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 enum sched_tunable_scaling 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 #endif 117 118 #ifdef CONFIG_CFS_BANDWIDTH 119 /* 120 * Amount of runtime to allocate from global (tg) to local (per-cfs_rq) pool 121 * each time a cfs_rq requests quota. 122 * 123 * Note: in the case that the slice exceeds the runtime remaining (either due 124 * to consumption or the quota being specified to be smaller than the slice) 125 * we will always only issue the remaining available time. 126 * 127 * (default: 5 msec, units: microseconds) 128 */ 129 unsigned int sysctl_sched_cfs_bandwidth_slice = 5000UL; 130 #endif 131 132 static inline void update_load_add(struct load_weight *lw, unsigned long inc) 133 { 134 lw->weight += inc; 135 lw->inv_weight = 0; 136 } 137 138 static inline void update_load_sub(struct load_weight *lw, unsigned long dec) 139 { 140 lw->weight -= dec; 141 lw->inv_weight = 0; 142 } 143 144 static inline void update_load_set(struct load_weight *lw, unsigned long w) 145 { 146 lw->weight = w; 147 lw->inv_weight = 0; 148 } 149 150 /* 151 * Increase the granularity value when there are more CPUs, 152 * because with more CPUs the 'effective latency' as visible 153 * to users decreases. But the relationship is not linear, 154 * so pick a second-best guess by going with the log2 of the 155 * number of CPUs. 156 * 157 * This idea comes from the SD scheduler of Con Kolivas: 158 */ 159 static unsigned int get_update_sysctl_factor(void) 160 { 161 unsigned int cpus = min_t(unsigned int, num_online_cpus(), 8); 162 unsigned int factor; 163 164 switch (sysctl_sched_tunable_scaling) { 165 case SCHED_TUNABLESCALING_NONE: 166 factor = 1; 167 break; 168 case SCHED_TUNABLESCALING_LINEAR: 169 factor = cpus; 170 break; 171 case SCHED_TUNABLESCALING_LOG: 172 default: 173 factor = 1 + ilog2(cpus); 174 break; 175 } 176 177 return factor; 178 } 179 180 static void update_sysctl(void) 181 { 182 unsigned int factor = get_update_sysctl_factor(); 183 184 #define SET_SYSCTL(name) \ 185 (sysctl_##name = (factor) * normalized_sysctl_##name) 186 SET_SYSCTL(sched_min_granularity); 187 SET_SYSCTL(sched_latency); 188 SET_SYSCTL(sched_wakeup_granularity); 189 #undef SET_SYSCTL 190 } 191 192 void __init sched_init_granularity(void) 193 { 194 update_sysctl(); 195 } 196 197 #define WMULT_CONST (~0U) 198 #define WMULT_SHIFT 32 199 200 static void __update_inv_weight(struct load_weight *lw) 201 { 202 unsigned long w; 203 204 if (likely(lw->inv_weight)) 205 return; 206 207 w = scale_load_down(lw->weight); 208 209 if (BITS_PER_LONG > 32 && unlikely(w >= WMULT_CONST)) 210 lw->inv_weight = 1; 211 else if (unlikely(!w)) 212 lw->inv_weight = WMULT_CONST; 213 else 214 lw->inv_weight = WMULT_CONST / w; 215 } 216 217 /* 218 * delta_exec * weight / lw.weight 219 * OR 220 * (delta_exec * (weight * lw->inv_weight)) >> WMULT_SHIFT 221 * 222 * Either weight := NICE_0_LOAD and lw \e sched_prio_to_wmult[], in which case 223 * we're guaranteed shift stays positive because inv_weight is guaranteed to 224 * fit 32 bits, and NICE_0_LOAD gives another 10 bits; therefore shift >= 22. 225 * 226 * Or, weight =< lw.weight (because lw.weight is the runqueue weight), thus 227 * weight/lw.weight <= 1, and therefore our shift will also be positive. 228 */ 229 static u64 __calc_delta(u64 delta_exec, unsigned long weight, struct load_weight *lw) 230 { 231 u64 fact = scale_load_down(weight); 232 int shift = WMULT_SHIFT; 233 234 __update_inv_weight(lw); 235 236 if (unlikely(fact >> 32)) { 237 while (fact >> 32) { 238 fact >>= 1; 239 shift--; 240 } 241 } 242 243 fact = mul_u32_u32(fact, lw->inv_weight); 244 245 while (fact >> 32) { 246 fact >>= 1; 247 shift--; 248 } 249 250 return mul_u64_u32_shr(delta_exec, fact, shift); 251 } 252 253 254 const struct sched_class fair_sched_class; 255 256 /************************************************************** 257 * CFS operations on generic schedulable entities: 258 */ 259 260 #ifdef CONFIG_FAIR_GROUP_SCHED 261 static inline struct task_struct *task_of(struct sched_entity *se) 262 { 263 SCHED_WARN_ON(!entity_is_task(se)); 264 return container_of(se, struct task_struct, se); 265 } 266 267 /* Walk up scheduling entities hierarchy */ 268 #define for_each_sched_entity(se) \ 269 for (; se; se = se->parent) 270 271 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p) 272 { 273 return p->se.cfs_rq; 274 } 275 276 /* runqueue on which this entity is (to be) queued */ 277 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se) 278 { 279 return se->cfs_rq; 280 } 281 282 /* runqueue "owned" by this group */ 283 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp) 284 { 285 return grp->my_q; 286 } 287 288 static inline void cfs_rq_tg_path(struct cfs_rq *cfs_rq, char *path, int len) 289 { 290 if (!path) 291 return; 292 293 if (cfs_rq && task_group_is_autogroup(cfs_rq->tg)) 294 autogroup_path(cfs_rq->tg, path, len); 295 else if (cfs_rq && cfs_rq->tg->css.cgroup) 296 cgroup_path(cfs_rq->tg->css.cgroup, path, len); 297 else 298 strlcpy(path, "(null)", len); 299 } 300 301 static inline bool list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq) 302 { 303 struct rq *rq = rq_of(cfs_rq); 304 int cpu = cpu_of(rq); 305 306 if (cfs_rq->on_list) 307 return rq->tmp_alone_branch == &rq->leaf_cfs_rq_list; 308 309 cfs_rq->on_list = 1; 310 311 /* 312 * Ensure we either appear before our parent (if already 313 * enqueued) or force our parent to appear after us when it is 314 * enqueued. The fact that we always enqueue bottom-up 315 * reduces this to two cases and a special case for the root 316 * cfs_rq. Furthermore, it also means that we will always reset 317 * tmp_alone_branch either when the branch is connected 318 * to a tree or when we reach the top of the tree 319 */ 320 if (cfs_rq->tg->parent && 321 cfs_rq->tg->parent->cfs_rq[cpu]->on_list) { 322 /* 323 * If parent is already on the list, we add the child 324 * just before. Thanks to circular linked property of 325 * the list, this means to put the child at the tail 326 * of the list that starts by parent. 327 */ 328 list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list, 329 &(cfs_rq->tg->parent->cfs_rq[cpu]->leaf_cfs_rq_list)); 330 /* 331 * The branch is now connected to its tree so we can 332 * reset tmp_alone_branch to the beginning of the 333 * list. 334 */ 335 rq->tmp_alone_branch = &rq->leaf_cfs_rq_list; 336 return true; 337 } 338 339 if (!cfs_rq->tg->parent) { 340 /* 341 * cfs rq without parent should be put 342 * at the tail of the list. 343 */ 344 list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list, 345 &rq->leaf_cfs_rq_list); 346 /* 347 * We have reach the top of a tree so we can reset 348 * tmp_alone_branch to the beginning of the list. 349 */ 350 rq->tmp_alone_branch = &rq->leaf_cfs_rq_list; 351 return true; 352 } 353 354 /* 355 * The parent has not already been added so we want to 356 * make sure that it will be put after us. 357 * tmp_alone_branch points to the begin of the branch 358 * where we will add parent. 359 */ 360 list_add_rcu(&cfs_rq->leaf_cfs_rq_list, rq->tmp_alone_branch); 361 /* 362 * update tmp_alone_branch to points to the new begin 363 * of the branch 364 */ 365 rq->tmp_alone_branch = &cfs_rq->leaf_cfs_rq_list; 366 return false; 367 } 368 369 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq) 370 { 371 if (cfs_rq->on_list) { 372 struct rq *rq = rq_of(cfs_rq); 373 374 /* 375 * With cfs_rq being unthrottled/throttled during an enqueue, 376 * it can happen the tmp_alone_branch points the a leaf that 377 * we finally want to del. In this case, tmp_alone_branch moves 378 * to the prev element but it will point to rq->leaf_cfs_rq_list 379 * at the end of the enqueue. 380 */ 381 if (rq->tmp_alone_branch == &cfs_rq->leaf_cfs_rq_list) 382 rq->tmp_alone_branch = cfs_rq->leaf_cfs_rq_list.prev; 383 384 list_del_rcu(&cfs_rq->leaf_cfs_rq_list); 385 cfs_rq->on_list = 0; 386 } 387 } 388 389 static inline void assert_list_leaf_cfs_rq(struct rq *rq) 390 { 391 SCHED_WARN_ON(rq->tmp_alone_branch != &rq->leaf_cfs_rq_list); 392 } 393 394 /* Iterate thr' all leaf cfs_rq's on a runqueue */ 395 #define for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos) \ 396 list_for_each_entry_safe(cfs_rq, pos, &rq->leaf_cfs_rq_list, \ 397 leaf_cfs_rq_list) 398 399 /* Do the two (enqueued) entities belong to the same group ? */ 400 static inline struct cfs_rq * 401 is_same_group(struct sched_entity *se, struct sched_entity *pse) 402 { 403 if (se->cfs_rq == pse->cfs_rq) 404 return se->cfs_rq; 405 406 return NULL; 407 } 408 409 static inline struct sched_entity *parent_entity(struct sched_entity *se) 410 { 411 return se->parent; 412 } 413 414 static void 415 find_matching_se(struct sched_entity **se, struct sched_entity **pse) 416 { 417 int se_depth, pse_depth; 418 419 /* 420 * preemption test can be made between sibling entities who are in the 421 * same cfs_rq i.e who have a common parent. Walk up the hierarchy of 422 * both tasks until we find their ancestors who are siblings of common 423 * parent. 424 */ 425 426 /* First walk up until both entities are at same depth */ 427 se_depth = (*se)->depth; 428 pse_depth = (*pse)->depth; 429 430 while (se_depth > pse_depth) { 431 se_depth--; 432 *se = parent_entity(*se); 433 } 434 435 while (pse_depth > se_depth) { 436 pse_depth--; 437 *pse = parent_entity(*pse); 438 } 439 440 while (!is_same_group(*se, *pse)) { 441 *se = parent_entity(*se); 442 *pse = parent_entity(*pse); 443 } 444 } 445 446 #else /* !CONFIG_FAIR_GROUP_SCHED */ 447 448 static inline struct task_struct *task_of(struct sched_entity *se) 449 { 450 return container_of(se, struct task_struct, se); 451 } 452 453 #define for_each_sched_entity(se) \ 454 for (; se; se = NULL) 455 456 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p) 457 { 458 return &task_rq(p)->cfs; 459 } 460 461 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se) 462 { 463 struct task_struct *p = task_of(se); 464 struct rq *rq = task_rq(p); 465 466 return &rq->cfs; 467 } 468 469 /* runqueue "owned" by this group */ 470 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp) 471 { 472 return NULL; 473 } 474 475 static inline void cfs_rq_tg_path(struct cfs_rq *cfs_rq, char *path, int len) 476 { 477 if (path) 478 strlcpy(path, "(null)", len); 479 } 480 481 static inline bool list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq) 482 { 483 return true; 484 } 485 486 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq) 487 { 488 } 489 490 static inline void assert_list_leaf_cfs_rq(struct rq *rq) 491 { 492 } 493 494 #define for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos) \ 495 for (cfs_rq = &rq->cfs, pos = NULL; cfs_rq; cfs_rq = pos) 496 497 static inline struct sched_entity *parent_entity(struct sched_entity *se) 498 { 499 return NULL; 500 } 501 502 static inline void 503 find_matching_se(struct sched_entity **se, struct sched_entity **pse) 504 { 505 } 506 507 #endif /* CONFIG_FAIR_GROUP_SCHED */ 508 509 static __always_inline 510 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec); 511 512 /************************************************************** 513 * Scheduling class tree data structure manipulation methods: 514 */ 515 516 static inline u64 max_vruntime(u64 max_vruntime, u64 vruntime) 517 { 518 s64 delta = (s64)(vruntime - max_vruntime); 519 if (delta > 0) 520 max_vruntime = vruntime; 521 522 return max_vruntime; 523 } 524 525 static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime) 526 { 527 s64 delta = (s64)(vruntime - min_vruntime); 528 if (delta < 0) 529 min_vruntime = vruntime; 530 531 return min_vruntime; 532 } 533 534 static inline int entity_before(struct sched_entity *a, 535 struct sched_entity *b) 536 { 537 return (s64)(a->vruntime - b->vruntime) < 0; 538 } 539 540 static void update_min_vruntime(struct cfs_rq *cfs_rq) 541 { 542 struct sched_entity *curr = cfs_rq->curr; 543 struct rb_node *leftmost = rb_first_cached(&cfs_rq->tasks_timeline); 544 545 u64 vruntime = cfs_rq->min_vruntime; 546 547 if (curr) { 548 if (curr->on_rq) 549 vruntime = curr->vruntime; 550 else 551 curr = NULL; 552 } 553 554 if (leftmost) { /* non-empty tree */ 555 struct sched_entity *se; 556 se = rb_entry(leftmost, struct sched_entity, run_node); 557 558 if (!curr) 559 vruntime = se->vruntime; 560 else 561 vruntime = min_vruntime(vruntime, se->vruntime); 562 } 563 564 /* ensure we never gain time by being placed backwards. */ 565 cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime); 566 #ifndef CONFIG_64BIT 567 smp_wmb(); 568 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime; 569 #endif 570 } 571 572 /* 573 * Enqueue an entity into the rb-tree: 574 */ 575 static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se) 576 { 577 struct rb_node **link = &cfs_rq->tasks_timeline.rb_root.rb_node; 578 struct rb_node *parent = NULL; 579 struct sched_entity *entry; 580 bool leftmost = true; 581 582 /* 583 * Find the right place in the rbtree: 584 */ 585 while (*link) { 586 parent = *link; 587 entry = rb_entry(parent, struct sched_entity, run_node); 588 /* 589 * We dont care about collisions. Nodes with 590 * the same key stay together. 591 */ 592 if (entity_before(se, entry)) { 593 link = &parent->rb_left; 594 } else { 595 link = &parent->rb_right; 596 leftmost = false; 597 } 598 } 599 600 rb_link_node(&se->run_node, parent, link); 601 rb_insert_color_cached(&se->run_node, 602 &cfs_rq->tasks_timeline, leftmost); 603 } 604 605 static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se) 606 { 607 rb_erase_cached(&se->run_node, &cfs_rq->tasks_timeline); 608 } 609 610 struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq) 611 { 612 struct rb_node *left = rb_first_cached(&cfs_rq->tasks_timeline); 613 614 if (!left) 615 return NULL; 616 617 return rb_entry(left, struct sched_entity, run_node); 618 } 619 620 static struct sched_entity *__pick_next_entity(struct sched_entity *se) 621 { 622 struct rb_node *next = rb_next(&se->run_node); 623 624 if (!next) 625 return NULL; 626 627 return rb_entry(next, struct sched_entity, run_node); 628 } 629 630 #ifdef CONFIG_SCHED_DEBUG 631 struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq) 632 { 633 struct rb_node *last = rb_last(&cfs_rq->tasks_timeline.rb_root); 634 635 if (!last) 636 return NULL; 637 638 return rb_entry(last, struct sched_entity, run_node); 639 } 640 641 /************************************************************** 642 * Scheduling class statistics methods: 643 */ 644 645 int sched_proc_update_handler(struct ctl_table *table, int write, 646 void *buffer, size_t *lenp, loff_t *ppos) 647 { 648 int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos); 649 unsigned int factor = get_update_sysctl_factor(); 650 651 if (ret || !write) 652 return ret; 653 654 sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency, 655 sysctl_sched_min_granularity); 656 657 #define WRT_SYSCTL(name) \ 658 (normalized_sysctl_##name = sysctl_##name / (factor)) 659 WRT_SYSCTL(sched_min_granularity); 660 WRT_SYSCTL(sched_latency); 661 WRT_SYSCTL(sched_wakeup_granularity); 662 #undef WRT_SYSCTL 663 664 return 0; 665 } 666 #endif 667 668 /* 669 * delta /= w 670 */ 671 static inline u64 calc_delta_fair(u64 delta, struct sched_entity *se) 672 { 673 if (unlikely(se->load.weight != NICE_0_LOAD)) 674 delta = __calc_delta(delta, NICE_0_LOAD, &se->load); 675 676 return delta; 677 } 678 679 /* 680 * The idea is to set a period in which each task runs once. 681 * 682 * When there are too many tasks (sched_nr_latency) we have to stretch 683 * this period because otherwise the slices get too small. 684 * 685 * p = (nr <= nl) ? l : l*nr/nl 686 */ 687 static u64 __sched_period(unsigned long nr_running) 688 { 689 if (unlikely(nr_running > sched_nr_latency)) 690 return nr_running * sysctl_sched_min_granularity; 691 else 692 return sysctl_sched_latency; 693 } 694 695 /* 696 * We calculate the wall-time slice from the period by taking a part 697 * proportional to the weight. 698 * 699 * s = p*P[w/rw] 700 */ 701 static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se) 702 { 703 u64 slice = __sched_period(cfs_rq->nr_running + !se->on_rq); 704 705 for_each_sched_entity(se) { 706 struct load_weight *load; 707 struct load_weight lw; 708 709 cfs_rq = cfs_rq_of(se); 710 load = &cfs_rq->load; 711 712 if (unlikely(!se->on_rq)) { 713 lw = cfs_rq->load; 714 715 update_load_add(&lw, se->load.weight); 716 load = &lw; 717 } 718 slice = __calc_delta(slice, se->load.weight, load); 719 } 720 return slice; 721 } 722 723 /* 724 * We calculate the vruntime slice of a to-be-inserted task. 725 * 726 * vs = s/w 727 */ 728 static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se) 729 { 730 return calc_delta_fair(sched_slice(cfs_rq, se), se); 731 } 732 733 #include "pelt.h" 734 #ifdef CONFIG_SMP 735 736 static int select_idle_sibling(struct task_struct *p, int prev_cpu, int cpu); 737 static unsigned long task_h_load(struct task_struct *p); 738 static unsigned long capacity_of(int cpu); 739 740 /* Give new sched_entity start runnable values to heavy its load in infant time */ 741 void init_entity_runnable_average(struct sched_entity *se) 742 { 743 struct sched_avg *sa = &se->avg; 744 745 memset(sa, 0, sizeof(*sa)); 746 747 /* 748 * Tasks are initialized with full load to be seen as heavy tasks until 749 * they get a chance to stabilize to their real load level. 750 * Group entities are initialized with zero load to reflect the fact that 751 * nothing has been attached to the task group yet. 752 */ 753 if (entity_is_task(se)) 754 sa->load_avg = scale_load_down(se->load.weight); 755 756 /* when this task enqueue'ed, it will contribute to its cfs_rq's load_avg */ 757 } 758 759 static void attach_entity_cfs_rq(struct sched_entity *se); 760 761 /* 762 * With new tasks being created, their initial util_avgs are extrapolated 763 * based on the cfs_rq's current util_avg: 764 * 765 * util_avg = cfs_rq->util_avg / (cfs_rq->load_avg + 1) * se.load.weight 766 * 767 * However, in many cases, the above util_avg does not give a desired 768 * value. Moreover, the sum of the util_avgs may be divergent, such 769 * as when the series is a harmonic series. 770 * 771 * To solve this problem, we also cap the util_avg of successive tasks to 772 * only 1/2 of the left utilization budget: 773 * 774 * util_avg_cap = (cpu_scale - cfs_rq->avg.util_avg) / 2^n 775 * 776 * where n denotes the nth task and cpu_scale the CPU capacity. 777 * 778 * For example, for a CPU with 1024 of capacity, a simplest series from 779 * the beginning would be like: 780 * 781 * task util_avg: 512, 256, 128, 64, 32, 16, 8, ... 782 * cfs_rq util_avg: 512, 768, 896, 960, 992, 1008, 1016, ... 783 * 784 * Finally, that extrapolated util_avg is clamped to the cap (util_avg_cap) 785 * if util_avg > util_avg_cap. 786 */ 787 void post_init_entity_util_avg(struct task_struct *p) 788 { 789 struct sched_entity *se = &p->se; 790 struct cfs_rq *cfs_rq = cfs_rq_of(se); 791 struct sched_avg *sa = &se->avg; 792 long cpu_scale = arch_scale_cpu_capacity(cpu_of(rq_of(cfs_rq))); 793 long cap = (long)(cpu_scale - cfs_rq->avg.util_avg) / 2; 794 795 if (cap > 0) { 796 if (cfs_rq->avg.util_avg != 0) { 797 sa->util_avg = cfs_rq->avg.util_avg * se->load.weight; 798 sa->util_avg /= (cfs_rq->avg.load_avg + 1); 799 800 if (sa->util_avg > cap) 801 sa->util_avg = cap; 802 } else { 803 sa->util_avg = cap; 804 } 805 } 806 807 sa->runnable_avg = sa->util_avg; 808 809 if (p->sched_class != &fair_sched_class) { 810 /* 811 * For !fair tasks do: 812 * 813 update_cfs_rq_load_avg(now, cfs_rq); 814 attach_entity_load_avg(cfs_rq, se); 815 switched_from_fair(rq, p); 816 * 817 * such that the next switched_to_fair() has the 818 * expected state. 819 */ 820 se->avg.last_update_time = cfs_rq_clock_pelt(cfs_rq); 821 return; 822 } 823 824 attach_entity_cfs_rq(se); 825 } 826 827 #else /* !CONFIG_SMP */ 828 void init_entity_runnable_average(struct sched_entity *se) 829 { 830 } 831 void post_init_entity_util_avg(struct task_struct *p) 832 { 833 } 834 static void update_tg_load_avg(struct cfs_rq *cfs_rq) 835 { 836 } 837 #endif /* CONFIG_SMP */ 838 839 /* 840 * Update the current task's runtime statistics. 841 */ 842 static void update_curr(struct cfs_rq *cfs_rq) 843 { 844 struct sched_entity *curr = cfs_rq->curr; 845 u64 now = rq_clock_task(rq_of(cfs_rq)); 846 u64 delta_exec; 847 848 if (unlikely(!curr)) 849 return; 850 851 delta_exec = now - curr->exec_start; 852 if (unlikely((s64)delta_exec <= 0)) 853 return; 854 855 curr->exec_start = now; 856 857 schedstat_set(curr->statistics.exec_max, 858 max(delta_exec, curr->statistics.exec_max)); 859 860 curr->sum_exec_runtime += delta_exec; 861 schedstat_add(cfs_rq->exec_clock, delta_exec); 862 863 curr->vruntime += calc_delta_fair(delta_exec, curr); 864 update_min_vruntime(cfs_rq); 865 866 if (entity_is_task(curr)) { 867 struct task_struct *curtask = task_of(curr); 868 869 trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime); 870 cgroup_account_cputime(curtask, delta_exec); 871 account_group_exec_runtime(curtask, delta_exec); 872 } 873 874 account_cfs_rq_runtime(cfs_rq, delta_exec); 875 } 876 877 static void update_curr_fair(struct rq *rq) 878 { 879 update_curr(cfs_rq_of(&rq->curr->se)); 880 } 881 882 static inline void 883 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se) 884 { 885 u64 wait_start, prev_wait_start; 886 887 if (!schedstat_enabled()) 888 return; 889 890 wait_start = rq_clock(rq_of(cfs_rq)); 891 prev_wait_start = schedstat_val(se->statistics.wait_start); 892 893 if (entity_is_task(se) && task_on_rq_migrating(task_of(se)) && 894 likely(wait_start > prev_wait_start)) 895 wait_start -= prev_wait_start; 896 897 __schedstat_set(se->statistics.wait_start, wait_start); 898 } 899 900 static inline void 901 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se) 902 { 903 struct task_struct *p; 904 u64 delta; 905 906 if (!schedstat_enabled()) 907 return; 908 909 delta = rq_clock(rq_of(cfs_rq)) - schedstat_val(se->statistics.wait_start); 910 911 if (entity_is_task(se)) { 912 p = task_of(se); 913 if (task_on_rq_migrating(p)) { 914 /* 915 * Preserve migrating task's wait time so wait_start 916 * time stamp can be adjusted to accumulate wait time 917 * prior to migration. 918 */ 919 __schedstat_set(se->statistics.wait_start, delta); 920 return; 921 } 922 trace_sched_stat_wait(p, delta); 923 } 924 925 __schedstat_set(se->statistics.wait_max, 926 max(schedstat_val(se->statistics.wait_max), delta)); 927 __schedstat_inc(se->statistics.wait_count); 928 __schedstat_add(se->statistics.wait_sum, delta); 929 __schedstat_set(se->statistics.wait_start, 0); 930 } 931 932 static inline void 933 update_stats_enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se) 934 { 935 struct task_struct *tsk = NULL; 936 u64 sleep_start, block_start; 937 938 if (!schedstat_enabled()) 939 return; 940 941 sleep_start = schedstat_val(se->statistics.sleep_start); 942 block_start = schedstat_val(se->statistics.block_start); 943 944 if (entity_is_task(se)) 945 tsk = task_of(se); 946 947 if (sleep_start) { 948 u64 delta = rq_clock(rq_of(cfs_rq)) - sleep_start; 949 950 if ((s64)delta < 0) 951 delta = 0; 952 953 if (unlikely(delta > schedstat_val(se->statistics.sleep_max))) 954 __schedstat_set(se->statistics.sleep_max, delta); 955 956 __schedstat_set(se->statistics.sleep_start, 0); 957 __schedstat_add(se->statistics.sum_sleep_runtime, delta); 958 959 if (tsk) { 960 account_scheduler_latency(tsk, delta >> 10, 1); 961 trace_sched_stat_sleep(tsk, delta); 962 } 963 } 964 if (block_start) { 965 u64 delta = rq_clock(rq_of(cfs_rq)) - block_start; 966 967 if ((s64)delta < 0) 968 delta = 0; 969 970 if (unlikely(delta > schedstat_val(se->statistics.block_max))) 971 __schedstat_set(se->statistics.block_max, delta); 972 973 __schedstat_set(se->statistics.block_start, 0); 974 __schedstat_add(se->statistics.sum_sleep_runtime, delta); 975 976 if (tsk) { 977 if (tsk->in_iowait) { 978 __schedstat_add(se->statistics.iowait_sum, delta); 979 __schedstat_inc(se->statistics.iowait_count); 980 trace_sched_stat_iowait(tsk, delta); 981 } 982 983 trace_sched_stat_blocked(tsk, delta); 984 985 /* 986 * Blocking time is in units of nanosecs, so shift by 987 * 20 to get a milliseconds-range estimation of the 988 * amount of time that the task spent sleeping: 989 */ 990 if (unlikely(prof_on == SLEEP_PROFILING)) { 991 profile_hits(SLEEP_PROFILING, 992 (void *)get_wchan(tsk), 993 delta >> 20); 994 } 995 account_scheduler_latency(tsk, delta >> 10, 0); 996 } 997 } 998 } 999 1000 /* 1001 * Task is being enqueued - update stats: 1002 */ 1003 static inline void 1004 update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags) 1005 { 1006 if (!schedstat_enabled()) 1007 return; 1008 1009 /* 1010 * Are we enqueueing a waiting task? (for current tasks 1011 * a dequeue/enqueue event is a NOP) 1012 */ 1013 if (se != cfs_rq->curr) 1014 update_stats_wait_start(cfs_rq, se); 1015 1016 if (flags & ENQUEUE_WAKEUP) 1017 update_stats_enqueue_sleeper(cfs_rq, se); 1018 } 1019 1020 static inline void 1021 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags) 1022 { 1023 1024 if (!schedstat_enabled()) 1025 return; 1026 1027 /* 1028 * Mark the end of the wait period if dequeueing a 1029 * waiting task: 1030 */ 1031 if (se != cfs_rq->curr) 1032 update_stats_wait_end(cfs_rq, se); 1033 1034 if ((flags & DEQUEUE_SLEEP) && entity_is_task(se)) { 1035 struct task_struct *tsk = task_of(se); 1036 1037 if (tsk->state & TASK_INTERRUPTIBLE) 1038 __schedstat_set(se->statistics.sleep_start, 1039 rq_clock(rq_of(cfs_rq))); 1040 if (tsk->state & TASK_UNINTERRUPTIBLE) 1041 __schedstat_set(se->statistics.block_start, 1042 rq_clock(rq_of(cfs_rq))); 1043 } 1044 } 1045 1046 /* 1047 * We are picking a new current task - update its stats: 1048 */ 1049 static inline void 1050 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se) 1051 { 1052 /* 1053 * We are starting a new run period: 1054 */ 1055 se->exec_start = rq_clock_task(rq_of(cfs_rq)); 1056 } 1057 1058 /************************************************** 1059 * Scheduling class queueing methods: 1060 */ 1061 1062 #ifdef CONFIG_NUMA_BALANCING 1063 /* 1064 * Approximate time to scan a full NUMA task in ms. The task scan period is 1065 * calculated based on the tasks virtual memory size and 1066 * numa_balancing_scan_size. 1067 */ 1068 unsigned int sysctl_numa_balancing_scan_period_min = 1000; 1069 unsigned int sysctl_numa_balancing_scan_period_max = 60000; 1070 1071 /* Portion of address space to scan in MB */ 1072 unsigned int sysctl_numa_balancing_scan_size = 256; 1073 1074 /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */ 1075 unsigned int sysctl_numa_balancing_scan_delay = 1000; 1076 1077 struct numa_group { 1078 refcount_t refcount; 1079 1080 spinlock_t lock; /* nr_tasks, tasks */ 1081 int nr_tasks; 1082 pid_t gid; 1083 int active_nodes; 1084 1085 struct rcu_head rcu; 1086 unsigned long total_faults; 1087 unsigned long max_faults_cpu; 1088 /* 1089 * Faults_cpu is used to decide whether memory should move 1090 * towards the CPU. As a consequence, these stats are weighted 1091 * more by CPU use than by memory faults. 1092 */ 1093 unsigned long *faults_cpu; 1094 unsigned long faults[]; 1095 }; 1096 1097 /* 1098 * For functions that can be called in multiple contexts that permit reading 1099 * ->numa_group (see struct task_struct for locking rules). 1100 */ 1101 static struct numa_group *deref_task_numa_group(struct task_struct *p) 1102 { 1103 return rcu_dereference_check(p->numa_group, p == current || 1104 (lockdep_is_held(&task_rq(p)->lock) && !READ_ONCE(p->on_cpu))); 1105 } 1106 1107 static struct numa_group *deref_curr_numa_group(struct task_struct *p) 1108 { 1109 return rcu_dereference_protected(p->numa_group, p == current); 1110 } 1111 1112 static inline unsigned long group_faults_priv(struct numa_group *ng); 1113 static inline unsigned long group_faults_shared(struct numa_group *ng); 1114 1115 static unsigned int task_nr_scan_windows(struct task_struct *p) 1116 { 1117 unsigned long rss = 0; 1118 unsigned long nr_scan_pages; 1119 1120 /* 1121 * Calculations based on RSS as non-present and empty pages are skipped 1122 * by the PTE scanner and NUMA hinting faults should be trapped based 1123 * on resident pages 1124 */ 1125 nr_scan_pages = sysctl_numa_balancing_scan_size << (20 - PAGE_SHIFT); 1126 rss = get_mm_rss(p->mm); 1127 if (!rss) 1128 rss = nr_scan_pages; 1129 1130 rss = round_up(rss, nr_scan_pages); 1131 return rss / nr_scan_pages; 1132 } 1133 1134 /* For sanitys sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */ 1135 #define MAX_SCAN_WINDOW 2560 1136 1137 static unsigned int task_scan_min(struct task_struct *p) 1138 { 1139 unsigned int scan_size = READ_ONCE(sysctl_numa_balancing_scan_size); 1140 unsigned int scan, floor; 1141 unsigned int windows = 1; 1142 1143 if (scan_size < MAX_SCAN_WINDOW) 1144 windows = MAX_SCAN_WINDOW / scan_size; 1145 floor = 1000 / windows; 1146 1147 scan = sysctl_numa_balancing_scan_period_min / task_nr_scan_windows(p); 1148 return max_t(unsigned int, floor, scan); 1149 } 1150 1151 static unsigned int task_scan_start(struct task_struct *p) 1152 { 1153 unsigned long smin = task_scan_min(p); 1154 unsigned long period = smin; 1155 struct numa_group *ng; 1156 1157 /* Scale the maximum scan period with the amount of shared memory. */ 1158 rcu_read_lock(); 1159 ng = rcu_dereference(p->numa_group); 1160 if (ng) { 1161 unsigned long shared = group_faults_shared(ng); 1162 unsigned long private = group_faults_priv(ng); 1163 1164 period *= refcount_read(&ng->refcount); 1165 period *= shared + 1; 1166 period /= private + shared + 1; 1167 } 1168 rcu_read_unlock(); 1169 1170 return max(smin, period); 1171 } 1172 1173 static unsigned int task_scan_max(struct task_struct *p) 1174 { 1175 unsigned long smin = task_scan_min(p); 1176 unsigned long smax; 1177 struct numa_group *ng; 1178 1179 /* Watch for min being lower than max due to floor calculations */ 1180 smax = sysctl_numa_balancing_scan_period_max / task_nr_scan_windows(p); 1181 1182 /* Scale the maximum scan period with the amount of shared memory. */ 1183 ng = deref_curr_numa_group(p); 1184 if (ng) { 1185 unsigned long shared = group_faults_shared(ng); 1186 unsigned long private = group_faults_priv(ng); 1187 unsigned long period = smax; 1188 1189 period *= refcount_read(&ng->refcount); 1190 period *= shared + 1; 1191 period /= private + shared + 1; 1192 1193 smax = max(smax, period); 1194 } 1195 1196 return max(smin, smax); 1197 } 1198 1199 static void account_numa_enqueue(struct rq *rq, struct task_struct *p) 1200 { 1201 rq->nr_numa_running += (p->numa_preferred_nid != NUMA_NO_NODE); 1202 rq->nr_preferred_running += (p->numa_preferred_nid == task_node(p)); 1203 } 1204 1205 static void account_numa_dequeue(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 /* Shared or private faults. */ 1212 #define NR_NUMA_HINT_FAULT_TYPES 2 1213 1214 /* Memory and CPU locality */ 1215 #define NR_NUMA_HINT_FAULT_STATS (NR_NUMA_HINT_FAULT_TYPES * 2) 1216 1217 /* Averaged statistics, and temporary buffers. */ 1218 #define NR_NUMA_HINT_FAULT_BUCKETS (NR_NUMA_HINT_FAULT_STATS * 2) 1219 1220 pid_t task_numa_group_id(struct task_struct *p) 1221 { 1222 struct numa_group *ng; 1223 pid_t gid = 0; 1224 1225 rcu_read_lock(); 1226 ng = rcu_dereference(p->numa_group); 1227 if (ng) 1228 gid = ng->gid; 1229 rcu_read_unlock(); 1230 1231 return gid; 1232 } 1233 1234 /* 1235 * The averaged statistics, shared & private, memory & CPU, 1236 * occupy the first half of the array. The second half of the 1237 * array is for current counters, which are averaged into the 1238 * first set by task_numa_placement. 1239 */ 1240 static inline int task_faults_idx(enum numa_faults_stats s, int nid, int priv) 1241 { 1242 return NR_NUMA_HINT_FAULT_TYPES * (s * nr_node_ids + nid) + priv; 1243 } 1244 1245 static inline unsigned long task_faults(struct task_struct *p, int nid) 1246 { 1247 if (!p->numa_faults) 1248 return 0; 1249 1250 return p->numa_faults[task_faults_idx(NUMA_MEM, nid, 0)] + 1251 p->numa_faults[task_faults_idx(NUMA_MEM, nid, 1)]; 1252 } 1253 1254 static inline unsigned long group_faults(struct task_struct *p, int nid) 1255 { 1256 struct numa_group *ng = deref_task_numa_group(p); 1257 1258 if (!ng) 1259 return 0; 1260 1261 return ng->faults[task_faults_idx(NUMA_MEM, nid, 0)] + 1262 ng->faults[task_faults_idx(NUMA_MEM, nid, 1)]; 1263 } 1264 1265 static inline unsigned long group_faults_cpu(struct numa_group *group, int nid) 1266 { 1267 return group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 0)] + 1268 group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 1)]; 1269 } 1270 1271 static inline unsigned long group_faults_priv(struct numa_group *ng) 1272 { 1273 unsigned long faults = 0; 1274 int node; 1275 1276 for_each_online_node(node) { 1277 faults += ng->faults[task_faults_idx(NUMA_MEM, node, 1)]; 1278 } 1279 1280 return faults; 1281 } 1282 1283 static inline unsigned long group_faults_shared(struct numa_group *ng) 1284 { 1285 unsigned long faults = 0; 1286 int node; 1287 1288 for_each_online_node(node) { 1289 faults += ng->faults[task_faults_idx(NUMA_MEM, node, 0)]; 1290 } 1291 1292 return faults; 1293 } 1294 1295 /* 1296 * A node triggering more than 1/3 as many NUMA faults as the maximum is 1297 * considered part of a numa group's pseudo-interleaving set. Migrations 1298 * between these nodes are slowed down, to allow things to settle down. 1299 */ 1300 #define ACTIVE_NODE_FRACTION 3 1301 1302 static bool numa_is_active_node(int nid, struct numa_group *ng) 1303 { 1304 return group_faults_cpu(ng, nid) * ACTIVE_NODE_FRACTION > ng->max_faults_cpu; 1305 } 1306 1307 /* Handle placement on systems where not all nodes are directly connected. */ 1308 static unsigned long score_nearby_nodes(struct task_struct *p, int nid, 1309 int maxdist, bool task) 1310 { 1311 unsigned long score = 0; 1312 int node; 1313 1314 /* 1315 * All nodes are directly connected, and the same distance 1316 * from each other. No need for fancy placement algorithms. 1317 */ 1318 if (sched_numa_topology_type == NUMA_DIRECT) 1319 return 0; 1320 1321 /* 1322 * This code is called for each node, introducing N^2 complexity, 1323 * which should be ok given the number of nodes rarely exceeds 8. 1324 */ 1325 for_each_online_node(node) { 1326 unsigned long faults; 1327 int dist = node_distance(nid, node); 1328 1329 /* 1330 * The furthest away nodes in the system are not interesting 1331 * for placement; nid was already counted. 1332 */ 1333 if (dist == sched_max_numa_distance || node == nid) 1334 continue; 1335 1336 /* 1337 * On systems with a backplane NUMA topology, compare groups 1338 * of nodes, and move tasks towards the group with the most 1339 * memory accesses. When comparing two nodes at distance 1340 * "hoplimit", only nodes closer by than "hoplimit" are part 1341 * of each group. Skip other nodes. 1342 */ 1343 if (sched_numa_topology_type == NUMA_BACKPLANE && 1344 dist >= maxdist) 1345 continue; 1346 1347 /* Add up the faults from nearby nodes. */ 1348 if (task) 1349 faults = task_faults(p, node); 1350 else 1351 faults = group_faults(p, node); 1352 1353 /* 1354 * On systems with a glueless mesh NUMA topology, there are 1355 * no fixed "groups of nodes". Instead, nodes that are not 1356 * directly connected bounce traffic through intermediate 1357 * nodes; a numa_group can occupy any set of nodes. 1358 * The further away a node is, the less the faults count. 1359 * This seems to result in good task placement. 1360 */ 1361 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) { 1362 faults *= (sched_max_numa_distance - dist); 1363 faults /= (sched_max_numa_distance - LOCAL_DISTANCE); 1364 } 1365 1366 score += faults; 1367 } 1368 1369 return score; 1370 } 1371 1372 /* 1373 * These return the fraction of accesses done by a particular task, or 1374 * task group, on a particular numa node. The group weight is given a 1375 * larger multiplier, in order to group tasks together that are almost 1376 * evenly spread out between numa nodes. 1377 */ 1378 static inline unsigned long task_weight(struct task_struct *p, int nid, 1379 int dist) 1380 { 1381 unsigned long faults, total_faults; 1382 1383 if (!p->numa_faults) 1384 return 0; 1385 1386 total_faults = p->total_numa_faults; 1387 1388 if (!total_faults) 1389 return 0; 1390 1391 faults = task_faults(p, nid); 1392 faults += score_nearby_nodes(p, nid, dist, true); 1393 1394 return 1000 * faults / total_faults; 1395 } 1396 1397 static inline unsigned long group_weight(struct task_struct *p, int nid, 1398 int dist) 1399 { 1400 struct numa_group *ng = deref_task_numa_group(p); 1401 unsigned long faults, total_faults; 1402 1403 if (!ng) 1404 return 0; 1405 1406 total_faults = ng->total_faults; 1407 1408 if (!total_faults) 1409 return 0; 1410 1411 faults = group_faults(p, nid); 1412 faults += score_nearby_nodes(p, nid, dist, false); 1413 1414 return 1000 * faults / total_faults; 1415 } 1416 1417 bool should_numa_migrate_memory(struct task_struct *p, struct page * page, 1418 int src_nid, int dst_cpu) 1419 { 1420 struct numa_group *ng = deref_curr_numa_group(p); 1421 int dst_nid = cpu_to_node(dst_cpu); 1422 int last_cpupid, this_cpupid; 1423 1424 this_cpupid = cpu_pid_to_cpupid(dst_cpu, current->pid); 1425 last_cpupid = page_cpupid_xchg_last(page, this_cpupid); 1426 1427 /* 1428 * Allow first faults or private faults to migrate immediately early in 1429 * the lifetime of a task. The magic number 4 is based on waiting for 1430 * two full passes of the "multi-stage node selection" test that is 1431 * executed below. 1432 */ 1433 if ((p->numa_preferred_nid == NUMA_NO_NODE || p->numa_scan_seq <= 4) && 1434 (cpupid_pid_unset(last_cpupid) || cpupid_match_pid(p, last_cpupid))) 1435 return true; 1436 1437 /* 1438 * Multi-stage node selection is used in conjunction with a periodic 1439 * migration fault to build a temporal task<->page relation. By using 1440 * a two-stage filter we remove short/unlikely relations. 1441 * 1442 * Using P(p) ~ n_p / n_t as per frequentist probability, we can equate 1443 * a task's usage of a particular page (n_p) per total usage of this 1444 * page (n_t) (in a given time-span) to a probability. 1445 * 1446 * Our periodic faults will sample this probability and getting the 1447 * same result twice in a row, given these samples are fully 1448 * independent, is then given by P(n)^2, provided our sample period 1449 * is sufficiently short compared to the usage pattern. 1450 * 1451 * This quadric squishes small probabilities, making it less likely we 1452 * act on an unlikely task<->page relation. 1453 */ 1454 if (!cpupid_pid_unset(last_cpupid) && 1455 cpupid_to_nid(last_cpupid) != dst_nid) 1456 return false; 1457 1458 /* Always allow migrate on private faults */ 1459 if (cpupid_match_pid(p, last_cpupid)) 1460 return true; 1461 1462 /* A shared fault, but p->numa_group has not been set up yet. */ 1463 if (!ng) 1464 return true; 1465 1466 /* 1467 * Destination node is much more heavily used than the source 1468 * node? Allow migration. 1469 */ 1470 if (group_faults_cpu(ng, dst_nid) > group_faults_cpu(ng, src_nid) * 1471 ACTIVE_NODE_FRACTION) 1472 return true; 1473 1474 /* 1475 * Distribute memory according to CPU & memory use on each node, 1476 * with 3/4 hysteresis to avoid unnecessary memory migrations: 1477 * 1478 * faults_cpu(dst) 3 faults_cpu(src) 1479 * --------------- * - > --------------- 1480 * faults_mem(dst) 4 faults_mem(src) 1481 */ 1482 return group_faults_cpu(ng, dst_nid) * group_faults(p, src_nid) * 3 > 1483 group_faults_cpu(ng, src_nid) * group_faults(p, dst_nid) * 4; 1484 } 1485 1486 /* 1487 * 'numa_type' describes the node at the moment of load balancing. 1488 */ 1489 enum numa_type { 1490 /* The node has spare capacity that can be used to run more tasks. */ 1491 node_has_spare = 0, 1492 /* 1493 * The node is fully used and the tasks don't compete for more CPU 1494 * cycles. Nevertheless, some tasks might wait before running. 1495 */ 1496 node_fully_busy, 1497 /* 1498 * The node is overloaded and can't provide expected CPU cycles to all 1499 * tasks. 1500 */ 1501 node_overloaded 1502 }; 1503 1504 /* Cached statistics for all CPUs within a node */ 1505 struct numa_stats { 1506 unsigned long load; 1507 unsigned long runnable; 1508 unsigned long util; 1509 /* Total compute capacity of CPUs on a node */ 1510 unsigned long compute_capacity; 1511 unsigned int nr_running; 1512 unsigned int weight; 1513 enum numa_type node_type; 1514 int idle_cpu; 1515 }; 1516 1517 static inline bool is_core_idle(int cpu) 1518 { 1519 #ifdef CONFIG_SCHED_SMT 1520 int sibling; 1521 1522 for_each_cpu(sibling, cpu_smt_mask(cpu)) { 1523 if (cpu == sibling) 1524 continue; 1525 1526 if (!idle_cpu(cpu)) 1527 return false; 1528 } 1529 #endif 1530 1531 return true; 1532 } 1533 1534 struct task_numa_env { 1535 struct task_struct *p; 1536 1537 int src_cpu, src_nid; 1538 int dst_cpu, dst_nid; 1539 1540 struct numa_stats src_stats, dst_stats; 1541 1542 int imbalance_pct; 1543 int dist; 1544 1545 struct task_struct *best_task; 1546 long best_imp; 1547 int best_cpu; 1548 }; 1549 1550 static unsigned long cpu_load(struct rq *rq); 1551 static unsigned long cpu_runnable(struct rq *rq); 1552 static unsigned long cpu_util(int cpu); 1553 static inline long adjust_numa_imbalance(int imbalance, int nr_running); 1554 1555 static inline enum 1556 numa_type numa_classify(unsigned int imbalance_pct, 1557 struct numa_stats *ns) 1558 { 1559 if ((ns->nr_running > ns->weight) && 1560 (((ns->compute_capacity * 100) < (ns->util * imbalance_pct)) || 1561 ((ns->compute_capacity * imbalance_pct) < (ns->runnable * 100)))) 1562 return node_overloaded; 1563 1564 if ((ns->nr_running < ns->weight) || 1565 (((ns->compute_capacity * 100) > (ns->util * imbalance_pct)) && 1566 ((ns->compute_capacity * imbalance_pct) > (ns->runnable * 100)))) 1567 return node_has_spare; 1568 1569 return node_fully_busy; 1570 } 1571 1572 #ifdef CONFIG_SCHED_SMT 1573 /* Forward declarations of select_idle_sibling helpers */ 1574 static inline bool test_idle_cores(int cpu, bool def); 1575 static inline int numa_idle_core(int idle_core, int cpu) 1576 { 1577 if (!static_branch_likely(&sched_smt_present) || 1578 idle_core >= 0 || !test_idle_cores(cpu, false)) 1579 return idle_core; 1580 1581 /* 1582 * Prefer cores instead of packing HT siblings 1583 * and triggering future load balancing. 1584 */ 1585 if (is_core_idle(cpu)) 1586 idle_core = cpu; 1587 1588 return idle_core; 1589 } 1590 #else 1591 static inline int numa_idle_core(int idle_core, int cpu) 1592 { 1593 return idle_core; 1594 } 1595 #endif 1596 1597 /* 1598 * Gather all necessary information to make NUMA balancing placement 1599 * decisions that are compatible with standard load balancer. This 1600 * borrows code and logic from update_sg_lb_stats but sharing a 1601 * common implementation is impractical. 1602 */ 1603 static void update_numa_stats(struct task_numa_env *env, 1604 struct numa_stats *ns, int nid, 1605 bool find_idle) 1606 { 1607 int cpu, idle_core = -1; 1608 1609 memset(ns, 0, sizeof(*ns)); 1610 ns->idle_cpu = -1; 1611 1612 rcu_read_lock(); 1613 for_each_cpu(cpu, cpumask_of_node(nid)) { 1614 struct rq *rq = cpu_rq(cpu); 1615 1616 ns->load += cpu_load(rq); 1617 ns->runnable += cpu_runnable(rq); 1618 ns->util += cpu_util(cpu); 1619 ns->nr_running += rq->cfs.h_nr_running; 1620 ns->compute_capacity += capacity_of(cpu); 1621 1622 if (find_idle && !rq->nr_running && idle_cpu(cpu)) { 1623 if (READ_ONCE(rq->numa_migrate_on) || 1624 !cpumask_test_cpu(cpu, env->p->cpus_ptr)) 1625 continue; 1626 1627 if (ns->idle_cpu == -1) 1628 ns->idle_cpu = cpu; 1629 1630 idle_core = numa_idle_core(idle_core, cpu); 1631 } 1632 } 1633 rcu_read_unlock(); 1634 1635 ns->weight = cpumask_weight(cpumask_of_node(nid)); 1636 1637 ns->node_type = numa_classify(env->imbalance_pct, ns); 1638 1639 if (idle_core >= 0) 1640 ns->idle_cpu = idle_core; 1641 } 1642 1643 static void task_numa_assign(struct task_numa_env *env, 1644 struct task_struct *p, long imp) 1645 { 1646 struct rq *rq = cpu_rq(env->dst_cpu); 1647 1648 /* Check if run-queue part of active NUMA balance. */ 1649 if (env->best_cpu != env->dst_cpu && xchg(&rq->numa_migrate_on, 1)) { 1650 int cpu; 1651 int start = env->dst_cpu; 1652 1653 /* Find alternative idle CPU. */ 1654 for_each_cpu_wrap(cpu, cpumask_of_node(env->dst_nid), start) { 1655 if (cpu == env->best_cpu || !idle_cpu(cpu) || 1656 !cpumask_test_cpu(cpu, env->p->cpus_ptr)) { 1657 continue; 1658 } 1659 1660 env->dst_cpu = cpu; 1661 rq = cpu_rq(env->dst_cpu); 1662 if (!xchg(&rq->numa_migrate_on, 1)) 1663 goto assign; 1664 } 1665 1666 /* Failed to find an alternative idle CPU */ 1667 return; 1668 } 1669 1670 assign: 1671 /* 1672 * Clear previous best_cpu/rq numa-migrate flag, since task now 1673 * found a better CPU to move/swap. 1674 */ 1675 if (env->best_cpu != -1 && env->best_cpu != env->dst_cpu) { 1676 rq = cpu_rq(env->best_cpu); 1677 WRITE_ONCE(rq->numa_migrate_on, 0); 1678 } 1679 1680 if (env->best_task) 1681 put_task_struct(env->best_task); 1682 if (p) 1683 get_task_struct(p); 1684 1685 env->best_task = p; 1686 env->best_imp = imp; 1687 env->best_cpu = env->dst_cpu; 1688 } 1689 1690 static bool load_too_imbalanced(long src_load, long dst_load, 1691 struct task_numa_env *env) 1692 { 1693 long imb, old_imb; 1694 long orig_src_load, orig_dst_load; 1695 long src_capacity, dst_capacity; 1696 1697 /* 1698 * The load is corrected for the CPU capacity available on each node. 1699 * 1700 * src_load dst_load 1701 * ------------ vs --------- 1702 * src_capacity dst_capacity 1703 */ 1704 src_capacity = env->src_stats.compute_capacity; 1705 dst_capacity = env->dst_stats.compute_capacity; 1706 1707 imb = abs(dst_load * src_capacity - src_load * dst_capacity); 1708 1709 orig_src_load = env->src_stats.load; 1710 orig_dst_load = env->dst_stats.load; 1711 1712 old_imb = abs(orig_dst_load * src_capacity - orig_src_load * dst_capacity); 1713 1714 /* Would this change make things worse? */ 1715 return (imb > old_imb); 1716 } 1717 1718 /* 1719 * Maximum NUMA importance can be 1998 (2*999); 1720 * SMALLIMP @ 30 would be close to 1998/64. 1721 * Used to deter task migration. 1722 */ 1723 #define SMALLIMP 30 1724 1725 /* 1726 * This checks if the overall compute and NUMA accesses of the system would 1727 * be improved if the source tasks was migrated to the target dst_cpu taking 1728 * into account that it might be best if task running on the dst_cpu should 1729 * be exchanged with the source task 1730 */ 1731 static bool task_numa_compare(struct task_numa_env *env, 1732 long taskimp, long groupimp, bool maymove) 1733 { 1734 struct numa_group *cur_ng, *p_ng = deref_curr_numa_group(env->p); 1735 struct rq *dst_rq = cpu_rq(env->dst_cpu); 1736 long imp = p_ng ? groupimp : taskimp; 1737 struct task_struct *cur; 1738 long src_load, dst_load; 1739 int dist = env->dist; 1740 long moveimp = imp; 1741 long load; 1742 bool stopsearch = false; 1743 1744 if (READ_ONCE(dst_rq->numa_migrate_on)) 1745 return false; 1746 1747 rcu_read_lock(); 1748 cur = rcu_dereference(dst_rq->curr); 1749 if (cur && ((cur->flags & PF_EXITING) || is_idle_task(cur))) 1750 cur = NULL; 1751 1752 /* 1753 * Because we have preemption enabled we can get migrated around and 1754 * end try selecting ourselves (current == env->p) as a swap candidate. 1755 */ 1756 if (cur == env->p) { 1757 stopsearch = true; 1758 goto unlock; 1759 } 1760 1761 if (!cur) { 1762 if (maymove && moveimp >= env->best_imp) 1763 goto assign; 1764 else 1765 goto unlock; 1766 } 1767 1768 /* Skip this swap candidate if cannot move to the source cpu. */ 1769 if (!cpumask_test_cpu(env->src_cpu, cur->cpus_ptr)) 1770 goto unlock; 1771 1772 /* 1773 * Skip this swap candidate if it is not moving to its preferred 1774 * node and the best task is. 1775 */ 1776 if (env->best_task && 1777 env->best_task->numa_preferred_nid == env->src_nid && 1778 cur->numa_preferred_nid != env->src_nid) { 1779 goto unlock; 1780 } 1781 1782 /* 1783 * "imp" is the fault differential for the source task between the 1784 * source and destination node. Calculate the total differential for 1785 * the source task and potential destination task. The more negative 1786 * the value is, the more remote accesses that would be expected to 1787 * be incurred if the tasks were swapped. 1788 * 1789 * If dst and source tasks are in the same NUMA group, or not 1790 * in any group then look only at task weights. 1791 */ 1792 cur_ng = rcu_dereference(cur->numa_group); 1793 if (cur_ng == p_ng) { 1794 imp = taskimp + task_weight(cur, env->src_nid, dist) - 1795 task_weight(cur, env->dst_nid, dist); 1796 /* 1797 * Add some hysteresis to prevent swapping the 1798 * tasks within a group over tiny differences. 1799 */ 1800 if (cur_ng) 1801 imp -= imp / 16; 1802 } else { 1803 /* 1804 * Compare the group weights. If a task is all by itself 1805 * (not part of a group), use the task weight instead. 1806 */ 1807 if (cur_ng && p_ng) 1808 imp += group_weight(cur, env->src_nid, dist) - 1809 group_weight(cur, env->dst_nid, dist); 1810 else 1811 imp += task_weight(cur, env->src_nid, dist) - 1812 task_weight(cur, env->dst_nid, dist); 1813 } 1814 1815 /* Discourage picking a task already on its preferred node */ 1816 if (cur->numa_preferred_nid == env->dst_nid) 1817 imp -= imp / 16; 1818 1819 /* 1820 * Encourage picking a task that moves to its preferred node. 1821 * This potentially makes imp larger than it's maximum of 1822 * 1998 (see SMALLIMP and task_weight for why) but in this 1823 * case, it does not matter. 1824 */ 1825 if (cur->numa_preferred_nid == env->src_nid) 1826 imp += imp / 8; 1827 1828 if (maymove && moveimp > imp && moveimp > env->best_imp) { 1829 imp = moveimp; 1830 cur = NULL; 1831 goto assign; 1832 } 1833 1834 /* 1835 * Prefer swapping with a task moving to its preferred node over a 1836 * task that is not. 1837 */ 1838 if (env->best_task && cur->numa_preferred_nid == env->src_nid && 1839 env->best_task->numa_preferred_nid != env->src_nid) { 1840 goto assign; 1841 } 1842 1843 /* 1844 * If the NUMA importance is less than SMALLIMP, 1845 * task migration might only result in ping pong 1846 * of tasks and also hurt performance due to cache 1847 * misses. 1848 */ 1849 if (imp < SMALLIMP || imp <= env->best_imp + SMALLIMP / 2) 1850 goto unlock; 1851 1852 /* 1853 * In the overloaded case, try and keep the load balanced. 1854 */ 1855 load = task_h_load(env->p) - task_h_load(cur); 1856 if (!load) 1857 goto assign; 1858 1859 dst_load = env->dst_stats.load + load; 1860 src_load = env->src_stats.load - load; 1861 1862 if (load_too_imbalanced(src_load, dst_load, env)) 1863 goto unlock; 1864 1865 assign: 1866 /* Evaluate an idle CPU for a task numa move. */ 1867 if (!cur) { 1868 int cpu = env->dst_stats.idle_cpu; 1869 1870 /* Nothing cached so current CPU went idle since the search. */ 1871 if (cpu < 0) 1872 cpu = env->dst_cpu; 1873 1874 /* 1875 * If the CPU is no longer truly idle and the previous best CPU 1876 * is, keep using it. 1877 */ 1878 if (!idle_cpu(cpu) && env->best_cpu >= 0 && 1879 idle_cpu(env->best_cpu)) { 1880 cpu = env->best_cpu; 1881 } 1882 1883 env->dst_cpu = cpu; 1884 } 1885 1886 task_numa_assign(env, cur, imp); 1887 1888 /* 1889 * If a move to idle is allowed because there is capacity or load 1890 * balance improves then stop the search. While a better swap 1891 * candidate may exist, a search is not free. 1892 */ 1893 if (maymove && !cur && env->best_cpu >= 0 && idle_cpu(env->best_cpu)) 1894 stopsearch = true; 1895 1896 /* 1897 * If a swap candidate must be identified and the current best task 1898 * moves its preferred node then stop the search. 1899 */ 1900 if (!maymove && env->best_task && 1901 env->best_task->numa_preferred_nid == env->src_nid) { 1902 stopsearch = true; 1903 } 1904 unlock: 1905 rcu_read_unlock(); 1906 1907 return stopsearch; 1908 } 1909 1910 static void task_numa_find_cpu(struct task_numa_env *env, 1911 long taskimp, long groupimp) 1912 { 1913 bool maymove = false; 1914 int cpu; 1915 1916 /* 1917 * If dst node has spare capacity, then check if there is an 1918 * imbalance that would be overruled by the load balancer. 1919 */ 1920 if (env->dst_stats.node_type == node_has_spare) { 1921 unsigned int imbalance; 1922 int src_running, dst_running; 1923 1924 /* 1925 * Would movement cause an imbalance? Note that if src has 1926 * more running tasks that the imbalance is ignored as the 1927 * move improves the imbalance from the perspective of the 1928 * CPU load balancer. 1929 * */ 1930 src_running = env->src_stats.nr_running - 1; 1931 dst_running = env->dst_stats.nr_running + 1; 1932 imbalance = max(0, dst_running - src_running); 1933 imbalance = adjust_numa_imbalance(imbalance, dst_running); 1934 1935 /* Use idle CPU if there is no imbalance */ 1936 if (!imbalance) { 1937 maymove = true; 1938 if (env->dst_stats.idle_cpu >= 0) { 1939 env->dst_cpu = env->dst_stats.idle_cpu; 1940 task_numa_assign(env, NULL, 0); 1941 return; 1942 } 1943 } 1944 } else { 1945 long src_load, dst_load, load; 1946 /* 1947 * If the improvement from just moving env->p direction is better 1948 * than swapping tasks around, check if a move is possible. 1949 */ 1950 load = task_h_load(env->p); 1951 dst_load = env->dst_stats.load + load; 1952 src_load = env->src_stats.load - load; 1953 maymove = !load_too_imbalanced(src_load, dst_load, env); 1954 } 1955 1956 for_each_cpu(cpu, cpumask_of_node(env->dst_nid)) { 1957 /* Skip this CPU if the source task cannot migrate */ 1958 if (!cpumask_test_cpu(cpu, env->p->cpus_ptr)) 1959 continue; 1960 1961 env->dst_cpu = cpu; 1962 if (task_numa_compare(env, taskimp, groupimp, maymove)) 1963 break; 1964 } 1965 } 1966 1967 static int task_numa_migrate(struct task_struct *p) 1968 { 1969 struct task_numa_env env = { 1970 .p = p, 1971 1972 .src_cpu = task_cpu(p), 1973 .src_nid = task_node(p), 1974 1975 .imbalance_pct = 112, 1976 1977 .best_task = NULL, 1978 .best_imp = 0, 1979 .best_cpu = -1, 1980 }; 1981 unsigned long taskweight, groupweight; 1982 struct sched_domain *sd; 1983 long taskimp, groupimp; 1984 struct numa_group *ng; 1985 struct rq *best_rq; 1986 int nid, ret, dist; 1987 1988 /* 1989 * Pick the lowest SD_NUMA domain, as that would have the smallest 1990 * imbalance and would be the first to start moving tasks about. 1991 * 1992 * And we want to avoid any moving of tasks about, as that would create 1993 * random movement of tasks -- counter the numa conditions we're trying 1994 * to satisfy here. 1995 */ 1996 rcu_read_lock(); 1997 sd = rcu_dereference(per_cpu(sd_numa, env.src_cpu)); 1998 if (sd) 1999 env.imbalance_pct = 100 + (sd->imbalance_pct - 100) / 2; 2000 rcu_read_unlock(); 2001 2002 /* 2003 * Cpusets can break the scheduler domain tree into smaller 2004 * balance domains, some of which do not cross NUMA boundaries. 2005 * Tasks that are "trapped" in such domains cannot be migrated 2006 * elsewhere, so there is no point in (re)trying. 2007 */ 2008 if (unlikely(!sd)) { 2009 sched_setnuma(p, task_node(p)); 2010 return -EINVAL; 2011 } 2012 2013 env.dst_nid = p->numa_preferred_nid; 2014 dist = env.dist = node_distance(env.src_nid, env.dst_nid); 2015 taskweight = task_weight(p, env.src_nid, dist); 2016 groupweight = group_weight(p, env.src_nid, dist); 2017 update_numa_stats(&env, &env.src_stats, env.src_nid, false); 2018 taskimp = task_weight(p, env.dst_nid, dist) - taskweight; 2019 groupimp = group_weight(p, env.dst_nid, dist) - groupweight; 2020 update_numa_stats(&env, &env.dst_stats, env.dst_nid, true); 2021 2022 /* Try to find a spot on the preferred nid. */ 2023 task_numa_find_cpu(&env, taskimp, groupimp); 2024 2025 /* 2026 * Look at other nodes in these cases: 2027 * - there is no space available on the preferred_nid 2028 * - the task is part of a numa_group that is interleaved across 2029 * multiple NUMA nodes; in order to better consolidate the group, 2030 * we need to check other locations. 2031 */ 2032 ng = deref_curr_numa_group(p); 2033 if (env.best_cpu == -1 || (ng && ng->active_nodes > 1)) { 2034 for_each_online_node(nid) { 2035 if (nid == env.src_nid || nid == p->numa_preferred_nid) 2036 continue; 2037 2038 dist = node_distance(env.src_nid, env.dst_nid); 2039 if (sched_numa_topology_type == NUMA_BACKPLANE && 2040 dist != env.dist) { 2041 taskweight = task_weight(p, env.src_nid, dist); 2042 groupweight = group_weight(p, env.src_nid, dist); 2043 } 2044 2045 /* Only consider nodes where both task and groups benefit */ 2046 taskimp = task_weight(p, nid, dist) - taskweight; 2047 groupimp = group_weight(p, nid, dist) - groupweight; 2048 if (taskimp < 0 && groupimp < 0) 2049 continue; 2050 2051 env.dist = dist; 2052 env.dst_nid = nid; 2053 update_numa_stats(&env, &env.dst_stats, env.dst_nid, true); 2054 task_numa_find_cpu(&env, taskimp, groupimp); 2055 } 2056 } 2057 2058 /* 2059 * If the task is part of a workload that spans multiple NUMA nodes, 2060 * and is migrating into one of the workload's active nodes, remember 2061 * this node as the task's preferred numa node, so the workload can 2062 * settle down. 2063 * A task that migrated to a second choice node will be better off 2064 * trying for a better one later. Do not set the preferred node here. 2065 */ 2066 if (ng) { 2067 if (env.best_cpu == -1) 2068 nid = env.src_nid; 2069 else 2070 nid = cpu_to_node(env.best_cpu); 2071 2072 if (nid != p->numa_preferred_nid) 2073 sched_setnuma(p, nid); 2074 } 2075 2076 /* No better CPU than the current one was found. */ 2077 if (env.best_cpu == -1) { 2078 trace_sched_stick_numa(p, env.src_cpu, NULL, -1); 2079 return -EAGAIN; 2080 } 2081 2082 best_rq = cpu_rq(env.best_cpu); 2083 if (env.best_task == NULL) { 2084 ret = migrate_task_to(p, env.best_cpu); 2085 WRITE_ONCE(best_rq->numa_migrate_on, 0); 2086 if (ret != 0) 2087 trace_sched_stick_numa(p, env.src_cpu, NULL, env.best_cpu); 2088 return ret; 2089 } 2090 2091 ret = migrate_swap(p, env.best_task, env.best_cpu, env.src_cpu); 2092 WRITE_ONCE(best_rq->numa_migrate_on, 0); 2093 2094 if (ret != 0) 2095 trace_sched_stick_numa(p, env.src_cpu, env.best_task, env.best_cpu); 2096 put_task_struct(env.best_task); 2097 return ret; 2098 } 2099 2100 /* Attempt to migrate a task to a CPU on the preferred node. */ 2101 static void numa_migrate_preferred(struct task_struct *p) 2102 { 2103 unsigned long interval = HZ; 2104 2105 /* This task has no NUMA fault statistics yet */ 2106 if (unlikely(p->numa_preferred_nid == NUMA_NO_NODE || !p->numa_faults)) 2107 return; 2108 2109 /* Periodically retry migrating the task to the preferred node */ 2110 interval = min(interval, msecs_to_jiffies(p->numa_scan_period) / 16); 2111 p->numa_migrate_retry = jiffies + interval; 2112 2113 /* Success if task is already running on preferred CPU */ 2114 if (task_node(p) == p->numa_preferred_nid) 2115 return; 2116 2117 /* Otherwise, try migrate to a CPU on the preferred node */ 2118 task_numa_migrate(p); 2119 } 2120 2121 /* 2122 * Find out how many nodes on the workload is actively running on. Do this by 2123 * tracking the nodes from which NUMA hinting faults are triggered. This can 2124 * be different from the set of nodes where the workload's memory is currently 2125 * located. 2126 */ 2127 static void numa_group_count_active_nodes(struct numa_group *numa_group) 2128 { 2129 unsigned long faults, max_faults = 0; 2130 int nid, active_nodes = 0; 2131 2132 for_each_online_node(nid) { 2133 faults = group_faults_cpu(numa_group, nid); 2134 if (faults > max_faults) 2135 max_faults = faults; 2136 } 2137 2138 for_each_online_node(nid) { 2139 faults = group_faults_cpu(numa_group, nid); 2140 if (faults * ACTIVE_NODE_FRACTION > max_faults) 2141 active_nodes++; 2142 } 2143 2144 numa_group->max_faults_cpu = max_faults; 2145 numa_group->active_nodes = active_nodes; 2146 } 2147 2148 /* 2149 * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS 2150 * increments. The more local the fault statistics are, the higher the scan 2151 * period will be for the next scan window. If local/(local+remote) ratio is 2152 * below NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS) 2153 * the scan period will decrease. Aim for 70% local accesses. 2154 */ 2155 #define NUMA_PERIOD_SLOTS 10 2156 #define NUMA_PERIOD_THRESHOLD 7 2157 2158 /* 2159 * Increase the scan period (slow down scanning) if the majority of 2160 * our memory is already on our local node, or if the majority of 2161 * the page accesses are shared with other processes. 2162 * Otherwise, decrease the scan period. 2163 */ 2164 static void update_task_scan_period(struct task_struct *p, 2165 unsigned long shared, unsigned long private) 2166 { 2167 unsigned int period_slot; 2168 int lr_ratio, ps_ratio; 2169 int diff; 2170 2171 unsigned long remote = p->numa_faults_locality[0]; 2172 unsigned long local = p->numa_faults_locality[1]; 2173 2174 /* 2175 * If there were no record hinting faults then either the task is 2176 * completely idle or all activity is areas that are not of interest 2177 * to automatic numa balancing. Related to that, if there were failed 2178 * migration then it implies we are migrating too quickly or the local 2179 * node is overloaded. In either case, scan slower 2180 */ 2181 if (local + shared == 0 || p->numa_faults_locality[2]) { 2182 p->numa_scan_period = min(p->numa_scan_period_max, 2183 p->numa_scan_period << 1); 2184 2185 p->mm->numa_next_scan = jiffies + 2186 msecs_to_jiffies(p->numa_scan_period); 2187 2188 return; 2189 } 2190 2191 /* 2192 * Prepare to scale scan period relative to the current period. 2193 * == NUMA_PERIOD_THRESHOLD scan period stays the same 2194 * < NUMA_PERIOD_THRESHOLD scan period decreases (scan faster) 2195 * >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower) 2196 */ 2197 period_slot = DIV_ROUND_UP(p->numa_scan_period, NUMA_PERIOD_SLOTS); 2198 lr_ratio = (local * NUMA_PERIOD_SLOTS) / (local + remote); 2199 ps_ratio = (private * NUMA_PERIOD_SLOTS) / (private + shared); 2200 2201 if (ps_ratio >= NUMA_PERIOD_THRESHOLD) { 2202 /* 2203 * Most memory accesses are local. There is no need to 2204 * do fast NUMA scanning, since memory is already local. 2205 */ 2206 int slot = ps_ratio - NUMA_PERIOD_THRESHOLD; 2207 if (!slot) 2208 slot = 1; 2209 diff = slot * period_slot; 2210 } else if (lr_ratio >= NUMA_PERIOD_THRESHOLD) { 2211 /* 2212 * Most memory accesses are shared with other tasks. 2213 * There is no point in continuing fast NUMA scanning, 2214 * since other tasks may just move the memory elsewhere. 2215 */ 2216 int slot = lr_ratio - NUMA_PERIOD_THRESHOLD; 2217 if (!slot) 2218 slot = 1; 2219 diff = slot * period_slot; 2220 } else { 2221 /* 2222 * Private memory faults exceed (SLOTS-THRESHOLD)/SLOTS, 2223 * yet they are not on the local NUMA node. Speed up 2224 * NUMA scanning to get the memory moved over. 2225 */ 2226 int ratio = max(lr_ratio, ps_ratio); 2227 diff = -(NUMA_PERIOD_THRESHOLD - ratio) * period_slot; 2228 } 2229 2230 p->numa_scan_period = clamp(p->numa_scan_period + diff, 2231 task_scan_min(p), task_scan_max(p)); 2232 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality)); 2233 } 2234 2235 /* 2236 * Get the fraction of time the task has been running since the last 2237 * NUMA placement cycle. The scheduler keeps similar statistics, but 2238 * decays those on a 32ms period, which is orders of magnitude off 2239 * from the dozens-of-seconds NUMA balancing period. Use the scheduler 2240 * stats only if the task is so new there are no NUMA statistics yet. 2241 */ 2242 static u64 numa_get_avg_runtime(struct task_struct *p, u64 *period) 2243 { 2244 u64 runtime, delta, now; 2245 /* Use the start of this time slice to avoid calculations. */ 2246 now = p->se.exec_start; 2247 runtime = p->se.sum_exec_runtime; 2248 2249 if (p->last_task_numa_placement) { 2250 delta = runtime - p->last_sum_exec_runtime; 2251 *period = now - p->last_task_numa_placement; 2252 2253 /* Avoid time going backwards, prevent potential divide error: */ 2254 if (unlikely((s64)*period < 0)) 2255 *period = 0; 2256 } else { 2257 delta = p->se.avg.load_sum; 2258 *period = LOAD_AVG_MAX; 2259 } 2260 2261 p->last_sum_exec_runtime = runtime; 2262 p->last_task_numa_placement = now; 2263 2264 return delta; 2265 } 2266 2267 /* 2268 * Determine the preferred nid for a task in a numa_group. This needs to 2269 * be done in a way that produces consistent results with group_weight, 2270 * otherwise workloads might not converge. 2271 */ 2272 static int preferred_group_nid(struct task_struct *p, int nid) 2273 { 2274 nodemask_t nodes; 2275 int dist; 2276 2277 /* Direct connections between all NUMA nodes. */ 2278 if (sched_numa_topology_type == NUMA_DIRECT) 2279 return nid; 2280 2281 /* 2282 * On a system with glueless mesh NUMA topology, group_weight 2283 * scores nodes according to the number of NUMA hinting faults on 2284 * both the node itself, and on nearby nodes. 2285 */ 2286 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) { 2287 unsigned long score, max_score = 0; 2288 int node, max_node = nid; 2289 2290 dist = sched_max_numa_distance; 2291 2292 for_each_online_node(node) { 2293 score = group_weight(p, node, dist); 2294 if (score > max_score) { 2295 max_score = score; 2296 max_node = node; 2297 } 2298 } 2299 return max_node; 2300 } 2301 2302 /* 2303 * Finding the preferred nid in a system with NUMA backplane 2304 * interconnect topology is more involved. The goal is to locate 2305 * tasks from numa_groups near each other in the system, and 2306 * untangle workloads from different sides of the system. This requires 2307 * searching down the hierarchy of node groups, recursively searching 2308 * inside the highest scoring group of nodes. The nodemask tricks 2309 * keep the complexity of the search down. 2310 */ 2311 nodes = node_online_map; 2312 for (dist = sched_max_numa_distance; dist > LOCAL_DISTANCE; dist--) { 2313 unsigned long max_faults = 0; 2314 nodemask_t max_group = NODE_MASK_NONE; 2315 int a, b; 2316 2317 /* Are there nodes at this distance from each other? */ 2318 if (!find_numa_distance(dist)) 2319 continue; 2320 2321 for_each_node_mask(a, nodes) { 2322 unsigned long faults = 0; 2323 nodemask_t this_group; 2324 nodes_clear(this_group); 2325 2326 /* Sum group's NUMA faults; includes a==b case. */ 2327 for_each_node_mask(b, nodes) { 2328 if (node_distance(a, b) < dist) { 2329 faults += group_faults(p, b); 2330 node_set(b, this_group); 2331 node_clear(b, nodes); 2332 } 2333 } 2334 2335 /* Remember the top group. */ 2336 if (faults > max_faults) { 2337 max_faults = faults; 2338 max_group = this_group; 2339 /* 2340 * subtle: at the smallest distance there is 2341 * just one node left in each "group", the 2342 * winner is the preferred nid. 2343 */ 2344 nid = a; 2345 } 2346 } 2347 /* Next round, evaluate the nodes within max_group. */ 2348 if (!max_faults) 2349 break; 2350 nodes = max_group; 2351 } 2352 return nid; 2353 } 2354 2355 static void task_numa_placement(struct task_struct *p) 2356 { 2357 int seq, nid, max_nid = NUMA_NO_NODE; 2358 unsigned long max_faults = 0; 2359 unsigned long fault_types[2] = { 0, 0 }; 2360 unsigned long total_faults; 2361 u64 runtime, period; 2362 spinlock_t *group_lock = NULL; 2363 struct numa_group *ng; 2364 2365 /* 2366 * The p->mm->numa_scan_seq field gets updated without 2367 * exclusive access. Use READ_ONCE() here to ensure 2368 * that the field is read in a single access: 2369 */ 2370 seq = READ_ONCE(p->mm->numa_scan_seq); 2371 if (p->numa_scan_seq == seq) 2372 return; 2373 p->numa_scan_seq = seq; 2374 p->numa_scan_period_max = task_scan_max(p); 2375 2376 total_faults = p->numa_faults_locality[0] + 2377 p->numa_faults_locality[1]; 2378 runtime = numa_get_avg_runtime(p, &period); 2379 2380 /* If the task is part of a group prevent parallel updates to group stats */ 2381 ng = deref_curr_numa_group(p); 2382 if (ng) { 2383 group_lock = &ng->lock; 2384 spin_lock_irq(group_lock); 2385 } 2386 2387 /* Find the node with the highest number of faults */ 2388 for_each_online_node(nid) { 2389 /* Keep track of the offsets in numa_faults array */ 2390 int mem_idx, membuf_idx, cpu_idx, cpubuf_idx; 2391 unsigned long faults = 0, group_faults = 0; 2392 int priv; 2393 2394 for (priv = 0; priv < NR_NUMA_HINT_FAULT_TYPES; priv++) { 2395 long diff, f_diff, f_weight; 2396 2397 mem_idx = task_faults_idx(NUMA_MEM, nid, priv); 2398 membuf_idx = task_faults_idx(NUMA_MEMBUF, nid, priv); 2399 cpu_idx = task_faults_idx(NUMA_CPU, nid, priv); 2400 cpubuf_idx = task_faults_idx(NUMA_CPUBUF, nid, priv); 2401 2402 /* Decay existing window, copy faults since last scan */ 2403 diff = p->numa_faults[membuf_idx] - p->numa_faults[mem_idx] / 2; 2404 fault_types[priv] += p->numa_faults[membuf_idx]; 2405 p->numa_faults[membuf_idx] = 0; 2406 2407 /* 2408 * Normalize the faults_from, so all tasks in a group 2409 * count according to CPU use, instead of by the raw 2410 * number of faults. Tasks with little runtime have 2411 * little over-all impact on throughput, and thus their 2412 * faults are less important. 2413 */ 2414 f_weight = div64_u64(runtime << 16, period + 1); 2415 f_weight = (f_weight * p->numa_faults[cpubuf_idx]) / 2416 (total_faults + 1); 2417 f_diff = f_weight - p->numa_faults[cpu_idx] / 2; 2418 p->numa_faults[cpubuf_idx] = 0; 2419 2420 p->numa_faults[mem_idx] += diff; 2421 p->numa_faults[cpu_idx] += f_diff; 2422 faults += p->numa_faults[mem_idx]; 2423 p->total_numa_faults += diff; 2424 if (ng) { 2425 /* 2426 * safe because we can only change our own group 2427 * 2428 * mem_idx represents the offset for a given 2429 * nid and priv in a specific region because it 2430 * is at the beginning of the numa_faults array. 2431 */ 2432 ng->faults[mem_idx] += diff; 2433 ng->faults_cpu[mem_idx] += f_diff; 2434 ng->total_faults += diff; 2435 group_faults += ng->faults[mem_idx]; 2436 } 2437 } 2438 2439 if (!ng) { 2440 if (faults > max_faults) { 2441 max_faults = faults; 2442 max_nid = nid; 2443 } 2444 } else if (group_faults > max_faults) { 2445 max_faults = group_faults; 2446 max_nid = nid; 2447 } 2448 } 2449 2450 if (ng) { 2451 numa_group_count_active_nodes(ng); 2452 spin_unlock_irq(group_lock); 2453 max_nid = preferred_group_nid(p, max_nid); 2454 } 2455 2456 if (max_faults) { 2457 /* Set the new preferred node */ 2458 if (max_nid != p->numa_preferred_nid) 2459 sched_setnuma(p, max_nid); 2460 } 2461 2462 update_task_scan_period(p, fault_types[0], fault_types[1]); 2463 } 2464 2465 static inline int get_numa_group(struct numa_group *grp) 2466 { 2467 return refcount_inc_not_zero(&grp->refcount); 2468 } 2469 2470 static inline void put_numa_group(struct numa_group *grp) 2471 { 2472 if (refcount_dec_and_test(&grp->refcount)) 2473 kfree_rcu(grp, rcu); 2474 } 2475 2476 static void task_numa_group(struct task_struct *p, int cpupid, int flags, 2477 int *priv) 2478 { 2479 struct numa_group *grp, *my_grp; 2480 struct task_struct *tsk; 2481 bool join = false; 2482 int cpu = cpupid_to_cpu(cpupid); 2483 int i; 2484 2485 if (unlikely(!deref_curr_numa_group(p))) { 2486 unsigned int size = sizeof(struct numa_group) + 2487 4*nr_node_ids*sizeof(unsigned long); 2488 2489 grp = kzalloc(size, GFP_KERNEL | __GFP_NOWARN); 2490 if (!grp) 2491 return; 2492 2493 refcount_set(&grp->refcount, 1); 2494 grp->active_nodes = 1; 2495 grp->max_faults_cpu = 0; 2496 spin_lock_init(&grp->lock); 2497 grp->gid = p->pid; 2498 /* Second half of the array tracks nids where faults happen */ 2499 grp->faults_cpu = grp->faults + NR_NUMA_HINT_FAULT_TYPES * 2500 nr_node_ids; 2501 2502 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) 2503 grp->faults[i] = p->numa_faults[i]; 2504 2505 grp->total_faults = p->total_numa_faults; 2506 2507 grp->nr_tasks++; 2508 rcu_assign_pointer(p->numa_group, grp); 2509 } 2510 2511 rcu_read_lock(); 2512 tsk = READ_ONCE(cpu_rq(cpu)->curr); 2513 2514 if (!cpupid_match_pid(tsk, cpupid)) 2515 goto no_join; 2516 2517 grp = rcu_dereference(tsk->numa_group); 2518 if (!grp) 2519 goto no_join; 2520 2521 my_grp = deref_curr_numa_group(p); 2522 if (grp == my_grp) 2523 goto no_join; 2524 2525 /* 2526 * Only join the other group if its bigger; if we're the bigger group, 2527 * the other task will join us. 2528 */ 2529 if (my_grp->nr_tasks > grp->nr_tasks) 2530 goto no_join; 2531 2532 /* 2533 * Tie-break on the grp address. 2534 */ 2535 if (my_grp->nr_tasks == grp->nr_tasks && my_grp > grp) 2536 goto no_join; 2537 2538 /* Always join threads in the same process. */ 2539 if (tsk->mm == current->mm) 2540 join = true; 2541 2542 /* Simple filter to avoid false positives due to PID collisions */ 2543 if (flags & TNF_SHARED) 2544 join = true; 2545 2546 /* Update priv based on whether false sharing was detected */ 2547 *priv = !join; 2548 2549 if (join && !get_numa_group(grp)) 2550 goto no_join; 2551 2552 rcu_read_unlock(); 2553 2554 if (!join) 2555 return; 2556 2557 BUG_ON(irqs_disabled()); 2558 double_lock_irq(&my_grp->lock, &grp->lock); 2559 2560 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) { 2561 my_grp->faults[i] -= p->numa_faults[i]; 2562 grp->faults[i] += p->numa_faults[i]; 2563 } 2564 my_grp->total_faults -= p->total_numa_faults; 2565 grp->total_faults += p->total_numa_faults; 2566 2567 my_grp->nr_tasks--; 2568 grp->nr_tasks++; 2569 2570 spin_unlock(&my_grp->lock); 2571 spin_unlock_irq(&grp->lock); 2572 2573 rcu_assign_pointer(p->numa_group, grp); 2574 2575 put_numa_group(my_grp); 2576 return; 2577 2578 no_join: 2579 rcu_read_unlock(); 2580 return; 2581 } 2582 2583 /* 2584 * Get rid of NUMA staticstics associated with a task (either current or dead). 2585 * If @final is set, the task is dead and has reached refcount zero, so we can 2586 * safely free all relevant data structures. Otherwise, there might be 2587 * concurrent reads from places like load balancing and procfs, and we should 2588 * reset the data back to default state without freeing ->numa_faults. 2589 */ 2590 void task_numa_free(struct task_struct *p, bool final) 2591 { 2592 /* safe: p either is current or is being freed by current */ 2593 struct numa_group *grp = rcu_dereference_raw(p->numa_group); 2594 unsigned long *numa_faults = p->numa_faults; 2595 unsigned long flags; 2596 int i; 2597 2598 if (!numa_faults) 2599 return; 2600 2601 if (grp) { 2602 spin_lock_irqsave(&grp->lock, flags); 2603 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) 2604 grp->faults[i] -= p->numa_faults[i]; 2605 grp->total_faults -= p->total_numa_faults; 2606 2607 grp->nr_tasks--; 2608 spin_unlock_irqrestore(&grp->lock, flags); 2609 RCU_INIT_POINTER(p->numa_group, NULL); 2610 put_numa_group(grp); 2611 } 2612 2613 if (final) { 2614 p->numa_faults = NULL; 2615 kfree(numa_faults); 2616 } else { 2617 p->total_numa_faults = 0; 2618 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) 2619 numa_faults[i] = 0; 2620 } 2621 } 2622 2623 /* 2624 * Got a PROT_NONE fault for a page on @node. 2625 */ 2626 void task_numa_fault(int last_cpupid, int mem_node, int pages, int flags) 2627 { 2628 struct task_struct *p = current; 2629 bool migrated = flags & TNF_MIGRATED; 2630 int cpu_node = task_node(current); 2631 int local = !!(flags & TNF_FAULT_LOCAL); 2632 struct numa_group *ng; 2633 int priv; 2634 2635 if (!static_branch_likely(&sched_numa_balancing)) 2636 return; 2637 2638 /* for example, ksmd faulting in a user's mm */ 2639 if (!p->mm) 2640 return; 2641 2642 /* Allocate buffer to track faults on a per-node basis */ 2643 if (unlikely(!p->numa_faults)) { 2644 int size = sizeof(*p->numa_faults) * 2645 NR_NUMA_HINT_FAULT_BUCKETS * nr_node_ids; 2646 2647 p->numa_faults = kzalloc(size, GFP_KERNEL|__GFP_NOWARN); 2648 if (!p->numa_faults) 2649 return; 2650 2651 p->total_numa_faults = 0; 2652 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality)); 2653 } 2654 2655 /* 2656 * First accesses are treated as private, otherwise consider accesses 2657 * to be private if the accessing pid has not changed 2658 */ 2659 if (unlikely(last_cpupid == (-1 & LAST_CPUPID_MASK))) { 2660 priv = 1; 2661 } else { 2662 priv = cpupid_match_pid(p, last_cpupid); 2663 if (!priv && !(flags & TNF_NO_GROUP)) 2664 task_numa_group(p, last_cpupid, flags, &priv); 2665 } 2666 2667 /* 2668 * If a workload spans multiple NUMA nodes, a shared fault that 2669 * occurs wholly within the set of nodes that the workload is 2670 * actively using should be counted as local. This allows the 2671 * scan rate to slow down when a workload has settled down. 2672 */ 2673 ng = deref_curr_numa_group(p); 2674 if (!priv && !local && ng && ng->active_nodes > 1 && 2675 numa_is_active_node(cpu_node, ng) && 2676 numa_is_active_node(mem_node, ng)) 2677 local = 1; 2678 2679 /* 2680 * Retry to migrate task to preferred node periodically, in case it 2681 * previously failed, or the scheduler moved us. 2682 */ 2683 if (time_after(jiffies, p->numa_migrate_retry)) { 2684 task_numa_placement(p); 2685 numa_migrate_preferred(p); 2686 } 2687 2688 if (migrated) 2689 p->numa_pages_migrated += pages; 2690 if (flags & TNF_MIGRATE_FAIL) 2691 p->numa_faults_locality[2] += pages; 2692 2693 p->numa_faults[task_faults_idx(NUMA_MEMBUF, mem_node, priv)] += pages; 2694 p->numa_faults[task_faults_idx(NUMA_CPUBUF, cpu_node, priv)] += pages; 2695 p->numa_faults_locality[local] += pages; 2696 } 2697 2698 static void reset_ptenuma_scan(struct task_struct *p) 2699 { 2700 /* 2701 * We only did a read acquisition of the mmap sem, so 2702 * p->mm->numa_scan_seq is written to without exclusive access 2703 * and the update is not guaranteed to be atomic. That's not 2704 * much of an issue though, since this is just used for 2705 * statistical sampling. Use READ_ONCE/WRITE_ONCE, which are not 2706 * expensive, to avoid any form of compiler optimizations: 2707 */ 2708 WRITE_ONCE(p->mm->numa_scan_seq, READ_ONCE(p->mm->numa_scan_seq) + 1); 2709 p->mm->numa_scan_offset = 0; 2710 } 2711 2712 /* 2713 * The expensive part of numa migration is done from task_work context. 2714 * Triggered from task_tick_numa(). 2715 */ 2716 static void task_numa_work(struct callback_head *work) 2717 { 2718 unsigned long migrate, next_scan, now = jiffies; 2719 struct task_struct *p = current; 2720 struct mm_struct *mm = p->mm; 2721 u64 runtime = p->se.sum_exec_runtime; 2722 struct vm_area_struct *vma; 2723 unsigned long start, end; 2724 unsigned long nr_pte_updates = 0; 2725 long pages, virtpages; 2726 2727 SCHED_WARN_ON(p != container_of(work, struct task_struct, numa_work)); 2728 2729 work->next = work; 2730 /* 2731 * Who cares about NUMA placement when they're dying. 2732 * 2733 * NOTE: make sure not to dereference p->mm before this check, 2734 * exit_task_work() happens _after_ exit_mm() so we could be called 2735 * without p->mm even though we still had it when we enqueued this 2736 * work. 2737 */ 2738 if (p->flags & PF_EXITING) 2739 return; 2740 2741 if (!mm->numa_next_scan) { 2742 mm->numa_next_scan = now + 2743 msecs_to_jiffies(sysctl_numa_balancing_scan_delay); 2744 } 2745 2746 /* 2747 * Enforce maximal scan/migration frequency.. 2748 */ 2749 migrate = mm->numa_next_scan; 2750 if (time_before(now, migrate)) 2751 return; 2752 2753 if (p->numa_scan_period == 0) { 2754 p->numa_scan_period_max = task_scan_max(p); 2755 p->numa_scan_period = task_scan_start(p); 2756 } 2757 2758 next_scan = now + msecs_to_jiffies(p->numa_scan_period); 2759 if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate) 2760 return; 2761 2762 /* 2763 * Delay this task enough that another task of this mm will likely win 2764 * the next time around. 2765 */ 2766 p->node_stamp += 2 * TICK_NSEC; 2767 2768 start = mm->numa_scan_offset; 2769 pages = sysctl_numa_balancing_scan_size; 2770 pages <<= 20 - PAGE_SHIFT; /* MB in pages */ 2771 virtpages = pages * 8; /* Scan up to this much virtual space */ 2772 if (!pages) 2773 return; 2774 2775 2776 if (!mmap_read_trylock(mm)) 2777 return; 2778 vma = find_vma(mm, start); 2779 if (!vma) { 2780 reset_ptenuma_scan(p); 2781 start = 0; 2782 vma = mm->mmap; 2783 } 2784 for (; vma; vma = vma->vm_next) { 2785 if (!vma_migratable(vma) || !vma_policy_mof(vma) || 2786 is_vm_hugetlb_page(vma) || (vma->vm_flags & VM_MIXEDMAP)) { 2787 continue; 2788 } 2789 2790 /* 2791 * Shared library pages mapped by multiple processes are not 2792 * migrated as it is expected they are cache replicated. Avoid 2793 * hinting faults in read-only file-backed mappings or the vdso 2794 * as migrating the pages will be of marginal benefit. 2795 */ 2796 if (!vma->vm_mm || 2797 (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ))) 2798 continue; 2799 2800 /* 2801 * Skip inaccessible VMAs to avoid any confusion between 2802 * PROT_NONE and NUMA hinting ptes 2803 */ 2804 if (!vma_is_accessible(vma)) 2805 continue; 2806 2807 do { 2808 start = max(start, vma->vm_start); 2809 end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE); 2810 end = min(end, vma->vm_end); 2811 nr_pte_updates = change_prot_numa(vma, start, end); 2812 2813 /* 2814 * Try to scan sysctl_numa_balancing_size worth of 2815 * hpages that have at least one present PTE that 2816 * is not already pte-numa. If the VMA contains 2817 * areas that are unused or already full of prot_numa 2818 * PTEs, scan up to virtpages, to skip through those 2819 * areas faster. 2820 */ 2821 if (nr_pte_updates) 2822 pages -= (end - start) >> PAGE_SHIFT; 2823 virtpages -= (end - start) >> PAGE_SHIFT; 2824 2825 start = end; 2826 if (pages <= 0 || virtpages <= 0) 2827 goto out; 2828 2829 cond_resched(); 2830 } while (end != vma->vm_end); 2831 } 2832 2833 out: 2834 /* 2835 * It is possible to reach the end of the VMA list but the last few 2836 * VMAs are not guaranteed to the vma_migratable. If they are not, we 2837 * would find the !migratable VMA on the next scan but not reset the 2838 * scanner to the start so check it now. 2839 */ 2840 if (vma) 2841 mm->numa_scan_offset = start; 2842 else 2843 reset_ptenuma_scan(p); 2844 mmap_read_unlock(mm); 2845 2846 /* 2847 * Make sure tasks use at least 32x as much time to run other code 2848 * than they used here, to limit NUMA PTE scanning overhead to 3% max. 2849 * Usually update_task_scan_period slows down scanning enough; on an 2850 * overloaded system we need to limit overhead on a per task basis. 2851 */ 2852 if (unlikely(p->se.sum_exec_runtime != runtime)) { 2853 u64 diff = p->se.sum_exec_runtime - runtime; 2854 p->node_stamp += 32 * diff; 2855 } 2856 } 2857 2858 void init_numa_balancing(unsigned long clone_flags, struct task_struct *p) 2859 { 2860 int mm_users = 0; 2861 struct mm_struct *mm = p->mm; 2862 2863 if (mm) { 2864 mm_users = atomic_read(&mm->mm_users); 2865 if (mm_users == 1) { 2866 mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay); 2867 mm->numa_scan_seq = 0; 2868 } 2869 } 2870 p->node_stamp = 0; 2871 p->numa_scan_seq = mm ? mm->numa_scan_seq : 0; 2872 p->numa_scan_period = sysctl_numa_balancing_scan_delay; 2873 /* Protect against double add, see task_tick_numa and task_numa_work */ 2874 p->numa_work.next = &p->numa_work; 2875 p->numa_faults = NULL; 2876 RCU_INIT_POINTER(p->numa_group, NULL); 2877 p->last_task_numa_placement = 0; 2878 p->last_sum_exec_runtime = 0; 2879 2880 init_task_work(&p->numa_work, task_numa_work); 2881 2882 /* New address space, reset the preferred nid */ 2883 if (!(clone_flags & CLONE_VM)) { 2884 p->numa_preferred_nid = NUMA_NO_NODE; 2885 return; 2886 } 2887 2888 /* 2889 * New thread, keep existing numa_preferred_nid which should be copied 2890 * already by arch_dup_task_struct but stagger when scans start. 2891 */ 2892 if (mm) { 2893 unsigned int delay; 2894 2895 delay = min_t(unsigned int, task_scan_max(current), 2896 current->numa_scan_period * mm_users * NSEC_PER_MSEC); 2897 delay += 2 * TICK_NSEC; 2898 p->node_stamp = delay; 2899 } 2900 } 2901 2902 /* 2903 * Drive the periodic memory faults.. 2904 */ 2905 static void task_tick_numa(struct rq *rq, struct task_struct *curr) 2906 { 2907 struct callback_head *work = &curr->numa_work; 2908 u64 period, now; 2909 2910 /* 2911 * We don't care about NUMA placement if we don't have memory. 2912 */ 2913 if ((curr->flags & (PF_EXITING | PF_KTHREAD)) || work->next != work) 2914 return; 2915 2916 /* 2917 * Using runtime rather than walltime has the dual advantage that 2918 * we (mostly) drive the selection from busy threads and that the 2919 * task needs to have done some actual work before we bother with 2920 * NUMA placement. 2921 */ 2922 now = curr->se.sum_exec_runtime; 2923 period = (u64)curr->numa_scan_period * NSEC_PER_MSEC; 2924 2925 if (now > curr->node_stamp + period) { 2926 if (!curr->node_stamp) 2927 curr->numa_scan_period = task_scan_start(curr); 2928 curr->node_stamp += period; 2929 2930 if (!time_before(jiffies, curr->mm->numa_next_scan)) 2931 task_work_add(curr, work, TWA_RESUME); 2932 } 2933 } 2934 2935 static void update_scan_period(struct task_struct *p, int new_cpu) 2936 { 2937 int src_nid = cpu_to_node(task_cpu(p)); 2938 int dst_nid = cpu_to_node(new_cpu); 2939 2940 if (!static_branch_likely(&sched_numa_balancing)) 2941 return; 2942 2943 if (!p->mm || !p->numa_faults || (p->flags & PF_EXITING)) 2944 return; 2945 2946 if (src_nid == dst_nid) 2947 return; 2948 2949 /* 2950 * Allow resets if faults have been trapped before one scan 2951 * has completed. This is most likely due to a new task that 2952 * is pulled cross-node due to wakeups or load balancing. 2953 */ 2954 if (p->numa_scan_seq) { 2955 /* 2956 * Avoid scan adjustments if moving to the preferred 2957 * node or if the task was not previously running on 2958 * the preferred node. 2959 */ 2960 if (dst_nid == p->numa_preferred_nid || 2961 (p->numa_preferred_nid != NUMA_NO_NODE && 2962 src_nid != p->numa_preferred_nid)) 2963 return; 2964 } 2965 2966 p->numa_scan_period = task_scan_start(p); 2967 } 2968 2969 #else 2970 static void task_tick_numa(struct rq *rq, struct task_struct *curr) 2971 { 2972 } 2973 2974 static inline void account_numa_enqueue(struct rq *rq, struct task_struct *p) 2975 { 2976 } 2977 2978 static inline void account_numa_dequeue(struct rq *rq, struct task_struct *p) 2979 { 2980 } 2981 2982 static inline void update_scan_period(struct task_struct *p, int new_cpu) 2983 { 2984 } 2985 2986 #endif /* CONFIG_NUMA_BALANCING */ 2987 2988 static void 2989 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se) 2990 { 2991 update_load_add(&cfs_rq->load, se->load.weight); 2992 #ifdef CONFIG_SMP 2993 if (entity_is_task(se)) { 2994 struct rq *rq = rq_of(cfs_rq); 2995 2996 account_numa_enqueue(rq, task_of(se)); 2997 list_add(&se->group_node, &rq->cfs_tasks); 2998 } 2999 #endif 3000 cfs_rq->nr_running++; 3001 } 3002 3003 static void 3004 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se) 3005 { 3006 update_load_sub(&cfs_rq->load, se->load.weight); 3007 #ifdef CONFIG_SMP 3008 if (entity_is_task(se)) { 3009 account_numa_dequeue(rq_of(cfs_rq), task_of(se)); 3010 list_del_init(&se->group_node); 3011 } 3012 #endif 3013 cfs_rq->nr_running--; 3014 } 3015 3016 /* 3017 * Signed add and clamp on underflow. 3018 * 3019 * Explicitly do a load-store to ensure the intermediate value never hits 3020 * memory. This allows lockless observations without ever seeing the negative 3021 * values. 3022 */ 3023 #define add_positive(_ptr, _val) do { \ 3024 typeof(_ptr) ptr = (_ptr); \ 3025 typeof(_val) val = (_val); \ 3026 typeof(*ptr) res, var = READ_ONCE(*ptr); \ 3027 \ 3028 res = var + val; \ 3029 \ 3030 if (val < 0 && res > var) \ 3031 res = 0; \ 3032 \ 3033 WRITE_ONCE(*ptr, res); \ 3034 } while (0) 3035 3036 /* 3037 * Unsigned subtract and clamp on underflow. 3038 * 3039 * Explicitly do a load-store to ensure the intermediate value never hits 3040 * memory. This allows lockless observations without ever seeing the negative 3041 * values. 3042 */ 3043 #define sub_positive(_ptr, _val) do { \ 3044 typeof(_ptr) ptr = (_ptr); \ 3045 typeof(*ptr) val = (_val); \ 3046 typeof(*ptr) res, var = READ_ONCE(*ptr); \ 3047 res = var - val; \ 3048 if (res > var) \ 3049 res = 0; \ 3050 WRITE_ONCE(*ptr, res); \ 3051 } while (0) 3052 3053 /* 3054 * Remove and clamp on negative, from a local variable. 3055 * 3056 * A variant of sub_positive(), which does not use explicit load-store 3057 * and is thus optimized for local variable updates. 3058 */ 3059 #define lsub_positive(_ptr, _val) do { \ 3060 typeof(_ptr) ptr = (_ptr); \ 3061 *ptr -= min_t(typeof(*ptr), *ptr, _val); \ 3062 } while (0) 3063 3064 #ifdef CONFIG_SMP 3065 static inline void 3066 enqueue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) 3067 { 3068 cfs_rq->avg.load_avg += se->avg.load_avg; 3069 cfs_rq->avg.load_sum += se_weight(se) * se->avg.load_sum; 3070 } 3071 3072 static inline void 3073 dequeue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) 3074 { 3075 sub_positive(&cfs_rq->avg.load_avg, se->avg.load_avg); 3076 sub_positive(&cfs_rq->avg.load_sum, se_weight(se) * se->avg.load_sum); 3077 } 3078 #else 3079 static inline void 3080 enqueue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { } 3081 static inline void 3082 dequeue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { } 3083 #endif 3084 3085 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, 3086 unsigned long weight) 3087 { 3088 if (se->on_rq) { 3089 /* commit outstanding execution time */ 3090 if (cfs_rq->curr == se) 3091 update_curr(cfs_rq); 3092 update_load_sub(&cfs_rq->load, se->load.weight); 3093 } 3094 dequeue_load_avg(cfs_rq, se); 3095 3096 update_load_set(&se->load, weight); 3097 3098 #ifdef CONFIG_SMP 3099 do { 3100 u32 divider = get_pelt_divider(&se->avg); 3101 3102 se->avg.load_avg = div_u64(se_weight(se) * se->avg.load_sum, divider); 3103 } while (0); 3104 #endif 3105 3106 enqueue_load_avg(cfs_rq, se); 3107 if (se->on_rq) 3108 update_load_add(&cfs_rq->load, se->load.weight); 3109 3110 } 3111 3112 void reweight_task(struct task_struct *p, int prio) 3113 { 3114 struct sched_entity *se = &p->se; 3115 struct cfs_rq *cfs_rq = cfs_rq_of(se); 3116 struct load_weight *load = &se->load; 3117 unsigned long weight = scale_load(sched_prio_to_weight[prio]); 3118 3119 reweight_entity(cfs_rq, se, weight); 3120 load->inv_weight = sched_prio_to_wmult[prio]; 3121 } 3122 3123 #ifdef CONFIG_FAIR_GROUP_SCHED 3124 #ifdef CONFIG_SMP 3125 /* 3126 * All this does is approximate the hierarchical proportion which includes that 3127 * global sum we all love to hate. 3128 * 3129 * That is, the weight of a group entity, is the proportional share of the 3130 * group weight based on the group runqueue weights. That is: 3131 * 3132 * tg->weight * grq->load.weight 3133 * ge->load.weight = ----------------------------- (1) 3134 * \Sum grq->load.weight 3135 * 3136 * Now, because computing that sum is prohibitively expensive to compute (been 3137 * there, done that) we approximate it with this average stuff. The average 3138 * moves slower and therefore the approximation is cheaper and more stable. 3139 * 3140 * So instead of the above, we substitute: 3141 * 3142 * grq->load.weight -> grq->avg.load_avg (2) 3143 * 3144 * which yields the following: 3145 * 3146 * tg->weight * grq->avg.load_avg 3147 * ge->load.weight = ------------------------------ (3) 3148 * tg->load_avg 3149 * 3150 * Where: tg->load_avg ~= \Sum grq->avg.load_avg 3151 * 3152 * That is shares_avg, and it is right (given the approximation (2)). 3153 * 3154 * The problem with it is that because the average is slow -- it was designed 3155 * to be exactly that of course -- this leads to transients in boundary 3156 * conditions. In specific, the case where the group was idle and we start the 3157 * one task. It takes time for our CPU's grq->avg.load_avg to build up, 3158 * yielding bad latency etc.. 3159 * 3160 * Now, in that special case (1) reduces to: 3161 * 3162 * tg->weight * grq->load.weight 3163 * ge->load.weight = ----------------------------- = tg->weight (4) 3164 * grp->load.weight 3165 * 3166 * That is, the sum collapses because all other CPUs are idle; the UP scenario. 3167 * 3168 * So what we do is modify our approximation (3) to approach (4) in the (near) 3169 * UP case, like: 3170 * 3171 * ge->load.weight = 3172 * 3173 * tg->weight * grq->load.weight 3174 * --------------------------------------------------- (5) 3175 * tg->load_avg - grq->avg.load_avg + grq->load.weight 3176 * 3177 * But because grq->load.weight can drop to 0, resulting in a divide by zero, 3178 * we need to use grq->avg.load_avg as its lower bound, which then gives: 3179 * 3180 * 3181 * tg->weight * grq->load.weight 3182 * ge->load.weight = ----------------------------- (6) 3183 * tg_load_avg' 3184 * 3185 * Where: 3186 * 3187 * tg_load_avg' = tg->load_avg - grq->avg.load_avg + 3188 * max(grq->load.weight, grq->avg.load_avg) 3189 * 3190 * And that is shares_weight and is icky. In the (near) UP case it approaches 3191 * (4) while in the normal case it approaches (3). It consistently 3192 * overestimates the ge->load.weight and therefore: 3193 * 3194 * \Sum ge->load.weight >= tg->weight 3195 * 3196 * hence icky! 3197 */ 3198 static long calc_group_shares(struct cfs_rq *cfs_rq) 3199 { 3200 long tg_weight, tg_shares, load, shares; 3201 struct task_group *tg = cfs_rq->tg; 3202 3203 tg_shares = READ_ONCE(tg->shares); 3204 3205 load = max(scale_load_down(cfs_rq->load.weight), cfs_rq->avg.load_avg); 3206 3207 tg_weight = atomic_long_read(&tg->load_avg); 3208 3209 /* Ensure tg_weight >= load */ 3210 tg_weight -= cfs_rq->tg_load_avg_contrib; 3211 tg_weight += load; 3212 3213 shares = (tg_shares * load); 3214 if (tg_weight) 3215 shares /= tg_weight; 3216 3217 /* 3218 * MIN_SHARES has to be unscaled here to support per-CPU partitioning 3219 * of a group with small tg->shares value. It is a floor value which is 3220 * assigned as a minimum load.weight to the sched_entity representing 3221 * the group on a CPU. 3222 * 3223 * E.g. on 64-bit for a group with tg->shares of scale_load(15)=15*1024 3224 * on an 8-core system with 8 tasks each runnable on one CPU shares has 3225 * to be 15*1024*1/8=1920 instead of scale_load(MIN_SHARES)=2*1024. In 3226 * case no task is runnable on a CPU MIN_SHARES=2 should be returned 3227 * instead of 0. 3228 */ 3229 return clamp_t(long, shares, MIN_SHARES, tg_shares); 3230 } 3231 #endif /* CONFIG_SMP */ 3232 3233 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq); 3234 3235 /* 3236 * Recomputes the group entity based on the current state of its group 3237 * runqueue. 3238 */ 3239 static void update_cfs_group(struct sched_entity *se) 3240 { 3241 struct cfs_rq *gcfs_rq = group_cfs_rq(se); 3242 long shares; 3243 3244 if (!gcfs_rq) 3245 return; 3246 3247 if (throttled_hierarchy(gcfs_rq)) 3248 return; 3249 3250 #ifndef CONFIG_SMP 3251 shares = READ_ONCE(gcfs_rq->tg->shares); 3252 3253 if (likely(se->load.weight == shares)) 3254 return; 3255 #else 3256 shares = calc_group_shares(gcfs_rq); 3257 #endif 3258 3259 reweight_entity(cfs_rq_of(se), se, shares); 3260 } 3261 3262 #else /* CONFIG_FAIR_GROUP_SCHED */ 3263 static inline void update_cfs_group(struct sched_entity *se) 3264 { 3265 } 3266 #endif /* CONFIG_FAIR_GROUP_SCHED */ 3267 3268 static inline void cfs_rq_util_change(struct cfs_rq *cfs_rq, int flags) 3269 { 3270 struct rq *rq = rq_of(cfs_rq); 3271 3272 if (&rq->cfs == cfs_rq) { 3273 /* 3274 * There are a few boundary cases this might miss but it should 3275 * get called often enough that that should (hopefully) not be 3276 * a real problem. 3277 * 3278 * It will not get called when we go idle, because the idle 3279 * thread is a different class (!fair), nor will the utilization 3280 * number include things like RT tasks. 3281 * 3282 * As is, the util number is not freq-invariant (we'd have to 3283 * implement arch_scale_freq_capacity() for that). 3284 * 3285 * See cpu_util(). 3286 */ 3287 cpufreq_update_util(rq, flags); 3288 } 3289 } 3290 3291 #ifdef CONFIG_SMP 3292 #ifdef CONFIG_FAIR_GROUP_SCHED 3293 /** 3294 * update_tg_load_avg - update the tg's load avg 3295 * @cfs_rq: the cfs_rq whose avg changed 3296 * 3297 * This function 'ensures': tg->load_avg := \Sum tg->cfs_rq[]->avg.load. 3298 * However, because tg->load_avg is a global value there are performance 3299 * considerations. 3300 * 3301 * In order to avoid having to look at the other cfs_rq's, we use a 3302 * differential update where we store the last value we propagated. This in 3303 * turn allows skipping updates if the differential is 'small'. 3304 * 3305 * Updating tg's load_avg is necessary before update_cfs_share(). 3306 */ 3307 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq) 3308 { 3309 long delta = cfs_rq->avg.load_avg - cfs_rq->tg_load_avg_contrib; 3310 3311 /* 3312 * No need to update load_avg for root_task_group as it is not used. 3313 */ 3314 if (cfs_rq->tg == &root_task_group) 3315 return; 3316 3317 if (abs(delta) > cfs_rq->tg_load_avg_contrib / 64) { 3318 atomic_long_add(delta, &cfs_rq->tg->load_avg); 3319 cfs_rq->tg_load_avg_contrib = cfs_rq->avg.load_avg; 3320 } 3321 } 3322 3323 /* 3324 * Called within set_task_rq() right before setting a task's CPU. The 3325 * caller only guarantees p->pi_lock is held; no other assumptions, 3326 * including the state of rq->lock, should be made. 3327 */ 3328 void set_task_rq_fair(struct sched_entity *se, 3329 struct cfs_rq *prev, struct cfs_rq *next) 3330 { 3331 u64 p_last_update_time; 3332 u64 n_last_update_time; 3333 3334 if (!sched_feat(ATTACH_AGE_LOAD)) 3335 return; 3336 3337 /* 3338 * We are supposed to update the task to "current" time, then its up to 3339 * date and ready to go to new CPU/cfs_rq. But we have difficulty in 3340 * getting what current time is, so simply throw away the out-of-date 3341 * time. This will result in the wakee task is less decayed, but giving 3342 * the wakee more load sounds not bad. 3343 */ 3344 if (!(se->avg.last_update_time && prev)) 3345 return; 3346 3347 #ifndef CONFIG_64BIT 3348 { 3349 u64 p_last_update_time_copy; 3350 u64 n_last_update_time_copy; 3351 3352 do { 3353 p_last_update_time_copy = prev->load_last_update_time_copy; 3354 n_last_update_time_copy = next->load_last_update_time_copy; 3355 3356 smp_rmb(); 3357 3358 p_last_update_time = prev->avg.last_update_time; 3359 n_last_update_time = next->avg.last_update_time; 3360 3361 } while (p_last_update_time != p_last_update_time_copy || 3362 n_last_update_time != n_last_update_time_copy); 3363 } 3364 #else 3365 p_last_update_time = prev->avg.last_update_time; 3366 n_last_update_time = next->avg.last_update_time; 3367 #endif 3368 __update_load_avg_blocked_se(p_last_update_time, se); 3369 se->avg.last_update_time = n_last_update_time; 3370 } 3371 3372 3373 /* 3374 * When on migration a sched_entity joins/leaves the PELT hierarchy, we need to 3375 * propagate its contribution. The key to this propagation is the invariant 3376 * that for each group: 3377 * 3378 * ge->avg == grq->avg (1) 3379 * 3380 * _IFF_ we look at the pure running and runnable sums. Because they 3381 * represent the very same entity, just at different points in the hierarchy. 3382 * 3383 * Per the above update_tg_cfs_util() and update_tg_cfs_runnable() are trivial 3384 * and simply copies the running/runnable sum over (but still wrong, because 3385 * the group entity and group rq do not have their PELT windows aligned). 3386 * 3387 * However, update_tg_cfs_load() is more complex. So we have: 3388 * 3389 * ge->avg.load_avg = ge->load.weight * ge->avg.runnable_avg (2) 3390 * 3391 * And since, like util, the runnable part should be directly transferable, 3392 * the following would _appear_ to be the straight forward approach: 3393 * 3394 * grq->avg.load_avg = grq->load.weight * grq->avg.runnable_avg (3) 3395 * 3396 * And per (1) we have: 3397 * 3398 * ge->avg.runnable_avg == grq->avg.runnable_avg 3399 * 3400 * Which gives: 3401 * 3402 * ge->load.weight * grq->avg.load_avg 3403 * ge->avg.load_avg = ----------------------------------- (4) 3404 * grq->load.weight 3405 * 3406 * Except that is wrong! 3407 * 3408 * Because while for entities historical weight is not important and we 3409 * really only care about our future and therefore can consider a pure 3410 * runnable sum, runqueues can NOT do this. 3411 * 3412 * We specifically want runqueues to have a load_avg that includes 3413 * historical weights. Those represent the blocked load, the load we expect 3414 * to (shortly) return to us. This only works by keeping the weights as 3415 * integral part of the sum. We therefore cannot decompose as per (3). 3416 * 3417 * Another reason this doesn't work is that runnable isn't a 0-sum entity. 3418 * Imagine a rq with 2 tasks that each are runnable 2/3 of the time. Then the 3419 * rq itself is runnable anywhere between 2/3 and 1 depending on how the 3420 * runnable section of these tasks overlap (or not). If they were to perfectly 3421 * align the rq as a whole would be runnable 2/3 of the time. If however we 3422 * always have at least 1 runnable task, the rq as a whole is always runnable. 3423 * 3424 * So we'll have to approximate.. :/ 3425 * 3426 * Given the constraint: 3427 * 3428 * ge->avg.running_sum <= ge->avg.runnable_sum <= LOAD_AVG_MAX 3429 * 3430 * We can construct a rule that adds runnable to a rq by assuming minimal 3431 * overlap. 3432 * 3433 * On removal, we'll assume each task is equally runnable; which yields: 3434 * 3435 * grq->avg.runnable_sum = grq->avg.load_sum / grq->load.weight 3436 * 3437 * XXX: only do this for the part of runnable > running ? 3438 * 3439 */ 3440 3441 static inline void 3442 update_tg_cfs_util(struct cfs_rq *cfs_rq, struct sched_entity *se, struct cfs_rq *gcfs_rq) 3443 { 3444 long delta = gcfs_rq->avg.util_avg - se->avg.util_avg; 3445 u32 divider; 3446 3447 /* Nothing to update */ 3448 if (!delta) 3449 return; 3450 3451 /* 3452 * cfs_rq->avg.period_contrib can be used for both cfs_rq and se. 3453 * See ___update_load_avg() for details. 3454 */ 3455 divider = get_pelt_divider(&cfs_rq->avg); 3456 3457 /* Set new sched_entity's utilization */ 3458 se->avg.util_avg = gcfs_rq->avg.util_avg; 3459 se->avg.util_sum = se->avg.util_avg * divider; 3460 3461 /* Update parent cfs_rq utilization */ 3462 add_positive(&cfs_rq->avg.util_avg, delta); 3463 cfs_rq->avg.util_sum = cfs_rq->avg.util_avg * divider; 3464 } 3465 3466 static inline void 3467 update_tg_cfs_runnable(struct cfs_rq *cfs_rq, struct sched_entity *se, struct cfs_rq *gcfs_rq) 3468 { 3469 long delta = gcfs_rq->avg.runnable_avg - se->avg.runnable_avg; 3470 u32 divider; 3471 3472 /* Nothing to update */ 3473 if (!delta) 3474 return; 3475 3476 /* 3477 * cfs_rq->avg.period_contrib can be used for both cfs_rq and se. 3478 * See ___update_load_avg() for details. 3479 */ 3480 divider = get_pelt_divider(&cfs_rq->avg); 3481 3482 /* Set new sched_entity's runnable */ 3483 se->avg.runnable_avg = gcfs_rq->avg.runnable_avg; 3484 se->avg.runnable_sum = se->avg.runnable_avg * divider; 3485 3486 /* Update parent cfs_rq runnable */ 3487 add_positive(&cfs_rq->avg.runnable_avg, delta); 3488 cfs_rq->avg.runnable_sum = cfs_rq->avg.runnable_avg * divider; 3489 } 3490 3491 static inline void 3492 update_tg_cfs_load(struct cfs_rq *cfs_rq, struct sched_entity *se, struct cfs_rq *gcfs_rq) 3493 { 3494 long delta_avg, running_sum, runnable_sum = gcfs_rq->prop_runnable_sum; 3495 unsigned long load_avg; 3496 u64 load_sum = 0; 3497 s64 delta_sum; 3498 u32 divider; 3499 3500 if (!runnable_sum) 3501 return; 3502 3503 gcfs_rq->prop_runnable_sum = 0; 3504 3505 /* 3506 * cfs_rq->avg.period_contrib can be used for both cfs_rq and se. 3507 * See ___update_load_avg() for details. 3508 */ 3509 divider = get_pelt_divider(&cfs_rq->avg); 3510 3511 if (runnable_sum >= 0) { 3512 /* 3513 * Add runnable; clip at LOAD_AVG_MAX. Reflects that until 3514 * the CPU is saturated running == runnable. 3515 */ 3516 runnable_sum += se->avg.load_sum; 3517 runnable_sum = min_t(long, runnable_sum, divider); 3518 } else { 3519 /* 3520 * Estimate the new unweighted runnable_sum of the gcfs_rq by 3521 * assuming all tasks are equally runnable. 3522 */ 3523 if (scale_load_down(gcfs_rq->load.weight)) { 3524 load_sum = div_s64(gcfs_rq->avg.load_sum, 3525 scale_load_down(gcfs_rq->load.weight)); 3526 } 3527 3528 /* But make sure to not inflate se's runnable */ 3529 runnable_sum = min(se->avg.load_sum, load_sum); 3530 } 3531 3532 /* 3533 * runnable_sum can't be lower than running_sum 3534 * Rescale running sum to be in the same range as runnable sum 3535 * running_sum is in [0 : LOAD_AVG_MAX << SCHED_CAPACITY_SHIFT] 3536 * runnable_sum is in [0 : LOAD_AVG_MAX] 3537 */ 3538 running_sum = se->avg.util_sum >> SCHED_CAPACITY_SHIFT; 3539 runnable_sum = max(runnable_sum, running_sum); 3540 3541 load_sum = (s64)se_weight(se) * runnable_sum; 3542 load_avg = div_s64(load_sum, divider); 3543 3544 delta_sum = load_sum - (s64)se_weight(se) * se->avg.load_sum; 3545 delta_avg = load_avg - se->avg.load_avg; 3546 3547 se->avg.load_sum = runnable_sum; 3548 se->avg.load_avg = load_avg; 3549 add_positive(&cfs_rq->avg.load_avg, delta_avg); 3550 add_positive(&cfs_rq->avg.load_sum, delta_sum); 3551 } 3552 3553 static inline void add_tg_cfs_propagate(struct cfs_rq *cfs_rq, long runnable_sum) 3554 { 3555 cfs_rq->propagate = 1; 3556 cfs_rq->prop_runnable_sum += runnable_sum; 3557 } 3558 3559 /* Update task and its cfs_rq load average */ 3560 static inline int propagate_entity_load_avg(struct sched_entity *se) 3561 { 3562 struct cfs_rq *cfs_rq, *gcfs_rq; 3563 3564 if (entity_is_task(se)) 3565 return 0; 3566 3567 gcfs_rq = group_cfs_rq(se); 3568 if (!gcfs_rq->propagate) 3569 return 0; 3570 3571 gcfs_rq->propagate = 0; 3572 3573 cfs_rq = cfs_rq_of(se); 3574 3575 add_tg_cfs_propagate(cfs_rq, gcfs_rq->prop_runnable_sum); 3576 3577 update_tg_cfs_util(cfs_rq, se, gcfs_rq); 3578 update_tg_cfs_runnable(cfs_rq, se, gcfs_rq); 3579 update_tg_cfs_load(cfs_rq, se, gcfs_rq); 3580 3581 trace_pelt_cfs_tp(cfs_rq); 3582 trace_pelt_se_tp(se); 3583 3584 return 1; 3585 } 3586 3587 /* 3588 * Check if we need to update the load and the utilization of a blocked 3589 * group_entity: 3590 */ 3591 static inline bool skip_blocked_update(struct sched_entity *se) 3592 { 3593 struct cfs_rq *gcfs_rq = group_cfs_rq(se); 3594 3595 /* 3596 * If sched_entity still have not zero load or utilization, we have to 3597 * decay it: 3598 */ 3599 if (se->avg.load_avg || se->avg.util_avg) 3600 return false; 3601 3602 /* 3603 * If there is a pending propagation, we have to update the load and 3604 * the utilization of the sched_entity: 3605 */ 3606 if (gcfs_rq->propagate) 3607 return false; 3608 3609 /* 3610 * Otherwise, the load and the utilization of the sched_entity is 3611 * already zero and there is no pending propagation, so it will be a 3612 * waste of time to try to decay it: 3613 */ 3614 return true; 3615 } 3616 3617 #else /* CONFIG_FAIR_GROUP_SCHED */ 3618 3619 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq) {} 3620 3621 static inline int propagate_entity_load_avg(struct sched_entity *se) 3622 { 3623 return 0; 3624 } 3625 3626 static inline void add_tg_cfs_propagate(struct cfs_rq *cfs_rq, long runnable_sum) {} 3627 3628 #endif /* CONFIG_FAIR_GROUP_SCHED */ 3629 3630 /** 3631 * update_cfs_rq_load_avg - update the cfs_rq's load/util averages 3632 * @now: current time, as per cfs_rq_clock_pelt() 3633 * @cfs_rq: cfs_rq to update 3634 * 3635 * The cfs_rq avg is the direct sum of all its entities (blocked and runnable) 3636 * avg. The immediate corollary is that all (fair) tasks must be attached, see 3637 * post_init_entity_util_avg(). 3638 * 3639 * cfs_rq->avg is used for task_h_load() and update_cfs_share() for example. 3640 * 3641 * Returns true if the load decayed or we removed load. 3642 * 3643 * Since both these conditions indicate a changed cfs_rq->avg.load we should 3644 * call update_tg_load_avg() when this function returns true. 3645 */ 3646 static inline int 3647 update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq) 3648 { 3649 unsigned long removed_load = 0, removed_util = 0, removed_runnable = 0; 3650 struct sched_avg *sa = &cfs_rq->avg; 3651 int decayed = 0; 3652 3653 if (cfs_rq->removed.nr) { 3654 unsigned long r; 3655 u32 divider = get_pelt_divider(&cfs_rq->avg); 3656 3657 raw_spin_lock(&cfs_rq->removed.lock); 3658 swap(cfs_rq->removed.util_avg, removed_util); 3659 swap(cfs_rq->removed.load_avg, removed_load); 3660 swap(cfs_rq->removed.runnable_avg, removed_runnable); 3661 cfs_rq->removed.nr = 0; 3662 raw_spin_unlock(&cfs_rq->removed.lock); 3663 3664 r = removed_load; 3665 sub_positive(&sa->load_avg, r); 3666 sub_positive(&sa->load_sum, r * divider); 3667 3668 r = removed_util; 3669 sub_positive(&sa->util_avg, r); 3670 sub_positive(&sa->util_sum, r * divider); 3671 3672 r = removed_runnable; 3673 sub_positive(&sa->runnable_avg, r); 3674 sub_positive(&sa->runnable_sum, r * divider); 3675 3676 /* 3677 * removed_runnable is the unweighted version of removed_load so we 3678 * can use it to estimate removed_load_sum. 3679 */ 3680 add_tg_cfs_propagate(cfs_rq, 3681 -(long)(removed_runnable * divider) >> SCHED_CAPACITY_SHIFT); 3682 3683 decayed = 1; 3684 } 3685 3686 decayed |= __update_load_avg_cfs_rq(now, cfs_rq); 3687 3688 #ifndef CONFIG_64BIT 3689 smp_wmb(); 3690 cfs_rq->load_last_update_time_copy = sa->last_update_time; 3691 #endif 3692 3693 return decayed; 3694 } 3695 3696 /** 3697 * attach_entity_load_avg - attach this entity to its cfs_rq load avg 3698 * @cfs_rq: cfs_rq to attach to 3699 * @se: sched_entity to attach 3700 * 3701 * Must call update_cfs_rq_load_avg() before this, since we rely on 3702 * cfs_rq->avg.last_update_time being current. 3703 */ 3704 static void attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) 3705 { 3706 /* 3707 * cfs_rq->avg.period_contrib can be used for both cfs_rq and se. 3708 * See ___update_load_avg() for details. 3709 */ 3710 u32 divider = get_pelt_divider(&cfs_rq->avg); 3711 3712 /* 3713 * When we attach the @se to the @cfs_rq, we must align the decay 3714 * window because without that, really weird and wonderful things can 3715 * happen. 3716 * 3717 * XXX illustrate 3718 */ 3719 se->avg.last_update_time = cfs_rq->avg.last_update_time; 3720 se->avg.period_contrib = cfs_rq->avg.period_contrib; 3721 3722 /* 3723 * Hell(o) Nasty stuff.. we need to recompute _sum based on the new 3724 * period_contrib. This isn't strictly correct, but since we're 3725 * entirely outside of the PELT hierarchy, nobody cares if we truncate 3726 * _sum a little. 3727 */ 3728 se->avg.util_sum = se->avg.util_avg * divider; 3729 3730 se->avg.runnable_sum = se->avg.runnable_avg * divider; 3731 3732 se->avg.load_sum = divider; 3733 if (se_weight(se)) { 3734 se->avg.load_sum = 3735 div_u64(se->avg.load_avg * se->avg.load_sum, se_weight(se)); 3736 } 3737 3738 enqueue_load_avg(cfs_rq, se); 3739 cfs_rq->avg.util_avg += se->avg.util_avg; 3740 cfs_rq->avg.util_sum += se->avg.util_sum; 3741 cfs_rq->avg.runnable_avg += se->avg.runnable_avg; 3742 cfs_rq->avg.runnable_sum += se->avg.runnable_sum; 3743 3744 add_tg_cfs_propagate(cfs_rq, se->avg.load_sum); 3745 3746 cfs_rq_util_change(cfs_rq, 0); 3747 3748 trace_pelt_cfs_tp(cfs_rq); 3749 } 3750 3751 /** 3752 * detach_entity_load_avg - detach this entity from its cfs_rq load avg 3753 * @cfs_rq: cfs_rq to detach from 3754 * @se: sched_entity to detach 3755 * 3756 * Must call update_cfs_rq_load_avg() before this, since we rely on 3757 * cfs_rq->avg.last_update_time being current. 3758 */ 3759 static void detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) 3760 { 3761 dequeue_load_avg(cfs_rq, se); 3762 sub_positive(&cfs_rq->avg.util_avg, se->avg.util_avg); 3763 sub_positive(&cfs_rq->avg.util_sum, se->avg.util_sum); 3764 sub_positive(&cfs_rq->avg.runnable_avg, se->avg.runnable_avg); 3765 sub_positive(&cfs_rq->avg.runnable_sum, se->avg.runnable_sum); 3766 3767 add_tg_cfs_propagate(cfs_rq, -se->avg.load_sum); 3768 3769 cfs_rq_util_change(cfs_rq, 0); 3770 3771 trace_pelt_cfs_tp(cfs_rq); 3772 } 3773 3774 /* 3775 * Optional action to be done while updating the load average 3776 */ 3777 #define UPDATE_TG 0x1 3778 #define SKIP_AGE_LOAD 0x2 3779 #define DO_ATTACH 0x4 3780 3781 /* Update task and its cfs_rq load average */ 3782 static inline void update_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags) 3783 { 3784 u64 now = cfs_rq_clock_pelt(cfs_rq); 3785 int decayed; 3786 3787 /* 3788 * Track task load average for carrying it to new CPU after migrated, and 3789 * track group sched_entity load average for task_h_load calc in migration 3790 */ 3791 if (se->avg.last_update_time && !(flags & SKIP_AGE_LOAD)) 3792 __update_load_avg_se(now, cfs_rq, se); 3793 3794 decayed = update_cfs_rq_load_avg(now, cfs_rq); 3795 decayed |= propagate_entity_load_avg(se); 3796 3797 if (!se->avg.last_update_time && (flags & DO_ATTACH)) { 3798 3799 /* 3800 * DO_ATTACH means we're here from enqueue_entity(). 3801 * !last_update_time means we've passed through 3802 * migrate_task_rq_fair() indicating we migrated. 3803 * 3804 * IOW we're enqueueing a task on a new CPU. 3805 */ 3806 attach_entity_load_avg(cfs_rq, se); 3807 update_tg_load_avg(cfs_rq); 3808 3809 } else if (decayed) { 3810 cfs_rq_util_change(cfs_rq, 0); 3811 3812 if (flags & UPDATE_TG) 3813 update_tg_load_avg(cfs_rq); 3814 } 3815 } 3816 3817 #ifndef CONFIG_64BIT 3818 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq) 3819 { 3820 u64 last_update_time_copy; 3821 u64 last_update_time; 3822 3823 do { 3824 last_update_time_copy = cfs_rq->load_last_update_time_copy; 3825 smp_rmb(); 3826 last_update_time = cfs_rq->avg.last_update_time; 3827 } while (last_update_time != last_update_time_copy); 3828 3829 return last_update_time; 3830 } 3831 #else 3832 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq) 3833 { 3834 return cfs_rq->avg.last_update_time; 3835 } 3836 #endif 3837 3838 /* 3839 * Synchronize entity load avg of dequeued entity without locking 3840 * the previous rq. 3841 */ 3842 static void sync_entity_load_avg(struct sched_entity *se) 3843 { 3844 struct cfs_rq *cfs_rq = cfs_rq_of(se); 3845 u64 last_update_time; 3846 3847 last_update_time = cfs_rq_last_update_time(cfs_rq); 3848 __update_load_avg_blocked_se(last_update_time, se); 3849 } 3850 3851 /* 3852 * Task first catches up with cfs_rq, and then subtract 3853 * itself from the cfs_rq (task must be off the queue now). 3854 */ 3855 static void remove_entity_load_avg(struct sched_entity *se) 3856 { 3857 struct cfs_rq *cfs_rq = cfs_rq_of(se); 3858 unsigned long flags; 3859 3860 /* 3861 * tasks cannot exit without having gone through wake_up_new_task() -> 3862 * post_init_entity_util_avg() which will have added things to the 3863 * cfs_rq, so we can remove unconditionally. 3864 */ 3865 3866 sync_entity_load_avg(se); 3867 3868 raw_spin_lock_irqsave(&cfs_rq->removed.lock, flags); 3869 ++cfs_rq->removed.nr; 3870 cfs_rq->removed.util_avg += se->avg.util_avg; 3871 cfs_rq->removed.load_avg += se->avg.load_avg; 3872 cfs_rq->removed.runnable_avg += se->avg.runnable_avg; 3873 raw_spin_unlock_irqrestore(&cfs_rq->removed.lock, flags); 3874 } 3875 3876 static inline unsigned long cfs_rq_runnable_avg(struct cfs_rq *cfs_rq) 3877 { 3878 return cfs_rq->avg.runnable_avg; 3879 } 3880 3881 static inline unsigned long cfs_rq_load_avg(struct cfs_rq *cfs_rq) 3882 { 3883 return cfs_rq->avg.load_avg; 3884 } 3885 3886 static int newidle_balance(struct rq *this_rq, struct rq_flags *rf); 3887 3888 static inline unsigned long task_util(struct task_struct *p) 3889 { 3890 return READ_ONCE(p->se.avg.util_avg); 3891 } 3892 3893 static inline unsigned long _task_util_est(struct task_struct *p) 3894 { 3895 struct util_est ue = READ_ONCE(p->se.avg.util_est); 3896 3897 return (max(ue.ewma, ue.enqueued) | UTIL_AVG_UNCHANGED); 3898 } 3899 3900 static inline unsigned long task_util_est(struct task_struct *p) 3901 { 3902 return max(task_util(p), _task_util_est(p)); 3903 } 3904 3905 #ifdef CONFIG_UCLAMP_TASK 3906 static inline unsigned long uclamp_task_util(struct task_struct *p) 3907 { 3908 return clamp(task_util_est(p), 3909 uclamp_eff_value(p, UCLAMP_MIN), 3910 uclamp_eff_value(p, UCLAMP_MAX)); 3911 } 3912 #else 3913 static inline unsigned long uclamp_task_util(struct task_struct *p) 3914 { 3915 return task_util_est(p); 3916 } 3917 #endif 3918 3919 static inline void util_est_enqueue(struct cfs_rq *cfs_rq, 3920 struct task_struct *p) 3921 { 3922 unsigned int enqueued; 3923 3924 if (!sched_feat(UTIL_EST)) 3925 return; 3926 3927 /* Update root cfs_rq's estimated utilization */ 3928 enqueued = cfs_rq->avg.util_est.enqueued; 3929 enqueued += _task_util_est(p); 3930 WRITE_ONCE(cfs_rq->avg.util_est.enqueued, enqueued); 3931 3932 trace_sched_util_est_cfs_tp(cfs_rq); 3933 } 3934 3935 /* 3936 * Check if a (signed) value is within a specified (unsigned) margin, 3937 * based on the observation that: 3938 * 3939 * abs(x) < y := (unsigned)(x + y - 1) < (2 * y - 1) 3940 * 3941 * NOTE: this only works when value + maring < INT_MAX. 3942 */ 3943 static inline bool within_margin(int value, int margin) 3944 { 3945 return ((unsigned int)(value + margin - 1) < (2 * margin - 1)); 3946 } 3947 3948 static void 3949 util_est_dequeue(struct cfs_rq *cfs_rq, struct task_struct *p, bool task_sleep) 3950 { 3951 long last_ewma_diff; 3952 struct util_est ue; 3953 int cpu; 3954 3955 if (!sched_feat(UTIL_EST)) 3956 return; 3957 3958 /* Update root cfs_rq's estimated utilization */ 3959 ue.enqueued = cfs_rq->avg.util_est.enqueued; 3960 ue.enqueued -= min_t(unsigned int, ue.enqueued, _task_util_est(p)); 3961 WRITE_ONCE(cfs_rq->avg.util_est.enqueued, ue.enqueued); 3962 3963 trace_sched_util_est_cfs_tp(cfs_rq); 3964 3965 /* 3966 * Skip update of task's estimated utilization when the task has not 3967 * yet completed an activation, e.g. being migrated. 3968 */ 3969 if (!task_sleep) 3970 return; 3971 3972 /* 3973 * If the PELT values haven't changed since enqueue time, 3974 * skip the util_est update. 3975 */ 3976 ue = p->se.avg.util_est; 3977 if (ue.enqueued & UTIL_AVG_UNCHANGED) 3978 return; 3979 3980 /* 3981 * Reset EWMA on utilization increases, the moving average is used only 3982 * to smooth utilization decreases. 3983 */ 3984 ue.enqueued = (task_util(p) | UTIL_AVG_UNCHANGED); 3985 if (sched_feat(UTIL_EST_FASTUP)) { 3986 if (ue.ewma < ue.enqueued) { 3987 ue.ewma = ue.enqueued; 3988 goto done; 3989 } 3990 } 3991 3992 /* 3993 * Skip update of task's estimated utilization when its EWMA is 3994 * already ~1% close to its last activation value. 3995 */ 3996 last_ewma_diff = ue.enqueued - ue.ewma; 3997 if (within_margin(last_ewma_diff, (SCHED_CAPACITY_SCALE / 100))) 3998 return; 3999 4000 /* 4001 * To avoid overestimation of actual task utilization, skip updates if 4002 * we cannot grant there is idle time in this CPU. 4003 */ 4004 cpu = cpu_of(rq_of(cfs_rq)); 4005 if (task_util(p) > capacity_orig_of(cpu)) 4006 return; 4007 4008 /* 4009 * Update Task's estimated utilization 4010 * 4011 * When *p completes an activation we can consolidate another sample 4012 * of the task size. This is done by storing the current PELT value 4013 * as ue.enqueued and by using this value to update the Exponential 4014 * Weighted Moving Average (EWMA): 4015 * 4016 * ewma(t) = w * task_util(p) + (1-w) * ewma(t-1) 4017 * = w * task_util(p) + ewma(t-1) - w * ewma(t-1) 4018 * = w * (task_util(p) - ewma(t-1)) + ewma(t-1) 4019 * = w * ( last_ewma_diff ) + ewma(t-1) 4020 * = w * (last_ewma_diff + ewma(t-1) / w) 4021 * 4022 * Where 'w' is the weight of new samples, which is configured to be 4023 * 0.25, thus making w=1/4 ( >>= UTIL_EST_WEIGHT_SHIFT) 4024 */ 4025 ue.ewma <<= UTIL_EST_WEIGHT_SHIFT; 4026 ue.ewma += last_ewma_diff; 4027 ue.ewma >>= UTIL_EST_WEIGHT_SHIFT; 4028 done: 4029 WRITE_ONCE(p->se.avg.util_est, ue); 4030 4031 trace_sched_util_est_se_tp(&p->se); 4032 } 4033 4034 static inline int task_fits_capacity(struct task_struct *p, long capacity) 4035 { 4036 return fits_capacity(uclamp_task_util(p), capacity); 4037 } 4038 4039 static inline void update_misfit_status(struct task_struct *p, struct rq *rq) 4040 { 4041 if (!static_branch_unlikely(&sched_asym_cpucapacity)) 4042 return; 4043 4044 if (!p) { 4045 rq->misfit_task_load = 0; 4046 return; 4047 } 4048 4049 if (task_fits_capacity(p, capacity_of(cpu_of(rq)))) { 4050 rq->misfit_task_load = 0; 4051 return; 4052 } 4053 4054 /* 4055 * Make sure that misfit_task_load will not be null even if 4056 * task_h_load() returns 0. 4057 */ 4058 rq->misfit_task_load = max_t(unsigned long, task_h_load(p), 1); 4059 } 4060 4061 #else /* CONFIG_SMP */ 4062 4063 #define UPDATE_TG 0x0 4064 #define SKIP_AGE_LOAD 0x0 4065 #define DO_ATTACH 0x0 4066 4067 static inline void update_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int not_used1) 4068 { 4069 cfs_rq_util_change(cfs_rq, 0); 4070 } 4071 4072 static inline void remove_entity_load_avg(struct sched_entity *se) {} 4073 4074 static inline void 4075 attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {} 4076 static inline void 4077 detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {} 4078 4079 static inline int newidle_balance(struct rq *rq, struct rq_flags *rf) 4080 { 4081 return 0; 4082 } 4083 4084 static inline void 4085 util_est_enqueue(struct cfs_rq *cfs_rq, struct task_struct *p) {} 4086 4087 static inline void 4088 util_est_dequeue(struct cfs_rq *cfs_rq, struct task_struct *p, 4089 bool task_sleep) {} 4090 static inline void update_misfit_status(struct task_struct *p, struct rq *rq) {} 4091 4092 #endif /* CONFIG_SMP */ 4093 4094 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se) 4095 { 4096 #ifdef CONFIG_SCHED_DEBUG 4097 s64 d = se->vruntime - cfs_rq->min_vruntime; 4098 4099 if (d < 0) 4100 d = -d; 4101 4102 if (d > 3*sysctl_sched_latency) 4103 schedstat_inc(cfs_rq->nr_spread_over); 4104 #endif 4105 } 4106 4107 static void 4108 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial) 4109 { 4110 u64 vruntime = cfs_rq->min_vruntime; 4111 4112 /* 4113 * The 'current' period is already promised to the current tasks, 4114 * however the extra weight of the new task will slow them down a 4115 * little, place the new task so that it fits in the slot that 4116 * stays open at the end. 4117 */ 4118 if (initial && sched_feat(START_DEBIT)) 4119 vruntime += sched_vslice(cfs_rq, se); 4120 4121 /* sleeps up to a single latency don't count. */ 4122 if (!initial) { 4123 unsigned long thresh = sysctl_sched_latency; 4124 4125 /* 4126 * Halve their sleep time's effect, to allow 4127 * for a gentler effect of sleepers: 4128 */ 4129 if (sched_feat(GENTLE_FAIR_SLEEPERS)) 4130 thresh >>= 1; 4131 4132 vruntime -= thresh; 4133 } 4134 4135 /* ensure we never gain time by being placed backwards. */ 4136 se->vruntime = max_vruntime(se->vruntime, vruntime); 4137 } 4138 4139 static void check_enqueue_throttle(struct cfs_rq *cfs_rq); 4140 4141 static inline void check_schedstat_required(void) 4142 { 4143 #ifdef CONFIG_SCHEDSTATS 4144 if (schedstat_enabled()) 4145 return; 4146 4147 /* Force schedstat enabled if a dependent tracepoint is active */ 4148 if (trace_sched_stat_wait_enabled() || 4149 trace_sched_stat_sleep_enabled() || 4150 trace_sched_stat_iowait_enabled() || 4151 trace_sched_stat_blocked_enabled() || 4152 trace_sched_stat_runtime_enabled()) { 4153 printk_deferred_once("Scheduler tracepoints stat_sleep, stat_iowait, " 4154 "stat_blocked and stat_runtime require the " 4155 "kernel parameter schedstats=enable or " 4156 "kernel.sched_schedstats=1\n"); 4157 } 4158 #endif 4159 } 4160 4161 static inline bool cfs_bandwidth_used(void); 4162 4163 /* 4164 * MIGRATION 4165 * 4166 * dequeue 4167 * update_curr() 4168 * update_min_vruntime() 4169 * vruntime -= min_vruntime 4170 * 4171 * enqueue 4172 * update_curr() 4173 * update_min_vruntime() 4174 * vruntime += min_vruntime 4175 * 4176 * this way the vruntime transition between RQs is done when both 4177 * min_vruntime are up-to-date. 4178 * 4179 * WAKEUP (remote) 4180 * 4181 * ->migrate_task_rq_fair() (p->state == TASK_WAKING) 4182 * vruntime -= min_vruntime 4183 * 4184 * enqueue 4185 * update_curr() 4186 * update_min_vruntime() 4187 * vruntime += min_vruntime 4188 * 4189 * this way we don't have the most up-to-date min_vruntime on the originating 4190 * CPU and an up-to-date min_vruntime on the destination CPU. 4191 */ 4192 4193 static void 4194 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags) 4195 { 4196 bool renorm = !(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_MIGRATED); 4197 bool curr = cfs_rq->curr == se; 4198 4199 /* 4200 * If we're the current task, we must renormalise before calling 4201 * update_curr(). 4202 */ 4203 if (renorm && curr) 4204 se->vruntime += cfs_rq->min_vruntime; 4205 4206 update_curr(cfs_rq); 4207 4208 /* 4209 * Otherwise, renormalise after, such that we're placed at the current 4210 * moment in time, instead of some random moment in the past. Being 4211 * placed in the past could significantly boost this task to the 4212 * fairness detriment of existing tasks. 4213 */ 4214 if (renorm && !curr) 4215 se->vruntime += cfs_rq->min_vruntime; 4216 4217 /* 4218 * When enqueuing a sched_entity, we must: 4219 * - Update loads to have both entity and cfs_rq synced with now. 4220 * - Add its load to cfs_rq->runnable_avg 4221 * - For group_entity, update its weight to reflect the new share of 4222 * its group cfs_rq 4223 * - Add its new weight to cfs_rq->load.weight 4224 */ 4225 update_load_avg(cfs_rq, se, UPDATE_TG | DO_ATTACH); 4226 se_update_runnable(se); 4227 update_cfs_group(se); 4228 account_entity_enqueue(cfs_rq, se); 4229 4230 if (flags & ENQUEUE_WAKEUP) 4231 place_entity(cfs_rq, se, 0); 4232 4233 check_schedstat_required(); 4234 update_stats_enqueue(cfs_rq, se, flags); 4235 check_spread(cfs_rq, se); 4236 if (!curr) 4237 __enqueue_entity(cfs_rq, se); 4238 se->on_rq = 1; 4239 4240 /* 4241 * When bandwidth control is enabled, cfs might have been removed 4242 * because of a parent been throttled but cfs->nr_running > 1. Try to 4243 * add it unconditionnally. 4244 */ 4245 if (cfs_rq->nr_running == 1 || cfs_bandwidth_used()) 4246 list_add_leaf_cfs_rq(cfs_rq); 4247 4248 if (cfs_rq->nr_running == 1) 4249 check_enqueue_throttle(cfs_rq); 4250 } 4251 4252 static void __clear_buddies_last(struct sched_entity *se) 4253 { 4254 for_each_sched_entity(se) { 4255 struct cfs_rq *cfs_rq = cfs_rq_of(se); 4256 if (cfs_rq->last != se) 4257 break; 4258 4259 cfs_rq->last = NULL; 4260 } 4261 } 4262 4263 static void __clear_buddies_next(struct sched_entity *se) 4264 { 4265 for_each_sched_entity(se) { 4266 struct cfs_rq *cfs_rq = cfs_rq_of(se); 4267 if (cfs_rq->next != se) 4268 break; 4269 4270 cfs_rq->next = NULL; 4271 } 4272 } 4273 4274 static void __clear_buddies_skip(struct sched_entity *se) 4275 { 4276 for_each_sched_entity(se) { 4277 struct cfs_rq *cfs_rq = cfs_rq_of(se); 4278 if (cfs_rq->skip != se) 4279 break; 4280 4281 cfs_rq->skip = NULL; 4282 } 4283 } 4284 4285 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se) 4286 { 4287 if (cfs_rq->last == se) 4288 __clear_buddies_last(se); 4289 4290 if (cfs_rq->next == se) 4291 __clear_buddies_next(se); 4292 4293 if (cfs_rq->skip == se) 4294 __clear_buddies_skip(se); 4295 } 4296 4297 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq); 4298 4299 static void 4300 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags) 4301 { 4302 /* 4303 * Update run-time statistics of the 'current'. 4304 */ 4305 update_curr(cfs_rq); 4306 4307 /* 4308 * When dequeuing a sched_entity, we must: 4309 * - Update loads to have both entity and cfs_rq synced with now. 4310 * - Subtract its load from the cfs_rq->runnable_avg. 4311 * - Subtract its previous weight from cfs_rq->load.weight. 4312 * - For group entity, update its weight to reflect the new share 4313 * of its group cfs_rq. 4314 */ 4315 update_load_avg(cfs_rq, se, UPDATE_TG); 4316 se_update_runnable(se); 4317 4318 update_stats_dequeue(cfs_rq, se, flags); 4319 4320 clear_buddies(cfs_rq, se); 4321 4322 if (se != cfs_rq->curr) 4323 __dequeue_entity(cfs_rq, se); 4324 se->on_rq = 0; 4325 account_entity_dequeue(cfs_rq, se); 4326 4327 /* 4328 * Normalize after update_curr(); which will also have moved 4329 * min_vruntime if @se is the one holding it back. But before doing 4330 * update_min_vruntime() again, which will discount @se's position and 4331 * can move min_vruntime forward still more. 4332 */ 4333 if (!(flags & DEQUEUE_SLEEP)) 4334 se->vruntime -= cfs_rq->min_vruntime; 4335 4336 /* return excess runtime on last dequeue */ 4337 return_cfs_rq_runtime(cfs_rq); 4338 4339 update_cfs_group(se); 4340 4341 /* 4342 * Now advance min_vruntime if @se was the entity holding it back, 4343 * except when: DEQUEUE_SAVE && !DEQUEUE_MOVE, in this case we'll be 4344 * put back on, and if we advance min_vruntime, we'll be placed back 4345 * further than we started -- ie. we'll be penalized. 4346 */ 4347 if ((flags & (DEQUEUE_SAVE | DEQUEUE_MOVE)) != DEQUEUE_SAVE) 4348 update_min_vruntime(cfs_rq); 4349 } 4350 4351 /* 4352 * Preempt the current task with a newly woken task if needed: 4353 */ 4354 static void 4355 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr) 4356 { 4357 unsigned long ideal_runtime, delta_exec; 4358 struct sched_entity *se; 4359 s64 delta; 4360 4361 ideal_runtime = sched_slice(cfs_rq, curr); 4362 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime; 4363 if (delta_exec > ideal_runtime) { 4364 resched_curr(rq_of(cfs_rq)); 4365 /* 4366 * The current task ran long enough, ensure it doesn't get 4367 * re-elected due to buddy favours. 4368 */ 4369 clear_buddies(cfs_rq, curr); 4370 return; 4371 } 4372 4373 /* 4374 * Ensure that a task that missed wakeup preemption by a 4375 * narrow margin doesn't have to wait for a full slice. 4376 * This also mitigates buddy induced latencies under load. 4377 */ 4378 if (delta_exec < sysctl_sched_min_granularity) 4379 return; 4380 4381 se = __pick_first_entity(cfs_rq); 4382 delta = curr->vruntime - se->vruntime; 4383 4384 if (delta < 0) 4385 return; 4386 4387 if (delta > ideal_runtime) 4388 resched_curr(rq_of(cfs_rq)); 4389 } 4390 4391 static void 4392 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se) 4393 { 4394 /* 'current' is not kept within the tree. */ 4395 if (se->on_rq) { 4396 /* 4397 * Any task has to be enqueued before it get to execute on 4398 * a CPU. So account for the time it spent waiting on the 4399 * runqueue. 4400 */ 4401 update_stats_wait_end(cfs_rq, se); 4402 __dequeue_entity(cfs_rq, se); 4403 update_load_avg(cfs_rq, se, UPDATE_TG); 4404 } 4405 4406 update_stats_curr_start(cfs_rq, se); 4407 cfs_rq->curr = se; 4408 4409 /* 4410 * Track our maximum slice length, if the CPU's load is at 4411 * least twice that of our own weight (i.e. dont track it 4412 * when there are only lesser-weight tasks around): 4413 */ 4414 if (schedstat_enabled() && 4415 rq_of(cfs_rq)->cfs.load.weight >= 2*se->load.weight) { 4416 schedstat_set(se->statistics.slice_max, 4417 max((u64)schedstat_val(se->statistics.slice_max), 4418 se->sum_exec_runtime - se->prev_sum_exec_runtime)); 4419 } 4420 4421 se->prev_sum_exec_runtime = se->sum_exec_runtime; 4422 } 4423 4424 static int 4425 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se); 4426 4427 /* 4428 * Pick the next process, keeping these things in mind, in this order: 4429 * 1) keep things fair between processes/task groups 4430 * 2) pick the "next" process, since someone really wants that to run 4431 * 3) pick the "last" process, for cache locality 4432 * 4) do not run the "skip" process, if something else is available 4433 */ 4434 static struct sched_entity * 4435 pick_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *curr) 4436 { 4437 struct sched_entity *left = __pick_first_entity(cfs_rq); 4438 struct sched_entity *se; 4439 4440 /* 4441 * If curr is set we have to see if its left of the leftmost entity 4442 * still in the tree, provided there was anything in the tree at all. 4443 */ 4444 if (!left || (curr && entity_before(curr, left))) 4445 left = curr; 4446 4447 se = left; /* ideally we run the leftmost entity */ 4448 4449 /* 4450 * Avoid running the skip buddy, if running something else can 4451 * be done without getting too unfair. 4452 */ 4453 if (cfs_rq->skip == se) { 4454 struct sched_entity *second; 4455 4456 if (se == curr) { 4457 second = __pick_first_entity(cfs_rq); 4458 } else { 4459 second = __pick_next_entity(se); 4460 if (!second || (curr && entity_before(curr, second))) 4461 second = curr; 4462 } 4463 4464 if (second && wakeup_preempt_entity(second, left) < 1) 4465 se = second; 4466 } 4467 4468 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1) { 4469 /* 4470 * Someone really wants this to run. If it's not unfair, run it. 4471 */ 4472 se = cfs_rq->next; 4473 } else if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1) { 4474 /* 4475 * Prefer last buddy, try to return the CPU to a preempted task. 4476 */ 4477 se = cfs_rq->last; 4478 } 4479 4480 clear_buddies(cfs_rq, se); 4481 4482 return se; 4483 } 4484 4485 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq); 4486 4487 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev) 4488 { 4489 /* 4490 * If still on the runqueue then deactivate_task() 4491 * was not called and update_curr() has to be done: 4492 */ 4493 if (prev->on_rq) 4494 update_curr(cfs_rq); 4495 4496 /* throttle cfs_rqs exceeding runtime */ 4497 check_cfs_rq_runtime(cfs_rq); 4498 4499 check_spread(cfs_rq, prev); 4500 4501 if (prev->on_rq) { 4502 update_stats_wait_start(cfs_rq, prev); 4503 /* Put 'current' back into the tree. */ 4504 __enqueue_entity(cfs_rq, prev); 4505 /* in !on_rq case, update occurred at dequeue */ 4506 update_load_avg(cfs_rq, prev, 0); 4507 } 4508 cfs_rq->curr = NULL; 4509 } 4510 4511 static void 4512 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued) 4513 { 4514 /* 4515 * Update run-time statistics of the 'current'. 4516 */ 4517 update_curr(cfs_rq); 4518 4519 /* 4520 * Ensure that runnable average is periodically updated. 4521 */ 4522 update_load_avg(cfs_rq, curr, UPDATE_TG); 4523 update_cfs_group(curr); 4524 4525 #ifdef CONFIG_SCHED_HRTICK 4526 /* 4527 * queued ticks are scheduled to match the slice, so don't bother 4528 * validating it and just reschedule. 4529 */ 4530 if (queued) { 4531 resched_curr(rq_of(cfs_rq)); 4532 return; 4533 } 4534 /* 4535 * don't let the period tick interfere with the hrtick preemption 4536 */ 4537 if (!sched_feat(DOUBLE_TICK) && 4538 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer)) 4539 return; 4540 #endif 4541 4542 if (cfs_rq->nr_running > 1) 4543 check_preempt_tick(cfs_rq, curr); 4544 } 4545 4546 4547 /************************************************** 4548 * CFS bandwidth control machinery 4549 */ 4550 4551 #ifdef CONFIG_CFS_BANDWIDTH 4552 4553 #ifdef CONFIG_JUMP_LABEL 4554 static struct static_key __cfs_bandwidth_used; 4555 4556 static inline bool cfs_bandwidth_used(void) 4557 { 4558 return static_key_false(&__cfs_bandwidth_used); 4559 } 4560 4561 void cfs_bandwidth_usage_inc(void) 4562 { 4563 static_key_slow_inc_cpuslocked(&__cfs_bandwidth_used); 4564 } 4565 4566 void cfs_bandwidth_usage_dec(void) 4567 { 4568 static_key_slow_dec_cpuslocked(&__cfs_bandwidth_used); 4569 } 4570 #else /* CONFIG_JUMP_LABEL */ 4571 static bool cfs_bandwidth_used(void) 4572 { 4573 return true; 4574 } 4575 4576 void cfs_bandwidth_usage_inc(void) {} 4577 void cfs_bandwidth_usage_dec(void) {} 4578 #endif /* CONFIG_JUMP_LABEL */ 4579 4580 /* 4581 * default period for cfs group bandwidth. 4582 * default: 0.1s, units: nanoseconds 4583 */ 4584 static inline u64 default_cfs_period(void) 4585 { 4586 return 100000000ULL; 4587 } 4588 4589 static inline u64 sched_cfs_bandwidth_slice(void) 4590 { 4591 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC; 4592 } 4593 4594 /* 4595 * Replenish runtime according to assigned quota. We use sched_clock_cpu 4596 * directly instead of rq->clock to avoid adding additional synchronization 4597 * around rq->lock. 4598 * 4599 * requires cfs_b->lock 4600 */ 4601 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b) 4602 { 4603 if (cfs_b->quota != RUNTIME_INF) 4604 cfs_b->runtime = cfs_b->quota; 4605 } 4606 4607 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg) 4608 { 4609 return &tg->cfs_bandwidth; 4610 } 4611 4612 /* returns 0 on failure to allocate runtime */ 4613 static int __assign_cfs_rq_runtime(struct cfs_bandwidth *cfs_b, 4614 struct cfs_rq *cfs_rq, u64 target_runtime) 4615 { 4616 u64 min_amount, amount = 0; 4617 4618 lockdep_assert_held(&cfs_b->lock); 4619 4620 /* note: this is a positive sum as runtime_remaining <= 0 */ 4621 min_amount = target_runtime - cfs_rq->runtime_remaining; 4622 4623 if (cfs_b->quota == RUNTIME_INF) 4624 amount = min_amount; 4625 else { 4626 start_cfs_bandwidth(cfs_b); 4627 4628 if (cfs_b->runtime > 0) { 4629 amount = min(cfs_b->runtime, min_amount); 4630 cfs_b->runtime -= amount; 4631 cfs_b->idle = 0; 4632 } 4633 } 4634 4635 cfs_rq->runtime_remaining += amount; 4636 4637 return cfs_rq->runtime_remaining > 0; 4638 } 4639 4640 /* returns 0 on failure to allocate runtime */ 4641 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq) 4642 { 4643 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg); 4644 int ret; 4645 4646 raw_spin_lock(&cfs_b->lock); 4647 ret = __assign_cfs_rq_runtime(cfs_b, cfs_rq, sched_cfs_bandwidth_slice()); 4648 raw_spin_unlock(&cfs_b->lock); 4649 4650 return ret; 4651 } 4652 4653 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) 4654 { 4655 /* dock delta_exec before expiring quota (as it could span periods) */ 4656 cfs_rq->runtime_remaining -= delta_exec; 4657 4658 if (likely(cfs_rq->runtime_remaining > 0)) 4659 return; 4660 4661 if (cfs_rq->throttled) 4662 return; 4663 /* 4664 * if we're unable to extend our runtime we resched so that the active 4665 * hierarchy can be throttled 4666 */ 4667 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr)) 4668 resched_curr(rq_of(cfs_rq)); 4669 } 4670 4671 static __always_inline 4672 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) 4673 { 4674 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled) 4675 return; 4676 4677 __account_cfs_rq_runtime(cfs_rq, delta_exec); 4678 } 4679 4680 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq) 4681 { 4682 return cfs_bandwidth_used() && cfs_rq->throttled; 4683 } 4684 4685 /* check whether cfs_rq, or any parent, is throttled */ 4686 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq) 4687 { 4688 return cfs_bandwidth_used() && cfs_rq->throttle_count; 4689 } 4690 4691 /* 4692 * Ensure that neither of the group entities corresponding to src_cpu or 4693 * dest_cpu are members of a throttled hierarchy when performing group 4694 * load-balance operations. 4695 */ 4696 static inline int throttled_lb_pair(struct task_group *tg, 4697 int src_cpu, int dest_cpu) 4698 { 4699 struct cfs_rq *src_cfs_rq, *dest_cfs_rq; 4700 4701 src_cfs_rq = tg->cfs_rq[src_cpu]; 4702 dest_cfs_rq = tg->cfs_rq[dest_cpu]; 4703 4704 return throttled_hierarchy(src_cfs_rq) || 4705 throttled_hierarchy(dest_cfs_rq); 4706 } 4707 4708 static int tg_unthrottle_up(struct task_group *tg, void *data) 4709 { 4710 struct rq *rq = data; 4711 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)]; 4712 4713 cfs_rq->throttle_count--; 4714 if (!cfs_rq->throttle_count) { 4715 cfs_rq->throttled_clock_task_time += rq_clock_task(rq) - 4716 cfs_rq->throttled_clock_task; 4717 4718 /* Add cfs_rq with already running entity in the list */ 4719 if (cfs_rq->nr_running >= 1) 4720 list_add_leaf_cfs_rq(cfs_rq); 4721 } 4722 4723 return 0; 4724 } 4725 4726 static int tg_throttle_down(struct task_group *tg, void *data) 4727 { 4728 struct rq *rq = data; 4729 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)]; 4730 4731 /* group is entering throttled state, stop time */ 4732 if (!cfs_rq->throttle_count) { 4733 cfs_rq->throttled_clock_task = rq_clock_task(rq); 4734 list_del_leaf_cfs_rq(cfs_rq); 4735 } 4736 cfs_rq->throttle_count++; 4737 4738 return 0; 4739 } 4740 4741 static bool throttle_cfs_rq(struct cfs_rq *cfs_rq) 4742 { 4743 struct rq *rq = rq_of(cfs_rq); 4744 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg); 4745 struct sched_entity *se; 4746 long task_delta, idle_task_delta, dequeue = 1; 4747 4748 raw_spin_lock(&cfs_b->lock); 4749 /* This will start the period timer if necessary */ 4750 if (__assign_cfs_rq_runtime(cfs_b, cfs_rq, 1)) { 4751 /* 4752 * We have raced with bandwidth becoming available, and if we 4753 * actually throttled the timer might not unthrottle us for an 4754 * entire period. We additionally needed to make sure that any 4755 * subsequent check_cfs_rq_runtime calls agree not to throttle 4756 * us, as we may commit to do cfs put_prev+pick_next, so we ask 4757 * for 1ns of runtime rather than just check cfs_b. 4758 */ 4759 dequeue = 0; 4760 } else { 4761 list_add_tail_rcu(&cfs_rq->throttled_list, 4762 &cfs_b->throttled_cfs_rq); 4763 } 4764 raw_spin_unlock(&cfs_b->lock); 4765 4766 if (!dequeue) 4767 return false; /* Throttle no longer required. */ 4768 4769 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))]; 4770 4771 /* freeze hierarchy runnable averages while throttled */ 4772 rcu_read_lock(); 4773 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq); 4774 rcu_read_unlock(); 4775 4776 task_delta = cfs_rq->h_nr_running; 4777 idle_task_delta = cfs_rq->idle_h_nr_running; 4778 for_each_sched_entity(se) { 4779 struct cfs_rq *qcfs_rq = cfs_rq_of(se); 4780 /* throttled entity or throttle-on-deactivate */ 4781 if (!se->on_rq) 4782 break; 4783 4784 if (dequeue) { 4785 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP); 4786 } else { 4787 update_load_avg(qcfs_rq, se, 0); 4788 se_update_runnable(se); 4789 } 4790 4791 qcfs_rq->h_nr_running -= task_delta; 4792 qcfs_rq->idle_h_nr_running -= idle_task_delta; 4793 4794 if (qcfs_rq->load.weight) 4795 dequeue = 0; 4796 } 4797 4798 if (!se) 4799 sub_nr_running(rq, task_delta); 4800 4801 /* 4802 * Note: distribution will already see us throttled via the 4803 * throttled-list. rq->lock protects completion. 4804 */ 4805 cfs_rq->throttled = 1; 4806 cfs_rq->throttled_clock = rq_clock(rq); 4807 return true; 4808 } 4809 4810 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq) 4811 { 4812 struct rq *rq = rq_of(cfs_rq); 4813 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg); 4814 struct sched_entity *se; 4815 long task_delta, idle_task_delta; 4816 4817 se = cfs_rq->tg->se[cpu_of(rq)]; 4818 4819 cfs_rq->throttled = 0; 4820 4821 update_rq_clock(rq); 4822 4823 raw_spin_lock(&cfs_b->lock); 4824 cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock; 4825 list_del_rcu(&cfs_rq->throttled_list); 4826 raw_spin_unlock(&cfs_b->lock); 4827 4828 /* update hierarchical throttle state */ 4829 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq); 4830 4831 if (!cfs_rq->load.weight) 4832 return; 4833 4834 task_delta = cfs_rq->h_nr_running; 4835 idle_task_delta = cfs_rq->idle_h_nr_running; 4836 for_each_sched_entity(se) { 4837 if (se->on_rq) 4838 break; 4839 cfs_rq = cfs_rq_of(se); 4840 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP); 4841 4842 cfs_rq->h_nr_running += task_delta; 4843 cfs_rq->idle_h_nr_running += idle_task_delta; 4844 4845 /* end evaluation on encountering a throttled cfs_rq */ 4846 if (cfs_rq_throttled(cfs_rq)) 4847 goto unthrottle_throttle; 4848 } 4849 4850 for_each_sched_entity(se) { 4851 cfs_rq = cfs_rq_of(se); 4852 4853 update_load_avg(cfs_rq, se, UPDATE_TG); 4854 se_update_runnable(se); 4855 4856 cfs_rq->h_nr_running += task_delta; 4857 cfs_rq->idle_h_nr_running += idle_task_delta; 4858 4859 4860 /* end evaluation on encountering a throttled cfs_rq */ 4861 if (cfs_rq_throttled(cfs_rq)) 4862 goto unthrottle_throttle; 4863 4864 /* 4865 * One parent has been throttled and cfs_rq removed from the 4866 * list. Add it back to not break the leaf list. 4867 */ 4868 if (throttled_hierarchy(cfs_rq)) 4869 list_add_leaf_cfs_rq(cfs_rq); 4870 } 4871 4872 /* At this point se is NULL and we are at root level*/ 4873 add_nr_running(rq, task_delta); 4874 4875 unthrottle_throttle: 4876 /* 4877 * The cfs_rq_throttled() breaks in the above iteration can result in 4878 * incomplete leaf list maintenance, resulting in triggering the 4879 * assertion below. 4880 */ 4881 for_each_sched_entity(se) { 4882 cfs_rq = cfs_rq_of(se); 4883 4884 if (list_add_leaf_cfs_rq(cfs_rq)) 4885 break; 4886 } 4887 4888 assert_list_leaf_cfs_rq(rq); 4889 4890 /* Determine whether we need to wake up potentially idle CPU: */ 4891 if (rq->curr == rq->idle && rq->cfs.nr_running) 4892 resched_curr(rq); 4893 } 4894 4895 static void distribute_cfs_runtime(struct cfs_bandwidth *cfs_b) 4896 { 4897 struct cfs_rq *cfs_rq; 4898 u64 runtime, remaining = 1; 4899 4900 rcu_read_lock(); 4901 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq, 4902 throttled_list) { 4903 struct rq *rq = rq_of(cfs_rq); 4904 struct rq_flags rf; 4905 4906 rq_lock_irqsave(rq, &rf); 4907 if (!cfs_rq_throttled(cfs_rq)) 4908 goto next; 4909 4910 /* By the above check, this should never be true */ 4911 SCHED_WARN_ON(cfs_rq->runtime_remaining > 0); 4912 4913 raw_spin_lock(&cfs_b->lock); 4914 runtime = -cfs_rq->runtime_remaining + 1; 4915 if (runtime > cfs_b->runtime) 4916 runtime = cfs_b->runtime; 4917 cfs_b->runtime -= runtime; 4918 remaining = cfs_b->runtime; 4919 raw_spin_unlock(&cfs_b->lock); 4920 4921 cfs_rq->runtime_remaining += runtime; 4922 4923 /* we check whether we're throttled above */ 4924 if (cfs_rq->runtime_remaining > 0) 4925 unthrottle_cfs_rq(cfs_rq); 4926 4927 next: 4928 rq_unlock_irqrestore(rq, &rf); 4929 4930 if (!remaining) 4931 break; 4932 } 4933 rcu_read_unlock(); 4934 } 4935 4936 /* 4937 * Responsible for refilling a task_group's bandwidth and unthrottling its 4938 * cfs_rqs as appropriate. If there has been no activity within the last 4939 * period the timer is deactivated until scheduling resumes; cfs_b->idle is 4940 * used to track this state. 4941 */ 4942 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun, unsigned long flags) 4943 { 4944 int throttled; 4945 4946 /* no need to continue the timer with no bandwidth constraint */ 4947 if (cfs_b->quota == RUNTIME_INF) 4948 goto out_deactivate; 4949 4950 throttled = !list_empty(&cfs_b->throttled_cfs_rq); 4951 cfs_b->nr_periods += overrun; 4952 4953 /* 4954 * idle depends on !throttled (for the case of a large deficit), and if 4955 * we're going inactive then everything else can be deferred 4956 */ 4957 if (cfs_b->idle && !throttled) 4958 goto out_deactivate; 4959 4960 __refill_cfs_bandwidth_runtime(cfs_b); 4961 4962 if (!throttled) { 4963 /* mark as potentially idle for the upcoming period */ 4964 cfs_b->idle = 1; 4965 return 0; 4966 } 4967 4968 /* account preceding periods in which throttling occurred */ 4969 cfs_b->nr_throttled += overrun; 4970 4971 /* 4972 * This check is repeated as we release cfs_b->lock while we unthrottle. 4973 */ 4974 while (throttled && cfs_b->runtime > 0) { 4975 raw_spin_unlock_irqrestore(&cfs_b->lock, flags); 4976 /* we can't nest cfs_b->lock while distributing bandwidth */ 4977 distribute_cfs_runtime(cfs_b); 4978 raw_spin_lock_irqsave(&cfs_b->lock, flags); 4979 4980 throttled = !list_empty(&cfs_b->throttled_cfs_rq); 4981 } 4982 4983 /* 4984 * While we are ensured activity in the period following an 4985 * unthrottle, this also covers the case in which the new bandwidth is 4986 * insufficient to cover the existing bandwidth deficit. (Forcing the 4987 * timer to remain active while there are any throttled entities.) 4988 */ 4989 cfs_b->idle = 0; 4990 4991 return 0; 4992 4993 out_deactivate: 4994 return 1; 4995 } 4996 4997 /* a cfs_rq won't donate quota below this amount */ 4998 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC; 4999 /* minimum remaining period time to redistribute slack quota */ 5000 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC; 5001 /* how long we wait to gather additional slack before distributing */ 5002 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC; 5003 5004 /* 5005 * Are we near the end of the current quota period? 5006 * 5007 * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the 5008 * hrtimer base being cleared by hrtimer_start. In the case of 5009 * migrate_hrtimers, base is never cleared, so we are fine. 5010 */ 5011 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire) 5012 { 5013 struct hrtimer *refresh_timer = &cfs_b->period_timer; 5014 u64 remaining; 5015 5016 /* if the call-back is running a quota refresh is already occurring */ 5017 if (hrtimer_callback_running(refresh_timer)) 5018 return 1; 5019 5020 /* is a quota refresh about to occur? */ 5021 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer)); 5022 if (remaining < min_expire) 5023 return 1; 5024 5025 return 0; 5026 } 5027 5028 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b) 5029 { 5030 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration; 5031 5032 /* if there's a quota refresh soon don't bother with slack */ 5033 if (runtime_refresh_within(cfs_b, min_left)) 5034 return; 5035 5036 /* don't push forwards an existing deferred unthrottle */ 5037 if (cfs_b->slack_started) 5038 return; 5039 cfs_b->slack_started = true; 5040 5041 hrtimer_start(&cfs_b->slack_timer, 5042 ns_to_ktime(cfs_bandwidth_slack_period), 5043 HRTIMER_MODE_REL); 5044 } 5045 5046 /* we know any runtime found here is valid as update_curr() precedes return */ 5047 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq) 5048 { 5049 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg); 5050 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime; 5051 5052 if (slack_runtime <= 0) 5053 return; 5054 5055 raw_spin_lock(&cfs_b->lock); 5056 if (cfs_b->quota != RUNTIME_INF) { 5057 cfs_b->runtime += slack_runtime; 5058 5059 /* we are under rq->lock, defer unthrottling using a timer */ 5060 if (cfs_b->runtime > sched_cfs_bandwidth_slice() && 5061 !list_empty(&cfs_b->throttled_cfs_rq)) 5062 start_cfs_slack_bandwidth(cfs_b); 5063 } 5064 raw_spin_unlock(&cfs_b->lock); 5065 5066 /* even if it's not valid for return we don't want to try again */ 5067 cfs_rq->runtime_remaining -= slack_runtime; 5068 } 5069 5070 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) 5071 { 5072 if (!cfs_bandwidth_used()) 5073 return; 5074 5075 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running) 5076 return; 5077 5078 __return_cfs_rq_runtime(cfs_rq); 5079 } 5080 5081 /* 5082 * This is done with a timer (instead of inline with bandwidth return) since 5083 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs. 5084 */ 5085 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b) 5086 { 5087 u64 runtime = 0, slice = sched_cfs_bandwidth_slice(); 5088 unsigned long flags; 5089 5090 /* confirm we're still not at a refresh boundary */ 5091 raw_spin_lock_irqsave(&cfs_b->lock, flags); 5092 cfs_b->slack_started = false; 5093 5094 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) { 5095 raw_spin_unlock_irqrestore(&cfs_b->lock, flags); 5096 return; 5097 } 5098 5099 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice) 5100 runtime = cfs_b->runtime; 5101 5102 raw_spin_unlock_irqrestore(&cfs_b->lock, flags); 5103 5104 if (!runtime) 5105 return; 5106 5107 distribute_cfs_runtime(cfs_b); 5108 5109 raw_spin_lock_irqsave(&cfs_b->lock, flags); 5110 raw_spin_unlock_irqrestore(&cfs_b->lock, flags); 5111 } 5112 5113 /* 5114 * When a group wakes up we want to make sure that its quota is not already 5115 * expired/exceeded, otherwise it may be allowed to steal additional ticks of 5116 * runtime as update_curr() throttling can not not trigger until it's on-rq. 5117 */ 5118 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) 5119 { 5120 if (!cfs_bandwidth_used()) 5121 return; 5122 5123 /* an active group must be handled by the update_curr()->put() path */ 5124 if (!cfs_rq->runtime_enabled || cfs_rq->curr) 5125 return; 5126 5127 /* ensure the group is not already throttled */ 5128 if (cfs_rq_throttled(cfs_rq)) 5129 return; 5130 5131 /* update runtime allocation */ 5132 account_cfs_rq_runtime(cfs_rq, 0); 5133 if (cfs_rq->runtime_remaining <= 0) 5134 throttle_cfs_rq(cfs_rq); 5135 } 5136 5137 static void sync_throttle(struct task_group *tg, int cpu) 5138 { 5139 struct cfs_rq *pcfs_rq, *cfs_rq; 5140 5141 if (!cfs_bandwidth_used()) 5142 return; 5143 5144 if (!tg->parent) 5145 return; 5146 5147 cfs_rq = tg->cfs_rq[cpu]; 5148 pcfs_rq = tg->parent->cfs_rq[cpu]; 5149 5150 cfs_rq->throttle_count = pcfs_rq->throttle_count; 5151 cfs_rq->throttled_clock_task = rq_clock_task(cpu_rq(cpu)); 5152 } 5153 5154 /* conditionally throttle active cfs_rq's from put_prev_entity() */ 5155 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) 5156 { 5157 if (!cfs_bandwidth_used()) 5158 return false; 5159 5160 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0)) 5161 return false; 5162 5163 /* 5164 * it's possible for a throttled entity to be forced into a running 5165 * state (e.g. set_curr_task), in this case we're finished. 5166 */ 5167 if (cfs_rq_throttled(cfs_rq)) 5168 return true; 5169 5170 return throttle_cfs_rq(cfs_rq); 5171 } 5172 5173 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer) 5174 { 5175 struct cfs_bandwidth *cfs_b = 5176 container_of(timer, struct cfs_bandwidth, slack_timer); 5177 5178 do_sched_cfs_slack_timer(cfs_b); 5179 5180 return HRTIMER_NORESTART; 5181 } 5182 5183 extern const u64 max_cfs_quota_period; 5184 5185 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer) 5186 { 5187 struct cfs_bandwidth *cfs_b = 5188 container_of(timer, struct cfs_bandwidth, period_timer); 5189 unsigned long flags; 5190 int overrun; 5191 int idle = 0; 5192 int count = 0; 5193 5194 raw_spin_lock_irqsave(&cfs_b->lock, flags); 5195 for (;;) { 5196 overrun = hrtimer_forward_now(timer, cfs_b->period); 5197 if (!overrun) 5198 break; 5199 5200 idle = do_sched_cfs_period_timer(cfs_b, overrun, flags); 5201 5202 if (++count > 3) { 5203 u64 new, old = ktime_to_ns(cfs_b->period); 5204 5205 /* 5206 * Grow period by a factor of 2 to avoid losing precision. 5207 * Precision loss in the quota/period ratio can cause __cfs_schedulable 5208 * to fail. 5209 */ 5210 new = old * 2; 5211 if (new < max_cfs_quota_period) { 5212 cfs_b->period = ns_to_ktime(new); 5213 cfs_b->quota *= 2; 5214 5215 pr_warn_ratelimited( 5216 "cfs_period_timer[cpu%d]: period too short, scaling up (new cfs_period_us = %lld, cfs_quota_us = %lld)\n", 5217 smp_processor_id(), 5218 div_u64(new, NSEC_PER_USEC), 5219 div_u64(cfs_b->quota, NSEC_PER_USEC)); 5220 } else { 5221 pr_warn_ratelimited( 5222 "cfs_period_timer[cpu%d]: period too short, but cannot scale up without losing precision (cfs_period_us = %lld, cfs_quota_us = %lld)\n", 5223 smp_processor_id(), 5224 div_u64(old, NSEC_PER_USEC), 5225 div_u64(cfs_b->quota, NSEC_PER_USEC)); 5226 } 5227 5228 /* reset count so we don't come right back in here */ 5229 count = 0; 5230 } 5231 } 5232 if (idle) 5233 cfs_b->period_active = 0; 5234 raw_spin_unlock_irqrestore(&cfs_b->lock, flags); 5235 5236 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART; 5237 } 5238 5239 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) 5240 { 5241 raw_spin_lock_init(&cfs_b->lock); 5242 cfs_b->runtime = 0; 5243 cfs_b->quota = RUNTIME_INF; 5244 cfs_b->period = ns_to_ktime(default_cfs_period()); 5245 5246 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq); 5247 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_PINNED); 5248 cfs_b->period_timer.function = sched_cfs_period_timer; 5249 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); 5250 cfs_b->slack_timer.function = sched_cfs_slack_timer; 5251 cfs_b->slack_started = false; 5252 } 5253 5254 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) 5255 { 5256 cfs_rq->runtime_enabled = 0; 5257 INIT_LIST_HEAD(&cfs_rq->throttled_list); 5258 } 5259 5260 void start_cfs_bandwidth(struct cfs_bandwidth *cfs_b) 5261 { 5262 lockdep_assert_held(&cfs_b->lock); 5263 5264 if (cfs_b->period_active) 5265 return; 5266 5267 cfs_b->period_active = 1; 5268 hrtimer_forward_now(&cfs_b->period_timer, cfs_b->period); 5269 hrtimer_start_expires(&cfs_b->period_timer, HRTIMER_MODE_ABS_PINNED); 5270 } 5271 5272 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) 5273 { 5274 /* init_cfs_bandwidth() was not called */ 5275 if (!cfs_b->throttled_cfs_rq.next) 5276 return; 5277 5278 hrtimer_cancel(&cfs_b->period_timer); 5279 hrtimer_cancel(&cfs_b->slack_timer); 5280 } 5281 5282 /* 5283 * Both these CPU hotplug callbacks race against unregister_fair_sched_group() 5284 * 5285 * The race is harmless, since modifying bandwidth settings of unhooked group 5286 * bits doesn't do much. 5287 */ 5288 5289 /* cpu online calback */ 5290 static void __maybe_unused update_runtime_enabled(struct rq *rq) 5291 { 5292 struct task_group *tg; 5293 5294 lockdep_assert_held(&rq->lock); 5295 5296 rcu_read_lock(); 5297 list_for_each_entry_rcu(tg, &task_groups, list) { 5298 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth; 5299 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)]; 5300 5301 raw_spin_lock(&cfs_b->lock); 5302 cfs_rq->runtime_enabled = cfs_b->quota != RUNTIME_INF; 5303 raw_spin_unlock(&cfs_b->lock); 5304 } 5305 rcu_read_unlock(); 5306 } 5307 5308 /* cpu offline callback */ 5309 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq) 5310 { 5311 struct task_group *tg; 5312 5313 lockdep_assert_held(&rq->lock); 5314 5315 rcu_read_lock(); 5316 list_for_each_entry_rcu(tg, &task_groups, list) { 5317 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)]; 5318 5319 if (!cfs_rq->runtime_enabled) 5320 continue; 5321 5322 /* 5323 * clock_task is not advancing so we just need to make sure 5324 * there's some valid quota amount 5325 */ 5326 cfs_rq->runtime_remaining = 1; 5327 /* 5328 * Offline rq is schedulable till CPU is completely disabled 5329 * in take_cpu_down(), so we prevent new cfs throttling here. 5330 */ 5331 cfs_rq->runtime_enabled = 0; 5332 5333 if (cfs_rq_throttled(cfs_rq)) 5334 unthrottle_cfs_rq(cfs_rq); 5335 } 5336 rcu_read_unlock(); 5337 } 5338 5339 #else /* CONFIG_CFS_BANDWIDTH */ 5340 5341 static inline bool cfs_bandwidth_used(void) 5342 { 5343 return false; 5344 } 5345 5346 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) {} 5347 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) { return false; } 5348 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {} 5349 static inline void sync_throttle(struct task_group *tg, int cpu) {} 5350 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {} 5351 5352 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq) 5353 { 5354 return 0; 5355 } 5356 5357 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq) 5358 { 5359 return 0; 5360 } 5361 5362 static inline int throttled_lb_pair(struct task_group *tg, 5363 int src_cpu, int dest_cpu) 5364 { 5365 return 0; 5366 } 5367 5368 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {} 5369 5370 #ifdef CONFIG_FAIR_GROUP_SCHED 5371 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {} 5372 #endif 5373 5374 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg) 5375 { 5376 return NULL; 5377 } 5378 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {} 5379 static inline void update_runtime_enabled(struct rq *rq) {} 5380 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {} 5381 5382 #endif /* CONFIG_CFS_BANDWIDTH */ 5383 5384 /************************************************** 5385 * CFS operations on tasks: 5386 */ 5387 5388 #ifdef CONFIG_SCHED_HRTICK 5389 static void hrtick_start_fair(struct rq *rq, struct task_struct *p) 5390 { 5391 struct sched_entity *se = &p->se; 5392 struct cfs_rq *cfs_rq = cfs_rq_of(se); 5393 5394 SCHED_WARN_ON(task_rq(p) != rq); 5395 5396 if (rq->cfs.h_nr_running > 1) { 5397 u64 slice = sched_slice(cfs_rq, se); 5398 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime; 5399 s64 delta = slice - ran; 5400 5401 if (delta < 0) { 5402 if (rq->curr == p) 5403 resched_curr(rq); 5404 return; 5405 } 5406 hrtick_start(rq, delta); 5407 } 5408 } 5409 5410 /* 5411 * called from enqueue/dequeue and updates the hrtick when the 5412 * current task is from our class and nr_running is low enough 5413 * to matter. 5414 */ 5415 static void hrtick_update(struct rq *rq) 5416 { 5417 struct task_struct *curr = rq->curr; 5418 5419 if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class) 5420 return; 5421 5422 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency) 5423 hrtick_start_fair(rq, curr); 5424 } 5425 #else /* !CONFIG_SCHED_HRTICK */ 5426 static inline void 5427 hrtick_start_fair(struct rq *rq, struct task_struct *p) 5428 { 5429 } 5430 5431 static inline void hrtick_update(struct rq *rq) 5432 { 5433 } 5434 #endif 5435 5436 #ifdef CONFIG_SMP 5437 static inline unsigned long cpu_util(int cpu); 5438 5439 static inline bool cpu_overutilized(int cpu) 5440 { 5441 return !fits_capacity(cpu_util(cpu), capacity_of(cpu)); 5442 } 5443 5444 static inline void update_overutilized_status(struct rq *rq) 5445 { 5446 if (!READ_ONCE(rq->rd->overutilized) && cpu_overutilized(rq->cpu)) { 5447 WRITE_ONCE(rq->rd->overutilized, SG_OVERUTILIZED); 5448 trace_sched_overutilized_tp(rq->rd, SG_OVERUTILIZED); 5449 } 5450 } 5451 #else 5452 static inline void update_overutilized_status(struct rq *rq) { } 5453 #endif 5454 5455 /* Runqueue only has SCHED_IDLE tasks enqueued */ 5456 static int sched_idle_rq(struct rq *rq) 5457 { 5458 return unlikely(rq->nr_running == rq->cfs.idle_h_nr_running && 5459 rq->nr_running); 5460 } 5461 5462 #ifdef CONFIG_SMP 5463 static int sched_idle_cpu(int cpu) 5464 { 5465 return sched_idle_rq(cpu_rq(cpu)); 5466 } 5467 #endif 5468 5469 /* 5470 * The enqueue_task method is called before nr_running is 5471 * increased. Here we update the fair scheduling stats and 5472 * then put the task into the rbtree: 5473 */ 5474 static void 5475 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags) 5476 { 5477 struct cfs_rq *cfs_rq; 5478 struct sched_entity *se = &p->se; 5479 int idle_h_nr_running = task_has_idle_policy(p); 5480 5481 /* 5482 * The code below (indirectly) updates schedutil which looks at 5483 * the cfs_rq utilization to select a frequency. 5484 * Let's add the task's estimated utilization to the cfs_rq's 5485 * estimated utilization, before we update schedutil. 5486 */ 5487 util_est_enqueue(&rq->cfs, p); 5488 5489 /* 5490 * If in_iowait is set, the code below may not trigger any cpufreq 5491 * utilization updates, so do it here explicitly with the IOWAIT flag 5492 * passed. 5493 */ 5494 if (p->in_iowait) 5495 cpufreq_update_util(rq, SCHED_CPUFREQ_IOWAIT); 5496 5497 for_each_sched_entity(se) { 5498 if (se->on_rq) 5499 break; 5500 cfs_rq = cfs_rq_of(se); 5501 enqueue_entity(cfs_rq, se, flags); 5502 5503 cfs_rq->h_nr_running++; 5504 cfs_rq->idle_h_nr_running += idle_h_nr_running; 5505 5506 /* end evaluation on encountering a throttled cfs_rq */ 5507 if (cfs_rq_throttled(cfs_rq)) 5508 goto enqueue_throttle; 5509 5510 flags = ENQUEUE_WAKEUP; 5511 } 5512 5513 for_each_sched_entity(se) { 5514 cfs_rq = cfs_rq_of(se); 5515 5516 update_load_avg(cfs_rq, se, UPDATE_TG); 5517 se_update_runnable(se); 5518 update_cfs_group(se); 5519 5520 cfs_rq->h_nr_running++; 5521 cfs_rq->idle_h_nr_running += idle_h_nr_running; 5522 5523 /* end evaluation on encountering a throttled cfs_rq */ 5524 if (cfs_rq_throttled(cfs_rq)) 5525 goto enqueue_throttle; 5526 5527 /* 5528 * One parent has been throttled and cfs_rq removed from the 5529 * list. Add it back to not break the leaf list. 5530 */ 5531 if (throttled_hierarchy(cfs_rq)) 5532 list_add_leaf_cfs_rq(cfs_rq); 5533 } 5534 5535 /* At this point se is NULL and we are at root level*/ 5536 add_nr_running(rq, 1); 5537 5538 /* 5539 * Since new tasks are assigned an initial util_avg equal to 5540 * half of the spare capacity of their CPU, tiny tasks have the 5541 * ability to cross the overutilized threshold, which will 5542 * result in the load balancer ruining all the task placement 5543 * done by EAS. As a way to mitigate that effect, do not account 5544 * for the first enqueue operation of new tasks during the 5545 * overutilized flag detection. 5546 * 5547 * A better way of solving this problem would be to wait for 5548 * the PELT signals of tasks to converge before taking them 5549 * into account, but that is not straightforward to implement, 5550 * and the following generally works well enough in practice. 5551 */ 5552 if (flags & ENQUEUE_WAKEUP) 5553 update_overutilized_status(rq); 5554 5555 enqueue_throttle: 5556 if (cfs_bandwidth_used()) { 5557 /* 5558 * When bandwidth control is enabled; the cfs_rq_throttled() 5559 * breaks in the above iteration can result in incomplete 5560 * leaf list maintenance, resulting in triggering the assertion 5561 * below. 5562 */ 5563 for_each_sched_entity(se) { 5564 cfs_rq = cfs_rq_of(se); 5565 5566 if (list_add_leaf_cfs_rq(cfs_rq)) 5567 break; 5568 } 5569 } 5570 5571 assert_list_leaf_cfs_rq(rq); 5572 5573 hrtick_update(rq); 5574 } 5575 5576 static void set_next_buddy(struct sched_entity *se); 5577 5578 /* 5579 * The dequeue_task method is called before nr_running is 5580 * decreased. We remove the task from the rbtree and 5581 * update the fair scheduling stats: 5582 */ 5583 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags) 5584 { 5585 struct cfs_rq *cfs_rq; 5586 struct sched_entity *se = &p->se; 5587 int task_sleep = flags & DEQUEUE_SLEEP; 5588 int idle_h_nr_running = task_has_idle_policy(p); 5589 bool was_sched_idle = sched_idle_rq(rq); 5590 5591 for_each_sched_entity(se) { 5592 cfs_rq = cfs_rq_of(se); 5593 dequeue_entity(cfs_rq, se, flags); 5594 5595 cfs_rq->h_nr_running--; 5596 cfs_rq->idle_h_nr_running -= idle_h_nr_running; 5597 5598 /* end evaluation on encountering a throttled cfs_rq */ 5599 if (cfs_rq_throttled(cfs_rq)) 5600 goto dequeue_throttle; 5601 5602 /* Don't dequeue parent if it has other entities besides us */ 5603 if (cfs_rq->load.weight) { 5604 /* Avoid re-evaluating load for this entity: */ 5605 se = parent_entity(se); 5606 /* 5607 * Bias pick_next to pick a task from this cfs_rq, as 5608 * p is sleeping when it is within its sched_slice. 5609 */ 5610 if (task_sleep && se && !throttled_hierarchy(cfs_rq)) 5611 set_next_buddy(se); 5612 break; 5613 } 5614 flags |= DEQUEUE_SLEEP; 5615 } 5616 5617 for_each_sched_entity(se) { 5618 cfs_rq = cfs_rq_of(se); 5619 5620 update_load_avg(cfs_rq, se, UPDATE_TG); 5621 se_update_runnable(se); 5622 update_cfs_group(se); 5623 5624 cfs_rq->h_nr_running--; 5625 cfs_rq->idle_h_nr_running -= idle_h_nr_running; 5626 5627 /* end evaluation on encountering a throttled cfs_rq */ 5628 if (cfs_rq_throttled(cfs_rq)) 5629 goto dequeue_throttle; 5630 5631 } 5632 5633 /* At this point se is NULL and we are at root level*/ 5634 sub_nr_running(rq, 1); 5635 5636 /* balance early to pull high priority tasks */ 5637 if (unlikely(!was_sched_idle && sched_idle_rq(rq))) 5638 rq->next_balance = jiffies; 5639 5640 dequeue_throttle: 5641 util_est_dequeue(&rq->cfs, p, task_sleep); 5642 hrtick_update(rq); 5643 } 5644 5645 #ifdef CONFIG_SMP 5646 5647 /* Working cpumask for: load_balance, load_balance_newidle. */ 5648 DEFINE_PER_CPU(cpumask_var_t, load_balance_mask); 5649 DEFINE_PER_CPU(cpumask_var_t, select_idle_mask); 5650 5651 #ifdef CONFIG_NO_HZ_COMMON 5652 5653 static struct { 5654 cpumask_var_t idle_cpus_mask; 5655 atomic_t nr_cpus; 5656 int has_blocked; /* Idle CPUS has blocked load */ 5657 unsigned long next_balance; /* in jiffy units */ 5658 unsigned long next_blocked; /* Next update of blocked load in jiffies */ 5659 } nohz ____cacheline_aligned; 5660 5661 #endif /* CONFIG_NO_HZ_COMMON */ 5662 5663 static unsigned long cpu_load(struct rq *rq) 5664 { 5665 return cfs_rq_load_avg(&rq->cfs); 5666 } 5667 5668 /* 5669 * cpu_load_without - compute CPU load without any contributions from *p 5670 * @cpu: the CPU which load is requested 5671 * @p: the task which load should be discounted 5672 * 5673 * The load of a CPU is defined by the load of tasks currently enqueued on that 5674 * CPU as well as tasks which are currently sleeping after an execution on that 5675 * CPU. 5676 * 5677 * This method returns the load of the specified CPU by discounting the load of 5678 * the specified task, whenever the task is currently contributing to the CPU 5679 * load. 5680 */ 5681 static unsigned long cpu_load_without(struct rq *rq, struct task_struct *p) 5682 { 5683 struct cfs_rq *cfs_rq; 5684 unsigned int load; 5685 5686 /* Task has no contribution or is new */ 5687 if (cpu_of(rq) != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time)) 5688 return cpu_load(rq); 5689 5690 cfs_rq = &rq->cfs; 5691 load = READ_ONCE(cfs_rq->avg.load_avg); 5692 5693 /* Discount task's util from CPU's util */ 5694 lsub_positive(&load, task_h_load(p)); 5695 5696 return load; 5697 } 5698 5699 static unsigned long cpu_runnable(struct rq *rq) 5700 { 5701 return cfs_rq_runnable_avg(&rq->cfs); 5702 } 5703 5704 static unsigned long cpu_runnable_without(struct rq *rq, struct task_struct *p) 5705 { 5706 struct cfs_rq *cfs_rq; 5707 unsigned int runnable; 5708 5709 /* Task has no contribution or is new */ 5710 if (cpu_of(rq) != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time)) 5711 return cpu_runnable(rq); 5712 5713 cfs_rq = &rq->cfs; 5714 runnable = READ_ONCE(cfs_rq->avg.runnable_avg); 5715 5716 /* Discount task's runnable from CPU's runnable */ 5717 lsub_positive(&runnable, p->se.avg.runnable_avg); 5718 5719 return runnable; 5720 } 5721 5722 static unsigned long capacity_of(int cpu) 5723 { 5724 return cpu_rq(cpu)->cpu_capacity; 5725 } 5726 5727 static void record_wakee(struct task_struct *p) 5728 { 5729 /* 5730 * Only decay a single time; tasks that have less then 1 wakeup per 5731 * jiffy will not have built up many flips. 5732 */ 5733 if (time_after(jiffies, current->wakee_flip_decay_ts + HZ)) { 5734 current->wakee_flips >>= 1; 5735 current->wakee_flip_decay_ts = jiffies; 5736 } 5737 5738 if (current->last_wakee != p) { 5739 current->last_wakee = p; 5740 current->wakee_flips++; 5741 } 5742 } 5743 5744 /* 5745 * Detect M:N waker/wakee relationships via a switching-frequency heuristic. 5746 * 5747 * A waker of many should wake a different task than the one last awakened 5748 * at a frequency roughly N times higher than one of its wakees. 5749 * 5750 * In order to determine whether we should let the load spread vs consolidating 5751 * to shared cache, we look for a minimum 'flip' frequency of llc_size in one 5752 * partner, and a factor of lls_size higher frequency in the other. 5753 * 5754 * With both conditions met, we can be relatively sure that the relationship is 5755 * non-monogamous, with partner count exceeding socket size. 5756 * 5757 * Waker/wakee being client/server, worker/dispatcher, interrupt source or 5758 * whatever is irrelevant, spread criteria is apparent partner count exceeds 5759 * socket size. 5760 */ 5761 static int wake_wide(struct task_struct *p) 5762 { 5763 unsigned int master = current->wakee_flips; 5764 unsigned int slave = p->wakee_flips; 5765 int factor = __this_cpu_read(sd_llc_size); 5766 5767 if (master < slave) 5768 swap(master, slave); 5769 if (slave < factor || master < slave * factor) 5770 return 0; 5771 return 1; 5772 } 5773 5774 /* 5775 * The purpose of wake_affine() is to quickly determine on which CPU we can run 5776 * soonest. For the purpose of speed we only consider the waking and previous 5777 * CPU. 5778 * 5779 * wake_affine_idle() - only considers 'now', it check if the waking CPU is 5780 * cache-affine and is (or will be) idle. 5781 * 5782 * wake_affine_weight() - considers the weight to reflect the average 5783 * scheduling latency of the CPUs. This seems to work 5784 * for the overloaded case. 5785 */ 5786 static int 5787 wake_affine_idle(int this_cpu, int prev_cpu, int sync) 5788 { 5789 /* 5790 * If this_cpu is idle, it implies the wakeup is from interrupt 5791 * context. Only allow the move if cache is shared. Otherwise an 5792 * interrupt intensive workload could force all tasks onto one 5793 * node depending on the IO topology or IRQ affinity settings. 5794 * 5795 * If the prev_cpu is idle and cache affine then avoid a migration. 5796 * There is no guarantee that the cache hot data from an interrupt 5797 * is more important than cache hot data on the prev_cpu and from 5798 * a cpufreq perspective, it's better to have higher utilisation 5799 * on one CPU. 5800 */ 5801 if (available_idle_cpu(this_cpu) && cpus_share_cache(this_cpu, prev_cpu)) 5802 return available_idle_cpu(prev_cpu) ? prev_cpu : this_cpu; 5803 5804 if (sync && cpu_rq(this_cpu)->nr_running == 1) 5805 return this_cpu; 5806 5807 return nr_cpumask_bits; 5808 } 5809 5810 static int 5811 wake_affine_weight(struct sched_domain *sd, struct task_struct *p, 5812 int this_cpu, int prev_cpu, int sync) 5813 { 5814 s64 this_eff_load, prev_eff_load; 5815 unsigned long task_load; 5816 5817 this_eff_load = cpu_load(cpu_rq(this_cpu)); 5818 5819 if (sync) { 5820 unsigned long current_load = task_h_load(current); 5821 5822 if (current_load > this_eff_load) 5823 return this_cpu; 5824 5825 this_eff_load -= current_load; 5826 } 5827 5828 task_load = task_h_load(p); 5829 5830 this_eff_load += task_load; 5831 if (sched_feat(WA_BIAS)) 5832 this_eff_load *= 100; 5833 this_eff_load *= capacity_of(prev_cpu); 5834 5835 prev_eff_load = cpu_load(cpu_rq(prev_cpu)); 5836 prev_eff_load -= task_load; 5837 if (sched_feat(WA_BIAS)) 5838 prev_eff_load *= 100 + (sd->imbalance_pct - 100) / 2; 5839 prev_eff_load *= capacity_of(this_cpu); 5840 5841 /* 5842 * If sync, adjust the weight of prev_eff_load such that if 5843 * prev_eff == this_eff that select_idle_sibling() will consider 5844 * stacking the wakee on top of the waker if no other CPU is 5845 * idle. 5846 */ 5847 if (sync) 5848 prev_eff_load += 1; 5849 5850 return this_eff_load < prev_eff_load ? this_cpu : nr_cpumask_bits; 5851 } 5852 5853 static int wake_affine(struct sched_domain *sd, struct task_struct *p, 5854 int this_cpu, int prev_cpu, int sync) 5855 { 5856 int target = nr_cpumask_bits; 5857 5858 if (sched_feat(WA_IDLE)) 5859 target = wake_affine_idle(this_cpu, prev_cpu, sync); 5860 5861 if (sched_feat(WA_WEIGHT) && target == nr_cpumask_bits) 5862 target = wake_affine_weight(sd, p, this_cpu, prev_cpu, sync); 5863 5864 schedstat_inc(p->se.statistics.nr_wakeups_affine_attempts); 5865 if (target == nr_cpumask_bits) 5866 return prev_cpu; 5867 5868 schedstat_inc(sd->ttwu_move_affine); 5869 schedstat_inc(p->se.statistics.nr_wakeups_affine); 5870 return target; 5871 } 5872 5873 static struct sched_group * 5874 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu); 5875 5876 /* 5877 * find_idlest_group_cpu - find the idlest CPU among the CPUs in the group. 5878 */ 5879 static int 5880 find_idlest_group_cpu(struct sched_group *group, struct task_struct *p, int this_cpu) 5881 { 5882 unsigned long load, min_load = ULONG_MAX; 5883 unsigned int min_exit_latency = UINT_MAX; 5884 u64 latest_idle_timestamp = 0; 5885 int least_loaded_cpu = this_cpu; 5886 int shallowest_idle_cpu = -1; 5887 int i; 5888 5889 /* Check if we have any choice: */ 5890 if (group->group_weight == 1) 5891 return cpumask_first(sched_group_span(group)); 5892 5893 /* Traverse only the allowed CPUs */ 5894 for_each_cpu_and(i, sched_group_span(group), p->cpus_ptr) { 5895 if (sched_idle_cpu(i)) 5896 return i; 5897 5898 if (available_idle_cpu(i)) { 5899 struct rq *rq = cpu_rq(i); 5900 struct cpuidle_state *idle = idle_get_state(rq); 5901 if (idle && idle->exit_latency < min_exit_latency) { 5902 /* 5903 * We give priority to a CPU whose idle state 5904 * has the smallest exit latency irrespective 5905 * of any idle timestamp. 5906 */ 5907 min_exit_latency = idle->exit_latency; 5908 latest_idle_timestamp = rq->idle_stamp; 5909 shallowest_idle_cpu = i; 5910 } else if ((!idle || idle->exit_latency == min_exit_latency) && 5911 rq->idle_stamp > latest_idle_timestamp) { 5912 /* 5913 * If equal or no active idle state, then 5914 * the most recently idled CPU might have 5915 * a warmer cache. 5916 */ 5917 latest_idle_timestamp = rq->idle_stamp; 5918 shallowest_idle_cpu = i; 5919 } 5920 } else if (shallowest_idle_cpu == -1) { 5921 load = cpu_load(cpu_rq(i)); 5922 if (load < min_load) { 5923 min_load = load; 5924 least_loaded_cpu = i; 5925 } 5926 } 5927 } 5928 5929 return shallowest_idle_cpu != -1 ? shallowest_idle_cpu : least_loaded_cpu; 5930 } 5931 5932 static inline int find_idlest_cpu(struct sched_domain *sd, struct task_struct *p, 5933 int cpu, int prev_cpu, int sd_flag) 5934 { 5935 int new_cpu = cpu; 5936 5937 if (!cpumask_intersects(sched_domain_span(sd), p->cpus_ptr)) 5938 return prev_cpu; 5939 5940 /* 5941 * We need task's util for cpu_util_without, sync it up to 5942 * prev_cpu's last_update_time. 5943 */ 5944 if (!(sd_flag & SD_BALANCE_FORK)) 5945 sync_entity_load_avg(&p->se); 5946 5947 while (sd) { 5948 struct sched_group *group; 5949 struct sched_domain *tmp; 5950 int weight; 5951 5952 if (!(sd->flags & sd_flag)) { 5953 sd = sd->child; 5954 continue; 5955 } 5956 5957 group = find_idlest_group(sd, p, cpu); 5958 if (!group) { 5959 sd = sd->child; 5960 continue; 5961 } 5962 5963 new_cpu = find_idlest_group_cpu(group, p, cpu); 5964 if (new_cpu == cpu) { 5965 /* Now try balancing at a lower domain level of 'cpu': */ 5966 sd = sd->child; 5967 continue; 5968 } 5969 5970 /* Now try balancing at a lower domain level of 'new_cpu': */ 5971 cpu = new_cpu; 5972 weight = sd->span_weight; 5973 sd = NULL; 5974 for_each_domain(cpu, tmp) { 5975 if (weight <= tmp->span_weight) 5976 break; 5977 if (tmp->flags & sd_flag) 5978 sd = tmp; 5979 } 5980 } 5981 5982 return new_cpu; 5983 } 5984 5985 #ifdef CONFIG_SCHED_SMT 5986 DEFINE_STATIC_KEY_FALSE(sched_smt_present); 5987 EXPORT_SYMBOL_GPL(sched_smt_present); 5988 5989 static inline void set_idle_cores(int cpu, int val) 5990 { 5991 struct sched_domain_shared *sds; 5992 5993 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu)); 5994 if (sds) 5995 WRITE_ONCE(sds->has_idle_cores, val); 5996 } 5997 5998 static inline bool test_idle_cores(int cpu, bool def) 5999 { 6000 struct sched_domain_shared *sds; 6001 6002 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu)); 6003 if (sds) 6004 return READ_ONCE(sds->has_idle_cores); 6005 6006 return def; 6007 } 6008 6009 /* 6010 * Scans the local SMT mask to see if the entire core is idle, and records this 6011 * information in sd_llc_shared->has_idle_cores. 6012 * 6013 * Since SMT siblings share all cache levels, inspecting this limited remote 6014 * state should be fairly cheap. 6015 */ 6016 void __update_idle_core(struct rq *rq) 6017 { 6018 int core = cpu_of(rq); 6019 int cpu; 6020 6021 rcu_read_lock(); 6022 if (test_idle_cores(core, true)) 6023 goto unlock; 6024 6025 for_each_cpu(cpu, cpu_smt_mask(core)) { 6026 if (cpu == core) 6027 continue; 6028 6029 if (!available_idle_cpu(cpu)) 6030 goto unlock; 6031 } 6032 6033 set_idle_cores(core, 1); 6034 unlock: 6035 rcu_read_unlock(); 6036 } 6037 6038 /* 6039 * Scan the entire LLC domain for idle cores; this dynamically switches off if 6040 * there are no idle cores left in the system; tracked through 6041 * sd_llc->shared->has_idle_cores and enabled through update_idle_core() above. 6042 */ 6043 static int select_idle_core(struct task_struct *p, struct sched_domain *sd, int target) 6044 { 6045 struct cpumask *cpus = this_cpu_cpumask_var_ptr(select_idle_mask); 6046 int core, cpu; 6047 6048 if (!static_branch_likely(&sched_smt_present)) 6049 return -1; 6050 6051 if (!test_idle_cores(target, false)) 6052 return -1; 6053 6054 cpumask_and(cpus, sched_domain_span(sd), p->cpus_ptr); 6055 6056 for_each_cpu_wrap(core, cpus, target) { 6057 bool idle = true; 6058 6059 for_each_cpu(cpu, cpu_smt_mask(core)) { 6060 if (!available_idle_cpu(cpu)) { 6061 idle = false; 6062 break; 6063 } 6064 } 6065 cpumask_andnot(cpus, cpus, cpu_smt_mask(core)); 6066 6067 if (idle) 6068 return core; 6069 } 6070 6071 /* 6072 * Failed to find an idle core; stop looking for one. 6073 */ 6074 set_idle_cores(target, 0); 6075 6076 return -1; 6077 } 6078 6079 /* 6080 * Scan the local SMT mask for idle CPUs. 6081 */ 6082 static int select_idle_smt(struct task_struct *p, struct sched_domain *sd, int target) 6083 { 6084 int cpu; 6085 6086 if (!static_branch_likely(&sched_smt_present)) 6087 return -1; 6088 6089 for_each_cpu(cpu, cpu_smt_mask(target)) { 6090 if (!cpumask_test_cpu(cpu, p->cpus_ptr) || 6091 !cpumask_test_cpu(cpu, sched_domain_span(sd))) 6092 continue; 6093 if (available_idle_cpu(cpu) || sched_idle_cpu(cpu)) 6094 return cpu; 6095 } 6096 6097 return -1; 6098 } 6099 6100 #else /* CONFIG_SCHED_SMT */ 6101 6102 static inline int select_idle_core(struct task_struct *p, struct sched_domain *sd, int target) 6103 { 6104 return -1; 6105 } 6106 6107 static inline int select_idle_smt(struct task_struct *p, struct sched_domain *sd, int target) 6108 { 6109 return -1; 6110 } 6111 6112 #endif /* CONFIG_SCHED_SMT */ 6113 6114 /* 6115 * Scan the LLC domain for idle CPUs; this is dynamically regulated by 6116 * comparing the average scan cost (tracked in sd->avg_scan_cost) against the 6117 * average idle time for this rq (as found in rq->avg_idle). 6118 */ 6119 static int select_idle_cpu(struct task_struct *p, struct sched_domain *sd, int target) 6120 { 6121 struct cpumask *cpus = this_cpu_cpumask_var_ptr(select_idle_mask); 6122 struct sched_domain *this_sd; 6123 u64 avg_cost, avg_idle; 6124 u64 time; 6125 int this = smp_processor_id(); 6126 int cpu, nr = INT_MAX; 6127 6128 this_sd = rcu_dereference(*this_cpu_ptr(&sd_llc)); 6129 if (!this_sd) 6130 return -1; 6131 6132 /* 6133 * Due to large variance we need a large fuzz factor; hackbench in 6134 * particularly is sensitive here. 6135 */ 6136 avg_idle = this_rq()->avg_idle / 512; 6137 avg_cost = this_sd->avg_scan_cost + 1; 6138 6139 if (sched_feat(SIS_AVG_CPU) && avg_idle < avg_cost) 6140 return -1; 6141 6142 if (sched_feat(SIS_PROP)) { 6143 u64 span_avg = sd->span_weight * avg_idle; 6144 if (span_avg > 4*avg_cost) 6145 nr = div_u64(span_avg, avg_cost); 6146 else 6147 nr = 4; 6148 } 6149 6150 time = cpu_clock(this); 6151 6152 cpumask_and(cpus, sched_domain_span(sd), p->cpus_ptr); 6153 6154 for_each_cpu_wrap(cpu, cpus, target) { 6155 if (!--nr) 6156 return -1; 6157 if (available_idle_cpu(cpu) || sched_idle_cpu(cpu)) 6158 break; 6159 } 6160 6161 time = cpu_clock(this) - time; 6162 update_avg(&this_sd->avg_scan_cost, time); 6163 6164 return cpu; 6165 } 6166 6167 /* 6168 * Scan the asym_capacity domain for idle CPUs; pick the first idle one on which 6169 * the task fits. If no CPU is big enough, but there are idle ones, try to 6170 * maximize capacity. 6171 */ 6172 static int 6173 select_idle_capacity(struct task_struct *p, struct sched_domain *sd, int target) 6174 { 6175 unsigned long task_util, best_cap = 0; 6176 int cpu, best_cpu = -1; 6177 struct cpumask *cpus; 6178 6179 cpus = this_cpu_cpumask_var_ptr(select_idle_mask); 6180 cpumask_and(cpus, sched_domain_span(sd), p->cpus_ptr); 6181 6182 task_util = uclamp_task_util(p); 6183 6184 for_each_cpu_wrap(cpu, cpus, target) { 6185 unsigned long cpu_cap = capacity_of(cpu); 6186 6187 if (!available_idle_cpu(cpu) && !sched_idle_cpu(cpu)) 6188 continue; 6189 if (fits_capacity(task_util, cpu_cap)) 6190 return cpu; 6191 6192 if (cpu_cap > best_cap) { 6193 best_cap = cpu_cap; 6194 best_cpu = cpu; 6195 } 6196 } 6197 6198 return best_cpu; 6199 } 6200 6201 static inline bool asym_fits_capacity(int task_util, int cpu) 6202 { 6203 if (static_branch_unlikely(&sched_asym_cpucapacity)) 6204 return fits_capacity(task_util, capacity_of(cpu)); 6205 6206 return true; 6207 } 6208 6209 /* 6210 * Try and locate an idle core/thread in the LLC cache domain. 6211 */ 6212 static int select_idle_sibling(struct task_struct *p, int prev, int target) 6213 { 6214 struct sched_domain *sd; 6215 unsigned long task_util; 6216 int i, recent_used_cpu; 6217 6218 /* 6219 * On asymmetric system, update task utilization because we will check 6220 * that the task fits with cpu's capacity. 6221 */ 6222 if (static_branch_unlikely(&sched_asym_cpucapacity)) { 6223 sync_entity_load_avg(&p->se); 6224 task_util = uclamp_task_util(p); 6225 } 6226 6227 if ((available_idle_cpu(target) || sched_idle_cpu(target)) && 6228 asym_fits_capacity(task_util, target)) 6229 return target; 6230 6231 /* 6232 * If the previous CPU is cache affine and idle, don't be stupid: 6233 */ 6234 if (prev != target && cpus_share_cache(prev, target) && 6235 (available_idle_cpu(prev) || sched_idle_cpu(prev)) && 6236 asym_fits_capacity(task_util, prev)) 6237 return prev; 6238 6239 /* 6240 * Allow a per-cpu kthread to stack with the wakee if the 6241 * kworker thread and the tasks previous CPUs are the same. 6242 * The assumption is that the wakee queued work for the 6243 * per-cpu kthread that is now complete and the wakeup is 6244 * essentially a sync wakeup. An obvious example of this 6245 * pattern is IO completions. 6246 */ 6247 if (is_per_cpu_kthread(current) && 6248 prev == smp_processor_id() && 6249 this_rq()->nr_running <= 1) { 6250 return prev; 6251 } 6252 6253 /* Check a recently used CPU as a potential idle candidate: */ 6254 recent_used_cpu = p->recent_used_cpu; 6255 if (recent_used_cpu != prev && 6256 recent_used_cpu != target && 6257 cpus_share_cache(recent_used_cpu, target) && 6258 (available_idle_cpu(recent_used_cpu) || sched_idle_cpu(recent_used_cpu)) && 6259 cpumask_test_cpu(p->recent_used_cpu, p->cpus_ptr) && 6260 asym_fits_capacity(task_util, recent_used_cpu)) { 6261 /* 6262 * Replace recent_used_cpu with prev as it is a potential 6263 * candidate for the next wake: 6264 */ 6265 p->recent_used_cpu = prev; 6266 return recent_used_cpu; 6267 } 6268 6269 /* 6270 * For asymmetric CPU capacity systems, our domain of interest is 6271 * sd_asym_cpucapacity rather than sd_llc. 6272 */ 6273 if (static_branch_unlikely(&sched_asym_cpucapacity)) { 6274 sd = rcu_dereference(per_cpu(sd_asym_cpucapacity, target)); 6275 /* 6276 * On an asymmetric CPU capacity system where an exclusive 6277 * cpuset defines a symmetric island (i.e. one unique 6278 * capacity_orig value through the cpuset), the key will be set 6279 * but the CPUs within that cpuset will not have a domain with 6280 * SD_ASYM_CPUCAPACITY. These should follow the usual symmetric 6281 * capacity path. 6282 */ 6283 if (sd) { 6284 i = select_idle_capacity(p, sd, target); 6285 return ((unsigned)i < nr_cpumask_bits) ? i : target; 6286 } 6287 } 6288 6289 sd = rcu_dereference(per_cpu(sd_llc, target)); 6290 if (!sd) 6291 return target; 6292 6293 i = select_idle_core(p, sd, target); 6294 if ((unsigned)i < nr_cpumask_bits) 6295 return i; 6296 6297 i = select_idle_cpu(p, sd, target); 6298 if ((unsigned)i < nr_cpumask_bits) 6299 return i; 6300 6301 i = select_idle_smt(p, sd, target); 6302 if ((unsigned)i < nr_cpumask_bits) 6303 return i; 6304 6305 return target; 6306 } 6307 6308 /** 6309 * Amount of capacity of a CPU that is (estimated to be) used by CFS tasks 6310 * @cpu: the CPU to get the utilization of 6311 * 6312 * The unit of the return value must be the one of capacity so we can compare 6313 * the utilization with the capacity of the CPU that is available for CFS task 6314 * (ie cpu_capacity). 6315 * 6316 * cfs_rq.avg.util_avg is the sum of running time of runnable tasks plus the 6317 * recent utilization of currently non-runnable tasks on a CPU. It represents 6318 * the amount of utilization of a CPU in the range [0..capacity_orig] where 6319 * capacity_orig is the cpu_capacity available at the highest frequency 6320 * (arch_scale_freq_capacity()). 6321 * The utilization of a CPU converges towards a sum equal to or less than the 6322 * current capacity (capacity_curr <= capacity_orig) of the CPU because it is 6323 * the running time on this CPU scaled by capacity_curr. 6324 * 6325 * The estimated utilization of a CPU is defined to be the maximum between its 6326 * cfs_rq.avg.util_avg and the sum of the estimated utilization of the tasks 6327 * currently RUNNABLE on that CPU. 6328 * This allows to properly represent the expected utilization of a CPU which 6329 * has just got a big task running since a long sleep period. At the same time 6330 * however it preserves the benefits of the "blocked utilization" in 6331 * describing the potential for other tasks waking up on the same CPU. 6332 * 6333 * Nevertheless, cfs_rq.avg.util_avg can be higher than capacity_curr or even 6334 * higher than capacity_orig because of unfortunate rounding in 6335 * cfs.avg.util_avg or just after migrating tasks and new task wakeups until 6336 * the average stabilizes with the new running time. We need to check that the 6337 * utilization stays within the range of [0..capacity_orig] and cap it if 6338 * necessary. Without utilization capping, a group could be seen as overloaded 6339 * (CPU0 utilization at 121% + CPU1 utilization at 80%) whereas CPU1 has 20% of 6340 * available capacity. We allow utilization to overshoot capacity_curr (but not 6341 * capacity_orig) as it useful for predicting the capacity required after task 6342 * migrations (scheduler-driven DVFS). 6343 * 6344 * Return: the (estimated) utilization for the specified CPU 6345 */ 6346 static inline unsigned long cpu_util(int cpu) 6347 { 6348 struct cfs_rq *cfs_rq; 6349 unsigned int util; 6350 6351 cfs_rq = &cpu_rq(cpu)->cfs; 6352 util = READ_ONCE(cfs_rq->avg.util_avg); 6353 6354 if (sched_feat(UTIL_EST)) 6355 util = max(util, READ_ONCE(cfs_rq->avg.util_est.enqueued)); 6356 6357 return min_t(unsigned long, util, capacity_orig_of(cpu)); 6358 } 6359 6360 /* 6361 * cpu_util_without: compute cpu utilization without any contributions from *p 6362 * @cpu: the CPU which utilization is requested 6363 * @p: the task which utilization should be discounted 6364 * 6365 * The utilization of a CPU is defined by the utilization of tasks currently 6366 * enqueued on that CPU as well as tasks which are currently sleeping after an 6367 * execution on that CPU. 6368 * 6369 * This method returns the utilization of the specified CPU by discounting the 6370 * utilization of the specified task, whenever the task is currently 6371 * contributing to the CPU utilization. 6372 */ 6373 static unsigned long cpu_util_without(int cpu, struct task_struct *p) 6374 { 6375 struct cfs_rq *cfs_rq; 6376 unsigned int util; 6377 6378 /* Task has no contribution or is new */ 6379 if (cpu != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time)) 6380 return cpu_util(cpu); 6381 6382 cfs_rq = &cpu_rq(cpu)->cfs; 6383 util = READ_ONCE(cfs_rq->avg.util_avg); 6384 6385 /* Discount task's util from CPU's util */ 6386 lsub_positive(&util, task_util(p)); 6387 6388 /* 6389 * Covered cases: 6390 * 6391 * a) if *p is the only task sleeping on this CPU, then: 6392 * cpu_util (== task_util) > util_est (== 0) 6393 * and thus we return: 6394 * cpu_util_without = (cpu_util - task_util) = 0 6395 * 6396 * b) if other tasks are SLEEPING on this CPU, which is now exiting 6397 * IDLE, then: 6398 * cpu_util >= task_util 6399 * cpu_util > util_est (== 0) 6400 * and thus we discount *p's blocked utilization to return: 6401 * cpu_util_without = (cpu_util - task_util) >= 0 6402 * 6403 * c) if other tasks are RUNNABLE on that CPU and 6404 * util_est > cpu_util 6405 * then we use util_est since it returns a more restrictive 6406 * estimation of the spare capacity on that CPU, by just 6407 * considering the expected utilization of tasks already 6408 * runnable on that CPU. 6409 * 6410 * Cases a) and b) are covered by the above code, while case c) is 6411 * covered by the following code when estimated utilization is 6412 * enabled. 6413 */ 6414 if (sched_feat(UTIL_EST)) { 6415 unsigned int estimated = 6416 READ_ONCE(cfs_rq->avg.util_est.enqueued); 6417 6418 /* 6419 * Despite the following checks we still have a small window 6420 * for a possible race, when an execl's select_task_rq_fair() 6421 * races with LB's detach_task(): 6422 * 6423 * detach_task() 6424 * p->on_rq = TASK_ON_RQ_MIGRATING; 6425 * ---------------------------------- A 6426 * deactivate_task() \ 6427 * dequeue_task() + RaceTime 6428 * util_est_dequeue() / 6429 * ---------------------------------- B 6430 * 6431 * The additional check on "current == p" it's required to 6432 * properly fix the execl regression and it helps in further 6433 * reducing the chances for the above race. 6434 */ 6435 if (unlikely(task_on_rq_queued(p) || current == p)) 6436 lsub_positive(&estimated, _task_util_est(p)); 6437 6438 util = max(util, estimated); 6439 } 6440 6441 /* 6442 * Utilization (estimated) can exceed the CPU capacity, thus let's 6443 * clamp to the maximum CPU capacity to ensure consistency with 6444 * the cpu_util call. 6445 */ 6446 return min_t(unsigned long, util, capacity_orig_of(cpu)); 6447 } 6448 6449 /* 6450 * Predicts what cpu_util(@cpu) would return if @p was migrated (and enqueued) 6451 * to @dst_cpu. 6452 */ 6453 static unsigned long cpu_util_next(int cpu, struct task_struct *p, int dst_cpu) 6454 { 6455 struct cfs_rq *cfs_rq = &cpu_rq(cpu)->cfs; 6456 unsigned long util_est, util = READ_ONCE(cfs_rq->avg.util_avg); 6457 6458 /* 6459 * If @p migrates from @cpu to another, remove its contribution. Or, 6460 * if @p migrates from another CPU to @cpu, add its contribution. In 6461 * the other cases, @cpu is not impacted by the migration, so the 6462 * util_avg should already be correct. 6463 */ 6464 if (task_cpu(p) == cpu && dst_cpu != cpu) 6465 sub_positive(&util, task_util(p)); 6466 else if (task_cpu(p) != cpu && dst_cpu == cpu) 6467 util += task_util(p); 6468 6469 if (sched_feat(UTIL_EST)) { 6470 util_est = READ_ONCE(cfs_rq->avg.util_est.enqueued); 6471 6472 /* 6473 * During wake-up, the task isn't enqueued yet and doesn't 6474 * appear in the cfs_rq->avg.util_est.enqueued of any rq, 6475 * so just add it (if needed) to "simulate" what will be 6476 * cpu_util() after the task has been enqueued. 6477 */ 6478 if (dst_cpu == cpu) 6479 util_est += _task_util_est(p); 6480 6481 util = max(util, util_est); 6482 } 6483 6484 return min(util, capacity_orig_of(cpu)); 6485 } 6486 6487 /* 6488 * compute_energy(): Estimates the energy that @pd would consume if @p was 6489 * migrated to @dst_cpu. compute_energy() predicts what will be the utilization 6490 * landscape of @pd's CPUs after the task migration, and uses the Energy Model 6491 * to compute what would be the energy if we decided to actually migrate that 6492 * task. 6493 */ 6494 static long 6495 compute_energy(struct task_struct *p, int dst_cpu, struct perf_domain *pd) 6496 { 6497 struct cpumask *pd_mask = perf_domain_span(pd); 6498 unsigned long cpu_cap = arch_scale_cpu_capacity(cpumask_first(pd_mask)); 6499 unsigned long max_util = 0, sum_util = 0; 6500 int cpu; 6501 6502 /* 6503 * The capacity state of CPUs of the current rd can be driven by CPUs 6504 * of another rd if they belong to the same pd. So, account for the 6505 * utilization of these CPUs too by masking pd with cpu_online_mask 6506 * instead of the rd span. 6507 * 6508 * If an entire pd is outside of the current rd, it will not appear in 6509 * its pd list and will not be accounted by compute_energy(). 6510 */ 6511 for_each_cpu_and(cpu, pd_mask, cpu_online_mask) { 6512 unsigned long cpu_util, util_cfs = cpu_util_next(cpu, p, dst_cpu); 6513 struct task_struct *tsk = cpu == dst_cpu ? p : NULL; 6514 6515 /* 6516 * Busy time computation: utilization clamping is not 6517 * required since the ratio (sum_util / cpu_capacity) 6518 * is already enough to scale the EM reported power 6519 * consumption at the (eventually clamped) cpu_capacity. 6520 */ 6521 sum_util += schedutil_cpu_util(cpu, util_cfs, cpu_cap, 6522 ENERGY_UTIL, NULL); 6523 6524 /* 6525 * Performance domain frequency: utilization clamping 6526 * must be considered since it affects the selection 6527 * of the performance domain frequency. 6528 * NOTE: in case RT tasks are running, by default the 6529 * FREQUENCY_UTIL's utilization can be max OPP. 6530 */ 6531 cpu_util = schedutil_cpu_util(cpu, util_cfs, cpu_cap, 6532 FREQUENCY_UTIL, tsk); 6533 max_util = max(max_util, cpu_util); 6534 } 6535 6536 return em_cpu_energy(pd->em_pd, max_util, sum_util); 6537 } 6538 6539 /* 6540 * find_energy_efficient_cpu(): Find most energy-efficient target CPU for the 6541 * waking task. find_energy_efficient_cpu() looks for the CPU with maximum 6542 * spare capacity in each performance domain and uses it as a potential 6543 * candidate to execute the task. Then, it uses the Energy Model to figure 6544 * out which of the CPU candidates is the most energy-efficient. 6545 * 6546 * The rationale for this heuristic is as follows. In a performance domain, 6547 * all the most energy efficient CPU candidates (according to the Energy 6548 * Model) are those for which we'll request a low frequency. When there are 6549 * several CPUs for which the frequency request will be the same, we don't 6550 * have enough data to break the tie between them, because the Energy Model 6551 * only includes active power costs. With this model, if we assume that 6552 * frequency requests follow utilization (e.g. using schedutil), the CPU with 6553 * the maximum spare capacity in a performance domain is guaranteed to be among 6554 * the best candidates of the performance domain. 6555 * 6556 * In practice, it could be preferable from an energy standpoint to pack 6557 * small tasks on a CPU in order to let other CPUs go in deeper idle states, 6558 * but that could also hurt our chances to go cluster idle, and we have no 6559 * ways to tell with the current Energy Model if this is actually a good 6560 * idea or not. So, find_energy_efficient_cpu() basically favors 6561 * cluster-packing, and spreading inside a cluster. That should at least be 6562 * a good thing for latency, and this is consistent with the idea that most 6563 * of the energy savings of EAS come from the asymmetry of the system, and 6564 * not so much from breaking the tie between identical CPUs. That's also the 6565 * reason why EAS is enabled in the topology code only for systems where 6566 * SD_ASYM_CPUCAPACITY is set. 6567 * 6568 * NOTE: Forkees are not accepted in the energy-aware wake-up path because 6569 * they don't have any useful utilization data yet and it's not possible to 6570 * forecast their impact on energy consumption. Consequently, they will be 6571 * placed by find_idlest_cpu() on the least loaded CPU, which might turn out 6572 * to be energy-inefficient in some use-cases. The alternative would be to 6573 * bias new tasks towards specific types of CPUs first, or to try to infer 6574 * their util_avg from the parent task, but those heuristics could hurt 6575 * other use-cases too. So, until someone finds a better way to solve this, 6576 * let's keep things simple by re-using the existing slow path. 6577 */ 6578 static int find_energy_efficient_cpu(struct task_struct *p, int prev_cpu) 6579 { 6580 unsigned long prev_delta = ULONG_MAX, best_delta = ULONG_MAX; 6581 struct root_domain *rd = cpu_rq(smp_processor_id())->rd; 6582 unsigned long cpu_cap, util, base_energy = 0; 6583 int cpu, best_energy_cpu = prev_cpu; 6584 struct sched_domain *sd; 6585 struct perf_domain *pd; 6586 6587 rcu_read_lock(); 6588 pd = rcu_dereference(rd->pd); 6589 if (!pd || READ_ONCE(rd->overutilized)) 6590 goto fail; 6591 6592 /* 6593 * Energy-aware wake-up happens on the lowest sched_domain starting 6594 * from sd_asym_cpucapacity spanning over this_cpu and prev_cpu. 6595 */ 6596 sd = rcu_dereference(*this_cpu_ptr(&sd_asym_cpucapacity)); 6597 while (sd && !cpumask_test_cpu(prev_cpu, sched_domain_span(sd))) 6598 sd = sd->parent; 6599 if (!sd) 6600 goto fail; 6601 6602 sync_entity_load_avg(&p->se); 6603 if (!task_util_est(p)) 6604 goto unlock; 6605 6606 for (; pd; pd = pd->next) { 6607 unsigned long cur_delta, spare_cap, max_spare_cap = 0; 6608 unsigned long base_energy_pd; 6609 int max_spare_cap_cpu = -1; 6610 6611 /* Compute the 'base' energy of the pd, without @p */ 6612 base_energy_pd = compute_energy(p, -1, pd); 6613 base_energy += base_energy_pd; 6614 6615 for_each_cpu_and(cpu, perf_domain_span(pd), sched_domain_span(sd)) { 6616 if (!cpumask_test_cpu(cpu, p->cpus_ptr)) 6617 continue; 6618 6619 util = cpu_util_next(cpu, p, cpu); 6620 cpu_cap = capacity_of(cpu); 6621 spare_cap = cpu_cap; 6622 lsub_positive(&spare_cap, util); 6623 6624 /* 6625 * Skip CPUs that cannot satisfy the capacity request. 6626 * IOW, placing the task there would make the CPU 6627 * overutilized. Take uclamp into account to see how 6628 * much capacity we can get out of the CPU; this is 6629 * aligned with schedutil_cpu_util(). 6630 */ 6631 util = uclamp_rq_util_with(cpu_rq(cpu), util, p); 6632 if (!fits_capacity(util, cpu_cap)) 6633 continue; 6634 6635 /* Always use prev_cpu as a candidate. */ 6636 if (cpu == prev_cpu) { 6637 prev_delta = compute_energy(p, prev_cpu, pd); 6638 prev_delta -= base_energy_pd; 6639 best_delta = min(best_delta, prev_delta); 6640 } 6641 6642 /* 6643 * Find the CPU with the maximum spare capacity in 6644 * the performance domain 6645 */ 6646 if (spare_cap > max_spare_cap) { 6647 max_spare_cap = spare_cap; 6648 max_spare_cap_cpu = cpu; 6649 } 6650 } 6651 6652 /* Evaluate the energy impact of using this CPU. */ 6653 if (max_spare_cap_cpu >= 0 && max_spare_cap_cpu != prev_cpu) { 6654 cur_delta = compute_energy(p, max_spare_cap_cpu, pd); 6655 cur_delta -= base_energy_pd; 6656 if (cur_delta < best_delta) { 6657 best_delta = cur_delta; 6658 best_energy_cpu = max_spare_cap_cpu; 6659 } 6660 } 6661 } 6662 unlock: 6663 rcu_read_unlock(); 6664 6665 /* 6666 * Pick the best CPU if prev_cpu cannot be used, or if it saves at 6667 * least 6% of the energy used by prev_cpu. 6668 */ 6669 if (prev_delta == ULONG_MAX) 6670 return best_energy_cpu; 6671 6672 if ((prev_delta - best_delta) > ((prev_delta + base_energy) >> 4)) 6673 return best_energy_cpu; 6674 6675 return prev_cpu; 6676 6677 fail: 6678 rcu_read_unlock(); 6679 6680 return -1; 6681 } 6682 6683 /* 6684 * select_task_rq_fair: Select target runqueue for the waking task in domains 6685 * that have the 'sd_flag' flag set. In practice, this is SD_BALANCE_WAKE, 6686 * SD_BALANCE_FORK, or SD_BALANCE_EXEC. 6687 * 6688 * Balances load by selecting the idlest CPU in the idlest group, or under 6689 * certain conditions an idle sibling CPU if the domain has SD_WAKE_AFFINE set. 6690 * 6691 * Returns the target CPU number. 6692 * 6693 * preempt must be disabled. 6694 */ 6695 static int 6696 select_task_rq_fair(struct task_struct *p, int prev_cpu, int sd_flag, int wake_flags) 6697 { 6698 struct sched_domain *tmp, *sd = NULL; 6699 int cpu = smp_processor_id(); 6700 int new_cpu = prev_cpu; 6701 int want_affine = 0; 6702 int sync = (wake_flags & WF_SYNC) && !(current->flags & PF_EXITING); 6703 6704 if (sd_flag & SD_BALANCE_WAKE) { 6705 record_wakee(p); 6706 6707 if (sched_energy_enabled()) { 6708 new_cpu = find_energy_efficient_cpu(p, prev_cpu); 6709 if (new_cpu >= 0) 6710 return new_cpu; 6711 new_cpu = prev_cpu; 6712 } 6713 6714 want_affine = !wake_wide(p) && cpumask_test_cpu(cpu, p->cpus_ptr); 6715 } 6716 6717 rcu_read_lock(); 6718 for_each_domain(cpu, tmp) { 6719 /* 6720 * If both 'cpu' and 'prev_cpu' are part of this domain, 6721 * cpu is a valid SD_WAKE_AFFINE target. 6722 */ 6723 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) && 6724 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) { 6725 if (cpu != prev_cpu) 6726 new_cpu = wake_affine(tmp, p, cpu, prev_cpu, sync); 6727 6728 sd = NULL; /* Prefer wake_affine over balance flags */ 6729 break; 6730 } 6731 6732 if (tmp->flags & sd_flag) 6733 sd = tmp; 6734 else if (!want_affine) 6735 break; 6736 } 6737 6738 if (unlikely(sd)) { 6739 /* Slow path */ 6740 new_cpu = find_idlest_cpu(sd, p, cpu, prev_cpu, sd_flag); 6741 } else if (sd_flag & SD_BALANCE_WAKE) { /* XXX always ? */ 6742 /* Fast path */ 6743 6744 new_cpu = select_idle_sibling(p, prev_cpu, new_cpu); 6745 6746 if (want_affine) 6747 current->recent_used_cpu = cpu; 6748 } 6749 rcu_read_unlock(); 6750 6751 return new_cpu; 6752 } 6753 6754 static void detach_entity_cfs_rq(struct sched_entity *se); 6755 6756 /* 6757 * Called immediately before a task is migrated to a new CPU; task_cpu(p) and 6758 * cfs_rq_of(p) references at time of call are still valid and identify the 6759 * previous CPU. The caller guarantees p->pi_lock or task_rq(p)->lock is held. 6760 */ 6761 static void migrate_task_rq_fair(struct task_struct *p, int new_cpu) 6762 { 6763 /* 6764 * As blocked tasks retain absolute vruntime the migration needs to 6765 * deal with this by subtracting the old and adding the new 6766 * min_vruntime -- the latter is done by enqueue_entity() when placing 6767 * the task on the new runqueue. 6768 */ 6769 if (p->state == TASK_WAKING) { 6770 struct sched_entity *se = &p->se; 6771 struct cfs_rq *cfs_rq = cfs_rq_of(se); 6772 u64 min_vruntime; 6773 6774 #ifndef CONFIG_64BIT 6775 u64 min_vruntime_copy; 6776 6777 do { 6778 min_vruntime_copy = cfs_rq->min_vruntime_copy; 6779 smp_rmb(); 6780 min_vruntime = cfs_rq->min_vruntime; 6781 } while (min_vruntime != min_vruntime_copy); 6782 #else 6783 min_vruntime = cfs_rq->min_vruntime; 6784 #endif 6785 6786 se->vruntime -= min_vruntime; 6787 } 6788 6789 if (p->on_rq == TASK_ON_RQ_MIGRATING) { 6790 /* 6791 * In case of TASK_ON_RQ_MIGRATING we in fact hold the 'old' 6792 * rq->lock and can modify state directly. 6793 */ 6794 lockdep_assert_held(&task_rq(p)->lock); 6795 detach_entity_cfs_rq(&p->se); 6796 6797 } else { 6798 /* 6799 * We are supposed to update the task to "current" time, then 6800 * its up to date and ready to go to new CPU/cfs_rq. But we 6801 * have difficulty in getting what current time is, so simply 6802 * throw away the out-of-date time. This will result in the 6803 * wakee task is less decayed, but giving the wakee more load 6804 * sounds not bad. 6805 */ 6806 remove_entity_load_avg(&p->se); 6807 } 6808 6809 /* Tell new CPU we are migrated */ 6810 p->se.avg.last_update_time = 0; 6811 6812 /* We have migrated, no longer consider this task hot */ 6813 p->se.exec_start = 0; 6814 6815 update_scan_period(p, new_cpu); 6816 } 6817 6818 static void task_dead_fair(struct task_struct *p) 6819 { 6820 remove_entity_load_avg(&p->se); 6821 } 6822 6823 static int 6824 balance_fair(struct rq *rq, struct task_struct *prev, struct rq_flags *rf) 6825 { 6826 if (rq->nr_running) 6827 return 1; 6828 6829 return newidle_balance(rq, rf) != 0; 6830 } 6831 #endif /* CONFIG_SMP */ 6832 6833 static unsigned long wakeup_gran(struct sched_entity *se) 6834 { 6835 unsigned long gran = sysctl_sched_wakeup_granularity; 6836 6837 /* 6838 * Since its curr running now, convert the gran from real-time 6839 * to virtual-time in his units. 6840 * 6841 * By using 'se' instead of 'curr' we penalize light tasks, so 6842 * they get preempted easier. That is, if 'se' < 'curr' then 6843 * the resulting gran will be larger, therefore penalizing the 6844 * lighter, if otoh 'se' > 'curr' then the resulting gran will 6845 * be smaller, again penalizing the lighter task. 6846 * 6847 * This is especially important for buddies when the leftmost 6848 * task is higher priority than the buddy. 6849 */ 6850 return calc_delta_fair(gran, se); 6851 } 6852 6853 /* 6854 * Should 'se' preempt 'curr'. 6855 * 6856 * |s1 6857 * |s2 6858 * |s3 6859 * g 6860 * |<--->|c 6861 * 6862 * w(c, s1) = -1 6863 * w(c, s2) = 0 6864 * w(c, s3) = 1 6865 * 6866 */ 6867 static int 6868 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se) 6869 { 6870 s64 gran, vdiff = curr->vruntime - se->vruntime; 6871 6872 if (vdiff <= 0) 6873 return -1; 6874 6875 gran = wakeup_gran(se); 6876 if (vdiff > gran) 6877 return 1; 6878 6879 return 0; 6880 } 6881 6882 static void set_last_buddy(struct sched_entity *se) 6883 { 6884 if (entity_is_task(se) && unlikely(task_has_idle_policy(task_of(se)))) 6885 return; 6886 6887 for_each_sched_entity(se) { 6888 if (SCHED_WARN_ON(!se->on_rq)) 6889 return; 6890 cfs_rq_of(se)->last = se; 6891 } 6892 } 6893 6894 static void set_next_buddy(struct sched_entity *se) 6895 { 6896 if (entity_is_task(se) && unlikely(task_has_idle_policy(task_of(se)))) 6897 return; 6898 6899 for_each_sched_entity(se) { 6900 if (SCHED_WARN_ON(!se->on_rq)) 6901 return; 6902 cfs_rq_of(se)->next = se; 6903 } 6904 } 6905 6906 static void set_skip_buddy(struct sched_entity *se) 6907 { 6908 for_each_sched_entity(se) 6909 cfs_rq_of(se)->skip = se; 6910 } 6911 6912 /* 6913 * Preempt the current task with a newly woken task if needed: 6914 */ 6915 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags) 6916 { 6917 struct task_struct *curr = rq->curr; 6918 struct sched_entity *se = &curr->se, *pse = &p->se; 6919 struct cfs_rq *cfs_rq = task_cfs_rq(curr); 6920 int scale = cfs_rq->nr_running >= sched_nr_latency; 6921 int next_buddy_marked = 0; 6922 6923 if (unlikely(se == pse)) 6924 return; 6925 6926 /* 6927 * This is possible from callers such as attach_tasks(), in which we 6928 * unconditionally check_prempt_curr() after an enqueue (which may have 6929 * lead to a throttle). This both saves work and prevents false 6930 * next-buddy nomination below. 6931 */ 6932 if (unlikely(throttled_hierarchy(cfs_rq_of(pse)))) 6933 return; 6934 6935 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) { 6936 set_next_buddy(pse); 6937 next_buddy_marked = 1; 6938 } 6939 6940 /* 6941 * We can come here with TIF_NEED_RESCHED already set from new task 6942 * wake up path. 6943 * 6944 * Note: this also catches the edge-case of curr being in a throttled 6945 * group (e.g. via set_curr_task), since update_curr() (in the 6946 * enqueue of curr) will have resulted in resched being set. This 6947 * prevents us from potentially nominating it as a false LAST_BUDDY 6948 * below. 6949 */ 6950 if (test_tsk_need_resched(curr)) 6951 return; 6952 6953 /* Idle tasks are by definition preempted by non-idle tasks. */ 6954 if (unlikely(task_has_idle_policy(curr)) && 6955 likely(!task_has_idle_policy(p))) 6956 goto preempt; 6957 6958 /* 6959 * Batch and idle tasks do not preempt non-idle tasks (their preemption 6960 * is driven by the tick): 6961 */ 6962 if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION)) 6963 return; 6964 6965 find_matching_se(&se, &pse); 6966 update_curr(cfs_rq_of(se)); 6967 BUG_ON(!pse); 6968 if (wakeup_preempt_entity(se, pse) == 1) { 6969 /* 6970 * Bias pick_next to pick the sched entity that is 6971 * triggering this preemption. 6972 */ 6973 if (!next_buddy_marked) 6974 set_next_buddy(pse); 6975 goto preempt; 6976 } 6977 6978 return; 6979 6980 preempt: 6981 resched_curr(rq); 6982 /* 6983 * Only set the backward buddy when the current task is still 6984 * on the rq. This can happen when a wakeup gets interleaved 6985 * with schedule on the ->pre_schedule() or idle_balance() 6986 * point, either of which can * drop the rq lock. 6987 * 6988 * Also, during early boot the idle thread is in the fair class, 6989 * for obvious reasons its a bad idea to schedule back to it. 6990 */ 6991 if (unlikely(!se->on_rq || curr == rq->idle)) 6992 return; 6993 6994 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se)) 6995 set_last_buddy(se); 6996 } 6997 6998 struct task_struct * 6999 pick_next_task_fair(struct rq *rq, struct task_struct *prev, struct rq_flags *rf) 7000 { 7001 struct cfs_rq *cfs_rq = &rq->cfs; 7002 struct sched_entity *se; 7003 struct task_struct *p; 7004 int new_tasks; 7005 7006 again: 7007 if (!sched_fair_runnable(rq)) 7008 goto idle; 7009 7010 #ifdef CONFIG_FAIR_GROUP_SCHED 7011 if (!prev || prev->sched_class != &fair_sched_class) 7012 goto simple; 7013 7014 /* 7015 * Because of the set_next_buddy() in dequeue_task_fair() it is rather 7016 * likely that a next task is from the same cgroup as the current. 7017 * 7018 * Therefore attempt to avoid putting and setting the entire cgroup 7019 * hierarchy, only change the part that actually changes. 7020 */ 7021 7022 do { 7023 struct sched_entity *curr = cfs_rq->curr; 7024 7025 /* 7026 * Since we got here without doing put_prev_entity() we also 7027 * have to consider cfs_rq->curr. If it is still a runnable 7028 * entity, update_curr() will update its vruntime, otherwise 7029 * forget we've ever seen it. 7030 */ 7031 if (curr) { 7032 if (curr->on_rq) 7033 update_curr(cfs_rq); 7034 else 7035 curr = NULL; 7036 7037 /* 7038 * This call to check_cfs_rq_runtime() will do the 7039 * throttle and dequeue its entity in the parent(s). 7040 * Therefore the nr_running test will indeed 7041 * be correct. 7042 */ 7043 if (unlikely(check_cfs_rq_runtime(cfs_rq))) { 7044 cfs_rq = &rq->cfs; 7045 7046 if (!cfs_rq->nr_running) 7047 goto idle; 7048 7049 goto simple; 7050 } 7051 } 7052 7053 se = pick_next_entity(cfs_rq, curr); 7054 cfs_rq = group_cfs_rq(se); 7055 } while (cfs_rq); 7056 7057 p = task_of(se); 7058 7059 /* 7060 * Since we haven't yet done put_prev_entity and if the selected task 7061 * is a different task than we started out with, try and touch the 7062 * least amount of cfs_rqs. 7063 */ 7064 if (prev != p) { 7065 struct sched_entity *pse = &prev->se; 7066 7067 while (!(cfs_rq = is_same_group(se, pse))) { 7068 int se_depth = se->depth; 7069 int pse_depth = pse->depth; 7070 7071 if (se_depth <= pse_depth) { 7072 put_prev_entity(cfs_rq_of(pse), pse); 7073 pse = parent_entity(pse); 7074 } 7075 if (se_depth >= pse_depth) { 7076 set_next_entity(cfs_rq_of(se), se); 7077 se = parent_entity(se); 7078 } 7079 } 7080 7081 put_prev_entity(cfs_rq, pse); 7082 set_next_entity(cfs_rq, se); 7083 } 7084 7085 goto done; 7086 simple: 7087 #endif 7088 if (prev) 7089 put_prev_task(rq, prev); 7090 7091 do { 7092 se = pick_next_entity(cfs_rq, NULL); 7093 set_next_entity(cfs_rq, se); 7094 cfs_rq = group_cfs_rq(se); 7095 } while (cfs_rq); 7096 7097 p = task_of(se); 7098 7099 done: __maybe_unused; 7100 #ifdef CONFIG_SMP 7101 /* 7102 * Move the next running task to the front of 7103 * the list, so our cfs_tasks list becomes MRU 7104 * one. 7105 */ 7106 list_move(&p->se.group_node, &rq->cfs_tasks); 7107 #endif 7108 7109 if (hrtick_enabled(rq)) 7110 hrtick_start_fair(rq, p); 7111 7112 update_misfit_status(p, rq); 7113 7114 return p; 7115 7116 idle: 7117 if (!rf) 7118 return NULL; 7119 7120 new_tasks = newidle_balance(rq, rf); 7121 7122 /* 7123 * Because newidle_balance() releases (and re-acquires) rq->lock, it is 7124 * possible for any higher priority task to appear. In that case we 7125 * must re-start the pick_next_entity() loop. 7126 */ 7127 if (new_tasks < 0) 7128 return RETRY_TASK; 7129 7130 if (new_tasks > 0) 7131 goto again; 7132 7133 /* 7134 * rq is about to be idle, check if we need to update the 7135 * lost_idle_time of clock_pelt 7136 */ 7137 update_idle_rq_clock_pelt(rq); 7138 7139 return NULL; 7140 } 7141 7142 static struct task_struct *__pick_next_task_fair(struct rq *rq) 7143 { 7144 return pick_next_task_fair(rq, NULL, NULL); 7145 } 7146 7147 /* 7148 * Account for a descheduled task: 7149 */ 7150 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev) 7151 { 7152 struct sched_entity *se = &prev->se; 7153 struct cfs_rq *cfs_rq; 7154 7155 for_each_sched_entity(se) { 7156 cfs_rq = cfs_rq_of(se); 7157 put_prev_entity(cfs_rq, se); 7158 } 7159 } 7160 7161 /* 7162 * sched_yield() is very simple 7163 * 7164 * The magic of dealing with the ->skip buddy is in pick_next_entity. 7165 */ 7166 static void yield_task_fair(struct rq *rq) 7167 { 7168 struct task_struct *curr = rq->curr; 7169 struct cfs_rq *cfs_rq = task_cfs_rq(curr); 7170 struct sched_entity *se = &curr->se; 7171 7172 /* 7173 * Are we the only task in the tree? 7174 */ 7175 if (unlikely(rq->nr_running == 1)) 7176 return; 7177 7178 clear_buddies(cfs_rq, se); 7179 7180 if (curr->policy != SCHED_BATCH) { 7181 update_rq_clock(rq); 7182 /* 7183 * Update run-time statistics of the 'current'. 7184 */ 7185 update_curr(cfs_rq); 7186 /* 7187 * Tell update_rq_clock() that we've just updated, 7188 * so we don't do microscopic update in schedule() 7189 * and double the fastpath cost. 7190 */ 7191 rq_clock_skip_update(rq); 7192 } 7193 7194 set_skip_buddy(se); 7195 } 7196 7197 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p) 7198 { 7199 struct sched_entity *se = &p->se; 7200 7201 /* throttled hierarchies are not runnable */ 7202 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se))) 7203 return false; 7204 7205 /* Tell the scheduler that we'd really like pse to run next. */ 7206 set_next_buddy(se); 7207 7208 yield_task_fair(rq); 7209 7210 return true; 7211 } 7212 7213 #ifdef CONFIG_SMP 7214 /************************************************** 7215 * Fair scheduling class load-balancing methods. 7216 * 7217 * BASICS 7218 * 7219 * The purpose of load-balancing is to achieve the same basic fairness the 7220 * per-CPU scheduler provides, namely provide a proportional amount of compute 7221 * time to each task. This is expressed in the following equation: 7222 * 7223 * W_i,n/P_i == W_j,n/P_j for all i,j (1) 7224 * 7225 * Where W_i,n is the n-th weight average for CPU i. The instantaneous weight 7226 * W_i,0 is defined as: 7227 * 7228 * W_i,0 = \Sum_j w_i,j (2) 7229 * 7230 * Where w_i,j is the weight of the j-th runnable task on CPU i. This weight 7231 * is derived from the nice value as per sched_prio_to_weight[]. 7232 * 7233 * The weight average is an exponential decay average of the instantaneous 7234 * weight: 7235 * 7236 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3) 7237 * 7238 * C_i is the compute capacity of CPU i, typically it is the 7239 * fraction of 'recent' time available for SCHED_OTHER task execution. But it 7240 * can also include other factors [XXX]. 7241 * 7242 * To achieve this balance we define a measure of imbalance which follows 7243 * directly from (1): 7244 * 7245 * imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j } (4) 7246 * 7247 * We them move tasks around to minimize the imbalance. In the continuous 7248 * function space it is obvious this converges, in the discrete case we get 7249 * a few fun cases generally called infeasible weight scenarios. 7250 * 7251 * [XXX expand on: 7252 * - infeasible weights; 7253 * - local vs global optima in the discrete case. ] 7254 * 7255 * 7256 * SCHED DOMAINS 7257 * 7258 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2) 7259 * for all i,j solution, we create a tree of CPUs that follows the hardware 7260 * topology where each level pairs two lower groups (or better). This results 7261 * in O(log n) layers. Furthermore we reduce the number of CPUs going up the 7262 * tree to only the first of the previous level and we decrease the frequency 7263 * of load-balance at each level inv. proportional to the number of CPUs in 7264 * the groups. 7265 * 7266 * This yields: 7267 * 7268 * log_2 n 1 n 7269 * \Sum { --- * --- * 2^i } = O(n) (5) 7270 * i = 0 2^i 2^i 7271 * `- size of each group 7272 * | | `- number of CPUs doing load-balance 7273 * | `- freq 7274 * `- sum over all levels 7275 * 7276 * Coupled with a limit on how many tasks we can migrate every balance pass, 7277 * this makes (5) the runtime complexity of the balancer. 7278 * 7279 * An important property here is that each CPU is still (indirectly) connected 7280 * to every other CPU in at most O(log n) steps: 7281 * 7282 * The adjacency matrix of the resulting graph is given by: 7283 * 7284 * log_2 n 7285 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6) 7286 * k = 0 7287 * 7288 * And you'll find that: 7289 * 7290 * A^(log_2 n)_i,j != 0 for all i,j (7) 7291 * 7292 * Showing there's indeed a path between every CPU in at most O(log n) steps. 7293 * The task movement gives a factor of O(m), giving a convergence complexity 7294 * of: 7295 * 7296 * O(nm log n), n := nr_cpus, m := nr_tasks (8) 7297 * 7298 * 7299 * WORK CONSERVING 7300 * 7301 * In order to avoid CPUs going idle while there's still work to do, new idle 7302 * balancing is more aggressive and has the newly idle CPU iterate up the domain 7303 * tree itself instead of relying on other CPUs to bring it work. 7304 * 7305 * This adds some complexity to both (5) and (8) but it reduces the total idle 7306 * time. 7307 * 7308 * [XXX more?] 7309 * 7310 * 7311 * CGROUPS 7312 * 7313 * Cgroups make a horror show out of (2), instead of a simple sum we get: 7314 * 7315 * s_k,i 7316 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9) 7317 * S_k 7318 * 7319 * Where 7320 * 7321 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10) 7322 * 7323 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on CPU i. 7324 * 7325 * The big problem is S_k, its a global sum needed to compute a local (W_i) 7326 * property. 7327 * 7328 * [XXX write more on how we solve this.. _after_ merging pjt's patches that 7329 * rewrite all of this once again.] 7330 */ 7331 7332 static unsigned long __read_mostly max_load_balance_interval = HZ/10; 7333 7334 enum fbq_type { regular, remote, all }; 7335 7336 /* 7337 * 'group_type' describes the group of CPUs at the moment of load balancing. 7338 * 7339 * The enum is ordered by pulling priority, with the group with lowest priority 7340 * first so the group_type can simply be compared when selecting the busiest 7341 * group. See update_sd_pick_busiest(). 7342 */ 7343 enum group_type { 7344 /* The group has spare capacity that can be used to run more tasks. */ 7345 group_has_spare = 0, 7346 /* 7347 * The group is fully used and the tasks don't compete for more CPU 7348 * cycles. Nevertheless, some tasks might wait before running. 7349 */ 7350 group_fully_busy, 7351 /* 7352 * SD_ASYM_CPUCAPACITY only: One task doesn't fit with CPU's capacity 7353 * and must be migrated to a more powerful CPU. 7354 */ 7355 group_misfit_task, 7356 /* 7357 * SD_ASYM_PACKING only: One local CPU with higher capacity is available, 7358 * and the task should be migrated to it instead of running on the 7359 * current CPU. 7360 */ 7361 group_asym_packing, 7362 /* 7363 * The tasks' affinity constraints previously prevented the scheduler 7364 * from balancing the load across the system. 7365 */ 7366 group_imbalanced, 7367 /* 7368 * The CPU is overloaded and can't provide expected CPU cycles to all 7369 * tasks. 7370 */ 7371 group_overloaded 7372 }; 7373 7374 enum migration_type { 7375 migrate_load = 0, 7376 migrate_util, 7377 migrate_task, 7378 migrate_misfit 7379 }; 7380 7381 #define LBF_ALL_PINNED 0x01 7382 #define LBF_NEED_BREAK 0x02 7383 #define LBF_DST_PINNED 0x04 7384 #define LBF_SOME_PINNED 0x08 7385 #define LBF_NOHZ_STATS 0x10 7386 #define LBF_NOHZ_AGAIN 0x20 7387 7388 struct lb_env { 7389 struct sched_domain *sd; 7390 7391 struct rq *src_rq; 7392 int src_cpu; 7393 7394 int dst_cpu; 7395 struct rq *dst_rq; 7396 7397 struct cpumask *dst_grpmask; 7398 int new_dst_cpu; 7399 enum cpu_idle_type idle; 7400 long imbalance; 7401 /* The set of CPUs under consideration for load-balancing */ 7402 struct cpumask *cpus; 7403 7404 unsigned int flags; 7405 7406 unsigned int loop; 7407 unsigned int loop_break; 7408 unsigned int loop_max; 7409 7410 enum fbq_type fbq_type; 7411 enum migration_type migration_type; 7412 struct list_head tasks; 7413 }; 7414 7415 /* 7416 * Is this task likely cache-hot: 7417 */ 7418 static int task_hot(struct task_struct *p, struct lb_env *env) 7419 { 7420 s64 delta; 7421 7422 lockdep_assert_held(&env->src_rq->lock); 7423 7424 if (p->sched_class != &fair_sched_class) 7425 return 0; 7426 7427 if (unlikely(task_has_idle_policy(p))) 7428 return 0; 7429 7430 /* SMT siblings share cache */ 7431 if (env->sd->flags & SD_SHARE_CPUCAPACITY) 7432 return 0; 7433 7434 /* 7435 * Buddy candidates are cache hot: 7436 */ 7437 if (sched_feat(CACHE_HOT_BUDDY) && env->dst_rq->nr_running && 7438 (&p->se == cfs_rq_of(&p->se)->next || 7439 &p->se == cfs_rq_of(&p->se)->last)) 7440 return 1; 7441 7442 if (sysctl_sched_migration_cost == -1) 7443 return 1; 7444 if (sysctl_sched_migration_cost == 0) 7445 return 0; 7446 7447 delta = rq_clock_task(env->src_rq) - p->se.exec_start; 7448 7449 return delta < (s64)sysctl_sched_migration_cost; 7450 } 7451 7452 #ifdef CONFIG_NUMA_BALANCING 7453 /* 7454 * Returns 1, if task migration degrades locality 7455 * Returns 0, if task migration improves locality i.e migration preferred. 7456 * Returns -1, if task migration is not affected by locality. 7457 */ 7458 static int migrate_degrades_locality(struct task_struct *p, struct lb_env *env) 7459 { 7460 struct numa_group *numa_group = rcu_dereference(p->numa_group); 7461 unsigned long src_weight, dst_weight; 7462 int src_nid, dst_nid, dist; 7463 7464 if (!static_branch_likely(&sched_numa_balancing)) 7465 return -1; 7466 7467 if (!p->numa_faults || !(env->sd->flags & SD_NUMA)) 7468 return -1; 7469 7470 src_nid = cpu_to_node(env->src_cpu); 7471 dst_nid = cpu_to_node(env->dst_cpu); 7472 7473 if (src_nid == dst_nid) 7474 return -1; 7475 7476 /* Migrating away from the preferred node is always bad. */ 7477 if (src_nid == p->numa_preferred_nid) { 7478 if (env->src_rq->nr_running > env->src_rq->nr_preferred_running) 7479 return 1; 7480 else 7481 return -1; 7482 } 7483 7484 /* Encourage migration to the preferred node. */ 7485 if (dst_nid == p->numa_preferred_nid) 7486 return 0; 7487 7488 /* Leaving a core idle is often worse than degrading locality. */ 7489 if (env->idle == CPU_IDLE) 7490 return -1; 7491 7492 dist = node_distance(src_nid, dst_nid); 7493 if (numa_group) { 7494 src_weight = group_weight(p, src_nid, dist); 7495 dst_weight = group_weight(p, dst_nid, dist); 7496 } else { 7497 src_weight = task_weight(p, src_nid, dist); 7498 dst_weight = task_weight(p, dst_nid, dist); 7499 } 7500 7501 return dst_weight < src_weight; 7502 } 7503 7504 #else 7505 static inline int migrate_degrades_locality(struct task_struct *p, 7506 struct lb_env *env) 7507 { 7508 return -1; 7509 } 7510 #endif 7511 7512 /* 7513 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu? 7514 */ 7515 static 7516 int can_migrate_task(struct task_struct *p, struct lb_env *env) 7517 { 7518 int tsk_cache_hot; 7519 7520 lockdep_assert_held(&env->src_rq->lock); 7521 7522 /* 7523 * We do not migrate tasks that are: 7524 * 1) throttled_lb_pair, or 7525 * 2) cannot be migrated to this CPU due to cpus_ptr, or 7526 * 3) running (obviously), or 7527 * 4) are cache-hot on their current CPU. 7528 */ 7529 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu)) 7530 return 0; 7531 7532 if (!cpumask_test_cpu(env->dst_cpu, p->cpus_ptr)) { 7533 int cpu; 7534 7535 schedstat_inc(p->se.statistics.nr_failed_migrations_affine); 7536 7537 env->flags |= LBF_SOME_PINNED; 7538 7539 /* 7540 * Remember if this task can be migrated to any other CPU in 7541 * our sched_group. We may want to revisit it if we couldn't 7542 * meet load balance goals by pulling other tasks on src_cpu. 7543 * 7544 * Avoid computing new_dst_cpu for NEWLY_IDLE or if we have 7545 * already computed one in current iteration. 7546 */ 7547 if (env->idle == CPU_NEWLY_IDLE || (env->flags & LBF_DST_PINNED)) 7548 return 0; 7549 7550 /* Prevent to re-select dst_cpu via env's CPUs: */ 7551 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) { 7552 if (cpumask_test_cpu(cpu, p->cpus_ptr)) { 7553 env->flags |= LBF_DST_PINNED; 7554 env->new_dst_cpu = cpu; 7555 break; 7556 } 7557 } 7558 7559 return 0; 7560 } 7561 7562 /* Record that we found atleast one task that could run on dst_cpu */ 7563 env->flags &= ~LBF_ALL_PINNED; 7564 7565 if (task_running(env->src_rq, p)) { 7566 schedstat_inc(p->se.statistics.nr_failed_migrations_running); 7567 return 0; 7568 } 7569 7570 /* 7571 * Aggressive migration if: 7572 * 1) destination numa is preferred 7573 * 2) task is cache cold, or 7574 * 3) too many balance attempts have failed. 7575 */ 7576 tsk_cache_hot = migrate_degrades_locality(p, env); 7577 if (tsk_cache_hot == -1) 7578 tsk_cache_hot = task_hot(p, env); 7579 7580 if (tsk_cache_hot <= 0 || 7581 env->sd->nr_balance_failed > env->sd->cache_nice_tries) { 7582 if (tsk_cache_hot == 1) { 7583 schedstat_inc(env->sd->lb_hot_gained[env->idle]); 7584 schedstat_inc(p->se.statistics.nr_forced_migrations); 7585 } 7586 return 1; 7587 } 7588 7589 schedstat_inc(p->se.statistics.nr_failed_migrations_hot); 7590 return 0; 7591 } 7592 7593 /* 7594 * detach_task() -- detach the task for the migration specified in env 7595 */ 7596 static void detach_task(struct task_struct *p, struct lb_env *env) 7597 { 7598 lockdep_assert_held(&env->src_rq->lock); 7599 7600 deactivate_task(env->src_rq, p, DEQUEUE_NOCLOCK); 7601 set_task_cpu(p, env->dst_cpu); 7602 } 7603 7604 /* 7605 * detach_one_task() -- tries to dequeue exactly one task from env->src_rq, as 7606 * part of active balancing operations within "domain". 7607 * 7608 * Returns a task if successful and NULL otherwise. 7609 */ 7610 static struct task_struct *detach_one_task(struct lb_env *env) 7611 { 7612 struct task_struct *p; 7613 7614 lockdep_assert_held(&env->src_rq->lock); 7615 7616 list_for_each_entry_reverse(p, 7617 &env->src_rq->cfs_tasks, se.group_node) { 7618 if (!can_migrate_task(p, env)) 7619 continue; 7620 7621 detach_task(p, env); 7622 7623 /* 7624 * Right now, this is only the second place where 7625 * lb_gained[env->idle] is updated (other is detach_tasks) 7626 * so we can safely collect stats here rather than 7627 * inside detach_tasks(). 7628 */ 7629 schedstat_inc(env->sd->lb_gained[env->idle]); 7630 return p; 7631 } 7632 return NULL; 7633 } 7634 7635 static const unsigned int sched_nr_migrate_break = 32; 7636 7637 /* 7638 * detach_tasks() -- tries to detach up to imbalance load/util/tasks from 7639 * busiest_rq, as part of a balancing operation within domain "sd". 7640 * 7641 * Returns number of detached tasks if successful and 0 otherwise. 7642 */ 7643 static int detach_tasks(struct lb_env *env) 7644 { 7645 struct list_head *tasks = &env->src_rq->cfs_tasks; 7646 unsigned long util, load; 7647 struct task_struct *p; 7648 int detached = 0; 7649 7650 lockdep_assert_held(&env->src_rq->lock); 7651 7652 if (env->imbalance <= 0) 7653 return 0; 7654 7655 while (!list_empty(tasks)) { 7656 /* 7657 * We don't want to steal all, otherwise we may be treated likewise, 7658 * which could at worst lead to a livelock crash. 7659 */ 7660 if (env->idle != CPU_NOT_IDLE && env->src_rq->nr_running <= 1) 7661 break; 7662 7663 p = list_last_entry(tasks, struct task_struct, se.group_node); 7664 7665 env->loop++; 7666 /* We've more or less seen every task there is, call it quits */ 7667 if (env->loop > env->loop_max) 7668 break; 7669 7670 /* take a breather every nr_migrate tasks */ 7671 if (env->loop > env->loop_break) { 7672 env->loop_break += sched_nr_migrate_break; 7673 env->flags |= LBF_NEED_BREAK; 7674 break; 7675 } 7676 7677 if (!can_migrate_task(p, env)) 7678 goto next; 7679 7680 switch (env->migration_type) { 7681 case migrate_load: 7682 /* 7683 * Depending of the number of CPUs and tasks and the 7684 * cgroup hierarchy, task_h_load() can return a null 7685 * value. Make sure that env->imbalance decreases 7686 * otherwise detach_tasks() will stop only after 7687 * detaching up to loop_max tasks. 7688 */ 7689 load = max_t(unsigned long, task_h_load(p), 1); 7690 7691 if (sched_feat(LB_MIN) && 7692 load < 16 && !env->sd->nr_balance_failed) 7693 goto next; 7694 7695 /* 7696 * Make sure that we don't migrate too much load. 7697 * Nevertheless, let relax the constraint if 7698 * scheduler fails to find a good waiting task to 7699 * migrate. 7700 */ 7701 7702 if ((load >> env->sd->nr_balance_failed) > env->imbalance) 7703 goto next; 7704 7705 env->imbalance -= load; 7706 break; 7707 7708 case migrate_util: 7709 util = task_util_est(p); 7710 7711 if (util > env->imbalance) 7712 goto next; 7713 7714 env->imbalance -= util; 7715 break; 7716 7717 case migrate_task: 7718 env->imbalance--; 7719 break; 7720 7721 case migrate_misfit: 7722 /* This is not a misfit task */ 7723 if (task_fits_capacity(p, capacity_of(env->src_cpu))) 7724 goto next; 7725 7726 env->imbalance = 0; 7727 break; 7728 } 7729 7730 detach_task(p, env); 7731 list_add(&p->se.group_node, &env->tasks); 7732 7733 detached++; 7734 7735 #ifdef CONFIG_PREEMPTION 7736 /* 7737 * NEWIDLE balancing is a source of latency, so preemptible 7738 * kernels will stop after the first task is detached to minimize 7739 * the critical section. 7740 */ 7741 if (env->idle == CPU_NEWLY_IDLE) 7742 break; 7743 #endif 7744 7745 /* 7746 * We only want to steal up to the prescribed amount of 7747 * load/util/tasks. 7748 */ 7749 if (env->imbalance <= 0) 7750 break; 7751 7752 continue; 7753 next: 7754 list_move(&p->se.group_node, tasks); 7755 } 7756 7757 /* 7758 * Right now, this is one of only two places we collect this stat 7759 * so we can safely collect detach_one_task() stats here rather 7760 * than inside detach_one_task(). 7761 */ 7762 schedstat_add(env->sd->lb_gained[env->idle], detached); 7763 7764 return detached; 7765 } 7766 7767 /* 7768 * attach_task() -- attach the task detached by detach_task() to its new rq. 7769 */ 7770 static void attach_task(struct rq *rq, struct task_struct *p) 7771 { 7772 lockdep_assert_held(&rq->lock); 7773 7774 BUG_ON(task_rq(p) != rq); 7775 activate_task(rq, p, ENQUEUE_NOCLOCK); 7776 check_preempt_curr(rq, p, 0); 7777 } 7778 7779 /* 7780 * attach_one_task() -- attaches the task returned from detach_one_task() to 7781 * its new rq. 7782 */ 7783 static void attach_one_task(struct rq *rq, struct task_struct *p) 7784 { 7785 struct rq_flags rf; 7786 7787 rq_lock(rq, &rf); 7788 update_rq_clock(rq); 7789 attach_task(rq, p); 7790 rq_unlock(rq, &rf); 7791 } 7792 7793 /* 7794 * attach_tasks() -- attaches all tasks detached by detach_tasks() to their 7795 * new rq. 7796 */ 7797 static void attach_tasks(struct lb_env *env) 7798 { 7799 struct list_head *tasks = &env->tasks; 7800 struct task_struct *p; 7801 struct rq_flags rf; 7802 7803 rq_lock(env->dst_rq, &rf); 7804 update_rq_clock(env->dst_rq); 7805 7806 while (!list_empty(tasks)) { 7807 p = list_first_entry(tasks, struct task_struct, se.group_node); 7808 list_del_init(&p->se.group_node); 7809 7810 attach_task(env->dst_rq, p); 7811 } 7812 7813 rq_unlock(env->dst_rq, &rf); 7814 } 7815 7816 #ifdef CONFIG_NO_HZ_COMMON 7817 static inline bool cfs_rq_has_blocked(struct cfs_rq *cfs_rq) 7818 { 7819 if (cfs_rq->avg.load_avg) 7820 return true; 7821 7822 if (cfs_rq->avg.util_avg) 7823 return true; 7824 7825 return false; 7826 } 7827 7828 static inline bool others_have_blocked(struct rq *rq) 7829 { 7830 if (READ_ONCE(rq->avg_rt.util_avg)) 7831 return true; 7832 7833 if (READ_ONCE(rq->avg_dl.util_avg)) 7834 return true; 7835 7836 if (thermal_load_avg(rq)) 7837 return true; 7838 7839 #ifdef CONFIG_HAVE_SCHED_AVG_IRQ 7840 if (READ_ONCE(rq->avg_irq.util_avg)) 7841 return true; 7842 #endif 7843 7844 return false; 7845 } 7846 7847 static inline void update_blocked_load_status(struct rq *rq, bool has_blocked) 7848 { 7849 rq->last_blocked_load_update_tick = jiffies; 7850 7851 if (!has_blocked) 7852 rq->has_blocked_load = 0; 7853 } 7854 #else 7855 static inline bool cfs_rq_has_blocked(struct cfs_rq *cfs_rq) { return false; } 7856 static inline bool others_have_blocked(struct rq *rq) { return false; } 7857 static inline void update_blocked_load_status(struct rq *rq, bool has_blocked) {} 7858 #endif 7859 7860 static bool __update_blocked_others(struct rq *rq, bool *done) 7861 { 7862 const struct sched_class *curr_class; 7863 u64 now = rq_clock_pelt(rq); 7864 unsigned long thermal_pressure; 7865 bool decayed; 7866 7867 /* 7868 * update_load_avg() can call cpufreq_update_util(). Make sure that RT, 7869 * DL and IRQ signals have been updated before updating CFS. 7870 */ 7871 curr_class = rq->curr->sched_class; 7872 7873 thermal_pressure = arch_scale_thermal_pressure(cpu_of(rq)); 7874 7875 decayed = update_rt_rq_load_avg(now, rq, curr_class == &rt_sched_class) | 7876 update_dl_rq_load_avg(now, rq, curr_class == &dl_sched_class) | 7877 update_thermal_load_avg(rq_clock_thermal(rq), rq, thermal_pressure) | 7878 update_irq_load_avg(rq, 0); 7879 7880 if (others_have_blocked(rq)) 7881 *done = false; 7882 7883 return decayed; 7884 } 7885 7886 #ifdef CONFIG_FAIR_GROUP_SCHED 7887 7888 static inline bool cfs_rq_is_decayed(struct cfs_rq *cfs_rq) 7889 { 7890 if (cfs_rq->load.weight) 7891 return false; 7892 7893 if (cfs_rq->avg.load_sum) 7894 return false; 7895 7896 if (cfs_rq->avg.util_sum) 7897 return false; 7898 7899 if (cfs_rq->avg.runnable_sum) 7900 return false; 7901 7902 return true; 7903 } 7904 7905 static bool __update_blocked_fair(struct rq *rq, bool *done) 7906 { 7907 struct cfs_rq *cfs_rq, *pos; 7908 bool decayed = false; 7909 int cpu = cpu_of(rq); 7910 7911 /* 7912 * Iterates the task_group tree in a bottom up fashion, see 7913 * list_add_leaf_cfs_rq() for details. 7914 */ 7915 for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos) { 7916 struct sched_entity *se; 7917 7918 if (update_cfs_rq_load_avg(cfs_rq_clock_pelt(cfs_rq), cfs_rq)) { 7919 update_tg_load_avg(cfs_rq); 7920 7921 if (cfs_rq == &rq->cfs) 7922 decayed = true; 7923 } 7924 7925 /* Propagate pending load changes to the parent, if any: */ 7926 se = cfs_rq->tg->se[cpu]; 7927 if (se && !skip_blocked_update(se)) 7928 update_load_avg(cfs_rq_of(se), se, 0); 7929 7930 /* 7931 * There can be a lot of idle CPU cgroups. Don't let fully 7932 * decayed cfs_rqs linger on the list. 7933 */ 7934 if (cfs_rq_is_decayed(cfs_rq)) 7935 list_del_leaf_cfs_rq(cfs_rq); 7936 7937 /* Don't need periodic decay once load/util_avg are null */ 7938 if (cfs_rq_has_blocked(cfs_rq)) 7939 *done = false; 7940 } 7941 7942 return decayed; 7943 } 7944 7945 /* 7946 * Compute the hierarchical load factor for cfs_rq and all its ascendants. 7947 * This needs to be done in a top-down fashion because the load of a child 7948 * group is a fraction of its parents load. 7949 */ 7950 static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq) 7951 { 7952 struct rq *rq = rq_of(cfs_rq); 7953 struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)]; 7954 unsigned long now = jiffies; 7955 unsigned long load; 7956 7957 if (cfs_rq->last_h_load_update == now) 7958 return; 7959 7960 WRITE_ONCE(cfs_rq->h_load_next, NULL); 7961 for_each_sched_entity(se) { 7962 cfs_rq = cfs_rq_of(se); 7963 WRITE_ONCE(cfs_rq->h_load_next, se); 7964 if (cfs_rq->last_h_load_update == now) 7965 break; 7966 } 7967 7968 if (!se) { 7969 cfs_rq->h_load = cfs_rq_load_avg(cfs_rq); 7970 cfs_rq->last_h_load_update = now; 7971 } 7972 7973 while ((se = READ_ONCE(cfs_rq->h_load_next)) != NULL) { 7974 load = cfs_rq->h_load; 7975 load = div64_ul(load * se->avg.load_avg, 7976 cfs_rq_load_avg(cfs_rq) + 1); 7977 cfs_rq = group_cfs_rq(se); 7978 cfs_rq->h_load = load; 7979 cfs_rq->last_h_load_update = now; 7980 } 7981 } 7982 7983 static unsigned long task_h_load(struct task_struct *p) 7984 { 7985 struct cfs_rq *cfs_rq = task_cfs_rq(p); 7986 7987 update_cfs_rq_h_load(cfs_rq); 7988 return div64_ul(p->se.avg.load_avg * cfs_rq->h_load, 7989 cfs_rq_load_avg(cfs_rq) + 1); 7990 } 7991 #else 7992 static bool __update_blocked_fair(struct rq *rq, bool *done) 7993 { 7994 struct cfs_rq *cfs_rq = &rq->cfs; 7995 bool decayed; 7996 7997 decayed = update_cfs_rq_load_avg(cfs_rq_clock_pelt(cfs_rq), cfs_rq); 7998 if (cfs_rq_has_blocked(cfs_rq)) 7999 *done = false; 8000 8001 return decayed; 8002 } 8003 8004 static unsigned long task_h_load(struct task_struct *p) 8005 { 8006 return p->se.avg.load_avg; 8007 } 8008 #endif 8009 8010 static void update_blocked_averages(int cpu) 8011 { 8012 bool decayed = false, done = true; 8013 struct rq *rq = cpu_rq(cpu); 8014 struct rq_flags rf; 8015 8016 rq_lock_irqsave(rq, &rf); 8017 update_rq_clock(rq); 8018 8019 decayed |= __update_blocked_others(rq, &done); 8020 decayed |= __update_blocked_fair(rq, &done); 8021 8022 update_blocked_load_status(rq, !done); 8023 if (decayed) 8024 cpufreq_update_util(rq, 0); 8025 rq_unlock_irqrestore(rq, &rf); 8026 } 8027 8028 /********** Helpers for find_busiest_group ************************/ 8029 8030 /* 8031 * sg_lb_stats - stats of a sched_group required for load_balancing 8032 */ 8033 struct sg_lb_stats { 8034 unsigned long avg_load; /*Avg load across the CPUs of the group */ 8035 unsigned long group_load; /* Total load over the CPUs of the group */ 8036 unsigned long group_capacity; 8037 unsigned long group_util; /* Total utilization over the CPUs of the group */ 8038 unsigned long group_runnable; /* Total runnable time over the CPUs of the group */ 8039 unsigned int sum_nr_running; /* Nr of tasks running in the group */ 8040 unsigned int sum_h_nr_running; /* Nr of CFS tasks running in the group */ 8041 unsigned int idle_cpus; 8042 unsigned int group_weight; 8043 enum group_type group_type; 8044 unsigned int group_asym_packing; /* Tasks should be moved to preferred CPU */ 8045 unsigned long group_misfit_task_load; /* A CPU has a task too big for its capacity */ 8046 #ifdef CONFIG_NUMA_BALANCING 8047 unsigned int nr_numa_running; 8048 unsigned int nr_preferred_running; 8049 #endif 8050 }; 8051 8052 /* 8053 * sd_lb_stats - Structure to store the statistics of a sched_domain 8054 * during load balancing. 8055 */ 8056 struct sd_lb_stats { 8057 struct sched_group *busiest; /* Busiest group in this sd */ 8058 struct sched_group *local; /* Local group in this sd */ 8059 unsigned long total_load; /* Total load of all groups in sd */ 8060 unsigned long total_capacity; /* Total capacity of all groups in sd */ 8061 unsigned long avg_load; /* Average load across all groups in sd */ 8062 unsigned int prefer_sibling; /* tasks should go to sibling first */ 8063 8064 struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */ 8065 struct sg_lb_stats local_stat; /* Statistics of the local group */ 8066 }; 8067 8068 static inline void init_sd_lb_stats(struct sd_lb_stats *sds) 8069 { 8070 /* 8071 * Skimp on the clearing to avoid duplicate work. We can avoid clearing 8072 * local_stat because update_sg_lb_stats() does a full clear/assignment. 8073 * We must however set busiest_stat::group_type and 8074 * busiest_stat::idle_cpus to the worst busiest group because 8075 * update_sd_pick_busiest() reads these before assignment. 8076 */ 8077 *sds = (struct sd_lb_stats){ 8078 .busiest = NULL, 8079 .local = NULL, 8080 .total_load = 0UL, 8081 .total_capacity = 0UL, 8082 .busiest_stat = { 8083 .idle_cpus = UINT_MAX, 8084 .group_type = group_has_spare, 8085 }, 8086 }; 8087 } 8088 8089 static unsigned long scale_rt_capacity(int cpu) 8090 { 8091 struct rq *rq = cpu_rq(cpu); 8092 unsigned long max = arch_scale_cpu_capacity(cpu); 8093 unsigned long used, free; 8094 unsigned long irq; 8095 8096 irq = cpu_util_irq(rq); 8097 8098 if (unlikely(irq >= max)) 8099 return 1; 8100 8101 /* 8102 * avg_rt.util_avg and avg_dl.util_avg track binary signals 8103 * (running and not running) with weights 0 and 1024 respectively. 8104 * avg_thermal.load_avg tracks thermal pressure and the weighted 8105 * average uses the actual delta max capacity(load). 8106 */ 8107 used = READ_ONCE(rq->avg_rt.util_avg); 8108 used += READ_ONCE(rq->avg_dl.util_avg); 8109 used += thermal_load_avg(rq); 8110 8111 if (unlikely(used >= max)) 8112 return 1; 8113 8114 free = max - used; 8115 8116 return scale_irq_capacity(free, irq, max); 8117 } 8118 8119 static void update_cpu_capacity(struct sched_domain *sd, int cpu) 8120 { 8121 unsigned long capacity = scale_rt_capacity(cpu); 8122 struct sched_group *sdg = sd->groups; 8123 8124 cpu_rq(cpu)->cpu_capacity_orig = arch_scale_cpu_capacity(cpu); 8125 8126 if (!capacity) 8127 capacity = 1; 8128 8129 cpu_rq(cpu)->cpu_capacity = capacity; 8130 trace_sched_cpu_capacity_tp(cpu_rq(cpu)); 8131 8132 sdg->sgc->capacity = capacity; 8133 sdg->sgc->min_capacity = capacity; 8134 sdg->sgc->max_capacity = capacity; 8135 } 8136 8137 void update_group_capacity(struct sched_domain *sd, int cpu) 8138 { 8139 struct sched_domain *child = sd->child; 8140 struct sched_group *group, *sdg = sd->groups; 8141 unsigned long capacity, min_capacity, max_capacity; 8142 unsigned long interval; 8143 8144 interval = msecs_to_jiffies(sd->balance_interval); 8145 interval = clamp(interval, 1UL, max_load_balance_interval); 8146 sdg->sgc->next_update = jiffies + interval; 8147 8148 if (!child) { 8149 update_cpu_capacity(sd, cpu); 8150 return; 8151 } 8152 8153 capacity = 0; 8154 min_capacity = ULONG_MAX; 8155 max_capacity = 0; 8156 8157 if (child->flags & SD_OVERLAP) { 8158 /* 8159 * SD_OVERLAP domains cannot assume that child groups 8160 * span the current group. 8161 */ 8162 8163 for_each_cpu(cpu, sched_group_span(sdg)) { 8164 unsigned long cpu_cap = capacity_of(cpu); 8165 8166 capacity += cpu_cap; 8167 min_capacity = min(cpu_cap, min_capacity); 8168 max_capacity = max(cpu_cap, max_capacity); 8169 } 8170 } else { 8171 /* 8172 * !SD_OVERLAP domains can assume that child groups 8173 * span the current group. 8174 */ 8175 8176 group = child->groups; 8177 do { 8178 struct sched_group_capacity *sgc = group->sgc; 8179 8180 capacity += sgc->capacity; 8181 min_capacity = min(sgc->min_capacity, min_capacity); 8182 max_capacity = max(sgc->max_capacity, max_capacity); 8183 group = group->next; 8184 } while (group != child->groups); 8185 } 8186 8187 sdg->sgc->capacity = capacity; 8188 sdg->sgc->min_capacity = min_capacity; 8189 sdg->sgc->max_capacity = max_capacity; 8190 } 8191 8192 /* 8193 * Check whether the capacity of the rq has been noticeably reduced by side 8194 * activity. The imbalance_pct is used for the threshold. 8195 * Return true is the capacity is reduced 8196 */ 8197 static inline int 8198 check_cpu_capacity(struct rq *rq, struct sched_domain *sd) 8199 { 8200 return ((rq->cpu_capacity * sd->imbalance_pct) < 8201 (rq->cpu_capacity_orig * 100)); 8202 } 8203 8204 /* 8205 * Check whether a rq has a misfit task and if it looks like we can actually 8206 * help that task: we can migrate the task to a CPU of higher capacity, or 8207 * the task's current CPU is heavily pressured. 8208 */ 8209 static inline int check_misfit_status(struct rq *rq, struct sched_domain *sd) 8210 { 8211 return rq->misfit_task_load && 8212 (rq->cpu_capacity_orig < rq->rd->max_cpu_capacity || 8213 check_cpu_capacity(rq, sd)); 8214 } 8215 8216 /* 8217 * Group imbalance indicates (and tries to solve) the problem where balancing 8218 * groups is inadequate due to ->cpus_ptr constraints. 8219 * 8220 * Imagine a situation of two groups of 4 CPUs each and 4 tasks each with a 8221 * cpumask covering 1 CPU of the first group and 3 CPUs of the second group. 8222 * Something like: 8223 * 8224 * { 0 1 2 3 } { 4 5 6 7 } 8225 * * * * * 8226 * 8227 * If we were to balance group-wise we'd place two tasks in the first group and 8228 * two tasks in the second group. Clearly this is undesired as it will overload 8229 * cpu 3 and leave one of the CPUs in the second group unused. 8230 * 8231 * The current solution to this issue is detecting the skew in the first group 8232 * by noticing the lower domain failed to reach balance and had difficulty 8233 * moving tasks due to affinity constraints. 8234 * 8235 * When this is so detected; this group becomes a candidate for busiest; see 8236 * update_sd_pick_busiest(). And calculate_imbalance() and 8237 * find_busiest_group() avoid some of the usual balance conditions to allow it 8238 * to create an effective group imbalance. 8239 * 8240 * This is a somewhat tricky proposition since the next run might not find the 8241 * group imbalance and decide the groups need to be balanced again. A most 8242 * subtle and fragile situation. 8243 */ 8244 8245 static inline int sg_imbalanced(struct sched_group *group) 8246 { 8247 return group->sgc->imbalance; 8248 } 8249 8250 /* 8251 * group_has_capacity returns true if the group has spare capacity that could 8252 * be used by some tasks. 8253 * We consider that a group has spare capacity if the * number of task is 8254 * smaller than the number of CPUs or if the utilization is lower than the 8255 * available capacity for CFS tasks. 8256 * For the latter, we use a threshold to stabilize the state, to take into 8257 * account the variance of the tasks' load and to return true if the available 8258 * capacity in meaningful for the load balancer. 8259 * As an example, an available capacity of 1% can appear but it doesn't make 8260 * any benefit for the load balance. 8261 */ 8262 static inline bool 8263 group_has_capacity(unsigned int imbalance_pct, struct sg_lb_stats *sgs) 8264 { 8265 if (sgs->sum_nr_running < sgs->group_weight) 8266 return true; 8267 8268 if ((sgs->group_capacity * imbalance_pct) < 8269 (sgs->group_runnable * 100)) 8270 return false; 8271 8272 if ((sgs->group_capacity * 100) > 8273 (sgs->group_util * imbalance_pct)) 8274 return true; 8275 8276 return false; 8277 } 8278 8279 /* 8280 * group_is_overloaded returns true if the group has more tasks than it can 8281 * handle. 8282 * group_is_overloaded is not equals to !group_has_capacity because a group 8283 * with the exact right number of tasks, has no more spare capacity but is not 8284 * overloaded so both group_has_capacity and group_is_overloaded return 8285 * false. 8286 */ 8287 static inline bool 8288 group_is_overloaded(unsigned int imbalance_pct, struct sg_lb_stats *sgs) 8289 { 8290 if (sgs->sum_nr_running <= sgs->group_weight) 8291 return false; 8292 8293 if ((sgs->group_capacity * 100) < 8294 (sgs->group_util * imbalance_pct)) 8295 return true; 8296 8297 if ((sgs->group_capacity * imbalance_pct) < 8298 (sgs->group_runnable * 100)) 8299 return true; 8300 8301 return false; 8302 } 8303 8304 /* 8305 * group_smaller_min_cpu_capacity: Returns true if sched_group sg has smaller 8306 * per-CPU capacity than sched_group ref. 8307 */ 8308 static inline bool 8309 group_smaller_min_cpu_capacity(struct sched_group *sg, struct sched_group *ref) 8310 { 8311 return fits_capacity(sg->sgc->min_capacity, ref->sgc->min_capacity); 8312 } 8313 8314 /* 8315 * group_smaller_max_cpu_capacity: Returns true if sched_group sg has smaller 8316 * per-CPU capacity_orig than sched_group ref. 8317 */ 8318 static inline bool 8319 group_smaller_max_cpu_capacity(struct sched_group *sg, struct sched_group *ref) 8320 { 8321 return fits_capacity(sg->sgc->max_capacity, ref->sgc->max_capacity); 8322 } 8323 8324 static inline enum 8325 group_type group_classify(unsigned int imbalance_pct, 8326 struct sched_group *group, 8327 struct sg_lb_stats *sgs) 8328 { 8329 if (group_is_overloaded(imbalance_pct, sgs)) 8330 return group_overloaded; 8331 8332 if (sg_imbalanced(group)) 8333 return group_imbalanced; 8334 8335 if (sgs->group_asym_packing) 8336 return group_asym_packing; 8337 8338 if (sgs->group_misfit_task_load) 8339 return group_misfit_task; 8340 8341 if (!group_has_capacity(imbalance_pct, sgs)) 8342 return group_fully_busy; 8343 8344 return group_has_spare; 8345 } 8346 8347 static bool update_nohz_stats(struct rq *rq, bool force) 8348 { 8349 #ifdef CONFIG_NO_HZ_COMMON 8350 unsigned int cpu = rq->cpu; 8351 8352 if (!rq->has_blocked_load) 8353 return false; 8354 8355 if (!cpumask_test_cpu(cpu, nohz.idle_cpus_mask)) 8356 return false; 8357 8358 if (!force && !time_after(jiffies, rq->last_blocked_load_update_tick)) 8359 return true; 8360 8361 update_blocked_averages(cpu); 8362 8363 return rq->has_blocked_load; 8364 #else 8365 return false; 8366 #endif 8367 } 8368 8369 /** 8370 * update_sg_lb_stats - Update sched_group's statistics for load balancing. 8371 * @env: The load balancing environment. 8372 * @group: sched_group whose statistics are to be updated. 8373 * @sgs: variable to hold the statistics for this group. 8374 * @sg_status: Holds flag indicating the status of the sched_group 8375 */ 8376 static inline void update_sg_lb_stats(struct lb_env *env, 8377 struct sched_group *group, 8378 struct sg_lb_stats *sgs, 8379 int *sg_status) 8380 { 8381 int i, nr_running, local_group; 8382 8383 memset(sgs, 0, sizeof(*sgs)); 8384 8385 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_span(group)); 8386 8387 for_each_cpu_and(i, sched_group_span(group), env->cpus) { 8388 struct rq *rq = cpu_rq(i); 8389 8390 if ((env->flags & LBF_NOHZ_STATS) && update_nohz_stats(rq, false)) 8391 env->flags |= LBF_NOHZ_AGAIN; 8392 8393 sgs->group_load += cpu_load(rq); 8394 sgs->group_util += cpu_util(i); 8395 sgs->group_runnable += cpu_runnable(rq); 8396 sgs->sum_h_nr_running += rq->cfs.h_nr_running; 8397 8398 nr_running = rq->nr_running; 8399 sgs->sum_nr_running += nr_running; 8400 8401 if (nr_running > 1) 8402 *sg_status |= SG_OVERLOAD; 8403 8404 if (cpu_overutilized(i)) 8405 *sg_status |= SG_OVERUTILIZED; 8406 8407 #ifdef CONFIG_NUMA_BALANCING 8408 sgs->nr_numa_running += rq->nr_numa_running; 8409 sgs->nr_preferred_running += rq->nr_preferred_running; 8410 #endif 8411 /* 8412 * No need to call idle_cpu() if nr_running is not 0 8413 */ 8414 if (!nr_running && idle_cpu(i)) { 8415 sgs->idle_cpus++; 8416 /* Idle cpu can't have misfit task */ 8417 continue; 8418 } 8419 8420 if (local_group) 8421 continue; 8422 8423 /* Check for a misfit task on the cpu */ 8424 if (env->sd->flags & SD_ASYM_CPUCAPACITY && 8425 sgs->group_misfit_task_load < rq->misfit_task_load) { 8426 sgs->group_misfit_task_load = rq->misfit_task_load; 8427 *sg_status |= SG_OVERLOAD; 8428 } 8429 } 8430 8431 /* Check if dst CPU is idle and preferred to this group */ 8432 if (env->sd->flags & SD_ASYM_PACKING && 8433 env->idle != CPU_NOT_IDLE && 8434 sgs->sum_h_nr_running && 8435 sched_asym_prefer(env->dst_cpu, group->asym_prefer_cpu)) { 8436 sgs->group_asym_packing = 1; 8437 } 8438 8439 sgs->group_capacity = group->sgc->capacity; 8440 8441 sgs->group_weight = group->group_weight; 8442 8443 sgs->group_type = group_classify(env->sd->imbalance_pct, group, sgs); 8444 8445 /* Computing avg_load makes sense only when group is overloaded */ 8446 if (sgs->group_type == group_overloaded) 8447 sgs->avg_load = (sgs->group_load * SCHED_CAPACITY_SCALE) / 8448 sgs->group_capacity; 8449 } 8450 8451 /** 8452 * update_sd_pick_busiest - return 1 on busiest group 8453 * @env: The load balancing environment. 8454 * @sds: sched_domain statistics 8455 * @sg: sched_group candidate to be checked for being the busiest 8456 * @sgs: sched_group statistics 8457 * 8458 * Determine if @sg is a busier group than the previously selected 8459 * busiest group. 8460 * 8461 * Return: %true if @sg is a busier group than the previously selected 8462 * busiest group. %false otherwise. 8463 */ 8464 static bool update_sd_pick_busiest(struct lb_env *env, 8465 struct sd_lb_stats *sds, 8466 struct sched_group *sg, 8467 struct sg_lb_stats *sgs) 8468 { 8469 struct sg_lb_stats *busiest = &sds->busiest_stat; 8470 8471 /* Make sure that there is at least one task to pull */ 8472 if (!sgs->sum_h_nr_running) 8473 return false; 8474 8475 /* 8476 * Don't try to pull misfit tasks we can't help. 8477 * We can use max_capacity here as reduction in capacity on some 8478 * CPUs in the group should either be possible to resolve 8479 * internally or be covered by avg_load imbalance (eventually). 8480 */ 8481 if (sgs->group_type == group_misfit_task && 8482 (!group_smaller_max_cpu_capacity(sg, sds->local) || 8483 sds->local_stat.group_type != group_has_spare)) 8484 return false; 8485 8486 if (sgs->group_type > busiest->group_type) 8487 return true; 8488 8489 if (sgs->group_type < busiest->group_type) 8490 return false; 8491 8492 /* 8493 * The candidate and the current busiest group are the same type of 8494 * group. Let check which one is the busiest according to the type. 8495 */ 8496 8497 switch (sgs->group_type) { 8498 case group_overloaded: 8499 /* Select the overloaded group with highest avg_load. */ 8500 if (sgs->avg_load <= busiest->avg_load) 8501 return false; 8502 break; 8503 8504 case group_imbalanced: 8505 /* 8506 * Select the 1st imbalanced group as we don't have any way to 8507 * choose one more than another. 8508 */ 8509 return false; 8510 8511 case group_asym_packing: 8512 /* Prefer to move from lowest priority CPU's work */ 8513 if (sched_asym_prefer(sg->asym_prefer_cpu, sds->busiest->asym_prefer_cpu)) 8514 return false; 8515 break; 8516 8517 case group_misfit_task: 8518 /* 8519 * If we have more than one misfit sg go with the biggest 8520 * misfit. 8521 */ 8522 if (sgs->group_misfit_task_load < busiest->group_misfit_task_load) 8523 return false; 8524 break; 8525 8526 case group_fully_busy: 8527 /* 8528 * Select the fully busy group with highest avg_load. In 8529 * theory, there is no need to pull task from such kind of 8530 * group because tasks have all compute capacity that they need 8531 * but we can still improve the overall throughput by reducing 8532 * contention when accessing shared HW resources. 8533 * 8534 * XXX for now avg_load is not computed and always 0 so we 8535 * select the 1st one. 8536 */ 8537 if (sgs->avg_load <= busiest->avg_load) 8538 return false; 8539 break; 8540 8541 case group_has_spare: 8542 /* 8543 * Select not overloaded group with lowest number of idle cpus 8544 * and highest number of running tasks. We could also compare 8545 * the spare capacity which is more stable but it can end up 8546 * that the group has less spare capacity but finally more idle 8547 * CPUs which means less opportunity to pull tasks. 8548 */ 8549 if (sgs->idle_cpus > busiest->idle_cpus) 8550 return false; 8551 else if ((sgs->idle_cpus == busiest->idle_cpus) && 8552 (sgs->sum_nr_running <= busiest->sum_nr_running)) 8553 return false; 8554 8555 break; 8556 } 8557 8558 /* 8559 * Candidate sg has no more than one task per CPU and has higher 8560 * per-CPU capacity. Migrating tasks to less capable CPUs may harm 8561 * throughput. Maximize throughput, power/energy consequences are not 8562 * considered. 8563 */ 8564 if ((env->sd->flags & SD_ASYM_CPUCAPACITY) && 8565 (sgs->group_type <= group_fully_busy) && 8566 (group_smaller_min_cpu_capacity(sds->local, sg))) 8567 return false; 8568 8569 return true; 8570 } 8571 8572 #ifdef CONFIG_NUMA_BALANCING 8573 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs) 8574 { 8575 if (sgs->sum_h_nr_running > sgs->nr_numa_running) 8576 return regular; 8577 if (sgs->sum_h_nr_running > sgs->nr_preferred_running) 8578 return remote; 8579 return all; 8580 } 8581 8582 static inline enum fbq_type fbq_classify_rq(struct rq *rq) 8583 { 8584 if (rq->nr_running > rq->nr_numa_running) 8585 return regular; 8586 if (rq->nr_running > rq->nr_preferred_running) 8587 return remote; 8588 return all; 8589 } 8590 #else 8591 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs) 8592 { 8593 return all; 8594 } 8595 8596 static inline enum fbq_type fbq_classify_rq(struct rq *rq) 8597 { 8598 return regular; 8599 } 8600 #endif /* CONFIG_NUMA_BALANCING */ 8601 8602 8603 struct sg_lb_stats; 8604 8605 /* 8606 * task_running_on_cpu - return 1 if @p is running on @cpu. 8607 */ 8608 8609 static unsigned int task_running_on_cpu(int cpu, struct task_struct *p) 8610 { 8611 /* Task has no contribution or is new */ 8612 if (cpu != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time)) 8613 return 0; 8614 8615 if (task_on_rq_queued(p)) 8616 return 1; 8617 8618 return 0; 8619 } 8620 8621 /** 8622 * idle_cpu_without - would a given CPU be idle without p ? 8623 * @cpu: the processor on which idleness is tested. 8624 * @p: task which should be ignored. 8625 * 8626 * Return: 1 if the CPU would be idle. 0 otherwise. 8627 */ 8628 static int idle_cpu_without(int cpu, struct task_struct *p) 8629 { 8630 struct rq *rq = cpu_rq(cpu); 8631 8632 if (rq->curr != rq->idle && rq->curr != p) 8633 return 0; 8634 8635 /* 8636 * rq->nr_running can't be used but an updated version without the 8637 * impact of p on cpu must be used instead. The updated nr_running 8638 * be computed and tested before calling idle_cpu_without(). 8639 */ 8640 8641 #ifdef CONFIG_SMP 8642 if (rq->ttwu_pending) 8643 return 0; 8644 #endif 8645 8646 return 1; 8647 } 8648 8649 /* 8650 * update_sg_wakeup_stats - Update sched_group's statistics for wakeup. 8651 * @sd: The sched_domain level to look for idlest group. 8652 * @group: sched_group whose statistics are to be updated. 8653 * @sgs: variable to hold the statistics for this group. 8654 * @p: The task for which we look for the idlest group/CPU. 8655 */ 8656 static inline void update_sg_wakeup_stats(struct sched_domain *sd, 8657 struct sched_group *group, 8658 struct sg_lb_stats *sgs, 8659 struct task_struct *p) 8660 { 8661 int i, nr_running; 8662 8663 memset(sgs, 0, sizeof(*sgs)); 8664 8665 for_each_cpu(i, sched_group_span(group)) { 8666 struct rq *rq = cpu_rq(i); 8667 unsigned int local; 8668 8669 sgs->group_load += cpu_load_without(rq, p); 8670 sgs->group_util += cpu_util_without(i, p); 8671 sgs->group_runnable += cpu_runnable_without(rq, p); 8672 local = task_running_on_cpu(i, p); 8673 sgs->sum_h_nr_running += rq->cfs.h_nr_running - local; 8674 8675 nr_running = rq->nr_running - local; 8676 sgs->sum_nr_running += nr_running; 8677 8678 /* 8679 * No need to call idle_cpu_without() if nr_running is not 0 8680 */ 8681 if (!nr_running && idle_cpu_without(i, p)) 8682 sgs->idle_cpus++; 8683 8684 } 8685 8686 /* Check if task fits in the group */ 8687 if (sd->flags & SD_ASYM_CPUCAPACITY && 8688 !task_fits_capacity(p, group->sgc->max_capacity)) { 8689 sgs->group_misfit_task_load = 1; 8690 } 8691 8692 sgs->group_capacity = group->sgc->capacity; 8693 8694 sgs->group_weight = group->group_weight; 8695 8696 sgs->group_type = group_classify(sd->imbalance_pct, group, sgs); 8697 8698 /* 8699 * Computing avg_load makes sense only when group is fully busy or 8700 * overloaded 8701 */ 8702 if (sgs->group_type == group_fully_busy || 8703 sgs->group_type == group_overloaded) 8704 sgs->avg_load = (sgs->group_load * SCHED_CAPACITY_SCALE) / 8705 sgs->group_capacity; 8706 } 8707 8708 static bool update_pick_idlest(struct sched_group *idlest, 8709 struct sg_lb_stats *idlest_sgs, 8710 struct sched_group *group, 8711 struct sg_lb_stats *sgs) 8712 { 8713 if (sgs->group_type < idlest_sgs->group_type) 8714 return true; 8715 8716 if (sgs->group_type > idlest_sgs->group_type) 8717 return false; 8718 8719 /* 8720 * The candidate and the current idlest group are the same type of 8721 * group. Let check which one is the idlest according to the type. 8722 */ 8723 8724 switch (sgs->group_type) { 8725 case group_overloaded: 8726 case group_fully_busy: 8727 /* Select the group with lowest avg_load. */ 8728 if (idlest_sgs->avg_load <= sgs->avg_load) 8729 return false; 8730 break; 8731 8732 case group_imbalanced: 8733 case group_asym_packing: 8734 /* Those types are not used in the slow wakeup path */ 8735 return false; 8736 8737 case group_misfit_task: 8738 /* Select group with the highest max capacity */ 8739 if (idlest->sgc->max_capacity >= group->sgc->max_capacity) 8740 return false; 8741 break; 8742 8743 case group_has_spare: 8744 /* Select group with most idle CPUs */ 8745 if (idlest_sgs->idle_cpus > sgs->idle_cpus) 8746 return false; 8747 8748 /* Select group with lowest group_util */ 8749 if (idlest_sgs->idle_cpus == sgs->idle_cpus && 8750 idlest_sgs->group_util <= sgs->group_util) 8751 return false; 8752 8753 break; 8754 } 8755 8756 return true; 8757 } 8758 8759 /* 8760 * find_idlest_group() finds and returns the least busy CPU group within the 8761 * domain. 8762 * 8763 * Assumes p is allowed on at least one CPU in sd. 8764 */ 8765 static struct sched_group * 8766 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu) 8767 { 8768 struct sched_group *idlest = NULL, *local = NULL, *group = sd->groups; 8769 struct sg_lb_stats local_sgs, tmp_sgs; 8770 struct sg_lb_stats *sgs; 8771 unsigned long imbalance; 8772 struct sg_lb_stats idlest_sgs = { 8773 .avg_load = UINT_MAX, 8774 .group_type = group_overloaded, 8775 }; 8776 8777 imbalance = scale_load_down(NICE_0_LOAD) * 8778 (sd->imbalance_pct-100) / 100; 8779 8780 do { 8781 int local_group; 8782 8783 /* Skip over this group if it has no CPUs allowed */ 8784 if (!cpumask_intersects(sched_group_span(group), 8785 p->cpus_ptr)) 8786 continue; 8787 8788 local_group = cpumask_test_cpu(this_cpu, 8789 sched_group_span(group)); 8790 8791 if (local_group) { 8792 sgs = &local_sgs; 8793 local = group; 8794 } else { 8795 sgs = &tmp_sgs; 8796 } 8797 8798 update_sg_wakeup_stats(sd, group, sgs, p); 8799 8800 if (!local_group && update_pick_idlest(idlest, &idlest_sgs, group, sgs)) { 8801 idlest = group; 8802 idlest_sgs = *sgs; 8803 } 8804 8805 } while (group = group->next, group != sd->groups); 8806 8807 8808 /* There is no idlest group to push tasks to */ 8809 if (!idlest) 8810 return NULL; 8811 8812 /* The local group has been skipped because of CPU affinity */ 8813 if (!local) 8814 return idlest; 8815 8816 /* 8817 * If the local group is idler than the selected idlest group 8818 * don't try and push the task. 8819 */ 8820 if (local_sgs.group_type < idlest_sgs.group_type) 8821 return NULL; 8822 8823 /* 8824 * If the local group is busier than the selected idlest group 8825 * try and push the task. 8826 */ 8827 if (local_sgs.group_type > idlest_sgs.group_type) 8828 return idlest; 8829 8830 switch (local_sgs.group_type) { 8831 case group_overloaded: 8832 case group_fully_busy: 8833 /* 8834 * When comparing groups across NUMA domains, it's possible for 8835 * the local domain to be very lightly loaded relative to the 8836 * remote domains but "imbalance" skews the comparison making 8837 * remote CPUs look much more favourable. When considering 8838 * cross-domain, add imbalance to the load on the remote node 8839 * and consider staying local. 8840 */ 8841 8842 if ((sd->flags & SD_NUMA) && 8843 ((idlest_sgs.avg_load + imbalance) >= local_sgs.avg_load)) 8844 return NULL; 8845 8846 /* 8847 * If the local group is less loaded than the selected 8848 * idlest group don't try and push any tasks. 8849 */ 8850 if (idlest_sgs.avg_load >= (local_sgs.avg_load + imbalance)) 8851 return NULL; 8852 8853 if (100 * local_sgs.avg_load <= sd->imbalance_pct * idlest_sgs.avg_load) 8854 return NULL; 8855 break; 8856 8857 case group_imbalanced: 8858 case group_asym_packing: 8859 /* Those type are not used in the slow wakeup path */ 8860 return NULL; 8861 8862 case group_misfit_task: 8863 /* Select group with the highest max capacity */ 8864 if (local->sgc->max_capacity >= idlest->sgc->max_capacity) 8865 return NULL; 8866 break; 8867 8868 case group_has_spare: 8869 if (sd->flags & SD_NUMA) { 8870 #ifdef CONFIG_NUMA_BALANCING 8871 int idlest_cpu; 8872 /* 8873 * If there is spare capacity at NUMA, try to select 8874 * the preferred node 8875 */ 8876 if (cpu_to_node(this_cpu) == p->numa_preferred_nid) 8877 return NULL; 8878 8879 idlest_cpu = cpumask_first(sched_group_span(idlest)); 8880 if (cpu_to_node(idlest_cpu) == p->numa_preferred_nid) 8881 return idlest; 8882 #endif 8883 /* 8884 * Otherwise, keep the task on this node to stay close 8885 * its wakeup source and improve locality. If there is 8886 * a real need of migration, periodic load balance will 8887 * take care of it. 8888 */ 8889 if (local_sgs.idle_cpus) 8890 return NULL; 8891 } 8892 8893 /* 8894 * Select group with highest number of idle CPUs. We could also 8895 * compare the utilization which is more stable but it can end 8896 * up that the group has less spare capacity but finally more 8897 * idle CPUs which means more opportunity to run task. 8898 */ 8899 if (local_sgs.idle_cpus >= idlest_sgs.idle_cpus) 8900 return NULL; 8901 break; 8902 } 8903 8904 return idlest; 8905 } 8906 8907 /** 8908 * update_sd_lb_stats - Update sched_domain's statistics for load balancing. 8909 * @env: The load balancing environment. 8910 * @sds: variable to hold the statistics for this sched_domain. 8911 */ 8912 8913 static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds) 8914 { 8915 struct sched_domain *child = env->sd->child; 8916 struct sched_group *sg = env->sd->groups; 8917 struct sg_lb_stats *local = &sds->local_stat; 8918 struct sg_lb_stats tmp_sgs; 8919 int sg_status = 0; 8920 8921 #ifdef CONFIG_NO_HZ_COMMON 8922 if (env->idle == CPU_NEWLY_IDLE && READ_ONCE(nohz.has_blocked)) 8923 env->flags |= LBF_NOHZ_STATS; 8924 #endif 8925 8926 do { 8927 struct sg_lb_stats *sgs = &tmp_sgs; 8928 int local_group; 8929 8930 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_span(sg)); 8931 if (local_group) { 8932 sds->local = sg; 8933 sgs = local; 8934 8935 if (env->idle != CPU_NEWLY_IDLE || 8936 time_after_eq(jiffies, sg->sgc->next_update)) 8937 update_group_capacity(env->sd, env->dst_cpu); 8938 } 8939 8940 update_sg_lb_stats(env, sg, sgs, &sg_status); 8941 8942 if (local_group) 8943 goto next_group; 8944 8945 8946 if (update_sd_pick_busiest(env, sds, sg, sgs)) { 8947 sds->busiest = sg; 8948 sds->busiest_stat = *sgs; 8949 } 8950 8951 next_group: 8952 /* Now, start updating sd_lb_stats */ 8953 sds->total_load += sgs->group_load; 8954 sds->total_capacity += sgs->group_capacity; 8955 8956 sg = sg->next; 8957 } while (sg != env->sd->groups); 8958 8959 /* Tag domain that child domain prefers tasks go to siblings first */ 8960 sds->prefer_sibling = child && child->flags & SD_PREFER_SIBLING; 8961 8962 #ifdef CONFIG_NO_HZ_COMMON 8963 if ((env->flags & LBF_NOHZ_AGAIN) && 8964 cpumask_subset(nohz.idle_cpus_mask, sched_domain_span(env->sd))) { 8965 8966 WRITE_ONCE(nohz.next_blocked, 8967 jiffies + msecs_to_jiffies(LOAD_AVG_PERIOD)); 8968 } 8969 #endif 8970 8971 if (env->sd->flags & SD_NUMA) 8972 env->fbq_type = fbq_classify_group(&sds->busiest_stat); 8973 8974 if (!env->sd->parent) { 8975 struct root_domain *rd = env->dst_rq->rd; 8976 8977 /* update overload indicator if we are at root domain */ 8978 WRITE_ONCE(rd->overload, sg_status & SG_OVERLOAD); 8979 8980 /* Update over-utilization (tipping point, U >= 0) indicator */ 8981 WRITE_ONCE(rd->overutilized, sg_status & SG_OVERUTILIZED); 8982 trace_sched_overutilized_tp(rd, sg_status & SG_OVERUTILIZED); 8983 } else if (sg_status & SG_OVERUTILIZED) { 8984 struct root_domain *rd = env->dst_rq->rd; 8985 8986 WRITE_ONCE(rd->overutilized, SG_OVERUTILIZED); 8987 trace_sched_overutilized_tp(rd, SG_OVERUTILIZED); 8988 } 8989 } 8990 8991 static inline long adjust_numa_imbalance(int imbalance, int nr_running) 8992 { 8993 unsigned int imbalance_min; 8994 8995 /* 8996 * Allow a small imbalance based on a simple pair of communicating 8997 * tasks that remain local when the source domain is almost idle. 8998 */ 8999 imbalance_min = 2; 9000 if (nr_running <= imbalance_min) 9001 return 0; 9002 9003 return imbalance; 9004 } 9005 9006 /** 9007 * calculate_imbalance - Calculate the amount of imbalance present within the 9008 * groups of a given sched_domain during load balance. 9009 * @env: load balance environment 9010 * @sds: statistics of the sched_domain whose imbalance is to be calculated. 9011 */ 9012 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds) 9013 { 9014 struct sg_lb_stats *local, *busiest; 9015 9016 local = &sds->local_stat; 9017 busiest = &sds->busiest_stat; 9018 9019 if (busiest->group_type == group_misfit_task) { 9020 /* Set imbalance to allow misfit tasks to be balanced. */ 9021 env->migration_type = migrate_misfit; 9022 env->imbalance = 1; 9023 return; 9024 } 9025 9026 if (busiest->group_type == group_asym_packing) { 9027 /* 9028 * In case of asym capacity, we will try to migrate all load to 9029 * the preferred CPU. 9030 */ 9031 env->migration_type = migrate_task; 9032 env->imbalance = busiest->sum_h_nr_running; 9033 return; 9034 } 9035 9036 if (busiest->group_type == group_imbalanced) { 9037 /* 9038 * In the group_imb case we cannot rely on group-wide averages 9039 * to ensure CPU-load equilibrium, try to move any task to fix 9040 * the imbalance. The next load balance will take care of 9041 * balancing back the system. 9042 */ 9043 env->migration_type = migrate_task; 9044 env->imbalance = 1; 9045 return; 9046 } 9047 9048 /* 9049 * Try to use spare capacity of local group without overloading it or 9050 * emptying busiest. 9051 */ 9052 if (local->group_type == group_has_spare) { 9053 if ((busiest->group_type > group_fully_busy) && 9054 !(env->sd->flags & SD_SHARE_PKG_RESOURCES)) { 9055 /* 9056 * If busiest is overloaded, try to fill spare 9057 * capacity. This might end up creating spare capacity 9058 * in busiest or busiest still being overloaded but 9059 * there is no simple way to directly compute the 9060 * amount of load to migrate in order to balance the 9061 * system. 9062 */ 9063 env->migration_type = migrate_util; 9064 env->imbalance = max(local->group_capacity, local->group_util) - 9065 local->group_util; 9066 9067 /* 9068 * In some cases, the group's utilization is max or even 9069 * higher than capacity because of migrations but the 9070 * local CPU is (newly) idle. There is at least one 9071 * waiting task in this overloaded busiest group. Let's 9072 * try to pull it. 9073 */ 9074 if (env->idle != CPU_NOT_IDLE && env->imbalance == 0) { 9075 env->migration_type = migrate_task; 9076 env->imbalance = 1; 9077 } 9078 9079 return; 9080 } 9081 9082 if (busiest->group_weight == 1 || sds->prefer_sibling) { 9083 unsigned int nr_diff = busiest->sum_nr_running; 9084 /* 9085 * When prefer sibling, evenly spread running tasks on 9086 * groups. 9087 */ 9088 env->migration_type = migrate_task; 9089 lsub_positive(&nr_diff, local->sum_nr_running); 9090 env->imbalance = nr_diff >> 1; 9091 } else { 9092 9093 /* 9094 * If there is no overload, we just want to even the number of 9095 * idle cpus. 9096 */ 9097 env->migration_type = migrate_task; 9098 env->imbalance = max_t(long, 0, (local->idle_cpus - 9099 busiest->idle_cpus) >> 1); 9100 } 9101 9102 /* Consider allowing a small imbalance between NUMA groups */ 9103 if (env->sd->flags & SD_NUMA) 9104 env->imbalance = adjust_numa_imbalance(env->imbalance, 9105 busiest->sum_nr_running); 9106 9107 return; 9108 } 9109 9110 /* 9111 * Local is fully busy but has to take more load to relieve the 9112 * busiest group 9113 */ 9114 if (local->group_type < group_overloaded) { 9115 /* 9116 * Local will become overloaded so the avg_load metrics are 9117 * finally needed. 9118 */ 9119 9120 local->avg_load = (local->group_load * SCHED_CAPACITY_SCALE) / 9121 local->group_capacity; 9122 9123 sds->avg_load = (sds->total_load * SCHED_CAPACITY_SCALE) / 9124 sds->total_capacity; 9125 /* 9126 * If the local group is more loaded than the selected 9127 * busiest group don't try to pull any tasks. 9128 */ 9129 if (local->avg_load >= busiest->avg_load) { 9130 env->imbalance = 0; 9131 return; 9132 } 9133 } 9134 9135 /* 9136 * Both group are or will become overloaded and we're trying to get all 9137 * the CPUs to the average_load, so we don't want to push ourselves 9138 * above the average load, nor do we wish to reduce the max loaded CPU 9139 * below the average load. At the same time, we also don't want to 9140 * reduce the group load below the group capacity. Thus we look for 9141 * the minimum possible imbalance. 9142 */ 9143 env->migration_type = migrate_load; 9144 env->imbalance = min( 9145 (busiest->avg_load - sds->avg_load) * busiest->group_capacity, 9146 (sds->avg_load - local->avg_load) * local->group_capacity 9147 ) / SCHED_CAPACITY_SCALE; 9148 } 9149 9150 /******* find_busiest_group() helpers end here *********************/ 9151 9152 /* 9153 * Decision matrix according to the local and busiest group type: 9154 * 9155 * busiest \ local has_spare fully_busy misfit asym imbalanced overloaded 9156 * has_spare nr_idle balanced N/A N/A balanced balanced 9157 * fully_busy nr_idle nr_idle N/A N/A balanced balanced 9158 * misfit_task force N/A N/A N/A force force 9159 * asym_packing force force N/A N/A force force 9160 * imbalanced force force N/A N/A force force 9161 * overloaded force force N/A N/A force avg_load 9162 * 9163 * N/A : Not Applicable because already filtered while updating 9164 * statistics. 9165 * balanced : The system is balanced for these 2 groups. 9166 * force : Calculate the imbalance as load migration is probably needed. 9167 * avg_load : Only if imbalance is significant enough. 9168 * nr_idle : dst_cpu is not busy and the number of idle CPUs is quite 9169 * different in groups. 9170 */ 9171 9172 /** 9173 * find_busiest_group - Returns the busiest group within the sched_domain 9174 * if there is an imbalance. 9175 * 9176 * Also calculates the amount of runnable load which should be moved 9177 * to restore balance. 9178 * 9179 * @env: The load balancing environment. 9180 * 9181 * Return: - The busiest group if imbalance exists. 9182 */ 9183 static struct sched_group *find_busiest_group(struct lb_env *env) 9184 { 9185 struct sg_lb_stats *local, *busiest; 9186 struct sd_lb_stats sds; 9187 9188 init_sd_lb_stats(&sds); 9189 9190 /* 9191 * Compute the various statistics relevant for load balancing at 9192 * this level. 9193 */ 9194 update_sd_lb_stats(env, &sds); 9195 9196 if (sched_energy_enabled()) { 9197 struct root_domain *rd = env->dst_rq->rd; 9198 9199 if (rcu_dereference(rd->pd) && !READ_ONCE(rd->overutilized)) 9200 goto out_balanced; 9201 } 9202 9203 local = &sds.local_stat; 9204 busiest = &sds.busiest_stat; 9205 9206 /* There is no busy sibling group to pull tasks from */ 9207 if (!sds.busiest) 9208 goto out_balanced; 9209 9210 /* Misfit tasks should be dealt with regardless of the avg load */ 9211 if (busiest->group_type == group_misfit_task) 9212 goto force_balance; 9213 9214 /* ASYM feature bypasses nice load balance check */ 9215 if (busiest->group_type == group_asym_packing) 9216 goto force_balance; 9217 9218 /* 9219 * If the busiest group is imbalanced the below checks don't 9220 * work because they assume all things are equal, which typically 9221 * isn't true due to cpus_ptr constraints and the like. 9222 */ 9223 if (busiest->group_type == group_imbalanced) 9224 goto force_balance; 9225 9226 /* 9227 * If the local group is busier than the selected busiest group 9228 * don't try and pull any tasks. 9229 */ 9230 if (local->group_type > busiest->group_type) 9231 goto out_balanced; 9232 9233 /* 9234 * When groups are overloaded, use the avg_load to ensure fairness 9235 * between tasks. 9236 */ 9237 if (local->group_type == group_overloaded) { 9238 /* 9239 * If the local group is more loaded than the selected 9240 * busiest group don't try to pull any tasks. 9241 */ 9242 if (local->avg_load >= busiest->avg_load) 9243 goto out_balanced; 9244 9245 /* XXX broken for overlapping NUMA groups */ 9246 sds.avg_load = (sds.total_load * SCHED_CAPACITY_SCALE) / 9247 sds.total_capacity; 9248 9249 /* 9250 * Don't pull any tasks if this group is already above the 9251 * domain average load. 9252 */ 9253 if (local->avg_load >= sds.avg_load) 9254 goto out_balanced; 9255 9256 /* 9257 * If the busiest group is more loaded, use imbalance_pct to be 9258 * conservative. 9259 */ 9260 if (100 * busiest->avg_load <= 9261 env->sd->imbalance_pct * local->avg_load) 9262 goto out_balanced; 9263 } 9264 9265 /* Try to move all excess tasks to child's sibling domain */ 9266 if (sds.prefer_sibling && local->group_type == group_has_spare && 9267 busiest->sum_nr_running > local->sum_nr_running + 1) 9268 goto force_balance; 9269 9270 if (busiest->group_type != group_overloaded) { 9271 if (env->idle == CPU_NOT_IDLE) 9272 /* 9273 * If the busiest group is not overloaded (and as a 9274 * result the local one too) but this CPU is already 9275 * busy, let another idle CPU try to pull task. 9276 */ 9277 goto out_balanced; 9278 9279 if (busiest->group_weight > 1 && 9280 local->idle_cpus <= (busiest->idle_cpus + 1)) 9281 /* 9282 * If the busiest group is not overloaded 9283 * and there is no imbalance between this and busiest 9284 * group wrt idle CPUs, it is balanced. The imbalance 9285 * becomes significant if the diff is greater than 1 9286 * otherwise we might end up to just move the imbalance 9287 * on another group. Of course this applies only if 9288 * there is more than 1 CPU per group. 9289 */ 9290 goto out_balanced; 9291 9292 if (busiest->sum_h_nr_running == 1) 9293 /* 9294 * busiest doesn't have any tasks waiting to run 9295 */ 9296 goto out_balanced; 9297 } 9298 9299 force_balance: 9300 /* Looks like there is an imbalance. Compute it */ 9301 calculate_imbalance(env, &sds); 9302 return env->imbalance ? sds.busiest : NULL; 9303 9304 out_balanced: 9305 env->imbalance = 0; 9306 return NULL; 9307 } 9308 9309 /* 9310 * find_busiest_queue - find the busiest runqueue among the CPUs in the group. 9311 */ 9312 static struct rq *find_busiest_queue(struct lb_env *env, 9313 struct sched_group *group) 9314 { 9315 struct rq *busiest = NULL, *rq; 9316 unsigned long busiest_util = 0, busiest_load = 0, busiest_capacity = 1; 9317 unsigned int busiest_nr = 0; 9318 int i; 9319 9320 for_each_cpu_and(i, sched_group_span(group), env->cpus) { 9321 unsigned long capacity, load, util; 9322 unsigned int nr_running; 9323 enum fbq_type rt; 9324 9325 rq = cpu_rq(i); 9326 rt = fbq_classify_rq(rq); 9327 9328 /* 9329 * We classify groups/runqueues into three groups: 9330 * - regular: there are !numa tasks 9331 * - remote: there are numa tasks that run on the 'wrong' node 9332 * - all: there is no distinction 9333 * 9334 * In order to avoid migrating ideally placed numa tasks, 9335 * ignore those when there's better options. 9336 * 9337 * If we ignore the actual busiest queue to migrate another 9338 * task, the next balance pass can still reduce the busiest 9339 * queue by moving tasks around inside the node. 9340 * 9341 * If we cannot move enough load due to this classification 9342 * the next pass will adjust the group classification and 9343 * allow migration of more tasks. 9344 * 9345 * Both cases only affect the total convergence complexity. 9346 */ 9347 if (rt > env->fbq_type) 9348 continue; 9349 9350 capacity = capacity_of(i); 9351 nr_running = rq->cfs.h_nr_running; 9352 9353 /* 9354 * For ASYM_CPUCAPACITY domains, don't pick a CPU that could 9355 * eventually lead to active_balancing high->low capacity. 9356 * Higher per-CPU capacity is considered better than balancing 9357 * average load. 9358 */ 9359 if (env->sd->flags & SD_ASYM_CPUCAPACITY && 9360 capacity_of(env->dst_cpu) < capacity && 9361 nr_running == 1) 9362 continue; 9363 9364 switch (env->migration_type) { 9365 case migrate_load: 9366 /* 9367 * When comparing with load imbalance, use cpu_load() 9368 * which is not scaled with the CPU capacity. 9369 */ 9370 load = cpu_load(rq); 9371 9372 if (nr_running == 1 && load > env->imbalance && 9373 !check_cpu_capacity(rq, env->sd)) 9374 break; 9375 9376 /* 9377 * For the load comparisons with the other CPUs, 9378 * consider the cpu_load() scaled with the CPU 9379 * capacity, so that the load can be moved away 9380 * from the CPU that is potentially running at a 9381 * lower capacity. 9382 * 9383 * Thus we're looking for max(load_i / capacity_i), 9384 * crosswise multiplication to rid ourselves of the 9385 * division works out to: 9386 * load_i * capacity_j > load_j * capacity_i; 9387 * where j is our previous maximum. 9388 */ 9389 if (load * busiest_capacity > busiest_load * capacity) { 9390 busiest_load = load; 9391 busiest_capacity = capacity; 9392 busiest = rq; 9393 } 9394 break; 9395 9396 case migrate_util: 9397 util = cpu_util(cpu_of(rq)); 9398 9399 /* 9400 * Don't try to pull utilization from a CPU with one 9401 * running task. Whatever its utilization, we will fail 9402 * detach the task. 9403 */ 9404 if (nr_running <= 1) 9405 continue; 9406 9407 if (busiest_util < util) { 9408 busiest_util = util; 9409 busiest = rq; 9410 } 9411 break; 9412 9413 case migrate_task: 9414 if (busiest_nr < nr_running) { 9415 busiest_nr = nr_running; 9416 busiest = rq; 9417 } 9418 break; 9419 9420 case migrate_misfit: 9421 /* 9422 * For ASYM_CPUCAPACITY domains with misfit tasks we 9423 * simply seek the "biggest" misfit task. 9424 */ 9425 if (rq->misfit_task_load > busiest_load) { 9426 busiest_load = rq->misfit_task_load; 9427 busiest = rq; 9428 } 9429 9430 break; 9431 9432 } 9433 } 9434 9435 return busiest; 9436 } 9437 9438 /* 9439 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but 9440 * so long as it is large enough. 9441 */ 9442 #define MAX_PINNED_INTERVAL 512 9443 9444 static inline bool 9445 asym_active_balance(struct lb_env *env) 9446 { 9447 /* 9448 * ASYM_PACKING needs to force migrate tasks from busy but 9449 * lower priority CPUs in order to pack all tasks in the 9450 * highest priority CPUs. 9451 */ 9452 return env->idle != CPU_NOT_IDLE && (env->sd->flags & SD_ASYM_PACKING) && 9453 sched_asym_prefer(env->dst_cpu, env->src_cpu); 9454 } 9455 9456 static inline bool 9457 voluntary_active_balance(struct lb_env *env) 9458 { 9459 struct sched_domain *sd = env->sd; 9460 9461 if (asym_active_balance(env)) 9462 return 1; 9463 9464 /* 9465 * The dst_cpu is idle and the src_cpu CPU has only 1 CFS task. 9466 * It's worth migrating the task if the src_cpu's capacity is reduced 9467 * because of other sched_class or IRQs if more capacity stays 9468 * available on dst_cpu. 9469 */ 9470 if ((env->idle != CPU_NOT_IDLE) && 9471 (env->src_rq->cfs.h_nr_running == 1)) { 9472 if ((check_cpu_capacity(env->src_rq, sd)) && 9473 (capacity_of(env->src_cpu)*sd->imbalance_pct < capacity_of(env->dst_cpu)*100)) 9474 return 1; 9475 } 9476 9477 if (env->migration_type == migrate_misfit) 9478 return 1; 9479 9480 return 0; 9481 } 9482 9483 static int need_active_balance(struct lb_env *env) 9484 { 9485 struct sched_domain *sd = env->sd; 9486 9487 if (voluntary_active_balance(env)) 9488 return 1; 9489 9490 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2); 9491 } 9492 9493 static int active_load_balance_cpu_stop(void *data); 9494 9495 static int should_we_balance(struct lb_env *env) 9496 { 9497 struct sched_group *sg = env->sd->groups; 9498 int cpu; 9499 9500 /* 9501 * Ensure the balancing environment is consistent; can happen 9502 * when the softirq triggers 'during' hotplug. 9503 */ 9504 if (!cpumask_test_cpu(env->dst_cpu, env->cpus)) 9505 return 0; 9506 9507 /* 9508 * In the newly idle case, we will allow all the CPUs 9509 * to do the newly idle load balance. 9510 */ 9511 if (env->idle == CPU_NEWLY_IDLE) 9512 return 1; 9513 9514 /* Try to find first idle CPU */ 9515 for_each_cpu_and(cpu, group_balance_mask(sg), env->cpus) { 9516 if (!idle_cpu(cpu)) 9517 continue; 9518 9519 /* Are we the first idle CPU? */ 9520 return cpu == env->dst_cpu; 9521 } 9522 9523 /* Are we the first CPU of this group ? */ 9524 return group_balance_cpu(sg) == env->dst_cpu; 9525 } 9526 9527 /* 9528 * Check this_cpu to ensure it is balanced within domain. Attempt to move 9529 * tasks if there is an imbalance. 9530 */ 9531 static int load_balance(int this_cpu, struct rq *this_rq, 9532 struct sched_domain *sd, enum cpu_idle_type idle, 9533 int *continue_balancing) 9534 { 9535 int ld_moved, cur_ld_moved, active_balance = 0; 9536 struct sched_domain *sd_parent = sd->parent; 9537 struct sched_group *group; 9538 struct rq *busiest; 9539 struct rq_flags rf; 9540 struct cpumask *cpus = this_cpu_cpumask_var_ptr(load_balance_mask); 9541 9542 struct lb_env env = { 9543 .sd = sd, 9544 .dst_cpu = this_cpu, 9545 .dst_rq = this_rq, 9546 .dst_grpmask = sched_group_span(sd->groups), 9547 .idle = idle, 9548 .loop_break = sched_nr_migrate_break, 9549 .cpus = cpus, 9550 .fbq_type = all, 9551 .tasks = LIST_HEAD_INIT(env.tasks), 9552 }; 9553 9554 cpumask_and(cpus, sched_domain_span(sd), cpu_active_mask); 9555 9556 schedstat_inc(sd->lb_count[idle]); 9557 9558 redo: 9559 if (!should_we_balance(&env)) { 9560 *continue_balancing = 0; 9561 goto out_balanced; 9562 } 9563 9564 group = find_busiest_group(&env); 9565 if (!group) { 9566 schedstat_inc(sd->lb_nobusyg[idle]); 9567 goto out_balanced; 9568 } 9569 9570 busiest = find_busiest_queue(&env, group); 9571 if (!busiest) { 9572 schedstat_inc(sd->lb_nobusyq[idle]); 9573 goto out_balanced; 9574 } 9575 9576 BUG_ON(busiest == env.dst_rq); 9577 9578 schedstat_add(sd->lb_imbalance[idle], env.imbalance); 9579 9580 env.src_cpu = busiest->cpu; 9581 env.src_rq = busiest; 9582 9583 ld_moved = 0; 9584 if (busiest->nr_running > 1) { 9585 /* 9586 * Attempt to move tasks. If find_busiest_group has found 9587 * an imbalance but busiest->nr_running <= 1, the group is 9588 * still unbalanced. ld_moved simply stays zero, so it is 9589 * correctly treated as an imbalance. 9590 */ 9591 env.flags |= LBF_ALL_PINNED; 9592 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running); 9593 9594 more_balance: 9595 rq_lock_irqsave(busiest, &rf); 9596 update_rq_clock(busiest); 9597 9598 /* 9599 * cur_ld_moved - load moved in current iteration 9600 * ld_moved - cumulative load moved across iterations 9601 */ 9602 cur_ld_moved = detach_tasks(&env); 9603 9604 /* 9605 * We've detached some tasks from busiest_rq. Every 9606 * task is masked "TASK_ON_RQ_MIGRATING", so we can safely 9607 * unlock busiest->lock, and we are able to be sure 9608 * that nobody can manipulate the tasks in parallel. 9609 * See task_rq_lock() family for the details. 9610 */ 9611 9612 rq_unlock(busiest, &rf); 9613 9614 if (cur_ld_moved) { 9615 attach_tasks(&env); 9616 ld_moved += cur_ld_moved; 9617 } 9618 9619 local_irq_restore(rf.flags); 9620 9621 if (env.flags & LBF_NEED_BREAK) { 9622 env.flags &= ~LBF_NEED_BREAK; 9623 goto more_balance; 9624 } 9625 9626 /* 9627 * Revisit (affine) tasks on src_cpu that couldn't be moved to 9628 * us and move them to an alternate dst_cpu in our sched_group 9629 * where they can run. The upper limit on how many times we 9630 * iterate on same src_cpu is dependent on number of CPUs in our 9631 * sched_group. 9632 * 9633 * This changes load balance semantics a bit on who can move 9634 * load to a given_cpu. In addition to the given_cpu itself 9635 * (or a ilb_cpu acting on its behalf where given_cpu is 9636 * nohz-idle), we now have balance_cpu in a position to move 9637 * load to given_cpu. In rare situations, this may cause 9638 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding 9639 * _independently_ and at _same_ time to move some load to 9640 * given_cpu) causing exceess load to be moved to given_cpu. 9641 * This however should not happen so much in practice and 9642 * moreover subsequent load balance cycles should correct the 9643 * excess load moved. 9644 */ 9645 if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) { 9646 9647 /* Prevent to re-select dst_cpu via env's CPUs */ 9648 __cpumask_clear_cpu(env.dst_cpu, env.cpus); 9649 9650 env.dst_rq = cpu_rq(env.new_dst_cpu); 9651 env.dst_cpu = env.new_dst_cpu; 9652 env.flags &= ~LBF_DST_PINNED; 9653 env.loop = 0; 9654 env.loop_break = sched_nr_migrate_break; 9655 9656 /* 9657 * Go back to "more_balance" rather than "redo" since we 9658 * need to continue with same src_cpu. 9659 */ 9660 goto more_balance; 9661 } 9662 9663 /* 9664 * We failed to reach balance because of affinity. 9665 */ 9666 if (sd_parent) { 9667 int *group_imbalance = &sd_parent->groups->sgc->imbalance; 9668 9669 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0) 9670 *group_imbalance = 1; 9671 } 9672 9673 /* All tasks on this runqueue were pinned by CPU affinity */ 9674 if (unlikely(env.flags & LBF_ALL_PINNED)) { 9675 __cpumask_clear_cpu(cpu_of(busiest), cpus); 9676 /* 9677 * Attempting to continue load balancing at the current 9678 * sched_domain level only makes sense if there are 9679 * active CPUs remaining as possible busiest CPUs to 9680 * pull load from which are not contained within the 9681 * destination group that is receiving any migrated 9682 * load. 9683 */ 9684 if (!cpumask_subset(cpus, env.dst_grpmask)) { 9685 env.loop = 0; 9686 env.loop_break = sched_nr_migrate_break; 9687 goto redo; 9688 } 9689 goto out_all_pinned; 9690 } 9691 } 9692 9693 if (!ld_moved) { 9694 schedstat_inc(sd->lb_failed[idle]); 9695 /* 9696 * Increment the failure counter only on periodic balance. 9697 * We do not want newidle balance, which can be very 9698 * frequent, pollute the failure counter causing 9699 * excessive cache_hot migrations and active balances. 9700 */ 9701 if (idle != CPU_NEWLY_IDLE) 9702 sd->nr_balance_failed++; 9703 9704 if (need_active_balance(&env)) { 9705 unsigned long flags; 9706 9707 raw_spin_lock_irqsave(&busiest->lock, flags); 9708 9709 /* 9710 * Don't kick the active_load_balance_cpu_stop, 9711 * if the curr task on busiest CPU can't be 9712 * moved to this_cpu: 9713 */ 9714 if (!cpumask_test_cpu(this_cpu, busiest->curr->cpus_ptr)) { 9715 raw_spin_unlock_irqrestore(&busiest->lock, 9716 flags); 9717 env.flags |= LBF_ALL_PINNED; 9718 goto out_one_pinned; 9719 } 9720 9721 /* 9722 * ->active_balance synchronizes accesses to 9723 * ->active_balance_work. Once set, it's cleared 9724 * only after active load balance is finished. 9725 */ 9726 if (!busiest->active_balance) { 9727 busiest->active_balance = 1; 9728 busiest->push_cpu = this_cpu; 9729 active_balance = 1; 9730 } 9731 raw_spin_unlock_irqrestore(&busiest->lock, flags); 9732 9733 if (active_balance) { 9734 stop_one_cpu_nowait(cpu_of(busiest), 9735 active_load_balance_cpu_stop, busiest, 9736 &busiest->active_balance_work); 9737 } 9738 9739 /* We've kicked active balancing, force task migration. */ 9740 sd->nr_balance_failed = sd->cache_nice_tries+1; 9741 } 9742 } else 9743 sd->nr_balance_failed = 0; 9744 9745 if (likely(!active_balance) || voluntary_active_balance(&env)) { 9746 /* We were unbalanced, so reset the balancing interval */ 9747 sd->balance_interval = sd->min_interval; 9748 } else { 9749 /* 9750 * If we've begun active balancing, start to back off. This 9751 * case may not be covered by the all_pinned logic if there 9752 * is only 1 task on the busy runqueue (because we don't call 9753 * detach_tasks). 9754 */ 9755 if (sd->balance_interval < sd->max_interval) 9756 sd->balance_interval *= 2; 9757 } 9758 9759 goto out; 9760 9761 out_balanced: 9762 /* 9763 * We reach balance although we may have faced some affinity 9764 * constraints. Clear the imbalance flag only if other tasks got 9765 * a chance to move and fix the imbalance. 9766 */ 9767 if (sd_parent && !(env.flags & LBF_ALL_PINNED)) { 9768 int *group_imbalance = &sd_parent->groups->sgc->imbalance; 9769 9770 if (*group_imbalance) 9771 *group_imbalance = 0; 9772 } 9773 9774 out_all_pinned: 9775 /* 9776 * We reach balance because all tasks are pinned at this level so 9777 * we can't migrate them. Let the imbalance flag set so parent level 9778 * can try to migrate them. 9779 */ 9780 schedstat_inc(sd->lb_balanced[idle]); 9781 9782 sd->nr_balance_failed = 0; 9783 9784 out_one_pinned: 9785 ld_moved = 0; 9786 9787 /* 9788 * newidle_balance() disregards balance intervals, so we could 9789 * repeatedly reach this code, which would lead to balance_interval 9790 * skyrocketting in a short amount of time. Skip the balance_interval 9791 * increase logic to avoid that. 9792 */ 9793 if (env.idle == CPU_NEWLY_IDLE) 9794 goto out; 9795 9796 /* tune up the balancing interval */ 9797 if ((env.flags & LBF_ALL_PINNED && 9798 sd->balance_interval < MAX_PINNED_INTERVAL) || 9799 sd->balance_interval < sd->max_interval) 9800 sd->balance_interval *= 2; 9801 out: 9802 return ld_moved; 9803 } 9804 9805 static inline unsigned long 9806 get_sd_balance_interval(struct sched_domain *sd, int cpu_busy) 9807 { 9808 unsigned long interval = sd->balance_interval; 9809 9810 if (cpu_busy) 9811 interval *= sd->busy_factor; 9812 9813 /* scale ms to jiffies */ 9814 interval = msecs_to_jiffies(interval); 9815 9816 /* 9817 * Reduce likelihood of busy balancing at higher domains racing with 9818 * balancing at lower domains by preventing their balancing periods 9819 * from being multiples of each other. 9820 */ 9821 if (cpu_busy) 9822 interval -= 1; 9823 9824 interval = clamp(interval, 1UL, max_load_balance_interval); 9825 9826 return interval; 9827 } 9828 9829 static inline void 9830 update_next_balance(struct sched_domain *sd, unsigned long *next_balance) 9831 { 9832 unsigned long interval, next; 9833 9834 /* used by idle balance, so cpu_busy = 0 */ 9835 interval = get_sd_balance_interval(sd, 0); 9836 next = sd->last_balance + interval; 9837 9838 if (time_after(*next_balance, next)) 9839 *next_balance = next; 9840 } 9841 9842 /* 9843 * active_load_balance_cpu_stop is run by the CPU stopper. It pushes 9844 * running tasks off the busiest CPU onto idle CPUs. It requires at 9845 * least 1 task to be running on each physical CPU where possible, and 9846 * avoids physical / logical imbalances. 9847 */ 9848 static int active_load_balance_cpu_stop(void *data) 9849 { 9850 struct rq *busiest_rq = data; 9851 int busiest_cpu = cpu_of(busiest_rq); 9852 int target_cpu = busiest_rq->push_cpu; 9853 struct rq *target_rq = cpu_rq(target_cpu); 9854 struct sched_domain *sd; 9855 struct task_struct *p = NULL; 9856 struct rq_flags rf; 9857 9858 rq_lock_irq(busiest_rq, &rf); 9859 /* 9860 * Between queueing the stop-work and running it is a hole in which 9861 * CPUs can become inactive. We should not move tasks from or to 9862 * inactive CPUs. 9863 */ 9864 if (!cpu_active(busiest_cpu) || !cpu_active(target_cpu)) 9865 goto out_unlock; 9866 9867 /* Make sure the requested CPU hasn't gone down in the meantime: */ 9868 if (unlikely(busiest_cpu != smp_processor_id() || 9869 !busiest_rq->active_balance)) 9870 goto out_unlock; 9871 9872 /* Is there any task to move? */ 9873 if (busiest_rq->nr_running <= 1) 9874 goto out_unlock; 9875 9876 /* 9877 * This condition is "impossible", if it occurs 9878 * we need to fix it. Originally reported by 9879 * Bjorn Helgaas on a 128-CPU setup. 9880 */ 9881 BUG_ON(busiest_rq == target_rq); 9882 9883 /* Search for an sd spanning us and the target CPU. */ 9884 rcu_read_lock(); 9885 for_each_domain(target_cpu, sd) { 9886 if (cpumask_test_cpu(busiest_cpu, sched_domain_span(sd))) 9887 break; 9888 } 9889 9890 if (likely(sd)) { 9891 struct lb_env env = { 9892 .sd = sd, 9893 .dst_cpu = target_cpu, 9894 .dst_rq = target_rq, 9895 .src_cpu = busiest_rq->cpu, 9896 .src_rq = busiest_rq, 9897 .idle = CPU_IDLE, 9898 /* 9899 * can_migrate_task() doesn't need to compute new_dst_cpu 9900 * for active balancing. Since we have CPU_IDLE, but no 9901 * @dst_grpmask we need to make that test go away with lying 9902 * about DST_PINNED. 9903 */ 9904 .flags = LBF_DST_PINNED, 9905 }; 9906 9907 schedstat_inc(sd->alb_count); 9908 update_rq_clock(busiest_rq); 9909 9910 p = detach_one_task(&env); 9911 if (p) { 9912 schedstat_inc(sd->alb_pushed); 9913 /* Active balancing done, reset the failure counter. */ 9914 sd->nr_balance_failed = 0; 9915 } else { 9916 schedstat_inc(sd->alb_failed); 9917 } 9918 } 9919 rcu_read_unlock(); 9920 out_unlock: 9921 busiest_rq->active_balance = 0; 9922 rq_unlock(busiest_rq, &rf); 9923 9924 if (p) 9925 attach_one_task(target_rq, p); 9926 9927 local_irq_enable(); 9928 9929 return 0; 9930 } 9931 9932 static DEFINE_SPINLOCK(balancing); 9933 9934 /* 9935 * Scale the max load_balance interval with the number of CPUs in the system. 9936 * This trades load-balance latency on larger machines for less cross talk. 9937 */ 9938 void update_max_interval(void) 9939 { 9940 max_load_balance_interval = HZ*num_online_cpus()/10; 9941 } 9942 9943 /* 9944 * It checks each scheduling domain to see if it is due to be balanced, 9945 * and initiates a balancing operation if so. 9946 * 9947 * Balancing parameters are set up in init_sched_domains. 9948 */ 9949 static void rebalance_domains(struct rq *rq, enum cpu_idle_type idle) 9950 { 9951 int continue_balancing = 1; 9952 int cpu = rq->cpu; 9953 int busy = idle != CPU_IDLE && !sched_idle_cpu(cpu); 9954 unsigned long interval; 9955 struct sched_domain *sd; 9956 /* Earliest time when we have to do rebalance again */ 9957 unsigned long next_balance = jiffies + 60*HZ; 9958 int update_next_balance = 0; 9959 int need_serialize, need_decay = 0; 9960 u64 max_cost = 0; 9961 9962 rcu_read_lock(); 9963 for_each_domain(cpu, sd) { 9964 /* 9965 * Decay the newidle max times here because this is a regular 9966 * visit to all the domains. Decay ~1% per second. 9967 */ 9968 if (time_after(jiffies, sd->next_decay_max_lb_cost)) { 9969 sd->max_newidle_lb_cost = 9970 (sd->max_newidle_lb_cost * 253) / 256; 9971 sd->next_decay_max_lb_cost = jiffies + HZ; 9972 need_decay = 1; 9973 } 9974 max_cost += sd->max_newidle_lb_cost; 9975 9976 /* 9977 * Stop the load balance at this level. There is another 9978 * CPU in our sched group which is doing load balancing more 9979 * actively. 9980 */ 9981 if (!continue_balancing) { 9982 if (need_decay) 9983 continue; 9984 break; 9985 } 9986 9987 interval = get_sd_balance_interval(sd, busy); 9988 9989 need_serialize = sd->flags & SD_SERIALIZE; 9990 if (need_serialize) { 9991 if (!spin_trylock(&balancing)) 9992 goto out; 9993 } 9994 9995 if (time_after_eq(jiffies, sd->last_balance + interval)) { 9996 if (load_balance(cpu, rq, sd, idle, &continue_balancing)) { 9997 /* 9998 * The LBF_DST_PINNED logic could have changed 9999 * env->dst_cpu, so we can't know our idle 10000 * state even if we migrated tasks. Update it. 10001 */ 10002 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE; 10003 busy = idle != CPU_IDLE && !sched_idle_cpu(cpu); 10004 } 10005 sd->last_balance = jiffies; 10006 interval = get_sd_balance_interval(sd, busy); 10007 } 10008 if (need_serialize) 10009 spin_unlock(&balancing); 10010 out: 10011 if (time_after(next_balance, sd->last_balance + interval)) { 10012 next_balance = sd->last_balance + interval; 10013 update_next_balance = 1; 10014 } 10015 } 10016 if (need_decay) { 10017 /* 10018 * Ensure the rq-wide value also decays but keep it at a 10019 * reasonable floor to avoid funnies with rq->avg_idle. 10020 */ 10021 rq->max_idle_balance_cost = 10022 max((u64)sysctl_sched_migration_cost, max_cost); 10023 } 10024 rcu_read_unlock(); 10025 10026 /* 10027 * next_balance will be updated only when there is a need. 10028 * When the cpu is attached to null domain for ex, it will not be 10029 * updated. 10030 */ 10031 if (likely(update_next_balance)) { 10032 rq->next_balance = next_balance; 10033 10034 #ifdef CONFIG_NO_HZ_COMMON 10035 /* 10036 * If this CPU has been elected to perform the nohz idle 10037 * balance. Other idle CPUs have already rebalanced with 10038 * nohz_idle_balance() and nohz.next_balance has been 10039 * updated accordingly. This CPU is now running the idle load 10040 * balance for itself and we need to update the 10041 * nohz.next_balance accordingly. 10042 */ 10043 if ((idle == CPU_IDLE) && time_after(nohz.next_balance, rq->next_balance)) 10044 nohz.next_balance = rq->next_balance; 10045 #endif 10046 } 10047 } 10048 10049 static inline int on_null_domain(struct rq *rq) 10050 { 10051 return unlikely(!rcu_dereference_sched(rq->sd)); 10052 } 10053 10054 #ifdef CONFIG_NO_HZ_COMMON 10055 /* 10056 * idle load balancing details 10057 * - When one of the busy CPUs notice that there may be an idle rebalancing 10058 * needed, they will kick the idle load balancer, which then does idle 10059 * load balancing for all the idle CPUs. 10060 * - HK_FLAG_MISC CPUs are used for this task, because HK_FLAG_SCHED not set 10061 * anywhere yet. 10062 */ 10063 10064 static inline int find_new_ilb(void) 10065 { 10066 int ilb; 10067 10068 for_each_cpu_and(ilb, nohz.idle_cpus_mask, 10069 housekeeping_cpumask(HK_FLAG_MISC)) { 10070 if (idle_cpu(ilb)) 10071 return ilb; 10072 } 10073 10074 return nr_cpu_ids; 10075 } 10076 10077 /* 10078 * Kick a CPU to do the nohz balancing, if it is time for it. We pick any 10079 * idle CPU in the HK_FLAG_MISC housekeeping set (if there is one). 10080 */ 10081 static void kick_ilb(unsigned int flags) 10082 { 10083 int ilb_cpu; 10084 10085 /* 10086 * Increase nohz.next_balance only when if full ilb is triggered but 10087 * not if we only update stats. 10088 */ 10089 if (flags & NOHZ_BALANCE_KICK) 10090 nohz.next_balance = jiffies+1; 10091 10092 ilb_cpu = find_new_ilb(); 10093 10094 if (ilb_cpu >= nr_cpu_ids) 10095 return; 10096 10097 /* 10098 * Access to rq::nohz_csd is serialized by NOHZ_KICK_MASK; he who sets 10099 * the first flag owns it; cleared by nohz_csd_func(). 10100 */ 10101 flags = atomic_fetch_or(flags, nohz_flags(ilb_cpu)); 10102 if (flags & NOHZ_KICK_MASK) 10103 return; 10104 10105 /* 10106 * This way we generate an IPI on the target CPU which 10107 * is idle. And the softirq performing nohz idle load balance 10108 * will be run before returning from the IPI. 10109 */ 10110 smp_call_function_single_async(ilb_cpu, &cpu_rq(ilb_cpu)->nohz_csd); 10111 } 10112 10113 /* 10114 * Current decision point for kicking the idle load balancer in the presence 10115 * of idle CPUs in the system. 10116 */ 10117 static void nohz_balancer_kick(struct rq *rq) 10118 { 10119 unsigned long now = jiffies; 10120 struct sched_domain_shared *sds; 10121 struct sched_domain *sd; 10122 int nr_busy, i, cpu = rq->cpu; 10123 unsigned int flags = 0; 10124 10125 if (unlikely(rq->idle_balance)) 10126 return; 10127 10128 /* 10129 * We may be recently in ticked or tickless idle mode. At the first 10130 * busy tick after returning from idle, we will update the busy stats. 10131 */ 10132 nohz_balance_exit_idle(rq); 10133 10134 /* 10135 * None are in tickless mode and hence no need for NOHZ idle load 10136 * balancing. 10137 */ 10138 if (likely(!atomic_read(&nohz.nr_cpus))) 10139 return; 10140 10141 if (READ_ONCE(nohz.has_blocked) && 10142 time_after(now, READ_ONCE(nohz.next_blocked))) 10143 flags = NOHZ_STATS_KICK; 10144 10145 if (time_before(now, nohz.next_balance)) 10146 goto out; 10147 10148 if (rq->nr_running >= 2) { 10149 flags = NOHZ_KICK_MASK; 10150 goto out; 10151 } 10152 10153 rcu_read_lock(); 10154 10155 sd = rcu_dereference(rq->sd); 10156 if (sd) { 10157 /* 10158 * If there's a CFS task and the current CPU has reduced 10159 * capacity; kick the ILB to see if there's a better CPU to run 10160 * on. 10161 */ 10162 if (rq->cfs.h_nr_running >= 1 && check_cpu_capacity(rq, sd)) { 10163 flags = NOHZ_KICK_MASK; 10164 goto unlock; 10165 } 10166 } 10167 10168 sd = rcu_dereference(per_cpu(sd_asym_packing, cpu)); 10169 if (sd) { 10170 /* 10171 * When ASYM_PACKING; see if there's a more preferred CPU 10172 * currently idle; in which case, kick the ILB to move tasks 10173 * around. 10174 */ 10175 for_each_cpu_and(i, sched_domain_span(sd), nohz.idle_cpus_mask) { 10176 if (sched_asym_prefer(i, cpu)) { 10177 flags = NOHZ_KICK_MASK; 10178 goto unlock; 10179 } 10180 } 10181 } 10182 10183 sd = rcu_dereference(per_cpu(sd_asym_cpucapacity, cpu)); 10184 if (sd) { 10185 /* 10186 * When ASYM_CPUCAPACITY; see if there's a higher capacity CPU 10187 * to run the misfit task on. 10188 */ 10189 if (check_misfit_status(rq, sd)) { 10190 flags = NOHZ_KICK_MASK; 10191 goto unlock; 10192 } 10193 10194 /* 10195 * For asymmetric systems, we do not want to nicely balance 10196 * cache use, instead we want to embrace asymmetry and only 10197 * ensure tasks have enough CPU capacity. 10198 * 10199 * Skip the LLC logic because it's not relevant in that case. 10200 */ 10201 goto unlock; 10202 } 10203 10204 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu)); 10205 if (sds) { 10206 /* 10207 * If there is an imbalance between LLC domains (IOW we could 10208 * increase the overall cache use), we need some less-loaded LLC 10209 * domain to pull some load. Likewise, we may need to spread 10210 * load within the current LLC domain (e.g. packed SMT cores but 10211 * other CPUs are idle). We can't really know from here how busy 10212 * the others are - so just get a nohz balance going if it looks 10213 * like this LLC domain has tasks we could move. 10214 */ 10215 nr_busy = atomic_read(&sds->nr_busy_cpus); 10216 if (nr_busy > 1) { 10217 flags = NOHZ_KICK_MASK; 10218 goto unlock; 10219 } 10220 } 10221 unlock: 10222 rcu_read_unlock(); 10223 out: 10224 if (flags) 10225 kick_ilb(flags); 10226 } 10227 10228 static void set_cpu_sd_state_busy(int cpu) 10229 { 10230 struct sched_domain *sd; 10231 10232 rcu_read_lock(); 10233 sd = rcu_dereference(per_cpu(sd_llc, cpu)); 10234 10235 if (!sd || !sd->nohz_idle) 10236 goto unlock; 10237 sd->nohz_idle = 0; 10238 10239 atomic_inc(&sd->shared->nr_busy_cpus); 10240 unlock: 10241 rcu_read_unlock(); 10242 } 10243 10244 void nohz_balance_exit_idle(struct rq *rq) 10245 { 10246 SCHED_WARN_ON(rq != this_rq()); 10247 10248 if (likely(!rq->nohz_tick_stopped)) 10249 return; 10250 10251 rq->nohz_tick_stopped = 0; 10252 cpumask_clear_cpu(rq->cpu, nohz.idle_cpus_mask); 10253 atomic_dec(&nohz.nr_cpus); 10254 10255 set_cpu_sd_state_busy(rq->cpu); 10256 } 10257 10258 static void set_cpu_sd_state_idle(int cpu) 10259 { 10260 struct sched_domain *sd; 10261 10262 rcu_read_lock(); 10263 sd = rcu_dereference(per_cpu(sd_llc, cpu)); 10264 10265 if (!sd || sd->nohz_idle) 10266 goto unlock; 10267 sd->nohz_idle = 1; 10268 10269 atomic_dec(&sd->shared->nr_busy_cpus); 10270 unlock: 10271 rcu_read_unlock(); 10272 } 10273 10274 /* 10275 * This routine will record that the CPU is going idle with tick stopped. 10276 * This info will be used in performing idle load balancing in the future. 10277 */ 10278 void nohz_balance_enter_idle(int cpu) 10279 { 10280 struct rq *rq = cpu_rq(cpu); 10281 10282 SCHED_WARN_ON(cpu != smp_processor_id()); 10283 10284 /* If this CPU is going down, then nothing needs to be done: */ 10285 if (!cpu_active(cpu)) 10286 return; 10287 10288 /* Spare idle load balancing on CPUs that don't want to be disturbed: */ 10289 if (!housekeeping_cpu(cpu, HK_FLAG_SCHED)) 10290 return; 10291 10292 /* 10293 * Can be set safely without rq->lock held 10294 * If a clear happens, it will have evaluated last additions because 10295 * rq->lock is held during the check and the clear 10296 */ 10297 rq->has_blocked_load = 1; 10298 10299 /* 10300 * The tick is still stopped but load could have been added in the 10301 * meantime. We set the nohz.has_blocked flag to trig a check of the 10302 * *_avg. The CPU is already part of nohz.idle_cpus_mask so the clear 10303 * of nohz.has_blocked can only happen after checking the new load 10304 */ 10305 if (rq->nohz_tick_stopped) 10306 goto out; 10307 10308 /* If we're a completely isolated CPU, we don't play: */ 10309 if (on_null_domain(rq)) 10310 return; 10311 10312 rq->nohz_tick_stopped = 1; 10313 10314 cpumask_set_cpu(cpu, nohz.idle_cpus_mask); 10315 atomic_inc(&nohz.nr_cpus); 10316 10317 /* 10318 * Ensures that if nohz_idle_balance() fails to observe our 10319 * @idle_cpus_mask store, it must observe the @has_blocked 10320 * store. 10321 */ 10322 smp_mb__after_atomic(); 10323 10324 set_cpu_sd_state_idle(cpu); 10325 10326 out: 10327 /* 10328 * Each time a cpu enter idle, we assume that it has blocked load and 10329 * enable the periodic update of the load of idle cpus 10330 */ 10331 WRITE_ONCE(nohz.has_blocked, 1); 10332 } 10333 10334 /* 10335 * Internal function that runs load balance for all idle cpus. The load balance 10336 * can be a simple update of blocked load or a complete load balance with 10337 * tasks movement depending of flags. 10338 * The function returns false if the loop has stopped before running 10339 * through all idle CPUs. 10340 */ 10341 static bool _nohz_idle_balance(struct rq *this_rq, unsigned int flags, 10342 enum cpu_idle_type idle) 10343 { 10344 /* Earliest time when we have to do rebalance again */ 10345 unsigned long now = jiffies; 10346 unsigned long next_balance = now + 60*HZ; 10347 bool has_blocked_load = false; 10348 int update_next_balance = 0; 10349 int this_cpu = this_rq->cpu; 10350 int balance_cpu; 10351 int ret = false; 10352 struct rq *rq; 10353 10354 SCHED_WARN_ON((flags & NOHZ_KICK_MASK) == NOHZ_BALANCE_KICK); 10355 10356 /* 10357 * We assume there will be no idle load after this update and clear 10358 * the has_blocked flag. If a cpu enters idle in the mean time, it will 10359 * set the has_blocked flag and trig another update of idle load. 10360 * Because a cpu that becomes idle, is added to idle_cpus_mask before 10361 * setting the flag, we are sure to not clear the state and not 10362 * check the load of an idle cpu. 10363 */ 10364 WRITE_ONCE(nohz.has_blocked, 0); 10365 10366 /* 10367 * Ensures that if we miss the CPU, we must see the has_blocked 10368 * store from nohz_balance_enter_idle(). 10369 */ 10370 smp_mb(); 10371 10372 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) { 10373 if (balance_cpu == this_cpu || !idle_cpu(balance_cpu)) 10374 continue; 10375 10376 /* 10377 * If this CPU gets work to do, stop the load balancing 10378 * work being done for other CPUs. Next load 10379 * balancing owner will pick it up. 10380 */ 10381 if (need_resched()) { 10382 has_blocked_load = true; 10383 goto abort; 10384 } 10385 10386 rq = cpu_rq(balance_cpu); 10387 10388 has_blocked_load |= update_nohz_stats(rq, true); 10389 10390 /* 10391 * If time for next balance is due, 10392 * do the balance. 10393 */ 10394 if (time_after_eq(jiffies, rq->next_balance)) { 10395 struct rq_flags rf; 10396 10397 rq_lock_irqsave(rq, &rf); 10398 update_rq_clock(rq); 10399 rq_unlock_irqrestore(rq, &rf); 10400 10401 if (flags & NOHZ_BALANCE_KICK) 10402 rebalance_domains(rq, CPU_IDLE); 10403 } 10404 10405 if (time_after(next_balance, rq->next_balance)) { 10406 next_balance = rq->next_balance; 10407 update_next_balance = 1; 10408 } 10409 } 10410 10411 /* 10412 * next_balance will be updated only when there is a need. 10413 * When the CPU is attached to null domain for ex, it will not be 10414 * updated. 10415 */ 10416 if (likely(update_next_balance)) 10417 nohz.next_balance = next_balance; 10418 10419 /* Newly idle CPU doesn't need an update */ 10420 if (idle != CPU_NEWLY_IDLE) { 10421 update_blocked_averages(this_cpu); 10422 has_blocked_load |= this_rq->has_blocked_load; 10423 } 10424 10425 if (flags & NOHZ_BALANCE_KICK) 10426 rebalance_domains(this_rq, CPU_IDLE); 10427 10428 WRITE_ONCE(nohz.next_blocked, 10429 now + msecs_to_jiffies(LOAD_AVG_PERIOD)); 10430 10431 /* The full idle balance loop has been done */ 10432 ret = true; 10433 10434 abort: 10435 /* There is still blocked load, enable periodic update */ 10436 if (has_blocked_load) 10437 WRITE_ONCE(nohz.has_blocked, 1); 10438 10439 return ret; 10440 } 10441 10442 /* 10443 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the 10444 * rebalancing for all the cpus for whom scheduler ticks are stopped. 10445 */ 10446 static bool nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle) 10447 { 10448 unsigned int flags = this_rq->nohz_idle_balance; 10449 10450 if (!flags) 10451 return false; 10452 10453 this_rq->nohz_idle_balance = 0; 10454 10455 if (idle != CPU_IDLE) 10456 return false; 10457 10458 _nohz_idle_balance(this_rq, flags, idle); 10459 10460 return true; 10461 } 10462 10463 static void nohz_newidle_balance(struct rq *this_rq) 10464 { 10465 int this_cpu = this_rq->cpu; 10466 10467 /* 10468 * This CPU doesn't want to be disturbed by scheduler 10469 * housekeeping 10470 */ 10471 if (!housekeeping_cpu(this_cpu, HK_FLAG_SCHED)) 10472 return; 10473 10474 /* Will wake up very soon. No time for doing anything else*/ 10475 if (this_rq->avg_idle < sysctl_sched_migration_cost) 10476 return; 10477 10478 /* Don't need to update blocked load of idle CPUs*/ 10479 if (!READ_ONCE(nohz.has_blocked) || 10480 time_before(jiffies, READ_ONCE(nohz.next_blocked))) 10481 return; 10482 10483 raw_spin_unlock(&this_rq->lock); 10484 /* 10485 * This CPU is going to be idle and blocked load of idle CPUs 10486 * need to be updated. Run the ilb locally as it is a good 10487 * candidate for ilb instead of waking up another idle CPU. 10488 * Kick an normal ilb if we failed to do the update. 10489 */ 10490 if (!_nohz_idle_balance(this_rq, NOHZ_STATS_KICK, CPU_NEWLY_IDLE)) 10491 kick_ilb(NOHZ_STATS_KICK); 10492 raw_spin_lock(&this_rq->lock); 10493 } 10494 10495 #else /* !CONFIG_NO_HZ_COMMON */ 10496 static inline void nohz_balancer_kick(struct rq *rq) { } 10497 10498 static inline bool nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle) 10499 { 10500 return false; 10501 } 10502 10503 static inline void nohz_newidle_balance(struct rq *this_rq) { } 10504 #endif /* CONFIG_NO_HZ_COMMON */ 10505 10506 /* 10507 * idle_balance is called by schedule() if this_cpu is about to become 10508 * idle. Attempts to pull tasks from other CPUs. 10509 * 10510 * Returns: 10511 * < 0 - we released the lock and there are !fair tasks present 10512 * 0 - failed, no new tasks 10513 * > 0 - success, new (fair) tasks present 10514 */ 10515 static int newidle_balance(struct rq *this_rq, struct rq_flags *rf) 10516 { 10517 unsigned long next_balance = jiffies + HZ; 10518 int this_cpu = this_rq->cpu; 10519 struct sched_domain *sd; 10520 int pulled_task = 0; 10521 u64 curr_cost = 0; 10522 10523 update_misfit_status(NULL, this_rq); 10524 /* 10525 * We must set idle_stamp _before_ calling idle_balance(), such that we 10526 * measure the duration of idle_balance() as idle time. 10527 */ 10528 this_rq->idle_stamp = rq_clock(this_rq); 10529 10530 /* 10531 * Do not pull tasks towards !active CPUs... 10532 */ 10533 if (!cpu_active(this_cpu)) 10534 return 0; 10535 10536 /* 10537 * This is OK, because current is on_cpu, which avoids it being picked 10538 * for load-balance and preemption/IRQs are still disabled avoiding 10539 * further scheduler activity on it and we're being very careful to 10540 * re-start the picking loop. 10541 */ 10542 rq_unpin_lock(this_rq, rf); 10543 10544 if (this_rq->avg_idle < sysctl_sched_migration_cost || 10545 !READ_ONCE(this_rq->rd->overload)) { 10546 10547 rcu_read_lock(); 10548 sd = rcu_dereference_check_sched_domain(this_rq->sd); 10549 if (sd) 10550 update_next_balance(sd, &next_balance); 10551 rcu_read_unlock(); 10552 10553 nohz_newidle_balance(this_rq); 10554 10555 goto out; 10556 } 10557 10558 raw_spin_unlock(&this_rq->lock); 10559 10560 update_blocked_averages(this_cpu); 10561 rcu_read_lock(); 10562 for_each_domain(this_cpu, sd) { 10563 int continue_balancing = 1; 10564 u64 t0, domain_cost; 10565 10566 if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost) { 10567 update_next_balance(sd, &next_balance); 10568 break; 10569 } 10570 10571 if (sd->flags & SD_BALANCE_NEWIDLE) { 10572 t0 = sched_clock_cpu(this_cpu); 10573 10574 pulled_task = load_balance(this_cpu, this_rq, 10575 sd, CPU_NEWLY_IDLE, 10576 &continue_balancing); 10577 10578 domain_cost = sched_clock_cpu(this_cpu) - t0; 10579 if (domain_cost > sd->max_newidle_lb_cost) 10580 sd->max_newidle_lb_cost = domain_cost; 10581 10582 curr_cost += domain_cost; 10583 } 10584 10585 update_next_balance(sd, &next_balance); 10586 10587 /* 10588 * Stop searching for tasks to pull if there are 10589 * now runnable tasks on this rq. 10590 */ 10591 if (pulled_task || this_rq->nr_running > 0) 10592 break; 10593 } 10594 rcu_read_unlock(); 10595 10596 raw_spin_lock(&this_rq->lock); 10597 10598 if (curr_cost > this_rq->max_idle_balance_cost) 10599 this_rq->max_idle_balance_cost = curr_cost; 10600 10601 out: 10602 /* 10603 * While browsing the domains, we released the rq lock, a task could 10604 * have been enqueued in the meantime. Since we're not going idle, 10605 * pretend we pulled a task. 10606 */ 10607 if (this_rq->cfs.h_nr_running && !pulled_task) 10608 pulled_task = 1; 10609 10610 /* Move the next balance forward */ 10611 if (time_after(this_rq->next_balance, next_balance)) 10612 this_rq->next_balance = next_balance; 10613 10614 /* Is there a task of a high priority class? */ 10615 if (this_rq->nr_running != this_rq->cfs.h_nr_running) 10616 pulled_task = -1; 10617 10618 if (pulled_task) 10619 this_rq->idle_stamp = 0; 10620 10621 rq_repin_lock(this_rq, rf); 10622 10623 return pulled_task; 10624 } 10625 10626 /* 10627 * run_rebalance_domains is triggered when needed from the scheduler tick. 10628 * Also triggered for nohz idle balancing (with nohz_balancing_kick set). 10629 */ 10630 static __latent_entropy void run_rebalance_domains(struct softirq_action *h) 10631 { 10632 struct rq *this_rq = this_rq(); 10633 enum cpu_idle_type idle = this_rq->idle_balance ? 10634 CPU_IDLE : CPU_NOT_IDLE; 10635 10636 /* 10637 * If this CPU has a pending nohz_balance_kick, then do the 10638 * balancing on behalf of the other idle CPUs whose ticks are 10639 * stopped. Do nohz_idle_balance *before* rebalance_domains to 10640 * give the idle CPUs a chance to load balance. Else we may 10641 * load balance only within the local sched_domain hierarchy 10642 * and abort nohz_idle_balance altogether if we pull some load. 10643 */ 10644 if (nohz_idle_balance(this_rq, idle)) 10645 return; 10646 10647 /* normal load balance */ 10648 update_blocked_averages(this_rq->cpu); 10649 rebalance_domains(this_rq, idle); 10650 } 10651 10652 /* 10653 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing. 10654 */ 10655 void trigger_load_balance(struct rq *rq) 10656 { 10657 /* Don't need to rebalance while attached to NULL domain */ 10658 if (unlikely(on_null_domain(rq))) 10659 return; 10660 10661 if (time_after_eq(jiffies, rq->next_balance)) 10662 raise_softirq(SCHED_SOFTIRQ); 10663 10664 nohz_balancer_kick(rq); 10665 } 10666 10667 static void rq_online_fair(struct rq *rq) 10668 { 10669 update_sysctl(); 10670 10671 update_runtime_enabled(rq); 10672 } 10673 10674 static void rq_offline_fair(struct rq *rq) 10675 { 10676 update_sysctl(); 10677 10678 /* Ensure any throttled groups are reachable by pick_next_task */ 10679 unthrottle_offline_cfs_rqs(rq); 10680 } 10681 10682 #endif /* CONFIG_SMP */ 10683 10684 /* 10685 * scheduler tick hitting a task of our scheduling class. 10686 * 10687 * NOTE: This function can be called remotely by the tick offload that 10688 * goes along full dynticks. Therefore no local assumption can be made 10689 * and everything must be accessed through the @rq and @curr passed in 10690 * parameters. 10691 */ 10692 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued) 10693 { 10694 struct cfs_rq *cfs_rq; 10695 struct sched_entity *se = &curr->se; 10696 10697 for_each_sched_entity(se) { 10698 cfs_rq = cfs_rq_of(se); 10699 entity_tick(cfs_rq, se, queued); 10700 } 10701 10702 if (static_branch_unlikely(&sched_numa_balancing)) 10703 task_tick_numa(rq, curr); 10704 10705 update_misfit_status(curr, rq); 10706 update_overutilized_status(task_rq(curr)); 10707 } 10708 10709 /* 10710 * called on fork with the child task as argument from the parent's context 10711 * - child not yet on the tasklist 10712 * - preemption disabled 10713 */ 10714 static void task_fork_fair(struct task_struct *p) 10715 { 10716 struct cfs_rq *cfs_rq; 10717 struct sched_entity *se = &p->se, *curr; 10718 struct rq *rq = this_rq(); 10719 struct rq_flags rf; 10720 10721 rq_lock(rq, &rf); 10722 update_rq_clock(rq); 10723 10724 cfs_rq = task_cfs_rq(current); 10725 curr = cfs_rq->curr; 10726 if (curr) { 10727 update_curr(cfs_rq); 10728 se->vruntime = curr->vruntime; 10729 } 10730 place_entity(cfs_rq, se, 1); 10731 10732 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) { 10733 /* 10734 * Upon rescheduling, sched_class::put_prev_task() will place 10735 * 'current' within the tree based on its new key value. 10736 */ 10737 swap(curr->vruntime, se->vruntime); 10738 resched_curr(rq); 10739 } 10740 10741 se->vruntime -= cfs_rq->min_vruntime; 10742 rq_unlock(rq, &rf); 10743 } 10744 10745 /* 10746 * Priority of the task has changed. Check to see if we preempt 10747 * the current task. 10748 */ 10749 static void 10750 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio) 10751 { 10752 if (!task_on_rq_queued(p)) 10753 return; 10754 10755 if (rq->cfs.nr_running == 1) 10756 return; 10757 10758 /* 10759 * Reschedule if we are currently running on this runqueue and 10760 * our priority decreased, or if we are not currently running on 10761 * this runqueue and our priority is higher than the current's 10762 */ 10763 if (rq->curr == p) { 10764 if (p->prio > oldprio) 10765 resched_curr(rq); 10766 } else 10767 check_preempt_curr(rq, p, 0); 10768 } 10769 10770 static inline bool vruntime_normalized(struct task_struct *p) 10771 { 10772 struct sched_entity *se = &p->se; 10773 10774 /* 10775 * In both the TASK_ON_RQ_QUEUED and TASK_ON_RQ_MIGRATING cases, 10776 * the dequeue_entity(.flags=0) will already have normalized the 10777 * vruntime. 10778 */ 10779 if (p->on_rq) 10780 return true; 10781 10782 /* 10783 * When !on_rq, vruntime of the task has usually NOT been normalized. 10784 * But there are some cases where it has already been normalized: 10785 * 10786 * - A forked child which is waiting for being woken up by 10787 * wake_up_new_task(). 10788 * - A task which has been woken up by try_to_wake_up() and 10789 * waiting for actually being woken up by sched_ttwu_pending(). 10790 */ 10791 if (!se->sum_exec_runtime || 10792 (p->state == TASK_WAKING && p->sched_remote_wakeup)) 10793 return true; 10794 10795 return false; 10796 } 10797 10798 #ifdef CONFIG_FAIR_GROUP_SCHED 10799 /* 10800 * Propagate the changes of the sched_entity across the tg tree to make it 10801 * visible to the root 10802 */ 10803 static void propagate_entity_cfs_rq(struct sched_entity *se) 10804 { 10805 struct cfs_rq *cfs_rq; 10806 10807 /* Start to propagate at parent */ 10808 se = se->parent; 10809 10810 for_each_sched_entity(se) { 10811 cfs_rq = cfs_rq_of(se); 10812 10813 if (cfs_rq_throttled(cfs_rq)) 10814 break; 10815 10816 update_load_avg(cfs_rq, se, UPDATE_TG); 10817 } 10818 } 10819 #else 10820 static void propagate_entity_cfs_rq(struct sched_entity *se) { } 10821 #endif 10822 10823 static void detach_entity_cfs_rq(struct sched_entity *se) 10824 { 10825 struct cfs_rq *cfs_rq = cfs_rq_of(se); 10826 10827 /* Catch up with the cfs_rq and remove our load when we leave */ 10828 update_load_avg(cfs_rq, se, 0); 10829 detach_entity_load_avg(cfs_rq, se); 10830 update_tg_load_avg(cfs_rq); 10831 propagate_entity_cfs_rq(se); 10832 } 10833 10834 static void attach_entity_cfs_rq(struct sched_entity *se) 10835 { 10836 struct cfs_rq *cfs_rq = cfs_rq_of(se); 10837 10838 #ifdef CONFIG_FAIR_GROUP_SCHED 10839 /* 10840 * Since the real-depth could have been changed (only FAIR 10841 * class maintain depth value), reset depth properly. 10842 */ 10843 se->depth = se->parent ? se->parent->depth + 1 : 0; 10844 #endif 10845 10846 /* Synchronize entity with its cfs_rq */ 10847 update_load_avg(cfs_rq, se, sched_feat(ATTACH_AGE_LOAD) ? 0 : SKIP_AGE_LOAD); 10848 attach_entity_load_avg(cfs_rq, se); 10849 update_tg_load_avg(cfs_rq); 10850 propagate_entity_cfs_rq(se); 10851 } 10852 10853 static void detach_task_cfs_rq(struct task_struct *p) 10854 { 10855 struct sched_entity *se = &p->se; 10856 struct cfs_rq *cfs_rq = cfs_rq_of(se); 10857 10858 if (!vruntime_normalized(p)) { 10859 /* 10860 * Fix up our vruntime so that the current sleep doesn't 10861 * cause 'unlimited' sleep bonus. 10862 */ 10863 place_entity(cfs_rq, se, 0); 10864 se->vruntime -= cfs_rq->min_vruntime; 10865 } 10866 10867 detach_entity_cfs_rq(se); 10868 } 10869 10870 static void attach_task_cfs_rq(struct task_struct *p) 10871 { 10872 struct sched_entity *se = &p->se; 10873 struct cfs_rq *cfs_rq = cfs_rq_of(se); 10874 10875 attach_entity_cfs_rq(se); 10876 10877 if (!vruntime_normalized(p)) 10878 se->vruntime += cfs_rq->min_vruntime; 10879 } 10880 10881 static void switched_from_fair(struct rq *rq, struct task_struct *p) 10882 { 10883 detach_task_cfs_rq(p); 10884 } 10885 10886 static void switched_to_fair(struct rq *rq, struct task_struct *p) 10887 { 10888 attach_task_cfs_rq(p); 10889 10890 if (task_on_rq_queued(p)) { 10891 /* 10892 * We were most likely switched from sched_rt, so 10893 * kick off the schedule if running, otherwise just see 10894 * if we can still preempt the current task. 10895 */ 10896 if (rq->curr == p) 10897 resched_curr(rq); 10898 else 10899 check_preempt_curr(rq, p, 0); 10900 } 10901 } 10902 10903 /* Account for a task changing its policy or group. 10904 * 10905 * This routine is mostly called to set cfs_rq->curr field when a task 10906 * migrates between groups/classes. 10907 */ 10908 static void set_next_task_fair(struct rq *rq, struct task_struct *p, bool first) 10909 { 10910 struct sched_entity *se = &p->se; 10911 10912 #ifdef CONFIG_SMP 10913 if (task_on_rq_queued(p)) { 10914 /* 10915 * Move the next running task to the front of the list, so our 10916 * cfs_tasks list becomes MRU one. 10917 */ 10918 list_move(&se->group_node, &rq->cfs_tasks); 10919 } 10920 #endif 10921 10922 for_each_sched_entity(se) { 10923 struct cfs_rq *cfs_rq = cfs_rq_of(se); 10924 10925 set_next_entity(cfs_rq, se); 10926 /* ensure bandwidth has been allocated on our new cfs_rq */ 10927 account_cfs_rq_runtime(cfs_rq, 0); 10928 } 10929 } 10930 10931 void init_cfs_rq(struct cfs_rq *cfs_rq) 10932 { 10933 cfs_rq->tasks_timeline = RB_ROOT_CACHED; 10934 cfs_rq->min_vruntime = (u64)(-(1LL << 20)); 10935 #ifndef CONFIG_64BIT 10936 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime; 10937 #endif 10938 #ifdef CONFIG_SMP 10939 raw_spin_lock_init(&cfs_rq->removed.lock); 10940 #endif 10941 } 10942 10943 #ifdef CONFIG_FAIR_GROUP_SCHED 10944 static void task_set_group_fair(struct task_struct *p) 10945 { 10946 struct sched_entity *se = &p->se; 10947 10948 set_task_rq(p, task_cpu(p)); 10949 se->depth = se->parent ? se->parent->depth + 1 : 0; 10950 } 10951 10952 static void task_move_group_fair(struct task_struct *p) 10953 { 10954 detach_task_cfs_rq(p); 10955 set_task_rq(p, task_cpu(p)); 10956 10957 #ifdef CONFIG_SMP 10958 /* Tell se's cfs_rq has been changed -- migrated */ 10959 p->se.avg.last_update_time = 0; 10960 #endif 10961 attach_task_cfs_rq(p); 10962 } 10963 10964 static void task_change_group_fair(struct task_struct *p, int type) 10965 { 10966 switch (type) { 10967 case TASK_SET_GROUP: 10968 task_set_group_fair(p); 10969 break; 10970 10971 case TASK_MOVE_GROUP: 10972 task_move_group_fair(p); 10973 break; 10974 } 10975 } 10976 10977 void free_fair_sched_group(struct task_group *tg) 10978 { 10979 int i; 10980 10981 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg)); 10982 10983 for_each_possible_cpu(i) { 10984 if (tg->cfs_rq) 10985 kfree(tg->cfs_rq[i]); 10986 if (tg->se) 10987 kfree(tg->se[i]); 10988 } 10989 10990 kfree(tg->cfs_rq); 10991 kfree(tg->se); 10992 } 10993 10994 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent) 10995 { 10996 struct sched_entity *se; 10997 struct cfs_rq *cfs_rq; 10998 int i; 10999 11000 tg->cfs_rq = kcalloc(nr_cpu_ids, sizeof(cfs_rq), GFP_KERNEL); 11001 if (!tg->cfs_rq) 11002 goto err; 11003 tg->se = kcalloc(nr_cpu_ids, sizeof(se), GFP_KERNEL); 11004 if (!tg->se) 11005 goto err; 11006 11007 tg->shares = NICE_0_LOAD; 11008 11009 init_cfs_bandwidth(tg_cfs_bandwidth(tg)); 11010 11011 for_each_possible_cpu(i) { 11012 cfs_rq = kzalloc_node(sizeof(struct cfs_rq), 11013 GFP_KERNEL, cpu_to_node(i)); 11014 if (!cfs_rq) 11015 goto err; 11016 11017 se = kzalloc_node(sizeof(struct sched_entity), 11018 GFP_KERNEL, cpu_to_node(i)); 11019 if (!se) 11020 goto err_free_rq; 11021 11022 init_cfs_rq(cfs_rq); 11023 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]); 11024 init_entity_runnable_average(se); 11025 } 11026 11027 return 1; 11028 11029 err_free_rq: 11030 kfree(cfs_rq); 11031 err: 11032 return 0; 11033 } 11034 11035 void online_fair_sched_group(struct task_group *tg) 11036 { 11037 struct sched_entity *se; 11038 struct rq_flags rf; 11039 struct rq *rq; 11040 int i; 11041 11042 for_each_possible_cpu(i) { 11043 rq = cpu_rq(i); 11044 se = tg->se[i]; 11045 rq_lock_irq(rq, &rf); 11046 update_rq_clock(rq); 11047 attach_entity_cfs_rq(se); 11048 sync_throttle(tg, i); 11049 rq_unlock_irq(rq, &rf); 11050 } 11051 } 11052 11053 void unregister_fair_sched_group(struct task_group *tg) 11054 { 11055 unsigned long flags; 11056 struct rq *rq; 11057 int cpu; 11058 11059 for_each_possible_cpu(cpu) { 11060 if (tg->se[cpu]) 11061 remove_entity_load_avg(tg->se[cpu]); 11062 11063 /* 11064 * Only empty task groups can be destroyed; so we can speculatively 11065 * check on_list without danger of it being re-added. 11066 */ 11067 if (!tg->cfs_rq[cpu]->on_list) 11068 continue; 11069 11070 rq = cpu_rq(cpu); 11071 11072 raw_spin_lock_irqsave(&rq->lock, flags); 11073 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]); 11074 raw_spin_unlock_irqrestore(&rq->lock, flags); 11075 } 11076 } 11077 11078 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq, 11079 struct sched_entity *se, int cpu, 11080 struct sched_entity *parent) 11081 { 11082 struct rq *rq = cpu_rq(cpu); 11083 11084 cfs_rq->tg = tg; 11085 cfs_rq->rq = rq; 11086 init_cfs_rq_runtime(cfs_rq); 11087 11088 tg->cfs_rq[cpu] = cfs_rq; 11089 tg->se[cpu] = se; 11090 11091 /* se could be NULL for root_task_group */ 11092 if (!se) 11093 return; 11094 11095 if (!parent) { 11096 se->cfs_rq = &rq->cfs; 11097 se->depth = 0; 11098 } else { 11099 se->cfs_rq = parent->my_q; 11100 se->depth = parent->depth + 1; 11101 } 11102 11103 se->my_q = cfs_rq; 11104 /* guarantee group entities always have weight */ 11105 update_load_set(&se->load, NICE_0_LOAD); 11106 se->parent = parent; 11107 } 11108 11109 static DEFINE_MUTEX(shares_mutex); 11110 11111 int sched_group_set_shares(struct task_group *tg, unsigned long shares) 11112 { 11113 int i; 11114 11115 /* 11116 * We can't change the weight of the root cgroup. 11117 */ 11118 if (!tg->se[0]) 11119 return -EINVAL; 11120 11121 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES)); 11122 11123 mutex_lock(&shares_mutex); 11124 if (tg->shares == shares) 11125 goto done; 11126 11127 tg->shares = shares; 11128 for_each_possible_cpu(i) { 11129 struct rq *rq = cpu_rq(i); 11130 struct sched_entity *se = tg->se[i]; 11131 struct rq_flags rf; 11132 11133 /* Propagate contribution to hierarchy */ 11134 rq_lock_irqsave(rq, &rf); 11135 update_rq_clock(rq); 11136 for_each_sched_entity(se) { 11137 update_load_avg(cfs_rq_of(se), se, UPDATE_TG); 11138 update_cfs_group(se); 11139 } 11140 rq_unlock_irqrestore(rq, &rf); 11141 } 11142 11143 done: 11144 mutex_unlock(&shares_mutex); 11145 return 0; 11146 } 11147 #else /* CONFIG_FAIR_GROUP_SCHED */ 11148 11149 void free_fair_sched_group(struct task_group *tg) { } 11150 11151 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent) 11152 { 11153 return 1; 11154 } 11155 11156 void online_fair_sched_group(struct task_group *tg) { } 11157 11158 void unregister_fair_sched_group(struct task_group *tg) { } 11159 11160 #endif /* CONFIG_FAIR_GROUP_SCHED */ 11161 11162 11163 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task) 11164 { 11165 struct sched_entity *se = &task->se; 11166 unsigned int rr_interval = 0; 11167 11168 /* 11169 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise 11170 * idle runqueue: 11171 */ 11172 if (rq->cfs.load.weight) 11173 rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se)); 11174 11175 return rr_interval; 11176 } 11177 11178 /* 11179 * All the scheduling class methods: 11180 */ 11181 const struct sched_class fair_sched_class 11182 __section("__fair_sched_class") = { 11183 .enqueue_task = enqueue_task_fair, 11184 .dequeue_task = dequeue_task_fair, 11185 .yield_task = yield_task_fair, 11186 .yield_to_task = yield_to_task_fair, 11187 11188 .check_preempt_curr = check_preempt_wakeup, 11189 11190 .pick_next_task = __pick_next_task_fair, 11191 .put_prev_task = put_prev_task_fair, 11192 .set_next_task = set_next_task_fair, 11193 11194 #ifdef CONFIG_SMP 11195 .balance = balance_fair, 11196 .select_task_rq = select_task_rq_fair, 11197 .migrate_task_rq = migrate_task_rq_fair, 11198 11199 .rq_online = rq_online_fair, 11200 .rq_offline = rq_offline_fair, 11201 11202 .task_dead = task_dead_fair, 11203 .set_cpus_allowed = set_cpus_allowed_common, 11204 #endif 11205 11206 .task_tick = task_tick_fair, 11207 .task_fork = task_fork_fair, 11208 11209 .prio_changed = prio_changed_fair, 11210 .switched_from = switched_from_fair, 11211 .switched_to = switched_to_fair, 11212 11213 .get_rr_interval = get_rr_interval_fair, 11214 11215 .update_curr = update_curr_fair, 11216 11217 #ifdef CONFIG_FAIR_GROUP_SCHED 11218 .task_change_group = task_change_group_fair, 11219 #endif 11220 11221 #ifdef CONFIG_UCLAMP_TASK 11222 .uclamp_enabled = 1, 11223 #endif 11224 }; 11225 11226 #ifdef CONFIG_SCHED_DEBUG 11227 void print_cfs_stats(struct seq_file *m, int cpu) 11228 { 11229 struct cfs_rq *cfs_rq, *pos; 11230 11231 rcu_read_lock(); 11232 for_each_leaf_cfs_rq_safe(cpu_rq(cpu), cfs_rq, pos) 11233 print_cfs_rq(m, cpu, cfs_rq); 11234 rcu_read_unlock(); 11235 } 11236 11237 #ifdef CONFIG_NUMA_BALANCING 11238 void show_numa_stats(struct task_struct *p, struct seq_file *m) 11239 { 11240 int node; 11241 unsigned long tsf = 0, tpf = 0, gsf = 0, gpf = 0; 11242 struct numa_group *ng; 11243 11244 rcu_read_lock(); 11245 ng = rcu_dereference(p->numa_group); 11246 for_each_online_node(node) { 11247 if (p->numa_faults) { 11248 tsf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 0)]; 11249 tpf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 1)]; 11250 } 11251 if (ng) { 11252 gsf = ng->faults[task_faults_idx(NUMA_MEM, node, 0)], 11253 gpf = ng->faults[task_faults_idx(NUMA_MEM, node, 1)]; 11254 } 11255 print_numa_stats(m, node, tsf, tpf, gsf, gpf); 11256 } 11257 rcu_read_unlock(); 11258 } 11259 #endif /* CONFIG_NUMA_BALANCING */ 11260 #endif /* CONFIG_SCHED_DEBUG */ 11261 11262 __init void init_sched_fair_class(void) 11263 { 11264 #ifdef CONFIG_SMP 11265 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains); 11266 11267 #ifdef CONFIG_NO_HZ_COMMON 11268 nohz.next_balance = jiffies; 11269 nohz.next_blocked = jiffies; 11270 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT); 11271 #endif 11272 #endif /* SMP */ 11273 11274 } 11275 11276 /* 11277 * Helper functions to facilitate extracting info from tracepoints. 11278 */ 11279 11280 const struct sched_avg *sched_trace_cfs_rq_avg(struct cfs_rq *cfs_rq) 11281 { 11282 #ifdef CONFIG_SMP 11283 return cfs_rq ? &cfs_rq->avg : NULL; 11284 #else 11285 return NULL; 11286 #endif 11287 } 11288 EXPORT_SYMBOL_GPL(sched_trace_cfs_rq_avg); 11289 11290 char *sched_trace_cfs_rq_path(struct cfs_rq *cfs_rq, char *str, int len) 11291 { 11292 if (!cfs_rq) { 11293 if (str) 11294 strlcpy(str, "(null)", len); 11295 else 11296 return NULL; 11297 } 11298 11299 cfs_rq_tg_path(cfs_rq, str, len); 11300 return str; 11301 } 11302 EXPORT_SYMBOL_GPL(sched_trace_cfs_rq_path); 11303 11304 int sched_trace_cfs_rq_cpu(struct cfs_rq *cfs_rq) 11305 { 11306 return cfs_rq ? cpu_of(rq_of(cfs_rq)) : -1; 11307 } 11308 EXPORT_SYMBOL_GPL(sched_trace_cfs_rq_cpu); 11309 11310 const struct sched_avg *sched_trace_rq_avg_rt(struct rq *rq) 11311 { 11312 #ifdef CONFIG_SMP 11313 return rq ? &rq->avg_rt : NULL; 11314 #else 11315 return NULL; 11316 #endif 11317 } 11318 EXPORT_SYMBOL_GPL(sched_trace_rq_avg_rt); 11319 11320 const struct sched_avg *sched_trace_rq_avg_dl(struct rq *rq) 11321 { 11322 #ifdef CONFIG_SMP 11323 return rq ? &rq->avg_dl : NULL; 11324 #else 11325 return NULL; 11326 #endif 11327 } 11328 EXPORT_SYMBOL_GPL(sched_trace_rq_avg_dl); 11329 11330 const struct sched_avg *sched_trace_rq_avg_irq(struct rq *rq) 11331 { 11332 #if defined(CONFIG_SMP) && defined(CONFIG_HAVE_SCHED_AVG_IRQ) 11333 return rq ? &rq->avg_irq : NULL; 11334 #else 11335 return NULL; 11336 #endif 11337 } 11338 EXPORT_SYMBOL_GPL(sched_trace_rq_avg_irq); 11339 11340 int sched_trace_rq_cpu(struct rq *rq) 11341 { 11342 return rq ? cpu_of(rq) : -1; 11343 } 11344 EXPORT_SYMBOL_GPL(sched_trace_rq_cpu); 11345 11346 int sched_trace_rq_cpu_capacity(struct rq *rq) 11347 { 11348 return rq ? 11349 #ifdef CONFIG_SMP 11350 rq->cpu_capacity 11351 #else 11352 SCHED_CAPACITY_SCALE 11353 #endif 11354 : -1; 11355 } 11356 EXPORT_SYMBOL_GPL(sched_trace_rq_cpu_capacity); 11357 11358 const struct cpumask *sched_trace_rd_span(struct root_domain *rd) 11359 { 11360 #ifdef CONFIG_SMP 11361 return rd ? rd->span : NULL; 11362 #else 11363 return NULL; 11364 #endif 11365 } 11366 EXPORT_SYMBOL_GPL(sched_trace_rd_span); 11367 11368 int sched_trace_rq_nr_running(struct rq *rq) 11369 { 11370 return rq ? rq->nr_running : -1; 11371 } 11372 EXPORT_SYMBOL_GPL(sched_trace_rq_nr_running); 11373