core.c revision 20ab65e33f469c35f3dabde3445b668aa9c943ee
1/* 2 * kernel/sched/core.c 3 * 4 * Kernel scheduler and related syscalls 5 * 6 * Copyright (C) 1991-2002 Linus Torvalds 7 * 8 * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and 9 * make semaphores SMP safe 10 * 1998-11-19 Implemented schedule_timeout() and related stuff 11 * by Andrea Arcangeli 12 * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar: 13 * hybrid priority-list and round-robin design with 14 * an array-switch method of distributing timeslices 15 * and per-CPU runqueues. Cleanups and useful suggestions 16 * by Davide Libenzi, preemptible kernel bits by Robert Love. 17 * 2003-09-03 Interactivity tuning by Con Kolivas. 18 * 2004-04-02 Scheduler domains code by Nick Piggin 19 * 2007-04-15 Work begun on replacing all interactivity tuning with a 20 * fair scheduling design by Con Kolivas. 21 * 2007-05-05 Load balancing (smp-nice) and other improvements 22 * by Peter Williams 23 * 2007-05-06 Interactivity improvements to CFS by Mike Galbraith 24 * 2007-07-01 Group scheduling enhancements by Srivatsa Vaddagiri 25 * 2007-11-29 RT balancing improvements by Steven Rostedt, Gregory Haskins, 26 * Thomas Gleixner, Mike Kravetz 27 */ 28 29#include <linux/mm.h> 30#include <linux/module.h> 31#include <linux/nmi.h> 32#include <linux/init.h> 33#include <linux/uaccess.h> 34#include <linux/highmem.h> 35#include <asm/mmu_context.h> 36#include <linux/interrupt.h> 37#include <linux/capability.h> 38#include <linux/completion.h> 39#include <linux/kernel_stat.h> 40#include <linux/debug_locks.h> 41#include <linux/perf_event.h> 42#include <linux/security.h> 43#include <linux/notifier.h> 44#include <linux/profile.h> 45#include <linux/freezer.h> 46#include <linux/vmalloc.h> 47#include <linux/blkdev.h> 48#include <linux/delay.h> 49#include <linux/pid_namespace.h> 50#include <linux/smp.h> 51#include <linux/threads.h> 52#include <linux/timer.h> 53#include <linux/rcupdate.h> 54#include <linux/cpu.h> 55#include <linux/cpuset.h> 56#include <linux/percpu.h> 57#include <linux/proc_fs.h> 58#include <linux/seq_file.h> 59#include <linux/sysctl.h> 60#include <linux/syscalls.h> 61#include <linux/times.h> 62#include <linux/tsacct_kern.h> 63#include <linux/kprobes.h> 64#include <linux/delayacct.h> 65#include <linux/unistd.h> 66#include <linux/pagemap.h> 67#include <linux/hrtimer.h> 68#include <linux/tick.h> 69#include <linux/debugfs.h> 70#include <linux/ctype.h> 71#include <linux/ftrace.h> 72#include <linux/slab.h> 73#include <linux/init_task.h> 74#include <linux/binfmts.h> 75 76#include <asm/switch_to.h> 77#include <asm/tlb.h> 78#include <asm/irq_regs.h> 79#include <asm/mutex.h> 80#ifdef CONFIG_PARAVIRT 81#include <asm/paravirt.h> 82#endif 83 84#include "sched.h" 85#include "../workqueue_sched.h" 86#include "../smpboot.h" 87 88#define CREATE_TRACE_POINTS 89#include <trace/events/sched.h> 90 91void start_bandwidth_timer(struct hrtimer *period_timer, ktime_t period) 92{ 93 unsigned long delta; 94 ktime_t soft, hard, now; 95 96 for (;;) { 97 if (hrtimer_active(period_timer)) 98 break; 99 100 now = hrtimer_cb_get_time(period_timer); 101 hrtimer_forward(period_timer, now, period); 102 103 soft = hrtimer_get_softexpires(period_timer); 104 hard = hrtimer_get_expires(period_timer); 105 delta = ktime_to_ns(ktime_sub(hard, soft)); 106 __hrtimer_start_range_ns(period_timer, soft, delta, 107 HRTIMER_MODE_ABS_PINNED, 0); 108 } 109} 110 111DEFINE_MUTEX(sched_domains_mutex); 112DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues); 113 114static void update_rq_clock_task(struct rq *rq, s64 delta); 115 116void update_rq_clock(struct rq *rq) 117{ 118 s64 delta; 119 120 if (rq->skip_clock_update > 0) 121 return; 122 123 delta = sched_clock_cpu(cpu_of(rq)) - rq->clock; 124 rq->clock += delta; 125 update_rq_clock_task(rq, delta); 126} 127 128/* 129 * Debugging: various feature bits 130 */ 131 132#define SCHED_FEAT(name, enabled) \ 133 (1UL << __SCHED_FEAT_##name) * enabled | 134 135const_debug unsigned int sysctl_sched_features = 136#include "features.h" 137 0; 138 139#undef SCHED_FEAT 140 141#ifdef CONFIG_SCHED_DEBUG 142#define SCHED_FEAT(name, enabled) \ 143 #name , 144 145static const char * const sched_feat_names[] = { 146#include "features.h" 147}; 148 149#undef SCHED_FEAT 150 151static int sched_feat_show(struct seq_file *m, void *v) 152{ 153 int i; 154 155 for (i = 0; i < __SCHED_FEAT_NR; i++) { 156 if (!(sysctl_sched_features & (1UL << i))) 157 seq_puts(m, "NO_"); 158 seq_printf(m, "%s ", sched_feat_names[i]); 159 } 160 seq_puts(m, "\n"); 161 162 return 0; 163} 164 165#ifdef HAVE_JUMP_LABEL 166 167#define jump_label_key__true STATIC_KEY_INIT_TRUE 168#define jump_label_key__false STATIC_KEY_INIT_FALSE 169 170#define SCHED_FEAT(name, enabled) \ 171 jump_label_key__##enabled , 172 173struct static_key sched_feat_keys[__SCHED_FEAT_NR] = { 174#include "features.h" 175}; 176 177#undef SCHED_FEAT 178 179static void sched_feat_disable(int i) 180{ 181 if (static_key_enabled(&sched_feat_keys[i])) 182 static_key_slow_dec(&sched_feat_keys[i]); 183} 184 185static void sched_feat_enable(int i) 186{ 187 if (!static_key_enabled(&sched_feat_keys[i])) 188 static_key_slow_inc(&sched_feat_keys[i]); 189} 190#else 191static void sched_feat_disable(int i) { }; 192static void sched_feat_enable(int i) { }; 193#endif /* HAVE_JUMP_LABEL */ 194 195static ssize_t 196sched_feat_write(struct file *filp, const char __user *ubuf, 197 size_t cnt, loff_t *ppos) 198{ 199 char buf[64]; 200 char *cmp; 201 int neg = 0; 202 int i; 203 204 if (cnt > 63) 205 cnt = 63; 206 207 if (copy_from_user(&buf, ubuf, cnt)) 208 return -EFAULT; 209 210 buf[cnt] = 0; 211 cmp = strstrip(buf); 212 213 if (strncmp(cmp, "NO_", 3) == 0) { 214 neg = 1; 215 cmp += 3; 216 } 217 218 for (i = 0; i < __SCHED_FEAT_NR; i++) { 219 if (strcmp(cmp, sched_feat_names[i]) == 0) { 220 if (neg) { 221 sysctl_sched_features &= ~(1UL << i); 222 sched_feat_disable(i); 223 } else { 224 sysctl_sched_features |= (1UL << i); 225 sched_feat_enable(i); 226 } 227 break; 228 } 229 } 230 231 if (i == __SCHED_FEAT_NR) 232 return -EINVAL; 233 234 *ppos += cnt; 235 236 return cnt; 237} 238 239static int sched_feat_open(struct inode *inode, struct file *filp) 240{ 241 return single_open(filp, sched_feat_show, NULL); 242} 243 244static const struct file_operations sched_feat_fops = { 245 .open = sched_feat_open, 246 .write = sched_feat_write, 247 .read = seq_read, 248 .llseek = seq_lseek, 249 .release = single_release, 250}; 251 252static __init int sched_init_debug(void) 253{ 254 debugfs_create_file("sched_features", 0644, NULL, NULL, 255 &sched_feat_fops); 256 257 return 0; 258} 259late_initcall(sched_init_debug); 260#endif /* CONFIG_SCHED_DEBUG */ 261 262/* 263 * Number of tasks to iterate in a single balance run. 264 * Limited because this is done with IRQs disabled. 265 */ 266const_debug unsigned int sysctl_sched_nr_migrate = 32; 267 268/* 269 * period over which we average the RT time consumption, measured 270 * in ms. 271 * 272 * default: 1s 273 */ 274const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC; 275 276/* 277 * period over which we measure -rt task cpu usage in us. 278 * default: 1s 279 */ 280unsigned int sysctl_sched_rt_period = 1000000; 281 282__read_mostly int scheduler_running; 283 284/* 285 * part of the period that we allow rt tasks to run in us. 286 * default: 0.95s 287 */ 288int sysctl_sched_rt_runtime = 950000; 289 290 291 292/* 293 * __task_rq_lock - lock the rq @p resides on. 294 */ 295static inline struct rq *__task_rq_lock(struct task_struct *p) 296 __acquires(rq->lock) 297{ 298 struct rq *rq; 299 300 lockdep_assert_held(&p->pi_lock); 301 302 for (;;) { 303 rq = task_rq(p); 304 raw_spin_lock(&rq->lock); 305 if (likely(rq == task_rq(p))) 306 return rq; 307 raw_spin_unlock(&rq->lock); 308 } 309} 310 311/* 312 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on. 313 */ 314static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags) 315 __acquires(p->pi_lock) 316 __acquires(rq->lock) 317{ 318 struct rq *rq; 319 320 for (;;) { 321 raw_spin_lock_irqsave(&p->pi_lock, *flags); 322 rq = task_rq(p); 323 raw_spin_lock(&rq->lock); 324 if (likely(rq == task_rq(p))) 325 return rq; 326 raw_spin_unlock(&rq->lock); 327 raw_spin_unlock_irqrestore(&p->pi_lock, *flags); 328 } 329} 330 331static void __task_rq_unlock(struct rq *rq) 332 __releases(rq->lock) 333{ 334 raw_spin_unlock(&rq->lock); 335} 336 337static inline void 338task_rq_unlock(struct rq *rq, struct task_struct *p, unsigned long *flags) 339 __releases(rq->lock) 340 __releases(p->pi_lock) 341{ 342 raw_spin_unlock(&rq->lock); 343 raw_spin_unlock_irqrestore(&p->pi_lock, *flags); 344} 345 346/* 347 * this_rq_lock - lock this runqueue and disable interrupts. 348 */ 349static struct rq *this_rq_lock(void) 350 __acquires(rq->lock) 351{ 352 struct rq *rq; 353 354 local_irq_disable(); 355 rq = this_rq(); 356 raw_spin_lock(&rq->lock); 357 358 return rq; 359} 360 361#ifdef CONFIG_SCHED_HRTICK 362/* 363 * Use HR-timers to deliver accurate preemption points. 364 * 365 * Its all a bit involved since we cannot program an hrt while holding the 366 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a 367 * reschedule event. 368 * 369 * When we get rescheduled we reprogram the hrtick_timer outside of the 370 * rq->lock. 371 */ 372 373static void hrtick_clear(struct rq *rq) 374{ 375 if (hrtimer_active(&rq->hrtick_timer)) 376 hrtimer_cancel(&rq->hrtick_timer); 377} 378 379/* 380 * High-resolution timer tick. 381 * Runs from hardirq context with interrupts disabled. 382 */ 383static enum hrtimer_restart hrtick(struct hrtimer *timer) 384{ 385 struct rq *rq = container_of(timer, struct rq, hrtick_timer); 386 387 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id()); 388 389 raw_spin_lock(&rq->lock); 390 update_rq_clock(rq); 391 rq->curr->sched_class->task_tick(rq, rq->curr, 1); 392 raw_spin_unlock(&rq->lock); 393 394 return HRTIMER_NORESTART; 395} 396 397#ifdef CONFIG_SMP 398/* 399 * called from hardirq (IPI) context 400 */ 401static void __hrtick_start(void *arg) 402{ 403 struct rq *rq = arg; 404 405 raw_spin_lock(&rq->lock); 406 hrtimer_restart(&rq->hrtick_timer); 407 rq->hrtick_csd_pending = 0; 408 raw_spin_unlock(&rq->lock); 409} 410 411/* 412 * Called to set the hrtick timer state. 413 * 414 * called with rq->lock held and irqs disabled 415 */ 416void hrtick_start(struct rq *rq, u64 delay) 417{ 418 struct hrtimer *timer = &rq->hrtick_timer; 419 ktime_t time = ktime_add_ns(timer->base->get_time(), delay); 420 421 hrtimer_set_expires(timer, time); 422 423 if (rq == this_rq()) { 424 hrtimer_restart(timer); 425 } else if (!rq->hrtick_csd_pending) { 426 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd, 0); 427 rq->hrtick_csd_pending = 1; 428 } 429} 430 431static int 432hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu) 433{ 434 int cpu = (int)(long)hcpu; 435 436 switch (action) { 437 case CPU_UP_CANCELED: 438 case CPU_UP_CANCELED_FROZEN: 439 case CPU_DOWN_PREPARE: 440 case CPU_DOWN_PREPARE_FROZEN: 441 case CPU_DEAD: 442 case CPU_DEAD_FROZEN: 443 hrtick_clear(cpu_rq(cpu)); 444 return NOTIFY_OK; 445 } 446 447 return NOTIFY_DONE; 448} 449 450static __init void init_hrtick(void) 451{ 452 hotcpu_notifier(hotplug_hrtick, 0); 453} 454#else 455/* 456 * Called to set the hrtick timer state. 457 * 458 * called with rq->lock held and irqs disabled 459 */ 460void hrtick_start(struct rq *rq, u64 delay) 461{ 462 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0, 463 HRTIMER_MODE_REL_PINNED, 0); 464} 465 466static inline void init_hrtick(void) 467{ 468} 469#endif /* CONFIG_SMP */ 470 471static void init_rq_hrtick(struct rq *rq) 472{ 473#ifdef CONFIG_SMP 474 rq->hrtick_csd_pending = 0; 475 476 rq->hrtick_csd.flags = 0; 477 rq->hrtick_csd.func = __hrtick_start; 478 rq->hrtick_csd.info = rq; 479#endif 480 481 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); 482 rq->hrtick_timer.function = hrtick; 483} 484#else /* CONFIG_SCHED_HRTICK */ 485static inline void hrtick_clear(struct rq *rq) 486{ 487} 488 489static inline void init_rq_hrtick(struct rq *rq) 490{ 491} 492 493static inline void init_hrtick(void) 494{ 495} 496#endif /* CONFIG_SCHED_HRTICK */ 497 498/* 499 * resched_task - mark a task 'to be rescheduled now'. 500 * 501 * On UP this means the setting of the need_resched flag, on SMP it 502 * might also involve a cross-CPU call to trigger the scheduler on 503 * the target CPU. 504 */ 505#ifdef CONFIG_SMP 506 507#ifndef tsk_is_polling 508#define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG) 509#endif 510 511void resched_task(struct task_struct *p) 512{ 513 int cpu; 514 515 assert_raw_spin_locked(&task_rq(p)->lock); 516 517 if (test_tsk_need_resched(p)) 518 return; 519 520 set_tsk_need_resched(p); 521 522 cpu = task_cpu(p); 523 if (cpu == smp_processor_id()) 524 return; 525 526 /* NEED_RESCHED must be visible before we test polling */ 527 smp_mb(); 528 if (!tsk_is_polling(p)) 529 smp_send_reschedule(cpu); 530} 531 532void resched_cpu(int cpu) 533{ 534 struct rq *rq = cpu_rq(cpu); 535 unsigned long flags; 536 537 if (!raw_spin_trylock_irqsave(&rq->lock, flags)) 538 return; 539 resched_task(cpu_curr(cpu)); 540 raw_spin_unlock_irqrestore(&rq->lock, flags); 541} 542 543#ifdef CONFIG_NO_HZ 544/* 545 * In the semi idle case, use the nearest busy cpu for migrating timers 546 * from an idle cpu. This is good for power-savings. 547 * 548 * We don't do similar optimization for completely idle system, as 549 * selecting an idle cpu will add more delays to the timers than intended 550 * (as that cpu's timer base may not be uptodate wrt jiffies etc). 551 */ 552int get_nohz_timer_target(void) 553{ 554 int cpu = smp_processor_id(); 555 int i; 556 struct sched_domain *sd; 557 558 rcu_read_lock(); 559 for_each_domain(cpu, sd) { 560 for_each_cpu(i, sched_domain_span(sd)) { 561 if (!idle_cpu(i)) { 562 cpu = i; 563 goto unlock; 564 } 565 } 566 } 567unlock: 568 rcu_read_unlock(); 569 return cpu; 570} 571/* 572 * When add_timer_on() enqueues a timer into the timer wheel of an 573 * idle CPU then this timer might expire before the next timer event 574 * which is scheduled to wake up that CPU. In case of a completely 575 * idle system the next event might even be infinite time into the 576 * future. wake_up_idle_cpu() ensures that the CPU is woken up and 577 * leaves the inner idle loop so the newly added timer is taken into 578 * account when the CPU goes back to idle and evaluates the timer 579 * wheel for the next timer event. 580 */ 581void wake_up_idle_cpu(int cpu) 582{ 583 struct rq *rq = cpu_rq(cpu); 584 585 if (cpu == smp_processor_id()) 586 return; 587 588 /* 589 * This is safe, as this function is called with the timer 590 * wheel base lock of (cpu) held. When the CPU is on the way 591 * to idle and has not yet set rq->curr to idle then it will 592 * be serialized on the timer wheel base lock and take the new 593 * timer into account automatically. 594 */ 595 if (rq->curr != rq->idle) 596 return; 597 598 /* 599 * We can set TIF_RESCHED on the idle task of the other CPU 600 * lockless. The worst case is that the other CPU runs the 601 * idle task through an additional NOOP schedule() 602 */ 603 set_tsk_need_resched(rq->idle); 604 605 /* NEED_RESCHED must be visible before we test polling */ 606 smp_mb(); 607 if (!tsk_is_polling(rq->idle)) 608 smp_send_reschedule(cpu); 609} 610 611static inline bool got_nohz_idle_kick(void) 612{ 613 int cpu = smp_processor_id(); 614 return idle_cpu(cpu) && test_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu)); 615} 616 617#else /* CONFIG_NO_HZ */ 618 619static inline bool got_nohz_idle_kick(void) 620{ 621 return false; 622} 623 624#endif /* CONFIG_NO_HZ */ 625 626void sched_avg_update(struct rq *rq) 627{ 628 s64 period = sched_avg_period(); 629 630 while ((s64)(rq->clock - rq->age_stamp) > period) { 631 /* 632 * Inline assembly required to prevent the compiler 633 * optimising this loop into a divmod call. 634 * See __iter_div_u64_rem() for another example of this. 635 */ 636 asm("" : "+rm" (rq->age_stamp)); 637 rq->age_stamp += period; 638 rq->rt_avg /= 2; 639 } 640} 641 642#else /* !CONFIG_SMP */ 643void resched_task(struct task_struct *p) 644{ 645 assert_raw_spin_locked(&task_rq(p)->lock); 646 set_tsk_need_resched(p); 647} 648#endif /* CONFIG_SMP */ 649 650#if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \ 651 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH))) 652/* 653 * Iterate task_group tree rooted at *from, calling @down when first entering a 654 * node and @up when leaving it for the final time. 655 * 656 * Caller must hold rcu_lock or sufficient equivalent. 657 */ 658int walk_tg_tree_from(struct task_group *from, 659 tg_visitor down, tg_visitor up, void *data) 660{ 661 struct task_group *parent, *child; 662 int ret; 663 664 parent = from; 665 666down: 667 ret = (*down)(parent, data); 668 if (ret) 669 goto out; 670 list_for_each_entry_rcu(child, &parent->children, siblings) { 671 parent = child; 672 goto down; 673 674up: 675 continue; 676 } 677 ret = (*up)(parent, data); 678 if (ret || parent == from) 679 goto out; 680 681 child = parent; 682 parent = parent->parent; 683 if (parent) 684 goto up; 685out: 686 return ret; 687} 688 689int tg_nop(struct task_group *tg, void *data) 690{ 691 return 0; 692} 693#endif 694 695static void set_load_weight(struct task_struct *p) 696{ 697 int prio = p->static_prio - MAX_RT_PRIO; 698 struct load_weight *load = &p->se.load; 699 700 /* 701 * SCHED_IDLE tasks get minimal weight: 702 */ 703 if (p->policy == SCHED_IDLE) { 704 load->weight = scale_load(WEIGHT_IDLEPRIO); 705 load->inv_weight = WMULT_IDLEPRIO; 706 return; 707 } 708 709 load->weight = scale_load(prio_to_weight[prio]); 710 load->inv_weight = prio_to_wmult[prio]; 711} 712 713static void enqueue_task(struct rq *rq, struct task_struct *p, int flags) 714{ 715 update_rq_clock(rq); 716 sched_info_queued(p); 717 p->sched_class->enqueue_task(rq, p, flags); 718} 719 720static void dequeue_task(struct rq *rq, struct task_struct *p, int flags) 721{ 722 update_rq_clock(rq); 723 sched_info_dequeued(p); 724 p->sched_class->dequeue_task(rq, p, flags); 725} 726 727void activate_task(struct rq *rq, struct task_struct *p, int flags) 728{ 729 if (task_contributes_to_load(p)) 730 rq->nr_uninterruptible--; 731 732 enqueue_task(rq, p, flags); 733} 734 735void deactivate_task(struct rq *rq, struct task_struct *p, int flags) 736{ 737 if (task_contributes_to_load(p)) 738 rq->nr_uninterruptible++; 739 740 dequeue_task(rq, p, flags); 741} 742 743#ifdef CONFIG_IRQ_TIME_ACCOUNTING 744 745/* 746 * There are no locks covering percpu hardirq/softirq time. 747 * They are only modified in account_system_vtime, on corresponding CPU 748 * with interrupts disabled. So, writes are safe. 749 * They are read and saved off onto struct rq in update_rq_clock(). 750 * This may result in other CPU reading this CPU's irq time and can 751 * race with irq/account_system_vtime on this CPU. We would either get old 752 * or new value with a side effect of accounting a slice of irq time to wrong 753 * task when irq is in progress while we read rq->clock. That is a worthy 754 * compromise in place of having locks on each irq in account_system_time. 755 */ 756static DEFINE_PER_CPU(u64, cpu_hardirq_time); 757static DEFINE_PER_CPU(u64, cpu_softirq_time); 758 759static DEFINE_PER_CPU(u64, irq_start_time); 760static int sched_clock_irqtime; 761 762void enable_sched_clock_irqtime(void) 763{ 764 sched_clock_irqtime = 1; 765} 766 767void disable_sched_clock_irqtime(void) 768{ 769 sched_clock_irqtime = 0; 770} 771 772#ifndef CONFIG_64BIT 773static DEFINE_PER_CPU(seqcount_t, irq_time_seq); 774 775static inline void irq_time_write_begin(void) 776{ 777 __this_cpu_inc(irq_time_seq.sequence); 778 smp_wmb(); 779} 780 781static inline void irq_time_write_end(void) 782{ 783 smp_wmb(); 784 __this_cpu_inc(irq_time_seq.sequence); 785} 786 787static inline u64 irq_time_read(int cpu) 788{ 789 u64 irq_time; 790 unsigned seq; 791 792 do { 793 seq = read_seqcount_begin(&per_cpu(irq_time_seq, cpu)); 794 irq_time = per_cpu(cpu_softirq_time, cpu) + 795 per_cpu(cpu_hardirq_time, cpu); 796 } while (read_seqcount_retry(&per_cpu(irq_time_seq, cpu), seq)); 797 798 return irq_time; 799} 800#else /* CONFIG_64BIT */ 801static inline void irq_time_write_begin(void) 802{ 803} 804 805static inline void irq_time_write_end(void) 806{ 807} 808 809static inline u64 irq_time_read(int cpu) 810{ 811 return per_cpu(cpu_softirq_time, cpu) + per_cpu(cpu_hardirq_time, cpu); 812} 813#endif /* CONFIG_64BIT */ 814 815/* 816 * Called before incrementing preempt_count on {soft,}irq_enter 817 * and before decrementing preempt_count on {soft,}irq_exit. 818 */ 819void account_system_vtime(struct task_struct *curr) 820{ 821 unsigned long flags; 822 s64 delta; 823 int cpu; 824 825 if (!sched_clock_irqtime) 826 return; 827 828 local_irq_save(flags); 829 830 cpu = smp_processor_id(); 831 delta = sched_clock_cpu(cpu) - __this_cpu_read(irq_start_time); 832 __this_cpu_add(irq_start_time, delta); 833 834 irq_time_write_begin(); 835 /* 836 * We do not account for softirq time from ksoftirqd here. 837 * We want to continue accounting softirq time to ksoftirqd thread 838 * in that case, so as not to confuse scheduler with a special task 839 * that do not consume any time, but still wants to run. 840 */ 841 if (hardirq_count()) 842 __this_cpu_add(cpu_hardirq_time, delta); 843 else if (in_serving_softirq() && curr != this_cpu_ksoftirqd()) 844 __this_cpu_add(cpu_softirq_time, delta); 845 846 irq_time_write_end(); 847 local_irq_restore(flags); 848} 849EXPORT_SYMBOL_GPL(account_system_vtime); 850 851#endif /* CONFIG_IRQ_TIME_ACCOUNTING */ 852 853#ifdef CONFIG_PARAVIRT 854static inline u64 steal_ticks(u64 steal) 855{ 856 if (unlikely(steal > NSEC_PER_SEC)) 857 return div_u64(steal, TICK_NSEC); 858 859 return __iter_div_u64_rem(steal, TICK_NSEC, &steal); 860} 861#endif 862 863static void update_rq_clock_task(struct rq *rq, s64 delta) 864{ 865/* 866 * In theory, the compile should just see 0 here, and optimize out the call 867 * to sched_rt_avg_update. But I don't trust it... 868 */ 869#if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING) 870 s64 steal = 0, irq_delta = 0; 871#endif 872#ifdef CONFIG_IRQ_TIME_ACCOUNTING 873 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time; 874 875 /* 876 * Since irq_time is only updated on {soft,}irq_exit, we might run into 877 * this case when a previous update_rq_clock() happened inside a 878 * {soft,}irq region. 879 * 880 * When this happens, we stop ->clock_task and only update the 881 * prev_irq_time stamp to account for the part that fit, so that a next 882 * update will consume the rest. This ensures ->clock_task is 883 * monotonic. 884 * 885 * It does however cause some slight miss-attribution of {soft,}irq 886 * time, a more accurate solution would be to update the irq_time using 887 * the current rq->clock timestamp, except that would require using 888 * atomic ops. 889 */ 890 if (irq_delta > delta) 891 irq_delta = delta; 892 893 rq->prev_irq_time += irq_delta; 894 delta -= irq_delta; 895#endif 896#ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING 897 if (static_key_false((¶virt_steal_rq_enabled))) { 898 u64 st; 899 900 steal = paravirt_steal_clock(cpu_of(rq)); 901 steal -= rq->prev_steal_time_rq; 902 903 if (unlikely(steal > delta)) 904 steal = delta; 905 906 st = steal_ticks(steal); 907 steal = st * TICK_NSEC; 908 909 rq->prev_steal_time_rq += steal; 910 911 delta -= steal; 912 } 913#endif 914 915 rq->clock_task += delta; 916 917#if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING) 918 if ((irq_delta + steal) && sched_feat(NONTASK_POWER)) 919 sched_rt_avg_update(rq, irq_delta + steal); 920#endif 921} 922 923#ifdef CONFIG_IRQ_TIME_ACCOUNTING 924static int irqtime_account_hi_update(void) 925{ 926 u64 *cpustat = kcpustat_this_cpu->cpustat; 927 unsigned long flags; 928 u64 latest_ns; 929 int ret = 0; 930 931 local_irq_save(flags); 932 latest_ns = this_cpu_read(cpu_hardirq_time); 933 if (nsecs_to_cputime64(latest_ns) > cpustat[CPUTIME_IRQ]) 934 ret = 1; 935 local_irq_restore(flags); 936 return ret; 937} 938 939static int irqtime_account_si_update(void) 940{ 941 u64 *cpustat = kcpustat_this_cpu->cpustat; 942 unsigned long flags; 943 u64 latest_ns; 944 int ret = 0; 945 946 local_irq_save(flags); 947 latest_ns = this_cpu_read(cpu_softirq_time); 948 if (nsecs_to_cputime64(latest_ns) > cpustat[CPUTIME_SOFTIRQ]) 949 ret = 1; 950 local_irq_restore(flags); 951 return ret; 952} 953 954#else /* CONFIG_IRQ_TIME_ACCOUNTING */ 955 956#define sched_clock_irqtime (0) 957 958#endif 959 960void sched_set_stop_task(int cpu, struct task_struct *stop) 961{ 962 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 }; 963 struct task_struct *old_stop = cpu_rq(cpu)->stop; 964 965 if (stop) { 966 /* 967 * Make it appear like a SCHED_FIFO task, its something 968 * userspace knows about and won't get confused about. 969 * 970 * Also, it will make PI more or less work without too 971 * much confusion -- but then, stop work should not 972 * rely on PI working anyway. 973 */ 974 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m); 975 976 stop->sched_class = &stop_sched_class; 977 } 978 979 cpu_rq(cpu)->stop = stop; 980 981 if (old_stop) { 982 /* 983 * Reset it back to a normal scheduling class so that 984 * it can die in pieces. 985 */ 986 old_stop->sched_class = &rt_sched_class; 987 } 988} 989 990/* 991 * __normal_prio - return the priority that is based on the static prio 992 */ 993static inline int __normal_prio(struct task_struct *p) 994{ 995 return p->static_prio; 996} 997 998/* 999 * Calculate the expected normal priority: i.e. priority 1000 * without taking RT-inheritance into account. Might be 1001 * boosted by interactivity modifiers. Changes upon fork, 1002 * setprio syscalls, and whenever the interactivity 1003 * estimator recalculates. 1004 */ 1005static inline int normal_prio(struct task_struct *p) 1006{ 1007 int prio; 1008 1009 if (task_has_rt_policy(p)) 1010 prio = MAX_RT_PRIO-1 - p->rt_priority; 1011 else 1012 prio = __normal_prio(p); 1013 return prio; 1014} 1015 1016/* 1017 * Calculate the current priority, i.e. the priority 1018 * taken into account by the scheduler. This value might 1019 * be boosted by RT tasks, or might be boosted by 1020 * interactivity modifiers. Will be RT if the task got 1021 * RT-boosted. If not then it returns p->normal_prio. 1022 */ 1023static int effective_prio(struct task_struct *p) 1024{ 1025 p->normal_prio = normal_prio(p); 1026 /* 1027 * If we are RT tasks or we were boosted to RT priority, 1028 * keep the priority unchanged. Otherwise, update priority 1029 * to the normal priority: 1030 */ 1031 if (!rt_prio(p->prio)) 1032 return p->normal_prio; 1033 return p->prio; 1034} 1035 1036/** 1037 * task_curr - is this task currently executing on a CPU? 1038 * @p: the task in question. 1039 */ 1040inline int task_curr(const struct task_struct *p) 1041{ 1042 return cpu_curr(task_cpu(p)) == p; 1043} 1044 1045static inline void check_class_changed(struct rq *rq, struct task_struct *p, 1046 const struct sched_class *prev_class, 1047 int oldprio) 1048{ 1049 if (prev_class != p->sched_class) { 1050 if (prev_class->switched_from) 1051 prev_class->switched_from(rq, p); 1052 p->sched_class->switched_to(rq, p); 1053 } else if (oldprio != p->prio) 1054 p->sched_class->prio_changed(rq, p, oldprio); 1055} 1056 1057void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags) 1058{ 1059 const struct sched_class *class; 1060 1061 if (p->sched_class == rq->curr->sched_class) { 1062 rq->curr->sched_class->check_preempt_curr(rq, p, flags); 1063 } else { 1064 for_each_class(class) { 1065 if (class == rq->curr->sched_class) 1066 break; 1067 if (class == p->sched_class) { 1068 resched_task(rq->curr); 1069 break; 1070 } 1071 } 1072 } 1073 1074 /* 1075 * A queue event has occurred, and we're going to schedule. In 1076 * this case, we can save a useless back to back clock update. 1077 */ 1078 if (rq->curr->on_rq && test_tsk_need_resched(rq->curr)) 1079 rq->skip_clock_update = 1; 1080} 1081 1082#ifdef CONFIG_SMP 1083void set_task_cpu(struct task_struct *p, unsigned int new_cpu) 1084{ 1085#ifdef CONFIG_SCHED_DEBUG 1086 /* 1087 * We should never call set_task_cpu() on a blocked task, 1088 * ttwu() will sort out the placement. 1089 */ 1090 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING && 1091 !(task_thread_info(p)->preempt_count & PREEMPT_ACTIVE)); 1092 1093#ifdef CONFIG_LOCKDEP 1094 /* 1095 * The caller should hold either p->pi_lock or rq->lock, when changing 1096 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks. 1097 * 1098 * sched_move_task() holds both and thus holding either pins the cgroup, 1099 * see task_group(). 1100 * 1101 * Furthermore, all task_rq users should acquire both locks, see 1102 * task_rq_lock(). 1103 */ 1104 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) || 1105 lockdep_is_held(&task_rq(p)->lock))); 1106#endif 1107#endif 1108 1109 trace_sched_migrate_task(p, new_cpu); 1110 1111 if (task_cpu(p) != new_cpu) { 1112 p->se.nr_migrations++; 1113 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, NULL, 0); 1114 } 1115 1116 __set_task_cpu(p, new_cpu); 1117} 1118 1119struct migration_arg { 1120 struct task_struct *task; 1121 int dest_cpu; 1122}; 1123 1124static int migration_cpu_stop(void *data); 1125 1126/* 1127 * wait_task_inactive - wait for a thread to unschedule. 1128 * 1129 * If @match_state is nonzero, it's the @p->state value just checked and 1130 * not expected to change. If it changes, i.e. @p might have woken up, 1131 * then return zero. When we succeed in waiting for @p to be off its CPU, 1132 * we return a positive number (its total switch count). If a second call 1133 * a short while later returns the same number, the caller can be sure that 1134 * @p has remained unscheduled the whole time. 1135 * 1136 * The caller must ensure that the task *will* unschedule sometime soon, 1137 * else this function might spin for a *long* time. This function can't 1138 * be called with interrupts off, or it may introduce deadlock with 1139 * smp_call_function() if an IPI is sent by the same process we are 1140 * waiting to become inactive. 1141 */ 1142unsigned long wait_task_inactive(struct task_struct *p, long match_state) 1143{ 1144 unsigned long flags; 1145 int running, on_rq; 1146 unsigned long ncsw; 1147 struct rq *rq; 1148 1149 for (;;) { 1150 /* 1151 * We do the initial early heuristics without holding 1152 * any task-queue locks at all. We'll only try to get 1153 * the runqueue lock when things look like they will 1154 * work out! 1155 */ 1156 rq = task_rq(p); 1157 1158 /* 1159 * If the task is actively running on another CPU 1160 * still, just relax and busy-wait without holding 1161 * any locks. 1162 * 1163 * NOTE! Since we don't hold any locks, it's not 1164 * even sure that "rq" stays as the right runqueue! 1165 * But we don't care, since "task_running()" will 1166 * return false if the runqueue has changed and p 1167 * is actually now running somewhere else! 1168 */ 1169 while (task_running(rq, p)) { 1170 if (match_state && unlikely(p->state != match_state)) 1171 return 0; 1172 cpu_relax(); 1173 } 1174 1175 /* 1176 * Ok, time to look more closely! We need the rq 1177 * lock now, to be *sure*. If we're wrong, we'll 1178 * just go back and repeat. 1179 */ 1180 rq = task_rq_lock(p, &flags); 1181 trace_sched_wait_task(p); 1182 running = task_running(rq, p); 1183 on_rq = p->on_rq; 1184 ncsw = 0; 1185 if (!match_state || p->state == match_state) 1186 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */ 1187 task_rq_unlock(rq, p, &flags); 1188 1189 /* 1190 * If it changed from the expected state, bail out now. 1191 */ 1192 if (unlikely(!ncsw)) 1193 break; 1194 1195 /* 1196 * Was it really running after all now that we 1197 * checked with the proper locks actually held? 1198 * 1199 * Oops. Go back and try again.. 1200 */ 1201 if (unlikely(running)) { 1202 cpu_relax(); 1203 continue; 1204 } 1205 1206 /* 1207 * It's not enough that it's not actively running, 1208 * it must be off the runqueue _entirely_, and not 1209 * preempted! 1210 * 1211 * So if it was still runnable (but just not actively 1212 * running right now), it's preempted, and we should 1213 * yield - it could be a while. 1214 */ 1215 if (unlikely(on_rq)) { 1216 ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ); 1217 1218 set_current_state(TASK_UNINTERRUPTIBLE); 1219 schedule_hrtimeout(&to, HRTIMER_MODE_REL); 1220 continue; 1221 } 1222 1223 /* 1224 * Ahh, all good. It wasn't running, and it wasn't 1225 * runnable, which means that it will never become 1226 * running in the future either. We're all done! 1227 */ 1228 break; 1229 } 1230 1231 return ncsw; 1232} 1233 1234/*** 1235 * kick_process - kick a running thread to enter/exit the kernel 1236 * @p: the to-be-kicked thread 1237 * 1238 * Cause a process which is running on another CPU to enter 1239 * kernel-mode, without any delay. (to get signals handled.) 1240 * 1241 * NOTE: this function doesn't have to take the runqueue lock, 1242 * because all it wants to ensure is that the remote task enters 1243 * the kernel. If the IPI races and the task has been migrated 1244 * to another CPU then no harm is done and the purpose has been 1245 * achieved as well. 1246 */ 1247void kick_process(struct task_struct *p) 1248{ 1249 int cpu; 1250 1251 preempt_disable(); 1252 cpu = task_cpu(p); 1253 if ((cpu != smp_processor_id()) && task_curr(p)) 1254 smp_send_reschedule(cpu); 1255 preempt_enable(); 1256} 1257EXPORT_SYMBOL_GPL(kick_process); 1258#endif /* CONFIG_SMP */ 1259 1260#ifdef CONFIG_SMP 1261/* 1262 * ->cpus_allowed is protected by both rq->lock and p->pi_lock 1263 */ 1264static int select_fallback_rq(int cpu, struct task_struct *p) 1265{ 1266 const struct cpumask *nodemask = cpumask_of_node(cpu_to_node(cpu)); 1267 enum { cpuset, possible, fail } state = cpuset; 1268 int dest_cpu; 1269 1270 /* Look for allowed, online CPU in same node. */ 1271 for_each_cpu(dest_cpu, nodemask) { 1272 if (!cpu_online(dest_cpu)) 1273 continue; 1274 if (!cpu_active(dest_cpu)) 1275 continue; 1276 if (cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p))) 1277 return dest_cpu; 1278 } 1279 1280 for (;;) { 1281 /* Any allowed, online CPU? */ 1282 for_each_cpu(dest_cpu, tsk_cpus_allowed(p)) { 1283 if (!cpu_online(dest_cpu)) 1284 continue; 1285 if (!cpu_active(dest_cpu)) 1286 continue; 1287 goto out; 1288 } 1289 1290 switch (state) { 1291 case cpuset: 1292 /* No more Mr. Nice Guy. */ 1293 cpuset_cpus_allowed_fallback(p); 1294 state = possible; 1295 break; 1296 1297 case possible: 1298 do_set_cpus_allowed(p, cpu_possible_mask); 1299 state = fail; 1300 break; 1301 1302 case fail: 1303 BUG(); 1304 break; 1305 } 1306 } 1307 1308out: 1309 if (state != cpuset) { 1310 /* 1311 * Don't tell them about moving exiting tasks or 1312 * kernel threads (both mm NULL), since they never 1313 * leave kernel. 1314 */ 1315 if (p->mm && printk_ratelimit()) { 1316 printk_sched("process %d (%s) no longer affine to cpu%d\n", 1317 task_pid_nr(p), p->comm, cpu); 1318 } 1319 } 1320 1321 return dest_cpu; 1322} 1323 1324/* 1325 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable. 1326 */ 1327static inline 1328int select_task_rq(struct task_struct *p, int sd_flags, int wake_flags) 1329{ 1330 int cpu = p->sched_class->select_task_rq(p, sd_flags, wake_flags); 1331 1332 /* 1333 * In order not to call set_task_cpu() on a blocking task we need 1334 * to rely on ttwu() to place the task on a valid ->cpus_allowed 1335 * cpu. 1336 * 1337 * Since this is common to all placement strategies, this lives here. 1338 * 1339 * [ this allows ->select_task() to simply return task_cpu(p) and 1340 * not worry about this generic constraint ] 1341 */ 1342 if (unlikely(!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)) || 1343 !cpu_online(cpu))) 1344 cpu = select_fallback_rq(task_cpu(p), p); 1345 1346 return cpu; 1347} 1348 1349static void update_avg(u64 *avg, u64 sample) 1350{ 1351 s64 diff = sample - *avg; 1352 *avg += diff >> 3; 1353} 1354#endif 1355 1356static void 1357ttwu_stat(struct task_struct *p, int cpu, int wake_flags) 1358{ 1359#ifdef CONFIG_SCHEDSTATS 1360 struct rq *rq = this_rq(); 1361 1362#ifdef CONFIG_SMP 1363 int this_cpu = smp_processor_id(); 1364 1365 if (cpu == this_cpu) { 1366 schedstat_inc(rq, ttwu_local); 1367 schedstat_inc(p, se.statistics.nr_wakeups_local); 1368 } else { 1369 struct sched_domain *sd; 1370 1371 schedstat_inc(p, se.statistics.nr_wakeups_remote); 1372 rcu_read_lock(); 1373 for_each_domain(this_cpu, sd) { 1374 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) { 1375 schedstat_inc(sd, ttwu_wake_remote); 1376 break; 1377 } 1378 } 1379 rcu_read_unlock(); 1380 } 1381 1382 if (wake_flags & WF_MIGRATED) 1383 schedstat_inc(p, se.statistics.nr_wakeups_migrate); 1384 1385#endif /* CONFIG_SMP */ 1386 1387 schedstat_inc(rq, ttwu_count); 1388 schedstat_inc(p, se.statistics.nr_wakeups); 1389 1390 if (wake_flags & WF_SYNC) 1391 schedstat_inc(p, se.statistics.nr_wakeups_sync); 1392 1393#endif /* CONFIG_SCHEDSTATS */ 1394} 1395 1396static void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags) 1397{ 1398 activate_task(rq, p, en_flags); 1399 p->on_rq = 1; 1400 1401 /* if a worker is waking up, notify workqueue */ 1402 if (p->flags & PF_WQ_WORKER) 1403 wq_worker_waking_up(p, cpu_of(rq)); 1404} 1405 1406/* 1407 * Mark the task runnable and perform wakeup-preemption. 1408 */ 1409static void 1410ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags) 1411{ 1412 trace_sched_wakeup(p, true); 1413 check_preempt_curr(rq, p, wake_flags); 1414 1415 p->state = TASK_RUNNING; 1416#ifdef CONFIG_SMP 1417 if (p->sched_class->task_woken) 1418 p->sched_class->task_woken(rq, p); 1419 1420 if (rq->idle_stamp) { 1421 u64 delta = rq->clock - rq->idle_stamp; 1422 u64 max = 2*sysctl_sched_migration_cost; 1423 1424 if (delta > max) 1425 rq->avg_idle = max; 1426 else 1427 update_avg(&rq->avg_idle, delta); 1428 rq->idle_stamp = 0; 1429 } 1430#endif 1431} 1432 1433static void 1434ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags) 1435{ 1436#ifdef CONFIG_SMP 1437 if (p->sched_contributes_to_load) 1438 rq->nr_uninterruptible--; 1439#endif 1440 1441 ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_WAKING); 1442 ttwu_do_wakeup(rq, p, wake_flags); 1443} 1444 1445/* 1446 * Called in case the task @p isn't fully descheduled from its runqueue, 1447 * in this case we must do a remote wakeup. Its a 'light' wakeup though, 1448 * since all we need to do is flip p->state to TASK_RUNNING, since 1449 * the task is still ->on_rq. 1450 */ 1451static int ttwu_remote(struct task_struct *p, int wake_flags) 1452{ 1453 struct rq *rq; 1454 int ret = 0; 1455 1456 rq = __task_rq_lock(p); 1457 if (p->on_rq) { 1458 ttwu_do_wakeup(rq, p, wake_flags); 1459 ret = 1; 1460 } 1461 __task_rq_unlock(rq); 1462 1463 return ret; 1464} 1465 1466#ifdef CONFIG_SMP 1467static void sched_ttwu_pending(void) 1468{ 1469 struct rq *rq = this_rq(); 1470 struct llist_node *llist = llist_del_all(&rq->wake_list); 1471 struct task_struct *p; 1472 1473 raw_spin_lock(&rq->lock); 1474 1475 while (llist) { 1476 p = llist_entry(llist, struct task_struct, wake_entry); 1477 llist = llist_next(llist); 1478 ttwu_do_activate(rq, p, 0); 1479 } 1480 1481 raw_spin_unlock(&rq->lock); 1482} 1483 1484void scheduler_ipi(void) 1485{ 1486 if (llist_empty(&this_rq()->wake_list) && !got_nohz_idle_kick()) 1487 return; 1488 1489 /* 1490 * Not all reschedule IPI handlers call irq_enter/irq_exit, since 1491 * traditionally all their work was done from the interrupt return 1492 * path. Now that we actually do some work, we need to make sure 1493 * we do call them. 1494 * 1495 * Some archs already do call them, luckily irq_enter/exit nest 1496 * properly. 1497 * 1498 * Arguably we should visit all archs and update all handlers, 1499 * however a fair share of IPIs are still resched only so this would 1500 * somewhat pessimize the simple resched case. 1501 */ 1502 irq_enter(); 1503 sched_ttwu_pending(); 1504 1505 /* 1506 * Check if someone kicked us for doing the nohz idle load balance. 1507 */ 1508 if (unlikely(got_nohz_idle_kick() && !need_resched())) { 1509 this_rq()->idle_balance = 1; 1510 raise_softirq_irqoff(SCHED_SOFTIRQ); 1511 } 1512 irq_exit(); 1513} 1514 1515static void ttwu_queue_remote(struct task_struct *p, int cpu) 1516{ 1517 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list)) 1518 smp_send_reschedule(cpu); 1519} 1520 1521#ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW 1522static int ttwu_activate_remote(struct task_struct *p, int wake_flags) 1523{ 1524 struct rq *rq; 1525 int ret = 0; 1526 1527 rq = __task_rq_lock(p); 1528 if (p->on_cpu) { 1529 ttwu_activate(rq, p, ENQUEUE_WAKEUP); 1530 ttwu_do_wakeup(rq, p, wake_flags); 1531 ret = 1; 1532 } 1533 __task_rq_unlock(rq); 1534 1535 return ret; 1536 1537} 1538#endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */ 1539 1540bool cpus_share_cache(int this_cpu, int that_cpu) 1541{ 1542 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu); 1543} 1544#endif /* CONFIG_SMP */ 1545 1546static void ttwu_queue(struct task_struct *p, int cpu) 1547{ 1548 struct rq *rq = cpu_rq(cpu); 1549 1550#if defined(CONFIG_SMP) 1551 if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) { 1552 sched_clock_cpu(cpu); /* sync clocks x-cpu */ 1553 ttwu_queue_remote(p, cpu); 1554 return; 1555 } 1556#endif 1557 1558 raw_spin_lock(&rq->lock); 1559 ttwu_do_activate(rq, p, 0); 1560 raw_spin_unlock(&rq->lock); 1561} 1562 1563/** 1564 * try_to_wake_up - wake up a thread 1565 * @p: the thread to be awakened 1566 * @state: the mask of task states that can be woken 1567 * @wake_flags: wake modifier flags (WF_*) 1568 * 1569 * Put it on the run-queue if it's not already there. The "current" 1570 * thread is always on the run-queue (except when the actual 1571 * re-schedule is in progress), and as such you're allowed to do 1572 * the simpler "current->state = TASK_RUNNING" to mark yourself 1573 * runnable without the overhead of this. 1574 * 1575 * Returns %true if @p was woken up, %false if it was already running 1576 * or @state didn't match @p's state. 1577 */ 1578static int 1579try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags) 1580{ 1581 unsigned long flags; 1582 int cpu, success = 0; 1583 1584 smp_wmb(); 1585 raw_spin_lock_irqsave(&p->pi_lock, flags); 1586 if (!(p->state & state)) 1587 goto out; 1588 1589 success = 1; /* we're going to change ->state */ 1590 cpu = task_cpu(p); 1591 1592 if (p->on_rq && ttwu_remote(p, wake_flags)) 1593 goto stat; 1594 1595#ifdef CONFIG_SMP 1596 /* 1597 * If the owning (remote) cpu is still in the middle of schedule() with 1598 * this task as prev, wait until its done referencing the task. 1599 */ 1600 while (p->on_cpu) { 1601#ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW 1602 /* 1603 * In case the architecture enables interrupts in 1604 * context_switch(), we cannot busy wait, since that 1605 * would lead to deadlocks when an interrupt hits and 1606 * tries to wake up @prev. So bail and do a complete 1607 * remote wakeup. 1608 */ 1609 if (ttwu_activate_remote(p, wake_flags)) 1610 goto stat; 1611#else 1612 cpu_relax(); 1613#endif 1614 } 1615 /* 1616 * Pairs with the smp_wmb() in finish_lock_switch(). 1617 */ 1618 smp_rmb(); 1619 1620 p->sched_contributes_to_load = !!task_contributes_to_load(p); 1621 p->state = TASK_WAKING; 1622 1623 if (p->sched_class->task_waking) 1624 p->sched_class->task_waking(p); 1625 1626 cpu = select_task_rq(p, SD_BALANCE_WAKE, wake_flags); 1627 if (task_cpu(p) != cpu) { 1628 wake_flags |= WF_MIGRATED; 1629 set_task_cpu(p, cpu); 1630 } 1631#endif /* CONFIG_SMP */ 1632 1633 ttwu_queue(p, cpu); 1634stat: 1635 ttwu_stat(p, cpu, wake_flags); 1636out: 1637 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 1638 1639 return success; 1640} 1641 1642/** 1643 * try_to_wake_up_local - try to wake up a local task with rq lock held 1644 * @p: the thread to be awakened 1645 * 1646 * Put @p on the run-queue if it's not already there. The caller must 1647 * ensure that this_rq() is locked, @p is bound to this_rq() and not 1648 * the current task. 1649 */ 1650static void try_to_wake_up_local(struct task_struct *p) 1651{ 1652 struct rq *rq = task_rq(p); 1653 1654 BUG_ON(rq != this_rq()); 1655 BUG_ON(p == current); 1656 lockdep_assert_held(&rq->lock); 1657 1658 if (!raw_spin_trylock(&p->pi_lock)) { 1659 raw_spin_unlock(&rq->lock); 1660 raw_spin_lock(&p->pi_lock); 1661 raw_spin_lock(&rq->lock); 1662 } 1663 1664 if (!(p->state & TASK_NORMAL)) 1665 goto out; 1666 1667 if (!p->on_rq) 1668 ttwu_activate(rq, p, ENQUEUE_WAKEUP); 1669 1670 ttwu_do_wakeup(rq, p, 0); 1671 ttwu_stat(p, smp_processor_id(), 0); 1672out: 1673 raw_spin_unlock(&p->pi_lock); 1674} 1675 1676/** 1677 * wake_up_process - Wake up a specific process 1678 * @p: The process to be woken up. 1679 * 1680 * Attempt to wake up the nominated process and move it to the set of runnable 1681 * processes. Returns 1 if the process was woken up, 0 if it was already 1682 * running. 1683 * 1684 * It may be assumed that this function implies a write memory barrier before 1685 * changing the task state if and only if any tasks are woken up. 1686 */ 1687int wake_up_process(struct task_struct *p) 1688{ 1689 return try_to_wake_up(p, TASK_ALL, 0); 1690} 1691EXPORT_SYMBOL(wake_up_process); 1692 1693int wake_up_state(struct task_struct *p, unsigned int state) 1694{ 1695 return try_to_wake_up(p, state, 0); 1696} 1697 1698/* 1699 * Perform scheduler related setup for a newly forked process p. 1700 * p is forked by current. 1701 * 1702 * __sched_fork() is basic setup used by init_idle() too: 1703 */ 1704static void __sched_fork(struct task_struct *p) 1705{ 1706 p->on_rq = 0; 1707 1708 p->se.on_rq = 0; 1709 p->se.exec_start = 0; 1710 p->se.sum_exec_runtime = 0; 1711 p->se.prev_sum_exec_runtime = 0; 1712 p->se.nr_migrations = 0; 1713 p->se.vruntime = 0; 1714 INIT_LIST_HEAD(&p->se.group_node); 1715 1716#ifdef CONFIG_SCHEDSTATS 1717 memset(&p->se.statistics, 0, sizeof(p->se.statistics)); 1718#endif 1719 1720 INIT_LIST_HEAD(&p->rt.run_list); 1721 1722#ifdef CONFIG_PREEMPT_NOTIFIERS 1723 INIT_HLIST_HEAD(&p->preempt_notifiers); 1724#endif 1725} 1726 1727/* 1728 * fork()/clone()-time setup: 1729 */ 1730void sched_fork(struct task_struct *p) 1731{ 1732 unsigned long flags; 1733 int cpu = get_cpu(); 1734 1735 __sched_fork(p); 1736 /* 1737 * We mark the process as running here. This guarantees that 1738 * nobody will actually run it, and a signal or other external 1739 * event cannot wake it up and insert it on the runqueue either. 1740 */ 1741 p->state = TASK_RUNNING; 1742 1743 /* 1744 * Make sure we do not leak PI boosting priority to the child. 1745 */ 1746 p->prio = current->normal_prio; 1747 1748 /* 1749 * Revert to default priority/policy on fork if requested. 1750 */ 1751 if (unlikely(p->sched_reset_on_fork)) { 1752 if (task_has_rt_policy(p)) { 1753 p->policy = SCHED_NORMAL; 1754 p->static_prio = NICE_TO_PRIO(0); 1755 p->rt_priority = 0; 1756 } else if (PRIO_TO_NICE(p->static_prio) < 0) 1757 p->static_prio = NICE_TO_PRIO(0); 1758 1759 p->prio = p->normal_prio = __normal_prio(p); 1760 set_load_weight(p); 1761 1762 /* 1763 * We don't need the reset flag anymore after the fork. It has 1764 * fulfilled its duty: 1765 */ 1766 p->sched_reset_on_fork = 0; 1767 } 1768 1769 if (!rt_prio(p->prio)) 1770 p->sched_class = &fair_sched_class; 1771 1772 if (p->sched_class->task_fork) 1773 p->sched_class->task_fork(p); 1774 1775 /* 1776 * The child is not yet in the pid-hash so no cgroup attach races, 1777 * and the cgroup is pinned to this child due to cgroup_fork() 1778 * is ran before sched_fork(). 1779 * 1780 * Silence PROVE_RCU. 1781 */ 1782 raw_spin_lock_irqsave(&p->pi_lock, flags); 1783 set_task_cpu(p, cpu); 1784 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 1785 1786#if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT) 1787 if (likely(sched_info_on())) 1788 memset(&p->sched_info, 0, sizeof(p->sched_info)); 1789#endif 1790#if defined(CONFIG_SMP) 1791 p->on_cpu = 0; 1792#endif 1793#ifdef CONFIG_PREEMPT_COUNT 1794 /* Want to start with kernel preemption disabled. */ 1795 task_thread_info(p)->preempt_count = 1; 1796#endif 1797#ifdef CONFIG_SMP 1798 plist_node_init(&p->pushable_tasks, MAX_PRIO); 1799#endif 1800 1801 put_cpu(); 1802} 1803 1804/* 1805 * wake_up_new_task - wake up a newly created task for the first time. 1806 * 1807 * This function will do some initial scheduler statistics housekeeping 1808 * that must be done for every newly created context, then puts the task 1809 * on the runqueue and wakes it. 1810 */ 1811void wake_up_new_task(struct task_struct *p) 1812{ 1813 unsigned long flags; 1814 struct rq *rq; 1815 1816 raw_spin_lock_irqsave(&p->pi_lock, flags); 1817#ifdef CONFIG_SMP 1818 /* 1819 * Fork balancing, do it here and not earlier because: 1820 * - cpus_allowed can change in the fork path 1821 * - any previously selected cpu might disappear through hotplug 1822 */ 1823 set_task_cpu(p, select_task_rq(p, SD_BALANCE_FORK, 0)); 1824#endif 1825 1826 rq = __task_rq_lock(p); 1827 activate_task(rq, p, 0); 1828 p->on_rq = 1; 1829 trace_sched_wakeup_new(p, true); 1830 check_preempt_curr(rq, p, WF_FORK); 1831#ifdef CONFIG_SMP 1832 if (p->sched_class->task_woken) 1833 p->sched_class->task_woken(rq, p); 1834#endif 1835 task_rq_unlock(rq, p, &flags); 1836} 1837 1838#ifdef CONFIG_PREEMPT_NOTIFIERS 1839 1840/** 1841 * preempt_notifier_register - tell me when current is being preempted & rescheduled 1842 * @notifier: notifier struct to register 1843 */ 1844void preempt_notifier_register(struct preempt_notifier *notifier) 1845{ 1846 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers); 1847} 1848EXPORT_SYMBOL_GPL(preempt_notifier_register); 1849 1850/** 1851 * preempt_notifier_unregister - no longer interested in preemption notifications 1852 * @notifier: notifier struct to unregister 1853 * 1854 * This is safe to call from within a preemption notifier. 1855 */ 1856void preempt_notifier_unregister(struct preempt_notifier *notifier) 1857{ 1858 hlist_del(¬ifier->link); 1859} 1860EXPORT_SYMBOL_GPL(preempt_notifier_unregister); 1861 1862static void fire_sched_in_preempt_notifiers(struct task_struct *curr) 1863{ 1864 struct preempt_notifier *notifier; 1865 struct hlist_node *node; 1866 1867 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link) 1868 notifier->ops->sched_in(notifier, raw_smp_processor_id()); 1869} 1870 1871static void 1872fire_sched_out_preempt_notifiers(struct task_struct *curr, 1873 struct task_struct *next) 1874{ 1875 struct preempt_notifier *notifier; 1876 struct hlist_node *node; 1877 1878 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link) 1879 notifier->ops->sched_out(notifier, next); 1880} 1881 1882#else /* !CONFIG_PREEMPT_NOTIFIERS */ 1883 1884static void fire_sched_in_preempt_notifiers(struct task_struct *curr) 1885{ 1886} 1887 1888static void 1889fire_sched_out_preempt_notifiers(struct task_struct *curr, 1890 struct task_struct *next) 1891{ 1892} 1893 1894#endif /* CONFIG_PREEMPT_NOTIFIERS */ 1895 1896/** 1897 * prepare_task_switch - prepare to switch tasks 1898 * @rq: the runqueue preparing to switch 1899 * @prev: the current task that is being switched out 1900 * @next: the task we are going to switch to. 1901 * 1902 * This is called with the rq lock held and interrupts off. It must 1903 * be paired with a subsequent finish_task_switch after the context 1904 * switch. 1905 * 1906 * prepare_task_switch sets up locking and calls architecture specific 1907 * hooks. 1908 */ 1909static inline void 1910prepare_task_switch(struct rq *rq, struct task_struct *prev, 1911 struct task_struct *next) 1912{ 1913 trace_sched_switch(prev, next); 1914 sched_info_switch(prev, next); 1915 perf_event_task_sched_out(prev, next); 1916 fire_sched_out_preempt_notifiers(prev, next); 1917 prepare_lock_switch(rq, next); 1918 prepare_arch_switch(next); 1919} 1920 1921/** 1922 * finish_task_switch - clean up after a task-switch 1923 * @rq: runqueue associated with task-switch 1924 * @prev: the thread we just switched away from. 1925 * 1926 * finish_task_switch must be called after the context switch, paired 1927 * with a prepare_task_switch call before the context switch. 1928 * finish_task_switch will reconcile locking set up by prepare_task_switch, 1929 * and do any other architecture-specific cleanup actions. 1930 * 1931 * Note that we may have delayed dropping an mm in context_switch(). If 1932 * so, we finish that here outside of the runqueue lock. (Doing it 1933 * with the lock held can cause deadlocks; see schedule() for 1934 * details.) 1935 */ 1936static void finish_task_switch(struct rq *rq, struct task_struct *prev) 1937 __releases(rq->lock) 1938{ 1939 struct mm_struct *mm = rq->prev_mm; 1940 long prev_state; 1941 1942 rq->prev_mm = NULL; 1943 1944 /* 1945 * A task struct has one reference for the use as "current". 1946 * If a task dies, then it sets TASK_DEAD in tsk->state and calls 1947 * schedule one last time. The schedule call will never return, and 1948 * the scheduled task must drop that reference. 1949 * The test for TASK_DEAD must occur while the runqueue locks are 1950 * still held, otherwise prev could be scheduled on another cpu, die 1951 * there before we look at prev->state, and then the reference would 1952 * be dropped twice. 1953 * Manfred Spraul <manfred@colorfullife.com> 1954 */ 1955 prev_state = prev->state; 1956 finish_arch_switch(prev); 1957#ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW 1958 local_irq_disable(); 1959#endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */ 1960 perf_event_task_sched_in(prev, current); 1961#ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW 1962 local_irq_enable(); 1963#endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */ 1964 finish_lock_switch(rq, prev); 1965 finish_arch_post_lock_switch(); 1966 1967 fire_sched_in_preempt_notifiers(current); 1968 if (mm) 1969 mmdrop(mm); 1970 if (unlikely(prev_state == TASK_DEAD)) { 1971 /* 1972 * Remove function-return probe instances associated with this 1973 * task and put them back on the free list. 1974 */ 1975 kprobe_flush_task(prev); 1976 put_task_struct(prev); 1977 } 1978} 1979 1980#ifdef CONFIG_SMP 1981 1982/* assumes rq->lock is held */ 1983static inline void pre_schedule(struct rq *rq, struct task_struct *prev) 1984{ 1985 if (prev->sched_class->pre_schedule) 1986 prev->sched_class->pre_schedule(rq, prev); 1987} 1988 1989/* rq->lock is NOT held, but preemption is disabled */ 1990static inline void post_schedule(struct rq *rq) 1991{ 1992 if (rq->post_schedule) { 1993 unsigned long flags; 1994 1995 raw_spin_lock_irqsave(&rq->lock, flags); 1996 if (rq->curr->sched_class->post_schedule) 1997 rq->curr->sched_class->post_schedule(rq); 1998 raw_spin_unlock_irqrestore(&rq->lock, flags); 1999 2000 rq->post_schedule = 0; 2001 } 2002} 2003 2004#else 2005 2006static inline void pre_schedule(struct rq *rq, struct task_struct *p) 2007{ 2008} 2009 2010static inline void post_schedule(struct rq *rq) 2011{ 2012} 2013 2014#endif 2015 2016/** 2017 * schedule_tail - first thing a freshly forked thread must call. 2018 * @prev: the thread we just switched away from. 2019 */ 2020asmlinkage void schedule_tail(struct task_struct *prev) 2021 __releases(rq->lock) 2022{ 2023 struct rq *rq = this_rq(); 2024 2025 finish_task_switch(rq, prev); 2026 2027 /* 2028 * FIXME: do we need to worry about rq being invalidated by the 2029 * task_switch? 2030 */ 2031 post_schedule(rq); 2032 2033#ifdef __ARCH_WANT_UNLOCKED_CTXSW 2034 /* In this case, finish_task_switch does not reenable preemption */ 2035 preempt_enable(); 2036#endif 2037 if (current->set_child_tid) 2038 put_user(task_pid_vnr(current), current->set_child_tid); 2039} 2040 2041/* 2042 * context_switch - switch to the new MM and the new 2043 * thread's register state. 2044 */ 2045static inline void 2046context_switch(struct rq *rq, struct task_struct *prev, 2047 struct task_struct *next) 2048{ 2049 struct mm_struct *mm, *oldmm; 2050 2051 prepare_task_switch(rq, prev, next); 2052 2053 mm = next->mm; 2054 oldmm = prev->active_mm; 2055 /* 2056 * For paravirt, this is coupled with an exit in switch_to to 2057 * combine the page table reload and the switch backend into 2058 * one hypercall. 2059 */ 2060 arch_start_context_switch(prev); 2061 2062 if (!mm) { 2063 next->active_mm = oldmm; 2064 atomic_inc(&oldmm->mm_count); 2065 enter_lazy_tlb(oldmm, next); 2066 } else 2067 switch_mm(oldmm, mm, next); 2068 2069 if (!prev->mm) { 2070 prev->active_mm = NULL; 2071 rq->prev_mm = oldmm; 2072 } 2073 /* 2074 * Since the runqueue lock will be released by the next 2075 * task (which is an invalid locking op but in the case 2076 * of the scheduler it's an obvious special-case), so we 2077 * do an early lockdep release here: 2078 */ 2079#ifndef __ARCH_WANT_UNLOCKED_CTXSW 2080 spin_release(&rq->lock.dep_map, 1, _THIS_IP_); 2081#endif 2082 2083 /* Here we just switch the register state and the stack. */ 2084 rcu_switch(prev, next); 2085 switch_to(prev, next, prev); 2086 2087 barrier(); 2088 /* 2089 * this_rq must be evaluated again because prev may have moved 2090 * CPUs since it called schedule(), thus the 'rq' on its stack 2091 * frame will be invalid. 2092 */ 2093 finish_task_switch(this_rq(), prev); 2094} 2095 2096/* 2097 * nr_running, nr_uninterruptible and nr_context_switches: 2098 * 2099 * externally visible scheduler statistics: current number of runnable 2100 * threads, current number of uninterruptible-sleeping threads, total 2101 * number of context switches performed since bootup. 2102 */ 2103unsigned long nr_running(void) 2104{ 2105 unsigned long i, sum = 0; 2106 2107 for_each_online_cpu(i) 2108 sum += cpu_rq(i)->nr_running; 2109 2110 return sum; 2111} 2112 2113unsigned long nr_uninterruptible(void) 2114{ 2115 unsigned long i, sum = 0; 2116 2117 for_each_possible_cpu(i) 2118 sum += cpu_rq(i)->nr_uninterruptible; 2119 2120 /* 2121 * Since we read the counters lockless, it might be slightly 2122 * inaccurate. Do not allow it to go below zero though: 2123 */ 2124 if (unlikely((long)sum < 0)) 2125 sum = 0; 2126 2127 return sum; 2128} 2129 2130unsigned long long nr_context_switches(void) 2131{ 2132 int i; 2133 unsigned long long sum = 0; 2134 2135 for_each_possible_cpu(i) 2136 sum += cpu_rq(i)->nr_switches; 2137 2138 return sum; 2139} 2140 2141unsigned long nr_iowait(void) 2142{ 2143 unsigned long i, sum = 0; 2144 2145 for_each_possible_cpu(i) 2146 sum += atomic_read(&cpu_rq(i)->nr_iowait); 2147 2148 return sum; 2149} 2150 2151unsigned long nr_iowait_cpu(int cpu) 2152{ 2153 struct rq *this = cpu_rq(cpu); 2154 return atomic_read(&this->nr_iowait); 2155} 2156 2157unsigned long this_cpu_load(void) 2158{ 2159 struct rq *this = this_rq(); 2160 return this->cpu_load[0]; 2161} 2162 2163 2164/* 2165 * Global load-average calculations 2166 * 2167 * We take a distributed and async approach to calculating the global load-avg 2168 * in order to minimize overhead. 2169 * 2170 * The global load average is an exponentially decaying average of nr_running + 2171 * nr_uninterruptible. 2172 * 2173 * Once every LOAD_FREQ: 2174 * 2175 * nr_active = 0; 2176 * for_each_possible_cpu(cpu) 2177 * nr_active += cpu_of(cpu)->nr_running + cpu_of(cpu)->nr_uninterruptible; 2178 * 2179 * avenrun[n] = avenrun[0] * exp_n + nr_active * (1 - exp_n) 2180 * 2181 * Due to a number of reasons the above turns in the mess below: 2182 * 2183 * - for_each_possible_cpu() is prohibitively expensive on machines with 2184 * serious number of cpus, therefore we need to take a distributed approach 2185 * to calculating nr_active. 2186 * 2187 * \Sum_i x_i(t) = \Sum_i x_i(t) - x_i(t_0) | x_i(t_0) := 0 2188 * = \Sum_i { \Sum_j=1 x_i(t_j) - x_i(t_j-1) } 2189 * 2190 * So assuming nr_active := 0 when we start out -- true per definition, we 2191 * can simply take per-cpu deltas and fold those into a global accumulate 2192 * to obtain the same result. See calc_load_fold_active(). 2193 * 2194 * Furthermore, in order to avoid synchronizing all per-cpu delta folding 2195 * across the machine, we assume 10 ticks is sufficient time for every 2196 * cpu to have completed this task. 2197 * 2198 * This places an upper-bound on the IRQ-off latency of the machine. Then 2199 * again, being late doesn't loose the delta, just wrecks the sample. 2200 * 2201 * - cpu_rq()->nr_uninterruptible isn't accurately tracked per-cpu because 2202 * this would add another cross-cpu cacheline miss and atomic operation 2203 * to the wakeup path. Instead we increment on whatever cpu the task ran 2204 * when it went into uninterruptible state and decrement on whatever cpu 2205 * did the wakeup. This means that only the sum of nr_uninterruptible over 2206 * all cpus yields the correct result. 2207 * 2208 * This covers the NO_HZ=n code, for extra head-aches, see the comment below. 2209 */ 2210 2211/* Variables and functions for calc_load */ 2212static atomic_long_t calc_load_tasks; 2213static unsigned long calc_load_update; 2214unsigned long avenrun[3]; 2215EXPORT_SYMBOL(avenrun); /* should be removed */ 2216 2217/** 2218 * get_avenrun - get the load average array 2219 * @loads: pointer to dest load array 2220 * @offset: offset to add 2221 * @shift: shift count to shift the result left 2222 * 2223 * These values are estimates at best, so no need for locking. 2224 */ 2225void get_avenrun(unsigned long *loads, unsigned long offset, int shift) 2226{ 2227 loads[0] = (avenrun[0] + offset) << shift; 2228 loads[1] = (avenrun[1] + offset) << shift; 2229 loads[2] = (avenrun[2] + offset) << shift; 2230} 2231 2232static long calc_load_fold_active(struct rq *this_rq) 2233{ 2234 long nr_active, delta = 0; 2235 2236 nr_active = this_rq->nr_running; 2237 nr_active += (long) this_rq->nr_uninterruptible; 2238 2239 if (nr_active != this_rq->calc_load_active) { 2240 delta = nr_active - this_rq->calc_load_active; 2241 this_rq->calc_load_active = nr_active; 2242 } 2243 2244 return delta; 2245} 2246 2247/* 2248 * a1 = a0 * e + a * (1 - e) 2249 */ 2250static unsigned long 2251calc_load(unsigned long load, unsigned long exp, unsigned long active) 2252{ 2253 load *= exp; 2254 load += active * (FIXED_1 - exp); 2255 load += 1UL << (FSHIFT - 1); 2256 return load >> FSHIFT; 2257} 2258 2259#ifdef CONFIG_NO_HZ 2260/* 2261 * Handle NO_HZ for the global load-average. 2262 * 2263 * Since the above described distributed algorithm to compute the global 2264 * load-average relies on per-cpu sampling from the tick, it is affected by 2265 * NO_HZ. 2266 * 2267 * The basic idea is to fold the nr_active delta into a global idle-delta upon 2268 * entering NO_HZ state such that we can include this as an 'extra' cpu delta 2269 * when we read the global state. 2270 * 2271 * Obviously reality has to ruin such a delightfully simple scheme: 2272 * 2273 * - When we go NO_HZ idle during the window, we can negate our sample 2274 * contribution, causing under-accounting. 2275 * 2276 * We avoid this by keeping two idle-delta counters and flipping them 2277 * when the window starts, thus separating old and new NO_HZ load. 2278 * 2279 * The only trick is the slight shift in index flip for read vs write. 2280 * 2281 * 0s 5s 10s 15s 2282 * +10 +10 +10 +10 2283 * |-|-----------|-|-----------|-|-----------|-| 2284 * r:0 0 1 1 0 0 1 1 0 2285 * w:0 1 1 0 0 1 1 0 0 2286 * 2287 * This ensures we'll fold the old idle contribution in this window while 2288 * accumlating the new one. 2289 * 2290 * - When we wake up from NO_HZ idle during the window, we push up our 2291 * contribution, since we effectively move our sample point to a known 2292 * busy state. 2293 * 2294 * This is solved by pushing the window forward, and thus skipping the 2295 * sample, for this cpu (effectively using the idle-delta for this cpu which 2296 * was in effect at the time the window opened). This also solves the issue 2297 * of having to deal with a cpu having been in NOHZ idle for multiple 2298 * LOAD_FREQ intervals. 2299 * 2300 * When making the ILB scale, we should try to pull this in as well. 2301 */ 2302static atomic_long_t calc_load_idle[2]; 2303static int calc_load_idx; 2304 2305static inline int calc_load_write_idx(void) 2306{ 2307 int idx = calc_load_idx; 2308 2309 /* 2310 * See calc_global_nohz(), if we observe the new index, we also 2311 * need to observe the new update time. 2312 */ 2313 smp_rmb(); 2314 2315 /* 2316 * If the folding window started, make sure we start writing in the 2317 * next idle-delta. 2318 */ 2319 if (!time_before(jiffies, calc_load_update)) 2320 idx++; 2321 2322 return idx & 1; 2323} 2324 2325static inline int calc_load_read_idx(void) 2326{ 2327 return calc_load_idx & 1; 2328} 2329 2330void calc_load_enter_idle(void) 2331{ 2332 struct rq *this_rq = this_rq(); 2333 long delta; 2334 2335 /* 2336 * We're going into NOHZ mode, if there's any pending delta, fold it 2337 * into the pending idle delta. 2338 */ 2339 delta = calc_load_fold_active(this_rq); 2340 if (delta) { 2341 int idx = calc_load_write_idx(); 2342 atomic_long_add(delta, &calc_load_idle[idx]); 2343 } 2344} 2345 2346void calc_load_exit_idle(void) 2347{ 2348 struct rq *this_rq = this_rq(); 2349 2350 /* 2351 * If we're still before the sample window, we're done. 2352 */ 2353 if (time_before(jiffies, this_rq->calc_load_update)) 2354 return; 2355 2356 /* 2357 * We woke inside or after the sample window, this means we're already 2358 * accounted through the nohz accounting, so skip the entire deal and 2359 * sync up for the next window. 2360 */ 2361 this_rq->calc_load_update = calc_load_update; 2362 if (time_before(jiffies, this_rq->calc_load_update + 10)) 2363 this_rq->calc_load_update += LOAD_FREQ; 2364} 2365 2366static long calc_load_fold_idle(void) 2367{ 2368 int idx = calc_load_read_idx(); 2369 long delta = 0; 2370 2371 if (atomic_long_read(&calc_load_idle[idx])) 2372 delta = atomic_long_xchg(&calc_load_idle[idx], 0); 2373 2374 return delta; 2375} 2376 2377/** 2378 * fixed_power_int - compute: x^n, in O(log n) time 2379 * 2380 * @x: base of the power 2381 * @frac_bits: fractional bits of @x 2382 * @n: power to raise @x to. 2383 * 2384 * By exploiting the relation between the definition of the natural power 2385 * function: x^n := x*x*...*x (x multiplied by itself for n times), and 2386 * the binary encoding of numbers used by computers: n := \Sum n_i * 2^i, 2387 * (where: n_i \elem {0, 1}, the binary vector representing n), 2388 * we find: x^n := x^(\Sum n_i * 2^i) := \Prod x^(n_i * 2^i), which is 2389 * of course trivially computable in O(log_2 n), the length of our binary 2390 * vector. 2391 */ 2392static unsigned long 2393fixed_power_int(unsigned long x, unsigned int frac_bits, unsigned int n) 2394{ 2395 unsigned long result = 1UL << frac_bits; 2396 2397 if (n) for (;;) { 2398 if (n & 1) { 2399 result *= x; 2400 result += 1UL << (frac_bits - 1); 2401 result >>= frac_bits; 2402 } 2403 n >>= 1; 2404 if (!n) 2405 break; 2406 x *= x; 2407 x += 1UL << (frac_bits - 1); 2408 x >>= frac_bits; 2409 } 2410 2411 return result; 2412} 2413 2414/* 2415 * a1 = a0 * e + a * (1 - e) 2416 * 2417 * a2 = a1 * e + a * (1 - e) 2418 * = (a0 * e + a * (1 - e)) * e + a * (1 - e) 2419 * = a0 * e^2 + a * (1 - e) * (1 + e) 2420 * 2421 * a3 = a2 * e + a * (1 - e) 2422 * = (a0 * e^2 + a * (1 - e) * (1 + e)) * e + a * (1 - e) 2423 * = a0 * e^3 + a * (1 - e) * (1 + e + e^2) 2424 * 2425 * ... 2426 * 2427 * an = a0 * e^n + a * (1 - e) * (1 + e + ... + e^n-1) [1] 2428 * = a0 * e^n + a * (1 - e) * (1 - e^n)/(1 - e) 2429 * = a0 * e^n + a * (1 - e^n) 2430 * 2431 * [1] application of the geometric series: 2432 * 2433 * n 1 - x^(n+1) 2434 * S_n := \Sum x^i = ------------- 2435 * i=0 1 - x 2436 */ 2437static unsigned long 2438calc_load_n(unsigned long load, unsigned long exp, 2439 unsigned long active, unsigned int n) 2440{ 2441 2442 return calc_load(load, fixed_power_int(exp, FSHIFT, n), active); 2443} 2444 2445/* 2446 * NO_HZ can leave us missing all per-cpu ticks calling 2447 * calc_load_account_active(), but since an idle CPU folds its delta into 2448 * calc_load_tasks_idle per calc_load_account_idle(), all we need to do is fold 2449 * in the pending idle delta if our idle period crossed a load cycle boundary. 2450 * 2451 * Once we've updated the global active value, we need to apply the exponential 2452 * weights adjusted to the number of cycles missed. 2453 */ 2454static void calc_global_nohz(void) 2455{ 2456 long delta, active, n; 2457 2458 if (!time_before(jiffies, calc_load_update + 10)) { 2459 /* 2460 * Catch-up, fold however many we are behind still 2461 */ 2462 delta = jiffies - calc_load_update - 10; 2463 n = 1 + (delta / LOAD_FREQ); 2464 2465 active = atomic_long_read(&calc_load_tasks); 2466 active = active > 0 ? active * FIXED_1 : 0; 2467 2468 avenrun[0] = calc_load_n(avenrun[0], EXP_1, active, n); 2469 avenrun[1] = calc_load_n(avenrun[1], EXP_5, active, n); 2470 avenrun[2] = calc_load_n(avenrun[2], EXP_15, active, n); 2471 2472 calc_load_update += n * LOAD_FREQ; 2473 } 2474 2475 /* 2476 * Flip the idle index... 2477 * 2478 * Make sure we first write the new time then flip the index, so that 2479 * calc_load_write_idx() will see the new time when it reads the new 2480 * index, this avoids a double flip messing things up. 2481 */ 2482 smp_wmb(); 2483 calc_load_idx++; 2484} 2485#else /* !CONFIG_NO_HZ */ 2486 2487static inline long calc_load_fold_idle(void) { return 0; } 2488static inline void calc_global_nohz(void) { } 2489 2490#endif /* CONFIG_NO_HZ */ 2491 2492/* 2493 * calc_load - update the avenrun load estimates 10 ticks after the 2494 * CPUs have updated calc_load_tasks. 2495 */ 2496void calc_global_load(unsigned long ticks) 2497{ 2498 long active, delta; 2499 2500 if (time_before(jiffies, calc_load_update + 10)) 2501 return; 2502 2503 /* 2504 * Fold the 'old' idle-delta to include all NO_HZ cpus. 2505 */ 2506 delta = calc_load_fold_idle(); 2507 if (delta) 2508 atomic_long_add(delta, &calc_load_tasks); 2509 2510 active = atomic_long_read(&calc_load_tasks); 2511 active = active > 0 ? active * FIXED_1 : 0; 2512 2513 avenrun[0] = calc_load(avenrun[0], EXP_1, active); 2514 avenrun[1] = calc_load(avenrun[1], EXP_5, active); 2515 avenrun[2] = calc_load(avenrun[2], EXP_15, active); 2516 2517 calc_load_update += LOAD_FREQ; 2518 2519 /* 2520 * In case we idled for multiple LOAD_FREQ intervals, catch up in bulk. 2521 */ 2522 calc_global_nohz(); 2523} 2524 2525/* 2526 * Called from update_cpu_load() to periodically update this CPU's 2527 * active count. 2528 */ 2529static void calc_load_account_active(struct rq *this_rq) 2530{ 2531 long delta; 2532 2533 if (time_before(jiffies, this_rq->calc_load_update)) 2534 return; 2535 2536 delta = calc_load_fold_active(this_rq); 2537 if (delta) 2538 atomic_long_add(delta, &calc_load_tasks); 2539 2540 this_rq->calc_load_update += LOAD_FREQ; 2541} 2542 2543/* 2544 * End of global load-average stuff 2545 */ 2546 2547/* 2548 * The exact cpuload at various idx values, calculated at every tick would be 2549 * load = (2^idx - 1) / 2^idx * load + 1 / 2^idx * cur_load 2550 * 2551 * If a cpu misses updates for n-1 ticks (as it was idle) and update gets called 2552 * on nth tick when cpu may be busy, then we have: 2553 * load = ((2^idx - 1) / 2^idx)^(n-1) * load 2554 * load = (2^idx - 1) / 2^idx) * load + 1 / 2^idx * cur_load 2555 * 2556 * decay_load_missed() below does efficient calculation of 2557 * load = ((2^idx - 1) / 2^idx)^(n-1) * load 2558 * avoiding 0..n-1 loop doing load = ((2^idx - 1) / 2^idx) * load 2559 * 2560 * The calculation is approximated on a 128 point scale. 2561 * degrade_zero_ticks is the number of ticks after which load at any 2562 * particular idx is approximated to be zero. 2563 * degrade_factor is a precomputed table, a row for each load idx. 2564 * Each column corresponds to degradation factor for a power of two ticks, 2565 * based on 128 point scale. 2566 * Example: 2567 * row 2, col 3 (=12) says that the degradation at load idx 2 after 2568 * 8 ticks is 12/128 (which is an approximation of exact factor 3^8/4^8). 2569 * 2570 * With this power of 2 load factors, we can degrade the load n times 2571 * by looking at 1 bits in n and doing as many mult/shift instead of 2572 * n mult/shifts needed by the exact degradation. 2573 */ 2574#define DEGRADE_SHIFT 7 2575static const unsigned char 2576 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128}; 2577static const unsigned char 2578 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = { 2579 {0, 0, 0, 0, 0, 0, 0, 0}, 2580 {64, 32, 8, 0, 0, 0, 0, 0}, 2581 {96, 72, 40, 12, 1, 0, 0}, 2582 {112, 98, 75, 43, 15, 1, 0}, 2583 {120, 112, 98, 76, 45, 16, 2} }; 2584 2585/* 2586 * Update cpu_load for any missed ticks, due to tickless idle. The backlog 2587 * would be when CPU is idle and so we just decay the old load without 2588 * adding any new load. 2589 */ 2590static unsigned long 2591decay_load_missed(unsigned long load, unsigned long missed_updates, int idx) 2592{ 2593 int j = 0; 2594 2595 if (!missed_updates) 2596 return load; 2597 2598 if (missed_updates >= degrade_zero_ticks[idx]) 2599 return 0; 2600 2601 if (idx == 1) 2602 return load >> missed_updates; 2603 2604 while (missed_updates) { 2605 if (missed_updates % 2) 2606 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT; 2607 2608 missed_updates >>= 1; 2609 j++; 2610 } 2611 return load; 2612} 2613 2614/* 2615 * Update rq->cpu_load[] statistics. This function is usually called every 2616 * scheduler tick (TICK_NSEC). With tickless idle this will not be called 2617 * every tick. We fix it up based on jiffies. 2618 */ 2619static void __update_cpu_load(struct rq *this_rq, unsigned long this_load, 2620 unsigned long pending_updates) 2621{ 2622 int i, scale; 2623 2624 this_rq->nr_load_updates++; 2625 2626 /* Update our load: */ 2627 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */ 2628 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) { 2629 unsigned long old_load, new_load; 2630 2631 /* scale is effectively 1 << i now, and >> i divides by scale */ 2632 2633 old_load = this_rq->cpu_load[i]; 2634 old_load = decay_load_missed(old_load, pending_updates - 1, i); 2635 new_load = this_load; 2636 /* 2637 * Round up the averaging division if load is increasing. This 2638 * prevents us from getting stuck on 9 if the load is 10, for 2639 * example. 2640 */ 2641 if (new_load > old_load) 2642 new_load += scale - 1; 2643 2644 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i; 2645 } 2646 2647 sched_avg_update(this_rq); 2648} 2649 2650#ifdef CONFIG_NO_HZ 2651/* 2652 * There is no sane way to deal with nohz on smp when using jiffies because the 2653 * cpu doing the jiffies update might drift wrt the cpu doing the jiffy reading 2654 * causing off-by-one errors in observed deltas; {0,2} instead of {1,1}. 2655 * 2656 * Therefore we cannot use the delta approach from the regular tick since that 2657 * would seriously skew the load calculation. However we'll make do for those 2658 * updates happening while idle (nohz_idle_balance) or coming out of idle 2659 * (tick_nohz_idle_exit). 2660 * 2661 * This means we might still be one tick off for nohz periods. 2662 */ 2663 2664/* 2665 * Called from nohz_idle_balance() to update the load ratings before doing the 2666 * idle balance. 2667 */ 2668void update_idle_cpu_load(struct rq *this_rq) 2669{ 2670 unsigned long curr_jiffies = ACCESS_ONCE(jiffies); 2671 unsigned long load = this_rq->load.weight; 2672 unsigned long pending_updates; 2673 2674 /* 2675 * bail if there's load or we're actually up-to-date. 2676 */ 2677 if (load || curr_jiffies == this_rq->last_load_update_tick) 2678 return; 2679 2680 pending_updates = curr_jiffies - this_rq->last_load_update_tick; 2681 this_rq->last_load_update_tick = curr_jiffies; 2682 2683 __update_cpu_load(this_rq, load, pending_updates); 2684} 2685 2686/* 2687 * Called from tick_nohz_idle_exit() -- try and fix up the ticks we missed. 2688 */ 2689void update_cpu_load_nohz(void) 2690{ 2691 struct rq *this_rq = this_rq(); 2692 unsigned long curr_jiffies = ACCESS_ONCE(jiffies); 2693 unsigned long pending_updates; 2694 2695 if (curr_jiffies == this_rq->last_load_update_tick) 2696 return; 2697 2698 raw_spin_lock(&this_rq->lock); 2699 pending_updates = curr_jiffies - this_rq->last_load_update_tick; 2700 if (pending_updates) { 2701 this_rq->last_load_update_tick = curr_jiffies; 2702 /* 2703 * We were idle, this means load 0, the current load might be 2704 * !0 due to remote wakeups and the sort. 2705 */ 2706 __update_cpu_load(this_rq, 0, pending_updates); 2707 } 2708 raw_spin_unlock(&this_rq->lock); 2709} 2710#endif /* CONFIG_NO_HZ */ 2711 2712/* 2713 * Called from scheduler_tick() 2714 */ 2715static void update_cpu_load_active(struct rq *this_rq) 2716{ 2717 /* 2718 * See the mess around update_idle_cpu_load() / update_cpu_load_nohz(). 2719 */ 2720 this_rq->last_load_update_tick = jiffies; 2721 __update_cpu_load(this_rq, this_rq->load.weight, 1); 2722 2723 calc_load_account_active(this_rq); 2724} 2725 2726#ifdef CONFIG_SMP 2727 2728/* 2729 * sched_exec - execve() is a valuable balancing opportunity, because at 2730 * this point the task has the smallest effective memory and cache footprint. 2731 */ 2732void sched_exec(void) 2733{ 2734 struct task_struct *p = current; 2735 unsigned long flags; 2736 int dest_cpu; 2737 2738 raw_spin_lock_irqsave(&p->pi_lock, flags); 2739 dest_cpu = p->sched_class->select_task_rq(p, SD_BALANCE_EXEC, 0); 2740 if (dest_cpu == smp_processor_id()) 2741 goto unlock; 2742 2743 if (likely(cpu_active(dest_cpu))) { 2744 struct migration_arg arg = { p, dest_cpu }; 2745 2746 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 2747 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg); 2748 return; 2749 } 2750unlock: 2751 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 2752} 2753 2754#endif 2755 2756DEFINE_PER_CPU(struct kernel_stat, kstat); 2757DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat); 2758 2759EXPORT_PER_CPU_SYMBOL(kstat); 2760EXPORT_PER_CPU_SYMBOL(kernel_cpustat); 2761 2762/* 2763 * Return any ns on the sched_clock that have not yet been accounted in 2764 * @p in case that task is currently running. 2765 * 2766 * Called with task_rq_lock() held on @rq. 2767 */ 2768static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq) 2769{ 2770 u64 ns = 0; 2771 2772 if (task_current(rq, p)) { 2773 update_rq_clock(rq); 2774 ns = rq->clock_task - p->se.exec_start; 2775 if ((s64)ns < 0) 2776 ns = 0; 2777 } 2778 2779 return ns; 2780} 2781 2782unsigned long long task_delta_exec(struct task_struct *p) 2783{ 2784 unsigned long flags; 2785 struct rq *rq; 2786 u64 ns = 0; 2787 2788 rq = task_rq_lock(p, &flags); 2789 ns = do_task_delta_exec(p, rq); 2790 task_rq_unlock(rq, p, &flags); 2791 2792 return ns; 2793} 2794 2795/* 2796 * Return accounted runtime for the task. 2797 * In case the task is currently running, return the runtime plus current's 2798 * pending runtime that have not been accounted yet. 2799 */ 2800unsigned long long task_sched_runtime(struct task_struct *p) 2801{ 2802 unsigned long flags; 2803 struct rq *rq; 2804 u64 ns = 0; 2805 2806 rq = task_rq_lock(p, &flags); 2807 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq); 2808 task_rq_unlock(rq, p, &flags); 2809 2810 return ns; 2811} 2812 2813#ifdef CONFIG_CGROUP_CPUACCT 2814struct cgroup_subsys cpuacct_subsys; 2815struct cpuacct root_cpuacct; 2816#endif 2817 2818static inline void task_group_account_field(struct task_struct *p, int index, 2819 u64 tmp) 2820{ 2821#ifdef CONFIG_CGROUP_CPUACCT 2822 struct kernel_cpustat *kcpustat; 2823 struct cpuacct *ca; 2824#endif 2825 /* 2826 * Since all updates are sure to touch the root cgroup, we 2827 * get ourselves ahead and touch it first. If the root cgroup 2828 * is the only cgroup, then nothing else should be necessary. 2829 * 2830 */ 2831 __get_cpu_var(kernel_cpustat).cpustat[index] += tmp; 2832 2833#ifdef CONFIG_CGROUP_CPUACCT 2834 if (unlikely(!cpuacct_subsys.active)) 2835 return; 2836 2837 rcu_read_lock(); 2838 ca = task_ca(p); 2839 while (ca && (ca != &root_cpuacct)) { 2840 kcpustat = this_cpu_ptr(ca->cpustat); 2841 kcpustat->cpustat[index] += tmp; 2842 ca = parent_ca(ca); 2843 } 2844 rcu_read_unlock(); 2845#endif 2846} 2847 2848 2849/* 2850 * Account user cpu time to a process. 2851 * @p: the process that the cpu time gets accounted to 2852 * @cputime: the cpu time spent in user space since the last update 2853 * @cputime_scaled: cputime scaled by cpu frequency 2854 */ 2855void account_user_time(struct task_struct *p, cputime_t cputime, 2856 cputime_t cputime_scaled) 2857{ 2858 int index; 2859 2860 /* Add user time to process. */ 2861 p->utime += cputime; 2862 p->utimescaled += cputime_scaled; 2863 account_group_user_time(p, cputime); 2864 2865 index = (TASK_NICE(p) > 0) ? CPUTIME_NICE : CPUTIME_USER; 2866 2867 /* Add user time to cpustat. */ 2868 task_group_account_field(p, index, (__force u64) cputime); 2869 2870 /* Account for user time used */ 2871 acct_update_integrals(p); 2872} 2873 2874/* 2875 * Account guest cpu time to a process. 2876 * @p: the process that the cpu time gets accounted to 2877 * @cputime: the cpu time spent in virtual machine since the last update 2878 * @cputime_scaled: cputime scaled by cpu frequency 2879 */ 2880static void account_guest_time(struct task_struct *p, cputime_t cputime, 2881 cputime_t cputime_scaled) 2882{ 2883 u64 *cpustat = kcpustat_this_cpu->cpustat; 2884 2885 /* Add guest time to process. */ 2886 p->utime += cputime; 2887 p->utimescaled += cputime_scaled; 2888 account_group_user_time(p, cputime); 2889 p->gtime += cputime; 2890 2891 /* Add guest time to cpustat. */ 2892 if (TASK_NICE(p) > 0) { 2893 cpustat[CPUTIME_NICE] += (__force u64) cputime; 2894 cpustat[CPUTIME_GUEST_NICE] += (__force u64) cputime; 2895 } else { 2896 cpustat[CPUTIME_USER] += (__force u64) cputime; 2897 cpustat[CPUTIME_GUEST] += (__force u64) cputime; 2898 } 2899} 2900 2901/* 2902 * Account system cpu time to a process and desired cpustat field 2903 * @p: the process that the cpu time gets accounted to 2904 * @cputime: the cpu time spent in kernel space since the last update 2905 * @cputime_scaled: cputime scaled by cpu frequency 2906 * @target_cputime64: pointer to cpustat field that has to be updated 2907 */ 2908static inline 2909void __account_system_time(struct task_struct *p, cputime_t cputime, 2910 cputime_t cputime_scaled, int index) 2911{ 2912 /* Add system time to process. */ 2913 p->stime += cputime; 2914 p->stimescaled += cputime_scaled; 2915 account_group_system_time(p, cputime); 2916 2917 /* Add system time to cpustat. */ 2918 task_group_account_field(p, index, (__force u64) cputime); 2919 2920 /* Account for system time used */ 2921 acct_update_integrals(p); 2922} 2923 2924/* 2925 * Account system cpu time to a process. 2926 * @p: the process that the cpu time gets accounted to 2927 * @hardirq_offset: the offset to subtract from hardirq_count() 2928 * @cputime: the cpu time spent in kernel space since the last update 2929 * @cputime_scaled: cputime scaled by cpu frequency 2930 */ 2931void account_system_time(struct task_struct *p, int hardirq_offset, 2932 cputime_t cputime, cputime_t cputime_scaled) 2933{ 2934 int index; 2935 2936 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) { 2937 account_guest_time(p, cputime, cputime_scaled); 2938 return; 2939 } 2940 2941 if (hardirq_count() - hardirq_offset) 2942 index = CPUTIME_IRQ; 2943 else if (in_serving_softirq()) 2944 index = CPUTIME_SOFTIRQ; 2945 else 2946 index = CPUTIME_SYSTEM; 2947 2948 __account_system_time(p, cputime, cputime_scaled, index); 2949} 2950 2951/* 2952 * Account for involuntary wait time. 2953 * @cputime: the cpu time spent in involuntary wait 2954 */ 2955void account_steal_time(cputime_t cputime) 2956{ 2957 u64 *cpustat = kcpustat_this_cpu->cpustat; 2958 2959 cpustat[CPUTIME_STEAL] += (__force u64) cputime; 2960} 2961 2962/* 2963 * Account for idle time. 2964 * @cputime: the cpu time spent in idle wait 2965 */ 2966void account_idle_time(cputime_t cputime) 2967{ 2968 u64 *cpustat = kcpustat_this_cpu->cpustat; 2969 struct rq *rq = this_rq(); 2970 2971 if (atomic_read(&rq->nr_iowait) > 0) 2972 cpustat[CPUTIME_IOWAIT] += (__force u64) cputime; 2973 else 2974 cpustat[CPUTIME_IDLE] += (__force u64) cputime; 2975} 2976 2977static __always_inline bool steal_account_process_tick(void) 2978{ 2979#ifdef CONFIG_PARAVIRT 2980 if (static_key_false(¶virt_steal_enabled)) { 2981 u64 steal, st = 0; 2982 2983 steal = paravirt_steal_clock(smp_processor_id()); 2984 steal -= this_rq()->prev_steal_time; 2985 2986 st = steal_ticks(steal); 2987 this_rq()->prev_steal_time += st * TICK_NSEC; 2988 2989 account_steal_time(st); 2990 return st; 2991 } 2992#endif 2993 return false; 2994} 2995 2996#ifndef CONFIG_VIRT_CPU_ACCOUNTING 2997 2998#ifdef CONFIG_IRQ_TIME_ACCOUNTING 2999/* 3000 * Account a tick to a process and cpustat 3001 * @p: the process that the cpu time gets accounted to 3002 * @user_tick: is the tick from userspace 3003 * @rq: the pointer to rq 3004 * 3005 * Tick demultiplexing follows the order 3006 * - pending hardirq update 3007 * - pending softirq update 3008 * - user_time 3009 * - idle_time 3010 * - system time 3011 * - check for guest_time 3012 * - else account as system_time 3013 * 3014 * Check for hardirq is done both for system and user time as there is 3015 * no timer going off while we are on hardirq and hence we may never get an 3016 * opportunity to update it solely in system time. 3017 * p->stime and friends are only updated on system time and not on irq 3018 * softirq as those do not count in task exec_runtime any more. 3019 */ 3020static void irqtime_account_process_tick(struct task_struct *p, int user_tick, 3021 struct rq *rq) 3022{ 3023 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy); 3024 u64 *cpustat = kcpustat_this_cpu->cpustat; 3025 3026 if (steal_account_process_tick()) 3027 return; 3028 3029 if (irqtime_account_hi_update()) { 3030 cpustat[CPUTIME_IRQ] += (__force u64) cputime_one_jiffy; 3031 } else if (irqtime_account_si_update()) { 3032 cpustat[CPUTIME_SOFTIRQ] += (__force u64) cputime_one_jiffy; 3033 } else if (this_cpu_ksoftirqd() == p) { 3034 /* 3035 * ksoftirqd time do not get accounted in cpu_softirq_time. 3036 * So, we have to handle it separately here. 3037 * Also, p->stime needs to be updated for ksoftirqd. 3038 */ 3039 __account_system_time(p, cputime_one_jiffy, one_jiffy_scaled, 3040 CPUTIME_SOFTIRQ); 3041 } else if (user_tick) { 3042 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled); 3043 } else if (p == rq->idle) { 3044 account_idle_time(cputime_one_jiffy); 3045 } else if (p->flags & PF_VCPU) { /* System time or guest time */ 3046 account_guest_time(p, cputime_one_jiffy, one_jiffy_scaled); 3047 } else { 3048 __account_system_time(p, cputime_one_jiffy, one_jiffy_scaled, 3049 CPUTIME_SYSTEM); 3050 } 3051} 3052 3053static void irqtime_account_idle_ticks(int ticks) 3054{ 3055 int i; 3056 struct rq *rq = this_rq(); 3057 3058 for (i = 0; i < ticks; i++) 3059 irqtime_account_process_tick(current, 0, rq); 3060} 3061#else /* CONFIG_IRQ_TIME_ACCOUNTING */ 3062static void irqtime_account_idle_ticks(int ticks) {} 3063static void irqtime_account_process_tick(struct task_struct *p, int user_tick, 3064 struct rq *rq) {} 3065#endif /* CONFIG_IRQ_TIME_ACCOUNTING */ 3066 3067/* 3068 * Account a single tick of cpu time. 3069 * @p: the process that the cpu time gets accounted to 3070 * @user_tick: indicates if the tick is a user or a system tick 3071 */ 3072void account_process_tick(struct task_struct *p, int user_tick) 3073{ 3074 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy); 3075 struct rq *rq = this_rq(); 3076 3077 if (sched_clock_irqtime) { 3078 irqtime_account_process_tick(p, user_tick, rq); 3079 return; 3080 } 3081 3082 if (steal_account_process_tick()) 3083 return; 3084 3085 if (user_tick) 3086 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled); 3087 else if ((p != rq->idle) || (irq_count() != HARDIRQ_OFFSET)) 3088 account_system_time(p, HARDIRQ_OFFSET, cputime_one_jiffy, 3089 one_jiffy_scaled); 3090 else 3091 account_idle_time(cputime_one_jiffy); 3092} 3093 3094/* 3095 * Account multiple ticks of steal time. 3096 * @p: the process from which the cpu time has been stolen 3097 * @ticks: number of stolen ticks 3098 */ 3099void account_steal_ticks(unsigned long ticks) 3100{ 3101 account_steal_time(jiffies_to_cputime(ticks)); 3102} 3103 3104/* 3105 * Account multiple ticks of idle time. 3106 * @ticks: number of stolen ticks 3107 */ 3108void account_idle_ticks(unsigned long ticks) 3109{ 3110 3111 if (sched_clock_irqtime) { 3112 irqtime_account_idle_ticks(ticks); 3113 return; 3114 } 3115 3116 account_idle_time(jiffies_to_cputime(ticks)); 3117} 3118 3119#endif 3120 3121/* 3122 * Use precise platform statistics if available: 3123 */ 3124#ifdef CONFIG_VIRT_CPU_ACCOUNTING 3125void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st) 3126{ 3127 *ut = p->utime; 3128 *st = p->stime; 3129} 3130 3131void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st) 3132{ 3133 struct task_cputime cputime; 3134 3135 thread_group_cputime(p, &cputime); 3136 3137 *ut = cputime.utime; 3138 *st = cputime.stime; 3139} 3140#else 3141 3142#ifndef nsecs_to_cputime 3143# define nsecs_to_cputime(__nsecs) nsecs_to_jiffies(__nsecs) 3144#endif 3145 3146static cputime_t scale_utime(cputime_t utime, cputime_t rtime, cputime_t total) 3147{ 3148 u64 temp = (__force u64) rtime; 3149 3150 temp *= (__force u64) utime; 3151 3152 if (sizeof(cputime_t) == 4) 3153 temp = div_u64(temp, (__force u32) total); 3154 else 3155 temp = div64_u64(temp, (__force u64) total); 3156 3157 return (__force cputime_t) temp; 3158} 3159 3160void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st) 3161{ 3162 cputime_t rtime, utime = p->utime, total = utime + p->stime; 3163 3164 /* 3165 * Use CFS's precise accounting: 3166 */ 3167 rtime = nsecs_to_cputime(p->se.sum_exec_runtime); 3168 3169 if (total) 3170 utime = scale_utime(utime, rtime, total); 3171 else 3172 utime = rtime; 3173 3174 /* 3175 * Compare with previous values, to keep monotonicity: 3176 */ 3177 p->prev_utime = max(p->prev_utime, utime); 3178 p->prev_stime = max(p->prev_stime, rtime - p->prev_utime); 3179 3180 *ut = p->prev_utime; 3181 *st = p->prev_stime; 3182} 3183 3184/* 3185 * Must be called with siglock held. 3186 */ 3187void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st) 3188{ 3189 struct signal_struct *sig = p->signal; 3190 struct task_cputime cputime; 3191 cputime_t rtime, utime, total; 3192 3193 thread_group_cputime(p, &cputime); 3194 3195 total = cputime.utime + cputime.stime; 3196 rtime = nsecs_to_cputime(cputime.sum_exec_runtime); 3197 3198 if (total) 3199 utime = scale_utime(cputime.utime, rtime, total); 3200 else 3201 utime = rtime; 3202 3203 sig->prev_utime = max(sig->prev_utime, utime); 3204 sig->prev_stime = max(sig->prev_stime, rtime - sig->prev_utime); 3205 3206 *ut = sig->prev_utime; 3207 *st = sig->prev_stime; 3208} 3209#endif 3210 3211/* 3212 * This function gets called by the timer code, with HZ frequency. 3213 * We call it with interrupts disabled. 3214 */ 3215void scheduler_tick(void) 3216{ 3217 int cpu = smp_processor_id(); 3218 struct rq *rq = cpu_rq(cpu); 3219 struct task_struct *curr = rq->curr; 3220 3221 sched_clock_tick(); 3222 3223 raw_spin_lock(&rq->lock); 3224 update_rq_clock(rq); 3225 update_cpu_load_active(rq); 3226 curr->sched_class->task_tick(rq, curr, 0); 3227 raw_spin_unlock(&rq->lock); 3228 3229 perf_event_task_tick(); 3230 3231#ifdef CONFIG_SMP 3232 rq->idle_balance = idle_cpu(cpu); 3233 trigger_load_balance(rq, cpu); 3234#endif 3235} 3236 3237notrace unsigned long get_parent_ip(unsigned long addr) 3238{ 3239 if (in_lock_functions(addr)) { 3240 addr = CALLER_ADDR2; 3241 if (in_lock_functions(addr)) 3242 addr = CALLER_ADDR3; 3243 } 3244 return addr; 3245} 3246 3247#if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \ 3248 defined(CONFIG_PREEMPT_TRACER)) 3249 3250void __kprobes add_preempt_count(int val) 3251{ 3252#ifdef CONFIG_DEBUG_PREEMPT 3253 /* 3254 * Underflow? 3255 */ 3256 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0))) 3257 return; 3258#endif 3259 preempt_count() += val; 3260#ifdef CONFIG_DEBUG_PREEMPT 3261 /* 3262 * Spinlock count overflowing soon? 3263 */ 3264 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >= 3265 PREEMPT_MASK - 10); 3266#endif 3267 if (preempt_count() == val) 3268 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1)); 3269} 3270EXPORT_SYMBOL(add_preempt_count); 3271 3272void __kprobes sub_preempt_count(int val) 3273{ 3274#ifdef CONFIG_DEBUG_PREEMPT 3275 /* 3276 * Underflow? 3277 */ 3278 if (DEBUG_LOCKS_WARN_ON(val > preempt_count())) 3279 return; 3280 /* 3281 * Is the spinlock portion underflowing? 3282 */ 3283 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) && 3284 !(preempt_count() & PREEMPT_MASK))) 3285 return; 3286#endif 3287 3288 if (preempt_count() == val) 3289 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1)); 3290 preempt_count() -= val; 3291} 3292EXPORT_SYMBOL(sub_preempt_count); 3293 3294#endif 3295 3296/* 3297 * Print scheduling while atomic bug: 3298 */ 3299static noinline void __schedule_bug(struct task_struct *prev) 3300{ 3301 if (oops_in_progress) 3302 return; 3303 3304 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n", 3305 prev->comm, prev->pid, preempt_count()); 3306 3307 debug_show_held_locks(prev); 3308 print_modules(); 3309 if (irqs_disabled()) 3310 print_irqtrace_events(prev); 3311 dump_stack(); 3312 add_taint(TAINT_WARN); 3313} 3314 3315/* 3316 * Various schedule()-time debugging checks and statistics: 3317 */ 3318static inline void schedule_debug(struct task_struct *prev) 3319{ 3320 /* 3321 * Test if we are atomic. Since do_exit() needs to call into 3322 * schedule() atomically, we ignore that path for now. 3323 * Otherwise, whine if we are scheduling when we should not be. 3324 */ 3325 if (unlikely(in_atomic_preempt_off() && !prev->exit_state)) 3326 __schedule_bug(prev); 3327 rcu_sleep_check(); 3328 3329 profile_hit(SCHED_PROFILING, __builtin_return_address(0)); 3330 3331 schedstat_inc(this_rq(), sched_count); 3332} 3333 3334static void put_prev_task(struct rq *rq, struct task_struct *prev) 3335{ 3336 if (prev->on_rq || rq->skip_clock_update < 0) 3337 update_rq_clock(rq); 3338 prev->sched_class->put_prev_task(rq, prev); 3339} 3340 3341/* 3342 * Pick up the highest-prio task: 3343 */ 3344static inline struct task_struct * 3345pick_next_task(struct rq *rq) 3346{ 3347 const struct sched_class *class; 3348 struct task_struct *p; 3349 3350 /* 3351 * Optimization: we know that if all tasks are in 3352 * the fair class we can call that function directly: 3353 */ 3354 if (likely(rq->nr_running == rq->cfs.h_nr_running)) { 3355 p = fair_sched_class.pick_next_task(rq); 3356 if (likely(p)) 3357 return p; 3358 } 3359 3360 for_each_class(class) { 3361 p = class->pick_next_task(rq); 3362 if (p) 3363 return p; 3364 } 3365 3366 BUG(); /* the idle class will always have a runnable task */ 3367} 3368 3369/* 3370 * __schedule() is the main scheduler function. 3371 */ 3372static void __sched __schedule(void) 3373{ 3374 struct task_struct *prev, *next; 3375 unsigned long *switch_count; 3376 struct rq *rq; 3377 int cpu; 3378 3379need_resched: 3380 preempt_disable(); 3381 cpu = smp_processor_id(); 3382 rq = cpu_rq(cpu); 3383 rcu_note_context_switch(cpu); 3384 prev = rq->curr; 3385 3386 schedule_debug(prev); 3387 3388 if (sched_feat(HRTICK)) 3389 hrtick_clear(rq); 3390 3391 raw_spin_lock_irq(&rq->lock); 3392 3393 switch_count = &prev->nivcsw; 3394 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) { 3395 if (unlikely(signal_pending_state(prev->state, prev))) { 3396 prev->state = TASK_RUNNING; 3397 } else { 3398 deactivate_task(rq, prev, DEQUEUE_SLEEP); 3399 prev->on_rq = 0; 3400 3401 /* 3402 * If a worker went to sleep, notify and ask workqueue 3403 * whether it wants to wake up a task to maintain 3404 * concurrency. 3405 */ 3406 if (prev->flags & PF_WQ_WORKER) { 3407 struct task_struct *to_wakeup; 3408 3409 to_wakeup = wq_worker_sleeping(prev, cpu); 3410 if (to_wakeup) 3411 try_to_wake_up_local(to_wakeup); 3412 } 3413 } 3414 switch_count = &prev->nvcsw; 3415 } 3416 3417 pre_schedule(rq, prev); 3418 3419 if (unlikely(!rq->nr_running)) 3420 idle_balance(cpu, rq); 3421 3422 put_prev_task(rq, prev); 3423 next = pick_next_task(rq); 3424 clear_tsk_need_resched(prev); 3425 rq->skip_clock_update = 0; 3426 3427 if (likely(prev != next)) { 3428 rq->nr_switches++; 3429 rq->curr = next; 3430 ++*switch_count; 3431 3432 context_switch(rq, prev, next); /* unlocks the rq */ 3433 /* 3434 * The context switch have flipped the stack from under us 3435 * and restored the local variables which were saved when 3436 * this task called schedule() in the past. prev == current 3437 * is still correct, but it can be moved to another cpu/rq. 3438 */ 3439 cpu = smp_processor_id(); 3440 rq = cpu_rq(cpu); 3441 } else 3442 raw_spin_unlock_irq(&rq->lock); 3443 3444 post_schedule(rq); 3445 3446 sched_preempt_enable_no_resched(); 3447 if (need_resched()) 3448 goto need_resched; 3449} 3450 3451static inline void sched_submit_work(struct task_struct *tsk) 3452{ 3453 if (!tsk->state || tsk_is_pi_blocked(tsk)) 3454 return; 3455 /* 3456 * If we are going to sleep and we have plugged IO queued, 3457 * make sure to submit it to avoid deadlocks. 3458 */ 3459 if (blk_needs_flush_plug(tsk)) 3460 blk_schedule_flush_plug(tsk); 3461} 3462 3463asmlinkage void __sched schedule(void) 3464{ 3465 struct task_struct *tsk = current; 3466 3467 sched_submit_work(tsk); 3468 __schedule(); 3469} 3470EXPORT_SYMBOL(schedule); 3471 3472#ifdef CONFIG_RCU_USER_QS 3473asmlinkage void __sched schedule_user(void) 3474{ 3475 /* 3476 * If we come here after a random call to set_need_resched(), 3477 * or we have been woken up remotely but the IPI has not yet arrived, 3478 * we haven't yet exited the RCU idle mode. Do it here manually until 3479 * we find a better solution. 3480 */ 3481 rcu_user_exit(); 3482 schedule(); 3483 rcu_user_enter(); 3484} 3485#endif 3486 3487/** 3488 * schedule_preempt_disabled - called with preemption disabled 3489 * 3490 * Returns with preemption disabled. Note: preempt_count must be 1 3491 */ 3492void __sched schedule_preempt_disabled(void) 3493{ 3494 sched_preempt_enable_no_resched(); 3495 schedule(); 3496 preempt_disable(); 3497} 3498 3499#ifdef CONFIG_MUTEX_SPIN_ON_OWNER 3500 3501static inline bool owner_running(struct mutex *lock, struct task_struct *owner) 3502{ 3503 if (lock->owner != owner) 3504 return false; 3505 3506 /* 3507 * Ensure we emit the owner->on_cpu, dereference _after_ checking 3508 * lock->owner still matches owner, if that fails, owner might 3509 * point to free()d memory, if it still matches, the rcu_read_lock() 3510 * ensures the memory stays valid. 3511 */ 3512 barrier(); 3513 3514 return owner->on_cpu; 3515} 3516 3517/* 3518 * Look out! "owner" is an entirely speculative pointer 3519 * access and not reliable. 3520 */ 3521int mutex_spin_on_owner(struct mutex *lock, struct task_struct *owner) 3522{ 3523 if (!sched_feat(OWNER_SPIN)) 3524 return 0; 3525 3526 rcu_read_lock(); 3527 while (owner_running(lock, owner)) { 3528 if (need_resched()) 3529 break; 3530 3531 arch_mutex_cpu_relax(); 3532 } 3533 rcu_read_unlock(); 3534 3535 /* 3536 * We break out the loop above on need_resched() and when the 3537 * owner changed, which is a sign for heavy contention. Return 3538 * success only when lock->owner is NULL. 3539 */ 3540 return lock->owner == NULL; 3541} 3542#endif 3543 3544#ifdef CONFIG_PREEMPT 3545/* 3546 * this is the entry point to schedule() from in-kernel preemption 3547 * off of preempt_enable. Kernel preemptions off return from interrupt 3548 * occur there and call schedule directly. 3549 */ 3550asmlinkage void __sched notrace preempt_schedule(void) 3551{ 3552 struct thread_info *ti = current_thread_info(); 3553 3554 /* 3555 * If there is a non-zero preempt_count or interrupts are disabled, 3556 * we do not want to preempt the current task. Just return.. 3557 */ 3558 if (likely(ti->preempt_count || irqs_disabled())) 3559 return; 3560 3561 do { 3562 add_preempt_count_notrace(PREEMPT_ACTIVE); 3563 __schedule(); 3564 sub_preempt_count_notrace(PREEMPT_ACTIVE); 3565 3566 /* 3567 * Check again in case we missed a preemption opportunity 3568 * between schedule and now. 3569 */ 3570 barrier(); 3571 } while (need_resched()); 3572} 3573EXPORT_SYMBOL(preempt_schedule); 3574 3575/* 3576 * this is the entry point to schedule() from kernel preemption 3577 * off of irq context. 3578 * Note, that this is called and return with irqs disabled. This will 3579 * protect us against recursive calling from irq. 3580 */ 3581asmlinkage void __sched preempt_schedule_irq(void) 3582{ 3583 struct thread_info *ti = current_thread_info(); 3584 3585 /* Catch callers which need to be fixed */ 3586 BUG_ON(ti->preempt_count || !irqs_disabled()); 3587 3588 rcu_user_exit(); 3589 do { 3590 add_preempt_count(PREEMPT_ACTIVE); 3591 local_irq_enable(); 3592 __schedule(); 3593 local_irq_disable(); 3594 sub_preempt_count(PREEMPT_ACTIVE); 3595 3596 /* 3597 * Check again in case we missed a preemption opportunity 3598 * between schedule and now. 3599 */ 3600 barrier(); 3601 } while (need_resched()); 3602} 3603 3604#endif /* CONFIG_PREEMPT */ 3605 3606int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags, 3607 void *key) 3608{ 3609 return try_to_wake_up(curr->private, mode, wake_flags); 3610} 3611EXPORT_SYMBOL(default_wake_function); 3612 3613/* 3614 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just 3615 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve 3616 * number) then we wake all the non-exclusive tasks and one exclusive task. 3617 * 3618 * There are circumstances in which we can try to wake a task which has already 3619 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns 3620 * zero in this (rare) case, and we handle it by continuing to scan the queue. 3621 */ 3622static void __wake_up_common(wait_queue_head_t *q, unsigned int mode, 3623 int nr_exclusive, int wake_flags, void *key) 3624{ 3625 wait_queue_t *curr, *next; 3626 3627 list_for_each_entry_safe(curr, next, &q->task_list, task_list) { 3628 unsigned flags = curr->flags; 3629 3630 if (curr->func(curr, mode, wake_flags, key) && 3631 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive) 3632 break; 3633 } 3634} 3635 3636/** 3637 * __wake_up - wake up threads blocked on a waitqueue. 3638 * @q: the waitqueue 3639 * @mode: which threads 3640 * @nr_exclusive: how many wake-one or wake-many threads to wake up 3641 * @key: is directly passed to the wakeup function 3642 * 3643 * It may be assumed that this function implies a write memory barrier before 3644 * changing the task state if and only if any tasks are woken up. 3645 */ 3646void __wake_up(wait_queue_head_t *q, unsigned int mode, 3647 int nr_exclusive, void *key) 3648{ 3649 unsigned long flags; 3650 3651 spin_lock_irqsave(&q->lock, flags); 3652 __wake_up_common(q, mode, nr_exclusive, 0, key); 3653 spin_unlock_irqrestore(&q->lock, flags); 3654} 3655EXPORT_SYMBOL(__wake_up); 3656 3657/* 3658 * Same as __wake_up but called with the spinlock in wait_queue_head_t held. 3659 */ 3660void __wake_up_locked(wait_queue_head_t *q, unsigned int mode, int nr) 3661{ 3662 __wake_up_common(q, mode, nr, 0, NULL); 3663} 3664EXPORT_SYMBOL_GPL(__wake_up_locked); 3665 3666void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key) 3667{ 3668 __wake_up_common(q, mode, 1, 0, key); 3669} 3670EXPORT_SYMBOL_GPL(__wake_up_locked_key); 3671 3672/** 3673 * __wake_up_sync_key - wake up threads blocked on a waitqueue. 3674 * @q: the waitqueue 3675 * @mode: which threads 3676 * @nr_exclusive: how many wake-one or wake-many threads to wake up 3677 * @key: opaque value to be passed to wakeup targets 3678 * 3679 * The sync wakeup differs that the waker knows that it will schedule 3680 * away soon, so while the target thread will be woken up, it will not 3681 * be migrated to another CPU - ie. the two threads are 'synchronized' 3682 * with each other. This can prevent needless bouncing between CPUs. 3683 * 3684 * On UP it can prevent extra preemption. 3685 * 3686 * It may be assumed that this function implies a write memory barrier before 3687 * changing the task state if and only if any tasks are woken up. 3688 */ 3689void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode, 3690 int nr_exclusive, void *key) 3691{ 3692 unsigned long flags; 3693 int wake_flags = WF_SYNC; 3694 3695 if (unlikely(!q)) 3696 return; 3697 3698 if (unlikely(!nr_exclusive)) 3699 wake_flags = 0; 3700 3701 spin_lock_irqsave(&q->lock, flags); 3702 __wake_up_common(q, mode, nr_exclusive, wake_flags, key); 3703 spin_unlock_irqrestore(&q->lock, flags); 3704} 3705EXPORT_SYMBOL_GPL(__wake_up_sync_key); 3706 3707/* 3708 * __wake_up_sync - see __wake_up_sync_key() 3709 */ 3710void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive) 3711{ 3712 __wake_up_sync_key(q, mode, nr_exclusive, NULL); 3713} 3714EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */ 3715 3716/** 3717 * complete: - signals a single thread waiting on this completion 3718 * @x: holds the state of this particular completion 3719 * 3720 * This will wake up a single thread waiting on this completion. Threads will be 3721 * awakened in the same order in which they were queued. 3722 * 3723 * See also complete_all(), wait_for_completion() and related routines. 3724 * 3725 * It may be assumed that this function implies a write memory barrier before 3726 * changing the task state if and only if any tasks are woken up. 3727 */ 3728void complete(struct completion *x) 3729{ 3730 unsigned long flags; 3731 3732 spin_lock_irqsave(&x->wait.lock, flags); 3733 x->done++; 3734 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL); 3735 spin_unlock_irqrestore(&x->wait.lock, flags); 3736} 3737EXPORT_SYMBOL(complete); 3738 3739/** 3740 * complete_all: - signals all threads waiting on this completion 3741 * @x: holds the state of this particular completion 3742 * 3743 * This will wake up all threads waiting on this particular completion event. 3744 * 3745 * It may be assumed that this function implies a write memory barrier before 3746 * changing the task state if and only if any tasks are woken up. 3747 */ 3748void complete_all(struct completion *x) 3749{ 3750 unsigned long flags; 3751 3752 spin_lock_irqsave(&x->wait.lock, flags); 3753 x->done += UINT_MAX/2; 3754 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL); 3755 spin_unlock_irqrestore(&x->wait.lock, flags); 3756} 3757EXPORT_SYMBOL(complete_all); 3758 3759static inline long __sched 3760do_wait_for_common(struct completion *x, long timeout, int state) 3761{ 3762 if (!x->done) { 3763 DECLARE_WAITQUEUE(wait, current); 3764 3765 __add_wait_queue_tail_exclusive(&x->wait, &wait); 3766 do { 3767 if (signal_pending_state(state, current)) { 3768 timeout = -ERESTARTSYS; 3769 break; 3770 } 3771 __set_current_state(state); 3772 spin_unlock_irq(&x->wait.lock); 3773 timeout = schedule_timeout(timeout); 3774 spin_lock_irq(&x->wait.lock); 3775 } while (!x->done && timeout); 3776 __remove_wait_queue(&x->wait, &wait); 3777 if (!x->done) 3778 return timeout; 3779 } 3780 x->done--; 3781 return timeout ?: 1; 3782} 3783 3784static long __sched 3785wait_for_common(struct completion *x, long timeout, int state) 3786{ 3787 might_sleep(); 3788 3789 spin_lock_irq(&x->wait.lock); 3790 timeout = do_wait_for_common(x, timeout, state); 3791 spin_unlock_irq(&x->wait.lock); 3792 return timeout; 3793} 3794 3795/** 3796 * wait_for_completion: - waits for completion of a task 3797 * @x: holds the state of this particular completion 3798 * 3799 * This waits to be signaled for completion of a specific task. It is NOT 3800 * interruptible and there is no timeout. 3801 * 3802 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout 3803 * and interrupt capability. Also see complete(). 3804 */ 3805void __sched wait_for_completion(struct completion *x) 3806{ 3807 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE); 3808} 3809EXPORT_SYMBOL(wait_for_completion); 3810 3811/** 3812 * wait_for_completion_timeout: - waits for completion of a task (w/timeout) 3813 * @x: holds the state of this particular completion 3814 * @timeout: timeout value in jiffies 3815 * 3816 * This waits for either a completion of a specific task to be signaled or for a 3817 * specified timeout to expire. The timeout is in jiffies. It is not 3818 * interruptible. 3819 * 3820 * The return value is 0 if timed out, and positive (at least 1, or number of 3821 * jiffies left till timeout) if completed. 3822 */ 3823unsigned long __sched 3824wait_for_completion_timeout(struct completion *x, unsigned long timeout) 3825{ 3826 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE); 3827} 3828EXPORT_SYMBOL(wait_for_completion_timeout); 3829 3830/** 3831 * wait_for_completion_interruptible: - waits for completion of a task (w/intr) 3832 * @x: holds the state of this particular completion 3833 * 3834 * This waits for completion of a specific task to be signaled. It is 3835 * interruptible. 3836 * 3837 * The return value is -ERESTARTSYS if interrupted, 0 if completed. 3838 */ 3839int __sched wait_for_completion_interruptible(struct completion *x) 3840{ 3841 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE); 3842 if (t == -ERESTARTSYS) 3843 return t; 3844 return 0; 3845} 3846EXPORT_SYMBOL(wait_for_completion_interruptible); 3847 3848/** 3849 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr)) 3850 * @x: holds the state of this particular completion 3851 * @timeout: timeout value in jiffies 3852 * 3853 * This waits for either a completion of a specific task to be signaled or for a 3854 * specified timeout to expire. It is interruptible. The timeout is in jiffies. 3855 * 3856 * The return value is -ERESTARTSYS if interrupted, 0 if timed out, 3857 * positive (at least 1, or number of jiffies left till timeout) if completed. 3858 */ 3859long __sched 3860wait_for_completion_interruptible_timeout(struct completion *x, 3861 unsigned long timeout) 3862{ 3863 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE); 3864} 3865EXPORT_SYMBOL(wait_for_completion_interruptible_timeout); 3866 3867/** 3868 * wait_for_completion_killable: - waits for completion of a task (killable) 3869 * @x: holds the state of this particular completion 3870 * 3871 * This waits to be signaled for completion of a specific task. It can be 3872 * interrupted by a kill signal. 3873 * 3874 * The return value is -ERESTARTSYS if interrupted, 0 if completed. 3875 */ 3876int __sched wait_for_completion_killable(struct completion *x) 3877{ 3878 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE); 3879 if (t == -ERESTARTSYS) 3880 return t; 3881 return 0; 3882} 3883EXPORT_SYMBOL(wait_for_completion_killable); 3884 3885/** 3886 * wait_for_completion_killable_timeout: - waits for completion of a task (w/(to,killable)) 3887 * @x: holds the state of this particular completion 3888 * @timeout: timeout value in jiffies 3889 * 3890 * This waits for either a completion of a specific task to be 3891 * signaled or for a specified timeout to expire. It can be 3892 * interrupted by a kill signal. The timeout is in jiffies. 3893 * 3894 * The return value is -ERESTARTSYS if interrupted, 0 if timed out, 3895 * positive (at least 1, or number of jiffies left till timeout) if completed. 3896 */ 3897long __sched 3898wait_for_completion_killable_timeout(struct completion *x, 3899 unsigned long timeout) 3900{ 3901 return wait_for_common(x, timeout, TASK_KILLABLE); 3902} 3903EXPORT_SYMBOL(wait_for_completion_killable_timeout); 3904 3905/** 3906 * try_wait_for_completion - try to decrement a completion without blocking 3907 * @x: completion structure 3908 * 3909 * Returns: 0 if a decrement cannot be done without blocking 3910 * 1 if a decrement succeeded. 3911 * 3912 * If a completion is being used as a counting completion, 3913 * attempt to decrement the counter without blocking. This 3914 * enables us to avoid waiting if the resource the completion 3915 * is protecting is not available. 3916 */ 3917bool try_wait_for_completion(struct completion *x) 3918{ 3919 unsigned long flags; 3920 int ret = 1; 3921 3922 spin_lock_irqsave(&x->wait.lock, flags); 3923 if (!x->done) 3924 ret = 0; 3925 else 3926 x->done--; 3927 spin_unlock_irqrestore(&x->wait.lock, flags); 3928 return ret; 3929} 3930EXPORT_SYMBOL(try_wait_for_completion); 3931 3932/** 3933 * completion_done - Test to see if a completion has any waiters 3934 * @x: completion structure 3935 * 3936 * Returns: 0 if there are waiters (wait_for_completion() in progress) 3937 * 1 if there are no waiters. 3938 * 3939 */ 3940bool completion_done(struct completion *x) 3941{ 3942 unsigned long flags; 3943 int ret = 1; 3944 3945 spin_lock_irqsave(&x->wait.lock, flags); 3946 if (!x->done) 3947 ret = 0; 3948 spin_unlock_irqrestore(&x->wait.lock, flags); 3949 return ret; 3950} 3951EXPORT_SYMBOL(completion_done); 3952 3953static long __sched 3954sleep_on_common(wait_queue_head_t *q, int state, long timeout) 3955{ 3956 unsigned long flags; 3957 wait_queue_t wait; 3958 3959 init_waitqueue_entry(&wait, current); 3960 3961 __set_current_state(state); 3962 3963 spin_lock_irqsave(&q->lock, flags); 3964 __add_wait_queue(q, &wait); 3965 spin_unlock(&q->lock); 3966 timeout = schedule_timeout(timeout); 3967 spin_lock_irq(&q->lock); 3968 __remove_wait_queue(q, &wait); 3969 spin_unlock_irqrestore(&q->lock, flags); 3970 3971 return timeout; 3972} 3973 3974void __sched interruptible_sleep_on(wait_queue_head_t *q) 3975{ 3976 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT); 3977} 3978EXPORT_SYMBOL(interruptible_sleep_on); 3979 3980long __sched 3981interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout) 3982{ 3983 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout); 3984} 3985EXPORT_SYMBOL(interruptible_sleep_on_timeout); 3986 3987void __sched sleep_on(wait_queue_head_t *q) 3988{ 3989 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT); 3990} 3991EXPORT_SYMBOL(sleep_on); 3992 3993long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout) 3994{ 3995 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout); 3996} 3997EXPORT_SYMBOL(sleep_on_timeout); 3998 3999#ifdef CONFIG_RT_MUTEXES 4000 4001/* 4002 * rt_mutex_setprio - set the current priority of a task 4003 * @p: task 4004 * @prio: prio value (kernel-internal form) 4005 * 4006 * This function changes the 'effective' priority of a task. It does 4007 * not touch ->normal_prio like __setscheduler(). 4008 * 4009 * Used by the rt_mutex code to implement priority inheritance logic. 4010 */ 4011void rt_mutex_setprio(struct task_struct *p, int prio) 4012{ 4013 int oldprio, on_rq, running; 4014 struct rq *rq; 4015 const struct sched_class *prev_class; 4016 4017 BUG_ON(prio < 0 || prio > MAX_PRIO); 4018 4019 rq = __task_rq_lock(p); 4020 4021 /* 4022 * Idle task boosting is a nono in general. There is one 4023 * exception, when PREEMPT_RT and NOHZ is active: 4024 * 4025 * The idle task calls get_next_timer_interrupt() and holds 4026 * the timer wheel base->lock on the CPU and another CPU wants 4027 * to access the timer (probably to cancel it). We can safely 4028 * ignore the boosting request, as the idle CPU runs this code 4029 * with interrupts disabled and will complete the lock 4030 * protected section without being interrupted. So there is no 4031 * real need to boost. 4032 */ 4033 if (unlikely(p == rq->idle)) { 4034 WARN_ON(p != rq->curr); 4035 WARN_ON(p->pi_blocked_on); 4036 goto out_unlock; 4037 } 4038 4039 trace_sched_pi_setprio(p, prio); 4040 oldprio = p->prio; 4041 prev_class = p->sched_class; 4042 on_rq = p->on_rq; 4043 running = task_current(rq, p); 4044 if (on_rq) 4045 dequeue_task(rq, p, 0); 4046 if (running) 4047 p->sched_class->put_prev_task(rq, p); 4048 4049 if (rt_prio(prio)) 4050 p->sched_class = &rt_sched_class; 4051 else 4052 p->sched_class = &fair_sched_class; 4053 4054 p->prio = prio; 4055 4056 if (running) 4057 p->sched_class->set_curr_task(rq); 4058 if (on_rq) 4059 enqueue_task(rq, p, oldprio < prio ? ENQUEUE_HEAD : 0); 4060 4061 check_class_changed(rq, p, prev_class, oldprio); 4062out_unlock: 4063 __task_rq_unlock(rq); 4064} 4065#endif 4066void set_user_nice(struct task_struct *p, long nice) 4067{ 4068 int old_prio, delta, on_rq; 4069 unsigned long flags; 4070 struct rq *rq; 4071 4072 if (TASK_NICE(p) == nice || nice < -20 || nice > 19) 4073 return; 4074 /* 4075 * We have to be careful, if called from sys_setpriority(), 4076 * the task might be in the middle of scheduling on another CPU. 4077 */ 4078 rq = task_rq_lock(p, &flags); 4079 /* 4080 * The RT priorities are set via sched_setscheduler(), but we still 4081 * allow the 'normal' nice value to be set - but as expected 4082 * it wont have any effect on scheduling until the task is 4083 * SCHED_FIFO/SCHED_RR: 4084 */ 4085 if (task_has_rt_policy(p)) { 4086 p->static_prio = NICE_TO_PRIO(nice); 4087 goto out_unlock; 4088 } 4089 on_rq = p->on_rq; 4090 if (on_rq) 4091 dequeue_task(rq, p, 0); 4092 4093 p->static_prio = NICE_TO_PRIO(nice); 4094 set_load_weight(p); 4095 old_prio = p->prio; 4096 p->prio = effective_prio(p); 4097 delta = p->prio - old_prio; 4098 4099 if (on_rq) { 4100 enqueue_task(rq, p, 0); 4101 /* 4102 * If the task increased its priority or is running and 4103 * lowered its priority, then reschedule its CPU: 4104 */ 4105 if (delta < 0 || (delta > 0 && task_running(rq, p))) 4106 resched_task(rq->curr); 4107 } 4108out_unlock: 4109 task_rq_unlock(rq, p, &flags); 4110} 4111EXPORT_SYMBOL(set_user_nice); 4112 4113/* 4114 * can_nice - check if a task can reduce its nice value 4115 * @p: task 4116 * @nice: nice value 4117 */ 4118int can_nice(const struct task_struct *p, const int nice) 4119{ 4120 /* convert nice value [19,-20] to rlimit style value [1,40] */ 4121 int nice_rlim = 20 - nice; 4122 4123 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) || 4124 capable(CAP_SYS_NICE)); 4125} 4126 4127#ifdef __ARCH_WANT_SYS_NICE 4128 4129/* 4130 * sys_nice - change the priority of the current process. 4131 * @increment: priority increment 4132 * 4133 * sys_setpriority is a more generic, but much slower function that 4134 * does similar things. 4135 */ 4136SYSCALL_DEFINE1(nice, int, increment) 4137{ 4138 long nice, retval; 4139 4140 /* 4141 * Setpriority might change our priority at the same moment. 4142 * We don't have to worry. Conceptually one call occurs first 4143 * and we have a single winner. 4144 */ 4145 if (increment < -40) 4146 increment = -40; 4147 if (increment > 40) 4148 increment = 40; 4149 4150 nice = TASK_NICE(current) + increment; 4151 if (nice < -20) 4152 nice = -20; 4153 if (nice > 19) 4154 nice = 19; 4155 4156 if (increment < 0 && !can_nice(current, nice)) 4157 return -EPERM; 4158 4159 retval = security_task_setnice(current, nice); 4160 if (retval) 4161 return retval; 4162 4163 set_user_nice(current, nice); 4164 return 0; 4165} 4166 4167#endif 4168 4169/** 4170 * task_prio - return the priority value of a given task. 4171 * @p: the task in question. 4172 * 4173 * This is the priority value as seen by users in /proc. 4174 * RT tasks are offset by -200. Normal tasks are centered 4175 * around 0, value goes from -16 to +15. 4176 */ 4177int task_prio(const struct task_struct *p) 4178{ 4179 return p->prio - MAX_RT_PRIO; 4180} 4181 4182/** 4183 * task_nice - return the nice value of a given task. 4184 * @p: the task in question. 4185 */ 4186int task_nice(const struct task_struct *p) 4187{ 4188 return TASK_NICE(p); 4189} 4190EXPORT_SYMBOL(task_nice); 4191 4192/** 4193 * idle_cpu - is a given cpu idle currently? 4194 * @cpu: the processor in question. 4195 */ 4196int idle_cpu(int cpu) 4197{ 4198 struct rq *rq = cpu_rq(cpu); 4199 4200 if (rq->curr != rq->idle) 4201 return 0; 4202 4203 if (rq->nr_running) 4204 return 0; 4205 4206#ifdef CONFIG_SMP 4207 if (!llist_empty(&rq->wake_list)) 4208 return 0; 4209#endif 4210 4211 return 1; 4212} 4213 4214/** 4215 * idle_task - return the idle task for a given cpu. 4216 * @cpu: the processor in question. 4217 */ 4218struct task_struct *idle_task(int cpu) 4219{ 4220 return cpu_rq(cpu)->idle; 4221} 4222 4223/** 4224 * find_process_by_pid - find a process with a matching PID value. 4225 * @pid: the pid in question. 4226 */ 4227static struct task_struct *find_process_by_pid(pid_t pid) 4228{ 4229 return pid ? find_task_by_vpid(pid) : current; 4230} 4231 4232/* Actually do priority change: must hold rq lock. */ 4233static void 4234__setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio) 4235{ 4236 p->policy = policy; 4237 p->rt_priority = prio; 4238 p->normal_prio = normal_prio(p); 4239 /* we are holding p->pi_lock already */ 4240 p->prio = rt_mutex_getprio(p); 4241 if (rt_prio(p->prio)) 4242 p->sched_class = &rt_sched_class; 4243 else 4244 p->sched_class = &fair_sched_class; 4245 set_load_weight(p); 4246} 4247 4248/* 4249 * check the target process has a UID that matches the current process's 4250 */ 4251static bool check_same_owner(struct task_struct *p) 4252{ 4253 const struct cred *cred = current_cred(), *pcred; 4254 bool match; 4255 4256 rcu_read_lock(); 4257 pcred = __task_cred(p); 4258 match = (uid_eq(cred->euid, pcred->euid) || 4259 uid_eq(cred->euid, pcred->uid)); 4260 rcu_read_unlock(); 4261 return match; 4262} 4263 4264static int __sched_setscheduler(struct task_struct *p, int policy, 4265 const struct sched_param *param, bool user) 4266{ 4267 int retval, oldprio, oldpolicy = -1, on_rq, running; 4268 unsigned long flags; 4269 const struct sched_class *prev_class; 4270 struct rq *rq; 4271 int reset_on_fork; 4272 4273 /* may grab non-irq protected spin_locks */ 4274 BUG_ON(in_interrupt()); 4275recheck: 4276 /* double check policy once rq lock held */ 4277 if (policy < 0) { 4278 reset_on_fork = p->sched_reset_on_fork; 4279 policy = oldpolicy = p->policy; 4280 } else { 4281 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK); 4282 policy &= ~SCHED_RESET_ON_FORK; 4283 4284 if (policy != SCHED_FIFO && policy != SCHED_RR && 4285 policy != SCHED_NORMAL && policy != SCHED_BATCH && 4286 policy != SCHED_IDLE) 4287 return -EINVAL; 4288 } 4289 4290 /* 4291 * Valid priorities for SCHED_FIFO and SCHED_RR are 4292 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL, 4293 * SCHED_BATCH and SCHED_IDLE is 0. 4294 */ 4295 if (param->sched_priority < 0 || 4296 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) || 4297 (!p->mm && param->sched_priority > MAX_RT_PRIO-1)) 4298 return -EINVAL; 4299 if (rt_policy(policy) != (param->sched_priority != 0)) 4300 return -EINVAL; 4301 4302 /* 4303 * Allow unprivileged RT tasks to decrease priority: 4304 */ 4305 if (user && !capable(CAP_SYS_NICE)) { 4306 if (rt_policy(policy)) { 4307 unsigned long rlim_rtprio = 4308 task_rlimit(p, RLIMIT_RTPRIO); 4309 4310 /* can't set/change the rt policy */ 4311 if (policy != p->policy && !rlim_rtprio) 4312 return -EPERM; 4313 4314 /* can't increase priority */ 4315 if (param->sched_priority > p->rt_priority && 4316 param->sched_priority > rlim_rtprio) 4317 return -EPERM; 4318 } 4319 4320 /* 4321 * Treat SCHED_IDLE as nice 20. Only allow a switch to 4322 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it. 4323 */ 4324 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) { 4325 if (!can_nice(p, TASK_NICE(p))) 4326 return -EPERM; 4327 } 4328 4329 /* can't change other user's priorities */ 4330 if (!check_same_owner(p)) 4331 return -EPERM; 4332 4333 /* Normal users shall not reset the sched_reset_on_fork flag */ 4334 if (p->sched_reset_on_fork && !reset_on_fork) 4335 return -EPERM; 4336 } 4337 4338 if (user) { 4339 retval = security_task_setscheduler(p); 4340 if (retval) 4341 return retval; 4342 } 4343 4344 /* 4345 * make sure no PI-waiters arrive (or leave) while we are 4346 * changing the priority of the task: 4347 * 4348 * To be able to change p->policy safely, the appropriate 4349 * runqueue lock must be held. 4350 */ 4351 rq = task_rq_lock(p, &flags); 4352 4353 /* 4354 * Changing the policy of the stop threads its a very bad idea 4355 */ 4356 if (p == rq->stop) { 4357 task_rq_unlock(rq, p, &flags); 4358 return -EINVAL; 4359 } 4360 4361 /* 4362 * If not changing anything there's no need to proceed further: 4363 */ 4364 if (unlikely(policy == p->policy && (!rt_policy(policy) || 4365 param->sched_priority == p->rt_priority))) { 4366 task_rq_unlock(rq, p, &flags); 4367 return 0; 4368 } 4369 4370#ifdef CONFIG_RT_GROUP_SCHED 4371 if (user) { 4372 /* 4373 * Do not allow realtime tasks into groups that have no runtime 4374 * assigned. 4375 */ 4376 if (rt_bandwidth_enabled() && rt_policy(policy) && 4377 task_group(p)->rt_bandwidth.rt_runtime == 0 && 4378 !task_group_is_autogroup(task_group(p))) { 4379 task_rq_unlock(rq, p, &flags); 4380 return -EPERM; 4381 } 4382 } 4383#endif 4384 4385 /* recheck policy now with rq lock held */ 4386 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) { 4387 policy = oldpolicy = -1; 4388 task_rq_unlock(rq, p, &flags); 4389 goto recheck; 4390 } 4391 on_rq = p->on_rq; 4392 running = task_current(rq, p); 4393 if (on_rq) 4394 dequeue_task(rq, p, 0); 4395 if (running) 4396 p->sched_class->put_prev_task(rq, p); 4397 4398 p->sched_reset_on_fork = reset_on_fork; 4399 4400 oldprio = p->prio; 4401 prev_class = p->sched_class; 4402 __setscheduler(rq, p, policy, param->sched_priority); 4403 4404 if (running) 4405 p->sched_class->set_curr_task(rq); 4406 if (on_rq) 4407 enqueue_task(rq, p, 0); 4408 4409 check_class_changed(rq, p, prev_class, oldprio); 4410 task_rq_unlock(rq, p, &flags); 4411 4412 rt_mutex_adjust_pi(p); 4413 4414 return 0; 4415} 4416 4417/** 4418 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread. 4419 * @p: the task in question. 4420 * @policy: new policy. 4421 * @param: structure containing the new RT priority. 4422 * 4423 * NOTE that the task may be already dead. 4424 */ 4425int sched_setscheduler(struct task_struct *p, int policy, 4426 const struct sched_param *param) 4427{ 4428 return __sched_setscheduler(p, policy, param, true); 4429} 4430EXPORT_SYMBOL_GPL(sched_setscheduler); 4431 4432/** 4433 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace. 4434 * @p: the task in question. 4435 * @policy: new policy. 4436 * @param: structure containing the new RT priority. 4437 * 4438 * Just like sched_setscheduler, only don't bother checking if the 4439 * current context has permission. For example, this is needed in 4440 * stop_machine(): we create temporary high priority worker threads, 4441 * but our caller might not have that capability. 4442 */ 4443int sched_setscheduler_nocheck(struct task_struct *p, int policy, 4444 const struct sched_param *param) 4445{ 4446 return __sched_setscheduler(p, policy, param, false); 4447} 4448 4449static int 4450do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param) 4451{ 4452 struct sched_param lparam; 4453 struct task_struct *p; 4454 int retval; 4455 4456 if (!param || pid < 0) 4457 return -EINVAL; 4458 if (copy_from_user(&lparam, param, sizeof(struct sched_param))) 4459 return -EFAULT; 4460 4461 rcu_read_lock(); 4462 retval = -ESRCH; 4463 p = find_process_by_pid(pid); 4464 if (p != NULL) 4465 retval = sched_setscheduler(p, policy, &lparam); 4466 rcu_read_unlock(); 4467 4468 return retval; 4469} 4470 4471/** 4472 * sys_sched_setscheduler - set/change the scheduler policy and RT priority 4473 * @pid: the pid in question. 4474 * @policy: new policy. 4475 * @param: structure containing the new RT priority. 4476 */ 4477SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy, 4478 struct sched_param __user *, param) 4479{ 4480 /* negative values for policy are not valid */ 4481 if (policy < 0) 4482 return -EINVAL; 4483 4484 return do_sched_setscheduler(pid, policy, param); 4485} 4486 4487/** 4488 * sys_sched_setparam - set/change the RT priority of a thread 4489 * @pid: the pid in question. 4490 * @param: structure containing the new RT priority. 4491 */ 4492SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param) 4493{ 4494 return do_sched_setscheduler(pid, -1, param); 4495} 4496 4497/** 4498 * sys_sched_getscheduler - get the policy (scheduling class) of a thread 4499 * @pid: the pid in question. 4500 */ 4501SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid) 4502{ 4503 struct task_struct *p; 4504 int retval; 4505 4506 if (pid < 0) 4507 return -EINVAL; 4508 4509 retval = -ESRCH; 4510 rcu_read_lock(); 4511 p = find_process_by_pid(pid); 4512 if (p) { 4513 retval = security_task_getscheduler(p); 4514 if (!retval) 4515 retval = p->policy 4516 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0); 4517 } 4518 rcu_read_unlock(); 4519 return retval; 4520} 4521 4522/** 4523 * sys_sched_getparam - get the RT priority of a thread 4524 * @pid: the pid in question. 4525 * @param: structure containing the RT priority. 4526 */ 4527SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param) 4528{ 4529 struct sched_param lp; 4530 struct task_struct *p; 4531 int retval; 4532 4533 if (!param || pid < 0) 4534 return -EINVAL; 4535 4536 rcu_read_lock(); 4537 p = find_process_by_pid(pid); 4538 retval = -ESRCH; 4539 if (!p) 4540 goto out_unlock; 4541 4542 retval = security_task_getscheduler(p); 4543 if (retval) 4544 goto out_unlock; 4545 4546 lp.sched_priority = p->rt_priority; 4547 rcu_read_unlock(); 4548 4549 /* 4550 * This one might sleep, we cannot do it with a spinlock held ... 4551 */ 4552 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0; 4553 4554 return retval; 4555 4556out_unlock: 4557 rcu_read_unlock(); 4558 return retval; 4559} 4560 4561long sched_setaffinity(pid_t pid, const struct cpumask *in_mask) 4562{ 4563 cpumask_var_t cpus_allowed, new_mask; 4564 struct task_struct *p; 4565 int retval; 4566 4567 get_online_cpus(); 4568 rcu_read_lock(); 4569 4570 p = find_process_by_pid(pid); 4571 if (!p) { 4572 rcu_read_unlock(); 4573 put_online_cpus(); 4574 return -ESRCH; 4575 } 4576 4577 /* Prevent p going away */ 4578 get_task_struct(p); 4579 rcu_read_unlock(); 4580 4581 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) { 4582 retval = -ENOMEM; 4583 goto out_put_task; 4584 } 4585 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) { 4586 retval = -ENOMEM; 4587 goto out_free_cpus_allowed; 4588 } 4589 retval = -EPERM; 4590 if (!check_same_owner(p) && !ns_capable(task_user_ns(p), CAP_SYS_NICE)) 4591 goto out_unlock; 4592 4593 retval = security_task_setscheduler(p); 4594 if (retval) 4595 goto out_unlock; 4596 4597 cpuset_cpus_allowed(p, cpus_allowed); 4598 cpumask_and(new_mask, in_mask, cpus_allowed); 4599again: 4600 retval = set_cpus_allowed_ptr(p, new_mask); 4601 4602 if (!retval) { 4603 cpuset_cpus_allowed(p, cpus_allowed); 4604 if (!cpumask_subset(new_mask, cpus_allowed)) { 4605 /* 4606 * We must have raced with a concurrent cpuset 4607 * update. Just reset the cpus_allowed to the 4608 * cpuset's cpus_allowed 4609 */ 4610 cpumask_copy(new_mask, cpus_allowed); 4611 goto again; 4612 } 4613 } 4614out_unlock: 4615 free_cpumask_var(new_mask); 4616out_free_cpus_allowed: 4617 free_cpumask_var(cpus_allowed); 4618out_put_task: 4619 put_task_struct(p); 4620 put_online_cpus(); 4621 return retval; 4622} 4623 4624static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len, 4625 struct cpumask *new_mask) 4626{ 4627 if (len < cpumask_size()) 4628 cpumask_clear(new_mask); 4629 else if (len > cpumask_size()) 4630 len = cpumask_size(); 4631 4632 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0; 4633} 4634 4635/** 4636 * sys_sched_setaffinity - set the cpu affinity of a process 4637 * @pid: pid of the process 4638 * @len: length in bytes of the bitmask pointed to by user_mask_ptr 4639 * @user_mask_ptr: user-space pointer to the new cpu mask 4640 */ 4641SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len, 4642 unsigned long __user *, user_mask_ptr) 4643{ 4644 cpumask_var_t new_mask; 4645 int retval; 4646 4647 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) 4648 return -ENOMEM; 4649 4650 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask); 4651 if (retval == 0) 4652 retval = sched_setaffinity(pid, new_mask); 4653 free_cpumask_var(new_mask); 4654 return retval; 4655} 4656 4657long sched_getaffinity(pid_t pid, struct cpumask *mask) 4658{ 4659 struct task_struct *p; 4660 unsigned long flags; 4661 int retval; 4662 4663 get_online_cpus(); 4664 rcu_read_lock(); 4665 4666 retval = -ESRCH; 4667 p = find_process_by_pid(pid); 4668 if (!p) 4669 goto out_unlock; 4670 4671 retval = security_task_getscheduler(p); 4672 if (retval) 4673 goto out_unlock; 4674 4675 raw_spin_lock_irqsave(&p->pi_lock, flags); 4676 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask); 4677 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 4678 4679out_unlock: 4680 rcu_read_unlock(); 4681 put_online_cpus(); 4682 4683 return retval; 4684} 4685 4686/** 4687 * sys_sched_getaffinity - get the cpu affinity of a process 4688 * @pid: pid of the process 4689 * @len: length in bytes of the bitmask pointed to by user_mask_ptr 4690 * @user_mask_ptr: user-space pointer to hold the current cpu mask 4691 */ 4692SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len, 4693 unsigned long __user *, user_mask_ptr) 4694{ 4695 int ret; 4696 cpumask_var_t mask; 4697 4698 if ((len * BITS_PER_BYTE) < nr_cpu_ids) 4699 return -EINVAL; 4700 if (len & (sizeof(unsigned long)-1)) 4701 return -EINVAL; 4702 4703 if (!alloc_cpumask_var(&mask, GFP_KERNEL)) 4704 return -ENOMEM; 4705 4706 ret = sched_getaffinity(pid, mask); 4707 if (ret == 0) { 4708 size_t retlen = min_t(size_t, len, cpumask_size()); 4709 4710 if (copy_to_user(user_mask_ptr, mask, retlen)) 4711 ret = -EFAULT; 4712 else 4713 ret = retlen; 4714 } 4715 free_cpumask_var(mask); 4716 4717 return ret; 4718} 4719 4720/** 4721 * sys_sched_yield - yield the current processor to other threads. 4722 * 4723 * This function yields the current CPU to other tasks. If there are no 4724 * other threads running on this CPU then this function will return. 4725 */ 4726SYSCALL_DEFINE0(sched_yield) 4727{ 4728 struct rq *rq = this_rq_lock(); 4729 4730 schedstat_inc(rq, yld_count); 4731 current->sched_class->yield_task(rq); 4732 4733 /* 4734 * Since we are going to call schedule() anyway, there's 4735 * no need to preempt or enable interrupts: 4736 */ 4737 __release(rq->lock); 4738 spin_release(&rq->lock.dep_map, 1, _THIS_IP_); 4739 do_raw_spin_unlock(&rq->lock); 4740 sched_preempt_enable_no_resched(); 4741 4742 schedule(); 4743 4744 return 0; 4745} 4746 4747static inline int should_resched(void) 4748{ 4749 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE); 4750} 4751 4752static void __cond_resched(void) 4753{ 4754 add_preempt_count(PREEMPT_ACTIVE); 4755 __schedule(); 4756 sub_preempt_count(PREEMPT_ACTIVE); 4757} 4758 4759int __sched _cond_resched(void) 4760{ 4761 if (should_resched()) { 4762 __cond_resched(); 4763 return 1; 4764 } 4765 return 0; 4766} 4767EXPORT_SYMBOL(_cond_resched); 4768 4769/* 4770 * __cond_resched_lock() - if a reschedule is pending, drop the given lock, 4771 * call schedule, and on return reacquire the lock. 4772 * 4773 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level 4774 * operations here to prevent schedule() from being called twice (once via 4775 * spin_unlock(), once by hand). 4776 */ 4777int __cond_resched_lock(spinlock_t *lock) 4778{ 4779 int resched = should_resched(); 4780 int ret = 0; 4781 4782 lockdep_assert_held(lock); 4783 4784 if (spin_needbreak(lock) || resched) { 4785 spin_unlock(lock); 4786 if (resched) 4787 __cond_resched(); 4788 else 4789 cpu_relax(); 4790 ret = 1; 4791 spin_lock(lock); 4792 } 4793 return ret; 4794} 4795EXPORT_SYMBOL(__cond_resched_lock); 4796 4797int __sched __cond_resched_softirq(void) 4798{ 4799 BUG_ON(!in_softirq()); 4800 4801 if (should_resched()) { 4802 local_bh_enable(); 4803 __cond_resched(); 4804 local_bh_disable(); 4805 return 1; 4806 } 4807 return 0; 4808} 4809EXPORT_SYMBOL(__cond_resched_softirq); 4810 4811/** 4812 * yield - yield the current processor to other threads. 4813 * 4814 * Do not ever use this function, there's a 99% chance you're doing it wrong. 4815 * 4816 * The scheduler is at all times free to pick the calling task as the most 4817 * eligible task to run, if removing the yield() call from your code breaks 4818 * it, its already broken. 4819 * 4820 * Typical broken usage is: 4821 * 4822 * while (!event) 4823 * yield(); 4824 * 4825 * where one assumes that yield() will let 'the other' process run that will 4826 * make event true. If the current task is a SCHED_FIFO task that will never 4827 * happen. Never use yield() as a progress guarantee!! 4828 * 4829 * If you want to use yield() to wait for something, use wait_event(). 4830 * If you want to use yield() to be 'nice' for others, use cond_resched(). 4831 * If you still want to use yield(), do not! 4832 */ 4833void __sched yield(void) 4834{ 4835 set_current_state(TASK_RUNNING); 4836 sys_sched_yield(); 4837} 4838EXPORT_SYMBOL(yield); 4839 4840/** 4841 * yield_to - yield the current processor to another thread in 4842 * your thread group, or accelerate that thread toward the 4843 * processor it's on. 4844 * @p: target task 4845 * @preempt: whether task preemption is allowed or not 4846 * 4847 * It's the caller's job to ensure that the target task struct 4848 * can't go away on us before we can do any checks. 4849 * 4850 * Returns true if we indeed boosted the target task. 4851 */ 4852bool __sched yield_to(struct task_struct *p, bool preempt) 4853{ 4854 struct task_struct *curr = current; 4855 struct rq *rq, *p_rq; 4856 unsigned long flags; 4857 bool yielded = 0; 4858 4859 local_irq_save(flags); 4860 rq = this_rq(); 4861 4862again: 4863 p_rq = task_rq(p); 4864 double_rq_lock(rq, p_rq); 4865 while (task_rq(p) != p_rq) { 4866 double_rq_unlock(rq, p_rq); 4867 goto again; 4868 } 4869 4870 if (!curr->sched_class->yield_to_task) 4871 goto out; 4872 4873 if (curr->sched_class != p->sched_class) 4874 goto out; 4875 4876 if (task_running(p_rq, p) || p->state) 4877 goto out; 4878 4879 yielded = curr->sched_class->yield_to_task(rq, p, preempt); 4880 if (yielded) { 4881 schedstat_inc(rq, yld_count); 4882 /* 4883 * Make p's CPU reschedule; pick_next_entity takes care of 4884 * fairness. 4885 */ 4886 if (preempt && rq != p_rq) 4887 resched_task(p_rq->curr); 4888 } else { 4889 /* 4890 * We might have set it in task_yield_fair(), but are 4891 * not going to schedule(), so don't want to skip 4892 * the next update. 4893 */ 4894 rq->skip_clock_update = 0; 4895 } 4896 4897out: 4898 double_rq_unlock(rq, p_rq); 4899 local_irq_restore(flags); 4900 4901 if (yielded) 4902 schedule(); 4903 4904 return yielded; 4905} 4906EXPORT_SYMBOL_GPL(yield_to); 4907 4908/* 4909 * This task is about to go to sleep on IO. Increment rq->nr_iowait so 4910 * that process accounting knows that this is a task in IO wait state. 4911 */ 4912void __sched io_schedule(void) 4913{ 4914 struct rq *rq = raw_rq(); 4915 4916 delayacct_blkio_start(); 4917 atomic_inc(&rq->nr_iowait); 4918 blk_flush_plug(current); 4919 current->in_iowait = 1; 4920 schedule(); 4921 current->in_iowait = 0; 4922 atomic_dec(&rq->nr_iowait); 4923 delayacct_blkio_end(); 4924} 4925EXPORT_SYMBOL(io_schedule); 4926 4927long __sched io_schedule_timeout(long timeout) 4928{ 4929 struct rq *rq = raw_rq(); 4930 long ret; 4931 4932 delayacct_blkio_start(); 4933 atomic_inc(&rq->nr_iowait); 4934 blk_flush_plug(current); 4935 current->in_iowait = 1; 4936 ret = schedule_timeout(timeout); 4937 current->in_iowait = 0; 4938 atomic_dec(&rq->nr_iowait); 4939 delayacct_blkio_end(); 4940 return ret; 4941} 4942 4943/** 4944 * sys_sched_get_priority_max - return maximum RT priority. 4945 * @policy: scheduling class. 4946 * 4947 * this syscall returns the maximum rt_priority that can be used 4948 * by a given scheduling class. 4949 */ 4950SYSCALL_DEFINE1(sched_get_priority_max, int, policy) 4951{ 4952 int ret = -EINVAL; 4953 4954 switch (policy) { 4955 case SCHED_FIFO: 4956 case SCHED_RR: 4957 ret = MAX_USER_RT_PRIO-1; 4958 break; 4959 case SCHED_NORMAL: 4960 case SCHED_BATCH: 4961 case SCHED_IDLE: 4962 ret = 0; 4963 break; 4964 } 4965 return ret; 4966} 4967 4968/** 4969 * sys_sched_get_priority_min - return minimum RT priority. 4970 * @policy: scheduling class. 4971 * 4972 * this syscall returns the minimum rt_priority that can be used 4973 * by a given scheduling class. 4974 */ 4975SYSCALL_DEFINE1(sched_get_priority_min, int, policy) 4976{ 4977 int ret = -EINVAL; 4978 4979 switch (policy) { 4980 case SCHED_FIFO: 4981 case SCHED_RR: 4982 ret = 1; 4983 break; 4984 case SCHED_NORMAL: 4985 case SCHED_BATCH: 4986 case SCHED_IDLE: 4987 ret = 0; 4988 } 4989 return ret; 4990} 4991 4992/** 4993 * sys_sched_rr_get_interval - return the default timeslice of a process. 4994 * @pid: pid of the process. 4995 * @interval: userspace pointer to the timeslice value. 4996 * 4997 * this syscall writes the default timeslice value of a given process 4998 * into the user-space timespec buffer. A value of '0' means infinity. 4999 */ 5000SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid, 5001 struct timespec __user *, interval) 5002{ 5003 struct task_struct *p; 5004 unsigned int time_slice; 5005 unsigned long flags; 5006 struct rq *rq; 5007 int retval; 5008 struct timespec t; 5009 5010 if (pid < 0) 5011 return -EINVAL; 5012 5013 retval = -ESRCH; 5014 rcu_read_lock(); 5015 p = find_process_by_pid(pid); 5016 if (!p) 5017 goto out_unlock; 5018 5019 retval = security_task_getscheduler(p); 5020 if (retval) 5021 goto out_unlock; 5022 5023 rq = task_rq_lock(p, &flags); 5024 time_slice = p->sched_class->get_rr_interval(rq, p); 5025 task_rq_unlock(rq, p, &flags); 5026 5027 rcu_read_unlock(); 5028 jiffies_to_timespec(time_slice, &t); 5029 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0; 5030 return retval; 5031 5032out_unlock: 5033 rcu_read_unlock(); 5034 return retval; 5035} 5036 5037static const char stat_nam[] = TASK_STATE_TO_CHAR_STR; 5038 5039void sched_show_task(struct task_struct *p) 5040{ 5041 unsigned long free = 0; 5042 unsigned state; 5043 5044 state = p->state ? __ffs(p->state) + 1 : 0; 5045 printk(KERN_INFO "%-15.15s %c", p->comm, 5046 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?'); 5047#if BITS_PER_LONG == 32 5048 if (state == TASK_RUNNING) 5049 printk(KERN_CONT " running "); 5050 else 5051 printk(KERN_CONT " %08lx ", thread_saved_pc(p)); 5052#else 5053 if (state == TASK_RUNNING) 5054 printk(KERN_CONT " running task "); 5055 else 5056 printk(KERN_CONT " %016lx ", thread_saved_pc(p)); 5057#endif 5058#ifdef CONFIG_DEBUG_STACK_USAGE 5059 free = stack_not_used(p); 5060#endif 5061 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free, 5062 task_pid_nr(p), task_pid_nr(rcu_dereference(p->real_parent)), 5063 (unsigned long)task_thread_info(p)->flags); 5064 5065 show_stack(p, NULL); 5066} 5067 5068void show_state_filter(unsigned long state_filter) 5069{ 5070 struct task_struct *g, *p; 5071 5072#if BITS_PER_LONG == 32 5073 printk(KERN_INFO 5074 " task PC stack pid father\n"); 5075#else 5076 printk(KERN_INFO 5077 " task PC stack pid father\n"); 5078#endif 5079 rcu_read_lock(); 5080 do_each_thread(g, p) { 5081 /* 5082 * reset the NMI-timeout, listing all files on a slow 5083 * console might take a lot of time: 5084 */ 5085 touch_nmi_watchdog(); 5086 if (!state_filter || (p->state & state_filter)) 5087 sched_show_task(p); 5088 } while_each_thread(g, p); 5089 5090 touch_all_softlockup_watchdogs(); 5091 5092#ifdef CONFIG_SCHED_DEBUG 5093 sysrq_sched_debug_show(); 5094#endif 5095 rcu_read_unlock(); 5096 /* 5097 * Only show locks if all tasks are dumped: 5098 */ 5099 if (!state_filter) 5100 debug_show_all_locks(); 5101} 5102 5103void __cpuinit init_idle_bootup_task(struct task_struct *idle) 5104{ 5105 idle->sched_class = &idle_sched_class; 5106} 5107 5108/** 5109 * init_idle - set up an idle thread for a given CPU 5110 * @idle: task in question 5111 * @cpu: cpu the idle task belongs to 5112 * 5113 * NOTE: this function does not set the idle thread's NEED_RESCHED 5114 * flag, to make booting more robust. 5115 */ 5116void __cpuinit init_idle(struct task_struct *idle, int cpu) 5117{ 5118 struct rq *rq = cpu_rq(cpu); 5119 unsigned long flags; 5120 5121 raw_spin_lock_irqsave(&rq->lock, flags); 5122 5123 __sched_fork(idle); 5124 idle->state = TASK_RUNNING; 5125 idle->se.exec_start = sched_clock(); 5126 5127 do_set_cpus_allowed(idle, cpumask_of(cpu)); 5128 /* 5129 * We're having a chicken and egg problem, even though we are 5130 * holding rq->lock, the cpu isn't yet set to this cpu so the 5131 * lockdep check in task_group() will fail. 5132 * 5133 * Similar case to sched_fork(). / Alternatively we could 5134 * use task_rq_lock() here and obtain the other rq->lock. 5135 * 5136 * Silence PROVE_RCU 5137 */ 5138 rcu_read_lock(); 5139 __set_task_cpu(idle, cpu); 5140 rcu_read_unlock(); 5141 5142 rq->curr = rq->idle = idle; 5143#if defined(CONFIG_SMP) 5144 idle->on_cpu = 1; 5145#endif 5146 raw_spin_unlock_irqrestore(&rq->lock, flags); 5147 5148 /* Set the preempt count _outside_ the spinlocks! */ 5149 task_thread_info(idle)->preempt_count = 0; 5150 5151 /* 5152 * The idle tasks have their own, simple scheduling class: 5153 */ 5154 idle->sched_class = &idle_sched_class; 5155 ftrace_graph_init_idle_task(idle, cpu); 5156#if defined(CONFIG_SMP) 5157 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu); 5158#endif 5159} 5160 5161#ifdef CONFIG_SMP 5162void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask) 5163{ 5164 if (p->sched_class && p->sched_class->set_cpus_allowed) 5165 p->sched_class->set_cpus_allowed(p, new_mask); 5166 5167 cpumask_copy(&p->cpus_allowed, new_mask); 5168 p->nr_cpus_allowed = cpumask_weight(new_mask); 5169} 5170 5171/* 5172 * This is how migration works: 5173 * 5174 * 1) we invoke migration_cpu_stop() on the target CPU using 5175 * stop_one_cpu(). 5176 * 2) stopper starts to run (implicitly forcing the migrated thread 5177 * off the CPU) 5178 * 3) it checks whether the migrated task is still in the wrong runqueue. 5179 * 4) if it's in the wrong runqueue then the migration thread removes 5180 * it and puts it into the right queue. 5181 * 5) stopper completes and stop_one_cpu() returns and the migration 5182 * is done. 5183 */ 5184 5185/* 5186 * Change a given task's CPU affinity. Migrate the thread to a 5187 * proper CPU and schedule it away if the CPU it's executing on 5188 * is removed from the allowed bitmask. 5189 * 5190 * NOTE: the caller must have a valid reference to the task, the 5191 * task must not exit() & deallocate itself prematurely. The 5192 * call is not atomic; no spinlocks may be held. 5193 */ 5194int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask) 5195{ 5196 unsigned long flags; 5197 struct rq *rq; 5198 unsigned int dest_cpu; 5199 int ret = 0; 5200 5201 rq = task_rq_lock(p, &flags); 5202 5203 if (cpumask_equal(&p->cpus_allowed, new_mask)) 5204 goto out; 5205 5206 if (!cpumask_intersects(new_mask, cpu_active_mask)) { 5207 ret = -EINVAL; 5208 goto out; 5209 } 5210 5211 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current)) { 5212 ret = -EINVAL; 5213 goto out; 5214 } 5215 5216 do_set_cpus_allowed(p, new_mask); 5217 5218 /* Can the task run on the task's current CPU? If so, we're done */ 5219 if (cpumask_test_cpu(task_cpu(p), new_mask)) 5220 goto out; 5221 5222 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask); 5223 if (p->on_rq) { 5224 struct migration_arg arg = { p, dest_cpu }; 5225 /* Need help from migration thread: drop lock and wait. */ 5226 task_rq_unlock(rq, p, &flags); 5227 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg); 5228 tlb_migrate_finish(p->mm); 5229 return 0; 5230 } 5231out: 5232 task_rq_unlock(rq, p, &flags); 5233 5234 return ret; 5235} 5236EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr); 5237 5238/* 5239 * Move (not current) task off this cpu, onto dest cpu. We're doing 5240 * this because either it can't run here any more (set_cpus_allowed() 5241 * away from this CPU, or CPU going down), or because we're 5242 * attempting to rebalance this task on exec (sched_exec). 5243 * 5244 * So we race with normal scheduler movements, but that's OK, as long 5245 * as the task is no longer on this CPU. 5246 * 5247 * Returns non-zero if task was successfully migrated. 5248 */ 5249static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu) 5250{ 5251 struct rq *rq_dest, *rq_src; 5252 int ret = 0; 5253 5254 if (unlikely(!cpu_active(dest_cpu))) 5255 return ret; 5256 5257 rq_src = cpu_rq(src_cpu); 5258 rq_dest = cpu_rq(dest_cpu); 5259 5260 raw_spin_lock(&p->pi_lock); 5261 double_rq_lock(rq_src, rq_dest); 5262 /* Already moved. */ 5263 if (task_cpu(p) != src_cpu) 5264 goto done; 5265 /* Affinity changed (again). */ 5266 if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p))) 5267 goto fail; 5268 5269 /* 5270 * If we're not on a rq, the next wake-up will ensure we're 5271 * placed properly. 5272 */ 5273 if (p->on_rq) { 5274 dequeue_task(rq_src, p, 0); 5275 set_task_cpu(p, dest_cpu); 5276 enqueue_task(rq_dest, p, 0); 5277 check_preempt_curr(rq_dest, p, 0); 5278 } 5279done: 5280 ret = 1; 5281fail: 5282 double_rq_unlock(rq_src, rq_dest); 5283 raw_spin_unlock(&p->pi_lock); 5284 return ret; 5285} 5286 5287/* 5288 * migration_cpu_stop - this will be executed by a highprio stopper thread 5289 * and performs thread migration by bumping thread off CPU then 5290 * 'pushing' onto another runqueue. 5291 */ 5292static int migration_cpu_stop(void *data) 5293{ 5294 struct migration_arg *arg = data; 5295 5296 /* 5297 * The original target cpu might have gone down and we might 5298 * be on another cpu but it doesn't matter. 5299 */ 5300 local_irq_disable(); 5301 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu); 5302 local_irq_enable(); 5303 return 0; 5304} 5305 5306#ifdef CONFIG_HOTPLUG_CPU 5307 5308/* 5309 * Ensures that the idle task is using init_mm right before its cpu goes 5310 * offline. 5311 */ 5312void idle_task_exit(void) 5313{ 5314 struct mm_struct *mm = current->active_mm; 5315 5316 BUG_ON(cpu_online(smp_processor_id())); 5317 5318 if (mm != &init_mm) 5319 switch_mm(mm, &init_mm, current); 5320 mmdrop(mm); 5321} 5322 5323/* 5324 * Since this CPU is going 'away' for a while, fold any nr_active delta 5325 * we might have. Assumes we're called after migrate_tasks() so that the 5326 * nr_active count is stable. 5327 * 5328 * Also see the comment "Global load-average calculations". 5329 */ 5330static void calc_load_migrate(struct rq *rq) 5331{ 5332 long delta = calc_load_fold_active(rq); 5333 if (delta) 5334 atomic_long_add(delta, &calc_load_tasks); 5335} 5336 5337/* 5338 * Migrate all tasks from the rq, sleeping tasks will be migrated by 5339 * try_to_wake_up()->select_task_rq(). 5340 * 5341 * Called with rq->lock held even though we'er in stop_machine() and 5342 * there's no concurrency possible, we hold the required locks anyway 5343 * because of lock validation efforts. 5344 */ 5345static void migrate_tasks(unsigned int dead_cpu) 5346{ 5347 struct rq *rq = cpu_rq(dead_cpu); 5348 struct task_struct *next, *stop = rq->stop; 5349 int dest_cpu; 5350 5351 /* 5352 * Fudge the rq selection such that the below task selection loop 5353 * doesn't get stuck on the currently eligible stop task. 5354 * 5355 * We're currently inside stop_machine() and the rq is either stuck 5356 * in the stop_machine_cpu_stop() loop, or we're executing this code, 5357 * either way we should never end up calling schedule() until we're 5358 * done here. 5359 */ 5360 rq->stop = NULL; 5361 5362 for ( ; ; ) { 5363 /* 5364 * There's this thread running, bail when that's the only 5365 * remaining thread. 5366 */ 5367 if (rq->nr_running == 1) 5368 break; 5369 5370 next = pick_next_task(rq); 5371 BUG_ON(!next); 5372 next->sched_class->put_prev_task(rq, next); 5373 5374 /* Find suitable destination for @next, with force if needed. */ 5375 dest_cpu = select_fallback_rq(dead_cpu, next); 5376 raw_spin_unlock(&rq->lock); 5377 5378 __migrate_task(next, dead_cpu, dest_cpu); 5379 5380 raw_spin_lock(&rq->lock); 5381 } 5382 5383 rq->stop = stop; 5384} 5385 5386#endif /* CONFIG_HOTPLUG_CPU */ 5387 5388#if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL) 5389 5390static struct ctl_table sd_ctl_dir[] = { 5391 { 5392 .procname = "sched_domain", 5393 .mode = 0555, 5394 }, 5395 {} 5396}; 5397 5398static struct ctl_table sd_ctl_root[] = { 5399 { 5400 .procname = "kernel", 5401 .mode = 0555, 5402 .child = sd_ctl_dir, 5403 }, 5404 {} 5405}; 5406 5407static struct ctl_table *sd_alloc_ctl_entry(int n) 5408{ 5409 struct ctl_table *entry = 5410 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL); 5411 5412 return entry; 5413} 5414 5415static void sd_free_ctl_entry(struct ctl_table **tablep) 5416{ 5417 struct ctl_table *entry; 5418 5419 /* 5420 * In the intermediate directories, both the child directory and 5421 * procname are dynamically allocated and could fail but the mode 5422 * will always be set. In the lowest directory the names are 5423 * static strings and all have proc handlers. 5424 */ 5425 for (entry = *tablep; entry->mode; entry++) { 5426 if (entry->child) 5427 sd_free_ctl_entry(&entry->child); 5428 if (entry->proc_handler == NULL) 5429 kfree(entry->procname); 5430 } 5431 5432 kfree(*tablep); 5433 *tablep = NULL; 5434} 5435 5436static void 5437set_table_entry(struct ctl_table *entry, 5438 const char *procname, void *data, int maxlen, 5439 umode_t mode, proc_handler *proc_handler) 5440{ 5441 entry->procname = procname; 5442 entry->data = data; 5443 entry->maxlen = maxlen; 5444 entry->mode = mode; 5445 entry->proc_handler = proc_handler; 5446} 5447 5448static struct ctl_table * 5449sd_alloc_ctl_domain_table(struct sched_domain *sd) 5450{ 5451 struct ctl_table *table = sd_alloc_ctl_entry(13); 5452 5453 if (table == NULL) 5454 return NULL; 5455 5456 set_table_entry(&table[0], "min_interval", &sd->min_interval, 5457 sizeof(long), 0644, proc_doulongvec_minmax); 5458 set_table_entry(&table[1], "max_interval", &sd->max_interval, 5459 sizeof(long), 0644, proc_doulongvec_minmax); 5460 set_table_entry(&table[2], "busy_idx", &sd->busy_idx, 5461 sizeof(int), 0644, proc_dointvec_minmax); 5462 set_table_entry(&table[3], "idle_idx", &sd->idle_idx, 5463 sizeof(int), 0644, proc_dointvec_minmax); 5464 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx, 5465 sizeof(int), 0644, proc_dointvec_minmax); 5466 set_table_entry(&table[5], "wake_idx", &sd->wake_idx, 5467 sizeof(int), 0644, proc_dointvec_minmax); 5468 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx, 5469 sizeof(int), 0644, proc_dointvec_minmax); 5470 set_table_entry(&table[7], "busy_factor", &sd->busy_factor, 5471 sizeof(int), 0644, proc_dointvec_minmax); 5472 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct, 5473 sizeof(int), 0644, proc_dointvec_minmax); 5474 set_table_entry(&table[9], "cache_nice_tries", 5475 &sd->cache_nice_tries, 5476 sizeof(int), 0644, proc_dointvec_minmax); 5477 set_table_entry(&table[10], "flags", &sd->flags, 5478 sizeof(int), 0644, proc_dointvec_minmax); 5479 set_table_entry(&table[11], "name", sd->name, 5480 CORENAME_MAX_SIZE, 0444, proc_dostring); 5481 /* &table[12] is terminator */ 5482 5483 return table; 5484} 5485 5486static ctl_table *sd_alloc_ctl_cpu_table(int cpu) 5487{ 5488 struct ctl_table *entry, *table; 5489 struct sched_domain *sd; 5490 int domain_num = 0, i; 5491 char buf[32]; 5492 5493 for_each_domain(cpu, sd) 5494 domain_num++; 5495 entry = table = sd_alloc_ctl_entry(domain_num + 1); 5496 if (table == NULL) 5497 return NULL; 5498 5499 i = 0; 5500 for_each_domain(cpu, sd) { 5501 snprintf(buf, 32, "domain%d", i); 5502 entry->procname = kstrdup(buf, GFP_KERNEL); 5503 entry->mode = 0555; 5504 entry->child = sd_alloc_ctl_domain_table(sd); 5505 entry++; 5506 i++; 5507 } 5508 return table; 5509} 5510 5511static struct ctl_table_header *sd_sysctl_header; 5512static void register_sched_domain_sysctl(void) 5513{ 5514 int i, cpu_num = num_possible_cpus(); 5515 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1); 5516 char buf[32]; 5517 5518 WARN_ON(sd_ctl_dir[0].child); 5519 sd_ctl_dir[0].child = entry; 5520 5521 if (entry == NULL) 5522 return; 5523 5524 for_each_possible_cpu(i) { 5525 snprintf(buf, 32, "cpu%d", i); 5526 entry->procname = kstrdup(buf, GFP_KERNEL); 5527 entry->mode = 0555; 5528 entry->child = sd_alloc_ctl_cpu_table(i); 5529 entry++; 5530 } 5531 5532 WARN_ON(sd_sysctl_header); 5533 sd_sysctl_header = register_sysctl_table(sd_ctl_root); 5534} 5535 5536/* may be called multiple times per register */ 5537static void unregister_sched_domain_sysctl(void) 5538{ 5539 if (sd_sysctl_header) 5540 unregister_sysctl_table(sd_sysctl_header); 5541 sd_sysctl_header = NULL; 5542 if (sd_ctl_dir[0].child) 5543 sd_free_ctl_entry(&sd_ctl_dir[0].child); 5544} 5545#else 5546static void register_sched_domain_sysctl(void) 5547{ 5548} 5549static void unregister_sched_domain_sysctl(void) 5550{ 5551} 5552#endif 5553 5554static void set_rq_online(struct rq *rq) 5555{ 5556 if (!rq->online) { 5557 const struct sched_class *class; 5558 5559 cpumask_set_cpu(rq->cpu, rq->rd->online); 5560 rq->online = 1; 5561 5562 for_each_class(class) { 5563 if (class->rq_online) 5564 class->rq_online(rq); 5565 } 5566 } 5567} 5568 5569static void set_rq_offline(struct rq *rq) 5570{ 5571 if (rq->online) { 5572 const struct sched_class *class; 5573 5574 for_each_class(class) { 5575 if (class->rq_offline) 5576 class->rq_offline(rq); 5577 } 5578 5579 cpumask_clear_cpu(rq->cpu, rq->rd->online); 5580 rq->online = 0; 5581 } 5582} 5583 5584/* 5585 * migration_call - callback that gets triggered when a CPU is added. 5586 * Here we can start up the necessary migration thread for the new CPU. 5587 */ 5588static int __cpuinit 5589migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu) 5590{ 5591 int cpu = (long)hcpu; 5592 unsigned long flags; 5593 struct rq *rq = cpu_rq(cpu); 5594 5595 switch (action & ~CPU_TASKS_FROZEN) { 5596 5597 case CPU_UP_PREPARE: 5598 rq->calc_load_update = calc_load_update; 5599 break; 5600 5601 case CPU_ONLINE: 5602 /* Update our root-domain */ 5603 raw_spin_lock_irqsave(&rq->lock, flags); 5604 if (rq->rd) { 5605 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span)); 5606 5607 set_rq_online(rq); 5608 } 5609 raw_spin_unlock_irqrestore(&rq->lock, flags); 5610 break; 5611 5612#ifdef CONFIG_HOTPLUG_CPU 5613 case CPU_DYING: 5614 sched_ttwu_pending(); 5615 /* Update our root-domain */ 5616 raw_spin_lock_irqsave(&rq->lock, flags); 5617 if (rq->rd) { 5618 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span)); 5619 set_rq_offline(rq); 5620 } 5621 migrate_tasks(cpu); 5622 BUG_ON(rq->nr_running != 1); /* the migration thread */ 5623 raw_spin_unlock_irqrestore(&rq->lock, flags); 5624 break; 5625 5626 case CPU_DEAD: 5627 calc_load_migrate(rq); 5628 break; 5629#endif 5630 } 5631 5632 update_max_interval(); 5633 5634 return NOTIFY_OK; 5635} 5636 5637/* 5638 * Register at high priority so that task migration (migrate_all_tasks) 5639 * happens before everything else. This has to be lower priority than 5640 * the notifier in the perf_event subsystem, though. 5641 */ 5642static struct notifier_block __cpuinitdata migration_notifier = { 5643 .notifier_call = migration_call, 5644 .priority = CPU_PRI_MIGRATION, 5645}; 5646 5647static int __cpuinit sched_cpu_active(struct notifier_block *nfb, 5648 unsigned long action, void *hcpu) 5649{ 5650 switch (action & ~CPU_TASKS_FROZEN) { 5651 case CPU_STARTING: 5652 case CPU_DOWN_FAILED: 5653 set_cpu_active((long)hcpu, true); 5654 return NOTIFY_OK; 5655 default: 5656 return NOTIFY_DONE; 5657 } 5658} 5659 5660static int __cpuinit sched_cpu_inactive(struct notifier_block *nfb, 5661 unsigned long action, void *hcpu) 5662{ 5663 switch (action & ~CPU_TASKS_FROZEN) { 5664 case CPU_DOWN_PREPARE: 5665 set_cpu_active((long)hcpu, false); 5666 return NOTIFY_OK; 5667 default: 5668 return NOTIFY_DONE; 5669 } 5670} 5671 5672static int __init migration_init(void) 5673{ 5674 void *cpu = (void *)(long)smp_processor_id(); 5675 int err; 5676 5677 /* Initialize migration for the boot CPU */ 5678 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu); 5679 BUG_ON(err == NOTIFY_BAD); 5680 migration_call(&migration_notifier, CPU_ONLINE, cpu); 5681 register_cpu_notifier(&migration_notifier); 5682 5683 /* Register cpu active notifiers */ 5684 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE); 5685 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE); 5686 5687 return 0; 5688} 5689early_initcall(migration_init); 5690#endif 5691 5692#ifdef CONFIG_SMP 5693 5694static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */ 5695 5696#ifdef CONFIG_SCHED_DEBUG 5697 5698static __read_mostly int sched_debug_enabled; 5699 5700static int __init sched_debug_setup(char *str) 5701{ 5702 sched_debug_enabled = 1; 5703 5704 return 0; 5705} 5706early_param("sched_debug", sched_debug_setup); 5707 5708static inline bool sched_debug(void) 5709{ 5710 return sched_debug_enabled; 5711} 5712 5713static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level, 5714 struct cpumask *groupmask) 5715{ 5716 struct sched_group *group = sd->groups; 5717 char str[256]; 5718 5719 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd)); 5720 cpumask_clear(groupmask); 5721 5722 printk(KERN_DEBUG "%*s domain %d: ", level, "", level); 5723 5724 if (!(sd->flags & SD_LOAD_BALANCE)) { 5725 printk("does not load-balance\n"); 5726 if (sd->parent) 5727 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain" 5728 " has parent"); 5729 return -1; 5730 } 5731 5732 printk(KERN_CONT "span %s level %s\n", str, sd->name); 5733 5734 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) { 5735 printk(KERN_ERR "ERROR: domain->span does not contain " 5736 "CPU%d\n", cpu); 5737 } 5738 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) { 5739 printk(KERN_ERR "ERROR: domain->groups does not contain" 5740 " CPU%d\n", cpu); 5741 } 5742 5743 printk(KERN_DEBUG "%*s groups:", level + 1, ""); 5744 do { 5745 if (!group) { 5746 printk("\n"); 5747 printk(KERN_ERR "ERROR: group is NULL\n"); 5748 break; 5749 } 5750 5751 /* 5752 * Even though we initialize ->power to something semi-sane, 5753 * we leave power_orig unset. This allows us to detect if 5754 * domain iteration is still funny without causing /0 traps. 5755 */ 5756 if (!group->sgp->power_orig) { 5757 printk(KERN_CONT "\n"); 5758 printk(KERN_ERR "ERROR: domain->cpu_power not " 5759 "set\n"); 5760 break; 5761 } 5762 5763 if (!cpumask_weight(sched_group_cpus(group))) { 5764 printk(KERN_CONT "\n"); 5765 printk(KERN_ERR "ERROR: empty group\n"); 5766 break; 5767 } 5768 5769 if (!(sd->flags & SD_OVERLAP) && 5770 cpumask_intersects(groupmask, sched_group_cpus(group))) { 5771 printk(KERN_CONT "\n"); 5772 printk(KERN_ERR "ERROR: repeated CPUs\n"); 5773 break; 5774 } 5775 5776 cpumask_or(groupmask, groupmask, sched_group_cpus(group)); 5777 5778 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group)); 5779 5780 printk(KERN_CONT " %s", str); 5781 if (group->sgp->power != SCHED_POWER_SCALE) { 5782 printk(KERN_CONT " (cpu_power = %d)", 5783 group->sgp->power); 5784 } 5785 5786 group = group->next; 5787 } while (group != sd->groups); 5788 printk(KERN_CONT "\n"); 5789 5790 if (!cpumask_equal(sched_domain_span(sd), groupmask)) 5791 printk(KERN_ERR "ERROR: groups don't span domain->span\n"); 5792 5793 if (sd->parent && 5794 !cpumask_subset(groupmask, sched_domain_span(sd->parent))) 5795 printk(KERN_ERR "ERROR: parent span is not a superset " 5796 "of domain->span\n"); 5797 return 0; 5798} 5799 5800static void sched_domain_debug(struct sched_domain *sd, int cpu) 5801{ 5802 int level = 0; 5803 5804 if (!sched_debug_enabled) 5805 return; 5806 5807 if (!sd) { 5808 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu); 5809 return; 5810 } 5811 5812 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu); 5813 5814 for (;;) { 5815 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask)) 5816 break; 5817 level++; 5818 sd = sd->parent; 5819 if (!sd) 5820 break; 5821 } 5822} 5823#else /* !CONFIG_SCHED_DEBUG */ 5824# define sched_domain_debug(sd, cpu) do { } while (0) 5825static inline bool sched_debug(void) 5826{ 5827 return false; 5828} 5829#endif /* CONFIG_SCHED_DEBUG */ 5830 5831static int sd_degenerate(struct sched_domain *sd) 5832{ 5833 if (cpumask_weight(sched_domain_span(sd)) == 1) 5834 return 1; 5835 5836 /* Following flags need at least 2 groups */ 5837 if (sd->flags & (SD_LOAD_BALANCE | 5838 SD_BALANCE_NEWIDLE | 5839 SD_BALANCE_FORK | 5840 SD_BALANCE_EXEC | 5841 SD_SHARE_CPUPOWER | 5842 SD_SHARE_PKG_RESOURCES)) { 5843 if (sd->groups != sd->groups->next) 5844 return 0; 5845 } 5846 5847 /* Following flags don't use groups */ 5848 if (sd->flags & (SD_WAKE_AFFINE)) 5849 return 0; 5850 5851 return 1; 5852} 5853 5854static int 5855sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent) 5856{ 5857 unsigned long cflags = sd->flags, pflags = parent->flags; 5858 5859 if (sd_degenerate(parent)) 5860 return 1; 5861 5862 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent))) 5863 return 0; 5864 5865 /* Flags needing groups don't count if only 1 group in parent */ 5866 if (parent->groups == parent->groups->next) { 5867 pflags &= ~(SD_LOAD_BALANCE | 5868 SD_BALANCE_NEWIDLE | 5869 SD_BALANCE_FORK | 5870 SD_BALANCE_EXEC | 5871 SD_SHARE_CPUPOWER | 5872 SD_SHARE_PKG_RESOURCES); 5873 if (nr_node_ids == 1) 5874 pflags &= ~SD_SERIALIZE; 5875 } 5876 if (~cflags & pflags) 5877 return 0; 5878 5879 return 1; 5880} 5881 5882static void free_rootdomain(struct rcu_head *rcu) 5883{ 5884 struct root_domain *rd = container_of(rcu, struct root_domain, rcu); 5885 5886 cpupri_cleanup(&rd->cpupri); 5887 free_cpumask_var(rd->rto_mask); 5888 free_cpumask_var(rd->online); 5889 free_cpumask_var(rd->span); 5890 kfree(rd); 5891} 5892 5893static void rq_attach_root(struct rq *rq, struct root_domain *rd) 5894{ 5895 struct root_domain *old_rd = NULL; 5896 unsigned long flags; 5897 5898 raw_spin_lock_irqsave(&rq->lock, flags); 5899 5900 if (rq->rd) { 5901 old_rd = rq->rd; 5902 5903 if (cpumask_test_cpu(rq->cpu, old_rd->online)) 5904 set_rq_offline(rq); 5905 5906 cpumask_clear_cpu(rq->cpu, old_rd->span); 5907 5908 /* 5909 * If we dont want to free the old_rt yet then 5910 * set old_rd to NULL to skip the freeing later 5911 * in this function: 5912 */ 5913 if (!atomic_dec_and_test(&old_rd->refcount)) 5914 old_rd = NULL; 5915 } 5916 5917 atomic_inc(&rd->refcount); 5918 rq->rd = rd; 5919 5920 cpumask_set_cpu(rq->cpu, rd->span); 5921 if (cpumask_test_cpu(rq->cpu, cpu_active_mask)) 5922 set_rq_online(rq); 5923 5924 raw_spin_unlock_irqrestore(&rq->lock, flags); 5925 5926 if (old_rd) 5927 call_rcu_sched(&old_rd->rcu, free_rootdomain); 5928} 5929 5930static int init_rootdomain(struct root_domain *rd) 5931{ 5932 memset(rd, 0, sizeof(*rd)); 5933 5934 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL)) 5935 goto out; 5936 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL)) 5937 goto free_span; 5938 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL)) 5939 goto free_online; 5940 5941 if (cpupri_init(&rd->cpupri) != 0) 5942 goto free_rto_mask; 5943 return 0; 5944 5945free_rto_mask: 5946 free_cpumask_var(rd->rto_mask); 5947free_online: 5948 free_cpumask_var(rd->online); 5949free_span: 5950 free_cpumask_var(rd->span); 5951out: 5952 return -ENOMEM; 5953} 5954 5955/* 5956 * By default the system creates a single root-domain with all cpus as 5957 * members (mimicking the global state we have today). 5958 */ 5959struct root_domain def_root_domain; 5960 5961static void init_defrootdomain(void) 5962{ 5963 init_rootdomain(&def_root_domain); 5964 5965 atomic_set(&def_root_domain.refcount, 1); 5966} 5967 5968static struct root_domain *alloc_rootdomain(void) 5969{ 5970 struct root_domain *rd; 5971 5972 rd = kmalloc(sizeof(*rd), GFP_KERNEL); 5973 if (!rd) 5974 return NULL; 5975 5976 if (init_rootdomain(rd) != 0) { 5977 kfree(rd); 5978 return NULL; 5979 } 5980 5981 return rd; 5982} 5983 5984static void free_sched_groups(struct sched_group *sg, int free_sgp) 5985{ 5986 struct sched_group *tmp, *first; 5987 5988 if (!sg) 5989 return; 5990 5991 first = sg; 5992 do { 5993 tmp = sg->next; 5994 5995 if (free_sgp && atomic_dec_and_test(&sg->sgp->ref)) 5996 kfree(sg->sgp); 5997 5998 kfree(sg); 5999 sg = tmp; 6000 } while (sg != first); 6001} 6002 6003static void free_sched_domain(struct rcu_head *rcu) 6004{ 6005 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu); 6006 6007 /* 6008 * If its an overlapping domain it has private groups, iterate and 6009 * nuke them all. 6010 */ 6011 if (sd->flags & SD_OVERLAP) { 6012 free_sched_groups(sd->groups, 1); 6013 } else if (atomic_dec_and_test(&sd->groups->ref)) { 6014 kfree(sd->groups->sgp); 6015 kfree(sd->groups); 6016 } 6017 kfree(sd); 6018} 6019 6020static void destroy_sched_domain(struct sched_domain *sd, int cpu) 6021{ 6022 call_rcu(&sd->rcu, free_sched_domain); 6023} 6024 6025static void destroy_sched_domains(struct sched_domain *sd, int cpu) 6026{ 6027 for (; sd; sd = sd->parent) 6028 destroy_sched_domain(sd, cpu); 6029} 6030 6031/* 6032 * Keep a special pointer to the highest sched_domain that has 6033 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this 6034 * allows us to avoid some pointer chasing select_idle_sibling(). 6035 * 6036 * Also keep a unique ID per domain (we use the first cpu number in 6037 * the cpumask of the domain), this allows us to quickly tell if 6038 * two cpus are in the same cache domain, see cpus_share_cache(). 6039 */ 6040DEFINE_PER_CPU(struct sched_domain *, sd_llc); 6041DEFINE_PER_CPU(int, sd_llc_id); 6042 6043static void update_top_cache_domain(int cpu) 6044{ 6045 struct sched_domain *sd; 6046 int id = cpu; 6047 6048 sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES); 6049 if (sd) 6050 id = cpumask_first(sched_domain_span(sd)); 6051 6052 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd); 6053 per_cpu(sd_llc_id, cpu) = id; 6054} 6055 6056/* 6057 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must 6058 * hold the hotplug lock. 6059 */ 6060static void 6061cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu) 6062{ 6063 struct rq *rq = cpu_rq(cpu); 6064 struct sched_domain *tmp; 6065 6066 /* Remove the sched domains which do not contribute to scheduling. */ 6067 for (tmp = sd; tmp; ) { 6068 struct sched_domain *parent = tmp->parent; 6069 if (!parent) 6070 break; 6071 6072 if (sd_parent_degenerate(tmp, parent)) { 6073 tmp->parent = parent->parent; 6074 if (parent->parent) 6075 parent->parent->child = tmp; 6076 destroy_sched_domain(parent, cpu); 6077 } else 6078 tmp = tmp->parent; 6079 } 6080 6081 if (sd && sd_degenerate(sd)) { 6082 tmp = sd; 6083 sd = sd->parent; 6084 destroy_sched_domain(tmp, cpu); 6085 if (sd) 6086 sd->child = NULL; 6087 } 6088 6089 sched_domain_debug(sd, cpu); 6090 6091 rq_attach_root(rq, rd); 6092 tmp = rq->sd; 6093 rcu_assign_pointer(rq->sd, sd); 6094 destroy_sched_domains(tmp, cpu); 6095 6096 update_top_cache_domain(cpu); 6097} 6098 6099/* cpus with isolated domains */ 6100static cpumask_var_t cpu_isolated_map; 6101 6102/* Setup the mask of cpus configured for isolated domains */ 6103static int __init isolated_cpu_setup(char *str) 6104{ 6105 alloc_bootmem_cpumask_var(&cpu_isolated_map); 6106 cpulist_parse(str, cpu_isolated_map); 6107 return 1; 6108} 6109 6110__setup("isolcpus=", isolated_cpu_setup); 6111 6112static const struct cpumask *cpu_cpu_mask(int cpu) 6113{ 6114 return cpumask_of_node(cpu_to_node(cpu)); 6115} 6116 6117struct sd_data { 6118 struct sched_domain **__percpu sd; 6119 struct sched_group **__percpu sg; 6120 struct sched_group_power **__percpu sgp; 6121}; 6122 6123struct s_data { 6124 struct sched_domain ** __percpu sd; 6125 struct root_domain *rd; 6126}; 6127 6128enum s_alloc { 6129 sa_rootdomain, 6130 sa_sd, 6131 sa_sd_storage, 6132 sa_none, 6133}; 6134 6135struct sched_domain_topology_level; 6136 6137typedef struct sched_domain *(*sched_domain_init_f)(struct sched_domain_topology_level *tl, int cpu); 6138typedef const struct cpumask *(*sched_domain_mask_f)(int cpu); 6139 6140#define SDTL_OVERLAP 0x01 6141 6142struct sched_domain_topology_level { 6143 sched_domain_init_f init; 6144 sched_domain_mask_f mask; 6145 int flags; 6146 int numa_level; 6147 struct sd_data data; 6148}; 6149 6150/* 6151 * Build an iteration mask that can exclude certain CPUs from the upwards 6152 * domain traversal. 6153 * 6154 * Asymmetric node setups can result in situations where the domain tree is of 6155 * unequal depth, make sure to skip domains that already cover the entire 6156 * range. 6157 * 6158 * In that case build_sched_domains() will have terminated the iteration early 6159 * and our sibling sd spans will be empty. Domains should always include the 6160 * cpu they're built on, so check that. 6161 * 6162 */ 6163static void build_group_mask(struct sched_domain *sd, struct sched_group *sg) 6164{ 6165 const struct cpumask *span = sched_domain_span(sd); 6166 struct sd_data *sdd = sd->private; 6167 struct sched_domain *sibling; 6168 int i; 6169 6170 for_each_cpu(i, span) { 6171 sibling = *per_cpu_ptr(sdd->sd, i); 6172 if (!cpumask_test_cpu(i, sched_domain_span(sibling))) 6173 continue; 6174 6175 cpumask_set_cpu(i, sched_group_mask(sg)); 6176 } 6177} 6178 6179/* 6180 * Return the canonical balance cpu for this group, this is the first cpu 6181 * of this group that's also in the iteration mask. 6182 */ 6183int group_balance_cpu(struct sched_group *sg) 6184{ 6185 return cpumask_first_and(sched_group_cpus(sg), sched_group_mask(sg)); 6186} 6187 6188static int 6189build_overlap_sched_groups(struct sched_domain *sd, int cpu) 6190{ 6191 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg; 6192 const struct cpumask *span = sched_domain_span(sd); 6193 struct cpumask *covered = sched_domains_tmpmask; 6194 struct sd_data *sdd = sd->private; 6195 struct sched_domain *child; 6196 int i; 6197 6198 cpumask_clear(covered); 6199 6200 for_each_cpu(i, span) { 6201 struct cpumask *sg_span; 6202 6203 if (cpumask_test_cpu(i, covered)) 6204 continue; 6205 6206 child = *per_cpu_ptr(sdd->sd, i); 6207 6208 /* See the comment near build_group_mask(). */ 6209 if (!cpumask_test_cpu(i, sched_domain_span(child))) 6210 continue; 6211 6212 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(), 6213 GFP_KERNEL, cpu_to_node(cpu)); 6214 6215 if (!sg) 6216 goto fail; 6217 6218 sg_span = sched_group_cpus(sg); 6219 if (child->child) { 6220 child = child->child; 6221 cpumask_copy(sg_span, sched_domain_span(child)); 6222 } else 6223 cpumask_set_cpu(i, sg_span); 6224 6225 cpumask_or(covered, covered, sg_span); 6226 6227 sg->sgp = *per_cpu_ptr(sdd->sgp, i); 6228 if (atomic_inc_return(&sg->sgp->ref) == 1) 6229 build_group_mask(sd, sg); 6230 6231 /* 6232 * Initialize sgp->power such that even if we mess up the 6233 * domains and no possible iteration will get us here, we won't 6234 * die on a /0 trap. 6235 */ 6236 sg->sgp->power = SCHED_POWER_SCALE * cpumask_weight(sg_span); 6237 6238 /* 6239 * Make sure the first group of this domain contains the 6240 * canonical balance cpu. Otherwise the sched_domain iteration 6241 * breaks. See update_sg_lb_stats(). 6242 */ 6243 if ((!groups && cpumask_test_cpu(cpu, sg_span)) || 6244 group_balance_cpu(sg) == cpu) 6245 groups = sg; 6246 6247 if (!first) 6248 first = sg; 6249 if (last) 6250 last->next = sg; 6251 last = sg; 6252 last->next = first; 6253 } 6254 sd->groups = groups; 6255 6256 return 0; 6257 6258fail: 6259 free_sched_groups(first, 0); 6260 6261 return -ENOMEM; 6262} 6263 6264static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg) 6265{ 6266 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu); 6267 struct sched_domain *child = sd->child; 6268 6269 if (child) 6270 cpu = cpumask_first(sched_domain_span(child)); 6271 6272 if (sg) { 6273 *sg = *per_cpu_ptr(sdd->sg, cpu); 6274 (*sg)->sgp = *per_cpu_ptr(sdd->sgp, cpu); 6275 atomic_set(&(*sg)->sgp->ref, 1); /* for claim_allocations */ 6276 } 6277 6278 return cpu; 6279} 6280 6281/* 6282 * build_sched_groups will build a circular linked list of the groups 6283 * covered by the given span, and will set each group's ->cpumask correctly, 6284 * and ->cpu_power to 0. 6285 * 6286 * Assumes the sched_domain tree is fully constructed 6287 */ 6288static int 6289build_sched_groups(struct sched_domain *sd, int cpu) 6290{ 6291 struct sched_group *first = NULL, *last = NULL; 6292 struct sd_data *sdd = sd->private; 6293 const struct cpumask *span = sched_domain_span(sd); 6294 struct cpumask *covered; 6295 int i; 6296 6297 get_group(cpu, sdd, &sd->groups); 6298 atomic_inc(&sd->groups->ref); 6299 6300 if (cpu != cpumask_first(sched_domain_span(sd))) 6301 return 0; 6302 6303 lockdep_assert_held(&sched_domains_mutex); 6304 covered = sched_domains_tmpmask; 6305 6306 cpumask_clear(covered); 6307 6308 for_each_cpu(i, span) { 6309 struct sched_group *sg; 6310 int group = get_group(i, sdd, &sg); 6311 int j; 6312 6313 if (cpumask_test_cpu(i, covered)) 6314 continue; 6315 6316 cpumask_clear(sched_group_cpus(sg)); 6317 sg->sgp->power = 0; 6318 cpumask_setall(sched_group_mask(sg)); 6319 6320 for_each_cpu(j, span) { 6321 if (get_group(j, sdd, NULL) != group) 6322 continue; 6323 6324 cpumask_set_cpu(j, covered); 6325 cpumask_set_cpu(j, sched_group_cpus(sg)); 6326 } 6327 6328 if (!first) 6329 first = sg; 6330 if (last) 6331 last->next = sg; 6332 last = sg; 6333 } 6334 last->next = first; 6335 6336 return 0; 6337} 6338 6339/* 6340 * Initialize sched groups cpu_power. 6341 * 6342 * cpu_power indicates the capacity of sched group, which is used while 6343 * distributing the load between different sched groups in a sched domain. 6344 * Typically cpu_power for all the groups in a sched domain will be same unless 6345 * there are asymmetries in the topology. If there are asymmetries, group 6346 * having more cpu_power will pickup more load compared to the group having 6347 * less cpu_power. 6348 */ 6349static void init_sched_groups_power(int cpu, struct sched_domain *sd) 6350{ 6351 struct sched_group *sg = sd->groups; 6352 6353 WARN_ON(!sd || !sg); 6354 6355 do { 6356 sg->group_weight = cpumask_weight(sched_group_cpus(sg)); 6357 sg = sg->next; 6358 } while (sg != sd->groups); 6359 6360 if (cpu != group_balance_cpu(sg)) 6361 return; 6362 6363 update_group_power(sd, cpu); 6364 atomic_set(&sg->sgp->nr_busy_cpus, sg->group_weight); 6365} 6366 6367int __weak arch_sd_sibling_asym_packing(void) 6368{ 6369 return 0*SD_ASYM_PACKING; 6370} 6371 6372/* 6373 * Initializers for schedule domains 6374 * Non-inlined to reduce accumulated stack pressure in build_sched_domains() 6375 */ 6376 6377#ifdef CONFIG_SCHED_DEBUG 6378# define SD_INIT_NAME(sd, type) sd->name = #type 6379#else 6380# define SD_INIT_NAME(sd, type) do { } while (0) 6381#endif 6382 6383#define SD_INIT_FUNC(type) \ 6384static noinline struct sched_domain * \ 6385sd_init_##type(struct sched_domain_topology_level *tl, int cpu) \ 6386{ \ 6387 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); \ 6388 *sd = SD_##type##_INIT; \ 6389 SD_INIT_NAME(sd, type); \ 6390 sd->private = &tl->data; \ 6391 return sd; \ 6392} 6393 6394SD_INIT_FUNC(CPU) 6395#ifdef CONFIG_SCHED_SMT 6396 SD_INIT_FUNC(SIBLING) 6397#endif 6398#ifdef CONFIG_SCHED_MC 6399 SD_INIT_FUNC(MC) 6400#endif 6401#ifdef CONFIG_SCHED_BOOK 6402 SD_INIT_FUNC(BOOK) 6403#endif 6404 6405static int default_relax_domain_level = -1; 6406int sched_domain_level_max; 6407 6408static int __init setup_relax_domain_level(char *str) 6409{ 6410 if (kstrtoint(str, 0, &default_relax_domain_level)) 6411 pr_warn("Unable to set relax_domain_level\n"); 6412 6413 return 1; 6414} 6415__setup("relax_domain_level=", setup_relax_domain_level); 6416 6417static void set_domain_attribute(struct sched_domain *sd, 6418 struct sched_domain_attr *attr) 6419{ 6420 int request; 6421 6422 if (!attr || attr->relax_domain_level < 0) { 6423 if (default_relax_domain_level < 0) 6424 return; 6425 else 6426 request = default_relax_domain_level; 6427 } else 6428 request = attr->relax_domain_level; 6429 if (request < sd->level) { 6430 /* turn off idle balance on this domain */ 6431 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE); 6432 } else { 6433 /* turn on idle balance on this domain */ 6434 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE); 6435 } 6436} 6437 6438static void __sdt_free(const struct cpumask *cpu_map); 6439static int __sdt_alloc(const struct cpumask *cpu_map); 6440 6441static void __free_domain_allocs(struct s_data *d, enum s_alloc what, 6442 const struct cpumask *cpu_map) 6443{ 6444 switch (what) { 6445 case sa_rootdomain: 6446 if (!atomic_read(&d->rd->refcount)) 6447 free_rootdomain(&d->rd->rcu); /* fall through */ 6448 case sa_sd: 6449 free_percpu(d->sd); /* fall through */ 6450 case sa_sd_storage: 6451 __sdt_free(cpu_map); /* fall through */ 6452 case sa_none: 6453 break; 6454 } 6455} 6456 6457static enum s_alloc __visit_domain_allocation_hell(struct s_data *d, 6458 const struct cpumask *cpu_map) 6459{ 6460 memset(d, 0, sizeof(*d)); 6461 6462 if (__sdt_alloc(cpu_map)) 6463 return sa_sd_storage; 6464 d->sd = alloc_percpu(struct sched_domain *); 6465 if (!d->sd) 6466 return sa_sd_storage; 6467 d->rd = alloc_rootdomain(); 6468 if (!d->rd) 6469 return sa_sd; 6470 return sa_rootdomain; 6471} 6472 6473/* 6474 * NULL the sd_data elements we've used to build the sched_domain and 6475 * sched_group structure so that the subsequent __free_domain_allocs() 6476 * will not free the data we're using. 6477 */ 6478static void claim_allocations(int cpu, struct sched_domain *sd) 6479{ 6480 struct sd_data *sdd = sd->private; 6481 6482 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd); 6483 *per_cpu_ptr(sdd->sd, cpu) = NULL; 6484 6485 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref)) 6486 *per_cpu_ptr(sdd->sg, cpu) = NULL; 6487 6488 if (atomic_read(&(*per_cpu_ptr(sdd->sgp, cpu))->ref)) 6489 *per_cpu_ptr(sdd->sgp, cpu) = NULL; 6490} 6491 6492#ifdef CONFIG_SCHED_SMT 6493static const struct cpumask *cpu_smt_mask(int cpu) 6494{ 6495 return topology_thread_cpumask(cpu); 6496} 6497#endif 6498 6499/* 6500 * Topology list, bottom-up. 6501 */ 6502static struct sched_domain_topology_level default_topology[] = { 6503#ifdef CONFIG_SCHED_SMT 6504 { sd_init_SIBLING, cpu_smt_mask, }, 6505#endif 6506#ifdef CONFIG_SCHED_MC 6507 { sd_init_MC, cpu_coregroup_mask, }, 6508#endif 6509#ifdef CONFIG_SCHED_BOOK 6510 { sd_init_BOOK, cpu_book_mask, }, 6511#endif 6512 { sd_init_CPU, cpu_cpu_mask, }, 6513 { NULL, }, 6514}; 6515 6516static struct sched_domain_topology_level *sched_domain_topology = default_topology; 6517 6518#ifdef CONFIG_NUMA 6519 6520static int sched_domains_numa_levels; 6521static int *sched_domains_numa_distance; 6522static struct cpumask ***sched_domains_numa_masks; 6523static int sched_domains_curr_level; 6524 6525static inline int sd_local_flags(int level) 6526{ 6527 if (sched_domains_numa_distance[level] > RECLAIM_DISTANCE) 6528 return 0; 6529 6530 return SD_BALANCE_EXEC | SD_BALANCE_FORK | SD_WAKE_AFFINE; 6531} 6532 6533static struct sched_domain * 6534sd_numa_init(struct sched_domain_topology_level *tl, int cpu) 6535{ 6536 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); 6537 int level = tl->numa_level; 6538 int sd_weight = cpumask_weight( 6539 sched_domains_numa_masks[level][cpu_to_node(cpu)]); 6540 6541 *sd = (struct sched_domain){ 6542 .min_interval = sd_weight, 6543 .max_interval = 2*sd_weight, 6544 .busy_factor = 32, 6545 .imbalance_pct = 125, 6546 .cache_nice_tries = 2, 6547 .busy_idx = 3, 6548 .idle_idx = 2, 6549 .newidle_idx = 0, 6550 .wake_idx = 0, 6551 .forkexec_idx = 0, 6552 6553 .flags = 1*SD_LOAD_BALANCE 6554 | 1*SD_BALANCE_NEWIDLE 6555 | 0*SD_BALANCE_EXEC 6556 | 0*SD_BALANCE_FORK 6557 | 0*SD_BALANCE_WAKE 6558 | 0*SD_WAKE_AFFINE 6559 | 0*SD_PREFER_LOCAL 6560 | 0*SD_SHARE_CPUPOWER 6561 | 0*SD_SHARE_PKG_RESOURCES 6562 | 1*SD_SERIALIZE 6563 | 0*SD_PREFER_SIBLING 6564 | sd_local_flags(level) 6565 , 6566 .last_balance = jiffies, 6567 .balance_interval = sd_weight, 6568 }; 6569 SD_INIT_NAME(sd, NUMA); 6570 sd->private = &tl->data; 6571 6572 /* 6573 * Ugly hack to pass state to sd_numa_mask()... 6574 */ 6575 sched_domains_curr_level = tl->numa_level; 6576 6577 return sd; 6578} 6579 6580static const struct cpumask *sd_numa_mask(int cpu) 6581{ 6582 return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)]; 6583} 6584 6585static void sched_numa_warn(const char *str) 6586{ 6587 static int done = false; 6588 int i,j; 6589 6590 if (done) 6591 return; 6592 6593 done = true; 6594 6595 printk(KERN_WARNING "ERROR: %s\n\n", str); 6596 6597 for (i = 0; i < nr_node_ids; i++) { 6598 printk(KERN_WARNING " "); 6599 for (j = 0; j < nr_node_ids; j++) 6600 printk(KERN_CONT "%02d ", node_distance(i,j)); 6601 printk(KERN_CONT "\n"); 6602 } 6603 printk(KERN_WARNING "\n"); 6604} 6605 6606static bool find_numa_distance(int distance) 6607{ 6608 int i; 6609 6610 if (distance == node_distance(0, 0)) 6611 return true; 6612 6613 for (i = 0; i < sched_domains_numa_levels; i++) { 6614 if (sched_domains_numa_distance[i] == distance) 6615 return true; 6616 } 6617 6618 return false; 6619} 6620 6621static void sched_init_numa(void) 6622{ 6623 int next_distance, curr_distance = node_distance(0, 0); 6624 struct sched_domain_topology_level *tl; 6625 int level = 0; 6626 int i, j, k; 6627 6628 sched_domains_numa_distance = kzalloc(sizeof(int) * nr_node_ids, GFP_KERNEL); 6629 if (!sched_domains_numa_distance) 6630 return; 6631 6632 /* 6633 * O(nr_nodes^2) deduplicating selection sort -- in order to find the 6634 * unique distances in the node_distance() table. 6635 * 6636 * Assumes node_distance(0,j) includes all distances in 6637 * node_distance(i,j) in order to avoid cubic time. 6638 */ 6639 next_distance = curr_distance; 6640 for (i = 0; i < nr_node_ids; i++) { 6641 for (j = 0; j < nr_node_ids; j++) { 6642 for (k = 0; k < nr_node_ids; k++) { 6643 int distance = node_distance(i, k); 6644 6645 if (distance > curr_distance && 6646 (distance < next_distance || 6647 next_distance == curr_distance)) 6648 next_distance = distance; 6649 6650 /* 6651 * While not a strong assumption it would be nice to know 6652 * about cases where if node A is connected to B, B is not 6653 * equally connected to A. 6654 */ 6655 if (sched_debug() && node_distance(k, i) != distance) 6656 sched_numa_warn("Node-distance not symmetric"); 6657 6658 if (sched_debug() && i && !find_numa_distance(distance)) 6659 sched_numa_warn("Node-0 not representative"); 6660 } 6661 if (next_distance != curr_distance) { 6662 sched_domains_numa_distance[level++] = next_distance; 6663 sched_domains_numa_levels = level; 6664 curr_distance = next_distance; 6665 } else break; 6666 } 6667 6668 /* 6669 * In case of sched_debug() we verify the above assumption. 6670 */ 6671 if (!sched_debug()) 6672 break; 6673 } 6674 /* 6675 * 'level' contains the number of unique distances, excluding the 6676 * identity distance node_distance(i,i). 6677 * 6678 * The sched_domains_nume_distance[] array includes the actual distance 6679 * numbers. 6680 */ 6681 6682 sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL); 6683 if (!sched_domains_numa_masks) 6684 return; 6685 6686 /* 6687 * Now for each level, construct a mask per node which contains all 6688 * cpus of nodes that are that many hops away from us. 6689 */ 6690 for (i = 0; i < level; i++) { 6691 sched_domains_numa_masks[i] = 6692 kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL); 6693 if (!sched_domains_numa_masks[i]) 6694 return; 6695 6696 for (j = 0; j < nr_node_ids; j++) { 6697 struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL); 6698 if (!mask) 6699 return; 6700 6701 sched_domains_numa_masks[i][j] = mask; 6702 6703 for (k = 0; k < nr_node_ids; k++) { 6704 if (node_distance(j, k) > sched_domains_numa_distance[i]) 6705 continue; 6706 6707 cpumask_or(mask, mask, cpumask_of_node(k)); 6708 } 6709 } 6710 } 6711 6712 tl = kzalloc((ARRAY_SIZE(default_topology) + level) * 6713 sizeof(struct sched_domain_topology_level), GFP_KERNEL); 6714 if (!tl) 6715 return; 6716 6717 /* 6718 * Copy the default topology bits.. 6719 */ 6720 for (i = 0; default_topology[i].init; i++) 6721 tl[i] = default_topology[i]; 6722 6723 /* 6724 * .. and append 'j' levels of NUMA goodness. 6725 */ 6726 for (j = 0; j < level; i++, j++) { 6727 tl[i] = (struct sched_domain_topology_level){ 6728 .init = sd_numa_init, 6729 .mask = sd_numa_mask, 6730 .flags = SDTL_OVERLAP, 6731 .numa_level = j, 6732 }; 6733 } 6734 6735 sched_domain_topology = tl; 6736} 6737#else 6738static inline void sched_init_numa(void) 6739{ 6740} 6741#endif /* CONFIG_NUMA */ 6742 6743static int __sdt_alloc(const struct cpumask *cpu_map) 6744{ 6745 struct sched_domain_topology_level *tl; 6746 int j; 6747 6748 for (tl = sched_domain_topology; tl->init; tl++) { 6749 struct sd_data *sdd = &tl->data; 6750 6751 sdd->sd = alloc_percpu(struct sched_domain *); 6752 if (!sdd->sd) 6753 return -ENOMEM; 6754 6755 sdd->sg = alloc_percpu(struct sched_group *); 6756 if (!sdd->sg) 6757 return -ENOMEM; 6758 6759 sdd->sgp = alloc_percpu(struct sched_group_power *); 6760 if (!sdd->sgp) 6761 return -ENOMEM; 6762 6763 for_each_cpu(j, cpu_map) { 6764 struct sched_domain *sd; 6765 struct sched_group *sg; 6766 struct sched_group_power *sgp; 6767 6768 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(), 6769 GFP_KERNEL, cpu_to_node(j)); 6770 if (!sd) 6771 return -ENOMEM; 6772 6773 *per_cpu_ptr(sdd->sd, j) = sd; 6774 6775 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(), 6776 GFP_KERNEL, cpu_to_node(j)); 6777 if (!sg) 6778 return -ENOMEM; 6779 6780 sg->next = sg; 6781 6782 *per_cpu_ptr(sdd->sg, j) = sg; 6783 6784 sgp = kzalloc_node(sizeof(struct sched_group_power) + cpumask_size(), 6785 GFP_KERNEL, cpu_to_node(j)); 6786 if (!sgp) 6787 return -ENOMEM; 6788 6789 *per_cpu_ptr(sdd->sgp, j) = sgp; 6790 } 6791 } 6792 6793 return 0; 6794} 6795 6796static void __sdt_free(const struct cpumask *cpu_map) 6797{ 6798 struct sched_domain_topology_level *tl; 6799 int j; 6800 6801 for (tl = sched_domain_topology; tl->init; tl++) { 6802 struct sd_data *sdd = &tl->data; 6803 6804 for_each_cpu(j, cpu_map) { 6805 struct sched_domain *sd; 6806 6807 if (sdd->sd) { 6808 sd = *per_cpu_ptr(sdd->sd, j); 6809 if (sd && (sd->flags & SD_OVERLAP)) 6810 free_sched_groups(sd->groups, 0); 6811 kfree(*per_cpu_ptr(sdd->sd, j)); 6812 } 6813 6814 if (sdd->sg) 6815 kfree(*per_cpu_ptr(sdd->sg, j)); 6816 if (sdd->sgp) 6817 kfree(*per_cpu_ptr(sdd->sgp, j)); 6818 } 6819 free_percpu(sdd->sd); 6820 sdd->sd = NULL; 6821 free_percpu(sdd->sg); 6822 sdd->sg = NULL; 6823 free_percpu(sdd->sgp); 6824 sdd->sgp = NULL; 6825 } 6826} 6827 6828struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl, 6829 struct s_data *d, const struct cpumask *cpu_map, 6830 struct sched_domain_attr *attr, struct sched_domain *child, 6831 int cpu) 6832{ 6833 struct sched_domain *sd = tl->init(tl, cpu); 6834 if (!sd) 6835 return child; 6836 6837 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu)); 6838 if (child) { 6839 sd->level = child->level + 1; 6840 sched_domain_level_max = max(sched_domain_level_max, sd->level); 6841 child->parent = sd; 6842 } 6843 sd->child = child; 6844 set_domain_attribute(sd, attr); 6845 6846 return sd; 6847} 6848 6849/* 6850 * Build sched domains for a given set of cpus and attach the sched domains 6851 * to the individual cpus 6852 */ 6853static int build_sched_domains(const struct cpumask *cpu_map, 6854 struct sched_domain_attr *attr) 6855{ 6856 enum s_alloc alloc_state = sa_none; 6857 struct sched_domain *sd; 6858 struct s_data d; 6859 int i, ret = -ENOMEM; 6860 6861 alloc_state = __visit_domain_allocation_hell(&d, cpu_map); 6862 if (alloc_state != sa_rootdomain) 6863 goto error; 6864 6865 /* Set up domains for cpus specified by the cpu_map. */ 6866 for_each_cpu(i, cpu_map) { 6867 struct sched_domain_topology_level *tl; 6868 6869 sd = NULL; 6870 for (tl = sched_domain_topology; tl->init; tl++) { 6871 sd = build_sched_domain(tl, &d, cpu_map, attr, sd, i); 6872 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP)) 6873 sd->flags |= SD_OVERLAP; 6874 if (cpumask_equal(cpu_map, sched_domain_span(sd))) 6875 break; 6876 } 6877 6878 while (sd->child) 6879 sd = sd->child; 6880 6881 *per_cpu_ptr(d.sd, i) = sd; 6882 } 6883 6884 /* Build the groups for the domains */ 6885 for_each_cpu(i, cpu_map) { 6886 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) { 6887 sd->span_weight = cpumask_weight(sched_domain_span(sd)); 6888 if (sd->flags & SD_OVERLAP) { 6889 if (build_overlap_sched_groups(sd, i)) 6890 goto error; 6891 } else { 6892 if (build_sched_groups(sd, i)) 6893 goto error; 6894 } 6895 } 6896 } 6897 6898 /* Calculate CPU power for physical packages and nodes */ 6899 for (i = nr_cpumask_bits-1; i >= 0; i--) { 6900 if (!cpumask_test_cpu(i, cpu_map)) 6901 continue; 6902 6903 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) { 6904 claim_allocations(i, sd); 6905 init_sched_groups_power(i, sd); 6906 } 6907 } 6908 6909 /* Attach the domains */ 6910 rcu_read_lock(); 6911 for_each_cpu(i, cpu_map) { 6912 sd = *per_cpu_ptr(d.sd, i); 6913 cpu_attach_domain(sd, d.rd, i); 6914 } 6915 rcu_read_unlock(); 6916 6917 ret = 0; 6918error: 6919 __free_domain_allocs(&d, alloc_state, cpu_map); 6920 return ret; 6921} 6922 6923static cpumask_var_t *doms_cur; /* current sched domains */ 6924static int ndoms_cur; /* number of sched domains in 'doms_cur' */ 6925static struct sched_domain_attr *dattr_cur; 6926 /* attribues of custom domains in 'doms_cur' */ 6927 6928/* 6929 * Special case: If a kmalloc of a doms_cur partition (array of 6930 * cpumask) fails, then fallback to a single sched domain, 6931 * as determined by the single cpumask fallback_doms. 6932 */ 6933static cpumask_var_t fallback_doms; 6934 6935/* 6936 * arch_update_cpu_topology lets virtualized architectures update the 6937 * cpu core maps. It is supposed to return 1 if the topology changed 6938 * or 0 if it stayed the same. 6939 */ 6940int __attribute__((weak)) arch_update_cpu_topology(void) 6941{ 6942 return 0; 6943} 6944 6945cpumask_var_t *alloc_sched_domains(unsigned int ndoms) 6946{ 6947 int i; 6948 cpumask_var_t *doms; 6949 6950 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL); 6951 if (!doms) 6952 return NULL; 6953 for (i = 0; i < ndoms; i++) { 6954 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) { 6955 free_sched_domains(doms, i); 6956 return NULL; 6957 } 6958 } 6959 return doms; 6960} 6961 6962void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms) 6963{ 6964 unsigned int i; 6965 for (i = 0; i < ndoms; i++) 6966 free_cpumask_var(doms[i]); 6967 kfree(doms); 6968} 6969 6970/* 6971 * Set up scheduler domains and groups. Callers must hold the hotplug lock. 6972 * For now this just excludes isolated cpus, but could be used to 6973 * exclude other special cases in the future. 6974 */ 6975static int init_sched_domains(const struct cpumask *cpu_map) 6976{ 6977 int err; 6978 6979 arch_update_cpu_topology(); 6980 ndoms_cur = 1; 6981 doms_cur = alloc_sched_domains(ndoms_cur); 6982 if (!doms_cur) 6983 doms_cur = &fallback_doms; 6984 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map); 6985 err = build_sched_domains(doms_cur[0], NULL); 6986 register_sched_domain_sysctl(); 6987 6988 return err; 6989} 6990 6991/* 6992 * Detach sched domains from a group of cpus specified in cpu_map 6993 * These cpus will now be attached to the NULL domain 6994 */ 6995static void detach_destroy_domains(const struct cpumask *cpu_map) 6996{ 6997 int i; 6998 6999 rcu_read_lock(); 7000 for_each_cpu(i, cpu_map) 7001 cpu_attach_domain(NULL, &def_root_domain, i); 7002 rcu_read_unlock(); 7003} 7004 7005/* handle null as "default" */ 7006static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur, 7007 struct sched_domain_attr *new, int idx_new) 7008{ 7009 struct sched_domain_attr tmp; 7010 7011 /* fast path */ 7012 if (!new && !cur) 7013 return 1; 7014 7015 tmp = SD_ATTR_INIT; 7016 return !memcmp(cur ? (cur + idx_cur) : &tmp, 7017 new ? (new + idx_new) : &tmp, 7018 sizeof(struct sched_domain_attr)); 7019} 7020 7021/* 7022 * Partition sched domains as specified by the 'ndoms_new' 7023 * cpumasks in the array doms_new[] of cpumasks. This compares 7024 * doms_new[] to the current sched domain partitioning, doms_cur[]. 7025 * It destroys each deleted domain and builds each new domain. 7026 * 7027 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'. 7028 * The masks don't intersect (don't overlap.) We should setup one 7029 * sched domain for each mask. CPUs not in any of the cpumasks will 7030 * not be load balanced. If the same cpumask appears both in the 7031 * current 'doms_cur' domains and in the new 'doms_new', we can leave 7032 * it as it is. 7033 * 7034 * The passed in 'doms_new' should be allocated using 7035 * alloc_sched_domains. This routine takes ownership of it and will 7036 * free_sched_domains it when done with it. If the caller failed the 7037 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1, 7038 * and partition_sched_domains() will fallback to the single partition 7039 * 'fallback_doms', it also forces the domains to be rebuilt. 7040 * 7041 * If doms_new == NULL it will be replaced with cpu_online_mask. 7042 * ndoms_new == 0 is a special case for destroying existing domains, 7043 * and it will not create the default domain. 7044 * 7045 * Call with hotplug lock held 7046 */ 7047void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[], 7048 struct sched_domain_attr *dattr_new) 7049{ 7050 int i, j, n; 7051 int new_topology; 7052 7053 mutex_lock(&sched_domains_mutex); 7054 7055 /* always unregister in case we don't destroy any domains */ 7056 unregister_sched_domain_sysctl(); 7057 7058 /* Let architecture update cpu core mappings. */ 7059 new_topology = arch_update_cpu_topology(); 7060 7061 n = doms_new ? ndoms_new : 0; 7062 7063 /* Destroy deleted domains */ 7064 for (i = 0; i < ndoms_cur; i++) { 7065 for (j = 0; j < n && !new_topology; j++) { 7066 if (cpumask_equal(doms_cur[i], doms_new[j]) 7067 && dattrs_equal(dattr_cur, i, dattr_new, j)) 7068 goto match1; 7069 } 7070 /* no match - a current sched domain not in new doms_new[] */ 7071 detach_destroy_domains(doms_cur[i]); 7072match1: 7073 ; 7074 } 7075 7076 if (doms_new == NULL) { 7077 ndoms_cur = 0; 7078 doms_new = &fallback_doms; 7079 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map); 7080 WARN_ON_ONCE(dattr_new); 7081 } 7082 7083 /* Build new domains */ 7084 for (i = 0; i < ndoms_new; i++) { 7085 for (j = 0; j < ndoms_cur && !new_topology; j++) { 7086 if (cpumask_equal(doms_new[i], doms_cur[j]) 7087 && dattrs_equal(dattr_new, i, dattr_cur, j)) 7088 goto match2; 7089 } 7090 /* no match - add a new doms_new */ 7091 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL); 7092match2: 7093 ; 7094 } 7095 7096 /* Remember the new sched domains */ 7097 if (doms_cur != &fallback_doms) 7098 free_sched_domains(doms_cur, ndoms_cur); 7099 kfree(dattr_cur); /* kfree(NULL) is safe */ 7100 doms_cur = doms_new; 7101 dattr_cur = dattr_new; 7102 ndoms_cur = ndoms_new; 7103 7104 register_sched_domain_sysctl(); 7105 7106 mutex_unlock(&sched_domains_mutex); 7107} 7108 7109static int num_cpus_frozen; /* used to mark begin/end of suspend/resume */ 7110 7111/* 7112 * Update cpusets according to cpu_active mask. If cpusets are 7113 * disabled, cpuset_update_active_cpus() becomes a simple wrapper 7114 * around partition_sched_domains(). 7115 * 7116 * If we come here as part of a suspend/resume, don't touch cpusets because we 7117 * want to restore it back to its original state upon resume anyway. 7118 */ 7119static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action, 7120 void *hcpu) 7121{ 7122 switch (action) { 7123 case CPU_ONLINE_FROZEN: 7124 case CPU_DOWN_FAILED_FROZEN: 7125 7126 /* 7127 * num_cpus_frozen tracks how many CPUs are involved in suspend 7128 * resume sequence. As long as this is not the last online 7129 * operation in the resume sequence, just build a single sched 7130 * domain, ignoring cpusets. 7131 */ 7132 num_cpus_frozen--; 7133 if (likely(num_cpus_frozen)) { 7134 partition_sched_domains(1, NULL, NULL); 7135 break; 7136 } 7137 7138 /* 7139 * This is the last CPU online operation. So fall through and 7140 * restore the original sched domains by considering the 7141 * cpuset configurations. 7142 */ 7143 7144 case CPU_ONLINE: 7145 case CPU_DOWN_FAILED: 7146 cpuset_update_active_cpus(true); 7147 break; 7148 default: 7149 return NOTIFY_DONE; 7150 } 7151 return NOTIFY_OK; 7152} 7153 7154static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action, 7155 void *hcpu) 7156{ 7157 switch (action) { 7158 case CPU_DOWN_PREPARE: 7159 cpuset_update_active_cpus(false); 7160 break; 7161 case CPU_DOWN_PREPARE_FROZEN: 7162 num_cpus_frozen++; 7163 partition_sched_domains(1, NULL, NULL); 7164 break; 7165 default: 7166 return NOTIFY_DONE; 7167 } 7168 return NOTIFY_OK; 7169} 7170 7171void __init sched_init_smp(void) 7172{ 7173 cpumask_var_t non_isolated_cpus; 7174 7175 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL); 7176 alloc_cpumask_var(&fallback_doms, GFP_KERNEL); 7177 7178 sched_init_numa(); 7179 7180 get_online_cpus(); 7181 mutex_lock(&sched_domains_mutex); 7182 init_sched_domains(cpu_active_mask); 7183 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map); 7184 if (cpumask_empty(non_isolated_cpus)) 7185 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus); 7186 mutex_unlock(&sched_domains_mutex); 7187 put_online_cpus(); 7188 7189 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE); 7190 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE); 7191 7192 /* RT runtime code needs to handle some hotplug events */ 7193 hotcpu_notifier(update_runtime, 0); 7194 7195 init_hrtick(); 7196 7197 /* Move init over to a non-isolated CPU */ 7198 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0) 7199 BUG(); 7200 sched_init_granularity(); 7201 free_cpumask_var(non_isolated_cpus); 7202 7203 init_sched_rt_class(); 7204} 7205#else 7206void __init sched_init_smp(void) 7207{ 7208 sched_init_granularity(); 7209} 7210#endif /* CONFIG_SMP */ 7211 7212const_debug unsigned int sysctl_timer_migration = 1; 7213 7214int in_sched_functions(unsigned long addr) 7215{ 7216 return in_lock_functions(addr) || 7217 (addr >= (unsigned long)__sched_text_start 7218 && addr < (unsigned long)__sched_text_end); 7219} 7220 7221#ifdef CONFIG_CGROUP_SCHED 7222struct task_group root_task_group; 7223LIST_HEAD(task_groups); 7224#endif 7225 7226DECLARE_PER_CPU(cpumask_var_t, load_balance_tmpmask); 7227 7228void __init sched_init(void) 7229{ 7230 int i, j; 7231 unsigned long alloc_size = 0, ptr; 7232 7233#ifdef CONFIG_FAIR_GROUP_SCHED 7234 alloc_size += 2 * nr_cpu_ids * sizeof(void **); 7235#endif 7236#ifdef CONFIG_RT_GROUP_SCHED 7237 alloc_size += 2 * nr_cpu_ids * sizeof(void **); 7238#endif 7239#ifdef CONFIG_CPUMASK_OFFSTACK 7240 alloc_size += num_possible_cpus() * cpumask_size(); 7241#endif 7242 if (alloc_size) { 7243 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT); 7244 7245#ifdef CONFIG_FAIR_GROUP_SCHED 7246 root_task_group.se = (struct sched_entity **)ptr; 7247 ptr += nr_cpu_ids * sizeof(void **); 7248 7249 root_task_group.cfs_rq = (struct cfs_rq **)ptr; 7250 ptr += nr_cpu_ids * sizeof(void **); 7251 7252#endif /* CONFIG_FAIR_GROUP_SCHED */ 7253#ifdef CONFIG_RT_GROUP_SCHED 7254 root_task_group.rt_se = (struct sched_rt_entity **)ptr; 7255 ptr += nr_cpu_ids * sizeof(void **); 7256 7257 root_task_group.rt_rq = (struct rt_rq **)ptr; 7258 ptr += nr_cpu_ids * sizeof(void **); 7259 7260#endif /* CONFIG_RT_GROUP_SCHED */ 7261#ifdef CONFIG_CPUMASK_OFFSTACK 7262 for_each_possible_cpu(i) { 7263 per_cpu(load_balance_tmpmask, i) = (void *)ptr; 7264 ptr += cpumask_size(); 7265 } 7266#endif /* CONFIG_CPUMASK_OFFSTACK */ 7267 } 7268 7269#ifdef CONFIG_SMP 7270 init_defrootdomain(); 7271#endif 7272 7273 init_rt_bandwidth(&def_rt_bandwidth, 7274 global_rt_period(), global_rt_runtime()); 7275 7276#ifdef CONFIG_RT_GROUP_SCHED 7277 init_rt_bandwidth(&root_task_group.rt_bandwidth, 7278 global_rt_period(), global_rt_runtime()); 7279#endif /* CONFIG_RT_GROUP_SCHED */ 7280 7281#ifdef CONFIG_CGROUP_SCHED 7282 list_add(&root_task_group.list, &task_groups); 7283 INIT_LIST_HEAD(&root_task_group.children); 7284 INIT_LIST_HEAD(&root_task_group.siblings); 7285 autogroup_init(&init_task); 7286 7287#endif /* CONFIG_CGROUP_SCHED */ 7288 7289#ifdef CONFIG_CGROUP_CPUACCT 7290 root_cpuacct.cpustat = &kernel_cpustat; 7291 root_cpuacct.cpuusage = alloc_percpu(u64); 7292 /* Too early, not expected to fail */ 7293 BUG_ON(!root_cpuacct.cpuusage); 7294#endif 7295 for_each_possible_cpu(i) { 7296 struct rq *rq; 7297 7298 rq = cpu_rq(i); 7299 raw_spin_lock_init(&rq->lock); 7300 rq->nr_running = 0; 7301 rq->calc_load_active = 0; 7302 rq->calc_load_update = jiffies + LOAD_FREQ; 7303 init_cfs_rq(&rq->cfs); 7304 init_rt_rq(&rq->rt, rq); 7305#ifdef CONFIG_FAIR_GROUP_SCHED 7306 root_task_group.shares = ROOT_TASK_GROUP_LOAD; 7307 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list); 7308 /* 7309 * How much cpu bandwidth does root_task_group get? 7310 * 7311 * In case of task-groups formed thr' the cgroup filesystem, it 7312 * gets 100% of the cpu resources in the system. This overall 7313 * system cpu resource is divided among the tasks of 7314 * root_task_group and its child task-groups in a fair manner, 7315 * based on each entity's (task or task-group's) weight 7316 * (se->load.weight). 7317 * 7318 * In other words, if root_task_group has 10 tasks of weight 7319 * 1024) and two child groups A0 and A1 (of weight 1024 each), 7320 * then A0's share of the cpu resource is: 7321 * 7322 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33% 7323 * 7324 * We achieve this by letting root_task_group's tasks sit 7325 * directly in rq->cfs (i.e root_task_group->se[] = NULL). 7326 */ 7327 init_cfs_bandwidth(&root_task_group.cfs_bandwidth); 7328 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL); 7329#endif /* CONFIG_FAIR_GROUP_SCHED */ 7330 7331 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime; 7332#ifdef CONFIG_RT_GROUP_SCHED 7333 INIT_LIST_HEAD(&rq->leaf_rt_rq_list); 7334 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL); 7335#endif 7336 7337 for (j = 0; j < CPU_LOAD_IDX_MAX; j++) 7338 rq->cpu_load[j] = 0; 7339 7340 rq->last_load_update_tick = jiffies; 7341 7342#ifdef CONFIG_SMP 7343 rq->sd = NULL; 7344 rq->rd = NULL; 7345 rq->cpu_power = SCHED_POWER_SCALE; 7346 rq->post_schedule = 0; 7347 rq->active_balance = 0; 7348 rq->next_balance = jiffies; 7349 rq->push_cpu = 0; 7350 rq->cpu = i; 7351 rq->online = 0; 7352 rq->idle_stamp = 0; 7353 rq->avg_idle = 2*sysctl_sched_migration_cost; 7354 7355 INIT_LIST_HEAD(&rq->cfs_tasks); 7356 7357 rq_attach_root(rq, &def_root_domain); 7358#ifdef CONFIG_NO_HZ 7359 rq->nohz_flags = 0; 7360#endif 7361#endif 7362 init_rq_hrtick(rq); 7363 atomic_set(&rq->nr_iowait, 0); 7364 } 7365 7366 set_load_weight(&init_task); 7367 7368#ifdef CONFIG_PREEMPT_NOTIFIERS 7369 INIT_HLIST_HEAD(&init_task.preempt_notifiers); 7370#endif 7371 7372#ifdef CONFIG_RT_MUTEXES 7373 plist_head_init(&init_task.pi_waiters); 7374#endif 7375 7376 /* 7377 * The boot idle thread does lazy MMU switching as well: 7378 */ 7379 atomic_inc(&init_mm.mm_count); 7380 enter_lazy_tlb(&init_mm, current); 7381 7382 /* 7383 * Make us the idle thread. Technically, schedule() should not be 7384 * called from this thread, however somewhere below it might be, 7385 * but because we are the idle thread, we just pick up running again 7386 * when this runqueue becomes "idle". 7387 */ 7388 init_idle(current, smp_processor_id()); 7389 7390 calc_load_update = jiffies + LOAD_FREQ; 7391 7392 /* 7393 * During early bootup we pretend to be a normal task: 7394 */ 7395 current->sched_class = &fair_sched_class; 7396 7397#ifdef CONFIG_SMP 7398 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT); 7399 /* May be allocated at isolcpus cmdline parse time */ 7400 if (cpu_isolated_map == NULL) 7401 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT); 7402 idle_thread_set_boot_cpu(); 7403#endif 7404 init_sched_fair_class(); 7405 7406 scheduler_running = 1; 7407} 7408 7409#ifdef CONFIG_DEBUG_ATOMIC_SLEEP 7410static inline int preempt_count_equals(int preempt_offset) 7411{ 7412 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth(); 7413 7414 return (nested == preempt_offset); 7415} 7416 7417void __might_sleep(const char *file, int line, int preempt_offset) 7418{ 7419 static unsigned long prev_jiffy; /* ratelimiting */ 7420 7421 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */ 7422 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) || 7423 system_state != SYSTEM_RUNNING || oops_in_progress) 7424 return; 7425 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy) 7426 return; 7427 prev_jiffy = jiffies; 7428 7429 printk(KERN_ERR 7430 "BUG: sleeping function called from invalid context at %s:%d\n", 7431 file, line); 7432 printk(KERN_ERR 7433 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n", 7434 in_atomic(), irqs_disabled(), 7435 current->pid, current->comm); 7436 7437 debug_show_held_locks(current); 7438 if (irqs_disabled()) 7439 print_irqtrace_events(current); 7440 dump_stack(); 7441} 7442EXPORT_SYMBOL(__might_sleep); 7443#endif 7444 7445#ifdef CONFIG_MAGIC_SYSRQ 7446static void normalize_task(struct rq *rq, struct task_struct *p) 7447{ 7448 const struct sched_class *prev_class = p->sched_class; 7449 int old_prio = p->prio; 7450 int on_rq; 7451 7452 on_rq = p->on_rq; 7453 if (on_rq) 7454 dequeue_task(rq, p, 0); 7455 __setscheduler(rq, p, SCHED_NORMAL, 0); 7456 if (on_rq) { 7457 enqueue_task(rq, p, 0); 7458 resched_task(rq->curr); 7459 } 7460 7461 check_class_changed(rq, p, prev_class, old_prio); 7462} 7463 7464void normalize_rt_tasks(void) 7465{ 7466 struct task_struct *g, *p; 7467 unsigned long flags; 7468 struct rq *rq; 7469 7470 read_lock_irqsave(&tasklist_lock, flags); 7471 do_each_thread(g, p) { 7472 /* 7473 * Only normalize user tasks: 7474 */ 7475 if (!p->mm) 7476 continue; 7477 7478 p->se.exec_start = 0; 7479#ifdef CONFIG_SCHEDSTATS 7480 p->se.statistics.wait_start = 0; 7481 p->se.statistics.sleep_start = 0; 7482 p->se.statistics.block_start = 0; 7483#endif 7484 7485 if (!rt_task(p)) { 7486 /* 7487 * Renice negative nice level userspace 7488 * tasks back to 0: 7489 */ 7490 if (TASK_NICE(p) < 0 && p->mm) 7491 set_user_nice(p, 0); 7492 continue; 7493 } 7494 7495 raw_spin_lock(&p->pi_lock); 7496 rq = __task_rq_lock(p); 7497 7498 normalize_task(rq, p); 7499 7500 __task_rq_unlock(rq); 7501 raw_spin_unlock(&p->pi_lock); 7502 } while_each_thread(g, p); 7503 7504 read_unlock_irqrestore(&tasklist_lock, flags); 7505} 7506 7507#endif /* CONFIG_MAGIC_SYSRQ */ 7508 7509#if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) 7510/* 7511 * These functions are only useful for the IA64 MCA handling, or kdb. 7512 * 7513 * They can only be called when the whole system has been 7514 * stopped - every CPU needs to be quiescent, and no scheduling 7515 * activity can take place. Using them for anything else would 7516 * be a serious bug, and as a result, they aren't even visible 7517 * under any other configuration. 7518 */ 7519 7520/** 7521 * curr_task - return the current task for a given cpu. 7522 * @cpu: the processor in question. 7523 * 7524 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED! 7525 */ 7526struct task_struct *curr_task(int cpu) 7527{ 7528 return cpu_curr(cpu); 7529} 7530 7531#endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */ 7532 7533#ifdef CONFIG_IA64 7534/** 7535 * set_curr_task - set the current task for a given cpu. 7536 * @cpu: the processor in question. 7537 * @p: the task pointer to set. 7538 * 7539 * Description: This function must only be used when non-maskable interrupts 7540 * are serviced on a separate stack. It allows the architecture to switch the 7541 * notion of the current task on a cpu in a non-blocking manner. This function 7542 * must be called with all CPU's synchronized, and interrupts disabled, the 7543 * and caller must save the original value of the current task (see 7544 * curr_task() above) and restore that value before reenabling interrupts and 7545 * re-starting the system. 7546 * 7547 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED! 7548 */ 7549void set_curr_task(int cpu, struct task_struct *p) 7550{ 7551 cpu_curr(cpu) = p; 7552} 7553 7554#endif 7555 7556#ifdef CONFIG_CGROUP_SCHED 7557/* task_group_lock serializes the addition/removal of task groups */ 7558static DEFINE_SPINLOCK(task_group_lock); 7559 7560static void free_sched_group(struct task_group *tg) 7561{ 7562 free_fair_sched_group(tg); 7563 free_rt_sched_group(tg); 7564 autogroup_free(tg); 7565 kfree(tg); 7566} 7567 7568/* allocate runqueue etc for a new task group */ 7569struct task_group *sched_create_group(struct task_group *parent) 7570{ 7571 struct task_group *tg; 7572 unsigned long flags; 7573 7574 tg = kzalloc(sizeof(*tg), GFP_KERNEL); 7575 if (!tg) 7576 return ERR_PTR(-ENOMEM); 7577 7578 if (!alloc_fair_sched_group(tg, parent)) 7579 goto err; 7580 7581 if (!alloc_rt_sched_group(tg, parent)) 7582 goto err; 7583 7584 spin_lock_irqsave(&task_group_lock, flags); 7585 list_add_rcu(&tg->list, &task_groups); 7586 7587 WARN_ON(!parent); /* root should already exist */ 7588 7589 tg->parent = parent; 7590 INIT_LIST_HEAD(&tg->children); 7591 list_add_rcu(&tg->siblings, &parent->children); 7592 spin_unlock_irqrestore(&task_group_lock, flags); 7593 7594 return tg; 7595 7596err: 7597 free_sched_group(tg); 7598 return ERR_PTR(-ENOMEM); 7599} 7600 7601/* rcu callback to free various structures associated with a task group */ 7602static void free_sched_group_rcu(struct rcu_head *rhp) 7603{ 7604 /* now it should be safe to free those cfs_rqs */ 7605 free_sched_group(container_of(rhp, struct task_group, rcu)); 7606} 7607 7608/* Destroy runqueue etc associated with a task group */ 7609void sched_destroy_group(struct task_group *tg) 7610{ 7611 unsigned long flags; 7612 int i; 7613 7614 /* end participation in shares distribution */ 7615 for_each_possible_cpu(i) 7616 unregister_fair_sched_group(tg, i); 7617 7618 spin_lock_irqsave(&task_group_lock, flags); 7619 list_del_rcu(&tg->list); 7620 list_del_rcu(&tg->siblings); 7621 spin_unlock_irqrestore(&task_group_lock, flags); 7622 7623 /* wait for possible concurrent references to cfs_rqs complete */ 7624 call_rcu(&tg->rcu, free_sched_group_rcu); 7625} 7626 7627/* change task's runqueue when it moves between groups. 7628 * The caller of this function should have put the task in its new group 7629 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to 7630 * reflect its new group. 7631 */ 7632void sched_move_task(struct task_struct *tsk) 7633{ 7634 struct task_group *tg; 7635 int on_rq, running; 7636 unsigned long flags; 7637 struct rq *rq; 7638 7639 rq = task_rq_lock(tsk, &flags); 7640 7641 running = task_current(rq, tsk); 7642 on_rq = tsk->on_rq; 7643 7644 if (on_rq) 7645 dequeue_task(rq, tsk, 0); 7646 if (unlikely(running)) 7647 tsk->sched_class->put_prev_task(rq, tsk); 7648 7649 tg = container_of(task_subsys_state_check(tsk, cpu_cgroup_subsys_id, 7650 lockdep_is_held(&tsk->sighand->siglock)), 7651 struct task_group, css); 7652 tg = autogroup_task_group(tsk, tg); 7653 tsk->sched_task_group = tg; 7654 7655#ifdef CONFIG_FAIR_GROUP_SCHED 7656 if (tsk->sched_class->task_move_group) 7657 tsk->sched_class->task_move_group(tsk, on_rq); 7658 else 7659#endif 7660 set_task_rq(tsk, task_cpu(tsk)); 7661 7662 if (unlikely(running)) 7663 tsk->sched_class->set_curr_task(rq); 7664 if (on_rq) 7665 enqueue_task(rq, tsk, 0); 7666 7667 task_rq_unlock(rq, tsk, &flags); 7668} 7669#endif /* CONFIG_CGROUP_SCHED */ 7670 7671#if defined(CONFIG_RT_GROUP_SCHED) || defined(CONFIG_CFS_BANDWIDTH) 7672static unsigned long to_ratio(u64 period, u64 runtime) 7673{ 7674 if (runtime == RUNTIME_INF) 7675 return 1ULL << 20; 7676 7677 return div64_u64(runtime << 20, period); 7678} 7679#endif 7680 7681#ifdef CONFIG_RT_GROUP_SCHED 7682/* 7683 * Ensure that the real time constraints are schedulable. 7684 */ 7685static DEFINE_MUTEX(rt_constraints_mutex); 7686 7687/* Must be called with tasklist_lock held */ 7688static inline int tg_has_rt_tasks(struct task_group *tg) 7689{ 7690 struct task_struct *g, *p; 7691 7692 do_each_thread(g, p) { 7693 if (rt_task(p) && task_rq(p)->rt.tg == tg) 7694 return 1; 7695 } while_each_thread(g, p); 7696 7697 return 0; 7698} 7699 7700struct rt_schedulable_data { 7701 struct task_group *tg; 7702 u64 rt_period; 7703 u64 rt_runtime; 7704}; 7705 7706static int tg_rt_schedulable(struct task_group *tg, void *data) 7707{ 7708 struct rt_schedulable_data *d = data; 7709 struct task_group *child; 7710 unsigned long total, sum = 0; 7711 u64 period, runtime; 7712 7713 period = ktime_to_ns(tg->rt_bandwidth.rt_period); 7714 runtime = tg->rt_bandwidth.rt_runtime; 7715 7716 if (tg == d->tg) { 7717 period = d->rt_period; 7718 runtime = d->rt_runtime; 7719 } 7720 7721 /* 7722 * Cannot have more runtime than the period. 7723 */ 7724 if (runtime > period && runtime != RUNTIME_INF) 7725 return -EINVAL; 7726 7727 /* 7728 * Ensure we don't starve existing RT tasks. 7729 */ 7730 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg)) 7731 return -EBUSY; 7732 7733 total = to_ratio(period, runtime); 7734 7735 /* 7736 * Nobody can have more than the global setting allows. 7737 */ 7738 if (total > to_ratio(global_rt_period(), global_rt_runtime())) 7739 return -EINVAL; 7740 7741 /* 7742 * The sum of our children's runtime should not exceed our own. 7743 */ 7744 list_for_each_entry_rcu(child, &tg->children, siblings) { 7745 period = ktime_to_ns(child->rt_bandwidth.rt_period); 7746 runtime = child->rt_bandwidth.rt_runtime; 7747 7748 if (child == d->tg) { 7749 period = d->rt_period; 7750 runtime = d->rt_runtime; 7751 } 7752 7753 sum += to_ratio(period, runtime); 7754 } 7755 7756 if (sum > total) 7757 return -EINVAL; 7758 7759 return 0; 7760} 7761 7762static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime) 7763{ 7764 int ret; 7765 7766 struct rt_schedulable_data data = { 7767 .tg = tg, 7768 .rt_period = period, 7769 .rt_runtime = runtime, 7770 }; 7771 7772 rcu_read_lock(); 7773 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data); 7774 rcu_read_unlock(); 7775 7776 return ret; 7777} 7778 7779static int tg_set_rt_bandwidth(struct task_group *tg, 7780 u64 rt_period, u64 rt_runtime) 7781{ 7782 int i, err = 0; 7783 7784 mutex_lock(&rt_constraints_mutex); 7785 read_lock(&tasklist_lock); 7786 err = __rt_schedulable(tg, rt_period, rt_runtime); 7787 if (err) 7788 goto unlock; 7789 7790 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock); 7791 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period); 7792 tg->rt_bandwidth.rt_runtime = rt_runtime; 7793 7794 for_each_possible_cpu(i) { 7795 struct rt_rq *rt_rq = tg->rt_rq[i]; 7796 7797 raw_spin_lock(&rt_rq->rt_runtime_lock); 7798 rt_rq->rt_runtime = rt_runtime; 7799 raw_spin_unlock(&rt_rq->rt_runtime_lock); 7800 } 7801 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock); 7802unlock: 7803 read_unlock(&tasklist_lock); 7804 mutex_unlock(&rt_constraints_mutex); 7805 7806 return err; 7807} 7808 7809int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us) 7810{ 7811 u64 rt_runtime, rt_period; 7812 7813 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period); 7814 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC; 7815 if (rt_runtime_us < 0) 7816 rt_runtime = RUNTIME_INF; 7817 7818 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime); 7819} 7820 7821long sched_group_rt_runtime(struct task_group *tg) 7822{ 7823 u64 rt_runtime_us; 7824 7825 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF) 7826 return -1; 7827 7828 rt_runtime_us = tg->rt_bandwidth.rt_runtime; 7829 do_div(rt_runtime_us, NSEC_PER_USEC); 7830 return rt_runtime_us; 7831} 7832 7833int sched_group_set_rt_period(struct task_group *tg, long rt_period_us) 7834{ 7835 u64 rt_runtime, rt_period; 7836 7837 rt_period = (u64)rt_period_us * NSEC_PER_USEC; 7838 rt_runtime = tg->rt_bandwidth.rt_runtime; 7839 7840 if (rt_period == 0) 7841 return -EINVAL; 7842 7843 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime); 7844} 7845 7846long sched_group_rt_period(struct task_group *tg) 7847{ 7848 u64 rt_period_us; 7849 7850 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period); 7851 do_div(rt_period_us, NSEC_PER_USEC); 7852 return rt_period_us; 7853} 7854 7855static int sched_rt_global_constraints(void) 7856{ 7857 u64 runtime, period; 7858 int ret = 0; 7859 7860 if (sysctl_sched_rt_period <= 0) 7861 return -EINVAL; 7862 7863 runtime = global_rt_runtime(); 7864 period = global_rt_period(); 7865 7866 /* 7867 * Sanity check on the sysctl variables. 7868 */ 7869 if (runtime > period && runtime != RUNTIME_INF) 7870 return -EINVAL; 7871 7872 mutex_lock(&rt_constraints_mutex); 7873 read_lock(&tasklist_lock); 7874 ret = __rt_schedulable(NULL, 0, 0); 7875 read_unlock(&tasklist_lock); 7876 mutex_unlock(&rt_constraints_mutex); 7877 7878 return ret; 7879} 7880 7881int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk) 7882{ 7883 /* Don't accept realtime tasks when there is no way for them to run */ 7884 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0) 7885 return 0; 7886 7887 return 1; 7888} 7889 7890#else /* !CONFIG_RT_GROUP_SCHED */ 7891static int sched_rt_global_constraints(void) 7892{ 7893 unsigned long flags; 7894 int i; 7895 7896 if (sysctl_sched_rt_period <= 0) 7897 return -EINVAL; 7898 7899 /* 7900 * There's always some RT tasks in the root group 7901 * -- migration, kstopmachine etc.. 7902 */ 7903 if (sysctl_sched_rt_runtime == 0) 7904 return -EBUSY; 7905 7906 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags); 7907 for_each_possible_cpu(i) { 7908 struct rt_rq *rt_rq = &cpu_rq(i)->rt; 7909 7910 raw_spin_lock(&rt_rq->rt_runtime_lock); 7911 rt_rq->rt_runtime = global_rt_runtime(); 7912 raw_spin_unlock(&rt_rq->rt_runtime_lock); 7913 } 7914 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags); 7915 7916 return 0; 7917} 7918#endif /* CONFIG_RT_GROUP_SCHED */ 7919 7920int sched_rt_handler(struct ctl_table *table, int write, 7921 void __user *buffer, size_t *lenp, 7922 loff_t *ppos) 7923{ 7924 int ret; 7925 int old_period, old_runtime; 7926 static DEFINE_MUTEX(mutex); 7927 7928 mutex_lock(&mutex); 7929 old_period = sysctl_sched_rt_period; 7930 old_runtime = sysctl_sched_rt_runtime; 7931 7932 ret = proc_dointvec(table, write, buffer, lenp, ppos); 7933 7934 if (!ret && write) { 7935 ret = sched_rt_global_constraints(); 7936 if (ret) { 7937 sysctl_sched_rt_period = old_period; 7938 sysctl_sched_rt_runtime = old_runtime; 7939 } else { 7940 def_rt_bandwidth.rt_runtime = global_rt_runtime(); 7941 def_rt_bandwidth.rt_period = 7942 ns_to_ktime(global_rt_period()); 7943 } 7944 } 7945 mutex_unlock(&mutex); 7946 7947 return ret; 7948} 7949 7950#ifdef CONFIG_CGROUP_SCHED 7951 7952/* return corresponding task_group object of a cgroup */ 7953static inline struct task_group *cgroup_tg(struct cgroup *cgrp) 7954{ 7955 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id), 7956 struct task_group, css); 7957} 7958 7959static struct cgroup_subsys_state *cpu_cgroup_create(struct cgroup *cgrp) 7960{ 7961 struct task_group *tg, *parent; 7962 7963 if (!cgrp->parent) { 7964 /* This is early initialization for the top cgroup */ 7965 return &root_task_group.css; 7966 } 7967 7968 parent = cgroup_tg(cgrp->parent); 7969 tg = sched_create_group(parent); 7970 if (IS_ERR(tg)) 7971 return ERR_PTR(-ENOMEM); 7972 7973 return &tg->css; 7974} 7975 7976static void cpu_cgroup_destroy(struct cgroup *cgrp) 7977{ 7978 struct task_group *tg = cgroup_tg(cgrp); 7979 7980 sched_destroy_group(tg); 7981} 7982 7983static int cpu_cgroup_can_attach(struct cgroup *cgrp, 7984 struct cgroup_taskset *tset) 7985{ 7986 struct task_struct *task; 7987 7988 cgroup_taskset_for_each(task, cgrp, tset) { 7989#ifdef CONFIG_RT_GROUP_SCHED 7990 if (!sched_rt_can_attach(cgroup_tg(cgrp), task)) 7991 return -EINVAL; 7992#else 7993 /* We don't support RT-tasks being in separate groups */ 7994 if (task->sched_class != &fair_sched_class) 7995 return -EINVAL; 7996#endif 7997 } 7998 return 0; 7999} 8000 8001static void cpu_cgroup_attach(struct cgroup *cgrp, 8002 struct cgroup_taskset *tset) 8003{ 8004 struct task_struct *task; 8005 8006 cgroup_taskset_for_each(task, cgrp, tset) 8007 sched_move_task(task); 8008} 8009 8010static void 8011cpu_cgroup_exit(struct cgroup *cgrp, struct cgroup *old_cgrp, 8012 struct task_struct *task) 8013{ 8014 /* 8015 * cgroup_exit() is called in the copy_process() failure path. 8016 * Ignore this case since the task hasn't ran yet, this avoids 8017 * trying to poke a half freed task state from generic code. 8018 */ 8019 if (!(task->flags & PF_EXITING)) 8020 return; 8021 8022 sched_move_task(task); 8023} 8024 8025#ifdef CONFIG_FAIR_GROUP_SCHED 8026static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype, 8027 u64 shareval) 8028{ 8029 return sched_group_set_shares(cgroup_tg(cgrp), scale_load(shareval)); 8030} 8031 8032static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft) 8033{ 8034 struct task_group *tg = cgroup_tg(cgrp); 8035 8036 return (u64) scale_load_down(tg->shares); 8037} 8038 8039#ifdef CONFIG_CFS_BANDWIDTH 8040static DEFINE_MUTEX(cfs_constraints_mutex); 8041 8042const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */ 8043const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */ 8044 8045static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime); 8046 8047static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota) 8048{ 8049 int i, ret = 0, runtime_enabled, runtime_was_enabled; 8050 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth; 8051 8052 if (tg == &root_task_group) 8053 return -EINVAL; 8054 8055 /* 8056 * Ensure we have at some amount of bandwidth every period. This is 8057 * to prevent reaching a state of large arrears when throttled via 8058 * entity_tick() resulting in prolonged exit starvation. 8059 */ 8060 if (quota < min_cfs_quota_period || period < min_cfs_quota_period) 8061 return -EINVAL; 8062 8063 /* 8064 * Likewise, bound things on the otherside by preventing insane quota 8065 * periods. This also allows us to normalize in computing quota 8066 * feasibility. 8067 */ 8068 if (period > max_cfs_quota_period) 8069 return -EINVAL; 8070 8071 mutex_lock(&cfs_constraints_mutex); 8072 ret = __cfs_schedulable(tg, period, quota); 8073 if (ret) 8074 goto out_unlock; 8075 8076 runtime_enabled = quota != RUNTIME_INF; 8077 runtime_was_enabled = cfs_b->quota != RUNTIME_INF; 8078 account_cfs_bandwidth_used(runtime_enabled, runtime_was_enabled); 8079 raw_spin_lock_irq(&cfs_b->lock); 8080 cfs_b->period = ns_to_ktime(period); 8081 cfs_b->quota = quota; 8082 8083 __refill_cfs_bandwidth_runtime(cfs_b); 8084 /* restart the period timer (if active) to handle new period expiry */ 8085 if (runtime_enabled && cfs_b->timer_active) { 8086 /* force a reprogram */ 8087 cfs_b->timer_active = 0; 8088 __start_cfs_bandwidth(cfs_b); 8089 } 8090 raw_spin_unlock_irq(&cfs_b->lock); 8091 8092 for_each_possible_cpu(i) { 8093 struct cfs_rq *cfs_rq = tg->cfs_rq[i]; 8094 struct rq *rq = cfs_rq->rq; 8095 8096 raw_spin_lock_irq(&rq->lock); 8097 cfs_rq->runtime_enabled = runtime_enabled; 8098 cfs_rq->runtime_remaining = 0; 8099 8100 if (cfs_rq->throttled) 8101 unthrottle_cfs_rq(cfs_rq); 8102 raw_spin_unlock_irq(&rq->lock); 8103 } 8104out_unlock: 8105 mutex_unlock(&cfs_constraints_mutex); 8106 8107 return ret; 8108} 8109 8110int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us) 8111{ 8112 u64 quota, period; 8113 8114 period = ktime_to_ns(tg->cfs_bandwidth.period); 8115 if (cfs_quota_us < 0) 8116 quota = RUNTIME_INF; 8117 else 8118 quota = (u64)cfs_quota_us * NSEC_PER_USEC; 8119 8120 return tg_set_cfs_bandwidth(tg, period, quota); 8121} 8122 8123long tg_get_cfs_quota(struct task_group *tg) 8124{ 8125 u64 quota_us; 8126 8127 if (tg->cfs_bandwidth.quota == RUNTIME_INF) 8128 return -1; 8129 8130 quota_us = tg->cfs_bandwidth.quota; 8131 do_div(quota_us, NSEC_PER_USEC); 8132 8133 return quota_us; 8134} 8135 8136int tg_set_cfs_period(struct task_group *tg, long cfs_period_us) 8137{ 8138 u64 quota, period; 8139 8140 period = (u64)cfs_period_us * NSEC_PER_USEC; 8141 quota = tg->cfs_bandwidth.quota; 8142 8143 return tg_set_cfs_bandwidth(tg, period, quota); 8144} 8145 8146long tg_get_cfs_period(struct task_group *tg) 8147{ 8148 u64 cfs_period_us; 8149 8150 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period); 8151 do_div(cfs_period_us, NSEC_PER_USEC); 8152 8153 return cfs_period_us; 8154} 8155 8156static s64 cpu_cfs_quota_read_s64(struct cgroup *cgrp, struct cftype *cft) 8157{ 8158 return tg_get_cfs_quota(cgroup_tg(cgrp)); 8159} 8160 8161static int cpu_cfs_quota_write_s64(struct cgroup *cgrp, struct cftype *cftype, 8162 s64 cfs_quota_us) 8163{ 8164 return tg_set_cfs_quota(cgroup_tg(cgrp), cfs_quota_us); 8165} 8166 8167static u64 cpu_cfs_period_read_u64(struct cgroup *cgrp, struct cftype *cft) 8168{ 8169 return tg_get_cfs_period(cgroup_tg(cgrp)); 8170} 8171 8172static int cpu_cfs_period_write_u64(struct cgroup *cgrp, struct cftype *cftype, 8173 u64 cfs_period_us) 8174{ 8175 return tg_set_cfs_period(cgroup_tg(cgrp), cfs_period_us); 8176} 8177 8178struct cfs_schedulable_data { 8179 struct task_group *tg; 8180 u64 period, quota; 8181}; 8182 8183/* 8184 * normalize group quota/period to be quota/max_period 8185 * note: units are usecs 8186 */ 8187static u64 normalize_cfs_quota(struct task_group *tg, 8188 struct cfs_schedulable_data *d) 8189{ 8190 u64 quota, period; 8191 8192 if (tg == d->tg) { 8193 period = d->period; 8194 quota = d->quota; 8195 } else { 8196 period = tg_get_cfs_period(tg); 8197 quota = tg_get_cfs_quota(tg); 8198 } 8199 8200 /* note: these should typically be equivalent */ 8201 if (quota == RUNTIME_INF || quota == -1) 8202 return RUNTIME_INF; 8203 8204 return to_ratio(period, quota); 8205} 8206 8207static int tg_cfs_schedulable_down(struct task_group *tg, void *data) 8208{ 8209 struct cfs_schedulable_data *d = data; 8210 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth; 8211 s64 quota = 0, parent_quota = -1; 8212 8213 if (!tg->parent) { 8214 quota = RUNTIME_INF; 8215 } else { 8216 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth; 8217 8218 quota = normalize_cfs_quota(tg, d); 8219 parent_quota = parent_b->hierarchal_quota; 8220 8221 /* 8222 * ensure max(child_quota) <= parent_quota, inherit when no 8223 * limit is set 8224 */ 8225 if (quota == RUNTIME_INF) 8226 quota = parent_quota; 8227 else if (parent_quota != RUNTIME_INF && quota > parent_quota) 8228 return -EINVAL; 8229 } 8230 cfs_b->hierarchal_quota = quota; 8231 8232 return 0; 8233} 8234 8235static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota) 8236{ 8237 int ret; 8238 struct cfs_schedulable_data data = { 8239 .tg = tg, 8240 .period = period, 8241 .quota = quota, 8242 }; 8243 8244 if (quota != RUNTIME_INF) { 8245 do_div(data.period, NSEC_PER_USEC); 8246 do_div(data.quota, NSEC_PER_USEC); 8247 } 8248 8249 rcu_read_lock(); 8250 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data); 8251 rcu_read_unlock(); 8252 8253 return ret; 8254} 8255 8256static int cpu_stats_show(struct cgroup *cgrp, struct cftype *cft, 8257 struct cgroup_map_cb *cb) 8258{ 8259 struct task_group *tg = cgroup_tg(cgrp); 8260 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth; 8261 8262 cb->fill(cb, "nr_periods", cfs_b->nr_periods); 8263 cb->fill(cb, "nr_throttled", cfs_b->nr_throttled); 8264 cb->fill(cb, "throttled_time", cfs_b->throttled_time); 8265 8266 return 0; 8267} 8268#endif /* CONFIG_CFS_BANDWIDTH */ 8269#endif /* CONFIG_FAIR_GROUP_SCHED */ 8270 8271#ifdef CONFIG_RT_GROUP_SCHED 8272static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft, 8273 s64 val) 8274{ 8275 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val); 8276} 8277 8278static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft) 8279{ 8280 return sched_group_rt_runtime(cgroup_tg(cgrp)); 8281} 8282 8283static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype, 8284 u64 rt_period_us) 8285{ 8286 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us); 8287} 8288 8289static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft) 8290{ 8291 return sched_group_rt_period(cgroup_tg(cgrp)); 8292} 8293#endif /* CONFIG_RT_GROUP_SCHED */ 8294 8295static struct cftype cpu_files[] = { 8296#ifdef CONFIG_FAIR_GROUP_SCHED 8297 { 8298 .name = "shares", 8299 .read_u64 = cpu_shares_read_u64, 8300 .write_u64 = cpu_shares_write_u64, 8301 }, 8302#endif 8303#ifdef CONFIG_CFS_BANDWIDTH 8304 { 8305 .name = "cfs_quota_us", 8306 .read_s64 = cpu_cfs_quota_read_s64, 8307 .write_s64 = cpu_cfs_quota_write_s64, 8308 }, 8309 { 8310 .name = "cfs_period_us", 8311 .read_u64 = cpu_cfs_period_read_u64, 8312 .write_u64 = cpu_cfs_period_write_u64, 8313 }, 8314 { 8315 .name = "stat", 8316 .read_map = cpu_stats_show, 8317 }, 8318#endif 8319#ifdef CONFIG_RT_GROUP_SCHED 8320 { 8321 .name = "rt_runtime_us", 8322 .read_s64 = cpu_rt_runtime_read, 8323 .write_s64 = cpu_rt_runtime_write, 8324 }, 8325 { 8326 .name = "rt_period_us", 8327 .read_u64 = cpu_rt_period_read_uint, 8328 .write_u64 = cpu_rt_period_write_uint, 8329 }, 8330#endif 8331 { } /* terminate */ 8332}; 8333 8334struct cgroup_subsys cpu_cgroup_subsys = { 8335 .name = "cpu", 8336 .create = cpu_cgroup_create, 8337 .destroy = cpu_cgroup_destroy, 8338 .can_attach = cpu_cgroup_can_attach, 8339 .attach = cpu_cgroup_attach, 8340 .exit = cpu_cgroup_exit, 8341 .subsys_id = cpu_cgroup_subsys_id, 8342 .base_cftypes = cpu_files, 8343 .early_init = 1, 8344}; 8345 8346#endif /* CONFIG_CGROUP_SCHED */ 8347 8348#ifdef CONFIG_CGROUP_CPUACCT 8349 8350/* 8351 * CPU accounting code for task groups. 8352 * 8353 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh 8354 * (balbir@in.ibm.com). 8355 */ 8356 8357/* create a new cpu accounting group */ 8358static struct cgroup_subsys_state *cpuacct_create(struct cgroup *cgrp) 8359{ 8360 struct cpuacct *ca; 8361 8362 if (!cgrp->parent) 8363 return &root_cpuacct.css; 8364 8365 ca = kzalloc(sizeof(*ca), GFP_KERNEL); 8366 if (!ca) 8367 goto out; 8368 8369 ca->cpuusage = alloc_percpu(u64); 8370 if (!ca->cpuusage) 8371 goto out_free_ca; 8372 8373 ca->cpustat = alloc_percpu(struct kernel_cpustat); 8374 if (!ca->cpustat) 8375 goto out_free_cpuusage; 8376 8377 return &ca->css; 8378 8379out_free_cpuusage: 8380 free_percpu(ca->cpuusage); 8381out_free_ca: 8382 kfree(ca); 8383out: 8384 return ERR_PTR(-ENOMEM); 8385} 8386 8387/* destroy an existing cpu accounting group */ 8388static void cpuacct_destroy(struct cgroup *cgrp) 8389{ 8390 struct cpuacct *ca = cgroup_ca(cgrp); 8391 8392 free_percpu(ca->cpustat); 8393 free_percpu(ca->cpuusage); 8394 kfree(ca); 8395} 8396 8397static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu) 8398{ 8399 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu); 8400 u64 data; 8401 8402#ifndef CONFIG_64BIT 8403 /* 8404 * Take rq->lock to make 64-bit read safe on 32-bit platforms. 8405 */ 8406 raw_spin_lock_irq(&cpu_rq(cpu)->lock); 8407 data = *cpuusage; 8408 raw_spin_unlock_irq(&cpu_rq(cpu)->lock); 8409#else 8410 data = *cpuusage; 8411#endif 8412 8413 return data; 8414} 8415 8416static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val) 8417{ 8418 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu); 8419 8420#ifndef CONFIG_64BIT 8421 /* 8422 * Take rq->lock to make 64-bit write safe on 32-bit platforms. 8423 */ 8424 raw_spin_lock_irq(&cpu_rq(cpu)->lock); 8425 *cpuusage = val; 8426 raw_spin_unlock_irq(&cpu_rq(cpu)->lock); 8427#else 8428 *cpuusage = val; 8429#endif 8430} 8431 8432/* return total cpu usage (in nanoseconds) of a group */ 8433static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft) 8434{ 8435 struct cpuacct *ca = cgroup_ca(cgrp); 8436 u64 totalcpuusage = 0; 8437 int i; 8438 8439 for_each_present_cpu(i) 8440 totalcpuusage += cpuacct_cpuusage_read(ca, i); 8441 8442 return totalcpuusage; 8443} 8444 8445static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype, 8446 u64 reset) 8447{ 8448 struct cpuacct *ca = cgroup_ca(cgrp); 8449 int err = 0; 8450 int i; 8451 8452 if (reset) { 8453 err = -EINVAL; 8454 goto out; 8455 } 8456 8457 for_each_present_cpu(i) 8458 cpuacct_cpuusage_write(ca, i, 0); 8459 8460out: 8461 return err; 8462} 8463 8464static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft, 8465 struct seq_file *m) 8466{ 8467 struct cpuacct *ca = cgroup_ca(cgroup); 8468 u64 percpu; 8469 int i; 8470 8471 for_each_present_cpu(i) { 8472 percpu = cpuacct_cpuusage_read(ca, i); 8473 seq_printf(m, "%llu ", (unsigned long long) percpu); 8474 } 8475 seq_printf(m, "\n"); 8476 return 0; 8477} 8478 8479static const char *cpuacct_stat_desc[] = { 8480 [CPUACCT_STAT_USER] = "user", 8481 [CPUACCT_STAT_SYSTEM] = "system", 8482}; 8483 8484static int cpuacct_stats_show(struct cgroup *cgrp, struct cftype *cft, 8485 struct cgroup_map_cb *cb) 8486{ 8487 struct cpuacct *ca = cgroup_ca(cgrp); 8488 int cpu; 8489 s64 val = 0; 8490 8491 for_each_online_cpu(cpu) { 8492 struct kernel_cpustat *kcpustat = per_cpu_ptr(ca->cpustat, cpu); 8493 val += kcpustat->cpustat[CPUTIME_USER]; 8494 val += kcpustat->cpustat[CPUTIME_NICE]; 8495 } 8496 val = cputime64_to_clock_t(val); 8497 cb->fill(cb, cpuacct_stat_desc[CPUACCT_STAT_USER], val); 8498 8499 val = 0; 8500 for_each_online_cpu(cpu) { 8501 struct kernel_cpustat *kcpustat = per_cpu_ptr(ca->cpustat, cpu); 8502 val += kcpustat->cpustat[CPUTIME_SYSTEM]; 8503 val += kcpustat->cpustat[CPUTIME_IRQ]; 8504 val += kcpustat->cpustat[CPUTIME_SOFTIRQ]; 8505 } 8506 8507 val = cputime64_to_clock_t(val); 8508 cb->fill(cb, cpuacct_stat_desc[CPUACCT_STAT_SYSTEM], val); 8509 8510 return 0; 8511} 8512 8513static struct cftype files[] = { 8514 { 8515 .name = "usage", 8516 .read_u64 = cpuusage_read, 8517 .write_u64 = cpuusage_write, 8518 }, 8519 { 8520 .name = "usage_percpu", 8521 .read_seq_string = cpuacct_percpu_seq_read, 8522 }, 8523 { 8524 .name = "stat", 8525 .read_map = cpuacct_stats_show, 8526 }, 8527 { } /* terminate */ 8528}; 8529 8530/* 8531 * charge this task's execution time to its accounting group. 8532 * 8533 * called with rq->lock held. 8534 */ 8535void cpuacct_charge(struct task_struct *tsk, u64 cputime) 8536{ 8537 struct cpuacct *ca; 8538 int cpu; 8539 8540 if (unlikely(!cpuacct_subsys.active)) 8541 return; 8542 8543 cpu = task_cpu(tsk); 8544 8545 rcu_read_lock(); 8546 8547 ca = task_ca(tsk); 8548 8549 for (; ca; ca = parent_ca(ca)) { 8550 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu); 8551 *cpuusage += cputime; 8552 } 8553 8554 rcu_read_unlock(); 8555} 8556 8557struct cgroup_subsys cpuacct_subsys = { 8558 .name = "cpuacct", 8559 .create = cpuacct_create, 8560 .destroy = cpuacct_destroy, 8561 .subsys_id = cpuacct_subsys_id, 8562 .base_cftypes = files, 8563}; 8564#endif /* CONFIG_CGROUP_CPUACCT */ 8565