 { 7711 struct task_struct *p; 7712 7713 lockdep_assert_held(&env->src_rq->lock); 7714 7715 list_for_each_entry_reverse(p, 7716 &env->src_rq->cfs_tasks, se.group_node) { 7717 if (!can_migrate_task(p, env)) 7718 continue; 7719 7720 detach_task(p, env); 7721 7722 /* 7723 * Right now, this is only the second place where 7724 * lb_gained[env->idle] is updated (other is detach_tasks) 7725 * so we can safely collect stats here rather than 7726 * inside detach_tasks(). 7727 */ 7728 schedstat_inc(env->sd->lb_gained[env->idle]); 7729 return p; 7730 } 7731 return NULL; 7732 } 7733 7734 static const unsigned int sched_nr_migrate_break = 32; 7735 7736 /* 7737 * detach_tasks() -- tries to detach up to imbalance load/util/tasks from 7738 * busiest_rq, as part of a balancing operation within domain "sd". 7739 * 7740 * Returns number of detached tasks if successful and 0 otherwise. 7741 */ 7742 static int detach_tasks(struct lb_env *env) 7743 { 7744 struct list_head *tasks = &env->src_rq->cfs_tasks; 7745 unsigned long util, load; 7746 struct task_struct *p; 7747 int detached = 0; 7748 7749 lockdep_assert_held(&env->src_rq->lock); 7750 7751 /* 7752 * Source run queue has been emptied by another CPU, clear 7753 * LBF_ALL_PINNED flag as we will not test any task. 7754 */ 7755 if (env->src_rq->nr_running <= 1) { 7756 env->flags &= ~LBF_ALL_PINNED; 7757 return 0; 7758 } 7759 7760 if (env->imbalance <= 0) 7761 return 0; 7762 7763 while (!list_empty(tasks)) { 7764 /* 7765 * We don't want to steal all, otherwise we may be treated likewise, 7766 * which could at worst lead to a livelock crash. 7767 */ 7768 if (env->idle != CPU_NOT_IDLE && env->src_rq->nr_running <= 1) 7769 break; 7770 7771 p = list_last_entry(tasks, struct task_struct, se.group_node); 7772 7773 env->loop++; 7774 /* We've more or less seen every task there is, call it quits */ 7775 if (env->loop > env->loop_max) 7776 break; 7777 7778 /* take a breather every nr_migrate tasks */ 7779 if (env->loop > env->loop_break) { 7780 env->loop_break += sched_nr_migrate_break; 7781 env->flags |= LBF_NEED_BREAK; 7782 break; 7783 } 7784 7785 if (!can_migrate_task(p, env)) 7786 goto next; 7787 7788 switch (env->migration_type) { 7789 case migrate_load: 7790 /* 7791 * Depending of the number of CPUs and tasks and the 7792 * cgroup hierarchy, task_h_load() can return a null 7793 * value. Make sure that env->imbalance decreases 7794 * otherwise detach_tasks() will stop only after 7795 * detaching up to loop_max tasks. 7796 */ 7797 load = max_t(unsigned long, task_h_load(p), 1); 7798 7799 if (sched_feat(LB_MIN) && 7800 load < 16 && !env->sd->nr_balance_failed) 7801 goto next; 7802 7803 /* 7804 * Make sure that we don't migrate too much load. 7805 * Nevertheless, let relax the constraint if 7806 * scheduler fails to find a good waiting task to 7807 * migrate. 7808 */ 7809 if (shr_bound(load, env->sd->nr_balance_failed) > env->imbalance) 7810 goto next; 7811 7812 env->imbalance -= load; 7813 break; 7814 7815 case migrate_util: 7816 util = task_util_est(p); 7817 7818 if (util > env->imbalance) 7819 goto next; 7820 7821 env->imbalance -= util; 7822 break; 7823 7824 case migrate_task: 7825 env->imbalance--; 7826 break; 7827 7828 case migrate_misfit: 7829 /* This is not a misfit task */ 7830 if (task_fits_capacity(p, capacity_of(env->src_cpu))) 7831 goto next; 7832 7833 env->imbalance = 0; 7834 break; 7835 } 7836 7837 detach_task(p, env); 7838 list_add(&p->se.group_node, &env->tasks); 7839 7840 detached++; 7841 7842 #ifdef CONFIG_PREEMPTION 7843 /* 7844 * NEWIDLE balancing is a source of latency, so preemptible 7845 * kernels will stop after the first task is detached to minimize 7846 * the critical section. 7847 */ 7848 if (env->idle == CPU_NEWLY_IDLE) 7849 break; 7850 #endif 7851 7852 /* 7853 * We only want to steal up to the prescribed amount of 7854 * load/util/tasks. 7855 */ 7856 if (env->imbalance <= 0) 7857 break; 7858 7859 continue; 7860 next: 7861 list_move(&p->se.group_node, tasks); 7862 } 7863 7864 /* 7865 * Right now, this is one of only two places we collect this stat 7866 * so we can safely collect detach_one_task() stats here rather 7867 * than inside detach_one_task(). 7868 */ 7869 schedstat_add(env->sd->lb_gained[env->idle], detached); 7870 7871 return detached; 7872 } 7873 7874 /* 7875 * attach_task() -- attach the task detached by detach_task() to its new rq. 7876 */ 7877 static void attach_task(struct rq *rq, struct task_struct *p) 7878 { 7879 lockdep_assert_held(&rq->lock); 7880 7881 BUG_ON(task_rq(p) != rq); 7882 activate_task(rq, p, ENQUEUE_NOCLOCK); 7883 check_preempt_curr(rq, p, 0); 7884 } 7885 7886 /* 7887 * attach_one_task() -- attaches the task returned from detach_one_task() to 7888 * its new rq. 7889 */ 7890 static void attach_one_task(struct rq *rq, struct task_struct *p) 7891 { 7892 struct rq_flags rf; 7893 7894 rq_lock(rq, &rf); 7895 update_rq_clock(rq); 7896 attach_task(rq, p); 7897 rq_unlock(rq, &rf); 7898 } 7899 7900 /* 7901 * attach_tasks() -- attaches all tasks detached by detach_tasks() to their 7902 * new rq. 7903 */ 7904 static void attach_tasks(struct lb_env *env) 7905 { 7906 struct list_head *tasks = &env->tasks; 7907 struct task_struct *p; 7908 struct rq_flags rf; 7909 7910 rq_lock(env->dst_rq, &rf); 7911 update_rq_clock(env->dst_rq); 7912 7913 while (!list_empty(tasks)) { 7914 p = list_first_entry(tasks, struct task_struct, se.group_node); 7915 list_del_init(&p->se.group_node); 7916 7917 attach_task(env->dst_rq, p); 7918 } 7919 7920 rq_unlock(env->dst_rq, &rf); 7921 } 7922 7923 #ifdef CONFIG_NO_HZ_COMMON 7924 static inline bool cfs_rq_has_blocked(struct cfs_rq *cfs_rq) 7925 { 7926 if (cfs_rq->avg.load_avg) 7927 return true; 7928 7929 if (cfs_rq->avg.util_avg) 7930 return true; 7931 7932 return false; 7933 } 7934 7935 static inline bool others_have_blocked(struct rq *rq) 7936 { 7937 if (READ_ONCE(rq->avg_rt.util_avg)) 7938 return true; 7939 7940 if (READ_ONCE(rq->avg_dl.util_avg)) 7941 return true; 7942 7943 if (thermal_load_avg(rq)) 7944 return true; 7945 7946 #ifdef CONFIG_HAVE_SCHED_AVG_IRQ 7947 if (READ_ONCE(rq->avg_irq.util_avg)) 7948 return true; 7949 #endif 7950 7951 return false; 7952 } 7953 7954 static inline void update_blocked_load_tick(struct rq *rq) 7955 { 7956 WRITE_ONCE(rq->last_blocked_load_update_tick, jiffies); 7957 } 7958 7959 static inline void update_blocked_load_status(struct rq *rq, bool has_blocked) 7960 { 7961 if (!has_blocked) 7962 rq->has_blocked_load = 0; 7963 } 7964 #else 7965 static inline bool cfs_rq_has_blocked(struct cfs_rq *cfs_rq) { return false; } 7966 static inline bool others_have_blocked(struct rq *rq) { return false; } 7967 static inline void update_blocked_load_tick(struct rq *rq) {} 7968 static inline void update_blocked_load_status(struct rq *rq, bool has_blocked) {} 7969 #endif 7970 7971 static bool __update_blocked_others(struct rq *rq, bool *done) 7972 { 7973 const struct sched_class *curr_class; 7974 u64 now = rq_clock_pelt(rq); 7975 unsigned long thermal_pressure; 7976 bool decayed; 7977 7978 /* 7979 * update_load_avg() can call cpufreq_update_util(). Make sure that RT, 7980 * DL and IRQ signals have been updated before updating CFS. 7981 */ 7982 curr_class = rq->curr->sched_class; 7983 7984 thermal_pressure = arch_scale_thermal_pressure(cpu_of(rq)); 7985 7986 decayed = update_rt_rq_load_avg(now, rq, curr_class == &rt_sched_class) | 7987 update_dl_rq_load_avg(now, rq, curr_class == &dl_sched_class) | 7988 update_thermal_load_avg(rq_clock_thermal(rq), rq, thermal_pressure) | 7989 update_irq_load_avg(rq, 0); 7990 7991 if (others_have_blocked(rq)) 7992 *done = false; 7993 7994 return decayed; 7995 } 7996 7997 #ifdef CONFIG_FAIR_GROUP_SCHED 7998 7999 static inline bool cfs_rq_is_decayed(struct cfs_rq *cfs_rq) 8000 { 8001 if (cfs_rq->load.weight) 8002 return false; 8003 8004 if (cfs_rq->avg.load_sum) 8005 return false; 8006 8007 if (cfs_rq->avg.util_sum) 8008 return false; 8009 8010 if (cfs_rq->avg.runnable_sum) 8011 return false; 8012 8013 return true; 8014 } 8015 8016 static bool __update_blocked_fair(struct rq *rq, bool *done) 8017 { 8018 struct cfs_rq *cfs_rq, *pos; 8019 bool decayed = false; 8020 int cpu = cpu_of(rq); 8021 8022 /* 8023 * Iterates the task_group tree in a bottom up fashion, see 8024 * list_add_leaf_cfs_rq() for details. 8025 */ 8026 for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos) { 8027 struct sched_entity *se; 8028 8029 if (update_cfs_rq_load_avg(cfs_rq_clock_pelt(cfs_rq), cfs_rq)) { 8030 update_tg_load_avg(cfs_rq); 8031 8032 if (cfs_rq == &rq->cfs) 8033 decayed = true; 8034 } 8035 8036 /* Propagate pending load changes to the parent, if any: */ 8037 se = cfs_rq->tg->se[cpu]; 8038 if (se && !skip_blocked_update(se)) 8039 update_load_avg(cfs_rq_of(se), se, UPDATE_TG); 8040 8041 /* 8042 * There can be a lot of idle CPU cgroups. Don't let fully 8043 * decayed cfs_rqs linger on the list. 8044 */ 8045 if (cfs_rq_is_decayed(cfs_rq)) 8046 list_del_leaf_cfs_rq(cfs_rq); 8047 8048 /* Don't need periodic decay once load/util_avg are null */ 8049 if (cfs_rq_has_blocked(cfs_rq)) 8050 *done = false; 8051 } 8052 8053 return decayed; 8054 } 8055 8056 /* 8057 * Compute the hierarchical load factor for cfs_rq and all its ascendants. 8058 * This needs to be done in a top-down fashion because the load of a child 8059 * group is a fraction of its parents load. 8060 */ 8061 static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq) 8062 { 8063 struct rq *rq = rq_of(cfs_rq); 8064 struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)]; 8065 unsigned long now = jiffies; 8066 unsigned long load; 8067 8068 if (cfs_rq->last_h_load_update == now) 8069 return; 8070 8071 WRITE_ONCE(cfs_rq->h_load_next, NULL); 8072 for_each_sched_entity(se) { 8073 cfs_rq = cfs_rq_of(se); 8074 WRITE_ONCE(cfs_rq->h_load_next, se); 8075 if (cfs_rq->last_h_load_update == now) 8076 break; 8077 } 8078 8079 if (!se) { 8080 cfs_rq->h_load = cfs_rq_load_avg(cfs_rq); 8081 cfs_rq->last_h_load_update = now; 8082 } 8083 8084 while ((se = READ_ONCE(cfs_rq->h_load_next)) != NULL) { 8085 load = cfs_rq->h_load; 8086 load = div64_ul(load * se->avg.load_avg, 8087 cfs_rq_load_avg(cfs_rq) + 1); 8088 cfs_rq = group_cfs_rq(se); 8089 cfs_rq->h_load = load; 8090 cfs_rq->last_h_load_update = now; 8091 } 8092 } 8093 8094 static unsigned long task_h_load(struct task_struct *p) 8095 { 8096 struct cfs_rq *cfs_rq = task_cfs_rq(p); 8097 8098 update_cfs_rq_h_load(cfs_rq); 8099 return div64_ul(p->se.avg.load_avg * cfs_rq->h_load, 8100 cfs_rq_load_avg(cfs_rq) + 1); 8101 } 8102 #else 8103 static bool __update_blocked_fair(struct rq *rq, bool *done) 8104 { 8105 struct cfs_rq *cfs_rq = &rq->cfs; 8106 bool decayed; 8107 8108 decayed = update_cfs_rq_load_avg(cfs_rq_clock_pelt(cfs_rq), cfs_rq); 8109 if (cfs_rq_has_blocked(cfs_rq)) 8110 *done = false; 8111 8112 return decayed; 8113 } 8114 8115 static unsigned long task_h_load(struct task_struct *p) 8116 { 8117 return p->se.avg.load_avg; 8118 } 8119 #endif 8120 8121 static void update_blocked_averages(int cpu) 8122 { 8123 bool decayed = false, done = true; 8124 struct rq *rq = cpu_rq(cpu); 8125 struct rq_flags rf; 8126 8127 rq_lock_irqsave(rq, &rf); 8128 update_blocked_load_tick(rq); 8129 update_rq_clock(rq); 8130 8131 decayed |= __update_blocked_others(rq, &done); 8132 decayed |= __update_blocked_fair(rq, &done); 8133 8134 update_blocked_load_status(rq, !done); 8135 if (decayed) 8136 cpufreq_update_util(rq, 0); 8137 rq_unlock_irqrestore(rq, &rf); 8138 } 8139 8140 /********** Helpers for find_busiest_group ************************/ 8141 8142 /* 8143 * sg_lb_stats - stats of a sched_group required for load_balancing 8144 */ 8145 struct sg_lb_stats { 8146 unsigned long avg_load; /*Avg load across the CPUs of the group */ 8147 unsigned long group_load; /* Total load over the CPUs of the group */ 8148 unsigned long group_capacity; 8149 unsigned long group_util; /* Total utilization over the CPUs of the group */ 8150 unsigned long group_runnable; /* Total runnable time over the CPUs of the group */ 8151 unsigned int sum_nr_running; /* Nr of tasks running in the group */ 8152 unsigned int sum_h_nr_running; /* Nr of CFS tasks running in the group */ 8153 unsigned int idle_cpus; 8154 unsigned int group_weight; 8155 enum group_type group_type; 8156 unsigned int group_asym_packing; /* Tasks should be moved to preferred CPU */ 8157 unsigned long group_misfit_task_load; /* A CPU has a task too big for its capacity */ 8158 #ifdef CONFIG_NUMA_BALANCING 8159 unsigned int nr_numa_running; 8160 unsigned int nr_preferred_running; 8161 #endif 8162 }; 8163 8164 /* 8165 * sd_lb_stats - Structure to store the statistics of a sched_domain 8166 * during load balancing. 8167 */ 8168 struct sd_lb_stats { 8169 struct sched_group *busiest; /* Busiest group in this sd */ 8170 struct sched_group *local; /* Local group in this sd */ 8171 unsigned long total_load; /* Total load of all groups in sd */ 8172 unsigned long total_capacity; /* Total capacity of all groups in sd */ 8173 unsigned long avg_load; /* Average load across all groups in sd */ 8174 unsigned int prefer_sibling; /* tasks should go to sibling first */ 8175 8176 struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */ 8177 struct sg_lb_stats local_stat; /* Statistics of the local group */ 8178 }; 8179 8180 static inline void init_sd_lb_stats(struct sd_lb_stats *sds) 8181 { 8182 /* 8183 * Skimp on the clearing to avoid duplicate work. We can avoid clearing 8184 * local_stat because update_sg_lb_stats() does a full clear/assignment. 8185 * We must however set busiest_stat::group_type and 8186 * busiest_stat::idle_cpus to the worst busiest group because 8187 * update_sd_pick_busiest() reads these before assignment. 8188 */ 8189 *sds = (struct sd_lb_stats){ 8190 .busiest = NULL, 8191 .local = NULL, 8192 .total_load = 0UL, 8193 .total_capacity = 0UL, 8194 .busiest_stat = { 8195 .idle_cpus = UINT_MAX, 8196 .group_type = group_has_spare, 8197 }, 8198 }; 8199 } 8200 8201 static unsigned long scale_rt_capacity(int cpu) 8202 { 8203 struct rq *rq = cpu_rq(cpu); 8204 unsigned long max = arch_scale_cpu_capacity(cpu); 8205 unsigned long used, free; 8206 unsigned long irq; 8207 8208 irq = cpu_util_irq(rq); 8209 8210 if (unlikely(irq >= max)) 8211 return 1; 8212 8213 /* 8214 * avg_rt.util_avg and avg_dl.util_avg track binary signals 8215 * (running and not running) with weights 0 and 1024 respectively. 8216 * avg_thermal.load_avg tracks thermal pressure and the weighted 8217 * average uses the actual delta max capacity(load). 8218 */ 8219 used = READ_ONCE(rq->avg_rt.util_avg); 8220 used += READ_ONCE(rq->avg_dl.util_avg); 8221 used += thermal_load_avg(rq); 8222 8223 if (unlikely(used >= max)) 8224 return 1; 8225 8226 free = max - used; 8227 8228 return scale_irq_capacity(free, irq, max); 8229 } 8230 8231 static void update_cpu_capacity(struct sched_domain *sd, int cpu) 8232 { 8233 unsigned long capacity = scale_rt_capacity(cpu); 8234 struct sched_group *sdg = sd->groups; 8235 8236 cpu_rq(cpu)->cpu_capacity_orig = arch_scale_cpu_capacity(cpu); 8237 8238 if (!capacity) 8239 capacity = 1; 8240 8241 cpu_rq(cpu)->cpu_capacity = capacity; 8242 trace_sched_cpu_capacity_tp(cpu_rq(cpu)); 8243 8244 sdg->sgc->capacity = capacity; 8245 sdg->sgc->min_capacity = capacity; 8246 sdg->sgc->max_capacity = capacity; 8247 } 8248 8249 void update_group_capacity(struct sched_domain *sd, int cpu) 8250 { 8251 struct sched_domain *child = sd->child; 8252 struct sched_group *group, *sdg = sd->groups; 8253 unsigned long capacity, min_capacity, max_capacity; 8254 unsigned long interval; 8255 8256 interval = msecs_to_jiffies(sd->balance_interval); 8257 interval = clamp(interval, 1UL, max_load_balance_interval); 8258 sdg->sgc->next_update = jiffies + interval; 8259 8260 if (!child) { 8261 update_cpu_capacity(sd, cpu); 8262 return; 8263 } 8264 8265 capacity = 0; 8266 min_capacity = ULONG_MAX; 8267 max_capacity = 0; 8268 8269 if (child->flags & SD_OVERLAP) { 8270 /* 8271 * SD_OVERLAP domains cannot assume that child groups 8272 * span the current group. 8273 */ 8274 8275 for_each_cpu(cpu, sched_group_span(sdg)) { 8276 unsigned long cpu_cap = capacity_of(cpu); 8277 8278 capacity += cpu_cap; 8279 min_capacity = min(cpu_cap, min_capacity); 8280 max_capacity = max(cpu_cap, max_capacity); 8281 } 8282 } else { 8283 /* 8284 * !SD_OVERLAP domains can assume that child groups 8285 * span the current group. 8286 */ 8287 8288 group = child->groups; 8289 do { 8290 struct sched_group_capacity *sgc = group->sgc; 8291 8292 capacity += sgc->capacity; 8293 min_capacity = min(sgc->min_capacity, min_capacity); 8294 max_capacity = max(sgc->max_capacity, max_capacity); 8295 group = group->next; 8296 } while (group != child->groups); 8297 } 8298 8299 sdg->sgc->capacity = capacity; 8300 sdg->sgc->min_capacity = min_capacity; 8301 sdg->sgc->max_capacity = max_capacity; 8302 } 8303 8304 /* 8305 * Check whether the capacity of the rq has been noticeably reduced by side 8306 * activity. The imbalance_pct is used for the threshold. 8307 * Return true is the capacity is reduced 8308 */ 8309 static inline int 8310 check_cpu_capacity(struct rq *rq, struct sched_domain *sd) 8311 { 8312 return ((rq->cpu_capacity * sd->imbalance_pct) < 8313 (rq->cpu_capacity_orig * 100)); 8314 } 8315 8316 /* 8317 * Check whether a rq has a misfit task and if it looks like we can actually 8318 * help that task: we can migrate the task to a CPU of higher capacity, or 8319 * the task's current CPU is heavily pressured. 8320 */ 8321 static inline int check_misfit_status(struct rq *rq, struct sched_domain *sd) 8322 { 8323 return rq->misfit_task_load && 8324 (rq->cpu_capacity_orig < rq->rd->max_cpu_capacity || 8325 check_cpu_capacity(rq, sd)); 8326 } 8327 8328 /* 8329 * Group imbalance indicates (and tries to solve) the problem where balancing 8330 * groups is inadequate due to ->cpus_ptr constraints. 8331 * 8332 * Imagine a situation of two groups of 4 CPUs each and 4 tasks each with a 8333 * cpumask covering 1 CPU of the first group and 3 CPUs of the second group. 8334 * Something like: 8335 * 8336 * { 0 1 2 3 } { 4 5 6 7 } 8337 * * * * * 8338 * 8339 * If we were to balance group-wise we'd place two tasks in the first group and 8340 * two tasks in the second group. Clearly this is undesired as it will overload 8341 * cpu 3 and leave one of the CPUs in the second group unused. 8342 * 8343 * The current solution to this issue is detecting the skew in the first group 8344 * by noticing the lower domain failed to reach balance and had difficulty 8345 * moving tasks due to affinity constraints. 8346 * 8347 * When this is so detected; this group becomes a candidate for busiest; see 8348 * update_sd_pick_busiest(). And calculate_imbalance() and 8349 * find_busiest_group() avoid some of the usual balance conditions to allow it 8350 * to create an effective group imbalance. 8351 * 8352 * This is a somewhat tricky proposition since the next run might not find the 8353 * group imbalance and decide the groups need to be balanced again. A most 8354 * subtle and fragile situation. 8355 */ 8356 8357 static inline int sg_imbalanced(struct sched_group *group) 8358 { 8359 return group->sgc->imbalance; 8360 } 8361 8362 /* 8363 * group_has_capacity returns true if the group has spare capacity that could 8364 * be used by some tasks. 8365 * We consider that a group has spare capacity if the * number of task is 8366 * smaller than the number of CPUs or if the utilization is lower than the 8367 * available capacity for CFS tasks. 8368 * For the latter, we use a threshold to stabilize the state, to take into 8369 * account the variance of the tasks' load and to return true if the available 8370 * capacity in meaningful for the load balancer. 8371 * As an example, an available capacity of 1% can appear but it doesn't make 8372 * any benefit for the load balance. 8373 */ 8374 static inline bool 8375 group_has_capacity(unsigned int imbalance_pct, struct sg_lb_stats *sgs) 8376 { 8377 if (sgs->sum_nr_running < sgs->group_weight) 8378 return true; 8379 8380 if ((sgs->group_capacity * imbalance_pct) < 8381 (sgs->group_runnable * 100)) 8382 return false; 8383 8384 if ((sgs->group_capacity * 100) > 8385 (sgs->group_util * imbalance_pct)) 8386 return true; 8387 8388 return false; 8389 } 8390 8391 /* 8392 * group_is_overloaded returns true if the group has more tasks than it can 8393 * handle. 8394 * group_is_overloaded is not equals to !group_has_capacity because a group 8395 * with the exact right number of tasks, has no more spare capacity but is not 8396 * overloaded so both group_has_capacity and group_is_overloaded return 8397 * false. 8398 */ 8399 static inline bool 8400 group_is_overloaded(unsigned int imbalance_pct, struct sg_lb_stats *sgs) 8401 { 8402 if (sgs->sum_nr_running <= sgs->group_weight) 8403 return false; 8404 8405 if ((sgs->group_capacity * 100) < 8406 (sgs->group_util * imbalance_pct)) 8407 return true; 8408 8409 if ((sgs->group_capacity * imbalance_pct) < 8410 (sgs->group_runnable * 100)) 8411 return true; 8412 8413 return false; 8414 } 8415 8416 static inline enum 8417 group_type group_classify(unsigned int imbalance_pct, 8418 struct sched_group *group, 8419 struct sg_lb_stats *sgs) 8420 { 8421 if (group_is_overloaded(imbalance_pct, sgs)) 8422 return group_overloaded; 8423 8424 if (sg_imbalanced(group)) 8425 return group_imbalanced; 8426 8427 if (sgs->group_asym_packing) 8428 return group_asym_packing; 8429 8430 if (sgs->group_misfit_task_load) 8431 return group_misfit_task; 8432 8433 if (!group_has_capacity(imbalance_pct, sgs)) 8434 return group_fully_busy; 8435 8436 return group_has_spare; 8437 } 8438 8439 /** 8440 * update_sg_lb_stats - Update sched_group's statistics for load balancing. 8441 * @env: The load balancing environment. 8442 * @group: sched_group whose statistics are to be updated. 8443 * @sgs: variable to hold the statistics for this group. 8444 * @sg_status: Holds flag indicating the status of the sched_group 8445 */ 8446 static inline void update_sg_lb_stats(struct lb_env *env, 8447 struct sched_group *group, 8448 struct sg_lb_stats *sgs, 8449 int *sg_status) 8450 { 8451 int i, nr_running, local_group; 8452 8453 memset(sgs, 0, sizeof(*sgs)); 8454 8455 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_span(group)); 8456 8457 for_each_cpu_and(i, sched_group_span(group), env->cpus) { 8458 struct rq *rq = cpu_rq(i); 8459 8460 sgs->group_load += cpu_load(rq); 8461 sgs->group_util += cpu_util(i); 8462 sgs->group_runnable += cpu_runnable(rq); 8463 sgs->sum_h_nr_running += rq->cfs.h_nr_running; 8464 8465 nr_running = rq->nr_running; 8466 sgs->sum_nr_running += nr_running; 8467 8468 if (nr_running > 1) 8469 *sg_status |= SG_OVERLOAD; 8470 8471 if (cpu_overutilized(i)) 8472 *sg_status |= SG_OVERUTILIZED; 8473 8474 #ifdef CONFIG_NUMA_BALANCING 8475 sgs->nr_numa_running += rq->nr_numa_running; 8476 sgs->nr_preferred_running += rq->nr_preferred_running; 8477 #endif 8478 /* 8479 * No need to call idle_cpu() if nr_running is not 0 8480 */ 8481 if (!nr_running && idle_cpu(i)) { 8482 sgs->idle_cpus++; 8483 /* Idle cpu can't have misfit task */ 8484 continue; 8485 } 8486 8487 if (local_group) 8488 continue; 8489 8490 /* Check for a misfit task on the cpu */ 8491 if (env->sd->flags & SD_ASYM_CPUCAPACITY && 8492 sgs->group_misfit_task_load < rq->misfit_task_load) { 8493 sgs->group_misfit_task_load = rq->misfit_task_load; 8494 *sg_status |= SG_OVERLOAD; 8495 } 8496 } 8497 8498 /* Check if dst CPU is idle and preferred to this group */ 8499 if (env->sd->flags & SD_ASYM_PACKING && 8500 env->idle != CPU_NOT_IDLE && 8501 sgs->sum_h_nr_running && 8502 sched_asym_prefer(env->dst_cpu, group->asym_prefer_cpu)) { 8503 sgs->group_asym_packing = 1; 8504 } 8505 8506 sgs->group_capacity = group->sgc->capacity; 8507 8508 sgs->group_weight = group->group_weight; 8509 8510 sgs->group_type = group_classify(env->sd->imbalance_pct, group, sgs); 8511 8512 /* Computing avg_load makes sense only when group is overloaded */ 8513 if (sgs->group_type == group_overloaded) 8514 sgs->avg_load = (sgs->group_load * SCHED_CAPACITY_SCALE) / 8515 sgs->group_capacity; 8516 } 8517 8518 /** 8519 * update_sd_pick_busiest - return 1 on busiest group 8520 * @env: The load balancing environment. 8521 * @sds: sched_domain statistics 8522 * @sg: sched_group candidate to be checked for being the busiest 8523 * @sgs: sched_group statistics 8524 * 8525 * Determine if @sg is a busier group than the previously selected 8526 * busiest group. 8527 * 8528 * Return: %true if @sg is a busier group than the previously selected 8529 * busiest group. %false otherwise. 8530 */ 8531 static bool update_sd_pick_busiest(struct lb_env *env, 8532 struct sd_lb_stats *sds, 8533 struct sched_group *sg, 8534 struct sg_lb_stats *sgs) 8535 { 8536 struct sg_lb_stats *busiest = &sds->busiest_stat; 8537 8538 /* Make sure that there is at least one task to pull */ 8539 if (!sgs->sum_h_nr_running) 8540 return false; 8541 8542 /* 8543 * Don't try to pull misfit tasks we can't help. 8544 * We can use max_capacity here as reduction in capacity on some 8545 * CPUs in the group should either be possible to resolve 8546 * internally or be covered by avg_load imbalance (eventually). 8547 */ 8548 if (sgs->group_type == group_misfit_task && 8549 (!capacity_greater(capacity_of(env->dst_cpu), sg->sgc->max_capacity) || 8550 sds->local_stat.group_type != group_has_spare)) 8551 return false; 8552 8553 if (sgs->group_type > busiest->group_type) 8554 return true; 8555 8556 if (sgs->group_type < busiest->group_type) 8557 return false; 8558 8559 /* 8560 * The candidate and the current busiest group are the same type of 8561 * group. Let check which one is the busiest according to the type. 8562 */ 8563 8564 switch (sgs->group_type) { 8565 case group_overloaded: 8566 /* Select the overloaded group with highest avg_load. */ 8567 if (sgs->avg_load <= busiest->avg_load) 8568 return false; 8569 break; 8570 8571 case group_imbalanced: 8572 /* 8573 * Select the 1st imbalanced group as we don't have any way to 8574 * choose one more than another. 8575 */ 8576 return false; 8577 8578 case group_asym_packing: 8579 /* Prefer to move from lowest priority CPU's work */ 8580 if (sched_asym_prefer(sg->asym_prefer_cpu, sds->busiest->asym_prefer_cpu)) 8581 return false; 8582 break; 8583 8584 case group_misfit_task: 8585 /* 8586 * If we have more than one misfit sg go with the biggest 8587 * misfit. 8588 */ 8589 if (sgs->group_misfit_task_load < busiest->group_misfit_task_load) 8590 return false; 8591 break; 8592 8593 case group_fully_busy: 8594 /* 8595 * Select the fully busy group with highest avg_load. In 8596 * theory, there is no need to pull task from such kind of 8597 * group because tasks have all compute capacity that they need 8598 * but we can still improve the overall throughput by reducing 8599 * contention when accessing shared HW resources. 8600 * 8601 * XXX for now avg_load is not computed and always 0 so we 8602 * select the 1st one. 8603 */ 8604 if (sgs->avg_load <= busiest->avg_load) 8605 return false; 8606 break; 8607 8608 case group_has_spare: 8609 /* 8610 * Select not overloaded group with lowest number of idle cpus 8611 * and highest number of running tasks. We could also compare 8612 * the spare capacity which is more stable but it can end up 8613 * that the group has less spare capacity but finally more idle 8614 * CPUs which means less opportunity to pull tasks. 8615 */ 8616 if (sgs->idle_cpus > busiest->idle_cpus) 8617 return false; 8618 else if ((sgs->idle_cpus == busiest->idle_cpus) && 8619 (sgs->sum_nr_running <= busiest->sum_nr_running)) 8620 return false; 8621 8622 break; 8623 } 8624 8625 /* 8626 * Candidate sg has no more than one task per CPU and has higher 8627 * per-CPU capacity. Migrating tasks to less capable CPUs may harm 8628 * throughput. Maximize throughput, power/energy consequences are not 8629 * considered. 8630 */ 8631 if ((env->sd->flags & SD_ASYM_CPUCAPACITY) && 8632 (sgs->group_type <= group_fully_busy) && 8633 (capacity_greater(sg->sgc->min_capacity, capacity_of(env->dst_cpu)))) 8634 return false; 8635 8636 return true; 8637 } 8638 8639 #ifdef CONFIG_NUMA_BALANCING 8640 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs) 8641 { 8642 if (sgs->sum_h_nr_running > sgs->nr_numa_running) 8643 return regular; 8644 if (sgs->sum_h_nr_running > sgs->nr_preferred_running) 8645 return remote; 8646 return all; 8647 } 8648 8649 static inline enum fbq_type fbq_classify_rq(struct rq *rq) 8650 { 8651 if (rq->nr_running > rq->nr_numa_running) 8652 return regular; 8653 if (rq->nr_running > rq->nr_preferred_running) 8654 return remote; 8655 return all; 8656 } 8657 #else 8658 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs) 8659 { 8660 return all; 8661 } 8662 8663 static inline enum fbq_type fbq_classify_rq(struct rq *rq) 8664 { 8665 return regular; 8666 } 8667 #endif /* CONFIG_NUMA_BALANCING */ 8668 8669 8670 struct sg_lb_stats; 8671 8672 /* 8673 * task_running_on_cpu - return 1 if @p is running on @cpu. 8674 */ 8675 8676 static unsigned int task_running_on_cpu(int cpu, struct task_struct *p) 8677 { 8678 /* Task has no contribution or is new */ 8679 if (cpu != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time)) 8680 return 0; 8681 8682 if (task_on_rq_queued(p)) 8683 return 1; 8684 8685 return 0; 8686 } 8687 8688 /** 8689 * idle_cpu_without - would a given CPU be idle without p ? 8690 * @cpu: the processor on which idleness is tested. 8691 * @p: task which should be ignored. 8692 * 8693 * Return: 1 if the CPU would be idle. 0 otherwise. 8694 */ 8695 static int idle_cpu_without(int cpu, struct task_struct *p) 8696 { 8697 struct rq *rq = cpu_rq(cpu); 8698 8699 if (rq->curr != rq->idle && rq->curr != p) 8700 return 0; 8701 8702 /* 8703 * rq->nr_running can't be used but an updated version without the 8704 * impact of p on cpu must be used instead. The updated nr_running 8705 * be computed and tested before calling idle_cpu_without(). 8706 */ 8707 8708 #ifdef CONFIG_SMP 8709 if (rq->ttwu_pending) 8710 return 0; 8711 #endif 8712 8713 return 1; 8714 } 8715 8716 /* 8717 * update_sg_wakeup_stats - Update sched_group's statistics for wakeup. 8718 * @sd: The sched_domain level to look for idlest group. 8719 * @group: sched_group whose statistics are to be updated. 8720 * @sgs: variable to hold the statistics for this group. 8721 * @p: The task for which we look for the idlest group/CPU. 8722 */ 8723 static inline void update_sg_wakeup_stats(struct sched_domain *sd, 8724 struct sched_group *group, 8725 struct sg_lb_stats *sgs, 8726 struct task_struct *p) 8727 { 8728 int i, nr_running; 8729 8730 memset(sgs, 0, sizeof(*sgs)); 8731 8732 for_each_cpu(i, sched_group_span(group)) { 8733 struct rq *rq = cpu_rq(i); 8734 unsigned int local; 8735 8736 sgs->group_load += cpu_load_without(rq, p); 8737 sgs->group_util += cpu_util_without(i, p); 8738 sgs->group_runnable += cpu_runnable_without(rq, p); 8739 local = task_running_on_cpu(i, p); 8740 sgs->sum_h_nr_running += rq->cfs.h_nr_running - local; 8741 8742 nr_running = rq->nr_running - local; 8743 sgs->sum_nr_running += nr_running; 8744 8745 /* 8746 * No need to call idle_cpu_without() if nr_running is not 0 8747 */ 8748 if (!nr_running && idle_cpu_without(i, p)) 8749 sgs->idle_cpus++; 8750 8751 } 8752 8753 /* Check if task fits in the group */ 8754 if (sd->flags & SD_ASYM_CPUCAPACITY && 8755 !task_fits_capacity(p, group->sgc->max_capacity)) { 8756 sgs->group_misfit_task_load = 1; 8757 } 8758 8759 sgs->group_capacity = group->sgc->capacity; 8760 8761 sgs->group_weight = group->group_weight; 8762 8763 sgs->group_type = group_classify(sd->imbalance_pct, group, sgs); 8764 8765 /* 8766 * Computing avg_load makes sense only when group is fully busy or 8767 * overloaded 8768 */ 8769 if (sgs->group_type == group_fully_busy || 8770 sgs->group_type == group_overloaded) 8771 sgs->avg_load = (sgs->group_load * SCHED_CAPACITY_SCALE) / 8772 sgs->group_capacity; 8773 } 8774 8775 static bool update_pick_idlest(struct sched_group *idlest, 8776 struct sg_lb_stats *idlest_sgs, 8777 struct sched_group *group, 8778 struct sg_lb_stats *sgs) 8779 { 8780 if (sgs->group_type < idlest_sgs->group_type) 8781 return true; 8782 8783 if (sgs->group_type > idlest_sgs->group_type) 8784 return false; 8785 8786 /* 8787 * The candidate and the current idlest group are the same type of 8788 * group. Let check which one is the idlest according to the type. 8789 */ 8790 8791 switch (sgs->group_type) { 8792 case group_overloaded: 8793 case group_fully_busy: 8794 /* Select the group with lowest avg_load. */ 8795 if (idlest_sgs->avg_load <= sgs->avg_load) 8796 return false; 8797 break; 8798 8799 case group_imbalanced: 8800 case group_asym_packing: 8801 /* Those types are not used in the slow wakeup path */ 8802 return false; 8803 8804 case group_misfit_task: 8805 /* Select group with the highest max capacity */ 8806 if (idlest->sgc->max_capacity >= group->sgc->max_capacity) 8807 return false; 8808 break; 8809 8810 case group_has_spare: 8811 /* Select group with most idle CPUs */ 8812 if (idlest_sgs->idle_cpus > sgs->idle_cpus) 8813 return false; 8814 8815 /* Select group with lowest group_util */ 8816 if (idlest_sgs->idle_cpus == sgs->idle_cpus && 8817 idlest_sgs->group_util <= sgs->group_util) 8818 return false; 8819 8820 break; 8821 } 8822 8823 return true; 8824 } 8825 8826 /* 8827 * Allow a NUMA imbalance if busy CPUs is less than 25% of the domain. 8828 * This is an approximation as the number of running tasks may not be 8829 * related to the number of busy CPUs due to sched_setaffinity. 8830 */ 8831 static inline bool allow_numa_imbalance(int dst_running, int dst_weight) 8832 { 8833 return (dst_running < (dst_weight >> 2)); 8834 } 8835 8836 /* 8837 * find_idlest_group() finds and returns the least busy CPU group within the 8838 * domain. 8839 * 8840 * Assumes p is allowed on at least one CPU in sd. 8841 */ 8842 static struct sched_group * 8843 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu) 8844 { 8845 struct sched_group *idlest = NULL, *local = NULL, *group = sd->groups; 8846 struct sg_lb_stats local_sgs, tmp_sgs; 8847 struct sg_lb_stats *sgs; 8848 unsigned long imbalance; 8849 struct sg_lb_stats idlest_sgs = { 8850 .avg_load = UINT_MAX, 8851 .group_type = group_overloaded, 8852 }; 8853 8854 do { 8855 int local_group; 8856 8857 /* Skip over this group if it has no CPUs allowed */ 8858 if (!cpumask_intersects(sched_group_span(group), 8859 p->cpus_ptr)) 8860 continue; 8861 8862 local_group = cpumask_test_cpu(this_cpu, 8863 sched_group_span(group)); 8864 8865 if (local_group) { 8866 sgs = &local_sgs; 8867 local = group; 8868 } else { 8869 sgs = &tmp_sgs; 8870 } 8871 8872 update_sg_wakeup_stats(sd, group, sgs, p); 8873 8874 if (!local_group && update_pick_idlest(idlest, &idlest_sgs, group, sgs)) { 8875 idlest = group; 8876 idlest_sgs = *sgs; 8877 } 8878 8879 } while (group = group->next, group != sd->groups); 8880 8881 8882 /* There is no idlest group to push tasks to */ 8883 if (!idlest) 8884 return NULL; 8885 8886 /* The local group has been skipped because of CPU affinity */ 8887 if (!local) 8888 return idlest; 8889 8890 /* 8891 * If the local group is idler than the selected idlest group 8892 * don't try and push the task. 8893 */ 8894 if (local_sgs.group_type < idlest_sgs.group_type) 8895 return NULL; 8896 8897 /* 8898 * If the local group is busier than the selected idlest group 8899 * try and push the task. 8900 */ 8901 if (local_sgs.group_type > idlest_sgs.group_type) 8902 return idlest; 8903 8904 switch (local_sgs.group_type) { 8905 case group_overloaded: 8906 case group_fully_busy: 8907 8908 /* Calculate allowed imbalance based on load */ 8909 imbalance = scale_load_down(NICE_0_LOAD) * 8910 (sd->imbalance_pct-100) / 100; 8911 8912 /* 8913 * When comparing groups across NUMA domains, it's possible for 8914 * the local domain to be very lightly loaded relative to the 8915 * remote domains but "imbalance" skews the comparison making 8916 * remote CPUs look much more favourable. When considering 8917 * cross-domain, add imbalance to the load on the remote node 8918 * and consider staying local. 8919 */ 8920 8921 if ((sd->flags & SD_NUMA) && 8922 ((idlest_sgs.avg_load + imbalance) >= local_sgs.avg_load)) 8923 return NULL; 8924 8925 /* 8926 * If the local group is less loaded than the selected 8927 * idlest group don't try and push any tasks. 8928 */ 8929 if (idlest_sgs.avg_load >= (local_sgs.avg_load + imbalance)) 8930 return NULL; 8931 8932 if (100 * local_sgs.avg_load <= sd->imbalance_pct * idlest_sgs.avg_load) 8933 return NULL; 8934 break; 8935 8936 case group_imbalanced: 8937 case group_asym_packing: 8938 /* Those type are not used in the slow wakeup path */ 8939 return NULL; 8940 8941 case group_misfit_task: 8942 /* Select group with the highest max capacity */ 8943 if (local->sgc->max_capacity >= idlest->sgc->max_capacity) 8944 return NULL; 8945 break; 8946 8947 case group_has_spare: 8948 if (sd->flags & SD_NUMA) { 8949 #ifdef CONFIG_NUMA_BALANCING 8950 int idlest_cpu; 8951 /* 8952 * If there is spare capacity at NUMA, try to select 8953 * the preferred node 8954 */ 8955 if (cpu_to_node(this_cpu) == p->numa_preferred_nid) 8956 return NULL; 8957 8958 idlest_cpu = cpumask_first(sched_group_span(idlest)); 8959 if (cpu_to_node(idlest_cpu) == p->numa_preferred_nid) 8960 return idlest; 8961 #endif 8962 /* 8963 * Otherwise, keep the task on this node to stay close 8964 * its wakeup source and improve locality. If there is 8965 * a real need of migration, periodic load balance will 8966 * take care of it. 8967 */ 8968 if (allow_numa_imbalance(local_sgs.sum_nr_running, sd->span_weight)) 8969 return NULL; 8970 } 8971 8972 /* 8973 * Select group with highest number of idle CPUs. We could also 8974 * compare the utilization which is more stable but it can end 8975 * up that the group has less spare capacity but finally more 8976 * idle CPUs which means more opportunity to run task. 8977 */ 8978 if (local_sgs.idle_cpus >= idlest_sgs.idle_cpus) 8979 return NULL; 8980 break; 8981 } 8982 8983 return idlest; 8984 } 8985 8986 /** 8987 * update_sd_lb_stats - Update sched_domain's statistics for load balancing. 8988 * @env: The load balancing environment. 8989 * @sds: variable to hold the statistics for this sched_domain. 8990 */ 8991 8992 static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds) 8993 { 8994 struct sched_domain *child = env->sd->child; 8995 struct sched_group *sg = env->sd->groups; 8996 struct sg_lb_stats *local = &sds->local_stat; 8997 struct sg_lb_stats tmp_sgs; 8998 int sg_status = 0; 8999 9000 do { 9001 struct sg_lb_stats *sgs = &tmp_sgs; 9002 int local_group; 9003 9004 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_span(sg)); 9005 if (local_group) { 9006 sds->local = sg; 9007 sgs = local; 9008 9009 if (env->idle != CPU_NEWLY_IDLE || 9010 time_after_eq(jiffies, sg->sgc->next_update)) 9011 update_group_capacity(env->sd, env->dst_cpu); 9012 } 9013 9014 update_sg_lb_stats(env, sg, sgs, &sg_status); 9015 9016 if (local_group) 9017 goto next_group; 9018 9019 9020 if (update_sd_pick_busiest(env, sds, sg, sgs)) { 9021 sds->busiest = sg; 9022 sds->busiest_stat = *sgs; 9023 } 9024 9025 next_group: 9026 /* Now, start updating sd_lb_stats */ 9027 sds->total_load += sgs->group_load; 9028 sds->total_capacity += sgs->group_capacity; 9029 9030 sg = sg->next; 9031 } while (sg != env->sd->groups); 9032 9033 /* Tag domain that child domain prefers tasks go to siblings first */ 9034 sds->prefer_sibling = child && child->flags & SD_PREFER_SIBLING; 9035 9036 9037 if (env->sd->flags & SD_NUMA) 9038 env->fbq_type = fbq_classify_group(&sds->busiest_stat); 9039 9040 if (!env->sd->parent) { 9041 struct root_domain *rd = env->dst_rq->rd; 9042 9043 /* update overload indicator if we are at root domain */ 9044 WRITE_ONCE(rd->overload, sg_status & SG_OVERLOAD); 9045 9046 /* Update over-utilization (tipping point, U >= 0) indicator */ 9047 WRITE_ONCE(rd->overutilized, sg_status & SG_OVERUTILIZED); 9048 trace_sched_overutilized_tp(rd, sg_status & SG_OVERUTILIZED); 9049 } else if (sg_status & SG_OVERUTILIZED) { 9050 struct root_domain *rd = env->dst_rq->rd; 9051 9052 WRITE_ONCE(rd->overutilized, SG_OVERUTILIZED); 9053 trace_sched_overutilized_tp(rd, SG_OVERUTILIZED); 9054 } 9055 } 9056 9057 #define NUMA_IMBALANCE_MIN 2 9058 9059 static inline long adjust_numa_imbalance(int imbalance, 9060 int dst_running, int dst_weight) 9061 { 9062 if (!allow_numa_imbalance(dst_running, dst_weight)) 9063 return imbalance; 9064 9065 /* 9066 * Allow a small imbalance based on a simple pair of communicating 9067 * tasks that remain local when the destination is lightly loaded. 9068 */ 9069 if (imbalance <= NUMA_IMBALANCE_MIN) 9070 return 0; 9071 9072 return imbalance; 9073 } 9074 9075 /** 9076 * calculate_imbalance - Calculate the amount of imbalance present within the 9077 * groups of a given sched_domain during load balance. 9078 * @env: load balance environment 9079 * @sds: statistics of the sched_domain whose imbalance is to be calculated. 9080 */ 9081 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds) 9082 { 9083 struct sg_lb_stats *local, *busiest; 9084 9085 local = &sds->local_stat; 9086 busiest = &sds->busiest_stat; 9087 9088 if (busiest->group_type == group_misfit_task) { 9089 /* Set imbalance to allow misfit tasks to be balanced. */ 9090 env->migration_type = migrate_misfit; 9091 env->imbalance = 1; 9092 return; 9093 } 9094 9095 if (busiest->group_type == group_asym_packing) { 9096 /* 9097 * In case of asym capacity, we will try to migrate all load to 9098 * the preferred CPU. 9099 */ 9100 env->migration_type = migrate_task; 9101 env->imbalance = busiest->sum_h_nr_running; 9102 return; 9103 } 9104 9105 if (busiest->group_type == group_imbalanced) { 9106 /* 9107 * In the group_imb case we cannot rely on group-wide averages 9108 * to ensure CPU-load equilibrium, try to move any task to fix 9109 * the imbalance. The next load balance will take care of 9110 * balancing back the system. 9111 */ 9112 env->migration_type = migrate_task; 9113 env->imbalance = 1; 9114 return; 9115 } 9116 9117 /* 9118 * Try to use spare capacity of local group without overloading it or 9119 * emptying busiest. 9120 */ 9121 if (local->group_type == group_has_spare) { 9122 if ((busiest->group_type > group_fully_busy) && 9123 !(env->sd->flags & SD_SHARE_PKG_RESOURCES)) { 9124 /* 9125 * If busiest is overloaded, try to fill spare 9126 * capacity. This might end up creating spare capacity 9127 * in busiest or busiest still being overloaded but 9128 * there is no simple way to directly compute the 9129 * amount of load to migrate in order to balance the 9130 * system. 9131 */ 9132 env->migration_type = migrate_util; 9133 env->imbalance = max(local->group_capacity, local->group_util) - 9134 local->group_util; 9135 9136 /* 9137 * In some cases, the group's utilization is max or even 9138 * higher than capacity because of migrations but the 9139 * local CPU is (newly) idle. There is at least one 9140 * waiting task in this overloaded busiest group. Let's 9141 * try to pull it. 9142 */ 9143 if (env->idle != CPU_NOT_IDLE && env->imbalance == 0) { 9144 env->migration_type = migrate_task; 9145 env->imbalance = 1; 9146 } 9147 9148 return; 9149 } 9150 9151 if (busiest->group_weight == 1 || sds->prefer_sibling) { 9152 unsigned int nr_diff = busiest->sum_nr_running; 9153 /* 9154 * When prefer sibling, evenly spread running tasks on 9155 * groups. 9156 */ 9157 env->migration_type = migrate_task; 9158 lsub_positive(&nr_diff, local->sum_nr_running); 9159 env->imbalance = nr_diff >> 1; 9160 } else { 9161 9162 /* 9163 * If there is no overload, we just want to even the number of 9164 * idle cpus. 9165 */ 9166 env->migration_type = migrate_task; 9167 env->imbalance = max_t(long, 0, (local->idle_cpus - 9168 busiest->idle_cpus) >> 1); 9169 } 9170 9171 /* Consider allowing a small imbalance between NUMA groups */ 9172 if (env->sd->flags & SD_NUMA) { 9173 env->imbalance = adjust_numa_imbalance(env->imbalance, 9174 busiest->sum_nr_running, busiest->group_weight); 9175 } 9176 9177 return; 9178 } 9179 9180 /* 9181 * Local is fully busy but has to take more load to relieve the 9182 * busiest group 9183 */ 9184 if (local->group_type < group_overloaded) { 9185 /* 9186 * Local will become overloaded so the avg_load metrics are 9187 * finally needed. 9188 */ 9189 9190 local->avg_load = (local->group_load * SCHED_CAPACITY_SCALE) / 9191 local->group_capacity; 9192 9193 sds->avg_load = (sds->total_load * SCHED_CAPACITY_SCALE) / 9194 sds->total_capacity; 9195 /* 9196 * If the local group is more loaded than the selected 9197 * busiest group don't try to pull any tasks. 9198 */ 9199 if (local->avg_load >= busiest->avg_load) { 9200 env->imbalance = 0; 9201 return; 9202 } 9203 } 9204 9205 /* 9206 * Both group are or will become overloaded and we're trying to get all 9207 * the CPUs to the average_load, so we don't want to push ourselves 9208 * above the average load, nor do we wish to reduce the max loaded CPU 9209 * below the average load. At the same time, we also don't want to 9210 * reduce the group load below the group capacity. Thus we look for 9211 * the minimum possible imbalance. 9212 */ 9213 env->migration_type = migrate_load; 9214 env->imbalance = min( 9215 (busiest->avg_load - sds->avg_load) * busiest->group_capacity, 9216 (sds->avg_load - local->avg_load) * local->group_capacity 9217 ) / SCHED_CAPACITY_SCALE; 9218 } 9219 9220 /******* find_busiest_group() helpers end here *********************/ 9221 9222 /* 9223 * Decision matrix according to the local and busiest group type: 9224 * 9225 * busiest \ local has_spare fully_busy misfit asym imbalanced overloaded 9226 * has_spare nr_idle balanced N/A N/A balanced balanced 9227 * fully_busy nr_idle nr_idle N/A N/A balanced balanced 9228 * misfit_task force N/A N/A N/A force force 9229 * asym_packing force force N/A N/A force force 9230 * imbalanced force force N/A N/A force force 9231 * overloaded force force N/A N/A force avg_load 9232 * 9233 * N/A : Not Applicable because already filtered while updating 9234 * statistics. 9235 * balanced : The system is balanced for these 2 groups. 9236 * force : Calculate the imbalance as load migration is probably needed. 9237 * avg_load : Only if imbalance is significant enough. 9238 * nr_idle : dst_cpu is not busy and the number of idle CPUs is quite 9239 * different in groups. 9240 */ 9241 9242 /** 9243 * find_busiest_group - Returns the busiest group within the sched_domain 9244 * if there is an imbalance. 9245 * 9246 * Also calculates the amount of runnable load which should be moved 9247 * to restore balance. 9248 * 9249 * @env: The load balancing environment. 9250 * 9251 * Return: - The busiest group if imbalance exists. 9252 */ 9253 static struct sched_group *find_busiest_group(struct lb_env *env) 9254 { 9255 struct sg_lb_stats *local, *busiest; 9256 struct sd_lb_stats sds; 9257 9258 init_sd_lb_stats(&sds); 9259 9260 /* 9261 * Compute the various statistics relevant for load balancing at 9262 * this level. 9263 */ 9264 update_sd_lb_stats(env, &sds); 9265 9266 if (sched_energy_enabled()) { 9267 struct root_domain *rd = env->dst_rq->rd; 9268 9269 if (rcu_dereference(rd->pd) && !READ_ONCE(rd->overutilized)) 9270 goto out_balanced; 9271 } 9272 9273 local = &sds.local_stat; 9274 busiest = &sds.busiest_stat; 9275 9276 /* There is no busy sibling group to pull tasks from */ 9277 if (!sds.busiest) 9278 goto out_balanced; 9279 9280 /* Misfit tasks should be dealt with regardless of the avg load */ 9281 if (busiest->group_type == group_misfit_task) 9282 goto force_balance; 9283 9284 /* ASYM feature bypasses nice load balance check */ 9285 if (busiest->group_type == group_asym_packing) 9286 goto force_balance; 9287 9288 /* 9289 * If the busiest group is imbalanced the below checks don't 9290 * work because they assume all things are equal, which typically 9291 * isn't true due to cpus_ptr constraints and the like. 9292 */ 9293 if (busiest->group_type == group_imbalanced) 9294 goto force_balance; 9295 9296 /* 9297 * If the local group is busier than the selected busiest group 9298 * don't try and pull any tasks. 9299 */ 9300 if (local->group_type > busiest->group_type) 9301 goto out_balanced; 9302 9303 /* 9304 * When groups are overloaded, use the avg_load to ensure fairness 9305 * between tasks. 9306 */ 9307 if (local->group_type == group_overloaded) { 9308 /* 9309 * If the local group is more loaded than the selected 9310 * busiest group don't try to pull any tasks. 9311 */ 9312 if (local->avg_load >= busiest->avg_load) 9313 goto out_balanced; 9314 9315 /* XXX broken for overlapping NUMA groups */ 9316 sds.avg_load = (sds.total_load * SCHED_CAPACITY_SCALE) / 9317 sds.total_capacity; 9318 9319 /* 9320 * Don't pull any tasks if this group is already above the 9321 * domain average load. 9322 */ 9323 if (local->avg_load >= sds.avg_load) 9324 goto out_balanced; 9325 9326 /* 9327 * If the busiest group is more loaded, use imbalance_pct to be 9328 * conservative. 9329 */ 9330 if (100 * busiest->avg_load <= 9331 env->sd->imbalance_pct * local->avg_load) 9332 goto out_balanced; 9333 } 9334 9335 /* Try to move all excess tasks to child's sibling domain */ 9336 if (sds.prefer_sibling && local->group_type == group_has_spare && 9337 busiest->sum_nr_running > local->sum_nr_running + 1) 9338 goto force_balance; 9339 9340 if (busiest->group_type != group_overloaded) { 9341 if (env->idle == CPU_NOT_IDLE) 9342 /* 9343 * If the busiest group is not overloaded (and as a 9344 * result the local one too) but this CPU is already 9345 * busy, let another idle CPU try to pull task. 9346 */ 9347 goto out_balanced; 9348 9349 if (busiest->group_weight > 1 && 9350 local->idle_cpus <= (busiest->idle_cpus + 1)) 9351 /* 9352 * If the busiest group is not overloaded 9353 * and there is no imbalance between this and busiest 9354 * group wrt idle CPUs, it is balanced. The imbalance 9355 * becomes significant if the diff is greater than 1 9356 * otherwise we might end up to just move the imbalance 9357 * on another group. Of course this applies only if 9358 * there is more than 1 CPU per group. 9359 */ 9360 goto out_balanced; 9361 9362 if (busiest->sum_h_nr_running == 1) 9363 /* 9364 * busiest doesn't have any tasks waiting to run 9365 */ 9366 goto out_balanced; 9367 } 9368 9369 force_balance: 9370 /* Looks like there is an imbalance. Compute it */ 9371 calculate_imbalance(env, &sds); 9372 return env->imbalance ? sds.busiest : NULL; 9373 9374 out_balanced: 9375 env->imbalance = 0; 9376 return NULL; 9377 } 9378 9379 /* 9380 * find_busiest_queue - find the busiest runqueue among the CPUs in the group. 9381 */ 9382 static struct rq *find_busiest_queue(struct lb_env *env, 9383 struct sched_group *group) 9384 { 9385 struct rq *busiest = NULL, *rq; 9386 unsigned long busiest_util = 0, busiest_load = 0, busiest_capacity = 1; 9387 unsigned int busiest_nr = 0; 9388 int i; 9389 9390 for_each_cpu_and(i, sched_group_span(group), env->cpus) { 9391 unsigned long capacity, load, util; 9392 unsigned int nr_running; 9393 enum fbq_type rt; 9394 9395 rq = cpu_rq(i); 9396 rt = fbq_classify_rq(rq); 9397 9398 /* 9399 * We classify groups/runqueues into three groups: 9400 * - regular: there are !numa tasks 9401 * - remote: there are numa tasks that run on the 'wrong' node 9402 * - all: there is no distinction 9403 * 9404 * In order to avoid migrating ideally placed numa tasks, 9405 * ignore those when there's better options. 9406 * 9407 * If we ignore the actual busiest queue to migrate another 9408 * task, the next balance pass can still reduce the busiest 9409 * queue by moving tasks around inside the node. 9410 * 9411 * If we cannot move enough load due to this classification 9412 * the next pass will adjust the group classification and 9413 * allow migration of more tasks. 9414 * 9415 * Both cases only affect the total convergence complexity. 9416 */ 9417 if (rt > env->fbq_type) 9418 continue; 9419 9420 nr_running = rq->cfs.h_nr_running; 9421 if (!nr_running) 9422 continue; 9423 9424 capacity = capacity_of(i); 9425 9426 /* 9427 * For ASYM_CPUCAPACITY domains, don't pick a CPU that could 9428 * eventually lead to active_balancing high->low capacity. 9429 * Higher per-CPU capacity is considered better than balancing 9430 * average load. 9431 */ 9432 if (env->sd->flags & SD_ASYM_CPUCAPACITY && 9433 !capacity_greater(capacity_of(env->dst_cpu), capacity) && 9434 nr_running == 1) 9435 continue; 9436 9437 switch (env->migration_type) { 9438 case migrate_load: 9439 /* 9440 * When comparing with load imbalance, use cpu_load() 9441 * which is not scaled with the CPU capacity. 9442 */ 9443 load = cpu_load(rq); 9444 9445 if (nr_running == 1 && load > env->imbalance && 9446 !check_cpu_capacity(rq, env->sd)) 9447 break; 9448 9449 /* 9450 * For the load comparisons with the other CPUs, 9451 * consider the cpu_load() scaled with the CPU 9452 * capacity, so that the load can be moved away 9453 * from the CPU that is potentially running at a 9454 * lower capacity. 9455 * 9456 * Thus we're looking for max(load_i / capacity_i), 9457 * crosswise multiplication to rid ourselves of the 9458 * division works out to: 9459 * load_i * capacity_j > load_j * capacity_i; 9460 * where j is our previous maximum. 9461 */ 9462 if (load * busiest_capacity > busiest_load * capacity) { 9463 busiest_load = load; 9464 busiest_capacity = capacity; 9465 busiest = rq; 9466 } 9467 break; 9468 9469 case migrate_util: 9470 util = cpu_util(cpu_of(rq)); 9471 9472 /* 9473 * Don't try to pull utilization from a CPU with one 9474 * running task. Whatever its utilization, we will fail 9475 * detach the task. 9476 */ 9477 if (nr_running <= 1) 9478 continue; 9479 9480 if (busiest_util < util) { 9481 busiest_util = util; 9482 busiest = rq; 9483 } 9484 break; 9485 9486 case migrate_task: 9487 if (busiest_nr < nr_running) { 9488 busiest_nr = nr_running; 9489 busiest = rq; 9490 } 9491 break; 9492 9493 case migrate_misfit: 9494 /* 9495 * For ASYM_CPUCAPACITY domains with misfit tasks we 9496 * simply seek the "biggest" misfit task. 9497 */ 9498 if (rq->misfit_task_load > busiest_load) { 9499 busiest_load = rq->misfit_task_load; 9500 busiest = rq; 9501 } 9502 9503 break; 9504 9505 } 9506 } 9507 9508 return busiest; 9509 } 9510 9511 /* 9512 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but 9513 * so long as it is large enough. 9514 */ 9515 #define MAX_PINNED_INTERVAL 512 9516 9517 static inline bool 9518 asym_active_balance(struct lb_env *env) 9519 { 9520 /* 9521 * ASYM_PACKING needs to force migrate tasks from busy but 9522 * lower priority CPUs in order to pack all tasks in the 9523 * highest priority CPUs. 9524 */ 9525 return env->idle != CPU_NOT_IDLE && (env->sd->flags & SD_ASYM_PACKING) && 9526 sched_asym_prefer(env->dst_cpu, env->src_cpu); 9527 } 9528 9529 static inline bool 9530 imbalanced_active_balance(struct lb_env *env) 9531 { 9532 struct sched_domain *sd = env->sd; 9533 9534 /* 9535 * The imbalanced case includes the case of pinned tasks preventing a fair 9536 * distribution of the load on the system but also the even distribution of the 9537 * threads on a system with spare capacity 9538 */ 9539 if ((env->migration_type == migrate_task) && 9540 (sd->nr_balance_failed > sd->cache_nice_tries+2)) 9541 return 1; 9542 9543 return 0; 9544 } 9545 9546 static int need_active_balance(struct lb_env *env) 9547 { 9548 struct sched_domain *sd = env->sd; 9549 9550 if (asym_active_balance(env)) 9551 return 1; 9552 9553 if (imbalanced_active_balance(env)) 9554 return 1; 9555 9556 /* 9557 * The dst_cpu is idle and the src_cpu CPU has only 1 CFS task. 9558 * It's worth migrating the task if the src_cpu's capacity is reduced 9559 * because of other sched_class or IRQs if more capacity stays 9560 * available on dst_cpu. 9561 */ 9562 if ((env->idle != CPU_NOT_IDLE) && 9563 (env->src_rq->cfs.h_nr_running == 1)) { 9564 if ((check_cpu_capacity(env->src_rq, sd)) && 9565 (capacity_of(env->src_cpu)*sd->imbalance_pct < capacity_of(env->dst_cpu)*100)) 9566 return 1; 9567 } 9568 9569 if (env->migration_type == migrate_misfit) 9570 return 1; 9571 9572 return 0; 9573 } 9574 9575 static int active_load_balance_cpu_stop(void *data); 9576 9577 static int should_we_balance(struct lb_env *env) 9578 { 9579 struct sched_group *sg = env->sd->groups; 9580 int cpu; 9581 9582 /* 9583 * Ensure the balancing environment is consistent; can happen 9584 * when the softirq triggers 'during' hotplug. 9585 */ 9586 if (!cpumask_test_cpu(env->dst_cpu, env->cpus)) 9587 return 0; 9588 9589 /* 9590 * In the newly idle case, we will allow all the CPUs 9591 * to do the newly idle load balance. 9592 */ 9593 if (env->idle == CPU_NEWLY_IDLE) 9594 return 1; 9595 9596 /* Try to find first idle CPU */ 9597 for_each_cpu_and(cpu, group_balance_mask(sg), env->cpus) { 9598 if (!idle_cpu(cpu)) 9599 continue; 9600 9601 /* Are we the first idle CPU? */ 9602 return cpu == env->dst_cpu; 9603 } 9604 9605 /* Are we the first CPU of this group ? */ 9606 return group_balance_cpu(sg) == env->dst_cpu; 9607 } 9608 9609 /* 9610 * Check this_cpu to ensure it is balanced within domain. Attempt to move 9611 * tasks if there is an imbalance. 9612 */ 9613 static int load_balance(int this_cpu, struct rq *this_rq, 9614 struct sched_domain *sd, enum cpu_idle_type idle, 9615 int *continue_balancing) 9616 { 9617 int ld_moved, cur_ld_moved, active_balance = 0; 9618 struct sched_domain *sd_parent = sd->parent; 9619 struct sched_group *group; 9620 struct rq *busiest; 9621 struct rq_flags rf; 9622 struct cpumask *cpus = this_cpu_cpumask_var_ptr(load_balance_mask); 9623 9624 struct lb_env env = { 9625 .sd = sd, 9626 .dst_cpu = this_cpu, 9627 .dst_rq = this_rq, 9628 .dst_grpmask = sched_group_span(sd->groups), 9629 .idle = idle, 9630 .loop_break = sched_nr_migrate_break, 9631 .cpus = cpus, 9632 .fbq_type = all, 9633 .tasks = LIST_HEAD_INIT(env.tasks), 9634 }; 9635 9636 cpumask_and(cpus, sched_domain_span(sd), cpu_active_mask); 9637 9638 schedstat_inc(sd->lb_count[idle]); 9639 9640 redo: 9641 if (!should_we_balance(&env)) { 9642 *continue_balancing = 0; 9643 goto out_balanced; 9644 } 9645 9646 group = find_busiest_group(&env); 9647 if (!group) { 9648 schedstat_inc(sd->lb_nobusyg[idle]); 9649 goto out_balanced; 9650 } 9651 9652 busiest = find_busiest_queue(&env, group); 9653 if (!busiest) { 9654 schedstat_inc(sd->lb_nobusyq[idle]); 9655 goto out_balanced; 9656 } 9657 9658 BUG_ON(busiest == env.dst_rq); 9659 9660 schedstat_add(sd->lb_imbalance[idle], env.imbalance); 9661 9662 env.src_cpu = busiest->cpu; 9663 env.src_rq = busiest; 9664 9665 ld_moved = 0; 9666 /* Clear this flag as soon as we find a pullable task */ 9667 env.flags |= LBF_ALL_PINNED; 9668 if (busiest->nr_running > 1) { 9669 /* 9670 * Attempt to move tasks. If find_busiest_group has found 9671 * an imbalance but busiest->nr_running <= 1, the group is 9672 * still unbalanced. ld_moved simply stays zero, so it is 9673 * correctly treated as an imbalance. 9674 */ 9675 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running); 9676 9677 more_balance: 9678 rq_lock_irqsave(busiest, &rf); 9679 update_rq_clock(busiest); 9680 9681 /* 9682 * cur_ld_moved - load moved in current iteration 9683 * ld_moved - cumulative load moved across iterations 9684 */ 9685 cur_ld_moved = detach_tasks(&env); 9686 9687 /* 9688 * We've detached some tasks from busiest_rq. Every 9689 * task is masked "TASK_ON_RQ_MIGRATING", so we can safely 9690 * unlock busiest->lock, and we are able to be sure 9691 * that nobody can manipulate the tasks in parallel. 9692 * See task_rq_lock() family for the details. 9693 */ 9694 9695 rq_unlock(busiest, &rf); 9696 9697 if (cur_ld_moved) { 9698 attach_tasks(&env); 9699 ld_moved += cur_ld_moved; 9700 } 9701 9702 local_irq_restore(rf.flags); 9703 9704 if (env.flags & LBF_NEED_BREAK) { 9705 env.flags &= ~LBF_NEED_BREAK; 9706 goto more_balance; 9707 } 9708 9709 /* 9710 * Revisit (affine) tasks on src_cpu that couldn't be moved to 9711 * us and move them to an alternate dst_cpu in our sched_group 9712 * where they can run. The upper limit on how many times we 9713 * iterate on same src_cpu is dependent on number of CPUs in our 9714 * sched_group. 9715 * 9716 * This changes load balance semantics a bit on who can move 9717 * load to a given_cpu. In addition to the given_cpu itself 9718 * (or a ilb_cpu acting on its behalf where given_cpu is 9719 * nohz-idle), we now have balance_cpu in a position to move 9720 * load to given_cpu. In rare situations, this may cause 9721 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding 9722 * _independently_ and at _same_ time to move some load to 9723 * given_cpu) causing excess load to be moved to given_cpu. 9724 * This however should not happen so much in practice and 9725 * moreover subsequent load balance cycles should correct the 9726 * excess load moved. 9727 */ 9728 if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) { 9729 9730 /* Prevent to re-select dst_cpu via env's CPUs */ 9731 __cpumask_clear_cpu(env.dst_cpu, env.cpus); 9732 9733 env.dst_rq = cpu_rq(env.new_dst_cpu); 9734 env.dst_cpu = env.new_dst_cpu; 9735 env.flags &= ~LBF_DST_PINNED; 9736 env.loop = 0; 9737 env.loop_break = sched_nr_migrate_break; 9738 9739 /* 9740 * Go back to "more_balance" rather than "redo" since we 9741 * need to continue with same src_cpu. 9742 */ 9743 goto more_balance; 9744 } 9745 9746 /* 9747 * We failed to reach balance because of affinity. 9748 */ 9749 if (sd_parent) { 9750 int *group_imbalance = &sd_parent->groups->sgc->imbalance; 9751 9752 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0) 9753 *group_imbalance = 1; 9754 } 9755 9756 /* All tasks on this runqueue were pinned by CPU affinity */ 9757 if (unlikely(env.flags & LBF_ALL_PINNED)) { 9758 __cpumask_clear_cpu(cpu_of(busiest), cpus); 9759 /* 9760 * Attempting to continue load balancing at the current 9761 * sched_domain level only makes sense if there are 9762 * active CPUs remaining as possible busiest CPUs to 9763 * pull load from which are not contained within the 9764 * destination group that is receiving any migrated 9765 * load. 9766 */ 9767 if (!cpumask_subset(cpus, env.dst_grpmask)) { 9768 env.loop = 0; 9769 env.loop_break = sched_nr_migrate_break; 9770 goto redo; 9771 } 9772 goto out_all_pinned; 9773 } 9774 } 9775 9776 if (!ld_moved) { 9777 schedstat_inc(sd->lb_failed[idle]); 9778 /* 9779 * Increment the failure counter only on periodic balance. 9780 * We do not want newidle balance, which can be very 9781 * frequent, pollute the failure counter causing 9782 * excessive cache_hot migrations and active balances. 9783 */ 9784 if (idle != CPU_NEWLY_IDLE) 9785 sd->nr_balance_failed++; 9786 9787 if (need_active_balance(&env)) { 9788 unsigned long flags; 9789 9790 raw_spin_lock_irqsave(&busiest->lock, flags); 9791 9792 /* 9793 * Don't kick the active_load_balance_cpu_stop, 9794 * if the curr task on busiest CPU can't be 9795 * moved to this_cpu: 9796 */ 9797 if (!cpumask_test_cpu(this_cpu, busiest->curr->cpus_ptr)) { 9798 raw_spin_unlock_irqrestore(&busiest->lock, 9799 flags); 9800 goto out_one_pinned; 9801 } 9802 9803 /* Record that we found at least one task that could run on this_cpu */ 9804 env.flags &= ~LBF_ALL_PINNED; 9805 9806 /* 9807 * ->active_balance synchronizes accesses to 9808 * ->active_balance_work. Once set, it's cleared 9809 * only after active load balance is finished. 9810 */ 9811 if (!busiest->active_balance) { 9812 busiest->active_balance = 1; 9813 busiest->push_cpu = this_cpu; 9814 active_balance = 1; 9815 } 9816 raw_spin_unlock_irqrestore(&busiest->lock, flags); 9817 9818 if (active_balance) { 9819 stop_one_cpu_nowait(cpu_of(busiest), 9820 active_load_balance_cpu_stop, busiest, 9821 &busiest->active_balance_work); 9822 } 9823 } 9824 } else { 9825 sd->nr_balance_failed = 0; 9826 } 9827 9828 if (likely(!active_balance) || need_active_balance(&env)) { 9829 /* We were unbalanced, so reset the balancing interval */ 9830 sd->balance_interval = sd->min_interval; 9831 } 9832 9833 goto out; 9834 9835 out_balanced: 9836 /* 9837 * We reach balance although we may have faced some affinity 9838 * constraints. Clear the imbalance flag only if other tasks got 9839 * a chance to move and fix the imbalance. 9840 */ 9841 if (sd_parent && !(env.flags & LBF_ALL_PINNED)) { 9842 int *group_imbalance = &sd_parent->groups->sgc->imbalance; 9843 9844 if (*group_imbalance) 9845 *group_imbalance = 0; 9846 } 9847 9848 out_all_pinned: 9849 /* 9850 * We reach balance because all tasks are pinned at this level so 9851 * we can't migrate them. Let the imbalance flag set so parent level 9852 * can try to migrate them. 9853 */ 9854 schedstat_inc(sd->lb_balanced[idle]); 9855 9856 sd->nr_balance_failed = 0; 9857 9858 out_one_pinned: 9859 ld_moved = 0; 9860 9861 /* 9862 * newidle_balance() disregards balance intervals, so we could 9863 * repeatedly reach this code, which would lead to balance_interval 9864 * skyrocketing in a short amount of time. Skip the balance_interval 9865 * increase logic to avoid that. 9866 */ 9867 if (env.idle == CPU_NEWLY_IDLE) 9868 goto out; 9869 9870 /* tune up the balancing interval */ 9871 if ((env.flags & LBF_ALL_PINNED && 9872 sd->balance_interval < MAX_PINNED_INTERVAL) || 9873 sd->balance_interval < sd->max_interval) 9874 sd->balance_interval *= 2; 9875 out: 9876 return ld_moved; 9877 } 9878 9879 static inline unsigned long 9880 get_sd_balance_interval(struct sched_domain *sd, int cpu_busy) 9881 { 9882 unsigned long interval = sd->balance_interval; 9883 9884 if (cpu_busy) 9885 interval *= sd->busy_factor; 9886 9887 /* scale ms to jiffies */ 9888 interval = msecs_to_jiffies(interval); 9889 9890 /* 9891 * Reduce likelihood of busy balancing at higher domains racing with 9892 * balancing at lower domains by preventing their balancing periods 9893 * from being multiples of each other. 9894 */ 9895 if (cpu_busy) 9896 interval -= 1; 9897 9898 interval = clamp(interval, 1UL, max_load_balance_interval); 9899 9900 return interval; 9901 } 9902 9903 static inline void 9904 update_next_balance(struct sched_domain *sd, unsigned long *next_balance) 9905 { 9906 unsigned long interval, next; 9907 9908 /* used by idle balance, so cpu_busy = 0 */ 9909 interval = get_sd_balance_interval(sd, 0); 9910 next = sd->last_balance + interval; 9911 9912 if (time_after(*next_balance, next)) 9913 *next_balance = next; 9914 } 9915 9916 /* 9917 * active_load_balance_cpu_stop is run by the CPU stopper. It pushes 9918 * running tasks off the busiest CPU onto idle CPUs. It requires at 9919 * least 1 task to be running on each physical CPU where possible, and 9920 * avoids physical / logical imbalances. 9921 */ 9922 static int active_load_balance_cpu_stop(void *data) 9923 { 9924 struct rq *busiest_rq = data; 9925 int busiest_cpu = cpu_of(busiest_rq); 9926 int target_cpu = busiest_rq->push_cpu; 9927 struct rq *target_rq = cpu_rq(target_cpu); 9928 struct sched_domain *sd; 9929 struct task_struct *p = NULL; 9930 struct rq_flags rf; 9931 9932 rq_lock_irq(busiest_rq, &rf); 9933 /* 9934 * Between queueing the stop-work and running it is a hole in which 9935 * CPUs can become inactive. We should not move tasks from or to 9936 * inactive CPUs. 9937 */ 9938 if (!cpu_active(busiest_cpu) || !cpu_active(target_cpu)) 9939 goto out_unlock; 9940 9941 /* Make sure the requested CPU hasn't gone down in the meantime: */ 9942 if (unlikely(busiest_cpu != smp_processor_id() || 9943 !busiest_rq->active_balance)) 9944 goto out_unlock; 9945 9946 /* Is there any task to move? */ 9947 if (busiest_rq->nr_running <= 1) 9948 goto out_unlock; 9949 9950 /* 9951 * This condition is "impossible", if it occurs 9952 * we need to fix it. Originally reported by 9953 * Bjorn Helgaas on a 128-CPU setup. 9954 */ 9955 BUG_ON(busiest_rq == target_rq); 9956 9957 /* Search for an sd spanning us and the target CPU. */ 9958 rcu_read_lock(); 9959 for_each_domain(target_cpu, sd) { 9960 if (cpumask_test_cpu(busiest_cpu, sched_domain_span(sd))) 9961 break; 9962 } 9963 9964 if (likely(sd)) { 9965 struct lb_env env = { 9966 .sd = sd, 9967 .dst_cpu = target_cpu, 9968 .dst_rq = target_rq, 9969 .src_cpu = busiest_rq->cpu, 9970 .src_rq = busiest_rq, 9971 .idle = CPU_IDLE, 9972 .flags = LBF_ACTIVE_LB, 9973 }; 9974 9975 schedstat_inc(sd->alb_count); 9976 update_rq_clock(busiest_rq); 9977 9978 p = detach_one_task(&env); 9979 if (p) { 9980 schedstat_inc(sd->alb_pushed); 9981 /* Active balancing done, reset the failure counter. */ 9982 sd->nr_balance_failed = 0; 9983 } else { 9984 schedstat_inc(sd->alb_failed); 9985 } 9986 } 9987 rcu_read_unlock(); 9988 out_unlock: 9989 busiest_rq->active_balance = 0; 9990 rq_unlock(busiest_rq, &rf); 9991 9992 if (p) 9993 attach_one_task(target_rq, p); 9994 9995 local_irq_enable(); 9996 9997 return 0; 9998 } 9999 10000 static DEFINE_SPINLOCK(balancing); 10001 10002 /* 10003 * Scale the max load_balance interval with the number of CPUs in the system. 10004 * This trades load-balance latency on larger machines for less cross talk. 10005 */ 10006 void update_max_interval(void) 10007 { 10008 max_load_balance_interval = HZ*num_online_cpus()/10; 10009 } 10010 10011 /* 10012 * It checks each scheduling domain to see if it is due to be balanced, 10013 * and initiates a balancing operation if so. 10014 * 10015 * Balancing parameters are set up in init_sched_domains. 10016 */ 10017 static void rebalance_domains(struct rq *rq, enum cpu_idle_type idle) 10018 { 10019 int continue_balancing = 1; 10020 int cpu = rq->cpu; 10021 int busy = idle != CPU_IDLE && !sched_idle_cpu(cpu); 10022 unsigned long interval; 10023 struct sched_domain *sd; 10024 /* Earliest time when we have to do rebalance again */ 10025 unsigned long next_balance = jiffies + 60*HZ; 10026 int update_next_balance = 0; 10027 int need_serialize, need_decay = 0; 10028 u64 max_cost = 0; 10029 10030 rcu_read_lock(); 10031 for_each_domain(cpu, sd) { 10032 /* 10033 * Decay the newidle max times here because this is a regular 10034 * visit to all the domains. Decay ~1% per second. 10035 */ 10036 if (time_after(jiffies, sd->next_decay_max_lb_cost)) { 10037 sd->max_newidle_lb_cost = 10038 (sd->max_newidle_lb_cost * 253) / 256; 10039 sd->next_decay_max_lb_cost = jiffies + HZ; 10040 need_decay = 1; 10041 } 10042 max_cost += sd->max_newidle_lb_cost; 10043 10044 /* 10045 * Stop the load balance at this level. There is another 10046 * CPU in our sched group which is doing load balancing more 10047 * actively. 10048 */ 10049 if (!continue_balancing) { 10050 if (need_decay) 10051 continue; 10052 break; 10053 } 10054 10055 interval = get_sd_balance_interval(sd, busy); 10056 10057 need_serialize = sd->flags & SD_SERIALIZE; 10058 if (need_serialize) { 10059 if (!spin_trylock(&balancing)) 10060 goto out; 10061 } 10062 10063 if (time_after_eq(jiffies, sd->last_balance + interval)) { 10064 if (load_balance(cpu, rq, sd, idle, &continue_balancing)) { 10065 /* 10066 * The LBF_DST_PINNED logic could have changed 10067 * env->dst_cpu, so we can't know our idle 10068 * state even if we migrated tasks. Update it. 10069 */ 10070 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE; 10071 busy = idle != CPU_IDLE && !sched_idle_cpu(cpu); 10072 } 10073 sd->last_balance = jiffies; 10074 interval = get_sd_balance_interval(sd, busy); 10075 } 10076 if (need_serialize) 10077 spin_unlock(&balancing); 10078 out: 10079 if (time_after(next_balance, sd->last_balance + interval)) { 10080 next_balance = sd->last_balance + interval; 10081 update_next_balance = 1; 10082 } 10083 } 10084 if (need_decay) { 10085 /* 10086 * Ensure the rq-wide value also decays but keep it at a 10087 * reasonable floor to avoid funnies with rq->avg_idle. 10088 */ 10089 rq->max_idle_balance_cost = 10090 max((u64)sysctl_sched_migration_cost, max_cost); 10091 } 10092 rcu_read_unlock(); 10093 10094 /* 10095 * next_balance will be updated only when there is a need. 10096 * When the cpu is attached to null domain for ex, it will not be 10097 * updated. 10098 */ 10099 if (likely(update_next_balance)) 10100 rq->next_balance = next_balance; 10101 10102 } 10103 10104 static inline int on_null_domain(struct rq *rq) 10105 { 10106 return unlikely(!rcu_dereference_sched(rq->sd)); 10107 } 10108 10109 #ifdef CONFIG_NO_HZ_COMMON 10110 /* 10111 * idle load balancing details 10112 * - When one of the busy CPUs notice that there may be an idle rebalancing 10113 * needed, they will kick the idle load balancer, which then does idle 10114 * load balancing for all the idle CPUs. 10115 * - HK_FLAG_MISC CPUs are used for this task, because HK_FLAG_SCHED not set 10116 * anywhere yet. 10117 */ 10118 10119 static inline int find_new_ilb(void) 10120 { 10121 int ilb; 10122 10123 for_each_cpu_and(ilb, nohz.idle_cpus_mask, 10124 housekeeping_cpumask(HK_FLAG_MISC)) { 10125 10126 if (ilb == smp_processor_id()) 10127 continue; 10128 10129 if (idle_cpu(ilb)) 10130 return ilb; 10131 } 10132 10133 return nr_cpu_ids; 10134 } 10135 10136 /* 10137 * Kick a CPU to do the nohz balancing, if it is time for it. We pick any 10138 * idle CPU in the HK_FLAG_MISC housekeeping set (if there is one). 10139 */ 10140 static void kick_ilb(unsigned int flags) 10141 { 10142 int ilb_cpu; 10143 10144 /* 10145 * Increase nohz.next_balance only when if full ilb is triggered but 10146 * not if we only update stats. 10147 */ 10148 if (flags & NOHZ_BALANCE_KICK) 10149 nohz.next_balance = jiffies+1; 10150 10151 ilb_cpu = find_new_ilb(); 10152 10153 if (ilb_cpu >= nr_cpu_ids) 10154 return; 10155 10156 /* 10157 * Access to rq::nohz_csd is serialized by NOHZ_KICK_MASK; he who sets 10158 * the first flag owns it; cleared by nohz_csd_func(). 10159 */ 10160 flags = atomic_fetch_or(flags, nohz_flags(ilb_cpu)); 10161 if (flags & NOHZ_KICK_MASK) 10162 return; 10163 10164 /* 10165 * This way we generate an IPI on the target CPU which 10166 * is idle. And the softirq performing nohz idle load balance 10167 * will be run before returning from the IPI. 10168 */ 10169 smp_call_function_single_async(ilb_cpu, &cpu_rq(ilb_cpu)->nohz_csd); 10170 } 10171 10172 /* 10173 * Current decision point for kicking the idle load balancer in the presence 10174 * of idle CPUs in the system. 10175 */ 10176 static void nohz_balancer_kick(struct rq *rq) 10177 { 10178 unsigned long now = jiffies; 10179 struct sched_domain_shared *sds; 10180 struct sched_domain *sd; 10181 int nr_busy, i, cpu = rq->cpu; 10182 unsigned int flags = 0; 10183 10184 if (unlikely(rq->idle_balance)) 10185 return; 10186 10187 /* 10188 * We may be recently in ticked or tickless idle mode. At the first 10189 * busy tick after returning from idle, we will update the busy stats. 10190 */ 10191 nohz_balance_exit_idle(rq); 10192 10193 /* 10194 * None are in tickless mode and hence no need for NOHZ idle load 10195 * balancing. 10196 */ 10197 if (likely(!atomic_read(&nohz.nr_cpus))) 10198 return; 10199 10200 if (READ_ONCE(nohz.has_blocked) && 10201 time_after(now, READ_ONCE(nohz.next_blocked))) 10202 flags = NOHZ_STATS_KICK; 10203 10204 if (time_before(now, nohz.next_balance)) 10205 goto out; 10206 10207 if (rq->nr_running >= 2) { 10208 flags = NOHZ_KICK_MASK; 10209 goto out; 10210 } 10211 10212 rcu_read_lock(); 10213 10214 sd = rcu_dereference(rq->sd); 10215 if (sd) { 10216 /* 10217 * If there's a CFS task and the current CPU has reduced 10218 * capacity; kick the ILB to see if there's a better CPU to run 10219 * on. 10220 */ 10221 if (rq->cfs.h_nr_running >= 1 && check_cpu_capacity(rq, sd)) { 10222 flags = NOHZ_KICK_MASK; 10223 goto unlock; 10224 } 10225 } 10226 10227 sd = rcu_dereference(per_cpu(sd_asym_packing, cpu)); 10228 if (sd) { 10229 /* 10230 * When ASYM_PACKING; see if there's a more preferred CPU 10231 * currently idle; in which case, kick the ILB to move tasks 10232 * around. 10233 */ 10234 for_each_cpu_and(i, sched_domain_span(sd), nohz.idle_cpus_mask) { 10235 if (sched_asym_prefer(i, cpu)) { 10236 flags = NOHZ_KICK_MASK; 10237 goto unlock; 10238 } 10239 } 10240 } 10241 10242 sd = rcu_dereference(per_cpu(sd_asym_cpucapacity, cpu)); 10243 if (sd) { 10244 /* 10245 * When ASYM_CPUCAPACITY; see if there's a higher capacity CPU 10246 * to run the misfit task on. 10247 */ 10248 if (check_misfit_status(rq, sd)) { 10249 flags = NOHZ_KICK_MASK; 10250 goto unlock; 10251 } 10252 10253 /* 10254 * For asymmetric systems, we do not want to nicely balance 10255 * cache use, instead we want to embrace asymmetry and only 10256 * ensure tasks have enough CPU capacity. 10257 * 10258 * Skip the LLC logic because it's not relevant in that case. 10259 */ 10260 goto unlock; 10261 } 10262 10263 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu)); 10264 if (sds) { 10265 /* 10266 * If there is an imbalance between LLC domains (IOW we could 10267 * increase the overall cache use), we need some less-loaded LLC 10268 * domain to pull some load. Likewise, we may need to spread 10269 * load within the current LLC domain (e.g. packed SMT cores but 10270 * other CPUs are idle). We can't really know from here how busy 10271 * the others are - so just get a nohz balance going if it looks 10272 * like this LLC domain has tasks we could move. 10273 */ 10274 nr_busy = atomic_read(&sds->nr_busy_cpus); 10275 if (nr_busy > 1) { 10276 flags = NOHZ_KICK_MASK; 10277 goto unlock; 10278 } 10279 } 10280 unlock: 10281 rcu_read_unlock(); 10282 out: 10283 if (flags) 10284 kick_ilb(flags); 10285 } 10286 10287 static void set_cpu_sd_state_busy(int cpu) 10288 { 10289 struct sched_domain *sd; 10290 10291 rcu_read_lock(); 10292 sd = rcu_dereference(per_cpu(sd_llc, cpu)); 10293 10294 if (!sd || !sd->nohz_idle) 10295 goto unlock; 10296 sd->nohz_idle = 0; 10297 10298 atomic_inc(&sd->shared->nr_busy_cpus); 10299 unlock: 10300 rcu_read_unlock(); 10301 } 10302 10303 void nohz_balance_exit_idle(struct rq *rq) 10304 { 10305 SCHED_WARN_ON(rq != this_rq()); 10306 10307 if (likely(!rq->nohz_tick_stopped)) 10308 return; 10309 10310 rq->nohz_tick_stopped = 0; 10311 cpumask_clear_cpu(rq->cpu, nohz.idle_cpus_mask); 10312 atomic_dec(&nohz.nr_cpus); 10313 10314 set_cpu_sd_state_busy(rq->cpu); 10315 } 10316 10317 static void set_cpu_sd_state_idle(int cpu) 10318 { 10319 struct sched_domain *sd; 10320 10321 rcu_read_lock(); 10322 sd = rcu_dereference(per_cpu(sd_llc, cpu)); 10323 10324 if (!sd || sd->nohz_idle) 10325 goto unlock; 10326 sd->nohz_idle = 1; 10327 10328 atomic_dec(&sd->shared->nr_busy_cpus); 10329 unlock: 10330 rcu_read_unlock(); 10331 } 10332 10333 /* 10334 * This routine will record that the CPU is going idle with tick stopped. 10335 * This info will be used in performing idle load balancing in the future. 10336 */ 10337 void nohz_balance_enter_idle(int cpu) 10338 { 10339 struct rq *rq = cpu_rq(cpu); 10340 10341 SCHED_WARN_ON(cpu != smp_processor_id()); 10342 10343 /* If this CPU is going down, then nothing needs to be done: */ 10344 if (!cpu_active(cpu)) 10345 return; 10346 10347 /* Spare idle load balancing on CPUs that don't want to be disturbed: */ 10348 if (!housekeeping_cpu(cpu, HK_FLAG_SCHED)) 10349 return; 10350 10351 /* 10352 * Can be set safely without rq->lock held 10353 * If a clear happens, it will have evaluated last additions because 10354 * rq->lock is held during the check and the clear 10355 */ 10356 rq->has_blocked_load = 1; 10357 10358 /* 10359 * The tick is still stopped but load could have been added in the 10360 * meantime. We set the nohz.has_blocked flag to trig a check of the 10361 * *_avg. The CPU is already part of nohz.idle_cpus_mask so the clear 10362 * of nohz.has_blocked can only happen after checking the new load 10363 */ 10364 if (rq->nohz_tick_stopped) 10365 goto out; 10366 10367 /* If we're a completely isolated CPU, we don't play: */ 10368 if (on_null_domain(rq)) 10369 return; 10370 10371 rq->nohz_tick_stopped = 1; 10372 10373 cpumask_set_cpu(cpu, nohz.idle_cpus_mask); 10374 atomic_inc(&nohz.nr_cpus); 10375 10376 /* 10377 * Ensures that if nohz_idle_balance() fails to observe our 10378 * @idle_cpus_mask store, it must observe the @has_blocked 10379 * store. 10380 */ 10381 smp_mb__after_atomic(); 10382 10383 set_cpu_sd_state_idle(cpu); 10384 10385 out: 10386 /* 10387 * Each time a cpu enter idle, we assume that it has blocked load and 10388 * enable the periodic update of the load of idle cpus 10389 */ 10390 WRITE_ONCE(nohz.has_blocked, 1); 10391 } 10392 10393 static bool update_nohz_stats(struct rq *rq) 10394 { 10395 unsigned int cpu = rq->cpu; 10396 10397 if (!rq->has_blocked_load) 10398 return false; 10399 10400 if (!cpumask_test_cpu(cpu, nohz.idle_cpus_mask)) 10401 return false; 10402 10403 if (!time_after(jiffies, READ_ONCE(rq->last_blocked_load_update_tick))) 10404 return true; 10405 10406 update_blocked_averages(cpu); 10407 10408 return rq->has_blocked_load; 10409 } 10410 10411 /* 10412 * Internal function that runs load balance for all idle cpus. The load balance 10413 * can be a simple update of blocked load or a complete load balance with 10414 * tasks movement depending of flags. 10415 */ 10416 static void _nohz_idle_balance(struct rq *this_rq, unsigned int flags, 10417 enum cpu_idle_type idle) 10418 { 10419 /* Earliest time when we have to do rebalance again */ 10420 unsigned long now = jiffies; 10421 unsigned long next_balance = now + 60*HZ; 10422 bool has_blocked_load = false; 10423 int update_next_balance = 0; 10424 int this_cpu = this_rq->cpu; 10425 int balance_cpu; 10426 struct rq *rq; 10427 10428 SCHED_WARN_ON((flags & NOHZ_KICK_MASK) == NOHZ_BALANCE_KICK); 10429 10430 /* 10431 * We assume there will be no idle load after this update and clear 10432 * the has_blocked flag. If a cpu enters idle in the mean time, it will 10433 * set the has_blocked flag and trig another update of idle load. 10434 * Because a cpu that becomes idle, is added to idle_cpus_mask before 10435 * setting the flag, we are sure to not clear the state and not 10436 * check the load of an idle cpu. 10437 */ 10438 WRITE_ONCE(nohz.has_blocked, 0); 10439 10440 /* 10441 * Ensures that if we miss the CPU, we must see the has_blocked 10442 * store from nohz_balance_enter_idle(). 10443 */ 10444 smp_mb(); 10445 10446 /* 10447 * Start with the next CPU after this_cpu so we will end with this_cpu and let a 10448 * chance for other idle cpu to pull load. 10449 */ 10450 for_each_cpu_wrap(balance_cpu, nohz.idle_cpus_mask, this_cpu+1) { 10451 if (!idle_cpu(balance_cpu)) 10452 continue; 10453 10454 /* 10455 * If this CPU gets work to do, stop the load balancing 10456 * work being done for other CPUs. Next load 10457 * balancing owner will pick it up. 10458 */ 10459 if (need_resched()) { 10460 has_blocked_load = true; 10461 goto abort; 10462 } 10463 10464 rq = cpu_rq(balance_cpu); 10465 10466 has_blocked_load |= update_nohz_stats(rq); 10467 10468 /* 10469 * If time for next balance is due, 10470 * do the balance. 10471 */ 10472 if (time_after_eq(jiffies, rq->next_balance)) { 10473 struct rq_flags rf; 10474 10475 rq_lock_irqsave(rq, &rf); 10476 update_rq_clock(rq); 10477 rq_unlock_irqrestore(rq, &rf); 10478 10479 if (flags & NOHZ_BALANCE_KICK) 10480 rebalance_domains(rq, CPU_IDLE); 10481 } 10482 10483 if (time_after(next_balance, rq->next_balance)) { 10484 next_balance = rq->next_balance; 10485 update_next_balance = 1; 10486 } 10487 } 10488 10489 /* 10490 * next_balance will be updated only when there is a need. 10491 * When the CPU is attached to null domain for ex, it will not be 10492 * updated. 10493 */ 10494 if (likely(update_next_balance)) 10495 nohz.next_balance = next_balance; 10496 10497 WRITE_ONCE(nohz.next_blocked, 10498 now + msecs_to_jiffies(LOAD_AVG_PERIOD)); 10499 10500 abort: 10501 /* There is still blocked load, enable periodic update */ 10502 if (has_blocked_load) 10503 WRITE_ONCE(nohz.has_blocked, 1); 10504 } 10505 10506 /* 10507 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the 10508 * rebalancing for all the cpus for whom scheduler ticks are stopped. 10509 */ 10510 static bool nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle) 10511 { 10512 unsigned int flags = this_rq->nohz_idle_balance; 10513 10514 if (!flags) 10515 return false; 10516 10517 this_rq->nohz_idle_balance = 0; 10518 10519 if (idle != CPU_IDLE) 10520 return false; 10521 10522 _nohz_idle_balance(this_rq, flags, idle); 10523 10524 return true; 10525 } 10526 10527 /* 10528 * Check if we need to run the ILB for updating blocked load before entering 10529 * idle state. 10530 */ 10531 void nohz_run_idle_balance(int cpu) 10532 { 10533 unsigned int flags; 10534 10535 flags = atomic_fetch_andnot(NOHZ_NEWILB_KICK, nohz_flags(cpu)); 10536 10537 /* 10538 * Update the blocked load only if no SCHED_SOFTIRQ is about to happen 10539 * (ie NOHZ_STATS_KICK set) and will do the same. 10540 */ 10541 if ((flags == NOHZ_NEWILB_KICK) && !need_resched()) 10542 _nohz_idle_balance(cpu_rq(cpu), NOHZ_STATS_KICK, CPU_IDLE); 10543 } 10544 10545 static void nohz_newidle_balance(struct rq *this_rq) 10546 { 10547 int this_cpu = this_rq->cpu; 10548 10549 /* 10550 * This CPU doesn't want to be disturbed by scheduler 10551 * housekeeping 10552 */ 10553 if (!housekeeping_cpu(this_cpu, HK_FLAG_SCHED)) 10554 return; 10555 10556 /* Will wake up very soon. No time for doing anything else*/ 10557 if (this_rq->avg_idle < sysctl_sched_migration_cost) 10558 return; 10559 10560 /* Don't need to update blocked load of idle CPUs*/ 10561 if (!READ_ONCE(nohz.has_blocked) || 10562 time_before(jiffies, READ_ONCE(nohz.next_blocked))) 10563 return; 10564 10565 /* 10566 * Set the need to trigger ILB in order to update blocked load 10567 * before entering idle state. 10568 */ 10569 atomic_or(NOHZ_NEWILB_KICK, nohz_flags(this_cpu)); 10570 } 10571 10572 #else /* !CONFIG_NO_HZ_COMMON */ 10573 static inline void nohz_balancer_kick(struct rq *rq) { } 10574 10575 static inline bool nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle) 10576 { 10577 return false; 10578 } 10579 10580 static inline void nohz_newidle_balance(struct rq *this_rq) { } 10581 #endif /* CONFIG_NO_HZ_COMMON */ 10582 10583 /* 10584 * newidle_balance is called by schedule() if this_cpu is about to become 10585 * idle. Attempts to pull tasks from other CPUs. 10586 * 10587 * Returns: 10588 * < 0 - we released the lock and there are !fair tasks present 10589 * 0 - failed, no new tasks 10590 * > 0 - success, new (fair) tasks present 10591 */ 10592 static int newidle_balance(struct rq *this_rq, struct rq_flags *rf) 10593 { 10594 unsigned long next_balance = jiffies + HZ; 10595 int this_cpu = this_rq->cpu; 10596 struct sched_domain *sd; 10597 int pulled_task = 0; 10598 u64 curr_cost = 0; 10599 10600 update_misfit_status(NULL, this_rq); 10601 /* 10602 * We must set idle_stamp _before_ calling idle_balance(), such that we 10603 * measure the duration of idle_balance() as idle time. 10604 */ 10605 this_rq->idle_stamp = rq_clock(this_rq); 10606 10607 /* 10608 * Do not pull tasks towards !active CPUs... 10609 */ 10610 if (!cpu_active(this_cpu)) 10611 return 0; 10612 10613 /* 10614 * This is OK, because current is on_cpu, which avoids it being picked 10615 * for load-balance and preemption/IRQs are still disabled avoiding 10616 * further scheduler activity on it and we're being very careful to 10617 * re-start the picking loop. 10618 */ 10619 rq_unpin_lock(this_rq, rf); 10620 10621 if (this_rq->avg_idle < sysctl_sched_migration_cost || 10622 !READ_ONCE(this_rq->rd->overload)) { 10623 10624 rcu_read_lock(); 10625 sd = rcu_dereference_check_sched_domain(this_rq->sd); 10626 if (sd) 10627 update_next_balance(sd, &next_balance); 10628 rcu_read_unlock(); 10629 10630 goto out; 10631 } 10632 10633 raw_spin_unlock(&this_rq->lock); 10634 10635 update_blocked_averages(this_cpu); 10636 rcu_read_lock(); 10637 for_each_domain(this_cpu, sd) { 10638 int continue_balancing = 1; 10639 u64 t0, domain_cost; 10640 10641 if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost) { 10642 update_next_balance(sd, &next_balance); 10643 break; 10644 } 10645 10646 if (sd->flags & SD_BALANCE_NEWIDLE) { 10647 t0 = sched_clock_cpu(this_cpu); 10648 10649 pulled_task = load_balance(this_cpu, this_rq, 10650 sd, CPU_NEWLY_IDLE, 10651 &continue_balancing); 10652 10653 domain_cost = sched_clock_cpu(this_cpu) - t0; 10654 if (domain_cost > sd->max_newidle_lb_cost) 10655 sd->max_newidle_lb_cost = domain_cost; 10656 10657 curr_cost += domain_cost; 10658 } 10659 10660 update_next_balance(sd, &next_balance); 10661 10662 /* 10663 * Stop searching for tasks to pull if there are 10664 * now runnable tasks on this rq. 10665 */ 10666 if (pulled_task || this_rq->nr_running > 0) 10667 break; 10668 } 10669 rcu_read_unlock(); 10670 10671 raw_spin_lock(&this_rq->lock); 10672 10673 if (curr_cost > this_rq->max_idle_balance_cost) 10674 this_rq->max_idle_balance_cost = curr_cost; 10675 10676 /* 10677 * While browsing the domains, we released the rq lock, a task could 10678 * have been enqueued in the meantime. Since we're not going idle, 10679 * pretend we pulled a task. 10680 */ 10681 if (this_rq->cfs.h_nr_running && !pulled_task) 10682 pulled_task = 1; 10683 10684 /* Is there a task of a high priority class? */ 10685 if (this_rq->nr_running != this_rq->cfs.h_nr_running) 10686 pulled_task = -1; 10687 10688 out: 10689 /* Move the next balance forward */ 10690 if (time_after(this_rq->next_balance, next_balance)) 10691 this_rq->next_balance = next_balance; 10692 10693 if (pulled_task) 10694 this_rq->idle_stamp = 0; 10695 else 10696 nohz_newidle_balance(this_rq); 10697 10698 rq_repin_lock(this_rq, rf); 10699 10700 return pulled_task; 10701 } 10702 10703 /* 10704 * run_rebalance_domains is triggered when needed from the scheduler tick. 10705 * Also triggered for nohz idle balancing (with nohz_balancing_kick set). 10706 */ 10707 static __latent_entropy void run_rebalance_domains(struct softirq_action *h) 10708 { 10709 struct rq *this_rq = this_rq(); 10710 enum cpu_idle_type idle = this_rq->idle_balance ? 10711 CPU_IDLE : CPU_NOT_IDLE; 10712 10713 /* 10714 * If this CPU has a pending nohz_balance_kick, then do the 10715 * balancing on behalf of the other idle CPUs whose ticks are 10716 * stopped. Do nohz_idle_balance *before* rebalance_domains to 10717 * give the idle CPUs a chance to load balance. Else we may 10718 * load balance only within the local sched_domain hierarchy 10719 * and abort nohz_idle_balance altogether if we pull some load. 10720 */ 10721 if (nohz_idle_balance(this_rq, idle)) 10722 return; 10723 10724 /* normal load balance */ 10725 update_blocked_averages(this_rq->cpu); 10726 rebalance_domains(this_rq, idle); 10727 } 10728 10729 /* 10730 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing. 10731 */ 10732 void trigger_load_balance(struct rq *rq) 10733 { 10734 /* 10735 * Don't need to rebalance while attached to NULL domain or 10736 * runqueue CPU is not active 10737 */ 10738 if (unlikely(on_null_domain(rq) || !cpu_active(cpu_of(rq)))) 10739 return; 10740 10741 if (time_after_eq(jiffies, rq->next_balance)) 10742 raise_softirq(SCHED_SOFTIRQ); 10743 10744 nohz_balancer_kick(rq); 10745 } 10746 10747 static void rq_online_fair(struct rq *rq) 10748 { 10749 update_sysctl(); 10750 10751 update_runtime_enabled(rq); 10752 } 10753 10754 static void rq_offline_fair(struct rq *rq) 10755 { 10756 update_sysctl(); 10757 10758 /* Ensure any throttled groups are reachable by pick_next_task */ 10759 unthrottle_offline_cfs_rqs(rq); 10760 } 10761 10762 #endif /* CONFIG_SMP */ 10763 10764 /* 10765 * scheduler tick hitting a task of our scheduling class. 10766 * 10767 * NOTE: This function can be called remotely by the tick offload that 10768 * goes along full dynticks. Therefore no local assumption can be made 10769 * and everything must be accessed through the @rq and @curr passed in 10770 * parameters. 10771 */ 10772 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued) 10773 { 10774 struct cfs_rq *cfs_rq; 10775 struct sched_entity *se = &curr->se; 10776 10777 for_each_sched_entity(se) { 10778 cfs_rq = cfs_rq_of(se); 10779 entity_tick(cfs_rq, se, queued); 10780 } 10781 10782 if (static_branch_unlikely(&sched_numa_balancing)) 10783 task_tick_numa(rq, curr); 10784 10785 update_misfit_status(curr, rq); 10786 update_overutilized_status(task_rq(curr)); 10787 } 10788 10789 /* 10790 * called on fork with the child task as argument from the parent's context 10791 * - child not yet on the tasklist 10792 * - preemption disabled 10793 */ 10794 static void task_fork_fair(struct task_struct *p) 10795 { 10796 struct cfs_rq *cfs_rq; 10797 struct sched_entity *se = &p->se, *curr; 10798 struct rq *rq = this_rq(); 10799 struct rq_flags rf; 10800 10801 rq_lock(rq, &rf); 10802 update_rq_clock(rq); 10803 10804 cfs_rq = task_cfs_rq(current); 10805 curr = cfs_rq->curr; 10806 if (curr) { 10807 update_curr(cfs_rq); 10808 se->vruntime = curr->vruntime; 10809 } 10810 place_entity(cfs_rq, se, 1); 10811 10812 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) { 10813 /* 10814 * Upon rescheduling, sched_class::put_prev_task() will place 10815 * 'current' within the tree based on its new key value. 10816 */ 10817 swap(curr->vruntime, se->vruntime); 10818 resched_curr(rq); 10819 } 10820 10821 se->vruntime -= cfs_rq->min_vruntime; 10822 rq_unlock(rq, &rf); 10823 } 10824 10825 /* 10826 * Priority of the task has changed. Check to see if we preempt 10827 * the current task. 10828 */ 10829 static void 10830 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio) 10831 { 10832 if (!task_on_rq_queued(p)) 10833 return; 10834 10835 if (rq->cfs.nr_running == 1) 10836 return; 10837 10838 /* 10839 * Reschedule if we are currently running on this runqueue and 10840 * our priority decreased, or if we are not currently running on 10841 * this runqueue and our priority is higher than the current's 10842 */ 10843 if (task_current(rq, p)) { 10844 if (p->prio > oldprio) 10845 resched_curr(rq); 10846 } else 10847 check_preempt_curr(rq, p, 0); 10848 } 10849 10850 static inline bool vruntime_normalized(struct task_struct *p) 10851 { 10852 struct sched_entity *se = &p->se; 10853 10854 /* 10855 * In both the TASK_ON_RQ_QUEUED and TASK_ON_RQ_MIGRATING cases, 10856 * the dequeue_entity(.flags=0) will already have normalized the 10857 * vruntime. 10858 */ 10859 if (p->on_rq) 10860 return true; 10861 10862 /* 10863 * When !on_rq, vruntime of the task has usually NOT been normalized. 10864 * But there are some cases where it has already been normalized: 10865 * 10866 * - A forked child which is waiting for being woken up by 10867 * wake_up_new_task(). 10868 * - A task which has been woken up by try_to_wake_up() and 10869 * waiting for actually being woken up by sched_ttwu_pending(). 10870 */ 10871 if (!se->sum_exec_runtime || 10872 (p->state == TASK_WAKING && p->sched_remote_wakeup)) 10873 return true; 10874 10875 return false; 10876 } 10877 10878 #ifdef CONFIG_FAIR_GROUP_SCHED 10879 /* 10880 * Propagate the changes of the sched_entity across the tg tree to make it 10881 * visible to the root 10882 */ 10883 static void propagate_entity_cfs_rq(struct sched_entity *se) 10884 { 10885 struct cfs_rq *cfs_rq; 10886 10887 list_add_leaf_cfs_rq(cfs_rq_of(se)); 10888 10889 /* Start to propagate at parent */ 10890 se = se->parent; 10891 10892 for_each_sched_entity(se) { 10893 cfs_rq = cfs_rq_of(se); 10894 10895 if (!cfs_rq_throttled(cfs_rq)){ 10896 update_load_avg(cfs_rq, se, UPDATE_TG); 10897 list_add_leaf_cfs_rq(cfs_rq); 10898 continue; 10899 } 10900 10901 if (list_add_leaf_cfs_rq(cfs_rq)) 10902 break; 10903 } 10904 } 10905 #else 10906 static void propagate_entity_cfs_rq(struct sched_entity *se) { } 10907 #endif 10908 10909 static void detach_entity_cfs_rq(struct sched_entity *se) 10910 { 10911 struct cfs_rq *cfs_rq = cfs_rq_of(se); 10912 10913 /* Catch up with the cfs_rq and remove our load when we leave */ 10914 update_load_avg(cfs_rq, se, 0); 10915 detach_entity_load_avg(cfs_rq, se); 10916 update_tg_load_avg(cfs_rq); 10917 propagate_entity_cfs_rq(se); 10918 } 10919 10920 static void attach_entity_cfs_rq(struct sched_entity *se) 10921 { 10922 struct cfs_rq *cfs_rq = cfs_rq_of(se); 10923 10924 #ifdef CONFIG_FAIR_GROUP_SCHED 10925 /* 10926 * Since the real-depth could have been changed (only FAIR 10927 * class maintain depth value), reset depth properly. 10928 */ 10929 se->depth = se->parent ? se->parent->depth + 1 : 0; 10930 #endif 10931 10932 /* Synchronize entity with its cfs_rq */ 10933 update_load_avg(cfs_rq, se, sched_feat(ATTACH_AGE_LOAD) ? 0 : SKIP_AGE_LOAD); 10934 attach_entity_load_avg(cfs_rq, se); 10935 update_tg_load_avg(cfs_rq); 10936 propagate_entity_cfs_rq(se); 10937 } 10938 10939 static void detach_task_cfs_rq(struct task_struct *p) 10940 { 10941 struct sched_entity *se = &p->se; 10942 struct cfs_rq *cfs_rq = cfs_rq_of(se); 10943 10944 if (!vruntime_normalized(p)) { 10945 /* 10946 * Fix up our vruntime so that the current sleep doesn't 10947 * cause 'unlimited' sleep bonus. 10948 */ 10949 place_entity(cfs_rq, se, 0); 10950 se->vruntime -= cfs_rq->min_vruntime; 10951 } 10952 10953 detach_entity_cfs_rq(se); 10954 } 10955 10956 static void attach_task_cfs_rq(struct task_struct *p) 10957 { 10958 struct sched_entity *se = &p->se; 10959 struct cfs_rq *cfs_rq = cfs_rq_of(se); 10960 10961 attach_entity_cfs_rq(se); 10962 10963 if (!vruntime_normalized(p)) 10964 se->vruntime += cfs_rq->min_vruntime; 10965 } 10966 10967 static void switched_from_fair(struct rq *rq, struct task_struct *p) 10968 { 10969 detach_task_cfs_rq(p); 10970 } 10971 10972 static void switched_to_fair(struct rq *rq, struct task_struct *p) 10973 { 10974 attach_task_cfs_rq(p); 10975 10976 if (task_on_rq_queued(p)) { 10977 /* 10978 * We were most likely switched from sched_rt, so 10979 * kick off the schedule if running, otherwise just see 10980 * if we can still preempt the current task. 10981 */ 10982 if (task_current(rq, p)) 10983 resched_curr(rq); 10984 else 10985 check_preempt_curr(rq, p, 0); 10986 } 10987 } 10988 10989 /* Account for a task changing its policy or group. 10990 * 10991 * This routine is mostly called to set cfs_rq->curr field when a task 10992 * migrates between groups/classes. 10993 */ 10994 static void set_next_task_fair(struct rq *rq, struct task_struct *p, bool first) 10995 { 10996 struct sched_entity *se = &p->se; 10997 10998 #ifdef CONFIG_SMP 10999 if (task_on_rq_queued(p)) { 11000 /* 11001 * Move the next running task to the front of the list, so our 11002 * cfs_tasks list becomes MRU one. 11003 */ 11004 list_move(&se->group_node, &rq->cfs_tasks); 11005 } 11006 #endif 11007 11008 for_each_sched_entity(se) { 11009 struct cfs_rq *cfs_rq = cfs_rq_of(se); 11010 11011 set_next_entity(cfs_rq, se); 11012 /* ensure bandwidth has been allocated on our new cfs_rq */ 11013 account_cfs_rq_runtime(cfs_rq, 0); 11014 } 11015 } 11016 11017 void init_cfs_rq(struct cfs_rq *cfs_rq) 11018 { 11019 cfs_rq->tasks_timeline = RB_ROOT_CACHED; 11020 cfs_rq->min_vruntime = (u64)(-(1LL << 20)); 11021 #ifndef CONFIG_64BIT 11022 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime; 11023 #endif 11024 #ifdef CONFIG_SMP 11025 raw_spin_lock_init(&cfs_rq->removed.lock); 11026 #endif 11027 } 11028 11029 #ifdef CONFIG_FAIR_GROUP_SCHED 11030 static void task_set_group_fair(struct task_struct *p) 11031 { 11032 struct sched_entity *se = &p->se; 11033 11034 set_task_rq(p, task_cpu(p)); 11035 se->depth = se->parent ? se->parent->depth + 1 : 0; 11036 } 11037 11038 static void task_move_group_fair(struct task_struct *p) 11039 { 11040 detach_task_cfs_rq(p); 11041 set_task_rq(p, task_cpu(p)); 11042 11043 #ifdef CONFIG_SMP 11044 /* Tell se's cfs_rq has been changed -- migrated */ 11045 p->se.avg.last_update_time = 0; 11046 #endif 11047 attach_task_cfs_rq(p); 11048 } 11049 11050 static void task_change_group_fair(struct task_struct *p, int type) 11051 { 11052 switch (type) { 11053 case TASK_SET_GROUP: 11054 task_set_group_fair(p); 11055 break; 11056 11057 case TASK_MOVE_GROUP: 11058 task_move_group_fair(p); 11059 break; 11060 } 11061 } 11062 11063 void free_fair_sched_group(struct task_group *tg) 11064 { 11065 int i; 11066 11067 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg)); 11068 11069 for_each_possible_cpu(i) { 11070 if (tg->cfs_rq) 11071 kfree(tg->cfs_rq[i]); 11072 if (tg->se) 11073 kfree(tg->se[i]); 11074 } 11075 11076 kfree(tg->cfs_rq); 11077 kfree(tg->se); 11078 } 11079 11080 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent) 11081 { 11082 struct sched_entity *se; 11083 struct cfs_rq *cfs_rq; 11084 int i; 11085 11086 tg->cfs_rq = kcalloc(nr_cpu_ids, sizeof(cfs_rq), GFP_KERNEL); 11087 if (!tg->cfs_rq) 11088 goto err; 11089 tg->se = kcalloc(nr_cpu_ids, sizeof(se), GFP_KERNEL); 11090 if (!tg->se) 11091 goto err; 11092 11093 tg->shares = NICE_0_LOAD; 11094 11095 init_cfs_bandwidth(tg_cfs_bandwidth(tg)); 11096 11097 for_each_possible_cpu(i) { 11098 cfs_rq = kzalloc_node(sizeof(struct cfs_rq), 11099 GFP_KERNEL, cpu_to_node(i)); 11100 if (!cfs_rq) 11101 goto err; 11102 11103 se = kzalloc_node(sizeof(struct sched_entity), 11104 GFP_KERNEL, cpu_to_node(i)); 11105 if (!se) 11106 goto err_free_rq; 11107 11108 init_cfs_rq(cfs_rq); 11109 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]); 11110 init_entity_runnable_average(se); 11111 } 11112 11113 return 1; 11114 11115 err_free_rq: 11116 kfree(cfs_rq); 11117 err: 11118 return 0; 11119 } 11120 11121 void online_fair_sched_group(struct task_group *tg) 11122 { 11123 struct sched_entity *se; 11124 struct rq_flags rf; 11125 struct rq *rq; 11126 int i; 11127 11128 for_each_possible_cpu(i) { 11129 rq = cpu_rq(i); 11130 se = tg->se[i]; 11131 rq_lock_irq(rq, &rf); 11132 update_rq_clock(rq); 11133 attach_entity_cfs_rq(se); 11134 sync_throttle(tg, i); 11135 rq_unlock_irq(rq, &rf); 11136 } 11137 } 11138 11139 void unregister_fair_sched_group(struct task_group *tg) 11140 { 11141 unsigned long flags; 11142 struct rq *rq; 11143 int cpu; 11144 11145 for_each_possible_cpu(cpu) { 11146 if (tg->se[cpu]) 11147 remove_entity_load_avg(tg->se[cpu]); 11148 11149 /* 11150 * Only empty task groups can be destroyed; so we can speculatively 11151 * check on_list without danger of it being re-added. 11152 */ 11153 if (!tg->cfs_rq[cpu]->on_list) 11154 continue; 11155 11156 rq = cpu_rq(cpu); 11157 11158 raw_spin_lock_irqsave(&rq->lock, flags); 11159 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]); 11160 raw_spin_unlock_irqrestore(&rq->lock, flags); 11161 } 11162 } 11163 11164 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq, 11165 struct sched_entity *se, int cpu, 11166 struct sched_entity *parent) 11167 { 11168 struct rq *rq = cpu_rq(cpu); 11169 11170 cfs_rq->tg = tg; 11171 cfs_rq->rq = rq; 11172 init_cfs_rq_runtime(cfs_rq); 11173 11174 tg->cfs_rq[cpu] = cfs_rq; 11175 tg->se[cpu] = se; 11176 11177 /* se could be NULL for root_task_group */ 11178 if (!se) 11179 return; 11180 11181 if (!parent) { 11182 se->cfs_rq = &rq->cfs; 11183 se->depth = 0; 11184 } else { 11185 se->cfs_rq = parent->my_q; 11186 se->depth = parent->depth + 1; 11187 } 11188 11189 se->my_q = cfs_rq; 11190 /* guarantee group entities always have weight */ 11191 update_load_set(&se->load, NICE_0_LOAD); 11192 se->parent = parent; 11193 } 11194 11195 static DEFINE_MUTEX(shares_mutex); 11196 11197 int sched_group_set_shares(struct task_group *tg, unsigned long shares) 11198 { 11199 int i; 11200 11201 /* 11202 * We can't change the weight of the root cgroup. 11203 */ 11204 if (!tg->se[0]) 11205 return -EINVAL; 11206 11207 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES)); 11208 11209 mutex_lock(&shares_mutex); 11210 if (tg->shares == shares) 11211 goto done; 11212 11213 tg->shares = shares; 11214 for_each_possible_cpu(i) { 11215 struct rq *rq = cpu_rq(i); 11216 struct sched_entity *se = tg->se[i]; 11217 struct rq_flags rf; 11218 11219 /* Propagate contribution to hierarchy */ 11220 rq_lock_irqsave(rq, &rf); 11221 update_rq_clock(rq); 11222 for_each_sched_entity(se) { 11223 update_load_avg(cfs_rq_of(se), se, UPDATE_TG); 11224 update_cfs_group(se); 11225 } 11226 rq_unlock_irqrestore(rq, &rf); 11227 } 11228 11229 done: 11230 mutex_unlock(&shares_mutex); 11231 return 0; 11232 } 11233 #else /* CONFIG_FAIR_GROUP_SCHED */ 11234 11235 void free_fair_sched_group(struct task_group *tg) { } 11236 11237 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent) 11238 { 11239 return 1; 11240 } 11241 11242 void online_fair_sched_group(struct task_group *tg) { } 11243 11244 void unregister_fair_sched_group(struct task_group *tg) { } 11245 11246 #endif /* CONFIG_FAIR_GROUP_SCHED */ 11247 11248 11249 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task) 11250 { 11251 struct sched_entity *se = &task->se; 11252 unsigned int rr_interval = 0; 11253 11254 /* 11255 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise 11256 * idle runqueue: 11257 */ 11258 if (rq->cfs.load.weight) 11259 rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se)); 11260 11261 return rr_interval; 11262 } 11263 11264 /* 11265 * All the scheduling class methods: 11266 */ 11267 DEFINE_SCHED_CLASS(fair) = { 11268 11269 .enqueue_task = enqueue_task_fair, 11270 .dequeue_task = dequeue_task_fair, 11271 .yield_task = yield_task_fair, 11272 .yield_to_task = yield_to_task_fair, 11273 11274 .check_preempt_curr = check_preempt_wakeup, 11275 11276 .pick_next_task = __pick_next_task_fair, 11277 .put_prev_task = put_prev_task_fair, 11278 .set_next_task = set_next_task_fair, 11279 11280 #ifdef CONFIG_SMP 11281 .balance = balance_fair, 11282 .select_task_rq = select_task_rq_fair, 11283 .migrate_task_rq = migrate_task_rq_fair, 11284 11285 .rq_online = rq_online_fair, 11286 .rq_offline = rq_offline_fair, 11287 11288 .task_dead = task_dead_fair, 11289 .set_cpus_allowed = set_cpus_allowed_common, 11290 #endif 11291 11292 .task_tick = task_tick_fair, 11293 .task_fork = task_fork_fair, 11294 11295 .prio_changed = prio_changed_fair, 11296 .switched_from = switched_from_fair, 11297 .switched_to = switched_to_fair, 11298 11299 .get_rr_interval = get_rr_interval_fair, 11300 11301 .update_curr = update_curr_fair, 11302 11303 #ifdef CONFIG_FAIR_GROUP_SCHED 11304 .task_change_group = task_change_group_fair, 11305 #endif 11306 11307 #ifdef CONFIG_UCLAMP_TASK 11308 .uclamp_enabled = 1, 11309 #endif 11310 }; 11311 11312 #ifdef CONFIG_SCHED_DEBUG 11313 void print_cfs_stats(struct seq_file *m, int cpu) 11314 { 11315 struct cfs_rq *cfs_rq, *pos; 11316 11317 rcu_read_lock(); 11318 for_each_leaf_cfs_rq_safe(cpu_rq(cpu), cfs_rq, pos) 11319 print_cfs_rq(m, cpu, cfs_rq); 11320 rcu_read_unlock(); 11321 } 11322 11323 #ifdef CONFIG_NUMA_BALANCING 11324 void show_numa_stats(struct task_struct *p, struct seq_file *m) 11325 { 11326 int node; 11327 unsigned long tsf = 0, tpf = 0, gsf = 0, gpf = 0; 11328 struct numa_group *ng; 11329 11330 rcu_read_lock(); 11331 ng = rcu_dereference(p->numa_group); 11332 for_each_online_node(node) { 11333 if (p->numa_faults) { 11334 tsf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 0)]; 11335 tpf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 1)]; 11336 } 11337 if (ng) { 11338 gsf = ng->faults[task_faults_idx(NUMA_MEM, node, 0)], 11339 gpf = ng->faults[task_faults_idx(NUMA_MEM, node, 1)]; 11340 } 11341 print_numa_stats(m, node, tsf, tpf, gsf, gpf); 11342 } 11343 rcu_read_unlock(); 11344 } 11345 #endif /* CONFIG_NUMA_BALANCING */ 11346 #endif /* CONFIG_SCHED_DEBUG */ 11347 11348 __init void init_sched_fair_class(void) 11349 { 11350 #ifdef CONFIG_SMP 11351 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains); 11352 11353 #ifdef CONFIG_NO_HZ_COMMON 11354 nohz.next_balance = jiffies; 11355 nohz.next_blocked = jiffies; 11356 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT); 11357 #endif 11358 #endif /* SMP */ 11359 11360 } 11361 11362 /* 11363 * Helper functions to facilitate extracting info from tracepoints. 11364 */ 11365 11366 const struct sched_avg *sched_trace_cfs_rq_avg(struct cfs_rq *cfs_rq) 11367 { 11368 #ifdef CONFIG_SMP 11369 return cfs_rq ? &cfs_rq->avg : NULL; 11370 #else 11371 return NULL; 11372 #endif 11373 } 11374 EXPORT_SYMBOL_GPL(sched_trace_cfs_rq_avg); 11375 11376 char *sched_trace_cfs_rq_path(struct cfs_rq *cfs_rq, char *str, int len) 11377 { 11378 if (!cfs_rq) { 11379 if (str) 11380 strlcpy(str, "(null)", len); 11381 else 11382 return NULL; 11383 } 11384 11385 cfs_rq_tg_path(cfs_rq, str, len); 11386 return str; 11387 } 11388 EXPORT_SYMBOL_GPL(sched_trace_cfs_rq_path); 11389 11390 int sched_trace_cfs_rq_cpu(struct cfs_rq *cfs_rq) 11391 { 11392 return cfs_rq ? cpu_of(rq_of(cfs_rq)) : -1; 11393 } 11394 EXPORT_SYMBOL_GPL(sched_trace_cfs_rq_cpu); 11395 11396 const struct sched_avg *sched_trace_rq_avg_rt(struct rq *rq) 11397 { 11398 #ifdef CONFIG_SMP 11399 return rq ? &rq->avg_rt : NULL; 11400 #else 11401 return NULL; 11402 #endif 11403 } 11404 EXPORT_SYMBOL_GPL(sched_trace_rq_avg_rt); 11405 11406 const struct sched_avg *sched_trace_rq_avg_dl(struct rq *rq) 11407 { 11408 #ifdef CONFIG_SMP 11409 return rq ? &rq->avg_dl : NULL; 11410 #else 11411 return NULL; 11412 #endif 11413 } 11414 EXPORT_SYMBOL_GPL(sched_trace_rq_avg_dl); 11415 11416 const struct sched_avg *sched_trace_rq_avg_irq(struct rq *rq) 11417 { 11418 #if defined(CONFIG_SMP) && defined(CONFIG_HAVE_SCHED_AVG_IRQ) 11419 return rq ? &rq->avg_irq : NULL; 11420 #else 11421 return NULL; 11422 #endif 11423 } 11424 EXPORT_SYMBOL_GPL(sched_trace_rq_avg_irq); 11425 11426 int sched_trace_rq_cpu(struct rq *rq) 11427 { 11428 return rq ? cpu_of(rq) : -1; 11429 } 11430 EXPORT_SYMBOL_GPL(sched_trace_rq_cpu); 11431 11432 int sched_trace_rq_cpu_capacity(struct rq *rq) 11433 { 11434 return rq ? 11435 #ifdef CONFIG_SMP 11436 rq->cpu_capacity 11437 #else 11438 SCHED_CAPACITY_SCALE 11439 #endif 11440 : -1; 11441 } 11442 EXPORT_SYMBOL_GPL(sched_trace_rq_cpu_capacity); 11443 11444 const struct cpumask *sched_trace_rd_span(struct root_domain *rd) 11445 { 11446 #ifdef CONFIG_SMP 11447 return rd ? rd->span : NULL; 11448 #else 11449 return NULL; 11450 #endif 11451 } 11452 EXPORT_SYMBOL_GPL(sched_trace_rd_span); 11453 11454 int sched_trace_rq_nr_running(struct rq *rq) 11455 { 11456 return rq ? rq->nr_running : -1; 11457 } 11458 EXPORT_SYMBOL_GPL(sched_trace_rq_nr_running); 11459