core.c revision 7e8d16b6cbccb2f5da579f5085479fb82ba851b8
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_work.next = &p->numa_work; 1637#endif /* CONFIG_NUMA_BALANCING */ 1638} 1639 1640#ifdef CONFIG_NUMA_BALANCING 1641#ifdef CONFIG_SCHED_DEBUG 1642void set_numabalancing_state(bool enabled) 1643{ 1644 if (enabled) 1645 sched_feat_set("NUMA"); 1646 else 1647 sched_feat_set("NO_NUMA"); 1648} 1649#else 1650__read_mostly bool numabalancing_enabled; 1651 1652void set_numabalancing_state(bool enabled) 1653{ 1654 numabalancing_enabled = enabled; 1655} 1656#endif /* CONFIG_SCHED_DEBUG */ 1657#endif /* CONFIG_NUMA_BALANCING */ 1658 1659/* 1660 * fork()/clone()-time setup: 1661 */ 1662void sched_fork(struct task_struct *p) 1663{ 1664 unsigned long flags; 1665 int cpu = get_cpu(); 1666 1667 __sched_fork(p); 1668 /* 1669 * We mark the process as running here. This guarantees that 1670 * nobody will actually run it, and a signal or other external 1671 * event cannot wake it up and insert it on the runqueue either. 1672 */ 1673 p->state = TASK_RUNNING; 1674 1675 /* 1676 * Make sure we do not leak PI boosting priority to the child. 1677 */ 1678 p->prio = current->normal_prio; 1679 1680 /* 1681 * Revert to default priority/policy on fork if requested. 1682 */ 1683 if (unlikely(p->sched_reset_on_fork)) { 1684 if (task_has_rt_policy(p)) { 1685 p->policy = SCHED_NORMAL; 1686 p->static_prio = NICE_TO_PRIO(0); 1687 p->rt_priority = 0; 1688 } else if (PRIO_TO_NICE(p->static_prio) < 0) 1689 p->static_prio = NICE_TO_PRIO(0); 1690 1691 p->prio = p->normal_prio = __normal_prio(p); 1692 set_load_weight(p); 1693 1694 /* 1695 * We don't need the reset flag anymore after the fork. It has 1696 * fulfilled its duty: 1697 */ 1698 p->sched_reset_on_fork = 0; 1699 } 1700 1701 if (!rt_prio(p->prio)) 1702 p->sched_class = &fair_sched_class; 1703 1704 if (p->sched_class->task_fork) 1705 p->sched_class->task_fork(p); 1706 1707 /* 1708 * The child is not yet in the pid-hash so no cgroup attach races, 1709 * and the cgroup is pinned to this child due to cgroup_fork() 1710 * is ran before sched_fork(). 1711 * 1712 * Silence PROVE_RCU. 1713 */ 1714 raw_spin_lock_irqsave(&p->pi_lock, flags); 1715 set_task_cpu(p, cpu); 1716 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 1717 1718#if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT) 1719 if (likely(sched_info_on())) 1720 memset(&p->sched_info, 0, sizeof(p->sched_info)); 1721#endif 1722#if defined(CONFIG_SMP) 1723 p->on_cpu = 0; 1724#endif 1725 init_task_preempt_count(p); 1726#ifdef CONFIG_SMP 1727 plist_node_init(&p->pushable_tasks, MAX_PRIO); 1728#endif 1729 1730 put_cpu(); 1731} 1732 1733/* 1734 * wake_up_new_task - wake up a newly created task for the first time. 1735 * 1736 * This function will do some initial scheduler statistics housekeeping 1737 * that must be done for every newly created context, then puts the task 1738 * on the runqueue and wakes it. 1739 */ 1740void wake_up_new_task(struct task_struct *p) 1741{ 1742 unsigned long flags; 1743 struct rq *rq; 1744 1745 raw_spin_lock_irqsave(&p->pi_lock, flags); 1746#ifdef CONFIG_SMP 1747 /* 1748 * Fork balancing, do it here and not earlier because: 1749 * - cpus_allowed can change in the fork path 1750 * - any previously selected cpu might disappear through hotplug 1751 */ 1752 set_task_cpu(p, select_task_rq(p, SD_BALANCE_FORK, 0)); 1753#endif 1754 1755 /* Initialize new task's runnable average */ 1756 init_task_runnable_average(p); 1757 rq = __task_rq_lock(p); 1758 activate_task(rq, p, 0); 1759 p->on_rq = 1; 1760 trace_sched_wakeup_new(p, true); 1761 check_preempt_curr(rq, p, WF_FORK); 1762#ifdef CONFIG_SMP 1763 if (p->sched_class->task_woken) 1764 p->sched_class->task_woken(rq, p); 1765#endif 1766 task_rq_unlock(rq, p, &flags); 1767} 1768 1769#ifdef CONFIG_PREEMPT_NOTIFIERS 1770 1771/** 1772 * preempt_notifier_register - tell me when current is being preempted & rescheduled 1773 * @notifier: notifier struct to register 1774 */ 1775void preempt_notifier_register(struct preempt_notifier *notifier) 1776{ 1777 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers); 1778} 1779EXPORT_SYMBOL_GPL(preempt_notifier_register); 1780 1781/** 1782 * preempt_notifier_unregister - no longer interested in preemption notifications 1783 * @notifier: notifier struct to unregister 1784 * 1785 * This is safe to call from within a preemption notifier. 1786 */ 1787void preempt_notifier_unregister(struct preempt_notifier *notifier) 1788{ 1789 hlist_del(¬ifier->link); 1790} 1791EXPORT_SYMBOL_GPL(preempt_notifier_unregister); 1792 1793static void fire_sched_in_preempt_notifiers(struct task_struct *curr) 1794{ 1795 struct preempt_notifier *notifier; 1796 1797 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link) 1798 notifier->ops->sched_in(notifier, raw_smp_processor_id()); 1799} 1800 1801static void 1802fire_sched_out_preempt_notifiers(struct task_struct *curr, 1803 struct task_struct *next) 1804{ 1805 struct preempt_notifier *notifier; 1806 1807 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link) 1808 notifier->ops->sched_out(notifier, next); 1809} 1810 1811#else /* !CONFIG_PREEMPT_NOTIFIERS */ 1812 1813static void fire_sched_in_preempt_notifiers(struct task_struct *curr) 1814{ 1815} 1816 1817static void 1818fire_sched_out_preempt_notifiers(struct task_struct *curr, 1819 struct task_struct *next) 1820{ 1821} 1822 1823#endif /* CONFIG_PREEMPT_NOTIFIERS */ 1824 1825/** 1826 * prepare_task_switch - prepare to switch tasks 1827 * @rq: the runqueue preparing to switch 1828 * @prev: the current task that is being switched out 1829 * @next: the task we are going to switch to. 1830 * 1831 * This is called with the rq lock held and interrupts off. It must 1832 * be paired with a subsequent finish_task_switch after the context 1833 * switch. 1834 * 1835 * prepare_task_switch sets up locking and calls architecture specific 1836 * hooks. 1837 */ 1838static inline void 1839prepare_task_switch(struct rq *rq, struct task_struct *prev, 1840 struct task_struct *next) 1841{ 1842 trace_sched_switch(prev, next); 1843 sched_info_switch(rq, prev, next); 1844 perf_event_task_sched_out(prev, next); 1845 fire_sched_out_preempt_notifiers(prev, next); 1846 prepare_lock_switch(rq, next); 1847 prepare_arch_switch(next); 1848} 1849 1850/** 1851 * finish_task_switch - clean up after a task-switch 1852 * @rq: runqueue associated with task-switch 1853 * @prev: the thread we just switched away from. 1854 * 1855 * finish_task_switch must be called after the context switch, paired 1856 * with a prepare_task_switch call before the context switch. 1857 * finish_task_switch will reconcile locking set up by prepare_task_switch, 1858 * and do any other architecture-specific cleanup actions. 1859 * 1860 * Note that we may have delayed dropping an mm in context_switch(). If 1861 * so, we finish that here outside of the runqueue lock. (Doing it 1862 * with the lock held can cause deadlocks; see schedule() for 1863 * details.) 1864 */ 1865static void finish_task_switch(struct rq *rq, struct task_struct *prev) 1866 __releases(rq->lock) 1867{ 1868 struct mm_struct *mm = rq->prev_mm; 1869 long prev_state; 1870 1871 rq->prev_mm = NULL; 1872 1873 /* 1874 * A task struct has one reference for the use as "current". 1875 * If a task dies, then it sets TASK_DEAD in tsk->state and calls 1876 * schedule one last time. The schedule call will never return, and 1877 * the scheduled task must drop that reference. 1878 * The test for TASK_DEAD must occur while the runqueue locks are 1879 * still held, otherwise prev could be scheduled on another cpu, die 1880 * there before we look at prev->state, and then the reference would 1881 * be dropped twice. 1882 * Manfred Spraul <manfred@colorfullife.com> 1883 */ 1884 prev_state = prev->state; 1885 vtime_task_switch(prev); 1886 finish_arch_switch(prev); 1887 perf_event_task_sched_in(prev, current); 1888 finish_lock_switch(rq, prev); 1889 finish_arch_post_lock_switch(); 1890 1891 fire_sched_in_preempt_notifiers(current); 1892 if (mm) 1893 mmdrop(mm); 1894 if (unlikely(prev_state == TASK_DEAD)) { 1895 /* 1896 * Remove function-return probe instances associated with this 1897 * task and put them back on the free list. 1898 */ 1899 kprobe_flush_task(prev); 1900 put_task_struct(prev); 1901 } 1902 1903 tick_nohz_task_switch(current); 1904} 1905 1906#ifdef CONFIG_SMP 1907 1908/* assumes rq->lock is held */ 1909static inline void pre_schedule(struct rq *rq, struct task_struct *prev) 1910{ 1911 if (prev->sched_class->pre_schedule) 1912 prev->sched_class->pre_schedule(rq, prev); 1913} 1914 1915/* rq->lock is NOT held, but preemption is disabled */ 1916static inline void post_schedule(struct rq *rq) 1917{ 1918 if (rq->post_schedule) { 1919 unsigned long flags; 1920 1921 raw_spin_lock_irqsave(&rq->lock, flags); 1922 if (rq->curr->sched_class->post_schedule) 1923 rq->curr->sched_class->post_schedule(rq); 1924 raw_spin_unlock_irqrestore(&rq->lock, flags); 1925 1926 rq->post_schedule = 0; 1927 } 1928} 1929 1930#else 1931 1932static inline void pre_schedule(struct rq *rq, struct task_struct *p) 1933{ 1934} 1935 1936static inline void post_schedule(struct rq *rq) 1937{ 1938} 1939 1940#endif 1941 1942/** 1943 * schedule_tail - first thing a freshly forked thread must call. 1944 * @prev: the thread we just switched away from. 1945 */ 1946asmlinkage void schedule_tail(struct task_struct *prev) 1947 __releases(rq->lock) 1948{ 1949 struct rq *rq = this_rq(); 1950 1951 finish_task_switch(rq, prev); 1952 1953 /* 1954 * FIXME: do we need to worry about rq being invalidated by the 1955 * task_switch? 1956 */ 1957 post_schedule(rq); 1958 1959#ifdef __ARCH_WANT_UNLOCKED_CTXSW 1960 /* In this case, finish_task_switch does not reenable preemption */ 1961 preempt_enable(); 1962#endif 1963 if (current->set_child_tid) 1964 put_user(task_pid_vnr(current), current->set_child_tid); 1965} 1966 1967/* 1968 * context_switch - switch to the new MM and the new 1969 * thread's register state. 1970 */ 1971static inline void 1972context_switch(struct rq *rq, struct task_struct *prev, 1973 struct task_struct *next) 1974{ 1975 struct mm_struct *mm, *oldmm; 1976 1977 prepare_task_switch(rq, prev, next); 1978 1979 mm = next->mm; 1980 oldmm = prev->active_mm; 1981 /* 1982 * For paravirt, this is coupled with an exit in switch_to to 1983 * combine the page table reload and the switch backend into 1984 * one hypercall. 1985 */ 1986 arch_start_context_switch(prev); 1987 1988 if (!mm) { 1989 next->active_mm = oldmm; 1990 atomic_inc(&oldmm->mm_count); 1991 enter_lazy_tlb(oldmm, next); 1992 } else 1993 switch_mm(oldmm, mm, next); 1994 1995 if (!prev->mm) { 1996 prev->active_mm = NULL; 1997 rq->prev_mm = oldmm; 1998 } 1999 /* 2000 * Since the runqueue lock will be released by the next 2001 * task (which is an invalid locking op but in the case 2002 * of the scheduler it's an obvious special-case), so we 2003 * do an early lockdep release here: 2004 */ 2005#ifndef __ARCH_WANT_UNLOCKED_CTXSW 2006 spin_release(&rq->lock.dep_map, 1, _THIS_IP_); 2007#endif 2008 2009 context_tracking_task_switch(prev, next); 2010 /* Here we just switch the register state and the stack. */ 2011 switch_to(prev, next, prev); 2012 2013 barrier(); 2014 /* 2015 * this_rq must be evaluated again because prev may have moved 2016 * CPUs since it called schedule(), thus the 'rq' on its stack 2017 * frame will be invalid. 2018 */ 2019 finish_task_switch(this_rq(), prev); 2020} 2021 2022/* 2023 * nr_running and nr_context_switches: 2024 * 2025 * externally visible scheduler statistics: current number of runnable 2026 * threads, total number of context switches performed since bootup. 2027 */ 2028unsigned long nr_running(void) 2029{ 2030 unsigned long i, sum = 0; 2031 2032 for_each_online_cpu(i) 2033 sum += cpu_rq(i)->nr_running; 2034 2035 return sum; 2036} 2037 2038unsigned long long nr_context_switches(void) 2039{ 2040 int i; 2041 unsigned long long sum = 0; 2042 2043 for_each_possible_cpu(i) 2044 sum += cpu_rq(i)->nr_switches; 2045 2046 return sum; 2047} 2048 2049unsigned long nr_iowait(void) 2050{ 2051 unsigned long i, sum = 0; 2052 2053 for_each_possible_cpu(i) 2054 sum += atomic_read(&cpu_rq(i)->nr_iowait); 2055 2056 return sum; 2057} 2058 2059unsigned long nr_iowait_cpu(int cpu) 2060{ 2061 struct rq *this = cpu_rq(cpu); 2062 return atomic_read(&this->nr_iowait); 2063} 2064 2065#ifdef CONFIG_SMP 2066 2067/* 2068 * sched_exec - execve() is a valuable balancing opportunity, because at 2069 * this point the task has the smallest effective memory and cache footprint. 2070 */ 2071void sched_exec(void) 2072{ 2073 struct task_struct *p = current; 2074 unsigned long flags; 2075 int dest_cpu; 2076 2077 raw_spin_lock_irqsave(&p->pi_lock, flags); 2078 dest_cpu = p->sched_class->select_task_rq(p, SD_BALANCE_EXEC, 0); 2079 if (dest_cpu == smp_processor_id()) 2080 goto unlock; 2081 2082 if (likely(cpu_active(dest_cpu))) { 2083 struct migration_arg arg = { p, dest_cpu }; 2084 2085 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 2086 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg); 2087 return; 2088 } 2089unlock: 2090 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 2091} 2092 2093#endif 2094 2095DEFINE_PER_CPU(struct kernel_stat, kstat); 2096DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat); 2097 2098EXPORT_PER_CPU_SYMBOL(kstat); 2099EXPORT_PER_CPU_SYMBOL(kernel_cpustat); 2100 2101/* 2102 * Return any ns on the sched_clock that have not yet been accounted in 2103 * @p in case that task is currently running. 2104 * 2105 * Called with task_rq_lock() held on @rq. 2106 */ 2107static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq) 2108{ 2109 u64 ns = 0; 2110 2111 if (task_current(rq, p)) { 2112 update_rq_clock(rq); 2113 ns = rq_clock_task(rq) - p->se.exec_start; 2114 if ((s64)ns < 0) 2115 ns = 0; 2116 } 2117 2118 return ns; 2119} 2120 2121unsigned long long task_delta_exec(struct task_struct *p) 2122{ 2123 unsigned long flags; 2124 struct rq *rq; 2125 u64 ns = 0; 2126 2127 rq = task_rq_lock(p, &flags); 2128 ns = do_task_delta_exec(p, rq); 2129 task_rq_unlock(rq, p, &flags); 2130 2131 return ns; 2132} 2133 2134/* 2135 * Return accounted runtime for the task. 2136 * In case the task is currently running, return the runtime plus current's 2137 * pending runtime that have not been accounted yet. 2138 */ 2139unsigned long long task_sched_runtime(struct task_struct *p) 2140{ 2141 unsigned long flags; 2142 struct rq *rq; 2143 u64 ns = 0; 2144 2145 rq = task_rq_lock(p, &flags); 2146 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq); 2147 task_rq_unlock(rq, p, &flags); 2148 2149 return ns; 2150} 2151 2152/* 2153 * This function gets called by the timer code, with HZ frequency. 2154 * We call it with interrupts disabled. 2155 */ 2156void scheduler_tick(void) 2157{ 2158 int cpu = smp_processor_id(); 2159 struct rq *rq = cpu_rq(cpu); 2160 struct task_struct *curr = rq->curr; 2161 2162 sched_clock_tick(); 2163 2164 raw_spin_lock(&rq->lock); 2165 update_rq_clock(rq); 2166 curr->sched_class->task_tick(rq, curr, 0); 2167 update_cpu_load_active(rq); 2168 raw_spin_unlock(&rq->lock); 2169 2170 perf_event_task_tick(); 2171 2172#ifdef CONFIG_SMP 2173 rq->idle_balance = idle_cpu(cpu); 2174 trigger_load_balance(rq, cpu); 2175#endif 2176 rq_last_tick_reset(rq); 2177} 2178 2179#ifdef CONFIG_NO_HZ_FULL 2180/** 2181 * scheduler_tick_max_deferment 2182 * 2183 * Keep at least one tick per second when a single 2184 * active task is running because the scheduler doesn't 2185 * yet completely support full dynticks environment. 2186 * 2187 * This makes sure that uptime, CFS vruntime, load 2188 * balancing, etc... continue to move forward, even 2189 * with a very low granularity. 2190 * 2191 * Return: Maximum deferment in nanoseconds. 2192 */ 2193u64 scheduler_tick_max_deferment(void) 2194{ 2195 struct rq *rq = this_rq(); 2196 unsigned long next, now = ACCESS_ONCE(jiffies); 2197 2198 next = rq->last_sched_tick + HZ; 2199 2200 if (time_before_eq(next, now)) 2201 return 0; 2202 2203 return jiffies_to_usecs(next - now) * NSEC_PER_USEC; 2204} 2205#endif 2206 2207notrace unsigned long get_parent_ip(unsigned long addr) 2208{ 2209 if (in_lock_functions(addr)) { 2210 addr = CALLER_ADDR2; 2211 if (in_lock_functions(addr)) 2212 addr = CALLER_ADDR3; 2213 } 2214 return addr; 2215} 2216 2217#if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \ 2218 defined(CONFIG_PREEMPT_TRACER)) 2219 2220void __kprobes preempt_count_add(int val) 2221{ 2222#ifdef CONFIG_DEBUG_PREEMPT 2223 /* 2224 * Underflow? 2225 */ 2226 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0))) 2227 return; 2228#endif 2229 __preempt_count_add(val); 2230#ifdef CONFIG_DEBUG_PREEMPT 2231 /* 2232 * Spinlock count overflowing soon? 2233 */ 2234 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >= 2235 PREEMPT_MASK - 10); 2236#endif 2237 if (preempt_count() == val) 2238 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1)); 2239} 2240EXPORT_SYMBOL(preempt_count_add); 2241 2242void __kprobes preempt_count_sub(int val) 2243{ 2244#ifdef CONFIG_DEBUG_PREEMPT 2245 /* 2246 * Underflow? 2247 */ 2248 if (DEBUG_LOCKS_WARN_ON(val > preempt_count())) 2249 return; 2250 /* 2251 * Is the spinlock portion underflowing? 2252 */ 2253 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) && 2254 !(preempt_count() & PREEMPT_MASK))) 2255 return; 2256#endif 2257 2258 if (preempt_count() == val) 2259 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1)); 2260 __preempt_count_sub(val); 2261} 2262EXPORT_SYMBOL(preempt_count_sub); 2263 2264#endif 2265 2266/* 2267 * Print scheduling while atomic bug: 2268 */ 2269static noinline void __schedule_bug(struct task_struct *prev) 2270{ 2271 if (oops_in_progress) 2272 return; 2273 2274 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n", 2275 prev->comm, prev->pid, preempt_count()); 2276 2277 debug_show_held_locks(prev); 2278 print_modules(); 2279 if (irqs_disabled()) 2280 print_irqtrace_events(prev); 2281 dump_stack(); 2282 add_taint(TAINT_WARN, LOCKDEP_STILL_OK); 2283} 2284 2285/* 2286 * Various schedule()-time debugging checks and statistics: 2287 */ 2288static inline void schedule_debug(struct task_struct *prev) 2289{ 2290 /* 2291 * Test if we are atomic. Since do_exit() needs to call into 2292 * schedule() atomically, we ignore that path for now. 2293 * Otherwise, whine if we are scheduling when we should not be. 2294 */ 2295 if (unlikely(in_atomic_preempt_off() && !prev->exit_state)) 2296 __schedule_bug(prev); 2297 rcu_sleep_check(); 2298 2299 profile_hit(SCHED_PROFILING, __builtin_return_address(0)); 2300 2301 schedstat_inc(this_rq(), sched_count); 2302} 2303 2304static void put_prev_task(struct rq *rq, struct task_struct *prev) 2305{ 2306 if (prev->on_rq || rq->skip_clock_update < 0) 2307 update_rq_clock(rq); 2308 prev->sched_class->put_prev_task(rq, prev); 2309} 2310 2311/* 2312 * Pick up the highest-prio task: 2313 */ 2314static inline struct task_struct * 2315pick_next_task(struct rq *rq) 2316{ 2317 const struct sched_class *class; 2318 struct task_struct *p; 2319 2320 /* 2321 * Optimization: we know that if all tasks are in 2322 * the fair class we can call that function directly: 2323 */ 2324 if (likely(rq->nr_running == rq->cfs.h_nr_running)) { 2325 p = fair_sched_class.pick_next_task(rq); 2326 if (likely(p)) 2327 return p; 2328 } 2329 2330 for_each_class(class) { 2331 p = class->pick_next_task(rq); 2332 if (p) 2333 return p; 2334 } 2335 2336 BUG(); /* the idle class will always have a runnable task */ 2337} 2338 2339/* 2340 * __schedule() is the main scheduler function. 2341 * 2342 * The main means of driving the scheduler and thus entering this function are: 2343 * 2344 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc. 2345 * 2346 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return 2347 * paths. For example, see arch/x86/entry_64.S. 2348 * 2349 * To drive preemption between tasks, the scheduler sets the flag in timer 2350 * interrupt handler scheduler_tick(). 2351 * 2352 * 3. Wakeups don't really cause entry into schedule(). They add a 2353 * task to the run-queue and that's it. 2354 * 2355 * Now, if the new task added to the run-queue preempts the current 2356 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets 2357 * called on the nearest possible occasion: 2358 * 2359 * - If the kernel is preemptible (CONFIG_PREEMPT=y): 2360 * 2361 * - in syscall or exception context, at the next outmost 2362 * preempt_enable(). (this might be as soon as the wake_up()'s 2363 * spin_unlock()!) 2364 * 2365 * - in IRQ context, return from interrupt-handler to 2366 * preemptible context 2367 * 2368 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set) 2369 * then at the next: 2370 * 2371 * - cond_resched() call 2372 * - explicit schedule() call 2373 * - return from syscall or exception to user-space 2374 * - return from interrupt-handler to user-space 2375 */ 2376static void __sched __schedule(void) 2377{ 2378 struct task_struct *prev, *next; 2379 unsigned long *switch_count; 2380 struct rq *rq; 2381 int cpu; 2382 2383need_resched: 2384 preempt_disable(); 2385 cpu = smp_processor_id(); 2386 rq = cpu_rq(cpu); 2387 rcu_note_context_switch(cpu); 2388 prev = rq->curr; 2389 2390 schedule_debug(prev); 2391 2392 if (sched_feat(HRTICK)) 2393 hrtick_clear(rq); 2394 2395 /* 2396 * Make sure that signal_pending_state()->signal_pending() below 2397 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE) 2398 * done by the caller to avoid the race with signal_wake_up(). 2399 */ 2400 smp_mb__before_spinlock(); 2401 raw_spin_lock_irq(&rq->lock); 2402 2403 switch_count = &prev->nivcsw; 2404 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) { 2405 if (unlikely(signal_pending_state(prev->state, prev))) { 2406 prev->state = TASK_RUNNING; 2407 } else { 2408 deactivate_task(rq, prev, DEQUEUE_SLEEP); 2409 prev->on_rq = 0; 2410 2411 /* 2412 * If a worker went to sleep, notify and ask workqueue 2413 * whether it wants to wake up a task to maintain 2414 * concurrency. 2415 */ 2416 if (prev->flags & PF_WQ_WORKER) { 2417 struct task_struct *to_wakeup; 2418 2419 to_wakeup = wq_worker_sleeping(prev, cpu); 2420 if (to_wakeup) 2421 try_to_wake_up_local(to_wakeup); 2422 } 2423 } 2424 switch_count = &prev->nvcsw; 2425 } 2426 2427 pre_schedule(rq, prev); 2428 2429 if (unlikely(!rq->nr_running)) 2430 idle_balance(cpu, rq); 2431 2432 put_prev_task(rq, prev); 2433 next = pick_next_task(rq); 2434 clear_tsk_need_resched(prev); 2435 clear_preempt_need_resched(); 2436 rq->skip_clock_update = 0; 2437 2438 if (likely(prev != next)) { 2439 rq->nr_switches++; 2440 rq->curr = next; 2441 ++*switch_count; 2442 2443 context_switch(rq, prev, next); /* unlocks the rq */ 2444 /* 2445 * The context switch have flipped the stack from under us 2446 * and restored the local variables which were saved when 2447 * this task called schedule() in the past. prev == current 2448 * is still correct, but it can be moved to another cpu/rq. 2449 */ 2450 cpu = smp_processor_id(); 2451 rq = cpu_rq(cpu); 2452 } else 2453 raw_spin_unlock_irq(&rq->lock); 2454 2455 post_schedule(rq); 2456 2457 sched_preempt_enable_no_resched(); 2458 if (need_resched()) 2459 goto need_resched; 2460} 2461 2462static inline void sched_submit_work(struct task_struct *tsk) 2463{ 2464 if (!tsk->state || tsk_is_pi_blocked(tsk)) 2465 return; 2466 /* 2467 * If we are going to sleep and we have plugged IO queued, 2468 * make sure to submit it to avoid deadlocks. 2469 */ 2470 if (blk_needs_flush_plug(tsk)) 2471 blk_schedule_flush_plug(tsk); 2472} 2473 2474asmlinkage void __sched schedule(void) 2475{ 2476 struct task_struct *tsk = current; 2477 2478 sched_submit_work(tsk); 2479 __schedule(); 2480} 2481EXPORT_SYMBOL(schedule); 2482 2483#ifdef CONFIG_CONTEXT_TRACKING 2484asmlinkage void __sched schedule_user(void) 2485{ 2486 /* 2487 * If we come here after a random call to set_need_resched(), 2488 * or we have been woken up remotely but the IPI has not yet arrived, 2489 * we haven't yet exited the RCU idle mode. Do it here manually until 2490 * we find a better solution. 2491 */ 2492 user_exit(); 2493 schedule(); 2494 user_enter(); 2495} 2496#endif 2497 2498/** 2499 * schedule_preempt_disabled - called with preemption disabled 2500 * 2501 * Returns with preemption disabled. Note: preempt_count must be 1 2502 */ 2503void __sched schedule_preempt_disabled(void) 2504{ 2505 sched_preempt_enable_no_resched(); 2506 schedule(); 2507 preempt_disable(); 2508} 2509 2510#ifdef CONFIG_PREEMPT 2511/* 2512 * this is the entry point to schedule() from in-kernel preemption 2513 * off of preempt_enable. Kernel preemptions off return from interrupt 2514 * occur there and call schedule directly. 2515 */ 2516asmlinkage void __sched notrace preempt_schedule(void) 2517{ 2518 /* 2519 * If there is a non-zero preempt_count or interrupts are disabled, 2520 * we do not want to preempt the current task. Just return.. 2521 */ 2522 if (likely(!preemptible())) 2523 return; 2524 2525 do { 2526 __preempt_count_add(PREEMPT_ACTIVE); 2527 __schedule(); 2528 __preempt_count_sub(PREEMPT_ACTIVE); 2529 2530 /* 2531 * Check again in case we missed a preemption opportunity 2532 * between schedule and now. 2533 */ 2534 barrier(); 2535 } while (need_resched()); 2536} 2537EXPORT_SYMBOL(preempt_schedule); 2538 2539/* 2540 * this is the entry point to schedule() from kernel preemption 2541 * off of irq context. 2542 * Note, that this is called and return with irqs disabled. This will 2543 * protect us against recursive calling from irq. 2544 */ 2545asmlinkage void __sched preempt_schedule_irq(void) 2546{ 2547 enum ctx_state prev_state; 2548 2549 /* Catch callers which need to be fixed */ 2550 BUG_ON(preempt_count() || !irqs_disabled()); 2551 2552 prev_state = exception_enter(); 2553 2554 do { 2555 __preempt_count_add(PREEMPT_ACTIVE); 2556 local_irq_enable(); 2557 __schedule(); 2558 local_irq_disable(); 2559 __preempt_count_sub(PREEMPT_ACTIVE); 2560 2561 /* 2562 * Check again in case we missed a preemption opportunity 2563 * between schedule and now. 2564 */ 2565 barrier(); 2566 } while (need_resched()); 2567 2568 exception_exit(prev_state); 2569} 2570 2571#endif /* CONFIG_PREEMPT */ 2572 2573int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags, 2574 void *key) 2575{ 2576 return try_to_wake_up(curr->private, mode, wake_flags); 2577} 2578EXPORT_SYMBOL(default_wake_function); 2579 2580/* 2581 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just 2582 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve 2583 * number) then we wake all the non-exclusive tasks and one exclusive task. 2584 * 2585 * There are circumstances in which we can try to wake a task which has already 2586 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns 2587 * zero in this (rare) case, and we handle it by continuing to scan the queue. 2588 */ 2589static void __wake_up_common(wait_queue_head_t *q, unsigned int mode, 2590 int nr_exclusive, int wake_flags, void *key) 2591{ 2592 wait_queue_t *curr, *next; 2593 2594 list_for_each_entry_safe(curr, next, &q->task_list, task_list) { 2595 unsigned flags = curr->flags; 2596 2597 if (curr->func(curr, mode, wake_flags, key) && 2598 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive) 2599 break; 2600 } 2601} 2602 2603/** 2604 * __wake_up - wake up threads blocked on a waitqueue. 2605 * @q: the waitqueue 2606 * @mode: which threads 2607 * @nr_exclusive: how many wake-one or wake-many threads to wake up 2608 * @key: is directly passed to the wakeup function 2609 * 2610 * It may be assumed that this function implies a write memory barrier before 2611 * changing the task state if and only if any tasks are woken up. 2612 */ 2613void __wake_up(wait_queue_head_t *q, unsigned int mode, 2614 int nr_exclusive, void *key) 2615{ 2616 unsigned long flags; 2617 2618 spin_lock_irqsave(&q->lock, flags); 2619 __wake_up_common(q, mode, nr_exclusive, 0, key); 2620 spin_unlock_irqrestore(&q->lock, flags); 2621} 2622EXPORT_SYMBOL(__wake_up); 2623 2624/* 2625 * Same as __wake_up but called with the spinlock in wait_queue_head_t held. 2626 */ 2627void __wake_up_locked(wait_queue_head_t *q, unsigned int mode, int nr) 2628{ 2629 __wake_up_common(q, mode, nr, 0, NULL); 2630} 2631EXPORT_SYMBOL_GPL(__wake_up_locked); 2632 2633void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key) 2634{ 2635 __wake_up_common(q, mode, 1, 0, key); 2636} 2637EXPORT_SYMBOL_GPL(__wake_up_locked_key); 2638 2639/** 2640 * __wake_up_sync_key - wake up threads blocked on a waitqueue. 2641 * @q: the waitqueue 2642 * @mode: which threads 2643 * @nr_exclusive: how many wake-one or wake-many threads to wake up 2644 * @key: opaque value to be passed to wakeup targets 2645 * 2646 * The sync wakeup differs that the waker knows that it will schedule 2647 * away soon, so while the target thread will be woken up, it will not 2648 * be migrated to another CPU - ie. the two threads are 'synchronized' 2649 * with each other. This can prevent needless bouncing between CPUs. 2650 * 2651 * On UP it can prevent extra preemption. 2652 * 2653 * It may be assumed that this function implies a write memory barrier before 2654 * changing the task state if and only if any tasks are woken up. 2655 */ 2656void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode, 2657 int nr_exclusive, void *key) 2658{ 2659 unsigned long flags; 2660 int wake_flags = WF_SYNC; 2661 2662 if (unlikely(!q)) 2663 return; 2664 2665 if (unlikely(nr_exclusive != 1)) 2666 wake_flags = 0; 2667 2668 spin_lock_irqsave(&q->lock, flags); 2669 __wake_up_common(q, mode, nr_exclusive, wake_flags, key); 2670 spin_unlock_irqrestore(&q->lock, flags); 2671} 2672EXPORT_SYMBOL_GPL(__wake_up_sync_key); 2673 2674/* 2675 * __wake_up_sync - see __wake_up_sync_key() 2676 */ 2677void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive) 2678{ 2679 __wake_up_sync_key(q, mode, nr_exclusive, NULL); 2680} 2681EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */ 2682 2683/** 2684 * complete: - signals a single thread waiting on this completion 2685 * @x: holds the state of this particular completion 2686 * 2687 * This will wake up a single thread waiting on this completion. Threads will be 2688 * awakened in the same order in which they were queued. 2689 * 2690 * See also complete_all(), wait_for_completion() and related routines. 2691 * 2692 * It may be assumed that this function implies a write memory barrier before 2693 * changing the task state if and only if any tasks are woken up. 2694 */ 2695void complete(struct completion *x) 2696{ 2697 unsigned long flags; 2698 2699 spin_lock_irqsave(&x->wait.lock, flags); 2700 x->done++; 2701 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL); 2702 spin_unlock_irqrestore(&x->wait.lock, flags); 2703} 2704EXPORT_SYMBOL(complete); 2705 2706/** 2707 * complete_all: - signals all threads waiting on this completion 2708 * @x: holds the state of this particular completion 2709 * 2710 * This will wake up all threads waiting on this particular completion event. 2711 * 2712 * It may be assumed that this function implies a write memory barrier before 2713 * changing the task state if and only if any tasks are woken up. 2714 */ 2715void complete_all(struct completion *x) 2716{ 2717 unsigned long flags; 2718 2719 spin_lock_irqsave(&x->wait.lock, flags); 2720 x->done += UINT_MAX/2; 2721 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL); 2722 spin_unlock_irqrestore(&x->wait.lock, flags); 2723} 2724EXPORT_SYMBOL(complete_all); 2725 2726static inline long __sched 2727do_wait_for_common(struct completion *x, 2728 long (*action)(long), long timeout, int state) 2729{ 2730 if (!x->done) { 2731 DECLARE_WAITQUEUE(wait, current); 2732 2733 __add_wait_queue_tail_exclusive(&x->wait, &wait); 2734 do { 2735 if (signal_pending_state(state, current)) { 2736 timeout = -ERESTARTSYS; 2737 break; 2738 } 2739 __set_current_state(state); 2740 spin_unlock_irq(&x->wait.lock); 2741 timeout = action(timeout); 2742 spin_lock_irq(&x->wait.lock); 2743 } while (!x->done && timeout); 2744 __remove_wait_queue(&x->wait, &wait); 2745 if (!x->done) 2746 return timeout; 2747 } 2748 x->done--; 2749 return timeout ?: 1; 2750} 2751 2752static inline long __sched 2753__wait_for_common(struct completion *x, 2754 long (*action)(long), long timeout, int state) 2755{ 2756 might_sleep(); 2757 2758 spin_lock_irq(&x->wait.lock); 2759 timeout = do_wait_for_common(x, action, timeout, state); 2760 spin_unlock_irq(&x->wait.lock); 2761 return timeout; 2762} 2763 2764static long __sched 2765wait_for_common(struct completion *x, long timeout, int state) 2766{ 2767 return __wait_for_common(x, schedule_timeout, timeout, state); 2768} 2769 2770static long __sched 2771wait_for_common_io(struct completion *x, long timeout, int state) 2772{ 2773 return __wait_for_common(x, io_schedule_timeout, timeout, state); 2774} 2775 2776/** 2777 * wait_for_completion: - waits for completion of a task 2778 * @x: holds the state of this particular completion 2779 * 2780 * This waits to be signaled for completion of a specific task. It is NOT 2781 * interruptible and there is no timeout. 2782 * 2783 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout 2784 * and interrupt capability. Also see complete(). 2785 */ 2786void __sched wait_for_completion(struct completion *x) 2787{ 2788 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE); 2789} 2790EXPORT_SYMBOL(wait_for_completion); 2791 2792/** 2793 * wait_for_completion_timeout: - waits for completion of a task (w/timeout) 2794 * @x: holds the state of this particular completion 2795 * @timeout: timeout value in jiffies 2796 * 2797 * This waits for either a completion of a specific task to be signaled or for a 2798 * specified timeout to expire. The timeout is in jiffies. It is not 2799 * interruptible. 2800 * 2801 * Return: 0 if timed out, and positive (at least 1, or number of jiffies left 2802 * till timeout) if completed. 2803 */ 2804unsigned long __sched 2805wait_for_completion_timeout(struct completion *x, unsigned long timeout) 2806{ 2807 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE); 2808} 2809EXPORT_SYMBOL(wait_for_completion_timeout); 2810 2811/** 2812 * wait_for_completion_io: - waits for completion of a task 2813 * @x: holds the state of this particular completion 2814 * 2815 * This waits to be signaled for completion of a specific task. It is NOT 2816 * interruptible and there is no timeout. The caller is accounted as waiting 2817 * for IO. 2818 */ 2819void __sched wait_for_completion_io(struct completion *x) 2820{ 2821 wait_for_common_io(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE); 2822} 2823EXPORT_SYMBOL(wait_for_completion_io); 2824 2825/** 2826 * wait_for_completion_io_timeout: - waits for completion of a task (w/timeout) 2827 * @x: holds the state of this particular completion 2828 * @timeout: timeout value in jiffies 2829 * 2830 * This waits for either a completion of a specific task to be signaled or for a 2831 * specified timeout to expire. The timeout is in jiffies. It is not 2832 * interruptible. The caller is accounted as waiting for IO. 2833 * 2834 * Return: 0 if timed out, and positive (at least 1, or number of jiffies left 2835 * till timeout) if completed. 2836 */ 2837unsigned long __sched 2838wait_for_completion_io_timeout(struct completion *x, unsigned long timeout) 2839{ 2840 return wait_for_common_io(x, timeout, TASK_UNINTERRUPTIBLE); 2841} 2842EXPORT_SYMBOL(wait_for_completion_io_timeout); 2843 2844/** 2845 * wait_for_completion_interruptible: - waits for completion of a task (w/intr) 2846 * @x: holds the state of this particular completion 2847 * 2848 * This waits for completion of a specific task to be signaled. It is 2849 * interruptible. 2850 * 2851 * Return: -ERESTARTSYS if interrupted, 0 if completed. 2852 */ 2853int __sched wait_for_completion_interruptible(struct completion *x) 2854{ 2855 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE); 2856 if (t == -ERESTARTSYS) 2857 return t; 2858 return 0; 2859} 2860EXPORT_SYMBOL(wait_for_completion_interruptible); 2861 2862/** 2863 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr)) 2864 * @x: holds the state of this particular completion 2865 * @timeout: timeout value in jiffies 2866 * 2867 * This waits for either a completion of a specific task to be signaled or for a 2868 * specified timeout to expire. It is interruptible. The timeout is in jiffies. 2869 * 2870 * Return: -ERESTARTSYS if interrupted, 0 if timed out, positive (at least 1, 2871 * or number of jiffies left till timeout) if completed. 2872 */ 2873long __sched 2874wait_for_completion_interruptible_timeout(struct completion *x, 2875 unsigned long timeout) 2876{ 2877 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE); 2878} 2879EXPORT_SYMBOL(wait_for_completion_interruptible_timeout); 2880 2881/** 2882 * wait_for_completion_killable: - waits for completion of a task (killable) 2883 * @x: holds the state of this particular completion 2884 * 2885 * This waits to be signaled for completion of a specific task. It can be 2886 * interrupted by a kill signal. 2887 * 2888 * Return: -ERESTARTSYS if interrupted, 0 if completed. 2889 */ 2890int __sched wait_for_completion_killable(struct completion *x) 2891{ 2892 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE); 2893 if (t == -ERESTARTSYS) 2894 return t; 2895 return 0; 2896} 2897EXPORT_SYMBOL(wait_for_completion_killable); 2898 2899/** 2900 * wait_for_completion_killable_timeout: - waits for completion of a task (w/(to,killable)) 2901 * @x: holds the state of this particular completion 2902 * @timeout: timeout value in jiffies 2903 * 2904 * This waits for either a completion of a specific task to be 2905 * signaled or for a specified timeout to expire. It can be 2906 * interrupted by a kill signal. The timeout is in jiffies. 2907 * 2908 * Return: -ERESTARTSYS if interrupted, 0 if timed out, positive (at least 1, 2909 * or number of jiffies left till timeout) if completed. 2910 */ 2911long __sched 2912wait_for_completion_killable_timeout(struct completion *x, 2913 unsigned long timeout) 2914{ 2915 return wait_for_common(x, timeout, TASK_KILLABLE); 2916} 2917EXPORT_SYMBOL(wait_for_completion_killable_timeout); 2918 2919/** 2920 * try_wait_for_completion - try to decrement a completion without blocking 2921 * @x: completion structure 2922 * 2923 * Return: 0 if a decrement cannot be done without blocking 2924 * 1 if a decrement succeeded. 2925 * 2926 * If a completion is being used as a counting completion, 2927 * attempt to decrement the counter without blocking. This 2928 * enables us to avoid waiting if the resource the completion 2929 * is protecting is not available. 2930 */ 2931bool try_wait_for_completion(struct completion *x) 2932{ 2933 unsigned long flags; 2934 int ret = 1; 2935 2936 spin_lock_irqsave(&x->wait.lock, flags); 2937 if (!x->done) 2938 ret = 0; 2939 else 2940 x->done--; 2941 spin_unlock_irqrestore(&x->wait.lock, flags); 2942 return ret; 2943} 2944EXPORT_SYMBOL(try_wait_for_completion); 2945 2946/** 2947 * completion_done - Test to see if a completion has any waiters 2948 * @x: completion structure 2949 * 2950 * Return: 0 if there are waiters (wait_for_completion() in progress) 2951 * 1 if there are no waiters. 2952 * 2953 */ 2954bool completion_done(struct completion *x) 2955{ 2956 unsigned long flags; 2957 int ret = 1; 2958 2959 spin_lock_irqsave(&x->wait.lock, flags); 2960 if (!x->done) 2961 ret = 0; 2962 spin_unlock_irqrestore(&x->wait.lock, flags); 2963 return ret; 2964} 2965EXPORT_SYMBOL(completion_done); 2966 2967static long __sched 2968sleep_on_common(wait_queue_head_t *q, int state, long timeout) 2969{ 2970 unsigned long flags; 2971 wait_queue_t wait; 2972 2973 init_waitqueue_entry(&wait, current); 2974 2975 __set_current_state(state); 2976 2977 spin_lock_irqsave(&q->lock, flags); 2978 __add_wait_queue(q, &wait); 2979 spin_unlock(&q->lock); 2980 timeout = schedule_timeout(timeout); 2981 spin_lock_irq(&q->lock); 2982 __remove_wait_queue(q, &wait); 2983 spin_unlock_irqrestore(&q->lock, flags); 2984 2985 return timeout; 2986} 2987 2988void __sched interruptible_sleep_on(wait_queue_head_t *q) 2989{ 2990 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT); 2991} 2992EXPORT_SYMBOL(interruptible_sleep_on); 2993 2994long __sched 2995interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout) 2996{ 2997 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout); 2998} 2999EXPORT_SYMBOL(interruptible_sleep_on_timeout); 3000 3001void __sched sleep_on(wait_queue_head_t *q) 3002{ 3003 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT); 3004} 3005EXPORT_SYMBOL(sleep_on); 3006 3007long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout) 3008{ 3009 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout); 3010} 3011EXPORT_SYMBOL(sleep_on_timeout); 3012 3013#ifdef CONFIG_RT_MUTEXES 3014 3015/* 3016 * rt_mutex_setprio - set the current priority of a task 3017 * @p: task 3018 * @prio: prio value (kernel-internal form) 3019 * 3020 * This function changes the 'effective' priority of a task. It does 3021 * not touch ->normal_prio like __setscheduler(). 3022 * 3023 * Used by the rt_mutex code to implement priority inheritance logic. 3024 */ 3025void rt_mutex_setprio(struct task_struct *p, int prio) 3026{ 3027 int oldprio, on_rq, running; 3028 struct rq *rq; 3029 const struct sched_class *prev_class; 3030 3031 BUG_ON(prio < 0 || prio > MAX_PRIO); 3032 3033 rq = __task_rq_lock(p); 3034 3035 /* 3036 * Idle task boosting is a nono in general. There is one 3037 * exception, when PREEMPT_RT and NOHZ is active: 3038 * 3039 * The idle task calls get_next_timer_interrupt() and holds 3040 * the timer wheel base->lock on the CPU and another CPU wants 3041 * to access the timer (probably to cancel it). We can safely 3042 * ignore the boosting request, as the idle CPU runs this code 3043 * with interrupts disabled and will complete the lock 3044 * protected section without being interrupted. So there is no 3045 * real need to boost. 3046 */ 3047 if (unlikely(p == rq->idle)) { 3048 WARN_ON(p != rq->curr); 3049 WARN_ON(p->pi_blocked_on); 3050 goto out_unlock; 3051 } 3052 3053 trace_sched_pi_setprio(p, prio); 3054 oldprio = p->prio; 3055 prev_class = p->sched_class; 3056 on_rq = p->on_rq; 3057 running = task_current(rq, p); 3058 if (on_rq) 3059 dequeue_task(rq, p, 0); 3060 if (running) 3061 p->sched_class->put_prev_task(rq, p); 3062 3063 if (rt_prio(prio)) 3064 p->sched_class = &rt_sched_class; 3065 else 3066 p->sched_class = &fair_sched_class; 3067 3068 p->prio = prio; 3069 3070 if (running) 3071 p->sched_class->set_curr_task(rq); 3072 if (on_rq) 3073 enqueue_task(rq, p, oldprio < prio ? ENQUEUE_HEAD : 0); 3074 3075 check_class_changed(rq, p, prev_class, oldprio); 3076out_unlock: 3077 __task_rq_unlock(rq); 3078} 3079#endif 3080void set_user_nice(struct task_struct *p, long nice) 3081{ 3082 int old_prio, delta, on_rq; 3083 unsigned long flags; 3084 struct rq *rq; 3085 3086 if (TASK_NICE(p) == nice || nice < -20 || nice > 19) 3087 return; 3088 /* 3089 * We have to be careful, if called from sys_setpriority(), 3090 * the task might be in the middle of scheduling on another CPU. 3091 */ 3092 rq = task_rq_lock(p, &flags); 3093 /* 3094 * The RT priorities are set via sched_setscheduler(), but we still 3095 * allow the 'normal' nice value to be set - but as expected 3096 * it wont have any effect on scheduling until the task is 3097 * SCHED_FIFO/SCHED_RR: 3098 */ 3099 if (task_has_rt_policy(p)) { 3100 p->static_prio = NICE_TO_PRIO(nice); 3101 goto out_unlock; 3102 } 3103 on_rq = p->on_rq; 3104 if (on_rq) 3105 dequeue_task(rq, p, 0); 3106 3107 p->static_prio = NICE_TO_PRIO(nice); 3108 set_load_weight(p); 3109 old_prio = p->prio; 3110 p->prio = effective_prio(p); 3111 delta = p->prio - old_prio; 3112 3113 if (on_rq) { 3114 enqueue_task(rq, p, 0); 3115 /* 3116 * If the task increased its priority or is running and 3117 * lowered its priority, then reschedule its CPU: 3118 */ 3119 if (delta < 0 || (delta > 0 && task_running(rq, p))) 3120 resched_task(rq->curr); 3121 } 3122out_unlock: 3123 task_rq_unlock(rq, p, &flags); 3124} 3125EXPORT_SYMBOL(set_user_nice); 3126 3127/* 3128 * can_nice - check if a task can reduce its nice value 3129 * @p: task 3130 * @nice: nice value 3131 */ 3132int can_nice(const struct task_struct *p, const int nice) 3133{ 3134 /* convert nice value [19,-20] to rlimit style value [1,40] */ 3135 int nice_rlim = 20 - nice; 3136 3137 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) || 3138 capable(CAP_SYS_NICE)); 3139} 3140 3141#ifdef __ARCH_WANT_SYS_NICE 3142 3143/* 3144 * sys_nice - change the priority of the current process. 3145 * @increment: priority increment 3146 * 3147 * sys_setpriority is a more generic, but much slower function that 3148 * does similar things. 3149 */ 3150SYSCALL_DEFINE1(nice, int, increment) 3151{ 3152 long nice, retval; 3153 3154 /* 3155 * Setpriority might change our priority at the same moment. 3156 * We don't have to worry. Conceptually one call occurs first 3157 * and we have a single winner. 3158 */ 3159 if (increment < -40) 3160 increment = -40; 3161 if (increment > 40) 3162 increment = 40; 3163 3164 nice = TASK_NICE(current) + increment; 3165 if (nice < -20) 3166 nice = -20; 3167 if (nice > 19) 3168 nice = 19; 3169 3170 if (increment < 0 && !can_nice(current, nice)) 3171 return -EPERM; 3172 3173 retval = security_task_setnice(current, nice); 3174 if (retval) 3175 return retval; 3176 3177 set_user_nice(current, nice); 3178 return 0; 3179} 3180 3181#endif 3182 3183/** 3184 * task_prio - return the priority value of a given task. 3185 * @p: the task in question. 3186 * 3187 * Return: The priority value as seen by users in /proc. 3188 * RT tasks are offset by -200. Normal tasks are centered 3189 * around 0, value goes from -16 to +15. 3190 */ 3191int task_prio(const struct task_struct *p) 3192{ 3193 return p->prio - MAX_RT_PRIO; 3194} 3195 3196/** 3197 * task_nice - return the nice value of a given task. 3198 * @p: the task in question. 3199 * 3200 * Return: The nice value [ -20 ... 0 ... 19 ]. 3201 */ 3202int task_nice(const struct task_struct *p) 3203{ 3204 return TASK_NICE(p); 3205} 3206EXPORT_SYMBOL(task_nice); 3207 3208/** 3209 * idle_cpu - is a given cpu idle currently? 3210 * @cpu: the processor in question. 3211 * 3212 * Return: 1 if the CPU is currently idle. 0 otherwise. 3213 */ 3214int idle_cpu(int cpu) 3215{ 3216 struct rq *rq = cpu_rq(cpu); 3217 3218 if (rq->curr != rq->idle) 3219 return 0; 3220 3221 if (rq->nr_running) 3222 return 0; 3223 3224#ifdef CONFIG_SMP 3225 if (!llist_empty(&rq->wake_list)) 3226 return 0; 3227#endif 3228 3229 return 1; 3230} 3231 3232/** 3233 * idle_task - return the idle task for a given cpu. 3234 * @cpu: the processor in question. 3235 * 3236 * Return: The idle task for the cpu @cpu. 3237 */ 3238struct task_struct *idle_task(int cpu) 3239{ 3240 return cpu_rq(cpu)->idle; 3241} 3242 3243/** 3244 * find_process_by_pid - find a process with a matching PID value. 3245 * @pid: the pid in question. 3246 * 3247 * The task of @pid, if found. %NULL otherwise. 3248 */ 3249static struct task_struct *find_process_by_pid(pid_t pid) 3250{ 3251 return pid ? find_task_by_vpid(pid) : current; 3252} 3253 3254/* Actually do priority change: must hold rq lock. */ 3255static void 3256__setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio) 3257{ 3258 p->policy = policy; 3259 p->rt_priority = prio; 3260 p->normal_prio = normal_prio(p); 3261 /* we are holding p->pi_lock already */ 3262 p->prio = rt_mutex_getprio(p); 3263 if (rt_prio(p->prio)) 3264 p->sched_class = &rt_sched_class; 3265 else 3266 p->sched_class = &fair_sched_class; 3267 set_load_weight(p); 3268} 3269 3270/* 3271 * check the target process has a UID that matches the current process's 3272 */ 3273static bool check_same_owner(struct task_struct *p) 3274{ 3275 const struct cred *cred = current_cred(), *pcred; 3276 bool match; 3277 3278 rcu_read_lock(); 3279 pcred = __task_cred(p); 3280 match = (uid_eq(cred->euid, pcred->euid) || 3281 uid_eq(cred->euid, pcred->uid)); 3282 rcu_read_unlock(); 3283 return match; 3284} 3285 3286static int __sched_setscheduler(struct task_struct *p, int policy, 3287 const struct sched_param *param, bool user) 3288{ 3289 int retval, oldprio, oldpolicy = -1, on_rq, running; 3290 unsigned long flags; 3291 const struct sched_class *prev_class; 3292 struct rq *rq; 3293 int reset_on_fork; 3294 3295 /* may grab non-irq protected spin_locks */ 3296 BUG_ON(in_interrupt()); 3297recheck: 3298 /* double check policy once rq lock held */ 3299 if (policy < 0) { 3300 reset_on_fork = p->sched_reset_on_fork; 3301 policy = oldpolicy = p->policy; 3302 } else { 3303 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK); 3304 policy &= ~SCHED_RESET_ON_FORK; 3305 3306 if (policy != SCHED_FIFO && policy != SCHED_RR && 3307 policy != SCHED_NORMAL && policy != SCHED_BATCH && 3308 policy != SCHED_IDLE) 3309 return -EINVAL; 3310 } 3311 3312 /* 3313 * Valid priorities for SCHED_FIFO and SCHED_RR are 3314 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL, 3315 * SCHED_BATCH and SCHED_IDLE is 0. 3316 */ 3317 if (param->sched_priority < 0 || 3318 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) || 3319 (!p->mm && param->sched_priority > MAX_RT_PRIO-1)) 3320 return -EINVAL; 3321 if (rt_policy(policy) != (param->sched_priority != 0)) 3322 return -EINVAL; 3323 3324 /* 3325 * Allow unprivileged RT tasks to decrease priority: 3326 */ 3327 if (user && !capable(CAP_SYS_NICE)) { 3328 if (rt_policy(policy)) { 3329 unsigned long rlim_rtprio = 3330 task_rlimit(p, RLIMIT_RTPRIO); 3331 3332 /* can't set/change the rt policy */ 3333 if (policy != p->policy && !rlim_rtprio) 3334 return -EPERM; 3335 3336 /* can't increase priority */ 3337 if (param->sched_priority > p->rt_priority && 3338 param->sched_priority > rlim_rtprio) 3339 return -EPERM; 3340 } 3341 3342 /* 3343 * Treat SCHED_IDLE as nice 20. Only allow a switch to 3344 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it. 3345 */ 3346 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) { 3347 if (!can_nice(p, TASK_NICE(p))) 3348 return -EPERM; 3349 } 3350 3351 /* can't change other user's priorities */ 3352 if (!check_same_owner(p)) 3353 return -EPERM; 3354 3355 /* Normal users shall not reset the sched_reset_on_fork flag */ 3356 if (p->sched_reset_on_fork && !reset_on_fork) 3357 return -EPERM; 3358 } 3359 3360 if (user) { 3361 retval = security_task_setscheduler(p); 3362 if (retval) 3363 return retval; 3364 } 3365 3366 /* 3367 * make sure no PI-waiters arrive (or leave) while we are 3368 * changing the priority of the task: 3369 * 3370 * To be able to change p->policy safely, the appropriate 3371 * runqueue lock must be held. 3372 */ 3373 rq = task_rq_lock(p, &flags); 3374 3375 /* 3376 * Changing the policy of the stop threads its a very bad idea 3377 */ 3378 if (p == rq->stop) { 3379 task_rq_unlock(rq, p, &flags); 3380 return -EINVAL; 3381 } 3382 3383 /* 3384 * If not changing anything there's no need to proceed further: 3385 */ 3386 if (unlikely(policy == p->policy && (!rt_policy(policy) || 3387 param->sched_priority == p->rt_priority))) { 3388 task_rq_unlock(rq, p, &flags); 3389 return 0; 3390 } 3391 3392#ifdef CONFIG_RT_GROUP_SCHED 3393 if (user) { 3394 /* 3395 * Do not allow realtime tasks into groups that have no runtime 3396 * assigned. 3397 */ 3398 if (rt_bandwidth_enabled() && rt_policy(policy) && 3399 task_group(p)->rt_bandwidth.rt_runtime == 0 && 3400 !task_group_is_autogroup(task_group(p))) { 3401 task_rq_unlock(rq, p, &flags); 3402 return -EPERM; 3403 } 3404 } 3405#endif 3406 3407 /* recheck policy now with rq lock held */ 3408 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) { 3409 policy = oldpolicy = -1; 3410 task_rq_unlock(rq, p, &flags); 3411 goto recheck; 3412 } 3413 on_rq = p->on_rq; 3414 running = task_current(rq, p); 3415 if (on_rq) 3416 dequeue_task(rq, p, 0); 3417 if (running) 3418 p->sched_class->put_prev_task(rq, p); 3419 3420 p->sched_reset_on_fork = reset_on_fork; 3421 3422 oldprio = p->prio; 3423 prev_class = p->sched_class; 3424 __setscheduler(rq, p, policy, param->sched_priority); 3425 3426 if (running) 3427 p->sched_class->set_curr_task(rq); 3428 if (on_rq) 3429 enqueue_task(rq, p, 0); 3430 3431 check_class_changed(rq, p, prev_class, oldprio); 3432 task_rq_unlock(rq, p, &flags); 3433 3434 rt_mutex_adjust_pi(p); 3435 3436 return 0; 3437} 3438 3439/** 3440 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread. 3441 * @p: the task in question. 3442 * @policy: new policy. 3443 * @param: structure containing the new RT priority. 3444 * 3445 * Return: 0 on success. An error code otherwise. 3446 * 3447 * NOTE that the task may be already dead. 3448 */ 3449int sched_setscheduler(struct task_struct *p, int policy, 3450 const struct sched_param *param) 3451{ 3452 return __sched_setscheduler(p, policy, param, true); 3453} 3454EXPORT_SYMBOL_GPL(sched_setscheduler); 3455 3456/** 3457 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace. 3458 * @p: the task in question. 3459 * @policy: new policy. 3460 * @param: structure containing the new RT priority. 3461 * 3462 * Just like sched_setscheduler, only don't bother checking if the 3463 * current context has permission. For example, this is needed in 3464 * stop_machine(): we create temporary high priority worker threads, 3465 * but our caller might not have that capability. 3466 * 3467 * Return: 0 on success. An error code otherwise. 3468 */ 3469int sched_setscheduler_nocheck(struct task_struct *p, int policy, 3470 const struct sched_param *param) 3471{ 3472 return __sched_setscheduler(p, policy, param, false); 3473} 3474 3475static int 3476do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param) 3477{ 3478 struct sched_param lparam; 3479 struct task_struct *p; 3480 int retval; 3481 3482 if (!param || pid < 0) 3483 return -EINVAL; 3484 if (copy_from_user(&lparam, param, sizeof(struct sched_param))) 3485 return -EFAULT; 3486 3487 rcu_read_lock(); 3488 retval = -ESRCH; 3489 p = find_process_by_pid(pid); 3490 if (p != NULL) 3491 retval = sched_setscheduler(p, policy, &lparam); 3492 rcu_read_unlock(); 3493 3494 return retval; 3495} 3496 3497/** 3498 * sys_sched_setscheduler - set/change the scheduler policy and RT priority 3499 * @pid: the pid in question. 3500 * @policy: new policy. 3501 * @param: structure containing the new RT priority. 3502 * 3503 * Return: 0 on success. An error code otherwise. 3504 */ 3505SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy, 3506 struct sched_param __user *, param) 3507{ 3508 /* negative values for policy are not valid */ 3509 if (policy < 0) 3510 return -EINVAL; 3511 3512 return do_sched_setscheduler(pid, policy, param); 3513} 3514 3515/** 3516 * sys_sched_setparam - set/change the RT priority of a thread 3517 * @pid: the pid in question. 3518 * @param: structure containing the new RT priority. 3519 * 3520 * Return: 0 on success. An error code otherwise. 3521 */ 3522SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param) 3523{ 3524 return do_sched_setscheduler(pid, -1, param); 3525} 3526 3527/** 3528 * sys_sched_getscheduler - get the policy (scheduling class) of a thread 3529 * @pid: the pid in question. 3530 * 3531 * Return: On success, the policy of the thread. Otherwise, a negative error 3532 * code. 3533 */ 3534SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid) 3535{ 3536 struct task_struct *p; 3537 int retval; 3538 3539 if (pid < 0) 3540 return -EINVAL; 3541 3542 retval = -ESRCH; 3543 rcu_read_lock(); 3544 p = find_process_by_pid(pid); 3545 if (p) { 3546 retval = security_task_getscheduler(p); 3547 if (!retval) 3548 retval = p->policy 3549 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0); 3550 } 3551 rcu_read_unlock(); 3552 return retval; 3553} 3554 3555/** 3556 * sys_sched_getparam - get the RT priority of a thread 3557 * @pid: the pid in question. 3558 * @param: structure containing the RT priority. 3559 * 3560 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error 3561 * code. 3562 */ 3563SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param) 3564{ 3565 struct sched_param lp; 3566 struct task_struct *p; 3567 int retval; 3568 3569 if (!param || pid < 0) 3570 return -EINVAL; 3571 3572 rcu_read_lock(); 3573 p = find_process_by_pid(pid); 3574 retval = -ESRCH; 3575 if (!p) 3576 goto out_unlock; 3577 3578 retval = security_task_getscheduler(p); 3579 if (retval) 3580 goto out_unlock; 3581 3582 lp.sched_priority = p->rt_priority; 3583 rcu_read_unlock(); 3584 3585 /* 3586 * This one might sleep, we cannot do it with a spinlock held ... 3587 */ 3588 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0; 3589 3590 return retval; 3591 3592out_unlock: 3593 rcu_read_unlock(); 3594 return retval; 3595} 3596 3597long sched_setaffinity(pid_t pid, const struct cpumask *in_mask) 3598{ 3599 cpumask_var_t cpus_allowed, new_mask; 3600 struct task_struct *p; 3601 int retval; 3602 3603 get_online_cpus(); 3604 rcu_read_lock(); 3605 3606 p = find_process_by_pid(pid); 3607 if (!p) { 3608 rcu_read_unlock(); 3609 put_online_cpus(); 3610 return -ESRCH; 3611 } 3612 3613 /* Prevent p going away */ 3614 get_task_struct(p); 3615 rcu_read_unlock(); 3616 3617 if (p->flags & PF_NO_SETAFFINITY) { 3618 retval = -EINVAL; 3619 goto out_put_task; 3620 } 3621 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) { 3622 retval = -ENOMEM; 3623 goto out_put_task; 3624 } 3625 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) { 3626 retval = -ENOMEM; 3627 goto out_free_cpus_allowed; 3628 } 3629 retval = -EPERM; 3630 if (!check_same_owner(p)) { 3631 rcu_read_lock(); 3632 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) { 3633 rcu_read_unlock(); 3634 goto out_unlock; 3635 } 3636 rcu_read_unlock(); 3637 } 3638 3639 retval = security_task_setscheduler(p); 3640 if (retval) 3641 goto out_unlock; 3642 3643 cpuset_cpus_allowed(p, cpus_allowed); 3644 cpumask_and(new_mask, in_mask, cpus_allowed); 3645again: 3646 retval = set_cpus_allowed_ptr(p, new_mask); 3647 3648 if (!retval) { 3649 cpuset_cpus_allowed(p, cpus_allowed); 3650 if (!cpumask_subset(new_mask, cpus_allowed)) { 3651 /* 3652 * We must have raced with a concurrent cpuset 3653 * update. Just reset the cpus_allowed to the 3654 * cpuset's cpus_allowed 3655 */ 3656 cpumask_copy(new_mask, cpus_allowed); 3657 goto again; 3658 } 3659 } 3660out_unlock: 3661 free_cpumask_var(new_mask); 3662out_free_cpus_allowed: 3663 free_cpumask_var(cpus_allowed); 3664out_put_task: 3665 put_task_struct(p); 3666 put_online_cpus(); 3667 return retval; 3668} 3669 3670static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len, 3671 struct cpumask *new_mask) 3672{ 3673 if (len < cpumask_size()) 3674 cpumask_clear(new_mask); 3675 else if (len > cpumask_size()) 3676 len = cpumask_size(); 3677 3678 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0; 3679} 3680 3681/** 3682 * sys_sched_setaffinity - set the cpu affinity of a process 3683 * @pid: pid of the process 3684 * @len: length in bytes of the bitmask pointed to by user_mask_ptr 3685 * @user_mask_ptr: user-space pointer to the new cpu mask 3686 * 3687 * Return: 0 on success. An error code otherwise. 3688 */ 3689SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len, 3690 unsigned long __user *, user_mask_ptr) 3691{ 3692 cpumask_var_t new_mask; 3693 int retval; 3694 3695 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) 3696 return -ENOMEM; 3697 3698 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask); 3699 if (retval == 0) 3700 retval = sched_setaffinity(pid, new_mask); 3701 free_cpumask_var(new_mask); 3702 return retval; 3703} 3704 3705long sched_getaffinity(pid_t pid, struct cpumask *mask) 3706{ 3707 struct task_struct *p; 3708 unsigned long flags; 3709 int retval; 3710 3711 get_online_cpus(); 3712 rcu_read_lock(); 3713 3714 retval = -ESRCH; 3715 p = find_process_by_pid(pid); 3716 if (!p) 3717 goto out_unlock; 3718 3719 retval = security_task_getscheduler(p); 3720 if (retval) 3721 goto out_unlock; 3722 3723 raw_spin_lock_irqsave(&p->pi_lock, flags); 3724 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask); 3725 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 3726 3727out_unlock: 3728 rcu_read_unlock(); 3729 put_online_cpus(); 3730 3731 return retval; 3732} 3733 3734/** 3735 * sys_sched_getaffinity - get the cpu affinity of a process 3736 * @pid: pid of the process 3737 * @len: length in bytes of the bitmask pointed to by user_mask_ptr 3738 * @user_mask_ptr: user-space pointer to hold the current cpu mask 3739 * 3740 * Return: 0 on success. An error code otherwise. 3741 */ 3742SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len, 3743 unsigned long __user *, user_mask_ptr) 3744{ 3745 int ret; 3746 cpumask_var_t mask; 3747 3748 if ((len * BITS_PER_BYTE) < nr_cpu_ids) 3749 return -EINVAL; 3750 if (len & (sizeof(unsigned long)-1)) 3751 return -EINVAL; 3752 3753 if (!alloc_cpumask_var(&mask, GFP_KERNEL)) 3754 return -ENOMEM; 3755 3756 ret = sched_getaffinity(pid, mask); 3757 if (ret == 0) { 3758 size_t retlen = min_t(size_t, len, cpumask_size()); 3759 3760 if (copy_to_user(user_mask_ptr, mask, retlen)) 3761 ret = -EFAULT; 3762 else 3763 ret = retlen; 3764 } 3765 free_cpumask_var(mask); 3766 3767 return ret; 3768} 3769 3770/** 3771 * sys_sched_yield - yield the current processor to other threads. 3772 * 3773 * This function yields the current CPU to other tasks. If there are no 3774 * other threads running on this CPU then this function will return. 3775 * 3776 * Return: 0. 3777 */ 3778SYSCALL_DEFINE0(sched_yield) 3779{ 3780 struct rq *rq = this_rq_lock(); 3781 3782 schedstat_inc(rq, yld_count); 3783 current->sched_class->yield_task(rq); 3784 3785 /* 3786 * Since we are going to call schedule() anyway, there's 3787 * no need to preempt or enable interrupts: 3788 */ 3789 __release(rq->lock); 3790 spin_release(&rq->lock.dep_map, 1, _THIS_IP_); 3791 do_raw_spin_unlock(&rq->lock); 3792 sched_preempt_enable_no_resched(); 3793 3794 schedule(); 3795 3796 return 0; 3797} 3798 3799static void __cond_resched(void) 3800{ 3801 __preempt_count_add(PREEMPT_ACTIVE); 3802 __schedule(); 3803 __preempt_count_sub(PREEMPT_ACTIVE); 3804} 3805 3806int __sched _cond_resched(void) 3807{ 3808 if (should_resched()) { 3809 __cond_resched(); 3810 return 1; 3811 } 3812 return 0; 3813} 3814EXPORT_SYMBOL(_cond_resched); 3815 3816/* 3817 * __cond_resched_lock() - if a reschedule is pending, drop the given lock, 3818 * call schedule, and on return reacquire the lock. 3819 * 3820 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level 3821 * operations here to prevent schedule() from being called twice (once via 3822 * spin_unlock(), once by hand). 3823 */ 3824int __cond_resched_lock(spinlock_t *lock) 3825{ 3826 int resched = should_resched(); 3827 int ret = 0; 3828 3829 lockdep_assert_held(lock); 3830 3831 if (spin_needbreak(lock) || resched) { 3832 spin_unlock(lock); 3833 if (resched) 3834 __cond_resched(); 3835 else 3836 cpu_relax(); 3837 ret = 1; 3838 spin_lock(lock); 3839 } 3840 return ret; 3841} 3842EXPORT_SYMBOL(__cond_resched_lock); 3843 3844int __sched __cond_resched_softirq(void) 3845{ 3846 BUG_ON(!in_softirq()); 3847 3848 if (should_resched()) { 3849 local_bh_enable(); 3850 __cond_resched(); 3851 local_bh_disable(); 3852 return 1; 3853 } 3854 return 0; 3855} 3856EXPORT_SYMBOL(__cond_resched_softirq); 3857 3858/** 3859 * yield - yield the current processor to other threads. 3860 * 3861 * Do not ever use this function, there's a 99% chance you're doing it wrong. 3862 * 3863 * The scheduler is at all times free to pick the calling task as the most 3864 * eligible task to run, if removing the yield() call from your code breaks 3865 * it, its already broken. 3866 * 3867 * Typical broken usage is: 3868 * 3869 * while (!event) 3870 * yield(); 3871 * 3872 * where one assumes that yield() will let 'the other' process run that will 3873 * make event true. If the current task is a SCHED_FIFO task that will never 3874 * happen. Never use yield() as a progress guarantee!! 3875 * 3876 * If you want to use yield() to wait for something, use wait_event(). 3877 * If you want to use yield() to be 'nice' for others, use cond_resched(). 3878 * If you still want to use yield(), do not! 3879 */ 3880void __sched yield(void) 3881{ 3882 set_current_state(TASK_RUNNING); 3883 sys_sched_yield(); 3884} 3885EXPORT_SYMBOL(yield); 3886 3887/** 3888 * yield_to - yield the current processor to another thread in 3889 * your thread group, or accelerate that thread toward the 3890 * processor it's on. 3891 * @p: target task 3892 * @preempt: whether task preemption is allowed or not 3893 * 3894 * It's the caller's job to ensure that the target task struct 3895 * can't go away on us before we can do any checks. 3896 * 3897 * Return: 3898 * true (>0) if we indeed boosted the target task. 3899 * false (0) if we failed to boost the target. 3900 * -ESRCH if there's no task to yield to. 3901 */ 3902bool __sched yield_to(struct task_struct *p, bool preempt) 3903{ 3904 struct task_struct *curr = current; 3905 struct rq *rq, *p_rq; 3906 unsigned long flags; 3907 int yielded = 0; 3908 3909 local_irq_save(flags); 3910 rq = this_rq(); 3911 3912again: 3913 p_rq = task_rq(p); 3914 /* 3915 * If we're the only runnable task on the rq and target rq also 3916 * has only one task, there's absolutely no point in yielding. 3917 */ 3918 if (rq->nr_running == 1 && p_rq->nr_running == 1) { 3919 yielded = -ESRCH; 3920 goto out_irq; 3921 } 3922 3923 double_rq_lock(rq, p_rq); 3924 while (task_rq(p) != p_rq) { 3925 double_rq_unlock(rq, p_rq); 3926 goto again; 3927 } 3928 3929 if (!curr->sched_class->yield_to_task) 3930 goto out_unlock; 3931 3932 if (curr->sched_class != p->sched_class) 3933 goto out_unlock; 3934 3935 if (task_running(p_rq, p) || p->state) 3936 goto out_unlock; 3937 3938 yielded = curr->sched_class->yield_to_task(rq, p, preempt); 3939 if (yielded) { 3940 schedstat_inc(rq, yld_count); 3941 /* 3942 * Make p's CPU reschedule; pick_next_entity takes care of 3943 * fairness. 3944 */ 3945 if (preempt && rq != p_rq) 3946 resched_task(p_rq->curr); 3947 } 3948 3949out_unlock: 3950 double_rq_unlock(rq, p_rq); 3951out_irq: 3952 local_irq_restore(flags); 3953 3954 if (yielded > 0) 3955 schedule(); 3956 3957 return yielded; 3958} 3959EXPORT_SYMBOL_GPL(yield_to); 3960 3961/* 3962 * This task is about to go to sleep on IO. Increment rq->nr_iowait so 3963 * that process accounting knows that this is a task in IO wait state. 3964 */ 3965void __sched io_schedule(void) 3966{ 3967 struct rq *rq = raw_rq(); 3968 3969 delayacct_blkio_start(); 3970 atomic_inc(&rq->nr_iowait); 3971 blk_flush_plug(current); 3972 current->in_iowait = 1; 3973 schedule(); 3974 current->in_iowait = 0; 3975 atomic_dec(&rq->nr_iowait); 3976 delayacct_blkio_end(); 3977} 3978EXPORT_SYMBOL(io_schedule); 3979 3980long __sched io_schedule_timeout(long timeout) 3981{ 3982 struct rq *rq = raw_rq(); 3983 long ret; 3984 3985 delayacct_blkio_start(); 3986 atomic_inc(&rq->nr_iowait); 3987 blk_flush_plug(current); 3988 current->in_iowait = 1; 3989 ret = schedule_timeout(timeout); 3990 current->in_iowait = 0; 3991 atomic_dec(&rq->nr_iowait); 3992 delayacct_blkio_end(); 3993 return ret; 3994} 3995 3996/** 3997 * sys_sched_get_priority_max - return maximum RT priority. 3998 * @policy: scheduling class. 3999 * 4000 * Return: On success, this syscall returns the maximum 4001 * rt_priority that can be used by a given scheduling class. 4002 * On failure, a negative error code is returned. 4003 */ 4004SYSCALL_DEFINE1(sched_get_priority_max, int, policy) 4005{ 4006 int ret = -EINVAL; 4007 4008 switch (policy) { 4009 case SCHED_FIFO: 4010 case SCHED_RR: 4011 ret = MAX_USER_RT_PRIO-1; 4012 break; 4013 case SCHED_NORMAL: 4014 case SCHED_BATCH: 4015 case SCHED_IDLE: 4016 ret = 0; 4017 break; 4018 } 4019 return ret; 4020} 4021 4022/** 4023 * sys_sched_get_priority_min - return minimum RT priority. 4024 * @policy: scheduling class. 4025 * 4026 * Return: On success, this syscall returns the minimum 4027 * rt_priority that can be used by a given scheduling class. 4028 * On failure, a negative error code is returned. 4029 */ 4030SYSCALL_DEFINE1(sched_get_priority_min, int, policy) 4031{ 4032 int ret = -EINVAL; 4033 4034 switch (policy) { 4035 case SCHED_FIFO: 4036 case SCHED_RR: 4037 ret = 1; 4038 break; 4039 case SCHED_NORMAL: 4040 case SCHED_BATCH: 4041 case SCHED_IDLE: 4042 ret = 0; 4043 } 4044 return ret; 4045} 4046 4047/** 4048 * sys_sched_rr_get_interval - return the default timeslice of a process. 4049 * @pid: pid of the process. 4050 * @interval: userspace pointer to the timeslice value. 4051 * 4052 * this syscall writes the default timeslice value of a given process 4053 * into the user-space timespec buffer. A value of '0' means infinity. 4054 * 4055 * Return: On success, 0 and the timeslice is in @interval. Otherwise, 4056 * an error code. 4057 */ 4058SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid, 4059 struct timespec __user *, interval) 4060{ 4061 struct task_struct *p; 4062 unsigned int time_slice; 4063 unsigned long flags; 4064 struct rq *rq; 4065 int retval; 4066 struct timespec t; 4067 4068 if (pid < 0) 4069 return -EINVAL; 4070 4071 retval = -ESRCH; 4072 rcu_read_lock(); 4073 p = find_process_by_pid(pid); 4074 if (!p) 4075 goto out_unlock; 4076 4077 retval = security_task_getscheduler(p); 4078 if (retval) 4079 goto out_unlock; 4080 4081 rq = task_rq_lock(p, &flags); 4082 time_slice = p->sched_class->get_rr_interval(rq, p); 4083 task_rq_unlock(rq, p, &flags); 4084 4085 rcu_read_unlock(); 4086 jiffies_to_timespec(time_slice, &t); 4087 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0; 4088 return retval; 4089 4090out_unlock: 4091 rcu_read_unlock(); 4092 return retval; 4093} 4094 4095static const char stat_nam[] = TASK_STATE_TO_CHAR_STR; 4096 4097void sched_show_task(struct task_struct *p) 4098{ 4099 unsigned long free = 0; 4100 int ppid; 4101 unsigned state; 4102 4103 state = p->state ? __ffs(p->state) + 1 : 0; 4104 printk(KERN_INFO "%-15.15s %c", p->comm, 4105 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?'); 4106#if BITS_PER_LONG == 32 4107 if (state == TASK_RUNNING) 4108 printk(KERN_CONT " running "); 4109 else 4110 printk(KERN_CONT " %08lx ", thread_saved_pc(p)); 4111#else 4112 if (state == TASK_RUNNING) 4113 printk(KERN_CONT " running task "); 4114 else 4115 printk(KERN_CONT " %016lx ", thread_saved_pc(p)); 4116#endif 4117#ifdef CONFIG_DEBUG_STACK_USAGE 4118 free = stack_not_used(p); 4119#endif 4120 rcu_read_lock(); 4121 ppid = task_pid_nr(rcu_dereference(p->real_parent)); 4122 rcu_read_unlock(); 4123 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free, 4124 task_pid_nr(p), ppid, 4125 (unsigned long)task_thread_info(p)->flags); 4126 4127 print_worker_info(KERN_INFO, p); 4128 show_stack(p, NULL); 4129} 4130 4131void show_state_filter(unsigned long state_filter) 4132{ 4133 struct task_struct *g, *p; 4134 4135#if BITS_PER_LONG == 32 4136 printk(KERN_INFO 4137 " task PC stack pid father\n"); 4138#else 4139 printk(KERN_INFO 4140 " task PC stack pid father\n"); 4141#endif 4142 rcu_read_lock(); 4143 do_each_thread(g, p) { 4144 /* 4145 * reset the NMI-timeout, listing all files on a slow 4146 * console might take a lot of time: 4147 */ 4148 touch_nmi_watchdog(); 4149 if (!state_filter || (p->state & state_filter)) 4150 sched_show_task(p); 4151 } while_each_thread(g, p); 4152 4153 touch_all_softlockup_watchdogs(); 4154 4155#ifdef CONFIG_SCHED_DEBUG 4156 sysrq_sched_debug_show(); 4157#endif 4158 rcu_read_unlock(); 4159 /* 4160 * Only show locks if all tasks are dumped: 4161 */ 4162 if (!state_filter) 4163 debug_show_all_locks(); 4164} 4165 4166void init_idle_bootup_task(struct task_struct *idle) 4167{ 4168 idle->sched_class = &idle_sched_class; 4169} 4170 4171/** 4172 * init_idle - set up an idle thread for a given CPU 4173 * @idle: task in question 4174 * @cpu: cpu the idle task belongs to 4175 * 4176 * NOTE: this function does not set the idle thread's NEED_RESCHED 4177 * flag, to make booting more robust. 4178 */ 4179void init_idle(struct task_struct *idle, int cpu) 4180{ 4181 struct rq *rq = cpu_rq(cpu); 4182 unsigned long flags; 4183 4184 raw_spin_lock_irqsave(&rq->lock, flags); 4185 4186 __sched_fork(idle); 4187 idle->state = TASK_RUNNING; 4188 idle->se.exec_start = sched_clock(); 4189 4190 do_set_cpus_allowed(idle, cpumask_of(cpu)); 4191 /* 4192 * We're having a chicken and egg problem, even though we are 4193 * holding rq->lock, the cpu isn't yet set to this cpu so the 4194 * lockdep check in task_group() will fail. 4195 * 4196 * Similar case to sched_fork(). / Alternatively we could 4197 * use task_rq_lock() here and obtain the other rq->lock. 4198 * 4199 * Silence PROVE_RCU 4200 */ 4201 rcu_read_lock(); 4202 __set_task_cpu(idle, cpu); 4203 rcu_read_unlock(); 4204 4205 rq->curr = rq->idle = idle; 4206#if defined(CONFIG_SMP) 4207 idle->on_cpu = 1; 4208#endif 4209 raw_spin_unlock_irqrestore(&rq->lock, flags); 4210 4211 /* Set the preempt count _outside_ the spinlocks! */ 4212 init_idle_preempt_count(idle, cpu); 4213 4214 /* 4215 * The idle tasks have their own, simple scheduling class: 4216 */ 4217 idle->sched_class = &idle_sched_class; 4218 ftrace_graph_init_idle_task(idle, cpu); 4219 vtime_init_idle(idle, cpu); 4220#if defined(CONFIG_SMP) 4221 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu); 4222#endif 4223} 4224 4225#ifdef CONFIG_SMP 4226void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask) 4227{ 4228 if (p->sched_class && p->sched_class->set_cpus_allowed) 4229 p->sched_class->set_cpus_allowed(p, new_mask); 4230 4231 cpumask_copy(&p->cpus_allowed, new_mask); 4232 p->nr_cpus_allowed = cpumask_weight(new_mask); 4233} 4234 4235/* 4236 * This is how migration works: 4237 * 4238 * 1) we invoke migration_cpu_stop() on the target CPU using 4239 * stop_one_cpu(). 4240 * 2) stopper starts to run (implicitly forcing the migrated thread 4241 * off the CPU) 4242 * 3) it checks whether the migrated task is still in the wrong runqueue. 4243 * 4) if it's in the wrong runqueue then the migration thread removes 4244 * it and puts it into the right queue. 4245 * 5) stopper completes and stop_one_cpu() returns and the migration 4246 * is done. 4247 */ 4248 4249/* 4250 * Change a given task's CPU affinity. Migrate the thread to a 4251 * proper CPU and schedule it away if the CPU it's executing on 4252 * is removed from the allowed bitmask. 4253 * 4254 * NOTE: the caller must have a valid reference to the task, the 4255 * task must not exit() & deallocate itself prematurely. The 4256 * call is not atomic; no spinlocks may be held. 4257 */ 4258int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask) 4259{ 4260 unsigned long flags; 4261 struct rq *rq; 4262 unsigned int dest_cpu; 4263 int ret = 0; 4264 4265 rq = task_rq_lock(p, &flags); 4266 4267 if (cpumask_equal(&p->cpus_allowed, new_mask)) 4268 goto out; 4269 4270 if (!cpumask_intersects(new_mask, cpu_active_mask)) { 4271 ret = -EINVAL; 4272 goto out; 4273 } 4274 4275 do_set_cpus_allowed(p, new_mask); 4276 4277 /* Can the task run on the task's current CPU? If so, we're done */ 4278 if (cpumask_test_cpu(task_cpu(p), new_mask)) 4279 goto out; 4280 4281 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask); 4282 if (p->on_rq) { 4283 struct migration_arg arg = { p, dest_cpu }; 4284 /* Need help from migration thread: drop lock and wait. */ 4285 task_rq_unlock(rq, p, &flags); 4286 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg); 4287 tlb_migrate_finish(p->mm); 4288 return 0; 4289 } 4290out: 4291 task_rq_unlock(rq, p, &flags); 4292 4293 return ret; 4294} 4295EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr); 4296 4297/* 4298 * Move (not current) task off this cpu, onto dest cpu. We're doing 4299 * this because either it can't run here any more (set_cpus_allowed() 4300 * away from this CPU, or CPU going down), or because we're 4301 * attempting to rebalance this task on exec (sched_exec). 4302 * 4303 * So we race with normal scheduler movements, but that's OK, as long 4304 * as the task is no longer on this CPU. 4305 * 4306 * Returns non-zero if task was successfully migrated. 4307 */ 4308static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu) 4309{ 4310 struct rq *rq_dest, *rq_src; 4311 int ret = 0; 4312 4313 if (unlikely(!cpu_active(dest_cpu))) 4314 return ret; 4315 4316 rq_src = cpu_rq(src_cpu); 4317 rq_dest = cpu_rq(dest_cpu); 4318 4319 raw_spin_lock(&p->pi_lock); 4320 double_rq_lock(rq_src, rq_dest); 4321 /* Already moved. */ 4322 if (task_cpu(p) != src_cpu) 4323 goto done; 4324 /* Affinity changed (again). */ 4325 if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p))) 4326 goto fail; 4327 4328 /* 4329 * If we're not on a rq, the next wake-up will ensure we're 4330 * placed properly. 4331 */ 4332 if (p->on_rq) { 4333 dequeue_task(rq_src, p, 0); 4334 set_task_cpu(p, dest_cpu); 4335 enqueue_task(rq_dest, p, 0); 4336 check_preempt_curr(rq_dest, p, 0); 4337 } 4338done: 4339 ret = 1; 4340fail: 4341 double_rq_unlock(rq_src, rq_dest); 4342 raw_spin_unlock(&p->pi_lock); 4343 return ret; 4344} 4345 4346/* 4347 * migration_cpu_stop - this will be executed by a highprio stopper thread 4348 * and performs thread migration by bumping thread off CPU then 4349 * 'pushing' onto another runqueue. 4350 */ 4351static int migration_cpu_stop(void *data) 4352{ 4353 struct migration_arg *arg = data; 4354 4355 /* 4356 * The original target cpu might have gone down and we might 4357 * be on another cpu but it doesn't matter. 4358 */ 4359 local_irq_disable(); 4360 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu); 4361 local_irq_enable(); 4362 return 0; 4363} 4364 4365#ifdef CONFIG_HOTPLUG_CPU 4366 4367/* 4368 * Ensures that the idle task is using init_mm right before its cpu goes 4369 * offline. 4370 */ 4371void idle_task_exit(void) 4372{ 4373 struct mm_struct *mm = current->active_mm; 4374 4375 BUG_ON(cpu_online(smp_processor_id())); 4376 4377 if (mm != &init_mm) 4378 switch_mm(mm, &init_mm, current); 4379 mmdrop(mm); 4380} 4381 4382/* 4383 * Since this CPU is going 'away' for a while, fold any nr_active delta 4384 * we might have. Assumes we're called after migrate_tasks() so that the 4385 * nr_active count is stable. 4386 * 4387 * Also see the comment "Global load-average calculations". 4388 */ 4389static void calc_load_migrate(struct rq *rq) 4390{ 4391 long delta = calc_load_fold_active(rq); 4392 if (delta) 4393 atomic_long_add(delta, &calc_load_tasks); 4394} 4395 4396/* 4397 * Migrate all tasks from the rq, sleeping tasks will be migrated by 4398 * try_to_wake_up()->select_task_rq(). 4399 * 4400 * Called with rq->lock held even though we'er in stop_machine() and 4401 * there's no concurrency possible, we hold the required locks anyway 4402 * because of lock validation efforts. 4403 */ 4404static void migrate_tasks(unsigned int dead_cpu) 4405{ 4406 struct rq *rq = cpu_rq(dead_cpu); 4407 struct task_struct *next, *stop = rq->stop; 4408 int dest_cpu; 4409 4410 /* 4411 * Fudge the rq selection such that the below task selection loop 4412 * doesn't get stuck on the currently eligible stop task. 4413 * 4414 * We're currently inside stop_machine() and the rq is either stuck 4415 * in the stop_machine_cpu_stop() loop, or we're executing this code, 4416 * either way we should never end up calling schedule() until we're 4417 * done here. 4418 */ 4419 rq->stop = NULL; 4420 4421 /* 4422 * put_prev_task() and pick_next_task() sched 4423 * class method both need to have an up-to-date 4424 * value of rq->clock[_task] 4425 */ 4426 update_rq_clock(rq); 4427 4428 for ( ; ; ) { 4429 /* 4430 * There's this thread running, bail when that's the only 4431 * remaining thread. 4432 */ 4433 if (rq->nr_running == 1) 4434 break; 4435 4436 next = pick_next_task(rq); 4437 BUG_ON(!next); 4438 next->sched_class->put_prev_task(rq, next); 4439 4440 /* Find suitable destination for @next, with force if needed. */ 4441 dest_cpu = select_fallback_rq(dead_cpu, next); 4442 raw_spin_unlock(&rq->lock); 4443 4444 __migrate_task(next, dead_cpu, dest_cpu); 4445 4446 raw_spin_lock(&rq->lock); 4447 } 4448 4449 rq->stop = stop; 4450} 4451 4452#endif /* CONFIG_HOTPLUG_CPU */ 4453 4454#if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL) 4455 4456static struct ctl_table sd_ctl_dir[] = { 4457 { 4458 .procname = "sched_domain", 4459 .mode = 0555, 4460 }, 4461 {} 4462}; 4463 4464static struct ctl_table sd_ctl_root[] = { 4465 { 4466 .procname = "kernel", 4467 .mode = 0555, 4468 .child = sd_ctl_dir, 4469 }, 4470 {} 4471}; 4472 4473static struct ctl_table *sd_alloc_ctl_entry(int n) 4474{ 4475 struct ctl_table *entry = 4476 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL); 4477 4478 return entry; 4479} 4480 4481static void sd_free_ctl_entry(struct ctl_table **tablep) 4482{ 4483 struct ctl_table *entry; 4484 4485 /* 4486 * In the intermediate directories, both the child directory and 4487 * procname are dynamically allocated and could fail but the mode 4488 * will always be set. In the lowest directory the names are 4489 * static strings and all have proc handlers. 4490 */ 4491 for (entry = *tablep; entry->mode; entry++) { 4492 if (entry->child) 4493 sd_free_ctl_entry(&entry->child); 4494 if (entry->proc_handler == NULL) 4495 kfree(entry->procname); 4496 } 4497 4498 kfree(*tablep); 4499 *tablep = NULL; 4500} 4501 4502static int min_load_idx = 0; 4503static int max_load_idx = CPU_LOAD_IDX_MAX-1; 4504 4505static void 4506set_table_entry(struct ctl_table *entry, 4507 const char *procname, void *data, int maxlen, 4508 umode_t mode, proc_handler *proc_handler, 4509 bool load_idx) 4510{ 4511 entry->procname = procname; 4512 entry->data = data; 4513 entry->maxlen = maxlen; 4514 entry->mode = mode; 4515 entry->proc_handler = proc_handler; 4516 4517 if (load_idx) { 4518 entry->extra1 = &min_load_idx; 4519 entry->extra2 = &max_load_idx; 4520 } 4521} 4522 4523static struct ctl_table * 4524sd_alloc_ctl_domain_table(struct sched_domain *sd) 4525{ 4526 struct ctl_table *table = sd_alloc_ctl_entry(13); 4527 4528 if (table == NULL) 4529 return NULL; 4530 4531 set_table_entry(&table[0], "min_interval", &sd->min_interval, 4532 sizeof(long), 0644, proc_doulongvec_minmax, false); 4533 set_table_entry(&table[1], "max_interval", &sd->max_interval, 4534 sizeof(long), 0644, proc_doulongvec_minmax, false); 4535 set_table_entry(&table[2], "busy_idx", &sd->busy_idx, 4536 sizeof(int), 0644, proc_dointvec_minmax, true); 4537 set_table_entry(&table[3], "idle_idx", &sd->idle_idx, 4538 sizeof(int), 0644, proc_dointvec_minmax, true); 4539 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx, 4540 sizeof(int), 0644, proc_dointvec_minmax, true); 4541 set_table_entry(&table[5], "wake_idx", &sd->wake_idx, 4542 sizeof(int), 0644, proc_dointvec_minmax, true); 4543 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx, 4544 sizeof(int), 0644, proc_dointvec_minmax, true); 4545 set_table_entry(&table[7], "busy_factor", &sd->busy_factor, 4546 sizeof(int), 0644, proc_dointvec_minmax, false); 4547 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct, 4548 sizeof(int), 0644, proc_dointvec_minmax, false); 4549 set_table_entry(&table[9], "cache_nice_tries", 4550 &sd->cache_nice_tries, 4551 sizeof(int), 0644, proc_dointvec_minmax, false); 4552 set_table_entry(&table[10], "flags", &sd->flags, 4553 sizeof(int), 0644, proc_dointvec_minmax, false); 4554 set_table_entry(&table[11], "name", sd->name, 4555 CORENAME_MAX_SIZE, 0444, proc_dostring, false); 4556 /* &table[12] is terminator */ 4557 4558 return table; 4559} 4560 4561static struct ctl_table *sd_alloc_ctl_cpu_table(int cpu) 4562{ 4563 struct ctl_table *entry, *table; 4564 struct sched_domain *sd; 4565 int domain_num = 0, i; 4566 char buf[32]; 4567 4568 for_each_domain(cpu, sd) 4569 domain_num++; 4570 entry = table = sd_alloc_ctl_entry(domain_num + 1); 4571 if (table == NULL) 4572 return NULL; 4573 4574 i = 0; 4575 for_each_domain(cpu, sd) { 4576 snprintf(buf, 32, "domain%d", i); 4577 entry->procname = kstrdup(buf, GFP_KERNEL); 4578 entry->mode = 0555; 4579 entry->child = sd_alloc_ctl_domain_table(sd); 4580 entry++; 4581 i++; 4582 } 4583 return table; 4584} 4585 4586static struct ctl_table_header *sd_sysctl_header; 4587static void register_sched_domain_sysctl(void) 4588{ 4589 int i, cpu_num = num_possible_cpus(); 4590 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1); 4591 char buf[32]; 4592 4593 WARN_ON(sd_ctl_dir[0].child); 4594 sd_ctl_dir[0].child = entry; 4595 4596 if (entry == NULL) 4597 return; 4598 4599 for_each_possible_cpu(i) { 4600 snprintf(buf, 32, "cpu%d", i); 4601 entry->procname = kstrdup(buf, GFP_KERNEL); 4602 entry->mode = 0555; 4603 entry->child = sd_alloc_ctl_cpu_table(i); 4604 entry++; 4605 } 4606 4607 WARN_ON(sd_sysctl_header); 4608 sd_sysctl_header = register_sysctl_table(sd_ctl_root); 4609} 4610 4611/* may be called multiple times per register */ 4612static void unregister_sched_domain_sysctl(void) 4613{ 4614 if (sd_sysctl_header) 4615 unregister_sysctl_table(sd_sysctl_header); 4616 sd_sysctl_header = NULL; 4617 if (sd_ctl_dir[0].child) 4618 sd_free_ctl_entry(&sd_ctl_dir[0].child); 4619} 4620#else 4621static void register_sched_domain_sysctl(void) 4622{ 4623} 4624static void unregister_sched_domain_sysctl(void) 4625{ 4626} 4627#endif 4628 4629static void set_rq_online(struct rq *rq) 4630{ 4631 if (!rq->online) { 4632 const struct sched_class *class; 4633 4634 cpumask_set_cpu(rq->cpu, rq->rd->online); 4635 rq->online = 1; 4636 4637 for_each_class(class) { 4638 if (class->rq_online) 4639 class->rq_online(rq); 4640 } 4641 } 4642} 4643 4644static void set_rq_offline(struct rq *rq) 4645{ 4646 if (rq->online) { 4647 const struct sched_class *class; 4648 4649 for_each_class(class) { 4650 if (class->rq_offline) 4651 class->rq_offline(rq); 4652 } 4653 4654 cpumask_clear_cpu(rq->cpu, rq->rd->online); 4655 rq->online = 0; 4656 } 4657} 4658 4659/* 4660 * migration_call - callback that gets triggered when a CPU is added. 4661 * Here we can start up the necessary migration thread for the new CPU. 4662 */ 4663static int 4664migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu) 4665{ 4666 int cpu = (long)hcpu; 4667 unsigned long flags; 4668 struct rq *rq = cpu_rq(cpu); 4669 4670 switch (action & ~CPU_TASKS_FROZEN) { 4671 4672 case CPU_UP_PREPARE: 4673 rq->calc_load_update = calc_load_update; 4674 break; 4675 4676 case CPU_ONLINE: 4677 /* Update our root-domain */ 4678 raw_spin_lock_irqsave(&rq->lock, flags); 4679 if (rq->rd) { 4680 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span)); 4681 4682 set_rq_online(rq); 4683 } 4684 raw_spin_unlock_irqrestore(&rq->lock, flags); 4685 break; 4686 4687#ifdef CONFIG_HOTPLUG_CPU 4688 case CPU_DYING: 4689 sched_ttwu_pending(); 4690 /* Update our root-domain */ 4691 raw_spin_lock_irqsave(&rq->lock, flags); 4692 if (rq->rd) { 4693 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span)); 4694 set_rq_offline(rq); 4695 } 4696 migrate_tasks(cpu); 4697 BUG_ON(rq->nr_running != 1); /* the migration thread */ 4698 raw_spin_unlock_irqrestore(&rq->lock, flags); 4699 break; 4700 4701 case CPU_DEAD: 4702 calc_load_migrate(rq); 4703 break; 4704#endif 4705 } 4706 4707 update_max_interval(); 4708 4709 return NOTIFY_OK; 4710} 4711 4712/* 4713 * Register at high priority so that task migration (migrate_all_tasks) 4714 * happens before everything else. This has to be lower priority than 4715 * the notifier in the perf_event subsystem, though. 4716 */ 4717static struct notifier_block migration_notifier = { 4718 .notifier_call = migration_call, 4719 .priority = CPU_PRI_MIGRATION, 4720}; 4721 4722static int sched_cpu_active(struct notifier_block *nfb, 4723 unsigned long action, void *hcpu) 4724{ 4725 switch (action & ~CPU_TASKS_FROZEN) { 4726 case CPU_STARTING: 4727 case CPU_DOWN_FAILED: 4728 set_cpu_active((long)hcpu, true); 4729 return NOTIFY_OK; 4730 default: 4731 return NOTIFY_DONE; 4732 } 4733} 4734 4735static int sched_cpu_inactive(struct notifier_block *nfb, 4736 unsigned long action, void *hcpu) 4737{ 4738 switch (action & ~CPU_TASKS_FROZEN) { 4739 case CPU_DOWN_PREPARE: 4740 set_cpu_active((long)hcpu, false); 4741 return NOTIFY_OK; 4742 default: 4743 return NOTIFY_DONE; 4744 } 4745} 4746 4747static int __init migration_init(void) 4748{ 4749 void *cpu = (void *)(long)smp_processor_id(); 4750 int err; 4751 4752 /* Initialize migration for the boot CPU */ 4753 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu); 4754 BUG_ON(err == NOTIFY_BAD); 4755 migration_call(&migration_notifier, CPU_ONLINE, cpu); 4756 register_cpu_notifier(&migration_notifier); 4757 4758 /* Register cpu active notifiers */ 4759 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE); 4760 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE); 4761 4762 return 0; 4763} 4764early_initcall(migration_init); 4765#endif 4766 4767#ifdef CONFIG_SMP 4768 4769static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */ 4770 4771#ifdef CONFIG_SCHED_DEBUG 4772 4773static __read_mostly int sched_debug_enabled; 4774 4775static int __init sched_debug_setup(char *str) 4776{ 4777 sched_debug_enabled = 1; 4778 4779 return 0; 4780} 4781early_param("sched_debug", sched_debug_setup); 4782 4783static inline bool sched_debug(void) 4784{ 4785 return sched_debug_enabled; 4786} 4787 4788static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level, 4789 struct cpumask *groupmask) 4790{ 4791 struct sched_group *group = sd->groups; 4792 char str[256]; 4793 4794 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd)); 4795 cpumask_clear(groupmask); 4796 4797 printk(KERN_DEBUG "%*s domain %d: ", level, "", level); 4798 4799 if (!(sd->flags & SD_LOAD_BALANCE)) { 4800 printk("does not load-balance\n"); 4801 if (sd->parent) 4802 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain" 4803 " has parent"); 4804 return -1; 4805 } 4806 4807 printk(KERN_CONT "span %s level %s\n", str, sd->name); 4808 4809 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) { 4810 printk(KERN_ERR "ERROR: domain->span does not contain " 4811 "CPU%d\n", cpu); 4812 } 4813 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) { 4814 printk(KERN_ERR "ERROR: domain->groups does not contain" 4815 " CPU%d\n", cpu); 4816 } 4817 4818 printk(KERN_DEBUG "%*s groups:", level + 1, ""); 4819 do { 4820 if (!group) { 4821 printk("\n"); 4822 printk(KERN_ERR "ERROR: group is NULL\n"); 4823 break; 4824 } 4825 4826 /* 4827 * Even though we initialize ->power to something semi-sane, 4828 * we leave power_orig unset. This allows us to detect if 4829 * domain iteration is still funny without causing /0 traps. 4830 */ 4831 if (!group->sgp->power_orig) { 4832 printk(KERN_CONT "\n"); 4833 printk(KERN_ERR "ERROR: domain->cpu_power not " 4834 "set\n"); 4835 break; 4836 } 4837 4838 if (!cpumask_weight(sched_group_cpus(group))) { 4839 printk(KERN_CONT "\n"); 4840 printk(KERN_ERR "ERROR: empty group\n"); 4841 break; 4842 } 4843 4844 if (!(sd->flags & SD_OVERLAP) && 4845 cpumask_intersects(groupmask, sched_group_cpus(group))) { 4846 printk(KERN_CONT "\n"); 4847 printk(KERN_ERR "ERROR: repeated CPUs\n"); 4848 break; 4849 } 4850 4851 cpumask_or(groupmask, groupmask, sched_group_cpus(group)); 4852 4853 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group)); 4854 4855 printk(KERN_CONT " %s", str); 4856 if (group->sgp->power != SCHED_POWER_SCALE) { 4857 printk(KERN_CONT " (cpu_power = %d)", 4858 group->sgp->power); 4859 } 4860 4861 group = group->next; 4862 } while (group != sd->groups); 4863 printk(KERN_CONT "\n"); 4864 4865 if (!cpumask_equal(sched_domain_span(sd), groupmask)) 4866 printk(KERN_ERR "ERROR: groups don't span domain->span\n"); 4867 4868 if (sd->parent && 4869 !cpumask_subset(groupmask, sched_domain_span(sd->parent))) 4870 printk(KERN_ERR "ERROR: parent span is not a superset " 4871 "of domain->span\n"); 4872 return 0; 4873} 4874 4875static void sched_domain_debug(struct sched_domain *sd, int cpu) 4876{ 4877 int level = 0; 4878 4879 if (!sched_debug_enabled) 4880 return; 4881 4882 if (!sd) { 4883 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu); 4884 return; 4885 } 4886 4887 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu); 4888 4889 for (;;) { 4890 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask)) 4891 break; 4892 level++; 4893 sd = sd->parent; 4894 if (!sd) 4895 break; 4896 } 4897} 4898#else /* !CONFIG_SCHED_DEBUG */ 4899# define sched_domain_debug(sd, cpu) do { } while (0) 4900static inline bool sched_debug(void) 4901{ 4902 return false; 4903} 4904#endif /* CONFIG_SCHED_DEBUG */ 4905 4906static int sd_degenerate(struct sched_domain *sd) 4907{ 4908 if (cpumask_weight(sched_domain_span(sd)) == 1) 4909 return 1; 4910 4911 /* Following flags need at least 2 groups */ 4912 if (sd->flags & (SD_LOAD_BALANCE | 4913 SD_BALANCE_NEWIDLE | 4914 SD_BALANCE_FORK | 4915 SD_BALANCE_EXEC | 4916 SD_SHARE_CPUPOWER | 4917 SD_SHARE_PKG_RESOURCES)) { 4918 if (sd->groups != sd->groups->next) 4919 return 0; 4920 } 4921 4922 /* Following flags don't use groups */ 4923 if (sd->flags & (SD_WAKE_AFFINE)) 4924 return 0; 4925 4926 return 1; 4927} 4928 4929static int 4930sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent) 4931{ 4932 unsigned long cflags = sd->flags, pflags = parent->flags; 4933 4934 if (sd_degenerate(parent)) 4935 return 1; 4936 4937 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent))) 4938 return 0; 4939 4940 /* Flags needing groups don't count if only 1 group in parent */ 4941 if (parent->groups == parent->groups->next) { 4942 pflags &= ~(SD_LOAD_BALANCE | 4943 SD_BALANCE_NEWIDLE | 4944 SD_BALANCE_FORK | 4945 SD_BALANCE_EXEC | 4946 SD_SHARE_CPUPOWER | 4947 SD_SHARE_PKG_RESOURCES | 4948 SD_PREFER_SIBLING); 4949 if (nr_node_ids == 1) 4950 pflags &= ~SD_SERIALIZE; 4951 } 4952 if (~cflags & pflags) 4953 return 0; 4954 4955 return 1; 4956} 4957 4958static void free_rootdomain(struct rcu_head *rcu) 4959{ 4960 struct root_domain *rd = container_of(rcu, struct root_domain, rcu); 4961 4962 cpupri_cleanup(&rd->cpupri); 4963 free_cpumask_var(rd->rto_mask); 4964 free_cpumask_var(rd->online); 4965 free_cpumask_var(rd->span); 4966 kfree(rd); 4967} 4968 4969static void rq_attach_root(struct rq *rq, struct root_domain *rd) 4970{ 4971 struct root_domain *old_rd = NULL; 4972 unsigned long flags; 4973 4974 raw_spin_lock_irqsave(&rq->lock, flags); 4975 4976 if (rq->rd) { 4977 old_rd = rq->rd; 4978 4979 if (cpumask_test_cpu(rq->cpu, old_rd->online)) 4980 set_rq_offline(rq); 4981 4982 cpumask_clear_cpu(rq->cpu, old_rd->span); 4983 4984 /* 4985 * If we dont want to free the old_rt yet then 4986 * set old_rd to NULL to skip the freeing later 4987 * in this function: 4988 */ 4989 if (!atomic_dec_and_test(&old_rd->refcount)) 4990 old_rd = NULL; 4991 } 4992 4993 atomic_inc(&rd->refcount); 4994 rq->rd = rd; 4995 4996 cpumask_set_cpu(rq->cpu, rd->span); 4997 if (cpumask_test_cpu(rq->cpu, cpu_active_mask)) 4998 set_rq_online(rq); 4999 5000 raw_spin_unlock_irqrestore(&rq->lock, flags); 5001 5002 if (old_rd) 5003 call_rcu_sched(&old_rd->rcu, free_rootdomain); 5004} 5005 5006static int init_rootdomain(struct root_domain *rd) 5007{ 5008 memset(rd, 0, sizeof(*rd)); 5009 5010 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL)) 5011 goto out; 5012 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL)) 5013 goto free_span; 5014 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL)) 5015 goto free_online; 5016 5017 if (cpupri_init(&rd->cpupri) != 0) 5018 goto free_rto_mask; 5019 return 0; 5020 5021free_rto_mask: 5022 free_cpumask_var(rd->rto_mask); 5023free_online: 5024 free_cpumask_var(rd->online); 5025free_span: 5026 free_cpumask_var(rd->span); 5027out: 5028 return -ENOMEM; 5029} 5030 5031/* 5032 * By default the system creates a single root-domain with all cpus as 5033 * members (mimicking the global state we have today). 5034 */ 5035struct root_domain def_root_domain; 5036 5037static void init_defrootdomain(void) 5038{ 5039 init_rootdomain(&def_root_domain); 5040 5041 atomic_set(&def_root_domain.refcount, 1); 5042} 5043 5044static struct root_domain *alloc_rootdomain(void) 5045{ 5046 struct root_domain *rd; 5047 5048 rd = kmalloc(sizeof(*rd), GFP_KERNEL); 5049 if (!rd) 5050 return NULL; 5051 5052 if (init_rootdomain(rd) != 0) { 5053 kfree(rd); 5054 return NULL; 5055 } 5056 5057 return rd; 5058} 5059 5060static void free_sched_groups(struct sched_group *sg, int free_sgp) 5061{ 5062 struct sched_group *tmp, *first; 5063 5064 if (!sg) 5065 return; 5066 5067 first = sg; 5068 do { 5069 tmp = sg->next; 5070 5071 if (free_sgp && atomic_dec_and_test(&sg->sgp->ref)) 5072 kfree(sg->sgp); 5073 5074 kfree(sg); 5075 sg = tmp; 5076 } while (sg != first); 5077} 5078 5079static void free_sched_domain(struct rcu_head *rcu) 5080{ 5081 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu); 5082 5083 /* 5084 * If its an overlapping domain it has private groups, iterate and 5085 * nuke them all. 5086 */ 5087 if (sd->flags & SD_OVERLAP) { 5088 free_sched_groups(sd->groups, 1); 5089 } else if (atomic_dec_and_test(&sd->groups->ref)) { 5090 kfree(sd->groups->sgp); 5091 kfree(sd->groups); 5092 } 5093 kfree(sd); 5094} 5095 5096static void destroy_sched_domain(struct sched_domain *sd, int cpu) 5097{ 5098 call_rcu(&sd->rcu, free_sched_domain); 5099} 5100 5101static void destroy_sched_domains(struct sched_domain *sd, int cpu) 5102{ 5103 for (; sd; sd = sd->parent) 5104 destroy_sched_domain(sd, cpu); 5105} 5106 5107/* 5108 * Keep a special pointer to the highest sched_domain that has 5109 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this 5110 * allows us to avoid some pointer chasing select_idle_sibling(). 5111 * 5112 * Also keep a unique ID per domain (we use the first cpu number in 5113 * the cpumask of the domain), this allows us to quickly tell if 5114 * two cpus are in the same cache domain, see cpus_share_cache(). 5115 */ 5116DEFINE_PER_CPU(struct sched_domain *, sd_llc); 5117DEFINE_PER_CPU(int, sd_llc_size); 5118DEFINE_PER_CPU(int, sd_llc_id); 5119 5120static void update_top_cache_domain(int cpu) 5121{ 5122 struct sched_domain *sd; 5123 int id = cpu; 5124 int size = 1; 5125 5126 sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES); 5127 if (sd) { 5128 id = cpumask_first(sched_domain_span(sd)); 5129 size = cpumask_weight(sched_domain_span(sd)); 5130 } 5131 5132 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd); 5133 per_cpu(sd_llc_size, cpu) = size; 5134 per_cpu(sd_llc_id, cpu) = id; 5135} 5136 5137/* 5138 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must 5139 * hold the hotplug lock. 5140 */ 5141static void 5142cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu) 5143{ 5144 struct rq *rq = cpu_rq(cpu); 5145 struct sched_domain *tmp; 5146 5147 /* Remove the sched domains which do not contribute to scheduling. */ 5148 for (tmp = sd; tmp; ) { 5149 struct sched_domain *parent = tmp->parent; 5150 if (!parent) 5151 break; 5152 5153 if (sd_parent_degenerate(tmp, parent)) { 5154 tmp->parent = parent->parent; 5155 if (parent->parent) 5156 parent->parent->child = tmp; 5157 /* 5158 * Transfer SD_PREFER_SIBLING down in case of a 5159 * degenerate parent; the spans match for this 5160 * so the property transfers. 5161 */ 5162 if (parent->flags & SD_PREFER_SIBLING) 5163 tmp->flags |= SD_PREFER_SIBLING; 5164 destroy_sched_domain(parent, cpu); 5165 } else 5166 tmp = tmp->parent; 5167 } 5168 5169 if (sd && sd_degenerate(sd)) { 5170 tmp = sd; 5171 sd = sd->parent; 5172 destroy_sched_domain(tmp, cpu); 5173 if (sd) 5174 sd->child = NULL; 5175 } 5176 5177 sched_domain_debug(sd, cpu); 5178 5179 rq_attach_root(rq, rd); 5180 tmp = rq->sd; 5181 rcu_assign_pointer(rq->sd, sd); 5182 destroy_sched_domains(tmp, cpu); 5183 5184 update_top_cache_domain(cpu); 5185} 5186 5187/* cpus with isolated domains */ 5188static cpumask_var_t cpu_isolated_map; 5189 5190/* Setup the mask of cpus configured for isolated domains */ 5191static int __init isolated_cpu_setup(char *str) 5192{ 5193 alloc_bootmem_cpumask_var(&cpu_isolated_map); 5194 cpulist_parse(str, cpu_isolated_map); 5195 return 1; 5196} 5197 5198__setup("isolcpus=", isolated_cpu_setup); 5199 5200static const struct cpumask *cpu_cpu_mask(int cpu) 5201{ 5202 return cpumask_of_node(cpu_to_node(cpu)); 5203} 5204 5205struct sd_data { 5206 struct sched_domain **__percpu sd; 5207 struct sched_group **__percpu sg; 5208 struct sched_group_power **__percpu sgp; 5209}; 5210 5211struct s_data { 5212 struct sched_domain ** __percpu sd; 5213 struct root_domain *rd; 5214}; 5215 5216enum s_alloc { 5217 sa_rootdomain, 5218 sa_sd, 5219 sa_sd_storage, 5220 sa_none, 5221}; 5222 5223struct sched_domain_topology_level; 5224 5225typedef struct sched_domain *(*sched_domain_init_f)(struct sched_domain_topology_level *tl, int cpu); 5226typedef const struct cpumask *(*sched_domain_mask_f)(int cpu); 5227 5228#define SDTL_OVERLAP 0x01 5229 5230struct sched_domain_topology_level { 5231 sched_domain_init_f init; 5232 sched_domain_mask_f mask; 5233 int flags; 5234 int numa_level; 5235 struct sd_data data; 5236}; 5237 5238/* 5239 * Build an iteration mask that can exclude certain CPUs from the upwards 5240 * domain traversal. 5241 * 5242 * Asymmetric node setups can result in situations where the domain tree is of 5243 * unequal depth, make sure to skip domains that already cover the entire 5244 * range. 5245 * 5246 * In that case build_sched_domains() will have terminated the iteration early 5247 * and our sibling sd spans will be empty. Domains should always include the 5248 * cpu they're built on, so check that. 5249 * 5250 */ 5251static void build_group_mask(struct sched_domain *sd, struct sched_group *sg) 5252{ 5253 const struct cpumask *span = sched_domain_span(sd); 5254 struct sd_data *sdd = sd->private; 5255 struct sched_domain *sibling; 5256 int i; 5257 5258 for_each_cpu(i, span) { 5259 sibling = *per_cpu_ptr(sdd->sd, i); 5260 if (!cpumask_test_cpu(i, sched_domain_span(sibling))) 5261 continue; 5262 5263 cpumask_set_cpu(i, sched_group_mask(sg)); 5264 } 5265} 5266 5267/* 5268 * Return the canonical balance cpu for this group, this is the first cpu 5269 * of this group that's also in the iteration mask. 5270 */ 5271int group_balance_cpu(struct sched_group *sg) 5272{ 5273 return cpumask_first_and(sched_group_cpus(sg), sched_group_mask(sg)); 5274} 5275 5276static int 5277build_overlap_sched_groups(struct sched_domain *sd, int cpu) 5278{ 5279 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg; 5280 const struct cpumask *span = sched_domain_span(sd); 5281 struct cpumask *covered = sched_domains_tmpmask; 5282 struct sd_data *sdd = sd->private; 5283 struct sched_domain *child; 5284 int i; 5285 5286 cpumask_clear(covered); 5287 5288 for_each_cpu(i, span) { 5289 struct cpumask *sg_span; 5290 5291 if (cpumask_test_cpu(i, covered)) 5292 continue; 5293 5294 child = *per_cpu_ptr(sdd->sd, i); 5295 5296 /* See the comment near build_group_mask(). */ 5297 if (!cpumask_test_cpu(i, sched_domain_span(child))) 5298 continue; 5299 5300 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(), 5301 GFP_KERNEL, cpu_to_node(cpu)); 5302 5303 if (!sg) 5304 goto fail; 5305 5306 sg_span = sched_group_cpus(sg); 5307 if (child->child) { 5308 child = child->child; 5309 cpumask_copy(sg_span, sched_domain_span(child)); 5310 } else 5311 cpumask_set_cpu(i, sg_span); 5312 5313 cpumask_or(covered, covered, sg_span); 5314 5315 sg->sgp = *per_cpu_ptr(sdd->sgp, i); 5316 if (atomic_inc_return(&sg->sgp->ref) == 1) 5317 build_group_mask(sd, sg); 5318 5319 /* 5320 * Initialize sgp->power such that even if we mess up the 5321 * domains and no possible iteration will get us here, we won't 5322 * die on a /0 trap. 5323 */ 5324 sg->sgp->power = SCHED_POWER_SCALE * cpumask_weight(sg_span); 5325 5326 /* 5327 * Make sure the first group of this domain contains the 5328 * canonical balance cpu. Otherwise the sched_domain iteration 5329 * breaks. See update_sg_lb_stats(). 5330 */ 5331 if ((!groups && cpumask_test_cpu(cpu, sg_span)) || 5332 group_balance_cpu(sg) == cpu) 5333 groups = sg; 5334 5335 if (!first) 5336 first = sg; 5337 if (last) 5338 last->next = sg; 5339 last = sg; 5340 last->next = first; 5341 } 5342 sd->groups = groups; 5343 5344 return 0; 5345 5346fail: 5347 free_sched_groups(first, 0); 5348 5349 return -ENOMEM; 5350} 5351 5352static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg) 5353{ 5354 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu); 5355 struct sched_domain *child = sd->child; 5356 5357 if (child) 5358 cpu = cpumask_first(sched_domain_span(child)); 5359 5360 if (sg) { 5361 *sg = *per_cpu_ptr(sdd->sg, cpu); 5362 (*sg)->sgp = *per_cpu_ptr(sdd->sgp, cpu); 5363 atomic_set(&(*sg)->sgp->ref, 1); /* for claim_allocations */ 5364 } 5365 5366 return cpu; 5367} 5368 5369/* 5370 * build_sched_groups will build a circular linked list of the groups 5371 * covered by the given span, and will set each group's ->cpumask correctly, 5372 * and ->cpu_power to 0. 5373 * 5374 * Assumes the sched_domain tree is fully constructed 5375 */ 5376static int 5377build_sched_groups(struct sched_domain *sd, int cpu) 5378{ 5379 struct sched_group *first = NULL, *last = NULL; 5380 struct sd_data *sdd = sd->private; 5381 const struct cpumask *span = sched_domain_span(sd); 5382 struct cpumask *covered; 5383 int i; 5384 5385 get_group(cpu, sdd, &sd->groups); 5386 atomic_inc(&sd->groups->ref); 5387 5388 if (cpu != cpumask_first(span)) 5389 return 0; 5390 5391 lockdep_assert_held(&sched_domains_mutex); 5392 covered = sched_domains_tmpmask; 5393 5394 cpumask_clear(covered); 5395 5396 for_each_cpu(i, span) { 5397 struct sched_group *sg; 5398 int group, j; 5399 5400 if (cpumask_test_cpu(i, covered)) 5401 continue; 5402 5403 group = get_group(i, sdd, &sg); 5404 cpumask_clear(sched_group_cpus(sg)); 5405 sg->sgp->power = 0; 5406 cpumask_setall(sched_group_mask(sg)); 5407 5408 for_each_cpu(j, span) { 5409 if (get_group(j, sdd, NULL) != group) 5410 continue; 5411 5412 cpumask_set_cpu(j, covered); 5413 cpumask_set_cpu(j, sched_group_cpus(sg)); 5414 } 5415 5416 if (!first) 5417 first = sg; 5418 if (last) 5419 last->next = sg; 5420 last = sg; 5421 } 5422 last->next = first; 5423 5424 return 0; 5425} 5426 5427/* 5428 * Initialize sched groups cpu_power. 5429 * 5430 * cpu_power indicates the capacity of sched group, which is used while 5431 * distributing the load between different sched groups in a sched domain. 5432 * Typically cpu_power for all the groups in a sched domain will be same unless 5433 * there are asymmetries in the topology. If there are asymmetries, group 5434 * having more cpu_power will pickup more load compared to the group having 5435 * less cpu_power. 5436 */ 5437static void init_sched_groups_power(int cpu, struct sched_domain *sd) 5438{ 5439 struct sched_group *sg = sd->groups; 5440 5441 WARN_ON(!sg); 5442 5443 do { 5444 sg->group_weight = cpumask_weight(sched_group_cpus(sg)); 5445 sg = sg->next; 5446 } while (sg != sd->groups); 5447 5448 if (cpu != group_balance_cpu(sg)) 5449 return; 5450 5451 update_group_power(sd, cpu); 5452 atomic_set(&sg->sgp->nr_busy_cpus, sg->group_weight); 5453} 5454 5455int __weak arch_sd_sibling_asym_packing(void) 5456{ 5457 return 0*SD_ASYM_PACKING; 5458} 5459 5460/* 5461 * Initializers for schedule domains 5462 * Non-inlined to reduce accumulated stack pressure in build_sched_domains() 5463 */ 5464 5465#ifdef CONFIG_SCHED_DEBUG 5466# define SD_INIT_NAME(sd, type) sd->name = #type 5467#else 5468# define SD_INIT_NAME(sd, type) do { } while (0) 5469#endif 5470 5471#define SD_INIT_FUNC(type) \ 5472static noinline struct sched_domain * \ 5473sd_init_##type(struct sched_domain_topology_level *tl, int cpu) \ 5474{ \ 5475 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); \ 5476 *sd = SD_##type##_INIT; \ 5477 SD_INIT_NAME(sd, type); \ 5478 sd->private = &tl->data; \ 5479 return sd; \ 5480} 5481 5482SD_INIT_FUNC(CPU) 5483#ifdef CONFIG_SCHED_SMT 5484 SD_INIT_FUNC(SIBLING) 5485#endif 5486#ifdef CONFIG_SCHED_MC 5487 SD_INIT_FUNC(MC) 5488#endif 5489#ifdef CONFIG_SCHED_BOOK 5490 SD_INIT_FUNC(BOOK) 5491#endif 5492 5493static int default_relax_domain_level = -1; 5494int sched_domain_level_max; 5495 5496static int __init setup_relax_domain_level(char *str) 5497{ 5498 if (kstrtoint(str, 0, &default_relax_domain_level)) 5499 pr_warn("Unable to set relax_domain_level\n"); 5500 5501 return 1; 5502} 5503__setup("relax_domain_level=", setup_relax_domain_level); 5504 5505static void set_domain_attribute(struct sched_domain *sd, 5506 struct sched_domain_attr *attr) 5507{ 5508 int request; 5509 5510 if (!attr || attr->relax_domain_level < 0) { 5511 if (default_relax_domain_level < 0) 5512 return; 5513 else 5514 request = default_relax_domain_level; 5515 } else 5516 request = attr->relax_domain_level; 5517 if (request < sd->level) { 5518 /* turn off idle balance on this domain */ 5519 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE); 5520 } else { 5521 /* turn on idle balance on this domain */ 5522 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE); 5523 } 5524} 5525 5526static void __sdt_free(const struct cpumask *cpu_map); 5527static int __sdt_alloc(const struct cpumask *cpu_map); 5528 5529static void __free_domain_allocs(struct s_data *d, enum s_alloc what, 5530 const struct cpumask *cpu_map) 5531{ 5532 switch (what) { 5533 case sa_rootdomain: 5534 if (!atomic_read(&d->rd->refcount)) 5535 free_rootdomain(&d->rd->rcu); /* fall through */ 5536 case sa_sd: 5537 free_percpu(d->sd); /* fall through */ 5538 case sa_sd_storage: 5539 __sdt_free(cpu_map); /* fall through */ 5540 case sa_none: 5541 break; 5542 } 5543} 5544 5545static enum s_alloc __visit_domain_allocation_hell(struct s_data *d, 5546 const struct cpumask *cpu_map) 5547{ 5548 memset(d, 0, sizeof(*d)); 5549 5550 if (__sdt_alloc(cpu_map)) 5551 return sa_sd_storage; 5552 d->sd = alloc_percpu(struct sched_domain *); 5553 if (!d->sd) 5554 return sa_sd_storage; 5555 d->rd = alloc_rootdomain(); 5556 if (!d->rd) 5557 return sa_sd; 5558 return sa_rootdomain; 5559} 5560 5561/* 5562 * NULL the sd_data elements we've used to build the sched_domain and 5563 * sched_group structure so that the subsequent __free_domain_allocs() 5564 * will not free the data we're using. 5565 */ 5566static void claim_allocations(int cpu, struct sched_domain *sd) 5567{ 5568 struct sd_data *sdd = sd->private; 5569 5570 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd); 5571 *per_cpu_ptr(sdd->sd, cpu) = NULL; 5572 5573 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref)) 5574 *per_cpu_ptr(sdd->sg, cpu) = NULL; 5575 5576 if (atomic_read(&(*per_cpu_ptr(sdd->sgp, cpu))->ref)) 5577 *per_cpu_ptr(sdd->sgp, cpu) = NULL; 5578} 5579 5580#ifdef CONFIG_SCHED_SMT 5581static const struct cpumask *cpu_smt_mask(int cpu) 5582{ 5583 return topology_thread_cpumask(cpu); 5584} 5585#endif 5586 5587/* 5588 * Topology list, bottom-up. 5589 */ 5590static struct sched_domain_topology_level default_topology[] = { 5591#ifdef CONFIG_SCHED_SMT 5592 { sd_init_SIBLING, cpu_smt_mask, }, 5593#endif 5594#ifdef CONFIG_SCHED_MC 5595 { sd_init_MC, cpu_coregroup_mask, }, 5596#endif 5597#ifdef CONFIG_SCHED_BOOK 5598 { sd_init_BOOK, cpu_book_mask, }, 5599#endif 5600 { sd_init_CPU, cpu_cpu_mask, }, 5601 { NULL, }, 5602}; 5603 5604static struct sched_domain_topology_level *sched_domain_topology = default_topology; 5605 5606#define for_each_sd_topology(tl) \ 5607 for (tl = sched_domain_topology; tl->init; tl++) 5608 5609#ifdef CONFIG_NUMA 5610 5611static int sched_domains_numa_levels; 5612static int *sched_domains_numa_distance; 5613static struct cpumask ***sched_domains_numa_masks; 5614static int sched_domains_curr_level; 5615 5616static inline int sd_local_flags(int level) 5617{ 5618 if (sched_domains_numa_distance[level] > RECLAIM_DISTANCE) 5619 return 0; 5620 5621 return SD_BALANCE_EXEC | SD_BALANCE_FORK | SD_WAKE_AFFINE; 5622} 5623 5624static struct sched_domain * 5625sd_numa_init(struct sched_domain_topology_level *tl, int cpu) 5626{ 5627 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); 5628 int level = tl->numa_level; 5629 int sd_weight = cpumask_weight( 5630 sched_domains_numa_masks[level][cpu_to_node(cpu)]); 5631 5632 *sd = (struct sched_domain){ 5633 .min_interval = sd_weight, 5634 .max_interval = 2*sd_weight, 5635 .busy_factor = 32, 5636 .imbalance_pct = 125, 5637 .cache_nice_tries = 2, 5638 .busy_idx = 3, 5639 .idle_idx = 2, 5640 .newidle_idx = 0, 5641 .wake_idx = 0, 5642 .forkexec_idx = 0, 5643 5644 .flags = 1*SD_LOAD_BALANCE 5645 | 1*SD_BALANCE_NEWIDLE 5646 | 0*SD_BALANCE_EXEC 5647 | 0*SD_BALANCE_FORK 5648 | 0*SD_BALANCE_WAKE 5649 | 0*SD_WAKE_AFFINE 5650 | 0*SD_SHARE_CPUPOWER 5651 | 0*SD_SHARE_PKG_RESOURCES 5652 | 1*SD_SERIALIZE 5653 | 0*SD_PREFER_SIBLING 5654 | sd_local_flags(level) 5655 , 5656 .last_balance = jiffies, 5657 .balance_interval = sd_weight, 5658 }; 5659 SD_INIT_NAME(sd, NUMA); 5660 sd->private = &tl->data; 5661 5662 /* 5663 * Ugly hack to pass state to sd_numa_mask()... 5664 */ 5665 sched_domains_curr_level = tl->numa_level; 5666 5667 return sd; 5668} 5669 5670static const struct cpumask *sd_numa_mask(int cpu) 5671{ 5672 return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)]; 5673} 5674 5675static void sched_numa_warn(const char *str) 5676{ 5677 static int done = false; 5678 int i,j; 5679 5680 if (done) 5681 return; 5682 5683 done = true; 5684 5685 printk(KERN_WARNING "ERROR: %s\n\n", str); 5686 5687 for (i = 0; i < nr_node_ids; i++) { 5688 printk(KERN_WARNING " "); 5689 for (j = 0; j < nr_node_ids; j++) 5690 printk(KERN_CONT "%02d ", node_distance(i,j)); 5691 printk(KERN_CONT "\n"); 5692 } 5693 printk(KERN_WARNING "\n"); 5694} 5695 5696static bool find_numa_distance(int distance) 5697{ 5698 int i; 5699 5700 if (distance == node_distance(0, 0)) 5701 return true; 5702 5703 for (i = 0; i < sched_domains_numa_levels; i++) { 5704 if (sched_domains_numa_distance[i] == distance) 5705 return true; 5706 } 5707 5708 return false; 5709} 5710 5711static void sched_init_numa(void) 5712{ 5713 int next_distance, curr_distance = node_distance(0, 0); 5714 struct sched_domain_topology_level *tl; 5715 int level = 0; 5716 int i, j, k; 5717 5718 sched_domains_numa_distance = kzalloc(sizeof(int) * nr_node_ids, GFP_KERNEL); 5719 if (!sched_domains_numa_distance) 5720 return; 5721 5722 /* 5723 * O(nr_nodes^2) deduplicating selection sort -- in order to find the 5724 * unique distances in the node_distance() table. 5725 * 5726 * Assumes node_distance(0,j) includes all distances in 5727 * node_distance(i,j) in order to avoid cubic time. 5728 */ 5729 next_distance = curr_distance; 5730 for (i = 0; i < nr_node_ids; i++) { 5731 for (j = 0; j < nr_node_ids; j++) { 5732 for (k = 0; k < nr_node_ids; k++) { 5733 int distance = node_distance(i, k); 5734 5735 if (distance > curr_distance && 5736 (distance < next_distance || 5737 next_distance == curr_distance)) 5738 next_distance = distance; 5739 5740 /* 5741 * While not a strong assumption it would be nice to know 5742 * about cases where if node A is connected to B, B is not 5743 * equally connected to A. 5744 */ 5745 if (sched_debug() && node_distance(k, i) != distance) 5746 sched_numa_warn("Node-distance not symmetric"); 5747 5748 if (sched_debug() && i && !find_numa_distance(distance)) 5749 sched_numa_warn("Node-0 not representative"); 5750 } 5751 if (next_distance != curr_distance) { 5752 sched_domains_numa_distance[level++] = next_distance; 5753 sched_domains_numa_levels = level; 5754 curr_distance = next_distance; 5755 } else break; 5756 } 5757 5758 /* 5759 * In case of sched_debug() we verify the above assumption. 5760 */ 5761 if (!sched_debug()) 5762 break; 5763 } 5764 /* 5765 * 'level' contains the number of unique distances, excluding the 5766 * identity distance node_distance(i,i). 5767 * 5768 * The sched_domains_numa_distance[] array includes the actual distance 5769 * numbers. 5770 */ 5771 5772 /* 5773 * Here, we should temporarily reset sched_domains_numa_levels to 0. 5774 * If it fails to allocate memory for array sched_domains_numa_masks[][], 5775 * the array will contain less then 'level' members. This could be 5776 * dangerous when we use it to iterate array sched_domains_numa_masks[][] 5777 * in other functions. 5778 * 5779 * We reset it to 'level' at the end of this function. 5780 */ 5781 sched_domains_numa_levels = 0; 5782 5783 sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL); 5784 if (!sched_domains_numa_masks) 5785 return; 5786 5787 /* 5788 * Now for each level, construct a mask per node which contains all 5789 * cpus of nodes that are that many hops away from us. 5790 */ 5791 for (i = 0; i < level; i++) { 5792 sched_domains_numa_masks[i] = 5793 kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL); 5794 if (!sched_domains_numa_masks[i]) 5795 return; 5796 5797 for (j = 0; j < nr_node_ids; j++) { 5798 struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL); 5799 if (!mask) 5800 return; 5801 5802 sched_domains_numa_masks[i][j] = mask; 5803 5804 for (k = 0; k < nr_node_ids; k++) { 5805 if (node_distance(j, k) > sched_domains_numa_distance[i]) 5806 continue; 5807 5808 cpumask_or(mask, mask, cpumask_of_node(k)); 5809 } 5810 } 5811 } 5812 5813 tl = kzalloc((ARRAY_SIZE(default_topology) + level) * 5814 sizeof(struct sched_domain_topology_level), GFP_KERNEL); 5815 if (!tl) 5816 return; 5817 5818 /* 5819 * Copy the default topology bits.. 5820 */ 5821 for (i = 0; default_topology[i].init; i++) 5822 tl[i] = default_topology[i]; 5823 5824 /* 5825 * .. and append 'j' levels of NUMA goodness. 5826 */ 5827 for (j = 0; j < level; i++, j++) { 5828 tl[i] = (struct sched_domain_topology_level){ 5829 .init = sd_numa_init, 5830 .mask = sd_numa_mask, 5831 .flags = SDTL_OVERLAP, 5832 .numa_level = j, 5833 }; 5834 } 5835 5836 sched_domain_topology = tl; 5837 5838 sched_domains_numa_levels = level; 5839} 5840 5841static void sched_domains_numa_masks_set(int cpu) 5842{ 5843 int i, j; 5844 int node = cpu_to_node(cpu); 5845 5846 for (i = 0; i < sched_domains_numa_levels; i++) { 5847 for (j = 0; j < nr_node_ids; j++) { 5848 if (node_distance(j, node) <= sched_domains_numa_distance[i]) 5849 cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]); 5850 } 5851 } 5852} 5853 5854static void sched_domains_numa_masks_clear(int cpu) 5855{ 5856 int i, j; 5857 for (i = 0; i < sched_domains_numa_levels; i++) { 5858 for (j = 0; j < nr_node_ids; j++) 5859 cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]); 5860 } 5861} 5862 5863/* 5864 * Update sched_domains_numa_masks[level][node] array when new cpus 5865 * are onlined. 5866 */ 5867static int sched_domains_numa_masks_update(struct notifier_block *nfb, 5868 unsigned long action, 5869 void *hcpu) 5870{ 5871 int cpu = (long)hcpu; 5872 5873 switch (action & ~CPU_TASKS_FROZEN) { 5874 case CPU_ONLINE: 5875 sched_domains_numa_masks_set(cpu); 5876 break; 5877 5878 case CPU_DEAD: 5879 sched_domains_numa_masks_clear(cpu); 5880 break; 5881 5882 default: 5883 return NOTIFY_DONE; 5884 } 5885 5886 return NOTIFY_OK; 5887} 5888#else 5889static inline void sched_init_numa(void) 5890{ 5891} 5892 5893static int sched_domains_numa_masks_update(struct notifier_block *nfb, 5894 unsigned long action, 5895 void *hcpu) 5896{ 5897 return 0; 5898} 5899#endif /* CONFIG_NUMA */ 5900 5901static int __sdt_alloc(const struct cpumask *cpu_map) 5902{ 5903 struct sched_domain_topology_level *tl; 5904 int j; 5905 5906 for_each_sd_topology(tl) { 5907 struct sd_data *sdd = &tl->data; 5908 5909 sdd->sd = alloc_percpu(struct sched_domain *); 5910 if (!sdd->sd) 5911 return -ENOMEM; 5912 5913 sdd->sg = alloc_percpu(struct sched_group *); 5914 if (!sdd->sg) 5915 return -ENOMEM; 5916 5917 sdd->sgp = alloc_percpu(struct sched_group_power *); 5918 if (!sdd->sgp) 5919 return -ENOMEM; 5920 5921 for_each_cpu(j, cpu_map) { 5922 struct sched_domain *sd; 5923 struct sched_group *sg; 5924 struct sched_group_power *sgp; 5925 5926 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(), 5927 GFP_KERNEL, cpu_to_node(j)); 5928 if (!sd) 5929 return -ENOMEM; 5930 5931 *per_cpu_ptr(sdd->sd, j) = sd; 5932 5933 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(), 5934 GFP_KERNEL, cpu_to_node(j)); 5935 if (!sg) 5936 return -ENOMEM; 5937 5938 sg->next = sg; 5939 5940 *per_cpu_ptr(sdd->sg, j) = sg; 5941 5942 sgp = kzalloc_node(sizeof(struct sched_group_power) + cpumask_size(), 5943 GFP_KERNEL, cpu_to_node(j)); 5944 if (!sgp) 5945 return -ENOMEM; 5946 5947 *per_cpu_ptr(sdd->sgp, j) = sgp; 5948 } 5949 } 5950 5951 return 0; 5952} 5953 5954static void __sdt_free(const struct cpumask *cpu_map) 5955{ 5956 struct sched_domain_topology_level *tl; 5957 int j; 5958 5959 for_each_sd_topology(tl) { 5960 struct sd_data *sdd = &tl->data; 5961 5962 for_each_cpu(j, cpu_map) { 5963 struct sched_domain *sd; 5964 5965 if (sdd->sd) { 5966 sd = *per_cpu_ptr(sdd->sd, j); 5967 if (sd && (sd->flags & SD_OVERLAP)) 5968 free_sched_groups(sd->groups, 0); 5969 kfree(*per_cpu_ptr(sdd->sd, j)); 5970 } 5971 5972 if (sdd->sg) 5973 kfree(*per_cpu_ptr(sdd->sg, j)); 5974 if (sdd->sgp) 5975 kfree(*per_cpu_ptr(sdd->sgp, j)); 5976 } 5977 free_percpu(sdd->sd); 5978 sdd->sd = NULL; 5979 free_percpu(sdd->sg); 5980 sdd->sg = NULL; 5981 free_percpu(sdd->sgp); 5982 sdd->sgp = NULL; 5983 } 5984} 5985 5986struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl, 5987 const struct cpumask *cpu_map, struct sched_domain_attr *attr, 5988 struct sched_domain *child, int cpu) 5989{ 5990 struct sched_domain *sd = tl->init(tl, cpu); 5991 if (!sd) 5992 return child; 5993 5994 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu)); 5995 if (child) { 5996 sd->level = child->level + 1; 5997 sched_domain_level_max = max(sched_domain_level_max, sd->level); 5998 child->parent = sd; 5999 sd->child = child; 6000 } 6001 set_domain_attribute(sd, attr); 6002 6003 return sd; 6004} 6005 6006/* 6007 * Build sched domains for a given set of cpus and attach the sched domains 6008 * to the individual cpus 6009 */ 6010static int build_sched_domains(const struct cpumask *cpu_map, 6011 struct sched_domain_attr *attr) 6012{ 6013 enum s_alloc alloc_state; 6014 struct sched_domain *sd; 6015 struct s_data d; 6016 int i, ret = -ENOMEM; 6017 6018 alloc_state = __visit_domain_allocation_hell(&d, cpu_map); 6019 if (alloc_state != sa_rootdomain) 6020 goto error; 6021 6022 /* Set up domains for cpus specified by the cpu_map. */ 6023 for_each_cpu(i, cpu_map) { 6024 struct sched_domain_topology_level *tl; 6025 6026 sd = NULL; 6027 for_each_sd_topology(tl) { 6028 sd = build_sched_domain(tl, cpu_map, attr, sd, i); 6029 if (tl == sched_domain_topology) 6030 *per_cpu_ptr(d.sd, i) = sd; 6031 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP)) 6032 sd->flags |= SD_OVERLAP; 6033 if (cpumask_equal(cpu_map, sched_domain_span(sd))) 6034 break; 6035 } 6036 } 6037 6038 /* Build the groups for the domains */ 6039 for_each_cpu(i, cpu_map) { 6040 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) { 6041 sd->span_weight = cpumask_weight(sched_domain_span(sd)); 6042 if (sd->flags & SD_OVERLAP) { 6043 if (build_overlap_sched_groups(sd, i)) 6044 goto error; 6045 } else { 6046 if (build_sched_groups(sd, i)) 6047 goto error; 6048 } 6049 } 6050 } 6051 6052 /* Calculate CPU power for physical packages and nodes */ 6053 for (i = nr_cpumask_bits-1; i >= 0; i--) { 6054 if (!cpumask_test_cpu(i, cpu_map)) 6055 continue; 6056 6057 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) { 6058 claim_allocations(i, sd); 6059 init_sched_groups_power(i, sd); 6060 } 6061 } 6062 6063 /* Attach the domains */ 6064 rcu_read_lock(); 6065 for_each_cpu(i, cpu_map) { 6066 sd = *per_cpu_ptr(d.sd, i); 6067 cpu_attach_domain(sd, d.rd, i); 6068 } 6069 rcu_read_unlock(); 6070 6071 ret = 0; 6072error: 6073 __free_domain_allocs(&d, alloc_state, cpu_map); 6074 return ret; 6075} 6076 6077static cpumask_var_t *doms_cur; /* current sched domains */ 6078static int ndoms_cur; /* number of sched domains in 'doms_cur' */ 6079static struct sched_domain_attr *dattr_cur; 6080 /* attribues of custom domains in 'doms_cur' */ 6081 6082/* 6083 * Special case: If a kmalloc of a doms_cur partition (array of 6084 * cpumask) fails, then fallback to a single sched domain, 6085 * as determined by the single cpumask fallback_doms. 6086 */ 6087static cpumask_var_t fallback_doms; 6088 6089/* 6090 * arch_update_cpu_topology lets virtualized architectures update the 6091 * cpu core maps. It is supposed to return 1 if the topology changed 6092 * or 0 if it stayed the same. 6093 */ 6094int __attribute__((weak)) arch_update_cpu_topology(void) 6095{ 6096 return 0; 6097} 6098 6099cpumask_var_t *alloc_sched_domains(unsigned int ndoms) 6100{ 6101 int i; 6102 cpumask_var_t *doms; 6103 6104 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL); 6105 if (!doms) 6106 return NULL; 6107 for (i = 0; i < ndoms; i++) { 6108 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) { 6109 free_sched_domains(doms, i); 6110 return NULL; 6111 } 6112 } 6113 return doms; 6114} 6115 6116void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms) 6117{ 6118 unsigned int i; 6119 for (i = 0; i < ndoms; i++) 6120 free_cpumask_var(doms[i]); 6121 kfree(doms); 6122} 6123 6124/* 6125 * Set up scheduler domains and groups. Callers must hold the hotplug lock. 6126 * For now this just excludes isolated cpus, but could be used to 6127 * exclude other special cases in the future. 6128 */ 6129static int init_sched_domains(const struct cpumask *cpu_map) 6130{ 6131 int err; 6132 6133 arch_update_cpu_topology(); 6134 ndoms_cur = 1; 6135 doms_cur = alloc_sched_domains(ndoms_cur); 6136 if (!doms_cur) 6137 doms_cur = &fallback_doms; 6138 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map); 6139 err = build_sched_domains(doms_cur[0], NULL); 6140 register_sched_domain_sysctl(); 6141 6142 return err; 6143} 6144 6145/* 6146 * Detach sched domains from a group of cpus specified in cpu_map 6147 * These cpus will now be attached to the NULL domain 6148 */ 6149static void detach_destroy_domains(const struct cpumask *cpu_map) 6150{ 6151 int i; 6152 6153 rcu_read_lock(); 6154 for_each_cpu(i, cpu_map) 6155 cpu_attach_domain(NULL, &def_root_domain, i); 6156 rcu_read_unlock(); 6157} 6158 6159/* handle null as "default" */ 6160static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur, 6161 struct sched_domain_attr *new, int idx_new) 6162{ 6163 struct sched_domain_attr tmp; 6164 6165 /* fast path */ 6166 if (!new && !cur) 6167 return 1; 6168 6169 tmp = SD_ATTR_INIT; 6170 return !memcmp(cur ? (cur + idx_cur) : &tmp, 6171 new ? (new + idx_new) : &tmp, 6172 sizeof(struct sched_domain_attr)); 6173} 6174 6175/* 6176 * Partition sched domains as specified by the 'ndoms_new' 6177 * cpumasks in the array doms_new[] of cpumasks. This compares 6178 * doms_new[] to the current sched domain partitioning, doms_cur[]. 6179 * It destroys each deleted domain and builds each new domain. 6180 * 6181 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'. 6182 * The masks don't intersect (don't overlap.) We should setup one 6183 * sched domain for each mask. CPUs not in any of the cpumasks will 6184 * not be load balanced. If the same cpumask appears both in the 6185 * current 'doms_cur' domains and in the new 'doms_new', we can leave 6186 * it as it is. 6187 * 6188 * The passed in 'doms_new' should be allocated using 6189 * alloc_sched_domains. This routine takes ownership of it and will 6190 * free_sched_domains it when done with it. If the caller failed the 6191 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1, 6192 * and partition_sched_domains() will fallback to the single partition 6193 * 'fallback_doms', it also forces the domains to be rebuilt. 6194 * 6195 * If doms_new == NULL it will be replaced with cpu_online_mask. 6196 * ndoms_new == 0 is a special case for destroying existing domains, 6197 * and it will not create the default domain. 6198 * 6199 * Call with hotplug lock held 6200 */ 6201void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[], 6202 struct sched_domain_attr *dattr_new) 6203{ 6204 int i, j, n; 6205 int new_topology; 6206 6207 mutex_lock(&sched_domains_mutex); 6208 6209 /* always unregister in case we don't destroy any domains */ 6210 unregister_sched_domain_sysctl(); 6211 6212 /* Let architecture update cpu core mappings. */ 6213 new_topology = arch_update_cpu_topology(); 6214 6215 n = doms_new ? ndoms_new : 0; 6216 6217 /* Destroy deleted domains */ 6218 for (i = 0; i < ndoms_cur; i++) { 6219 for (j = 0; j < n && !new_topology; j++) { 6220 if (cpumask_equal(doms_cur[i], doms_new[j]) 6221 && dattrs_equal(dattr_cur, i, dattr_new, j)) 6222 goto match1; 6223 } 6224 /* no match - a current sched domain not in new doms_new[] */ 6225 detach_destroy_domains(doms_cur[i]); 6226match1: 6227 ; 6228 } 6229 6230 n = ndoms_cur; 6231 if (doms_new == NULL) { 6232 n = 0; 6233 doms_new = &fallback_doms; 6234 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map); 6235 WARN_ON_ONCE(dattr_new); 6236 } 6237 6238 /* Build new domains */ 6239 for (i = 0; i < ndoms_new; i++) { 6240 for (j = 0; j < n && !new_topology; j++) { 6241 if (cpumask_equal(doms_new[i], doms_cur[j]) 6242 && dattrs_equal(dattr_new, i, dattr_cur, j)) 6243 goto match2; 6244 } 6245 /* no match - add a new doms_new */ 6246 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL); 6247match2: 6248 ; 6249 } 6250 6251 /* Remember the new sched domains */ 6252 if (doms_cur != &fallback_doms) 6253 free_sched_domains(doms_cur, ndoms_cur); 6254 kfree(dattr_cur); /* kfree(NULL) is safe */ 6255 doms_cur = doms_new; 6256 dattr_cur = dattr_new; 6257 ndoms_cur = ndoms_new; 6258 6259 register_sched_domain_sysctl(); 6260 6261 mutex_unlock(&sched_domains_mutex); 6262} 6263 6264static int num_cpus_frozen; /* used to mark begin/end of suspend/resume */ 6265 6266/* 6267 * Update cpusets according to cpu_active mask. If cpusets are 6268 * disabled, cpuset_update_active_cpus() becomes a simple wrapper 6269 * around partition_sched_domains(). 6270 * 6271 * If we come here as part of a suspend/resume, don't touch cpusets because we 6272 * want to restore it back to its original state upon resume anyway. 6273 */ 6274static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action, 6275 void *hcpu) 6276{ 6277 switch (action) { 6278 case CPU_ONLINE_FROZEN: 6279 case CPU_DOWN_FAILED_FROZEN: 6280 6281 /* 6282 * num_cpus_frozen tracks how many CPUs are involved in suspend 6283 * resume sequence. As long as this is not the last online 6284 * operation in the resume sequence, just build a single sched 6285 * domain, ignoring cpusets. 6286 */ 6287 num_cpus_frozen--; 6288 if (likely(num_cpus_frozen)) { 6289 partition_sched_domains(1, NULL, NULL); 6290 break; 6291 } 6292 6293 /* 6294 * This is the last CPU online operation. So fall through and 6295 * restore the original sched domains by considering the 6296 * cpuset configurations. 6297 */ 6298 6299 case CPU_ONLINE: 6300 case CPU_DOWN_FAILED: 6301 cpuset_update_active_cpus(true); 6302 break; 6303 default: 6304 return NOTIFY_DONE; 6305 } 6306 return NOTIFY_OK; 6307} 6308 6309static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action, 6310 void *hcpu) 6311{ 6312 switch (action) { 6313 case CPU_DOWN_PREPARE: 6314 cpuset_update_active_cpus(false); 6315 break; 6316 case CPU_DOWN_PREPARE_FROZEN: 6317 num_cpus_frozen++; 6318 partition_sched_domains(1, NULL, NULL); 6319 break; 6320 default: 6321 return NOTIFY_DONE; 6322 } 6323 return NOTIFY_OK; 6324} 6325 6326void __init sched_init_smp(void) 6327{ 6328 cpumask_var_t non_isolated_cpus; 6329 6330 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL); 6331 alloc_cpumask_var(&fallback_doms, GFP_KERNEL); 6332 6333 sched_init_numa(); 6334 6335 get_online_cpus(); 6336 mutex_lock(&sched_domains_mutex); 6337 init_sched_domains(cpu_active_mask); 6338 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map); 6339 if (cpumask_empty(non_isolated_cpus)) 6340 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus); 6341 mutex_unlock(&sched_domains_mutex); 6342 put_online_cpus(); 6343 6344 hotcpu_notifier(sched_domains_numa_masks_update, CPU_PRI_SCHED_ACTIVE); 6345 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE); 6346 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE); 6347 6348 init_hrtick(); 6349 6350 /* Move init over to a non-isolated CPU */ 6351 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0) 6352 BUG(); 6353 sched_init_granularity(); 6354 free_cpumask_var(non_isolated_cpus); 6355 6356 init_sched_rt_class(); 6357} 6358#else 6359void __init sched_init_smp(void) 6360{ 6361 sched_init_granularity(); 6362} 6363#endif /* CONFIG_SMP */ 6364 6365const_debug unsigned int sysctl_timer_migration = 1; 6366 6367int in_sched_functions(unsigned long addr) 6368{ 6369 return in_lock_functions(addr) || 6370 (addr >= (unsigned long)__sched_text_start 6371 && addr < (unsigned long)__sched_text_end); 6372} 6373 6374#ifdef CONFIG_CGROUP_SCHED 6375/* 6376 * Default task group. 6377 * Every task in system belongs to this group at bootup. 6378 */ 6379struct task_group root_task_group; 6380LIST_HEAD(task_groups); 6381#endif 6382 6383DECLARE_PER_CPU(cpumask_var_t, load_balance_mask); 6384 6385void __init sched_init(void) 6386{ 6387 int i, j; 6388 unsigned long alloc_size = 0, ptr; 6389 6390#ifdef CONFIG_FAIR_GROUP_SCHED 6391 alloc_size += 2 * nr_cpu_ids * sizeof(void **); 6392#endif 6393#ifdef CONFIG_RT_GROUP_SCHED 6394 alloc_size += 2 * nr_cpu_ids * sizeof(void **); 6395#endif 6396#ifdef CONFIG_CPUMASK_OFFSTACK 6397 alloc_size += num_possible_cpus() * cpumask_size(); 6398#endif 6399 if (alloc_size) { 6400 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT); 6401 6402#ifdef CONFIG_FAIR_GROUP_SCHED 6403 root_task_group.se = (struct sched_entity **)ptr; 6404 ptr += nr_cpu_ids * sizeof(void **); 6405 6406 root_task_group.cfs_rq = (struct cfs_rq **)ptr; 6407 ptr += nr_cpu_ids * sizeof(void **); 6408 6409#endif /* CONFIG_FAIR_GROUP_SCHED */ 6410#ifdef CONFIG_RT_GROUP_SCHED 6411 root_task_group.rt_se = (struct sched_rt_entity **)ptr; 6412 ptr += nr_cpu_ids * sizeof(void **); 6413 6414 root_task_group.rt_rq = (struct rt_rq **)ptr; 6415 ptr += nr_cpu_ids * sizeof(void **); 6416 6417#endif /* CONFIG_RT_GROUP_SCHED */ 6418#ifdef CONFIG_CPUMASK_OFFSTACK 6419 for_each_possible_cpu(i) { 6420 per_cpu(load_balance_mask, i) = (void *)ptr; 6421 ptr += cpumask_size(); 6422 } 6423#endif /* CONFIG_CPUMASK_OFFSTACK */ 6424 } 6425 6426#ifdef CONFIG_SMP 6427 init_defrootdomain(); 6428#endif 6429 6430 init_rt_bandwidth(&def_rt_bandwidth, 6431 global_rt_period(), global_rt_runtime()); 6432 6433#ifdef CONFIG_RT_GROUP_SCHED 6434 init_rt_bandwidth(&root_task_group.rt_bandwidth, 6435 global_rt_period(), global_rt_runtime()); 6436#endif /* CONFIG_RT_GROUP_SCHED */ 6437 6438#ifdef CONFIG_CGROUP_SCHED 6439 list_add(&root_task_group.list, &task_groups); 6440 INIT_LIST_HEAD(&root_task_group.children); 6441 INIT_LIST_HEAD(&root_task_group.siblings); 6442 autogroup_init(&init_task); 6443 6444#endif /* CONFIG_CGROUP_SCHED */ 6445 6446 for_each_possible_cpu(i) { 6447 struct rq *rq; 6448 6449 rq = cpu_rq(i); 6450 raw_spin_lock_init(&rq->lock); 6451 rq->nr_running = 0; 6452 rq->calc_load_active = 0; 6453 rq->calc_load_update = jiffies + LOAD_FREQ; 6454 init_cfs_rq(&rq->cfs); 6455 init_rt_rq(&rq->rt, rq); 6456#ifdef CONFIG_FAIR_GROUP_SCHED 6457 root_task_group.shares = ROOT_TASK_GROUP_LOAD; 6458 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list); 6459 /* 6460 * How much cpu bandwidth does root_task_group get? 6461 * 6462 * In case of task-groups formed thr' the cgroup filesystem, it 6463 * gets 100% of the cpu resources in the system. This overall 6464 * system cpu resource is divided among the tasks of 6465 * root_task_group and its child task-groups in a fair manner, 6466 * based on each entity's (task or task-group's) weight 6467 * (se->load.weight). 6468 * 6469 * In other words, if root_task_group has 10 tasks of weight 6470 * 1024) and two child groups A0 and A1 (of weight 1024 each), 6471 * then A0's share of the cpu resource is: 6472 * 6473 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33% 6474 * 6475 * We achieve this by letting root_task_group's tasks sit 6476 * directly in rq->cfs (i.e root_task_group->se[] = NULL). 6477 */ 6478 init_cfs_bandwidth(&root_task_group.cfs_bandwidth); 6479 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL); 6480#endif /* CONFIG_FAIR_GROUP_SCHED */ 6481 6482 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime; 6483#ifdef CONFIG_RT_GROUP_SCHED 6484 INIT_LIST_HEAD(&rq->leaf_rt_rq_list); 6485 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL); 6486#endif 6487 6488 for (j = 0; j < CPU_LOAD_IDX_MAX; j++) 6489 rq->cpu_load[j] = 0; 6490 6491 rq->last_load_update_tick = jiffies; 6492 6493#ifdef CONFIG_SMP 6494 rq->sd = NULL; 6495 rq->rd = NULL; 6496 rq->cpu_power = SCHED_POWER_SCALE; 6497 rq->post_schedule = 0; 6498 rq->active_balance = 0; 6499 rq->next_balance = jiffies; 6500 rq->push_cpu = 0; 6501 rq->cpu = i; 6502 rq->online = 0; 6503 rq->idle_stamp = 0; 6504 rq->avg_idle = 2*sysctl_sched_migration_cost; 6505 rq->max_idle_balance_cost = sysctl_sched_migration_cost; 6506 6507 INIT_LIST_HEAD(&rq->cfs_tasks); 6508 6509 rq_attach_root(rq, &def_root_domain); 6510#ifdef CONFIG_NO_HZ_COMMON 6511 rq->nohz_flags = 0; 6512#endif 6513#ifdef CONFIG_NO_HZ_FULL 6514 rq->last_sched_tick = 0; 6515#endif 6516#endif 6517 init_rq_hrtick(rq); 6518 atomic_set(&rq->nr_iowait, 0); 6519 } 6520 6521 set_load_weight(&init_task); 6522 6523#ifdef CONFIG_PREEMPT_NOTIFIERS 6524 INIT_HLIST_HEAD(&init_task.preempt_notifiers); 6525#endif 6526 6527#ifdef CONFIG_RT_MUTEXES 6528 plist_head_init(&init_task.pi_waiters); 6529#endif 6530 6531 /* 6532 * The boot idle thread does lazy MMU switching as well: 6533 */ 6534 atomic_inc(&init_mm.mm_count); 6535 enter_lazy_tlb(&init_mm, current); 6536 6537 /* 6538 * Make us the idle thread. Technically, schedule() should not be 6539 * called from this thread, however somewhere below it might be, 6540 * but because we are the idle thread, we just pick up running again 6541 * when this runqueue becomes "idle". 6542 */ 6543 init_idle(current, smp_processor_id()); 6544 6545 calc_load_update = jiffies + LOAD_FREQ; 6546 6547 /* 6548 * During early bootup we pretend to be a normal task: 6549 */ 6550 current->sched_class = &fair_sched_class; 6551 6552#ifdef CONFIG_SMP 6553 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT); 6554 /* May be allocated at isolcpus cmdline parse time */ 6555 if (cpu_isolated_map == NULL) 6556 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT); 6557 idle_thread_set_boot_cpu(); 6558#endif 6559 init_sched_fair_class(); 6560 6561 scheduler_running = 1; 6562} 6563 6564#ifdef CONFIG_DEBUG_ATOMIC_SLEEP 6565static inline int preempt_count_equals(int preempt_offset) 6566{ 6567 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth(); 6568 6569 return (nested == preempt_offset); 6570} 6571 6572void __might_sleep(const char *file, int line, int preempt_offset) 6573{ 6574 static unsigned long prev_jiffy; /* ratelimiting */ 6575 6576 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */ 6577 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) || 6578 system_state != SYSTEM_RUNNING || oops_in_progress) 6579 return; 6580 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy) 6581 return; 6582 prev_jiffy = jiffies; 6583 6584 printk(KERN_ERR 6585 "BUG: sleeping function called from invalid context at %s:%d\n", 6586 file, line); 6587 printk(KERN_ERR 6588 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n", 6589 in_atomic(), irqs_disabled(), 6590 current->pid, current->comm); 6591 6592 debug_show_held_locks(current); 6593 if (irqs_disabled()) 6594 print_irqtrace_events(current); 6595 dump_stack(); 6596} 6597EXPORT_SYMBOL(__might_sleep); 6598#endif 6599 6600#ifdef CONFIG_MAGIC_SYSRQ 6601static void normalize_task(struct rq *rq, struct task_struct *p) 6602{ 6603 const struct sched_class *prev_class = p->sched_class; 6604 int old_prio = p->prio; 6605 int on_rq; 6606 6607 on_rq = p->on_rq; 6608 if (on_rq) 6609 dequeue_task(rq, p, 0); 6610 __setscheduler(rq, p, SCHED_NORMAL, 0); 6611 if (on_rq) { 6612 enqueue_task(rq, p, 0); 6613 resched_task(rq->curr); 6614 } 6615 6616 check_class_changed(rq, p, prev_class, old_prio); 6617} 6618 6619void normalize_rt_tasks(void) 6620{ 6621 struct task_struct *g, *p; 6622 unsigned long flags; 6623 struct rq *rq; 6624 6625 read_lock_irqsave(&tasklist_lock, flags); 6626 do_each_thread(g, p) { 6627 /* 6628 * Only normalize user tasks: 6629 */ 6630 if (!p->mm) 6631 continue; 6632 6633 p->se.exec_start = 0; 6634#ifdef CONFIG_SCHEDSTATS 6635 p->se.statistics.wait_start = 0; 6636 p->se.statistics.sleep_start = 0; 6637 p->se.statistics.block_start = 0; 6638#endif 6639 6640 if (!rt_task(p)) { 6641 /* 6642 * Renice negative nice level userspace 6643 * tasks back to 0: 6644 */ 6645 if (TASK_NICE(p) < 0 && p->mm) 6646 set_user_nice(p, 0); 6647 continue; 6648 } 6649 6650 raw_spin_lock(&p->pi_lock); 6651 rq = __task_rq_lock(p); 6652 6653 normalize_task(rq, p); 6654 6655 __task_rq_unlock(rq); 6656 raw_spin_unlock(&p->pi_lock); 6657 } while_each_thread(g, p); 6658 6659 read_unlock_irqrestore(&tasklist_lock, flags); 6660} 6661 6662#endif /* CONFIG_MAGIC_SYSRQ */ 6663 6664#if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) 6665/* 6666 * These functions are only useful for the IA64 MCA handling, or kdb. 6667 * 6668 * They can only be called when the whole system has been 6669 * stopped - every CPU needs to be quiescent, and no scheduling 6670 * activity can take place. Using them for anything else would 6671 * be a serious bug, and as a result, they aren't even visible 6672 * under any other configuration. 6673 */ 6674 6675/** 6676 * curr_task - return the current task for a given cpu. 6677 * @cpu: the processor in question. 6678 * 6679 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED! 6680 * 6681 * Return: The current task for @cpu. 6682 */ 6683struct task_struct *curr_task(int cpu) 6684{ 6685 return cpu_curr(cpu); 6686} 6687 6688#endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */ 6689 6690#ifdef CONFIG_IA64 6691/** 6692 * set_curr_task - set the current task for a given cpu. 6693 * @cpu: the processor in question. 6694 * @p: the task pointer to set. 6695 * 6696 * Description: This function must only be used when non-maskable interrupts 6697 * are serviced on a separate stack. It allows the architecture to switch the 6698 * notion of the current task on a cpu in a non-blocking manner. This function 6699 * must be called with all CPU's synchronized, and interrupts disabled, the 6700 * and caller must save the original value of the current task (see 6701 * curr_task() above) and restore that value before reenabling interrupts and 6702 * re-starting the system. 6703 * 6704 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED! 6705 */ 6706void set_curr_task(int cpu, struct task_struct *p) 6707{ 6708 cpu_curr(cpu) = p; 6709} 6710 6711#endif 6712 6713#ifdef CONFIG_CGROUP_SCHED 6714/* task_group_lock serializes the addition/removal of task groups */ 6715static DEFINE_SPINLOCK(task_group_lock); 6716 6717static void free_sched_group(struct task_group *tg) 6718{ 6719 free_fair_sched_group(tg); 6720 free_rt_sched_group(tg); 6721 autogroup_free(tg); 6722 kfree(tg); 6723} 6724 6725/* allocate runqueue etc for a new task group */ 6726struct task_group *sched_create_group(struct task_group *parent) 6727{ 6728 struct task_group *tg; 6729 6730 tg = kzalloc(sizeof(*tg), GFP_KERNEL); 6731 if (!tg) 6732 return ERR_PTR(-ENOMEM); 6733 6734 if (!alloc_fair_sched_group(tg, parent)) 6735 goto err; 6736 6737 if (!alloc_rt_sched_group(tg, parent)) 6738 goto err; 6739 6740 return tg; 6741 6742err: 6743 free_sched_group(tg); 6744 return ERR_PTR(-ENOMEM); 6745} 6746 6747void sched_online_group(struct task_group *tg, struct task_group *parent) 6748{ 6749 unsigned long flags; 6750 6751 spin_lock_irqsave(&task_group_lock, flags); 6752 list_add_rcu(&tg->list, &task_groups); 6753 6754 WARN_ON(!parent); /* root should already exist */ 6755 6756 tg->parent = parent; 6757 INIT_LIST_HEAD(&tg->children); 6758 list_add_rcu(&tg->siblings, &parent->children); 6759 spin_unlock_irqrestore(&task_group_lock, flags); 6760} 6761 6762/* rcu callback to free various structures associated with a task group */ 6763static void free_sched_group_rcu(struct rcu_head *rhp) 6764{ 6765 /* now it should be safe to free those cfs_rqs */ 6766 free_sched_group(container_of(rhp, struct task_group, rcu)); 6767} 6768 6769/* Destroy runqueue etc associated with a task group */ 6770void sched_destroy_group(struct task_group *tg) 6771{ 6772 /* wait for possible concurrent references to cfs_rqs complete */ 6773 call_rcu(&tg->rcu, free_sched_group_rcu); 6774} 6775 6776void sched_offline_group(struct task_group *tg) 6777{ 6778 unsigned long flags; 6779 int i; 6780 6781 /* end participation in shares distribution */ 6782 for_each_possible_cpu(i) 6783 unregister_fair_sched_group(tg, i); 6784 6785 spin_lock_irqsave(&task_group_lock, flags); 6786 list_del_rcu(&tg->list); 6787 list_del_rcu(&tg->siblings); 6788 spin_unlock_irqrestore(&task_group_lock, flags); 6789} 6790 6791/* change task's runqueue when it moves between groups. 6792 * The caller of this function should have put the task in its new group 6793 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to 6794 * reflect its new group. 6795 */ 6796void sched_move_task(struct task_struct *tsk) 6797{ 6798 struct task_group *tg; 6799 int on_rq, running; 6800 unsigned long flags; 6801 struct rq *rq; 6802 6803 rq = task_rq_lock(tsk, &flags); 6804 6805 running = task_current(rq, tsk); 6806 on_rq = tsk->on_rq; 6807 6808 if (on_rq) 6809 dequeue_task(rq, tsk, 0); 6810 if (unlikely(running)) 6811 tsk->sched_class->put_prev_task(rq, tsk); 6812 6813 tg = container_of(task_css_check(tsk, cpu_cgroup_subsys_id, 6814 lockdep_is_held(&tsk->sighand->siglock)), 6815 struct task_group, css); 6816 tg = autogroup_task_group(tsk, tg); 6817 tsk->sched_task_group = tg; 6818 6819#ifdef CONFIG_FAIR_GROUP_SCHED 6820 if (tsk->sched_class->task_move_group) 6821 tsk->sched_class->task_move_group(tsk, on_rq); 6822 else 6823#endif 6824 set_task_rq(tsk, task_cpu(tsk)); 6825 6826 if (unlikely(running)) 6827 tsk->sched_class->set_curr_task(rq); 6828 if (on_rq) 6829 enqueue_task(rq, tsk, 0); 6830 6831 task_rq_unlock(rq, tsk, &flags); 6832} 6833#endif /* CONFIG_CGROUP_SCHED */ 6834 6835#if defined(CONFIG_RT_GROUP_SCHED) || defined(CONFIG_CFS_BANDWIDTH) 6836static unsigned long to_ratio(u64 period, u64 runtime) 6837{ 6838 if (runtime == RUNTIME_INF) 6839 return 1ULL << 20; 6840 6841 return div64_u64(runtime << 20, period); 6842} 6843#endif 6844 6845#ifdef CONFIG_RT_GROUP_SCHED 6846/* 6847 * Ensure that the real time constraints are schedulable. 6848 */ 6849static DEFINE_MUTEX(rt_constraints_mutex); 6850 6851/* Must be called with tasklist_lock held */ 6852static inline int tg_has_rt_tasks(struct task_group *tg) 6853{ 6854 struct task_struct *g, *p; 6855 6856 do_each_thread(g, p) { 6857 if (rt_task(p) && task_rq(p)->rt.tg == tg) 6858 return 1; 6859 } while_each_thread(g, p); 6860 6861 return 0; 6862} 6863 6864struct rt_schedulable_data { 6865 struct task_group *tg; 6866 u64 rt_period; 6867 u64 rt_runtime; 6868}; 6869 6870static int tg_rt_schedulable(struct task_group *tg, void *data) 6871{ 6872 struct rt_schedulable_data *d = data; 6873 struct task_group *child; 6874 unsigned long total, sum = 0; 6875 u64 period, runtime; 6876 6877 period = ktime_to_ns(tg->rt_bandwidth.rt_period); 6878 runtime = tg->rt_bandwidth.rt_runtime; 6879 6880 if (tg == d->tg) { 6881 period = d->rt_period; 6882 runtime = d->rt_runtime; 6883 } 6884 6885 /* 6886 * Cannot have more runtime than the period. 6887 */ 6888 if (runtime > period && runtime != RUNTIME_INF) 6889 return -EINVAL; 6890 6891 /* 6892 * Ensure we don't starve existing RT tasks. 6893 */ 6894 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg)) 6895 return -EBUSY; 6896 6897 total = to_ratio(period, runtime); 6898 6899 /* 6900 * Nobody can have more than the global setting allows. 6901 */ 6902 if (total > to_ratio(global_rt_period(), global_rt_runtime())) 6903 return -EINVAL; 6904 6905 /* 6906 * The sum of our children's runtime should not exceed our own. 6907 */ 6908 list_for_each_entry_rcu(child, &tg->children, siblings) { 6909 period = ktime_to_ns(child->rt_bandwidth.rt_period); 6910 runtime = child->rt_bandwidth.rt_runtime; 6911 6912 if (child == d->tg) { 6913 period = d->rt_period; 6914 runtime = d->rt_runtime; 6915 } 6916 6917 sum += to_ratio(period, runtime); 6918 } 6919 6920 if (sum > total) 6921 return -EINVAL; 6922 6923 return 0; 6924} 6925 6926static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime) 6927{ 6928 int ret; 6929 6930 struct rt_schedulable_data data = { 6931 .tg = tg, 6932 .rt_period = period, 6933 .rt_runtime = runtime, 6934 }; 6935 6936 rcu_read_lock(); 6937 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data); 6938 rcu_read_unlock(); 6939 6940 return ret; 6941} 6942 6943static int tg_set_rt_bandwidth(struct task_group *tg, 6944 u64 rt_period, u64 rt_runtime) 6945{ 6946 int i, err = 0; 6947 6948 mutex_lock(&rt_constraints_mutex); 6949 read_lock(&tasklist_lock); 6950 err = __rt_schedulable(tg, rt_period, rt_runtime); 6951 if (err) 6952 goto unlock; 6953 6954 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock); 6955 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period); 6956 tg->rt_bandwidth.rt_runtime = rt_runtime; 6957 6958 for_each_possible_cpu(i) { 6959 struct rt_rq *rt_rq = tg->rt_rq[i]; 6960 6961 raw_spin_lock(&rt_rq->rt_runtime_lock); 6962 rt_rq->rt_runtime = rt_runtime; 6963 raw_spin_unlock(&rt_rq->rt_runtime_lock); 6964 } 6965 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock); 6966unlock: 6967 read_unlock(&tasklist_lock); 6968 mutex_unlock(&rt_constraints_mutex); 6969 6970 return err; 6971} 6972 6973static int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us) 6974{ 6975 u64 rt_runtime, rt_period; 6976 6977 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period); 6978 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC; 6979 if (rt_runtime_us < 0) 6980 rt_runtime = RUNTIME_INF; 6981 6982 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime); 6983} 6984 6985static long sched_group_rt_runtime(struct task_group *tg) 6986{ 6987 u64 rt_runtime_us; 6988 6989 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF) 6990 return -1; 6991 6992 rt_runtime_us = tg->rt_bandwidth.rt_runtime; 6993 do_div(rt_runtime_us, NSEC_PER_USEC); 6994 return rt_runtime_us; 6995} 6996 6997static int sched_group_set_rt_period(struct task_group *tg, long rt_period_us) 6998{ 6999 u64 rt_runtime, rt_period; 7000 7001 rt_period = (u64)rt_period_us * NSEC_PER_USEC; 7002 rt_runtime = tg->rt_bandwidth.rt_runtime; 7003 7004 if (rt_period == 0) 7005 return -EINVAL; 7006 7007 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime); 7008} 7009 7010static long sched_group_rt_period(struct task_group *tg) 7011{ 7012 u64 rt_period_us; 7013 7014 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period); 7015 do_div(rt_period_us, NSEC_PER_USEC); 7016 return rt_period_us; 7017} 7018 7019static int sched_rt_global_constraints(void) 7020{ 7021 u64 runtime, period; 7022 int ret = 0; 7023 7024 if (sysctl_sched_rt_period <= 0) 7025 return -EINVAL; 7026 7027 runtime = global_rt_runtime(); 7028 period = global_rt_period(); 7029 7030 /* 7031 * Sanity check on the sysctl variables. 7032 */ 7033 if (runtime > period && runtime != RUNTIME_INF) 7034 return -EINVAL; 7035 7036 mutex_lock(&rt_constraints_mutex); 7037 read_lock(&tasklist_lock); 7038 ret = __rt_schedulable(NULL, 0, 0); 7039 read_unlock(&tasklist_lock); 7040 mutex_unlock(&rt_constraints_mutex); 7041 7042 return ret; 7043} 7044 7045static int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk) 7046{ 7047 /* Don't accept realtime tasks when there is no way for them to run */ 7048 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0) 7049 return 0; 7050 7051 return 1; 7052} 7053 7054#else /* !CONFIG_RT_GROUP_SCHED */ 7055static int sched_rt_global_constraints(void) 7056{ 7057 unsigned long flags; 7058 int i; 7059 7060 if (sysctl_sched_rt_period <= 0) 7061 return -EINVAL; 7062 7063 /* 7064 * There's always some RT tasks in the root group 7065 * -- migration, kstopmachine etc.. 7066 */ 7067 if (sysctl_sched_rt_runtime == 0) 7068 return -EBUSY; 7069 7070 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags); 7071 for_each_possible_cpu(i) { 7072 struct rt_rq *rt_rq = &cpu_rq(i)->rt; 7073 7074 raw_spin_lock(&rt_rq->rt_runtime_lock); 7075 rt_rq->rt_runtime = global_rt_runtime(); 7076 raw_spin_unlock(&rt_rq->rt_runtime_lock); 7077 } 7078 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags); 7079 7080 return 0; 7081} 7082#endif /* CONFIG_RT_GROUP_SCHED */ 7083 7084int sched_rr_handler(struct ctl_table *table, int write, 7085 void __user *buffer, size_t *lenp, 7086 loff_t *ppos) 7087{ 7088 int ret; 7089 static DEFINE_MUTEX(mutex); 7090 7091 mutex_lock(&mutex); 7092 ret = proc_dointvec(table, write, buffer, lenp, ppos); 7093 /* make sure that internally we keep jiffies */ 7094 /* also, writing zero resets timeslice to default */ 7095 if (!ret && write) { 7096 sched_rr_timeslice = sched_rr_timeslice <= 0 ? 7097 RR_TIMESLICE : msecs_to_jiffies(sched_rr_timeslice); 7098 } 7099 mutex_unlock(&mutex); 7100 return ret; 7101} 7102 7103int sched_rt_handler(struct ctl_table *table, int write, 7104 void __user *buffer, size_t *lenp, 7105 loff_t *ppos) 7106{ 7107 int ret; 7108 int old_period, old_runtime; 7109 static DEFINE_MUTEX(mutex); 7110 7111 mutex_lock(&mutex); 7112 old_period = sysctl_sched_rt_period; 7113 old_runtime = sysctl_sched_rt_runtime; 7114 7115 ret = proc_dointvec(table, write, buffer, lenp, ppos); 7116 7117 if (!ret && write) { 7118 ret = sched_rt_global_constraints(); 7119 if (ret) { 7120 sysctl_sched_rt_period = old_period; 7121 sysctl_sched_rt_runtime = old_runtime; 7122 } else { 7123 def_rt_bandwidth.rt_runtime = global_rt_runtime(); 7124 def_rt_bandwidth.rt_period = 7125 ns_to_ktime(global_rt_period()); 7126 } 7127 } 7128 mutex_unlock(&mutex); 7129 7130 return ret; 7131} 7132 7133#ifdef CONFIG_CGROUP_SCHED 7134 7135static inline struct task_group *css_tg(struct cgroup_subsys_state *css) 7136{ 7137 return css ? container_of(css, struct task_group, css) : NULL; 7138} 7139 7140static struct cgroup_subsys_state * 7141cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css) 7142{ 7143 struct task_group *parent = css_tg(parent_css); 7144 struct task_group *tg; 7145 7146 if (!parent) { 7147 /* This is early initialization for the top cgroup */ 7148 return &root_task_group.css; 7149 } 7150 7151 tg = sched_create_group(parent); 7152 if (IS_ERR(tg)) 7153 return ERR_PTR(-ENOMEM); 7154 7155 return &tg->css; 7156} 7157 7158static int cpu_cgroup_css_online(struct cgroup_subsys_state *css) 7159{ 7160 struct task_group *tg = css_tg(css); 7161 struct task_group *parent = css_tg(css_parent(css)); 7162 7163 if (parent) 7164 sched_online_group(tg, parent); 7165 return 0; 7166} 7167 7168static void cpu_cgroup_css_free(struct cgroup_subsys_state *css) 7169{ 7170 struct task_group *tg = css_tg(css); 7171 7172 sched_destroy_group(tg); 7173} 7174 7175static void cpu_cgroup_css_offline(struct cgroup_subsys_state *css) 7176{ 7177 struct task_group *tg = css_tg(css); 7178 7179 sched_offline_group(tg); 7180} 7181 7182static int cpu_cgroup_can_attach(struct cgroup_subsys_state *css, 7183 struct cgroup_taskset *tset) 7184{ 7185 struct task_struct *task; 7186 7187 cgroup_taskset_for_each(task, css, tset) { 7188#ifdef CONFIG_RT_GROUP_SCHED 7189 if (!sched_rt_can_attach(css_tg(css), task)) 7190 return -EINVAL; 7191#else 7192 /* We don't support RT-tasks being in separate groups */ 7193 if (task->sched_class != &fair_sched_class) 7194 return -EINVAL; 7195#endif 7196 } 7197 return 0; 7198} 7199 7200static void cpu_cgroup_attach(struct cgroup_subsys_state *css, 7201 struct cgroup_taskset *tset) 7202{ 7203 struct task_struct *task; 7204 7205 cgroup_taskset_for_each(task, css, tset) 7206 sched_move_task(task); 7207} 7208 7209static void cpu_cgroup_exit(struct cgroup_subsys_state *css, 7210 struct cgroup_subsys_state *old_css, 7211 struct task_struct *task) 7212{ 7213 /* 7214 * cgroup_exit() is called in the copy_process() failure path. 7215 * Ignore this case since the task hasn't ran yet, this avoids 7216 * trying to poke a half freed task state from generic code. 7217 */ 7218 if (!(task->flags & PF_EXITING)) 7219 return; 7220 7221 sched_move_task(task); 7222} 7223 7224#ifdef CONFIG_FAIR_GROUP_SCHED 7225static int cpu_shares_write_u64(struct cgroup_subsys_state *css, 7226 struct cftype *cftype, u64 shareval) 7227{ 7228 return sched_group_set_shares(css_tg(css), scale_load(shareval)); 7229} 7230 7231static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css, 7232 struct cftype *cft) 7233{ 7234 struct task_group *tg = css_tg(css); 7235 7236 return (u64) scale_load_down(tg->shares); 7237} 7238 7239#ifdef CONFIG_CFS_BANDWIDTH 7240static DEFINE_MUTEX(cfs_constraints_mutex); 7241 7242const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */ 7243const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */ 7244 7245static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime); 7246 7247static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota) 7248{ 7249 int i, ret = 0, runtime_enabled, runtime_was_enabled; 7250 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth; 7251 7252 if (tg == &root_task_group) 7253 return -EINVAL; 7254 7255 /* 7256 * Ensure we have at some amount of bandwidth every period. This is 7257 * to prevent reaching a state of large arrears when throttled via 7258 * entity_tick() resulting in prolonged exit starvation. 7259 */ 7260 if (quota < min_cfs_quota_period || period < min_cfs_quota_period) 7261 return -EINVAL; 7262 7263 /* 7264 * Likewise, bound things on the otherside by preventing insane quota 7265 * periods. This also allows us to normalize in computing quota 7266 * feasibility. 7267 */ 7268 if (period > max_cfs_quota_period) 7269 return -EINVAL; 7270 7271 mutex_lock(&cfs_constraints_mutex); 7272 ret = __cfs_schedulable(tg, period, quota); 7273 if (ret) 7274 goto out_unlock; 7275 7276 runtime_enabled = quota != RUNTIME_INF; 7277 runtime_was_enabled = cfs_b->quota != RUNTIME_INF; 7278 account_cfs_bandwidth_used(runtime_enabled, runtime_was_enabled); 7279 raw_spin_lock_irq(&cfs_b->lock); 7280 cfs_b->period = ns_to_ktime(period); 7281 cfs_b->quota = quota; 7282 7283 __refill_cfs_bandwidth_runtime(cfs_b); 7284 /* restart the period timer (if active) to handle new period expiry */ 7285 if (runtime_enabled && cfs_b->timer_active) { 7286 /* force a reprogram */ 7287 cfs_b->timer_active = 0; 7288 __start_cfs_bandwidth(cfs_b); 7289 } 7290 raw_spin_unlock_irq(&cfs_b->lock); 7291 7292 for_each_possible_cpu(i) { 7293 struct cfs_rq *cfs_rq = tg->cfs_rq[i]; 7294 struct rq *rq = cfs_rq->rq; 7295 7296 raw_spin_lock_irq(&rq->lock); 7297 cfs_rq->runtime_enabled = runtime_enabled; 7298 cfs_rq->runtime_remaining = 0; 7299 7300 if (cfs_rq->throttled) 7301 unthrottle_cfs_rq(cfs_rq); 7302 raw_spin_unlock_irq(&rq->lock); 7303 } 7304out_unlock: 7305 mutex_unlock(&cfs_constraints_mutex); 7306 7307 return ret; 7308} 7309 7310int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us) 7311{ 7312 u64 quota, period; 7313 7314 period = ktime_to_ns(tg->cfs_bandwidth.period); 7315 if (cfs_quota_us < 0) 7316 quota = RUNTIME_INF; 7317 else 7318 quota = (u64)cfs_quota_us * NSEC_PER_USEC; 7319 7320 return tg_set_cfs_bandwidth(tg, period, quota); 7321} 7322 7323long tg_get_cfs_quota(struct task_group *tg) 7324{ 7325 u64 quota_us; 7326 7327 if (tg->cfs_bandwidth.quota == RUNTIME_INF) 7328 return -1; 7329 7330 quota_us = tg->cfs_bandwidth.quota; 7331 do_div(quota_us, NSEC_PER_USEC); 7332 7333 return quota_us; 7334} 7335 7336int tg_set_cfs_period(struct task_group *tg, long cfs_period_us) 7337{ 7338 u64 quota, period; 7339 7340 period = (u64)cfs_period_us * NSEC_PER_USEC; 7341 quota = tg->cfs_bandwidth.quota; 7342 7343 return tg_set_cfs_bandwidth(tg, period, quota); 7344} 7345 7346long tg_get_cfs_period(struct task_group *tg) 7347{ 7348 u64 cfs_period_us; 7349 7350 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period); 7351 do_div(cfs_period_us, NSEC_PER_USEC); 7352 7353 return cfs_period_us; 7354} 7355 7356static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css, 7357 struct cftype *cft) 7358{ 7359 return tg_get_cfs_quota(css_tg(css)); 7360} 7361 7362static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css, 7363 struct cftype *cftype, s64 cfs_quota_us) 7364{ 7365 return tg_set_cfs_quota(css_tg(css), cfs_quota_us); 7366} 7367 7368static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css, 7369 struct cftype *cft) 7370{ 7371 return tg_get_cfs_period(css_tg(css)); 7372} 7373 7374static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css, 7375 struct cftype *cftype, u64 cfs_period_us) 7376{ 7377 return tg_set_cfs_period(css_tg(css), cfs_period_us); 7378} 7379 7380struct cfs_schedulable_data { 7381 struct task_group *tg; 7382 u64 period, quota; 7383}; 7384 7385/* 7386 * normalize group quota/period to be quota/max_period 7387 * note: units are usecs 7388 */ 7389static u64 normalize_cfs_quota(struct task_group *tg, 7390 struct cfs_schedulable_data *d) 7391{ 7392 u64 quota, period; 7393 7394 if (tg == d->tg) { 7395 period = d->period; 7396 quota = d->quota; 7397 } else { 7398 period = tg_get_cfs_period(tg); 7399 quota = tg_get_cfs_quota(tg); 7400 } 7401 7402 /* note: these should typically be equivalent */ 7403 if (quota == RUNTIME_INF || quota == -1) 7404 return RUNTIME_INF; 7405 7406 return to_ratio(period, quota); 7407} 7408 7409static int tg_cfs_schedulable_down(struct task_group *tg, void *data) 7410{ 7411 struct cfs_schedulable_data *d = data; 7412 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth; 7413 s64 quota = 0, parent_quota = -1; 7414 7415 if (!tg->parent) { 7416 quota = RUNTIME_INF; 7417 } else { 7418 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth; 7419 7420 quota = normalize_cfs_quota(tg, d); 7421 parent_quota = parent_b->hierarchal_quota; 7422 7423 /* 7424 * ensure max(child_quota) <= parent_quota, inherit when no 7425 * limit is set 7426 */ 7427 if (quota == RUNTIME_INF) 7428 quota = parent_quota; 7429 else if (parent_quota != RUNTIME_INF && quota > parent_quota) 7430 return -EINVAL; 7431 } 7432 cfs_b->hierarchal_quota = quota; 7433 7434 return 0; 7435} 7436 7437static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota) 7438{ 7439 int ret; 7440 struct cfs_schedulable_data data = { 7441 .tg = tg, 7442 .period = period, 7443 .quota = quota, 7444 }; 7445 7446 if (quota != RUNTIME_INF) { 7447 do_div(data.period, NSEC_PER_USEC); 7448 do_div(data.quota, NSEC_PER_USEC); 7449 } 7450 7451 rcu_read_lock(); 7452 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data); 7453 rcu_read_unlock(); 7454 7455 return ret; 7456} 7457 7458static int cpu_stats_show(struct cgroup_subsys_state *css, struct cftype *cft, 7459 struct cgroup_map_cb *cb) 7460{ 7461 struct task_group *tg = css_tg(css); 7462 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth; 7463 7464 cb->fill(cb, "nr_periods", cfs_b->nr_periods); 7465 cb->fill(cb, "nr_throttled", cfs_b->nr_throttled); 7466 cb->fill(cb, "throttled_time", cfs_b->throttled_time); 7467 7468 return 0; 7469} 7470#endif /* CONFIG_CFS_BANDWIDTH */ 7471#endif /* CONFIG_FAIR_GROUP_SCHED */ 7472 7473#ifdef CONFIG_RT_GROUP_SCHED 7474static int cpu_rt_runtime_write(struct cgroup_subsys_state *css, 7475 struct cftype *cft, s64 val) 7476{ 7477 return sched_group_set_rt_runtime(css_tg(css), val); 7478} 7479 7480static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css, 7481 struct cftype *cft) 7482{ 7483 return sched_group_rt_runtime(css_tg(css)); 7484} 7485 7486static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css, 7487 struct cftype *cftype, u64 rt_period_us) 7488{ 7489 return sched_group_set_rt_period(css_tg(css), rt_period_us); 7490} 7491 7492static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css, 7493 struct cftype *cft) 7494{ 7495 return sched_group_rt_period(css_tg(css)); 7496} 7497#endif /* CONFIG_RT_GROUP_SCHED */ 7498 7499static struct cftype cpu_files[] = { 7500#ifdef CONFIG_FAIR_GROUP_SCHED 7501 { 7502 .name = "shares", 7503 .read_u64 = cpu_shares_read_u64, 7504 .write_u64 = cpu_shares_write_u64, 7505 }, 7506#endif 7507#ifdef CONFIG_CFS_BANDWIDTH 7508 { 7509 .name = "cfs_quota_us", 7510 .read_s64 = cpu_cfs_quota_read_s64, 7511 .write_s64 = cpu_cfs_quota_write_s64, 7512 }, 7513 { 7514 .name = "cfs_period_us", 7515 .read_u64 = cpu_cfs_period_read_u64, 7516 .write_u64 = cpu_cfs_period_write_u64, 7517 }, 7518 { 7519 .name = "stat", 7520 .read_map = cpu_stats_show, 7521 }, 7522#endif 7523#ifdef CONFIG_RT_GROUP_SCHED 7524 { 7525 .name = "rt_runtime_us", 7526 .read_s64 = cpu_rt_runtime_read, 7527 .write_s64 = cpu_rt_runtime_write, 7528 }, 7529 { 7530 .name = "rt_period_us", 7531 .read_u64 = cpu_rt_period_read_uint, 7532 .write_u64 = cpu_rt_period_write_uint, 7533 }, 7534#endif 7535 { } /* terminate */ 7536}; 7537 7538struct cgroup_subsys cpu_cgroup_subsys = { 7539 .name = "cpu", 7540 .css_alloc = cpu_cgroup_css_alloc, 7541 .css_free = cpu_cgroup_css_free, 7542 .css_online = cpu_cgroup_css_online, 7543 .css_offline = cpu_cgroup_css_offline, 7544 .can_attach = cpu_cgroup_can_attach, 7545 .attach = cpu_cgroup_attach, 7546 .exit = cpu_cgroup_exit, 7547 .subsys_id = cpu_cgroup_subsys_id, 7548 .base_cftypes = cpu_files, 7549 .early_init = 1, 7550}; 7551 7552#endif /* CONFIG_CGROUP_SCHED */ 7553 7554void dump_cpu_task(int cpu) 7555{ 7556 pr_info("Task dump for CPU %d:\n", cpu); 7557 sched_show_task(cpu_curr(cpu)); 7558} 7559