core.c revision bdb43806589096ac4272fe1307e789846ac08d7c
1/* 2 * kernel/sched/core.c 3 * 4 * Kernel scheduler and related syscalls 5 * 6 * Copyright (C) 1991-2002 Linus Torvalds 7 * 8 * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and 9 * make semaphores SMP safe 10 * 1998-11-19 Implemented schedule_timeout() and related stuff 11 * by Andrea Arcangeli 12 * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar: 13 * hybrid priority-list and round-robin design with 14 * an array-switch method of distributing timeslices 15 * and per-CPU runqueues. Cleanups and useful suggestions 16 * by Davide Libenzi, preemptible kernel bits by Robert Love. 17 * 2003-09-03 Interactivity tuning by Con Kolivas. 18 * 2004-04-02 Scheduler domains code by Nick Piggin 19 * 2007-04-15 Work begun on replacing all interactivity tuning with a 20 * fair scheduling design by Con Kolivas. 21 * 2007-05-05 Load balancing (smp-nice) and other improvements 22 * by Peter Williams 23 * 2007-05-06 Interactivity improvements to CFS by Mike Galbraith 24 * 2007-07-01 Group scheduling enhancements by Srivatsa Vaddagiri 25 * 2007-11-29 RT balancing improvements by Steven Rostedt, Gregory Haskins, 26 * Thomas Gleixner, Mike Kravetz 27 */ 28 29#include <linux/mm.h> 30#include <linux/module.h> 31#include <linux/nmi.h> 32#include <linux/init.h> 33#include <linux/uaccess.h> 34#include <linux/highmem.h> 35#include <asm/mmu_context.h> 36#include <linux/interrupt.h> 37#include <linux/capability.h> 38#include <linux/completion.h> 39#include <linux/kernel_stat.h> 40#include <linux/debug_locks.h> 41#include <linux/perf_event.h> 42#include <linux/security.h> 43#include <linux/notifier.h> 44#include <linux/profile.h> 45#include <linux/freezer.h> 46#include <linux/vmalloc.h> 47#include <linux/blkdev.h> 48#include <linux/delay.h> 49#include <linux/pid_namespace.h> 50#include <linux/smp.h> 51#include <linux/threads.h> 52#include <linux/timer.h> 53#include <linux/rcupdate.h> 54#include <linux/cpu.h> 55#include <linux/cpuset.h> 56#include <linux/percpu.h> 57#include <linux/proc_fs.h> 58#include <linux/seq_file.h> 59#include <linux/sysctl.h> 60#include <linux/syscalls.h> 61#include <linux/times.h> 62#include <linux/tsacct_kern.h> 63#include <linux/kprobes.h> 64#include <linux/delayacct.h> 65#include <linux/unistd.h> 66#include <linux/pagemap.h> 67#include <linux/hrtimer.h> 68#include <linux/tick.h> 69#include <linux/debugfs.h> 70#include <linux/ctype.h> 71#include <linux/ftrace.h> 72#include <linux/slab.h> 73#include <linux/init_task.h> 74#include <linux/binfmts.h> 75#include <linux/context_tracking.h> 76 77#include <asm/switch_to.h> 78#include <asm/tlb.h> 79#include <asm/irq_regs.h> 80#include <asm/mutex.h> 81#ifdef CONFIG_PARAVIRT 82#include <asm/paravirt.h> 83#endif 84 85#include "sched.h" 86#include "../workqueue_internal.h" 87#include "../smpboot.h" 88 89#define CREATE_TRACE_POINTS 90#include <trace/events/sched.h> 91 92void start_bandwidth_timer(struct hrtimer *period_timer, ktime_t period) 93{ 94 unsigned long delta; 95 ktime_t soft, hard, now; 96 97 for (;;) { 98 if (hrtimer_active(period_timer)) 99 break; 100 101 now = hrtimer_cb_get_time(period_timer); 102 hrtimer_forward(period_timer, now, period); 103 104 soft = hrtimer_get_softexpires(period_timer); 105 hard = hrtimer_get_expires(period_timer); 106 delta = ktime_to_ns(ktime_sub(hard, soft)); 107 __hrtimer_start_range_ns(period_timer, soft, delta, 108 HRTIMER_MODE_ABS_PINNED, 0); 109 } 110} 111 112DEFINE_MUTEX(sched_domains_mutex); 113DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues); 114 115static void update_rq_clock_task(struct rq *rq, s64 delta); 116 117void update_rq_clock(struct rq *rq) 118{ 119 s64 delta; 120 121 if (rq->skip_clock_update > 0) 122 return; 123 124 delta = sched_clock_cpu(cpu_of(rq)) - rq->clock; 125 rq->clock += delta; 126 update_rq_clock_task(rq, delta); 127} 128 129/* 130 * Debugging: various feature bits 131 */ 132 133#define SCHED_FEAT(name, enabled) \ 134 (1UL << __SCHED_FEAT_##name) * enabled | 135 136const_debug unsigned int sysctl_sched_features = 137#include "features.h" 138 0; 139 140#undef SCHED_FEAT 141 142#ifdef CONFIG_SCHED_DEBUG 143#define SCHED_FEAT(name, enabled) \ 144 #name , 145 146static const char * const sched_feat_names[] = { 147#include "features.h" 148}; 149 150#undef SCHED_FEAT 151 152static int sched_feat_show(struct seq_file *m, void *v) 153{ 154 int i; 155 156 for (i = 0; i < __SCHED_FEAT_NR; i++) { 157 if (!(sysctl_sched_features & (1UL << i))) 158 seq_puts(m, "NO_"); 159 seq_printf(m, "%s ", sched_feat_names[i]); 160 } 161 seq_puts(m, "\n"); 162 163 return 0; 164} 165 166#ifdef HAVE_JUMP_LABEL 167 168#define jump_label_key__true STATIC_KEY_INIT_TRUE 169#define jump_label_key__false STATIC_KEY_INIT_FALSE 170 171#define SCHED_FEAT(name, enabled) \ 172 jump_label_key__##enabled , 173 174struct static_key sched_feat_keys[__SCHED_FEAT_NR] = { 175#include "features.h" 176}; 177 178#undef SCHED_FEAT 179 180static void sched_feat_disable(int i) 181{ 182 if (static_key_enabled(&sched_feat_keys[i])) 183 static_key_slow_dec(&sched_feat_keys[i]); 184} 185 186static void sched_feat_enable(int i) 187{ 188 if (!static_key_enabled(&sched_feat_keys[i])) 189 static_key_slow_inc(&sched_feat_keys[i]); 190} 191#else 192static void sched_feat_disable(int i) { }; 193static void sched_feat_enable(int i) { }; 194#endif /* HAVE_JUMP_LABEL */ 195 196static int sched_feat_set(char *cmp) 197{ 198 int i; 199 int neg = 0; 200 201 if (strncmp(cmp, "NO_", 3) == 0) { 202 neg = 1; 203 cmp += 3; 204 } 205 206 for (i = 0; i < __SCHED_FEAT_NR; i++) { 207 if (strcmp(cmp, sched_feat_names[i]) == 0) { 208 if (neg) { 209 sysctl_sched_features &= ~(1UL << i); 210 sched_feat_disable(i); 211 } else { 212 sysctl_sched_features |= (1UL << i); 213 sched_feat_enable(i); 214 } 215 break; 216 } 217 } 218 219 return i; 220} 221 222static ssize_t 223sched_feat_write(struct file *filp, const char __user *ubuf, 224 size_t cnt, loff_t *ppos) 225{ 226 char buf[64]; 227 char *cmp; 228 int i; 229 230 if (cnt > 63) 231 cnt = 63; 232 233 if (copy_from_user(&buf, ubuf, cnt)) 234 return -EFAULT; 235 236 buf[cnt] = 0; 237 cmp = strstrip(buf); 238 239 i = sched_feat_set(cmp); 240 if (i == __SCHED_FEAT_NR) 241 return -EINVAL; 242 243 *ppos += cnt; 244 245 return cnt; 246} 247 248static int sched_feat_open(struct inode *inode, struct file *filp) 249{ 250 return single_open(filp, sched_feat_show, NULL); 251} 252 253static const struct file_operations sched_feat_fops = { 254 .open = sched_feat_open, 255 .write = sched_feat_write, 256 .read = seq_read, 257 .llseek = seq_lseek, 258 .release = single_release, 259}; 260 261static __init int sched_init_debug(void) 262{ 263 debugfs_create_file("sched_features", 0644, NULL, NULL, 264 &sched_feat_fops); 265 266 return 0; 267} 268late_initcall(sched_init_debug); 269#endif /* CONFIG_SCHED_DEBUG */ 270 271/* 272 * Number of tasks to iterate in a single balance run. 273 * Limited because this is done with IRQs disabled. 274 */ 275const_debug unsigned int sysctl_sched_nr_migrate = 32; 276 277/* 278 * period over which we average the RT time consumption, measured 279 * in ms. 280 * 281 * default: 1s 282 */ 283const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC; 284 285/* 286 * period over which we measure -rt task cpu usage in us. 287 * default: 1s 288 */ 289unsigned int sysctl_sched_rt_period = 1000000; 290 291__read_mostly int scheduler_running; 292 293/* 294 * part of the period that we allow rt tasks to run in us. 295 * default: 0.95s 296 */ 297int sysctl_sched_rt_runtime = 950000; 298 299 300 301/* 302 * __task_rq_lock - lock the rq @p resides on. 303 */ 304static inline struct rq *__task_rq_lock(struct task_struct *p) 305 __acquires(rq->lock) 306{ 307 struct rq *rq; 308 309 lockdep_assert_held(&p->pi_lock); 310 311 for (;;) { 312 rq = task_rq(p); 313 raw_spin_lock(&rq->lock); 314 if (likely(rq == task_rq(p))) 315 return rq; 316 raw_spin_unlock(&rq->lock); 317 } 318} 319 320/* 321 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on. 322 */ 323static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags) 324 __acquires(p->pi_lock) 325 __acquires(rq->lock) 326{ 327 struct rq *rq; 328 329 for (;;) { 330 raw_spin_lock_irqsave(&p->pi_lock, *flags); 331 rq = task_rq(p); 332 raw_spin_lock(&rq->lock); 333 if (likely(rq == task_rq(p))) 334 return rq; 335 raw_spin_unlock(&rq->lock); 336 raw_spin_unlock_irqrestore(&p->pi_lock, *flags); 337 } 338} 339 340static void __task_rq_unlock(struct rq *rq) 341 __releases(rq->lock) 342{ 343 raw_spin_unlock(&rq->lock); 344} 345 346static inline void 347task_rq_unlock(struct rq *rq, struct task_struct *p, unsigned long *flags) 348 __releases(rq->lock) 349 __releases(p->pi_lock) 350{ 351 raw_spin_unlock(&rq->lock); 352 raw_spin_unlock_irqrestore(&p->pi_lock, *flags); 353} 354 355/* 356 * this_rq_lock - lock this runqueue and disable interrupts. 357 */ 358static struct rq *this_rq_lock(void) 359 __acquires(rq->lock) 360{ 361 struct rq *rq; 362 363 local_irq_disable(); 364 rq = this_rq(); 365 raw_spin_lock(&rq->lock); 366 367 return rq; 368} 369 370#ifdef CONFIG_SCHED_HRTICK 371/* 372 * Use HR-timers to deliver accurate preemption points. 373 */ 374 375static void hrtick_clear(struct rq *rq) 376{ 377 if (hrtimer_active(&rq->hrtick_timer)) 378 hrtimer_cancel(&rq->hrtick_timer); 379} 380 381/* 382 * High-resolution timer tick. 383 * Runs from hardirq context with interrupts disabled. 384 */ 385static enum hrtimer_restart hrtick(struct hrtimer *timer) 386{ 387 struct rq *rq = container_of(timer, struct rq, hrtick_timer); 388 389 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id()); 390 391 raw_spin_lock(&rq->lock); 392 update_rq_clock(rq); 393 rq->curr->sched_class->task_tick(rq, rq->curr, 1); 394 raw_spin_unlock(&rq->lock); 395 396 return HRTIMER_NORESTART; 397} 398 399#ifdef CONFIG_SMP 400 401static int __hrtick_restart(struct rq *rq) 402{ 403 struct hrtimer *timer = &rq->hrtick_timer; 404 ktime_t time = hrtimer_get_softexpires(timer); 405 406 return __hrtimer_start_range_ns(timer, time, 0, HRTIMER_MODE_ABS_PINNED, 0); 407} 408 409/* 410 * called from hardirq (IPI) context 411 */ 412static void __hrtick_start(void *arg) 413{ 414 struct rq *rq = arg; 415 416 raw_spin_lock(&rq->lock); 417 __hrtick_restart(rq); 418 rq->hrtick_csd_pending = 0; 419 raw_spin_unlock(&rq->lock); 420} 421 422/* 423 * Called to set the hrtick timer state. 424 * 425 * called with rq->lock held and irqs disabled 426 */ 427void hrtick_start(struct rq *rq, u64 delay) 428{ 429 struct hrtimer *timer = &rq->hrtick_timer; 430 ktime_t time = ktime_add_ns(timer->base->get_time(), delay); 431 432 hrtimer_set_expires(timer, time); 433 434 if (rq == this_rq()) { 435 __hrtick_restart(rq); 436 } else if (!rq->hrtick_csd_pending) { 437 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd, 0); 438 rq->hrtick_csd_pending = 1; 439 } 440} 441 442static int 443hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu) 444{ 445 int cpu = (int)(long)hcpu; 446 447 switch (action) { 448 case CPU_UP_CANCELED: 449 case CPU_UP_CANCELED_FROZEN: 450 case CPU_DOWN_PREPARE: 451 case CPU_DOWN_PREPARE_FROZEN: 452 case CPU_DEAD: 453 case CPU_DEAD_FROZEN: 454 hrtick_clear(cpu_rq(cpu)); 455 return NOTIFY_OK; 456 } 457 458 return NOTIFY_DONE; 459} 460 461static __init void init_hrtick(void) 462{ 463 hotcpu_notifier(hotplug_hrtick, 0); 464} 465#else 466/* 467 * Called to set the hrtick timer state. 468 * 469 * called with rq->lock held and irqs disabled 470 */ 471void hrtick_start(struct rq *rq, u64 delay) 472{ 473 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0, 474 HRTIMER_MODE_REL_PINNED, 0); 475} 476 477static inline void init_hrtick(void) 478{ 479} 480#endif /* CONFIG_SMP */ 481 482static void init_rq_hrtick(struct rq *rq) 483{ 484#ifdef CONFIG_SMP 485 rq->hrtick_csd_pending = 0; 486 487 rq->hrtick_csd.flags = 0; 488 rq->hrtick_csd.func = __hrtick_start; 489 rq->hrtick_csd.info = rq; 490#endif 491 492 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); 493 rq->hrtick_timer.function = hrtick; 494} 495#else /* CONFIG_SCHED_HRTICK */ 496static inline void hrtick_clear(struct rq *rq) 497{ 498} 499 500static inline void init_rq_hrtick(struct rq *rq) 501{ 502} 503 504static inline void init_hrtick(void) 505{ 506} 507#endif /* CONFIG_SCHED_HRTICK */ 508 509/* 510 * resched_task - mark a task 'to be rescheduled now'. 511 * 512 * On UP this means the setting of the need_resched flag, on SMP it 513 * might also involve a cross-CPU call to trigger the scheduler on 514 * the target CPU. 515 */ 516void resched_task(struct task_struct *p) 517{ 518 int cpu; 519 520 lockdep_assert_held(&task_rq(p)->lock); 521 522 if (test_tsk_need_resched(p)) 523 return; 524 525 set_tsk_need_resched(p); 526 527 cpu = task_cpu(p); 528 if (cpu == smp_processor_id()) { 529 set_preempt_need_resched(); 530 return; 531 } 532 533 /* NEED_RESCHED must be visible before we test polling */ 534 smp_mb(); 535 if (!tsk_is_polling(p)) 536 smp_send_reschedule(cpu); 537} 538 539void resched_cpu(int cpu) 540{ 541 struct rq *rq = cpu_rq(cpu); 542 unsigned long flags; 543 544 if (!raw_spin_trylock_irqsave(&rq->lock, flags)) 545 return; 546 resched_task(cpu_curr(cpu)); 547 raw_spin_unlock_irqrestore(&rq->lock, flags); 548} 549 550#ifdef CONFIG_SMP 551#ifdef CONFIG_NO_HZ_COMMON 552/* 553 * In the semi idle case, use the nearest busy cpu for migrating timers 554 * from an idle cpu. This is good for power-savings. 555 * 556 * We don't do similar optimization for completely idle system, as 557 * selecting an idle cpu will add more delays to the timers than intended 558 * (as that cpu's timer base may not be uptodate wrt jiffies etc). 559 */ 560int get_nohz_timer_target(void) 561{ 562 int cpu = smp_processor_id(); 563 int i; 564 struct sched_domain *sd; 565 566 rcu_read_lock(); 567 for_each_domain(cpu, sd) { 568 for_each_cpu(i, sched_domain_span(sd)) { 569 if (!idle_cpu(i)) { 570 cpu = i; 571 goto unlock; 572 } 573 } 574 } 575unlock: 576 rcu_read_unlock(); 577 return cpu; 578} 579/* 580 * When add_timer_on() enqueues a timer into the timer wheel of an 581 * idle CPU then this timer might expire before the next timer event 582 * which is scheduled to wake up that CPU. In case of a completely 583 * idle system the next event might even be infinite time into the 584 * future. wake_up_idle_cpu() ensures that the CPU is woken up and 585 * leaves the inner idle loop so the newly added timer is taken into 586 * account when the CPU goes back to idle and evaluates the timer 587 * wheel for the next timer event. 588 */ 589static void wake_up_idle_cpu(int cpu) 590{ 591 struct rq *rq = cpu_rq(cpu); 592 593 if (cpu == smp_processor_id()) 594 return; 595 596 /* 597 * This is safe, as this function is called with the timer 598 * wheel base lock of (cpu) held. When the CPU is on the way 599 * to idle and has not yet set rq->curr to idle then it will 600 * be serialized on the timer wheel base lock and take the new 601 * timer into account automatically. 602 */ 603 if (rq->curr != rq->idle) 604 return; 605 606 /* 607 * We can set TIF_RESCHED on the idle task of the other CPU 608 * lockless. The worst case is that the other CPU runs the 609 * idle task through an additional NOOP schedule() 610 */ 611 set_tsk_need_resched(rq->idle); 612 613 /* NEED_RESCHED must be visible before we test polling */ 614 smp_mb(); 615 if (!tsk_is_polling(rq->idle)) 616 smp_send_reschedule(cpu); 617} 618 619static bool wake_up_full_nohz_cpu(int cpu) 620{ 621 if (tick_nohz_full_cpu(cpu)) { 622 if (cpu != smp_processor_id() || 623 tick_nohz_tick_stopped()) 624 smp_send_reschedule(cpu); 625 return true; 626 } 627 628 return false; 629} 630 631void wake_up_nohz_cpu(int cpu) 632{ 633 if (!wake_up_full_nohz_cpu(cpu)) 634 wake_up_idle_cpu(cpu); 635} 636 637static inline bool got_nohz_idle_kick(void) 638{ 639 int cpu = smp_processor_id(); 640 641 if (!test_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu))) 642 return false; 643 644 if (idle_cpu(cpu) && !need_resched()) 645 return true; 646 647 /* 648 * We can't run Idle Load Balance on this CPU for this time so we 649 * cancel it and clear NOHZ_BALANCE_KICK 650 */ 651 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu)); 652 return false; 653} 654 655#else /* CONFIG_NO_HZ_COMMON */ 656 657static inline bool got_nohz_idle_kick(void) 658{ 659 return false; 660} 661 662#endif /* CONFIG_NO_HZ_COMMON */ 663 664#ifdef CONFIG_NO_HZ_FULL 665bool sched_can_stop_tick(void) 666{ 667 struct rq *rq; 668 669 rq = this_rq(); 670 671 /* Make sure rq->nr_running update is visible after the IPI */ 672 smp_rmb(); 673 674 /* More than one running task need preemption */ 675 if (rq->nr_running > 1) 676 return false; 677 678 return true; 679} 680#endif /* CONFIG_NO_HZ_FULL */ 681 682void sched_avg_update(struct rq *rq) 683{ 684 s64 period = sched_avg_period(); 685 686 while ((s64)(rq_clock(rq) - rq->age_stamp) > period) { 687 /* 688 * Inline assembly required to prevent the compiler 689 * optimising this loop into a divmod call. 690 * See __iter_div_u64_rem() for another example of this. 691 */ 692 asm("" : "+rm" (rq->age_stamp)); 693 rq->age_stamp += period; 694 rq->rt_avg /= 2; 695 } 696} 697 698#endif /* CONFIG_SMP */ 699 700#if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \ 701 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH))) 702/* 703 * Iterate task_group tree rooted at *from, calling @down when first entering a 704 * node and @up when leaving it for the final time. 705 * 706 * Caller must hold rcu_lock or sufficient equivalent. 707 */ 708int walk_tg_tree_from(struct task_group *from, 709 tg_visitor down, tg_visitor up, void *data) 710{ 711 struct task_group *parent, *child; 712 int ret; 713 714 parent = from; 715 716down: 717 ret = (*down)(parent, data); 718 if (ret) 719 goto out; 720 list_for_each_entry_rcu(child, &parent->children, siblings) { 721 parent = child; 722 goto down; 723 724up: 725 continue; 726 } 727 ret = (*up)(parent, data); 728 if (ret || parent == from) 729 goto out; 730 731 child = parent; 732 parent = parent->parent; 733 if (parent) 734 goto up; 735out: 736 return ret; 737} 738 739int tg_nop(struct task_group *tg, void *data) 740{ 741 return 0; 742} 743#endif 744 745static void set_load_weight(struct task_struct *p) 746{ 747 int prio = p->static_prio - MAX_RT_PRIO; 748 struct load_weight *load = &p->se.load; 749 750 /* 751 * SCHED_IDLE tasks get minimal weight: 752 */ 753 if (p->policy == SCHED_IDLE) { 754 load->weight = scale_load(WEIGHT_IDLEPRIO); 755 load->inv_weight = WMULT_IDLEPRIO; 756 return; 757 } 758 759 load->weight = scale_load(prio_to_weight[prio]); 760 load->inv_weight = prio_to_wmult[prio]; 761} 762 763static void enqueue_task(struct rq *rq, struct task_struct *p, int flags) 764{ 765 update_rq_clock(rq); 766 sched_info_queued(rq, p); 767 p->sched_class->enqueue_task(rq, p, flags); 768} 769 770static void dequeue_task(struct rq *rq, struct task_struct *p, int flags) 771{ 772 update_rq_clock(rq); 773 sched_info_dequeued(rq, p); 774 p->sched_class->dequeue_task(rq, p, flags); 775} 776 777void activate_task(struct rq *rq, struct task_struct *p, int flags) 778{ 779 if (task_contributes_to_load(p)) 780 rq->nr_uninterruptible--; 781 782 enqueue_task(rq, p, flags); 783} 784 785void deactivate_task(struct rq *rq, struct task_struct *p, int flags) 786{ 787 if (task_contributes_to_load(p)) 788 rq->nr_uninterruptible++; 789 790 dequeue_task(rq, p, flags); 791} 792 793static void update_rq_clock_task(struct rq *rq, s64 delta) 794{ 795/* 796 * In theory, the compile should just see 0 here, and optimize out the call 797 * to sched_rt_avg_update. But I don't trust it... 798 */ 799#if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING) 800 s64 steal = 0, irq_delta = 0; 801#endif 802#ifdef CONFIG_IRQ_TIME_ACCOUNTING 803 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time; 804 805 /* 806 * Since irq_time is only updated on {soft,}irq_exit, we might run into 807 * this case when a previous update_rq_clock() happened inside a 808 * {soft,}irq region. 809 * 810 * When this happens, we stop ->clock_task and only update the 811 * prev_irq_time stamp to account for the part that fit, so that a next 812 * update will consume the rest. This ensures ->clock_task is 813 * monotonic. 814 * 815 * It does however cause some slight miss-attribution of {soft,}irq 816 * time, a more accurate solution would be to update the irq_time using 817 * the current rq->clock timestamp, except that would require using 818 * atomic ops. 819 */ 820 if (irq_delta > delta) 821 irq_delta = delta; 822 823 rq->prev_irq_time += irq_delta; 824 delta -= irq_delta; 825#endif 826#ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING 827 if (static_key_false((¶virt_steal_rq_enabled))) { 828 u64 st; 829 830 steal = paravirt_steal_clock(cpu_of(rq)); 831 steal -= rq->prev_steal_time_rq; 832 833 if (unlikely(steal > delta)) 834 steal = delta; 835 836 st = steal_ticks(steal); 837 steal = st * TICK_NSEC; 838 839 rq->prev_steal_time_rq += steal; 840 841 delta -= steal; 842 } 843#endif 844 845 rq->clock_task += delta; 846 847#if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING) 848 if ((irq_delta + steal) && sched_feat(NONTASK_POWER)) 849 sched_rt_avg_update(rq, irq_delta + steal); 850#endif 851} 852 853void sched_set_stop_task(int cpu, struct task_struct *stop) 854{ 855 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 }; 856 struct task_struct *old_stop = cpu_rq(cpu)->stop; 857 858 if (stop) { 859 /* 860 * Make it appear like a SCHED_FIFO task, its something 861 * userspace knows about and won't get confused about. 862 * 863 * Also, it will make PI more or less work without too 864 * much confusion -- but then, stop work should not 865 * rely on PI working anyway. 866 */ 867 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m); 868 869 stop->sched_class = &stop_sched_class; 870 } 871 872 cpu_rq(cpu)->stop = stop; 873 874 if (old_stop) { 875 /* 876 * Reset it back to a normal scheduling class so that 877 * it can die in pieces. 878 */ 879 old_stop->sched_class = &rt_sched_class; 880 } 881} 882 883/* 884 * __normal_prio - return the priority that is based on the static prio 885 */ 886static inline int __normal_prio(struct task_struct *p) 887{ 888 return p->static_prio; 889} 890 891/* 892 * Calculate the expected normal priority: i.e. priority 893 * without taking RT-inheritance into account. Might be 894 * boosted by interactivity modifiers. Changes upon fork, 895 * setprio syscalls, and whenever the interactivity 896 * estimator recalculates. 897 */ 898static inline int normal_prio(struct task_struct *p) 899{ 900 int prio; 901 902 if (task_has_rt_policy(p)) 903 prio = MAX_RT_PRIO-1 - p->rt_priority; 904 else 905 prio = __normal_prio(p); 906 return prio; 907} 908 909/* 910 * Calculate the current priority, i.e. the priority 911 * taken into account by the scheduler. This value might 912 * be boosted by RT tasks, or might be boosted by 913 * interactivity modifiers. Will be RT if the task got 914 * RT-boosted. If not then it returns p->normal_prio. 915 */ 916static int effective_prio(struct task_struct *p) 917{ 918 p->normal_prio = normal_prio(p); 919 /* 920 * If we are RT tasks or we were boosted to RT priority, 921 * keep the priority unchanged. Otherwise, update priority 922 * to the normal priority: 923 */ 924 if (!rt_prio(p->prio)) 925 return p->normal_prio; 926 return p->prio; 927} 928 929/** 930 * task_curr - is this task currently executing on a CPU? 931 * @p: the task in question. 932 * 933 * Return: 1 if the task is currently executing. 0 otherwise. 934 */ 935inline int task_curr(const struct task_struct *p) 936{ 937 return cpu_curr(task_cpu(p)) == p; 938} 939 940static inline void check_class_changed(struct rq *rq, struct task_struct *p, 941 const struct sched_class *prev_class, 942 int oldprio) 943{ 944 if (prev_class != p->sched_class) { 945 if (prev_class->switched_from) 946 prev_class->switched_from(rq, p); 947 p->sched_class->switched_to(rq, p); 948 } else if (oldprio != p->prio) 949 p->sched_class->prio_changed(rq, p, oldprio); 950} 951 952void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags) 953{ 954 const struct sched_class *class; 955 956 if (p->sched_class == rq->curr->sched_class) { 957 rq->curr->sched_class->check_preempt_curr(rq, p, flags); 958 } else { 959 for_each_class(class) { 960 if (class == rq->curr->sched_class) 961 break; 962 if (class == p->sched_class) { 963 resched_task(rq->curr); 964 break; 965 } 966 } 967 } 968 969 /* 970 * A queue event has occurred, and we're going to schedule. In 971 * this case, we can save a useless back to back clock update. 972 */ 973 if (rq->curr->on_rq && test_tsk_need_resched(rq->curr)) 974 rq->skip_clock_update = 1; 975} 976 977#ifdef CONFIG_SMP 978void set_task_cpu(struct task_struct *p, unsigned int new_cpu) 979{ 980#ifdef CONFIG_SCHED_DEBUG 981 /* 982 * We should never call set_task_cpu() on a blocked task, 983 * ttwu() will sort out the placement. 984 */ 985 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING && 986 !(task_preempt_count(p) & PREEMPT_ACTIVE)); 987 988#ifdef CONFIG_LOCKDEP 989 /* 990 * The caller should hold either p->pi_lock or rq->lock, when changing 991 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks. 992 * 993 * sched_move_task() holds both and thus holding either pins the cgroup, 994 * see task_group(). 995 * 996 * Furthermore, all task_rq users should acquire both locks, see 997 * task_rq_lock(). 998 */ 999 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) || 1000 lockdep_is_held(&task_rq(p)->lock))); 1001#endif 1002#endif 1003 1004 trace_sched_migrate_task(p, new_cpu); 1005 1006 if (task_cpu(p) != new_cpu) { 1007 if (p->sched_class->migrate_task_rq) 1008 p->sched_class->migrate_task_rq(p, new_cpu); 1009 p->se.nr_migrations++; 1010 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, NULL, 0); 1011 } 1012 1013 __set_task_cpu(p, new_cpu); 1014} 1015 1016struct migration_arg { 1017 struct task_struct *task; 1018 int dest_cpu; 1019}; 1020 1021static int migration_cpu_stop(void *data); 1022 1023/* 1024 * wait_task_inactive - wait for a thread to unschedule. 1025 * 1026 * If @match_state is nonzero, it's the @p->state value just checked and 1027 * not expected to change. If it changes, i.e. @p might have woken up, 1028 * then return zero. When we succeed in waiting for @p to be off its CPU, 1029 * we return a positive number (its total switch count). If a second call 1030 * a short while later returns the same number, the caller can be sure that 1031 * @p has remained unscheduled the whole time. 1032 * 1033 * The caller must ensure that the task *will* unschedule sometime soon, 1034 * else this function might spin for a *long* time. This function can't 1035 * be called with interrupts off, or it may introduce deadlock with 1036 * smp_call_function() if an IPI is sent by the same process we are 1037 * waiting to become inactive. 1038 */ 1039unsigned long wait_task_inactive(struct task_struct *p, long match_state) 1040{ 1041 unsigned long flags; 1042 int running, on_rq; 1043 unsigned long ncsw; 1044 struct rq *rq; 1045 1046 for (;;) { 1047 /* 1048 * We do the initial early heuristics without holding 1049 * any task-queue locks at all. We'll only try to get 1050 * the runqueue lock when things look like they will 1051 * work out! 1052 */ 1053 rq = task_rq(p); 1054 1055 /* 1056 * If the task is actively running on another CPU 1057 * still, just relax and busy-wait without holding 1058 * any locks. 1059 * 1060 * NOTE! Since we don't hold any locks, it's not 1061 * even sure that "rq" stays as the right runqueue! 1062 * But we don't care, since "task_running()" will 1063 * return false if the runqueue has changed and p 1064 * is actually now running somewhere else! 1065 */ 1066 while (task_running(rq, p)) { 1067 if (match_state && unlikely(p->state != match_state)) 1068 return 0; 1069 cpu_relax(); 1070 } 1071 1072 /* 1073 * Ok, time to look more closely! We need the rq 1074 * lock now, to be *sure*. If we're wrong, we'll 1075 * just go back and repeat. 1076 */ 1077 rq = task_rq_lock(p, &flags); 1078 trace_sched_wait_task(p); 1079 running = task_running(rq, p); 1080 on_rq = p->on_rq; 1081 ncsw = 0; 1082 if (!match_state || p->state == match_state) 1083 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */ 1084 task_rq_unlock(rq, p, &flags); 1085 1086 /* 1087 * If it changed from the expected state, bail out now. 1088 */ 1089 if (unlikely(!ncsw)) 1090 break; 1091 1092 /* 1093 * Was it really running after all now that we 1094 * checked with the proper locks actually held? 1095 * 1096 * Oops. Go back and try again.. 1097 */ 1098 if (unlikely(running)) { 1099 cpu_relax(); 1100 continue; 1101 } 1102 1103 /* 1104 * It's not enough that it's not actively running, 1105 * it must be off the runqueue _entirely_, and not 1106 * preempted! 1107 * 1108 * So if it was still runnable (but just not actively 1109 * running right now), it's preempted, and we should 1110 * yield - it could be a while. 1111 */ 1112 if (unlikely(on_rq)) { 1113 ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ); 1114 1115 set_current_state(TASK_UNINTERRUPTIBLE); 1116 schedule_hrtimeout(&to, HRTIMER_MODE_REL); 1117 continue; 1118 } 1119 1120 /* 1121 * Ahh, all good. It wasn't running, and it wasn't 1122 * runnable, which means that it will never become 1123 * running in the future either. We're all done! 1124 */ 1125 break; 1126 } 1127 1128 return ncsw; 1129} 1130 1131/*** 1132 * kick_process - kick a running thread to enter/exit the kernel 1133 * @p: the to-be-kicked thread 1134 * 1135 * Cause a process which is running on another CPU to enter 1136 * kernel-mode, without any delay. (to get signals handled.) 1137 * 1138 * NOTE: this function doesn't have to take the runqueue lock, 1139 * because all it wants to ensure is that the remote task enters 1140 * the kernel. If the IPI races and the task has been migrated 1141 * to another CPU then no harm is done and the purpose has been 1142 * achieved as well. 1143 */ 1144void kick_process(struct task_struct *p) 1145{ 1146 int cpu; 1147 1148 preempt_disable(); 1149 cpu = task_cpu(p); 1150 if ((cpu != smp_processor_id()) && task_curr(p)) 1151 smp_send_reschedule(cpu); 1152 preempt_enable(); 1153} 1154EXPORT_SYMBOL_GPL(kick_process); 1155#endif /* CONFIG_SMP */ 1156 1157#ifdef CONFIG_SMP 1158/* 1159 * ->cpus_allowed is protected by both rq->lock and p->pi_lock 1160 */ 1161static int select_fallback_rq(int cpu, struct task_struct *p) 1162{ 1163 int nid = cpu_to_node(cpu); 1164 const struct cpumask *nodemask = NULL; 1165 enum { cpuset, possible, fail } state = cpuset; 1166 int dest_cpu; 1167 1168 /* 1169 * If the node that the cpu is on has been offlined, cpu_to_node() 1170 * will return -1. There is no cpu on the node, and we should 1171 * select the cpu on the other node. 1172 */ 1173 if (nid != -1) { 1174 nodemask = cpumask_of_node(nid); 1175 1176 /* Look for allowed, online CPU in same node. */ 1177 for_each_cpu(dest_cpu, nodemask) { 1178 if (!cpu_online(dest_cpu)) 1179 continue; 1180 if (!cpu_active(dest_cpu)) 1181 continue; 1182 if (cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p))) 1183 return dest_cpu; 1184 } 1185 } 1186 1187 for (;;) { 1188 /* Any allowed, online CPU? */ 1189 for_each_cpu(dest_cpu, tsk_cpus_allowed(p)) { 1190 if (!cpu_online(dest_cpu)) 1191 continue; 1192 if (!cpu_active(dest_cpu)) 1193 continue; 1194 goto out; 1195 } 1196 1197 switch (state) { 1198 case cpuset: 1199 /* No more Mr. Nice Guy. */ 1200 cpuset_cpus_allowed_fallback(p); 1201 state = possible; 1202 break; 1203 1204 case possible: 1205 do_set_cpus_allowed(p, cpu_possible_mask); 1206 state = fail; 1207 break; 1208 1209 case fail: 1210 BUG(); 1211 break; 1212 } 1213 } 1214 1215out: 1216 if (state != cpuset) { 1217 /* 1218 * Don't tell them about moving exiting tasks or 1219 * kernel threads (both mm NULL), since they never 1220 * leave kernel. 1221 */ 1222 if (p->mm && printk_ratelimit()) { 1223 printk_sched("process %d (%s) no longer affine to cpu%d\n", 1224 task_pid_nr(p), p->comm, cpu); 1225 } 1226 } 1227 1228 return dest_cpu; 1229} 1230 1231/* 1232 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable. 1233 */ 1234static inline 1235int select_task_rq(struct task_struct *p, int sd_flags, int wake_flags) 1236{ 1237 int cpu = p->sched_class->select_task_rq(p, sd_flags, wake_flags); 1238 1239 /* 1240 * In order not to call set_task_cpu() on a blocking task we need 1241 * to rely on ttwu() to place the task on a valid ->cpus_allowed 1242 * cpu. 1243 * 1244 * Since this is common to all placement strategies, this lives here. 1245 * 1246 * [ this allows ->select_task() to simply return task_cpu(p) and 1247 * not worry about this generic constraint ] 1248 */ 1249 if (unlikely(!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)) || 1250 !cpu_online(cpu))) 1251 cpu = select_fallback_rq(task_cpu(p), p); 1252 1253 return cpu; 1254} 1255 1256static void update_avg(u64 *avg, u64 sample) 1257{ 1258 s64 diff = sample - *avg; 1259 *avg += diff >> 3; 1260} 1261#endif 1262 1263static void 1264ttwu_stat(struct task_struct *p, int cpu, int wake_flags) 1265{ 1266#ifdef CONFIG_SCHEDSTATS 1267 struct rq *rq = this_rq(); 1268 1269#ifdef CONFIG_SMP 1270 int this_cpu = smp_processor_id(); 1271 1272 if (cpu == this_cpu) { 1273 schedstat_inc(rq, ttwu_local); 1274 schedstat_inc(p, se.statistics.nr_wakeups_local); 1275 } else { 1276 struct sched_domain *sd; 1277 1278 schedstat_inc(p, se.statistics.nr_wakeups_remote); 1279 rcu_read_lock(); 1280 for_each_domain(this_cpu, sd) { 1281 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) { 1282 schedstat_inc(sd, ttwu_wake_remote); 1283 break; 1284 } 1285 } 1286 rcu_read_unlock(); 1287 } 1288 1289 if (wake_flags & WF_MIGRATED) 1290 schedstat_inc(p, se.statistics.nr_wakeups_migrate); 1291 1292#endif /* CONFIG_SMP */ 1293 1294 schedstat_inc(rq, ttwu_count); 1295 schedstat_inc(p, se.statistics.nr_wakeups); 1296 1297 if (wake_flags & WF_SYNC) 1298 schedstat_inc(p, se.statistics.nr_wakeups_sync); 1299 1300#endif /* CONFIG_SCHEDSTATS */ 1301} 1302 1303static void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags) 1304{ 1305 activate_task(rq, p, en_flags); 1306 p->on_rq = 1; 1307 1308 /* if a worker is waking up, notify workqueue */ 1309 if (p->flags & PF_WQ_WORKER) 1310 wq_worker_waking_up(p, cpu_of(rq)); 1311} 1312 1313/* 1314 * Mark the task runnable and perform wakeup-preemption. 1315 */ 1316static void 1317ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags) 1318{ 1319 check_preempt_curr(rq, p, wake_flags); 1320 trace_sched_wakeup(p, true); 1321 1322 p->state = TASK_RUNNING; 1323#ifdef CONFIG_SMP 1324 if (p->sched_class->task_woken) 1325 p->sched_class->task_woken(rq, p); 1326 1327 if (rq->idle_stamp) { 1328 u64 delta = rq_clock(rq) - rq->idle_stamp; 1329 u64 max = 2*rq->max_idle_balance_cost; 1330 1331 update_avg(&rq->avg_idle, delta); 1332 1333 if (rq->avg_idle > max) 1334 rq->avg_idle = max; 1335 1336 rq->idle_stamp = 0; 1337 } 1338#endif 1339} 1340 1341static void 1342ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags) 1343{ 1344#ifdef CONFIG_SMP 1345 if (p->sched_contributes_to_load) 1346 rq->nr_uninterruptible--; 1347#endif 1348 1349 ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_WAKING); 1350 ttwu_do_wakeup(rq, p, wake_flags); 1351} 1352 1353/* 1354 * Called in case the task @p isn't fully descheduled from its runqueue, 1355 * in this case we must do a remote wakeup. Its a 'light' wakeup though, 1356 * since all we need to do is flip p->state to TASK_RUNNING, since 1357 * the task is still ->on_rq. 1358 */ 1359static int ttwu_remote(struct task_struct *p, int wake_flags) 1360{ 1361 struct rq *rq; 1362 int ret = 0; 1363 1364 rq = __task_rq_lock(p); 1365 if (p->on_rq) { 1366 /* check_preempt_curr() may use rq clock */ 1367 update_rq_clock(rq); 1368 ttwu_do_wakeup(rq, p, wake_flags); 1369 ret = 1; 1370 } 1371 __task_rq_unlock(rq); 1372 1373 return ret; 1374} 1375 1376#ifdef CONFIG_SMP 1377static void sched_ttwu_pending(void) 1378{ 1379 struct rq *rq = this_rq(); 1380 struct llist_node *llist = llist_del_all(&rq->wake_list); 1381 struct task_struct *p; 1382 1383 raw_spin_lock(&rq->lock); 1384 1385 while (llist) { 1386 p = llist_entry(llist, struct task_struct, wake_entry); 1387 llist = llist_next(llist); 1388 ttwu_do_activate(rq, p, 0); 1389 } 1390 1391 raw_spin_unlock(&rq->lock); 1392} 1393 1394void scheduler_ipi(void) 1395{ 1396 /* 1397 * Fold TIF_NEED_RESCHED into the preempt_count; anybody setting 1398 * TIF_NEED_RESCHED remotely (for the first time) will also send 1399 * this IPI. 1400 */ 1401 if (tif_need_resched()) 1402 set_preempt_need_resched(); 1403 1404 if (llist_empty(&this_rq()->wake_list) 1405 && !tick_nohz_full_cpu(smp_processor_id()) 1406 && !got_nohz_idle_kick()) 1407 return; 1408 1409 /* 1410 * Not all reschedule IPI handlers call irq_enter/irq_exit, since 1411 * traditionally all their work was done from the interrupt return 1412 * path. Now that we actually do some work, we need to make sure 1413 * we do call them. 1414 * 1415 * Some archs already do call them, luckily irq_enter/exit nest 1416 * properly. 1417 * 1418 * Arguably we should visit all archs and update all handlers, 1419 * however a fair share of IPIs are still resched only so this would 1420 * somewhat pessimize the simple resched case. 1421 */ 1422 irq_enter(); 1423 tick_nohz_full_check(); 1424 sched_ttwu_pending(); 1425 1426 /* 1427 * Check if someone kicked us for doing the nohz idle load balance. 1428 */ 1429 if (unlikely(got_nohz_idle_kick())) { 1430 this_rq()->idle_balance = 1; 1431 raise_softirq_irqoff(SCHED_SOFTIRQ); 1432 } 1433 irq_exit(); 1434} 1435 1436static void ttwu_queue_remote(struct task_struct *p, int cpu) 1437{ 1438 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list)) 1439 smp_send_reschedule(cpu); 1440} 1441 1442bool cpus_share_cache(int this_cpu, int that_cpu) 1443{ 1444 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu); 1445} 1446#endif /* CONFIG_SMP */ 1447 1448static void ttwu_queue(struct task_struct *p, int cpu) 1449{ 1450 struct rq *rq = cpu_rq(cpu); 1451 1452#if defined(CONFIG_SMP) 1453 if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) { 1454 sched_clock_cpu(cpu); /* sync clocks x-cpu */ 1455 ttwu_queue_remote(p, cpu); 1456 return; 1457 } 1458#endif 1459 1460 raw_spin_lock(&rq->lock); 1461 ttwu_do_activate(rq, p, 0); 1462 raw_spin_unlock(&rq->lock); 1463} 1464 1465/** 1466 * try_to_wake_up - wake up a thread 1467 * @p: the thread to be awakened 1468 * @state: the mask of task states that can be woken 1469 * @wake_flags: wake modifier flags (WF_*) 1470 * 1471 * Put it on the run-queue if it's not already there. The "current" 1472 * thread is always on the run-queue (except when the actual 1473 * re-schedule is in progress), and as such you're allowed to do 1474 * the simpler "current->state = TASK_RUNNING" to mark yourself 1475 * runnable without the overhead of this. 1476 * 1477 * Return: %true if @p was woken up, %false if it was already running. 1478 * or @state didn't match @p's state. 1479 */ 1480static int 1481try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags) 1482{ 1483 unsigned long flags; 1484 int cpu, success = 0; 1485 1486 /* 1487 * If we are going to wake up a thread waiting for CONDITION we 1488 * need to ensure that CONDITION=1 done by the caller can not be 1489 * reordered with p->state check below. This pairs with mb() in 1490 * set_current_state() the waiting thread does. 1491 */ 1492 smp_mb__before_spinlock(); 1493 raw_spin_lock_irqsave(&p->pi_lock, flags); 1494 if (!(p->state & state)) 1495 goto out; 1496 1497 success = 1; /* we're going to change ->state */ 1498 cpu = task_cpu(p); 1499 1500 if (p->on_rq && ttwu_remote(p, wake_flags)) 1501 goto stat; 1502 1503#ifdef CONFIG_SMP 1504 /* 1505 * If the owning (remote) cpu is still in the middle of schedule() with 1506 * this task as prev, wait until its done referencing the task. 1507 */ 1508 while (p->on_cpu) 1509 cpu_relax(); 1510 /* 1511 * Pairs with the smp_wmb() in finish_lock_switch(). 1512 */ 1513 smp_rmb(); 1514 1515 p->sched_contributes_to_load = !!task_contributes_to_load(p); 1516 p->state = TASK_WAKING; 1517 1518 if (p->sched_class->task_waking) 1519 p->sched_class->task_waking(p); 1520 1521 cpu = select_task_rq(p, SD_BALANCE_WAKE, wake_flags); 1522 if (task_cpu(p) != cpu) { 1523 wake_flags |= WF_MIGRATED; 1524 set_task_cpu(p, cpu); 1525 } 1526#endif /* CONFIG_SMP */ 1527 1528 ttwu_queue(p, cpu); 1529stat: 1530 ttwu_stat(p, cpu, wake_flags); 1531out: 1532 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 1533 1534 return success; 1535} 1536 1537/** 1538 * try_to_wake_up_local - try to wake up a local task with rq lock held 1539 * @p: the thread to be awakened 1540 * 1541 * Put @p on the run-queue if it's not already there. The caller must 1542 * ensure that this_rq() is locked, @p is bound to this_rq() and not 1543 * the current task. 1544 */ 1545static void try_to_wake_up_local(struct task_struct *p) 1546{ 1547 struct rq *rq = task_rq(p); 1548 1549 if (WARN_ON_ONCE(rq != this_rq()) || 1550 WARN_ON_ONCE(p == current)) 1551 return; 1552 1553 lockdep_assert_held(&rq->lock); 1554 1555 if (!raw_spin_trylock(&p->pi_lock)) { 1556 raw_spin_unlock(&rq->lock); 1557 raw_spin_lock(&p->pi_lock); 1558 raw_spin_lock(&rq->lock); 1559 } 1560 1561 if (!(p->state & TASK_NORMAL)) 1562 goto out; 1563 1564 if (!p->on_rq) 1565 ttwu_activate(rq, p, ENQUEUE_WAKEUP); 1566 1567 ttwu_do_wakeup(rq, p, 0); 1568 ttwu_stat(p, smp_processor_id(), 0); 1569out: 1570 raw_spin_unlock(&p->pi_lock); 1571} 1572 1573/** 1574 * wake_up_process - Wake up a specific process 1575 * @p: The process to be woken up. 1576 * 1577 * Attempt to wake up the nominated process and move it to the set of runnable 1578 * processes. 1579 * 1580 * Return: 1 if the process was woken up, 0 if it was already running. 1581 * 1582 * It may be assumed that this function implies a write memory barrier before 1583 * changing the task state if and only if any tasks are woken up. 1584 */ 1585int wake_up_process(struct task_struct *p) 1586{ 1587 WARN_ON(task_is_stopped_or_traced(p)); 1588 return try_to_wake_up(p, TASK_NORMAL, 0); 1589} 1590EXPORT_SYMBOL(wake_up_process); 1591 1592int wake_up_state(struct task_struct *p, unsigned int state) 1593{ 1594 return try_to_wake_up(p, state, 0); 1595} 1596 1597/* 1598 * Perform scheduler related setup for a newly forked process p. 1599 * p is forked by current. 1600 * 1601 * __sched_fork() is basic setup used by init_idle() too: 1602 */ 1603static void __sched_fork(struct task_struct *p) 1604{ 1605 p->on_rq = 0; 1606 1607 p->se.on_rq = 0; 1608 p->se.exec_start = 0; 1609 p->se.sum_exec_runtime = 0; 1610 p->se.prev_sum_exec_runtime = 0; 1611 p->se.nr_migrations = 0; 1612 p->se.vruntime = 0; 1613 INIT_LIST_HEAD(&p->se.group_node); 1614 1615#ifdef CONFIG_SCHEDSTATS 1616 memset(&p->se.statistics, 0, sizeof(p->se.statistics)); 1617#endif 1618 1619 INIT_LIST_HEAD(&p->rt.run_list); 1620 1621#ifdef CONFIG_PREEMPT_NOTIFIERS 1622 INIT_HLIST_HEAD(&p->preempt_notifiers); 1623#endif 1624 1625#ifdef CONFIG_NUMA_BALANCING 1626 if (p->mm && atomic_read(&p->mm->mm_users) == 1) { 1627 p->mm->numa_next_scan = jiffies; 1628 p->mm->numa_next_reset = jiffies; 1629 p->mm->numa_scan_seq = 0; 1630 } 1631 1632 p->node_stamp = 0ULL; 1633 p->numa_scan_seq = p->mm ? p->mm->numa_scan_seq : 0; 1634 p->numa_migrate_seq = p->mm ? p->mm->numa_scan_seq - 1 : 0; 1635 p->numa_scan_period = sysctl_numa_balancing_scan_delay; 1636 p->numa_work.next = &p->numa_work; 1637#endif /* CONFIG_NUMA_BALANCING */ 1638} 1639 1640#ifdef CONFIG_NUMA_BALANCING 1641#ifdef CONFIG_SCHED_DEBUG 1642void set_numabalancing_state(bool enabled) 1643{ 1644 if (enabled) 1645 sched_feat_set("NUMA"); 1646 else 1647 sched_feat_set("NO_NUMA"); 1648} 1649#else 1650__read_mostly bool numabalancing_enabled; 1651 1652void set_numabalancing_state(bool enabled) 1653{ 1654 numabalancing_enabled = enabled; 1655} 1656#endif /* CONFIG_SCHED_DEBUG */ 1657#endif /* CONFIG_NUMA_BALANCING */ 1658 1659/* 1660 * fork()/clone()-time setup: 1661 */ 1662void sched_fork(struct task_struct *p) 1663{ 1664 unsigned long flags; 1665 int cpu = get_cpu(); 1666 1667 __sched_fork(p); 1668 /* 1669 * We mark the process as running here. This guarantees that 1670 * nobody will actually run it, and a signal or other external 1671 * event cannot wake it up and insert it on the runqueue either. 1672 */ 1673 p->state = TASK_RUNNING; 1674 1675 /* 1676 * Make sure we do not leak PI boosting priority to the child. 1677 */ 1678 p->prio = current->normal_prio; 1679 1680 /* 1681 * Revert to default priority/policy on fork if requested. 1682 */ 1683 if (unlikely(p->sched_reset_on_fork)) { 1684 if (task_has_rt_policy(p)) { 1685 p->policy = SCHED_NORMAL; 1686 p->static_prio = NICE_TO_PRIO(0); 1687 p->rt_priority = 0; 1688 } else if (PRIO_TO_NICE(p->static_prio) < 0) 1689 p->static_prio = NICE_TO_PRIO(0); 1690 1691 p->prio = p->normal_prio = __normal_prio(p); 1692 set_load_weight(p); 1693 1694 /* 1695 * We don't need the reset flag anymore after the fork. It has 1696 * fulfilled its duty: 1697 */ 1698 p->sched_reset_on_fork = 0; 1699 } 1700 1701 if (!rt_prio(p->prio)) 1702 p->sched_class = &fair_sched_class; 1703 1704 if (p->sched_class->task_fork) 1705 p->sched_class->task_fork(p); 1706 1707 /* 1708 * The child is not yet in the pid-hash so no cgroup attach races, 1709 * and the cgroup is pinned to this child due to cgroup_fork() 1710 * is ran before sched_fork(). 1711 * 1712 * Silence PROVE_RCU. 1713 */ 1714 raw_spin_lock_irqsave(&p->pi_lock, flags); 1715 set_task_cpu(p, cpu); 1716 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 1717 1718#if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT) 1719 if (likely(sched_info_on())) 1720 memset(&p->sched_info, 0, sizeof(p->sched_info)); 1721#endif 1722#if defined(CONFIG_SMP) 1723 p->on_cpu = 0; 1724#endif 1725#ifdef CONFIG_PREEMPT_COUNT 1726 init_task_preempt_count(p); 1727#endif 1728#ifdef CONFIG_SMP 1729 plist_node_init(&p->pushable_tasks, MAX_PRIO); 1730#endif 1731 1732 put_cpu(); 1733} 1734 1735/* 1736 * wake_up_new_task - wake up a newly created task for the first time. 1737 * 1738 * This function will do some initial scheduler statistics housekeeping 1739 * that must be done for every newly created context, then puts the task 1740 * on the runqueue and wakes it. 1741 */ 1742void wake_up_new_task(struct task_struct *p) 1743{ 1744 unsigned long flags; 1745 struct rq *rq; 1746 1747 raw_spin_lock_irqsave(&p->pi_lock, flags); 1748#ifdef CONFIG_SMP 1749 /* 1750 * Fork balancing, do it here and not earlier because: 1751 * - cpus_allowed can change in the fork path 1752 * - any previously selected cpu might disappear through hotplug 1753 */ 1754 set_task_cpu(p, select_task_rq(p, SD_BALANCE_FORK, 0)); 1755#endif 1756 1757 /* Initialize new task's runnable average */ 1758 init_task_runnable_average(p); 1759 rq = __task_rq_lock(p); 1760 activate_task(rq, p, 0); 1761 p->on_rq = 1; 1762 trace_sched_wakeup_new(p, true); 1763 check_preempt_curr(rq, p, WF_FORK); 1764#ifdef CONFIG_SMP 1765 if (p->sched_class->task_woken) 1766 p->sched_class->task_woken(rq, p); 1767#endif 1768 task_rq_unlock(rq, p, &flags); 1769} 1770 1771#ifdef CONFIG_PREEMPT_NOTIFIERS 1772 1773/** 1774 * preempt_notifier_register - tell me when current is being preempted & rescheduled 1775 * @notifier: notifier struct to register 1776 */ 1777void preempt_notifier_register(struct preempt_notifier *notifier) 1778{ 1779 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers); 1780} 1781EXPORT_SYMBOL_GPL(preempt_notifier_register); 1782 1783/** 1784 * preempt_notifier_unregister - no longer interested in preemption notifications 1785 * @notifier: notifier struct to unregister 1786 * 1787 * This is safe to call from within a preemption notifier. 1788 */ 1789void preempt_notifier_unregister(struct preempt_notifier *notifier) 1790{ 1791 hlist_del(¬ifier->link); 1792} 1793EXPORT_SYMBOL_GPL(preempt_notifier_unregister); 1794 1795static void fire_sched_in_preempt_notifiers(struct task_struct *curr) 1796{ 1797 struct preempt_notifier *notifier; 1798 1799 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link) 1800 notifier->ops->sched_in(notifier, raw_smp_processor_id()); 1801} 1802 1803static void 1804fire_sched_out_preempt_notifiers(struct task_struct *curr, 1805 struct task_struct *next) 1806{ 1807 struct preempt_notifier *notifier; 1808 1809 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link) 1810 notifier->ops->sched_out(notifier, next); 1811} 1812 1813#else /* !CONFIG_PREEMPT_NOTIFIERS */ 1814 1815static void fire_sched_in_preempt_notifiers(struct task_struct *curr) 1816{ 1817} 1818 1819static void 1820fire_sched_out_preempt_notifiers(struct task_struct *curr, 1821 struct task_struct *next) 1822{ 1823} 1824 1825#endif /* CONFIG_PREEMPT_NOTIFIERS */ 1826 1827/** 1828 * prepare_task_switch - prepare to switch tasks 1829 * @rq: the runqueue preparing to switch 1830 * @prev: the current task that is being switched out 1831 * @next: the task we are going to switch to. 1832 * 1833 * This is called with the rq lock held and interrupts off. It must 1834 * be paired with a subsequent finish_task_switch after the context 1835 * switch. 1836 * 1837 * prepare_task_switch sets up locking and calls architecture specific 1838 * hooks. 1839 */ 1840static inline void 1841prepare_task_switch(struct rq *rq, struct task_struct *prev, 1842 struct task_struct *next) 1843{ 1844 trace_sched_switch(prev, next); 1845 sched_info_switch(rq, prev, next); 1846 perf_event_task_sched_out(prev, next); 1847 fire_sched_out_preempt_notifiers(prev, next); 1848 prepare_lock_switch(rq, next); 1849 prepare_arch_switch(next); 1850} 1851 1852/** 1853 * finish_task_switch - clean up after a task-switch 1854 * @rq: runqueue associated with task-switch 1855 * @prev: the thread we just switched away from. 1856 * 1857 * finish_task_switch must be called after the context switch, paired 1858 * with a prepare_task_switch call before the context switch. 1859 * finish_task_switch will reconcile locking set up by prepare_task_switch, 1860 * and do any other architecture-specific cleanup actions. 1861 * 1862 * Note that we may have delayed dropping an mm in context_switch(). If 1863 * so, we finish that here outside of the runqueue lock. (Doing it 1864 * with the lock held can cause deadlocks; see schedule() for 1865 * details.) 1866 */ 1867static void finish_task_switch(struct rq *rq, struct task_struct *prev) 1868 __releases(rq->lock) 1869{ 1870 struct mm_struct *mm = rq->prev_mm; 1871 long prev_state; 1872 1873 rq->prev_mm = NULL; 1874 1875 /* 1876 * A task struct has one reference for the use as "current". 1877 * If a task dies, then it sets TASK_DEAD in tsk->state and calls 1878 * schedule one last time. The schedule call will never return, and 1879 * the scheduled task must drop that reference. 1880 * The test for TASK_DEAD must occur while the runqueue locks are 1881 * still held, otherwise prev could be scheduled on another cpu, die 1882 * there before we look at prev->state, and then the reference would 1883 * be dropped twice. 1884 * Manfred Spraul <manfred@colorfullife.com> 1885 */ 1886 prev_state = prev->state; 1887 vtime_task_switch(prev); 1888 finish_arch_switch(prev); 1889 perf_event_task_sched_in(prev, current); 1890 finish_lock_switch(rq, prev); 1891 finish_arch_post_lock_switch(); 1892 1893 fire_sched_in_preempt_notifiers(current); 1894 if (mm) 1895 mmdrop(mm); 1896 if (unlikely(prev_state == TASK_DEAD)) { 1897 /* 1898 * Remove function-return probe instances associated with this 1899 * task and put them back on the free list. 1900 */ 1901 kprobe_flush_task(prev); 1902 put_task_struct(prev); 1903 } 1904 1905 tick_nohz_task_switch(current); 1906} 1907 1908#ifdef CONFIG_SMP 1909 1910/* assumes rq->lock is held */ 1911static inline void pre_schedule(struct rq *rq, struct task_struct *prev) 1912{ 1913 if (prev->sched_class->pre_schedule) 1914 prev->sched_class->pre_schedule(rq, prev); 1915} 1916 1917/* rq->lock is NOT held, but preemption is disabled */ 1918static inline void post_schedule(struct rq *rq) 1919{ 1920 if (rq->post_schedule) { 1921 unsigned long flags; 1922 1923 raw_spin_lock_irqsave(&rq->lock, flags); 1924 if (rq->curr->sched_class->post_schedule) 1925 rq->curr->sched_class->post_schedule(rq); 1926 raw_spin_unlock_irqrestore(&rq->lock, flags); 1927 1928 rq->post_schedule = 0; 1929 } 1930} 1931 1932#else 1933 1934static inline void pre_schedule(struct rq *rq, struct task_struct *p) 1935{ 1936} 1937 1938static inline void post_schedule(struct rq *rq) 1939{ 1940} 1941 1942#endif 1943 1944/** 1945 * schedule_tail - first thing a freshly forked thread must call. 1946 * @prev: the thread we just switched away from. 1947 */ 1948asmlinkage void schedule_tail(struct task_struct *prev) 1949 __releases(rq->lock) 1950{ 1951 struct rq *rq = this_rq(); 1952 1953 finish_task_switch(rq, prev); 1954 1955 /* 1956 * FIXME: do we need to worry about rq being invalidated by the 1957 * task_switch? 1958 */ 1959 post_schedule(rq); 1960 1961#ifdef __ARCH_WANT_UNLOCKED_CTXSW 1962 /* In this case, finish_task_switch does not reenable preemption */ 1963 preempt_enable(); 1964#endif 1965 if (current->set_child_tid) 1966 put_user(task_pid_vnr(current), current->set_child_tid); 1967} 1968 1969/* 1970 * context_switch - switch to the new MM and the new 1971 * thread's register state. 1972 */ 1973static inline void 1974context_switch(struct rq *rq, struct task_struct *prev, 1975 struct task_struct *next) 1976{ 1977 struct mm_struct *mm, *oldmm; 1978 1979 prepare_task_switch(rq, prev, next); 1980 1981 mm = next->mm; 1982 oldmm = prev->active_mm; 1983 /* 1984 * For paravirt, this is coupled with an exit in switch_to to 1985 * combine the page table reload and the switch backend into 1986 * one hypercall. 1987 */ 1988 arch_start_context_switch(prev); 1989 1990 if (!mm) { 1991 next->active_mm = oldmm; 1992 atomic_inc(&oldmm->mm_count); 1993 enter_lazy_tlb(oldmm, next); 1994 } else 1995 switch_mm(oldmm, mm, next); 1996 1997 if (!prev->mm) { 1998 prev->active_mm = NULL; 1999 rq->prev_mm = oldmm; 2000 } 2001 /* 2002 * Since the runqueue lock will be released by the next 2003 * task (which is an invalid locking op but in the case 2004 * of the scheduler it's an obvious special-case), so we 2005 * do an early lockdep release here: 2006 */ 2007#ifndef __ARCH_WANT_UNLOCKED_CTXSW 2008 spin_release(&rq->lock.dep_map, 1, _THIS_IP_); 2009#endif 2010 2011 context_tracking_task_switch(prev, next); 2012 /* Here we just switch the register state and the stack. */ 2013 switch_to(prev, next, prev); 2014 2015 barrier(); 2016 /* 2017 * this_rq must be evaluated again because prev may have moved 2018 * CPUs since it called schedule(), thus the 'rq' on its stack 2019 * frame will be invalid. 2020 */ 2021 finish_task_switch(this_rq(), prev); 2022} 2023 2024/* 2025 * nr_running and nr_context_switches: 2026 * 2027 * externally visible scheduler statistics: current number of runnable 2028 * threads, total number of context switches performed since bootup. 2029 */ 2030unsigned long nr_running(void) 2031{ 2032 unsigned long i, sum = 0; 2033 2034 for_each_online_cpu(i) 2035 sum += cpu_rq(i)->nr_running; 2036 2037 return sum; 2038} 2039 2040unsigned long long nr_context_switches(void) 2041{ 2042 int i; 2043 unsigned long long sum = 0; 2044 2045 for_each_possible_cpu(i) 2046 sum += cpu_rq(i)->nr_switches; 2047 2048 return sum; 2049} 2050 2051unsigned long nr_iowait(void) 2052{ 2053 unsigned long i, sum = 0; 2054 2055 for_each_possible_cpu(i) 2056 sum += atomic_read(&cpu_rq(i)->nr_iowait); 2057 2058 return sum; 2059} 2060 2061unsigned long nr_iowait_cpu(int cpu) 2062{ 2063 struct rq *this = cpu_rq(cpu); 2064 return atomic_read(&this->nr_iowait); 2065} 2066 2067#ifdef CONFIG_SMP 2068 2069/* 2070 * sched_exec - execve() is a valuable balancing opportunity, because at 2071 * this point the task has the smallest effective memory and cache footprint. 2072 */ 2073void sched_exec(void) 2074{ 2075 struct task_struct *p = current; 2076 unsigned long flags; 2077 int dest_cpu; 2078 2079 raw_spin_lock_irqsave(&p->pi_lock, flags); 2080 dest_cpu = p->sched_class->select_task_rq(p, SD_BALANCE_EXEC, 0); 2081 if (dest_cpu == smp_processor_id()) 2082 goto unlock; 2083 2084 if (likely(cpu_active(dest_cpu))) { 2085 struct migration_arg arg = { p, dest_cpu }; 2086 2087 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 2088 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg); 2089 return; 2090 } 2091unlock: 2092 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 2093} 2094 2095#endif 2096 2097DEFINE_PER_CPU(struct kernel_stat, kstat); 2098DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat); 2099 2100EXPORT_PER_CPU_SYMBOL(kstat); 2101EXPORT_PER_CPU_SYMBOL(kernel_cpustat); 2102 2103/* 2104 * Return any ns on the sched_clock that have not yet been accounted in 2105 * @p in case that task is currently running. 2106 * 2107 * Called with task_rq_lock() held on @rq. 2108 */ 2109static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq) 2110{ 2111 u64 ns = 0; 2112 2113 if (task_current(rq, p)) { 2114 update_rq_clock(rq); 2115 ns = rq_clock_task(rq) - p->se.exec_start; 2116 if ((s64)ns < 0) 2117 ns = 0; 2118 } 2119 2120 return ns; 2121} 2122 2123unsigned long long task_delta_exec(struct task_struct *p) 2124{ 2125 unsigned long flags; 2126 struct rq *rq; 2127 u64 ns = 0; 2128 2129 rq = task_rq_lock(p, &flags); 2130 ns = do_task_delta_exec(p, rq); 2131 task_rq_unlock(rq, p, &flags); 2132 2133 return ns; 2134} 2135 2136/* 2137 * Return accounted runtime for the task. 2138 * In case the task is currently running, return the runtime plus current's 2139 * pending runtime that have not been accounted yet. 2140 */ 2141unsigned long long task_sched_runtime(struct task_struct *p) 2142{ 2143 unsigned long flags; 2144 struct rq *rq; 2145 u64 ns = 0; 2146 2147 rq = task_rq_lock(p, &flags); 2148 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq); 2149 task_rq_unlock(rq, p, &flags); 2150 2151 return ns; 2152} 2153 2154/* 2155 * This function gets called by the timer code, with HZ frequency. 2156 * We call it with interrupts disabled. 2157 */ 2158void scheduler_tick(void) 2159{ 2160 int cpu = smp_processor_id(); 2161 struct rq *rq = cpu_rq(cpu); 2162 struct task_struct *curr = rq->curr; 2163 2164 sched_clock_tick(); 2165 2166 raw_spin_lock(&rq->lock); 2167 update_rq_clock(rq); 2168 curr->sched_class->task_tick(rq, curr, 0); 2169 update_cpu_load_active(rq); 2170 raw_spin_unlock(&rq->lock); 2171 2172 perf_event_task_tick(); 2173 2174#ifdef CONFIG_SMP 2175 rq->idle_balance = idle_cpu(cpu); 2176 trigger_load_balance(rq, cpu); 2177#endif 2178 rq_last_tick_reset(rq); 2179} 2180 2181#ifdef CONFIG_NO_HZ_FULL 2182/** 2183 * scheduler_tick_max_deferment 2184 * 2185 * Keep at least one tick per second when a single 2186 * active task is running because the scheduler doesn't 2187 * yet completely support full dynticks environment. 2188 * 2189 * This makes sure that uptime, CFS vruntime, load 2190 * balancing, etc... continue to move forward, even 2191 * with a very low granularity. 2192 * 2193 * Return: Maximum deferment in nanoseconds. 2194 */ 2195u64 scheduler_tick_max_deferment(void) 2196{ 2197 struct rq *rq = this_rq(); 2198 unsigned long next, now = ACCESS_ONCE(jiffies); 2199 2200 next = rq->last_sched_tick + HZ; 2201 2202 if (time_before_eq(next, now)) 2203 return 0; 2204 2205 return jiffies_to_usecs(next - now) * NSEC_PER_USEC; 2206} 2207#endif 2208 2209notrace unsigned long get_parent_ip(unsigned long addr) 2210{ 2211 if (in_lock_functions(addr)) { 2212 addr = CALLER_ADDR2; 2213 if (in_lock_functions(addr)) 2214 addr = CALLER_ADDR3; 2215 } 2216 return addr; 2217} 2218 2219#if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \ 2220 defined(CONFIG_PREEMPT_TRACER)) 2221 2222void __kprobes preempt_count_add(int val) 2223{ 2224#ifdef CONFIG_DEBUG_PREEMPT 2225 /* 2226 * Underflow? 2227 */ 2228 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0))) 2229 return; 2230#endif 2231 __preempt_count_add(val); 2232#ifdef CONFIG_DEBUG_PREEMPT 2233 /* 2234 * Spinlock count overflowing soon? 2235 */ 2236 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >= 2237 PREEMPT_MASK - 10); 2238#endif 2239 if (preempt_count() == val) 2240 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1)); 2241} 2242EXPORT_SYMBOL(preempt_count_add); 2243 2244void __kprobes preempt_count_sub(int val) 2245{ 2246#ifdef CONFIG_DEBUG_PREEMPT 2247 /* 2248 * Underflow? 2249 */ 2250 if (DEBUG_LOCKS_WARN_ON(val > preempt_count())) 2251 return; 2252 /* 2253 * Is the spinlock portion underflowing? 2254 */ 2255 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) && 2256 !(preempt_count() & PREEMPT_MASK))) 2257 return; 2258#endif 2259 2260 if (preempt_count() == val) 2261 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1)); 2262 __preempt_count_sub(val); 2263} 2264EXPORT_SYMBOL(preempt_count_sub); 2265 2266#endif 2267 2268/* 2269 * Print scheduling while atomic bug: 2270 */ 2271static noinline void __schedule_bug(struct task_struct *prev) 2272{ 2273 if (oops_in_progress) 2274 return; 2275 2276 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n", 2277 prev->comm, prev->pid, preempt_count()); 2278 2279 debug_show_held_locks(prev); 2280 print_modules(); 2281 if (irqs_disabled()) 2282 print_irqtrace_events(prev); 2283 dump_stack(); 2284 add_taint(TAINT_WARN, LOCKDEP_STILL_OK); 2285} 2286 2287/* 2288 * Various schedule()-time debugging checks and statistics: 2289 */ 2290static inline void schedule_debug(struct task_struct *prev) 2291{ 2292 /* 2293 * Test if we are atomic. Since do_exit() needs to call into 2294 * schedule() atomically, we ignore that path for now. 2295 * Otherwise, whine if we are scheduling when we should not be. 2296 */ 2297 if (unlikely(in_atomic_preempt_off() && !prev->exit_state)) 2298 __schedule_bug(prev); 2299 rcu_sleep_check(); 2300 2301 profile_hit(SCHED_PROFILING, __builtin_return_address(0)); 2302 2303 schedstat_inc(this_rq(), sched_count); 2304} 2305 2306static void put_prev_task(struct rq *rq, struct task_struct *prev) 2307{ 2308 if (prev->on_rq || rq->skip_clock_update < 0) 2309 update_rq_clock(rq); 2310 prev->sched_class->put_prev_task(rq, prev); 2311} 2312 2313/* 2314 * Pick up the highest-prio task: 2315 */ 2316static inline struct task_struct * 2317pick_next_task(struct rq *rq) 2318{ 2319 const struct sched_class *class; 2320 struct task_struct *p; 2321 2322 /* 2323 * Optimization: we know that if all tasks are in 2324 * the fair class we can call that function directly: 2325 */ 2326 if (likely(rq->nr_running == rq->cfs.h_nr_running)) { 2327 p = fair_sched_class.pick_next_task(rq); 2328 if (likely(p)) 2329 return p; 2330 } 2331 2332 for_each_class(class) { 2333 p = class->pick_next_task(rq); 2334 if (p) 2335 return p; 2336 } 2337 2338 BUG(); /* the idle class will always have a runnable task */ 2339} 2340 2341/* 2342 * __schedule() is the main scheduler function. 2343 * 2344 * The main means of driving the scheduler and thus entering this function are: 2345 * 2346 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc. 2347 * 2348 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return 2349 * paths. For example, see arch/x86/entry_64.S. 2350 * 2351 * To drive preemption between tasks, the scheduler sets the flag in timer 2352 * interrupt handler scheduler_tick(). 2353 * 2354 * 3. Wakeups don't really cause entry into schedule(). They add a 2355 * task to the run-queue and that's it. 2356 * 2357 * Now, if the new task added to the run-queue preempts the current 2358 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets 2359 * called on the nearest possible occasion: 2360 * 2361 * - If the kernel is preemptible (CONFIG_PREEMPT=y): 2362 * 2363 * - in syscall or exception context, at the next outmost 2364 * preempt_enable(). (this might be as soon as the wake_up()'s 2365 * spin_unlock()!) 2366 * 2367 * - in IRQ context, return from interrupt-handler to 2368 * preemptible context 2369 * 2370 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set) 2371 * then at the next: 2372 * 2373 * - cond_resched() call 2374 * - explicit schedule() call 2375 * - return from syscall or exception to user-space 2376 * - return from interrupt-handler to user-space 2377 */ 2378static void __sched __schedule(void) 2379{ 2380 struct task_struct *prev, *next; 2381 unsigned long *switch_count; 2382 struct rq *rq; 2383 int cpu; 2384 2385need_resched: 2386 preempt_disable(); 2387 cpu = smp_processor_id(); 2388 rq = cpu_rq(cpu); 2389 rcu_note_context_switch(cpu); 2390 prev = rq->curr; 2391 2392 schedule_debug(prev); 2393 2394 if (sched_feat(HRTICK)) 2395 hrtick_clear(rq); 2396 2397 /* 2398 * Make sure that signal_pending_state()->signal_pending() below 2399 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE) 2400 * done by the caller to avoid the race with signal_wake_up(). 2401 */ 2402 smp_mb__before_spinlock(); 2403 raw_spin_lock_irq(&rq->lock); 2404 2405 switch_count = &prev->nivcsw; 2406 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) { 2407 if (unlikely(signal_pending_state(prev->state, prev))) { 2408 prev->state = TASK_RUNNING; 2409 } else { 2410 deactivate_task(rq, prev, DEQUEUE_SLEEP); 2411 prev->on_rq = 0; 2412 2413 /* 2414 * If a worker went to sleep, notify and ask workqueue 2415 * whether it wants to wake up a task to maintain 2416 * concurrency. 2417 */ 2418 if (prev->flags & PF_WQ_WORKER) { 2419 struct task_struct *to_wakeup; 2420 2421 to_wakeup = wq_worker_sleeping(prev, cpu); 2422 if (to_wakeup) 2423 try_to_wake_up_local(to_wakeup); 2424 } 2425 } 2426 switch_count = &prev->nvcsw; 2427 } 2428 2429 pre_schedule(rq, prev); 2430 2431 if (unlikely(!rq->nr_running)) 2432 idle_balance(cpu, rq); 2433 2434 put_prev_task(rq, prev); 2435 next = pick_next_task(rq); 2436 clear_tsk_need_resched(prev); 2437 clear_preempt_need_resched(); 2438 rq->skip_clock_update = 0; 2439 2440 if (likely(prev != next)) { 2441 rq->nr_switches++; 2442 rq->curr = next; 2443 ++*switch_count; 2444 2445 context_switch(rq, prev, next); /* unlocks the rq */ 2446 /* 2447 * The context switch have flipped the stack from under us 2448 * and restored the local variables which were saved when 2449 * this task called schedule() in the past. prev == current 2450 * is still correct, but it can be moved to another cpu/rq. 2451 */ 2452 cpu = smp_processor_id(); 2453 rq = cpu_rq(cpu); 2454 } else 2455 raw_spin_unlock_irq(&rq->lock); 2456 2457 post_schedule(rq); 2458 2459 sched_preempt_enable_no_resched(); 2460 if (need_resched()) 2461 goto need_resched; 2462} 2463 2464static inline void sched_submit_work(struct task_struct *tsk) 2465{ 2466 if (!tsk->state || tsk_is_pi_blocked(tsk)) 2467 return; 2468 /* 2469 * If we are going to sleep and we have plugged IO queued, 2470 * make sure to submit it to avoid deadlocks. 2471 */ 2472 if (blk_needs_flush_plug(tsk)) 2473 blk_schedule_flush_plug(tsk); 2474} 2475 2476asmlinkage void __sched schedule(void) 2477{ 2478 struct task_struct *tsk = current; 2479 2480 sched_submit_work(tsk); 2481 __schedule(); 2482} 2483EXPORT_SYMBOL(schedule); 2484 2485#ifdef CONFIG_CONTEXT_TRACKING 2486asmlinkage void __sched schedule_user(void) 2487{ 2488 /* 2489 * If we come here after a random call to set_need_resched(), 2490 * or we have been woken up remotely but the IPI has not yet arrived, 2491 * we haven't yet exited the RCU idle mode. Do it here manually until 2492 * we find a better solution. 2493 */ 2494 user_exit(); 2495 schedule(); 2496 user_enter(); 2497} 2498#endif 2499 2500/** 2501 * schedule_preempt_disabled - called with preemption disabled 2502 * 2503 * Returns with preemption disabled. Note: preempt_count must be 1 2504 */ 2505void __sched schedule_preempt_disabled(void) 2506{ 2507 sched_preempt_enable_no_resched(); 2508 schedule(); 2509 preempt_disable(); 2510} 2511 2512#ifdef CONFIG_PREEMPT 2513/* 2514 * this is the entry point to schedule() from in-kernel preemption 2515 * off of preempt_enable. Kernel preemptions off return from interrupt 2516 * occur there and call schedule directly. 2517 */ 2518asmlinkage void __sched notrace preempt_schedule(void) 2519{ 2520 /* 2521 * If there is a non-zero preempt_count or interrupts are disabled, 2522 * we do not want to preempt the current task. Just return.. 2523 */ 2524 if (likely(!preemptible())) 2525 return; 2526 2527 do { 2528 __preempt_count_add(PREEMPT_ACTIVE); 2529 __schedule(); 2530 __preempt_count_sub(PREEMPT_ACTIVE); 2531 2532 /* 2533 * Check again in case we missed a preemption opportunity 2534 * between schedule and now. 2535 */ 2536 barrier(); 2537 } while (need_resched()); 2538} 2539EXPORT_SYMBOL(preempt_schedule); 2540 2541/* 2542 * this is the entry point to schedule() from kernel preemption 2543 * off of irq context. 2544 * Note, that this is called and return with irqs disabled. This will 2545 * protect us against recursive calling from irq. 2546 */ 2547asmlinkage void __sched preempt_schedule_irq(void) 2548{ 2549 enum ctx_state prev_state; 2550 2551 /* Catch callers which need to be fixed */ 2552 BUG_ON(preempt_count() || !irqs_disabled()); 2553 2554 prev_state = exception_enter(); 2555 2556 do { 2557 __preempt_count_add(PREEMPT_ACTIVE); 2558 local_irq_enable(); 2559 __schedule(); 2560 local_irq_disable(); 2561 __preempt_count_sub(PREEMPT_ACTIVE); 2562 2563 /* 2564 * Check again in case we missed a preemption opportunity 2565 * between schedule and now. 2566 */ 2567 barrier(); 2568 } while (need_resched()); 2569 2570 exception_exit(prev_state); 2571} 2572 2573#endif /* CONFIG_PREEMPT */ 2574 2575int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags, 2576 void *key) 2577{ 2578 return try_to_wake_up(curr->private, mode, wake_flags); 2579} 2580EXPORT_SYMBOL(default_wake_function); 2581 2582/* 2583 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just 2584 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve 2585 * number) then we wake all the non-exclusive tasks and one exclusive task. 2586 * 2587 * There are circumstances in which we can try to wake a task which has already 2588 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns 2589 * zero in this (rare) case, and we handle it by continuing to scan the queue. 2590 */ 2591static void __wake_up_common(wait_queue_head_t *q, unsigned int mode, 2592 int nr_exclusive, int wake_flags, void *key) 2593{ 2594 wait_queue_t *curr, *next; 2595 2596 list_for_each_entry_safe(curr, next, &q->task_list, task_list) { 2597 unsigned flags = curr->flags; 2598 2599 if (curr->func(curr, mode, wake_flags, key) && 2600 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive) 2601 break; 2602 } 2603} 2604 2605/** 2606 * __wake_up - wake up threads blocked on a waitqueue. 2607 * @q: the waitqueue 2608 * @mode: which threads 2609 * @nr_exclusive: how many wake-one or wake-many threads to wake up 2610 * @key: is directly passed to the wakeup function 2611 * 2612 * It may be assumed that this function implies a write memory barrier before 2613 * changing the task state if and only if any tasks are woken up. 2614 */ 2615void __wake_up(wait_queue_head_t *q, unsigned int mode, 2616 int nr_exclusive, void *key) 2617{ 2618 unsigned long flags; 2619 2620 spin_lock_irqsave(&q->lock, flags); 2621 __wake_up_common(q, mode, nr_exclusive, 0, key); 2622 spin_unlock_irqrestore(&q->lock, flags); 2623} 2624EXPORT_SYMBOL(__wake_up); 2625 2626/* 2627 * Same as __wake_up but called with the spinlock in wait_queue_head_t held. 2628 */ 2629void __wake_up_locked(wait_queue_head_t *q, unsigned int mode, int nr) 2630{ 2631 __wake_up_common(q, mode, nr, 0, NULL); 2632} 2633EXPORT_SYMBOL_GPL(__wake_up_locked); 2634 2635void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key) 2636{ 2637 __wake_up_common(q, mode, 1, 0, key); 2638} 2639EXPORT_SYMBOL_GPL(__wake_up_locked_key); 2640 2641/** 2642 * __wake_up_sync_key - wake up threads blocked on a waitqueue. 2643 * @q: the waitqueue 2644 * @mode: which threads 2645 * @nr_exclusive: how many wake-one or wake-many threads to wake up 2646 * @key: opaque value to be passed to wakeup targets 2647 * 2648 * The sync wakeup differs that the waker knows that it will schedule 2649 * away soon, so while the target thread will be woken up, it will not 2650 * be migrated to another CPU - ie. the two threads are 'synchronized' 2651 * with each other. This can prevent needless bouncing between CPUs. 2652 * 2653 * On UP it can prevent extra preemption. 2654 * 2655 * It may be assumed that this function implies a write memory barrier before 2656 * changing the task state if and only if any tasks are woken up. 2657 */ 2658void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode, 2659 int nr_exclusive, void *key) 2660{ 2661 unsigned long flags; 2662 int wake_flags = WF_SYNC; 2663 2664 if (unlikely(!q)) 2665 return; 2666 2667 if (unlikely(nr_exclusive != 1)) 2668 wake_flags = 0; 2669 2670 spin_lock_irqsave(&q->lock, flags); 2671 __wake_up_common(q, mode, nr_exclusive, wake_flags, key); 2672 spin_unlock_irqrestore(&q->lock, flags); 2673} 2674EXPORT_SYMBOL_GPL(__wake_up_sync_key); 2675 2676/* 2677 * __wake_up_sync - see __wake_up_sync_key() 2678 */ 2679void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive) 2680{ 2681 __wake_up_sync_key(q, mode, nr_exclusive, NULL); 2682} 2683EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */ 2684 2685/** 2686 * complete: - signals a single thread waiting on this completion 2687 * @x: holds the state of this particular completion 2688 * 2689 * This will wake up a single thread waiting on this completion. Threads will be 2690 * awakened in the same order in which they were queued. 2691 * 2692 * See also complete_all(), wait_for_completion() and related routines. 2693 * 2694 * It may be assumed that this function implies a write memory barrier before 2695 * changing the task state if and only if any tasks are woken up. 2696 */ 2697void complete(struct completion *x) 2698{ 2699 unsigned long flags; 2700 2701 spin_lock_irqsave(&x->wait.lock, flags); 2702 x->done++; 2703 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL); 2704 spin_unlock_irqrestore(&x->wait.lock, flags); 2705} 2706EXPORT_SYMBOL(complete); 2707 2708/** 2709 * complete_all: - signals all threads waiting on this completion 2710 * @x: holds the state of this particular completion 2711 * 2712 * This will wake up all threads waiting on this particular completion event. 2713 * 2714 * It may be assumed that this function implies a write memory barrier before 2715 * changing the task state if and only if any tasks are woken up. 2716 */ 2717void complete_all(struct completion *x) 2718{ 2719 unsigned long flags; 2720 2721 spin_lock_irqsave(&x->wait.lock, flags); 2722 x->done += UINT_MAX/2; 2723 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL); 2724 spin_unlock_irqrestore(&x->wait.lock, flags); 2725} 2726EXPORT_SYMBOL(complete_all); 2727 2728static inline long __sched 2729do_wait_for_common(struct completion *x, 2730 long (*action)(long), long timeout, int state) 2731{ 2732 if (!x->done) { 2733 DECLARE_WAITQUEUE(wait, current); 2734 2735 __add_wait_queue_tail_exclusive(&x->wait, &wait); 2736 do { 2737 if (signal_pending_state(state, current)) { 2738 timeout = -ERESTARTSYS; 2739 break; 2740 } 2741 __set_current_state(state); 2742 spin_unlock_irq(&x->wait.lock); 2743 timeout = action(timeout); 2744 spin_lock_irq(&x->wait.lock); 2745 } while (!x->done && timeout); 2746 __remove_wait_queue(&x->wait, &wait); 2747 if (!x->done) 2748 return timeout; 2749 } 2750 x->done--; 2751 return timeout ?: 1; 2752} 2753 2754static inline long __sched 2755__wait_for_common(struct completion *x, 2756 long (*action)(long), long timeout, int state) 2757{ 2758 might_sleep(); 2759 2760 spin_lock_irq(&x->wait.lock); 2761 timeout = do_wait_for_common(x, action, timeout, state); 2762 spin_unlock_irq(&x->wait.lock); 2763 return timeout; 2764} 2765 2766static long __sched 2767wait_for_common(struct completion *x, long timeout, int state) 2768{ 2769 return __wait_for_common(x, schedule_timeout, timeout, state); 2770} 2771 2772static long __sched 2773wait_for_common_io(struct completion *x, long timeout, int state) 2774{ 2775 return __wait_for_common(x, io_schedule_timeout, timeout, state); 2776} 2777 2778/** 2779 * wait_for_completion: - waits for completion of a task 2780 * @x: holds the state of this particular completion 2781 * 2782 * This waits to be signaled for completion of a specific task. It is NOT 2783 * interruptible and there is no timeout. 2784 * 2785 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout 2786 * and interrupt capability. Also see complete(). 2787 */ 2788void __sched wait_for_completion(struct completion *x) 2789{ 2790 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE); 2791} 2792EXPORT_SYMBOL(wait_for_completion); 2793 2794/** 2795 * wait_for_completion_timeout: - waits for completion of a task (w/timeout) 2796 * @x: holds the state of this particular completion 2797 * @timeout: timeout value in jiffies 2798 * 2799 * This waits for either a completion of a specific task to be signaled or for a 2800 * specified timeout to expire. The timeout is in jiffies. It is not 2801 * interruptible. 2802 * 2803 * Return: 0 if timed out, and positive (at least 1, or number of jiffies left 2804 * till timeout) if completed. 2805 */ 2806unsigned long __sched 2807wait_for_completion_timeout(struct completion *x, unsigned long timeout) 2808{ 2809 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE); 2810} 2811EXPORT_SYMBOL(wait_for_completion_timeout); 2812 2813/** 2814 * wait_for_completion_io: - waits for completion of a task 2815 * @x: holds the state of this particular completion 2816 * 2817 * This waits to be signaled for completion of a specific task. It is NOT 2818 * interruptible and there is no timeout. The caller is accounted as waiting 2819 * for IO. 2820 */ 2821void __sched wait_for_completion_io(struct completion *x) 2822{ 2823 wait_for_common_io(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE); 2824} 2825EXPORT_SYMBOL(wait_for_completion_io); 2826 2827/** 2828 * wait_for_completion_io_timeout: - waits for completion of a task (w/timeout) 2829 * @x: holds the state of this particular completion 2830 * @timeout: timeout value in jiffies 2831 * 2832 * This waits for either a completion of a specific task to be signaled or for a 2833 * specified timeout to expire. The timeout is in jiffies. It is not 2834 * interruptible. The caller is accounted as waiting for IO. 2835 * 2836 * Return: 0 if timed out, and positive (at least 1, or number of jiffies left 2837 * till timeout) if completed. 2838 */ 2839unsigned long __sched 2840wait_for_completion_io_timeout(struct completion *x, unsigned long timeout) 2841{ 2842 return wait_for_common_io(x, timeout, TASK_UNINTERRUPTIBLE); 2843} 2844EXPORT_SYMBOL(wait_for_completion_io_timeout); 2845 2846/** 2847 * wait_for_completion_interruptible: - waits for completion of a task (w/intr) 2848 * @x: holds the state of this particular completion 2849 * 2850 * This waits for completion of a specific task to be signaled. It is 2851 * interruptible. 2852 * 2853 * Return: -ERESTARTSYS if interrupted, 0 if completed. 2854 */ 2855int __sched wait_for_completion_interruptible(struct completion *x) 2856{ 2857 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE); 2858 if (t == -ERESTARTSYS) 2859 return t; 2860 return 0; 2861} 2862EXPORT_SYMBOL(wait_for_completion_interruptible); 2863 2864/** 2865 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr)) 2866 * @x: holds the state of this particular completion 2867 * @timeout: timeout value in jiffies 2868 * 2869 * This waits for either a completion of a specific task to be signaled or for a 2870 * specified timeout to expire. It is interruptible. The timeout is in jiffies. 2871 * 2872 * Return: -ERESTARTSYS if interrupted, 0 if timed out, positive (at least 1, 2873 * or number of jiffies left till timeout) if completed. 2874 */ 2875long __sched 2876wait_for_completion_interruptible_timeout(struct completion *x, 2877 unsigned long timeout) 2878{ 2879 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE); 2880} 2881EXPORT_SYMBOL(wait_for_completion_interruptible_timeout); 2882 2883/** 2884 * wait_for_completion_killable: - waits for completion of a task (killable) 2885 * @x: holds the state of this particular completion 2886 * 2887 * This waits to be signaled for completion of a specific task. It can be 2888 * interrupted by a kill signal. 2889 * 2890 * Return: -ERESTARTSYS if interrupted, 0 if completed. 2891 */ 2892int __sched wait_for_completion_killable(struct completion *x) 2893{ 2894 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE); 2895 if (t == -ERESTARTSYS) 2896 return t; 2897 return 0; 2898} 2899EXPORT_SYMBOL(wait_for_completion_killable); 2900 2901/** 2902 * wait_for_completion_killable_timeout: - waits for completion of a task (w/(to,killable)) 2903 * @x: holds the state of this particular completion 2904 * @timeout: timeout value in jiffies 2905 * 2906 * This waits for either a completion of a specific task to be 2907 * signaled or for a specified timeout to expire. It can be 2908 * interrupted by a kill signal. The timeout is in jiffies. 2909 * 2910 * Return: -ERESTARTSYS if interrupted, 0 if timed out, positive (at least 1, 2911 * or number of jiffies left till timeout) if completed. 2912 */ 2913long __sched 2914wait_for_completion_killable_timeout(struct completion *x, 2915 unsigned long timeout) 2916{ 2917 return wait_for_common(x, timeout, TASK_KILLABLE); 2918} 2919EXPORT_SYMBOL(wait_for_completion_killable_timeout); 2920 2921/** 2922 * try_wait_for_completion - try to decrement a completion without blocking 2923 * @x: completion structure 2924 * 2925 * Return: 0 if a decrement cannot be done without blocking 2926 * 1 if a decrement succeeded. 2927 * 2928 * If a completion is being used as a counting completion, 2929 * attempt to decrement the counter without blocking. This 2930 * enables us to avoid waiting if the resource the completion 2931 * is protecting is not available. 2932 */ 2933bool try_wait_for_completion(struct completion *x) 2934{ 2935 unsigned long flags; 2936 int ret = 1; 2937 2938 spin_lock_irqsave(&x->wait.lock, flags); 2939 if (!x->done) 2940 ret = 0; 2941 else 2942 x->done--; 2943 spin_unlock_irqrestore(&x->wait.lock, flags); 2944 return ret; 2945} 2946EXPORT_SYMBOL(try_wait_for_completion); 2947 2948/** 2949 * completion_done - Test to see if a completion has any waiters 2950 * @x: completion structure 2951 * 2952 * Return: 0 if there are waiters (wait_for_completion() in progress) 2953 * 1 if there are no waiters. 2954 * 2955 */ 2956bool completion_done(struct completion *x) 2957{ 2958 unsigned long flags; 2959 int ret = 1; 2960 2961 spin_lock_irqsave(&x->wait.lock, flags); 2962 if (!x->done) 2963 ret = 0; 2964 spin_unlock_irqrestore(&x->wait.lock, flags); 2965 return ret; 2966} 2967EXPORT_SYMBOL(completion_done); 2968 2969static long __sched 2970sleep_on_common(wait_queue_head_t *q, int state, long timeout) 2971{ 2972 unsigned long flags; 2973 wait_queue_t wait; 2974 2975 init_waitqueue_entry(&wait, current); 2976 2977 __set_current_state(state); 2978 2979 spin_lock_irqsave(&q->lock, flags); 2980 __add_wait_queue(q, &wait); 2981 spin_unlock(&q->lock); 2982 timeout = schedule_timeout(timeout); 2983 spin_lock_irq(&q->lock); 2984 __remove_wait_queue(q, &wait); 2985 spin_unlock_irqrestore(&q->lock, flags); 2986 2987 return timeout; 2988} 2989 2990void __sched interruptible_sleep_on(wait_queue_head_t *q) 2991{ 2992 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT); 2993} 2994EXPORT_SYMBOL(interruptible_sleep_on); 2995 2996long __sched 2997interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout) 2998{ 2999 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout); 3000} 3001EXPORT_SYMBOL(interruptible_sleep_on_timeout); 3002 3003void __sched sleep_on(wait_queue_head_t *q) 3004{ 3005 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT); 3006} 3007EXPORT_SYMBOL(sleep_on); 3008 3009long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout) 3010{ 3011 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout); 3012} 3013EXPORT_SYMBOL(sleep_on_timeout); 3014 3015#ifdef CONFIG_RT_MUTEXES 3016 3017/* 3018 * rt_mutex_setprio - set the current priority of a task 3019 * @p: task 3020 * @prio: prio value (kernel-internal form) 3021 * 3022 * This function changes the 'effective' priority of a task. It does 3023 * not touch ->normal_prio like __setscheduler(). 3024 * 3025 * Used by the rt_mutex code to implement priority inheritance logic. 3026 */ 3027void rt_mutex_setprio(struct task_struct *p, int prio) 3028{ 3029 int oldprio, on_rq, running; 3030 struct rq *rq; 3031 const struct sched_class *prev_class; 3032 3033 BUG_ON(prio < 0 || prio > MAX_PRIO); 3034 3035 rq = __task_rq_lock(p); 3036 3037 /* 3038 * Idle task boosting is a nono in general. There is one 3039 * exception, when PREEMPT_RT and NOHZ is active: 3040 * 3041 * The idle task calls get_next_timer_interrupt() and holds 3042 * the timer wheel base->lock on the CPU and another CPU wants 3043 * to access the timer (probably to cancel it). We can safely 3044 * ignore the boosting request, as the idle CPU runs this code 3045 * with interrupts disabled and will complete the lock 3046 * protected section without being interrupted. So there is no 3047 * real need to boost. 3048 */ 3049 if (unlikely(p == rq->idle)) { 3050 WARN_ON(p != rq->curr); 3051 WARN_ON(p->pi_blocked_on); 3052 goto out_unlock; 3053 } 3054 3055 trace_sched_pi_setprio(p, prio); 3056 oldprio = p->prio; 3057 prev_class = p->sched_class; 3058 on_rq = p->on_rq; 3059 running = task_current(rq, p); 3060 if (on_rq) 3061 dequeue_task(rq, p, 0); 3062 if (running) 3063 p->sched_class->put_prev_task(rq, p); 3064 3065 if (rt_prio(prio)) 3066 p->sched_class = &rt_sched_class; 3067 else 3068 p->sched_class = &fair_sched_class; 3069 3070 p->prio = prio; 3071 3072 if (running) 3073 p->sched_class->set_curr_task(rq); 3074 if (on_rq) 3075 enqueue_task(rq, p, oldprio < prio ? ENQUEUE_HEAD : 0); 3076 3077 check_class_changed(rq, p, prev_class, oldprio); 3078out_unlock: 3079 __task_rq_unlock(rq); 3080} 3081#endif 3082void set_user_nice(struct task_struct *p, long nice) 3083{ 3084 int old_prio, delta, on_rq; 3085 unsigned long flags; 3086 struct rq *rq; 3087 3088 if (TASK_NICE(p) == nice || nice < -20 || nice > 19) 3089 return; 3090 /* 3091 * We have to be careful, if called from sys_setpriority(), 3092 * the task might be in the middle of scheduling on another CPU. 3093 */ 3094 rq = task_rq_lock(p, &flags); 3095 /* 3096 * The RT priorities are set via sched_setscheduler(), but we still 3097 * allow the 'normal' nice value to be set - but as expected 3098 * it wont have any effect on scheduling until the task is 3099 * SCHED_FIFO/SCHED_RR: 3100 */ 3101 if (task_has_rt_policy(p)) { 3102 p->static_prio = NICE_TO_PRIO(nice); 3103 goto out_unlock; 3104 } 3105 on_rq = p->on_rq; 3106 if (on_rq) 3107 dequeue_task(rq, p, 0); 3108 3109 p->static_prio = NICE_TO_PRIO(nice); 3110 set_load_weight(p); 3111 old_prio = p->prio; 3112 p->prio = effective_prio(p); 3113 delta = p->prio - old_prio; 3114 3115 if (on_rq) { 3116 enqueue_task(rq, p, 0); 3117 /* 3118 * If the task increased its priority or is running and 3119 * lowered its priority, then reschedule its CPU: 3120 */ 3121 if (delta < 0 || (delta > 0 && task_running(rq, p))) 3122 resched_task(rq->curr); 3123 } 3124out_unlock: 3125 task_rq_unlock(rq, p, &flags); 3126} 3127EXPORT_SYMBOL(set_user_nice); 3128 3129/* 3130 * can_nice - check if a task can reduce its nice value 3131 * @p: task 3132 * @nice: nice value 3133 */ 3134int can_nice(const struct task_struct *p, const int nice) 3135{ 3136 /* convert nice value [19,-20] to rlimit style value [1,40] */ 3137 int nice_rlim = 20 - nice; 3138 3139 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) || 3140 capable(CAP_SYS_NICE)); 3141} 3142 3143#ifdef __ARCH_WANT_SYS_NICE 3144 3145/* 3146 * sys_nice - change the priority of the current process. 3147 * @increment: priority increment 3148 * 3149 * sys_setpriority is a more generic, but much slower function that 3150 * does similar things. 3151 */ 3152SYSCALL_DEFINE1(nice, int, increment) 3153{ 3154 long nice, retval; 3155 3156 /* 3157 * Setpriority might change our priority at the same moment. 3158 * We don't have to worry. Conceptually one call occurs first 3159 * and we have a single winner. 3160 */ 3161 if (increment < -40) 3162 increment = -40; 3163 if (increment > 40) 3164 increment = 40; 3165 3166 nice = TASK_NICE(current) + increment; 3167 if (nice < -20) 3168 nice = -20; 3169 if (nice > 19) 3170 nice = 19; 3171 3172 if (increment < 0 && !can_nice(current, nice)) 3173 return -EPERM; 3174 3175 retval = security_task_setnice(current, nice); 3176 if (retval) 3177 return retval; 3178 3179 set_user_nice(current, nice); 3180 return 0; 3181} 3182 3183#endif 3184 3185/** 3186 * task_prio - return the priority value of a given task. 3187 * @p: the task in question. 3188 * 3189 * Return: The priority value as seen by users in /proc. 3190 * RT tasks are offset by -200. Normal tasks are centered 3191 * around 0, value goes from -16 to +15. 3192 */ 3193int task_prio(const struct task_struct *p) 3194{ 3195 return p->prio - MAX_RT_PRIO; 3196} 3197 3198/** 3199 * task_nice - return the nice value of a given task. 3200 * @p: the task in question. 3201 * 3202 * Return: The nice value [ -20 ... 0 ... 19 ]. 3203 */ 3204int task_nice(const struct task_struct *p) 3205{ 3206 return TASK_NICE(p); 3207} 3208EXPORT_SYMBOL(task_nice); 3209 3210/** 3211 * idle_cpu - is a given cpu idle currently? 3212 * @cpu: the processor in question. 3213 * 3214 * Return: 1 if the CPU is currently idle. 0 otherwise. 3215 */ 3216int idle_cpu(int cpu) 3217{ 3218 struct rq *rq = cpu_rq(cpu); 3219 3220 if (rq->curr != rq->idle) 3221 return 0; 3222 3223 if (rq->nr_running) 3224 return 0; 3225 3226#ifdef CONFIG_SMP 3227 if (!llist_empty(&rq->wake_list)) 3228 return 0; 3229#endif 3230 3231 return 1; 3232} 3233 3234/** 3235 * idle_task - return the idle task for a given cpu. 3236 * @cpu: the processor in question. 3237 * 3238 * Return: The idle task for the cpu @cpu. 3239 */ 3240struct task_struct *idle_task(int cpu) 3241{ 3242 return cpu_rq(cpu)->idle; 3243} 3244 3245/** 3246 * find_process_by_pid - find a process with a matching PID value. 3247 * @pid: the pid in question. 3248 * 3249 * The task of @pid, if found. %NULL otherwise. 3250 */ 3251static struct task_struct *find_process_by_pid(pid_t pid) 3252{ 3253 return pid ? find_task_by_vpid(pid) : current; 3254} 3255 3256/* Actually do priority change: must hold rq lock. */ 3257static void 3258__setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio) 3259{ 3260 p->policy = policy; 3261 p->rt_priority = prio; 3262 p->normal_prio = normal_prio(p); 3263 /* we are holding p->pi_lock already */ 3264 p->prio = rt_mutex_getprio(p); 3265 if (rt_prio(p->prio)) 3266 p->sched_class = &rt_sched_class; 3267 else 3268 p->sched_class = &fair_sched_class; 3269 set_load_weight(p); 3270} 3271 3272/* 3273 * check the target process has a UID that matches the current process's 3274 */ 3275static bool check_same_owner(struct task_struct *p) 3276{ 3277 const struct cred *cred = current_cred(), *pcred; 3278 bool match; 3279 3280 rcu_read_lock(); 3281 pcred = __task_cred(p); 3282 match = (uid_eq(cred->euid, pcred->euid) || 3283 uid_eq(cred->euid, pcred->uid)); 3284 rcu_read_unlock(); 3285 return match; 3286} 3287 3288static int __sched_setscheduler(struct task_struct *p, int policy, 3289 const struct sched_param *param, bool user) 3290{ 3291 int retval, oldprio, oldpolicy = -1, on_rq, running; 3292 unsigned long flags; 3293 const struct sched_class *prev_class; 3294 struct rq *rq; 3295 int reset_on_fork; 3296 3297 /* may grab non-irq protected spin_locks */ 3298 BUG_ON(in_interrupt()); 3299recheck: 3300 /* double check policy once rq lock held */ 3301 if (policy < 0) { 3302 reset_on_fork = p->sched_reset_on_fork; 3303 policy = oldpolicy = p->policy; 3304 } else { 3305 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK); 3306 policy &= ~SCHED_RESET_ON_FORK; 3307 3308 if (policy != SCHED_FIFO && policy != SCHED_RR && 3309 policy != SCHED_NORMAL && policy != SCHED_BATCH && 3310 policy != SCHED_IDLE) 3311 return -EINVAL; 3312 } 3313 3314 /* 3315 * Valid priorities for SCHED_FIFO and SCHED_RR are 3316 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL, 3317 * SCHED_BATCH and SCHED_IDLE is 0. 3318 */ 3319 if (param->sched_priority < 0 || 3320 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) || 3321 (!p->mm && param->sched_priority > MAX_RT_PRIO-1)) 3322 return -EINVAL; 3323 if (rt_policy(policy) != (param->sched_priority != 0)) 3324 return -EINVAL; 3325 3326 /* 3327 * Allow unprivileged RT tasks to decrease priority: 3328 */ 3329 if (user && !capable(CAP_SYS_NICE)) { 3330 if (rt_policy(policy)) { 3331 unsigned long rlim_rtprio = 3332 task_rlimit(p, RLIMIT_RTPRIO); 3333 3334 /* can't set/change the rt policy */ 3335 if (policy != p->policy && !rlim_rtprio) 3336 return -EPERM; 3337 3338 /* can't increase priority */ 3339 if (param->sched_priority > p->rt_priority && 3340 param->sched_priority > rlim_rtprio) 3341 return -EPERM; 3342 } 3343 3344 /* 3345 * Treat SCHED_IDLE as nice 20. Only allow a switch to 3346 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it. 3347 */ 3348 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) { 3349 if (!can_nice(p, TASK_NICE(p))) 3350 return -EPERM; 3351 } 3352 3353 /* can't change other user's priorities */ 3354 if (!check_same_owner(p)) 3355 return -EPERM; 3356 3357 /* Normal users shall not reset the sched_reset_on_fork flag */ 3358 if (p->sched_reset_on_fork && !reset_on_fork) 3359 return -EPERM; 3360 } 3361 3362 if (user) { 3363 retval = security_task_setscheduler(p); 3364 if (retval) 3365 return retval; 3366 } 3367 3368 /* 3369 * make sure no PI-waiters arrive (or leave) while we are 3370 * changing the priority of the task: 3371 * 3372 * To be able to change p->policy safely, the appropriate 3373 * runqueue lock must be held. 3374 */ 3375 rq = task_rq_lock(p, &flags); 3376 3377 /* 3378 * Changing the policy of the stop threads its a very bad idea 3379 */ 3380 if (p == rq->stop) { 3381 task_rq_unlock(rq, p, &flags); 3382 return -EINVAL; 3383 } 3384 3385 /* 3386 * If not changing anything there's no need to proceed further: 3387 */ 3388 if (unlikely(policy == p->policy && (!rt_policy(policy) || 3389 param->sched_priority == p->rt_priority))) { 3390 task_rq_unlock(rq, p, &flags); 3391 return 0; 3392 } 3393 3394#ifdef CONFIG_RT_GROUP_SCHED 3395 if (user) { 3396 /* 3397 * Do not allow realtime tasks into groups that have no runtime 3398 * assigned. 3399 */ 3400 if (rt_bandwidth_enabled() && rt_policy(policy) && 3401 task_group(p)->rt_bandwidth.rt_runtime == 0 && 3402 !task_group_is_autogroup(task_group(p))) { 3403 task_rq_unlock(rq, p, &flags); 3404 return -EPERM; 3405 } 3406 } 3407#endif 3408 3409 /* recheck policy now with rq lock held */ 3410 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) { 3411 policy = oldpolicy = -1; 3412 task_rq_unlock(rq, p, &flags); 3413 goto recheck; 3414 } 3415 on_rq = p->on_rq; 3416 running = task_current(rq, p); 3417 if (on_rq) 3418 dequeue_task(rq, p, 0); 3419 if (running) 3420 p->sched_class->put_prev_task(rq, p); 3421 3422 p->sched_reset_on_fork = reset_on_fork; 3423 3424 oldprio = p->prio; 3425 prev_class = p->sched_class; 3426 __setscheduler(rq, p, policy, param->sched_priority); 3427 3428 if (running) 3429 p->sched_class->set_curr_task(rq); 3430 if (on_rq) 3431 enqueue_task(rq, p, 0); 3432 3433 check_class_changed(rq, p, prev_class, oldprio); 3434 task_rq_unlock(rq, p, &flags); 3435 3436 rt_mutex_adjust_pi(p); 3437 3438 return 0; 3439} 3440 3441/** 3442 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread. 3443 * @p: the task in question. 3444 * @policy: new policy. 3445 * @param: structure containing the new RT priority. 3446 * 3447 * Return: 0 on success. An error code otherwise. 3448 * 3449 * NOTE that the task may be already dead. 3450 */ 3451int sched_setscheduler(struct task_struct *p, int policy, 3452 const struct sched_param *param) 3453{ 3454 return __sched_setscheduler(p, policy, param, true); 3455} 3456EXPORT_SYMBOL_GPL(sched_setscheduler); 3457 3458/** 3459 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace. 3460 * @p: the task in question. 3461 * @policy: new policy. 3462 * @param: structure containing the new RT priority. 3463 * 3464 * Just like sched_setscheduler, only don't bother checking if the 3465 * current context has permission. For example, this is needed in 3466 * stop_machine(): we create temporary high priority worker threads, 3467 * but our caller might not have that capability. 3468 * 3469 * Return: 0 on success. An error code otherwise. 3470 */ 3471int sched_setscheduler_nocheck(struct task_struct *p, int policy, 3472 const struct sched_param *param) 3473{ 3474 return __sched_setscheduler(p, policy, param, false); 3475} 3476 3477static int 3478do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param) 3479{ 3480 struct sched_param lparam; 3481 struct task_struct *p; 3482 int retval; 3483 3484 if (!param || pid < 0) 3485 return -EINVAL; 3486 if (copy_from_user(&lparam, param, sizeof(struct sched_param))) 3487 return -EFAULT; 3488 3489 rcu_read_lock(); 3490 retval = -ESRCH; 3491 p = find_process_by_pid(pid); 3492 if (p != NULL) 3493 retval = sched_setscheduler(p, policy, &lparam); 3494 rcu_read_unlock(); 3495 3496 return retval; 3497} 3498 3499/** 3500 * sys_sched_setscheduler - set/change the scheduler policy and RT priority 3501 * @pid: the pid in question. 3502 * @policy: new policy. 3503 * @param: structure containing the new RT priority. 3504 * 3505 * Return: 0 on success. An error code otherwise. 3506 */ 3507SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy, 3508 struct sched_param __user *, param) 3509{ 3510 /* negative values for policy are not valid */ 3511 if (policy < 0) 3512 return -EINVAL; 3513 3514 return do_sched_setscheduler(pid, policy, param); 3515} 3516 3517/** 3518 * sys_sched_setparam - set/change the RT priority of a thread 3519 * @pid: the pid in question. 3520 * @param: structure containing the new RT priority. 3521 * 3522 * Return: 0 on success. An error code otherwise. 3523 */ 3524SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param) 3525{ 3526 return do_sched_setscheduler(pid, -1, param); 3527} 3528 3529/** 3530 * sys_sched_getscheduler - get the policy (scheduling class) of a thread 3531 * @pid: the pid in question. 3532 * 3533 * Return: On success, the policy of the thread. Otherwise, a negative error 3534 * code. 3535 */ 3536SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid) 3537{ 3538 struct task_struct *p; 3539 int retval; 3540 3541 if (pid < 0) 3542 return -EINVAL; 3543 3544 retval = -ESRCH; 3545 rcu_read_lock(); 3546 p = find_process_by_pid(pid); 3547 if (p) { 3548 retval = security_task_getscheduler(p); 3549 if (!retval) 3550 retval = p->policy 3551 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0); 3552 } 3553 rcu_read_unlock(); 3554 return retval; 3555} 3556 3557/** 3558 * sys_sched_getparam - get the RT priority of a thread 3559 * @pid: the pid in question. 3560 * @param: structure containing the RT priority. 3561 * 3562 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error 3563 * code. 3564 */ 3565SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param) 3566{ 3567 struct sched_param lp; 3568 struct task_struct *p; 3569 int retval; 3570 3571 if (!param || pid < 0) 3572 return -EINVAL; 3573 3574 rcu_read_lock(); 3575 p = find_process_by_pid(pid); 3576 retval = -ESRCH; 3577 if (!p) 3578 goto out_unlock; 3579 3580 retval = security_task_getscheduler(p); 3581 if (retval) 3582 goto out_unlock; 3583 3584 lp.sched_priority = p->rt_priority; 3585 rcu_read_unlock(); 3586 3587 /* 3588 * This one might sleep, we cannot do it with a spinlock held ... 3589 */ 3590 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0; 3591 3592 return retval; 3593 3594out_unlock: 3595 rcu_read_unlock(); 3596 return retval; 3597} 3598 3599long sched_setaffinity(pid_t pid, const struct cpumask *in_mask) 3600{ 3601 cpumask_var_t cpus_allowed, new_mask; 3602 struct task_struct *p; 3603 int retval; 3604 3605 get_online_cpus(); 3606 rcu_read_lock(); 3607 3608 p = find_process_by_pid(pid); 3609 if (!p) { 3610 rcu_read_unlock(); 3611 put_online_cpus(); 3612 return -ESRCH; 3613 } 3614 3615 /* Prevent p going away */ 3616 get_task_struct(p); 3617 rcu_read_unlock(); 3618 3619 if (p->flags & PF_NO_SETAFFINITY) { 3620 retval = -EINVAL; 3621 goto out_put_task; 3622 } 3623 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) { 3624 retval = -ENOMEM; 3625 goto out_put_task; 3626 } 3627 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) { 3628 retval = -ENOMEM; 3629 goto out_free_cpus_allowed; 3630 } 3631 retval = -EPERM; 3632 if (!check_same_owner(p)) { 3633 rcu_read_lock(); 3634 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) { 3635 rcu_read_unlock(); 3636 goto out_unlock; 3637 } 3638 rcu_read_unlock(); 3639 } 3640 3641 retval = security_task_setscheduler(p); 3642 if (retval) 3643 goto out_unlock; 3644 3645 cpuset_cpus_allowed(p, cpus_allowed); 3646 cpumask_and(new_mask, in_mask, cpus_allowed); 3647again: 3648 retval = set_cpus_allowed_ptr(p, new_mask); 3649 3650 if (!retval) { 3651 cpuset_cpus_allowed(p, cpus_allowed); 3652 if (!cpumask_subset(new_mask, cpus_allowed)) { 3653 /* 3654 * We must have raced with a concurrent cpuset 3655 * update. Just reset the cpus_allowed to the 3656 * cpuset's cpus_allowed 3657 */ 3658 cpumask_copy(new_mask, cpus_allowed); 3659 goto again; 3660 } 3661 } 3662out_unlock: 3663 free_cpumask_var(new_mask); 3664out_free_cpus_allowed: 3665 free_cpumask_var(cpus_allowed); 3666out_put_task: 3667 put_task_struct(p); 3668 put_online_cpus(); 3669 return retval; 3670} 3671 3672static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len, 3673 struct cpumask *new_mask) 3674{ 3675 if (len < cpumask_size()) 3676 cpumask_clear(new_mask); 3677 else if (len > cpumask_size()) 3678 len = cpumask_size(); 3679 3680 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0; 3681} 3682 3683/** 3684 * sys_sched_setaffinity - set the cpu affinity of a process 3685 * @pid: pid of the process 3686 * @len: length in bytes of the bitmask pointed to by user_mask_ptr 3687 * @user_mask_ptr: user-space pointer to the new cpu mask 3688 * 3689 * Return: 0 on success. An error code otherwise. 3690 */ 3691SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len, 3692 unsigned long __user *, user_mask_ptr) 3693{ 3694 cpumask_var_t new_mask; 3695 int retval; 3696 3697 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) 3698 return -ENOMEM; 3699 3700 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask); 3701 if (retval == 0) 3702 retval = sched_setaffinity(pid, new_mask); 3703 free_cpumask_var(new_mask); 3704 return retval; 3705} 3706 3707long sched_getaffinity(pid_t pid, struct cpumask *mask) 3708{ 3709 struct task_struct *p; 3710 unsigned long flags; 3711 int retval; 3712 3713 get_online_cpus(); 3714 rcu_read_lock(); 3715 3716 retval = -ESRCH; 3717 p = find_process_by_pid(pid); 3718 if (!p) 3719 goto out_unlock; 3720 3721 retval = security_task_getscheduler(p); 3722 if (retval) 3723 goto out_unlock; 3724 3725 raw_spin_lock_irqsave(&p->pi_lock, flags); 3726 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask); 3727 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 3728 3729out_unlock: 3730 rcu_read_unlock(); 3731 put_online_cpus(); 3732 3733 return retval; 3734} 3735 3736/** 3737 * sys_sched_getaffinity - get the cpu affinity of a process 3738 * @pid: pid of the process 3739 * @len: length in bytes of the bitmask pointed to by user_mask_ptr 3740 * @user_mask_ptr: user-space pointer to hold the current cpu mask 3741 * 3742 * Return: 0 on success. An error code otherwise. 3743 */ 3744SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len, 3745 unsigned long __user *, user_mask_ptr) 3746{ 3747 int ret; 3748 cpumask_var_t mask; 3749 3750 if ((len * BITS_PER_BYTE) < nr_cpu_ids) 3751 return -EINVAL; 3752 if (len & (sizeof(unsigned long)-1)) 3753 return -EINVAL; 3754 3755 if (!alloc_cpumask_var(&mask, GFP_KERNEL)) 3756 return -ENOMEM; 3757 3758 ret = sched_getaffinity(pid, mask); 3759 if (ret == 0) { 3760 size_t retlen = min_t(size_t, len, cpumask_size()); 3761 3762 if (copy_to_user(user_mask_ptr, mask, retlen)) 3763 ret = -EFAULT; 3764 else 3765 ret = retlen; 3766 } 3767 free_cpumask_var(mask); 3768 3769 return ret; 3770} 3771 3772/** 3773 * sys_sched_yield - yield the current processor to other threads. 3774 * 3775 * This function yields the current CPU to other tasks. If there are no 3776 * other threads running on this CPU then this function will return. 3777 * 3778 * Return: 0. 3779 */ 3780SYSCALL_DEFINE0(sched_yield) 3781{ 3782 struct rq *rq = this_rq_lock(); 3783 3784 schedstat_inc(rq, yld_count); 3785 current->sched_class->yield_task(rq); 3786 3787 /* 3788 * Since we are going to call schedule() anyway, there's 3789 * no need to preempt or enable interrupts: 3790 */ 3791 __release(rq->lock); 3792 spin_release(&rq->lock.dep_map, 1, _THIS_IP_); 3793 do_raw_spin_unlock(&rq->lock); 3794 sched_preempt_enable_no_resched(); 3795 3796 schedule(); 3797 3798 return 0; 3799} 3800 3801static void __cond_resched(void) 3802{ 3803 __preempt_count_add(PREEMPT_ACTIVE); 3804 __schedule(); 3805 __preempt_count_sub(PREEMPT_ACTIVE); 3806} 3807 3808int __sched _cond_resched(void) 3809{ 3810 if (should_resched()) { 3811 __cond_resched(); 3812 return 1; 3813 } 3814 return 0; 3815} 3816EXPORT_SYMBOL(_cond_resched); 3817 3818/* 3819 * __cond_resched_lock() - if a reschedule is pending, drop the given lock, 3820 * call schedule, and on return reacquire the lock. 3821 * 3822 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level 3823 * operations here to prevent schedule() from being called twice (once via 3824 * spin_unlock(), once by hand). 3825 */ 3826int __cond_resched_lock(spinlock_t *lock) 3827{ 3828 int resched = should_resched(); 3829 int ret = 0; 3830 3831 lockdep_assert_held(lock); 3832 3833 if (spin_needbreak(lock) || resched) { 3834 spin_unlock(lock); 3835 if (resched) 3836 __cond_resched(); 3837 else 3838 cpu_relax(); 3839 ret = 1; 3840 spin_lock(lock); 3841 } 3842 return ret; 3843} 3844EXPORT_SYMBOL(__cond_resched_lock); 3845 3846int __sched __cond_resched_softirq(void) 3847{ 3848 BUG_ON(!in_softirq()); 3849 3850 if (should_resched()) { 3851 local_bh_enable(); 3852 __cond_resched(); 3853 local_bh_disable(); 3854 return 1; 3855 } 3856 return 0; 3857} 3858EXPORT_SYMBOL(__cond_resched_softirq); 3859 3860/** 3861 * yield - yield the current processor to other threads. 3862 * 3863 * Do not ever use this function, there's a 99% chance you're doing it wrong. 3864 * 3865 * The scheduler is at all times free to pick the calling task as the most 3866 * eligible task to run, if removing the yield() call from your code breaks 3867 * it, its already broken. 3868 * 3869 * Typical broken usage is: 3870 * 3871 * while (!event) 3872 * yield(); 3873 * 3874 * where one assumes that yield() will let 'the other' process run that will 3875 * make event true. If the current task is a SCHED_FIFO task that will never 3876 * happen. Never use yield() as a progress guarantee!! 3877 * 3878 * If you want to use yield() to wait for something, use wait_event(). 3879 * If you want to use yield() to be 'nice' for others, use cond_resched(). 3880 * If you still want to use yield(), do not! 3881 */ 3882void __sched yield(void) 3883{ 3884 set_current_state(TASK_RUNNING); 3885 sys_sched_yield(); 3886} 3887EXPORT_SYMBOL(yield); 3888 3889/** 3890 * yield_to - yield the current processor to another thread in 3891 * your thread group, or accelerate that thread toward the 3892 * processor it's on. 3893 * @p: target task 3894 * @preempt: whether task preemption is allowed or not 3895 * 3896 * It's the caller's job to ensure that the target task struct 3897 * can't go away on us before we can do any checks. 3898 * 3899 * Return: 3900 * true (>0) if we indeed boosted the target task. 3901 * false (0) if we failed to boost the target. 3902 * -ESRCH if there's no task to yield to. 3903 */ 3904bool __sched yield_to(struct task_struct *p, bool preempt) 3905{ 3906 struct task_struct *curr = current; 3907 struct rq *rq, *p_rq; 3908 unsigned long flags; 3909 int yielded = 0; 3910 3911 local_irq_save(flags); 3912 rq = this_rq(); 3913 3914again: 3915 p_rq = task_rq(p); 3916 /* 3917 * If we're the only runnable task on the rq and target rq also 3918 * has only one task, there's absolutely no point in yielding. 3919 */ 3920 if (rq->nr_running == 1 && p_rq->nr_running == 1) { 3921 yielded = -ESRCH; 3922 goto out_irq; 3923 } 3924 3925 double_rq_lock(rq, p_rq); 3926 while (task_rq(p) != p_rq) { 3927 double_rq_unlock(rq, p_rq); 3928 goto again; 3929 } 3930 3931 if (!curr->sched_class->yield_to_task) 3932 goto out_unlock; 3933 3934 if (curr->sched_class != p->sched_class) 3935 goto out_unlock; 3936 3937 if (task_running(p_rq, p) || p->state) 3938 goto out_unlock; 3939 3940 yielded = curr->sched_class->yield_to_task(rq, p, preempt); 3941 if (yielded) { 3942 schedstat_inc(rq, yld_count); 3943 /* 3944 * Make p's CPU reschedule; pick_next_entity takes care of 3945 * fairness. 3946 */ 3947 if (preempt && rq != p_rq) 3948 resched_task(p_rq->curr); 3949 } 3950 3951out_unlock: 3952 double_rq_unlock(rq, p_rq); 3953out_irq: 3954 local_irq_restore(flags); 3955 3956 if (yielded > 0) 3957 schedule(); 3958 3959 return yielded; 3960} 3961EXPORT_SYMBOL_GPL(yield_to); 3962 3963/* 3964 * This task is about to go to sleep on IO. Increment rq->nr_iowait so 3965 * that process accounting knows that this is a task in IO wait state. 3966 */ 3967void __sched io_schedule(void) 3968{ 3969 struct rq *rq = raw_rq(); 3970 3971 delayacct_blkio_start(); 3972 atomic_inc(&rq->nr_iowait); 3973 blk_flush_plug(current); 3974 current->in_iowait = 1; 3975 schedule(); 3976 current->in_iowait = 0; 3977 atomic_dec(&rq->nr_iowait); 3978 delayacct_blkio_end(); 3979} 3980EXPORT_SYMBOL(io_schedule); 3981 3982long __sched io_schedule_timeout(long timeout) 3983{ 3984 struct rq *rq = raw_rq(); 3985 long ret; 3986 3987 delayacct_blkio_start(); 3988 atomic_inc(&rq->nr_iowait); 3989 blk_flush_plug(current); 3990 current->in_iowait = 1; 3991 ret = schedule_timeout(timeout); 3992 current->in_iowait = 0; 3993 atomic_dec(&rq->nr_iowait); 3994 delayacct_blkio_end(); 3995 return ret; 3996} 3997 3998/** 3999 * sys_sched_get_priority_max - return maximum RT priority. 4000 * @policy: scheduling class. 4001 * 4002 * Return: On success, this syscall returns the maximum 4003 * rt_priority that can be used by a given scheduling class. 4004 * On failure, a negative error code is returned. 4005 */ 4006SYSCALL_DEFINE1(sched_get_priority_max, int, policy) 4007{ 4008 int ret = -EINVAL; 4009 4010 switch (policy) { 4011 case SCHED_FIFO: 4012 case SCHED_RR: 4013 ret = MAX_USER_RT_PRIO-1; 4014 break; 4015 case SCHED_NORMAL: 4016 case SCHED_BATCH: 4017 case SCHED_IDLE: 4018 ret = 0; 4019 break; 4020 } 4021 return ret; 4022} 4023 4024/** 4025 * sys_sched_get_priority_min - return minimum RT priority. 4026 * @policy: scheduling class. 4027 * 4028 * Return: On success, this syscall returns the minimum 4029 * rt_priority that can be used by a given scheduling class. 4030 * On failure, a negative error code is returned. 4031 */ 4032SYSCALL_DEFINE1(sched_get_priority_min, int, policy) 4033{ 4034 int ret = -EINVAL; 4035 4036 switch (policy) { 4037 case SCHED_FIFO: 4038 case SCHED_RR: 4039 ret = 1; 4040 break; 4041 case SCHED_NORMAL: 4042 case SCHED_BATCH: 4043 case SCHED_IDLE: 4044 ret = 0; 4045 } 4046 return ret; 4047} 4048 4049/** 4050 * sys_sched_rr_get_interval - return the default timeslice of a process. 4051 * @pid: pid of the process. 4052 * @interval: userspace pointer to the timeslice value. 4053 * 4054 * this syscall writes the default timeslice value of a given process 4055 * into the user-space timespec buffer. A value of '0' means infinity. 4056 * 4057 * Return: On success, 0 and the timeslice is in @interval. Otherwise, 4058 * an error code. 4059 */ 4060SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid, 4061 struct timespec __user *, interval) 4062{ 4063 struct task_struct *p; 4064 unsigned int time_slice; 4065 unsigned long flags; 4066 struct rq *rq; 4067 int retval; 4068 struct timespec t; 4069 4070 if (pid < 0) 4071 return -EINVAL; 4072 4073 retval = -ESRCH; 4074 rcu_read_lock(); 4075 p = find_process_by_pid(pid); 4076 if (!p) 4077 goto out_unlock; 4078 4079 retval = security_task_getscheduler(p); 4080 if (retval) 4081 goto out_unlock; 4082 4083 rq = task_rq_lock(p, &flags); 4084 time_slice = p->sched_class->get_rr_interval(rq, p); 4085 task_rq_unlock(rq, p, &flags); 4086 4087 rcu_read_unlock(); 4088 jiffies_to_timespec(time_slice, &t); 4089 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0; 4090 return retval; 4091 4092out_unlock: 4093 rcu_read_unlock(); 4094 return retval; 4095} 4096 4097static const char stat_nam[] = TASK_STATE_TO_CHAR_STR; 4098 4099void sched_show_task(struct task_struct *p) 4100{ 4101 unsigned long free = 0; 4102 int ppid; 4103 unsigned state; 4104 4105 state = p->state ? __ffs(p->state) + 1 : 0; 4106 printk(KERN_INFO "%-15.15s %c", p->comm, 4107 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?'); 4108#if BITS_PER_LONG == 32 4109 if (state == TASK_RUNNING) 4110 printk(KERN_CONT " running "); 4111 else 4112 printk(KERN_CONT " %08lx ", thread_saved_pc(p)); 4113#else 4114 if (state == TASK_RUNNING) 4115 printk(KERN_CONT " running task "); 4116 else 4117 printk(KERN_CONT " %016lx ", thread_saved_pc(p)); 4118#endif 4119#ifdef CONFIG_DEBUG_STACK_USAGE 4120 free = stack_not_used(p); 4121#endif 4122 rcu_read_lock(); 4123 ppid = task_pid_nr(rcu_dereference(p->real_parent)); 4124 rcu_read_unlock(); 4125 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free, 4126 task_pid_nr(p), ppid, 4127 (unsigned long)task_thread_info(p)->flags); 4128 4129 print_worker_info(KERN_INFO, p); 4130 show_stack(p, NULL); 4131} 4132 4133void show_state_filter(unsigned long state_filter) 4134{ 4135 struct task_struct *g, *p; 4136 4137#if BITS_PER_LONG == 32 4138 printk(KERN_INFO 4139 " task PC stack pid father\n"); 4140#else 4141 printk(KERN_INFO 4142 " task PC stack pid father\n"); 4143#endif 4144 rcu_read_lock(); 4145 do_each_thread(g, p) { 4146 /* 4147 * reset the NMI-timeout, listing all files on a slow 4148 * console might take a lot of time: 4149 */ 4150 touch_nmi_watchdog(); 4151 if (!state_filter || (p->state & state_filter)) 4152 sched_show_task(p); 4153 } while_each_thread(g, p); 4154 4155 touch_all_softlockup_watchdogs(); 4156 4157#ifdef CONFIG_SCHED_DEBUG 4158 sysrq_sched_debug_show(); 4159#endif 4160 rcu_read_unlock(); 4161 /* 4162 * Only show locks if all tasks are dumped: 4163 */ 4164 if (!state_filter) 4165 debug_show_all_locks(); 4166} 4167 4168void init_idle_bootup_task(struct task_struct *idle) 4169{ 4170 idle->sched_class = &idle_sched_class; 4171} 4172 4173/** 4174 * init_idle - set up an idle thread for a given CPU 4175 * @idle: task in question 4176 * @cpu: cpu the idle task belongs to 4177 * 4178 * NOTE: this function does not set the idle thread's NEED_RESCHED 4179 * flag, to make booting more robust. 4180 */ 4181void init_idle(struct task_struct *idle, int cpu) 4182{ 4183 struct rq *rq = cpu_rq(cpu); 4184 unsigned long flags; 4185 4186 raw_spin_lock_irqsave(&rq->lock, flags); 4187 4188 __sched_fork(idle); 4189 idle->state = TASK_RUNNING; 4190 idle->se.exec_start = sched_clock(); 4191 4192 do_set_cpus_allowed(idle, cpumask_of(cpu)); 4193 /* 4194 * We're having a chicken and egg problem, even though we are 4195 * holding rq->lock, the cpu isn't yet set to this cpu so the 4196 * lockdep check in task_group() will fail. 4197 * 4198 * Similar case to sched_fork(). / Alternatively we could 4199 * use task_rq_lock() here and obtain the other rq->lock. 4200 * 4201 * Silence PROVE_RCU 4202 */ 4203 rcu_read_lock(); 4204 __set_task_cpu(idle, cpu); 4205 rcu_read_unlock(); 4206 4207 rq->curr = rq->idle = idle; 4208#if defined(CONFIG_SMP) 4209 idle->on_cpu = 1; 4210#endif 4211 raw_spin_unlock_irqrestore(&rq->lock, flags); 4212 4213 /* Set the preempt count _outside_ the spinlocks! */ 4214 init_idle_preempt_count(idle, cpu); 4215 4216 /* 4217 * The idle tasks have their own, simple scheduling class: 4218 */ 4219 idle->sched_class = &idle_sched_class; 4220 ftrace_graph_init_idle_task(idle, cpu); 4221 vtime_init_idle(idle, cpu); 4222#if defined(CONFIG_SMP) 4223 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu); 4224#endif 4225} 4226 4227#ifdef CONFIG_SMP 4228void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask) 4229{ 4230 if (p->sched_class && p->sched_class->set_cpus_allowed) 4231 p->sched_class->set_cpus_allowed(p, new_mask); 4232 4233 cpumask_copy(&p->cpus_allowed, new_mask); 4234 p->nr_cpus_allowed = cpumask_weight(new_mask); 4235} 4236 4237/* 4238 * This is how migration works: 4239 * 4240 * 1) we invoke migration_cpu_stop() on the target CPU using 4241 * stop_one_cpu(). 4242 * 2) stopper starts to run (implicitly forcing the migrated thread 4243 * off the CPU) 4244 * 3) it checks whether the migrated task is still in the wrong runqueue. 4245 * 4) if it's in the wrong runqueue then the migration thread removes 4246 * it and puts it into the right queue. 4247 * 5) stopper completes and stop_one_cpu() returns and the migration 4248 * is done. 4249 */ 4250 4251/* 4252 * Change a given task's CPU affinity. Migrate the thread to a 4253 * proper CPU and schedule it away if the CPU it's executing on 4254 * is removed from the allowed bitmask. 4255 * 4256 * NOTE: the caller must have a valid reference to the task, the 4257 * task must not exit() & deallocate itself prematurely. The 4258 * call is not atomic; no spinlocks may be held. 4259 */ 4260int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask) 4261{ 4262 unsigned long flags; 4263 struct rq *rq; 4264 unsigned int dest_cpu; 4265 int ret = 0; 4266 4267 rq = task_rq_lock(p, &flags); 4268 4269 if (cpumask_equal(&p->cpus_allowed, new_mask)) 4270 goto out; 4271 4272 if (!cpumask_intersects(new_mask, cpu_active_mask)) { 4273 ret = -EINVAL; 4274 goto out; 4275 } 4276 4277 do_set_cpus_allowed(p, new_mask); 4278 4279 /* Can the task run on the task's current CPU? If so, we're done */ 4280 if (cpumask_test_cpu(task_cpu(p), new_mask)) 4281 goto out; 4282 4283 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask); 4284 if (p->on_rq) { 4285 struct migration_arg arg = { p, dest_cpu }; 4286 /* Need help from migration thread: drop lock and wait. */ 4287 task_rq_unlock(rq, p, &flags); 4288 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg); 4289 tlb_migrate_finish(p->mm); 4290 return 0; 4291 } 4292out: 4293 task_rq_unlock(rq, p, &flags); 4294 4295 return ret; 4296} 4297EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr); 4298 4299/* 4300 * Move (not current) task off this cpu, onto dest cpu. We're doing 4301 * this because either it can't run here any more (set_cpus_allowed() 4302 * away from this CPU, or CPU going down), or because we're 4303 * attempting to rebalance this task on exec (sched_exec). 4304 * 4305 * So we race with normal scheduler movements, but that's OK, as long 4306 * as the task is no longer on this CPU. 4307 * 4308 * Returns non-zero if task was successfully migrated. 4309 */ 4310static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu) 4311{ 4312 struct rq *rq_dest, *rq_src; 4313 int ret = 0; 4314 4315 if (unlikely(!cpu_active(dest_cpu))) 4316 return ret; 4317 4318 rq_src = cpu_rq(src_cpu); 4319 rq_dest = cpu_rq(dest_cpu); 4320 4321 raw_spin_lock(&p->pi_lock); 4322 double_rq_lock(rq_src, rq_dest); 4323 /* Already moved. */ 4324 if (task_cpu(p) != src_cpu) 4325 goto done; 4326 /* Affinity changed (again). */ 4327 if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p))) 4328 goto fail; 4329 4330 /* 4331 * If we're not on a rq, the next wake-up will ensure we're 4332 * placed properly. 4333 */ 4334 if (p->on_rq) { 4335 dequeue_task(rq_src, p, 0); 4336 set_task_cpu(p, dest_cpu); 4337 enqueue_task(rq_dest, p, 0); 4338 check_preempt_curr(rq_dest, p, 0); 4339 } 4340done: 4341 ret = 1; 4342fail: 4343 double_rq_unlock(rq_src, rq_dest); 4344 raw_spin_unlock(&p->pi_lock); 4345 return ret; 4346} 4347 4348/* 4349 * migration_cpu_stop - this will be executed by a highprio stopper thread 4350 * and performs thread migration by bumping thread off CPU then 4351 * 'pushing' onto another runqueue. 4352 */ 4353static int migration_cpu_stop(void *data) 4354{ 4355 struct migration_arg *arg = data; 4356 4357 /* 4358 * The original target cpu might have gone down and we might 4359 * be on another cpu but it doesn't matter. 4360 */ 4361 local_irq_disable(); 4362 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu); 4363 local_irq_enable(); 4364 return 0; 4365} 4366 4367#ifdef CONFIG_HOTPLUG_CPU 4368 4369/* 4370 * Ensures that the idle task is using init_mm right before its cpu goes 4371 * offline. 4372 */ 4373void idle_task_exit(void) 4374{ 4375 struct mm_struct *mm = current->active_mm; 4376 4377 BUG_ON(cpu_online(smp_processor_id())); 4378 4379 if (mm != &init_mm) 4380 switch_mm(mm, &init_mm, current); 4381 mmdrop(mm); 4382} 4383 4384/* 4385 * Since this CPU is going 'away' for a while, fold any nr_active delta 4386 * we might have. Assumes we're called after migrate_tasks() so that the 4387 * nr_active count is stable. 4388 * 4389 * Also see the comment "Global load-average calculations". 4390 */ 4391static void calc_load_migrate(struct rq *rq) 4392{ 4393 long delta = calc_load_fold_active(rq); 4394 if (delta) 4395 atomic_long_add(delta, &calc_load_tasks); 4396} 4397 4398/* 4399 * Migrate all tasks from the rq, sleeping tasks will be migrated by 4400 * try_to_wake_up()->select_task_rq(). 4401 * 4402 * Called with rq->lock held even though we'er in stop_machine() and 4403 * there's no concurrency possible, we hold the required locks anyway 4404 * because of lock validation efforts. 4405 */ 4406static void migrate_tasks(unsigned int dead_cpu) 4407{ 4408 struct rq *rq = cpu_rq(dead_cpu); 4409 struct task_struct *next, *stop = rq->stop; 4410 int dest_cpu; 4411 4412 /* 4413 * Fudge the rq selection such that the below task selection loop 4414 * doesn't get stuck on the currently eligible stop task. 4415 * 4416 * We're currently inside stop_machine() and the rq is either stuck 4417 * in the stop_machine_cpu_stop() loop, or we're executing this code, 4418 * either way we should never end up calling schedule() until we're 4419 * done here. 4420 */ 4421 rq->stop = NULL; 4422 4423 /* 4424 * put_prev_task() and pick_next_task() sched 4425 * class method both need to have an up-to-date 4426 * value of rq->clock[_task] 4427 */ 4428 update_rq_clock(rq); 4429 4430 for ( ; ; ) { 4431 /* 4432 * There's this thread running, bail when that's the only 4433 * remaining thread. 4434 */ 4435 if (rq->nr_running == 1) 4436 break; 4437 4438 next = pick_next_task(rq); 4439 BUG_ON(!next); 4440 next->sched_class->put_prev_task(rq, next); 4441 4442 /* Find suitable destination for @next, with force if needed. */ 4443 dest_cpu = select_fallback_rq(dead_cpu, next); 4444 raw_spin_unlock(&rq->lock); 4445 4446 __migrate_task(next, dead_cpu, dest_cpu); 4447 4448 raw_spin_lock(&rq->lock); 4449 } 4450 4451 rq->stop = stop; 4452} 4453 4454#endif /* CONFIG_HOTPLUG_CPU */ 4455 4456#if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL) 4457 4458static struct ctl_table sd_ctl_dir[] = { 4459 { 4460 .procname = "sched_domain", 4461 .mode = 0555, 4462 }, 4463 {} 4464}; 4465 4466static struct ctl_table sd_ctl_root[] = { 4467 { 4468 .procname = "kernel", 4469 .mode = 0555, 4470 .child = sd_ctl_dir, 4471 }, 4472 {} 4473}; 4474 4475static struct ctl_table *sd_alloc_ctl_entry(int n) 4476{ 4477 struct ctl_table *entry = 4478 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL); 4479 4480 return entry; 4481} 4482 4483static void sd_free_ctl_entry(struct ctl_table **tablep) 4484{ 4485 struct ctl_table *entry; 4486 4487 /* 4488 * In the intermediate directories, both the child directory and 4489 * procname are dynamically allocated and could fail but the mode 4490 * will always be set. In the lowest directory the names are 4491 * static strings and all have proc handlers. 4492 */ 4493 for (entry = *tablep; entry->mode; entry++) { 4494 if (entry->child) 4495 sd_free_ctl_entry(&entry->child); 4496 if (entry->proc_handler == NULL) 4497 kfree(entry->procname); 4498 } 4499 4500 kfree(*tablep); 4501 *tablep = NULL; 4502} 4503 4504static int min_load_idx = 0; 4505static int max_load_idx = CPU_LOAD_IDX_MAX-1; 4506 4507static void 4508set_table_entry(struct ctl_table *entry, 4509 const char *procname, void *data, int maxlen, 4510 umode_t mode, proc_handler *proc_handler, 4511 bool load_idx) 4512{ 4513 entry->procname = procname; 4514 entry->data = data; 4515 entry->maxlen = maxlen; 4516 entry->mode = mode; 4517 entry->proc_handler = proc_handler; 4518 4519 if (load_idx) { 4520 entry->extra1 = &min_load_idx; 4521 entry->extra2 = &max_load_idx; 4522 } 4523} 4524 4525static struct ctl_table * 4526sd_alloc_ctl_domain_table(struct sched_domain *sd) 4527{ 4528 struct ctl_table *table = sd_alloc_ctl_entry(13); 4529 4530 if (table == NULL) 4531 return NULL; 4532 4533 set_table_entry(&table[0], "min_interval", &sd->min_interval, 4534 sizeof(long), 0644, proc_doulongvec_minmax, false); 4535 set_table_entry(&table[1], "max_interval", &sd->max_interval, 4536 sizeof(long), 0644, proc_doulongvec_minmax, false); 4537 set_table_entry(&table[2], "busy_idx", &sd->busy_idx, 4538 sizeof(int), 0644, proc_dointvec_minmax, true); 4539 set_table_entry(&table[3], "idle_idx", &sd->idle_idx, 4540 sizeof(int), 0644, proc_dointvec_minmax, true); 4541 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx, 4542 sizeof(int), 0644, proc_dointvec_minmax, true); 4543 set_table_entry(&table[5], "wake_idx", &sd->wake_idx, 4544 sizeof(int), 0644, proc_dointvec_minmax, true); 4545 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx, 4546 sizeof(int), 0644, proc_dointvec_minmax, true); 4547 set_table_entry(&table[7], "busy_factor", &sd->busy_factor, 4548 sizeof(int), 0644, proc_dointvec_minmax, false); 4549 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct, 4550 sizeof(int), 0644, proc_dointvec_minmax, false); 4551 set_table_entry(&table[9], "cache_nice_tries", 4552 &sd->cache_nice_tries, 4553 sizeof(int), 0644, proc_dointvec_minmax, false); 4554 set_table_entry(&table[10], "flags", &sd->flags, 4555 sizeof(int), 0644, proc_dointvec_minmax, false); 4556 set_table_entry(&table[11], "name", sd->name, 4557 CORENAME_MAX_SIZE, 0444, proc_dostring, false); 4558 /* &table[12] is terminator */ 4559 4560 return table; 4561} 4562 4563static struct ctl_table *sd_alloc_ctl_cpu_table(int cpu) 4564{ 4565 struct ctl_table *entry, *table; 4566 struct sched_domain *sd; 4567 int domain_num = 0, i; 4568 char buf[32]; 4569 4570 for_each_domain(cpu, sd) 4571 domain_num++; 4572 entry = table = sd_alloc_ctl_entry(domain_num + 1); 4573 if (table == NULL) 4574 return NULL; 4575 4576 i = 0; 4577 for_each_domain(cpu, sd) { 4578 snprintf(buf, 32, "domain%d", i); 4579 entry->procname = kstrdup(buf, GFP_KERNEL); 4580 entry->mode = 0555; 4581 entry->child = sd_alloc_ctl_domain_table(sd); 4582 entry++; 4583 i++; 4584 } 4585 return table; 4586} 4587 4588static struct ctl_table_header *sd_sysctl_header; 4589static void register_sched_domain_sysctl(void) 4590{ 4591 int i, cpu_num = num_possible_cpus(); 4592 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1); 4593 char buf[32]; 4594 4595 WARN_ON(sd_ctl_dir[0].child); 4596 sd_ctl_dir[0].child = entry; 4597 4598 if (entry == NULL) 4599 return; 4600 4601 for_each_possible_cpu(i) { 4602 snprintf(buf, 32, "cpu%d", i); 4603 entry->procname = kstrdup(buf, GFP_KERNEL); 4604 entry->mode = 0555; 4605 entry->child = sd_alloc_ctl_cpu_table(i); 4606 entry++; 4607 } 4608 4609 WARN_ON(sd_sysctl_header); 4610 sd_sysctl_header = register_sysctl_table(sd_ctl_root); 4611} 4612 4613/* may be called multiple times per register */ 4614static void unregister_sched_domain_sysctl(void) 4615{ 4616 if (sd_sysctl_header) 4617 unregister_sysctl_table(sd_sysctl_header); 4618 sd_sysctl_header = NULL; 4619 if (sd_ctl_dir[0].child) 4620 sd_free_ctl_entry(&sd_ctl_dir[0].child); 4621} 4622#else 4623static void register_sched_domain_sysctl(void) 4624{ 4625} 4626static void unregister_sched_domain_sysctl(void) 4627{ 4628} 4629#endif 4630 4631static void set_rq_online(struct rq *rq) 4632{ 4633 if (!rq->online) { 4634 const struct sched_class *class; 4635 4636 cpumask_set_cpu(rq->cpu, rq->rd->online); 4637 rq->online = 1; 4638 4639 for_each_class(class) { 4640 if (class->rq_online) 4641 class->rq_online(rq); 4642 } 4643 } 4644} 4645 4646static void set_rq_offline(struct rq *rq) 4647{ 4648 if (rq->online) { 4649 const struct sched_class *class; 4650 4651 for_each_class(class) { 4652 if (class->rq_offline) 4653 class->rq_offline(rq); 4654 } 4655 4656 cpumask_clear_cpu(rq->cpu, rq->rd->online); 4657 rq->online = 0; 4658 } 4659} 4660 4661/* 4662 * migration_call - callback that gets triggered when a CPU is added. 4663 * Here we can start up the necessary migration thread for the new CPU. 4664 */ 4665static int 4666migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu) 4667{ 4668 int cpu = (long)hcpu; 4669 unsigned long flags; 4670 struct rq *rq = cpu_rq(cpu); 4671 4672 switch (action & ~CPU_TASKS_FROZEN) { 4673 4674 case CPU_UP_PREPARE: 4675 rq->calc_load_update = calc_load_update; 4676 break; 4677 4678 case CPU_ONLINE: 4679 /* Update our root-domain */ 4680 raw_spin_lock_irqsave(&rq->lock, flags); 4681 if (rq->rd) { 4682 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span)); 4683 4684 set_rq_online(rq); 4685 } 4686 raw_spin_unlock_irqrestore(&rq->lock, flags); 4687 break; 4688 4689#ifdef CONFIG_HOTPLUG_CPU 4690 case CPU_DYING: 4691 sched_ttwu_pending(); 4692 /* Update our root-domain */ 4693 raw_spin_lock_irqsave(&rq->lock, flags); 4694 if (rq->rd) { 4695 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span)); 4696 set_rq_offline(rq); 4697 } 4698 migrate_tasks(cpu); 4699 BUG_ON(rq->nr_running != 1); /* the migration thread */ 4700 raw_spin_unlock_irqrestore(&rq->lock, flags); 4701 break; 4702 4703 case CPU_DEAD: 4704 calc_load_migrate(rq); 4705 break; 4706#endif 4707 } 4708 4709 update_max_interval(); 4710 4711 return NOTIFY_OK; 4712} 4713 4714/* 4715 * Register at high priority so that task migration (migrate_all_tasks) 4716 * happens before everything else. This has to be lower priority than 4717 * the notifier in the perf_event subsystem, though. 4718 */ 4719static struct notifier_block migration_notifier = { 4720 .notifier_call = migration_call, 4721 .priority = CPU_PRI_MIGRATION, 4722}; 4723 4724static int sched_cpu_active(struct notifier_block *nfb, 4725 unsigned long action, void *hcpu) 4726{ 4727 switch (action & ~CPU_TASKS_FROZEN) { 4728 case CPU_STARTING: 4729 case CPU_DOWN_FAILED: 4730 set_cpu_active((long)hcpu, true); 4731 return NOTIFY_OK; 4732 default: 4733 return NOTIFY_DONE; 4734 } 4735} 4736 4737static int sched_cpu_inactive(struct notifier_block *nfb, 4738 unsigned long action, void *hcpu) 4739{ 4740 switch (action & ~CPU_TASKS_FROZEN) { 4741 case CPU_DOWN_PREPARE: 4742 set_cpu_active((long)hcpu, false); 4743 return NOTIFY_OK; 4744 default: 4745 return NOTIFY_DONE; 4746 } 4747} 4748 4749static int __init migration_init(void) 4750{ 4751 void *cpu = (void *)(long)smp_processor_id(); 4752 int err; 4753 4754 /* Initialize migration for the boot CPU */ 4755 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu); 4756 BUG_ON(err == NOTIFY_BAD); 4757 migration_call(&migration_notifier, CPU_ONLINE, cpu); 4758 register_cpu_notifier(&migration_notifier); 4759 4760 /* Register cpu active notifiers */ 4761 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE); 4762 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE); 4763 4764 return 0; 4765} 4766early_initcall(migration_init); 4767#endif 4768 4769#ifdef CONFIG_SMP 4770 4771static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */ 4772 4773#ifdef CONFIG_SCHED_DEBUG 4774 4775static __read_mostly int sched_debug_enabled; 4776 4777static int __init sched_debug_setup(char *str) 4778{ 4779 sched_debug_enabled = 1; 4780 4781 return 0; 4782} 4783early_param("sched_debug", sched_debug_setup); 4784 4785static inline bool sched_debug(void) 4786{ 4787 return sched_debug_enabled; 4788} 4789 4790static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level, 4791 struct cpumask *groupmask) 4792{ 4793 struct sched_group *group = sd->groups; 4794 char str[256]; 4795 4796 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd)); 4797 cpumask_clear(groupmask); 4798 4799 printk(KERN_DEBUG "%*s domain %d: ", level, "", level); 4800 4801 if (!(sd->flags & SD_LOAD_BALANCE)) { 4802 printk("does not load-balance\n"); 4803 if (sd->parent) 4804 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain" 4805 " has parent"); 4806 return -1; 4807 } 4808 4809 printk(KERN_CONT "span %s level %s\n", str, sd->name); 4810 4811 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) { 4812 printk(KERN_ERR "ERROR: domain->span does not contain " 4813 "CPU%d\n", cpu); 4814 } 4815 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) { 4816 printk(KERN_ERR "ERROR: domain->groups does not contain" 4817 " CPU%d\n", cpu); 4818 } 4819 4820 printk(KERN_DEBUG "%*s groups:", level + 1, ""); 4821 do { 4822 if (!group) { 4823 printk("\n"); 4824 printk(KERN_ERR "ERROR: group is NULL\n"); 4825 break; 4826 } 4827 4828 /* 4829 * Even though we initialize ->power to something semi-sane, 4830 * we leave power_orig unset. This allows us to detect if 4831 * domain iteration is still funny without causing /0 traps. 4832 */ 4833 if (!group->sgp->power_orig) { 4834 printk(KERN_CONT "\n"); 4835 printk(KERN_ERR "ERROR: domain->cpu_power not " 4836 "set\n"); 4837 break; 4838 } 4839 4840 if (!cpumask_weight(sched_group_cpus(group))) { 4841 printk(KERN_CONT "\n"); 4842 printk(KERN_ERR "ERROR: empty group\n"); 4843 break; 4844 } 4845 4846 if (!(sd->flags & SD_OVERLAP) && 4847 cpumask_intersects(groupmask, sched_group_cpus(group))) { 4848 printk(KERN_CONT "\n"); 4849 printk(KERN_ERR "ERROR: repeated CPUs\n"); 4850 break; 4851 } 4852 4853 cpumask_or(groupmask, groupmask, sched_group_cpus(group)); 4854 4855 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group)); 4856 4857 printk(KERN_CONT " %s", str); 4858 if (group->sgp->power != SCHED_POWER_SCALE) { 4859 printk(KERN_CONT " (cpu_power = %d)", 4860 group->sgp->power); 4861 } 4862 4863 group = group->next; 4864 } while (group != sd->groups); 4865 printk(KERN_CONT "\n"); 4866 4867 if (!cpumask_equal(sched_domain_span(sd), groupmask)) 4868 printk(KERN_ERR "ERROR: groups don't span domain->span\n"); 4869 4870 if (sd->parent && 4871 !cpumask_subset(groupmask, sched_domain_span(sd->parent))) 4872 printk(KERN_ERR "ERROR: parent span is not a superset " 4873 "of domain->span\n"); 4874 return 0; 4875} 4876 4877static void sched_domain_debug(struct sched_domain *sd, int cpu) 4878{ 4879 int level = 0; 4880 4881 if (!sched_debug_enabled) 4882 return; 4883 4884 if (!sd) { 4885 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu); 4886 return; 4887 } 4888 4889 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu); 4890 4891 for (;;) { 4892 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask)) 4893 break; 4894 level++; 4895 sd = sd->parent; 4896 if (!sd) 4897 break; 4898 } 4899} 4900#else /* !CONFIG_SCHED_DEBUG */ 4901# define sched_domain_debug(sd, cpu) do { } while (0) 4902static inline bool sched_debug(void) 4903{ 4904 return false; 4905} 4906#endif /* CONFIG_SCHED_DEBUG */ 4907 4908static int sd_degenerate(struct sched_domain *sd) 4909{ 4910 if (cpumask_weight(sched_domain_span(sd)) == 1) 4911 return 1; 4912 4913 /* Following flags need at least 2 groups */ 4914 if (sd->flags & (SD_LOAD_BALANCE | 4915 SD_BALANCE_NEWIDLE | 4916 SD_BALANCE_FORK | 4917 SD_BALANCE_EXEC | 4918 SD_SHARE_CPUPOWER | 4919 SD_SHARE_PKG_RESOURCES)) { 4920 if (sd->groups != sd->groups->next) 4921 return 0; 4922 } 4923 4924 /* Following flags don't use groups */ 4925 if (sd->flags & (SD_WAKE_AFFINE)) 4926 return 0; 4927 4928 return 1; 4929} 4930 4931static int 4932sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent) 4933{ 4934 unsigned long cflags = sd->flags, pflags = parent->flags; 4935 4936 if (sd_degenerate(parent)) 4937 return 1; 4938 4939 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent))) 4940 return 0; 4941 4942 /* Flags needing groups don't count if only 1 group in parent */ 4943 if (parent->groups == parent->groups->next) { 4944 pflags &= ~(SD_LOAD_BALANCE | 4945 SD_BALANCE_NEWIDLE | 4946 SD_BALANCE_FORK | 4947 SD_BALANCE_EXEC | 4948 SD_SHARE_CPUPOWER | 4949 SD_SHARE_PKG_RESOURCES | 4950 SD_PREFER_SIBLING); 4951 if (nr_node_ids == 1) 4952 pflags &= ~SD_SERIALIZE; 4953 } 4954 if (~cflags & pflags) 4955 return 0; 4956 4957 return 1; 4958} 4959 4960static void free_rootdomain(struct rcu_head *rcu) 4961{ 4962 struct root_domain *rd = container_of(rcu, struct root_domain, rcu); 4963 4964 cpupri_cleanup(&rd->cpupri); 4965 free_cpumask_var(rd->rto_mask); 4966 free_cpumask_var(rd->online); 4967 free_cpumask_var(rd->span); 4968 kfree(rd); 4969} 4970 4971static void rq_attach_root(struct rq *rq, struct root_domain *rd) 4972{ 4973 struct root_domain *old_rd = NULL; 4974 unsigned long flags; 4975 4976 raw_spin_lock_irqsave(&rq->lock, flags); 4977 4978 if (rq->rd) { 4979 old_rd = rq->rd; 4980 4981 if (cpumask_test_cpu(rq->cpu, old_rd->online)) 4982 set_rq_offline(rq); 4983 4984 cpumask_clear_cpu(rq->cpu, old_rd->span); 4985 4986 /* 4987 * If we dont want to free the old_rt yet then 4988 * set old_rd to NULL to skip the freeing later 4989 * in this function: 4990 */ 4991 if (!atomic_dec_and_test(&old_rd->refcount)) 4992 old_rd = NULL; 4993 } 4994 4995 atomic_inc(&rd->refcount); 4996 rq->rd = rd; 4997 4998 cpumask_set_cpu(rq->cpu, rd->span); 4999 if (cpumask_test_cpu(rq->cpu, cpu_active_mask)) 5000 set_rq_online(rq); 5001 5002 raw_spin_unlock_irqrestore(&rq->lock, flags); 5003 5004 if (old_rd) 5005 call_rcu_sched(&old_rd->rcu, free_rootdomain); 5006} 5007 5008static int init_rootdomain(struct root_domain *rd) 5009{ 5010 memset(rd, 0, sizeof(*rd)); 5011 5012 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL)) 5013 goto out; 5014 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL)) 5015 goto free_span; 5016 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL)) 5017 goto free_online; 5018 5019 if (cpupri_init(&rd->cpupri) != 0) 5020 goto free_rto_mask; 5021 return 0; 5022 5023free_rto_mask: 5024 free_cpumask_var(rd->rto_mask); 5025free_online: 5026 free_cpumask_var(rd->online); 5027free_span: 5028 free_cpumask_var(rd->span); 5029out: 5030 return -ENOMEM; 5031} 5032 5033/* 5034 * By default the system creates a single root-domain with all cpus as 5035 * members (mimicking the global state we have today). 5036 */ 5037struct root_domain def_root_domain; 5038 5039static void init_defrootdomain(void) 5040{ 5041 init_rootdomain(&def_root_domain); 5042 5043 atomic_set(&def_root_domain.refcount, 1); 5044} 5045 5046static struct root_domain *alloc_rootdomain(void) 5047{ 5048 struct root_domain *rd; 5049 5050 rd = kmalloc(sizeof(*rd), GFP_KERNEL); 5051 if (!rd) 5052 return NULL; 5053 5054 if (init_rootdomain(rd) != 0) { 5055 kfree(rd); 5056 return NULL; 5057 } 5058 5059 return rd; 5060} 5061 5062static void free_sched_groups(struct sched_group *sg, int free_sgp) 5063{ 5064 struct sched_group *tmp, *first; 5065 5066 if (!sg) 5067 return; 5068 5069 first = sg; 5070 do { 5071 tmp = sg->next; 5072 5073 if (free_sgp && atomic_dec_and_test(&sg->sgp->ref)) 5074 kfree(sg->sgp); 5075 5076 kfree(sg); 5077 sg = tmp; 5078 } while (sg != first); 5079} 5080 5081static void free_sched_domain(struct rcu_head *rcu) 5082{ 5083 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu); 5084 5085 /* 5086 * If its an overlapping domain it has private groups, iterate and 5087 * nuke them all. 5088 */ 5089 if (sd->flags & SD_OVERLAP) { 5090 free_sched_groups(sd->groups, 1); 5091 } else if (atomic_dec_and_test(&sd->groups->ref)) { 5092 kfree(sd->groups->sgp); 5093 kfree(sd->groups); 5094 } 5095 kfree(sd); 5096} 5097 5098static void destroy_sched_domain(struct sched_domain *sd, int cpu) 5099{ 5100 call_rcu(&sd->rcu, free_sched_domain); 5101} 5102 5103static void destroy_sched_domains(struct sched_domain *sd, int cpu) 5104{ 5105 for (; sd; sd = sd->parent) 5106 destroy_sched_domain(sd, cpu); 5107} 5108 5109/* 5110 * Keep a special pointer to the highest sched_domain that has 5111 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this 5112 * allows us to avoid some pointer chasing select_idle_sibling(). 5113 * 5114 * Also keep a unique ID per domain (we use the first cpu number in 5115 * the cpumask of the domain), this allows us to quickly tell if 5116 * two cpus are in the same cache domain, see cpus_share_cache(). 5117 */ 5118DEFINE_PER_CPU(struct sched_domain *, sd_llc); 5119DEFINE_PER_CPU(int, sd_llc_size); 5120DEFINE_PER_CPU(int, sd_llc_id); 5121 5122static void update_top_cache_domain(int cpu) 5123{ 5124 struct sched_domain *sd; 5125 int id = cpu; 5126 int size = 1; 5127 5128 sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES); 5129 if (sd) { 5130 id = cpumask_first(sched_domain_span(sd)); 5131 size = cpumask_weight(sched_domain_span(sd)); 5132 } 5133 5134 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd); 5135 per_cpu(sd_llc_size, cpu) = size; 5136 per_cpu(sd_llc_id, cpu) = id; 5137} 5138 5139/* 5140 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must 5141 * hold the hotplug lock. 5142 */ 5143static void 5144cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu) 5145{ 5146 struct rq *rq = cpu_rq(cpu); 5147 struct sched_domain *tmp; 5148 5149 /* Remove the sched domains which do not contribute to scheduling. */ 5150 for (tmp = sd; tmp; ) { 5151 struct sched_domain *parent = tmp->parent; 5152 if (!parent) 5153 break; 5154 5155 if (sd_parent_degenerate(tmp, parent)) { 5156 tmp->parent = parent->parent; 5157 if (parent->parent) 5158 parent->parent->child = tmp; 5159 /* 5160 * Transfer SD_PREFER_SIBLING down in case of a 5161 * degenerate parent; the spans match for this 5162 * so the property transfers. 5163 */ 5164 if (parent->flags & SD_PREFER_SIBLING) 5165 tmp->flags |= SD_PREFER_SIBLING; 5166 destroy_sched_domain(parent, cpu); 5167 } else 5168 tmp = tmp->parent; 5169 } 5170 5171 if (sd && sd_degenerate(sd)) { 5172 tmp = sd; 5173 sd = sd->parent; 5174 destroy_sched_domain(tmp, cpu); 5175 if (sd) 5176 sd->child = NULL; 5177 } 5178 5179 sched_domain_debug(sd, cpu); 5180 5181 rq_attach_root(rq, rd); 5182 tmp = rq->sd; 5183 rcu_assign_pointer(rq->sd, sd); 5184 destroy_sched_domains(tmp, cpu); 5185 5186 update_top_cache_domain(cpu); 5187} 5188 5189/* cpus with isolated domains */ 5190static cpumask_var_t cpu_isolated_map; 5191 5192/* Setup the mask of cpus configured for isolated domains */ 5193static int __init isolated_cpu_setup(char *str) 5194{ 5195 alloc_bootmem_cpumask_var(&cpu_isolated_map); 5196 cpulist_parse(str, cpu_isolated_map); 5197 return 1; 5198} 5199 5200__setup("isolcpus=", isolated_cpu_setup); 5201 5202static const struct cpumask *cpu_cpu_mask(int cpu) 5203{ 5204 return cpumask_of_node(cpu_to_node(cpu)); 5205} 5206 5207struct sd_data { 5208 struct sched_domain **__percpu sd; 5209 struct sched_group **__percpu sg; 5210 struct sched_group_power **__percpu sgp; 5211}; 5212 5213struct s_data { 5214 struct sched_domain ** __percpu sd; 5215 struct root_domain *rd; 5216}; 5217 5218enum s_alloc { 5219 sa_rootdomain, 5220 sa_sd, 5221 sa_sd_storage, 5222 sa_none, 5223}; 5224 5225struct sched_domain_topology_level; 5226 5227typedef struct sched_domain *(*sched_domain_init_f)(struct sched_domain_topology_level *tl, int cpu); 5228typedef const struct cpumask *(*sched_domain_mask_f)(int cpu); 5229 5230#define SDTL_OVERLAP 0x01 5231 5232struct sched_domain_topology_level { 5233 sched_domain_init_f init; 5234 sched_domain_mask_f mask; 5235 int flags; 5236 int numa_level; 5237 struct sd_data data; 5238}; 5239 5240/* 5241 * Build an iteration mask that can exclude certain CPUs from the upwards 5242 * domain traversal. 5243 * 5244 * Asymmetric node setups can result in situations where the domain tree is of 5245 * unequal depth, make sure to skip domains that already cover the entire 5246 * range. 5247 * 5248 * In that case build_sched_domains() will have terminated the iteration early 5249 * and our sibling sd spans will be empty. Domains should always include the 5250 * cpu they're built on, so check that. 5251 * 5252 */ 5253static void build_group_mask(struct sched_domain *sd, struct sched_group *sg) 5254{ 5255 const struct cpumask *span = sched_domain_span(sd); 5256 struct sd_data *sdd = sd->private; 5257 struct sched_domain *sibling; 5258 int i; 5259 5260 for_each_cpu(i, span) { 5261 sibling = *per_cpu_ptr(sdd->sd, i); 5262 if (!cpumask_test_cpu(i, sched_domain_span(sibling))) 5263 continue; 5264 5265 cpumask_set_cpu(i, sched_group_mask(sg)); 5266 } 5267} 5268 5269/* 5270 * Return the canonical balance cpu for this group, this is the first cpu 5271 * of this group that's also in the iteration mask. 5272 */ 5273int group_balance_cpu(struct sched_group *sg) 5274{ 5275 return cpumask_first_and(sched_group_cpus(sg), sched_group_mask(sg)); 5276} 5277 5278static int 5279build_overlap_sched_groups(struct sched_domain *sd, int cpu) 5280{ 5281 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg; 5282 const struct cpumask *span = sched_domain_span(sd); 5283 struct cpumask *covered = sched_domains_tmpmask; 5284 struct sd_data *sdd = sd->private; 5285 struct sched_domain *child; 5286 int i; 5287 5288 cpumask_clear(covered); 5289 5290 for_each_cpu(i, span) { 5291 struct cpumask *sg_span; 5292 5293 if (cpumask_test_cpu(i, covered)) 5294 continue; 5295 5296 child = *per_cpu_ptr(sdd->sd, i); 5297 5298 /* See the comment near build_group_mask(). */ 5299 if (!cpumask_test_cpu(i, sched_domain_span(child))) 5300 continue; 5301 5302 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(), 5303 GFP_KERNEL, cpu_to_node(cpu)); 5304 5305 if (!sg) 5306 goto fail; 5307 5308 sg_span = sched_group_cpus(sg); 5309 if (child->child) { 5310 child = child->child; 5311 cpumask_copy(sg_span, sched_domain_span(child)); 5312 } else 5313 cpumask_set_cpu(i, sg_span); 5314 5315 cpumask_or(covered, covered, sg_span); 5316 5317 sg->sgp = *per_cpu_ptr(sdd->sgp, i); 5318 if (atomic_inc_return(&sg->sgp->ref) == 1) 5319 build_group_mask(sd, sg); 5320 5321 /* 5322 * Initialize sgp->power such that even if we mess up the 5323 * domains and no possible iteration will get us here, we won't 5324 * die on a /0 trap. 5325 */ 5326 sg->sgp->power = SCHED_POWER_SCALE * cpumask_weight(sg_span); 5327 5328 /* 5329 * Make sure the first group of this domain contains the 5330 * canonical balance cpu. Otherwise the sched_domain iteration 5331 * breaks. See update_sg_lb_stats(). 5332 */ 5333 if ((!groups && cpumask_test_cpu(cpu, sg_span)) || 5334 group_balance_cpu(sg) == cpu) 5335 groups = sg; 5336 5337 if (!first) 5338 first = sg; 5339 if (last) 5340 last->next = sg; 5341 last = sg; 5342 last->next = first; 5343 } 5344 sd->groups = groups; 5345 5346 return 0; 5347 5348fail: 5349 free_sched_groups(first, 0); 5350 5351 return -ENOMEM; 5352} 5353 5354static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg) 5355{ 5356 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu); 5357 struct sched_domain *child = sd->child; 5358 5359 if (child) 5360 cpu = cpumask_first(sched_domain_span(child)); 5361 5362 if (sg) { 5363 *sg = *per_cpu_ptr(sdd->sg, cpu); 5364 (*sg)->sgp = *per_cpu_ptr(sdd->sgp, cpu); 5365 atomic_set(&(*sg)->sgp->ref, 1); /* for claim_allocations */ 5366 } 5367 5368 return cpu; 5369} 5370 5371/* 5372 * build_sched_groups will build a circular linked list of the groups 5373 * covered by the given span, and will set each group's ->cpumask correctly, 5374 * and ->cpu_power to 0. 5375 * 5376 * Assumes the sched_domain tree is fully constructed 5377 */ 5378static int 5379build_sched_groups(struct sched_domain *sd, int cpu) 5380{ 5381 struct sched_group *first = NULL, *last = NULL; 5382 struct sd_data *sdd = sd->private; 5383 const struct cpumask *span = sched_domain_span(sd); 5384 struct cpumask *covered; 5385 int i; 5386 5387 get_group(cpu, sdd, &sd->groups); 5388 atomic_inc(&sd->groups->ref); 5389 5390 if (cpu != cpumask_first(span)) 5391 return 0; 5392 5393 lockdep_assert_held(&sched_domains_mutex); 5394 covered = sched_domains_tmpmask; 5395 5396 cpumask_clear(covered); 5397 5398 for_each_cpu(i, span) { 5399 struct sched_group *sg; 5400 int group, j; 5401 5402 if (cpumask_test_cpu(i, covered)) 5403 continue; 5404 5405 group = get_group(i, sdd, &sg); 5406 cpumask_clear(sched_group_cpus(sg)); 5407 sg->sgp->power = 0; 5408 cpumask_setall(sched_group_mask(sg)); 5409 5410 for_each_cpu(j, span) { 5411 if (get_group(j, sdd, NULL) != group) 5412 continue; 5413 5414 cpumask_set_cpu(j, covered); 5415 cpumask_set_cpu(j, sched_group_cpus(sg)); 5416 } 5417 5418 if (!first) 5419 first = sg; 5420 if (last) 5421 last->next = sg; 5422 last = sg; 5423 } 5424 last->next = first; 5425 5426 return 0; 5427} 5428 5429/* 5430 * Initialize sched groups cpu_power. 5431 * 5432 * cpu_power indicates the capacity of sched group, which is used while 5433 * distributing the load between different sched groups in a sched domain. 5434 * Typically cpu_power for all the groups in a sched domain will be same unless 5435 * there are asymmetries in the topology. If there are asymmetries, group 5436 * having more cpu_power will pickup more load compared to the group having 5437 * less cpu_power. 5438 */ 5439static void init_sched_groups_power(int cpu, struct sched_domain *sd) 5440{ 5441 struct sched_group *sg = sd->groups; 5442 5443 WARN_ON(!sg); 5444 5445 do { 5446 sg->group_weight = cpumask_weight(sched_group_cpus(sg)); 5447 sg = sg->next; 5448 } while (sg != sd->groups); 5449 5450 if (cpu != group_balance_cpu(sg)) 5451 return; 5452 5453 update_group_power(sd, cpu); 5454 atomic_set(&sg->sgp->nr_busy_cpus, sg->group_weight); 5455} 5456 5457int __weak arch_sd_sibling_asym_packing(void) 5458{ 5459 return 0*SD_ASYM_PACKING; 5460} 5461 5462/* 5463 * Initializers for schedule domains 5464 * Non-inlined to reduce accumulated stack pressure in build_sched_domains() 5465 */ 5466 5467#ifdef CONFIG_SCHED_DEBUG 5468# define SD_INIT_NAME(sd, type) sd->name = #type 5469#else 5470# define SD_INIT_NAME(sd, type) do { } while (0) 5471#endif 5472 5473#define SD_INIT_FUNC(type) \ 5474static noinline struct sched_domain * \ 5475sd_init_##type(struct sched_domain_topology_level *tl, int cpu) \ 5476{ \ 5477 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); \ 5478 *sd = SD_##type##_INIT; \ 5479 SD_INIT_NAME(sd, type); \ 5480 sd->private = &tl->data; \ 5481 return sd; \ 5482} 5483 5484SD_INIT_FUNC(CPU) 5485#ifdef CONFIG_SCHED_SMT 5486 SD_INIT_FUNC(SIBLING) 5487#endif 5488#ifdef CONFIG_SCHED_MC 5489 SD_INIT_FUNC(MC) 5490#endif 5491#ifdef CONFIG_SCHED_BOOK 5492 SD_INIT_FUNC(BOOK) 5493#endif 5494 5495static int default_relax_domain_level = -1; 5496int sched_domain_level_max; 5497 5498static int __init setup_relax_domain_level(char *str) 5499{ 5500 if (kstrtoint(str, 0, &default_relax_domain_level)) 5501 pr_warn("Unable to set relax_domain_level\n"); 5502 5503 return 1; 5504} 5505__setup("relax_domain_level=", setup_relax_domain_level); 5506 5507static void set_domain_attribute(struct sched_domain *sd, 5508 struct sched_domain_attr *attr) 5509{ 5510 int request; 5511 5512 if (!attr || attr->relax_domain_level < 0) { 5513 if (default_relax_domain_level < 0) 5514 return; 5515 else 5516 request = default_relax_domain_level; 5517 } else 5518 request = attr->relax_domain_level; 5519 if (request < sd->level) { 5520 /* turn off idle balance on this domain */ 5521 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE); 5522 } else { 5523 /* turn on idle balance on this domain */ 5524 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE); 5525 } 5526} 5527 5528static void __sdt_free(const struct cpumask *cpu_map); 5529static int __sdt_alloc(const struct cpumask *cpu_map); 5530 5531static void __free_domain_allocs(struct s_data *d, enum s_alloc what, 5532 const struct cpumask *cpu_map) 5533{ 5534 switch (what) { 5535 case sa_rootdomain: 5536 if (!atomic_read(&d->rd->refcount)) 5537 free_rootdomain(&d->rd->rcu); /* fall through */ 5538 case sa_sd: 5539 free_percpu(d->sd); /* fall through */ 5540 case sa_sd_storage: 5541 __sdt_free(cpu_map); /* fall through */ 5542 case sa_none: 5543 break; 5544 } 5545} 5546 5547static enum s_alloc __visit_domain_allocation_hell(struct s_data *d, 5548 const struct cpumask *cpu_map) 5549{ 5550 memset(d, 0, sizeof(*d)); 5551 5552 if (__sdt_alloc(cpu_map)) 5553 return sa_sd_storage; 5554 d->sd = alloc_percpu(struct sched_domain *); 5555 if (!d->sd) 5556 return sa_sd_storage; 5557 d->rd = alloc_rootdomain(); 5558 if (!d->rd) 5559 return sa_sd; 5560 return sa_rootdomain; 5561} 5562 5563/* 5564 * NULL the sd_data elements we've used to build the sched_domain and 5565 * sched_group structure so that the subsequent __free_domain_allocs() 5566 * will not free the data we're using. 5567 */ 5568static void claim_allocations(int cpu, struct sched_domain *sd) 5569{ 5570 struct sd_data *sdd = sd->private; 5571 5572 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd); 5573 *per_cpu_ptr(sdd->sd, cpu) = NULL; 5574 5575 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref)) 5576 *per_cpu_ptr(sdd->sg, cpu) = NULL; 5577 5578 if (atomic_read(&(*per_cpu_ptr(sdd->sgp, cpu))->ref)) 5579 *per_cpu_ptr(sdd->sgp, cpu) = NULL; 5580} 5581 5582#ifdef CONFIG_SCHED_SMT 5583static const struct cpumask *cpu_smt_mask(int cpu) 5584{ 5585 return topology_thread_cpumask(cpu); 5586} 5587#endif 5588 5589/* 5590 * Topology list, bottom-up. 5591 */ 5592static struct sched_domain_topology_level default_topology[] = { 5593#ifdef CONFIG_SCHED_SMT 5594 { sd_init_SIBLING, cpu_smt_mask, }, 5595#endif 5596#ifdef CONFIG_SCHED_MC 5597 { sd_init_MC, cpu_coregroup_mask, }, 5598#endif 5599#ifdef CONFIG_SCHED_BOOK 5600 { sd_init_BOOK, cpu_book_mask, }, 5601#endif 5602 { sd_init_CPU, cpu_cpu_mask, }, 5603 { NULL, }, 5604}; 5605 5606static struct sched_domain_topology_level *sched_domain_topology = default_topology; 5607 5608#define for_each_sd_topology(tl) \ 5609 for (tl = sched_domain_topology; tl->init; tl++) 5610 5611#ifdef CONFIG_NUMA 5612 5613static int sched_domains_numa_levels; 5614static int *sched_domains_numa_distance; 5615static struct cpumask ***sched_domains_numa_masks; 5616static int sched_domains_curr_level; 5617 5618static inline int sd_local_flags(int level) 5619{ 5620 if (sched_domains_numa_distance[level] > RECLAIM_DISTANCE) 5621 return 0; 5622 5623 return SD_BALANCE_EXEC | SD_BALANCE_FORK | SD_WAKE_AFFINE; 5624} 5625 5626static struct sched_domain * 5627sd_numa_init(struct sched_domain_topology_level *tl, int cpu) 5628{ 5629 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); 5630 int level = tl->numa_level; 5631 int sd_weight = cpumask_weight( 5632 sched_domains_numa_masks[level][cpu_to_node(cpu)]); 5633 5634 *sd = (struct sched_domain){ 5635 .min_interval = sd_weight, 5636 .max_interval = 2*sd_weight, 5637 .busy_factor = 32, 5638 .imbalance_pct = 125, 5639 .cache_nice_tries = 2, 5640 .busy_idx = 3, 5641 .idle_idx = 2, 5642 .newidle_idx = 0, 5643 .wake_idx = 0, 5644 .forkexec_idx = 0, 5645 5646 .flags = 1*SD_LOAD_BALANCE 5647 | 1*SD_BALANCE_NEWIDLE 5648 | 0*SD_BALANCE_EXEC 5649 | 0*SD_BALANCE_FORK 5650 | 0*SD_BALANCE_WAKE 5651 | 0*SD_WAKE_AFFINE 5652 | 0*SD_SHARE_CPUPOWER 5653 | 0*SD_SHARE_PKG_RESOURCES 5654 | 1*SD_SERIALIZE 5655 | 0*SD_PREFER_SIBLING 5656 | sd_local_flags(level) 5657 , 5658 .last_balance = jiffies, 5659 .balance_interval = sd_weight, 5660 }; 5661 SD_INIT_NAME(sd, NUMA); 5662 sd->private = &tl->data; 5663 5664 /* 5665 * Ugly hack to pass state to sd_numa_mask()... 5666 */ 5667 sched_domains_curr_level = tl->numa_level; 5668 5669 return sd; 5670} 5671 5672static const struct cpumask *sd_numa_mask(int cpu) 5673{ 5674 return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)]; 5675} 5676 5677static void sched_numa_warn(const char *str) 5678{ 5679 static int done = false; 5680 int i,j; 5681 5682 if (done) 5683 return; 5684 5685 done = true; 5686 5687 printk(KERN_WARNING "ERROR: %s\n\n", str); 5688 5689 for (i = 0; i < nr_node_ids; i++) { 5690 printk(KERN_WARNING " "); 5691 for (j = 0; j < nr_node_ids; j++) 5692 printk(KERN_CONT "%02d ", node_distance(i,j)); 5693 printk(KERN_CONT "\n"); 5694 } 5695 printk(KERN_WARNING "\n"); 5696} 5697 5698static bool find_numa_distance(int distance) 5699{ 5700 int i; 5701 5702 if (distance == node_distance(0, 0)) 5703 return true; 5704 5705 for (i = 0; i < sched_domains_numa_levels; i++) { 5706 if (sched_domains_numa_distance[i] == distance) 5707 return true; 5708 } 5709 5710 return false; 5711} 5712 5713static void sched_init_numa(void) 5714{ 5715 int next_distance, curr_distance = node_distance(0, 0); 5716 struct sched_domain_topology_level *tl; 5717 int level = 0; 5718 int i, j, k; 5719 5720 sched_domains_numa_distance = kzalloc(sizeof(int) * nr_node_ids, GFP_KERNEL); 5721 if (!sched_domains_numa_distance) 5722 return; 5723 5724 /* 5725 * O(nr_nodes^2) deduplicating selection sort -- in order to find the 5726 * unique distances in the node_distance() table. 5727 * 5728 * Assumes node_distance(0,j) includes all distances in 5729 * node_distance(i,j) in order to avoid cubic time. 5730 */ 5731 next_distance = curr_distance; 5732 for (i = 0; i < nr_node_ids; i++) { 5733 for (j = 0; j < nr_node_ids; j++) { 5734 for (k = 0; k < nr_node_ids; k++) { 5735 int distance = node_distance(i, k); 5736 5737 if (distance > curr_distance && 5738 (distance < next_distance || 5739 next_distance == curr_distance)) 5740 next_distance = distance; 5741 5742 /* 5743 * While not a strong assumption it would be nice to know 5744 * about cases where if node A is connected to B, B is not 5745 * equally connected to A. 5746 */ 5747 if (sched_debug() && node_distance(k, i) != distance) 5748 sched_numa_warn("Node-distance not symmetric"); 5749 5750 if (sched_debug() && i && !find_numa_distance(distance)) 5751 sched_numa_warn("Node-0 not representative"); 5752 } 5753 if (next_distance != curr_distance) { 5754 sched_domains_numa_distance[level++] = next_distance; 5755 sched_domains_numa_levels = level; 5756 curr_distance = next_distance; 5757 } else break; 5758 } 5759 5760 /* 5761 * In case of sched_debug() we verify the above assumption. 5762 */ 5763 if (!sched_debug()) 5764 break; 5765 } 5766 /* 5767 * 'level' contains the number of unique distances, excluding the 5768 * identity distance node_distance(i,i). 5769 * 5770 * The sched_domains_numa_distance[] array includes the actual distance 5771 * numbers. 5772 */ 5773 5774 /* 5775 * Here, we should temporarily reset sched_domains_numa_levels to 0. 5776 * If it fails to allocate memory for array sched_domains_numa_masks[][], 5777 * the array will contain less then 'level' members. This could be 5778 * dangerous when we use it to iterate array sched_domains_numa_masks[][] 5779 * in other functions. 5780 * 5781 * We reset it to 'level' at the end of this function. 5782 */ 5783 sched_domains_numa_levels = 0; 5784 5785 sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL); 5786 if (!sched_domains_numa_masks) 5787 return; 5788 5789 /* 5790 * Now for each level, construct a mask per node which contains all 5791 * cpus of nodes that are that many hops away from us. 5792 */ 5793 for (i = 0; i < level; i++) { 5794 sched_domains_numa_masks[i] = 5795 kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL); 5796 if (!sched_domains_numa_masks[i]) 5797 return; 5798 5799 for (j = 0; j < nr_node_ids; j++) { 5800 struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL); 5801 if (!mask) 5802 return; 5803 5804 sched_domains_numa_masks[i][j] = mask; 5805 5806 for (k = 0; k < nr_node_ids; k++) { 5807 if (node_distance(j, k) > sched_domains_numa_distance[i]) 5808 continue; 5809 5810 cpumask_or(mask, mask, cpumask_of_node(k)); 5811 } 5812 } 5813 } 5814 5815 tl = kzalloc((ARRAY_SIZE(default_topology) + level) * 5816 sizeof(struct sched_domain_topology_level), GFP_KERNEL); 5817 if (!tl) 5818 return; 5819 5820 /* 5821 * Copy the default topology bits.. 5822 */ 5823 for (i = 0; default_topology[i].init; i++) 5824 tl[i] = default_topology[i]; 5825 5826 /* 5827 * .. and append 'j' levels of NUMA goodness. 5828 */ 5829 for (j = 0; j < level; i++, j++) { 5830 tl[i] = (struct sched_domain_topology_level){ 5831 .init = sd_numa_init, 5832 .mask = sd_numa_mask, 5833 .flags = SDTL_OVERLAP, 5834 .numa_level = j, 5835 }; 5836 } 5837 5838 sched_domain_topology = tl; 5839 5840 sched_domains_numa_levels = level; 5841} 5842 5843static void sched_domains_numa_masks_set(int cpu) 5844{ 5845 int i, j; 5846 int node = cpu_to_node(cpu); 5847 5848 for (i = 0; i < sched_domains_numa_levels; i++) { 5849 for (j = 0; j < nr_node_ids; j++) { 5850 if (node_distance(j, node) <= sched_domains_numa_distance[i]) 5851 cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]); 5852 } 5853 } 5854} 5855 5856static void sched_domains_numa_masks_clear(int cpu) 5857{ 5858 int i, j; 5859 for (i = 0; i < sched_domains_numa_levels; i++) { 5860 for (j = 0; j < nr_node_ids; j++) 5861 cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]); 5862 } 5863} 5864 5865/* 5866 * Update sched_domains_numa_masks[level][node] array when new cpus 5867 * are onlined. 5868 */ 5869static int sched_domains_numa_masks_update(struct notifier_block *nfb, 5870 unsigned long action, 5871 void *hcpu) 5872{ 5873 int cpu = (long)hcpu; 5874 5875 switch (action & ~CPU_TASKS_FROZEN) { 5876 case CPU_ONLINE: 5877 sched_domains_numa_masks_set(cpu); 5878 break; 5879 5880 case CPU_DEAD: 5881 sched_domains_numa_masks_clear(cpu); 5882 break; 5883 5884 default: 5885 return NOTIFY_DONE; 5886 } 5887 5888 return NOTIFY_OK; 5889} 5890#else 5891static inline void sched_init_numa(void) 5892{ 5893} 5894 5895static int sched_domains_numa_masks_update(struct notifier_block *nfb, 5896 unsigned long action, 5897 void *hcpu) 5898{ 5899 return 0; 5900} 5901#endif /* CONFIG_NUMA */ 5902 5903static int __sdt_alloc(const struct cpumask *cpu_map) 5904{ 5905 struct sched_domain_topology_level *tl; 5906 int j; 5907 5908 for_each_sd_topology(tl) { 5909 struct sd_data *sdd = &tl->data; 5910 5911 sdd->sd = alloc_percpu(struct sched_domain *); 5912 if (!sdd->sd) 5913 return -ENOMEM; 5914 5915 sdd->sg = alloc_percpu(struct sched_group *); 5916 if (!sdd->sg) 5917 return -ENOMEM; 5918 5919 sdd->sgp = alloc_percpu(struct sched_group_power *); 5920 if (!sdd->sgp) 5921 return -ENOMEM; 5922 5923 for_each_cpu(j, cpu_map) { 5924 struct sched_domain *sd; 5925 struct sched_group *sg; 5926 struct sched_group_power *sgp; 5927 5928 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(), 5929 GFP_KERNEL, cpu_to_node(j)); 5930 if (!sd) 5931 return -ENOMEM; 5932 5933 *per_cpu_ptr(sdd->sd, j) = sd; 5934 5935 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(), 5936 GFP_KERNEL, cpu_to_node(j)); 5937 if (!sg) 5938 return -ENOMEM; 5939 5940 sg->next = sg; 5941 5942 *per_cpu_ptr(sdd->sg, j) = sg; 5943 5944 sgp = kzalloc_node(sizeof(struct sched_group_power) + cpumask_size(), 5945 GFP_KERNEL, cpu_to_node(j)); 5946 if (!sgp) 5947 return -ENOMEM; 5948 5949 *per_cpu_ptr(sdd->sgp, j) = sgp; 5950 } 5951 } 5952 5953 return 0; 5954} 5955 5956static void __sdt_free(const struct cpumask *cpu_map) 5957{ 5958 struct sched_domain_topology_level *tl; 5959 int j; 5960 5961 for_each_sd_topology(tl) { 5962 struct sd_data *sdd = &tl->data; 5963 5964 for_each_cpu(j, cpu_map) { 5965 struct sched_domain *sd; 5966 5967 if (sdd->sd) { 5968 sd = *per_cpu_ptr(sdd->sd, j); 5969 if (sd && (sd->flags & SD_OVERLAP)) 5970 free_sched_groups(sd->groups, 0); 5971 kfree(*per_cpu_ptr(sdd->sd, j)); 5972 } 5973 5974 if (sdd->sg) 5975 kfree(*per_cpu_ptr(sdd->sg, j)); 5976 if (sdd->sgp) 5977 kfree(*per_cpu_ptr(sdd->sgp, j)); 5978 } 5979 free_percpu(sdd->sd); 5980 sdd->sd = NULL; 5981 free_percpu(sdd->sg); 5982 sdd->sg = NULL; 5983 free_percpu(sdd->sgp); 5984 sdd->sgp = NULL; 5985 } 5986} 5987 5988struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl, 5989 const struct cpumask *cpu_map, struct sched_domain_attr *attr, 5990 struct sched_domain *child, int cpu) 5991{ 5992 struct sched_domain *sd = tl->init(tl, cpu); 5993 if (!sd) 5994 return child; 5995 5996 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu)); 5997 if (child) { 5998 sd->level = child->level + 1; 5999 sched_domain_level_max = max(sched_domain_level_max, sd->level); 6000 child->parent = sd; 6001 sd->child = child; 6002 } 6003 set_domain_attribute(sd, attr); 6004 6005 return sd; 6006} 6007 6008/* 6009 * Build sched domains for a given set of cpus and attach the sched domains 6010 * to the individual cpus 6011 */ 6012static int build_sched_domains(const struct cpumask *cpu_map, 6013 struct sched_domain_attr *attr) 6014{ 6015 enum s_alloc alloc_state; 6016 struct sched_domain *sd; 6017 struct s_data d; 6018 int i, ret = -ENOMEM; 6019 6020 alloc_state = __visit_domain_allocation_hell(&d, cpu_map); 6021 if (alloc_state != sa_rootdomain) 6022 goto error; 6023 6024 /* Set up domains for cpus specified by the cpu_map. */ 6025 for_each_cpu(i, cpu_map) { 6026 struct sched_domain_topology_level *tl; 6027 6028 sd = NULL; 6029 for_each_sd_topology(tl) { 6030 sd = build_sched_domain(tl, cpu_map, attr, sd, i); 6031 if (tl == sched_domain_topology) 6032 *per_cpu_ptr(d.sd, i) = sd; 6033 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP)) 6034 sd->flags |= SD_OVERLAP; 6035 if (cpumask_equal(cpu_map, sched_domain_span(sd))) 6036 break; 6037 } 6038 } 6039 6040 /* Build the groups for the domains */ 6041 for_each_cpu(i, cpu_map) { 6042 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) { 6043 sd->span_weight = cpumask_weight(sched_domain_span(sd)); 6044 if (sd->flags & SD_OVERLAP) { 6045 if (build_overlap_sched_groups(sd, i)) 6046 goto error; 6047 } else { 6048 if (build_sched_groups(sd, i)) 6049 goto error; 6050 } 6051 } 6052 } 6053 6054 /* Calculate CPU power for physical packages and nodes */ 6055 for (i = nr_cpumask_bits-1; i >= 0; i--) { 6056 if (!cpumask_test_cpu(i, cpu_map)) 6057 continue; 6058 6059 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) { 6060 claim_allocations(i, sd); 6061 init_sched_groups_power(i, sd); 6062 } 6063 } 6064 6065 /* Attach the domains */ 6066 rcu_read_lock(); 6067 for_each_cpu(i, cpu_map) { 6068 sd = *per_cpu_ptr(d.sd, i); 6069 cpu_attach_domain(sd, d.rd, i); 6070 } 6071 rcu_read_unlock(); 6072 6073 ret = 0; 6074error: 6075 __free_domain_allocs(&d, alloc_state, cpu_map); 6076 return ret; 6077} 6078 6079static cpumask_var_t *doms_cur; /* current sched domains */ 6080static int ndoms_cur; /* number of sched domains in 'doms_cur' */ 6081static struct sched_domain_attr *dattr_cur; 6082 /* attribues of custom domains in 'doms_cur' */ 6083 6084/* 6085 * Special case: If a kmalloc of a doms_cur partition (array of 6086 * cpumask) fails, then fallback to a single sched domain, 6087 * as determined by the single cpumask fallback_doms. 6088 */ 6089static cpumask_var_t fallback_doms; 6090 6091/* 6092 * arch_update_cpu_topology lets virtualized architectures update the 6093 * cpu core maps. It is supposed to return 1 if the topology changed 6094 * or 0 if it stayed the same. 6095 */ 6096int __attribute__((weak)) arch_update_cpu_topology(void) 6097{ 6098 return 0; 6099} 6100 6101cpumask_var_t *alloc_sched_domains(unsigned int ndoms) 6102{ 6103 int i; 6104 cpumask_var_t *doms; 6105 6106 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL); 6107 if (!doms) 6108 return NULL; 6109 for (i = 0; i < ndoms; i++) { 6110 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) { 6111 free_sched_domains(doms, i); 6112 return NULL; 6113 } 6114 } 6115 return doms; 6116} 6117 6118void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms) 6119{ 6120 unsigned int i; 6121 for (i = 0; i < ndoms; i++) 6122 free_cpumask_var(doms[i]); 6123 kfree(doms); 6124} 6125 6126/* 6127 * Set up scheduler domains and groups. Callers must hold the hotplug lock. 6128 * For now this just excludes isolated cpus, but could be used to 6129 * exclude other special cases in the future. 6130 */ 6131static int init_sched_domains(const struct cpumask *cpu_map) 6132{ 6133 int err; 6134 6135 arch_update_cpu_topology(); 6136 ndoms_cur = 1; 6137 doms_cur = alloc_sched_domains(ndoms_cur); 6138 if (!doms_cur) 6139 doms_cur = &fallback_doms; 6140 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map); 6141 err = build_sched_domains(doms_cur[0], NULL); 6142 register_sched_domain_sysctl(); 6143 6144 return err; 6145} 6146 6147/* 6148 * Detach sched domains from a group of cpus specified in cpu_map 6149 * These cpus will now be attached to the NULL domain 6150 */ 6151static void detach_destroy_domains(const struct cpumask *cpu_map) 6152{ 6153 int i; 6154 6155 rcu_read_lock(); 6156 for_each_cpu(i, cpu_map) 6157 cpu_attach_domain(NULL, &def_root_domain, i); 6158 rcu_read_unlock(); 6159} 6160 6161/* handle null as "default" */ 6162static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur, 6163 struct sched_domain_attr *new, int idx_new) 6164{ 6165 struct sched_domain_attr tmp; 6166 6167 /* fast path */ 6168 if (!new && !cur) 6169 return 1; 6170 6171 tmp = SD_ATTR_INIT; 6172 return !memcmp(cur ? (cur + idx_cur) : &tmp, 6173 new ? (new + idx_new) : &tmp, 6174 sizeof(struct sched_domain_attr)); 6175} 6176 6177/* 6178 * Partition sched domains as specified by the 'ndoms_new' 6179 * cpumasks in the array doms_new[] of cpumasks. This compares 6180 * doms_new[] to the current sched domain partitioning, doms_cur[]. 6181 * It destroys each deleted domain and builds each new domain. 6182 * 6183 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'. 6184 * The masks don't intersect (don't overlap.) We should setup one 6185 * sched domain for each mask. CPUs not in any of the cpumasks will 6186 * not be load balanced. If the same cpumask appears both in the 6187 * current 'doms_cur' domains and in the new 'doms_new', we can leave 6188 * it as it is. 6189 * 6190 * The passed in 'doms_new' should be allocated using 6191 * alloc_sched_domains. This routine takes ownership of it and will 6192 * free_sched_domains it when done with it. If the caller failed the 6193 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1, 6194 * and partition_sched_domains() will fallback to the single partition 6195 * 'fallback_doms', it also forces the domains to be rebuilt. 6196 * 6197 * If doms_new == NULL it will be replaced with cpu_online_mask. 6198 * ndoms_new == 0 is a special case for destroying existing domains, 6199 * and it will not create the default domain. 6200 * 6201 * Call with hotplug lock held 6202 */ 6203void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[], 6204 struct sched_domain_attr *dattr_new) 6205{ 6206 int i, j, n; 6207 int new_topology; 6208 6209 mutex_lock(&sched_domains_mutex); 6210 6211 /* always unregister in case we don't destroy any domains */ 6212 unregister_sched_domain_sysctl(); 6213 6214 /* Let architecture update cpu core mappings. */ 6215 new_topology = arch_update_cpu_topology(); 6216 6217 n = doms_new ? ndoms_new : 0; 6218 6219 /* Destroy deleted domains */ 6220 for (i = 0; i < ndoms_cur; i++) { 6221 for (j = 0; j < n && !new_topology; j++) { 6222 if (cpumask_equal(doms_cur[i], doms_new[j]) 6223 && dattrs_equal(dattr_cur, i, dattr_new, j)) 6224 goto match1; 6225 } 6226 /* no match - a current sched domain not in new doms_new[] */ 6227 detach_destroy_domains(doms_cur[i]); 6228match1: 6229 ; 6230 } 6231 6232 n = ndoms_cur; 6233 if (doms_new == NULL) { 6234 n = 0; 6235 doms_new = &fallback_doms; 6236 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map); 6237 WARN_ON_ONCE(dattr_new); 6238 } 6239 6240 /* Build new domains */ 6241 for (i = 0; i < ndoms_new; i++) { 6242 for (j = 0; j < n && !new_topology; j++) { 6243 if (cpumask_equal(doms_new[i], doms_cur[j]) 6244 && dattrs_equal(dattr_new, i, dattr_cur, j)) 6245 goto match2; 6246 } 6247 /* no match - add a new doms_new */ 6248 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL); 6249match2: 6250 ; 6251 } 6252 6253 /* Remember the new sched domains */ 6254 if (doms_cur != &fallback_doms) 6255 free_sched_domains(doms_cur, ndoms_cur); 6256 kfree(dattr_cur); /* kfree(NULL) is safe */ 6257 doms_cur = doms_new; 6258 dattr_cur = dattr_new; 6259 ndoms_cur = ndoms_new; 6260 6261 register_sched_domain_sysctl(); 6262 6263 mutex_unlock(&sched_domains_mutex); 6264} 6265 6266static int num_cpus_frozen; /* used to mark begin/end of suspend/resume */ 6267 6268/* 6269 * Update cpusets according to cpu_active mask. If cpusets are 6270 * disabled, cpuset_update_active_cpus() becomes a simple wrapper 6271 * around partition_sched_domains(). 6272 * 6273 * If we come here as part of a suspend/resume, don't touch cpusets because we 6274 * want to restore it back to its original state upon resume anyway. 6275 */ 6276static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action, 6277 void *hcpu) 6278{ 6279 switch (action) { 6280 case CPU_ONLINE_FROZEN: 6281 case CPU_DOWN_FAILED_FROZEN: 6282 6283 /* 6284 * num_cpus_frozen tracks how many CPUs are involved in suspend 6285 * resume sequence. As long as this is not the last online 6286 * operation in the resume sequence, just build a single sched 6287 * domain, ignoring cpusets. 6288 */ 6289 num_cpus_frozen--; 6290 if (likely(num_cpus_frozen)) { 6291 partition_sched_domains(1, NULL, NULL); 6292 break; 6293 } 6294 6295 /* 6296 * This is the last CPU online operation. So fall through and 6297 * restore the original sched domains by considering the 6298 * cpuset configurations. 6299 */ 6300 6301 case CPU_ONLINE: 6302 case CPU_DOWN_FAILED: 6303 cpuset_update_active_cpus(true); 6304 break; 6305 default: 6306 return NOTIFY_DONE; 6307 } 6308 return NOTIFY_OK; 6309} 6310 6311static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action, 6312 void *hcpu) 6313{ 6314 switch (action) { 6315 case CPU_DOWN_PREPARE: 6316 cpuset_update_active_cpus(false); 6317 break; 6318 case CPU_DOWN_PREPARE_FROZEN: 6319 num_cpus_frozen++; 6320 partition_sched_domains(1, NULL, NULL); 6321 break; 6322 default: 6323 return NOTIFY_DONE; 6324 } 6325 return NOTIFY_OK; 6326} 6327 6328void __init sched_init_smp(void) 6329{ 6330 cpumask_var_t non_isolated_cpus; 6331 6332 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL); 6333 alloc_cpumask_var(&fallback_doms, GFP_KERNEL); 6334 6335 sched_init_numa(); 6336 6337 get_online_cpus(); 6338 mutex_lock(&sched_domains_mutex); 6339 init_sched_domains(cpu_active_mask); 6340 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map); 6341 if (cpumask_empty(non_isolated_cpus)) 6342 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus); 6343 mutex_unlock(&sched_domains_mutex); 6344 put_online_cpus(); 6345 6346 hotcpu_notifier(sched_domains_numa_masks_update, CPU_PRI_SCHED_ACTIVE); 6347 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE); 6348 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE); 6349 6350 init_hrtick(); 6351 6352 /* Move init over to a non-isolated CPU */ 6353 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0) 6354 BUG(); 6355 sched_init_granularity(); 6356 free_cpumask_var(non_isolated_cpus); 6357 6358 init_sched_rt_class(); 6359} 6360#else 6361void __init sched_init_smp(void) 6362{ 6363 sched_init_granularity(); 6364} 6365#endif /* CONFIG_SMP */ 6366 6367const_debug unsigned int sysctl_timer_migration = 1; 6368 6369int in_sched_functions(unsigned long addr) 6370{ 6371 return in_lock_functions(addr) || 6372 (addr >= (unsigned long)__sched_text_start 6373 && addr < (unsigned long)__sched_text_end); 6374} 6375 6376#ifdef CONFIG_CGROUP_SCHED 6377/* 6378 * Default task group. 6379 * Every task in system belongs to this group at bootup. 6380 */ 6381struct task_group root_task_group; 6382LIST_HEAD(task_groups); 6383#endif 6384 6385DECLARE_PER_CPU(cpumask_var_t, load_balance_mask); 6386 6387void __init sched_init(void) 6388{ 6389 int i, j; 6390 unsigned long alloc_size = 0, ptr; 6391 6392#ifdef CONFIG_FAIR_GROUP_SCHED 6393 alloc_size += 2 * nr_cpu_ids * sizeof(void **); 6394#endif 6395#ifdef CONFIG_RT_GROUP_SCHED 6396 alloc_size += 2 * nr_cpu_ids * sizeof(void **); 6397#endif 6398#ifdef CONFIG_CPUMASK_OFFSTACK 6399 alloc_size += num_possible_cpus() * cpumask_size(); 6400#endif 6401 if (alloc_size) { 6402 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT); 6403 6404#ifdef CONFIG_FAIR_GROUP_SCHED 6405 root_task_group.se = (struct sched_entity **)ptr; 6406 ptr += nr_cpu_ids * sizeof(void **); 6407 6408 root_task_group.cfs_rq = (struct cfs_rq **)ptr; 6409 ptr += nr_cpu_ids * sizeof(void **); 6410 6411#endif /* CONFIG_FAIR_GROUP_SCHED */ 6412#ifdef CONFIG_RT_GROUP_SCHED 6413 root_task_group.rt_se = (struct sched_rt_entity **)ptr; 6414 ptr += nr_cpu_ids * sizeof(void **); 6415 6416 root_task_group.rt_rq = (struct rt_rq **)ptr; 6417 ptr += nr_cpu_ids * sizeof(void **); 6418 6419#endif /* CONFIG_RT_GROUP_SCHED */ 6420#ifdef CONFIG_CPUMASK_OFFSTACK 6421 for_each_possible_cpu(i) { 6422 per_cpu(load_balance_mask, i) = (void *)ptr; 6423 ptr += cpumask_size(); 6424 } 6425#endif /* CONFIG_CPUMASK_OFFSTACK */ 6426 } 6427 6428#ifdef CONFIG_SMP 6429 init_defrootdomain(); 6430#endif 6431 6432 init_rt_bandwidth(&def_rt_bandwidth, 6433 global_rt_period(), global_rt_runtime()); 6434 6435#ifdef CONFIG_RT_GROUP_SCHED 6436 init_rt_bandwidth(&root_task_group.rt_bandwidth, 6437 global_rt_period(), global_rt_runtime()); 6438#endif /* CONFIG_RT_GROUP_SCHED */ 6439 6440#ifdef CONFIG_CGROUP_SCHED 6441 list_add(&root_task_group.list, &task_groups); 6442 INIT_LIST_HEAD(&root_task_group.children); 6443 INIT_LIST_HEAD(&root_task_group.siblings); 6444 autogroup_init(&init_task); 6445 6446#endif /* CONFIG_CGROUP_SCHED */ 6447 6448 for_each_possible_cpu(i) { 6449 struct rq *rq; 6450 6451 rq = cpu_rq(i); 6452 raw_spin_lock_init(&rq->lock); 6453 rq->nr_running = 0; 6454 rq->calc_load_active = 0; 6455 rq->calc_load_update = jiffies + LOAD_FREQ; 6456 init_cfs_rq(&rq->cfs); 6457 init_rt_rq(&rq->rt, rq); 6458#ifdef CONFIG_FAIR_GROUP_SCHED 6459 root_task_group.shares = ROOT_TASK_GROUP_LOAD; 6460 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list); 6461 /* 6462 * How much cpu bandwidth does root_task_group get? 6463 * 6464 * In case of task-groups formed thr' the cgroup filesystem, it 6465 * gets 100% of the cpu resources in the system. This overall 6466 * system cpu resource is divided among the tasks of 6467 * root_task_group and its child task-groups in a fair manner, 6468 * based on each entity's (task or task-group's) weight 6469 * (se->load.weight). 6470 * 6471 * In other words, if root_task_group has 10 tasks of weight 6472 * 1024) and two child groups A0 and A1 (of weight 1024 each), 6473 * then A0's share of the cpu resource is: 6474 * 6475 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33% 6476 * 6477 * We achieve this by letting root_task_group's tasks sit 6478 * directly in rq->cfs (i.e root_task_group->se[] = NULL). 6479 */ 6480 init_cfs_bandwidth(&root_task_group.cfs_bandwidth); 6481 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL); 6482#endif /* CONFIG_FAIR_GROUP_SCHED */ 6483 6484 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime; 6485#ifdef CONFIG_RT_GROUP_SCHED 6486 INIT_LIST_HEAD(&rq->leaf_rt_rq_list); 6487 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL); 6488#endif 6489 6490 for (j = 0; j < CPU_LOAD_IDX_MAX; j++) 6491 rq->cpu_load[j] = 0; 6492 6493 rq->last_load_update_tick = jiffies; 6494 6495#ifdef CONFIG_SMP 6496 rq->sd = NULL; 6497 rq->rd = NULL; 6498 rq->cpu_power = SCHED_POWER_SCALE; 6499 rq->post_schedule = 0; 6500 rq->active_balance = 0; 6501 rq->next_balance = jiffies; 6502 rq->push_cpu = 0; 6503 rq->cpu = i; 6504 rq->online = 0; 6505 rq->idle_stamp = 0; 6506 rq->avg_idle = 2*sysctl_sched_migration_cost; 6507 rq->max_idle_balance_cost = sysctl_sched_migration_cost; 6508 6509 INIT_LIST_HEAD(&rq->cfs_tasks); 6510 6511 rq_attach_root(rq, &def_root_domain); 6512#ifdef CONFIG_NO_HZ_COMMON 6513 rq->nohz_flags = 0; 6514#endif 6515#ifdef CONFIG_NO_HZ_FULL 6516 rq->last_sched_tick = 0; 6517#endif 6518#endif 6519 init_rq_hrtick(rq); 6520 atomic_set(&rq->nr_iowait, 0); 6521 } 6522 6523 set_load_weight(&init_task); 6524 6525#ifdef CONFIG_PREEMPT_NOTIFIERS 6526 INIT_HLIST_HEAD(&init_task.preempt_notifiers); 6527#endif 6528 6529#ifdef CONFIG_RT_MUTEXES 6530 plist_head_init(&init_task.pi_waiters); 6531#endif 6532 6533 /* 6534 * The boot idle thread does lazy MMU switching as well: 6535 */ 6536 atomic_inc(&init_mm.mm_count); 6537 enter_lazy_tlb(&init_mm, current); 6538 6539 /* 6540 * Make us the idle thread. Technically, schedule() should not be 6541 * called from this thread, however somewhere below it might be, 6542 * but because we are the idle thread, we just pick up running again 6543 * when this runqueue becomes "idle". 6544 */ 6545 init_idle(current, smp_processor_id()); 6546 6547 calc_load_update = jiffies + LOAD_FREQ; 6548 6549 /* 6550 * During early bootup we pretend to be a normal task: 6551 */ 6552 current->sched_class = &fair_sched_class; 6553 6554#ifdef CONFIG_SMP 6555 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT); 6556 /* May be allocated at isolcpus cmdline parse time */ 6557 if (cpu_isolated_map == NULL) 6558 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT); 6559 idle_thread_set_boot_cpu(); 6560#endif 6561 init_sched_fair_class(); 6562 6563 scheduler_running = 1; 6564} 6565 6566#ifdef CONFIG_DEBUG_ATOMIC_SLEEP 6567static inline int preempt_count_equals(int preempt_offset) 6568{ 6569 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth(); 6570 6571 return (nested == preempt_offset); 6572} 6573 6574void __might_sleep(const char *file, int line, int preempt_offset) 6575{ 6576 static unsigned long prev_jiffy; /* ratelimiting */ 6577 6578 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */ 6579 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) || 6580 system_state != SYSTEM_RUNNING || oops_in_progress) 6581 return; 6582 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy) 6583 return; 6584 prev_jiffy = jiffies; 6585 6586 printk(KERN_ERR 6587 "BUG: sleeping function called from invalid context at %s:%d\n", 6588 file, line); 6589 printk(KERN_ERR 6590 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n", 6591 in_atomic(), irqs_disabled(), 6592 current->pid, current->comm); 6593 6594 debug_show_held_locks(current); 6595 if (irqs_disabled()) 6596 print_irqtrace_events(current); 6597 dump_stack(); 6598} 6599EXPORT_SYMBOL(__might_sleep); 6600#endif 6601 6602#ifdef CONFIG_MAGIC_SYSRQ 6603static void normalize_task(struct rq *rq, struct task_struct *p) 6604{ 6605 const struct sched_class *prev_class = p->sched_class; 6606 int old_prio = p->prio; 6607 int on_rq; 6608 6609 on_rq = p->on_rq; 6610 if (on_rq) 6611 dequeue_task(rq, p, 0); 6612 __setscheduler(rq, p, SCHED_NORMAL, 0); 6613 if (on_rq) { 6614 enqueue_task(rq, p, 0); 6615 resched_task(rq->curr); 6616 } 6617 6618 check_class_changed(rq, p, prev_class, old_prio); 6619} 6620 6621void normalize_rt_tasks(void) 6622{ 6623 struct task_struct *g, *p; 6624 unsigned long flags; 6625 struct rq *rq; 6626 6627 read_lock_irqsave(&tasklist_lock, flags); 6628 do_each_thread(g, p) { 6629 /* 6630 * Only normalize user tasks: 6631 */ 6632 if (!p->mm) 6633 continue; 6634 6635 p->se.exec_start = 0; 6636#ifdef CONFIG_SCHEDSTATS 6637 p->se.statistics.wait_start = 0; 6638 p->se.statistics.sleep_start = 0; 6639 p->se.statistics.block_start = 0; 6640#endif 6641 6642 if (!rt_task(p)) { 6643 /* 6644 * Renice negative nice level userspace 6645 * tasks back to 0: 6646 */ 6647 if (TASK_NICE(p) < 0 && p->mm) 6648 set_user_nice(p, 0); 6649 continue; 6650 } 6651 6652 raw_spin_lock(&p->pi_lock); 6653 rq = __task_rq_lock(p); 6654 6655 normalize_task(rq, p); 6656 6657 __task_rq_unlock(rq); 6658 raw_spin_unlock(&p->pi_lock); 6659 } while_each_thread(g, p); 6660 6661 read_unlock_irqrestore(&tasklist_lock, flags); 6662} 6663 6664#endif /* CONFIG_MAGIC_SYSRQ */ 6665 6666#if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) 6667/* 6668 * These functions are only useful for the IA64 MCA handling, or kdb. 6669 * 6670 * They can only be called when the whole system has been 6671 * stopped - every CPU needs to be quiescent, and no scheduling 6672 * activity can take place. Using them for anything else would 6673 * be a serious bug, and as a result, they aren't even visible 6674 * under any other configuration. 6675 */ 6676 6677/** 6678 * curr_task - return the current task for a given cpu. 6679 * @cpu: the processor in question. 6680 * 6681 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED! 6682 * 6683 * Return: The current task for @cpu. 6684 */ 6685struct task_struct *curr_task(int cpu) 6686{ 6687 return cpu_curr(cpu); 6688} 6689 6690#endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */ 6691 6692#ifdef CONFIG_IA64 6693/** 6694 * set_curr_task - set the current task for a given cpu. 6695 * @cpu: the processor in question. 6696 * @p: the task pointer to set. 6697 * 6698 * Description: This function must only be used when non-maskable interrupts 6699 * are serviced on a separate stack. It allows the architecture to switch the 6700 * notion of the current task on a cpu in a non-blocking manner. This function 6701 * must be called with all CPU's synchronized, and interrupts disabled, the 6702 * and caller must save the original value of the current task (see 6703 * curr_task() above) and restore that value before reenabling interrupts and 6704 * re-starting the system. 6705 * 6706 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED! 6707 */ 6708void set_curr_task(int cpu, struct task_struct *p) 6709{ 6710 cpu_curr(cpu) = p; 6711} 6712 6713#endif 6714 6715#ifdef CONFIG_CGROUP_SCHED 6716/* task_group_lock serializes the addition/removal of task groups */ 6717static DEFINE_SPINLOCK(task_group_lock); 6718 6719static void free_sched_group(struct task_group *tg) 6720{ 6721 free_fair_sched_group(tg); 6722 free_rt_sched_group(tg); 6723 autogroup_free(tg); 6724 kfree(tg); 6725} 6726 6727/* allocate runqueue etc for a new task group */ 6728struct task_group *sched_create_group(struct task_group *parent) 6729{ 6730 struct task_group *tg; 6731 6732 tg = kzalloc(sizeof(*tg), GFP_KERNEL); 6733 if (!tg) 6734 return ERR_PTR(-ENOMEM); 6735 6736 if (!alloc_fair_sched_group(tg, parent)) 6737 goto err; 6738 6739 if (!alloc_rt_sched_group(tg, parent)) 6740 goto err; 6741 6742 return tg; 6743 6744err: 6745 free_sched_group(tg); 6746 return ERR_PTR(-ENOMEM); 6747} 6748 6749void sched_online_group(struct task_group *tg, struct task_group *parent) 6750{ 6751 unsigned long flags; 6752 6753 spin_lock_irqsave(&task_group_lock, flags); 6754 list_add_rcu(&tg->list, &task_groups); 6755 6756 WARN_ON(!parent); /* root should already exist */ 6757 6758 tg->parent = parent; 6759 INIT_LIST_HEAD(&tg->children); 6760 list_add_rcu(&tg->siblings, &parent->children); 6761 spin_unlock_irqrestore(&task_group_lock, flags); 6762} 6763 6764/* rcu callback to free various structures associated with a task group */ 6765static void free_sched_group_rcu(struct rcu_head *rhp) 6766{ 6767 /* now it should be safe to free those cfs_rqs */ 6768 free_sched_group(container_of(rhp, struct task_group, rcu)); 6769} 6770 6771/* Destroy runqueue etc associated with a task group */ 6772void sched_destroy_group(struct task_group *tg) 6773{ 6774 /* wait for possible concurrent references to cfs_rqs complete */ 6775 call_rcu(&tg->rcu, free_sched_group_rcu); 6776} 6777 6778void sched_offline_group(struct task_group *tg) 6779{ 6780 unsigned long flags; 6781 int i; 6782 6783 /* end participation in shares distribution */ 6784 for_each_possible_cpu(i) 6785 unregister_fair_sched_group(tg, i); 6786 6787 spin_lock_irqsave(&task_group_lock, flags); 6788 list_del_rcu(&tg->list); 6789 list_del_rcu(&tg->siblings); 6790 spin_unlock_irqrestore(&task_group_lock, flags); 6791} 6792 6793/* change task's runqueue when it moves between groups. 6794 * The caller of this function should have put the task in its new group 6795 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to 6796 * reflect its new group. 6797 */ 6798void sched_move_task(struct task_struct *tsk) 6799{ 6800 struct task_group *tg; 6801 int on_rq, running; 6802 unsigned long flags; 6803 struct rq *rq; 6804 6805 rq = task_rq_lock(tsk, &flags); 6806 6807 running = task_current(rq, tsk); 6808 on_rq = tsk->on_rq; 6809 6810 if (on_rq) 6811 dequeue_task(rq, tsk, 0); 6812 if (unlikely(running)) 6813 tsk->sched_class->put_prev_task(rq, tsk); 6814 6815 tg = container_of(task_css_check(tsk, cpu_cgroup_subsys_id, 6816 lockdep_is_held(&tsk->sighand->siglock)), 6817 struct task_group, css); 6818 tg = autogroup_task_group(tsk, tg); 6819 tsk->sched_task_group = tg; 6820 6821#ifdef CONFIG_FAIR_GROUP_SCHED 6822 if (tsk->sched_class->task_move_group) 6823 tsk->sched_class->task_move_group(tsk, on_rq); 6824 else 6825#endif 6826 set_task_rq(tsk, task_cpu(tsk)); 6827 6828 if (unlikely(running)) 6829 tsk->sched_class->set_curr_task(rq); 6830 if (on_rq) 6831 enqueue_task(rq, tsk, 0); 6832 6833 task_rq_unlock(rq, tsk, &flags); 6834} 6835#endif /* CONFIG_CGROUP_SCHED */ 6836 6837#if defined(CONFIG_RT_GROUP_SCHED) || defined(CONFIG_CFS_BANDWIDTH) 6838static unsigned long to_ratio(u64 period, u64 runtime) 6839{ 6840 if (runtime == RUNTIME_INF) 6841 return 1ULL << 20; 6842 6843 return div64_u64(runtime << 20, period); 6844} 6845#endif 6846 6847#ifdef CONFIG_RT_GROUP_SCHED 6848/* 6849 * Ensure that the real time constraints are schedulable. 6850 */ 6851static DEFINE_MUTEX(rt_constraints_mutex); 6852 6853/* Must be called with tasklist_lock held */ 6854static inline int tg_has_rt_tasks(struct task_group *tg) 6855{ 6856 struct task_struct *g, *p; 6857 6858 do_each_thread(g, p) { 6859 if (rt_task(p) && task_rq(p)->rt.tg == tg) 6860 return 1; 6861 } while_each_thread(g, p); 6862 6863 return 0; 6864} 6865 6866struct rt_schedulable_data { 6867 struct task_group *tg; 6868 u64 rt_period; 6869 u64 rt_runtime; 6870}; 6871 6872static int tg_rt_schedulable(struct task_group *tg, void *data) 6873{ 6874 struct rt_schedulable_data *d = data; 6875 struct task_group *child; 6876 unsigned long total, sum = 0; 6877 u64 period, runtime; 6878 6879 period = ktime_to_ns(tg->rt_bandwidth.rt_period); 6880 runtime = tg->rt_bandwidth.rt_runtime; 6881 6882 if (tg == d->tg) { 6883 period = d->rt_period; 6884 runtime = d->rt_runtime; 6885 } 6886 6887 /* 6888 * Cannot have more runtime than the period. 6889 */ 6890 if (runtime > period && runtime != RUNTIME_INF) 6891 return -EINVAL; 6892 6893 /* 6894 * Ensure we don't starve existing RT tasks. 6895 */ 6896 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg)) 6897 return -EBUSY; 6898 6899 total = to_ratio(period, runtime); 6900 6901 /* 6902 * Nobody can have more than the global setting allows. 6903 */ 6904 if (total > to_ratio(global_rt_period(), global_rt_runtime())) 6905 return -EINVAL; 6906 6907 /* 6908 * The sum of our children's runtime should not exceed our own. 6909 */ 6910 list_for_each_entry_rcu(child, &tg->children, siblings) { 6911 period = ktime_to_ns(child->rt_bandwidth.rt_period); 6912 runtime = child->rt_bandwidth.rt_runtime; 6913 6914 if (child == d->tg) { 6915 period = d->rt_period; 6916 runtime = d->rt_runtime; 6917 } 6918 6919 sum += to_ratio(period, runtime); 6920 } 6921 6922 if (sum > total) 6923 return -EINVAL; 6924 6925 return 0; 6926} 6927 6928static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime) 6929{ 6930 int ret; 6931 6932 struct rt_schedulable_data data = { 6933 .tg = tg, 6934 .rt_period = period, 6935 .rt_runtime = runtime, 6936 }; 6937 6938 rcu_read_lock(); 6939 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data); 6940 rcu_read_unlock(); 6941 6942 return ret; 6943} 6944 6945static int tg_set_rt_bandwidth(struct task_group *tg, 6946 u64 rt_period, u64 rt_runtime) 6947{ 6948 int i, err = 0; 6949 6950 mutex_lock(&rt_constraints_mutex); 6951 read_lock(&tasklist_lock); 6952 err = __rt_schedulable(tg, rt_period, rt_runtime); 6953 if (err) 6954 goto unlock; 6955 6956 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock); 6957 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period); 6958 tg->rt_bandwidth.rt_runtime = rt_runtime; 6959 6960 for_each_possible_cpu(i) { 6961 struct rt_rq *rt_rq = tg->rt_rq[i]; 6962 6963 raw_spin_lock(&rt_rq->rt_runtime_lock); 6964 rt_rq->rt_runtime = rt_runtime; 6965 raw_spin_unlock(&rt_rq->rt_runtime_lock); 6966 } 6967 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock); 6968unlock: 6969 read_unlock(&tasklist_lock); 6970 mutex_unlock(&rt_constraints_mutex); 6971 6972 return err; 6973} 6974 6975static int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us) 6976{ 6977 u64 rt_runtime, rt_period; 6978 6979 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period); 6980 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC; 6981 if (rt_runtime_us < 0) 6982 rt_runtime = RUNTIME_INF; 6983 6984 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime); 6985} 6986 6987static long sched_group_rt_runtime(struct task_group *tg) 6988{ 6989 u64 rt_runtime_us; 6990 6991 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF) 6992 return -1; 6993 6994 rt_runtime_us = tg->rt_bandwidth.rt_runtime; 6995 do_div(rt_runtime_us, NSEC_PER_USEC); 6996 return rt_runtime_us; 6997} 6998 6999static int sched_group_set_rt_period(struct task_group *tg, long rt_period_us) 7000{ 7001 u64 rt_runtime, rt_period; 7002 7003 rt_period = (u64)rt_period_us * NSEC_PER_USEC; 7004 rt_runtime = tg->rt_bandwidth.rt_runtime; 7005 7006 if (rt_period == 0) 7007 return -EINVAL; 7008 7009 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime); 7010} 7011 7012static long sched_group_rt_period(struct task_group *tg) 7013{ 7014 u64 rt_period_us; 7015 7016 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period); 7017 do_div(rt_period_us, NSEC_PER_USEC); 7018 return rt_period_us; 7019} 7020 7021static int sched_rt_global_constraints(void) 7022{ 7023 u64 runtime, period; 7024 int ret = 0; 7025 7026 if (sysctl_sched_rt_period <= 0) 7027 return -EINVAL; 7028 7029 runtime = global_rt_runtime(); 7030 period = global_rt_period(); 7031 7032 /* 7033 * Sanity check on the sysctl variables. 7034 */ 7035 if (runtime > period && runtime != RUNTIME_INF) 7036 return -EINVAL; 7037 7038 mutex_lock(&rt_constraints_mutex); 7039 read_lock(&tasklist_lock); 7040 ret = __rt_schedulable(NULL, 0, 0); 7041 read_unlock(&tasklist_lock); 7042 mutex_unlock(&rt_constraints_mutex); 7043 7044 return ret; 7045} 7046 7047static int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk) 7048{ 7049 /* Don't accept realtime tasks when there is no way for them to run */ 7050 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0) 7051 return 0; 7052 7053 return 1; 7054} 7055 7056#else /* !CONFIG_RT_GROUP_SCHED */ 7057static int sched_rt_global_constraints(void) 7058{ 7059 unsigned long flags; 7060 int i; 7061 7062 if (sysctl_sched_rt_period <= 0) 7063 return -EINVAL; 7064 7065 /* 7066 * There's always some RT tasks in the root group 7067 * -- migration, kstopmachine etc.. 7068 */ 7069 if (sysctl_sched_rt_runtime == 0) 7070 return -EBUSY; 7071 7072 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags); 7073 for_each_possible_cpu(i) { 7074 struct rt_rq *rt_rq = &cpu_rq(i)->rt; 7075 7076 raw_spin_lock(&rt_rq->rt_runtime_lock); 7077 rt_rq->rt_runtime = global_rt_runtime(); 7078 raw_spin_unlock(&rt_rq->rt_runtime_lock); 7079 } 7080 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags); 7081 7082 return 0; 7083} 7084#endif /* CONFIG_RT_GROUP_SCHED */ 7085 7086int sched_rr_handler(struct ctl_table *table, int write, 7087 void __user *buffer, size_t *lenp, 7088 loff_t *ppos) 7089{ 7090 int ret; 7091 static DEFINE_MUTEX(mutex); 7092 7093 mutex_lock(&mutex); 7094 ret = proc_dointvec(table, write, buffer, lenp, ppos); 7095 /* make sure that internally we keep jiffies */ 7096 /* also, writing zero resets timeslice to default */ 7097 if (!ret && write) { 7098 sched_rr_timeslice = sched_rr_timeslice <= 0 ? 7099 RR_TIMESLICE : msecs_to_jiffies(sched_rr_timeslice); 7100 } 7101 mutex_unlock(&mutex); 7102 return ret; 7103} 7104 7105int sched_rt_handler(struct ctl_table *table, int write, 7106 void __user *buffer, size_t *lenp, 7107 loff_t *ppos) 7108{ 7109 int ret; 7110 int old_period, old_runtime; 7111 static DEFINE_MUTEX(mutex); 7112 7113 mutex_lock(&mutex); 7114 old_period = sysctl_sched_rt_period; 7115 old_runtime = sysctl_sched_rt_runtime; 7116 7117 ret = proc_dointvec(table, write, buffer, lenp, ppos); 7118 7119 if (!ret && write) { 7120 ret = sched_rt_global_constraints(); 7121 if (ret) { 7122 sysctl_sched_rt_period = old_period; 7123 sysctl_sched_rt_runtime = old_runtime; 7124 } else { 7125 def_rt_bandwidth.rt_runtime = global_rt_runtime(); 7126 def_rt_bandwidth.rt_period = 7127 ns_to_ktime(global_rt_period()); 7128 } 7129 } 7130 mutex_unlock(&mutex); 7131 7132 return ret; 7133} 7134 7135#ifdef CONFIG_CGROUP_SCHED 7136 7137static inline struct task_group *css_tg(struct cgroup_subsys_state *css) 7138{ 7139 return css ? container_of(css, struct task_group, css) : NULL; 7140} 7141 7142static struct cgroup_subsys_state * 7143cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css) 7144{ 7145 struct task_group *parent = css_tg(parent_css); 7146 struct task_group *tg; 7147 7148 if (!parent) { 7149 /* This is early initialization for the top cgroup */ 7150 return &root_task_group.css; 7151 } 7152 7153 tg = sched_create_group(parent); 7154 if (IS_ERR(tg)) 7155 return ERR_PTR(-ENOMEM); 7156 7157 return &tg->css; 7158} 7159 7160static int cpu_cgroup_css_online(struct cgroup_subsys_state *css) 7161{ 7162 struct task_group *tg = css_tg(css); 7163 struct task_group *parent = css_tg(css_parent(css)); 7164 7165 if (parent) 7166 sched_online_group(tg, parent); 7167 return 0; 7168} 7169 7170static void cpu_cgroup_css_free(struct cgroup_subsys_state *css) 7171{ 7172 struct task_group *tg = css_tg(css); 7173 7174 sched_destroy_group(tg); 7175} 7176 7177static void cpu_cgroup_css_offline(struct cgroup_subsys_state *css) 7178{ 7179 struct task_group *tg = css_tg(css); 7180 7181 sched_offline_group(tg); 7182} 7183 7184static int cpu_cgroup_can_attach(struct cgroup_subsys_state *css, 7185 struct cgroup_taskset *tset) 7186{ 7187 struct task_struct *task; 7188 7189 cgroup_taskset_for_each(task, css, tset) { 7190#ifdef CONFIG_RT_GROUP_SCHED 7191 if (!sched_rt_can_attach(css_tg(css), task)) 7192 return -EINVAL; 7193#else 7194 /* We don't support RT-tasks being in separate groups */ 7195 if (task->sched_class != &fair_sched_class) 7196 return -EINVAL; 7197#endif 7198 } 7199 return 0; 7200} 7201 7202static void cpu_cgroup_attach(struct cgroup_subsys_state *css, 7203 struct cgroup_taskset *tset) 7204{ 7205 struct task_struct *task; 7206 7207 cgroup_taskset_for_each(task, css, tset) 7208 sched_move_task(task); 7209} 7210 7211static void cpu_cgroup_exit(struct cgroup_subsys_state *css, 7212 struct cgroup_subsys_state *old_css, 7213 struct task_struct *task) 7214{ 7215 /* 7216 * cgroup_exit() is called in the copy_process() failure path. 7217 * Ignore this case since the task hasn't ran yet, this avoids 7218 * trying to poke a half freed task state from generic code. 7219 */ 7220 if (!(task->flags & PF_EXITING)) 7221 return; 7222 7223 sched_move_task(task); 7224} 7225 7226#ifdef CONFIG_FAIR_GROUP_SCHED 7227static int cpu_shares_write_u64(struct cgroup_subsys_state *css, 7228 struct cftype *cftype, u64 shareval) 7229{ 7230 return sched_group_set_shares(css_tg(css), scale_load(shareval)); 7231} 7232 7233static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css, 7234 struct cftype *cft) 7235{ 7236 struct task_group *tg = css_tg(css); 7237 7238 return (u64) scale_load_down(tg->shares); 7239} 7240 7241#ifdef CONFIG_CFS_BANDWIDTH 7242static DEFINE_MUTEX(cfs_constraints_mutex); 7243 7244const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */ 7245const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */ 7246 7247static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime); 7248 7249static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota) 7250{ 7251 int i, ret = 0, runtime_enabled, runtime_was_enabled; 7252 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth; 7253 7254 if (tg == &root_task_group) 7255 return -EINVAL; 7256 7257 /* 7258 * Ensure we have at some amount of bandwidth every period. This is 7259 * to prevent reaching a state of large arrears when throttled via 7260 * entity_tick() resulting in prolonged exit starvation. 7261 */ 7262 if (quota < min_cfs_quota_period || period < min_cfs_quota_period) 7263 return -EINVAL; 7264 7265 /* 7266 * Likewise, bound things on the otherside by preventing insane quota 7267 * periods. This also allows us to normalize in computing quota 7268 * feasibility. 7269 */ 7270 if (period > max_cfs_quota_period) 7271 return -EINVAL; 7272 7273 mutex_lock(&cfs_constraints_mutex); 7274 ret = __cfs_schedulable(tg, period, quota); 7275 if (ret) 7276 goto out_unlock; 7277 7278 runtime_enabled = quota != RUNTIME_INF; 7279 runtime_was_enabled = cfs_b->quota != RUNTIME_INF; 7280 account_cfs_bandwidth_used(runtime_enabled, runtime_was_enabled); 7281 raw_spin_lock_irq(&cfs_b->lock); 7282 cfs_b->period = ns_to_ktime(period); 7283 cfs_b->quota = quota; 7284 7285 __refill_cfs_bandwidth_runtime(cfs_b); 7286 /* restart the period timer (if active) to handle new period expiry */ 7287 if (runtime_enabled && cfs_b->timer_active) { 7288 /* force a reprogram */ 7289 cfs_b->timer_active = 0; 7290 __start_cfs_bandwidth(cfs_b); 7291 } 7292 raw_spin_unlock_irq(&cfs_b->lock); 7293 7294 for_each_possible_cpu(i) { 7295 struct cfs_rq *cfs_rq = tg->cfs_rq[i]; 7296 struct rq *rq = cfs_rq->rq; 7297 7298 raw_spin_lock_irq(&rq->lock); 7299 cfs_rq->runtime_enabled = runtime_enabled; 7300 cfs_rq->runtime_remaining = 0; 7301 7302 if (cfs_rq->throttled) 7303 unthrottle_cfs_rq(cfs_rq); 7304 raw_spin_unlock_irq(&rq->lock); 7305 } 7306out_unlock: 7307 mutex_unlock(&cfs_constraints_mutex); 7308 7309 return ret; 7310} 7311 7312int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us) 7313{ 7314 u64 quota, period; 7315 7316 period = ktime_to_ns(tg->cfs_bandwidth.period); 7317 if (cfs_quota_us < 0) 7318 quota = RUNTIME_INF; 7319 else 7320 quota = (u64)cfs_quota_us * NSEC_PER_USEC; 7321 7322 return tg_set_cfs_bandwidth(tg, period, quota); 7323} 7324 7325long tg_get_cfs_quota(struct task_group *tg) 7326{ 7327 u64 quota_us; 7328 7329 if (tg->cfs_bandwidth.quota == RUNTIME_INF) 7330 return -1; 7331 7332 quota_us = tg->cfs_bandwidth.quota; 7333 do_div(quota_us, NSEC_PER_USEC); 7334 7335 return quota_us; 7336} 7337 7338int tg_set_cfs_period(struct task_group *tg, long cfs_period_us) 7339{ 7340 u64 quota, period; 7341 7342 period = (u64)cfs_period_us * NSEC_PER_USEC; 7343 quota = tg->cfs_bandwidth.quota; 7344 7345 return tg_set_cfs_bandwidth(tg, period, quota); 7346} 7347 7348long tg_get_cfs_period(struct task_group *tg) 7349{ 7350 u64 cfs_period_us; 7351 7352 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period); 7353 do_div(cfs_period_us, NSEC_PER_USEC); 7354 7355 return cfs_period_us; 7356} 7357 7358static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css, 7359 struct cftype *cft) 7360{ 7361 return tg_get_cfs_quota(css_tg(css)); 7362} 7363 7364static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css, 7365 struct cftype *cftype, s64 cfs_quota_us) 7366{ 7367 return tg_set_cfs_quota(css_tg(css), cfs_quota_us); 7368} 7369 7370static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css, 7371 struct cftype *cft) 7372{ 7373 return tg_get_cfs_period(css_tg(css)); 7374} 7375 7376static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css, 7377 struct cftype *cftype, u64 cfs_period_us) 7378{ 7379 return tg_set_cfs_period(css_tg(css), cfs_period_us); 7380} 7381 7382struct cfs_schedulable_data { 7383 struct task_group *tg; 7384 u64 period, quota; 7385}; 7386 7387/* 7388 * normalize group quota/period to be quota/max_period 7389 * note: units are usecs 7390 */ 7391static u64 normalize_cfs_quota(struct task_group *tg, 7392 struct cfs_schedulable_data *d) 7393{ 7394 u64 quota, period; 7395 7396 if (tg == d->tg) { 7397 period = d->period; 7398 quota = d->quota; 7399 } else { 7400 period = tg_get_cfs_period(tg); 7401 quota = tg_get_cfs_quota(tg); 7402 } 7403 7404 /* note: these should typically be equivalent */ 7405 if (quota == RUNTIME_INF || quota == -1) 7406 return RUNTIME_INF; 7407 7408 return to_ratio(period, quota); 7409} 7410 7411static int tg_cfs_schedulable_down(struct task_group *tg, void *data) 7412{ 7413 struct cfs_schedulable_data *d = data; 7414 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth; 7415 s64 quota = 0, parent_quota = -1; 7416 7417 if (!tg->parent) { 7418 quota = RUNTIME_INF; 7419 } else { 7420 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth; 7421 7422 quota = normalize_cfs_quota(tg, d); 7423 parent_quota = parent_b->hierarchal_quota; 7424 7425 /* 7426 * ensure max(child_quota) <= parent_quota, inherit when no 7427 * limit is set 7428 */ 7429 if (quota == RUNTIME_INF) 7430 quota = parent_quota; 7431 else if (parent_quota != RUNTIME_INF && quota > parent_quota) 7432 return -EINVAL; 7433 } 7434 cfs_b->hierarchal_quota = quota; 7435 7436 return 0; 7437} 7438 7439static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota) 7440{ 7441 int ret; 7442 struct cfs_schedulable_data data = { 7443 .tg = tg, 7444 .period = period, 7445 .quota = quota, 7446 }; 7447 7448 if (quota != RUNTIME_INF) { 7449 do_div(data.period, NSEC_PER_USEC); 7450 do_div(data.quota, NSEC_PER_USEC); 7451 } 7452 7453 rcu_read_lock(); 7454 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data); 7455 rcu_read_unlock(); 7456 7457 return ret; 7458} 7459 7460static int cpu_stats_show(struct cgroup_subsys_state *css, struct cftype *cft, 7461 struct cgroup_map_cb *cb) 7462{ 7463 struct task_group *tg = css_tg(css); 7464 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth; 7465 7466 cb->fill(cb, "nr_periods", cfs_b->nr_periods); 7467 cb->fill(cb, "nr_throttled", cfs_b->nr_throttled); 7468 cb->fill(cb, "throttled_time", cfs_b->throttled_time); 7469 7470 return 0; 7471} 7472#endif /* CONFIG_CFS_BANDWIDTH */ 7473#endif /* CONFIG_FAIR_GROUP_SCHED */ 7474 7475#ifdef CONFIG_RT_GROUP_SCHED 7476static int cpu_rt_runtime_write(struct cgroup_subsys_state *css, 7477 struct cftype *cft, s64 val) 7478{ 7479 return sched_group_set_rt_runtime(css_tg(css), val); 7480} 7481 7482static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css, 7483 struct cftype *cft) 7484{ 7485 return sched_group_rt_runtime(css_tg(css)); 7486} 7487 7488static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css, 7489 struct cftype *cftype, u64 rt_period_us) 7490{ 7491 return sched_group_set_rt_period(css_tg(css), rt_period_us); 7492} 7493 7494static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css, 7495 struct cftype *cft) 7496{ 7497 return sched_group_rt_period(css_tg(css)); 7498} 7499#endif /* CONFIG_RT_GROUP_SCHED */ 7500 7501static struct cftype cpu_files[] = { 7502#ifdef CONFIG_FAIR_GROUP_SCHED 7503 { 7504 .name = "shares", 7505 .read_u64 = cpu_shares_read_u64, 7506 .write_u64 = cpu_shares_write_u64, 7507 }, 7508#endif 7509#ifdef CONFIG_CFS_BANDWIDTH 7510 { 7511 .name = "cfs_quota_us", 7512 .read_s64 = cpu_cfs_quota_read_s64, 7513 .write_s64 = cpu_cfs_quota_write_s64, 7514 }, 7515 { 7516 .name = "cfs_period_us", 7517 .read_u64 = cpu_cfs_period_read_u64, 7518 .write_u64 = cpu_cfs_period_write_u64, 7519 }, 7520 { 7521 .name = "stat", 7522 .read_map = cpu_stats_show, 7523 }, 7524#endif 7525#ifdef CONFIG_RT_GROUP_SCHED 7526 { 7527 .name = "rt_runtime_us", 7528 .read_s64 = cpu_rt_runtime_read, 7529 .write_s64 = cpu_rt_runtime_write, 7530 }, 7531 { 7532 .name = "rt_period_us", 7533 .read_u64 = cpu_rt_period_read_uint, 7534 .write_u64 = cpu_rt_period_write_uint, 7535 }, 7536#endif 7537 { } /* terminate */ 7538}; 7539 7540struct cgroup_subsys cpu_cgroup_subsys = { 7541 .name = "cpu", 7542 .css_alloc = cpu_cgroup_css_alloc, 7543 .css_free = cpu_cgroup_css_free, 7544 .css_online = cpu_cgroup_css_online, 7545 .css_offline = cpu_cgroup_css_offline, 7546 .can_attach = cpu_cgroup_can_attach, 7547 .attach = cpu_cgroup_attach, 7548 .exit = cpu_cgroup_exit, 7549 .subsys_id = cpu_cgroup_subsys_id, 7550 .base_cftypes = cpu_files, 7551 .early_init = 1, 7552}; 7553 7554#endif /* CONFIG_CGROUP_SCHED */ 7555 7556void dump_cpu_task(int cpu) 7557{ 7558 pr_info("Task dump for CPU %d:\n", cpu); 7559 sched_show_task(cpu_curr(cpu)); 7560} 7561