core.c revision 1e3646ffc64b232cb14a5ef01d7b98997c1b73f9
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 1016static void __migrate_swap_task(struct task_struct *p, int cpu) 1017{ 1018 if (p->on_rq) { 1019 struct rq *src_rq, *dst_rq; 1020 1021 src_rq = task_rq(p); 1022 dst_rq = cpu_rq(cpu); 1023 1024 deactivate_task(src_rq, p, 0); 1025 set_task_cpu(p, cpu); 1026 activate_task(dst_rq, p, 0); 1027 check_preempt_curr(dst_rq, p, 0); 1028 } else { 1029 /* 1030 * Task isn't running anymore; make it appear like we migrated 1031 * it before it went to sleep. This means on wakeup we make the 1032 * previous cpu our targer instead of where it really is. 1033 */ 1034 p->wake_cpu = cpu; 1035 } 1036} 1037 1038struct migration_swap_arg { 1039 struct task_struct *src_task, *dst_task; 1040 int src_cpu, dst_cpu; 1041}; 1042 1043static int migrate_swap_stop(void *data) 1044{ 1045 struct migration_swap_arg *arg = data; 1046 struct rq *src_rq, *dst_rq; 1047 int ret = -EAGAIN; 1048 1049 src_rq = cpu_rq(arg->src_cpu); 1050 dst_rq = cpu_rq(arg->dst_cpu); 1051 1052 double_rq_lock(src_rq, dst_rq); 1053 if (task_cpu(arg->dst_task) != arg->dst_cpu) 1054 goto unlock; 1055 1056 if (task_cpu(arg->src_task) != arg->src_cpu) 1057 goto unlock; 1058 1059 if (!cpumask_test_cpu(arg->dst_cpu, tsk_cpus_allowed(arg->src_task))) 1060 goto unlock; 1061 1062 if (!cpumask_test_cpu(arg->src_cpu, tsk_cpus_allowed(arg->dst_task))) 1063 goto unlock; 1064 1065 __migrate_swap_task(arg->src_task, arg->dst_cpu); 1066 __migrate_swap_task(arg->dst_task, arg->src_cpu); 1067 1068 ret = 0; 1069 1070unlock: 1071 double_rq_unlock(src_rq, dst_rq); 1072 1073 return ret; 1074} 1075 1076/* 1077 * Cross migrate two tasks 1078 */ 1079int migrate_swap(struct task_struct *cur, struct task_struct *p) 1080{ 1081 struct migration_swap_arg arg; 1082 int ret = -EINVAL; 1083 1084 get_online_cpus(); 1085 1086 arg = (struct migration_swap_arg){ 1087 .src_task = cur, 1088 .src_cpu = task_cpu(cur), 1089 .dst_task = p, 1090 .dst_cpu = task_cpu(p), 1091 }; 1092 1093 if (arg.src_cpu == arg.dst_cpu) 1094 goto out; 1095 1096 if (!cpu_active(arg.src_cpu) || !cpu_active(arg.dst_cpu)) 1097 goto out; 1098 1099 if (!cpumask_test_cpu(arg.dst_cpu, tsk_cpus_allowed(arg.src_task))) 1100 goto out; 1101 1102 if (!cpumask_test_cpu(arg.src_cpu, tsk_cpus_allowed(arg.dst_task))) 1103 goto out; 1104 1105 ret = stop_two_cpus(arg.dst_cpu, arg.src_cpu, migrate_swap_stop, &arg); 1106 1107out: 1108 put_online_cpus(); 1109 return ret; 1110} 1111 1112struct migration_arg { 1113 struct task_struct *task; 1114 int dest_cpu; 1115}; 1116 1117static int migration_cpu_stop(void *data); 1118 1119/* 1120 * wait_task_inactive - wait for a thread to unschedule. 1121 * 1122 * If @match_state is nonzero, it's the @p->state value just checked and 1123 * not expected to change. If it changes, i.e. @p might have woken up, 1124 * then return zero. When we succeed in waiting for @p to be off its CPU, 1125 * we return a positive number (its total switch count). If a second call 1126 * a short while later returns the same number, the caller can be sure that 1127 * @p has remained unscheduled the whole time. 1128 * 1129 * The caller must ensure that the task *will* unschedule sometime soon, 1130 * else this function might spin for a *long* time. This function can't 1131 * be called with interrupts off, or it may introduce deadlock with 1132 * smp_call_function() if an IPI is sent by the same process we are 1133 * waiting to become inactive. 1134 */ 1135unsigned long wait_task_inactive(struct task_struct *p, long match_state) 1136{ 1137 unsigned long flags; 1138 int running, on_rq; 1139 unsigned long ncsw; 1140 struct rq *rq; 1141 1142 for (;;) { 1143 /* 1144 * We do the initial early heuristics without holding 1145 * any task-queue locks at all. We'll only try to get 1146 * the runqueue lock when things look like they will 1147 * work out! 1148 */ 1149 rq = task_rq(p); 1150 1151 /* 1152 * If the task is actively running on another CPU 1153 * still, just relax and busy-wait without holding 1154 * any locks. 1155 * 1156 * NOTE! Since we don't hold any locks, it's not 1157 * even sure that "rq" stays as the right runqueue! 1158 * But we don't care, since "task_running()" will 1159 * return false if the runqueue has changed and p 1160 * is actually now running somewhere else! 1161 */ 1162 while (task_running(rq, p)) { 1163 if (match_state && unlikely(p->state != match_state)) 1164 return 0; 1165 cpu_relax(); 1166 } 1167 1168 /* 1169 * Ok, time to look more closely! We need the rq 1170 * lock now, to be *sure*. If we're wrong, we'll 1171 * just go back and repeat. 1172 */ 1173 rq = task_rq_lock(p, &flags); 1174 trace_sched_wait_task(p); 1175 running = task_running(rq, p); 1176 on_rq = p->on_rq; 1177 ncsw = 0; 1178 if (!match_state || p->state == match_state) 1179 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */ 1180 task_rq_unlock(rq, p, &flags); 1181 1182 /* 1183 * If it changed from the expected state, bail out now. 1184 */ 1185 if (unlikely(!ncsw)) 1186 break; 1187 1188 /* 1189 * Was it really running after all now that we 1190 * checked with the proper locks actually held? 1191 * 1192 * Oops. Go back and try again.. 1193 */ 1194 if (unlikely(running)) { 1195 cpu_relax(); 1196 continue; 1197 } 1198 1199 /* 1200 * It's not enough that it's not actively running, 1201 * it must be off the runqueue _entirely_, and not 1202 * preempted! 1203 * 1204 * So if it was still runnable (but just not actively 1205 * running right now), it's preempted, and we should 1206 * yield - it could be a while. 1207 */ 1208 if (unlikely(on_rq)) { 1209 ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ); 1210 1211 set_current_state(TASK_UNINTERRUPTIBLE); 1212 schedule_hrtimeout(&to, HRTIMER_MODE_REL); 1213 continue; 1214 } 1215 1216 /* 1217 * Ahh, all good. It wasn't running, and it wasn't 1218 * runnable, which means that it will never become 1219 * running in the future either. We're all done! 1220 */ 1221 break; 1222 } 1223 1224 return ncsw; 1225} 1226 1227/*** 1228 * kick_process - kick a running thread to enter/exit the kernel 1229 * @p: the to-be-kicked thread 1230 * 1231 * Cause a process which is running on another CPU to enter 1232 * kernel-mode, without any delay. (to get signals handled.) 1233 * 1234 * NOTE: this function doesn't have to take the runqueue lock, 1235 * because all it wants to ensure is that the remote task enters 1236 * the kernel. If the IPI races and the task has been migrated 1237 * to another CPU then no harm is done and the purpose has been 1238 * achieved as well. 1239 */ 1240void kick_process(struct task_struct *p) 1241{ 1242 int cpu; 1243 1244 preempt_disable(); 1245 cpu = task_cpu(p); 1246 if ((cpu != smp_processor_id()) && task_curr(p)) 1247 smp_send_reschedule(cpu); 1248 preempt_enable(); 1249} 1250EXPORT_SYMBOL_GPL(kick_process); 1251#endif /* CONFIG_SMP */ 1252 1253#ifdef CONFIG_SMP 1254/* 1255 * ->cpus_allowed is protected by both rq->lock and p->pi_lock 1256 */ 1257static int select_fallback_rq(int cpu, struct task_struct *p) 1258{ 1259 int nid = cpu_to_node(cpu); 1260 const struct cpumask *nodemask = NULL; 1261 enum { cpuset, possible, fail } state = cpuset; 1262 int dest_cpu; 1263 1264 /* 1265 * If the node that the cpu is on has been offlined, cpu_to_node() 1266 * will return -1. There is no cpu on the node, and we should 1267 * select the cpu on the other node. 1268 */ 1269 if (nid != -1) { 1270 nodemask = cpumask_of_node(nid); 1271 1272 /* Look for allowed, online CPU in same node. */ 1273 for_each_cpu(dest_cpu, nodemask) { 1274 if (!cpu_online(dest_cpu)) 1275 continue; 1276 if (!cpu_active(dest_cpu)) 1277 continue; 1278 if (cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p))) 1279 return dest_cpu; 1280 } 1281 } 1282 1283 for (;;) { 1284 /* Any allowed, online CPU? */ 1285 for_each_cpu(dest_cpu, tsk_cpus_allowed(p)) { 1286 if (!cpu_online(dest_cpu)) 1287 continue; 1288 if (!cpu_active(dest_cpu)) 1289 continue; 1290 goto out; 1291 } 1292 1293 switch (state) { 1294 case cpuset: 1295 /* No more Mr. Nice Guy. */ 1296 cpuset_cpus_allowed_fallback(p); 1297 state = possible; 1298 break; 1299 1300 case possible: 1301 do_set_cpus_allowed(p, cpu_possible_mask); 1302 state = fail; 1303 break; 1304 1305 case fail: 1306 BUG(); 1307 break; 1308 } 1309 } 1310 1311out: 1312 if (state != cpuset) { 1313 /* 1314 * Don't tell them about moving exiting tasks or 1315 * kernel threads (both mm NULL), since they never 1316 * leave kernel. 1317 */ 1318 if (p->mm && printk_ratelimit()) { 1319 printk_sched("process %d (%s) no longer affine to cpu%d\n", 1320 task_pid_nr(p), p->comm, cpu); 1321 } 1322 } 1323 1324 return dest_cpu; 1325} 1326 1327/* 1328 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable. 1329 */ 1330static inline 1331int select_task_rq(struct task_struct *p, int cpu, int sd_flags, int wake_flags) 1332{ 1333 cpu = p->sched_class->select_task_rq(p, cpu, sd_flags, wake_flags); 1334 1335 /* 1336 * In order not to call set_task_cpu() on a blocking task we need 1337 * to rely on ttwu() to place the task on a valid ->cpus_allowed 1338 * cpu. 1339 * 1340 * Since this is common to all placement strategies, this lives here. 1341 * 1342 * [ this allows ->select_task() to simply return task_cpu(p) and 1343 * not worry about this generic constraint ] 1344 */ 1345 if (unlikely(!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)) || 1346 !cpu_online(cpu))) 1347 cpu = select_fallback_rq(task_cpu(p), p); 1348 1349 return cpu; 1350} 1351 1352static void update_avg(u64 *avg, u64 sample) 1353{ 1354 s64 diff = sample - *avg; 1355 *avg += diff >> 3; 1356} 1357#endif 1358 1359static void 1360ttwu_stat(struct task_struct *p, int cpu, int wake_flags) 1361{ 1362#ifdef CONFIG_SCHEDSTATS 1363 struct rq *rq = this_rq(); 1364 1365#ifdef CONFIG_SMP 1366 int this_cpu = smp_processor_id(); 1367 1368 if (cpu == this_cpu) { 1369 schedstat_inc(rq, ttwu_local); 1370 schedstat_inc(p, se.statistics.nr_wakeups_local); 1371 } else { 1372 struct sched_domain *sd; 1373 1374 schedstat_inc(p, se.statistics.nr_wakeups_remote); 1375 rcu_read_lock(); 1376 for_each_domain(this_cpu, sd) { 1377 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) { 1378 schedstat_inc(sd, ttwu_wake_remote); 1379 break; 1380 } 1381 } 1382 rcu_read_unlock(); 1383 } 1384 1385 if (wake_flags & WF_MIGRATED) 1386 schedstat_inc(p, se.statistics.nr_wakeups_migrate); 1387 1388#endif /* CONFIG_SMP */ 1389 1390 schedstat_inc(rq, ttwu_count); 1391 schedstat_inc(p, se.statistics.nr_wakeups); 1392 1393 if (wake_flags & WF_SYNC) 1394 schedstat_inc(p, se.statistics.nr_wakeups_sync); 1395 1396#endif /* CONFIG_SCHEDSTATS */ 1397} 1398 1399static void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags) 1400{ 1401 activate_task(rq, p, en_flags); 1402 p->on_rq = 1; 1403 1404 /* if a worker is waking up, notify workqueue */ 1405 if (p->flags & PF_WQ_WORKER) 1406 wq_worker_waking_up(p, cpu_of(rq)); 1407} 1408 1409/* 1410 * Mark the task runnable and perform wakeup-preemption. 1411 */ 1412static void 1413ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags) 1414{ 1415 check_preempt_curr(rq, p, wake_flags); 1416 trace_sched_wakeup(p, true); 1417 1418 p->state = TASK_RUNNING; 1419#ifdef CONFIG_SMP 1420 if (p->sched_class->task_woken) 1421 p->sched_class->task_woken(rq, p); 1422 1423 if (rq->idle_stamp) { 1424 u64 delta = rq_clock(rq) - rq->idle_stamp; 1425 u64 max = 2*rq->max_idle_balance_cost; 1426 1427 update_avg(&rq->avg_idle, delta); 1428 1429 if (rq->avg_idle > max) 1430 rq->avg_idle = max; 1431 1432 rq->idle_stamp = 0; 1433 } 1434#endif 1435} 1436 1437static void 1438ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags) 1439{ 1440#ifdef CONFIG_SMP 1441 if (p->sched_contributes_to_load) 1442 rq->nr_uninterruptible--; 1443#endif 1444 1445 ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_WAKING); 1446 ttwu_do_wakeup(rq, p, wake_flags); 1447} 1448 1449/* 1450 * Called in case the task @p isn't fully descheduled from its runqueue, 1451 * in this case we must do a remote wakeup. Its a 'light' wakeup though, 1452 * since all we need to do is flip p->state to TASK_RUNNING, since 1453 * the task is still ->on_rq. 1454 */ 1455static int ttwu_remote(struct task_struct *p, int wake_flags) 1456{ 1457 struct rq *rq; 1458 int ret = 0; 1459 1460 rq = __task_rq_lock(p); 1461 if (p->on_rq) { 1462 /* check_preempt_curr() may use rq clock */ 1463 update_rq_clock(rq); 1464 ttwu_do_wakeup(rq, p, wake_flags); 1465 ret = 1; 1466 } 1467 __task_rq_unlock(rq); 1468 1469 return ret; 1470} 1471 1472#ifdef CONFIG_SMP 1473static void sched_ttwu_pending(void) 1474{ 1475 struct rq *rq = this_rq(); 1476 struct llist_node *llist = llist_del_all(&rq->wake_list); 1477 struct task_struct *p; 1478 1479 raw_spin_lock(&rq->lock); 1480 1481 while (llist) { 1482 p = llist_entry(llist, struct task_struct, wake_entry); 1483 llist = llist_next(llist); 1484 ttwu_do_activate(rq, p, 0); 1485 } 1486 1487 raw_spin_unlock(&rq->lock); 1488} 1489 1490void scheduler_ipi(void) 1491{ 1492 /* 1493 * Fold TIF_NEED_RESCHED into the preempt_count; anybody setting 1494 * TIF_NEED_RESCHED remotely (for the first time) will also send 1495 * this IPI. 1496 */ 1497 if (tif_need_resched()) 1498 set_preempt_need_resched(); 1499 1500 if (llist_empty(&this_rq()->wake_list) 1501 && !tick_nohz_full_cpu(smp_processor_id()) 1502 && !got_nohz_idle_kick()) 1503 return; 1504 1505 /* 1506 * Not all reschedule IPI handlers call irq_enter/irq_exit, since 1507 * traditionally all their work was done from the interrupt return 1508 * path. Now that we actually do some work, we need to make sure 1509 * we do call them. 1510 * 1511 * Some archs already do call them, luckily irq_enter/exit nest 1512 * properly. 1513 * 1514 * Arguably we should visit all archs and update all handlers, 1515 * however a fair share of IPIs are still resched only so this would 1516 * somewhat pessimize the simple resched case. 1517 */ 1518 irq_enter(); 1519 tick_nohz_full_check(); 1520 sched_ttwu_pending(); 1521 1522 /* 1523 * Check if someone kicked us for doing the nohz idle load balance. 1524 */ 1525 if (unlikely(got_nohz_idle_kick())) { 1526 this_rq()->idle_balance = 1; 1527 raise_softirq_irqoff(SCHED_SOFTIRQ); 1528 } 1529 irq_exit(); 1530} 1531 1532static void ttwu_queue_remote(struct task_struct *p, int cpu) 1533{ 1534 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list)) 1535 smp_send_reschedule(cpu); 1536} 1537 1538bool cpus_share_cache(int this_cpu, int that_cpu) 1539{ 1540 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu); 1541} 1542#endif /* CONFIG_SMP */ 1543 1544static void ttwu_queue(struct task_struct *p, int cpu) 1545{ 1546 struct rq *rq = cpu_rq(cpu); 1547 1548#if defined(CONFIG_SMP) 1549 if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) { 1550 sched_clock_cpu(cpu); /* sync clocks x-cpu */ 1551 ttwu_queue_remote(p, cpu); 1552 return; 1553 } 1554#endif 1555 1556 raw_spin_lock(&rq->lock); 1557 ttwu_do_activate(rq, p, 0); 1558 raw_spin_unlock(&rq->lock); 1559} 1560 1561/** 1562 * try_to_wake_up - wake up a thread 1563 * @p: the thread to be awakened 1564 * @state: the mask of task states that can be woken 1565 * @wake_flags: wake modifier flags (WF_*) 1566 * 1567 * Put it on the run-queue if it's not already there. The "current" 1568 * thread is always on the run-queue (except when the actual 1569 * re-schedule is in progress), and as such you're allowed to do 1570 * the simpler "current->state = TASK_RUNNING" to mark yourself 1571 * runnable without the overhead of this. 1572 * 1573 * Return: %true if @p was woken up, %false if it was already running. 1574 * or @state didn't match @p's state. 1575 */ 1576static int 1577try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags) 1578{ 1579 unsigned long flags; 1580 int cpu, success = 0; 1581 1582 /* 1583 * If we are going to wake up a thread waiting for CONDITION we 1584 * need to ensure that CONDITION=1 done by the caller can not be 1585 * reordered with p->state check below. This pairs with mb() in 1586 * set_current_state() the waiting thread does. 1587 */ 1588 smp_mb__before_spinlock(); 1589 raw_spin_lock_irqsave(&p->pi_lock, flags); 1590 if (!(p->state & state)) 1591 goto out; 1592 1593 success = 1; /* we're going to change ->state */ 1594 cpu = task_cpu(p); 1595 1596 if (p->on_rq && ttwu_remote(p, wake_flags)) 1597 goto stat; 1598 1599#ifdef CONFIG_SMP 1600 /* 1601 * If the owning (remote) cpu is still in the middle of schedule() with 1602 * this task as prev, wait until its done referencing the task. 1603 */ 1604 while (p->on_cpu) 1605 cpu_relax(); 1606 /* 1607 * Pairs with the smp_wmb() in finish_lock_switch(). 1608 */ 1609 smp_rmb(); 1610 1611 p->sched_contributes_to_load = !!task_contributes_to_load(p); 1612 p->state = TASK_WAKING; 1613 1614 if (p->sched_class->task_waking) 1615 p->sched_class->task_waking(p); 1616 1617 cpu = select_task_rq(p, p->wake_cpu, SD_BALANCE_WAKE, wake_flags); 1618 if (task_cpu(p) != cpu) { 1619 wake_flags |= WF_MIGRATED; 1620 set_task_cpu(p, cpu); 1621 } 1622#endif /* CONFIG_SMP */ 1623 1624 ttwu_queue(p, cpu); 1625stat: 1626 ttwu_stat(p, cpu, wake_flags); 1627out: 1628 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 1629 1630 return success; 1631} 1632 1633/** 1634 * try_to_wake_up_local - try to wake up a local task with rq lock held 1635 * @p: the thread to be awakened 1636 * 1637 * Put @p on the run-queue if it's not already there. The caller must 1638 * ensure that this_rq() is locked, @p is bound to this_rq() and not 1639 * the current task. 1640 */ 1641static void try_to_wake_up_local(struct task_struct *p) 1642{ 1643 struct rq *rq = task_rq(p); 1644 1645 if (WARN_ON_ONCE(rq != this_rq()) || 1646 WARN_ON_ONCE(p == current)) 1647 return; 1648 1649 lockdep_assert_held(&rq->lock); 1650 1651 if (!raw_spin_trylock(&p->pi_lock)) { 1652 raw_spin_unlock(&rq->lock); 1653 raw_spin_lock(&p->pi_lock); 1654 raw_spin_lock(&rq->lock); 1655 } 1656 1657 if (!(p->state & TASK_NORMAL)) 1658 goto out; 1659 1660 if (!p->on_rq) 1661 ttwu_activate(rq, p, ENQUEUE_WAKEUP); 1662 1663 ttwu_do_wakeup(rq, p, 0); 1664 ttwu_stat(p, smp_processor_id(), 0); 1665out: 1666 raw_spin_unlock(&p->pi_lock); 1667} 1668 1669/** 1670 * wake_up_process - Wake up a specific process 1671 * @p: The process to be woken up. 1672 * 1673 * Attempt to wake up the nominated process and move it to the set of runnable 1674 * processes. 1675 * 1676 * Return: 1 if the process was woken up, 0 if it was already running. 1677 * 1678 * It may be assumed that this function implies a write memory barrier before 1679 * changing the task state if and only if any tasks are woken up. 1680 */ 1681int wake_up_process(struct task_struct *p) 1682{ 1683 WARN_ON(task_is_stopped_or_traced(p)); 1684 return try_to_wake_up(p, TASK_NORMAL, 0); 1685} 1686EXPORT_SYMBOL(wake_up_process); 1687 1688int wake_up_state(struct task_struct *p, unsigned int state) 1689{ 1690 return try_to_wake_up(p, state, 0); 1691} 1692 1693/* 1694 * Perform scheduler related setup for a newly forked process p. 1695 * p is forked by current. 1696 * 1697 * __sched_fork() is basic setup used by init_idle() too: 1698 */ 1699static void __sched_fork(unsigned long clone_flags, struct task_struct *p) 1700{ 1701 p->on_rq = 0; 1702 1703 p->se.on_rq = 0; 1704 p->se.exec_start = 0; 1705 p->se.sum_exec_runtime = 0; 1706 p->se.prev_sum_exec_runtime = 0; 1707 p->se.nr_migrations = 0; 1708 p->se.vruntime = 0; 1709 INIT_LIST_HEAD(&p->se.group_node); 1710 1711#ifdef CONFIG_SCHEDSTATS 1712 memset(&p->se.statistics, 0, sizeof(p->se.statistics)); 1713#endif 1714 1715 INIT_LIST_HEAD(&p->rt.run_list); 1716 1717#ifdef CONFIG_PREEMPT_NOTIFIERS 1718 INIT_HLIST_HEAD(&p->preempt_notifiers); 1719#endif 1720 1721#ifdef CONFIG_NUMA_BALANCING 1722 if (p->mm && atomic_read(&p->mm->mm_users) == 1) { 1723 p->mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay); 1724 p->mm->numa_scan_seq = 0; 1725 } 1726 1727 if (clone_flags & CLONE_VM) 1728 p->numa_preferred_nid = current->numa_preferred_nid; 1729 else 1730 p->numa_preferred_nid = -1; 1731 1732 p->node_stamp = 0ULL; 1733 p->numa_scan_seq = p->mm ? p->mm->numa_scan_seq : 0; 1734 p->numa_scan_period = sysctl_numa_balancing_scan_delay; 1735 p->numa_work.next = &p->numa_work; 1736 p->numa_faults = NULL; 1737 p->numa_faults_buffer = NULL; 1738 1739 INIT_LIST_HEAD(&p->numa_entry); 1740 p->numa_group = NULL; 1741#endif /* CONFIG_NUMA_BALANCING */ 1742} 1743 1744#ifdef CONFIG_NUMA_BALANCING 1745#ifdef CONFIG_SCHED_DEBUG 1746void set_numabalancing_state(bool enabled) 1747{ 1748 if (enabled) 1749 sched_feat_set("NUMA"); 1750 else 1751 sched_feat_set("NO_NUMA"); 1752} 1753#else 1754__read_mostly bool numabalancing_enabled; 1755 1756void set_numabalancing_state(bool enabled) 1757{ 1758 numabalancing_enabled = enabled; 1759} 1760#endif /* CONFIG_SCHED_DEBUG */ 1761#endif /* CONFIG_NUMA_BALANCING */ 1762 1763/* 1764 * fork()/clone()-time setup: 1765 */ 1766void sched_fork(unsigned long clone_flags, struct task_struct *p) 1767{ 1768 unsigned long flags; 1769 int cpu = get_cpu(); 1770 1771 __sched_fork(clone_flags, p); 1772 /* 1773 * We mark the process as running here. This guarantees that 1774 * nobody will actually run it, and a signal or other external 1775 * event cannot wake it up and insert it on the runqueue either. 1776 */ 1777 p->state = TASK_RUNNING; 1778 1779 /* 1780 * Make sure we do not leak PI boosting priority to the child. 1781 */ 1782 p->prio = current->normal_prio; 1783 1784 /* 1785 * Revert to default priority/policy on fork if requested. 1786 */ 1787 if (unlikely(p->sched_reset_on_fork)) { 1788 if (task_has_rt_policy(p)) { 1789 p->policy = SCHED_NORMAL; 1790 p->static_prio = NICE_TO_PRIO(0); 1791 p->rt_priority = 0; 1792 } else if (PRIO_TO_NICE(p->static_prio) < 0) 1793 p->static_prio = NICE_TO_PRIO(0); 1794 1795 p->prio = p->normal_prio = __normal_prio(p); 1796 set_load_weight(p); 1797 1798 /* 1799 * We don't need the reset flag anymore after the fork. It has 1800 * fulfilled its duty: 1801 */ 1802 p->sched_reset_on_fork = 0; 1803 } 1804 1805 if (!rt_prio(p->prio)) 1806 p->sched_class = &fair_sched_class; 1807 1808 if (p->sched_class->task_fork) 1809 p->sched_class->task_fork(p); 1810 1811 /* 1812 * The child is not yet in the pid-hash so no cgroup attach races, 1813 * and the cgroup is pinned to this child due to cgroup_fork() 1814 * is ran before sched_fork(). 1815 * 1816 * Silence PROVE_RCU. 1817 */ 1818 raw_spin_lock_irqsave(&p->pi_lock, flags); 1819 set_task_cpu(p, cpu); 1820 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 1821 1822#if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT) 1823 if (likely(sched_info_on())) 1824 memset(&p->sched_info, 0, sizeof(p->sched_info)); 1825#endif 1826#if defined(CONFIG_SMP) 1827 p->on_cpu = 0; 1828#endif 1829 init_task_preempt_count(p); 1830#ifdef CONFIG_SMP 1831 plist_node_init(&p->pushable_tasks, MAX_PRIO); 1832#endif 1833 1834 put_cpu(); 1835} 1836 1837/* 1838 * wake_up_new_task - wake up a newly created task for the first time. 1839 * 1840 * This function will do some initial scheduler statistics housekeeping 1841 * that must be done for every newly created context, then puts the task 1842 * on the runqueue and wakes it. 1843 */ 1844void wake_up_new_task(struct task_struct *p) 1845{ 1846 unsigned long flags; 1847 struct rq *rq; 1848 1849 raw_spin_lock_irqsave(&p->pi_lock, flags); 1850#ifdef CONFIG_SMP 1851 /* 1852 * Fork balancing, do it here and not earlier because: 1853 * - cpus_allowed can change in the fork path 1854 * - any previously selected cpu might disappear through hotplug 1855 */ 1856 set_task_cpu(p, select_task_rq(p, task_cpu(p), SD_BALANCE_FORK, 0)); 1857#endif 1858 1859 /* Initialize new task's runnable average */ 1860 init_task_runnable_average(p); 1861 rq = __task_rq_lock(p); 1862 activate_task(rq, p, 0); 1863 p->on_rq = 1; 1864 trace_sched_wakeup_new(p, true); 1865 check_preempt_curr(rq, p, WF_FORK); 1866#ifdef CONFIG_SMP 1867 if (p->sched_class->task_woken) 1868 p->sched_class->task_woken(rq, p); 1869#endif 1870 task_rq_unlock(rq, p, &flags); 1871} 1872 1873#ifdef CONFIG_PREEMPT_NOTIFIERS 1874 1875/** 1876 * preempt_notifier_register - tell me when current is being preempted & rescheduled 1877 * @notifier: notifier struct to register 1878 */ 1879void preempt_notifier_register(struct preempt_notifier *notifier) 1880{ 1881 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers); 1882} 1883EXPORT_SYMBOL_GPL(preempt_notifier_register); 1884 1885/** 1886 * preempt_notifier_unregister - no longer interested in preemption notifications 1887 * @notifier: notifier struct to unregister 1888 * 1889 * This is safe to call from within a preemption notifier. 1890 */ 1891void preempt_notifier_unregister(struct preempt_notifier *notifier) 1892{ 1893 hlist_del(¬ifier->link); 1894} 1895EXPORT_SYMBOL_GPL(preempt_notifier_unregister); 1896 1897static void fire_sched_in_preempt_notifiers(struct task_struct *curr) 1898{ 1899 struct preempt_notifier *notifier; 1900 1901 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link) 1902 notifier->ops->sched_in(notifier, raw_smp_processor_id()); 1903} 1904 1905static void 1906fire_sched_out_preempt_notifiers(struct task_struct *curr, 1907 struct task_struct *next) 1908{ 1909 struct preempt_notifier *notifier; 1910 1911 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link) 1912 notifier->ops->sched_out(notifier, next); 1913} 1914 1915#else /* !CONFIG_PREEMPT_NOTIFIERS */ 1916 1917static void fire_sched_in_preempt_notifiers(struct task_struct *curr) 1918{ 1919} 1920 1921static void 1922fire_sched_out_preempt_notifiers(struct task_struct *curr, 1923 struct task_struct *next) 1924{ 1925} 1926 1927#endif /* CONFIG_PREEMPT_NOTIFIERS */ 1928 1929/** 1930 * prepare_task_switch - prepare to switch tasks 1931 * @rq: the runqueue preparing to switch 1932 * @prev: the current task that is being switched out 1933 * @next: the task we are going to switch to. 1934 * 1935 * This is called with the rq lock held and interrupts off. It must 1936 * be paired with a subsequent finish_task_switch after the context 1937 * switch. 1938 * 1939 * prepare_task_switch sets up locking and calls architecture specific 1940 * hooks. 1941 */ 1942static inline void 1943prepare_task_switch(struct rq *rq, struct task_struct *prev, 1944 struct task_struct *next) 1945{ 1946 trace_sched_switch(prev, next); 1947 sched_info_switch(rq, prev, next); 1948 perf_event_task_sched_out(prev, next); 1949 fire_sched_out_preempt_notifiers(prev, next); 1950 prepare_lock_switch(rq, next); 1951 prepare_arch_switch(next); 1952} 1953 1954/** 1955 * finish_task_switch - clean up after a task-switch 1956 * @rq: runqueue associated with task-switch 1957 * @prev: the thread we just switched away from. 1958 * 1959 * finish_task_switch must be called after the context switch, paired 1960 * with a prepare_task_switch call before the context switch. 1961 * finish_task_switch will reconcile locking set up by prepare_task_switch, 1962 * and do any other architecture-specific cleanup actions. 1963 * 1964 * Note that we may have delayed dropping an mm in context_switch(). If 1965 * so, we finish that here outside of the runqueue lock. (Doing it 1966 * with the lock held can cause deadlocks; see schedule() for 1967 * details.) 1968 */ 1969static void finish_task_switch(struct rq *rq, struct task_struct *prev) 1970 __releases(rq->lock) 1971{ 1972 struct mm_struct *mm = rq->prev_mm; 1973 long prev_state; 1974 1975 rq->prev_mm = NULL; 1976 1977 /* 1978 * A task struct has one reference for the use as "current". 1979 * If a task dies, then it sets TASK_DEAD in tsk->state and calls 1980 * schedule one last time. The schedule call will never return, and 1981 * the scheduled task must drop that reference. 1982 * The test for TASK_DEAD must occur while the runqueue locks are 1983 * still held, otherwise prev could be scheduled on another cpu, die 1984 * there before we look at prev->state, and then the reference would 1985 * be dropped twice. 1986 * Manfred Spraul <manfred@colorfullife.com> 1987 */ 1988 prev_state = prev->state; 1989 vtime_task_switch(prev); 1990 finish_arch_switch(prev); 1991 perf_event_task_sched_in(prev, current); 1992 finish_lock_switch(rq, prev); 1993 finish_arch_post_lock_switch(); 1994 1995 fire_sched_in_preempt_notifiers(current); 1996 if (mm) 1997 mmdrop(mm); 1998 if (unlikely(prev_state == TASK_DEAD)) { 1999 task_numa_free(prev); 2000 2001 /* 2002 * Remove function-return probe instances associated with this 2003 * task and put them back on the free list. 2004 */ 2005 kprobe_flush_task(prev); 2006 put_task_struct(prev); 2007 } 2008 2009 tick_nohz_task_switch(current); 2010} 2011 2012#ifdef CONFIG_SMP 2013 2014/* assumes rq->lock is held */ 2015static inline void pre_schedule(struct rq *rq, struct task_struct *prev) 2016{ 2017 if (prev->sched_class->pre_schedule) 2018 prev->sched_class->pre_schedule(rq, prev); 2019} 2020 2021/* rq->lock is NOT held, but preemption is disabled */ 2022static inline void post_schedule(struct rq *rq) 2023{ 2024 if (rq->post_schedule) { 2025 unsigned long flags; 2026 2027 raw_spin_lock_irqsave(&rq->lock, flags); 2028 if (rq->curr->sched_class->post_schedule) 2029 rq->curr->sched_class->post_schedule(rq); 2030 raw_spin_unlock_irqrestore(&rq->lock, flags); 2031 2032 rq->post_schedule = 0; 2033 } 2034} 2035 2036#else 2037 2038static inline void pre_schedule(struct rq *rq, struct task_struct *p) 2039{ 2040} 2041 2042static inline void post_schedule(struct rq *rq) 2043{ 2044} 2045 2046#endif 2047 2048/** 2049 * schedule_tail - first thing a freshly forked thread must call. 2050 * @prev: the thread we just switched away from. 2051 */ 2052asmlinkage void schedule_tail(struct task_struct *prev) 2053 __releases(rq->lock) 2054{ 2055 struct rq *rq = this_rq(); 2056 2057 finish_task_switch(rq, prev); 2058 2059 /* 2060 * FIXME: do we need to worry about rq being invalidated by the 2061 * task_switch? 2062 */ 2063 post_schedule(rq); 2064 2065#ifdef __ARCH_WANT_UNLOCKED_CTXSW 2066 /* In this case, finish_task_switch does not reenable preemption */ 2067 preempt_enable(); 2068#endif 2069 if (current->set_child_tid) 2070 put_user(task_pid_vnr(current), current->set_child_tid); 2071} 2072 2073/* 2074 * context_switch - switch to the new MM and the new 2075 * thread's register state. 2076 */ 2077static inline void 2078context_switch(struct rq *rq, struct task_struct *prev, 2079 struct task_struct *next) 2080{ 2081 struct mm_struct *mm, *oldmm; 2082 2083 prepare_task_switch(rq, prev, next); 2084 2085 mm = next->mm; 2086 oldmm = prev->active_mm; 2087 /* 2088 * For paravirt, this is coupled with an exit in switch_to to 2089 * combine the page table reload and the switch backend into 2090 * one hypercall. 2091 */ 2092 arch_start_context_switch(prev); 2093 2094 if (!mm) { 2095 next->active_mm = oldmm; 2096 atomic_inc(&oldmm->mm_count); 2097 enter_lazy_tlb(oldmm, next); 2098 } else 2099 switch_mm(oldmm, mm, next); 2100 2101 if (!prev->mm) { 2102 prev->active_mm = NULL; 2103 rq->prev_mm = oldmm; 2104 } 2105 /* 2106 * Since the runqueue lock will be released by the next 2107 * task (which is an invalid locking op but in the case 2108 * of the scheduler it's an obvious special-case), so we 2109 * do an early lockdep release here: 2110 */ 2111#ifndef __ARCH_WANT_UNLOCKED_CTXSW 2112 spin_release(&rq->lock.dep_map, 1, _THIS_IP_); 2113#endif 2114 2115 context_tracking_task_switch(prev, next); 2116 /* Here we just switch the register state and the stack. */ 2117 switch_to(prev, next, prev); 2118 2119 barrier(); 2120 /* 2121 * this_rq must be evaluated again because prev may have moved 2122 * CPUs since it called schedule(), thus the 'rq' on its stack 2123 * frame will be invalid. 2124 */ 2125 finish_task_switch(this_rq(), prev); 2126} 2127 2128/* 2129 * nr_running and nr_context_switches: 2130 * 2131 * externally visible scheduler statistics: current number of runnable 2132 * threads, total number of context switches performed since bootup. 2133 */ 2134unsigned long nr_running(void) 2135{ 2136 unsigned long i, sum = 0; 2137 2138 for_each_online_cpu(i) 2139 sum += cpu_rq(i)->nr_running; 2140 2141 return sum; 2142} 2143 2144unsigned long long nr_context_switches(void) 2145{ 2146 int i; 2147 unsigned long long sum = 0; 2148 2149 for_each_possible_cpu(i) 2150 sum += cpu_rq(i)->nr_switches; 2151 2152 return sum; 2153} 2154 2155unsigned long nr_iowait(void) 2156{ 2157 unsigned long i, sum = 0; 2158 2159 for_each_possible_cpu(i) 2160 sum += atomic_read(&cpu_rq(i)->nr_iowait); 2161 2162 return sum; 2163} 2164 2165unsigned long nr_iowait_cpu(int cpu) 2166{ 2167 struct rq *this = cpu_rq(cpu); 2168 return atomic_read(&this->nr_iowait); 2169} 2170 2171#ifdef CONFIG_SMP 2172 2173/* 2174 * sched_exec - execve() is a valuable balancing opportunity, because at 2175 * this point the task has the smallest effective memory and cache footprint. 2176 */ 2177void sched_exec(void) 2178{ 2179 struct task_struct *p = current; 2180 unsigned long flags; 2181 int dest_cpu; 2182 2183 raw_spin_lock_irqsave(&p->pi_lock, flags); 2184 dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), SD_BALANCE_EXEC, 0); 2185 if (dest_cpu == smp_processor_id()) 2186 goto unlock; 2187 2188 if (likely(cpu_active(dest_cpu))) { 2189 struct migration_arg arg = { p, dest_cpu }; 2190 2191 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 2192 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg); 2193 return; 2194 } 2195unlock: 2196 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 2197} 2198 2199#endif 2200 2201DEFINE_PER_CPU(struct kernel_stat, kstat); 2202DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat); 2203 2204EXPORT_PER_CPU_SYMBOL(kstat); 2205EXPORT_PER_CPU_SYMBOL(kernel_cpustat); 2206 2207/* 2208 * Return any ns on the sched_clock that have not yet been accounted in 2209 * @p in case that task is currently running. 2210 * 2211 * Called with task_rq_lock() held on @rq. 2212 */ 2213static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq) 2214{ 2215 u64 ns = 0; 2216 2217 if (task_current(rq, p)) { 2218 update_rq_clock(rq); 2219 ns = rq_clock_task(rq) - p->se.exec_start; 2220 if ((s64)ns < 0) 2221 ns = 0; 2222 } 2223 2224 return ns; 2225} 2226 2227unsigned long long task_delta_exec(struct task_struct *p) 2228{ 2229 unsigned long flags; 2230 struct rq *rq; 2231 u64 ns = 0; 2232 2233 rq = task_rq_lock(p, &flags); 2234 ns = do_task_delta_exec(p, rq); 2235 task_rq_unlock(rq, p, &flags); 2236 2237 return ns; 2238} 2239 2240/* 2241 * Return accounted runtime for the task. 2242 * In case the task is currently running, return the runtime plus current's 2243 * pending runtime that have not been accounted yet. 2244 */ 2245unsigned long long task_sched_runtime(struct task_struct *p) 2246{ 2247 unsigned long flags; 2248 struct rq *rq; 2249 u64 ns = 0; 2250 2251 rq = task_rq_lock(p, &flags); 2252 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq); 2253 task_rq_unlock(rq, p, &flags); 2254 2255 return ns; 2256} 2257 2258/* 2259 * This function gets called by the timer code, with HZ frequency. 2260 * We call it with interrupts disabled. 2261 */ 2262void scheduler_tick(void) 2263{ 2264 int cpu = smp_processor_id(); 2265 struct rq *rq = cpu_rq(cpu); 2266 struct task_struct *curr = rq->curr; 2267 2268 sched_clock_tick(); 2269 2270 raw_spin_lock(&rq->lock); 2271 update_rq_clock(rq); 2272 curr->sched_class->task_tick(rq, curr, 0); 2273 update_cpu_load_active(rq); 2274 raw_spin_unlock(&rq->lock); 2275 2276 perf_event_task_tick(); 2277 2278#ifdef CONFIG_SMP 2279 rq->idle_balance = idle_cpu(cpu); 2280 trigger_load_balance(rq, cpu); 2281#endif 2282 rq_last_tick_reset(rq); 2283} 2284 2285#ifdef CONFIG_NO_HZ_FULL 2286/** 2287 * scheduler_tick_max_deferment 2288 * 2289 * Keep at least one tick per second when a single 2290 * active task is running because the scheduler doesn't 2291 * yet completely support full dynticks environment. 2292 * 2293 * This makes sure that uptime, CFS vruntime, load 2294 * balancing, etc... continue to move forward, even 2295 * with a very low granularity. 2296 * 2297 * Return: Maximum deferment in nanoseconds. 2298 */ 2299u64 scheduler_tick_max_deferment(void) 2300{ 2301 struct rq *rq = this_rq(); 2302 unsigned long next, now = ACCESS_ONCE(jiffies); 2303 2304 next = rq->last_sched_tick + HZ; 2305 2306 if (time_before_eq(next, now)) 2307 return 0; 2308 2309 return jiffies_to_usecs(next - now) * NSEC_PER_USEC; 2310} 2311#endif 2312 2313notrace unsigned long get_parent_ip(unsigned long addr) 2314{ 2315 if (in_lock_functions(addr)) { 2316 addr = CALLER_ADDR2; 2317 if (in_lock_functions(addr)) 2318 addr = CALLER_ADDR3; 2319 } 2320 return addr; 2321} 2322 2323#if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \ 2324 defined(CONFIG_PREEMPT_TRACER)) 2325 2326void __kprobes preempt_count_add(int val) 2327{ 2328#ifdef CONFIG_DEBUG_PREEMPT 2329 /* 2330 * Underflow? 2331 */ 2332 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0))) 2333 return; 2334#endif 2335 __preempt_count_add(val); 2336#ifdef CONFIG_DEBUG_PREEMPT 2337 /* 2338 * Spinlock count overflowing soon? 2339 */ 2340 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >= 2341 PREEMPT_MASK - 10); 2342#endif 2343 if (preempt_count() == val) 2344 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1)); 2345} 2346EXPORT_SYMBOL(preempt_count_add); 2347 2348void __kprobes preempt_count_sub(int val) 2349{ 2350#ifdef CONFIG_DEBUG_PREEMPT 2351 /* 2352 * Underflow? 2353 */ 2354 if (DEBUG_LOCKS_WARN_ON(val > preempt_count())) 2355 return; 2356 /* 2357 * Is the spinlock portion underflowing? 2358 */ 2359 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) && 2360 !(preempt_count() & PREEMPT_MASK))) 2361 return; 2362#endif 2363 2364 if (preempt_count() == val) 2365 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1)); 2366 __preempt_count_sub(val); 2367} 2368EXPORT_SYMBOL(preempt_count_sub); 2369 2370#endif 2371 2372/* 2373 * Print scheduling while atomic bug: 2374 */ 2375static noinline void __schedule_bug(struct task_struct *prev) 2376{ 2377 if (oops_in_progress) 2378 return; 2379 2380 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n", 2381 prev->comm, prev->pid, preempt_count()); 2382 2383 debug_show_held_locks(prev); 2384 print_modules(); 2385 if (irqs_disabled()) 2386 print_irqtrace_events(prev); 2387 dump_stack(); 2388 add_taint(TAINT_WARN, LOCKDEP_STILL_OK); 2389} 2390 2391/* 2392 * Various schedule()-time debugging checks and statistics: 2393 */ 2394static inline void schedule_debug(struct task_struct *prev) 2395{ 2396 /* 2397 * Test if we are atomic. Since do_exit() needs to call into 2398 * schedule() atomically, we ignore that path for now. 2399 * Otherwise, whine if we are scheduling when we should not be. 2400 */ 2401 if (unlikely(in_atomic_preempt_off() && !prev->exit_state)) 2402 __schedule_bug(prev); 2403 rcu_sleep_check(); 2404 2405 profile_hit(SCHED_PROFILING, __builtin_return_address(0)); 2406 2407 schedstat_inc(this_rq(), sched_count); 2408} 2409 2410static void put_prev_task(struct rq *rq, struct task_struct *prev) 2411{ 2412 if (prev->on_rq || rq->skip_clock_update < 0) 2413 update_rq_clock(rq); 2414 prev->sched_class->put_prev_task(rq, prev); 2415} 2416 2417/* 2418 * Pick up the highest-prio task: 2419 */ 2420static inline struct task_struct * 2421pick_next_task(struct rq *rq) 2422{ 2423 const struct sched_class *class; 2424 struct task_struct *p; 2425 2426 /* 2427 * Optimization: we know that if all tasks are in 2428 * the fair class we can call that function directly: 2429 */ 2430 if (likely(rq->nr_running == rq->cfs.h_nr_running)) { 2431 p = fair_sched_class.pick_next_task(rq); 2432 if (likely(p)) 2433 return p; 2434 } 2435 2436 for_each_class(class) { 2437 p = class->pick_next_task(rq); 2438 if (p) 2439 return p; 2440 } 2441 2442 BUG(); /* the idle class will always have a runnable task */ 2443} 2444 2445/* 2446 * __schedule() is the main scheduler function. 2447 * 2448 * The main means of driving the scheduler and thus entering this function are: 2449 * 2450 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc. 2451 * 2452 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return 2453 * paths. For example, see arch/x86/entry_64.S. 2454 * 2455 * To drive preemption between tasks, the scheduler sets the flag in timer 2456 * interrupt handler scheduler_tick(). 2457 * 2458 * 3. Wakeups don't really cause entry into schedule(). They add a 2459 * task to the run-queue and that's it. 2460 * 2461 * Now, if the new task added to the run-queue preempts the current 2462 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets 2463 * called on the nearest possible occasion: 2464 * 2465 * - If the kernel is preemptible (CONFIG_PREEMPT=y): 2466 * 2467 * - in syscall or exception context, at the next outmost 2468 * preempt_enable(). (this might be as soon as the wake_up()'s 2469 * spin_unlock()!) 2470 * 2471 * - in IRQ context, return from interrupt-handler to 2472 * preemptible context 2473 * 2474 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set) 2475 * then at the next: 2476 * 2477 * - cond_resched() call 2478 * - explicit schedule() call 2479 * - return from syscall or exception to user-space 2480 * - return from interrupt-handler to user-space 2481 */ 2482static void __sched __schedule(void) 2483{ 2484 struct task_struct *prev, *next; 2485 unsigned long *switch_count; 2486 struct rq *rq; 2487 int cpu; 2488 2489need_resched: 2490 preempt_disable(); 2491 cpu = smp_processor_id(); 2492 rq = cpu_rq(cpu); 2493 rcu_note_context_switch(cpu); 2494 prev = rq->curr; 2495 2496 schedule_debug(prev); 2497 2498 if (sched_feat(HRTICK)) 2499 hrtick_clear(rq); 2500 2501 /* 2502 * Make sure that signal_pending_state()->signal_pending() below 2503 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE) 2504 * done by the caller to avoid the race with signal_wake_up(). 2505 */ 2506 smp_mb__before_spinlock(); 2507 raw_spin_lock_irq(&rq->lock); 2508 2509 switch_count = &prev->nivcsw; 2510 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) { 2511 if (unlikely(signal_pending_state(prev->state, prev))) { 2512 prev->state = TASK_RUNNING; 2513 } else { 2514 deactivate_task(rq, prev, DEQUEUE_SLEEP); 2515 prev->on_rq = 0; 2516 2517 /* 2518 * If a worker went to sleep, notify and ask workqueue 2519 * whether it wants to wake up a task to maintain 2520 * concurrency. 2521 */ 2522 if (prev->flags & PF_WQ_WORKER) { 2523 struct task_struct *to_wakeup; 2524 2525 to_wakeup = wq_worker_sleeping(prev, cpu); 2526 if (to_wakeup) 2527 try_to_wake_up_local(to_wakeup); 2528 } 2529 } 2530 switch_count = &prev->nvcsw; 2531 } 2532 2533 pre_schedule(rq, prev); 2534 2535 if (unlikely(!rq->nr_running)) 2536 idle_balance(cpu, rq); 2537 2538 put_prev_task(rq, prev); 2539 next = pick_next_task(rq); 2540 clear_tsk_need_resched(prev); 2541 clear_preempt_need_resched(); 2542 rq->skip_clock_update = 0; 2543 2544 if (likely(prev != next)) { 2545 rq->nr_switches++; 2546 rq->curr = next; 2547 ++*switch_count; 2548 2549 context_switch(rq, prev, next); /* unlocks the rq */ 2550 /* 2551 * The context switch have flipped the stack from under us 2552 * and restored the local variables which were saved when 2553 * this task called schedule() in the past. prev == current 2554 * is still correct, but it can be moved to another cpu/rq. 2555 */ 2556 cpu = smp_processor_id(); 2557 rq = cpu_rq(cpu); 2558 } else 2559 raw_spin_unlock_irq(&rq->lock); 2560 2561 post_schedule(rq); 2562 2563 sched_preempt_enable_no_resched(); 2564 if (need_resched()) 2565 goto need_resched; 2566} 2567 2568static inline void sched_submit_work(struct task_struct *tsk) 2569{ 2570 if (!tsk->state || tsk_is_pi_blocked(tsk)) 2571 return; 2572 /* 2573 * If we are going to sleep and we have plugged IO queued, 2574 * make sure to submit it to avoid deadlocks. 2575 */ 2576 if (blk_needs_flush_plug(tsk)) 2577 blk_schedule_flush_plug(tsk); 2578} 2579 2580asmlinkage void __sched schedule(void) 2581{ 2582 struct task_struct *tsk = current; 2583 2584 sched_submit_work(tsk); 2585 __schedule(); 2586} 2587EXPORT_SYMBOL(schedule); 2588 2589#ifdef CONFIG_CONTEXT_TRACKING 2590asmlinkage void __sched schedule_user(void) 2591{ 2592 /* 2593 * If we come here after a random call to set_need_resched(), 2594 * or we have been woken up remotely but the IPI has not yet arrived, 2595 * we haven't yet exited the RCU idle mode. Do it here manually until 2596 * we find a better solution. 2597 */ 2598 user_exit(); 2599 schedule(); 2600 user_enter(); 2601} 2602#endif 2603 2604/** 2605 * schedule_preempt_disabled - called with preemption disabled 2606 * 2607 * Returns with preemption disabled. Note: preempt_count must be 1 2608 */ 2609void __sched schedule_preempt_disabled(void) 2610{ 2611 sched_preempt_enable_no_resched(); 2612 schedule(); 2613 preempt_disable(); 2614} 2615 2616#ifdef CONFIG_PREEMPT 2617/* 2618 * this is the entry point to schedule() from in-kernel preemption 2619 * off of preempt_enable. Kernel preemptions off return from interrupt 2620 * occur there and call schedule directly. 2621 */ 2622asmlinkage void __sched notrace preempt_schedule(void) 2623{ 2624 /* 2625 * If there is a non-zero preempt_count or interrupts are disabled, 2626 * we do not want to preempt the current task. Just return.. 2627 */ 2628 if (likely(!preemptible())) 2629 return; 2630 2631 do { 2632 __preempt_count_add(PREEMPT_ACTIVE); 2633 __schedule(); 2634 __preempt_count_sub(PREEMPT_ACTIVE); 2635 2636 /* 2637 * Check again in case we missed a preemption opportunity 2638 * between schedule and now. 2639 */ 2640 barrier(); 2641 } while (need_resched()); 2642} 2643EXPORT_SYMBOL(preempt_schedule); 2644 2645/* 2646 * this is the entry point to schedule() from kernel preemption 2647 * off of irq context. 2648 * Note, that this is called and return with irqs disabled. This will 2649 * protect us against recursive calling from irq. 2650 */ 2651asmlinkage void __sched preempt_schedule_irq(void) 2652{ 2653 enum ctx_state prev_state; 2654 2655 /* Catch callers which need to be fixed */ 2656 BUG_ON(preempt_count() || !irqs_disabled()); 2657 2658 prev_state = exception_enter(); 2659 2660 do { 2661 __preempt_count_add(PREEMPT_ACTIVE); 2662 local_irq_enable(); 2663 __schedule(); 2664 local_irq_disable(); 2665 __preempt_count_sub(PREEMPT_ACTIVE); 2666 2667 /* 2668 * Check again in case we missed a preemption opportunity 2669 * between schedule and now. 2670 */ 2671 barrier(); 2672 } while (need_resched()); 2673 2674 exception_exit(prev_state); 2675} 2676 2677#endif /* CONFIG_PREEMPT */ 2678 2679int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags, 2680 void *key) 2681{ 2682 return try_to_wake_up(curr->private, mode, wake_flags); 2683} 2684EXPORT_SYMBOL(default_wake_function); 2685 2686/* 2687 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just 2688 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve 2689 * number) then we wake all the non-exclusive tasks and one exclusive task. 2690 * 2691 * There are circumstances in which we can try to wake a task which has already 2692 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns 2693 * zero in this (rare) case, and we handle it by continuing to scan the queue. 2694 */ 2695static void __wake_up_common(wait_queue_head_t *q, unsigned int mode, 2696 int nr_exclusive, int wake_flags, void *key) 2697{ 2698 wait_queue_t *curr, *next; 2699 2700 list_for_each_entry_safe(curr, next, &q->task_list, task_list) { 2701 unsigned flags = curr->flags; 2702 2703 if (curr->func(curr, mode, wake_flags, key) && 2704 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive) 2705 break; 2706 } 2707} 2708 2709/** 2710 * __wake_up - wake up threads blocked on a waitqueue. 2711 * @q: the waitqueue 2712 * @mode: which threads 2713 * @nr_exclusive: how many wake-one or wake-many threads to wake up 2714 * @key: is directly passed to the wakeup function 2715 * 2716 * It may be assumed that this function implies a write memory barrier before 2717 * changing the task state if and only if any tasks are woken up. 2718 */ 2719void __wake_up(wait_queue_head_t *q, unsigned int mode, 2720 int nr_exclusive, void *key) 2721{ 2722 unsigned long flags; 2723 2724 spin_lock_irqsave(&q->lock, flags); 2725 __wake_up_common(q, mode, nr_exclusive, 0, key); 2726 spin_unlock_irqrestore(&q->lock, flags); 2727} 2728EXPORT_SYMBOL(__wake_up); 2729 2730/* 2731 * Same as __wake_up but called with the spinlock in wait_queue_head_t held. 2732 */ 2733void __wake_up_locked(wait_queue_head_t *q, unsigned int mode, int nr) 2734{ 2735 __wake_up_common(q, mode, nr, 0, NULL); 2736} 2737EXPORT_SYMBOL_GPL(__wake_up_locked); 2738 2739void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key) 2740{ 2741 __wake_up_common(q, mode, 1, 0, key); 2742} 2743EXPORT_SYMBOL_GPL(__wake_up_locked_key); 2744 2745/** 2746 * __wake_up_sync_key - wake up threads blocked on a waitqueue. 2747 * @q: the waitqueue 2748 * @mode: which threads 2749 * @nr_exclusive: how many wake-one or wake-many threads to wake up 2750 * @key: opaque value to be passed to wakeup targets 2751 * 2752 * The sync wakeup differs that the waker knows that it will schedule 2753 * away soon, so while the target thread will be woken up, it will not 2754 * be migrated to another CPU - ie. the two threads are 'synchronized' 2755 * with each other. This can prevent needless bouncing between CPUs. 2756 * 2757 * On UP it can prevent extra preemption. 2758 * 2759 * It may be assumed that this function implies a write memory barrier before 2760 * changing the task state if and only if any tasks are woken up. 2761 */ 2762void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode, 2763 int nr_exclusive, void *key) 2764{ 2765 unsigned long flags; 2766 int wake_flags = WF_SYNC; 2767 2768 if (unlikely(!q)) 2769 return; 2770 2771 if (unlikely(nr_exclusive != 1)) 2772 wake_flags = 0; 2773 2774 spin_lock_irqsave(&q->lock, flags); 2775 __wake_up_common(q, mode, nr_exclusive, wake_flags, key); 2776 spin_unlock_irqrestore(&q->lock, flags); 2777} 2778EXPORT_SYMBOL_GPL(__wake_up_sync_key); 2779 2780/* 2781 * __wake_up_sync - see __wake_up_sync_key() 2782 */ 2783void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive) 2784{ 2785 __wake_up_sync_key(q, mode, nr_exclusive, NULL); 2786} 2787EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */ 2788 2789/** 2790 * complete: - signals a single thread waiting on this completion 2791 * @x: holds the state of this particular completion 2792 * 2793 * This will wake up a single thread waiting on this completion. Threads will be 2794 * awakened in the same order in which they were queued. 2795 * 2796 * See also complete_all(), wait_for_completion() and related routines. 2797 * 2798 * It may be assumed that this function implies a write memory barrier before 2799 * changing the task state if and only if any tasks are woken up. 2800 */ 2801void complete(struct completion *x) 2802{ 2803 unsigned long flags; 2804 2805 spin_lock_irqsave(&x->wait.lock, flags); 2806 x->done++; 2807 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL); 2808 spin_unlock_irqrestore(&x->wait.lock, flags); 2809} 2810EXPORT_SYMBOL(complete); 2811 2812/** 2813 * complete_all: - signals all threads waiting on this completion 2814 * @x: holds the state of this particular completion 2815 * 2816 * This will wake up all threads waiting on this particular completion event. 2817 * 2818 * It may be assumed that this function implies a write memory barrier before 2819 * changing the task state if and only if any tasks are woken up. 2820 */ 2821void complete_all(struct completion *x) 2822{ 2823 unsigned long flags; 2824 2825 spin_lock_irqsave(&x->wait.lock, flags); 2826 x->done += UINT_MAX/2; 2827 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL); 2828 spin_unlock_irqrestore(&x->wait.lock, flags); 2829} 2830EXPORT_SYMBOL(complete_all); 2831 2832static inline long __sched 2833do_wait_for_common(struct completion *x, 2834 long (*action)(long), long timeout, int state) 2835{ 2836 if (!x->done) { 2837 DECLARE_WAITQUEUE(wait, current); 2838 2839 __add_wait_queue_tail_exclusive(&x->wait, &wait); 2840 do { 2841 if (signal_pending_state(state, current)) { 2842 timeout = -ERESTARTSYS; 2843 break; 2844 } 2845 __set_current_state(state); 2846 spin_unlock_irq(&x->wait.lock); 2847 timeout = action(timeout); 2848 spin_lock_irq(&x->wait.lock); 2849 } while (!x->done && timeout); 2850 __remove_wait_queue(&x->wait, &wait); 2851 if (!x->done) 2852 return timeout; 2853 } 2854 x->done--; 2855 return timeout ?: 1; 2856} 2857 2858static inline long __sched 2859__wait_for_common(struct completion *x, 2860 long (*action)(long), long timeout, int state) 2861{ 2862 might_sleep(); 2863 2864 spin_lock_irq(&x->wait.lock); 2865 timeout = do_wait_for_common(x, action, timeout, state); 2866 spin_unlock_irq(&x->wait.lock); 2867 return timeout; 2868} 2869 2870static long __sched 2871wait_for_common(struct completion *x, long timeout, int state) 2872{ 2873 return __wait_for_common(x, schedule_timeout, timeout, state); 2874} 2875 2876static long __sched 2877wait_for_common_io(struct completion *x, long timeout, int state) 2878{ 2879 return __wait_for_common(x, io_schedule_timeout, timeout, state); 2880} 2881 2882/** 2883 * wait_for_completion: - waits for completion of a task 2884 * @x: holds the state of this particular completion 2885 * 2886 * This waits to be signaled for completion of a specific task. It is NOT 2887 * interruptible and there is no timeout. 2888 * 2889 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout 2890 * and interrupt capability. Also see complete(). 2891 */ 2892void __sched wait_for_completion(struct completion *x) 2893{ 2894 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE); 2895} 2896EXPORT_SYMBOL(wait_for_completion); 2897 2898/** 2899 * wait_for_completion_timeout: - waits for completion of a task (w/timeout) 2900 * @x: holds the state of this particular completion 2901 * @timeout: timeout value in jiffies 2902 * 2903 * This waits for either a completion of a specific task to be signaled or for a 2904 * specified timeout to expire. The timeout is in jiffies. It is not 2905 * interruptible. 2906 * 2907 * Return: 0 if timed out, and positive (at least 1, or number of jiffies left 2908 * till timeout) if completed. 2909 */ 2910unsigned long __sched 2911wait_for_completion_timeout(struct completion *x, unsigned long timeout) 2912{ 2913 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE); 2914} 2915EXPORT_SYMBOL(wait_for_completion_timeout); 2916 2917/** 2918 * wait_for_completion_io: - waits for completion of a task 2919 * @x: holds the state of this particular completion 2920 * 2921 * This waits to be signaled for completion of a specific task. It is NOT 2922 * interruptible and there is no timeout. The caller is accounted as waiting 2923 * for IO. 2924 */ 2925void __sched wait_for_completion_io(struct completion *x) 2926{ 2927 wait_for_common_io(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE); 2928} 2929EXPORT_SYMBOL(wait_for_completion_io); 2930 2931/** 2932 * wait_for_completion_io_timeout: - waits for completion of a task (w/timeout) 2933 * @x: holds the state of this particular completion 2934 * @timeout: timeout value in jiffies 2935 * 2936 * This waits for either a completion of a specific task to be signaled or for a 2937 * specified timeout to expire. The timeout is in jiffies. It is not 2938 * interruptible. The caller is accounted as waiting for IO. 2939 * 2940 * Return: 0 if timed out, and positive (at least 1, or number of jiffies left 2941 * till timeout) if completed. 2942 */ 2943unsigned long __sched 2944wait_for_completion_io_timeout(struct completion *x, unsigned long timeout) 2945{ 2946 return wait_for_common_io(x, timeout, TASK_UNINTERRUPTIBLE); 2947} 2948EXPORT_SYMBOL(wait_for_completion_io_timeout); 2949 2950/** 2951 * wait_for_completion_interruptible: - waits for completion of a task (w/intr) 2952 * @x: holds the state of this particular completion 2953 * 2954 * This waits for completion of a specific task to be signaled. It is 2955 * interruptible. 2956 * 2957 * Return: -ERESTARTSYS if interrupted, 0 if completed. 2958 */ 2959int __sched wait_for_completion_interruptible(struct completion *x) 2960{ 2961 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE); 2962 if (t == -ERESTARTSYS) 2963 return t; 2964 return 0; 2965} 2966EXPORT_SYMBOL(wait_for_completion_interruptible); 2967 2968/** 2969 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr)) 2970 * @x: holds the state of this particular completion 2971 * @timeout: timeout value in jiffies 2972 * 2973 * This waits for either a completion of a specific task to be signaled or for a 2974 * specified timeout to expire. It is interruptible. The timeout is in jiffies. 2975 * 2976 * Return: -ERESTARTSYS if interrupted, 0 if timed out, positive (at least 1, 2977 * or number of jiffies left till timeout) if completed. 2978 */ 2979long __sched 2980wait_for_completion_interruptible_timeout(struct completion *x, 2981 unsigned long timeout) 2982{ 2983 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE); 2984} 2985EXPORT_SYMBOL(wait_for_completion_interruptible_timeout); 2986 2987/** 2988 * wait_for_completion_killable: - waits for completion of a task (killable) 2989 * @x: holds the state of this particular completion 2990 * 2991 * This waits to be signaled for completion of a specific task. It can be 2992 * interrupted by a kill signal. 2993 * 2994 * Return: -ERESTARTSYS if interrupted, 0 if completed. 2995 */ 2996int __sched wait_for_completion_killable(struct completion *x) 2997{ 2998 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE); 2999 if (t == -ERESTARTSYS) 3000 return t; 3001 return 0; 3002} 3003EXPORT_SYMBOL(wait_for_completion_killable); 3004 3005/** 3006 * wait_for_completion_killable_timeout: - waits for completion of a task (w/(to,killable)) 3007 * @x: holds the state of this particular completion 3008 * @timeout: timeout value in jiffies 3009 * 3010 * This waits for either a completion of a specific task to be 3011 * signaled or for a specified timeout to expire. It can be 3012 * interrupted by a kill signal. The timeout is in jiffies. 3013 * 3014 * Return: -ERESTARTSYS if interrupted, 0 if timed out, positive (at least 1, 3015 * or number of jiffies left till timeout) if completed. 3016 */ 3017long __sched 3018wait_for_completion_killable_timeout(struct completion *x, 3019 unsigned long timeout) 3020{ 3021 return wait_for_common(x, timeout, TASK_KILLABLE); 3022} 3023EXPORT_SYMBOL(wait_for_completion_killable_timeout); 3024 3025/** 3026 * try_wait_for_completion - try to decrement a completion without blocking 3027 * @x: completion structure 3028 * 3029 * Return: 0 if a decrement cannot be done without blocking 3030 * 1 if a decrement succeeded. 3031 * 3032 * If a completion is being used as a counting completion, 3033 * attempt to decrement the counter without blocking. This 3034 * enables us to avoid waiting if the resource the completion 3035 * is protecting is not available. 3036 */ 3037bool try_wait_for_completion(struct completion *x) 3038{ 3039 unsigned long flags; 3040 int ret = 1; 3041 3042 spin_lock_irqsave(&x->wait.lock, flags); 3043 if (!x->done) 3044 ret = 0; 3045 else 3046 x->done--; 3047 spin_unlock_irqrestore(&x->wait.lock, flags); 3048 return ret; 3049} 3050EXPORT_SYMBOL(try_wait_for_completion); 3051 3052/** 3053 * completion_done - Test to see if a completion has any waiters 3054 * @x: completion structure 3055 * 3056 * Return: 0 if there are waiters (wait_for_completion() in progress) 3057 * 1 if there are no waiters. 3058 * 3059 */ 3060bool completion_done(struct completion *x) 3061{ 3062 unsigned long flags; 3063 int ret = 1; 3064 3065 spin_lock_irqsave(&x->wait.lock, flags); 3066 if (!x->done) 3067 ret = 0; 3068 spin_unlock_irqrestore(&x->wait.lock, flags); 3069 return ret; 3070} 3071EXPORT_SYMBOL(completion_done); 3072 3073static long __sched 3074sleep_on_common(wait_queue_head_t *q, int state, long timeout) 3075{ 3076 unsigned long flags; 3077 wait_queue_t wait; 3078 3079 init_waitqueue_entry(&wait, current); 3080 3081 __set_current_state(state); 3082 3083 spin_lock_irqsave(&q->lock, flags); 3084 __add_wait_queue(q, &wait); 3085 spin_unlock(&q->lock); 3086 timeout = schedule_timeout(timeout); 3087 spin_lock_irq(&q->lock); 3088 __remove_wait_queue(q, &wait); 3089 spin_unlock_irqrestore(&q->lock, flags); 3090 3091 return timeout; 3092} 3093 3094void __sched interruptible_sleep_on(wait_queue_head_t *q) 3095{ 3096 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT); 3097} 3098EXPORT_SYMBOL(interruptible_sleep_on); 3099 3100long __sched 3101interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout) 3102{ 3103 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout); 3104} 3105EXPORT_SYMBOL(interruptible_sleep_on_timeout); 3106 3107void __sched sleep_on(wait_queue_head_t *q) 3108{ 3109 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT); 3110} 3111EXPORT_SYMBOL(sleep_on); 3112 3113long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout) 3114{ 3115 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout); 3116} 3117EXPORT_SYMBOL(sleep_on_timeout); 3118 3119#ifdef CONFIG_RT_MUTEXES 3120 3121/* 3122 * rt_mutex_setprio - set the current priority of a task 3123 * @p: task 3124 * @prio: prio value (kernel-internal form) 3125 * 3126 * This function changes the 'effective' priority of a task. It does 3127 * not touch ->normal_prio like __setscheduler(). 3128 * 3129 * Used by the rt_mutex code to implement priority inheritance logic. 3130 */ 3131void rt_mutex_setprio(struct task_struct *p, int prio) 3132{ 3133 int oldprio, on_rq, running; 3134 struct rq *rq; 3135 const struct sched_class *prev_class; 3136 3137 BUG_ON(prio < 0 || prio > MAX_PRIO); 3138 3139 rq = __task_rq_lock(p); 3140 3141 /* 3142 * Idle task boosting is a nono in general. There is one 3143 * exception, when PREEMPT_RT and NOHZ is active: 3144 * 3145 * The idle task calls get_next_timer_interrupt() and holds 3146 * the timer wheel base->lock on the CPU and another CPU wants 3147 * to access the timer (probably to cancel it). We can safely 3148 * ignore the boosting request, as the idle CPU runs this code 3149 * with interrupts disabled and will complete the lock 3150 * protected section without being interrupted. So there is no 3151 * real need to boost. 3152 */ 3153 if (unlikely(p == rq->idle)) { 3154 WARN_ON(p != rq->curr); 3155 WARN_ON(p->pi_blocked_on); 3156 goto out_unlock; 3157 } 3158 3159 trace_sched_pi_setprio(p, prio); 3160 oldprio = p->prio; 3161 prev_class = p->sched_class; 3162 on_rq = p->on_rq; 3163 running = task_current(rq, p); 3164 if (on_rq) 3165 dequeue_task(rq, p, 0); 3166 if (running) 3167 p->sched_class->put_prev_task(rq, p); 3168 3169 if (rt_prio(prio)) 3170 p->sched_class = &rt_sched_class; 3171 else 3172 p->sched_class = &fair_sched_class; 3173 3174 p->prio = prio; 3175 3176 if (running) 3177 p->sched_class->set_curr_task(rq); 3178 if (on_rq) 3179 enqueue_task(rq, p, oldprio < prio ? ENQUEUE_HEAD : 0); 3180 3181 check_class_changed(rq, p, prev_class, oldprio); 3182out_unlock: 3183 __task_rq_unlock(rq); 3184} 3185#endif 3186void set_user_nice(struct task_struct *p, long nice) 3187{ 3188 int old_prio, delta, on_rq; 3189 unsigned long flags; 3190 struct rq *rq; 3191 3192 if (TASK_NICE(p) == nice || nice < -20 || nice > 19) 3193 return; 3194 /* 3195 * We have to be careful, if called from sys_setpriority(), 3196 * the task might be in the middle of scheduling on another CPU. 3197 */ 3198 rq = task_rq_lock(p, &flags); 3199 /* 3200 * The RT priorities are set via sched_setscheduler(), but we still 3201 * allow the 'normal' nice value to be set - but as expected 3202 * it wont have any effect on scheduling until the task is 3203 * SCHED_FIFO/SCHED_RR: 3204 */ 3205 if (task_has_rt_policy(p)) { 3206 p->static_prio = NICE_TO_PRIO(nice); 3207 goto out_unlock; 3208 } 3209 on_rq = p->on_rq; 3210 if (on_rq) 3211 dequeue_task(rq, p, 0); 3212 3213 p->static_prio = NICE_TO_PRIO(nice); 3214 set_load_weight(p); 3215 old_prio = p->prio; 3216 p->prio = effective_prio(p); 3217 delta = p->prio - old_prio; 3218 3219 if (on_rq) { 3220 enqueue_task(rq, p, 0); 3221 /* 3222 * If the task increased its priority or is running and 3223 * lowered its priority, then reschedule its CPU: 3224 */ 3225 if (delta < 0 || (delta > 0 && task_running(rq, p))) 3226 resched_task(rq->curr); 3227 } 3228out_unlock: 3229 task_rq_unlock(rq, p, &flags); 3230} 3231EXPORT_SYMBOL(set_user_nice); 3232 3233/* 3234 * can_nice - check if a task can reduce its nice value 3235 * @p: task 3236 * @nice: nice value 3237 */ 3238int can_nice(const struct task_struct *p, const int nice) 3239{ 3240 /* convert nice value [19,-20] to rlimit style value [1,40] */ 3241 int nice_rlim = 20 - nice; 3242 3243 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) || 3244 capable(CAP_SYS_NICE)); 3245} 3246 3247#ifdef __ARCH_WANT_SYS_NICE 3248 3249/* 3250 * sys_nice - change the priority of the current process. 3251 * @increment: priority increment 3252 * 3253 * sys_setpriority is a more generic, but much slower function that 3254 * does similar things. 3255 */ 3256SYSCALL_DEFINE1(nice, int, increment) 3257{ 3258 long nice, retval; 3259 3260 /* 3261 * Setpriority might change our priority at the same moment. 3262 * We don't have to worry. Conceptually one call occurs first 3263 * and we have a single winner. 3264 */ 3265 if (increment < -40) 3266 increment = -40; 3267 if (increment > 40) 3268 increment = 40; 3269 3270 nice = TASK_NICE(current) + increment; 3271 if (nice < -20) 3272 nice = -20; 3273 if (nice > 19) 3274 nice = 19; 3275 3276 if (increment < 0 && !can_nice(current, nice)) 3277 return -EPERM; 3278 3279 retval = security_task_setnice(current, nice); 3280 if (retval) 3281 return retval; 3282 3283 set_user_nice(current, nice); 3284 return 0; 3285} 3286 3287#endif 3288 3289/** 3290 * task_prio - return the priority value of a given task. 3291 * @p: the task in question. 3292 * 3293 * Return: The priority value as seen by users in /proc. 3294 * RT tasks are offset by -200. Normal tasks are centered 3295 * around 0, value goes from -16 to +15. 3296 */ 3297int task_prio(const struct task_struct *p) 3298{ 3299 return p->prio - MAX_RT_PRIO; 3300} 3301 3302/** 3303 * task_nice - return the nice value of a given task. 3304 * @p: the task in question. 3305 * 3306 * Return: The nice value [ -20 ... 0 ... 19 ]. 3307 */ 3308int task_nice(const struct task_struct *p) 3309{ 3310 return TASK_NICE(p); 3311} 3312EXPORT_SYMBOL(task_nice); 3313 3314/** 3315 * idle_cpu - is a given cpu idle currently? 3316 * @cpu: the processor in question. 3317 * 3318 * Return: 1 if the CPU is currently idle. 0 otherwise. 3319 */ 3320int idle_cpu(int cpu) 3321{ 3322 struct rq *rq = cpu_rq(cpu); 3323 3324 if (rq->curr != rq->idle) 3325 return 0; 3326 3327 if (rq->nr_running) 3328 return 0; 3329 3330#ifdef CONFIG_SMP 3331 if (!llist_empty(&rq->wake_list)) 3332 return 0; 3333#endif 3334 3335 return 1; 3336} 3337 3338/** 3339 * idle_task - return the idle task for a given cpu. 3340 * @cpu: the processor in question. 3341 * 3342 * Return: The idle task for the cpu @cpu. 3343 */ 3344struct task_struct *idle_task(int cpu) 3345{ 3346 return cpu_rq(cpu)->idle; 3347} 3348 3349/** 3350 * find_process_by_pid - find a process with a matching PID value. 3351 * @pid: the pid in question. 3352 * 3353 * The task of @pid, if found. %NULL otherwise. 3354 */ 3355static struct task_struct *find_process_by_pid(pid_t pid) 3356{ 3357 return pid ? find_task_by_vpid(pid) : current; 3358} 3359 3360/* Actually do priority change: must hold rq lock. */ 3361static void 3362__setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio) 3363{ 3364 p->policy = policy; 3365 p->rt_priority = prio; 3366 p->normal_prio = normal_prio(p); 3367 /* we are holding p->pi_lock already */ 3368 p->prio = rt_mutex_getprio(p); 3369 if (rt_prio(p->prio)) 3370 p->sched_class = &rt_sched_class; 3371 else 3372 p->sched_class = &fair_sched_class; 3373 set_load_weight(p); 3374} 3375 3376/* 3377 * check the target process has a UID that matches the current process's 3378 */ 3379static bool check_same_owner(struct task_struct *p) 3380{ 3381 const struct cred *cred = current_cred(), *pcred; 3382 bool match; 3383 3384 rcu_read_lock(); 3385 pcred = __task_cred(p); 3386 match = (uid_eq(cred->euid, pcred->euid) || 3387 uid_eq(cred->euid, pcred->uid)); 3388 rcu_read_unlock(); 3389 return match; 3390} 3391 3392static int __sched_setscheduler(struct task_struct *p, int policy, 3393 const struct sched_param *param, bool user) 3394{ 3395 int retval, oldprio, oldpolicy = -1, on_rq, running; 3396 unsigned long flags; 3397 const struct sched_class *prev_class; 3398 struct rq *rq; 3399 int reset_on_fork; 3400 3401 /* may grab non-irq protected spin_locks */ 3402 BUG_ON(in_interrupt()); 3403recheck: 3404 /* double check policy once rq lock held */ 3405 if (policy < 0) { 3406 reset_on_fork = p->sched_reset_on_fork; 3407 policy = oldpolicy = p->policy; 3408 } else { 3409 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK); 3410 policy &= ~SCHED_RESET_ON_FORK; 3411 3412 if (policy != SCHED_FIFO && policy != SCHED_RR && 3413 policy != SCHED_NORMAL && policy != SCHED_BATCH && 3414 policy != SCHED_IDLE) 3415 return -EINVAL; 3416 } 3417 3418 /* 3419 * Valid priorities for SCHED_FIFO and SCHED_RR are 3420 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL, 3421 * SCHED_BATCH and SCHED_IDLE is 0. 3422 */ 3423 if (param->sched_priority < 0 || 3424 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) || 3425 (!p->mm && param->sched_priority > MAX_RT_PRIO-1)) 3426 return -EINVAL; 3427 if (rt_policy(policy) != (param->sched_priority != 0)) 3428 return -EINVAL; 3429 3430 /* 3431 * Allow unprivileged RT tasks to decrease priority: 3432 */ 3433 if (user && !capable(CAP_SYS_NICE)) { 3434 if (rt_policy(policy)) { 3435 unsigned long rlim_rtprio = 3436 task_rlimit(p, RLIMIT_RTPRIO); 3437 3438 /* can't set/change the rt policy */ 3439 if (policy != p->policy && !rlim_rtprio) 3440 return -EPERM; 3441 3442 /* can't increase priority */ 3443 if (param->sched_priority > p->rt_priority && 3444 param->sched_priority > rlim_rtprio) 3445 return -EPERM; 3446 } 3447 3448 /* 3449 * Treat SCHED_IDLE as nice 20. Only allow a switch to 3450 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it. 3451 */ 3452 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) { 3453 if (!can_nice(p, TASK_NICE(p))) 3454 return -EPERM; 3455 } 3456 3457 /* can't change other user's priorities */ 3458 if (!check_same_owner(p)) 3459 return -EPERM; 3460 3461 /* Normal users shall not reset the sched_reset_on_fork flag */ 3462 if (p->sched_reset_on_fork && !reset_on_fork) 3463 return -EPERM; 3464 } 3465 3466 if (user) { 3467 retval = security_task_setscheduler(p); 3468 if (retval) 3469 return retval; 3470 } 3471 3472 /* 3473 * make sure no PI-waiters arrive (or leave) while we are 3474 * changing the priority of the task: 3475 * 3476 * To be able to change p->policy safely, the appropriate 3477 * runqueue lock must be held. 3478 */ 3479 rq = task_rq_lock(p, &flags); 3480 3481 /* 3482 * Changing the policy of the stop threads its a very bad idea 3483 */ 3484 if (p == rq->stop) { 3485 task_rq_unlock(rq, p, &flags); 3486 return -EINVAL; 3487 } 3488 3489 /* 3490 * If not changing anything there's no need to proceed further: 3491 */ 3492 if (unlikely(policy == p->policy && (!rt_policy(policy) || 3493 param->sched_priority == p->rt_priority))) { 3494 task_rq_unlock(rq, p, &flags); 3495 return 0; 3496 } 3497 3498#ifdef CONFIG_RT_GROUP_SCHED 3499 if (user) { 3500 /* 3501 * Do not allow realtime tasks into groups that have no runtime 3502 * assigned. 3503 */ 3504 if (rt_bandwidth_enabled() && rt_policy(policy) && 3505 task_group(p)->rt_bandwidth.rt_runtime == 0 && 3506 !task_group_is_autogroup(task_group(p))) { 3507 task_rq_unlock(rq, p, &flags); 3508 return -EPERM; 3509 } 3510 } 3511#endif 3512 3513 /* recheck policy now with rq lock held */ 3514 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) { 3515 policy = oldpolicy = -1; 3516 task_rq_unlock(rq, p, &flags); 3517 goto recheck; 3518 } 3519 on_rq = p->on_rq; 3520 running = task_current(rq, p); 3521 if (on_rq) 3522 dequeue_task(rq, p, 0); 3523 if (running) 3524 p->sched_class->put_prev_task(rq, p); 3525 3526 p->sched_reset_on_fork = reset_on_fork; 3527 3528 oldprio = p->prio; 3529 prev_class = p->sched_class; 3530 __setscheduler(rq, p, policy, param->sched_priority); 3531 3532 if (running) 3533 p->sched_class->set_curr_task(rq); 3534 if (on_rq) 3535 enqueue_task(rq, p, 0); 3536 3537 check_class_changed(rq, p, prev_class, oldprio); 3538 task_rq_unlock(rq, p, &flags); 3539 3540 rt_mutex_adjust_pi(p); 3541 3542 return 0; 3543} 3544 3545/** 3546 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread. 3547 * @p: the task in question. 3548 * @policy: new policy. 3549 * @param: structure containing the new RT priority. 3550 * 3551 * Return: 0 on success. An error code otherwise. 3552 * 3553 * NOTE that the task may be already dead. 3554 */ 3555int sched_setscheduler(struct task_struct *p, int policy, 3556 const struct sched_param *param) 3557{ 3558 return __sched_setscheduler(p, policy, param, true); 3559} 3560EXPORT_SYMBOL_GPL(sched_setscheduler); 3561 3562/** 3563 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace. 3564 * @p: the task in question. 3565 * @policy: new policy. 3566 * @param: structure containing the new RT priority. 3567 * 3568 * Just like sched_setscheduler, only don't bother checking if the 3569 * current context has permission. For example, this is needed in 3570 * stop_machine(): we create temporary high priority worker threads, 3571 * but our caller might not have that capability. 3572 * 3573 * Return: 0 on success. An error code otherwise. 3574 */ 3575int sched_setscheduler_nocheck(struct task_struct *p, int policy, 3576 const struct sched_param *param) 3577{ 3578 return __sched_setscheduler(p, policy, param, false); 3579} 3580 3581static int 3582do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param) 3583{ 3584 struct sched_param lparam; 3585 struct task_struct *p; 3586 int retval; 3587 3588 if (!param || pid < 0) 3589 return -EINVAL; 3590 if (copy_from_user(&lparam, param, sizeof(struct sched_param))) 3591 return -EFAULT; 3592 3593 rcu_read_lock(); 3594 retval = -ESRCH; 3595 p = find_process_by_pid(pid); 3596 if (p != NULL) 3597 retval = sched_setscheduler(p, policy, &lparam); 3598 rcu_read_unlock(); 3599 3600 return retval; 3601} 3602 3603/** 3604 * sys_sched_setscheduler - set/change the scheduler policy and RT priority 3605 * @pid: the pid in question. 3606 * @policy: new policy. 3607 * @param: structure containing the new RT priority. 3608 * 3609 * Return: 0 on success. An error code otherwise. 3610 */ 3611SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy, 3612 struct sched_param __user *, param) 3613{ 3614 /* negative values for policy are not valid */ 3615 if (policy < 0) 3616 return -EINVAL; 3617 3618 return do_sched_setscheduler(pid, policy, param); 3619} 3620 3621/** 3622 * sys_sched_setparam - set/change the RT priority of a thread 3623 * @pid: the pid in question. 3624 * @param: structure containing the new RT priority. 3625 * 3626 * Return: 0 on success. An error code otherwise. 3627 */ 3628SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param) 3629{ 3630 return do_sched_setscheduler(pid, -1, param); 3631} 3632 3633/** 3634 * sys_sched_getscheduler - get the policy (scheduling class) of a thread 3635 * @pid: the pid in question. 3636 * 3637 * Return: On success, the policy of the thread. Otherwise, a negative error 3638 * code. 3639 */ 3640SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid) 3641{ 3642 struct task_struct *p; 3643 int retval; 3644 3645 if (pid < 0) 3646 return -EINVAL; 3647 3648 retval = -ESRCH; 3649 rcu_read_lock(); 3650 p = find_process_by_pid(pid); 3651 if (p) { 3652 retval = security_task_getscheduler(p); 3653 if (!retval) 3654 retval = p->policy 3655 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0); 3656 } 3657 rcu_read_unlock(); 3658 return retval; 3659} 3660 3661/** 3662 * sys_sched_getparam - get the RT priority of a thread 3663 * @pid: the pid in question. 3664 * @param: structure containing the RT priority. 3665 * 3666 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error 3667 * code. 3668 */ 3669SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param) 3670{ 3671 struct sched_param lp; 3672 struct task_struct *p; 3673 int retval; 3674 3675 if (!param || pid < 0) 3676 return -EINVAL; 3677 3678 rcu_read_lock(); 3679 p = find_process_by_pid(pid); 3680 retval = -ESRCH; 3681 if (!p) 3682 goto out_unlock; 3683 3684 retval = security_task_getscheduler(p); 3685 if (retval) 3686 goto out_unlock; 3687 3688 lp.sched_priority = p->rt_priority; 3689 rcu_read_unlock(); 3690 3691 /* 3692 * This one might sleep, we cannot do it with a spinlock held ... 3693 */ 3694 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0; 3695 3696 return retval; 3697 3698out_unlock: 3699 rcu_read_unlock(); 3700 return retval; 3701} 3702 3703long sched_setaffinity(pid_t pid, const struct cpumask *in_mask) 3704{ 3705 cpumask_var_t cpus_allowed, new_mask; 3706 struct task_struct *p; 3707 int retval; 3708 3709 get_online_cpus(); 3710 rcu_read_lock(); 3711 3712 p = find_process_by_pid(pid); 3713 if (!p) { 3714 rcu_read_unlock(); 3715 put_online_cpus(); 3716 return -ESRCH; 3717 } 3718 3719 /* Prevent p going away */ 3720 get_task_struct(p); 3721 rcu_read_unlock(); 3722 3723 if (p->flags & PF_NO_SETAFFINITY) { 3724 retval = -EINVAL; 3725 goto out_put_task; 3726 } 3727 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) { 3728 retval = -ENOMEM; 3729 goto out_put_task; 3730 } 3731 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) { 3732 retval = -ENOMEM; 3733 goto out_free_cpus_allowed; 3734 } 3735 retval = -EPERM; 3736 if (!check_same_owner(p)) { 3737 rcu_read_lock(); 3738 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) { 3739 rcu_read_unlock(); 3740 goto out_unlock; 3741 } 3742 rcu_read_unlock(); 3743 } 3744 3745 retval = security_task_setscheduler(p); 3746 if (retval) 3747 goto out_unlock; 3748 3749 cpuset_cpus_allowed(p, cpus_allowed); 3750 cpumask_and(new_mask, in_mask, cpus_allowed); 3751again: 3752 retval = set_cpus_allowed_ptr(p, new_mask); 3753 3754 if (!retval) { 3755 cpuset_cpus_allowed(p, cpus_allowed); 3756 if (!cpumask_subset(new_mask, cpus_allowed)) { 3757 /* 3758 * We must have raced with a concurrent cpuset 3759 * update. Just reset the cpus_allowed to the 3760 * cpuset's cpus_allowed 3761 */ 3762 cpumask_copy(new_mask, cpus_allowed); 3763 goto again; 3764 } 3765 } 3766out_unlock: 3767 free_cpumask_var(new_mask); 3768out_free_cpus_allowed: 3769 free_cpumask_var(cpus_allowed); 3770out_put_task: 3771 put_task_struct(p); 3772 put_online_cpus(); 3773 return retval; 3774} 3775 3776static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len, 3777 struct cpumask *new_mask) 3778{ 3779 if (len < cpumask_size()) 3780 cpumask_clear(new_mask); 3781 else if (len > cpumask_size()) 3782 len = cpumask_size(); 3783 3784 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0; 3785} 3786 3787/** 3788 * sys_sched_setaffinity - set the cpu affinity of a process 3789 * @pid: pid of the process 3790 * @len: length in bytes of the bitmask pointed to by user_mask_ptr 3791 * @user_mask_ptr: user-space pointer to the new cpu mask 3792 * 3793 * Return: 0 on success. An error code otherwise. 3794 */ 3795SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len, 3796 unsigned long __user *, user_mask_ptr) 3797{ 3798 cpumask_var_t new_mask; 3799 int retval; 3800 3801 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) 3802 return -ENOMEM; 3803 3804 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask); 3805 if (retval == 0) 3806 retval = sched_setaffinity(pid, new_mask); 3807 free_cpumask_var(new_mask); 3808 return retval; 3809} 3810 3811long sched_getaffinity(pid_t pid, struct cpumask *mask) 3812{ 3813 struct task_struct *p; 3814 unsigned long flags; 3815 int retval; 3816 3817 get_online_cpus(); 3818 rcu_read_lock(); 3819 3820 retval = -ESRCH; 3821 p = find_process_by_pid(pid); 3822 if (!p) 3823 goto out_unlock; 3824 3825 retval = security_task_getscheduler(p); 3826 if (retval) 3827 goto out_unlock; 3828 3829 raw_spin_lock_irqsave(&p->pi_lock, flags); 3830 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask); 3831 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 3832 3833out_unlock: 3834 rcu_read_unlock(); 3835 put_online_cpus(); 3836 3837 return retval; 3838} 3839 3840/** 3841 * sys_sched_getaffinity - get the cpu affinity of a process 3842 * @pid: pid of the process 3843 * @len: length in bytes of the bitmask pointed to by user_mask_ptr 3844 * @user_mask_ptr: user-space pointer to hold the current cpu mask 3845 * 3846 * Return: 0 on success. An error code otherwise. 3847 */ 3848SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len, 3849 unsigned long __user *, user_mask_ptr) 3850{ 3851 int ret; 3852 cpumask_var_t mask; 3853 3854 if ((len * BITS_PER_BYTE) < nr_cpu_ids) 3855 return -EINVAL; 3856 if (len & (sizeof(unsigned long)-1)) 3857 return -EINVAL; 3858 3859 if (!alloc_cpumask_var(&mask, GFP_KERNEL)) 3860 return -ENOMEM; 3861 3862 ret = sched_getaffinity(pid, mask); 3863 if (ret == 0) { 3864 size_t retlen = min_t(size_t, len, cpumask_size()); 3865 3866 if (copy_to_user(user_mask_ptr, mask, retlen)) 3867 ret = -EFAULT; 3868 else 3869 ret = retlen; 3870 } 3871 free_cpumask_var(mask); 3872 3873 return ret; 3874} 3875 3876/** 3877 * sys_sched_yield - yield the current processor to other threads. 3878 * 3879 * This function yields the current CPU to other tasks. If there are no 3880 * other threads running on this CPU then this function will return. 3881 * 3882 * Return: 0. 3883 */ 3884SYSCALL_DEFINE0(sched_yield) 3885{ 3886 struct rq *rq = this_rq_lock(); 3887 3888 schedstat_inc(rq, yld_count); 3889 current->sched_class->yield_task(rq); 3890 3891 /* 3892 * Since we are going to call schedule() anyway, there's 3893 * no need to preempt or enable interrupts: 3894 */ 3895 __release(rq->lock); 3896 spin_release(&rq->lock.dep_map, 1, _THIS_IP_); 3897 do_raw_spin_unlock(&rq->lock); 3898 sched_preempt_enable_no_resched(); 3899 3900 schedule(); 3901 3902 return 0; 3903} 3904 3905static void __cond_resched(void) 3906{ 3907 __preempt_count_add(PREEMPT_ACTIVE); 3908 __schedule(); 3909 __preempt_count_sub(PREEMPT_ACTIVE); 3910} 3911 3912int __sched _cond_resched(void) 3913{ 3914 if (should_resched()) { 3915 __cond_resched(); 3916 return 1; 3917 } 3918 return 0; 3919} 3920EXPORT_SYMBOL(_cond_resched); 3921 3922/* 3923 * __cond_resched_lock() - if a reschedule is pending, drop the given lock, 3924 * call schedule, and on return reacquire the lock. 3925 * 3926 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level 3927 * operations here to prevent schedule() from being called twice (once via 3928 * spin_unlock(), once by hand). 3929 */ 3930int __cond_resched_lock(spinlock_t *lock) 3931{ 3932 int resched = should_resched(); 3933 int ret = 0; 3934 3935 lockdep_assert_held(lock); 3936 3937 if (spin_needbreak(lock) || resched) { 3938 spin_unlock(lock); 3939 if (resched) 3940 __cond_resched(); 3941 else 3942 cpu_relax(); 3943 ret = 1; 3944 spin_lock(lock); 3945 } 3946 return ret; 3947} 3948EXPORT_SYMBOL(__cond_resched_lock); 3949 3950int __sched __cond_resched_softirq(void) 3951{ 3952 BUG_ON(!in_softirq()); 3953 3954 if (should_resched()) { 3955 local_bh_enable(); 3956 __cond_resched(); 3957 local_bh_disable(); 3958 return 1; 3959 } 3960 return 0; 3961} 3962EXPORT_SYMBOL(__cond_resched_softirq); 3963 3964/** 3965 * yield - yield the current processor to other threads. 3966 * 3967 * Do not ever use this function, there's a 99% chance you're doing it wrong. 3968 * 3969 * The scheduler is at all times free to pick the calling task as the most 3970 * eligible task to run, if removing the yield() call from your code breaks 3971 * it, its already broken. 3972 * 3973 * Typical broken usage is: 3974 * 3975 * while (!event) 3976 * yield(); 3977 * 3978 * where one assumes that yield() will let 'the other' process run that will 3979 * make event true. If the current task is a SCHED_FIFO task that will never 3980 * happen. Never use yield() as a progress guarantee!! 3981 * 3982 * If you want to use yield() to wait for something, use wait_event(). 3983 * If you want to use yield() to be 'nice' for others, use cond_resched(). 3984 * If you still want to use yield(), do not! 3985 */ 3986void __sched yield(void) 3987{ 3988 set_current_state(TASK_RUNNING); 3989 sys_sched_yield(); 3990} 3991EXPORT_SYMBOL(yield); 3992 3993/** 3994 * yield_to - yield the current processor to another thread in 3995 * your thread group, or accelerate that thread toward the 3996 * processor it's on. 3997 * @p: target task 3998 * @preempt: whether task preemption is allowed or not 3999 * 4000 * It's the caller's job to ensure that the target task struct 4001 * can't go away on us before we can do any checks. 4002 * 4003 * Return: 4004 * true (>0) if we indeed boosted the target task. 4005 * false (0) if we failed to boost the target. 4006 * -ESRCH if there's no task to yield to. 4007 */ 4008bool __sched yield_to(struct task_struct *p, bool preempt) 4009{ 4010 struct task_struct *curr = current; 4011 struct rq *rq, *p_rq; 4012 unsigned long flags; 4013 int yielded = 0; 4014 4015 local_irq_save(flags); 4016 rq = this_rq(); 4017 4018again: 4019 p_rq = task_rq(p); 4020 /* 4021 * If we're the only runnable task on the rq and target rq also 4022 * has only one task, there's absolutely no point in yielding. 4023 */ 4024 if (rq->nr_running == 1 && p_rq->nr_running == 1) { 4025 yielded = -ESRCH; 4026 goto out_irq; 4027 } 4028 4029 double_rq_lock(rq, p_rq); 4030 while (task_rq(p) != p_rq) { 4031 double_rq_unlock(rq, p_rq); 4032 goto again; 4033 } 4034 4035 if (!curr->sched_class->yield_to_task) 4036 goto out_unlock; 4037 4038 if (curr->sched_class != p->sched_class) 4039 goto out_unlock; 4040 4041 if (task_running(p_rq, p) || p->state) 4042 goto out_unlock; 4043 4044 yielded = curr->sched_class->yield_to_task(rq, p, preempt); 4045 if (yielded) { 4046 schedstat_inc(rq, yld_count); 4047 /* 4048 * Make p's CPU reschedule; pick_next_entity takes care of 4049 * fairness. 4050 */ 4051 if (preempt && rq != p_rq) 4052 resched_task(p_rq->curr); 4053 } 4054 4055out_unlock: 4056 double_rq_unlock(rq, p_rq); 4057out_irq: 4058 local_irq_restore(flags); 4059 4060 if (yielded > 0) 4061 schedule(); 4062 4063 return yielded; 4064} 4065EXPORT_SYMBOL_GPL(yield_to); 4066 4067/* 4068 * This task is about to go to sleep on IO. Increment rq->nr_iowait so 4069 * that process accounting knows that this is a task in IO wait state. 4070 */ 4071void __sched io_schedule(void) 4072{ 4073 struct rq *rq = raw_rq(); 4074 4075 delayacct_blkio_start(); 4076 atomic_inc(&rq->nr_iowait); 4077 blk_flush_plug(current); 4078 current->in_iowait = 1; 4079 schedule(); 4080 current->in_iowait = 0; 4081 atomic_dec(&rq->nr_iowait); 4082 delayacct_blkio_end(); 4083} 4084EXPORT_SYMBOL(io_schedule); 4085 4086long __sched io_schedule_timeout(long timeout) 4087{ 4088 struct rq *rq = raw_rq(); 4089 long ret; 4090 4091 delayacct_blkio_start(); 4092 atomic_inc(&rq->nr_iowait); 4093 blk_flush_plug(current); 4094 current->in_iowait = 1; 4095 ret = schedule_timeout(timeout); 4096 current->in_iowait = 0; 4097 atomic_dec(&rq->nr_iowait); 4098 delayacct_blkio_end(); 4099 return ret; 4100} 4101 4102/** 4103 * sys_sched_get_priority_max - return maximum RT priority. 4104 * @policy: scheduling class. 4105 * 4106 * Return: On success, this syscall returns the maximum 4107 * rt_priority that can be used by a given scheduling class. 4108 * On failure, a negative error code is returned. 4109 */ 4110SYSCALL_DEFINE1(sched_get_priority_max, int, policy) 4111{ 4112 int ret = -EINVAL; 4113 4114 switch (policy) { 4115 case SCHED_FIFO: 4116 case SCHED_RR: 4117 ret = MAX_USER_RT_PRIO-1; 4118 break; 4119 case SCHED_NORMAL: 4120 case SCHED_BATCH: 4121 case SCHED_IDLE: 4122 ret = 0; 4123 break; 4124 } 4125 return ret; 4126} 4127 4128/** 4129 * sys_sched_get_priority_min - return minimum RT priority. 4130 * @policy: scheduling class. 4131 * 4132 * Return: On success, this syscall returns the minimum 4133 * rt_priority that can be used by a given scheduling class. 4134 * On failure, a negative error code is returned. 4135 */ 4136SYSCALL_DEFINE1(sched_get_priority_min, int, policy) 4137{ 4138 int ret = -EINVAL; 4139 4140 switch (policy) { 4141 case SCHED_FIFO: 4142 case SCHED_RR: 4143 ret = 1; 4144 break; 4145 case SCHED_NORMAL: 4146 case SCHED_BATCH: 4147 case SCHED_IDLE: 4148 ret = 0; 4149 } 4150 return ret; 4151} 4152 4153/** 4154 * sys_sched_rr_get_interval - return the default timeslice of a process. 4155 * @pid: pid of the process. 4156 * @interval: userspace pointer to the timeslice value. 4157 * 4158 * this syscall writes the default timeslice value of a given process 4159 * into the user-space timespec buffer. A value of '0' means infinity. 4160 * 4161 * Return: On success, 0 and the timeslice is in @interval. Otherwise, 4162 * an error code. 4163 */ 4164SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid, 4165 struct timespec __user *, interval) 4166{ 4167 struct task_struct *p; 4168 unsigned int time_slice; 4169 unsigned long flags; 4170 struct rq *rq; 4171 int retval; 4172 struct timespec t; 4173 4174 if (pid < 0) 4175 return -EINVAL; 4176 4177 retval = -ESRCH; 4178 rcu_read_lock(); 4179 p = find_process_by_pid(pid); 4180 if (!p) 4181 goto out_unlock; 4182 4183 retval = security_task_getscheduler(p); 4184 if (retval) 4185 goto out_unlock; 4186 4187 rq = task_rq_lock(p, &flags); 4188 time_slice = p->sched_class->get_rr_interval(rq, p); 4189 task_rq_unlock(rq, p, &flags); 4190 4191 rcu_read_unlock(); 4192 jiffies_to_timespec(time_slice, &t); 4193 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0; 4194 return retval; 4195 4196out_unlock: 4197 rcu_read_unlock(); 4198 return retval; 4199} 4200 4201static const char stat_nam[] = TASK_STATE_TO_CHAR_STR; 4202 4203void sched_show_task(struct task_struct *p) 4204{ 4205 unsigned long free = 0; 4206 int ppid; 4207 unsigned state; 4208 4209 state = p->state ? __ffs(p->state) + 1 : 0; 4210 printk(KERN_INFO "%-15.15s %c", p->comm, 4211 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?'); 4212#if BITS_PER_LONG == 32 4213 if (state == TASK_RUNNING) 4214 printk(KERN_CONT " running "); 4215 else 4216 printk(KERN_CONT " %08lx ", thread_saved_pc(p)); 4217#else 4218 if (state == TASK_RUNNING) 4219 printk(KERN_CONT " running task "); 4220 else 4221 printk(KERN_CONT " %016lx ", thread_saved_pc(p)); 4222#endif 4223#ifdef CONFIG_DEBUG_STACK_USAGE 4224 free = stack_not_used(p); 4225#endif 4226 rcu_read_lock(); 4227 ppid = task_pid_nr(rcu_dereference(p->real_parent)); 4228 rcu_read_unlock(); 4229 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free, 4230 task_pid_nr(p), ppid, 4231 (unsigned long)task_thread_info(p)->flags); 4232 4233 print_worker_info(KERN_INFO, p); 4234 show_stack(p, NULL); 4235} 4236 4237void show_state_filter(unsigned long state_filter) 4238{ 4239 struct task_struct *g, *p; 4240 4241#if BITS_PER_LONG == 32 4242 printk(KERN_INFO 4243 " task PC stack pid father\n"); 4244#else 4245 printk(KERN_INFO 4246 " task PC stack pid father\n"); 4247#endif 4248 rcu_read_lock(); 4249 do_each_thread(g, p) { 4250 /* 4251 * reset the NMI-timeout, listing all files on a slow 4252 * console might take a lot of time: 4253 */ 4254 touch_nmi_watchdog(); 4255 if (!state_filter || (p->state & state_filter)) 4256 sched_show_task(p); 4257 } while_each_thread(g, p); 4258 4259 touch_all_softlockup_watchdogs(); 4260 4261#ifdef CONFIG_SCHED_DEBUG 4262 sysrq_sched_debug_show(); 4263#endif 4264 rcu_read_unlock(); 4265 /* 4266 * Only show locks if all tasks are dumped: 4267 */ 4268 if (!state_filter) 4269 debug_show_all_locks(); 4270} 4271 4272void init_idle_bootup_task(struct task_struct *idle) 4273{ 4274 idle->sched_class = &idle_sched_class; 4275} 4276 4277/** 4278 * init_idle - set up an idle thread for a given CPU 4279 * @idle: task in question 4280 * @cpu: cpu the idle task belongs to 4281 * 4282 * NOTE: this function does not set the idle thread's NEED_RESCHED 4283 * flag, to make booting more robust. 4284 */ 4285void init_idle(struct task_struct *idle, int cpu) 4286{ 4287 struct rq *rq = cpu_rq(cpu); 4288 unsigned long flags; 4289 4290 raw_spin_lock_irqsave(&rq->lock, flags); 4291 4292 __sched_fork(0, idle); 4293 idle->state = TASK_RUNNING; 4294 idle->se.exec_start = sched_clock(); 4295 4296 do_set_cpus_allowed(idle, cpumask_of(cpu)); 4297 /* 4298 * We're having a chicken and egg problem, even though we are 4299 * holding rq->lock, the cpu isn't yet set to this cpu so the 4300 * lockdep check in task_group() will fail. 4301 * 4302 * Similar case to sched_fork(). / Alternatively we could 4303 * use task_rq_lock() here and obtain the other rq->lock. 4304 * 4305 * Silence PROVE_RCU 4306 */ 4307 rcu_read_lock(); 4308 __set_task_cpu(idle, cpu); 4309 rcu_read_unlock(); 4310 4311 rq->curr = rq->idle = idle; 4312#if defined(CONFIG_SMP) 4313 idle->on_cpu = 1; 4314#endif 4315 raw_spin_unlock_irqrestore(&rq->lock, flags); 4316 4317 /* Set the preempt count _outside_ the spinlocks! */ 4318 init_idle_preempt_count(idle, cpu); 4319 4320 /* 4321 * The idle tasks have their own, simple scheduling class: 4322 */ 4323 idle->sched_class = &idle_sched_class; 4324 ftrace_graph_init_idle_task(idle, cpu); 4325 vtime_init_idle(idle, cpu); 4326#if defined(CONFIG_SMP) 4327 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu); 4328#endif 4329} 4330 4331#ifdef CONFIG_SMP 4332void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask) 4333{ 4334 if (p->sched_class && p->sched_class->set_cpus_allowed) 4335 p->sched_class->set_cpus_allowed(p, new_mask); 4336 4337 cpumask_copy(&p->cpus_allowed, new_mask); 4338 p->nr_cpus_allowed = cpumask_weight(new_mask); 4339} 4340 4341/* 4342 * This is how migration works: 4343 * 4344 * 1) we invoke migration_cpu_stop() on the target CPU using 4345 * stop_one_cpu(). 4346 * 2) stopper starts to run (implicitly forcing the migrated thread 4347 * off the CPU) 4348 * 3) it checks whether the migrated task is still in the wrong runqueue. 4349 * 4) if it's in the wrong runqueue then the migration thread removes 4350 * it and puts it into the right queue. 4351 * 5) stopper completes and stop_one_cpu() returns and the migration 4352 * is done. 4353 */ 4354 4355/* 4356 * Change a given task's CPU affinity. Migrate the thread to a 4357 * proper CPU and schedule it away if the CPU it's executing on 4358 * is removed from the allowed bitmask. 4359 * 4360 * NOTE: the caller must have a valid reference to the task, the 4361 * task must not exit() & deallocate itself prematurely. The 4362 * call is not atomic; no spinlocks may be held. 4363 */ 4364int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask) 4365{ 4366 unsigned long flags; 4367 struct rq *rq; 4368 unsigned int dest_cpu; 4369 int ret = 0; 4370 4371 rq = task_rq_lock(p, &flags); 4372 4373 if (cpumask_equal(&p->cpus_allowed, new_mask)) 4374 goto out; 4375 4376 if (!cpumask_intersects(new_mask, cpu_active_mask)) { 4377 ret = -EINVAL; 4378 goto out; 4379 } 4380 4381 do_set_cpus_allowed(p, new_mask); 4382 4383 /* Can the task run on the task's current CPU? If so, we're done */ 4384 if (cpumask_test_cpu(task_cpu(p), new_mask)) 4385 goto out; 4386 4387 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask); 4388 if (p->on_rq) { 4389 struct migration_arg arg = { p, dest_cpu }; 4390 /* Need help from migration thread: drop lock and wait. */ 4391 task_rq_unlock(rq, p, &flags); 4392 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg); 4393 tlb_migrate_finish(p->mm); 4394 return 0; 4395 } 4396out: 4397 task_rq_unlock(rq, p, &flags); 4398 4399 return ret; 4400} 4401EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr); 4402 4403/* 4404 * Move (not current) task off this cpu, onto dest cpu. We're doing 4405 * this because either it can't run here any more (set_cpus_allowed() 4406 * away from this CPU, or CPU going down), or because we're 4407 * attempting to rebalance this task on exec (sched_exec). 4408 * 4409 * So we race with normal scheduler movements, but that's OK, as long 4410 * as the task is no longer on this CPU. 4411 * 4412 * Returns non-zero if task was successfully migrated. 4413 */ 4414static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu) 4415{ 4416 struct rq *rq_dest, *rq_src; 4417 int ret = 0; 4418 4419 if (unlikely(!cpu_active(dest_cpu))) 4420 return ret; 4421 4422 rq_src = cpu_rq(src_cpu); 4423 rq_dest = cpu_rq(dest_cpu); 4424 4425 raw_spin_lock(&p->pi_lock); 4426 double_rq_lock(rq_src, rq_dest); 4427 /* Already moved. */ 4428 if (task_cpu(p) != src_cpu) 4429 goto done; 4430 /* Affinity changed (again). */ 4431 if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p))) 4432 goto fail; 4433 4434 /* 4435 * If we're not on a rq, the next wake-up will ensure we're 4436 * placed properly. 4437 */ 4438 if (p->on_rq) { 4439 dequeue_task(rq_src, p, 0); 4440 set_task_cpu(p, dest_cpu); 4441 enqueue_task(rq_dest, p, 0); 4442 check_preempt_curr(rq_dest, p, 0); 4443 } 4444done: 4445 ret = 1; 4446fail: 4447 double_rq_unlock(rq_src, rq_dest); 4448 raw_spin_unlock(&p->pi_lock); 4449 return ret; 4450} 4451 4452#ifdef CONFIG_NUMA_BALANCING 4453/* Migrate current task p to target_cpu */ 4454int migrate_task_to(struct task_struct *p, int target_cpu) 4455{ 4456 struct migration_arg arg = { p, target_cpu }; 4457 int curr_cpu = task_cpu(p); 4458 4459 if (curr_cpu == target_cpu) 4460 return 0; 4461 4462 if (!cpumask_test_cpu(target_cpu, tsk_cpus_allowed(p))) 4463 return -EINVAL; 4464 4465 /* TODO: This is not properly updating schedstats */ 4466 4467 return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg); 4468} 4469 4470/* 4471 * Requeue a task on a given node and accurately track the number of NUMA 4472 * tasks on the runqueues 4473 */ 4474void sched_setnuma(struct task_struct *p, int nid) 4475{ 4476 struct rq *rq; 4477 unsigned long flags; 4478 bool on_rq, running; 4479 4480 rq = task_rq_lock(p, &flags); 4481 on_rq = p->on_rq; 4482 running = task_current(rq, p); 4483 4484 if (on_rq) 4485 dequeue_task(rq, p, 0); 4486 if (running) 4487 p->sched_class->put_prev_task(rq, p); 4488 4489 p->numa_preferred_nid = nid; 4490 4491 if (running) 4492 p->sched_class->set_curr_task(rq); 4493 if (on_rq) 4494 enqueue_task(rq, p, 0); 4495 task_rq_unlock(rq, p, &flags); 4496} 4497#endif 4498 4499/* 4500 * migration_cpu_stop - this will be executed by a highprio stopper thread 4501 * and performs thread migration by bumping thread off CPU then 4502 * 'pushing' onto another runqueue. 4503 */ 4504static int migration_cpu_stop(void *data) 4505{ 4506 struct migration_arg *arg = data; 4507 4508 /* 4509 * The original target cpu might have gone down and we might 4510 * be on another cpu but it doesn't matter. 4511 */ 4512 local_irq_disable(); 4513 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu); 4514 local_irq_enable(); 4515 return 0; 4516} 4517 4518#ifdef CONFIG_HOTPLUG_CPU 4519 4520/* 4521 * Ensures that the idle task is using init_mm right before its cpu goes 4522 * offline. 4523 */ 4524void idle_task_exit(void) 4525{ 4526 struct mm_struct *mm = current->active_mm; 4527 4528 BUG_ON(cpu_online(smp_processor_id())); 4529 4530 if (mm != &init_mm) 4531 switch_mm(mm, &init_mm, current); 4532 mmdrop(mm); 4533} 4534 4535/* 4536 * Since this CPU is going 'away' for a while, fold any nr_active delta 4537 * we might have. Assumes we're called after migrate_tasks() so that the 4538 * nr_active count is stable. 4539 * 4540 * Also see the comment "Global load-average calculations". 4541 */ 4542static void calc_load_migrate(struct rq *rq) 4543{ 4544 long delta = calc_load_fold_active(rq); 4545 if (delta) 4546 atomic_long_add(delta, &calc_load_tasks); 4547} 4548 4549/* 4550 * Migrate all tasks from the rq, sleeping tasks will be migrated by 4551 * try_to_wake_up()->select_task_rq(). 4552 * 4553 * Called with rq->lock held even though we'er in stop_machine() and 4554 * there's no concurrency possible, we hold the required locks anyway 4555 * because of lock validation efforts. 4556 */ 4557static void migrate_tasks(unsigned int dead_cpu) 4558{ 4559 struct rq *rq = cpu_rq(dead_cpu); 4560 struct task_struct *next, *stop = rq->stop; 4561 int dest_cpu; 4562 4563 /* 4564 * Fudge the rq selection such that the below task selection loop 4565 * doesn't get stuck on the currently eligible stop task. 4566 * 4567 * We're currently inside stop_machine() and the rq is either stuck 4568 * in the stop_machine_cpu_stop() loop, or we're executing this code, 4569 * either way we should never end up calling schedule() until we're 4570 * done here. 4571 */ 4572 rq->stop = NULL; 4573 4574 /* 4575 * put_prev_task() and pick_next_task() sched 4576 * class method both need to have an up-to-date 4577 * value of rq->clock[_task] 4578 */ 4579 update_rq_clock(rq); 4580 4581 for ( ; ; ) { 4582 /* 4583 * There's this thread running, bail when that's the only 4584 * remaining thread. 4585 */ 4586 if (rq->nr_running == 1) 4587 break; 4588 4589 next = pick_next_task(rq); 4590 BUG_ON(!next); 4591 next->sched_class->put_prev_task(rq, next); 4592 4593 /* Find suitable destination for @next, with force if needed. */ 4594 dest_cpu = select_fallback_rq(dead_cpu, next); 4595 raw_spin_unlock(&rq->lock); 4596 4597 __migrate_task(next, dead_cpu, dest_cpu); 4598 4599 raw_spin_lock(&rq->lock); 4600 } 4601 4602 rq->stop = stop; 4603} 4604 4605#endif /* CONFIG_HOTPLUG_CPU */ 4606 4607#if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL) 4608 4609static struct ctl_table sd_ctl_dir[] = { 4610 { 4611 .procname = "sched_domain", 4612 .mode = 0555, 4613 }, 4614 {} 4615}; 4616 4617static struct ctl_table sd_ctl_root[] = { 4618 { 4619 .procname = "kernel", 4620 .mode = 0555, 4621 .child = sd_ctl_dir, 4622 }, 4623 {} 4624}; 4625 4626static struct ctl_table *sd_alloc_ctl_entry(int n) 4627{ 4628 struct ctl_table *entry = 4629 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL); 4630 4631 return entry; 4632} 4633 4634static void sd_free_ctl_entry(struct ctl_table **tablep) 4635{ 4636 struct ctl_table *entry; 4637 4638 /* 4639 * In the intermediate directories, both the child directory and 4640 * procname are dynamically allocated and could fail but the mode 4641 * will always be set. In the lowest directory the names are 4642 * static strings and all have proc handlers. 4643 */ 4644 for (entry = *tablep; entry->mode; entry++) { 4645 if (entry->child) 4646 sd_free_ctl_entry(&entry->child); 4647 if (entry->proc_handler == NULL) 4648 kfree(entry->procname); 4649 } 4650 4651 kfree(*tablep); 4652 *tablep = NULL; 4653} 4654 4655static int min_load_idx = 0; 4656static int max_load_idx = CPU_LOAD_IDX_MAX-1; 4657 4658static void 4659set_table_entry(struct ctl_table *entry, 4660 const char *procname, void *data, int maxlen, 4661 umode_t mode, proc_handler *proc_handler, 4662 bool load_idx) 4663{ 4664 entry->procname = procname; 4665 entry->data = data; 4666 entry->maxlen = maxlen; 4667 entry->mode = mode; 4668 entry->proc_handler = proc_handler; 4669 4670 if (load_idx) { 4671 entry->extra1 = &min_load_idx; 4672 entry->extra2 = &max_load_idx; 4673 } 4674} 4675 4676static struct ctl_table * 4677sd_alloc_ctl_domain_table(struct sched_domain *sd) 4678{ 4679 struct ctl_table *table = sd_alloc_ctl_entry(13); 4680 4681 if (table == NULL) 4682 return NULL; 4683 4684 set_table_entry(&table[0], "min_interval", &sd->min_interval, 4685 sizeof(long), 0644, proc_doulongvec_minmax, false); 4686 set_table_entry(&table[1], "max_interval", &sd->max_interval, 4687 sizeof(long), 0644, proc_doulongvec_minmax, false); 4688 set_table_entry(&table[2], "busy_idx", &sd->busy_idx, 4689 sizeof(int), 0644, proc_dointvec_minmax, true); 4690 set_table_entry(&table[3], "idle_idx", &sd->idle_idx, 4691 sizeof(int), 0644, proc_dointvec_minmax, true); 4692 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx, 4693 sizeof(int), 0644, proc_dointvec_minmax, true); 4694 set_table_entry(&table[5], "wake_idx", &sd->wake_idx, 4695 sizeof(int), 0644, proc_dointvec_minmax, true); 4696 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx, 4697 sizeof(int), 0644, proc_dointvec_minmax, true); 4698 set_table_entry(&table[7], "busy_factor", &sd->busy_factor, 4699 sizeof(int), 0644, proc_dointvec_minmax, false); 4700 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct, 4701 sizeof(int), 0644, proc_dointvec_minmax, false); 4702 set_table_entry(&table[9], "cache_nice_tries", 4703 &sd->cache_nice_tries, 4704 sizeof(int), 0644, proc_dointvec_minmax, false); 4705 set_table_entry(&table[10], "flags", &sd->flags, 4706 sizeof(int), 0644, proc_dointvec_minmax, false); 4707 set_table_entry(&table[11], "name", sd->name, 4708 CORENAME_MAX_SIZE, 0444, proc_dostring, false); 4709 /* &table[12] is terminator */ 4710 4711 return table; 4712} 4713 4714static struct ctl_table *sd_alloc_ctl_cpu_table(int cpu) 4715{ 4716 struct ctl_table *entry, *table; 4717 struct sched_domain *sd; 4718 int domain_num = 0, i; 4719 char buf[32]; 4720 4721 for_each_domain(cpu, sd) 4722 domain_num++; 4723 entry = table = sd_alloc_ctl_entry(domain_num + 1); 4724 if (table == NULL) 4725 return NULL; 4726 4727 i = 0; 4728 for_each_domain(cpu, sd) { 4729 snprintf(buf, 32, "domain%d", i); 4730 entry->procname = kstrdup(buf, GFP_KERNEL); 4731 entry->mode = 0555; 4732 entry->child = sd_alloc_ctl_domain_table(sd); 4733 entry++; 4734 i++; 4735 } 4736 return table; 4737} 4738 4739static struct ctl_table_header *sd_sysctl_header; 4740static void register_sched_domain_sysctl(void) 4741{ 4742 int i, cpu_num = num_possible_cpus(); 4743 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1); 4744 char buf[32]; 4745 4746 WARN_ON(sd_ctl_dir[0].child); 4747 sd_ctl_dir[0].child = entry; 4748 4749 if (entry == NULL) 4750 return; 4751 4752 for_each_possible_cpu(i) { 4753 snprintf(buf, 32, "cpu%d", i); 4754 entry->procname = kstrdup(buf, GFP_KERNEL); 4755 entry->mode = 0555; 4756 entry->child = sd_alloc_ctl_cpu_table(i); 4757 entry++; 4758 } 4759 4760 WARN_ON(sd_sysctl_header); 4761 sd_sysctl_header = register_sysctl_table(sd_ctl_root); 4762} 4763 4764/* may be called multiple times per register */ 4765static void unregister_sched_domain_sysctl(void) 4766{ 4767 if (sd_sysctl_header) 4768 unregister_sysctl_table(sd_sysctl_header); 4769 sd_sysctl_header = NULL; 4770 if (sd_ctl_dir[0].child) 4771 sd_free_ctl_entry(&sd_ctl_dir[0].child); 4772} 4773#else 4774static void register_sched_domain_sysctl(void) 4775{ 4776} 4777static void unregister_sched_domain_sysctl(void) 4778{ 4779} 4780#endif 4781 4782static void set_rq_online(struct rq *rq) 4783{ 4784 if (!rq->online) { 4785 const struct sched_class *class; 4786 4787 cpumask_set_cpu(rq->cpu, rq->rd->online); 4788 rq->online = 1; 4789 4790 for_each_class(class) { 4791 if (class->rq_online) 4792 class->rq_online(rq); 4793 } 4794 } 4795} 4796 4797static void set_rq_offline(struct rq *rq) 4798{ 4799 if (rq->online) { 4800 const struct sched_class *class; 4801 4802 for_each_class(class) { 4803 if (class->rq_offline) 4804 class->rq_offline(rq); 4805 } 4806 4807 cpumask_clear_cpu(rq->cpu, rq->rd->online); 4808 rq->online = 0; 4809 } 4810} 4811 4812/* 4813 * migration_call - callback that gets triggered when a CPU is added. 4814 * Here we can start up the necessary migration thread for the new CPU. 4815 */ 4816static int 4817migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu) 4818{ 4819 int cpu = (long)hcpu; 4820 unsigned long flags; 4821 struct rq *rq = cpu_rq(cpu); 4822 4823 switch (action & ~CPU_TASKS_FROZEN) { 4824 4825 case CPU_UP_PREPARE: 4826 rq->calc_load_update = calc_load_update; 4827 break; 4828 4829 case CPU_ONLINE: 4830 /* Update our root-domain */ 4831 raw_spin_lock_irqsave(&rq->lock, flags); 4832 if (rq->rd) { 4833 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span)); 4834 4835 set_rq_online(rq); 4836 } 4837 raw_spin_unlock_irqrestore(&rq->lock, flags); 4838 break; 4839 4840#ifdef CONFIG_HOTPLUG_CPU 4841 case CPU_DYING: 4842 sched_ttwu_pending(); 4843 /* Update our root-domain */ 4844 raw_spin_lock_irqsave(&rq->lock, flags); 4845 if (rq->rd) { 4846 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span)); 4847 set_rq_offline(rq); 4848 } 4849 migrate_tasks(cpu); 4850 BUG_ON(rq->nr_running != 1); /* the migration thread */ 4851 raw_spin_unlock_irqrestore(&rq->lock, flags); 4852 break; 4853 4854 case CPU_DEAD: 4855 calc_load_migrate(rq); 4856 break; 4857#endif 4858 } 4859 4860 update_max_interval(); 4861 4862 return NOTIFY_OK; 4863} 4864 4865/* 4866 * Register at high priority so that task migration (migrate_all_tasks) 4867 * happens before everything else. This has to be lower priority than 4868 * the notifier in the perf_event subsystem, though. 4869 */ 4870static struct notifier_block migration_notifier = { 4871 .notifier_call = migration_call, 4872 .priority = CPU_PRI_MIGRATION, 4873}; 4874 4875static int sched_cpu_active(struct notifier_block *nfb, 4876 unsigned long action, void *hcpu) 4877{ 4878 switch (action & ~CPU_TASKS_FROZEN) { 4879 case CPU_STARTING: 4880 case CPU_DOWN_FAILED: 4881 set_cpu_active((long)hcpu, true); 4882 return NOTIFY_OK; 4883 default: 4884 return NOTIFY_DONE; 4885 } 4886} 4887 4888static int sched_cpu_inactive(struct notifier_block *nfb, 4889 unsigned long action, void *hcpu) 4890{ 4891 switch (action & ~CPU_TASKS_FROZEN) { 4892 case CPU_DOWN_PREPARE: 4893 set_cpu_active((long)hcpu, false); 4894 return NOTIFY_OK; 4895 default: 4896 return NOTIFY_DONE; 4897 } 4898} 4899 4900static int __init migration_init(void) 4901{ 4902 void *cpu = (void *)(long)smp_processor_id(); 4903 int err; 4904 4905 /* Initialize migration for the boot CPU */ 4906 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu); 4907 BUG_ON(err == NOTIFY_BAD); 4908 migration_call(&migration_notifier, CPU_ONLINE, cpu); 4909 register_cpu_notifier(&migration_notifier); 4910 4911 /* Register cpu active notifiers */ 4912 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE); 4913 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE); 4914 4915 return 0; 4916} 4917early_initcall(migration_init); 4918#endif 4919 4920#ifdef CONFIG_SMP 4921 4922static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */ 4923 4924#ifdef CONFIG_SCHED_DEBUG 4925 4926static __read_mostly int sched_debug_enabled; 4927 4928static int __init sched_debug_setup(char *str) 4929{ 4930 sched_debug_enabled = 1; 4931 4932 return 0; 4933} 4934early_param("sched_debug", sched_debug_setup); 4935 4936static inline bool sched_debug(void) 4937{ 4938 return sched_debug_enabled; 4939} 4940 4941static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level, 4942 struct cpumask *groupmask) 4943{ 4944 struct sched_group *group = sd->groups; 4945 char str[256]; 4946 4947 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd)); 4948 cpumask_clear(groupmask); 4949 4950 printk(KERN_DEBUG "%*s domain %d: ", level, "", level); 4951 4952 if (!(sd->flags & SD_LOAD_BALANCE)) { 4953 printk("does not load-balance\n"); 4954 if (sd->parent) 4955 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain" 4956 " has parent"); 4957 return -1; 4958 } 4959 4960 printk(KERN_CONT "span %s level %s\n", str, sd->name); 4961 4962 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) { 4963 printk(KERN_ERR "ERROR: domain->span does not contain " 4964 "CPU%d\n", cpu); 4965 } 4966 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) { 4967 printk(KERN_ERR "ERROR: domain->groups does not contain" 4968 " CPU%d\n", cpu); 4969 } 4970 4971 printk(KERN_DEBUG "%*s groups:", level + 1, ""); 4972 do { 4973 if (!group) { 4974 printk("\n"); 4975 printk(KERN_ERR "ERROR: group is NULL\n"); 4976 break; 4977 } 4978 4979 /* 4980 * Even though we initialize ->power to something semi-sane, 4981 * we leave power_orig unset. This allows us to detect if 4982 * domain iteration is still funny without causing /0 traps. 4983 */ 4984 if (!group->sgp->power_orig) { 4985 printk(KERN_CONT "\n"); 4986 printk(KERN_ERR "ERROR: domain->cpu_power not " 4987 "set\n"); 4988 break; 4989 } 4990 4991 if (!cpumask_weight(sched_group_cpus(group))) { 4992 printk(KERN_CONT "\n"); 4993 printk(KERN_ERR "ERROR: empty group\n"); 4994 break; 4995 } 4996 4997 if (!(sd->flags & SD_OVERLAP) && 4998 cpumask_intersects(groupmask, sched_group_cpus(group))) { 4999 printk(KERN_CONT "\n"); 5000 printk(KERN_ERR "ERROR: repeated CPUs\n"); 5001 break; 5002 } 5003 5004 cpumask_or(groupmask, groupmask, sched_group_cpus(group)); 5005 5006 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group)); 5007 5008 printk(KERN_CONT " %s", str); 5009 if (group->sgp->power != SCHED_POWER_SCALE) { 5010 printk(KERN_CONT " (cpu_power = %d)", 5011 group->sgp->power); 5012 } 5013 5014 group = group->next; 5015 } while (group != sd->groups); 5016 printk(KERN_CONT "\n"); 5017 5018 if (!cpumask_equal(sched_domain_span(sd), groupmask)) 5019 printk(KERN_ERR "ERROR: groups don't span domain->span\n"); 5020 5021 if (sd->parent && 5022 !cpumask_subset(groupmask, sched_domain_span(sd->parent))) 5023 printk(KERN_ERR "ERROR: parent span is not a superset " 5024 "of domain->span\n"); 5025 return 0; 5026} 5027 5028static void sched_domain_debug(struct sched_domain *sd, int cpu) 5029{ 5030 int level = 0; 5031 5032 if (!sched_debug_enabled) 5033 return; 5034 5035 if (!sd) { 5036 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu); 5037 return; 5038 } 5039 5040 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu); 5041 5042 for (;;) { 5043 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask)) 5044 break; 5045 level++; 5046 sd = sd->parent; 5047 if (!sd) 5048 break; 5049 } 5050} 5051#else /* !CONFIG_SCHED_DEBUG */ 5052# define sched_domain_debug(sd, cpu) do { } while (0) 5053static inline bool sched_debug(void) 5054{ 5055 return false; 5056} 5057#endif /* CONFIG_SCHED_DEBUG */ 5058 5059static int sd_degenerate(struct sched_domain *sd) 5060{ 5061 if (cpumask_weight(sched_domain_span(sd)) == 1) 5062 return 1; 5063 5064 /* Following flags need at least 2 groups */ 5065 if (sd->flags & (SD_LOAD_BALANCE | 5066 SD_BALANCE_NEWIDLE | 5067 SD_BALANCE_FORK | 5068 SD_BALANCE_EXEC | 5069 SD_SHARE_CPUPOWER | 5070 SD_SHARE_PKG_RESOURCES)) { 5071 if (sd->groups != sd->groups->next) 5072 return 0; 5073 } 5074 5075 /* Following flags don't use groups */ 5076 if (sd->flags & (SD_WAKE_AFFINE)) 5077 return 0; 5078 5079 return 1; 5080} 5081 5082static int 5083sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent) 5084{ 5085 unsigned long cflags = sd->flags, pflags = parent->flags; 5086 5087 if (sd_degenerate(parent)) 5088 return 1; 5089 5090 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent))) 5091 return 0; 5092 5093 /* Flags needing groups don't count if only 1 group in parent */ 5094 if (parent->groups == parent->groups->next) { 5095 pflags &= ~(SD_LOAD_BALANCE | 5096 SD_BALANCE_NEWIDLE | 5097 SD_BALANCE_FORK | 5098 SD_BALANCE_EXEC | 5099 SD_SHARE_CPUPOWER | 5100 SD_SHARE_PKG_RESOURCES | 5101 SD_PREFER_SIBLING); 5102 if (nr_node_ids == 1) 5103 pflags &= ~SD_SERIALIZE; 5104 } 5105 if (~cflags & pflags) 5106 return 0; 5107 5108 return 1; 5109} 5110 5111static void free_rootdomain(struct rcu_head *rcu) 5112{ 5113 struct root_domain *rd = container_of(rcu, struct root_domain, rcu); 5114 5115 cpupri_cleanup(&rd->cpupri); 5116 free_cpumask_var(rd->rto_mask); 5117 free_cpumask_var(rd->online); 5118 free_cpumask_var(rd->span); 5119 kfree(rd); 5120} 5121 5122static void rq_attach_root(struct rq *rq, struct root_domain *rd) 5123{ 5124 struct root_domain *old_rd = NULL; 5125 unsigned long flags; 5126 5127 raw_spin_lock_irqsave(&rq->lock, flags); 5128 5129 if (rq->rd) { 5130 old_rd = rq->rd; 5131 5132 if (cpumask_test_cpu(rq->cpu, old_rd->online)) 5133 set_rq_offline(rq); 5134 5135 cpumask_clear_cpu(rq->cpu, old_rd->span); 5136 5137 /* 5138 * If we dont want to free the old_rt yet then 5139 * set old_rd to NULL to skip the freeing later 5140 * in this function: 5141 */ 5142 if (!atomic_dec_and_test(&old_rd->refcount)) 5143 old_rd = NULL; 5144 } 5145 5146 atomic_inc(&rd->refcount); 5147 rq->rd = rd; 5148 5149 cpumask_set_cpu(rq->cpu, rd->span); 5150 if (cpumask_test_cpu(rq->cpu, cpu_active_mask)) 5151 set_rq_online(rq); 5152 5153 raw_spin_unlock_irqrestore(&rq->lock, flags); 5154 5155 if (old_rd) 5156 call_rcu_sched(&old_rd->rcu, free_rootdomain); 5157} 5158 5159static int init_rootdomain(struct root_domain *rd) 5160{ 5161 memset(rd, 0, sizeof(*rd)); 5162 5163 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL)) 5164 goto out; 5165 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL)) 5166 goto free_span; 5167 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL)) 5168 goto free_online; 5169 5170 if (cpupri_init(&rd->cpupri) != 0) 5171 goto free_rto_mask; 5172 return 0; 5173 5174free_rto_mask: 5175 free_cpumask_var(rd->rto_mask); 5176free_online: 5177 free_cpumask_var(rd->online); 5178free_span: 5179 free_cpumask_var(rd->span); 5180out: 5181 return -ENOMEM; 5182} 5183 5184/* 5185 * By default the system creates a single root-domain with all cpus as 5186 * members (mimicking the global state we have today). 5187 */ 5188struct root_domain def_root_domain; 5189 5190static void init_defrootdomain(void) 5191{ 5192 init_rootdomain(&def_root_domain); 5193 5194 atomic_set(&def_root_domain.refcount, 1); 5195} 5196 5197static struct root_domain *alloc_rootdomain(void) 5198{ 5199 struct root_domain *rd; 5200 5201 rd = kmalloc(sizeof(*rd), GFP_KERNEL); 5202 if (!rd) 5203 return NULL; 5204 5205 if (init_rootdomain(rd) != 0) { 5206 kfree(rd); 5207 return NULL; 5208 } 5209 5210 return rd; 5211} 5212 5213static void free_sched_groups(struct sched_group *sg, int free_sgp) 5214{ 5215 struct sched_group *tmp, *first; 5216 5217 if (!sg) 5218 return; 5219 5220 first = sg; 5221 do { 5222 tmp = sg->next; 5223 5224 if (free_sgp && atomic_dec_and_test(&sg->sgp->ref)) 5225 kfree(sg->sgp); 5226 5227 kfree(sg); 5228 sg = tmp; 5229 } while (sg != first); 5230} 5231 5232static void free_sched_domain(struct rcu_head *rcu) 5233{ 5234 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu); 5235 5236 /* 5237 * If its an overlapping domain it has private groups, iterate and 5238 * nuke them all. 5239 */ 5240 if (sd->flags & SD_OVERLAP) { 5241 free_sched_groups(sd->groups, 1); 5242 } else if (atomic_dec_and_test(&sd->groups->ref)) { 5243 kfree(sd->groups->sgp); 5244 kfree(sd->groups); 5245 } 5246 kfree(sd); 5247} 5248 5249static void destroy_sched_domain(struct sched_domain *sd, int cpu) 5250{ 5251 call_rcu(&sd->rcu, free_sched_domain); 5252} 5253 5254static void destroy_sched_domains(struct sched_domain *sd, int cpu) 5255{ 5256 for (; sd; sd = sd->parent) 5257 destroy_sched_domain(sd, cpu); 5258} 5259 5260/* 5261 * Keep a special pointer to the highest sched_domain that has 5262 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this 5263 * allows us to avoid some pointer chasing select_idle_sibling(). 5264 * 5265 * Also keep a unique ID per domain (we use the first cpu number in 5266 * the cpumask of the domain), this allows us to quickly tell if 5267 * two cpus are in the same cache domain, see cpus_share_cache(). 5268 */ 5269DEFINE_PER_CPU(struct sched_domain *, sd_llc); 5270DEFINE_PER_CPU(int, sd_llc_size); 5271DEFINE_PER_CPU(int, sd_llc_id); 5272DEFINE_PER_CPU(struct sched_domain *, sd_numa); 5273 5274static void update_top_cache_domain(int cpu) 5275{ 5276 struct sched_domain *sd; 5277 int id = cpu; 5278 int size = 1; 5279 5280 sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES); 5281 if (sd) { 5282 id = cpumask_first(sched_domain_span(sd)); 5283 size = cpumask_weight(sched_domain_span(sd)); 5284 } 5285 5286 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd); 5287 per_cpu(sd_llc_size, cpu) = size; 5288 per_cpu(sd_llc_id, cpu) = id; 5289 5290 sd = lowest_flag_domain(cpu, SD_NUMA); 5291 rcu_assign_pointer(per_cpu(sd_numa, cpu), sd); 5292} 5293 5294/* 5295 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must 5296 * hold the hotplug lock. 5297 */ 5298static void 5299cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu) 5300{ 5301 struct rq *rq = cpu_rq(cpu); 5302 struct sched_domain *tmp; 5303 5304 /* Remove the sched domains which do not contribute to scheduling. */ 5305 for (tmp = sd; tmp; ) { 5306 struct sched_domain *parent = tmp->parent; 5307 if (!parent) 5308 break; 5309 5310 if (sd_parent_degenerate(tmp, parent)) { 5311 tmp->parent = parent->parent; 5312 if (parent->parent) 5313 parent->parent->child = tmp; 5314 /* 5315 * Transfer SD_PREFER_SIBLING down in case of a 5316 * degenerate parent; the spans match for this 5317 * so the property transfers. 5318 */ 5319 if (parent->flags & SD_PREFER_SIBLING) 5320 tmp->flags |= SD_PREFER_SIBLING; 5321 destroy_sched_domain(parent, cpu); 5322 } else 5323 tmp = tmp->parent; 5324 } 5325 5326 if (sd && sd_degenerate(sd)) { 5327 tmp = sd; 5328 sd = sd->parent; 5329 destroy_sched_domain(tmp, cpu); 5330 if (sd) 5331 sd->child = NULL; 5332 } 5333 5334 sched_domain_debug(sd, cpu); 5335 5336 rq_attach_root(rq, rd); 5337 tmp = rq->sd; 5338 rcu_assign_pointer(rq->sd, sd); 5339 destroy_sched_domains(tmp, cpu); 5340 5341 update_top_cache_domain(cpu); 5342} 5343 5344/* cpus with isolated domains */ 5345static cpumask_var_t cpu_isolated_map; 5346 5347/* Setup the mask of cpus configured for isolated domains */ 5348static int __init isolated_cpu_setup(char *str) 5349{ 5350 alloc_bootmem_cpumask_var(&cpu_isolated_map); 5351 cpulist_parse(str, cpu_isolated_map); 5352 return 1; 5353} 5354 5355__setup("isolcpus=", isolated_cpu_setup); 5356 5357static const struct cpumask *cpu_cpu_mask(int cpu) 5358{ 5359 return cpumask_of_node(cpu_to_node(cpu)); 5360} 5361 5362struct sd_data { 5363 struct sched_domain **__percpu sd; 5364 struct sched_group **__percpu sg; 5365 struct sched_group_power **__percpu sgp; 5366}; 5367 5368struct s_data { 5369 struct sched_domain ** __percpu sd; 5370 struct root_domain *rd; 5371}; 5372 5373enum s_alloc { 5374 sa_rootdomain, 5375 sa_sd, 5376 sa_sd_storage, 5377 sa_none, 5378}; 5379 5380struct sched_domain_topology_level; 5381 5382typedef struct sched_domain *(*sched_domain_init_f)(struct sched_domain_topology_level *tl, int cpu); 5383typedef const struct cpumask *(*sched_domain_mask_f)(int cpu); 5384 5385#define SDTL_OVERLAP 0x01 5386 5387struct sched_domain_topology_level { 5388 sched_domain_init_f init; 5389 sched_domain_mask_f mask; 5390 int flags; 5391 int numa_level; 5392 struct sd_data data; 5393}; 5394 5395/* 5396 * Build an iteration mask that can exclude certain CPUs from the upwards 5397 * domain traversal. 5398 * 5399 * Asymmetric node setups can result in situations where the domain tree is of 5400 * unequal depth, make sure to skip domains that already cover the entire 5401 * range. 5402 * 5403 * In that case build_sched_domains() will have terminated the iteration early 5404 * and our sibling sd spans will be empty. Domains should always include the 5405 * cpu they're built on, so check that. 5406 * 5407 */ 5408static void build_group_mask(struct sched_domain *sd, struct sched_group *sg) 5409{ 5410 const struct cpumask *span = sched_domain_span(sd); 5411 struct sd_data *sdd = sd->private; 5412 struct sched_domain *sibling; 5413 int i; 5414 5415 for_each_cpu(i, span) { 5416 sibling = *per_cpu_ptr(sdd->sd, i); 5417 if (!cpumask_test_cpu(i, sched_domain_span(sibling))) 5418 continue; 5419 5420 cpumask_set_cpu(i, sched_group_mask(sg)); 5421 } 5422} 5423 5424/* 5425 * Return the canonical balance cpu for this group, this is the first cpu 5426 * of this group that's also in the iteration mask. 5427 */ 5428int group_balance_cpu(struct sched_group *sg) 5429{ 5430 return cpumask_first_and(sched_group_cpus(sg), sched_group_mask(sg)); 5431} 5432 5433static int 5434build_overlap_sched_groups(struct sched_domain *sd, int cpu) 5435{ 5436 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg; 5437 const struct cpumask *span = sched_domain_span(sd); 5438 struct cpumask *covered = sched_domains_tmpmask; 5439 struct sd_data *sdd = sd->private; 5440 struct sched_domain *child; 5441 int i; 5442 5443 cpumask_clear(covered); 5444 5445 for_each_cpu(i, span) { 5446 struct cpumask *sg_span; 5447 5448 if (cpumask_test_cpu(i, covered)) 5449 continue; 5450 5451 child = *per_cpu_ptr(sdd->sd, i); 5452 5453 /* See the comment near build_group_mask(). */ 5454 if (!cpumask_test_cpu(i, sched_domain_span(child))) 5455 continue; 5456 5457 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(), 5458 GFP_KERNEL, cpu_to_node(cpu)); 5459 5460 if (!sg) 5461 goto fail; 5462 5463 sg_span = sched_group_cpus(sg); 5464 if (child->child) { 5465 child = child->child; 5466 cpumask_copy(sg_span, sched_domain_span(child)); 5467 } else 5468 cpumask_set_cpu(i, sg_span); 5469 5470 cpumask_or(covered, covered, sg_span); 5471 5472 sg->sgp = *per_cpu_ptr(sdd->sgp, i); 5473 if (atomic_inc_return(&sg->sgp->ref) == 1) 5474 build_group_mask(sd, sg); 5475 5476 /* 5477 * Initialize sgp->power such that even if we mess up the 5478 * domains and no possible iteration will get us here, we won't 5479 * die on a /0 trap. 5480 */ 5481 sg->sgp->power = SCHED_POWER_SCALE * cpumask_weight(sg_span); 5482 5483 /* 5484 * Make sure the first group of this domain contains the 5485 * canonical balance cpu. Otherwise the sched_domain iteration 5486 * breaks. See update_sg_lb_stats(). 5487 */ 5488 if ((!groups && cpumask_test_cpu(cpu, sg_span)) || 5489 group_balance_cpu(sg) == cpu) 5490 groups = sg; 5491 5492 if (!first) 5493 first = sg; 5494 if (last) 5495 last->next = sg; 5496 last = sg; 5497 last->next = first; 5498 } 5499 sd->groups = groups; 5500 5501 return 0; 5502 5503fail: 5504 free_sched_groups(first, 0); 5505 5506 return -ENOMEM; 5507} 5508 5509static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg) 5510{ 5511 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu); 5512 struct sched_domain *child = sd->child; 5513 5514 if (child) 5515 cpu = cpumask_first(sched_domain_span(child)); 5516 5517 if (sg) { 5518 *sg = *per_cpu_ptr(sdd->sg, cpu); 5519 (*sg)->sgp = *per_cpu_ptr(sdd->sgp, cpu); 5520 atomic_set(&(*sg)->sgp->ref, 1); /* for claim_allocations */ 5521 } 5522 5523 return cpu; 5524} 5525 5526/* 5527 * build_sched_groups will build a circular linked list of the groups 5528 * covered by the given span, and will set each group's ->cpumask correctly, 5529 * and ->cpu_power to 0. 5530 * 5531 * Assumes the sched_domain tree is fully constructed 5532 */ 5533static int 5534build_sched_groups(struct sched_domain *sd, int cpu) 5535{ 5536 struct sched_group *first = NULL, *last = NULL; 5537 struct sd_data *sdd = sd->private; 5538 const struct cpumask *span = sched_domain_span(sd); 5539 struct cpumask *covered; 5540 int i; 5541 5542 get_group(cpu, sdd, &sd->groups); 5543 atomic_inc(&sd->groups->ref); 5544 5545 if (cpu != cpumask_first(span)) 5546 return 0; 5547 5548 lockdep_assert_held(&sched_domains_mutex); 5549 covered = sched_domains_tmpmask; 5550 5551 cpumask_clear(covered); 5552 5553 for_each_cpu(i, span) { 5554 struct sched_group *sg; 5555 int group, j; 5556 5557 if (cpumask_test_cpu(i, covered)) 5558 continue; 5559 5560 group = get_group(i, sdd, &sg); 5561 cpumask_clear(sched_group_cpus(sg)); 5562 sg->sgp->power = 0; 5563 cpumask_setall(sched_group_mask(sg)); 5564 5565 for_each_cpu(j, span) { 5566 if (get_group(j, sdd, NULL) != group) 5567 continue; 5568 5569 cpumask_set_cpu(j, covered); 5570 cpumask_set_cpu(j, sched_group_cpus(sg)); 5571 } 5572 5573 if (!first) 5574 first = sg; 5575 if (last) 5576 last->next = sg; 5577 last = sg; 5578 } 5579 last->next = first; 5580 5581 return 0; 5582} 5583 5584/* 5585 * Initialize sched groups cpu_power. 5586 * 5587 * cpu_power indicates the capacity of sched group, which is used while 5588 * distributing the load between different sched groups in a sched domain. 5589 * Typically cpu_power for all the groups in a sched domain will be same unless 5590 * there are asymmetries in the topology. If there are asymmetries, group 5591 * having more cpu_power will pickup more load compared to the group having 5592 * less cpu_power. 5593 */ 5594static void init_sched_groups_power(int cpu, struct sched_domain *sd) 5595{ 5596 struct sched_group *sg = sd->groups; 5597 5598 WARN_ON(!sg); 5599 5600 do { 5601 sg->group_weight = cpumask_weight(sched_group_cpus(sg)); 5602 sg = sg->next; 5603 } while (sg != sd->groups); 5604 5605 if (cpu != group_balance_cpu(sg)) 5606 return; 5607 5608 update_group_power(sd, cpu); 5609 atomic_set(&sg->sgp->nr_busy_cpus, sg->group_weight); 5610} 5611 5612int __weak arch_sd_sibling_asym_packing(void) 5613{ 5614 return 0*SD_ASYM_PACKING; 5615} 5616 5617/* 5618 * Initializers for schedule domains 5619 * Non-inlined to reduce accumulated stack pressure in build_sched_domains() 5620 */ 5621 5622#ifdef CONFIG_SCHED_DEBUG 5623# define SD_INIT_NAME(sd, type) sd->name = #type 5624#else 5625# define SD_INIT_NAME(sd, type) do { } while (0) 5626#endif 5627 5628#define SD_INIT_FUNC(type) \ 5629static noinline struct sched_domain * \ 5630sd_init_##type(struct sched_domain_topology_level *tl, int cpu) \ 5631{ \ 5632 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); \ 5633 *sd = SD_##type##_INIT; \ 5634 SD_INIT_NAME(sd, type); \ 5635 sd->private = &tl->data; \ 5636 return sd; \ 5637} 5638 5639SD_INIT_FUNC(CPU) 5640#ifdef CONFIG_SCHED_SMT 5641 SD_INIT_FUNC(SIBLING) 5642#endif 5643#ifdef CONFIG_SCHED_MC 5644 SD_INIT_FUNC(MC) 5645#endif 5646#ifdef CONFIG_SCHED_BOOK 5647 SD_INIT_FUNC(BOOK) 5648#endif 5649 5650static int default_relax_domain_level = -1; 5651int sched_domain_level_max; 5652 5653static int __init setup_relax_domain_level(char *str) 5654{ 5655 if (kstrtoint(str, 0, &default_relax_domain_level)) 5656 pr_warn("Unable to set relax_domain_level\n"); 5657 5658 return 1; 5659} 5660__setup("relax_domain_level=", setup_relax_domain_level); 5661 5662static void set_domain_attribute(struct sched_domain *sd, 5663 struct sched_domain_attr *attr) 5664{ 5665 int request; 5666 5667 if (!attr || attr->relax_domain_level < 0) { 5668 if (default_relax_domain_level < 0) 5669 return; 5670 else 5671 request = default_relax_domain_level; 5672 } else 5673 request = attr->relax_domain_level; 5674 if (request < sd->level) { 5675 /* turn off idle balance on this domain */ 5676 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE); 5677 } else { 5678 /* turn on idle balance on this domain */ 5679 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE); 5680 } 5681} 5682 5683static void __sdt_free(const struct cpumask *cpu_map); 5684static int __sdt_alloc(const struct cpumask *cpu_map); 5685 5686static void __free_domain_allocs(struct s_data *d, enum s_alloc what, 5687 const struct cpumask *cpu_map) 5688{ 5689 switch (what) { 5690 case sa_rootdomain: 5691 if (!atomic_read(&d->rd->refcount)) 5692 free_rootdomain(&d->rd->rcu); /* fall through */ 5693 case sa_sd: 5694 free_percpu(d->sd); /* fall through */ 5695 case sa_sd_storage: 5696 __sdt_free(cpu_map); /* fall through */ 5697 case sa_none: 5698 break; 5699 } 5700} 5701 5702static enum s_alloc __visit_domain_allocation_hell(struct s_data *d, 5703 const struct cpumask *cpu_map) 5704{ 5705 memset(d, 0, sizeof(*d)); 5706 5707 if (__sdt_alloc(cpu_map)) 5708 return sa_sd_storage; 5709 d->sd = alloc_percpu(struct sched_domain *); 5710 if (!d->sd) 5711 return sa_sd_storage; 5712 d->rd = alloc_rootdomain(); 5713 if (!d->rd) 5714 return sa_sd; 5715 return sa_rootdomain; 5716} 5717 5718/* 5719 * NULL the sd_data elements we've used to build the sched_domain and 5720 * sched_group structure so that the subsequent __free_domain_allocs() 5721 * will not free the data we're using. 5722 */ 5723static void claim_allocations(int cpu, struct sched_domain *sd) 5724{ 5725 struct sd_data *sdd = sd->private; 5726 5727 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd); 5728 *per_cpu_ptr(sdd->sd, cpu) = NULL; 5729 5730 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref)) 5731 *per_cpu_ptr(sdd->sg, cpu) = NULL; 5732 5733 if (atomic_read(&(*per_cpu_ptr(sdd->sgp, cpu))->ref)) 5734 *per_cpu_ptr(sdd->sgp, cpu) = NULL; 5735} 5736 5737#ifdef CONFIG_SCHED_SMT 5738static const struct cpumask *cpu_smt_mask(int cpu) 5739{ 5740 return topology_thread_cpumask(cpu); 5741} 5742#endif 5743 5744/* 5745 * Topology list, bottom-up. 5746 */ 5747static struct sched_domain_topology_level default_topology[] = { 5748#ifdef CONFIG_SCHED_SMT 5749 { sd_init_SIBLING, cpu_smt_mask, }, 5750#endif 5751#ifdef CONFIG_SCHED_MC 5752 { sd_init_MC, cpu_coregroup_mask, }, 5753#endif 5754#ifdef CONFIG_SCHED_BOOK 5755 { sd_init_BOOK, cpu_book_mask, }, 5756#endif 5757 { sd_init_CPU, cpu_cpu_mask, }, 5758 { NULL, }, 5759}; 5760 5761static struct sched_domain_topology_level *sched_domain_topology = default_topology; 5762 5763#define for_each_sd_topology(tl) \ 5764 for (tl = sched_domain_topology; tl->init; tl++) 5765 5766#ifdef CONFIG_NUMA 5767 5768static int sched_domains_numa_levels; 5769static int *sched_domains_numa_distance; 5770static struct cpumask ***sched_domains_numa_masks; 5771static int sched_domains_curr_level; 5772 5773static inline int sd_local_flags(int level) 5774{ 5775 if (sched_domains_numa_distance[level] > RECLAIM_DISTANCE) 5776 return 0; 5777 5778 return SD_BALANCE_EXEC | SD_BALANCE_FORK | SD_WAKE_AFFINE; 5779} 5780 5781static struct sched_domain * 5782sd_numa_init(struct sched_domain_topology_level *tl, int cpu) 5783{ 5784 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); 5785 int level = tl->numa_level; 5786 int sd_weight = cpumask_weight( 5787 sched_domains_numa_masks[level][cpu_to_node(cpu)]); 5788 5789 *sd = (struct sched_domain){ 5790 .min_interval = sd_weight, 5791 .max_interval = 2*sd_weight, 5792 .busy_factor = 32, 5793 .imbalance_pct = 125, 5794 .cache_nice_tries = 2, 5795 .busy_idx = 3, 5796 .idle_idx = 2, 5797 .newidle_idx = 0, 5798 .wake_idx = 0, 5799 .forkexec_idx = 0, 5800 5801 .flags = 1*SD_LOAD_BALANCE 5802 | 1*SD_BALANCE_NEWIDLE 5803 | 0*SD_BALANCE_EXEC 5804 | 0*SD_BALANCE_FORK 5805 | 0*SD_BALANCE_WAKE 5806 | 0*SD_WAKE_AFFINE 5807 | 0*SD_SHARE_CPUPOWER 5808 | 0*SD_SHARE_PKG_RESOURCES 5809 | 1*SD_SERIALIZE 5810 | 0*SD_PREFER_SIBLING 5811 | 1*SD_NUMA 5812 | sd_local_flags(level) 5813 , 5814 .last_balance = jiffies, 5815 .balance_interval = sd_weight, 5816 }; 5817 SD_INIT_NAME(sd, NUMA); 5818 sd->private = &tl->data; 5819 5820 /* 5821 * Ugly hack to pass state to sd_numa_mask()... 5822 */ 5823 sched_domains_curr_level = tl->numa_level; 5824 5825 return sd; 5826} 5827 5828static const struct cpumask *sd_numa_mask(int cpu) 5829{ 5830 return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)]; 5831} 5832 5833static void sched_numa_warn(const char *str) 5834{ 5835 static int done = false; 5836 int i,j; 5837 5838 if (done) 5839 return; 5840 5841 done = true; 5842 5843 printk(KERN_WARNING "ERROR: %s\n\n", str); 5844 5845 for (i = 0; i < nr_node_ids; i++) { 5846 printk(KERN_WARNING " "); 5847 for (j = 0; j < nr_node_ids; j++) 5848 printk(KERN_CONT "%02d ", node_distance(i,j)); 5849 printk(KERN_CONT "\n"); 5850 } 5851 printk(KERN_WARNING "\n"); 5852} 5853 5854static bool find_numa_distance(int distance) 5855{ 5856 int i; 5857 5858 if (distance == node_distance(0, 0)) 5859 return true; 5860 5861 for (i = 0; i < sched_domains_numa_levels; i++) { 5862 if (sched_domains_numa_distance[i] == distance) 5863 return true; 5864 } 5865 5866 return false; 5867} 5868 5869static void sched_init_numa(void) 5870{ 5871 int next_distance, curr_distance = node_distance(0, 0); 5872 struct sched_domain_topology_level *tl; 5873 int level = 0; 5874 int i, j, k; 5875 5876 sched_domains_numa_distance = kzalloc(sizeof(int) * nr_node_ids, GFP_KERNEL); 5877 if (!sched_domains_numa_distance) 5878 return; 5879 5880 /* 5881 * O(nr_nodes^2) deduplicating selection sort -- in order to find the 5882 * unique distances in the node_distance() table. 5883 * 5884 * Assumes node_distance(0,j) includes all distances in 5885 * node_distance(i,j) in order to avoid cubic time. 5886 */ 5887 next_distance = curr_distance; 5888 for (i = 0; i < nr_node_ids; i++) { 5889 for (j = 0; j < nr_node_ids; j++) { 5890 for (k = 0; k < nr_node_ids; k++) { 5891 int distance = node_distance(i, k); 5892 5893 if (distance > curr_distance && 5894 (distance < next_distance || 5895 next_distance == curr_distance)) 5896 next_distance = distance; 5897 5898 /* 5899 * While not a strong assumption it would be nice to know 5900 * about cases where if node A is connected to B, B is not 5901 * equally connected to A. 5902 */ 5903 if (sched_debug() && node_distance(k, i) != distance) 5904 sched_numa_warn("Node-distance not symmetric"); 5905 5906 if (sched_debug() && i && !find_numa_distance(distance)) 5907 sched_numa_warn("Node-0 not representative"); 5908 } 5909 if (next_distance != curr_distance) { 5910 sched_domains_numa_distance[level++] = next_distance; 5911 sched_domains_numa_levels = level; 5912 curr_distance = next_distance; 5913 } else break; 5914 } 5915 5916 /* 5917 * In case of sched_debug() we verify the above assumption. 5918 */ 5919 if (!sched_debug()) 5920 break; 5921 } 5922 /* 5923 * 'level' contains the number of unique distances, excluding the 5924 * identity distance node_distance(i,i). 5925 * 5926 * The sched_domains_numa_distance[] array includes the actual distance 5927 * numbers. 5928 */ 5929 5930 /* 5931 * Here, we should temporarily reset sched_domains_numa_levels to 0. 5932 * If it fails to allocate memory for array sched_domains_numa_masks[][], 5933 * the array will contain less then 'level' members. This could be 5934 * dangerous when we use it to iterate array sched_domains_numa_masks[][] 5935 * in other functions. 5936 * 5937 * We reset it to 'level' at the end of this function. 5938 */ 5939 sched_domains_numa_levels = 0; 5940 5941 sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL); 5942 if (!sched_domains_numa_masks) 5943 return; 5944 5945 /* 5946 * Now for each level, construct a mask per node which contains all 5947 * cpus of nodes that are that many hops away from us. 5948 */ 5949 for (i = 0; i < level; i++) { 5950 sched_domains_numa_masks[i] = 5951 kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL); 5952 if (!sched_domains_numa_masks[i]) 5953 return; 5954 5955 for (j = 0; j < nr_node_ids; j++) { 5956 struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL); 5957 if (!mask) 5958 return; 5959 5960 sched_domains_numa_masks[i][j] = mask; 5961 5962 for (k = 0; k < nr_node_ids; k++) { 5963 if (node_distance(j, k) > sched_domains_numa_distance[i]) 5964 continue; 5965 5966 cpumask_or(mask, mask, cpumask_of_node(k)); 5967 } 5968 } 5969 } 5970 5971 tl = kzalloc((ARRAY_SIZE(default_topology) + level) * 5972 sizeof(struct sched_domain_topology_level), GFP_KERNEL); 5973 if (!tl) 5974 return; 5975 5976 /* 5977 * Copy the default topology bits.. 5978 */ 5979 for (i = 0; default_topology[i].init; i++) 5980 tl[i] = default_topology[i]; 5981 5982 /* 5983 * .. and append 'j' levels of NUMA goodness. 5984 */ 5985 for (j = 0; j < level; i++, j++) { 5986 tl[i] = (struct sched_domain_topology_level){ 5987 .init = sd_numa_init, 5988 .mask = sd_numa_mask, 5989 .flags = SDTL_OVERLAP, 5990 .numa_level = j, 5991 }; 5992 } 5993 5994 sched_domain_topology = tl; 5995 5996 sched_domains_numa_levels = level; 5997} 5998 5999static void sched_domains_numa_masks_set(int cpu) 6000{ 6001 int i, j; 6002 int node = cpu_to_node(cpu); 6003 6004 for (i = 0; i < sched_domains_numa_levels; i++) { 6005 for (j = 0; j < nr_node_ids; j++) { 6006 if (node_distance(j, node) <= sched_domains_numa_distance[i]) 6007 cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]); 6008 } 6009 } 6010} 6011 6012static void sched_domains_numa_masks_clear(int cpu) 6013{ 6014 int i, j; 6015 for (i = 0; i < sched_domains_numa_levels; i++) { 6016 for (j = 0; j < nr_node_ids; j++) 6017 cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]); 6018 } 6019} 6020 6021/* 6022 * Update sched_domains_numa_masks[level][node] array when new cpus 6023 * are onlined. 6024 */ 6025static int sched_domains_numa_masks_update(struct notifier_block *nfb, 6026 unsigned long action, 6027 void *hcpu) 6028{ 6029 int cpu = (long)hcpu; 6030 6031 switch (action & ~CPU_TASKS_FROZEN) { 6032 case CPU_ONLINE: 6033 sched_domains_numa_masks_set(cpu); 6034 break; 6035 6036 case CPU_DEAD: 6037 sched_domains_numa_masks_clear(cpu); 6038 break; 6039 6040 default: 6041 return NOTIFY_DONE; 6042 } 6043 6044 return NOTIFY_OK; 6045} 6046#else 6047static inline void sched_init_numa(void) 6048{ 6049} 6050 6051static int sched_domains_numa_masks_update(struct notifier_block *nfb, 6052 unsigned long action, 6053 void *hcpu) 6054{ 6055 return 0; 6056} 6057#endif /* CONFIG_NUMA */ 6058 6059static int __sdt_alloc(const struct cpumask *cpu_map) 6060{ 6061 struct sched_domain_topology_level *tl; 6062 int j; 6063 6064 for_each_sd_topology(tl) { 6065 struct sd_data *sdd = &tl->data; 6066 6067 sdd->sd = alloc_percpu(struct sched_domain *); 6068 if (!sdd->sd) 6069 return -ENOMEM; 6070 6071 sdd->sg = alloc_percpu(struct sched_group *); 6072 if (!sdd->sg) 6073 return -ENOMEM; 6074 6075 sdd->sgp = alloc_percpu(struct sched_group_power *); 6076 if (!sdd->sgp) 6077 return -ENOMEM; 6078 6079 for_each_cpu(j, cpu_map) { 6080 struct sched_domain *sd; 6081 struct sched_group *sg; 6082 struct sched_group_power *sgp; 6083 6084 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(), 6085 GFP_KERNEL, cpu_to_node(j)); 6086 if (!sd) 6087 return -ENOMEM; 6088 6089 *per_cpu_ptr(sdd->sd, j) = sd; 6090 6091 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(), 6092 GFP_KERNEL, cpu_to_node(j)); 6093 if (!sg) 6094 return -ENOMEM; 6095 6096 sg->next = sg; 6097 6098 *per_cpu_ptr(sdd->sg, j) = sg; 6099 6100 sgp = kzalloc_node(sizeof(struct sched_group_power) + cpumask_size(), 6101 GFP_KERNEL, cpu_to_node(j)); 6102 if (!sgp) 6103 return -ENOMEM; 6104 6105 *per_cpu_ptr(sdd->sgp, j) = sgp; 6106 } 6107 } 6108 6109 return 0; 6110} 6111 6112static void __sdt_free(const struct cpumask *cpu_map) 6113{ 6114 struct sched_domain_topology_level *tl; 6115 int j; 6116 6117 for_each_sd_topology(tl) { 6118 struct sd_data *sdd = &tl->data; 6119 6120 for_each_cpu(j, cpu_map) { 6121 struct sched_domain *sd; 6122 6123 if (sdd->sd) { 6124 sd = *per_cpu_ptr(sdd->sd, j); 6125 if (sd && (sd->flags & SD_OVERLAP)) 6126 free_sched_groups(sd->groups, 0); 6127 kfree(*per_cpu_ptr(sdd->sd, j)); 6128 } 6129 6130 if (sdd->sg) 6131 kfree(*per_cpu_ptr(sdd->sg, j)); 6132 if (sdd->sgp) 6133 kfree(*per_cpu_ptr(sdd->sgp, j)); 6134 } 6135 free_percpu(sdd->sd); 6136 sdd->sd = NULL; 6137 free_percpu(sdd->sg); 6138 sdd->sg = NULL; 6139 free_percpu(sdd->sgp); 6140 sdd->sgp = NULL; 6141 } 6142} 6143 6144struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl, 6145 const struct cpumask *cpu_map, struct sched_domain_attr *attr, 6146 struct sched_domain *child, int cpu) 6147{ 6148 struct sched_domain *sd = tl->init(tl, cpu); 6149 if (!sd) 6150 return child; 6151 6152 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu)); 6153 if (child) { 6154 sd->level = child->level + 1; 6155 sched_domain_level_max = max(sched_domain_level_max, sd->level); 6156 child->parent = sd; 6157 sd->child = child; 6158 } 6159 set_domain_attribute(sd, attr); 6160 6161 return sd; 6162} 6163 6164/* 6165 * Build sched domains for a given set of cpus and attach the sched domains 6166 * to the individual cpus 6167 */ 6168static int build_sched_domains(const struct cpumask *cpu_map, 6169 struct sched_domain_attr *attr) 6170{ 6171 enum s_alloc alloc_state; 6172 struct sched_domain *sd; 6173 struct s_data d; 6174 int i, ret = -ENOMEM; 6175 6176 alloc_state = __visit_domain_allocation_hell(&d, cpu_map); 6177 if (alloc_state != sa_rootdomain) 6178 goto error; 6179 6180 /* Set up domains for cpus specified by the cpu_map. */ 6181 for_each_cpu(i, cpu_map) { 6182 struct sched_domain_topology_level *tl; 6183 6184 sd = NULL; 6185 for_each_sd_topology(tl) { 6186 sd = build_sched_domain(tl, cpu_map, attr, sd, i); 6187 if (tl == sched_domain_topology) 6188 *per_cpu_ptr(d.sd, i) = sd; 6189 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP)) 6190 sd->flags |= SD_OVERLAP; 6191 if (cpumask_equal(cpu_map, sched_domain_span(sd))) 6192 break; 6193 } 6194 } 6195 6196 /* Build the groups for the domains */ 6197 for_each_cpu(i, cpu_map) { 6198 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) { 6199 sd->span_weight = cpumask_weight(sched_domain_span(sd)); 6200 if (sd->flags & SD_OVERLAP) { 6201 if (build_overlap_sched_groups(sd, i)) 6202 goto error; 6203 } else { 6204 if (build_sched_groups(sd, i)) 6205 goto error; 6206 } 6207 } 6208 } 6209 6210 /* Calculate CPU power for physical packages and nodes */ 6211 for (i = nr_cpumask_bits-1; i >= 0; i--) { 6212 if (!cpumask_test_cpu(i, cpu_map)) 6213 continue; 6214 6215 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) { 6216 claim_allocations(i, sd); 6217 init_sched_groups_power(i, sd); 6218 } 6219 } 6220 6221 /* Attach the domains */ 6222 rcu_read_lock(); 6223 for_each_cpu(i, cpu_map) { 6224 sd = *per_cpu_ptr(d.sd, i); 6225 cpu_attach_domain(sd, d.rd, i); 6226 } 6227 rcu_read_unlock(); 6228 6229 ret = 0; 6230error: 6231 __free_domain_allocs(&d, alloc_state, cpu_map); 6232 return ret; 6233} 6234 6235static cpumask_var_t *doms_cur; /* current sched domains */ 6236static int ndoms_cur; /* number of sched domains in 'doms_cur' */ 6237static struct sched_domain_attr *dattr_cur; 6238 /* attribues of custom domains in 'doms_cur' */ 6239 6240/* 6241 * Special case: If a kmalloc of a doms_cur partition (array of 6242 * cpumask) fails, then fallback to a single sched domain, 6243 * as determined by the single cpumask fallback_doms. 6244 */ 6245static cpumask_var_t fallback_doms; 6246 6247/* 6248 * arch_update_cpu_topology lets virtualized architectures update the 6249 * cpu core maps. It is supposed to return 1 if the topology changed 6250 * or 0 if it stayed the same. 6251 */ 6252int __attribute__((weak)) arch_update_cpu_topology(void) 6253{ 6254 return 0; 6255} 6256 6257cpumask_var_t *alloc_sched_domains(unsigned int ndoms) 6258{ 6259 int i; 6260 cpumask_var_t *doms; 6261 6262 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL); 6263 if (!doms) 6264 return NULL; 6265 for (i = 0; i < ndoms; i++) { 6266 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) { 6267 free_sched_domains(doms, i); 6268 return NULL; 6269 } 6270 } 6271 return doms; 6272} 6273 6274void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms) 6275{ 6276 unsigned int i; 6277 for (i = 0; i < ndoms; i++) 6278 free_cpumask_var(doms[i]); 6279 kfree(doms); 6280} 6281 6282/* 6283 * Set up scheduler domains and groups. Callers must hold the hotplug lock. 6284 * For now this just excludes isolated cpus, but could be used to 6285 * exclude other special cases in the future. 6286 */ 6287static int init_sched_domains(const struct cpumask *cpu_map) 6288{ 6289 int err; 6290 6291 arch_update_cpu_topology(); 6292 ndoms_cur = 1; 6293 doms_cur = alloc_sched_domains(ndoms_cur); 6294 if (!doms_cur) 6295 doms_cur = &fallback_doms; 6296 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map); 6297 err = build_sched_domains(doms_cur[0], NULL); 6298 register_sched_domain_sysctl(); 6299 6300 return err; 6301} 6302 6303/* 6304 * Detach sched domains from a group of cpus specified in cpu_map 6305 * These cpus will now be attached to the NULL domain 6306 */ 6307static void detach_destroy_domains(const struct cpumask *cpu_map) 6308{ 6309 int i; 6310 6311 rcu_read_lock(); 6312 for_each_cpu(i, cpu_map) 6313 cpu_attach_domain(NULL, &def_root_domain, i); 6314 rcu_read_unlock(); 6315} 6316 6317/* handle null as "default" */ 6318static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur, 6319 struct sched_domain_attr *new, int idx_new) 6320{ 6321 struct sched_domain_attr tmp; 6322 6323 /* fast path */ 6324 if (!new && !cur) 6325 return 1; 6326 6327 tmp = SD_ATTR_INIT; 6328 return !memcmp(cur ? (cur + idx_cur) : &tmp, 6329 new ? (new + idx_new) : &tmp, 6330 sizeof(struct sched_domain_attr)); 6331} 6332 6333/* 6334 * Partition sched domains as specified by the 'ndoms_new' 6335 * cpumasks in the array doms_new[] of cpumasks. This compares 6336 * doms_new[] to the current sched domain partitioning, doms_cur[]. 6337 * It destroys each deleted domain and builds each new domain. 6338 * 6339 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'. 6340 * The masks don't intersect (don't overlap.) We should setup one 6341 * sched domain for each mask. CPUs not in any of the cpumasks will 6342 * not be load balanced. If the same cpumask appears both in the 6343 * current 'doms_cur' domains and in the new 'doms_new', we can leave 6344 * it as it is. 6345 * 6346 * The passed in 'doms_new' should be allocated using 6347 * alloc_sched_domains. This routine takes ownership of it and will 6348 * free_sched_domains it when done with it. If the caller failed the 6349 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1, 6350 * and partition_sched_domains() will fallback to the single partition 6351 * 'fallback_doms', it also forces the domains to be rebuilt. 6352 * 6353 * If doms_new == NULL it will be replaced with cpu_online_mask. 6354 * ndoms_new == 0 is a special case for destroying existing domains, 6355 * and it will not create the default domain. 6356 * 6357 * Call with hotplug lock held 6358 */ 6359void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[], 6360 struct sched_domain_attr *dattr_new) 6361{ 6362 int i, j, n; 6363 int new_topology; 6364 6365 mutex_lock(&sched_domains_mutex); 6366 6367 /* always unregister in case we don't destroy any domains */ 6368 unregister_sched_domain_sysctl(); 6369 6370 /* Let architecture update cpu core mappings. */ 6371 new_topology = arch_update_cpu_topology(); 6372 6373 n = doms_new ? ndoms_new : 0; 6374 6375 /* Destroy deleted domains */ 6376 for (i = 0; i < ndoms_cur; i++) { 6377 for (j = 0; j < n && !new_topology; j++) { 6378 if (cpumask_equal(doms_cur[i], doms_new[j]) 6379 && dattrs_equal(dattr_cur, i, dattr_new, j)) 6380 goto match1; 6381 } 6382 /* no match - a current sched domain not in new doms_new[] */ 6383 detach_destroy_domains(doms_cur[i]); 6384match1: 6385 ; 6386 } 6387 6388 n = ndoms_cur; 6389 if (doms_new == NULL) { 6390 n = 0; 6391 doms_new = &fallback_doms; 6392 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map); 6393 WARN_ON_ONCE(dattr_new); 6394 } 6395 6396 /* Build new domains */ 6397 for (i = 0; i < ndoms_new; i++) { 6398 for (j = 0; j < n && !new_topology; j++) { 6399 if (cpumask_equal(doms_new[i], doms_cur[j]) 6400 && dattrs_equal(dattr_new, i, dattr_cur, j)) 6401 goto match2; 6402 } 6403 /* no match - add a new doms_new */ 6404 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL); 6405match2: 6406 ; 6407 } 6408 6409 /* Remember the new sched domains */ 6410 if (doms_cur != &fallback_doms) 6411 free_sched_domains(doms_cur, ndoms_cur); 6412 kfree(dattr_cur); /* kfree(NULL) is safe */ 6413 doms_cur = doms_new; 6414 dattr_cur = dattr_new; 6415 ndoms_cur = ndoms_new; 6416 6417 register_sched_domain_sysctl(); 6418 6419 mutex_unlock(&sched_domains_mutex); 6420} 6421 6422static int num_cpus_frozen; /* used to mark begin/end of suspend/resume */ 6423 6424/* 6425 * Update cpusets according to cpu_active mask. If cpusets are 6426 * disabled, cpuset_update_active_cpus() becomes a simple wrapper 6427 * around partition_sched_domains(). 6428 * 6429 * If we come here as part of a suspend/resume, don't touch cpusets because we 6430 * want to restore it back to its original state upon resume anyway. 6431 */ 6432static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action, 6433 void *hcpu) 6434{ 6435 switch (action) { 6436 case CPU_ONLINE_FROZEN: 6437 case CPU_DOWN_FAILED_FROZEN: 6438 6439 /* 6440 * num_cpus_frozen tracks how many CPUs are involved in suspend 6441 * resume sequence. As long as this is not the last online 6442 * operation in the resume sequence, just build a single sched 6443 * domain, ignoring cpusets. 6444 */ 6445 num_cpus_frozen--; 6446 if (likely(num_cpus_frozen)) { 6447 partition_sched_domains(1, NULL, NULL); 6448 break; 6449 } 6450 6451 /* 6452 * This is the last CPU online operation. So fall through and 6453 * restore the original sched domains by considering the 6454 * cpuset configurations. 6455 */ 6456 6457 case CPU_ONLINE: 6458 case CPU_DOWN_FAILED: 6459 cpuset_update_active_cpus(true); 6460 break; 6461 default: 6462 return NOTIFY_DONE; 6463 } 6464 return NOTIFY_OK; 6465} 6466 6467static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action, 6468 void *hcpu) 6469{ 6470 switch (action) { 6471 case CPU_DOWN_PREPARE: 6472 cpuset_update_active_cpus(false); 6473 break; 6474 case CPU_DOWN_PREPARE_FROZEN: 6475 num_cpus_frozen++; 6476 partition_sched_domains(1, NULL, NULL); 6477 break; 6478 default: 6479 return NOTIFY_DONE; 6480 } 6481 return NOTIFY_OK; 6482} 6483 6484void __init sched_init_smp(void) 6485{ 6486 cpumask_var_t non_isolated_cpus; 6487 6488 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL); 6489 alloc_cpumask_var(&fallback_doms, GFP_KERNEL); 6490 6491 sched_init_numa(); 6492 6493 get_online_cpus(); 6494 mutex_lock(&sched_domains_mutex); 6495 init_sched_domains(cpu_active_mask); 6496 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map); 6497 if (cpumask_empty(non_isolated_cpus)) 6498 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus); 6499 mutex_unlock(&sched_domains_mutex); 6500 put_online_cpus(); 6501 6502 hotcpu_notifier(sched_domains_numa_masks_update, CPU_PRI_SCHED_ACTIVE); 6503 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE); 6504 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE); 6505 6506 init_hrtick(); 6507 6508 /* Move init over to a non-isolated CPU */ 6509 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0) 6510 BUG(); 6511 sched_init_granularity(); 6512 free_cpumask_var(non_isolated_cpus); 6513 6514 init_sched_rt_class(); 6515} 6516#else 6517void __init sched_init_smp(void) 6518{ 6519 sched_init_granularity(); 6520} 6521#endif /* CONFIG_SMP */ 6522 6523const_debug unsigned int sysctl_timer_migration = 1; 6524 6525int in_sched_functions(unsigned long addr) 6526{ 6527 return in_lock_functions(addr) || 6528 (addr >= (unsigned long)__sched_text_start 6529 && addr < (unsigned long)__sched_text_end); 6530} 6531 6532#ifdef CONFIG_CGROUP_SCHED 6533/* 6534 * Default task group. 6535 * Every task in system belongs to this group at bootup. 6536 */ 6537struct task_group root_task_group; 6538LIST_HEAD(task_groups); 6539#endif 6540 6541DECLARE_PER_CPU(cpumask_var_t, load_balance_mask); 6542 6543void __init sched_init(void) 6544{ 6545 int i, j; 6546 unsigned long alloc_size = 0, ptr; 6547 6548#ifdef CONFIG_FAIR_GROUP_SCHED 6549 alloc_size += 2 * nr_cpu_ids * sizeof(void **); 6550#endif 6551#ifdef CONFIG_RT_GROUP_SCHED 6552 alloc_size += 2 * nr_cpu_ids * sizeof(void **); 6553#endif 6554#ifdef CONFIG_CPUMASK_OFFSTACK 6555 alloc_size += num_possible_cpus() * cpumask_size(); 6556#endif 6557 if (alloc_size) { 6558 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT); 6559 6560#ifdef CONFIG_FAIR_GROUP_SCHED 6561 root_task_group.se = (struct sched_entity **)ptr; 6562 ptr += nr_cpu_ids * sizeof(void **); 6563 6564 root_task_group.cfs_rq = (struct cfs_rq **)ptr; 6565 ptr += nr_cpu_ids * sizeof(void **); 6566 6567#endif /* CONFIG_FAIR_GROUP_SCHED */ 6568#ifdef CONFIG_RT_GROUP_SCHED 6569 root_task_group.rt_se = (struct sched_rt_entity **)ptr; 6570 ptr += nr_cpu_ids * sizeof(void **); 6571 6572 root_task_group.rt_rq = (struct rt_rq **)ptr; 6573 ptr += nr_cpu_ids * sizeof(void **); 6574 6575#endif /* CONFIG_RT_GROUP_SCHED */ 6576#ifdef CONFIG_CPUMASK_OFFSTACK 6577 for_each_possible_cpu(i) { 6578 per_cpu(load_balance_mask, i) = (void *)ptr; 6579 ptr += cpumask_size(); 6580 } 6581#endif /* CONFIG_CPUMASK_OFFSTACK */ 6582 } 6583 6584#ifdef CONFIG_SMP 6585 init_defrootdomain(); 6586#endif 6587 6588 init_rt_bandwidth(&def_rt_bandwidth, 6589 global_rt_period(), global_rt_runtime()); 6590 6591#ifdef CONFIG_RT_GROUP_SCHED 6592 init_rt_bandwidth(&root_task_group.rt_bandwidth, 6593 global_rt_period(), global_rt_runtime()); 6594#endif /* CONFIG_RT_GROUP_SCHED */ 6595 6596#ifdef CONFIG_CGROUP_SCHED 6597 list_add(&root_task_group.list, &task_groups); 6598 INIT_LIST_HEAD(&root_task_group.children); 6599 INIT_LIST_HEAD(&root_task_group.siblings); 6600 autogroup_init(&init_task); 6601 6602#endif /* CONFIG_CGROUP_SCHED */ 6603 6604 for_each_possible_cpu(i) { 6605 struct rq *rq; 6606 6607 rq = cpu_rq(i); 6608 raw_spin_lock_init(&rq->lock); 6609 rq->nr_running = 0; 6610 rq->calc_load_active = 0; 6611 rq->calc_load_update = jiffies + LOAD_FREQ; 6612 init_cfs_rq(&rq->cfs); 6613 init_rt_rq(&rq->rt, rq); 6614#ifdef CONFIG_FAIR_GROUP_SCHED 6615 root_task_group.shares = ROOT_TASK_GROUP_LOAD; 6616 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list); 6617 /* 6618 * How much cpu bandwidth does root_task_group get? 6619 * 6620 * In case of task-groups formed thr' the cgroup filesystem, it 6621 * gets 100% of the cpu resources in the system. This overall 6622 * system cpu resource is divided among the tasks of 6623 * root_task_group and its child task-groups in a fair manner, 6624 * based on each entity's (task or task-group's) weight 6625 * (se->load.weight). 6626 * 6627 * In other words, if root_task_group has 10 tasks of weight 6628 * 1024) and two child groups A0 and A1 (of weight 1024 each), 6629 * then A0's share of the cpu resource is: 6630 * 6631 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33% 6632 * 6633 * We achieve this by letting root_task_group's tasks sit 6634 * directly in rq->cfs (i.e root_task_group->se[] = NULL). 6635 */ 6636 init_cfs_bandwidth(&root_task_group.cfs_bandwidth); 6637 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL); 6638#endif /* CONFIG_FAIR_GROUP_SCHED */ 6639 6640 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime; 6641#ifdef CONFIG_RT_GROUP_SCHED 6642 INIT_LIST_HEAD(&rq->leaf_rt_rq_list); 6643 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL); 6644#endif 6645 6646 for (j = 0; j < CPU_LOAD_IDX_MAX; j++) 6647 rq->cpu_load[j] = 0; 6648 6649 rq->last_load_update_tick = jiffies; 6650 6651#ifdef CONFIG_SMP 6652 rq->sd = NULL; 6653 rq->rd = NULL; 6654 rq->cpu_power = SCHED_POWER_SCALE; 6655 rq->post_schedule = 0; 6656 rq->active_balance = 0; 6657 rq->next_balance = jiffies; 6658 rq->push_cpu = 0; 6659 rq->cpu = i; 6660 rq->online = 0; 6661 rq->idle_stamp = 0; 6662 rq->avg_idle = 2*sysctl_sched_migration_cost; 6663 rq->max_idle_balance_cost = sysctl_sched_migration_cost; 6664 6665 INIT_LIST_HEAD(&rq->cfs_tasks); 6666 6667 rq_attach_root(rq, &def_root_domain); 6668#ifdef CONFIG_NO_HZ_COMMON 6669 rq->nohz_flags = 0; 6670#endif 6671#ifdef CONFIG_NO_HZ_FULL 6672 rq->last_sched_tick = 0; 6673#endif 6674#endif 6675 init_rq_hrtick(rq); 6676 atomic_set(&rq->nr_iowait, 0); 6677 } 6678 6679 set_load_weight(&init_task); 6680 6681#ifdef CONFIG_PREEMPT_NOTIFIERS 6682 INIT_HLIST_HEAD(&init_task.preempt_notifiers); 6683#endif 6684 6685#ifdef CONFIG_RT_MUTEXES 6686 plist_head_init(&init_task.pi_waiters); 6687#endif 6688 6689 /* 6690 * The boot idle thread does lazy MMU switching as well: 6691 */ 6692 atomic_inc(&init_mm.mm_count); 6693 enter_lazy_tlb(&init_mm, current); 6694 6695 /* 6696 * Make us the idle thread. Technically, schedule() should not be 6697 * called from this thread, however somewhere below it might be, 6698 * but because we are the idle thread, we just pick up running again 6699 * when this runqueue becomes "idle". 6700 */ 6701 init_idle(current, smp_processor_id()); 6702 6703 calc_load_update = jiffies + LOAD_FREQ; 6704 6705 /* 6706 * During early bootup we pretend to be a normal task: 6707 */ 6708 current->sched_class = &fair_sched_class; 6709 6710#ifdef CONFIG_SMP 6711 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT); 6712 /* May be allocated at isolcpus cmdline parse time */ 6713 if (cpu_isolated_map == NULL) 6714 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT); 6715 idle_thread_set_boot_cpu(); 6716#endif 6717 init_sched_fair_class(); 6718 6719 scheduler_running = 1; 6720} 6721 6722#ifdef CONFIG_DEBUG_ATOMIC_SLEEP 6723static inline int preempt_count_equals(int preempt_offset) 6724{ 6725 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth(); 6726 6727 return (nested == preempt_offset); 6728} 6729 6730void __might_sleep(const char *file, int line, int preempt_offset) 6731{ 6732 static unsigned long prev_jiffy; /* ratelimiting */ 6733 6734 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */ 6735 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) || 6736 system_state != SYSTEM_RUNNING || oops_in_progress) 6737 return; 6738 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy) 6739 return; 6740 prev_jiffy = jiffies; 6741 6742 printk(KERN_ERR 6743 "BUG: sleeping function called from invalid context at %s:%d\n", 6744 file, line); 6745 printk(KERN_ERR 6746 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n", 6747 in_atomic(), irqs_disabled(), 6748 current->pid, current->comm); 6749 6750 debug_show_held_locks(current); 6751 if (irqs_disabled()) 6752 print_irqtrace_events(current); 6753 dump_stack(); 6754} 6755EXPORT_SYMBOL(__might_sleep); 6756#endif 6757 6758#ifdef CONFIG_MAGIC_SYSRQ 6759static void normalize_task(struct rq *rq, struct task_struct *p) 6760{ 6761 const struct sched_class *prev_class = p->sched_class; 6762 int old_prio = p->prio; 6763 int on_rq; 6764 6765 on_rq = p->on_rq; 6766 if (on_rq) 6767 dequeue_task(rq, p, 0); 6768 __setscheduler(rq, p, SCHED_NORMAL, 0); 6769 if (on_rq) { 6770 enqueue_task(rq, p, 0); 6771 resched_task(rq->curr); 6772 } 6773 6774 check_class_changed(rq, p, prev_class, old_prio); 6775} 6776 6777void normalize_rt_tasks(void) 6778{ 6779 struct task_struct *g, *p; 6780 unsigned long flags; 6781 struct rq *rq; 6782 6783 read_lock_irqsave(&tasklist_lock, flags); 6784 do_each_thread(g, p) { 6785 /* 6786 * Only normalize user tasks: 6787 */ 6788 if (!p->mm) 6789 continue; 6790 6791 p->se.exec_start = 0; 6792#ifdef CONFIG_SCHEDSTATS 6793 p->se.statistics.wait_start = 0; 6794 p->se.statistics.sleep_start = 0; 6795 p->se.statistics.block_start = 0; 6796#endif 6797 6798 if (!rt_task(p)) { 6799 /* 6800 * Renice negative nice level userspace 6801 * tasks back to 0: 6802 */ 6803 if (TASK_NICE(p) < 0 && p->mm) 6804 set_user_nice(p, 0); 6805 continue; 6806 } 6807 6808 raw_spin_lock(&p->pi_lock); 6809 rq = __task_rq_lock(p); 6810 6811 normalize_task(rq, p); 6812 6813 __task_rq_unlock(rq); 6814 raw_spin_unlock(&p->pi_lock); 6815 } while_each_thread(g, p); 6816 6817 read_unlock_irqrestore(&tasklist_lock, flags); 6818} 6819 6820#endif /* CONFIG_MAGIC_SYSRQ */ 6821 6822#if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) 6823/* 6824 * These functions are only useful for the IA64 MCA handling, or kdb. 6825 * 6826 * They can only be called when the whole system has been 6827 * stopped - every CPU needs to be quiescent, and no scheduling 6828 * activity can take place. Using them for anything else would 6829 * be a serious bug, and as a result, they aren't even visible 6830 * under any other configuration. 6831 */ 6832 6833/** 6834 * curr_task - return the current task for a given cpu. 6835 * @cpu: the processor in question. 6836 * 6837 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED! 6838 * 6839 * Return: The current task for @cpu. 6840 */ 6841struct task_struct *curr_task(int cpu) 6842{ 6843 return cpu_curr(cpu); 6844} 6845 6846#endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */ 6847 6848#ifdef CONFIG_IA64 6849/** 6850 * set_curr_task - set the current task for a given cpu. 6851 * @cpu: the processor in question. 6852 * @p: the task pointer to set. 6853 * 6854 * Description: This function must only be used when non-maskable interrupts 6855 * are serviced on a separate stack. It allows the architecture to switch the 6856 * notion of the current task on a cpu in a non-blocking manner. This function 6857 * must be called with all CPU's synchronized, and interrupts disabled, the 6858 * and caller must save the original value of the current task (see 6859 * curr_task() above) and restore that value before reenabling interrupts and 6860 * re-starting the system. 6861 * 6862 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED! 6863 */ 6864void set_curr_task(int cpu, struct task_struct *p) 6865{ 6866 cpu_curr(cpu) = p; 6867} 6868 6869#endif 6870 6871#ifdef CONFIG_CGROUP_SCHED 6872/* task_group_lock serializes the addition/removal of task groups */ 6873static DEFINE_SPINLOCK(task_group_lock); 6874 6875static void free_sched_group(struct task_group *tg) 6876{ 6877 free_fair_sched_group(tg); 6878 free_rt_sched_group(tg); 6879 autogroup_free(tg); 6880 kfree(tg); 6881} 6882 6883/* allocate runqueue etc for a new task group */ 6884struct task_group *sched_create_group(struct task_group *parent) 6885{ 6886 struct task_group *tg; 6887 6888 tg = kzalloc(sizeof(*tg), GFP_KERNEL); 6889 if (!tg) 6890 return ERR_PTR(-ENOMEM); 6891 6892 if (!alloc_fair_sched_group(tg, parent)) 6893 goto err; 6894 6895 if (!alloc_rt_sched_group(tg, parent)) 6896 goto err; 6897 6898 return tg; 6899 6900err: 6901 free_sched_group(tg); 6902 return ERR_PTR(-ENOMEM); 6903} 6904 6905void sched_online_group(struct task_group *tg, struct task_group *parent) 6906{ 6907 unsigned long flags; 6908 6909 spin_lock_irqsave(&task_group_lock, flags); 6910 list_add_rcu(&tg->list, &task_groups); 6911 6912 WARN_ON(!parent); /* root should already exist */ 6913 6914 tg->parent = parent; 6915 INIT_LIST_HEAD(&tg->children); 6916 list_add_rcu(&tg->siblings, &parent->children); 6917 spin_unlock_irqrestore(&task_group_lock, flags); 6918} 6919 6920/* rcu callback to free various structures associated with a task group */ 6921static void free_sched_group_rcu(struct rcu_head *rhp) 6922{ 6923 /* now it should be safe to free those cfs_rqs */ 6924 free_sched_group(container_of(rhp, struct task_group, rcu)); 6925} 6926 6927/* Destroy runqueue etc associated with a task group */ 6928void sched_destroy_group(struct task_group *tg) 6929{ 6930 /* wait for possible concurrent references to cfs_rqs complete */ 6931 call_rcu(&tg->rcu, free_sched_group_rcu); 6932} 6933 6934void sched_offline_group(struct task_group *tg) 6935{ 6936 unsigned long flags; 6937 int i; 6938 6939 /* end participation in shares distribution */ 6940 for_each_possible_cpu(i) 6941 unregister_fair_sched_group(tg, i); 6942 6943 spin_lock_irqsave(&task_group_lock, flags); 6944 list_del_rcu(&tg->list); 6945 list_del_rcu(&tg->siblings); 6946 spin_unlock_irqrestore(&task_group_lock, flags); 6947} 6948 6949/* change task's runqueue when it moves between groups. 6950 * The caller of this function should have put the task in its new group 6951 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to 6952 * reflect its new group. 6953 */ 6954void sched_move_task(struct task_struct *tsk) 6955{ 6956 struct task_group *tg; 6957 int on_rq, running; 6958 unsigned long flags; 6959 struct rq *rq; 6960 6961 rq = task_rq_lock(tsk, &flags); 6962 6963 running = task_current(rq, tsk); 6964 on_rq = tsk->on_rq; 6965 6966 if (on_rq) 6967 dequeue_task(rq, tsk, 0); 6968 if (unlikely(running)) 6969 tsk->sched_class->put_prev_task(rq, tsk); 6970 6971 tg = container_of(task_css_check(tsk, cpu_cgroup_subsys_id, 6972 lockdep_is_held(&tsk->sighand->siglock)), 6973 struct task_group, css); 6974 tg = autogroup_task_group(tsk, tg); 6975 tsk->sched_task_group = tg; 6976 6977#ifdef CONFIG_FAIR_GROUP_SCHED 6978 if (tsk->sched_class->task_move_group) 6979 tsk->sched_class->task_move_group(tsk, on_rq); 6980 else 6981#endif 6982 set_task_rq(tsk, task_cpu(tsk)); 6983 6984 if (unlikely(running)) 6985 tsk->sched_class->set_curr_task(rq); 6986 if (on_rq) 6987 enqueue_task(rq, tsk, 0); 6988 6989 task_rq_unlock(rq, tsk, &flags); 6990} 6991#endif /* CONFIG_CGROUP_SCHED */ 6992 6993#if defined(CONFIG_RT_GROUP_SCHED) || defined(CONFIG_CFS_BANDWIDTH) 6994static unsigned long to_ratio(u64 period, u64 runtime) 6995{ 6996 if (runtime == RUNTIME_INF) 6997 return 1ULL << 20; 6998 6999 return div64_u64(runtime << 20, period); 7000} 7001#endif 7002 7003#ifdef CONFIG_RT_GROUP_SCHED 7004/* 7005 * Ensure that the real time constraints are schedulable. 7006 */ 7007static DEFINE_MUTEX(rt_constraints_mutex); 7008 7009/* Must be called with tasklist_lock held */ 7010static inline int tg_has_rt_tasks(struct task_group *tg) 7011{ 7012 struct task_struct *g, *p; 7013 7014 do_each_thread(g, p) { 7015 if (rt_task(p) && task_rq(p)->rt.tg == tg) 7016 return 1; 7017 } while_each_thread(g, p); 7018 7019 return 0; 7020} 7021 7022struct rt_schedulable_data { 7023 struct task_group *tg; 7024 u64 rt_period; 7025 u64 rt_runtime; 7026}; 7027 7028static int tg_rt_schedulable(struct task_group *tg, void *data) 7029{ 7030 struct rt_schedulable_data *d = data; 7031 struct task_group *child; 7032 unsigned long total, sum = 0; 7033 u64 period, runtime; 7034 7035 period = ktime_to_ns(tg->rt_bandwidth.rt_period); 7036 runtime = tg->rt_bandwidth.rt_runtime; 7037 7038 if (tg == d->tg) { 7039 period = d->rt_period; 7040 runtime = d->rt_runtime; 7041 } 7042 7043 /* 7044 * Cannot have more runtime than the period. 7045 */ 7046 if (runtime > period && runtime != RUNTIME_INF) 7047 return -EINVAL; 7048 7049 /* 7050 * Ensure we don't starve existing RT tasks. 7051 */ 7052 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg)) 7053 return -EBUSY; 7054 7055 total = to_ratio(period, runtime); 7056 7057 /* 7058 * Nobody can have more than the global setting allows. 7059 */ 7060 if (total > to_ratio(global_rt_period(), global_rt_runtime())) 7061 return -EINVAL; 7062 7063 /* 7064 * The sum of our children's runtime should not exceed our own. 7065 */ 7066 list_for_each_entry_rcu(child, &tg->children, siblings) { 7067 period = ktime_to_ns(child->rt_bandwidth.rt_period); 7068 runtime = child->rt_bandwidth.rt_runtime; 7069 7070 if (child == d->tg) { 7071 period = d->rt_period; 7072 runtime = d->rt_runtime; 7073 } 7074 7075 sum += to_ratio(period, runtime); 7076 } 7077 7078 if (sum > total) 7079 return -EINVAL; 7080 7081 return 0; 7082} 7083 7084static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime) 7085{ 7086 int ret; 7087 7088 struct rt_schedulable_data data = { 7089 .tg = tg, 7090 .rt_period = period, 7091 .rt_runtime = runtime, 7092 }; 7093 7094 rcu_read_lock(); 7095 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data); 7096 rcu_read_unlock(); 7097 7098 return ret; 7099} 7100 7101static int tg_set_rt_bandwidth(struct task_group *tg, 7102 u64 rt_period, u64 rt_runtime) 7103{ 7104 int i, err = 0; 7105 7106 mutex_lock(&rt_constraints_mutex); 7107 read_lock(&tasklist_lock); 7108 err = __rt_schedulable(tg, rt_period, rt_runtime); 7109 if (err) 7110 goto unlock; 7111 7112 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock); 7113 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period); 7114 tg->rt_bandwidth.rt_runtime = rt_runtime; 7115 7116 for_each_possible_cpu(i) { 7117 struct rt_rq *rt_rq = tg->rt_rq[i]; 7118 7119 raw_spin_lock(&rt_rq->rt_runtime_lock); 7120 rt_rq->rt_runtime = rt_runtime; 7121 raw_spin_unlock(&rt_rq->rt_runtime_lock); 7122 } 7123 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock); 7124unlock: 7125 read_unlock(&tasklist_lock); 7126 mutex_unlock(&rt_constraints_mutex); 7127 7128 return err; 7129} 7130 7131static int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us) 7132{ 7133 u64 rt_runtime, rt_period; 7134 7135 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period); 7136 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC; 7137 if (rt_runtime_us < 0) 7138 rt_runtime = RUNTIME_INF; 7139 7140 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime); 7141} 7142 7143static long sched_group_rt_runtime(struct task_group *tg) 7144{ 7145 u64 rt_runtime_us; 7146 7147 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF) 7148 return -1; 7149 7150 rt_runtime_us = tg->rt_bandwidth.rt_runtime; 7151 do_div(rt_runtime_us, NSEC_PER_USEC); 7152 return rt_runtime_us; 7153} 7154 7155static int sched_group_set_rt_period(struct task_group *tg, long rt_period_us) 7156{ 7157 u64 rt_runtime, rt_period; 7158 7159 rt_period = (u64)rt_period_us * NSEC_PER_USEC; 7160 rt_runtime = tg->rt_bandwidth.rt_runtime; 7161 7162 if (rt_period == 0) 7163 return -EINVAL; 7164 7165 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime); 7166} 7167 7168static long sched_group_rt_period(struct task_group *tg) 7169{ 7170 u64 rt_period_us; 7171 7172 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period); 7173 do_div(rt_period_us, NSEC_PER_USEC); 7174 return rt_period_us; 7175} 7176 7177static int sched_rt_global_constraints(void) 7178{ 7179 u64 runtime, period; 7180 int ret = 0; 7181 7182 if (sysctl_sched_rt_period <= 0) 7183 return -EINVAL; 7184 7185 runtime = global_rt_runtime(); 7186 period = global_rt_period(); 7187 7188 /* 7189 * Sanity check on the sysctl variables. 7190 */ 7191 if (runtime > period && runtime != RUNTIME_INF) 7192 return -EINVAL; 7193 7194 mutex_lock(&rt_constraints_mutex); 7195 read_lock(&tasklist_lock); 7196 ret = __rt_schedulable(NULL, 0, 0); 7197 read_unlock(&tasklist_lock); 7198 mutex_unlock(&rt_constraints_mutex); 7199 7200 return ret; 7201} 7202 7203static int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk) 7204{ 7205 /* Don't accept realtime tasks when there is no way for them to run */ 7206 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0) 7207 return 0; 7208 7209 return 1; 7210} 7211 7212#else /* !CONFIG_RT_GROUP_SCHED */ 7213static int sched_rt_global_constraints(void) 7214{ 7215 unsigned long flags; 7216 int i; 7217 7218 if (sysctl_sched_rt_period <= 0) 7219 return -EINVAL; 7220 7221 /* 7222 * There's always some RT tasks in the root group 7223 * -- migration, kstopmachine etc.. 7224 */ 7225 if (sysctl_sched_rt_runtime == 0) 7226 return -EBUSY; 7227 7228 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags); 7229 for_each_possible_cpu(i) { 7230 struct rt_rq *rt_rq = &cpu_rq(i)->rt; 7231 7232 raw_spin_lock(&rt_rq->rt_runtime_lock); 7233 rt_rq->rt_runtime = global_rt_runtime(); 7234 raw_spin_unlock(&rt_rq->rt_runtime_lock); 7235 } 7236 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags); 7237 7238 return 0; 7239} 7240#endif /* CONFIG_RT_GROUP_SCHED */ 7241 7242int sched_rr_handler(struct ctl_table *table, int write, 7243 void __user *buffer, size_t *lenp, 7244 loff_t *ppos) 7245{ 7246 int ret; 7247 static DEFINE_MUTEX(mutex); 7248 7249 mutex_lock(&mutex); 7250 ret = proc_dointvec(table, write, buffer, lenp, ppos); 7251 /* make sure that internally we keep jiffies */ 7252 /* also, writing zero resets timeslice to default */ 7253 if (!ret && write) { 7254 sched_rr_timeslice = sched_rr_timeslice <= 0 ? 7255 RR_TIMESLICE : msecs_to_jiffies(sched_rr_timeslice); 7256 } 7257 mutex_unlock(&mutex); 7258 return ret; 7259} 7260 7261int sched_rt_handler(struct ctl_table *table, int write, 7262 void __user *buffer, size_t *lenp, 7263 loff_t *ppos) 7264{ 7265 int ret; 7266 int old_period, old_runtime; 7267 static DEFINE_MUTEX(mutex); 7268 7269 mutex_lock(&mutex); 7270 old_period = sysctl_sched_rt_period; 7271 old_runtime = sysctl_sched_rt_runtime; 7272 7273 ret = proc_dointvec(table, write, buffer, lenp, ppos); 7274 7275 if (!ret && write) { 7276 ret = sched_rt_global_constraints(); 7277 if (ret) { 7278 sysctl_sched_rt_period = old_period; 7279 sysctl_sched_rt_runtime = old_runtime; 7280 } else { 7281 def_rt_bandwidth.rt_runtime = global_rt_runtime(); 7282 def_rt_bandwidth.rt_period = 7283 ns_to_ktime(global_rt_period()); 7284 } 7285 } 7286 mutex_unlock(&mutex); 7287 7288 return ret; 7289} 7290 7291#ifdef CONFIG_CGROUP_SCHED 7292 7293static inline struct task_group *css_tg(struct cgroup_subsys_state *css) 7294{ 7295 return css ? container_of(css, struct task_group, css) : NULL; 7296} 7297 7298static struct cgroup_subsys_state * 7299cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css) 7300{ 7301 struct task_group *parent = css_tg(parent_css); 7302 struct task_group *tg; 7303 7304 if (!parent) { 7305 /* This is early initialization for the top cgroup */ 7306 return &root_task_group.css; 7307 } 7308 7309 tg = sched_create_group(parent); 7310 if (IS_ERR(tg)) 7311 return ERR_PTR(-ENOMEM); 7312 7313 return &tg->css; 7314} 7315 7316static int cpu_cgroup_css_online(struct cgroup_subsys_state *css) 7317{ 7318 struct task_group *tg = css_tg(css); 7319 struct task_group *parent = css_tg(css_parent(css)); 7320 7321 if (parent) 7322 sched_online_group(tg, parent); 7323 return 0; 7324} 7325 7326static void cpu_cgroup_css_free(struct cgroup_subsys_state *css) 7327{ 7328 struct task_group *tg = css_tg(css); 7329 7330 sched_destroy_group(tg); 7331} 7332 7333static void cpu_cgroup_css_offline(struct cgroup_subsys_state *css) 7334{ 7335 struct task_group *tg = css_tg(css); 7336 7337 sched_offline_group(tg); 7338} 7339 7340static int cpu_cgroup_can_attach(struct cgroup_subsys_state *css, 7341 struct cgroup_taskset *tset) 7342{ 7343 struct task_struct *task; 7344 7345 cgroup_taskset_for_each(task, css, tset) { 7346#ifdef CONFIG_RT_GROUP_SCHED 7347 if (!sched_rt_can_attach(css_tg(css), task)) 7348 return -EINVAL; 7349#else 7350 /* We don't support RT-tasks being in separate groups */ 7351 if (task->sched_class != &fair_sched_class) 7352 return -EINVAL; 7353#endif 7354 } 7355 return 0; 7356} 7357 7358static void cpu_cgroup_attach(struct cgroup_subsys_state *css, 7359 struct cgroup_taskset *tset) 7360{ 7361 struct task_struct *task; 7362 7363 cgroup_taskset_for_each(task, css, tset) 7364 sched_move_task(task); 7365} 7366 7367static void cpu_cgroup_exit(struct cgroup_subsys_state *css, 7368 struct cgroup_subsys_state *old_css, 7369 struct task_struct *task) 7370{ 7371 /* 7372 * cgroup_exit() is called in the copy_process() failure path. 7373 * Ignore this case since the task hasn't ran yet, this avoids 7374 * trying to poke a half freed task state from generic code. 7375 */ 7376 if (!(task->flags & PF_EXITING)) 7377 return; 7378 7379 sched_move_task(task); 7380} 7381 7382#ifdef CONFIG_FAIR_GROUP_SCHED 7383static int cpu_shares_write_u64(struct cgroup_subsys_state *css, 7384 struct cftype *cftype, u64 shareval) 7385{ 7386 return sched_group_set_shares(css_tg(css), scale_load(shareval)); 7387} 7388 7389static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css, 7390 struct cftype *cft) 7391{ 7392 struct task_group *tg = css_tg(css); 7393 7394 return (u64) scale_load_down(tg->shares); 7395} 7396 7397#ifdef CONFIG_CFS_BANDWIDTH 7398static DEFINE_MUTEX(cfs_constraints_mutex); 7399 7400const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */ 7401const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */ 7402 7403static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime); 7404 7405static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota) 7406{ 7407 int i, ret = 0, runtime_enabled, runtime_was_enabled; 7408 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth; 7409 7410 if (tg == &root_task_group) 7411 return -EINVAL; 7412 7413 /* 7414 * Ensure we have at some amount of bandwidth every period. This is 7415 * to prevent reaching a state of large arrears when throttled via 7416 * entity_tick() resulting in prolonged exit starvation. 7417 */ 7418 if (quota < min_cfs_quota_period || period < min_cfs_quota_period) 7419 return -EINVAL; 7420 7421 /* 7422 * Likewise, bound things on the otherside by preventing insane quota 7423 * periods. This also allows us to normalize in computing quota 7424 * feasibility. 7425 */ 7426 if (period > max_cfs_quota_period) 7427 return -EINVAL; 7428 7429 mutex_lock(&cfs_constraints_mutex); 7430 ret = __cfs_schedulable(tg, period, quota); 7431 if (ret) 7432 goto out_unlock; 7433 7434 runtime_enabled = quota != RUNTIME_INF; 7435 runtime_was_enabled = cfs_b->quota != RUNTIME_INF; 7436 account_cfs_bandwidth_used(runtime_enabled, runtime_was_enabled); 7437 raw_spin_lock_irq(&cfs_b->lock); 7438 cfs_b->period = ns_to_ktime(period); 7439 cfs_b->quota = quota; 7440 7441 __refill_cfs_bandwidth_runtime(cfs_b); 7442 /* restart the period timer (if active) to handle new period expiry */ 7443 if (runtime_enabled && cfs_b->timer_active) { 7444 /* force a reprogram */ 7445 cfs_b->timer_active = 0; 7446 __start_cfs_bandwidth(cfs_b); 7447 } 7448 raw_spin_unlock_irq(&cfs_b->lock); 7449 7450 for_each_possible_cpu(i) { 7451 struct cfs_rq *cfs_rq = tg->cfs_rq[i]; 7452 struct rq *rq = cfs_rq->rq; 7453 7454 raw_spin_lock_irq(&rq->lock); 7455 cfs_rq->runtime_enabled = runtime_enabled; 7456 cfs_rq->runtime_remaining = 0; 7457 7458 if (cfs_rq->throttled) 7459 unthrottle_cfs_rq(cfs_rq); 7460 raw_spin_unlock_irq(&rq->lock); 7461 } 7462out_unlock: 7463 mutex_unlock(&cfs_constraints_mutex); 7464 7465 return ret; 7466} 7467 7468int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us) 7469{ 7470 u64 quota, period; 7471 7472 period = ktime_to_ns(tg->cfs_bandwidth.period); 7473 if (cfs_quota_us < 0) 7474 quota = RUNTIME_INF; 7475 else 7476 quota = (u64)cfs_quota_us * NSEC_PER_USEC; 7477 7478 return tg_set_cfs_bandwidth(tg, period, quota); 7479} 7480 7481long tg_get_cfs_quota(struct task_group *tg) 7482{ 7483 u64 quota_us; 7484 7485 if (tg->cfs_bandwidth.quota == RUNTIME_INF) 7486 return -1; 7487 7488 quota_us = tg->cfs_bandwidth.quota; 7489 do_div(quota_us, NSEC_PER_USEC); 7490 7491 return quota_us; 7492} 7493 7494int tg_set_cfs_period(struct task_group *tg, long cfs_period_us) 7495{ 7496 u64 quota, period; 7497 7498 period = (u64)cfs_period_us * NSEC_PER_USEC; 7499 quota = tg->cfs_bandwidth.quota; 7500 7501 return tg_set_cfs_bandwidth(tg, period, quota); 7502} 7503 7504long tg_get_cfs_period(struct task_group *tg) 7505{ 7506 u64 cfs_period_us; 7507 7508 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period); 7509 do_div(cfs_period_us, NSEC_PER_USEC); 7510 7511 return cfs_period_us; 7512} 7513 7514static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css, 7515 struct cftype *cft) 7516{ 7517 return tg_get_cfs_quota(css_tg(css)); 7518} 7519 7520static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css, 7521 struct cftype *cftype, s64 cfs_quota_us) 7522{ 7523 return tg_set_cfs_quota(css_tg(css), cfs_quota_us); 7524} 7525 7526static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css, 7527 struct cftype *cft) 7528{ 7529 return tg_get_cfs_period(css_tg(css)); 7530} 7531 7532static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css, 7533 struct cftype *cftype, u64 cfs_period_us) 7534{ 7535 return tg_set_cfs_period(css_tg(css), cfs_period_us); 7536} 7537 7538struct cfs_schedulable_data { 7539 struct task_group *tg; 7540 u64 period, quota; 7541}; 7542 7543/* 7544 * normalize group quota/period to be quota/max_period 7545 * note: units are usecs 7546 */ 7547static u64 normalize_cfs_quota(struct task_group *tg, 7548 struct cfs_schedulable_data *d) 7549{ 7550 u64 quota, period; 7551 7552 if (tg == d->tg) { 7553 period = d->period; 7554 quota = d->quota; 7555 } else { 7556 period = tg_get_cfs_period(tg); 7557 quota = tg_get_cfs_quota(tg); 7558 } 7559 7560 /* note: these should typically be equivalent */ 7561 if (quota == RUNTIME_INF || quota == -1) 7562 return RUNTIME_INF; 7563 7564 return to_ratio(period, quota); 7565} 7566 7567static int tg_cfs_schedulable_down(struct task_group *tg, void *data) 7568{ 7569 struct cfs_schedulable_data *d = data; 7570 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth; 7571 s64 quota = 0, parent_quota = -1; 7572 7573 if (!tg->parent) { 7574 quota = RUNTIME_INF; 7575 } else { 7576 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth; 7577 7578 quota = normalize_cfs_quota(tg, d); 7579 parent_quota = parent_b->hierarchal_quota; 7580 7581 /* 7582 * ensure max(child_quota) <= parent_quota, inherit when no 7583 * limit is set 7584 */ 7585 if (quota == RUNTIME_INF) 7586 quota = parent_quota; 7587 else if (parent_quota != RUNTIME_INF && quota > parent_quota) 7588 return -EINVAL; 7589 } 7590 cfs_b->hierarchal_quota = quota; 7591 7592 return 0; 7593} 7594 7595static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota) 7596{ 7597 int ret; 7598 struct cfs_schedulable_data data = { 7599 .tg = tg, 7600 .period = period, 7601 .quota = quota, 7602 }; 7603 7604 if (quota != RUNTIME_INF) { 7605 do_div(data.period, NSEC_PER_USEC); 7606 do_div(data.quota, NSEC_PER_USEC); 7607 } 7608 7609 rcu_read_lock(); 7610 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data); 7611 rcu_read_unlock(); 7612 7613 return ret; 7614} 7615 7616static int cpu_stats_show(struct cgroup_subsys_state *css, struct cftype *cft, 7617 struct cgroup_map_cb *cb) 7618{ 7619 struct task_group *tg = css_tg(css); 7620 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth; 7621 7622 cb->fill(cb, "nr_periods", cfs_b->nr_periods); 7623 cb->fill(cb, "nr_throttled", cfs_b->nr_throttled); 7624 cb->fill(cb, "throttled_time", cfs_b->throttled_time); 7625 7626 return 0; 7627} 7628#endif /* CONFIG_CFS_BANDWIDTH */ 7629#endif /* CONFIG_FAIR_GROUP_SCHED */ 7630 7631#ifdef CONFIG_RT_GROUP_SCHED 7632static int cpu_rt_runtime_write(struct cgroup_subsys_state *css, 7633 struct cftype *cft, s64 val) 7634{ 7635 return sched_group_set_rt_runtime(css_tg(css), val); 7636} 7637 7638static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css, 7639 struct cftype *cft) 7640{ 7641 return sched_group_rt_runtime(css_tg(css)); 7642} 7643 7644static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css, 7645 struct cftype *cftype, u64 rt_period_us) 7646{ 7647 return sched_group_set_rt_period(css_tg(css), rt_period_us); 7648} 7649 7650static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css, 7651 struct cftype *cft) 7652{ 7653 return sched_group_rt_period(css_tg(css)); 7654} 7655#endif /* CONFIG_RT_GROUP_SCHED */ 7656 7657static struct cftype cpu_files[] = { 7658#ifdef CONFIG_FAIR_GROUP_SCHED 7659 { 7660 .name = "shares", 7661 .read_u64 = cpu_shares_read_u64, 7662 .write_u64 = cpu_shares_write_u64, 7663 }, 7664#endif 7665#ifdef CONFIG_CFS_BANDWIDTH 7666 { 7667 .name = "cfs_quota_us", 7668 .read_s64 = cpu_cfs_quota_read_s64, 7669 .write_s64 = cpu_cfs_quota_write_s64, 7670 }, 7671 { 7672 .name = "cfs_period_us", 7673 .read_u64 = cpu_cfs_period_read_u64, 7674 .write_u64 = cpu_cfs_period_write_u64, 7675 }, 7676 { 7677 .name = "stat", 7678 .read_map = cpu_stats_show, 7679 }, 7680#endif 7681#ifdef CONFIG_RT_GROUP_SCHED 7682 { 7683 .name = "rt_runtime_us", 7684 .read_s64 = cpu_rt_runtime_read, 7685 .write_s64 = cpu_rt_runtime_write, 7686 }, 7687 { 7688 .name = "rt_period_us", 7689 .read_u64 = cpu_rt_period_read_uint, 7690 .write_u64 = cpu_rt_period_write_uint, 7691 }, 7692#endif 7693 { } /* terminate */ 7694}; 7695 7696struct cgroup_subsys cpu_cgroup_subsys = { 7697 .name = "cpu", 7698 .css_alloc = cpu_cgroup_css_alloc, 7699 .css_free = cpu_cgroup_css_free, 7700 .css_online = cpu_cgroup_css_online, 7701 .css_offline = cpu_cgroup_css_offline, 7702 .can_attach = cpu_cgroup_can_attach, 7703 .attach = cpu_cgroup_attach, 7704 .exit = cpu_cgroup_exit, 7705 .subsys_id = cpu_cgroup_subsys_id, 7706 .base_cftypes = cpu_files, 7707 .early_init = 1, 7708}; 7709 7710#endif /* CONFIG_CGROUP_SCHED */ 7711 7712void dump_cpu_task(int cpu) 7713{ 7714 pr_info("Task dump for CPU %d:\n", cpu); 7715 sched_show_task(cpu_curr(cpu)); 7716} 7717