core.c revision c5405a495e88d93cf9b4f4cc91507c7f4afcb901
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->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->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 ttwu_do_wakeup(rq, p, wake_flags); 1369 ret = 1; 1370 } 1371 __task_rq_unlock(rq); 1372 1373 return ret; 1374} 1375 1376#ifdef CONFIG_SMP 1377static void sched_ttwu_pending(void) 1378{ 1379 struct rq *rq = this_rq(); 1380 struct llist_node *llist = llist_del_all(&rq->wake_list); 1381 struct task_struct *p; 1382 1383 raw_spin_lock(&rq->lock); 1384 1385 while (llist) { 1386 p = llist_entry(llist, struct task_struct, wake_entry); 1387 llist = llist_next(llist); 1388 ttwu_do_activate(rq, p, 0); 1389 } 1390 1391 raw_spin_unlock(&rq->lock); 1392} 1393 1394void scheduler_ipi(void) 1395{ 1396 if (llist_empty(&this_rq()->wake_list) && !got_nohz_idle_kick() 1397 && !tick_nohz_full_cpu(smp_processor_id())) 1398 return; 1399 1400 /* 1401 * Not all reschedule IPI handlers call irq_enter/irq_exit, since 1402 * traditionally all their work was done from the interrupt return 1403 * path. Now that we actually do some work, we need to make sure 1404 * we do call them. 1405 * 1406 * Some archs already do call them, luckily irq_enter/exit nest 1407 * properly. 1408 * 1409 * Arguably we should visit all archs and update all handlers, 1410 * however a fair share of IPIs are still resched only so this would 1411 * somewhat pessimize the simple resched case. 1412 */ 1413 irq_enter(); 1414 tick_nohz_full_check(); 1415 sched_ttwu_pending(); 1416 1417 /* 1418 * Check if someone kicked us for doing the nohz idle load balance. 1419 */ 1420 if (unlikely(got_nohz_idle_kick() && !need_resched())) { 1421 this_rq()->idle_balance = 1; 1422 raise_softirq_irqoff(SCHED_SOFTIRQ); 1423 } 1424 irq_exit(); 1425} 1426 1427static void ttwu_queue_remote(struct task_struct *p, int cpu) 1428{ 1429 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list)) 1430 smp_send_reschedule(cpu); 1431} 1432 1433bool cpus_share_cache(int this_cpu, int that_cpu) 1434{ 1435 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu); 1436} 1437#endif /* CONFIG_SMP */ 1438 1439static void ttwu_queue(struct task_struct *p, int cpu) 1440{ 1441 struct rq *rq = cpu_rq(cpu); 1442 1443#if defined(CONFIG_SMP) 1444 if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) { 1445 sched_clock_cpu(cpu); /* sync clocks x-cpu */ 1446 ttwu_queue_remote(p, cpu); 1447 return; 1448 } 1449#endif 1450 1451 raw_spin_lock(&rq->lock); 1452 ttwu_do_activate(rq, p, 0); 1453 raw_spin_unlock(&rq->lock); 1454} 1455 1456/** 1457 * try_to_wake_up - wake up a thread 1458 * @p: the thread to be awakened 1459 * @state: the mask of task states that can be woken 1460 * @wake_flags: wake modifier flags (WF_*) 1461 * 1462 * Put it on the run-queue if it's not already there. The "current" 1463 * thread is always on the run-queue (except when the actual 1464 * re-schedule is in progress), and as such you're allowed to do 1465 * the simpler "current->state = TASK_RUNNING" to mark yourself 1466 * runnable without the overhead of this. 1467 * 1468 * Returns %true if @p was woken up, %false if it was already running 1469 * or @state didn't match @p's state. 1470 */ 1471static int 1472try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags) 1473{ 1474 unsigned long flags; 1475 int cpu, success = 0; 1476 1477 smp_wmb(); 1478 raw_spin_lock_irqsave(&p->pi_lock, flags); 1479 if (!(p->state & state)) 1480 goto out; 1481 1482 success = 1; /* we're going to change ->state */ 1483 cpu = task_cpu(p); 1484 1485 if (p->on_rq && ttwu_remote(p, wake_flags)) 1486 goto stat; 1487 1488#ifdef CONFIG_SMP 1489 /* 1490 * If the owning (remote) cpu is still in the middle of schedule() with 1491 * this task as prev, wait until its done referencing the task. 1492 */ 1493 while (p->on_cpu) 1494 cpu_relax(); 1495 /* 1496 * Pairs with the smp_wmb() in finish_lock_switch(). 1497 */ 1498 smp_rmb(); 1499 1500 p->sched_contributes_to_load = !!task_contributes_to_load(p); 1501 p->state = TASK_WAKING; 1502 1503 if (p->sched_class->task_waking) 1504 p->sched_class->task_waking(p); 1505 1506 cpu = select_task_rq(p, SD_BALANCE_WAKE, wake_flags); 1507 if (task_cpu(p) != cpu) { 1508 wake_flags |= WF_MIGRATED; 1509 set_task_cpu(p, cpu); 1510 } 1511#endif /* CONFIG_SMP */ 1512 1513 ttwu_queue(p, cpu); 1514stat: 1515 ttwu_stat(p, cpu, wake_flags); 1516out: 1517 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 1518 1519 return success; 1520} 1521 1522/** 1523 * try_to_wake_up_local - try to wake up a local task with rq lock held 1524 * @p: the thread to be awakened 1525 * 1526 * Put @p on the run-queue if it's not already there. The caller must 1527 * ensure that this_rq() is locked, @p is bound to this_rq() and not 1528 * the current task. 1529 */ 1530static void try_to_wake_up_local(struct task_struct *p) 1531{ 1532 struct rq *rq = task_rq(p); 1533 1534 if (WARN_ON_ONCE(rq != this_rq()) || 1535 WARN_ON_ONCE(p == current)) 1536 return; 1537 1538 lockdep_assert_held(&rq->lock); 1539 1540 if (!raw_spin_trylock(&p->pi_lock)) { 1541 raw_spin_unlock(&rq->lock); 1542 raw_spin_lock(&p->pi_lock); 1543 raw_spin_lock(&rq->lock); 1544 } 1545 1546 if (!(p->state & TASK_NORMAL)) 1547 goto out; 1548 1549 if (!p->on_rq) 1550 ttwu_activate(rq, p, ENQUEUE_WAKEUP); 1551 1552 ttwu_do_wakeup(rq, p, 0); 1553 ttwu_stat(p, smp_processor_id(), 0); 1554out: 1555 raw_spin_unlock(&p->pi_lock); 1556} 1557 1558/** 1559 * wake_up_process - Wake up a specific process 1560 * @p: The process to be woken up. 1561 * 1562 * Attempt to wake up the nominated process and move it to the set of runnable 1563 * processes. Returns 1 if the process was woken up, 0 if it was already 1564 * running. 1565 * 1566 * It may be assumed that this function implies a write memory barrier before 1567 * changing the task state if and only if any tasks are woken up. 1568 */ 1569int wake_up_process(struct task_struct *p) 1570{ 1571 WARN_ON(task_is_stopped_or_traced(p)); 1572 return try_to_wake_up(p, TASK_NORMAL, 0); 1573} 1574EXPORT_SYMBOL(wake_up_process); 1575 1576int wake_up_state(struct task_struct *p, unsigned int state) 1577{ 1578 return try_to_wake_up(p, state, 0); 1579} 1580 1581/* 1582 * Perform scheduler related setup for a newly forked process p. 1583 * p is forked by current. 1584 * 1585 * __sched_fork() is basic setup used by init_idle() too: 1586 */ 1587static void __sched_fork(struct task_struct *p) 1588{ 1589 p->on_rq = 0; 1590 1591 p->se.on_rq = 0; 1592 p->se.exec_start = 0; 1593 p->se.sum_exec_runtime = 0; 1594 p->se.prev_sum_exec_runtime = 0; 1595 p->se.nr_migrations = 0; 1596 p->se.vruntime = 0; 1597 INIT_LIST_HEAD(&p->se.group_node); 1598 1599/* 1600 * Load-tracking only depends on SMP, FAIR_GROUP_SCHED dependency below may be 1601 * removed when useful for applications beyond shares distribution (e.g. 1602 * load-balance). 1603 */ 1604#if defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED) 1605 p->se.avg.runnable_avg_period = 0; 1606 p->se.avg.runnable_avg_sum = 0; 1607#endif 1608#ifdef CONFIG_SCHEDSTATS 1609 memset(&p->se.statistics, 0, sizeof(p->se.statistics)); 1610#endif 1611 1612 INIT_LIST_HEAD(&p->rt.run_list); 1613 1614#ifdef CONFIG_PREEMPT_NOTIFIERS 1615 INIT_HLIST_HEAD(&p->preempt_notifiers); 1616#endif 1617 1618#ifdef CONFIG_NUMA_BALANCING 1619 if (p->mm && atomic_read(&p->mm->mm_users) == 1) { 1620 p->mm->numa_next_scan = jiffies; 1621 p->mm->numa_next_reset = jiffies; 1622 p->mm->numa_scan_seq = 0; 1623 } 1624 1625 p->node_stamp = 0ULL; 1626 p->numa_scan_seq = p->mm ? p->mm->numa_scan_seq : 0; 1627 p->numa_migrate_seq = p->mm ? p->mm->numa_scan_seq - 1 : 0; 1628 p->numa_scan_period = sysctl_numa_balancing_scan_delay; 1629 p->numa_work.next = &p->numa_work; 1630#endif /* CONFIG_NUMA_BALANCING */ 1631} 1632 1633#ifdef CONFIG_NUMA_BALANCING 1634#ifdef CONFIG_SCHED_DEBUG 1635void set_numabalancing_state(bool enabled) 1636{ 1637 if (enabled) 1638 sched_feat_set("NUMA"); 1639 else 1640 sched_feat_set("NO_NUMA"); 1641} 1642#else 1643__read_mostly bool numabalancing_enabled; 1644 1645void set_numabalancing_state(bool enabled) 1646{ 1647 numabalancing_enabled = enabled; 1648} 1649#endif /* CONFIG_SCHED_DEBUG */ 1650#endif /* CONFIG_NUMA_BALANCING */ 1651 1652/* 1653 * fork()/clone()-time setup: 1654 */ 1655void sched_fork(struct task_struct *p) 1656{ 1657 unsigned long flags; 1658 int cpu = get_cpu(); 1659 1660 __sched_fork(p); 1661 /* 1662 * We mark the process as running here. This guarantees that 1663 * nobody will actually run it, and a signal or other external 1664 * event cannot wake it up and insert it on the runqueue either. 1665 */ 1666 p->state = TASK_RUNNING; 1667 1668 /* 1669 * Make sure we do not leak PI boosting priority to the child. 1670 */ 1671 p->prio = current->normal_prio; 1672 1673 /* 1674 * Revert to default priority/policy on fork if requested. 1675 */ 1676 if (unlikely(p->sched_reset_on_fork)) { 1677 if (task_has_rt_policy(p)) { 1678 p->policy = SCHED_NORMAL; 1679 p->static_prio = NICE_TO_PRIO(0); 1680 p->rt_priority = 0; 1681 } else if (PRIO_TO_NICE(p->static_prio) < 0) 1682 p->static_prio = NICE_TO_PRIO(0); 1683 1684 p->prio = p->normal_prio = __normal_prio(p); 1685 set_load_weight(p); 1686 1687 /* 1688 * We don't need the reset flag anymore after the fork. It has 1689 * fulfilled its duty: 1690 */ 1691 p->sched_reset_on_fork = 0; 1692 } 1693 1694 if (!rt_prio(p->prio)) 1695 p->sched_class = &fair_sched_class; 1696 1697 if (p->sched_class->task_fork) 1698 p->sched_class->task_fork(p); 1699 1700 /* 1701 * The child is not yet in the pid-hash so no cgroup attach races, 1702 * and the cgroup is pinned to this child due to cgroup_fork() 1703 * is ran before sched_fork(). 1704 * 1705 * Silence PROVE_RCU. 1706 */ 1707 raw_spin_lock_irqsave(&p->pi_lock, flags); 1708 set_task_cpu(p, cpu); 1709 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 1710 1711#if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT) 1712 if (likely(sched_info_on())) 1713 memset(&p->sched_info, 0, sizeof(p->sched_info)); 1714#endif 1715#if defined(CONFIG_SMP) 1716 p->on_cpu = 0; 1717#endif 1718#ifdef CONFIG_PREEMPT_COUNT 1719 /* Want to start with kernel preemption disabled. */ 1720 task_thread_info(p)->preempt_count = 1; 1721#endif 1722#ifdef CONFIG_SMP 1723 plist_node_init(&p->pushable_tasks, MAX_PRIO); 1724#endif 1725 1726 put_cpu(); 1727} 1728 1729/* 1730 * wake_up_new_task - wake up a newly created task for the first time. 1731 * 1732 * This function will do some initial scheduler statistics housekeeping 1733 * that must be done for every newly created context, then puts the task 1734 * on the runqueue and wakes it. 1735 */ 1736void wake_up_new_task(struct task_struct *p) 1737{ 1738 unsigned long flags; 1739 struct rq *rq; 1740 1741 raw_spin_lock_irqsave(&p->pi_lock, flags); 1742#ifdef CONFIG_SMP 1743 /* 1744 * Fork balancing, do it here and not earlier because: 1745 * - cpus_allowed can change in the fork path 1746 * - any previously selected cpu might disappear through hotplug 1747 */ 1748 set_task_cpu(p, select_task_rq(p, SD_BALANCE_FORK, 0)); 1749#endif 1750 1751 rq = __task_rq_lock(p); 1752 activate_task(rq, p, 0); 1753 p->on_rq = 1; 1754 trace_sched_wakeup_new(p, true); 1755 check_preempt_curr(rq, p, WF_FORK); 1756#ifdef CONFIG_SMP 1757 if (p->sched_class->task_woken) 1758 p->sched_class->task_woken(rq, p); 1759#endif 1760 task_rq_unlock(rq, p, &flags); 1761} 1762 1763#ifdef CONFIG_PREEMPT_NOTIFIERS 1764 1765/** 1766 * preempt_notifier_register - tell me when current is being preempted & rescheduled 1767 * @notifier: notifier struct to register 1768 */ 1769void preempt_notifier_register(struct preempt_notifier *notifier) 1770{ 1771 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers); 1772} 1773EXPORT_SYMBOL_GPL(preempt_notifier_register); 1774 1775/** 1776 * preempt_notifier_unregister - no longer interested in preemption notifications 1777 * @notifier: notifier struct to unregister 1778 * 1779 * This is safe to call from within a preemption notifier. 1780 */ 1781void preempt_notifier_unregister(struct preempt_notifier *notifier) 1782{ 1783 hlist_del(¬ifier->link); 1784} 1785EXPORT_SYMBOL_GPL(preempt_notifier_unregister); 1786 1787static void fire_sched_in_preempt_notifiers(struct task_struct *curr) 1788{ 1789 struct preempt_notifier *notifier; 1790 1791 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link) 1792 notifier->ops->sched_in(notifier, raw_smp_processor_id()); 1793} 1794 1795static void 1796fire_sched_out_preempt_notifiers(struct task_struct *curr, 1797 struct task_struct *next) 1798{ 1799 struct preempt_notifier *notifier; 1800 1801 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link) 1802 notifier->ops->sched_out(notifier, next); 1803} 1804 1805#else /* !CONFIG_PREEMPT_NOTIFIERS */ 1806 1807static void fire_sched_in_preempt_notifiers(struct task_struct *curr) 1808{ 1809} 1810 1811static void 1812fire_sched_out_preempt_notifiers(struct task_struct *curr, 1813 struct task_struct *next) 1814{ 1815} 1816 1817#endif /* CONFIG_PREEMPT_NOTIFIERS */ 1818 1819/** 1820 * prepare_task_switch - prepare to switch tasks 1821 * @rq: the runqueue preparing to switch 1822 * @prev: the current task that is being switched out 1823 * @next: the task we are going to switch to. 1824 * 1825 * This is called with the rq lock held and interrupts off. It must 1826 * be paired with a subsequent finish_task_switch after the context 1827 * switch. 1828 * 1829 * prepare_task_switch sets up locking and calls architecture specific 1830 * hooks. 1831 */ 1832static inline void 1833prepare_task_switch(struct rq *rq, struct task_struct *prev, 1834 struct task_struct *next) 1835{ 1836 trace_sched_switch(prev, next); 1837 sched_info_switch(prev, next); 1838 perf_event_task_sched_out(prev, next); 1839 fire_sched_out_preempt_notifiers(prev, next); 1840 prepare_lock_switch(rq, next); 1841 prepare_arch_switch(next); 1842} 1843 1844/** 1845 * finish_task_switch - clean up after a task-switch 1846 * @rq: runqueue associated with task-switch 1847 * @prev: the thread we just switched away from. 1848 * 1849 * finish_task_switch must be called after the context switch, paired 1850 * with a prepare_task_switch call before the context switch. 1851 * finish_task_switch will reconcile locking set up by prepare_task_switch, 1852 * and do any other architecture-specific cleanup actions. 1853 * 1854 * Note that we may have delayed dropping an mm in context_switch(). If 1855 * so, we finish that here outside of the runqueue lock. (Doing it 1856 * with the lock held can cause deadlocks; see schedule() for 1857 * details.) 1858 */ 1859static void finish_task_switch(struct rq *rq, struct task_struct *prev) 1860 __releases(rq->lock) 1861{ 1862 struct mm_struct *mm = rq->prev_mm; 1863 long prev_state; 1864 1865 rq->prev_mm = NULL; 1866 1867 /* 1868 * A task struct has one reference for the use as "current". 1869 * If a task dies, then it sets TASK_DEAD in tsk->state and calls 1870 * schedule one last time. The schedule call will never return, and 1871 * the scheduled task must drop that reference. 1872 * The test for TASK_DEAD must occur while the runqueue locks are 1873 * still held, otherwise prev could be scheduled on another cpu, die 1874 * there before we look at prev->state, and then the reference would 1875 * be dropped twice. 1876 * Manfred Spraul <manfred@colorfullife.com> 1877 */ 1878 prev_state = prev->state; 1879 vtime_task_switch(prev); 1880 finish_arch_switch(prev); 1881 perf_event_task_sched_in(prev, current); 1882 finish_lock_switch(rq, prev); 1883 finish_arch_post_lock_switch(); 1884 1885 fire_sched_in_preempt_notifiers(current); 1886 if (mm) 1887 mmdrop(mm); 1888 if (unlikely(prev_state == TASK_DEAD)) { 1889 /* 1890 * Remove function-return probe instances associated with this 1891 * task and put them back on the free list. 1892 */ 1893 kprobe_flush_task(prev); 1894 put_task_struct(prev); 1895 } 1896 1897 tick_nohz_task_switch(current); 1898} 1899 1900#ifdef CONFIG_SMP 1901 1902/* assumes rq->lock is held */ 1903static inline void pre_schedule(struct rq *rq, struct task_struct *prev) 1904{ 1905 if (prev->sched_class->pre_schedule) 1906 prev->sched_class->pre_schedule(rq, prev); 1907} 1908 1909/* rq->lock is NOT held, but preemption is disabled */ 1910static inline void post_schedule(struct rq *rq) 1911{ 1912 if (rq->post_schedule) { 1913 unsigned long flags; 1914 1915 raw_spin_lock_irqsave(&rq->lock, flags); 1916 if (rq->curr->sched_class->post_schedule) 1917 rq->curr->sched_class->post_schedule(rq); 1918 raw_spin_unlock_irqrestore(&rq->lock, flags); 1919 1920 rq->post_schedule = 0; 1921 } 1922} 1923 1924#else 1925 1926static inline void pre_schedule(struct rq *rq, struct task_struct *p) 1927{ 1928} 1929 1930static inline void post_schedule(struct rq *rq) 1931{ 1932} 1933 1934#endif 1935 1936/** 1937 * schedule_tail - first thing a freshly forked thread must call. 1938 * @prev: the thread we just switched away from. 1939 */ 1940asmlinkage void schedule_tail(struct task_struct *prev) 1941 __releases(rq->lock) 1942{ 1943 struct rq *rq = this_rq(); 1944 1945 finish_task_switch(rq, prev); 1946 1947 /* 1948 * FIXME: do we need to worry about rq being invalidated by the 1949 * task_switch? 1950 */ 1951 post_schedule(rq); 1952 1953#ifdef __ARCH_WANT_UNLOCKED_CTXSW 1954 /* In this case, finish_task_switch does not reenable preemption */ 1955 preempt_enable(); 1956#endif 1957 if (current->set_child_tid) 1958 put_user(task_pid_vnr(current), current->set_child_tid); 1959} 1960 1961/* 1962 * context_switch - switch to the new MM and the new 1963 * thread's register state. 1964 */ 1965static inline void 1966context_switch(struct rq *rq, struct task_struct *prev, 1967 struct task_struct *next) 1968{ 1969 struct mm_struct *mm, *oldmm; 1970 1971 prepare_task_switch(rq, prev, next); 1972 1973 mm = next->mm; 1974 oldmm = prev->active_mm; 1975 /* 1976 * For paravirt, this is coupled with an exit in switch_to to 1977 * combine the page table reload and the switch backend into 1978 * one hypercall. 1979 */ 1980 arch_start_context_switch(prev); 1981 1982 if (!mm) { 1983 next->active_mm = oldmm; 1984 atomic_inc(&oldmm->mm_count); 1985 enter_lazy_tlb(oldmm, next); 1986 } else 1987 switch_mm(oldmm, mm, next); 1988 1989 if (!prev->mm) { 1990 prev->active_mm = NULL; 1991 rq->prev_mm = oldmm; 1992 } 1993 /* 1994 * Since the runqueue lock will be released by the next 1995 * task (which is an invalid locking op but in the case 1996 * of the scheduler it's an obvious special-case), so we 1997 * do an early lockdep release here: 1998 */ 1999#ifndef __ARCH_WANT_UNLOCKED_CTXSW 2000 spin_release(&rq->lock.dep_map, 1, _THIS_IP_); 2001#endif 2002 2003 context_tracking_task_switch(prev, next); 2004 /* Here we just switch the register state and the stack. */ 2005 switch_to(prev, next, prev); 2006 2007 barrier(); 2008 /* 2009 * this_rq must be evaluated again because prev may have moved 2010 * CPUs since it called schedule(), thus the 'rq' on its stack 2011 * frame will be invalid. 2012 */ 2013 finish_task_switch(this_rq(), prev); 2014} 2015 2016/* 2017 * nr_running and nr_context_switches: 2018 * 2019 * externally visible scheduler statistics: current number of runnable 2020 * threads, total number of context switches performed since bootup. 2021 */ 2022unsigned long nr_running(void) 2023{ 2024 unsigned long i, sum = 0; 2025 2026 for_each_online_cpu(i) 2027 sum += cpu_rq(i)->nr_running; 2028 2029 return sum; 2030} 2031 2032unsigned long long nr_context_switches(void) 2033{ 2034 int i; 2035 unsigned long long sum = 0; 2036 2037 for_each_possible_cpu(i) 2038 sum += cpu_rq(i)->nr_switches; 2039 2040 return sum; 2041} 2042 2043unsigned long nr_iowait(void) 2044{ 2045 unsigned long i, sum = 0; 2046 2047 for_each_possible_cpu(i) 2048 sum += atomic_read(&cpu_rq(i)->nr_iowait); 2049 2050 return sum; 2051} 2052 2053unsigned long nr_iowait_cpu(int cpu) 2054{ 2055 struct rq *this = cpu_rq(cpu); 2056 return atomic_read(&this->nr_iowait); 2057} 2058 2059#ifdef CONFIG_SMP 2060 2061/* 2062 * sched_exec - execve() is a valuable balancing opportunity, because at 2063 * this point the task has the smallest effective memory and cache footprint. 2064 */ 2065void sched_exec(void) 2066{ 2067 struct task_struct *p = current; 2068 unsigned long flags; 2069 int dest_cpu; 2070 2071 raw_spin_lock_irqsave(&p->pi_lock, flags); 2072 dest_cpu = p->sched_class->select_task_rq(p, SD_BALANCE_EXEC, 0); 2073 if (dest_cpu == smp_processor_id()) 2074 goto unlock; 2075 2076 if (likely(cpu_active(dest_cpu))) { 2077 struct migration_arg arg = { p, dest_cpu }; 2078 2079 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 2080 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg); 2081 return; 2082 } 2083unlock: 2084 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 2085} 2086 2087#endif 2088 2089DEFINE_PER_CPU(struct kernel_stat, kstat); 2090DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat); 2091 2092EXPORT_PER_CPU_SYMBOL(kstat); 2093EXPORT_PER_CPU_SYMBOL(kernel_cpustat); 2094 2095/* 2096 * Return any ns on the sched_clock that have not yet been accounted in 2097 * @p in case that task is currently running. 2098 * 2099 * Called with task_rq_lock() held on @rq. 2100 */ 2101static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq) 2102{ 2103 u64 ns = 0; 2104 2105 if (task_current(rq, p)) { 2106 update_rq_clock(rq); 2107 ns = rq->clock_task - p->se.exec_start; 2108 if ((s64)ns < 0) 2109 ns = 0; 2110 } 2111 2112 return ns; 2113} 2114 2115unsigned long long task_delta_exec(struct task_struct *p) 2116{ 2117 unsigned long flags; 2118 struct rq *rq; 2119 u64 ns = 0; 2120 2121 rq = task_rq_lock(p, &flags); 2122 ns = do_task_delta_exec(p, rq); 2123 task_rq_unlock(rq, p, &flags); 2124 2125 return ns; 2126} 2127 2128/* 2129 * Return accounted runtime for the task. 2130 * In case the task is currently running, return the runtime plus current's 2131 * pending runtime that have not been accounted yet. 2132 */ 2133unsigned long long task_sched_runtime(struct task_struct *p) 2134{ 2135 unsigned long flags; 2136 struct rq *rq; 2137 u64 ns = 0; 2138 2139 rq = task_rq_lock(p, &flags); 2140 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq); 2141 task_rq_unlock(rq, p, &flags); 2142 2143 return ns; 2144} 2145 2146/* 2147 * This function gets called by the timer code, with HZ frequency. 2148 * We call it with interrupts disabled. 2149 */ 2150void scheduler_tick(void) 2151{ 2152 int cpu = smp_processor_id(); 2153 struct rq *rq = cpu_rq(cpu); 2154 struct task_struct *curr = rq->curr; 2155 2156 sched_clock_tick(); 2157 2158 raw_spin_lock(&rq->lock); 2159 update_rq_clock(rq); 2160 update_cpu_load_active(rq); 2161 curr->sched_class->task_tick(rq, curr, 0); 2162 raw_spin_unlock(&rq->lock); 2163 2164 perf_event_task_tick(); 2165 2166#ifdef CONFIG_SMP 2167 rq->idle_balance = idle_cpu(cpu); 2168 trigger_load_balance(rq, cpu); 2169#endif 2170 rq_last_tick_reset(rq); 2171} 2172 2173#ifdef CONFIG_NO_HZ_FULL 2174/** 2175 * scheduler_tick_max_deferment 2176 * 2177 * Keep at least one tick per second when a single 2178 * active task is running because the scheduler doesn't 2179 * yet completely support full dynticks environment. 2180 * 2181 * This makes sure that uptime, CFS vruntime, load 2182 * balancing, etc... continue to move forward, even 2183 * with a very low granularity. 2184 */ 2185u64 scheduler_tick_max_deferment(void) 2186{ 2187 struct rq *rq = this_rq(); 2188 unsigned long next, now = ACCESS_ONCE(jiffies); 2189 2190 next = rq->last_sched_tick + HZ; 2191 2192 if (time_before_eq(next, now)) 2193 return 0; 2194 2195 return jiffies_to_usecs(next - now) * NSEC_PER_USEC; 2196} 2197#endif 2198 2199notrace unsigned long get_parent_ip(unsigned long addr) 2200{ 2201 if (in_lock_functions(addr)) { 2202 addr = CALLER_ADDR2; 2203 if (in_lock_functions(addr)) 2204 addr = CALLER_ADDR3; 2205 } 2206 return addr; 2207} 2208 2209#if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \ 2210 defined(CONFIG_PREEMPT_TRACER)) 2211 2212void __kprobes add_preempt_count(int val) 2213{ 2214#ifdef CONFIG_DEBUG_PREEMPT 2215 /* 2216 * Underflow? 2217 */ 2218 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0))) 2219 return; 2220#endif 2221 preempt_count() += val; 2222#ifdef CONFIG_DEBUG_PREEMPT 2223 /* 2224 * Spinlock count overflowing soon? 2225 */ 2226 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >= 2227 PREEMPT_MASK - 10); 2228#endif 2229 if (preempt_count() == val) 2230 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1)); 2231} 2232EXPORT_SYMBOL(add_preempt_count); 2233 2234void __kprobes sub_preempt_count(int val) 2235{ 2236#ifdef CONFIG_DEBUG_PREEMPT 2237 /* 2238 * Underflow? 2239 */ 2240 if (DEBUG_LOCKS_WARN_ON(val > preempt_count())) 2241 return; 2242 /* 2243 * Is the spinlock portion underflowing? 2244 */ 2245 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) && 2246 !(preempt_count() & PREEMPT_MASK))) 2247 return; 2248#endif 2249 2250 if (preempt_count() == val) 2251 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1)); 2252 preempt_count() -= val; 2253} 2254EXPORT_SYMBOL(sub_preempt_count); 2255 2256#endif 2257 2258/* 2259 * Print scheduling while atomic bug: 2260 */ 2261static noinline void __schedule_bug(struct task_struct *prev) 2262{ 2263 if (oops_in_progress) 2264 return; 2265 2266 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n", 2267 prev->comm, prev->pid, preempt_count()); 2268 2269 debug_show_held_locks(prev); 2270 print_modules(); 2271 if (irqs_disabled()) 2272 print_irqtrace_events(prev); 2273 dump_stack(); 2274 add_taint(TAINT_WARN, LOCKDEP_STILL_OK); 2275} 2276 2277/* 2278 * Various schedule()-time debugging checks and statistics: 2279 */ 2280static inline void schedule_debug(struct task_struct *prev) 2281{ 2282 /* 2283 * Test if we are atomic. Since do_exit() needs to call into 2284 * schedule() atomically, we ignore that path for now. 2285 * Otherwise, whine if we are scheduling when we should not be. 2286 */ 2287 if (unlikely(in_atomic_preempt_off() && !prev->exit_state)) 2288 __schedule_bug(prev); 2289 rcu_sleep_check(); 2290 2291 profile_hit(SCHED_PROFILING, __builtin_return_address(0)); 2292 2293 schedstat_inc(this_rq(), sched_count); 2294} 2295 2296static void put_prev_task(struct rq *rq, struct task_struct *prev) 2297{ 2298 if (prev->on_rq || rq->skip_clock_update < 0) 2299 update_rq_clock(rq); 2300 prev->sched_class->put_prev_task(rq, prev); 2301} 2302 2303/* 2304 * Pick up the highest-prio task: 2305 */ 2306static inline struct task_struct * 2307pick_next_task(struct rq *rq) 2308{ 2309 const struct sched_class *class; 2310 struct task_struct *p; 2311 2312 /* 2313 * Optimization: we know that if all tasks are in 2314 * the fair class we can call that function directly: 2315 */ 2316 if (likely(rq->nr_running == rq->cfs.h_nr_running)) { 2317 p = fair_sched_class.pick_next_task(rq); 2318 if (likely(p)) 2319 return p; 2320 } 2321 2322 for_each_class(class) { 2323 p = class->pick_next_task(rq); 2324 if (p) 2325 return p; 2326 } 2327 2328 BUG(); /* the idle class will always have a runnable task */ 2329} 2330 2331/* 2332 * __schedule() is the main scheduler function. 2333 * 2334 * The main means of driving the scheduler and thus entering this function are: 2335 * 2336 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc. 2337 * 2338 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return 2339 * paths. For example, see arch/x86/entry_64.S. 2340 * 2341 * To drive preemption between tasks, the scheduler sets the flag in timer 2342 * interrupt handler scheduler_tick(). 2343 * 2344 * 3. Wakeups don't really cause entry into schedule(). They add a 2345 * task to the run-queue and that's it. 2346 * 2347 * Now, if the new task added to the run-queue preempts the current 2348 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets 2349 * called on the nearest possible occasion: 2350 * 2351 * - If the kernel is preemptible (CONFIG_PREEMPT=y): 2352 * 2353 * - in syscall or exception context, at the next outmost 2354 * preempt_enable(). (this might be as soon as the wake_up()'s 2355 * spin_unlock()!) 2356 * 2357 * - in IRQ context, return from interrupt-handler to 2358 * preemptible context 2359 * 2360 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set) 2361 * then at the next: 2362 * 2363 * - cond_resched() call 2364 * - explicit schedule() call 2365 * - return from syscall or exception to user-space 2366 * - return from interrupt-handler to user-space 2367 */ 2368static void __sched __schedule(void) 2369{ 2370 struct task_struct *prev, *next; 2371 unsigned long *switch_count; 2372 struct rq *rq; 2373 int cpu; 2374 2375need_resched: 2376 preempt_disable(); 2377 cpu = smp_processor_id(); 2378 rq = cpu_rq(cpu); 2379 rcu_note_context_switch(cpu); 2380 prev = rq->curr; 2381 2382 schedule_debug(prev); 2383 2384 if (sched_feat(HRTICK)) 2385 hrtick_clear(rq); 2386 2387 raw_spin_lock_irq(&rq->lock); 2388 2389 switch_count = &prev->nivcsw; 2390 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) { 2391 if (unlikely(signal_pending_state(prev->state, prev))) { 2392 prev->state = TASK_RUNNING; 2393 } else { 2394 deactivate_task(rq, prev, DEQUEUE_SLEEP); 2395 prev->on_rq = 0; 2396 2397 /* 2398 * If a worker went to sleep, notify and ask workqueue 2399 * whether it wants to wake up a task to maintain 2400 * concurrency. 2401 */ 2402 if (prev->flags & PF_WQ_WORKER) { 2403 struct task_struct *to_wakeup; 2404 2405 to_wakeup = wq_worker_sleeping(prev, cpu); 2406 if (to_wakeup) 2407 try_to_wake_up_local(to_wakeup); 2408 } 2409 } 2410 switch_count = &prev->nvcsw; 2411 } 2412 2413 pre_schedule(rq, prev); 2414 2415 if (unlikely(!rq->nr_running)) 2416 idle_balance(cpu, rq); 2417 2418 put_prev_task(rq, prev); 2419 next = pick_next_task(rq); 2420 clear_tsk_need_resched(prev); 2421 rq->skip_clock_update = 0; 2422 2423 if (likely(prev != next)) { 2424 rq->nr_switches++; 2425 rq->curr = next; 2426 ++*switch_count; 2427 2428 context_switch(rq, prev, next); /* unlocks the rq */ 2429 /* 2430 * The context switch have flipped the stack from under us 2431 * and restored the local variables which were saved when 2432 * this task called schedule() in the past. prev == current 2433 * is still correct, but it can be moved to another cpu/rq. 2434 */ 2435 cpu = smp_processor_id(); 2436 rq = cpu_rq(cpu); 2437 } else 2438 raw_spin_unlock_irq(&rq->lock); 2439 2440 post_schedule(rq); 2441 2442 sched_preempt_enable_no_resched(); 2443 if (need_resched()) 2444 goto need_resched; 2445} 2446 2447static inline void sched_submit_work(struct task_struct *tsk) 2448{ 2449 if (!tsk->state || tsk_is_pi_blocked(tsk)) 2450 return; 2451 /* 2452 * If we are going to sleep and we have plugged IO queued, 2453 * make sure to submit it to avoid deadlocks. 2454 */ 2455 if (blk_needs_flush_plug(tsk)) 2456 blk_schedule_flush_plug(tsk); 2457} 2458 2459asmlinkage void __sched schedule(void) 2460{ 2461 struct task_struct *tsk = current; 2462 2463 sched_submit_work(tsk); 2464 __schedule(); 2465} 2466EXPORT_SYMBOL(schedule); 2467 2468#ifdef CONFIG_CONTEXT_TRACKING 2469asmlinkage void __sched schedule_user(void) 2470{ 2471 /* 2472 * If we come here after a random call to set_need_resched(), 2473 * or we have been woken up remotely but the IPI has not yet arrived, 2474 * we haven't yet exited the RCU idle mode. Do it here manually until 2475 * we find a better solution. 2476 */ 2477 user_exit(); 2478 schedule(); 2479 user_enter(); 2480} 2481#endif 2482 2483/** 2484 * schedule_preempt_disabled - called with preemption disabled 2485 * 2486 * Returns with preemption disabled. Note: preempt_count must be 1 2487 */ 2488void __sched schedule_preempt_disabled(void) 2489{ 2490 sched_preempt_enable_no_resched(); 2491 schedule(); 2492 preempt_disable(); 2493} 2494 2495#ifdef CONFIG_PREEMPT 2496/* 2497 * this is the entry point to schedule() from in-kernel preemption 2498 * off of preempt_enable. Kernel preemptions off return from interrupt 2499 * occur there and call schedule directly. 2500 */ 2501asmlinkage void __sched notrace preempt_schedule(void) 2502{ 2503 struct thread_info *ti = current_thread_info(); 2504 2505 /* 2506 * If there is a non-zero preempt_count or interrupts are disabled, 2507 * we do not want to preempt the current task. Just return.. 2508 */ 2509 if (likely(ti->preempt_count || irqs_disabled())) 2510 return; 2511 2512 do { 2513 add_preempt_count_notrace(PREEMPT_ACTIVE); 2514 __schedule(); 2515 sub_preempt_count_notrace(PREEMPT_ACTIVE); 2516 2517 /* 2518 * Check again in case we missed a preemption opportunity 2519 * between schedule and now. 2520 */ 2521 barrier(); 2522 } while (need_resched()); 2523} 2524EXPORT_SYMBOL(preempt_schedule); 2525 2526/* 2527 * this is the entry point to schedule() from kernel preemption 2528 * off of irq context. 2529 * Note, that this is called and return with irqs disabled. This will 2530 * protect us against recursive calling from irq. 2531 */ 2532asmlinkage void __sched preempt_schedule_irq(void) 2533{ 2534 struct thread_info *ti = current_thread_info(); 2535 enum ctx_state prev_state; 2536 2537 /* Catch callers which need to be fixed */ 2538 BUG_ON(ti->preempt_count || !irqs_disabled()); 2539 2540 prev_state = exception_enter(); 2541 2542 do { 2543 add_preempt_count(PREEMPT_ACTIVE); 2544 local_irq_enable(); 2545 __schedule(); 2546 local_irq_disable(); 2547 sub_preempt_count(PREEMPT_ACTIVE); 2548 2549 /* 2550 * Check again in case we missed a preemption opportunity 2551 * between schedule and now. 2552 */ 2553 barrier(); 2554 } while (need_resched()); 2555 2556 exception_exit(prev_state); 2557} 2558 2559#endif /* CONFIG_PREEMPT */ 2560 2561int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags, 2562 void *key) 2563{ 2564 return try_to_wake_up(curr->private, mode, wake_flags); 2565} 2566EXPORT_SYMBOL(default_wake_function); 2567 2568/* 2569 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just 2570 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve 2571 * number) then we wake all the non-exclusive tasks and one exclusive task. 2572 * 2573 * There are circumstances in which we can try to wake a task which has already 2574 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns 2575 * zero in this (rare) case, and we handle it by continuing to scan the queue. 2576 */ 2577static void __wake_up_common(wait_queue_head_t *q, unsigned int mode, 2578 int nr_exclusive, int wake_flags, void *key) 2579{ 2580 wait_queue_t *curr, *next; 2581 2582 list_for_each_entry_safe(curr, next, &q->task_list, task_list) { 2583 unsigned flags = curr->flags; 2584 2585 if (curr->func(curr, mode, wake_flags, key) && 2586 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive) 2587 break; 2588 } 2589} 2590 2591/** 2592 * __wake_up - wake up threads blocked on a waitqueue. 2593 * @q: the waitqueue 2594 * @mode: which threads 2595 * @nr_exclusive: how many wake-one or wake-many threads to wake up 2596 * @key: is directly passed to the wakeup function 2597 * 2598 * It may be assumed that this function implies a write memory barrier before 2599 * changing the task state if and only if any tasks are woken up. 2600 */ 2601void __wake_up(wait_queue_head_t *q, unsigned int mode, 2602 int nr_exclusive, void *key) 2603{ 2604 unsigned long flags; 2605 2606 spin_lock_irqsave(&q->lock, flags); 2607 __wake_up_common(q, mode, nr_exclusive, 0, key); 2608 spin_unlock_irqrestore(&q->lock, flags); 2609} 2610EXPORT_SYMBOL(__wake_up); 2611 2612/* 2613 * Same as __wake_up but called with the spinlock in wait_queue_head_t held. 2614 */ 2615void __wake_up_locked(wait_queue_head_t *q, unsigned int mode, int nr) 2616{ 2617 __wake_up_common(q, mode, nr, 0, NULL); 2618} 2619EXPORT_SYMBOL_GPL(__wake_up_locked); 2620 2621void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key) 2622{ 2623 __wake_up_common(q, mode, 1, 0, key); 2624} 2625EXPORT_SYMBOL_GPL(__wake_up_locked_key); 2626 2627/** 2628 * __wake_up_sync_key - wake up threads blocked on a waitqueue. 2629 * @q: the waitqueue 2630 * @mode: which threads 2631 * @nr_exclusive: how many wake-one or wake-many threads to wake up 2632 * @key: opaque value to be passed to wakeup targets 2633 * 2634 * The sync wakeup differs that the waker knows that it will schedule 2635 * away soon, so while the target thread will be woken up, it will not 2636 * be migrated to another CPU - ie. the two threads are 'synchronized' 2637 * with each other. This can prevent needless bouncing between CPUs. 2638 * 2639 * On UP it can prevent extra preemption. 2640 * 2641 * It may be assumed that this function implies a write memory barrier before 2642 * changing the task state if and only if any tasks are woken up. 2643 */ 2644void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode, 2645 int nr_exclusive, void *key) 2646{ 2647 unsigned long flags; 2648 int wake_flags = WF_SYNC; 2649 2650 if (unlikely(!q)) 2651 return; 2652 2653 if (unlikely(!nr_exclusive)) 2654 wake_flags = 0; 2655 2656 spin_lock_irqsave(&q->lock, flags); 2657 __wake_up_common(q, mode, nr_exclusive, wake_flags, key); 2658 spin_unlock_irqrestore(&q->lock, flags); 2659} 2660EXPORT_SYMBOL_GPL(__wake_up_sync_key); 2661 2662/* 2663 * __wake_up_sync - see __wake_up_sync_key() 2664 */ 2665void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive) 2666{ 2667 __wake_up_sync_key(q, mode, nr_exclusive, NULL); 2668} 2669EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */ 2670 2671/** 2672 * complete: - signals a single thread waiting on this completion 2673 * @x: holds the state of this particular completion 2674 * 2675 * This will wake up a single thread waiting on this completion. Threads will be 2676 * awakened in the same order in which they were queued. 2677 * 2678 * See also complete_all(), wait_for_completion() and related routines. 2679 * 2680 * It may be assumed that this function implies a write memory barrier before 2681 * changing the task state if and only if any tasks are woken up. 2682 */ 2683void complete(struct completion *x) 2684{ 2685 unsigned long flags; 2686 2687 spin_lock_irqsave(&x->wait.lock, flags); 2688 x->done++; 2689 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL); 2690 spin_unlock_irqrestore(&x->wait.lock, flags); 2691} 2692EXPORT_SYMBOL(complete); 2693 2694/** 2695 * complete_all: - signals all threads waiting on this completion 2696 * @x: holds the state of this particular completion 2697 * 2698 * This will wake up all threads waiting on this particular completion event. 2699 * 2700 * It may be assumed that this function implies a write memory barrier before 2701 * changing the task state if and only if any tasks are woken up. 2702 */ 2703void complete_all(struct completion *x) 2704{ 2705 unsigned long flags; 2706 2707 spin_lock_irqsave(&x->wait.lock, flags); 2708 x->done += UINT_MAX/2; 2709 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL); 2710 spin_unlock_irqrestore(&x->wait.lock, flags); 2711} 2712EXPORT_SYMBOL(complete_all); 2713 2714static inline long __sched 2715do_wait_for_common(struct completion *x, 2716 long (*action)(long), long timeout, int state) 2717{ 2718 if (!x->done) { 2719 DECLARE_WAITQUEUE(wait, current); 2720 2721 __add_wait_queue_tail_exclusive(&x->wait, &wait); 2722 do { 2723 if (signal_pending_state(state, current)) { 2724 timeout = -ERESTARTSYS; 2725 break; 2726 } 2727 __set_current_state(state); 2728 spin_unlock_irq(&x->wait.lock); 2729 timeout = action(timeout); 2730 spin_lock_irq(&x->wait.lock); 2731 } while (!x->done && timeout); 2732 __remove_wait_queue(&x->wait, &wait); 2733 if (!x->done) 2734 return timeout; 2735 } 2736 x->done--; 2737 return timeout ?: 1; 2738} 2739 2740static inline long __sched 2741__wait_for_common(struct completion *x, 2742 long (*action)(long), long timeout, int state) 2743{ 2744 might_sleep(); 2745 2746 spin_lock_irq(&x->wait.lock); 2747 timeout = do_wait_for_common(x, action, timeout, state); 2748 spin_unlock_irq(&x->wait.lock); 2749 return timeout; 2750} 2751 2752static long __sched 2753wait_for_common(struct completion *x, long timeout, int state) 2754{ 2755 return __wait_for_common(x, schedule_timeout, timeout, state); 2756} 2757 2758static long __sched 2759wait_for_common_io(struct completion *x, long timeout, int state) 2760{ 2761 return __wait_for_common(x, io_schedule_timeout, timeout, state); 2762} 2763 2764/** 2765 * wait_for_completion: - waits for completion of a task 2766 * @x: holds the state of this particular completion 2767 * 2768 * This waits to be signaled for completion of a specific task. It is NOT 2769 * interruptible and there is no timeout. 2770 * 2771 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout 2772 * and interrupt capability. Also see complete(). 2773 */ 2774void __sched wait_for_completion(struct completion *x) 2775{ 2776 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE); 2777} 2778EXPORT_SYMBOL(wait_for_completion); 2779 2780/** 2781 * wait_for_completion_timeout: - waits for completion of a task (w/timeout) 2782 * @x: holds the state of this particular completion 2783 * @timeout: timeout value in jiffies 2784 * 2785 * This waits for either a completion of a specific task to be signaled or for a 2786 * specified timeout to expire. The timeout is in jiffies. It is not 2787 * interruptible. 2788 * 2789 * The return value is 0 if timed out, and positive (at least 1, or number of 2790 * jiffies left till timeout) if completed. 2791 */ 2792unsigned long __sched 2793wait_for_completion_timeout(struct completion *x, unsigned long timeout) 2794{ 2795 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE); 2796} 2797EXPORT_SYMBOL(wait_for_completion_timeout); 2798 2799/** 2800 * wait_for_completion_io: - waits for completion of a task 2801 * @x: holds the state of this particular completion 2802 * 2803 * This waits to be signaled for completion of a specific task. It is NOT 2804 * interruptible and there is no timeout. The caller is accounted as waiting 2805 * for IO. 2806 */ 2807void __sched wait_for_completion_io(struct completion *x) 2808{ 2809 wait_for_common_io(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE); 2810} 2811EXPORT_SYMBOL(wait_for_completion_io); 2812 2813/** 2814 * wait_for_completion_io_timeout: - waits for completion of a task (w/timeout) 2815 * @x: holds the state of this particular completion 2816 * @timeout: timeout value in jiffies 2817 * 2818 * This waits for either a completion of a specific task to be signaled or for a 2819 * specified timeout to expire. The timeout is in jiffies. It is not 2820 * interruptible. The caller is accounted as waiting for IO. 2821 * 2822 * The return value is 0 if timed out, and positive (at least 1, or number of 2823 * jiffies left till timeout) if completed. 2824 */ 2825unsigned long __sched 2826wait_for_completion_io_timeout(struct completion *x, unsigned long timeout) 2827{ 2828 return wait_for_common_io(x, timeout, TASK_UNINTERRUPTIBLE); 2829} 2830EXPORT_SYMBOL(wait_for_completion_io_timeout); 2831 2832/** 2833 * wait_for_completion_interruptible: - waits for completion of a task (w/intr) 2834 * @x: holds the state of this particular completion 2835 * 2836 * This waits for completion of a specific task to be signaled. It is 2837 * interruptible. 2838 * 2839 * The return value is -ERESTARTSYS if interrupted, 0 if completed. 2840 */ 2841int __sched wait_for_completion_interruptible(struct completion *x) 2842{ 2843 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE); 2844 if (t == -ERESTARTSYS) 2845 return t; 2846 return 0; 2847} 2848EXPORT_SYMBOL(wait_for_completion_interruptible); 2849 2850/** 2851 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr)) 2852 * @x: holds the state of this particular completion 2853 * @timeout: timeout value in jiffies 2854 * 2855 * This waits for either a completion of a specific task to be signaled or for a 2856 * specified timeout to expire. It is interruptible. The timeout is in jiffies. 2857 * 2858 * The return value is -ERESTARTSYS if interrupted, 0 if timed out, 2859 * positive (at least 1, or number of jiffies left till timeout) if completed. 2860 */ 2861long __sched 2862wait_for_completion_interruptible_timeout(struct completion *x, 2863 unsigned long timeout) 2864{ 2865 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE); 2866} 2867EXPORT_SYMBOL(wait_for_completion_interruptible_timeout); 2868 2869/** 2870 * wait_for_completion_killable: - waits for completion of a task (killable) 2871 * @x: holds the state of this particular completion 2872 * 2873 * This waits to be signaled for completion of a specific task. It can be 2874 * interrupted by a kill signal. 2875 * 2876 * The return value is -ERESTARTSYS if interrupted, 0 if completed. 2877 */ 2878int __sched wait_for_completion_killable(struct completion *x) 2879{ 2880 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE); 2881 if (t == -ERESTARTSYS) 2882 return t; 2883 return 0; 2884} 2885EXPORT_SYMBOL(wait_for_completion_killable); 2886 2887/** 2888 * wait_for_completion_killable_timeout: - waits for completion of a task (w/(to,killable)) 2889 * @x: holds the state of this particular completion 2890 * @timeout: timeout value in jiffies 2891 * 2892 * This waits for either a completion of a specific task to be 2893 * signaled or for a specified timeout to expire. It can be 2894 * interrupted by a kill signal. The timeout is in jiffies. 2895 * 2896 * The return value is -ERESTARTSYS if interrupted, 0 if timed out, 2897 * positive (at least 1, or number of jiffies left till timeout) if completed. 2898 */ 2899long __sched 2900wait_for_completion_killable_timeout(struct completion *x, 2901 unsigned long timeout) 2902{ 2903 return wait_for_common(x, timeout, TASK_KILLABLE); 2904} 2905EXPORT_SYMBOL(wait_for_completion_killable_timeout); 2906 2907/** 2908 * try_wait_for_completion - try to decrement a completion without blocking 2909 * @x: completion structure 2910 * 2911 * Returns: 0 if a decrement cannot be done without blocking 2912 * 1 if a decrement succeeded. 2913 * 2914 * If a completion is being used as a counting completion, 2915 * attempt to decrement the counter without blocking. This 2916 * enables us to avoid waiting if the resource the completion 2917 * is protecting is not available. 2918 */ 2919bool try_wait_for_completion(struct completion *x) 2920{ 2921 unsigned long flags; 2922 int ret = 1; 2923 2924 spin_lock_irqsave(&x->wait.lock, flags); 2925 if (!x->done) 2926 ret = 0; 2927 else 2928 x->done--; 2929 spin_unlock_irqrestore(&x->wait.lock, flags); 2930 return ret; 2931} 2932EXPORT_SYMBOL(try_wait_for_completion); 2933 2934/** 2935 * completion_done - Test to see if a completion has any waiters 2936 * @x: completion structure 2937 * 2938 * Returns: 0 if there are waiters (wait_for_completion() in progress) 2939 * 1 if there are no waiters. 2940 * 2941 */ 2942bool completion_done(struct completion *x) 2943{ 2944 unsigned long flags; 2945 int ret = 1; 2946 2947 spin_lock_irqsave(&x->wait.lock, flags); 2948 if (!x->done) 2949 ret = 0; 2950 spin_unlock_irqrestore(&x->wait.lock, flags); 2951 return ret; 2952} 2953EXPORT_SYMBOL(completion_done); 2954 2955static long __sched 2956sleep_on_common(wait_queue_head_t *q, int state, long timeout) 2957{ 2958 unsigned long flags; 2959 wait_queue_t wait; 2960 2961 init_waitqueue_entry(&wait, current); 2962 2963 __set_current_state(state); 2964 2965 spin_lock_irqsave(&q->lock, flags); 2966 __add_wait_queue(q, &wait); 2967 spin_unlock(&q->lock); 2968 timeout = schedule_timeout(timeout); 2969 spin_lock_irq(&q->lock); 2970 __remove_wait_queue(q, &wait); 2971 spin_unlock_irqrestore(&q->lock, flags); 2972 2973 return timeout; 2974} 2975 2976void __sched interruptible_sleep_on(wait_queue_head_t *q) 2977{ 2978 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT); 2979} 2980EXPORT_SYMBOL(interruptible_sleep_on); 2981 2982long __sched 2983interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout) 2984{ 2985 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout); 2986} 2987EXPORT_SYMBOL(interruptible_sleep_on_timeout); 2988 2989void __sched sleep_on(wait_queue_head_t *q) 2990{ 2991 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT); 2992} 2993EXPORT_SYMBOL(sleep_on); 2994 2995long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout) 2996{ 2997 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout); 2998} 2999EXPORT_SYMBOL(sleep_on_timeout); 3000 3001#ifdef CONFIG_RT_MUTEXES 3002 3003/* 3004 * rt_mutex_setprio - set the current priority of a task 3005 * @p: task 3006 * @prio: prio value (kernel-internal form) 3007 * 3008 * This function changes the 'effective' priority of a task. It does 3009 * not touch ->normal_prio like __setscheduler(). 3010 * 3011 * Used by the rt_mutex code to implement priority inheritance logic. 3012 */ 3013void rt_mutex_setprio(struct task_struct *p, int prio) 3014{ 3015 int oldprio, on_rq, running; 3016 struct rq *rq; 3017 const struct sched_class *prev_class; 3018 3019 BUG_ON(prio < 0 || prio > MAX_PRIO); 3020 3021 rq = __task_rq_lock(p); 3022 3023 /* 3024 * Idle task boosting is a nono in general. There is one 3025 * exception, when PREEMPT_RT and NOHZ is active: 3026 * 3027 * The idle task calls get_next_timer_interrupt() and holds 3028 * the timer wheel base->lock on the CPU and another CPU wants 3029 * to access the timer (probably to cancel it). We can safely 3030 * ignore the boosting request, as the idle CPU runs this code 3031 * with interrupts disabled and will complete the lock 3032 * protected section without being interrupted. So there is no 3033 * real need to boost. 3034 */ 3035 if (unlikely(p == rq->idle)) { 3036 WARN_ON(p != rq->curr); 3037 WARN_ON(p->pi_blocked_on); 3038 goto out_unlock; 3039 } 3040 3041 trace_sched_pi_setprio(p, prio); 3042 oldprio = p->prio; 3043 prev_class = p->sched_class; 3044 on_rq = p->on_rq; 3045 running = task_current(rq, p); 3046 if (on_rq) 3047 dequeue_task(rq, p, 0); 3048 if (running) 3049 p->sched_class->put_prev_task(rq, p); 3050 3051 if (rt_prio(prio)) 3052 p->sched_class = &rt_sched_class; 3053 else 3054 p->sched_class = &fair_sched_class; 3055 3056 p->prio = prio; 3057 3058 if (running) 3059 p->sched_class->set_curr_task(rq); 3060 if (on_rq) 3061 enqueue_task(rq, p, oldprio < prio ? ENQUEUE_HEAD : 0); 3062 3063 check_class_changed(rq, p, prev_class, oldprio); 3064out_unlock: 3065 __task_rq_unlock(rq); 3066} 3067#endif 3068void set_user_nice(struct task_struct *p, long nice) 3069{ 3070 int old_prio, delta, on_rq; 3071 unsigned long flags; 3072 struct rq *rq; 3073 3074 if (TASK_NICE(p) == nice || nice < -20 || nice > 19) 3075 return; 3076 /* 3077 * We have to be careful, if called from sys_setpriority(), 3078 * the task might be in the middle of scheduling on another CPU. 3079 */ 3080 rq = task_rq_lock(p, &flags); 3081 /* 3082 * The RT priorities are set via sched_setscheduler(), but we still 3083 * allow the 'normal' nice value to be set - but as expected 3084 * it wont have any effect on scheduling until the task is 3085 * SCHED_FIFO/SCHED_RR: 3086 */ 3087 if (task_has_rt_policy(p)) { 3088 p->static_prio = NICE_TO_PRIO(nice); 3089 goto out_unlock; 3090 } 3091 on_rq = p->on_rq; 3092 if (on_rq) 3093 dequeue_task(rq, p, 0); 3094 3095 p->static_prio = NICE_TO_PRIO(nice); 3096 set_load_weight(p); 3097 old_prio = p->prio; 3098 p->prio = effective_prio(p); 3099 delta = p->prio - old_prio; 3100 3101 if (on_rq) { 3102 enqueue_task(rq, p, 0); 3103 /* 3104 * If the task increased its priority or is running and 3105 * lowered its priority, then reschedule its CPU: 3106 */ 3107 if (delta < 0 || (delta > 0 && task_running(rq, p))) 3108 resched_task(rq->curr); 3109 } 3110out_unlock: 3111 task_rq_unlock(rq, p, &flags); 3112} 3113EXPORT_SYMBOL(set_user_nice); 3114 3115/* 3116 * can_nice - check if a task can reduce its nice value 3117 * @p: task 3118 * @nice: nice value 3119 */ 3120int can_nice(const struct task_struct *p, const int nice) 3121{ 3122 /* convert nice value [19,-20] to rlimit style value [1,40] */ 3123 int nice_rlim = 20 - nice; 3124 3125 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) || 3126 capable(CAP_SYS_NICE)); 3127} 3128 3129#ifdef __ARCH_WANT_SYS_NICE 3130 3131/* 3132 * sys_nice - change the priority of the current process. 3133 * @increment: priority increment 3134 * 3135 * sys_setpriority is a more generic, but much slower function that 3136 * does similar things. 3137 */ 3138SYSCALL_DEFINE1(nice, int, increment) 3139{ 3140 long nice, retval; 3141 3142 /* 3143 * Setpriority might change our priority at the same moment. 3144 * We don't have to worry. Conceptually one call occurs first 3145 * and we have a single winner. 3146 */ 3147 if (increment < -40) 3148 increment = -40; 3149 if (increment > 40) 3150 increment = 40; 3151 3152 nice = TASK_NICE(current) + increment; 3153 if (nice < -20) 3154 nice = -20; 3155 if (nice > 19) 3156 nice = 19; 3157 3158 if (increment < 0 && !can_nice(current, nice)) 3159 return -EPERM; 3160 3161 retval = security_task_setnice(current, nice); 3162 if (retval) 3163 return retval; 3164 3165 set_user_nice(current, nice); 3166 return 0; 3167} 3168 3169#endif 3170 3171/** 3172 * task_prio - return the priority value of a given task. 3173 * @p: the task in question. 3174 * 3175 * This is the priority value as seen by users in /proc. 3176 * RT tasks are offset by -200. Normal tasks are centered 3177 * around 0, value goes from -16 to +15. 3178 */ 3179int task_prio(const struct task_struct *p) 3180{ 3181 return p->prio - MAX_RT_PRIO; 3182} 3183 3184/** 3185 * task_nice - return the nice value of a given task. 3186 * @p: the task in question. 3187 */ 3188int task_nice(const struct task_struct *p) 3189{ 3190 return TASK_NICE(p); 3191} 3192EXPORT_SYMBOL(task_nice); 3193 3194/** 3195 * idle_cpu - is a given cpu idle currently? 3196 * @cpu: the processor in question. 3197 */ 3198int idle_cpu(int cpu) 3199{ 3200 struct rq *rq = cpu_rq(cpu); 3201 3202 if (rq->curr != rq->idle) 3203 return 0; 3204 3205 if (rq->nr_running) 3206 return 0; 3207 3208#ifdef CONFIG_SMP 3209 if (!llist_empty(&rq->wake_list)) 3210 return 0; 3211#endif 3212 3213 return 1; 3214} 3215 3216/** 3217 * idle_task - return the idle task for a given cpu. 3218 * @cpu: the processor in question. 3219 */ 3220struct task_struct *idle_task(int cpu) 3221{ 3222 return cpu_rq(cpu)->idle; 3223} 3224 3225/** 3226 * find_process_by_pid - find a process with a matching PID value. 3227 * @pid: the pid in question. 3228 */ 3229static struct task_struct *find_process_by_pid(pid_t pid) 3230{ 3231 return pid ? find_task_by_vpid(pid) : current; 3232} 3233 3234/* Actually do priority change: must hold rq lock. */ 3235static void 3236__setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio) 3237{ 3238 p->policy = policy; 3239 p->rt_priority = prio; 3240 p->normal_prio = normal_prio(p); 3241 /* we are holding p->pi_lock already */ 3242 p->prio = rt_mutex_getprio(p); 3243 if (rt_prio(p->prio)) 3244 p->sched_class = &rt_sched_class; 3245 else 3246 p->sched_class = &fair_sched_class; 3247 set_load_weight(p); 3248} 3249 3250/* 3251 * check the target process has a UID that matches the current process's 3252 */ 3253static bool check_same_owner(struct task_struct *p) 3254{ 3255 const struct cred *cred = current_cred(), *pcred; 3256 bool match; 3257 3258 rcu_read_lock(); 3259 pcred = __task_cred(p); 3260 match = (uid_eq(cred->euid, pcred->euid) || 3261 uid_eq(cred->euid, pcred->uid)); 3262 rcu_read_unlock(); 3263 return match; 3264} 3265 3266static int __sched_setscheduler(struct task_struct *p, int policy, 3267 const struct sched_param *param, bool user) 3268{ 3269 int retval, oldprio, oldpolicy = -1, on_rq, running; 3270 unsigned long flags; 3271 const struct sched_class *prev_class; 3272 struct rq *rq; 3273 int reset_on_fork; 3274 3275 /* may grab non-irq protected spin_locks */ 3276 BUG_ON(in_interrupt()); 3277recheck: 3278 /* double check policy once rq lock held */ 3279 if (policy < 0) { 3280 reset_on_fork = p->sched_reset_on_fork; 3281 policy = oldpolicy = p->policy; 3282 } else { 3283 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK); 3284 policy &= ~SCHED_RESET_ON_FORK; 3285 3286 if (policy != SCHED_FIFO && policy != SCHED_RR && 3287 policy != SCHED_NORMAL && policy != SCHED_BATCH && 3288 policy != SCHED_IDLE) 3289 return -EINVAL; 3290 } 3291 3292 /* 3293 * Valid priorities for SCHED_FIFO and SCHED_RR are 3294 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL, 3295 * SCHED_BATCH and SCHED_IDLE is 0. 3296 */ 3297 if (param->sched_priority < 0 || 3298 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) || 3299 (!p->mm && param->sched_priority > MAX_RT_PRIO-1)) 3300 return -EINVAL; 3301 if (rt_policy(policy) != (param->sched_priority != 0)) 3302 return -EINVAL; 3303 3304 /* 3305 * Allow unprivileged RT tasks to decrease priority: 3306 */ 3307 if (user && !capable(CAP_SYS_NICE)) { 3308 if (rt_policy(policy)) { 3309 unsigned long rlim_rtprio = 3310 task_rlimit(p, RLIMIT_RTPRIO); 3311 3312 /* can't set/change the rt policy */ 3313 if (policy != p->policy && !rlim_rtprio) 3314 return -EPERM; 3315 3316 /* can't increase priority */ 3317 if (param->sched_priority > p->rt_priority && 3318 param->sched_priority > rlim_rtprio) 3319 return -EPERM; 3320 } 3321 3322 /* 3323 * Treat SCHED_IDLE as nice 20. Only allow a switch to 3324 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it. 3325 */ 3326 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) { 3327 if (!can_nice(p, TASK_NICE(p))) 3328 return -EPERM; 3329 } 3330 3331 /* can't change other user's priorities */ 3332 if (!check_same_owner(p)) 3333 return -EPERM; 3334 3335 /* Normal users shall not reset the sched_reset_on_fork flag */ 3336 if (p->sched_reset_on_fork && !reset_on_fork) 3337 return -EPERM; 3338 } 3339 3340 if (user) { 3341 retval = security_task_setscheduler(p); 3342 if (retval) 3343 return retval; 3344 } 3345 3346 /* 3347 * make sure no PI-waiters arrive (or leave) while we are 3348 * changing the priority of the task: 3349 * 3350 * To be able to change p->policy safely, the appropriate 3351 * runqueue lock must be held. 3352 */ 3353 rq = task_rq_lock(p, &flags); 3354 3355 /* 3356 * Changing the policy of the stop threads its a very bad idea 3357 */ 3358 if (p == rq->stop) { 3359 task_rq_unlock(rq, p, &flags); 3360 return -EINVAL; 3361 } 3362 3363 /* 3364 * If not changing anything there's no need to proceed further: 3365 */ 3366 if (unlikely(policy == p->policy && (!rt_policy(policy) || 3367 param->sched_priority == p->rt_priority))) { 3368 task_rq_unlock(rq, p, &flags); 3369 return 0; 3370 } 3371 3372#ifdef CONFIG_RT_GROUP_SCHED 3373 if (user) { 3374 /* 3375 * Do not allow realtime tasks into groups that have no runtime 3376 * assigned. 3377 */ 3378 if (rt_bandwidth_enabled() && rt_policy(policy) && 3379 task_group(p)->rt_bandwidth.rt_runtime == 0 && 3380 !task_group_is_autogroup(task_group(p))) { 3381 task_rq_unlock(rq, p, &flags); 3382 return -EPERM; 3383 } 3384 } 3385#endif 3386 3387 /* recheck policy now with rq lock held */ 3388 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) { 3389 policy = oldpolicy = -1; 3390 task_rq_unlock(rq, p, &flags); 3391 goto recheck; 3392 } 3393 on_rq = p->on_rq; 3394 running = task_current(rq, p); 3395 if (on_rq) 3396 dequeue_task(rq, p, 0); 3397 if (running) 3398 p->sched_class->put_prev_task(rq, p); 3399 3400 p->sched_reset_on_fork = reset_on_fork; 3401 3402 oldprio = p->prio; 3403 prev_class = p->sched_class; 3404 __setscheduler(rq, p, policy, param->sched_priority); 3405 3406 if (running) 3407 p->sched_class->set_curr_task(rq); 3408 if (on_rq) 3409 enqueue_task(rq, p, 0); 3410 3411 check_class_changed(rq, p, prev_class, oldprio); 3412 task_rq_unlock(rq, p, &flags); 3413 3414 rt_mutex_adjust_pi(p); 3415 3416 return 0; 3417} 3418 3419/** 3420 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread. 3421 * @p: the task in question. 3422 * @policy: new policy. 3423 * @param: structure containing the new RT priority. 3424 * 3425 * NOTE that the task may be already dead. 3426 */ 3427int sched_setscheduler(struct task_struct *p, int policy, 3428 const struct sched_param *param) 3429{ 3430 return __sched_setscheduler(p, policy, param, true); 3431} 3432EXPORT_SYMBOL_GPL(sched_setscheduler); 3433 3434/** 3435 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace. 3436 * @p: the task in question. 3437 * @policy: new policy. 3438 * @param: structure containing the new RT priority. 3439 * 3440 * Just like sched_setscheduler, only don't bother checking if the 3441 * current context has permission. For example, this is needed in 3442 * stop_machine(): we create temporary high priority worker threads, 3443 * but our caller might not have that capability. 3444 */ 3445int sched_setscheduler_nocheck(struct task_struct *p, int policy, 3446 const struct sched_param *param) 3447{ 3448 return __sched_setscheduler(p, policy, param, false); 3449} 3450 3451static int 3452do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param) 3453{ 3454 struct sched_param lparam; 3455 struct task_struct *p; 3456 int retval; 3457 3458 if (!param || pid < 0) 3459 return -EINVAL; 3460 if (copy_from_user(&lparam, param, sizeof(struct sched_param))) 3461 return -EFAULT; 3462 3463 rcu_read_lock(); 3464 retval = -ESRCH; 3465 p = find_process_by_pid(pid); 3466 if (p != NULL) 3467 retval = sched_setscheduler(p, policy, &lparam); 3468 rcu_read_unlock(); 3469 3470 return retval; 3471} 3472 3473/** 3474 * sys_sched_setscheduler - set/change the scheduler policy and RT priority 3475 * @pid: the pid in question. 3476 * @policy: new policy. 3477 * @param: structure containing the new RT priority. 3478 */ 3479SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy, 3480 struct sched_param __user *, param) 3481{ 3482 /* negative values for policy are not valid */ 3483 if (policy < 0) 3484 return -EINVAL; 3485 3486 return do_sched_setscheduler(pid, policy, param); 3487} 3488 3489/** 3490 * sys_sched_setparam - set/change the RT priority of a thread 3491 * @pid: the pid in question. 3492 * @param: structure containing the new RT priority. 3493 */ 3494SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param) 3495{ 3496 return do_sched_setscheduler(pid, -1, param); 3497} 3498 3499/** 3500 * sys_sched_getscheduler - get the policy (scheduling class) of a thread 3501 * @pid: the pid in question. 3502 */ 3503SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid) 3504{ 3505 struct task_struct *p; 3506 int retval; 3507 3508 if (pid < 0) 3509 return -EINVAL; 3510 3511 retval = -ESRCH; 3512 rcu_read_lock(); 3513 p = find_process_by_pid(pid); 3514 if (p) { 3515 retval = security_task_getscheduler(p); 3516 if (!retval) 3517 retval = p->policy 3518 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0); 3519 } 3520 rcu_read_unlock(); 3521 return retval; 3522} 3523 3524/** 3525 * sys_sched_getparam - get the RT priority of a thread 3526 * @pid: the pid in question. 3527 * @param: structure containing the RT priority. 3528 */ 3529SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param) 3530{ 3531 struct sched_param lp; 3532 struct task_struct *p; 3533 int retval; 3534 3535 if (!param || pid < 0) 3536 return -EINVAL; 3537 3538 rcu_read_lock(); 3539 p = find_process_by_pid(pid); 3540 retval = -ESRCH; 3541 if (!p) 3542 goto out_unlock; 3543 3544 retval = security_task_getscheduler(p); 3545 if (retval) 3546 goto out_unlock; 3547 3548 lp.sched_priority = p->rt_priority; 3549 rcu_read_unlock(); 3550 3551 /* 3552 * This one might sleep, we cannot do it with a spinlock held ... 3553 */ 3554 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0; 3555 3556 return retval; 3557 3558out_unlock: 3559 rcu_read_unlock(); 3560 return retval; 3561} 3562 3563long sched_setaffinity(pid_t pid, const struct cpumask *in_mask) 3564{ 3565 cpumask_var_t cpus_allowed, new_mask; 3566 struct task_struct *p; 3567 int retval; 3568 3569 get_online_cpus(); 3570 rcu_read_lock(); 3571 3572 p = find_process_by_pid(pid); 3573 if (!p) { 3574 rcu_read_unlock(); 3575 put_online_cpus(); 3576 return -ESRCH; 3577 } 3578 3579 /* Prevent p going away */ 3580 get_task_struct(p); 3581 rcu_read_unlock(); 3582 3583 if (p->flags & PF_NO_SETAFFINITY) { 3584 retval = -EINVAL; 3585 goto out_put_task; 3586 } 3587 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) { 3588 retval = -ENOMEM; 3589 goto out_put_task; 3590 } 3591 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) { 3592 retval = -ENOMEM; 3593 goto out_free_cpus_allowed; 3594 } 3595 retval = -EPERM; 3596 if (!check_same_owner(p)) { 3597 rcu_read_lock(); 3598 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) { 3599 rcu_read_unlock(); 3600 goto out_unlock; 3601 } 3602 rcu_read_unlock(); 3603 } 3604 3605 retval = security_task_setscheduler(p); 3606 if (retval) 3607 goto out_unlock; 3608 3609 cpuset_cpus_allowed(p, cpus_allowed); 3610 cpumask_and(new_mask, in_mask, cpus_allowed); 3611again: 3612 retval = set_cpus_allowed_ptr(p, new_mask); 3613 3614 if (!retval) { 3615 cpuset_cpus_allowed(p, cpus_allowed); 3616 if (!cpumask_subset(new_mask, cpus_allowed)) { 3617 /* 3618 * We must have raced with a concurrent cpuset 3619 * update. Just reset the cpus_allowed to the 3620 * cpuset's cpus_allowed 3621 */ 3622 cpumask_copy(new_mask, cpus_allowed); 3623 goto again; 3624 } 3625 } 3626out_unlock: 3627 free_cpumask_var(new_mask); 3628out_free_cpus_allowed: 3629 free_cpumask_var(cpus_allowed); 3630out_put_task: 3631 put_task_struct(p); 3632 put_online_cpus(); 3633 return retval; 3634} 3635 3636static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len, 3637 struct cpumask *new_mask) 3638{ 3639 if (len < cpumask_size()) 3640 cpumask_clear(new_mask); 3641 else if (len > cpumask_size()) 3642 len = cpumask_size(); 3643 3644 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0; 3645} 3646 3647/** 3648 * sys_sched_setaffinity - set the cpu affinity of a process 3649 * @pid: pid of the process 3650 * @len: length in bytes of the bitmask pointed to by user_mask_ptr 3651 * @user_mask_ptr: user-space pointer to the new cpu mask 3652 */ 3653SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len, 3654 unsigned long __user *, user_mask_ptr) 3655{ 3656 cpumask_var_t new_mask; 3657 int retval; 3658 3659 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) 3660 return -ENOMEM; 3661 3662 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask); 3663 if (retval == 0) 3664 retval = sched_setaffinity(pid, new_mask); 3665 free_cpumask_var(new_mask); 3666 return retval; 3667} 3668 3669long sched_getaffinity(pid_t pid, struct cpumask *mask) 3670{ 3671 struct task_struct *p; 3672 unsigned long flags; 3673 int retval; 3674 3675 get_online_cpus(); 3676 rcu_read_lock(); 3677 3678 retval = -ESRCH; 3679 p = find_process_by_pid(pid); 3680 if (!p) 3681 goto out_unlock; 3682 3683 retval = security_task_getscheduler(p); 3684 if (retval) 3685 goto out_unlock; 3686 3687 raw_spin_lock_irqsave(&p->pi_lock, flags); 3688 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask); 3689 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 3690 3691out_unlock: 3692 rcu_read_unlock(); 3693 put_online_cpus(); 3694 3695 return retval; 3696} 3697 3698/** 3699 * sys_sched_getaffinity - get the cpu affinity of a process 3700 * @pid: pid of the process 3701 * @len: length in bytes of the bitmask pointed to by user_mask_ptr 3702 * @user_mask_ptr: user-space pointer to hold the current cpu mask 3703 */ 3704SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len, 3705 unsigned long __user *, user_mask_ptr) 3706{ 3707 int ret; 3708 cpumask_var_t mask; 3709 3710 if ((len * BITS_PER_BYTE) < nr_cpu_ids) 3711 return -EINVAL; 3712 if (len & (sizeof(unsigned long)-1)) 3713 return -EINVAL; 3714 3715 if (!alloc_cpumask_var(&mask, GFP_KERNEL)) 3716 return -ENOMEM; 3717 3718 ret = sched_getaffinity(pid, mask); 3719 if (ret == 0) { 3720 size_t retlen = min_t(size_t, len, cpumask_size()); 3721 3722 if (copy_to_user(user_mask_ptr, mask, retlen)) 3723 ret = -EFAULT; 3724 else 3725 ret = retlen; 3726 } 3727 free_cpumask_var(mask); 3728 3729 return ret; 3730} 3731 3732/** 3733 * sys_sched_yield - yield the current processor to other threads. 3734 * 3735 * This function yields the current CPU to other tasks. If there are no 3736 * other threads running on this CPU then this function will return. 3737 */ 3738SYSCALL_DEFINE0(sched_yield) 3739{ 3740 struct rq *rq = this_rq_lock(); 3741 3742 schedstat_inc(rq, yld_count); 3743 current->sched_class->yield_task(rq); 3744 3745 /* 3746 * Since we are going to call schedule() anyway, there's 3747 * no need to preempt or enable interrupts: 3748 */ 3749 __release(rq->lock); 3750 spin_release(&rq->lock.dep_map, 1, _THIS_IP_); 3751 do_raw_spin_unlock(&rq->lock); 3752 sched_preempt_enable_no_resched(); 3753 3754 schedule(); 3755 3756 return 0; 3757} 3758 3759static inline int should_resched(void) 3760{ 3761 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE); 3762} 3763 3764static void __cond_resched(void) 3765{ 3766 add_preempt_count(PREEMPT_ACTIVE); 3767 __schedule(); 3768 sub_preempt_count(PREEMPT_ACTIVE); 3769} 3770 3771int __sched _cond_resched(void) 3772{ 3773 if (should_resched()) { 3774 __cond_resched(); 3775 return 1; 3776 } 3777 return 0; 3778} 3779EXPORT_SYMBOL(_cond_resched); 3780 3781/* 3782 * __cond_resched_lock() - if a reschedule is pending, drop the given lock, 3783 * call schedule, and on return reacquire the lock. 3784 * 3785 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level 3786 * operations here to prevent schedule() from being called twice (once via 3787 * spin_unlock(), once by hand). 3788 */ 3789int __cond_resched_lock(spinlock_t *lock) 3790{ 3791 int resched = should_resched(); 3792 int ret = 0; 3793 3794 lockdep_assert_held(lock); 3795 3796 if (spin_needbreak(lock) || resched) { 3797 spin_unlock(lock); 3798 if (resched) 3799 __cond_resched(); 3800 else 3801 cpu_relax(); 3802 ret = 1; 3803 spin_lock(lock); 3804 } 3805 return ret; 3806} 3807EXPORT_SYMBOL(__cond_resched_lock); 3808 3809int __sched __cond_resched_softirq(void) 3810{ 3811 BUG_ON(!in_softirq()); 3812 3813 if (should_resched()) { 3814 local_bh_enable(); 3815 __cond_resched(); 3816 local_bh_disable(); 3817 return 1; 3818 } 3819 return 0; 3820} 3821EXPORT_SYMBOL(__cond_resched_softirq); 3822 3823/** 3824 * yield - yield the current processor to other threads. 3825 * 3826 * Do not ever use this function, there's a 99% chance you're doing it wrong. 3827 * 3828 * The scheduler is at all times free to pick the calling task as the most 3829 * eligible task to run, if removing the yield() call from your code breaks 3830 * it, its already broken. 3831 * 3832 * Typical broken usage is: 3833 * 3834 * while (!event) 3835 * yield(); 3836 * 3837 * where one assumes that yield() will let 'the other' process run that will 3838 * make event true. If the current task is a SCHED_FIFO task that will never 3839 * happen. Never use yield() as a progress guarantee!! 3840 * 3841 * If you want to use yield() to wait for something, use wait_event(). 3842 * If you want to use yield() to be 'nice' for others, use cond_resched(). 3843 * If you still want to use yield(), do not! 3844 */ 3845void __sched yield(void) 3846{ 3847 set_current_state(TASK_RUNNING); 3848 sys_sched_yield(); 3849} 3850EXPORT_SYMBOL(yield); 3851 3852/** 3853 * yield_to - yield the current processor to another thread in 3854 * your thread group, or accelerate that thread toward the 3855 * processor it's on. 3856 * @p: target task 3857 * @preempt: whether task preemption is allowed or not 3858 * 3859 * It's the caller's job to ensure that the target task struct 3860 * can't go away on us before we can do any checks. 3861 * 3862 * Returns: 3863 * true (>0) if we indeed boosted the target task. 3864 * false (0) if we failed to boost the target. 3865 * -ESRCH if there's no task to yield to. 3866 */ 3867bool __sched yield_to(struct task_struct *p, bool preempt) 3868{ 3869 struct task_struct *curr = current; 3870 struct rq *rq, *p_rq; 3871 unsigned long flags; 3872 int yielded = 0; 3873 3874 local_irq_save(flags); 3875 rq = this_rq(); 3876 3877again: 3878 p_rq = task_rq(p); 3879 /* 3880 * If we're the only runnable task on the rq and target rq also 3881 * has only one task, there's absolutely no point in yielding. 3882 */ 3883 if (rq->nr_running == 1 && p_rq->nr_running == 1) { 3884 yielded = -ESRCH; 3885 goto out_irq; 3886 } 3887 3888 double_rq_lock(rq, p_rq); 3889 while (task_rq(p) != p_rq) { 3890 double_rq_unlock(rq, p_rq); 3891 goto again; 3892 } 3893 3894 if (!curr->sched_class->yield_to_task) 3895 goto out_unlock; 3896 3897 if (curr->sched_class != p->sched_class) 3898 goto out_unlock; 3899 3900 if (task_running(p_rq, p) || p->state) 3901 goto out_unlock; 3902 3903 yielded = curr->sched_class->yield_to_task(rq, p, preempt); 3904 if (yielded) { 3905 schedstat_inc(rq, yld_count); 3906 /* 3907 * Make p's CPU reschedule; pick_next_entity takes care of 3908 * fairness. 3909 */ 3910 if (preempt && rq != p_rq) 3911 resched_task(p_rq->curr); 3912 } 3913 3914out_unlock: 3915 double_rq_unlock(rq, p_rq); 3916out_irq: 3917 local_irq_restore(flags); 3918 3919 if (yielded > 0) 3920 schedule(); 3921 3922 return yielded; 3923} 3924EXPORT_SYMBOL_GPL(yield_to); 3925 3926/* 3927 * This task is about to go to sleep on IO. Increment rq->nr_iowait so 3928 * that process accounting knows that this is a task in IO wait state. 3929 */ 3930void __sched io_schedule(void) 3931{ 3932 struct rq *rq = raw_rq(); 3933 3934 delayacct_blkio_start(); 3935 atomic_inc(&rq->nr_iowait); 3936 blk_flush_plug(current); 3937 current->in_iowait = 1; 3938 schedule(); 3939 current->in_iowait = 0; 3940 atomic_dec(&rq->nr_iowait); 3941 delayacct_blkio_end(); 3942} 3943EXPORT_SYMBOL(io_schedule); 3944 3945long __sched io_schedule_timeout(long timeout) 3946{ 3947 struct rq *rq = raw_rq(); 3948 long ret; 3949 3950 delayacct_blkio_start(); 3951 atomic_inc(&rq->nr_iowait); 3952 blk_flush_plug(current); 3953 current->in_iowait = 1; 3954 ret = schedule_timeout(timeout); 3955 current->in_iowait = 0; 3956 atomic_dec(&rq->nr_iowait); 3957 delayacct_blkio_end(); 3958 return ret; 3959} 3960 3961/** 3962 * sys_sched_get_priority_max - return maximum RT priority. 3963 * @policy: scheduling class. 3964 * 3965 * this syscall returns the maximum rt_priority that can be used 3966 * by a given scheduling class. 3967 */ 3968SYSCALL_DEFINE1(sched_get_priority_max, int, policy) 3969{ 3970 int ret = -EINVAL; 3971 3972 switch (policy) { 3973 case SCHED_FIFO: 3974 case SCHED_RR: 3975 ret = MAX_USER_RT_PRIO-1; 3976 break; 3977 case SCHED_NORMAL: 3978 case SCHED_BATCH: 3979 case SCHED_IDLE: 3980 ret = 0; 3981 break; 3982 } 3983 return ret; 3984} 3985 3986/** 3987 * sys_sched_get_priority_min - return minimum RT priority. 3988 * @policy: scheduling class. 3989 * 3990 * this syscall returns the minimum rt_priority that can be used 3991 * by a given scheduling class. 3992 */ 3993SYSCALL_DEFINE1(sched_get_priority_min, int, policy) 3994{ 3995 int ret = -EINVAL; 3996 3997 switch (policy) { 3998 case SCHED_FIFO: 3999 case SCHED_RR: 4000 ret = 1; 4001 break; 4002 case SCHED_NORMAL: 4003 case SCHED_BATCH: 4004 case SCHED_IDLE: 4005 ret = 0; 4006 } 4007 return ret; 4008} 4009 4010/** 4011 * sys_sched_rr_get_interval - return the default timeslice of a process. 4012 * @pid: pid of the process. 4013 * @interval: userspace pointer to the timeslice value. 4014 * 4015 * this syscall writes the default timeslice value of a given process 4016 * into the user-space timespec buffer. A value of '0' means infinity. 4017 */ 4018SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid, 4019 struct timespec __user *, interval) 4020{ 4021 struct task_struct *p; 4022 unsigned int time_slice; 4023 unsigned long flags; 4024 struct rq *rq; 4025 int retval; 4026 struct timespec t; 4027 4028 if (pid < 0) 4029 return -EINVAL; 4030 4031 retval = -ESRCH; 4032 rcu_read_lock(); 4033 p = find_process_by_pid(pid); 4034 if (!p) 4035 goto out_unlock; 4036 4037 retval = security_task_getscheduler(p); 4038 if (retval) 4039 goto out_unlock; 4040 4041 rq = task_rq_lock(p, &flags); 4042 time_slice = p->sched_class->get_rr_interval(rq, p); 4043 task_rq_unlock(rq, p, &flags); 4044 4045 rcu_read_unlock(); 4046 jiffies_to_timespec(time_slice, &t); 4047 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0; 4048 return retval; 4049 4050out_unlock: 4051 rcu_read_unlock(); 4052 return retval; 4053} 4054 4055static const char stat_nam[] = TASK_STATE_TO_CHAR_STR; 4056 4057void sched_show_task(struct task_struct *p) 4058{ 4059 unsigned long free = 0; 4060 int ppid; 4061 unsigned state; 4062 4063 state = p->state ? __ffs(p->state) + 1 : 0; 4064 printk(KERN_INFO "%-15.15s %c", p->comm, 4065 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?'); 4066#if BITS_PER_LONG == 32 4067 if (state == TASK_RUNNING) 4068 printk(KERN_CONT " running "); 4069 else 4070 printk(KERN_CONT " %08lx ", thread_saved_pc(p)); 4071#else 4072 if (state == TASK_RUNNING) 4073 printk(KERN_CONT " running task "); 4074 else 4075 printk(KERN_CONT " %016lx ", thread_saved_pc(p)); 4076#endif 4077#ifdef CONFIG_DEBUG_STACK_USAGE 4078 free = stack_not_used(p); 4079#endif 4080 rcu_read_lock(); 4081 ppid = task_pid_nr(rcu_dereference(p->real_parent)); 4082 rcu_read_unlock(); 4083 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free, 4084 task_pid_nr(p), ppid, 4085 (unsigned long)task_thread_info(p)->flags); 4086 4087 print_worker_info(KERN_INFO, p); 4088 show_stack(p, NULL); 4089} 4090 4091void show_state_filter(unsigned long state_filter) 4092{ 4093 struct task_struct *g, *p; 4094 4095#if BITS_PER_LONG == 32 4096 printk(KERN_INFO 4097 " task PC stack pid father\n"); 4098#else 4099 printk(KERN_INFO 4100 " task PC stack pid father\n"); 4101#endif 4102 rcu_read_lock(); 4103 do_each_thread(g, p) { 4104 /* 4105 * reset the NMI-timeout, listing all files on a slow 4106 * console might take a lot of time: 4107 */ 4108 touch_nmi_watchdog(); 4109 if (!state_filter || (p->state & state_filter)) 4110 sched_show_task(p); 4111 } while_each_thread(g, p); 4112 4113 touch_all_softlockup_watchdogs(); 4114 4115#ifdef CONFIG_SCHED_DEBUG 4116 sysrq_sched_debug_show(); 4117#endif 4118 rcu_read_unlock(); 4119 /* 4120 * Only show locks if all tasks are dumped: 4121 */ 4122 if (!state_filter) 4123 debug_show_all_locks(); 4124} 4125 4126void __cpuinit init_idle_bootup_task(struct task_struct *idle) 4127{ 4128 idle->sched_class = &idle_sched_class; 4129} 4130 4131/** 4132 * init_idle - set up an idle thread for a given CPU 4133 * @idle: task in question 4134 * @cpu: cpu the idle task belongs to 4135 * 4136 * NOTE: this function does not set the idle thread's NEED_RESCHED 4137 * flag, to make booting more robust. 4138 */ 4139void __cpuinit init_idle(struct task_struct *idle, int cpu) 4140{ 4141 struct rq *rq = cpu_rq(cpu); 4142 unsigned long flags; 4143 4144 raw_spin_lock_irqsave(&rq->lock, flags); 4145 4146 __sched_fork(idle); 4147 idle->state = TASK_RUNNING; 4148 idle->se.exec_start = sched_clock(); 4149 4150 do_set_cpus_allowed(idle, cpumask_of(cpu)); 4151 /* 4152 * We're having a chicken and egg problem, even though we are 4153 * holding rq->lock, the cpu isn't yet set to this cpu so the 4154 * lockdep check in task_group() will fail. 4155 * 4156 * Similar case to sched_fork(). / Alternatively we could 4157 * use task_rq_lock() here and obtain the other rq->lock. 4158 * 4159 * Silence PROVE_RCU 4160 */ 4161 rcu_read_lock(); 4162 __set_task_cpu(idle, cpu); 4163 rcu_read_unlock(); 4164 4165 rq->curr = rq->idle = idle; 4166#if defined(CONFIG_SMP) 4167 idle->on_cpu = 1; 4168#endif 4169 raw_spin_unlock_irqrestore(&rq->lock, flags); 4170 4171 /* Set the preempt count _outside_ the spinlocks! */ 4172 task_thread_info(idle)->preempt_count = 0; 4173 4174 /* 4175 * The idle tasks have their own, simple scheduling class: 4176 */ 4177 idle->sched_class = &idle_sched_class; 4178 ftrace_graph_init_idle_task(idle, cpu); 4179 vtime_init_idle(idle); 4180#if defined(CONFIG_SMP) 4181 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu); 4182#endif 4183} 4184 4185#ifdef CONFIG_SMP 4186void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask) 4187{ 4188 if (p->sched_class && p->sched_class->set_cpus_allowed) 4189 p->sched_class->set_cpus_allowed(p, new_mask); 4190 4191 cpumask_copy(&p->cpus_allowed, new_mask); 4192 p->nr_cpus_allowed = cpumask_weight(new_mask); 4193} 4194 4195/* 4196 * This is how migration works: 4197 * 4198 * 1) we invoke migration_cpu_stop() on the target CPU using 4199 * stop_one_cpu(). 4200 * 2) stopper starts to run (implicitly forcing the migrated thread 4201 * off the CPU) 4202 * 3) it checks whether the migrated task is still in the wrong runqueue. 4203 * 4) if it's in the wrong runqueue then the migration thread removes 4204 * it and puts it into the right queue. 4205 * 5) stopper completes and stop_one_cpu() returns and the migration 4206 * is done. 4207 */ 4208 4209/* 4210 * Change a given task's CPU affinity. Migrate the thread to a 4211 * proper CPU and schedule it away if the CPU it's executing on 4212 * is removed from the allowed bitmask. 4213 * 4214 * NOTE: the caller must have a valid reference to the task, the 4215 * task must not exit() & deallocate itself prematurely. The 4216 * call is not atomic; no spinlocks may be held. 4217 */ 4218int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask) 4219{ 4220 unsigned long flags; 4221 struct rq *rq; 4222 unsigned int dest_cpu; 4223 int ret = 0; 4224 4225 rq = task_rq_lock(p, &flags); 4226 4227 if (cpumask_equal(&p->cpus_allowed, new_mask)) 4228 goto out; 4229 4230 if (!cpumask_intersects(new_mask, cpu_active_mask)) { 4231 ret = -EINVAL; 4232 goto out; 4233 } 4234 4235 do_set_cpus_allowed(p, new_mask); 4236 4237 /* Can the task run on the task's current CPU? If so, we're done */ 4238 if (cpumask_test_cpu(task_cpu(p), new_mask)) 4239 goto out; 4240 4241 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask); 4242 if (p->on_rq) { 4243 struct migration_arg arg = { p, dest_cpu }; 4244 /* Need help from migration thread: drop lock and wait. */ 4245 task_rq_unlock(rq, p, &flags); 4246 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg); 4247 tlb_migrate_finish(p->mm); 4248 return 0; 4249 } 4250out: 4251 task_rq_unlock(rq, p, &flags); 4252 4253 return ret; 4254} 4255EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr); 4256 4257/* 4258 * Move (not current) task off this cpu, onto dest cpu. We're doing 4259 * this because either it can't run here any more (set_cpus_allowed() 4260 * away from this CPU, or CPU going down), or because we're 4261 * attempting to rebalance this task on exec (sched_exec). 4262 * 4263 * So we race with normal scheduler movements, but that's OK, as long 4264 * as the task is no longer on this CPU. 4265 * 4266 * Returns non-zero if task was successfully migrated. 4267 */ 4268static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu) 4269{ 4270 struct rq *rq_dest, *rq_src; 4271 int ret = 0; 4272 4273 if (unlikely(!cpu_active(dest_cpu))) 4274 return ret; 4275 4276 rq_src = cpu_rq(src_cpu); 4277 rq_dest = cpu_rq(dest_cpu); 4278 4279 raw_spin_lock(&p->pi_lock); 4280 double_rq_lock(rq_src, rq_dest); 4281 /* Already moved. */ 4282 if (task_cpu(p) != src_cpu) 4283 goto done; 4284 /* Affinity changed (again). */ 4285 if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p))) 4286 goto fail; 4287 4288 /* 4289 * If we're not on a rq, the next wake-up will ensure we're 4290 * placed properly. 4291 */ 4292 if (p->on_rq) { 4293 dequeue_task(rq_src, p, 0); 4294 set_task_cpu(p, dest_cpu); 4295 enqueue_task(rq_dest, p, 0); 4296 check_preempt_curr(rq_dest, p, 0); 4297 } 4298done: 4299 ret = 1; 4300fail: 4301 double_rq_unlock(rq_src, rq_dest); 4302 raw_spin_unlock(&p->pi_lock); 4303 return ret; 4304} 4305 4306/* 4307 * migration_cpu_stop - this will be executed by a highprio stopper thread 4308 * and performs thread migration by bumping thread off CPU then 4309 * 'pushing' onto another runqueue. 4310 */ 4311static int migration_cpu_stop(void *data) 4312{ 4313 struct migration_arg *arg = data; 4314 4315 /* 4316 * The original target cpu might have gone down and we might 4317 * be on another cpu but it doesn't matter. 4318 */ 4319 local_irq_disable(); 4320 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu); 4321 local_irq_enable(); 4322 return 0; 4323} 4324 4325#ifdef CONFIG_HOTPLUG_CPU 4326 4327/* 4328 * Ensures that the idle task is using init_mm right before its cpu goes 4329 * offline. 4330 */ 4331void idle_task_exit(void) 4332{ 4333 struct mm_struct *mm = current->active_mm; 4334 4335 BUG_ON(cpu_online(smp_processor_id())); 4336 4337 if (mm != &init_mm) 4338 switch_mm(mm, &init_mm, current); 4339 mmdrop(mm); 4340} 4341 4342/* 4343 * Since this CPU is going 'away' for a while, fold any nr_active delta 4344 * we might have. Assumes we're called after migrate_tasks() so that the 4345 * nr_active count is stable. 4346 * 4347 * Also see the comment "Global load-average calculations". 4348 */ 4349static void calc_load_migrate(struct rq *rq) 4350{ 4351 long delta = calc_load_fold_active(rq); 4352 if (delta) 4353 atomic_long_add(delta, &calc_load_tasks); 4354} 4355 4356/* 4357 * Migrate all tasks from the rq, sleeping tasks will be migrated by 4358 * try_to_wake_up()->select_task_rq(). 4359 * 4360 * Called with rq->lock held even though we'er in stop_machine() and 4361 * there's no concurrency possible, we hold the required locks anyway 4362 * because of lock validation efforts. 4363 */ 4364static void migrate_tasks(unsigned int dead_cpu) 4365{ 4366 struct rq *rq = cpu_rq(dead_cpu); 4367 struct task_struct *next, *stop = rq->stop; 4368 int dest_cpu; 4369 4370 /* 4371 * Fudge the rq selection such that the below task selection loop 4372 * doesn't get stuck on the currently eligible stop task. 4373 * 4374 * We're currently inside stop_machine() and the rq is either stuck 4375 * in the stop_machine_cpu_stop() loop, or we're executing this code, 4376 * either way we should never end up calling schedule() until we're 4377 * done here. 4378 */ 4379 rq->stop = NULL; 4380 4381 for ( ; ; ) { 4382 /* 4383 * There's this thread running, bail when that's the only 4384 * remaining thread. 4385 */ 4386 if (rq->nr_running == 1) 4387 break; 4388 4389 next = pick_next_task(rq); 4390 BUG_ON(!next); 4391 next->sched_class->put_prev_task(rq, next); 4392 4393 /* Find suitable destination for @next, with force if needed. */ 4394 dest_cpu = select_fallback_rq(dead_cpu, next); 4395 raw_spin_unlock(&rq->lock); 4396 4397 __migrate_task(next, dead_cpu, dest_cpu); 4398 4399 raw_spin_lock(&rq->lock); 4400 } 4401 4402 rq->stop = stop; 4403} 4404 4405#endif /* CONFIG_HOTPLUG_CPU */ 4406 4407#if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL) 4408 4409static struct ctl_table sd_ctl_dir[] = { 4410 { 4411 .procname = "sched_domain", 4412 .mode = 0555, 4413 }, 4414 {} 4415}; 4416 4417static struct ctl_table sd_ctl_root[] = { 4418 { 4419 .procname = "kernel", 4420 .mode = 0555, 4421 .child = sd_ctl_dir, 4422 }, 4423 {} 4424}; 4425 4426static struct ctl_table *sd_alloc_ctl_entry(int n) 4427{ 4428 struct ctl_table *entry = 4429 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL); 4430 4431 return entry; 4432} 4433 4434static void sd_free_ctl_entry(struct ctl_table **tablep) 4435{ 4436 struct ctl_table *entry; 4437 4438 /* 4439 * In the intermediate directories, both the child directory and 4440 * procname are dynamically allocated and could fail but the mode 4441 * will always be set. In the lowest directory the names are 4442 * static strings and all have proc handlers. 4443 */ 4444 for (entry = *tablep; entry->mode; entry++) { 4445 if (entry->child) 4446 sd_free_ctl_entry(&entry->child); 4447 if (entry->proc_handler == NULL) 4448 kfree(entry->procname); 4449 } 4450 4451 kfree(*tablep); 4452 *tablep = NULL; 4453} 4454 4455static int min_load_idx = 0; 4456static int max_load_idx = CPU_LOAD_IDX_MAX-1; 4457 4458static void 4459set_table_entry(struct ctl_table *entry, 4460 const char *procname, void *data, int maxlen, 4461 umode_t mode, proc_handler *proc_handler, 4462 bool load_idx) 4463{ 4464 entry->procname = procname; 4465 entry->data = data; 4466 entry->maxlen = maxlen; 4467 entry->mode = mode; 4468 entry->proc_handler = proc_handler; 4469 4470 if (load_idx) { 4471 entry->extra1 = &min_load_idx; 4472 entry->extra2 = &max_load_idx; 4473 } 4474} 4475 4476static struct ctl_table * 4477sd_alloc_ctl_domain_table(struct sched_domain *sd) 4478{ 4479 struct ctl_table *table = sd_alloc_ctl_entry(13); 4480 4481 if (table == NULL) 4482 return NULL; 4483 4484 set_table_entry(&table[0], "min_interval", &sd->min_interval, 4485 sizeof(long), 0644, proc_doulongvec_minmax, false); 4486 set_table_entry(&table[1], "max_interval", &sd->max_interval, 4487 sizeof(long), 0644, proc_doulongvec_minmax, false); 4488 set_table_entry(&table[2], "busy_idx", &sd->busy_idx, 4489 sizeof(int), 0644, proc_dointvec_minmax, true); 4490 set_table_entry(&table[3], "idle_idx", &sd->idle_idx, 4491 sizeof(int), 0644, proc_dointvec_minmax, true); 4492 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx, 4493 sizeof(int), 0644, proc_dointvec_minmax, true); 4494 set_table_entry(&table[5], "wake_idx", &sd->wake_idx, 4495 sizeof(int), 0644, proc_dointvec_minmax, true); 4496 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx, 4497 sizeof(int), 0644, proc_dointvec_minmax, true); 4498 set_table_entry(&table[7], "busy_factor", &sd->busy_factor, 4499 sizeof(int), 0644, proc_dointvec_minmax, false); 4500 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct, 4501 sizeof(int), 0644, proc_dointvec_minmax, false); 4502 set_table_entry(&table[9], "cache_nice_tries", 4503 &sd->cache_nice_tries, 4504 sizeof(int), 0644, proc_dointvec_minmax, false); 4505 set_table_entry(&table[10], "flags", &sd->flags, 4506 sizeof(int), 0644, proc_dointvec_minmax, false); 4507 set_table_entry(&table[11], "name", sd->name, 4508 CORENAME_MAX_SIZE, 0444, proc_dostring, false); 4509 /* &table[12] is terminator */ 4510 4511 return table; 4512} 4513 4514static ctl_table *sd_alloc_ctl_cpu_table(int cpu) 4515{ 4516 struct ctl_table *entry, *table; 4517 struct sched_domain *sd; 4518 int domain_num = 0, i; 4519 char buf[32]; 4520 4521 for_each_domain(cpu, sd) 4522 domain_num++; 4523 entry = table = sd_alloc_ctl_entry(domain_num + 1); 4524 if (table == NULL) 4525 return NULL; 4526 4527 i = 0; 4528 for_each_domain(cpu, sd) { 4529 snprintf(buf, 32, "domain%d", i); 4530 entry->procname = kstrdup(buf, GFP_KERNEL); 4531 entry->mode = 0555; 4532 entry->child = sd_alloc_ctl_domain_table(sd); 4533 entry++; 4534 i++; 4535 } 4536 return table; 4537} 4538 4539static struct ctl_table_header *sd_sysctl_header; 4540static void register_sched_domain_sysctl(void) 4541{ 4542 int i, cpu_num = num_possible_cpus(); 4543 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1); 4544 char buf[32]; 4545 4546 WARN_ON(sd_ctl_dir[0].child); 4547 sd_ctl_dir[0].child = entry; 4548 4549 if (entry == NULL) 4550 return; 4551 4552 for_each_possible_cpu(i) { 4553 snprintf(buf, 32, "cpu%d", i); 4554 entry->procname = kstrdup(buf, GFP_KERNEL); 4555 entry->mode = 0555; 4556 entry->child = sd_alloc_ctl_cpu_table(i); 4557 entry++; 4558 } 4559 4560 WARN_ON(sd_sysctl_header); 4561 sd_sysctl_header = register_sysctl_table(sd_ctl_root); 4562} 4563 4564/* may be called multiple times per register */ 4565static void unregister_sched_domain_sysctl(void) 4566{ 4567 if (sd_sysctl_header) 4568 unregister_sysctl_table(sd_sysctl_header); 4569 sd_sysctl_header = NULL; 4570 if (sd_ctl_dir[0].child) 4571 sd_free_ctl_entry(&sd_ctl_dir[0].child); 4572} 4573#else 4574static void register_sched_domain_sysctl(void) 4575{ 4576} 4577static void unregister_sched_domain_sysctl(void) 4578{ 4579} 4580#endif 4581 4582static void set_rq_online(struct rq *rq) 4583{ 4584 if (!rq->online) { 4585 const struct sched_class *class; 4586 4587 cpumask_set_cpu(rq->cpu, rq->rd->online); 4588 rq->online = 1; 4589 4590 for_each_class(class) { 4591 if (class->rq_online) 4592 class->rq_online(rq); 4593 } 4594 } 4595} 4596 4597static void set_rq_offline(struct rq *rq) 4598{ 4599 if (rq->online) { 4600 const struct sched_class *class; 4601 4602 for_each_class(class) { 4603 if (class->rq_offline) 4604 class->rq_offline(rq); 4605 } 4606 4607 cpumask_clear_cpu(rq->cpu, rq->rd->online); 4608 rq->online = 0; 4609 } 4610} 4611 4612/* 4613 * migration_call - callback that gets triggered when a CPU is added. 4614 * Here we can start up the necessary migration thread for the new CPU. 4615 */ 4616static int __cpuinit 4617migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu) 4618{ 4619 int cpu = (long)hcpu; 4620 unsigned long flags; 4621 struct rq *rq = cpu_rq(cpu); 4622 4623 switch (action & ~CPU_TASKS_FROZEN) { 4624 4625 case CPU_UP_PREPARE: 4626 rq->calc_load_update = calc_load_update; 4627 break; 4628 4629 case CPU_ONLINE: 4630 /* Update our root-domain */ 4631 raw_spin_lock_irqsave(&rq->lock, flags); 4632 if (rq->rd) { 4633 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span)); 4634 4635 set_rq_online(rq); 4636 } 4637 raw_spin_unlock_irqrestore(&rq->lock, flags); 4638 break; 4639 4640#ifdef CONFIG_HOTPLUG_CPU 4641 case CPU_DYING: 4642 sched_ttwu_pending(); 4643 /* Update our root-domain */ 4644 raw_spin_lock_irqsave(&rq->lock, flags); 4645 if (rq->rd) { 4646 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span)); 4647 set_rq_offline(rq); 4648 } 4649 migrate_tasks(cpu); 4650 BUG_ON(rq->nr_running != 1); /* the migration thread */ 4651 raw_spin_unlock_irqrestore(&rq->lock, flags); 4652 break; 4653 4654 case CPU_DEAD: 4655 calc_load_migrate(rq); 4656 break; 4657#endif 4658 } 4659 4660 update_max_interval(); 4661 4662 return NOTIFY_OK; 4663} 4664 4665/* 4666 * Register at high priority so that task migration (migrate_all_tasks) 4667 * happens before everything else. This has to be lower priority than 4668 * the notifier in the perf_event subsystem, though. 4669 */ 4670static struct notifier_block __cpuinitdata migration_notifier = { 4671 .notifier_call = migration_call, 4672 .priority = CPU_PRI_MIGRATION, 4673}; 4674 4675static int __cpuinit sched_cpu_active(struct notifier_block *nfb, 4676 unsigned long action, void *hcpu) 4677{ 4678 switch (action & ~CPU_TASKS_FROZEN) { 4679 case CPU_STARTING: 4680 case CPU_DOWN_FAILED: 4681 set_cpu_active((long)hcpu, true); 4682 return NOTIFY_OK; 4683 default: 4684 return NOTIFY_DONE; 4685 } 4686} 4687 4688static int __cpuinit sched_cpu_inactive(struct notifier_block *nfb, 4689 unsigned long action, void *hcpu) 4690{ 4691 switch (action & ~CPU_TASKS_FROZEN) { 4692 case CPU_DOWN_PREPARE: 4693 set_cpu_active((long)hcpu, false); 4694 return NOTIFY_OK; 4695 default: 4696 return NOTIFY_DONE; 4697 } 4698} 4699 4700static int __init migration_init(void) 4701{ 4702 void *cpu = (void *)(long)smp_processor_id(); 4703 int err; 4704 4705 /* Initialize migration for the boot CPU */ 4706 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu); 4707 BUG_ON(err == NOTIFY_BAD); 4708 migration_call(&migration_notifier, CPU_ONLINE, cpu); 4709 register_cpu_notifier(&migration_notifier); 4710 4711 /* Register cpu active notifiers */ 4712 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE); 4713 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE); 4714 4715 return 0; 4716} 4717early_initcall(migration_init); 4718#endif 4719 4720#ifdef CONFIG_SMP 4721 4722static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */ 4723 4724#ifdef CONFIG_SCHED_DEBUG 4725 4726static __read_mostly int sched_debug_enabled; 4727 4728static int __init sched_debug_setup(char *str) 4729{ 4730 sched_debug_enabled = 1; 4731 4732 return 0; 4733} 4734early_param("sched_debug", sched_debug_setup); 4735 4736static inline bool sched_debug(void) 4737{ 4738 return sched_debug_enabled; 4739} 4740 4741static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level, 4742 struct cpumask *groupmask) 4743{ 4744 struct sched_group *group = sd->groups; 4745 char str[256]; 4746 4747 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd)); 4748 cpumask_clear(groupmask); 4749 4750 printk(KERN_DEBUG "%*s domain %d: ", level, "", level); 4751 4752 if (!(sd->flags & SD_LOAD_BALANCE)) { 4753 printk("does not load-balance\n"); 4754 if (sd->parent) 4755 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain" 4756 " has parent"); 4757 return -1; 4758 } 4759 4760 printk(KERN_CONT "span %s level %s\n", str, sd->name); 4761 4762 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) { 4763 printk(KERN_ERR "ERROR: domain->span does not contain " 4764 "CPU%d\n", cpu); 4765 } 4766 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) { 4767 printk(KERN_ERR "ERROR: domain->groups does not contain" 4768 " CPU%d\n", cpu); 4769 } 4770 4771 printk(KERN_DEBUG "%*s groups:", level + 1, ""); 4772 do { 4773 if (!group) { 4774 printk("\n"); 4775 printk(KERN_ERR "ERROR: group is NULL\n"); 4776 break; 4777 } 4778 4779 /* 4780 * Even though we initialize ->power to something semi-sane, 4781 * we leave power_orig unset. This allows us to detect if 4782 * domain iteration is still funny without causing /0 traps. 4783 */ 4784 if (!group->sgp->power_orig) { 4785 printk(KERN_CONT "\n"); 4786 printk(KERN_ERR "ERROR: domain->cpu_power not " 4787 "set\n"); 4788 break; 4789 } 4790 4791 if (!cpumask_weight(sched_group_cpus(group))) { 4792 printk(KERN_CONT "\n"); 4793 printk(KERN_ERR "ERROR: empty group\n"); 4794 break; 4795 } 4796 4797 if (!(sd->flags & SD_OVERLAP) && 4798 cpumask_intersects(groupmask, sched_group_cpus(group))) { 4799 printk(KERN_CONT "\n"); 4800 printk(KERN_ERR "ERROR: repeated CPUs\n"); 4801 break; 4802 } 4803 4804 cpumask_or(groupmask, groupmask, sched_group_cpus(group)); 4805 4806 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group)); 4807 4808 printk(KERN_CONT " %s", str); 4809 if (group->sgp->power != SCHED_POWER_SCALE) { 4810 printk(KERN_CONT " (cpu_power = %d)", 4811 group->sgp->power); 4812 } 4813 4814 group = group->next; 4815 } while (group != sd->groups); 4816 printk(KERN_CONT "\n"); 4817 4818 if (!cpumask_equal(sched_domain_span(sd), groupmask)) 4819 printk(KERN_ERR "ERROR: groups don't span domain->span\n"); 4820 4821 if (sd->parent && 4822 !cpumask_subset(groupmask, sched_domain_span(sd->parent))) 4823 printk(KERN_ERR "ERROR: parent span is not a superset " 4824 "of domain->span\n"); 4825 return 0; 4826} 4827 4828static void sched_domain_debug(struct sched_domain *sd, int cpu) 4829{ 4830 int level = 0; 4831 4832 if (!sched_debug_enabled) 4833 return; 4834 4835 if (!sd) { 4836 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu); 4837 return; 4838 } 4839 4840 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu); 4841 4842 for (;;) { 4843 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask)) 4844 break; 4845 level++; 4846 sd = sd->parent; 4847 if (!sd) 4848 break; 4849 } 4850} 4851#else /* !CONFIG_SCHED_DEBUG */ 4852# define sched_domain_debug(sd, cpu) do { } while (0) 4853static inline bool sched_debug(void) 4854{ 4855 return false; 4856} 4857#endif /* CONFIG_SCHED_DEBUG */ 4858 4859static int sd_degenerate(struct sched_domain *sd) 4860{ 4861 if (cpumask_weight(sched_domain_span(sd)) == 1) 4862 return 1; 4863 4864 /* Following flags need at least 2 groups */ 4865 if (sd->flags & (SD_LOAD_BALANCE | 4866 SD_BALANCE_NEWIDLE | 4867 SD_BALANCE_FORK | 4868 SD_BALANCE_EXEC | 4869 SD_SHARE_CPUPOWER | 4870 SD_SHARE_PKG_RESOURCES)) { 4871 if (sd->groups != sd->groups->next) 4872 return 0; 4873 } 4874 4875 /* Following flags don't use groups */ 4876 if (sd->flags & (SD_WAKE_AFFINE)) 4877 return 0; 4878 4879 return 1; 4880} 4881 4882static int 4883sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent) 4884{ 4885 unsigned long cflags = sd->flags, pflags = parent->flags; 4886 4887 if (sd_degenerate(parent)) 4888 return 1; 4889 4890 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent))) 4891 return 0; 4892 4893 /* Flags needing groups don't count if only 1 group in parent */ 4894 if (parent->groups == parent->groups->next) { 4895 pflags &= ~(SD_LOAD_BALANCE | 4896 SD_BALANCE_NEWIDLE | 4897 SD_BALANCE_FORK | 4898 SD_BALANCE_EXEC | 4899 SD_SHARE_CPUPOWER | 4900 SD_SHARE_PKG_RESOURCES); 4901 if (nr_node_ids == 1) 4902 pflags &= ~SD_SERIALIZE; 4903 } 4904 if (~cflags & pflags) 4905 return 0; 4906 4907 return 1; 4908} 4909 4910static void free_rootdomain(struct rcu_head *rcu) 4911{ 4912 struct root_domain *rd = container_of(rcu, struct root_domain, rcu); 4913 4914 cpupri_cleanup(&rd->cpupri); 4915 free_cpumask_var(rd->rto_mask); 4916 free_cpumask_var(rd->online); 4917 free_cpumask_var(rd->span); 4918 kfree(rd); 4919} 4920 4921static void rq_attach_root(struct rq *rq, struct root_domain *rd) 4922{ 4923 struct root_domain *old_rd = NULL; 4924 unsigned long flags; 4925 4926 raw_spin_lock_irqsave(&rq->lock, flags); 4927 4928 if (rq->rd) { 4929 old_rd = rq->rd; 4930 4931 if (cpumask_test_cpu(rq->cpu, old_rd->online)) 4932 set_rq_offline(rq); 4933 4934 cpumask_clear_cpu(rq->cpu, old_rd->span); 4935 4936 /* 4937 * If we dont want to free the old_rt yet then 4938 * set old_rd to NULL to skip the freeing later 4939 * in this function: 4940 */ 4941 if (!atomic_dec_and_test(&old_rd->refcount)) 4942 old_rd = NULL; 4943 } 4944 4945 atomic_inc(&rd->refcount); 4946 rq->rd = rd; 4947 4948 cpumask_set_cpu(rq->cpu, rd->span); 4949 if (cpumask_test_cpu(rq->cpu, cpu_active_mask)) 4950 set_rq_online(rq); 4951 4952 raw_spin_unlock_irqrestore(&rq->lock, flags); 4953 4954 if (old_rd) 4955 call_rcu_sched(&old_rd->rcu, free_rootdomain); 4956} 4957 4958static int init_rootdomain(struct root_domain *rd) 4959{ 4960 memset(rd, 0, sizeof(*rd)); 4961 4962 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL)) 4963 goto out; 4964 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL)) 4965 goto free_span; 4966 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL)) 4967 goto free_online; 4968 4969 if (cpupri_init(&rd->cpupri) != 0) 4970 goto free_rto_mask; 4971 return 0; 4972 4973free_rto_mask: 4974 free_cpumask_var(rd->rto_mask); 4975free_online: 4976 free_cpumask_var(rd->online); 4977free_span: 4978 free_cpumask_var(rd->span); 4979out: 4980 return -ENOMEM; 4981} 4982 4983/* 4984 * By default the system creates a single root-domain with all cpus as 4985 * members (mimicking the global state we have today). 4986 */ 4987struct root_domain def_root_domain; 4988 4989static void init_defrootdomain(void) 4990{ 4991 init_rootdomain(&def_root_domain); 4992 4993 atomic_set(&def_root_domain.refcount, 1); 4994} 4995 4996static struct root_domain *alloc_rootdomain(void) 4997{ 4998 struct root_domain *rd; 4999 5000 rd = kmalloc(sizeof(*rd), GFP_KERNEL); 5001 if (!rd) 5002 return NULL; 5003 5004 if (init_rootdomain(rd) != 0) { 5005 kfree(rd); 5006 return NULL; 5007 } 5008 5009 return rd; 5010} 5011 5012static void free_sched_groups(struct sched_group *sg, int free_sgp) 5013{ 5014 struct sched_group *tmp, *first; 5015 5016 if (!sg) 5017 return; 5018 5019 first = sg; 5020 do { 5021 tmp = sg->next; 5022 5023 if (free_sgp && atomic_dec_and_test(&sg->sgp->ref)) 5024 kfree(sg->sgp); 5025 5026 kfree(sg); 5027 sg = tmp; 5028 } while (sg != first); 5029} 5030 5031static void free_sched_domain(struct rcu_head *rcu) 5032{ 5033 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu); 5034 5035 /* 5036 * If its an overlapping domain it has private groups, iterate and 5037 * nuke them all. 5038 */ 5039 if (sd->flags & SD_OVERLAP) { 5040 free_sched_groups(sd->groups, 1); 5041 } else if (atomic_dec_and_test(&sd->groups->ref)) { 5042 kfree(sd->groups->sgp); 5043 kfree(sd->groups); 5044 } 5045 kfree(sd); 5046} 5047 5048static void destroy_sched_domain(struct sched_domain *sd, int cpu) 5049{ 5050 call_rcu(&sd->rcu, free_sched_domain); 5051} 5052 5053static void destroy_sched_domains(struct sched_domain *sd, int cpu) 5054{ 5055 for (; sd; sd = sd->parent) 5056 destroy_sched_domain(sd, cpu); 5057} 5058 5059/* 5060 * Keep a special pointer to the highest sched_domain that has 5061 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this 5062 * allows us to avoid some pointer chasing select_idle_sibling(). 5063 * 5064 * Also keep a unique ID per domain (we use the first cpu number in 5065 * the cpumask of the domain), this allows us to quickly tell if 5066 * two cpus are in the same cache domain, see cpus_share_cache(). 5067 */ 5068DEFINE_PER_CPU(struct sched_domain *, sd_llc); 5069DEFINE_PER_CPU(int, sd_llc_id); 5070 5071static void update_top_cache_domain(int cpu) 5072{ 5073 struct sched_domain *sd; 5074 int id = cpu; 5075 5076 sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES); 5077 if (sd) 5078 id = cpumask_first(sched_domain_span(sd)); 5079 5080 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd); 5081 per_cpu(sd_llc_id, cpu) = id; 5082} 5083 5084/* 5085 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must 5086 * hold the hotplug lock. 5087 */ 5088static void 5089cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu) 5090{ 5091 struct rq *rq = cpu_rq(cpu); 5092 struct sched_domain *tmp; 5093 5094 /* Remove the sched domains which do not contribute to scheduling. */ 5095 for (tmp = sd; tmp; ) { 5096 struct sched_domain *parent = tmp->parent; 5097 if (!parent) 5098 break; 5099 5100 if (sd_parent_degenerate(tmp, parent)) { 5101 tmp->parent = parent->parent; 5102 if (parent->parent) 5103 parent->parent->child = tmp; 5104 destroy_sched_domain(parent, cpu); 5105 } else 5106 tmp = tmp->parent; 5107 } 5108 5109 if (sd && sd_degenerate(sd)) { 5110 tmp = sd; 5111 sd = sd->parent; 5112 destroy_sched_domain(tmp, cpu); 5113 if (sd) 5114 sd->child = NULL; 5115 } 5116 5117 sched_domain_debug(sd, cpu); 5118 5119 rq_attach_root(rq, rd); 5120 tmp = rq->sd; 5121 rcu_assign_pointer(rq->sd, sd); 5122 destroy_sched_domains(tmp, cpu); 5123 5124 update_top_cache_domain(cpu); 5125} 5126 5127/* cpus with isolated domains */ 5128static cpumask_var_t cpu_isolated_map; 5129 5130/* Setup the mask of cpus configured for isolated domains */ 5131static int __init isolated_cpu_setup(char *str) 5132{ 5133 alloc_bootmem_cpumask_var(&cpu_isolated_map); 5134 cpulist_parse(str, cpu_isolated_map); 5135 return 1; 5136} 5137 5138__setup("isolcpus=", isolated_cpu_setup); 5139 5140static const struct cpumask *cpu_cpu_mask(int cpu) 5141{ 5142 return cpumask_of_node(cpu_to_node(cpu)); 5143} 5144 5145struct sd_data { 5146 struct sched_domain **__percpu sd; 5147 struct sched_group **__percpu sg; 5148 struct sched_group_power **__percpu sgp; 5149}; 5150 5151struct s_data { 5152 struct sched_domain ** __percpu sd; 5153 struct root_domain *rd; 5154}; 5155 5156enum s_alloc { 5157 sa_rootdomain, 5158 sa_sd, 5159 sa_sd_storage, 5160 sa_none, 5161}; 5162 5163struct sched_domain_topology_level; 5164 5165typedef struct sched_domain *(*sched_domain_init_f)(struct sched_domain_topology_level *tl, int cpu); 5166typedef const struct cpumask *(*sched_domain_mask_f)(int cpu); 5167 5168#define SDTL_OVERLAP 0x01 5169 5170struct sched_domain_topology_level { 5171 sched_domain_init_f init; 5172 sched_domain_mask_f mask; 5173 int flags; 5174 int numa_level; 5175 struct sd_data data; 5176}; 5177 5178/* 5179 * Build an iteration mask that can exclude certain CPUs from the upwards 5180 * domain traversal. 5181 * 5182 * Asymmetric node setups can result in situations where the domain tree is of 5183 * unequal depth, make sure to skip domains that already cover the entire 5184 * range. 5185 * 5186 * In that case build_sched_domains() will have terminated the iteration early 5187 * and our sibling sd spans will be empty. Domains should always include the 5188 * cpu they're built on, so check that. 5189 * 5190 */ 5191static void build_group_mask(struct sched_domain *sd, struct sched_group *sg) 5192{ 5193 const struct cpumask *span = sched_domain_span(sd); 5194 struct sd_data *sdd = sd->private; 5195 struct sched_domain *sibling; 5196 int i; 5197 5198 for_each_cpu(i, span) { 5199 sibling = *per_cpu_ptr(sdd->sd, i); 5200 if (!cpumask_test_cpu(i, sched_domain_span(sibling))) 5201 continue; 5202 5203 cpumask_set_cpu(i, sched_group_mask(sg)); 5204 } 5205} 5206 5207/* 5208 * Return the canonical balance cpu for this group, this is the first cpu 5209 * of this group that's also in the iteration mask. 5210 */ 5211int group_balance_cpu(struct sched_group *sg) 5212{ 5213 return cpumask_first_and(sched_group_cpus(sg), sched_group_mask(sg)); 5214} 5215 5216static int 5217build_overlap_sched_groups(struct sched_domain *sd, int cpu) 5218{ 5219 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg; 5220 const struct cpumask *span = sched_domain_span(sd); 5221 struct cpumask *covered = sched_domains_tmpmask; 5222 struct sd_data *sdd = sd->private; 5223 struct sched_domain *child; 5224 int i; 5225 5226 cpumask_clear(covered); 5227 5228 for_each_cpu(i, span) { 5229 struct cpumask *sg_span; 5230 5231 if (cpumask_test_cpu(i, covered)) 5232 continue; 5233 5234 child = *per_cpu_ptr(sdd->sd, i); 5235 5236 /* See the comment near build_group_mask(). */ 5237 if (!cpumask_test_cpu(i, sched_domain_span(child))) 5238 continue; 5239 5240 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(), 5241 GFP_KERNEL, cpu_to_node(cpu)); 5242 5243 if (!sg) 5244 goto fail; 5245 5246 sg_span = sched_group_cpus(sg); 5247 if (child->child) { 5248 child = child->child; 5249 cpumask_copy(sg_span, sched_domain_span(child)); 5250 } else 5251 cpumask_set_cpu(i, sg_span); 5252 5253 cpumask_or(covered, covered, sg_span); 5254 5255 sg->sgp = *per_cpu_ptr(sdd->sgp, i); 5256 if (atomic_inc_return(&sg->sgp->ref) == 1) 5257 build_group_mask(sd, sg); 5258 5259 /* 5260 * Initialize sgp->power such that even if we mess up the 5261 * domains and no possible iteration will get us here, we won't 5262 * die on a /0 trap. 5263 */ 5264 sg->sgp->power = SCHED_POWER_SCALE * cpumask_weight(sg_span); 5265 5266 /* 5267 * Make sure the first group of this domain contains the 5268 * canonical balance cpu. Otherwise the sched_domain iteration 5269 * breaks. See update_sg_lb_stats(). 5270 */ 5271 if ((!groups && cpumask_test_cpu(cpu, sg_span)) || 5272 group_balance_cpu(sg) == cpu) 5273 groups = sg; 5274 5275 if (!first) 5276 first = sg; 5277 if (last) 5278 last->next = sg; 5279 last = sg; 5280 last->next = first; 5281 } 5282 sd->groups = groups; 5283 5284 return 0; 5285 5286fail: 5287 free_sched_groups(first, 0); 5288 5289 return -ENOMEM; 5290} 5291 5292static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg) 5293{ 5294 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu); 5295 struct sched_domain *child = sd->child; 5296 5297 if (child) 5298 cpu = cpumask_first(sched_domain_span(child)); 5299 5300 if (sg) { 5301 *sg = *per_cpu_ptr(sdd->sg, cpu); 5302 (*sg)->sgp = *per_cpu_ptr(sdd->sgp, cpu); 5303 atomic_set(&(*sg)->sgp->ref, 1); /* for claim_allocations */ 5304 } 5305 5306 return cpu; 5307} 5308 5309/* 5310 * build_sched_groups will build a circular linked list of the groups 5311 * covered by the given span, and will set each group's ->cpumask correctly, 5312 * and ->cpu_power to 0. 5313 * 5314 * Assumes the sched_domain tree is fully constructed 5315 */ 5316static int 5317build_sched_groups(struct sched_domain *sd, int cpu) 5318{ 5319 struct sched_group *first = NULL, *last = NULL; 5320 struct sd_data *sdd = sd->private; 5321 const struct cpumask *span = sched_domain_span(sd); 5322 struct cpumask *covered; 5323 int i; 5324 5325 get_group(cpu, sdd, &sd->groups); 5326 atomic_inc(&sd->groups->ref); 5327 5328 if (cpu != cpumask_first(sched_domain_span(sd))) 5329 return 0; 5330 5331 lockdep_assert_held(&sched_domains_mutex); 5332 covered = sched_domains_tmpmask; 5333 5334 cpumask_clear(covered); 5335 5336 for_each_cpu(i, span) { 5337 struct sched_group *sg; 5338 int group = get_group(i, sdd, &sg); 5339 int j; 5340 5341 if (cpumask_test_cpu(i, covered)) 5342 continue; 5343 5344 cpumask_clear(sched_group_cpus(sg)); 5345 sg->sgp->power = 0; 5346 cpumask_setall(sched_group_mask(sg)); 5347 5348 for_each_cpu(j, span) { 5349 if (get_group(j, sdd, NULL) != group) 5350 continue; 5351 5352 cpumask_set_cpu(j, covered); 5353 cpumask_set_cpu(j, sched_group_cpus(sg)); 5354 } 5355 5356 if (!first) 5357 first = sg; 5358 if (last) 5359 last->next = sg; 5360 last = sg; 5361 } 5362 last->next = first; 5363 5364 return 0; 5365} 5366 5367/* 5368 * Initialize sched groups cpu_power. 5369 * 5370 * cpu_power indicates the capacity of sched group, which is used while 5371 * distributing the load between different sched groups in a sched domain. 5372 * Typically cpu_power for all the groups in a sched domain will be same unless 5373 * there are asymmetries in the topology. If there are asymmetries, group 5374 * having more cpu_power will pickup more load compared to the group having 5375 * less cpu_power. 5376 */ 5377static void init_sched_groups_power(int cpu, struct sched_domain *sd) 5378{ 5379 struct sched_group *sg = sd->groups; 5380 5381 WARN_ON(!sd || !sg); 5382 5383 do { 5384 sg->group_weight = cpumask_weight(sched_group_cpus(sg)); 5385 sg = sg->next; 5386 } while (sg != sd->groups); 5387 5388 if (cpu != group_balance_cpu(sg)) 5389 return; 5390 5391 update_group_power(sd, cpu); 5392 atomic_set(&sg->sgp->nr_busy_cpus, sg->group_weight); 5393} 5394 5395int __weak arch_sd_sibling_asym_packing(void) 5396{ 5397 return 0*SD_ASYM_PACKING; 5398} 5399 5400/* 5401 * Initializers for schedule domains 5402 * Non-inlined to reduce accumulated stack pressure in build_sched_domains() 5403 */ 5404 5405#ifdef CONFIG_SCHED_DEBUG 5406# define SD_INIT_NAME(sd, type) sd->name = #type 5407#else 5408# define SD_INIT_NAME(sd, type) do { } while (0) 5409#endif 5410 5411#define SD_INIT_FUNC(type) \ 5412static noinline struct sched_domain * \ 5413sd_init_##type(struct sched_domain_topology_level *tl, int cpu) \ 5414{ \ 5415 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); \ 5416 *sd = SD_##type##_INIT; \ 5417 SD_INIT_NAME(sd, type); \ 5418 sd->private = &tl->data; \ 5419 return sd; \ 5420} 5421 5422SD_INIT_FUNC(CPU) 5423#ifdef CONFIG_SCHED_SMT 5424 SD_INIT_FUNC(SIBLING) 5425#endif 5426#ifdef CONFIG_SCHED_MC 5427 SD_INIT_FUNC(MC) 5428#endif 5429#ifdef CONFIG_SCHED_BOOK 5430 SD_INIT_FUNC(BOOK) 5431#endif 5432 5433static int default_relax_domain_level = -1; 5434int sched_domain_level_max; 5435 5436static int __init setup_relax_domain_level(char *str) 5437{ 5438 if (kstrtoint(str, 0, &default_relax_domain_level)) 5439 pr_warn("Unable to set relax_domain_level\n"); 5440 5441 return 1; 5442} 5443__setup("relax_domain_level=", setup_relax_domain_level); 5444 5445static void set_domain_attribute(struct sched_domain *sd, 5446 struct sched_domain_attr *attr) 5447{ 5448 int request; 5449 5450 if (!attr || attr->relax_domain_level < 0) { 5451 if (default_relax_domain_level < 0) 5452 return; 5453 else 5454 request = default_relax_domain_level; 5455 } else 5456 request = attr->relax_domain_level; 5457 if (request < sd->level) { 5458 /* turn off idle balance on this domain */ 5459 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE); 5460 } else { 5461 /* turn on idle balance on this domain */ 5462 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE); 5463 } 5464} 5465 5466static void __sdt_free(const struct cpumask *cpu_map); 5467static int __sdt_alloc(const struct cpumask *cpu_map); 5468 5469static void __free_domain_allocs(struct s_data *d, enum s_alloc what, 5470 const struct cpumask *cpu_map) 5471{ 5472 switch (what) { 5473 case sa_rootdomain: 5474 if (!atomic_read(&d->rd->refcount)) 5475 free_rootdomain(&d->rd->rcu); /* fall through */ 5476 case sa_sd: 5477 free_percpu(d->sd); /* fall through */ 5478 case sa_sd_storage: 5479 __sdt_free(cpu_map); /* fall through */ 5480 case sa_none: 5481 break; 5482 } 5483} 5484 5485static enum s_alloc __visit_domain_allocation_hell(struct s_data *d, 5486 const struct cpumask *cpu_map) 5487{ 5488 memset(d, 0, sizeof(*d)); 5489 5490 if (__sdt_alloc(cpu_map)) 5491 return sa_sd_storage; 5492 d->sd = alloc_percpu(struct sched_domain *); 5493 if (!d->sd) 5494 return sa_sd_storage; 5495 d->rd = alloc_rootdomain(); 5496 if (!d->rd) 5497 return sa_sd; 5498 return sa_rootdomain; 5499} 5500 5501/* 5502 * NULL the sd_data elements we've used to build the sched_domain and 5503 * sched_group structure so that the subsequent __free_domain_allocs() 5504 * will not free the data we're using. 5505 */ 5506static void claim_allocations(int cpu, struct sched_domain *sd) 5507{ 5508 struct sd_data *sdd = sd->private; 5509 5510 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd); 5511 *per_cpu_ptr(sdd->sd, cpu) = NULL; 5512 5513 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref)) 5514 *per_cpu_ptr(sdd->sg, cpu) = NULL; 5515 5516 if (atomic_read(&(*per_cpu_ptr(sdd->sgp, cpu))->ref)) 5517 *per_cpu_ptr(sdd->sgp, cpu) = NULL; 5518} 5519 5520#ifdef CONFIG_SCHED_SMT 5521static const struct cpumask *cpu_smt_mask(int cpu) 5522{ 5523 return topology_thread_cpumask(cpu); 5524} 5525#endif 5526 5527/* 5528 * Topology list, bottom-up. 5529 */ 5530static struct sched_domain_topology_level default_topology[] = { 5531#ifdef CONFIG_SCHED_SMT 5532 { sd_init_SIBLING, cpu_smt_mask, }, 5533#endif 5534#ifdef CONFIG_SCHED_MC 5535 { sd_init_MC, cpu_coregroup_mask, }, 5536#endif 5537#ifdef CONFIG_SCHED_BOOK 5538 { sd_init_BOOK, cpu_book_mask, }, 5539#endif 5540 { sd_init_CPU, cpu_cpu_mask, }, 5541 { NULL, }, 5542}; 5543 5544static struct sched_domain_topology_level *sched_domain_topology = default_topology; 5545 5546#ifdef CONFIG_NUMA 5547 5548static int sched_domains_numa_levels; 5549static int *sched_domains_numa_distance; 5550static struct cpumask ***sched_domains_numa_masks; 5551static int sched_domains_curr_level; 5552 5553static inline int sd_local_flags(int level) 5554{ 5555 if (sched_domains_numa_distance[level] > RECLAIM_DISTANCE) 5556 return 0; 5557 5558 return SD_BALANCE_EXEC | SD_BALANCE_FORK | SD_WAKE_AFFINE; 5559} 5560 5561static struct sched_domain * 5562sd_numa_init(struct sched_domain_topology_level *tl, int cpu) 5563{ 5564 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); 5565 int level = tl->numa_level; 5566 int sd_weight = cpumask_weight( 5567 sched_domains_numa_masks[level][cpu_to_node(cpu)]); 5568 5569 *sd = (struct sched_domain){ 5570 .min_interval = sd_weight, 5571 .max_interval = 2*sd_weight, 5572 .busy_factor = 32, 5573 .imbalance_pct = 125, 5574 .cache_nice_tries = 2, 5575 .busy_idx = 3, 5576 .idle_idx = 2, 5577 .newidle_idx = 0, 5578 .wake_idx = 0, 5579 .forkexec_idx = 0, 5580 5581 .flags = 1*SD_LOAD_BALANCE 5582 | 1*SD_BALANCE_NEWIDLE 5583 | 0*SD_BALANCE_EXEC 5584 | 0*SD_BALANCE_FORK 5585 | 0*SD_BALANCE_WAKE 5586 | 0*SD_WAKE_AFFINE 5587 | 0*SD_SHARE_CPUPOWER 5588 | 0*SD_SHARE_PKG_RESOURCES 5589 | 1*SD_SERIALIZE 5590 | 0*SD_PREFER_SIBLING 5591 | sd_local_flags(level) 5592 , 5593 .last_balance = jiffies, 5594 .balance_interval = sd_weight, 5595 }; 5596 SD_INIT_NAME(sd, NUMA); 5597 sd->private = &tl->data; 5598 5599 /* 5600 * Ugly hack to pass state to sd_numa_mask()... 5601 */ 5602 sched_domains_curr_level = tl->numa_level; 5603 5604 return sd; 5605} 5606 5607static const struct cpumask *sd_numa_mask(int cpu) 5608{ 5609 return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)]; 5610} 5611 5612static void sched_numa_warn(const char *str) 5613{ 5614 static int done = false; 5615 int i,j; 5616 5617 if (done) 5618 return; 5619 5620 done = true; 5621 5622 printk(KERN_WARNING "ERROR: %s\n\n", str); 5623 5624 for (i = 0; i < nr_node_ids; i++) { 5625 printk(KERN_WARNING " "); 5626 for (j = 0; j < nr_node_ids; j++) 5627 printk(KERN_CONT "%02d ", node_distance(i,j)); 5628 printk(KERN_CONT "\n"); 5629 } 5630 printk(KERN_WARNING "\n"); 5631} 5632 5633static bool find_numa_distance(int distance) 5634{ 5635 int i; 5636 5637 if (distance == node_distance(0, 0)) 5638 return true; 5639 5640 for (i = 0; i < sched_domains_numa_levels; i++) { 5641 if (sched_domains_numa_distance[i] == distance) 5642 return true; 5643 } 5644 5645 return false; 5646} 5647 5648static void sched_init_numa(void) 5649{ 5650 int next_distance, curr_distance = node_distance(0, 0); 5651 struct sched_domain_topology_level *tl; 5652 int level = 0; 5653 int i, j, k; 5654 5655 sched_domains_numa_distance = kzalloc(sizeof(int) * nr_node_ids, GFP_KERNEL); 5656 if (!sched_domains_numa_distance) 5657 return; 5658 5659 /* 5660 * O(nr_nodes^2) deduplicating selection sort -- in order to find the 5661 * unique distances in the node_distance() table. 5662 * 5663 * Assumes node_distance(0,j) includes all distances in 5664 * node_distance(i,j) in order to avoid cubic time. 5665 */ 5666 next_distance = curr_distance; 5667 for (i = 0; i < nr_node_ids; i++) { 5668 for (j = 0; j < nr_node_ids; j++) { 5669 for (k = 0; k < nr_node_ids; k++) { 5670 int distance = node_distance(i, k); 5671 5672 if (distance > curr_distance && 5673 (distance < next_distance || 5674 next_distance == curr_distance)) 5675 next_distance = distance; 5676 5677 /* 5678 * While not a strong assumption it would be nice to know 5679 * about cases where if node A is connected to B, B is not 5680 * equally connected to A. 5681 */ 5682 if (sched_debug() && node_distance(k, i) != distance) 5683 sched_numa_warn("Node-distance not symmetric"); 5684 5685 if (sched_debug() && i && !find_numa_distance(distance)) 5686 sched_numa_warn("Node-0 not representative"); 5687 } 5688 if (next_distance != curr_distance) { 5689 sched_domains_numa_distance[level++] = next_distance; 5690 sched_domains_numa_levels = level; 5691 curr_distance = next_distance; 5692 } else break; 5693 } 5694 5695 /* 5696 * In case of sched_debug() we verify the above assumption. 5697 */ 5698 if (!sched_debug()) 5699 break; 5700 } 5701 /* 5702 * 'level' contains the number of unique distances, excluding the 5703 * identity distance node_distance(i,i). 5704 * 5705 * The sched_domains_numa_distance[] array includes the actual distance 5706 * numbers. 5707 */ 5708 5709 /* 5710 * Here, we should temporarily reset sched_domains_numa_levels to 0. 5711 * If it fails to allocate memory for array sched_domains_numa_masks[][], 5712 * the array will contain less then 'level' members. This could be 5713 * dangerous when we use it to iterate array sched_domains_numa_masks[][] 5714 * in other functions. 5715 * 5716 * We reset it to 'level' at the end of this function. 5717 */ 5718 sched_domains_numa_levels = 0; 5719 5720 sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL); 5721 if (!sched_domains_numa_masks) 5722 return; 5723 5724 /* 5725 * Now for each level, construct a mask per node which contains all 5726 * cpus of nodes that are that many hops away from us. 5727 */ 5728 for (i = 0; i < level; i++) { 5729 sched_domains_numa_masks[i] = 5730 kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL); 5731 if (!sched_domains_numa_masks[i]) 5732 return; 5733 5734 for (j = 0; j < nr_node_ids; j++) { 5735 struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL); 5736 if (!mask) 5737 return; 5738 5739 sched_domains_numa_masks[i][j] = mask; 5740 5741 for (k = 0; k < nr_node_ids; k++) { 5742 if (node_distance(j, k) > sched_domains_numa_distance[i]) 5743 continue; 5744 5745 cpumask_or(mask, mask, cpumask_of_node(k)); 5746 } 5747 } 5748 } 5749 5750 tl = kzalloc((ARRAY_SIZE(default_topology) + level) * 5751 sizeof(struct sched_domain_topology_level), GFP_KERNEL); 5752 if (!tl) 5753 return; 5754 5755 /* 5756 * Copy the default topology bits.. 5757 */ 5758 for (i = 0; default_topology[i].init; i++) 5759 tl[i] = default_topology[i]; 5760 5761 /* 5762 * .. and append 'j' levels of NUMA goodness. 5763 */ 5764 for (j = 0; j < level; i++, j++) { 5765 tl[i] = (struct sched_domain_topology_level){ 5766 .init = sd_numa_init, 5767 .mask = sd_numa_mask, 5768 .flags = SDTL_OVERLAP, 5769 .numa_level = j, 5770 }; 5771 } 5772 5773 sched_domain_topology = tl; 5774 5775 sched_domains_numa_levels = level; 5776} 5777 5778static void sched_domains_numa_masks_set(int cpu) 5779{ 5780 int i, j; 5781 int node = cpu_to_node(cpu); 5782 5783 for (i = 0; i < sched_domains_numa_levels; i++) { 5784 for (j = 0; j < nr_node_ids; j++) { 5785 if (node_distance(j, node) <= sched_domains_numa_distance[i]) 5786 cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]); 5787 } 5788 } 5789} 5790 5791static void sched_domains_numa_masks_clear(int cpu) 5792{ 5793 int i, j; 5794 for (i = 0; i < sched_domains_numa_levels; i++) { 5795 for (j = 0; j < nr_node_ids; j++) 5796 cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]); 5797 } 5798} 5799 5800/* 5801 * Update sched_domains_numa_masks[level][node] array when new cpus 5802 * are onlined. 5803 */ 5804static int sched_domains_numa_masks_update(struct notifier_block *nfb, 5805 unsigned long action, 5806 void *hcpu) 5807{ 5808 int cpu = (long)hcpu; 5809 5810 switch (action & ~CPU_TASKS_FROZEN) { 5811 case CPU_ONLINE: 5812 sched_domains_numa_masks_set(cpu); 5813 break; 5814 5815 case CPU_DEAD: 5816 sched_domains_numa_masks_clear(cpu); 5817 break; 5818 5819 default: 5820 return NOTIFY_DONE; 5821 } 5822 5823 return NOTIFY_OK; 5824} 5825#else 5826static inline void sched_init_numa(void) 5827{ 5828} 5829 5830static int sched_domains_numa_masks_update(struct notifier_block *nfb, 5831 unsigned long action, 5832 void *hcpu) 5833{ 5834 return 0; 5835} 5836#endif /* CONFIG_NUMA */ 5837 5838static int __sdt_alloc(const struct cpumask *cpu_map) 5839{ 5840 struct sched_domain_topology_level *tl; 5841 int j; 5842 5843 for (tl = sched_domain_topology; tl->init; tl++) { 5844 struct sd_data *sdd = &tl->data; 5845 5846 sdd->sd = alloc_percpu(struct sched_domain *); 5847 if (!sdd->sd) 5848 return -ENOMEM; 5849 5850 sdd->sg = alloc_percpu(struct sched_group *); 5851 if (!sdd->sg) 5852 return -ENOMEM; 5853 5854 sdd->sgp = alloc_percpu(struct sched_group_power *); 5855 if (!sdd->sgp) 5856 return -ENOMEM; 5857 5858 for_each_cpu(j, cpu_map) { 5859 struct sched_domain *sd; 5860 struct sched_group *sg; 5861 struct sched_group_power *sgp; 5862 5863 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(), 5864 GFP_KERNEL, cpu_to_node(j)); 5865 if (!sd) 5866 return -ENOMEM; 5867 5868 *per_cpu_ptr(sdd->sd, j) = sd; 5869 5870 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(), 5871 GFP_KERNEL, cpu_to_node(j)); 5872 if (!sg) 5873 return -ENOMEM; 5874 5875 sg->next = sg; 5876 5877 *per_cpu_ptr(sdd->sg, j) = sg; 5878 5879 sgp = kzalloc_node(sizeof(struct sched_group_power) + cpumask_size(), 5880 GFP_KERNEL, cpu_to_node(j)); 5881 if (!sgp) 5882 return -ENOMEM; 5883 5884 *per_cpu_ptr(sdd->sgp, j) = sgp; 5885 } 5886 } 5887 5888 return 0; 5889} 5890 5891static void __sdt_free(const struct cpumask *cpu_map) 5892{ 5893 struct sched_domain_topology_level *tl; 5894 int j; 5895 5896 for (tl = sched_domain_topology; tl->init; tl++) { 5897 struct sd_data *sdd = &tl->data; 5898 5899 for_each_cpu(j, cpu_map) { 5900 struct sched_domain *sd; 5901 5902 if (sdd->sd) { 5903 sd = *per_cpu_ptr(sdd->sd, j); 5904 if (sd && (sd->flags & SD_OVERLAP)) 5905 free_sched_groups(sd->groups, 0); 5906 kfree(*per_cpu_ptr(sdd->sd, j)); 5907 } 5908 5909 if (sdd->sg) 5910 kfree(*per_cpu_ptr(sdd->sg, j)); 5911 if (sdd->sgp) 5912 kfree(*per_cpu_ptr(sdd->sgp, j)); 5913 } 5914 free_percpu(sdd->sd); 5915 sdd->sd = NULL; 5916 free_percpu(sdd->sg); 5917 sdd->sg = NULL; 5918 free_percpu(sdd->sgp); 5919 sdd->sgp = NULL; 5920 } 5921} 5922 5923struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl, 5924 struct s_data *d, const struct cpumask *cpu_map, 5925 struct sched_domain_attr *attr, struct sched_domain *child, 5926 int cpu) 5927{ 5928 struct sched_domain *sd = tl->init(tl, cpu); 5929 if (!sd) 5930 return child; 5931 5932 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu)); 5933 if (child) { 5934 sd->level = child->level + 1; 5935 sched_domain_level_max = max(sched_domain_level_max, sd->level); 5936 child->parent = sd; 5937 } 5938 sd->child = child; 5939 set_domain_attribute(sd, attr); 5940 5941 return sd; 5942} 5943 5944/* 5945 * Build sched domains for a given set of cpus and attach the sched domains 5946 * to the individual cpus 5947 */ 5948static int build_sched_domains(const struct cpumask *cpu_map, 5949 struct sched_domain_attr *attr) 5950{ 5951 enum s_alloc alloc_state = sa_none; 5952 struct sched_domain *sd; 5953 struct s_data d; 5954 int i, ret = -ENOMEM; 5955 5956 alloc_state = __visit_domain_allocation_hell(&d, cpu_map); 5957 if (alloc_state != sa_rootdomain) 5958 goto error; 5959 5960 /* Set up domains for cpus specified by the cpu_map. */ 5961 for_each_cpu(i, cpu_map) { 5962 struct sched_domain_topology_level *tl; 5963 5964 sd = NULL; 5965 for (tl = sched_domain_topology; tl->init; tl++) { 5966 sd = build_sched_domain(tl, &d, cpu_map, attr, sd, i); 5967 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP)) 5968 sd->flags |= SD_OVERLAP; 5969 if (cpumask_equal(cpu_map, sched_domain_span(sd))) 5970 break; 5971 } 5972 5973 while (sd->child) 5974 sd = sd->child; 5975 5976 *per_cpu_ptr(d.sd, i) = sd; 5977 } 5978 5979 /* Build the groups for the domains */ 5980 for_each_cpu(i, cpu_map) { 5981 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) { 5982 sd->span_weight = cpumask_weight(sched_domain_span(sd)); 5983 if (sd->flags & SD_OVERLAP) { 5984 if (build_overlap_sched_groups(sd, i)) 5985 goto error; 5986 } else { 5987 if (build_sched_groups(sd, i)) 5988 goto error; 5989 } 5990 } 5991 } 5992 5993 /* Calculate CPU power for physical packages and nodes */ 5994 for (i = nr_cpumask_bits-1; i >= 0; i--) { 5995 if (!cpumask_test_cpu(i, cpu_map)) 5996 continue; 5997 5998 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) { 5999 claim_allocations(i, sd); 6000 init_sched_groups_power(i, sd); 6001 } 6002 } 6003 6004 /* Attach the domains */ 6005 rcu_read_lock(); 6006 for_each_cpu(i, cpu_map) { 6007 sd = *per_cpu_ptr(d.sd, i); 6008 cpu_attach_domain(sd, d.rd, i); 6009 } 6010 rcu_read_unlock(); 6011 6012 ret = 0; 6013error: 6014 __free_domain_allocs(&d, alloc_state, cpu_map); 6015 return ret; 6016} 6017 6018static cpumask_var_t *doms_cur; /* current sched domains */ 6019static int ndoms_cur; /* number of sched domains in 'doms_cur' */ 6020static struct sched_domain_attr *dattr_cur; 6021 /* attribues of custom domains in 'doms_cur' */ 6022 6023/* 6024 * Special case: If a kmalloc of a doms_cur partition (array of 6025 * cpumask) fails, then fallback to a single sched domain, 6026 * as determined by the single cpumask fallback_doms. 6027 */ 6028static cpumask_var_t fallback_doms; 6029 6030/* 6031 * arch_update_cpu_topology lets virtualized architectures update the 6032 * cpu core maps. It is supposed to return 1 if the topology changed 6033 * or 0 if it stayed the same. 6034 */ 6035int __attribute__((weak)) arch_update_cpu_topology(void) 6036{ 6037 return 0; 6038} 6039 6040cpumask_var_t *alloc_sched_domains(unsigned int ndoms) 6041{ 6042 int i; 6043 cpumask_var_t *doms; 6044 6045 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL); 6046 if (!doms) 6047 return NULL; 6048 for (i = 0; i < ndoms; i++) { 6049 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) { 6050 free_sched_domains(doms, i); 6051 return NULL; 6052 } 6053 } 6054 return doms; 6055} 6056 6057void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms) 6058{ 6059 unsigned int i; 6060 for (i = 0; i < ndoms; i++) 6061 free_cpumask_var(doms[i]); 6062 kfree(doms); 6063} 6064 6065/* 6066 * Set up scheduler domains and groups. Callers must hold the hotplug lock. 6067 * For now this just excludes isolated cpus, but could be used to 6068 * exclude other special cases in the future. 6069 */ 6070static int init_sched_domains(const struct cpumask *cpu_map) 6071{ 6072 int err; 6073 6074 arch_update_cpu_topology(); 6075 ndoms_cur = 1; 6076 doms_cur = alloc_sched_domains(ndoms_cur); 6077 if (!doms_cur) 6078 doms_cur = &fallback_doms; 6079 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map); 6080 err = build_sched_domains(doms_cur[0], NULL); 6081 register_sched_domain_sysctl(); 6082 6083 return err; 6084} 6085 6086/* 6087 * Detach sched domains from a group of cpus specified in cpu_map 6088 * These cpus will now be attached to the NULL domain 6089 */ 6090static void detach_destroy_domains(const struct cpumask *cpu_map) 6091{ 6092 int i; 6093 6094 rcu_read_lock(); 6095 for_each_cpu(i, cpu_map) 6096 cpu_attach_domain(NULL, &def_root_domain, i); 6097 rcu_read_unlock(); 6098} 6099 6100/* handle null as "default" */ 6101static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur, 6102 struct sched_domain_attr *new, int idx_new) 6103{ 6104 struct sched_domain_attr tmp; 6105 6106 /* fast path */ 6107 if (!new && !cur) 6108 return 1; 6109 6110 tmp = SD_ATTR_INIT; 6111 return !memcmp(cur ? (cur + idx_cur) : &tmp, 6112 new ? (new + idx_new) : &tmp, 6113 sizeof(struct sched_domain_attr)); 6114} 6115 6116/* 6117 * Partition sched domains as specified by the 'ndoms_new' 6118 * cpumasks in the array doms_new[] of cpumasks. This compares 6119 * doms_new[] to the current sched domain partitioning, doms_cur[]. 6120 * It destroys each deleted domain and builds each new domain. 6121 * 6122 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'. 6123 * The masks don't intersect (don't overlap.) We should setup one 6124 * sched domain for each mask. CPUs not in any of the cpumasks will 6125 * not be load balanced. If the same cpumask appears both in the 6126 * current 'doms_cur' domains and in the new 'doms_new', we can leave 6127 * it as it is. 6128 * 6129 * The passed in 'doms_new' should be allocated using 6130 * alloc_sched_domains. This routine takes ownership of it and will 6131 * free_sched_domains it when done with it. If the caller failed the 6132 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1, 6133 * and partition_sched_domains() will fallback to the single partition 6134 * 'fallback_doms', it also forces the domains to be rebuilt. 6135 * 6136 * If doms_new == NULL it will be replaced with cpu_online_mask. 6137 * ndoms_new == 0 is a special case for destroying existing domains, 6138 * and it will not create the default domain. 6139 * 6140 * Call with hotplug lock held 6141 */ 6142void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[], 6143 struct sched_domain_attr *dattr_new) 6144{ 6145 int i, j, n; 6146 int new_topology; 6147 6148 mutex_lock(&sched_domains_mutex); 6149 6150 /* always unregister in case we don't destroy any domains */ 6151 unregister_sched_domain_sysctl(); 6152 6153 /* Let architecture update cpu core mappings. */ 6154 new_topology = arch_update_cpu_topology(); 6155 6156 n = doms_new ? ndoms_new : 0; 6157 6158 /* Destroy deleted domains */ 6159 for (i = 0; i < ndoms_cur; i++) { 6160 for (j = 0; j < n && !new_topology; j++) { 6161 if (cpumask_equal(doms_cur[i], doms_new[j]) 6162 && dattrs_equal(dattr_cur, i, dattr_new, j)) 6163 goto match1; 6164 } 6165 /* no match - a current sched domain not in new doms_new[] */ 6166 detach_destroy_domains(doms_cur[i]); 6167match1: 6168 ; 6169 } 6170 6171 if (doms_new == NULL) { 6172 ndoms_cur = 0; 6173 doms_new = &fallback_doms; 6174 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map); 6175 WARN_ON_ONCE(dattr_new); 6176 } 6177 6178 /* Build new domains */ 6179 for (i = 0; i < ndoms_new; i++) { 6180 for (j = 0; j < ndoms_cur && !new_topology; j++) { 6181 if (cpumask_equal(doms_new[i], doms_cur[j]) 6182 && dattrs_equal(dattr_new, i, dattr_cur, j)) 6183 goto match2; 6184 } 6185 /* no match - add a new doms_new */ 6186 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL); 6187match2: 6188 ; 6189 } 6190 6191 /* Remember the new sched domains */ 6192 if (doms_cur != &fallback_doms) 6193 free_sched_domains(doms_cur, ndoms_cur); 6194 kfree(dattr_cur); /* kfree(NULL) is safe */ 6195 doms_cur = doms_new; 6196 dattr_cur = dattr_new; 6197 ndoms_cur = ndoms_new; 6198 6199 register_sched_domain_sysctl(); 6200 6201 mutex_unlock(&sched_domains_mutex); 6202} 6203 6204static int num_cpus_frozen; /* used to mark begin/end of suspend/resume */ 6205 6206/* 6207 * Update cpusets according to cpu_active mask. If cpusets are 6208 * disabled, cpuset_update_active_cpus() becomes a simple wrapper 6209 * around partition_sched_domains(). 6210 * 6211 * If we come here as part of a suspend/resume, don't touch cpusets because we 6212 * want to restore it back to its original state upon resume anyway. 6213 */ 6214static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action, 6215 void *hcpu) 6216{ 6217 switch (action) { 6218 case CPU_ONLINE_FROZEN: 6219 case CPU_DOWN_FAILED_FROZEN: 6220 6221 /* 6222 * num_cpus_frozen tracks how many CPUs are involved in suspend 6223 * resume sequence. As long as this is not the last online 6224 * operation in the resume sequence, just build a single sched 6225 * domain, ignoring cpusets. 6226 */ 6227 num_cpus_frozen--; 6228 if (likely(num_cpus_frozen)) { 6229 partition_sched_domains(1, NULL, NULL); 6230 break; 6231 } 6232 6233 /* 6234 * This is the last CPU online operation. So fall through and 6235 * restore the original sched domains by considering the 6236 * cpuset configurations. 6237 */ 6238 6239 case CPU_ONLINE: 6240 case CPU_DOWN_FAILED: 6241 cpuset_update_active_cpus(true); 6242 break; 6243 default: 6244 return NOTIFY_DONE; 6245 } 6246 return NOTIFY_OK; 6247} 6248 6249static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action, 6250 void *hcpu) 6251{ 6252 switch (action) { 6253 case CPU_DOWN_PREPARE: 6254 cpuset_update_active_cpus(false); 6255 break; 6256 case CPU_DOWN_PREPARE_FROZEN: 6257 num_cpus_frozen++; 6258 partition_sched_domains(1, NULL, NULL); 6259 break; 6260 default: 6261 return NOTIFY_DONE; 6262 } 6263 return NOTIFY_OK; 6264} 6265 6266void __init sched_init_smp(void) 6267{ 6268 cpumask_var_t non_isolated_cpus; 6269 6270 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL); 6271 alloc_cpumask_var(&fallback_doms, GFP_KERNEL); 6272 6273 sched_init_numa(); 6274 6275 get_online_cpus(); 6276 mutex_lock(&sched_domains_mutex); 6277 init_sched_domains(cpu_active_mask); 6278 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map); 6279 if (cpumask_empty(non_isolated_cpus)) 6280 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus); 6281 mutex_unlock(&sched_domains_mutex); 6282 put_online_cpus(); 6283 6284 hotcpu_notifier(sched_domains_numa_masks_update, CPU_PRI_SCHED_ACTIVE); 6285 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE); 6286 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE); 6287 6288 init_hrtick(); 6289 6290 /* Move init over to a non-isolated CPU */ 6291 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0) 6292 BUG(); 6293 sched_init_granularity(); 6294 free_cpumask_var(non_isolated_cpus); 6295 6296 init_sched_rt_class(); 6297} 6298#else 6299void __init sched_init_smp(void) 6300{ 6301 sched_init_granularity(); 6302} 6303#endif /* CONFIG_SMP */ 6304 6305const_debug unsigned int sysctl_timer_migration = 1; 6306 6307int in_sched_functions(unsigned long addr) 6308{ 6309 return in_lock_functions(addr) || 6310 (addr >= (unsigned long)__sched_text_start 6311 && addr < (unsigned long)__sched_text_end); 6312} 6313 6314#ifdef CONFIG_CGROUP_SCHED 6315/* 6316 * Default task group. 6317 * Every task in system belongs to this group at bootup. 6318 */ 6319struct task_group root_task_group; 6320LIST_HEAD(task_groups); 6321#endif 6322 6323DECLARE_PER_CPU(cpumask_var_t, load_balance_mask); 6324 6325void __init sched_init(void) 6326{ 6327 int i, j; 6328 unsigned long alloc_size = 0, ptr; 6329 6330#ifdef CONFIG_FAIR_GROUP_SCHED 6331 alloc_size += 2 * nr_cpu_ids * sizeof(void **); 6332#endif 6333#ifdef CONFIG_RT_GROUP_SCHED 6334 alloc_size += 2 * nr_cpu_ids * sizeof(void **); 6335#endif 6336#ifdef CONFIG_CPUMASK_OFFSTACK 6337 alloc_size += num_possible_cpus() * cpumask_size(); 6338#endif 6339 if (alloc_size) { 6340 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT); 6341 6342#ifdef CONFIG_FAIR_GROUP_SCHED 6343 root_task_group.se = (struct sched_entity **)ptr; 6344 ptr += nr_cpu_ids * sizeof(void **); 6345 6346 root_task_group.cfs_rq = (struct cfs_rq **)ptr; 6347 ptr += nr_cpu_ids * sizeof(void **); 6348 6349#endif /* CONFIG_FAIR_GROUP_SCHED */ 6350#ifdef CONFIG_RT_GROUP_SCHED 6351 root_task_group.rt_se = (struct sched_rt_entity **)ptr; 6352 ptr += nr_cpu_ids * sizeof(void **); 6353 6354 root_task_group.rt_rq = (struct rt_rq **)ptr; 6355 ptr += nr_cpu_ids * sizeof(void **); 6356 6357#endif /* CONFIG_RT_GROUP_SCHED */ 6358#ifdef CONFIG_CPUMASK_OFFSTACK 6359 for_each_possible_cpu(i) { 6360 per_cpu(load_balance_mask, i) = (void *)ptr; 6361 ptr += cpumask_size(); 6362 } 6363#endif /* CONFIG_CPUMASK_OFFSTACK */ 6364 } 6365 6366#ifdef CONFIG_SMP 6367 init_defrootdomain(); 6368#endif 6369 6370 init_rt_bandwidth(&def_rt_bandwidth, 6371 global_rt_period(), global_rt_runtime()); 6372 6373#ifdef CONFIG_RT_GROUP_SCHED 6374 init_rt_bandwidth(&root_task_group.rt_bandwidth, 6375 global_rt_period(), global_rt_runtime()); 6376#endif /* CONFIG_RT_GROUP_SCHED */ 6377 6378#ifdef CONFIG_CGROUP_SCHED 6379 list_add(&root_task_group.list, &task_groups); 6380 INIT_LIST_HEAD(&root_task_group.children); 6381 INIT_LIST_HEAD(&root_task_group.siblings); 6382 autogroup_init(&init_task); 6383 6384#endif /* CONFIG_CGROUP_SCHED */ 6385 6386 for_each_possible_cpu(i) { 6387 struct rq *rq; 6388 6389 rq = cpu_rq(i); 6390 raw_spin_lock_init(&rq->lock); 6391 rq->nr_running = 0; 6392 rq->calc_load_active = 0; 6393 rq->calc_load_update = jiffies + LOAD_FREQ; 6394 init_cfs_rq(&rq->cfs); 6395 init_rt_rq(&rq->rt, rq); 6396#ifdef CONFIG_FAIR_GROUP_SCHED 6397 root_task_group.shares = ROOT_TASK_GROUP_LOAD; 6398 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list); 6399 /* 6400 * How much cpu bandwidth does root_task_group get? 6401 * 6402 * In case of task-groups formed thr' the cgroup filesystem, it 6403 * gets 100% of the cpu resources in the system. This overall 6404 * system cpu resource is divided among the tasks of 6405 * root_task_group and its child task-groups in a fair manner, 6406 * based on each entity's (task or task-group's) weight 6407 * (se->load.weight). 6408 * 6409 * In other words, if root_task_group has 10 tasks of weight 6410 * 1024) and two child groups A0 and A1 (of weight 1024 each), 6411 * then A0's share of the cpu resource is: 6412 * 6413 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33% 6414 * 6415 * We achieve this by letting root_task_group's tasks sit 6416 * directly in rq->cfs (i.e root_task_group->se[] = NULL). 6417 */ 6418 init_cfs_bandwidth(&root_task_group.cfs_bandwidth); 6419 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL); 6420#endif /* CONFIG_FAIR_GROUP_SCHED */ 6421 6422 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime; 6423#ifdef CONFIG_RT_GROUP_SCHED 6424 INIT_LIST_HEAD(&rq->leaf_rt_rq_list); 6425 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL); 6426#endif 6427 6428 for (j = 0; j < CPU_LOAD_IDX_MAX; j++) 6429 rq->cpu_load[j] = 0; 6430 6431 rq->last_load_update_tick = jiffies; 6432 6433#ifdef CONFIG_SMP 6434 rq->sd = NULL; 6435 rq->rd = NULL; 6436 rq->cpu_power = SCHED_POWER_SCALE; 6437 rq->post_schedule = 0; 6438 rq->active_balance = 0; 6439 rq->next_balance = jiffies; 6440 rq->push_cpu = 0; 6441 rq->cpu = i; 6442 rq->online = 0; 6443 rq->idle_stamp = 0; 6444 rq->avg_idle = 2*sysctl_sched_migration_cost; 6445 6446 INIT_LIST_HEAD(&rq->cfs_tasks); 6447 6448 rq_attach_root(rq, &def_root_domain); 6449#ifdef CONFIG_NO_HZ_COMMON 6450 rq->nohz_flags = 0; 6451#endif 6452#ifdef CONFIG_NO_HZ_FULL 6453 rq->last_sched_tick = 0; 6454#endif 6455#endif 6456 init_rq_hrtick(rq); 6457 atomic_set(&rq->nr_iowait, 0); 6458 } 6459 6460 set_load_weight(&init_task); 6461 6462#ifdef CONFIG_PREEMPT_NOTIFIERS 6463 INIT_HLIST_HEAD(&init_task.preempt_notifiers); 6464#endif 6465 6466#ifdef CONFIG_RT_MUTEXES 6467 plist_head_init(&init_task.pi_waiters); 6468#endif 6469 6470 /* 6471 * The boot idle thread does lazy MMU switching as well: 6472 */ 6473 atomic_inc(&init_mm.mm_count); 6474 enter_lazy_tlb(&init_mm, current); 6475 6476 /* 6477 * Make us the idle thread. Technically, schedule() should not be 6478 * called from this thread, however somewhere below it might be, 6479 * but because we are the idle thread, we just pick up running again 6480 * when this runqueue becomes "idle". 6481 */ 6482 init_idle(current, smp_processor_id()); 6483 6484 calc_load_update = jiffies + LOAD_FREQ; 6485 6486 /* 6487 * During early bootup we pretend to be a normal task: 6488 */ 6489 current->sched_class = &fair_sched_class; 6490 6491#ifdef CONFIG_SMP 6492 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT); 6493 /* May be allocated at isolcpus cmdline parse time */ 6494 if (cpu_isolated_map == NULL) 6495 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT); 6496 idle_thread_set_boot_cpu(); 6497#endif 6498 init_sched_fair_class(); 6499 6500 scheduler_running = 1; 6501} 6502 6503#ifdef CONFIG_DEBUG_ATOMIC_SLEEP 6504static inline int preempt_count_equals(int preempt_offset) 6505{ 6506 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth(); 6507 6508 return (nested == preempt_offset); 6509} 6510 6511void __might_sleep(const char *file, int line, int preempt_offset) 6512{ 6513 static unsigned long prev_jiffy; /* ratelimiting */ 6514 6515 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */ 6516 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) || 6517 system_state != SYSTEM_RUNNING || oops_in_progress) 6518 return; 6519 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy) 6520 return; 6521 prev_jiffy = jiffies; 6522 6523 printk(KERN_ERR 6524 "BUG: sleeping function called from invalid context at %s:%d\n", 6525 file, line); 6526 printk(KERN_ERR 6527 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n", 6528 in_atomic(), irqs_disabled(), 6529 current->pid, current->comm); 6530 6531 debug_show_held_locks(current); 6532 if (irqs_disabled()) 6533 print_irqtrace_events(current); 6534 dump_stack(); 6535} 6536EXPORT_SYMBOL(__might_sleep); 6537#endif 6538 6539#ifdef CONFIG_MAGIC_SYSRQ 6540static void normalize_task(struct rq *rq, struct task_struct *p) 6541{ 6542 const struct sched_class *prev_class = p->sched_class; 6543 int old_prio = p->prio; 6544 int on_rq; 6545 6546 on_rq = p->on_rq; 6547 if (on_rq) 6548 dequeue_task(rq, p, 0); 6549 __setscheduler(rq, p, SCHED_NORMAL, 0); 6550 if (on_rq) { 6551 enqueue_task(rq, p, 0); 6552 resched_task(rq->curr); 6553 } 6554 6555 check_class_changed(rq, p, prev_class, old_prio); 6556} 6557 6558void normalize_rt_tasks(void) 6559{ 6560 struct task_struct *g, *p; 6561 unsigned long flags; 6562 struct rq *rq; 6563 6564 read_lock_irqsave(&tasklist_lock, flags); 6565 do_each_thread(g, p) { 6566 /* 6567 * Only normalize user tasks: 6568 */ 6569 if (!p->mm) 6570 continue; 6571 6572 p->se.exec_start = 0; 6573#ifdef CONFIG_SCHEDSTATS 6574 p->se.statistics.wait_start = 0; 6575 p->se.statistics.sleep_start = 0; 6576 p->se.statistics.block_start = 0; 6577#endif 6578 6579 if (!rt_task(p)) { 6580 /* 6581 * Renice negative nice level userspace 6582 * tasks back to 0: 6583 */ 6584 if (TASK_NICE(p) < 0 && p->mm) 6585 set_user_nice(p, 0); 6586 continue; 6587 } 6588 6589 raw_spin_lock(&p->pi_lock); 6590 rq = __task_rq_lock(p); 6591 6592 normalize_task(rq, p); 6593 6594 __task_rq_unlock(rq); 6595 raw_spin_unlock(&p->pi_lock); 6596 } while_each_thread(g, p); 6597 6598 read_unlock_irqrestore(&tasklist_lock, flags); 6599} 6600 6601#endif /* CONFIG_MAGIC_SYSRQ */ 6602 6603#if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) 6604/* 6605 * These functions are only useful for the IA64 MCA handling, or kdb. 6606 * 6607 * They can only be called when the whole system has been 6608 * stopped - every CPU needs to be quiescent, and no scheduling 6609 * activity can take place. Using them for anything else would 6610 * be a serious bug, and as a result, they aren't even visible 6611 * under any other configuration. 6612 */ 6613 6614/** 6615 * curr_task - return the current task for a given cpu. 6616 * @cpu: the processor in question. 6617 * 6618 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED! 6619 */ 6620struct task_struct *curr_task(int cpu) 6621{ 6622 return cpu_curr(cpu); 6623} 6624 6625#endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */ 6626 6627#ifdef CONFIG_IA64 6628/** 6629 * set_curr_task - set the current task for a given cpu. 6630 * @cpu: the processor in question. 6631 * @p: the task pointer to set. 6632 * 6633 * Description: This function must only be used when non-maskable interrupts 6634 * are serviced on a separate stack. It allows the architecture to switch the 6635 * notion of the current task on a cpu in a non-blocking manner. This function 6636 * must be called with all CPU's synchronized, and interrupts disabled, the 6637 * and caller must save the original value of the current task (see 6638 * curr_task() above) and restore that value before reenabling interrupts and 6639 * re-starting the system. 6640 * 6641 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED! 6642 */ 6643void set_curr_task(int cpu, struct task_struct *p) 6644{ 6645 cpu_curr(cpu) = p; 6646} 6647 6648#endif 6649 6650#ifdef CONFIG_CGROUP_SCHED 6651/* task_group_lock serializes the addition/removal of task groups */ 6652static DEFINE_SPINLOCK(task_group_lock); 6653 6654static void free_sched_group(struct task_group *tg) 6655{ 6656 free_fair_sched_group(tg); 6657 free_rt_sched_group(tg); 6658 autogroup_free(tg); 6659 kfree(tg); 6660} 6661 6662/* allocate runqueue etc for a new task group */ 6663struct task_group *sched_create_group(struct task_group *parent) 6664{ 6665 struct task_group *tg; 6666 6667 tg = kzalloc(sizeof(*tg), GFP_KERNEL); 6668 if (!tg) 6669 return ERR_PTR(-ENOMEM); 6670 6671 if (!alloc_fair_sched_group(tg, parent)) 6672 goto err; 6673 6674 if (!alloc_rt_sched_group(tg, parent)) 6675 goto err; 6676 6677 return tg; 6678 6679err: 6680 free_sched_group(tg); 6681 return ERR_PTR(-ENOMEM); 6682} 6683 6684void sched_online_group(struct task_group *tg, struct task_group *parent) 6685{ 6686 unsigned long flags; 6687 6688 spin_lock_irqsave(&task_group_lock, flags); 6689 list_add_rcu(&tg->list, &task_groups); 6690 6691 WARN_ON(!parent); /* root should already exist */ 6692 6693 tg->parent = parent; 6694 INIT_LIST_HEAD(&tg->children); 6695 list_add_rcu(&tg->siblings, &parent->children); 6696 spin_unlock_irqrestore(&task_group_lock, flags); 6697} 6698 6699/* rcu callback to free various structures associated with a task group */ 6700static void free_sched_group_rcu(struct rcu_head *rhp) 6701{ 6702 /* now it should be safe to free those cfs_rqs */ 6703 free_sched_group(container_of(rhp, struct task_group, rcu)); 6704} 6705 6706/* Destroy runqueue etc associated with a task group */ 6707void sched_destroy_group(struct task_group *tg) 6708{ 6709 /* wait for possible concurrent references to cfs_rqs complete */ 6710 call_rcu(&tg->rcu, free_sched_group_rcu); 6711} 6712 6713void sched_offline_group(struct task_group *tg) 6714{ 6715 unsigned long flags; 6716 int i; 6717 6718 /* end participation in shares distribution */ 6719 for_each_possible_cpu(i) 6720 unregister_fair_sched_group(tg, i); 6721 6722 spin_lock_irqsave(&task_group_lock, flags); 6723 list_del_rcu(&tg->list); 6724 list_del_rcu(&tg->siblings); 6725 spin_unlock_irqrestore(&task_group_lock, flags); 6726} 6727 6728/* change task's runqueue when it moves between groups. 6729 * The caller of this function should have put the task in its new group 6730 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to 6731 * reflect its new group. 6732 */ 6733void sched_move_task(struct task_struct *tsk) 6734{ 6735 struct task_group *tg; 6736 int on_rq, running; 6737 unsigned long flags; 6738 struct rq *rq; 6739 6740 rq = task_rq_lock(tsk, &flags); 6741 6742 running = task_current(rq, tsk); 6743 on_rq = tsk->on_rq; 6744 6745 if (on_rq) 6746 dequeue_task(rq, tsk, 0); 6747 if (unlikely(running)) 6748 tsk->sched_class->put_prev_task(rq, tsk); 6749 6750 tg = container_of(task_subsys_state_check(tsk, cpu_cgroup_subsys_id, 6751 lockdep_is_held(&tsk->sighand->siglock)), 6752 struct task_group, css); 6753 tg = autogroup_task_group(tsk, tg); 6754 tsk->sched_task_group = tg; 6755 6756#ifdef CONFIG_FAIR_GROUP_SCHED 6757 if (tsk->sched_class->task_move_group) 6758 tsk->sched_class->task_move_group(tsk, on_rq); 6759 else 6760#endif 6761 set_task_rq(tsk, task_cpu(tsk)); 6762 6763 if (unlikely(running)) 6764 tsk->sched_class->set_curr_task(rq); 6765 if (on_rq) 6766 enqueue_task(rq, tsk, 0); 6767 6768 task_rq_unlock(rq, tsk, &flags); 6769} 6770#endif /* CONFIG_CGROUP_SCHED */ 6771 6772#if defined(CONFIG_RT_GROUP_SCHED) || defined(CONFIG_CFS_BANDWIDTH) 6773static unsigned long to_ratio(u64 period, u64 runtime) 6774{ 6775 if (runtime == RUNTIME_INF) 6776 return 1ULL << 20; 6777 6778 return div64_u64(runtime << 20, period); 6779} 6780#endif 6781 6782#ifdef CONFIG_RT_GROUP_SCHED 6783/* 6784 * Ensure that the real time constraints are schedulable. 6785 */ 6786static DEFINE_MUTEX(rt_constraints_mutex); 6787 6788/* Must be called with tasklist_lock held */ 6789static inline int tg_has_rt_tasks(struct task_group *tg) 6790{ 6791 struct task_struct *g, *p; 6792 6793 do_each_thread(g, p) { 6794 if (rt_task(p) && task_rq(p)->rt.tg == tg) 6795 return 1; 6796 } while_each_thread(g, p); 6797 6798 return 0; 6799} 6800 6801struct rt_schedulable_data { 6802 struct task_group *tg; 6803 u64 rt_period; 6804 u64 rt_runtime; 6805}; 6806 6807static int tg_rt_schedulable(struct task_group *tg, void *data) 6808{ 6809 struct rt_schedulable_data *d = data; 6810 struct task_group *child; 6811 unsigned long total, sum = 0; 6812 u64 period, runtime; 6813 6814 period = ktime_to_ns(tg->rt_bandwidth.rt_period); 6815 runtime = tg->rt_bandwidth.rt_runtime; 6816 6817 if (tg == d->tg) { 6818 period = d->rt_period; 6819 runtime = d->rt_runtime; 6820 } 6821 6822 /* 6823 * Cannot have more runtime than the period. 6824 */ 6825 if (runtime > period && runtime != RUNTIME_INF) 6826 return -EINVAL; 6827 6828 /* 6829 * Ensure we don't starve existing RT tasks. 6830 */ 6831 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg)) 6832 return -EBUSY; 6833 6834 total = to_ratio(period, runtime); 6835 6836 /* 6837 * Nobody can have more than the global setting allows. 6838 */ 6839 if (total > to_ratio(global_rt_period(), global_rt_runtime())) 6840 return -EINVAL; 6841 6842 /* 6843 * The sum of our children's runtime should not exceed our own. 6844 */ 6845 list_for_each_entry_rcu(child, &tg->children, siblings) { 6846 period = ktime_to_ns(child->rt_bandwidth.rt_period); 6847 runtime = child->rt_bandwidth.rt_runtime; 6848 6849 if (child == d->tg) { 6850 period = d->rt_period; 6851 runtime = d->rt_runtime; 6852 } 6853 6854 sum += to_ratio(period, runtime); 6855 } 6856 6857 if (sum > total) 6858 return -EINVAL; 6859 6860 return 0; 6861} 6862 6863static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime) 6864{ 6865 int ret; 6866 6867 struct rt_schedulable_data data = { 6868 .tg = tg, 6869 .rt_period = period, 6870 .rt_runtime = runtime, 6871 }; 6872 6873 rcu_read_lock(); 6874 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data); 6875 rcu_read_unlock(); 6876 6877 return ret; 6878} 6879 6880static int tg_set_rt_bandwidth(struct task_group *tg, 6881 u64 rt_period, u64 rt_runtime) 6882{ 6883 int i, err = 0; 6884 6885 mutex_lock(&rt_constraints_mutex); 6886 read_lock(&tasklist_lock); 6887 err = __rt_schedulable(tg, rt_period, rt_runtime); 6888 if (err) 6889 goto unlock; 6890 6891 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock); 6892 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period); 6893 tg->rt_bandwidth.rt_runtime = rt_runtime; 6894 6895 for_each_possible_cpu(i) { 6896 struct rt_rq *rt_rq = tg->rt_rq[i]; 6897 6898 raw_spin_lock(&rt_rq->rt_runtime_lock); 6899 rt_rq->rt_runtime = rt_runtime; 6900 raw_spin_unlock(&rt_rq->rt_runtime_lock); 6901 } 6902 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock); 6903unlock: 6904 read_unlock(&tasklist_lock); 6905 mutex_unlock(&rt_constraints_mutex); 6906 6907 return err; 6908} 6909 6910static int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us) 6911{ 6912 u64 rt_runtime, rt_period; 6913 6914 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period); 6915 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC; 6916 if (rt_runtime_us < 0) 6917 rt_runtime = RUNTIME_INF; 6918 6919 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime); 6920} 6921 6922static long sched_group_rt_runtime(struct task_group *tg) 6923{ 6924 u64 rt_runtime_us; 6925 6926 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF) 6927 return -1; 6928 6929 rt_runtime_us = tg->rt_bandwidth.rt_runtime; 6930 do_div(rt_runtime_us, NSEC_PER_USEC); 6931 return rt_runtime_us; 6932} 6933 6934static int sched_group_set_rt_period(struct task_group *tg, long rt_period_us) 6935{ 6936 u64 rt_runtime, rt_period; 6937 6938 rt_period = (u64)rt_period_us * NSEC_PER_USEC; 6939 rt_runtime = tg->rt_bandwidth.rt_runtime; 6940 6941 if (rt_period == 0) 6942 return -EINVAL; 6943 6944 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime); 6945} 6946 6947static long sched_group_rt_period(struct task_group *tg) 6948{ 6949 u64 rt_period_us; 6950 6951 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period); 6952 do_div(rt_period_us, NSEC_PER_USEC); 6953 return rt_period_us; 6954} 6955 6956static int sched_rt_global_constraints(void) 6957{ 6958 u64 runtime, period; 6959 int ret = 0; 6960 6961 if (sysctl_sched_rt_period <= 0) 6962 return -EINVAL; 6963 6964 runtime = global_rt_runtime(); 6965 period = global_rt_period(); 6966 6967 /* 6968 * Sanity check on the sysctl variables. 6969 */ 6970 if (runtime > period && runtime != RUNTIME_INF) 6971 return -EINVAL; 6972 6973 mutex_lock(&rt_constraints_mutex); 6974 read_lock(&tasklist_lock); 6975 ret = __rt_schedulable(NULL, 0, 0); 6976 read_unlock(&tasklist_lock); 6977 mutex_unlock(&rt_constraints_mutex); 6978 6979 return ret; 6980} 6981 6982static int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk) 6983{ 6984 /* Don't accept realtime tasks when there is no way for them to run */ 6985 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0) 6986 return 0; 6987 6988 return 1; 6989} 6990 6991#else /* !CONFIG_RT_GROUP_SCHED */ 6992static int sched_rt_global_constraints(void) 6993{ 6994 unsigned long flags; 6995 int i; 6996 6997 if (sysctl_sched_rt_period <= 0) 6998 return -EINVAL; 6999 7000 /* 7001 * There's always some RT tasks in the root group 7002 * -- migration, kstopmachine etc.. 7003 */ 7004 if (sysctl_sched_rt_runtime == 0) 7005 return -EBUSY; 7006 7007 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags); 7008 for_each_possible_cpu(i) { 7009 struct rt_rq *rt_rq = &cpu_rq(i)->rt; 7010 7011 raw_spin_lock(&rt_rq->rt_runtime_lock); 7012 rt_rq->rt_runtime = global_rt_runtime(); 7013 raw_spin_unlock(&rt_rq->rt_runtime_lock); 7014 } 7015 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags); 7016 7017 return 0; 7018} 7019#endif /* CONFIG_RT_GROUP_SCHED */ 7020 7021int sched_rr_handler(struct ctl_table *table, int write, 7022 void __user *buffer, size_t *lenp, 7023 loff_t *ppos) 7024{ 7025 int ret; 7026 static DEFINE_MUTEX(mutex); 7027 7028 mutex_lock(&mutex); 7029 ret = proc_dointvec(table, write, buffer, lenp, ppos); 7030 /* make sure that internally we keep jiffies */ 7031 /* also, writing zero resets timeslice to default */ 7032 if (!ret && write) { 7033 sched_rr_timeslice = sched_rr_timeslice <= 0 ? 7034 RR_TIMESLICE : msecs_to_jiffies(sched_rr_timeslice); 7035 } 7036 mutex_unlock(&mutex); 7037 return ret; 7038} 7039 7040int sched_rt_handler(struct ctl_table *table, int write, 7041 void __user *buffer, size_t *lenp, 7042 loff_t *ppos) 7043{ 7044 int ret; 7045 int old_period, old_runtime; 7046 static DEFINE_MUTEX(mutex); 7047 7048 mutex_lock(&mutex); 7049 old_period = sysctl_sched_rt_period; 7050 old_runtime = sysctl_sched_rt_runtime; 7051 7052 ret = proc_dointvec(table, write, buffer, lenp, ppos); 7053 7054 if (!ret && write) { 7055 ret = sched_rt_global_constraints(); 7056 if (ret) { 7057 sysctl_sched_rt_period = old_period; 7058 sysctl_sched_rt_runtime = old_runtime; 7059 } else { 7060 def_rt_bandwidth.rt_runtime = global_rt_runtime(); 7061 def_rt_bandwidth.rt_period = 7062 ns_to_ktime(global_rt_period()); 7063 } 7064 } 7065 mutex_unlock(&mutex); 7066 7067 return ret; 7068} 7069 7070#ifdef CONFIG_CGROUP_SCHED 7071 7072/* return corresponding task_group object of a cgroup */ 7073static inline struct task_group *cgroup_tg(struct cgroup *cgrp) 7074{ 7075 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id), 7076 struct task_group, css); 7077} 7078 7079static struct cgroup_subsys_state *cpu_cgroup_css_alloc(struct cgroup *cgrp) 7080{ 7081 struct task_group *tg, *parent; 7082 7083 if (!cgrp->parent) { 7084 /* This is early initialization for the top cgroup */ 7085 return &root_task_group.css; 7086 } 7087 7088 parent = cgroup_tg(cgrp->parent); 7089 tg = sched_create_group(parent); 7090 if (IS_ERR(tg)) 7091 return ERR_PTR(-ENOMEM); 7092 7093 return &tg->css; 7094} 7095 7096static int cpu_cgroup_css_online(struct cgroup *cgrp) 7097{ 7098 struct task_group *tg = cgroup_tg(cgrp); 7099 struct task_group *parent; 7100 7101 if (!cgrp->parent) 7102 return 0; 7103 7104 parent = cgroup_tg(cgrp->parent); 7105 sched_online_group(tg, parent); 7106 return 0; 7107} 7108 7109static void cpu_cgroup_css_free(struct cgroup *cgrp) 7110{ 7111 struct task_group *tg = cgroup_tg(cgrp); 7112 7113 sched_destroy_group(tg); 7114} 7115 7116static void cpu_cgroup_css_offline(struct cgroup *cgrp) 7117{ 7118 struct task_group *tg = cgroup_tg(cgrp); 7119 7120 sched_offline_group(tg); 7121} 7122 7123static int cpu_cgroup_can_attach(struct cgroup *cgrp, 7124 struct cgroup_taskset *tset) 7125{ 7126 struct task_struct *task; 7127 7128 cgroup_taskset_for_each(task, cgrp, tset) { 7129#ifdef CONFIG_RT_GROUP_SCHED 7130 if (!sched_rt_can_attach(cgroup_tg(cgrp), task)) 7131 return -EINVAL; 7132#else 7133 /* We don't support RT-tasks being in separate groups */ 7134 if (task->sched_class != &fair_sched_class) 7135 return -EINVAL; 7136#endif 7137 } 7138 return 0; 7139} 7140 7141static void cpu_cgroup_attach(struct cgroup *cgrp, 7142 struct cgroup_taskset *tset) 7143{ 7144 struct task_struct *task; 7145 7146 cgroup_taskset_for_each(task, cgrp, tset) 7147 sched_move_task(task); 7148} 7149 7150static void 7151cpu_cgroup_exit(struct cgroup *cgrp, struct cgroup *old_cgrp, 7152 struct task_struct *task) 7153{ 7154 /* 7155 * cgroup_exit() is called in the copy_process() failure path. 7156 * Ignore this case since the task hasn't ran yet, this avoids 7157 * trying to poke a half freed task state from generic code. 7158 */ 7159 if (!(task->flags & PF_EXITING)) 7160 return; 7161 7162 sched_move_task(task); 7163} 7164 7165#ifdef CONFIG_FAIR_GROUP_SCHED 7166static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype, 7167 u64 shareval) 7168{ 7169 return sched_group_set_shares(cgroup_tg(cgrp), scale_load(shareval)); 7170} 7171 7172static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft) 7173{ 7174 struct task_group *tg = cgroup_tg(cgrp); 7175 7176 return (u64) scale_load_down(tg->shares); 7177} 7178 7179#ifdef CONFIG_CFS_BANDWIDTH 7180static DEFINE_MUTEX(cfs_constraints_mutex); 7181 7182const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */ 7183const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */ 7184 7185static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime); 7186 7187static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota) 7188{ 7189 int i, ret = 0, runtime_enabled, runtime_was_enabled; 7190 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth; 7191 7192 if (tg == &root_task_group) 7193 return -EINVAL; 7194 7195 /* 7196 * Ensure we have at some amount of bandwidth every period. This is 7197 * to prevent reaching a state of large arrears when throttled via 7198 * entity_tick() resulting in prolonged exit starvation. 7199 */ 7200 if (quota < min_cfs_quota_period || period < min_cfs_quota_period) 7201 return -EINVAL; 7202 7203 /* 7204 * Likewise, bound things on the otherside by preventing insane quota 7205 * periods. This also allows us to normalize in computing quota 7206 * feasibility. 7207 */ 7208 if (period > max_cfs_quota_period) 7209 return -EINVAL; 7210 7211 mutex_lock(&cfs_constraints_mutex); 7212 ret = __cfs_schedulable(tg, period, quota); 7213 if (ret) 7214 goto out_unlock; 7215 7216 runtime_enabled = quota != RUNTIME_INF; 7217 runtime_was_enabled = cfs_b->quota != RUNTIME_INF; 7218 account_cfs_bandwidth_used(runtime_enabled, runtime_was_enabled); 7219 raw_spin_lock_irq(&cfs_b->lock); 7220 cfs_b->period = ns_to_ktime(period); 7221 cfs_b->quota = quota; 7222 7223 __refill_cfs_bandwidth_runtime(cfs_b); 7224 /* restart the period timer (if active) to handle new period expiry */ 7225 if (runtime_enabled && cfs_b->timer_active) { 7226 /* force a reprogram */ 7227 cfs_b->timer_active = 0; 7228 __start_cfs_bandwidth(cfs_b); 7229 } 7230 raw_spin_unlock_irq(&cfs_b->lock); 7231 7232 for_each_possible_cpu(i) { 7233 struct cfs_rq *cfs_rq = tg->cfs_rq[i]; 7234 struct rq *rq = cfs_rq->rq; 7235 7236 raw_spin_lock_irq(&rq->lock); 7237 cfs_rq->runtime_enabled = runtime_enabled; 7238 cfs_rq->runtime_remaining = 0; 7239 7240 if (cfs_rq->throttled) 7241 unthrottle_cfs_rq(cfs_rq); 7242 raw_spin_unlock_irq(&rq->lock); 7243 } 7244out_unlock: 7245 mutex_unlock(&cfs_constraints_mutex); 7246 7247 return ret; 7248} 7249 7250int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us) 7251{ 7252 u64 quota, period; 7253 7254 period = ktime_to_ns(tg->cfs_bandwidth.period); 7255 if (cfs_quota_us < 0) 7256 quota = RUNTIME_INF; 7257 else 7258 quota = (u64)cfs_quota_us * NSEC_PER_USEC; 7259 7260 return tg_set_cfs_bandwidth(tg, period, quota); 7261} 7262 7263long tg_get_cfs_quota(struct task_group *tg) 7264{ 7265 u64 quota_us; 7266 7267 if (tg->cfs_bandwidth.quota == RUNTIME_INF) 7268 return -1; 7269 7270 quota_us = tg->cfs_bandwidth.quota; 7271 do_div(quota_us, NSEC_PER_USEC); 7272 7273 return quota_us; 7274} 7275 7276int tg_set_cfs_period(struct task_group *tg, long cfs_period_us) 7277{ 7278 u64 quota, period; 7279 7280 period = (u64)cfs_period_us * NSEC_PER_USEC; 7281 quota = tg->cfs_bandwidth.quota; 7282 7283 return tg_set_cfs_bandwidth(tg, period, quota); 7284} 7285 7286long tg_get_cfs_period(struct task_group *tg) 7287{ 7288 u64 cfs_period_us; 7289 7290 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period); 7291 do_div(cfs_period_us, NSEC_PER_USEC); 7292 7293 return cfs_period_us; 7294} 7295 7296static s64 cpu_cfs_quota_read_s64(struct cgroup *cgrp, struct cftype *cft) 7297{ 7298 return tg_get_cfs_quota(cgroup_tg(cgrp)); 7299} 7300 7301static int cpu_cfs_quota_write_s64(struct cgroup *cgrp, struct cftype *cftype, 7302 s64 cfs_quota_us) 7303{ 7304 return tg_set_cfs_quota(cgroup_tg(cgrp), cfs_quota_us); 7305} 7306 7307static u64 cpu_cfs_period_read_u64(struct cgroup *cgrp, struct cftype *cft) 7308{ 7309 return tg_get_cfs_period(cgroup_tg(cgrp)); 7310} 7311 7312static int cpu_cfs_period_write_u64(struct cgroup *cgrp, struct cftype *cftype, 7313 u64 cfs_period_us) 7314{ 7315 return tg_set_cfs_period(cgroup_tg(cgrp), cfs_period_us); 7316} 7317 7318struct cfs_schedulable_data { 7319 struct task_group *tg; 7320 u64 period, quota; 7321}; 7322 7323/* 7324 * normalize group quota/period to be quota/max_period 7325 * note: units are usecs 7326 */ 7327static u64 normalize_cfs_quota(struct task_group *tg, 7328 struct cfs_schedulable_data *d) 7329{ 7330 u64 quota, period; 7331 7332 if (tg == d->tg) { 7333 period = d->period; 7334 quota = d->quota; 7335 } else { 7336 period = tg_get_cfs_period(tg); 7337 quota = tg_get_cfs_quota(tg); 7338 } 7339 7340 /* note: these should typically be equivalent */ 7341 if (quota == RUNTIME_INF || quota == -1) 7342 return RUNTIME_INF; 7343 7344 return to_ratio(period, quota); 7345} 7346 7347static int tg_cfs_schedulable_down(struct task_group *tg, void *data) 7348{ 7349 struct cfs_schedulable_data *d = data; 7350 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth; 7351 s64 quota = 0, parent_quota = -1; 7352 7353 if (!tg->parent) { 7354 quota = RUNTIME_INF; 7355 } else { 7356 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth; 7357 7358 quota = normalize_cfs_quota(tg, d); 7359 parent_quota = parent_b->hierarchal_quota; 7360 7361 /* 7362 * ensure max(child_quota) <= parent_quota, inherit when no 7363 * limit is set 7364 */ 7365 if (quota == RUNTIME_INF) 7366 quota = parent_quota; 7367 else if (parent_quota != RUNTIME_INF && quota > parent_quota) 7368 return -EINVAL; 7369 } 7370 cfs_b->hierarchal_quota = quota; 7371 7372 return 0; 7373} 7374 7375static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota) 7376{ 7377 int ret; 7378 struct cfs_schedulable_data data = { 7379 .tg = tg, 7380 .period = period, 7381 .quota = quota, 7382 }; 7383 7384 if (quota != RUNTIME_INF) { 7385 do_div(data.period, NSEC_PER_USEC); 7386 do_div(data.quota, NSEC_PER_USEC); 7387 } 7388 7389 rcu_read_lock(); 7390 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data); 7391 rcu_read_unlock(); 7392 7393 return ret; 7394} 7395 7396static int cpu_stats_show(struct cgroup *cgrp, struct cftype *cft, 7397 struct cgroup_map_cb *cb) 7398{ 7399 struct task_group *tg = cgroup_tg(cgrp); 7400 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth; 7401 7402 cb->fill(cb, "nr_periods", cfs_b->nr_periods); 7403 cb->fill(cb, "nr_throttled", cfs_b->nr_throttled); 7404 cb->fill(cb, "throttled_time", cfs_b->throttled_time); 7405 7406 return 0; 7407} 7408#endif /* CONFIG_CFS_BANDWIDTH */ 7409#endif /* CONFIG_FAIR_GROUP_SCHED */ 7410 7411#ifdef CONFIG_RT_GROUP_SCHED 7412static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft, 7413 s64 val) 7414{ 7415 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val); 7416} 7417 7418static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft) 7419{ 7420 return sched_group_rt_runtime(cgroup_tg(cgrp)); 7421} 7422 7423static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype, 7424 u64 rt_period_us) 7425{ 7426 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us); 7427} 7428 7429static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft) 7430{ 7431 return sched_group_rt_period(cgroup_tg(cgrp)); 7432} 7433#endif /* CONFIG_RT_GROUP_SCHED */ 7434 7435static struct cftype cpu_files[] = { 7436#ifdef CONFIG_FAIR_GROUP_SCHED 7437 { 7438 .name = "shares", 7439 .read_u64 = cpu_shares_read_u64, 7440 .write_u64 = cpu_shares_write_u64, 7441 }, 7442#endif 7443#ifdef CONFIG_CFS_BANDWIDTH 7444 { 7445 .name = "cfs_quota_us", 7446 .read_s64 = cpu_cfs_quota_read_s64, 7447 .write_s64 = cpu_cfs_quota_write_s64, 7448 }, 7449 { 7450 .name = "cfs_period_us", 7451 .read_u64 = cpu_cfs_period_read_u64, 7452 .write_u64 = cpu_cfs_period_write_u64, 7453 }, 7454 { 7455 .name = "stat", 7456 .read_map = cpu_stats_show, 7457 }, 7458#endif 7459#ifdef CONFIG_RT_GROUP_SCHED 7460 { 7461 .name = "rt_runtime_us", 7462 .read_s64 = cpu_rt_runtime_read, 7463 .write_s64 = cpu_rt_runtime_write, 7464 }, 7465 { 7466 .name = "rt_period_us", 7467 .read_u64 = cpu_rt_period_read_uint, 7468 .write_u64 = cpu_rt_period_write_uint, 7469 }, 7470#endif 7471 { } /* terminate */ 7472}; 7473 7474struct cgroup_subsys cpu_cgroup_subsys = { 7475 .name = "cpu", 7476 .css_alloc = cpu_cgroup_css_alloc, 7477 .css_free = cpu_cgroup_css_free, 7478 .css_online = cpu_cgroup_css_online, 7479 .css_offline = cpu_cgroup_css_offline, 7480 .can_attach = cpu_cgroup_can_attach, 7481 .attach = cpu_cgroup_attach, 7482 .exit = cpu_cgroup_exit, 7483 .subsys_id = cpu_cgroup_subsys_id, 7484 .base_cftypes = cpu_files, 7485 .early_init = 1, 7486}; 7487 7488#endif /* CONFIG_CGROUP_SCHED */ 7489 7490void dump_cpu_task(int cpu) 7491{ 7492 pr_info("Task dump for CPU %d:\n", cpu); 7493 sched_show_task(cpu_curr(cpu)); 7494} 7495