core.c revision 688b7585d16ab57a17aa4422a3b290b3a55fa679
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 1016struct migration_arg { 1017 struct task_struct *task; 1018 int dest_cpu; 1019}; 1020 1021static int migration_cpu_stop(void *data); 1022 1023/* 1024 * wait_task_inactive - wait for a thread to unschedule. 1025 * 1026 * If @match_state is nonzero, it's the @p->state value just checked and 1027 * not expected to change. If it changes, i.e. @p might have woken up, 1028 * then return zero. When we succeed in waiting for @p to be off its CPU, 1029 * we return a positive number (its total switch count). If a second call 1030 * a short while later returns the same number, the caller can be sure that 1031 * @p has remained unscheduled the whole time. 1032 * 1033 * The caller must ensure that the task *will* unschedule sometime soon, 1034 * else this function might spin for a *long* time. This function can't 1035 * be called with interrupts off, or it may introduce deadlock with 1036 * smp_call_function() if an IPI is sent by the same process we are 1037 * waiting to become inactive. 1038 */ 1039unsigned long wait_task_inactive(struct task_struct *p, long match_state) 1040{ 1041 unsigned long flags; 1042 int running, on_rq; 1043 unsigned long ncsw; 1044 struct rq *rq; 1045 1046 for (;;) { 1047 /* 1048 * We do the initial early heuristics without holding 1049 * any task-queue locks at all. We'll only try to get 1050 * the runqueue lock when things look like they will 1051 * work out! 1052 */ 1053 rq = task_rq(p); 1054 1055 /* 1056 * If the task is actively running on another CPU 1057 * still, just relax and busy-wait without holding 1058 * any locks. 1059 * 1060 * NOTE! Since we don't hold any locks, it's not 1061 * even sure that "rq" stays as the right runqueue! 1062 * But we don't care, since "task_running()" will 1063 * return false if the runqueue has changed and p 1064 * is actually now running somewhere else! 1065 */ 1066 while (task_running(rq, p)) { 1067 if (match_state && unlikely(p->state != match_state)) 1068 return 0; 1069 cpu_relax(); 1070 } 1071 1072 /* 1073 * Ok, time to look more closely! We need the rq 1074 * lock now, to be *sure*. If we're wrong, we'll 1075 * just go back and repeat. 1076 */ 1077 rq = task_rq_lock(p, &flags); 1078 trace_sched_wait_task(p); 1079 running = task_running(rq, p); 1080 on_rq = p->on_rq; 1081 ncsw = 0; 1082 if (!match_state || p->state == match_state) 1083 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */ 1084 task_rq_unlock(rq, p, &flags); 1085 1086 /* 1087 * If it changed from the expected state, bail out now. 1088 */ 1089 if (unlikely(!ncsw)) 1090 break; 1091 1092 /* 1093 * Was it really running after all now that we 1094 * checked with the proper locks actually held? 1095 * 1096 * Oops. Go back and try again.. 1097 */ 1098 if (unlikely(running)) { 1099 cpu_relax(); 1100 continue; 1101 } 1102 1103 /* 1104 * It's not enough that it's not actively running, 1105 * it must be off the runqueue _entirely_, and not 1106 * preempted! 1107 * 1108 * So if it was still runnable (but just not actively 1109 * running right now), it's preempted, and we should 1110 * yield - it could be a while. 1111 */ 1112 if (unlikely(on_rq)) { 1113 ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ); 1114 1115 set_current_state(TASK_UNINTERRUPTIBLE); 1116 schedule_hrtimeout(&to, HRTIMER_MODE_REL); 1117 continue; 1118 } 1119 1120 /* 1121 * Ahh, all good. It wasn't running, and it wasn't 1122 * runnable, which means that it will never become 1123 * running in the future either. We're all done! 1124 */ 1125 break; 1126 } 1127 1128 return ncsw; 1129} 1130 1131/*** 1132 * kick_process - kick a running thread to enter/exit the kernel 1133 * @p: the to-be-kicked thread 1134 * 1135 * Cause a process which is running on another CPU to enter 1136 * kernel-mode, without any delay. (to get signals handled.) 1137 * 1138 * NOTE: this function doesn't have to take the runqueue lock, 1139 * because all it wants to ensure is that the remote task enters 1140 * the kernel. If the IPI races and the task has been migrated 1141 * to another CPU then no harm is done and the purpose has been 1142 * achieved as well. 1143 */ 1144void kick_process(struct task_struct *p) 1145{ 1146 int cpu; 1147 1148 preempt_disable(); 1149 cpu = task_cpu(p); 1150 if ((cpu != smp_processor_id()) && task_curr(p)) 1151 smp_send_reschedule(cpu); 1152 preempt_enable(); 1153} 1154EXPORT_SYMBOL_GPL(kick_process); 1155#endif /* CONFIG_SMP */ 1156 1157#ifdef CONFIG_SMP 1158/* 1159 * ->cpus_allowed is protected by both rq->lock and p->pi_lock 1160 */ 1161static int select_fallback_rq(int cpu, struct task_struct *p) 1162{ 1163 int nid = cpu_to_node(cpu); 1164 const struct cpumask *nodemask = NULL; 1165 enum { cpuset, possible, fail } state = cpuset; 1166 int dest_cpu; 1167 1168 /* 1169 * If the node that the cpu is on has been offlined, cpu_to_node() 1170 * will return -1. There is no cpu on the node, and we should 1171 * select the cpu on the other node. 1172 */ 1173 if (nid != -1) { 1174 nodemask = cpumask_of_node(nid); 1175 1176 /* Look for allowed, online CPU in same node. */ 1177 for_each_cpu(dest_cpu, nodemask) { 1178 if (!cpu_online(dest_cpu)) 1179 continue; 1180 if (!cpu_active(dest_cpu)) 1181 continue; 1182 if (cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p))) 1183 return dest_cpu; 1184 } 1185 } 1186 1187 for (;;) { 1188 /* Any allowed, online CPU? */ 1189 for_each_cpu(dest_cpu, tsk_cpus_allowed(p)) { 1190 if (!cpu_online(dest_cpu)) 1191 continue; 1192 if (!cpu_active(dest_cpu)) 1193 continue; 1194 goto out; 1195 } 1196 1197 switch (state) { 1198 case cpuset: 1199 /* No more Mr. Nice Guy. */ 1200 cpuset_cpus_allowed_fallback(p); 1201 state = possible; 1202 break; 1203 1204 case possible: 1205 do_set_cpus_allowed(p, cpu_possible_mask); 1206 state = fail; 1207 break; 1208 1209 case fail: 1210 BUG(); 1211 break; 1212 } 1213 } 1214 1215out: 1216 if (state != cpuset) { 1217 /* 1218 * Don't tell them about moving exiting tasks or 1219 * kernel threads (both mm NULL), since they never 1220 * leave kernel. 1221 */ 1222 if (p->mm && printk_ratelimit()) { 1223 printk_sched("process %d (%s) no longer affine to cpu%d\n", 1224 task_pid_nr(p), p->comm, cpu); 1225 } 1226 } 1227 1228 return dest_cpu; 1229} 1230 1231/* 1232 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable. 1233 */ 1234static inline 1235int select_task_rq(struct task_struct *p, int sd_flags, int wake_flags) 1236{ 1237 int cpu = p->sched_class->select_task_rq(p, sd_flags, wake_flags); 1238 1239 /* 1240 * In order not to call set_task_cpu() on a blocking task we need 1241 * to rely on ttwu() to place the task on a valid ->cpus_allowed 1242 * cpu. 1243 * 1244 * Since this is common to all placement strategies, this lives here. 1245 * 1246 * [ this allows ->select_task() to simply return task_cpu(p) and 1247 * not worry about this generic constraint ] 1248 */ 1249 if (unlikely(!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)) || 1250 !cpu_online(cpu))) 1251 cpu = select_fallback_rq(task_cpu(p), p); 1252 1253 return cpu; 1254} 1255 1256static void update_avg(u64 *avg, u64 sample) 1257{ 1258 s64 diff = sample - *avg; 1259 *avg += diff >> 3; 1260} 1261#endif 1262 1263static void 1264ttwu_stat(struct task_struct *p, int cpu, int wake_flags) 1265{ 1266#ifdef CONFIG_SCHEDSTATS 1267 struct rq *rq = this_rq(); 1268 1269#ifdef CONFIG_SMP 1270 int this_cpu = smp_processor_id(); 1271 1272 if (cpu == this_cpu) { 1273 schedstat_inc(rq, ttwu_local); 1274 schedstat_inc(p, se.statistics.nr_wakeups_local); 1275 } else { 1276 struct sched_domain *sd; 1277 1278 schedstat_inc(p, se.statistics.nr_wakeups_remote); 1279 rcu_read_lock(); 1280 for_each_domain(this_cpu, sd) { 1281 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) { 1282 schedstat_inc(sd, ttwu_wake_remote); 1283 break; 1284 } 1285 } 1286 rcu_read_unlock(); 1287 } 1288 1289 if (wake_flags & WF_MIGRATED) 1290 schedstat_inc(p, se.statistics.nr_wakeups_migrate); 1291 1292#endif /* CONFIG_SMP */ 1293 1294 schedstat_inc(rq, ttwu_count); 1295 schedstat_inc(p, se.statistics.nr_wakeups); 1296 1297 if (wake_flags & WF_SYNC) 1298 schedstat_inc(p, se.statistics.nr_wakeups_sync); 1299 1300#endif /* CONFIG_SCHEDSTATS */ 1301} 1302 1303static void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags) 1304{ 1305 activate_task(rq, p, en_flags); 1306 p->on_rq = 1; 1307 1308 /* if a worker is waking up, notify workqueue */ 1309 if (p->flags & PF_WQ_WORKER) 1310 wq_worker_waking_up(p, cpu_of(rq)); 1311} 1312 1313/* 1314 * Mark the task runnable and perform wakeup-preemption. 1315 */ 1316static void 1317ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags) 1318{ 1319 check_preempt_curr(rq, p, wake_flags); 1320 trace_sched_wakeup(p, true); 1321 1322 p->state = TASK_RUNNING; 1323#ifdef CONFIG_SMP 1324 if (p->sched_class->task_woken) 1325 p->sched_class->task_woken(rq, p); 1326 1327 if (rq->idle_stamp) { 1328 u64 delta = rq_clock(rq) - rq->idle_stamp; 1329 u64 max = 2*rq->max_idle_balance_cost; 1330 1331 update_avg(&rq->avg_idle, delta); 1332 1333 if (rq->avg_idle > max) 1334 rq->avg_idle = max; 1335 1336 rq->idle_stamp = 0; 1337 } 1338#endif 1339} 1340 1341static void 1342ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags) 1343{ 1344#ifdef CONFIG_SMP 1345 if (p->sched_contributes_to_load) 1346 rq->nr_uninterruptible--; 1347#endif 1348 1349 ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_WAKING); 1350 ttwu_do_wakeup(rq, p, wake_flags); 1351} 1352 1353/* 1354 * Called in case the task @p isn't fully descheduled from its runqueue, 1355 * in this case we must do a remote wakeup. Its a 'light' wakeup though, 1356 * since all we need to do is flip p->state to TASK_RUNNING, since 1357 * the task is still ->on_rq. 1358 */ 1359static int ttwu_remote(struct task_struct *p, int wake_flags) 1360{ 1361 struct rq *rq; 1362 int ret = 0; 1363 1364 rq = __task_rq_lock(p); 1365 if (p->on_rq) { 1366 /* check_preempt_curr() may use rq clock */ 1367 update_rq_clock(rq); 1368 ttwu_do_wakeup(rq, p, wake_flags); 1369 ret = 1; 1370 } 1371 __task_rq_unlock(rq); 1372 1373 return ret; 1374} 1375 1376#ifdef CONFIG_SMP 1377static void sched_ttwu_pending(void) 1378{ 1379 struct rq *rq = this_rq(); 1380 struct llist_node *llist = llist_del_all(&rq->wake_list); 1381 struct task_struct *p; 1382 1383 raw_spin_lock(&rq->lock); 1384 1385 while (llist) { 1386 p = llist_entry(llist, struct task_struct, wake_entry); 1387 llist = llist_next(llist); 1388 ttwu_do_activate(rq, p, 0); 1389 } 1390 1391 raw_spin_unlock(&rq->lock); 1392} 1393 1394void scheduler_ipi(void) 1395{ 1396 /* 1397 * Fold TIF_NEED_RESCHED into the preempt_count; anybody setting 1398 * TIF_NEED_RESCHED remotely (for the first time) will also send 1399 * this IPI. 1400 */ 1401 if (tif_need_resched()) 1402 set_preempt_need_resched(); 1403 1404 if (llist_empty(&this_rq()->wake_list) 1405 && !tick_nohz_full_cpu(smp_processor_id()) 1406 && !got_nohz_idle_kick()) 1407 return; 1408 1409 /* 1410 * Not all reschedule IPI handlers call irq_enter/irq_exit, since 1411 * traditionally all their work was done from the interrupt return 1412 * path. Now that we actually do some work, we need to make sure 1413 * we do call them. 1414 * 1415 * Some archs already do call them, luckily irq_enter/exit nest 1416 * properly. 1417 * 1418 * Arguably we should visit all archs and update all handlers, 1419 * however a fair share of IPIs are still resched only so this would 1420 * somewhat pessimize the simple resched case. 1421 */ 1422 irq_enter(); 1423 tick_nohz_full_check(); 1424 sched_ttwu_pending(); 1425 1426 /* 1427 * Check if someone kicked us for doing the nohz idle load balance. 1428 */ 1429 if (unlikely(got_nohz_idle_kick())) { 1430 this_rq()->idle_balance = 1; 1431 raise_softirq_irqoff(SCHED_SOFTIRQ); 1432 } 1433 irq_exit(); 1434} 1435 1436static void ttwu_queue_remote(struct task_struct *p, int cpu) 1437{ 1438 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list)) 1439 smp_send_reschedule(cpu); 1440} 1441 1442bool cpus_share_cache(int this_cpu, int that_cpu) 1443{ 1444 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu); 1445} 1446#endif /* CONFIG_SMP */ 1447 1448static void ttwu_queue(struct task_struct *p, int cpu) 1449{ 1450 struct rq *rq = cpu_rq(cpu); 1451 1452#if defined(CONFIG_SMP) 1453 if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) { 1454 sched_clock_cpu(cpu); /* sync clocks x-cpu */ 1455 ttwu_queue_remote(p, cpu); 1456 return; 1457 } 1458#endif 1459 1460 raw_spin_lock(&rq->lock); 1461 ttwu_do_activate(rq, p, 0); 1462 raw_spin_unlock(&rq->lock); 1463} 1464 1465/** 1466 * try_to_wake_up - wake up a thread 1467 * @p: the thread to be awakened 1468 * @state: the mask of task states that can be woken 1469 * @wake_flags: wake modifier flags (WF_*) 1470 * 1471 * Put it on the run-queue if it's not already there. The "current" 1472 * thread is always on the run-queue (except when the actual 1473 * re-schedule is in progress), and as such you're allowed to do 1474 * the simpler "current->state = TASK_RUNNING" to mark yourself 1475 * runnable without the overhead of this. 1476 * 1477 * Return: %true if @p was woken up, %false if it was already running. 1478 * or @state didn't match @p's state. 1479 */ 1480static int 1481try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags) 1482{ 1483 unsigned long flags; 1484 int cpu, success = 0; 1485 1486 /* 1487 * If we are going to wake up a thread waiting for CONDITION we 1488 * need to ensure that CONDITION=1 done by the caller can not be 1489 * reordered with p->state check below. This pairs with mb() in 1490 * set_current_state() the waiting thread does. 1491 */ 1492 smp_mb__before_spinlock(); 1493 raw_spin_lock_irqsave(&p->pi_lock, flags); 1494 if (!(p->state & state)) 1495 goto out; 1496 1497 success = 1; /* we're going to change ->state */ 1498 cpu = task_cpu(p); 1499 1500 if (p->on_rq && ttwu_remote(p, wake_flags)) 1501 goto stat; 1502 1503#ifdef CONFIG_SMP 1504 /* 1505 * If the owning (remote) cpu is still in the middle of schedule() with 1506 * this task as prev, wait until its done referencing the task. 1507 */ 1508 while (p->on_cpu) 1509 cpu_relax(); 1510 /* 1511 * Pairs with the smp_wmb() in finish_lock_switch(). 1512 */ 1513 smp_rmb(); 1514 1515 p->sched_contributes_to_load = !!task_contributes_to_load(p); 1516 p->state = TASK_WAKING; 1517 1518 if (p->sched_class->task_waking) 1519 p->sched_class->task_waking(p); 1520 1521 cpu = select_task_rq(p, SD_BALANCE_WAKE, wake_flags); 1522 if (task_cpu(p) != cpu) { 1523 wake_flags |= WF_MIGRATED; 1524 set_task_cpu(p, cpu); 1525 } 1526#endif /* CONFIG_SMP */ 1527 1528 ttwu_queue(p, cpu); 1529stat: 1530 ttwu_stat(p, cpu, wake_flags); 1531out: 1532 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 1533 1534 return success; 1535} 1536 1537/** 1538 * try_to_wake_up_local - try to wake up a local task with rq lock held 1539 * @p: the thread to be awakened 1540 * 1541 * Put @p on the run-queue if it's not already there. The caller must 1542 * ensure that this_rq() is locked, @p is bound to this_rq() and not 1543 * the current task. 1544 */ 1545static void try_to_wake_up_local(struct task_struct *p) 1546{ 1547 struct rq *rq = task_rq(p); 1548 1549 if (WARN_ON_ONCE(rq != this_rq()) || 1550 WARN_ON_ONCE(p == current)) 1551 return; 1552 1553 lockdep_assert_held(&rq->lock); 1554 1555 if (!raw_spin_trylock(&p->pi_lock)) { 1556 raw_spin_unlock(&rq->lock); 1557 raw_spin_lock(&p->pi_lock); 1558 raw_spin_lock(&rq->lock); 1559 } 1560 1561 if (!(p->state & TASK_NORMAL)) 1562 goto out; 1563 1564 if (!p->on_rq) 1565 ttwu_activate(rq, p, ENQUEUE_WAKEUP); 1566 1567 ttwu_do_wakeup(rq, p, 0); 1568 ttwu_stat(p, smp_processor_id(), 0); 1569out: 1570 raw_spin_unlock(&p->pi_lock); 1571} 1572 1573/** 1574 * wake_up_process - Wake up a specific process 1575 * @p: The process to be woken up. 1576 * 1577 * Attempt to wake up the nominated process and move it to the set of runnable 1578 * processes. 1579 * 1580 * Return: 1 if the process was woken up, 0 if it was already running. 1581 * 1582 * It may be assumed that this function implies a write memory barrier before 1583 * changing the task state if and only if any tasks are woken up. 1584 */ 1585int wake_up_process(struct task_struct *p) 1586{ 1587 WARN_ON(task_is_stopped_or_traced(p)); 1588 return try_to_wake_up(p, TASK_NORMAL, 0); 1589} 1590EXPORT_SYMBOL(wake_up_process); 1591 1592int wake_up_state(struct task_struct *p, unsigned int state) 1593{ 1594 return try_to_wake_up(p, state, 0); 1595} 1596 1597/* 1598 * Perform scheduler related setup for a newly forked process p. 1599 * p is forked by current. 1600 * 1601 * __sched_fork() is basic setup used by init_idle() too: 1602 */ 1603static void __sched_fork(struct task_struct *p) 1604{ 1605 p->on_rq = 0; 1606 1607 p->se.on_rq = 0; 1608 p->se.exec_start = 0; 1609 p->se.sum_exec_runtime = 0; 1610 p->se.prev_sum_exec_runtime = 0; 1611 p->se.nr_migrations = 0; 1612 p->se.vruntime = 0; 1613 INIT_LIST_HEAD(&p->se.group_node); 1614 1615#ifdef CONFIG_SCHEDSTATS 1616 memset(&p->se.statistics, 0, sizeof(p->se.statistics)); 1617#endif 1618 1619 INIT_LIST_HEAD(&p->rt.run_list); 1620 1621#ifdef CONFIG_PREEMPT_NOTIFIERS 1622 INIT_HLIST_HEAD(&p->preempt_notifiers); 1623#endif 1624 1625#ifdef CONFIG_NUMA_BALANCING 1626 if (p->mm && atomic_read(&p->mm->mm_users) == 1) { 1627 p->mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay); 1628 p->mm->numa_next_reset = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_period_reset); 1629 p->mm->numa_scan_seq = 0; 1630 } 1631 1632 p->node_stamp = 0ULL; 1633 p->numa_scan_seq = p->mm ? p->mm->numa_scan_seq : 0; 1634 p->numa_migrate_seq = p->mm ? p->mm->numa_scan_seq - 1 : 0; 1635 p->numa_scan_period = sysctl_numa_balancing_scan_delay; 1636 p->numa_preferred_nid = -1; 1637 p->numa_work.next = &p->numa_work; 1638 p->numa_faults = NULL; 1639#endif /* CONFIG_NUMA_BALANCING */ 1640} 1641 1642#ifdef CONFIG_NUMA_BALANCING 1643#ifdef CONFIG_SCHED_DEBUG 1644void set_numabalancing_state(bool enabled) 1645{ 1646 if (enabled) 1647 sched_feat_set("NUMA"); 1648 else 1649 sched_feat_set("NO_NUMA"); 1650} 1651#else 1652__read_mostly bool numabalancing_enabled; 1653 1654void set_numabalancing_state(bool enabled) 1655{ 1656 numabalancing_enabled = enabled; 1657} 1658#endif /* CONFIG_SCHED_DEBUG */ 1659#endif /* CONFIG_NUMA_BALANCING */ 1660 1661/* 1662 * fork()/clone()-time setup: 1663 */ 1664void sched_fork(struct task_struct *p) 1665{ 1666 unsigned long flags; 1667 int cpu = get_cpu(); 1668 1669 __sched_fork(p); 1670 /* 1671 * We mark the process as running here. This guarantees that 1672 * nobody will actually run it, and a signal or other external 1673 * event cannot wake it up and insert it on the runqueue either. 1674 */ 1675 p->state = TASK_RUNNING; 1676 1677 /* 1678 * Make sure we do not leak PI boosting priority to the child. 1679 */ 1680 p->prio = current->normal_prio; 1681 1682 /* 1683 * Revert to default priority/policy on fork if requested. 1684 */ 1685 if (unlikely(p->sched_reset_on_fork)) { 1686 if (task_has_rt_policy(p)) { 1687 p->policy = SCHED_NORMAL; 1688 p->static_prio = NICE_TO_PRIO(0); 1689 p->rt_priority = 0; 1690 } else if (PRIO_TO_NICE(p->static_prio) < 0) 1691 p->static_prio = NICE_TO_PRIO(0); 1692 1693 p->prio = p->normal_prio = __normal_prio(p); 1694 set_load_weight(p); 1695 1696 /* 1697 * We don't need the reset flag anymore after the fork. It has 1698 * fulfilled its duty: 1699 */ 1700 p->sched_reset_on_fork = 0; 1701 } 1702 1703 if (!rt_prio(p->prio)) 1704 p->sched_class = &fair_sched_class; 1705 1706 if (p->sched_class->task_fork) 1707 p->sched_class->task_fork(p); 1708 1709 /* 1710 * The child is not yet in the pid-hash so no cgroup attach races, 1711 * and the cgroup is pinned to this child due to cgroup_fork() 1712 * is ran before sched_fork(). 1713 * 1714 * Silence PROVE_RCU. 1715 */ 1716 raw_spin_lock_irqsave(&p->pi_lock, flags); 1717 set_task_cpu(p, cpu); 1718 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 1719 1720#if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT) 1721 if (likely(sched_info_on())) 1722 memset(&p->sched_info, 0, sizeof(p->sched_info)); 1723#endif 1724#if defined(CONFIG_SMP) 1725 p->on_cpu = 0; 1726#endif 1727 init_task_preempt_count(p); 1728#ifdef CONFIG_SMP 1729 plist_node_init(&p->pushable_tasks, MAX_PRIO); 1730#endif 1731 1732 put_cpu(); 1733} 1734 1735/* 1736 * wake_up_new_task - wake up a newly created task for the first time. 1737 * 1738 * This function will do some initial scheduler statistics housekeeping 1739 * that must be done for every newly created context, then puts the task 1740 * on the runqueue and wakes it. 1741 */ 1742void wake_up_new_task(struct task_struct *p) 1743{ 1744 unsigned long flags; 1745 struct rq *rq; 1746 1747 raw_spin_lock_irqsave(&p->pi_lock, flags); 1748#ifdef CONFIG_SMP 1749 /* 1750 * Fork balancing, do it here and not earlier because: 1751 * - cpus_allowed can change in the fork path 1752 * - any previously selected cpu might disappear through hotplug 1753 */ 1754 set_task_cpu(p, select_task_rq(p, SD_BALANCE_FORK, 0)); 1755#endif 1756 1757 /* Initialize new task's runnable average */ 1758 init_task_runnable_average(p); 1759 rq = __task_rq_lock(p); 1760 activate_task(rq, p, 0); 1761 p->on_rq = 1; 1762 trace_sched_wakeup_new(p, true); 1763 check_preempt_curr(rq, p, WF_FORK); 1764#ifdef CONFIG_SMP 1765 if (p->sched_class->task_woken) 1766 p->sched_class->task_woken(rq, p); 1767#endif 1768 task_rq_unlock(rq, p, &flags); 1769} 1770 1771#ifdef CONFIG_PREEMPT_NOTIFIERS 1772 1773/** 1774 * preempt_notifier_register - tell me when current is being preempted & rescheduled 1775 * @notifier: notifier struct to register 1776 */ 1777void preempt_notifier_register(struct preempt_notifier *notifier) 1778{ 1779 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers); 1780} 1781EXPORT_SYMBOL_GPL(preempt_notifier_register); 1782 1783/** 1784 * preempt_notifier_unregister - no longer interested in preemption notifications 1785 * @notifier: notifier struct to unregister 1786 * 1787 * This is safe to call from within a preemption notifier. 1788 */ 1789void preempt_notifier_unregister(struct preempt_notifier *notifier) 1790{ 1791 hlist_del(¬ifier->link); 1792} 1793EXPORT_SYMBOL_GPL(preempt_notifier_unregister); 1794 1795static void fire_sched_in_preempt_notifiers(struct task_struct *curr) 1796{ 1797 struct preempt_notifier *notifier; 1798 1799 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link) 1800 notifier->ops->sched_in(notifier, raw_smp_processor_id()); 1801} 1802 1803static void 1804fire_sched_out_preempt_notifiers(struct task_struct *curr, 1805 struct task_struct *next) 1806{ 1807 struct preempt_notifier *notifier; 1808 1809 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link) 1810 notifier->ops->sched_out(notifier, next); 1811} 1812 1813#else /* !CONFIG_PREEMPT_NOTIFIERS */ 1814 1815static void fire_sched_in_preempt_notifiers(struct task_struct *curr) 1816{ 1817} 1818 1819static void 1820fire_sched_out_preempt_notifiers(struct task_struct *curr, 1821 struct task_struct *next) 1822{ 1823} 1824 1825#endif /* CONFIG_PREEMPT_NOTIFIERS */ 1826 1827/** 1828 * prepare_task_switch - prepare to switch tasks 1829 * @rq: the runqueue preparing to switch 1830 * @prev: the current task that is being switched out 1831 * @next: the task we are going to switch to. 1832 * 1833 * This is called with the rq lock held and interrupts off. It must 1834 * be paired with a subsequent finish_task_switch after the context 1835 * switch. 1836 * 1837 * prepare_task_switch sets up locking and calls architecture specific 1838 * hooks. 1839 */ 1840static inline void 1841prepare_task_switch(struct rq *rq, struct task_struct *prev, 1842 struct task_struct *next) 1843{ 1844 trace_sched_switch(prev, next); 1845 sched_info_switch(rq, prev, next); 1846 perf_event_task_sched_out(prev, next); 1847 fire_sched_out_preempt_notifiers(prev, next); 1848 prepare_lock_switch(rq, next); 1849 prepare_arch_switch(next); 1850} 1851 1852/** 1853 * finish_task_switch - clean up after a task-switch 1854 * @rq: runqueue associated with task-switch 1855 * @prev: the thread we just switched away from. 1856 * 1857 * finish_task_switch must be called after the context switch, paired 1858 * with a prepare_task_switch call before the context switch. 1859 * finish_task_switch will reconcile locking set up by prepare_task_switch, 1860 * and do any other architecture-specific cleanup actions. 1861 * 1862 * Note that we may have delayed dropping an mm in context_switch(). If 1863 * so, we finish that here outside of the runqueue lock. (Doing it 1864 * with the lock held can cause deadlocks; see schedule() for 1865 * details.) 1866 */ 1867static void finish_task_switch(struct rq *rq, struct task_struct *prev) 1868 __releases(rq->lock) 1869{ 1870 struct mm_struct *mm = rq->prev_mm; 1871 long prev_state; 1872 1873 rq->prev_mm = NULL; 1874 1875 /* 1876 * A task struct has one reference for the use as "current". 1877 * If a task dies, then it sets TASK_DEAD in tsk->state and calls 1878 * schedule one last time. The schedule call will never return, and 1879 * the scheduled task must drop that reference. 1880 * The test for TASK_DEAD must occur while the runqueue locks are 1881 * still held, otherwise prev could be scheduled on another cpu, die 1882 * there before we look at prev->state, and then the reference would 1883 * be dropped twice. 1884 * Manfred Spraul <manfred@colorfullife.com> 1885 */ 1886 prev_state = prev->state; 1887 vtime_task_switch(prev); 1888 finish_arch_switch(prev); 1889 perf_event_task_sched_in(prev, current); 1890 finish_lock_switch(rq, prev); 1891 finish_arch_post_lock_switch(); 1892 1893 fire_sched_in_preempt_notifiers(current); 1894 if (mm) 1895 mmdrop(mm); 1896 if (unlikely(prev_state == TASK_DEAD)) { 1897 task_numa_free(prev); 1898 1899 /* 1900 * Remove function-return probe instances associated with this 1901 * task and put them back on the free list. 1902 */ 1903 kprobe_flush_task(prev); 1904 put_task_struct(prev); 1905 } 1906 1907 tick_nohz_task_switch(current); 1908} 1909 1910#ifdef CONFIG_SMP 1911 1912/* assumes rq->lock is held */ 1913static inline void pre_schedule(struct rq *rq, struct task_struct *prev) 1914{ 1915 if (prev->sched_class->pre_schedule) 1916 prev->sched_class->pre_schedule(rq, prev); 1917} 1918 1919/* rq->lock is NOT held, but preemption is disabled */ 1920static inline void post_schedule(struct rq *rq) 1921{ 1922 if (rq->post_schedule) { 1923 unsigned long flags; 1924 1925 raw_spin_lock_irqsave(&rq->lock, flags); 1926 if (rq->curr->sched_class->post_schedule) 1927 rq->curr->sched_class->post_schedule(rq); 1928 raw_spin_unlock_irqrestore(&rq->lock, flags); 1929 1930 rq->post_schedule = 0; 1931 } 1932} 1933 1934#else 1935 1936static inline void pre_schedule(struct rq *rq, struct task_struct *p) 1937{ 1938} 1939 1940static inline void post_schedule(struct rq *rq) 1941{ 1942} 1943 1944#endif 1945 1946/** 1947 * schedule_tail - first thing a freshly forked thread must call. 1948 * @prev: the thread we just switched away from. 1949 */ 1950asmlinkage void schedule_tail(struct task_struct *prev) 1951 __releases(rq->lock) 1952{ 1953 struct rq *rq = this_rq(); 1954 1955 finish_task_switch(rq, prev); 1956 1957 /* 1958 * FIXME: do we need to worry about rq being invalidated by the 1959 * task_switch? 1960 */ 1961 post_schedule(rq); 1962 1963#ifdef __ARCH_WANT_UNLOCKED_CTXSW 1964 /* In this case, finish_task_switch does not reenable preemption */ 1965 preempt_enable(); 1966#endif 1967 if (current->set_child_tid) 1968 put_user(task_pid_vnr(current), current->set_child_tid); 1969} 1970 1971/* 1972 * context_switch - switch to the new MM and the new 1973 * thread's register state. 1974 */ 1975static inline void 1976context_switch(struct rq *rq, struct task_struct *prev, 1977 struct task_struct *next) 1978{ 1979 struct mm_struct *mm, *oldmm; 1980 1981 prepare_task_switch(rq, prev, next); 1982 1983 mm = next->mm; 1984 oldmm = prev->active_mm; 1985 /* 1986 * For paravirt, this is coupled with an exit in switch_to to 1987 * combine the page table reload and the switch backend into 1988 * one hypercall. 1989 */ 1990 arch_start_context_switch(prev); 1991 1992 if (!mm) { 1993 next->active_mm = oldmm; 1994 atomic_inc(&oldmm->mm_count); 1995 enter_lazy_tlb(oldmm, next); 1996 } else 1997 switch_mm(oldmm, mm, next); 1998 1999 if (!prev->mm) { 2000 prev->active_mm = NULL; 2001 rq->prev_mm = oldmm; 2002 } 2003 /* 2004 * Since the runqueue lock will be released by the next 2005 * task (which is an invalid locking op but in the case 2006 * of the scheduler it's an obvious special-case), so we 2007 * do an early lockdep release here: 2008 */ 2009#ifndef __ARCH_WANT_UNLOCKED_CTXSW 2010 spin_release(&rq->lock.dep_map, 1, _THIS_IP_); 2011#endif 2012 2013 context_tracking_task_switch(prev, next); 2014 /* Here we just switch the register state and the stack. */ 2015 switch_to(prev, next, prev); 2016 2017 barrier(); 2018 /* 2019 * this_rq must be evaluated again because prev may have moved 2020 * CPUs since it called schedule(), thus the 'rq' on its stack 2021 * frame will be invalid. 2022 */ 2023 finish_task_switch(this_rq(), prev); 2024} 2025 2026/* 2027 * nr_running and nr_context_switches: 2028 * 2029 * externally visible scheduler statistics: current number of runnable 2030 * threads, total number of context switches performed since bootup. 2031 */ 2032unsigned long nr_running(void) 2033{ 2034 unsigned long i, sum = 0; 2035 2036 for_each_online_cpu(i) 2037 sum += cpu_rq(i)->nr_running; 2038 2039 return sum; 2040} 2041 2042unsigned long long nr_context_switches(void) 2043{ 2044 int i; 2045 unsigned long long sum = 0; 2046 2047 for_each_possible_cpu(i) 2048 sum += cpu_rq(i)->nr_switches; 2049 2050 return sum; 2051} 2052 2053unsigned long nr_iowait(void) 2054{ 2055 unsigned long i, sum = 0; 2056 2057 for_each_possible_cpu(i) 2058 sum += atomic_read(&cpu_rq(i)->nr_iowait); 2059 2060 return sum; 2061} 2062 2063unsigned long nr_iowait_cpu(int cpu) 2064{ 2065 struct rq *this = cpu_rq(cpu); 2066 return atomic_read(&this->nr_iowait); 2067} 2068 2069#ifdef CONFIG_SMP 2070 2071/* 2072 * sched_exec - execve() is a valuable balancing opportunity, because at 2073 * this point the task has the smallest effective memory and cache footprint. 2074 */ 2075void sched_exec(void) 2076{ 2077 struct task_struct *p = current; 2078 unsigned long flags; 2079 int dest_cpu; 2080 2081 raw_spin_lock_irqsave(&p->pi_lock, flags); 2082 dest_cpu = p->sched_class->select_task_rq(p, SD_BALANCE_EXEC, 0); 2083 if (dest_cpu == smp_processor_id()) 2084 goto unlock; 2085 2086 if (likely(cpu_active(dest_cpu))) { 2087 struct migration_arg arg = { p, dest_cpu }; 2088 2089 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 2090 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg); 2091 return; 2092 } 2093unlock: 2094 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 2095} 2096 2097#endif 2098 2099DEFINE_PER_CPU(struct kernel_stat, kstat); 2100DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat); 2101 2102EXPORT_PER_CPU_SYMBOL(kstat); 2103EXPORT_PER_CPU_SYMBOL(kernel_cpustat); 2104 2105/* 2106 * Return any ns on the sched_clock that have not yet been accounted in 2107 * @p in case that task is currently running. 2108 * 2109 * Called with task_rq_lock() held on @rq. 2110 */ 2111static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq) 2112{ 2113 u64 ns = 0; 2114 2115 if (task_current(rq, p)) { 2116 update_rq_clock(rq); 2117 ns = rq_clock_task(rq) - p->se.exec_start; 2118 if ((s64)ns < 0) 2119 ns = 0; 2120 } 2121 2122 return ns; 2123} 2124 2125unsigned long long task_delta_exec(struct task_struct *p) 2126{ 2127 unsigned long flags; 2128 struct rq *rq; 2129 u64 ns = 0; 2130 2131 rq = task_rq_lock(p, &flags); 2132 ns = do_task_delta_exec(p, rq); 2133 task_rq_unlock(rq, p, &flags); 2134 2135 return ns; 2136} 2137 2138/* 2139 * Return accounted runtime for the task. 2140 * In case the task is currently running, return the runtime plus current's 2141 * pending runtime that have not been accounted yet. 2142 */ 2143unsigned long long task_sched_runtime(struct task_struct *p) 2144{ 2145 unsigned long flags; 2146 struct rq *rq; 2147 u64 ns = 0; 2148 2149 rq = task_rq_lock(p, &flags); 2150 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq); 2151 task_rq_unlock(rq, p, &flags); 2152 2153 return ns; 2154} 2155 2156/* 2157 * This function gets called by the timer code, with HZ frequency. 2158 * We call it with interrupts disabled. 2159 */ 2160void scheduler_tick(void) 2161{ 2162 int cpu = smp_processor_id(); 2163 struct rq *rq = cpu_rq(cpu); 2164 struct task_struct *curr = rq->curr; 2165 2166 sched_clock_tick(); 2167 2168 raw_spin_lock(&rq->lock); 2169 update_rq_clock(rq); 2170 curr->sched_class->task_tick(rq, curr, 0); 2171 update_cpu_load_active(rq); 2172 raw_spin_unlock(&rq->lock); 2173 2174 perf_event_task_tick(); 2175 2176#ifdef CONFIG_SMP 2177 rq->idle_balance = idle_cpu(cpu); 2178 trigger_load_balance(rq, cpu); 2179#endif 2180 rq_last_tick_reset(rq); 2181} 2182 2183#ifdef CONFIG_NO_HZ_FULL 2184/** 2185 * scheduler_tick_max_deferment 2186 * 2187 * Keep at least one tick per second when a single 2188 * active task is running because the scheduler doesn't 2189 * yet completely support full dynticks environment. 2190 * 2191 * This makes sure that uptime, CFS vruntime, load 2192 * balancing, etc... continue to move forward, even 2193 * with a very low granularity. 2194 * 2195 * Return: Maximum deferment in nanoseconds. 2196 */ 2197u64 scheduler_tick_max_deferment(void) 2198{ 2199 struct rq *rq = this_rq(); 2200 unsigned long next, now = ACCESS_ONCE(jiffies); 2201 2202 next = rq->last_sched_tick + HZ; 2203 2204 if (time_before_eq(next, now)) 2205 return 0; 2206 2207 return jiffies_to_usecs(next - now) * NSEC_PER_USEC; 2208} 2209#endif 2210 2211notrace unsigned long get_parent_ip(unsigned long addr) 2212{ 2213 if (in_lock_functions(addr)) { 2214 addr = CALLER_ADDR2; 2215 if (in_lock_functions(addr)) 2216 addr = CALLER_ADDR3; 2217 } 2218 return addr; 2219} 2220 2221#if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \ 2222 defined(CONFIG_PREEMPT_TRACER)) 2223 2224void __kprobes preempt_count_add(int val) 2225{ 2226#ifdef CONFIG_DEBUG_PREEMPT 2227 /* 2228 * Underflow? 2229 */ 2230 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0))) 2231 return; 2232#endif 2233 __preempt_count_add(val); 2234#ifdef CONFIG_DEBUG_PREEMPT 2235 /* 2236 * Spinlock count overflowing soon? 2237 */ 2238 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >= 2239 PREEMPT_MASK - 10); 2240#endif 2241 if (preempt_count() == val) 2242 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1)); 2243} 2244EXPORT_SYMBOL(preempt_count_add); 2245 2246void __kprobes preempt_count_sub(int val) 2247{ 2248#ifdef CONFIG_DEBUG_PREEMPT 2249 /* 2250 * Underflow? 2251 */ 2252 if (DEBUG_LOCKS_WARN_ON(val > preempt_count())) 2253 return; 2254 /* 2255 * Is the spinlock portion underflowing? 2256 */ 2257 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) && 2258 !(preempt_count() & PREEMPT_MASK))) 2259 return; 2260#endif 2261 2262 if (preempt_count() == val) 2263 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1)); 2264 __preempt_count_sub(val); 2265} 2266EXPORT_SYMBOL(preempt_count_sub); 2267 2268#endif 2269 2270/* 2271 * Print scheduling while atomic bug: 2272 */ 2273static noinline void __schedule_bug(struct task_struct *prev) 2274{ 2275 if (oops_in_progress) 2276 return; 2277 2278 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n", 2279 prev->comm, prev->pid, preempt_count()); 2280 2281 debug_show_held_locks(prev); 2282 print_modules(); 2283 if (irqs_disabled()) 2284 print_irqtrace_events(prev); 2285 dump_stack(); 2286 add_taint(TAINT_WARN, LOCKDEP_STILL_OK); 2287} 2288 2289/* 2290 * Various schedule()-time debugging checks and statistics: 2291 */ 2292static inline void schedule_debug(struct task_struct *prev) 2293{ 2294 /* 2295 * Test if we are atomic. Since do_exit() needs to call into 2296 * schedule() atomically, we ignore that path for now. 2297 * Otherwise, whine if we are scheduling when we should not be. 2298 */ 2299 if (unlikely(in_atomic_preempt_off() && !prev->exit_state)) 2300 __schedule_bug(prev); 2301 rcu_sleep_check(); 2302 2303 profile_hit(SCHED_PROFILING, __builtin_return_address(0)); 2304 2305 schedstat_inc(this_rq(), sched_count); 2306} 2307 2308static void put_prev_task(struct rq *rq, struct task_struct *prev) 2309{ 2310 if (prev->on_rq || rq->skip_clock_update < 0) 2311 update_rq_clock(rq); 2312 prev->sched_class->put_prev_task(rq, prev); 2313} 2314 2315/* 2316 * Pick up the highest-prio task: 2317 */ 2318static inline struct task_struct * 2319pick_next_task(struct rq *rq) 2320{ 2321 const struct sched_class *class; 2322 struct task_struct *p; 2323 2324 /* 2325 * Optimization: we know that if all tasks are in 2326 * the fair class we can call that function directly: 2327 */ 2328 if (likely(rq->nr_running == rq->cfs.h_nr_running)) { 2329 p = fair_sched_class.pick_next_task(rq); 2330 if (likely(p)) 2331 return p; 2332 } 2333 2334 for_each_class(class) { 2335 p = class->pick_next_task(rq); 2336 if (p) 2337 return p; 2338 } 2339 2340 BUG(); /* the idle class will always have a runnable task */ 2341} 2342 2343/* 2344 * __schedule() is the main scheduler function. 2345 * 2346 * The main means of driving the scheduler and thus entering this function are: 2347 * 2348 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc. 2349 * 2350 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return 2351 * paths. For example, see arch/x86/entry_64.S. 2352 * 2353 * To drive preemption between tasks, the scheduler sets the flag in timer 2354 * interrupt handler scheduler_tick(). 2355 * 2356 * 3. Wakeups don't really cause entry into schedule(). They add a 2357 * task to the run-queue and that's it. 2358 * 2359 * Now, if the new task added to the run-queue preempts the current 2360 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets 2361 * called on the nearest possible occasion: 2362 * 2363 * - If the kernel is preemptible (CONFIG_PREEMPT=y): 2364 * 2365 * - in syscall or exception context, at the next outmost 2366 * preempt_enable(). (this might be as soon as the wake_up()'s 2367 * spin_unlock()!) 2368 * 2369 * - in IRQ context, return from interrupt-handler to 2370 * preemptible context 2371 * 2372 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set) 2373 * then at the next: 2374 * 2375 * - cond_resched() call 2376 * - explicit schedule() call 2377 * - return from syscall or exception to user-space 2378 * - return from interrupt-handler to user-space 2379 */ 2380static void __sched __schedule(void) 2381{ 2382 struct task_struct *prev, *next; 2383 unsigned long *switch_count; 2384 struct rq *rq; 2385 int cpu; 2386 2387need_resched: 2388 preempt_disable(); 2389 cpu = smp_processor_id(); 2390 rq = cpu_rq(cpu); 2391 rcu_note_context_switch(cpu); 2392 prev = rq->curr; 2393 2394 schedule_debug(prev); 2395 2396 if (sched_feat(HRTICK)) 2397 hrtick_clear(rq); 2398 2399 /* 2400 * Make sure that signal_pending_state()->signal_pending() below 2401 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE) 2402 * done by the caller to avoid the race with signal_wake_up(). 2403 */ 2404 smp_mb__before_spinlock(); 2405 raw_spin_lock_irq(&rq->lock); 2406 2407 switch_count = &prev->nivcsw; 2408 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) { 2409 if (unlikely(signal_pending_state(prev->state, prev))) { 2410 prev->state = TASK_RUNNING; 2411 } else { 2412 deactivate_task(rq, prev, DEQUEUE_SLEEP); 2413 prev->on_rq = 0; 2414 2415 /* 2416 * If a worker went to sleep, notify and ask workqueue 2417 * whether it wants to wake up a task to maintain 2418 * concurrency. 2419 */ 2420 if (prev->flags & PF_WQ_WORKER) { 2421 struct task_struct *to_wakeup; 2422 2423 to_wakeup = wq_worker_sleeping(prev, cpu); 2424 if (to_wakeup) 2425 try_to_wake_up_local(to_wakeup); 2426 } 2427 } 2428 switch_count = &prev->nvcsw; 2429 } 2430 2431 pre_schedule(rq, prev); 2432 2433 if (unlikely(!rq->nr_running)) 2434 idle_balance(cpu, rq); 2435 2436 put_prev_task(rq, prev); 2437 next = pick_next_task(rq); 2438 clear_tsk_need_resched(prev); 2439 clear_preempt_need_resched(); 2440 rq->skip_clock_update = 0; 2441 2442 if (likely(prev != next)) { 2443 rq->nr_switches++; 2444 rq->curr = next; 2445 ++*switch_count; 2446 2447 context_switch(rq, prev, next); /* unlocks the rq */ 2448 /* 2449 * The context switch have flipped the stack from under us 2450 * and restored the local variables which were saved when 2451 * this task called schedule() in the past. prev == current 2452 * is still correct, but it can be moved to another cpu/rq. 2453 */ 2454 cpu = smp_processor_id(); 2455 rq = cpu_rq(cpu); 2456 } else 2457 raw_spin_unlock_irq(&rq->lock); 2458 2459 post_schedule(rq); 2460 2461 sched_preempt_enable_no_resched(); 2462 if (need_resched()) 2463 goto need_resched; 2464} 2465 2466static inline void sched_submit_work(struct task_struct *tsk) 2467{ 2468 if (!tsk->state || tsk_is_pi_blocked(tsk)) 2469 return; 2470 /* 2471 * If we are going to sleep and we have plugged IO queued, 2472 * make sure to submit it to avoid deadlocks. 2473 */ 2474 if (blk_needs_flush_plug(tsk)) 2475 blk_schedule_flush_plug(tsk); 2476} 2477 2478asmlinkage void __sched schedule(void) 2479{ 2480 struct task_struct *tsk = current; 2481 2482 sched_submit_work(tsk); 2483 __schedule(); 2484} 2485EXPORT_SYMBOL(schedule); 2486 2487#ifdef CONFIG_CONTEXT_TRACKING 2488asmlinkage void __sched schedule_user(void) 2489{ 2490 /* 2491 * If we come here after a random call to set_need_resched(), 2492 * or we have been woken up remotely but the IPI has not yet arrived, 2493 * we haven't yet exited the RCU idle mode. Do it here manually until 2494 * we find a better solution. 2495 */ 2496 user_exit(); 2497 schedule(); 2498 user_enter(); 2499} 2500#endif 2501 2502/** 2503 * schedule_preempt_disabled - called with preemption disabled 2504 * 2505 * Returns with preemption disabled. Note: preempt_count must be 1 2506 */ 2507void __sched schedule_preempt_disabled(void) 2508{ 2509 sched_preempt_enable_no_resched(); 2510 schedule(); 2511 preempt_disable(); 2512} 2513 2514#ifdef CONFIG_PREEMPT 2515/* 2516 * this is the entry point to schedule() from in-kernel preemption 2517 * off of preempt_enable. Kernel preemptions off return from interrupt 2518 * occur there and call schedule directly. 2519 */ 2520asmlinkage void __sched notrace preempt_schedule(void) 2521{ 2522 /* 2523 * If there is a non-zero preempt_count or interrupts are disabled, 2524 * we do not want to preempt the current task. Just return.. 2525 */ 2526 if (likely(!preemptible())) 2527 return; 2528 2529 do { 2530 __preempt_count_add(PREEMPT_ACTIVE); 2531 __schedule(); 2532 __preempt_count_sub(PREEMPT_ACTIVE); 2533 2534 /* 2535 * Check again in case we missed a preemption opportunity 2536 * between schedule and now. 2537 */ 2538 barrier(); 2539 } while (need_resched()); 2540} 2541EXPORT_SYMBOL(preempt_schedule); 2542 2543/* 2544 * this is the entry point to schedule() from kernel preemption 2545 * off of irq context. 2546 * Note, that this is called and return with irqs disabled. This will 2547 * protect us against recursive calling from irq. 2548 */ 2549asmlinkage void __sched preempt_schedule_irq(void) 2550{ 2551 enum ctx_state prev_state; 2552 2553 /* Catch callers which need to be fixed */ 2554 BUG_ON(preempt_count() || !irqs_disabled()); 2555 2556 prev_state = exception_enter(); 2557 2558 do { 2559 __preempt_count_add(PREEMPT_ACTIVE); 2560 local_irq_enable(); 2561 __schedule(); 2562 local_irq_disable(); 2563 __preempt_count_sub(PREEMPT_ACTIVE); 2564 2565 /* 2566 * Check again in case we missed a preemption opportunity 2567 * between schedule and now. 2568 */ 2569 barrier(); 2570 } while (need_resched()); 2571 2572 exception_exit(prev_state); 2573} 2574 2575#endif /* CONFIG_PREEMPT */ 2576 2577int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags, 2578 void *key) 2579{ 2580 return try_to_wake_up(curr->private, mode, wake_flags); 2581} 2582EXPORT_SYMBOL(default_wake_function); 2583 2584/* 2585 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just 2586 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve 2587 * number) then we wake all the non-exclusive tasks and one exclusive task. 2588 * 2589 * There are circumstances in which we can try to wake a task which has already 2590 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns 2591 * zero in this (rare) case, and we handle it by continuing to scan the queue. 2592 */ 2593static void __wake_up_common(wait_queue_head_t *q, unsigned int mode, 2594 int nr_exclusive, int wake_flags, void *key) 2595{ 2596 wait_queue_t *curr, *next; 2597 2598 list_for_each_entry_safe(curr, next, &q->task_list, task_list) { 2599 unsigned flags = curr->flags; 2600 2601 if (curr->func(curr, mode, wake_flags, key) && 2602 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive) 2603 break; 2604 } 2605} 2606 2607/** 2608 * __wake_up - wake up threads blocked on a waitqueue. 2609 * @q: the waitqueue 2610 * @mode: which threads 2611 * @nr_exclusive: how many wake-one or wake-many threads to wake up 2612 * @key: is directly passed to the wakeup function 2613 * 2614 * It may be assumed that this function implies a write memory barrier before 2615 * changing the task state if and only if any tasks are woken up. 2616 */ 2617void __wake_up(wait_queue_head_t *q, unsigned int mode, 2618 int nr_exclusive, void *key) 2619{ 2620 unsigned long flags; 2621 2622 spin_lock_irqsave(&q->lock, flags); 2623 __wake_up_common(q, mode, nr_exclusive, 0, key); 2624 spin_unlock_irqrestore(&q->lock, flags); 2625} 2626EXPORT_SYMBOL(__wake_up); 2627 2628/* 2629 * Same as __wake_up but called with the spinlock in wait_queue_head_t held. 2630 */ 2631void __wake_up_locked(wait_queue_head_t *q, unsigned int mode, int nr) 2632{ 2633 __wake_up_common(q, mode, nr, 0, NULL); 2634} 2635EXPORT_SYMBOL_GPL(__wake_up_locked); 2636 2637void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key) 2638{ 2639 __wake_up_common(q, mode, 1, 0, key); 2640} 2641EXPORT_SYMBOL_GPL(__wake_up_locked_key); 2642 2643/** 2644 * __wake_up_sync_key - wake up threads blocked on a waitqueue. 2645 * @q: the waitqueue 2646 * @mode: which threads 2647 * @nr_exclusive: how many wake-one or wake-many threads to wake up 2648 * @key: opaque value to be passed to wakeup targets 2649 * 2650 * The sync wakeup differs that the waker knows that it will schedule 2651 * away soon, so while the target thread will be woken up, it will not 2652 * be migrated to another CPU - ie. the two threads are 'synchronized' 2653 * with each other. This can prevent needless bouncing between CPUs. 2654 * 2655 * On UP it can prevent extra preemption. 2656 * 2657 * It may be assumed that this function implies a write memory barrier before 2658 * changing the task state if and only if any tasks are woken up. 2659 */ 2660void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode, 2661 int nr_exclusive, void *key) 2662{ 2663 unsigned long flags; 2664 int wake_flags = WF_SYNC; 2665 2666 if (unlikely(!q)) 2667 return; 2668 2669 if (unlikely(nr_exclusive != 1)) 2670 wake_flags = 0; 2671 2672 spin_lock_irqsave(&q->lock, flags); 2673 __wake_up_common(q, mode, nr_exclusive, wake_flags, key); 2674 spin_unlock_irqrestore(&q->lock, flags); 2675} 2676EXPORT_SYMBOL_GPL(__wake_up_sync_key); 2677 2678/* 2679 * __wake_up_sync - see __wake_up_sync_key() 2680 */ 2681void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive) 2682{ 2683 __wake_up_sync_key(q, mode, nr_exclusive, NULL); 2684} 2685EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */ 2686 2687/** 2688 * complete: - signals a single thread waiting on this completion 2689 * @x: holds the state of this particular completion 2690 * 2691 * This will wake up a single thread waiting on this completion. Threads will be 2692 * awakened in the same order in which they were queued. 2693 * 2694 * See also complete_all(), wait_for_completion() and related routines. 2695 * 2696 * It may be assumed that this function implies a write memory barrier before 2697 * changing the task state if and only if any tasks are woken up. 2698 */ 2699void complete(struct completion *x) 2700{ 2701 unsigned long flags; 2702 2703 spin_lock_irqsave(&x->wait.lock, flags); 2704 x->done++; 2705 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL); 2706 spin_unlock_irqrestore(&x->wait.lock, flags); 2707} 2708EXPORT_SYMBOL(complete); 2709 2710/** 2711 * complete_all: - signals all threads waiting on this completion 2712 * @x: holds the state of this particular completion 2713 * 2714 * This will wake up all threads waiting on this particular completion event. 2715 * 2716 * It may be assumed that this function implies a write memory barrier before 2717 * changing the task state if and only if any tasks are woken up. 2718 */ 2719void complete_all(struct completion *x) 2720{ 2721 unsigned long flags; 2722 2723 spin_lock_irqsave(&x->wait.lock, flags); 2724 x->done += UINT_MAX/2; 2725 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL); 2726 spin_unlock_irqrestore(&x->wait.lock, flags); 2727} 2728EXPORT_SYMBOL(complete_all); 2729 2730static inline long __sched 2731do_wait_for_common(struct completion *x, 2732 long (*action)(long), long timeout, int state) 2733{ 2734 if (!x->done) { 2735 DECLARE_WAITQUEUE(wait, current); 2736 2737 __add_wait_queue_tail_exclusive(&x->wait, &wait); 2738 do { 2739 if (signal_pending_state(state, current)) { 2740 timeout = -ERESTARTSYS; 2741 break; 2742 } 2743 __set_current_state(state); 2744 spin_unlock_irq(&x->wait.lock); 2745 timeout = action(timeout); 2746 spin_lock_irq(&x->wait.lock); 2747 } while (!x->done && timeout); 2748 __remove_wait_queue(&x->wait, &wait); 2749 if (!x->done) 2750 return timeout; 2751 } 2752 x->done--; 2753 return timeout ?: 1; 2754} 2755 2756static inline long __sched 2757__wait_for_common(struct completion *x, 2758 long (*action)(long), long timeout, int state) 2759{ 2760 might_sleep(); 2761 2762 spin_lock_irq(&x->wait.lock); 2763 timeout = do_wait_for_common(x, action, timeout, state); 2764 spin_unlock_irq(&x->wait.lock); 2765 return timeout; 2766} 2767 2768static long __sched 2769wait_for_common(struct completion *x, long timeout, int state) 2770{ 2771 return __wait_for_common(x, schedule_timeout, timeout, state); 2772} 2773 2774static long __sched 2775wait_for_common_io(struct completion *x, long timeout, int state) 2776{ 2777 return __wait_for_common(x, io_schedule_timeout, timeout, state); 2778} 2779 2780/** 2781 * wait_for_completion: - waits for completion of a task 2782 * @x: holds the state of this particular completion 2783 * 2784 * This waits to be signaled for completion of a specific task. It is NOT 2785 * interruptible and there is no timeout. 2786 * 2787 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout 2788 * and interrupt capability. Also see complete(). 2789 */ 2790void __sched wait_for_completion(struct completion *x) 2791{ 2792 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE); 2793} 2794EXPORT_SYMBOL(wait_for_completion); 2795 2796/** 2797 * wait_for_completion_timeout: - waits for completion of a task (w/timeout) 2798 * @x: holds the state of this particular completion 2799 * @timeout: timeout value in jiffies 2800 * 2801 * This waits for either a completion of a specific task to be signaled or for a 2802 * specified timeout to expire. The timeout is in jiffies. It is not 2803 * interruptible. 2804 * 2805 * Return: 0 if timed out, and positive (at least 1, or number of jiffies left 2806 * till timeout) if completed. 2807 */ 2808unsigned long __sched 2809wait_for_completion_timeout(struct completion *x, unsigned long timeout) 2810{ 2811 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE); 2812} 2813EXPORT_SYMBOL(wait_for_completion_timeout); 2814 2815/** 2816 * wait_for_completion_io: - waits for completion of a task 2817 * @x: holds the state of this particular completion 2818 * 2819 * This waits to be signaled for completion of a specific task. It is NOT 2820 * interruptible and there is no timeout. The caller is accounted as waiting 2821 * for IO. 2822 */ 2823void __sched wait_for_completion_io(struct completion *x) 2824{ 2825 wait_for_common_io(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE); 2826} 2827EXPORT_SYMBOL(wait_for_completion_io); 2828 2829/** 2830 * wait_for_completion_io_timeout: - waits for completion of a task (w/timeout) 2831 * @x: holds the state of this particular completion 2832 * @timeout: timeout value in jiffies 2833 * 2834 * This waits for either a completion of a specific task to be signaled or for a 2835 * specified timeout to expire. The timeout is in jiffies. It is not 2836 * interruptible. The caller is accounted as waiting for IO. 2837 * 2838 * Return: 0 if timed out, and positive (at least 1, or number of jiffies left 2839 * till timeout) if completed. 2840 */ 2841unsigned long __sched 2842wait_for_completion_io_timeout(struct completion *x, unsigned long timeout) 2843{ 2844 return wait_for_common_io(x, timeout, TASK_UNINTERRUPTIBLE); 2845} 2846EXPORT_SYMBOL(wait_for_completion_io_timeout); 2847 2848/** 2849 * wait_for_completion_interruptible: - waits for completion of a task (w/intr) 2850 * @x: holds the state of this particular completion 2851 * 2852 * This waits for completion of a specific task to be signaled. It is 2853 * interruptible. 2854 * 2855 * Return: -ERESTARTSYS if interrupted, 0 if completed. 2856 */ 2857int __sched wait_for_completion_interruptible(struct completion *x) 2858{ 2859 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE); 2860 if (t == -ERESTARTSYS) 2861 return t; 2862 return 0; 2863} 2864EXPORT_SYMBOL(wait_for_completion_interruptible); 2865 2866/** 2867 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr)) 2868 * @x: holds the state of this particular completion 2869 * @timeout: timeout value in jiffies 2870 * 2871 * This waits for either a completion of a specific task to be signaled or for a 2872 * specified timeout to expire. It is interruptible. The timeout is in jiffies. 2873 * 2874 * Return: -ERESTARTSYS if interrupted, 0 if timed out, positive (at least 1, 2875 * or number of jiffies left till timeout) if completed. 2876 */ 2877long __sched 2878wait_for_completion_interruptible_timeout(struct completion *x, 2879 unsigned long timeout) 2880{ 2881 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE); 2882} 2883EXPORT_SYMBOL(wait_for_completion_interruptible_timeout); 2884 2885/** 2886 * wait_for_completion_killable: - waits for completion of a task (killable) 2887 * @x: holds the state of this particular completion 2888 * 2889 * This waits to be signaled for completion of a specific task. It can be 2890 * interrupted by a kill signal. 2891 * 2892 * Return: -ERESTARTSYS if interrupted, 0 if completed. 2893 */ 2894int __sched wait_for_completion_killable(struct completion *x) 2895{ 2896 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE); 2897 if (t == -ERESTARTSYS) 2898 return t; 2899 return 0; 2900} 2901EXPORT_SYMBOL(wait_for_completion_killable); 2902 2903/** 2904 * wait_for_completion_killable_timeout: - waits for completion of a task (w/(to,killable)) 2905 * @x: holds the state of this particular completion 2906 * @timeout: timeout value in jiffies 2907 * 2908 * This waits for either a completion of a specific task to be 2909 * signaled or for a specified timeout to expire. It can be 2910 * interrupted by a kill signal. The timeout is in jiffies. 2911 * 2912 * Return: -ERESTARTSYS if interrupted, 0 if timed out, positive (at least 1, 2913 * or number of jiffies left till timeout) if completed. 2914 */ 2915long __sched 2916wait_for_completion_killable_timeout(struct completion *x, 2917 unsigned long timeout) 2918{ 2919 return wait_for_common(x, timeout, TASK_KILLABLE); 2920} 2921EXPORT_SYMBOL(wait_for_completion_killable_timeout); 2922 2923/** 2924 * try_wait_for_completion - try to decrement a completion without blocking 2925 * @x: completion structure 2926 * 2927 * Return: 0 if a decrement cannot be done without blocking 2928 * 1 if a decrement succeeded. 2929 * 2930 * If a completion is being used as a counting completion, 2931 * attempt to decrement the counter without blocking. This 2932 * enables us to avoid waiting if the resource the completion 2933 * is protecting is not available. 2934 */ 2935bool try_wait_for_completion(struct completion *x) 2936{ 2937 unsigned long flags; 2938 int ret = 1; 2939 2940 spin_lock_irqsave(&x->wait.lock, flags); 2941 if (!x->done) 2942 ret = 0; 2943 else 2944 x->done--; 2945 spin_unlock_irqrestore(&x->wait.lock, flags); 2946 return ret; 2947} 2948EXPORT_SYMBOL(try_wait_for_completion); 2949 2950/** 2951 * completion_done - Test to see if a completion has any waiters 2952 * @x: completion structure 2953 * 2954 * Return: 0 if there are waiters (wait_for_completion() in progress) 2955 * 1 if there are no waiters. 2956 * 2957 */ 2958bool completion_done(struct completion *x) 2959{ 2960 unsigned long flags; 2961 int ret = 1; 2962 2963 spin_lock_irqsave(&x->wait.lock, flags); 2964 if (!x->done) 2965 ret = 0; 2966 spin_unlock_irqrestore(&x->wait.lock, flags); 2967 return ret; 2968} 2969EXPORT_SYMBOL(completion_done); 2970 2971static long __sched 2972sleep_on_common(wait_queue_head_t *q, int state, long timeout) 2973{ 2974 unsigned long flags; 2975 wait_queue_t wait; 2976 2977 init_waitqueue_entry(&wait, current); 2978 2979 __set_current_state(state); 2980 2981 spin_lock_irqsave(&q->lock, flags); 2982 __add_wait_queue(q, &wait); 2983 spin_unlock(&q->lock); 2984 timeout = schedule_timeout(timeout); 2985 spin_lock_irq(&q->lock); 2986 __remove_wait_queue(q, &wait); 2987 spin_unlock_irqrestore(&q->lock, flags); 2988 2989 return timeout; 2990} 2991 2992void __sched interruptible_sleep_on(wait_queue_head_t *q) 2993{ 2994 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT); 2995} 2996EXPORT_SYMBOL(interruptible_sleep_on); 2997 2998long __sched 2999interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout) 3000{ 3001 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout); 3002} 3003EXPORT_SYMBOL(interruptible_sleep_on_timeout); 3004 3005void __sched sleep_on(wait_queue_head_t *q) 3006{ 3007 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT); 3008} 3009EXPORT_SYMBOL(sleep_on); 3010 3011long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout) 3012{ 3013 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout); 3014} 3015EXPORT_SYMBOL(sleep_on_timeout); 3016 3017#ifdef CONFIG_RT_MUTEXES 3018 3019/* 3020 * rt_mutex_setprio - set the current priority of a task 3021 * @p: task 3022 * @prio: prio value (kernel-internal form) 3023 * 3024 * This function changes the 'effective' priority of a task. It does 3025 * not touch ->normal_prio like __setscheduler(). 3026 * 3027 * Used by the rt_mutex code to implement priority inheritance logic. 3028 */ 3029void rt_mutex_setprio(struct task_struct *p, int prio) 3030{ 3031 int oldprio, on_rq, running; 3032 struct rq *rq; 3033 const struct sched_class *prev_class; 3034 3035 BUG_ON(prio < 0 || prio > MAX_PRIO); 3036 3037 rq = __task_rq_lock(p); 3038 3039 /* 3040 * Idle task boosting is a nono in general. There is one 3041 * exception, when PREEMPT_RT and NOHZ is active: 3042 * 3043 * The idle task calls get_next_timer_interrupt() and holds 3044 * the timer wheel base->lock on the CPU and another CPU wants 3045 * to access the timer (probably to cancel it). We can safely 3046 * ignore the boosting request, as the idle CPU runs this code 3047 * with interrupts disabled and will complete the lock 3048 * protected section without being interrupted. So there is no 3049 * real need to boost. 3050 */ 3051 if (unlikely(p == rq->idle)) { 3052 WARN_ON(p != rq->curr); 3053 WARN_ON(p->pi_blocked_on); 3054 goto out_unlock; 3055 } 3056 3057 trace_sched_pi_setprio(p, prio); 3058 oldprio = p->prio; 3059 prev_class = p->sched_class; 3060 on_rq = p->on_rq; 3061 running = task_current(rq, p); 3062 if (on_rq) 3063 dequeue_task(rq, p, 0); 3064 if (running) 3065 p->sched_class->put_prev_task(rq, p); 3066 3067 if (rt_prio(prio)) 3068 p->sched_class = &rt_sched_class; 3069 else 3070 p->sched_class = &fair_sched_class; 3071 3072 p->prio = prio; 3073 3074 if (running) 3075 p->sched_class->set_curr_task(rq); 3076 if (on_rq) 3077 enqueue_task(rq, p, oldprio < prio ? ENQUEUE_HEAD : 0); 3078 3079 check_class_changed(rq, p, prev_class, oldprio); 3080out_unlock: 3081 __task_rq_unlock(rq); 3082} 3083#endif 3084void set_user_nice(struct task_struct *p, long nice) 3085{ 3086 int old_prio, delta, on_rq; 3087 unsigned long flags; 3088 struct rq *rq; 3089 3090 if (TASK_NICE(p) == nice || nice < -20 || nice > 19) 3091 return; 3092 /* 3093 * We have to be careful, if called from sys_setpriority(), 3094 * the task might be in the middle of scheduling on another CPU. 3095 */ 3096 rq = task_rq_lock(p, &flags); 3097 /* 3098 * The RT priorities are set via sched_setscheduler(), but we still 3099 * allow the 'normal' nice value to be set - but as expected 3100 * it wont have any effect on scheduling until the task is 3101 * SCHED_FIFO/SCHED_RR: 3102 */ 3103 if (task_has_rt_policy(p)) { 3104 p->static_prio = NICE_TO_PRIO(nice); 3105 goto out_unlock; 3106 } 3107 on_rq = p->on_rq; 3108 if (on_rq) 3109 dequeue_task(rq, p, 0); 3110 3111 p->static_prio = NICE_TO_PRIO(nice); 3112 set_load_weight(p); 3113 old_prio = p->prio; 3114 p->prio = effective_prio(p); 3115 delta = p->prio - old_prio; 3116 3117 if (on_rq) { 3118 enqueue_task(rq, p, 0); 3119 /* 3120 * If the task increased its priority or is running and 3121 * lowered its priority, then reschedule its CPU: 3122 */ 3123 if (delta < 0 || (delta > 0 && task_running(rq, p))) 3124 resched_task(rq->curr); 3125 } 3126out_unlock: 3127 task_rq_unlock(rq, p, &flags); 3128} 3129EXPORT_SYMBOL(set_user_nice); 3130 3131/* 3132 * can_nice - check if a task can reduce its nice value 3133 * @p: task 3134 * @nice: nice value 3135 */ 3136int can_nice(const struct task_struct *p, const int nice) 3137{ 3138 /* convert nice value [19,-20] to rlimit style value [1,40] */ 3139 int nice_rlim = 20 - nice; 3140 3141 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) || 3142 capable(CAP_SYS_NICE)); 3143} 3144 3145#ifdef __ARCH_WANT_SYS_NICE 3146 3147/* 3148 * sys_nice - change the priority of the current process. 3149 * @increment: priority increment 3150 * 3151 * sys_setpriority is a more generic, but much slower function that 3152 * does similar things. 3153 */ 3154SYSCALL_DEFINE1(nice, int, increment) 3155{ 3156 long nice, retval; 3157 3158 /* 3159 * Setpriority might change our priority at the same moment. 3160 * We don't have to worry. Conceptually one call occurs first 3161 * and we have a single winner. 3162 */ 3163 if (increment < -40) 3164 increment = -40; 3165 if (increment > 40) 3166 increment = 40; 3167 3168 nice = TASK_NICE(current) + increment; 3169 if (nice < -20) 3170 nice = -20; 3171 if (nice > 19) 3172 nice = 19; 3173 3174 if (increment < 0 && !can_nice(current, nice)) 3175 return -EPERM; 3176 3177 retval = security_task_setnice(current, nice); 3178 if (retval) 3179 return retval; 3180 3181 set_user_nice(current, nice); 3182 return 0; 3183} 3184 3185#endif 3186 3187/** 3188 * task_prio - return the priority value of a given task. 3189 * @p: the task in question. 3190 * 3191 * Return: The priority value as seen by users in /proc. 3192 * RT tasks are offset by -200. Normal tasks are centered 3193 * around 0, value goes from -16 to +15. 3194 */ 3195int task_prio(const struct task_struct *p) 3196{ 3197 return p->prio - MAX_RT_PRIO; 3198} 3199 3200/** 3201 * task_nice - return the nice value of a given task. 3202 * @p: the task in question. 3203 * 3204 * Return: The nice value [ -20 ... 0 ... 19 ]. 3205 */ 3206int task_nice(const struct task_struct *p) 3207{ 3208 return TASK_NICE(p); 3209} 3210EXPORT_SYMBOL(task_nice); 3211 3212/** 3213 * idle_cpu - is a given cpu idle currently? 3214 * @cpu: the processor in question. 3215 * 3216 * Return: 1 if the CPU is currently idle. 0 otherwise. 3217 */ 3218int idle_cpu(int cpu) 3219{ 3220 struct rq *rq = cpu_rq(cpu); 3221 3222 if (rq->curr != rq->idle) 3223 return 0; 3224 3225 if (rq->nr_running) 3226 return 0; 3227 3228#ifdef CONFIG_SMP 3229 if (!llist_empty(&rq->wake_list)) 3230 return 0; 3231#endif 3232 3233 return 1; 3234} 3235 3236/** 3237 * idle_task - return the idle task for a given cpu. 3238 * @cpu: the processor in question. 3239 * 3240 * Return: The idle task for the cpu @cpu. 3241 */ 3242struct task_struct *idle_task(int cpu) 3243{ 3244 return cpu_rq(cpu)->idle; 3245} 3246 3247/** 3248 * find_process_by_pid - find a process with a matching PID value. 3249 * @pid: the pid in question. 3250 * 3251 * The task of @pid, if found. %NULL otherwise. 3252 */ 3253static struct task_struct *find_process_by_pid(pid_t pid) 3254{ 3255 return pid ? find_task_by_vpid(pid) : current; 3256} 3257 3258/* Actually do priority change: must hold rq lock. */ 3259static void 3260__setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio) 3261{ 3262 p->policy = policy; 3263 p->rt_priority = prio; 3264 p->normal_prio = normal_prio(p); 3265 /* we are holding p->pi_lock already */ 3266 p->prio = rt_mutex_getprio(p); 3267 if (rt_prio(p->prio)) 3268 p->sched_class = &rt_sched_class; 3269 else 3270 p->sched_class = &fair_sched_class; 3271 set_load_weight(p); 3272} 3273 3274/* 3275 * check the target process has a UID that matches the current process's 3276 */ 3277static bool check_same_owner(struct task_struct *p) 3278{ 3279 const struct cred *cred = current_cred(), *pcred; 3280 bool match; 3281 3282 rcu_read_lock(); 3283 pcred = __task_cred(p); 3284 match = (uid_eq(cred->euid, pcred->euid) || 3285 uid_eq(cred->euid, pcred->uid)); 3286 rcu_read_unlock(); 3287 return match; 3288} 3289 3290static int __sched_setscheduler(struct task_struct *p, int policy, 3291 const struct sched_param *param, bool user) 3292{ 3293 int retval, oldprio, oldpolicy = -1, on_rq, running; 3294 unsigned long flags; 3295 const struct sched_class *prev_class; 3296 struct rq *rq; 3297 int reset_on_fork; 3298 3299 /* may grab non-irq protected spin_locks */ 3300 BUG_ON(in_interrupt()); 3301recheck: 3302 /* double check policy once rq lock held */ 3303 if (policy < 0) { 3304 reset_on_fork = p->sched_reset_on_fork; 3305 policy = oldpolicy = p->policy; 3306 } else { 3307 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK); 3308 policy &= ~SCHED_RESET_ON_FORK; 3309 3310 if (policy != SCHED_FIFO && policy != SCHED_RR && 3311 policy != SCHED_NORMAL && policy != SCHED_BATCH && 3312 policy != SCHED_IDLE) 3313 return -EINVAL; 3314 } 3315 3316 /* 3317 * Valid priorities for SCHED_FIFO and SCHED_RR are 3318 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL, 3319 * SCHED_BATCH and SCHED_IDLE is 0. 3320 */ 3321 if (param->sched_priority < 0 || 3322 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) || 3323 (!p->mm && param->sched_priority > MAX_RT_PRIO-1)) 3324 return -EINVAL; 3325 if (rt_policy(policy) != (param->sched_priority != 0)) 3326 return -EINVAL; 3327 3328 /* 3329 * Allow unprivileged RT tasks to decrease priority: 3330 */ 3331 if (user && !capable(CAP_SYS_NICE)) { 3332 if (rt_policy(policy)) { 3333 unsigned long rlim_rtprio = 3334 task_rlimit(p, RLIMIT_RTPRIO); 3335 3336 /* can't set/change the rt policy */ 3337 if (policy != p->policy && !rlim_rtprio) 3338 return -EPERM; 3339 3340 /* can't increase priority */ 3341 if (param->sched_priority > p->rt_priority && 3342 param->sched_priority > rlim_rtprio) 3343 return -EPERM; 3344 } 3345 3346 /* 3347 * Treat SCHED_IDLE as nice 20. Only allow a switch to 3348 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it. 3349 */ 3350 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) { 3351 if (!can_nice(p, TASK_NICE(p))) 3352 return -EPERM; 3353 } 3354 3355 /* can't change other user's priorities */ 3356 if (!check_same_owner(p)) 3357 return -EPERM; 3358 3359 /* Normal users shall not reset the sched_reset_on_fork flag */ 3360 if (p->sched_reset_on_fork && !reset_on_fork) 3361 return -EPERM; 3362 } 3363 3364 if (user) { 3365 retval = security_task_setscheduler(p); 3366 if (retval) 3367 return retval; 3368 } 3369 3370 /* 3371 * make sure no PI-waiters arrive (or leave) while we are 3372 * changing the priority of the task: 3373 * 3374 * To be able to change p->policy safely, the appropriate 3375 * runqueue lock must be held. 3376 */ 3377 rq = task_rq_lock(p, &flags); 3378 3379 /* 3380 * Changing the policy of the stop threads its a very bad idea 3381 */ 3382 if (p == rq->stop) { 3383 task_rq_unlock(rq, p, &flags); 3384 return -EINVAL; 3385 } 3386 3387 /* 3388 * If not changing anything there's no need to proceed further: 3389 */ 3390 if (unlikely(policy == p->policy && (!rt_policy(policy) || 3391 param->sched_priority == p->rt_priority))) { 3392 task_rq_unlock(rq, p, &flags); 3393 return 0; 3394 } 3395 3396#ifdef CONFIG_RT_GROUP_SCHED 3397 if (user) { 3398 /* 3399 * Do not allow realtime tasks into groups that have no runtime 3400 * assigned. 3401 */ 3402 if (rt_bandwidth_enabled() && rt_policy(policy) && 3403 task_group(p)->rt_bandwidth.rt_runtime == 0 && 3404 !task_group_is_autogroup(task_group(p))) { 3405 task_rq_unlock(rq, p, &flags); 3406 return -EPERM; 3407 } 3408 } 3409#endif 3410 3411 /* recheck policy now with rq lock held */ 3412 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) { 3413 policy = oldpolicy = -1; 3414 task_rq_unlock(rq, p, &flags); 3415 goto recheck; 3416 } 3417 on_rq = p->on_rq; 3418 running = task_current(rq, p); 3419 if (on_rq) 3420 dequeue_task(rq, p, 0); 3421 if (running) 3422 p->sched_class->put_prev_task(rq, p); 3423 3424 p->sched_reset_on_fork = reset_on_fork; 3425 3426 oldprio = p->prio; 3427 prev_class = p->sched_class; 3428 __setscheduler(rq, p, policy, param->sched_priority); 3429 3430 if (running) 3431 p->sched_class->set_curr_task(rq); 3432 if (on_rq) 3433 enqueue_task(rq, p, 0); 3434 3435 check_class_changed(rq, p, prev_class, oldprio); 3436 task_rq_unlock(rq, p, &flags); 3437 3438 rt_mutex_adjust_pi(p); 3439 3440 return 0; 3441} 3442 3443/** 3444 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread. 3445 * @p: the task in question. 3446 * @policy: new policy. 3447 * @param: structure containing the new RT priority. 3448 * 3449 * Return: 0 on success. An error code otherwise. 3450 * 3451 * NOTE that the task may be already dead. 3452 */ 3453int sched_setscheduler(struct task_struct *p, int policy, 3454 const struct sched_param *param) 3455{ 3456 return __sched_setscheduler(p, policy, param, true); 3457} 3458EXPORT_SYMBOL_GPL(sched_setscheduler); 3459 3460/** 3461 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace. 3462 * @p: the task in question. 3463 * @policy: new policy. 3464 * @param: structure containing the new RT priority. 3465 * 3466 * Just like sched_setscheduler, only don't bother checking if the 3467 * current context has permission. For example, this is needed in 3468 * stop_machine(): we create temporary high priority worker threads, 3469 * but our caller might not have that capability. 3470 * 3471 * Return: 0 on success. An error code otherwise. 3472 */ 3473int sched_setscheduler_nocheck(struct task_struct *p, int policy, 3474 const struct sched_param *param) 3475{ 3476 return __sched_setscheduler(p, policy, param, false); 3477} 3478 3479static int 3480do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param) 3481{ 3482 struct sched_param lparam; 3483 struct task_struct *p; 3484 int retval; 3485 3486 if (!param || pid < 0) 3487 return -EINVAL; 3488 if (copy_from_user(&lparam, param, sizeof(struct sched_param))) 3489 return -EFAULT; 3490 3491 rcu_read_lock(); 3492 retval = -ESRCH; 3493 p = find_process_by_pid(pid); 3494 if (p != NULL) 3495 retval = sched_setscheduler(p, policy, &lparam); 3496 rcu_read_unlock(); 3497 3498 return retval; 3499} 3500 3501/** 3502 * sys_sched_setscheduler - set/change the scheduler policy and RT priority 3503 * @pid: the pid in question. 3504 * @policy: new policy. 3505 * @param: structure containing the new RT priority. 3506 * 3507 * Return: 0 on success. An error code otherwise. 3508 */ 3509SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy, 3510 struct sched_param __user *, param) 3511{ 3512 /* negative values for policy are not valid */ 3513 if (policy < 0) 3514 return -EINVAL; 3515 3516 return do_sched_setscheduler(pid, policy, param); 3517} 3518 3519/** 3520 * sys_sched_setparam - set/change the RT priority of a thread 3521 * @pid: the pid in question. 3522 * @param: structure containing the new RT priority. 3523 * 3524 * Return: 0 on success. An error code otherwise. 3525 */ 3526SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param) 3527{ 3528 return do_sched_setscheduler(pid, -1, param); 3529} 3530 3531/** 3532 * sys_sched_getscheduler - get the policy (scheduling class) of a thread 3533 * @pid: the pid in question. 3534 * 3535 * Return: On success, the policy of the thread. Otherwise, a negative error 3536 * code. 3537 */ 3538SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid) 3539{ 3540 struct task_struct *p; 3541 int retval; 3542 3543 if (pid < 0) 3544 return -EINVAL; 3545 3546 retval = -ESRCH; 3547 rcu_read_lock(); 3548 p = find_process_by_pid(pid); 3549 if (p) { 3550 retval = security_task_getscheduler(p); 3551 if (!retval) 3552 retval = p->policy 3553 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0); 3554 } 3555 rcu_read_unlock(); 3556 return retval; 3557} 3558 3559/** 3560 * sys_sched_getparam - get the RT priority of a thread 3561 * @pid: the pid in question. 3562 * @param: structure containing the RT priority. 3563 * 3564 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error 3565 * code. 3566 */ 3567SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param) 3568{ 3569 struct sched_param lp; 3570 struct task_struct *p; 3571 int retval; 3572 3573 if (!param || pid < 0) 3574 return -EINVAL; 3575 3576 rcu_read_lock(); 3577 p = find_process_by_pid(pid); 3578 retval = -ESRCH; 3579 if (!p) 3580 goto out_unlock; 3581 3582 retval = security_task_getscheduler(p); 3583 if (retval) 3584 goto out_unlock; 3585 3586 lp.sched_priority = p->rt_priority; 3587 rcu_read_unlock(); 3588 3589 /* 3590 * This one might sleep, we cannot do it with a spinlock held ... 3591 */ 3592 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0; 3593 3594 return retval; 3595 3596out_unlock: 3597 rcu_read_unlock(); 3598 return retval; 3599} 3600 3601long sched_setaffinity(pid_t pid, const struct cpumask *in_mask) 3602{ 3603 cpumask_var_t cpus_allowed, new_mask; 3604 struct task_struct *p; 3605 int retval; 3606 3607 get_online_cpus(); 3608 rcu_read_lock(); 3609 3610 p = find_process_by_pid(pid); 3611 if (!p) { 3612 rcu_read_unlock(); 3613 put_online_cpus(); 3614 return -ESRCH; 3615 } 3616 3617 /* Prevent p going away */ 3618 get_task_struct(p); 3619 rcu_read_unlock(); 3620 3621 if (p->flags & PF_NO_SETAFFINITY) { 3622 retval = -EINVAL; 3623 goto out_put_task; 3624 } 3625 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) { 3626 retval = -ENOMEM; 3627 goto out_put_task; 3628 } 3629 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) { 3630 retval = -ENOMEM; 3631 goto out_free_cpus_allowed; 3632 } 3633 retval = -EPERM; 3634 if (!check_same_owner(p)) { 3635 rcu_read_lock(); 3636 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) { 3637 rcu_read_unlock(); 3638 goto out_unlock; 3639 } 3640 rcu_read_unlock(); 3641 } 3642 3643 retval = security_task_setscheduler(p); 3644 if (retval) 3645 goto out_unlock; 3646 3647 cpuset_cpus_allowed(p, cpus_allowed); 3648 cpumask_and(new_mask, in_mask, cpus_allowed); 3649again: 3650 retval = set_cpus_allowed_ptr(p, new_mask); 3651 3652 if (!retval) { 3653 cpuset_cpus_allowed(p, cpus_allowed); 3654 if (!cpumask_subset(new_mask, cpus_allowed)) { 3655 /* 3656 * We must have raced with a concurrent cpuset 3657 * update. Just reset the cpus_allowed to the 3658 * cpuset's cpus_allowed 3659 */ 3660 cpumask_copy(new_mask, cpus_allowed); 3661 goto again; 3662 } 3663 } 3664out_unlock: 3665 free_cpumask_var(new_mask); 3666out_free_cpus_allowed: 3667 free_cpumask_var(cpus_allowed); 3668out_put_task: 3669 put_task_struct(p); 3670 put_online_cpus(); 3671 return retval; 3672} 3673 3674static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len, 3675 struct cpumask *new_mask) 3676{ 3677 if (len < cpumask_size()) 3678 cpumask_clear(new_mask); 3679 else if (len > cpumask_size()) 3680 len = cpumask_size(); 3681 3682 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0; 3683} 3684 3685/** 3686 * sys_sched_setaffinity - set the cpu affinity of a process 3687 * @pid: pid of the process 3688 * @len: length in bytes of the bitmask pointed to by user_mask_ptr 3689 * @user_mask_ptr: user-space pointer to the new cpu mask 3690 * 3691 * Return: 0 on success. An error code otherwise. 3692 */ 3693SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len, 3694 unsigned long __user *, user_mask_ptr) 3695{ 3696 cpumask_var_t new_mask; 3697 int retval; 3698 3699 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) 3700 return -ENOMEM; 3701 3702 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask); 3703 if (retval == 0) 3704 retval = sched_setaffinity(pid, new_mask); 3705 free_cpumask_var(new_mask); 3706 return retval; 3707} 3708 3709long sched_getaffinity(pid_t pid, struct cpumask *mask) 3710{ 3711 struct task_struct *p; 3712 unsigned long flags; 3713 int retval; 3714 3715 get_online_cpus(); 3716 rcu_read_lock(); 3717 3718 retval = -ESRCH; 3719 p = find_process_by_pid(pid); 3720 if (!p) 3721 goto out_unlock; 3722 3723 retval = security_task_getscheduler(p); 3724 if (retval) 3725 goto out_unlock; 3726 3727 raw_spin_lock_irqsave(&p->pi_lock, flags); 3728 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask); 3729 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 3730 3731out_unlock: 3732 rcu_read_unlock(); 3733 put_online_cpus(); 3734 3735 return retval; 3736} 3737 3738/** 3739 * sys_sched_getaffinity - get the cpu affinity of a process 3740 * @pid: pid of the process 3741 * @len: length in bytes of the bitmask pointed to by user_mask_ptr 3742 * @user_mask_ptr: user-space pointer to hold the current cpu mask 3743 * 3744 * Return: 0 on success. An error code otherwise. 3745 */ 3746SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len, 3747 unsigned long __user *, user_mask_ptr) 3748{ 3749 int ret; 3750 cpumask_var_t mask; 3751 3752 if ((len * BITS_PER_BYTE) < nr_cpu_ids) 3753 return -EINVAL; 3754 if (len & (sizeof(unsigned long)-1)) 3755 return -EINVAL; 3756 3757 if (!alloc_cpumask_var(&mask, GFP_KERNEL)) 3758 return -ENOMEM; 3759 3760 ret = sched_getaffinity(pid, mask); 3761 if (ret == 0) { 3762 size_t retlen = min_t(size_t, len, cpumask_size()); 3763 3764 if (copy_to_user(user_mask_ptr, mask, retlen)) 3765 ret = -EFAULT; 3766 else 3767 ret = retlen; 3768 } 3769 free_cpumask_var(mask); 3770 3771 return ret; 3772} 3773 3774/** 3775 * sys_sched_yield - yield the current processor to other threads. 3776 * 3777 * This function yields the current CPU to other tasks. If there are no 3778 * other threads running on this CPU then this function will return. 3779 * 3780 * Return: 0. 3781 */ 3782SYSCALL_DEFINE0(sched_yield) 3783{ 3784 struct rq *rq = this_rq_lock(); 3785 3786 schedstat_inc(rq, yld_count); 3787 current->sched_class->yield_task(rq); 3788 3789 /* 3790 * Since we are going to call schedule() anyway, there's 3791 * no need to preempt or enable interrupts: 3792 */ 3793 __release(rq->lock); 3794 spin_release(&rq->lock.dep_map, 1, _THIS_IP_); 3795 do_raw_spin_unlock(&rq->lock); 3796 sched_preempt_enable_no_resched(); 3797 3798 schedule(); 3799 3800 return 0; 3801} 3802 3803static void __cond_resched(void) 3804{ 3805 __preempt_count_add(PREEMPT_ACTIVE); 3806 __schedule(); 3807 __preempt_count_sub(PREEMPT_ACTIVE); 3808} 3809 3810int __sched _cond_resched(void) 3811{ 3812 if (should_resched()) { 3813 __cond_resched(); 3814 return 1; 3815 } 3816 return 0; 3817} 3818EXPORT_SYMBOL(_cond_resched); 3819 3820/* 3821 * __cond_resched_lock() - if a reschedule is pending, drop the given lock, 3822 * call schedule, and on return reacquire the lock. 3823 * 3824 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level 3825 * operations here to prevent schedule() from being called twice (once via 3826 * spin_unlock(), once by hand). 3827 */ 3828int __cond_resched_lock(spinlock_t *lock) 3829{ 3830 int resched = should_resched(); 3831 int ret = 0; 3832 3833 lockdep_assert_held(lock); 3834 3835 if (spin_needbreak(lock) || resched) { 3836 spin_unlock(lock); 3837 if (resched) 3838 __cond_resched(); 3839 else 3840 cpu_relax(); 3841 ret = 1; 3842 spin_lock(lock); 3843 } 3844 return ret; 3845} 3846EXPORT_SYMBOL(__cond_resched_lock); 3847 3848int __sched __cond_resched_softirq(void) 3849{ 3850 BUG_ON(!in_softirq()); 3851 3852 if (should_resched()) { 3853 local_bh_enable(); 3854 __cond_resched(); 3855 local_bh_disable(); 3856 return 1; 3857 } 3858 return 0; 3859} 3860EXPORT_SYMBOL(__cond_resched_softirq); 3861 3862/** 3863 * yield - yield the current processor to other threads. 3864 * 3865 * Do not ever use this function, there's a 99% chance you're doing it wrong. 3866 * 3867 * The scheduler is at all times free to pick the calling task as the most 3868 * eligible task to run, if removing the yield() call from your code breaks 3869 * it, its already broken. 3870 * 3871 * Typical broken usage is: 3872 * 3873 * while (!event) 3874 * yield(); 3875 * 3876 * where one assumes that yield() will let 'the other' process run that will 3877 * make event true. If the current task is a SCHED_FIFO task that will never 3878 * happen. Never use yield() as a progress guarantee!! 3879 * 3880 * If you want to use yield() to wait for something, use wait_event(). 3881 * If you want to use yield() to be 'nice' for others, use cond_resched(). 3882 * If you still want to use yield(), do not! 3883 */ 3884void __sched yield(void) 3885{ 3886 set_current_state(TASK_RUNNING); 3887 sys_sched_yield(); 3888} 3889EXPORT_SYMBOL(yield); 3890 3891/** 3892 * yield_to - yield the current processor to another thread in 3893 * your thread group, or accelerate that thread toward the 3894 * processor it's on. 3895 * @p: target task 3896 * @preempt: whether task preemption is allowed or not 3897 * 3898 * It's the caller's job to ensure that the target task struct 3899 * can't go away on us before we can do any checks. 3900 * 3901 * Return: 3902 * true (>0) if we indeed boosted the target task. 3903 * false (0) if we failed to boost the target. 3904 * -ESRCH if there's no task to yield to. 3905 */ 3906bool __sched yield_to(struct task_struct *p, bool preempt) 3907{ 3908 struct task_struct *curr = current; 3909 struct rq *rq, *p_rq; 3910 unsigned long flags; 3911 int yielded = 0; 3912 3913 local_irq_save(flags); 3914 rq = this_rq(); 3915 3916again: 3917 p_rq = task_rq(p); 3918 /* 3919 * If we're the only runnable task on the rq and target rq also 3920 * has only one task, there's absolutely no point in yielding. 3921 */ 3922 if (rq->nr_running == 1 && p_rq->nr_running == 1) { 3923 yielded = -ESRCH; 3924 goto out_irq; 3925 } 3926 3927 double_rq_lock(rq, p_rq); 3928 while (task_rq(p) != p_rq) { 3929 double_rq_unlock(rq, p_rq); 3930 goto again; 3931 } 3932 3933 if (!curr->sched_class->yield_to_task) 3934 goto out_unlock; 3935 3936 if (curr->sched_class != p->sched_class) 3937 goto out_unlock; 3938 3939 if (task_running(p_rq, p) || p->state) 3940 goto out_unlock; 3941 3942 yielded = curr->sched_class->yield_to_task(rq, p, preempt); 3943 if (yielded) { 3944 schedstat_inc(rq, yld_count); 3945 /* 3946 * Make p's CPU reschedule; pick_next_entity takes care of 3947 * fairness. 3948 */ 3949 if (preempt && rq != p_rq) 3950 resched_task(p_rq->curr); 3951 } 3952 3953out_unlock: 3954 double_rq_unlock(rq, p_rq); 3955out_irq: 3956 local_irq_restore(flags); 3957 3958 if (yielded > 0) 3959 schedule(); 3960 3961 return yielded; 3962} 3963EXPORT_SYMBOL_GPL(yield_to); 3964 3965/* 3966 * This task is about to go to sleep on IO. Increment rq->nr_iowait so 3967 * that process accounting knows that this is a task in IO wait state. 3968 */ 3969void __sched io_schedule(void) 3970{ 3971 struct rq *rq = raw_rq(); 3972 3973 delayacct_blkio_start(); 3974 atomic_inc(&rq->nr_iowait); 3975 blk_flush_plug(current); 3976 current->in_iowait = 1; 3977 schedule(); 3978 current->in_iowait = 0; 3979 atomic_dec(&rq->nr_iowait); 3980 delayacct_blkio_end(); 3981} 3982EXPORT_SYMBOL(io_schedule); 3983 3984long __sched io_schedule_timeout(long timeout) 3985{ 3986 struct rq *rq = raw_rq(); 3987 long ret; 3988 3989 delayacct_blkio_start(); 3990 atomic_inc(&rq->nr_iowait); 3991 blk_flush_plug(current); 3992 current->in_iowait = 1; 3993 ret = schedule_timeout(timeout); 3994 current->in_iowait = 0; 3995 atomic_dec(&rq->nr_iowait); 3996 delayacct_blkio_end(); 3997 return ret; 3998} 3999 4000/** 4001 * sys_sched_get_priority_max - return maximum RT priority. 4002 * @policy: scheduling class. 4003 * 4004 * Return: On success, this syscall returns the maximum 4005 * rt_priority that can be used by a given scheduling class. 4006 * On failure, a negative error code is returned. 4007 */ 4008SYSCALL_DEFINE1(sched_get_priority_max, int, policy) 4009{ 4010 int ret = -EINVAL; 4011 4012 switch (policy) { 4013 case SCHED_FIFO: 4014 case SCHED_RR: 4015 ret = MAX_USER_RT_PRIO-1; 4016 break; 4017 case SCHED_NORMAL: 4018 case SCHED_BATCH: 4019 case SCHED_IDLE: 4020 ret = 0; 4021 break; 4022 } 4023 return ret; 4024} 4025 4026/** 4027 * sys_sched_get_priority_min - return minimum RT priority. 4028 * @policy: scheduling class. 4029 * 4030 * Return: On success, this syscall returns the minimum 4031 * rt_priority that can be used by a given scheduling class. 4032 * On failure, a negative error code is returned. 4033 */ 4034SYSCALL_DEFINE1(sched_get_priority_min, int, policy) 4035{ 4036 int ret = -EINVAL; 4037 4038 switch (policy) { 4039 case SCHED_FIFO: 4040 case SCHED_RR: 4041 ret = 1; 4042 break; 4043 case SCHED_NORMAL: 4044 case SCHED_BATCH: 4045 case SCHED_IDLE: 4046 ret = 0; 4047 } 4048 return ret; 4049} 4050 4051/** 4052 * sys_sched_rr_get_interval - return the default timeslice of a process. 4053 * @pid: pid of the process. 4054 * @interval: userspace pointer to the timeslice value. 4055 * 4056 * this syscall writes the default timeslice value of a given process 4057 * into the user-space timespec buffer. A value of '0' means infinity. 4058 * 4059 * Return: On success, 0 and the timeslice is in @interval. Otherwise, 4060 * an error code. 4061 */ 4062SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid, 4063 struct timespec __user *, interval) 4064{ 4065 struct task_struct *p; 4066 unsigned int time_slice; 4067 unsigned long flags; 4068 struct rq *rq; 4069 int retval; 4070 struct timespec t; 4071 4072 if (pid < 0) 4073 return -EINVAL; 4074 4075 retval = -ESRCH; 4076 rcu_read_lock(); 4077 p = find_process_by_pid(pid); 4078 if (!p) 4079 goto out_unlock; 4080 4081 retval = security_task_getscheduler(p); 4082 if (retval) 4083 goto out_unlock; 4084 4085 rq = task_rq_lock(p, &flags); 4086 time_slice = p->sched_class->get_rr_interval(rq, p); 4087 task_rq_unlock(rq, p, &flags); 4088 4089 rcu_read_unlock(); 4090 jiffies_to_timespec(time_slice, &t); 4091 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0; 4092 return retval; 4093 4094out_unlock: 4095 rcu_read_unlock(); 4096 return retval; 4097} 4098 4099static const char stat_nam[] = TASK_STATE_TO_CHAR_STR; 4100 4101void sched_show_task(struct task_struct *p) 4102{ 4103 unsigned long free = 0; 4104 int ppid; 4105 unsigned state; 4106 4107 state = p->state ? __ffs(p->state) + 1 : 0; 4108 printk(KERN_INFO "%-15.15s %c", p->comm, 4109 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?'); 4110#if BITS_PER_LONG == 32 4111 if (state == TASK_RUNNING) 4112 printk(KERN_CONT " running "); 4113 else 4114 printk(KERN_CONT " %08lx ", thread_saved_pc(p)); 4115#else 4116 if (state == TASK_RUNNING) 4117 printk(KERN_CONT " running task "); 4118 else 4119 printk(KERN_CONT " %016lx ", thread_saved_pc(p)); 4120#endif 4121#ifdef CONFIG_DEBUG_STACK_USAGE 4122 free = stack_not_used(p); 4123#endif 4124 rcu_read_lock(); 4125 ppid = task_pid_nr(rcu_dereference(p->real_parent)); 4126 rcu_read_unlock(); 4127 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free, 4128 task_pid_nr(p), ppid, 4129 (unsigned long)task_thread_info(p)->flags); 4130 4131 print_worker_info(KERN_INFO, p); 4132 show_stack(p, NULL); 4133} 4134 4135void show_state_filter(unsigned long state_filter) 4136{ 4137 struct task_struct *g, *p; 4138 4139#if BITS_PER_LONG == 32 4140 printk(KERN_INFO 4141 " task PC stack pid father\n"); 4142#else 4143 printk(KERN_INFO 4144 " task PC stack pid father\n"); 4145#endif 4146 rcu_read_lock(); 4147 do_each_thread(g, p) { 4148 /* 4149 * reset the NMI-timeout, listing all files on a slow 4150 * console might take a lot of time: 4151 */ 4152 touch_nmi_watchdog(); 4153 if (!state_filter || (p->state & state_filter)) 4154 sched_show_task(p); 4155 } while_each_thread(g, p); 4156 4157 touch_all_softlockup_watchdogs(); 4158 4159#ifdef CONFIG_SCHED_DEBUG 4160 sysrq_sched_debug_show(); 4161#endif 4162 rcu_read_unlock(); 4163 /* 4164 * Only show locks if all tasks are dumped: 4165 */ 4166 if (!state_filter) 4167 debug_show_all_locks(); 4168} 4169 4170void init_idle_bootup_task(struct task_struct *idle) 4171{ 4172 idle->sched_class = &idle_sched_class; 4173} 4174 4175/** 4176 * init_idle - set up an idle thread for a given CPU 4177 * @idle: task in question 4178 * @cpu: cpu the idle task belongs to 4179 * 4180 * NOTE: this function does not set the idle thread's NEED_RESCHED 4181 * flag, to make booting more robust. 4182 */ 4183void init_idle(struct task_struct *idle, int cpu) 4184{ 4185 struct rq *rq = cpu_rq(cpu); 4186 unsigned long flags; 4187 4188 raw_spin_lock_irqsave(&rq->lock, flags); 4189 4190 __sched_fork(idle); 4191 idle->state = TASK_RUNNING; 4192 idle->se.exec_start = sched_clock(); 4193 4194 do_set_cpus_allowed(idle, cpumask_of(cpu)); 4195 /* 4196 * We're having a chicken and egg problem, even though we are 4197 * holding rq->lock, the cpu isn't yet set to this cpu so the 4198 * lockdep check in task_group() will fail. 4199 * 4200 * Similar case to sched_fork(). / Alternatively we could 4201 * use task_rq_lock() here and obtain the other rq->lock. 4202 * 4203 * Silence PROVE_RCU 4204 */ 4205 rcu_read_lock(); 4206 __set_task_cpu(idle, cpu); 4207 rcu_read_unlock(); 4208 4209 rq->curr = rq->idle = idle; 4210#if defined(CONFIG_SMP) 4211 idle->on_cpu = 1; 4212#endif 4213 raw_spin_unlock_irqrestore(&rq->lock, flags); 4214 4215 /* Set the preempt count _outside_ the spinlocks! */ 4216 init_idle_preempt_count(idle, cpu); 4217 4218 /* 4219 * The idle tasks have their own, simple scheduling class: 4220 */ 4221 idle->sched_class = &idle_sched_class; 4222 ftrace_graph_init_idle_task(idle, cpu); 4223 vtime_init_idle(idle, cpu); 4224#if defined(CONFIG_SMP) 4225 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu); 4226#endif 4227} 4228 4229#ifdef CONFIG_SMP 4230void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask) 4231{ 4232 if (p->sched_class && p->sched_class->set_cpus_allowed) 4233 p->sched_class->set_cpus_allowed(p, new_mask); 4234 4235 cpumask_copy(&p->cpus_allowed, new_mask); 4236 p->nr_cpus_allowed = cpumask_weight(new_mask); 4237} 4238 4239/* 4240 * This is how migration works: 4241 * 4242 * 1) we invoke migration_cpu_stop() on the target CPU using 4243 * stop_one_cpu(). 4244 * 2) stopper starts to run (implicitly forcing the migrated thread 4245 * off the CPU) 4246 * 3) it checks whether the migrated task is still in the wrong runqueue. 4247 * 4) if it's in the wrong runqueue then the migration thread removes 4248 * it and puts it into the right queue. 4249 * 5) stopper completes and stop_one_cpu() returns and the migration 4250 * is done. 4251 */ 4252 4253/* 4254 * Change a given task's CPU affinity. Migrate the thread to a 4255 * proper CPU and schedule it away if the CPU it's executing on 4256 * is removed from the allowed bitmask. 4257 * 4258 * NOTE: the caller must have a valid reference to the task, the 4259 * task must not exit() & deallocate itself prematurely. The 4260 * call is not atomic; no spinlocks may be held. 4261 */ 4262int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask) 4263{ 4264 unsigned long flags; 4265 struct rq *rq; 4266 unsigned int dest_cpu; 4267 int ret = 0; 4268 4269 rq = task_rq_lock(p, &flags); 4270 4271 if (cpumask_equal(&p->cpus_allowed, new_mask)) 4272 goto out; 4273 4274 if (!cpumask_intersects(new_mask, cpu_active_mask)) { 4275 ret = -EINVAL; 4276 goto out; 4277 } 4278 4279 do_set_cpus_allowed(p, new_mask); 4280 4281 /* Can the task run on the task's current CPU? If so, we're done */ 4282 if (cpumask_test_cpu(task_cpu(p), new_mask)) 4283 goto out; 4284 4285 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask); 4286 if (p->on_rq) { 4287 struct migration_arg arg = { p, dest_cpu }; 4288 /* Need help from migration thread: drop lock and wait. */ 4289 task_rq_unlock(rq, p, &flags); 4290 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg); 4291 tlb_migrate_finish(p->mm); 4292 return 0; 4293 } 4294out: 4295 task_rq_unlock(rq, p, &flags); 4296 4297 return ret; 4298} 4299EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr); 4300 4301/* 4302 * Move (not current) task off this cpu, onto dest cpu. We're doing 4303 * this because either it can't run here any more (set_cpus_allowed() 4304 * away from this CPU, or CPU going down), or because we're 4305 * attempting to rebalance this task on exec (sched_exec). 4306 * 4307 * So we race with normal scheduler movements, but that's OK, as long 4308 * as the task is no longer on this CPU. 4309 * 4310 * Returns non-zero if task was successfully migrated. 4311 */ 4312static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu) 4313{ 4314 struct rq *rq_dest, *rq_src; 4315 int ret = 0; 4316 4317 if (unlikely(!cpu_active(dest_cpu))) 4318 return ret; 4319 4320 rq_src = cpu_rq(src_cpu); 4321 rq_dest = cpu_rq(dest_cpu); 4322 4323 raw_spin_lock(&p->pi_lock); 4324 double_rq_lock(rq_src, rq_dest); 4325 /* Already moved. */ 4326 if (task_cpu(p) != src_cpu) 4327 goto done; 4328 /* Affinity changed (again). */ 4329 if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p))) 4330 goto fail; 4331 4332 /* 4333 * If we're not on a rq, the next wake-up will ensure we're 4334 * placed properly. 4335 */ 4336 if (p->on_rq) { 4337 dequeue_task(rq_src, p, 0); 4338 set_task_cpu(p, dest_cpu); 4339 enqueue_task(rq_dest, p, 0); 4340 check_preempt_curr(rq_dest, p, 0); 4341 } 4342done: 4343 ret = 1; 4344fail: 4345 double_rq_unlock(rq_src, rq_dest); 4346 raw_spin_unlock(&p->pi_lock); 4347 return ret; 4348} 4349 4350/* 4351 * migration_cpu_stop - this will be executed by a highprio stopper thread 4352 * and performs thread migration by bumping thread off CPU then 4353 * 'pushing' onto another runqueue. 4354 */ 4355static int migration_cpu_stop(void *data) 4356{ 4357 struct migration_arg *arg = data; 4358 4359 /* 4360 * The original target cpu might have gone down and we might 4361 * be on another cpu but it doesn't matter. 4362 */ 4363 local_irq_disable(); 4364 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu); 4365 local_irq_enable(); 4366 return 0; 4367} 4368 4369#ifdef CONFIG_HOTPLUG_CPU 4370 4371/* 4372 * Ensures that the idle task is using init_mm right before its cpu goes 4373 * offline. 4374 */ 4375void idle_task_exit(void) 4376{ 4377 struct mm_struct *mm = current->active_mm; 4378 4379 BUG_ON(cpu_online(smp_processor_id())); 4380 4381 if (mm != &init_mm) 4382 switch_mm(mm, &init_mm, current); 4383 mmdrop(mm); 4384} 4385 4386/* 4387 * Since this CPU is going 'away' for a while, fold any nr_active delta 4388 * we might have. Assumes we're called after migrate_tasks() so that the 4389 * nr_active count is stable. 4390 * 4391 * Also see the comment "Global load-average calculations". 4392 */ 4393static void calc_load_migrate(struct rq *rq) 4394{ 4395 long delta = calc_load_fold_active(rq); 4396 if (delta) 4397 atomic_long_add(delta, &calc_load_tasks); 4398} 4399 4400/* 4401 * Migrate all tasks from the rq, sleeping tasks will be migrated by 4402 * try_to_wake_up()->select_task_rq(). 4403 * 4404 * Called with rq->lock held even though we'er in stop_machine() and 4405 * there's no concurrency possible, we hold the required locks anyway 4406 * because of lock validation efforts. 4407 */ 4408static void migrate_tasks(unsigned int dead_cpu) 4409{ 4410 struct rq *rq = cpu_rq(dead_cpu); 4411 struct task_struct *next, *stop = rq->stop; 4412 int dest_cpu; 4413 4414 /* 4415 * Fudge the rq selection such that the below task selection loop 4416 * doesn't get stuck on the currently eligible stop task. 4417 * 4418 * We're currently inside stop_machine() and the rq is either stuck 4419 * in the stop_machine_cpu_stop() loop, or we're executing this code, 4420 * either way we should never end up calling schedule() until we're 4421 * done here. 4422 */ 4423 rq->stop = NULL; 4424 4425 /* 4426 * put_prev_task() and pick_next_task() sched 4427 * class method both need to have an up-to-date 4428 * value of rq->clock[_task] 4429 */ 4430 update_rq_clock(rq); 4431 4432 for ( ; ; ) { 4433 /* 4434 * There's this thread running, bail when that's the only 4435 * remaining thread. 4436 */ 4437 if (rq->nr_running == 1) 4438 break; 4439 4440 next = pick_next_task(rq); 4441 BUG_ON(!next); 4442 next->sched_class->put_prev_task(rq, next); 4443 4444 /* Find suitable destination for @next, with force if needed. */ 4445 dest_cpu = select_fallback_rq(dead_cpu, next); 4446 raw_spin_unlock(&rq->lock); 4447 4448 __migrate_task(next, dead_cpu, dest_cpu); 4449 4450 raw_spin_lock(&rq->lock); 4451 } 4452 4453 rq->stop = stop; 4454} 4455 4456#endif /* CONFIG_HOTPLUG_CPU */ 4457 4458#if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL) 4459 4460static struct ctl_table sd_ctl_dir[] = { 4461 { 4462 .procname = "sched_domain", 4463 .mode = 0555, 4464 }, 4465 {} 4466}; 4467 4468static struct ctl_table sd_ctl_root[] = { 4469 { 4470 .procname = "kernel", 4471 .mode = 0555, 4472 .child = sd_ctl_dir, 4473 }, 4474 {} 4475}; 4476 4477static struct ctl_table *sd_alloc_ctl_entry(int n) 4478{ 4479 struct ctl_table *entry = 4480 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL); 4481 4482 return entry; 4483} 4484 4485static void sd_free_ctl_entry(struct ctl_table **tablep) 4486{ 4487 struct ctl_table *entry; 4488 4489 /* 4490 * In the intermediate directories, both the child directory and 4491 * procname are dynamically allocated and could fail but the mode 4492 * will always be set. In the lowest directory the names are 4493 * static strings and all have proc handlers. 4494 */ 4495 for (entry = *tablep; entry->mode; entry++) { 4496 if (entry->child) 4497 sd_free_ctl_entry(&entry->child); 4498 if (entry->proc_handler == NULL) 4499 kfree(entry->procname); 4500 } 4501 4502 kfree(*tablep); 4503 *tablep = NULL; 4504} 4505 4506static int min_load_idx = 0; 4507static int max_load_idx = CPU_LOAD_IDX_MAX-1; 4508 4509static void 4510set_table_entry(struct ctl_table *entry, 4511 const char *procname, void *data, int maxlen, 4512 umode_t mode, proc_handler *proc_handler, 4513 bool load_idx) 4514{ 4515 entry->procname = procname; 4516 entry->data = data; 4517 entry->maxlen = maxlen; 4518 entry->mode = mode; 4519 entry->proc_handler = proc_handler; 4520 4521 if (load_idx) { 4522 entry->extra1 = &min_load_idx; 4523 entry->extra2 = &max_load_idx; 4524 } 4525} 4526 4527static struct ctl_table * 4528sd_alloc_ctl_domain_table(struct sched_domain *sd) 4529{ 4530 struct ctl_table *table = sd_alloc_ctl_entry(13); 4531 4532 if (table == NULL) 4533 return NULL; 4534 4535 set_table_entry(&table[0], "min_interval", &sd->min_interval, 4536 sizeof(long), 0644, proc_doulongvec_minmax, false); 4537 set_table_entry(&table[1], "max_interval", &sd->max_interval, 4538 sizeof(long), 0644, proc_doulongvec_minmax, false); 4539 set_table_entry(&table[2], "busy_idx", &sd->busy_idx, 4540 sizeof(int), 0644, proc_dointvec_minmax, true); 4541 set_table_entry(&table[3], "idle_idx", &sd->idle_idx, 4542 sizeof(int), 0644, proc_dointvec_minmax, true); 4543 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx, 4544 sizeof(int), 0644, proc_dointvec_minmax, true); 4545 set_table_entry(&table[5], "wake_idx", &sd->wake_idx, 4546 sizeof(int), 0644, proc_dointvec_minmax, true); 4547 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx, 4548 sizeof(int), 0644, proc_dointvec_minmax, true); 4549 set_table_entry(&table[7], "busy_factor", &sd->busy_factor, 4550 sizeof(int), 0644, proc_dointvec_minmax, false); 4551 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct, 4552 sizeof(int), 0644, proc_dointvec_minmax, false); 4553 set_table_entry(&table[9], "cache_nice_tries", 4554 &sd->cache_nice_tries, 4555 sizeof(int), 0644, proc_dointvec_minmax, false); 4556 set_table_entry(&table[10], "flags", &sd->flags, 4557 sizeof(int), 0644, proc_dointvec_minmax, false); 4558 set_table_entry(&table[11], "name", sd->name, 4559 CORENAME_MAX_SIZE, 0444, proc_dostring, false); 4560 /* &table[12] is terminator */ 4561 4562 return table; 4563} 4564 4565static struct ctl_table *sd_alloc_ctl_cpu_table(int cpu) 4566{ 4567 struct ctl_table *entry, *table; 4568 struct sched_domain *sd; 4569 int domain_num = 0, i; 4570 char buf[32]; 4571 4572 for_each_domain(cpu, sd) 4573 domain_num++; 4574 entry = table = sd_alloc_ctl_entry(domain_num + 1); 4575 if (table == NULL) 4576 return NULL; 4577 4578 i = 0; 4579 for_each_domain(cpu, sd) { 4580 snprintf(buf, 32, "domain%d", i); 4581 entry->procname = kstrdup(buf, GFP_KERNEL); 4582 entry->mode = 0555; 4583 entry->child = sd_alloc_ctl_domain_table(sd); 4584 entry++; 4585 i++; 4586 } 4587 return table; 4588} 4589 4590static struct ctl_table_header *sd_sysctl_header; 4591static void register_sched_domain_sysctl(void) 4592{ 4593 int i, cpu_num = num_possible_cpus(); 4594 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1); 4595 char buf[32]; 4596 4597 WARN_ON(sd_ctl_dir[0].child); 4598 sd_ctl_dir[0].child = entry; 4599 4600 if (entry == NULL) 4601 return; 4602 4603 for_each_possible_cpu(i) { 4604 snprintf(buf, 32, "cpu%d", i); 4605 entry->procname = kstrdup(buf, GFP_KERNEL); 4606 entry->mode = 0555; 4607 entry->child = sd_alloc_ctl_cpu_table(i); 4608 entry++; 4609 } 4610 4611 WARN_ON(sd_sysctl_header); 4612 sd_sysctl_header = register_sysctl_table(sd_ctl_root); 4613} 4614 4615/* may be called multiple times per register */ 4616static void unregister_sched_domain_sysctl(void) 4617{ 4618 if (sd_sysctl_header) 4619 unregister_sysctl_table(sd_sysctl_header); 4620 sd_sysctl_header = NULL; 4621 if (sd_ctl_dir[0].child) 4622 sd_free_ctl_entry(&sd_ctl_dir[0].child); 4623} 4624#else 4625static void register_sched_domain_sysctl(void) 4626{ 4627} 4628static void unregister_sched_domain_sysctl(void) 4629{ 4630} 4631#endif 4632 4633static void set_rq_online(struct rq *rq) 4634{ 4635 if (!rq->online) { 4636 const struct sched_class *class; 4637 4638 cpumask_set_cpu(rq->cpu, rq->rd->online); 4639 rq->online = 1; 4640 4641 for_each_class(class) { 4642 if (class->rq_online) 4643 class->rq_online(rq); 4644 } 4645 } 4646} 4647 4648static void set_rq_offline(struct rq *rq) 4649{ 4650 if (rq->online) { 4651 const struct sched_class *class; 4652 4653 for_each_class(class) { 4654 if (class->rq_offline) 4655 class->rq_offline(rq); 4656 } 4657 4658 cpumask_clear_cpu(rq->cpu, rq->rd->online); 4659 rq->online = 0; 4660 } 4661} 4662 4663/* 4664 * migration_call - callback that gets triggered when a CPU is added. 4665 * Here we can start up the necessary migration thread for the new CPU. 4666 */ 4667static int 4668migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu) 4669{ 4670 int cpu = (long)hcpu; 4671 unsigned long flags; 4672 struct rq *rq = cpu_rq(cpu); 4673 4674 switch (action & ~CPU_TASKS_FROZEN) { 4675 4676 case CPU_UP_PREPARE: 4677 rq->calc_load_update = calc_load_update; 4678 break; 4679 4680 case CPU_ONLINE: 4681 /* Update our root-domain */ 4682 raw_spin_lock_irqsave(&rq->lock, flags); 4683 if (rq->rd) { 4684 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span)); 4685 4686 set_rq_online(rq); 4687 } 4688 raw_spin_unlock_irqrestore(&rq->lock, flags); 4689 break; 4690 4691#ifdef CONFIG_HOTPLUG_CPU 4692 case CPU_DYING: 4693 sched_ttwu_pending(); 4694 /* Update our root-domain */ 4695 raw_spin_lock_irqsave(&rq->lock, flags); 4696 if (rq->rd) { 4697 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span)); 4698 set_rq_offline(rq); 4699 } 4700 migrate_tasks(cpu); 4701 BUG_ON(rq->nr_running != 1); /* the migration thread */ 4702 raw_spin_unlock_irqrestore(&rq->lock, flags); 4703 break; 4704 4705 case CPU_DEAD: 4706 calc_load_migrate(rq); 4707 break; 4708#endif 4709 } 4710 4711 update_max_interval(); 4712 4713 return NOTIFY_OK; 4714} 4715 4716/* 4717 * Register at high priority so that task migration (migrate_all_tasks) 4718 * happens before everything else. This has to be lower priority than 4719 * the notifier in the perf_event subsystem, though. 4720 */ 4721static struct notifier_block migration_notifier = { 4722 .notifier_call = migration_call, 4723 .priority = CPU_PRI_MIGRATION, 4724}; 4725 4726static int sched_cpu_active(struct notifier_block *nfb, 4727 unsigned long action, void *hcpu) 4728{ 4729 switch (action & ~CPU_TASKS_FROZEN) { 4730 case CPU_STARTING: 4731 case CPU_DOWN_FAILED: 4732 set_cpu_active((long)hcpu, true); 4733 return NOTIFY_OK; 4734 default: 4735 return NOTIFY_DONE; 4736 } 4737} 4738 4739static int sched_cpu_inactive(struct notifier_block *nfb, 4740 unsigned long action, void *hcpu) 4741{ 4742 switch (action & ~CPU_TASKS_FROZEN) { 4743 case CPU_DOWN_PREPARE: 4744 set_cpu_active((long)hcpu, false); 4745 return NOTIFY_OK; 4746 default: 4747 return NOTIFY_DONE; 4748 } 4749} 4750 4751static int __init migration_init(void) 4752{ 4753 void *cpu = (void *)(long)smp_processor_id(); 4754 int err; 4755 4756 /* Initialize migration for the boot CPU */ 4757 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu); 4758 BUG_ON(err == NOTIFY_BAD); 4759 migration_call(&migration_notifier, CPU_ONLINE, cpu); 4760 register_cpu_notifier(&migration_notifier); 4761 4762 /* Register cpu active notifiers */ 4763 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE); 4764 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE); 4765 4766 return 0; 4767} 4768early_initcall(migration_init); 4769#endif 4770 4771#ifdef CONFIG_SMP 4772 4773static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */ 4774 4775#ifdef CONFIG_SCHED_DEBUG 4776 4777static __read_mostly int sched_debug_enabled; 4778 4779static int __init sched_debug_setup(char *str) 4780{ 4781 sched_debug_enabled = 1; 4782 4783 return 0; 4784} 4785early_param("sched_debug", sched_debug_setup); 4786 4787static inline bool sched_debug(void) 4788{ 4789 return sched_debug_enabled; 4790} 4791 4792static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level, 4793 struct cpumask *groupmask) 4794{ 4795 struct sched_group *group = sd->groups; 4796 char str[256]; 4797 4798 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd)); 4799 cpumask_clear(groupmask); 4800 4801 printk(KERN_DEBUG "%*s domain %d: ", level, "", level); 4802 4803 if (!(sd->flags & SD_LOAD_BALANCE)) { 4804 printk("does not load-balance\n"); 4805 if (sd->parent) 4806 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain" 4807 " has parent"); 4808 return -1; 4809 } 4810 4811 printk(KERN_CONT "span %s level %s\n", str, sd->name); 4812 4813 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) { 4814 printk(KERN_ERR "ERROR: domain->span does not contain " 4815 "CPU%d\n", cpu); 4816 } 4817 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) { 4818 printk(KERN_ERR "ERROR: domain->groups does not contain" 4819 " CPU%d\n", cpu); 4820 } 4821 4822 printk(KERN_DEBUG "%*s groups:", level + 1, ""); 4823 do { 4824 if (!group) { 4825 printk("\n"); 4826 printk(KERN_ERR "ERROR: group is NULL\n"); 4827 break; 4828 } 4829 4830 /* 4831 * Even though we initialize ->power to something semi-sane, 4832 * we leave power_orig unset. This allows us to detect if 4833 * domain iteration is still funny without causing /0 traps. 4834 */ 4835 if (!group->sgp->power_orig) { 4836 printk(KERN_CONT "\n"); 4837 printk(KERN_ERR "ERROR: domain->cpu_power not " 4838 "set\n"); 4839 break; 4840 } 4841 4842 if (!cpumask_weight(sched_group_cpus(group))) { 4843 printk(KERN_CONT "\n"); 4844 printk(KERN_ERR "ERROR: empty group\n"); 4845 break; 4846 } 4847 4848 if (!(sd->flags & SD_OVERLAP) && 4849 cpumask_intersects(groupmask, sched_group_cpus(group))) { 4850 printk(KERN_CONT "\n"); 4851 printk(KERN_ERR "ERROR: repeated CPUs\n"); 4852 break; 4853 } 4854 4855 cpumask_or(groupmask, groupmask, sched_group_cpus(group)); 4856 4857 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group)); 4858 4859 printk(KERN_CONT " %s", str); 4860 if (group->sgp->power != SCHED_POWER_SCALE) { 4861 printk(KERN_CONT " (cpu_power = %d)", 4862 group->sgp->power); 4863 } 4864 4865 group = group->next; 4866 } while (group != sd->groups); 4867 printk(KERN_CONT "\n"); 4868 4869 if (!cpumask_equal(sched_domain_span(sd), groupmask)) 4870 printk(KERN_ERR "ERROR: groups don't span domain->span\n"); 4871 4872 if (sd->parent && 4873 !cpumask_subset(groupmask, sched_domain_span(sd->parent))) 4874 printk(KERN_ERR "ERROR: parent span is not a superset " 4875 "of domain->span\n"); 4876 return 0; 4877} 4878 4879static void sched_domain_debug(struct sched_domain *sd, int cpu) 4880{ 4881 int level = 0; 4882 4883 if (!sched_debug_enabled) 4884 return; 4885 4886 if (!sd) { 4887 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu); 4888 return; 4889 } 4890 4891 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu); 4892 4893 for (;;) { 4894 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask)) 4895 break; 4896 level++; 4897 sd = sd->parent; 4898 if (!sd) 4899 break; 4900 } 4901} 4902#else /* !CONFIG_SCHED_DEBUG */ 4903# define sched_domain_debug(sd, cpu) do { } while (0) 4904static inline bool sched_debug(void) 4905{ 4906 return false; 4907} 4908#endif /* CONFIG_SCHED_DEBUG */ 4909 4910static int sd_degenerate(struct sched_domain *sd) 4911{ 4912 if (cpumask_weight(sched_domain_span(sd)) == 1) 4913 return 1; 4914 4915 /* Following flags need at least 2 groups */ 4916 if (sd->flags & (SD_LOAD_BALANCE | 4917 SD_BALANCE_NEWIDLE | 4918 SD_BALANCE_FORK | 4919 SD_BALANCE_EXEC | 4920 SD_SHARE_CPUPOWER | 4921 SD_SHARE_PKG_RESOURCES)) { 4922 if (sd->groups != sd->groups->next) 4923 return 0; 4924 } 4925 4926 /* Following flags don't use groups */ 4927 if (sd->flags & (SD_WAKE_AFFINE)) 4928 return 0; 4929 4930 return 1; 4931} 4932 4933static int 4934sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent) 4935{ 4936 unsigned long cflags = sd->flags, pflags = parent->flags; 4937 4938 if (sd_degenerate(parent)) 4939 return 1; 4940 4941 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent))) 4942 return 0; 4943 4944 /* Flags needing groups don't count if only 1 group in parent */ 4945 if (parent->groups == parent->groups->next) { 4946 pflags &= ~(SD_LOAD_BALANCE | 4947 SD_BALANCE_NEWIDLE | 4948 SD_BALANCE_FORK | 4949 SD_BALANCE_EXEC | 4950 SD_SHARE_CPUPOWER | 4951 SD_SHARE_PKG_RESOURCES | 4952 SD_PREFER_SIBLING); 4953 if (nr_node_ids == 1) 4954 pflags &= ~SD_SERIALIZE; 4955 } 4956 if (~cflags & pflags) 4957 return 0; 4958 4959 return 1; 4960} 4961 4962static void free_rootdomain(struct rcu_head *rcu) 4963{ 4964 struct root_domain *rd = container_of(rcu, struct root_domain, rcu); 4965 4966 cpupri_cleanup(&rd->cpupri); 4967 free_cpumask_var(rd->rto_mask); 4968 free_cpumask_var(rd->online); 4969 free_cpumask_var(rd->span); 4970 kfree(rd); 4971} 4972 4973static void rq_attach_root(struct rq *rq, struct root_domain *rd) 4974{ 4975 struct root_domain *old_rd = NULL; 4976 unsigned long flags; 4977 4978 raw_spin_lock_irqsave(&rq->lock, flags); 4979 4980 if (rq->rd) { 4981 old_rd = rq->rd; 4982 4983 if (cpumask_test_cpu(rq->cpu, old_rd->online)) 4984 set_rq_offline(rq); 4985 4986 cpumask_clear_cpu(rq->cpu, old_rd->span); 4987 4988 /* 4989 * If we dont want to free the old_rt yet then 4990 * set old_rd to NULL to skip the freeing later 4991 * in this function: 4992 */ 4993 if (!atomic_dec_and_test(&old_rd->refcount)) 4994 old_rd = NULL; 4995 } 4996 4997 atomic_inc(&rd->refcount); 4998 rq->rd = rd; 4999 5000 cpumask_set_cpu(rq->cpu, rd->span); 5001 if (cpumask_test_cpu(rq->cpu, cpu_active_mask)) 5002 set_rq_online(rq); 5003 5004 raw_spin_unlock_irqrestore(&rq->lock, flags); 5005 5006 if (old_rd) 5007 call_rcu_sched(&old_rd->rcu, free_rootdomain); 5008} 5009 5010static int init_rootdomain(struct root_domain *rd) 5011{ 5012 memset(rd, 0, sizeof(*rd)); 5013 5014 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL)) 5015 goto out; 5016 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL)) 5017 goto free_span; 5018 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL)) 5019 goto free_online; 5020 5021 if (cpupri_init(&rd->cpupri) != 0) 5022 goto free_rto_mask; 5023 return 0; 5024 5025free_rto_mask: 5026 free_cpumask_var(rd->rto_mask); 5027free_online: 5028 free_cpumask_var(rd->online); 5029free_span: 5030 free_cpumask_var(rd->span); 5031out: 5032 return -ENOMEM; 5033} 5034 5035/* 5036 * By default the system creates a single root-domain with all cpus as 5037 * members (mimicking the global state we have today). 5038 */ 5039struct root_domain def_root_domain; 5040 5041static void init_defrootdomain(void) 5042{ 5043 init_rootdomain(&def_root_domain); 5044 5045 atomic_set(&def_root_domain.refcount, 1); 5046} 5047 5048static struct root_domain *alloc_rootdomain(void) 5049{ 5050 struct root_domain *rd; 5051 5052 rd = kmalloc(sizeof(*rd), GFP_KERNEL); 5053 if (!rd) 5054 return NULL; 5055 5056 if (init_rootdomain(rd) != 0) { 5057 kfree(rd); 5058 return NULL; 5059 } 5060 5061 return rd; 5062} 5063 5064static void free_sched_groups(struct sched_group *sg, int free_sgp) 5065{ 5066 struct sched_group *tmp, *first; 5067 5068 if (!sg) 5069 return; 5070 5071 first = sg; 5072 do { 5073 tmp = sg->next; 5074 5075 if (free_sgp && atomic_dec_and_test(&sg->sgp->ref)) 5076 kfree(sg->sgp); 5077 5078 kfree(sg); 5079 sg = tmp; 5080 } while (sg != first); 5081} 5082 5083static void free_sched_domain(struct rcu_head *rcu) 5084{ 5085 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu); 5086 5087 /* 5088 * If its an overlapping domain it has private groups, iterate and 5089 * nuke them all. 5090 */ 5091 if (sd->flags & SD_OVERLAP) { 5092 free_sched_groups(sd->groups, 1); 5093 } else if (atomic_dec_and_test(&sd->groups->ref)) { 5094 kfree(sd->groups->sgp); 5095 kfree(sd->groups); 5096 } 5097 kfree(sd); 5098} 5099 5100static void destroy_sched_domain(struct sched_domain *sd, int cpu) 5101{ 5102 call_rcu(&sd->rcu, free_sched_domain); 5103} 5104 5105static void destroy_sched_domains(struct sched_domain *sd, int cpu) 5106{ 5107 for (; sd; sd = sd->parent) 5108 destroy_sched_domain(sd, cpu); 5109} 5110 5111/* 5112 * Keep a special pointer to the highest sched_domain that has 5113 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this 5114 * allows us to avoid some pointer chasing select_idle_sibling(). 5115 * 5116 * Also keep a unique ID per domain (we use the first cpu number in 5117 * the cpumask of the domain), this allows us to quickly tell if 5118 * two cpus are in the same cache domain, see cpus_share_cache(). 5119 */ 5120DEFINE_PER_CPU(struct sched_domain *, sd_llc); 5121DEFINE_PER_CPU(int, sd_llc_size); 5122DEFINE_PER_CPU(int, sd_llc_id); 5123 5124static void update_top_cache_domain(int cpu) 5125{ 5126 struct sched_domain *sd; 5127 int id = cpu; 5128 int size = 1; 5129 5130 sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES); 5131 if (sd) { 5132 id = cpumask_first(sched_domain_span(sd)); 5133 size = cpumask_weight(sched_domain_span(sd)); 5134 } 5135 5136 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd); 5137 per_cpu(sd_llc_size, cpu) = size; 5138 per_cpu(sd_llc_id, cpu) = id; 5139} 5140 5141/* 5142 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must 5143 * hold the hotplug lock. 5144 */ 5145static void 5146cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu) 5147{ 5148 struct rq *rq = cpu_rq(cpu); 5149 struct sched_domain *tmp; 5150 5151 /* Remove the sched domains which do not contribute to scheduling. */ 5152 for (tmp = sd; tmp; ) { 5153 struct sched_domain *parent = tmp->parent; 5154 if (!parent) 5155 break; 5156 5157 if (sd_parent_degenerate(tmp, parent)) { 5158 tmp->parent = parent->parent; 5159 if (parent->parent) 5160 parent->parent->child = tmp; 5161 /* 5162 * Transfer SD_PREFER_SIBLING down in case of a 5163 * degenerate parent; the spans match for this 5164 * so the property transfers. 5165 */ 5166 if (parent->flags & SD_PREFER_SIBLING) 5167 tmp->flags |= SD_PREFER_SIBLING; 5168 destroy_sched_domain(parent, cpu); 5169 } else 5170 tmp = tmp->parent; 5171 } 5172 5173 if (sd && sd_degenerate(sd)) { 5174 tmp = sd; 5175 sd = sd->parent; 5176 destroy_sched_domain(tmp, cpu); 5177 if (sd) 5178 sd->child = NULL; 5179 } 5180 5181 sched_domain_debug(sd, cpu); 5182 5183 rq_attach_root(rq, rd); 5184 tmp = rq->sd; 5185 rcu_assign_pointer(rq->sd, sd); 5186 destroy_sched_domains(tmp, cpu); 5187 5188 update_top_cache_domain(cpu); 5189} 5190 5191/* cpus with isolated domains */ 5192static cpumask_var_t cpu_isolated_map; 5193 5194/* Setup the mask of cpus configured for isolated domains */ 5195static int __init isolated_cpu_setup(char *str) 5196{ 5197 alloc_bootmem_cpumask_var(&cpu_isolated_map); 5198 cpulist_parse(str, cpu_isolated_map); 5199 return 1; 5200} 5201 5202__setup("isolcpus=", isolated_cpu_setup); 5203 5204static const struct cpumask *cpu_cpu_mask(int cpu) 5205{ 5206 return cpumask_of_node(cpu_to_node(cpu)); 5207} 5208 5209struct sd_data { 5210 struct sched_domain **__percpu sd; 5211 struct sched_group **__percpu sg; 5212 struct sched_group_power **__percpu sgp; 5213}; 5214 5215struct s_data { 5216 struct sched_domain ** __percpu sd; 5217 struct root_domain *rd; 5218}; 5219 5220enum s_alloc { 5221 sa_rootdomain, 5222 sa_sd, 5223 sa_sd_storage, 5224 sa_none, 5225}; 5226 5227struct sched_domain_topology_level; 5228 5229typedef struct sched_domain *(*sched_domain_init_f)(struct sched_domain_topology_level *tl, int cpu); 5230typedef const struct cpumask *(*sched_domain_mask_f)(int cpu); 5231 5232#define SDTL_OVERLAP 0x01 5233 5234struct sched_domain_topology_level { 5235 sched_domain_init_f init; 5236 sched_domain_mask_f mask; 5237 int flags; 5238 int numa_level; 5239 struct sd_data data; 5240}; 5241 5242/* 5243 * Build an iteration mask that can exclude certain CPUs from the upwards 5244 * domain traversal. 5245 * 5246 * Asymmetric node setups can result in situations where the domain tree is of 5247 * unequal depth, make sure to skip domains that already cover the entire 5248 * range. 5249 * 5250 * In that case build_sched_domains() will have terminated the iteration early 5251 * and our sibling sd spans will be empty. Domains should always include the 5252 * cpu they're built on, so check that. 5253 * 5254 */ 5255static void build_group_mask(struct sched_domain *sd, struct sched_group *sg) 5256{ 5257 const struct cpumask *span = sched_domain_span(sd); 5258 struct sd_data *sdd = sd->private; 5259 struct sched_domain *sibling; 5260 int i; 5261 5262 for_each_cpu(i, span) { 5263 sibling = *per_cpu_ptr(sdd->sd, i); 5264 if (!cpumask_test_cpu(i, sched_domain_span(sibling))) 5265 continue; 5266 5267 cpumask_set_cpu(i, sched_group_mask(sg)); 5268 } 5269} 5270 5271/* 5272 * Return the canonical balance cpu for this group, this is the first cpu 5273 * of this group that's also in the iteration mask. 5274 */ 5275int group_balance_cpu(struct sched_group *sg) 5276{ 5277 return cpumask_first_and(sched_group_cpus(sg), sched_group_mask(sg)); 5278} 5279 5280static int 5281build_overlap_sched_groups(struct sched_domain *sd, int cpu) 5282{ 5283 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg; 5284 const struct cpumask *span = sched_domain_span(sd); 5285 struct cpumask *covered = sched_domains_tmpmask; 5286 struct sd_data *sdd = sd->private; 5287 struct sched_domain *child; 5288 int i; 5289 5290 cpumask_clear(covered); 5291 5292 for_each_cpu(i, span) { 5293 struct cpumask *sg_span; 5294 5295 if (cpumask_test_cpu(i, covered)) 5296 continue; 5297 5298 child = *per_cpu_ptr(sdd->sd, i); 5299 5300 /* See the comment near build_group_mask(). */ 5301 if (!cpumask_test_cpu(i, sched_domain_span(child))) 5302 continue; 5303 5304 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(), 5305 GFP_KERNEL, cpu_to_node(cpu)); 5306 5307 if (!sg) 5308 goto fail; 5309 5310 sg_span = sched_group_cpus(sg); 5311 if (child->child) { 5312 child = child->child; 5313 cpumask_copy(sg_span, sched_domain_span(child)); 5314 } else 5315 cpumask_set_cpu(i, sg_span); 5316 5317 cpumask_or(covered, covered, sg_span); 5318 5319 sg->sgp = *per_cpu_ptr(sdd->sgp, i); 5320 if (atomic_inc_return(&sg->sgp->ref) == 1) 5321 build_group_mask(sd, sg); 5322 5323 /* 5324 * Initialize sgp->power such that even if we mess up the 5325 * domains and no possible iteration will get us here, we won't 5326 * die on a /0 trap. 5327 */ 5328 sg->sgp->power = SCHED_POWER_SCALE * cpumask_weight(sg_span); 5329 5330 /* 5331 * Make sure the first group of this domain contains the 5332 * canonical balance cpu. Otherwise the sched_domain iteration 5333 * breaks. See update_sg_lb_stats(). 5334 */ 5335 if ((!groups && cpumask_test_cpu(cpu, sg_span)) || 5336 group_balance_cpu(sg) == cpu) 5337 groups = sg; 5338 5339 if (!first) 5340 first = sg; 5341 if (last) 5342 last->next = sg; 5343 last = sg; 5344 last->next = first; 5345 } 5346 sd->groups = groups; 5347 5348 return 0; 5349 5350fail: 5351 free_sched_groups(first, 0); 5352 5353 return -ENOMEM; 5354} 5355 5356static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg) 5357{ 5358 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu); 5359 struct sched_domain *child = sd->child; 5360 5361 if (child) 5362 cpu = cpumask_first(sched_domain_span(child)); 5363 5364 if (sg) { 5365 *sg = *per_cpu_ptr(sdd->sg, cpu); 5366 (*sg)->sgp = *per_cpu_ptr(sdd->sgp, cpu); 5367 atomic_set(&(*sg)->sgp->ref, 1); /* for claim_allocations */ 5368 } 5369 5370 return cpu; 5371} 5372 5373/* 5374 * build_sched_groups will build a circular linked list of the groups 5375 * covered by the given span, and will set each group's ->cpumask correctly, 5376 * and ->cpu_power to 0. 5377 * 5378 * Assumes the sched_domain tree is fully constructed 5379 */ 5380static int 5381build_sched_groups(struct sched_domain *sd, int cpu) 5382{ 5383 struct sched_group *first = NULL, *last = NULL; 5384 struct sd_data *sdd = sd->private; 5385 const struct cpumask *span = sched_domain_span(sd); 5386 struct cpumask *covered; 5387 int i; 5388 5389 get_group(cpu, sdd, &sd->groups); 5390 atomic_inc(&sd->groups->ref); 5391 5392 if (cpu != cpumask_first(span)) 5393 return 0; 5394 5395 lockdep_assert_held(&sched_domains_mutex); 5396 covered = sched_domains_tmpmask; 5397 5398 cpumask_clear(covered); 5399 5400 for_each_cpu(i, span) { 5401 struct sched_group *sg; 5402 int group, j; 5403 5404 if (cpumask_test_cpu(i, covered)) 5405 continue; 5406 5407 group = get_group(i, sdd, &sg); 5408 cpumask_clear(sched_group_cpus(sg)); 5409 sg->sgp->power = 0; 5410 cpumask_setall(sched_group_mask(sg)); 5411 5412 for_each_cpu(j, span) { 5413 if (get_group(j, sdd, NULL) != group) 5414 continue; 5415 5416 cpumask_set_cpu(j, covered); 5417 cpumask_set_cpu(j, sched_group_cpus(sg)); 5418 } 5419 5420 if (!first) 5421 first = sg; 5422 if (last) 5423 last->next = sg; 5424 last = sg; 5425 } 5426 last->next = first; 5427 5428 return 0; 5429} 5430 5431/* 5432 * Initialize sched groups cpu_power. 5433 * 5434 * cpu_power indicates the capacity of sched group, which is used while 5435 * distributing the load between different sched groups in a sched domain. 5436 * Typically cpu_power for all the groups in a sched domain will be same unless 5437 * there are asymmetries in the topology. If there are asymmetries, group 5438 * having more cpu_power will pickup more load compared to the group having 5439 * less cpu_power. 5440 */ 5441static void init_sched_groups_power(int cpu, struct sched_domain *sd) 5442{ 5443 struct sched_group *sg = sd->groups; 5444 5445 WARN_ON(!sg); 5446 5447 do { 5448 sg->group_weight = cpumask_weight(sched_group_cpus(sg)); 5449 sg = sg->next; 5450 } while (sg != sd->groups); 5451 5452 if (cpu != group_balance_cpu(sg)) 5453 return; 5454 5455 update_group_power(sd, cpu); 5456 atomic_set(&sg->sgp->nr_busy_cpus, sg->group_weight); 5457} 5458 5459int __weak arch_sd_sibling_asym_packing(void) 5460{ 5461 return 0*SD_ASYM_PACKING; 5462} 5463 5464/* 5465 * Initializers for schedule domains 5466 * Non-inlined to reduce accumulated stack pressure in build_sched_domains() 5467 */ 5468 5469#ifdef CONFIG_SCHED_DEBUG 5470# define SD_INIT_NAME(sd, type) sd->name = #type 5471#else 5472# define SD_INIT_NAME(sd, type) do { } while (0) 5473#endif 5474 5475#define SD_INIT_FUNC(type) \ 5476static noinline struct sched_domain * \ 5477sd_init_##type(struct sched_domain_topology_level *tl, int cpu) \ 5478{ \ 5479 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); \ 5480 *sd = SD_##type##_INIT; \ 5481 SD_INIT_NAME(sd, type); \ 5482 sd->private = &tl->data; \ 5483 return sd; \ 5484} 5485 5486SD_INIT_FUNC(CPU) 5487#ifdef CONFIG_SCHED_SMT 5488 SD_INIT_FUNC(SIBLING) 5489#endif 5490#ifdef CONFIG_SCHED_MC 5491 SD_INIT_FUNC(MC) 5492#endif 5493#ifdef CONFIG_SCHED_BOOK 5494 SD_INIT_FUNC(BOOK) 5495#endif 5496 5497static int default_relax_domain_level = -1; 5498int sched_domain_level_max; 5499 5500static int __init setup_relax_domain_level(char *str) 5501{ 5502 if (kstrtoint(str, 0, &default_relax_domain_level)) 5503 pr_warn("Unable to set relax_domain_level\n"); 5504 5505 return 1; 5506} 5507__setup("relax_domain_level=", setup_relax_domain_level); 5508 5509static void set_domain_attribute(struct sched_domain *sd, 5510 struct sched_domain_attr *attr) 5511{ 5512 int request; 5513 5514 if (!attr || attr->relax_domain_level < 0) { 5515 if (default_relax_domain_level < 0) 5516 return; 5517 else 5518 request = default_relax_domain_level; 5519 } else 5520 request = attr->relax_domain_level; 5521 if (request < sd->level) { 5522 /* turn off idle balance on this domain */ 5523 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE); 5524 } else { 5525 /* turn on idle balance on this domain */ 5526 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE); 5527 } 5528} 5529 5530static void __sdt_free(const struct cpumask *cpu_map); 5531static int __sdt_alloc(const struct cpumask *cpu_map); 5532 5533static void __free_domain_allocs(struct s_data *d, enum s_alloc what, 5534 const struct cpumask *cpu_map) 5535{ 5536 switch (what) { 5537 case sa_rootdomain: 5538 if (!atomic_read(&d->rd->refcount)) 5539 free_rootdomain(&d->rd->rcu); /* fall through */ 5540 case sa_sd: 5541 free_percpu(d->sd); /* fall through */ 5542 case sa_sd_storage: 5543 __sdt_free(cpu_map); /* fall through */ 5544 case sa_none: 5545 break; 5546 } 5547} 5548 5549static enum s_alloc __visit_domain_allocation_hell(struct s_data *d, 5550 const struct cpumask *cpu_map) 5551{ 5552 memset(d, 0, sizeof(*d)); 5553 5554 if (__sdt_alloc(cpu_map)) 5555 return sa_sd_storage; 5556 d->sd = alloc_percpu(struct sched_domain *); 5557 if (!d->sd) 5558 return sa_sd_storage; 5559 d->rd = alloc_rootdomain(); 5560 if (!d->rd) 5561 return sa_sd; 5562 return sa_rootdomain; 5563} 5564 5565/* 5566 * NULL the sd_data elements we've used to build the sched_domain and 5567 * sched_group structure so that the subsequent __free_domain_allocs() 5568 * will not free the data we're using. 5569 */ 5570static void claim_allocations(int cpu, struct sched_domain *sd) 5571{ 5572 struct sd_data *sdd = sd->private; 5573 5574 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd); 5575 *per_cpu_ptr(sdd->sd, cpu) = NULL; 5576 5577 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref)) 5578 *per_cpu_ptr(sdd->sg, cpu) = NULL; 5579 5580 if (atomic_read(&(*per_cpu_ptr(sdd->sgp, cpu))->ref)) 5581 *per_cpu_ptr(sdd->sgp, cpu) = NULL; 5582} 5583 5584#ifdef CONFIG_SCHED_SMT 5585static const struct cpumask *cpu_smt_mask(int cpu) 5586{ 5587 return topology_thread_cpumask(cpu); 5588} 5589#endif 5590 5591/* 5592 * Topology list, bottom-up. 5593 */ 5594static struct sched_domain_topology_level default_topology[] = { 5595#ifdef CONFIG_SCHED_SMT 5596 { sd_init_SIBLING, cpu_smt_mask, }, 5597#endif 5598#ifdef CONFIG_SCHED_MC 5599 { sd_init_MC, cpu_coregroup_mask, }, 5600#endif 5601#ifdef CONFIG_SCHED_BOOK 5602 { sd_init_BOOK, cpu_book_mask, }, 5603#endif 5604 { sd_init_CPU, cpu_cpu_mask, }, 5605 { NULL, }, 5606}; 5607 5608static struct sched_domain_topology_level *sched_domain_topology = default_topology; 5609 5610#define for_each_sd_topology(tl) \ 5611 for (tl = sched_domain_topology; tl->init; tl++) 5612 5613#ifdef CONFIG_NUMA 5614 5615static int sched_domains_numa_levels; 5616static int *sched_domains_numa_distance; 5617static struct cpumask ***sched_domains_numa_masks; 5618static int sched_domains_curr_level; 5619 5620static inline int sd_local_flags(int level) 5621{ 5622 if (sched_domains_numa_distance[level] > RECLAIM_DISTANCE) 5623 return 0; 5624 5625 return SD_BALANCE_EXEC | SD_BALANCE_FORK | SD_WAKE_AFFINE; 5626} 5627 5628static struct sched_domain * 5629sd_numa_init(struct sched_domain_topology_level *tl, int cpu) 5630{ 5631 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); 5632 int level = tl->numa_level; 5633 int sd_weight = cpumask_weight( 5634 sched_domains_numa_masks[level][cpu_to_node(cpu)]); 5635 5636 *sd = (struct sched_domain){ 5637 .min_interval = sd_weight, 5638 .max_interval = 2*sd_weight, 5639 .busy_factor = 32, 5640 .imbalance_pct = 125, 5641 .cache_nice_tries = 2, 5642 .busy_idx = 3, 5643 .idle_idx = 2, 5644 .newidle_idx = 0, 5645 .wake_idx = 0, 5646 .forkexec_idx = 0, 5647 5648 .flags = 1*SD_LOAD_BALANCE 5649 | 1*SD_BALANCE_NEWIDLE 5650 | 0*SD_BALANCE_EXEC 5651 | 0*SD_BALANCE_FORK 5652 | 0*SD_BALANCE_WAKE 5653 | 0*SD_WAKE_AFFINE 5654 | 0*SD_SHARE_CPUPOWER 5655 | 0*SD_SHARE_PKG_RESOURCES 5656 | 1*SD_SERIALIZE 5657 | 0*SD_PREFER_SIBLING 5658 | sd_local_flags(level) 5659 , 5660 .last_balance = jiffies, 5661 .balance_interval = sd_weight, 5662 }; 5663 SD_INIT_NAME(sd, NUMA); 5664 sd->private = &tl->data; 5665 5666 /* 5667 * Ugly hack to pass state to sd_numa_mask()... 5668 */ 5669 sched_domains_curr_level = tl->numa_level; 5670 5671 return sd; 5672} 5673 5674static const struct cpumask *sd_numa_mask(int cpu) 5675{ 5676 return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)]; 5677} 5678 5679static void sched_numa_warn(const char *str) 5680{ 5681 static int done = false; 5682 int i,j; 5683 5684 if (done) 5685 return; 5686 5687 done = true; 5688 5689 printk(KERN_WARNING "ERROR: %s\n\n", str); 5690 5691 for (i = 0; i < nr_node_ids; i++) { 5692 printk(KERN_WARNING " "); 5693 for (j = 0; j < nr_node_ids; j++) 5694 printk(KERN_CONT "%02d ", node_distance(i,j)); 5695 printk(KERN_CONT "\n"); 5696 } 5697 printk(KERN_WARNING "\n"); 5698} 5699 5700static bool find_numa_distance(int distance) 5701{ 5702 int i; 5703 5704 if (distance == node_distance(0, 0)) 5705 return true; 5706 5707 for (i = 0; i < sched_domains_numa_levels; i++) { 5708 if (sched_domains_numa_distance[i] == distance) 5709 return true; 5710 } 5711 5712 return false; 5713} 5714 5715static void sched_init_numa(void) 5716{ 5717 int next_distance, curr_distance = node_distance(0, 0); 5718 struct sched_domain_topology_level *tl; 5719 int level = 0; 5720 int i, j, k; 5721 5722 sched_domains_numa_distance = kzalloc(sizeof(int) * nr_node_ids, GFP_KERNEL); 5723 if (!sched_domains_numa_distance) 5724 return; 5725 5726 /* 5727 * O(nr_nodes^2) deduplicating selection sort -- in order to find the 5728 * unique distances in the node_distance() table. 5729 * 5730 * Assumes node_distance(0,j) includes all distances in 5731 * node_distance(i,j) in order to avoid cubic time. 5732 */ 5733 next_distance = curr_distance; 5734 for (i = 0; i < nr_node_ids; i++) { 5735 for (j = 0; j < nr_node_ids; j++) { 5736 for (k = 0; k < nr_node_ids; k++) { 5737 int distance = node_distance(i, k); 5738 5739 if (distance > curr_distance && 5740 (distance < next_distance || 5741 next_distance == curr_distance)) 5742 next_distance = distance; 5743 5744 /* 5745 * While not a strong assumption it would be nice to know 5746 * about cases where if node A is connected to B, B is not 5747 * equally connected to A. 5748 */ 5749 if (sched_debug() && node_distance(k, i) != distance) 5750 sched_numa_warn("Node-distance not symmetric"); 5751 5752 if (sched_debug() && i && !find_numa_distance(distance)) 5753 sched_numa_warn("Node-0 not representative"); 5754 } 5755 if (next_distance != curr_distance) { 5756 sched_domains_numa_distance[level++] = next_distance; 5757 sched_domains_numa_levels = level; 5758 curr_distance = next_distance; 5759 } else break; 5760 } 5761 5762 /* 5763 * In case of sched_debug() we verify the above assumption. 5764 */ 5765 if (!sched_debug()) 5766 break; 5767 } 5768 /* 5769 * 'level' contains the number of unique distances, excluding the 5770 * identity distance node_distance(i,i). 5771 * 5772 * The sched_domains_numa_distance[] array includes the actual distance 5773 * numbers. 5774 */ 5775 5776 /* 5777 * Here, we should temporarily reset sched_domains_numa_levels to 0. 5778 * If it fails to allocate memory for array sched_domains_numa_masks[][], 5779 * the array will contain less then 'level' members. This could be 5780 * dangerous when we use it to iterate array sched_domains_numa_masks[][] 5781 * in other functions. 5782 * 5783 * We reset it to 'level' at the end of this function. 5784 */ 5785 sched_domains_numa_levels = 0; 5786 5787 sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL); 5788 if (!sched_domains_numa_masks) 5789 return; 5790 5791 /* 5792 * Now for each level, construct a mask per node which contains all 5793 * cpus of nodes that are that many hops away from us. 5794 */ 5795 for (i = 0; i < level; i++) { 5796 sched_domains_numa_masks[i] = 5797 kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL); 5798 if (!sched_domains_numa_masks[i]) 5799 return; 5800 5801 for (j = 0; j < nr_node_ids; j++) { 5802 struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL); 5803 if (!mask) 5804 return; 5805 5806 sched_domains_numa_masks[i][j] = mask; 5807 5808 for (k = 0; k < nr_node_ids; k++) { 5809 if (node_distance(j, k) > sched_domains_numa_distance[i]) 5810 continue; 5811 5812 cpumask_or(mask, mask, cpumask_of_node(k)); 5813 } 5814 } 5815 } 5816 5817 tl = kzalloc((ARRAY_SIZE(default_topology) + level) * 5818 sizeof(struct sched_domain_topology_level), GFP_KERNEL); 5819 if (!tl) 5820 return; 5821 5822 /* 5823 * Copy the default topology bits.. 5824 */ 5825 for (i = 0; default_topology[i].init; i++) 5826 tl[i] = default_topology[i]; 5827 5828 /* 5829 * .. and append 'j' levels of NUMA goodness. 5830 */ 5831 for (j = 0; j < level; i++, j++) { 5832 tl[i] = (struct sched_domain_topology_level){ 5833 .init = sd_numa_init, 5834 .mask = sd_numa_mask, 5835 .flags = SDTL_OVERLAP, 5836 .numa_level = j, 5837 }; 5838 } 5839 5840 sched_domain_topology = tl; 5841 5842 sched_domains_numa_levels = level; 5843} 5844 5845static void sched_domains_numa_masks_set(int cpu) 5846{ 5847 int i, j; 5848 int node = cpu_to_node(cpu); 5849 5850 for (i = 0; i < sched_domains_numa_levels; i++) { 5851 for (j = 0; j < nr_node_ids; j++) { 5852 if (node_distance(j, node) <= sched_domains_numa_distance[i]) 5853 cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]); 5854 } 5855 } 5856} 5857 5858static void sched_domains_numa_masks_clear(int cpu) 5859{ 5860 int i, j; 5861 for (i = 0; i < sched_domains_numa_levels; i++) { 5862 for (j = 0; j < nr_node_ids; j++) 5863 cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]); 5864 } 5865} 5866 5867/* 5868 * Update sched_domains_numa_masks[level][node] array when new cpus 5869 * are onlined. 5870 */ 5871static int sched_domains_numa_masks_update(struct notifier_block *nfb, 5872 unsigned long action, 5873 void *hcpu) 5874{ 5875 int cpu = (long)hcpu; 5876 5877 switch (action & ~CPU_TASKS_FROZEN) { 5878 case CPU_ONLINE: 5879 sched_domains_numa_masks_set(cpu); 5880 break; 5881 5882 case CPU_DEAD: 5883 sched_domains_numa_masks_clear(cpu); 5884 break; 5885 5886 default: 5887 return NOTIFY_DONE; 5888 } 5889 5890 return NOTIFY_OK; 5891} 5892#else 5893static inline void sched_init_numa(void) 5894{ 5895} 5896 5897static int sched_domains_numa_masks_update(struct notifier_block *nfb, 5898 unsigned long action, 5899 void *hcpu) 5900{ 5901 return 0; 5902} 5903#endif /* CONFIG_NUMA */ 5904 5905static int __sdt_alloc(const struct cpumask *cpu_map) 5906{ 5907 struct sched_domain_topology_level *tl; 5908 int j; 5909 5910 for_each_sd_topology(tl) { 5911 struct sd_data *sdd = &tl->data; 5912 5913 sdd->sd = alloc_percpu(struct sched_domain *); 5914 if (!sdd->sd) 5915 return -ENOMEM; 5916 5917 sdd->sg = alloc_percpu(struct sched_group *); 5918 if (!sdd->sg) 5919 return -ENOMEM; 5920 5921 sdd->sgp = alloc_percpu(struct sched_group_power *); 5922 if (!sdd->sgp) 5923 return -ENOMEM; 5924 5925 for_each_cpu(j, cpu_map) { 5926 struct sched_domain *sd; 5927 struct sched_group *sg; 5928 struct sched_group_power *sgp; 5929 5930 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(), 5931 GFP_KERNEL, cpu_to_node(j)); 5932 if (!sd) 5933 return -ENOMEM; 5934 5935 *per_cpu_ptr(sdd->sd, j) = sd; 5936 5937 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(), 5938 GFP_KERNEL, cpu_to_node(j)); 5939 if (!sg) 5940 return -ENOMEM; 5941 5942 sg->next = sg; 5943 5944 *per_cpu_ptr(sdd->sg, j) = sg; 5945 5946 sgp = kzalloc_node(sizeof(struct sched_group_power) + cpumask_size(), 5947 GFP_KERNEL, cpu_to_node(j)); 5948 if (!sgp) 5949 return -ENOMEM; 5950 5951 *per_cpu_ptr(sdd->sgp, j) = sgp; 5952 } 5953 } 5954 5955 return 0; 5956} 5957 5958static void __sdt_free(const struct cpumask *cpu_map) 5959{ 5960 struct sched_domain_topology_level *tl; 5961 int j; 5962 5963 for_each_sd_topology(tl) { 5964 struct sd_data *sdd = &tl->data; 5965 5966 for_each_cpu(j, cpu_map) { 5967 struct sched_domain *sd; 5968 5969 if (sdd->sd) { 5970 sd = *per_cpu_ptr(sdd->sd, j); 5971 if (sd && (sd->flags & SD_OVERLAP)) 5972 free_sched_groups(sd->groups, 0); 5973 kfree(*per_cpu_ptr(sdd->sd, j)); 5974 } 5975 5976 if (sdd->sg) 5977 kfree(*per_cpu_ptr(sdd->sg, j)); 5978 if (sdd->sgp) 5979 kfree(*per_cpu_ptr(sdd->sgp, j)); 5980 } 5981 free_percpu(sdd->sd); 5982 sdd->sd = NULL; 5983 free_percpu(sdd->sg); 5984 sdd->sg = NULL; 5985 free_percpu(sdd->sgp); 5986 sdd->sgp = NULL; 5987 } 5988} 5989 5990struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl, 5991 const struct cpumask *cpu_map, struct sched_domain_attr *attr, 5992 struct sched_domain *child, int cpu) 5993{ 5994 struct sched_domain *sd = tl->init(tl, cpu); 5995 if (!sd) 5996 return child; 5997 5998 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu)); 5999 if (child) { 6000 sd->level = child->level + 1; 6001 sched_domain_level_max = max(sched_domain_level_max, sd->level); 6002 child->parent = sd; 6003 sd->child = child; 6004 } 6005 set_domain_attribute(sd, attr); 6006 6007 return sd; 6008} 6009 6010/* 6011 * Build sched domains for a given set of cpus and attach the sched domains 6012 * to the individual cpus 6013 */ 6014static int build_sched_domains(const struct cpumask *cpu_map, 6015 struct sched_domain_attr *attr) 6016{ 6017 enum s_alloc alloc_state; 6018 struct sched_domain *sd; 6019 struct s_data d; 6020 int i, ret = -ENOMEM; 6021 6022 alloc_state = __visit_domain_allocation_hell(&d, cpu_map); 6023 if (alloc_state != sa_rootdomain) 6024 goto error; 6025 6026 /* Set up domains for cpus specified by the cpu_map. */ 6027 for_each_cpu(i, cpu_map) { 6028 struct sched_domain_topology_level *tl; 6029 6030 sd = NULL; 6031 for_each_sd_topology(tl) { 6032 sd = build_sched_domain(tl, cpu_map, attr, sd, i); 6033 if (tl == sched_domain_topology) 6034 *per_cpu_ptr(d.sd, i) = sd; 6035 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP)) 6036 sd->flags |= SD_OVERLAP; 6037 if (cpumask_equal(cpu_map, sched_domain_span(sd))) 6038 break; 6039 } 6040 } 6041 6042 /* Build the groups for the domains */ 6043 for_each_cpu(i, cpu_map) { 6044 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) { 6045 sd->span_weight = cpumask_weight(sched_domain_span(sd)); 6046 if (sd->flags & SD_OVERLAP) { 6047 if (build_overlap_sched_groups(sd, i)) 6048 goto error; 6049 } else { 6050 if (build_sched_groups(sd, i)) 6051 goto error; 6052 } 6053 } 6054 } 6055 6056 /* Calculate CPU power for physical packages and nodes */ 6057 for (i = nr_cpumask_bits-1; i >= 0; i--) { 6058 if (!cpumask_test_cpu(i, cpu_map)) 6059 continue; 6060 6061 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) { 6062 claim_allocations(i, sd); 6063 init_sched_groups_power(i, sd); 6064 } 6065 } 6066 6067 /* Attach the domains */ 6068 rcu_read_lock(); 6069 for_each_cpu(i, cpu_map) { 6070 sd = *per_cpu_ptr(d.sd, i); 6071 cpu_attach_domain(sd, d.rd, i); 6072 } 6073 rcu_read_unlock(); 6074 6075 ret = 0; 6076error: 6077 __free_domain_allocs(&d, alloc_state, cpu_map); 6078 return ret; 6079} 6080 6081static cpumask_var_t *doms_cur; /* current sched domains */ 6082static int ndoms_cur; /* number of sched domains in 'doms_cur' */ 6083static struct sched_domain_attr *dattr_cur; 6084 /* attribues of custom domains in 'doms_cur' */ 6085 6086/* 6087 * Special case: If a kmalloc of a doms_cur partition (array of 6088 * cpumask) fails, then fallback to a single sched domain, 6089 * as determined by the single cpumask fallback_doms. 6090 */ 6091static cpumask_var_t fallback_doms; 6092 6093/* 6094 * arch_update_cpu_topology lets virtualized architectures update the 6095 * cpu core maps. It is supposed to return 1 if the topology changed 6096 * or 0 if it stayed the same. 6097 */ 6098int __attribute__((weak)) arch_update_cpu_topology(void) 6099{ 6100 return 0; 6101} 6102 6103cpumask_var_t *alloc_sched_domains(unsigned int ndoms) 6104{ 6105 int i; 6106 cpumask_var_t *doms; 6107 6108 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL); 6109 if (!doms) 6110 return NULL; 6111 for (i = 0; i < ndoms; i++) { 6112 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) { 6113 free_sched_domains(doms, i); 6114 return NULL; 6115 } 6116 } 6117 return doms; 6118} 6119 6120void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms) 6121{ 6122 unsigned int i; 6123 for (i = 0; i < ndoms; i++) 6124 free_cpumask_var(doms[i]); 6125 kfree(doms); 6126} 6127 6128/* 6129 * Set up scheduler domains and groups. Callers must hold the hotplug lock. 6130 * For now this just excludes isolated cpus, but could be used to 6131 * exclude other special cases in the future. 6132 */ 6133static int init_sched_domains(const struct cpumask *cpu_map) 6134{ 6135 int err; 6136 6137 arch_update_cpu_topology(); 6138 ndoms_cur = 1; 6139 doms_cur = alloc_sched_domains(ndoms_cur); 6140 if (!doms_cur) 6141 doms_cur = &fallback_doms; 6142 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map); 6143 err = build_sched_domains(doms_cur[0], NULL); 6144 register_sched_domain_sysctl(); 6145 6146 return err; 6147} 6148 6149/* 6150 * Detach sched domains from a group of cpus specified in cpu_map 6151 * These cpus will now be attached to the NULL domain 6152 */ 6153static void detach_destroy_domains(const struct cpumask *cpu_map) 6154{ 6155 int i; 6156 6157 rcu_read_lock(); 6158 for_each_cpu(i, cpu_map) 6159 cpu_attach_domain(NULL, &def_root_domain, i); 6160 rcu_read_unlock(); 6161} 6162 6163/* handle null as "default" */ 6164static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur, 6165 struct sched_domain_attr *new, int idx_new) 6166{ 6167 struct sched_domain_attr tmp; 6168 6169 /* fast path */ 6170 if (!new && !cur) 6171 return 1; 6172 6173 tmp = SD_ATTR_INIT; 6174 return !memcmp(cur ? (cur + idx_cur) : &tmp, 6175 new ? (new + idx_new) : &tmp, 6176 sizeof(struct sched_domain_attr)); 6177} 6178 6179/* 6180 * Partition sched domains as specified by the 'ndoms_new' 6181 * cpumasks in the array doms_new[] of cpumasks. This compares 6182 * doms_new[] to the current sched domain partitioning, doms_cur[]. 6183 * It destroys each deleted domain and builds each new domain. 6184 * 6185 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'. 6186 * The masks don't intersect (don't overlap.) We should setup one 6187 * sched domain for each mask. CPUs not in any of the cpumasks will 6188 * not be load balanced. If the same cpumask appears both in the 6189 * current 'doms_cur' domains and in the new 'doms_new', we can leave 6190 * it as it is. 6191 * 6192 * The passed in 'doms_new' should be allocated using 6193 * alloc_sched_domains. This routine takes ownership of it and will 6194 * free_sched_domains it when done with it. If the caller failed the 6195 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1, 6196 * and partition_sched_domains() will fallback to the single partition 6197 * 'fallback_doms', it also forces the domains to be rebuilt. 6198 * 6199 * If doms_new == NULL it will be replaced with cpu_online_mask. 6200 * ndoms_new == 0 is a special case for destroying existing domains, 6201 * and it will not create the default domain. 6202 * 6203 * Call with hotplug lock held 6204 */ 6205void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[], 6206 struct sched_domain_attr *dattr_new) 6207{ 6208 int i, j, n; 6209 int new_topology; 6210 6211 mutex_lock(&sched_domains_mutex); 6212 6213 /* always unregister in case we don't destroy any domains */ 6214 unregister_sched_domain_sysctl(); 6215 6216 /* Let architecture update cpu core mappings. */ 6217 new_topology = arch_update_cpu_topology(); 6218 6219 n = doms_new ? ndoms_new : 0; 6220 6221 /* Destroy deleted domains */ 6222 for (i = 0; i < ndoms_cur; i++) { 6223 for (j = 0; j < n && !new_topology; j++) { 6224 if (cpumask_equal(doms_cur[i], doms_new[j]) 6225 && dattrs_equal(dattr_cur, i, dattr_new, j)) 6226 goto match1; 6227 } 6228 /* no match - a current sched domain not in new doms_new[] */ 6229 detach_destroy_domains(doms_cur[i]); 6230match1: 6231 ; 6232 } 6233 6234 n = ndoms_cur; 6235 if (doms_new == NULL) { 6236 n = 0; 6237 doms_new = &fallback_doms; 6238 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map); 6239 WARN_ON_ONCE(dattr_new); 6240 } 6241 6242 /* Build new domains */ 6243 for (i = 0; i < ndoms_new; i++) { 6244 for (j = 0; j < n && !new_topology; j++) { 6245 if (cpumask_equal(doms_new[i], doms_cur[j]) 6246 && dattrs_equal(dattr_new, i, dattr_cur, j)) 6247 goto match2; 6248 } 6249 /* no match - add a new doms_new */ 6250 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL); 6251match2: 6252 ; 6253 } 6254 6255 /* Remember the new sched domains */ 6256 if (doms_cur != &fallback_doms) 6257 free_sched_domains(doms_cur, ndoms_cur); 6258 kfree(dattr_cur); /* kfree(NULL) is safe */ 6259 doms_cur = doms_new; 6260 dattr_cur = dattr_new; 6261 ndoms_cur = ndoms_new; 6262 6263 register_sched_domain_sysctl(); 6264 6265 mutex_unlock(&sched_domains_mutex); 6266} 6267 6268static int num_cpus_frozen; /* used to mark begin/end of suspend/resume */ 6269 6270/* 6271 * Update cpusets according to cpu_active mask. If cpusets are 6272 * disabled, cpuset_update_active_cpus() becomes a simple wrapper 6273 * around partition_sched_domains(). 6274 * 6275 * If we come here as part of a suspend/resume, don't touch cpusets because we 6276 * want to restore it back to its original state upon resume anyway. 6277 */ 6278static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action, 6279 void *hcpu) 6280{ 6281 switch (action) { 6282 case CPU_ONLINE_FROZEN: 6283 case CPU_DOWN_FAILED_FROZEN: 6284 6285 /* 6286 * num_cpus_frozen tracks how many CPUs are involved in suspend 6287 * resume sequence. As long as this is not the last online 6288 * operation in the resume sequence, just build a single sched 6289 * domain, ignoring cpusets. 6290 */ 6291 num_cpus_frozen--; 6292 if (likely(num_cpus_frozen)) { 6293 partition_sched_domains(1, NULL, NULL); 6294 break; 6295 } 6296 6297 /* 6298 * This is the last CPU online operation. So fall through and 6299 * restore the original sched domains by considering the 6300 * cpuset configurations. 6301 */ 6302 6303 case CPU_ONLINE: 6304 case CPU_DOWN_FAILED: 6305 cpuset_update_active_cpus(true); 6306 break; 6307 default: 6308 return NOTIFY_DONE; 6309 } 6310 return NOTIFY_OK; 6311} 6312 6313static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action, 6314 void *hcpu) 6315{ 6316 switch (action) { 6317 case CPU_DOWN_PREPARE: 6318 cpuset_update_active_cpus(false); 6319 break; 6320 case CPU_DOWN_PREPARE_FROZEN: 6321 num_cpus_frozen++; 6322 partition_sched_domains(1, NULL, NULL); 6323 break; 6324 default: 6325 return NOTIFY_DONE; 6326 } 6327 return NOTIFY_OK; 6328} 6329 6330void __init sched_init_smp(void) 6331{ 6332 cpumask_var_t non_isolated_cpus; 6333 6334 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL); 6335 alloc_cpumask_var(&fallback_doms, GFP_KERNEL); 6336 6337 sched_init_numa(); 6338 6339 get_online_cpus(); 6340 mutex_lock(&sched_domains_mutex); 6341 init_sched_domains(cpu_active_mask); 6342 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map); 6343 if (cpumask_empty(non_isolated_cpus)) 6344 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus); 6345 mutex_unlock(&sched_domains_mutex); 6346 put_online_cpus(); 6347 6348 hotcpu_notifier(sched_domains_numa_masks_update, CPU_PRI_SCHED_ACTIVE); 6349 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE); 6350 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE); 6351 6352 init_hrtick(); 6353 6354 /* Move init over to a non-isolated CPU */ 6355 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0) 6356 BUG(); 6357 sched_init_granularity(); 6358 free_cpumask_var(non_isolated_cpus); 6359 6360 init_sched_rt_class(); 6361} 6362#else 6363void __init sched_init_smp(void) 6364{ 6365 sched_init_granularity(); 6366} 6367#endif /* CONFIG_SMP */ 6368 6369const_debug unsigned int sysctl_timer_migration = 1; 6370 6371int in_sched_functions(unsigned long addr) 6372{ 6373 return in_lock_functions(addr) || 6374 (addr >= (unsigned long)__sched_text_start 6375 && addr < (unsigned long)__sched_text_end); 6376} 6377 6378#ifdef CONFIG_CGROUP_SCHED 6379/* 6380 * Default task group. 6381 * Every task in system belongs to this group at bootup. 6382 */ 6383struct task_group root_task_group; 6384LIST_HEAD(task_groups); 6385#endif 6386 6387DECLARE_PER_CPU(cpumask_var_t, load_balance_mask); 6388 6389void __init sched_init(void) 6390{ 6391 int i, j; 6392 unsigned long alloc_size = 0, ptr; 6393 6394#ifdef CONFIG_FAIR_GROUP_SCHED 6395 alloc_size += 2 * nr_cpu_ids * sizeof(void **); 6396#endif 6397#ifdef CONFIG_RT_GROUP_SCHED 6398 alloc_size += 2 * nr_cpu_ids * sizeof(void **); 6399#endif 6400#ifdef CONFIG_CPUMASK_OFFSTACK 6401 alloc_size += num_possible_cpus() * cpumask_size(); 6402#endif 6403 if (alloc_size) { 6404 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT); 6405 6406#ifdef CONFIG_FAIR_GROUP_SCHED 6407 root_task_group.se = (struct sched_entity **)ptr; 6408 ptr += nr_cpu_ids * sizeof(void **); 6409 6410 root_task_group.cfs_rq = (struct cfs_rq **)ptr; 6411 ptr += nr_cpu_ids * sizeof(void **); 6412 6413#endif /* CONFIG_FAIR_GROUP_SCHED */ 6414#ifdef CONFIG_RT_GROUP_SCHED 6415 root_task_group.rt_se = (struct sched_rt_entity **)ptr; 6416 ptr += nr_cpu_ids * sizeof(void **); 6417 6418 root_task_group.rt_rq = (struct rt_rq **)ptr; 6419 ptr += nr_cpu_ids * sizeof(void **); 6420 6421#endif /* CONFIG_RT_GROUP_SCHED */ 6422#ifdef CONFIG_CPUMASK_OFFSTACK 6423 for_each_possible_cpu(i) { 6424 per_cpu(load_balance_mask, i) = (void *)ptr; 6425 ptr += cpumask_size(); 6426 } 6427#endif /* CONFIG_CPUMASK_OFFSTACK */ 6428 } 6429 6430#ifdef CONFIG_SMP 6431 init_defrootdomain(); 6432#endif 6433 6434 init_rt_bandwidth(&def_rt_bandwidth, 6435 global_rt_period(), global_rt_runtime()); 6436 6437#ifdef CONFIG_RT_GROUP_SCHED 6438 init_rt_bandwidth(&root_task_group.rt_bandwidth, 6439 global_rt_period(), global_rt_runtime()); 6440#endif /* CONFIG_RT_GROUP_SCHED */ 6441 6442#ifdef CONFIG_CGROUP_SCHED 6443 list_add(&root_task_group.list, &task_groups); 6444 INIT_LIST_HEAD(&root_task_group.children); 6445 INIT_LIST_HEAD(&root_task_group.siblings); 6446 autogroup_init(&init_task); 6447 6448#endif /* CONFIG_CGROUP_SCHED */ 6449 6450 for_each_possible_cpu(i) { 6451 struct rq *rq; 6452 6453 rq = cpu_rq(i); 6454 raw_spin_lock_init(&rq->lock); 6455 rq->nr_running = 0; 6456 rq->calc_load_active = 0; 6457 rq->calc_load_update = jiffies + LOAD_FREQ; 6458 init_cfs_rq(&rq->cfs); 6459 init_rt_rq(&rq->rt, rq); 6460#ifdef CONFIG_FAIR_GROUP_SCHED 6461 root_task_group.shares = ROOT_TASK_GROUP_LOAD; 6462 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list); 6463 /* 6464 * How much cpu bandwidth does root_task_group get? 6465 * 6466 * In case of task-groups formed thr' the cgroup filesystem, it 6467 * gets 100% of the cpu resources in the system. This overall 6468 * system cpu resource is divided among the tasks of 6469 * root_task_group and its child task-groups in a fair manner, 6470 * based on each entity's (task or task-group's) weight 6471 * (se->load.weight). 6472 * 6473 * In other words, if root_task_group has 10 tasks of weight 6474 * 1024) and two child groups A0 and A1 (of weight 1024 each), 6475 * then A0's share of the cpu resource is: 6476 * 6477 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33% 6478 * 6479 * We achieve this by letting root_task_group's tasks sit 6480 * directly in rq->cfs (i.e root_task_group->se[] = NULL). 6481 */ 6482 init_cfs_bandwidth(&root_task_group.cfs_bandwidth); 6483 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL); 6484#endif /* CONFIG_FAIR_GROUP_SCHED */ 6485 6486 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime; 6487#ifdef CONFIG_RT_GROUP_SCHED 6488 INIT_LIST_HEAD(&rq->leaf_rt_rq_list); 6489 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL); 6490#endif 6491 6492 for (j = 0; j < CPU_LOAD_IDX_MAX; j++) 6493 rq->cpu_load[j] = 0; 6494 6495 rq->last_load_update_tick = jiffies; 6496 6497#ifdef CONFIG_SMP 6498 rq->sd = NULL; 6499 rq->rd = NULL; 6500 rq->cpu_power = SCHED_POWER_SCALE; 6501 rq->post_schedule = 0; 6502 rq->active_balance = 0; 6503 rq->next_balance = jiffies; 6504 rq->push_cpu = 0; 6505 rq->cpu = i; 6506 rq->online = 0; 6507 rq->idle_stamp = 0; 6508 rq->avg_idle = 2*sysctl_sched_migration_cost; 6509 rq->max_idle_balance_cost = sysctl_sched_migration_cost; 6510 6511 INIT_LIST_HEAD(&rq->cfs_tasks); 6512 6513 rq_attach_root(rq, &def_root_domain); 6514#ifdef CONFIG_NO_HZ_COMMON 6515 rq->nohz_flags = 0; 6516#endif 6517#ifdef CONFIG_NO_HZ_FULL 6518 rq->last_sched_tick = 0; 6519#endif 6520#endif 6521 init_rq_hrtick(rq); 6522 atomic_set(&rq->nr_iowait, 0); 6523 } 6524 6525 set_load_weight(&init_task); 6526 6527#ifdef CONFIG_PREEMPT_NOTIFIERS 6528 INIT_HLIST_HEAD(&init_task.preempt_notifiers); 6529#endif 6530 6531#ifdef CONFIG_RT_MUTEXES 6532 plist_head_init(&init_task.pi_waiters); 6533#endif 6534 6535 /* 6536 * The boot idle thread does lazy MMU switching as well: 6537 */ 6538 atomic_inc(&init_mm.mm_count); 6539 enter_lazy_tlb(&init_mm, current); 6540 6541 /* 6542 * Make us the idle thread. Technically, schedule() should not be 6543 * called from this thread, however somewhere below it might be, 6544 * but because we are the idle thread, we just pick up running again 6545 * when this runqueue becomes "idle". 6546 */ 6547 init_idle(current, smp_processor_id()); 6548 6549 calc_load_update = jiffies + LOAD_FREQ; 6550 6551 /* 6552 * During early bootup we pretend to be a normal task: 6553 */ 6554 current->sched_class = &fair_sched_class; 6555 6556#ifdef CONFIG_SMP 6557 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT); 6558 /* May be allocated at isolcpus cmdline parse time */ 6559 if (cpu_isolated_map == NULL) 6560 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT); 6561 idle_thread_set_boot_cpu(); 6562#endif 6563 init_sched_fair_class(); 6564 6565 scheduler_running = 1; 6566} 6567 6568#ifdef CONFIG_DEBUG_ATOMIC_SLEEP 6569static inline int preempt_count_equals(int preempt_offset) 6570{ 6571 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth(); 6572 6573 return (nested == preempt_offset); 6574} 6575 6576void __might_sleep(const char *file, int line, int preempt_offset) 6577{ 6578 static unsigned long prev_jiffy; /* ratelimiting */ 6579 6580 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */ 6581 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) || 6582 system_state != SYSTEM_RUNNING || oops_in_progress) 6583 return; 6584 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy) 6585 return; 6586 prev_jiffy = jiffies; 6587 6588 printk(KERN_ERR 6589 "BUG: sleeping function called from invalid context at %s:%d\n", 6590 file, line); 6591 printk(KERN_ERR 6592 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n", 6593 in_atomic(), irqs_disabled(), 6594 current->pid, current->comm); 6595 6596 debug_show_held_locks(current); 6597 if (irqs_disabled()) 6598 print_irqtrace_events(current); 6599 dump_stack(); 6600} 6601EXPORT_SYMBOL(__might_sleep); 6602#endif 6603 6604#ifdef CONFIG_MAGIC_SYSRQ 6605static void normalize_task(struct rq *rq, struct task_struct *p) 6606{ 6607 const struct sched_class *prev_class = p->sched_class; 6608 int old_prio = p->prio; 6609 int on_rq; 6610 6611 on_rq = p->on_rq; 6612 if (on_rq) 6613 dequeue_task(rq, p, 0); 6614 __setscheduler(rq, p, SCHED_NORMAL, 0); 6615 if (on_rq) { 6616 enqueue_task(rq, p, 0); 6617 resched_task(rq->curr); 6618 } 6619 6620 check_class_changed(rq, p, prev_class, old_prio); 6621} 6622 6623void normalize_rt_tasks(void) 6624{ 6625 struct task_struct *g, *p; 6626 unsigned long flags; 6627 struct rq *rq; 6628 6629 read_lock_irqsave(&tasklist_lock, flags); 6630 do_each_thread(g, p) { 6631 /* 6632 * Only normalize user tasks: 6633 */ 6634 if (!p->mm) 6635 continue; 6636 6637 p->se.exec_start = 0; 6638#ifdef CONFIG_SCHEDSTATS 6639 p->se.statistics.wait_start = 0; 6640 p->se.statistics.sleep_start = 0; 6641 p->se.statistics.block_start = 0; 6642#endif 6643 6644 if (!rt_task(p)) { 6645 /* 6646 * Renice negative nice level userspace 6647 * tasks back to 0: 6648 */ 6649 if (TASK_NICE(p) < 0 && p->mm) 6650 set_user_nice(p, 0); 6651 continue; 6652 } 6653 6654 raw_spin_lock(&p->pi_lock); 6655 rq = __task_rq_lock(p); 6656 6657 normalize_task(rq, p); 6658 6659 __task_rq_unlock(rq); 6660 raw_spin_unlock(&p->pi_lock); 6661 } while_each_thread(g, p); 6662 6663 read_unlock_irqrestore(&tasklist_lock, flags); 6664} 6665 6666#endif /* CONFIG_MAGIC_SYSRQ */ 6667 6668#if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) 6669/* 6670 * These functions are only useful for the IA64 MCA handling, or kdb. 6671 * 6672 * They can only be called when the whole system has been 6673 * stopped - every CPU needs to be quiescent, and no scheduling 6674 * activity can take place. Using them for anything else would 6675 * be a serious bug, and as a result, they aren't even visible 6676 * under any other configuration. 6677 */ 6678 6679/** 6680 * curr_task - return the current task for a given cpu. 6681 * @cpu: the processor in question. 6682 * 6683 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED! 6684 * 6685 * Return: The current task for @cpu. 6686 */ 6687struct task_struct *curr_task(int cpu) 6688{ 6689 return cpu_curr(cpu); 6690} 6691 6692#endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */ 6693 6694#ifdef CONFIG_IA64 6695/** 6696 * set_curr_task - set the current task for a given cpu. 6697 * @cpu: the processor in question. 6698 * @p: the task pointer to set. 6699 * 6700 * Description: This function must only be used when non-maskable interrupts 6701 * are serviced on a separate stack. It allows the architecture to switch the 6702 * notion of the current task on a cpu in a non-blocking manner. This function 6703 * must be called with all CPU's synchronized, and interrupts disabled, the 6704 * and caller must save the original value of the current task (see 6705 * curr_task() above) and restore that value before reenabling interrupts and 6706 * re-starting the system. 6707 * 6708 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED! 6709 */ 6710void set_curr_task(int cpu, struct task_struct *p) 6711{ 6712 cpu_curr(cpu) = p; 6713} 6714 6715#endif 6716 6717#ifdef CONFIG_CGROUP_SCHED 6718/* task_group_lock serializes the addition/removal of task groups */ 6719static DEFINE_SPINLOCK(task_group_lock); 6720 6721static void free_sched_group(struct task_group *tg) 6722{ 6723 free_fair_sched_group(tg); 6724 free_rt_sched_group(tg); 6725 autogroup_free(tg); 6726 kfree(tg); 6727} 6728 6729/* allocate runqueue etc for a new task group */ 6730struct task_group *sched_create_group(struct task_group *parent) 6731{ 6732 struct task_group *tg; 6733 6734 tg = kzalloc(sizeof(*tg), GFP_KERNEL); 6735 if (!tg) 6736 return ERR_PTR(-ENOMEM); 6737 6738 if (!alloc_fair_sched_group(tg, parent)) 6739 goto err; 6740 6741 if (!alloc_rt_sched_group(tg, parent)) 6742 goto err; 6743 6744 return tg; 6745 6746err: 6747 free_sched_group(tg); 6748 return ERR_PTR(-ENOMEM); 6749} 6750 6751void sched_online_group(struct task_group *tg, struct task_group *parent) 6752{ 6753 unsigned long flags; 6754 6755 spin_lock_irqsave(&task_group_lock, flags); 6756 list_add_rcu(&tg->list, &task_groups); 6757 6758 WARN_ON(!parent); /* root should already exist */ 6759 6760 tg->parent = parent; 6761 INIT_LIST_HEAD(&tg->children); 6762 list_add_rcu(&tg->siblings, &parent->children); 6763 spin_unlock_irqrestore(&task_group_lock, flags); 6764} 6765 6766/* rcu callback to free various structures associated with a task group */ 6767static void free_sched_group_rcu(struct rcu_head *rhp) 6768{ 6769 /* now it should be safe to free those cfs_rqs */ 6770 free_sched_group(container_of(rhp, struct task_group, rcu)); 6771} 6772 6773/* Destroy runqueue etc associated with a task group */ 6774void sched_destroy_group(struct task_group *tg) 6775{ 6776 /* wait for possible concurrent references to cfs_rqs complete */ 6777 call_rcu(&tg->rcu, free_sched_group_rcu); 6778} 6779 6780void sched_offline_group(struct task_group *tg) 6781{ 6782 unsigned long flags; 6783 int i; 6784 6785 /* end participation in shares distribution */ 6786 for_each_possible_cpu(i) 6787 unregister_fair_sched_group(tg, i); 6788 6789 spin_lock_irqsave(&task_group_lock, flags); 6790 list_del_rcu(&tg->list); 6791 list_del_rcu(&tg->siblings); 6792 spin_unlock_irqrestore(&task_group_lock, flags); 6793} 6794 6795/* change task's runqueue when it moves between groups. 6796 * The caller of this function should have put the task in its new group 6797 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to 6798 * reflect its new group. 6799 */ 6800void sched_move_task(struct task_struct *tsk) 6801{ 6802 struct task_group *tg; 6803 int on_rq, running; 6804 unsigned long flags; 6805 struct rq *rq; 6806 6807 rq = task_rq_lock(tsk, &flags); 6808 6809 running = task_current(rq, tsk); 6810 on_rq = tsk->on_rq; 6811 6812 if (on_rq) 6813 dequeue_task(rq, tsk, 0); 6814 if (unlikely(running)) 6815 tsk->sched_class->put_prev_task(rq, tsk); 6816 6817 tg = container_of(task_css_check(tsk, cpu_cgroup_subsys_id, 6818 lockdep_is_held(&tsk->sighand->siglock)), 6819 struct task_group, css); 6820 tg = autogroup_task_group(tsk, tg); 6821 tsk->sched_task_group = tg; 6822 6823#ifdef CONFIG_FAIR_GROUP_SCHED 6824 if (tsk->sched_class->task_move_group) 6825 tsk->sched_class->task_move_group(tsk, on_rq); 6826 else 6827#endif 6828 set_task_rq(tsk, task_cpu(tsk)); 6829 6830 if (unlikely(running)) 6831 tsk->sched_class->set_curr_task(rq); 6832 if (on_rq) 6833 enqueue_task(rq, tsk, 0); 6834 6835 task_rq_unlock(rq, tsk, &flags); 6836} 6837#endif /* CONFIG_CGROUP_SCHED */ 6838 6839#if defined(CONFIG_RT_GROUP_SCHED) || defined(CONFIG_CFS_BANDWIDTH) 6840static unsigned long to_ratio(u64 period, u64 runtime) 6841{ 6842 if (runtime == RUNTIME_INF) 6843 return 1ULL << 20; 6844 6845 return div64_u64(runtime << 20, period); 6846} 6847#endif 6848 6849#ifdef CONFIG_RT_GROUP_SCHED 6850/* 6851 * Ensure that the real time constraints are schedulable. 6852 */ 6853static DEFINE_MUTEX(rt_constraints_mutex); 6854 6855/* Must be called with tasklist_lock held */ 6856static inline int tg_has_rt_tasks(struct task_group *tg) 6857{ 6858 struct task_struct *g, *p; 6859 6860 do_each_thread(g, p) { 6861 if (rt_task(p) && task_rq(p)->rt.tg == tg) 6862 return 1; 6863 } while_each_thread(g, p); 6864 6865 return 0; 6866} 6867 6868struct rt_schedulable_data { 6869 struct task_group *tg; 6870 u64 rt_period; 6871 u64 rt_runtime; 6872}; 6873 6874static int tg_rt_schedulable(struct task_group *tg, void *data) 6875{ 6876 struct rt_schedulable_data *d = data; 6877 struct task_group *child; 6878 unsigned long total, sum = 0; 6879 u64 period, runtime; 6880 6881 period = ktime_to_ns(tg->rt_bandwidth.rt_period); 6882 runtime = tg->rt_bandwidth.rt_runtime; 6883 6884 if (tg == d->tg) { 6885 period = d->rt_period; 6886 runtime = d->rt_runtime; 6887 } 6888 6889 /* 6890 * Cannot have more runtime than the period. 6891 */ 6892 if (runtime > period && runtime != RUNTIME_INF) 6893 return -EINVAL; 6894 6895 /* 6896 * Ensure we don't starve existing RT tasks. 6897 */ 6898 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg)) 6899 return -EBUSY; 6900 6901 total = to_ratio(period, runtime); 6902 6903 /* 6904 * Nobody can have more than the global setting allows. 6905 */ 6906 if (total > to_ratio(global_rt_period(), global_rt_runtime())) 6907 return -EINVAL; 6908 6909 /* 6910 * The sum of our children's runtime should not exceed our own. 6911 */ 6912 list_for_each_entry_rcu(child, &tg->children, siblings) { 6913 period = ktime_to_ns(child->rt_bandwidth.rt_period); 6914 runtime = child->rt_bandwidth.rt_runtime; 6915 6916 if (child == d->tg) { 6917 period = d->rt_period; 6918 runtime = d->rt_runtime; 6919 } 6920 6921 sum += to_ratio(period, runtime); 6922 } 6923 6924 if (sum > total) 6925 return -EINVAL; 6926 6927 return 0; 6928} 6929 6930static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime) 6931{ 6932 int ret; 6933 6934 struct rt_schedulable_data data = { 6935 .tg = tg, 6936 .rt_period = period, 6937 .rt_runtime = runtime, 6938 }; 6939 6940 rcu_read_lock(); 6941 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data); 6942 rcu_read_unlock(); 6943 6944 return ret; 6945} 6946 6947static int tg_set_rt_bandwidth(struct task_group *tg, 6948 u64 rt_period, u64 rt_runtime) 6949{ 6950 int i, err = 0; 6951 6952 mutex_lock(&rt_constraints_mutex); 6953 read_lock(&tasklist_lock); 6954 err = __rt_schedulable(tg, rt_period, rt_runtime); 6955 if (err) 6956 goto unlock; 6957 6958 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock); 6959 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period); 6960 tg->rt_bandwidth.rt_runtime = rt_runtime; 6961 6962 for_each_possible_cpu(i) { 6963 struct rt_rq *rt_rq = tg->rt_rq[i]; 6964 6965 raw_spin_lock(&rt_rq->rt_runtime_lock); 6966 rt_rq->rt_runtime = rt_runtime; 6967 raw_spin_unlock(&rt_rq->rt_runtime_lock); 6968 } 6969 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock); 6970unlock: 6971 read_unlock(&tasklist_lock); 6972 mutex_unlock(&rt_constraints_mutex); 6973 6974 return err; 6975} 6976 6977static int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us) 6978{ 6979 u64 rt_runtime, rt_period; 6980 6981 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period); 6982 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC; 6983 if (rt_runtime_us < 0) 6984 rt_runtime = RUNTIME_INF; 6985 6986 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime); 6987} 6988 6989static long sched_group_rt_runtime(struct task_group *tg) 6990{ 6991 u64 rt_runtime_us; 6992 6993 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF) 6994 return -1; 6995 6996 rt_runtime_us = tg->rt_bandwidth.rt_runtime; 6997 do_div(rt_runtime_us, NSEC_PER_USEC); 6998 return rt_runtime_us; 6999} 7000 7001static int sched_group_set_rt_period(struct task_group *tg, long rt_period_us) 7002{ 7003 u64 rt_runtime, rt_period; 7004 7005 rt_period = (u64)rt_period_us * NSEC_PER_USEC; 7006 rt_runtime = tg->rt_bandwidth.rt_runtime; 7007 7008 if (rt_period == 0) 7009 return -EINVAL; 7010 7011 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime); 7012} 7013 7014static long sched_group_rt_period(struct task_group *tg) 7015{ 7016 u64 rt_period_us; 7017 7018 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period); 7019 do_div(rt_period_us, NSEC_PER_USEC); 7020 return rt_period_us; 7021} 7022 7023static int sched_rt_global_constraints(void) 7024{ 7025 u64 runtime, period; 7026 int ret = 0; 7027 7028 if (sysctl_sched_rt_period <= 0) 7029 return -EINVAL; 7030 7031 runtime = global_rt_runtime(); 7032 period = global_rt_period(); 7033 7034 /* 7035 * Sanity check on the sysctl variables. 7036 */ 7037 if (runtime > period && runtime != RUNTIME_INF) 7038 return -EINVAL; 7039 7040 mutex_lock(&rt_constraints_mutex); 7041 read_lock(&tasklist_lock); 7042 ret = __rt_schedulable(NULL, 0, 0); 7043 read_unlock(&tasklist_lock); 7044 mutex_unlock(&rt_constraints_mutex); 7045 7046 return ret; 7047} 7048 7049static int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk) 7050{ 7051 /* Don't accept realtime tasks when there is no way for them to run */ 7052 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0) 7053 return 0; 7054 7055 return 1; 7056} 7057 7058#else /* !CONFIG_RT_GROUP_SCHED */ 7059static int sched_rt_global_constraints(void) 7060{ 7061 unsigned long flags; 7062 int i; 7063 7064 if (sysctl_sched_rt_period <= 0) 7065 return -EINVAL; 7066 7067 /* 7068 * There's always some RT tasks in the root group 7069 * -- migration, kstopmachine etc.. 7070 */ 7071 if (sysctl_sched_rt_runtime == 0) 7072 return -EBUSY; 7073 7074 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags); 7075 for_each_possible_cpu(i) { 7076 struct rt_rq *rt_rq = &cpu_rq(i)->rt; 7077 7078 raw_spin_lock(&rt_rq->rt_runtime_lock); 7079 rt_rq->rt_runtime = global_rt_runtime(); 7080 raw_spin_unlock(&rt_rq->rt_runtime_lock); 7081 } 7082 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags); 7083 7084 return 0; 7085} 7086#endif /* CONFIG_RT_GROUP_SCHED */ 7087 7088int sched_rr_handler(struct ctl_table *table, int write, 7089 void __user *buffer, size_t *lenp, 7090 loff_t *ppos) 7091{ 7092 int ret; 7093 static DEFINE_MUTEX(mutex); 7094 7095 mutex_lock(&mutex); 7096 ret = proc_dointvec(table, write, buffer, lenp, ppos); 7097 /* make sure that internally we keep jiffies */ 7098 /* also, writing zero resets timeslice to default */ 7099 if (!ret && write) { 7100 sched_rr_timeslice = sched_rr_timeslice <= 0 ? 7101 RR_TIMESLICE : msecs_to_jiffies(sched_rr_timeslice); 7102 } 7103 mutex_unlock(&mutex); 7104 return ret; 7105} 7106 7107int sched_rt_handler(struct ctl_table *table, int write, 7108 void __user *buffer, size_t *lenp, 7109 loff_t *ppos) 7110{ 7111 int ret; 7112 int old_period, old_runtime; 7113 static DEFINE_MUTEX(mutex); 7114 7115 mutex_lock(&mutex); 7116 old_period = sysctl_sched_rt_period; 7117 old_runtime = sysctl_sched_rt_runtime; 7118 7119 ret = proc_dointvec(table, write, buffer, lenp, ppos); 7120 7121 if (!ret && write) { 7122 ret = sched_rt_global_constraints(); 7123 if (ret) { 7124 sysctl_sched_rt_period = old_period; 7125 sysctl_sched_rt_runtime = old_runtime; 7126 } else { 7127 def_rt_bandwidth.rt_runtime = global_rt_runtime(); 7128 def_rt_bandwidth.rt_period = 7129 ns_to_ktime(global_rt_period()); 7130 } 7131 } 7132 mutex_unlock(&mutex); 7133 7134 return ret; 7135} 7136 7137#ifdef CONFIG_CGROUP_SCHED 7138 7139static inline struct task_group *css_tg(struct cgroup_subsys_state *css) 7140{ 7141 return css ? container_of(css, struct task_group, css) : NULL; 7142} 7143 7144static struct cgroup_subsys_state * 7145cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css) 7146{ 7147 struct task_group *parent = css_tg(parent_css); 7148 struct task_group *tg; 7149 7150 if (!parent) { 7151 /* This is early initialization for the top cgroup */ 7152 return &root_task_group.css; 7153 } 7154 7155 tg = sched_create_group(parent); 7156 if (IS_ERR(tg)) 7157 return ERR_PTR(-ENOMEM); 7158 7159 return &tg->css; 7160} 7161 7162static int cpu_cgroup_css_online(struct cgroup_subsys_state *css) 7163{ 7164 struct task_group *tg = css_tg(css); 7165 struct task_group *parent = css_tg(css_parent(css)); 7166 7167 if (parent) 7168 sched_online_group(tg, parent); 7169 return 0; 7170} 7171 7172static void cpu_cgroup_css_free(struct cgroup_subsys_state *css) 7173{ 7174 struct task_group *tg = css_tg(css); 7175 7176 sched_destroy_group(tg); 7177} 7178 7179static void cpu_cgroup_css_offline(struct cgroup_subsys_state *css) 7180{ 7181 struct task_group *tg = css_tg(css); 7182 7183 sched_offline_group(tg); 7184} 7185 7186static int cpu_cgroup_can_attach(struct cgroup_subsys_state *css, 7187 struct cgroup_taskset *tset) 7188{ 7189 struct task_struct *task; 7190 7191 cgroup_taskset_for_each(task, css, tset) { 7192#ifdef CONFIG_RT_GROUP_SCHED 7193 if (!sched_rt_can_attach(css_tg(css), task)) 7194 return -EINVAL; 7195#else 7196 /* We don't support RT-tasks being in separate groups */ 7197 if (task->sched_class != &fair_sched_class) 7198 return -EINVAL; 7199#endif 7200 } 7201 return 0; 7202} 7203 7204static void cpu_cgroup_attach(struct cgroup_subsys_state *css, 7205 struct cgroup_taskset *tset) 7206{ 7207 struct task_struct *task; 7208 7209 cgroup_taskset_for_each(task, css, tset) 7210 sched_move_task(task); 7211} 7212 7213static void cpu_cgroup_exit(struct cgroup_subsys_state *css, 7214 struct cgroup_subsys_state *old_css, 7215 struct task_struct *task) 7216{ 7217 /* 7218 * cgroup_exit() is called in the copy_process() failure path. 7219 * Ignore this case since the task hasn't ran yet, this avoids 7220 * trying to poke a half freed task state from generic code. 7221 */ 7222 if (!(task->flags & PF_EXITING)) 7223 return; 7224 7225 sched_move_task(task); 7226} 7227 7228#ifdef CONFIG_FAIR_GROUP_SCHED 7229static int cpu_shares_write_u64(struct cgroup_subsys_state *css, 7230 struct cftype *cftype, u64 shareval) 7231{ 7232 return sched_group_set_shares(css_tg(css), scale_load(shareval)); 7233} 7234 7235static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css, 7236 struct cftype *cft) 7237{ 7238 struct task_group *tg = css_tg(css); 7239 7240 return (u64) scale_load_down(tg->shares); 7241} 7242 7243#ifdef CONFIG_CFS_BANDWIDTH 7244static DEFINE_MUTEX(cfs_constraints_mutex); 7245 7246const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */ 7247const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */ 7248 7249static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime); 7250 7251static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota) 7252{ 7253 int i, ret = 0, runtime_enabled, runtime_was_enabled; 7254 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth; 7255 7256 if (tg == &root_task_group) 7257 return -EINVAL; 7258 7259 /* 7260 * Ensure we have at some amount of bandwidth every period. This is 7261 * to prevent reaching a state of large arrears when throttled via 7262 * entity_tick() resulting in prolonged exit starvation. 7263 */ 7264 if (quota < min_cfs_quota_period || period < min_cfs_quota_period) 7265 return -EINVAL; 7266 7267 /* 7268 * Likewise, bound things on the otherside by preventing insane quota 7269 * periods. This also allows us to normalize in computing quota 7270 * feasibility. 7271 */ 7272 if (period > max_cfs_quota_period) 7273 return -EINVAL; 7274 7275 mutex_lock(&cfs_constraints_mutex); 7276 ret = __cfs_schedulable(tg, period, quota); 7277 if (ret) 7278 goto out_unlock; 7279 7280 runtime_enabled = quota != RUNTIME_INF; 7281 runtime_was_enabled = cfs_b->quota != RUNTIME_INF; 7282 account_cfs_bandwidth_used(runtime_enabled, runtime_was_enabled); 7283 raw_spin_lock_irq(&cfs_b->lock); 7284 cfs_b->period = ns_to_ktime(period); 7285 cfs_b->quota = quota; 7286 7287 __refill_cfs_bandwidth_runtime(cfs_b); 7288 /* restart the period timer (if active) to handle new period expiry */ 7289 if (runtime_enabled && cfs_b->timer_active) { 7290 /* force a reprogram */ 7291 cfs_b->timer_active = 0; 7292 __start_cfs_bandwidth(cfs_b); 7293 } 7294 raw_spin_unlock_irq(&cfs_b->lock); 7295 7296 for_each_possible_cpu(i) { 7297 struct cfs_rq *cfs_rq = tg->cfs_rq[i]; 7298 struct rq *rq = cfs_rq->rq; 7299 7300 raw_spin_lock_irq(&rq->lock); 7301 cfs_rq->runtime_enabled = runtime_enabled; 7302 cfs_rq->runtime_remaining = 0; 7303 7304 if (cfs_rq->throttled) 7305 unthrottle_cfs_rq(cfs_rq); 7306 raw_spin_unlock_irq(&rq->lock); 7307 } 7308out_unlock: 7309 mutex_unlock(&cfs_constraints_mutex); 7310 7311 return ret; 7312} 7313 7314int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us) 7315{ 7316 u64 quota, period; 7317 7318 period = ktime_to_ns(tg->cfs_bandwidth.period); 7319 if (cfs_quota_us < 0) 7320 quota = RUNTIME_INF; 7321 else 7322 quota = (u64)cfs_quota_us * NSEC_PER_USEC; 7323 7324 return tg_set_cfs_bandwidth(tg, period, quota); 7325} 7326 7327long tg_get_cfs_quota(struct task_group *tg) 7328{ 7329 u64 quota_us; 7330 7331 if (tg->cfs_bandwidth.quota == RUNTIME_INF) 7332 return -1; 7333 7334 quota_us = tg->cfs_bandwidth.quota; 7335 do_div(quota_us, NSEC_PER_USEC); 7336 7337 return quota_us; 7338} 7339 7340int tg_set_cfs_period(struct task_group *tg, long cfs_period_us) 7341{ 7342 u64 quota, period; 7343 7344 period = (u64)cfs_period_us * NSEC_PER_USEC; 7345 quota = tg->cfs_bandwidth.quota; 7346 7347 return tg_set_cfs_bandwidth(tg, period, quota); 7348} 7349 7350long tg_get_cfs_period(struct task_group *tg) 7351{ 7352 u64 cfs_period_us; 7353 7354 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period); 7355 do_div(cfs_period_us, NSEC_PER_USEC); 7356 7357 return cfs_period_us; 7358} 7359 7360static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css, 7361 struct cftype *cft) 7362{ 7363 return tg_get_cfs_quota(css_tg(css)); 7364} 7365 7366static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css, 7367 struct cftype *cftype, s64 cfs_quota_us) 7368{ 7369 return tg_set_cfs_quota(css_tg(css), cfs_quota_us); 7370} 7371 7372static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css, 7373 struct cftype *cft) 7374{ 7375 return tg_get_cfs_period(css_tg(css)); 7376} 7377 7378static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css, 7379 struct cftype *cftype, u64 cfs_period_us) 7380{ 7381 return tg_set_cfs_period(css_tg(css), cfs_period_us); 7382} 7383 7384struct cfs_schedulable_data { 7385 struct task_group *tg; 7386 u64 period, quota; 7387}; 7388 7389/* 7390 * normalize group quota/period to be quota/max_period 7391 * note: units are usecs 7392 */ 7393static u64 normalize_cfs_quota(struct task_group *tg, 7394 struct cfs_schedulable_data *d) 7395{ 7396 u64 quota, period; 7397 7398 if (tg == d->tg) { 7399 period = d->period; 7400 quota = d->quota; 7401 } else { 7402 period = tg_get_cfs_period(tg); 7403 quota = tg_get_cfs_quota(tg); 7404 } 7405 7406 /* note: these should typically be equivalent */ 7407 if (quota == RUNTIME_INF || quota == -1) 7408 return RUNTIME_INF; 7409 7410 return to_ratio(period, quota); 7411} 7412 7413static int tg_cfs_schedulable_down(struct task_group *tg, void *data) 7414{ 7415 struct cfs_schedulable_data *d = data; 7416 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth; 7417 s64 quota = 0, parent_quota = -1; 7418 7419 if (!tg->parent) { 7420 quota = RUNTIME_INF; 7421 } else { 7422 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth; 7423 7424 quota = normalize_cfs_quota(tg, d); 7425 parent_quota = parent_b->hierarchal_quota; 7426 7427 /* 7428 * ensure max(child_quota) <= parent_quota, inherit when no 7429 * limit is set 7430 */ 7431 if (quota == RUNTIME_INF) 7432 quota = parent_quota; 7433 else if (parent_quota != RUNTIME_INF && quota > parent_quota) 7434 return -EINVAL; 7435 } 7436 cfs_b->hierarchal_quota = quota; 7437 7438 return 0; 7439} 7440 7441static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota) 7442{ 7443 int ret; 7444 struct cfs_schedulable_data data = { 7445 .tg = tg, 7446 .period = period, 7447 .quota = quota, 7448 }; 7449 7450 if (quota != RUNTIME_INF) { 7451 do_div(data.period, NSEC_PER_USEC); 7452 do_div(data.quota, NSEC_PER_USEC); 7453 } 7454 7455 rcu_read_lock(); 7456 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data); 7457 rcu_read_unlock(); 7458 7459 return ret; 7460} 7461 7462static int cpu_stats_show(struct cgroup_subsys_state *css, struct cftype *cft, 7463 struct cgroup_map_cb *cb) 7464{ 7465 struct task_group *tg = css_tg(css); 7466 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth; 7467 7468 cb->fill(cb, "nr_periods", cfs_b->nr_periods); 7469 cb->fill(cb, "nr_throttled", cfs_b->nr_throttled); 7470 cb->fill(cb, "throttled_time", cfs_b->throttled_time); 7471 7472 return 0; 7473} 7474#endif /* CONFIG_CFS_BANDWIDTH */ 7475#endif /* CONFIG_FAIR_GROUP_SCHED */ 7476 7477#ifdef CONFIG_RT_GROUP_SCHED 7478static int cpu_rt_runtime_write(struct cgroup_subsys_state *css, 7479 struct cftype *cft, s64 val) 7480{ 7481 return sched_group_set_rt_runtime(css_tg(css), val); 7482} 7483 7484static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css, 7485 struct cftype *cft) 7486{ 7487 return sched_group_rt_runtime(css_tg(css)); 7488} 7489 7490static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css, 7491 struct cftype *cftype, u64 rt_period_us) 7492{ 7493 return sched_group_set_rt_period(css_tg(css), rt_period_us); 7494} 7495 7496static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css, 7497 struct cftype *cft) 7498{ 7499 return sched_group_rt_period(css_tg(css)); 7500} 7501#endif /* CONFIG_RT_GROUP_SCHED */ 7502 7503static struct cftype cpu_files[] = { 7504#ifdef CONFIG_FAIR_GROUP_SCHED 7505 { 7506 .name = "shares", 7507 .read_u64 = cpu_shares_read_u64, 7508 .write_u64 = cpu_shares_write_u64, 7509 }, 7510#endif 7511#ifdef CONFIG_CFS_BANDWIDTH 7512 { 7513 .name = "cfs_quota_us", 7514 .read_s64 = cpu_cfs_quota_read_s64, 7515 .write_s64 = cpu_cfs_quota_write_s64, 7516 }, 7517 { 7518 .name = "cfs_period_us", 7519 .read_u64 = cpu_cfs_period_read_u64, 7520 .write_u64 = cpu_cfs_period_write_u64, 7521 }, 7522 { 7523 .name = "stat", 7524 .read_map = cpu_stats_show, 7525 }, 7526#endif 7527#ifdef CONFIG_RT_GROUP_SCHED 7528 { 7529 .name = "rt_runtime_us", 7530 .read_s64 = cpu_rt_runtime_read, 7531 .write_s64 = cpu_rt_runtime_write, 7532 }, 7533 { 7534 .name = "rt_period_us", 7535 .read_u64 = cpu_rt_period_read_uint, 7536 .write_u64 = cpu_rt_period_write_uint, 7537 }, 7538#endif 7539 { } /* terminate */ 7540}; 7541 7542struct cgroup_subsys cpu_cgroup_subsys = { 7543 .name = "cpu", 7544 .css_alloc = cpu_cgroup_css_alloc, 7545 .css_free = cpu_cgroup_css_free, 7546 .css_online = cpu_cgroup_css_online, 7547 .css_offline = cpu_cgroup_css_offline, 7548 .can_attach = cpu_cgroup_can_attach, 7549 .attach = cpu_cgroup_attach, 7550 .exit = cpu_cgroup_exit, 7551 .subsys_id = cpu_cgroup_subsys_id, 7552 .base_cftypes = cpu_files, 7553 .early_init = 1, 7554}; 7555 7556#endif /* CONFIG_CGROUP_SCHED */ 7557 7558void dump_cpu_task(int cpu) 7559{ 7560 pr_info("Task dump for CPU %d:\n", cpu); 7561 sched_show_task(cpu_curr(cpu)); 7562} 7563