core.c revision 1c792db7f7957e2e34b9a164f08200e36a25dfd0
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 75#include <asm/tlb.h> 76#include <asm/irq_regs.h> 77#ifdef CONFIG_PARAVIRT 78#include <asm/paravirt.h> 79#endif 80 81#include "sched.h" 82#include "../workqueue_sched.h" 83 84#define CREATE_TRACE_POINTS 85#include <trace/events/sched.h> 86 87void start_bandwidth_timer(struct hrtimer *period_timer, ktime_t period) 88{ 89 unsigned long delta; 90 ktime_t soft, hard, now; 91 92 for (;;) { 93 if (hrtimer_active(period_timer)) 94 break; 95 96 now = hrtimer_cb_get_time(period_timer); 97 hrtimer_forward(period_timer, now, period); 98 99 soft = hrtimer_get_softexpires(period_timer); 100 hard = hrtimer_get_expires(period_timer); 101 delta = ktime_to_ns(ktime_sub(hard, soft)); 102 __hrtimer_start_range_ns(period_timer, soft, delta, 103 HRTIMER_MODE_ABS_PINNED, 0); 104 } 105} 106 107DEFINE_MUTEX(sched_domains_mutex); 108DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues); 109 110static void update_rq_clock_task(struct rq *rq, s64 delta); 111 112void update_rq_clock(struct rq *rq) 113{ 114 s64 delta; 115 116 if (rq->skip_clock_update > 0) 117 return; 118 119 delta = sched_clock_cpu(cpu_of(rq)) - rq->clock; 120 rq->clock += delta; 121 update_rq_clock_task(rq, delta); 122} 123 124/* 125 * Debugging: various feature bits 126 */ 127 128#define SCHED_FEAT(name, enabled) \ 129 (1UL << __SCHED_FEAT_##name) * enabled | 130 131const_debug unsigned int sysctl_sched_features = 132#include "features.h" 133 0; 134 135#undef SCHED_FEAT 136 137#ifdef CONFIG_SCHED_DEBUG 138#define SCHED_FEAT(name, enabled) \ 139 #name , 140 141static __read_mostly char *sched_feat_names[] = { 142#include "features.h" 143 NULL 144}; 145 146#undef SCHED_FEAT 147 148static int sched_feat_show(struct seq_file *m, void *v) 149{ 150 int i; 151 152 for (i = 0; sched_feat_names[i]; i++) { 153 if (!(sysctl_sched_features & (1UL << i))) 154 seq_puts(m, "NO_"); 155 seq_printf(m, "%s ", sched_feat_names[i]); 156 } 157 seq_puts(m, "\n"); 158 159 return 0; 160} 161 162static ssize_t 163sched_feat_write(struct file *filp, const char __user *ubuf, 164 size_t cnt, loff_t *ppos) 165{ 166 char buf[64]; 167 char *cmp; 168 int neg = 0; 169 int i; 170 171 if (cnt > 63) 172 cnt = 63; 173 174 if (copy_from_user(&buf, ubuf, cnt)) 175 return -EFAULT; 176 177 buf[cnt] = 0; 178 cmp = strstrip(buf); 179 180 if (strncmp(cmp, "NO_", 3) == 0) { 181 neg = 1; 182 cmp += 3; 183 } 184 185 for (i = 0; sched_feat_names[i]; i++) { 186 if (strcmp(cmp, sched_feat_names[i]) == 0) { 187 if (neg) 188 sysctl_sched_features &= ~(1UL << i); 189 else 190 sysctl_sched_features |= (1UL << i); 191 break; 192 } 193 } 194 195 if (!sched_feat_names[i]) 196 return -EINVAL; 197 198 *ppos += cnt; 199 200 return cnt; 201} 202 203static int sched_feat_open(struct inode *inode, struct file *filp) 204{ 205 return single_open(filp, sched_feat_show, NULL); 206} 207 208static const struct file_operations sched_feat_fops = { 209 .open = sched_feat_open, 210 .write = sched_feat_write, 211 .read = seq_read, 212 .llseek = seq_lseek, 213 .release = single_release, 214}; 215 216static __init int sched_init_debug(void) 217{ 218 debugfs_create_file("sched_features", 0644, NULL, NULL, 219 &sched_feat_fops); 220 221 return 0; 222} 223late_initcall(sched_init_debug); 224 225#endif 226 227/* 228 * Number of tasks to iterate in a single balance run. 229 * Limited because this is done with IRQs disabled. 230 */ 231const_debug unsigned int sysctl_sched_nr_migrate = 32; 232 233/* 234 * period over which we average the RT time consumption, measured 235 * in ms. 236 * 237 * default: 1s 238 */ 239const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC; 240 241/* 242 * period over which we measure -rt task cpu usage in us. 243 * default: 1s 244 */ 245unsigned int sysctl_sched_rt_period = 1000000; 246 247__read_mostly int scheduler_running; 248 249/* 250 * part of the period that we allow rt tasks to run in us. 251 * default: 0.95s 252 */ 253int sysctl_sched_rt_runtime = 950000; 254 255 256 257/* 258 * __task_rq_lock - lock the rq @p resides on. 259 */ 260static inline struct rq *__task_rq_lock(struct task_struct *p) 261 __acquires(rq->lock) 262{ 263 struct rq *rq; 264 265 lockdep_assert_held(&p->pi_lock); 266 267 for (;;) { 268 rq = task_rq(p); 269 raw_spin_lock(&rq->lock); 270 if (likely(rq == task_rq(p))) 271 return rq; 272 raw_spin_unlock(&rq->lock); 273 } 274} 275 276/* 277 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on. 278 */ 279static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags) 280 __acquires(p->pi_lock) 281 __acquires(rq->lock) 282{ 283 struct rq *rq; 284 285 for (;;) { 286 raw_spin_lock_irqsave(&p->pi_lock, *flags); 287 rq = task_rq(p); 288 raw_spin_lock(&rq->lock); 289 if (likely(rq == task_rq(p))) 290 return rq; 291 raw_spin_unlock(&rq->lock); 292 raw_spin_unlock_irqrestore(&p->pi_lock, *flags); 293 } 294} 295 296static void __task_rq_unlock(struct rq *rq) 297 __releases(rq->lock) 298{ 299 raw_spin_unlock(&rq->lock); 300} 301 302static inline void 303task_rq_unlock(struct rq *rq, struct task_struct *p, unsigned long *flags) 304 __releases(rq->lock) 305 __releases(p->pi_lock) 306{ 307 raw_spin_unlock(&rq->lock); 308 raw_spin_unlock_irqrestore(&p->pi_lock, *flags); 309} 310 311/* 312 * this_rq_lock - lock this runqueue and disable interrupts. 313 */ 314static struct rq *this_rq_lock(void) 315 __acquires(rq->lock) 316{ 317 struct rq *rq; 318 319 local_irq_disable(); 320 rq = this_rq(); 321 raw_spin_lock(&rq->lock); 322 323 return rq; 324} 325 326#ifdef CONFIG_SCHED_HRTICK 327/* 328 * Use HR-timers to deliver accurate preemption points. 329 * 330 * Its all a bit involved since we cannot program an hrt while holding the 331 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a 332 * reschedule event. 333 * 334 * When we get rescheduled we reprogram the hrtick_timer outside of the 335 * rq->lock. 336 */ 337 338static void hrtick_clear(struct rq *rq) 339{ 340 if (hrtimer_active(&rq->hrtick_timer)) 341 hrtimer_cancel(&rq->hrtick_timer); 342} 343 344/* 345 * High-resolution timer tick. 346 * Runs from hardirq context with interrupts disabled. 347 */ 348static enum hrtimer_restart hrtick(struct hrtimer *timer) 349{ 350 struct rq *rq = container_of(timer, struct rq, hrtick_timer); 351 352 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id()); 353 354 raw_spin_lock(&rq->lock); 355 update_rq_clock(rq); 356 rq->curr->sched_class->task_tick(rq, rq->curr, 1); 357 raw_spin_unlock(&rq->lock); 358 359 return HRTIMER_NORESTART; 360} 361 362#ifdef CONFIG_SMP 363/* 364 * called from hardirq (IPI) context 365 */ 366static void __hrtick_start(void *arg) 367{ 368 struct rq *rq = arg; 369 370 raw_spin_lock(&rq->lock); 371 hrtimer_restart(&rq->hrtick_timer); 372 rq->hrtick_csd_pending = 0; 373 raw_spin_unlock(&rq->lock); 374} 375 376/* 377 * Called to set the hrtick timer state. 378 * 379 * called with rq->lock held and irqs disabled 380 */ 381void hrtick_start(struct rq *rq, u64 delay) 382{ 383 struct hrtimer *timer = &rq->hrtick_timer; 384 ktime_t time = ktime_add_ns(timer->base->get_time(), delay); 385 386 hrtimer_set_expires(timer, time); 387 388 if (rq == this_rq()) { 389 hrtimer_restart(timer); 390 } else if (!rq->hrtick_csd_pending) { 391 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd, 0); 392 rq->hrtick_csd_pending = 1; 393 } 394} 395 396static int 397hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu) 398{ 399 int cpu = (int)(long)hcpu; 400 401 switch (action) { 402 case CPU_UP_CANCELED: 403 case CPU_UP_CANCELED_FROZEN: 404 case CPU_DOWN_PREPARE: 405 case CPU_DOWN_PREPARE_FROZEN: 406 case CPU_DEAD: 407 case CPU_DEAD_FROZEN: 408 hrtick_clear(cpu_rq(cpu)); 409 return NOTIFY_OK; 410 } 411 412 return NOTIFY_DONE; 413} 414 415static __init void init_hrtick(void) 416{ 417 hotcpu_notifier(hotplug_hrtick, 0); 418} 419#else 420/* 421 * Called to set the hrtick timer state. 422 * 423 * called with rq->lock held and irqs disabled 424 */ 425void hrtick_start(struct rq *rq, u64 delay) 426{ 427 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0, 428 HRTIMER_MODE_REL_PINNED, 0); 429} 430 431static inline void init_hrtick(void) 432{ 433} 434#endif /* CONFIG_SMP */ 435 436static void init_rq_hrtick(struct rq *rq) 437{ 438#ifdef CONFIG_SMP 439 rq->hrtick_csd_pending = 0; 440 441 rq->hrtick_csd.flags = 0; 442 rq->hrtick_csd.func = __hrtick_start; 443 rq->hrtick_csd.info = rq; 444#endif 445 446 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); 447 rq->hrtick_timer.function = hrtick; 448} 449#else /* CONFIG_SCHED_HRTICK */ 450static inline void hrtick_clear(struct rq *rq) 451{ 452} 453 454static inline void init_rq_hrtick(struct rq *rq) 455{ 456} 457 458static inline void init_hrtick(void) 459{ 460} 461#endif /* CONFIG_SCHED_HRTICK */ 462 463/* 464 * resched_task - mark a task 'to be rescheduled now'. 465 * 466 * On UP this means the setting of the need_resched flag, on SMP it 467 * might also involve a cross-CPU call to trigger the scheduler on 468 * the target CPU. 469 */ 470#ifdef CONFIG_SMP 471 472#ifndef tsk_is_polling 473#define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG) 474#endif 475 476void resched_task(struct task_struct *p) 477{ 478 int cpu; 479 480 assert_raw_spin_locked(&task_rq(p)->lock); 481 482 if (test_tsk_need_resched(p)) 483 return; 484 485 set_tsk_need_resched(p); 486 487 cpu = task_cpu(p); 488 if (cpu == smp_processor_id()) 489 return; 490 491 /* NEED_RESCHED must be visible before we test polling */ 492 smp_mb(); 493 if (!tsk_is_polling(p)) 494 smp_send_reschedule(cpu); 495} 496 497void resched_cpu(int cpu) 498{ 499 struct rq *rq = cpu_rq(cpu); 500 unsigned long flags; 501 502 if (!raw_spin_trylock_irqsave(&rq->lock, flags)) 503 return; 504 resched_task(cpu_curr(cpu)); 505 raw_spin_unlock_irqrestore(&rq->lock, flags); 506} 507 508#ifdef CONFIG_NO_HZ 509/* 510 * In the semi idle case, use the nearest busy cpu for migrating timers 511 * from an idle cpu. This is good for power-savings. 512 * 513 * We don't do similar optimization for completely idle system, as 514 * selecting an idle cpu will add more delays to the timers than intended 515 * (as that cpu's timer base may not be uptodate wrt jiffies etc). 516 */ 517int get_nohz_timer_target(void) 518{ 519 int cpu = smp_processor_id(); 520 int i; 521 struct sched_domain *sd; 522 523 rcu_read_lock(); 524 for_each_domain(cpu, sd) { 525 for_each_cpu(i, sched_domain_span(sd)) { 526 if (!idle_cpu(i)) { 527 cpu = i; 528 goto unlock; 529 } 530 } 531 } 532unlock: 533 rcu_read_unlock(); 534 return cpu; 535} 536/* 537 * When add_timer_on() enqueues a timer into the timer wheel of an 538 * idle CPU then this timer might expire before the next timer event 539 * which is scheduled to wake up that CPU. In case of a completely 540 * idle system the next event might even be infinite time into the 541 * future. wake_up_idle_cpu() ensures that the CPU is woken up and 542 * leaves the inner idle loop so the newly added timer is taken into 543 * account when the CPU goes back to idle and evaluates the timer 544 * wheel for the next timer event. 545 */ 546void wake_up_idle_cpu(int cpu) 547{ 548 struct rq *rq = cpu_rq(cpu); 549 550 if (cpu == smp_processor_id()) 551 return; 552 553 /* 554 * This is safe, as this function is called with the timer 555 * wheel base lock of (cpu) held. When the CPU is on the way 556 * to idle and has not yet set rq->curr to idle then it will 557 * be serialized on the timer wheel base lock and take the new 558 * timer into account automatically. 559 */ 560 if (rq->curr != rq->idle) 561 return; 562 563 /* 564 * We can set TIF_RESCHED on the idle task of the other CPU 565 * lockless. The worst case is that the other CPU runs the 566 * idle task through an additional NOOP schedule() 567 */ 568 set_tsk_need_resched(rq->idle); 569 570 /* NEED_RESCHED must be visible before we test polling */ 571 smp_mb(); 572 if (!tsk_is_polling(rq->idle)) 573 smp_send_reschedule(cpu); 574} 575 576static inline bool got_nohz_idle_kick(void) 577{ 578 int cpu = smp_processor_id(); 579 return idle_cpu(cpu) && test_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu)); 580} 581 582#else /* CONFIG_NO_HZ */ 583 584static inline bool got_nohz_idle_kick(void) 585{ 586 return false; 587} 588 589#endif /* CONFIG_NO_HZ */ 590 591void sched_avg_update(struct rq *rq) 592{ 593 s64 period = sched_avg_period(); 594 595 while ((s64)(rq->clock - rq->age_stamp) > period) { 596 /* 597 * Inline assembly required to prevent the compiler 598 * optimising this loop into a divmod call. 599 * See __iter_div_u64_rem() for another example of this. 600 */ 601 asm("" : "+rm" (rq->age_stamp)); 602 rq->age_stamp += period; 603 rq->rt_avg /= 2; 604 } 605} 606 607#else /* !CONFIG_SMP */ 608void resched_task(struct task_struct *p) 609{ 610 assert_raw_spin_locked(&task_rq(p)->lock); 611 set_tsk_need_resched(p); 612} 613#endif /* CONFIG_SMP */ 614 615#if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \ 616 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH))) 617/* 618 * Iterate task_group tree rooted at *from, calling @down when first entering a 619 * node and @up when leaving it for the final time. 620 * 621 * Caller must hold rcu_lock or sufficient equivalent. 622 */ 623int walk_tg_tree_from(struct task_group *from, 624 tg_visitor down, tg_visitor up, void *data) 625{ 626 struct task_group *parent, *child; 627 int ret; 628 629 parent = from; 630 631down: 632 ret = (*down)(parent, data); 633 if (ret) 634 goto out; 635 list_for_each_entry_rcu(child, &parent->children, siblings) { 636 parent = child; 637 goto down; 638 639up: 640 continue; 641 } 642 ret = (*up)(parent, data); 643 if (ret || parent == from) 644 goto out; 645 646 child = parent; 647 parent = parent->parent; 648 if (parent) 649 goto up; 650out: 651 return ret; 652} 653 654int tg_nop(struct task_group *tg, void *data) 655{ 656 return 0; 657} 658#endif 659 660void update_cpu_load(struct rq *this_rq); 661 662static void set_load_weight(struct task_struct *p) 663{ 664 int prio = p->static_prio - MAX_RT_PRIO; 665 struct load_weight *load = &p->se.load; 666 667 /* 668 * SCHED_IDLE tasks get minimal weight: 669 */ 670 if (p->policy == SCHED_IDLE) { 671 load->weight = scale_load(WEIGHT_IDLEPRIO); 672 load->inv_weight = WMULT_IDLEPRIO; 673 return; 674 } 675 676 load->weight = scale_load(prio_to_weight[prio]); 677 load->inv_weight = prio_to_wmult[prio]; 678} 679 680static void enqueue_task(struct rq *rq, struct task_struct *p, int flags) 681{ 682 update_rq_clock(rq); 683 sched_info_queued(p); 684 p->sched_class->enqueue_task(rq, p, flags); 685} 686 687static void dequeue_task(struct rq *rq, struct task_struct *p, int flags) 688{ 689 update_rq_clock(rq); 690 sched_info_dequeued(p); 691 p->sched_class->dequeue_task(rq, p, flags); 692} 693 694/* 695 * activate_task - move a task to the runqueue. 696 */ 697void activate_task(struct rq *rq, struct task_struct *p, int flags) 698{ 699 if (task_contributes_to_load(p)) 700 rq->nr_uninterruptible--; 701 702 enqueue_task(rq, p, flags); 703} 704 705/* 706 * deactivate_task - remove a task from the runqueue. 707 */ 708void deactivate_task(struct rq *rq, struct task_struct *p, int flags) 709{ 710 if (task_contributes_to_load(p)) 711 rq->nr_uninterruptible++; 712 713 dequeue_task(rq, p, flags); 714} 715 716#ifdef CONFIG_IRQ_TIME_ACCOUNTING 717 718/* 719 * There are no locks covering percpu hardirq/softirq time. 720 * They are only modified in account_system_vtime, on corresponding CPU 721 * with interrupts disabled. So, writes are safe. 722 * They are read and saved off onto struct rq in update_rq_clock(). 723 * This may result in other CPU reading this CPU's irq time and can 724 * race with irq/account_system_vtime on this CPU. We would either get old 725 * or new value with a side effect of accounting a slice of irq time to wrong 726 * task when irq is in progress while we read rq->clock. That is a worthy 727 * compromise in place of having locks on each irq in account_system_time. 728 */ 729static DEFINE_PER_CPU(u64, cpu_hardirq_time); 730static DEFINE_PER_CPU(u64, cpu_softirq_time); 731 732static DEFINE_PER_CPU(u64, irq_start_time); 733static int sched_clock_irqtime; 734 735void enable_sched_clock_irqtime(void) 736{ 737 sched_clock_irqtime = 1; 738} 739 740void disable_sched_clock_irqtime(void) 741{ 742 sched_clock_irqtime = 0; 743} 744 745#ifndef CONFIG_64BIT 746static DEFINE_PER_CPU(seqcount_t, irq_time_seq); 747 748static inline void irq_time_write_begin(void) 749{ 750 __this_cpu_inc(irq_time_seq.sequence); 751 smp_wmb(); 752} 753 754static inline void irq_time_write_end(void) 755{ 756 smp_wmb(); 757 __this_cpu_inc(irq_time_seq.sequence); 758} 759 760static inline u64 irq_time_read(int cpu) 761{ 762 u64 irq_time; 763 unsigned seq; 764 765 do { 766 seq = read_seqcount_begin(&per_cpu(irq_time_seq, cpu)); 767 irq_time = per_cpu(cpu_softirq_time, cpu) + 768 per_cpu(cpu_hardirq_time, cpu); 769 } while (read_seqcount_retry(&per_cpu(irq_time_seq, cpu), seq)); 770 771 return irq_time; 772} 773#else /* CONFIG_64BIT */ 774static inline void irq_time_write_begin(void) 775{ 776} 777 778static inline void irq_time_write_end(void) 779{ 780} 781 782static inline u64 irq_time_read(int cpu) 783{ 784 return per_cpu(cpu_softirq_time, cpu) + per_cpu(cpu_hardirq_time, cpu); 785} 786#endif /* CONFIG_64BIT */ 787 788/* 789 * Called before incrementing preempt_count on {soft,}irq_enter 790 * and before decrementing preempt_count on {soft,}irq_exit. 791 */ 792void account_system_vtime(struct task_struct *curr) 793{ 794 unsigned long flags; 795 s64 delta; 796 int cpu; 797 798 if (!sched_clock_irqtime) 799 return; 800 801 local_irq_save(flags); 802 803 cpu = smp_processor_id(); 804 delta = sched_clock_cpu(cpu) - __this_cpu_read(irq_start_time); 805 __this_cpu_add(irq_start_time, delta); 806 807 irq_time_write_begin(); 808 /* 809 * We do not account for softirq time from ksoftirqd here. 810 * We want to continue accounting softirq time to ksoftirqd thread 811 * in that case, so as not to confuse scheduler with a special task 812 * that do not consume any time, but still wants to run. 813 */ 814 if (hardirq_count()) 815 __this_cpu_add(cpu_hardirq_time, delta); 816 else if (in_serving_softirq() && curr != this_cpu_ksoftirqd()) 817 __this_cpu_add(cpu_softirq_time, delta); 818 819 irq_time_write_end(); 820 local_irq_restore(flags); 821} 822EXPORT_SYMBOL_GPL(account_system_vtime); 823 824#endif /* CONFIG_IRQ_TIME_ACCOUNTING */ 825 826#ifdef CONFIG_PARAVIRT 827static inline u64 steal_ticks(u64 steal) 828{ 829 if (unlikely(steal > NSEC_PER_SEC)) 830 return div_u64(steal, TICK_NSEC); 831 832 return __iter_div_u64_rem(steal, TICK_NSEC, &steal); 833} 834#endif 835 836static void update_rq_clock_task(struct rq *rq, s64 delta) 837{ 838/* 839 * In theory, the compile should just see 0 here, and optimize out the call 840 * to sched_rt_avg_update. But I don't trust it... 841 */ 842#if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING) 843 s64 steal = 0, irq_delta = 0; 844#endif 845#ifdef CONFIG_IRQ_TIME_ACCOUNTING 846 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time; 847 848 /* 849 * Since irq_time is only updated on {soft,}irq_exit, we might run into 850 * this case when a previous update_rq_clock() happened inside a 851 * {soft,}irq region. 852 * 853 * When this happens, we stop ->clock_task and only update the 854 * prev_irq_time stamp to account for the part that fit, so that a next 855 * update will consume the rest. This ensures ->clock_task is 856 * monotonic. 857 * 858 * It does however cause some slight miss-attribution of {soft,}irq 859 * time, a more accurate solution would be to update the irq_time using 860 * the current rq->clock timestamp, except that would require using 861 * atomic ops. 862 */ 863 if (irq_delta > delta) 864 irq_delta = delta; 865 866 rq->prev_irq_time += irq_delta; 867 delta -= irq_delta; 868#endif 869#ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING 870 if (static_branch((¶virt_steal_rq_enabled))) { 871 u64 st; 872 873 steal = paravirt_steal_clock(cpu_of(rq)); 874 steal -= rq->prev_steal_time_rq; 875 876 if (unlikely(steal > delta)) 877 steal = delta; 878 879 st = steal_ticks(steal); 880 steal = st * TICK_NSEC; 881 882 rq->prev_steal_time_rq += steal; 883 884 delta -= steal; 885 } 886#endif 887 888 rq->clock_task += delta; 889 890#if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING) 891 if ((irq_delta + steal) && sched_feat(NONTASK_POWER)) 892 sched_rt_avg_update(rq, irq_delta + steal); 893#endif 894} 895 896#ifdef CONFIG_IRQ_TIME_ACCOUNTING 897static int irqtime_account_hi_update(void) 898{ 899 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat; 900 unsigned long flags; 901 u64 latest_ns; 902 int ret = 0; 903 904 local_irq_save(flags); 905 latest_ns = this_cpu_read(cpu_hardirq_time); 906 if (cputime64_gt(nsecs_to_cputime64(latest_ns), cpustat->irq)) 907 ret = 1; 908 local_irq_restore(flags); 909 return ret; 910} 911 912static int irqtime_account_si_update(void) 913{ 914 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat; 915 unsigned long flags; 916 u64 latest_ns; 917 int ret = 0; 918 919 local_irq_save(flags); 920 latest_ns = this_cpu_read(cpu_softirq_time); 921 if (cputime64_gt(nsecs_to_cputime64(latest_ns), cpustat->softirq)) 922 ret = 1; 923 local_irq_restore(flags); 924 return ret; 925} 926 927#else /* CONFIG_IRQ_TIME_ACCOUNTING */ 928 929#define sched_clock_irqtime (0) 930 931#endif 932 933void sched_set_stop_task(int cpu, struct task_struct *stop) 934{ 935 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 }; 936 struct task_struct *old_stop = cpu_rq(cpu)->stop; 937 938 if (stop) { 939 /* 940 * Make it appear like a SCHED_FIFO task, its something 941 * userspace knows about and won't get confused about. 942 * 943 * Also, it will make PI more or less work without too 944 * much confusion -- but then, stop work should not 945 * rely on PI working anyway. 946 */ 947 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m); 948 949 stop->sched_class = &stop_sched_class; 950 } 951 952 cpu_rq(cpu)->stop = stop; 953 954 if (old_stop) { 955 /* 956 * Reset it back to a normal scheduling class so that 957 * it can die in pieces. 958 */ 959 old_stop->sched_class = &rt_sched_class; 960 } 961} 962 963/* 964 * __normal_prio - return the priority that is based on the static prio 965 */ 966static inline int __normal_prio(struct task_struct *p) 967{ 968 return p->static_prio; 969} 970 971/* 972 * Calculate the expected normal priority: i.e. priority 973 * without taking RT-inheritance into account. Might be 974 * boosted by interactivity modifiers. Changes upon fork, 975 * setprio syscalls, and whenever the interactivity 976 * estimator recalculates. 977 */ 978static inline int normal_prio(struct task_struct *p) 979{ 980 int prio; 981 982 if (task_has_rt_policy(p)) 983 prio = MAX_RT_PRIO-1 - p->rt_priority; 984 else 985 prio = __normal_prio(p); 986 return prio; 987} 988 989/* 990 * Calculate the current priority, i.e. the priority 991 * taken into account by the scheduler. This value might 992 * be boosted by RT tasks, or might be boosted by 993 * interactivity modifiers. Will be RT if the task got 994 * RT-boosted. If not then it returns p->normal_prio. 995 */ 996static int effective_prio(struct task_struct *p) 997{ 998 p->normal_prio = normal_prio(p); 999 /* 1000 * If we are RT tasks or we were boosted to RT priority, 1001 * keep the priority unchanged. Otherwise, update priority 1002 * to the normal priority: 1003 */ 1004 if (!rt_prio(p->prio)) 1005 return p->normal_prio; 1006 return p->prio; 1007} 1008 1009/** 1010 * task_curr - is this task currently executing on a CPU? 1011 * @p: the task in question. 1012 */ 1013inline int task_curr(const struct task_struct *p) 1014{ 1015 return cpu_curr(task_cpu(p)) == p; 1016} 1017 1018static inline void check_class_changed(struct rq *rq, struct task_struct *p, 1019 const struct sched_class *prev_class, 1020 int oldprio) 1021{ 1022 if (prev_class != p->sched_class) { 1023 if (prev_class->switched_from) 1024 prev_class->switched_from(rq, p); 1025 p->sched_class->switched_to(rq, p); 1026 } else if (oldprio != p->prio) 1027 p->sched_class->prio_changed(rq, p, oldprio); 1028} 1029 1030void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags) 1031{ 1032 const struct sched_class *class; 1033 1034 if (p->sched_class == rq->curr->sched_class) { 1035 rq->curr->sched_class->check_preempt_curr(rq, p, flags); 1036 } else { 1037 for_each_class(class) { 1038 if (class == rq->curr->sched_class) 1039 break; 1040 if (class == p->sched_class) { 1041 resched_task(rq->curr); 1042 break; 1043 } 1044 } 1045 } 1046 1047 /* 1048 * A queue event has occurred, and we're going to schedule. In 1049 * this case, we can save a useless back to back clock update. 1050 */ 1051 if (rq->curr->on_rq && test_tsk_need_resched(rq->curr)) 1052 rq->skip_clock_update = 1; 1053} 1054 1055#ifdef CONFIG_SMP 1056void set_task_cpu(struct task_struct *p, unsigned int new_cpu) 1057{ 1058#ifdef CONFIG_SCHED_DEBUG 1059 /* 1060 * We should never call set_task_cpu() on a blocked task, 1061 * ttwu() will sort out the placement. 1062 */ 1063 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING && 1064 !(task_thread_info(p)->preempt_count & PREEMPT_ACTIVE)); 1065 1066#ifdef CONFIG_LOCKDEP 1067 /* 1068 * The caller should hold either p->pi_lock or rq->lock, when changing 1069 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks. 1070 * 1071 * sched_move_task() holds both and thus holding either pins the cgroup, 1072 * see set_task_rq(). 1073 * 1074 * Furthermore, all task_rq users should acquire both locks, see 1075 * task_rq_lock(). 1076 */ 1077 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) || 1078 lockdep_is_held(&task_rq(p)->lock))); 1079#endif 1080#endif 1081 1082 trace_sched_migrate_task(p, new_cpu); 1083 1084 if (task_cpu(p) != new_cpu) { 1085 p->se.nr_migrations++; 1086 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, NULL, 0); 1087 } 1088 1089 __set_task_cpu(p, new_cpu); 1090} 1091 1092struct migration_arg { 1093 struct task_struct *task; 1094 int dest_cpu; 1095}; 1096 1097static int migration_cpu_stop(void *data); 1098 1099/* 1100 * wait_task_inactive - wait for a thread to unschedule. 1101 * 1102 * If @match_state is nonzero, it's the @p->state value just checked and 1103 * not expected to change. If it changes, i.e. @p might have woken up, 1104 * then return zero. When we succeed in waiting for @p to be off its CPU, 1105 * we return a positive number (its total switch count). If a second call 1106 * a short while later returns the same number, the caller can be sure that 1107 * @p has remained unscheduled the whole time. 1108 * 1109 * The caller must ensure that the task *will* unschedule sometime soon, 1110 * else this function might spin for a *long* time. This function can't 1111 * be called with interrupts off, or it may introduce deadlock with 1112 * smp_call_function() if an IPI is sent by the same process we are 1113 * waiting to become inactive. 1114 */ 1115unsigned long wait_task_inactive(struct task_struct *p, long match_state) 1116{ 1117 unsigned long flags; 1118 int running, on_rq; 1119 unsigned long ncsw; 1120 struct rq *rq; 1121 1122 for (;;) { 1123 /* 1124 * We do the initial early heuristics without holding 1125 * any task-queue locks at all. We'll only try to get 1126 * the runqueue lock when things look like they will 1127 * work out! 1128 */ 1129 rq = task_rq(p); 1130 1131 /* 1132 * If the task is actively running on another CPU 1133 * still, just relax and busy-wait without holding 1134 * any locks. 1135 * 1136 * NOTE! Since we don't hold any locks, it's not 1137 * even sure that "rq" stays as the right runqueue! 1138 * But we don't care, since "task_running()" will 1139 * return false if the runqueue has changed and p 1140 * is actually now running somewhere else! 1141 */ 1142 while (task_running(rq, p)) { 1143 if (match_state && unlikely(p->state != match_state)) 1144 return 0; 1145 cpu_relax(); 1146 } 1147 1148 /* 1149 * Ok, time to look more closely! We need the rq 1150 * lock now, to be *sure*. If we're wrong, we'll 1151 * just go back and repeat. 1152 */ 1153 rq = task_rq_lock(p, &flags); 1154 trace_sched_wait_task(p); 1155 running = task_running(rq, p); 1156 on_rq = p->on_rq; 1157 ncsw = 0; 1158 if (!match_state || p->state == match_state) 1159 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */ 1160 task_rq_unlock(rq, p, &flags); 1161 1162 /* 1163 * If it changed from the expected state, bail out now. 1164 */ 1165 if (unlikely(!ncsw)) 1166 break; 1167 1168 /* 1169 * Was it really running after all now that we 1170 * checked with the proper locks actually held? 1171 * 1172 * Oops. Go back and try again.. 1173 */ 1174 if (unlikely(running)) { 1175 cpu_relax(); 1176 continue; 1177 } 1178 1179 /* 1180 * It's not enough that it's not actively running, 1181 * it must be off the runqueue _entirely_, and not 1182 * preempted! 1183 * 1184 * So if it was still runnable (but just not actively 1185 * running right now), it's preempted, and we should 1186 * yield - it could be a while. 1187 */ 1188 if (unlikely(on_rq)) { 1189 ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ); 1190 1191 set_current_state(TASK_UNINTERRUPTIBLE); 1192 schedule_hrtimeout(&to, HRTIMER_MODE_REL); 1193 continue; 1194 } 1195 1196 /* 1197 * Ahh, all good. It wasn't running, and it wasn't 1198 * runnable, which means that it will never become 1199 * running in the future either. We're all done! 1200 */ 1201 break; 1202 } 1203 1204 return ncsw; 1205} 1206 1207/*** 1208 * kick_process - kick a running thread to enter/exit the kernel 1209 * @p: the to-be-kicked thread 1210 * 1211 * Cause a process which is running on another CPU to enter 1212 * kernel-mode, without any delay. (to get signals handled.) 1213 * 1214 * NOTE: this function doesn't have to take the runqueue lock, 1215 * because all it wants to ensure is that the remote task enters 1216 * the kernel. If the IPI races and the task has been migrated 1217 * to another CPU then no harm is done and the purpose has been 1218 * achieved as well. 1219 */ 1220void kick_process(struct task_struct *p) 1221{ 1222 int cpu; 1223 1224 preempt_disable(); 1225 cpu = task_cpu(p); 1226 if ((cpu != smp_processor_id()) && task_curr(p)) 1227 smp_send_reschedule(cpu); 1228 preempt_enable(); 1229} 1230EXPORT_SYMBOL_GPL(kick_process); 1231#endif /* CONFIG_SMP */ 1232 1233#ifdef CONFIG_SMP 1234/* 1235 * ->cpus_allowed is protected by both rq->lock and p->pi_lock 1236 */ 1237static int select_fallback_rq(int cpu, struct task_struct *p) 1238{ 1239 int dest_cpu; 1240 const struct cpumask *nodemask = cpumask_of_node(cpu_to_node(cpu)); 1241 1242 /* Look for allowed, online CPU in same node. */ 1243 for_each_cpu_and(dest_cpu, nodemask, cpu_active_mask) 1244 if (cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p))) 1245 return dest_cpu; 1246 1247 /* Any allowed, online CPU? */ 1248 dest_cpu = cpumask_any_and(tsk_cpus_allowed(p), cpu_active_mask); 1249 if (dest_cpu < nr_cpu_ids) 1250 return dest_cpu; 1251 1252 /* No more Mr. Nice Guy. */ 1253 dest_cpu = cpuset_cpus_allowed_fallback(p); 1254 /* 1255 * Don't tell them about moving exiting tasks or 1256 * kernel threads (both mm NULL), since they never 1257 * leave kernel. 1258 */ 1259 if (p->mm && printk_ratelimit()) { 1260 printk(KERN_INFO "process %d (%s) no longer affine to cpu%d\n", 1261 task_pid_nr(p), p->comm, cpu); 1262 } 1263 1264 return dest_cpu; 1265} 1266 1267/* 1268 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable. 1269 */ 1270static inline 1271int select_task_rq(struct task_struct *p, int sd_flags, int wake_flags) 1272{ 1273 int cpu = p->sched_class->select_task_rq(p, sd_flags, wake_flags); 1274 1275 /* 1276 * In order not to call set_task_cpu() on a blocking task we need 1277 * to rely on ttwu() to place the task on a valid ->cpus_allowed 1278 * cpu. 1279 * 1280 * Since this is common to all placement strategies, this lives here. 1281 * 1282 * [ this allows ->select_task() to simply return task_cpu(p) and 1283 * not worry about this generic constraint ] 1284 */ 1285 if (unlikely(!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)) || 1286 !cpu_online(cpu))) 1287 cpu = select_fallback_rq(task_cpu(p), p); 1288 1289 return cpu; 1290} 1291 1292static void update_avg(u64 *avg, u64 sample) 1293{ 1294 s64 diff = sample - *avg; 1295 *avg += diff >> 3; 1296} 1297#endif 1298 1299static void 1300ttwu_stat(struct task_struct *p, int cpu, int wake_flags) 1301{ 1302#ifdef CONFIG_SCHEDSTATS 1303 struct rq *rq = this_rq(); 1304 1305#ifdef CONFIG_SMP 1306 int this_cpu = smp_processor_id(); 1307 1308 if (cpu == this_cpu) { 1309 schedstat_inc(rq, ttwu_local); 1310 schedstat_inc(p, se.statistics.nr_wakeups_local); 1311 } else { 1312 struct sched_domain *sd; 1313 1314 schedstat_inc(p, se.statistics.nr_wakeups_remote); 1315 rcu_read_lock(); 1316 for_each_domain(this_cpu, sd) { 1317 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) { 1318 schedstat_inc(sd, ttwu_wake_remote); 1319 break; 1320 } 1321 } 1322 rcu_read_unlock(); 1323 } 1324 1325 if (wake_flags & WF_MIGRATED) 1326 schedstat_inc(p, se.statistics.nr_wakeups_migrate); 1327 1328#endif /* CONFIG_SMP */ 1329 1330 schedstat_inc(rq, ttwu_count); 1331 schedstat_inc(p, se.statistics.nr_wakeups); 1332 1333 if (wake_flags & WF_SYNC) 1334 schedstat_inc(p, se.statistics.nr_wakeups_sync); 1335 1336#endif /* CONFIG_SCHEDSTATS */ 1337} 1338 1339static void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags) 1340{ 1341 activate_task(rq, p, en_flags); 1342 p->on_rq = 1; 1343 1344 /* if a worker is waking up, notify workqueue */ 1345 if (p->flags & PF_WQ_WORKER) 1346 wq_worker_waking_up(p, cpu_of(rq)); 1347} 1348 1349/* 1350 * Mark the task runnable and perform wakeup-preemption. 1351 */ 1352static void 1353ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags) 1354{ 1355 trace_sched_wakeup(p, true); 1356 check_preempt_curr(rq, p, wake_flags); 1357 1358 p->state = TASK_RUNNING; 1359#ifdef CONFIG_SMP 1360 if (p->sched_class->task_woken) 1361 p->sched_class->task_woken(rq, p); 1362 1363 if (rq->idle_stamp) { 1364 u64 delta = rq->clock - rq->idle_stamp; 1365 u64 max = 2*sysctl_sched_migration_cost; 1366 1367 if (delta > max) 1368 rq->avg_idle = max; 1369 else 1370 update_avg(&rq->avg_idle, delta); 1371 rq->idle_stamp = 0; 1372 } 1373#endif 1374} 1375 1376static void 1377ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags) 1378{ 1379#ifdef CONFIG_SMP 1380 if (p->sched_contributes_to_load) 1381 rq->nr_uninterruptible--; 1382#endif 1383 1384 ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_WAKING); 1385 ttwu_do_wakeup(rq, p, wake_flags); 1386} 1387 1388/* 1389 * Called in case the task @p isn't fully descheduled from its runqueue, 1390 * in this case we must do a remote wakeup. Its a 'light' wakeup though, 1391 * since all we need to do is flip p->state to TASK_RUNNING, since 1392 * the task is still ->on_rq. 1393 */ 1394static int ttwu_remote(struct task_struct *p, int wake_flags) 1395{ 1396 struct rq *rq; 1397 int ret = 0; 1398 1399 rq = __task_rq_lock(p); 1400 if (p->on_rq) { 1401 ttwu_do_wakeup(rq, p, wake_flags); 1402 ret = 1; 1403 } 1404 __task_rq_unlock(rq); 1405 1406 return ret; 1407} 1408 1409#ifdef CONFIG_SMP 1410static void sched_ttwu_pending(void) 1411{ 1412 struct rq *rq = this_rq(); 1413 struct llist_node *llist = llist_del_all(&rq->wake_list); 1414 struct task_struct *p; 1415 1416 raw_spin_lock(&rq->lock); 1417 1418 while (llist) { 1419 p = llist_entry(llist, struct task_struct, wake_entry); 1420 llist = llist_next(llist); 1421 ttwu_do_activate(rq, p, 0); 1422 } 1423 1424 raw_spin_unlock(&rq->lock); 1425} 1426 1427void scheduler_ipi(void) 1428{ 1429 if (llist_empty(&this_rq()->wake_list) && !got_nohz_idle_kick()) 1430 return; 1431 1432 /* 1433 * Not all reschedule IPI handlers call irq_enter/irq_exit, since 1434 * traditionally all their work was done from the interrupt return 1435 * path. Now that we actually do some work, we need to make sure 1436 * we do call them. 1437 * 1438 * Some archs already do call them, luckily irq_enter/exit nest 1439 * properly. 1440 * 1441 * Arguably we should visit all archs and update all handlers, 1442 * however a fair share of IPIs are still resched only so this would 1443 * somewhat pessimize the simple resched case. 1444 */ 1445 irq_enter(); 1446 sched_ttwu_pending(); 1447 1448 /* 1449 * Check if someone kicked us for doing the nohz idle load balance. 1450 */ 1451 if (unlikely(got_nohz_idle_kick() && !need_resched())) { 1452 this_rq()->idle_balance = 1; 1453 raise_softirq_irqoff(SCHED_SOFTIRQ); 1454 } 1455 irq_exit(); 1456} 1457 1458static void ttwu_queue_remote(struct task_struct *p, int cpu) 1459{ 1460 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list)) 1461 smp_send_reschedule(cpu); 1462} 1463 1464#ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW 1465static int ttwu_activate_remote(struct task_struct *p, int wake_flags) 1466{ 1467 struct rq *rq; 1468 int ret = 0; 1469 1470 rq = __task_rq_lock(p); 1471 if (p->on_cpu) { 1472 ttwu_activate(rq, p, ENQUEUE_WAKEUP); 1473 ttwu_do_wakeup(rq, p, wake_flags); 1474 ret = 1; 1475 } 1476 __task_rq_unlock(rq); 1477 1478 return ret; 1479 1480} 1481#endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */ 1482#endif /* CONFIG_SMP */ 1483 1484static void ttwu_queue(struct task_struct *p, int cpu) 1485{ 1486 struct rq *rq = cpu_rq(cpu); 1487 1488#if defined(CONFIG_SMP) 1489 if (sched_feat(TTWU_QUEUE) && cpu != smp_processor_id()) { 1490 sched_clock_cpu(cpu); /* sync clocks x-cpu */ 1491 ttwu_queue_remote(p, cpu); 1492 return; 1493 } 1494#endif 1495 1496 raw_spin_lock(&rq->lock); 1497 ttwu_do_activate(rq, p, 0); 1498 raw_spin_unlock(&rq->lock); 1499} 1500 1501/** 1502 * try_to_wake_up - wake up a thread 1503 * @p: the thread to be awakened 1504 * @state: the mask of task states that can be woken 1505 * @wake_flags: wake modifier flags (WF_*) 1506 * 1507 * Put it on the run-queue if it's not already there. The "current" 1508 * thread is always on the run-queue (except when the actual 1509 * re-schedule is in progress), and as such you're allowed to do 1510 * the simpler "current->state = TASK_RUNNING" to mark yourself 1511 * runnable without the overhead of this. 1512 * 1513 * Returns %true if @p was woken up, %false if it was already running 1514 * or @state didn't match @p's state. 1515 */ 1516static int 1517try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags) 1518{ 1519 unsigned long flags; 1520 int cpu, success = 0; 1521 1522 smp_wmb(); 1523 raw_spin_lock_irqsave(&p->pi_lock, flags); 1524 if (!(p->state & state)) 1525 goto out; 1526 1527 success = 1; /* we're going to change ->state */ 1528 cpu = task_cpu(p); 1529 1530 if (p->on_rq && ttwu_remote(p, wake_flags)) 1531 goto stat; 1532 1533#ifdef CONFIG_SMP 1534 /* 1535 * If the owning (remote) cpu is still in the middle of schedule() with 1536 * this task as prev, wait until its done referencing the task. 1537 */ 1538 while (p->on_cpu) { 1539#ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW 1540 /* 1541 * In case the architecture enables interrupts in 1542 * context_switch(), we cannot busy wait, since that 1543 * would lead to deadlocks when an interrupt hits and 1544 * tries to wake up @prev. So bail and do a complete 1545 * remote wakeup. 1546 */ 1547 if (ttwu_activate_remote(p, wake_flags)) 1548 goto stat; 1549#else 1550 cpu_relax(); 1551#endif 1552 } 1553 /* 1554 * Pairs with the smp_wmb() in finish_lock_switch(). 1555 */ 1556 smp_rmb(); 1557 1558 p->sched_contributes_to_load = !!task_contributes_to_load(p); 1559 p->state = TASK_WAKING; 1560 1561 if (p->sched_class->task_waking) 1562 p->sched_class->task_waking(p); 1563 1564 cpu = select_task_rq(p, SD_BALANCE_WAKE, wake_flags); 1565 if (task_cpu(p) != cpu) { 1566 wake_flags |= WF_MIGRATED; 1567 set_task_cpu(p, cpu); 1568 } 1569#endif /* CONFIG_SMP */ 1570 1571 ttwu_queue(p, cpu); 1572stat: 1573 ttwu_stat(p, cpu, wake_flags); 1574out: 1575 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 1576 1577 return success; 1578} 1579 1580/** 1581 * try_to_wake_up_local - try to wake up a local task with rq lock held 1582 * @p: the thread to be awakened 1583 * 1584 * Put @p on the run-queue if it's not already there. The caller must 1585 * ensure that this_rq() is locked, @p is bound to this_rq() and not 1586 * the current task. 1587 */ 1588static void try_to_wake_up_local(struct task_struct *p) 1589{ 1590 struct rq *rq = task_rq(p); 1591 1592 BUG_ON(rq != this_rq()); 1593 BUG_ON(p == current); 1594 lockdep_assert_held(&rq->lock); 1595 1596 if (!raw_spin_trylock(&p->pi_lock)) { 1597 raw_spin_unlock(&rq->lock); 1598 raw_spin_lock(&p->pi_lock); 1599 raw_spin_lock(&rq->lock); 1600 } 1601 1602 if (!(p->state & TASK_NORMAL)) 1603 goto out; 1604 1605 if (!p->on_rq) 1606 ttwu_activate(rq, p, ENQUEUE_WAKEUP); 1607 1608 ttwu_do_wakeup(rq, p, 0); 1609 ttwu_stat(p, smp_processor_id(), 0); 1610out: 1611 raw_spin_unlock(&p->pi_lock); 1612} 1613 1614/** 1615 * wake_up_process - Wake up a specific process 1616 * @p: The process to be woken up. 1617 * 1618 * Attempt to wake up the nominated process and move it to the set of runnable 1619 * processes. Returns 1 if the process was woken up, 0 if it was already 1620 * running. 1621 * 1622 * It may be assumed that this function implies a write memory barrier before 1623 * changing the task state if and only if any tasks are woken up. 1624 */ 1625int wake_up_process(struct task_struct *p) 1626{ 1627 return try_to_wake_up(p, TASK_ALL, 0); 1628} 1629EXPORT_SYMBOL(wake_up_process); 1630 1631int wake_up_state(struct task_struct *p, unsigned int state) 1632{ 1633 return try_to_wake_up(p, state, 0); 1634} 1635 1636/* 1637 * Perform scheduler related setup for a newly forked process p. 1638 * p is forked by current. 1639 * 1640 * __sched_fork() is basic setup used by init_idle() too: 1641 */ 1642static void __sched_fork(struct task_struct *p) 1643{ 1644 p->on_rq = 0; 1645 1646 p->se.on_rq = 0; 1647 p->se.exec_start = 0; 1648 p->se.sum_exec_runtime = 0; 1649 p->se.prev_sum_exec_runtime = 0; 1650 p->se.nr_migrations = 0; 1651 p->se.vruntime = 0; 1652 INIT_LIST_HEAD(&p->se.group_node); 1653 1654#ifdef CONFIG_SCHEDSTATS 1655 memset(&p->se.statistics, 0, sizeof(p->se.statistics)); 1656#endif 1657 1658 INIT_LIST_HEAD(&p->rt.run_list); 1659 1660#ifdef CONFIG_PREEMPT_NOTIFIERS 1661 INIT_HLIST_HEAD(&p->preempt_notifiers); 1662#endif 1663} 1664 1665/* 1666 * fork()/clone()-time setup: 1667 */ 1668void sched_fork(struct task_struct *p) 1669{ 1670 unsigned long flags; 1671 int cpu = get_cpu(); 1672 1673 __sched_fork(p); 1674 /* 1675 * We mark the process as running here. This guarantees that 1676 * nobody will actually run it, and a signal or other external 1677 * event cannot wake it up and insert it on the runqueue either. 1678 */ 1679 p->state = TASK_RUNNING; 1680 1681 /* 1682 * Make sure we do not leak PI boosting priority to the child. 1683 */ 1684 p->prio = current->normal_prio; 1685 1686 /* 1687 * Revert to default priority/policy on fork if requested. 1688 */ 1689 if (unlikely(p->sched_reset_on_fork)) { 1690 if (task_has_rt_policy(p)) { 1691 p->policy = SCHED_NORMAL; 1692 p->static_prio = NICE_TO_PRIO(0); 1693 p->rt_priority = 0; 1694 } else if (PRIO_TO_NICE(p->static_prio) < 0) 1695 p->static_prio = NICE_TO_PRIO(0); 1696 1697 p->prio = p->normal_prio = __normal_prio(p); 1698 set_load_weight(p); 1699 1700 /* 1701 * We don't need the reset flag anymore after the fork. It has 1702 * fulfilled its duty: 1703 */ 1704 p->sched_reset_on_fork = 0; 1705 } 1706 1707 if (!rt_prio(p->prio)) 1708 p->sched_class = &fair_sched_class; 1709 1710 if (p->sched_class->task_fork) 1711 p->sched_class->task_fork(p); 1712 1713 /* 1714 * The child is not yet in the pid-hash so no cgroup attach races, 1715 * and the cgroup is pinned to this child due to cgroup_fork() 1716 * is ran before sched_fork(). 1717 * 1718 * Silence PROVE_RCU. 1719 */ 1720 raw_spin_lock_irqsave(&p->pi_lock, flags); 1721 set_task_cpu(p, cpu); 1722 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 1723 1724#if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT) 1725 if (likely(sched_info_on())) 1726 memset(&p->sched_info, 0, sizeof(p->sched_info)); 1727#endif 1728#if defined(CONFIG_SMP) 1729 p->on_cpu = 0; 1730#endif 1731#ifdef CONFIG_PREEMPT_COUNT 1732 /* Want to start with kernel preemption disabled. */ 1733 task_thread_info(p)->preempt_count = 1; 1734#endif 1735#ifdef CONFIG_SMP 1736 plist_node_init(&p->pushable_tasks, MAX_PRIO); 1737#endif 1738 1739 put_cpu(); 1740} 1741 1742/* 1743 * wake_up_new_task - wake up a newly created task for the first time. 1744 * 1745 * This function will do some initial scheduler statistics housekeeping 1746 * that must be done for every newly created context, then puts the task 1747 * on the runqueue and wakes it. 1748 */ 1749void wake_up_new_task(struct task_struct *p) 1750{ 1751 unsigned long flags; 1752 struct rq *rq; 1753 1754 raw_spin_lock_irqsave(&p->pi_lock, flags); 1755#ifdef CONFIG_SMP 1756 /* 1757 * Fork balancing, do it here and not earlier because: 1758 * - cpus_allowed can change in the fork path 1759 * - any previously selected cpu might disappear through hotplug 1760 */ 1761 set_task_cpu(p, select_task_rq(p, SD_BALANCE_FORK, 0)); 1762#endif 1763 1764 rq = __task_rq_lock(p); 1765 activate_task(rq, p, 0); 1766 p->on_rq = 1; 1767 trace_sched_wakeup_new(p, true); 1768 check_preempt_curr(rq, p, WF_FORK); 1769#ifdef CONFIG_SMP 1770 if (p->sched_class->task_woken) 1771 p->sched_class->task_woken(rq, p); 1772#endif 1773 task_rq_unlock(rq, p, &flags); 1774} 1775 1776#ifdef CONFIG_PREEMPT_NOTIFIERS 1777 1778/** 1779 * preempt_notifier_register - tell me when current is being preempted & rescheduled 1780 * @notifier: notifier struct to register 1781 */ 1782void preempt_notifier_register(struct preempt_notifier *notifier) 1783{ 1784 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers); 1785} 1786EXPORT_SYMBOL_GPL(preempt_notifier_register); 1787 1788/** 1789 * preempt_notifier_unregister - no longer interested in preemption notifications 1790 * @notifier: notifier struct to unregister 1791 * 1792 * This is safe to call from within a preemption notifier. 1793 */ 1794void preempt_notifier_unregister(struct preempt_notifier *notifier) 1795{ 1796 hlist_del(¬ifier->link); 1797} 1798EXPORT_SYMBOL_GPL(preempt_notifier_unregister); 1799 1800static void fire_sched_in_preempt_notifiers(struct task_struct *curr) 1801{ 1802 struct preempt_notifier *notifier; 1803 struct hlist_node *node; 1804 1805 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link) 1806 notifier->ops->sched_in(notifier, raw_smp_processor_id()); 1807} 1808 1809static void 1810fire_sched_out_preempt_notifiers(struct task_struct *curr, 1811 struct task_struct *next) 1812{ 1813 struct preempt_notifier *notifier; 1814 struct hlist_node *node; 1815 1816 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link) 1817 notifier->ops->sched_out(notifier, next); 1818} 1819 1820#else /* !CONFIG_PREEMPT_NOTIFIERS */ 1821 1822static void fire_sched_in_preempt_notifiers(struct task_struct *curr) 1823{ 1824} 1825 1826static void 1827fire_sched_out_preempt_notifiers(struct task_struct *curr, 1828 struct task_struct *next) 1829{ 1830} 1831 1832#endif /* CONFIG_PREEMPT_NOTIFIERS */ 1833 1834/** 1835 * prepare_task_switch - prepare to switch tasks 1836 * @rq: the runqueue preparing to switch 1837 * @prev: the current task that is being switched out 1838 * @next: the task we are going to switch to. 1839 * 1840 * This is called with the rq lock held and interrupts off. It must 1841 * be paired with a subsequent finish_task_switch after the context 1842 * switch. 1843 * 1844 * prepare_task_switch sets up locking and calls architecture specific 1845 * hooks. 1846 */ 1847static inline void 1848prepare_task_switch(struct rq *rq, struct task_struct *prev, 1849 struct task_struct *next) 1850{ 1851 sched_info_switch(prev, next); 1852 perf_event_task_sched_out(prev, next); 1853 fire_sched_out_preempt_notifiers(prev, next); 1854 prepare_lock_switch(rq, next); 1855 prepare_arch_switch(next); 1856 trace_sched_switch(prev, next); 1857} 1858 1859/** 1860 * finish_task_switch - clean up after a task-switch 1861 * @rq: runqueue associated with task-switch 1862 * @prev: the thread we just switched away from. 1863 * 1864 * finish_task_switch must be called after the context switch, paired 1865 * with a prepare_task_switch call before the context switch. 1866 * finish_task_switch will reconcile locking set up by prepare_task_switch, 1867 * and do any other architecture-specific cleanup actions. 1868 * 1869 * Note that we may have delayed dropping an mm in context_switch(). If 1870 * so, we finish that here outside of the runqueue lock. (Doing it 1871 * with the lock held can cause deadlocks; see schedule() for 1872 * details.) 1873 */ 1874static void finish_task_switch(struct rq *rq, struct task_struct *prev) 1875 __releases(rq->lock) 1876{ 1877 struct mm_struct *mm = rq->prev_mm; 1878 long prev_state; 1879 1880 rq->prev_mm = NULL; 1881 1882 /* 1883 * A task struct has one reference for the use as "current". 1884 * If a task dies, then it sets TASK_DEAD in tsk->state and calls 1885 * schedule one last time. The schedule call will never return, and 1886 * the scheduled task must drop that reference. 1887 * The test for TASK_DEAD must occur while the runqueue locks are 1888 * still held, otherwise prev could be scheduled on another cpu, die 1889 * there before we look at prev->state, and then the reference would 1890 * be dropped twice. 1891 * Manfred Spraul <manfred@colorfullife.com> 1892 */ 1893 prev_state = prev->state; 1894 finish_arch_switch(prev); 1895#ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW 1896 local_irq_disable(); 1897#endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */ 1898 perf_event_task_sched_in(prev, current); 1899#ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW 1900 local_irq_enable(); 1901#endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */ 1902 finish_lock_switch(rq, prev); 1903 1904 fire_sched_in_preempt_notifiers(current); 1905 if (mm) 1906 mmdrop(mm); 1907 if (unlikely(prev_state == TASK_DEAD)) { 1908 /* 1909 * Remove function-return probe instances associated with this 1910 * task and put them back on the free list. 1911 */ 1912 kprobe_flush_task(prev); 1913 put_task_struct(prev); 1914 } 1915} 1916 1917#ifdef CONFIG_SMP 1918 1919/* assumes rq->lock is held */ 1920static inline void pre_schedule(struct rq *rq, struct task_struct *prev) 1921{ 1922 if (prev->sched_class->pre_schedule) 1923 prev->sched_class->pre_schedule(rq, prev); 1924} 1925 1926/* rq->lock is NOT held, but preemption is disabled */ 1927static inline void post_schedule(struct rq *rq) 1928{ 1929 if (rq->post_schedule) { 1930 unsigned long flags; 1931 1932 raw_spin_lock_irqsave(&rq->lock, flags); 1933 if (rq->curr->sched_class->post_schedule) 1934 rq->curr->sched_class->post_schedule(rq); 1935 raw_spin_unlock_irqrestore(&rq->lock, flags); 1936 1937 rq->post_schedule = 0; 1938 } 1939} 1940 1941#else 1942 1943static inline void pre_schedule(struct rq *rq, struct task_struct *p) 1944{ 1945} 1946 1947static inline void post_schedule(struct rq *rq) 1948{ 1949} 1950 1951#endif 1952 1953/** 1954 * schedule_tail - first thing a freshly forked thread must call. 1955 * @prev: the thread we just switched away from. 1956 */ 1957asmlinkage void schedule_tail(struct task_struct *prev) 1958 __releases(rq->lock) 1959{ 1960 struct rq *rq = this_rq(); 1961 1962 finish_task_switch(rq, prev); 1963 1964 /* 1965 * FIXME: do we need to worry about rq being invalidated by the 1966 * task_switch? 1967 */ 1968 post_schedule(rq); 1969 1970#ifdef __ARCH_WANT_UNLOCKED_CTXSW 1971 /* In this case, finish_task_switch does not reenable preemption */ 1972 preempt_enable(); 1973#endif 1974 if (current->set_child_tid) 1975 put_user(task_pid_vnr(current), current->set_child_tid); 1976} 1977 1978/* 1979 * context_switch - switch to the new MM and the new 1980 * thread's register state. 1981 */ 1982static inline void 1983context_switch(struct rq *rq, struct task_struct *prev, 1984 struct task_struct *next) 1985{ 1986 struct mm_struct *mm, *oldmm; 1987 1988 prepare_task_switch(rq, prev, next); 1989 1990 mm = next->mm; 1991 oldmm = prev->active_mm; 1992 /* 1993 * For paravirt, this is coupled with an exit in switch_to to 1994 * combine the page table reload and the switch backend into 1995 * one hypercall. 1996 */ 1997 arch_start_context_switch(prev); 1998 1999 if (!mm) { 2000 next->active_mm = oldmm; 2001 atomic_inc(&oldmm->mm_count); 2002 enter_lazy_tlb(oldmm, next); 2003 } else 2004 switch_mm(oldmm, mm, next); 2005 2006 if (!prev->mm) { 2007 prev->active_mm = NULL; 2008 rq->prev_mm = oldmm; 2009 } 2010 /* 2011 * Since the runqueue lock will be released by the next 2012 * task (which is an invalid locking op but in the case 2013 * of the scheduler it's an obvious special-case), so we 2014 * do an early lockdep release here: 2015 */ 2016#ifndef __ARCH_WANT_UNLOCKED_CTXSW 2017 spin_release(&rq->lock.dep_map, 1, _THIS_IP_); 2018#endif 2019 2020 /* Here we just switch the register state and the stack. */ 2021 switch_to(prev, next, prev); 2022 2023 barrier(); 2024 /* 2025 * this_rq must be evaluated again because prev may have moved 2026 * CPUs since it called schedule(), thus the 'rq' on its stack 2027 * frame will be invalid. 2028 */ 2029 finish_task_switch(this_rq(), prev); 2030} 2031 2032/* 2033 * nr_running, nr_uninterruptible and nr_context_switches: 2034 * 2035 * externally visible scheduler statistics: current number of runnable 2036 * threads, current number of uninterruptible-sleeping threads, total 2037 * number of context switches performed since bootup. 2038 */ 2039unsigned long nr_running(void) 2040{ 2041 unsigned long i, sum = 0; 2042 2043 for_each_online_cpu(i) 2044 sum += cpu_rq(i)->nr_running; 2045 2046 return sum; 2047} 2048 2049unsigned long nr_uninterruptible(void) 2050{ 2051 unsigned long i, sum = 0; 2052 2053 for_each_possible_cpu(i) 2054 sum += cpu_rq(i)->nr_uninterruptible; 2055 2056 /* 2057 * Since we read the counters lockless, it might be slightly 2058 * inaccurate. Do not allow it to go below zero though: 2059 */ 2060 if (unlikely((long)sum < 0)) 2061 sum = 0; 2062 2063 return sum; 2064} 2065 2066unsigned long long nr_context_switches(void) 2067{ 2068 int i; 2069 unsigned long long sum = 0; 2070 2071 for_each_possible_cpu(i) 2072 sum += cpu_rq(i)->nr_switches; 2073 2074 return sum; 2075} 2076 2077unsigned long nr_iowait(void) 2078{ 2079 unsigned long i, sum = 0; 2080 2081 for_each_possible_cpu(i) 2082 sum += atomic_read(&cpu_rq(i)->nr_iowait); 2083 2084 return sum; 2085} 2086 2087unsigned long nr_iowait_cpu(int cpu) 2088{ 2089 struct rq *this = cpu_rq(cpu); 2090 return atomic_read(&this->nr_iowait); 2091} 2092 2093unsigned long this_cpu_load(void) 2094{ 2095 struct rq *this = this_rq(); 2096 return this->cpu_load[0]; 2097} 2098 2099 2100/* Variables and functions for calc_load */ 2101static atomic_long_t calc_load_tasks; 2102static unsigned long calc_load_update; 2103unsigned long avenrun[3]; 2104EXPORT_SYMBOL(avenrun); 2105 2106static long calc_load_fold_active(struct rq *this_rq) 2107{ 2108 long nr_active, delta = 0; 2109 2110 nr_active = this_rq->nr_running; 2111 nr_active += (long) this_rq->nr_uninterruptible; 2112 2113 if (nr_active != this_rq->calc_load_active) { 2114 delta = nr_active - this_rq->calc_load_active; 2115 this_rq->calc_load_active = nr_active; 2116 } 2117 2118 return delta; 2119} 2120 2121static unsigned long 2122calc_load(unsigned long load, unsigned long exp, unsigned long active) 2123{ 2124 load *= exp; 2125 load += active * (FIXED_1 - exp); 2126 load += 1UL << (FSHIFT - 1); 2127 return load >> FSHIFT; 2128} 2129 2130#ifdef CONFIG_NO_HZ 2131/* 2132 * For NO_HZ we delay the active fold to the next LOAD_FREQ update. 2133 * 2134 * When making the ILB scale, we should try to pull this in as well. 2135 */ 2136static atomic_long_t calc_load_tasks_idle; 2137 2138void calc_load_account_idle(struct rq *this_rq) 2139{ 2140 long delta; 2141 2142 delta = calc_load_fold_active(this_rq); 2143 if (delta) 2144 atomic_long_add(delta, &calc_load_tasks_idle); 2145} 2146 2147static long calc_load_fold_idle(void) 2148{ 2149 long delta = 0; 2150 2151 /* 2152 * Its got a race, we don't care... 2153 */ 2154 if (atomic_long_read(&calc_load_tasks_idle)) 2155 delta = atomic_long_xchg(&calc_load_tasks_idle, 0); 2156 2157 return delta; 2158} 2159 2160/** 2161 * fixed_power_int - compute: x^n, in O(log n) time 2162 * 2163 * @x: base of the power 2164 * @frac_bits: fractional bits of @x 2165 * @n: power to raise @x to. 2166 * 2167 * By exploiting the relation between the definition of the natural power 2168 * function: x^n := x*x*...*x (x multiplied by itself for n times), and 2169 * the binary encoding of numbers used by computers: n := \Sum n_i * 2^i, 2170 * (where: n_i \elem {0, 1}, the binary vector representing n), 2171 * we find: x^n := x^(\Sum n_i * 2^i) := \Prod x^(n_i * 2^i), which is 2172 * of course trivially computable in O(log_2 n), the length of our binary 2173 * vector. 2174 */ 2175static unsigned long 2176fixed_power_int(unsigned long x, unsigned int frac_bits, unsigned int n) 2177{ 2178 unsigned long result = 1UL << frac_bits; 2179 2180 if (n) for (;;) { 2181 if (n & 1) { 2182 result *= x; 2183 result += 1UL << (frac_bits - 1); 2184 result >>= frac_bits; 2185 } 2186 n >>= 1; 2187 if (!n) 2188 break; 2189 x *= x; 2190 x += 1UL << (frac_bits - 1); 2191 x >>= frac_bits; 2192 } 2193 2194 return result; 2195} 2196 2197/* 2198 * a1 = a0 * e + a * (1 - e) 2199 * 2200 * a2 = a1 * e + a * (1 - e) 2201 * = (a0 * e + a * (1 - e)) * e + a * (1 - e) 2202 * = a0 * e^2 + a * (1 - e) * (1 + e) 2203 * 2204 * a3 = a2 * e + a * (1 - e) 2205 * = (a0 * e^2 + a * (1 - e) * (1 + e)) * e + a * (1 - e) 2206 * = a0 * e^3 + a * (1 - e) * (1 + e + e^2) 2207 * 2208 * ... 2209 * 2210 * an = a0 * e^n + a * (1 - e) * (1 + e + ... + e^n-1) [1] 2211 * = a0 * e^n + a * (1 - e) * (1 - e^n)/(1 - e) 2212 * = a0 * e^n + a * (1 - e^n) 2213 * 2214 * [1] application of the geometric series: 2215 * 2216 * n 1 - x^(n+1) 2217 * S_n := \Sum x^i = ------------- 2218 * i=0 1 - x 2219 */ 2220static unsigned long 2221calc_load_n(unsigned long load, unsigned long exp, 2222 unsigned long active, unsigned int n) 2223{ 2224 2225 return calc_load(load, fixed_power_int(exp, FSHIFT, n), active); 2226} 2227 2228/* 2229 * NO_HZ can leave us missing all per-cpu ticks calling 2230 * calc_load_account_active(), but since an idle CPU folds its delta into 2231 * calc_load_tasks_idle per calc_load_account_idle(), all we need to do is fold 2232 * in the pending idle delta if our idle period crossed a load cycle boundary. 2233 * 2234 * Once we've updated the global active value, we need to apply the exponential 2235 * weights adjusted to the number of cycles missed. 2236 */ 2237static void calc_global_nohz(unsigned long ticks) 2238{ 2239 long delta, active, n; 2240 2241 if (time_before(jiffies, calc_load_update)) 2242 return; 2243 2244 /* 2245 * If we crossed a calc_load_update boundary, make sure to fold 2246 * any pending idle changes, the respective CPUs might have 2247 * missed the tick driven calc_load_account_active() update 2248 * due to NO_HZ. 2249 */ 2250 delta = calc_load_fold_idle(); 2251 if (delta) 2252 atomic_long_add(delta, &calc_load_tasks); 2253 2254 /* 2255 * If we were idle for multiple load cycles, apply them. 2256 */ 2257 if (ticks >= LOAD_FREQ) { 2258 n = ticks / LOAD_FREQ; 2259 2260 active = atomic_long_read(&calc_load_tasks); 2261 active = active > 0 ? active * FIXED_1 : 0; 2262 2263 avenrun[0] = calc_load_n(avenrun[0], EXP_1, active, n); 2264 avenrun[1] = calc_load_n(avenrun[1], EXP_5, active, n); 2265 avenrun[2] = calc_load_n(avenrun[2], EXP_15, active, n); 2266 2267 calc_load_update += n * LOAD_FREQ; 2268 } 2269 2270 /* 2271 * Its possible the remainder of the above division also crosses 2272 * a LOAD_FREQ period, the regular check in calc_global_load() 2273 * which comes after this will take care of that. 2274 * 2275 * Consider us being 11 ticks before a cycle completion, and us 2276 * sleeping for 4*LOAD_FREQ + 22 ticks, then the above code will 2277 * age us 4 cycles, and the test in calc_global_load() will 2278 * pick up the final one. 2279 */ 2280} 2281#else 2282void calc_load_account_idle(struct rq *this_rq) 2283{ 2284} 2285 2286static inline long calc_load_fold_idle(void) 2287{ 2288 return 0; 2289} 2290 2291static void calc_global_nohz(unsigned long ticks) 2292{ 2293} 2294#endif 2295 2296/** 2297 * get_avenrun - get the load average array 2298 * @loads: pointer to dest load array 2299 * @offset: offset to add 2300 * @shift: shift count to shift the result left 2301 * 2302 * These values are estimates at best, so no need for locking. 2303 */ 2304void get_avenrun(unsigned long *loads, unsigned long offset, int shift) 2305{ 2306 loads[0] = (avenrun[0] + offset) << shift; 2307 loads[1] = (avenrun[1] + offset) << shift; 2308 loads[2] = (avenrun[2] + offset) << shift; 2309} 2310 2311/* 2312 * calc_load - update the avenrun load estimates 10 ticks after the 2313 * CPUs have updated calc_load_tasks. 2314 */ 2315void calc_global_load(unsigned long ticks) 2316{ 2317 long active; 2318 2319 calc_global_nohz(ticks); 2320 2321 if (time_before(jiffies, calc_load_update + 10)) 2322 return; 2323 2324 active = atomic_long_read(&calc_load_tasks); 2325 active = active > 0 ? active * FIXED_1 : 0; 2326 2327 avenrun[0] = calc_load(avenrun[0], EXP_1, active); 2328 avenrun[1] = calc_load(avenrun[1], EXP_5, active); 2329 avenrun[2] = calc_load(avenrun[2], EXP_15, active); 2330 2331 calc_load_update += LOAD_FREQ; 2332} 2333 2334/* 2335 * Called from update_cpu_load() to periodically update this CPU's 2336 * active count. 2337 */ 2338static void calc_load_account_active(struct rq *this_rq) 2339{ 2340 long delta; 2341 2342 if (time_before(jiffies, this_rq->calc_load_update)) 2343 return; 2344 2345 delta = calc_load_fold_active(this_rq); 2346 delta += calc_load_fold_idle(); 2347 if (delta) 2348 atomic_long_add(delta, &calc_load_tasks); 2349 2350 this_rq->calc_load_update += LOAD_FREQ; 2351} 2352 2353/* 2354 * The exact cpuload at various idx values, calculated at every tick would be 2355 * load = (2^idx - 1) / 2^idx * load + 1 / 2^idx * cur_load 2356 * 2357 * If a cpu misses updates for n-1 ticks (as it was idle) and update gets called 2358 * on nth tick when cpu may be busy, then we have: 2359 * load = ((2^idx - 1) / 2^idx)^(n-1) * load 2360 * load = (2^idx - 1) / 2^idx) * load + 1 / 2^idx * cur_load 2361 * 2362 * decay_load_missed() below does efficient calculation of 2363 * load = ((2^idx - 1) / 2^idx)^(n-1) * load 2364 * avoiding 0..n-1 loop doing load = ((2^idx - 1) / 2^idx) * load 2365 * 2366 * The calculation is approximated on a 128 point scale. 2367 * degrade_zero_ticks is the number of ticks after which load at any 2368 * particular idx is approximated to be zero. 2369 * degrade_factor is a precomputed table, a row for each load idx. 2370 * Each column corresponds to degradation factor for a power of two ticks, 2371 * based on 128 point scale. 2372 * Example: 2373 * row 2, col 3 (=12) says that the degradation at load idx 2 after 2374 * 8 ticks is 12/128 (which is an approximation of exact factor 3^8/4^8). 2375 * 2376 * With this power of 2 load factors, we can degrade the load n times 2377 * by looking at 1 bits in n and doing as many mult/shift instead of 2378 * n mult/shifts needed by the exact degradation. 2379 */ 2380#define DEGRADE_SHIFT 7 2381static const unsigned char 2382 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128}; 2383static const unsigned char 2384 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = { 2385 {0, 0, 0, 0, 0, 0, 0, 0}, 2386 {64, 32, 8, 0, 0, 0, 0, 0}, 2387 {96, 72, 40, 12, 1, 0, 0}, 2388 {112, 98, 75, 43, 15, 1, 0}, 2389 {120, 112, 98, 76, 45, 16, 2} }; 2390 2391/* 2392 * Update cpu_load for any missed ticks, due to tickless idle. The backlog 2393 * would be when CPU is idle and so we just decay the old load without 2394 * adding any new load. 2395 */ 2396static unsigned long 2397decay_load_missed(unsigned long load, unsigned long missed_updates, int idx) 2398{ 2399 int j = 0; 2400 2401 if (!missed_updates) 2402 return load; 2403 2404 if (missed_updates >= degrade_zero_ticks[idx]) 2405 return 0; 2406 2407 if (idx == 1) 2408 return load >> missed_updates; 2409 2410 while (missed_updates) { 2411 if (missed_updates % 2) 2412 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT; 2413 2414 missed_updates >>= 1; 2415 j++; 2416 } 2417 return load; 2418} 2419 2420/* 2421 * Update rq->cpu_load[] statistics. This function is usually called every 2422 * scheduler tick (TICK_NSEC). With tickless idle this will not be called 2423 * every tick. We fix it up based on jiffies. 2424 */ 2425void update_cpu_load(struct rq *this_rq) 2426{ 2427 unsigned long this_load = this_rq->load.weight; 2428 unsigned long curr_jiffies = jiffies; 2429 unsigned long pending_updates; 2430 int i, scale; 2431 2432 this_rq->nr_load_updates++; 2433 2434 /* Avoid repeated calls on same jiffy, when moving in and out of idle */ 2435 if (curr_jiffies == this_rq->last_load_update_tick) 2436 return; 2437 2438 pending_updates = curr_jiffies - this_rq->last_load_update_tick; 2439 this_rq->last_load_update_tick = curr_jiffies; 2440 2441 /* Update our load: */ 2442 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */ 2443 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) { 2444 unsigned long old_load, new_load; 2445 2446 /* scale is effectively 1 << i now, and >> i divides by scale */ 2447 2448 old_load = this_rq->cpu_load[i]; 2449 old_load = decay_load_missed(old_load, pending_updates - 1, i); 2450 new_load = this_load; 2451 /* 2452 * Round up the averaging division if load is increasing. This 2453 * prevents us from getting stuck on 9 if the load is 10, for 2454 * example. 2455 */ 2456 if (new_load > old_load) 2457 new_load += scale - 1; 2458 2459 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i; 2460 } 2461 2462 sched_avg_update(this_rq); 2463} 2464 2465static void update_cpu_load_active(struct rq *this_rq) 2466{ 2467 update_cpu_load(this_rq); 2468 2469 calc_load_account_active(this_rq); 2470} 2471 2472#ifdef CONFIG_SMP 2473 2474/* 2475 * sched_exec - execve() is a valuable balancing opportunity, because at 2476 * this point the task has the smallest effective memory and cache footprint. 2477 */ 2478void sched_exec(void) 2479{ 2480 struct task_struct *p = current; 2481 unsigned long flags; 2482 int dest_cpu; 2483 2484 raw_spin_lock_irqsave(&p->pi_lock, flags); 2485 dest_cpu = p->sched_class->select_task_rq(p, SD_BALANCE_EXEC, 0); 2486 if (dest_cpu == smp_processor_id()) 2487 goto unlock; 2488 2489 if (likely(cpu_active(dest_cpu))) { 2490 struct migration_arg arg = { p, dest_cpu }; 2491 2492 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 2493 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg); 2494 return; 2495 } 2496unlock: 2497 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 2498} 2499 2500#endif 2501 2502DEFINE_PER_CPU(struct kernel_stat, kstat); 2503 2504EXPORT_PER_CPU_SYMBOL(kstat); 2505 2506/* 2507 * Return any ns on the sched_clock that have not yet been accounted in 2508 * @p in case that task is currently running. 2509 * 2510 * Called with task_rq_lock() held on @rq. 2511 */ 2512static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq) 2513{ 2514 u64 ns = 0; 2515 2516 if (task_current(rq, p)) { 2517 update_rq_clock(rq); 2518 ns = rq->clock_task - p->se.exec_start; 2519 if ((s64)ns < 0) 2520 ns = 0; 2521 } 2522 2523 return ns; 2524} 2525 2526unsigned long long task_delta_exec(struct task_struct *p) 2527{ 2528 unsigned long flags; 2529 struct rq *rq; 2530 u64 ns = 0; 2531 2532 rq = task_rq_lock(p, &flags); 2533 ns = do_task_delta_exec(p, rq); 2534 task_rq_unlock(rq, p, &flags); 2535 2536 return ns; 2537} 2538 2539/* 2540 * Return accounted runtime for the task. 2541 * In case the task is currently running, return the runtime plus current's 2542 * pending runtime that have not been accounted yet. 2543 */ 2544unsigned long long task_sched_runtime(struct task_struct *p) 2545{ 2546 unsigned long flags; 2547 struct rq *rq; 2548 u64 ns = 0; 2549 2550 rq = task_rq_lock(p, &flags); 2551 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq); 2552 task_rq_unlock(rq, p, &flags); 2553 2554 return ns; 2555} 2556 2557/* 2558 * Account user cpu time to a process. 2559 * @p: the process that the cpu time gets accounted to 2560 * @cputime: the cpu time spent in user space since the last update 2561 * @cputime_scaled: cputime scaled by cpu frequency 2562 */ 2563void account_user_time(struct task_struct *p, cputime_t cputime, 2564 cputime_t cputime_scaled) 2565{ 2566 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat; 2567 cputime64_t tmp; 2568 2569 /* Add user time to process. */ 2570 p->utime = cputime_add(p->utime, cputime); 2571 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled); 2572 account_group_user_time(p, cputime); 2573 2574 /* Add user time to cpustat. */ 2575 tmp = cputime_to_cputime64(cputime); 2576 if (TASK_NICE(p) > 0) 2577 cpustat->nice = cputime64_add(cpustat->nice, tmp); 2578 else 2579 cpustat->user = cputime64_add(cpustat->user, tmp); 2580 2581 cpuacct_update_stats(p, CPUACCT_STAT_USER, cputime); 2582 /* Account for user time used */ 2583 acct_update_integrals(p); 2584} 2585 2586/* 2587 * Account guest cpu time to a process. 2588 * @p: the process that the cpu time gets accounted to 2589 * @cputime: the cpu time spent in virtual machine since the last update 2590 * @cputime_scaled: cputime scaled by cpu frequency 2591 */ 2592static void account_guest_time(struct task_struct *p, cputime_t cputime, 2593 cputime_t cputime_scaled) 2594{ 2595 cputime64_t tmp; 2596 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat; 2597 2598 tmp = cputime_to_cputime64(cputime); 2599 2600 /* Add guest time to process. */ 2601 p->utime = cputime_add(p->utime, cputime); 2602 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled); 2603 account_group_user_time(p, cputime); 2604 p->gtime = cputime_add(p->gtime, cputime); 2605 2606 /* Add guest time to cpustat. */ 2607 if (TASK_NICE(p) > 0) { 2608 cpustat->nice = cputime64_add(cpustat->nice, tmp); 2609 cpustat->guest_nice = cputime64_add(cpustat->guest_nice, tmp); 2610 } else { 2611 cpustat->user = cputime64_add(cpustat->user, tmp); 2612 cpustat->guest = cputime64_add(cpustat->guest, tmp); 2613 } 2614} 2615 2616/* 2617 * Account system cpu time to a process and desired cpustat field 2618 * @p: the process that the cpu time gets accounted to 2619 * @cputime: the cpu time spent in kernel space since the last update 2620 * @cputime_scaled: cputime scaled by cpu frequency 2621 * @target_cputime64: pointer to cpustat field that has to be updated 2622 */ 2623static inline 2624void __account_system_time(struct task_struct *p, cputime_t cputime, 2625 cputime_t cputime_scaled, cputime64_t *target_cputime64) 2626{ 2627 cputime64_t tmp = cputime_to_cputime64(cputime); 2628 2629 /* Add system time to process. */ 2630 p->stime = cputime_add(p->stime, cputime); 2631 p->stimescaled = cputime_add(p->stimescaled, cputime_scaled); 2632 account_group_system_time(p, cputime); 2633 2634 /* Add system time to cpustat. */ 2635 *target_cputime64 = cputime64_add(*target_cputime64, tmp); 2636 cpuacct_update_stats(p, CPUACCT_STAT_SYSTEM, cputime); 2637 2638 /* Account for system time used */ 2639 acct_update_integrals(p); 2640} 2641 2642/* 2643 * Account system cpu time to a process. 2644 * @p: the process that the cpu time gets accounted to 2645 * @hardirq_offset: the offset to subtract from hardirq_count() 2646 * @cputime: the cpu time spent in kernel space since the last update 2647 * @cputime_scaled: cputime scaled by cpu frequency 2648 */ 2649void account_system_time(struct task_struct *p, int hardirq_offset, 2650 cputime_t cputime, cputime_t cputime_scaled) 2651{ 2652 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat; 2653 cputime64_t *target_cputime64; 2654 2655 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) { 2656 account_guest_time(p, cputime, cputime_scaled); 2657 return; 2658 } 2659 2660 if (hardirq_count() - hardirq_offset) 2661 target_cputime64 = &cpustat->irq; 2662 else if (in_serving_softirq()) 2663 target_cputime64 = &cpustat->softirq; 2664 else 2665 target_cputime64 = &cpustat->system; 2666 2667 __account_system_time(p, cputime, cputime_scaled, target_cputime64); 2668} 2669 2670/* 2671 * Account for involuntary wait time. 2672 * @cputime: the cpu time spent in involuntary wait 2673 */ 2674void account_steal_time(cputime_t cputime) 2675{ 2676 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat; 2677 cputime64_t cputime64 = cputime_to_cputime64(cputime); 2678 2679 cpustat->steal = cputime64_add(cpustat->steal, cputime64); 2680} 2681 2682/* 2683 * Account for idle time. 2684 * @cputime: the cpu time spent in idle wait 2685 */ 2686void account_idle_time(cputime_t cputime) 2687{ 2688 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat; 2689 cputime64_t cputime64 = cputime_to_cputime64(cputime); 2690 struct rq *rq = this_rq(); 2691 2692 if (atomic_read(&rq->nr_iowait) > 0) 2693 cpustat->iowait = cputime64_add(cpustat->iowait, cputime64); 2694 else 2695 cpustat->idle = cputime64_add(cpustat->idle, cputime64); 2696} 2697 2698static __always_inline bool steal_account_process_tick(void) 2699{ 2700#ifdef CONFIG_PARAVIRT 2701 if (static_branch(¶virt_steal_enabled)) { 2702 u64 steal, st = 0; 2703 2704 steal = paravirt_steal_clock(smp_processor_id()); 2705 steal -= this_rq()->prev_steal_time; 2706 2707 st = steal_ticks(steal); 2708 this_rq()->prev_steal_time += st * TICK_NSEC; 2709 2710 account_steal_time(st); 2711 return st; 2712 } 2713#endif 2714 return false; 2715} 2716 2717#ifndef CONFIG_VIRT_CPU_ACCOUNTING 2718 2719#ifdef CONFIG_IRQ_TIME_ACCOUNTING 2720/* 2721 * Account a tick to a process and cpustat 2722 * @p: the process that the cpu time gets accounted to 2723 * @user_tick: is the tick from userspace 2724 * @rq: the pointer to rq 2725 * 2726 * Tick demultiplexing follows the order 2727 * - pending hardirq update 2728 * - pending softirq update 2729 * - user_time 2730 * - idle_time 2731 * - system time 2732 * - check for guest_time 2733 * - else account as system_time 2734 * 2735 * Check for hardirq is done both for system and user time as there is 2736 * no timer going off while we are on hardirq and hence we may never get an 2737 * opportunity to update it solely in system time. 2738 * p->stime and friends are only updated on system time and not on irq 2739 * softirq as those do not count in task exec_runtime any more. 2740 */ 2741static void irqtime_account_process_tick(struct task_struct *p, int user_tick, 2742 struct rq *rq) 2743{ 2744 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy); 2745 cputime64_t tmp = cputime_to_cputime64(cputime_one_jiffy); 2746 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat; 2747 2748 if (steal_account_process_tick()) 2749 return; 2750 2751 if (irqtime_account_hi_update()) { 2752 cpustat->irq = cputime64_add(cpustat->irq, tmp); 2753 } else if (irqtime_account_si_update()) { 2754 cpustat->softirq = cputime64_add(cpustat->softirq, tmp); 2755 } else if (this_cpu_ksoftirqd() == p) { 2756 /* 2757 * ksoftirqd time do not get accounted in cpu_softirq_time. 2758 * So, we have to handle it separately here. 2759 * Also, p->stime needs to be updated for ksoftirqd. 2760 */ 2761 __account_system_time(p, cputime_one_jiffy, one_jiffy_scaled, 2762 &cpustat->softirq); 2763 } else if (user_tick) { 2764 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled); 2765 } else if (p == rq->idle) { 2766 account_idle_time(cputime_one_jiffy); 2767 } else if (p->flags & PF_VCPU) { /* System time or guest time */ 2768 account_guest_time(p, cputime_one_jiffy, one_jiffy_scaled); 2769 } else { 2770 __account_system_time(p, cputime_one_jiffy, one_jiffy_scaled, 2771 &cpustat->system); 2772 } 2773} 2774 2775static void irqtime_account_idle_ticks(int ticks) 2776{ 2777 int i; 2778 struct rq *rq = this_rq(); 2779 2780 for (i = 0; i < ticks; i++) 2781 irqtime_account_process_tick(current, 0, rq); 2782} 2783#else /* CONFIG_IRQ_TIME_ACCOUNTING */ 2784static void irqtime_account_idle_ticks(int ticks) {} 2785static void irqtime_account_process_tick(struct task_struct *p, int user_tick, 2786 struct rq *rq) {} 2787#endif /* CONFIG_IRQ_TIME_ACCOUNTING */ 2788 2789/* 2790 * Account a single tick of cpu time. 2791 * @p: the process that the cpu time gets accounted to 2792 * @user_tick: indicates if the tick is a user or a system tick 2793 */ 2794void account_process_tick(struct task_struct *p, int user_tick) 2795{ 2796 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy); 2797 struct rq *rq = this_rq(); 2798 2799 if (sched_clock_irqtime) { 2800 irqtime_account_process_tick(p, user_tick, rq); 2801 return; 2802 } 2803 2804 if (steal_account_process_tick()) 2805 return; 2806 2807 if (user_tick) 2808 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled); 2809 else if ((p != rq->idle) || (irq_count() != HARDIRQ_OFFSET)) 2810 account_system_time(p, HARDIRQ_OFFSET, cputime_one_jiffy, 2811 one_jiffy_scaled); 2812 else 2813 account_idle_time(cputime_one_jiffy); 2814} 2815 2816/* 2817 * Account multiple ticks of steal time. 2818 * @p: the process from which the cpu time has been stolen 2819 * @ticks: number of stolen ticks 2820 */ 2821void account_steal_ticks(unsigned long ticks) 2822{ 2823 account_steal_time(jiffies_to_cputime(ticks)); 2824} 2825 2826/* 2827 * Account multiple ticks of idle time. 2828 * @ticks: number of stolen ticks 2829 */ 2830void account_idle_ticks(unsigned long ticks) 2831{ 2832 2833 if (sched_clock_irqtime) { 2834 irqtime_account_idle_ticks(ticks); 2835 return; 2836 } 2837 2838 account_idle_time(jiffies_to_cputime(ticks)); 2839} 2840 2841#endif 2842 2843/* 2844 * Use precise platform statistics if available: 2845 */ 2846#ifdef CONFIG_VIRT_CPU_ACCOUNTING 2847void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st) 2848{ 2849 *ut = p->utime; 2850 *st = p->stime; 2851} 2852 2853void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st) 2854{ 2855 struct task_cputime cputime; 2856 2857 thread_group_cputime(p, &cputime); 2858 2859 *ut = cputime.utime; 2860 *st = cputime.stime; 2861} 2862#else 2863 2864#ifndef nsecs_to_cputime 2865# define nsecs_to_cputime(__nsecs) nsecs_to_jiffies(__nsecs) 2866#endif 2867 2868void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st) 2869{ 2870 cputime_t rtime, utime = p->utime, total = cputime_add(utime, p->stime); 2871 2872 /* 2873 * Use CFS's precise accounting: 2874 */ 2875 rtime = nsecs_to_cputime(p->se.sum_exec_runtime); 2876 2877 if (total) { 2878 u64 temp = rtime; 2879 2880 temp *= utime; 2881 do_div(temp, total); 2882 utime = (cputime_t)temp; 2883 } else 2884 utime = rtime; 2885 2886 /* 2887 * Compare with previous values, to keep monotonicity: 2888 */ 2889 p->prev_utime = max(p->prev_utime, utime); 2890 p->prev_stime = max(p->prev_stime, cputime_sub(rtime, p->prev_utime)); 2891 2892 *ut = p->prev_utime; 2893 *st = p->prev_stime; 2894} 2895 2896/* 2897 * Must be called with siglock held. 2898 */ 2899void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st) 2900{ 2901 struct signal_struct *sig = p->signal; 2902 struct task_cputime cputime; 2903 cputime_t rtime, utime, total; 2904 2905 thread_group_cputime(p, &cputime); 2906 2907 total = cputime_add(cputime.utime, cputime.stime); 2908 rtime = nsecs_to_cputime(cputime.sum_exec_runtime); 2909 2910 if (total) { 2911 u64 temp = rtime; 2912 2913 temp *= cputime.utime; 2914 do_div(temp, total); 2915 utime = (cputime_t)temp; 2916 } else 2917 utime = rtime; 2918 2919 sig->prev_utime = max(sig->prev_utime, utime); 2920 sig->prev_stime = max(sig->prev_stime, 2921 cputime_sub(rtime, sig->prev_utime)); 2922 2923 *ut = sig->prev_utime; 2924 *st = sig->prev_stime; 2925} 2926#endif 2927 2928/* 2929 * This function gets called by the timer code, with HZ frequency. 2930 * We call it with interrupts disabled. 2931 */ 2932void scheduler_tick(void) 2933{ 2934 int cpu = smp_processor_id(); 2935 struct rq *rq = cpu_rq(cpu); 2936 struct task_struct *curr = rq->curr; 2937 2938 sched_clock_tick(); 2939 2940 raw_spin_lock(&rq->lock); 2941 update_rq_clock(rq); 2942 update_cpu_load_active(rq); 2943 curr->sched_class->task_tick(rq, curr, 0); 2944 raw_spin_unlock(&rq->lock); 2945 2946 perf_event_task_tick(); 2947 2948#ifdef CONFIG_SMP 2949 rq->idle_balance = idle_cpu(cpu); 2950 trigger_load_balance(rq, cpu); 2951#endif 2952} 2953 2954notrace unsigned long get_parent_ip(unsigned long addr) 2955{ 2956 if (in_lock_functions(addr)) { 2957 addr = CALLER_ADDR2; 2958 if (in_lock_functions(addr)) 2959 addr = CALLER_ADDR3; 2960 } 2961 return addr; 2962} 2963 2964#if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \ 2965 defined(CONFIG_PREEMPT_TRACER)) 2966 2967void __kprobes add_preempt_count(int val) 2968{ 2969#ifdef CONFIG_DEBUG_PREEMPT 2970 /* 2971 * Underflow? 2972 */ 2973 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0))) 2974 return; 2975#endif 2976 preempt_count() += val; 2977#ifdef CONFIG_DEBUG_PREEMPT 2978 /* 2979 * Spinlock count overflowing soon? 2980 */ 2981 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >= 2982 PREEMPT_MASK - 10); 2983#endif 2984 if (preempt_count() == val) 2985 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1)); 2986} 2987EXPORT_SYMBOL(add_preempt_count); 2988 2989void __kprobes sub_preempt_count(int val) 2990{ 2991#ifdef CONFIG_DEBUG_PREEMPT 2992 /* 2993 * Underflow? 2994 */ 2995 if (DEBUG_LOCKS_WARN_ON(val > preempt_count())) 2996 return; 2997 /* 2998 * Is the spinlock portion underflowing? 2999 */ 3000 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) && 3001 !(preempt_count() & PREEMPT_MASK))) 3002 return; 3003#endif 3004 3005 if (preempt_count() == val) 3006 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1)); 3007 preempt_count() -= val; 3008} 3009EXPORT_SYMBOL(sub_preempt_count); 3010 3011#endif 3012 3013/* 3014 * Print scheduling while atomic bug: 3015 */ 3016static noinline void __schedule_bug(struct task_struct *prev) 3017{ 3018 struct pt_regs *regs = get_irq_regs(); 3019 3020 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n", 3021 prev->comm, prev->pid, preempt_count()); 3022 3023 debug_show_held_locks(prev); 3024 print_modules(); 3025 if (irqs_disabled()) 3026 print_irqtrace_events(prev); 3027 3028 if (regs) 3029 show_regs(regs); 3030 else 3031 dump_stack(); 3032} 3033 3034/* 3035 * Various schedule()-time debugging checks and statistics: 3036 */ 3037static inline void schedule_debug(struct task_struct *prev) 3038{ 3039 /* 3040 * Test if we are atomic. Since do_exit() needs to call into 3041 * schedule() atomically, we ignore that path for now. 3042 * Otherwise, whine if we are scheduling when we should not be. 3043 */ 3044 if (unlikely(in_atomic_preempt_off() && !prev->exit_state)) 3045 __schedule_bug(prev); 3046 rcu_sleep_check(); 3047 3048 profile_hit(SCHED_PROFILING, __builtin_return_address(0)); 3049 3050 schedstat_inc(this_rq(), sched_count); 3051} 3052 3053static void put_prev_task(struct rq *rq, struct task_struct *prev) 3054{ 3055 if (prev->on_rq || rq->skip_clock_update < 0) 3056 update_rq_clock(rq); 3057 prev->sched_class->put_prev_task(rq, prev); 3058} 3059 3060/* 3061 * Pick up the highest-prio task: 3062 */ 3063static inline struct task_struct * 3064pick_next_task(struct rq *rq) 3065{ 3066 const struct sched_class *class; 3067 struct task_struct *p; 3068 3069 /* 3070 * Optimization: we know that if all tasks are in 3071 * the fair class we can call that function directly: 3072 */ 3073 if (likely(rq->nr_running == rq->cfs.h_nr_running)) { 3074 p = fair_sched_class.pick_next_task(rq); 3075 if (likely(p)) 3076 return p; 3077 } 3078 3079 for_each_class(class) { 3080 p = class->pick_next_task(rq); 3081 if (p) 3082 return p; 3083 } 3084 3085 BUG(); /* the idle class will always have a runnable task */ 3086} 3087 3088/* 3089 * __schedule() is the main scheduler function. 3090 */ 3091static void __sched __schedule(void) 3092{ 3093 struct task_struct *prev, *next; 3094 unsigned long *switch_count; 3095 struct rq *rq; 3096 int cpu; 3097 3098need_resched: 3099 preempt_disable(); 3100 cpu = smp_processor_id(); 3101 rq = cpu_rq(cpu); 3102 rcu_note_context_switch(cpu); 3103 prev = rq->curr; 3104 3105 schedule_debug(prev); 3106 3107 if (sched_feat(HRTICK)) 3108 hrtick_clear(rq); 3109 3110 raw_spin_lock_irq(&rq->lock); 3111 3112 switch_count = &prev->nivcsw; 3113 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) { 3114 if (unlikely(signal_pending_state(prev->state, prev))) { 3115 prev->state = TASK_RUNNING; 3116 } else { 3117 deactivate_task(rq, prev, DEQUEUE_SLEEP); 3118 prev->on_rq = 0; 3119 3120 /* 3121 * If a worker went to sleep, notify and ask workqueue 3122 * whether it wants to wake up a task to maintain 3123 * concurrency. 3124 */ 3125 if (prev->flags & PF_WQ_WORKER) { 3126 struct task_struct *to_wakeup; 3127 3128 to_wakeup = wq_worker_sleeping(prev, cpu); 3129 if (to_wakeup) 3130 try_to_wake_up_local(to_wakeup); 3131 } 3132 } 3133 switch_count = &prev->nvcsw; 3134 } 3135 3136 pre_schedule(rq, prev); 3137 3138 if (unlikely(!rq->nr_running)) 3139 idle_balance(cpu, rq); 3140 3141 put_prev_task(rq, prev); 3142 next = pick_next_task(rq); 3143 clear_tsk_need_resched(prev); 3144 rq->skip_clock_update = 0; 3145 3146 if (likely(prev != next)) { 3147 rq->nr_switches++; 3148 rq->curr = next; 3149 ++*switch_count; 3150 3151 context_switch(rq, prev, next); /* unlocks the rq */ 3152 /* 3153 * The context switch have flipped the stack from under us 3154 * and restored the local variables which were saved when 3155 * this task called schedule() in the past. prev == current 3156 * is still correct, but it can be moved to another cpu/rq. 3157 */ 3158 cpu = smp_processor_id(); 3159 rq = cpu_rq(cpu); 3160 } else 3161 raw_spin_unlock_irq(&rq->lock); 3162 3163 post_schedule(rq); 3164 3165 preempt_enable_no_resched(); 3166 if (need_resched()) 3167 goto need_resched; 3168} 3169 3170static inline void sched_submit_work(struct task_struct *tsk) 3171{ 3172 if (!tsk->state) 3173 return; 3174 /* 3175 * If we are going to sleep and we have plugged IO queued, 3176 * make sure to submit it to avoid deadlocks. 3177 */ 3178 if (blk_needs_flush_plug(tsk)) 3179 blk_schedule_flush_plug(tsk); 3180} 3181 3182asmlinkage void __sched schedule(void) 3183{ 3184 struct task_struct *tsk = current; 3185 3186 sched_submit_work(tsk); 3187 __schedule(); 3188} 3189EXPORT_SYMBOL(schedule); 3190 3191#ifdef CONFIG_MUTEX_SPIN_ON_OWNER 3192 3193static inline bool owner_running(struct mutex *lock, struct task_struct *owner) 3194{ 3195 if (lock->owner != owner) 3196 return false; 3197 3198 /* 3199 * Ensure we emit the owner->on_cpu, dereference _after_ checking 3200 * lock->owner still matches owner, if that fails, owner might 3201 * point to free()d memory, if it still matches, the rcu_read_lock() 3202 * ensures the memory stays valid. 3203 */ 3204 barrier(); 3205 3206 return owner->on_cpu; 3207} 3208 3209/* 3210 * Look out! "owner" is an entirely speculative pointer 3211 * access and not reliable. 3212 */ 3213int mutex_spin_on_owner(struct mutex *lock, struct task_struct *owner) 3214{ 3215 if (!sched_feat(OWNER_SPIN)) 3216 return 0; 3217 3218 rcu_read_lock(); 3219 while (owner_running(lock, owner)) { 3220 if (need_resched()) 3221 break; 3222 3223 arch_mutex_cpu_relax(); 3224 } 3225 rcu_read_unlock(); 3226 3227 /* 3228 * We break out the loop above on need_resched() and when the 3229 * owner changed, which is a sign for heavy contention. Return 3230 * success only when lock->owner is NULL. 3231 */ 3232 return lock->owner == NULL; 3233} 3234#endif 3235 3236#ifdef CONFIG_PREEMPT 3237/* 3238 * this is the entry point to schedule() from in-kernel preemption 3239 * off of preempt_enable. Kernel preemptions off return from interrupt 3240 * occur there and call schedule directly. 3241 */ 3242asmlinkage void __sched notrace preempt_schedule(void) 3243{ 3244 struct thread_info *ti = current_thread_info(); 3245 3246 /* 3247 * If there is a non-zero preempt_count or interrupts are disabled, 3248 * we do not want to preempt the current task. Just return.. 3249 */ 3250 if (likely(ti->preempt_count || irqs_disabled())) 3251 return; 3252 3253 do { 3254 add_preempt_count_notrace(PREEMPT_ACTIVE); 3255 __schedule(); 3256 sub_preempt_count_notrace(PREEMPT_ACTIVE); 3257 3258 /* 3259 * Check again in case we missed a preemption opportunity 3260 * between schedule and now. 3261 */ 3262 barrier(); 3263 } while (need_resched()); 3264} 3265EXPORT_SYMBOL(preempt_schedule); 3266 3267/* 3268 * this is the entry point to schedule() from kernel preemption 3269 * off of irq context. 3270 * Note, that this is called and return with irqs disabled. This will 3271 * protect us against recursive calling from irq. 3272 */ 3273asmlinkage void __sched preempt_schedule_irq(void) 3274{ 3275 struct thread_info *ti = current_thread_info(); 3276 3277 /* Catch callers which need to be fixed */ 3278 BUG_ON(ti->preempt_count || !irqs_disabled()); 3279 3280 do { 3281 add_preempt_count(PREEMPT_ACTIVE); 3282 local_irq_enable(); 3283 __schedule(); 3284 local_irq_disable(); 3285 sub_preempt_count(PREEMPT_ACTIVE); 3286 3287 /* 3288 * Check again in case we missed a preemption opportunity 3289 * between schedule and now. 3290 */ 3291 barrier(); 3292 } while (need_resched()); 3293} 3294 3295#endif /* CONFIG_PREEMPT */ 3296 3297int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags, 3298 void *key) 3299{ 3300 return try_to_wake_up(curr->private, mode, wake_flags); 3301} 3302EXPORT_SYMBOL(default_wake_function); 3303 3304/* 3305 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just 3306 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve 3307 * number) then we wake all the non-exclusive tasks and one exclusive task. 3308 * 3309 * There are circumstances in which we can try to wake a task which has already 3310 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns 3311 * zero in this (rare) case, and we handle it by continuing to scan the queue. 3312 */ 3313static void __wake_up_common(wait_queue_head_t *q, unsigned int mode, 3314 int nr_exclusive, int wake_flags, void *key) 3315{ 3316 wait_queue_t *curr, *next; 3317 3318 list_for_each_entry_safe(curr, next, &q->task_list, task_list) { 3319 unsigned flags = curr->flags; 3320 3321 if (curr->func(curr, mode, wake_flags, key) && 3322 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive) 3323 break; 3324 } 3325} 3326 3327/** 3328 * __wake_up - wake up threads blocked on a waitqueue. 3329 * @q: the waitqueue 3330 * @mode: which threads 3331 * @nr_exclusive: how many wake-one or wake-many threads to wake up 3332 * @key: is directly passed to the wakeup function 3333 * 3334 * It may be assumed that this function implies a write memory barrier before 3335 * changing the task state if and only if any tasks are woken up. 3336 */ 3337void __wake_up(wait_queue_head_t *q, unsigned int mode, 3338 int nr_exclusive, void *key) 3339{ 3340 unsigned long flags; 3341 3342 spin_lock_irqsave(&q->lock, flags); 3343 __wake_up_common(q, mode, nr_exclusive, 0, key); 3344 spin_unlock_irqrestore(&q->lock, flags); 3345} 3346EXPORT_SYMBOL(__wake_up); 3347 3348/* 3349 * Same as __wake_up but called with the spinlock in wait_queue_head_t held. 3350 */ 3351void __wake_up_locked(wait_queue_head_t *q, unsigned int mode) 3352{ 3353 __wake_up_common(q, mode, 1, 0, NULL); 3354} 3355EXPORT_SYMBOL_GPL(__wake_up_locked); 3356 3357void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key) 3358{ 3359 __wake_up_common(q, mode, 1, 0, key); 3360} 3361EXPORT_SYMBOL_GPL(__wake_up_locked_key); 3362 3363/** 3364 * __wake_up_sync_key - wake up threads blocked on a waitqueue. 3365 * @q: the waitqueue 3366 * @mode: which threads 3367 * @nr_exclusive: how many wake-one or wake-many threads to wake up 3368 * @key: opaque value to be passed to wakeup targets 3369 * 3370 * The sync wakeup differs that the waker knows that it will schedule 3371 * away soon, so while the target thread will be woken up, it will not 3372 * be migrated to another CPU - ie. the two threads are 'synchronized' 3373 * with each other. This can prevent needless bouncing between CPUs. 3374 * 3375 * On UP it can prevent extra preemption. 3376 * 3377 * It may be assumed that this function implies a write memory barrier before 3378 * changing the task state if and only if any tasks are woken up. 3379 */ 3380void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode, 3381 int nr_exclusive, void *key) 3382{ 3383 unsigned long flags; 3384 int wake_flags = WF_SYNC; 3385 3386 if (unlikely(!q)) 3387 return; 3388 3389 if (unlikely(!nr_exclusive)) 3390 wake_flags = 0; 3391 3392 spin_lock_irqsave(&q->lock, flags); 3393 __wake_up_common(q, mode, nr_exclusive, wake_flags, key); 3394 spin_unlock_irqrestore(&q->lock, flags); 3395} 3396EXPORT_SYMBOL_GPL(__wake_up_sync_key); 3397 3398/* 3399 * __wake_up_sync - see __wake_up_sync_key() 3400 */ 3401void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive) 3402{ 3403 __wake_up_sync_key(q, mode, nr_exclusive, NULL); 3404} 3405EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */ 3406 3407/** 3408 * complete: - signals a single thread waiting on this completion 3409 * @x: holds the state of this particular completion 3410 * 3411 * This will wake up a single thread waiting on this completion. Threads will be 3412 * awakened in the same order in which they were queued. 3413 * 3414 * See also complete_all(), wait_for_completion() and related routines. 3415 * 3416 * It may be assumed that this function implies a write memory barrier before 3417 * changing the task state if and only if any tasks are woken up. 3418 */ 3419void complete(struct completion *x) 3420{ 3421 unsigned long flags; 3422 3423 spin_lock_irqsave(&x->wait.lock, flags); 3424 x->done++; 3425 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL); 3426 spin_unlock_irqrestore(&x->wait.lock, flags); 3427} 3428EXPORT_SYMBOL(complete); 3429 3430/** 3431 * complete_all: - signals all threads waiting on this completion 3432 * @x: holds the state of this particular completion 3433 * 3434 * This will wake up all threads waiting on this particular completion event. 3435 * 3436 * It may be assumed that this function implies a write memory barrier before 3437 * changing the task state if and only if any tasks are woken up. 3438 */ 3439void complete_all(struct completion *x) 3440{ 3441 unsigned long flags; 3442 3443 spin_lock_irqsave(&x->wait.lock, flags); 3444 x->done += UINT_MAX/2; 3445 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL); 3446 spin_unlock_irqrestore(&x->wait.lock, flags); 3447} 3448EXPORT_SYMBOL(complete_all); 3449 3450static inline long __sched 3451do_wait_for_common(struct completion *x, long timeout, int state) 3452{ 3453 if (!x->done) { 3454 DECLARE_WAITQUEUE(wait, current); 3455 3456 __add_wait_queue_tail_exclusive(&x->wait, &wait); 3457 do { 3458 if (signal_pending_state(state, current)) { 3459 timeout = -ERESTARTSYS; 3460 break; 3461 } 3462 __set_current_state(state); 3463 spin_unlock_irq(&x->wait.lock); 3464 timeout = schedule_timeout(timeout); 3465 spin_lock_irq(&x->wait.lock); 3466 } while (!x->done && timeout); 3467 __remove_wait_queue(&x->wait, &wait); 3468 if (!x->done) 3469 return timeout; 3470 } 3471 x->done--; 3472 return timeout ?: 1; 3473} 3474 3475static long __sched 3476wait_for_common(struct completion *x, long timeout, int state) 3477{ 3478 might_sleep(); 3479 3480 spin_lock_irq(&x->wait.lock); 3481 timeout = do_wait_for_common(x, timeout, state); 3482 spin_unlock_irq(&x->wait.lock); 3483 return timeout; 3484} 3485 3486/** 3487 * wait_for_completion: - waits for completion of a task 3488 * @x: holds the state of this particular completion 3489 * 3490 * This waits to be signaled for completion of a specific task. It is NOT 3491 * interruptible and there is no timeout. 3492 * 3493 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout 3494 * and interrupt capability. Also see complete(). 3495 */ 3496void __sched wait_for_completion(struct completion *x) 3497{ 3498 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE); 3499} 3500EXPORT_SYMBOL(wait_for_completion); 3501 3502/** 3503 * wait_for_completion_timeout: - waits for completion of a task (w/timeout) 3504 * @x: holds the state of this particular completion 3505 * @timeout: timeout value in jiffies 3506 * 3507 * This waits for either a completion of a specific task to be signaled or for a 3508 * specified timeout to expire. The timeout is in jiffies. It is not 3509 * interruptible. 3510 * 3511 * The return value is 0 if timed out, and positive (at least 1, or number of 3512 * jiffies left till timeout) if completed. 3513 */ 3514unsigned long __sched 3515wait_for_completion_timeout(struct completion *x, unsigned long timeout) 3516{ 3517 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE); 3518} 3519EXPORT_SYMBOL(wait_for_completion_timeout); 3520 3521/** 3522 * wait_for_completion_interruptible: - waits for completion of a task (w/intr) 3523 * @x: holds the state of this particular completion 3524 * 3525 * This waits for completion of a specific task to be signaled. It is 3526 * interruptible. 3527 * 3528 * The return value is -ERESTARTSYS if interrupted, 0 if completed. 3529 */ 3530int __sched wait_for_completion_interruptible(struct completion *x) 3531{ 3532 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE); 3533 if (t == -ERESTARTSYS) 3534 return t; 3535 return 0; 3536} 3537EXPORT_SYMBOL(wait_for_completion_interruptible); 3538 3539/** 3540 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr)) 3541 * @x: holds the state of this particular completion 3542 * @timeout: timeout value in jiffies 3543 * 3544 * This waits for either a completion of a specific task to be signaled or for a 3545 * specified timeout to expire. It is interruptible. The timeout is in jiffies. 3546 * 3547 * The return value is -ERESTARTSYS if interrupted, 0 if timed out, 3548 * positive (at least 1, or number of jiffies left till timeout) if completed. 3549 */ 3550long __sched 3551wait_for_completion_interruptible_timeout(struct completion *x, 3552 unsigned long timeout) 3553{ 3554 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE); 3555} 3556EXPORT_SYMBOL(wait_for_completion_interruptible_timeout); 3557 3558/** 3559 * wait_for_completion_killable: - waits for completion of a task (killable) 3560 * @x: holds the state of this particular completion 3561 * 3562 * This waits to be signaled for completion of a specific task. It can be 3563 * interrupted by a kill signal. 3564 * 3565 * The return value is -ERESTARTSYS if interrupted, 0 if completed. 3566 */ 3567int __sched wait_for_completion_killable(struct completion *x) 3568{ 3569 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE); 3570 if (t == -ERESTARTSYS) 3571 return t; 3572 return 0; 3573} 3574EXPORT_SYMBOL(wait_for_completion_killable); 3575 3576/** 3577 * wait_for_completion_killable_timeout: - waits for completion of a task (w/(to,killable)) 3578 * @x: holds the state of this particular completion 3579 * @timeout: timeout value in jiffies 3580 * 3581 * This waits for either a completion of a specific task to be 3582 * signaled or for a specified timeout to expire. It can be 3583 * interrupted by a kill signal. The timeout is in jiffies. 3584 * 3585 * The return value is -ERESTARTSYS if interrupted, 0 if timed out, 3586 * positive (at least 1, or number of jiffies left till timeout) if completed. 3587 */ 3588long __sched 3589wait_for_completion_killable_timeout(struct completion *x, 3590 unsigned long timeout) 3591{ 3592 return wait_for_common(x, timeout, TASK_KILLABLE); 3593} 3594EXPORT_SYMBOL(wait_for_completion_killable_timeout); 3595 3596/** 3597 * try_wait_for_completion - try to decrement a completion without blocking 3598 * @x: completion structure 3599 * 3600 * Returns: 0 if a decrement cannot be done without blocking 3601 * 1 if a decrement succeeded. 3602 * 3603 * If a completion is being used as a counting completion, 3604 * attempt to decrement the counter without blocking. This 3605 * enables us to avoid waiting if the resource the completion 3606 * is protecting is not available. 3607 */ 3608bool try_wait_for_completion(struct completion *x) 3609{ 3610 unsigned long flags; 3611 int ret = 1; 3612 3613 spin_lock_irqsave(&x->wait.lock, flags); 3614 if (!x->done) 3615 ret = 0; 3616 else 3617 x->done--; 3618 spin_unlock_irqrestore(&x->wait.lock, flags); 3619 return ret; 3620} 3621EXPORT_SYMBOL(try_wait_for_completion); 3622 3623/** 3624 * completion_done - Test to see if a completion has any waiters 3625 * @x: completion structure 3626 * 3627 * Returns: 0 if there are waiters (wait_for_completion() in progress) 3628 * 1 if there are no waiters. 3629 * 3630 */ 3631bool completion_done(struct completion *x) 3632{ 3633 unsigned long flags; 3634 int ret = 1; 3635 3636 spin_lock_irqsave(&x->wait.lock, flags); 3637 if (!x->done) 3638 ret = 0; 3639 spin_unlock_irqrestore(&x->wait.lock, flags); 3640 return ret; 3641} 3642EXPORT_SYMBOL(completion_done); 3643 3644static long __sched 3645sleep_on_common(wait_queue_head_t *q, int state, long timeout) 3646{ 3647 unsigned long flags; 3648 wait_queue_t wait; 3649 3650 init_waitqueue_entry(&wait, current); 3651 3652 __set_current_state(state); 3653 3654 spin_lock_irqsave(&q->lock, flags); 3655 __add_wait_queue(q, &wait); 3656 spin_unlock(&q->lock); 3657 timeout = schedule_timeout(timeout); 3658 spin_lock_irq(&q->lock); 3659 __remove_wait_queue(q, &wait); 3660 spin_unlock_irqrestore(&q->lock, flags); 3661 3662 return timeout; 3663} 3664 3665void __sched interruptible_sleep_on(wait_queue_head_t *q) 3666{ 3667 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT); 3668} 3669EXPORT_SYMBOL(interruptible_sleep_on); 3670 3671long __sched 3672interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout) 3673{ 3674 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout); 3675} 3676EXPORT_SYMBOL(interruptible_sleep_on_timeout); 3677 3678void __sched sleep_on(wait_queue_head_t *q) 3679{ 3680 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT); 3681} 3682EXPORT_SYMBOL(sleep_on); 3683 3684long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout) 3685{ 3686 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout); 3687} 3688EXPORT_SYMBOL(sleep_on_timeout); 3689 3690#ifdef CONFIG_RT_MUTEXES 3691 3692/* 3693 * rt_mutex_setprio - set the current priority of a task 3694 * @p: task 3695 * @prio: prio value (kernel-internal form) 3696 * 3697 * This function changes the 'effective' priority of a task. It does 3698 * not touch ->normal_prio like __setscheduler(). 3699 * 3700 * Used by the rt_mutex code to implement priority inheritance logic. 3701 */ 3702void rt_mutex_setprio(struct task_struct *p, int prio) 3703{ 3704 int oldprio, on_rq, running; 3705 struct rq *rq; 3706 const struct sched_class *prev_class; 3707 3708 BUG_ON(prio < 0 || prio > MAX_PRIO); 3709 3710 rq = __task_rq_lock(p); 3711 3712 trace_sched_pi_setprio(p, prio); 3713 oldprio = p->prio; 3714 prev_class = p->sched_class; 3715 on_rq = p->on_rq; 3716 running = task_current(rq, p); 3717 if (on_rq) 3718 dequeue_task(rq, p, 0); 3719 if (running) 3720 p->sched_class->put_prev_task(rq, p); 3721 3722 if (rt_prio(prio)) 3723 p->sched_class = &rt_sched_class; 3724 else 3725 p->sched_class = &fair_sched_class; 3726 3727 p->prio = prio; 3728 3729 if (running) 3730 p->sched_class->set_curr_task(rq); 3731 if (on_rq) 3732 enqueue_task(rq, p, oldprio < prio ? ENQUEUE_HEAD : 0); 3733 3734 check_class_changed(rq, p, prev_class, oldprio); 3735 __task_rq_unlock(rq); 3736} 3737 3738#endif 3739 3740void set_user_nice(struct task_struct *p, long nice) 3741{ 3742 int old_prio, delta, on_rq; 3743 unsigned long flags; 3744 struct rq *rq; 3745 3746 if (TASK_NICE(p) == nice || nice < -20 || nice > 19) 3747 return; 3748 /* 3749 * We have to be careful, if called from sys_setpriority(), 3750 * the task might be in the middle of scheduling on another CPU. 3751 */ 3752 rq = task_rq_lock(p, &flags); 3753 /* 3754 * The RT priorities are set via sched_setscheduler(), but we still 3755 * allow the 'normal' nice value to be set - but as expected 3756 * it wont have any effect on scheduling until the task is 3757 * SCHED_FIFO/SCHED_RR: 3758 */ 3759 if (task_has_rt_policy(p)) { 3760 p->static_prio = NICE_TO_PRIO(nice); 3761 goto out_unlock; 3762 } 3763 on_rq = p->on_rq; 3764 if (on_rq) 3765 dequeue_task(rq, p, 0); 3766 3767 p->static_prio = NICE_TO_PRIO(nice); 3768 set_load_weight(p); 3769 old_prio = p->prio; 3770 p->prio = effective_prio(p); 3771 delta = p->prio - old_prio; 3772 3773 if (on_rq) { 3774 enqueue_task(rq, p, 0); 3775 /* 3776 * If the task increased its priority or is running and 3777 * lowered its priority, then reschedule its CPU: 3778 */ 3779 if (delta < 0 || (delta > 0 && task_running(rq, p))) 3780 resched_task(rq->curr); 3781 } 3782out_unlock: 3783 task_rq_unlock(rq, p, &flags); 3784} 3785EXPORT_SYMBOL(set_user_nice); 3786 3787/* 3788 * can_nice - check if a task can reduce its nice value 3789 * @p: task 3790 * @nice: nice value 3791 */ 3792int can_nice(const struct task_struct *p, const int nice) 3793{ 3794 /* convert nice value [19,-20] to rlimit style value [1,40] */ 3795 int nice_rlim = 20 - nice; 3796 3797 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) || 3798 capable(CAP_SYS_NICE)); 3799} 3800 3801#ifdef __ARCH_WANT_SYS_NICE 3802 3803/* 3804 * sys_nice - change the priority of the current process. 3805 * @increment: priority increment 3806 * 3807 * sys_setpriority is a more generic, but much slower function that 3808 * does similar things. 3809 */ 3810SYSCALL_DEFINE1(nice, int, increment) 3811{ 3812 long nice, retval; 3813 3814 /* 3815 * Setpriority might change our priority at the same moment. 3816 * We don't have to worry. Conceptually one call occurs first 3817 * and we have a single winner. 3818 */ 3819 if (increment < -40) 3820 increment = -40; 3821 if (increment > 40) 3822 increment = 40; 3823 3824 nice = TASK_NICE(current) + increment; 3825 if (nice < -20) 3826 nice = -20; 3827 if (nice > 19) 3828 nice = 19; 3829 3830 if (increment < 0 && !can_nice(current, nice)) 3831 return -EPERM; 3832 3833 retval = security_task_setnice(current, nice); 3834 if (retval) 3835 return retval; 3836 3837 set_user_nice(current, nice); 3838 return 0; 3839} 3840 3841#endif 3842 3843/** 3844 * task_prio - return the priority value of a given task. 3845 * @p: the task in question. 3846 * 3847 * This is the priority value as seen by users in /proc. 3848 * RT tasks are offset by -200. Normal tasks are centered 3849 * around 0, value goes from -16 to +15. 3850 */ 3851int task_prio(const struct task_struct *p) 3852{ 3853 return p->prio - MAX_RT_PRIO; 3854} 3855 3856/** 3857 * task_nice - return the nice value of a given task. 3858 * @p: the task in question. 3859 */ 3860int task_nice(const struct task_struct *p) 3861{ 3862 return TASK_NICE(p); 3863} 3864EXPORT_SYMBOL(task_nice); 3865 3866/** 3867 * idle_cpu - is a given cpu idle currently? 3868 * @cpu: the processor in question. 3869 */ 3870int idle_cpu(int cpu) 3871{ 3872 struct rq *rq = cpu_rq(cpu); 3873 3874 if (rq->curr != rq->idle) 3875 return 0; 3876 3877 if (rq->nr_running) 3878 return 0; 3879 3880#ifdef CONFIG_SMP 3881 if (!llist_empty(&rq->wake_list)) 3882 return 0; 3883#endif 3884 3885 return 1; 3886} 3887 3888/** 3889 * idle_task - return the idle task for a given cpu. 3890 * @cpu: the processor in question. 3891 */ 3892struct task_struct *idle_task(int cpu) 3893{ 3894 return cpu_rq(cpu)->idle; 3895} 3896 3897/** 3898 * find_process_by_pid - find a process with a matching PID value. 3899 * @pid: the pid in question. 3900 */ 3901static struct task_struct *find_process_by_pid(pid_t pid) 3902{ 3903 return pid ? find_task_by_vpid(pid) : current; 3904} 3905 3906/* Actually do priority change: must hold rq lock. */ 3907static void 3908__setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio) 3909{ 3910 p->policy = policy; 3911 p->rt_priority = prio; 3912 p->normal_prio = normal_prio(p); 3913 /* we are holding p->pi_lock already */ 3914 p->prio = rt_mutex_getprio(p); 3915 if (rt_prio(p->prio)) 3916 p->sched_class = &rt_sched_class; 3917 else 3918 p->sched_class = &fair_sched_class; 3919 set_load_weight(p); 3920} 3921 3922/* 3923 * check the target process has a UID that matches the current process's 3924 */ 3925static bool check_same_owner(struct task_struct *p) 3926{ 3927 const struct cred *cred = current_cred(), *pcred; 3928 bool match; 3929 3930 rcu_read_lock(); 3931 pcred = __task_cred(p); 3932 if (cred->user->user_ns == pcred->user->user_ns) 3933 match = (cred->euid == pcred->euid || 3934 cred->euid == pcred->uid); 3935 else 3936 match = false; 3937 rcu_read_unlock(); 3938 return match; 3939} 3940 3941static int __sched_setscheduler(struct task_struct *p, int policy, 3942 const struct sched_param *param, bool user) 3943{ 3944 int retval, oldprio, oldpolicy = -1, on_rq, running; 3945 unsigned long flags; 3946 const struct sched_class *prev_class; 3947 struct rq *rq; 3948 int reset_on_fork; 3949 3950 /* may grab non-irq protected spin_locks */ 3951 BUG_ON(in_interrupt()); 3952recheck: 3953 /* double check policy once rq lock held */ 3954 if (policy < 0) { 3955 reset_on_fork = p->sched_reset_on_fork; 3956 policy = oldpolicy = p->policy; 3957 } else { 3958 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK); 3959 policy &= ~SCHED_RESET_ON_FORK; 3960 3961 if (policy != SCHED_FIFO && policy != SCHED_RR && 3962 policy != SCHED_NORMAL && policy != SCHED_BATCH && 3963 policy != SCHED_IDLE) 3964 return -EINVAL; 3965 } 3966 3967 /* 3968 * Valid priorities for SCHED_FIFO and SCHED_RR are 3969 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL, 3970 * SCHED_BATCH and SCHED_IDLE is 0. 3971 */ 3972 if (param->sched_priority < 0 || 3973 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) || 3974 (!p->mm && param->sched_priority > MAX_RT_PRIO-1)) 3975 return -EINVAL; 3976 if (rt_policy(policy) != (param->sched_priority != 0)) 3977 return -EINVAL; 3978 3979 /* 3980 * Allow unprivileged RT tasks to decrease priority: 3981 */ 3982 if (user && !capable(CAP_SYS_NICE)) { 3983 if (rt_policy(policy)) { 3984 unsigned long rlim_rtprio = 3985 task_rlimit(p, RLIMIT_RTPRIO); 3986 3987 /* can't set/change the rt policy */ 3988 if (policy != p->policy && !rlim_rtprio) 3989 return -EPERM; 3990 3991 /* can't increase priority */ 3992 if (param->sched_priority > p->rt_priority && 3993 param->sched_priority > rlim_rtprio) 3994 return -EPERM; 3995 } 3996 3997 /* 3998 * Treat SCHED_IDLE as nice 20. Only allow a switch to 3999 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it. 4000 */ 4001 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) { 4002 if (!can_nice(p, TASK_NICE(p))) 4003 return -EPERM; 4004 } 4005 4006 /* can't change other user's priorities */ 4007 if (!check_same_owner(p)) 4008 return -EPERM; 4009 4010 /* Normal users shall not reset the sched_reset_on_fork flag */ 4011 if (p->sched_reset_on_fork && !reset_on_fork) 4012 return -EPERM; 4013 } 4014 4015 if (user) { 4016 retval = security_task_setscheduler(p); 4017 if (retval) 4018 return retval; 4019 } 4020 4021 /* 4022 * make sure no PI-waiters arrive (or leave) while we are 4023 * changing the priority of the task: 4024 * 4025 * To be able to change p->policy safely, the appropriate 4026 * runqueue lock must be held. 4027 */ 4028 rq = task_rq_lock(p, &flags); 4029 4030 /* 4031 * Changing the policy of the stop threads its a very bad idea 4032 */ 4033 if (p == rq->stop) { 4034 task_rq_unlock(rq, p, &flags); 4035 return -EINVAL; 4036 } 4037 4038 /* 4039 * If not changing anything there's no need to proceed further: 4040 */ 4041 if (unlikely(policy == p->policy && (!rt_policy(policy) || 4042 param->sched_priority == p->rt_priority))) { 4043 4044 __task_rq_unlock(rq); 4045 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 4046 return 0; 4047 } 4048 4049#ifdef CONFIG_RT_GROUP_SCHED 4050 if (user) { 4051 /* 4052 * Do not allow realtime tasks into groups that have no runtime 4053 * assigned. 4054 */ 4055 if (rt_bandwidth_enabled() && rt_policy(policy) && 4056 task_group(p)->rt_bandwidth.rt_runtime == 0 && 4057 !task_group_is_autogroup(task_group(p))) { 4058 task_rq_unlock(rq, p, &flags); 4059 return -EPERM; 4060 } 4061 } 4062#endif 4063 4064 /* recheck policy now with rq lock held */ 4065 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) { 4066 policy = oldpolicy = -1; 4067 task_rq_unlock(rq, p, &flags); 4068 goto recheck; 4069 } 4070 on_rq = p->on_rq; 4071 running = task_current(rq, p); 4072 if (on_rq) 4073 deactivate_task(rq, p, 0); 4074 if (running) 4075 p->sched_class->put_prev_task(rq, p); 4076 4077 p->sched_reset_on_fork = reset_on_fork; 4078 4079 oldprio = p->prio; 4080 prev_class = p->sched_class; 4081 __setscheduler(rq, p, policy, param->sched_priority); 4082 4083 if (running) 4084 p->sched_class->set_curr_task(rq); 4085 if (on_rq) 4086 activate_task(rq, p, 0); 4087 4088 check_class_changed(rq, p, prev_class, oldprio); 4089 task_rq_unlock(rq, p, &flags); 4090 4091 rt_mutex_adjust_pi(p); 4092 4093 return 0; 4094} 4095 4096/** 4097 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread. 4098 * @p: the task in question. 4099 * @policy: new policy. 4100 * @param: structure containing the new RT priority. 4101 * 4102 * NOTE that the task may be already dead. 4103 */ 4104int sched_setscheduler(struct task_struct *p, int policy, 4105 const struct sched_param *param) 4106{ 4107 return __sched_setscheduler(p, policy, param, true); 4108} 4109EXPORT_SYMBOL_GPL(sched_setscheduler); 4110 4111/** 4112 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace. 4113 * @p: the task in question. 4114 * @policy: new policy. 4115 * @param: structure containing the new RT priority. 4116 * 4117 * Just like sched_setscheduler, only don't bother checking if the 4118 * current context has permission. For example, this is needed in 4119 * stop_machine(): we create temporary high priority worker threads, 4120 * but our caller might not have that capability. 4121 */ 4122int sched_setscheduler_nocheck(struct task_struct *p, int policy, 4123 const struct sched_param *param) 4124{ 4125 return __sched_setscheduler(p, policy, param, false); 4126} 4127 4128static int 4129do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param) 4130{ 4131 struct sched_param lparam; 4132 struct task_struct *p; 4133 int retval; 4134 4135 if (!param || pid < 0) 4136 return -EINVAL; 4137 if (copy_from_user(&lparam, param, sizeof(struct sched_param))) 4138 return -EFAULT; 4139 4140 rcu_read_lock(); 4141 retval = -ESRCH; 4142 p = find_process_by_pid(pid); 4143 if (p != NULL) 4144 retval = sched_setscheduler(p, policy, &lparam); 4145 rcu_read_unlock(); 4146 4147 return retval; 4148} 4149 4150/** 4151 * sys_sched_setscheduler - set/change the scheduler policy and RT priority 4152 * @pid: the pid in question. 4153 * @policy: new policy. 4154 * @param: structure containing the new RT priority. 4155 */ 4156SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy, 4157 struct sched_param __user *, param) 4158{ 4159 /* negative values for policy are not valid */ 4160 if (policy < 0) 4161 return -EINVAL; 4162 4163 return do_sched_setscheduler(pid, policy, param); 4164} 4165 4166/** 4167 * sys_sched_setparam - set/change the RT priority of a thread 4168 * @pid: the pid in question. 4169 * @param: structure containing the new RT priority. 4170 */ 4171SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param) 4172{ 4173 return do_sched_setscheduler(pid, -1, param); 4174} 4175 4176/** 4177 * sys_sched_getscheduler - get the policy (scheduling class) of a thread 4178 * @pid: the pid in question. 4179 */ 4180SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid) 4181{ 4182 struct task_struct *p; 4183 int retval; 4184 4185 if (pid < 0) 4186 return -EINVAL; 4187 4188 retval = -ESRCH; 4189 rcu_read_lock(); 4190 p = find_process_by_pid(pid); 4191 if (p) { 4192 retval = security_task_getscheduler(p); 4193 if (!retval) 4194 retval = p->policy 4195 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0); 4196 } 4197 rcu_read_unlock(); 4198 return retval; 4199} 4200 4201/** 4202 * sys_sched_getparam - get the RT priority of a thread 4203 * @pid: the pid in question. 4204 * @param: structure containing the RT priority. 4205 */ 4206SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param) 4207{ 4208 struct sched_param lp; 4209 struct task_struct *p; 4210 int retval; 4211 4212 if (!param || pid < 0) 4213 return -EINVAL; 4214 4215 rcu_read_lock(); 4216 p = find_process_by_pid(pid); 4217 retval = -ESRCH; 4218 if (!p) 4219 goto out_unlock; 4220 4221 retval = security_task_getscheduler(p); 4222 if (retval) 4223 goto out_unlock; 4224 4225 lp.sched_priority = p->rt_priority; 4226 rcu_read_unlock(); 4227 4228 /* 4229 * This one might sleep, we cannot do it with a spinlock held ... 4230 */ 4231 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0; 4232 4233 return retval; 4234 4235out_unlock: 4236 rcu_read_unlock(); 4237 return retval; 4238} 4239 4240long sched_setaffinity(pid_t pid, const struct cpumask *in_mask) 4241{ 4242 cpumask_var_t cpus_allowed, new_mask; 4243 struct task_struct *p; 4244 int retval; 4245 4246 get_online_cpus(); 4247 rcu_read_lock(); 4248 4249 p = find_process_by_pid(pid); 4250 if (!p) { 4251 rcu_read_unlock(); 4252 put_online_cpus(); 4253 return -ESRCH; 4254 } 4255 4256 /* Prevent p going away */ 4257 get_task_struct(p); 4258 rcu_read_unlock(); 4259 4260 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) { 4261 retval = -ENOMEM; 4262 goto out_put_task; 4263 } 4264 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) { 4265 retval = -ENOMEM; 4266 goto out_free_cpus_allowed; 4267 } 4268 retval = -EPERM; 4269 if (!check_same_owner(p) && !task_ns_capable(p, CAP_SYS_NICE)) 4270 goto out_unlock; 4271 4272 retval = security_task_setscheduler(p); 4273 if (retval) 4274 goto out_unlock; 4275 4276 cpuset_cpus_allowed(p, cpus_allowed); 4277 cpumask_and(new_mask, in_mask, cpus_allowed); 4278again: 4279 retval = set_cpus_allowed_ptr(p, new_mask); 4280 4281 if (!retval) { 4282 cpuset_cpus_allowed(p, cpus_allowed); 4283 if (!cpumask_subset(new_mask, cpus_allowed)) { 4284 /* 4285 * We must have raced with a concurrent cpuset 4286 * update. Just reset the cpus_allowed to the 4287 * cpuset's cpus_allowed 4288 */ 4289 cpumask_copy(new_mask, cpus_allowed); 4290 goto again; 4291 } 4292 } 4293out_unlock: 4294 free_cpumask_var(new_mask); 4295out_free_cpus_allowed: 4296 free_cpumask_var(cpus_allowed); 4297out_put_task: 4298 put_task_struct(p); 4299 put_online_cpus(); 4300 return retval; 4301} 4302 4303static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len, 4304 struct cpumask *new_mask) 4305{ 4306 if (len < cpumask_size()) 4307 cpumask_clear(new_mask); 4308 else if (len > cpumask_size()) 4309 len = cpumask_size(); 4310 4311 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0; 4312} 4313 4314/** 4315 * sys_sched_setaffinity - set the cpu affinity of a process 4316 * @pid: pid of the process 4317 * @len: length in bytes of the bitmask pointed to by user_mask_ptr 4318 * @user_mask_ptr: user-space pointer to the new cpu mask 4319 */ 4320SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len, 4321 unsigned long __user *, user_mask_ptr) 4322{ 4323 cpumask_var_t new_mask; 4324 int retval; 4325 4326 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) 4327 return -ENOMEM; 4328 4329 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask); 4330 if (retval == 0) 4331 retval = sched_setaffinity(pid, new_mask); 4332 free_cpumask_var(new_mask); 4333 return retval; 4334} 4335 4336long sched_getaffinity(pid_t pid, struct cpumask *mask) 4337{ 4338 struct task_struct *p; 4339 unsigned long flags; 4340 int retval; 4341 4342 get_online_cpus(); 4343 rcu_read_lock(); 4344 4345 retval = -ESRCH; 4346 p = find_process_by_pid(pid); 4347 if (!p) 4348 goto out_unlock; 4349 4350 retval = security_task_getscheduler(p); 4351 if (retval) 4352 goto out_unlock; 4353 4354 raw_spin_lock_irqsave(&p->pi_lock, flags); 4355 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask); 4356 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 4357 4358out_unlock: 4359 rcu_read_unlock(); 4360 put_online_cpus(); 4361 4362 return retval; 4363} 4364 4365/** 4366 * sys_sched_getaffinity - get the cpu affinity of a process 4367 * @pid: pid of the process 4368 * @len: length in bytes of the bitmask pointed to by user_mask_ptr 4369 * @user_mask_ptr: user-space pointer to hold the current cpu mask 4370 */ 4371SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len, 4372 unsigned long __user *, user_mask_ptr) 4373{ 4374 int ret; 4375 cpumask_var_t mask; 4376 4377 if ((len * BITS_PER_BYTE) < nr_cpu_ids) 4378 return -EINVAL; 4379 if (len & (sizeof(unsigned long)-1)) 4380 return -EINVAL; 4381 4382 if (!alloc_cpumask_var(&mask, GFP_KERNEL)) 4383 return -ENOMEM; 4384 4385 ret = sched_getaffinity(pid, mask); 4386 if (ret == 0) { 4387 size_t retlen = min_t(size_t, len, cpumask_size()); 4388 4389 if (copy_to_user(user_mask_ptr, mask, retlen)) 4390 ret = -EFAULT; 4391 else 4392 ret = retlen; 4393 } 4394 free_cpumask_var(mask); 4395 4396 return ret; 4397} 4398 4399/** 4400 * sys_sched_yield - yield the current processor to other threads. 4401 * 4402 * This function yields the current CPU to other tasks. If there are no 4403 * other threads running on this CPU then this function will return. 4404 */ 4405SYSCALL_DEFINE0(sched_yield) 4406{ 4407 struct rq *rq = this_rq_lock(); 4408 4409 schedstat_inc(rq, yld_count); 4410 current->sched_class->yield_task(rq); 4411 4412 /* 4413 * Since we are going to call schedule() anyway, there's 4414 * no need to preempt or enable interrupts: 4415 */ 4416 __release(rq->lock); 4417 spin_release(&rq->lock.dep_map, 1, _THIS_IP_); 4418 do_raw_spin_unlock(&rq->lock); 4419 preempt_enable_no_resched(); 4420 4421 schedule(); 4422 4423 return 0; 4424} 4425 4426static inline int should_resched(void) 4427{ 4428 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE); 4429} 4430 4431static void __cond_resched(void) 4432{ 4433 add_preempt_count(PREEMPT_ACTIVE); 4434 __schedule(); 4435 sub_preempt_count(PREEMPT_ACTIVE); 4436} 4437 4438int __sched _cond_resched(void) 4439{ 4440 if (should_resched()) { 4441 __cond_resched(); 4442 return 1; 4443 } 4444 return 0; 4445} 4446EXPORT_SYMBOL(_cond_resched); 4447 4448/* 4449 * __cond_resched_lock() - if a reschedule is pending, drop the given lock, 4450 * call schedule, and on return reacquire the lock. 4451 * 4452 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level 4453 * operations here to prevent schedule() from being called twice (once via 4454 * spin_unlock(), once by hand). 4455 */ 4456int __cond_resched_lock(spinlock_t *lock) 4457{ 4458 int resched = should_resched(); 4459 int ret = 0; 4460 4461 lockdep_assert_held(lock); 4462 4463 if (spin_needbreak(lock) || resched) { 4464 spin_unlock(lock); 4465 if (resched) 4466 __cond_resched(); 4467 else 4468 cpu_relax(); 4469 ret = 1; 4470 spin_lock(lock); 4471 } 4472 return ret; 4473} 4474EXPORT_SYMBOL(__cond_resched_lock); 4475 4476int __sched __cond_resched_softirq(void) 4477{ 4478 BUG_ON(!in_softirq()); 4479 4480 if (should_resched()) { 4481 local_bh_enable(); 4482 __cond_resched(); 4483 local_bh_disable(); 4484 return 1; 4485 } 4486 return 0; 4487} 4488EXPORT_SYMBOL(__cond_resched_softirq); 4489 4490/** 4491 * yield - yield the current processor to other threads. 4492 * 4493 * This is a shortcut for kernel-space yielding - it marks the 4494 * thread runnable and calls sys_sched_yield(). 4495 */ 4496void __sched yield(void) 4497{ 4498 set_current_state(TASK_RUNNING); 4499 sys_sched_yield(); 4500} 4501EXPORT_SYMBOL(yield); 4502 4503/** 4504 * yield_to - yield the current processor to another thread in 4505 * your thread group, or accelerate that thread toward the 4506 * processor it's on. 4507 * @p: target task 4508 * @preempt: whether task preemption is allowed or not 4509 * 4510 * It's the caller's job to ensure that the target task struct 4511 * can't go away on us before we can do any checks. 4512 * 4513 * Returns true if we indeed boosted the target task. 4514 */ 4515bool __sched yield_to(struct task_struct *p, bool preempt) 4516{ 4517 struct task_struct *curr = current; 4518 struct rq *rq, *p_rq; 4519 unsigned long flags; 4520 bool yielded = 0; 4521 4522 local_irq_save(flags); 4523 rq = this_rq(); 4524 4525again: 4526 p_rq = task_rq(p); 4527 double_rq_lock(rq, p_rq); 4528 while (task_rq(p) != p_rq) { 4529 double_rq_unlock(rq, p_rq); 4530 goto again; 4531 } 4532 4533 if (!curr->sched_class->yield_to_task) 4534 goto out; 4535 4536 if (curr->sched_class != p->sched_class) 4537 goto out; 4538 4539 if (task_running(p_rq, p) || p->state) 4540 goto out; 4541 4542 yielded = curr->sched_class->yield_to_task(rq, p, preempt); 4543 if (yielded) { 4544 schedstat_inc(rq, yld_count); 4545 /* 4546 * Make p's CPU reschedule; pick_next_entity takes care of 4547 * fairness. 4548 */ 4549 if (preempt && rq != p_rq) 4550 resched_task(p_rq->curr); 4551 } else { 4552 /* 4553 * We might have set it in task_yield_fair(), but are 4554 * not going to schedule(), so don't want to skip 4555 * the next update. 4556 */ 4557 rq->skip_clock_update = 0; 4558 } 4559 4560out: 4561 double_rq_unlock(rq, p_rq); 4562 local_irq_restore(flags); 4563 4564 if (yielded) 4565 schedule(); 4566 4567 return yielded; 4568} 4569EXPORT_SYMBOL_GPL(yield_to); 4570 4571/* 4572 * This task is about to go to sleep on IO. Increment rq->nr_iowait so 4573 * that process accounting knows that this is a task in IO wait state. 4574 */ 4575void __sched io_schedule(void) 4576{ 4577 struct rq *rq = raw_rq(); 4578 4579 delayacct_blkio_start(); 4580 atomic_inc(&rq->nr_iowait); 4581 blk_flush_plug(current); 4582 current->in_iowait = 1; 4583 schedule(); 4584 current->in_iowait = 0; 4585 atomic_dec(&rq->nr_iowait); 4586 delayacct_blkio_end(); 4587} 4588EXPORT_SYMBOL(io_schedule); 4589 4590long __sched io_schedule_timeout(long timeout) 4591{ 4592 struct rq *rq = raw_rq(); 4593 long ret; 4594 4595 delayacct_blkio_start(); 4596 atomic_inc(&rq->nr_iowait); 4597 blk_flush_plug(current); 4598 current->in_iowait = 1; 4599 ret = schedule_timeout(timeout); 4600 current->in_iowait = 0; 4601 atomic_dec(&rq->nr_iowait); 4602 delayacct_blkio_end(); 4603 return ret; 4604} 4605 4606/** 4607 * sys_sched_get_priority_max - return maximum RT priority. 4608 * @policy: scheduling class. 4609 * 4610 * this syscall returns the maximum rt_priority that can be used 4611 * by a given scheduling class. 4612 */ 4613SYSCALL_DEFINE1(sched_get_priority_max, int, policy) 4614{ 4615 int ret = -EINVAL; 4616 4617 switch (policy) { 4618 case SCHED_FIFO: 4619 case SCHED_RR: 4620 ret = MAX_USER_RT_PRIO-1; 4621 break; 4622 case SCHED_NORMAL: 4623 case SCHED_BATCH: 4624 case SCHED_IDLE: 4625 ret = 0; 4626 break; 4627 } 4628 return ret; 4629} 4630 4631/** 4632 * sys_sched_get_priority_min - return minimum RT priority. 4633 * @policy: scheduling class. 4634 * 4635 * this syscall returns the minimum rt_priority that can be used 4636 * by a given scheduling class. 4637 */ 4638SYSCALL_DEFINE1(sched_get_priority_min, int, policy) 4639{ 4640 int ret = -EINVAL; 4641 4642 switch (policy) { 4643 case SCHED_FIFO: 4644 case SCHED_RR: 4645 ret = 1; 4646 break; 4647 case SCHED_NORMAL: 4648 case SCHED_BATCH: 4649 case SCHED_IDLE: 4650 ret = 0; 4651 } 4652 return ret; 4653} 4654 4655/** 4656 * sys_sched_rr_get_interval - return the default timeslice of a process. 4657 * @pid: pid of the process. 4658 * @interval: userspace pointer to the timeslice value. 4659 * 4660 * this syscall writes the default timeslice value of a given process 4661 * into the user-space timespec buffer. A value of '0' means infinity. 4662 */ 4663SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid, 4664 struct timespec __user *, interval) 4665{ 4666 struct task_struct *p; 4667 unsigned int time_slice; 4668 unsigned long flags; 4669 struct rq *rq; 4670 int retval; 4671 struct timespec t; 4672 4673 if (pid < 0) 4674 return -EINVAL; 4675 4676 retval = -ESRCH; 4677 rcu_read_lock(); 4678 p = find_process_by_pid(pid); 4679 if (!p) 4680 goto out_unlock; 4681 4682 retval = security_task_getscheduler(p); 4683 if (retval) 4684 goto out_unlock; 4685 4686 rq = task_rq_lock(p, &flags); 4687 time_slice = p->sched_class->get_rr_interval(rq, p); 4688 task_rq_unlock(rq, p, &flags); 4689 4690 rcu_read_unlock(); 4691 jiffies_to_timespec(time_slice, &t); 4692 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0; 4693 return retval; 4694 4695out_unlock: 4696 rcu_read_unlock(); 4697 return retval; 4698} 4699 4700static const char stat_nam[] = TASK_STATE_TO_CHAR_STR; 4701 4702void sched_show_task(struct task_struct *p) 4703{ 4704 unsigned long free = 0; 4705 unsigned state; 4706 4707 state = p->state ? __ffs(p->state) + 1 : 0; 4708 printk(KERN_INFO "%-15.15s %c", p->comm, 4709 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?'); 4710#if BITS_PER_LONG == 32 4711 if (state == TASK_RUNNING) 4712 printk(KERN_CONT " running "); 4713 else 4714 printk(KERN_CONT " %08lx ", thread_saved_pc(p)); 4715#else 4716 if (state == TASK_RUNNING) 4717 printk(KERN_CONT " running task "); 4718 else 4719 printk(KERN_CONT " %016lx ", thread_saved_pc(p)); 4720#endif 4721#ifdef CONFIG_DEBUG_STACK_USAGE 4722 free = stack_not_used(p); 4723#endif 4724 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free, 4725 task_pid_nr(p), task_pid_nr(p->real_parent), 4726 (unsigned long)task_thread_info(p)->flags); 4727 4728 show_stack(p, NULL); 4729} 4730 4731void show_state_filter(unsigned long state_filter) 4732{ 4733 struct task_struct *g, *p; 4734 4735#if BITS_PER_LONG == 32 4736 printk(KERN_INFO 4737 " task PC stack pid father\n"); 4738#else 4739 printk(KERN_INFO 4740 " task PC stack pid father\n"); 4741#endif 4742 rcu_read_lock(); 4743 do_each_thread(g, p) { 4744 /* 4745 * reset the NMI-timeout, listing all files on a slow 4746 * console might take a lot of time: 4747 */ 4748 touch_nmi_watchdog(); 4749 if (!state_filter || (p->state & state_filter)) 4750 sched_show_task(p); 4751 } while_each_thread(g, p); 4752 4753 touch_all_softlockup_watchdogs(); 4754 4755#ifdef CONFIG_SCHED_DEBUG 4756 sysrq_sched_debug_show(); 4757#endif 4758 rcu_read_unlock(); 4759 /* 4760 * Only show locks if all tasks are dumped: 4761 */ 4762 if (!state_filter) 4763 debug_show_all_locks(); 4764} 4765 4766void __cpuinit init_idle_bootup_task(struct task_struct *idle) 4767{ 4768 idle->sched_class = &idle_sched_class; 4769} 4770 4771/** 4772 * init_idle - set up an idle thread for a given CPU 4773 * @idle: task in question 4774 * @cpu: cpu the idle task belongs to 4775 * 4776 * NOTE: this function does not set the idle thread's NEED_RESCHED 4777 * flag, to make booting more robust. 4778 */ 4779void __cpuinit init_idle(struct task_struct *idle, int cpu) 4780{ 4781 struct rq *rq = cpu_rq(cpu); 4782 unsigned long flags; 4783 4784 raw_spin_lock_irqsave(&rq->lock, flags); 4785 4786 __sched_fork(idle); 4787 idle->state = TASK_RUNNING; 4788 idle->se.exec_start = sched_clock(); 4789 4790 do_set_cpus_allowed(idle, cpumask_of(cpu)); 4791 /* 4792 * We're having a chicken and egg problem, even though we are 4793 * holding rq->lock, the cpu isn't yet set to this cpu so the 4794 * lockdep check in task_group() will fail. 4795 * 4796 * Similar case to sched_fork(). / Alternatively we could 4797 * use task_rq_lock() here and obtain the other rq->lock. 4798 * 4799 * Silence PROVE_RCU 4800 */ 4801 rcu_read_lock(); 4802 __set_task_cpu(idle, cpu); 4803 rcu_read_unlock(); 4804 4805 rq->curr = rq->idle = idle; 4806#if defined(CONFIG_SMP) 4807 idle->on_cpu = 1; 4808#endif 4809 raw_spin_unlock_irqrestore(&rq->lock, flags); 4810 4811 /* Set the preempt count _outside_ the spinlocks! */ 4812 task_thread_info(idle)->preempt_count = 0; 4813 4814 /* 4815 * The idle tasks have their own, simple scheduling class: 4816 */ 4817 idle->sched_class = &idle_sched_class; 4818 ftrace_graph_init_idle_task(idle, cpu); 4819#if defined(CONFIG_SMP) 4820 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu); 4821#endif 4822} 4823 4824#ifdef CONFIG_SMP 4825void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask) 4826{ 4827 if (p->sched_class && p->sched_class->set_cpus_allowed) 4828 p->sched_class->set_cpus_allowed(p, new_mask); 4829 4830 cpumask_copy(&p->cpus_allowed, new_mask); 4831 p->rt.nr_cpus_allowed = cpumask_weight(new_mask); 4832} 4833 4834/* 4835 * This is how migration works: 4836 * 4837 * 1) we invoke migration_cpu_stop() on the target CPU using 4838 * stop_one_cpu(). 4839 * 2) stopper starts to run (implicitly forcing the migrated thread 4840 * off the CPU) 4841 * 3) it checks whether the migrated task is still in the wrong runqueue. 4842 * 4) if it's in the wrong runqueue then the migration thread removes 4843 * it and puts it into the right queue. 4844 * 5) stopper completes and stop_one_cpu() returns and the migration 4845 * is done. 4846 */ 4847 4848/* 4849 * Change a given task's CPU affinity. Migrate the thread to a 4850 * proper CPU and schedule it away if the CPU it's executing on 4851 * is removed from the allowed bitmask. 4852 * 4853 * NOTE: the caller must have a valid reference to the task, the 4854 * task must not exit() & deallocate itself prematurely. The 4855 * call is not atomic; no spinlocks may be held. 4856 */ 4857int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask) 4858{ 4859 unsigned long flags; 4860 struct rq *rq; 4861 unsigned int dest_cpu; 4862 int ret = 0; 4863 4864 rq = task_rq_lock(p, &flags); 4865 4866 if (cpumask_equal(&p->cpus_allowed, new_mask)) 4867 goto out; 4868 4869 if (!cpumask_intersects(new_mask, cpu_active_mask)) { 4870 ret = -EINVAL; 4871 goto out; 4872 } 4873 4874 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current)) { 4875 ret = -EINVAL; 4876 goto out; 4877 } 4878 4879 do_set_cpus_allowed(p, new_mask); 4880 4881 /* Can the task run on the task's current CPU? If so, we're done */ 4882 if (cpumask_test_cpu(task_cpu(p), new_mask)) 4883 goto out; 4884 4885 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask); 4886 if (p->on_rq) { 4887 struct migration_arg arg = { p, dest_cpu }; 4888 /* Need help from migration thread: drop lock and wait. */ 4889 task_rq_unlock(rq, p, &flags); 4890 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg); 4891 tlb_migrate_finish(p->mm); 4892 return 0; 4893 } 4894out: 4895 task_rq_unlock(rq, p, &flags); 4896 4897 return ret; 4898} 4899EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr); 4900 4901/* 4902 * Move (not current) task off this cpu, onto dest cpu. We're doing 4903 * this because either it can't run here any more (set_cpus_allowed() 4904 * away from this CPU, or CPU going down), or because we're 4905 * attempting to rebalance this task on exec (sched_exec). 4906 * 4907 * So we race with normal scheduler movements, but that's OK, as long 4908 * as the task is no longer on this CPU. 4909 * 4910 * Returns non-zero if task was successfully migrated. 4911 */ 4912static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu) 4913{ 4914 struct rq *rq_dest, *rq_src; 4915 int ret = 0; 4916 4917 if (unlikely(!cpu_active(dest_cpu))) 4918 return ret; 4919 4920 rq_src = cpu_rq(src_cpu); 4921 rq_dest = cpu_rq(dest_cpu); 4922 4923 raw_spin_lock(&p->pi_lock); 4924 double_rq_lock(rq_src, rq_dest); 4925 /* Already moved. */ 4926 if (task_cpu(p) != src_cpu) 4927 goto done; 4928 /* Affinity changed (again). */ 4929 if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p))) 4930 goto fail; 4931 4932 /* 4933 * If we're not on a rq, the next wake-up will ensure we're 4934 * placed properly. 4935 */ 4936 if (p->on_rq) { 4937 deactivate_task(rq_src, p, 0); 4938 set_task_cpu(p, dest_cpu); 4939 activate_task(rq_dest, p, 0); 4940 check_preempt_curr(rq_dest, p, 0); 4941 } 4942done: 4943 ret = 1; 4944fail: 4945 double_rq_unlock(rq_src, rq_dest); 4946 raw_spin_unlock(&p->pi_lock); 4947 return ret; 4948} 4949 4950/* 4951 * migration_cpu_stop - this will be executed by a highprio stopper thread 4952 * and performs thread migration by bumping thread off CPU then 4953 * 'pushing' onto another runqueue. 4954 */ 4955static int migration_cpu_stop(void *data) 4956{ 4957 struct migration_arg *arg = data; 4958 4959 /* 4960 * The original target cpu might have gone down and we might 4961 * be on another cpu but it doesn't matter. 4962 */ 4963 local_irq_disable(); 4964 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu); 4965 local_irq_enable(); 4966 return 0; 4967} 4968 4969#ifdef CONFIG_HOTPLUG_CPU 4970 4971/* 4972 * Ensures that the idle task is using init_mm right before its cpu goes 4973 * offline. 4974 */ 4975void idle_task_exit(void) 4976{ 4977 struct mm_struct *mm = current->active_mm; 4978 4979 BUG_ON(cpu_online(smp_processor_id())); 4980 4981 if (mm != &init_mm) 4982 switch_mm(mm, &init_mm, current); 4983 mmdrop(mm); 4984} 4985 4986/* 4987 * While a dead CPU has no uninterruptible tasks queued at this point, 4988 * it might still have a nonzero ->nr_uninterruptible counter, because 4989 * for performance reasons the counter is not stricly tracking tasks to 4990 * their home CPUs. So we just add the counter to another CPU's counter, 4991 * to keep the global sum constant after CPU-down: 4992 */ 4993static void migrate_nr_uninterruptible(struct rq *rq_src) 4994{ 4995 struct rq *rq_dest = cpu_rq(cpumask_any(cpu_active_mask)); 4996 4997 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible; 4998 rq_src->nr_uninterruptible = 0; 4999} 5000 5001/* 5002 * remove the tasks which were accounted by rq from calc_load_tasks. 5003 */ 5004static void calc_global_load_remove(struct rq *rq) 5005{ 5006 atomic_long_sub(rq->calc_load_active, &calc_load_tasks); 5007 rq->calc_load_active = 0; 5008} 5009 5010/* 5011 * Migrate all tasks from the rq, sleeping tasks will be migrated by 5012 * try_to_wake_up()->select_task_rq(). 5013 * 5014 * Called with rq->lock held even though we'er in stop_machine() and 5015 * there's no concurrency possible, we hold the required locks anyway 5016 * because of lock validation efforts. 5017 */ 5018static void migrate_tasks(unsigned int dead_cpu) 5019{ 5020 struct rq *rq = cpu_rq(dead_cpu); 5021 struct task_struct *next, *stop = rq->stop; 5022 int dest_cpu; 5023 5024 /* 5025 * Fudge the rq selection such that the below task selection loop 5026 * doesn't get stuck on the currently eligible stop task. 5027 * 5028 * We're currently inside stop_machine() and the rq is either stuck 5029 * in the stop_machine_cpu_stop() loop, or we're executing this code, 5030 * either way we should never end up calling schedule() until we're 5031 * done here. 5032 */ 5033 rq->stop = NULL; 5034 5035 /* Ensure any throttled groups are reachable by pick_next_task */ 5036 unthrottle_offline_cfs_rqs(rq); 5037 5038 for ( ; ; ) { 5039 /* 5040 * There's this thread running, bail when that's the only 5041 * remaining thread. 5042 */ 5043 if (rq->nr_running == 1) 5044 break; 5045 5046 next = pick_next_task(rq); 5047 BUG_ON(!next); 5048 next->sched_class->put_prev_task(rq, next); 5049 5050 /* Find suitable destination for @next, with force if needed. */ 5051 dest_cpu = select_fallback_rq(dead_cpu, next); 5052 raw_spin_unlock(&rq->lock); 5053 5054 __migrate_task(next, dead_cpu, dest_cpu); 5055 5056 raw_spin_lock(&rq->lock); 5057 } 5058 5059 rq->stop = stop; 5060} 5061 5062#endif /* CONFIG_HOTPLUG_CPU */ 5063 5064#if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL) 5065 5066static struct ctl_table sd_ctl_dir[] = { 5067 { 5068 .procname = "sched_domain", 5069 .mode = 0555, 5070 }, 5071 {} 5072}; 5073 5074static struct ctl_table sd_ctl_root[] = { 5075 { 5076 .procname = "kernel", 5077 .mode = 0555, 5078 .child = sd_ctl_dir, 5079 }, 5080 {} 5081}; 5082 5083static struct ctl_table *sd_alloc_ctl_entry(int n) 5084{ 5085 struct ctl_table *entry = 5086 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL); 5087 5088 return entry; 5089} 5090 5091static void sd_free_ctl_entry(struct ctl_table **tablep) 5092{ 5093 struct ctl_table *entry; 5094 5095 /* 5096 * In the intermediate directories, both the child directory and 5097 * procname are dynamically allocated and could fail but the mode 5098 * will always be set. In the lowest directory the names are 5099 * static strings and all have proc handlers. 5100 */ 5101 for (entry = *tablep; entry->mode; entry++) { 5102 if (entry->child) 5103 sd_free_ctl_entry(&entry->child); 5104 if (entry->proc_handler == NULL) 5105 kfree(entry->procname); 5106 } 5107 5108 kfree(*tablep); 5109 *tablep = NULL; 5110} 5111 5112static void 5113set_table_entry(struct ctl_table *entry, 5114 const char *procname, void *data, int maxlen, 5115 mode_t mode, proc_handler *proc_handler) 5116{ 5117 entry->procname = procname; 5118 entry->data = data; 5119 entry->maxlen = maxlen; 5120 entry->mode = mode; 5121 entry->proc_handler = proc_handler; 5122} 5123 5124static struct ctl_table * 5125sd_alloc_ctl_domain_table(struct sched_domain *sd) 5126{ 5127 struct ctl_table *table = sd_alloc_ctl_entry(13); 5128 5129 if (table == NULL) 5130 return NULL; 5131 5132 set_table_entry(&table[0], "min_interval", &sd->min_interval, 5133 sizeof(long), 0644, proc_doulongvec_minmax); 5134 set_table_entry(&table[1], "max_interval", &sd->max_interval, 5135 sizeof(long), 0644, proc_doulongvec_minmax); 5136 set_table_entry(&table[2], "busy_idx", &sd->busy_idx, 5137 sizeof(int), 0644, proc_dointvec_minmax); 5138 set_table_entry(&table[3], "idle_idx", &sd->idle_idx, 5139 sizeof(int), 0644, proc_dointvec_minmax); 5140 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx, 5141 sizeof(int), 0644, proc_dointvec_minmax); 5142 set_table_entry(&table[5], "wake_idx", &sd->wake_idx, 5143 sizeof(int), 0644, proc_dointvec_minmax); 5144 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx, 5145 sizeof(int), 0644, proc_dointvec_minmax); 5146 set_table_entry(&table[7], "busy_factor", &sd->busy_factor, 5147 sizeof(int), 0644, proc_dointvec_minmax); 5148 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct, 5149 sizeof(int), 0644, proc_dointvec_minmax); 5150 set_table_entry(&table[9], "cache_nice_tries", 5151 &sd->cache_nice_tries, 5152 sizeof(int), 0644, proc_dointvec_minmax); 5153 set_table_entry(&table[10], "flags", &sd->flags, 5154 sizeof(int), 0644, proc_dointvec_minmax); 5155 set_table_entry(&table[11], "name", sd->name, 5156 CORENAME_MAX_SIZE, 0444, proc_dostring); 5157 /* &table[12] is terminator */ 5158 5159 return table; 5160} 5161 5162static ctl_table *sd_alloc_ctl_cpu_table(int cpu) 5163{ 5164 struct ctl_table *entry, *table; 5165 struct sched_domain *sd; 5166 int domain_num = 0, i; 5167 char buf[32]; 5168 5169 for_each_domain(cpu, sd) 5170 domain_num++; 5171 entry = table = sd_alloc_ctl_entry(domain_num + 1); 5172 if (table == NULL) 5173 return NULL; 5174 5175 i = 0; 5176 for_each_domain(cpu, sd) { 5177 snprintf(buf, 32, "domain%d", i); 5178 entry->procname = kstrdup(buf, GFP_KERNEL); 5179 entry->mode = 0555; 5180 entry->child = sd_alloc_ctl_domain_table(sd); 5181 entry++; 5182 i++; 5183 } 5184 return table; 5185} 5186 5187static struct ctl_table_header *sd_sysctl_header; 5188static void register_sched_domain_sysctl(void) 5189{ 5190 int i, cpu_num = num_possible_cpus(); 5191 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1); 5192 char buf[32]; 5193 5194 WARN_ON(sd_ctl_dir[0].child); 5195 sd_ctl_dir[0].child = entry; 5196 5197 if (entry == NULL) 5198 return; 5199 5200 for_each_possible_cpu(i) { 5201 snprintf(buf, 32, "cpu%d", i); 5202 entry->procname = kstrdup(buf, GFP_KERNEL); 5203 entry->mode = 0555; 5204 entry->child = sd_alloc_ctl_cpu_table(i); 5205 entry++; 5206 } 5207 5208 WARN_ON(sd_sysctl_header); 5209 sd_sysctl_header = register_sysctl_table(sd_ctl_root); 5210} 5211 5212/* may be called multiple times per register */ 5213static void unregister_sched_domain_sysctl(void) 5214{ 5215 if (sd_sysctl_header) 5216 unregister_sysctl_table(sd_sysctl_header); 5217 sd_sysctl_header = NULL; 5218 if (sd_ctl_dir[0].child) 5219 sd_free_ctl_entry(&sd_ctl_dir[0].child); 5220} 5221#else 5222static void register_sched_domain_sysctl(void) 5223{ 5224} 5225static void unregister_sched_domain_sysctl(void) 5226{ 5227} 5228#endif 5229 5230static void set_rq_online(struct rq *rq) 5231{ 5232 if (!rq->online) { 5233 const struct sched_class *class; 5234 5235 cpumask_set_cpu(rq->cpu, rq->rd->online); 5236 rq->online = 1; 5237 5238 for_each_class(class) { 5239 if (class->rq_online) 5240 class->rq_online(rq); 5241 } 5242 } 5243} 5244 5245static void set_rq_offline(struct rq *rq) 5246{ 5247 if (rq->online) { 5248 const struct sched_class *class; 5249 5250 for_each_class(class) { 5251 if (class->rq_offline) 5252 class->rq_offline(rq); 5253 } 5254 5255 cpumask_clear_cpu(rq->cpu, rq->rd->online); 5256 rq->online = 0; 5257 } 5258} 5259 5260/* 5261 * migration_call - callback that gets triggered when a CPU is added. 5262 * Here we can start up the necessary migration thread for the new CPU. 5263 */ 5264static int __cpuinit 5265migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu) 5266{ 5267 int cpu = (long)hcpu; 5268 unsigned long flags; 5269 struct rq *rq = cpu_rq(cpu); 5270 5271 switch (action & ~CPU_TASKS_FROZEN) { 5272 5273 case CPU_UP_PREPARE: 5274 rq->calc_load_update = calc_load_update; 5275 break; 5276 5277 case CPU_ONLINE: 5278 /* Update our root-domain */ 5279 raw_spin_lock_irqsave(&rq->lock, flags); 5280 if (rq->rd) { 5281 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span)); 5282 5283 set_rq_online(rq); 5284 } 5285 raw_spin_unlock_irqrestore(&rq->lock, flags); 5286 break; 5287 5288#ifdef CONFIG_HOTPLUG_CPU 5289 case CPU_DYING: 5290 sched_ttwu_pending(); 5291 /* Update our root-domain */ 5292 raw_spin_lock_irqsave(&rq->lock, flags); 5293 if (rq->rd) { 5294 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span)); 5295 set_rq_offline(rq); 5296 } 5297 migrate_tasks(cpu); 5298 BUG_ON(rq->nr_running != 1); /* the migration thread */ 5299 raw_spin_unlock_irqrestore(&rq->lock, flags); 5300 5301 migrate_nr_uninterruptible(rq); 5302 calc_global_load_remove(rq); 5303 break; 5304#endif 5305 } 5306 5307 update_max_interval(); 5308 5309 return NOTIFY_OK; 5310} 5311 5312/* 5313 * Register at high priority so that task migration (migrate_all_tasks) 5314 * happens before everything else. This has to be lower priority than 5315 * the notifier in the perf_event subsystem, though. 5316 */ 5317static struct notifier_block __cpuinitdata migration_notifier = { 5318 .notifier_call = migration_call, 5319 .priority = CPU_PRI_MIGRATION, 5320}; 5321 5322static int __cpuinit sched_cpu_active(struct notifier_block *nfb, 5323 unsigned long action, void *hcpu) 5324{ 5325 switch (action & ~CPU_TASKS_FROZEN) { 5326 case CPU_ONLINE: 5327 case CPU_DOWN_FAILED: 5328 set_cpu_active((long)hcpu, true); 5329 return NOTIFY_OK; 5330 default: 5331 return NOTIFY_DONE; 5332 } 5333} 5334 5335static int __cpuinit sched_cpu_inactive(struct notifier_block *nfb, 5336 unsigned long action, void *hcpu) 5337{ 5338 switch (action & ~CPU_TASKS_FROZEN) { 5339 case CPU_DOWN_PREPARE: 5340 set_cpu_active((long)hcpu, false); 5341 return NOTIFY_OK; 5342 default: 5343 return NOTIFY_DONE; 5344 } 5345} 5346 5347static int __init migration_init(void) 5348{ 5349 void *cpu = (void *)(long)smp_processor_id(); 5350 int err; 5351 5352 /* Initialize migration for the boot CPU */ 5353 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu); 5354 BUG_ON(err == NOTIFY_BAD); 5355 migration_call(&migration_notifier, CPU_ONLINE, cpu); 5356 register_cpu_notifier(&migration_notifier); 5357 5358 /* Register cpu active notifiers */ 5359 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE); 5360 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE); 5361 5362 return 0; 5363} 5364early_initcall(migration_init); 5365#endif 5366 5367#ifdef CONFIG_SMP 5368 5369static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */ 5370 5371#ifdef CONFIG_SCHED_DEBUG 5372 5373static __read_mostly int sched_domain_debug_enabled; 5374 5375static int __init sched_domain_debug_setup(char *str) 5376{ 5377 sched_domain_debug_enabled = 1; 5378 5379 return 0; 5380} 5381early_param("sched_debug", sched_domain_debug_setup); 5382 5383static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level, 5384 struct cpumask *groupmask) 5385{ 5386 struct sched_group *group = sd->groups; 5387 char str[256]; 5388 5389 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd)); 5390 cpumask_clear(groupmask); 5391 5392 printk(KERN_DEBUG "%*s domain %d: ", level, "", level); 5393 5394 if (!(sd->flags & SD_LOAD_BALANCE)) { 5395 printk("does not load-balance\n"); 5396 if (sd->parent) 5397 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain" 5398 " has parent"); 5399 return -1; 5400 } 5401 5402 printk(KERN_CONT "span %s level %s\n", str, sd->name); 5403 5404 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) { 5405 printk(KERN_ERR "ERROR: domain->span does not contain " 5406 "CPU%d\n", cpu); 5407 } 5408 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) { 5409 printk(KERN_ERR "ERROR: domain->groups does not contain" 5410 " CPU%d\n", cpu); 5411 } 5412 5413 printk(KERN_DEBUG "%*s groups:", level + 1, ""); 5414 do { 5415 if (!group) { 5416 printk("\n"); 5417 printk(KERN_ERR "ERROR: group is NULL\n"); 5418 break; 5419 } 5420 5421 if (!group->sgp->power) { 5422 printk(KERN_CONT "\n"); 5423 printk(KERN_ERR "ERROR: domain->cpu_power not " 5424 "set\n"); 5425 break; 5426 } 5427 5428 if (!cpumask_weight(sched_group_cpus(group))) { 5429 printk(KERN_CONT "\n"); 5430 printk(KERN_ERR "ERROR: empty group\n"); 5431 break; 5432 } 5433 5434 if (cpumask_intersects(groupmask, sched_group_cpus(group))) { 5435 printk(KERN_CONT "\n"); 5436 printk(KERN_ERR "ERROR: repeated CPUs\n"); 5437 break; 5438 } 5439 5440 cpumask_or(groupmask, groupmask, sched_group_cpus(group)); 5441 5442 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group)); 5443 5444 printk(KERN_CONT " %s", str); 5445 if (group->sgp->power != SCHED_POWER_SCALE) { 5446 printk(KERN_CONT " (cpu_power = %d)", 5447 group->sgp->power); 5448 } 5449 5450 group = group->next; 5451 } while (group != sd->groups); 5452 printk(KERN_CONT "\n"); 5453 5454 if (!cpumask_equal(sched_domain_span(sd), groupmask)) 5455 printk(KERN_ERR "ERROR: groups don't span domain->span\n"); 5456 5457 if (sd->parent && 5458 !cpumask_subset(groupmask, sched_domain_span(sd->parent))) 5459 printk(KERN_ERR "ERROR: parent span is not a superset " 5460 "of domain->span\n"); 5461 return 0; 5462} 5463 5464static void sched_domain_debug(struct sched_domain *sd, int cpu) 5465{ 5466 int level = 0; 5467 5468 if (!sched_domain_debug_enabled) 5469 return; 5470 5471 if (!sd) { 5472 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu); 5473 return; 5474 } 5475 5476 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu); 5477 5478 for (;;) { 5479 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask)) 5480 break; 5481 level++; 5482 sd = sd->parent; 5483 if (!sd) 5484 break; 5485 } 5486} 5487#else /* !CONFIG_SCHED_DEBUG */ 5488# define sched_domain_debug(sd, cpu) do { } while (0) 5489#endif /* CONFIG_SCHED_DEBUG */ 5490 5491static int sd_degenerate(struct sched_domain *sd) 5492{ 5493 if (cpumask_weight(sched_domain_span(sd)) == 1) 5494 return 1; 5495 5496 /* Following flags need at least 2 groups */ 5497 if (sd->flags & (SD_LOAD_BALANCE | 5498 SD_BALANCE_NEWIDLE | 5499 SD_BALANCE_FORK | 5500 SD_BALANCE_EXEC | 5501 SD_SHARE_CPUPOWER | 5502 SD_SHARE_PKG_RESOURCES)) { 5503 if (sd->groups != sd->groups->next) 5504 return 0; 5505 } 5506 5507 /* Following flags don't use groups */ 5508 if (sd->flags & (SD_WAKE_AFFINE)) 5509 return 0; 5510 5511 return 1; 5512} 5513 5514static int 5515sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent) 5516{ 5517 unsigned long cflags = sd->flags, pflags = parent->flags; 5518 5519 if (sd_degenerate(parent)) 5520 return 1; 5521 5522 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent))) 5523 return 0; 5524 5525 /* Flags needing groups don't count if only 1 group in parent */ 5526 if (parent->groups == parent->groups->next) { 5527 pflags &= ~(SD_LOAD_BALANCE | 5528 SD_BALANCE_NEWIDLE | 5529 SD_BALANCE_FORK | 5530 SD_BALANCE_EXEC | 5531 SD_SHARE_CPUPOWER | 5532 SD_SHARE_PKG_RESOURCES); 5533 if (nr_node_ids == 1) 5534 pflags &= ~SD_SERIALIZE; 5535 } 5536 if (~cflags & pflags) 5537 return 0; 5538 5539 return 1; 5540} 5541 5542static void free_rootdomain(struct rcu_head *rcu) 5543{ 5544 struct root_domain *rd = container_of(rcu, struct root_domain, rcu); 5545 5546 cpupri_cleanup(&rd->cpupri); 5547 free_cpumask_var(rd->rto_mask); 5548 free_cpumask_var(rd->online); 5549 free_cpumask_var(rd->span); 5550 kfree(rd); 5551} 5552 5553static void rq_attach_root(struct rq *rq, struct root_domain *rd) 5554{ 5555 struct root_domain *old_rd = NULL; 5556 unsigned long flags; 5557 5558 raw_spin_lock_irqsave(&rq->lock, flags); 5559 5560 if (rq->rd) { 5561 old_rd = rq->rd; 5562 5563 if (cpumask_test_cpu(rq->cpu, old_rd->online)) 5564 set_rq_offline(rq); 5565 5566 cpumask_clear_cpu(rq->cpu, old_rd->span); 5567 5568 /* 5569 * If we dont want to free the old_rt yet then 5570 * set old_rd to NULL to skip the freeing later 5571 * in this function: 5572 */ 5573 if (!atomic_dec_and_test(&old_rd->refcount)) 5574 old_rd = NULL; 5575 } 5576 5577 atomic_inc(&rd->refcount); 5578 rq->rd = rd; 5579 5580 cpumask_set_cpu(rq->cpu, rd->span); 5581 if (cpumask_test_cpu(rq->cpu, cpu_active_mask)) 5582 set_rq_online(rq); 5583 5584 raw_spin_unlock_irqrestore(&rq->lock, flags); 5585 5586 if (old_rd) 5587 call_rcu_sched(&old_rd->rcu, free_rootdomain); 5588} 5589 5590static int init_rootdomain(struct root_domain *rd) 5591{ 5592 memset(rd, 0, sizeof(*rd)); 5593 5594 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL)) 5595 goto out; 5596 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL)) 5597 goto free_span; 5598 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL)) 5599 goto free_online; 5600 5601 if (cpupri_init(&rd->cpupri) != 0) 5602 goto free_rto_mask; 5603 return 0; 5604 5605free_rto_mask: 5606 free_cpumask_var(rd->rto_mask); 5607free_online: 5608 free_cpumask_var(rd->online); 5609free_span: 5610 free_cpumask_var(rd->span); 5611out: 5612 return -ENOMEM; 5613} 5614 5615/* 5616 * By default the system creates a single root-domain with all cpus as 5617 * members (mimicking the global state we have today). 5618 */ 5619struct root_domain def_root_domain; 5620 5621static void init_defrootdomain(void) 5622{ 5623 init_rootdomain(&def_root_domain); 5624 5625 atomic_set(&def_root_domain.refcount, 1); 5626} 5627 5628static struct root_domain *alloc_rootdomain(void) 5629{ 5630 struct root_domain *rd; 5631 5632 rd = kmalloc(sizeof(*rd), GFP_KERNEL); 5633 if (!rd) 5634 return NULL; 5635 5636 if (init_rootdomain(rd) != 0) { 5637 kfree(rd); 5638 return NULL; 5639 } 5640 5641 return rd; 5642} 5643 5644static void free_sched_groups(struct sched_group *sg, int free_sgp) 5645{ 5646 struct sched_group *tmp, *first; 5647 5648 if (!sg) 5649 return; 5650 5651 first = sg; 5652 do { 5653 tmp = sg->next; 5654 5655 if (free_sgp && atomic_dec_and_test(&sg->sgp->ref)) 5656 kfree(sg->sgp); 5657 5658 kfree(sg); 5659 sg = tmp; 5660 } while (sg != first); 5661} 5662 5663static void free_sched_domain(struct rcu_head *rcu) 5664{ 5665 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu); 5666 5667 /* 5668 * If its an overlapping domain it has private groups, iterate and 5669 * nuke them all. 5670 */ 5671 if (sd->flags & SD_OVERLAP) { 5672 free_sched_groups(sd->groups, 1); 5673 } else if (atomic_dec_and_test(&sd->groups->ref)) { 5674 kfree(sd->groups->sgp); 5675 kfree(sd->groups); 5676 } 5677 kfree(sd); 5678} 5679 5680static void destroy_sched_domain(struct sched_domain *sd, int cpu) 5681{ 5682 call_rcu(&sd->rcu, free_sched_domain); 5683} 5684 5685static void destroy_sched_domains(struct sched_domain *sd, int cpu) 5686{ 5687 for (; sd; sd = sd->parent) 5688 destroy_sched_domain(sd, cpu); 5689} 5690 5691/* 5692 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must 5693 * hold the hotplug lock. 5694 */ 5695static void 5696cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu) 5697{ 5698 struct rq *rq = cpu_rq(cpu); 5699 struct sched_domain *tmp; 5700 5701 /* Remove the sched domains which do not contribute to scheduling. */ 5702 for (tmp = sd; tmp; ) { 5703 struct sched_domain *parent = tmp->parent; 5704 if (!parent) 5705 break; 5706 5707 if (sd_parent_degenerate(tmp, parent)) { 5708 tmp->parent = parent->parent; 5709 if (parent->parent) 5710 parent->parent->child = tmp; 5711 destroy_sched_domain(parent, cpu); 5712 } else 5713 tmp = tmp->parent; 5714 } 5715 5716 if (sd && sd_degenerate(sd)) { 5717 tmp = sd; 5718 sd = sd->parent; 5719 destroy_sched_domain(tmp, cpu); 5720 if (sd) 5721 sd->child = NULL; 5722 } 5723 5724 sched_domain_debug(sd, cpu); 5725 5726 rq_attach_root(rq, rd); 5727 tmp = rq->sd; 5728 rcu_assign_pointer(rq->sd, sd); 5729 destroy_sched_domains(tmp, cpu); 5730} 5731 5732/* cpus with isolated domains */ 5733static cpumask_var_t cpu_isolated_map; 5734 5735/* Setup the mask of cpus configured for isolated domains */ 5736static int __init isolated_cpu_setup(char *str) 5737{ 5738 alloc_bootmem_cpumask_var(&cpu_isolated_map); 5739 cpulist_parse(str, cpu_isolated_map); 5740 return 1; 5741} 5742 5743__setup("isolcpus=", isolated_cpu_setup); 5744 5745#ifdef CONFIG_NUMA 5746 5747/** 5748 * find_next_best_node - find the next node to include in a sched_domain 5749 * @node: node whose sched_domain we're building 5750 * @used_nodes: nodes already in the sched_domain 5751 * 5752 * Find the next node to include in a given scheduling domain. Simply 5753 * finds the closest node not already in the @used_nodes map. 5754 * 5755 * Should use nodemask_t. 5756 */ 5757static int find_next_best_node(int node, nodemask_t *used_nodes) 5758{ 5759 int i, n, val, min_val, best_node = -1; 5760 5761 min_val = INT_MAX; 5762 5763 for (i = 0; i < nr_node_ids; i++) { 5764 /* Start at @node */ 5765 n = (node + i) % nr_node_ids; 5766 5767 if (!nr_cpus_node(n)) 5768 continue; 5769 5770 /* Skip already used nodes */ 5771 if (node_isset(n, *used_nodes)) 5772 continue; 5773 5774 /* Simple min distance search */ 5775 val = node_distance(node, n); 5776 5777 if (val < min_val) { 5778 min_val = val; 5779 best_node = n; 5780 } 5781 } 5782 5783 if (best_node != -1) 5784 node_set(best_node, *used_nodes); 5785 return best_node; 5786} 5787 5788/** 5789 * sched_domain_node_span - get a cpumask for a node's sched_domain 5790 * @node: node whose cpumask we're constructing 5791 * @span: resulting cpumask 5792 * 5793 * Given a node, construct a good cpumask for its sched_domain to span. It 5794 * should be one that prevents unnecessary balancing, but also spreads tasks 5795 * out optimally. 5796 */ 5797static void sched_domain_node_span(int node, struct cpumask *span) 5798{ 5799 nodemask_t used_nodes; 5800 int i; 5801 5802 cpumask_clear(span); 5803 nodes_clear(used_nodes); 5804 5805 cpumask_or(span, span, cpumask_of_node(node)); 5806 node_set(node, used_nodes); 5807 5808 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) { 5809 int next_node = find_next_best_node(node, &used_nodes); 5810 if (next_node < 0) 5811 break; 5812 cpumask_or(span, span, cpumask_of_node(next_node)); 5813 } 5814} 5815 5816static const struct cpumask *cpu_node_mask(int cpu) 5817{ 5818 lockdep_assert_held(&sched_domains_mutex); 5819 5820 sched_domain_node_span(cpu_to_node(cpu), sched_domains_tmpmask); 5821 5822 return sched_domains_tmpmask; 5823} 5824 5825static const struct cpumask *cpu_allnodes_mask(int cpu) 5826{ 5827 return cpu_possible_mask; 5828} 5829#endif /* CONFIG_NUMA */ 5830 5831static const struct cpumask *cpu_cpu_mask(int cpu) 5832{ 5833 return cpumask_of_node(cpu_to_node(cpu)); 5834} 5835 5836int sched_smt_power_savings = 0, sched_mc_power_savings = 0; 5837 5838struct sd_data { 5839 struct sched_domain **__percpu sd; 5840 struct sched_group **__percpu sg; 5841 struct sched_group_power **__percpu sgp; 5842}; 5843 5844struct s_data { 5845 struct sched_domain ** __percpu sd; 5846 struct root_domain *rd; 5847}; 5848 5849enum s_alloc { 5850 sa_rootdomain, 5851 sa_sd, 5852 sa_sd_storage, 5853 sa_none, 5854}; 5855 5856struct sched_domain_topology_level; 5857 5858typedef struct sched_domain *(*sched_domain_init_f)(struct sched_domain_topology_level *tl, int cpu); 5859typedef const struct cpumask *(*sched_domain_mask_f)(int cpu); 5860 5861#define SDTL_OVERLAP 0x01 5862 5863struct sched_domain_topology_level { 5864 sched_domain_init_f init; 5865 sched_domain_mask_f mask; 5866 int flags; 5867 struct sd_data data; 5868}; 5869 5870static int 5871build_overlap_sched_groups(struct sched_domain *sd, int cpu) 5872{ 5873 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg; 5874 const struct cpumask *span = sched_domain_span(sd); 5875 struct cpumask *covered = sched_domains_tmpmask; 5876 struct sd_data *sdd = sd->private; 5877 struct sched_domain *child; 5878 int i; 5879 5880 cpumask_clear(covered); 5881 5882 for_each_cpu(i, span) { 5883 struct cpumask *sg_span; 5884 5885 if (cpumask_test_cpu(i, covered)) 5886 continue; 5887 5888 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(), 5889 GFP_KERNEL, cpu_to_node(cpu)); 5890 5891 if (!sg) 5892 goto fail; 5893 5894 sg_span = sched_group_cpus(sg); 5895 5896 child = *per_cpu_ptr(sdd->sd, i); 5897 if (child->child) { 5898 child = child->child; 5899 cpumask_copy(sg_span, sched_domain_span(child)); 5900 } else 5901 cpumask_set_cpu(i, sg_span); 5902 5903 cpumask_or(covered, covered, sg_span); 5904 5905 sg->sgp = *per_cpu_ptr(sdd->sgp, cpumask_first(sg_span)); 5906 atomic_inc(&sg->sgp->ref); 5907 5908 if (cpumask_test_cpu(cpu, sg_span)) 5909 groups = sg; 5910 5911 if (!first) 5912 first = sg; 5913 if (last) 5914 last->next = sg; 5915 last = sg; 5916 last->next = first; 5917 } 5918 sd->groups = groups; 5919 5920 return 0; 5921 5922fail: 5923 free_sched_groups(first, 0); 5924 5925 return -ENOMEM; 5926} 5927 5928static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg) 5929{ 5930 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu); 5931 struct sched_domain *child = sd->child; 5932 5933 if (child) 5934 cpu = cpumask_first(sched_domain_span(child)); 5935 5936 if (sg) { 5937 *sg = *per_cpu_ptr(sdd->sg, cpu); 5938 (*sg)->sgp = *per_cpu_ptr(sdd->sgp, cpu); 5939 atomic_set(&(*sg)->sgp->ref, 1); /* for claim_allocations */ 5940 } 5941 5942 return cpu; 5943} 5944 5945/* 5946 * build_sched_groups will build a circular linked list of the groups 5947 * covered by the given span, and will set each group's ->cpumask correctly, 5948 * and ->cpu_power to 0. 5949 * 5950 * Assumes the sched_domain tree is fully constructed 5951 */ 5952static int 5953build_sched_groups(struct sched_domain *sd, int cpu) 5954{ 5955 struct sched_group *first = NULL, *last = NULL; 5956 struct sd_data *sdd = sd->private; 5957 const struct cpumask *span = sched_domain_span(sd); 5958 struct cpumask *covered; 5959 int i; 5960 5961 get_group(cpu, sdd, &sd->groups); 5962 atomic_inc(&sd->groups->ref); 5963 5964 if (cpu != cpumask_first(sched_domain_span(sd))) 5965 return 0; 5966 5967 lockdep_assert_held(&sched_domains_mutex); 5968 covered = sched_domains_tmpmask; 5969 5970 cpumask_clear(covered); 5971 5972 for_each_cpu(i, span) { 5973 struct sched_group *sg; 5974 int group = get_group(i, sdd, &sg); 5975 int j; 5976 5977 if (cpumask_test_cpu(i, covered)) 5978 continue; 5979 5980 cpumask_clear(sched_group_cpus(sg)); 5981 sg->sgp->power = 0; 5982 5983 for_each_cpu(j, span) { 5984 if (get_group(j, sdd, NULL) != group) 5985 continue; 5986 5987 cpumask_set_cpu(j, covered); 5988 cpumask_set_cpu(j, sched_group_cpus(sg)); 5989 } 5990 5991 if (!first) 5992 first = sg; 5993 if (last) 5994 last->next = sg; 5995 last = sg; 5996 } 5997 last->next = first; 5998 5999 return 0; 6000} 6001 6002/* 6003 * Initialize sched groups cpu_power. 6004 * 6005 * cpu_power indicates the capacity of sched group, which is used while 6006 * distributing the load between different sched groups in a sched domain. 6007 * Typically cpu_power for all the groups in a sched domain will be same unless 6008 * there are asymmetries in the topology. If there are asymmetries, group 6009 * having more cpu_power will pickup more load compared to the group having 6010 * less cpu_power. 6011 */ 6012static void init_sched_groups_power(int cpu, struct sched_domain *sd) 6013{ 6014 struct sched_group *sg = sd->groups; 6015 6016 WARN_ON(!sd || !sg); 6017 6018 do { 6019 sg->group_weight = cpumask_weight(sched_group_cpus(sg)); 6020 sg = sg->next; 6021 } while (sg != sd->groups); 6022 6023 if (cpu != group_first_cpu(sg)) 6024 return; 6025 6026 update_group_power(sd, cpu); 6027} 6028 6029int __weak arch_sd_sibling_asym_packing(void) 6030{ 6031 return 0*SD_ASYM_PACKING; 6032} 6033 6034/* 6035 * Initializers for schedule domains 6036 * Non-inlined to reduce accumulated stack pressure in build_sched_domains() 6037 */ 6038 6039#ifdef CONFIG_SCHED_DEBUG 6040# define SD_INIT_NAME(sd, type) sd->name = #type 6041#else 6042# define SD_INIT_NAME(sd, type) do { } while (0) 6043#endif 6044 6045#define SD_INIT_FUNC(type) \ 6046static noinline struct sched_domain * \ 6047sd_init_##type(struct sched_domain_topology_level *tl, int cpu) \ 6048{ \ 6049 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); \ 6050 *sd = SD_##type##_INIT; \ 6051 SD_INIT_NAME(sd, type); \ 6052 sd->private = &tl->data; \ 6053 return sd; \ 6054} 6055 6056SD_INIT_FUNC(CPU) 6057#ifdef CONFIG_NUMA 6058 SD_INIT_FUNC(ALLNODES) 6059 SD_INIT_FUNC(NODE) 6060#endif 6061#ifdef CONFIG_SCHED_SMT 6062 SD_INIT_FUNC(SIBLING) 6063#endif 6064#ifdef CONFIG_SCHED_MC 6065 SD_INIT_FUNC(MC) 6066#endif 6067#ifdef CONFIG_SCHED_BOOK 6068 SD_INIT_FUNC(BOOK) 6069#endif 6070 6071static int default_relax_domain_level = -1; 6072int sched_domain_level_max; 6073 6074static int __init setup_relax_domain_level(char *str) 6075{ 6076 unsigned long val; 6077 6078 val = simple_strtoul(str, NULL, 0); 6079 if (val < sched_domain_level_max) 6080 default_relax_domain_level = val; 6081 6082 return 1; 6083} 6084__setup("relax_domain_level=", setup_relax_domain_level); 6085 6086static void set_domain_attribute(struct sched_domain *sd, 6087 struct sched_domain_attr *attr) 6088{ 6089 int request; 6090 6091 if (!attr || attr->relax_domain_level < 0) { 6092 if (default_relax_domain_level < 0) 6093 return; 6094 else 6095 request = default_relax_domain_level; 6096 } else 6097 request = attr->relax_domain_level; 6098 if (request < sd->level) { 6099 /* turn off idle balance on this domain */ 6100 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE); 6101 } else { 6102 /* turn on idle balance on this domain */ 6103 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE); 6104 } 6105} 6106 6107static void __sdt_free(const struct cpumask *cpu_map); 6108static int __sdt_alloc(const struct cpumask *cpu_map); 6109 6110static void __free_domain_allocs(struct s_data *d, enum s_alloc what, 6111 const struct cpumask *cpu_map) 6112{ 6113 switch (what) { 6114 case sa_rootdomain: 6115 if (!atomic_read(&d->rd->refcount)) 6116 free_rootdomain(&d->rd->rcu); /* fall through */ 6117 case sa_sd: 6118 free_percpu(d->sd); /* fall through */ 6119 case sa_sd_storage: 6120 __sdt_free(cpu_map); /* fall through */ 6121 case sa_none: 6122 break; 6123 } 6124} 6125 6126static enum s_alloc __visit_domain_allocation_hell(struct s_data *d, 6127 const struct cpumask *cpu_map) 6128{ 6129 memset(d, 0, sizeof(*d)); 6130 6131 if (__sdt_alloc(cpu_map)) 6132 return sa_sd_storage; 6133 d->sd = alloc_percpu(struct sched_domain *); 6134 if (!d->sd) 6135 return sa_sd_storage; 6136 d->rd = alloc_rootdomain(); 6137 if (!d->rd) 6138 return sa_sd; 6139 return sa_rootdomain; 6140} 6141 6142/* 6143 * NULL the sd_data elements we've used to build the sched_domain and 6144 * sched_group structure so that the subsequent __free_domain_allocs() 6145 * will not free the data we're using. 6146 */ 6147static void claim_allocations(int cpu, struct sched_domain *sd) 6148{ 6149 struct sd_data *sdd = sd->private; 6150 6151 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd); 6152 *per_cpu_ptr(sdd->sd, cpu) = NULL; 6153 6154 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref)) 6155 *per_cpu_ptr(sdd->sg, cpu) = NULL; 6156 6157 if (atomic_read(&(*per_cpu_ptr(sdd->sgp, cpu))->ref)) 6158 *per_cpu_ptr(sdd->sgp, cpu) = NULL; 6159} 6160 6161#ifdef CONFIG_SCHED_SMT 6162static const struct cpumask *cpu_smt_mask(int cpu) 6163{ 6164 return topology_thread_cpumask(cpu); 6165} 6166#endif 6167 6168/* 6169 * Topology list, bottom-up. 6170 */ 6171static struct sched_domain_topology_level default_topology[] = { 6172#ifdef CONFIG_SCHED_SMT 6173 { sd_init_SIBLING, cpu_smt_mask, }, 6174#endif 6175#ifdef CONFIG_SCHED_MC 6176 { sd_init_MC, cpu_coregroup_mask, }, 6177#endif 6178#ifdef CONFIG_SCHED_BOOK 6179 { sd_init_BOOK, cpu_book_mask, }, 6180#endif 6181 { sd_init_CPU, cpu_cpu_mask, }, 6182#ifdef CONFIG_NUMA 6183 { sd_init_NODE, cpu_node_mask, SDTL_OVERLAP, }, 6184 { sd_init_ALLNODES, cpu_allnodes_mask, }, 6185#endif 6186 { NULL, }, 6187}; 6188 6189static struct sched_domain_topology_level *sched_domain_topology = default_topology; 6190 6191static int __sdt_alloc(const struct cpumask *cpu_map) 6192{ 6193 struct sched_domain_topology_level *tl; 6194 int j; 6195 6196 for (tl = sched_domain_topology; tl->init; tl++) { 6197 struct sd_data *sdd = &tl->data; 6198 6199 sdd->sd = alloc_percpu(struct sched_domain *); 6200 if (!sdd->sd) 6201 return -ENOMEM; 6202 6203 sdd->sg = alloc_percpu(struct sched_group *); 6204 if (!sdd->sg) 6205 return -ENOMEM; 6206 6207 sdd->sgp = alloc_percpu(struct sched_group_power *); 6208 if (!sdd->sgp) 6209 return -ENOMEM; 6210 6211 for_each_cpu(j, cpu_map) { 6212 struct sched_domain *sd; 6213 struct sched_group *sg; 6214 struct sched_group_power *sgp; 6215 6216 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(), 6217 GFP_KERNEL, cpu_to_node(j)); 6218 if (!sd) 6219 return -ENOMEM; 6220 6221 *per_cpu_ptr(sdd->sd, j) = sd; 6222 6223 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(), 6224 GFP_KERNEL, cpu_to_node(j)); 6225 if (!sg) 6226 return -ENOMEM; 6227 6228 *per_cpu_ptr(sdd->sg, j) = sg; 6229 6230 sgp = kzalloc_node(sizeof(struct sched_group_power), 6231 GFP_KERNEL, cpu_to_node(j)); 6232 if (!sgp) 6233 return -ENOMEM; 6234 6235 *per_cpu_ptr(sdd->sgp, j) = sgp; 6236 } 6237 } 6238 6239 return 0; 6240} 6241 6242static void __sdt_free(const struct cpumask *cpu_map) 6243{ 6244 struct sched_domain_topology_level *tl; 6245 int j; 6246 6247 for (tl = sched_domain_topology; tl->init; tl++) { 6248 struct sd_data *sdd = &tl->data; 6249 6250 for_each_cpu(j, cpu_map) { 6251 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, j); 6252 if (sd && (sd->flags & SD_OVERLAP)) 6253 free_sched_groups(sd->groups, 0); 6254 kfree(*per_cpu_ptr(sdd->sd, j)); 6255 kfree(*per_cpu_ptr(sdd->sg, j)); 6256 kfree(*per_cpu_ptr(sdd->sgp, j)); 6257 } 6258 free_percpu(sdd->sd); 6259 free_percpu(sdd->sg); 6260 free_percpu(sdd->sgp); 6261 } 6262} 6263 6264struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl, 6265 struct s_data *d, const struct cpumask *cpu_map, 6266 struct sched_domain_attr *attr, struct sched_domain *child, 6267 int cpu) 6268{ 6269 struct sched_domain *sd = tl->init(tl, cpu); 6270 if (!sd) 6271 return child; 6272 6273 set_domain_attribute(sd, attr); 6274 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu)); 6275 if (child) { 6276 sd->level = child->level + 1; 6277 sched_domain_level_max = max(sched_domain_level_max, sd->level); 6278 child->parent = sd; 6279 } 6280 sd->child = child; 6281 6282 return sd; 6283} 6284 6285/* 6286 * Build sched domains for a given set of cpus and attach the sched domains 6287 * to the individual cpus 6288 */ 6289static int build_sched_domains(const struct cpumask *cpu_map, 6290 struct sched_domain_attr *attr) 6291{ 6292 enum s_alloc alloc_state = sa_none; 6293 struct sched_domain *sd; 6294 struct s_data d; 6295 int i, ret = -ENOMEM; 6296 6297 alloc_state = __visit_domain_allocation_hell(&d, cpu_map); 6298 if (alloc_state != sa_rootdomain) 6299 goto error; 6300 6301 /* Set up domains for cpus specified by the cpu_map. */ 6302 for_each_cpu(i, cpu_map) { 6303 struct sched_domain_topology_level *tl; 6304 6305 sd = NULL; 6306 for (tl = sched_domain_topology; tl->init; tl++) { 6307 sd = build_sched_domain(tl, &d, cpu_map, attr, sd, i); 6308 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP)) 6309 sd->flags |= SD_OVERLAP; 6310 if (cpumask_equal(cpu_map, sched_domain_span(sd))) 6311 break; 6312 } 6313 6314 while (sd->child) 6315 sd = sd->child; 6316 6317 *per_cpu_ptr(d.sd, i) = sd; 6318 } 6319 6320 /* Build the groups for the domains */ 6321 for_each_cpu(i, cpu_map) { 6322 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) { 6323 sd->span_weight = cpumask_weight(sched_domain_span(sd)); 6324 if (sd->flags & SD_OVERLAP) { 6325 if (build_overlap_sched_groups(sd, i)) 6326 goto error; 6327 } else { 6328 if (build_sched_groups(sd, i)) 6329 goto error; 6330 } 6331 } 6332 } 6333 6334 /* Calculate CPU power for physical packages and nodes */ 6335 for (i = nr_cpumask_bits-1; i >= 0; i--) { 6336 if (!cpumask_test_cpu(i, cpu_map)) 6337 continue; 6338 6339 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) { 6340 claim_allocations(i, sd); 6341 init_sched_groups_power(i, sd); 6342 } 6343 } 6344 6345 /* Attach the domains */ 6346 rcu_read_lock(); 6347 for_each_cpu(i, cpu_map) { 6348 sd = *per_cpu_ptr(d.sd, i); 6349 cpu_attach_domain(sd, d.rd, i); 6350 } 6351 rcu_read_unlock(); 6352 6353 ret = 0; 6354error: 6355 __free_domain_allocs(&d, alloc_state, cpu_map); 6356 return ret; 6357} 6358 6359static cpumask_var_t *doms_cur; /* current sched domains */ 6360static int ndoms_cur; /* number of sched domains in 'doms_cur' */ 6361static struct sched_domain_attr *dattr_cur; 6362 /* attribues of custom domains in 'doms_cur' */ 6363 6364/* 6365 * Special case: If a kmalloc of a doms_cur partition (array of 6366 * cpumask) fails, then fallback to a single sched domain, 6367 * as determined by the single cpumask fallback_doms. 6368 */ 6369static cpumask_var_t fallback_doms; 6370 6371/* 6372 * arch_update_cpu_topology lets virtualized architectures update the 6373 * cpu core maps. It is supposed to return 1 if the topology changed 6374 * or 0 if it stayed the same. 6375 */ 6376int __attribute__((weak)) arch_update_cpu_topology(void) 6377{ 6378 return 0; 6379} 6380 6381cpumask_var_t *alloc_sched_domains(unsigned int ndoms) 6382{ 6383 int i; 6384 cpumask_var_t *doms; 6385 6386 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL); 6387 if (!doms) 6388 return NULL; 6389 for (i = 0; i < ndoms; i++) { 6390 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) { 6391 free_sched_domains(doms, i); 6392 return NULL; 6393 } 6394 } 6395 return doms; 6396} 6397 6398void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms) 6399{ 6400 unsigned int i; 6401 for (i = 0; i < ndoms; i++) 6402 free_cpumask_var(doms[i]); 6403 kfree(doms); 6404} 6405 6406/* 6407 * Set up scheduler domains and groups. Callers must hold the hotplug lock. 6408 * For now this just excludes isolated cpus, but could be used to 6409 * exclude other special cases in the future. 6410 */ 6411static int init_sched_domains(const struct cpumask *cpu_map) 6412{ 6413 int err; 6414 6415 arch_update_cpu_topology(); 6416 ndoms_cur = 1; 6417 doms_cur = alloc_sched_domains(ndoms_cur); 6418 if (!doms_cur) 6419 doms_cur = &fallback_doms; 6420 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map); 6421 dattr_cur = NULL; 6422 err = build_sched_domains(doms_cur[0], NULL); 6423 register_sched_domain_sysctl(); 6424 6425 return err; 6426} 6427 6428/* 6429 * Detach sched domains from a group of cpus specified in cpu_map 6430 * These cpus will now be attached to the NULL domain 6431 */ 6432static void detach_destroy_domains(const struct cpumask *cpu_map) 6433{ 6434 int i; 6435 6436 rcu_read_lock(); 6437 for_each_cpu(i, cpu_map) 6438 cpu_attach_domain(NULL, &def_root_domain, i); 6439 rcu_read_unlock(); 6440} 6441 6442/* handle null as "default" */ 6443static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur, 6444 struct sched_domain_attr *new, int idx_new) 6445{ 6446 struct sched_domain_attr tmp; 6447 6448 /* fast path */ 6449 if (!new && !cur) 6450 return 1; 6451 6452 tmp = SD_ATTR_INIT; 6453 return !memcmp(cur ? (cur + idx_cur) : &tmp, 6454 new ? (new + idx_new) : &tmp, 6455 sizeof(struct sched_domain_attr)); 6456} 6457 6458/* 6459 * Partition sched domains as specified by the 'ndoms_new' 6460 * cpumasks in the array doms_new[] of cpumasks. This compares 6461 * doms_new[] to the current sched domain partitioning, doms_cur[]. 6462 * It destroys each deleted domain and builds each new domain. 6463 * 6464 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'. 6465 * The masks don't intersect (don't overlap.) We should setup one 6466 * sched domain for each mask. CPUs not in any of the cpumasks will 6467 * not be load balanced. If the same cpumask appears both in the 6468 * current 'doms_cur' domains and in the new 'doms_new', we can leave 6469 * it as it is. 6470 * 6471 * The passed in 'doms_new' should be allocated using 6472 * alloc_sched_domains. This routine takes ownership of it and will 6473 * free_sched_domains it when done with it. If the caller failed the 6474 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1, 6475 * and partition_sched_domains() will fallback to the single partition 6476 * 'fallback_doms', it also forces the domains to be rebuilt. 6477 * 6478 * If doms_new == NULL it will be replaced with cpu_online_mask. 6479 * ndoms_new == 0 is a special case for destroying existing domains, 6480 * and it will not create the default domain. 6481 * 6482 * Call with hotplug lock held 6483 */ 6484void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[], 6485 struct sched_domain_attr *dattr_new) 6486{ 6487 int i, j, n; 6488 int new_topology; 6489 6490 mutex_lock(&sched_domains_mutex); 6491 6492 /* always unregister in case we don't destroy any domains */ 6493 unregister_sched_domain_sysctl(); 6494 6495 /* Let architecture update cpu core mappings. */ 6496 new_topology = arch_update_cpu_topology(); 6497 6498 n = doms_new ? ndoms_new : 0; 6499 6500 /* Destroy deleted domains */ 6501 for (i = 0; i < ndoms_cur; i++) { 6502 for (j = 0; j < n && !new_topology; j++) { 6503 if (cpumask_equal(doms_cur[i], doms_new[j]) 6504 && dattrs_equal(dattr_cur, i, dattr_new, j)) 6505 goto match1; 6506 } 6507 /* no match - a current sched domain not in new doms_new[] */ 6508 detach_destroy_domains(doms_cur[i]); 6509match1: 6510 ; 6511 } 6512 6513 if (doms_new == NULL) { 6514 ndoms_cur = 0; 6515 doms_new = &fallback_doms; 6516 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map); 6517 WARN_ON_ONCE(dattr_new); 6518 } 6519 6520 /* Build new domains */ 6521 for (i = 0; i < ndoms_new; i++) { 6522 for (j = 0; j < ndoms_cur && !new_topology; j++) { 6523 if (cpumask_equal(doms_new[i], doms_cur[j]) 6524 && dattrs_equal(dattr_new, i, dattr_cur, j)) 6525 goto match2; 6526 } 6527 /* no match - add a new doms_new */ 6528 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL); 6529match2: 6530 ; 6531 } 6532 6533 /* Remember the new sched domains */ 6534 if (doms_cur != &fallback_doms) 6535 free_sched_domains(doms_cur, ndoms_cur); 6536 kfree(dattr_cur); /* kfree(NULL) is safe */ 6537 doms_cur = doms_new; 6538 dattr_cur = dattr_new; 6539 ndoms_cur = ndoms_new; 6540 6541 register_sched_domain_sysctl(); 6542 6543 mutex_unlock(&sched_domains_mutex); 6544} 6545 6546#if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT) 6547static void reinit_sched_domains(void) 6548{ 6549 get_online_cpus(); 6550 6551 /* Destroy domains first to force the rebuild */ 6552 partition_sched_domains(0, NULL, NULL); 6553 6554 rebuild_sched_domains(); 6555 put_online_cpus(); 6556} 6557 6558static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt) 6559{ 6560 unsigned int level = 0; 6561 6562 if (sscanf(buf, "%u", &level) != 1) 6563 return -EINVAL; 6564 6565 /* 6566 * level is always be positive so don't check for 6567 * level < POWERSAVINGS_BALANCE_NONE which is 0 6568 * What happens on 0 or 1 byte write, 6569 * need to check for count as well? 6570 */ 6571 6572 if (level >= MAX_POWERSAVINGS_BALANCE_LEVELS) 6573 return -EINVAL; 6574 6575 if (smt) 6576 sched_smt_power_savings = level; 6577 else 6578 sched_mc_power_savings = level; 6579 6580 reinit_sched_domains(); 6581 6582 return count; 6583} 6584 6585#ifdef CONFIG_SCHED_MC 6586static ssize_t sched_mc_power_savings_show(struct sysdev_class *class, 6587 struct sysdev_class_attribute *attr, 6588 char *page) 6589{ 6590 return sprintf(page, "%u\n", sched_mc_power_savings); 6591} 6592static ssize_t sched_mc_power_savings_store(struct sysdev_class *class, 6593 struct sysdev_class_attribute *attr, 6594 const char *buf, size_t count) 6595{ 6596 return sched_power_savings_store(buf, count, 0); 6597} 6598static SYSDEV_CLASS_ATTR(sched_mc_power_savings, 0644, 6599 sched_mc_power_savings_show, 6600 sched_mc_power_savings_store); 6601#endif 6602 6603#ifdef CONFIG_SCHED_SMT 6604static ssize_t sched_smt_power_savings_show(struct sysdev_class *dev, 6605 struct sysdev_class_attribute *attr, 6606 char *page) 6607{ 6608 return sprintf(page, "%u\n", sched_smt_power_savings); 6609} 6610static ssize_t sched_smt_power_savings_store(struct sysdev_class *dev, 6611 struct sysdev_class_attribute *attr, 6612 const char *buf, size_t count) 6613{ 6614 return sched_power_savings_store(buf, count, 1); 6615} 6616static SYSDEV_CLASS_ATTR(sched_smt_power_savings, 0644, 6617 sched_smt_power_savings_show, 6618 sched_smt_power_savings_store); 6619#endif 6620 6621int __init sched_create_sysfs_power_savings_entries(struct sysdev_class *cls) 6622{ 6623 int err = 0; 6624 6625#ifdef CONFIG_SCHED_SMT 6626 if (smt_capable()) 6627 err = sysfs_create_file(&cls->kset.kobj, 6628 &attr_sched_smt_power_savings.attr); 6629#endif 6630#ifdef CONFIG_SCHED_MC 6631 if (!err && mc_capable()) 6632 err = sysfs_create_file(&cls->kset.kobj, 6633 &attr_sched_mc_power_savings.attr); 6634#endif 6635 return err; 6636} 6637#endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */ 6638 6639/* 6640 * Update cpusets according to cpu_active mask. If cpusets are 6641 * disabled, cpuset_update_active_cpus() becomes a simple wrapper 6642 * around partition_sched_domains(). 6643 */ 6644static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action, 6645 void *hcpu) 6646{ 6647 switch (action & ~CPU_TASKS_FROZEN) { 6648 case CPU_ONLINE: 6649 case CPU_DOWN_FAILED: 6650 cpuset_update_active_cpus(); 6651 return NOTIFY_OK; 6652 default: 6653 return NOTIFY_DONE; 6654 } 6655} 6656 6657static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action, 6658 void *hcpu) 6659{ 6660 switch (action & ~CPU_TASKS_FROZEN) { 6661 case CPU_DOWN_PREPARE: 6662 cpuset_update_active_cpus(); 6663 return NOTIFY_OK; 6664 default: 6665 return NOTIFY_DONE; 6666 } 6667} 6668 6669void __init sched_init_smp(void) 6670{ 6671 cpumask_var_t non_isolated_cpus; 6672 6673 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL); 6674 alloc_cpumask_var(&fallback_doms, GFP_KERNEL); 6675 6676 get_online_cpus(); 6677 mutex_lock(&sched_domains_mutex); 6678 init_sched_domains(cpu_active_mask); 6679 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map); 6680 if (cpumask_empty(non_isolated_cpus)) 6681 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus); 6682 mutex_unlock(&sched_domains_mutex); 6683 put_online_cpus(); 6684 6685 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE); 6686 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE); 6687 6688 /* RT runtime code needs to handle some hotplug events */ 6689 hotcpu_notifier(update_runtime, 0); 6690 6691 init_hrtick(); 6692 6693 /* Move init over to a non-isolated CPU */ 6694 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0) 6695 BUG(); 6696 sched_init_granularity(); 6697 free_cpumask_var(non_isolated_cpus); 6698 6699 init_sched_rt_class(); 6700} 6701#else 6702void __init sched_init_smp(void) 6703{ 6704 sched_init_granularity(); 6705} 6706#endif /* CONFIG_SMP */ 6707 6708const_debug unsigned int sysctl_timer_migration = 1; 6709 6710int in_sched_functions(unsigned long addr) 6711{ 6712 return in_lock_functions(addr) || 6713 (addr >= (unsigned long)__sched_text_start 6714 && addr < (unsigned long)__sched_text_end); 6715} 6716 6717#ifdef CONFIG_CGROUP_SCHED 6718struct task_group root_task_group; 6719#endif 6720 6721DECLARE_PER_CPU(cpumask_var_t, load_balance_tmpmask); 6722 6723void __init sched_init(void) 6724{ 6725 int i, j; 6726 unsigned long alloc_size = 0, ptr; 6727 6728#ifdef CONFIG_FAIR_GROUP_SCHED 6729 alloc_size += 2 * nr_cpu_ids * sizeof(void **); 6730#endif 6731#ifdef CONFIG_RT_GROUP_SCHED 6732 alloc_size += 2 * nr_cpu_ids * sizeof(void **); 6733#endif 6734#ifdef CONFIG_CPUMASK_OFFSTACK 6735 alloc_size += num_possible_cpus() * cpumask_size(); 6736#endif 6737 if (alloc_size) { 6738 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT); 6739 6740#ifdef CONFIG_FAIR_GROUP_SCHED 6741 root_task_group.se = (struct sched_entity **)ptr; 6742 ptr += nr_cpu_ids * sizeof(void **); 6743 6744 root_task_group.cfs_rq = (struct cfs_rq **)ptr; 6745 ptr += nr_cpu_ids * sizeof(void **); 6746 6747#endif /* CONFIG_FAIR_GROUP_SCHED */ 6748#ifdef CONFIG_RT_GROUP_SCHED 6749 root_task_group.rt_se = (struct sched_rt_entity **)ptr; 6750 ptr += nr_cpu_ids * sizeof(void **); 6751 6752 root_task_group.rt_rq = (struct rt_rq **)ptr; 6753 ptr += nr_cpu_ids * sizeof(void **); 6754 6755#endif /* CONFIG_RT_GROUP_SCHED */ 6756#ifdef CONFIG_CPUMASK_OFFSTACK 6757 for_each_possible_cpu(i) { 6758 per_cpu(load_balance_tmpmask, i) = (void *)ptr; 6759 ptr += cpumask_size(); 6760 } 6761#endif /* CONFIG_CPUMASK_OFFSTACK */ 6762 } 6763 6764#ifdef CONFIG_SMP 6765 init_defrootdomain(); 6766#endif 6767 6768 init_rt_bandwidth(&def_rt_bandwidth, 6769 global_rt_period(), global_rt_runtime()); 6770 6771#ifdef CONFIG_RT_GROUP_SCHED 6772 init_rt_bandwidth(&root_task_group.rt_bandwidth, 6773 global_rt_period(), global_rt_runtime()); 6774#endif /* CONFIG_RT_GROUP_SCHED */ 6775 6776#ifdef CONFIG_CGROUP_SCHED 6777 list_add(&root_task_group.list, &task_groups); 6778 INIT_LIST_HEAD(&root_task_group.children); 6779 INIT_LIST_HEAD(&root_task_group.siblings); 6780 autogroup_init(&init_task); 6781#endif /* CONFIG_CGROUP_SCHED */ 6782 6783 for_each_possible_cpu(i) { 6784 struct rq *rq; 6785 6786 rq = cpu_rq(i); 6787 raw_spin_lock_init(&rq->lock); 6788 rq->nr_running = 0; 6789 rq->calc_load_active = 0; 6790 rq->calc_load_update = jiffies + LOAD_FREQ; 6791 init_cfs_rq(&rq->cfs); 6792 init_rt_rq(&rq->rt, rq); 6793#ifdef CONFIG_FAIR_GROUP_SCHED 6794 root_task_group.shares = ROOT_TASK_GROUP_LOAD; 6795 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list); 6796 /* 6797 * How much cpu bandwidth does root_task_group get? 6798 * 6799 * In case of task-groups formed thr' the cgroup filesystem, it 6800 * gets 100% of the cpu resources in the system. This overall 6801 * system cpu resource is divided among the tasks of 6802 * root_task_group and its child task-groups in a fair manner, 6803 * based on each entity's (task or task-group's) weight 6804 * (se->load.weight). 6805 * 6806 * In other words, if root_task_group has 10 tasks of weight 6807 * 1024) and two child groups A0 and A1 (of weight 1024 each), 6808 * then A0's share of the cpu resource is: 6809 * 6810 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33% 6811 * 6812 * We achieve this by letting root_task_group's tasks sit 6813 * directly in rq->cfs (i.e root_task_group->se[] = NULL). 6814 */ 6815 init_cfs_bandwidth(&root_task_group.cfs_bandwidth); 6816 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL); 6817#endif /* CONFIG_FAIR_GROUP_SCHED */ 6818 6819 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime; 6820#ifdef CONFIG_RT_GROUP_SCHED 6821 INIT_LIST_HEAD(&rq->leaf_rt_rq_list); 6822 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL); 6823#endif 6824 6825 for (j = 0; j < CPU_LOAD_IDX_MAX; j++) 6826 rq->cpu_load[j] = 0; 6827 6828 rq->last_load_update_tick = jiffies; 6829 6830#ifdef CONFIG_SMP 6831 rq->sd = NULL; 6832 rq->rd = NULL; 6833 rq->cpu_power = SCHED_POWER_SCALE; 6834 rq->post_schedule = 0; 6835 rq->active_balance = 0; 6836 rq->next_balance = jiffies; 6837 rq->push_cpu = 0; 6838 rq->cpu = i; 6839 rq->online = 0; 6840 rq->idle_stamp = 0; 6841 rq->avg_idle = 2*sysctl_sched_migration_cost; 6842 rq_attach_root(rq, &def_root_domain); 6843#ifdef CONFIG_NO_HZ 6844 rq->nohz_flags = 0; 6845#endif 6846#endif 6847 init_rq_hrtick(rq); 6848 atomic_set(&rq->nr_iowait, 0); 6849 } 6850 6851 set_load_weight(&init_task); 6852 6853#ifdef CONFIG_PREEMPT_NOTIFIERS 6854 INIT_HLIST_HEAD(&init_task.preempt_notifiers); 6855#endif 6856 6857#ifdef CONFIG_RT_MUTEXES 6858 plist_head_init(&init_task.pi_waiters); 6859#endif 6860 6861 /* 6862 * The boot idle thread does lazy MMU switching as well: 6863 */ 6864 atomic_inc(&init_mm.mm_count); 6865 enter_lazy_tlb(&init_mm, current); 6866 6867 /* 6868 * Make us the idle thread. Technically, schedule() should not be 6869 * called from this thread, however somewhere below it might be, 6870 * but because we are the idle thread, we just pick up running again 6871 * when this runqueue becomes "idle". 6872 */ 6873 init_idle(current, smp_processor_id()); 6874 6875 calc_load_update = jiffies + LOAD_FREQ; 6876 6877 /* 6878 * During early bootup we pretend to be a normal task: 6879 */ 6880 current->sched_class = &fair_sched_class; 6881 6882#ifdef CONFIG_SMP 6883 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT); 6884 /* May be allocated at isolcpus cmdline parse time */ 6885 if (cpu_isolated_map == NULL) 6886 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT); 6887#endif 6888 init_sched_fair_class(); 6889 6890 scheduler_running = 1; 6891} 6892 6893#ifdef CONFIG_DEBUG_ATOMIC_SLEEP 6894static inline int preempt_count_equals(int preempt_offset) 6895{ 6896 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth(); 6897 6898 return (nested == preempt_offset); 6899} 6900 6901void __might_sleep(const char *file, int line, int preempt_offset) 6902{ 6903 static unsigned long prev_jiffy; /* ratelimiting */ 6904 6905 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */ 6906 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) || 6907 system_state != SYSTEM_RUNNING || oops_in_progress) 6908 return; 6909 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy) 6910 return; 6911 prev_jiffy = jiffies; 6912 6913 printk(KERN_ERR 6914 "BUG: sleeping function called from invalid context at %s:%d\n", 6915 file, line); 6916 printk(KERN_ERR 6917 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n", 6918 in_atomic(), irqs_disabled(), 6919 current->pid, current->comm); 6920 6921 debug_show_held_locks(current); 6922 if (irqs_disabled()) 6923 print_irqtrace_events(current); 6924 dump_stack(); 6925} 6926EXPORT_SYMBOL(__might_sleep); 6927#endif 6928 6929#ifdef CONFIG_MAGIC_SYSRQ 6930static void normalize_task(struct rq *rq, struct task_struct *p) 6931{ 6932 const struct sched_class *prev_class = p->sched_class; 6933 int old_prio = p->prio; 6934 int on_rq; 6935 6936 on_rq = p->on_rq; 6937 if (on_rq) 6938 deactivate_task(rq, p, 0); 6939 __setscheduler(rq, p, SCHED_NORMAL, 0); 6940 if (on_rq) { 6941 activate_task(rq, p, 0); 6942 resched_task(rq->curr); 6943 } 6944 6945 check_class_changed(rq, p, prev_class, old_prio); 6946} 6947 6948void normalize_rt_tasks(void) 6949{ 6950 struct task_struct *g, *p; 6951 unsigned long flags; 6952 struct rq *rq; 6953 6954 read_lock_irqsave(&tasklist_lock, flags); 6955 do_each_thread(g, p) { 6956 /* 6957 * Only normalize user tasks: 6958 */ 6959 if (!p->mm) 6960 continue; 6961 6962 p->se.exec_start = 0; 6963#ifdef CONFIG_SCHEDSTATS 6964 p->se.statistics.wait_start = 0; 6965 p->se.statistics.sleep_start = 0; 6966 p->se.statistics.block_start = 0; 6967#endif 6968 6969 if (!rt_task(p)) { 6970 /* 6971 * Renice negative nice level userspace 6972 * tasks back to 0: 6973 */ 6974 if (TASK_NICE(p) < 0 && p->mm) 6975 set_user_nice(p, 0); 6976 continue; 6977 } 6978 6979 raw_spin_lock(&p->pi_lock); 6980 rq = __task_rq_lock(p); 6981 6982 normalize_task(rq, p); 6983 6984 __task_rq_unlock(rq); 6985 raw_spin_unlock(&p->pi_lock); 6986 } while_each_thread(g, p); 6987 6988 read_unlock_irqrestore(&tasklist_lock, flags); 6989} 6990 6991#endif /* CONFIG_MAGIC_SYSRQ */ 6992 6993#if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) 6994/* 6995 * These functions are only useful for the IA64 MCA handling, or kdb. 6996 * 6997 * They can only be called when the whole system has been 6998 * stopped - every CPU needs to be quiescent, and no scheduling 6999 * activity can take place. Using them for anything else would 7000 * be a serious bug, and as a result, they aren't even visible 7001 * under any other configuration. 7002 */ 7003 7004/** 7005 * curr_task - return the current task for a given cpu. 7006 * @cpu: the processor in question. 7007 * 7008 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED! 7009 */ 7010struct task_struct *curr_task(int cpu) 7011{ 7012 return cpu_curr(cpu); 7013} 7014 7015#endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */ 7016 7017#ifdef CONFIG_IA64 7018/** 7019 * set_curr_task - set the current task for a given cpu. 7020 * @cpu: the processor in question. 7021 * @p: the task pointer to set. 7022 * 7023 * Description: This function must only be used when non-maskable interrupts 7024 * are serviced on a separate stack. It allows the architecture to switch the 7025 * notion of the current task on a cpu in a non-blocking manner. This function 7026 * must be called with all CPU's synchronized, and interrupts disabled, the 7027 * and caller must save the original value of the current task (see 7028 * curr_task() above) and restore that value before reenabling interrupts and 7029 * re-starting the system. 7030 * 7031 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED! 7032 */ 7033void set_curr_task(int cpu, struct task_struct *p) 7034{ 7035 cpu_curr(cpu) = p; 7036} 7037 7038#endif 7039 7040#ifdef CONFIG_RT_GROUP_SCHED 7041#else /* !CONFIG_RT_GROUP_SCHED */ 7042#endif /* CONFIG_RT_GROUP_SCHED */ 7043 7044#ifdef CONFIG_CGROUP_SCHED 7045/* task_group_lock serializes the addition/removal of task groups */ 7046static DEFINE_SPINLOCK(task_group_lock); 7047 7048static void free_sched_group(struct task_group *tg) 7049{ 7050 free_fair_sched_group(tg); 7051 free_rt_sched_group(tg); 7052 autogroup_free(tg); 7053 kfree(tg); 7054} 7055 7056/* allocate runqueue etc for a new task group */ 7057struct task_group *sched_create_group(struct task_group *parent) 7058{ 7059 struct task_group *tg; 7060 unsigned long flags; 7061 7062 tg = kzalloc(sizeof(*tg), GFP_KERNEL); 7063 if (!tg) 7064 return ERR_PTR(-ENOMEM); 7065 7066 if (!alloc_fair_sched_group(tg, parent)) 7067 goto err; 7068 7069 if (!alloc_rt_sched_group(tg, parent)) 7070 goto err; 7071 7072 spin_lock_irqsave(&task_group_lock, flags); 7073 list_add_rcu(&tg->list, &task_groups); 7074 7075 WARN_ON(!parent); /* root should already exist */ 7076 7077 tg->parent = parent; 7078 INIT_LIST_HEAD(&tg->children); 7079 list_add_rcu(&tg->siblings, &parent->children); 7080 spin_unlock_irqrestore(&task_group_lock, flags); 7081 7082 return tg; 7083 7084err: 7085 free_sched_group(tg); 7086 return ERR_PTR(-ENOMEM); 7087} 7088 7089/* rcu callback to free various structures associated with a task group */ 7090static void free_sched_group_rcu(struct rcu_head *rhp) 7091{ 7092 /* now it should be safe to free those cfs_rqs */ 7093 free_sched_group(container_of(rhp, struct task_group, rcu)); 7094} 7095 7096/* Destroy runqueue etc associated with a task group */ 7097void sched_destroy_group(struct task_group *tg) 7098{ 7099 unsigned long flags; 7100 int i; 7101 7102 /* end participation in shares distribution */ 7103 for_each_possible_cpu(i) 7104 unregister_fair_sched_group(tg, i); 7105 7106 spin_lock_irqsave(&task_group_lock, flags); 7107 list_del_rcu(&tg->list); 7108 list_del_rcu(&tg->siblings); 7109 spin_unlock_irqrestore(&task_group_lock, flags); 7110 7111 /* wait for possible concurrent references to cfs_rqs complete */ 7112 call_rcu(&tg->rcu, free_sched_group_rcu); 7113} 7114 7115/* change task's runqueue when it moves between groups. 7116 * The caller of this function should have put the task in its new group 7117 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to 7118 * reflect its new group. 7119 */ 7120void sched_move_task(struct task_struct *tsk) 7121{ 7122 int on_rq, running; 7123 unsigned long flags; 7124 struct rq *rq; 7125 7126 rq = task_rq_lock(tsk, &flags); 7127 7128 running = task_current(rq, tsk); 7129 on_rq = tsk->on_rq; 7130 7131 if (on_rq) 7132 dequeue_task(rq, tsk, 0); 7133 if (unlikely(running)) 7134 tsk->sched_class->put_prev_task(rq, tsk); 7135 7136#ifdef CONFIG_FAIR_GROUP_SCHED 7137 if (tsk->sched_class->task_move_group) 7138 tsk->sched_class->task_move_group(tsk, on_rq); 7139 else 7140#endif 7141 set_task_rq(tsk, task_cpu(tsk)); 7142 7143 if (unlikely(running)) 7144 tsk->sched_class->set_curr_task(rq); 7145 if (on_rq) 7146 enqueue_task(rq, tsk, 0); 7147 7148 task_rq_unlock(rq, tsk, &flags); 7149} 7150#endif /* CONFIG_CGROUP_SCHED */ 7151 7152#ifdef CONFIG_FAIR_GROUP_SCHED 7153#endif 7154 7155#if defined(CONFIG_RT_GROUP_SCHED) || defined(CONFIG_CFS_BANDWIDTH) 7156static unsigned long to_ratio(u64 period, u64 runtime) 7157{ 7158 if (runtime == RUNTIME_INF) 7159 return 1ULL << 20; 7160 7161 return div64_u64(runtime << 20, period); 7162} 7163#endif 7164 7165#ifdef CONFIG_RT_GROUP_SCHED 7166/* 7167 * Ensure that the real time constraints are schedulable. 7168 */ 7169static DEFINE_MUTEX(rt_constraints_mutex); 7170 7171/* Must be called with tasklist_lock held */ 7172static inline int tg_has_rt_tasks(struct task_group *tg) 7173{ 7174 struct task_struct *g, *p; 7175 7176 do_each_thread(g, p) { 7177 if (rt_task(p) && task_rq(p)->rt.tg == tg) 7178 return 1; 7179 } while_each_thread(g, p); 7180 7181 return 0; 7182} 7183 7184struct rt_schedulable_data { 7185 struct task_group *tg; 7186 u64 rt_period; 7187 u64 rt_runtime; 7188}; 7189 7190static int tg_rt_schedulable(struct task_group *tg, void *data) 7191{ 7192 struct rt_schedulable_data *d = data; 7193 struct task_group *child; 7194 unsigned long total, sum = 0; 7195 u64 period, runtime; 7196 7197 period = ktime_to_ns(tg->rt_bandwidth.rt_period); 7198 runtime = tg->rt_bandwidth.rt_runtime; 7199 7200 if (tg == d->tg) { 7201 period = d->rt_period; 7202 runtime = d->rt_runtime; 7203 } 7204 7205 /* 7206 * Cannot have more runtime than the period. 7207 */ 7208 if (runtime > period && runtime != RUNTIME_INF) 7209 return -EINVAL; 7210 7211 /* 7212 * Ensure we don't starve existing RT tasks. 7213 */ 7214 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg)) 7215 return -EBUSY; 7216 7217 total = to_ratio(period, runtime); 7218 7219 /* 7220 * Nobody can have more than the global setting allows. 7221 */ 7222 if (total > to_ratio(global_rt_period(), global_rt_runtime())) 7223 return -EINVAL; 7224 7225 /* 7226 * The sum of our children's runtime should not exceed our own. 7227 */ 7228 list_for_each_entry_rcu(child, &tg->children, siblings) { 7229 period = ktime_to_ns(child->rt_bandwidth.rt_period); 7230 runtime = child->rt_bandwidth.rt_runtime; 7231 7232 if (child == d->tg) { 7233 period = d->rt_period; 7234 runtime = d->rt_runtime; 7235 } 7236 7237 sum += to_ratio(period, runtime); 7238 } 7239 7240 if (sum > total) 7241 return -EINVAL; 7242 7243 return 0; 7244} 7245 7246static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime) 7247{ 7248 int ret; 7249 7250 struct rt_schedulable_data data = { 7251 .tg = tg, 7252 .rt_period = period, 7253 .rt_runtime = runtime, 7254 }; 7255 7256 rcu_read_lock(); 7257 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data); 7258 rcu_read_unlock(); 7259 7260 return ret; 7261} 7262 7263static int tg_set_rt_bandwidth(struct task_group *tg, 7264 u64 rt_period, u64 rt_runtime) 7265{ 7266 int i, err = 0; 7267 7268 mutex_lock(&rt_constraints_mutex); 7269 read_lock(&tasklist_lock); 7270 err = __rt_schedulable(tg, rt_period, rt_runtime); 7271 if (err) 7272 goto unlock; 7273 7274 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock); 7275 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period); 7276 tg->rt_bandwidth.rt_runtime = rt_runtime; 7277 7278 for_each_possible_cpu(i) { 7279 struct rt_rq *rt_rq = tg->rt_rq[i]; 7280 7281 raw_spin_lock(&rt_rq->rt_runtime_lock); 7282 rt_rq->rt_runtime = rt_runtime; 7283 raw_spin_unlock(&rt_rq->rt_runtime_lock); 7284 } 7285 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock); 7286unlock: 7287 read_unlock(&tasklist_lock); 7288 mutex_unlock(&rt_constraints_mutex); 7289 7290 return err; 7291} 7292 7293int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us) 7294{ 7295 u64 rt_runtime, rt_period; 7296 7297 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period); 7298 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC; 7299 if (rt_runtime_us < 0) 7300 rt_runtime = RUNTIME_INF; 7301 7302 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime); 7303} 7304 7305long sched_group_rt_runtime(struct task_group *tg) 7306{ 7307 u64 rt_runtime_us; 7308 7309 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF) 7310 return -1; 7311 7312 rt_runtime_us = tg->rt_bandwidth.rt_runtime; 7313 do_div(rt_runtime_us, NSEC_PER_USEC); 7314 return rt_runtime_us; 7315} 7316 7317int sched_group_set_rt_period(struct task_group *tg, long rt_period_us) 7318{ 7319 u64 rt_runtime, rt_period; 7320 7321 rt_period = (u64)rt_period_us * NSEC_PER_USEC; 7322 rt_runtime = tg->rt_bandwidth.rt_runtime; 7323 7324 if (rt_period == 0) 7325 return -EINVAL; 7326 7327 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime); 7328} 7329 7330long sched_group_rt_period(struct task_group *tg) 7331{ 7332 u64 rt_period_us; 7333 7334 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period); 7335 do_div(rt_period_us, NSEC_PER_USEC); 7336 return rt_period_us; 7337} 7338 7339static int sched_rt_global_constraints(void) 7340{ 7341 u64 runtime, period; 7342 int ret = 0; 7343 7344 if (sysctl_sched_rt_period <= 0) 7345 return -EINVAL; 7346 7347 runtime = global_rt_runtime(); 7348 period = global_rt_period(); 7349 7350 /* 7351 * Sanity check on the sysctl variables. 7352 */ 7353 if (runtime > period && runtime != RUNTIME_INF) 7354 return -EINVAL; 7355 7356 mutex_lock(&rt_constraints_mutex); 7357 read_lock(&tasklist_lock); 7358 ret = __rt_schedulable(NULL, 0, 0); 7359 read_unlock(&tasklist_lock); 7360 mutex_unlock(&rt_constraints_mutex); 7361 7362 return ret; 7363} 7364 7365int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk) 7366{ 7367 /* Don't accept realtime tasks when there is no way for them to run */ 7368 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0) 7369 return 0; 7370 7371 return 1; 7372} 7373 7374#else /* !CONFIG_RT_GROUP_SCHED */ 7375static int sched_rt_global_constraints(void) 7376{ 7377 unsigned long flags; 7378 int i; 7379 7380 if (sysctl_sched_rt_period <= 0) 7381 return -EINVAL; 7382 7383 /* 7384 * There's always some RT tasks in the root group 7385 * -- migration, kstopmachine etc.. 7386 */ 7387 if (sysctl_sched_rt_runtime == 0) 7388 return -EBUSY; 7389 7390 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags); 7391 for_each_possible_cpu(i) { 7392 struct rt_rq *rt_rq = &cpu_rq(i)->rt; 7393 7394 raw_spin_lock(&rt_rq->rt_runtime_lock); 7395 rt_rq->rt_runtime = global_rt_runtime(); 7396 raw_spin_unlock(&rt_rq->rt_runtime_lock); 7397 } 7398 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags); 7399 7400 return 0; 7401} 7402#endif /* CONFIG_RT_GROUP_SCHED */ 7403 7404int sched_rt_handler(struct ctl_table *table, int write, 7405 void __user *buffer, size_t *lenp, 7406 loff_t *ppos) 7407{ 7408 int ret; 7409 int old_period, old_runtime; 7410 static DEFINE_MUTEX(mutex); 7411 7412 mutex_lock(&mutex); 7413 old_period = sysctl_sched_rt_period; 7414 old_runtime = sysctl_sched_rt_runtime; 7415 7416 ret = proc_dointvec(table, write, buffer, lenp, ppos); 7417 7418 if (!ret && write) { 7419 ret = sched_rt_global_constraints(); 7420 if (ret) { 7421 sysctl_sched_rt_period = old_period; 7422 sysctl_sched_rt_runtime = old_runtime; 7423 } else { 7424 def_rt_bandwidth.rt_runtime = global_rt_runtime(); 7425 def_rt_bandwidth.rt_period = 7426 ns_to_ktime(global_rt_period()); 7427 } 7428 } 7429 mutex_unlock(&mutex); 7430 7431 return ret; 7432} 7433 7434#ifdef CONFIG_CGROUP_SCHED 7435 7436/* return corresponding task_group object of a cgroup */ 7437static inline struct task_group *cgroup_tg(struct cgroup *cgrp) 7438{ 7439 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id), 7440 struct task_group, css); 7441} 7442 7443static struct cgroup_subsys_state * 7444cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp) 7445{ 7446 struct task_group *tg, *parent; 7447 7448 if (!cgrp->parent) { 7449 /* This is early initialization for the top cgroup */ 7450 return &root_task_group.css; 7451 } 7452 7453 parent = cgroup_tg(cgrp->parent); 7454 tg = sched_create_group(parent); 7455 if (IS_ERR(tg)) 7456 return ERR_PTR(-ENOMEM); 7457 7458 return &tg->css; 7459} 7460 7461static void 7462cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp) 7463{ 7464 struct task_group *tg = cgroup_tg(cgrp); 7465 7466 sched_destroy_group(tg); 7467} 7468 7469static int 7470cpu_cgroup_can_attach_task(struct cgroup *cgrp, struct task_struct *tsk) 7471{ 7472#ifdef CONFIG_RT_GROUP_SCHED 7473 if (!sched_rt_can_attach(cgroup_tg(cgrp), tsk)) 7474 return -EINVAL; 7475#else 7476 /* We don't support RT-tasks being in separate groups */ 7477 if (tsk->sched_class != &fair_sched_class) 7478 return -EINVAL; 7479#endif 7480 return 0; 7481} 7482 7483static void 7484cpu_cgroup_attach_task(struct cgroup *cgrp, struct task_struct *tsk) 7485{ 7486 sched_move_task(tsk); 7487} 7488 7489static void 7490cpu_cgroup_exit(struct cgroup_subsys *ss, struct cgroup *cgrp, 7491 struct cgroup *old_cgrp, struct task_struct *task) 7492{ 7493 /* 7494 * cgroup_exit() is called in the copy_process() failure path. 7495 * Ignore this case since the task hasn't ran yet, this avoids 7496 * trying to poke a half freed task state from generic code. 7497 */ 7498 if (!(task->flags & PF_EXITING)) 7499 return; 7500 7501 sched_move_task(task); 7502} 7503 7504#ifdef CONFIG_FAIR_GROUP_SCHED 7505static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype, 7506 u64 shareval) 7507{ 7508 return sched_group_set_shares(cgroup_tg(cgrp), scale_load(shareval)); 7509} 7510 7511static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft) 7512{ 7513 struct task_group *tg = cgroup_tg(cgrp); 7514 7515 return (u64) scale_load_down(tg->shares); 7516} 7517 7518#ifdef CONFIG_CFS_BANDWIDTH 7519static DEFINE_MUTEX(cfs_constraints_mutex); 7520 7521const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */ 7522const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */ 7523 7524static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime); 7525 7526static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota) 7527{ 7528 int i, ret = 0, runtime_enabled, runtime_was_enabled; 7529 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth; 7530 7531 if (tg == &root_task_group) 7532 return -EINVAL; 7533 7534 /* 7535 * Ensure we have at some amount of bandwidth every period. This is 7536 * to prevent reaching a state of large arrears when throttled via 7537 * entity_tick() resulting in prolonged exit starvation. 7538 */ 7539 if (quota < min_cfs_quota_period || period < min_cfs_quota_period) 7540 return -EINVAL; 7541 7542 /* 7543 * Likewise, bound things on the otherside by preventing insane quota 7544 * periods. This also allows us to normalize in computing quota 7545 * feasibility. 7546 */ 7547 if (period > max_cfs_quota_period) 7548 return -EINVAL; 7549 7550 mutex_lock(&cfs_constraints_mutex); 7551 ret = __cfs_schedulable(tg, period, quota); 7552 if (ret) 7553 goto out_unlock; 7554 7555 runtime_enabled = quota != RUNTIME_INF; 7556 runtime_was_enabled = cfs_b->quota != RUNTIME_INF; 7557 account_cfs_bandwidth_used(runtime_enabled, runtime_was_enabled); 7558 raw_spin_lock_irq(&cfs_b->lock); 7559 cfs_b->period = ns_to_ktime(period); 7560 cfs_b->quota = quota; 7561 7562 __refill_cfs_bandwidth_runtime(cfs_b); 7563 /* restart the period timer (if active) to handle new period expiry */ 7564 if (runtime_enabled && cfs_b->timer_active) { 7565 /* force a reprogram */ 7566 cfs_b->timer_active = 0; 7567 __start_cfs_bandwidth(cfs_b); 7568 } 7569 raw_spin_unlock_irq(&cfs_b->lock); 7570 7571 for_each_possible_cpu(i) { 7572 struct cfs_rq *cfs_rq = tg->cfs_rq[i]; 7573 struct rq *rq = cfs_rq->rq; 7574 7575 raw_spin_lock_irq(&rq->lock); 7576 cfs_rq->runtime_enabled = runtime_enabled; 7577 cfs_rq->runtime_remaining = 0; 7578 7579 if (cfs_rq->throttled) 7580 unthrottle_cfs_rq(cfs_rq); 7581 raw_spin_unlock_irq(&rq->lock); 7582 } 7583out_unlock: 7584 mutex_unlock(&cfs_constraints_mutex); 7585 7586 return ret; 7587} 7588 7589int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us) 7590{ 7591 u64 quota, period; 7592 7593 period = ktime_to_ns(tg->cfs_bandwidth.period); 7594 if (cfs_quota_us < 0) 7595 quota = RUNTIME_INF; 7596 else 7597 quota = (u64)cfs_quota_us * NSEC_PER_USEC; 7598 7599 return tg_set_cfs_bandwidth(tg, period, quota); 7600} 7601 7602long tg_get_cfs_quota(struct task_group *tg) 7603{ 7604 u64 quota_us; 7605 7606 if (tg->cfs_bandwidth.quota == RUNTIME_INF) 7607 return -1; 7608 7609 quota_us = tg->cfs_bandwidth.quota; 7610 do_div(quota_us, NSEC_PER_USEC); 7611 7612 return quota_us; 7613} 7614 7615int tg_set_cfs_period(struct task_group *tg, long cfs_period_us) 7616{ 7617 u64 quota, period; 7618 7619 period = (u64)cfs_period_us * NSEC_PER_USEC; 7620 quota = tg->cfs_bandwidth.quota; 7621 7622 if (period <= 0) 7623 return -EINVAL; 7624 7625 return tg_set_cfs_bandwidth(tg, period, quota); 7626} 7627 7628long tg_get_cfs_period(struct task_group *tg) 7629{ 7630 u64 cfs_period_us; 7631 7632 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period); 7633 do_div(cfs_period_us, NSEC_PER_USEC); 7634 7635 return cfs_period_us; 7636} 7637 7638static s64 cpu_cfs_quota_read_s64(struct cgroup *cgrp, struct cftype *cft) 7639{ 7640 return tg_get_cfs_quota(cgroup_tg(cgrp)); 7641} 7642 7643static int cpu_cfs_quota_write_s64(struct cgroup *cgrp, struct cftype *cftype, 7644 s64 cfs_quota_us) 7645{ 7646 return tg_set_cfs_quota(cgroup_tg(cgrp), cfs_quota_us); 7647} 7648 7649static u64 cpu_cfs_period_read_u64(struct cgroup *cgrp, struct cftype *cft) 7650{ 7651 return tg_get_cfs_period(cgroup_tg(cgrp)); 7652} 7653 7654static int cpu_cfs_period_write_u64(struct cgroup *cgrp, struct cftype *cftype, 7655 u64 cfs_period_us) 7656{ 7657 return tg_set_cfs_period(cgroup_tg(cgrp), cfs_period_us); 7658} 7659 7660struct cfs_schedulable_data { 7661 struct task_group *tg; 7662 u64 period, quota; 7663}; 7664 7665/* 7666 * normalize group quota/period to be quota/max_period 7667 * note: units are usecs 7668 */ 7669static u64 normalize_cfs_quota(struct task_group *tg, 7670 struct cfs_schedulable_data *d) 7671{ 7672 u64 quota, period; 7673 7674 if (tg == d->tg) { 7675 period = d->period; 7676 quota = d->quota; 7677 } else { 7678 period = tg_get_cfs_period(tg); 7679 quota = tg_get_cfs_quota(tg); 7680 } 7681 7682 /* note: these should typically be equivalent */ 7683 if (quota == RUNTIME_INF || quota == -1) 7684 return RUNTIME_INF; 7685 7686 return to_ratio(period, quota); 7687} 7688 7689static int tg_cfs_schedulable_down(struct task_group *tg, void *data) 7690{ 7691 struct cfs_schedulable_data *d = data; 7692 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth; 7693 s64 quota = 0, parent_quota = -1; 7694 7695 if (!tg->parent) { 7696 quota = RUNTIME_INF; 7697 } else { 7698 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth; 7699 7700 quota = normalize_cfs_quota(tg, d); 7701 parent_quota = parent_b->hierarchal_quota; 7702 7703 /* 7704 * ensure max(child_quota) <= parent_quota, inherit when no 7705 * limit is set 7706 */ 7707 if (quota == RUNTIME_INF) 7708 quota = parent_quota; 7709 else if (parent_quota != RUNTIME_INF && quota > parent_quota) 7710 return -EINVAL; 7711 } 7712 cfs_b->hierarchal_quota = quota; 7713 7714 return 0; 7715} 7716 7717static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota) 7718{ 7719 int ret; 7720 struct cfs_schedulable_data data = { 7721 .tg = tg, 7722 .period = period, 7723 .quota = quota, 7724 }; 7725 7726 if (quota != RUNTIME_INF) { 7727 do_div(data.period, NSEC_PER_USEC); 7728 do_div(data.quota, NSEC_PER_USEC); 7729 } 7730 7731 rcu_read_lock(); 7732 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data); 7733 rcu_read_unlock(); 7734 7735 return ret; 7736} 7737 7738static int cpu_stats_show(struct cgroup *cgrp, struct cftype *cft, 7739 struct cgroup_map_cb *cb) 7740{ 7741 struct task_group *tg = cgroup_tg(cgrp); 7742 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth; 7743 7744 cb->fill(cb, "nr_periods", cfs_b->nr_periods); 7745 cb->fill(cb, "nr_throttled", cfs_b->nr_throttled); 7746 cb->fill(cb, "throttled_time", cfs_b->throttled_time); 7747 7748 return 0; 7749} 7750#endif /* CONFIG_CFS_BANDWIDTH */ 7751#endif /* CONFIG_FAIR_GROUP_SCHED */ 7752 7753#ifdef CONFIG_RT_GROUP_SCHED 7754static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft, 7755 s64 val) 7756{ 7757 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val); 7758} 7759 7760static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft) 7761{ 7762 return sched_group_rt_runtime(cgroup_tg(cgrp)); 7763} 7764 7765static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype, 7766 u64 rt_period_us) 7767{ 7768 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us); 7769} 7770 7771static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft) 7772{ 7773 return sched_group_rt_period(cgroup_tg(cgrp)); 7774} 7775#endif /* CONFIG_RT_GROUP_SCHED */ 7776 7777static struct cftype cpu_files[] = { 7778#ifdef CONFIG_FAIR_GROUP_SCHED 7779 { 7780 .name = "shares", 7781 .read_u64 = cpu_shares_read_u64, 7782 .write_u64 = cpu_shares_write_u64, 7783 }, 7784#endif 7785#ifdef CONFIG_CFS_BANDWIDTH 7786 { 7787 .name = "cfs_quota_us", 7788 .read_s64 = cpu_cfs_quota_read_s64, 7789 .write_s64 = cpu_cfs_quota_write_s64, 7790 }, 7791 { 7792 .name = "cfs_period_us", 7793 .read_u64 = cpu_cfs_period_read_u64, 7794 .write_u64 = cpu_cfs_period_write_u64, 7795 }, 7796 { 7797 .name = "stat", 7798 .read_map = cpu_stats_show, 7799 }, 7800#endif 7801#ifdef CONFIG_RT_GROUP_SCHED 7802 { 7803 .name = "rt_runtime_us", 7804 .read_s64 = cpu_rt_runtime_read, 7805 .write_s64 = cpu_rt_runtime_write, 7806 }, 7807 { 7808 .name = "rt_period_us", 7809 .read_u64 = cpu_rt_period_read_uint, 7810 .write_u64 = cpu_rt_period_write_uint, 7811 }, 7812#endif 7813}; 7814 7815static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont) 7816{ 7817 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files)); 7818} 7819 7820struct cgroup_subsys cpu_cgroup_subsys = { 7821 .name = "cpu", 7822 .create = cpu_cgroup_create, 7823 .destroy = cpu_cgroup_destroy, 7824 .can_attach_task = cpu_cgroup_can_attach_task, 7825 .attach_task = cpu_cgroup_attach_task, 7826 .exit = cpu_cgroup_exit, 7827 .populate = cpu_cgroup_populate, 7828 .subsys_id = cpu_cgroup_subsys_id, 7829 .early_init = 1, 7830}; 7831 7832#endif /* CONFIG_CGROUP_SCHED */ 7833 7834#ifdef CONFIG_CGROUP_CPUACCT 7835 7836/* 7837 * CPU accounting code for task groups. 7838 * 7839 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh 7840 * (balbir@in.ibm.com). 7841 */ 7842 7843/* track cpu usage of a group of tasks and its child groups */ 7844struct cpuacct { 7845 struct cgroup_subsys_state css; 7846 /* cpuusage holds pointer to a u64-type object on every cpu */ 7847 u64 __percpu *cpuusage; 7848 struct percpu_counter cpustat[CPUACCT_STAT_NSTATS]; 7849 struct cpuacct *parent; 7850}; 7851 7852struct cgroup_subsys cpuacct_subsys; 7853 7854/* return cpu accounting group corresponding to this container */ 7855static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp) 7856{ 7857 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id), 7858 struct cpuacct, css); 7859} 7860 7861/* return cpu accounting group to which this task belongs */ 7862static inline struct cpuacct *task_ca(struct task_struct *tsk) 7863{ 7864 return container_of(task_subsys_state(tsk, cpuacct_subsys_id), 7865 struct cpuacct, css); 7866} 7867 7868/* create a new cpu accounting group */ 7869static struct cgroup_subsys_state *cpuacct_create( 7870 struct cgroup_subsys *ss, struct cgroup *cgrp) 7871{ 7872 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL); 7873 int i; 7874 7875 if (!ca) 7876 goto out; 7877 7878 ca->cpuusage = alloc_percpu(u64); 7879 if (!ca->cpuusage) 7880 goto out_free_ca; 7881 7882 for (i = 0; i < CPUACCT_STAT_NSTATS; i++) 7883 if (percpu_counter_init(&ca->cpustat[i], 0)) 7884 goto out_free_counters; 7885 7886 if (cgrp->parent) 7887 ca->parent = cgroup_ca(cgrp->parent); 7888 7889 return &ca->css; 7890 7891out_free_counters: 7892 while (--i >= 0) 7893 percpu_counter_destroy(&ca->cpustat[i]); 7894 free_percpu(ca->cpuusage); 7895out_free_ca: 7896 kfree(ca); 7897out: 7898 return ERR_PTR(-ENOMEM); 7899} 7900 7901/* destroy an existing cpu accounting group */ 7902static void 7903cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp) 7904{ 7905 struct cpuacct *ca = cgroup_ca(cgrp); 7906 int i; 7907 7908 for (i = 0; i < CPUACCT_STAT_NSTATS; i++) 7909 percpu_counter_destroy(&ca->cpustat[i]); 7910 free_percpu(ca->cpuusage); 7911 kfree(ca); 7912} 7913 7914static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu) 7915{ 7916 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu); 7917 u64 data; 7918 7919#ifndef CONFIG_64BIT 7920 /* 7921 * Take rq->lock to make 64-bit read safe on 32-bit platforms. 7922 */ 7923 raw_spin_lock_irq(&cpu_rq(cpu)->lock); 7924 data = *cpuusage; 7925 raw_spin_unlock_irq(&cpu_rq(cpu)->lock); 7926#else 7927 data = *cpuusage; 7928#endif 7929 7930 return data; 7931} 7932 7933static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val) 7934{ 7935 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu); 7936 7937#ifndef CONFIG_64BIT 7938 /* 7939 * Take rq->lock to make 64-bit write safe on 32-bit platforms. 7940 */ 7941 raw_spin_lock_irq(&cpu_rq(cpu)->lock); 7942 *cpuusage = val; 7943 raw_spin_unlock_irq(&cpu_rq(cpu)->lock); 7944#else 7945 *cpuusage = val; 7946#endif 7947} 7948 7949/* return total cpu usage (in nanoseconds) of a group */ 7950static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft) 7951{ 7952 struct cpuacct *ca = cgroup_ca(cgrp); 7953 u64 totalcpuusage = 0; 7954 int i; 7955 7956 for_each_present_cpu(i) 7957 totalcpuusage += cpuacct_cpuusage_read(ca, i); 7958 7959 return totalcpuusage; 7960} 7961 7962static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype, 7963 u64 reset) 7964{ 7965 struct cpuacct *ca = cgroup_ca(cgrp); 7966 int err = 0; 7967 int i; 7968 7969 if (reset) { 7970 err = -EINVAL; 7971 goto out; 7972 } 7973 7974 for_each_present_cpu(i) 7975 cpuacct_cpuusage_write(ca, i, 0); 7976 7977out: 7978 return err; 7979} 7980 7981static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft, 7982 struct seq_file *m) 7983{ 7984 struct cpuacct *ca = cgroup_ca(cgroup); 7985 u64 percpu; 7986 int i; 7987 7988 for_each_present_cpu(i) { 7989 percpu = cpuacct_cpuusage_read(ca, i); 7990 seq_printf(m, "%llu ", (unsigned long long) percpu); 7991 } 7992 seq_printf(m, "\n"); 7993 return 0; 7994} 7995 7996static const char *cpuacct_stat_desc[] = { 7997 [CPUACCT_STAT_USER] = "user", 7998 [CPUACCT_STAT_SYSTEM] = "system", 7999}; 8000 8001static int cpuacct_stats_show(struct cgroup *cgrp, struct cftype *cft, 8002 struct cgroup_map_cb *cb) 8003{ 8004 struct cpuacct *ca = cgroup_ca(cgrp); 8005 int i; 8006 8007 for (i = 0; i < CPUACCT_STAT_NSTATS; i++) { 8008 s64 val = percpu_counter_read(&ca->cpustat[i]); 8009 val = cputime64_to_clock_t(val); 8010 cb->fill(cb, cpuacct_stat_desc[i], val); 8011 } 8012 return 0; 8013} 8014 8015static struct cftype files[] = { 8016 { 8017 .name = "usage", 8018 .read_u64 = cpuusage_read, 8019 .write_u64 = cpuusage_write, 8020 }, 8021 { 8022 .name = "usage_percpu", 8023 .read_seq_string = cpuacct_percpu_seq_read, 8024 }, 8025 { 8026 .name = "stat", 8027 .read_map = cpuacct_stats_show, 8028 }, 8029}; 8030 8031static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp) 8032{ 8033 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files)); 8034} 8035 8036/* 8037 * charge this task's execution time to its accounting group. 8038 * 8039 * called with rq->lock held. 8040 */ 8041void cpuacct_charge(struct task_struct *tsk, u64 cputime) 8042{ 8043 struct cpuacct *ca; 8044 int cpu; 8045 8046 if (unlikely(!cpuacct_subsys.active)) 8047 return; 8048 8049 cpu = task_cpu(tsk); 8050 8051 rcu_read_lock(); 8052 8053 ca = task_ca(tsk); 8054 8055 for (; ca; ca = ca->parent) { 8056 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu); 8057 *cpuusage += cputime; 8058 } 8059 8060 rcu_read_unlock(); 8061} 8062 8063/* 8064 * When CONFIG_VIRT_CPU_ACCOUNTING is enabled one jiffy can be very large 8065 * in cputime_t units. As a result, cpuacct_update_stats calls 8066 * percpu_counter_add with values large enough to always overflow the 8067 * per cpu batch limit causing bad SMP scalability. 8068 * 8069 * To fix this we scale percpu_counter_batch by cputime_one_jiffy so we 8070 * batch the same amount of time with CONFIG_VIRT_CPU_ACCOUNTING disabled 8071 * and enabled. We cap it at INT_MAX which is the largest allowed batch value. 8072 */ 8073#ifdef CONFIG_SMP 8074#define CPUACCT_BATCH \ 8075 min_t(long, percpu_counter_batch * cputime_one_jiffy, INT_MAX) 8076#else 8077#define CPUACCT_BATCH 0 8078#endif 8079 8080/* 8081 * Charge the system/user time to the task's accounting group. 8082 */ 8083void cpuacct_update_stats(struct task_struct *tsk, 8084 enum cpuacct_stat_index idx, cputime_t val) 8085{ 8086 struct cpuacct *ca; 8087 int batch = CPUACCT_BATCH; 8088 8089 if (unlikely(!cpuacct_subsys.active)) 8090 return; 8091 8092 rcu_read_lock(); 8093 ca = task_ca(tsk); 8094 8095 do { 8096 __percpu_counter_add(&ca->cpustat[idx], val, batch); 8097 ca = ca->parent; 8098 } while (ca); 8099 rcu_read_unlock(); 8100} 8101 8102struct cgroup_subsys cpuacct_subsys = { 8103 .name = "cpuacct", 8104 .create = cpuacct_create, 8105 .destroy = cpuacct_destroy, 8106 .populate = cpuacct_populate, 8107 .subsys_id = cpuacct_subsys_id, 8108}; 8109#endif /* CONFIG_CGROUP_CPUACCT */ 8110