core.c revision cb04ff9ac424d0e689d9b612e9f73cb443ab4b7e
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 76#include <asm/switch_to.h> 77#include <asm/tlb.h> 78#include <asm/irq_regs.h> 79#include <asm/mutex.h> 80#ifdef CONFIG_PARAVIRT 81#include <asm/paravirt.h> 82#endif 83 84#include "sched.h" 85#include "../workqueue_sched.h" 86 87#define CREATE_TRACE_POINTS 88#include <trace/events/sched.h> 89 90void start_bandwidth_timer(struct hrtimer *period_timer, ktime_t period) 91{ 92 unsigned long delta; 93 ktime_t soft, hard, now; 94 95 for (;;) { 96 if (hrtimer_active(period_timer)) 97 break; 98 99 now = hrtimer_cb_get_time(period_timer); 100 hrtimer_forward(period_timer, now, period); 101 102 soft = hrtimer_get_softexpires(period_timer); 103 hard = hrtimer_get_expires(period_timer); 104 delta = ktime_to_ns(ktime_sub(hard, soft)); 105 __hrtimer_start_range_ns(period_timer, soft, delta, 106 HRTIMER_MODE_ABS_PINNED, 0); 107 } 108} 109 110DEFINE_MUTEX(sched_domains_mutex); 111DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues); 112 113static void update_rq_clock_task(struct rq *rq, s64 delta); 114 115void update_rq_clock(struct rq *rq) 116{ 117 s64 delta; 118 119 if (rq->skip_clock_update > 0) 120 return; 121 122 delta = sched_clock_cpu(cpu_of(rq)) - rq->clock; 123 rq->clock += delta; 124 update_rq_clock_task(rq, delta); 125} 126 127/* 128 * Debugging: various feature bits 129 */ 130 131#define SCHED_FEAT(name, enabled) \ 132 (1UL << __SCHED_FEAT_##name) * enabled | 133 134const_debug unsigned int sysctl_sched_features = 135#include "features.h" 136 0; 137 138#undef SCHED_FEAT 139 140#ifdef CONFIG_SCHED_DEBUG 141#define SCHED_FEAT(name, enabled) \ 142 #name , 143 144static __read_mostly char *sched_feat_names[] = { 145#include "features.h" 146 NULL 147}; 148 149#undef SCHED_FEAT 150 151static int sched_feat_show(struct seq_file *m, void *v) 152{ 153 int i; 154 155 for (i = 0; i < __SCHED_FEAT_NR; i++) { 156 if (!(sysctl_sched_features & (1UL << i))) 157 seq_puts(m, "NO_"); 158 seq_printf(m, "%s ", sched_feat_names[i]); 159 } 160 seq_puts(m, "\n"); 161 162 return 0; 163} 164 165#ifdef HAVE_JUMP_LABEL 166 167#define jump_label_key__true STATIC_KEY_INIT_TRUE 168#define jump_label_key__false STATIC_KEY_INIT_FALSE 169 170#define SCHED_FEAT(name, enabled) \ 171 jump_label_key__##enabled , 172 173struct static_key sched_feat_keys[__SCHED_FEAT_NR] = { 174#include "features.h" 175}; 176 177#undef SCHED_FEAT 178 179static void sched_feat_disable(int i) 180{ 181 if (static_key_enabled(&sched_feat_keys[i])) 182 static_key_slow_dec(&sched_feat_keys[i]); 183} 184 185static void sched_feat_enable(int i) 186{ 187 if (!static_key_enabled(&sched_feat_keys[i])) 188 static_key_slow_inc(&sched_feat_keys[i]); 189} 190#else 191static void sched_feat_disable(int i) { }; 192static void sched_feat_enable(int i) { }; 193#endif /* HAVE_JUMP_LABEL */ 194 195static ssize_t 196sched_feat_write(struct file *filp, const char __user *ubuf, 197 size_t cnt, loff_t *ppos) 198{ 199 char buf[64]; 200 char *cmp; 201 int neg = 0; 202 int i; 203 204 if (cnt > 63) 205 cnt = 63; 206 207 if (copy_from_user(&buf, ubuf, cnt)) 208 return -EFAULT; 209 210 buf[cnt] = 0; 211 cmp = strstrip(buf); 212 213 if (strncmp(cmp, "NO_", 3) == 0) { 214 neg = 1; 215 cmp += 3; 216 } 217 218 for (i = 0; i < __SCHED_FEAT_NR; i++) { 219 if (strcmp(cmp, sched_feat_names[i]) == 0) { 220 if (neg) { 221 sysctl_sched_features &= ~(1UL << i); 222 sched_feat_disable(i); 223 } else { 224 sysctl_sched_features |= (1UL << i); 225 sched_feat_enable(i); 226 } 227 break; 228 } 229 } 230 231 if (i == __SCHED_FEAT_NR) 232 return -EINVAL; 233 234 *ppos += cnt; 235 236 return cnt; 237} 238 239static int sched_feat_open(struct inode *inode, struct file *filp) 240{ 241 return single_open(filp, sched_feat_show, NULL); 242} 243 244static const struct file_operations sched_feat_fops = { 245 .open = sched_feat_open, 246 .write = sched_feat_write, 247 .read = seq_read, 248 .llseek = seq_lseek, 249 .release = single_release, 250}; 251 252static __init int sched_init_debug(void) 253{ 254 debugfs_create_file("sched_features", 0644, NULL, NULL, 255 &sched_feat_fops); 256 257 return 0; 258} 259late_initcall(sched_init_debug); 260#endif /* CONFIG_SCHED_DEBUG */ 261 262/* 263 * Number of tasks to iterate in a single balance run. 264 * Limited because this is done with IRQs disabled. 265 */ 266const_debug unsigned int sysctl_sched_nr_migrate = 32; 267 268/* 269 * period over which we average the RT time consumption, measured 270 * in ms. 271 * 272 * default: 1s 273 */ 274const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC; 275 276/* 277 * period over which we measure -rt task cpu usage in us. 278 * default: 1s 279 */ 280unsigned int sysctl_sched_rt_period = 1000000; 281 282__read_mostly int scheduler_running; 283 284/* 285 * part of the period that we allow rt tasks to run in us. 286 * default: 0.95s 287 */ 288int sysctl_sched_rt_runtime = 950000; 289 290 291 292/* 293 * __task_rq_lock - lock the rq @p resides on. 294 */ 295static inline struct rq *__task_rq_lock(struct task_struct *p) 296 __acquires(rq->lock) 297{ 298 struct rq *rq; 299 300 lockdep_assert_held(&p->pi_lock); 301 302 for (;;) { 303 rq = task_rq(p); 304 raw_spin_lock(&rq->lock); 305 if (likely(rq == task_rq(p))) 306 return rq; 307 raw_spin_unlock(&rq->lock); 308 } 309} 310 311/* 312 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on. 313 */ 314static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags) 315 __acquires(p->pi_lock) 316 __acquires(rq->lock) 317{ 318 struct rq *rq; 319 320 for (;;) { 321 raw_spin_lock_irqsave(&p->pi_lock, *flags); 322 rq = task_rq(p); 323 raw_spin_lock(&rq->lock); 324 if (likely(rq == task_rq(p))) 325 return rq; 326 raw_spin_unlock(&rq->lock); 327 raw_spin_unlock_irqrestore(&p->pi_lock, *flags); 328 } 329} 330 331static void __task_rq_unlock(struct rq *rq) 332 __releases(rq->lock) 333{ 334 raw_spin_unlock(&rq->lock); 335} 336 337static inline void 338task_rq_unlock(struct rq *rq, struct task_struct *p, unsigned long *flags) 339 __releases(rq->lock) 340 __releases(p->pi_lock) 341{ 342 raw_spin_unlock(&rq->lock); 343 raw_spin_unlock_irqrestore(&p->pi_lock, *flags); 344} 345 346/* 347 * this_rq_lock - lock this runqueue and disable interrupts. 348 */ 349static struct rq *this_rq_lock(void) 350 __acquires(rq->lock) 351{ 352 struct rq *rq; 353 354 local_irq_disable(); 355 rq = this_rq(); 356 raw_spin_lock(&rq->lock); 357 358 return rq; 359} 360 361#ifdef CONFIG_SCHED_HRTICK 362/* 363 * Use HR-timers to deliver accurate preemption points. 364 * 365 * Its all a bit involved since we cannot program an hrt while holding the 366 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a 367 * reschedule event. 368 * 369 * When we get rescheduled we reprogram the hrtick_timer outside of the 370 * rq->lock. 371 */ 372 373static void hrtick_clear(struct rq *rq) 374{ 375 if (hrtimer_active(&rq->hrtick_timer)) 376 hrtimer_cancel(&rq->hrtick_timer); 377} 378 379/* 380 * High-resolution timer tick. 381 * Runs from hardirq context with interrupts disabled. 382 */ 383static enum hrtimer_restart hrtick(struct hrtimer *timer) 384{ 385 struct rq *rq = container_of(timer, struct rq, hrtick_timer); 386 387 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id()); 388 389 raw_spin_lock(&rq->lock); 390 update_rq_clock(rq); 391 rq->curr->sched_class->task_tick(rq, rq->curr, 1); 392 raw_spin_unlock(&rq->lock); 393 394 return HRTIMER_NORESTART; 395} 396 397#ifdef CONFIG_SMP 398/* 399 * called from hardirq (IPI) context 400 */ 401static void __hrtick_start(void *arg) 402{ 403 struct rq *rq = arg; 404 405 raw_spin_lock(&rq->lock); 406 hrtimer_restart(&rq->hrtick_timer); 407 rq->hrtick_csd_pending = 0; 408 raw_spin_unlock(&rq->lock); 409} 410 411/* 412 * Called to set the hrtick timer state. 413 * 414 * called with rq->lock held and irqs disabled 415 */ 416void hrtick_start(struct rq *rq, u64 delay) 417{ 418 struct hrtimer *timer = &rq->hrtick_timer; 419 ktime_t time = ktime_add_ns(timer->base->get_time(), delay); 420 421 hrtimer_set_expires(timer, time); 422 423 if (rq == this_rq()) { 424 hrtimer_restart(timer); 425 } else if (!rq->hrtick_csd_pending) { 426 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd, 0); 427 rq->hrtick_csd_pending = 1; 428 } 429} 430 431static int 432hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu) 433{ 434 int cpu = (int)(long)hcpu; 435 436 switch (action) { 437 case CPU_UP_CANCELED: 438 case CPU_UP_CANCELED_FROZEN: 439 case CPU_DOWN_PREPARE: 440 case CPU_DOWN_PREPARE_FROZEN: 441 case CPU_DEAD: 442 case CPU_DEAD_FROZEN: 443 hrtick_clear(cpu_rq(cpu)); 444 return NOTIFY_OK; 445 } 446 447 return NOTIFY_DONE; 448} 449 450static __init void init_hrtick(void) 451{ 452 hotcpu_notifier(hotplug_hrtick, 0); 453} 454#else 455/* 456 * Called to set the hrtick timer state. 457 * 458 * called with rq->lock held and irqs disabled 459 */ 460void hrtick_start(struct rq *rq, u64 delay) 461{ 462 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0, 463 HRTIMER_MODE_REL_PINNED, 0); 464} 465 466static inline void init_hrtick(void) 467{ 468} 469#endif /* CONFIG_SMP */ 470 471static void init_rq_hrtick(struct rq *rq) 472{ 473#ifdef CONFIG_SMP 474 rq->hrtick_csd_pending = 0; 475 476 rq->hrtick_csd.flags = 0; 477 rq->hrtick_csd.func = __hrtick_start; 478 rq->hrtick_csd.info = rq; 479#endif 480 481 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); 482 rq->hrtick_timer.function = hrtick; 483} 484#else /* CONFIG_SCHED_HRTICK */ 485static inline void hrtick_clear(struct rq *rq) 486{ 487} 488 489static inline void init_rq_hrtick(struct rq *rq) 490{ 491} 492 493static inline void init_hrtick(void) 494{ 495} 496#endif /* CONFIG_SCHED_HRTICK */ 497 498/* 499 * resched_task - mark a task 'to be rescheduled now'. 500 * 501 * On UP this means the setting of the need_resched flag, on SMP it 502 * might also involve a cross-CPU call to trigger the scheduler on 503 * the target CPU. 504 */ 505#ifdef CONFIG_SMP 506 507#ifndef tsk_is_polling 508#define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG) 509#endif 510 511void resched_task(struct task_struct *p) 512{ 513 int cpu; 514 515 assert_raw_spin_locked(&task_rq(p)->lock); 516 517 if (test_tsk_need_resched(p)) 518 return; 519 520 set_tsk_need_resched(p); 521 522 cpu = task_cpu(p); 523 if (cpu == smp_processor_id()) 524 return; 525 526 /* NEED_RESCHED must be visible before we test polling */ 527 smp_mb(); 528 if (!tsk_is_polling(p)) 529 smp_send_reschedule(cpu); 530} 531 532void resched_cpu(int cpu) 533{ 534 struct rq *rq = cpu_rq(cpu); 535 unsigned long flags; 536 537 if (!raw_spin_trylock_irqsave(&rq->lock, flags)) 538 return; 539 resched_task(cpu_curr(cpu)); 540 raw_spin_unlock_irqrestore(&rq->lock, flags); 541} 542 543#ifdef CONFIG_NO_HZ 544/* 545 * In the semi idle case, use the nearest busy cpu for migrating timers 546 * from an idle cpu. This is good for power-savings. 547 * 548 * We don't do similar optimization for completely idle system, as 549 * selecting an idle cpu will add more delays to the timers than intended 550 * (as that cpu's timer base may not be uptodate wrt jiffies etc). 551 */ 552int get_nohz_timer_target(void) 553{ 554 int cpu = smp_processor_id(); 555 int i; 556 struct sched_domain *sd; 557 558 rcu_read_lock(); 559 for_each_domain(cpu, sd) { 560 for_each_cpu(i, sched_domain_span(sd)) { 561 if (!idle_cpu(i)) { 562 cpu = i; 563 goto unlock; 564 } 565 } 566 } 567unlock: 568 rcu_read_unlock(); 569 return cpu; 570} 571/* 572 * When add_timer_on() enqueues a timer into the timer wheel of an 573 * idle CPU then this timer might expire before the next timer event 574 * which is scheduled to wake up that CPU. In case of a completely 575 * idle system the next event might even be infinite time into the 576 * future. wake_up_idle_cpu() ensures that the CPU is woken up and 577 * leaves the inner idle loop so the newly added timer is taken into 578 * account when the CPU goes back to idle and evaluates the timer 579 * wheel for the next timer event. 580 */ 581void wake_up_idle_cpu(int cpu) 582{ 583 struct rq *rq = cpu_rq(cpu); 584 585 if (cpu == smp_processor_id()) 586 return; 587 588 /* 589 * This is safe, as this function is called with the timer 590 * wheel base lock of (cpu) held. When the CPU is on the way 591 * to idle and has not yet set rq->curr to idle then it will 592 * be serialized on the timer wheel base lock and take the new 593 * timer into account automatically. 594 */ 595 if (rq->curr != rq->idle) 596 return; 597 598 /* 599 * We can set TIF_RESCHED on the idle task of the other CPU 600 * lockless. The worst case is that the other CPU runs the 601 * idle task through an additional NOOP schedule() 602 */ 603 set_tsk_need_resched(rq->idle); 604 605 /* NEED_RESCHED must be visible before we test polling */ 606 smp_mb(); 607 if (!tsk_is_polling(rq->idle)) 608 smp_send_reschedule(cpu); 609} 610 611static inline bool got_nohz_idle_kick(void) 612{ 613 int cpu = smp_processor_id(); 614 return idle_cpu(cpu) && test_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu)); 615} 616 617#else /* CONFIG_NO_HZ */ 618 619static inline bool got_nohz_idle_kick(void) 620{ 621 return false; 622} 623 624#endif /* CONFIG_NO_HZ */ 625 626void sched_avg_update(struct rq *rq) 627{ 628 s64 period = sched_avg_period(); 629 630 while ((s64)(rq->clock - rq->age_stamp) > period) { 631 /* 632 * Inline assembly required to prevent the compiler 633 * optimising this loop into a divmod call. 634 * See __iter_div_u64_rem() for another example of this. 635 */ 636 asm("" : "+rm" (rq->age_stamp)); 637 rq->age_stamp += period; 638 rq->rt_avg /= 2; 639 } 640} 641 642#else /* !CONFIG_SMP */ 643void resched_task(struct task_struct *p) 644{ 645 assert_raw_spin_locked(&task_rq(p)->lock); 646 set_tsk_need_resched(p); 647} 648#endif /* CONFIG_SMP */ 649 650#if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \ 651 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH))) 652/* 653 * Iterate task_group tree rooted at *from, calling @down when first entering a 654 * node and @up when leaving it for the final time. 655 * 656 * Caller must hold rcu_lock or sufficient equivalent. 657 */ 658int walk_tg_tree_from(struct task_group *from, 659 tg_visitor down, tg_visitor up, void *data) 660{ 661 struct task_group *parent, *child; 662 int ret; 663 664 parent = from; 665 666down: 667 ret = (*down)(parent, data); 668 if (ret) 669 goto out; 670 list_for_each_entry_rcu(child, &parent->children, siblings) { 671 parent = child; 672 goto down; 673 674up: 675 continue; 676 } 677 ret = (*up)(parent, data); 678 if (ret || parent == from) 679 goto out; 680 681 child = parent; 682 parent = parent->parent; 683 if (parent) 684 goto up; 685out: 686 return ret; 687} 688 689int tg_nop(struct task_group *tg, void *data) 690{ 691 return 0; 692} 693#endif 694 695void update_cpu_load(struct rq *this_rq); 696 697static void set_load_weight(struct task_struct *p) 698{ 699 int prio = p->static_prio - MAX_RT_PRIO; 700 struct load_weight *load = &p->se.load; 701 702 /* 703 * SCHED_IDLE tasks get minimal weight: 704 */ 705 if (p->policy == SCHED_IDLE) { 706 load->weight = scale_load(WEIGHT_IDLEPRIO); 707 load->inv_weight = WMULT_IDLEPRIO; 708 return; 709 } 710 711 load->weight = scale_load(prio_to_weight[prio]); 712 load->inv_weight = prio_to_wmult[prio]; 713} 714 715static void enqueue_task(struct rq *rq, struct task_struct *p, int flags) 716{ 717 update_rq_clock(rq); 718 sched_info_queued(p); 719 p->sched_class->enqueue_task(rq, p, flags); 720} 721 722static void dequeue_task(struct rq *rq, struct task_struct *p, int flags) 723{ 724 update_rq_clock(rq); 725 sched_info_dequeued(p); 726 p->sched_class->dequeue_task(rq, p, flags); 727} 728 729void activate_task(struct rq *rq, struct task_struct *p, int flags) 730{ 731 if (task_contributes_to_load(p)) 732 rq->nr_uninterruptible--; 733 734 enqueue_task(rq, p, flags); 735} 736 737void deactivate_task(struct rq *rq, struct task_struct *p, int flags) 738{ 739 if (task_contributes_to_load(p)) 740 rq->nr_uninterruptible++; 741 742 dequeue_task(rq, p, flags); 743} 744 745#ifdef CONFIG_IRQ_TIME_ACCOUNTING 746 747/* 748 * There are no locks covering percpu hardirq/softirq time. 749 * They are only modified in account_system_vtime, on corresponding CPU 750 * with interrupts disabled. So, writes are safe. 751 * They are read and saved off onto struct rq in update_rq_clock(). 752 * This may result in other CPU reading this CPU's irq time and can 753 * race with irq/account_system_vtime on this CPU. We would either get old 754 * or new value with a side effect of accounting a slice of irq time to wrong 755 * task when irq is in progress while we read rq->clock. That is a worthy 756 * compromise in place of having locks on each irq in account_system_time. 757 */ 758static DEFINE_PER_CPU(u64, cpu_hardirq_time); 759static DEFINE_PER_CPU(u64, cpu_softirq_time); 760 761static DEFINE_PER_CPU(u64, irq_start_time); 762static int sched_clock_irqtime; 763 764void enable_sched_clock_irqtime(void) 765{ 766 sched_clock_irqtime = 1; 767} 768 769void disable_sched_clock_irqtime(void) 770{ 771 sched_clock_irqtime = 0; 772} 773 774#ifndef CONFIG_64BIT 775static DEFINE_PER_CPU(seqcount_t, irq_time_seq); 776 777static inline void irq_time_write_begin(void) 778{ 779 __this_cpu_inc(irq_time_seq.sequence); 780 smp_wmb(); 781} 782 783static inline void irq_time_write_end(void) 784{ 785 smp_wmb(); 786 __this_cpu_inc(irq_time_seq.sequence); 787} 788 789static inline u64 irq_time_read(int cpu) 790{ 791 u64 irq_time; 792 unsigned seq; 793 794 do { 795 seq = read_seqcount_begin(&per_cpu(irq_time_seq, cpu)); 796 irq_time = per_cpu(cpu_softirq_time, cpu) + 797 per_cpu(cpu_hardirq_time, cpu); 798 } while (read_seqcount_retry(&per_cpu(irq_time_seq, cpu), seq)); 799 800 return irq_time; 801} 802#else /* CONFIG_64BIT */ 803static inline void irq_time_write_begin(void) 804{ 805} 806 807static inline void irq_time_write_end(void) 808{ 809} 810 811static inline u64 irq_time_read(int cpu) 812{ 813 return per_cpu(cpu_softirq_time, cpu) + per_cpu(cpu_hardirq_time, cpu); 814} 815#endif /* CONFIG_64BIT */ 816 817/* 818 * Called before incrementing preempt_count on {soft,}irq_enter 819 * and before decrementing preempt_count on {soft,}irq_exit. 820 */ 821void account_system_vtime(struct task_struct *curr) 822{ 823 unsigned long flags; 824 s64 delta; 825 int cpu; 826 827 if (!sched_clock_irqtime) 828 return; 829 830 local_irq_save(flags); 831 832 cpu = smp_processor_id(); 833 delta = sched_clock_cpu(cpu) - __this_cpu_read(irq_start_time); 834 __this_cpu_add(irq_start_time, delta); 835 836 irq_time_write_begin(); 837 /* 838 * We do not account for softirq time from ksoftirqd here. 839 * We want to continue accounting softirq time to ksoftirqd thread 840 * in that case, so as not to confuse scheduler with a special task 841 * that do not consume any time, but still wants to run. 842 */ 843 if (hardirq_count()) 844 __this_cpu_add(cpu_hardirq_time, delta); 845 else if (in_serving_softirq() && curr != this_cpu_ksoftirqd()) 846 __this_cpu_add(cpu_softirq_time, delta); 847 848 irq_time_write_end(); 849 local_irq_restore(flags); 850} 851EXPORT_SYMBOL_GPL(account_system_vtime); 852 853#endif /* CONFIG_IRQ_TIME_ACCOUNTING */ 854 855#ifdef CONFIG_PARAVIRT 856static inline u64 steal_ticks(u64 steal) 857{ 858 if (unlikely(steal > NSEC_PER_SEC)) 859 return div_u64(steal, TICK_NSEC); 860 861 return __iter_div_u64_rem(steal, TICK_NSEC, &steal); 862} 863#endif 864 865static void update_rq_clock_task(struct rq *rq, s64 delta) 866{ 867/* 868 * In theory, the compile should just see 0 here, and optimize out the call 869 * to sched_rt_avg_update. But I don't trust it... 870 */ 871#if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING) 872 s64 steal = 0, irq_delta = 0; 873#endif 874#ifdef CONFIG_IRQ_TIME_ACCOUNTING 875 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time; 876 877 /* 878 * Since irq_time is only updated on {soft,}irq_exit, we might run into 879 * this case when a previous update_rq_clock() happened inside a 880 * {soft,}irq region. 881 * 882 * When this happens, we stop ->clock_task and only update the 883 * prev_irq_time stamp to account for the part that fit, so that a next 884 * update will consume the rest. This ensures ->clock_task is 885 * monotonic. 886 * 887 * It does however cause some slight miss-attribution of {soft,}irq 888 * time, a more accurate solution would be to update the irq_time using 889 * the current rq->clock timestamp, except that would require using 890 * atomic ops. 891 */ 892 if (irq_delta > delta) 893 irq_delta = delta; 894 895 rq->prev_irq_time += irq_delta; 896 delta -= irq_delta; 897#endif 898#ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING 899 if (static_key_false((¶virt_steal_rq_enabled))) { 900 u64 st; 901 902 steal = paravirt_steal_clock(cpu_of(rq)); 903 steal -= rq->prev_steal_time_rq; 904 905 if (unlikely(steal > delta)) 906 steal = delta; 907 908 st = steal_ticks(steal); 909 steal = st * TICK_NSEC; 910 911 rq->prev_steal_time_rq += steal; 912 913 delta -= steal; 914 } 915#endif 916 917 rq->clock_task += delta; 918 919#if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING) 920 if ((irq_delta + steal) && sched_feat(NONTASK_POWER)) 921 sched_rt_avg_update(rq, irq_delta + steal); 922#endif 923} 924 925#ifdef CONFIG_IRQ_TIME_ACCOUNTING 926static int irqtime_account_hi_update(void) 927{ 928 u64 *cpustat = kcpustat_this_cpu->cpustat; 929 unsigned long flags; 930 u64 latest_ns; 931 int ret = 0; 932 933 local_irq_save(flags); 934 latest_ns = this_cpu_read(cpu_hardirq_time); 935 if (nsecs_to_cputime64(latest_ns) > cpustat[CPUTIME_IRQ]) 936 ret = 1; 937 local_irq_restore(flags); 938 return ret; 939} 940 941static int irqtime_account_si_update(void) 942{ 943 u64 *cpustat = kcpustat_this_cpu->cpustat; 944 unsigned long flags; 945 u64 latest_ns; 946 int ret = 0; 947 948 local_irq_save(flags); 949 latest_ns = this_cpu_read(cpu_softirq_time); 950 if (nsecs_to_cputime64(latest_ns) > cpustat[CPUTIME_SOFTIRQ]) 951 ret = 1; 952 local_irq_restore(flags); 953 return ret; 954} 955 956#else /* CONFIG_IRQ_TIME_ACCOUNTING */ 957 958#define sched_clock_irqtime (0) 959 960#endif 961 962void sched_set_stop_task(int cpu, struct task_struct *stop) 963{ 964 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 }; 965 struct task_struct *old_stop = cpu_rq(cpu)->stop; 966 967 if (stop) { 968 /* 969 * Make it appear like a SCHED_FIFO task, its something 970 * userspace knows about and won't get confused about. 971 * 972 * Also, it will make PI more or less work without too 973 * much confusion -- but then, stop work should not 974 * rely on PI working anyway. 975 */ 976 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m); 977 978 stop->sched_class = &stop_sched_class; 979 } 980 981 cpu_rq(cpu)->stop = stop; 982 983 if (old_stop) { 984 /* 985 * Reset it back to a normal scheduling class so that 986 * it can die in pieces. 987 */ 988 old_stop->sched_class = &rt_sched_class; 989 } 990} 991 992/* 993 * __normal_prio - return the priority that is based on the static prio 994 */ 995static inline int __normal_prio(struct task_struct *p) 996{ 997 return p->static_prio; 998} 999 1000/* 1001 * Calculate the expected normal priority: i.e. priority 1002 * without taking RT-inheritance into account. Might be 1003 * boosted by interactivity modifiers. Changes upon fork, 1004 * setprio syscalls, and whenever the interactivity 1005 * estimator recalculates. 1006 */ 1007static inline int normal_prio(struct task_struct *p) 1008{ 1009 int prio; 1010 1011 if (task_has_rt_policy(p)) 1012 prio = MAX_RT_PRIO-1 - p->rt_priority; 1013 else 1014 prio = __normal_prio(p); 1015 return prio; 1016} 1017 1018/* 1019 * Calculate the current priority, i.e. the priority 1020 * taken into account by the scheduler. This value might 1021 * be boosted by RT tasks, or might be boosted by 1022 * interactivity modifiers. Will be RT if the task got 1023 * RT-boosted. If not then it returns p->normal_prio. 1024 */ 1025static int effective_prio(struct task_struct *p) 1026{ 1027 p->normal_prio = normal_prio(p); 1028 /* 1029 * If we are RT tasks or we were boosted to RT priority, 1030 * keep the priority unchanged. Otherwise, update priority 1031 * to the normal priority: 1032 */ 1033 if (!rt_prio(p->prio)) 1034 return p->normal_prio; 1035 return p->prio; 1036} 1037 1038/** 1039 * task_curr - is this task currently executing on a CPU? 1040 * @p: the task in question. 1041 */ 1042inline int task_curr(const struct task_struct *p) 1043{ 1044 return cpu_curr(task_cpu(p)) == p; 1045} 1046 1047static inline void check_class_changed(struct rq *rq, struct task_struct *p, 1048 const struct sched_class *prev_class, 1049 int oldprio) 1050{ 1051 if (prev_class != p->sched_class) { 1052 if (prev_class->switched_from) 1053 prev_class->switched_from(rq, p); 1054 p->sched_class->switched_to(rq, p); 1055 } else if (oldprio != p->prio) 1056 p->sched_class->prio_changed(rq, p, oldprio); 1057} 1058 1059void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags) 1060{ 1061 const struct sched_class *class; 1062 1063 if (p->sched_class == rq->curr->sched_class) { 1064 rq->curr->sched_class->check_preempt_curr(rq, p, flags); 1065 } else { 1066 for_each_class(class) { 1067 if (class == rq->curr->sched_class) 1068 break; 1069 if (class == p->sched_class) { 1070 resched_task(rq->curr); 1071 break; 1072 } 1073 } 1074 } 1075 1076 /* 1077 * A queue event has occurred, and we're going to schedule. In 1078 * this case, we can save a useless back to back clock update. 1079 */ 1080 if (rq->curr->on_rq && test_tsk_need_resched(rq->curr)) 1081 rq->skip_clock_update = 1; 1082} 1083 1084#ifdef CONFIG_SMP 1085void set_task_cpu(struct task_struct *p, unsigned int new_cpu) 1086{ 1087#ifdef CONFIG_SCHED_DEBUG 1088 /* 1089 * We should never call set_task_cpu() on a blocked task, 1090 * ttwu() will sort out the placement. 1091 */ 1092 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING && 1093 !(task_thread_info(p)->preempt_count & PREEMPT_ACTIVE)); 1094 1095#ifdef CONFIG_LOCKDEP 1096 /* 1097 * The caller should hold either p->pi_lock or rq->lock, when changing 1098 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks. 1099 * 1100 * sched_move_task() holds both and thus holding either pins the cgroup, 1101 * see set_task_rq(). 1102 * 1103 * Furthermore, all task_rq users should acquire both locks, see 1104 * task_rq_lock(). 1105 */ 1106 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) || 1107 lockdep_is_held(&task_rq(p)->lock))); 1108#endif 1109#endif 1110 1111 trace_sched_migrate_task(p, new_cpu); 1112 1113 if (task_cpu(p) != new_cpu) { 1114 p->se.nr_migrations++; 1115 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, NULL, 0); 1116 } 1117 1118 __set_task_cpu(p, new_cpu); 1119} 1120 1121struct migration_arg { 1122 struct task_struct *task; 1123 int dest_cpu; 1124}; 1125 1126static int migration_cpu_stop(void *data); 1127 1128/* 1129 * wait_task_inactive - wait for a thread to unschedule. 1130 * 1131 * If @match_state is nonzero, it's the @p->state value just checked and 1132 * not expected to change. If it changes, i.e. @p might have woken up, 1133 * then return zero. When we succeed in waiting for @p to be off its CPU, 1134 * we return a positive number (its total switch count). If a second call 1135 * a short while later returns the same number, the caller can be sure that 1136 * @p has remained unscheduled the whole time. 1137 * 1138 * The caller must ensure that the task *will* unschedule sometime soon, 1139 * else this function might spin for a *long* time. This function can't 1140 * be called with interrupts off, or it may introduce deadlock with 1141 * smp_call_function() if an IPI is sent by the same process we are 1142 * waiting to become inactive. 1143 */ 1144unsigned long wait_task_inactive(struct task_struct *p, long match_state) 1145{ 1146 unsigned long flags; 1147 int running, on_rq; 1148 unsigned long ncsw; 1149 struct rq *rq; 1150 1151 for (;;) { 1152 /* 1153 * We do the initial early heuristics without holding 1154 * any task-queue locks at all. We'll only try to get 1155 * the runqueue lock when things look like they will 1156 * work out! 1157 */ 1158 rq = task_rq(p); 1159 1160 /* 1161 * If the task is actively running on another CPU 1162 * still, just relax and busy-wait without holding 1163 * any locks. 1164 * 1165 * NOTE! Since we don't hold any locks, it's not 1166 * even sure that "rq" stays as the right runqueue! 1167 * But we don't care, since "task_running()" will 1168 * return false if the runqueue has changed and p 1169 * is actually now running somewhere else! 1170 */ 1171 while (task_running(rq, p)) { 1172 if (match_state && unlikely(p->state != match_state)) 1173 return 0; 1174 cpu_relax(); 1175 } 1176 1177 /* 1178 * Ok, time to look more closely! We need the rq 1179 * lock now, to be *sure*. If we're wrong, we'll 1180 * just go back and repeat. 1181 */ 1182 rq = task_rq_lock(p, &flags); 1183 trace_sched_wait_task(p); 1184 running = task_running(rq, p); 1185 on_rq = p->on_rq; 1186 ncsw = 0; 1187 if (!match_state || p->state == match_state) 1188 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */ 1189 task_rq_unlock(rq, p, &flags); 1190 1191 /* 1192 * If it changed from the expected state, bail out now. 1193 */ 1194 if (unlikely(!ncsw)) 1195 break; 1196 1197 /* 1198 * Was it really running after all now that we 1199 * checked with the proper locks actually held? 1200 * 1201 * Oops. Go back and try again.. 1202 */ 1203 if (unlikely(running)) { 1204 cpu_relax(); 1205 continue; 1206 } 1207 1208 /* 1209 * It's not enough that it's not actively running, 1210 * it must be off the runqueue _entirely_, and not 1211 * preempted! 1212 * 1213 * So if it was still runnable (but just not actively 1214 * running right now), it's preempted, and we should 1215 * yield - it could be a while. 1216 */ 1217 if (unlikely(on_rq)) { 1218 ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ); 1219 1220 set_current_state(TASK_UNINTERRUPTIBLE); 1221 schedule_hrtimeout(&to, HRTIMER_MODE_REL); 1222 continue; 1223 } 1224 1225 /* 1226 * Ahh, all good. It wasn't running, and it wasn't 1227 * runnable, which means that it will never become 1228 * running in the future either. We're all done! 1229 */ 1230 break; 1231 } 1232 1233 return ncsw; 1234} 1235 1236/*** 1237 * kick_process - kick a running thread to enter/exit the kernel 1238 * @p: the to-be-kicked thread 1239 * 1240 * Cause a process which is running on another CPU to enter 1241 * kernel-mode, without any delay. (to get signals handled.) 1242 * 1243 * NOTE: this function doesn't have to take the runqueue lock, 1244 * because all it wants to ensure is that the remote task enters 1245 * the kernel. If the IPI races and the task has been migrated 1246 * to another CPU then no harm is done and the purpose has been 1247 * achieved as well. 1248 */ 1249void kick_process(struct task_struct *p) 1250{ 1251 int cpu; 1252 1253 preempt_disable(); 1254 cpu = task_cpu(p); 1255 if ((cpu != smp_processor_id()) && task_curr(p)) 1256 smp_send_reschedule(cpu); 1257 preempt_enable(); 1258} 1259EXPORT_SYMBOL_GPL(kick_process); 1260#endif /* CONFIG_SMP */ 1261 1262#ifdef CONFIG_SMP 1263/* 1264 * ->cpus_allowed is protected by both rq->lock and p->pi_lock 1265 */ 1266static int select_fallback_rq(int cpu, struct task_struct *p) 1267{ 1268 const struct cpumask *nodemask = cpumask_of_node(cpu_to_node(cpu)); 1269 enum { cpuset, possible, fail } state = cpuset; 1270 int dest_cpu; 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 for (;;) { 1283 /* Any allowed, online CPU? */ 1284 for_each_cpu(dest_cpu, tsk_cpus_allowed(p)) { 1285 if (!cpu_online(dest_cpu)) 1286 continue; 1287 if (!cpu_active(dest_cpu)) 1288 continue; 1289 goto out; 1290 } 1291 1292 switch (state) { 1293 case cpuset: 1294 /* No more Mr. Nice Guy. */ 1295 cpuset_cpus_allowed_fallback(p); 1296 state = possible; 1297 break; 1298 1299 case possible: 1300 do_set_cpus_allowed(p, cpu_possible_mask); 1301 state = fail; 1302 break; 1303 1304 case fail: 1305 BUG(); 1306 break; 1307 } 1308 } 1309 1310out: 1311 if (state != cpuset) { 1312 /* 1313 * Don't tell them about moving exiting tasks or 1314 * kernel threads (both mm NULL), since they never 1315 * leave kernel. 1316 */ 1317 if (p->mm && printk_ratelimit()) { 1318 printk_sched("process %d (%s) no longer affine to cpu%d\n", 1319 task_pid_nr(p), p->comm, cpu); 1320 } 1321 } 1322 1323 return dest_cpu; 1324} 1325 1326/* 1327 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable. 1328 */ 1329static inline 1330int select_task_rq(struct task_struct *p, int sd_flags, int wake_flags) 1331{ 1332 int cpu = p->sched_class->select_task_rq(p, sd_flags, wake_flags); 1333 1334 /* 1335 * In order not to call set_task_cpu() on a blocking task we need 1336 * to rely on ttwu() to place the task on a valid ->cpus_allowed 1337 * cpu. 1338 * 1339 * Since this is common to all placement strategies, this lives here. 1340 * 1341 * [ this allows ->select_task() to simply return task_cpu(p) and 1342 * not worry about this generic constraint ] 1343 */ 1344 if (unlikely(!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)) || 1345 !cpu_online(cpu))) 1346 cpu = select_fallback_rq(task_cpu(p), p); 1347 1348 return cpu; 1349} 1350 1351static void update_avg(u64 *avg, u64 sample) 1352{ 1353 s64 diff = sample - *avg; 1354 *avg += diff >> 3; 1355} 1356#endif 1357 1358static void 1359ttwu_stat(struct task_struct *p, int cpu, int wake_flags) 1360{ 1361#ifdef CONFIG_SCHEDSTATS 1362 struct rq *rq = this_rq(); 1363 1364#ifdef CONFIG_SMP 1365 int this_cpu = smp_processor_id(); 1366 1367 if (cpu == this_cpu) { 1368 schedstat_inc(rq, ttwu_local); 1369 schedstat_inc(p, se.statistics.nr_wakeups_local); 1370 } else { 1371 struct sched_domain *sd; 1372 1373 schedstat_inc(p, se.statistics.nr_wakeups_remote); 1374 rcu_read_lock(); 1375 for_each_domain(this_cpu, sd) { 1376 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) { 1377 schedstat_inc(sd, ttwu_wake_remote); 1378 break; 1379 } 1380 } 1381 rcu_read_unlock(); 1382 } 1383 1384 if (wake_flags & WF_MIGRATED) 1385 schedstat_inc(p, se.statistics.nr_wakeups_migrate); 1386 1387#endif /* CONFIG_SMP */ 1388 1389 schedstat_inc(rq, ttwu_count); 1390 schedstat_inc(p, se.statistics.nr_wakeups); 1391 1392 if (wake_flags & WF_SYNC) 1393 schedstat_inc(p, se.statistics.nr_wakeups_sync); 1394 1395#endif /* CONFIG_SCHEDSTATS */ 1396} 1397 1398static void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags) 1399{ 1400 activate_task(rq, p, en_flags); 1401 p->on_rq = 1; 1402 1403 /* if a worker is waking up, notify workqueue */ 1404 if (p->flags & PF_WQ_WORKER) 1405 wq_worker_waking_up(p, cpu_of(rq)); 1406} 1407 1408/* 1409 * Mark the task runnable and perform wakeup-preemption. 1410 */ 1411static void 1412ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags) 1413{ 1414 trace_sched_wakeup(p, true); 1415 check_preempt_curr(rq, p, wake_flags); 1416 1417 p->state = TASK_RUNNING; 1418#ifdef CONFIG_SMP 1419 if (p->sched_class->task_woken) 1420 p->sched_class->task_woken(rq, p); 1421 1422 if (rq->idle_stamp) { 1423 u64 delta = rq->clock - rq->idle_stamp; 1424 u64 max = 2*sysctl_sched_migration_cost; 1425 1426 if (delta > max) 1427 rq->avg_idle = max; 1428 else 1429 update_avg(&rq->avg_idle, delta); 1430 rq->idle_stamp = 0; 1431 } 1432#endif 1433} 1434 1435static void 1436ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags) 1437{ 1438#ifdef CONFIG_SMP 1439 if (p->sched_contributes_to_load) 1440 rq->nr_uninterruptible--; 1441#endif 1442 1443 ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_WAKING); 1444 ttwu_do_wakeup(rq, p, wake_flags); 1445} 1446 1447/* 1448 * Called in case the task @p isn't fully descheduled from its runqueue, 1449 * in this case we must do a remote wakeup. Its a 'light' wakeup though, 1450 * since all we need to do is flip p->state to TASK_RUNNING, since 1451 * the task is still ->on_rq. 1452 */ 1453static int ttwu_remote(struct task_struct *p, int wake_flags) 1454{ 1455 struct rq *rq; 1456 int ret = 0; 1457 1458 rq = __task_rq_lock(p); 1459 if (p->on_rq) { 1460 ttwu_do_wakeup(rq, p, wake_flags); 1461 ret = 1; 1462 } 1463 __task_rq_unlock(rq); 1464 1465 return ret; 1466} 1467 1468#ifdef CONFIG_SMP 1469static void sched_ttwu_pending(void) 1470{ 1471 struct rq *rq = this_rq(); 1472 struct llist_node *llist = llist_del_all(&rq->wake_list); 1473 struct task_struct *p; 1474 1475 raw_spin_lock(&rq->lock); 1476 1477 while (llist) { 1478 p = llist_entry(llist, struct task_struct, wake_entry); 1479 llist = llist_next(llist); 1480 ttwu_do_activate(rq, p, 0); 1481 } 1482 1483 raw_spin_unlock(&rq->lock); 1484} 1485 1486void scheduler_ipi(void) 1487{ 1488 if (llist_empty(&this_rq()->wake_list) && !got_nohz_idle_kick()) 1489 return; 1490 1491 /* 1492 * Not all reschedule IPI handlers call irq_enter/irq_exit, since 1493 * traditionally all their work was done from the interrupt return 1494 * path. Now that we actually do some work, we need to make sure 1495 * we do call them. 1496 * 1497 * Some archs already do call them, luckily irq_enter/exit nest 1498 * properly. 1499 * 1500 * Arguably we should visit all archs and update all handlers, 1501 * however a fair share of IPIs are still resched only so this would 1502 * somewhat pessimize the simple resched case. 1503 */ 1504 irq_enter(); 1505 sched_ttwu_pending(); 1506 1507 /* 1508 * Check if someone kicked us for doing the nohz idle load balance. 1509 */ 1510 if (unlikely(got_nohz_idle_kick() && !need_resched())) { 1511 this_rq()->idle_balance = 1; 1512 raise_softirq_irqoff(SCHED_SOFTIRQ); 1513 } 1514 irq_exit(); 1515} 1516 1517static void ttwu_queue_remote(struct task_struct *p, int cpu) 1518{ 1519 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list)) 1520 smp_send_reschedule(cpu); 1521} 1522 1523#ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW 1524static int ttwu_activate_remote(struct task_struct *p, int wake_flags) 1525{ 1526 struct rq *rq; 1527 int ret = 0; 1528 1529 rq = __task_rq_lock(p); 1530 if (p->on_cpu) { 1531 ttwu_activate(rq, p, ENQUEUE_WAKEUP); 1532 ttwu_do_wakeup(rq, p, wake_flags); 1533 ret = 1; 1534 } 1535 __task_rq_unlock(rq); 1536 1537 return ret; 1538 1539} 1540#endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */ 1541 1542bool cpus_share_cache(int this_cpu, int that_cpu) 1543{ 1544 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu); 1545} 1546#endif /* CONFIG_SMP */ 1547 1548static void ttwu_queue(struct task_struct *p, int cpu) 1549{ 1550 struct rq *rq = cpu_rq(cpu); 1551 1552#if defined(CONFIG_SMP) 1553 if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) { 1554 sched_clock_cpu(cpu); /* sync clocks x-cpu */ 1555 ttwu_queue_remote(p, cpu); 1556 return; 1557 } 1558#endif 1559 1560 raw_spin_lock(&rq->lock); 1561 ttwu_do_activate(rq, p, 0); 1562 raw_spin_unlock(&rq->lock); 1563} 1564 1565/** 1566 * try_to_wake_up - wake up a thread 1567 * @p: the thread to be awakened 1568 * @state: the mask of task states that can be woken 1569 * @wake_flags: wake modifier flags (WF_*) 1570 * 1571 * Put it on the run-queue if it's not already there. The "current" 1572 * thread is always on the run-queue (except when the actual 1573 * re-schedule is in progress), and as such you're allowed to do 1574 * the simpler "current->state = TASK_RUNNING" to mark yourself 1575 * runnable without the overhead of this. 1576 * 1577 * Returns %true if @p was woken up, %false if it was already running 1578 * or @state didn't match @p's state. 1579 */ 1580static int 1581try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags) 1582{ 1583 unsigned long flags; 1584 int cpu, success = 0; 1585 1586 smp_wmb(); 1587 raw_spin_lock_irqsave(&p->pi_lock, flags); 1588 if (!(p->state & state)) 1589 goto out; 1590 1591 success = 1; /* we're going to change ->state */ 1592 cpu = task_cpu(p); 1593 1594 if (p->on_rq && ttwu_remote(p, wake_flags)) 1595 goto stat; 1596 1597#ifdef CONFIG_SMP 1598 /* 1599 * If the owning (remote) cpu is still in the middle of schedule() with 1600 * this task as prev, wait until its done referencing the task. 1601 */ 1602 while (p->on_cpu) { 1603#ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW 1604 /* 1605 * In case the architecture enables interrupts in 1606 * context_switch(), we cannot busy wait, since that 1607 * would lead to deadlocks when an interrupt hits and 1608 * tries to wake up @prev. So bail and do a complete 1609 * remote wakeup. 1610 */ 1611 if (ttwu_activate_remote(p, wake_flags)) 1612 goto stat; 1613#else 1614 cpu_relax(); 1615#endif 1616 } 1617 /* 1618 * Pairs with the smp_wmb() in finish_lock_switch(). 1619 */ 1620 smp_rmb(); 1621 1622 p->sched_contributes_to_load = !!task_contributes_to_load(p); 1623 p->state = TASK_WAKING; 1624 1625 if (p->sched_class->task_waking) 1626 p->sched_class->task_waking(p); 1627 1628 cpu = select_task_rq(p, SD_BALANCE_WAKE, wake_flags); 1629 if (task_cpu(p) != cpu) { 1630 wake_flags |= WF_MIGRATED; 1631 set_task_cpu(p, cpu); 1632 } 1633#endif /* CONFIG_SMP */ 1634 1635 ttwu_queue(p, cpu); 1636stat: 1637 ttwu_stat(p, cpu, wake_flags); 1638out: 1639 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 1640 1641 return success; 1642} 1643 1644/** 1645 * try_to_wake_up_local - try to wake up a local task with rq lock held 1646 * @p: the thread to be awakened 1647 * 1648 * Put @p on the run-queue if it's not already there. The caller must 1649 * ensure that this_rq() is locked, @p is bound to this_rq() and not 1650 * the current task. 1651 */ 1652static void try_to_wake_up_local(struct task_struct *p) 1653{ 1654 struct rq *rq = task_rq(p); 1655 1656 BUG_ON(rq != this_rq()); 1657 BUG_ON(p == current); 1658 lockdep_assert_held(&rq->lock); 1659 1660 if (!raw_spin_trylock(&p->pi_lock)) { 1661 raw_spin_unlock(&rq->lock); 1662 raw_spin_lock(&p->pi_lock); 1663 raw_spin_lock(&rq->lock); 1664 } 1665 1666 if (!(p->state & TASK_NORMAL)) 1667 goto out; 1668 1669 if (!p->on_rq) 1670 ttwu_activate(rq, p, ENQUEUE_WAKEUP); 1671 1672 ttwu_do_wakeup(rq, p, 0); 1673 ttwu_stat(p, smp_processor_id(), 0); 1674out: 1675 raw_spin_unlock(&p->pi_lock); 1676} 1677 1678/** 1679 * wake_up_process - Wake up a specific process 1680 * @p: The process to be woken up. 1681 * 1682 * Attempt to wake up the nominated process and move it to the set of runnable 1683 * processes. Returns 1 if the process was woken up, 0 if it was already 1684 * running. 1685 * 1686 * It may be assumed that this function implies a write memory barrier before 1687 * changing the task state if and only if any tasks are woken up. 1688 */ 1689int wake_up_process(struct task_struct *p) 1690{ 1691 return try_to_wake_up(p, TASK_ALL, 0); 1692} 1693EXPORT_SYMBOL(wake_up_process); 1694 1695int wake_up_state(struct task_struct *p, unsigned int state) 1696{ 1697 return try_to_wake_up(p, state, 0); 1698} 1699 1700/* 1701 * Perform scheduler related setup for a newly forked process p. 1702 * p is forked by current. 1703 * 1704 * __sched_fork() is basic setup used by init_idle() too: 1705 */ 1706static void __sched_fork(struct task_struct *p) 1707{ 1708 p->on_rq = 0; 1709 1710 p->se.on_rq = 0; 1711 p->se.exec_start = 0; 1712 p->se.sum_exec_runtime = 0; 1713 p->se.prev_sum_exec_runtime = 0; 1714 p->se.nr_migrations = 0; 1715 p->se.vruntime = 0; 1716 INIT_LIST_HEAD(&p->se.group_node); 1717 1718#ifdef CONFIG_SCHEDSTATS 1719 memset(&p->se.statistics, 0, sizeof(p->se.statistics)); 1720#endif 1721 1722 INIT_LIST_HEAD(&p->rt.run_list); 1723 1724#ifdef CONFIG_PREEMPT_NOTIFIERS 1725 INIT_HLIST_HEAD(&p->preempt_notifiers); 1726#endif 1727} 1728 1729/* 1730 * fork()/clone()-time setup: 1731 */ 1732void sched_fork(struct task_struct *p) 1733{ 1734 unsigned long flags; 1735 int cpu = get_cpu(); 1736 1737 __sched_fork(p); 1738 /* 1739 * We mark the process as running here. This guarantees that 1740 * nobody will actually run it, and a signal or other external 1741 * event cannot wake it up and insert it on the runqueue either. 1742 */ 1743 p->state = TASK_RUNNING; 1744 1745 /* 1746 * Make sure we do not leak PI boosting priority to the child. 1747 */ 1748 p->prio = current->normal_prio; 1749 1750 /* 1751 * Revert to default priority/policy on fork if requested. 1752 */ 1753 if (unlikely(p->sched_reset_on_fork)) { 1754 if (task_has_rt_policy(p)) { 1755 p->policy = SCHED_NORMAL; 1756 p->static_prio = NICE_TO_PRIO(0); 1757 p->rt_priority = 0; 1758 } else if (PRIO_TO_NICE(p->static_prio) < 0) 1759 p->static_prio = NICE_TO_PRIO(0); 1760 1761 p->prio = p->normal_prio = __normal_prio(p); 1762 set_load_weight(p); 1763 1764 /* 1765 * We don't need the reset flag anymore after the fork. It has 1766 * fulfilled its duty: 1767 */ 1768 p->sched_reset_on_fork = 0; 1769 } 1770 1771 if (!rt_prio(p->prio)) 1772 p->sched_class = &fair_sched_class; 1773 1774 if (p->sched_class->task_fork) 1775 p->sched_class->task_fork(p); 1776 1777 /* 1778 * The child is not yet in the pid-hash so no cgroup attach races, 1779 * and the cgroup is pinned to this child due to cgroup_fork() 1780 * is ran before sched_fork(). 1781 * 1782 * Silence PROVE_RCU. 1783 */ 1784 raw_spin_lock_irqsave(&p->pi_lock, flags); 1785 set_task_cpu(p, cpu); 1786 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 1787 1788#if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT) 1789 if (likely(sched_info_on())) 1790 memset(&p->sched_info, 0, sizeof(p->sched_info)); 1791#endif 1792#if defined(CONFIG_SMP) 1793 p->on_cpu = 0; 1794#endif 1795#ifdef CONFIG_PREEMPT_COUNT 1796 /* Want to start with kernel preemption disabled. */ 1797 task_thread_info(p)->preempt_count = 1; 1798#endif 1799#ifdef CONFIG_SMP 1800 plist_node_init(&p->pushable_tasks, MAX_PRIO); 1801#endif 1802 1803 put_cpu(); 1804} 1805 1806/* 1807 * wake_up_new_task - wake up a newly created task for the first time. 1808 * 1809 * This function will do some initial scheduler statistics housekeeping 1810 * that must be done for every newly created context, then puts the task 1811 * on the runqueue and wakes it. 1812 */ 1813void wake_up_new_task(struct task_struct *p) 1814{ 1815 unsigned long flags; 1816 struct rq *rq; 1817 1818 raw_spin_lock_irqsave(&p->pi_lock, flags); 1819#ifdef CONFIG_SMP 1820 /* 1821 * Fork balancing, do it here and not earlier because: 1822 * - cpus_allowed can change in the fork path 1823 * - any previously selected cpu might disappear through hotplug 1824 */ 1825 set_task_cpu(p, select_task_rq(p, SD_BALANCE_FORK, 0)); 1826#endif 1827 1828 rq = __task_rq_lock(p); 1829 activate_task(rq, p, 0); 1830 p->on_rq = 1; 1831 trace_sched_wakeup_new(p, true); 1832 check_preempt_curr(rq, p, WF_FORK); 1833#ifdef CONFIG_SMP 1834 if (p->sched_class->task_woken) 1835 p->sched_class->task_woken(rq, p); 1836#endif 1837 task_rq_unlock(rq, p, &flags); 1838} 1839 1840#ifdef CONFIG_PREEMPT_NOTIFIERS 1841 1842/** 1843 * preempt_notifier_register - tell me when current is being preempted & rescheduled 1844 * @notifier: notifier struct to register 1845 */ 1846void preempt_notifier_register(struct preempt_notifier *notifier) 1847{ 1848 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers); 1849} 1850EXPORT_SYMBOL_GPL(preempt_notifier_register); 1851 1852/** 1853 * preempt_notifier_unregister - no longer interested in preemption notifications 1854 * @notifier: notifier struct to unregister 1855 * 1856 * This is safe to call from within a preemption notifier. 1857 */ 1858void preempt_notifier_unregister(struct preempt_notifier *notifier) 1859{ 1860 hlist_del(¬ifier->link); 1861} 1862EXPORT_SYMBOL_GPL(preempt_notifier_unregister); 1863 1864static void fire_sched_in_preempt_notifiers(struct task_struct *curr) 1865{ 1866 struct preempt_notifier *notifier; 1867 struct hlist_node *node; 1868 1869 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link) 1870 notifier->ops->sched_in(notifier, raw_smp_processor_id()); 1871} 1872 1873static void 1874fire_sched_out_preempt_notifiers(struct task_struct *curr, 1875 struct task_struct *next) 1876{ 1877 struct preempt_notifier *notifier; 1878 struct hlist_node *node; 1879 1880 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link) 1881 notifier->ops->sched_out(notifier, next); 1882} 1883 1884#else /* !CONFIG_PREEMPT_NOTIFIERS */ 1885 1886static void fire_sched_in_preempt_notifiers(struct task_struct *curr) 1887{ 1888} 1889 1890static void 1891fire_sched_out_preempt_notifiers(struct task_struct *curr, 1892 struct task_struct *next) 1893{ 1894} 1895 1896#endif /* CONFIG_PREEMPT_NOTIFIERS */ 1897 1898/** 1899 * prepare_task_switch - prepare to switch tasks 1900 * @rq: the runqueue preparing to switch 1901 * @prev: the current task that is being switched out 1902 * @next: the task we are going to switch to. 1903 * 1904 * This is called with the rq lock held and interrupts off. It must 1905 * be paired with a subsequent finish_task_switch after the context 1906 * switch. 1907 * 1908 * prepare_task_switch sets up locking and calls architecture specific 1909 * hooks. 1910 */ 1911static inline void 1912prepare_task_switch(struct rq *rq, struct task_struct *prev, 1913 struct task_struct *next) 1914{ 1915 sched_info_switch(prev, next); 1916 perf_event_task_sched(prev, next); 1917 fire_sched_out_preempt_notifiers(prev, next); 1918 prepare_lock_switch(rq, next); 1919 prepare_arch_switch(next); 1920 trace_sched_switch(prev, next); 1921} 1922 1923/** 1924 * finish_task_switch - clean up after a task-switch 1925 * @rq: runqueue associated with task-switch 1926 * @prev: the thread we just switched away from. 1927 * 1928 * finish_task_switch must be called after the context switch, paired 1929 * with a prepare_task_switch call before the context switch. 1930 * finish_task_switch will reconcile locking set up by prepare_task_switch, 1931 * and do any other architecture-specific cleanup actions. 1932 * 1933 * Note that we may have delayed dropping an mm in context_switch(). If 1934 * so, we finish that here outside of the runqueue lock. (Doing it 1935 * with the lock held can cause deadlocks; see schedule() for 1936 * details.) 1937 */ 1938static void finish_task_switch(struct rq *rq, struct task_struct *prev) 1939 __releases(rq->lock) 1940{ 1941 struct mm_struct *mm = rq->prev_mm; 1942 long prev_state; 1943 1944 rq->prev_mm = NULL; 1945 1946 /* 1947 * A task struct has one reference for the use as "current". 1948 * If a task dies, then it sets TASK_DEAD in tsk->state and calls 1949 * schedule one last time. The schedule call will never return, and 1950 * the scheduled task must drop that reference. 1951 * The test for TASK_DEAD must occur while the runqueue locks are 1952 * still held, otherwise prev could be scheduled on another cpu, die 1953 * there before we look at prev->state, and then the reference would 1954 * be dropped twice. 1955 * Manfred Spraul <manfred@colorfullife.com> 1956 */ 1957 prev_state = prev->state; 1958 finish_arch_switch(prev); 1959 finish_lock_switch(rq, prev); 1960 finish_arch_post_lock_switch(); 1961 1962 fire_sched_in_preempt_notifiers(current); 1963 if (mm) 1964 mmdrop(mm); 1965 if (unlikely(prev_state == TASK_DEAD)) { 1966 /* 1967 * Remove function-return probe instances associated with this 1968 * task and put them back on the free list. 1969 */ 1970 kprobe_flush_task(prev); 1971 put_task_struct(prev); 1972 } 1973} 1974 1975#ifdef CONFIG_SMP 1976 1977/* assumes rq->lock is held */ 1978static inline void pre_schedule(struct rq *rq, struct task_struct *prev) 1979{ 1980 if (prev->sched_class->pre_schedule) 1981 prev->sched_class->pre_schedule(rq, prev); 1982} 1983 1984/* rq->lock is NOT held, but preemption is disabled */ 1985static inline void post_schedule(struct rq *rq) 1986{ 1987 if (rq->post_schedule) { 1988 unsigned long flags; 1989 1990 raw_spin_lock_irqsave(&rq->lock, flags); 1991 if (rq->curr->sched_class->post_schedule) 1992 rq->curr->sched_class->post_schedule(rq); 1993 raw_spin_unlock_irqrestore(&rq->lock, flags); 1994 1995 rq->post_schedule = 0; 1996 } 1997} 1998 1999#else 2000 2001static inline void pre_schedule(struct rq *rq, struct task_struct *p) 2002{ 2003} 2004 2005static inline void post_schedule(struct rq *rq) 2006{ 2007} 2008 2009#endif 2010 2011/** 2012 * schedule_tail - first thing a freshly forked thread must call. 2013 * @prev: the thread we just switched away from. 2014 */ 2015asmlinkage void schedule_tail(struct task_struct *prev) 2016 __releases(rq->lock) 2017{ 2018 struct rq *rq = this_rq(); 2019 2020 finish_task_switch(rq, prev); 2021 2022 /* 2023 * FIXME: do we need to worry about rq being invalidated by the 2024 * task_switch? 2025 */ 2026 post_schedule(rq); 2027 2028#ifdef __ARCH_WANT_UNLOCKED_CTXSW 2029 /* In this case, finish_task_switch does not reenable preemption */ 2030 preempt_enable(); 2031#endif 2032 if (current->set_child_tid) 2033 put_user(task_pid_vnr(current), current->set_child_tid); 2034} 2035 2036/* 2037 * context_switch - switch to the new MM and the new 2038 * thread's register state. 2039 */ 2040static inline void 2041context_switch(struct rq *rq, struct task_struct *prev, 2042 struct task_struct *next) 2043{ 2044 struct mm_struct *mm, *oldmm; 2045 2046 prepare_task_switch(rq, prev, next); 2047 2048 mm = next->mm; 2049 oldmm = prev->active_mm; 2050 /* 2051 * For paravirt, this is coupled with an exit in switch_to to 2052 * combine the page table reload and the switch backend into 2053 * one hypercall. 2054 */ 2055 arch_start_context_switch(prev); 2056 2057 if (!mm) { 2058 next->active_mm = oldmm; 2059 atomic_inc(&oldmm->mm_count); 2060 enter_lazy_tlb(oldmm, next); 2061 } else 2062 switch_mm(oldmm, mm, next); 2063 2064 if (!prev->mm) { 2065 prev->active_mm = NULL; 2066 rq->prev_mm = oldmm; 2067 } 2068 /* 2069 * Since the runqueue lock will be released by the next 2070 * task (which is an invalid locking op but in the case 2071 * of the scheduler it's an obvious special-case), so we 2072 * do an early lockdep release here: 2073 */ 2074#ifndef __ARCH_WANT_UNLOCKED_CTXSW 2075 spin_release(&rq->lock.dep_map, 1, _THIS_IP_); 2076#endif 2077 2078 /* Here we just switch the register state and the stack. */ 2079 switch_to(prev, next, prev); 2080 2081 barrier(); 2082 /* 2083 * this_rq must be evaluated again because prev may have moved 2084 * CPUs since it called schedule(), thus the 'rq' on its stack 2085 * frame will be invalid. 2086 */ 2087 finish_task_switch(this_rq(), prev); 2088} 2089 2090/* 2091 * nr_running, nr_uninterruptible and nr_context_switches: 2092 * 2093 * externally visible scheduler statistics: current number of runnable 2094 * threads, current number of uninterruptible-sleeping threads, total 2095 * number of context switches performed since bootup. 2096 */ 2097unsigned long nr_running(void) 2098{ 2099 unsigned long i, sum = 0; 2100 2101 for_each_online_cpu(i) 2102 sum += cpu_rq(i)->nr_running; 2103 2104 return sum; 2105} 2106 2107unsigned long nr_uninterruptible(void) 2108{ 2109 unsigned long i, sum = 0; 2110 2111 for_each_possible_cpu(i) 2112 sum += cpu_rq(i)->nr_uninterruptible; 2113 2114 /* 2115 * Since we read the counters lockless, it might be slightly 2116 * inaccurate. Do not allow it to go below zero though: 2117 */ 2118 if (unlikely((long)sum < 0)) 2119 sum = 0; 2120 2121 return sum; 2122} 2123 2124unsigned long long nr_context_switches(void) 2125{ 2126 int i; 2127 unsigned long long sum = 0; 2128 2129 for_each_possible_cpu(i) 2130 sum += cpu_rq(i)->nr_switches; 2131 2132 return sum; 2133} 2134 2135unsigned long nr_iowait(void) 2136{ 2137 unsigned long i, sum = 0; 2138 2139 for_each_possible_cpu(i) 2140 sum += atomic_read(&cpu_rq(i)->nr_iowait); 2141 2142 return sum; 2143} 2144 2145unsigned long nr_iowait_cpu(int cpu) 2146{ 2147 struct rq *this = cpu_rq(cpu); 2148 return atomic_read(&this->nr_iowait); 2149} 2150 2151unsigned long this_cpu_load(void) 2152{ 2153 struct rq *this = this_rq(); 2154 return this->cpu_load[0]; 2155} 2156 2157 2158/* Variables and functions for calc_load */ 2159static atomic_long_t calc_load_tasks; 2160static unsigned long calc_load_update; 2161unsigned long avenrun[3]; 2162EXPORT_SYMBOL(avenrun); 2163 2164static long calc_load_fold_active(struct rq *this_rq) 2165{ 2166 long nr_active, delta = 0; 2167 2168 nr_active = this_rq->nr_running; 2169 nr_active += (long) this_rq->nr_uninterruptible; 2170 2171 if (nr_active != this_rq->calc_load_active) { 2172 delta = nr_active - this_rq->calc_load_active; 2173 this_rq->calc_load_active = nr_active; 2174 } 2175 2176 return delta; 2177} 2178 2179static unsigned long 2180calc_load(unsigned long load, unsigned long exp, unsigned long active) 2181{ 2182 load *= exp; 2183 load += active * (FIXED_1 - exp); 2184 load += 1UL << (FSHIFT - 1); 2185 return load >> FSHIFT; 2186} 2187 2188#ifdef CONFIG_NO_HZ 2189/* 2190 * For NO_HZ we delay the active fold to the next LOAD_FREQ update. 2191 * 2192 * When making the ILB scale, we should try to pull this in as well. 2193 */ 2194static atomic_long_t calc_load_tasks_idle; 2195 2196void calc_load_account_idle(struct rq *this_rq) 2197{ 2198 long delta; 2199 2200 delta = calc_load_fold_active(this_rq); 2201 if (delta) 2202 atomic_long_add(delta, &calc_load_tasks_idle); 2203} 2204 2205static long calc_load_fold_idle(void) 2206{ 2207 long delta = 0; 2208 2209 /* 2210 * Its got a race, we don't care... 2211 */ 2212 if (atomic_long_read(&calc_load_tasks_idle)) 2213 delta = atomic_long_xchg(&calc_load_tasks_idle, 0); 2214 2215 return delta; 2216} 2217 2218/** 2219 * fixed_power_int - compute: x^n, in O(log n) time 2220 * 2221 * @x: base of the power 2222 * @frac_bits: fractional bits of @x 2223 * @n: power to raise @x to. 2224 * 2225 * By exploiting the relation between the definition of the natural power 2226 * function: x^n := x*x*...*x (x multiplied by itself for n times), and 2227 * the binary encoding of numbers used by computers: n := \Sum n_i * 2^i, 2228 * (where: n_i \elem {0, 1}, the binary vector representing n), 2229 * we find: x^n := x^(\Sum n_i * 2^i) := \Prod x^(n_i * 2^i), which is 2230 * of course trivially computable in O(log_2 n), the length of our binary 2231 * vector. 2232 */ 2233static unsigned long 2234fixed_power_int(unsigned long x, unsigned int frac_bits, unsigned int n) 2235{ 2236 unsigned long result = 1UL << frac_bits; 2237 2238 if (n) for (;;) { 2239 if (n & 1) { 2240 result *= x; 2241 result += 1UL << (frac_bits - 1); 2242 result >>= frac_bits; 2243 } 2244 n >>= 1; 2245 if (!n) 2246 break; 2247 x *= x; 2248 x += 1UL << (frac_bits - 1); 2249 x >>= frac_bits; 2250 } 2251 2252 return result; 2253} 2254 2255/* 2256 * a1 = a0 * e + a * (1 - e) 2257 * 2258 * a2 = a1 * e + a * (1 - e) 2259 * = (a0 * e + a * (1 - e)) * e + a * (1 - e) 2260 * = a0 * e^2 + a * (1 - e) * (1 + e) 2261 * 2262 * a3 = a2 * e + a * (1 - e) 2263 * = (a0 * e^2 + a * (1 - e) * (1 + e)) * e + a * (1 - e) 2264 * = a0 * e^3 + a * (1 - e) * (1 + e + e^2) 2265 * 2266 * ... 2267 * 2268 * an = a0 * e^n + a * (1 - e) * (1 + e + ... + e^n-1) [1] 2269 * = a0 * e^n + a * (1 - e) * (1 - e^n)/(1 - e) 2270 * = a0 * e^n + a * (1 - e^n) 2271 * 2272 * [1] application of the geometric series: 2273 * 2274 * n 1 - x^(n+1) 2275 * S_n := \Sum x^i = ------------- 2276 * i=0 1 - x 2277 */ 2278static unsigned long 2279calc_load_n(unsigned long load, unsigned long exp, 2280 unsigned long active, unsigned int n) 2281{ 2282 2283 return calc_load(load, fixed_power_int(exp, FSHIFT, n), active); 2284} 2285 2286/* 2287 * NO_HZ can leave us missing all per-cpu ticks calling 2288 * calc_load_account_active(), but since an idle CPU folds its delta into 2289 * calc_load_tasks_idle per calc_load_account_idle(), all we need to do is fold 2290 * in the pending idle delta if our idle period crossed a load cycle boundary. 2291 * 2292 * Once we've updated the global active value, we need to apply the exponential 2293 * weights adjusted to the number of cycles missed. 2294 */ 2295static void calc_global_nohz(void) 2296{ 2297 long delta, active, n; 2298 2299 /* 2300 * If we crossed a calc_load_update boundary, make sure to fold 2301 * any pending idle changes, the respective CPUs might have 2302 * missed the tick driven calc_load_account_active() update 2303 * due to NO_HZ. 2304 */ 2305 delta = calc_load_fold_idle(); 2306 if (delta) 2307 atomic_long_add(delta, &calc_load_tasks); 2308 2309 /* 2310 * It could be the one fold was all it took, we done! 2311 */ 2312 if (time_before(jiffies, calc_load_update + 10)) 2313 return; 2314 2315 /* 2316 * Catch-up, fold however many we are behind still 2317 */ 2318 delta = jiffies - calc_load_update - 10; 2319 n = 1 + (delta / LOAD_FREQ); 2320 2321 active = atomic_long_read(&calc_load_tasks); 2322 active = active > 0 ? active * FIXED_1 : 0; 2323 2324 avenrun[0] = calc_load_n(avenrun[0], EXP_1, active, n); 2325 avenrun[1] = calc_load_n(avenrun[1], EXP_5, active, n); 2326 avenrun[2] = calc_load_n(avenrun[2], EXP_15, active, n); 2327 2328 calc_load_update += n * LOAD_FREQ; 2329} 2330#else 2331void calc_load_account_idle(struct rq *this_rq) 2332{ 2333} 2334 2335static inline long calc_load_fold_idle(void) 2336{ 2337 return 0; 2338} 2339 2340static void calc_global_nohz(void) 2341{ 2342} 2343#endif 2344 2345/** 2346 * get_avenrun - get the load average array 2347 * @loads: pointer to dest load array 2348 * @offset: offset to add 2349 * @shift: shift count to shift the result left 2350 * 2351 * These values are estimates at best, so no need for locking. 2352 */ 2353void get_avenrun(unsigned long *loads, unsigned long offset, int shift) 2354{ 2355 loads[0] = (avenrun[0] + offset) << shift; 2356 loads[1] = (avenrun[1] + offset) << shift; 2357 loads[2] = (avenrun[2] + offset) << shift; 2358} 2359 2360/* 2361 * calc_load - update the avenrun load estimates 10 ticks after the 2362 * CPUs have updated calc_load_tasks. 2363 */ 2364void calc_global_load(unsigned long ticks) 2365{ 2366 long active; 2367 2368 if (time_before(jiffies, calc_load_update + 10)) 2369 return; 2370 2371 active = atomic_long_read(&calc_load_tasks); 2372 active = active > 0 ? active * FIXED_1 : 0; 2373 2374 avenrun[0] = calc_load(avenrun[0], EXP_1, active); 2375 avenrun[1] = calc_load(avenrun[1], EXP_5, active); 2376 avenrun[2] = calc_load(avenrun[2], EXP_15, active); 2377 2378 calc_load_update += LOAD_FREQ; 2379 2380 /* 2381 * Account one period with whatever state we found before 2382 * folding in the nohz state and ageing the entire idle period. 2383 * 2384 * This avoids loosing a sample when we go idle between 2385 * calc_load_account_active() (10 ticks ago) and now and thus 2386 * under-accounting. 2387 */ 2388 calc_global_nohz(); 2389} 2390 2391/* 2392 * Called from update_cpu_load() to periodically update this CPU's 2393 * active count. 2394 */ 2395static void calc_load_account_active(struct rq *this_rq) 2396{ 2397 long delta; 2398 2399 if (time_before(jiffies, this_rq->calc_load_update)) 2400 return; 2401 2402 delta = calc_load_fold_active(this_rq); 2403 delta += calc_load_fold_idle(); 2404 if (delta) 2405 atomic_long_add(delta, &calc_load_tasks); 2406 2407 this_rq->calc_load_update += LOAD_FREQ; 2408} 2409 2410/* 2411 * The exact cpuload at various idx values, calculated at every tick would be 2412 * load = (2^idx - 1) / 2^idx * load + 1 / 2^idx * cur_load 2413 * 2414 * If a cpu misses updates for n-1 ticks (as it was idle) and update gets called 2415 * on nth tick when cpu may be busy, then we have: 2416 * load = ((2^idx - 1) / 2^idx)^(n-1) * load 2417 * load = (2^idx - 1) / 2^idx) * load + 1 / 2^idx * cur_load 2418 * 2419 * decay_load_missed() below does efficient calculation of 2420 * load = ((2^idx - 1) / 2^idx)^(n-1) * load 2421 * avoiding 0..n-1 loop doing load = ((2^idx - 1) / 2^idx) * load 2422 * 2423 * The calculation is approximated on a 128 point scale. 2424 * degrade_zero_ticks is the number of ticks after which load at any 2425 * particular idx is approximated to be zero. 2426 * degrade_factor is a precomputed table, a row for each load idx. 2427 * Each column corresponds to degradation factor for a power of two ticks, 2428 * based on 128 point scale. 2429 * Example: 2430 * row 2, col 3 (=12) says that the degradation at load idx 2 after 2431 * 8 ticks is 12/128 (which is an approximation of exact factor 3^8/4^8). 2432 * 2433 * With this power of 2 load factors, we can degrade the load n times 2434 * by looking at 1 bits in n and doing as many mult/shift instead of 2435 * n mult/shifts needed by the exact degradation. 2436 */ 2437#define DEGRADE_SHIFT 7 2438static const unsigned char 2439 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128}; 2440static const unsigned char 2441 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = { 2442 {0, 0, 0, 0, 0, 0, 0, 0}, 2443 {64, 32, 8, 0, 0, 0, 0, 0}, 2444 {96, 72, 40, 12, 1, 0, 0}, 2445 {112, 98, 75, 43, 15, 1, 0}, 2446 {120, 112, 98, 76, 45, 16, 2} }; 2447 2448/* 2449 * Update cpu_load for any missed ticks, due to tickless idle. The backlog 2450 * would be when CPU is idle and so we just decay the old load without 2451 * adding any new load. 2452 */ 2453static unsigned long 2454decay_load_missed(unsigned long load, unsigned long missed_updates, int idx) 2455{ 2456 int j = 0; 2457 2458 if (!missed_updates) 2459 return load; 2460 2461 if (missed_updates >= degrade_zero_ticks[idx]) 2462 return 0; 2463 2464 if (idx == 1) 2465 return load >> missed_updates; 2466 2467 while (missed_updates) { 2468 if (missed_updates % 2) 2469 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT; 2470 2471 missed_updates >>= 1; 2472 j++; 2473 } 2474 return load; 2475} 2476 2477/* 2478 * Update rq->cpu_load[] statistics. This function is usually called every 2479 * scheduler tick (TICK_NSEC). With tickless idle this will not be called 2480 * every tick. We fix it up based on jiffies. 2481 */ 2482void update_cpu_load(struct rq *this_rq) 2483{ 2484 unsigned long this_load = this_rq->load.weight; 2485 unsigned long curr_jiffies = jiffies; 2486 unsigned long pending_updates; 2487 int i, scale; 2488 2489 this_rq->nr_load_updates++; 2490 2491 /* Avoid repeated calls on same jiffy, when moving in and out of idle */ 2492 if (curr_jiffies == this_rq->last_load_update_tick) 2493 return; 2494 2495 pending_updates = curr_jiffies - this_rq->last_load_update_tick; 2496 this_rq->last_load_update_tick = curr_jiffies; 2497 2498 /* Update our load: */ 2499 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */ 2500 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) { 2501 unsigned long old_load, new_load; 2502 2503 /* scale is effectively 1 << i now, and >> i divides by scale */ 2504 2505 old_load = this_rq->cpu_load[i]; 2506 old_load = decay_load_missed(old_load, pending_updates - 1, i); 2507 new_load = this_load; 2508 /* 2509 * Round up the averaging division if load is increasing. This 2510 * prevents us from getting stuck on 9 if the load is 10, for 2511 * example. 2512 */ 2513 if (new_load > old_load) 2514 new_load += scale - 1; 2515 2516 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i; 2517 } 2518 2519 sched_avg_update(this_rq); 2520} 2521 2522static void update_cpu_load_active(struct rq *this_rq) 2523{ 2524 update_cpu_load(this_rq); 2525 2526 calc_load_account_active(this_rq); 2527} 2528 2529#ifdef CONFIG_SMP 2530 2531/* 2532 * sched_exec - execve() is a valuable balancing opportunity, because at 2533 * this point the task has the smallest effective memory and cache footprint. 2534 */ 2535void sched_exec(void) 2536{ 2537 struct task_struct *p = current; 2538 unsigned long flags; 2539 int dest_cpu; 2540 2541 raw_spin_lock_irqsave(&p->pi_lock, flags); 2542 dest_cpu = p->sched_class->select_task_rq(p, SD_BALANCE_EXEC, 0); 2543 if (dest_cpu == smp_processor_id()) 2544 goto unlock; 2545 2546 if (likely(cpu_active(dest_cpu))) { 2547 struct migration_arg arg = { p, dest_cpu }; 2548 2549 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 2550 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg); 2551 return; 2552 } 2553unlock: 2554 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 2555} 2556 2557#endif 2558 2559DEFINE_PER_CPU(struct kernel_stat, kstat); 2560DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat); 2561 2562EXPORT_PER_CPU_SYMBOL(kstat); 2563EXPORT_PER_CPU_SYMBOL(kernel_cpustat); 2564 2565/* 2566 * Return any ns on the sched_clock that have not yet been accounted in 2567 * @p in case that task is currently running. 2568 * 2569 * Called with task_rq_lock() held on @rq. 2570 */ 2571static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq) 2572{ 2573 u64 ns = 0; 2574 2575 if (task_current(rq, p)) { 2576 update_rq_clock(rq); 2577 ns = rq->clock_task - p->se.exec_start; 2578 if ((s64)ns < 0) 2579 ns = 0; 2580 } 2581 2582 return ns; 2583} 2584 2585unsigned long long task_delta_exec(struct task_struct *p) 2586{ 2587 unsigned long flags; 2588 struct rq *rq; 2589 u64 ns = 0; 2590 2591 rq = task_rq_lock(p, &flags); 2592 ns = do_task_delta_exec(p, rq); 2593 task_rq_unlock(rq, p, &flags); 2594 2595 return ns; 2596} 2597 2598/* 2599 * Return accounted runtime for the task. 2600 * In case the task is currently running, return the runtime plus current's 2601 * pending runtime that have not been accounted yet. 2602 */ 2603unsigned long long task_sched_runtime(struct task_struct *p) 2604{ 2605 unsigned long flags; 2606 struct rq *rq; 2607 u64 ns = 0; 2608 2609 rq = task_rq_lock(p, &flags); 2610 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq); 2611 task_rq_unlock(rq, p, &flags); 2612 2613 return ns; 2614} 2615 2616#ifdef CONFIG_CGROUP_CPUACCT 2617struct cgroup_subsys cpuacct_subsys; 2618struct cpuacct root_cpuacct; 2619#endif 2620 2621static inline void task_group_account_field(struct task_struct *p, int index, 2622 u64 tmp) 2623{ 2624#ifdef CONFIG_CGROUP_CPUACCT 2625 struct kernel_cpustat *kcpustat; 2626 struct cpuacct *ca; 2627#endif 2628 /* 2629 * Since all updates are sure to touch the root cgroup, we 2630 * get ourselves ahead and touch it first. If the root cgroup 2631 * is the only cgroup, then nothing else should be necessary. 2632 * 2633 */ 2634 __get_cpu_var(kernel_cpustat).cpustat[index] += tmp; 2635 2636#ifdef CONFIG_CGROUP_CPUACCT 2637 if (unlikely(!cpuacct_subsys.active)) 2638 return; 2639 2640 rcu_read_lock(); 2641 ca = task_ca(p); 2642 while (ca && (ca != &root_cpuacct)) { 2643 kcpustat = this_cpu_ptr(ca->cpustat); 2644 kcpustat->cpustat[index] += tmp; 2645 ca = parent_ca(ca); 2646 } 2647 rcu_read_unlock(); 2648#endif 2649} 2650 2651 2652/* 2653 * Account user cpu time to a process. 2654 * @p: the process that the cpu time gets accounted to 2655 * @cputime: the cpu time spent in user space since the last update 2656 * @cputime_scaled: cputime scaled by cpu frequency 2657 */ 2658void account_user_time(struct task_struct *p, cputime_t cputime, 2659 cputime_t cputime_scaled) 2660{ 2661 int index; 2662 2663 /* Add user time to process. */ 2664 p->utime += cputime; 2665 p->utimescaled += cputime_scaled; 2666 account_group_user_time(p, cputime); 2667 2668 index = (TASK_NICE(p) > 0) ? CPUTIME_NICE : CPUTIME_USER; 2669 2670 /* Add user time to cpustat. */ 2671 task_group_account_field(p, index, (__force u64) cputime); 2672 2673 /* Account for user time used */ 2674 acct_update_integrals(p); 2675} 2676 2677/* 2678 * Account guest cpu time to a process. 2679 * @p: the process that the cpu time gets accounted to 2680 * @cputime: the cpu time spent in virtual machine since the last update 2681 * @cputime_scaled: cputime scaled by cpu frequency 2682 */ 2683static void account_guest_time(struct task_struct *p, cputime_t cputime, 2684 cputime_t cputime_scaled) 2685{ 2686 u64 *cpustat = kcpustat_this_cpu->cpustat; 2687 2688 /* Add guest time to process. */ 2689 p->utime += cputime; 2690 p->utimescaled += cputime_scaled; 2691 account_group_user_time(p, cputime); 2692 p->gtime += cputime; 2693 2694 /* Add guest time to cpustat. */ 2695 if (TASK_NICE(p) > 0) { 2696 cpustat[CPUTIME_NICE] += (__force u64) cputime; 2697 cpustat[CPUTIME_GUEST_NICE] += (__force u64) cputime; 2698 } else { 2699 cpustat[CPUTIME_USER] += (__force u64) cputime; 2700 cpustat[CPUTIME_GUEST] += (__force u64) cputime; 2701 } 2702} 2703 2704/* 2705 * Account system cpu time to a process and desired cpustat field 2706 * @p: the process that the cpu time gets accounted to 2707 * @cputime: the cpu time spent in kernel space since the last update 2708 * @cputime_scaled: cputime scaled by cpu frequency 2709 * @target_cputime64: pointer to cpustat field that has to be updated 2710 */ 2711static inline 2712void __account_system_time(struct task_struct *p, cputime_t cputime, 2713 cputime_t cputime_scaled, int index) 2714{ 2715 /* Add system time to process. */ 2716 p->stime += cputime; 2717 p->stimescaled += cputime_scaled; 2718 account_group_system_time(p, cputime); 2719 2720 /* Add system time to cpustat. */ 2721 task_group_account_field(p, index, (__force u64) cputime); 2722 2723 /* Account for system time used */ 2724 acct_update_integrals(p); 2725} 2726 2727/* 2728 * Account system cpu time to a process. 2729 * @p: the process that the cpu time gets accounted to 2730 * @hardirq_offset: the offset to subtract from hardirq_count() 2731 * @cputime: the cpu time spent in kernel space since the last update 2732 * @cputime_scaled: cputime scaled by cpu frequency 2733 */ 2734void account_system_time(struct task_struct *p, int hardirq_offset, 2735 cputime_t cputime, cputime_t cputime_scaled) 2736{ 2737 int index; 2738 2739 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) { 2740 account_guest_time(p, cputime, cputime_scaled); 2741 return; 2742 } 2743 2744 if (hardirq_count() - hardirq_offset) 2745 index = CPUTIME_IRQ; 2746 else if (in_serving_softirq()) 2747 index = CPUTIME_SOFTIRQ; 2748 else 2749 index = CPUTIME_SYSTEM; 2750 2751 __account_system_time(p, cputime, cputime_scaled, index); 2752} 2753 2754/* 2755 * Account for involuntary wait time. 2756 * @cputime: the cpu time spent in involuntary wait 2757 */ 2758void account_steal_time(cputime_t cputime) 2759{ 2760 u64 *cpustat = kcpustat_this_cpu->cpustat; 2761 2762 cpustat[CPUTIME_STEAL] += (__force u64) cputime; 2763} 2764 2765/* 2766 * Account for idle time. 2767 * @cputime: the cpu time spent in idle wait 2768 */ 2769void account_idle_time(cputime_t cputime) 2770{ 2771 u64 *cpustat = kcpustat_this_cpu->cpustat; 2772 struct rq *rq = this_rq(); 2773 2774 if (atomic_read(&rq->nr_iowait) > 0) 2775 cpustat[CPUTIME_IOWAIT] += (__force u64) cputime; 2776 else 2777 cpustat[CPUTIME_IDLE] += (__force u64) cputime; 2778} 2779 2780static __always_inline bool steal_account_process_tick(void) 2781{ 2782#ifdef CONFIG_PARAVIRT 2783 if (static_key_false(¶virt_steal_enabled)) { 2784 u64 steal, st = 0; 2785 2786 steal = paravirt_steal_clock(smp_processor_id()); 2787 steal -= this_rq()->prev_steal_time; 2788 2789 st = steal_ticks(steal); 2790 this_rq()->prev_steal_time += st * TICK_NSEC; 2791 2792 account_steal_time(st); 2793 return st; 2794 } 2795#endif 2796 return false; 2797} 2798 2799#ifndef CONFIG_VIRT_CPU_ACCOUNTING 2800 2801#ifdef CONFIG_IRQ_TIME_ACCOUNTING 2802/* 2803 * Account a tick to a process and cpustat 2804 * @p: the process that the cpu time gets accounted to 2805 * @user_tick: is the tick from userspace 2806 * @rq: the pointer to rq 2807 * 2808 * Tick demultiplexing follows the order 2809 * - pending hardirq update 2810 * - pending softirq update 2811 * - user_time 2812 * - idle_time 2813 * - system time 2814 * - check for guest_time 2815 * - else account as system_time 2816 * 2817 * Check for hardirq is done both for system and user time as there is 2818 * no timer going off while we are on hardirq and hence we may never get an 2819 * opportunity to update it solely in system time. 2820 * p->stime and friends are only updated on system time and not on irq 2821 * softirq as those do not count in task exec_runtime any more. 2822 */ 2823static void irqtime_account_process_tick(struct task_struct *p, int user_tick, 2824 struct rq *rq) 2825{ 2826 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy); 2827 u64 *cpustat = kcpustat_this_cpu->cpustat; 2828 2829 if (steal_account_process_tick()) 2830 return; 2831 2832 if (irqtime_account_hi_update()) { 2833 cpustat[CPUTIME_IRQ] += (__force u64) cputime_one_jiffy; 2834 } else if (irqtime_account_si_update()) { 2835 cpustat[CPUTIME_SOFTIRQ] += (__force u64) cputime_one_jiffy; 2836 } else if (this_cpu_ksoftirqd() == p) { 2837 /* 2838 * ksoftirqd time do not get accounted in cpu_softirq_time. 2839 * So, we have to handle it separately here. 2840 * Also, p->stime needs to be updated for ksoftirqd. 2841 */ 2842 __account_system_time(p, cputime_one_jiffy, one_jiffy_scaled, 2843 CPUTIME_SOFTIRQ); 2844 } else if (user_tick) { 2845 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled); 2846 } else if (p == rq->idle) { 2847 account_idle_time(cputime_one_jiffy); 2848 } else if (p->flags & PF_VCPU) { /* System time or guest time */ 2849 account_guest_time(p, cputime_one_jiffy, one_jiffy_scaled); 2850 } else { 2851 __account_system_time(p, cputime_one_jiffy, one_jiffy_scaled, 2852 CPUTIME_SYSTEM); 2853 } 2854} 2855 2856static void irqtime_account_idle_ticks(int ticks) 2857{ 2858 int i; 2859 struct rq *rq = this_rq(); 2860 2861 for (i = 0; i < ticks; i++) 2862 irqtime_account_process_tick(current, 0, rq); 2863} 2864#else /* CONFIG_IRQ_TIME_ACCOUNTING */ 2865static void irqtime_account_idle_ticks(int ticks) {} 2866static void irqtime_account_process_tick(struct task_struct *p, int user_tick, 2867 struct rq *rq) {} 2868#endif /* CONFIG_IRQ_TIME_ACCOUNTING */ 2869 2870/* 2871 * Account a single tick of cpu time. 2872 * @p: the process that the cpu time gets accounted to 2873 * @user_tick: indicates if the tick is a user or a system tick 2874 */ 2875void account_process_tick(struct task_struct *p, int user_tick) 2876{ 2877 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy); 2878 struct rq *rq = this_rq(); 2879 2880 if (sched_clock_irqtime) { 2881 irqtime_account_process_tick(p, user_tick, rq); 2882 return; 2883 } 2884 2885 if (steal_account_process_tick()) 2886 return; 2887 2888 if (user_tick) 2889 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled); 2890 else if ((p != rq->idle) || (irq_count() != HARDIRQ_OFFSET)) 2891 account_system_time(p, HARDIRQ_OFFSET, cputime_one_jiffy, 2892 one_jiffy_scaled); 2893 else 2894 account_idle_time(cputime_one_jiffy); 2895} 2896 2897/* 2898 * Account multiple ticks of steal time. 2899 * @p: the process from which the cpu time has been stolen 2900 * @ticks: number of stolen ticks 2901 */ 2902void account_steal_ticks(unsigned long ticks) 2903{ 2904 account_steal_time(jiffies_to_cputime(ticks)); 2905} 2906 2907/* 2908 * Account multiple ticks of idle time. 2909 * @ticks: number of stolen ticks 2910 */ 2911void account_idle_ticks(unsigned long ticks) 2912{ 2913 2914 if (sched_clock_irqtime) { 2915 irqtime_account_idle_ticks(ticks); 2916 return; 2917 } 2918 2919 account_idle_time(jiffies_to_cputime(ticks)); 2920} 2921 2922#endif 2923 2924/* 2925 * Use precise platform statistics if available: 2926 */ 2927#ifdef CONFIG_VIRT_CPU_ACCOUNTING 2928void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st) 2929{ 2930 *ut = p->utime; 2931 *st = p->stime; 2932} 2933 2934void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st) 2935{ 2936 struct task_cputime cputime; 2937 2938 thread_group_cputime(p, &cputime); 2939 2940 *ut = cputime.utime; 2941 *st = cputime.stime; 2942} 2943#else 2944 2945#ifndef nsecs_to_cputime 2946# define nsecs_to_cputime(__nsecs) nsecs_to_jiffies(__nsecs) 2947#endif 2948 2949void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st) 2950{ 2951 cputime_t rtime, utime = p->utime, total = utime + p->stime; 2952 2953 /* 2954 * Use CFS's precise accounting: 2955 */ 2956 rtime = nsecs_to_cputime(p->se.sum_exec_runtime); 2957 2958 if (total) { 2959 u64 temp = (__force u64) rtime; 2960 2961 temp *= (__force u64) utime; 2962 do_div(temp, (__force u32) total); 2963 utime = (__force cputime_t) temp; 2964 } else 2965 utime = rtime; 2966 2967 /* 2968 * Compare with previous values, to keep monotonicity: 2969 */ 2970 p->prev_utime = max(p->prev_utime, utime); 2971 p->prev_stime = max(p->prev_stime, rtime - p->prev_utime); 2972 2973 *ut = p->prev_utime; 2974 *st = p->prev_stime; 2975} 2976 2977/* 2978 * Must be called with siglock held. 2979 */ 2980void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st) 2981{ 2982 struct signal_struct *sig = p->signal; 2983 struct task_cputime cputime; 2984 cputime_t rtime, utime, total; 2985 2986 thread_group_cputime(p, &cputime); 2987 2988 total = cputime.utime + cputime.stime; 2989 rtime = nsecs_to_cputime(cputime.sum_exec_runtime); 2990 2991 if (total) { 2992 u64 temp = (__force u64) rtime; 2993 2994 temp *= (__force u64) cputime.utime; 2995 do_div(temp, (__force u32) total); 2996 utime = (__force cputime_t) temp; 2997 } else 2998 utime = rtime; 2999 3000 sig->prev_utime = max(sig->prev_utime, utime); 3001 sig->prev_stime = max(sig->prev_stime, rtime - sig->prev_utime); 3002 3003 *ut = sig->prev_utime; 3004 *st = sig->prev_stime; 3005} 3006#endif 3007 3008/* 3009 * This function gets called by the timer code, with HZ frequency. 3010 * We call it with interrupts disabled. 3011 */ 3012void scheduler_tick(void) 3013{ 3014 int cpu = smp_processor_id(); 3015 struct rq *rq = cpu_rq(cpu); 3016 struct task_struct *curr = rq->curr; 3017 3018 sched_clock_tick(); 3019 3020 raw_spin_lock(&rq->lock); 3021 update_rq_clock(rq); 3022 update_cpu_load_active(rq); 3023 curr->sched_class->task_tick(rq, curr, 0); 3024 raw_spin_unlock(&rq->lock); 3025 3026 perf_event_task_tick(); 3027 3028#ifdef CONFIG_SMP 3029 rq->idle_balance = idle_cpu(cpu); 3030 trigger_load_balance(rq, cpu); 3031#endif 3032} 3033 3034notrace unsigned long get_parent_ip(unsigned long addr) 3035{ 3036 if (in_lock_functions(addr)) { 3037 addr = CALLER_ADDR2; 3038 if (in_lock_functions(addr)) 3039 addr = CALLER_ADDR3; 3040 } 3041 return addr; 3042} 3043 3044#if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \ 3045 defined(CONFIG_PREEMPT_TRACER)) 3046 3047void __kprobes add_preempt_count(int val) 3048{ 3049#ifdef CONFIG_DEBUG_PREEMPT 3050 /* 3051 * Underflow? 3052 */ 3053 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0))) 3054 return; 3055#endif 3056 preempt_count() += val; 3057#ifdef CONFIG_DEBUG_PREEMPT 3058 /* 3059 * Spinlock count overflowing soon? 3060 */ 3061 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >= 3062 PREEMPT_MASK - 10); 3063#endif 3064 if (preempt_count() == val) 3065 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1)); 3066} 3067EXPORT_SYMBOL(add_preempt_count); 3068 3069void __kprobes sub_preempt_count(int val) 3070{ 3071#ifdef CONFIG_DEBUG_PREEMPT 3072 /* 3073 * Underflow? 3074 */ 3075 if (DEBUG_LOCKS_WARN_ON(val > preempt_count())) 3076 return; 3077 /* 3078 * Is the spinlock portion underflowing? 3079 */ 3080 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) && 3081 !(preempt_count() & PREEMPT_MASK))) 3082 return; 3083#endif 3084 3085 if (preempt_count() == val) 3086 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1)); 3087 preempt_count() -= val; 3088} 3089EXPORT_SYMBOL(sub_preempt_count); 3090 3091#endif 3092 3093/* 3094 * Print scheduling while atomic bug: 3095 */ 3096static noinline void __schedule_bug(struct task_struct *prev) 3097{ 3098 if (oops_in_progress) 3099 return; 3100 3101 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n", 3102 prev->comm, prev->pid, preempt_count()); 3103 3104 debug_show_held_locks(prev); 3105 print_modules(); 3106 if (irqs_disabled()) 3107 print_irqtrace_events(prev); 3108 dump_stack(); 3109} 3110 3111/* 3112 * Various schedule()-time debugging checks and statistics: 3113 */ 3114static inline void schedule_debug(struct task_struct *prev) 3115{ 3116 /* 3117 * Test if we are atomic. Since do_exit() needs to call into 3118 * schedule() atomically, we ignore that path for now. 3119 * Otherwise, whine if we are scheduling when we should not be. 3120 */ 3121 if (unlikely(in_atomic_preempt_off() && !prev->exit_state)) 3122 __schedule_bug(prev); 3123 rcu_sleep_check(); 3124 3125 profile_hit(SCHED_PROFILING, __builtin_return_address(0)); 3126 3127 schedstat_inc(this_rq(), sched_count); 3128} 3129 3130static void put_prev_task(struct rq *rq, struct task_struct *prev) 3131{ 3132 if (prev->on_rq || rq->skip_clock_update < 0) 3133 update_rq_clock(rq); 3134 prev->sched_class->put_prev_task(rq, prev); 3135} 3136 3137/* 3138 * Pick up the highest-prio task: 3139 */ 3140static inline struct task_struct * 3141pick_next_task(struct rq *rq) 3142{ 3143 const struct sched_class *class; 3144 struct task_struct *p; 3145 3146 /* 3147 * Optimization: we know that if all tasks are in 3148 * the fair class we can call that function directly: 3149 */ 3150 if (likely(rq->nr_running == rq->cfs.h_nr_running)) { 3151 p = fair_sched_class.pick_next_task(rq); 3152 if (likely(p)) 3153 return p; 3154 } 3155 3156 for_each_class(class) { 3157 p = class->pick_next_task(rq); 3158 if (p) 3159 return p; 3160 } 3161 3162 BUG(); /* the idle class will always have a runnable task */ 3163} 3164 3165/* 3166 * __schedule() is the main scheduler function. 3167 */ 3168static void __sched __schedule(void) 3169{ 3170 struct task_struct *prev, *next; 3171 unsigned long *switch_count; 3172 struct rq *rq; 3173 int cpu; 3174 3175need_resched: 3176 preempt_disable(); 3177 cpu = smp_processor_id(); 3178 rq = cpu_rq(cpu); 3179 rcu_note_context_switch(cpu); 3180 prev = rq->curr; 3181 3182 schedule_debug(prev); 3183 3184 if (sched_feat(HRTICK)) 3185 hrtick_clear(rq); 3186 3187 raw_spin_lock_irq(&rq->lock); 3188 3189 switch_count = &prev->nivcsw; 3190 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) { 3191 if (unlikely(signal_pending_state(prev->state, prev))) { 3192 prev->state = TASK_RUNNING; 3193 } else { 3194 deactivate_task(rq, prev, DEQUEUE_SLEEP); 3195 prev->on_rq = 0; 3196 3197 /* 3198 * If a worker went to sleep, notify and ask workqueue 3199 * whether it wants to wake up a task to maintain 3200 * concurrency. 3201 */ 3202 if (prev->flags & PF_WQ_WORKER) { 3203 struct task_struct *to_wakeup; 3204 3205 to_wakeup = wq_worker_sleeping(prev, cpu); 3206 if (to_wakeup) 3207 try_to_wake_up_local(to_wakeup); 3208 } 3209 } 3210 switch_count = &prev->nvcsw; 3211 } 3212 3213 pre_schedule(rq, prev); 3214 3215 if (unlikely(!rq->nr_running)) 3216 idle_balance(cpu, rq); 3217 3218 put_prev_task(rq, prev); 3219 next = pick_next_task(rq); 3220 clear_tsk_need_resched(prev); 3221 rq->skip_clock_update = 0; 3222 3223 if (likely(prev != next)) { 3224 rq->nr_switches++; 3225 rq->curr = next; 3226 ++*switch_count; 3227 3228 context_switch(rq, prev, next); /* unlocks the rq */ 3229 /* 3230 * The context switch have flipped the stack from under us 3231 * and restored the local variables which were saved when 3232 * this task called schedule() in the past. prev == current 3233 * is still correct, but it can be moved to another cpu/rq. 3234 */ 3235 cpu = smp_processor_id(); 3236 rq = cpu_rq(cpu); 3237 } else 3238 raw_spin_unlock_irq(&rq->lock); 3239 3240 post_schedule(rq); 3241 3242 sched_preempt_enable_no_resched(); 3243 if (need_resched()) 3244 goto need_resched; 3245} 3246 3247static inline void sched_submit_work(struct task_struct *tsk) 3248{ 3249 if (!tsk->state || tsk_is_pi_blocked(tsk)) 3250 return; 3251 /* 3252 * If we are going to sleep and we have plugged IO queued, 3253 * make sure to submit it to avoid deadlocks. 3254 */ 3255 if (blk_needs_flush_plug(tsk)) 3256 blk_schedule_flush_plug(tsk); 3257} 3258 3259asmlinkage void __sched schedule(void) 3260{ 3261 struct task_struct *tsk = current; 3262 3263 sched_submit_work(tsk); 3264 __schedule(); 3265} 3266EXPORT_SYMBOL(schedule); 3267 3268/** 3269 * schedule_preempt_disabled - called with preemption disabled 3270 * 3271 * Returns with preemption disabled. Note: preempt_count must be 1 3272 */ 3273void __sched schedule_preempt_disabled(void) 3274{ 3275 sched_preempt_enable_no_resched(); 3276 schedule(); 3277 preempt_disable(); 3278} 3279 3280#ifdef CONFIG_MUTEX_SPIN_ON_OWNER 3281 3282static inline bool owner_running(struct mutex *lock, struct task_struct *owner) 3283{ 3284 if (lock->owner != owner) 3285 return false; 3286 3287 /* 3288 * Ensure we emit the owner->on_cpu, dereference _after_ checking 3289 * lock->owner still matches owner, if that fails, owner might 3290 * point to free()d memory, if it still matches, the rcu_read_lock() 3291 * ensures the memory stays valid. 3292 */ 3293 barrier(); 3294 3295 return owner->on_cpu; 3296} 3297 3298/* 3299 * Look out! "owner" is an entirely speculative pointer 3300 * access and not reliable. 3301 */ 3302int mutex_spin_on_owner(struct mutex *lock, struct task_struct *owner) 3303{ 3304 if (!sched_feat(OWNER_SPIN)) 3305 return 0; 3306 3307 rcu_read_lock(); 3308 while (owner_running(lock, owner)) { 3309 if (need_resched()) 3310 break; 3311 3312 arch_mutex_cpu_relax(); 3313 } 3314 rcu_read_unlock(); 3315 3316 /* 3317 * We break out the loop above on need_resched() and when the 3318 * owner changed, which is a sign for heavy contention. Return 3319 * success only when lock->owner is NULL. 3320 */ 3321 return lock->owner == NULL; 3322} 3323#endif 3324 3325#ifdef CONFIG_PREEMPT 3326/* 3327 * this is the entry point to schedule() from in-kernel preemption 3328 * off of preempt_enable. Kernel preemptions off return from interrupt 3329 * occur there and call schedule directly. 3330 */ 3331asmlinkage void __sched notrace preempt_schedule(void) 3332{ 3333 struct thread_info *ti = current_thread_info(); 3334 3335 /* 3336 * If there is a non-zero preempt_count or interrupts are disabled, 3337 * we do not want to preempt the current task. Just return.. 3338 */ 3339 if (likely(ti->preempt_count || irqs_disabled())) 3340 return; 3341 3342 do { 3343 add_preempt_count_notrace(PREEMPT_ACTIVE); 3344 __schedule(); 3345 sub_preempt_count_notrace(PREEMPT_ACTIVE); 3346 3347 /* 3348 * Check again in case we missed a preemption opportunity 3349 * between schedule and now. 3350 */ 3351 barrier(); 3352 } while (need_resched()); 3353} 3354EXPORT_SYMBOL(preempt_schedule); 3355 3356/* 3357 * this is the entry point to schedule() from kernel preemption 3358 * off of irq context. 3359 * Note, that this is called and return with irqs disabled. This will 3360 * protect us against recursive calling from irq. 3361 */ 3362asmlinkage void __sched preempt_schedule_irq(void) 3363{ 3364 struct thread_info *ti = current_thread_info(); 3365 3366 /* Catch callers which need to be fixed */ 3367 BUG_ON(ti->preempt_count || !irqs_disabled()); 3368 3369 do { 3370 add_preempt_count(PREEMPT_ACTIVE); 3371 local_irq_enable(); 3372 __schedule(); 3373 local_irq_disable(); 3374 sub_preempt_count(PREEMPT_ACTIVE); 3375 3376 /* 3377 * Check again in case we missed a preemption opportunity 3378 * between schedule and now. 3379 */ 3380 barrier(); 3381 } while (need_resched()); 3382} 3383 3384#endif /* CONFIG_PREEMPT */ 3385 3386int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags, 3387 void *key) 3388{ 3389 return try_to_wake_up(curr->private, mode, wake_flags); 3390} 3391EXPORT_SYMBOL(default_wake_function); 3392 3393/* 3394 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just 3395 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve 3396 * number) then we wake all the non-exclusive tasks and one exclusive task. 3397 * 3398 * There are circumstances in which we can try to wake a task which has already 3399 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns 3400 * zero in this (rare) case, and we handle it by continuing to scan the queue. 3401 */ 3402static void __wake_up_common(wait_queue_head_t *q, unsigned int mode, 3403 int nr_exclusive, int wake_flags, void *key) 3404{ 3405 wait_queue_t *curr, *next; 3406 3407 list_for_each_entry_safe(curr, next, &q->task_list, task_list) { 3408 unsigned flags = curr->flags; 3409 3410 if (curr->func(curr, mode, wake_flags, key) && 3411 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive) 3412 break; 3413 } 3414} 3415 3416/** 3417 * __wake_up - wake up threads blocked on a waitqueue. 3418 * @q: the waitqueue 3419 * @mode: which threads 3420 * @nr_exclusive: how many wake-one or wake-many threads to wake up 3421 * @key: is directly passed to the wakeup function 3422 * 3423 * It may be assumed that this function implies a write memory barrier before 3424 * changing the task state if and only if any tasks are woken up. 3425 */ 3426void __wake_up(wait_queue_head_t *q, unsigned int mode, 3427 int nr_exclusive, void *key) 3428{ 3429 unsigned long flags; 3430 3431 spin_lock_irqsave(&q->lock, flags); 3432 __wake_up_common(q, mode, nr_exclusive, 0, key); 3433 spin_unlock_irqrestore(&q->lock, flags); 3434} 3435EXPORT_SYMBOL(__wake_up); 3436 3437/* 3438 * Same as __wake_up but called with the spinlock in wait_queue_head_t held. 3439 */ 3440void __wake_up_locked(wait_queue_head_t *q, unsigned int mode, int nr) 3441{ 3442 __wake_up_common(q, mode, nr, 0, NULL); 3443} 3444EXPORT_SYMBOL_GPL(__wake_up_locked); 3445 3446void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key) 3447{ 3448 __wake_up_common(q, mode, 1, 0, key); 3449} 3450EXPORT_SYMBOL_GPL(__wake_up_locked_key); 3451 3452/** 3453 * __wake_up_sync_key - wake up threads blocked on a waitqueue. 3454 * @q: the waitqueue 3455 * @mode: which threads 3456 * @nr_exclusive: how many wake-one or wake-many threads to wake up 3457 * @key: opaque value to be passed to wakeup targets 3458 * 3459 * The sync wakeup differs that the waker knows that it will schedule 3460 * away soon, so while the target thread will be woken up, it will not 3461 * be migrated to another CPU - ie. the two threads are 'synchronized' 3462 * with each other. This can prevent needless bouncing between CPUs. 3463 * 3464 * On UP it can prevent extra preemption. 3465 * 3466 * It may be assumed that this function implies a write memory barrier before 3467 * changing the task state if and only if any tasks are woken up. 3468 */ 3469void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode, 3470 int nr_exclusive, void *key) 3471{ 3472 unsigned long flags; 3473 int wake_flags = WF_SYNC; 3474 3475 if (unlikely(!q)) 3476 return; 3477 3478 if (unlikely(!nr_exclusive)) 3479 wake_flags = 0; 3480 3481 spin_lock_irqsave(&q->lock, flags); 3482 __wake_up_common(q, mode, nr_exclusive, wake_flags, key); 3483 spin_unlock_irqrestore(&q->lock, flags); 3484} 3485EXPORT_SYMBOL_GPL(__wake_up_sync_key); 3486 3487/* 3488 * __wake_up_sync - see __wake_up_sync_key() 3489 */ 3490void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive) 3491{ 3492 __wake_up_sync_key(q, mode, nr_exclusive, NULL); 3493} 3494EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */ 3495 3496/** 3497 * complete: - signals a single thread waiting on this completion 3498 * @x: holds the state of this particular completion 3499 * 3500 * This will wake up a single thread waiting on this completion. Threads will be 3501 * awakened in the same order in which they were queued. 3502 * 3503 * See also complete_all(), wait_for_completion() and related routines. 3504 * 3505 * It may be assumed that this function implies a write memory barrier before 3506 * changing the task state if and only if any tasks are woken up. 3507 */ 3508void complete(struct completion *x) 3509{ 3510 unsigned long flags; 3511 3512 spin_lock_irqsave(&x->wait.lock, flags); 3513 x->done++; 3514 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL); 3515 spin_unlock_irqrestore(&x->wait.lock, flags); 3516} 3517EXPORT_SYMBOL(complete); 3518 3519/** 3520 * complete_all: - signals all threads waiting on this completion 3521 * @x: holds the state of this particular completion 3522 * 3523 * This will wake up all threads waiting on this particular completion event. 3524 * 3525 * It may be assumed that this function implies a write memory barrier before 3526 * changing the task state if and only if any tasks are woken up. 3527 */ 3528void complete_all(struct completion *x) 3529{ 3530 unsigned long flags; 3531 3532 spin_lock_irqsave(&x->wait.lock, flags); 3533 x->done += UINT_MAX/2; 3534 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL); 3535 spin_unlock_irqrestore(&x->wait.lock, flags); 3536} 3537EXPORT_SYMBOL(complete_all); 3538 3539static inline long __sched 3540do_wait_for_common(struct completion *x, long timeout, int state) 3541{ 3542 if (!x->done) { 3543 DECLARE_WAITQUEUE(wait, current); 3544 3545 __add_wait_queue_tail_exclusive(&x->wait, &wait); 3546 do { 3547 if (signal_pending_state(state, current)) { 3548 timeout = -ERESTARTSYS; 3549 break; 3550 } 3551 __set_current_state(state); 3552 spin_unlock_irq(&x->wait.lock); 3553 timeout = schedule_timeout(timeout); 3554 spin_lock_irq(&x->wait.lock); 3555 } while (!x->done && timeout); 3556 __remove_wait_queue(&x->wait, &wait); 3557 if (!x->done) 3558 return timeout; 3559 } 3560 x->done--; 3561 return timeout ?: 1; 3562} 3563 3564static long __sched 3565wait_for_common(struct completion *x, long timeout, int state) 3566{ 3567 might_sleep(); 3568 3569 spin_lock_irq(&x->wait.lock); 3570 timeout = do_wait_for_common(x, timeout, state); 3571 spin_unlock_irq(&x->wait.lock); 3572 return timeout; 3573} 3574 3575/** 3576 * wait_for_completion: - waits for completion of a task 3577 * @x: holds the state of this particular completion 3578 * 3579 * This waits to be signaled for completion of a specific task. It is NOT 3580 * interruptible and there is no timeout. 3581 * 3582 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout 3583 * and interrupt capability. Also see complete(). 3584 */ 3585void __sched wait_for_completion(struct completion *x) 3586{ 3587 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE); 3588} 3589EXPORT_SYMBOL(wait_for_completion); 3590 3591/** 3592 * wait_for_completion_timeout: - waits for completion of a task (w/timeout) 3593 * @x: holds the state of this particular completion 3594 * @timeout: timeout value in jiffies 3595 * 3596 * This waits for either a completion of a specific task to be signaled or for a 3597 * specified timeout to expire. The timeout is in jiffies. It is not 3598 * interruptible. 3599 * 3600 * The return value is 0 if timed out, and positive (at least 1, or number of 3601 * jiffies left till timeout) if completed. 3602 */ 3603unsigned long __sched 3604wait_for_completion_timeout(struct completion *x, unsigned long timeout) 3605{ 3606 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE); 3607} 3608EXPORT_SYMBOL(wait_for_completion_timeout); 3609 3610/** 3611 * wait_for_completion_interruptible: - waits for completion of a task (w/intr) 3612 * @x: holds the state of this particular completion 3613 * 3614 * This waits for completion of a specific task to be signaled. It is 3615 * interruptible. 3616 * 3617 * The return value is -ERESTARTSYS if interrupted, 0 if completed. 3618 */ 3619int __sched wait_for_completion_interruptible(struct completion *x) 3620{ 3621 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE); 3622 if (t == -ERESTARTSYS) 3623 return t; 3624 return 0; 3625} 3626EXPORT_SYMBOL(wait_for_completion_interruptible); 3627 3628/** 3629 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr)) 3630 * @x: holds the state of this particular completion 3631 * @timeout: timeout value in jiffies 3632 * 3633 * This waits for either a completion of a specific task to be signaled or for a 3634 * specified timeout to expire. It is interruptible. The timeout is in jiffies. 3635 * 3636 * The return value is -ERESTARTSYS if interrupted, 0 if timed out, 3637 * positive (at least 1, or number of jiffies left till timeout) if completed. 3638 */ 3639long __sched 3640wait_for_completion_interruptible_timeout(struct completion *x, 3641 unsigned long timeout) 3642{ 3643 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE); 3644} 3645EXPORT_SYMBOL(wait_for_completion_interruptible_timeout); 3646 3647/** 3648 * wait_for_completion_killable: - waits for completion of a task (killable) 3649 * @x: holds the state of this particular completion 3650 * 3651 * This waits to be signaled for completion of a specific task. It can be 3652 * interrupted by a kill signal. 3653 * 3654 * The return value is -ERESTARTSYS if interrupted, 0 if completed. 3655 */ 3656int __sched wait_for_completion_killable(struct completion *x) 3657{ 3658 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE); 3659 if (t == -ERESTARTSYS) 3660 return t; 3661 return 0; 3662} 3663EXPORT_SYMBOL(wait_for_completion_killable); 3664 3665/** 3666 * wait_for_completion_killable_timeout: - waits for completion of a task (w/(to,killable)) 3667 * @x: holds the state of this particular completion 3668 * @timeout: timeout value in jiffies 3669 * 3670 * This waits for either a completion of a specific task to be 3671 * signaled or for a specified timeout to expire. It can be 3672 * interrupted by a kill signal. The timeout is in jiffies. 3673 * 3674 * The return value is -ERESTARTSYS if interrupted, 0 if timed out, 3675 * positive (at least 1, or number of jiffies left till timeout) if completed. 3676 */ 3677long __sched 3678wait_for_completion_killable_timeout(struct completion *x, 3679 unsigned long timeout) 3680{ 3681 return wait_for_common(x, timeout, TASK_KILLABLE); 3682} 3683EXPORT_SYMBOL(wait_for_completion_killable_timeout); 3684 3685/** 3686 * try_wait_for_completion - try to decrement a completion without blocking 3687 * @x: completion structure 3688 * 3689 * Returns: 0 if a decrement cannot be done without blocking 3690 * 1 if a decrement succeeded. 3691 * 3692 * If a completion is being used as a counting completion, 3693 * attempt to decrement the counter without blocking. This 3694 * enables us to avoid waiting if the resource the completion 3695 * is protecting is not available. 3696 */ 3697bool try_wait_for_completion(struct completion *x) 3698{ 3699 unsigned long flags; 3700 int ret = 1; 3701 3702 spin_lock_irqsave(&x->wait.lock, flags); 3703 if (!x->done) 3704 ret = 0; 3705 else 3706 x->done--; 3707 spin_unlock_irqrestore(&x->wait.lock, flags); 3708 return ret; 3709} 3710EXPORT_SYMBOL(try_wait_for_completion); 3711 3712/** 3713 * completion_done - Test to see if a completion has any waiters 3714 * @x: completion structure 3715 * 3716 * Returns: 0 if there are waiters (wait_for_completion() in progress) 3717 * 1 if there are no waiters. 3718 * 3719 */ 3720bool completion_done(struct completion *x) 3721{ 3722 unsigned long flags; 3723 int ret = 1; 3724 3725 spin_lock_irqsave(&x->wait.lock, flags); 3726 if (!x->done) 3727 ret = 0; 3728 spin_unlock_irqrestore(&x->wait.lock, flags); 3729 return ret; 3730} 3731EXPORT_SYMBOL(completion_done); 3732 3733static long __sched 3734sleep_on_common(wait_queue_head_t *q, int state, long timeout) 3735{ 3736 unsigned long flags; 3737 wait_queue_t wait; 3738 3739 init_waitqueue_entry(&wait, current); 3740 3741 __set_current_state(state); 3742 3743 spin_lock_irqsave(&q->lock, flags); 3744 __add_wait_queue(q, &wait); 3745 spin_unlock(&q->lock); 3746 timeout = schedule_timeout(timeout); 3747 spin_lock_irq(&q->lock); 3748 __remove_wait_queue(q, &wait); 3749 spin_unlock_irqrestore(&q->lock, flags); 3750 3751 return timeout; 3752} 3753 3754void __sched interruptible_sleep_on(wait_queue_head_t *q) 3755{ 3756 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT); 3757} 3758EXPORT_SYMBOL(interruptible_sleep_on); 3759 3760long __sched 3761interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout) 3762{ 3763 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout); 3764} 3765EXPORT_SYMBOL(interruptible_sleep_on_timeout); 3766 3767void __sched sleep_on(wait_queue_head_t *q) 3768{ 3769 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT); 3770} 3771EXPORT_SYMBOL(sleep_on); 3772 3773long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout) 3774{ 3775 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout); 3776} 3777EXPORT_SYMBOL(sleep_on_timeout); 3778 3779#ifdef CONFIG_RT_MUTEXES 3780 3781/* 3782 * rt_mutex_setprio - set the current priority of a task 3783 * @p: task 3784 * @prio: prio value (kernel-internal form) 3785 * 3786 * This function changes the 'effective' priority of a task. It does 3787 * not touch ->normal_prio like __setscheduler(). 3788 * 3789 * Used by the rt_mutex code to implement priority inheritance logic. 3790 */ 3791void rt_mutex_setprio(struct task_struct *p, int prio) 3792{ 3793 int oldprio, on_rq, running; 3794 struct rq *rq; 3795 const struct sched_class *prev_class; 3796 3797 BUG_ON(prio < 0 || prio > MAX_PRIO); 3798 3799 rq = __task_rq_lock(p); 3800 3801 /* 3802 * Idle task boosting is a nono in general. There is one 3803 * exception, when PREEMPT_RT and NOHZ is active: 3804 * 3805 * The idle task calls get_next_timer_interrupt() and holds 3806 * the timer wheel base->lock on the CPU and another CPU wants 3807 * to access the timer (probably to cancel it). We can safely 3808 * ignore the boosting request, as the idle CPU runs this code 3809 * with interrupts disabled and will complete the lock 3810 * protected section without being interrupted. So there is no 3811 * real need to boost. 3812 */ 3813 if (unlikely(p == rq->idle)) { 3814 WARN_ON(p != rq->curr); 3815 WARN_ON(p->pi_blocked_on); 3816 goto out_unlock; 3817 } 3818 3819 trace_sched_pi_setprio(p, prio); 3820 oldprio = p->prio; 3821 prev_class = p->sched_class; 3822 on_rq = p->on_rq; 3823 running = task_current(rq, p); 3824 if (on_rq) 3825 dequeue_task(rq, p, 0); 3826 if (running) 3827 p->sched_class->put_prev_task(rq, p); 3828 3829 if (rt_prio(prio)) 3830 p->sched_class = &rt_sched_class; 3831 else 3832 p->sched_class = &fair_sched_class; 3833 3834 p->prio = prio; 3835 3836 if (running) 3837 p->sched_class->set_curr_task(rq); 3838 if (on_rq) 3839 enqueue_task(rq, p, oldprio < prio ? ENQUEUE_HEAD : 0); 3840 3841 check_class_changed(rq, p, prev_class, oldprio); 3842out_unlock: 3843 __task_rq_unlock(rq); 3844} 3845#endif 3846void set_user_nice(struct task_struct *p, long nice) 3847{ 3848 int old_prio, delta, on_rq; 3849 unsigned long flags; 3850 struct rq *rq; 3851 3852 if (TASK_NICE(p) == nice || nice < -20 || nice > 19) 3853 return; 3854 /* 3855 * We have to be careful, if called from sys_setpriority(), 3856 * the task might be in the middle of scheduling on another CPU. 3857 */ 3858 rq = task_rq_lock(p, &flags); 3859 /* 3860 * The RT priorities are set via sched_setscheduler(), but we still 3861 * allow the 'normal' nice value to be set - but as expected 3862 * it wont have any effect on scheduling until the task is 3863 * SCHED_FIFO/SCHED_RR: 3864 */ 3865 if (task_has_rt_policy(p)) { 3866 p->static_prio = NICE_TO_PRIO(nice); 3867 goto out_unlock; 3868 } 3869 on_rq = p->on_rq; 3870 if (on_rq) 3871 dequeue_task(rq, p, 0); 3872 3873 p->static_prio = NICE_TO_PRIO(nice); 3874 set_load_weight(p); 3875 old_prio = p->prio; 3876 p->prio = effective_prio(p); 3877 delta = p->prio - old_prio; 3878 3879 if (on_rq) { 3880 enqueue_task(rq, p, 0); 3881 /* 3882 * If the task increased its priority or is running and 3883 * lowered its priority, then reschedule its CPU: 3884 */ 3885 if (delta < 0 || (delta > 0 && task_running(rq, p))) 3886 resched_task(rq->curr); 3887 } 3888out_unlock: 3889 task_rq_unlock(rq, p, &flags); 3890} 3891EXPORT_SYMBOL(set_user_nice); 3892 3893/* 3894 * can_nice - check if a task can reduce its nice value 3895 * @p: task 3896 * @nice: nice value 3897 */ 3898int can_nice(const struct task_struct *p, const int nice) 3899{ 3900 /* convert nice value [19,-20] to rlimit style value [1,40] */ 3901 int nice_rlim = 20 - nice; 3902 3903 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) || 3904 capable(CAP_SYS_NICE)); 3905} 3906 3907#ifdef __ARCH_WANT_SYS_NICE 3908 3909/* 3910 * sys_nice - change the priority of the current process. 3911 * @increment: priority increment 3912 * 3913 * sys_setpriority is a more generic, but much slower function that 3914 * does similar things. 3915 */ 3916SYSCALL_DEFINE1(nice, int, increment) 3917{ 3918 long nice, retval; 3919 3920 /* 3921 * Setpriority might change our priority at the same moment. 3922 * We don't have to worry. Conceptually one call occurs first 3923 * and we have a single winner. 3924 */ 3925 if (increment < -40) 3926 increment = -40; 3927 if (increment > 40) 3928 increment = 40; 3929 3930 nice = TASK_NICE(current) + increment; 3931 if (nice < -20) 3932 nice = -20; 3933 if (nice > 19) 3934 nice = 19; 3935 3936 if (increment < 0 && !can_nice(current, nice)) 3937 return -EPERM; 3938 3939 retval = security_task_setnice(current, nice); 3940 if (retval) 3941 return retval; 3942 3943 set_user_nice(current, nice); 3944 return 0; 3945} 3946 3947#endif 3948 3949/** 3950 * task_prio - return the priority value of a given task. 3951 * @p: the task in question. 3952 * 3953 * This is the priority value as seen by users in /proc. 3954 * RT tasks are offset by -200. Normal tasks are centered 3955 * around 0, value goes from -16 to +15. 3956 */ 3957int task_prio(const struct task_struct *p) 3958{ 3959 return p->prio - MAX_RT_PRIO; 3960} 3961 3962/** 3963 * task_nice - return the nice value of a given task. 3964 * @p: the task in question. 3965 */ 3966int task_nice(const struct task_struct *p) 3967{ 3968 return TASK_NICE(p); 3969} 3970EXPORT_SYMBOL(task_nice); 3971 3972/** 3973 * idle_cpu - is a given cpu idle currently? 3974 * @cpu: the processor in question. 3975 */ 3976int idle_cpu(int cpu) 3977{ 3978 struct rq *rq = cpu_rq(cpu); 3979 3980 if (rq->curr != rq->idle) 3981 return 0; 3982 3983 if (rq->nr_running) 3984 return 0; 3985 3986#ifdef CONFIG_SMP 3987 if (!llist_empty(&rq->wake_list)) 3988 return 0; 3989#endif 3990 3991 return 1; 3992} 3993 3994/** 3995 * idle_task - return the idle task for a given cpu. 3996 * @cpu: the processor in question. 3997 */ 3998struct task_struct *idle_task(int cpu) 3999{ 4000 return cpu_rq(cpu)->idle; 4001} 4002 4003/** 4004 * find_process_by_pid - find a process with a matching PID value. 4005 * @pid: the pid in question. 4006 */ 4007static struct task_struct *find_process_by_pid(pid_t pid) 4008{ 4009 return pid ? find_task_by_vpid(pid) : current; 4010} 4011 4012/* Actually do priority change: must hold rq lock. */ 4013static void 4014__setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio) 4015{ 4016 p->policy = policy; 4017 p->rt_priority = prio; 4018 p->normal_prio = normal_prio(p); 4019 /* we are holding p->pi_lock already */ 4020 p->prio = rt_mutex_getprio(p); 4021 if (rt_prio(p->prio)) 4022 p->sched_class = &rt_sched_class; 4023 else 4024 p->sched_class = &fair_sched_class; 4025 set_load_weight(p); 4026} 4027 4028/* 4029 * check the target process has a UID that matches the current process's 4030 */ 4031static bool check_same_owner(struct task_struct *p) 4032{ 4033 const struct cred *cred = current_cred(), *pcred; 4034 bool match; 4035 4036 rcu_read_lock(); 4037 pcred = __task_cred(p); 4038 if (cred->user->user_ns == pcred->user->user_ns) 4039 match = (cred->euid == pcred->euid || 4040 cred->euid == pcred->uid); 4041 else 4042 match = false; 4043 rcu_read_unlock(); 4044 return match; 4045} 4046 4047static int __sched_setscheduler(struct task_struct *p, int policy, 4048 const struct sched_param *param, bool user) 4049{ 4050 int retval, oldprio, oldpolicy = -1, on_rq, running; 4051 unsigned long flags; 4052 const struct sched_class *prev_class; 4053 struct rq *rq; 4054 int reset_on_fork; 4055 4056 /* may grab non-irq protected spin_locks */ 4057 BUG_ON(in_interrupt()); 4058recheck: 4059 /* double check policy once rq lock held */ 4060 if (policy < 0) { 4061 reset_on_fork = p->sched_reset_on_fork; 4062 policy = oldpolicy = p->policy; 4063 } else { 4064 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK); 4065 policy &= ~SCHED_RESET_ON_FORK; 4066 4067 if (policy != SCHED_FIFO && policy != SCHED_RR && 4068 policy != SCHED_NORMAL && policy != SCHED_BATCH && 4069 policy != SCHED_IDLE) 4070 return -EINVAL; 4071 } 4072 4073 /* 4074 * Valid priorities for SCHED_FIFO and SCHED_RR are 4075 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL, 4076 * SCHED_BATCH and SCHED_IDLE is 0. 4077 */ 4078 if (param->sched_priority < 0 || 4079 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) || 4080 (!p->mm && param->sched_priority > MAX_RT_PRIO-1)) 4081 return -EINVAL; 4082 if (rt_policy(policy) != (param->sched_priority != 0)) 4083 return -EINVAL; 4084 4085 /* 4086 * Allow unprivileged RT tasks to decrease priority: 4087 */ 4088 if (user && !capable(CAP_SYS_NICE)) { 4089 if (rt_policy(policy)) { 4090 unsigned long rlim_rtprio = 4091 task_rlimit(p, RLIMIT_RTPRIO); 4092 4093 /* can't set/change the rt policy */ 4094 if (policy != p->policy && !rlim_rtprio) 4095 return -EPERM; 4096 4097 /* can't increase priority */ 4098 if (param->sched_priority > p->rt_priority && 4099 param->sched_priority > rlim_rtprio) 4100 return -EPERM; 4101 } 4102 4103 /* 4104 * Treat SCHED_IDLE as nice 20. Only allow a switch to 4105 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it. 4106 */ 4107 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) { 4108 if (!can_nice(p, TASK_NICE(p))) 4109 return -EPERM; 4110 } 4111 4112 /* can't change other user's priorities */ 4113 if (!check_same_owner(p)) 4114 return -EPERM; 4115 4116 /* Normal users shall not reset the sched_reset_on_fork flag */ 4117 if (p->sched_reset_on_fork && !reset_on_fork) 4118 return -EPERM; 4119 } 4120 4121 if (user) { 4122 retval = security_task_setscheduler(p); 4123 if (retval) 4124 return retval; 4125 } 4126 4127 /* 4128 * make sure no PI-waiters arrive (or leave) while we are 4129 * changing the priority of the task: 4130 * 4131 * To be able to change p->policy safely, the appropriate 4132 * runqueue lock must be held. 4133 */ 4134 rq = task_rq_lock(p, &flags); 4135 4136 /* 4137 * Changing the policy of the stop threads its a very bad idea 4138 */ 4139 if (p == rq->stop) { 4140 task_rq_unlock(rq, p, &flags); 4141 return -EINVAL; 4142 } 4143 4144 /* 4145 * If not changing anything there's no need to proceed further: 4146 */ 4147 if (unlikely(policy == p->policy && (!rt_policy(policy) || 4148 param->sched_priority == p->rt_priority))) { 4149 4150 __task_rq_unlock(rq); 4151 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 4152 return 0; 4153 } 4154 4155#ifdef CONFIG_RT_GROUP_SCHED 4156 if (user) { 4157 /* 4158 * Do not allow realtime tasks into groups that have no runtime 4159 * assigned. 4160 */ 4161 if (rt_bandwidth_enabled() && rt_policy(policy) && 4162 task_group(p)->rt_bandwidth.rt_runtime == 0 && 4163 !task_group_is_autogroup(task_group(p))) { 4164 task_rq_unlock(rq, p, &flags); 4165 return -EPERM; 4166 } 4167 } 4168#endif 4169 4170 /* recheck policy now with rq lock held */ 4171 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) { 4172 policy = oldpolicy = -1; 4173 task_rq_unlock(rq, p, &flags); 4174 goto recheck; 4175 } 4176 on_rq = p->on_rq; 4177 running = task_current(rq, p); 4178 if (on_rq) 4179 dequeue_task(rq, p, 0); 4180 if (running) 4181 p->sched_class->put_prev_task(rq, p); 4182 4183 p->sched_reset_on_fork = reset_on_fork; 4184 4185 oldprio = p->prio; 4186 prev_class = p->sched_class; 4187 __setscheduler(rq, p, policy, param->sched_priority); 4188 4189 if (running) 4190 p->sched_class->set_curr_task(rq); 4191 if (on_rq) 4192 enqueue_task(rq, p, 0); 4193 4194 check_class_changed(rq, p, prev_class, oldprio); 4195 task_rq_unlock(rq, p, &flags); 4196 4197 rt_mutex_adjust_pi(p); 4198 4199 return 0; 4200} 4201 4202/** 4203 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread. 4204 * @p: the task in question. 4205 * @policy: new policy. 4206 * @param: structure containing the new RT priority. 4207 * 4208 * NOTE that the task may be already dead. 4209 */ 4210int sched_setscheduler(struct task_struct *p, int policy, 4211 const struct sched_param *param) 4212{ 4213 return __sched_setscheduler(p, policy, param, true); 4214} 4215EXPORT_SYMBOL_GPL(sched_setscheduler); 4216 4217/** 4218 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace. 4219 * @p: the task in question. 4220 * @policy: new policy. 4221 * @param: structure containing the new RT priority. 4222 * 4223 * Just like sched_setscheduler, only don't bother checking if the 4224 * current context has permission. For example, this is needed in 4225 * stop_machine(): we create temporary high priority worker threads, 4226 * but our caller might not have that capability. 4227 */ 4228int sched_setscheduler_nocheck(struct task_struct *p, int policy, 4229 const struct sched_param *param) 4230{ 4231 return __sched_setscheduler(p, policy, param, false); 4232} 4233 4234static int 4235do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param) 4236{ 4237 struct sched_param lparam; 4238 struct task_struct *p; 4239 int retval; 4240 4241 if (!param || pid < 0) 4242 return -EINVAL; 4243 if (copy_from_user(&lparam, param, sizeof(struct sched_param))) 4244 return -EFAULT; 4245 4246 rcu_read_lock(); 4247 retval = -ESRCH; 4248 p = find_process_by_pid(pid); 4249 if (p != NULL) 4250 retval = sched_setscheduler(p, policy, &lparam); 4251 rcu_read_unlock(); 4252 4253 return retval; 4254} 4255 4256/** 4257 * sys_sched_setscheduler - set/change the scheduler policy and RT priority 4258 * @pid: the pid in question. 4259 * @policy: new policy. 4260 * @param: structure containing the new RT priority. 4261 */ 4262SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy, 4263 struct sched_param __user *, param) 4264{ 4265 /* negative values for policy are not valid */ 4266 if (policy < 0) 4267 return -EINVAL; 4268 4269 return do_sched_setscheduler(pid, policy, param); 4270} 4271 4272/** 4273 * sys_sched_setparam - set/change the RT priority of a thread 4274 * @pid: the pid in question. 4275 * @param: structure containing the new RT priority. 4276 */ 4277SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param) 4278{ 4279 return do_sched_setscheduler(pid, -1, param); 4280} 4281 4282/** 4283 * sys_sched_getscheduler - get the policy (scheduling class) of a thread 4284 * @pid: the pid in question. 4285 */ 4286SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid) 4287{ 4288 struct task_struct *p; 4289 int retval; 4290 4291 if (pid < 0) 4292 return -EINVAL; 4293 4294 retval = -ESRCH; 4295 rcu_read_lock(); 4296 p = find_process_by_pid(pid); 4297 if (p) { 4298 retval = security_task_getscheduler(p); 4299 if (!retval) 4300 retval = p->policy 4301 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0); 4302 } 4303 rcu_read_unlock(); 4304 return retval; 4305} 4306 4307/** 4308 * sys_sched_getparam - get the RT priority of a thread 4309 * @pid: the pid in question. 4310 * @param: structure containing the RT priority. 4311 */ 4312SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param) 4313{ 4314 struct sched_param lp; 4315 struct task_struct *p; 4316 int retval; 4317 4318 if (!param || pid < 0) 4319 return -EINVAL; 4320 4321 rcu_read_lock(); 4322 p = find_process_by_pid(pid); 4323 retval = -ESRCH; 4324 if (!p) 4325 goto out_unlock; 4326 4327 retval = security_task_getscheduler(p); 4328 if (retval) 4329 goto out_unlock; 4330 4331 lp.sched_priority = p->rt_priority; 4332 rcu_read_unlock(); 4333 4334 /* 4335 * This one might sleep, we cannot do it with a spinlock held ... 4336 */ 4337 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0; 4338 4339 return retval; 4340 4341out_unlock: 4342 rcu_read_unlock(); 4343 return retval; 4344} 4345 4346long sched_setaffinity(pid_t pid, const struct cpumask *in_mask) 4347{ 4348 cpumask_var_t cpus_allowed, new_mask; 4349 struct task_struct *p; 4350 int retval; 4351 4352 get_online_cpus(); 4353 rcu_read_lock(); 4354 4355 p = find_process_by_pid(pid); 4356 if (!p) { 4357 rcu_read_unlock(); 4358 put_online_cpus(); 4359 return -ESRCH; 4360 } 4361 4362 /* Prevent p going away */ 4363 get_task_struct(p); 4364 rcu_read_unlock(); 4365 4366 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) { 4367 retval = -ENOMEM; 4368 goto out_put_task; 4369 } 4370 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) { 4371 retval = -ENOMEM; 4372 goto out_free_cpus_allowed; 4373 } 4374 retval = -EPERM; 4375 if (!check_same_owner(p) && !ns_capable(task_user_ns(p), CAP_SYS_NICE)) 4376 goto out_unlock; 4377 4378 retval = security_task_setscheduler(p); 4379 if (retval) 4380 goto out_unlock; 4381 4382 cpuset_cpus_allowed(p, cpus_allowed); 4383 cpumask_and(new_mask, in_mask, cpus_allowed); 4384again: 4385 retval = set_cpus_allowed_ptr(p, new_mask); 4386 4387 if (!retval) { 4388 cpuset_cpus_allowed(p, cpus_allowed); 4389 if (!cpumask_subset(new_mask, cpus_allowed)) { 4390 /* 4391 * We must have raced with a concurrent cpuset 4392 * update. Just reset the cpus_allowed to the 4393 * cpuset's cpus_allowed 4394 */ 4395 cpumask_copy(new_mask, cpus_allowed); 4396 goto again; 4397 } 4398 } 4399out_unlock: 4400 free_cpumask_var(new_mask); 4401out_free_cpus_allowed: 4402 free_cpumask_var(cpus_allowed); 4403out_put_task: 4404 put_task_struct(p); 4405 put_online_cpus(); 4406 return retval; 4407} 4408 4409static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len, 4410 struct cpumask *new_mask) 4411{ 4412 if (len < cpumask_size()) 4413 cpumask_clear(new_mask); 4414 else if (len > cpumask_size()) 4415 len = cpumask_size(); 4416 4417 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0; 4418} 4419 4420/** 4421 * sys_sched_setaffinity - set the cpu affinity of a process 4422 * @pid: pid of the process 4423 * @len: length in bytes of the bitmask pointed to by user_mask_ptr 4424 * @user_mask_ptr: user-space pointer to the new cpu mask 4425 */ 4426SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len, 4427 unsigned long __user *, user_mask_ptr) 4428{ 4429 cpumask_var_t new_mask; 4430 int retval; 4431 4432 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) 4433 return -ENOMEM; 4434 4435 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask); 4436 if (retval == 0) 4437 retval = sched_setaffinity(pid, new_mask); 4438 free_cpumask_var(new_mask); 4439 return retval; 4440} 4441 4442long sched_getaffinity(pid_t pid, struct cpumask *mask) 4443{ 4444 struct task_struct *p; 4445 unsigned long flags; 4446 int retval; 4447 4448 get_online_cpus(); 4449 rcu_read_lock(); 4450 4451 retval = -ESRCH; 4452 p = find_process_by_pid(pid); 4453 if (!p) 4454 goto out_unlock; 4455 4456 retval = security_task_getscheduler(p); 4457 if (retval) 4458 goto out_unlock; 4459 4460 raw_spin_lock_irqsave(&p->pi_lock, flags); 4461 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask); 4462 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 4463 4464out_unlock: 4465 rcu_read_unlock(); 4466 put_online_cpus(); 4467 4468 return retval; 4469} 4470 4471/** 4472 * sys_sched_getaffinity - get the cpu affinity of a process 4473 * @pid: pid of the process 4474 * @len: length in bytes of the bitmask pointed to by user_mask_ptr 4475 * @user_mask_ptr: user-space pointer to hold the current cpu mask 4476 */ 4477SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len, 4478 unsigned long __user *, user_mask_ptr) 4479{ 4480 int ret; 4481 cpumask_var_t mask; 4482 4483 if ((len * BITS_PER_BYTE) < nr_cpu_ids) 4484 return -EINVAL; 4485 if (len & (sizeof(unsigned long)-1)) 4486 return -EINVAL; 4487 4488 if (!alloc_cpumask_var(&mask, GFP_KERNEL)) 4489 return -ENOMEM; 4490 4491 ret = sched_getaffinity(pid, mask); 4492 if (ret == 0) { 4493 size_t retlen = min_t(size_t, len, cpumask_size()); 4494 4495 if (copy_to_user(user_mask_ptr, mask, retlen)) 4496 ret = -EFAULT; 4497 else 4498 ret = retlen; 4499 } 4500 free_cpumask_var(mask); 4501 4502 return ret; 4503} 4504 4505/** 4506 * sys_sched_yield - yield the current processor to other threads. 4507 * 4508 * This function yields the current CPU to other tasks. If there are no 4509 * other threads running on this CPU then this function will return. 4510 */ 4511SYSCALL_DEFINE0(sched_yield) 4512{ 4513 struct rq *rq = this_rq_lock(); 4514 4515 schedstat_inc(rq, yld_count); 4516 current->sched_class->yield_task(rq); 4517 4518 /* 4519 * Since we are going to call schedule() anyway, there's 4520 * no need to preempt or enable interrupts: 4521 */ 4522 __release(rq->lock); 4523 spin_release(&rq->lock.dep_map, 1, _THIS_IP_); 4524 do_raw_spin_unlock(&rq->lock); 4525 sched_preempt_enable_no_resched(); 4526 4527 schedule(); 4528 4529 return 0; 4530} 4531 4532static inline int should_resched(void) 4533{ 4534 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE); 4535} 4536 4537static void __cond_resched(void) 4538{ 4539 add_preempt_count(PREEMPT_ACTIVE); 4540 __schedule(); 4541 sub_preempt_count(PREEMPT_ACTIVE); 4542} 4543 4544int __sched _cond_resched(void) 4545{ 4546 if (should_resched()) { 4547 __cond_resched(); 4548 return 1; 4549 } 4550 return 0; 4551} 4552EXPORT_SYMBOL(_cond_resched); 4553 4554/* 4555 * __cond_resched_lock() - if a reschedule is pending, drop the given lock, 4556 * call schedule, and on return reacquire the lock. 4557 * 4558 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level 4559 * operations here to prevent schedule() from being called twice (once via 4560 * spin_unlock(), once by hand). 4561 */ 4562int __cond_resched_lock(spinlock_t *lock) 4563{ 4564 int resched = should_resched(); 4565 int ret = 0; 4566 4567 lockdep_assert_held(lock); 4568 4569 if (spin_needbreak(lock) || resched) { 4570 spin_unlock(lock); 4571 if (resched) 4572 __cond_resched(); 4573 else 4574 cpu_relax(); 4575 ret = 1; 4576 spin_lock(lock); 4577 } 4578 return ret; 4579} 4580EXPORT_SYMBOL(__cond_resched_lock); 4581 4582int __sched __cond_resched_softirq(void) 4583{ 4584 BUG_ON(!in_softirq()); 4585 4586 if (should_resched()) { 4587 local_bh_enable(); 4588 __cond_resched(); 4589 local_bh_disable(); 4590 return 1; 4591 } 4592 return 0; 4593} 4594EXPORT_SYMBOL(__cond_resched_softirq); 4595 4596/** 4597 * yield - yield the current processor to other threads. 4598 * 4599 * Do not ever use this function, there's a 99% chance you're doing it wrong. 4600 * 4601 * The scheduler is at all times free to pick the calling task as the most 4602 * eligible task to run, if removing the yield() call from your code breaks 4603 * it, its already broken. 4604 * 4605 * Typical broken usage is: 4606 * 4607 * while (!event) 4608 * yield(); 4609 * 4610 * where one assumes that yield() will let 'the other' process run that will 4611 * make event true. If the current task is a SCHED_FIFO task that will never 4612 * happen. Never use yield() as a progress guarantee!! 4613 * 4614 * If you want to use yield() to wait for something, use wait_event(). 4615 * If you want to use yield() to be 'nice' for others, use cond_resched(). 4616 * If you still want to use yield(), do not! 4617 */ 4618void __sched yield(void) 4619{ 4620 set_current_state(TASK_RUNNING); 4621 sys_sched_yield(); 4622} 4623EXPORT_SYMBOL(yield); 4624 4625/** 4626 * yield_to - yield the current processor to another thread in 4627 * your thread group, or accelerate that thread toward the 4628 * processor it's on. 4629 * @p: target task 4630 * @preempt: whether task preemption is allowed or not 4631 * 4632 * It's the caller's job to ensure that the target task struct 4633 * can't go away on us before we can do any checks. 4634 * 4635 * Returns true if we indeed boosted the target task. 4636 */ 4637bool __sched yield_to(struct task_struct *p, bool preempt) 4638{ 4639 struct task_struct *curr = current; 4640 struct rq *rq, *p_rq; 4641 unsigned long flags; 4642 bool yielded = 0; 4643 4644 local_irq_save(flags); 4645 rq = this_rq(); 4646 4647again: 4648 p_rq = task_rq(p); 4649 double_rq_lock(rq, p_rq); 4650 while (task_rq(p) != p_rq) { 4651 double_rq_unlock(rq, p_rq); 4652 goto again; 4653 } 4654 4655 if (!curr->sched_class->yield_to_task) 4656 goto out; 4657 4658 if (curr->sched_class != p->sched_class) 4659 goto out; 4660 4661 if (task_running(p_rq, p) || p->state) 4662 goto out; 4663 4664 yielded = curr->sched_class->yield_to_task(rq, p, preempt); 4665 if (yielded) { 4666 schedstat_inc(rq, yld_count); 4667 /* 4668 * Make p's CPU reschedule; pick_next_entity takes care of 4669 * fairness. 4670 */ 4671 if (preempt && rq != p_rq) 4672 resched_task(p_rq->curr); 4673 } else { 4674 /* 4675 * We might have set it in task_yield_fair(), but are 4676 * not going to schedule(), so don't want to skip 4677 * the next update. 4678 */ 4679 rq->skip_clock_update = 0; 4680 } 4681 4682out: 4683 double_rq_unlock(rq, p_rq); 4684 local_irq_restore(flags); 4685 4686 if (yielded) 4687 schedule(); 4688 4689 return yielded; 4690} 4691EXPORT_SYMBOL_GPL(yield_to); 4692 4693/* 4694 * This task is about to go to sleep on IO. Increment rq->nr_iowait so 4695 * that process accounting knows that this is a task in IO wait state. 4696 */ 4697void __sched io_schedule(void) 4698{ 4699 struct rq *rq = raw_rq(); 4700 4701 delayacct_blkio_start(); 4702 atomic_inc(&rq->nr_iowait); 4703 blk_flush_plug(current); 4704 current->in_iowait = 1; 4705 schedule(); 4706 current->in_iowait = 0; 4707 atomic_dec(&rq->nr_iowait); 4708 delayacct_blkio_end(); 4709} 4710EXPORT_SYMBOL(io_schedule); 4711 4712long __sched io_schedule_timeout(long timeout) 4713{ 4714 struct rq *rq = raw_rq(); 4715 long ret; 4716 4717 delayacct_blkio_start(); 4718 atomic_inc(&rq->nr_iowait); 4719 blk_flush_plug(current); 4720 current->in_iowait = 1; 4721 ret = schedule_timeout(timeout); 4722 current->in_iowait = 0; 4723 atomic_dec(&rq->nr_iowait); 4724 delayacct_blkio_end(); 4725 return ret; 4726} 4727 4728/** 4729 * sys_sched_get_priority_max - return maximum RT priority. 4730 * @policy: scheduling class. 4731 * 4732 * this syscall returns the maximum rt_priority that can be used 4733 * by a given scheduling class. 4734 */ 4735SYSCALL_DEFINE1(sched_get_priority_max, int, policy) 4736{ 4737 int ret = -EINVAL; 4738 4739 switch (policy) { 4740 case SCHED_FIFO: 4741 case SCHED_RR: 4742 ret = MAX_USER_RT_PRIO-1; 4743 break; 4744 case SCHED_NORMAL: 4745 case SCHED_BATCH: 4746 case SCHED_IDLE: 4747 ret = 0; 4748 break; 4749 } 4750 return ret; 4751} 4752 4753/** 4754 * sys_sched_get_priority_min - return minimum RT priority. 4755 * @policy: scheduling class. 4756 * 4757 * this syscall returns the minimum rt_priority that can be used 4758 * by a given scheduling class. 4759 */ 4760SYSCALL_DEFINE1(sched_get_priority_min, int, policy) 4761{ 4762 int ret = -EINVAL; 4763 4764 switch (policy) { 4765 case SCHED_FIFO: 4766 case SCHED_RR: 4767 ret = 1; 4768 break; 4769 case SCHED_NORMAL: 4770 case SCHED_BATCH: 4771 case SCHED_IDLE: 4772 ret = 0; 4773 } 4774 return ret; 4775} 4776 4777/** 4778 * sys_sched_rr_get_interval - return the default timeslice of a process. 4779 * @pid: pid of the process. 4780 * @interval: userspace pointer to the timeslice value. 4781 * 4782 * this syscall writes the default timeslice value of a given process 4783 * into the user-space timespec buffer. A value of '0' means infinity. 4784 */ 4785SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid, 4786 struct timespec __user *, interval) 4787{ 4788 struct task_struct *p; 4789 unsigned int time_slice; 4790 unsigned long flags; 4791 struct rq *rq; 4792 int retval; 4793 struct timespec t; 4794 4795 if (pid < 0) 4796 return -EINVAL; 4797 4798 retval = -ESRCH; 4799 rcu_read_lock(); 4800 p = find_process_by_pid(pid); 4801 if (!p) 4802 goto out_unlock; 4803 4804 retval = security_task_getscheduler(p); 4805 if (retval) 4806 goto out_unlock; 4807 4808 rq = task_rq_lock(p, &flags); 4809 time_slice = p->sched_class->get_rr_interval(rq, p); 4810 task_rq_unlock(rq, p, &flags); 4811 4812 rcu_read_unlock(); 4813 jiffies_to_timespec(time_slice, &t); 4814 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0; 4815 return retval; 4816 4817out_unlock: 4818 rcu_read_unlock(); 4819 return retval; 4820} 4821 4822static const char stat_nam[] = TASK_STATE_TO_CHAR_STR; 4823 4824void sched_show_task(struct task_struct *p) 4825{ 4826 unsigned long free = 0; 4827 unsigned state; 4828 4829 state = p->state ? __ffs(p->state) + 1 : 0; 4830 printk(KERN_INFO "%-15.15s %c", p->comm, 4831 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?'); 4832#if BITS_PER_LONG == 32 4833 if (state == TASK_RUNNING) 4834 printk(KERN_CONT " running "); 4835 else 4836 printk(KERN_CONT " %08lx ", thread_saved_pc(p)); 4837#else 4838 if (state == TASK_RUNNING) 4839 printk(KERN_CONT " running task "); 4840 else 4841 printk(KERN_CONT " %016lx ", thread_saved_pc(p)); 4842#endif 4843#ifdef CONFIG_DEBUG_STACK_USAGE 4844 free = stack_not_used(p); 4845#endif 4846 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free, 4847 task_pid_nr(p), task_pid_nr(rcu_dereference(p->real_parent)), 4848 (unsigned long)task_thread_info(p)->flags); 4849 4850 show_stack(p, NULL); 4851} 4852 4853void show_state_filter(unsigned long state_filter) 4854{ 4855 struct task_struct *g, *p; 4856 4857#if BITS_PER_LONG == 32 4858 printk(KERN_INFO 4859 " task PC stack pid father\n"); 4860#else 4861 printk(KERN_INFO 4862 " task PC stack pid father\n"); 4863#endif 4864 rcu_read_lock(); 4865 do_each_thread(g, p) { 4866 /* 4867 * reset the NMI-timeout, listing all files on a slow 4868 * console might take a lot of time: 4869 */ 4870 touch_nmi_watchdog(); 4871 if (!state_filter || (p->state & state_filter)) 4872 sched_show_task(p); 4873 } while_each_thread(g, p); 4874 4875 touch_all_softlockup_watchdogs(); 4876 4877#ifdef CONFIG_SCHED_DEBUG 4878 sysrq_sched_debug_show(); 4879#endif 4880 rcu_read_unlock(); 4881 /* 4882 * Only show locks if all tasks are dumped: 4883 */ 4884 if (!state_filter) 4885 debug_show_all_locks(); 4886} 4887 4888void __cpuinit init_idle_bootup_task(struct task_struct *idle) 4889{ 4890 idle->sched_class = &idle_sched_class; 4891} 4892 4893/** 4894 * init_idle - set up an idle thread for a given CPU 4895 * @idle: task in question 4896 * @cpu: cpu the idle task belongs to 4897 * 4898 * NOTE: this function does not set the idle thread's NEED_RESCHED 4899 * flag, to make booting more robust. 4900 */ 4901void __cpuinit init_idle(struct task_struct *idle, int cpu) 4902{ 4903 struct rq *rq = cpu_rq(cpu); 4904 unsigned long flags; 4905 4906 raw_spin_lock_irqsave(&rq->lock, flags); 4907 4908 __sched_fork(idle); 4909 idle->state = TASK_RUNNING; 4910 idle->se.exec_start = sched_clock(); 4911 4912 do_set_cpus_allowed(idle, cpumask_of(cpu)); 4913 /* 4914 * We're having a chicken and egg problem, even though we are 4915 * holding rq->lock, the cpu isn't yet set to this cpu so the 4916 * lockdep check in task_group() will fail. 4917 * 4918 * Similar case to sched_fork(). / Alternatively we could 4919 * use task_rq_lock() here and obtain the other rq->lock. 4920 * 4921 * Silence PROVE_RCU 4922 */ 4923 rcu_read_lock(); 4924 __set_task_cpu(idle, cpu); 4925 rcu_read_unlock(); 4926 4927 rq->curr = rq->idle = idle; 4928#if defined(CONFIG_SMP) 4929 idle->on_cpu = 1; 4930#endif 4931 raw_spin_unlock_irqrestore(&rq->lock, flags); 4932 4933 /* Set the preempt count _outside_ the spinlocks! */ 4934 task_thread_info(idle)->preempt_count = 0; 4935 4936 /* 4937 * The idle tasks have their own, simple scheduling class: 4938 */ 4939 idle->sched_class = &idle_sched_class; 4940 ftrace_graph_init_idle_task(idle, cpu); 4941#if defined(CONFIG_SMP) 4942 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu); 4943#endif 4944} 4945 4946#ifdef CONFIG_SMP 4947void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask) 4948{ 4949 if (p->sched_class && p->sched_class->set_cpus_allowed) 4950 p->sched_class->set_cpus_allowed(p, new_mask); 4951 4952 cpumask_copy(&p->cpus_allowed, new_mask); 4953 p->rt.nr_cpus_allowed = cpumask_weight(new_mask); 4954} 4955 4956/* 4957 * This is how migration works: 4958 * 4959 * 1) we invoke migration_cpu_stop() on the target CPU using 4960 * stop_one_cpu(). 4961 * 2) stopper starts to run (implicitly forcing the migrated thread 4962 * off the CPU) 4963 * 3) it checks whether the migrated task is still in the wrong runqueue. 4964 * 4) if it's in the wrong runqueue then the migration thread removes 4965 * it and puts it into the right queue. 4966 * 5) stopper completes and stop_one_cpu() returns and the migration 4967 * is done. 4968 */ 4969 4970/* 4971 * Change a given task's CPU affinity. Migrate the thread to a 4972 * proper CPU and schedule it away if the CPU it's executing on 4973 * is removed from the allowed bitmask. 4974 * 4975 * NOTE: the caller must have a valid reference to the task, the 4976 * task must not exit() & deallocate itself prematurely. The 4977 * call is not atomic; no spinlocks may be held. 4978 */ 4979int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask) 4980{ 4981 unsigned long flags; 4982 struct rq *rq; 4983 unsigned int dest_cpu; 4984 int ret = 0; 4985 4986 rq = task_rq_lock(p, &flags); 4987 4988 if (cpumask_equal(&p->cpus_allowed, new_mask)) 4989 goto out; 4990 4991 if (!cpumask_intersects(new_mask, cpu_active_mask)) { 4992 ret = -EINVAL; 4993 goto out; 4994 } 4995 4996 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current)) { 4997 ret = -EINVAL; 4998 goto out; 4999 } 5000 5001 do_set_cpus_allowed(p, new_mask); 5002 5003 /* Can the task run on the task's current CPU? If so, we're done */ 5004 if (cpumask_test_cpu(task_cpu(p), new_mask)) 5005 goto out; 5006 5007 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask); 5008 if (p->on_rq) { 5009 struct migration_arg arg = { p, dest_cpu }; 5010 /* Need help from migration thread: drop lock and wait. */ 5011 task_rq_unlock(rq, p, &flags); 5012 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg); 5013 tlb_migrate_finish(p->mm); 5014 return 0; 5015 } 5016out: 5017 task_rq_unlock(rq, p, &flags); 5018 5019 return ret; 5020} 5021EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr); 5022 5023/* 5024 * Move (not current) task off this cpu, onto dest cpu. We're doing 5025 * this because either it can't run here any more (set_cpus_allowed() 5026 * away from this CPU, or CPU going down), or because we're 5027 * attempting to rebalance this task on exec (sched_exec). 5028 * 5029 * So we race with normal scheduler movements, but that's OK, as long 5030 * as the task is no longer on this CPU. 5031 * 5032 * Returns non-zero if task was successfully migrated. 5033 */ 5034static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu) 5035{ 5036 struct rq *rq_dest, *rq_src; 5037 int ret = 0; 5038 5039 if (unlikely(!cpu_active(dest_cpu))) 5040 return ret; 5041 5042 rq_src = cpu_rq(src_cpu); 5043 rq_dest = cpu_rq(dest_cpu); 5044 5045 raw_spin_lock(&p->pi_lock); 5046 double_rq_lock(rq_src, rq_dest); 5047 /* Already moved. */ 5048 if (task_cpu(p) != src_cpu) 5049 goto done; 5050 /* Affinity changed (again). */ 5051 if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p))) 5052 goto fail; 5053 5054 /* 5055 * If we're not on a rq, the next wake-up will ensure we're 5056 * placed properly. 5057 */ 5058 if (p->on_rq) { 5059 dequeue_task(rq_src, p, 0); 5060 set_task_cpu(p, dest_cpu); 5061 enqueue_task(rq_dest, p, 0); 5062 check_preempt_curr(rq_dest, p, 0); 5063 } 5064done: 5065 ret = 1; 5066fail: 5067 double_rq_unlock(rq_src, rq_dest); 5068 raw_spin_unlock(&p->pi_lock); 5069 return ret; 5070} 5071 5072/* 5073 * migration_cpu_stop - this will be executed by a highprio stopper thread 5074 * and performs thread migration by bumping thread off CPU then 5075 * 'pushing' onto another runqueue. 5076 */ 5077static int migration_cpu_stop(void *data) 5078{ 5079 struct migration_arg *arg = data; 5080 5081 /* 5082 * The original target cpu might have gone down and we might 5083 * be on another cpu but it doesn't matter. 5084 */ 5085 local_irq_disable(); 5086 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu); 5087 local_irq_enable(); 5088 return 0; 5089} 5090 5091#ifdef CONFIG_HOTPLUG_CPU 5092 5093/* 5094 * Ensures that the idle task is using init_mm right before its cpu goes 5095 * offline. 5096 */ 5097void idle_task_exit(void) 5098{ 5099 struct mm_struct *mm = current->active_mm; 5100 5101 BUG_ON(cpu_online(smp_processor_id())); 5102 5103 if (mm != &init_mm) 5104 switch_mm(mm, &init_mm, current); 5105 mmdrop(mm); 5106} 5107 5108/* 5109 * While a dead CPU has no uninterruptible tasks queued at this point, 5110 * it might still have a nonzero ->nr_uninterruptible counter, because 5111 * for performance reasons the counter is not stricly tracking tasks to 5112 * their home CPUs. So we just add the counter to another CPU's counter, 5113 * to keep the global sum constant after CPU-down: 5114 */ 5115static void migrate_nr_uninterruptible(struct rq *rq_src) 5116{ 5117 struct rq *rq_dest = cpu_rq(cpumask_any(cpu_active_mask)); 5118 5119 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible; 5120 rq_src->nr_uninterruptible = 0; 5121} 5122 5123/* 5124 * remove the tasks which were accounted by rq from calc_load_tasks. 5125 */ 5126static void calc_global_load_remove(struct rq *rq) 5127{ 5128 atomic_long_sub(rq->calc_load_active, &calc_load_tasks); 5129 rq->calc_load_active = 0; 5130} 5131 5132/* 5133 * Migrate all tasks from the rq, sleeping tasks will be migrated by 5134 * try_to_wake_up()->select_task_rq(). 5135 * 5136 * Called with rq->lock held even though we'er in stop_machine() and 5137 * there's no concurrency possible, we hold the required locks anyway 5138 * because of lock validation efforts. 5139 */ 5140static void migrate_tasks(unsigned int dead_cpu) 5141{ 5142 struct rq *rq = cpu_rq(dead_cpu); 5143 struct task_struct *next, *stop = rq->stop; 5144 int dest_cpu; 5145 5146 /* 5147 * Fudge the rq selection such that the below task selection loop 5148 * doesn't get stuck on the currently eligible stop task. 5149 * 5150 * We're currently inside stop_machine() and the rq is either stuck 5151 * in the stop_machine_cpu_stop() loop, or we're executing this code, 5152 * either way we should never end up calling schedule() until we're 5153 * done here. 5154 */ 5155 rq->stop = NULL; 5156 5157 /* Ensure any throttled groups are reachable by pick_next_task */ 5158 unthrottle_offline_cfs_rqs(rq); 5159 5160 for ( ; ; ) { 5161 /* 5162 * There's this thread running, bail when that's the only 5163 * remaining thread. 5164 */ 5165 if (rq->nr_running == 1) 5166 break; 5167 5168 next = pick_next_task(rq); 5169 BUG_ON(!next); 5170 next->sched_class->put_prev_task(rq, next); 5171 5172 /* Find suitable destination for @next, with force if needed. */ 5173 dest_cpu = select_fallback_rq(dead_cpu, next); 5174 raw_spin_unlock(&rq->lock); 5175 5176 __migrate_task(next, dead_cpu, dest_cpu); 5177 5178 raw_spin_lock(&rq->lock); 5179 } 5180 5181 rq->stop = stop; 5182} 5183 5184#endif /* CONFIG_HOTPLUG_CPU */ 5185 5186#if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL) 5187 5188static struct ctl_table sd_ctl_dir[] = { 5189 { 5190 .procname = "sched_domain", 5191 .mode = 0555, 5192 }, 5193 {} 5194}; 5195 5196static struct ctl_table sd_ctl_root[] = { 5197 { 5198 .procname = "kernel", 5199 .mode = 0555, 5200 .child = sd_ctl_dir, 5201 }, 5202 {} 5203}; 5204 5205static struct ctl_table *sd_alloc_ctl_entry(int n) 5206{ 5207 struct ctl_table *entry = 5208 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL); 5209 5210 return entry; 5211} 5212 5213static void sd_free_ctl_entry(struct ctl_table **tablep) 5214{ 5215 struct ctl_table *entry; 5216 5217 /* 5218 * In the intermediate directories, both the child directory and 5219 * procname are dynamically allocated and could fail but the mode 5220 * will always be set. In the lowest directory the names are 5221 * static strings and all have proc handlers. 5222 */ 5223 for (entry = *tablep; entry->mode; entry++) { 5224 if (entry->child) 5225 sd_free_ctl_entry(&entry->child); 5226 if (entry->proc_handler == NULL) 5227 kfree(entry->procname); 5228 } 5229 5230 kfree(*tablep); 5231 *tablep = NULL; 5232} 5233 5234static void 5235set_table_entry(struct ctl_table *entry, 5236 const char *procname, void *data, int maxlen, 5237 umode_t mode, proc_handler *proc_handler) 5238{ 5239 entry->procname = procname; 5240 entry->data = data; 5241 entry->maxlen = maxlen; 5242 entry->mode = mode; 5243 entry->proc_handler = proc_handler; 5244} 5245 5246static struct ctl_table * 5247sd_alloc_ctl_domain_table(struct sched_domain *sd) 5248{ 5249 struct ctl_table *table = sd_alloc_ctl_entry(13); 5250 5251 if (table == NULL) 5252 return NULL; 5253 5254 set_table_entry(&table[0], "min_interval", &sd->min_interval, 5255 sizeof(long), 0644, proc_doulongvec_minmax); 5256 set_table_entry(&table[1], "max_interval", &sd->max_interval, 5257 sizeof(long), 0644, proc_doulongvec_minmax); 5258 set_table_entry(&table[2], "busy_idx", &sd->busy_idx, 5259 sizeof(int), 0644, proc_dointvec_minmax); 5260 set_table_entry(&table[3], "idle_idx", &sd->idle_idx, 5261 sizeof(int), 0644, proc_dointvec_minmax); 5262 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx, 5263 sizeof(int), 0644, proc_dointvec_minmax); 5264 set_table_entry(&table[5], "wake_idx", &sd->wake_idx, 5265 sizeof(int), 0644, proc_dointvec_minmax); 5266 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx, 5267 sizeof(int), 0644, proc_dointvec_minmax); 5268 set_table_entry(&table[7], "busy_factor", &sd->busy_factor, 5269 sizeof(int), 0644, proc_dointvec_minmax); 5270 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct, 5271 sizeof(int), 0644, proc_dointvec_minmax); 5272 set_table_entry(&table[9], "cache_nice_tries", 5273 &sd->cache_nice_tries, 5274 sizeof(int), 0644, proc_dointvec_minmax); 5275 set_table_entry(&table[10], "flags", &sd->flags, 5276 sizeof(int), 0644, proc_dointvec_minmax); 5277 set_table_entry(&table[11], "name", sd->name, 5278 CORENAME_MAX_SIZE, 0444, proc_dostring); 5279 /* &table[12] is terminator */ 5280 5281 return table; 5282} 5283 5284static ctl_table *sd_alloc_ctl_cpu_table(int cpu) 5285{ 5286 struct ctl_table *entry, *table; 5287 struct sched_domain *sd; 5288 int domain_num = 0, i; 5289 char buf[32]; 5290 5291 for_each_domain(cpu, sd) 5292 domain_num++; 5293 entry = table = sd_alloc_ctl_entry(domain_num + 1); 5294 if (table == NULL) 5295 return NULL; 5296 5297 i = 0; 5298 for_each_domain(cpu, sd) { 5299 snprintf(buf, 32, "domain%d", i); 5300 entry->procname = kstrdup(buf, GFP_KERNEL); 5301 entry->mode = 0555; 5302 entry->child = sd_alloc_ctl_domain_table(sd); 5303 entry++; 5304 i++; 5305 } 5306 return table; 5307} 5308 5309static struct ctl_table_header *sd_sysctl_header; 5310static void register_sched_domain_sysctl(void) 5311{ 5312 int i, cpu_num = num_possible_cpus(); 5313 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1); 5314 char buf[32]; 5315 5316 WARN_ON(sd_ctl_dir[0].child); 5317 sd_ctl_dir[0].child = entry; 5318 5319 if (entry == NULL) 5320 return; 5321 5322 for_each_possible_cpu(i) { 5323 snprintf(buf, 32, "cpu%d", i); 5324 entry->procname = kstrdup(buf, GFP_KERNEL); 5325 entry->mode = 0555; 5326 entry->child = sd_alloc_ctl_cpu_table(i); 5327 entry++; 5328 } 5329 5330 WARN_ON(sd_sysctl_header); 5331 sd_sysctl_header = register_sysctl_table(sd_ctl_root); 5332} 5333 5334/* may be called multiple times per register */ 5335static void unregister_sched_domain_sysctl(void) 5336{ 5337 if (sd_sysctl_header) 5338 unregister_sysctl_table(sd_sysctl_header); 5339 sd_sysctl_header = NULL; 5340 if (sd_ctl_dir[0].child) 5341 sd_free_ctl_entry(&sd_ctl_dir[0].child); 5342} 5343#else 5344static void register_sched_domain_sysctl(void) 5345{ 5346} 5347static void unregister_sched_domain_sysctl(void) 5348{ 5349} 5350#endif 5351 5352static void set_rq_online(struct rq *rq) 5353{ 5354 if (!rq->online) { 5355 const struct sched_class *class; 5356 5357 cpumask_set_cpu(rq->cpu, rq->rd->online); 5358 rq->online = 1; 5359 5360 for_each_class(class) { 5361 if (class->rq_online) 5362 class->rq_online(rq); 5363 } 5364 } 5365} 5366 5367static void set_rq_offline(struct rq *rq) 5368{ 5369 if (rq->online) { 5370 const struct sched_class *class; 5371 5372 for_each_class(class) { 5373 if (class->rq_offline) 5374 class->rq_offline(rq); 5375 } 5376 5377 cpumask_clear_cpu(rq->cpu, rq->rd->online); 5378 rq->online = 0; 5379 } 5380} 5381 5382/* 5383 * migration_call - callback that gets triggered when a CPU is added. 5384 * Here we can start up the necessary migration thread for the new CPU. 5385 */ 5386static int __cpuinit 5387migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu) 5388{ 5389 int cpu = (long)hcpu; 5390 unsigned long flags; 5391 struct rq *rq = cpu_rq(cpu); 5392 5393 switch (action & ~CPU_TASKS_FROZEN) { 5394 5395 case CPU_UP_PREPARE: 5396 rq->calc_load_update = calc_load_update; 5397 break; 5398 5399 case CPU_ONLINE: 5400 /* Update our root-domain */ 5401 raw_spin_lock_irqsave(&rq->lock, flags); 5402 if (rq->rd) { 5403 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span)); 5404 5405 set_rq_online(rq); 5406 } 5407 raw_spin_unlock_irqrestore(&rq->lock, flags); 5408 break; 5409 5410#ifdef CONFIG_HOTPLUG_CPU 5411 case CPU_DYING: 5412 sched_ttwu_pending(); 5413 /* Update our root-domain */ 5414 raw_spin_lock_irqsave(&rq->lock, flags); 5415 if (rq->rd) { 5416 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span)); 5417 set_rq_offline(rq); 5418 } 5419 migrate_tasks(cpu); 5420 BUG_ON(rq->nr_running != 1); /* the migration thread */ 5421 raw_spin_unlock_irqrestore(&rq->lock, flags); 5422 5423 migrate_nr_uninterruptible(rq); 5424 calc_global_load_remove(rq); 5425 break; 5426#endif 5427 } 5428 5429 update_max_interval(); 5430 5431 return NOTIFY_OK; 5432} 5433 5434/* 5435 * Register at high priority so that task migration (migrate_all_tasks) 5436 * happens before everything else. This has to be lower priority than 5437 * the notifier in the perf_event subsystem, though. 5438 */ 5439static struct notifier_block __cpuinitdata migration_notifier = { 5440 .notifier_call = migration_call, 5441 .priority = CPU_PRI_MIGRATION, 5442}; 5443 5444static int __cpuinit sched_cpu_active(struct notifier_block *nfb, 5445 unsigned long action, void *hcpu) 5446{ 5447 switch (action & ~CPU_TASKS_FROZEN) { 5448 case CPU_STARTING: 5449 case CPU_DOWN_FAILED: 5450 set_cpu_active((long)hcpu, true); 5451 return NOTIFY_OK; 5452 default: 5453 return NOTIFY_DONE; 5454 } 5455} 5456 5457static int __cpuinit sched_cpu_inactive(struct notifier_block *nfb, 5458 unsigned long action, void *hcpu) 5459{ 5460 switch (action & ~CPU_TASKS_FROZEN) { 5461 case CPU_DOWN_PREPARE: 5462 set_cpu_active((long)hcpu, false); 5463 return NOTIFY_OK; 5464 default: 5465 return NOTIFY_DONE; 5466 } 5467} 5468 5469static int __init migration_init(void) 5470{ 5471 void *cpu = (void *)(long)smp_processor_id(); 5472 int err; 5473 5474 /* Initialize migration for the boot CPU */ 5475 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu); 5476 BUG_ON(err == NOTIFY_BAD); 5477 migration_call(&migration_notifier, CPU_ONLINE, cpu); 5478 register_cpu_notifier(&migration_notifier); 5479 5480 /* Register cpu active notifiers */ 5481 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE); 5482 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE); 5483 5484 return 0; 5485} 5486early_initcall(migration_init); 5487#endif 5488 5489#ifdef CONFIG_SMP 5490 5491static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */ 5492 5493#ifdef CONFIG_SCHED_DEBUG 5494 5495static __read_mostly int sched_domain_debug_enabled; 5496 5497static int __init sched_domain_debug_setup(char *str) 5498{ 5499 sched_domain_debug_enabled = 1; 5500 5501 return 0; 5502} 5503early_param("sched_debug", sched_domain_debug_setup); 5504 5505static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level, 5506 struct cpumask *groupmask) 5507{ 5508 struct sched_group *group = sd->groups; 5509 char str[256]; 5510 5511 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd)); 5512 cpumask_clear(groupmask); 5513 5514 printk(KERN_DEBUG "%*s domain %d: ", level, "", level); 5515 5516 if (!(sd->flags & SD_LOAD_BALANCE)) { 5517 printk("does not load-balance\n"); 5518 if (sd->parent) 5519 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain" 5520 " has parent"); 5521 return -1; 5522 } 5523 5524 printk(KERN_CONT "span %s level %s\n", str, sd->name); 5525 5526 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) { 5527 printk(KERN_ERR "ERROR: domain->span does not contain " 5528 "CPU%d\n", cpu); 5529 } 5530 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) { 5531 printk(KERN_ERR "ERROR: domain->groups does not contain" 5532 " CPU%d\n", cpu); 5533 } 5534 5535 printk(KERN_DEBUG "%*s groups:", level + 1, ""); 5536 do { 5537 if (!group) { 5538 printk("\n"); 5539 printk(KERN_ERR "ERROR: group is NULL\n"); 5540 break; 5541 } 5542 5543 if (!group->sgp->power) { 5544 printk(KERN_CONT "\n"); 5545 printk(KERN_ERR "ERROR: domain->cpu_power not " 5546 "set\n"); 5547 break; 5548 } 5549 5550 if (!cpumask_weight(sched_group_cpus(group))) { 5551 printk(KERN_CONT "\n"); 5552 printk(KERN_ERR "ERROR: empty group\n"); 5553 break; 5554 } 5555 5556 if (cpumask_intersects(groupmask, sched_group_cpus(group))) { 5557 printk(KERN_CONT "\n"); 5558 printk(KERN_ERR "ERROR: repeated CPUs\n"); 5559 break; 5560 } 5561 5562 cpumask_or(groupmask, groupmask, sched_group_cpus(group)); 5563 5564 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group)); 5565 5566 printk(KERN_CONT " %s", str); 5567 if (group->sgp->power != SCHED_POWER_SCALE) { 5568 printk(KERN_CONT " (cpu_power = %d)", 5569 group->sgp->power); 5570 } 5571 5572 group = group->next; 5573 } while (group != sd->groups); 5574 printk(KERN_CONT "\n"); 5575 5576 if (!cpumask_equal(sched_domain_span(sd), groupmask)) 5577 printk(KERN_ERR "ERROR: groups don't span domain->span\n"); 5578 5579 if (sd->parent && 5580 !cpumask_subset(groupmask, sched_domain_span(sd->parent))) 5581 printk(KERN_ERR "ERROR: parent span is not a superset " 5582 "of domain->span\n"); 5583 return 0; 5584} 5585 5586static void sched_domain_debug(struct sched_domain *sd, int cpu) 5587{ 5588 int level = 0; 5589 5590 if (!sched_domain_debug_enabled) 5591 return; 5592 5593 if (!sd) { 5594 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu); 5595 return; 5596 } 5597 5598 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu); 5599 5600 for (;;) { 5601 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask)) 5602 break; 5603 level++; 5604 sd = sd->parent; 5605 if (!sd) 5606 break; 5607 } 5608} 5609#else /* !CONFIG_SCHED_DEBUG */ 5610# define sched_domain_debug(sd, cpu) do { } while (0) 5611#endif /* CONFIG_SCHED_DEBUG */ 5612 5613static int sd_degenerate(struct sched_domain *sd) 5614{ 5615 if (cpumask_weight(sched_domain_span(sd)) == 1) 5616 return 1; 5617 5618 /* Following flags need at least 2 groups */ 5619 if (sd->flags & (SD_LOAD_BALANCE | 5620 SD_BALANCE_NEWIDLE | 5621 SD_BALANCE_FORK | 5622 SD_BALANCE_EXEC | 5623 SD_SHARE_CPUPOWER | 5624 SD_SHARE_PKG_RESOURCES)) { 5625 if (sd->groups != sd->groups->next) 5626 return 0; 5627 } 5628 5629 /* Following flags don't use groups */ 5630 if (sd->flags & (SD_WAKE_AFFINE)) 5631 return 0; 5632 5633 return 1; 5634} 5635 5636static int 5637sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent) 5638{ 5639 unsigned long cflags = sd->flags, pflags = parent->flags; 5640 5641 if (sd_degenerate(parent)) 5642 return 1; 5643 5644 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent))) 5645 return 0; 5646 5647 /* Flags needing groups don't count if only 1 group in parent */ 5648 if (parent->groups == parent->groups->next) { 5649 pflags &= ~(SD_LOAD_BALANCE | 5650 SD_BALANCE_NEWIDLE | 5651 SD_BALANCE_FORK | 5652 SD_BALANCE_EXEC | 5653 SD_SHARE_CPUPOWER | 5654 SD_SHARE_PKG_RESOURCES); 5655 if (nr_node_ids == 1) 5656 pflags &= ~SD_SERIALIZE; 5657 } 5658 if (~cflags & pflags) 5659 return 0; 5660 5661 return 1; 5662} 5663 5664static void free_rootdomain(struct rcu_head *rcu) 5665{ 5666 struct root_domain *rd = container_of(rcu, struct root_domain, rcu); 5667 5668 cpupri_cleanup(&rd->cpupri); 5669 free_cpumask_var(rd->rto_mask); 5670 free_cpumask_var(rd->online); 5671 free_cpumask_var(rd->span); 5672 kfree(rd); 5673} 5674 5675static void rq_attach_root(struct rq *rq, struct root_domain *rd) 5676{ 5677 struct root_domain *old_rd = NULL; 5678 unsigned long flags; 5679 5680 raw_spin_lock_irqsave(&rq->lock, flags); 5681 5682 if (rq->rd) { 5683 old_rd = rq->rd; 5684 5685 if (cpumask_test_cpu(rq->cpu, old_rd->online)) 5686 set_rq_offline(rq); 5687 5688 cpumask_clear_cpu(rq->cpu, old_rd->span); 5689 5690 /* 5691 * If we dont want to free the old_rt yet then 5692 * set old_rd to NULL to skip the freeing later 5693 * in this function: 5694 */ 5695 if (!atomic_dec_and_test(&old_rd->refcount)) 5696 old_rd = NULL; 5697 } 5698 5699 atomic_inc(&rd->refcount); 5700 rq->rd = rd; 5701 5702 cpumask_set_cpu(rq->cpu, rd->span); 5703 if (cpumask_test_cpu(rq->cpu, cpu_active_mask)) 5704 set_rq_online(rq); 5705 5706 raw_spin_unlock_irqrestore(&rq->lock, flags); 5707 5708 if (old_rd) 5709 call_rcu_sched(&old_rd->rcu, free_rootdomain); 5710} 5711 5712static int init_rootdomain(struct root_domain *rd) 5713{ 5714 memset(rd, 0, sizeof(*rd)); 5715 5716 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL)) 5717 goto out; 5718 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL)) 5719 goto free_span; 5720 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL)) 5721 goto free_online; 5722 5723 if (cpupri_init(&rd->cpupri) != 0) 5724 goto free_rto_mask; 5725 return 0; 5726 5727free_rto_mask: 5728 free_cpumask_var(rd->rto_mask); 5729free_online: 5730 free_cpumask_var(rd->online); 5731free_span: 5732 free_cpumask_var(rd->span); 5733out: 5734 return -ENOMEM; 5735} 5736 5737/* 5738 * By default the system creates a single root-domain with all cpus as 5739 * members (mimicking the global state we have today). 5740 */ 5741struct root_domain def_root_domain; 5742 5743static void init_defrootdomain(void) 5744{ 5745 init_rootdomain(&def_root_domain); 5746 5747 atomic_set(&def_root_domain.refcount, 1); 5748} 5749 5750static struct root_domain *alloc_rootdomain(void) 5751{ 5752 struct root_domain *rd; 5753 5754 rd = kmalloc(sizeof(*rd), GFP_KERNEL); 5755 if (!rd) 5756 return NULL; 5757 5758 if (init_rootdomain(rd) != 0) { 5759 kfree(rd); 5760 return NULL; 5761 } 5762 5763 return rd; 5764} 5765 5766static void free_sched_groups(struct sched_group *sg, int free_sgp) 5767{ 5768 struct sched_group *tmp, *first; 5769 5770 if (!sg) 5771 return; 5772 5773 first = sg; 5774 do { 5775 tmp = sg->next; 5776 5777 if (free_sgp && atomic_dec_and_test(&sg->sgp->ref)) 5778 kfree(sg->sgp); 5779 5780 kfree(sg); 5781 sg = tmp; 5782 } while (sg != first); 5783} 5784 5785static void free_sched_domain(struct rcu_head *rcu) 5786{ 5787 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu); 5788 5789 /* 5790 * If its an overlapping domain it has private groups, iterate and 5791 * nuke them all. 5792 */ 5793 if (sd->flags & SD_OVERLAP) { 5794 free_sched_groups(sd->groups, 1); 5795 } else if (atomic_dec_and_test(&sd->groups->ref)) { 5796 kfree(sd->groups->sgp); 5797 kfree(sd->groups); 5798 } 5799 kfree(sd); 5800} 5801 5802static void destroy_sched_domain(struct sched_domain *sd, int cpu) 5803{ 5804 call_rcu(&sd->rcu, free_sched_domain); 5805} 5806 5807static void destroy_sched_domains(struct sched_domain *sd, int cpu) 5808{ 5809 for (; sd; sd = sd->parent) 5810 destroy_sched_domain(sd, cpu); 5811} 5812 5813/* 5814 * Keep a special pointer to the highest sched_domain that has 5815 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this 5816 * allows us to avoid some pointer chasing select_idle_sibling(). 5817 * 5818 * Also keep a unique ID per domain (we use the first cpu number in 5819 * the cpumask of the domain), this allows us to quickly tell if 5820 * two cpus are in the same cache domain, see cpus_share_cache(). 5821 */ 5822DEFINE_PER_CPU(struct sched_domain *, sd_llc); 5823DEFINE_PER_CPU(int, sd_llc_id); 5824 5825static void update_top_cache_domain(int cpu) 5826{ 5827 struct sched_domain *sd; 5828 int id = cpu; 5829 5830 sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES); 5831 if (sd) 5832 id = cpumask_first(sched_domain_span(sd)); 5833 5834 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd); 5835 per_cpu(sd_llc_id, cpu) = id; 5836} 5837 5838/* 5839 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must 5840 * hold the hotplug lock. 5841 */ 5842static void 5843cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu) 5844{ 5845 struct rq *rq = cpu_rq(cpu); 5846 struct sched_domain *tmp; 5847 5848 /* Remove the sched domains which do not contribute to scheduling. */ 5849 for (tmp = sd; tmp; ) { 5850 struct sched_domain *parent = tmp->parent; 5851 if (!parent) 5852 break; 5853 5854 if (sd_parent_degenerate(tmp, parent)) { 5855 tmp->parent = parent->parent; 5856 if (parent->parent) 5857 parent->parent->child = tmp; 5858 destroy_sched_domain(parent, cpu); 5859 } else 5860 tmp = tmp->parent; 5861 } 5862 5863 if (sd && sd_degenerate(sd)) { 5864 tmp = sd; 5865 sd = sd->parent; 5866 destroy_sched_domain(tmp, cpu); 5867 if (sd) 5868 sd->child = NULL; 5869 } 5870 5871 sched_domain_debug(sd, cpu); 5872 5873 rq_attach_root(rq, rd); 5874 tmp = rq->sd; 5875 rcu_assign_pointer(rq->sd, sd); 5876 destroy_sched_domains(tmp, cpu); 5877 5878 update_top_cache_domain(cpu); 5879} 5880 5881/* cpus with isolated domains */ 5882static cpumask_var_t cpu_isolated_map; 5883 5884/* Setup the mask of cpus configured for isolated domains */ 5885static int __init isolated_cpu_setup(char *str) 5886{ 5887 alloc_bootmem_cpumask_var(&cpu_isolated_map); 5888 cpulist_parse(str, cpu_isolated_map); 5889 return 1; 5890} 5891 5892__setup("isolcpus=", isolated_cpu_setup); 5893 5894#ifdef CONFIG_NUMA 5895 5896/** 5897 * find_next_best_node - find the next node to include in a sched_domain 5898 * @node: node whose sched_domain we're building 5899 * @used_nodes: nodes already in the sched_domain 5900 * 5901 * Find the next node to include in a given scheduling domain. Simply 5902 * finds the closest node not already in the @used_nodes map. 5903 * 5904 * Should use nodemask_t. 5905 */ 5906static int find_next_best_node(int node, nodemask_t *used_nodes) 5907{ 5908 int i, n, val, min_val, best_node = -1; 5909 5910 min_val = INT_MAX; 5911 5912 for (i = 0; i < nr_node_ids; i++) { 5913 /* Start at @node */ 5914 n = (node + i) % nr_node_ids; 5915 5916 if (!nr_cpus_node(n)) 5917 continue; 5918 5919 /* Skip already used nodes */ 5920 if (node_isset(n, *used_nodes)) 5921 continue; 5922 5923 /* Simple min distance search */ 5924 val = node_distance(node, n); 5925 5926 if (val < min_val) { 5927 min_val = val; 5928 best_node = n; 5929 } 5930 } 5931 5932 if (best_node != -1) 5933 node_set(best_node, *used_nodes); 5934 return best_node; 5935} 5936 5937/** 5938 * sched_domain_node_span - get a cpumask for a node's sched_domain 5939 * @node: node whose cpumask we're constructing 5940 * @span: resulting cpumask 5941 * 5942 * Given a node, construct a good cpumask for its sched_domain to span. It 5943 * should be one that prevents unnecessary balancing, but also spreads tasks 5944 * out optimally. 5945 */ 5946static void sched_domain_node_span(int node, struct cpumask *span) 5947{ 5948 nodemask_t used_nodes; 5949 int i; 5950 5951 cpumask_clear(span); 5952 nodes_clear(used_nodes); 5953 5954 cpumask_or(span, span, cpumask_of_node(node)); 5955 node_set(node, used_nodes); 5956 5957 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) { 5958 int next_node = find_next_best_node(node, &used_nodes); 5959 if (next_node < 0) 5960 break; 5961 cpumask_or(span, span, cpumask_of_node(next_node)); 5962 } 5963} 5964 5965static const struct cpumask *cpu_node_mask(int cpu) 5966{ 5967 lockdep_assert_held(&sched_domains_mutex); 5968 5969 sched_domain_node_span(cpu_to_node(cpu), sched_domains_tmpmask); 5970 5971 return sched_domains_tmpmask; 5972} 5973 5974static const struct cpumask *cpu_allnodes_mask(int cpu) 5975{ 5976 return cpu_possible_mask; 5977} 5978#endif /* CONFIG_NUMA */ 5979 5980static const struct cpumask *cpu_cpu_mask(int cpu) 5981{ 5982 return cpumask_of_node(cpu_to_node(cpu)); 5983} 5984 5985int sched_smt_power_savings = 0, sched_mc_power_savings = 0; 5986 5987struct sd_data { 5988 struct sched_domain **__percpu sd; 5989 struct sched_group **__percpu sg; 5990 struct sched_group_power **__percpu sgp; 5991}; 5992 5993struct s_data { 5994 struct sched_domain ** __percpu sd; 5995 struct root_domain *rd; 5996}; 5997 5998enum s_alloc { 5999 sa_rootdomain, 6000 sa_sd, 6001 sa_sd_storage, 6002 sa_none, 6003}; 6004 6005struct sched_domain_topology_level; 6006 6007typedef struct sched_domain *(*sched_domain_init_f)(struct sched_domain_topology_level *tl, int cpu); 6008typedef const struct cpumask *(*sched_domain_mask_f)(int cpu); 6009 6010#define SDTL_OVERLAP 0x01 6011 6012struct sched_domain_topology_level { 6013 sched_domain_init_f init; 6014 sched_domain_mask_f mask; 6015 int flags; 6016 struct sd_data data; 6017}; 6018 6019static int 6020build_overlap_sched_groups(struct sched_domain *sd, int cpu) 6021{ 6022 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg; 6023 const struct cpumask *span = sched_domain_span(sd); 6024 struct cpumask *covered = sched_domains_tmpmask; 6025 struct sd_data *sdd = sd->private; 6026 struct sched_domain *child; 6027 int i; 6028 6029 cpumask_clear(covered); 6030 6031 for_each_cpu(i, span) { 6032 struct cpumask *sg_span; 6033 6034 if (cpumask_test_cpu(i, covered)) 6035 continue; 6036 6037 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(), 6038 GFP_KERNEL, cpu_to_node(cpu)); 6039 6040 if (!sg) 6041 goto fail; 6042 6043 sg_span = sched_group_cpus(sg); 6044 6045 child = *per_cpu_ptr(sdd->sd, i); 6046 if (child->child) { 6047 child = child->child; 6048 cpumask_copy(sg_span, sched_domain_span(child)); 6049 } else 6050 cpumask_set_cpu(i, sg_span); 6051 6052 cpumask_or(covered, covered, sg_span); 6053 6054 sg->sgp = *per_cpu_ptr(sdd->sgp, cpumask_first(sg_span)); 6055 atomic_inc(&sg->sgp->ref); 6056 6057 if (cpumask_test_cpu(cpu, sg_span)) 6058 groups = sg; 6059 6060 if (!first) 6061 first = sg; 6062 if (last) 6063 last->next = sg; 6064 last = sg; 6065 last->next = first; 6066 } 6067 sd->groups = groups; 6068 6069 return 0; 6070 6071fail: 6072 free_sched_groups(first, 0); 6073 6074 return -ENOMEM; 6075} 6076 6077static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg) 6078{ 6079 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu); 6080 struct sched_domain *child = sd->child; 6081 6082 if (child) 6083 cpu = cpumask_first(sched_domain_span(child)); 6084 6085 if (sg) { 6086 *sg = *per_cpu_ptr(sdd->sg, cpu); 6087 (*sg)->sgp = *per_cpu_ptr(sdd->sgp, cpu); 6088 atomic_set(&(*sg)->sgp->ref, 1); /* for claim_allocations */ 6089 } 6090 6091 return cpu; 6092} 6093 6094/* 6095 * build_sched_groups will build a circular linked list of the groups 6096 * covered by the given span, and will set each group's ->cpumask correctly, 6097 * and ->cpu_power to 0. 6098 * 6099 * Assumes the sched_domain tree is fully constructed 6100 */ 6101static int 6102build_sched_groups(struct sched_domain *sd, int cpu) 6103{ 6104 struct sched_group *first = NULL, *last = NULL; 6105 struct sd_data *sdd = sd->private; 6106 const struct cpumask *span = sched_domain_span(sd); 6107 struct cpumask *covered; 6108 int i; 6109 6110 get_group(cpu, sdd, &sd->groups); 6111 atomic_inc(&sd->groups->ref); 6112 6113 if (cpu != cpumask_first(sched_domain_span(sd))) 6114 return 0; 6115 6116 lockdep_assert_held(&sched_domains_mutex); 6117 covered = sched_domains_tmpmask; 6118 6119 cpumask_clear(covered); 6120 6121 for_each_cpu(i, span) { 6122 struct sched_group *sg; 6123 int group = get_group(i, sdd, &sg); 6124 int j; 6125 6126 if (cpumask_test_cpu(i, covered)) 6127 continue; 6128 6129 cpumask_clear(sched_group_cpus(sg)); 6130 sg->sgp->power = 0; 6131 6132 for_each_cpu(j, span) { 6133 if (get_group(j, sdd, NULL) != group) 6134 continue; 6135 6136 cpumask_set_cpu(j, covered); 6137 cpumask_set_cpu(j, sched_group_cpus(sg)); 6138 } 6139 6140 if (!first) 6141 first = sg; 6142 if (last) 6143 last->next = sg; 6144 last = sg; 6145 } 6146 last->next = first; 6147 6148 return 0; 6149} 6150 6151/* 6152 * Initialize sched groups cpu_power. 6153 * 6154 * cpu_power indicates the capacity of sched group, which is used while 6155 * distributing the load between different sched groups in a sched domain. 6156 * Typically cpu_power for all the groups in a sched domain will be same unless 6157 * there are asymmetries in the topology. If there are asymmetries, group 6158 * having more cpu_power will pickup more load compared to the group having 6159 * less cpu_power. 6160 */ 6161static void init_sched_groups_power(int cpu, struct sched_domain *sd) 6162{ 6163 struct sched_group *sg = sd->groups; 6164 6165 WARN_ON(!sd || !sg); 6166 6167 do { 6168 sg->group_weight = cpumask_weight(sched_group_cpus(sg)); 6169 sg = sg->next; 6170 } while (sg != sd->groups); 6171 6172 if (cpu != group_first_cpu(sg)) 6173 return; 6174 6175 update_group_power(sd, cpu); 6176 atomic_set(&sg->sgp->nr_busy_cpus, sg->group_weight); 6177} 6178 6179int __weak arch_sd_sibling_asym_packing(void) 6180{ 6181 return 0*SD_ASYM_PACKING; 6182} 6183 6184/* 6185 * Initializers for schedule domains 6186 * Non-inlined to reduce accumulated stack pressure in build_sched_domains() 6187 */ 6188 6189#ifdef CONFIG_SCHED_DEBUG 6190# define SD_INIT_NAME(sd, type) sd->name = #type 6191#else 6192# define SD_INIT_NAME(sd, type) do { } while (0) 6193#endif 6194 6195#define SD_INIT_FUNC(type) \ 6196static noinline struct sched_domain * \ 6197sd_init_##type(struct sched_domain_topology_level *tl, int cpu) \ 6198{ \ 6199 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); \ 6200 *sd = SD_##type##_INIT; \ 6201 SD_INIT_NAME(sd, type); \ 6202 sd->private = &tl->data; \ 6203 return sd; \ 6204} 6205 6206SD_INIT_FUNC(CPU) 6207#ifdef CONFIG_NUMA 6208 SD_INIT_FUNC(ALLNODES) 6209 SD_INIT_FUNC(NODE) 6210#endif 6211#ifdef CONFIG_SCHED_SMT 6212 SD_INIT_FUNC(SIBLING) 6213#endif 6214#ifdef CONFIG_SCHED_MC 6215 SD_INIT_FUNC(MC) 6216#endif 6217#ifdef CONFIG_SCHED_BOOK 6218 SD_INIT_FUNC(BOOK) 6219#endif 6220 6221static int default_relax_domain_level = -1; 6222int sched_domain_level_max; 6223 6224static int __init setup_relax_domain_level(char *str) 6225{ 6226 unsigned long val; 6227 6228 val = simple_strtoul(str, NULL, 0); 6229 if (val < sched_domain_level_max) 6230 default_relax_domain_level = val; 6231 6232 return 1; 6233} 6234__setup("relax_domain_level=", setup_relax_domain_level); 6235 6236static void set_domain_attribute(struct sched_domain *sd, 6237 struct sched_domain_attr *attr) 6238{ 6239 int request; 6240 6241 if (!attr || attr->relax_domain_level < 0) { 6242 if (default_relax_domain_level < 0) 6243 return; 6244 else 6245 request = default_relax_domain_level; 6246 } else 6247 request = attr->relax_domain_level; 6248 if (request < sd->level) { 6249 /* turn off idle balance on this domain */ 6250 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE); 6251 } else { 6252 /* turn on idle balance on this domain */ 6253 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE); 6254 } 6255} 6256 6257static void __sdt_free(const struct cpumask *cpu_map); 6258static int __sdt_alloc(const struct cpumask *cpu_map); 6259 6260static void __free_domain_allocs(struct s_data *d, enum s_alloc what, 6261 const struct cpumask *cpu_map) 6262{ 6263 switch (what) { 6264 case sa_rootdomain: 6265 if (!atomic_read(&d->rd->refcount)) 6266 free_rootdomain(&d->rd->rcu); /* fall through */ 6267 case sa_sd: 6268 free_percpu(d->sd); /* fall through */ 6269 case sa_sd_storage: 6270 __sdt_free(cpu_map); /* fall through */ 6271 case sa_none: 6272 break; 6273 } 6274} 6275 6276static enum s_alloc __visit_domain_allocation_hell(struct s_data *d, 6277 const struct cpumask *cpu_map) 6278{ 6279 memset(d, 0, sizeof(*d)); 6280 6281 if (__sdt_alloc(cpu_map)) 6282 return sa_sd_storage; 6283 d->sd = alloc_percpu(struct sched_domain *); 6284 if (!d->sd) 6285 return sa_sd_storage; 6286 d->rd = alloc_rootdomain(); 6287 if (!d->rd) 6288 return sa_sd; 6289 return sa_rootdomain; 6290} 6291 6292/* 6293 * NULL the sd_data elements we've used to build the sched_domain and 6294 * sched_group structure so that the subsequent __free_domain_allocs() 6295 * will not free the data we're using. 6296 */ 6297static void claim_allocations(int cpu, struct sched_domain *sd) 6298{ 6299 struct sd_data *sdd = sd->private; 6300 6301 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd); 6302 *per_cpu_ptr(sdd->sd, cpu) = NULL; 6303 6304 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref)) 6305 *per_cpu_ptr(sdd->sg, cpu) = NULL; 6306 6307 if (atomic_read(&(*per_cpu_ptr(sdd->sgp, cpu))->ref)) 6308 *per_cpu_ptr(sdd->sgp, cpu) = NULL; 6309} 6310 6311#ifdef CONFIG_SCHED_SMT 6312static const struct cpumask *cpu_smt_mask(int cpu) 6313{ 6314 return topology_thread_cpumask(cpu); 6315} 6316#endif 6317 6318/* 6319 * Topology list, bottom-up. 6320 */ 6321static struct sched_domain_topology_level default_topology[] = { 6322#ifdef CONFIG_SCHED_SMT 6323 { sd_init_SIBLING, cpu_smt_mask, }, 6324#endif 6325#ifdef CONFIG_SCHED_MC 6326 { sd_init_MC, cpu_coregroup_mask, }, 6327#endif 6328#ifdef CONFIG_SCHED_BOOK 6329 { sd_init_BOOK, cpu_book_mask, }, 6330#endif 6331 { sd_init_CPU, cpu_cpu_mask, }, 6332#ifdef CONFIG_NUMA 6333 { sd_init_NODE, cpu_node_mask, SDTL_OVERLAP, }, 6334 { sd_init_ALLNODES, cpu_allnodes_mask, }, 6335#endif 6336 { NULL, }, 6337}; 6338 6339static struct sched_domain_topology_level *sched_domain_topology = default_topology; 6340 6341static int __sdt_alloc(const struct cpumask *cpu_map) 6342{ 6343 struct sched_domain_topology_level *tl; 6344 int j; 6345 6346 for (tl = sched_domain_topology; tl->init; tl++) { 6347 struct sd_data *sdd = &tl->data; 6348 6349 sdd->sd = alloc_percpu(struct sched_domain *); 6350 if (!sdd->sd) 6351 return -ENOMEM; 6352 6353 sdd->sg = alloc_percpu(struct sched_group *); 6354 if (!sdd->sg) 6355 return -ENOMEM; 6356 6357 sdd->sgp = alloc_percpu(struct sched_group_power *); 6358 if (!sdd->sgp) 6359 return -ENOMEM; 6360 6361 for_each_cpu(j, cpu_map) { 6362 struct sched_domain *sd; 6363 struct sched_group *sg; 6364 struct sched_group_power *sgp; 6365 6366 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(), 6367 GFP_KERNEL, cpu_to_node(j)); 6368 if (!sd) 6369 return -ENOMEM; 6370 6371 *per_cpu_ptr(sdd->sd, j) = sd; 6372 6373 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(), 6374 GFP_KERNEL, cpu_to_node(j)); 6375 if (!sg) 6376 return -ENOMEM; 6377 6378 *per_cpu_ptr(sdd->sg, j) = sg; 6379 6380 sgp = kzalloc_node(sizeof(struct sched_group_power), 6381 GFP_KERNEL, cpu_to_node(j)); 6382 if (!sgp) 6383 return -ENOMEM; 6384 6385 *per_cpu_ptr(sdd->sgp, j) = sgp; 6386 } 6387 } 6388 6389 return 0; 6390} 6391 6392static void __sdt_free(const struct cpumask *cpu_map) 6393{ 6394 struct sched_domain_topology_level *tl; 6395 int j; 6396 6397 for (tl = sched_domain_topology; tl->init; tl++) { 6398 struct sd_data *sdd = &tl->data; 6399 6400 for_each_cpu(j, cpu_map) { 6401 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, j); 6402 if (sd && (sd->flags & SD_OVERLAP)) 6403 free_sched_groups(sd->groups, 0); 6404 kfree(*per_cpu_ptr(sdd->sd, j)); 6405 kfree(*per_cpu_ptr(sdd->sg, j)); 6406 kfree(*per_cpu_ptr(sdd->sgp, j)); 6407 } 6408 free_percpu(sdd->sd); 6409 free_percpu(sdd->sg); 6410 free_percpu(sdd->sgp); 6411 } 6412} 6413 6414struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl, 6415 struct s_data *d, const struct cpumask *cpu_map, 6416 struct sched_domain_attr *attr, struct sched_domain *child, 6417 int cpu) 6418{ 6419 struct sched_domain *sd = tl->init(tl, cpu); 6420 if (!sd) 6421 return child; 6422 6423 set_domain_attribute(sd, attr); 6424 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu)); 6425 if (child) { 6426 sd->level = child->level + 1; 6427 sched_domain_level_max = max(sched_domain_level_max, sd->level); 6428 child->parent = sd; 6429 } 6430 sd->child = child; 6431 6432 return sd; 6433} 6434 6435/* 6436 * Build sched domains for a given set of cpus and attach the sched domains 6437 * to the individual cpus 6438 */ 6439static int build_sched_domains(const struct cpumask *cpu_map, 6440 struct sched_domain_attr *attr) 6441{ 6442 enum s_alloc alloc_state = sa_none; 6443 struct sched_domain *sd; 6444 struct s_data d; 6445 int i, ret = -ENOMEM; 6446 6447 alloc_state = __visit_domain_allocation_hell(&d, cpu_map); 6448 if (alloc_state != sa_rootdomain) 6449 goto error; 6450 6451 /* Set up domains for cpus specified by the cpu_map. */ 6452 for_each_cpu(i, cpu_map) { 6453 struct sched_domain_topology_level *tl; 6454 6455 sd = NULL; 6456 for (tl = sched_domain_topology; tl->init; tl++) { 6457 sd = build_sched_domain(tl, &d, cpu_map, attr, sd, i); 6458 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP)) 6459 sd->flags |= SD_OVERLAP; 6460 if (cpumask_equal(cpu_map, sched_domain_span(sd))) 6461 break; 6462 } 6463 6464 while (sd->child) 6465 sd = sd->child; 6466 6467 *per_cpu_ptr(d.sd, i) = sd; 6468 } 6469 6470 /* Build the groups for the domains */ 6471 for_each_cpu(i, cpu_map) { 6472 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) { 6473 sd->span_weight = cpumask_weight(sched_domain_span(sd)); 6474 if (sd->flags & SD_OVERLAP) { 6475 if (build_overlap_sched_groups(sd, i)) 6476 goto error; 6477 } else { 6478 if (build_sched_groups(sd, i)) 6479 goto error; 6480 } 6481 } 6482 } 6483 6484 /* Calculate CPU power for physical packages and nodes */ 6485 for (i = nr_cpumask_bits-1; i >= 0; i--) { 6486 if (!cpumask_test_cpu(i, cpu_map)) 6487 continue; 6488 6489 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) { 6490 claim_allocations(i, sd); 6491 init_sched_groups_power(i, sd); 6492 } 6493 } 6494 6495 /* Attach the domains */ 6496 rcu_read_lock(); 6497 for_each_cpu(i, cpu_map) { 6498 sd = *per_cpu_ptr(d.sd, i); 6499 cpu_attach_domain(sd, d.rd, i); 6500 } 6501 rcu_read_unlock(); 6502 6503 ret = 0; 6504error: 6505 __free_domain_allocs(&d, alloc_state, cpu_map); 6506 return ret; 6507} 6508 6509static cpumask_var_t *doms_cur; /* current sched domains */ 6510static int ndoms_cur; /* number of sched domains in 'doms_cur' */ 6511static struct sched_domain_attr *dattr_cur; 6512 /* attribues of custom domains in 'doms_cur' */ 6513 6514/* 6515 * Special case: If a kmalloc of a doms_cur partition (array of 6516 * cpumask) fails, then fallback to a single sched domain, 6517 * as determined by the single cpumask fallback_doms. 6518 */ 6519static cpumask_var_t fallback_doms; 6520 6521/* 6522 * arch_update_cpu_topology lets virtualized architectures update the 6523 * cpu core maps. It is supposed to return 1 if the topology changed 6524 * or 0 if it stayed the same. 6525 */ 6526int __attribute__((weak)) arch_update_cpu_topology(void) 6527{ 6528 return 0; 6529} 6530 6531cpumask_var_t *alloc_sched_domains(unsigned int ndoms) 6532{ 6533 int i; 6534 cpumask_var_t *doms; 6535 6536 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL); 6537 if (!doms) 6538 return NULL; 6539 for (i = 0; i < ndoms; i++) { 6540 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) { 6541 free_sched_domains(doms, i); 6542 return NULL; 6543 } 6544 } 6545 return doms; 6546} 6547 6548void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms) 6549{ 6550 unsigned int i; 6551 for (i = 0; i < ndoms; i++) 6552 free_cpumask_var(doms[i]); 6553 kfree(doms); 6554} 6555 6556/* 6557 * Set up scheduler domains and groups. Callers must hold the hotplug lock. 6558 * For now this just excludes isolated cpus, but could be used to 6559 * exclude other special cases in the future. 6560 */ 6561static int init_sched_domains(const struct cpumask *cpu_map) 6562{ 6563 int err; 6564 6565 arch_update_cpu_topology(); 6566 ndoms_cur = 1; 6567 doms_cur = alloc_sched_domains(ndoms_cur); 6568 if (!doms_cur) 6569 doms_cur = &fallback_doms; 6570 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map); 6571 dattr_cur = NULL; 6572 err = build_sched_domains(doms_cur[0], NULL); 6573 register_sched_domain_sysctl(); 6574 6575 return err; 6576} 6577 6578/* 6579 * Detach sched domains from a group of cpus specified in cpu_map 6580 * These cpus will now be attached to the NULL domain 6581 */ 6582static void detach_destroy_domains(const struct cpumask *cpu_map) 6583{ 6584 int i; 6585 6586 rcu_read_lock(); 6587 for_each_cpu(i, cpu_map) 6588 cpu_attach_domain(NULL, &def_root_domain, i); 6589 rcu_read_unlock(); 6590} 6591 6592/* handle null as "default" */ 6593static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur, 6594 struct sched_domain_attr *new, int idx_new) 6595{ 6596 struct sched_domain_attr tmp; 6597 6598 /* fast path */ 6599 if (!new && !cur) 6600 return 1; 6601 6602 tmp = SD_ATTR_INIT; 6603 return !memcmp(cur ? (cur + idx_cur) : &tmp, 6604 new ? (new + idx_new) : &tmp, 6605 sizeof(struct sched_domain_attr)); 6606} 6607 6608/* 6609 * Partition sched domains as specified by the 'ndoms_new' 6610 * cpumasks in the array doms_new[] of cpumasks. This compares 6611 * doms_new[] to the current sched domain partitioning, doms_cur[]. 6612 * It destroys each deleted domain and builds each new domain. 6613 * 6614 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'. 6615 * The masks don't intersect (don't overlap.) We should setup one 6616 * sched domain for each mask. CPUs not in any of the cpumasks will 6617 * not be load balanced. If the same cpumask appears both in the 6618 * current 'doms_cur' domains and in the new 'doms_new', we can leave 6619 * it as it is. 6620 * 6621 * The passed in 'doms_new' should be allocated using 6622 * alloc_sched_domains. This routine takes ownership of it and will 6623 * free_sched_domains it when done with it. If the caller failed the 6624 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1, 6625 * and partition_sched_domains() will fallback to the single partition 6626 * 'fallback_doms', it also forces the domains to be rebuilt. 6627 * 6628 * If doms_new == NULL it will be replaced with cpu_online_mask. 6629 * ndoms_new == 0 is a special case for destroying existing domains, 6630 * and it will not create the default domain. 6631 * 6632 * Call with hotplug lock held 6633 */ 6634void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[], 6635 struct sched_domain_attr *dattr_new) 6636{ 6637 int i, j, n; 6638 int new_topology; 6639 6640 mutex_lock(&sched_domains_mutex); 6641 6642 /* always unregister in case we don't destroy any domains */ 6643 unregister_sched_domain_sysctl(); 6644 6645 /* Let architecture update cpu core mappings. */ 6646 new_topology = arch_update_cpu_topology(); 6647 6648 n = doms_new ? ndoms_new : 0; 6649 6650 /* Destroy deleted domains */ 6651 for (i = 0; i < ndoms_cur; i++) { 6652 for (j = 0; j < n && !new_topology; j++) { 6653 if (cpumask_equal(doms_cur[i], doms_new[j]) 6654 && dattrs_equal(dattr_cur, i, dattr_new, j)) 6655 goto match1; 6656 } 6657 /* no match - a current sched domain not in new doms_new[] */ 6658 detach_destroy_domains(doms_cur[i]); 6659match1: 6660 ; 6661 } 6662 6663 if (doms_new == NULL) { 6664 ndoms_cur = 0; 6665 doms_new = &fallback_doms; 6666 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map); 6667 WARN_ON_ONCE(dattr_new); 6668 } 6669 6670 /* Build new domains */ 6671 for (i = 0; i < ndoms_new; i++) { 6672 for (j = 0; j < ndoms_cur && !new_topology; j++) { 6673 if (cpumask_equal(doms_new[i], doms_cur[j]) 6674 && dattrs_equal(dattr_new, i, dattr_cur, j)) 6675 goto match2; 6676 } 6677 /* no match - add a new doms_new */ 6678 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL); 6679match2: 6680 ; 6681 } 6682 6683 /* Remember the new sched domains */ 6684 if (doms_cur != &fallback_doms) 6685 free_sched_domains(doms_cur, ndoms_cur); 6686 kfree(dattr_cur); /* kfree(NULL) is safe */ 6687 doms_cur = doms_new; 6688 dattr_cur = dattr_new; 6689 ndoms_cur = ndoms_new; 6690 6691 register_sched_domain_sysctl(); 6692 6693 mutex_unlock(&sched_domains_mutex); 6694} 6695 6696#if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT) 6697static void reinit_sched_domains(void) 6698{ 6699 get_online_cpus(); 6700 6701 /* Destroy domains first to force the rebuild */ 6702 partition_sched_domains(0, NULL, NULL); 6703 6704 rebuild_sched_domains(); 6705 put_online_cpus(); 6706} 6707 6708static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt) 6709{ 6710 unsigned int level = 0; 6711 6712 if (sscanf(buf, "%u", &level) != 1) 6713 return -EINVAL; 6714 6715 /* 6716 * level is always be positive so don't check for 6717 * level < POWERSAVINGS_BALANCE_NONE which is 0 6718 * What happens on 0 or 1 byte write, 6719 * need to check for count as well? 6720 */ 6721 6722 if (level >= MAX_POWERSAVINGS_BALANCE_LEVELS) 6723 return -EINVAL; 6724 6725 if (smt) 6726 sched_smt_power_savings = level; 6727 else 6728 sched_mc_power_savings = level; 6729 6730 reinit_sched_domains(); 6731 6732 return count; 6733} 6734 6735#ifdef CONFIG_SCHED_MC 6736static ssize_t sched_mc_power_savings_show(struct device *dev, 6737 struct device_attribute *attr, 6738 char *buf) 6739{ 6740 return sprintf(buf, "%u\n", sched_mc_power_savings); 6741} 6742static ssize_t sched_mc_power_savings_store(struct device *dev, 6743 struct device_attribute *attr, 6744 const char *buf, size_t count) 6745{ 6746 return sched_power_savings_store(buf, count, 0); 6747} 6748static DEVICE_ATTR(sched_mc_power_savings, 0644, 6749 sched_mc_power_savings_show, 6750 sched_mc_power_savings_store); 6751#endif 6752 6753#ifdef CONFIG_SCHED_SMT 6754static ssize_t sched_smt_power_savings_show(struct device *dev, 6755 struct device_attribute *attr, 6756 char *buf) 6757{ 6758 return sprintf(buf, "%u\n", sched_smt_power_savings); 6759} 6760static ssize_t sched_smt_power_savings_store(struct device *dev, 6761 struct device_attribute *attr, 6762 const char *buf, size_t count) 6763{ 6764 return sched_power_savings_store(buf, count, 1); 6765} 6766static DEVICE_ATTR(sched_smt_power_savings, 0644, 6767 sched_smt_power_savings_show, 6768 sched_smt_power_savings_store); 6769#endif 6770 6771int __init sched_create_sysfs_power_savings_entries(struct device *dev) 6772{ 6773 int err = 0; 6774 6775#ifdef CONFIG_SCHED_SMT 6776 if (smt_capable()) 6777 err = device_create_file(dev, &dev_attr_sched_smt_power_savings); 6778#endif 6779#ifdef CONFIG_SCHED_MC 6780 if (!err && mc_capable()) 6781 err = device_create_file(dev, &dev_attr_sched_mc_power_savings); 6782#endif 6783 return err; 6784} 6785#endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */ 6786 6787/* 6788 * Update cpusets according to cpu_active mask. If cpusets are 6789 * disabled, cpuset_update_active_cpus() becomes a simple wrapper 6790 * around partition_sched_domains(). 6791 */ 6792static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action, 6793 void *hcpu) 6794{ 6795 switch (action & ~CPU_TASKS_FROZEN) { 6796 case CPU_ONLINE: 6797 case CPU_DOWN_FAILED: 6798 cpuset_update_active_cpus(); 6799 return NOTIFY_OK; 6800 default: 6801 return NOTIFY_DONE; 6802 } 6803} 6804 6805static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action, 6806 void *hcpu) 6807{ 6808 switch (action & ~CPU_TASKS_FROZEN) { 6809 case CPU_DOWN_PREPARE: 6810 cpuset_update_active_cpus(); 6811 return NOTIFY_OK; 6812 default: 6813 return NOTIFY_DONE; 6814 } 6815} 6816 6817void __init sched_init_smp(void) 6818{ 6819 cpumask_var_t non_isolated_cpus; 6820 6821 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL); 6822 alloc_cpumask_var(&fallback_doms, GFP_KERNEL); 6823 6824 get_online_cpus(); 6825 mutex_lock(&sched_domains_mutex); 6826 init_sched_domains(cpu_active_mask); 6827 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map); 6828 if (cpumask_empty(non_isolated_cpus)) 6829 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus); 6830 mutex_unlock(&sched_domains_mutex); 6831 put_online_cpus(); 6832 6833 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE); 6834 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE); 6835 6836 /* RT runtime code needs to handle some hotplug events */ 6837 hotcpu_notifier(update_runtime, 0); 6838 6839 init_hrtick(); 6840 6841 /* Move init over to a non-isolated CPU */ 6842 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0) 6843 BUG(); 6844 sched_init_granularity(); 6845 free_cpumask_var(non_isolated_cpus); 6846 6847 init_sched_rt_class(); 6848} 6849#else 6850void __init sched_init_smp(void) 6851{ 6852 sched_init_granularity(); 6853} 6854#endif /* CONFIG_SMP */ 6855 6856const_debug unsigned int sysctl_timer_migration = 1; 6857 6858int in_sched_functions(unsigned long addr) 6859{ 6860 return in_lock_functions(addr) || 6861 (addr >= (unsigned long)__sched_text_start 6862 && addr < (unsigned long)__sched_text_end); 6863} 6864 6865#ifdef CONFIG_CGROUP_SCHED 6866struct task_group root_task_group; 6867#endif 6868 6869DECLARE_PER_CPU(cpumask_var_t, load_balance_tmpmask); 6870 6871void __init sched_init(void) 6872{ 6873 int i, j; 6874 unsigned long alloc_size = 0, ptr; 6875 6876#ifdef CONFIG_FAIR_GROUP_SCHED 6877 alloc_size += 2 * nr_cpu_ids * sizeof(void **); 6878#endif 6879#ifdef CONFIG_RT_GROUP_SCHED 6880 alloc_size += 2 * nr_cpu_ids * sizeof(void **); 6881#endif 6882#ifdef CONFIG_CPUMASK_OFFSTACK 6883 alloc_size += num_possible_cpus() * cpumask_size(); 6884#endif 6885 if (alloc_size) { 6886 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT); 6887 6888#ifdef CONFIG_FAIR_GROUP_SCHED 6889 root_task_group.se = (struct sched_entity **)ptr; 6890 ptr += nr_cpu_ids * sizeof(void **); 6891 6892 root_task_group.cfs_rq = (struct cfs_rq **)ptr; 6893 ptr += nr_cpu_ids * sizeof(void **); 6894 6895#endif /* CONFIG_FAIR_GROUP_SCHED */ 6896#ifdef CONFIG_RT_GROUP_SCHED 6897 root_task_group.rt_se = (struct sched_rt_entity **)ptr; 6898 ptr += nr_cpu_ids * sizeof(void **); 6899 6900 root_task_group.rt_rq = (struct rt_rq **)ptr; 6901 ptr += nr_cpu_ids * sizeof(void **); 6902 6903#endif /* CONFIG_RT_GROUP_SCHED */ 6904#ifdef CONFIG_CPUMASK_OFFSTACK 6905 for_each_possible_cpu(i) { 6906 per_cpu(load_balance_tmpmask, i) = (void *)ptr; 6907 ptr += cpumask_size(); 6908 } 6909#endif /* CONFIG_CPUMASK_OFFSTACK */ 6910 } 6911 6912#ifdef CONFIG_SMP 6913 init_defrootdomain(); 6914#endif 6915 6916 init_rt_bandwidth(&def_rt_bandwidth, 6917 global_rt_period(), global_rt_runtime()); 6918 6919#ifdef CONFIG_RT_GROUP_SCHED 6920 init_rt_bandwidth(&root_task_group.rt_bandwidth, 6921 global_rt_period(), global_rt_runtime()); 6922#endif /* CONFIG_RT_GROUP_SCHED */ 6923 6924#ifdef CONFIG_CGROUP_SCHED 6925 list_add(&root_task_group.list, &task_groups); 6926 INIT_LIST_HEAD(&root_task_group.children); 6927 INIT_LIST_HEAD(&root_task_group.siblings); 6928 autogroup_init(&init_task); 6929 6930#endif /* CONFIG_CGROUP_SCHED */ 6931 6932#ifdef CONFIG_CGROUP_CPUACCT 6933 root_cpuacct.cpustat = &kernel_cpustat; 6934 root_cpuacct.cpuusage = alloc_percpu(u64); 6935 /* Too early, not expected to fail */ 6936 BUG_ON(!root_cpuacct.cpuusage); 6937#endif 6938 for_each_possible_cpu(i) { 6939 struct rq *rq; 6940 6941 rq = cpu_rq(i); 6942 raw_spin_lock_init(&rq->lock); 6943 rq->nr_running = 0; 6944 rq->calc_load_active = 0; 6945 rq->calc_load_update = jiffies + LOAD_FREQ; 6946 init_cfs_rq(&rq->cfs); 6947 init_rt_rq(&rq->rt, rq); 6948#ifdef CONFIG_FAIR_GROUP_SCHED 6949 root_task_group.shares = ROOT_TASK_GROUP_LOAD; 6950 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list); 6951 /* 6952 * How much cpu bandwidth does root_task_group get? 6953 * 6954 * In case of task-groups formed thr' the cgroup filesystem, it 6955 * gets 100% of the cpu resources in the system. This overall 6956 * system cpu resource is divided among the tasks of 6957 * root_task_group and its child task-groups in a fair manner, 6958 * based on each entity's (task or task-group's) weight 6959 * (se->load.weight). 6960 * 6961 * In other words, if root_task_group has 10 tasks of weight 6962 * 1024) and two child groups A0 and A1 (of weight 1024 each), 6963 * then A0's share of the cpu resource is: 6964 * 6965 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33% 6966 * 6967 * We achieve this by letting root_task_group's tasks sit 6968 * directly in rq->cfs (i.e root_task_group->se[] = NULL). 6969 */ 6970 init_cfs_bandwidth(&root_task_group.cfs_bandwidth); 6971 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL); 6972#endif /* CONFIG_FAIR_GROUP_SCHED */ 6973 6974 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime; 6975#ifdef CONFIG_RT_GROUP_SCHED 6976 INIT_LIST_HEAD(&rq->leaf_rt_rq_list); 6977 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL); 6978#endif 6979 6980 for (j = 0; j < CPU_LOAD_IDX_MAX; j++) 6981 rq->cpu_load[j] = 0; 6982 6983 rq->last_load_update_tick = jiffies; 6984 6985#ifdef CONFIG_SMP 6986 rq->sd = NULL; 6987 rq->rd = NULL; 6988 rq->cpu_power = SCHED_POWER_SCALE; 6989 rq->post_schedule = 0; 6990 rq->active_balance = 0; 6991 rq->next_balance = jiffies; 6992 rq->push_cpu = 0; 6993 rq->cpu = i; 6994 rq->online = 0; 6995 rq->idle_stamp = 0; 6996 rq->avg_idle = 2*sysctl_sched_migration_cost; 6997 6998 INIT_LIST_HEAD(&rq->cfs_tasks); 6999 7000 rq_attach_root(rq, &def_root_domain); 7001#ifdef CONFIG_NO_HZ 7002 rq->nohz_flags = 0; 7003#endif 7004#endif 7005 init_rq_hrtick(rq); 7006 atomic_set(&rq->nr_iowait, 0); 7007 } 7008 7009 set_load_weight(&init_task); 7010 7011#ifdef CONFIG_PREEMPT_NOTIFIERS 7012 INIT_HLIST_HEAD(&init_task.preempt_notifiers); 7013#endif 7014 7015#ifdef CONFIG_RT_MUTEXES 7016 plist_head_init(&init_task.pi_waiters); 7017#endif 7018 7019 /* 7020 * The boot idle thread does lazy MMU switching as well: 7021 */ 7022 atomic_inc(&init_mm.mm_count); 7023 enter_lazy_tlb(&init_mm, current); 7024 7025 /* 7026 * Make us the idle thread. Technically, schedule() should not be 7027 * called from this thread, however somewhere below it might be, 7028 * but because we are the idle thread, we just pick up running again 7029 * when this runqueue becomes "idle". 7030 */ 7031 init_idle(current, smp_processor_id()); 7032 7033 calc_load_update = jiffies + LOAD_FREQ; 7034 7035 /* 7036 * During early bootup we pretend to be a normal task: 7037 */ 7038 current->sched_class = &fair_sched_class; 7039 7040#ifdef CONFIG_SMP 7041 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT); 7042 /* May be allocated at isolcpus cmdline parse time */ 7043 if (cpu_isolated_map == NULL) 7044 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT); 7045#endif 7046 init_sched_fair_class(); 7047 7048 scheduler_running = 1; 7049} 7050 7051#ifdef CONFIG_DEBUG_ATOMIC_SLEEP 7052static inline int preempt_count_equals(int preempt_offset) 7053{ 7054 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth(); 7055 7056 return (nested == preempt_offset); 7057} 7058 7059void __might_sleep(const char *file, int line, int preempt_offset) 7060{ 7061 static unsigned long prev_jiffy; /* ratelimiting */ 7062 7063 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */ 7064 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) || 7065 system_state != SYSTEM_RUNNING || oops_in_progress) 7066 return; 7067 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy) 7068 return; 7069 prev_jiffy = jiffies; 7070 7071 printk(KERN_ERR 7072 "BUG: sleeping function called from invalid context at %s:%d\n", 7073 file, line); 7074 printk(KERN_ERR 7075 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n", 7076 in_atomic(), irqs_disabled(), 7077 current->pid, current->comm); 7078 7079 debug_show_held_locks(current); 7080 if (irqs_disabled()) 7081 print_irqtrace_events(current); 7082 dump_stack(); 7083} 7084EXPORT_SYMBOL(__might_sleep); 7085#endif 7086 7087#ifdef CONFIG_MAGIC_SYSRQ 7088static void normalize_task(struct rq *rq, struct task_struct *p) 7089{ 7090 const struct sched_class *prev_class = p->sched_class; 7091 int old_prio = p->prio; 7092 int on_rq; 7093 7094 on_rq = p->on_rq; 7095 if (on_rq) 7096 dequeue_task(rq, p, 0); 7097 __setscheduler(rq, p, SCHED_NORMAL, 0); 7098 if (on_rq) { 7099 enqueue_task(rq, p, 0); 7100 resched_task(rq->curr); 7101 } 7102 7103 check_class_changed(rq, p, prev_class, old_prio); 7104} 7105 7106void normalize_rt_tasks(void) 7107{ 7108 struct task_struct *g, *p; 7109 unsigned long flags; 7110 struct rq *rq; 7111 7112 read_lock_irqsave(&tasklist_lock, flags); 7113 do_each_thread(g, p) { 7114 /* 7115 * Only normalize user tasks: 7116 */ 7117 if (!p->mm) 7118 continue; 7119 7120 p->se.exec_start = 0; 7121#ifdef CONFIG_SCHEDSTATS 7122 p->se.statistics.wait_start = 0; 7123 p->se.statistics.sleep_start = 0; 7124 p->se.statistics.block_start = 0; 7125#endif 7126 7127 if (!rt_task(p)) { 7128 /* 7129 * Renice negative nice level userspace 7130 * tasks back to 0: 7131 */ 7132 if (TASK_NICE(p) < 0 && p->mm) 7133 set_user_nice(p, 0); 7134 continue; 7135 } 7136 7137 raw_spin_lock(&p->pi_lock); 7138 rq = __task_rq_lock(p); 7139 7140 normalize_task(rq, p); 7141 7142 __task_rq_unlock(rq); 7143 raw_spin_unlock(&p->pi_lock); 7144 } while_each_thread(g, p); 7145 7146 read_unlock_irqrestore(&tasklist_lock, flags); 7147} 7148 7149#endif /* CONFIG_MAGIC_SYSRQ */ 7150 7151#if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) 7152/* 7153 * These functions are only useful for the IA64 MCA handling, or kdb. 7154 * 7155 * They can only be called when the whole system has been 7156 * stopped - every CPU needs to be quiescent, and no scheduling 7157 * activity can take place. Using them for anything else would 7158 * be a serious bug, and as a result, they aren't even visible 7159 * under any other configuration. 7160 */ 7161 7162/** 7163 * curr_task - return the current task for a given cpu. 7164 * @cpu: the processor in question. 7165 * 7166 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED! 7167 */ 7168struct task_struct *curr_task(int cpu) 7169{ 7170 return cpu_curr(cpu); 7171} 7172 7173#endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */ 7174 7175#ifdef CONFIG_IA64 7176/** 7177 * set_curr_task - set the current task for a given cpu. 7178 * @cpu: the processor in question. 7179 * @p: the task pointer to set. 7180 * 7181 * Description: This function must only be used when non-maskable interrupts 7182 * are serviced on a separate stack. It allows the architecture to switch the 7183 * notion of the current task on a cpu in a non-blocking manner. This function 7184 * must be called with all CPU's synchronized, and interrupts disabled, the 7185 * and caller must save the original value of the current task (see 7186 * curr_task() above) and restore that value before reenabling interrupts and 7187 * re-starting the system. 7188 * 7189 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED! 7190 */ 7191void set_curr_task(int cpu, struct task_struct *p) 7192{ 7193 cpu_curr(cpu) = p; 7194} 7195 7196#endif 7197 7198#ifdef CONFIG_CGROUP_SCHED 7199/* task_group_lock serializes the addition/removal of task groups */ 7200static DEFINE_SPINLOCK(task_group_lock); 7201 7202static void free_sched_group(struct task_group *tg) 7203{ 7204 free_fair_sched_group(tg); 7205 free_rt_sched_group(tg); 7206 autogroup_free(tg); 7207 kfree(tg); 7208} 7209 7210/* allocate runqueue etc for a new task group */ 7211struct task_group *sched_create_group(struct task_group *parent) 7212{ 7213 struct task_group *tg; 7214 unsigned long flags; 7215 7216 tg = kzalloc(sizeof(*tg), GFP_KERNEL); 7217 if (!tg) 7218 return ERR_PTR(-ENOMEM); 7219 7220 if (!alloc_fair_sched_group(tg, parent)) 7221 goto err; 7222 7223 if (!alloc_rt_sched_group(tg, parent)) 7224 goto err; 7225 7226 spin_lock_irqsave(&task_group_lock, flags); 7227 list_add_rcu(&tg->list, &task_groups); 7228 7229 WARN_ON(!parent); /* root should already exist */ 7230 7231 tg->parent = parent; 7232 INIT_LIST_HEAD(&tg->children); 7233 list_add_rcu(&tg->siblings, &parent->children); 7234 spin_unlock_irqrestore(&task_group_lock, flags); 7235 7236 return tg; 7237 7238err: 7239 free_sched_group(tg); 7240 return ERR_PTR(-ENOMEM); 7241} 7242 7243/* rcu callback to free various structures associated with a task group */ 7244static void free_sched_group_rcu(struct rcu_head *rhp) 7245{ 7246 /* now it should be safe to free those cfs_rqs */ 7247 free_sched_group(container_of(rhp, struct task_group, rcu)); 7248} 7249 7250/* Destroy runqueue etc associated with a task group */ 7251void sched_destroy_group(struct task_group *tg) 7252{ 7253 unsigned long flags; 7254 int i; 7255 7256 /* end participation in shares distribution */ 7257 for_each_possible_cpu(i) 7258 unregister_fair_sched_group(tg, i); 7259 7260 spin_lock_irqsave(&task_group_lock, flags); 7261 list_del_rcu(&tg->list); 7262 list_del_rcu(&tg->siblings); 7263 spin_unlock_irqrestore(&task_group_lock, flags); 7264 7265 /* wait for possible concurrent references to cfs_rqs complete */ 7266 call_rcu(&tg->rcu, free_sched_group_rcu); 7267} 7268 7269/* change task's runqueue when it moves between groups. 7270 * The caller of this function should have put the task in its new group 7271 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to 7272 * reflect its new group. 7273 */ 7274void sched_move_task(struct task_struct *tsk) 7275{ 7276 int on_rq, running; 7277 unsigned long flags; 7278 struct rq *rq; 7279 7280 rq = task_rq_lock(tsk, &flags); 7281 7282 running = task_current(rq, tsk); 7283 on_rq = tsk->on_rq; 7284 7285 if (on_rq) 7286 dequeue_task(rq, tsk, 0); 7287 if (unlikely(running)) 7288 tsk->sched_class->put_prev_task(rq, tsk); 7289 7290#ifdef CONFIG_FAIR_GROUP_SCHED 7291 if (tsk->sched_class->task_move_group) 7292 tsk->sched_class->task_move_group(tsk, on_rq); 7293 else 7294#endif 7295 set_task_rq(tsk, task_cpu(tsk)); 7296 7297 if (unlikely(running)) 7298 tsk->sched_class->set_curr_task(rq); 7299 if (on_rq) 7300 enqueue_task(rq, tsk, 0); 7301 7302 task_rq_unlock(rq, tsk, &flags); 7303} 7304#endif /* CONFIG_CGROUP_SCHED */ 7305 7306#if defined(CONFIG_RT_GROUP_SCHED) || defined(CONFIG_CFS_BANDWIDTH) 7307static unsigned long to_ratio(u64 period, u64 runtime) 7308{ 7309 if (runtime == RUNTIME_INF) 7310 return 1ULL << 20; 7311 7312 return div64_u64(runtime << 20, period); 7313} 7314#endif 7315 7316#ifdef CONFIG_RT_GROUP_SCHED 7317/* 7318 * Ensure that the real time constraints are schedulable. 7319 */ 7320static DEFINE_MUTEX(rt_constraints_mutex); 7321 7322/* Must be called with tasklist_lock held */ 7323static inline int tg_has_rt_tasks(struct task_group *tg) 7324{ 7325 struct task_struct *g, *p; 7326 7327 do_each_thread(g, p) { 7328 if (rt_task(p) && task_rq(p)->rt.tg == tg) 7329 return 1; 7330 } while_each_thread(g, p); 7331 7332 return 0; 7333} 7334 7335struct rt_schedulable_data { 7336 struct task_group *tg; 7337 u64 rt_period; 7338 u64 rt_runtime; 7339}; 7340 7341static int tg_rt_schedulable(struct task_group *tg, void *data) 7342{ 7343 struct rt_schedulable_data *d = data; 7344 struct task_group *child; 7345 unsigned long total, sum = 0; 7346 u64 period, runtime; 7347 7348 period = ktime_to_ns(tg->rt_bandwidth.rt_period); 7349 runtime = tg->rt_bandwidth.rt_runtime; 7350 7351 if (tg == d->tg) { 7352 period = d->rt_period; 7353 runtime = d->rt_runtime; 7354 } 7355 7356 /* 7357 * Cannot have more runtime than the period. 7358 */ 7359 if (runtime > period && runtime != RUNTIME_INF) 7360 return -EINVAL; 7361 7362 /* 7363 * Ensure we don't starve existing RT tasks. 7364 */ 7365 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg)) 7366 return -EBUSY; 7367 7368 total = to_ratio(period, runtime); 7369 7370 /* 7371 * Nobody can have more than the global setting allows. 7372 */ 7373 if (total > to_ratio(global_rt_period(), global_rt_runtime())) 7374 return -EINVAL; 7375 7376 /* 7377 * The sum of our children's runtime should not exceed our own. 7378 */ 7379 list_for_each_entry_rcu(child, &tg->children, siblings) { 7380 period = ktime_to_ns(child->rt_bandwidth.rt_period); 7381 runtime = child->rt_bandwidth.rt_runtime; 7382 7383 if (child == d->tg) { 7384 period = d->rt_period; 7385 runtime = d->rt_runtime; 7386 } 7387 7388 sum += to_ratio(period, runtime); 7389 } 7390 7391 if (sum > total) 7392 return -EINVAL; 7393 7394 return 0; 7395} 7396 7397static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime) 7398{ 7399 int ret; 7400 7401 struct rt_schedulable_data data = { 7402 .tg = tg, 7403 .rt_period = period, 7404 .rt_runtime = runtime, 7405 }; 7406 7407 rcu_read_lock(); 7408 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data); 7409 rcu_read_unlock(); 7410 7411 return ret; 7412} 7413 7414static int tg_set_rt_bandwidth(struct task_group *tg, 7415 u64 rt_period, u64 rt_runtime) 7416{ 7417 int i, err = 0; 7418 7419 mutex_lock(&rt_constraints_mutex); 7420 read_lock(&tasklist_lock); 7421 err = __rt_schedulable(tg, rt_period, rt_runtime); 7422 if (err) 7423 goto unlock; 7424 7425 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock); 7426 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period); 7427 tg->rt_bandwidth.rt_runtime = rt_runtime; 7428 7429 for_each_possible_cpu(i) { 7430 struct rt_rq *rt_rq = tg->rt_rq[i]; 7431 7432 raw_spin_lock(&rt_rq->rt_runtime_lock); 7433 rt_rq->rt_runtime = rt_runtime; 7434 raw_spin_unlock(&rt_rq->rt_runtime_lock); 7435 } 7436 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock); 7437unlock: 7438 read_unlock(&tasklist_lock); 7439 mutex_unlock(&rt_constraints_mutex); 7440 7441 return err; 7442} 7443 7444int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us) 7445{ 7446 u64 rt_runtime, rt_period; 7447 7448 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period); 7449 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC; 7450 if (rt_runtime_us < 0) 7451 rt_runtime = RUNTIME_INF; 7452 7453 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime); 7454} 7455 7456long sched_group_rt_runtime(struct task_group *tg) 7457{ 7458 u64 rt_runtime_us; 7459 7460 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF) 7461 return -1; 7462 7463 rt_runtime_us = tg->rt_bandwidth.rt_runtime; 7464 do_div(rt_runtime_us, NSEC_PER_USEC); 7465 return rt_runtime_us; 7466} 7467 7468int sched_group_set_rt_period(struct task_group *tg, long rt_period_us) 7469{ 7470 u64 rt_runtime, rt_period; 7471 7472 rt_period = (u64)rt_period_us * NSEC_PER_USEC; 7473 rt_runtime = tg->rt_bandwidth.rt_runtime; 7474 7475 if (rt_period == 0) 7476 return -EINVAL; 7477 7478 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime); 7479} 7480 7481long sched_group_rt_period(struct task_group *tg) 7482{ 7483 u64 rt_period_us; 7484 7485 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period); 7486 do_div(rt_period_us, NSEC_PER_USEC); 7487 return rt_period_us; 7488} 7489 7490static int sched_rt_global_constraints(void) 7491{ 7492 u64 runtime, period; 7493 int ret = 0; 7494 7495 if (sysctl_sched_rt_period <= 0) 7496 return -EINVAL; 7497 7498 runtime = global_rt_runtime(); 7499 period = global_rt_period(); 7500 7501 /* 7502 * Sanity check on the sysctl variables. 7503 */ 7504 if (runtime > period && runtime != RUNTIME_INF) 7505 return -EINVAL; 7506 7507 mutex_lock(&rt_constraints_mutex); 7508 read_lock(&tasklist_lock); 7509 ret = __rt_schedulable(NULL, 0, 0); 7510 read_unlock(&tasklist_lock); 7511 mutex_unlock(&rt_constraints_mutex); 7512 7513 return ret; 7514} 7515 7516int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk) 7517{ 7518 /* Don't accept realtime tasks when there is no way for them to run */ 7519 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0) 7520 return 0; 7521 7522 return 1; 7523} 7524 7525#else /* !CONFIG_RT_GROUP_SCHED */ 7526static int sched_rt_global_constraints(void) 7527{ 7528 unsigned long flags; 7529 int i; 7530 7531 if (sysctl_sched_rt_period <= 0) 7532 return -EINVAL; 7533 7534 /* 7535 * There's always some RT tasks in the root group 7536 * -- migration, kstopmachine etc.. 7537 */ 7538 if (sysctl_sched_rt_runtime == 0) 7539 return -EBUSY; 7540 7541 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags); 7542 for_each_possible_cpu(i) { 7543 struct rt_rq *rt_rq = &cpu_rq(i)->rt; 7544 7545 raw_spin_lock(&rt_rq->rt_runtime_lock); 7546 rt_rq->rt_runtime = global_rt_runtime(); 7547 raw_spin_unlock(&rt_rq->rt_runtime_lock); 7548 } 7549 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags); 7550 7551 return 0; 7552} 7553#endif /* CONFIG_RT_GROUP_SCHED */ 7554 7555int sched_rt_handler(struct ctl_table *table, int write, 7556 void __user *buffer, size_t *lenp, 7557 loff_t *ppos) 7558{ 7559 int ret; 7560 int old_period, old_runtime; 7561 static DEFINE_MUTEX(mutex); 7562 7563 mutex_lock(&mutex); 7564 old_period = sysctl_sched_rt_period; 7565 old_runtime = sysctl_sched_rt_runtime; 7566 7567 ret = proc_dointvec(table, write, buffer, lenp, ppos); 7568 7569 if (!ret && write) { 7570 ret = sched_rt_global_constraints(); 7571 if (ret) { 7572 sysctl_sched_rt_period = old_period; 7573 sysctl_sched_rt_runtime = old_runtime; 7574 } else { 7575 def_rt_bandwidth.rt_runtime = global_rt_runtime(); 7576 def_rt_bandwidth.rt_period = 7577 ns_to_ktime(global_rt_period()); 7578 } 7579 } 7580 mutex_unlock(&mutex); 7581 7582 return ret; 7583} 7584 7585#ifdef CONFIG_CGROUP_SCHED 7586 7587/* return corresponding task_group object of a cgroup */ 7588static inline struct task_group *cgroup_tg(struct cgroup *cgrp) 7589{ 7590 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id), 7591 struct task_group, css); 7592} 7593 7594static struct cgroup_subsys_state *cpu_cgroup_create(struct cgroup *cgrp) 7595{ 7596 struct task_group *tg, *parent; 7597 7598 if (!cgrp->parent) { 7599 /* This is early initialization for the top cgroup */ 7600 return &root_task_group.css; 7601 } 7602 7603 parent = cgroup_tg(cgrp->parent); 7604 tg = sched_create_group(parent); 7605 if (IS_ERR(tg)) 7606 return ERR_PTR(-ENOMEM); 7607 7608 return &tg->css; 7609} 7610 7611static void cpu_cgroup_destroy(struct cgroup *cgrp) 7612{ 7613 struct task_group *tg = cgroup_tg(cgrp); 7614 7615 sched_destroy_group(tg); 7616} 7617 7618static int cpu_cgroup_can_attach(struct cgroup *cgrp, 7619 struct cgroup_taskset *tset) 7620{ 7621 struct task_struct *task; 7622 7623 cgroup_taskset_for_each(task, cgrp, tset) { 7624#ifdef CONFIG_RT_GROUP_SCHED 7625 if (!sched_rt_can_attach(cgroup_tg(cgrp), task)) 7626 return -EINVAL; 7627#else 7628 /* We don't support RT-tasks being in separate groups */ 7629 if (task->sched_class != &fair_sched_class) 7630 return -EINVAL; 7631#endif 7632 } 7633 return 0; 7634} 7635 7636static void cpu_cgroup_attach(struct cgroup *cgrp, 7637 struct cgroup_taskset *tset) 7638{ 7639 struct task_struct *task; 7640 7641 cgroup_taskset_for_each(task, cgrp, tset) 7642 sched_move_task(task); 7643} 7644 7645static void 7646cpu_cgroup_exit(struct cgroup *cgrp, struct cgroup *old_cgrp, 7647 struct task_struct *task) 7648{ 7649 /* 7650 * cgroup_exit() is called in the copy_process() failure path. 7651 * Ignore this case since the task hasn't ran yet, this avoids 7652 * trying to poke a half freed task state from generic code. 7653 */ 7654 if (!(task->flags & PF_EXITING)) 7655 return; 7656 7657 sched_move_task(task); 7658} 7659 7660#ifdef CONFIG_FAIR_GROUP_SCHED 7661static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype, 7662 u64 shareval) 7663{ 7664 return sched_group_set_shares(cgroup_tg(cgrp), scale_load(shareval)); 7665} 7666 7667static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft) 7668{ 7669 struct task_group *tg = cgroup_tg(cgrp); 7670 7671 return (u64) scale_load_down(tg->shares); 7672} 7673 7674#ifdef CONFIG_CFS_BANDWIDTH 7675static DEFINE_MUTEX(cfs_constraints_mutex); 7676 7677const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */ 7678const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */ 7679 7680static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime); 7681 7682static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota) 7683{ 7684 int i, ret = 0, runtime_enabled, runtime_was_enabled; 7685 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth; 7686 7687 if (tg == &root_task_group) 7688 return -EINVAL; 7689 7690 /* 7691 * Ensure we have at some amount of bandwidth every period. This is 7692 * to prevent reaching a state of large arrears when throttled via 7693 * entity_tick() resulting in prolonged exit starvation. 7694 */ 7695 if (quota < min_cfs_quota_period || period < min_cfs_quota_period) 7696 return -EINVAL; 7697 7698 /* 7699 * Likewise, bound things on the otherside by preventing insane quota 7700 * periods. This also allows us to normalize in computing quota 7701 * feasibility. 7702 */ 7703 if (period > max_cfs_quota_period) 7704 return -EINVAL; 7705 7706 mutex_lock(&cfs_constraints_mutex); 7707 ret = __cfs_schedulable(tg, period, quota); 7708 if (ret) 7709 goto out_unlock; 7710 7711 runtime_enabled = quota != RUNTIME_INF; 7712 runtime_was_enabled = cfs_b->quota != RUNTIME_INF; 7713 account_cfs_bandwidth_used(runtime_enabled, runtime_was_enabled); 7714 raw_spin_lock_irq(&cfs_b->lock); 7715 cfs_b->period = ns_to_ktime(period); 7716 cfs_b->quota = quota; 7717 7718 __refill_cfs_bandwidth_runtime(cfs_b); 7719 /* restart the period timer (if active) to handle new period expiry */ 7720 if (runtime_enabled && cfs_b->timer_active) { 7721 /* force a reprogram */ 7722 cfs_b->timer_active = 0; 7723 __start_cfs_bandwidth(cfs_b); 7724 } 7725 raw_spin_unlock_irq(&cfs_b->lock); 7726 7727 for_each_possible_cpu(i) { 7728 struct cfs_rq *cfs_rq = tg->cfs_rq[i]; 7729 struct rq *rq = cfs_rq->rq; 7730 7731 raw_spin_lock_irq(&rq->lock); 7732 cfs_rq->runtime_enabled = runtime_enabled; 7733 cfs_rq->runtime_remaining = 0; 7734 7735 if (cfs_rq->throttled) 7736 unthrottle_cfs_rq(cfs_rq); 7737 raw_spin_unlock_irq(&rq->lock); 7738 } 7739out_unlock: 7740 mutex_unlock(&cfs_constraints_mutex); 7741 7742 return ret; 7743} 7744 7745int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us) 7746{ 7747 u64 quota, period; 7748 7749 period = ktime_to_ns(tg->cfs_bandwidth.period); 7750 if (cfs_quota_us < 0) 7751 quota = RUNTIME_INF; 7752 else 7753 quota = (u64)cfs_quota_us * NSEC_PER_USEC; 7754 7755 return tg_set_cfs_bandwidth(tg, period, quota); 7756} 7757 7758long tg_get_cfs_quota(struct task_group *tg) 7759{ 7760 u64 quota_us; 7761 7762 if (tg->cfs_bandwidth.quota == RUNTIME_INF) 7763 return -1; 7764 7765 quota_us = tg->cfs_bandwidth.quota; 7766 do_div(quota_us, NSEC_PER_USEC); 7767 7768 return quota_us; 7769} 7770 7771int tg_set_cfs_period(struct task_group *tg, long cfs_period_us) 7772{ 7773 u64 quota, period; 7774 7775 period = (u64)cfs_period_us * NSEC_PER_USEC; 7776 quota = tg->cfs_bandwidth.quota; 7777 7778 return tg_set_cfs_bandwidth(tg, period, quota); 7779} 7780 7781long tg_get_cfs_period(struct task_group *tg) 7782{ 7783 u64 cfs_period_us; 7784 7785 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period); 7786 do_div(cfs_period_us, NSEC_PER_USEC); 7787 7788 return cfs_period_us; 7789} 7790 7791static s64 cpu_cfs_quota_read_s64(struct cgroup *cgrp, struct cftype *cft) 7792{ 7793 return tg_get_cfs_quota(cgroup_tg(cgrp)); 7794} 7795 7796static int cpu_cfs_quota_write_s64(struct cgroup *cgrp, struct cftype *cftype, 7797 s64 cfs_quota_us) 7798{ 7799 return tg_set_cfs_quota(cgroup_tg(cgrp), cfs_quota_us); 7800} 7801 7802static u64 cpu_cfs_period_read_u64(struct cgroup *cgrp, struct cftype *cft) 7803{ 7804 return tg_get_cfs_period(cgroup_tg(cgrp)); 7805} 7806 7807static int cpu_cfs_period_write_u64(struct cgroup *cgrp, struct cftype *cftype, 7808 u64 cfs_period_us) 7809{ 7810 return tg_set_cfs_period(cgroup_tg(cgrp), cfs_period_us); 7811} 7812 7813struct cfs_schedulable_data { 7814 struct task_group *tg; 7815 u64 period, quota; 7816}; 7817 7818/* 7819 * normalize group quota/period to be quota/max_period 7820 * note: units are usecs 7821 */ 7822static u64 normalize_cfs_quota(struct task_group *tg, 7823 struct cfs_schedulable_data *d) 7824{ 7825 u64 quota, period; 7826 7827 if (tg == d->tg) { 7828 period = d->period; 7829 quota = d->quota; 7830 } else { 7831 period = tg_get_cfs_period(tg); 7832 quota = tg_get_cfs_quota(tg); 7833 } 7834 7835 /* note: these should typically be equivalent */ 7836 if (quota == RUNTIME_INF || quota == -1) 7837 return RUNTIME_INF; 7838 7839 return to_ratio(period, quota); 7840} 7841 7842static int tg_cfs_schedulable_down(struct task_group *tg, void *data) 7843{ 7844 struct cfs_schedulable_data *d = data; 7845 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth; 7846 s64 quota = 0, parent_quota = -1; 7847 7848 if (!tg->parent) { 7849 quota = RUNTIME_INF; 7850 } else { 7851 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth; 7852 7853 quota = normalize_cfs_quota(tg, d); 7854 parent_quota = parent_b->hierarchal_quota; 7855 7856 /* 7857 * ensure max(child_quota) <= parent_quota, inherit when no 7858 * limit is set 7859 */ 7860 if (quota == RUNTIME_INF) 7861 quota = parent_quota; 7862 else if (parent_quota != RUNTIME_INF && quota > parent_quota) 7863 return -EINVAL; 7864 } 7865 cfs_b->hierarchal_quota = quota; 7866 7867 return 0; 7868} 7869 7870static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota) 7871{ 7872 int ret; 7873 struct cfs_schedulable_data data = { 7874 .tg = tg, 7875 .period = period, 7876 .quota = quota, 7877 }; 7878 7879 if (quota != RUNTIME_INF) { 7880 do_div(data.period, NSEC_PER_USEC); 7881 do_div(data.quota, NSEC_PER_USEC); 7882 } 7883 7884 rcu_read_lock(); 7885 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data); 7886 rcu_read_unlock(); 7887 7888 return ret; 7889} 7890 7891static int cpu_stats_show(struct cgroup *cgrp, struct cftype *cft, 7892 struct cgroup_map_cb *cb) 7893{ 7894 struct task_group *tg = cgroup_tg(cgrp); 7895 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth; 7896 7897 cb->fill(cb, "nr_periods", cfs_b->nr_periods); 7898 cb->fill(cb, "nr_throttled", cfs_b->nr_throttled); 7899 cb->fill(cb, "throttled_time", cfs_b->throttled_time); 7900 7901 return 0; 7902} 7903#endif /* CONFIG_CFS_BANDWIDTH */ 7904#endif /* CONFIG_FAIR_GROUP_SCHED */ 7905 7906#ifdef CONFIG_RT_GROUP_SCHED 7907static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft, 7908 s64 val) 7909{ 7910 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val); 7911} 7912 7913static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft) 7914{ 7915 return sched_group_rt_runtime(cgroup_tg(cgrp)); 7916} 7917 7918static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype, 7919 u64 rt_period_us) 7920{ 7921 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us); 7922} 7923 7924static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft) 7925{ 7926 return sched_group_rt_period(cgroup_tg(cgrp)); 7927} 7928#endif /* CONFIG_RT_GROUP_SCHED */ 7929 7930static struct cftype cpu_files[] = { 7931#ifdef CONFIG_FAIR_GROUP_SCHED 7932 { 7933 .name = "shares", 7934 .read_u64 = cpu_shares_read_u64, 7935 .write_u64 = cpu_shares_write_u64, 7936 }, 7937#endif 7938#ifdef CONFIG_CFS_BANDWIDTH 7939 { 7940 .name = "cfs_quota_us", 7941 .read_s64 = cpu_cfs_quota_read_s64, 7942 .write_s64 = cpu_cfs_quota_write_s64, 7943 }, 7944 { 7945 .name = "cfs_period_us", 7946 .read_u64 = cpu_cfs_period_read_u64, 7947 .write_u64 = cpu_cfs_period_write_u64, 7948 }, 7949 { 7950 .name = "stat", 7951 .read_map = cpu_stats_show, 7952 }, 7953#endif 7954#ifdef CONFIG_RT_GROUP_SCHED 7955 { 7956 .name = "rt_runtime_us", 7957 .read_s64 = cpu_rt_runtime_read, 7958 .write_s64 = cpu_rt_runtime_write, 7959 }, 7960 { 7961 .name = "rt_period_us", 7962 .read_u64 = cpu_rt_period_read_uint, 7963 .write_u64 = cpu_rt_period_write_uint, 7964 }, 7965#endif 7966}; 7967 7968static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont) 7969{ 7970 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files)); 7971} 7972 7973struct cgroup_subsys cpu_cgroup_subsys = { 7974 .name = "cpu", 7975 .create = cpu_cgroup_create, 7976 .destroy = cpu_cgroup_destroy, 7977 .can_attach = cpu_cgroup_can_attach, 7978 .attach = cpu_cgroup_attach, 7979 .exit = cpu_cgroup_exit, 7980 .populate = cpu_cgroup_populate, 7981 .subsys_id = cpu_cgroup_subsys_id, 7982 .early_init = 1, 7983}; 7984 7985#endif /* CONFIG_CGROUP_SCHED */ 7986 7987#ifdef CONFIG_CGROUP_CPUACCT 7988 7989/* 7990 * CPU accounting code for task groups. 7991 * 7992 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh 7993 * (balbir@in.ibm.com). 7994 */ 7995 7996/* create a new cpu accounting group */ 7997static struct cgroup_subsys_state *cpuacct_create(struct cgroup *cgrp) 7998{ 7999 struct cpuacct *ca; 8000 8001 if (!cgrp->parent) 8002 return &root_cpuacct.css; 8003 8004 ca = kzalloc(sizeof(*ca), GFP_KERNEL); 8005 if (!ca) 8006 goto out; 8007 8008 ca->cpuusage = alloc_percpu(u64); 8009 if (!ca->cpuusage) 8010 goto out_free_ca; 8011 8012 ca->cpustat = alloc_percpu(struct kernel_cpustat); 8013 if (!ca->cpustat) 8014 goto out_free_cpuusage; 8015 8016 return &ca->css; 8017 8018out_free_cpuusage: 8019 free_percpu(ca->cpuusage); 8020out_free_ca: 8021 kfree(ca); 8022out: 8023 return ERR_PTR(-ENOMEM); 8024} 8025 8026/* destroy an existing cpu accounting group */ 8027static void cpuacct_destroy(struct cgroup *cgrp) 8028{ 8029 struct cpuacct *ca = cgroup_ca(cgrp); 8030 8031 free_percpu(ca->cpustat); 8032 free_percpu(ca->cpuusage); 8033 kfree(ca); 8034} 8035 8036static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu) 8037{ 8038 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu); 8039 u64 data; 8040 8041#ifndef CONFIG_64BIT 8042 /* 8043 * Take rq->lock to make 64-bit read safe on 32-bit platforms. 8044 */ 8045 raw_spin_lock_irq(&cpu_rq(cpu)->lock); 8046 data = *cpuusage; 8047 raw_spin_unlock_irq(&cpu_rq(cpu)->lock); 8048#else 8049 data = *cpuusage; 8050#endif 8051 8052 return data; 8053} 8054 8055static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val) 8056{ 8057 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu); 8058 8059#ifndef CONFIG_64BIT 8060 /* 8061 * Take rq->lock to make 64-bit write safe on 32-bit platforms. 8062 */ 8063 raw_spin_lock_irq(&cpu_rq(cpu)->lock); 8064 *cpuusage = val; 8065 raw_spin_unlock_irq(&cpu_rq(cpu)->lock); 8066#else 8067 *cpuusage = val; 8068#endif 8069} 8070 8071/* return total cpu usage (in nanoseconds) of a group */ 8072static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft) 8073{ 8074 struct cpuacct *ca = cgroup_ca(cgrp); 8075 u64 totalcpuusage = 0; 8076 int i; 8077 8078 for_each_present_cpu(i) 8079 totalcpuusage += cpuacct_cpuusage_read(ca, i); 8080 8081 return totalcpuusage; 8082} 8083 8084static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype, 8085 u64 reset) 8086{ 8087 struct cpuacct *ca = cgroup_ca(cgrp); 8088 int err = 0; 8089 int i; 8090 8091 if (reset) { 8092 err = -EINVAL; 8093 goto out; 8094 } 8095 8096 for_each_present_cpu(i) 8097 cpuacct_cpuusage_write(ca, i, 0); 8098 8099out: 8100 return err; 8101} 8102 8103static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft, 8104 struct seq_file *m) 8105{ 8106 struct cpuacct *ca = cgroup_ca(cgroup); 8107 u64 percpu; 8108 int i; 8109 8110 for_each_present_cpu(i) { 8111 percpu = cpuacct_cpuusage_read(ca, i); 8112 seq_printf(m, "%llu ", (unsigned long long) percpu); 8113 } 8114 seq_printf(m, "\n"); 8115 return 0; 8116} 8117 8118static const char *cpuacct_stat_desc[] = { 8119 [CPUACCT_STAT_USER] = "user", 8120 [CPUACCT_STAT_SYSTEM] = "system", 8121}; 8122 8123static int cpuacct_stats_show(struct cgroup *cgrp, struct cftype *cft, 8124 struct cgroup_map_cb *cb) 8125{ 8126 struct cpuacct *ca = cgroup_ca(cgrp); 8127 int cpu; 8128 s64 val = 0; 8129 8130 for_each_online_cpu(cpu) { 8131 struct kernel_cpustat *kcpustat = per_cpu_ptr(ca->cpustat, cpu); 8132 val += kcpustat->cpustat[CPUTIME_USER]; 8133 val += kcpustat->cpustat[CPUTIME_NICE]; 8134 } 8135 val = cputime64_to_clock_t(val); 8136 cb->fill(cb, cpuacct_stat_desc[CPUACCT_STAT_USER], val); 8137 8138 val = 0; 8139 for_each_online_cpu(cpu) { 8140 struct kernel_cpustat *kcpustat = per_cpu_ptr(ca->cpustat, cpu); 8141 val += kcpustat->cpustat[CPUTIME_SYSTEM]; 8142 val += kcpustat->cpustat[CPUTIME_IRQ]; 8143 val += kcpustat->cpustat[CPUTIME_SOFTIRQ]; 8144 } 8145 8146 val = cputime64_to_clock_t(val); 8147 cb->fill(cb, cpuacct_stat_desc[CPUACCT_STAT_SYSTEM], val); 8148 8149 return 0; 8150} 8151 8152static struct cftype files[] = { 8153 { 8154 .name = "usage", 8155 .read_u64 = cpuusage_read, 8156 .write_u64 = cpuusage_write, 8157 }, 8158 { 8159 .name = "usage_percpu", 8160 .read_seq_string = cpuacct_percpu_seq_read, 8161 }, 8162 { 8163 .name = "stat", 8164 .read_map = cpuacct_stats_show, 8165 }, 8166}; 8167 8168static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp) 8169{ 8170 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files)); 8171} 8172 8173/* 8174 * charge this task's execution time to its accounting group. 8175 * 8176 * called with rq->lock held. 8177 */ 8178void cpuacct_charge(struct task_struct *tsk, u64 cputime) 8179{ 8180 struct cpuacct *ca; 8181 int cpu; 8182 8183 if (unlikely(!cpuacct_subsys.active)) 8184 return; 8185 8186 cpu = task_cpu(tsk); 8187 8188 rcu_read_lock(); 8189 8190 ca = task_ca(tsk); 8191 8192 for (; ca; ca = parent_ca(ca)) { 8193 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu); 8194 *cpuusage += cputime; 8195 } 8196 8197 rcu_read_unlock(); 8198} 8199 8200struct cgroup_subsys cpuacct_subsys = { 8201 .name = "cpuacct", 8202 .create = cpuacct_create, 8203 .destroy = cpuacct_destroy, 8204 .populate = cpuacct_populate, 8205 .subsys_id = cpuacct_subsys_id, 8206}; 8207#endif /* CONFIG_CGROUP_CPUACCT */ 8208