core.c revision 78becc27097585c6aec7043834cadde950ae79f2
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 * Its all a bit involved since we cannot program an hrt while holding the 375 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a 376 * reschedule event. 377 * 378 * When we get rescheduled we reprogram the hrtick_timer outside of the 379 * rq->lock. 380 */ 381 382static void hrtick_clear(struct rq *rq) 383{ 384 if (hrtimer_active(&rq->hrtick_timer)) 385 hrtimer_cancel(&rq->hrtick_timer); 386} 387 388/* 389 * High-resolution timer tick. 390 * Runs from hardirq context with interrupts disabled. 391 */ 392static enum hrtimer_restart hrtick(struct hrtimer *timer) 393{ 394 struct rq *rq = container_of(timer, struct rq, hrtick_timer); 395 396 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id()); 397 398 raw_spin_lock(&rq->lock); 399 update_rq_clock(rq); 400 rq->curr->sched_class->task_tick(rq, rq->curr, 1); 401 raw_spin_unlock(&rq->lock); 402 403 return HRTIMER_NORESTART; 404} 405 406#ifdef CONFIG_SMP 407/* 408 * called from hardirq (IPI) context 409 */ 410static void __hrtick_start(void *arg) 411{ 412 struct rq *rq = arg; 413 414 raw_spin_lock(&rq->lock); 415 hrtimer_restart(&rq->hrtick_timer); 416 rq->hrtick_csd_pending = 0; 417 raw_spin_unlock(&rq->lock); 418} 419 420/* 421 * Called to set the hrtick timer state. 422 * 423 * called with rq->lock held and irqs disabled 424 */ 425void hrtick_start(struct rq *rq, u64 delay) 426{ 427 struct hrtimer *timer = &rq->hrtick_timer; 428 ktime_t time = ktime_add_ns(timer->base->get_time(), delay); 429 430 hrtimer_set_expires(timer, time); 431 432 if (rq == this_rq()) { 433 hrtimer_restart(timer); 434 } else if (!rq->hrtick_csd_pending) { 435 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd, 0); 436 rq->hrtick_csd_pending = 1; 437 } 438} 439 440static int 441hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu) 442{ 443 int cpu = (int)(long)hcpu; 444 445 switch (action) { 446 case CPU_UP_CANCELED: 447 case CPU_UP_CANCELED_FROZEN: 448 case CPU_DOWN_PREPARE: 449 case CPU_DOWN_PREPARE_FROZEN: 450 case CPU_DEAD: 451 case CPU_DEAD_FROZEN: 452 hrtick_clear(cpu_rq(cpu)); 453 return NOTIFY_OK; 454 } 455 456 return NOTIFY_DONE; 457} 458 459static __init void init_hrtick(void) 460{ 461 hotcpu_notifier(hotplug_hrtick, 0); 462} 463#else 464/* 465 * Called to set the hrtick timer state. 466 * 467 * called with rq->lock held and irqs disabled 468 */ 469void hrtick_start(struct rq *rq, u64 delay) 470{ 471 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0, 472 HRTIMER_MODE_REL_PINNED, 0); 473} 474 475static inline void init_hrtick(void) 476{ 477} 478#endif /* CONFIG_SMP */ 479 480static void init_rq_hrtick(struct rq *rq) 481{ 482#ifdef CONFIG_SMP 483 rq->hrtick_csd_pending = 0; 484 485 rq->hrtick_csd.flags = 0; 486 rq->hrtick_csd.func = __hrtick_start; 487 rq->hrtick_csd.info = rq; 488#endif 489 490 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); 491 rq->hrtick_timer.function = hrtick; 492} 493#else /* CONFIG_SCHED_HRTICK */ 494static inline void hrtick_clear(struct rq *rq) 495{ 496} 497 498static inline void init_rq_hrtick(struct rq *rq) 499{ 500} 501 502static inline void init_hrtick(void) 503{ 504} 505#endif /* CONFIG_SCHED_HRTICK */ 506 507/* 508 * resched_task - mark a task 'to be rescheduled now'. 509 * 510 * On UP this means the setting of the need_resched flag, on SMP it 511 * might also involve a cross-CPU call to trigger the scheduler on 512 * the target CPU. 513 */ 514#ifdef CONFIG_SMP 515void resched_task(struct task_struct *p) 516{ 517 int cpu; 518 519 assert_raw_spin_locked(&task_rq(p)->lock); 520 521 if (test_tsk_need_resched(p)) 522 return; 523 524 set_tsk_need_resched(p); 525 526 cpu = task_cpu(p); 527 if (cpu == smp_processor_id()) 528 return; 529 530 /* NEED_RESCHED must be visible before we test polling */ 531 smp_mb(); 532 if (!tsk_is_polling(p)) 533 smp_send_reschedule(cpu); 534} 535 536void resched_cpu(int cpu) 537{ 538 struct rq *rq = cpu_rq(cpu); 539 unsigned long flags; 540 541 if (!raw_spin_trylock_irqsave(&rq->lock, flags)) 542 return; 543 resched_task(cpu_curr(cpu)); 544 raw_spin_unlock_irqrestore(&rq->lock, flags); 545} 546 547#ifdef CONFIG_NO_HZ_COMMON 548/* 549 * In the semi idle case, use the nearest busy cpu for migrating timers 550 * from an idle cpu. This is good for power-savings. 551 * 552 * We don't do similar optimization for completely idle system, as 553 * selecting an idle cpu will add more delays to the timers than intended 554 * (as that cpu's timer base may not be uptodate wrt jiffies etc). 555 */ 556int get_nohz_timer_target(void) 557{ 558 int cpu = smp_processor_id(); 559 int i; 560 struct sched_domain *sd; 561 562 rcu_read_lock(); 563 for_each_domain(cpu, sd) { 564 for_each_cpu(i, sched_domain_span(sd)) { 565 if (!idle_cpu(i)) { 566 cpu = i; 567 goto unlock; 568 } 569 } 570 } 571unlock: 572 rcu_read_unlock(); 573 return cpu; 574} 575/* 576 * When add_timer_on() enqueues a timer into the timer wheel of an 577 * idle CPU then this timer might expire before the next timer event 578 * which is scheduled to wake up that CPU. In case of a completely 579 * idle system the next event might even be infinite time into the 580 * future. wake_up_idle_cpu() ensures that the CPU is woken up and 581 * leaves the inner idle loop so the newly added timer is taken into 582 * account when the CPU goes back to idle and evaluates the timer 583 * wheel for the next timer event. 584 */ 585static void wake_up_idle_cpu(int cpu) 586{ 587 struct rq *rq = cpu_rq(cpu); 588 589 if (cpu == smp_processor_id()) 590 return; 591 592 /* 593 * This is safe, as this function is called with the timer 594 * wheel base lock of (cpu) held. When the CPU is on the way 595 * to idle and has not yet set rq->curr to idle then it will 596 * be serialized on the timer wheel base lock and take the new 597 * timer into account automatically. 598 */ 599 if (rq->curr != rq->idle) 600 return; 601 602 /* 603 * We can set TIF_RESCHED on the idle task of the other CPU 604 * lockless. The worst case is that the other CPU runs the 605 * idle task through an additional NOOP schedule() 606 */ 607 set_tsk_need_resched(rq->idle); 608 609 /* NEED_RESCHED must be visible before we test polling */ 610 smp_mb(); 611 if (!tsk_is_polling(rq->idle)) 612 smp_send_reschedule(cpu); 613} 614 615static bool wake_up_full_nohz_cpu(int cpu) 616{ 617 if (tick_nohz_full_cpu(cpu)) { 618 if (cpu != smp_processor_id() || 619 tick_nohz_tick_stopped()) 620 smp_send_reschedule(cpu); 621 return true; 622 } 623 624 return false; 625} 626 627void wake_up_nohz_cpu(int cpu) 628{ 629 if (!wake_up_full_nohz_cpu(cpu)) 630 wake_up_idle_cpu(cpu); 631} 632 633static inline bool got_nohz_idle_kick(void) 634{ 635 int cpu = smp_processor_id(); 636 return idle_cpu(cpu) && test_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu)); 637} 638 639#else /* CONFIG_NO_HZ_COMMON */ 640 641static inline bool got_nohz_idle_kick(void) 642{ 643 return false; 644} 645 646#endif /* CONFIG_NO_HZ_COMMON */ 647 648#ifdef CONFIG_NO_HZ_FULL 649bool sched_can_stop_tick(void) 650{ 651 struct rq *rq; 652 653 rq = this_rq(); 654 655 /* Make sure rq->nr_running update is visible after the IPI */ 656 smp_rmb(); 657 658 /* More than one running task need preemption */ 659 if (rq->nr_running > 1) 660 return false; 661 662 return true; 663} 664#endif /* CONFIG_NO_HZ_FULL */ 665 666void sched_avg_update(struct rq *rq) 667{ 668 s64 period = sched_avg_period(); 669 670 while ((s64)(rq_clock(rq) - rq->age_stamp) > period) { 671 /* 672 * Inline assembly required to prevent the compiler 673 * optimising this loop into a divmod call. 674 * See __iter_div_u64_rem() for another example of this. 675 */ 676 asm("" : "+rm" (rq->age_stamp)); 677 rq->age_stamp += period; 678 rq->rt_avg /= 2; 679 } 680} 681 682#else /* !CONFIG_SMP */ 683void resched_task(struct task_struct *p) 684{ 685 assert_raw_spin_locked(&task_rq(p)->lock); 686 set_tsk_need_resched(p); 687} 688#endif /* CONFIG_SMP */ 689 690#if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \ 691 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH))) 692/* 693 * Iterate task_group tree rooted at *from, calling @down when first entering a 694 * node and @up when leaving it for the final time. 695 * 696 * Caller must hold rcu_lock or sufficient equivalent. 697 */ 698int walk_tg_tree_from(struct task_group *from, 699 tg_visitor down, tg_visitor up, void *data) 700{ 701 struct task_group *parent, *child; 702 int ret; 703 704 parent = from; 705 706down: 707 ret = (*down)(parent, data); 708 if (ret) 709 goto out; 710 list_for_each_entry_rcu(child, &parent->children, siblings) { 711 parent = child; 712 goto down; 713 714up: 715 continue; 716 } 717 ret = (*up)(parent, data); 718 if (ret || parent == from) 719 goto out; 720 721 child = parent; 722 parent = parent->parent; 723 if (parent) 724 goto up; 725out: 726 return ret; 727} 728 729int tg_nop(struct task_group *tg, void *data) 730{ 731 return 0; 732} 733#endif 734 735static void set_load_weight(struct task_struct *p) 736{ 737 int prio = p->static_prio - MAX_RT_PRIO; 738 struct load_weight *load = &p->se.load; 739 740 /* 741 * SCHED_IDLE tasks get minimal weight: 742 */ 743 if (p->policy == SCHED_IDLE) { 744 load->weight = scale_load(WEIGHT_IDLEPRIO); 745 load->inv_weight = WMULT_IDLEPRIO; 746 return; 747 } 748 749 load->weight = scale_load(prio_to_weight[prio]); 750 load->inv_weight = prio_to_wmult[prio]; 751} 752 753static void enqueue_task(struct rq *rq, struct task_struct *p, int flags) 754{ 755 update_rq_clock(rq); 756 sched_info_queued(p); 757 p->sched_class->enqueue_task(rq, p, flags); 758} 759 760static void dequeue_task(struct rq *rq, struct task_struct *p, int flags) 761{ 762 update_rq_clock(rq); 763 sched_info_dequeued(p); 764 p->sched_class->dequeue_task(rq, p, flags); 765} 766 767void activate_task(struct rq *rq, struct task_struct *p, int flags) 768{ 769 if (task_contributes_to_load(p)) 770 rq->nr_uninterruptible--; 771 772 enqueue_task(rq, p, flags); 773} 774 775void deactivate_task(struct rq *rq, struct task_struct *p, int flags) 776{ 777 if (task_contributes_to_load(p)) 778 rq->nr_uninterruptible++; 779 780 dequeue_task(rq, p, flags); 781} 782 783static void update_rq_clock_task(struct rq *rq, s64 delta) 784{ 785/* 786 * In theory, the compile should just see 0 here, and optimize out the call 787 * to sched_rt_avg_update. But I don't trust it... 788 */ 789#if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING) 790 s64 steal = 0, irq_delta = 0; 791#endif 792#ifdef CONFIG_IRQ_TIME_ACCOUNTING 793 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time; 794 795 /* 796 * Since irq_time is only updated on {soft,}irq_exit, we might run into 797 * this case when a previous update_rq_clock() happened inside a 798 * {soft,}irq region. 799 * 800 * When this happens, we stop ->clock_task and only update the 801 * prev_irq_time stamp to account for the part that fit, so that a next 802 * update will consume the rest. This ensures ->clock_task is 803 * monotonic. 804 * 805 * It does however cause some slight miss-attribution of {soft,}irq 806 * time, a more accurate solution would be to update the irq_time using 807 * the current rq->clock timestamp, except that would require using 808 * atomic ops. 809 */ 810 if (irq_delta > delta) 811 irq_delta = delta; 812 813 rq->prev_irq_time += irq_delta; 814 delta -= irq_delta; 815#endif 816#ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING 817 if (static_key_false((¶virt_steal_rq_enabled))) { 818 u64 st; 819 820 steal = paravirt_steal_clock(cpu_of(rq)); 821 steal -= rq->prev_steal_time_rq; 822 823 if (unlikely(steal > delta)) 824 steal = delta; 825 826 st = steal_ticks(steal); 827 steal = st * TICK_NSEC; 828 829 rq->prev_steal_time_rq += steal; 830 831 delta -= steal; 832 } 833#endif 834 835 rq->clock_task += delta; 836 837#if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING) 838 if ((irq_delta + steal) && sched_feat(NONTASK_POWER)) 839 sched_rt_avg_update(rq, irq_delta + steal); 840#endif 841} 842 843void sched_set_stop_task(int cpu, struct task_struct *stop) 844{ 845 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 }; 846 struct task_struct *old_stop = cpu_rq(cpu)->stop; 847 848 if (stop) { 849 /* 850 * Make it appear like a SCHED_FIFO task, its something 851 * userspace knows about and won't get confused about. 852 * 853 * Also, it will make PI more or less work without too 854 * much confusion -- but then, stop work should not 855 * rely on PI working anyway. 856 */ 857 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m); 858 859 stop->sched_class = &stop_sched_class; 860 } 861 862 cpu_rq(cpu)->stop = stop; 863 864 if (old_stop) { 865 /* 866 * Reset it back to a normal scheduling class so that 867 * it can die in pieces. 868 */ 869 old_stop->sched_class = &rt_sched_class; 870 } 871} 872 873/* 874 * __normal_prio - return the priority that is based on the static prio 875 */ 876static inline int __normal_prio(struct task_struct *p) 877{ 878 return p->static_prio; 879} 880 881/* 882 * Calculate the expected normal priority: i.e. priority 883 * without taking RT-inheritance into account. Might be 884 * boosted by interactivity modifiers. Changes upon fork, 885 * setprio syscalls, and whenever the interactivity 886 * estimator recalculates. 887 */ 888static inline int normal_prio(struct task_struct *p) 889{ 890 int prio; 891 892 if (task_has_rt_policy(p)) 893 prio = MAX_RT_PRIO-1 - p->rt_priority; 894 else 895 prio = __normal_prio(p); 896 return prio; 897} 898 899/* 900 * Calculate the current priority, i.e. the priority 901 * taken into account by the scheduler. This value might 902 * be boosted by RT tasks, or might be boosted by 903 * interactivity modifiers. Will be RT if the task got 904 * RT-boosted. If not then it returns p->normal_prio. 905 */ 906static int effective_prio(struct task_struct *p) 907{ 908 p->normal_prio = normal_prio(p); 909 /* 910 * If we are RT tasks or we were boosted to RT priority, 911 * keep the priority unchanged. Otherwise, update priority 912 * to the normal priority: 913 */ 914 if (!rt_prio(p->prio)) 915 return p->normal_prio; 916 return p->prio; 917} 918 919/** 920 * task_curr - is this task currently executing on a CPU? 921 * @p: the task in question. 922 */ 923inline int task_curr(const struct task_struct *p) 924{ 925 return cpu_curr(task_cpu(p)) == p; 926} 927 928static inline void check_class_changed(struct rq *rq, struct task_struct *p, 929 const struct sched_class *prev_class, 930 int oldprio) 931{ 932 if (prev_class != p->sched_class) { 933 if (prev_class->switched_from) 934 prev_class->switched_from(rq, p); 935 p->sched_class->switched_to(rq, p); 936 } else if (oldprio != p->prio) 937 p->sched_class->prio_changed(rq, p, oldprio); 938} 939 940void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags) 941{ 942 const struct sched_class *class; 943 944 if (p->sched_class == rq->curr->sched_class) { 945 rq->curr->sched_class->check_preempt_curr(rq, p, flags); 946 } else { 947 for_each_class(class) { 948 if (class == rq->curr->sched_class) 949 break; 950 if (class == p->sched_class) { 951 resched_task(rq->curr); 952 break; 953 } 954 } 955 } 956 957 /* 958 * A queue event has occurred, and we're going to schedule. In 959 * this case, we can save a useless back to back clock update. 960 */ 961 if (rq->curr->on_rq && test_tsk_need_resched(rq->curr)) 962 rq->skip_clock_update = 1; 963} 964 965static ATOMIC_NOTIFIER_HEAD(task_migration_notifier); 966 967void register_task_migration_notifier(struct notifier_block *n) 968{ 969 atomic_notifier_chain_register(&task_migration_notifier, n); 970} 971 972#ifdef CONFIG_SMP 973void set_task_cpu(struct task_struct *p, unsigned int new_cpu) 974{ 975#ifdef CONFIG_SCHED_DEBUG 976 /* 977 * We should never call set_task_cpu() on a blocked task, 978 * ttwu() will sort out the placement. 979 */ 980 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING && 981 !(task_thread_info(p)->preempt_count & PREEMPT_ACTIVE)); 982 983#ifdef CONFIG_LOCKDEP 984 /* 985 * The caller should hold either p->pi_lock or rq->lock, when changing 986 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks. 987 * 988 * sched_move_task() holds both and thus holding either pins the cgroup, 989 * see task_group(). 990 * 991 * Furthermore, all task_rq users should acquire both locks, see 992 * task_rq_lock(). 993 */ 994 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) || 995 lockdep_is_held(&task_rq(p)->lock))); 996#endif 997#endif 998 999 trace_sched_migrate_task(p, new_cpu); 1000 1001 if (task_cpu(p) != new_cpu) { 1002 struct task_migration_notifier tmn; 1003 1004 if (p->sched_class->migrate_task_rq) 1005 p->sched_class->migrate_task_rq(p, new_cpu); 1006 p->se.nr_migrations++; 1007 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, NULL, 0); 1008 1009 tmn.task = p; 1010 tmn.from_cpu = task_cpu(p); 1011 tmn.to_cpu = new_cpu; 1012 1013 atomic_notifier_call_chain(&task_migration_notifier, 0, &tmn); 1014 } 1015 1016 __set_task_cpu(p, new_cpu); 1017} 1018 1019struct migration_arg { 1020 struct task_struct *task; 1021 int dest_cpu; 1022}; 1023 1024static int migration_cpu_stop(void *data); 1025 1026/* 1027 * wait_task_inactive - wait for a thread to unschedule. 1028 * 1029 * If @match_state is nonzero, it's the @p->state value just checked and 1030 * not expected to change. If it changes, i.e. @p might have woken up, 1031 * then return zero. When we succeed in waiting for @p to be off its CPU, 1032 * we return a positive number (its total switch count). If a second call 1033 * a short while later returns the same number, the caller can be sure that 1034 * @p has remained unscheduled the whole time. 1035 * 1036 * The caller must ensure that the task *will* unschedule sometime soon, 1037 * else this function might spin for a *long* time. This function can't 1038 * be called with interrupts off, or it may introduce deadlock with 1039 * smp_call_function() if an IPI is sent by the same process we are 1040 * waiting to become inactive. 1041 */ 1042unsigned long wait_task_inactive(struct task_struct *p, long match_state) 1043{ 1044 unsigned long flags; 1045 int running, on_rq; 1046 unsigned long ncsw; 1047 struct rq *rq; 1048 1049 for (;;) { 1050 /* 1051 * We do the initial early heuristics without holding 1052 * any task-queue locks at all. We'll only try to get 1053 * the runqueue lock when things look like they will 1054 * work out! 1055 */ 1056 rq = task_rq(p); 1057 1058 /* 1059 * If the task is actively running on another CPU 1060 * still, just relax and busy-wait without holding 1061 * any locks. 1062 * 1063 * NOTE! Since we don't hold any locks, it's not 1064 * even sure that "rq" stays as the right runqueue! 1065 * But we don't care, since "task_running()" will 1066 * return false if the runqueue has changed and p 1067 * is actually now running somewhere else! 1068 */ 1069 while (task_running(rq, p)) { 1070 if (match_state && unlikely(p->state != match_state)) 1071 return 0; 1072 cpu_relax(); 1073 } 1074 1075 /* 1076 * Ok, time to look more closely! We need the rq 1077 * lock now, to be *sure*. If we're wrong, we'll 1078 * just go back and repeat. 1079 */ 1080 rq = task_rq_lock(p, &flags); 1081 trace_sched_wait_task(p); 1082 running = task_running(rq, p); 1083 on_rq = p->on_rq; 1084 ncsw = 0; 1085 if (!match_state || p->state == match_state) 1086 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */ 1087 task_rq_unlock(rq, p, &flags); 1088 1089 /* 1090 * If it changed from the expected state, bail out now. 1091 */ 1092 if (unlikely(!ncsw)) 1093 break; 1094 1095 /* 1096 * Was it really running after all now that we 1097 * checked with the proper locks actually held? 1098 * 1099 * Oops. Go back and try again.. 1100 */ 1101 if (unlikely(running)) { 1102 cpu_relax(); 1103 continue; 1104 } 1105 1106 /* 1107 * It's not enough that it's not actively running, 1108 * it must be off the runqueue _entirely_, and not 1109 * preempted! 1110 * 1111 * So if it was still runnable (but just not actively 1112 * running right now), it's preempted, and we should 1113 * yield - it could be a while. 1114 */ 1115 if (unlikely(on_rq)) { 1116 ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ); 1117 1118 set_current_state(TASK_UNINTERRUPTIBLE); 1119 schedule_hrtimeout(&to, HRTIMER_MODE_REL); 1120 continue; 1121 } 1122 1123 /* 1124 * Ahh, all good. It wasn't running, and it wasn't 1125 * runnable, which means that it will never become 1126 * running in the future either. We're all done! 1127 */ 1128 break; 1129 } 1130 1131 return ncsw; 1132} 1133 1134/*** 1135 * kick_process - kick a running thread to enter/exit the kernel 1136 * @p: the to-be-kicked thread 1137 * 1138 * Cause a process which is running on another CPU to enter 1139 * kernel-mode, without any delay. (to get signals handled.) 1140 * 1141 * NOTE: this function doesn't have to take the runqueue lock, 1142 * because all it wants to ensure is that the remote task enters 1143 * the kernel. If the IPI races and the task has been migrated 1144 * to another CPU then no harm is done and the purpose has been 1145 * achieved as well. 1146 */ 1147void kick_process(struct task_struct *p) 1148{ 1149 int cpu; 1150 1151 preempt_disable(); 1152 cpu = task_cpu(p); 1153 if ((cpu != smp_processor_id()) && task_curr(p)) 1154 smp_send_reschedule(cpu); 1155 preempt_enable(); 1156} 1157EXPORT_SYMBOL_GPL(kick_process); 1158#endif /* CONFIG_SMP */ 1159 1160#ifdef CONFIG_SMP 1161/* 1162 * ->cpus_allowed is protected by both rq->lock and p->pi_lock 1163 */ 1164static int select_fallback_rq(int cpu, struct task_struct *p) 1165{ 1166 int nid = cpu_to_node(cpu); 1167 const struct cpumask *nodemask = NULL; 1168 enum { cpuset, possible, fail } state = cpuset; 1169 int dest_cpu; 1170 1171 /* 1172 * If the node that the cpu is on has been offlined, cpu_to_node() 1173 * will return -1. There is no cpu on the node, and we should 1174 * select the cpu on the other node. 1175 */ 1176 if (nid != -1) { 1177 nodemask = cpumask_of_node(nid); 1178 1179 /* Look for allowed, online CPU in same node. */ 1180 for_each_cpu(dest_cpu, nodemask) { 1181 if (!cpu_online(dest_cpu)) 1182 continue; 1183 if (!cpu_active(dest_cpu)) 1184 continue; 1185 if (cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p))) 1186 return dest_cpu; 1187 } 1188 } 1189 1190 for (;;) { 1191 /* Any allowed, online CPU? */ 1192 for_each_cpu(dest_cpu, tsk_cpus_allowed(p)) { 1193 if (!cpu_online(dest_cpu)) 1194 continue; 1195 if (!cpu_active(dest_cpu)) 1196 continue; 1197 goto out; 1198 } 1199 1200 switch (state) { 1201 case cpuset: 1202 /* No more Mr. Nice Guy. */ 1203 cpuset_cpus_allowed_fallback(p); 1204 state = possible; 1205 break; 1206 1207 case possible: 1208 do_set_cpus_allowed(p, cpu_possible_mask); 1209 state = fail; 1210 break; 1211 1212 case fail: 1213 BUG(); 1214 break; 1215 } 1216 } 1217 1218out: 1219 if (state != cpuset) { 1220 /* 1221 * Don't tell them about moving exiting tasks or 1222 * kernel threads (both mm NULL), since they never 1223 * leave kernel. 1224 */ 1225 if (p->mm && printk_ratelimit()) { 1226 printk_sched("process %d (%s) no longer affine to cpu%d\n", 1227 task_pid_nr(p), p->comm, cpu); 1228 } 1229 } 1230 1231 return dest_cpu; 1232} 1233 1234/* 1235 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable. 1236 */ 1237static inline 1238int select_task_rq(struct task_struct *p, int sd_flags, int wake_flags) 1239{ 1240 int cpu = p->sched_class->select_task_rq(p, sd_flags, wake_flags); 1241 1242 /* 1243 * In order not to call set_task_cpu() on a blocking task we need 1244 * to rely on ttwu() to place the task on a valid ->cpus_allowed 1245 * cpu. 1246 * 1247 * Since this is common to all placement strategies, this lives here. 1248 * 1249 * [ this allows ->select_task() to simply return task_cpu(p) and 1250 * not worry about this generic constraint ] 1251 */ 1252 if (unlikely(!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)) || 1253 !cpu_online(cpu))) 1254 cpu = select_fallback_rq(task_cpu(p), p); 1255 1256 return cpu; 1257} 1258 1259static void update_avg(u64 *avg, u64 sample) 1260{ 1261 s64 diff = sample - *avg; 1262 *avg += diff >> 3; 1263} 1264#endif 1265 1266static void 1267ttwu_stat(struct task_struct *p, int cpu, int wake_flags) 1268{ 1269#ifdef CONFIG_SCHEDSTATS 1270 struct rq *rq = this_rq(); 1271 1272#ifdef CONFIG_SMP 1273 int this_cpu = smp_processor_id(); 1274 1275 if (cpu == this_cpu) { 1276 schedstat_inc(rq, ttwu_local); 1277 schedstat_inc(p, se.statistics.nr_wakeups_local); 1278 } else { 1279 struct sched_domain *sd; 1280 1281 schedstat_inc(p, se.statistics.nr_wakeups_remote); 1282 rcu_read_lock(); 1283 for_each_domain(this_cpu, sd) { 1284 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) { 1285 schedstat_inc(sd, ttwu_wake_remote); 1286 break; 1287 } 1288 } 1289 rcu_read_unlock(); 1290 } 1291 1292 if (wake_flags & WF_MIGRATED) 1293 schedstat_inc(p, se.statistics.nr_wakeups_migrate); 1294 1295#endif /* CONFIG_SMP */ 1296 1297 schedstat_inc(rq, ttwu_count); 1298 schedstat_inc(p, se.statistics.nr_wakeups); 1299 1300 if (wake_flags & WF_SYNC) 1301 schedstat_inc(p, se.statistics.nr_wakeups_sync); 1302 1303#endif /* CONFIG_SCHEDSTATS */ 1304} 1305 1306static void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags) 1307{ 1308 activate_task(rq, p, en_flags); 1309 p->on_rq = 1; 1310 1311 /* if a worker is waking up, notify workqueue */ 1312 if (p->flags & PF_WQ_WORKER) 1313 wq_worker_waking_up(p, cpu_of(rq)); 1314} 1315 1316/* 1317 * Mark the task runnable and perform wakeup-preemption. 1318 */ 1319static void 1320ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags) 1321{ 1322 check_preempt_curr(rq, p, wake_flags); 1323 trace_sched_wakeup(p, true); 1324 1325 p->state = TASK_RUNNING; 1326#ifdef CONFIG_SMP 1327 if (p->sched_class->task_woken) 1328 p->sched_class->task_woken(rq, p); 1329 1330 if (rq->idle_stamp) { 1331 u64 delta = rq_clock(rq) - rq->idle_stamp; 1332 u64 max = 2*sysctl_sched_migration_cost; 1333 1334 if (delta > max) 1335 rq->avg_idle = max; 1336 else 1337 update_avg(&rq->avg_idle, delta); 1338 rq->idle_stamp = 0; 1339 } 1340#endif 1341} 1342 1343static void 1344ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags) 1345{ 1346#ifdef CONFIG_SMP 1347 if (p->sched_contributes_to_load) 1348 rq->nr_uninterruptible--; 1349#endif 1350 1351 ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_WAKING); 1352 ttwu_do_wakeup(rq, p, wake_flags); 1353} 1354 1355/* 1356 * Called in case the task @p isn't fully descheduled from its runqueue, 1357 * in this case we must do a remote wakeup. Its a 'light' wakeup though, 1358 * since all we need to do is flip p->state to TASK_RUNNING, since 1359 * the task is still ->on_rq. 1360 */ 1361static int ttwu_remote(struct task_struct *p, int wake_flags) 1362{ 1363 struct rq *rq; 1364 int ret = 0; 1365 1366 rq = __task_rq_lock(p); 1367 if (p->on_rq) { 1368 /* check_preempt_curr() may use rq clock */ 1369 update_rq_clock(rq); 1370 ttwu_do_wakeup(rq, p, wake_flags); 1371 ret = 1; 1372 } 1373 __task_rq_unlock(rq); 1374 1375 return ret; 1376} 1377 1378#ifdef CONFIG_SMP 1379static void sched_ttwu_pending(void) 1380{ 1381 struct rq *rq = this_rq(); 1382 struct llist_node *llist = llist_del_all(&rq->wake_list); 1383 struct task_struct *p; 1384 1385 raw_spin_lock(&rq->lock); 1386 1387 while (llist) { 1388 p = llist_entry(llist, struct task_struct, wake_entry); 1389 llist = llist_next(llist); 1390 ttwu_do_activate(rq, p, 0); 1391 } 1392 1393 raw_spin_unlock(&rq->lock); 1394} 1395 1396void scheduler_ipi(void) 1397{ 1398 if (llist_empty(&this_rq()->wake_list) && !got_nohz_idle_kick() 1399 && !tick_nohz_full_cpu(smp_processor_id())) 1400 return; 1401 1402 /* 1403 * Not all reschedule IPI handlers call irq_enter/irq_exit, since 1404 * traditionally all their work was done from the interrupt return 1405 * path. Now that we actually do some work, we need to make sure 1406 * we do call them. 1407 * 1408 * Some archs already do call them, luckily irq_enter/exit nest 1409 * properly. 1410 * 1411 * Arguably we should visit all archs and update all handlers, 1412 * however a fair share of IPIs are still resched only so this would 1413 * somewhat pessimize the simple resched case. 1414 */ 1415 irq_enter(); 1416 tick_nohz_full_check(); 1417 sched_ttwu_pending(); 1418 1419 /* 1420 * Check if someone kicked us for doing the nohz idle load balance. 1421 */ 1422 if (unlikely(got_nohz_idle_kick() && !need_resched())) { 1423 this_rq()->idle_balance = 1; 1424 raise_softirq_irqoff(SCHED_SOFTIRQ); 1425 } 1426 irq_exit(); 1427} 1428 1429static void ttwu_queue_remote(struct task_struct *p, int cpu) 1430{ 1431 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list)) 1432 smp_send_reschedule(cpu); 1433} 1434 1435bool cpus_share_cache(int this_cpu, int that_cpu) 1436{ 1437 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu); 1438} 1439#endif /* CONFIG_SMP */ 1440 1441static void ttwu_queue(struct task_struct *p, int cpu) 1442{ 1443 struct rq *rq = cpu_rq(cpu); 1444 1445#if defined(CONFIG_SMP) 1446 if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) { 1447 sched_clock_cpu(cpu); /* sync clocks x-cpu */ 1448 ttwu_queue_remote(p, cpu); 1449 return; 1450 } 1451#endif 1452 1453 raw_spin_lock(&rq->lock); 1454 ttwu_do_activate(rq, p, 0); 1455 raw_spin_unlock(&rq->lock); 1456} 1457 1458/** 1459 * try_to_wake_up - wake up a thread 1460 * @p: the thread to be awakened 1461 * @state: the mask of task states that can be woken 1462 * @wake_flags: wake modifier flags (WF_*) 1463 * 1464 * Put it on the run-queue if it's not already there. The "current" 1465 * thread is always on the run-queue (except when the actual 1466 * re-schedule is in progress), and as such you're allowed to do 1467 * the simpler "current->state = TASK_RUNNING" to mark yourself 1468 * runnable without the overhead of this. 1469 * 1470 * Returns %true if @p was woken up, %false if it was already running 1471 * or @state didn't match @p's state. 1472 */ 1473static int 1474try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags) 1475{ 1476 unsigned long flags; 1477 int cpu, success = 0; 1478 1479 smp_wmb(); 1480 raw_spin_lock_irqsave(&p->pi_lock, flags); 1481 if (!(p->state & state)) 1482 goto out; 1483 1484 success = 1; /* we're going to change ->state */ 1485 cpu = task_cpu(p); 1486 1487 if (p->on_rq && ttwu_remote(p, wake_flags)) 1488 goto stat; 1489 1490#ifdef CONFIG_SMP 1491 /* 1492 * If the owning (remote) cpu is still in the middle of schedule() with 1493 * this task as prev, wait until its done referencing the task. 1494 */ 1495 while (p->on_cpu) 1496 cpu_relax(); 1497 /* 1498 * Pairs with the smp_wmb() in finish_lock_switch(). 1499 */ 1500 smp_rmb(); 1501 1502 p->sched_contributes_to_load = !!task_contributes_to_load(p); 1503 p->state = TASK_WAKING; 1504 1505 if (p->sched_class->task_waking) 1506 p->sched_class->task_waking(p); 1507 1508 cpu = select_task_rq(p, SD_BALANCE_WAKE, wake_flags); 1509 if (task_cpu(p) != cpu) { 1510 wake_flags |= WF_MIGRATED; 1511 set_task_cpu(p, cpu); 1512 } 1513#endif /* CONFIG_SMP */ 1514 1515 ttwu_queue(p, cpu); 1516stat: 1517 ttwu_stat(p, cpu, wake_flags); 1518out: 1519 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 1520 1521 return success; 1522} 1523 1524/** 1525 * try_to_wake_up_local - try to wake up a local task with rq lock held 1526 * @p: the thread to be awakened 1527 * 1528 * Put @p on the run-queue if it's not already there. The caller must 1529 * ensure that this_rq() is locked, @p is bound to this_rq() and not 1530 * the current task. 1531 */ 1532static void try_to_wake_up_local(struct task_struct *p) 1533{ 1534 struct rq *rq = task_rq(p); 1535 1536 if (WARN_ON_ONCE(rq != this_rq()) || 1537 WARN_ON_ONCE(p == current)) 1538 return; 1539 1540 lockdep_assert_held(&rq->lock); 1541 1542 if (!raw_spin_trylock(&p->pi_lock)) { 1543 raw_spin_unlock(&rq->lock); 1544 raw_spin_lock(&p->pi_lock); 1545 raw_spin_lock(&rq->lock); 1546 } 1547 1548 if (!(p->state & TASK_NORMAL)) 1549 goto out; 1550 1551 if (!p->on_rq) 1552 ttwu_activate(rq, p, ENQUEUE_WAKEUP); 1553 1554 ttwu_do_wakeup(rq, p, 0); 1555 ttwu_stat(p, smp_processor_id(), 0); 1556out: 1557 raw_spin_unlock(&p->pi_lock); 1558} 1559 1560/** 1561 * wake_up_process - Wake up a specific process 1562 * @p: The process to be woken up. 1563 * 1564 * Attempt to wake up the nominated process and move it to the set of runnable 1565 * processes. Returns 1 if the process was woken up, 0 if it was already 1566 * running. 1567 * 1568 * It may be assumed that this function implies a write memory barrier before 1569 * changing the task state if and only if any tasks are woken up. 1570 */ 1571int wake_up_process(struct task_struct *p) 1572{ 1573 WARN_ON(task_is_stopped_or_traced(p)); 1574 return try_to_wake_up(p, TASK_NORMAL, 0); 1575} 1576EXPORT_SYMBOL(wake_up_process); 1577 1578int wake_up_state(struct task_struct *p, unsigned int state) 1579{ 1580 return try_to_wake_up(p, state, 0); 1581} 1582 1583/* 1584 * Perform scheduler related setup for a newly forked process p. 1585 * p is forked by current. 1586 * 1587 * __sched_fork() is basic setup used by init_idle() too: 1588 */ 1589static void __sched_fork(struct task_struct *p) 1590{ 1591 p->on_rq = 0; 1592 1593 p->se.on_rq = 0; 1594 p->se.exec_start = 0; 1595 p->se.sum_exec_runtime = 0; 1596 p->se.prev_sum_exec_runtime = 0; 1597 p->se.nr_migrations = 0; 1598 p->se.vruntime = 0; 1599 INIT_LIST_HEAD(&p->se.group_node); 1600 1601/* 1602 * Load-tracking only depends on SMP, FAIR_GROUP_SCHED dependency below may be 1603 * removed when useful for applications beyond shares distribution (e.g. 1604 * load-balance). 1605 */ 1606#if defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED) 1607 p->se.avg.runnable_avg_period = 0; 1608 p->se.avg.runnable_avg_sum = 0; 1609#endif 1610#ifdef CONFIG_SCHEDSTATS 1611 memset(&p->se.statistics, 0, sizeof(p->se.statistics)); 1612#endif 1613 1614 INIT_LIST_HEAD(&p->rt.run_list); 1615 1616#ifdef CONFIG_PREEMPT_NOTIFIERS 1617 INIT_HLIST_HEAD(&p->preempt_notifiers); 1618#endif 1619 1620#ifdef CONFIG_NUMA_BALANCING 1621 if (p->mm && atomic_read(&p->mm->mm_users) == 1) { 1622 p->mm->numa_next_scan = jiffies; 1623 p->mm->numa_next_reset = jiffies; 1624 p->mm->numa_scan_seq = 0; 1625 } 1626 1627 p->node_stamp = 0ULL; 1628 p->numa_scan_seq = p->mm ? p->mm->numa_scan_seq : 0; 1629 p->numa_migrate_seq = p->mm ? p->mm->numa_scan_seq - 1 : 0; 1630 p->numa_scan_period = sysctl_numa_balancing_scan_delay; 1631 p->numa_work.next = &p->numa_work; 1632#endif /* CONFIG_NUMA_BALANCING */ 1633} 1634 1635#ifdef CONFIG_NUMA_BALANCING 1636#ifdef CONFIG_SCHED_DEBUG 1637void set_numabalancing_state(bool enabled) 1638{ 1639 if (enabled) 1640 sched_feat_set("NUMA"); 1641 else 1642 sched_feat_set("NO_NUMA"); 1643} 1644#else 1645__read_mostly bool numabalancing_enabled; 1646 1647void set_numabalancing_state(bool enabled) 1648{ 1649 numabalancing_enabled = enabled; 1650} 1651#endif /* CONFIG_SCHED_DEBUG */ 1652#endif /* CONFIG_NUMA_BALANCING */ 1653 1654/* 1655 * fork()/clone()-time setup: 1656 */ 1657void sched_fork(struct task_struct *p) 1658{ 1659 unsigned long flags; 1660 int cpu = get_cpu(); 1661 1662 __sched_fork(p); 1663 /* 1664 * We mark the process as running here. This guarantees that 1665 * nobody will actually run it, and a signal or other external 1666 * event cannot wake it up and insert it on the runqueue either. 1667 */ 1668 p->state = TASK_RUNNING; 1669 1670 /* 1671 * Make sure we do not leak PI boosting priority to the child. 1672 */ 1673 p->prio = current->normal_prio; 1674 1675 /* 1676 * Revert to default priority/policy on fork if requested. 1677 */ 1678 if (unlikely(p->sched_reset_on_fork)) { 1679 if (task_has_rt_policy(p)) { 1680 p->policy = SCHED_NORMAL; 1681 p->static_prio = NICE_TO_PRIO(0); 1682 p->rt_priority = 0; 1683 } else if (PRIO_TO_NICE(p->static_prio) < 0) 1684 p->static_prio = NICE_TO_PRIO(0); 1685 1686 p->prio = p->normal_prio = __normal_prio(p); 1687 set_load_weight(p); 1688 1689 /* 1690 * We don't need the reset flag anymore after the fork. It has 1691 * fulfilled its duty: 1692 */ 1693 p->sched_reset_on_fork = 0; 1694 } 1695 1696 if (!rt_prio(p->prio)) 1697 p->sched_class = &fair_sched_class; 1698 1699 if (p->sched_class->task_fork) 1700 p->sched_class->task_fork(p); 1701 1702 /* 1703 * The child is not yet in the pid-hash so no cgroup attach races, 1704 * and the cgroup is pinned to this child due to cgroup_fork() 1705 * is ran before sched_fork(). 1706 * 1707 * Silence PROVE_RCU. 1708 */ 1709 raw_spin_lock_irqsave(&p->pi_lock, flags); 1710 set_task_cpu(p, cpu); 1711 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 1712 1713#if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT) 1714 if (likely(sched_info_on())) 1715 memset(&p->sched_info, 0, sizeof(p->sched_info)); 1716#endif 1717#if defined(CONFIG_SMP) 1718 p->on_cpu = 0; 1719#endif 1720#ifdef CONFIG_PREEMPT_COUNT 1721 /* Want to start with kernel preemption disabled. */ 1722 task_thread_info(p)->preempt_count = 1; 1723#endif 1724#ifdef CONFIG_SMP 1725 plist_node_init(&p->pushable_tasks, MAX_PRIO); 1726#endif 1727 1728 put_cpu(); 1729} 1730 1731/* 1732 * wake_up_new_task - wake up a newly created task for the first time. 1733 * 1734 * This function will do some initial scheduler statistics housekeeping 1735 * that must be done for every newly created context, then puts the task 1736 * on the runqueue and wakes it. 1737 */ 1738void wake_up_new_task(struct task_struct *p) 1739{ 1740 unsigned long flags; 1741 struct rq *rq; 1742 1743 raw_spin_lock_irqsave(&p->pi_lock, flags); 1744#ifdef CONFIG_SMP 1745 /* 1746 * Fork balancing, do it here and not earlier because: 1747 * - cpus_allowed can change in the fork path 1748 * - any previously selected cpu might disappear through hotplug 1749 */ 1750 set_task_cpu(p, select_task_rq(p, SD_BALANCE_FORK, 0)); 1751#endif 1752 1753 rq = __task_rq_lock(p); 1754 activate_task(rq, p, 0); 1755 p->on_rq = 1; 1756 trace_sched_wakeup_new(p, true); 1757 check_preempt_curr(rq, p, WF_FORK); 1758#ifdef CONFIG_SMP 1759 if (p->sched_class->task_woken) 1760 p->sched_class->task_woken(rq, p); 1761#endif 1762 task_rq_unlock(rq, p, &flags); 1763} 1764 1765#ifdef CONFIG_PREEMPT_NOTIFIERS 1766 1767/** 1768 * preempt_notifier_register - tell me when current is being preempted & rescheduled 1769 * @notifier: notifier struct to register 1770 */ 1771void preempt_notifier_register(struct preempt_notifier *notifier) 1772{ 1773 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers); 1774} 1775EXPORT_SYMBOL_GPL(preempt_notifier_register); 1776 1777/** 1778 * preempt_notifier_unregister - no longer interested in preemption notifications 1779 * @notifier: notifier struct to unregister 1780 * 1781 * This is safe to call from within a preemption notifier. 1782 */ 1783void preempt_notifier_unregister(struct preempt_notifier *notifier) 1784{ 1785 hlist_del(¬ifier->link); 1786} 1787EXPORT_SYMBOL_GPL(preempt_notifier_unregister); 1788 1789static void fire_sched_in_preempt_notifiers(struct task_struct *curr) 1790{ 1791 struct preempt_notifier *notifier; 1792 1793 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link) 1794 notifier->ops->sched_in(notifier, raw_smp_processor_id()); 1795} 1796 1797static void 1798fire_sched_out_preempt_notifiers(struct task_struct *curr, 1799 struct task_struct *next) 1800{ 1801 struct preempt_notifier *notifier; 1802 1803 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link) 1804 notifier->ops->sched_out(notifier, next); 1805} 1806 1807#else /* !CONFIG_PREEMPT_NOTIFIERS */ 1808 1809static void fire_sched_in_preempt_notifiers(struct task_struct *curr) 1810{ 1811} 1812 1813static void 1814fire_sched_out_preempt_notifiers(struct task_struct *curr, 1815 struct task_struct *next) 1816{ 1817} 1818 1819#endif /* CONFIG_PREEMPT_NOTIFIERS */ 1820 1821/** 1822 * prepare_task_switch - prepare to switch tasks 1823 * @rq: the runqueue preparing to switch 1824 * @prev: the current task that is being switched out 1825 * @next: the task we are going to switch to. 1826 * 1827 * This is called with the rq lock held and interrupts off. It must 1828 * be paired with a subsequent finish_task_switch after the context 1829 * switch. 1830 * 1831 * prepare_task_switch sets up locking and calls architecture specific 1832 * hooks. 1833 */ 1834static inline void 1835prepare_task_switch(struct rq *rq, struct task_struct *prev, 1836 struct task_struct *next) 1837{ 1838 trace_sched_switch(prev, next); 1839 sched_info_switch(prev, next); 1840 perf_event_task_sched_out(prev, next); 1841 fire_sched_out_preempt_notifiers(prev, next); 1842 prepare_lock_switch(rq, next); 1843 prepare_arch_switch(next); 1844} 1845 1846/** 1847 * finish_task_switch - clean up after a task-switch 1848 * @rq: runqueue associated with task-switch 1849 * @prev: the thread we just switched away from. 1850 * 1851 * finish_task_switch must be called after the context switch, paired 1852 * with a prepare_task_switch call before the context switch. 1853 * finish_task_switch will reconcile locking set up by prepare_task_switch, 1854 * and do any other architecture-specific cleanup actions. 1855 * 1856 * Note that we may have delayed dropping an mm in context_switch(). If 1857 * so, we finish that here outside of the runqueue lock. (Doing it 1858 * with the lock held can cause deadlocks; see schedule() for 1859 * details.) 1860 */ 1861static void finish_task_switch(struct rq *rq, struct task_struct *prev) 1862 __releases(rq->lock) 1863{ 1864 struct mm_struct *mm = rq->prev_mm; 1865 long prev_state; 1866 1867 rq->prev_mm = NULL; 1868 1869 /* 1870 * A task struct has one reference for the use as "current". 1871 * If a task dies, then it sets TASK_DEAD in tsk->state and calls 1872 * schedule one last time. The schedule call will never return, and 1873 * the scheduled task must drop that reference. 1874 * The test for TASK_DEAD must occur while the runqueue locks are 1875 * still held, otherwise prev could be scheduled on another cpu, die 1876 * there before we look at prev->state, and then the reference would 1877 * be dropped twice. 1878 * Manfred Spraul <manfred@colorfullife.com> 1879 */ 1880 prev_state = prev->state; 1881 vtime_task_switch(prev); 1882 finish_arch_switch(prev); 1883 perf_event_task_sched_in(prev, current); 1884 finish_lock_switch(rq, prev); 1885 finish_arch_post_lock_switch(); 1886 1887 fire_sched_in_preempt_notifiers(current); 1888 if (mm) 1889 mmdrop(mm); 1890 if (unlikely(prev_state == TASK_DEAD)) { 1891 /* 1892 * Remove function-return probe instances associated with this 1893 * task and put them back on the free list. 1894 */ 1895 kprobe_flush_task(prev); 1896 put_task_struct(prev); 1897 } 1898 1899 tick_nohz_task_switch(current); 1900} 1901 1902#ifdef CONFIG_SMP 1903 1904/* assumes rq->lock is held */ 1905static inline void pre_schedule(struct rq *rq, struct task_struct *prev) 1906{ 1907 if (prev->sched_class->pre_schedule) 1908 prev->sched_class->pre_schedule(rq, prev); 1909} 1910 1911/* rq->lock is NOT held, but preemption is disabled */ 1912static inline void post_schedule(struct rq *rq) 1913{ 1914 if (rq->post_schedule) { 1915 unsigned long flags; 1916 1917 raw_spin_lock_irqsave(&rq->lock, flags); 1918 if (rq->curr->sched_class->post_schedule) 1919 rq->curr->sched_class->post_schedule(rq); 1920 raw_spin_unlock_irqrestore(&rq->lock, flags); 1921 1922 rq->post_schedule = 0; 1923 } 1924} 1925 1926#else 1927 1928static inline void pre_schedule(struct rq *rq, struct task_struct *p) 1929{ 1930} 1931 1932static inline void post_schedule(struct rq *rq) 1933{ 1934} 1935 1936#endif 1937 1938/** 1939 * schedule_tail - first thing a freshly forked thread must call. 1940 * @prev: the thread we just switched away from. 1941 */ 1942asmlinkage void schedule_tail(struct task_struct *prev) 1943 __releases(rq->lock) 1944{ 1945 struct rq *rq = this_rq(); 1946 1947 finish_task_switch(rq, prev); 1948 1949 /* 1950 * FIXME: do we need to worry about rq being invalidated by the 1951 * task_switch? 1952 */ 1953 post_schedule(rq); 1954 1955#ifdef __ARCH_WANT_UNLOCKED_CTXSW 1956 /* In this case, finish_task_switch does not reenable preemption */ 1957 preempt_enable(); 1958#endif 1959 if (current->set_child_tid) 1960 put_user(task_pid_vnr(current), current->set_child_tid); 1961} 1962 1963/* 1964 * context_switch - switch to the new MM and the new 1965 * thread's register state. 1966 */ 1967static inline void 1968context_switch(struct rq *rq, struct task_struct *prev, 1969 struct task_struct *next) 1970{ 1971 struct mm_struct *mm, *oldmm; 1972 1973 prepare_task_switch(rq, prev, next); 1974 1975 mm = next->mm; 1976 oldmm = prev->active_mm; 1977 /* 1978 * For paravirt, this is coupled with an exit in switch_to to 1979 * combine the page table reload and the switch backend into 1980 * one hypercall. 1981 */ 1982 arch_start_context_switch(prev); 1983 1984 if (!mm) { 1985 next->active_mm = oldmm; 1986 atomic_inc(&oldmm->mm_count); 1987 enter_lazy_tlb(oldmm, next); 1988 } else 1989 switch_mm(oldmm, mm, next); 1990 1991 if (!prev->mm) { 1992 prev->active_mm = NULL; 1993 rq->prev_mm = oldmm; 1994 } 1995 /* 1996 * Since the runqueue lock will be released by the next 1997 * task (which is an invalid locking op but in the case 1998 * of the scheduler it's an obvious special-case), so we 1999 * do an early lockdep release here: 2000 */ 2001#ifndef __ARCH_WANT_UNLOCKED_CTXSW 2002 spin_release(&rq->lock.dep_map, 1, _THIS_IP_); 2003#endif 2004 2005 context_tracking_task_switch(prev, next); 2006 /* Here we just switch the register state and the stack. */ 2007 switch_to(prev, next, prev); 2008 2009 barrier(); 2010 /* 2011 * this_rq must be evaluated again because prev may have moved 2012 * CPUs since it called schedule(), thus the 'rq' on its stack 2013 * frame will be invalid. 2014 */ 2015 finish_task_switch(this_rq(), prev); 2016} 2017 2018/* 2019 * nr_running and nr_context_switches: 2020 * 2021 * externally visible scheduler statistics: current number of runnable 2022 * threads, total number of context switches performed since bootup. 2023 */ 2024unsigned long nr_running(void) 2025{ 2026 unsigned long i, sum = 0; 2027 2028 for_each_online_cpu(i) 2029 sum += cpu_rq(i)->nr_running; 2030 2031 return sum; 2032} 2033 2034unsigned long long nr_context_switches(void) 2035{ 2036 int i; 2037 unsigned long long sum = 0; 2038 2039 for_each_possible_cpu(i) 2040 sum += cpu_rq(i)->nr_switches; 2041 2042 return sum; 2043} 2044 2045unsigned long nr_iowait(void) 2046{ 2047 unsigned long i, sum = 0; 2048 2049 for_each_possible_cpu(i) 2050 sum += atomic_read(&cpu_rq(i)->nr_iowait); 2051 2052 return sum; 2053} 2054 2055unsigned long nr_iowait_cpu(int cpu) 2056{ 2057 struct rq *this = cpu_rq(cpu); 2058 return atomic_read(&this->nr_iowait); 2059} 2060 2061#ifdef CONFIG_SMP 2062 2063/* 2064 * sched_exec - execve() is a valuable balancing opportunity, because at 2065 * this point the task has the smallest effective memory and cache footprint. 2066 */ 2067void sched_exec(void) 2068{ 2069 struct task_struct *p = current; 2070 unsigned long flags; 2071 int dest_cpu; 2072 2073 raw_spin_lock_irqsave(&p->pi_lock, flags); 2074 dest_cpu = p->sched_class->select_task_rq(p, SD_BALANCE_EXEC, 0); 2075 if (dest_cpu == smp_processor_id()) 2076 goto unlock; 2077 2078 if (likely(cpu_active(dest_cpu))) { 2079 struct migration_arg arg = { p, dest_cpu }; 2080 2081 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 2082 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg); 2083 return; 2084 } 2085unlock: 2086 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 2087} 2088 2089#endif 2090 2091DEFINE_PER_CPU(struct kernel_stat, kstat); 2092DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat); 2093 2094EXPORT_PER_CPU_SYMBOL(kstat); 2095EXPORT_PER_CPU_SYMBOL(kernel_cpustat); 2096 2097/* 2098 * Return any ns on the sched_clock that have not yet been accounted in 2099 * @p in case that task is currently running. 2100 * 2101 * Called with task_rq_lock() held on @rq. 2102 */ 2103static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq) 2104{ 2105 u64 ns = 0; 2106 2107 if (task_current(rq, p)) { 2108 update_rq_clock(rq); 2109 ns = rq_clock_task(rq) - p->se.exec_start; 2110 if ((s64)ns < 0) 2111 ns = 0; 2112 } 2113 2114 return ns; 2115} 2116 2117unsigned long long task_delta_exec(struct task_struct *p) 2118{ 2119 unsigned long flags; 2120 struct rq *rq; 2121 u64 ns = 0; 2122 2123 rq = task_rq_lock(p, &flags); 2124 ns = do_task_delta_exec(p, rq); 2125 task_rq_unlock(rq, p, &flags); 2126 2127 return ns; 2128} 2129 2130/* 2131 * Return accounted runtime for the task. 2132 * In case the task is currently running, return the runtime plus current's 2133 * pending runtime that have not been accounted yet. 2134 */ 2135unsigned long long task_sched_runtime(struct task_struct *p) 2136{ 2137 unsigned long flags; 2138 struct rq *rq; 2139 u64 ns = 0; 2140 2141 rq = task_rq_lock(p, &flags); 2142 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq); 2143 task_rq_unlock(rq, p, &flags); 2144 2145 return ns; 2146} 2147 2148/* 2149 * This function gets called by the timer code, with HZ frequency. 2150 * We call it with interrupts disabled. 2151 */ 2152void scheduler_tick(void) 2153{ 2154 int cpu = smp_processor_id(); 2155 struct rq *rq = cpu_rq(cpu); 2156 struct task_struct *curr = rq->curr; 2157 2158 sched_clock_tick(); 2159 2160 raw_spin_lock(&rq->lock); 2161 update_rq_clock(rq); 2162 update_cpu_load_active(rq); 2163 curr->sched_class->task_tick(rq, curr, 0); 2164 raw_spin_unlock(&rq->lock); 2165 2166 perf_event_task_tick(); 2167 2168#ifdef CONFIG_SMP 2169 rq->idle_balance = idle_cpu(cpu); 2170 trigger_load_balance(rq, cpu); 2171#endif 2172 rq_last_tick_reset(rq); 2173} 2174 2175#ifdef CONFIG_NO_HZ_FULL 2176/** 2177 * scheduler_tick_max_deferment 2178 * 2179 * Keep at least one tick per second when a single 2180 * active task is running because the scheduler doesn't 2181 * yet completely support full dynticks environment. 2182 * 2183 * This makes sure that uptime, CFS vruntime, load 2184 * balancing, etc... continue to move forward, even 2185 * with a very low granularity. 2186 */ 2187u64 scheduler_tick_max_deferment(void) 2188{ 2189 struct rq *rq = this_rq(); 2190 unsigned long next, now = ACCESS_ONCE(jiffies); 2191 2192 next = rq->last_sched_tick + HZ; 2193 2194 if (time_before_eq(next, now)) 2195 return 0; 2196 2197 return jiffies_to_usecs(next - now) * NSEC_PER_USEC; 2198} 2199#endif 2200 2201notrace unsigned long get_parent_ip(unsigned long addr) 2202{ 2203 if (in_lock_functions(addr)) { 2204 addr = CALLER_ADDR2; 2205 if (in_lock_functions(addr)) 2206 addr = CALLER_ADDR3; 2207 } 2208 return addr; 2209} 2210 2211#if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \ 2212 defined(CONFIG_PREEMPT_TRACER)) 2213 2214void __kprobes add_preempt_count(int val) 2215{ 2216#ifdef CONFIG_DEBUG_PREEMPT 2217 /* 2218 * Underflow? 2219 */ 2220 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0))) 2221 return; 2222#endif 2223 preempt_count() += val; 2224#ifdef CONFIG_DEBUG_PREEMPT 2225 /* 2226 * Spinlock count overflowing soon? 2227 */ 2228 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >= 2229 PREEMPT_MASK - 10); 2230#endif 2231 if (preempt_count() == val) 2232 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1)); 2233} 2234EXPORT_SYMBOL(add_preempt_count); 2235 2236void __kprobes sub_preempt_count(int val) 2237{ 2238#ifdef CONFIG_DEBUG_PREEMPT 2239 /* 2240 * Underflow? 2241 */ 2242 if (DEBUG_LOCKS_WARN_ON(val > preempt_count())) 2243 return; 2244 /* 2245 * Is the spinlock portion underflowing? 2246 */ 2247 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) && 2248 !(preempt_count() & PREEMPT_MASK))) 2249 return; 2250#endif 2251 2252 if (preempt_count() == val) 2253 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1)); 2254 preempt_count() -= val; 2255} 2256EXPORT_SYMBOL(sub_preempt_count); 2257 2258#endif 2259 2260/* 2261 * Print scheduling while atomic bug: 2262 */ 2263static noinline void __schedule_bug(struct task_struct *prev) 2264{ 2265 if (oops_in_progress) 2266 return; 2267 2268 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n", 2269 prev->comm, prev->pid, preempt_count()); 2270 2271 debug_show_held_locks(prev); 2272 print_modules(); 2273 if (irqs_disabled()) 2274 print_irqtrace_events(prev); 2275 dump_stack(); 2276 add_taint(TAINT_WARN, LOCKDEP_STILL_OK); 2277} 2278 2279/* 2280 * Various schedule()-time debugging checks and statistics: 2281 */ 2282static inline void schedule_debug(struct task_struct *prev) 2283{ 2284 /* 2285 * Test if we are atomic. Since do_exit() needs to call into 2286 * schedule() atomically, we ignore that path for now. 2287 * Otherwise, whine if we are scheduling when we should not be. 2288 */ 2289 if (unlikely(in_atomic_preempt_off() && !prev->exit_state)) 2290 __schedule_bug(prev); 2291 rcu_sleep_check(); 2292 2293 profile_hit(SCHED_PROFILING, __builtin_return_address(0)); 2294 2295 schedstat_inc(this_rq(), sched_count); 2296} 2297 2298static void put_prev_task(struct rq *rq, struct task_struct *prev) 2299{ 2300 if (prev->on_rq || rq->skip_clock_update < 0) 2301 update_rq_clock(rq); 2302 prev->sched_class->put_prev_task(rq, prev); 2303} 2304 2305/* 2306 * Pick up the highest-prio task: 2307 */ 2308static inline struct task_struct * 2309pick_next_task(struct rq *rq) 2310{ 2311 const struct sched_class *class; 2312 struct task_struct *p; 2313 2314 /* 2315 * Optimization: we know that if all tasks are in 2316 * the fair class we can call that function directly: 2317 */ 2318 if (likely(rq->nr_running == rq->cfs.h_nr_running)) { 2319 p = fair_sched_class.pick_next_task(rq); 2320 if (likely(p)) 2321 return p; 2322 } 2323 2324 for_each_class(class) { 2325 p = class->pick_next_task(rq); 2326 if (p) 2327 return p; 2328 } 2329 2330 BUG(); /* the idle class will always have a runnable task */ 2331} 2332 2333/* 2334 * __schedule() is the main scheduler function. 2335 * 2336 * The main means of driving the scheduler and thus entering this function are: 2337 * 2338 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc. 2339 * 2340 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return 2341 * paths. For example, see arch/x86/entry_64.S. 2342 * 2343 * To drive preemption between tasks, the scheduler sets the flag in timer 2344 * interrupt handler scheduler_tick(). 2345 * 2346 * 3. Wakeups don't really cause entry into schedule(). They add a 2347 * task to the run-queue and that's it. 2348 * 2349 * Now, if the new task added to the run-queue preempts the current 2350 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets 2351 * called on the nearest possible occasion: 2352 * 2353 * - If the kernel is preemptible (CONFIG_PREEMPT=y): 2354 * 2355 * - in syscall or exception context, at the next outmost 2356 * preempt_enable(). (this might be as soon as the wake_up()'s 2357 * spin_unlock()!) 2358 * 2359 * - in IRQ context, return from interrupt-handler to 2360 * preemptible context 2361 * 2362 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set) 2363 * then at the next: 2364 * 2365 * - cond_resched() call 2366 * - explicit schedule() call 2367 * - return from syscall or exception to user-space 2368 * - return from interrupt-handler to user-space 2369 */ 2370static void __sched __schedule(void) 2371{ 2372 struct task_struct *prev, *next; 2373 unsigned long *switch_count; 2374 struct rq *rq; 2375 int cpu; 2376 2377need_resched: 2378 preempt_disable(); 2379 cpu = smp_processor_id(); 2380 rq = cpu_rq(cpu); 2381 rcu_note_context_switch(cpu); 2382 prev = rq->curr; 2383 2384 schedule_debug(prev); 2385 2386 if (sched_feat(HRTICK)) 2387 hrtick_clear(rq); 2388 2389 raw_spin_lock_irq(&rq->lock); 2390 2391 switch_count = &prev->nivcsw; 2392 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) { 2393 if (unlikely(signal_pending_state(prev->state, prev))) { 2394 prev->state = TASK_RUNNING; 2395 } else { 2396 deactivate_task(rq, prev, DEQUEUE_SLEEP); 2397 prev->on_rq = 0; 2398 2399 /* 2400 * If a worker went to sleep, notify and ask workqueue 2401 * whether it wants to wake up a task to maintain 2402 * concurrency. 2403 */ 2404 if (prev->flags & PF_WQ_WORKER) { 2405 struct task_struct *to_wakeup; 2406 2407 to_wakeup = wq_worker_sleeping(prev, cpu); 2408 if (to_wakeup) 2409 try_to_wake_up_local(to_wakeup); 2410 } 2411 } 2412 switch_count = &prev->nvcsw; 2413 } 2414 2415 pre_schedule(rq, prev); 2416 2417 if (unlikely(!rq->nr_running)) 2418 idle_balance(cpu, rq); 2419 2420 put_prev_task(rq, prev); 2421 next = pick_next_task(rq); 2422 clear_tsk_need_resched(prev); 2423 rq->skip_clock_update = 0; 2424 2425 if (likely(prev != next)) { 2426 rq->nr_switches++; 2427 rq->curr = next; 2428 ++*switch_count; 2429 2430 context_switch(rq, prev, next); /* unlocks the rq */ 2431 /* 2432 * The context switch have flipped the stack from under us 2433 * and restored the local variables which were saved when 2434 * this task called schedule() in the past. prev == current 2435 * is still correct, but it can be moved to another cpu/rq. 2436 */ 2437 cpu = smp_processor_id(); 2438 rq = cpu_rq(cpu); 2439 } else 2440 raw_spin_unlock_irq(&rq->lock); 2441 2442 post_schedule(rq); 2443 2444 sched_preempt_enable_no_resched(); 2445 if (need_resched()) 2446 goto need_resched; 2447} 2448 2449static inline void sched_submit_work(struct task_struct *tsk) 2450{ 2451 if (!tsk->state || tsk_is_pi_blocked(tsk)) 2452 return; 2453 /* 2454 * If we are going to sleep and we have plugged IO queued, 2455 * make sure to submit it to avoid deadlocks. 2456 */ 2457 if (blk_needs_flush_plug(tsk)) 2458 blk_schedule_flush_plug(tsk); 2459} 2460 2461asmlinkage void __sched schedule(void) 2462{ 2463 struct task_struct *tsk = current; 2464 2465 sched_submit_work(tsk); 2466 __schedule(); 2467} 2468EXPORT_SYMBOL(schedule); 2469 2470#ifdef CONFIG_CONTEXT_TRACKING 2471asmlinkage void __sched schedule_user(void) 2472{ 2473 /* 2474 * If we come here after a random call to set_need_resched(), 2475 * or we have been woken up remotely but the IPI has not yet arrived, 2476 * we haven't yet exited the RCU idle mode. Do it here manually until 2477 * we find a better solution. 2478 */ 2479 user_exit(); 2480 schedule(); 2481 user_enter(); 2482} 2483#endif 2484 2485/** 2486 * schedule_preempt_disabled - called with preemption disabled 2487 * 2488 * Returns with preemption disabled. Note: preempt_count must be 1 2489 */ 2490void __sched schedule_preempt_disabled(void) 2491{ 2492 sched_preempt_enable_no_resched(); 2493 schedule(); 2494 preempt_disable(); 2495} 2496 2497#ifdef CONFIG_PREEMPT 2498/* 2499 * this is the entry point to schedule() from in-kernel preemption 2500 * off of preempt_enable. Kernel preemptions off return from interrupt 2501 * occur there and call schedule directly. 2502 */ 2503asmlinkage void __sched notrace preempt_schedule(void) 2504{ 2505 struct thread_info *ti = current_thread_info(); 2506 2507 /* 2508 * If there is a non-zero preempt_count or interrupts are disabled, 2509 * we do not want to preempt the current task. Just return.. 2510 */ 2511 if (likely(ti->preempt_count || irqs_disabled())) 2512 return; 2513 2514 do { 2515 add_preempt_count_notrace(PREEMPT_ACTIVE); 2516 __schedule(); 2517 sub_preempt_count_notrace(PREEMPT_ACTIVE); 2518 2519 /* 2520 * Check again in case we missed a preemption opportunity 2521 * between schedule and now. 2522 */ 2523 barrier(); 2524 } while (need_resched()); 2525} 2526EXPORT_SYMBOL(preempt_schedule); 2527 2528/* 2529 * this is the entry point to schedule() from kernel preemption 2530 * off of irq context. 2531 * Note, that this is called and return with irqs disabled. This will 2532 * protect us against recursive calling from irq. 2533 */ 2534asmlinkage void __sched preempt_schedule_irq(void) 2535{ 2536 struct thread_info *ti = current_thread_info(); 2537 enum ctx_state prev_state; 2538 2539 /* Catch callers which need to be fixed */ 2540 BUG_ON(ti->preempt_count || !irqs_disabled()); 2541 2542 prev_state = exception_enter(); 2543 2544 do { 2545 add_preempt_count(PREEMPT_ACTIVE); 2546 local_irq_enable(); 2547 __schedule(); 2548 local_irq_disable(); 2549 sub_preempt_count(PREEMPT_ACTIVE); 2550 2551 /* 2552 * Check again in case we missed a preemption opportunity 2553 * between schedule and now. 2554 */ 2555 barrier(); 2556 } while (need_resched()); 2557 2558 exception_exit(prev_state); 2559} 2560 2561#endif /* CONFIG_PREEMPT */ 2562 2563int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags, 2564 void *key) 2565{ 2566 return try_to_wake_up(curr->private, mode, wake_flags); 2567} 2568EXPORT_SYMBOL(default_wake_function); 2569 2570/* 2571 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just 2572 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve 2573 * number) then we wake all the non-exclusive tasks and one exclusive task. 2574 * 2575 * There are circumstances in which we can try to wake a task which has already 2576 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns 2577 * zero in this (rare) case, and we handle it by continuing to scan the queue. 2578 */ 2579static void __wake_up_common(wait_queue_head_t *q, unsigned int mode, 2580 int nr_exclusive, int wake_flags, void *key) 2581{ 2582 wait_queue_t *curr, *next; 2583 2584 list_for_each_entry_safe(curr, next, &q->task_list, task_list) { 2585 unsigned flags = curr->flags; 2586 2587 if (curr->func(curr, mode, wake_flags, key) && 2588 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive) 2589 break; 2590 } 2591} 2592 2593/** 2594 * __wake_up - wake up threads blocked on a waitqueue. 2595 * @q: the waitqueue 2596 * @mode: which threads 2597 * @nr_exclusive: how many wake-one or wake-many threads to wake up 2598 * @key: is directly passed to the wakeup function 2599 * 2600 * It may be assumed that this function implies a write memory barrier before 2601 * changing the task state if and only if any tasks are woken up. 2602 */ 2603void __wake_up(wait_queue_head_t *q, unsigned int mode, 2604 int nr_exclusive, void *key) 2605{ 2606 unsigned long flags; 2607 2608 spin_lock_irqsave(&q->lock, flags); 2609 __wake_up_common(q, mode, nr_exclusive, 0, key); 2610 spin_unlock_irqrestore(&q->lock, flags); 2611} 2612EXPORT_SYMBOL(__wake_up); 2613 2614/* 2615 * Same as __wake_up but called with the spinlock in wait_queue_head_t held. 2616 */ 2617void __wake_up_locked(wait_queue_head_t *q, unsigned int mode, int nr) 2618{ 2619 __wake_up_common(q, mode, nr, 0, NULL); 2620} 2621EXPORT_SYMBOL_GPL(__wake_up_locked); 2622 2623void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key) 2624{ 2625 __wake_up_common(q, mode, 1, 0, key); 2626} 2627EXPORT_SYMBOL_GPL(__wake_up_locked_key); 2628 2629/** 2630 * __wake_up_sync_key - wake up threads blocked on a waitqueue. 2631 * @q: the waitqueue 2632 * @mode: which threads 2633 * @nr_exclusive: how many wake-one or wake-many threads to wake up 2634 * @key: opaque value to be passed to wakeup targets 2635 * 2636 * The sync wakeup differs that the waker knows that it will schedule 2637 * away soon, so while the target thread will be woken up, it will not 2638 * be migrated to another CPU - ie. the two threads are 'synchronized' 2639 * with each other. This can prevent needless bouncing between CPUs. 2640 * 2641 * On UP it can prevent extra preemption. 2642 * 2643 * It may be assumed that this function implies a write memory barrier before 2644 * changing the task state if and only if any tasks are woken up. 2645 */ 2646void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode, 2647 int nr_exclusive, void *key) 2648{ 2649 unsigned long flags; 2650 int wake_flags = WF_SYNC; 2651 2652 if (unlikely(!q)) 2653 return; 2654 2655 if (unlikely(!nr_exclusive)) 2656 wake_flags = 0; 2657 2658 spin_lock_irqsave(&q->lock, flags); 2659 __wake_up_common(q, mode, nr_exclusive, wake_flags, key); 2660 spin_unlock_irqrestore(&q->lock, flags); 2661} 2662EXPORT_SYMBOL_GPL(__wake_up_sync_key); 2663 2664/* 2665 * __wake_up_sync - see __wake_up_sync_key() 2666 */ 2667void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive) 2668{ 2669 __wake_up_sync_key(q, mode, nr_exclusive, NULL); 2670} 2671EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */ 2672 2673/** 2674 * complete: - signals a single thread waiting on this completion 2675 * @x: holds the state of this particular completion 2676 * 2677 * This will wake up a single thread waiting on this completion. Threads will be 2678 * awakened in the same order in which they were queued. 2679 * 2680 * See also complete_all(), wait_for_completion() and related routines. 2681 * 2682 * It may be assumed that this function implies a write memory barrier before 2683 * changing the task state if and only if any tasks are woken up. 2684 */ 2685void complete(struct completion *x) 2686{ 2687 unsigned long flags; 2688 2689 spin_lock_irqsave(&x->wait.lock, flags); 2690 x->done++; 2691 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL); 2692 spin_unlock_irqrestore(&x->wait.lock, flags); 2693} 2694EXPORT_SYMBOL(complete); 2695 2696/** 2697 * complete_all: - signals all threads waiting on this completion 2698 * @x: holds the state of this particular completion 2699 * 2700 * This will wake up all threads waiting on this particular completion event. 2701 * 2702 * It may be assumed that this function implies a write memory barrier before 2703 * changing the task state if and only if any tasks are woken up. 2704 */ 2705void complete_all(struct completion *x) 2706{ 2707 unsigned long flags; 2708 2709 spin_lock_irqsave(&x->wait.lock, flags); 2710 x->done += UINT_MAX/2; 2711 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL); 2712 spin_unlock_irqrestore(&x->wait.lock, flags); 2713} 2714EXPORT_SYMBOL(complete_all); 2715 2716static inline long __sched 2717do_wait_for_common(struct completion *x, 2718 long (*action)(long), long timeout, int state) 2719{ 2720 if (!x->done) { 2721 DECLARE_WAITQUEUE(wait, current); 2722 2723 __add_wait_queue_tail_exclusive(&x->wait, &wait); 2724 do { 2725 if (signal_pending_state(state, current)) { 2726 timeout = -ERESTARTSYS; 2727 break; 2728 } 2729 __set_current_state(state); 2730 spin_unlock_irq(&x->wait.lock); 2731 timeout = action(timeout); 2732 spin_lock_irq(&x->wait.lock); 2733 } while (!x->done && timeout); 2734 __remove_wait_queue(&x->wait, &wait); 2735 if (!x->done) 2736 return timeout; 2737 } 2738 x->done--; 2739 return timeout ?: 1; 2740} 2741 2742static inline long __sched 2743__wait_for_common(struct completion *x, 2744 long (*action)(long), long timeout, int state) 2745{ 2746 might_sleep(); 2747 2748 spin_lock_irq(&x->wait.lock); 2749 timeout = do_wait_for_common(x, action, timeout, state); 2750 spin_unlock_irq(&x->wait.lock); 2751 return timeout; 2752} 2753 2754static long __sched 2755wait_for_common(struct completion *x, long timeout, int state) 2756{ 2757 return __wait_for_common(x, schedule_timeout, timeout, state); 2758} 2759 2760static long __sched 2761wait_for_common_io(struct completion *x, long timeout, int state) 2762{ 2763 return __wait_for_common(x, io_schedule_timeout, timeout, state); 2764} 2765 2766/** 2767 * wait_for_completion: - waits for completion of a task 2768 * @x: holds the state of this particular completion 2769 * 2770 * This waits to be signaled for completion of a specific task. It is NOT 2771 * interruptible and there is no timeout. 2772 * 2773 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout 2774 * and interrupt capability. Also see complete(). 2775 */ 2776void __sched wait_for_completion(struct completion *x) 2777{ 2778 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE); 2779} 2780EXPORT_SYMBOL(wait_for_completion); 2781 2782/** 2783 * wait_for_completion_timeout: - waits for completion of a task (w/timeout) 2784 * @x: holds the state of this particular completion 2785 * @timeout: timeout value in jiffies 2786 * 2787 * This waits for either a completion of a specific task to be signaled or for a 2788 * specified timeout to expire. The timeout is in jiffies. It is not 2789 * interruptible. 2790 * 2791 * The return value is 0 if timed out, and positive (at least 1, or number of 2792 * jiffies left till timeout) if completed. 2793 */ 2794unsigned long __sched 2795wait_for_completion_timeout(struct completion *x, unsigned long timeout) 2796{ 2797 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE); 2798} 2799EXPORT_SYMBOL(wait_for_completion_timeout); 2800 2801/** 2802 * wait_for_completion_io: - waits for completion of a task 2803 * @x: holds the state of this particular completion 2804 * 2805 * This waits to be signaled for completion of a specific task. It is NOT 2806 * interruptible and there is no timeout. The caller is accounted as waiting 2807 * for IO. 2808 */ 2809void __sched wait_for_completion_io(struct completion *x) 2810{ 2811 wait_for_common_io(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE); 2812} 2813EXPORT_SYMBOL(wait_for_completion_io); 2814 2815/** 2816 * wait_for_completion_io_timeout: - waits for completion of a task (w/timeout) 2817 * @x: holds the state of this particular completion 2818 * @timeout: timeout value in jiffies 2819 * 2820 * This waits for either a completion of a specific task to be signaled or for a 2821 * specified timeout to expire. The timeout is in jiffies. It is not 2822 * interruptible. The caller is accounted as waiting for IO. 2823 * 2824 * The return value is 0 if timed out, and positive (at least 1, or number of 2825 * jiffies left till timeout) if completed. 2826 */ 2827unsigned long __sched 2828wait_for_completion_io_timeout(struct completion *x, unsigned long timeout) 2829{ 2830 return wait_for_common_io(x, timeout, TASK_UNINTERRUPTIBLE); 2831} 2832EXPORT_SYMBOL(wait_for_completion_io_timeout); 2833 2834/** 2835 * wait_for_completion_interruptible: - waits for completion of a task (w/intr) 2836 * @x: holds the state of this particular completion 2837 * 2838 * This waits for completion of a specific task to be signaled. It is 2839 * interruptible. 2840 * 2841 * The return value is -ERESTARTSYS if interrupted, 0 if completed. 2842 */ 2843int __sched wait_for_completion_interruptible(struct completion *x) 2844{ 2845 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE); 2846 if (t == -ERESTARTSYS) 2847 return t; 2848 return 0; 2849} 2850EXPORT_SYMBOL(wait_for_completion_interruptible); 2851 2852/** 2853 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr)) 2854 * @x: holds the state of this particular completion 2855 * @timeout: timeout value in jiffies 2856 * 2857 * This waits for either a completion of a specific task to be signaled or for a 2858 * specified timeout to expire. It is interruptible. The timeout is in jiffies. 2859 * 2860 * The return value is -ERESTARTSYS if interrupted, 0 if timed out, 2861 * positive (at least 1, or number of jiffies left till timeout) if completed. 2862 */ 2863long __sched 2864wait_for_completion_interruptible_timeout(struct completion *x, 2865 unsigned long timeout) 2866{ 2867 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE); 2868} 2869EXPORT_SYMBOL(wait_for_completion_interruptible_timeout); 2870 2871/** 2872 * wait_for_completion_killable: - waits for completion of a task (killable) 2873 * @x: holds the state of this particular completion 2874 * 2875 * This waits to be signaled for completion of a specific task. It can be 2876 * interrupted by a kill signal. 2877 * 2878 * The return value is -ERESTARTSYS if interrupted, 0 if completed. 2879 */ 2880int __sched wait_for_completion_killable(struct completion *x) 2881{ 2882 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE); 2883 if (t == -ERESTARTSYS) 2884 return t; 2885 return 0; 2886} 2887EXPORT_SYMBOL(wait_for_completion_killable); 2888 2889/** 2890 * wait_for_completion_killable_timeout: - waits for completion of a task (w/(to,killable)) 2891 * @x: holds the state of this particular completion 2892 * @timeout: timeout value in jiffies 2893 * 2894 * This waits for either a completion of a specific task to be 2895 * signaled or for a specified timeout to expire. It can be 2896 * interrupted by a kill signal. The timeout is in jiffies. 2897 * 2898 * The return value is -ERESTARTSYS if interrupted, 0 if timed out, 2899 * positive (at least 1, or number of jiffies left till timeout) if completed. 2900 */ 2901long __sched 2902wait_for_completion_killable_timeout(struct completion *x, 2903 unsigned long timeout) 2904{ 2905 return wait_for_common(x, timeout, TASK_KILLABLE); 2906} 2907EXPORT_SYMBOL(wait_for_completion_killable_timeout); 2908 2909/** 2910 * try_wait_for_completion - try to decrement a completion without blocking 2911 * @x: completion structure 2912 * 2913 * Returns: 0 if a decrement cannot be done without blocking 2914 * 1 if a decrement succeeded. 2915 * 2916 * If a completion is being used as a counting completion, 2917 * attempt to decrement the counter without blocking. This 2918 * enables us to avoid waiting if the resource the completion 2919 * is protecting is not available. 2920 */ 2921bool try_wait_for_completion(struct completion *x) 2922{ 2923 unsigned long flags; 2924 int ret = 1; 2925 2926 spin_lock_irqsave(&x->wait.lock, flags); 2927 if (!x->done) 2928 ret = 0; 2929 else 2930 x->done--; 2931 spin_unlock_irqrestore(&x->wait.lock, flags); 2932 return ret; 2933} 2934EXPORT_SYMBOL(try_wait_for_completion); 2935 2936/** 2937 * completion_done - Test to see if a completion has any waiters 2938 * @x: completion structure 2939 * 2940 * Returns: 0 if there are waiters (wait_for_completion() in progress) 2941 * 1 if there are no waiters. 2942 * 2943 */ 2944bool completion_done(struct completion *x) 2945{ 2946 unsigned long flags; 2947 int ret = 1; 2948 2949 spin_lock_irqsave(&x->wait.lock, flags); 2950 if (!x->done) 2951 ret = 0; 2952 spin_unlock_irqrestore(&x->wait.lock, flags); 2953 return ret; 2954} 2955EXPORT_SYMBOL(completion_done); 2956 2957static long __sched 2958sleep_on_common(wait_queue_head_t *q, int state, long timeout) 2959{ 2960 unsigned long flags; 2961 wait_queue_t wait; 2962 2963 init_waitqueue_entry(&wait, current); 2964 2965 __set_current_state(state); 2966 2967 spin_lock_irqsave(&q->lock, flags); 2968 __add_wait_queue(q, &wait); 2969 spin_unlock(&q->lock); 2970 timeout = schedule_timeout(timeout); 2971 spin_lock_irq(&q->lock); 2972 __remove_wait_queue(q, &wait); 2973 spin_unlock_irqrestore(&q->lock, flags); 2974 2975 return timeout; 2976} 2977 2978void __sched interruptible_sleep_on(wait_queue_head_t *q) 2979{ 2980 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT); 2981} 2982EXPORT_SYMBOL(interruptible_sleep_on); 2983 2984long __sched 2985interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout) 2986{ 2987 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout); 2988} 2989EXPORT_SYMBOL(interruptible_sleep_on_timeout); 2990 2991void __sched sleep_on(wait_queue_head_t *q) 2992{ 2993 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT); 2994} 2995EXPORT_SYMBOL(sleep_on); 2996 2997long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout) 2998{ 2999 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout); 3000} 3001EXPORT_SYMBOL(sleep_on_timeout); 3002 3003#ifdef CONFIG_RT_MUTEXES 3004 3005/* 3006 * rt_mutex_setprio - set the current priority of a task 3007 * @p: task 3008 * @prio: prio value (kernel-internal form) 3009 * 3010 * This function changes the 'effective' priority of a task. It does 3011 * not touch ->normal_prio like __setscheduler(). 3012 * 3013 * Used by the rt_mutex code to implement priority inheritance logic. 3014 */ 3015void rt_mutex_setprio(struct task_struct *p, int prio) 3016{ 3017 int oldprio, on_rq, running; 3018 struct rq *rq; 3019 const struct sched_class *prev_class; 3020 3021 BUG_ON(prio < 0 || prio > MAX_PRIO); 3022 3023 rq = __task_rq_lock(p); 3024 3025 /* 3026 * Idle task boosting is a nono in general. There is one 3027 * exception, when PREEMPT_RT and NOHZ is active: 3028 * 3029 * The idle task calls get_next_timer_interrupt() and holds 3030 * the timer wheel base->lock on the CPU and another CPU wants 3031 * to access the timer (probably to cancel it). We can safely 3032 * ignore the boosting request, as the idle CPU runs this code 3033 * with interrupts disabled and will complete the lock 3034 * protected section without being interrupted. So there is no 3035 * real need to boost. 3036 */ 3037 if (unlikely(p == rq->idle)) { 3038 WARN_ON(p != rq->curr); 3039 WARN_ON(p->pi_blocked_on); 3040 goto out_unlock; 3041 } 3042 3043 trace_sched_pi_setprio(p, prio); 3044 oldprio = p->prio; 3045 prev_class = p->sched_class; 3046 on_rq = p->on_rq; 3047 running = task_current(rq, p); 3048 if (on_rq) 3049 dequeue_task(rq, p, 0); 3050 if (running) 3051 p->sched_class->put_prev_task(rq, p); 3052 3053 if (rt_prio(prio)) 3054 p->sched_class = &rt_sched_class; 3055 else 3056 p->sched_class = &fair_sched_class; 3057 3058 p->prio = prio; 3059 3060 if (running) 3061 p->sched_class->set_curr_task(rq); 3062 if (on_rq) 3063 enqueue_task(rq, p, oldprio < prio ? ENQUEUE_HEAD : 0); 3064 3065 check_class_changed(rq, p, prev_class, oldprio); 3066out_unlock: 3067 __task_rq_unlock(rq); 3068} 3069#endif 3070void set_user_nice(struct task_struct *p, long nice) 3071{ 3072 int old_prio, delta, on_rq; 3073 unsigned long flags; 3074 struct rq *rq; 3075 3076 if (TASK_NICE(p) == nice || nice < -20 || nice > 19) 3077 return; 3078 /* 3079 * We have to be careful, if called from sys_setpriority(), 3080 * the task might be in the middle of scheduling on another CPU. 3081 */ 3082 rq = task_rq_lock(p, &flags); 3083 /* 3084 * The RT priorities are set via sched_setscheduler(), but we still 3085 * allow the 'normal' nice value to be set - but as expected 3086 * it wont have any effect on scheduling until the task is 3087 * SCHED_FIFO/SCHED_RR: 3088 */ 3089 if (task_has_rt_policy(p)) { 3090 p->static_prio = NICE_TO_PRIO(nice); 3091 goto out_unlock; 3092 } 3093 on_rq = p->on_rq; 3094 if (on_rq) 3095 dequeue_task(rq, p, 0); 3096 3097 p->static_prio = NICE_TO_PRIO(nice); 3098 set_load_weight(p); 3099 old_prio = p->prio; 3100 p->prio = effective_prio(p); 3101 delta = p->prio - old_prio; 3102 3103 if (on_rq) { 3104 enqueue_task(rq, p, 0); 3105 /* 3106 * If the task increased its priority or is running and 3107 * lowered its priority, then reschedule its CPU: 3108 */ 3109 if (delta < 0 || (delta > 0 && task_running(rq, p))) 3110 resched_task(rq->curr); 3111 } 3112out_unlock: 3113 task_rq_unlock(rq, p, &flags); 3114} 3115EXPORT_SYMBOL(set_user_nice); 3116 3117/* 3118 * can_nice - check if a task can reduce its nice value 3119 * @p: task 3120 * @nice: nice value 3121 */ 3122int can_nice(const struct task_struct *p, const int nice) 3123{ 3124 /* convert nice value [19,-20] to rlimit style value [1,40] */ 3125 int nice_rlim = 20 - nice; 3126 3127 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) || 3128 capable(CAP_SYS_NICE)); 3129} 3130 3131#ifdef __ARCH_WANT_SYS_NICE 3132 3133/* 3134 * sys_nice - change the priority of the current process. 3135 * @increment: priority increment 3136 * 3137 * sys_setpriority is a more generic, but much slower function that 3138 * does similar things. 3139 */ 3140SYSCALL_DEFINE1(nice, int, increment) 3141{ 3142 long nice, retval; 3143 3144 /* 3145 * Setpriority might change our priority at the same moment. 3146 * We don't have to worry. Conceptually one call occurs first 3147 * and we have a single winner. 3148 */ 3149 if (increment < -40) 3150 increment = -40; 3151 if (increment > 40) 3152 increment = 40; 3153 3154 nice = TASK_NICE(current) + increment; 3155 if (nice < -20) 3156 nice = -20; 3157 if (nice > 19) 3158 nice = 19; 3159 3160 if (increment < 0 && !can_nice(current, nice)) 3161 return -EPERM; 3162 3163 retval = security_task_setnice(current, nice); 3164 if (retval) 3165 return retval; 3166 3167 set_user_nice(current, nice); 3168 return 0; 3169} 3170 3171#endif 3172 3173/** 3174 * task_prio - return the priority value of a given task. 3175 * @p: the task in question. 3176 * 3177 * This is the priority value as seen by users in /proc. 3178 * RT tasks are offset by -200. Normal tasks are centered 3179 * around 0, value goes from -16 to +15. 3180 */ 3181int task_prio(const struct task_struct *p) 3182{ 3183 return p->prio - MAX_RT_PRIO; 3184} 3185 3186/** 3187 * task_nice - return the nice value of a given task. 3188 * @p: the task in question. 3189 */ 3190int task_nice(const struct task_struct *p) 3191{ 3192 return TASK_NICE(p); 3193} 3194EXPORT_SYMBOL(task_nice); 3195 3196/** 3197 * idle_cpu - is a given cpu idle currently? 3198 * @cpu: the processor in question. 3199 */ 3200int idle_cpu(int cpu) 3201{ 3202 struct rq *rq = cpu_rq(cpu); 3203 3204 if (rq->curr != rq->idle) 3205 return 0; 3206 3207 if (rq->nr_running) 3208 return 0; 3209 3210#ifdef CONFIG_SMP 3211 if (!llist_empty(&rq->wake_list)) 3212 return 0; 3213#endif 3214 3215 return 1; 3216} 3217 3218/** 3219 * idle_task - return the idle task for a given cpu. 3220 * @cpu: the processor in question. 3221 */ 3222struct task_struct *idle_task(int cpu) 3223{ 3224 return cpu_rq(cpu)->idle; 3225} 3226 3227/** 3228 * find_process_by_pid - find a process with a matching PID value. 3229 * @pid: the pid in question. 3230 */ 3231static struct task_struct *find_process_by_pid(pid_t pid) 3232{ 3233 return pid ? find_task_by_vpid(pid) : current; 3234} 3235 3236/* Actually do priority change: must hold rq lock. */ 3237static void 3238__setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio) 3239{ 3240 p->policy = policy; 3241 p->rt_priority = prio; 3242 p->normal_prio = normal_prio(p); 3243 /* we are holding p->pi_lock already */ 3244 p->prio = rt_mutex_getprio(p); 3245 if (rt_prio(p->prio)) 3246 p->sched_class = &rt_sched_class; 3247 else 3248 p->sched_class = &fair_sched_class; 3249 set_load_weight(p); 3250} 3251 3252/* 3253 * check the target process has a UID that matches the current process's 3254 */ 3255static bool check_same_owner(struct task_struct *p) 3256{ 3257 const struct cred *cred = current_cred(), *pcred; 3258 bool match; 3259 3260 rcu_read_lock(); 3261 pcred = __task_cred(p); 3262 match = (uid_eq(cred->euid, pcred->euid) || 3263 uid_eq(cred->euid, pcred->uid)); 3264 rcu_read_unlock(); 3265 return match; 3266} 3267 3268static int __sched_setscheduler(struct task_struct *p, int policy, 3269 const struct sched_param *param, bool user) 3270{ 3271 int retval, oldprio, oldpolicy = -1, on_rq, running; 3272 unsigned long flags; 3273 const struct sched_class *prev_class; 3274 struct rq *rq; 3275 int reset_on_fork; 3276 3277 /* may grab non-irq protected spin_locks */ 3278 BUG_ON(in_interrupt()); 3279recheck: 3280 /* double check policy once rq lock held */ 3281 if (policy < 0) { 3282 reset_on_fork = p->sched_reset_on_fork; 3283 policy = oldpolicy = p->policy; 3284 } else { 3285 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK); 3286 policy &= ~SCHED_RESET_ON_FORK; 3287 3288 if (policy != SCHED_FIFO && policy != SCHED_RR && 3289 policy != SCHED_NORMAL && policy != SCHED_BATCH && 3290 policy != SCHED_IDLE) 3291 return -EINVAL; 3292 } 3293 3294 /* 3295 * Valid priorities for SCHED_FIFO and SCHED_RR are 3296 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL, 3297 * SCHED_BATCH and SCHED_IDLE is 0. 3298 */ 3299 if (param->sched_priority < 0 || 3300 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) || 3301 (!p->mm && param->sched_priority > MAX_RT_PRIO-1)) 3302 return -EINVAL; 3303 if (rt_policy(policy) != (param->sched_priority != 0)) 3304 return -EINVAL; 3305 3306 /* 3307 * Allow unprivileged RT tasks to decrease priority: 3308 */ 3309 if (user && !capable(CAP_SYS_NICE)) { 3310 if (rt_policy(policy)) { 3311 unsigned long rlim_rtprio = 3312 task_rlimit(p, RLIMIT_RTPRIO); 3313 3314 /* can't set/change the rt policy */ 3315 if (policy != p->policy && !rlim_rtprio) 3316 return -EPERM; 3317 3318 /* can't increase priority */ 3319 if (param->sched_priority > p->rt_priority && 3320 param->sched_priority > rlim_rtprio) 3321 return -EPERM; 3322 } 3323 3324 /* 3325 * Treat SCHED_IDLE as nice 20. Only allow a switch to 3326 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it. 3327 */ 3328 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) { 3329 if (!can_nice(p, TASK_NICE(p))) 3330 return -EPERM; 3331 } 3332 3333 /* can't change other user's priorities */ 3334 if (!check_same_owner(p)) 3335 return -EPERM; 3336 3337 /* Normal users shall not reset the sched_reset_on_fork flag */ 3338 if (p->sched_reset_on_fork && !reset_on_fork) 3339 return -EPERM; 3340 } 3341 3342 if (user) { 3343 retval = security_task_setscheduler(p); 3344 if (retval) 3345 return retval; 3346 } 3347 3348 /* 3349 * make sure no PI-waiters arrive (or leave) while we are 3350 * changing the priority of the task: 3351 * 3352 * To be able to change p->policy safely, the appropriate 3353 * runqueue lock must be held. 3354 */ 3355 rq = task_rq_lock(p, &flags); 3356 3357 /* 3358 * Changing the policy of the stop threads its a very bad idea 3359 */ 3360 if (p == rq->stop) { 3361 task_rq_unlock(rq, p, &flags); 3362 return -EINVAL; 3363 } 3364 3365 /* 3366 * If not changing anything there's no need to proceed further: 3367 */ 3368 if (unlikely(policy == p->policy && (!rt_policy(policy) || 3369 param->sched_priority == p->rt_priority))) { 3370 task_rq_unlock(rq, p, &flags); 3371 return 0; 3372 } 3373 3374#ifdef CONFIG_RT_GROUP_SCHED 3375 if (user) { 3376 /* 3377 * Do not allow realtime tasks into groups that have no runtime 3378 * assigned. 3379 */ 3380 if (rt_bandwidth_enabled() && rt_policy(policy) && 3381 task_group(p)->rt_bandwidth.rt_runtime == 0 && 3382 !task_group_is_autogroup(task_group(p))) { 3383 task_rq_unlock(rq, p, &flags); 3384 return -EPERM; 3385 } 3386 } 3387#endif 3388 3389 /* recheck policy now with rq lock held */ 3390 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) { 3391 policy = oldpolicy = -1; 3392 task_rq_unlock(rq, p, &flags); 3393 goto recheck; 3394 } 3395 on_rq = p->on_rq; 3396 running = task_current(rq, p); 3397 if (on_rq) 3398 dequeue_task(rq, p, 0); 3399 if (running) 3400 p->sched_class->put_prev_task(rq, p); 3401 3402 p->sched_reset_on_fork = reset_on_fork; 3403 3404 oldprio = p->prio; 3405 prev_class = p->sched_class; 3406 __setscheduler(rq, p, policy, param->sched_priority); 3407 3408 if (running) 3409 p->sched_class->set_curr_task(rq); 3410 if (on_rq) 3411 enqueue_task(rq, p, 0); 3412 3413 check_class_changed(rq, p, prev_class, oldprio); 3414 task_rq_unlock(rq, p, &flags); 3415 3416 rt_mutex_adjust_pi(p); 3417 3418 return 0; 3419} 3420 3421/** 3422 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread. 3423 * @p: the task in question. 3424 * @policy: new policy. 3425 * @param: structure containing the new RT priority. 3426 * 3427 * NOTE that the task may be already dead. 3428 */ 3429int sched_setscheduler(struct task_struct *p, int policy, 3430 const struct sched_param *param) 3431{ 3432 return __sched_setscheduler(p, policy, param, true); 3433} 3434EXPORT_SYMBOL_GPL(sched_setscheduler); 3435 3436/** 3437 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace. 3438 * @p: the task in question. 3439 * @policy: new policy. 3440 * @param: structure containing the new RT priority. 3441 * 3442 * Just like sched_setscheduler, only don't bother checking if the 3443 * current context has permission. For example, this is needed in 3444 * stop_machine(): we create temporary high priority worker threads, 3445 * but our caller might not have that capability. 3446 */ 3447int sched_setscheduler_nocheck(struct task_struct *p, int policy, 3448 const struct sched_param *param) 3449{ 3450 return __sched_setscheduler(p, policy, param, false); 3451} 3452 3453static int 3454do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param) 3455{ 3456 struct sched_param lparam; 3457 struct task_struct *p; 3458 int retval; 3459 3460 if (!param || pid < 0) 3461 return -EINVAL; 3462 if (copy_from_user(&lparam, param, sizeof(struct sched_param))) 3463 return -EFAULT; 3464 3465 rcu_read_lock(); 3466 retval = -ESRCH; 3467 p = find_process_by_pid(pid); 3468 if (p != NULL) 3469 retval = sched_setscheduler(p, policy, &lparam); 3470 rcu_read_unlock(); 3471 3472 return retval; 3473} 3474 3475/** 3476 * sys_sched_setscheduler - set/change the scheduler policy and RT priority 3477 * @pid: the pid in question. 3478 * @policy: new policy. 3479 * @param: structure containing the new RT priority. 3480 */ 3481SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy, 3482 struct sched_param __user *, param) 3483{ 3484 /* negative values for policy are not valid */ 3485 if (policy < 0) 3486 return -EINVAL; 3487 3488 return do_sched_setscheduler(pid, policy, param); 3489} 3490 3491/** 3492 * sys_sched_setparam - set/change the RT priority of a thread 3493 * @pid: the pid in question. 3494 * @param: structure containing the new RT priority. 3495 */ 3496SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param) 3497{ 3498 return do_sched_setscheduler(pid, -1, param); 3499} 3500 3501/** 3502 * sys_sched_getscheduler - get the policy (scheduling class) of a thread 3503 * @pid: the pid in question. 3504 */ 3505SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid) 3506{ 3507 struct task_struct *p; 3508 int retval; 3509 3510 if (pid < 0) 3511 return -EINVAL; 3512 3513 retval = -ESRCH; 3514 rcu_read_lock(); 3515 p = find_process_by_pid(pid); 3516 if (p) { 3517 retval = security_task_getscheduler(p); 3518 if (!retval) 3519 retval = p->policy 3520 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0); 3521 } 3522 rcu_read_unlock(); 3523 return retval; 3524} 3525 3526/** 3527 * sys_sched_getparam - get the RT priority of a thread 3528 * @pid: the pid in question. 3529 * @param: structure containing the RT priority. 3530 */ 3531SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param) 3532{ 3533 struct sched_param lp; 3534 struct task_struct *p; 3535 int retval; 3536 3537 if (!param || pid < 0) 3538 return -EINVAL; 3539 3540 rcu_read_lock(); 3541 p = find_process_by_pid(pid); 3542 retval = -ESRCH; 3543 if (!p) 3544 goto out_unlock; 3545 3546 retval = security_task_getscheduler(p); 3547 if (retval) 3548 goto out_unlock; 3549 3550 lp.sched_priority = p->rt_priority; 3551 rcu_read_unlock(); 3552 3553 /* 3554 * This one might sleep, we cannot do it with a spinlock held ... 3555 */ 3556 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0; 3557 3558 return retval; 3559 3560out_unlock: 3561 rcu_read_unlock(); 3562 return retval; 3563} 3564 3565long sched_setaffinity(pid_t pid, const struct cpumask *in_mask) 3566{ 3567 cpumask_var_t cpus_allowed, new_mask; 3568 struct task_struct *p; 3569 int retval; 3570 3571 get_online_cpus(); 3572 rcu_read_lock(); 3573 3574 p = find_process_by_pid(pid); 3575 if (!p) { 3576 rcu_read_unlock(); 3577 put_online_cpus(); 3578 return -ESRCH; 3579 } 3580 3581 /* Prevent p going away */ 3582 get_task_struct(p); 3583 rcu_read_unlock(); 3584 3585 if (p->flags & PF_NO_SETAFFINITY) { 3586 retval = -EINVAL; 3587 goto out_put_task; 3588 } 3589 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) { 3590 retval = -ENOMEM; 3591 goto out_put_task; 3592 } 3593 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) { 3594 retval = -ENOMEM; 3595 goto out_free_cpus_allowed; 3596 } 3597 retval = -EPERM; 3598 if (!check_same_owner(p)) { 3599 rcu_read_lock(); 3600 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) { 3601 rcu_read_unlock(); 3602 goto out_unlock; 3603 } 3604 rcu_read_unlock(); 3605 } 3606 3607 retval = security_task_setscheduler(p); 3608 if (retval) 3609 goto out_unlock; 3610 3611 cpuset_cpus_allowed(p, cpus_allowed); 3612 cpumask_and(new_mask, in_mask, cpus_allowed); 3613again: 3614 retval = set_cpus_allowed_ptr(p, new_mask); 3615 3616 if (!retval) { 3617 cpuset_cpus_allowed(p, cpus_allowed); 3618 if (!cpumask_subset(new_mask, cpus_allowed)) { 3619 /* 3620 * We must have raced with a concurrent cpuset 3621 * update. Just reset the cpus_allowed to the 3622 * cpuset's cpus_allowed 3623 */ 3624 cpumask_copy(new_mask, cpus_allowed); 3625 goto again; 3626 } 3627 } 3628out_unlock: 3629 free_cpumask_var(new_mask); 3630out_free_cpus_allowed: 3631 free_cpumask_var(cpus_allowed); 3632out_put_task: 3633 put_task_struct(p); 3634 put_online_cpus(); 3635 return retval; 3636} 3637 3638static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len, 3639 struct cpumask *new_mask) 3640{ 3641 if (len < cpumask_size()) 3642 cpumask_clear(new_mask); 3643 else if (len > cpumask_size()) 3644 len = cpumask_size(); 3645 3646 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0; 3647} 3648 3649/** 3650 * sys_sched_setaffinity - set the cpu affinity of a process 3651 * @pid: pid of the process 3652 * @len: length in bytes of the bitmask pointed to by user_mask_ptr 3653 * @user_mask_ptr: user-space pointer to the new cpu mask 3654 */ 3655SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len, 3656 unsigned long __user *, user_mask_ptr) 3657{ 3658 cpumask_var_t new_mask; 3659 int retval; 3660 3661 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) 3662 return -ENOMEM; 3663 3664 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask); 3665 if (retval == 0) 3666 retval = sched_setaffinity(pid, new_mask); 3667 free_cpumask_var(new_mask); 3668 return retval; 3669} 3670 3671long sched_getaffinity(pid_t pid, struct cpumask *mask) 3672{ 3673 struct task_struct *p; 3674 unsigned long flags; 3675 int retval; 3676 3677 get_online_cpus(); 3678 rcu_read_lock(); 3679 3680 retval = -ESRCH; 3681 p = find_process_by_pid(pid); 3682 if (!p) 3683 goto out_unlock; 3684 3685 retval = security_task_getscheduler(p); 3686 if (retval) 3687 goto out_unlock; 3688 3689 raw_spin_lock_irqsave(&p->pi_lock, flags); 3690 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask); 3691 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 3692 3693out_unlock: 3694 rcu_read_unlock(); 3695 put_online_cpus(); 3696 3697 return retval; 3698} 3699 3700/** 3701 * sys_sched_getaffinity - get the cpu affinity of a process 3702 * @pid: pid of the process 3703 * @len: length in bytes of the bitmask pointed to by user_mask_ptr 3704 * @user_mask_ptr: user-space pointer to hold the current cpu mask 3705 */ 3706SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len, 3707 unsigned long __user *, user_mask_ptr) 3708{ 3709 int ret; 3710 cpumask_var_t mask; 3711 3712 if ((len * BITS_PER_BYTE) < nr_cpu_ids) 3713 return -EINVAL; 3714 if (len & (sizeof(unsigned long)-1)) 3715 return -EINVAL; 3716 3717 if (!alloc_cpumask_var(&mask, GFP_KERNEL)) 3718 return -ENOMEM; 3719 3720 ret = sched_getaffinity(pid, mask); 3721 if (ret == 0) { 3722 size_t retlen = min_t(size_t, len, cpumask_size()); 3723 3724 if (copy_to_user(user_mask_ptr, mask, retlen)) 3725 ret = -EFAULT; 3726 else 3727 ret = retlen; 3728 } 3729 free_cpumask_var(mask); 3730 3731 return ret; 3732} 3733 3734/** 3735 * sys_sched_yield - yield the current processor to other threads. 3736 * 3737 * This function yields the current CPU to other tasks. If there are no 3738 * other threads running on this CPU then this function will return. 3739 */ 3740SYSCALL_DEFINE0(sched_yield) 3741{ 3742 struct rq *rq = this_rq_lock(); 3743 3744 schedstat_inc(rq, yld_count); 3745 current->sched_class->yield_task(rq); 3746 3747 /* 3748 * Since we are going to call schedule() anyway, there's 3749 * no need to preempt or enable interrupts: 3750 */ 3751 __release(rq->lock); 3752 spin_release(&rq->lock.dep_map, 1, _THIS_IP_); 3753 do_raw_spin_unlock(&rq->lock); 3754 sched_preempt_enable_no_resched(); 3755 3756 schedule(); 3757 3758 return 0; 3759} 3760 3761static inline int should_resched(void) 3762{ 3763 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE); 3764} 3765 3766static void __cond_resched(void) 3767{ 3768 add_preempt_count(PREEMPT_ACTIVE); 3769 __schedule(); 3770 sub_preempt_count(PREEMPT_ACTIVE); 3771} 3772 3773int __sched _cond_resched(void) 3774{ 3775 if (should_resched()) { 3776 __cond_resched(); 3777 return 1; 3778 } 3779 return 0; 3780} 3781EXPORT_SYMBOL(_cond_resched); 3782 3783/* 3784 * __cond_resched_lock() - if a reschedule is pending, drop the given lock, 3785 * call schedule, and on return reacquire the lock. 3786 * 3787 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level 3788 * operations here to prevent schedule() from being called twice (once via 3789 * spin_unlock(), once by hand). 3790 */ 3791int __cond_resched_lock(spinlock_t *lock) 3792{ 3793 int resched = should_resched(); 3794 int ret = 0; 3795 3796 lockdep_assert_held(lock); 3797 3798 if (spin_needbreak(lock) || resched) { 3799 spin_unlock(lock); 3800 if (resched) 3801 __cond_resched(); 3802 else 3803 cpu_relax(); 3804 ret = 1; 3805 spin_lock(lock); 3806 } 3807 return ret; 3808} 3809EXPORT_SYMBOL(__cond_resched_lock); 3810 3811int __sched __cond_resched_softirq(void) 3812{ 3813 BUG_ON(!in_softirq()); 3814 3815 if (should_resched()) { 3816 local_bh_enable(); 3817 __cond_resched(); 3818 local_bh_disable(); 3819 return 1; 3820 } 3821 return 0; 3822} 3823EXPORT_SYMBOL(__cond_resched_softirq); 3824 3825/** 3826 * yield - yield the current processor to other threads. 3827 * 3828 * Do not ever use this function, there's a 99% chance you're doing it wrong. 3829 * 3830 * The scheduler is at all times free to pick the calling task as the most 3831 * eligible task to run, if removing the yield() call from your code breaks 3832 * it, its already broken. 3833 * 3834 * Typical broken usage is: 3835 * 3836 * while (!event) 3837 * yield(); 3838 * 3839 * where one assumes that yield() will let 'the other' process run that will 3840 * make event true. If the current task is a SCHED_FIFO task that will never 3841 * happen. Never use yield() as a progress guarantee!! 3842 * 3843 * If you want to use yield() to wait for something, use wait_event(). 3844 * If you want to use yield() to be 'nice' for others, use cond_resched(). 3845 * If you still want to use yield(), do not! 3846 */ 3847void __sched yield(void) 3848{ 3849 set_current_state(TASK_RUNNING); 3850 sys_sched_yield(); 3851} 3852EXPORT_SYMBOL(yield); 3853 3854/** 3855 * yield_to - yield the current processor to another thread in 3856 * your thread group, or accelerate that thread toward the 3857 * processor it's on. 3858 * @p: target task 3859 * @preempt: whether task preemption is allowed or not 3860 * 3861 * It's the caller's job to ensure that the target task struct 3862 * can't go away on us before we can do any checks. 3863 * 3864 * Returns: 3865 * true (>0) if we indeed boosted the target task. 3866 * false (0) if we failed to boost the target. 3867 * -ESRCH if there's no task to yield to. 3868 */ 3869bool __sched yield_to(struct task_struct *p, bool preempt) 3870{ 3871 struct task_struct *curr = current; 3872 struct rq *rq, *p_rq; 3873 unsigned long flags; 3874 int yielded = 0; 3875 3876 local_irq_save(flags); 3877 rq = this_rq(); 3878 3879again: 3880 p_rq = task_rq(p); 3881 /* 3882 * If we're the only runnable task on the rq and target rq also 3883 * has only one task, there's absolutely no point in yielding. 3884 */ 3885 if (rq->nr_running == 1 && p_rq->nr_running == 1) { 3886 yielded = -ESRCH; 3887 goto out_irq; 3888 } 3889 3890 double_rq_lock(rq, p_rq); 3891 while (task_rq(p) != p_rq) { 3892 double_rq_unlock(rq, p_rq); 3893 goto again; 3894 } 3895 3896 if (!curr->sched_class->yield_to_task) 3897 goto out_unlock; 3898 3899 if (curr->sched_class != p->sched_class) 3900 goto out_unlock; 3901 3902 if (task_running(p_rq, p) || p->state) 3903 goto out_unlock; 3904 3905 yielded = curr->sched_class->yield_to_task(rq, p, preempt); 3906 if (yielded) { 3907 schedstat_inc(rq, yld_count); 3908 /* 3909 * Make p's CPU reschedule; pick_next_entity takes care of 3910 * fairness. 3911 */ 3912 if (preempt && rq != p_rq) 3913 resched_task(p_rq->curr); 3914 } 3915 3916out_unlock: 3917 double_rq_unlock(rq, p_rq); 3918out_irq: 3919 local_irq_restore(flags); 3920 3921 if (yielded > 0) 3922 schedule(); 3923 3924 return yielded; 3925} 3926EXPORT_SYMBOL_GPL(yield_to); 3927 3928/* 3929 * This task is about to go to sleep on IO. Increment rq->nr_iowait so 3930 * that process accounting knows that this is a task in IO wait state. 3931 */ 3932void __sched io_schedule(void) 3933{ 3934 struct rq *rq = raw_rq(); 3935 3936 delayacct_blkio_start(); 3937 atomic_inc(&rq->nr_iowait); 3938 blk_flush_plug(current); 3939 current->in_iowait = 1; 3940 schedule(); 3941 current->in_iowait = 0; 3942 atomic_dec(&rq->nr_iowait); 3943 delayacct_blkio_end(); 3944} 3945EXPORT_SYMBOL(io_schedule); 3946 3947long __sched io_schedule_timeout(long timeout) 3948{ 3949 struct rq *rq = raw_rq(); 3950 long ret; 3951 3952 delayacct_blkio_start(); 3953 atomic_inc(&rq->nr_iowait); 3954 blk_flush_plug(current); 3955 current->in_iowait = 1; 3956 ret = schedule_timeout(timeout); 3957 current->in_iowait = 0; 3958 atomic_dec(&rq->nr_iowait); 3959 delayacct_blkio_end(); 3960 return ret; 3961} 3962 3963/** 3964 * sys_sched_get_priority_max - return maximum RT priority. 3965 * @policy: scheduling class. 3966 * 3967 * this syscall returns the maximum rt_priority that can be used 3968 * by a given scheduling class. 3969 */ 3970SYSCALL_DEFINE1(sched_get_priority_max, int, policy) 3971{ 3972 int ret = -EINVAL; 3973 3974 switch (policy) { 3975 case SCHED_FIFO: 3976 case SCHED_RR: 3977 ret = MAX_USER_RT_PRIO-1; 3978 break; 3979 case SCHED_NORMAL: 3980 case SCHED_BATCH: 3981 case SCHED_IDLE: 3982 ret = 0; 3983 break; 3984 } 3985 return ret; 3986} 3987 3988/** 3989 * sys_sched_get_priority_min - return minimum RT priority. 3990 * @policy: scheduling class. 3991 * 3992 * this syscall returns the minimum rt_priority that can be used 3993 * by a given scheduling class. 3994 */ 3995SYSCALL_DEFINE1(sched_get_priority_min, int, policy) 3996{ 3997 int ret = -EINVAL; 3998 3999 switch (policy) { 4000 case SCHED_FIFO: 4001 case SCHED_RR: 4002 ret = 1; 4003 break; 4004 case SCHED_NORMAL: 4005 case SCHED_BATCH: 4006 case SCHED_IDLE: 4007 ret = 0; 4008 } 4009 return ret; 4010} 4011 4012/** 4013 * sys_sched_rr_get_interval - return the default timeslice of a process. 4014 * @pid: pid of the process. 4015 * @interval: userspace pointer to the timeslice value. 4016 * 4017 * this syscall writes the default timeslice value of a given process 4018 * into the user-space timespec buffer. A value of '0' means infinity. 4019 */ 4020SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid, 4021 struct timespec __user *, interval) 4022{ 4023 struct task_struct *p; 4024 unsigned int time_slice; 4025 unsigned long flags; 4026 struct rq *rq; 4027 int retval; 4028 struct timespec t; 4029 4030 if (pid < 0) 4031 return -EINVAL; 4032 4033 retval = -ESRCH; 4034 rcu_read_lock(); 4035 p = find_process_by_pid(pid); 4036 if (!p) 4037 goto out_unlock; 4038 4039 retval = security_task_getscheduler(p); 4040 if (retval) 4041 goto out_unlock; 4042 4043 rq = task_rq_lock(p, &flags); 4044 time_slice = p->sched_class->get_rr_interval(rq, p); 4045 task_rq_unlock(rq, p, &flags); 4046 4047 rcu_read_unlock(); 4048 jiffies_to_timespec(time_slice, &t); 4049 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0; 4050 return retval; 4051 4052out_unlock: 4053 rcu_read_unlock(); 4054 return retval; 4055} 4056 4057static const char stat_nam[] = TASK_STATE_TO_CHAR_STR; 4058 4059void sched_show_task(struct task_struct *p) 4060{ 4061 unsigned long free = 0; 4062 int ppid; 4063 unsigned state; 4064 4065 state = p->state ? __ffs(p->state) + 1 : 0; 4066 printk(KERN_INFO "%-15.15s %c", p->comm, 4067 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?'); 4068#if BITS_PER_LONG == 32 4069 if (state == TASK_RUNNING) 4070 printk(KERN_CONT " running "); 4071 else 4072 printk(KERN_CONT " %08lx ", thread_saved_pc(p)); 4073#else 4074 if (state == TASK_RUNNING) 4075 printk(KERN_CONT " running task "); 4076 else 4077 printk(KERN_CONT " %016lx ", thread_saved_pc(p)); 4078#endif 4079#ifdef CONFIG_DEBUG_STACK_USAGE 4080 free = stack_not_used(p); 4081#endif 4082 rcu_read_lock(); 4083 ppid = task_pid_nr(rcu_dereference(p->real_parent)); 4084 rcu_read_unlock(); 4085 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free, 4086 task_pid_nr(p), ppid, 4087 (unsigned long)task_thread_info(p)->flags); 4088 4089 print_worker_info(KERN_INFO, p); 4090 show_stack(p, NULL); 4091} 4092 4093void show_state_filter(unsigned long state_filter) 4094{ 4095 struct task_struct *g, *p; 4096 4097#if BITS_PER_LONG == 32 4098 printk(KERN_INFO 4099 " task PC stack pid father\n"); 4100#else 4101 printk(KERN_INFO 4102 " task PC stack pid father\n"); 4103#endif 4104 rcu_read_lock(); 4105 do_each_thread(g, p) { 4106 /* 4107 * reset the NMI-timeout, listing all files on a slow 4108 * console might take a lot of time: 4109 */ 4110 touch_nmi_watchdog(); 4111 if (!state_filter || (p->state & state_filter)) 4112 sched_show_task(p); 4113 } while_each_thread(g, p); 4114 4115 touch_all_softlockup_watchdogs(); 4116 4117#ifdef CONFIG_SCHED_DEBUG 4118 sysrq_sched_debug_show(); 4119#endif 4120 rcu_read_unlock(); 4121 /* 4122 * Only show locks if all tasks are dumped: 4123 */ 4124 if (!state_filter) 4125 debug_show_all_locks(); 4126} 4127 4128void __cpuinit init_idle_bootup_task(struct task_struct *idle) 4129{ 4130 idle->sched_class = &idle_sched_class; 4131} 4132 4133/** 4134 * init_idle - set up an idle thread for a given CPU 4135 * @idle: task in question 4136 * @cpu: cpu the idle task belongs to 4137 * 4138 * NOTE: this function does not set the idle thread's NEED_RESCHED 4139 * flag, to make booting more robust. 4140 */ 4141void __cpuinit init_idle(struct task_struct *idle, int cpu) 4142{ 4143 struct rq *rq = cpu_rq(cpu); 4144 unsigned long flags; 4145 4146 raw_spin_lock_irqsave(&rq->lock, flags); 4147 4148 __sched_fork(idle); 4149 idle->state = TASK_RUNNING; 4150 idle->se.exec_start = sched_clock(); 4151 4152 do_set_cpus_allowed(idle, cpumask_of(cpu)); 4153 /* 4154 * We're having a chicken and egg problem, even though we are 4155 * holding rq->lock, the cpu isn't yet set to this cpu so the 4156 * lockdep check in task_group() will fail. 4157 * 4158 * Similar case to sched_fork(). / Alternatively we could 4159 * use task_rq_lock() here and obtain the other rq->lock. 4160 * 4161 * Silence PROVE_RCU 4162 */ 4163 rcu_read_lock(); 4164 __set_task_cpu(idle, cpu); 4165 rcu_read_unlock(); 4166 4167 rq->curr = rq->idle = idle; 4168#if defined(CONFIG_SMP) 4169 idle->on_cpu = 1; 4170#endif 4171 raw_spin_unlock_irqrestore(&rq->lock, flags); 4172 4173 /* Set the preempt count _outside_ the spinlocks! */ 4174 task_thread_info(idle)->preempt_count = 0; 4175 4176 /* 4177 * The idle tasks have their own, simple scheduling class: 4178 */ 4179 idle->sched_class = &idle_sched_class; 4180 ftrace_graph_init_idle_task(idle, cpu); 4181 vtime_init_idle(idle); 4182#if defined(CONFIG_SMP) 4183 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu); 4184#endif 4185} 4186 4187#ifdef CONFIG_SMP 4188void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask) 4189{ 4190 if (p->sched_class && p->sched_class->set_cpus_allowed) 4191 p->sched_class->set_cpus_allowed(p, new_mask); 4192 4193 cpumask_copy(&p->cpus_allowed, new_mask); 4194 p->nr_cpus_allowed = cpumask_weight(new_mask); 4195} 4196 4197/* 4198 * This is how migration works: 4199 * 4200 * 1) we invoke migration_cpu_stop() on the target CPU using 4201 * stop_one_cpu(). 4202 * 2) stopper starts to run (implicitly forcing the migrated thread 4203 * off the CPU) 4204 * 3) it checks whether the migrated task is still in the wrong runqueue. 4205 * 4) if it's in the wrong runqueue then the migration thread removes 4206 * it and puts it into the right queue. 4207 * 5) stopper completes and stop_one_cpu() returns and the migration 4208 * is done. 4209 */ 4210 4211/* 4212 * Change a given task's CPU affinity. Migrate the thread to a 4213 * proper CPU and schedule it away if the CPU it's executing on 4214 * is removed from the allowed bitmask. 4215 * 4216 * NOTE: the caller must have a valid reference to the task, the 4217 * task must not exit() & deallocate itself prematurely. The 4218 * call is not atomic; no spinlocks may be held. 4219 */ 4220int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask) 4221{ 4222 unsigned long flags; 4223 struct rq *rq; 4224 unsigned int dest_cpu; 4225 int ret = 0; 4226 4227 rq = task_rq_lock(p, &flags); 4228 4229 if (cpumask_equal(&p->cpus_allowed, new_mask)) 4230 goto out; 4231 4232 if (!cpumask_intersects(new_mask, cpu_active_mask)) { 4233 ret = -EINVAL; 4234 goto out; 4235 } 4236 4237 do_set_cpus_allowed(p, new_mask); 4238 4239 /* Can the task run on the task's current CPU? If so, we're done */ 4240 if (cpumask_test_cpu(task_cpu(p), new_mask)) 4241 goto out; 4242 4243 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask); 4244 if (p->on_rq) { 4245 struct migration_arg arg = { p, dest_cpu }; 4246 /* Need help from migration thread: drop lock and wait. */ 4247 task_rq_unlock(rq, p, &flags); 4248 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg); 4249 tlb_migrate_finish(p->mm); 4250 return 0; 4251 } 4252out: 4253 task_rq_unlock(rq, p, &flags); 4254 4255 return ret; 4256} 4257EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr); 4258 4259/* 4260 * Move (not current) task off this cpu, onto dest cpu. We're doing 4261 * this because either it can't run here any more (set_cpus_allowed() 4262 * away from this CPU, or CPU going down), or because we're 4263 * attempting to rebalance this task on exec (sched_exec). 4264 * 4265 * So we race with normal scheduler movements, but that's OK, as long 4266 * as the task is no longer on this CPU. 4267 * 4268 * Returns non-zero if task was successfully migrated. 4269 */ 4270static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu) 4271{ 4272 struct rq *rq_dest, *rq_src; 4273 int ret = 0; 4274 4275 if (unlikely(!cpu_active(dest_cpu))) 4276 return ret; 4277 4278 rq_src = cpu_rq(src_cpu); 4279 rq_dest = cpu_rq(dest_cpu); 4280 4281 raw_spin_lock(&p->pi_lock); 4282 double_rq_lock(rq_src, rq_dest); 4283 /* Already moved. */ 4284 if (task_cpu(p) != src_cpu) 4285 goto done; 4286 /* Affinity changed (again). */ 4287 if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p))) 4288 goto fail; 4289 4290 /* 4291 * If we're not on a rq, the next wake-up will ensure we're 4292 * placed properly. 4293 */ 4294 if (p->on_rq) { 4295 dequeue_task(rq_src, p, 0); 4296 set_task_cpu(p, dest_cpu); 4297 enqueue_task(rq_dest, p, 0); 4298 check_preempt_curr(rq_dest, p, 0); 4299 } 4300done: 4301 ret = 1; 4302fail: 4303 double_rq_unlock(rq_src, rq_dest); 4304 raw_spin_unlock(&p->pi_lock); 4305 return ret; 4306} 4307 4308/* 4309 * migration_cpu_stop - this will be executed by a highprio stopper thread 4310 * and performs thread migration by bumping thread off CPU then 4311 * 'pushing' onto another runqueue. 4312 */ 4313static int migration_cpu_stop(void *data) 4314{ 4315 struct migration_arg *arg = data; 4316 4317 /* 4318 * The original target cpu might have gone down and we might 4319 * be on another cpu but it doesn't matter. 4320 */ 4321 local_irq_disable(); 4322 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu); 4323 local_irq_enable(); 4324 return 0; 4325} 4326 4327#ifdef CONFIG_HOTPLUG_CPU 4328 4329/* 4330 * Ensures that the idle task is using init_mm right before its cpu goes 4331 * offline. 4332 */ 4333void idle_task_exit(void) 4334{ 4335 struct mm_struct *mm = current->active_mm; 4336 4337 BUG_ON(cpu_online(smp_processor_id())); 4338 4339 if (mm != &init_mm) 4340 switch_mm(mm, &init_mm, current); 4341 mmdrop(mm); 4342} 4343 4344/* 4345 * Since this CPU is going 'away' for a while, fold any nr_active delta 4346 * we might have. Assumes we're called after migrate_tasks() so that the 4347 * nr_active count is stable. 4348 * 4349 * Also see the comment "Global load-average calculations". 4350 */ 4351static void calc_load_migrate(struct rq *rq) 4352{ 4353 long delta = calc_load_fold_active(rq); 4354 if (delta) 4355 atomic_long_add(delta, &calc_load_tasks); 4356} 4357 4358/* 4359 * Migrate all tasks from the rq, sleeping tasks will be migrated by 4360 * try_to_wake_up()->select_task_rq(). 4361 * 4362 * Called with rq->lock held even though we'er in stop_machine() and 4363 * there's no concurrency possible, we hold the required locks anyway 4364 * because of lock validation efforts. 4365 */ 4366static void migrate_tasks(unsigned int dead_cpu) 4367{ 4368 struct rq *rq = cpu_rq(dead_cpu); 4369 struct task_struct *next, *stop = rq->stop; 4370 int dest_cpu; 4371 4372 /* 4373 * Fudge the rq selection such that the below task selection loop 4374 * doesn't get stuck on the currently eligible stop task. 4375 * 4376 * We're currently inside stop_machine() and the rq is either stuck 4377 * in the stop_machine_cpu_stop() loop, or we're executing this code, 4378 * either way we should never end up calling schedule() until we're 4379 * done here. 4380 */ 4381 rq->stop = NULL; 4382 4383 /* 4384 * put_prev_task() and pick_next_task() sched 4385 * class method both need to have an up-to-date 4386 * value of rq->clock[_task] 4387 */ 4388 update_rq_clock(rq); 4389 4390 for ( ; ; ) { 4391 /* 4392 * There's this thread running, bail when that's the only 4393 * remaining thread. 4394 */ 4395 if (rq->nr_running == 1) 4396 break; 4397 4398 next = pick_next_task(rq); 4399 BUG_ON(!next); 4400 next->sched_class->put_prev_task(rq, next); 4401 4402 /* Find suitable destination for @next, with force if needed. */ 4403 dest_cpu = select_fallback_rq(dead_cpu, next); 4404 raw_spin_unlock(&rq->lock); 4405 4406 __migrate_task(next, dead_cpu, dest_cpu); 4407 4408 raw_spin_lock(&rq->lock); 4409 } 4410 4411 rq->stop = stop; 4412} 4413 4414#endif /* CONFIG_HOTPLUG_CPU */ 4415 4416#if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL) 4417 4418static struct ctl_table sd_ctl_dir[] = { 4419 { 4420 .procname = "sched_domain", 4421 .mode = 0555, 4422 }, 4423 {} 4424}; 4425 4426static struct ctl_table sd_ctl_root[] = { 4427 { 4428 .procname = "kernel", 4429 .mode = 0555, 4430 .child = sd_ctl_dir, 4431 }, 4432 {} 4433}; 4434 4435static struct ctl_table *sd_alloc_ctl_entry(int n) 4436{ 4437 struct ctl_table *entry = 4438 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL); 4439 4440 return entry; 4441} 4442 4443static void sd_free_ctl_entry(struct ctl_table **tablep) 4444{ 4445 struct ctl_table *entry; 4446 4447 /* 4448 * In the intermediate directories, both the child directory and 4449 * procname are dynamically allocated and could fail but the mode 4450 * will always be set. In the lowest directory the names are 4451 * static strings and all have proc handlers. 4452 */ 4453 for (entry = *tablep; entry->mode; entry++) { 4454 if (entry->child) 4455 sd_free_ctl_entry(&entry->child); 4456 if (entry->proc_handler == NULL) 4457 kfree(entry->procname); 4458 } 4459 4460 kfree(*tablep); 4461 *tablep = NULL; 4462} 4463 4464static int min_load_idx = 0; 4465static int max_load_idx = CPU_LOAD_IDX_MAX-1; 4466 4467static void 4468set_table_entry(struct ctl_table *entry, 4469 const char *procname, void *data, int maxlen, 4470 umode_t mode, proc_handler *proc_handler, 4471 bool load_idx) 4472{ 4473 entry->procname = procname; 4474 entry->data = data; 4475 entry->maxlen = maxlen; 4476 entry->mode = mode; 4477 entry->proc_handler = proc_handler; 4478 4479 if (load_idx) { 4480 entry->extra1 = &min_load_idx; 4481 entry->extra2 = &max_load_idx; 4482 } 4483} 4484 4485static struct ctl_table * 4486sd_alloc_ctl_domain_table(struct sched_domain *sd) 4487{ 4488 struct ctl_table *table = sd_alloc_ctl_entry(13); 4489 4490 if (table == NULL) 4491 return NULL; 4492 4493 set_table_entry(&table[0], "min_interval", &sd->min_interval, 4494 sizeof(long), 0644, proc_doulongvec_minmax, false); 4495 set_table_entry(&table[1], "max_interval", &sd->max_interval, 4496 sizeof(long), 0644, proc_doulongvec_minmax, false); 4497 set_table_entry(&table[2], "busy_idx", &sd->busy_idx, 4498 sizeof(int), 0644, proc_dointvec_minmax, true); 4499 set_table_entry(&table[3], "idle_idx", &sd->idle_idx, 4500 sizeof(int), 0644, proc_dointvec_minmax, true); 4501 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx, 4502 sizeof(int), 0644, proc_dointvec_minmax, true); 4503 set_table_entry(&table[5], "wake_idx", &sd->wake_idx, 4504 sizeof(int), 0644, proc_dointvec_minmax, true); 4505 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx, 4506 sizeof(int), 0644, proc_dointvec_minmax, true); 4507 set_table_entry(&table[7], "busy_factor", &sd->busy_factor, 4508 sizeof(int), 0644, proc_dointvec_minmax, false); 4509 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct, 4510 sizeof(int), 0644, proc_dointvec_minmax, false); 4511 set_table_entry(&table[9], "cache_nice_tries", 4512 &sd->cache_nice_tries, 4513 sizeof(int), 0644, proc_dointvec_minmax, false); 4514 set_table_entry(&table[10], "flags", &sd->flags, 4515 sizeof(int), 0644, proc_dointvec_minmax, false); 4516 set_table_entry(&table[11], "name", sd->name, 4517 CORENAME_MAX_SIZE, 0444, proc_dostring, false); 4518 /* &table[12] is terminator */ 4519 4520 return table; 4521} 4522 4523static ctl_table *sd_alloc_ctl_cpu_table(int cpu) 4524{ 4525 struct ctl_table *entry, *table; 4526 struct sched_domain *sd; 4527 int domain_num = 0, i; 4528 char buf[32]; 4529 4530 for_each_domain(cpu, sd) 4531 domain_num++; 4532 entry = table = sd_alloc_ctl_entry(domain_num + 1); 4533 if (table == NULL) 4534 return NULL; 4535 4536 i = 0; 4537 for_each_domain(cpu, sd) { 4538 snprintf(buf, 32, "domain%d", i); 4539 entry->procname = kstrdup(buf, GFP_KERNEL); 4540 entry->mode = 0555; 4541 entry->child = sd_alloc_ctl_domain_table(sd); 4542 entry++; 4543 i++; 4544 } 4545 return table; 4546} 4547 4548static struct ctl_table_header *sd_sysctl_header; 4549static void register_sched_domain_sysctl(void) 4550{ 4551 int i, cpu_num = num_possible_cpus(); 4552 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1); 4553 char buf[32]; 4554 4555 WARN_ON(sd_ctl_dir[0].child); 4556 sd_ctl_dir[0].child = entry; 4557 4558 if (entry == NULL) 4559 return; 4560 4561 for_each_possible_cpu(i) { 4562 snprintf(buf, 32, "cpu%d", i); 4563 entry->procname = kstrdup(buf, GFP_KERNEL); 4564 entry->mode = 0555; 4565 entry->child = sd_alloc_ctl_cpu_table(i); 4566 entry++; 4567 } 4568 4569 WARN_ON(sd_sysctl_header); 4570 sd_sysctl_header = register_sysctl_table(sd_ctl_root); 4571} 4572 4573/* may be called multiple times per register */ 4574static void unregister_sched_domain_sysctl(void) 4575{ 4576 if (sd_sysctl_header) 4577 unregister_sysctl_table(sd_sysctl_header); 4578 sd_sysctl_header = NULL; 4579 if (sd_ctl_dir[0].child) 4580 sd_free_ctl_entry(&sd_ctl_dir[0].child); 4581} 4582#else 4583static void register_sched_domain_sysctl(void) 4584{ 4585} 4586static void unregister_sched_domain_sysctl(void) 4587{ 4588} 4589#endif 4590 4591static void set_rq_online(struct rq *rq) 4592{ 4593 if (!rq->online) { 4594 const struct sched_class *class; 4595 4596 cpumask_set_cpu(rq->cpu, rq->rd->online); 4597 rq->online = 1; 4598 4599 for_each_class(class) { 4600 if (class->rq_online) 4601 class->rq_online(rq); 4602 } 4603 } 4604} 4605 4606static void set_rq_offline(struct rq *rq) 4607{ 4608 if (rq->online) { 4609 const struct sched_class *class; 4610 4611 for_each_class(class) { 4612 if (class->rq_offline) 4613 class->rq_offline(rq); 4614 } 4615 4616 cpumask_clear_cpu(rq->cpu, rq->rd->online); 4617 rq->online = 0; 4618 } 4619} 4620 4621/* 4622 * migration_call - callback that gets triggered when a CPU is added. 4623 * Here we can start up the necessary migration thread for the new CPU. 4624 */ 4625static int __cpuinit 4626migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu) 4627{ 4628 int cpu = (long)hcpu; 4629 unsigned long flags; 4630 struct rq *rq = cpu_rq(cpu); 4631 4632 switch (action & ~CPU_TASKS_FROZEN) { 4633 4634 case CPU_UP_PREPARE: 4635 rq->calc_load_update = calc_load_update; 4636 break; 4637 4638 case CPU_ONLINE: 4639 /* Update our root-domain */ 4640 raw_spin_lock_irqsave(&rq->lock, flags); 4641 if (rq->rd) { 4642 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span)); 4643 4644 set_rq_online(rq); 4645 } 4646 raw_spin_unlock_irqrestore(&rq->lock, flags); 4647 break; 4648 4649#ifdef CONFIG_HOTPLUG_CPU 4650 case CPU_DYING: 4651 sched_ttwu_pending(); 4652 /* Update our root-domain */ 4653 raw_spin_lock_irqsave(&rq->lock, flags); 4654 if (rq->rd) { 4655 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span)); 4656 set_rq_offline(rq); 4657 } 4658 migrate_tasks(cpu); 4659 BUG_ON(rq->nr_running != 1); /* the migration thread */ 4660 raw_spin_unlock_irqrestore(&rq->lock, flags); 4661 break; 4662 4663 case CPU_DEAD: 4664 calc_load_migrate(rq); 4665 break; 4666#endif 4667 } 4668 4669 update_max_interval(); 4670 4671 return NOTIFY_OK; 4672} 4673 4674/* 4675 * Register at high priority so that task migration (migrate_all_tasks) 4676 * happens before everything else. This has to be lower priority than 4677 * the notifier in the perf_event subsystem, though. 4678 */ 4679static struct notifier_block __cpuinitdata migration_notifier = { 4680 .notifier_call = migration_call, 4681 .priority = CPU_PRI_MIGRATION, 4682}; 4683 4684static int __cpuinit sched_cpu_active(struct notifier_block *nfb, 4685 unsigned long action, void *hcpu) 4686{ 4687 switch (action & ~CPU_TASKS_FROZEN) { 4688 case CPU_STARTING: 4689 case CPU_DOWN_FAILED: 4690 set_cpu_active((long)hcpu, true); 4691 return NOTIFY_OK; 4692 default: 4693 return NOTIFY_DONE; 4694 } 4695} 4696 4697static int __cpuinit sched_cpu_inactive(struct notifier_block *nfb, 4698 unsigned long action, void *hcpu) 4699{ 4700 switch (action & ~CPU_TASKS_FROZEN) { 4701 case CPU_DOWN_PREPARE: 4702 set_cpu_active((long)hcpu, false); 4703 return NOTIFY_OK; 4704 default: 4705 return NOTIFY_DONE; 4706 } 4707} 4708 4709static int __init migration_init(void) 4710{ 4711 void *cpu = (void *)(long)smp_processor_id(); 4712 int err; 4713 4714 /* Initialize migration for the boot CPU */ 4715 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu); 4716 BUG_ON(err == NOTIFY_BAD); 4717 migration_call(&migration_notifier, CPU_ONLINE, cpu); 4718 register_cpu_notifier(&migration_notifier); 4719 4720 /* Register cpu active notifiers */ 4721 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE); 4722 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE); 4723 4724 return 0; 4725} 4726early_initcall(migration_init); 4727#endif 4728 4729#ifdef CONFIG_SMP 4730 4731static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */ 4732 4733#ifdef CONFIG_SCHED_DEBUG 4734 4735static __read_mostly int sched_debug_enabled; 4736 4737static int __init sched_debug_setup(char *str) 4738{ 4739 sched_debug_enabled = 1; 4740 4741 return 0; 4742} 4743early_param("sched_debug", sched_debug_setup); 4744 4745static inline bool sched_debug(void) 4746{ 4747 return sched_debug_enabled; 4748} 4749 4750static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level, 4751 struct cpumask *groupmask) 4752{ 4753 struct sched_group *group = sd->groups; 4754 char str[256]; 4755 4756 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd)); 4757 cpumask_clear(groupmask); 4758 4759 printk(KERN_DEBUG "%*s domain %d: ", level, "", level); 4760 4761 if (!(sd->flags & SD_LOAD_BALANCE)) { 4762 printk("does not load-balance\n"); 4763 if (sd->parent) 4764 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain" 4765 " has parent"); 4766 return -1; 4767 } 4768 4769 printk(KERN_CONT "span %s level %s\n", str, sd->name); 4770 4771 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) { 4772 printk(KERN_ERR "ERROR: domain->span does not contain " 4773 "CPU%d\n", cpu); 4774 } 4775 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) { 4776 printk(KERN_ERR "ERROR: domain->groups does not contain" 4777 " CPU%d\n", cpu); 4778 } 4779 4780 printk(KERN_DEBUG "%*s groups:", level + 1, ""); 4781 do { 4782 if (!group) { 4783 printk("\n"); 4784 printk(KERN_ERR "ERROR: group is NULL\n"); 4785 break; 4786 } 4787 4788 /* 4789 * Even though we initialize ->power to something semi-sane, 4790 * we leave power_orig unset. This allows us to detect if 4791 * domain iteration is still funny without causing /0 traps. 4792 */ 4793 if (!group->sgp->power_orig) { 4794 printk(KERN_CONT "\n"); 4795 printk(KERN_ERR "ERROR: domain->cpu_power not " 4796 "set\n"); 4797 break; 4798 } 4799 4800 if (!cpumask_weight(sched_group_cpus(group))) { 4801 printk(KERN_CONT "\n"); 4802 printk(KERN_ERR "ERROR: empty group\n"); 4803 break; 4804 } 4805 4806 if (!(sd->flags & SD_OVERLAP) && 4807 cpumask_intersects(groupmask, sched_group_cpus(group))) { 4808 printk(KERN_CONT "\n"); 4809 printk(KERN_ERR "ERROR: repeated CPUs\n"); 4810 break; 4811 } 4812 4813 cpumask_or(groupmask, groupmask, sched_group_cpus(group)); 4814 4815 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group)); 4816 4817 printk(KERN_CONT " %s", str); 4818 if (group->sgp->power != SCHED_POWER_SCALE) { 4819 printk(KERN_CONT " (cpu_power = %d)", 4820 group->sgp->power); 4821 } 4822 4823 group = group->next; 4824 } while (group != sd->groups); 4825 printk(KERN_CONT "\n"); 4826 4827 if (!cpumask_equal(sched_domain_span(sd), groupmask)) 4828 printk(KERN_ERR "ERROR: groups don't span domain->span\n"); 4829 4830 if (sd->parent && 4831 !cpumask_subset(groupmask, sched_domain_span(sd->parent))) 4832 printk(KERN_ERR "ERROR: parent span is not a superset " 4833 "of domain->span\n"); 4834 return 0; 4835} 4836 4837static void sched_domain_debug(struct sched_domain *sd, int cpu) 4838{ 4839 int level = 0; 4840 4841 if (!sched_debug_enabled) 4842 return; 4843 4844 if (!sd) { 4845 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu); 4846 return; 4847 } 4848 4849 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu); 4850 4851 for (;;) { 4852 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask)) 4853 break; 4854 level++; 4855 sd = sd->parent; 4856 if (!sd) 4857 break; 4858 } 4859} 4860#else /* !CONFIG_SCHED_DEBUG */ 4861# define sched_domain_debug(sd, cpu) do { } while (0) 4862static inline bool sched_debug(void) 4863{ 4864 return false; 4865} 4866#endif /* CONFIG_SCHED_DEBUG */ 4867 4868static int sd_degenerate(struct sched_domain *sd) 4869{ 4870 if (cpumask_weight(sched_domain_span(sd)) == 1) 4871 return 1; 4872 4873 /* Following flags need at least 2 groups */ 4874 if (sd->flags & (SD_LOAD_BALANCE | 4875 SD_BALANCE_NEWIDLE | 4876 SD_BALANCE_FORK | 4877 SD_BALANCE_EXEC | 4878 SD_SHARE_CPUPOWER | 4879 SD_SHARE_PKG_RESOURCES)) { 4880 if (sd->groups != sd->groups->next) 4881 return 0; 4882 } 4883 4884 /* Following flags don't use groups */ 4885 if (sd->flags & (SD_WAKE_AFFINE)) 4886 return 0; 4887 4888 return 1; 4889} 4890 4891static int 4892sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent) 4893{ 4894 unsigned long cflags = sd->flags, pflags = parent->flags; 4895 4896 if (sd_degenerate(parent)) 4897 return 1; 4898 4899 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent))) 4900 return 0; 4901 4902 /* Flags needing groups don't count if only 1 group in parent */ 4903 if (parent->groups == parent->groups->next) { 4904 pflags &= ~(SD_LOAD_BALANCE | 4905 SD_BALANCE_NEWIDLE | 4906 SD_BALANCE_FORK | 4907 SD_BALANCE_EXEC | 4908 SD_SHARE_CPUPOWER | 4909 SD_SHARE_PKG_RESOURCES); 4910 if (nr_node_ids == 1) 4911 pflags &= ~SD_SERIALIZE; 4912 } 4913 if (~cflags & pflags) 4914 return 0; 4915 4916 return 1; 4917} 4918 4919static void free_rootdomain(struct rcu_head *rcu) 4920{ 4921 struct root_domain *rd = container_of(rcu, struct root_domain, rcu); 4922 4923 cpupri_cleanup(&rd->cpupri); 4924 free_cpumask_var(rd->rto_mask); 4925 free_cpumask_var(rd->online); 4926 free_cpumask_var(rd->span); 4927 kfree(rd); 4928} 4929 4930static void rq_attach_root(struct rq *rq, struct root_domain *rd) 4931{ 4932 struct root_domain *old_rd = NULL; 4933 unsigned long flags; 4934 4935 raw_spin_lock_irqsave(&rq->lock, flags); 4936 4937 if (rq->rd) { 4938 old_rd = rq->rd; 4939 4940 if (cpumask_test_cpu(rq->cpu, old_rd->online)) 4941 set_rq_offline(rq); 4942 4943 cpumask_clear_cpu(rq->cpu, old_rd->span); 4944 4945 /* 4946 * If we dont want to free the old_rt yet then 4947 * set old_rd to NULL to skip the freeing later 4948 * in this function: 4949 */ 4950 if (!atomic_dec_and_test(&old_rd->refcount)) 4951 old_rd = NULL; 4952 } 4953 4954 atomic_inc(&rd->refcount); 4955 rq->rd = rd; 4956 4957 cpumask_set_cpu(rq->cpu, rd->span); 4958 if (cpumask_test_cpu(rq->cpu, cpu_active_mask)) 4959 set_rq_online(rq); 4960 4961 raw_spin_unlock_irqrestore(&rq->lock, flags); 4962 4963 if (old_rd) 4964 call_rcu_sched(&old_rd->rcu, free_rootdomain); 4965} 4966 4967static int init_rootdomain(struct root_domain *rd) 4968{ 4969 memset(rd, 0, sizeof(*rd)); 4970 4971 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL)) 4972 goto out; 4973 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL)) 4974 goto free_span; 4975 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL)) 4976 goto free_online; 4977 4978 if (cpupri_init(&rd->cpupri) != 0) 4979 goto free_rto_mask; 4980 return 0; 4981 4982free_rto_mask: 4983 free_cpumask_var(rd->rto_mask); 4984free_online: 4985 free_cpumask_var(rd->online); 4986free_span: 4987 free_cpumask_var(rd->span); 4988out: 4989 return -ENOMEM; 4990} 4991 4992/* 4993 * By default the system creates a single root-domain with all cpus as 4994 * members (mimicking the global state we have today). 4995 */ 4996struct root_domain def_root_domain; 4997 4998static void init_defrootdomain(void) 4999{ 5000 init_rootdomain(&def_root_domain); 5001 5002 atomic_set(&def_root_domain.refcount, 1); 5003} 5004 5005static struct root_domain *alloc_rootdomain(void) 5006{ 5007 struct root_domain *rd; 5008 5009 rd = kmalloc(sizeof(*rd), GFP_KERNEL); 5010 if (!rd) 5011 return NULL; 5012 5013 if (init_rootdomain(rd) != 0) { 5014 kfree(rd); 5015 return NULL; 5016 } 5017 5018 return rd; 5019} 5020 5021static void free_sched_groups(struct sched_group *sg, int free_sgp) 5022{ 5023 struct sched_group *tmp, *first; 5024 5025 if (!sg) 5026 return; 5027 5028 first = sg; 5029 do { 5030 tmp = sg->next; 5031 5032 if (free_sgp && atomic_dec_and_test(&sg->sgp->ref)) 5033 kfree(sg->sgp); 5034 5035 kfree(sg); 5036 sg = tmp; 5037 } while (sg != first); 5038} 5039 5040static void free_sched_domain(struct rcu_head *rcu) 5041{ 5042 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu); 5043 5044 /* 5045 * If its an overlapping domain it has private groups, iterate and 5046 * nuke them all. 5047 */ 5048 if (sd->flags & SD_OVERLAP) { 5049 free_sched_groups(sd->groups, 1); 5050 } else if (atomic_dec_and_test(&sd->groups->ref)) { 5051 kfree(sd->groups->sgp); 5052 kfree(sd->groups); 5053 } 5054 kfree(sd); 5055} 5056 5057static void destroy_sched_domain(struct sched_domain *sd, int cpu) 5058{ 5059 call_rcu(&sd->rcu, free_sched_domain); 5060} 5061 5062static void destroy_sched_domains(struct sched_domain *sd, int cpu) 5063{ 5064 for (; sd; sd = sd->parent) 5065 destroy_sched_domain(sd, cpu); 5066} 5067 5068/* 5069 * Keep a special pointer to the highest sched_domain that has 5070 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this 5071 * allows us to avoid some pointer chasing select_idle_sibling(). 5072 * 5073 * Also keep a unique ID per domain (we use the first cpu number in 5074 * the cpumask of the domain), this allows us to quickly tell if 5075 * two cpus are in the same cache domain, see cpus_share_cache(). 5076 */ 5077DEFINE_PER_CPU(struct sched_domain *, sd_llc); 5078DEFINE_PER_CPU(int, sd_llc_id); 5079 5080static void update_top_cache_domain(int cpu) 5081{ 5082 struct sched_domain *sd; 5083 int id = cpu; 5084 5085 sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES); 5086 if (sd) 5087 id = cpumask_first(sched_domain_span(sd)); 5088 5089 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd); 5090 per_cpu(sd_llc_id, cpu) = id; 5091} 5092 5093/* 5094 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must 5095 * hold the hotplug lock. 5096 */ 5097static void 5098cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu) 5099{ 5100 struct rq *rq = cpu_rq(cpu); 5101 struct sched_domain *tmp; 5102 5103 /* Remove the sched domains which do not contribute to scheduling. */ 5104 for (tmp = sd; tmp; ) { 5105 struct sched_domain *parent = tmp->parent; 5106 if (!parent) 5107 break; 5108 5109 if (sd_parent_degenerate(tmp, parent)) { 5110 tmp->parent = parent->parent; 5111 if (parent->parent) 5112 parent->parent->child = tmp; 5113 destroy_sched_domain(parent, cpu); 5114 } else 5115 tmp = tmp->parent; 5116 } 5117 5118 if (sd && sd_degenerate(sd)) { 5119 tmp = sd; 5120 sd = sd->parent; 5121 destroy_sched_domain(tmp, cpu); 5122 if (sd) 5123 sd->child = NULL; 5124 } 5125 5126 sched_domain_debug(sd, cpu); 5127 5128 rq_attach_root(rq, rd); 5129 tmp = rq->sd; 5130 rcu_assign_pointer(rq->sd, sd); 5131 destroy_sched_domains(tmp, cpu); 5132 5133 update_top_cache_domain(cpu); 5134} 5135 5136/* cpus with isolated domains */ 5137static cpumask_var_t cpu_isolated_map; 5138 5139/* Setup the mask of cpus configured for isolated domains */ 5140static int __init isolated_cpu_setup(char *str) 5141{ 5142 alloc_bootmem_cpumask_var(&cpu_isolated_map); 5143 cpulist_parse(str, cpu_isolated_map); 5144 return 1; 5145} 5146 5147__setup("isolcpus=", isolated_cpu_setup); 5148 5149static const struct cpumask *cpu_cpu_mask(int cpu) 5150{ 5151 return cpumask_of_node(cpu_to_node(cpu)); 5152} 5153 5154struct sd_data { 5155 struct sched_domain **__percpu sd; 5156 struct sched_group **__percpu sg; 5157 struct sched_group_power **__percpu sgp; 5158}; 5159 5160struct s_data { 5161 struct sched_domain ** __percpu sd; 5162 struct root_domain *rd; 5163}; 5164 5165enum s_alloc { 5166 sa_rootdomain, 5167 sa_sd, 5168 sa_sd_storage, 5169 sa_none, 5170}; 5171 5172struct sched_domain_topology_level; 5173 5174typedef struct sched_domain *(*sched_domain_init_f)(struct sched_domain_topology_level *tl, int cpu); 5175typedef const struct cpumask *(*sched_domain_mask_f)(int cpu); 5176 5177#define SDTL_OVERLAP 0x01 5178 5179struct sched_domain_topology_level { 5180 sched_domain_init_f init; 5181 sched_domain_mask_f mask; 5182 int flags; 5183 int numa_level; 5184 struct sd_data data; 5185}; 5186 5187/* 5188 * Build an iteration mask that can exclude certain CPUs from the upwards 5189 * domain traversal. 5190 * 5191 * Asymmetric node setups can result in situations where the domain tree is of 5192 * unequal depth, make sure to skip domains that already cover the entire 5193 * range. 5194 * 5195 * In that case build_sched_domains() will have terminated the iteration early 5196 * and our sibling sd spans will be empty. Domains should always include the 5197 * cpu they're built on, so check that. 5198 * 5199 */ 5200static void build_group_mask(struct sched_domain *sd, struct sched_group *sg) 5201{ 5202 const struct cpumask *span = sched_domain_span(sd); 5203 struct sd_data *sdd = sd->private; 5204 struct sched_domain *sibling; 5205 int i; 5206 5207 for_each_cpu(i, span) { 5208 sibling = *per_cpu_ptr(sdd->sd, i); 5209 if (!cpumask_test_cpu(i, sched_domain_span(sibling))) 5210 continue; 5211 5212 cpumask_set_cpu(i, sched_group_mask(sg)); 5213 } 5214} 5215 5216/* 5217 * Return the canonical balance cpu for this group, this is the first cpu 5218 * of this group that's also in the iteration mask. 5219 */ 5220int group_balance_cpu(struct sched_group *sg) 5221{ 5222 return cpumask_first_and(sched_group_cpus(sg), sched_group_mask(sg)); 5223} 5224 5225static int 5226build_overlap_sched_groups(struct sched_domain *sd, int cpu) 5227{ 5228 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg; 5229 const struct cpumask *span = sched_domain_span(sd); 5230 struct cpumask *covered = sched_domains_tmpmask; 5231 struct sd_data *sdd = sd->private; 5232 struct sched_domain *child; 5233 int i; 5234 5235 cpumask_clear(covered); 5236 5237 for_each_cpu(i, span) { 5238 struct cpumask *sg_span; 5239 5240 if (cpumask_test_cpu(i, covered)) 5241 continue; 5242 5243 child = *per_cpu_ptr(sdd->sd, i); 5244 5245 /* See the comment near build_group_mask(). */ 5246 if (!cpumask_test_cpu(i, sched_domain_span(child))) 5247 continue; 5248 5249 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(), 5250 GFP_KERNEL, cpu_to_node(cpu)); 5251 5252 if (!sg) 5253 goto fail; 5254 5255 sg_span = sched_group_cpus(sg); 5256 if (child->child) { 5257 child = child->child; 5258 cpumask_copy(sg_span, sched_domain_span(child)); 5259 } else 5260 cpumask_set_cpu(i, sg_span); 5261 5262 cpumask_or(covered, covered, sg_span); 5263 5264 sg->sgp = *per_cpu_ptr(sdd->sgp, i); 5265 if (atomic_inc_return(&sg->sgp->ref) == 1) 5266 build_group_mask(sd, sg); 5267 5268 /* 5269 * Initialize sgp->power such that even if we mess up the 5270 * domains and no possible iteration will get us here, we won't 5271 * die on a /0 trap. 5272 */ 5273 sg->sgp->power = SCHED_POWER_SCALE * cpumask_weight(sg_span); 5274 5275 /* 5276 * Make sure the first group of this domain contains the 5277 * canonical balance cpu. Otherwise the sched_domain iteration 5278 * breaks. See update_sg_lb_stats(). 5279 */ 5280 if ((!groups && cpumask_test_cpu(cpu, sg_span)) || 5281 group_balance_cpu(sg) == cpu) 5282 groups = sg; 5283 5284 if (!first) 5285 first = sg; 5286 if (last) 5287 last->next = sg; 5288 last = sg; 5289 last->next = first; 5290 } 5291 sd->groups = groups; 5292 5293 return 0; 5294 5295fail: 5296 free_sched_groups(first, 0); 5297 5298 return -ENOMEM; 5299} 5300 5301static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg) 5302{ 5303 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu); 5304 struct sched_domain *child = sd->child; 5305 5306 if (child) 5307 cpu = cpumask_first(sched_domain_span(child)); 5308 5309 if (sg) { 5310 *sg = *per_cpu_ptr(sdd->sg, cpu); 5311 (*sg)->sgp = *per_cpu_ptr(sdd->sgp, cpu); 5312 atomic_set(&(*sg)->sgp->ref, 1); /* for claim_allocations */ 5313 } 5314 5315 return cpu; 5316} 5317 5318/* 5319 * build_sched_groups will build a circular linked list of the groups 5320 * covered by the given span, and will set each group's ->cpumask correctly, 5321 * and ->cpu_power to 0. 5322 * 5323 * Assumes the sched_domain tree is fully constructed 5324 */ 5325static int 5326build_sched_groups(struct sched_domain *sd, int cpu) 5327{ 5328 struct sched_group *first = NULL, *last = NULL; 5329 struct sd_data *sdd = sd->private; 5330 const struct cpumask *span = sched_domain_span(sd); 5331 struct cpumask *covered; 5332 int i; 5333 5334 get_group(cpu, sdd, &sd->groups); 5335 atomic_inc(&sd->groups->ref); 5336 5337 if (cpu != cpumask_first(sched_domain_span(sd))) 5338 return 0; 5339 5340 lockdep_assert_held(&sched_domains_mutex); 5341 covered = sched_domains_tmpmask; 5342 5343 cpumask_clear(covered); 5344 5345 for_each_cpu(i, span) { 5346 struct sched_group *sg; 5347 int group = get_group(i, sdd, &sg); 5348 int j; 5349 5350 if (cpumask_test_cpu(i, covered)) 5351 continue; 5352 5353 cpumask_clear(sched_group_cpus(sg)); 5354 sg->sgp->power = 0; 5355 cpumask_setall(sched_group_mask(sg)); 5356 5357 for_each_cpu(j, span) { 5358 if (get_group(j, sdd, NULL) != group) 5359 continue; 5360 5361 cpumask_set_cpu(j, covered); 5362 cpumask_set_cpu(j, sched_group_cpus(sg)); 5363 } 5364 5365 if (!first) 5366 first = sg; 5367 if (last) 5368 last->next = sg; 5369 last = sg; 5370 } 5371 last->next = first; 5372 5373 return 0; 5374} 5375 5376/* 5377 * Initialize sched groups cpu_power. 5378 * 5379 * cpu_power indicates the capacity of sched group, which is used while 5380 * distributing the load between different sched groups in a sched domain. 5381 * Typically cpu_power for all the groups in a sched domain will be same unless 5382 * there are asymmetries in the topology. If there are asymmetries, group 5383 * having more cpu_power will pickup more load compared to the group having 5384 * less cpu_power. 5385 */ 5386static void init_sched_groups_power(int cpu, struct sched_domain *sd) 5387{ 5388 struct sched_group *sg = sd->groups; 5389 5390 WARN_ON(!sd || !sg); 5391 5392 do { 5393 sg->group_weight = cpumask_weight(sched_group_cpus(sg)); 5394 sg = sg->next; 5395 } while (sg != sd->groups); 5396 5397 if (cpu != group_balance_cpu(sg)) 5398 return; 5399 5400 update_group_power(sd, cpu); 5401 atomic_set(&sg->sgp->nr_busy_cpus, sg->group_weight); 5402} 5403 5404int __weak arch_sd_sibling_asym_packing(void) 5405{ 5406 return 0*SD_ASYM_PACKING; 5407} 5408 5409/* 5410 * Initializers for schedule domains 5411 * Non-inlined to reduce accumulated stack pressure in build_sched_domains() 5412 */ 5413 5414#ifdef CONFIG_SCHED_DEBUG 5415# define SD_INIT_NAME(sd, type) sd->name = #type 5416#else 5417# define SD_INIT_NAME(sd, type) do { } while (0) 5418#endif 5419 5420#define SD_INIT_FUNC(type) \ 5421static noinline struct sched_domain * \ 5422sd_init_##type(struct sched_domain_topology_level *tl, int cpu) \ 5423{ \ 5424 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); \ 5425 *sd = SD_##type##_INIT; \ 5426 SD_INIT_NAME(sd, type); \ 5427 sd->private = &tl->data; \ 5428 return sd; \ 5429} 5430 5431SD_INIT_FUNC(CPU) 5432#ifdef CONFIG_SCHED_SMT 5433 SD_INIT_FUNC(SIBLING) 5434#endif 5435#ifdef CONFIG_SCHED_MC 5436 SD_INIT_FUNC(MC) 5437#endif 5438#ifdef CONFIG_SCHED_BOOK 5439 SD_INIT_FUNC(BOOK) 5440#endif 5441 5442static int default_relax_domain_level = -1; 5443int sched_domain_level_max; 5444 5445static int __init setup_relax_domain_level(char *str) 5446{ 5447 if (kstrtoint(str, 0, &default_relax_domain_level)) 5448 pr_warn("Unable to set relax_domain_level\n"); 5449 5450 return 1; 5451} 5452__setup("relax_domain_level=", setup_relax_domain_level); 5453 5454static void set_domain_attribute(struct sched_domain *sd, 5455 struct sched_domain_attr *attr) 5456{ 5457 int request; 5458 5459 if (!attr || attr->relax_domain_level < 0) { 5460 if (default_relax_domain_level < 0) 5461 return; 5462 else 5463 request = default_relax_domain_level; 5464 } else 5465 request = attr->relax_domain_level; 5466 if (request < sd->level) { 5467 /* turn off idle balance on this domain */ 5468 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE); 5469 } else { 5470 /* turn on idle balance on this domain */ 5471 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE); 5472 } 5473} 5474 5475static void __sdt_free(const struct cpumask *cpu_map); 5476static int __sdt_alloc(const struct cpumask *cpu_map); 5477 5478static void __free_domain_allocs(struct s_data *d, enum s_alloc what, 5479 const struct cpumask *cpu_map) 5480{ 5481 switch (what) { 5482 case sa_rootdomain: 5483 if (!atomic_read(&d->rd->refcount)) 5484 free_rootdomain(&d->rd->rcu); /* fall through */ 5485 case sa_sd: 5486 free_percpu(d->sd); /* fall through */ 5487 case sa_sd_storage: 5488 __sdt_free(cpu_map); /* fall through */ 5489 case sa_none: 5490 break; 5491 } 5492} 5493 5494static enum s_alloc __visit_domain_allocation_hell(struct s_data *d, 5495 const struct cpumask *cpu_map) 5496{ 5497 memset(d, 0, sizeof(*d)); 5498 5499 if (__sdt_alloc(cpu_map)) 5500 return sa_sd_storage; 5501 d->sd = alloc_percpu(struct sched_domain *); 5502 if (!d->sd) 5503 return sa_sd_storage; 5504 d->rd = alloc_rootdomain(); 5505 if (!d->rd) 5506 return sa_sd; 5507 return sa_rootdomain; 5508} 5509 5510/* 5511 * NULL the sd_data elements we've used to build the sched_domain and 5512 * sched_group structure so that the subsequent __free_domain_allocs() 5513 * will not free the data we're using. 5514 */ 5515static void claim_allocations(int cpu, struct sched_domain *sd) 5516{ 5517 struct sd_data *sdd = sd->private; 5518 5519 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd); 5520 *per_cpu_ptr(sdd->sd, cpu) = NULL; 5521 5522 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref)) 5523 *per_cpu_ptr(sdd->sg, cpu) = NULL; 5524 5525 if (atomic_read(&(*per_cpu_ptr(sdd->sgp, cpu))->ref)) 5526 *per_cpu_ptr(sdd->sgp, cpu) = NULL; 5527} 5528 5529#ifdef CONFIG_SCHED_SMT 5530static const struct cpumask *cpu_smt_mask(int cpu) 5531{ 5532 return topology_thread_cpumask(cpu); 5533} 5534#endif 5535 5536/* 5537 * Topology list, bottom-up. 5538 */ 5539static struct sched_domain_topology_level default_topology[] = { 5540#ifdef CONFIG_SCHED_SMT 5541 { sd_init_SIBLING, cpu_smt_mask, }, 5542#endif 5543#ifdef CONFIG_SCHED_MC 5544 { sd_init_MC, cpu_coregroup_mask, }, 5545#endif 5546#ifdef CONFIG_SCHED_BOOK 5547 { sd_init_BOOK, cpu_book_mask, }, 5548#endif 5549 { sd_init_CPU, cpu_cpu_mask, }, 5550 { NULL, }, 5551}; 5552 5553static struct sched_domain_topology_level *sched_domain_topology = default_topology; 5554 5555#ifdef CONFIG_NUMA 5556 5557static int sched_domains_numa_levels; 5558static int *sched_domains_numa_distance; 5559static struct cpumask ***sched_domains_numa_masks; 5560static int sched_domains_curr_level; 5561 5562static inline int sd_local_flags(int level) 5563{ 5564 if (sched_domains_numa_distance[level] > RECLAIM_DISTANCE) 5565 return 0; 5566 5567 return SD_BALANCE_EXEC | SD_BALANCE_FORK | SD_WAKE_AFFINE; 5568} 5569 5570static struct sched_domain * 5571sd_numa_init(struct sched_domain_topology_level *tl, int cpu) 5572{ 5573 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); 5574 int level = tl->numa_level; 5575 int sd_weight = cpumask_weight( 5576 sched_domains_numa_masks[level][cpu_to_node(cpu)]); 5577 5578 *sd = (struct sched_domain){ 5579 .min_interval = sd_weight, 5580 .max_interval = 2*sd_weight, 5581 .busy_factor = 32, 5582 .imbalance_pct = 125, 5583 .cache_nice_tries = 2, 5584 .busy_idx = 3, 5585 .idle_idx = 2, 5586 .newidle_idx = 0, 5587 .wake_idx = 0, 5588 .forkexec_idx = 0, 5589 5590 .flags = 1*SD_LOAD_BALANCE 5591 | 1*SD_BALANCE_NEWIDLE 5592 | 0*SD_BALANCE_EXEC 5593 | 0*SD_BALANCE_FORK 5594 | 0*SD_BALANCE_WAKE 5595 | 0*SD_WAKE_AFFINE 5596 | 0*SD_SHARE_CPUPOWER 5597 | 0*SD_SHARE_PKG_RESOURCES 5598 | 1*SD_SERIALIZE 5599 | 0*SD_PREFER_SIBLING 5600 | sd_local_flags(level) 5601 , 5602 .last_balance = jiffies, 5603 .balance_interval = sd_weight, 5604 }; 5605 SD_INIT_NAME(sd, NUMA); 5606 sd->private = &tl->data; 5607 5608 /* 5609 * Ugly hack to pass state to sd_numa_mask()... 5610 */ 5611 sched_domains_curr_level = tl->numa_level; 5612 5613 return sd; 5614} 5615 5616static const struct cpumask *sd_numa_mask(int cpu) 5617{ 5618 return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)]; 5619} 5620 5621static void sched_numa_warn(const char *str) 5622{ 5623 static int done = false; 5624 int i,j; 5625 5626 if (done) 5627 return; 5628 5629 done = true; 5630 5631 printk(KERN_WARNING "ERROR: %s\n\n", str); 5632 5633 for (i = 0; i < nr_node_ids; i++) { 5634 printk(KERN_WARNING " "); 5635 for (j = 0; j < nr_node_ids; j++) 5636 printk(KERN_CONT "%02d ", node_distance(i,j)); 5637 printk(KERN_CONT "\n"); 5638 } 5639 printk(KERN_WARNING "\n"); 5640} 5641 5642static bool find_numa_distance(int distance) 5643{ 5644 int i; 5645 5646 if (distance == node_distance(0, 0)) 5647 return true; 5648 5649 for (i = 0; i < sched_domains_numa_levels; i++) { 5650 if (sched_domains_numa_distance[i] == distance) 5651 return true; 5652 } 5653 5654 return false; 5655} 5656 5657static void sched_init_numa(void) 5658{ 5659 int next_distance, curr_distance = node_distance(0, 0); 5660 struct sched_domain_topology_level *tl; 5661 int level = 0; 5662 int i, j, k; 5663 5664 sched_domains_numa_distance = kzalloc(sizeof(int) * nr_node_ids, GFP_KERNEL); 5665 if (!sched_domains_numa_distance) 5666 return; 5667 5668 /* 5669 * O(nr_nodes^2) deduplicating selection sort -- in order to find the 5670 * unique distances in the node_distance() table. 5671 * 5672 * Assumes node_distance(0,j) includes all distances in 5673 * node_distance(i,j) in order to avoid cubic time. 5674 */ 5675 next_distance = curr_distance; 5676 for (i = 0; i < nr_node_ids; i++) { 5677 for (j = 0; j < nr_node_ids; j++) { 5678 for (k = 0; k < nr_node_ids; k++) { 5679 int distance = node_distance(i, k); 5680 5681 if (distance > curr_distance && 5682 (distance < next_distance || 5683 next_distance == curr_distance)) 5684 next_distance = distance; 5685 5686 /* 5687 * While not a strong assumption it would be nice to know 5688 * about cases where if node A is connected to B, B is not 5689 * equally connected to A. 5690 */ 5691 if (sched_debug() && node_distance(k, i) != distance) 5692 sched_numa_warn("Node-distance not symmetric"); 5693 5694 if (sched_debug() && i && !find_numa_distance(distance)) 5695 sched_numa_warn("Node-0 not representative"); 5696 } 5697 if (next_distance != curr_distance) { 5698 sched_domains_numa_distance[level++] = next_distance; 5699 sched_domains_numa_levels = level; 5700 curr_distance = next_distance; 5701 } else break; 5702 } 5703 5704 /* 5705 * In case of sched_debug() we verify the above assumption. 5706 */ 5707 if (!sched_debug()) 5708 break; 5709 } 5710 /* 5711 * 'level' contains the number of unique distances, excluding the 5712 * identity distance node_distance(i,i). 5713 * 5714 * The sched_domains_numa_distance[] array includes the actual distance 5715 * numbers. 5716 */ 5717 5718 /* 5719 * Here, we should temporarily reset sched_domains_numa_levels to 0. 5720 * If it fails to allocate memory for array sched_domains_numa_masks[][], 5721 * the array will contain less then 'level' members. This could be 5722 * dangerous when we use it to iterate array sched_domains_numa_masks[][] 5723 * in other functions. 5724 * 5725 * We reset it to 'level' at the end of this function. 5726 */ 5727 sched_domains_numa_levels = 0; 5728 5729 sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL); 5730 if (!sched_domains_numa_masks) 5731 return; 5732 5733 /* 5734 * Now for each level, construct a mask per node which contains all 5735 * cpus of nodes that are that many hops away from us. 5736 */ 5737 for (i = 0; i < level; i++) { 5738 sched_domains_numa_masks[i] = 5739 kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL); 5740 if (!sched_domains_numa_masks[i]) 5741 return; 5742 5743 for (j = 0; j < nr_node_ids; j++) { 5744 struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL); 5745 if (!mask) 5746 return; 5747 5748 sched_domains_numa_masks[i][j] = mask; 5749 5750 for (k = 0; k < nr_node_ids; k++) { 5751 if (node_distance(j, k) > sched_domains_numa_distance[i]) 5752 continue; 5753 5754 cpumask_or(mask, mask, cpumask_of_node(k)); 5755 } 5756 } 5757 } 5758 5759 tl = kzalloc((ARRAY_SIZE(default_topology) + level) * 5760 sizeof(struct sched_domain_topology_level), GFP_KERNEL); 5761 if (!tl) 5762 return; 5763 5764 /* 5765 * Copy the default topology bits.. 5766 */ 5767 for (i = 0; default_topology[i].init; i++) 5768 tl[i] = default_topology[i]; 5769 5770 /* 5771 * .. and append 'j' levels of NUMA goodness. 5772 */ 5773 for (j = 0; j < level; i++, j++) { 5774 tl[i] = (struct sched_domain_topology_level){ 5775 .init = sd_numa_init, 5776 .mask = sd_numa_mask, 5777 .flags = SDTL_OVERLAP, 5778 .numa_level = j, 5779 }; 5780 } 5781 5782 sched_domain_topology = tl; 5783 5784 sched_domains_numa_levels = level; 5785} 5786 5787static void sched_domains_numa_masks_set(int cpu) 5788{ 5789 int i, j; 5790 int node = cpu_to_node(cpu); 5791 5792 for (i = 0; i < sched_domains_numa_levels; i++) { 5793 for (j = 0; j < nr_node_ids; j++) { 5794 if (node_distance(j, node) <= sched_domains_numa_distance[i]) 5795 cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]); 5796 } 5797 } 5798} 5799 5800static void sched_domains_numa_masks_clear(int cpu) 5801{ 5802 int i, j; 5803 for (i = 0; i < sched_domains_numa_levels; i++) { 5804 for (j = 0; j < nr_node_ids; j++) 5805 cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]); 5806 } 5807} 5808 5809/* 5810 * Update sched_domains_numa_masks[level][node] array when new cpus 5811 * are onlined. 5812 */ 5813static int sched_domains_numa_masks_update(struct notifier_block *nfb, 5814 unsigned long action, 5815 void *hcpu) 5816{ 5817 int cpu = (long)hcpu; 5818 5819 switch (action & ~CPU_TASKS_FROZEN) { 5820 case CPU_ONLINE: 5821 sched_domains_numa_masks_set(cpu); 5822 break; 5823 5824 case CPU_DEAD: 5825 sched_domains_numa_masks_clear(cpu); 5826 break; 5827 5828 default: 5829 return NOTIFY_DONE; 5830 } 5831 5832 return NOTIFY_OK; 5833} 5834#else 5835static inline void sched_init_numa(void) 5836{ 5837} 5838 5839static int sched_domains_numa_masks_update(struct notifier_block *nfb, 5840 unsigned long action, 5841 void *hcpu) 5842{ 5843 return 0; 5844} 5845#endif /* CONFIG_NUMA */ 5846 5847static int __sdt_alloc(const struct cpumask *cpu_map) 5848{ 5849 struct sched_domain_topology_level *tl; 5850 int j; 5851 5852 for (tl = sched_domain_topology; tl->init; tl++) { 5853 struct sd_data *sdd = &tl->data; 5854 5855 sdd->sd = alloc_percpu(struct sched_domain *); 5856 if (!sdd->sd) 5857 return -ENOMEM; 5858 5859 sdd->sg = alloc_percpu(struct sched_group *); 5860 if (!sdd->sg) 5861 return -ENOMEM; 5862 5863 sdd->sgp = alloc_percpu(struct sched_group_power *); 5864 if (!sdd->sgp) 5865 return -ENOMEM; 5866 5867 for_each_cpu(j, cpu_map) { 5868 struct sched_domain *sd; 5869 struct sched_group *sg; 5870 struct sched_group_power *sgp; 5871 5872 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(), 5873 GFP_KERNEL, cpu_to_node(j)); 5874 if (!sd) 5875 return -ENOMEM; 5876 5877 *per_cpu_ptr(sdd->sd, j) = sd; 5878 5879 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(), 5880 GFP_KERNEL, cpu_to_node(j)); 5881 if (!sg) 5882 return -ENOMEM; 5883 5884 sg->next = sg; 5885 5886 *per_cpu_ptr(sdd->sg, j) = sg; 5887 5888 sgp = kzalloc_node(sizeof(struct sched_group_power) + cpumask_size(), 5889 GFP_KERNEL, cpu_to_node(j)); 5890 if (!sgp) 5891 return -ENOMEM; 5892 5893 *per_cpu_ptr(sdd->sgp, j) = sgp; 5894 } 5895 } 5896 5897 return 0; 5898} 5899 5900static void __sdt_free(const struct cpumask *cpu_map) 5901{ 5902 struct sched_domain_topology_level *tl; 5903 int j; 5904 5905 for (tl = sched_domain_topology; tl->init; tl++) { 5906 struct sd_data *sdd = &tl->data; 5907 5908 for_each_cpu(j, cpu_map) { 5909 struct sched_domain *sd; 5910 5911 if (sdd->sd) { 5912 sd = *per_cpu_ptr(sdd->sd, j); 5913 if (sd && (sd->flags & SD_OVERLAP)) 5914 free_sched_groups(sd->groups, 0); 5915 kfree(*per_cpu_ptr(sdd->sd, j)); 5916 } 5917 5918 if (sdd->sg) 5919 kfree(*per_cpu_ptr(sdd->sg, j)); 5920 if (sdd->sgp) 5921 kfree(*per_cpu_ptr(sdd->sgp, j)); 5922 } 5923 free_percpu(sdd->sd); 5924 sdd->sd = NULL; 5925 free_percpu(sdd->sg); 5926 sdd->sg = NULL; 5927 free_percpu(sdd->sgp); 5928 sdd->sgp = NULL; 5929 } 5930} 5931 5932struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl, 5933 struct s_data *d, const struct cpumask *cpu_map, 5934 struct sched_domain_attr *attr, struct sched_domain *child, 5935 int cpu) 5936{ 5937 struct sched_domain *sd = tl->init(tl, cpu); 5938 if (!sd) 5939 return child; 5940 5941 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu)); 5942 if (child) { 5943 sd->level = child->level + 1; 5944 sched_domain_level_max = max(sched_domain_level_max, sd->level); 5945 child->parent = sd; 5946 } 5947 sd->child = child; 5948 set_domain_attribute(sd, attr); 5949 5950 return sd; 5951} 5952 5953/* 5954 * Build sched domains for a given set of cpus and attach the sched domains 5955 * to the individual cpus 5956 */ 5957static int build_sched_domains(const struct cpumask *cpu_map, 5958 struct sched_domain_attr *attr) 5959{ 5960 enum s_alloc alloc_state = sa_none; 5961 struct sched_domain *sd; 5962 struct s_data d; 5963 int i, ret = -ENOMEM; 5964 5965 alloc_state = __visit_domain_allocation_hell(&d, cpu_map); 5966 if (alloc_state != sa_rootdomain) 5967 goto error; 5968 5969 /* Set up domains for cpus specified by the cpu_map. */ 5970 for_each_cpu(i, cpu_map) { 5971 struct sched_domain_topology_level *tl; 5972 5973 sd = NULL; 5974 for (tl = sched_domain_topology; tl->init; tl++) { 5975 sd = build_sched_domain(tl, &d, cpu_map, attr, sd, i); 5976 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP)) 5977 sd->flags |= SD_OVERLAP; 5978 if (cpumask_equal(cpu_map, sched_domain_span(sd))) 5979 break; 5980 } 5981 5982 while (sd->child) 5983 sd = sd->child; 5984 5985 *per_cpu_ptr(d.sd, i) = sd; 5986 } 5987 5988 /* Build the groups for the domains */ 5989 for_each_cpu(i, cpu_map) { 5990 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) { 5991 sd->span_weight = cpumask_weight(sched_domain_span(sd)); 5992 if (sd->flags & SD_OVERLAP) { 5993 if (build_overlap_sched_groups(sd, i)) 5994 goto error; 5995 } else { 5996 if (build_sched_groups(sd, i)) 5997 goto error; 5998 } 5999 } 6000 } 6001 6002 /* Calculate CPU power for physical packages and nodes */ 6003 for (i = nr_cpumask_bits-1; i >= 0; i--) { 6004 if (!cpumask_test_cpu(i, cpu_map)) 6005 continue; 6006 6007 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) { 6008 claim_allocations(i, sd); 6009 init_sched_groups_power(i, sd); 6010 } 6011 } 6012 6013 /* Attach the domains */ 6014 rcu_read_lock(); 6015 for_each_cpu(i, cpu_map) { 6016 sd = *per_cpu_ptr(d.sd, i); 6017 cpu_attach_domain(sd, d.rd, i); 6018 } 6019 rcu_read_unlock(); 6020 6021 ret = 0; 6022error: 6023 __free_domain_allocs(&d, alloc_state, cpu_map); 6024 return ret; 6025} 6026 6027static cpumask_var_t *doms_cur; /* current sched domains */ 6028static int ndoms_cur; /* number of sched domains in 'doms_cur' */ 6029static struct sched_domain_attr *dattr_cur; 6030 /* attribues of custom domains in 'doms_cur' */ 6031 6032/* 6033 * Special case: If a kmalloc of a doms_cur partition (array of 6034 * cpumask) fails, then fallback to a single sched domain, 6035 * as determined by the single cpumask fallback_doms. 6036 */ 6037static cpumask_var_t fallback_doms; 6038 6039/* 6040 * arch_update_cpu_topology lets virtualized architectures update the 6041 * cpu core maps. It is supposed to return 1 if the topology changed 6042 * or 0 if it stayed the same. 6043 */ 6044int __attribute__((weak)) arch_update_cpu_topology(void) 6045{ 6046 return 0; 6047} 6048 6049cpumask_var_t *alloc_sched_domains(unsigned int ndoms) 6050{ 6051 int i; 6052 cpumask_var_t *doms; 6053 6054 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL); 6055 if (!doms) 6056 return NULL; 6057 for (i = 0; i < ndoms; i++) { 6058 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) { 6059 free_sched_domains(doms, i); 6060 return NULL; 6061 } 6062 } 6063 return doms; 6064} 6065 6066void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms) 6067{ 6068 unsigned int i; 6069 for (i = 0; i < ndoms; i++) 6070 free_cpumask_var(doms[i]); 6071 kfree(doms); 6072} 6073 6074/* 6075 * Set up scheduler domains and groups. Callers must hold the hotplug lock. 6076 * For now this just excludes isolated cpus, but could be used to 6077 * exclude other special cases in the future. 6078 */ 6079static int init_sched_domains(const struct cpumask *cpu_map) 6080{ 6081 int err; 6082 6083 arch_update_cpu_topology(); 6084 ndoms_cur = 1; 6085 doms_cur = alloc_sched_domains(ndoms_cur); 6086 if (!doms_cur) 6087 doms_cur = &fallback_doms; 6088 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map); 6089 err = build_sched_domains(doms_cur[0], NULL); 6090 register_sched_domain_sysctl(); 6091 6092 return err; 6093} 6094 6095/* 6096 * Detach sched domains from a group of cpus specified in cpu_map 6097 * These cpus will now be attached to the NULL domain 6098 */ 6099static void detach_destroy_domains(const struct cpumask *cpu_map) 6100{ 6101 int i; 6102 6103 rcu_read_lock(); 6104 for_each_cpu(i, cpu_map) 6105 cpu_attach_domain(NULL, &def_root_domain, i); 6106 rcu_read_unlock(); 6107} 6108 6109/* handle null as "default" */ 6110static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur, 6111 struct sched_domain_attr *new, int idx_new) 6112{ 6113 struct sched_domain_attr tmp; 6114 6115 /* fast path */ 6116 if (!new && !cur) 6117 return 1; 6118 6119 tmp = SD_ATTR_INIT; 6120 return !memcmp(cur ? (cur + idx_cur) : &tmp, 6121 new ? (new + idx_new) : &tmp, 6122 sizeof(struct sched_domain_attr)); 6123} 6124 6125/* 6126 * Partition sched domains as specified by the 'ndoms_new' 6127 * cpumasks in the array doms_new[] of cpumasks. This compares 6128 * doms_new[] to the current sched domain partitioning, doms_cur[]. 6129 * It destroys each deleted domain and builds each new domain. 6130 * 6131 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'. 6132 * The masks don't intersect (don't overlap.) We should setup one 6133 * sched domain for each mask. CPUs not in any of the cpumasks will 6134 * not be load balanced. If the same cpumask appears both in the 6135 * current 'doms_cur' domains and in the new 'doms_new', we can leave 6136 * it as it is. 6137 * 6138 * The passed in 'doms_new' should be allocated using 6139 * alloc_sched_domains. This routine takes ownership of it and will 6140 * free_sched_domains it when done with it. If the caller failed the 6141 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1, 6142 * and partition_sched_domains() will fallback to the single partition 6143 * 'fallback_doms', it also forces the domains to be rebuilt. 6144 * 6145 * If doms_new == NULL it will be replaced with cpu_online_mask. 6146 * ndoms_new == 0 is a special case for destroying existing domains, 6147 * and it will not create the default domain. 6148 * 6149 * Call with hotplug lock held 6150 */ 6151void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[], 6152 struct sched_domain_attr *dattr_new) 6153{ 6154 int i, j, n; 6155 int new_topology; 6156 6157 mutex_lock(&sched_domains_mutex); 6158 6159 /* always unregister in case we don't destroy any domains */ 6160 unregister_sched_domain_sysctl(); 6161 6162 /* Let architecture update cpu core mappings. */ 6163 new_topology = arch_update_cpu_topology(); 6164 6165 n = doms_new ? ndoms_new : 0; 6166 6167 /* Destroy deleted domains */ 6168 for (i = 0; i < ndoms_cur; i++) { 6169 for (j = 0; j < n && !new_topology; j++) { 6170 if (cpumask_equal(doms_cur[i], doms_new[j]) 6171 && dattrs_equal(dattr_cur, i, dattr_new, j)) 6172 goto match1; 6173 } 6174 /* no match - a current sched domain not in new doms_new[] */ 6175 detach_destroy_domains(doms_cur[i]); 6176match1: 6177 ; 6178 } 6179 6180 if (doms_new == NULL) { 6181 ndoms_cur = 0; 6182 doms_new = &fallback_doms; 6183 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map); 6184 WARN_ON_ONCE(dattr_new); 6185 } 6186 6187 /* Build new domains */ 6188 for (i = 0; i < ndoms_new; i++) { 6189 for (j = 0; j < ndoms_cur && !new_topology; j++) { 6190 if (cpumask_equal(doms_new[i], doms_cur[j]) 6191 && dattrs_equal(dattr_new, i, dattr_cur, j)) 6192 goto match2; 6193 } 6194 /* no match - add a new doms_new */ 6195 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL); 6196match2: 6197 ; 6198 } 6199 6200 /* Remember the new sched domains */ 6201 if (doms_cur != &fallback_doms) 6202 free_sched_domains(doms_cur, ndoms_cur); 6203 kfree(dattr_cur); /* kfree(NULL) is safe */ 6204 doms_cur = doms_new; 6205 dattr_cur = dattr_new; 6206 ndoms_cur = ndoms_new; 6207 6208 register_sched_domain_sysctl(); 6209 6210 mutex_unlock(&sched_domains_mutex); 6211} 6212 6213static int num_cpus_frozen; /* used to mark begin/end of suspend/resume */ 6214 6215/* 6216 * Update cpusets according to cpu_active mask. If cpusets are 6217 * disabled, cpuset_update_active_cpus() becomes a simple wrapper 6218 * around partition_sched_domains(). 6219 * 6220 * If we come here as part of a suspend/resume, don't touch cpusets because we 6221 * want to restore it back to its original state upon resume anyway. 6222 */ 6223static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action, 6224 void *hcpu) 6225{ 6226 switch (action) { 6227 case CPU_ONLINE_FROZEN: 6228 case CPU_DOWN_FAILED_FROZEN: 6229 6230 /* 6231 * num_cpus_frozen tracks how many CPUs are involved in suspend 6232 * resume sequence. As long as this is not the last online 6233 * operation in the resume sequence, just build a single sched 6234 * domain, ignoring cpusets. 6235 */ 6236 num_cpus_frozen--; 6237 if (likely(num_cpus_frozen)) { 6238 partition_sched_domains(1, NULL, NULL); 6239 break; 6240 } 6241 6242 /* 6243 * This is the last CPU online operation. So fall through and 6244 * restore the original sched domains by considering the 6245 * cpuset configurations. 6246 */ 6247 6248 case CPU_ONLINE: 6249 case CPU_DOWN_FAILED: 6250 cpuset_update_active_cpus(true); 6251 break; 6252 default: 6253 return NOTIFY_DONE; 6254 } 6255 return NOTIFY_OK; 6256} 6257 6258static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action, 6259 void *hcpu) 6260{ 6261 switch (action) { 6262 case CPU_DOWN_PREPARE: 6263 cpuset_update_active_cpus(false); 6264 break; 6265 case CPU_DOWN_PREPARE_FROZEN: 6266 num_cpus_frozen++; 6267 partition_sched_domains(1, NULL, NULL); 6268 break; 6269 default: 6270 return NOTIFY_DONE; 6271 } 6272 return NOTIFY_OK; 6273} 6274 6275void __init sched_init_smp(void) 6276{ 6277 cpumask_var_t non_isolated_cpus; 6278 6279 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL); 6280 alloc_cpumask_var(&fallback_doms, GFP_KERNEL); 6281 6282 sched_init_numa(); 6283 6284 get_online_cpus(); 6285 mutex_lock(&sched_domains_mutex); 6286 init_sched_domains(cpu_active_mask); 6287 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map); 6288 if (cpumask_empty(non_isolated_cpus)) 6289 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus); 6290 mutex_unlock(&sched_domains_mutex); 6291 put_online_cpus(); 6292 6293 hotcpu_notifier(sched_domains_numa_masks_update, CPU_PRI_SCHED_ACTIVE); 6294 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE); 6295 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE); 6296 6297 init_hrtick(); 6298 6299 /* Move init over to a non-isolated CPU */ 6300 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0) 6301 BUG(); 6302 sched_init_granularity(); 6303 free_cpumask_var(non_isolated_cpus); 6304 6305 init_sched_rt_class(); 6306} 6307#else 6308void __init sched_init_smp(void) 6309{ 6310 sched_init_granularity(); 6311} 6312#endif /* CONFIG_SMP */ 6313 6314const_debug unsigned int sysctl_timer_migration = 1; 6315 6316int in_sched_functions(unsigned long addr) 6317{ 6318 return in_lock_functions(addr) || 6319 (addr >= (unsigned long)__sched_text_start 6320 && addr < (unsigned long)__sched_text_end); 6321} 6322 6323#ifdef CONFIG_CGROUP_SCHED 6324/* 6325 * Default task group. 6326 * Every task in system belongs to this group at bootup. 6327 */ 6328struct task_group root_task_group; 6329LIST_HEAD(task_groups); 6330#endif 6331 6332DECLARE_PER_CPU(cpumask_var_t, load_balance_mask); 6333 6334void __init sched_init(void) 6335{ 6336 int i, j; 6337 unsigned long alloc_size = 0, ptr; 6338 6339#ifdef CONFIG_FAIR_GROUP_SCHED 6340 alloc_size += 2 * nr_cpu_ids * sizeof(void **); 6341#endif 6342#ifdef CONFIG_RT_GROUP_SCHED 6343 alloc_size += 2 * nr_cpu_ids * sizeof(void **); 6344#endif 6345#ifdef CONFIG_CPUMASK_OFFSTACK 6346 alloc_size += num_possible_cpus() * cpumask_size(); 6347#endif 6348 if (alloc_size) { 6349 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT); 6350 6351#ifdef CONFIG_FAIR_GROUP_SCHED 6352 root_task_group.se = (struct sched_entity **)ptr; 6353 ptr += nr_cpu_ids * sizeof(void **); 6354 6355 root_task_group.cfs_rq = (struct cfs_rq **)ptr; 6356 ptr += nr_cpu_ids * sizeof(void **); 6357 6358#endif /* CONFIG_FAIR_GROUP_SCHED */ 6359#ifdef CONFIG_RT_GROUP_SCHED 6360 root_task_group.rt_se = (struct sched_rt_entity **)ptr; 6361 ptr += nr_cpu_ids * sizeof(void **); 6362 6363 root_task_group.rt_rq = (struct rt_rq **)ptr; 6364 ptr += nr_cpu_ids * sizeof(void **); 6365 6366#endif /* CONFIG_RT_GROUP_SCHED */ 6367#ifdef CONFIG_CPUMASK_OFFSTACK 6368 for_each_possible_cpu(i) { 6369 per_cpu(load_balance_mask, i) = (void *)ptr; 6370 ptr += cpumask_size(); 6371 } 6372#endif /* CONFIG_CPUMASK_OFFSTACK */ 6373 } 6374 6375#ifdef CONFIG_SMP 6376 init_defrootdomain(); 6377#endif 6378 6379 init_rt_bandwidth(&def_rt_bandwidth, 6380 global_rt_period(), global_rt_runtime()); 6381 6382#ifdef CONFIG_RT_GROUP_SCHED 6383 init_rt_bandwidth(&root_task_group.rt_bandwidth, 6384 global_rt_period(), global_rt_runtime()); 6385#endif /* CONFIG_RT_GROUP_SCHED */ 6386 6387#ifdef CONFIG_CGROUP_SCHED 6388 list_add(&root_task_group.list, &task_groups); 6389 INIT_LIST_HEAD(&root_task_group.children); 6390 INIT_LIST_HEAD(&root_task_group.siblings); 6391 autogroup_init(&init_task); 6392 6393#endif /* CONFIG_CGROUP_SCHED */ 6394 6395 for_each_possible_cpu(i) { 6396 struct rq *rq; 6397 6398 rq = cpu_rq(i); 6399 raw_spin_lock_init(&rq->lock); 6400 rq->nr_running = 0; 6401 rq->calc_load_active = 0; 6402 rq->calc_load_update = jiffies + LOAD_FREQ; 6403 init_cfs_rq(&rq->cfs); 6404 init_rt_rq(&rq->rt, rq); 6405#ifdef CONFIG_FAIR_GROUP_SCHED 6406 root_task_group.shares = ROOT_TASK_GROUP_LOAD; 6407 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list); 6408 /* 6409 * How much cpu bandwidth does root_task_group get? 6410 * 6411 * In case of task-groups formed thr' the cgroup filesystem, it 6412 * gets 100% of the cpu resources in the system. This overall 6413 * system cpu resource is divided among the tasks of 6414 * root_task_group and its child task-groups in a fair manner, 6415 * based on each entity's (task or task-group's) weight 6416 * (se->load.weight). 6417 * 6418 * In other words, if root_task_group has 10 tasks of weight 6419 * 1024) and two child groups A0 and A1 (of weight 1024 each), 6420 * then A0's share of the cpu resource is: 6421 * 6422 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33% 6423 * 6424 * We achieve this by letting root_task_group's tasks sit 6425 * directly in rq->cfs (i.e root_task_group->se[] = NULL). 6426 */ 6427 init_cfs_bandwidth(&root_task_group.cfs_bandwidth); 6428 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL); 6429#endif /* CONFIG_FAIR_GROUP_SCHED */ 6430 6431 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime; 6432#ifdef CONFIG_RT_GROUP_SCHED 6433 INIT_LIST_HEAD(&rq->leaf_rt_rq_list); 6434 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL); 6435#endif 6436 6437 for (j = 0; j < CPU_LOAD_IDX_MAX; j++) 6438 rq->cpu_load[j] = 0; 6439 6440 rq->last_load_update_tick = jiffies; 6441 6442#ifdef CONFIG_SMP 6443 rq->sd = NULL; 6444 rq->rd = NULL; 6445 rq->cpu_power = SCHED_POWER_SCALE; 6446 rq->post_schedule = 0; 6447 rq->active_balance = 0; 6448 rq->next_balance = jiffies; 6449 rq->push_cpu = 0; 6450 rq->cpu = i; 6451 rq->online = 0; 6452 rq->idle_stamp = 0; 6453 rq->avg_idle = 2*sysctl_sched_migration_cost; 6454 6455 INIT_LIST_HEAD(&rq->cfs_tasks); 6456 6457 rq_attach_root(rq, &def_root_domain); 6458#ifdef CONFIG_NO_HZ_COMMON 6459 rq->nohz_flags = 0; 6460#endif 6461#ifdef CONFIG_NO_HZ_FULL 6462 rq->last_sched_tick = 0; 6463#endif 6464#endif 6465 init_rq_hrtick(rq); 6466 atomic_set(&rq->nr_iowait, 0); 6467 } 6468 6469 set_load_weight(&init_task); 6470 6471#ifdef CONFIG_PREEMPT_NOTIFIERS 6472 INIT_HLIST_HEAD(&init_task.preempt_notifiers); 6473#endif 6474 6475#ifdef CONFIG_RT_MUTEXES 6476 plist_head_init(&init_task.pi_waiters); 6477#endif 6478 6479 /* 6480 * The boot idle thread does lazy MMU switching as well: 6481 */ 6482 atomic_inc(&init_mm.mm_count); 6483 enter_lazy_tlb(&init_mm, current); 6484 6485 /* 6486 * Make us the idle thread. Technically, schedule() should not be 6487 * called from this thread, however somewhere below it might be, 6488 * but because we are the idle thread, we just pick up running again 6489 * when this runqueue becomes "idle". 6490 */ 6491 init_idle(current, smp_processor_id()); 6492 6493 calc_load_update = jiffies + LOAD_FREQ; 6494 6495 /* 6496 * During early bootup we pretend to be a normal task: 6497 */ 6498 current->sched_class = &fair_sched_class; 6499 6500#ifdef CONFIG_SMP 6501 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT); 6502 /* May be allocated at isolcpus cmdline parse time */ 6503 if (cpu_isolated_map == NULL) 6504 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT); 6505 idle_thread_set_boot_cpu(); 6506#endif 6507 init_sched_fair_class(); 6508 6509 scheduler_running = 1; 6510} 6511 6512#ifdef CONFIG_DEBUG_ATOMIC_SLEEP 6513static inline int preempt_count_equals(int preempt_offset) 6514{ 6515 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth(); 6516 6517 return (nested == preempt_offset); 6518} 6519 6520void __might_sleep(const char *file, int line, int preempt_offset) 6521{ 6522 static unsigned long prev_jiffy; /* ratelimiting */ 6523 6524 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */ 6525 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) || 6526 system_state != SYSTEM_RUNNING || oops_in_progress) 6527 return; 6528 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy) 6529 return; 6530 prev_jiffy = jiffies; 6531 6532 printk(KERN_ERR 6533 "BUG: sleeping function called from invalid context at %s:%d\n", 6534 file, line); 6535 printk(KERN_ERR 6536 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n", 6537 in_atomic(), irqs_disabled(), 6538 current->pid, current->comm); 6539 6540 debug_show_held_locks(current); 6541 if (irqs_disabled()) 6542 print_irqtrace_events(current); 6543 dump_stack(); 6544} 6545EXPORT_SYMBOL(__might_sleep); 6546#endif 6547 6548#ifdef CONFIG_MAGIC_SYSRQ 6549static void normalize_task(struct rq *rq, struct task_struct *p) 6550{ 6551 const struct sched_class *prev_class = p->sched_class; 6552 int old_prio = p->prio; 6553 int on_rq; 6554 6555 on_rq = p->on_rq; 6556 if (on_rq) 6557 dequeue_task(rq, p, 0); 6558 __setscheduler(rq, p, SCHED_NORMAL, 0); 6559 if (on_rq) { 6560 enqueue_task(rq, p, 0); 6561 resched_task(rq->curr); 6562 } 6563 6564 check_class_changed(rq, p, prev_class, old_prio); 6565} 6566 6567void normalize_rt_tasks(void) 6568{ 6569 struct task_struct *g, *p; 6570 unsigned long flags; 6571 struct rq *rq; 6572 6573 read_lock_irqsave(&tasklist_lock, flags); 6574 do_each_thread(g, p) { 6575 /* 6576 * Only normalize user tasks: 6577 */ 6578 if (!p->mm) 6579 continue; 6580 6581 p->se.exec_start = 0; 6582#ifdef CONFIG_SCHEDSTATS 6583 p->se.statistics.wait_start = 0; 6584 p->se.statistics.sleep_start = 0; 6585 p->se.statistics.block_start = 0; 6586#endif 6587 6588 if (!rt_task(p)) { 6589 /* 6590 * Renice negative nice level userspace 6591 * tasks back to 0: 6592 */ 6593 if (TASK_NICE(p) < 0 && p->mm) 6594 set_user_nice(p, 0); 6595 continue; 6596 } 6597 6598 raw_spin_lock(&p->pi_lock); 6599 rq = __task_rq_lock(p); 6600 6601 normalize_task(rq, p); 6602 6603 __task_rq_unlock(rq); 6604 raw_spin_unlock(&p->pi_lock); 6605 } while_each_thread(g, p); 6606 6607 read_unlock_irqrestore(&tasklist_lock, flags); 6608} 6609 6610#endif /* CONFIG_MAGIC_SYSRQ */ 6611 6612#if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) 6613/* 6614 * These functions are only useful for the IA64 MCA handling, or kdb. 6615 * 6616 * They can only be called when the whole system has been 6617 * stopped - every CPU needs to be quiescent, and no scheduling 6618 * activity can take place. Using them for anything else would 6619 * be a serious bug, and as a result, they aren't even visible 6620 * under any other configuration. 6621 */ 6622 6623/** 6624 * curr_task - return the current task for a given cpu. 6625 * @cpu: the processor in question. 6626 * 6627 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED! 6628 */ 6629struct task_struct *curr_task(int cpu) 6630{ 6631 return cpu_curr(cpu); 6632} 6633 6634#endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */ 6635 6636#ifdef CONFIG_IA64 6637/** 6638 * set_curr_task - set the current task for a given cpu. 6639 * @cpu: the processor in question. 6640 * @p: the task pointer to set. 6641 * 6642 * Description: This function must only be used when non-maskable interrupts 6643 * are serviced on a separate stack. It allows the architecture to switch the 6644 * notion of the current task on a cpu in a non-blocking manner. This function 6645 * must be called with all CPU's synchronized, and interrupts disabled, the 6646 * and caller must save the original value of the current task (see 6647 * curr_task() above) and restore that value before reenabling interrupts and 6648 * re-starting the system. 6649 * 6650 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED! 6651 */ 6652void set_curr_task(int cpu, struct task_struct *p) 6653{ 6654 cpu_curr(cpu) = p; 6655} 6656 6657#endif 6658 6659#ifdef CONFIG_CGROUP_SCHED 6660/* task_group_lock serializes the addition/removal of task groups */ 6661static DEFINE_SPINLOCK(task_group_lock); 6662 6663static void free_sched_group(struct task_group *tg) 6664{ 6665 free_fair_sched_group(tg); 6666 free_rt_sched_group(tg); 6667 autogroup_free(tg); 6668 kfree(tg); 6669} 6670 6671/* allocate runqueue etc for a new task group */ 6672struct task_group *sched_create_group(struct task_group *parent) 6673{ 6674 struct task_group *tg; 6675 6676 tg = kzalloc(sizeof(*tg), GFP_KERNEL); 6677 if (!tg) 6678 return ERR_PTR(-ENOMEM); 6679 6680 if (!alloc_fair_sched_group(tg, parent)) 6681 goto err; 6682 6683 if (!alloc_rt_sched_group(tg, parent)) 6684 goto err; 6685 6686 return tg; 6687 6688err: 6689 free_sched_group(tg); 6690 return ERR_PTR(-ENOMEM); 6691} 6692 6693void sched_online_group(struct task_group *tg, struct task_group *parent) 6694{ 6695 unsigned long flags; 6696 6697 spin_lock_irqsave(&task_group_lock, flags); 6698 list_add_rcu(&tg->list, &task_groups); 6699 6700 WARN_ON(!parent); /* root should already exist */ 6701 6702 tg->parent = parent; 6703 INIT_LIST_HEAD(&tg->children); 6704 list_add_rcu(&tg->siblings, &parent->children); 6705 spin_unlock_irqrestore(&task_group_lock, flags); 6706} 6707 6708/* rcu callback to free various structures associated with a task group */ 6709static void free_sched_group_rcu(struct rcu_head *rhp) 6710{ 6711 /* now it should be safe to free those cfs_rqs */ 6712 free_sched_group(container_of(rhp, struct task_group, rcu)); 6713} 6714 6715/* Destroy runqueue etc associated with a task group */ 6716void sched_destroy_group(struct task_group *tg) 6717{ 6718 /* wait for possible concurrent references to cfs_rqs complete */ 6719 call_rcu(&tg->rcu, free_sched_group_rcu); 6720} 6721 6722void sched_offline_group(struct task_group *tg) 6723{ 6724 unsigned long flags; 6725 int i; 6726 6727 /* end participation in shares distribution */ 6728 for_each_possible_cpu(i) 6729 unregister_fair_sched_group(tg, i); 6730 6731 spin_lock_irqsave(&task_group_lock, flags); 6732 list_del_rcu(&tg->list); 6733 list_del_rcu(&tg->siblings); 6734 spin_unlock_irqrestore(&task_group_lock, flags); 6735} 6736 6737/* change task's runqueue when it moves between groups. 6738 * The caller of this function should have put the task in its new group 6739 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to 6740 * reflect its new group. 6741 */ 6742void sched_move_task(struct task_struct *tsk) 6743{ 6744 struct task_group *tg; 6745 int on_rq, running; 6746 unsigned long flags; 6747 struct rq *rq; 6748 6749 rq = task_rq_lock(tsk, &flags); 6750 6751 running = task_current(rq, tsk); 6752 on_rq = tsk->on_rq; 6753 6754 if (on_rq) 6755 dequeue_task(rq, tsk, 0); 6756 if (unlikely(running)) 6757 tsk->sched_class->put_prev_task(rq, tsk); 6758 6759 tg = container_of(task_subsys_state_check(tsk, cpu_cgroup_subsys_id, 6760 lockdep_is_held(&tsk->sighand->siglock)), 6761 struct task_group, css); 6762 tg = autogroup_task_group(tsk, tg); 6763 tsk->sched_task_group = tg; 6764 6765#ifdef CONFIG_FAIR_GROUP_SCHED 6766 if (tsk->sched_class->task_move_group) 6767 tsk->sched_class->task_move_group(tsk, on_rq); 6768 else 6769#endif 6770 set_task_rq(tsk, task_cpu(tsk)); 6771 6772 if (unlikely(running)) 6773 tsk->sched_class->set_curr_task(rq); 6774 if (on_rq) 6775 enqueue_task(rq, tsk, 0); 6776 6777 task_rq_unlock(rq, tsk, &flags); 6778} 6779#endif /* CONFIG_CGROUP_SCHED */ 6780 6781#if defined(CONFIG_RT_GROUP_SCHED) || defined(CONFIG_CFS_BANDWIDTH) 6782static unsigned long to_ratio(u64 period, u64 runtime) 6783{ 6784 if (runtime == RUNTIME_INF) 6785 return 1ULL << 20; 6786 6787 return div64_u64(runtime << 20, period); 6788} 6789#endif 6790 6791#ifdef CONFIG_RT_GROUP_SCHED 6792/* 6793 * Ensure that the real time constraints are schedulable. 6794 */ 6795static DEFINE_MUTEX(rt_constraints_mutex); 6796 6797/* Must be called with tasklist_lock held */ 6798static inline int tg_has_rt_tasks(struct task_group *tg) 6799{ 6800 struct task_struct *g, *p; 6801 6802 do_each_thread(g, p) { 6803 if (rt_task(p) && task_rq(p)->rt.tg == tg) 6804 return 1; 6805 } while_each_thread(g, p); 6806 6807 return 0; 6808} 6809 6810struct rt_schedulable_data { 6811 struct task_group *tg; 6812 u64 rt_period; 6813 u64 rt_runtime; 6814}; 6815 6816static int tg_rt_schedulable(struct task_group *tg, void *data) 6817{ 6818 struct rt_schedulable_data *d = data; 6819 struct task_group *child; 6820 unsigned long total, sum = 0; 6821 u64 period, runtime; 6822 6823 period = ktime_to_ns(tg->rt_bandwidth.rt_period); 6824 runtime = tg->rt_bandwidth.rt_runtime; 6825 6826 if (tg == d->tg) { 6827 period = d->rt_period; 6828 runtime = d->rt_runtime; 6829 } 6830 6831 /* 6832 * Cannot have more runtime than the period. 6833 */ 6834 if (runtime > period && runtime != RUNTIME_INF) 6835 return -EINVAL; 6836 6837 /* 6838 * Ensure we don't starve existing RT tasks. 6839 */ 6840 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg)) 6841 return -EBUSY; 6842 6843 total = to_ratio(period, runtime); 6844 6845 /* 6846 * Nobody can have more than the global setting allows. 6847 */ 6848 if (total > to_ratio(global_rt_period(), global_rt_runtime())) 6849 return -EINVAL; 6850 6851 /* 6852 * The sum of our children's runtime should not exceed our own. 6853 */ 6854 list_for_each_entry_rcu(child, &tg->children, siblings) { 6855 period = ktime_to_ns(child->rt_bandwidth.rt_period); 6856 runtime = child->rt_bandwidth.rt_runtime; 6857 6858 if (child == d->tg) { 6859 period = d->rt_period; 6860 runtime = d->rt_runtime; 6861 } 6862 6863 sum += to_ratio(period, runtime); 6864 } 6865 6866 if (sum > total) 6867 return -EINVAL; 6868 6869 return 0; 6870} 6871 6872static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime) 6873{ 6874 int ret; 6875 6876 struct rt_schedulable_data data = { 6877 .tg = tg, 6878 .rt_period = period, 6879 .rt_runtime = runtime, 6880 }; 6881 6882 rcu_read_lock(); 6883 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data); 6884 rcu_read_unlock(); 6885 6886 return ret; 6887} 6888 6889static int tg_set_rt_bandwidth(struct task_group *tg, 6890 u64 rt_period, u64 rt_runtime) 6891{ 6892 int i, err = 0; 6893 6894 mutex_lock(&rt_constraints_mutex); 6895 read_lock(&tasklist_lock); 6896 err = __rt_schedulable(tg, rt_period, rt_runtime); 6897 if (err) 6898 goto unlock; 6899 6900 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock); 6901 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period); 6902 tg->rt_bandwidth.rt_runtime = rt_runtime; 6903 6904 for_each_possible_cpu(i) { 6905 struct rt_rq *rt_rq = tg->rt_rq[i]; 6906 6907 raw_spin_lock(&rt_rq->rt_runtime_lock); 6908 rt_rq->rt_runtime = rt_runtime; 6909 raw_spin_unlock(&rt_rq->rt_runtime_lock); 6910 } 6911 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock); 6912unlock: 6913 read_unlock(&tasklist_lock); 6914 mutex_unlock(&rt_constraints_mutex); 6915 6916 return err; 6917} 6918 6919static int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us) 6920{ 6921 u64 rt_runtime, rt_period; 6922 6923 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period); 6924 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC; 6925 if (rt_runtime_us < 0) 6926 rt_runtime = RUNTIME_INF; 6927 6928 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime); 6929} 6930 6931static long sched_group_rt_runtime(struct task_group *tg) 6932{ 6933 u64 rt_runtime_us; 6934 6935 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF) 6936 return -1; 6937 6938 rt_runtime_us = tg->rt_bandwidth.rt_runtime; 6939 do_div(rt_runtime_us, NSEC_PER_USEC); 6940 return rt_runtime_us; 6941} 6942 6943static int sched_group_set_rt_period(struct task_group *tg, long rt_period_us) 6944{ 6945 u64 rt_runtime, rt_period; 6946 6947 rt_period = (u64)rt_period_us * NSEC_PER_USEC; 6948 rt_runtime = tg->rt_bandwidth.rt_runtime; 6949 6950 if (rt_period == 0) 6951 return -EINVAL; 6952 6953 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime); 6954} 6955 6956static long sched_group_rt_period(struct task_group *tg) 6957{ 6958 u64 rt_period_us; 6959 6960 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period); 6961 do_div(rt_period_us, NSEC_PER_USEC); 6962 return rt_period_us; 6963} 6964 6965static int sched_rt_global_constraints(void) 6966{ 6967 u64 runtime, period; 6968 int ret = 0; 6969 6970 if (sysctl_sched_rt_period <= 0) 6971 return -EINVAL; 6972 6973 runtime = global_rt_runtime(); 6974 period = global_rt_period(); 6975 6976 /* 6977 * Sanity check on the sysctl variables. 6978 */ 6979 if (runtime > period && runtime != RUNTIME_INF) 6980 return -EINVAL; 6981 6982 mutex_lock(&rt_constraints_mutex); 6983 read_lock(&tasklist_lock); 6984 ret = __rt_schedulable(NULL, 0, 0); 6985 read_unlock(&tasklist_lock); 6986 mutex_unlock(&rt_constraints_mutex); 6987 6988 return ret; 6989} 6990 6991static int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk) 6992{ 6993 /* Don't accept realtime tasks when there is no way for them to run */ 6994 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0) 6995 return 0; 6996 6997 return 1; 6998} 6999 7000#else /* !CONFIG_RT_GROUP_SCHED */ 7001static int sched_rt_global_constraints(void) 7002{ 7003 unsigned long flags; 7004 int i; 7005 7006 if (sysctl_sched_rt_period <= 0) 7007 return -EINVAL; 7008 7009 /* 7010 * There's always some RT tasks in the root group 7011 * -- migration, kstopmachine etc.. 7012 */ 7013 if (sysctl_sched_rt_runtime == 0) 7014 return -EBUSY; 7015 7016 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags); 7017 for_each_possible_cpu(i) { 7018 struct rt_rq *rt_rq = &cpu_rq(i)->rt; 7019 7020 raw_spin_lock(&rt_rq->rt_runtime_lock); 7021 rt_rq->rt_runtime = global_rt_runtime(); 7022 raw_spin_unlock(&rt_rq->rt_runtime_lock); 7023 } 7024 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags); 7025 7026 return 0; 7027} 7028#endif /* CONFIG_RT_GROUP_SCHED */ 7029 7030int sched_rr_handler(struct ctl_table *table, int write, 7031 void __user *buffer, size_t *lenp, 7032 loff_t *ppos) 7033{ 7034 int ret; 7035 static DEFINE_MUTEX(mutex); 7036 7037 mutex_lock(&mutex); 7038 ret = proc_dointvec(table, write, buffer, lenp, ppos); 7039 /* make sure that internally we keep jiffies */ 7040 /* also, writing zero resets timeslice to default */ 7041 if (!ret && write) { 7042 sched_rr_timeslice = sched_rr_timeslice <= 0 ? 7043 RR_TIMESLICE : msecs_to_jiffies(sched_rr_timeslice); 7044 } 7045 mutex_unlock(&mutex); 7046 return ret; 7047} 7048 7049int sched_rt_handler(struct ctl_table *table, int write, 7050 void __user *buffer, size_t *lenp, 7051 loff_t *ppos) 7052{ 7053 int ret; 7054 int old_period, old_runtime; 7055 static DEFINE_MUTEX(mutex); 7056 7057 mutex_lock(&mutex); 7058 old_period = sysctl_sched_rt_period; 7059 old_runtime = sysctl_sched_rt_runtime; 7060 7061 ret = proc_dointvec(table, write, buffer, lenp, ppos); 7062 7063 if (!ret && write) { 7064 ret = sched_rt_global_constraints(); 7065 if (ret) { 7066 sysctl_sched_rt_period = old_period; 7067 sysctl_sched_rt_runtime = old_runtime; 7068 } else { 7069 def_rt_bandwidth.rt_runtime = global_rt_runtime(); 7070 def_rt_bandwidth.rt_period = 7071 ns_to_ktime(global_rt_period()); 7072 } 7073 } 7074 mutex_unlock(&mutex); 7075 7076 return ret; 7077} 7078 7079#ifdef CONFIG_CGROUP_SCHED 7080 7081/* return corresponding task_group object of a cgroup */ 7082static inline struct task_group *cgroup_tg(struct cgroup *cgrp) 7083{ 7084 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id), 7085 struct task_group, css); 7086} 7087 7088static struct cgroup_subsys_state *cpu_cgroup_css_alloc(struct cgroup *cgrp) 7089{ 7090 struct task_group *tg, *parent; 7091 7092 if (!cgrp->parent) { 7093 /* This is early initialization for the top cgroup */ 7094 return &root_task_group.css; 7095 } 7096 7097 parent = cgroup_tg(cgrp->parent); 7098 tg = sched_create_group(parent); 7099 if (IS_ERR(tg)) 7100 return ERR_PTR(-ENOMEM); 7101 7102 return &tg->css; 7103} 7104 7105static int cpu_cgroup_css_online(struct cgroup *cgrp) 7106{ 7107 struct task_group *tg = cgroup_tg(cgrp); 7108 struct task_group *parent; 7109 7110 if (!cgrp->parent) 7111 return 0; 7112 7113 parent = cgroup_tg(cgrp->parent); 7114 sched_online_group(tg, parent); 7115 return 0; 7116} 7117 7118static void cpu_cgroup_css_free(struct cgroup *cgrp) 7119{ 7120 struct task_group *tg = cgroup_tg(cgrp); 7121 7122 sched_destroy_group(tg); 7123} 7124 7125static void cpu_cgroup_css_offline(struct cgroup *cgrp) 7126{ 7127 struct task_group *tg = cgroup_tg(cgrp); 7128 7129 sched_offline_group(tg); 7130} 7131 7132static int cpu_cgroup_can_attach(struct cgroup *cgrp, 7133 struct cgroup_taskset *tset) 7134{ 7135 struct task_struct *task; 7136 7137 cgroup_taskset_for_each(task, cgrp, tset) { 7138#ifdef CONFIG_RT_GROUP_SCHED 7139 if (!sched_rt_can_attach(cgroup_tg(cgrp), task)) 7140 return -EINVAL; 7141#else 7142 /* We don't support RT-tasks being in separate groups */ 7143 if (task->sched_class != &fair_sched_class) 7144 return -EINVAL; 7145#endif 7146 } 7147 return 0; 7148} 7149 7150static void cpu_cgroup_attach(struct cgroup *cgrp, 7151 struct cgroup_taskset *tset) 7152{ 7153 struct task_struct *task; 7154 7155 cgroup_taskset_for_each(task, cgrp, tset) 7156 sched_move_task(task); 7157} 7158 7159static void 7160cpu_cgroup_exit(struct cgroup *cgrp, struct cgroup *old_cgrp, 7161 struct task_struct *task) 7162{ 7163 /* 7164 * cgroup_exit() is called in the copy_process() failure path. 7165 * Ignore this case since the task hasn't ran yet, this avoids 7166 * trying to poke a half freed task state from generic code. 7167 */ 7168 if (!(task->flags & PF_EXITING)) 7169 return; 7170 7171 sched_move_task(task); 7172} 7173 7174#ifdef CONFIG_FAIR_GROUP_SCHED 7175static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype, 7176 u64 shareval) 7177{ 7178 return sched_group_set_shares(cgroup_tg(cgrp), scale_load(shareval)); 7179} 7180 7181static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft) 7182{ 7183 struct task_group *tg = cgroup_tg(cgrp); 7184 7185 return (u64) scale_load_down(tg->shares); 7186} 7187 7188#ifdef CONFIG_CFS_BANDWIDTH 7189static DEFINE_MUTEX(cfs_constraints_mutex); 7190 7191const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */ 7192const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */ 7193 7194static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime); 7195 7196static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota) 7197{ 7198 int i, ret = 0, runtime_enabled, runtime_was_enabled; 7199 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth; 7200 7201 if (tg == &root_task_group) 7202 return -EINVAL; 7203 7204 /* 7205 * Ensure we have at some amount of bandwidth every period. This is 7206 * to prevent reaching a state of large arrears when throttled via 7207 * entity_tick() resulting in prolonged exit starvation. 7208 */ 7209 if (quota < min_cfs_quota_period || period < min_cfs_quota_period) 7210 return -EINVAL; 7211 7212 /* 7213 * Likewise, bound things on the otherside by preventing insane quota 7214 * periods. This also allows us to normalize in computing quota 7215 * feasibility. 7216 */ 7217 if (period > max_cfs_quota_period) 7218 return -EINVAL; 7219 7220 mutex_lock(&cfs_constraints_mutex); 7221 ret = __cfs_schedulable(tg, period, quota); 7222 if (ret) 7223 goto out_unlock; 7224 7225 runtime_enabled = quota != RUNTIME_INF; 7226 runtime_was_enabled = cfs_b->quota != RUNTIME_INF; 7227 account_cfs_bandwidth_used(runtime_enabled, runtime_was_enabled); 7228 raw_spin_lock_irq(&cfs_b->lock); 7229 cfs_b->period = ns_to_ktime(period); 7230 cfs_b->quota = quota; 7231 7232 __refill_cfs_bandwidth_runtime(cfs_b); 7233 /* restart the period timer (if active) to handle new period expiry */ 7234 if (runtime_enabled && cfs_b->timer_active) { 7235 /* force a reprogram */ 7236 cfs_b->timer_active = 0; 7237 __start_cfs_bandwidth(cfs_b); 7238 } 7239 raw_spin_unlock_irq(&cfs_b->lock); 7240 7241 for_each_possible_cpu(i) { 7242 struct cfs_rq *cfs_rq = tg->cfs_rq[i]; 7243 struct rq *rq = cfs_rq->rq; 7244 7245 raw_spin_lock_irq(&rq->lock); 7246 cfs_rq->runtime_enabled = runtime_enabled; 7247 cfs_rq->runtime_remaining = 0; 7248 7249 if (cfs_rq->throttled) 7250 unthrottle_cfs_rq(cfs_rq); 7251 raw_spin_unlock_irq(&rq->lock); 7252 } 7253out_unlock: 7254 mutex_unlock(&cfs_constraints_mutex); 7255 7256 return ret; 7257} 7258 7259int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us) 7260{ 7261 u64 quota, period; 7262 7263 period = ktime_to_ns(tg->cfs_bandwidth.period); 7264 if (cfs_quota_us < 0) 7265 quota = RUNTIME_INF; 7266 else 7267 quota = (u64)cfs_quota_us * NSEC_PER_USEC; 7268 7269 return tg_set_cfs_bandwidth(tg, period, quota); 7270} 7271 7272long tg_get_cfs_quota(struct task_group *tg) 7273{ 7274 u64 quota_us; 7275 7276 if (tg->cfs_bandwidth.quota == RUNTIME_INF) 7277 return -1; 7278 7279 quota_us = tg->cfs_bandwidth.quota; 7280 do_div(quota_us, NSEC_PER_USEC); 7281 7282 return quota_us; 7283} 7284 7285int tg_set_cfs_period(struct task_group *tg, long cfs_period_us) 7286{ 7287 u64 quota, period; 7288 7289 period = (u64)cfs_period_us * NSEC_PER_USEC; 7290 quota = tg->cfs_bandwidth.quota; 7291 7292 return tg_set_cfs_bandwidth(tg, period, quota); 7293} 7294 7295long tg_get_cfs_period(struct task_group *tg) 7296{ 7297 u64 cfs_period_us; 7298 7299 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period); 7300 do_div(cfs_period_us, NSEC_PER_USEC); 7301 7302 return cfs_period_us; 7303} 7304 7305static s64 cpu_cfs_quota_read_s64(struct cgroup *cgrp, struct cftype *cft) 7306{ 7307 return tg_get_cfs_quota(cgroup_tg(cgrp)); 7308} 7309 7310static int cpu_cfs_quota_write_s64(struct cgroup *cgrp, struct cftype *cftype, 7311 s64 cfs_quota_us) 7312{ 7313 return tg_set_cfs_quota(cgroup_tg(cgrp), cfs_quota_us); 7314} 7315 7316static u64 cpu_cfs_period_read_u64(struct cgroup *cgrp, struct cftype *cft) 7317{ 7318 return tg_get_cfs_period(cgroup_tg(cgrp)); 7319} 7320 7321static int cpu_cfs_period_write_u64(struct cgroup *cgrp, struct cftype *cftype, 7322 u64 cfs_period_us) 7323{ 7324 return tg_set_cfs_period(cgroup_tg(cgrp), cfs_period_us); 7325} 7326 7327struct cfs_schedulable_data { 7328 struct task_group *tg; 7329 u64 period, quota; 7330}; 7331 7332/* 7333 * normalize group quota/period to be quota/max_period 7334 * note: units are usecs 7335 */ 7336static u64 normalize_cfs_quota(struct task_group *tg, 7337 struct cfs_schedulable_data *d) 7338{ 7339 u64 quota, period; 7340 7341 if (tg == d->tg) { 7342 period = d->period; 7343 quota = d->quota; 7344 } else { 7345 period = tg_get_cfs_period(tg); 7346 quota = tg_get_cfs_quota(tg); 7347 } 7348 7349 /* note: these should typically be equivalent */ 7350 if (quota == RUNTIME_INF || quota == -1) 7351 return RUNTIME_INF; 7352 7353 return to_ratio(period, quota); 7354} 7355 7356static int tg_cfs_schedulable_down(struct task_group *tg, void *data) 7357{ 7358 struct cfs_schedulable_data *d = data; 7359 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth; 7360 s64 quota = 0, parent_quota = -1; 7361 7362 if (!tg->parent) { 7363 quota = RUNTIME_INF; 7364 } else { 7365 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth; 7366 7367 quota = normalize_cfs_quota(tg, d); 7368 parent_quota = parent_b->hierarchal_quota; 7369 7370 /* 7371 * ensure max(child_quota) <= parent_quota, inherit when no 7372 * limit is set 7373 */ 7374 if (quota == RUNTIME_INF) 7375 quota = parent_quota; 7376 else if (parent_quota != RUNTIME_INF && quota > parent_quota) 7377 return -EINVAL; 7378 } 7379 cfs_b->hierarchal_quota = quota; 7380 7381 return 0; 7382} 7383 7384static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota) 7385{ 7386 int ret; 7387 struct cfs_schedulable_data data = { 7388 .tg = tg, 7389 .period = period, 7390 .quota = quota, 7391 }; 7392 7393 if (quota != RUNTIME_INF) { 7394 do_div(data.period, NSEC_PER_USEC); 7395 do_div(data.quota, NSEC_PER_USEC); 7396 } 7397 7398 rcu_read_lock(); 7399 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data); 7400 rcu_read_unlock(); 7401 7402 return ret; 7403} 7404 7405static int cpu_stats_show(struct cgroup *cgrp, struct cftype *cft, 7406 struct cgroup_map_cb *cb) 7407{ 7408 struct task_group *tg = cgroup_tg(cgrp); 7409 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth; 7410 7411 cb->fill(cb, "nr_periods", cfs_b->nr_periods); 7412 cb->fill(cb, "nr_throttled", cfs_b->nr_throttled); 7413 cb->fill(cb, "throttled_time", cfs_b->throttled_time); 7414 7415 return 0; 7416} 7417#endif /* CONFIG_CFS_BANDWIDTH */ 7418#endif /* CONFIG_FAIR_GROUP_SCHED */ 7419 7420#ifdef CONFIG_RT_GROUP_SCHED 7421static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft, 7422 s64 val) 7423{ 7424 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val); 7425} 7426 7427static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft) 7428{ 7429 return sched_group_rt_runtime(cgroup_tg(cgrp)); 7430} 7431 7432static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype, 7433 u64 rt_period_us) 7434{ 7435 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us); 7436} 7437 7438static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft) 7439{ 7440 return sched_group_rt_period(cgroup_tg(cgrp)); 7441} 7442#endif /* CONFIG_RT_GROUP_SCHED */ 7443 7444static struct cftype cpu_files[] = { 7445#ifdef CONFIG_FAIR_GROUP_SCHED 7446 { 7447 .name = "shares", 7448 .read_u64 = cpu_shares_read_u64, 7449 .write_u64 = cpu_shares_write_u64, 7450 }, 7451#endif 7452#ifdef CONFIG_CFS_BANDWIDTH 7453 { 7454 .name = "cfs_quota_us", 7455 .read_s64 = cpu_cfs_quota_read_s64, 7456 .write_s64 = cpu_cfs_quota_write_s64, 7457 }, 7458 { 7459 .name = "cfs_period_us", 7460 .read_u64 = cpu_cfs_period_read_u64, 7461 .write_u64 = cpu_cfs_period_write_u64, 7462 }, 7463 { 7464 .name = "stat", 7465 .read_map = cpu_stats_show, 7466 }, 7467#endif 7468#ifdef CONFIG_RT_GROUP_SCHED 7469 { 7470 .name = "rt_runtime_us", 7471 .read_s64 = cpu_rt_runtime_read, 7472 .write_s64 = cpu_rt_runtime_write, 7473 }, 7474 { 7475 .name = "rt_period_us", 7476 .read_u64 = cpu_rt_period_read_uint, 7477 .write_u64 = cpu_rt_period_write_uint, 7478 }, 7479#endif 7480 { } /* terminate */ 7481}; 7482 7483struct cgroup_subsys cpu_cgroup_subsys = { 7484 .name = "cpu", 7485 .css_alloc = cpu_cgroup_css_alloc, 7486 .css_free = cpu_cgroup_css_free, 7487 .css_online = cpu_cgroup_css_online, 7488 .css_offline = cpu_cgroup_css_offline, 7489 .can_attach = cpu_cgroup_can_attach, 7490 .attach = cpu_cgroup_attach, 7491 .exit = cpu_cgroup_exit, 7492 .subsys_id = cpu_cgroup_subsys_id, 7493 .base_cftypes = cpu_files, 7494 .early_init = 1, 7495}; 7496 7497#endif /* CONFIG_CGROUP_SCHED */ 7498 7499void dump_cpu_task(int cpu) 7500{ 7501 pr_info("Task dump for CPU %d:\n", cpu); 7502 sched_show_task(cpu_curr(cpu)); 7503} 7504