core.c revision b8a216269ec0ce2e961d32e6d640d7010b8a818e
1/* 2 * kernel/sched/core.c 3 * 4 * Kernel scheduler and related syscalls 5 * 6 * Copyright (C) 1991-2002 Linus Torvalds 7 * 8 * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and 9 * make semaphores SMP safe 10 * 1998-11-19 Implemented schedule_timeout() and related stuff 11 * by Andrea Arcangeli 12 * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar: 13 * hybrid priority-list and round-robin design with 14 * an array-switch method of distributing timeslices 15 * and per-CPU runqueues. Cleanups and useful suggestions 16 * by Davide Libenzi, preemptible kernel bits by Robert Love. 17 * 2003-09-03 Interactivity tuning by Con Kolivas. 18 * 2004-04-02 Scheduler domains code by Nick Piggin 19 * 2007-04-15 Work begun on replacing all interactivity tuning with a 20 * fair scheduling design by Con Kolivas. 21 * 2007-05-05 Load balancing (smp-nice) and other improvements 22 * by Peter Williams 23 * 2007-05-06 Interactivity improvements to CFS by Mike Galbraith 24 * 2007-07-01 Group scheduling enhancements by Srivatsa Vaddagiri 25 * 2007-11-29 RT balancing improvements by Steven Rostedt, Gregory Haskins, 26 * Thomas Gleixner, Mike Kravetz 27 */ 28 29#include <linux/mm.h> 30#include <linux/module.h> 31#include <linux/nmi.h> 32#include <linux/init.h> 33#include <linux/uaccess.h> 34#include <linux/highmem.h> 35#include <asm/mmu_context.h> 36#include <linux/interrupt.h> 37#include <linux/capability.h> 38#include <linux/completion.h> 39#include <linux/kernel_stat.h> 40#include <linux/debug_locks.h> 41#include <linux/perf_event.h> 42#include <linux/security.h> 43#include <linux/notifier.h> 44#include <linux/profile.h> 45#include <linux/freezer.h> 46#include <linux/vmalloc.h> 47#include <linux/blkdev.h> 48#include <linux/delay.h> 49#include <linux/pid_namespace.h> 50#include <linux/smp.h> 51#include <linux/threads.h> 52#include <linux/timer.h> 53#include <linux/rcupdate.h> 54#include <linux/cpu.h> 55#include <linux/cpuset.h> 56#include <linux/percpu.h> 57#include <linux/proc_fs.h> 58#include <linux/seq_file.h> 59#include <linux/sysctl.h> 60#include <linux/syscalls.h> 61#include <linux/times.h> 62#include <linux/tsacct_kern.h> 63#include <linux/kprobes.h> 64#include <linux/delayacct.h> 65#include <linux/unistd.h> 66#include <linux/pagemap.h> 67#include <linux/hrtimer.h> 68#include <linux/tick.h> 69#include <linux/debugfs.h> 70#include <linux/ctype.h> 71#include <linux/ftrace.h> 72#include <linux/slab.h> 73#include <linux/init_task.h> 74#include <linux/binfmts.h> 75#include <linux/context_tracking.h> 76 77#include <asm/switch_to.h> 78#include <asm/tlb.h> 79#include <asm/irq_regs.h> 80#include <asm/mutex.h> 81#ifdef CONFIG_PARAVIRT 82#include <asm/paravirt.h> 83#endif 84 85#include "sched.h" 86#include "../workqueue_internal.h" 87#include "../smpboot.h" 88 89#define CREATE_TRACE_POINTS 90#include <trace/events/sched.h> 91 92void start_bandwidth_timer(struct hrtimer *period_timer, ktime_t period) 93{ 94 unsigned long delta; 95 ktime_t soft, hard, now; 96 97 for (;;) { 98 if (hrtimer_active(period_timer)) 99 break; 100 101 now = hrtimer_cb_get_time(period_timer); 102 hrtimer_forward(period_timer, now, period); 103 104 soft = hrtimer_get_softexpires(period_timer); 105 hard = hrtimer_get_expires(period_timer); 106 delta = ktime_to_ns(ktime_sub(hard, soft)); 107 __hrtimer_start_range_ns(period_timer, soft, delta, 108 HRTIMER_MODE_ABS_PINNED, 0); 109 } 110} 111 112DEFINE_MUTEX(sched_domains_mutex); 113DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues); 114 115static void update_rq_clock_task(struct rq *rq, s64 delta); 116 117void update_rq_clock(struct rq *rq) 118{ 119 s64 delta; 120 121 if (rq->skip_clock_update > 0) 122 return; 123 124 delta = sched_clock_cpu(cpu_of(rq)) - rq->clock; 125 rq->clock += delta; 126 update_rq_clock_task(rq, delta); 127} 128 129/* 130 * Debugging: various feature bits 131 */ 132 133#define SCHED_FEAT(name, enabled) \ 134 (1UL << __SCHED_FEAT_##name) * enabled | 135 136const_debug unsigned int sysctl_sched_features = 137#include "features.h" 138 0; 139 140#undef SCHED_FEAT 141 142#ifdef CONFIG_SCHED_DEBUG 143#define SCHED_FEAT(name, enabled) \ 144 #name , 145 146static const char * const sched_feat_names[] = { 147#include "features.h" 148}; 149 150#undef SCHED_FEAT 151 152static int sched_feat_show(struct seq_file *m, void *v) 153{ 154 int i; 155 156 for (i = 0; i < __SCHED_FEAT_NR; i++) { 157 if (!(sysctl_sched_features & (1UL << i))) 158 seq_puts(m, "NO_"); 159 seq_printf(m, "%s ", sched_feat_names[i]); 160 } 161 seq_puts(m, "\n"); 162 163 return 0; 164} 165 166#ifdef HAVE_JUMP_LABEL 167 168#define jump_label_key__true STATIC_KEY_INIT_TRUE 169#define jump_label_key__false STATIC_KEY_INIT_FALSE 170 171#define SCHED_FEAT(name, enabled) \ 172 jump_label_key__##enabled , 173 174struct static_key sched_feat_keys[__SCHED_FEAT_NR] = { 175#include "features.h" 176}; 177 178#undef SCHED_FEAT 179 180static void sched_feat_disable(int i) 181{ 182 if (static_key_enabled(&sched_feat_keys[i])) 183 static_key_slow_dec(&sched_feat_keys[i]); 184} 185 186static void sched_feat_enable(int i) 187{ 188 if (!static_key_enabled(&sched_feat_keys[i])) 189 static_key_slow_inc(&sched_feat_keys[i]); 190} 191#else 192static void sched_feat_disable(int i) { }; 193static void sched_feat_enable(int i) { }; 194#endif /* HAVE_JUMP_LABEL */ 195 196static int sched_feat_set(char *cmp) 197{ 198 int i; 199 int neg = 0; 200 201 if (strncmp(cmp, "NO_", 3) == 0) { 202 neg = 1; 203 cmp += 3; 204 } 205 206 for (i = 0; i < __SCHED_FEAT_NR; i++) { 207 if (strcmp(cmp, sched_feat_names[i]) == 0) { 208 if (neg) { 209 sysctl_sched_features &= ~(1UL << i); 210 sched_feat_disable(i); 211 } else { 212 sysctl_sched_features |= (1UL << i); 213 sched_feat_enable(i); 214 } 215 break; 216 } 217 } 218 219 return i; 220} 221 222static ssize_t 223sched_feat_write(struct file *filp, const char __user *ubuf, 224 size_t cnt, loff_t *ppos) 225{ 226 char buf[64]; 227 char *cmp; 228 int i; 229 230 if (cnt > 63) 231 cnt = 63; 232 233 if (copy_from_user(&buf, ubuf, cnt)) 234 return -EFAULT; 235 236 buf[cnt] = 0; 237 cmp = strstrip(buf); 238 239 i = sched_feat_set(cmp); 240 if (i == __SCHED_FEAT_NR) 241 return -EINVAL; 242 243 *ppos += cnt; 244 245 return cnt; 246} 247 248static int sched_feat_open(struct inode *inode, struct file *filp) 249{ 250 return single_open(filp, sched_feat_show, NULL); 251} 252 253static const struct file_operations sched_feat_fops = { 254 .open = sched_feat_open, 255 .write = sched_feat_write, 256 .read = seq_read, 257 .llseek = seq_lseek, 258 .release = single_release, 259}; 260 261static __init int sched_init_debug(void) 262{ 263 debugfs_create_file("sched_features", 0644, NULL, NULL, 264 &sched_feat_fops); 265 266 return 0; 267} 268late_initcall(sched_init_debug); 269#endif /* CONFIG_SCHED_DEBUG */ 270 271/* 272 * Number of tasks to iterate in a single balance run. 273 * Limited because this is done with IRQs disabled. 274 */ 275const_debug unsigned int sysctl_sched_nr_migrate = 32; 276 277/* 278 * period over which we average the RT time consumption, measured 279 * in ms. 280 * 281 * default: 1s 282 */ 283const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC; 284 285/* 286 * period over which we measure -rt task cpu usage in us. 287 * default: 1s 288 */ 289unsigned int sysctl_sched_rt_period = 1000000; 290 291__read_mostly int scheduler_running; 292 293/* 294 * part of the period that we allow rt tasks to run in us. 295 * default: 0.95s 296 */ 297int sysctl_sched_rt_runtime = 950000; 298 299 300 301/* 302 * __task_rq_lock - lock the rq @p resides on. 303 */ 304static inline struct rq *__task_rq_lock(struct task_struct *p) 305 __acquires(rq->lock) 306{ 307 struct rq *rq; 308 309 lockdep_assert_held(&p->pi_lock); 310 311 for (;;) { 312 rq = task_rq(p); 313 raw_spin_lock(&rq->lock); 314 if (likely(rq == task_rq(p))) 315 return rq; 316 raw_spin_unlock(&rq->lock); 317 } 318} 319 320/* 321 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on. 322 */ 323static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags) 324 __acquires(p->pi_lock) 325 __acquires(rq->lock) 326{ 327 struct rq *rq; 328 329 for (;;) { 330 raw_spin_lock_irqsave(&p->pi_lock, *flags); 331 rq = task_rq(p); 332 raw_spin_lock(&rq->lock); 333 if (likely(rq == task_rq(p))) 334 return rq; 335 raw_spin_unlock(&rq->lock); 336 raw_spin_unlock_irqrestore(&p->pi_lock, *flags); 337 } 338} 339 340static void __task_rq_unlock(struct rq *rq) 341 __releases(rq->lock) 342{ 343 raw_spin_unlock(&rq->lock); 344} 345 346static inline void 347task_rq_unlock(struct rq *rq, struct task_struct *p, unsigned long *flags) 348 __releases(rq->lock) 349 __releases(p->pi_lock) 350{ 351 raw_spin_unlock(&rq->lock); 352 raw_spin_unlock_irqrestore(&p->pi_lock, *flags); 353} 354 355/* 356 * this_rq_lock - lock this runqueue and disable interrupts. 357 */ 358static struct rq *this_rq_lock(void) 359 __acquires(rq->lock) 360{ 361 struct rq *rq; 362 363 local_irq_disable(); 364 rq = this_rq(); 365 raw_spin_lock(&rq->lock); 366 367 return rq; 368} 369 370#ifdef CONFIG_SCHED_HRTICK 371/* 372 * Use HR-timers to deliver accurate preemption points. 373 */ 374 375static void hrtick_clear(struct rq *rq) 376{ 377 if (hrtimer_active(&rq->hrtick_timer)) 378 hrtimer_cancel(&rq->hrtick_timer); 379} 380 381/* 382 * High-resolution timer tick. 383 * Runs from hardirq context with interrupts disabled. 384 */ 385static enum hrtimer_restart hrtick(struct hrtimer *timer) 386{ 387 struct rq *rq = container_of(timer, struct rq, hrtick_timer); 388 389 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id()); 390 391 raw_spin_lock(&rq->lock); 392 update_rq_clock(rq); 393 rq->curr->sched_class->task_tick(rq, rq->curr, 1); 394 raw_spin_unlock(&rq->lock); 395 396 return HRTIMER_NORESTART; 397} 398 399#ifdef CONFIG_SMP 400 401static int __hrtick_restart(struct rq *rq) 402{ 403 struct hrtimer *timer = &rq->hrtick_timer; 404 ktime_t time = hrtimer_get_softexpires(timer); 405 406 return __hrtimer_start_range_ns(timer, time, 0, HRTIMER_MODE_ABS_PINNED, 0); 407} 408 409/* 410 * called from hardirq (IPI) context 411 */ 412static void __hrtick_start(void *arg) 413{ 414 struct rq *rq = arg; 415 416 raw_spin_lock(&rq->lock); 417 __hrtick_restart(rq); 418 rq->hrtick_csd_pending = 0; 419 raw_spin_unlock(&rq->lock); 420} 421 422/* 423 * Called to set the hrtick timer state. 424 * 425 * called with rq->lock held and irqs disabled 426 */ 427void hrtick_start(struct rq *rq, u64 delay) 428{ 429 struct hrtimer *timer = &rq->hrtick_timer; 430 ktime_t time = ktime_add_ns(timer->base->get_time(), delay); 431 432 hrtimer_set_expires(timer, time); 433 434 if (rq == this_rq()) { 435 __hrtick_restart(rq); 436 } else if (!rq->hrtick_csd_pending) { 437 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd, 0); 438 rq->hrtick_csd_pending = 1; 439 } 440} 441 442static int 443hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu) 444{ 445 int cpu = (int)(long)hcpu; 446 447 switch (action) { 448 case CPU_UP_CANCELED: 449 case CPU_UP_CANCELED_FROZEN: 450 case CPU_DOWN_PREPARE: 451 case CPU_DOWN_PREPARE_FROZEN: 452 case CPU_DEAD: 453 case CPU_DEAD_FROZEN: 454 hrtick_clear(cpu_rq(cpu)); 455 return NOTIFY_OK; 456 } 457 458 return NOTIFY_DONE; 459} 460 461static __init void init_hrtick(void) 462{ 463 hotcpu_notifier(hotplug_hrtick, 0); 464} 465#else 466/* 467 * Called to set the hrtick timer state. 468 * 469 * called with rq->lock held and irqs disabled 470 */ 471void hrtick_start(struct rq *rq, u64 delay) 472{ 473 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0, 474 HRTIMER_MODE_REL_PINNED, 0); 475} 476 477static inline void init_hrtick(void) 478{ 479} 480#endif /* CONFIG_SMP */ 481 482static void init_rq_hrtick(struct rq *rq) 483{ 484#ifdef CONFIG_SMP 485 rq->hrtick_csd_pending = 0; 486 487 rq->hrtick_csd.flags = 0; 488 rq->hrtick_csd.func = __hrtick_start; 489 rq->hrtick_csd.info = rq; 490#endif 491 492 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); 493 rq->hrtick_timer.function = hrtick; 494} 495#else /* CONFIG_SCHED_HRTICK */ 496static inline void hrtick_clear(struct rq *rq) 497{ 498} 499 500static inline void init_rq_hrtick(struct rq *rq) 501{ 502} 503 504static inline void init_hrtick(void) 505{ 506} 507#endif /* CONFIG_SCHED_HRTICK */ 508 509/* 510 * resched_task - mark a task 'to be rescheduled now'. 511 * 512 * On UP this means the setting of the need_resched flag, on SMP it 513 * might also involve a cross-CPU call to trigger the scheduler on 514 * the target CPU. 515 */ 516void resched_task(struct task_struct *p) 517{ 518 int cpu; 519 520 lockdep_assert_held(&task_rq(p)->lock); 521 522 if (test_tsk_need_resched(p)) 523 return; 524 525 set_tsk_need_resched(p); 526 527 cpu = task_cpu(p); 528 if (cpu == smp_processor_id()) { 529 set_preempt_need_resched(); 530 return; 531 } 532 533 /* NEED_RESCHED must be visible before we test polling */ 534 smp_mb(); 535 if (!tsk_is_polling(p)) 536 smp_send_reschedule(cpu); 537} 538 539void resched_cpu(int cpu) 540{ 541 struct rq *rq = cpu_rq(cpu); 542 unsigned long flags; 543 544 if (!raw_spin_trylock_irqsave(&rq->lock, flags)) 545 return; 546 resched_task(cpu_curr(cpu)); 547 raw_spin_unlock_irqrestore(&rq->lock, flags); 548} 549 550#ifdef CONFIG_SMP 551#ifdef CONFIG_NO_HZ_COMMON 552/* 553 * In the semi idle case, use the nearest busy cpu for migrating timers 554 * from an idle cpu. This is good for power-savings. 555 * 556 * We don't do similar optimization for completely idle system, as 557 * selecting an idle cpu will add more delays to the timers than intended 558 * (as that cpu's timer base may not be uptodate wrt jiffies etc). 559 */ 560int get_nohz_timer_target(void) 561{ 562 int cpu = smp_processor_id(); 563 int i; 564 struct sched_domain *sd; 565 566 rcu_read_lock(); 567 for_each_domain(cpu, sd) { 568 for_each_cpu(i, sched_domain_span(sd)) { 569 if (!idle_cpu(i)) { 570 cpu = i; 571 goto unlock; 572 } 573 } 574 } 575unlock: 576 rcu_read_unlock(); 577 return cpu; 578} 579/* 580 * When add_timer_on() enqueues a timer into the timer wheel of an 581 * idle CPU then this timer might expire before the next timer event 582 * which is scheduled to wake up that CPU. In case of a completely 583 * idle system the next event might even be infinite time into the 584 * future. wake_up_idle_cpu() ensures that the CPU is woken up and 585 * leaves the inner idle loop so the newly added timer is taken into 586 * account when the CPU goes back to idle and evaluates the timer 587 * wheel for the next timer event. 588 */ 589static void wake_up_idle_cpu(int cpu) 590{ 591 struct rq *rq = cpu_rq(cpu); 592 593 if (cpu == smp_processor_id()) 594 return; 595 596 /* 597 * This is safe, as this function is called with the timer 598 * wheel base lock of (cpu) held. When the CPU is on the way 599 * to idle and has not yet set rq->curr to idle then it will 600 * be serialized on the timer wheel base lock and take the new 601 * timer into account automatically. 602 */ 603 if (rq->curr != rq->idle) 604 return; 605 606 /* 607 * We can set TIF_RESCHED on the idle task of the other CPU 608 * lockless. The worst case is that the other CPU runs the 609 * idle task through an additional NOOP schedule() 610 */ 611 set_tsk_need_resched(rq->idle); 612 613 /* NEED_RESCHED must be visible before we test polling */ 614 smp_mb(); 615 if (!tsk_is_polling(rq->idle)) 616 smp_send_reschedule(cpu); 617} 618 619static bool wake_up_full_nohz_cpu(int cpu) 620{ 621 if (tick_nohz_full_cpu(cpu)) { 622 if (cpu != smp_processor_id() || 623 tick_nohz_tick_stopped()) 624 smp_send_reschedule(cpu); 625 return true; 626 } 627 628 return false; 629} 630 631void wake_up_nohz_cpu(int cpu) 632{ 633 if (!wake_up_full_nohz_cpu(cpu)) 634 wake_up_idle_cpu(cpu); 635} 636 637static inline bool got_nohz_idle_kick(void) 638{ 639 int cpu = smp_processor_id(); 640 641 if (!test_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu))) 642 return false; 643 644 if (idle_cpu(cpu) && !need_resched()) 645 return true; 646 647 /* 648 * We can't run Idle Load Balance on this CPU for this time so we 649 * cancel it and clear NOHZ_BALANCE_KICK 650 */ 651 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu)); 652 return false; 653} 654 655#else /* CONFIG_NO_HZ_COMMON */ 656 657static inline bool got_nohz_idle_kick(void) 658{ 659 return false; 660} 661 662#endif /* CONFIG_NO_HZ_COMMON */ 663 664#ifdef CONFIG_NO_HZ_FULL 665bool sched_can_stop_tick(void) 666{ 667 struct rq *rq; 668 669 rq = this_rq(); 670 671 /* Make sure rq->nr_running update is visible after the IPI */ 672 smp_rmb(); 673 674 /* More than one running task need preemption */ 675 if (rq->nr_running > 1) 676 return false; 677 678 return true; 679} 680#endif /* CONFIG_NO_HZ_FULL */ 681 682void sched_avg_update(struct rq *rq) 683{ 684 s64 period = sched_avg_period(); 685 686 while ((s64)(rq_clock(rq) - rq->age_stamp) > period) { 687 /* 688 * Inline assembly required to prevent the compiler 689 * optimising this loop into a divmod call. 690 * See __iter_div_u64_rem() for another example of this. 691 */ 692 asm("" : "+rm" (rq->age_stamp)); 693 rq->age_stamp += period; 694 rq->rt_avg /= 2; 695 } 696} 697 698#endif /* CONFIG_SMP */ 699 700#if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \ 701 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH))) 702/* 703 * Iterate task_group tree rooted at *from, calling @down when first entering a 704 * node and @up when leaving it for the final time. 705 * 706 * Caller must hold rcu_lock or sufficient equivalent. 707 */ 708int walk_tg_tree_from(struct task_group *from, 709 tg_visitor down, tg_visitor up, void *data) 710{ 711 struct task_group *parent, *child; 712 int ret; 713 714 parent = from; 715 716down: 717 ret = (*down)(parent, data); 718 if (ret) 719 goto out; 720 list_for_each_entry_rcu(child, &parent->children, siblings) { 721 parent = child; 722 goto down; 723 724up: 725 continue; 726 } 727 ret = (*up)(parent, data); 728 if (ret || parent == from) 729 goto out; 730 731 child = parent; 732 parent = parent->parent; 733 if (parent) 734 goto up; 735out: 736 return ret; 737} 738 739int tg_nop(struct task_group *tg, void *data) 740{ 741 return 0; 742} 743#endif 744 745static void set_load_weight(struct task_struct *p) 746{ 747 int prio = p->static_prio - MAX_RT_PRIO; 748 struct load_weight *load = &p->se.load; 749 750 /* 751 * SCHED_IDLE tasks get minimal weight: 752 */ 753 if (p->policy == SCHED_IDLE) { 754 load->weight = scale_load(WEIGHT_IDLEPRIO); 755 load->inv_weight = WMULT_IDLEPRIO; 756 return; 757 } 758 759 load->weight = scale_load(prio_to_weight[prio]); 760 load->inv_weight = prio_to_wmult[prio]; 761} 762 763static void enqueue_task(struct rq *rq, struct task_struct *p, int flags) 764{ 765 update_rq_clock(rq); 766 sched_info_queued(rq, p); 767 p->sched_class->enqueue_task(rq, p, flags); 768} 769 770static void dequeue_task(struct rq *rq, struct task_struct *p, int flags) 771{ 772 update_rq_clock(rq); 773 sched_info_dequeued(rq, p); 774 p->sched_class->dequeue_task(rq, p, flags); 775} 776 777void activate_task(struct rq *rq, struct task_struct *p, int flags) 778{ 779 if (task_contributes_to_load(p)) 780 rq->nr_uninterruptible--; 781 782 enqueue_task(rq, p, flags); 783} 784 785void deactivate_task(struct rq *rq, struct task_struct *p, int flags) 786{ 787 if (task_contributes_to_load(p)) 788 rq->nr_uninterruptible++; 789 790 dequeue_task(rq, p, flags); 791} 792 793static void update_rq_clock_task(struct rq *rq, s64 delta) 794{ 795/* 796 * In theory, the compile should just see 0 here, and optimize out the call 797 * to sched_rt_avg_update. But I don't trust it... 798 */ 799#if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING) 800 s64 steal = 0, irq_delta = 0; 801#endif 802#ifdef CONFIG_IRQ_TIME_ACCOUNTING 803 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time; 804 805 /* 806 * Since irq_time is only updated on {soft,}irq_exit, we might run into 807 * this case when a previous update_rq_clock() happened inside a 808 * {soft,}irq region. 809 * 810 * When this happens, we stop ->clock_task and only update the 811 * prev_irq_time stamp to account for the part that fit, so that a next 812 * update will consume the rest. This ensures ->clock_task is 813 * monotonic. 814 * 815 * It does however cause some slight miss-attribution of {soft,}irq 816 * time, a more accurate solution would be to update the irq_time using 817 * the current rq->clock timestamp, except that would require using 818 * atomic ops. 819 */ 820 if (irq_delta > delta) 821 irq_delta = delta; 822 823 rq->prev_irq_time += irq_delta; 824 delta -= irq_delta; 825#endif 826#ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING 827 if (static_key_false((¶virt_steal_rq_enabled))) { 828 u64 st; 829 830 steal = paravirt_steal_clock(cpu_of(rq)); 831 steal -= rq->prev_steal_time_rq; 832 833 if (unlikely(steal > delta)) 834 steal = delta; 835 836 st = steal_ticks(steal); 837 steal = st * TICK_NSEC; 838 839 rq->prev_steal_time_rq += steal; 840 841 delta -= steal; 842 } 843#endif 844 845 rq->clock_task += delta; 846 847#if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING) 848 if ((irq_delta + steal) && sched_feat(NONTASK_POWER)) 849 sched_rt_avg_update(rq, irq_delta + steal); 850#endif 851} 852 853void sched_set_stop_task(int cpu, struct task_struct *stop) 854{ 855 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 }; 856 struct task_struct *old_stop = cpu_rq(cpu)->stop; 857 858 if (stop) { 859 /* 860 * Make it appear like a SCHED_FIFO task, its something 861 * userspace knows about and won't get confused about. 862 * 863 * Also, it will make PI more or less work without too 864 * much confusion -- but then, stop work should not 865 * rely on PI working anyway. 866 */ 867 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m); 868 869 stop->sched_class = &stop_sched_class; 870 } 871 872 cpu_rq(cpu)->stop = stop; 873 874 if (old_stop) { 875 /* 876 * Reset it back to a normal scheduling class so that 877 * it can die in pieces. 878 */ 879 old_stop->sched_class = &rt_sched_class; 880 } 881} 882 883/* 884 * __normal_prio - return the priority that is based on the static prio 885 */ 886static inline int __normal_prio(struct task_struct *p) 887{ 888 return p->static_prio; 889} 890 891/* 892 * Calculate the expected normal priority: i.e. priority 893 * without taking RT-inheritance into account. Might be 894 * boosted by interactivity modifiers. Changes upon fork, 895 * setprio syscalls, and whenever the interactivity 896 * estimator recalculates. 897 */ 898static inline int normal_prio(struct task_struct *p) 899{ 900 int prio; 901 902 if (task_has_rt_policy(p)) 903 prio = MAX_RT_PRIO-1 - p->rt_priority; 904 else 905 prio = __normal_prio(p); 906 return prio; 907} 908 909/* 910 * Calculate the current priority, i.e. the priority 911 * taken into account by the scheduler. This value might 912 * be boosted by RT tasks, or might be boosted by 913 * interactivity modifiers. Will be RT if the task got 914 * RT-boosted. If not then it returns p->normal_prio. 915 */ 916static int effective_prio(struct task_struct *p) 917{ 918 p->normal_prio = normal_prio(p); 919 /* 920 * If we are RT tasks or we were boosted to RT priority, 921 * keep the priority unchanged. Otherwise, update priority 922 * to the normal priority: 923 */ 924 if (!rt_prio(p->prio)) 925 return p->normal_prio; 926 return p->prio; 927} 928 929/** 930 * task_curr - is this task currently executing on a CPU? 931 * @p: the task in question. 932 * 933 * Return: 1 if the task is currently executing. 0 otherwise. 934 */ 935inline int task_curr(const struct task_struct *p) 936{ 937 return cpu_curr(task_cpu(p)) == p; 938} 939 940static inline void check_class_changed(struct rq *rq, struct task_struct *p, 941 const struct sched_class *prev_class, 942 int oldprio) 943{ 944 if (prev_class != p->sched_class) { 945 if (prev_class->switched_from) 946 prev_class->switched_from(rq, p); 947 p->sched_class->switched_to(rq, p); 948 } else if (oldprio != p->prio) 949 p->sched_class->prio_changed(rq, p, oldprio); 950} 951 952void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags) 953{ 954 const struct sched_class *class; 955 956 if (p->sched_class == rq->curr->sched_class) { 957 rq->curr->sched_class->check_preempt_curr(rq, p, flags); 958 } else { 959 for_each_class(class) { 960 if (class == rq->curr->sched_class) 961 break; 962 if (class == p->sched_class) { 963 resched_task(rq->curr); 964 break; 965 } 966 } 967 } 968 969 /* 970 * A queue event has occurred, and we're going to schedule. In 971 * this case, we can save a useless back to back clock update. 972 */ 973 if (rq->curr->on_rq && test_tsk_need_resched(rq->curr)) 974 rq->skip_clock_update = 1; 975} 976 977#ifdef CONFIG_SMP 978void set_task_cpu(struct task_struct *p, unsigned int new_cpu) 979{ 980#ifdef CONFIG_SCHED_DEBUG 981 /* 982 * We should never call set_task_cpu() on a blocked task, 983 * ttwu() will sort out the placement. 984 */ 985 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING && 986 !(task_preempt_count(p) & PREEMPT_ACTIVE)); 987 988#ifdef CONFIG_LOCKDEP 989 /* 990 * The caller should hold either p->pi_lock or rq->lock, when changing 991 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks. 992 * 993 * sched_move_task() holds both and thus holding either pins the cgroup, 994 * see task_group(). 995 * 996 * Furthermore, all task_rq users should acquire both locks, see 997 * task_rq_lock(). 998 */ 999 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) || 1000 lockdep_is_held(&task_rq(p)->lock))); 1001#endif 1002#endif 1003 1004 trace_sched_migrate_task(p, new_cpu); 1005 1006 if (task_cpu(p) != new_cpu) { 1007 if (p->sched_class->migrate_task_rq) 1008 p->sched_class->migrate_task_rq(p, new_cpu); 1009 p->se.nr_migrations++; 1010 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, NULL, 0); 1011 } 1012 1013 __set_task_cpu(p, new_cpu); 1014} 1015 1016static void __migrate_swap_task(struct task_struct *p, int cpu) 1017{ 1018 if (p->on_rq) { 1019 struct rq *src_rq, *dst_rq; 1020 1021 src_rq = task_rq(p); 1022 dst_rq = cpu_rq(cpu); 1023 1024 deactivate_task(src_rq, p, 0); 1025 set_task_cpu(p, cpu); 1026 activate_task(dst_rq, p, 0); 1027 check_preempt_curr(dst_rq, p, 0); 1028 } else { 1029 /* 1030 * Task isn't running anymore; make it appear like we migrated 1031 * it before it went to sleep. This means on wakeup we make the 1032 * previous cpu our targer instead of where it really is. 1033 */ 1034 p->wake_cpu = cpu; 1035 } 1036} 1037 1038struct migration_swap_arg { 1039 struct task_struct *src_task, *dst_task; 1040 int src_cpu, dst_cpu; 1041}; 1042 1043static int migrate_swap_stop(void *data) 1044{ 1045 struct migration_swap_arg *arg = data; 1046 struct rq *src_rq, *dst_rq; 1047 int ret = -EAGAIN; 1048 1049 src_rq = cpu_rq(arg->src_cpu); 1050 dst_rq = cpu_rq(arg->dst_cpu); 1051 1052 double_raw_lock(&arg->src_task->pi_lock, 1053 &arg->dst_task->pi_lock); 1054 double_rq_lock(src_rq, dst_rq); 1055 if (task_cpu(arg->dst_task) != arg->dst_cpu) 1056 goto unlock; 1057 1058 if (task_cpu(arg->src_task) != arg->src_cpu) 1059 goto unlock; 1060 1061 if (!cpumask_test_cpu(arg->dst_cpu, tsk_cpus_allowed(arg->src_task))) 1062 goto unlock; 1063 1064 if (!cpumask_test_cpu(arg->src_cpu, tsk_cpus_allowed(arg->dst_task))) 1065 goto unlock; 1066 1067 __migrate_swap_task(arg->src_task, arg->dst_cpu); 1068 __migrate_swap_task(arg->dst_task, arg->src_cpu); 1069 1070 ret = 0; 1071 1072unlock: 1073 double_rq_unlock(src_rq, dst_rq); 1074 raw_spin_unlock(&arg->dst_task->pi_lock); 1075 raw_spin_unlock(&arg->src_task->pi_lock); 1076 1077 return ret; 1078} 1079 1080/* 1081 * Cross migrate two tasks 1082 */ 1083int migrate_swap(struct task_struct *cur, struct task_struct *p) 1084{ 1085 struct migration_swap_arg arg; 1086 int ret = -EINVAL; 1087 1088 arg = (struct migration_swap_arg){ 1089 .src_task = cur, 1090 .src_cpu = task_cpu(cur), 1091 .dst_task = p, 1092 .dst_cpu = task_cpu(p), 1093 }; 1094 1095 if (arg.src_cpu == arg.dst_cpu) 1096 goto out; 1097 1098 /* 1099 * These three tests are all lockless; this is OK since all of them 1100 * will be re-checked with proper locks held further down the line. 1101 */ 1102 if (!cpu_active(arg.src_cpu) || !cpu_active(arg.dst_cpu)) 1103 goto out; 1104 1105 if (!cpumask_test_cpu(arg.dst_cpu, tsk_cpus_allowed(arg.src_task))) 1106 goto out; 1107 1108 if (!cpumask_test_cpu(arg.src_cpu, tsk_cpus_allowed(arg.dst_task))) 1109 goto out; 1110 1111 ret = stop_two_cpus(arg.dst_cpu, arg.src_cpu, migrate_swap_stop, &arg); 1112 1113out: 1114 return ret; 1115} 1116 1117struct migration_arg { 1118 struct task_struct *task; 1119 int dest_cpu; 1120}; 1121 1122static int migration_cpu_stop(void *data); 1123 1124/* 1125 * wait_task_inactive - wait for a thread to unschedule. 1126 * 1127 * If @match_state is nonzero, it's the @p->state value just checked and 1128 * not expected to change. If it changes, i.e. @p might have woken up, 1129 * then return zero. When we succeed in waiting for @p to be off its CPU, 1130 * we return a positive number (its total switch count). If a second call 1131 * a short while later returns the same number, the caller can be sure that 1132 * @p has remained unscheduled the whole time. 1133 * 1134 * The caller must ensure that the task *will* unschedule sometime soon, 1135 * else this function might spin for a *long* time. This function can't 1136 * be called with interrupts off, or it may introduce deadlock with 1137 * smp_call_function() if an IPI is sent by the same process we are 1138 * waiting to become inactive. 1139 */ 1140unsigned long wait_task_inactive(struct task_struct *p, long match_state) 1141{ 1142 unsigned long flags; 1143 int running, on_rq; 1144 unsigned long ncsw; 1145 struct rq *rq; 1146 1147 for (;;) { 1148 /* 1149 * We do the initial early heuristics without holding 1150 * any task-queue locks at all. We'll only try to get 1151 * the runqueue lock when things look like they will 1152 * work out! 1153 */ 1154 rq = task_rq(p); 1155 1156 /* 1157 * If the task is actively running on another CPU 1158 * still, just relax and busy-wait without holding 1159 * any locks. 1160 * 1161 * NOTE! Since we don't hold any locks, it's not 1162 * even sure that "rq" stays as the right runqueue! 1163 * But we don't care, since "task_running()" will 1164 * return false if the runqueue has changed and p 1165 * is actually now running somewhere else! 1166 */ 1167 while (task_running(rq, p)) { 1168 if (match_state && unlikely(p->state != match_state)) 1169 return 0; 1170 cpu_relax(); 1171 } 1172 1173 /* 1174 * Ok, time to look more closely! We need the rq 1175 * lock now, to be *sure*. If we're wrong, we'll 1176 * just go back and repeat. 1177 */ 1178 rq = task_rq_lock(p, &flags); 1179 trace_sched_wait_task(p); 1180 running = task_running(rq, p); 1181 on_rq = p->on_rq; 1182 ncsw = 0; 1183 if (!match_state || p->state == match_state) 1184 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */ 1185 task_rq_unlock(rq, p, &flags); 1186 1187 /* 1188 * If it changed from the expected state, bail out now. 1189 */ 1190 if (unlikely(!ncsw)) 1191 break; 1192 1193 /* 1194 * Was it really running after all now that we 1195 * checked with the proper locks actually held? 1196 * 1197 * Oops. Go back and try again.. 1198 */ 1199 if (unlikely(running)) { 1200 cpu_relax(); 1201 continue; 1202 } 1203 1204 /* 1205 * It's not enough that it's not actively running, 1206 * it must be off the runqueue _entirely_, and not 1207 * preempted! 1208 * 1209 * So if it was still runnable (but just not actively 1210 * running right now), it's preempted, and we should 1211 * yield - it could be a while. 1212 */ 1213 if (unlikely(on_rq)) { 1214 ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ); 1215 1216 set_current_state(TASK_UNINTERRUPTIBLE); 1217 schedule_hrtimeout(&to, HRTIMER_MODE_REL); 1218 continue; 1219 } 1220 1221 /* 1222 * Ahh, all good. It wasn't running, and it wasn't 1223 * runnable, which means that it will never become 1224 * running in the future either. We're all done! 1225 */ 1226 break; 1227 } 1228 1229 return ncsw; 1230} 1231 1232/*** 1233 * kick_process - kick a running thread to enter/exit the kernel 1234 * @p: the to-be-kicked thread 1235 * 1236 * Cause a process which is running on another CPU to enter 1237 * kernel-mode, without any delay. (to get signals handled.) 1238 * 1239 * NOTE: this function doesn't have to take the runqueue lock, 1240 * because all it wants to ensure is that the remote task enters 1241 * the kernel. If the IPI races and the task has been migrated 1242 * to another CPU then no harm is done and the purpose has been 1243 * achieved as well. 1244 */ 1245void kick_process(struct task_struct *p) 1246{ 1247 int cpu; 1248 1249 preempt_disable(); 1250 cpu = task_cpu(p); 1251 if ((cpu != smp_processor_id()) && task_curr(p)) 1252 smp_send_reschedule(cpu); 1253 preempt_enable(); 1254} 1255EXPORT_SYMBOL_GPL(kick_process); 1256#endif /* CONFIG_SMP */ 1257 1258#ifdef CONFIG_SMP 1259/* 1260 * ->cpus_allowed is protected by both rq->lock and p->pi_lock 1261 */ 1262static int select_fallback_rq(int cpu, struct task_struct *p) 1263{ 1264 int nid = cpu_to_node(cpu); 1265 const struct cpumask *nodemask = NULL; 1266 enum { cpuset, possible, fail } state = cpuset; 1267 int dest_cpu; 1268 1269 /* 1270 * If the node that the cpu is on has been offlined, cpu_to_node() 1271 * will return -1. There is no cpu on the node, and we should 1272 * select the cpu on the other node. 1273 */ 1274 if (nid != -1) { 1275 nodemask = cpumask_of_node(nid); 1276 1277 /* Look for allowed, online CPU in same node. */ 1278 for_each_cpu(dest_cpu, nodemask) { 1279 if (!cpu_online(dest_cpu)) 1280 continue; 1281 if (!cpu_active(dest_cpu)) 1282 continue; 1283 if (cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p))) 1284 return dest_cpu; 1285 } 1286 } 1287 1288 for (;;) { 1289 /* Any allowed, online CPU? */ 1290 for_each_cpu(dest_cpu, tsk_cpus_allowed(p)) { 1291 if (!cpu_online(dest_cpu)) 1292 continue; 1293 if (!cpu_active(dest_cpu)) 1294 continue; 1295 goto out; 1296 } 1297 1298 switch (state) { 1299 case cpuset: 1300 /* No more Mr. Nice Guy. */ 1301 cpuset_cpus_allowed_fallback(p); 1302 state = possible; 1303 break; 1304 1305 case possible: 1306 do_set_cpus_allowed(p, cpu_possible_mask); 1307 state = fail; 1308 break; 1309 1310 case fail: 1311 BUG(); 1312 break; 1313 } 1314 } 1315 1316out: 1317 if (state != cpuset) { 1318 /* 1319 * Don't tell them about moving exiting tasks or 1320 * kernel threads (both mm NULL), since they never 1321 * leave kernel. 1322 */ 1323 if (p->mm && printk_ratelimit()) { 1324 printk_sched("process %d (%s) no longer affine to cpu%d\n", 1325 task_pid_nr(p), p->comm, cpu); 1326 } 1327 } 1328 1329 return dest_cpu; 1330} 1331 1332/* 1333 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable. 1334 */ 1335static inline 1336int select_task_rq(struct task_struct *p, int cpu, int sd_flags, int wake_flags) 1337{ 1338 cpu = p->sched_class->select_task_rq(p, cpu, sd_flags, wake_flags); 1339 1340 /* 1341 * In order not to call set_task_cpu() on a blocking task we need 1342 * to rely on ttwu() to place the task on a valid ->cpus_allowed 1343 * cpu. 1344 * 1345 * Since this is common to all placement strategies, this lives here. 1346 * 1347 * [ this allows ->select_task() to simply return task_cpu(p) and 1348 * not worry about this generic constraint ] 1349 */ 1350 if (unlikely(!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)) || 1351 !cpu_online(cpu))) 1352 cpu = select_fallback_rq(task_cpu(p), p); 1353 1354 return cpu; 1355} 1356 1357static void update_avg(u64 *avg, u64 sample) 1358{ 1359 s64 diff = sample - *avg; 1360 *avg += diff >> 3; 1361} 1362#endif 1363 1364static void 1365ttwu_stat(struct task_struct *p, int cpu, int wake_flags) 1366{ 1367#ifdef CONFIG_SCHEDSTATS 1368 struct rq *rq = this_rq(); 1369 1370#ifdef CONFIG_SMP 1371 int this_cpu = smp_processor_id(); 1372 1373 if (cpu == this_cpu) { 1374 schedstat_inc(rq, ttwu_local); 1375 schedstat_inc(p, se.statistics.nr_wakeups_local); 1376 } else { 1377 struct sched_domain *sd; 1378 1379 schedstat_inc(p, se.statistics.nr_wakeups_remote); 1380 rcu_read_lock(); 1381 for_each_domain(this_cpu, sd) { 1382 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) { 1383 schedstat_inc(sd, ttwu_wake_remote); 1384 break; 1385 } 1386 } 1387 rcu_read_unlock(); 1388 } 1389 1390 if (wake_flags & WF_MIGRATED) 1391 schedstat_inc(p, se.statistics.nr_wakeups_migrate); 1392 1393#endif /* CONFIG_SMP */ 1394 1395 schedstat_inc(rq, ttwu_count); 1396 schedstat_inc(p, se.statistics.nr_wakeups); 1397 1398 if (wake_flags & WF_SYNC) 1399 schedstat_inc(p, se.statistics.nr_wakeups_sync); 1400 1401#endif /* CONFIG_SCHEDSTATS */ 1402} 1403 1404static void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags) 1405{ 1406 activate_task(rq, p, en_flags); 1407 p->on_rq = 1; 1408 1409 /* if a worker is waking up, notify workqueue */ 1410 if (p->flags & PF_WQ_WORKER) 1411 wq_worker_waking_up(p, cpu_of(rq)); 1412} 1413 1414/* 1415 * Mark the task runnable and perform wakeup-preemption. 1416 */ 1417static void 1418ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags) 1419{ 1420 check_preempt_curr(rq, p, wake_flags); 1421 trace_sched_wakeup(p, true); 1422 1423 p->state = TASK_RUNNING; 1424#ifdef CONFIG_SMP 1425 if (p->sched_class->task_woken) 1426 p->sched_class->task_woken(rq, p); 1427 1428 if (rq->idle_stamp) { 1429 u64 delta = rq_clock(rq) - rq->idle_stamp; 1430 u64 max = 2*rq->max_idle_balance_cost; 1431 1432 update_avg(&rq->avg_idle, delta); 1433 1434 if (rq->avg_idle > max) 1435 rq->avg_idle = max; 1436 1437 rq->idle_stamp = 0; 1438 } 1439#endif 1440} 1441 1442static void 1443ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags) 1444{ 1445#ifdef CONFIG_SMP 1446 if (p->sched_contributes_to_load) 1447 rq->nr_uninterruptible--; 1448#endif 1449 1450 ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_WAKING); 1451 ttwu_do_wakeup(rq, p, wake_flags); 1452} 1453 1454/* 1455 * Called in case the task @p isn't fully descheduled from its runqueue, 1456 * in this case we must do a remote wakeup. Its a 'light' wakeup though, 1457 * since all we need to do is flip p->state to TASK_RUNNING, since 1458 * the task is still ->on_rq. 1459 */ 1460static int ttwu_remote(struct task_struct *p, int wake_flags) 1461{ 1462 struct rq *rq; 1463 int ret = 0; 1464 1465 rq = __task_rq_lock(p); 1466 if (p->on_rq) { 1467 /* check_preempt_curr() may use rq clock */ 1468 update_rq_clock(rq); 1469 ttwu_do_wakeup(rq, p, wake_flags); 1470 ret = 1; 1471 } 1472 __task_rq_unlock(rq); 1473 1474 return ret; 1475} 1476 1477#ifdef CONFIG_SMP 1478static void sched_ttwu_pending(void) 1479{ 1480 struct rq *rq = this_rq(); 1481 struct llist_node *llist = llist_del_all(&rq->wake_list); 1482 struct task_struct *p; 1483 1484 raw_spin_lock(&rq->lock); 1485 1486 while (llist) { 1487 p = llist_entry(llist, struct task_struct, wake_entry); 1488 llist = llist_next(llist); 1489 ttwu_do_activate(rq, p, 0); 1490 } 1491 1492 raw_spin_unlock(&rq->lock); 1493} 1494 1495void scheduler_ipi(void) 1496{ 1497 /* 1498 * Fold TIF_NEED_RESCHED into the preempt_count; anybody setting 1499 * TIF_NEED_RESCHED remotely (for the first time) will also send 1500 * this IPI. 1501 */ 1502 if (tif_need_resched()) 1503 set_preempt_need_resched(); 1504 1505 if (llist_empty(&this_rq()->wake_list) 1506 && !tick_nohz_full_cpu(smp_processor_id()) 1507 && !got_nohz_idle_kick()) 1508 return; 1509 1510 /* 1511 * Not all reschedule IPI handlers call irq_enter/irq_exit, since 1512 * traditionally all their work was done from the interrupt return 1513 * path. Now that we actually do some work, we need to make sure 1514 * we do call them. 1515 * 1516 * Some archs already do call them, luckily irq_enter/exit nest 1517 * properly. 1518 * 1519 * Arguably we should visit all archs and update all handlers, 1520 * however a fair share of IPIs are still resched only so this would 1521 * somewhat pessimize the simple resched case. 1522 */ 1523 irq_enter(); 1524 tick_nohz_full_check(); 1525 sched_ttwu_pending(); 1526 1527 /* 1528 * Check if someone kicked us for doing the nohz idle load balance. 1529 */ 1530 if (unlikely(got_nohz_idle_kick())) { 1531 this_rq()->idle_balance = 1; 1532 raise_softirq_irqoff(SCHED_SOFTIRQ); 1533 } 1534 irq_exit(); 1535} 1536 1537static void ttwu_queue_remote(struct task_struct *p, int cpu) 1538{ 1539 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list)) 1540 smp_send_reschedule(cpu); 1541} 1542 1543bool cpus_share_cache(int this_cpu, int that_cpu) 1544{ 1545 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu); 1546} 1547#endif /* CONFIG_SMP */ 1548 1549static void ttwu_queue(struct task_struct *p, int cpu) 1550{ 1551 struct rq *rq = cpu_rq(cpu); 1552 1553#if defined(CONFIG_SMP) 1554 if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) { 1555 sched_clock_cpu(cpu); /* sync clocks x-cpu */ 1556 ttwu_queue_remote(p, cpu); 1557 return; 1558 } 1559#endif 1560 1561 raw_spin_lock(&rq->lock); 1562 ttwu_do_activate(rq, p, 0); 1563 raw_spin_unlock(&rq->lock); 1564} 1565 1566/** 1567 * try_to_wake_up - wake up a thread 1568 * @p: the thread to be awakened 1569 * @state: the mask of task states that can be woken 1570 * @wake_flags: wake modifier flags (WF_*) 1571 * 1572 * Put it on the run-queue if it's not already there. The "current" 1573 * thread is always on the run-queue (except when the actual 1574 * re-schedule is in progress), and as such you're allowed to do 1575 * the simpler "current->state = TASK_RUNNING" to mark yourself 1576 * runnable without the overhead of this. 1577 * 1578 * Return: %true if @p was woken up, %false if it was already running. 1579 * or @state didn't match @p's state. 1580 */ 1581static int 1582try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags) 1583{ 1584 unsigned long flags; 1585 int cpu, success = 0; 1586 1587 /* 1588 * If we are going to wake up a thread waiting for CONDITION we 1589 * need to ensure that CONDITION=1 done by the caller can not be 1590 * reordered with p->state check below. This pairs with mb() in 1591 * set_current_state() the waiting thread does. 1592 */ 1593 smp_mb__before_spinlock(); 1594 raw_spin_lock_irqsave(&p->pi_lock, flags); 1595 if (!(p->state & state)) 1596 goto out; 1597 1598 success = 1; /* we're going to change ->state */ 1599 cpu = task_cpu(p); 1600 1601 if (p->on_rq && ttwu_remote(p, wake_flags)) 1602 goto stat; 1603 1604#ifdef CONFIG_SMP 1605 /* 1606 * If the owning (remote) cpu is still in the middle of schedule() with 1607 * this task as prev, wait until its done referencing the task. 1608 */ 1609 while (p->on_cpu) 1610 cpu_relax(); 1611 /* 1612 * Pairs with the smp_wmb() in finish_lock_switch(). 1613 */ 1614 smp_rmb(); 1615 1616 p->sched_contributes_to_load = !!task_contributes_to_load(p); 1617 p->state = TASK_WAKING; 1618 1619 if (p->sched_class->task_waking) 1620 p->sched_class->task_waking(p); 1621 1622 cpu = select_task_rq(p, p->wake_cpu, SD_BALANCE_WAKE, wake_flags); 1623 if (task_cpu(p) != cpu) { 1624 wake_flags |= WF_MIGRATED; 1625 set_task_cpu(p, cpu); 1626 } 1627#endif /* CONFIG_SMP */ 1628 1629 ttwu_queue(p, cpu); 1630stat: 1631 ttwu_stat(p, cpu, wake_flags); 1632out: 1633 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 1634 1635 return success; 1636} 1637 1638/** 1639 * try_to_wake_up_local - try to wake up a local task with rq lock held 1640 * @p: the thread to be awakened 1641 * 1642 * Put @p on the run-queue if it's not already there. The caller must 1643 * ensure that this_rq() is locked, @p is bound to this_rq() and not 1644 * the current task. 1645 */ 1646static void try_to_wake_up_local(struct task_struct *p) 1647{ 1648 struct rq *rq = task_rq(p); 1649 1650 if (WARN_ON_ONCE(rq != this_rq()) || 1651 WARN_ON_ONCE(p == current)) 1652 return; 1653 1654 lockdep_assert_held(&rq->lock); 1655 1656 if (!raw_spin_trylock(&p->pi_lock)) { 1657 raw_spin_unlock(&rq->lock); 1658 raw_spin_lock(&p->pi_lock); 1659 raw_spin_lock(&rq->lock); 1660 } 1661 1662 if (!(p->state & TASK_NORMAL)) 1663 goto out; 1664 1665 if (!p->on_rq) 1666 ttwu_activate(rq, p, ENQUEUE_WAKEUP); 1667 1668 ttwu_do_wakeup(rq, p, 0); 1669 ttwu_stat(p, smp_processor_id(), 0); 1670out: 1671 raw_spin_unlock(&p->pi_lock); 1672} 1673 1674/** 1675 * wake_up_process - Wake up a specific process 1676 * @p: The process to be woken up. 1677 * 1678 * Attempt to wake up the nominated process and move it to the set of runnable 1679 * processes. 1680 * 1681 * Return: 1 if the process was woken up, 0 if it was already running. 1682 * 1683 * It may be assumed that this function implies a write memory barrier before 1684 * changing the task state if and only if any tasks are woken up. 1685 */ 1686int wake_up_process(struct task_struct *p) 1687{ 1688 WARN_ON(task_is_stopped_or_traced(p)); 1689 return try_to_wake_up(p, TASK_NORMAL, 0); 1690} 1691EXPORT_SYMBOL(wake_up_process); 1692 1693int wake_up_state(struct task_struct *p, unsigned int state) 1694{ 1695 return try_to_wake_up(p, state, 0); 1696} 1697 1698/* 1699 * Perform scheduler related setup for a newly forked process p. 1700 * p is forked by current. 1701 * 1702 * __sched_fork() is basic setup used by init_idle() too: 1703 */ 1704static void __sched_fork(unsigned long clone_flags, struct task_struct *p) 1705{ 1706 p->on_rq = 0; 1707 1708 p->se.on_rq = 0; 1709 p->se.exec_start = 0; 1710 p->se.sum_exec_runtime = 0; 1711 p->se.prev_sum_exec_runtime = 0; 1712 p->se.nr_migrations = 0; 1713 p->se.vruntime = 0; 1714 INIT_LIST_HEAD(&p->se.group_node); 1715 1716#ifdef CONFIG_SCHEDSTATS 1717 memset(&p->se.statistics, 0, sizeof(p->se.statistics)); 1718#endif 1719 1720 INIT_LIST_HEAD(&p->rt.run_list); 1721 1722#ifdef CONFIG_PREEMPT_NOTIFIERS 1723 INIT_HLIST_HEAD(&p->preempt_notifiers); 1724#endif 1725 1726#ifdef CONFIG_NUMA_BALANCING 1727 if (p->mm && atomic_read(&p->mm->mm_users) == 1) { 1728 p->mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay); 1729 p->mm->numa_scan_seq = 0; 1730 } 1731 1732 if (clone_flags & CLONE_VM) 1733 p->numa_preferred_nid = current->numa_preferred_nid; 1734 else 1735 p->numa_preferred_nid = -1; 1736 1737 p->node_stamp = 0ULL; 1738 p->numa_scan_seq = p->mm ? p->mm->numa_scan_seq : 0; 1739 p->numa_scan_period = sysctl_numa_balancing_scan_delay; 1740 p->numa_work.next = &p->numa_work; 1741 p->numa_faults = NULL; 1742 p->numa_faults_buffer = NULL; 1743 1744 INIT_LIST_HEAD(&p->numa_entry); 1745 p->numa_group = NULL; 1746#endif /* CONFIG_NUMA_BALANCING */ 1747} 1748 1749#ifdef CONFIG_NUMA_BALANCING 1750#ifdef CONFIG_SCHED_DEBUG 1751void set_numabalancing_state(bool enabled) 1752{ 1753 if (enabled) 1754 sched_feat_set("NUMA"); 1755 else 1756 sched_feat_set("NO_NUMA"); 1757} 1758#else 1759__read_mostly bool numabalancing_enabled; 1760 1761void set_numabalancing_state(bool enabled) 1762{ 1763 numabalancing_enabled = enabled; 1764} 1765#endif /* CONFIG_SCHED_DEBUG */ 1766#endif /* CONFIG_NUMA_BALANCING */ 1767 1768/* 1769 * fork()/clone()-time setup: 1770 */ 1771void sched_fork(unsigned long clone_flags, struct task_struct *p) 1772{ 1773 unsigned long flags; 1774 int cpu = get_cpu(); 1775 1776 __sched_fork(clone_flags, p); 1777 /* 1778 * We mark the process as running here. This guarantees that 1779 * nobody will actually run it, and a signal or other external 1780 * event cannot wake it up and insert it on the runqueue either. 1781 */ 1782 p->state = TASK_RUNNING; 1783 1784 /* 1785 * Make sure we do not leak PI boosting priority to the child. 1786 */ 1787 p->prio = current->normal_prio; 1788 1789 /* 1790 * Revert to default priority/policy on fork if requested. 1791 */ 1792 if (unlikely(p->sched_reset_on_fork)) { 1793 if (task_has_rt_policy(p)) { 1794 p->policy = SCHED_NORMAL; 1795 p->static_prio = NICE_TO_PRIO(0); 1796 p->rt_priority = 0; 1797 } else if (PRIO_TO_NICE(p->static_prio) < 0) 1798 p->static_prio = NICE_TO_PRIO(0); 1799 1800 p->prio = p->normal_prio = __normal_prio(p); 1801 set_load_weight(p); 1802 1803 /* 1804 * We don't need the reset flag anymore after the fork. It has 1805 * fulfilled its duty: 1806 */ 1807 p->sched_reset_on_fork = 0; 1808 } 1809 1810 if (!rt_prio(p->prio)) 1811 p->sched_class = &fair_sched_class; 1812 1813 if (p->sched_class->task_fork) 1814 p->sched_class->task_fork(p); 1815 1816 /* 1817 * The child is not yet in the pid-hash so no cgroup attach races, 1818 * and the cgroup is pinned to this child due to cgroup_fork() 1819 * is ran before sched_fork(). 1820 * 1821 * Silence PROVE_RCU. 1822 */ 1823 raw_spin_lock_irqsave(&p->pi_lock, flags); 1824 set_task_cpu(p, cpu); 1825 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 1826 1827#if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT) 1828 if (likely(sched_info_on())) 1829 memset(&p->sched_info, 0, sizeof(p->sched_info)); 1830#endif 1831#if defined(CONFIG_SMP) 1832 p->on_cpu = 0; 1833#endif 1834 init_task_preempt_count(p); 1835#ifdef CONFIG_SMP 1836 plist_node_init(&p->pushable_tasks, MAX_PRIO); 1837#endif 1838 1839 put_cpu(); 1840} 1841 1842/* 1843 * wake_up_new_task - wake up a newly created task for the first time. 1844 * 1845 * This function will do some initial scheduler statistics housekeeping 1846 * that must be done for every newly created context, then puts the task 1847 * on the runqueue and wakes it. 1848 */ 1849void wake_up_new_task(struct task_struct *p) 1850{ 1851 unsigned long flags; 1852 struct rq *rq; 1853 1854 raw_spin_lock_irqsave(&p->pi_lock, flags); 1855#ifdef CONFIG_SMP 1856 /* 1857 * Fork balancing, do it here and not earlier because: 1858 * - cpus_allowed can change in the fork path 1859 * - any previously selected cpu might disappear through hotplug 1860 */ 1861 set_task_cpu(p, select_task_rq(p, task_cpu(p), SD_BALANCE_FORK, 0)); 1862#endif 1863 1864 /* Initialize new task's runnable average */ 1865 init_task_runnable_average(p); 1866 rq = __task_rq_lock(p); 1867 activate_task(rq, p, 0); 1868 p->on_rq = 1; 1869 trace_sched_wakeup_new(p, true); 1870 check_preempt_curr(rq, p, WF_FORK); 1871#ifdef CONFIG_SMP 1872 if (p->sched_class->task_woken) 1873 p->sched_class->task_woken(rq, p); 1874#endif 1875 task_rq_unlock(rq, p, &flags); 1876} 1877 1878#ifdef CONFIG_PREEMPT_NOTIFIERS 1879 1880/** 1881 * preempt_notifier_register - tell me when current is being preempted & rescheduled 1882 * @notifier: notifier struct to register 1883 */ 1884void preempt_notifier_register(struct preempt_notifier *notifier) 1885{ 1886 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers); 1887} 1888EXPORT_SYMBOL_GPL(preempt_notifier_register); 1889 1890/** 1891 * preempt_notifier_unregister - no longer interested in preemption notifications 1892 * @notifier: notifier struct to unregister 1893 * 1894 * This is safe to call from within a preemption notifier. 1895 */ 1896void preempt_notifier_unregister(struct preempt_notifier *notifier) 1897{ 1898 hlist_del(¬ifier->link); 1899} 1900EXPORT_SYMBOL_GPL(preempt_notifier_unregister); 1901 1902static void fire_sched_in_preempt_notifiers(struct task_struct *curr) 1903{ 1904 struct preempt_notifier *notifier; 1905 1906 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link) 1907 notifier->ops->sched_in(notifier, raw_smp_processor_id()); 1908} 1909 1910static void 1911fire_sched_out_preempt_notifiers(struct task_struct *curr, 1912 struct task_struct *next) 1913{ 1914 struct preempt_notifier *notifier; 1915 1916 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link) 1917 notifier->ops->sched_out(notifier, next); 1918} 1919 1920#else /* !CONFIG_PREEMPT_NOTIFIERS */ 1921 1922static void fire_sched_in_preempt_notifiers(struct task_struct *curr) 1923{ 1924} 1925 1926static void 1927fire_sched_out_preempt_notifiers(struct task_struct *curr, 1928 struct task_struct *next) 1929{ 1930} 1931 1932#endif /* CONFIG_PREEMPT_NOTIFIERS */ 1933 1934/** 1935 * prepare_task_switch - prepare to switch tasks 1936 * @rq: the runqueue preparing to switch 1937 * @prev: the current task that is being switched out 1938 * @next: the task we are going to switch to. 1939 * 1940 * This is called with the rq lock held and interrupts off. It must 1941 * be paired with a subsequent finish_task_switch after the context 1942 * switch. 1943 * 1944 * prepare_task_switch sets up locking and calls architecture specific 1945 * hooks. 1946 */ 1947static inline void 1948prepare_task_switch(struct rq *rq, struct task_struct *prev, 1949 struct task_struct *next) 1950{ 1951 trace_sched_switch(prev, next); 1952 sched_info_switch(rq, prev, next); 1953 perf_event_task_sched_out(prev, next); 1954 fire_sched_out_preempt_notifiers(prev, next); 1955 prepare_lock_switch(rq, next); 1956 prepare_arch_switch(next); 1957} 1958 1959/** 1960 * finish_task_switch - clean up after a task-switch 1961 * @rq: runqueue associated with task-switch 1962 * @prev: the thread we just switched away from. 1963 * 1964 * finish_task_switch must be called after the context switch, paired 1965 * with a prepare_task_switch call before the context switch. 1966 * finish_task_switch will reconcile locking set up by prepare_task_switch, 1967 * and do any other architecture-specific cleanup actions. 1968 * 1969 * Note that we may have delayed dropping an mm in context_switch(). If 1970 * so, we finish that here outside of the runqueue lock. (Doing it 1971 * with the lock held can cause deadlocks; see schedule() for 1972 * details.) 1973 */ 1974static void finish_task_switch(struct rq *rq, struct task_struct *prev) 1975 __releases(rq->lock) 1976{ 1977 struct mm_struct *mm = rq->prev_mm; 1978 long prev_state; 1979 1980 rq->prev_mm = NULL; 1981 1982 /* 1983 * A task struct has one reference for the use as "current". 1984 * If a task dies, then it sets TASK_DEAD in tsk->state and calls 1985 * schedule one last time. The schedule call will never return, and 1986 * the scheduled task must drop that reference. 1987 * The test for TASK_DEAD must occur while the runqueue locks are 1988 * still held, otherwise prev could be scheduled on another cpu, die 1989 * there before we look at prev->state, and then the reference would 1990 * be dropped twice. 1991 * Manfred Spraul <manfred@colorfullife.com> 1992 */ 1993 prev_state = prev->state; 1994 vtime_task_switch(prev); 1995 finish_arch_switch(prev); 1996 perf_event_task_sched_in(prev, current); 1997 finish_lock_switch(rq, prev); 1998 finish_arch_post_lock_switch(); 1999 2000 fire_sched_in_preempt_notifiers(current); 2001 if (mm) 2002 mmdrop(mm); 2003 if (unlikely(prev_state == TASK_DEAD)) { 2004 task_numa_free(prev); 2005 2006 /* 2007 * Remove function-return probe instances associated with this 2008 * task and put them back on the free list. 2009 */ 2010 kprobe_flush_task(prev); 2011 put_task_struct(prev); 2012 } 2013 2014 tick_nohz_task_switch(current); 2015} 2016 2017#ifdef CONFIG_SMP 2018 2019/* assumes rq->lock is held */ 2020static inline void pre_schedule(struct rq *rq, struct task_struct *prev) 2021{ 2022 if (prev->sched_class->pre_schedule) 2023 prev->sched_class->pre_schedule(rq, prev); 2024} 2025 2026/* rq->lock is NOT held, but preemption is disabled */ 2027static inline void post_schedule(struct rq *rq) 2028{ 2029 if (rq->post_schedule) { 2030 unsigned long flags; 2031 2032 raw_spin_lock_irqsave(&rq->lock, flags); 2033 if (rq->curr->sched_class->post_schedule) 2034 rq->curr->sched_class->post_schedule(rq); 2035 raw_spin_unlock_irqrestore(&rq->lock, flags); 2036 2037 rq->post_schedule = 0; 2038 } 2039} 2040 2041#else 2042 2043static inline void pre_schedule(struct rq *rq, struct task_struct *p) 2044{ 2045} 2046 2047static inline void post_schedule(struct rq *rq) 2048{ 2049} 2050 2051#endif 2052 2053/** 2054 * schedule_tail - first thing a freshly forked thread must call. 2055 * @prev: the thread we just switched away from. 2056 */ 2057asmlinkage void schedule_tail(struct task_struct *prev) 2058 __releases(rq->lock) 2059{ 2060 struct rq *rq = this_rq(); 2061 2062 finish_task_switch(rq, prev); 2063 2064 /* 2065 * FIXME: do we need to worry about rq being invalidated by the 2066 * task_switch? 2067 */ 2068 post_schedule(rq); 2069 2070#ifdef __ARCH_WANT_UNLOCKED_CTXSW 2071 /* In this case, finish_task_switch does not reenable preemption */ 2072 preempt_enable(); 2073#endif 2074 if (current->set_child_tid) 2075 put_user(task_pid_vnr(current), current->set_child_tid); 2076} 2077 2078/* 2079 * context_switch - switch to the new MM and the new 2080 * thread's register state. 2081 */ 2082static inline void 2083context_switch(struct rq *rq, struct task_struct *prev, 2084 struct task_struct *next) 2085{ 2086 struct mm_struct *mm, *oldmm; 2087 2088 prepare_task_switch(rq, prev, next); 2089 2090 mm = next->mm; 2091 oldmm = prev->active_mm; 2092 /* 2093 * For paravirt, this is coupled with an exit in switch_to to 2094 * combine the page table reload and the switch backend into 2095 * one hypercall. 2096 */ 2097 arch_start_context_switch(prev); 2098 2099 if (!mm) { 2100 next->active_mm = oldmm; 2101 atomic_inc(&oldmm->mm_count); 2102 enter_lazy_tlb(oldmm, next); 2103 } else 2104 switch_mm(oldmm, mm, next); 2105 2106 if (!prev->mm) { 2107 prev->active_mm = NULL; 2108 rq->prev_mm = oldmm; 2109 } 2110 /* 2111 * Since the runqueue lock will be released by the next 2112 * task (which is an invalid locking op but in the case 2113 * of the scheduler it's an obvious special-case), so we 2114 * do an early lockdep release here: 2115 */ 2116#ifndef __ARCH_WANT_UNLOCKED_CTXSW 2117 spin_release(&rq->lock.dep_map, 1, _THIS_IP_); 2118#endif 2119 2120 context_tracking_task_switch(prev, next); 2121 /* Here we just switch the register state and the stack. */ 2122 switch_to(prev, next, prev); 2123 2124 barrier(); 2125 /* 2126 * this_rq must be evaluated again because prev may have moved 2127 * CPUs since it called schedule(), thus the 'rq' on its stack 2128 * frame will be invalid. 2129 */ 2130 finish_task_switch(this_rq(), prev); 2131} 2132 2133/* 2134 * nr_running and nr_context_switches: 2135 * 2136 * externally visible scheduler statistics: current number of runnable 2137 * threads, total number of context switches performed since bootup. 2138 */ 2139unsigned long nr_running(void) 2140{ 2141 unsigned long i, sum = 0; 2142 2143 for_each_online_cpu(i) 2144 sum += cpu_rq(i)->nr_running; 2145 2146 return sum; 2147} 2148 2149unsigned long long nr_context_switches(void) 2150{ 2151 int i; 2152 unsigned long long sum = 0; 2153 2154 for_each_possible_cpu(i) 2155 sum += cpu_rq(i)->nr_switches; 2156 2157 return sum; 2158} 2159 2160unsigned long nr_iowait(void) 2161{ 2162 unsigned long i, sum = 0; 2163 2164 for_each_possible_cpu(i) 2165 sum += atomic_read(&cpu_rq(i)->nr_iowait); 2166 2167 return sum; 2168} 2169 2170unsigned long nr_iowait_cpu(int cpu) 2171{ 2172 struct rq *this = cpu_rq(cpu); 2173 return atomic_read(&this->nr_iowait); 2174} 2175 2176#ifdef CONFIG_SMP 2177 2178/* 2179 * sched_exec - execve() is a valuable balancing opportunity, because at 2180 * this point the task has the smallest effective memory and cache footprint. 2181 */ 2182void sched_exec(void) 2183{ 2184 struct task_struct *p = current; 2185 unsigned long flags; 2186 int dest_cpu; 2187 2188 raw_spin_lock_irqsave(&p->pi_lock, flags); 2189 dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), SD_BALANCE_EXEC, 0); 2190 if (dest_cpu == smp_processor_id()) 2191 goto unlock; 2192 2193 if (likely(cpu_active(dest_cpu))) { 2194 struct migration_arg arg = { p, dest_cpu }; 2195 2196 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 2197 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg); 2198 return; 2199 } 2200unlock: 2201 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 2202} 2203 2204#endif 2205 2206DEFINE_PER_CPU(struct kernel_stat, kstat); 2207DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat); 2208 2209EXPORT_PER_CPU_SYMBOL(kstat); 2210EXPORT_PER_CPU_SYMBOL(kernel_cpustat); 2211 2212/* 2213 * Return any ns on the sched_clock that have not yet been accounted in 2214 * @p in case that task is currently running. 2215 * 2216 * Called with task_rq_lock() held on @rq. 2217 */ 2218static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq) 2219{ 2220 u64 ns = 0; 2221 2222 if (task_current(rq, p)) { 2223 update_rq_clock(rq); 2224 ns = rq_clock_task(rq) - p->se.exec_start; 2225 if ((s64)ns < 0) 2226 ns = 0; 2227 } 2228 2229 return ns; 2230} 2231 2232unsigned long long task_delta_exec(struct task_struct *p) 2233{ 2234 unsigned long flags; 2235 struct rq *rq; 2236 u64 ns = 0; 2237 2238 rq = task_rq_lock(p, &flags); 2239 ns = do_task_delta_exec(p, rq); 2240 task_rq_unlock(rq, p, &flags); 2241 2242 return ns; 2243} 2244 2245/* 2246 * Return accounted runtime for the task. 2247 * In case the task is currently running, return the runtime plus current's 2248 * pending runtime that have not been accounted yet. 2249 */ 2250unsigned long long task_sched_runtime(struct task_struct *p) 2251{ 2252 unsigned long flags; 2253 struct rq *rq; 2254 u64 ns = 0; 2255 2256 rq = task_rq_lock(p, &flags); 2257 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq); 2258 task_rq_unlock(rq, p, &flags); 2259 2260 return ns; 2261} 2262 2263/* 2264 * This function gets called by the timer code, with HZ frequency. 2265 * We call it with interrupts disabled. 2266 */ 2267void scheduler_tick(void) 2268{ 2269 int cpu = smp_processor_id(); 2270 struct rq *rq = cpu_rq(cpu); 2271 struct task_struct *curr = rq->curr; 2272 2273 sched_clock_tick(); 2274 2275 raw_spin_lock(&rq->lock); 2276 update_rq_clock(rq); 2277 curr->sched_class->task_tick(rq, curr, 0); 2278 update_cpu_load_active(rq); 2279 raw_spin_unlock(&rq->lock); 2280 2281 perf_event_task_tick(); 2282 2283#ifdef CONFIG_SMP 2284 rq->idle_balance = idle_cpu(cpu); 2285 trigger_load_balance(rq, cpu); 2286#endif 2287 rq_last_tick_reset(rq); 2288} 2289 2290#ifdef CONFIG_NO_HZ_FULL 2291/** 2292 * scheduler_tick_max_deferment 2293 * 2294 * Keep at least one tick per second when a single 2295 * active task is running because the scheduler doesn't 2296 * yet completely support full dynticks environment. 2297 * 2298 * This makes sure that uptime, CFS vruntime, load 2299 * balancing, etc... continue to move forward, even 2300 * with a very low granularity. 2301 * 2302 * Return: Maximum deferment in nanoseconds. 2303 */ 2304u64 scheduler_tick_max_deferment(void) 2305{ 2306 struct rq *rq = this_rq(); 2307 unsigned long next, now = ACCESS_ONCE(jiffies); 2308 2309 next = rq->last_sched_tick + HZ; 2310 2311 if (time_before_eq(next, now)) 2312 return 0; 2313 2314 return jiffies_to_usecs(next - now) * NSEC_PER_USEC; 2315} 2316#endif 2317 2318notrace unsigned long get_parent_ip(unsigned long addr) 2319{ 2320 if (in_lock_functions(addr)) { 2321 addr = CALLER_ADDR2; 2322 if (in_lock_functions(addr)) 2323 addr = CALLER_ADDR3; 2324 } 2325 return addr; 2326} 2327 2328#if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \ 2329 defined(CONFIG_PREEMPT_TRACER)) 2330 2331void __kprobes preempt_count_add(int val) 2332{ 2333#ifdef CONFIG_DEBUG_PREEMPT 2334 /* 2335 * Underflow? 2336 */ 2337 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0))) 2338 return; 2339#endif 2340 __preempt_count_add(val); 2341#ifdef CONFIG_DEBUG_PREEMPT 2342 /* 2343 * Spinlock count overflowing soon? 2344 */ 2345 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >= 2346 PREEMPT_MASK - 10); 2347#endif 2348 if (preempt_count() == val) 2349 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1)); 2350} 2351EXPORT_SYMBOL(preempt_count_add); 2352 2353void __kprobes preempt_count_sub(int val) 2354{ 2355#ifdef CONFIG_DEBUG_PREEMPT 2356 /* 2357 * Underflow? 2358 */ 2359 if (DEBUG_LOCKS_WARN_ON(val > preempt_count())) 2360 return; 2361 /* 2362 * Is the spinlock portion underflowing? 2363 */ 2364 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) && 2365 !(preempt_count() & PREEMPT_MASK))) 2366 return; 2367#endif 2368 2369 if (preempt_count() == val) 2370 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1)); 2371 __preempt_count_sub(val); 2372} 2373EXPORT_SYMBOL(preempt_count_sub); 2374 2375#endif 2376 2377/* 2378 * Print scheduling while atomic bug: 2379 */ 2380static noinline void __schedule_bug(struct task_struct *prev) 2381{ 2382 if (oops_in_progress) 2383 return; 2384 2385 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n", 2386 prev->comm, prev->pid, preempt_count()); 2387 2388 debug_show_held_locks(prev); 2389 print_modules(); 2390 if (irqs_disabled()) 2391 print_irqtrace_events(prev); 2392 dump_stack(); 2393 add_taint(TAINT_WARN, LOCKDEP_STILL_OK); 2394} 2395 2396/* 2397 * Various schedule()-time debugging checks and statistics: 2398 */ 2399static inline void schedule_debug(struct task_struct *prev) 2400{ 2401 /* 2402 * Test if we are atomic. Since do_exit() needs to call into 2403 * schedule() atomically, we ignore that path for now. 2404 * Otherwise, whine if we are scheduling when we should not be. 2405 */ 2406 if (unlikely(in_atomic_preempt_off() && !prev->exit_state)) 2407 __schedule_bug(prev); 2408 rcu_sleep_check(); 2409 2410 profile_hit(SCHED_PROFILING, __builtin_return_address(0)); 2411 2412 schedstat_inc(this_rq(), sched_count); 2413} 2414 2415static void put_prev_task(struct rq *rq, struct task_struct *prev) 2416{ 2417 if (prev->on_rq || rq->skip_clock_update < 0) 2418 update_rq_clock(rq); 2419 prev->sched_class->put_prev_task(rq, prev); 2420} 2421 2422/* 2423 * Pick up the highest-prio task: 2424 */ 2425static inline struct task_struct * 2426pick_next_task(struct rq *rq) 2427{ 2428 const struct sched_class *class; 2429 struct task_struct *p; 2430 2431 /* 2432 * Optimization: we know that if all tasks are in 2433 * the fair class we can call that function directly: 2434 */ 2435 if (likely(rq->nr_running == rq->cfs.h_nr_running)) { 2436 p = fair_sched_class.pick_next_task(rq); 2437 if (likely(p)) 2438 return p; 2439 } 2440 2441 for_each_class(class) { 2442 p = class->pick_next_task(rq); 2443 if (p) 2444 return p; 2445 } 2446 2447 BUG(); /* the idle class will always have a runnable task */ 2448} 2449 2450/* 2451 * __schedule() is the main scheduler function. 2452 * 2453 * The main means of driving the scheduler and thus entering this function are: 2454 * 2455 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc. 2456 * 2457 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return 2458 * paths. For example, see arch/x86/entry_64.S. 2459 * 2460 * To drive preemption between tasks, the scheduler sets the flag in timer 2461 * interrupt handler scheduler_tick(). 2462 * 2463 * 3. Wakeups don't really cause entry into schedule(). They add a 2464 * task to the run-queue and that's it. 2465 * 2466 * Now, if the new task added to the run-queue preempts the current 2467 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets 2468 * called on the nearest possible occasion: 2469 * 2470 * - If the kernel is preemptible (CONFIG_PREEMPT=y): 2471 * 2472 * - in syscall or exception context, at the next outmost 2473 * preempt_enable(). (this might be as soon as the wake_up()'s 2474 * spin_unlock()!) 2475 * 2476 * - in IRQ context, return from interrupt-handler to 2477 * preemptible context 2478 * 2479 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set) 2480 * then at the next: 2481 * 2482 * - cond_resched() call 2483 * - explicit schedule() call 2484 * - return from syscall or exception to user-space 2485 * - return from interrupt-handler to user-space 2486 */ 2487static void __sched __schedule(void) 2488{ 2489 struct task_struct *prev, *next; 2490 unsigned long *switch_count; 2491 struct rq *rq; 2492 int cpu; 2493 2494need_resched: 2495 preempt_disable(); 2496 cpu = smp_processor_id(); 2497 rq = cpu_rq(cpu); 2498 rcu_note_context_switch(cpu); 2499 prev = rq->curr; 2500 2501 schedule_debug(prev); 2502 2503 if (sched_feat(HRTICK)) 2504 hrtick_clear(rq); 2505 2506 /* 2507 * Make sure that signal_pending_state()->signal_pending() below 2508 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE) 2509 * done by the caller to avoid the race with signal_wake_up(). 2510 */ 2511 smp_mb__before_spinlock(); 2512 raw_spin_lock_irq(&rq->lock); 2513 2514 switch_count = &prev->nivcsw; 2515 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) { 2516 if (unlikely(signal_pending_state(prev->state, prev))) { 2517 prev->state = TASK_RUNNING; 2518 } else { 2519 deactivate_task(rq, prev, DEQUEUE_SLEEP); 2520 prev->on_rq = 0; 2521 2522 /* 2523 * If a worker went to sleep, notify and ask workqueue 2524 * whether it wants to wake up a task to maintain 2525 * concurrency. 2526 */ 2527 if (prev->flags & PF_WQ_WORKER) { 2528 struct task_struct *to_wakeup; 2529 2530 to_wakeup = wq_worker_sleeping(prev, cpu); 2531 if (to_wakeup) 2532 try_to_wake_up_local(to_wakeup); 2533 } 2534 } 2535 switch_count = &prev->nvcsw; 2536 } 2537 2538 pre_schedule(rq, prev); 2539 2540 if (unlikely(!rq->nr_running)) 2541 idle_balance(cpu, rq); 2542 2543 put_prev_task(rq, prev); 2544 next = pick_next_task(rq); 2545 clear_tsk_need_resched(prev); 2546 clear_preempt_need_resched(); 2547 rq->skip_clock_update = 0; 2548 2549 if (likely(prev != next)) { 2550 rq->nr_switches++; 2551 rq->curr = next; 2552 ++*switch_count; 2553 2554 context_switch(rq, prev, next); /* unlocks the rq */ 2555 /* 2556 * The context switch have flipped the stack from under us 2557 * and restored the local variables which were saved when 2558 * this task called schedule() in the past. prev == current 2559 * is still correct, but it can be moved to another cpu/rq. 2560 */ 2561 cpu = smp_processor_id(); 2562 rq = cpu_rq(cpu); 2563 } else 2564 raw_spin_unlock_irq(&rq->lock); 2565 2566 post_schedule(rq); 2567 2568 sched_preempt_enable_no_resched(); 2569 if (need_resched()) 2570 goto need_resched; 2571} 2572 2573static inline void sched_submit_work(struct task_struct *tsk) 2574{ 2575 if (!tsk->state || tsk_is_pi_blocked(tsk)) 2576 return; 2577 /* 2578 * If we are going to sleep and we have plugged IO queued, 2579 * make sure to submit it to avoid deadlocks. 2580 */ 2581 if (blk_needs_flush_plug(tsk)) 2582 blk_schedule_flush_plug(tsk); 2583} 2584 2585asmlinkage void __sched schedule(void) 2586{ 2587 struct task_struct *tsk = current; 2588 2589 sched_submit_work(tsk); 2590 __schedule(); 2591} 2592EXPORT_SYMBOL(schedule); 2593 2594#ifdef CONFIG_CONTEXT_TRACKING 2595asmlinkage void __sched schedule_user(void) 2596{ 2597 /* 2598 * If we come here after a random call to set_need_resched(), 2599 * or we have been woken up remotely but the IPI has not yet arrived, 2600 * we haven't yet exited the RCU idle mode. Do it here manually until 2601 * we find a better solution. 2602 */ 2603 user_exit(); 2604 schedule(); 2605 user_enter(); 2606} 2607#endif 2608 2609/** 2610 * schedule_preempt_disabled - called with preemption disabled 2611 * 2612 * Returns with preemption disabled. Note: preempt_count must be 1 2613 */ 2614void __sched schedule_preempt_disabled(void) 2615{ 2616 sched_preempt_enable_no_resched(); 2617 schedule(); 2618 preempt_disable(); 2619} 2620 2621#ifdef CONFIG_PREEMPT 2622/* 2623 * this is the entry point to schedule() from in-kernel preemption 2624 * off of preempt_enable. Kernel preemptions off return from interrupt 2625 * occur there and call schedule directly. 2626 */ 2627asmlinkage void __sched notrace preempt_schedule(void) 2628{ 2629 /* 2630 * If there is a non-zero preempt_count or interrupts are disabled, 2631 * we do not want to preempt the current task. Just return.. 2632 */ 2633 if (likely(!preemptible())) 2634 return; 2635 2636 do { 2637 __preempt_count_add(PREEMPT_ACTIVE); 2638 __schedule(); 2639 __preempt_count_sub(PREEMPT_ACTIVE); 2640 2641 /* 2642 * Check again in case we missed a preemption opportunity 2643 * between schedule and now. 2644 */ 2645 barrier(); 2646 } while (need_resched()); 2647} 2648EXPORT_SYMBOL(preempt_schedule); 2649 2650/* 2651 * this is the entry point to schedule() from kernel preemption 2652 * off of irq context. 2653 * Note, that this is called and return with irqs disabled. This will 2654 * protect us against recursive calling from irq. 2655 */ 2656asmlinkage void __sched preempt_schedule_irq(void) 2657{ 2658 enum ctx_state prev_state; 2659 2660 /* Catch callers which need to be fixed */ 2661 BUG_ON(preempt_count() || !irqs_disabled()); 2662 2663 prev_state = exception_enter(); 2664 2665 do { 2666 __preempt_count_add(PREEMPT_ACTIVE); 2667 local_irq_enable(); 2668 __schedule(); 2669 local_irq_disable(); 2670 __preempt_count_sub(PREEMPT_ACTIVE); 2671 2672 /* 2673 * Check again in case we missed a preemption opportunity 2674 * between schedule and now. 2675 */ 2676 barrier(); 2677 } while (need_resched()); 2678 2679 exception_exit(prev_state); 2680} 2681 2682#endif /* CONFIG_PREEMPT */ 2683 2684int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags, 2685 void *key) 2686{ 2687 return try_to_wake_up(curr->private, mode, wake_flags); 2688} 2689EXPORT_SYMBOL(default_wake_function); 2690 2691static long __sched 2692sleep_on_common(wait_queue_head_t *q, int state, long timeout) 2693{ 2694 unsigned long flags; 2695 wait_queue_t wait; 2696 2697 init_waitqueue_entry(&wait, current); 2698 2699 __set_current_state(state); 2700 2701 spin_lock_irqsave(&q->lock, flags); 2702 __add_wait_queue(q, &wait); 2703 spin_unlock(&q->lock); 2704 timeout = schedule_timeout(timeout); 2705 spin_lock_irq(&q->lock); 2706 __remove_wait_queue(q, &wait); 2707 spin_unlock_irqrestore(&q->lock, flags); 2708 2709 return timeout; 2710} 2711 2712void __sched interruptible_sleep_on(wait_queue_head_t *q) 2713{ 2714 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT); 2715} 2716EXPORT_SYMBOL(interruptible_sleep_on); 2717 2718long __sched 2719interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout) 2720{ 2721 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout); 2722} 2723EXPORT_SYMBOL(interruptible_sleep_on_timeout); 2724 2725void __sched sleep_on(wait_queue_head_t *q) 2726{ 2727 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT); 2728} 2729EXPORT_SYMBOL(sleep_on); 2730 2731long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout) 2732{ 2733 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout); 2734} 2735EXPORT_SYMBOL(sleep_on_timeout); 2736 2737#ifdef CONFIG_RT_MUTEXES 2738 2739/* 2740 * rt_mutex_setprio - set the current priority of a task 2741 * @p: task 2742 * @prio: prio value (kernel-internal form) 2743 * 2744 * This function changes the 'effective' priority of a task. It does 2745 * not touch ->normal_prio like __setscheduler(). 2746 * 2747 * Used by the rt_mutex code to implement priority inheritance logic. 2748 */ 2749void rt_mutex_setprio(struct task_struct *p, int prio) 2750{ 2751 int oldprio, on_rq, running; 2752 struct rq *rq; 2753 const struct sched_class *prev_class; 2754 2755 BUG_ON(prio < 0 || prio > MAX_PRIO); 2756 2757 rq = __task_rq_lock(p); 2758 2759 /* 2760 * Idle task boosting is a nono in general. There is one 2761 * exception, when PREEMPT_RT and NOHZ is active: 2762 * 2763 * The idle task calls get_next_timer_interrupt() and holds 2764 * the timer wheel base->lock on the CPU and another CPU wants 2765 * to access the timer (probably to cancel it). We can safely 2766 * ignore the boosting request, as the idle CPU runs this code 2767 * with interrupts disabled and will complete the lock 2768 * protected section without being interrupted. So there is no 2769 * real need to boost. 2770 */ 2771 if (unlikely(p == rq->idle)) { 2772 WARN_ON(p != rq->curr); 2773 WARN_ON(p->pi_blocked_on); 2774 goto out_unlock; 2775 } 2776 2777 trace_sched_pi_setprio(p, prio); 2778 oldprio = p->prio; 2779 prev_class = p->sched_class; 2780 on_rq = p->on_rq; 2781 running = task_current(rq, p); 2782 if (on_rq) 2783 dequeue_task(rq, p, 0); 2784 if (running) 2785 p->sched_class->put_prev_task(rq, p); 2786 2787 if (rt_prio(prio)) 2788 p->sched_class = &rt_sched_class; 2789 else 2790 p->sched_class = &fair_sched_class; 2791 2792 p->prio = prio; 2793 2794 if (running) 2795 p->sched_class->set_curr_task(rq); 2796 if (on_rq) 2797 enqueue_task(rq, p, oldprio < prio ? ENQUEUE_HEAD : 0); 2798 2799 check_class_changed(rq, p, prev_class, oldprio); 2800out_unlock: 2801 __task_rq_unlock(rq); 2802} 2803#endif 2804void set_user_nice(struct task_struct *p, long nice) 2805{ 2806 int old_prio, delta, on_rq; 2807 unsigned long flags; 2808 struct rq *rq; 2809 2810 if (TASK_NICE(p) == nice || nice < -20 || nice > 19) 2811 return; 2812 /* 2813 * We have to be careful, if called from sys_setpriority(), 2814 * the task might be in the middle of scheduling on another CPU. 2815 */ 2816 rq = task_rq_lock(p, &flags); 2817 /* 2818 * The RT priorities are set via sched_setscheduler(), but we still 2819 * allow the 'normal' nice value to be set - but as expected 2820 * it wont have any effect on scheduling until the task is 2821 * SCHED_FIFO/SCHED_RR: 2822 */ 2823 if (task_has_rt_policy(p)) { 2824 p->static_prio = NICE_TO_PRIO(nice); 2825 goto out_unlock; 2826 } 2827 on_rq = p->on_rq; 2828 if (on_rq) 2829 dequeue_task(rq, p, 0); 2830 2831 p->static_prio = NICE_TO_PRIO(nice); 2832 set_load_weight(p); 2833 old_prio = p->prio; 2834 p->prio = effective_prio(p); 2835 delta = p->prio - old_prio; 2836 2837 if (on_rq) { 2838 enqueue_task(rq, p, 0); 2839 /* 2840 * If the task increased its priority or is running and 2841 * lowered its priority, then reschedule its CPU: 2842 */ 2843 if (delta < 0 || (delta > 0 && task_running(rq, p))) 2844 resched_task(rq->curr); 2845 } 2846out_unlock: 2847 task_rq_unlock(rq, p, &flags); 2848} 2849EXPORT_SYMBOL(set_user_nice); 2850 2851/* 2852 * can_nice - check if a task can reduce its nice value 2853 * @p: task 2854 * @nice: nice value 2855 */ 2856int can_nice(const struct task_struct *p, const int nice) 2857{ 2858 /* convert nice value [19,-20] to rlimit style value [1,40] */ 2859 int nice_rlim = 20 - nice; 2860 2861 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) || 2862 capable(CAP_SYS_NICE)); 2863} 2864 2865#ifdef __ARCH_WANT_SYS_NICE 2866 2867/* 2868 * sys_nice - change the priority of the current process. 2869 * @increment: priority increment 2870 * 2871 * sys_setpriority is a more generic, but much slower function that 2872 * does similar things. 2873 */ 2874SYSCALL_DEFINE1(nice, int, increment) 2875{ 2876 long nice, retval; 2877 2878 /* 2879 * Setpriority might change our priority at the same moment. 2880 * We don't have to worry. Conceptually one call occurs first 2881 * and we have a single winner. 2882 */ 2883 if (increment < -40) 2884 increment = -40; 2885 if (increment > 40) 2886 increment = 40; 2887 2888 nice = TASK_NICE(current) + increment; 2889 if (nice < -20) 2890 nice = -20; 2891 if (nice > 19) 2892 nice = 19; 2893 2894 if (increment < 0 && !can_nice(current, nice)) 2895 return -EPERM; 2896 2897 retval = security_task_setnice(current, nice); 2898 if (retval) 2899 return retval; 2900 2901 set_user_nice(current, nice); 2902 return 0; 2903} 2904 2905#endif 2906 2907/** 2908 * task_prio - return the priority value of a given task. 2909 * @p: the task in question. 2910 * 2911 * Return: The priority value as seen by users in /proc. 2912 * RT tasks are offset by -200. Normal tasks are centered 2913 * around 0, value goes from -16 to +15. 2914 */ 2915int task_prio(const struct task_struct *p) 2916{ 2917 return p->prio - MAX_RT_PRIO; 2918} 2919 2920/** 2921 * task_nice - return the nice value of a given task. 2922 * @p: the task in question. 2923 * 2924 * Return: The nice value [ -20 ... 0 ... 19 ]. 2925 */ 2926int task_nice(const struct task_struct *p) 2927{ 2928 return TASK_NICE(p); 2929} 2930EXPORT_SYMBOL(task_nice); 2931 2932/** 2933 * idle_cpu - is a given cpu idle currently? 2934 * @cpu: the processor in question. 2935 * 2936 * Return: 1 if the CPU is currently idle. 0 otherwise. 2937 */ 2938int idle_cpu(int cpu) 2939{ 2940 struct rq *rq = cpu_rq(cpu); 2941 2942 if (rq->curr != rq->idle) 2943 return 0; 2944 2945 if (rq->nr_running) 2946 return 0; 2947 2948#ifdef CONFIG_SMP 2949 if (!llist_empty(&rq->wake_list)) 2950 return 0; 2951#endif 2952 2953 return 1; 2954} 2955 2956/** 2957 * idle_task - return the idle task for a given cpu. 2958 * @cpu: the processor in question. 2959 * 2960 * Return: The idle task for the cpu @cpu. 2961 */ 2962struct task_struct *idle_task(int cpu) 2963{ 2964 return cpu_rq(cpu)->idle; 2965} 2966 2967/** 2968 * find_process_by_pid - find a process with a matching PID value. 2969 * @pid: the pid in question. 2970 * 2971 * The task of @pid, if found. %NULL otherwise. 2972 */ 2973static struct task_struct *find_process_by_pid(pid_t pid) 2974{ 2975 return pid ? find_task_by_vpid(pid) : current; 2976} 2977 2978/* Actually do priority change: must hold rq lock. */ 2979static void 2980__setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio) 2981{ 2982 p->policy = policy; 2983 p->rt_priority = prio; 2984 p->normal_prio = normal_prio(p); 2985 /* we are holding p->pi_lock already */ 2986 p->prio = rt_mutex_getprio(p); 2987 if (rt_prio(p->prio)) 2988 p->sched_class = &rt_sched_class; 2989 else 2990 p->sched_class = &fair_sched_class; 2991 set_load_weight(p); 2992} 2993 2994/* 2995 * check the target process has a UID that matches the current process's 2996 */ 2997static bool check_same_owner(struct task_struct *p) 2998{ 2999 const struct cred *cred = current_cred(), *pcred; 3000 bool match; 3001 3002 rcu_read_lock(); 3003 pcred = __task_cred(p); 3004 match = (uid_eq(cred->euid, pcred->euid) || 3005 uid_eq(cred->euid, pcred->uid)); 3006 rcu_read_unlock(); 3007 return match; 3008} 3009 3010static int __sched_setscheduler(struct task_struct *p, int policy, 3011 const struct sched_param *param, bool user) 3012{ 3013 int retval, oldprio, oldpolicy = -1, on_rq, running; 3014 unsigned long flags; 3015 const struct sched_class *prev_class; 3016 struct rq *rq; 3017 int reset_on_fork; 3018 3019 /* may grab non-irq protected spin_locks */ 3020 BUG_ON(in_interrupt()); 3021recheck: 3022 /* double check policy once rq lock held */ 3023 if (policy < 0) { 3024 reset_on_fork = p->sched_reset_on_fork; 3025 policy = oldpolicy = p->policy; 3026 } else { 3027 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK); 3028 policy &= ~SCHED_RESET_ON_FORK; 3029 3030 if (policy != SCHED_FIFO && policy != SCHED_RR && 3031 policy != SCHED_NORMAL && policy != SCHED_BATCH && 3032 policy != SCHED_IDLE) 3033 return -EINVAL; 3034 } 3035 3036 /* 3037 * Valid priorities for SCHED_FIFO and SCHED_RR are 3038 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL, 3039 * SCHED_BATCH and SCHED_IDLE is 0. 3040 */ 3041 if (param->sched_priority < 0 || 3042 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) || 3043 (!p->mm && param->sched_priority > MAX_RT_PRIO-1)) 3044 return -EINVAL; 3045 if (rt_policy(policy) != (param->sched_priority != 0)) 3046 return -EINVAL; 3047 3048 /* 3049 * Allow unprivileged RT tasks to decrease priority: 3050 */ 3051 if (user && !capable(CAP_SYS_NICE)) { 3052 if (rt_policy(policy)) { 3053 unsigned long rlim_rtprio = 3054 task_rlimit(p, RLIMIT_RTPRIO); 3055 3056 /* can't set/change the rt policy */ 3057 if (policy != p->policy && !rlim_rtprio) 3058 return -EPERM; 3059 3060 /* can't increase priority */ 3061 if (param->sched_priority > p->rt_priority && 3062 param->sched_priority > rlim_rtprio) 3063 return -EPERM; 3064 } 3065 3066 /* 3067 * Treat SCHED_IDLE as nice 20. Only allow a switch to 3068 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it. 3069 */ 3070 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) { 3071 if (!can_nice(p, TASK_NICE(p))) 3072 return -EPERM; 3073 } 3074 3075 /* can't change other user's priorities */ 3076 if (!check_same_owner(p)) 3077 return -EPERM; 3078 3079 /* Normal users shall not reset the sched_reset_on_fork flag */ 3080 if (p->sched_reset_on_fork && !reset_on_fork) 3081 return -EPERM; 3082 } 3083 3084 if (user) { 3085 retval = security_task_setscheduler(p); 3086 if (retval) 3087 return retval; 3088 } 3089 3090 /* 3091 * make sure no PI-waiters arrive (or leave) while we are 3092 * changing the priority of the task: 3093 * 3094 * To be able to change p->policy safely, the appropriate 3095 * runqueue lock must be held. 3096 */ 3097 rq = task_rq_lock(p, &flags); 3098 3099 /* 3100 * Changing the policy of the stop threads its a very bad idea 3101 */ 3102 if (p == rq->stop) { 3103 task_rq_unlock(rq, p, &flags); 3104 return -EINVAL; 3105 } 3106 3107 /* 3108 * If not changing anything there's no need to proceed further: 3109 */ 3110 if (unlikely(policy == p->policy && (!rt_policy(policy) || 3111 param->sched_priority == p->rt_priority))) { 3112 task_rq_unlock(rq, p, &flags); 3113 return 0; 3114 } 3115 3116#ifdef CONFIG_RT_GROUP_SCHED 3117 if (user) { 3118 /* 3119 * Do not allow realtime tasks into groups that have no runtime 3120 * assigned. 3121 */ 3122 if (rt_bandwidth_enabled() && rt_policy(policy) && 3123 task_group(p)->rt_bandwidth.rt_runtime == 0 && 3124 !task_group_is_autogroup(task_group(p))) { 3125 task_rq_unlock(rq, p, &flags); 3126 return -EPERM; 3127 } 3128 } 3129#endif 3130 3131 /* recheck policy now with rq lock held */ 3132 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) { 3133 policy = oldpolicy = -1; 3134 task_rq_unlock(rq, p, &flags); 3135 goto recheck; 3136 } 3137 on_rq = p->on_rq; 3138 running = task_current(rq, p); 3139 if (on_rq) 3140 dequeue_task(rq, p, 0); 3141 if (running) 3142 p->sched_class->put_prev_task(rq, p); 3143 3144 p->sched_reset_on_fork = reset_on_fork; 3145 3146 oldprio = p->prio; 3147 prev_class = p->sched_class; 3148 __setscheduler(rq, p, policy, param->sched_priority); 3149 3150 if (running) 3151 p->sched_class->set_curr_task(rq); 3152 if (on_rq) 3153 enqueue_task(rq, p, 0); 3154 3155 check_class_changed(rq, p, prev_class, oldprio); 3156 task_rq_unlock(rq, p, &flags); 3157 3158 rt_mutex_adjust_pi(p); 3159 3160 return 0; 3161} 3162 3163/** 3164 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread. 3165 * @p: the task in question. 3166 * @policy: new policy. 3167 * @param: structure containing the new RT priority. 3168 * 3169 * Return: 0 on success. An error code otherwise. 3170 * 3171 * NOTE that the task may be already dead. 3172 */ 3173int sched_setscheduler(struct task_struct *p, int policy, 3174 const struct sched_param *param) 3175{ 3176 return __sched_setscheduler(p, policy, param, true); 3177} 3178EXPORT_SYMBOL_GPL(sched_setscheduler); 3179 3180/** 3181 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace. 3182 * @p: the task in question. 3183 * @policy: new policy. 3184 * @param: structure containing the new RT priority. 3185 * 3186 * Just like sched_setscheduler, only don't bother checking if the 3187 * current context has permission. For example, this is needed in 3188 * stop_machine(): we create temporary high priority worker threads, 3189 * but our caller might not have that capability. 3190 * 3191 * Return: 0 on success. An error code otherwise. 3192 */ 3193int sched_setscheduler_nocheck(struct task_struct *p, int policy, 3194 const struct sched_param *param) 3195{ 3196 return __sched_setscheduler(p, policy, param, false); 3197} 3198 3199static int 3200do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param) 3201{ 3202 struct sched_param lparam; 3203 struct task_struct *p; 3204 int retval; 3205 3206 if (!param || pid < 0) 3207 return -EINVAL; 3208 if (copy_from_user(&lparam, param, sizeof(struct sched_param))) 3209 return -EFAULT; 3210 3211 rcu_read_lock(); 3212 retval = -ESRCH; 3213 p = find_process_by_pid(pid); 3214 if (p != NULL) 3215 retval = sched_setscheduler(p, policy, &lparam); 3216 rcu_read_unlock(); 3217 3218 return retval; 3219} 3220 3221/** 3222 * sys_sched_setscheduler - set/change the scheduler policy and RT priority 3223 * @pid: the pid in question. 3224 * @policy: new policy. 3225 * @param: structure containing the new RT priority. 3226 * 3227 * Return: 0 on success. An error code otherwise. 3228 */ 3229SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy, 3230 struct sched_param __user *, param) 3231{ 3232 /* negative values for policy are not valid */ 3233 if (policy < 0) 3234 return -EINVAL; 3235 3236 return do_sched_setscheduler(pid, policy, param); 3237} 3238 3239/** 3240 * sys_sched_setparam - set/change the RT priority of a thread 3241 * @pid: the pid in question. 3242 * @param: structure containing the new RT priority. 3243 * 3244 * Return: 0 on success. An error code otherwise. 3245 */ 3246SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param) 3247{ 3248 return do_sched_setscheduler(pid, -1, param); 3249} 3250 3251/** 3252 * sys_sched_getscheduler - get the policy (scheduling class) of a thread 3253 * @pid: the pid in question. 3254 * 3255 * Return: On success, the policy of the thread. Otherwise, a negative error 3256 * code. 3257 */ 3258SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid) 3259{ 3260 struct task_struct *p; 3261 int retval; 3262 3263 if (pid < 0) 3264 return -EINVAL; 3265 3266 retval = -ESRCH; 3267 rcu_read_lock(); 3268 p = find_process_by_pid(pid); 3269 if (p) { 3270 retval = security_task_getscheduler(p); 3271 if (!retval) 3272 retval = p->policy 3273 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0); 3274 } 3275 rcu_read_unlock(); 3276 return retval; 3277} 3278 3279/** 3280 * sys_sched_getparam - get the RT priority of a thread 3281 * @pid: the pid in question. 3282 * @param: structure containing the RT priority. 3283 * 3284 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error 3285 * code. 3286 */ 3287SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param) 3288{ 3289 struct sched_param lp; 3290 struct task_struct *p; 3291 int retval; 3292 3293 if (!param || pid < 0) 3294 return -EINVAL; 3295 3296 rcu_read_lock(); 3297 p = find_process_by_pid(pid); 3298 retval = -ESRCH; 3299 if (!p) 3300 goto out_unlock; 3301 3302 retval = security_task_getscheduler(p); 3303 if (retval) 3304 goto out_unlock; 3305 3306 lp.sched_priority = p->rt_priority; 3307 rcu_read_unlock(); 3308 3309 /* 3310 * This one might sleep, we cannot do it with a spinlock held ... 3311 */ 3312 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0; 3313 3314 return retval; 3315 3316out_unlock: 3317 rcu_read_unlock(); 3318 return retval; 3319} 3320 3321long sched_setaffinity(pid_t pid, const struct cpumask *in_mask) 3322{ 3323 cpumask_var_t cpus_allowed, new_mask; 3324 struct task_struct *p; 3325 int retval; 3326 3327 rcu_read_lock(); 3328 3329 p = find_process_by_pid(pid); 3330 if (!p) { 3331 rcu_read_unlock(); 3332 return -ESRCH; 3333 } 3334 3335 /* Prevent p going away */ 3336 get_task_struct(p); 3337 rcu_read_unlock(); 3338 3339 if (p->flags & PF_NO_SETAFFINITY) { 3340 retval = -EINVAL; 3341 goto out_put_task; 3342 } 3343 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) { 3344 retval = -ENOMEM; 3345 goto out_put_task; 3346 } 3347 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) { 3348 retval = -ENOMEM; 3349 goto out_free_cpus_allowed; 3350 } 3351 retval = -EPERM; 3352 if (!check_same_owner(p)) { 3353 rcu_read_lock(); 3354 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) { 3355 rcu_read_unlock(); 3356 goto out_unlock; 3357 } 3358 rcu_read_unlock(); 3359 } 3360 3361 retval = security_task_setscheduler(p); 3362 if (retval) 3363 goto out_unlock; 3364 3365 cpuset_cpus_allowed(p, cpus_allowed); 3366 cpumask_and(new_mask, in_mask, cpus_allowed); 3367again: 3368 retval = set_cpus_allowed_ptr(p, new_mask); 3369 3370 if (!retval) { 3371 cpuset_cpus_allowed(p, cpus_allowed); 3372 if (!cpumask_subset(new_mask, cpus_allowed)) { 3373 /* 3374 * We must have raced with a concurrent cpuset 3375 * update. Just reset the cpus_allowed to the 3376 * cpuset's cpus_allowed 3377 */ 3378 cpumask_copy(new_mask, cpus_allowed); 3379 goto again; 3380 } 3381 } 3382out_unlock: 3383 free_cpumask_var(new_mask); 3384out_free_cpus_allowed: 3385 free_cpumask_var(cpus_allowed); 3386out_put_task: 3387 put_task_struct(p); 3388 return retval; 3389} 3390 3391static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len, 3392 struct cpumask *new_mask) 3393{ 3394 if (len < cpumask_size()) 3395 cpumask_clear(new_mask); 3396 else if (len > cpumask_size()) 3397 len = cpumask_size(); 3398 3399 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0; 3400} 3401 3402/** 3403 * sys_sched_setaffinity - set the cpu affinity of a process 3404 * @pid: pid of the process 3405 * @len: length in bytes of the bitmask pointed to by user_mask_ptr 3406 * @user_mask_ptr: user-space pointer to the new cpu mask 3407 * 3408 * Return: 0 on success. An error code otherwise. 3409 */ 3410SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len, 3411 unsigned long __user *, user_mask_ptr) 3412{ 3413 cpumask_var_t new_mask; 3414 int retval; 3415 3416 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) 3417 return -ENOMEM; 3418 3419 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask); 3420 if (retval == 0) 3421 retval = sched_setaffinity(pid, new_mask); 3422 free_cpumask_var(new_mask); 3423 return retval; 3424} 3425 3426long sched_getaffinity(pid_t pid, struct cpumask *mask) 3427{ 3428 struct task_struct *p; 3429 unsigned long flags; 3430 int retval; 3431 3432 rcu_read_lock(); 3433 3434 retval = -ESRCH; 3435 p = find_process_by_pid(pid); 3436 if (!p) 3437 goto out_unlock; 3438 3439 retval = security_task_getscheduler(p); 3440 if (retval) 3441 goto out_unlock; 3442 3443 raw_spin_lock_irqsave(&p->pi_lock, flags); 3444 cpumask_and(mask, &p->cpus_allowed, cpu_active_mask); 3445 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 3446 3447out_unlock: 3448 rcu_read_unlock(); 3449 3450 return retval; 3451} 3452 3453/** 3454 * sys_sched_getaffinity - get the cpu affinity of a process 3455 * @pid: pid of the process 3456 * @len: length in bytes of the bitmask pointed to by user_mask_ptr 3457 * @user_mask_ptr: user-space pointer to hold the current cpu mask 3458 * 3459 * Return: 0 on success. An error code otherwise. 3460 */ 3461SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len, 3462 unsigned long __user *, user_mask_ptr) 3463{ 3464 int ret; 3465 cpumask_var_t mask; 3466 3467 if ((len * BITS_PER_BYTE) < nr_cpu_ids) 3468 return -EINVAL; 3469 if (len & (sizeof(unsigned long)-1)) 3470 return -EINVAL; 3471 3472 if (!alloc_cpumask_var(&mask, GFP_KERNEL)) 3473 return -ENOMEM; 3474 3475 ret = sched_getaffinity(pid, mask); 3476 if (ret == 0) { 3477 size_t retlen = min_t(size_t, len, cpumask_size()); 3478 3479 if (copy_to_user(user_mask_ptr, mask, retlen)) 3480 ret = -EFAULT; 3481 else 3482 ret = retlen; 3483 } 3484 free_cpumask_var(mask); 3485 3486 return ret; 3487} 3488 3489/** 3490 * sys_sched_yield - yield the current processor to other threads. 3491 * 3492 * This function yields the current CPU to other tasks. If there are no 3493 * other threads running on this CPU then this function will return. 3494 * 3495 * Return: 0. 3496 */ 3497SYSCALL_DEFINE0(sched_yield) 3498{ 3499 struct rq *rq = this_rq_lock(); 3500 3501 schedstat_inc(rq, yld_count); 3502 current->sched_class->yield_task(rq); 3503 3504 /* 3505 * Since we are going to call schedule() anyway, there's 3506 * no need to preempt or enable interrupts: 3507 */ 3508 __release(rq->lock); 3509 spin_release(&rq->lock.dep_map, 1, _THIS_IP_); 3510 do_raw_spin_unlock(&rq->lock); 3511 sched_preempt_enable_no_resched(); 3512 3513 schedule(); 3514 3515 return 0; 3516} 3517 3518static void __cond_resched(void) 3519{ 3520 __preempt_count_add(PREEMPT_ACTIVE); 3521 __schedule(); 3522 __preempt_count_sub(PREEMPT_ACTIVE); 3523} 3524 3525int __sched _cond_resched(void) 3526{ 3527 if (should_resched()) { 3528 __cond_resched(); 3529 return 1; 3530 } 3531 return 0; 3532} 3533EXPORT_SYMBOL(_cond_resched); 3534 3535/* 3536 * __cond_resched_lock() - if a reschedule is pending, drop the given lock, 3537 * call schedule, and on return reacquire the lock. 3538 * 3539 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level 3540 * operations here to prevent schedule() from being called twice (once via 3541 * spin_unlock(), once by hand). 3542 */ 3543int __cond_resched_lock(spinlock_t *lock) 3544{ 3545 int resched = should_resched(); 3546 int ret = 0; 3547 3548 lockdep_assert_held(lock); 3549 3550 if (spin_needbreak(lock) || resched) { 3551 spin_unlock(lock); 3552 if (resched) 3553 __cond_resched(); 3554 else 3555 cpu_relax(); 3556 ret = 1; 3557 spin_lock(lock); 3558 } 3559 return ret; 3560} 3561EXPORT_SYMBOL(__cond_resched_lock); 3562 3563int __sched __cond_resched_softirq(void) 3564{ 3565 BUG_ON(!in_softirq()); 3566 3567 if (should_resched()) { 3568 local_bh_enable(); 3569 __cond_resched(); 3570 local_bh_disable(); 3571 return 1; 3572 } 3573 return 0; 3574} 3575EXPORT_SYMBOL(__cond_resched_softirq); 3576 3577/** 3578 * yield - yield the current processor to other threads. 3579 * 3580 * Do not ever use this function, there's a 99% chance you're doing it wrong. 3581 * 3582 * The scheduler is at all times free to pick the calling task as the most 3583 * eligible task to run, if removing the yield() call from your code breaks 3584 * it, its already broken. 3585 * 3586 * Typical broken usage is: 3587 * 3588 * while (!event) 3589 * yield(); 3590 * 3591 * where one assumes that yield() will let 'the other' process run that will 3592 * make event true. If the current task is a SCHED_FIFO task that will never 3593 * happen. Never use yield() as a progress guarantee!! 3594 * 3595 * If you want to use yield() to wait for something, use wait_event(). 3596 * If you want to use yield() to be 'nice' for others, use cond_resched(). 3597 * If you still want to use yield(), do not! 3598 */ 3599void __sched yield(void) 3600{ 3601 set_current_state(TASK_RUNNING); 3602 sys_sched_yield(); 3603} 3604EXPORT_SYMBOL(yield); 3605 3606/** 3607 * yield_to - yield the current processor to another thread in 3608 * your thread group, or accelerate that thread toward the 3609 * processor it's on. 3610 * @p: target task 3611 * @preempt: whether task preemption is allowed or not 3612 * 3613 * It's the caller's job to ensure that the target task struct 3614 * can't go away on us before we can do any checks. 3615 * 3616 * Return: 3617 * true (>0) if we indeed boosted the target task. 3618 * false (0) if we failed to boost the target. 3619 * -ESRCH if there's no task to yield to. 3620 */ 3621bool __sched yield_to(struct task_struct *p, bool preempt) 3622{ 3623 struct task_struct *curr = current; 3624 struct rq *rq, *p_rq; 3625 unsigned long flags; 3626 int yielded = 0; 3627 3628 local_irq_save(flags); 3629 rq = this_rq(); 3630 3631again: 3632 p_rq = task_rq(p); 3633 /* 3634 * If we're the only runnable task on the rq and target rq also 3635 * has only one task, there's absolutely no point in yielding. 3636 */ 3637 if (rq->nr_running == 1 && p_rq->nr_running == 1) { 3638 yielded = -ESRCH; 3639 goto out_irq; 3640 } 3641 3642 double_rq_lock(rq, p_rq); 3643 while (task_rq(p) != p_rq) { 3644 double_rq_unlock(rq, p_rq); 3645 goto again; 3646 } 3647 3648 if (!curr->sched_class->yield_to_task) 3649 goto out_unlock; 3650 3651 if (curr->sched_class != p->sched_class) 3652 goto out_unlock; 3653 3654 if (task_running(p_rq, p) || p->state) 3655 goto out_unlock; 3656 3657 yielded = curr->sched_class->yield_to_task(rq, p, preempt); 3658 if (yielded) { 3659 schedstat_inc(rq, yld_count); 3660 /* 3661 * Make p's CPU reschedule; pick_next_entity takes care of 3662 * fairness. 3663 */ 3664 if (preempt && rq != p_rq) 3665 resched_task(p_rq->curr); 3666 } 3667 3668out_unlock: 3669 double_rq_unlock(rq, p_rq); 3670out_irq: 3671 local_irq_restore(flags); 3672 3673 if (yielded > 0) 3674 schedule(); 3675 3676 return yielded; 3677} 3678EXPORT_SYMBOL_GPL(yield_to); 3679 3680/* 3681 * This task is about to go to sleep on IO. Increment rq->nr_iowait so 3682 * that process accounting knows that this is a task in IO wait state. 3683 */ 3684void __sched io_schedule(void) 3685{ 3686 struct rq *rq = raw_rq(); 3687 3688 delayacct_blkio_start(); 3689 atomic_inc(&rq->nr_iowait); 3690 blk_flush_plug(current); 3691 current->in_iowait = 1; 3692 schedule(); 3693 current->in_iowait = 0; 3694 atomic_dec(&rq->nr_iowait); 3695 delayacct_blkio_end(); 3696} 3697EXPORT_SYMBOL(io_schedule); 3698 3699long __sched io_schedule_timeout(long timeout) 3700{ 3701 struct rq *rq = raw_rq(); 3702 long ret; 3703 3704 delayacct_blkio_start(); 3705 atomic_inc(&rq->nr_iowait); 3706 blk_flush_plug(current); 3707 current->in_iowait = 1; 3708 ret = schedule_timeout(timeout); 3709 current->in_iowait = 0; 3710 atomic_dec(&rq->nr_iowait); 3711 delayacct_blkio_end(); 3712 return ret; 3713} 3714 3715/** 3716 * sys_sched_get_priority_max - return maximum RT priority. 3717 * @policy: scheduling class. 3718 * 3719 * Return: On success, this syscall returns the maximum 3720 * rt_priority that can be used by a given scheduling class. 3721 * On failure, a negative error code is returned. 3722 */ 3723SYSCALL_DEFINE1(sched_get_priority_max, int, policy) 3724{ 3725 int ret = -EINVAL; 3726 3727 switch (policy) { 3728 case SCHED_FIFO: 3729 case SCHED_RR: 3730 ret = MAX_USER_RT_PRIO-1; 3731 break; 3732 case SCHED_NORMAL: 3733 case SCHED_BATCH: 3734 case SCHED_IDLE: 3735 ret = 0; 3736 break; 3737 } 3738 return ret; 3739} 3740 3741/** 3742 * sys_sched_get_priority_min - return minimum RT priority. 3743 * @policy: scheduling class. 3744 * 3745 * Return: On success, this syscall returns the minimum 3746 * rt_priority that can be used by a given scheduling class. 3747 * On failure, a negative error code is returned. 3748 */ 3749SYSCALL_DEFINE1(sched_get_priority_min, int, policy) 3750{ 3751 int ret = -EINVAL; 3752 3753 switch (policy) { 3754 case SCHED_FIFO: 3755 case SCHED_RR: 3756 ret = 1; 3757 break; 3758 case SCHED_NORMAL: 3759 case SCHED_BATCH: 3760 case SCHED_IDLE: 3761 ret = 0; 3762 } 3763 return ret; 3764} 3765 3766/** 3767 * sys_sched_rr_get_interval - return the default timeslice of a process. 3768 * @pid: pid of the process. 3769 * @interval: userspace pointer to the timeslice value. 3770 * 3771 * this syscall writes the default timeslice value of a given process 3772 * into the user-space timespec buffer. A value of '0' means infinity. 3773 * 3774 * Return: On success, 0 and the timeslice is in @interval. Otherwise, 3775 * an error code. 3776 */ 3777SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid, 3778 struct timespec __user *, interval) 3779{ 3780 struct task_struct *p; 3781 unsigned int time_slice; 3782 unsigned long flags; 3783 struct rq *rq; 3784 int retval; 3785 struct timespec t; 3786 3787 if (pid < 0) 3788 return -EINVAL; 3789 3790 retval = -ESRCH; 3791 rcu_read_lock(); 3792 p = find_process_by_pid(pid); 3793 if (!p) 3794 goto out_unlock; 3795 3796 retval = security_task_getscheduler(p); 3797 if (retval) 3798 goto out_unlock; 3799 3800 rq = task_rq_lock(p, &flags); 3801 time_slice = p->sched_class->get_rr_interval(rq, p); 3802 task_rq_unlock(rq, p, &flags); 3803 3804 rcu_read_unlock(); 3805 jiffies_to_timespec(time_slice, &t); 3806 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0; 3807 return retval; 3808 3809out_unlock: 3810 rcu_read_unlock(); 3811 return retval; 3812} 3813 3814static const char stat_nam[] = TASK_STATE_TO_CHAR_STR; 3815 3816void sched_show_task(struct task_struct *p) 3817{ 3818 unsigned long free = 0; 3819 int ppid; 3820 unsigned state; 3821 3822 state = p->state ? __ffs(p->state) + 1 : 0; 3823 printk(KERN_INFO "%-15.15s %c", p->comm, 3824 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?'); 3825#if BITS_PER_LONG == 32 3826 if (state == TASK_RUNNING) 3827 printk(KERN_CONT " running "); 3828 else 3829 printk(KERN_CONT " %08lx ", thread_saved_pc(p)); 3830#else 3831 if (state == TASK_RUNNING) 3832 printk(KERN_CONT " running task "); 3833 else 3834 printk(KERN_CONT " %016lx ", thread_saved_pc(p)); 3835#endif 3836#ifdef CONFIG_DEBUG_STACK_USAGE 3837 free = stack_not_used(p); 3838#endif 3839 rcu_read_lock(); 3840 ppid = task_pid_nr(rcu_dereference(p->real_parent)); 3841 rcu_read_unlock(); 3842 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free, 3843 task_pid_nr(p), ppid, 3844 (unsigned long)task_thread_info(p)->flags); 3845 3846 print_worker_info(KERN_INFO, p); 3847 show_stack(p, NULL); 3848} 3849 3850void show_state_filter(unsigned long state_filter) 3851{ 3852 struct task_struct *g, *p; 3853 3854#if BITS_PER_LONG == 32 3855 printk(KERN_INFO 3856 " task PC stack pid father\n"); 3857#else 3858 printk(KERN_INFO 3859 " task PC stack pid father\n"); 3860#endif 3861 rcu_read_lock(); 3862 do_each_thread(g, p) { 3863 /* 3864 * reset the NMI-timeout, listing all files on a slow 3865 * console might take a lot of time: 3866 */ 3867 touch_nmi_watchdog(); 3868 if (!state_filter || (p->state & state_filter)) 3869 sched_show_task(p); 3870 } while_each_thread(g, p); 3871 3872 touch_all_softlockup_watchdogs(); 3873 3874#ifdef CONFIG_SCHED_DEBUG 3875 sysrq_sched_debug_show(); 3876#endif 3877 rcu_read_unlock(); 3878 /* 3879 * Only show locks if all tasks are dumped: 3880 */ 3881 if (!state_filter) 3882 debug_show_all_locks(); 3883} 3884 3885void init_idle_bootup_task(struct task_struct *idle) 3886{ 3887 idle->sched_class = &idle_sched_class; 3888} 3889 3890/** 3891 * init_idle - set up an idle thread for a given CPU 3892 * @idle: task in question 3893 * @cpu: cpu the idle task belongs to 3894 * 3895 * NOTE: this function does not set the idle thread's NEED_RESCHED 3896 * flag, to make booting more robust. 3897 */ 3898void init_idle(struct task_struct *idle, int cpu) 3899{ 3900 struct rq *rq = cpu_rq(cpu); 3901 unsigned long flags; 3902 3903 raw_spin_lock_irqsave(&rq->lock, flags); 3904 3905 __sched_fork(0, idle); 3906 idle->state = TASK_RUNNING; 3907 idle->se.exec_start = sched_clock(); 3908 3909 do_set_cpus_allowed(idle, cpumask_of(cpu)); 3910 /* 3911 * We're having a chicken and egg problem, even though we are 3912 * holding rq->lock, the cpu isn't yet set to this cpu so the 3913 * lockdep check in task_group() will fail. 3914 * 3915 * Similar case to sched_fork(). / Alternatively we could 3916 * use task_rq_lock() here and obtain the other rq->lock. 3917 * 3918 * Silence PROVE_RCU 3919 */ 3920 rcu_read_lock(); 3921 __set_task_cpu(idle, cpu); 3922 rcu_read_unlock(); 3923 3924 rq->curr = rq->idle = idle; 3925#if defined(CONFIG_SMP) 3926 idle->on_cpu = 1; 3927#endif 3928 raw_spin_unlock_irqrestore(&rq->lock, flags); 3929 3930 /* Set the preempt count _outside_ the spinlocks! */ 3931 init_idle_preempt_count(idle, cpu); 3932 3933 /* 3934 * The idle tasks have their own, simple scheduling class: 3935 */ 3936 idle->sched_class = &idle_sched_class; 3937 ftrace_graph_init_idle_task(idle, cpu); 3938 vtime_init_idle(idle, cpu); 3939#if defined(CONFIG_SMP) 3940 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu); 3941#endif 3942} 3943 3944#ifdef CONFIG_SMP 3945void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask) 3946{ 3947 if (p->sched_class && p->sched_class->set_cpus_allowed) 3948 p->sched_class->set_cpus_allowed(p, new_mask); 3949 3950 cpumask_copy(&p->cpus_allowed, new_mask); 3951 p->nr_cpus_allowed = cpumask_weight(new_mask); 3952} 3953 3954/* 3955 * This is how migration works: 3956 * 3957 * 1) we invoke migration_cpu_stop() on the target CPU using 3958 * stop_one_cpu(). 3959 * 2) stopper starts to run (implicitly forcing the migrated thread 3960 * off the CPU) 3961 * 3) it checks whether the migrated task is still in the wrong runqueue. 3962 * 4) if it's in the wrong runqueue then the migration thread removes 3963 * it and puts it into the right queue. 3964 * 5) stopper completes and stop_one_cpu() returns and the migration 3965 * is done. 3966 */ 3967 3968/* 3969 * Change a given task's CPU affinity. Migrate the thread to a 3970 * proper CPU and schedule it away if the CPU it's executing on 3971 * is removed from the allowed bitmask. 3972 * 3973 * NOTE: the caller must have a valid reference to the task, the 3974 * task must not exit() & deallocate itself prematurely. The 3975 * call is not atomic; no spinlocks may be held. 3976 */ 3977int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask) 3978{ 3979 unsigned long flags; 3980 struct rq *rq; 3981 unsigned int dest_cpu; 3982 int ret = 0; 3983 3984 rq = task_rq_lock(p, &flags); 3985 3986 if (cpumask_equal(&p->cpus_allowed, new_mask)) 3987 goto out; 3988 3989 if (!cpumask_intersects(new_mask, cpu_active_mask)) { 3990 ret = -EINVAL; 3991 goto out; 3992 } 3993 3994 do_set_cpus_allowed(p, new_mask); 3995 3996 /* Can the task run on the task's current CPU? If so, we're done */ 3997 if (cpumask_test_cpu(task_cpu(p), new_mask)) 3998 goto out; 3999 4000 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask); 4001 if (p->on_rq) { 4002 struct migration_arg arg = { p, dest_cpu }; 4003 /* Need help from migration thread: drop lock and wait. */ 4004 task_rq_unlock(rq, p, &flags); 4005 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg); 4006 tlb_migrate_finish(p->mm); 4007 return 0; 4008 } 4009out: 4010 task_rq_unlock(rq, p, &flags); 4011 4012 return ret; 4013} 4014EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr); 4015 4016/* 4017 * Move (not current) task off this cpu, onto dest cpu. We're doing 4018 * this because either it can't run here any more (set_cpus_allowed() 4019 * away from this CPU, or CPU going down), or because we're 4020 * attempting to rebalance this task on exec (sched_exec). 4021 * 4022 * So we race with normal scheduler movements, but that's OK, as long 4023 * as the task is no longer on this CPU. 4024 * 4025 * Returns non-zero if task was successfully migrated. 4026 */ 4027static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu) 4028{ 4029 struct rq *rq_dest, *rq_src; 4030 int ret = 0; 4031 4032 if (unlikely(!cpu_active(dest_cpu))) 4033 return ret; 4034 4035 rq_src = cpu_rq(src_cpu); 4036 rq_dest = cpu_rq(dest_cpu); 4037 4038 raw_spin_lock(&p->pi_lock); 4039 double_rq_lock(rq_src, rq_dest); 4040 /* Already moved. */ 4041 if (task_cpu(p) != src_cpu) 4042 goto done; 4043 /* Affinity changed (again). */ 4044 if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p))) 4045 goto fail; 4046 4047 /* 4048 * If we're not on a rq, the next wake-up will ensure we're 4049 * placed properly. 4050 */ 4051 if (p->on_rq) { 4052 dequeue_task(rq_src, p, 0); 4053 set_task_cpu(p, dest_cpu); 4054 enqueue_task(rq_dest, p, 0); 4055 check_preempt_curr(rq_dest, p, 0); 4056 } 4057done: 4058 ret = 1; 4059fail: 4060 double_rq_unlock(rq_src, rq_dest); 4061 raw_spin_unlock(&p->pi_lock); 4062 return ret; 4063} 4064 4065#ifdef CONFIG_NUMA_BALANCING 4066/* Migrate current task p to target_cpu */ 4067int migrate_task_to(struct task_struct *p, int target_cpu) 4068{ 4069 struct migration_arg arg = { p, target_cpu }; 4070 int curr_cpu = task_cpu(p); 4071 4072 if (curr_cpu == target_cpu) 4073 return 0; 4074 4075 if (!cpumask_test_cpu(target_cpu, tsk_cpus_allowed(p))) 4076 return -EINVAL; 4077 4078 /* TODO: This is not properly updating schedstats */ 4079 4080 return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg); 4081} 4082 4083/* 4084 * Requeue a task on a given node and accurately track the number of NUMA 4085 * tasks on the runqueues 4086 */ 4087void sched_setnuma(struct task_struct *p, int nid) 4088{ 4089 struct rq *rq; 4090 unsigned long flags; 4091 bool on_rq, running; 4092 4093 rq = task_rq_lock(p, &flags); 4094 on_rq = p->on_rq; 4095 running = task_current(rq, p); 4096 4097 if (on_rq) 4098 dequeue_task(rq, p, 0); 4099 if (running) 4100 p->sched_class->put_prev_task(rq, p); 4101 4102 p->numa_preferred_nid = nid; 4103 4104 if (running) 4105 p->sched_class->set_curr_task(rq); 4106 if (on_rq) 4107 enqueue_task(rq, p, 0); 4108 task_rq_unlock(rq, p, &flags); 4109} 4110#endif 4111 4112/* 4113 * migration_cpu_stop - this will be executed by a highprio stopper thread 4114 * and performs thread migration by bumping thread off CPU then 4115 * 'pushing' onto another runqueue. 4116 */ 4117static int migration_cpu_stop(void *data) 4118{ 4119 struct migration_arg *arg = data; 4120 4121 /* 4122 * The original target cpu might have gone down and we might 4123 * be on another cpu but it doesn't matter. 4124 */ 4125 local_irq_disable(); 4126 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu); 4127 local_irq_enable(); 4128 return 0; 4129} 4130 4131#ifdef CONFIG_HOTPLUG_CPU 4132 4133/* 4134 * Ensures that the idle task is using init_mm right before its cpu goes 4135 * offline. 4136 */ 4137void idle_task_exit(void) 4138{ 4139 struct mm_struct *mm = current->active_mm; 4140 4141 BUG_ON(cpu_online(smp_processor_id())); 4142 4143 if (mm != &init_mm) 4144 switch_mm(mm, &init_mm, current); 4145 mmdrop(mm); 4146} 4147 4148/* 4149 * Since this CPU is going 'away' for a while, fold any nr_active delta 4150 * we might have. Assumes we're called after migrate_tasks() so that the 4151 * nr_active count is stable. 4152 * 4153 * Also see the comment "Global load-average calculations". 4154 */ 4155static void calc_load_migrate(struct rq *rq) 4156{ 4157 long delta = calc_load_fold_active(rq); 4158 if (delta) 4159 atomic_long_add(delta, &calc_load_tasks); 4160} 4161 4162/* 4163 * Migrate all tasks from the rq, sleeping tasks will be migrated by 4164 * try_to_wake_up()->select_task_rq(). 4165 * 4166 * Called with rq->lock held even though we'er in stop_machine() and 4167 * there's no concurrency possible, we hold the required locks anyway 4168 * because of lock validation efforts. 4169 */ 4170static void migrate_tasks(unsigned int dead_cpu) 4171{ 4172 struct rq *rq = cpu_rq(dead_cpu); 4173 struct task_struct *next, *stop = rq->stop; 4174 int dest_cpu; 4175 4176 /* 4177 * Fudge the rq selection such that the below task selection loop 4178 * doesn't get stuck on the currently eligible stop task. 4179 * 4180 * We're currently inside stop_machine() and the rq is either stuck 4181 * in the stop_machine_cpu_stop() loop, or we're executing this code, 4182 * either way we should never end up calling schedule() until we're 4183 * done here. 4184 */ 4185 rq->stop = NULL; 4186 4187 /* 4188 * put_prev_task() and pick_next_task() sched 4189 * class method both need to have an up-to-date 4190 * value of rq->clock[_task] 4191 */ 4192 update_rq_clock(rq); 4193 4194 for ( ; ; ) { 4195 /* 4196 * There's this thread running, bail when that's the only 4197 * remaining thread. 4198 */ 4199 if (rq->nr_running == 1) 4200 break; 4201 4202 next = pick_next_task(rq); 4203 BUG_ON(!next); 4204 next->sched_class->put_prev_task(rq, next); 4205 4206 /* Find suitable destination for @next, with force if needed. */ 4207 dest_cpu = select_fallback_rq(dead_cpu, next); 4208 raw_spin_unlock(&rq->lock); 4209 4210 __migrate_task(next, dead_cpu, dest_cpu); 4211 4212 raw_spin_lock(&rq->lock); 4213 } 4214 4215 rq->stop = stop; 4216} 4217 4218#endif /* CONFIG_HOTPLUG_CPU */ 4219 4220#if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL) 4221 4222static struct ctl_table sd_ctl_dir[] = { 4223 { 4224 .procname = "sched_domain", 4225 .mode = 0555, 4226 }, 4227 {} 4228}; 4229 4230static struct ctl_table sd_ctl_root[] = { 4231 { 4232 .procname = "kernel", 4233 .mode = 0555, 4234 .child = sd_ctl_dir, 4235 }, 4236 {} 4237}; 4238 4239static struct ctl_table *sd_alloc_ctl_entry(int n) 4240{ 4241 struct ctl_table *entry = 4242 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL); 4243 4244 return entry; 4245} 4246 4247static void sd_free_ctl_entry(struct ctl_table **tablep) 4248{ 4249 struct ctl_table *entry; 4250 4251 /* 4252 * In the intermediate directories, both the child directory and 4253 * procname are dynamically allocated and could fail but the mode 4254 * will always be set. In the lowest directory the names are 4255 * static strings and all have proc handlers. 4256 */ 4257 for (entry = *tablep; entry->mode; entry++) { 4258 if (entry->child) 4259 sd_free_ctl_entry(&entry->child); 4260 if (entry->proc_handler == NULL) 4261 kfree(entry->procname); 4262 } 4263 4264 kfree(*tablep); 4265 *tablep = NULL; 4266} 4267 4268static int min_load_idx = 0; 4269static int max_load_idx = CPU_LOAD_IDX_MAX-1; 4270 4271static void 4272set_table_entry(struct ctl_table *entry, 4273 const char *procname, void *data, int maxlen, 4274 umode_t mode, proc_handler *proc_handler, 4275 bool load_idx) 4276{ 4277 entry->procname = procname; 4278 entry->data = data; 4279 entry->maxlen = maxlen; 4280 entry->mode = mode; 4281 entry->proc_handler = proc_handler; 4282 4283 if (load_idx) { 4284 entry->extra1 = &min_load_idx; 4285 entry->extra2 = &max_load_idx; 4286 } 4287} 4288 4289static struct ctl_table * 4290sd_alloc_ctl_domain_table(struct sched_domain *sd) 4291{ 4292 struct ctl_table *table = sd_alloc_ctl_entry(13); 4293 4294 if (table == NULL) 4295 return NULL; 4296 4297 set_table_entry(&table[0], "min_interval", &sd->min_interval, 4298 sizeof(long), 0644, proc_doulongvec_minmax, false); 4299 set_table_entry(&table[1], "max_interval", &sd->max_interval, 4300 sizeof(long), 0644, proc_doulongvec_minmax, false); 4301 set_table_entry(&table[2], "busy_idx", &sd->busy_idx, 4302 sizeof(int), 0644, proc_dointvec_minmax, true); 4303 set_table_entry(&table[3], "idle_idx", &sd->idle_idx, 4304 sizeof(int), 0644, proc_dointvec_minmax, true); 4305 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx, 4306 sizeof(int), 0644, proc_dointvec_minmax, true); 4307 set_table_entry(&table[5], "wake_idx", &sd->wake_idx, 4308 sizeof(int), 0644, proc_dointvec_minmax, true); 4309 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx, 4310 sizeof(int), 0644, proc_dointvec_minmax, true); 4311 set_table_entry(&table[7], "busy_factor", &sd->busy_factor, 4312 sizeof(int), 0644, proc_dointvec_minmax, false); 4313 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct, 4314 sizeof(int), 0644, proc_dointvec_minmax, false); 4315 set_table_entry(&table[9], "cache_nice_tries", 4316 &sd->cache_nice_tries, 4317 sizeof(int), 0644, proc_dointvec_minmax, false); 4318 set_table_entry(&table[10], "flags", &sd->flags, 4319 sizeof(int), 0644, proc_dointvec_minmax, false); 4320 set_table_entry(&table[11], "name", sd->name, 4321 CORENAME_MAX_SIZE, 0444, proc_dostring, false); 4322 /* &table[12] is terminator */ 4323 4324 return table; 4325} 4326 4327static struct ctl_table *sd_alloc_ctl_cpu_table(int cpu) 4328{ 4329 struct ctl_table *entry, *table; 4330 struct sched_domain *sd; 4331 int domain_num = 0, i; 4332 char buf[32]; 4333 4334 for_each_domain(cpu, sd) 4335 domain_num++; 4336 entry = table = sd_alloc_ctl_entry(domain_num + 1); 4337 if (table == NULL) 4338 return NULL; 4339 4340 i = 0; 4341 for_each_domain(cpu, sd) { 4342 snprintf(buf, 32, "domain%d", i); 4343 entry->procname = kstrdup(buf, GFP_KERNEL); 4344 entry->mode = 0555; 4345 entry->child = sd_alloc_ctl_domain_table(sd); 4346 entry++; 4347 i++; 4348 } 4349 return table; 4350} 4351 4352static struct ctl_table_header *sd_sysctl_header; 4353static void register_sched_domain_sysctl(void) 4354{ 4355 int i, cpu_num = num_possible_cpus(); 4356 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1); 4357 char buf[32]; 4358 4359 WARN_ON(sd_ctl_dir[0].child); 4360 sd_ctl_dir[0].child = entry; 4361 4362 if (entry == NULL) 4363 return; 4364 4365 for_each_possible_cpu(i) { 4366 snprintf(buf, 32, "cpu%d", i); 4367 entry->procname = kstrdup(buf, GFP_KERNEL); 4368 entry->mode = 0555; 4369 entry->child = sd_alloc_ctl_cpu_table(i); 4370 entry++; 4371 } 4372 4373 WARN_ON(sd_sysctl_header); 4374 sd_sysctl_header = register_sysctl_table(sd_ctl_root); 4375} 4376 4377/* may be called multiple times per register */ 4378static void unregister_sched_domain_sysctl(void) 4379{ 4380 if (sd_sysctl_header) 4381 unregister_sysctl_table(sd_sysctl_header); 4382 sd_sysctl_header = NULL; 4383 if (sd_ctl_dir[0].child) 4384 sd_free_ctl_entry(&sd_ctl_dir[0].child); 4385} 4386#else 4387static void register_sched_domain_sysctl(void) 4388{ 4389} 4390static void unregister_sched_domain_sysctl(void) 4391{ 4392} 4393#endif 4394 4395static void set_rq_online(struct rq *rq) 4396{ 4397 if (!rq->online) { 4398 const struct sched_class *class; 4399 4400 cpumask_set_cpu(rq->cpu, rq->rd->online); 4401 rq->online = 1; 4402 4403 for_each_class(class) { 4404 if (class->rq_online) 4405 class->rq_online(rq); 4406 } 4407 } 4408} 4409 4410static void set_rq_offline(struct rq *rq) 4411{ 4412 if (rq->online) { 4413 const struct sched_class *class; 4414 4415 for_each_class(class) { 4416 if (class->rq_offline) 4417 class->rq_offline(rq); 4418 } 4419 4420 cpumask_clear_cpu(rq->cpu, rq->rd->online); 4421 rq->online = 0; 4422 } 4423} 4424 4425/* 4426 * migration_call - callback that gets triggered when a CPU is added. 4427 * Here we can start up the necessary migration thread for the new CPU. 4428 */ 4429static int 4430migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu) 4431{ 4432 int cpu = (long)hcpu; 4433 unsigned long flags; 4434 struct rq *rq = cpu_rq(cpu); 4435 4436 switch (action & ~CPU_TASKS_FROZEN) { 4437 4438 case CPU_UP_PREPARE: 4439 rq->calc_load_update = calc_load_update; 4440 break; 4441 4442 case CPU_ONLINE: 4443 /* Update our root-domain */ 4444 raw_spin_lock_irqsave(&rq->lock, flags); 4445 if (rq->rd) { 4446 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span)); 4447 4448 set_rq_online(rq); 4449 } 4450 raw_spin_unlock_irqrestore(&rq->lock, flags); 4451 break; 4452 4453#ifdef CONFIG_HOTPLUG_CPU 4454 case CPU_DYING: 4455 sched_ttwu_pending(); 4456 /* Update our root-domain */ 4457 raw_spin_lock_irqsave(&rq->lock, flags); 4458 if (rq->rd) { 4459 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span)); 4460 set_rq_offline(rq); 4461 } 4462 migrate_tasks(cpu); 4463 BUG_ON(rq->nr_running != 1); /* the migration thread */ 4464 raw_spin_unlock_irqrestore(&rq->lock, flags); 4465 break; 4466 4467 case CPU_DEAD: 4468 calc_load_migrate(rq); 4469 break; 4470#endif 4471 } 4472 4473 update_max_interval(); 4474 4475 return NOTIFY_OK; 4476} 4477 4478/* 4479 * Register at high priority so that task migration (migrate_all_tasks) 4480 * happens before everything else. This has to be lower priority than 4481 * the notifier in the perf_event subsystem, though. 4482 */ 4483static struct notifier_block migration_notifier = { 4484 .notifier_call = migration_call, 4485 .priority = CPU_PRI_MIGRATION, 4486}; 4487 4488static int sched_cpu_active(struct notifier_block *nfb, 4489 unsigned long action, void *hcpu) 4490{ 4491 switch (action & ~CPU_TASKS_FROZEN) { 4492 case CPU_STARTING: 4493 case CPU_DOWN_FAILED: 4494 set_cpu_active((long)hcpu, true); 4495 return NOTIFY_OK; 4496 default: 4497 return NOTIFY_DONE; 4498 } 4499} 4500 4501static int sched_cpu_inactive(struct notifier_block *nfb, 4502 unsigned long action, void *hcpu) 4503{ 4504 switch (action & ~CPU_TASKS_FROZEN) { 4505 case CPU_DOWN_PREPARE: 4506 set_cpu_active((long)hcpu, false); 4507 return NOTIFY_OK; 4508 default: 4509 return NOTIFY_DONE; 4510 } 4511} 4512 4513static int __init migration_init(void) 4514{ 4515 void *cpu = (void *)(long)smp_processor_id(); 4516 int err; 4517 4518 /* Initialize migration for the boot CPU */ 4519 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu); 4520 BUG_ON(err == NOTIFY_BAD); 4521 migration_call(&migration_notifier, CPU_ONLINE, cpu); 4522 register_cpu_notifier(&migration_notifier); 4523 4524 /* Register cpu active notifiers */ 4525 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE); 4526 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE); 4527 4528 return 0; 4529} 4530early_initcall(migration_init); 4531#endif 4532 4533#ifdef CONFIG_SMP 4534 4535static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */ 4536 4537#ifdef CONFIG_SCHED_DEBUG 4538 4539static __read_mostly int sched_debug_enabled; 4540 4541static int __init sched_debug_setup(char *str) 4542{ 4543 sched_debug_enabled = 1; 4544 4545 return 0; 4546} 4547early_param("sched_debug", sched_debug_setup); 4548 4549static inline bool sched_debug(void) 4550{ 4551 return sched_debug_enabled; 4552} 4553 4554static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level, 4555 struct cpumask *groupmask) 4556{ 4557 struct sched_group *group = sd->groups; 4558 char str[256]; 4559 4560 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd)); 4561 cpumask_clear(groupmask); 4562 4563 printk(KERN_DEBUG "%*s domain %d: ", level, "", level); 4564 4565 if (!(sd->flags & SD_LOAD_BALANCE)) { 4566 printk("does not load-balance\n"); 4567 if (sd->parent) 4568 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain" 4569 " has parent"); 4570 return -1; 4571 } 4572 4573 printk(KERN_CONT "span %s level %s\n", str, sd->name); 4574 4575 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) { 4576 printk(KERN_ERR "ERROR: domain->span does not contain " 4577 "CPU%d\n", cpu); 4578 } 4579 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) { 4580 printk(KERN_ERR "ERROR: domain->groups does not contain" 4581 " CPU%d\n", cpu); 4582 } 4583 4584 printk(KERN_DEBUG "%*s groups:", level + 1, ""); 4585 do { 4586 if (!group) { 4587 printk("\n"); 4588 printk(KERN_ERR "ERROR: group is NULL\n"); 4589 break; 4590 } 4591 4592 /* 4593 * Even though we initialize ->power to something semi-sane, 4594 * we leave power_orig unset. This allows us to detect if 4595 * domain iteration is still funny without causing /0 traps. 4596 */ 4597 if (!group->sgp->power_orig) { 4598 printk(KERN_CONT "\n"); 4599 printk(KERN_ERR "ERROR: domain->cpu_power not " 4600 "set\n"); 4601 break; 4602 } 4603 4604 if (!cpumask_weight(sched_group_cpus(group))) { 4605 printk(KERN_CONT "\n"); 4606 printk(KERN_ERR "ERROR: empty group\n"); 4607 break; 4608 } 4609 4610 if (!(sd->flags & SD_OVERLAP) && 4611 cpumask_intersects(groupmask, sched_group_cpus(group))) { 4612 printk(KERN_CONT "\n"); 4613 printk(KERN_ERR "ERROR: repeated CPUs\n"); 4614 break; 4615 } 4616 4617 cpumask_or(groupmask, groupmask, sched_group_cpus(group)); 4618 4619 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group)); 4620 4621 printk(KERN_CONT " %s", str); 4622 if (group->sgp->power != SCHED_POWER_SCALE) { 4623 printk(KERN_CONT " (cpu_power = %d)", 4624 group->sgp->power); 4625 } 4626 4627 group = group->next; 4628 } while (group != sd->groups); 4629 printk(KERN_CONT "\n"); 4630 4631 if (!cpumask_equal(sched_domain_span(sd), groupmask)) 4632 printk(KERN_ERR "ERROR: groups don't span domain->span\n"); 4633 4634 if (sd->parent && 4635 !cpumask_subset(groupmask, sched_domain_span(sd->parent))) 4636 printk(KERN_ERR "ERROR: parent span is not a superset " 4637 "of domain->span\n"); 4638 return 0; 4639} 4640 4641static void sched_domain_debug(struct sched_domain *sd, int cpu) 4642{ 4643 int level = 0; 4644 4645 if (!sched_debug_enabled) 4646 return; 4647 4648 if (!sd) { 4649 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu); 4650 return; 4651 } 4652 4653 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu); 4654 4655 for (;;) { 4656 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask)) 4657 break; 4658 level++; 4659 sd = sd->parent; 4660 if (!sd) 4661 break; 4662 } 4663} 4664#else /* !CONFIG_SCHED_DEBUG */ 4665# define sched_domain_debug(sd, cpu) do { } while (0) 4666static inline bool sched_debug(void) 4667{ 4668 return false; 4669} 4670#endif /* CONFIG_SCHED_DEBUG */ 4671 4672static int sd_degenerate(struct sched_domain *sd) 4673{ 4674 if (cpumask_weight(sched_domain_span(sd)) == 1) 4675 return 1; 4676 4677 /* Following flags need at least 2 groups */ 4678 if (sd->flags & (SD_LOAD_BALANCE | 4679 SD_BALANCE_NEWIDLE | 4680 SD_BALANCE_FORK | 4681 SD_BALANCE_EXEC | 4682 SD_SHARE_CPUPOWER | 4683 SD_SHARE_PKG_RESOURCES)) { 4684 if (sd->groups != sd->groups->next) 4685 return 0; 4686 } 4687 4688 /* Following flags don't use groups */ 4689 if (sd->flags & (SD_WAKE_AFFINE)) 4690 return 0; 4691 4692 return 1; 4693} 4694 4695static int 4696sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent) 4697{ 4698 unsigned long cflags = sd->flags, pflags = parent->flags; 4699 4700 if (sd_degenerate(parent)) 4701 return 1; 4702 4703 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent))) 4704 return 0; 4705 4706 /* Flags needing groups don't count if only 1 group in parent */ 4707 if (parent->groups == parent->groups->next) { 4708 pflags &= ~(SD_LOAD_BALANCE | 4709 SD_BALANCE_NEWIDLE | 4710 SD_BALANCE_FORK | 4711 SD_BALANCE_EXEC | 4712 SD_SHARE_CPUPOWER | 4713 SD_SHARE_PKG_RESOURCES | 4714 SD_PREFER_SIBLING); 4715 if (nr_node_ids == 1) 4716 pflags &= ~SD_SERIALIZE; 4717 } 4718 if (~cflags & pflags) 4719 return 0; 4720 4721 return 1; 4722} 4723 4724static void free_rootdomain(struct rcu_head *rcu) 4725{ 4726 struct root_domain *rd = container_of(rcu, struct root_domain, rcu); 4727 4728 cpupri_cleanup(&rd->cpupri); 4729 free_cpumask_var(rd->rto_mask); 4730 free_cpumask_var(rd->online); 4731 free_cpumask_var(rd->span); 4732 kfree(rd); 4733} 4734 4735static void rq_attach_root(struct rq *rq, struct root_domain *rd) 4736{ 4737 struct root_domain *old_rd = NULL; 4738 unsigned long flags; 4739 4740 raw_spin_lock_irqsave(&rq->lock, flags); 4741 4742 if (rq->rd) { 4743 old_rd = rq->rd; 4744 4745 if (cpumask_test_cpu(rq->cpu, old_rd->online)) 4746 set_rq_offline(rq); 4747 4748 cpumask_clear_cpu(rq->cpu, old_rd->span); 4749 4750 /* 4751 * If we dont want to free the old_rt yet then 4752 * set old_rd to NULL to skip the freeing later 4753 * in this function: 4754 */ 4755 if (!atomic_dec_and_test(&old_rd->refcount)) 4756 old_rd = NULL; 4757 } 4758 4759 atomic_inc(&rd->refcount); 4760 rq->rd = rd; 4761 4762 cpumask_set_cpu(rq->cpu, rd->span); 4763 if (cpumask_test_cpu(rq->cpu, cpu_active_mask)) 4764 set_rq_online(rq); 4765 4766 raw_spin_unlock_irqrestore(&rq->lock, flags); 4767 4768 if (old_rd) 4769 call_rcu_sched(&old_rd->rcu, free_rootdomain); 4770} 4771 4772static int init_rootdomain(struct root_domain *rd) 4773{ 4774 memset(rd, 0, sizeof(*rd)); 4775 4776 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL)) 4777 goto out; 4778 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL)) 4779 goto free_span; 4780 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL)) 4781 goto free_online; 4782 4783 if (cpupri_init(&rd->cpupri) != 0) 4784 goto free_rto_mask; 4785 return 0; 4786 4787free_rto_mask: 4788 free_cpumask_var(rd->rto_mask); 4789free_online: 4790 free_cpumask_var(rd->online); 4791free_span: 4792 free_cpumask_var(rd->span); 4793out: 4794 return -ENOMEM; 4795} 4796 4797/* 4798 * By default the system creates a single root-domain with all cpus as 4799 * members (mimicking the global state we have today). 4800 */ 4801struct root_domain def_root_domain; 4802 4803static void init_defrootdomain(void) 4804{ 4805 init_rootdomain(&def_root_domain); 4806 4807 atomic_set(&def_root_domain.refcount, 1); 4808} 4809 4810static struct root_domain *alloc_rootdomain(void) 4811{ 4812 struct root_domain *rd; 4813 4814 rd = kmalloc(sizeof(*rd), GFP_KERNEL); 4815 if (!rd) 4816 return NULL; 4817 4818 if (init_rootdomain(rd) != 0) { 4819 kfree(rd); 4820 return NULL; 4821 } 4822 4823 return rd; 4824} 4825 4826static void free_sched_groups(struct sched_group *sg, int free_sgp) 4827{ 4828 struct sched_group *tmp, *first; 4829 4830 if (!sg) 4831 return; 4832 4833 first = sg; 4834 do { 4835 tmp = sg->next; 4836 4837 if (free_sgp && atomic_dec_and_test(&sg->sgp->ref)) 4838 kfree(sg->sgp); 4839 4840 kfree(sg); 4841 sg = tmp; 4842 } while (sg != first); 4843} 4844 4845static void free_sched_domain(struct rcu_head *rcu) 4846{ 4847 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu); 4848 4849 /* 4850 * If its an overlapping domain it has private groups, iterate and 4851 * nuke them all. 4852 */ 4853 if (sd->flags & SD_OVERLAP) { 4854 free_sched_groups(sd->groups, 1); 4855 } else if (atomic_dec_and_test(&sd->groups->ref)) { 4856 kfree(sd->groups->sgp); 4857 kfree(sd->groups); 4858 } 4859 kfree(sd); 4860} 4861 4862static void destroy_sched_domain(struct sched_domain *sd, int cpu) 4863{ 4864 call_rcu(&sd->rcu, free_sched_domain); 4865} 4866 4867static void destroy_sched_domains(struct sched_domain *sd, int cpu) 4868{ 4869 for (; sd; sd = sd->parent) 4870 destroy_sched_domain(sd, cpu); 4871} 4872 4873/* 4874 * Keep a special pointer to the highest sched_domain that has 4875 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this 4876 * allows us to avoid some pointer chasing select_idle_sibling(). 4877 * 4878 * Also keep a unique ID per domain (we use the first cpu number in 4879 * the cpumask of the domain), this allows us to quickly tell if 4880 * two cpus are in the same cache domain, see cpus_share_cache(). 4881 */ 4882DEFINE_PER_CPU(struct sched_domain *, sd_llc); 4883DEFINE_PER_CPU(int, sd_llc_size); 4884DEFINE_PER_CPU(int, sd_llc_id); 4885DEFINE_PER_CPU(struct sched_domain *, sd_numa); 4886 4887static void update_top_cache_domain(int cpu) 4888{ 4889 struct sched_domain *sd; 4890 int id = cpu; 4891 int size = 1; 4892 4893 sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES); 4894 if (sd) { 4895 id = cpumask_first(sched_domain_span(sd)); 4896 size = cpumask_weight(sched_domain_span(sd)); 4897 } 4898 4899 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd); 4900 per_cpu(sd_llc_size, cpu) = size; 4901 per_cpu(sd_llc_id, cpu) = id; 4902 4903 sd = lowest_flag_domain(cpu, SD_NUMA); 4904 rcu_assign_pointer(per_cpu(sd_numa, cpu), sd); 4905} 4906 4907/* 4908 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must 4909 * hold the hotplug lock. 4910 */ 4911static void 4912cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu) 4913{ 4914 struct rq *rq = cpu_rq(cpu); 4915 struct sched_domain *tmp; 4916 4917 /* Remove the sched domains which do not contribute to scheduling. */ 4918 for (tmp = sd; tmp; ) { 4919 struct sched_domain *parent = tmp->parent; 4920 if (!parent) 4921 break; 4922 4923 if (sd_parent_degenerate(tmp, parent)) { 4924 tmp->parent = parent->parent; 4925 if (parent->parent) 4926 parent->parent->child = tmp; 4927 /* 4928 * Transfer SD_PREFER_SIBLING down in case of a 4929 * degenerate parent; the spans match for this 4930 * so the property transfers. 4931 */ 4932 if (parent->flags & SD_PREFER_SIBLING) 4933 tmp->flags |= SD_PREFER_SIBLING; 4934 destroy_sched_domain(parent, cpu); 4935 } else 4936 tmp = tmp->parent; 4937 } 4938 4939 if (sd && sd_degenerate(sd)) { 4940 tmp = sd; 4941 sd = sd->parent; 4942 destroy_sched_domain(tmp, cpu); 4943 if (sd) 4944 sd->child = NULL; 4945 } 4946 4947 sched_domain_debug(sd, cpu); 4948 4949 rq_attach_root(rq, rd); 4950 tmp = rq->sd; 4951 rcu_assign_pointer(rq->sd, sd); 4952 destroy_sched_domains(tmp, cpu); 4953 4954 update_top_cache_domain(cpu); 4955} 4956 4957/* cpus with isolated domains */ 4958static cpumask_var_t cpu_isolated_map; 4959 4960/* Setup the mask of cpus configured for isolated domains */ 4961static int __init isolated_cpu_setup(char *str) 4962{ 4963 alloc_bootmem_cpumask_var(&cpu_isolated_map); 4964 cpulist_parse(str, cpu_isolated_map); 4965 return 1; 4966} 4967 4968__setup("isolcpus=", isolated_cpu_setup); 4969 4970static const struct cpumask *cpu_cpu_mask(int cpu) 4971{ 4972 return cpumask_of_node(cpu_to_node(cpu)); 4973} 4974 4975struct sd_data { 4976 struct sched_domain **__percpu sd; 4977 struct sched_group **__percpu sg; 4978 struct sched_group_power **__percpu sgp; 4979}; 4980 4981struct s_data { 4982 struct sched_domain ** __percpu sd; 4983 struct root_domain *rd; 4984}; 4985 4986enum s_alloc { 4987 sa_rootdomain, 4988 sa_sd, 4989 sa_sd_storage, 4990 sa_none, 4991}; 4992 4993struct sched_domain_topology_level; 4994 4995typedef struct sched_domain *(*sched_domain_init_f)(struct sched_domain_topology_level *tl, int cpu); 4996typedef const struct cpumask *(*sched_domain_mask_f)(int cpu); 4997 4998#define SDTL_OVERLAP 0x01 4999 5000struct sched_domain_topology_level { 5001 sched_domain_init_f init; 5002 sched_domain_mask_f mask; 5003 int flags; 5004 int numa_level; 5005 struct sd_data data; 5006}; 5007 5008/* 5009 * Build an iteration mask that can exclude certain CPUs from the upwards 5010 * domain traversal. 5011 * 5012 * Asymmetric node setups can result in situations where the domain tree is of 5013 * unequal depth, make sure to skip domains that already cover the entire 5014 * range. 5015 * 5016 * In that case build_sched_domains() will have terminated the iteration early 5017 * and our sibling sd spans will be empty. Domains should always include the 5018 * cpu they're built on, so check that. 5019 * 5020 */ 5021static void build_group_mask(struct sched_domain *sd, struct sched_group *sg) 5022{ 5023 const struct cpumask *span = sched_domain_span(sd); 5024 struct sd_data *sdd = sd->private; 5025 struct sched_domain *sibling; 5026 int i; 5027 5028 for_each_cpu(i, span) { 5029 sibling = *per_cpu_ptr(sdd->sd, i); 5030 if (!cpumask_test_cpu(i, sched_domain_span(sibling))) 5031 continue; 5032 5033 cpumask_set_cpu(i, sched_group_mask(sg)); 5034 } 5035} 5036 5037/* 5038 * Return the canonical balance cpu for this group, this is the first cpu 5039 * of this group that's also in the iteration mask. 5040 */ 5041int group_balance_cpu(struct sched_group *sg) 5042{ 5043 return cpumask_first_and(sched_group_cpus(sg), sched_group_mask(sg)); 5044} 5045 5046static int 5047build_overlap_sched_groups(struct sched_domain *sd, int cpu) 5048{ 5049 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg; 5050 const struct cpumask *span = sched_domain_span(sd); 5051 struct cpumask *covered = sched_domains_tmpmask; 5052 struct sd_data *sdd = sd->private; 5053 struct sched_domain *child; 5054 int i; 5055 5056 cpumask_clear(covered); 5057 5058 for_each_cpu(i, span) { 5059 struct cpumask *sg_span; 5060 5061 if (cpumask_test_cpu(i, covered)) 5062 continue; 5063 5064 child = *per_cpu_ptr(sdd->sd, i); 5065 5066 /* See the comment near build_group_mask(). */ 5067 if (!cpumask_test_cpu(i, sched_domain_span(child))) 5068 continue; 5069 5070 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(), 5071 GFP_KERNEL, cpu_to_node(cpu)); 5072 5073 if (!sg) 5074 goto fail; 5075 5076 sg_span = sched_group_cpus(sg); 5077 if (child->child) { 5078 child = child->child; 5079 cpumask_copy(sg_span, sched_domain_span(child)); 5080 } else 5081 cpumask_set_cpu(i, sg_span); 5082 5083 cpumask_or(covered, covered, sg_span); 5084 5085 sg->sgp = *per_cpu_ptr(sdd->sgp, i); 5086 if (atomic_inc_return(&sg->sgp->ref) == 1) 5087 build_group_mask(sd, sg); 5088 5089 /* 5090 * Initialize sgp->power such that even if we mess up the 5091 * domains and no possible iteration will get us here, we won't 5092 * die on a /0 trap. 5093 */ 5094 sg->sgp->power = SCHED_POWER_SCALE * cpumask_weight(sg_span); 5095 5096 /* 5097 * Make sure the first group of this domain contains the 5098 * canonical balance cpu. Otherwise the sched_domain iteration 5099 * breaks. See update_sg_lb_stats(). 5100 */ 5101 if ((!groups && cpumask_test_cpu(cpu, sg_span)) || 5102 group_balance_cpu(sg) == cpu) 5103 groups = sg; 5104 5105 if (!first) 5106 first = sg; 5107 if (last) 5108 last->next = sg; 5109 last = sg; 5110 last->next = first; 5111 } 5112 sd->groups = groups; 5113 5114 return 0; 5115 5116fail: 5117 free_sched_groups(first, 0); 5118 5119 return -ENOMEM; 5120} 5121 5122static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg) 5123{ 5124 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu); 5125 struct sched_domain *child = sd->child; 5126 5127 if (child) 5128 cpu = cpumask_first(sched_domain_span(child)); 5129 5130 if (sg) { 5131 *sg = *per_cpu_ptr(sdd->sg, cpu); 5132 (*sg)->sgp = *per_cpu_ptr(sdd->sgp, cpu); 5133 atomic_set(&(*sg)->sgp->ref, 1); /* for claim_allocations */ 5134 } 5135 5136 return cpu; 5137} 5138 5139/* 5140 * build_sched_groups will build a circular linked list of the groups 5141 * covered by the given span, and will set each group's ->cpumask correctly, 5142 * and ->cpu_power to 0. 5143 * 5144 * Assumes the sched_domain tree is fully constructed 5145 */ 5146static int 5147build_sched_groups(struct sched_domain *sd, int cpu) 5148{ 5149 struct sched_group *first = NULL, *last = NULL; 5150 struct sd_data *sdd = sd->private; 5151 const struct cpumask *span = sched_domain_span(sd); 5152 struct cpumask *covered; 5153 int i; 5154 5155 get_group(cpu, sdd, &sd->groups); 5156 atomic_inc(&sd->groups->ref); 5157 5158 if (cpu != cpumask_first(span)) 5159 return 0; 5160 5161 lockdep_assert_held(&sched_domains_mutex); 5162 covered = sched_domains_tmpmask; 5163 5164 cpumask_clear(covered); 5165 5166 for_each_cpu(i, span) { 5167 struct sched_group *sg; 5168 int group, j; 5169 5170 if (cpumask_test_cpu(i, covered)) 5171 continue; 5172 5173 group = get_group(i, sdd, &sg); 5174 cpumask_clear(sched_group_cpus(sg)); 5175 sg->sgp->power = 0; 5176 cpumask_setall(sched_group_mask(sg)); 5177 5178 for_each_cpu(j, span) { 5179 if (get_group(j, sdd, NULL) != group) 5180 continue; 5181 5182 cpumask_set_cpu(j, covered); 5183 cpumask_set_cpu(j, sched_group_cpus(sg)); 5184 } 5185 5186 if (!first) 5187 first = sg; 5188 if (last) 5189 last->next = sg; 5190 last = sg; 5191 } 5192 last->next = first; 5193 5194 return 0; 5195} 5196 5197/* 5198 * Initialize sched groups cpu_power. 5199 * 5200 * cpu_power indicates the capacity of sched group, which is used while 5201 * distributing the load between different sched groups in a sched domain. 5202 * Typically cpu_power for all the groups in a sched domain will be same unless 5203 * there are asymmetries in the topology. If there are asymmetries, group 5204 * having more cpu_power will pickup more load compared to the group having 5205 * less cpu_power. 5206 */ 5207static void init_sched_groups_power(int cpu, struct sched_domain *sd) 5208{ 5209 struct sched_group *sg = sd->groups; 5210 5211 WARN_ON(!sg); 5212 5213 do { 5214 sg->group_weight = cpumask_weight(sched_group_cpus(sg)); 5215 sg = sg->next; 5216 } while (sg != sd->groups); 5217 5218 if (cpu != group_balance_cpu(sg)) 5219 return; 5220 5221 update_group_power(sd, cpu); 5222 atomic_set(&sg->sgp->nr_busy_cpus, sg->group_weight); 5223} 5224 5225int __weak arch_sd_sibling_asym_packing(void) 5226{ 5227 return 0*SD_ASYM_PACKING; 5228} 5229 5230/* 5231 * Initializers for schedule domains 5232 * Non-inlined to reduce accumulated stack pressure in build_sched_domains() 5233 */ 5234 5235#ifdef CONFIG_SCHED_DEBUG 5236# define SD_INIT_NAME(sd, type) sd->name = #type 5237#else 5238# define SD_INIT_NAME(sd, type) do { } while (0) 5239#endif 5240 5241#define SD_INIT_FUNC(type) \ 5242static noinline struct sched_domain * \ 5243sd_init_##type(struct sched_domain_topology_level *tl, int cpu) \ 5244{ \ 5245 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); \ 5246 *sd = SD_##type##_INIT; \ 5247 SD_INIT_NAME(sd, type); \ 5248 sd->private = &tl->data; \ 5249 return sd; \ 5250} 5251 5252SD_INIT_FUNC(CPU) 5253#ifdef CONFIG_SCHED_SMT 5254 SD_INIT_FUNC(SIBLING) 5255#endif 5256#ifdef CONFIG_SCHED_MC 5257 SD_INIT_FUNC(MC) 5258#endif 5259#ifdef CONFIG_SCHED_BOOK 5260 SD_INIT_FUNC(BOOK) 5261#endif 5262 5263static int default_relax_domain_level = -1; 5264int sched_domain_level_max; 5265 5266static int __init setup_relax_domain_level(char *str) 5267{ 5268 if (kstrtoint(str, 0, &default_relax_domain_level)) 5269 pr_warn("Unable to set relax_domain_level\n"); 5270 5271 return 1; 5272} 5273__setup("relax_domain_level=", setup_relax_domain_level); 5274 5275static void set_domain_attribute(struct sched_domain *sd, 5276 struct sched_domain_attr *attr) 5277{ 5278 int request; 5279 5280 if (!attr || attr->relax_domain_level < 0) { 5281 if (default_relax_domain_level < 0) 5282 return; 5283 else 5284 request = default_relax_domain_level; 5285 } else 5286 request = attr->relax_domain_level; 5287 if (request < sd->level) { 5288 /* turn off idle balance on this domain */ 5289 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE); 5290 } else { 5291 /* turn on idle balance on this domain */ 5292 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE); 5293 } 5294} 5295 5296static void __sdt_free(const struct cpumask *cpu_map); 5297static int __sdt_alloc(const struct cpumask *cpu_map); 5298 5299static void __free_domain_allocs(struct s_data *d, enum s_alloc what, 5300 const struct cpumask *cpu_map) 5301{ 5302 switch (what) { 5303 case sa_rootdomain: 5304 if (!atomic_read(&d->rd->refcount)) 5305 free_rootdomain(&d->rd->rcu); /* fall through */ 5306 case sa_sd: 5307 free_percpu(d->sd); /* fall through */ 5308 case sa_sd_storage: 5309 __sdt_free(cpu_map); /* fall through */ 5310 case sa_none: 5311 break; 5312 } 5313} 5314 5315static enum s_alloc __visit_domain_allocation_hell(struct s_data *d, 5316 const struct cpumask *cpu_map) 5317{ 5318 memset(d, 0, sizeof(*d)); 5319 5320 if (__sdt_alloc(cpu_map)) 5321 return sa_sd_storage; 5322 d->sd = alloc_percpu(struct sched_domain *); 5323 if (!d->sd) 5324 return sa_sd_storage; 5325 d->rd = alloc_rootdomain(); 5326 if (!d->rd) 5327 return sa_sd; 5328 return sa_rootdomain; 5329} 5330 5331/* 5332 * NULL the sd_data elements we've used to build the sched_domain and 5333 * sched_group structure so that the subsequent __free_domain_allocs() 5334 * will not free the data we're using. 5335 */ 5336static void claim_allocations(int cpu, struct sched_domain *sd) 5337{ 5338 struct sd_data *sdd = sd->private; 5339 5340 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd); 5341 *per_cpu_ptr(sdd->sd, cpu) = NULL; 5342 5343 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref)) 5344 *per_cpu_ptr(sdd->sg, cpu) = NULL; 5345 5346 if (atomic_read(&(*per_cpu_ptr(sdd->sgp, cpu))->ref)) 5347 *per_cpu_ptr(sdd->sgp, cpu) = NULL; 5348} 5349 5350#ifdef CONFIG_SCHED_SMT 5351static const struct cpumask *cpu_smt_mask(int cpu) 5352{ 5353 return topology_thread_cpumask(cpu); 5354} 5355#endif 5356 5357/* 5358 * Topology list, bottom-up. 5359 */ 5360static struct sched_domain_topology_level default_topology[] = { 5361#ifdef CONFIG_SCHED_SMT 5362 { sd_init_SIBLING, cpu_smt_mask, }, 5363#endif 5364#ifdef CONFIG_SCHED_MC 5365 { sd_init_MC, cpu_coregroup_mask, }, 5366#endif 5367#ifdef CONFIG_SCHED_BOOK 5368 { sd_init_BOOK, cpu_book_mask, }, 5369#endif 5370 { sd_init_CPU, cpu_cpu_mask, }, 5371 { NULL, }, 5372}; 5373 5374static struct sched_domain_topology_level *sched_domain_topology = default_topology; 5375 5376#define for_each_sd_topology(tl) \ 5377 for (tl = sched_domain_topology; tl->init; tl++) 5378 5379#ifdef CONFIG_NUMA 5380 5381static int sched_domains_numa_levels; 5382static int *sched_domains_numa_distance; 5383static struct cpumask ***sched_domains_numa_masks; 5384static int sched_domains_curr_level; 5385 5386static inline int sd_local_flags(int level) 5387{ 5388 if (sched_domains_numa_distance[level] > RECLAIM_DISTANCE) 5389 return 0; 5390 5391 return SD_BALANCE_EXEC | SD_BALANCE_FORK | SD_WAKE_AFFINE; 5392} 5393 5394static struct sched_domain * 5395sd_numa_init(struct sched_domain_topology_level *tl, int cpu) 5396{ 5397 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); 5398 int level = tl->numa_level; 5399 int sd_weight = cpumask_weight( 5400 sched_domains_numa_masks[level][cpu_to_node(cpu)]); 5401 5402 *sd = (struct sched_domain){ 5403 .min_interval = sd_weight, 5404 .max_interval = 2*sd_weight, 5405 .busy_factor = 32, 5406 .imbalance_pct = 125, 5407 .cache_nice_tries = 2, 5408 .busy_idx = 3, 5409 .idle_idx = 2, 5410 .newidle_idx = 0, 5411 .wake_idx = 0, 5412 .forkexec_idx = 0, 5413 5414 .flags = 1*SD_LOAD_BALANCE 5415 | 1*SD_BALANCE_NEWIDLE 5416 | 0*SD_BALANCE_EXEC 5417 | 0*SD_BALANCE_FORK 5418 | 0*SD_BALANCE_WAKE 5419 | 0*SD_WAKE_AFFINE 5420 | 0*SD_SHARE_CPUPOWER 5421 | 0*SD_SHARE_PKG_RESOURCES 5422 | 1*SD_SERIALIZE 5423 | 0*SD_PREFER_SIBLING 5424 | 1*SD_NUMA 5425 | sd_local_flags(level) 5426 , 5427 .last_balance = jiffies, 5428 .balance_interval = sd_weight, 5429 }; 5430 SD_INIT_NAME(sd, NUMA); 5431 sd->private = &tl->data; 5432 5433 /* 5434 * Ugly hack to pass state to sd_numa_mask()... 5435 */ 5436 sched_domains_curr_level = tl->numa_level; 5437 5438 return sd; 5439} 5440 5441static const struct cpumask *sd_numa_mask(int cpu) 5442{ 5443 return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)]; 5444} 5445 5446static void sched_numa_warn(const char *str) 5447{ 5448 static int done = false; 5449 int i,j; 5450 5451 if (done) 5452 return; 5453 5454 done = true; 5455 5456 printk(KERN_WARNING "ERROR: %s\n\n", str); 5457 5458 for (i = 0; i < nr_node_ids; i++) { 5459 printk(KERN_WARNING " "); 5460 for (j = 0; j < nr_node_ids; j++) 5461 printk(KERN_CONT "%02d ", node_distance(i,j)); 5462 printk(KERN_CONT "\n"); 5463 } 5464 printk(KERN_WARNING "\n"); 5465} 5466 5467static bool find_numa_distance(int distance) 5468{ 5469 int i; 5470 5471 if (distance == node_distance(0, 0)) 5472 return true; 5473 5474 for (i = 0; i < sched_domains_numa_levels; i++) { 5475 if (sched_domains_numa_distance[i] == distance) 5476 return true; 5477 } 5478 5479 return false; 5480} 5481 5482static void sched_init_numa(void) 5483{ 5484 int next_distance, curr_distance = node_distance(0, 0); 5485 struct sched_domain_topology_level *tl; 5486 int level = 0; 5487 int i, j, k; 5488 5489 sched_domains_numa_distance = kzalloc(sizeof(int) * nr_node_ids, GFP_KERNEL); 5490 if (!sched_domains_numa_distance) 5491 return; 5492 5493 /* 5494 * O(nr_nodes^2) deduplicating selection sort -- in order to find the 5495 * unique distances in the node_distance() table. 5496 * 5497 * Assumes node_distance(0,j) includes all distances in 5498 * node_distance(i,j) in order to avoid cubic time. 5499 */ 5500 next_distance = curr_distance; 5501 for (i = 0; i < nr_node_ids; i++) { 5502 for (j = 0; j < nr_node_ids; j++) { 5503 for (k = 0; k < nr_node_ids; k++) { 5504 int distance = node_distance(i, k); 5505 5506 if (distance > curr_distance && 5507 (distance < next_distance || 5508 next_distance == curr_distance)) 5509 next_distance = distance; 5510 5511 /* 5512 * While not a strong assumption it would be nice to know 5513 * about cases where if node A is connected to B, B is not 5514 * equally connected to A. 5515 */ 5516 if (sched_debug() && node_distance(k, i) != distance) 5517 sched_numa_warn("Node-distance not symmetric"); 5518 5519 if (sched_debug() && i && !find_numa_distance(distance)) 5520 sched_numa_warn("Node-0 not representative"); 5521 } 5522 if (next_distance != curr_distance) { 5523 sched_domains_numa_distance[level++] = next_distance; 5524 sched_domains_numa_levels = level; 5525 curr_distance = next_distance; 5526 } else break; 5527 } 5528 5529 /* 5530 * In case of sched_debug() we verify the above assumption. 5531 */ 5532 if (!sched_debug()) 5533 break; 5534 } 5535 /* 5536 * 'level' contains the number of unique distances, excluding the 5537 * identity distance node_distance(i,i). 5538 * 5539 * The sched_domains_numa_distance[] array includes the actual distance 5540 * numbers. 5541 */ 5542 5543 /* 5544 * Here, we should temporarily reset sched_domains_numa_levels to 0. 5545 * If it fails to allocate memory for array sched_domains_numa_masks[][], 5546 * the array will contain less then 'level' members. This could be 5547 * dangerous when we use it to iterate array sched_domains_numa_masks[][] 5548 * in other functions. 5549 * 5550 * We reset it to 'level' at the end of this function. 5551 */ 5552 sched_domains_numa_levels = 0; 5553 5554 sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL); 5555 if (!sched_domains_numa_masks) 5556 return; 5557 5558 /* 5559 * Now for each level, construct a mask per node which contains all 5560 * cpus of nodes that are that many hops away from us. 5561 */ 5562 for (i = 0; i < level; i++) { 5563 sched_domains_numa_masks[i] = 5564 kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL); 5565 if (!sched_domains_numa_masks[i]) 5566 return; 5567 5568 for (j = 0; j < nr_node_ids; j++) { 5569 struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL); 5570 if (!mask) 5571 return; 5572 5573 sched_domains_numa_masks[i][j] = mask; 5574 5575 for (k = 0; k < nr_node_ids; k++) { 5576 if (node_distance(j, k) > sched_domains_numa_distance[i]) 5577 continue; 5578 5579 cpumask_or(mask, mask, cpumask_of_node(k)); 5580 } 5581 } 5582 } 5583 5584 tl = kzalloc((ARRAY_SIZE(default_topology) + level) * 5585 sizeof(struct sched_domain_topology_level), GFP_KERNEL); 5586 if (!tl) 5587 return; 5588 5589 /* 5590 * Copy the default topology bits.. 5591 */ 5592 for (i = 0; default_topology[i].init; i++) 5593 tl[i] = default_topology[i]; 5594 5595 /* 5596 * .. and append 'j' levels of NUMA goodness. 5597 */ 5598 for (j = 0; j < level; i++, j++) { 5599 tl[i] = (struct sched_domain_topology_level){ 5600 .init = sd_numa_init, 5601 .mask = sd_numa_mask, 5602 .flags = SDTL_OVERLAP, 5603 .numa_level = j, 5604 }; 5605 } 5606 5607 sched_domain_topology = tl; 5608 5609 sched_domains_numa_levels = level; 5610} 5611 5612static void sched_domains_numa_masks_set(int cpu) 5613{ 5614 int i, j; 5615 int node = cpu_to_node(cpu); 5616 5617 for (i = 0; i < sched_domains_numa_levels; i++) { 5618 for (j = 0; j < nr_node_ids; j++) { 5619 if (node_distance(j, node) <= sched_domains_numa_distance[i]) 5620 cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]); 5621 } 5622 } 5623} 5624 5625static void sched_domains_numa_masks_clear(int cpu) 5626{ 5627 int i, j; 5628 for (i = 0; i < sched_domains_numa_levels; i++) { 5629 for (j = 0; j < nr_node_ids; j++) 5630 cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]); 5631 } 5632} 5633 5634/* 5635 * Update sched_domains_numa_masks[level][node] array when new cpus 5636 * are onlined. 5637 */ 5638static int sched_domains_numa_masks_update(struct notifier_block *nfb, 5639 unsigned long action, 5640 void *hcpu) 5641{ 5642 int cpu = (long)hcpu; 5643 5644 switch (action & ~CPU_TASKS_FROZEN) { 5645 case CPU_ONLINE: 5646 sched_domains_numa_masks_set(cpu); 5647 break; 5648 5649 case CPU_DEAD: 5650 sched_domains_numa_masks_clear(cpu); 5651 break; 5652 5653 default: 5654 return NOTIFY_DONE; 5655 } 5656 5657 return NOTIFY_OK; 5658} 5659#else 5660static inline void sched_init_numa(void) 5661{ 5662} 5663 5664static int sched_domains_numa_masks_update(struct notifier_block *nfb, 5665 unsigned long action, 5666 void *hcpu) 5667{ 5668 return 0; 5669} 5670#endif /* CONFIG_NUMA */ 5671 5672static int __sdt_alloc(const struct cpumask *cpu_map) 5673{ 5674 struct sched_domain_topology_level *tl; 5675 int j; 5676 5677 for_each_sd_topology(tl) { 5678 struct sd_data *sdd = &tl->data; 5679 5680 sdd->sd = alloc_percpu(struct sched_domain *); 5681 if (!sdd->sd) 5682 return -ENOMEM; 5683 5684 sdd->sg = alloc_percpu(struct sched_group *); 5685 if (!sdd->sg) 5686 return -ENOMEM; 5687 5688 sdd->sgp = alloc_percpu(struct sched_group_power *); 5689 if (!sdd->sgp) 5690 return -ENOMEM; 5691 5692 for_each_cpu(j, cpu_map) { 5693 struct sched_domain *sd; 5694 struct sched_group *sg; 5695 struct sched_group_power *sgp; 5696 5697 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(), 5698 GFP_KERNEL, cpu_to_node(j)); 5699 if (!sd) 5700 return -ENOMEM; 5701 5702 *per_cpu_ptr(sdd->sd, j) = sd; 5703 5704 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(), 5705 GFP_KERNEL, cpu_to_node(j)); 5706 if (!sg) 5707 return -ENOMEM; 5708 5709 sg->next = sg; 5710 5711 *per_cpu_ptr(sdd->sg, j) = sg; 5712 5713 sgp = kzalloc_node(sizeof(struct sched_group_power) + cpumask_size(), 5714 GFP_KERNEL, cpu_to_node(j)); 5715 if (!sgp) 5716 return -ENOMEM; 5717 5718 *per_cpu_ptr(sdd->sgp, j) = sgp; 5719 } 5720 } 5721 5722 return 0; 5723} 5724 5725static void __sdt_free(const struct cpumask *cpu_map) 5726{ 5727 struct sched_domain_topology_level *tl; 5728 int j; 5729 5730 for_each_sd_topology(tl) { 5731 struct sd_data *sdd = &tl->data; 5732 5733 for_each_cpu(j, cpu_map) { 5734 struct sched_domain *sd; 5735 5736 if (sdd->sd) { 5737 sd = *per_cpu_ptr(sdd->sd, j); 5738 if (sd && (sd->flags & SD_OVERLAP)) 5739 free_sched_groups(sd->groups, 0); 5740 kfree(*per_cpu_ptr(sdd->sd, j)); 5741 } 5742 5743 if (sdd->sg) 5744 kfree(*per_cpu_ptr(sdd->sg, j)); 5745 if (sdd->sgp) 5746 kfree(*per_cpu_ptr(sdd->sgp, j)); 5747 } 5748 free_percpu(sdd->sd); 5749 sdd->sd = NULL; 5750 free_percpu(sdd->sg); 5751 sdd->sg = NULL; 5752 free_percpu(sdd->sgp); 5753 sdd->sgp = NULL; 5754 } 5755} 5756 5757struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl, 5758 const struct cpumask *cpu_map, struct sched_domain_attr *attr, 5759 struct sched_domain *child, int cpu) 5760{ 5761 struct sched_domain *sd = tl->init(tl, cpu); 5762 if (!sd) 5763 return child; 5764 5765 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu)); 5766 if (child) { 5767 sd->level = child->level + 1; 5768 sched_domain_level_max = max(sched_domain_level_max, sd->level); 5769 child->parent = sd; 5770 sd->child = child; 5771 } 5772 set_domain_attribute(sd, attr); 5773 5774 return sd; 5775} 5776 5777/* 5778 * Build sched domains for a given set of cpus and attach the sched domains 5779 * to the individual cpus 5780 */ 5781static int build_sched_domains(const struct cpumask *cpu_map, 5782 struct sched_domain_attr *attr) 5783{ 5784 enum s_alloc alloc_state; 5785 struct sched_domain *sd; 5786 struct s_data d; 5787 int i, ret = -ENOMEM; 5788 5789 alloc_state = __visit_domain_allocation_hell(&d, cpu_map); 5790 if (alloc_state != sa_rootdomain) 5791 goto error; 5792 5793 /* Set up domains for cpus specified by the cpu_map. */ 5794 for_each_cpu(i, cpu_map) { 5795 struct sched_domain_topology_level *tl; 5796 5797 sd = NULL; 5798 for_each_sd_topology(tl) { 5799 sd = build_sched_domain(tl, cpu_map, attr, sd, i); 5800 if (tl == sched_domain_topology) 5801 *per_cpu_ptr(d.sd, i) = sd; 5802 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP)) 5803 sd->flags |= SD_OVERLAP; 5804 if (cpumask_equal(cpu_map, sched_domain_span(sd))) 5805 break; 5806 } 5807 } 5808 5809 /* Build the groups for the domains */ 5810 for_each_cpu(i, cpu_map) { 5811 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) { 5812 sd->span_weight = cpumask_weight(sched_domain_span(sd)); 5813 if (sd->flags & SD_OVERLAP) { 5814 if (build_overlap_sched_groups(sd, i)) 5815 goto error; 5816 } else { 5817 if (build_sched_groups(sd, i)) 5818 goto error; 5819 } 5820 } 5821 } 5822 5823 /* Calculate CPU power for physical packages and nodes */ 5824 for (i = nr_cpumask_bits-1; i >= 0; i--) { 5825 if (!cpumask_test_cpu(i, cpu_map)) 5826 continue; 5827 5828 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) { 5829 claim_allocations(i, sd); 5830 init_sched_groups_power(i, sd); 5831 } 5832 } 5833 5834 /* Attach the domains */ 5835 rcu_read_lock(); 5836 for_each_cpu(i, cpu_map) { 5837 sd = *per_cpu_ptr(d.sd, i); 5838 cpu_attach_domain(sd, d.rd, i); 5839 } 5840 rcu_read_unlock(); 5841 5842 ret = 0; 5843error: 5844 __free_domain_allocs(&d, alloc_state, cpu_map); 5845 return ret; 5846} 5847 5848static cpumask_var_t *doms_cur; /* current sched domains */ 5849static int ndoms_cur; /* number of sched domains in 'doms_cur' */ 5850static struct sched_domain_attr *dattr_cur; 5851 /* attribues of custom domains in 'doms_cur' */ 5852 5853/* 5854 * Special case: If a kmalloc of a doms_cur partition (array of 5855 * cpumask) fails, then fallback to a single sched domain, 5856 * as determined by the single cpumask fallback_doms. 5857 */ 5858static cpumask_var_t fallback_doms; 5859 5860/* 5861 * arch_update_cpu_topology lets virtualized architectures update the 5862 * cpu core maps. It is supposed to return 1 if the topology changed 5863 * or 0 if it stayed the same. 5864 */ 5865int __attribute__((weak)) arch_update_cpu_topology(void) 5866{ 5867 return 0; 5868} 5869 5870cpumask_var_t *alloc_sched_domains(unsigned int ndoms) 5871{ 5872 int i; 5873 cpumask_var_t *doms; 5874 5875 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL); 5876 if (!doms) 5877 return NULL; 5878 for (i = 0; i < ndoms; i++) { 5879 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) { 5880 free_sched_domains(doms, i); 5881 return NULL; 5882 } 5883 } 5884 return doms; 5885} 5886 5887void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms) 5888{ 5889 unsigned int i; 5890 for (i = 0; i < ndoms; i++) 5891 free_cpumask_var(doms[i]); 5892 kfree(doms); 5893} 5894 5895/* 5896 * Set up scheduler domains and groups. Callers must hold the hotplug lock. 5897 * For now this just excludes isolated cpus, but could be used to 5898 * exclude other special cases in the future. 5899 */ 5900static int init_sched_domains(const struct cpumask *cpu_map) 5901{ 5902 int err; 5903 5904 arch_update_cpu_topology(); 5905 ndoms_cur = 1; 5906 doms_cur = alloc_sched_domains(ndoms_cur); 5907 if (!doms_cur) 5908 doms_cur = &fallback_doms; 5909 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map); 5910 err = build_sched_domains(doms_cur[0], NULL); 5911 register_sched_domain_sysctl(); 5912 5913 return err; 5914} 5915 5916/* 5917 * Detach sched domains from a group of cpus specified in cpu_map 5918 * These cpus will now be attached to the NULL domain 5919 */ 5920static void detach_destroy_domains(const struct cpumask *cpu_map) 5921{ 5922 int i; 5923 5924 rcu_read_lock(); 5925 for_each_cpu(i, cpu_map) 5926 cpu_attach_domain(NULL, &def_root_domain, i); 5927 rcu_read_unlock(); 5928} 5929 5930/* handle null as "default" */ 5931static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur, 5932 struct sched_domain_attr *new, int idx_new) 5933{ 5934 struct sched_domain_attr tmp; 5935 5936 /* fast path */ 5937 if (!new && !cur) 5938 return 1; 5939 5940 tmp = SD_ATTR_INIT; 5941 return !memcmp(cur ? (cur + idx_cur) : &tmp, 5942 new ? (new + idx_new) : &tmp, 5943 sizeof(struct sched_domain_attr)); 5944} 5945 5946/* 5947 * Partition sched domains as specified by the 'ndoms_new' 5948 * cpumasks in the array doms_new[] of cpumasks. This compares 5949 * doms_new[] to the current sched domain partitioning, doms_cur[]. 5950 * It destroys each deleted domain and builds each new domain. 5951 * 5952 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'. 5953 * The masks don't intersect (don't overlap.) We should setup one 5954 * sched domain for each mask. CPUs not in any of the cpumasks will 5955 * not be load balanced. If the same cpumask appears both in the 5956 * current 'doms_cur' domains and in the new 'doms_new', we can leave 5957 * it as it is. 5958 * 5959 * The passed in 'doms_new' should be allocated using 5960 * alloc_sched_domains. This routine takes ownership of it and will 5961 * free_sched_domains it when done with it. If the caller failed the 5962 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1, 5963 * and partition_sched_domains() will fallback to the single partition 5964 * 'fallback_doms', it also forces the domains to be rebuilt. 5965 * 5966 * If doms_new == NULL it will be replaced with cpu_online_mask. 5967 * ndoms_new == 0 is a special case for destroying existing domains, 5968 * and it will not create the default domain. 5969 * 5970 * Call with hotplug lock held 5971 */ 5972void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[], 5973 struct sched_domain_attr *dattr_new) 5974{ 5975 int i, j, n; 5976 int new_topology; 5977 5978 mutex_lock(&sched_domains_mutex); 5979 5980 /* always unregister in case we don't destroy any domains */ 5981 unregister_sched_domain_sysctl(); 5982 5983 /* Let architecture update cpu core mappings. */ 5984 new_topology = arch_update_cpu_topology(); 5985 5986 n = doms_new ? ndoms_new : 0; 5987 5988 /* Destroy deleted domains */ 5989 for (i = 0; i < ndoms_cur; i++) { 5990 for (j = 0; j < n && !new_topology; j++) { 5991 if (cpumask_equal(doms_cur[i], doms_new[j]) 5992 && dattrs_equal(dattr_cur, i, dattr_new, j)) 5993 goto match1; 5994 } 5995 /* no match - a current sched domain not in new doms_new[] */ 5996 detach_destroy_domains(doms_cur[i]); 5997match1: 5998 ; 5999 } 6000 6001 n = ndoms_cur; 6002 if (doms_new == NULL) { 6003 n = 0; 6004 doms_new = &fallback_doms; 6005 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map); 6006 WARN_ON_ONCE(dattr_new); 6007 } 6008 6009 /* Build new domains */ 6010 for (i = 0; i < ndoms_new; i++) { 6011 for (j = 0; j < n && !new_topology; j++) { 6012 if (cpumask_equal(doms_new[i], doms_cur[j]) 6013 && dattrs_equal(dattr_new, i, dattr_cur, j)) 6014 goto match2; 6015 } 6016 /* no match - add a new doms_new */ 6017 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL); 6018match2: 6019 ; 6020 } 6021 6022 /* Remember the new sched domains */ 6023 if (doms_cur != &fallback_doms) 6024 free_sched_domains(doms_cur, ndoms_cur); 6025 kfree(dattr_cur); /* kfree(NULL) is safe */ 6026 doms_cur = doms_new; 6027 dattr_cur = dattr_new; 6028 ndoms_cur = ndoms_new; 6029 6030 register_sched_domain_sysctl(); 6031 6032 mutex_unlock(&sched_domains_mutex); 6033} 6034 6035static int num_cpus_frozen; /* used to mark begin/end of suspend/resume */ 6036 6037/* 6038 * Update cpusets according to cpu_active mask. If cpusets are 6039 * disabled, cpuset_update_active_cpus() becomes a simple wrapper 6040 * around partition_sched_domains(). 6041 * 6042 * If we come here as part of a suspend/resume, don't touch cpusets because we 6043 * want to restore it back to its original state upon resume anyway. 6044 */ 6045static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action, 6046 void *hcpu) 6047{ 6048 switch (action) { 6049 case CPU_ONLINE_FROZEN: 6050 case CPU_DOWN_FAILED_FROZEN: 6051 6052 /* 6053 * num_cpus_frozen tracks how many CPUs are involved in suspend 6054 * resume sequence. As long as this is not the last online 6055 * operation in the resume sequence, just build a single sched 6056 * domain, ignoring cpusets. 6057 */ 6058 num_cpus_frozen--; 6059 if (likely(num_cpus_frozen)) { 6060 partition_sched_domains(1, NULL, NULL); 6061 break; 6062 } 6063 6064 /* 6065 * This is the last CPU online operation. So fall through and 6066 * restore the original sched domains by considering the 6067 * cpuset configurations. 6068 */ 6069 6070 case CPU_ONLINE: 6071 case CPU_DOWN_FAILED: 6072 cpuset_update_active_cpus(true); 6073 break; 6074 default: 6075 return NOTIFY_DONE; 6076 } 6077 return NOTIFY_OK; 6078} 6079 6080static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action, 6081 void *hcpu) 6082{ 6083 switch (action) { 6084 case CPU_DOWN_PREPARE: 6085 cpuset_update_active_cpus(false); 6086 break; 6087 case CPU_DOWN_PREPARE_FROZEN: 6088 num_cpus_frozen++; 6089 partition_sched_domains(1, NULL, NULL); 6090 break; 6091 default: 6092 return NOTIFY_DONE; 6093 } 6094 return NOTIFY_OK; 6095} 6096 6097void __init sched_init_smp(void) 6098{ 6099 cpumask_var_t non_isolated_cpus; 6100 6101 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL); 6102 alloc_cpumask_var(&fallback_doms, GFP_KERNEL); 6103 6104 sched_init_numa(); 6105 6106 /* 6107 * There's no userspace yet to cause hotplug operations; hence all the 6108 * cpu masks are stable and all blatant races in the below code cannot 6109 * happen. 6110 */ 6111 mutex_lock(&sched_domains_mutex); 6112 init_sched_domains(cpu_active_mask); 6113 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map); 6114 if (cpumask_empty(non_isolated_cpus)) 6115 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus); 6116 mutex_unlock(&sched_domains_mutex); 6117 6118 hotcpu_notifier(sched_domains_numa_masks_update, CPU_PRI_SCHED_ACTIVE); 6119 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE); 6120 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE); 6121 6122 init_hrtick(); 6123 6124 /* Move init over to a non-isolated CPU */ 6125 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0) 6126 BUG(); 6127 sched_init_granularity(); 6128 free_cpumask_var(non_isolated_cpus); 6129 6130 init_sched_rt_class(); 6131} 6132#else 6133void __init sched_init_smp(void) 6134{ 6135 sched_init_granularity(); 6136} 6137#endif /* CONFIG_SMP */ 6138 6139const_debug unsigned int sysctl_timer_migration = 1; 6140 6141int in_sched_functions(unsigned long addr) 6142{ 6143 return in_lock_functions(addr) || 6144 (addr >= (unsigned long)__sched_text_start 6145 && addr < (unsigned long)__sched_text_end); 6146} 6147 6148#ifdef CONFIG_CGROUP_SCHED 6149/* 6150 * Default task group. 6151 * Every task in system belongs to this group at bootup. 6152 */ 6153struct task_group root_task_group; 6154LIST_HEAD(task_groups); 6155#endif 6156 6157DECLARE_PER_CPU(cpumask_var_t, load_balance_mask); 6158 6159void __init sched_init(void) 6160{ 6161 int i, j; 6162 unsigned long alloc_size = 0, ptr; 6163 6164#ifdef CONFIG_FAIR_GROUP_SCHED 6165 alloc_size += 2 * nr_cpu_ids * sizeof(void **); 6166#endif 6167#ifdef CONFIG_RT_GROUP_SCHED 6168 alloc_size += 2 * nr_cpu_ids * sizeof(void **); 6169#endif 6170#ifdef CONFIG_CPUMASK_OFFSTACK 6171 alloc_size += num_possible_cpus() * cpumask_size(); 6172#endif 6173 if (alloc_size) { 6174 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT); 6175 6176#ifdef CONFIG_FAIR_GROUP_SCHED 6177 root_task_group.se = (struct sched_entity **)ptr; 6178 ptr += nr_cpu_ids * sizeof(void **); 6179 6180 root_task_group.cfs_rq = (struct cfs_rq **)ptr; 6181 ptr += nr_cpu_ids * sizeof(void **); 6182 6183#endif /* CONFIG_FAIR_GROUP_SCHED */ 6184#ifdef CONFIG_RT_GROUP_SCHED 6185 root_task_group.rt_se = (struct sched_rt_entity **)ptr; 6186 ptr += nr_cpu_ids * sizeof(void **); 6187 6188 root_task_group.rt_rq = (struct rt_rq **)ptr; 6189 ptr += nr_cpu_ids * sizeof(void **); 6190 6191#endif /* CONFIG_RT_GROUP_SCHED */ 6192#ifdef CONFIG_CPUMASK_OFFSTACK 6193 for_each_possible_cpu(i) { 6194 per_cpu(load_balance_mask, i) = (void *)ptr; 6195 ptr += cpumask_size(); 6196 } 6197#endif /* CONFIG_CPUMASK_OFFSTACK */ 6198 } 6199 6200#ifdef CONFIG_SMP 6201 init_defrootdomain(); 6202#endif 6203 6204 init_rt_bandwidth(&def_rt_bandwidth, 6205 global_rt_period(), global_rt_runtime()); 6206 6207#ifdef CONFIG_RT_GROUP_SCHED 6208 init_rt_bandwidth(&root_task_group.rt_bandwidth, 6209 global_rt_period(), global_rt_runtime()); 6210#endif /* CONFIG_RT_GROUP_SCHED */ 6211 6212#ifdef CONFIG_CGROUP_SCHED 6213 list_add(&root_task_group.list, &task_groups); 6214 INIT_LIST_HEAD(&root_task_group.children); 6215 INIT_LIST_HEAD(&root_task_group.siblings); 6216 autogroup_init(&init_task); 6217 6218#endif /* CONFIG_CGROUP_SCHED */ 6219 6220 for_each_possible_cpu(i) { 6221 struct rq *rq; 6222 6223 rq = cpu_rq(i); 6224 raw_spin_lock_init(&rq->lock); 6225 rq->nr_running = 0; 6226 rq->calc_load_active = 0; 6227 rq->calc_load_update = jiffies + LOAD_FREQ; 6228 init_cfs_rq(&rq->cfs); 6229 init_rt_rq(&rq->rt, rq); 6230#ifdef CONFIG_FAIR_GROUP_SCHED 6231 root_task_group.shares = ROOT_TASK_GROUP_LOAD; 6232 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list); 6233 /* 6234 * How much cpu bandwidth does root_task_group get? 6235 * 6236 * In case of task-groups formed thr' the cgroup filesystem, it 6237 * gets 100% of the cpu resources in the system. This overall 6238 * system cpu resource is divided among the tasks of 6239 * root_task_group and its child task-groups in a fair manner, 6240 * based on each entity's (task or task-group's) weight 6241 * (se->load.weight). 6242 * 6243 * In other words, if root_task_group has 10 tasks of weight 6244 * 1024) and two child groups A0 and A1 (of weight 1024 each), 6245 * then A0's share of the cpu resource is: 6246 * 6247 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33% 6248 * 6249 * We achieve this by letting root_task_group's tasks sit 6250 * directly in rq->cfs (i.e root_task_group->se[] = NULL). 6251 */ 6252 init_cfs_bandwidth(&root_task_group.cfs_bandwidth); 6253 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL); 6254#endif /* CONFIG_FAIR_GROUP_SCHED */ 6255 6256 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime; 6257#ifdef CONFIG_RT_GROUP_SCHED 6258 INIT_LIST_HEAD(&rq->leaf_rt_rq_list); 6259 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL); 6260#endif 6261 6262 for (j = 0; j < CPU_LOAD_IDX_MAX; j++) 6263 rq->cpu_load[j] = 0; 6264 6265 rq->last_load_update_tick = jiffies; 6266 6267#ifdef CONFIG_SMP 6268 rq->sd = NULL; 6269 rq->rd = NULL; 6270 rq->cpu_power = SCHED_POWER_SCALE; 6271 rq->post_schedule = 0; 6272 rq->active_balance = 0; 6273 rq->next_balance = jiffies; 6274 rq->push_cpu = 0; 6275 rq->cpu = i; 6276 rq->online = 0; 6277 rq->idle_stamp = 0; 6278 rq->avg_idle = 2*sysctl_sched_migration_cost; 6279 rq->max_idle_balance_cost = sysctl_sched_migration_cost; 6280 6281 INIT_LIST_HEAD(&rq->cfs_tasks); 6282 6283 rq_attach_root(rq, &def_root_domain); 6284#ifdef CONFIG_NO_HZ_COMMON 6285 rq->nohz_flags = 0; 6286#endif 6287#ifdef CONFIG_NO_HZ_FULL 6288 rq->last_sched_tick = 0; 6289#endif 6290#endif 6291 init_rq_hrtick(rq); 6292 atomic_set(&rq->nr_iowait, 0); 6293 } 6294 6295 set_load_weight(&init_task); 6296 6297#ifdef CONFIG_PREEMPT_NOTIFIERS 6298 INIT_HLIST_HEAD(&init_task.preempt_notifiers); 6299#endif 6300 6301#ifdef CONFIG_RT_MUTEXES 6302 plist_head_init(&init_task.pi_waiters); 6303#endif 6304 6305 /* 6306 * The boot idle thread does lazy MMU switching as well: 6307 */ 6308 atomic_inc(&init_mm.mm_count); 6309 enter_lazy_tlb(&init_mm, current); 6310 6311 /* 6312 * Make us the idle thread. Technically, schedule() should not be 6313 * called from this thread, however somewhere below it might be, 6314 * but because we are the idle thread, we just pick up running again 6315 * when this runqueue becomes "idle". 6316 */ 6317 init_idle(current, smp_processor_id()); 6318 6319 calc_load_update = jiffies + LOAD_FREQ; 6320 6321 /* 6322 * During early bootup we pretend to be a normal task: 6323 */ 6324 current->sched_class = &fair_sched_class; 6325 6326#ifdef CONFIG_SMP 6327 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT); 6328 /* May be allocated at isolcpus cmdline parse time */ 6329 if (cpu_isolated_map == NULL) 6330 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT); 6331 idle_thread_set_boot_cpu(); 6332#endif 6333 init_sched_fair_class(); 6334 6335 scheduler_running = 1; 6336} 6337 6338#ifdef CONFIG_DEBUG_ATOMIC_SLEEP 6339static inline int preempt_count_equals(int preempt_offset) 6340{ 6341 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth(); 6342 6343 return (nested == preempt_offset); 6344} 6345 6346void __might_sleep(const char *file, int line, int preempt_offset) 6347{ 6348 static unsigned long prev_jiffy; /* ratelimiting */ 6349 6350 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */ 6351 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) || 6352 system_state != SYSTEM_RUNNING || oops_in_progress) 6353 return; 6354 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy) 6355 return; 6356 prev_jiffy = jiffies; 6357 6358 printk(KERN_ERR 6359 "BUG: sleeping function called from invalid context at %s:%d\n", 6360 file, line); 6361 printk(KERN_ERR 6362 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n", 6363 in_atomic(), irqs_disabled(), 6364 current->pid, current->comm); 6365 6366 debug_show_held_locks(current); 6367 if (irqs_disabled()) 6368 print_irqtrace_events(current); 6369 dump_stack(); 6370} 6371EXPORT_SYMBOL(__might_sleep); 6372#endif 6373 6374#ifdef CONFIG_MAGIC_SYSRQ 6375static void normalize_task(struct rq *rq, struct task_struct *p) 6376{ 6377 const struct sched_class *prev_class = p->sched_class; 6378 int old_prio = p->prio; 6379 int on_rq; 6380 6381 on_rq = p->on_rq; 6382 if (on_rq) 6383 dequeue_task(rq, p, 0); 6384 __setscheduler(rq, p, SCHED_NORMAL, 0); 6385 if (on_rq) { 6386 enqueue_task(rq, p, 0); 6387 resched_task(rq->curr); 6388 } 6389 6390 check_class_changed(rq, p, prev_class, old_prio); 6391} 6392 6393void normalize_rt_tasks(void) 6394{ 6395 struct task_struct *g, *p; 6396 unsigned long flags; 6397 struct rq *rq; 6398 6399 read_lock_irqsave(&tasklist_lock, flags); 6400 do_each_thread(g, p) { 6401 /* 6402 * Only normalize user tasks: 6403 */ 6404 if (!p->mm) 6405 continue; 6406 6407 p->se.exec_start = 0; 6408#ifdef CONFIG_SCHEDSTATS 6409 p->se.statistics.wait_start = 0; 6410 p->se.statistics.sleep_start = 0; 6411 p->se.statistics.block_start = 0; 6412#endif 6413 6414 if (!rt_task(p)) { 6415 /* 6416 * Renice negative nice level userspace 6417 * tasks back to 0: 6418 */ 6419 if (TASK_NICE(p) < 0 && p->mm) 6420 set_user_nice(p, 0); 6421 continue; 6422 } 6423 6424 raw_spin_lock(&p->pi_lock); 6425 rq = __task_rq_lock(p); 6426 6427 normalize_task(rq, p); 6428 6429 __task_rq_unlock(rq); 6430 raw_spin_unlock(&p->pi_lock); 6431 } while_each_thread(g, p); 6432 6433 read_unlock_irqrestore(&tasklist_lock, flags); 6434} 6435 6436#endif /* CONFIG_MAGIC_SYSRQ */ 6437 6438#if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) 6439/* 6440 * These functions are only useful for the IA64 MCA handling, or kdb. 6441 * 6442 * They can only be called when the whole system has been 6443 * stopped - every CPU needs to be quiescent, and no scheduling 6444 * activity can take place. Using them for anything else would 6445 * be a serious bug, and as a result, they aren't even visible 6446 * under any other configuration. 6447 */ 6448 6449/** 6450 * curr_task - return the current task for a given cpu. 6451 * @cpu: the processor in question. 6452 * 6453 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED! 6454 * 6455 * Return: The current task for @cpu. 6456 */ 6457struct task_struct *curr_task(int cpu) 6458{ 6459 return cpu_curr(cpu); 6460} 6461 6462#endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */ 6463 6464#ifdef CONFIG_IA64 6465/** 6466 * set_curr_task - set the current task for a given cpu. 6467 * @cpu: the processor in question. 6468 * @p: the task pointer to set. 6469 * 6470 * Description: This function must only be used when non-maskable interrupts 6471 * are serviced on a separate stack. It allows the architecture to switch the 6472 * notion of the current task on a cpu in a non-blocking manner. This function 6473 * must be called with all CPU's synchronized, and interrupts disabled, the 6474 * and caller must save the original value of the current task (see 6475 * curr_task() above) and restore that value before reenabling interrupts and 6476 * re-starting the system. 6477 * 6478 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED! 6479 */ 6480void set_curr_task(int cpu, struct task_struct *p) 6481{ 6482 cpu_curr(cpu) = p; 6483} 6484 6485#endif 6486 6487#ifdef CONFIG_CGROUP_SCHED 6488/* task_group_lock serializes the addition/removal of task groups */ 6489static DEFINE_SPINLOCK(task_group_lock); 6490 6491static void free_sched_group(struct task_group *tg) 6492{ 6493 free_fair_sched_group(tg); 6494 free_rt_sched_group(tg); 6495 autogroup_free(tg); 6496 kfree(tg); 6497} 6498 6499/* allocate runqueue etc for a new task group */ 6500struct task_group *sched_create_group(struct task_group *parent) 6501{ 6502 struct task_group *tg; 6503 6504 tg = kzalloc(sizeof(*tg), GFP_KERNEL); 6505 if (!tg) 6506 return ERR_PTR(-ENOMEM); 6507 6508 if (!alloc_fair_sched_group(tg, parent)) 6509 goto err; 6510 6511 if (!alloc_rt_sched_group(tg, parent)) 6512 goto err; 6513 6514 return tg; 6515 6516err: 6517 free_sched_group(tg); 6518 return ERR_PTR(-ENOMEM); 6519} 6520 6521void sched_online_group(struct task_group *tg, struct task_group *parent) 6522{ 6523 unsigned long flags; 6524 6525 spin_lock_irqsave(&task_group_lock, flags); 6526 list_add_rcu(&tg->list, &task_groups); 6527 6528 WARN_ON(!parent); /* root should already exist */ 6529 6530 tg->parent = parent; 6531 INIT_LIST_HEAD(&tg->children); 6532 list_add_rcu(&tg->siblings, &parent->children); 6533 spin_unlock_irqrestore(&task_group_lock, flags); 6534} 6535 6536/* rcu callback to free various structures associated with a task group */ 6537static void free_sched_group_rcu(struct rcu_head *rhp) 6538{ 6539 /* now it should be safe to free those cfs_rqs */ 6540 free_sched_group(container_of(rhp, struct task_group, rcu)); 6541} 6542 6543/* Destroy runqueue etc associated with a task group */ 6544void sched_destroy_group(struct task_group *tg) 6545{ 6546 /* wait for possible concurrent references to cfs_rqs complete */ 6547 call_rcu(&tg->rcu, free_sched_group_rcu); 6548} 6549 6550void sched_offline_group(struct task_group *tg) 6551{ 6552 unsigned long flags; 6553 int i; 6554 6555 /* end participation in shares distribution */ 6556 for_each_possible_cpu(i) 6557 unregister_fair_sched_group(tg, i); 6558 6559 spin_lock_irqsave(&task_group_lock, flags); 6560 list_del_rcu(&tg->list); 6561 list_del_rcu(&tg->siblings); 6562 spin_unlock_irqrestore(&task_group_lock, flags); 6563} 6564 6565/* change task's runqueue when it moves between groups. 6566 * The caller of this function should have put the task in its new group 6567 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to 6568 * reflect its new group. 6569 */ 6570void sched_move_task(struct task_struct *tsk) 6571{ 6572 struct task_group *tg; 6573 int on_rq, running; 6574 unsigned long flags; 6575 struct rq *rq; 6576 6577 rq = task_rq_lock(tsk, &flags); 6578 6579 running = task_current(rq, tsk); 6580 on_rq = tsk->on_rq; 6581 6582 if (on_rq) 6583 dequeue_task(rq, tsk, 0); 6584 if (unlikely(running)) 6585 tsk->sched_class->put_prev_task(rq, tsk); 6586 6587 tg = container_of(task_css_check(tsk, cpu_cgroup_subsys_id, 6588 lockdep_is_held(&tsk->sighand->siglock)), 6589 struct task_group, css); 6590 tg = autogroup_task_group(tsk, tg); 6591 tsk->sched_task_group = tg; 6592 6593#ifdef CONFIG_FAIR_GROUP_SCHED 6594 if (tsk->sched_class->task_move_group) 6595 tsk->sched_class->task_move_group(tsk, on_rq); 6596 else 6597#endif 6598 set_task_rq(tsk, task_cpu(tsk)); 6599 6600 if (unlikely(running)) 6601 tsk->sched_class->set_curr_task(rq); 6602 if (on_rq) 6603 enqueue_task(rq, tsk, 0); 6604 6605 task_rq_unlock(rq, tsk, &flags); 6606} 6607#endif /* CONFIG_CGROUP_SCHED */ 6608 6609#if defined(CONFIG_RT_GROUP_SCHED) || defined(CONFIG_CFS_BANDWIDTH) 6610static unsigned long to_ratio(u64 period, u64 runtime) 6611{ 6612 if (runtime == RUNTIME_INF) 6613 return 1ULL << 20; 6614 6615 return div64_u64(runtime << 20, period); 6616} 6617#endif 6618 6619#ifdef CONFIG_RT_GROUP_SCHED 6620/* 6621 * Ensure that the real time constraints are schedulable. 6622 */ 6623static DEFINE_MUTEX(rt_constraints_mutex); 6624 6625/* Must be called with tasklist_lock held */ 6626static inline int tg_has_rt_tasks(struct task_group *tg) 6627{ 6628 struct task_struct *g, *p; 6629 6630 do_each_thread(g, p) { 6631 if (rt_task(p) && task_rq(p)->rt.tg == tg) 6632 return 1; 6633 } while_each_thread(g, p); 6634 6635 return 0; 6636} 6637 6638struct rt_schedulable_data { 6639 struct task_group *tg; 6640 u64 rt_period; 6641 u64 rt_runtime; 6642}; 6643 6644static int tg_rt_schedulable(struct task_group *tg, void *data) 6645{ 6646 struct rt_schedulable_data *d = data; 6647 struct task_group *child; 6648 unsigned long total, sum = 0; 6649 u64 period, runtime; 6650 6651 period = ktime_to_ns(tg->rt_bandwidth.rt_period); 6652 runtime = tg->rt_bandwidth.rt_runtime; 6653 6654 if (tg == d->tg) { 6655 period = d->rt_period; 6656 runtime = d->rt_runtime; 6657 } 6658 6659 /* 6660 * Cannot have more runtime than the period. 6661 */ 6662 if (runtime > period && runtime != RUNTIME_INF) 6663 return -EINVAL; 6664 6665 /* 6666 * Ensure we don't starve existing RT tasks. 6667 */ 6668 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg)) 6669 return -EBUSY; 6670 6671 total = to_ratio(period, runtime); 6672 6673 /* 6674 * Nobody can have more than the global setting allows. 6675 */ 6676 if (total > to_ratio(global_rt_period(), global_rt_runtime())) 6677 return -EINVAL; 6678 6679 /* 6680 * The sum of our children's runtime should not exceed our own. 6681 */ 6682 list_for_each_entry_rcu(child, &tg->children, siblings) { 6683 period = ktime_to_ns(child->rt_bandwidth.rt_period); 6684 runtime = child->rt_bandwidth.rt_runtime; 6685 6686 if (child == d->tg) { 6687 period = d->rt_period; 6688 runtime = d->rt_runtime; 6689 } 6690 6691 sum += to_ratio(period, runtime); 6692 } 6693 6694 if (sum > total) 6695 return -EINVAL; 6696 6697 return 0; 6698} 6699 6700static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime) 6701{ 6702 int ret; 6703 6704 struct rt_schedulable_data data = { 6705 .tg = tg, 6706 .rt_period = period, 6707 .rt_runtime = runtime, 6708 }; 6709 6710 rcu_read_lock(); 6711 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data); 6712 rcu_read_unlock(); 6713 6714 return ret; 6715} 6716 6717static int tg_set_rt_bandwidth(struct task_group *tg, 6718 u64 rt_period, u64 rt_runtime) 6719{ 6720 int i, err = 0; 6721 6722 mutex_lock(&rt_constraints_mutex); 6723 read_lock(&tasklist_lock); 6724 err = __rt_schedulable(tg, rt_period, rt_runtime); 6725 if (err) 6726 goto unlock; 6727 6728 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock); 6729 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period); 6730 tg->rt_bandwidth.rt_runtime = rt_runtime; 6731 6732 for_each_possible_cpu(i) { 6733 struct rt_rq *rt_rq = tg->rt_rq[i]; 6734 6735 raw_spin_lock(&rt_rq->rt_runtime_lock); 6736 rt_rq->rt_runtime = rt_runtime; 6737 raw_spin_unlock(&rt_rq->rt_runtime_lock); 6738 } 6739 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock); 6740unlock: 6741 read_unlock(&tasklist_lock); 6742 mutex_unlock(&rt_constraints_mutex); 6743 6744 return err; 6745} 6746 6747static int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us) 6748{ 6749 u64 rt_runtime, rt_period; 6750 6751 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period); 6752 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC; 6753 if (rt_runtime_us < 0) 6754 rt_runtime = RUNTIME_INF; 6755 6756 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime); 6757} 6758 6759static long sched_group_rt_runtime(struct task_group *tg) 6760{ 6761 u64 rt_runtime_us; 6762 6763 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF) 6764 return -1; 6765 6766 rt_runtime_us = tg->rt_bandwidth.rt_runtime; 6767 do_div(rt_runtime_us, NSEC_PER_USEC); 6768 return rt_runtime_us; 6769} 6770 6771static int sched_group_set_rt_period(struct task_group *tg, long rt_period_us) 6772{ 6773 u64 rt_runtime, rt_period; 6774 6775 rt_period = (u64)rt_period_us * NSEC_PER_USEC; 6776 rt_runtime = tg->rt_bandwidth.rt_runtime; 6777 6778 if (rt_period == 0) 6779 return -EINVAL; 6780 6781 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime); 6782} 6783 6784static long sched_group_rt_period(struct task_group *tg) 6785{ 6786 u64 rt_period_us; 6787 6788 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period); 6789 do_div(rt_period_us, NSEC_PER_USEC); 6790 return rt_period_us; 6791} 6792 6793static int sched_rt_global_constraints(void) 6794{ 6795 u64 runtime, period; 6796 int ret = 0; 6797 6798 if (sysctl_sched_rt_period <= 0) 6799 return -EINVAL; 6800 6801 runtime = global_rt_runtime(); 6802 period = global_rt_period(); 6803 6804 /* 6805 * Sanity check on the sysctl variables. 6806 */ 6807 if (runtime > period && runtime != RUNTIME_INF) 6808 return -EINVAL; 6809 6810 mutex_lock(&rt_constraints_mutex); 6811 read_lock(&tasklist_lock); 6812 ret = __rt_schedulable(NULL, 0, 0); 6813 read_unlock(&tasklist_lock); 6814 mutex_unlock(&rt_constraints_mutex); 6815 6816 return ret; 6817} 6818 6819static int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk) 6820{ 6821 /* Don't accept realtime tasks when there is no way for them to run */ 6822 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0) 6823 return 0; 6824 6825 return 1; 6826} 6827 6828#else /* !CONFIG_RT_GROUP_SCHED */ 6829static int sched_rt_global_constraints(void) 6830{ 6831 unsigned long flags; 6832 int i; 6833 6834 if (sysctl_sched_rt_period <= 0) 6835 return -EINVAL; 6836 6837 /* 6838 * There's always some RT tasks in the root group 6839 * -- migration, kstopmachine etc.. 6840 */ 6841 if (sysctl_sched_rt_runtime == 0) 6842 return -EBUSY; 6843 6844 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags); 6845 for_each_possible_cpu(i) { 6846 struct rt_rq *rt_rq = &cpu_rq(i)->rt; 6847 6848 raw_spin_lock(&rt_rq->rt_runtime_lock); 6849 rt_rq->rt_runtime = global_rt_runtime(); 6850 raw_spin_unlock(&rt_rq->rt_runtime_lock); 6851 } 6852 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags); 6853 6854 return 0; 6855} 6856#endif /* CONFIG_RT_GROUP_SCHED */ 6857 6858int sched_rr_handler(struct ctl_table *table, int write, 6859 void __user *buffer, size_t *lenp, 6860 loff_t *ppos) 6861{ 6862 int ret; 6863 static DEFINE_MUTEX(mutex); 6864 6865 mutex_lock(&mutex); 6866 ret = proc_dointvec(table, write, buffer, lenp, ppos); 6867 /* make sure that internally we keep jiffies */ 6868 /* also, writing zero resets timeslice to default */ 6869 if (!ret && write) { 6870 sched_rr_timeslice = sched_rr_timeslice <= 0 ? 6871 RR_TIMESLICE : msecs_to_jiffies(sched_rr_timeslice); 6872 } 6873 mutex_unlock(&mutex); 6874 return ret; 6875} 6876 6877int sched_rt_handler(struct ctl_table *table, int write, 6878 void __user *buffer, size_t *lenp, 6879 loff_t *ppos) 6880{ 6881 int ret; 6882 int old_period, old_runtime; 6883 static DEFINE_MUTEX(mutex); 6884 6885 mutex_lock(&mutex); 6886 old_period = sysctl_sched_rt_period; 6887 old_runtime = sysctl_sched_rt_runtime; 6888 6889 ret = proc_dointvec(table, write, buffer, lenp, ppos); 6890 6891 if (!ret && write) { 6892 ret = sched_rt_global_constraints(); 6893 if (ret) { 6894 sysctl_sched_rt_period = old_period; 6895 sysctl_sched_rt_runtime = old_runtime; 6896 } else { 6897 def_rt_bandwidth.rt_runtime = global_rt_runtime(); 6898 def_rt_bandwidth.rt_period = 6899 ns_to_ktime(global_rt_period()); 6900 } 6901 } 6902 mutex_unlock(&mutex); 6903 6904 return ret; 6905} 6906 6907#ifdef CONFIG_CGROUP_SCHED 6908 6909static inline struct task_group *css_tg(struct cgroup_subsys_state *css) 6910{ 6911 return css ? container_of(css, struct task_group, css) : NULL; 6912} 6913 6914static struct cgroup_subsys_state * 6915cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css) 6916{ 6917 struct task_group *parent = css_tg(parent_css); 6918 struct task_group *tg; 6919 6920 if (!parent) { 6921 /* This is early initialization for the top cgroup */ 6922 return &root_task_group.css; 6923 } 6924 6925 tg = sched_create_group(parent); 6926 if (IS_ERR(tg)) 6927 return ERR_PTR(-ENOMEM); 6928 6929 return &tg->css; 6930} 6931 6932static int cpu_cgroup_css_online(struct cgroup_subsys_state *css) 6933{ 6934 struct task_group *tg = css_tg(css); 6935 struct task_group *parent = css_tg(css_parent(css)); 6936 6937 if (parent) 6938 sched_online_group(tg, parent); 6939 return 0; 6940} 6941 6942static void cpu_cgroup_css_free(struct cgroup_subsys_state *css) 6943{ 6944 struct task_group *tg = css_tg(css); 6945 6946 sched_destroy_group(tg); 6947} 6948 6949static void cpu_cgroup_css_offline(struct cgroup_subsys_state *css) 6950{ 6951 struct task_group *tg = css_tg(css); 6952 6953 sched_offline_group(tg); 6954} 6955 6956static int cpu_cgroup_can_attach(struct cgroup_subsys_state *css, 6957 struct cgroup_taskset *tset) 6958{ 6959 struct task_struct *task; 6960 6961 cgroup_taskset_for_each(task, css, tset) { 6962#ifdef CONFIG_RT_GROUP_SCHED 6963 if (!sched_rt_can_attach(css_tg(css), task)) 6964 return -EINVAL; 6965#else 6966 /* We don't support RT-tasks being in separate groups */ 6967 if (task->sched_class != &fair_sched_class) 6968 return -EINVAL; 6969#endif 6970 } 6971 return 0; 6972} 6973 6974static void cpu_cgroup_attach(struct cgroup_subsys_state *css, 6975 struct cgroup_taskset *tset) 6976{ 6977 struct task_struct *task; 6978 6979 cgroup_taskset_for_each(task, css, tset) 6980 sched_move_task(task); 6981} 6982 6983static void cpu_cgroup_exit(struct cgroup_subsys_state *css, 6984 struct cgroup_subsys_state *old_css, 6985 struct task_struct *task) 6986{ 6987 /* 6988 * cgroup_exit() is called in the copy_process() failure path. 6989 * Ignore this case since the task hasn't ran yet, this avoids 6990 * trying to poke a half freed task state from generic code. 6991 */ 6992 if (!(task->flags & PF_EXITING)) 6993 return; 6994 6995 sched_move_task(task); 6996} 6997 6998#ifdef CONFIG_FAIR_GROUP_SCHED 6999static int cpu_shares_write_u64(struct cgroup_subsys_state *css, 7000 struct cftype *cftype, u64 shareval) 7001{ 7002 return sched_group_set_shares(css_tg(css), scale_load(shareval)); 7003} 7004 7005static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css, 7006 struct cftype *cft) 7007{ 7008 struct task_group *tg = css_tg(css); 7009 7010 return (u64) scale_load_down(tg->shares); 7011} 7012 7013#ifdef CONFIG_CFS_BANDWIDTH 7014static DEFINE_MUTEX(cfs_constraints_mutex); 7015 7016const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */ 7017const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */ 7018 7019static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime); 7020 7021static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota) 7022{ 7023 int i, ret = 0, runtime_enabled, runtime_was_enabled; 7024 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth; 7025 7026 if (tg == &root_task_group) 7027 return -EINVAL; 7028 7029 /* 7030 * Ensure we have at some amount of bandwidth every period. This is 7031 * to prevent reaching a state of large arrears when throttled via 7032 * entity_tick() resulting in prolonged exit starvation. 7033 */ 7034 if (quota < min_cfs_quota_period || period < min_cfs_quota_period) 7035 return -EINVAL; 7036 7037 /* 7038 * Likewise, bound things on the otherside by preventing insane quota 7039 * periods. This also allows us to normalize in computing quota 7040 * feasibility. 7041 */ 7042 if (period > max_cfs_quota_period) 7043 return -EINVAL; 7044 7045 mutex_lock(&cfs_constraints_mutex); 7046 ret = __cfs_schedulable(tg, period, quota); 7047 if (ret) 7048 goto out_unlock; 7049 7050 runtime_enabled = quota != RUNTIME_INF; 7051 runtime_was_enabled = cfs_b->quota != RUNTIME_INF; 7052 /* 7053 * If we need to toggle cfs_bandwidth_used, off->on must occur 7054 * before making related changes, and on->off must occur afterwards 7055 */ 7056 if (runtime_enabled && !runtime_was_enabled) 7057 cfs_bandwidth_usage_inc(); 7058 raw_spin_lock_irq(&cfs_b->lock); 7059 cfs_b->period = ns_to_ktime(period); 7060 cfs_b->quota = quota; 7061 7062 __refill_cfs_bandwidth_runtime(cfs_b); 7063 /* restart the period timer (if active) to handle new period expiry */ 7064 if (runtime_enabled && cfs_b->timer_active) { 7065 /* force a reprogram */ 7066 cfs_b->timer_active = 0; 7067 __start_cfs_bandwidth(cfs_b); 7068 } 7069 raw_spin_unlock_irq(&cfs_b->lock); 7070 7071 for_each_possible_cpu(i) { 7072 struct cfs_rq *cfs_rq = tg->cfs_rq[i]; 7073 struct rq *rq = cfs_rq->rq; 7074 7075 raw_spin_lock_irq(&rq->lock); 7076 cfs_rq->runtime_enabled = runtime_enabled; 7077 cfs_rq->runtime_remaining = 0; 7078 7079 if (cfs_rq->throttled) 7080 unthrottle_cfs_rq(cfs_rq); 7081 raw_spin_unlock_irq(&rq->lock); 7082 } 7083 if (runtime_was_enabled && !runtime_enabled) 7084 cfs_bandwidth_usage_dec(); 7085out_unlock: 7086 mutex_unlock(&cfs_constraints_mutex); 7087 7088 return ret; 7089} 7090 7091int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us) 7092{ 7093 u64 quota, period; 7094 7095 period = ktime_to_ns(tg->cfs_bandwidth.period); 7096 if (cfs_quota_us < 0) 7097 quota = RUNTIME_INF; 7098 else 7099 quota = (u64)cfs_quota_us * NSEC_PER_USEC; 7100 7101 return tg_set_cfs_bandwidth(tg, period, quota); 7102} 7103 7104long tg_get_cfs_quota(struct task_group *tg) 7105{ 7106 u64 quota_us; 7107 7108 if (tg->cfs_bandwidth.quota == RUNTIME_INF) 7109 return -1; 7110 7111 quota_us = tg->cfs_bandwidth.quota; 7112 do_div(quota_us, NSEC_PER_USEC); 7113 7114 return quota_us; 7115} 7116 7117int tg_set_cfs_period(struct task_group *tg, long cfs_period_us) 7118{ 7119 u64 quota, period; 7120 7121 period = (u64)cfs_period_us * NSEC_PER_USEC; 7122 quota = tg->cfs_bandwidth.quota; 7123 7124 return tg_set_cfs_bandwidth(tg, period, quota); 7125} 7126 7127long tg_get_cfs_period(struct task_group *tg) 7128{ 7129 u64 cfs_period_us; 7130 7131 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period); 7132 do_div(cfs_period_us, NSEC_PER_USEC); 7133 7134 return cfs_period_us; 7135} 7136 7137static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css, 7138 struct cftype *cft) 7139{ 7140 return tg_get_cfs_quota(css_tg(css)); 7141} 7142 7143static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css, 7144 struct cftype *cftype, s64 cfs_quota_us) 7145{ 7146 return tg_set_cfs_quota(css_tg(css), cfs_quota_us); 7147} 7148 7149static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css, 7150 struct cftype *cft) 7151{ 7152 return tg_get_cfs_period(css_tg(css)); 7153} 7154 7155static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css, 7156 struct cftype *cftype, u64 cfs_period_us) 7157{ 7158 return tg_set_cfs_period(css_tg(css), cfs_period_us); 7159} 7160 7161struct cfs_schedulable_data { 7162 struct task_group *tg; 7163 u64 period, quota; 7164}; 7165 7166/* 7167 * normalize group quota/period to be quota/max_period 7168 * note: units are usecs 7169 */ 7170static u64 normalize_cfs_quota(struct task_group *tg, 7171 struct cfs_schedulable_data *d) 7172{ 7173 u64 quota, period; 7174 7175 if (tg == d->tg) { 7176 period = d->period; 7177 quota = d->quota; 7178 } else { 7179 period = tg_get_cfs_period(tg); 7180 quota = tg_get_cfs_quota(tg); 7181 } 7182 7183 /* note: these should typically be equivalent */ 7184 if (quota == RUNTIME_INF || quota == -1) 7185 return RUNTIME_INF; 7186 7187 return to_ratio(period, quota); 7188} 7189 7190static int tg_cfs_schedulable_down(struct task_group *tg, void *data) 7191{ 7192 struct cfs_schedulable_data *d = data; 7193 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth; 7194 s64 quota = 0, parent_quota = -1; 7195 7196 if (!tg->parent) { 7197 quota = RUNTIME_INF; 7198 } else { 7199 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth; 7200 7201 quota = normalize_cfs_quota(tg, d); 7202 parent_quota = parent_b->hierarchal_quota; 7203 7204 /* 7205 * ensure max(child_quota) <= parent_quota, inherit when no 7206 * limit is set 7207 */ 7208 if (quota == RUNTIME_INF) 7209 quota = parent_quota; 7210 else if (parent_quota != RUNTIME_INF && quota > parent_quota) 7211 return -EINVAL; 7212 } 7213 cfs_b->hierarchal_quota = quota; 7214 7215 return 0; 7216} 7217 7218static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota) 7219{ 7220 int ret; 7221 struct cfs_schedulable_data data = { 7222 .tg = tg, 7223 .period = period, 7224 .quota = quota, 7225 }; 7226 7227 if (quota != RUNTIME_INF) { 7228 do_div(data.period, NSEC_PER_USEC); 7229 do_div(data.quota, NSEC_PER_USEC); 7230 } 7231 7232 rcu_read_lock(); 7233 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data); 7234 rcu_read_unlock(); 7235 7236 return ret; 7237} 7238 7239static int cpu_stats_show(struct cgroup_subsys_state *css, struct cftype *cft, 7240 struct cgroup_map_cb *cb) 7241{ 7242 struct task_group *tg = css_tg(css); 7243 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth; 7244 7245 cb->fill(cb, "nr_periods", cfs_b->nr_periods); 7246 cb->fill(cb, "nr_throttled", cfs_b->nr_throttled); 7247 cb->fill(cb, "throttled_time", cfs_b->throttled_time); 7248 7249 return 0; 7250} 7251#endif /* CONFIG_CFS_BANDWIDTH */ 7252#endif /* CONFIG_FAIR_GROUP_SCHED */ 7253 7254#ifdef CONFIG_RT_GROUP_SCHED 7255static int cpu_rt_runtime_write(struct cgroup_subsys_state *css, 7256 struct cftype *cft, s64 val) 7257{ 7258 return sched_group_set_rt_runtime(css_tg(css), val); 7259} 7260 7261static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css, 7262 struct cftype *cft) 7263{ 7264 return sched_group_rt_runtime(css_tg(css)); 7265} 7266 7267static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css, 7268 struct cftype *cftype, u64 rt_period_us) 7269{ 7270 return sched_group_set_rt_period(css_tg(css), rt_period_us); 7271} 7272 7273static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css, 7274 struct cftype *cft) 7275{ 7276 return sched_group_rt_period(css_tg(css)); 7277} 7278#endif /* CONFIG_RT_GROUP_SCHED */ 7279 7280static struct cftype cpu_files[] = { 7281#ifdef CONFIG_FAIR_GROUP_SCHED 7282 { 7283 .name = "shares", 7284 .read_u64 = cpu_shares_read_u64, 7285 .write_u64 = cpu_shares_write_u64, 7286 }, 7287#endif 7288#ifdef CONFIG_CFS_BANDWIDTH 7289 { 7290 .name = "cfs_quota_us", 7291 .read_s64 = cpu_cfs_quota_read_s64, 7292 .write_s64 = cpu_cfs_quota_write_s64, 7293 }, 7294 { 7295 .name = "cfs_period_us", 7296 .read_u64 = cpu_cfs_period_read_u64, 7297 .write_u64 = cpu_cfs_period_write_u64, 7298 }, 7299 { 7300 .name = "stat", 7301 .read_map = cpu_stats_show, 7302 }, 7303#endif 7304#ifdef CONFIG_RT_GROUP_SCHED 7305 { 7306 .name = "rt_runtime_us", 7307 .read_s64 = cpu_rt_runtime_read, 7308 .write_s64 = cpu_rt_runtime_write, 7309 }, 7310 { 7311 .name = "rt_period_us", 7312 .read_u64 = cpu_rt_period_read_uint, 7313 .write_u64 = cpu_rt_period_write_uint, 7314 }, 7315#endif 7316 { } /* terminate */ 7317}; 7318 7319struct cgroup_subsys cpu_cgroup_subsys = { 7320 .name = "cpu", 7321 .css_alloc = cpu_cgroup_css_alloc, 7322 .css_free = cpu_cgroup_css_free, 7323 .css_online = cpu_cgroup_css_online, 7324 .css_offline = cpu_cgroup_css_offline, 7325 .can_attach = cpu_cgroup_can_attach, 7326 .attach = cpu_cgroup_attach, 7327 .exit = cpu_cgroup_exit, 7328 .subsys_id = cpu_cgroup_subsys_id, 7329 .base_cftypes = cpu_files, 7330 .early_init = 1, 7331}; 7332 7333#endif /* CONFIG_CGROUP_SCHED */ 7334 7335void dump_cpu_task(int cpu) 7336{ 7337 pr_info("Task dump for CPU %d:\n", cpu); 7338 sched_show_task(cpu_curr(cpu)); 7339} 7340