core.c revision 9ee474f55664ff63111c843099d365e7ecffb56f
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) 0 509#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 743static void update_rq_clock_task(struct rq *rq, s64 delta) 744{ 745/* 746 * In theory, the compile should just see 0 here, and optimize out the call 747 * to sched_rt_avg_update. But I don't trust it... 748 */ 749#if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING) 750 s64 steal = 0, irq_delta = 0; 751#endif 752#ifdef CONFIG_IRQ_TIME_ACCOUNTING 753 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time; 754 755 /* 756 * Since irq_time is only updated on {soft,}irq_exit, we might run into 757 * this case when a previous update_rq_clock() happened inside a 758 * {soft,}irq region. 759 * 760 * When this happens, we stop ->clock_task and only update the 761 * prev_irq_time stamp to account for the part that fit, so that a next 762 * update will consume the rest. This ensures ->clock_task is 763 * monotonic. 764 * 765 * It does however cause some slight miss-attribution of {soft,}irq 766 * time, a more accurate solution would be to update the irq_time using 767 * the current rq->clock timestamp, except that would require using 768 * atomic ops. 769 */ 770 if (irq_delta > delta) 771 irq_delta = delta; 772 773 rq->prev_irq_time += irq_delta; 774 delta -= irq_delta; 775#endif 776#ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING 777 if (static_key_false((¶virt_steal_rq_enabled))) { 778 u64 st; 779 780 steal = paravirt_steal_clock(cpu_of(rq)); 781 steal -= rq->prev_steal_time_rq; 782 783 if (unlikely(steal > delta)) 784 steal = delta; 785 786 st = steal_ticks(steal); 787 steal = st * TICK_NSEC; 788 789 rq->prev_steal_time_rq += steal; 790 791 delta -= steal; 792 } 793#endif 794 795 rq->clock_task += delta; 796 797#if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING) 798 if ((irq_delta + steal) && sched_feat(NONTASK_POWER)) 799 sched_rt_avg_update(rq, irq_delta + steal); 800#endif 801} 802 803void sched_set_stop_task(int cpu, struct task_struct *stop) 804{ 805 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 }; 806 struct task_struct *old_stop = cpu_rq(cpu)->stop; 807 808 if (stop) { 809 /* 810 * Make it appear like a SCHED_FIFO task, its something 811 * userspace knows about and won't get confused about. 812 * 813 * Also, it will make PI more or less work without too 814 * much confusion -- but then, stop work should not 815 * rely on PI working anyway. 816 */ 817 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m); 818 819 stop->sched_class = &stop_sched_class; 820 } 821 822 cpu_rq(cpu)->stop = stop; 823 824 if (old_stop) { 825 /* 826 * Reset it back to a normal scheduling class so that 827 * it can die in pieces. 828 */ 829 old_stop->sched_class = &rt_sched_class; 830 } 831} 832 833/* 834 * __normal_prio - return the priority that is based on the static prio 835 */ 836static inline int __normal_prio(struct task_struct *p) 837{ 838 return p->static_prio; 839} 840 841/* 842 * Calculate the expected normal priority: i.e. priority 843 * without taking RT-inheritance into account. Might be 844 * boosted by interactivity modifiers. Changes upon fork, 845 * setprio syscalls, and whenever the interactivity 846 * estimator recalculates. 847 */ 848static inline int normal_prio(struct task_struct *p) 849{ 850 int prio; 851 852 if (task_has_rt_policy(p)) 853 prio = MAX_RT_PRIO-1 - p->rt_priority; 854 else 855 prio = __normal_prio(p); 856 return prio; 857} 858 859/* 860 * Calculate the current priority, i.e. the priority 861 * taken into account by the scheduler. This value might 862 * be boosted by RT tasks, or might be boosted by 863 * interactivity modifiers. Will be RT if the task got 864 * RT-boosted. If not then it returns p->normal_prio. 865 */ 866static int effective_prio(struct task_struct *p) 867{ 868 p->normal_prio = normal_prio(p); 869 /* 870 * If we are RT tasks or we were boosted to RT priority, 871 * keep the priority unchanged. Otherwise, update priority 872 * to the normal priority: 873 */ 874 if (!rt_prio(p->prio)) 875 return p->normal_prio; 876 return p->prio; 877} 878 879/** 880 * task_curr - is this task currently executing on a CPU? 881 * @p: the task in question. 882 */ 883inline int task_curr(const struct task_struct *p) 884{ 885 return cpu_curr(task_cpu(p)) == p; 886} 887 888static inline void check_class_changed(struct rq *rq, struct task_struct *p, 889 const struct sched_class *prev_class, 890 int oldprio) 891{ 892 if (prev_class != p->sched_class) { 893 if (prev_class->switched_from) 894 prev_class->switched_from(rq, p); 895 p->sched_class->switched_to(rq, p); 896 } else if (oldprio != p->prio) 897 p->sched_class->prio_changed(rq, p, oldprio); 898} 899 900void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags) 901{ 902 const struct sched_class *class; 903 904 if (p->sched_class == rq->curr->sched_class) { 905 rq->curr->sched_class->check_preempt_curr(rq, p, flags); 906 } else { 907 for_each_class(class) { 908 if (class == rq->curr->sched_class) 909 break; 910 if (class == p->sched_class) { 911 resched_task(rq->curr); 912 break; 913 } 914 } 915 } 916 917 /* 918 * A queue event has occurred, and we're going to schedule. In 919 * this case, we can save a useless back to back clock update. 920 */ 921 if (rq->curr->on_rq && test_tsk_need_resched(rq->curr)) 922 rq->skip_clock_update = 1; 923} 924 925#ifdef CONFIG_SMP 926void set_task_cpu(struct task_struct *p, unsigned int new_cpu) 927{ 928#ifdef CONFIG_SCHED_DEBUG 929 /* 930 * We should never call set_task_cpu() on a blocked task, 931 * ttwu() will sort out the placement. 932 */ 933 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING && 934 !(task_thread_info(p)->preempt_count & PREEMPT_ACTIVE)); 935 936#ifdef CONFIG_LOCKDEP 937 /* 938 * The caller should hold either p->pi_lock or rq->lock, when changing 939 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks. 940 * 941 * sched_move_task() holds both and thus holding either pins the cgroup, 942 * see task_group(). 943 * 944 * Furthermore, all task_rq users should acquire both locks, see 945 * task_rq_lock(). 946 */ 947 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) || 948 lockdep_is_held(&task_rq(p)->lock))); 949#endif 950#endif 951 952 trace_sched_migrate_task(p, new_cpu); 953 954 if (task_cpu(p) != new_cpu) { 955 p->se.nr_migrations++; 956 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, NULL, 0); 957 } 958 959 __set_task_cpu(p, new_cpu); 960} 961 962struct migration_arg { 963 struct task_struct *task; 964 int dest_cpu; 965}; 966 967static int migration_cpu_stop(void *data); 968 969/* 970 * wait_task_inactive - wait for a thread to unschedule. 971 * 972 * If @match_state is nonzero, it's the @p->state value just checked and 973 * not expected to change. If it changes, i.e. @p might have woken up, 974 * then return zero. When we succeed in waiting for @p to be off its CPU, 975 * we return a positive number (its total switch count). If a second call 976 * a short while later returns the same number, the caller can be sure that 977 * @p has remained unscheduled the whole time. 978 * 979 * The caller must ensure that the task *will* unschedule sometime soon, 980 * else this function might spin for a *long* time. This function can't 981 * be called with interrupts off, or it may introduce deadlock with 982 * smp_call_function() if an IPI is sent by the same process we are 983 * waiting to become inactive. 984 */ 985unsigned long wait_task_inactive(struct task_struct *p, long match_state) 986{ 987 unsigned long flags; 988 int running, on_rq; 989 unsigned long ncsw; 990 struct rq *rq; 991 992 for (;;) { 993 /* 994 * We do the initial early heuristics without holding 995 * any task-queue locks at all. We'll only try to get 996 * the runqueue lock when things look like they will 997 * work out! 998 */ 999 rq = task_rq(p); 1000 1001 /* 1002 * If the task is actively running on another CPU 1003 * still, just relax and busy-wait without holding 1004 * any locks. 1005 * 1006 * NOTE! Since we don't hold any locks, it's not 1007 * even sure that "rq" stays as the right runqueue! 1008 * But we don't care, since "task_running()" will 1009 * return false if the runqueue has changed and p 1010 * is actually now running somewhere else! 1011 */ 1012 while (task_running(rq, p)) { 1013 if (match_state && unlikely(p->state != match_state)) 1014 return 0; 1015 cpu_relax(); 1016 } 1017 1018 /* 1019 * Ok, time to look more closely! We need the rq 1020 * lock now, to be *sure*. If we're wrong, we'll 1021 * just go back and repeat. 1022 */ 1023 rq = task_rq_lock(p, &flags); 1024 trace_sched_wait_task(p); 1025 running = task_running(rq, p); 1026 on_rq = p->on_rq; 1027 ncsw = 0; 1028 if (!match_state || p->state == match_state) 1029 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */ 1030 task_rq_unlock(rq, p, &flags); 1031 1032 /* 1033 * If it changed from the expected state, bail out now. 1034 */ 1035 if (unlikely(!ncsw)) 1036 break; 1037 1038 /* 1039 * Was it really running after all now that we 1040 * checked with the proper locks actually held? 1041 * 1042 * Oops. Go back and try again.. 1043 */ 1044 if (unlikely(running)) { 1045 cpu_relax(); 1046 continue; 1047 } 1048 1049 /* 1050 * It's not enough that it's not actively running, 1051 * it must be off the runqueue _entirely_, and not 1052 * preempted! 1053 * 1054 * So if it was still runnable (but just not actively 1055 * running right now), it's preempted, and we should 1056 * yield - it could be a while. 1057 */ 1058 if (unlikely(on_rq)) { 1059 ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ); 1060 1061 set_current_state(TASK_UNINTERRUPTIBLE); 1062 schedule_hrtimeout(&to, HRTIMER_MODE_REL); 1063 continue; 1064 } 1065 1066 /* 1067 * Ahh, all good. It wasn't running, and it wasn't 1068 * runnable, which means that it will never become 1069 * running in the future either. We're all done! 1070 */ 1071 break; 1072 } 1073 1074 return ncsw; 1075} 1076 1077/*** 1078 * kick_process - kick a running thread to enter/exit the kernel 1079 * @p: the to-be-kicked thread 1080 * 1081 * Cause a process which is running on another CPU to enter 1082 * kernel-mode, without any delay. (to get signals handled.) 1083 * 1084 * NOTE: this function doesn't have to take the runqueue lock, 1085 * because all it wants to ensure is that the remote task enters 1086 * the kernel. If the IPI races and the task has been migrated 1087 * to another CPU then no harm is done and the purpose has been 1088 * achieved as well. 1089 */ 1090void kick_process(struct task_struct *p) 1091{ 1092 int cpu; 1093 1094 preempt_disable(); 1095 cpu = task_cpu(p); 1096 if ((cpu != smp_processor_id()) && task_curr(p)) 1097 smp_send_reschedule(cpu); 1098 preempt_enable(); 1099} 1100EXPORT_SYMBOL_GPL(kick_process); 1101#endif /* CONFIG_SMP */ 1102 1103#ifdef CONFIG_SMP 1104/* 1105 * ->cpus_allowed is protected by both rq->lock and p->pi_lock 1106 */ 1107static int select_fallback_rq(int cpu, struct task_struct *p) 1108{ 1109 const struct cpumask *nodemask = cpumask_of_node(cpu_to_node(cpu)); 1110 enum { cpuset, possible, fail } state = cpuset; 1111 int dest_cpu; 1112 1113 /* Look for allowed, online CPU in same node. */ 1114 for_each_cpu(dest_cpu, nodemask) { 1115 if (!cpu_online(dest_cpu)) 1116 continue; 1117 if (!cpu_active(dest_cpu)) 1118 continue; 1119 if (cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p))) 1120 return dest_cpu; 1121 } 1122 1123 for (;;) { 1124 /* Any allowed, online CPU? */ 1125 for_each_cpu(dest_cpu, tsk_cpus_allowed(p)) { 1126 if (!cpu_online(dest_cpu)) 1127 continue; 1128 if (!cpu_active(dest_cpu)) 1129 continue; 1130 goto out; 1131 } 1132 1133 switch (state) { 1134 case cpuset: 1135 /* No more Mr. Nice Guy. */ 1136 cpuset_cpus_allowed_fallback(p); 1137 state = possible; 1138 break; 1139 1140 case possible: 1141 do_set_cpus_allowed(p, cpu_possible_mask); 1142 state = fail; 1143 break; 1144 1145 case fail: 1146 BUG(); 1147 break; 1148 } 1149 } 1150 1151out: 1152 if (state != cpuset) { 1153 /* 1154 * Don't tell them about moving exiting tasks or 1155 * kernel threads (both mm NULL), since they never 1156 * leave kernel. 1157 */ 1158 if (p->mm && printk_ratelimit()) { 1159 printk_sched("process %d (%s) no longer affine to cpu%d\n", 1160 task_pid_nr(p), p->comm, cpu); 1161 } 1162 } 1163 1164 return dest_cpu; 1165} 1166 1167/* 1168 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable. 1169 */ 1170static inline 1171int select_task_rq(struct task_struct *p, int sd_flags, int wake_flags) 1172{ 1173 int cpu = p->sched_class->select_task_rq(p, sd_flags, wake_flags); 1174 1175 /* 1176 * In order not to call set_task_cpu() on a blocking task we need 1177 * to rely on ttwu() to place the task on a valid ->cpus_allowed 1178 * cpu. 1179 * 1180 * Since this is common to all placement strategies, this lives here. 1181 * 1182 * [ this allows ->select_task() to simply return task_cpu(p) and 1183 * not worry about this generic constraint ] 1184 */ 1185 if (unlikely(!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)) || 1186 !cpu_online(cpu))) 1187 cpu = select_fallback_rq(task_cpu(p), p); 1188 1189 return cpu; 1190} 1191 1192static void update_avg(u64 *avg, u64 sample) 1193{ 1194 s64 diff = sample - *avg; 1195 *avg += diff >> 3; 1196} 1197#endif 1198 1199static void 1200ttwu_stat(struct task_struct *p, int cpu, int wake_flags) 1201{ 1202#ifdef CONFIG_SCHEDSTATS 1203 struct rq *rq = this_rq(); 1204 1205#ifdef CONFIG_SMP 1206 int this_cpu = smp_processor_id(); 1207 1208 if (cpu == this_cpu) { 1209 schedstat_inc(rq, ttwu_local); 1210 schedstat_inc(p, se.statistics.nr_wakeups_local); 1211 } else { 1212 struct sched_domain *sd; 1213 1214 schedstat_inc(p, se.statistics.nr_wakeups_remote); 1215 rcu_read_lock(); 1216 for_each_domain(this_cpu, sd) { 1217 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) { 1218 schedstat_inc(sd, ttwu_wake_remote); 1219 break; 1220 } 1221 } 1222 rcu_read_unlock(); 1223 } 1224 1225 if (wake_flags & WF_MIGRATED) 1226 schedstat_inc(p, se.statistics.nr_wakeups_migrate); 1227 1228#endif /* CONFIG_SMP */ 1229 1230 schedstat_inc(rq, ttwu_count); 1231 schedstat_inc(p, se.statistics.nr_wakeups); 1232 1233 if (wake_flags & WF_SYNC) 1234 schedstat_inc(p, se.statistics.nr_wakeups_sync); 1235 1236#endif /* CONFIG_SCHEDSTATS */ 1237} 1238 1239static void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags) 1240{ 1241 activate_task(rq, p, en_flags); 1242 p->on_rq = 1; 1243 1244 /* if a worker is waking up, notify workqueue */ 1245 if (p->flags & PF_WQ_WORKER) 1246 wq_worker_waking_up(p, cpu_of(rq)); 1247} 1248 1249/* 1250 * Mark the task runnable and perform wakeup-preemption. 1251 */ 1252static void 1253ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags) 1254{ 1255 trace_sched_wakeup(p, true); 1256 check_preempt_curr(rq, p, wake_flags); 1257 1258 p->state = TASK_RUNNING; 1259#ifdef CONFIG_SMP 1260 if (p->sched_class->task_woken) 1261 p->sched_class->task_woken(rq, p); 1262 1263 if (rq->idle_stamp) { 1264 u64 delta = rq->clock - rq->idle_stamp; 1265 u64 max = 2*sysctl_sched_migration_cost; 1266 1267 if (delta > max) 1268 rq->avg_idle = max; 1269 else 1270 update_avg(&rq->avg_idle, delta); 1271 rq->idle_stamp = 0; 1272 } 1273#endif 1274} 1275 1276static void 1277ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags) 1278{ 1279#ifdef CONFIG_SMP 1280 if (p->sched_contributes_to_load) 1281 rq->nr_uninterruptible--; 1282#endif 1283 1284 ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_WAKING); 1285 ttwu_do_wakeup(rq, p, wake_flags); 1286} 1287 1288/* 1289 * Called in case the task @p isn't fully descheduled from its runqueue, 1290 * in this case we must do a remote wakeup. Its a 'light' wakeup though, 1291 * since all we need to do is flip p->state to TASK_RUNNING, since 1292 * the task is still ->on_rq. 1293 */ 1294static int ttwu_remote(struct task_struct *p, int wake_flags) 1295{ 1296 struct rq *rq; 1297 int ret = 0; 1298 1299 rq = __task_rq_lock(p); 1300 if (p->on_rq) { 1301 ttwu_do_wakeup(rq, p, wake_flags); 1302 ret = 1; 1303 } 1304 __task_rq_unlock(rq); 1305 1306 return ret; 1307} 1308 1309#ifdef CONFIG_SMP 1310static void sched_ttwu_pending(void) 1311{ 1312 struct rq *rq = this_rq(); 1313 struct llist_node *llist = llist_del_all(&rq->wake_list); 1314 struct task_struct *p; 1315 1316 raw_spin_lock(&rq->lock); 1317 1318 while (llist) { 1319 p = llist_entry(llist, struct task_struct, wake_entry); 1320 llist = llist_next(llist); 1321 ttwu_do_activate(rq, p, 0); 1322 } 1323 1324 raw_spin_unlock(&rq->lock); 1325} 1326 1327void scheduler_ipi(void) 1328{ 1329 if (llist_empty(&this_rq()->wake_list) && !got_nohz_idle_kick()) 1330 return; 1331 1332 /* 1333 * Not all reschedule IPI handlers call irq_enter/irq_exit, since 1334 * traditionally all their work was done from the interrupt return 1335 * path. Now that we actually do some work, we need to make sure 1336 * we do call them. 1337 * 1338 * Some archs already do call them, luckily irq_enter/exit nest 1339 * properly. 1340 * 1341 * Arguably we should visit all archs and update all handlers, 1342 * however a fair share of IPIs are still resched only so this would 1343 * somewhat pessimize the simple resched case. 1344 */ 1345 irq_enter(); 1346 sched_ttwu_pending(); 1347 1348 /* 1349 * Check if someone kicked us for doing the nohz idle load balance. 1350 */ 1351 if (unlikely(got_nohz_idle_kick() && !need_resched())) { 1352 this_rq()->idle_balance = 1; 1353 raise_softirq_irqoff(SCHED_SOFTIRQ); 1354 } 1355 irq_exit(); 1356} 1357 1358static void ttwu_queue_remote(struct task_struct *p, int cpu) 1359{ 1360 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list)) 1361 smp_send_reschedule(cpu); 1362} 1363 1364bool cpus_share_cache(int this_cpu, int that_cpu) 1365{ 1366 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu); 1367} 1368#endif /* CONFIG_SMP */ 1369 1370static void ttwu_queue(struct task_struct *p, int cpu) 1371{ 1372 struct rq *rq = cpu_rq(cpu); 1373 1374#if defined(CONFIG_SMP) 1375 if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) { 1376 sched_clock_cpu(cpu); /* sync clocks x-cpu */ 1377 ttwu_queue_remote(p, cpu); 1378 return; 1379 } 1380#endif 1381 1382 raw_spin_lock(&rq->lock); 1383 ttwu_do_activate(rq, p, 0); 1384 raw_spin_unlock(&rq->lock); 1385} 1386 1387/** 1388 * try_to_wake_up - wake up a thread 1389 * @p: the thread to be awakened 1390 * @state: the mask of task states that can be woken 1391 * @wake_flags: wake modifier flags (WF_*) 1392 * 1393 * Put it on the run-queue if it's not already there. The "current" 1394 * thread is always on the run-queue (except when the actual 1395 * re-schedule is in progress), and as such you're allowed to do 1396 * the simpler "current->state = TASK_RUNNING" to mark yourself 1397 * runnable without the overhead of this. 1398 * 1399 * Returns %true if @p was woken up, %false if it was already running 1400 * or @state didn't match @p's state. 1401 */ 1402static int 1403try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags) 1404{ 1405 unsigned long flags; 1406 int cpu, success = 0; 1407 1408 smp_wmb(); 1409 raw_spin_lock_irqsave(&p->pi_lock, flags); 1410 if (!(p->state & state)) 1411 goto out; 1412 1413 success = 1; /* we're going to change ->state */ 1414 cpu = task_cpu(p); 1415 1416 if (p->on_rq && ttwu_remote(p, wake_flags)) 1417 goto stat; 1418 1419#ifdef CONFIG_SMP 1420 /* 1421 * If the owning (remote) cpu is still in the middle of schedule() with 1422 * this task as prev, wait until its done referencing the task. 1423 */ 1424 while (p->on_cpu) 1425 cpu_relax(); 1426 /* 1427 * Pairs with the smp_wmb() in finish_lock_switch(). 1428 */ 1429 smp_rmb(); 1430 1431 p->sched_contributes_to_load = !!task_contributes_to_load(p); 1432 p->state = TASK_WAKING; 1433 1434 if (p->sched_class->task_waking) 1435 p->sched_class->task_waking(p); 1436 1437 cpu = select_task_rq(p, SD_BALANCE_WAKE, wake_flags); 1438 if (task_cpu(p) != cpu) { 1439 wake_flags |= WF_MIGRATED; 1440 set_task_cpu(p, cpu); 1441 } 1442#endif /* CONFIG_SMP */ 1443 1444 ttwu_queue(p, cpu); 1445stat: 1446 ttwu_stat(p, cpu, wake_flags); 1447out: 1448 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 1449 1450 return success; 1451} 1452 1453/** 1454 * try_to_wake_up_local - try to wake up a local task with rq lock held 1455 * @p: the thread to be awakened 1456 * 1457 * Put @p on the run-queue if it's not already there. The caller must 1458 * ensure that this_rq() is locked, @p is bound to this_rq() and not 1459 * the current task. 1460 */ 1461static void try_to_wake_up_local(struct task_struct *p) 1462{ 1463 struct rq *rq = task_rq(p); 1464 1465 BUG_ON(rq != this_rq()); 1466 BUG_ON(p == current); 1467 lockdep_assert_held(&rq->lock); 1468 1469 if (!raw_spin_trylock(&p->pi_lock)) { 1470 raw_spin_unlock(&rq->lock); 1471 raw_spin_lock(&p->pi_lock); 1472 raw_spin_lock(&rq->lock); 1473 } 1474 1475 if (!(p->state & TASK_NORMAL)) 1476 goto out; 1477 1478 if (!p->on_rq) 1479 ttwu_activate(rq, p, ENQUEUE_WAKEUP); 1480 1481 ttwu_do_wakeup(rq, p, 0); 1482 ttwu_stat(p, smp_processor_id(), 0); 1483out: 1484 raw_spin_unlock(&p->pi_lock); 1485} 1486 1487/** 1488 * wake_up_process - Wake up a specific process 1489 * @p: The process to be woken up. 1490 * 1491 * Attempt to wake up the nominated process and move it to the set of runnable 1492 * processes. Returns 1 if the process was woken up, 0 if it was already 1493 * running. 1494 * 1495 * It may be assumed that this function implies a write memory barrier before 1496 * changing the task state if and only if any tasks are woken up. 1497 */ 1498int wake_up_process(struct task_struct *p) 1499{ 1500 return try_to_wake_up(p, TASK_ALL, 0); 1501} 1502EXPORT_SYMBOL(wake_up_process); 1503 1504int wake_up_state(struct task_struct *p, unsigned int state) 1505{ 1506 return try_to_wake_up(p, state, 0); 1507} 1508 1509/* 1510 * Perform scheduler related setup for a newly forked process p. 1511 * p is forked by current. 1512 * 1513 * __sched_fork() is basic setup used by init_idle() too: 1514 */ 1515static void __sched_fork(struct task_struct *p) 1516{ 1517 p->on_rq = 0; 1518 1519 p->se.on_rq = 0; 1520 p->se.exec_start = 0; 1521 p->se.sum_exec_runtime = 0; 1522 p->se.prev_sum_exec_runtime = 0; 1523 p->se.nr_migrations = 0; 1524 p->se.vruntime = 0; 1525 INIT_LIST_HEAD(&p->se.group_node); 1526 1527#ifdef CONFIG_SMP 1528 p->se.avg.runnable_avg_period = 0; 1529 p->se.avg.runnable_avg_sum = 0; 1530#endif 1531#ifdef CONFIG_SCHEDSTATS 1532 memset(&p->se.statistics, 0, sizeof(p->se.statistics)); 1533#endif 1534 1535 INIT_LIST_HEAD(&p->rt.run_list); 1536 1537#ifdef CONFIG_PREEMPT_NOTIFIERS 1538 INIT_HLIST_HEAD(&p->preempt_notifiers); 1539#endif 1540} 1541 1542/* 1543 * fork()/clone()-time setup: 1544 */ 1545void sched_fork(struct task_struct *p) 1546{ 1547 unsigned long flags; 1548 int cpu = get_cpu(); 1549 1550 __sched_fork(p); 1551 /* 1552 * We mark the process as running here. This guarantees that 1553 * nobody will actually run it, and a signal or other external 1554 * event cannot wake it up and insert it on the runqueue either. 1555 */ 1556 p->state = TASK_RUNNING; 1557 1558 /* 1559 * Make sure we do not leak PI boosting priority to the child. 1560 */ 1561 p->prio = current->normal_prio; 1562 1563 /* 1564 * Revert to default priority/policy on fork if requested. 1565 */ 1566 if (unlikely(p->sched_reset_on_fork)) { 1567 if (task_has_rt_policy(p)) { 1568 p->policy = SCHED_NORMAL; 1569 p->static_prio = NICE_TO_PRIO(0); 1570 p->rt_priority = 0; 1571 } else if (PRIO_TO_NICE(p->static_prio) < 0) 1572 p->static_prio = NICE_TO_PRIO(0); 1573 1574 p->prio = p->normal_prio = __normal_prio(p); 1575 set_load_weight(p); 1576 1577 /* 1578 * We don't need the reset flag anymore after the fork. It has 1579 * fulfilled its duty: 1580 */ 1581 p->sched_reset_on_fork = 0; 1582 } 1583 1584 if (!rt_prio(p->prio)) 1585 p->sched_class = &fair_sched_class; 1586 1587 if (p->sched_class->task_fork) 1588 p->sched_class->task_fork(p); 1589 1590 /* 1591 * The child is not yet in the pid-hash so no cgroup attach races, 1592 * and the cgroup is pinned to this child due to cgroup_fork() 1593 * is ran before sched_fork(). 1594 * 1595 * Silence PROVE_RCU. 1596 */ 1597 raw_spin_lock_irqsave(&p->pi_lock, flags); 1598 set_task_cpu(p, cpu); 1599 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 1600 1601#if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT) 1602 if (likely(sched_info_on())) 1603 memset(&p->sched_info, 0, sizeof(p->sched_info)); 1604#endif 1605#if defined(CONFIG_SMP) 1606 p->on_cpu = 0; 1607#endif 1608#ifdef CONFIG_PREEMPT_COUNT 1609 /* Want to start with kernel preemption disabled. */ 1610 task_thread_info(p)->preempt_count = 1; 1611#endif 1612#ifdef CONFIG_SMP 1613 plist_node_init(&p->pushable_tasks, MAX_PRIO); 1614#endif 1615 1616 put_cpu(); 1617} 1618 1619/* 1620 * wake_up_new_task - wake up a newly created task for the first time. 1621 * 1622 * This function will do some initial scheduler statistics housekeeping 1623 * that must be done for every newly created context, then puts the task 1624 * on the runqueue and wakes it. 1625 */ 1626void wake_up_new_task(struct task_struct *p) 1627{ 1628 unsigned long flags; 1629 struct rq *rq; 1630 1631 raw_spin_lock_irqsave(&p->pi_lock, flags); 1632#ifdef CONFIG_SMP 1633 /* 1634 * Fork balancing, do it here and not earlier because: 1635 * - cpus_allowed can change in the fork path 1636 * - any previously selected cpu might disappear through hotplug 1637 */ 1638 set_task_cpu(p, select_task_rq(p, SD_BALANCE_FORK, 0)); 1639#endif 1640 1641 rq = __task_rq_lock(p); 1642 activate_task(rq, p, 0); 1643 p->on_rq = 1; 1644 trace_sched_wakeup_new(p, true); 1645 check_preempt_curr(rq, p, WF_FORK); 1646#ifdef CONFIG_SMP 1647 if (p->sched_class->task_woken) 1648 p->sched_class->task_woken(rq, p); 1649#endif 1650 task_rq_unlock(rq, p, &flags); 1651} 1652 1653#ifdef CONFIG_PREEMPT_NOTIFIERS 1654 1655/** 1656 * preempt_notifier_register - tell me when current is being preempted & rescheduled 1657 * @notifier: notifier struct to register 1658 */ 1659void preempt_notifier_register(struct preempt_notifier *notifier) 1660{ 1661 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers); 1662} 1663EXPORT_SYMBOL_GPL(preempt_notifier_register); 1664 1665/** 1666 * preempt_notifier_unregister - no longer interested in preemption notifications 1667 * @notifier: notifier struct to unregister 1668 * 1669 * This is safe to call from within a preemption notifier. 1670 */ 1671void preempt_notifier_unregister(struct preempt_notifier *notifier) 1672{ 1673 hlist_del(¬ifier->link); 1674} 1675EXPORT_SYMBOL_GPL(preempt_notifier_unregister); 1676 1677static void fire_sched_in_preempt_notifiers(struct task_struct *curr) 1678{ 1679 struct preempt_notifier *notifier; 1680 struct hlist_node *node; 1681 1682 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link) 1683 notifier->ops->sched_in(notifier, raw_smp_processor_id()); 1684} 1685 1686static void 1687fire_sched_out_preempt_notifiers(struct task_struct *curr, 1688 struct task_struct *next) 1689{ 1690 struct preempt_notifier *notifier; 1691 struct hlist_node *node; 1692 1693 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link) 1694 notifier->ops->sched_out(notifier, next); 1695} 1696 1697#else /* !CONFIG_PREEMPT_NOTIFIERS */ 1698 1699static void fire_sched_in_preempt_notifiers(struct task_struct *curr) 1700{ 1701} 1702 1703static void 1704fire_sched_out_preempt_notifiers(struct task_struct *curr, 1705 struct task_struct *next) 1706{ 1707} 1708 1709#endif /* CONFIG_PREEMPT_NOTIFIERS */ 1710 1711/** 1712 * prepare_task_switch - prepare to switch tasks 1713 * @rq: the runqueue preparing to switch 1714 * @prev: the current task that is being switched out 1715 * @next: the task we are going to switch to. 1716 * 1717 * This is called with the rq lock held and interrupts off. It must 1718 * be paired with a subsequent finish_task_switch after the context 1719 * switch. 1720 * 1721 * prepare_task_switch sets up locking and calls architecture specific 1722 * hooks. 1723 */ 1724static inline void 1725prepare_task_switch(struct rq *rq, struct task_struct *prev, 1726 struct task_struct *next) 1727{ 1728 trace_sched_switch(prev, next); 1729 sched_info_switch(prev, next); 1730 perf_event_task_sched_out(prev, next); 1731 fire_sched_out_preempt_notifiers(prev, next); 1732 prepare_lock_switch(rq, next); 1733 prepare_arch_switch(next); 1734} 1735 1736/** 1737 * finish_task_switch - clean up after a task-switch 1738 * @rq: runqueue associated with task-switch 1739 * @prev: the thread we just switched away from. 1740 * 1741 * finish_task_switch must be called after the context switch, paired 1742 * with a prepare_task_switch call before the context switch. 1743 * finish_task_switch will reconcile locking set up by prepare_task_switch, 1744 * and do any other architecture-specific cleanup actions. 1745 * 1746 * Note that we may have delayed dropping an mm in context_switch(). If 1747 * so, we finish that here outside of the runqueue lock. (Doing it 1748 * with the lock held can cause deadlocks; see schedule() for 1749 * details.) 1750 */ 1751static void finish_task_switch(struct rq *rq, struct task_struct *prev) 1752 __releases(rq->lock) 1753{ 1754 struct mm_struct *mm = rq->prev_mm; 1755 long prev_state; 1756 1757 rq->prev_mm = NULL; 1758 1759 /* 1760 * A task struct has one reference for the use as "current". 1761 * If a task dies, then it sets TASK_DEAD in tsk->state and calls 1762 * schedule one last time. The schedule call will never return, and 1763 * the scheduled task must drop that reference. 1764 * The test for TASK_DEAD must occur while the runqueue locks are 1765 * still held, otherwise prev could be scheduled on another cpu, die 1766 * there before we look at prev->state, and then the reference would 1767 * be dropped twice. 1768 * Manfred Spraul <manfred@colorfullife.com> 1769 */ 1770 prev_state = prev->state; 1771 vtime_task_switch(prev); 1772 finish_arch_switch(prev); 1773 perf_event_task_sched_in(prev, current); 1774 finish_lock_switch(rq, prev); 1775 finish_arch_post_lock_switch(); 1776 1777 fire_sched_in_preempt_notifiers(current); 1778 if (mm) 1779 mmdrop(mm); 1780 if (unlikely(prev_state == TASK_DEAD)) { 1781 /* 1782 * Remove function-return probe instances associated with this 1783 * task and put them back on the free list. 1784 */ 1785 kprobe_flush_task(prev); 1786 put_task_struct(prev); 1787 } 1788} 1789 1790#ifdef CONFIG_SMP 1791 1792/* assumes rq->lock is held */ 1793static inline void pre_schedule(struct rq *rq, struct task_struct *prev) 1794{ 1795 if (prev->sched_class->pre_schedule) 1796 prev->sched_class->pre_schedule(rq, prev); 1797} 1798 1799/* rq->lock is NOT held, but preemption is disabled */ 1800static inline void post_schedule(struct rq *rq) 1801{ 1802 if (rq->post_schedule) { 1803 unsigned long flags; 1804 1805 raw_spin_lock_irqsave(&rq->lock, flags); 1806 if (rq->curr->sched_class->post_schedule) 1807 rq->curr->sched_class->post_schedule(rq); 1808 raw_spin_unlock_irqrestore(&rq->lock, flags); 1809 1810 rq->post_schedule = 0; 1811 } 1812} 1813 1814#else 1815 1816static inline void pre_schedule(struct rq *rq, struct task_struct *p) 1817{ 1818} 1819 1820static inline void post_schedule(struct rq *rq) 1821{ 1822} 1823 1824#endif 1825 1826/** 1827 * schedule_tail - first thing a freshly forked thread must call. 1828 * @prev: the thread we just switched away from. 1829 */ 1830asmlinkage void schedule_tail(struct task_struct *prev) 1831 __releases(rq->lock) 1832{ 1833 struct rq *rq = this_rq(); 1834 1835 finish_task_switch(rq, prev); 1836 1837 /* 1838 * FIXME: do we need to worry about rq being invalidated by the 1839 * task_switch? 1840 */ 1841 post_schedule(rq); 1842 1843#ifdef __ARCH_WANT_UNLOCKED_CTXSW 1844 /* In this case, finish_task_switch does not reenable preemption */ 1845 preempt_enable(); 1846#endif 1847 if (current->set_child_tid) 1848 put_user(task_pid_vnr(current), current->set_child_tid); 1849} 1850 1851/* 1852 * context_switch - switch to the new MM and the new 1853 * thread's register state. 1854 */ 1855static inline void 1856context_switch(struct rq *rq, struct task_struct *prev, 1857 struct task_struct *next) 1858{ 1859 struct mm_struct *mm, *oldmm; 1860 1861 prepare_task_switch(rq, prev, next); 1862 1863 mm = next->mm; 1864 oldmm = prev->active_mm; 1865 /* 1866 * For paravirt, this is coupled with an exit in switch_to to 1867 * combine the page table reload and the switch backend into 1868 * one hypercall. 1869 */ 1870 arch_start_context_switch(prev); 1871 1872 if (!mm) { 1873 next->active_mm = oldmm; 1874 atomic_inc(&oldmm->mm_count); 1875 enter_lazy_tlb(oldmm, next); 1876 } else 1877 switch_mm(oldmm, mm, next); 1878 1879 if (!prev->mm) { 1880 prev->active_mm = NULL; 1881 rq->prev_mm = oldmm; 1882 } 1883 /* 1884 * Since the runqueue lock will be released by the next 1885 * task (which is an invalid locking op but in the case 1886 * of the scheduler it's an obvious special-case), so we 1887 * do an early lockdep release here: 1888 */ 1889#ifndef __ARCH_WANT_UNLOCKED_CTXSW 1890 spin_release(&rq->lock.dep_map, 1, _THIS_IP_); 1891#endif 1892 1893 /* Here we just switch the register state and the stack. */ 1894 rcu_switch(prev, next); 1895 switch_to(prev, next, prev); 1896 1897 barrier(); 1898 /* 1899 * this_rq must be evaluated again because prev may have moved 1900 * CPUs since it called schedule(), thus the 'rq' on its stack 1901 * frame will be invalid. 1902 */ 1903 finish_task_switch(this_rq(), prev); 1904} 1905 1906/* 1907 * nr_running, nr_uninterruptible and nr_context_switches: 1908 * 1909 * externally visible scheduler statistics: current number of runnable 1910 * threads, current number of uninterruptible-sleeping threads, total 1911 * number of context switches performed since bootup. 1912 */ 1913unsigned long nr_running(void) 1914{ 1915 unsigned long i, sum = 0; 1916 1917 for_each_online_cpu(i) 1918 sum += cpu_rq(i)->nr_running; 1919 1920 return sum; 1921} 1922 1923unsigned long nr_uninterruptible(void) 1924{ 1925 unsigned long i, sum = 0; 1926 1927 for_each_possible_cpu(i) 1928 sum += cpu_rq(i)->nr_uninterruptible; 1929 1930 /* 1931 * Since we read the counters lockless, it might be slightly 1932 * inaccurate. Do not allow it to go below zero though: 1933 */ 1934 if (unlikely((long)sum < 0)) 1935 sum = 0; 1936 1937 return sum; 1938} 1939 1940unsigned long long nr_context_switches(void) 1941{ 1942 int i; 1943 unsigned long long sum = 0; 1944 1945 for_each_possible_cpu(i) 1946 sum += cpu_rq(i)->nr_switches; 1947 1948 return sum; 1949} 1950 1951unsigned long nr_iowait(void) 1952{ 1953 unsigned long i, sum = 0; 1954 1955 for_each_possible_cpu(i) 1956 sum += atomic_read(&cpu_rq(i)->nr_iowait); 1957 1958 return sum; 1959} 1960 1961unsigned long nr_iowait_cpu(int cpu) 1962{ 1963 struct rq *this = cpu_rq(cpu); 1964 return atomic_read(&this->nr_iowait); 1965} 1966 1967unsigned long this_cpu_load(void) 1968{ 1969 struct rq *this = this_rq(); 1970 return this->cpu_load[0]; 1971} 1972 1973 1974/* 1975 * Global load-average calculations 1976 * 1977 * We take a distributed and async approach to calculating the global load-avg 1978 * in order to minimize overhead. 1979 * 1980 * The global load average is an exponentially decaying average of nr_running + 1981 * nr_uninterruptible. 1982 * 1983 * Once every LOAD_FREQ: 1984 * 1985 * nr_active = 0; 1986 * for_each_possible_cpu(cpu) 1987 * nr_active += cpu_of(cpu)->nr_running + cpu_of(cpu)->nr_uninterruptible; 1988 * 1989 * avenrun[n] = avenrun[0] * exp_n + nr_active * (1 - exp_n) 1990 * 1991 * Due to a number of reasons the above turns in the mess below: 1992 * 1993 * - for_each_possible_cpu() is prohibitively expensive on machines with 1994 * serious number of cpus, therefore we need to take a distributed approach 1995 * to calculating nr_active. 1996 * 1997 * \Sum_i x_i(t) = \Sum_i x_i(t) - x_i(t_0) | x_i(t_0) := 0 1998 * = \Sum_i { \Sum_j=1 x_i(t_j) - x_i(t_j-1) } 1999 * 2000 * So assuming nr_active := 0 when we start out -- true per definition, we 2001 * can simply take per-cpu deltas and fold those into a global accumulate 2002 * to obtain the same result. See calc_load_fold_active(). 2003 * 2004 * Furthermore, in order to avoid synchronizing all per-cpu delta folding 2005 * across the machine, we assume 10 ticks is sufficient time for every 2006 * cpu to have completed this task. 2007 * 2008 * This places an upper-bound on the IRQ-off latency of the machine. Then 2009 * again, being late doesn't loose the delta, just wrecks the sample. 2010 * 2011 * - cpu_rq()->nr_uninterruptible isn't accurately tracked per-cpu because 2012 * this would add another cross-cpu cacheline miss and atomic operation 2013 * to the wakeup path. Instead we increment on whatever cpu the task ran 2014 * when it went into uninterruptible state and decrement on whatever cpu 2015 * did the wakeup. This means that only the sum of nr_uninterruptible over 2016 * all cpus yields the correct result. 2017 * 2018 * This covers the NO_HZ=n code, for extra head-aches, see the comment below. 2019 */ 2020 2021/* Variables and functions for calc_load */ 2022static atomic_long_t calc_load_tasks; 2023static unsigned long calc_load_update; 2024unsigned long avenrun[3]; 2025EXPORT_SYMBOL(avenrun); /* should be removed */ 2026 2027/** 2028 * get_avenrun - get the load average array 2029 * @loads: pointer to dest load array 2030 * @offset: offset to add 2031 * @shift: shift count to shift the result left 2032 * 2033 * These values are estimates at best, so no need for locking. 2034 */ 2035void get_avenrun(unsigned long *loads, unsigned long offset, int shift) 2036{ 2037 loads[0] = (avenrun[0] + offset) << shift; 2038 loads[1] = (avenrun[1] + offset) << shift; 2039 loads[2] = (avenrun[2] + offset) << shift; 2040} 2041 2042static long calc_load_fold_active(struct rq *this_rq) 2043{ 2044 long nr_active, delta = 0; 2045 2046 nr_active = this_rq->nr_running; 2047 nr_active += (long) this_rq->nr_uninterruptible; 2048 2049 if (nr_active != this_rq->calc_load_active) { 2050 delta = nr_active - this_rq->calc_load_active; 2051 this_rq->calc_load_active = nr_active; 2052 } 2053 2054 return delta; 2055} 2056 2057/* 2058 * a1 = a0 * e + a * (1 - e) 2059 */ 2060static unsigned long 2061calc_load(unsigned long load, unsigned long exp, unsigned long active) 2062{ 2063 load *= exp; 2064 load += active * (FIXED_1 - exp); 2065 load += 1UL << (FSHIFT - 1); 2066 return load >> FSHIFT; 2067} 2068 2069#ifdef CONFIG_NO_HZ 2070/* 2071 * Handle NO_HZ for the global load-average. 2072 * 2073 * Since the above described distributed algorithm to compute the global 2074 * load-average relies on per-cpu sampling from the tick, it is affected by 2075 * NO_HZ. 2076 * 2077 * The basic idea is to fold the nr_active delta into a global idle-delta upon 2078 * entering NO_HZ state such that we can include this as an 'extra' cpu delta 2079 * when we read the global state. 2080 * 2081 * Obviously reality has to ruin such a delightfully simple scheme: 2082 * 2083 * - When we go NO_HZ idle during the window, we can negate our sample 2084 * contribution, causing under-accounting. 2085 * 2086 * We avoid this by keeping two idle-delta counters and flipping them 2087 * when the window starts, thus separating old and new NO_HZ load. 2088 * 2089 * The only trick is the slight shift in index flip for read vs write. 2090 * 2091 * 0s 5s 10s 15s 2092 * +10 +10 +10 +10 2093 * |-|-----------|-|-----------|-|-----------|-| 2094 * r:0 0 1 1 0 0 1 1 0 2095 * w:0 1 1 0 0 1 1 0 0 2096 * 2097 * This ensures we'll fold the old idle contribution in this window while 2098 * accumlating the new one. 2099 * 2100 * - When we wake up from NO_HZ idle during the window, we push up our 2101 * contribution, since we effectively move our sample point to a known 2102 * busy state. 2103 * 2104 * This is solved by pushing the window forward, and thus skipping the 2105 * sample, for this cpu (effectively using the idle-delta for this cpu which 2106 * was in effect at the time the window opened). This also solves the issue 2107 * of having to deal with a cpu having been in NOHZ idle for multiple 2108 * LOAD_FREQ intervals. 2109 * 2110 * When making the ILB scale, we should try to pull this in as well. 2111 */ 2112static atomic_long_t calc_load_idle[2]; 2113static int calc_load_idx; 2114 2115static inline int calc_load_write_idx(void) 2116{ 2117 int idx = calc_load_idx; 2118 2119 /* 2120 * See calc_global_nohz(), if we observe the new index, we also 2121 * need to observe the new update time. 2122 */ 2123 smp_rmb(); 2124 2125 /* 2126 * If the folding window started, make sure we start writing in the 2127 * next idle-delta. 2128 */ 2129 if (!time_before(jiffies, calc_load_update)) 2130 idx++; 2131 2132 return idx & 1; 2133} 2134 2135static inline int calc_load_read_idx(void) 2136{ 2137 return calc_load_idx & 1; 2138} 2139 2140void calc_load_enter_idle(void) 2141{ 2142 struct rq *this_rq = this_rq(); 2143 long delta; 2144 2145 /* 2146 * We're going into NOHZ mode, if there's any pending delta, fold it 2147 * into the pending idle delta. 2148 */ 2149 delta = calc_load_fold_active(this_rq); 2150 if (delta) { 2151 int idx = calc_load_write_idx(); 2152 atomic_long_add(delta, &calc_load_idle[idx]); 2153 } 2154} 2155 2156void calc_load_exit_idle(void) 2157{ 2158 struct rq *this_rq = this_rq(); 2159 2160 /* 2161 * If we're still before the sample window, we're done. 2162 */ 2163 if (time_before(jiffies, this_rq->calc_load_update)) 2164 return; 2165 2166 /* 2167 * We woke inside or after the sample window, this means we're already 2168 * accounted through the nohz accounting, so skip the entire deal and 2169 * sync up for the next window. 2170 */ 2171 this_rq->calc_load_update = calc_load_update; 2172 if (time_before(jiffies, this_rq->calc_load_update + 10)) 2173 this_rq->calc_load_update += LOAD_FREQ; 2174} 2175 2176static long calc_load_fold_idle(void) 2177{ 2178 int idx = calc_load_read_idx(); 2179 long delta = 0; 2180 2181 if (atomic_long_read(&calc_load_idle[idx])) 2182 delta = atomic_long_xchg(&calc_load_idle[idx], 0); 2183 2184 return delta; 2185} 2186 2187/** 2188 * fixed_power_int - compute: x^n, in O(log n) time 2189 * 2190 * @x: base of the power 2191 * @frac_bits: fractional bits of @x 2192 * @n: power to raise @x to. 2193 * 2194 * By exploiting the relation between the definition of the natural power 2195 * function: x^n := x*x*...*x (x multiplied by itself for n times), and 2196 * the binary encoding of numbers used by computers: n := \Sum n_i * 2^i, 2197 * (where: n_i \elem {0, 1}, the binary vector representing n), 2198 * we find: x^n := x^(\Sum n_i * 2^i) := \Prod x^(n_i * 2^i), which is 2199 * of course trivially computable in O(log_2 n), the length of our binary 2200 * vector. 2201 */ 2202static unsigned long 2203fixed_power_int(unsigned long x, unsigned int frac_bits, unsigned int n) 2204{ 2205 unsigned long result = 1UL << frac_bits; 2206 2207 if (n) for (;;) { 2208 if (n & 1) { 2209 result *= x; 2210 result += 1UL << (frac_bits - 1); 2211 result >>= frac_bits; 2212 } 2213 n >>= 1; 2214 if (!n) 2215 break; 2216 x *= x; 2217 x += 1UL << (frac_bits - 1); 2218 x >>= frac_bits; 2219 } 2220 2221 return result; 2222} 2223 2224/* 2225 * a1 = a0 * e + a * (1 - e) 2226 * 2227 * a2 = a1 * e + a * (1 - e) 2228 * = (a0 * e + a * (1 - e)) * e + a * (1 - e) 2229 * = a0 * e^2 + a * (1 - e) * (1 + e) 2230 * 2231 * a3 = a2 * e + a * (1 - e) 2232 * = (a0 * e^2 + a * (1 - e) * (1 + e)) * e + a * (1 - e) 2233 * = a0 * e^3 + a * (1 - e) * (1 + e + e^2) 2234 * 2235 * ... 2236 * 2237 * an = a0 * e^n + a * (1 - e) * (1 + e + ... + e^n-1) [1] 2238 * = a0 * e^n + a * (1 - e) * (1 - e^n)/(1 - e) 2239 * = a0 * e^n + a * (1 - e^n) 2240 * 2241 * [1] application of the geometric series: 2242 * 2243 * n 1 - x^(n+1) 2244 * S_n := \Sum x^i = ------------- 2245 * i=0 1 - x 2246 */ 2247static unsigned long 2248calc_load_n(unsigned long load, unsigned long exp, 2249 unsigned long active, unsigned int n) 2250{ 2251 2252 return calc_load(load, fixed_power_int(exp, FSHIFT, n), active); 2253} 2254 2255/* 2256 * NO_HZ can leave us missing all per-cpu ticks calling 2257 * calc_load_account_active(), but since an idle CPU folds its delta into 2258 * calc_load_tasks_idle per calc_load_account_idle(), all we need to do is fold 2259 * in the pending idle delta if our idle period crossed a load cycle boundary. 2260 * 2261 * Once we've updated the global active value, we need to apply the exponential 2262 * weights adjusted to the number of cycles missed. 2263 */ 2264static void calc_global_nohz(void) 2265{ 2266 long delta, active, n; 2267 2268 if (!time_before(jiffies, calc_load_update + 10)) { 2269 /* 2270 * Catch-up, fold however many we are behind still 2271 */ 2272 delta = jiffies - calc_load_update - 10; 2273 n = 1 + (delta / LOAD_FREQ); 2274 2275 active = atomic_long_read(&calc_load_tasks); 2276 active = active > 0 ? active * FIXED_1 : 0; 2277 2278 avenrun[0] = calc_load_n(avenrun[0], EXP_1, active, n); 2279 avenrun[1] = calc_load_n(avenrun[1], EXP_5, active, n); 2280 avenrun[2] = calc_load_n(avenrun[2], EXP_15, active, n); 2281 2282 calc_load_update += n * LOAD_FREQ; 2283 } 2284 2285 /* 2286 * Flip the idle index... 2287 * 2288 * Make sure we first write the new time then flip the index, so that 2289 * calc_load_write_idx() will see the new time when it reads the new 2290 * index, this avoids a double flip messing things up. 2291 */ 2292 smp_wmb(); 2293 calc_load_idx++; 2294} 2295#else /* !CONFIG_NO_HZ */ 2296 2297static inline long calc_load_fold_idle(void) { return 0; } 2298static inline void calc_global_nohz(void) { } 2299 2300#endif /* CONFIG_NO_HZ */ 2301 2302/* 2303 * calc_load - update the avenrun load estimates 10 ticks after the 2304 * CPUs have updated calc_load_tasks. 2305 */ 2306void calc_global_load(unsigned long ticks) 2307{ 2308 long active, delta; 2309 2310 if (time_before(jiffies, calc_load_update + 10)) 2311 return; 2312 2313 /* 2314 * Fold the 'old' idle-delta to include all NO_HZ cpus. 2315 */ 2316 delta = calc_load_fold_idle(); 2317 if (delta) 2318 atomic_long_add(delta, &calc_load_tasks); 2319 2320 active = atomic_long_read(&calc_load_tasks); 2321 active = active > 0 ? active * FIXED_1 : 0; 2322 2323 avenrun[0] = calc_load(avenrun[0], EXP_1, active); 2324 avenrun[1] = calc_load(avenrun[1], EXP_5, active); 2325 avenrun[2] = calc_load(avenrun[2], EXP_15, active); 2326 2327 calc_load_update += LOAD_FREQ; 2328 2329 /* 2330 * In case we idled for multiple LOAD_FREQ intervals, catch up in bulk. 2331 */ 2332 calc_global_nohz(); 2333} 2334 2335/* 2336 * Called from update_cpu_load() to periodically update this CPU's 2337 * active count. 2338 */ 2339static void calc_load_account_active(struct rq *this_rq) 2340{ 2341 long delta; 2342 2343 if (time_before(jiffies, this_rq->calc_load_update)) 2344 return; 2345 2346 delta = calc_load_fold_active(this_rq); 2347 if (delta) 2348 atomic_long_add(delta, &calc_load_tasks); 2349 2350 this_rq->calc_load_update += LOAD_FREQ; 2351} 2352 2353/* 2354 * End of global load-average stuff 2355 */ 2356 2357/* 2358 * The exact cpuload at various idx values, calculated at every tick would be 2359 * load = (2^idx - 1) / 2^idx * load + 1 / 2^idx * cur_load 2360 * 2361 * If a cpu misses updates for n-1 ticks (as it was idle) and update gets called 2362 * on nth tick when cpu may be busy, then we have: 2363 * load = ((2^idx - 1) / 2^idx)^(n-1) * load 2364 * load = (2^idx - 1) / 2^idx) * load + 1 / 2^idx * cur_load 2365 * 2366 * decay_load_missed() below does efficient calculation of 2367 * load = ((2^idx - 1) / 2^idx)^(n-1) * load 2368 * avoiding 0..n-1 loop doing load = ((2^idx - 1) / 2^idx) * load 2369 * 2370 * The calculation is approximated on a 128 point scale. 2371 * degrade_zero_ticks is the number of ticks after which load at any 2372 * particular idx is approximated to be zero. 2373 * degrade_factor is a precomputed table, a row for each load idx. 2374 * Each column corresponds to degradation factor for a power of two ticks, 2375 * based on 128 point scale. 2376 * Example: 2377 * row 2, col 3 (=12) says that the degradation at load idx 2 after 2378 * 8 ticks is 12/128 (which is an approximation of exact factor 3^8/4^8). 2379 * 2380 * With this power of 2 load factors, we can degrade the load n times 2381 * by looking at 1 bits in n and doing as many mult/shift instead of 2382 * n mult/shifts needed by the exact degradation. 2383 */ 2384#define DEGRADE_SHIFT 7 2385static const unsigned char 2386 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128}; 2387static const unsigned char 2388 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = { 2389 {0, 0, 0, 0, 0, 0, 0, 0}, 2390 {64, 32, 8, 0, 0, 0, 0, 0}, 2391 {96, 72, 40, 12, 1, 0, 0}, 2392 {112, 98, 75, 43, 15, 1, 0}, 2393 {120, 112, 98, 76, 45, 16, 2} }; 2394 2395/* 2396 * Update cpu_load for any missed ticks, due to tickless idle. The backlog 2397 * would be when CPU is idle and so we just decay the old load without 2398 * adding any new load. 2399 */ 2400static unsigned long 2401decay_load_missed(unsigned long load, unsigned long missed_updates, int idx) 2402{ 2403 int j = 0; 2404 2405 if (!missed_updates) 2406 return load; 2407 2408 if (missed_updates >= degrade_zero_ticks[idx]) 2409 return 0; 2410 2411 if (idx == 1) 2412 return load >> missed_updates; 2413 2414 while (missed_updates) { 2415 if (missed_updates % 2) 2416 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT; 2417 2418 missed_updates >>= 1; 2419 j++; 2420 } 2421 return load; 2422} 2423 2424/* 2425 * Update rq->cpu_load[] statistics. This function is usually called every 2426 * scheduler tick (TICK_NSEC). With tickless idle this will not be called 2427 * every tick. We fix it up based on jiffies. 2428 */ 2429static void __update_cpu_load(struct rq *this_rq, unsigned long this_load, 2430 unsigned long pending_updates) 2431{ 2432 int i, scale; 2433 2434 this_rq->nr_load_updates++; 2435 2436 /* Update our load: */ 2437 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */ 2438 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) { 2439 unsigned long old_load, new_load; 2440 2441 /* scale is effectively 1 << i now, and >> i divides by scale */ 2442 2443 old_load = this_rq->cpu_load[i]; 2444 old_load = decay_load_missed(old_load, pending_updates - 1, i); 2445 new_load = this_load; 2446 /* 2447 * Round up the averaging division if load is increasing. This 2448 * prevents us from getting stuck on 9 if the load is 10, for 2449 * example. 2450 */ 2451 if (new_load > old_load) 2452 new_load += scale - 1; 2453 2454 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i; 2455 } 2456 2457 sched_avg_update(this_rq); 2458} 2459 2460#ifdef CONFIG_NO_HZ 2461/* 2462 * There is no sane way to deal with nohz on smp when using jiffies because the 2463 * cpu doing the jiffies update might drift wrt the cpu doing the jiffy reading 2464 * causing off-by-one errors in observed deltas; {0,2} instead of {1,1}. 2465 * 2466 * Therefore we cannot use the delta approach from the regular tick since that 2467 * would seriously skew the load calculation. However we'll make do for those 2468 * updates happening while idle (nohz_idle_balance) or coming out of idle 2469 * (tick_nohz_idle_exit). 2470 * 2471 * This means we might still be one tick off for nohz periods. 2472 */ 2473 2474/* 2475 * Called from nohz_idle_balance() to update the load ratings before doing the 2476 * idle balance. 2477 */ 2478void update_idle_cpu_load(struct rq *this_rq) 2479{ 2480 unsigned long curr_jiffies = ACCESS_ONCE(jiffies); 2481 unsigned long load = this_rq->load.weight; 2482 unsigned long pending_updates; 2483 2484 /* 2485 * bail if there's load or we're actually up-to-date. 2486 */ 2487 if (load || curr_jiffies == this_rq->last_load_update_tick) 2488 return; 2489 2490 pending_updates = curr_jiffies - this_rq->last_load_update_tick; 2491 this_rq->last_load_update_tick = curr_jiffies; 2492 2493 __update_cpu_load(this_rq, load, pending_updates); 2494} 2495 2496/* 2497 * Called from tick_nohz_idle_exit() -- try and fix up the ticks we missed. 2498 */ 2499void update_cpu_load_nohz(void) 2500{ 2501 struct rq *this_rq = this_rq(); 2502 unsigned long curr_jiffies = ACCESS_ONCE(jiffies); 2503 unsigned long pending_updates; 2504 2505 if (curr_jiffies == this_rq->last_load_update_tick) 2506 return; 2507 2508 raw_spin_lock(&this_rq->lock); 2509 pending_updates = curr_jiffies - this_rq->last_load_update_tick; 2510 if (pending_updates) { 2511 this_rq->last_load_update_tick = curr_jiffies; 2512 /* 2513 * We were idle, this means load 0, the current load might be 2514 * !0 due to remote wakeups and the sort. 2515 */ 2516 __update_cpu_load(this_rq, 0, pending_updates); 2517 } 2518 raw_spin_unlock(&this_rq->lock); 2519} 2520#endif /* CONFIG_NO_HZ */ 2521 2522/* 2523 * Called from scheduler_tick() 2524 */ 2525static void update_cpu_load_active(struct rq *this_rq) 2526{ 2527 /* 2528 * See the mess around update_idle_cpu_load() / update_cpu_load_nohz(). 2529 */ 2530 this_rq->last_load_update_tick = jiffies; 2531 __update_cpu_load(this_rq, this_rq->load.weight, 1); 2532 2533 calc_load_account_active(this_rq); 2534} 2535 2536#ifdef CONFIG_SMP 2537 2538/* 2539 * sched_exec - execve() is a valuable balancing opportunity, because at 2540 * this point the task has the smallest effective memory and cache footprint. 2541 */ 2542void sched_exec(void) 2543{ 2544 struct task_struct *p = current; 2545 unsigned long flags; 2546 int dest_cpu; 2547 2548 raw_spin_lock_irqsave(&p->pi_lock, flags); 2549 dest_cpu = p->sched_class->select_task_rq(p, SD_BALANCE_EXEC, 0); 2550 if (dest_cpu == smp_processor_id()) 2551 goto unlock; 2552 2553 if (likely(cpu_active(dest_cpu))) { 2554 struct migration_arg arg = { p, dest_cpu }; 2555 2556 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 2557 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg); 2558 return; 2559 } 2560unlock: 2561 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 2562} 2563 2564#endif 2565 2566DEFINE_PER_CPU(struct kernel_stat, kstat); 2567DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat); 2568 2569EXPORT_PER_CPU_SYMBOL(kstat); 2570EXPORT_PER_CPU_SYMBOL(kernel_cpustat); 2571 2572/* 2573 * Return any ns on the sched_clock that have not yet been accounted in 2574 * @p in case that task is currently running. 2575 * 2576 * Called with task_rq_lock() held on @rq. 2577 */ 2578static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq) 2579{ 2580 u64 ns = 0; 2581 2582 if (task_current(rq, p)) { 2583 update_rq_clock(rq); 2584 ns = rq->clock_task - p->se.exec_start; 2585 if ((s64)ns < 0) 2586 ns = 0; 2587 } 2588 2589 return ns; 2590} 2591 2592unsigned long long task_delta_exec(struct task_struct *p) 2593{ 2594 unsigned long flags; 2595 struct rq *rq; 2596 u64 ns = 0; 2597 2598 rq = task_rq_lock(p, &flags); 2599 ns = do_task_delta_exec(p, rq); 2600 task_rq_unlock(rq, p, &flags); 2601 2602 return ns; 2603} 2604 2605/* 2606 * Return accounted runtime for the task. 2607 * In case the task is currently running, return the runtime plus current's 2608 * pending runtime that have not been accounted yet. 2609 */ 2610unsigned long long task_sched_runtime(struct task_struct *p) 2611{ 2612 unsigned long flags; 2613 struct rq *rq; 2614 u64 ns = 0; 2615 2616 rq = task_rq_lock(p, &flags); 2617 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq); 2618 task_rq_unlock(rq, p, &flags); 2619 2620 return ns; 2621} 2622 2623/* 2624 * This function gets called by the timer code, with HZ frequency. 2625 * We call it with interrupts disabled. 2626 */ 2627void scheduler_tick(void) 2628{ 2629 int cpu = smp_processor_id(); 2630 struct rq *rq = cpu_rq(cpu); 2631 struct task_struct *curr = rq->curr; 2632 2633 sched_clock_tick(); 2634 2635 raw_spin_lock(&rq->lock); 2636 update_rq_clock(rq); 2637 update_cpu_load_active(rq); 2638 curr->sched_class->task_tick(rq, curr, 0); 2639 raw_spin_unlock(&rq->lock); 2640 2641 perf_event_task_tick(); 2642 2643#ifdef CONFIG_SMP 2644 rq->idle_balance = idle_cpu(cpu); 2645 trigger_load_balance(rq, cpu); 2646#endif 2647} 2648 2649notrace unsigned long get_parent_ip(unsigned long addr) 2650{ 2651 if (in_lock_functions(addr)) { 2652 addr = CALLER_ADDR2; 2653 if (in_lock_functions(addr)) 2654 addr = CALLER_ADDR3; 2655 } 2656 return addr; 2657} 2658 2659#if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \ 2660 defined(CONFIG_PREEMPT_TRACER)) 2661 2662void __kprobes add_preempt_count(int val) 2663{ 2664#ifdef CONFIG_DEBUG_PREEMPT 2665 /* 2666 * Underflow? 2667 */ 2668 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0))) 2669 return; 2670#endif 2671 preempt_count() += val; 2672#ifdef CONFIG_DEBUG_PREEMPT 2673 /* 2674 * Spinlock count overflowing soon? 2675 */ 2676 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >= 2677 PREEMPT_MASK - 10); 2678#endif 2679 if (preempt_count() == val) 2680 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1)); 2681} 2682EXPORT_SYMBOL(add_preempt_count); 2683 2684void __kprobes sub_preempt_count(int val) 2685{ 2686#ifdef CONFIG_DEBUG_PREEMPT 2687 /* 2688 * Underflow? 2689 */ 2690 if (DEBUG_LOCKS_WARN_ON(val > preempt_count())) 2691 return; 2692 /* 2693 * Is the spinlock portion underflowing? 2694 */ 2695 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) && 2696 !(preempt_count() & PREEMPT_MASK))) 2697 return; 2698#endif 2699 2700 if (preempt_count() == val) 2701 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1)); 2702 preempt_count() -= val; 2703} 2704EXPORT_SYMBOL(sub_preempt_count); 2705 2706#endif 2707 2708/* 2709 * Print scheduling while atomic bug: 2710 */ 2711static noinline void __schedule_bug(struct task_struct *prev) 2712{ 2713 if (oops_in_progress) 2714 return; 2715 2716 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n", 2717 prev->comm, prev->pid, preempt_count()); 2718 2719 debug_show_held_locks(prev); 2720 print_modules(); 2721 if (irqs_disabled()) 2722 print_irqtrace_events(prev); 2723 dump_stack(); 2724 add_taint(TAINT_WARN); 2725} 2726 2727/* 2728 * Various schedule()-time debugging checks and statistics: 2729 */ 2730static inline void schedule_debug(struct task_struct *prev) 2731{ 2732 /* 2733 * Test if we are atomic. Since do_exit() needs to call into 2734 * schedule() atomically, we ignore that path for now. 2735 * Otherwise, whine if we are scheduling when we should not be. 2736 */ 2737 if (unlikely(in_atomic_preempt_off() && !prev->exit_state)) 2738 __schedule_bug(prev); 2739 rcu_sleep_check(); 2740 2741 profile_hit(SCHED_PROFILING, __builtin_return_address(0)); 2742 2743 schedstat_inc(this_rq(), sched_count); 2744} 2745 2746static void put_prev_task(struct rq *rq, struct task_struct *prev) 2747{ 2748 if (prev->on_rq || rq->skip_clock_update < 0) 2749 update_rq_clock(rq); 2750 prev->sched_class->put_prev_task(rq, prev); 2751} 2752 2753/* 2754 * Pick up the highest-prio task: 2755 */ 2756static inline struct task_struct * 2757pick_next_task(struct rq *rq) 2758{ 2759 const struct sched_class *class; 2760 struct task_struct *p; 2761 2762 /* 2763 * Optimization: we know that if all tasks are in 2764 * the fair class we can call that function directly: 2765 */ 2766 if (likely(rq->nr_running == rq->cfs.h_nr_running)) { 2767 p = fair_sched_class.pick_next_task(rq); 2768 if (likely(p)) 2769 return p; 2770 } 2771 2772 for_each_class(class) { 2773 p = class->pick_next_task(rq); 2774 if (p) 2775 return p; 2776 } 2777 2778 BUG(); /* the idle class will always have a runnable task */ 2779} 2780 2781/* 2782 * __schedule() is the main scheduler function. 2783 * 2784 * The main means of driving the scheduler and thus entering this function are: 2785 * 2786 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc. 2787 * 2788 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return 2789 * paths. For example, see arch/x86/entry_64.S. 2790 * 2791 * To drive preemption between tasks, the scheduler sets the flag in timer 2792 * interrupt handler scheduler_tick(). 2793 * 2794 * 3. Wakeups don't really cause entry into schedule(). They add a 2795 * task to the run-queue and that's it. 2796 * 2797 * Now, if the new task added to the run-queue preempts the current 2798 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets 2799 * called on the nearest possible occasion: 2800 * 2801 * - If the kernel is preemptible (CONFIG_PREEMPT=y): 2802 * 2803 * - in syscall or exception context, at the next outmost 2804 * preempt_enable(). (this might be as soon as the wake_up()'s 2805 * spin_unlock()!) 2806 * 2807 * - in IRQ context, return from interrupt-handler to 2808 * preemptible context 2809 * 2810 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set) 2811 * then at the next: 2812 * 2813 * - cond_resched() call 2814 * - explicit schedule() call 2815 * - return from syscall or exception to user-space 2816 * - return from interrupt-handler to user-space 2817 */ 2818static void __sched __schedule(void) 2819{ 2820 struct task_struct *prev, *next; 2821 unsigned long *switch_count; 2822 struct rq *rq; 2823 int cpu; 2824 2825need_resched: 2826 preempt_disable(); 2827 cpu = smp_processor_id(); 2828 rq = cpu_rq(cpu); 2829 rcu_note_context_switch(cpu); 2830 prev = rq->curr; 2831 2832 schedule_debug(prev); 2833 2834 if (sched_feat(HRTICK)) 2835 hrtick_clear(rq); 2836 2837 raw_spin_lock_irq(&rq->lock); 2838 2839 switch_count = &prev->nivcsw; 2840 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) { 2841 if (unlikely(signal_pending_state(prev->state, prev))) { 2842 prev->state = TASK_RUNNING; 2843 } else { 2844 deactivate_task(rq, prev, DEQUEUE_SLEEP); 2845 prev->on_rq = 0; 2846 2847 /* 2848 * If a worker went to sleep, notify and ask workqueue 2849 * whether it wants to wake up a task to maintain 2850 * concurrency. 2851 */ 2852 if (prev->flags & PF_WQ_WORKER) { 2853 struct task_struct *to_wakeup; 2854 2855 to_wakeup = wq_worker_sleeping(prev, cpu); 2856 if (to_wakeup) 2857 try_to_wake_up_local(to_wakeup); 2858 } 2859 } 2860 switch_count = &prev->nvcsw; 2861 } 2862 2863 pre_schedule(rq, prev); 2864 2865 if (unlikely(!rq->nr_running)) 2866 idle_balance(cpu, rq); 2867 2868 put_prev_task(rq, prev); 2869 next = pick_next_task(rq); 2870 clear_tsk_need_resched(prev); 2871 rq->skip_clock_update = 0; 2872 2873 if (likely(prev != next)) { 2874 rq->nr_switches++; 2875 rq->curr = next; 2876 ++*switch_count; 2877 2878 context_switch(rq, prev, next); /* unlocks the rq */ 2879 /* 2880 * The context switch have flipped the stack from under us 2881 * and restored the local variables which were saved when 2882 * this task called schedule() in the past. prev == current 2883 * is still correct, but it can be moved to another cpu/rq. 2884 */ 2885 cpu = smp_processor_id(); 2886 rq = cpu_rq(cpu); 2887 } else 2888 raw_spin_unlock_irq(&rq->lock); 2889 2890 post_schedule(rq); 2891 2892 sched_preempt_enable_no_resched(); 2893 if (need_resched()) 2894 goto need_resched; 2895} 2896 2897static inline void sched_submit_work(struct task_struct *tsk) 2898{ 2899 if (!tsk->state || tsk_is_pi_blocked(tsk)) 2900 return; 2901 /* 2902 * If we are going to sleep and we have plugged IO queued, 2903 * make sure to submit it to avoid deadlocks. 2904 */ 2905 if (blk_needs_flush_plug(tsk)) 2906 blk_schedule_flush_plug(tsk); 2907} 2908 2909asmlinkage void __sched schedule(void) 2910{ 2911 struct task_struct *tsk = current; 2912 2913 sched_submit_work(tsk); 2914 __schedule(); 2915} 2916EXPORT_SYMBOL(schedule); 2917 2918#ifdef CONFIG_RCU_USER_QS 2919asmlinkage void __sched schedule_user(void) 2920{ 2921 /* 2922 * If we come here after a random call to set_need_resched(), 2923 * or we have been woken up remotely but the IPI has not yet arrived, 2924 * we haven't yet exited the RCU idle mode. Do it here manually until 2925 * we find a better solution. 2926 */ 2927 rcu_user_exit(); 2928 schedule(); 2929 rcu_user_enter(); 2930} 2931#endif 2932 2933/** 2934 * schedule_preempt_disabled - called with preemption disabled 2935 * 2936 * Returns with preemption disabled. Note: preempt_count must be 1 2937 */ 2938void __sched schedule_preempt_disabled(void) 2939{ 2940 sched_preempt_enable_no_resched(); 2941 schedule(); 2942 preempt_disable(); 2943} 2944 2945#ifdef CONFIG_MUTEX_SPIN_ON_OWNER 2946 2947static inline bool owner_running(struct mutex *lock, struct task_struct *owner) 2948{ 2949 if (lock->owner != owner) 2950 return false; 2951 2952 /* 2953 * Ensure we emit the owner->on_cpu, dereference _after_ checking 2954 * lock->owner still matches owner, if that fails, owner might 2955 * point to free()d memory, if it still matches, the rcu_read_lock() 2956 * ensures the memory stays valid. 2957 */ 2958 barrier(); 2959 2960 return owner->on_cpu; 2961} 2962 2963/* 2964 * Look out! "owner" is an entirely speculative pointer 2965 * access and not reliable. 2966 */ 2967int mutex_spin_on_owner(struct mutex *lock, struct task_struct *owner) 2968{ 2969 if (!sched_feat(OWNER_SPIN)) 2970 return 0; 2971 2972 rcu_read_lock(); 2973 while (owner_running(lock, owner)) { 2974 if (need_resched()) 2975 break; 2976 2977 arch_mutex_cpu_relax(); 2978 } 2979 rcu_read_unlock(); 2980 2981 /* 2982 * We break out the loop above on need_resched() and when the 2983 * owner changed, which is a sign for heavy contention. Return 2984 * success only when lock->owner is NULL. 2985 */ 2986 return lock->owner == NULL; 2987} 2988#endif 2989 2990#ifdef CONFIG_PREEMPT 2991/* 2992 * this is the entry point to schedule() from in-kernel preemption 2993 * off of preempt_enable. Kernel preemptions off return from interrupt 2994 * occur there and call schedule directly. 2995 */ 2996asmlinkage void __sched notrace preempt_schedule(void) 2997{ 2998 struct thread_info *ti = current_thread_info(); 2999 3000 /* 3001 * If there is a non-zero preempt_count or interrupts are disabled, 3002 * we do not want to preempt the current task. Just return.. 3003 */ 3004 if (likely(ti->preempt_count || irqs_disabled())) 3005 return; 3006 3007 do { 3008 add_preempt_count_notrace(PREEMPT_ACTIVE); 3009 __schedule(); 3010 sub_preempt_count_notrace(PREEMPT_ACTIVE); 3011 3012 /* 3013 * Check again in case we missed a preemption opportunity 3014 * between schedule and now. 3015 */ 3016 barrier(); 3017 } while (need_resched()); 3018} 3019EXPORT_SYMBOL(preempt_schedule); 3020 3021/* 3022 * this is the entry point to schedule() from kernel preemption 3023 * off of irq context. 3024 * Note, that this is called and return with irqs disabled. This will 3025 * protect us against recursive calling from irq. 3026 */ 3027asmlinkage void __sched preempt_schedule_irq(void) 3028{ 3029 struct thread_info *ti = current_thread_info(); 3030 3031 /* Catch callers which need to be fixed */ 3032 BUG_ON(ti->preempt_count || !irqs_disabled()); 3033 3034 rcu_user_exit(); 3035 do { 3036 add_preempt_count(PREEMPT_ACTIVE); 3037 local_irq_enable(); 3038 __schedule(); 3039 local_irq_disable(); 3040 sub_preempt_count(PREEMPT_ACTIVE); 3041 3042 /* 3043 * Check again in case we missed a preemption opportunity 3044 * between schedule and now. 3045 */ 3046 barrier(); 3047 } while (need_resched()); 3048} 3049 3050#endif /* CONFIG_PREEMPT */ 3051 3052int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags, 3053 void *key) 3054{ 3055 return try_to_wake_up(curr->private, mode, wake_flags); 3056} 3057EXPORT_SYMBOL(default_wake_function); 3058 3059/* 3060 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just 3061 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve 3062 * number) then we wake all the non-exclusive tasks and one exclusive task. 3063 * 3064 * There are circumstances in which we can try to wake a task which has already 3065 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns 3066 * zero in this (rare) case, and we handle it by continuing to scan the queue. 3067 */ 3068static void __wake_up_common(wait_queue_head_t *q, unsigned int mode, 3069 int nr_exclusive, int wake_flags, void *key) 3070{ 3071 wait_queue_t *curr, *next; 3072 3073 list_for_each_entry_safe(curr, next, &q->task_list, task_list) { 3074 unsigned flags = curr->flags; 3075 3076 if (curr->func(curr, mode, wake_flags, key) && 3077 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive) 3078 break; 3079 } 3080} 3081 3082/** 3083 * __wake_up - wake up threads blocked on a waitqueue. 3084 * @q: the waitqueue 3085 * @mode: which threads 3086 * @nr_exclusive: how many wake-one or wake-many threads to wake up 3087 * @key: is directly passed to the wakeup function 3088 * 3089 * It may be assumed that this function implies a write memory barrier before 3090 * changing the task state if and only if any tasks are woken up. 3091 */ 3092void __wake_up(wait_queue_head_t *q, unsigned int mode, 3093 int nr_exclusive, void *key) 3094{ 3095 unsigned long flags; 3096 3097 spin_lock_irqsave(&q->lock, flags); 3098 __wake_up_common(q, mode, nr_exclusive, 0, key); 3099 spin_unlock_irqrestore(&q->lock, flags); 3100} 3101EXPORT_SYMBOL(__wake_up); 3102 3103/* 3104 * Same as __wake_up but called with the spinlock in wait_queue_head_t held. 3105 */ 3106void __wake_up_locked(wait_queue_head_t *q, unsigned int mode, int nr) 3107{ 3108 __wake_up_common(q, mode, nr, 0, NULL); 3109} 3110EXPORT_SYMBOL_GPL(__wake_up_locked); 3111 3112void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key) 3113{ 3114 __wake_up_common(q, mode, 1, 0, key); 3115} 3116EXPORT_SYMBOL_GPL(__wake_up_locked_key); 3117 3118/** 3119 * __wake_up_sync_key - wake up threads blocked on a waitqueue. 3120 * @q: the waitqueue 3121 * @mode: which threads 3122 * @nr_exclusive: how many wake-one or wake-many threads to wake up 3123 * @key: opaque value to be passed to wakeup targets 3124 * 3125 * The sync wakeup differs that the waker knows that it will schedule 3126 * away soon, so while the target thread will be woken up, it will not 3127 * be migrated to another CPU - ie. the two threads are 'synchronized' 3128 * with each other. This can prevent needless bouncing between CPUs. 3129 * 3130 * On UP it can prevent extra preemption. 3131 * 3132 * It may be assumed that this function implies a write memory barrier before 3133 * changing the task state if and only if any tasks are woken up. 3134 */ 3135void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode, 3136 int nr_exclusive, void *key) 3137{ 3138 unsigned long flags; 3139 int wake_flags = WF_SYNC; 3140 3141 if (unlikely(!q)) 3142 return; 3143 3144 if (unlikely(!nr_exclusive)) 3145 wake_flags = 0; 3146 3147 spin_lock_irqsave(&q->lock, flags); 3148 __wake_up_common(q, mode, nr_exclusive, wake_flags, key); 3149 spin_unlock_irqrestore(&q->lock, flags); 3150} 3151EXPORT_SYMBOL_GPL(__wake_up_sync_key); 3152 3153/* 3154 * __wake_up_sync - see __wake_up_sync_key() 3155 */ 3156void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive) 3157{ 3158 __wake_up_sync_key(q, mode, nr_exclusive, NULL); 3159} 3160EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */ 3161 3162/** 3163 * complete: - signals a single thread waiting on this completion 3164 * @x: holds the state of this particular completion 3165 * 3166 * This will wake up a single thread waiting on this completion. Threads will be 3167 * awakened in the same order in which they were queued. 3168 * 3169 * See also complete_all(), wait_for_completion() and related routines. 3170 * 3171 * It may be assumed that this function implies a write memory barrier before 3172 * changing the task state if and only if any tasks are woken up. 3173 */ 3174void complete(struct completion *x) 3175{ 3176 unsigned long flags; 3177 3178 spin_lock_irqsave(&x->wait.lock, flags); 3179 x->done++; 3180 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL); 3181 spin_unlock_irqrestore(&x->wait.lock, flags); 3182} 3183EXPORT_SYMBOL(complete); 3184 3185/** 3186 * complete_all: - signals all threads waiting on this completion 3187 * @x: holds the state of this particular completion 3188 * 3189 * This will wake up all threads waiting on this particular completion event. 3190 * 3191 * It may be assumed that this function implies a write memory barrier before 3192 * changing the task state if and only if any tasks are woken up. 3193 */ 3194void complete_all(struct completion *x) 3195{ 3196 unsigned long flags; 3197 3198 spin_lock_irqsave(&x->wait.lock, flags); 3199 x->done += UINT_MAX/2; 3200 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL); 3201 spin_unlock_irqrestore(&x->wait.lock, flags); 3202} 3203EXPORT_SYMBOL(complete_all); 3204 3205static inline long __sched 3206do_wait_for_common(struct completion *x, long timeout, int state) 3207{ 3208 if (!x->done) { 3209 DECLARE_WAITQUEUE(wait, current); 3210 3211 __add_wait_queue_tail_exclusive(&x->wait, &wait); 3212 do { 3213 if (signal_pending_state(state, current)) { 3214 timeout = -ERESTARTSYS; 3215 break; 3216 } 3217 __set_current_state(state); 3218 spin_unlock_irq(&x->wait.lock); 3219 timeout = schedule_timeout(timeout); 3220 spin_lock_irq(&x->wait.lock); 3221 } while (!x->done && timeout); 3222 __remove_wait_queue(&x->wait, &wait); 3223 if (!x->done) 3224 return timeout; 3225 } 3226 x->done--; 3227 return timeout ?: 1; 3228} 3229 3230static long __sched 3231wait_for_common(struct completion *x, long timeout, int state) 3232{ 3233 might_sleep(); 3234 3235 spin_lock_irq(&x->wait.lock); 3236 timeout = do_wait_for_common(x, timeout, state); 3237 spin_unlock_irq(&x->wait.lock); 3238 return timeout; 3239} 3240 3241/** 3242 * wait_for_completion: - waits for completion of a task 3243 * @x: holds the state of this particular completion 3244 * 3245 * This waits to be signaled for completion of a specific task. It is NOT 3246 * interruptible and there is no timeout. 3247 * 3248 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout 3249 * and interrupt capability. Also see complete(). 3250 */ 3251void __sched wait_for_completion(struct completion *x) 3252{ 3253 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE); 3254} 3255EXPORT_SYMBOL(wait_for_completion); 3256 3257/** 3258 * wait_for_completion_timeout: - waits for completion of a task (w/timeout) 3259 * @x: holds the state of this particular completion 3260 * @timeout: timeout value in jiffies 3261 * 3262 * This waits for either a completion of a specific task to be signaled or for a 3263 * specified timeout to expire. The timeout is in jiffies. It is not 3264 * interruptible. 3265 * 3266 * The return value is 0 if timed out, and positive (at least 1, or number of 3267 * jiffies left till timeout) if completed. 3268 */ 3269unsigned long __sched 3270wait_for_completion_timeout(struct completion *x, unsigned long timeout) 3271{ 3272 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE); 3273} 3274EXPORT_SYMBOL(wait_for_completion_timeout); 3275 3276/** 3277 * wait_for_completion_interruptible: - waits for completion of a task (w/intr) 3278 * @x: holds the state of this particular completion 3279 * 3280 * This waits for completion of a specific task to be signaled. It is 3281 * interruptible. 3282 * 3283 * The return value is -ERESTARTSYS if interrupted, 0 if completed. 3284 */ 3285int __sched wait_for_completion_interruptible(struct completion *x) 3286{ 3287 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE); 3288 if (t == -ERESTARTSYS) 3289 return t; 3290 return 0; 3291} 3292EXPORT_SYMBOL(wait_for_completion_interruptible); 3293 3294/** 3295 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr)) 3296 * @x: holds the state of this particular completion 3297 * @timeout: timeout value in jiffies 3298 * 3299 * This waits for either a completion of a specific task to be signaled or for a 3300 * specified timeout to expire. It is interruptible. The timeout is in jiffies. 3301 * 3302 * The return value is -ERESTARTSYS if interrupted, 0 if timed out, 3303 * positive (at least 1, or number of jiffies left till timeout) if completed. 3304 */ 3305long __sched 3306wait_for_completion_interruptible_timeout(struct completion *x, 3307 unsigned long timeout) 3308{ 3309 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE); 3310} 3311EXPORT_SYMBOL(wait_for_completion_interruptible_timeout); 3312 3313/** 3314 * wait_for_completion_killable: - waits for completion of a task (killable) 3315 * @x: holds the state of this particular completion 3316 * 3317 * This waits to be signaled for completion of a specific task. It can be 3318 * interrupted by a kill signal. 3319 * 3320 * The return value is -ERESTARTSYS if interrupted, 0 if completed. 3321 */ 3322int __sched wait_for_completion_killable(struct completion *x) 3323{ 3324 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE); 3325 if (t == -ERESTARTSYS) 3326 return t; 3327 return 0; 3328} 3329EXPORT_SYMBOL(wait_for_completion_killable); 3330 3331/** 3332 * wait_for_completion_killable_timeout: - waits for completion of a task (w/(to,killable)) 3333 * @x: holds the state of this particular completion 3334 * @timeout: timeout value in jiffies 3335 * 3336 * This waits for either a completion of a specific task to be 3337 * signaled or for a specified timeout to expire. It can be 3338 * interrupted by a kill signal. The timeout is in jiffies. 3339 * 3340 * The return value is -ERESTARTSYS if interrupted, 0 if timed out, 3341 * positive (at least 1, or number of jiffies left till timeout) if completed. 3342 */ 3343long __sched 3344wait_for_completion_killable_timeout(struct completion *x, 3345 unsigned long timeout) 3346{ 3347 return wait_for_common(x, timeout, TASK_KILLABLE); 3348} 3349EXPORT_SYMBOL(wait_for_completion_killable_timeout); 3350 3351/** 3352 * try_wait_for_completion - try to decrement a completion without blocking 3353 * @x: completion structure 3354 * 3355 * Returns: 0 if a decrement cannot be done without blocking 3356 * 1 if a decrement succeeded. 3357 * 3358 * If a completion is being used as a counting completion, 3359 * attempt to decrement the counter without blocking. This 3360 * enables us to avoid waiting if the resource the completion 3361 * is protecting is not available. 3362 */ 3363bool try_wait_for_completion(struct completion *x) 3364{ 3365 unsigned long flags; 3366 int ret = 1; 3367 3368 spin_lock_irqsave(&x->wait.lock, flags); 3369 if (!x->done) 3370 ret = 0; 3371 else 3372 x->done--; 3373 spin_unlock_irqrestore(&x->wait.lock, flags); 3374 return ret; 3375} 3376EXPORT_SYMBOL(try_wait_for_completion); 3377 3378/** 3379 * completion_done - Test to see if a completion has any waiters 3380 * @x: completion structure 3381 * 3382 * Returns: 0 if there are waiters (wait_for_completion() in progress) 3383 * 1 if there are no waiters. 3384 * 3385 */ 3386bool completion_done(struct completion *x) 3387{ 3388 unsigned long flags; 3389 int ret = 1; 3390 3391 spin_lock_irqsave(&x->wait.lock, flags); 3392 if (!x->done) 3393 ret = 0; 3394 spin_unlock_irqrestore(&x->wait.lock, flags); 3395 return ret; 3396} 3397EXPORT_SYMBOL(completion_done); 3398 3399static long __sched 3400sleep_on_common(wait_queue_head_t *q, int state, long timeout) 3401{ 3402 unsigned long flags; 3403 wait_queue_t wait; 3404 3405 init_waitqueue_entry(&wait, current); 3406 3407 __set_current_state(state); 3408 3409 spin_lock_irqsave(&q->lock, flags); 3410 __add_wait_queue(q, &wait); 3411 spin_unlock(&q->lock); 3412 timeout = schedule_timeout(timeout); 3413 spin_lock_irq(&q->lock); 3414 __remove_wait_queue(q, &wait); 3415 spin_unlock_irqrestore(&q->lock, flags); 3416 3417 return timeout; 3418} 3419 3420void __sched interruptible_sleep_on(wait_queue_head_t *q) 3421{ 3422 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT); 3423} 3424EXPORT_SYMBOL(interruptible_sleep_on); 3425 3426long __sched 3427interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout) 3428{ 3429 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout); 3430} 3431EXPORT_SYMBOL(interruptible_sleep_on_timeout); 3432 3433void __sched sleep_on(wait_queue_head_t *q) 3434{ 3435 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT); 3436} 3437EXPORT_SYMBOL(sleep_on); 3438 3439long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout) 3440{ 3441 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout); 3442} 3443EXPORT_SYMBOL(sleep_on_timeout); 3444 3445#ifdef CONFIG_RT_MUTEXES 3446 3447/* 3448 * rt_mutex_setprio - set the current priority of a task 3449 * @p: task 3450 * @prio: prio value (kernel-internal form) 3451 * 3452 * This function changes the 'effective' priority of a task. It does 3453 * not touch ->normal_prio like __setscheduler(). 3454 * 3455 * Used by the rt_mutex code to implement priority inheritance logic. 3456 */ 3457void rt_mutex_setprio(struct task_struct *p, int prio) 3458{ 3459 int oldprio, on_rq, running; 3460 struct rq *rq; 3461 const struct sched_class *prev_class; 3462 3463 BUG_ON(prio < 0 || prio > MAX_PRIO); 3464 3465 rq = __task_rq_lock(p); 3466 3467 /* 3468 * Idle task boosting is a nono in general. There is one 3469 * exception, when PREEMPT_RT and NOHZ is active: 3470 * 3471 * The idle task calls get_next_timer_interrupt() and holds 3472 * the timer wheel base->lock on the CPU and another CPU wants 3473 * to access the timer (probably to cancel it). We can safely 3474 * ignore the boosting request, as the idle CPU runs this code 3475 * with interrupts disabled and will complete the lock 3476 * protected section without being interrupted. So there is no 3477 * real need to boost. 3478 */ 3479 if (unlikely(p == rq->idle)) { 3480 WARN_ON(p != rq->curr); 3481 WARN_ON(p->pi_blocked_on); 3482 goto out_unlock; 3483 } 3484 3485 trace_sched_pi_setprio(p, prio); 3486 oldprio = p->prio; 3487 prev_class = p->sched_class; 3488 on_rq = p->on_rq; 3489 running = task_current(rq, p); 3490 if (on_rq) 3491 dequeue_task(rq, p, 0); 3492 if (running) 3493 p->sched_class->put_prev_task(rq, p); 3494 3495 if (rt_prio(prio)) 3496 p->sched_class = &rt_sched_class; 3497 else 3498 p->sched_class = &fair_sched_class; 3499 3500 p->prio = prio; 3501 3502 if (running) 3503 p->sched_class->set_curr_task(rq); 3504 if (on_rq) 3505 enqueue_task(rq, p, oldprio < prio ? ENQUEUE_HEAD : 0); 3506 3507 check_class_changed(rq, p, prev_class, oldprio); 3508out_unlock: 3509 __task_rq_unlock(rq); 3510} 3511#endif 3512void set_user_nice(struct task_struct *p, long nice) 3513{ 3514 int old_prio, delta, on_rq; 3515 unsigned long flags; 3516 struct rq *rq; 3517 3518 if (TASK_NICE(p) == nice || nice < -20 || nice > 19) 3519 return; 3520 /* 3521 * We have to be careful, if called from sys_setpriority(), 3522 * the task might be in the middle of scheduling on another CPU. 3523 */ 3524 rq = task_rq_lock(p, &flags); 3525 /* 3526 * The RT priorities are set via sched_setscheduler(), but we still 3527 * allow the 'normal' nice value to be set - but as expected 3528 * it wont have any effect on scheduling until the task is 3529 * SCHED_FIFO/SCHED_RR: 3530 */ 3531 if (task_has_rt_policy(p)) { 3532 p->static_prio = NICE_TO_PRIO(nice); 3533 goto out_unlock; 3534 } 3535 on_rq = p->on_rq; 3536 if (on_rq) 3537 dequeue_task(rq, p, 0); 3538 3539 p->static_prio = NICE_TO_PRIO(nice); 3540 set_load_weight(p); 3541 old_prio = p->prio; 3542 p->prio = effective_prio(p); 3543 delta = p->prio - old_prio; 3544 3545 if (on_rq) { 3546 enqueue_task(rq, p, 0); 3547 /* 3548 * If the task increased its priority or is running and 3549 * lowered its priority, then reschedule its CPU: 3550 */ 3551 if (delta < 0 || (delta > 0 && task_running(rq, p))) 3552 resched_task(rq->curr); 3553 } 3554out_unlock: 3555 task_rq_unlock(rq, p, &flags); 3556} 3557EXPORT_SYMBOL(set_user_nice); 3558 3559/* 3560 * can_nice - check if a task can reduce its nice value 3561 * @p: task 3562 * @nice: nice value 3563 */ 3564int can_nice(const struct task_struct *p, const int nice) 3565{ 3566 /* convert nice value [19,-20] to rlimit style value [1,40] */ 3567 int nice_rlim = 20 - nice; 3568 3569 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) || 3570 capable(CAP_SYS_NICE)); 3571} 3572 3573#ifdef __ARCH_WANT_SYS_NICE 3574 3575/* 3576 * sys_nice - change the priority of the current process. 3577 * @increment: priority increment 3578 * 3579 * sys_setpriority is a more generic, but much slower function that 3580 * does similar things. 3581 */ 3582SYSCALL_DEFINE1(nice, int, increment) 3583{ 3584 long nice, retval; 3585 3586 /* 3587 * Setpriority might change our priority at the same moment. 3588 * We don't have to worry. Conceptually one call occurs first 3589 * and we have a single winner. 3590 */ 3591 if (increment < -40) 3592 increment = -40; 3593 if (increment > 40) 3594 increment = 40; 3595 3596 nice = TASK_NICE(current) + increment; 3597 if (nice < -20) 3598 nice = -20; 3599 if (nice > 19) 3600 nice = 19; 3601 3602 if (increment < 0 && !can_nice(current, nice)) 3603 return -EPERM; 3604 3605 retval = security_task_setnice(current, nice); 3606 if (retval) 3607 return retval; 3608 3609 set_user_nice(current, nice); 3610 return 0; 3611} 3612 3613#endif 3614 3615/** 3616 * task_prio - return the priority value of a given task. 3617 * @p: the task in question. 3618 * 3619 * This is the priority value as seen by users in /proc. 3620 * RT tasks are offset by -200. Normal tasks are centered 3621 * around 0, value goes from -16 to +15. 3622 */ 3623int task_prio(const struct task_struct *p) 3624{ 3625 return p->prio - MAX_RT_PRIO; 3626} 3627 3628/** 3629 * task_nice - return the nice value of a given task. 3630 * @p: the task in question. 3631 */ 3632int task_nice(const struct task_struct *p) 3633{ 3634 return TASK_NICE(p); 3635} 3636EXPORT_SYMBOL(task_nice); 3637 3638/** 3639 * idle_cpu - is a given cpu idle currently? 3640 * @cpu: the processor in question. 3641 */ 3642int idle_cpu(int cpu) 3643{ 3644 struct rq *rq = cpu_rq(cpu); 3645 3646 if (rq->curr != rq->idle) 3647 return 0; 3648 3649 if (rq->nr_running) 3650 return 0; 3651 3652#ifdef CONFIG_SMP 3653 if (!llist_empty(&rq->wake_list)) 3654 return 0; 3655#endif 3656 3657 return 1; 3658} 3659 3660/** 3661 * idle_task - return the idle task for a given cpu. 3662 * @cpu: the processor in question. 3663 */ 3664struct task_struct *idle_task(int cpu) 3665{ 3666 return cpu_rq(cpu)->idle; 3667} 3668 3669/** 3670 * find_process_by_pid - find a process with a matching PID value. 3671 * @pid: the pid in question. 3672 */ 3673static struct task_struct *find_process_by_pid(pid_t pid) 3674{ 3675 return pid ? find_task_by_vpid(pid) : current; 3676} 3677 3678/* Actually do priority change: must hold rq lock. */ 3679static void 3680__setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio) 3681{ 3682 p->policy = policy; 3683 p->rt_priority = prio; 3684 p->normal_prio = normal_prio(p); 3685 /* we are holding p->pi_lock already */ 3686 p->prio = rt_mutex_getprio(p); 3687 if (rt_prio(p->prio)) 3688 p->sched_class = &rt_sched_class; 3689 else 3690 p->sched_class = &fair_sched_class; 3691 set_load_weight(p); 3692} 3693 3694/* 3695 * check the target process has a UID that matches the current process's 3696 */ 3697static bool check_same_owner(struct task_struct *p) 3698{ 3699 const struct cred *cred = current_cred(), *pcred; 3700 bool match; 3701 3702 rcu_read_lock(); 3703 pcred = __task_cred(p); 3704 match = (uid_eq(cred->euid, pcred->euid) || 3705 uid_eq(cred->euid, pcred->uid)); 3706 rcu_read_unlock(); 3707 return match; 3708} 3709 3710static int __sched_setscheduler(struct task_struct *p, int policy, 3711 const struct sched_param *param, bool user) 3712{ 3713 int retval, oldprio, oldpolicy = -1, on_rq, running; 3714 unsigned long flags; 3715 const struct sched_class *prev_class; 3716 struct rq *rq; 3717 int reset_on_fork; 3718 3719 /* may grab non-irq protected spin_locks */ 3720 BUG_ON(in_interrupt()); 3721recheck: 3722 /* double check policy once rq lock held */ 3723 if (policy < 0) { 3724 reset_on_fork = p->sched_reset_on_fork; 3725 policy = oldpolicy = p->policy; 3726 } else { 3727 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK); 3728 policy &= ~SCHED_RESET_ON_FORK; 3729 3730 if (policy != SCHED_FIFO && policy != SCHED_RR && 3731 policy != SCHED_NORMAL && policy != SCHED_BATCH && 3732 policy != SCHED_IDLE) 3733 return -EINVAL; 3734 } 3735 3736 /* 3737 * Valid priorities for SCHED_FIFO and SCHED_RR are 3738 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL, 3739 * SCHED_BATCH and SCHED_IDLE is 0. 3740 */ 3741 if (param->sched_priority < 0 || 3742 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) || 3743 (!p->mm && param->sched_priority > MAX_RT_PRIO-1)) 3744 return -EINVAL; 3745 if (rt_policy(policy) != (param->sched_priority != 0)) 3746 return -EINVAL; 3747 3748 /* 3749 * Allow unprivileged RT tasks to decrease priority: 3750 */ 3751 if (user && !capable(CAP_SYS_NICE)) { 3752 if (rt_policy(policy)) { 3753 unsigned long rlim_rtprio = 3754 task_rlimit(p, RLIMIT_RTPRIO); 3755 3756 /* can't set/change the rt policy */ 3757 if (policy != p->policy && !rlim_rtprio) 3758 return -EPERM; 3759 3760 /* can't increase priority */ 3761 if (param->sched_priority > p->rt_priority && 3762 param->sched_priority > rlim_rtprio) 3763 return -EPERM; 3764 } 3765 3766 /* 3767 * Treat SCHED_IDLE as nice 20. Only allow a switch to 3768 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it. 3769 */ 3770 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) { 3771 if (!can_nice(p, TASK_NICE(p))) 3772 return -EPERM; 3773 } 3774 3775 /* can't change other user's priorities */ 3776 if (!check_same_owner(p)) 3777 return -EPERM; 3778 3779 /* Normal users shall not reset the sched_reset_on_fork flag */ 3780 if (p->sched_reset_on_fork && !reset_on_fork) 3781 return -EPERM; 3782 } 3783 3784 if (user) { 3785 retval = security_task_setscheduler(p); 3786 if (retval) 3787 return retval; 3788 } 3789 3790 /* 3791 * make sure no PI-waiters arrive (or leave) while we are 3792 * changing the priority of the task: 3793 * 3794 * To be able to change p->policy safely, the appropriate 3795 * runqueue lock must be held. 3796 */ 3797 rq = task_rq_lock(p, &flags); 3798 3799 /* 3800 * Changing the policy of the stop threads its a very bad idea 3801 */ 3802 if (p == rq->stop) { 3803 task_rq_unlock(rq, p, &flags); 3804 return -EINVAL; 3805 } 3806 3807 /* 3808 * If not changing anything there's no need to proceed further: 3809 */ 3810 if (unlikely(policy == p->policy && (!rt_policy(policy) || 3811 param->sched_priority == p->rt_priority))) { 3812 task_rq_unlock(rq, p, &flags); 3813 return 0; 3814 } 3815 3816#ifdef CONFIG_RT_GROUP_SCHED 3817 if (user) { 3818 /* 3819 * Do not allow realtime tasks into groups that have no runtime 3820 * assigned. 3821 */ 3822 if (rt_bandwidth_enabled() && rt_policy(policy) && 3823 task_group(p)->rt_bandwidth.rt_runtime == 0 && 3824 !task_group_is_autogroup(task_group(p))) { 3825 task_rq_unlock(rq, p, &flags); 3826 return -EPERM; 3827 } 3828 } 3829#endif 3830 3831 /* recheck policy now with rq lock held */ 3832 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) { 3833 policy = oldpolicy = -1; 3834 task_rq_unlock(rq, p, &flags); 3835 goto recheck; 3836 } 3837 on_rq = p->on_rq; 3838 running = task_current(rq, p); 3839 if (on_rq) 3840 dequeue_task(rq, p, 0); 3841 if (running) 3842 p->sched_class->put_prev_task(rq, p); 3843 3844 p->sched_reset_on_fork = reset_on_fork; 3845 3846 oldprio = p->prio; 3847 prev_class = p->sched_class; 3848 __setscheduler(rq, p, policy, param->sched_priority); 3849 3850 if (running) 3851 p->sched_class->set_curr_task(rq); 3852 if (on_rq) 3853 enqueue_task(rq, p, 0); 3854 3855 check_class_changed(rq, p, prev_class, oldprio); 3856 task_rq_unlock(rq, p, &flags); 3857 3858 rt_mutex_adjust_pi(p); 3859 3860 return 0; 3861} 3862 3863/** 3864 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread. 3865 * @p: the task in question. 3866 * @policy: new policy. 3867 * @param: structure containing the new RT priority. 3868 * 3869 * NOTE that the task may be already dead. 3870 */ 3871int sched_setscheduler(struct task_struct *p, int policy, 3872 const struct sched_param *param) 3873{ 3874 return __sched_setscheduler(p, policy, param, true); 3875} 3876EXPORT_SYMBOL_GPL(sched_setscheduler); 3877 3878/** 3879 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace. 3880 * @p: the task in question. 3881 * @policy: new policy. 3882 * @param: structure containing the new RT priority. 3883 * 3884 * Just like sched_setscheduler, only don't bother checking if the 3885 * current context has permission. For example, this is needed in 3886 * stop_machine(): we create temporary high priority worker threads, 3887 * but our caller might not have that capability. 3888 */ 3889int sched_setscheduler_nocheck(struct task_struct *p, int policy, 3890 const struct sched_param *param) 3891{ 3892 return __sched_setscheduler(p, policy, param, false); 3893} 3894 3895static int 3896do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param) 3897{ 3898 struct sched_param lparam; 3899 struct task_struct *p; 3900 int retval; 3901 3902 if (!param || pid < 0) 3903 return -EINVAL; 3904 if (copy_from_user(&lparam, param, sizeof(struct sched_param))) 3905 return -EFAULT; 3906 3907 rcu_read_lock(); 3908 retval = -ESRCH; 3909 p = find_process_by_pid(pid); 3910 if (p != NULL) 3911 retval = sched_setscheduler(p, policy, &lparam); 3912 rcu_read_unlock(); 3913 3914 return retval; 3915} 3916 3917/** 3918 * sys_sched_setscheduler - set/change the scheduler policy and RT priority 3919 * @pid: the pid in question. 3920 * @policy: new policy. 3921 * @param: structure containing the new RT priority. 3922 */ 3923SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy, 3924 struct sched_param __user *, param) 3925{ 3926 /* negative values for policy are not valid */ 3927 if (policy < 0) 3928 return -EINVAL; 3929 3930 return do_sched_setscheduler(pid, policy, param); 3931} 3932 3933/** 3934 * sys_sched_setparam - set/change the RT priority of a thread 3935 * @pid: the pid in question. 3936 * @param: structure containing the new RT priority. 3937 */ 3938SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param) 3939{ 3940 return do_sched_setscheduler(pid, -1, param); 3941} 3942 3943/** 3944 * sys_sched_getscheduler - get the policy (scheduling class) of a thread 3945 * @pid: the pid in question. 3946 */ 3947SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid) 3948{ 3949 struct task_struct *p; 3950 int retval; 3951 3952 if (pid < 0) 3953 return -EINVAL; 3954 3955 retval = -ESRCH; 3956 rcu_read_lock(); 3957 p = find_process_by_pid(pid); 3958 if (p) { 3959 retval = security_task_getscheduler(p); 3960 if (!retval) 3961 retval = p->policy 3962 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0); 3963 } 3964 rcu_read_unlock(); 3965 return retval; 3966} 3967 3968/** 3969 * sys_sched_getparam - get the RT priority of a thread 3970 * @pid: the pid in question. 3971 * @param: structure containing the RT priority. 3972 */ 3973SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param) 3974{ 3975 struct sched_param lp; 3976 struct task_struct *p; 3977 int retval; 3978 3979 if (!param || pid < 0) 3980 return -EINVAL; 3981 3982 rcu_read_lock(); 3983 p = find_process_by_pid(pid); 3984 retval = -ESRCH; 3985 if (!p) 3986 goto out_unlock; 3987 3988 retval = security_task_getscheduler(p); 3989 if (retval) 3990 goto out_unlock; 3991 3992 lp.sched_priority = p->rt_priority; 3993 rcu_read_unlock(); 3994 3995 /* 3996 * This one might sleep, we cannot do it with a spinlock held ... 3997 */ 3998 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0; 3999 4000 return retval; 4001 4002out_unlock: 4003 rcu_read_unlock(); 4004 return retval; 4005} 4006 4007long sched_setaffinity(pid_t pid, const struct cpumask *in_mask) 4008{ 4009 cpumask_var_t cpus_allowed, new_mask; 4010 struct task_struct *p; 4011 int retval; 4012 4013 get_online_cpus(); 4014 rcu_read_lock(); 4015 4016 p = find_process_by_pid(pid); 4017 if (!p) { 4018 rcu_read_unlock(); 4019 put_online_cpus(); 4020 return -ESRCH; 4021 } 4022 4023 /* Prevent p going away */ 4024 get_task_struct(p); 4025 rcu_read_unlock(); 4026 4027 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) { 4028 retval = -ENOMEM; 4029 goto out_put_task; 4030 } 4031 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) { 4032 retval = -ENOMEM; 4033 goto out_free_cpus_allowed; 4034 } 4035 retval = -EPERM; 4036 if (!check_same_owner(p) && !ns_capable(task_user_ns(p), CAP_SYS_NICE)) 4037 goto out_unlock; 4038 4039 retval = security_task_setscheduler(p); 4040 if (retval) 4041 goto out_unlock; 4042 4043 cpuset_cpus_allowed(p, cpus_allowed); 4044 cpumask_and(new_mask, in_mask, cpus_allowed); 4045again: 4046 retval = set_cpus_allowed_ptr(p, new_mask); 4047 4048 if (!retval) { 4049 cpuset_cpus_allowed(p, cpus_allowed); 4050 if (!cpumask_subset(new_mask, cpus_allowed)) { 4051 /* 4052 * We must have raced with a concurrent cpuset 4053 * update. Just reset the cpus_allowed to the 4054 * cpuset's cpus_allowed 4055 */ 4056 cpumask_copy(new_mask, cpus_allowed); 4057 goto again; 4058 } 4059 } 4060out_unlock: 4061 free_cpumask_var(new_mask); 4062out_free_cpus_allowed: 4063 free_cpumask_var(cpus_allowed); 4064out_put_task: 4065 put_task_struct(p); 4066 put_online_cpus(); 4067 return retval; 4068} 4069 4070static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len, 4071 struct cpumask *new_mask) 4072{ 4073 if (len < cpumask_size()) 4074 cpumask_clear(new_mask); 4075 else if (len > cpumask_size()) 4076 len = cpumask_size(); 4077 4078 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0; 4079} 4080 4081/** 4082 * sys_sched_setaffinity - set the cpu affinity of a process 4083 * @pid: pid of the process 4084 * @len: length in bytes of the bitmask pointed to by user_mask_ptr 4085 * @user_mask_ptr: user-space pointer to the new cpu mask 4086 */ 4087SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len, 4088 unsigned long __user *, user_mask_ptr) 4089{ 4090 cpumask_var_t new_mask; 4091 int retval; 4092 4093 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) 4094 return -ENOMEM; 4095 4096 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask); 4097 if (retval == 0) 4098 retval = sched_setaffinity(pid, new_mask); 4099 free_cpumask_var(new_mask); 4100 return retval; 4101} 4102 4103long sched_getaffinity(pid_t pid, struct cpumask *mask) 4104{ 4105 struct task_struct *p; 4106 unsigned long flags; 4107 int retval; 4108 4109 get_online_cpus(); 4110 rcu_read_lock(); 4111 4112 retval = -ESRCH; 4113 p = find_process_by_pid(pid); 4114 if (!p) 4115 goto out_unlock; 4116 4117 retval = security_task_getscheduler(p); 4118 if (retval) 4119 goto out_unlock; 4120 4121 raw_spin_lock_irqsave(&p->pi_lock, flags); 4122 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask); 4123 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 4124 4125out_unlock: 4126 rcu_read_unlock(); 4127 put_online_cpus(); 4128 4129 return retval; 4130} 4131 4132/** 4133 * sys_sched_getaffinity - get the cpu affinity of a process 4134 * @pid: pid of the process 4135 * @len: length in bytes of the bitmask pointed to by user_mask_ptr 4136 * @user_mask_ptr: user-space pointer to hold the current cpu mask 4137 */ 4138SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len, 4139 unsigned long __user *, user_mask_ptr) 4140{ 4141 int ret; 4142 cpumask_var_t mask; 4143 4144 if ((len * BITS_PER_BYTE) < nr_cpu_ids) 4145 return -EINVAL; 4146 if (len & (sizeof(unsigned long)-1)) 4147 return -EINVAL; 4148 4149 if (!alloc_cpumask_var(&mask, GFP_KERNEL)) 4150 return -ENOMEM; 4151 4152 ret = sched_getaffinity(pid, mask); 4153 if (ret == 0) { 4154 size_t retlen = min_t(size_t, len, cpumask_size()); 4155 4156 if (copy_to_user(user_mask_ptr, mask, retlen)) 4157 ret = -EFAULT; 4158 else 4159 ret = retlen; 4160 } 4161 free_cpumask_var(mask); 4162 4163 return ret; 4164} 4165 4166/** 4167 * sys_sched_yield - yield the current processor to other threads. 4168 * 4169 * This function yields the current CPU to other tasks. If there are no 4170 * other threads running on this CPU then this function will return. 4171 */ 4172SYSCALL_DEFINE0(sched_yield) 4173{ 4174 struct rq *rq = this_rq_lock(); 4175 4176 schedstat_inc(rq, yld_count); 4177 current->sched_class->yield_task(rq); 4178 4179 /* 4180 * Since we are going to call schedule() anyway, there's 4181 * no need to preempt or enable interrupts: 4182 */ 4183 __release(rq->lock); 4184 spin_release(&rq->lock.dep_map, 1, _THIS_IP_); 4185 do_raw_spin_unlock(&rq->lock); 4186 sched_preempt_enable_no_resched(); 4187 4188 schedule(); 4189 4190 return 0; 4191} 4192 4193static inline int should_resched(void) 4194{ 4195 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE); 4196} 4197 4198static void __cond_resched(void) 4199{ 4200 add_preempt_count(PREEMPT_ACTIVE); 4201 __schedule(); 4202 sub_preempt_count(PREEMPT_ACTIVE); 4203} 4204 4205int __sched _cond_resched(void) 4206{ 4207 if (should_resched()) { 4208 __cond_resched(); 4209 return 1; 4210 } 4211 return 0; 4212} 4213EXPORT_SYMBOL(_cond_resched); 4214 4215/* 4216 * __cond_resched_lock() - if a reschedule is pending, drop the given lock, 4217 * call schedule, and on return reacquire the lock. 4218 * 4219 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level 4220 * operations here to prevent schedule() from being called twice (once via 4221 * spin_unlock(), once by hand). 4222 */ 4223int __cond_resched_lock(spinlock_t *lock) 4224{ 4225 int resched = should_resched(); 4226 int ret = 0; 4227 4228 lockdep_assert_held(lock); 4229 4230 if (spin_needbreak(lock) || resched) { 4231 spin_unlock(lock); 4232 if (resched) 4233 __cond_resched(); 4234 else 4235 cpu_relax(); 4236 ret = 1; 4237 spin_lock(lock); 4238 } 4239 return ret; 4240} 4241EXPORT_SYMBOL(__cond_resched_lock); 4242 4243int __sched __cond_resched_softirq(void) 4244{ 4245 BUG_ON(!in_softirq()); 4246 4247 if (should_resched()) { 4248 local_bh_enable(); 4249 __cond_resched(); 4250 local_bh_disable(); 4251 return 1; 4252 } 4253 return 0; 4254} 4255EXPORT_SYMBOL(__cond_resched_softirq); 4256 4257/** 4258 * yield - yield the current processor to other threads. 4259 * 4260 * Do not ever use this function, there's a 99% chance you're doing it wrong. 4261 * 4262 * The scheduler is at all times free to pick the calling task as the most 4263 * eligible task to run, if removing the yield() call from your code breaks 4264 * it, its already broken. 4265 * 4266 * Typical broken usage is: 4267 * 4268 * while (!event) 4269 * yield(); 4270 * 4271 * where one assumes that yield() will let 'the other' process run that will 4272 * make event true. If the current task is a SCHED_FIFO task that will never 4273 * happen. Never use yield() as a progress guarantee!! 4274 * 4275 * If you want to use yield() to wait for something, use wait_event(). 4276 * If you want to use yield() to be 'nice' for others, use cond_resched(). 4277 * If you still want to use yield(), do not! 4278 */ 4279void __sched yield(void) 4280{ 4281 set_current_state(TASK_RUNNING); 4282 sys_sched_yield(); 4283} 4284EXPORT_SYMBOL(yield); 4285 4286/** 4287 * yield_to - yield the current processor to another thread in 4288 * your thread group, or accelerate that thread toward the 4289 * processor it's on. 4290 * @p: target task 4291 * @preempt: whether task preemption is allowed or not 4292 * 4293 * It's the caller's job to ensure that the target task struct 4294 * can't go away on us before we can do any checks. 4295 * 4296 * Returns true if we indeed boosted the target task. 4297 */ 4298bool __sched yield_to(struct task_struct *p, bool preempt) 4299{ 4300 struct task_struct *curr = current; 4301 struct rq *rq, *p_rq; 4302 unsigned long flags; 4303 bool yielded = 0; 4304 4305 local_irq_save(flags); 4306 rq = this_rq(); 4307 4308again: 4309 p_rq = task_rq(p); 4310 double_rq_lock(rq, p_rq); 4311 while (task_rq(p) != p_rq) { 4312 double_rq_unlock(rq, p_rq); 4313 goto again; 4314 } 4315 4316 if (!curr->sched_class->yield_to_task) 4317 goto out; 4318 4319 if (curr->sched_class != p->sched_class) 4320 goto out; 4321 4322 if (task_running(p_rq, p) || p->state) 4323 goto out; 4324 4325 yielded = curr->sched_class->yield_to_task(rq, p, preempt); 4326 if (yielded) { 4327 schedstat_inc(rq, yld_count); 4328 /* 4329 * Make p's CPU reschedule; pick_next_entity takes care of 4330 * fairness. 4331 */ 4332 if (preempt && rq != p_rq) 4333 resched_task(p_rq->curr); 4334 } 4335 4336out: 4337 double_rq_unlock(rq, p_rq); 4338 local_irq_restore(flags); 4339 4340 if (yielded) 4341 schedule(); 4342 4343 return yielded; 4344} 4345EXPORT_SYMBOL_GPL(yield_to); 4346 4347/* 4348 * This task is about to go to sleep on IO. Increment rq->nr_iowait so 4349 * that process accounting knows that this is a task in IO wait state. 4350 */ 4351void __sched io_schedule(void) 4352{ 4353 struct rq *rq = raw_rq(); 4354 4355 delayacct_blkio_start(); 4356 atomic_inc(&rq->nr_iowait); 4357 blk_flush_plug(current); 4358 current->in_iowait = 1; 4359 schedule(); 4360 current->in_iowait = 0; 4361 atomic_dec(&rq->nr_iowait); 4362 delayacct_blkio_end(); 4363} 4364EXPORT_SYMBOL(io_schedule); 4365 4366long __sched io_schedule_timeout(long timeout) 4367{ 4368 struct rq *rq = raw_rq(); 4369 long ret; 4370 4371 delayacct_blkio_start(); 4372 atomic_inc(&rq->nr_iowait); 4373 blk_flush_plug(current); 4374 current->in_iowait = 1; 4375 ret = schedule_timeout(timeout); 4376 current->in_iowait = 0; 4377 atomic_dec(&rq->nr_iowait); 4378 delayacct_blkio_end(); 4379 return ret; 4380} 4381 4382/** 4383 * sys_sched_get_priority_max - return maximum RT priority. 4384 * @policy: scheduling class. 4385 * 4386 * this syscall returns the maximum rt_priority that can be used 4387 * by a given scheduling class. 4388 */ 4389SYSCALL_DEFINE1(sched_get_priority_max, int, policy) 4390{ 4391 int ret = -EINVAL; 4392 4393 switch (policy) { 4394 case SCHED_FIFO: 4395 case SCHED_RR: 4396 ret = MAX_USER_RT_PRIO-1; 4397 break; 4398 case SCHED_NORMAL: 4399 case SCHED_BATCH: 4400 case SCHED_IDLE: 4401 ret = 0; 4402 break; 4403 } 4404 return ret; 4405} 4406 4407/** 4408 * sys_sched_get_priority_min - return minimum RT priority. 4409 * @policy: scheduling class. 4410 * 4411 * this syscall returns the minimum rt_priority that can be used 4412 * by a given scheduling class. 4413 */ 4414SYSCALL_DEFINE1(sched_get_priority_min, int, policy) 4415{ 4416 int ret = -EINVAL; 4417 4418 switch (policy) { 4419 case SCHED_FIFO: 4420 case SCHED_RR: 4421 ret = 1; 4422 break; 4423 case SCHED_NORMAL: 4424 case SCHED_BATCH: 4425 case SCHED_IDLE: 4426 ret = 0; 4427 } 4428 return ret; 4429} 4430 4431/** 4432 * sys_sched_rr_get_interval - return the default timeslice of a process. 4433 * @pid: pid of the process. 4434 * @interval: userspace pointer to the timeslice value. 4435 * 4436 * this syscall writes the default timeslice value of a given process 4437 * into the user-space timespec buffer. A value of '0' means infinity. 4438 */ 4439SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid, 4440 struct timespec __user *, interval) 4441{ 4442 struct task_struct *p; 4443 unsigned int time_slice; 4444 unsigned long flags; 4445 struct rq *rq; 4446 int retval; 4447 struct timespec t; 4448 4449 if (pid < 0) 4450 return -EINVAL; 4451 4452 retval = -ESRCH; 4453 rcu_read_lock(); 4454 p = find_process_by_pid(pid); 4455 if (!p) 4456 goto out_unlock; 4457 4458 retval = security_task_getscheduler(p); 4459 if (retval) 4460 goto out_unlock; 4461 4462 rq = task_rq_lock(p, &flags); 4463 time_slice = p->sched_class->get_rr_interval(rq, p); 4464 task_rq_unlock(rq, p, &flags); 4465 4466 rcu_read_unlock(); 4467 jiffies_to_timespec(time_slice, &t); 4468 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0; 4469 return retval; 4470 4471out_unlock: 4472 rcu_read_unlock(); 4473 return retval; 4474} 4475 4476static const char stat_nam[] = TASK_STATE_TO_CHAR_STR; 4477 4478void sched_show_task(struct task_struct *p) 4479{ 4480 unsigned long free = 0; 4481 unsigned state; 4482 4483 state = p->state ? __ffs(p->state) + 1 : 0; 4484 printk(KERN_INFO "%-15.15s %c", p->comm, 4485 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?'); 4486#if BITS_PER_LONG == 32 4487 if (state == TASK_RUNNING) 4488 printk(KERN_CONT " running "); 4489 else 4490 printk(KERN_CONT " %08lx ", thread_saved_pc(p)); 4491#else 4492 if (state == TASK_RUNNING) 4493 printk(KERN_CONT " running task "); 4494 else 4495 printk(KERN_CONT " %016lx ", thread_saved_pc(p)); 4496#endif 4497#ifdef CONFIG_DEBUG_STACK_USAGE 4498 free = stack_not_used(p); 4499#endif 4500 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free, 4501 task_pid_nr(p), task_pid_nr(rcu_dereference(p->real_parent)), 4502 (unsigned long)task_thread_info(p)->flags); 4503 4504 show_stack(p, NULL); 4505} 4506 4507void show_state_filter(unsigned long state_filter) 4508{ 4509 struct task_struct *g, *p; 4510 4511#if BITS_PER_LONG == 32 4512 printk(KERN_INFO 4513 " task PC stack pid father\n"); 4514#else 4515 printk(KERN_INFO 4516 " task PC stack pid father\n"); 4517#endif 4518 rcu_read_lock(); 4519 do_each_thread(g, p) { 4520 /* 4521 * reset the NMI-timeout, listing all files on a slow 4522 * console might take a lot of time: 4523 */ 4524 touch_nmi_watchdog(); 4525 if (!state_filter || (p->state & state_filter)) 4526 sched_show_task(p); 4527 } while_each_thread(g, p); 4528 4529 touch_all_softlockup_watchdogs(); 4530 4531#ifdef CONFIG_SCHED_DEBUG 4532 sysrq_sched_debug_show(); 4533#endif 4534 rcu_read_unlock(); 4535 /* 4536 * Only show locks if all tasks are dumped: 4537 */ 4538 if (!state_filter) 4539 debug_show_all_locks(); 4540} 4541 4542void __cpuinit init_idle_bootup_task(struct task_struct *idle) 4543{ 4544 idle->sched_class = &idle_sched_class; 4545} 4546 4547/** 4548 * init_idle - set up an idle thread for a given CPU 4549 * @idle: task in question 4550 * @cpu: cpu the idle task belongs to 4551 * 4552 * NOTE: this function does not set the idle thread's NEED_RESCHED 4553 * flag, to make booting more robust. 4554 */ 4555void __cpuinit init_idle(struct task_struct *idle, int cpu) 4556{ 4557 struct rq *rq = cpu_rq(cpu); 4558 unsigned long flags; 4559 4560 raw_spin_lock_irqsave(&rq->lock, flags); 4561 4562 __sched_fork(idle); 4563 idle->state = TASK_RUNNING; 4564 idle->se.exec_start = sched_clock(); 4565 4566 do_set_cpus_allowed(idle, cpumask_of(cpu)); 4567 /* 4568 * We're having a chicken and egg problem, even though we are 4569 * holding rq->lock, the cpu isn't yet set to this cpu so the 4570 * lockdep check in task_group() will fail. 4571 * 4572 * Similar case to sched_fork(). / Alternatively we could 4573 * use task_rq_lock() here and obtain the other rq->lock. 4574 * 4575 * Silence PROVE_RCU 4576 */ 4577 rcu_read_lock(); 4578 __set_task_cpu(idle, cpu); 4579 rcu_read_unlock(); 4580 4581 rq->curr = rq->idle = idle; 4582#if defined(CONFIG_SMP) 4583 idle->on_cpu = 1; 4584#endif 4585 raw_spin_unlock_irqrestore(&rq->lock, flags); 4586 4587 /* Set the preempt count _outside_ the spinlocks! */ 4588 task_thread_info(idle)->preempt_count = 0; 4589 4590 /* 4591 * The idle tasks have their own, simple scheduling class: 4592 */ 4593 idle->sched_class = &idle_sched_class; 4594 ftrace_graph_init_idle_task(idle, cpu); 4595#if defined(CONFIG_SMP) 4596 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu); 4597#endif 4598} 4599 4600#ifdef CONFIG_SMP 4601void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask) 4602{ 4603 if (p->sched_class && p->sched_class->set_cpus_allowed) 4604 p->sched_class->set_cpus_allowed(p, new_mask); 4605 4606 cpumask_copy(&p->cpus_allowed, new_mask); 4607 p->nr_cpus_allowed = cpumask_weight(new_mask); 4608} 4609 4610/* 4611 * This is how migration works: 4612 * 4613 * 1) we invoke migration_cpu_stop() on the target CPU using 4614 * stop_one_cpu(). 4615 * 2) stopper starts to run (implicitly forcing the migrated thread 4616 * off the CPU) 4617 * 3) it checks whether the migrated task is still in the wrong runqueue. 4618 * 4) if it's in the wrong runqueue then the migration thread removes 4619 * it and puts it into the right queue. 4620 * 5) stopper completes and stop_one_cpu() returns and the migration 4621 * is done. 4622 */ 4623 4624/* 4625 * Change a given task's CPU affinity. Migrate the thread to a 4626 * proper CPU and schedule it away if the CPU it's executing on 4627 * is removed from the allowed bitmask. 4628 * 4629 * NOTE: the caller must have a valid reference to the task, the 4630 * task must not exit() & deallocate itself prematurely. The 4631 * call is not atomic; no spinlocks may be held. 4632 */ 4633int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask) 4634{ 4635 unsigned long flags; 4636 struct rq *rq; 4637 unsigned int dest_cpu; 4638 int ret = 0; 4639 4640 rq = task_rq_lock(p, &flags); 4641 4642 if (cpumask_equal(&p->cpus_allowed, new_mask)) 4643 goto out; 4644 4645 if (!cpumask_intersects(new_mask, cpu_active_mask)) { 4646 ret = -EINVAL; 4647 goto out; 4648 } 4649 4650 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current)) { 4651 ret = -EINVAL; 4652 goto out; 4653 } 4654 4655 do_set_cpus_allowed(p, new_mask); 4656 4657 /* Can the task run on the task's current CPU? If so, we're done */ 4658 if (cpumask_test_cpu(task_cpu(p), new_mask)) 4659 goto out; 4660 4661 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask); 4662 if (p->on_rq) { 4663 struct migration_arg arg = { p, dest_cpu }; 4664 /* Need help from migration thread: drop lock and wait. */ 4665 task_rq_unlock(rq, p, &flags); 4666 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg); 4667 tlb_migrate_finish(p->mm); 4668 return 0; 4669 } 4670out: 4671 task_rq_unlock(rq, p, &flags); 4672 4673 return ret; 4674} 4675EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr); 4676 4677/* 4678 * Move (not current) task off this cpu, onto dest cpu. We're doing 4679 * this because either it can't run here any more (set_cpus_allowed() 4680 * away from this CPU, or CPU going down), or because we're 4681 * attempting to rebalance this task on exec (sched_exec). 4682 * 4683 * So we race with normal scheduler movements, but that's OK, as long 4684 * as the task is no longer on this CPU. 4685 * 4686 * Returns non-zero if task was successfully migrated. 4687 */ 4688static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu) 4689{ 4690 struct rq *rq_dest, *rq_src; 4691 int ret = 0; 4692 4693 if (unlikely(!cpu_active(dest_cpu))) 4694 return ret; 4695 4696 rq_src = cpu_rq(src_cpu); 4697 rq_dest = cpu_rq(dest_cpu); 4698 4699 raw_spin_lock(&p->pi_lock); 4700 double_rq_lock(rq_src, rq_dest); 4701 /* Already moved. */ 4702 if (task_cpu(p) != src_cpu) 4703 goto done; 4704 /* Affinity changed (again). */ 4705 if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p))) 4706 goto fail; 4707 4708 /* 4709 * If we're not on a rq, the next wake-up will ensure we're 4710 * placed properly. 4711 */ 4712 if (p->on_rq) { 4713 dequeue_task(rq_src, p, 0); 4714 set_task_cpu(p, dest_cpu); 4715 enqueue_task(rq_dest, p, 0); 4716 check_preempt_curr(rq_dest, p, 0); 4717 } 4718done: 4719 ret = 1; 4720fail: 4721 double_rq_unlock(rq_src, rq_dest); 4722 raw_spin_unlock(&p->pi_lock); 4723 return ret; 4724} 4725 4726/* 4727 * migration_cpu_stop - this will be executed by a highprio stopper thread 4728 * and performs thread migration by bumping thread off CPU then 4729 * 'pushing' onto another runqueue. 4730 */ 4731static int migration_cpu_stop(void *data) 4732{ 4733 struct migration_arg *arg = data; 4734 4735 /* 4736 * The original target cpu might have gone down and we might 4737 * be on another cpu but it doesn't matter. 4738 */ 4739 local_irq_disable(); 4740 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu); 4741 local_irq_enable(); 4742 return 0; 4743} 4744 4745#ifdef CONFIG_HOTPLUG_CPU 4746 4747/* 4748 * Ensures that the idle task is using init_mm right before its cpu goes 4749 * offline. 4750 */ 4751void idle_task_exit(void) 4752{ 4753 struct mm_struct *mm = current->active_mm; 4754 4755 BUG_ON(cpu_online(smp_processor_id())); 4756 4757 if (mm != &init_mm) 4758 switch_mm(mm, &init_mm, current); 4759 mmdrop(mm); 4760} 4761 4762/* 4763 * Since this CPU is going 'away' for a while, fold any nr_active delta 4764 * we might have. Assumes we're called after migrate_tasks() so that the 4765 * nr_active count is stable. 4766 * 4767 * Also see the comment "Global load-average calculations". 4768 */ 4769static void calc_load_migrate(struct rq *rq) 4770{ 4771 long delta = calc_load_fold_active(rq); 4772 if (delta) 4773 atomic_long_add(delta, &calc_load_tasks); 4774} 4775 4776/* 4777 * Migrate all tasks from the rq, sleeping tasks will be migrated by 4778 * try_to_wake_up()->select_task_rq(). 4779 * 4780 * Called with rq->lock held even though we'er in stop_machine() and 4781 * there's no concurrency possible, we hold the required locks anyway 4782 * because of lock validation efforts. 4783 */ 4784static void migrate_tasks(unsigned int dead_cpu) 4785{ 4786 struct rq *rq = cpu_rq(dead_cpu); 4787 struct task_struct *next, *stop = rq->stop; 4788 int dest_cpu; 4789 4790 /* 4791 * Fudge the rq selection such that the below task selection loop 4792 * doesn't get stuck on the currently eligible stop task. 4793 * 4794 * We're currently inside stop_machine() and the rq is either stuck 4795 * in the stop_machine_cpu_stop() loop, or we're executing this code, 4796 * either way we should never end up calling schedule() until we're 4797 * done here. 4798 */ 4799 rq->stop = NULL; 4800 4801 for ( ; ; ) { 4802 /* 4803 * There's this thread running, bail when that's the only 4804 * remaining thread. 4805 */ 4806 if (rq->nr_running == 1) 4807 break; 4808 4809 next = pick_next_task(rq); 4810 BUG_ON(!next); 4811 next->sched_class->put_prev_task(rq, next); 4812 4813 /* Find suitable destination for @next, with force if needed. */ 4814 dest_cpu = select_fallback_rq(dead_cpu, next); 4815 raw_spin_unlock(&rq->lock); 4816 4817 __migrate_task(next, dead_cpu, dest_cpu); 4818 4819 raw_spin_lock(&rq->lock); 4820 } 4821 4822 rq->stop = stop; 4823} 4824 4825#endif /* CONFIG_HOTPLUG_CPU */ 4826 4827#if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL) 4828 4829static struct ctl_table sd_ctl_dir[] = { 4830 { 4831 .procname = "sched_domain", 4832 .mode = 0555, 4833 }, 4834 {} 4835}; 4836 4837static struct ctl_table sd_ctl_root[] = { 4838 { 4839 .procname = "kernel", 4840 .mode = 0555, 4841 .child = sd_ctl_dir, 4842 }, 4843 {} 4844}; 4845 4846static struct ctl_table *sd_alloc_ctl_entry(int n) 4847{ 4848 struct ctl_table *entry = 4849 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL); 4850 4851 return entry; 4852} 4853 4854static void sd_free_ctl_entry(struct ctl_table **tablep) 4855{ 4856 struct ctl_table *entry; 4857 4858 /* 4859 * In the intermediate directories, both the child directory and 4860 * procname are dynamically allocated and could fail but the mode 4861 * will always be set. In the lowest directory the names are 4862 * static strings and all have proc handlers. 4863 */ 4864 for (entry = *tablep; entry->mode; entry++) { 4865 if (entry->child) 4866 sd_free_ctl_entry(&entry->child); 4867 if (entry->proc_handler == NULL) 4868 kfree(entry->procname); 4869 } 4870 4871 kfree(*tablep); 4872 *tablep = NULL; 4873} 4874 4875static int min_load_idx = 0; 4876static int max_load_idx = CPU_LOAD_IDX_MAX; 4877 4878static void 4879set_table_entry(struct ctl_table *entry, 4880 const char *procname, void *data, int maxlen, 4881 umode_t mode, proc_handler *proc_handler, 4882 bool load_idx) 4883{ 4884 entry->procname = procname; 4885 entry->data = data; 4886 entry->maxlen = maxlen; 4887 entry->mode = mode; 4888 entry->proc_handler = proc_handler; 4889 4890 if (load_idx) { 4891 entry->extra1 = &min_load_idx; 4892 entry->extra2 = &max_load_idx; 4893 } 4894} 4895 4896static struct ctl_table * 4897sd_alloc_ctl_domain_table(struct sched_domain *sd) 4898{ 4899 struct ctl_table *table = sd_alloc_ctl_entry(13); 4900 4901 if (table == NULL) 4902 return NULL; 4903 4904 set_table_entry(&table[0], "min_interval", &sd->min_interval, 4905 sizeof(long), 0644, proc_doulongvec_minmax, false); 4906 set_table_entry(&table[1], "max_interval", &sd->max_interval, 4907 sizeof(long), 0644, proc_doulongvec_minmax, false); 4908 set_table_entry(&table[2], "busy_idx", &sd->busy_idx, 4909 sizeof(int), 0644, proc_dointvec_minmax, true); 4910 set_table_entry(&table[3], "idle_idx", &sd->idle_idx, 4911 sizeof(int), 0644, proc_dointvec_minmax, true); 4912 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx, 4913 sizeof(int), 0644, proc_dointvec_minmax, true); 4914 set_table_entry(&table[5], "wake_idx", &sd->wake_idx, 4915 sizeof(int), 0644, proc_dointvec_minmax, true); 4916 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx, 4917 sizeof(int), 0644, proc_dointvec_minmax, true); 4918 set_table_entry(&table[7], "busy_factor", &sd->busy_factor, 4919 sizeof(int), 0644, proc_dointvec_minmax, false); 4920 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct, 4921 sizeof(int), 0644, proc_dointvec_minmax, false); 4922 set_table_entry(&table[9], "cache_nice_tries", 4923 &sd->cache_nice_tries, 4924 sizeof(int), 0644, proc_dointvec_minmax, false); 4925 set_table_entry(&table[10], "flags", &sd->flags, 4926 sizeof(int), 0644, proc_dointvec_minmax, false); 4927 set_table_entry(&table[11], "name", sd->name, 4928 CORENAME_MAX_SIZE, 0444, proc_dostring, false); 4929 /* &table[12] is terminator */ 4930 4931 return table; 4932} 4933 4934static ctl_table *sd_alloc_ctl_cpu_table(int cpu) 4935{ 4936 struct ctl_table *entry, *table; 4937 struct sched_domain *sd; 4938 int domain_num = 0, i; 4939 char buf[32]; 4940 4941 for_each_domain(cpu, sd) 4942 domain_num++; 4943 entry = table = sd_alloc_ctl_entry(domain_num + 1); 4944 if (table == NULL) 4945 return NULL; 4946 4947 i = 0; 4948 for_each_domain(cpu, sd) { 4949 snprintf(buf, 32, "domain%d", i); 4950 entry->procname = kstrdup(buf, GFP_KERNEL); 4951 entry->mode = 0555; 4952 entry->child = sd_alloc_ctl_domain_table(sd); 4953 entry++; 4954 i++; 4955 } 4956 return table; 4957} 4958 4959static struct ctl_table_header *sd_sysctl_header; 4960static void register_sched_domain_sysctl(void) 4961{ 4962 int i, cpu_num = num_possible_cpus(); 4963 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1); 4964 char buf[32]; 4965 4966 WARN_ON(sd_ctl_dir[0].child); 4967 sd_ctl_dir[0].child = entry; 4968 4969 if (entry == NULL) 4970 return; 4971 4972 for_each_possible_cpu(i) { 4973 snprintf(buf, 32, "cpu%d", i); 4974 entry->procname = kstrdup(buf, GFP_KERNEL); 4975 entry->mode = 0555; 4976 entry->child = sd_alloc_ctl_cpu_table(i); 4977 entry++; 4978 } 4979 4980 WARN_ON(sd_sysctl_header); 4981 sd_sysctl_header = register_sysctl_table(sd_ctl_root); 4982} 4983 4984/* may be called multiple times per register */ 4985static void unregister_sched_domain_sysctl(void) 4986{ 4987 if (sd_sysctl_header) 4988 unregister_sysctl_table(sd_sysctl_header); 4989 sd_sysctl_header = NULL; 4990 if (sd_ctl_dir[0].child) 4991 sd_free_ctl_entry(&sd_ctl_dir[0].child); 4992} 4993#else 4994static void register_sched_domain_sysctl(void) 4995{ 4996} 4997static void unregister_sched_domain_sysctl(void) 4998{ 4999} 5000#endif 5001 5002static void set_rq_online(struct rq *rq) 5003{ 5004 if (!rq->online) { 5005 const struct sched_class *class; 5006 5007 cpumask_set_cpu(rq->cpu, rq->rd->online); 5008 rq->online = 1; 5009 5010 for_each_class(class) { 5011 if (class->rq_online) 5012 class->rq_online(rq); 5013 } 5014 } 5015} 5016 5017static void set_rq_offline(struct rq *rq) 5018{ 5019 if (rq->online) { 5020 const struct sched_class *class; 5021 5022 for_each_class(class) { 5023 if (class->rq_offline) 5024 class->rq_offline(rq); 5025 } 5026 5027 cpumask_clear_cpu(rq->cpu, rq->rd->online); 5028 rq->online = 0; 5029 } 5030} 5031 5032/* 5033 * migration_call - callback that gets triggered when a CPU is added. 5034 * Here we can start up the necessary migration thread for the new CPU. 5035 */ 5036static int __cpuinit 5037migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu) 5038{ 5039 int cpu = (long)hcpu; 5040 unsigned long flags; 5041 struct rq *rq = cpu_rq(cpu); 5042 5043 switch (action & ~CPU_TASKS_FROZEN) { 5044 5045 case CPU_UP_PREPARE: 5046 rq->calc_load_update = calc_load_update; 5047 break; 5048 5049 case CPU_ONLINE: 5050 /* Update our root-domain */ 5051 raw_spin_lock_irqsave(&rq->lock, flags); 5052 if (rq->rd) { 5053 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span)); 5054 5055 set_rq_online(rq); 5056 } 5057 raw_spin_unlock_irqrestore(&rq->lock, flags); 5058 break; 5059 5060#ifdef CONFIG_HOTPLUG_CPU 5061 case CPU_DYING: 5062 sched_ttwu_pending(); 5063 /* Update our root-domain */ 5064 raw_spin_lock_irqsave(&rq->lock, flags); 5065 if (rq->rd) { 5066 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span)); 5067 set_rq_offline(rq); 5068 } 5069 migrate_tasks(cpu); 5070 BUG_ON(rq->nr_running != 1); /* the migration thread */ 5071 raw_spin_unlock_irqrestore(&rq->lock, flags); 5072 break; 5073 5074 case CPU_DEAD: 5075 calc_load_migrate(rq); 5076 break; 5077#endif 5078 } 5079 5080 update_max_interval(); 5081 5082 return NOTIFY_OK; 5083} 5084 5085/* 5086 * Register at high priority so that task migration (migrate_all_tasks) 5087 * happens before everything else. This has to be lower priority than 5088 * the notifier in the perf_event subsystem, though. 5089 */ 5090static struct notifier_block __cpuinitdata migration_notifier = { 5091 .notifier_call = migration_call, 5092 .priority = CPU_PRI_MIGRATION, 5093}; 5094 5095static int __cpuinit sched_cpu_active(struct notifier_block *nfb, 5096 unsigned long action, void *hcpu) 5097{ 5098 switch (action & ~CPU_TASKS_FROZEN) { 5099 case CPU_STARTING: 5100 case CPU_DOWN_FAILED: 5101 set_cpu_active((long)hcpu, true); 5102 return NOTIFY_OK; 5103 default: 5104 return NOTIFY_DONE; 5105 } 5106} 5107 5108static int __cpuinit sched_cpu_inactive(struct notifier_block *nfb, 5109 unsigned long action, void *hcpu) 5110{ 5111 switch (action & ~CPU_TASKS_FROZEN) { 5112 case CPU_DOWN_PREPARE: 5113 set_cpu_active((long)hcpu, false); 5114 return NOTIFY_OK; 5115 default: 5116 return NOTIFY_DONE; 5117 } 5118} 5119 5120static int __init migration_init(void) 5121{ 5122 void *cpu = (void *)(long)smp_processor_id(); 5123 int err; 5124 5125 /* Initialize migration for the boot CPU */ 5126 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu); 5127 BUG_ON(err == NOTIFY_BAD); 5128 migration_call(&migration_notifier, CPU_ONLINE, cpu); 5129 register_cpu_notifier(&migration_notifier); 5130 5131 /* Register cpu active notifiers */ 5132 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE); 5133 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE); 5134 5135 return 0; 5136} 5137early_initcall(migration_init); 5138#endif 5139 5140#ifdef CONFIG_SMP 5141 5142static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */ 5143 5144#ifdef CONFIG_SCHED_DEBUG 5145 5146static __read_mostly int sched_debug_enabled; 5147 5148static int __init sched_debug_setup(char *str) 5149{ 5150 sched_debug_enabled = 1; 5151 5152 return 0; 5153} 5154early_param("sched_debug", sched_debug_setup); 5155 5156static inline bool sched_debug(void) 5157{ 5158 return sched_debug_enabled; 5159} 5160 5161static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level, 5162 struct cpumask *groupmask) 5163{ 5164 struct sched_group *group = sd->groups; 5165 char str[256]; 5166 5167 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd)); 5168 cpumask_clear(groupmask); 5169 5170 printk(KERN_DEBUG "%*s domain %d: ", level, "", level); 5171 5172 if (!(sd->flags & SD_LOAD_BALANCE)) { 5173 printk("does not load-balance\n"); 5174 if (sd->parent) 5175 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain" 5176 " has parent"); 5177 return -1; 5178 } 5179 5180 printk(KERN_CONT "span %s level %s\n", str, sd->name); 5181 5182 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) { 5183 printk(KERN_ERR "ERROR: domain->span does not contain " 5184 "CPU%d\n", cpu); 5185 } 5186 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) { 5187 printk(KERN_ERR "ERROR: domain->groups does not contain" 5188 " CPU%d\n", cpu); 5189 } 5190 5191 printk(KERN_DEBUG "%*s groups:", level + 1, ""); 5192 do { 5193 if (!group) { 5194 printk("\n"); 5195 printk(KERN_ERR "ERROR: group is NULL\n"); 5196 break; 5197 } 5198 5199 /* 5200 * Even though we initialize ->power to something semi-sane, 5201 * we leave power_orig unset. This allows us to detect if 5202 * domain iteration is still funny without causing /0 traps. 5203 */ 5204 if (!group->sgp->power_orig) { 5205 printk(KERN_CONT "\n"); 5206 printk(KERN_ERR "ERROR: domain->cpu_power not " 5207 "set\n"); 5208 break; 5209 } 5210 5211 if (!cpumask_weight(sched_group_cpus(group))) { 5212 printk(KERN_CONT "\n"); 5213 printk(KERN_ERR "ERROR: empty group\n"); 5214 break; 5215 } 5216 5217 if (!(sd->flags & SD_OVERLAP) && 5218 cpumask_intersects(groupmask, sched_group_cpus(group))) { 5219 printk(KERN_CONT "\n"); 5220 printk(KERN_ERR "ERROR: repeated CPUs\n"); 5221 break; 5222 } 5223 5224 cpumask_or(groupmask, groupmask, sched_group_cpus(group)); 5225 5226 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group)); 5227 5228 printk(KERN_CONT " %s", str); 5229 if (group->sgp->power != SCHED_POWER_SCALE) { 5230 printk(KERN_CONT " (cpu_power = %d)", 5231 group->sgp->power); 5232 } 5233 5234 group = group->next; 5235 } while (group != sd->groups); 5236 printk(KERN_CONT "\n"); 5237 5238 if (!cpumask_equal(sched_domain_span(sd), groupmask)) 5239 printk(KERN_ERR "ERROR: groups don't span domain->span\n"); 5240 5241 if (sd->parent && 5242 !cpumask_subset(groupmask, sched_domain_span(sd->parent))) 5243 printk(KERN_ERR "ERROR: parent span is not a superset " 5244 "of domain->span\n"); 5245 return 0; 5246} 5247 5248static void sched_domain_debug(struct sched_domain *sd, int cpu) 5249{ 5250 int level = 0; 5251 5252 if (!sched_debug_enabled) 5253 return; 5254 5255 if (!sd) { 5256 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu); 5257 return; 5258 } 5259 5260 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu); 5261 5262 for (;;) { 5263 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask)) 5264 break; 5265 level++; 5266 sd = sd->parent; 5267 if (!sd) 5268 break; 5269 } 5270} 5271#else /* !CONFIG_SCHED_DEBUG */ 5272# define sched_domain_debug(sd, cpu) do { } while (0) 5273static inline bool sched_debug(void) 5274{ 5275 return false; 5276} 5277#endif /* CONFIG_SCHED_DEBUG */ 5278 5279static int sd_degenerate(struct sched_domain *sd) 5280{ 5281 if (cpumask_weight(sched_domain_span(sd)) == 1) 5282 return 1; 5283 5284 /* Following flags need at least 2 groups */ 5285 if (sd->flags & (SD_LOAD_BALANCE | 5286 SD_BALANCE_NEWIDLE | 5287 SD_BALANCE_FORK | 5288 SD_BALANCE_EXEC | 5289 SD_SHARE_CPUPOWER | 5290 SD_SHARE_PKG_RESOURCES)) { 5291 if (sd->groups != sd->groups->next) 5292 return 0; 5293 } 5294 5295 /* Following flags don't use groups */ 5296 if (sd->flags & (SD_WAKE_AFFINE)) 5297 return 0; 5298 5299 return 1; 5300} 5301 5302static int 5303sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent) 5304{ 5305 unsigned long cflags = sd->flags, pflags = parent->flags; 5306 5307 if (sd_degenerate(parent)) 5308 return 1; 5309 5310 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent))) 5311 return 0; 5312 5313 /* Flags needing groups don't count if only 1 group in parent */ 5314 if (parent->groups == parent->groups->next) { 5315 pflags &= ~(SD_LOAD_BALANCE | 5316 SD_BALANCE_NEWIDLE | 5317 SD_BALANCE_FORK | 5318 SD_BALANCE_EXEC | 5319 SD_SHARE_CPUPOWER | 5320 SD_SHARE_PKG_RESOURCES); 5321 if (nr_node_ids == 1) 5322 pflags &= ~SD_SERIALIZE; 5323 } 5324 if (~cflags & pflags) 5325 return 0; 5326 5327 return 1; 5328} 5329 5330static void free_rootdomain(struct rcu_head *rcu) 5331{ 5332 struct root_domain *rd = container_of(rcu, struct root_domain, rcu); 5333 5334 cpupri_cleanup(&rd->cpupri); 5335 free_cpumask_var(rd->rto_mask); 5336 free_cpumask_var(rd->online); 5337 free_cpumask_var(rd->span); 5338 kfree(rd); 5339} 5340 5341static void rq_attach_root(struct rq *rq, struct root_domain *rd) 5342{ 5343 struct root_domain *old_rd = NULL; 5344 unsigned long flags; 5345 5346 raw_spin_lock_irqsave(&rq->lock, flags); 5347 5348 if (rq->rd) { 5349 old_rd = rq->rd; 5350 5351 if (cpumask_test_cpu(rq->cpu, old_rd->online)) 5352 set_rq_offline(rq); 5353 5354 cpumask_clear_cpu(rq->cpu, old_rd->span); 5355 5356 /* 5357 * If we dont want to free the old_rt yet then 5358 * set old_rd to NULL to skip the freeing later 5359 * in this function: 5360 */ 5361 if (!atomic_dec_and_test(&old_rd->refcount)) 5362 old_rd = NULL; 5363 } 5364 5365 atomic_inc(&rd->refcount); 5366 rq->rd = rd; 5367 5368 cpumask_set_cpu(rq->cpu, rd->span); 5369 if (cpumask_test_cpu(rq->cpu, cpu_active_mask)) 5370 set_rq_online(rq); 5371 5372 raw_spin_unlock_irqrestore(&rq->lock, flags); 5373 5374 if (old_rd) 5375 call_rcu_sched(&old_rd->rcu, free_rootdomain); 5376} 5377 5378static int init_rootdomain(struct root_domain *rd) 5379{ 5380 memset(rd, 0, sizeof(*rd)); 5381 5382 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL)) 5383 goto out; 5384 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL)) 5385 goto free_span; 5386 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL)) 5387 goto free_online; 5388 5389 if (cpupri_init(&rd->cpupri) != 0) 5390 goto free_rto_mask; 5391 return 0; 5392 5393free_rto_mask: 5394 free_cpumask_var(rd->rto_mask); 5395free_online: 5396 free_cpumask_var(rd->online); 5397free_span: 5398 free_cpumask_var(rd->span); 5399out: 5400 return -ENOMEM; 5401} 5402 5403/* 5404 * By default the system creates a single root-domain with all cpus as 5405 * members (mimicking the global state we have today). 5406 */ 5407struct root_domain def_root_domain; 5408 5409static void init_defrootdomain(void) 5410{ 5411 init_rootdomain(&def_root_domain); 5412 5413 atomic_set(&def_root_domain.refcount, 1); 5414} 5415 5416static struct root_domain *alloc_rootdomain(void) 5417{ 5418 struct root_domain *rd; 5419 5420 rd = kmalloc(sizeof(*rd), GFP_KERNEL); 5421 if (!rd) 5422 return NULL; 5423 5424 if (init_rootdomain(rd) != 0) { 5425 kfree(rd); 5426 return NULL; 5427 } 5428 5429 return rd; 5430} 5431 5432static void free_sched_groups(struct sched_group *sg, int free_sgp) 5433{ 5434 struct sched_group *tmp, *first; 5435 5436 if (!sg) 5437 return; 5438 5439 first = sg; 5440 do { 5441 tmp = sg->next; 5442 5443 if (free_sgp && atomic_dec_and_test(&sg->sgp->ref)) 5444 kfree(sg->sgp); 5445 5446 kfree(sg); 5447 sg = tmp; 5448 } while (sg != first); 5449} 5450 5451static void free_sched_domain(struct rcu_head *rcu) 5452{ 5453 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu); 5454 5455 /* 5456 * If its an overlapping domain it has private groups, iterate and 5457 * nuke them all. 5458 */ 5459 if (sd->flags & SD_OVERLAP) { 5460 free_sched_groups(sd->groups, 1); 5461 } else if (atomic_dec_and_test(&sd->groups->ref)) { 5462 kfree(sd->groups->sgp); 5463 kfree(sd->groups); 5464 } 5465 kfree(sd); 5466} 5467 5468static void destroy_sched_domain(struct sched_domain *sd, int cpu) 5469{ 5470 call_rcu(&sd->rcu, free_sched_domain); 5471} 5472 5473static void destroy_sched_domains(struct sched_domain *sd, int cpu) 5474{ 5475 for (; sd; sd = sd->parent) 5476 destroy_sched_domain(sd, cpu); 5477} 5478 5479/* 5480 * Keep a special pointer to the highest sched_domain that has 5481 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this 5482 * allows us to avoid some pointer chasing select_idle_sibling(). 5483 * 5484 * Also keep a unique ID per domain (we use the first cpu number in 5485 * the cpumask of the domain), this allows us to quickly tell if 5486 * two cpus are in the same cache domain, see cpus_share_cache(). 5487 */ 5488DEFINE_PER_CPU(struct sched_domain *, sd_llc); 5489DEFINE_PER_CPU(int, sd_llc_id); 5490 5491static void update_top_cache_domain(int cpu) 5492{ 5493 struct sched_domain *sd; 5494 int id = cpu; 5495 5496 sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES); 5497 if (sd) 5498 id = cpumask_first(sched_domain_span(sd)); 5499 5500 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd); 5501 per_cpu(sd_llc_id, cpu) = id; 5502} 5503 5504/* 5505 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must 5506 * hold the hotplug lock. 5507 */ 5508static void 5509cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu) 5510{ 5511 struct rq *rq = cpu_rq(cpu); 5512 struct sched_domain *tmp; 5513 5514 /* Remove the sched domains which do not contribute to scheduling. */ 5515 for (tmp = sd; tmp; ) { 5516 struct sched_domain *parent = tmp->parent; 5517 if (!parent) 5518 break; 5519 5520 if (sd_parent_degenerate(tmp, parent)) { 5521 tmp->parent = parent->parent; 5522 if (parent->parent) 5523 parent->parent->child = tmp; 5524 destroy_sched_domain(parent, cpu); 5525 } else 5526 tmp = tmp->parent; 5527 } 5528 5529 if (sd && sd_degenerate(sd)) { 5530 tmp = sd; 5531 sd = sd->parent; 5532 destroy_sched_domain(tmp, cpu); 5533 if (sd) 5534 sd->child = NULL; 5535 } 5536 5537 sched_domain_debug(sd, cpu); 5538 5539 rq_attach_root(rq, rd); 5540 tmp = rq->sd; 5541 rcu_assign_pointer(rq->sd, sd); 5542 destroy_sched_domains(tmp, cpu); 5543 5544 update_top_cache_domain(cpu); 5545} 5546 5547/* cpus with isolated domains */ 5548static cpumask_var_t cpu_isolated_map; 5549 5550/* Setup the mask of cpus configured for isolated domains */ 5551static int __init isolated_cpu_setup(char *str) 5552{ 5553 alloc_bootmem_cpumask_var(&cpu_isolated_map); 5554 cpulist_parse(str, cpu_isolated_map); 5555 return 1; 5556} 5557 5558__setup("isolcpus=", isolated_cpu_setup); 5559 5560static const struct cpumask *cpu_cpu_mask(int cpu) 5561{ 5562 return cpumask_of_node(cpu_to_node(cpu)); 5563} 5564 5565struct sd_data { 5566 struct sched_domain **__percpu sd; 5567 struct sched_group **__percpu sg; 5568 struct sched_group_power **__percpu sgp; 5569}; 5570 5571struct s_data { 5572 struct sched_domain ** __percpu sd; 5573 struct root_domain *rd; 5574}; 5575 5576enum s_alloc { 5577 sa_rootdomain, 5578 sa_sd, 5579 sa_sd_storage, 5580 sa_none, 5581}; 5582 5583struct sched_domain_topology_level; 5584 5585typedef struct sched_domain *(*sched_domain_init_f)(struct sched_domain_topology_level *tl, int cpu); 5586typedef const struct cpumask *(*sched_domain_mask_f)(int cpu); 5587 5588#define SDTL_OVERLAP 0x01 5589 5590struct sched_domain_topology_level { 5591 sched_domain_init_f init; 5592 sched_domain_mask_f mask; 5593 int flags; 5594 int numa_level; 5595 struct sd_data data; 5596}; 5597 5598/* 5599 * Build an iteration mask that can exclude certain CPUs from the upwards 5600 * domain traversal. 5601 * 5602 * Asymmetric node setups can result in situations where the domain tree is of 5603 * unequal depth, make sure to skip domains that already cover the entire 5604 * range. 5605 * 5606 * In that case build_sched_domains() will have terminated the iteration early 5607 * and our sibling sd spans will be empty. Domains should always include the 5608 * cpu they're built on, so check that. 5609 * 5610 */ 5611static void build_group_mask(struct sched_domain *sd, struct sched_group *sg) 5612{ 5613 const struct cpumask *span = sched_domain_span(sd); 5614 struct sd_data *sdd = sd->private; 5615 struct sched_domain *sibling; 5616 int i; 5617 5618 for_each_cpu(i, span) { 5619 sibling = *per_cpu_ptr(sdd->sd, i); 5620 if (!cpumask_test_cpu(i, sched_domain_span(sibling))) 5621 continue; 5622 5623 cpumask_set_cpu(i, sched_group_mask(sg)); 5624 } 5625} 5626 5627/* 5628 * Return the canonical balance cpu for this group, this is the first cpu 5629 * of this group that's also in the iteration mask. 5630 */ 5631int group_balance_cpu(struct sched_group *sg) 5632{ 5633 return cpumask_first_and(sched_group_cpus(sg), sched_group_mask(sg)); 5634} 5635 5636static int 5637build_overlap_sched_groups(struct sched_domain *sd, int cpu) 5638{ 5639 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg; 5640 const struct cpumask *span = sched_domain_span(sd); 5641 struct cpumask *covered = sched_domains_tmpmask; 5642 struct sd_data *sdd = sd->private; 5643 struct sched_domain *child; 5644 int i; 5645 5646 cpumask_clear(covered); 5647 5648 for_each_cpu(i, span) { 5649 struct cpumask *sg_span; 5650 5651 if (cpumask_test_cpu(i, covered)) 5652 continue; 5653 5654 child = *per_cpu_ptr(sdd->sd, i); 5655 5656 /* See the comment near build_group_mask(). */ 5657 if (!cpumask_test_cpu(i, sched_domain_span(child))) 5658 continue; 5659 5660 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(), 5661 GFP_KERNEL, cpu_to_node(cpu)); 5662 5663 if (!sg) 5664 goto fail; 5665 5666 sg_span = sched_group_cpus(sg); 5667 if (child->child) { 5668 child = child->child; 5669 cpumask_copy(sg_span, sched_domain_span(child)); 5670 } else 5671 cpumask_set_cpu(i, sg_span); 5672 5673 cpumask_or(covered, covered, sg_span); 5674 5675 sg->sgp = *per_cpu_ptr(sdd->sgp, i); 5676 if (atomic_inc_return(&sg->sgp->ref) == 1) 5677 build_group_mask(sd, sg); 5678 5679 /* 5680 * Initialize sgp->power such that even if we mess up the 5681 * domains and no possible iteration will get us here, we won't 5682 * die on a /0 trap. 5683 */ 5684 sg->sgp->power = SCHED_POWER_SCALE * cpumask_weight(sg_span); 5685 5686 /* 5687 * Make sure the first group of this domain contains the 5688 * canonical balance cpu. Otherwise the sched_domain iteration 5689 * breaks. See update_sg_lb_stats(). 5690 */ 5691 if ((!groups && cpumask_test_cpu(cpu, sg_span)) || 5692 group_balance_cpu(sg) == cpu) 5693 groups = sg; 5694 5695 if (!first) 5696 first = sg; 5697 if (last) 5698 last->next = sg; 5699 last = sg; 5700 last->next = first; 5701 } 5702 sd->groups = groups; 5703 5704 return 0; 5705 5706fail: 5707 free_sched_groups(first, 0); 5708 5709 return -ENOMEM; 5710} 5711 5712static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg) 5713{ 5714 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu); 5715 struct sched_domain *child = sd->child; 5716 5717 if (child) 5718 cpu = cpumask_first(sched_domain_span(child)); 5719 5720 if (sg) { 5721 *sg = *per_cpu_ptr(sdd->sg, cpu); 5722 (*sg)->sgp = *per_cpu_ptr(sdd->sgp, cpu); 5723 atomic_set(&(*sg)->sgp->ref, 1); /* for claim_allocations */ 5724 } 5725 5726 return cpu; 5727} 5728 5729/* 5730 * build_sched_groups will build a circular linked list of the groups 5731 * covered by the given span, and will set each group's ->cpumask correctly, 5732 * and ->cpu_power to 0. 5733 * 5734 * Assumes the sched_domain tree is fully constructed 5735 */ 5736static int 5737build_sched_groups(struct sched_domain *sd, int cpu) 5738{ 5739 struct sched_group *first = NULL, *last = NULL; 5740 struct sd_data *sdd = sd->private; 5741 const struct cpumask *span = sched_domain_span(sd); 5742 struct cpumask *covered; 5743 int i; 5744 5745 get_group(cpu, sdd, &sd->groups); 5746 atomic_inc(&sd->groups->ref); 5747 5748 if (cpu != cpumask_first(sched_domain_span(sd))) 5749 return 0; 5750 5751 lockdep_assert_held(&sched_domains_mutex); 5752 covered = sched_domains_tmpmask; 5753 5754 cpumask_clear(covered); 5755 5756 for_each_cpu(i, span) { 5757 struct sched_group *sg; 5758 int group = get_group(i, sdd, &sg); 5759 int j; 5760 5761 if (cpumask_test_cpu(i, covered)) 5762 continue; 5763 5764 cpumask_clear(sched_group_cpus(sg)); 5765 sg->sgp->power = 0; 5766 cpumask_setall(sched_group_mask(sg)); 5767 5768 for_each_cpu(j, span) { 5769 if (get_group(j, sdd, NULL) != group) 5770 continue; 5771 5772 cpumask_set_cpu(j, covered); 5773 cpumask_set_cpu(j, sched_group_cpus(sg)); 5774 } 5775 5776 if (!first) 5777 first = sg; 5778 if (last) 5779 last->next = sg; 5780 last = sg; 5781 } 5782 last->next = first; 5783 5784 return 0; 5785} 5786 5787/* 5788 * Initialize sched groups cpu_power. 5789 * 5790 * cpu_power indicates the capacity of sched group, which is used while 5791 * distributing the load between different sched groups in a sched domain. 5792 * Typically cpu_power for all the groups in a sched domain will be same unless 5793 * there are asymmetries in the topology. If there are asymmetries, group 5794 * having more cpu_power will pickup more load compared to the group having 5795 * less cpu_power. 5796 */ 5797static void init_sched_groups_power(int cpu, struct sched_domain *sd) 5798{ 5799 struct sched_group *sg = sd->groups; 5800 5801 WARN_ON(!sd || !sg); 5802 5803 do { 5804 sg->group_weight = cpumask_weight(sched_group_cpus(sg)); 5805 sg = sg->next; 5806 } while (sg != sd->groups); 5807 5808 if (cpu != group_balance_cpu(sg)) 5809 return; 5810 5811 update_group_power(sd, cpu); 5812 atomic_set(&sg->sgp->nr_busy_cpus, sg->group_weight); 5813} 5814 5815int __weak arch_sd_sibling_asym_packing(void) 5816{ 5817 return 0*SD_ASYM_PACKING; 5818} 5819 5820/* 5821 * Initializers for schedule domains 5822 * Non-inlined to reduce accumulated stack pressure in build_sched_domains() 5823 */ 5824 5825#ifdef CONFIG_SCHED_DEBUG 5826# define SD_INIT_NAME(sd, type) sd->name = #type 5827#else 5828# define SD_INIT_NAME(sd, type) do { } while (0) 5829#endif 5830 5831#define SD_INIT_FUNC(type) \ 5832static noinline struct sched_domain * \ 5833sd_init_##type(struct sched_domain_topology_level *tl, int cpu) \ 5834{ \ 5835 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); \ 5836 *sd = SD_##type##_INIT; \ 5837 SD_INIT_NAME(sd, type); \ 5838 sd->private = &tl->data; \ 5839 return sd; \ 5840} 5841 5842SD_INIT_FUNC(CPU) 5843#ifdef CONFIG_SCHED_SMT 5844 SD_INIT_FUNC(SIBLING) 5845#endif 5846#ifdef CONFIG_SCHED_MC 5847 SD_INIT_FUNC(MC) 5848#endif 5849#ifdef CONFIG_SCHED_BOOK 5850 SD_INIT_FUNC(BOOK) 5851#endif 5852 5853static int default_relax_domain_level = -1; 5854int sched_domain_level_max; 5855 5856static int __init setup_relax_domain_level(char *str) 5857{ 5858 if (kstrtoint(str, 0, &default_relax_domain_level)) 5859 pr_warn("Unable to set relax_domain_level\n"); 5860 5861 return 1; 5862} 5863__setup("relax_domain_level=", setup_relax_domain_level); 5864 5865static void set_domain_attribute(struct sched_domain *sd, 5866 struct sched_domain_attr *attr) 5867{ 5868 int request; 5869 5870 if (!attr || attr->relax_domain_level < 0) { 5871 if (default_relax_domain_level < 0) 5872 return; 5873 else 5874 request = default_relax_domain_level; 5875 } else 5876 request = attr->relax_domain_level; 5877 if (request < sd->level) { 5878 /* turn off idle balance on this domain */ 5879 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE); 5880 } else { 5881 /* turn on idle balance on this domain */ 5882 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE); 5883 } 5884} 5885 5886static void __sdt_free(const struct cpumask *cpu_map); 5887static int __sdt_alloc(const struct cpumask *cpu_map); 5888 5889static void __free_domain_allocs(struct s_data *d, enum s_alloc what, 5890 const struct cpumask *cpu_map) 5891{ 5892 switch (what) { 5893 case sa_rootdomain: 5894 if (!atomic_read(&d->rd->refcount)) 5895 free_rootdomain(&d->rd->rcu); /* fall through */ 5896 case sa_sd: 5897 free_percpu(d->sd); /* fall through */ 5898 case sa_sd_storage: 5899 __sdt_free(cpu_map); /* fall through */ 5900 case sa_none: 5901 break; 5902 } 5903} 5904 5905static enum s_alloc __visit_domain_allocation_hell(struct s_data *d, 5906 const struct cpumask *cpu_map) 5907{ 5908 memset(d, 0, sizeof(*d)); 5909 5910 if (__sdt_alloc(cpu_map)) 5911 return sa_sd_storage; 5912 d->sd = alloc_percpu(struct sched_domain *); 5913 if (!d->sd) 5914 return sa_sd_storage; 5915 d->rd = alloc_rootdomain(); 5916 if (!d->rd) 5917 return sa_sd; 5918 return sa_rootdomain; 5919} 5920 5921/* 5922 * NULL the sd_data elements we've used to build the sched_domain and 5923 * sched_group structure so that the subsequent __free_domain_allocs() 5924 * will not free the data we're using. 5925 */ 5926static void claim_allocations(int cpu, struct sched_domain *sd) 5927{ 5928 struct sd_data *sdd = sd->private; 5929 5930 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd); 5931 *per_cpu_ptr(sdd->sd, cpu) = NULL; 5932 5933 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref)) 5934 *per_cpu_ptr(sdd->sg, cpu) = NULL; 5935 5936 if (atomic_read(&(*per_cpu_ptr(sdd->sgp, cpu))->ref)) 5937 *per_cpu_ptr(sdd->sgp, cpu) = NULL; 5938} 5939 5940#ifdef CONFIG_SCHED_SMT 5941static const struct cpumask *cpu_smt_mask(int cpu) 5942{ 5943 return topology_thread_cpumask(cpu); 5944} 5945#endif 5946 5947/* 5948 * Topology list, bottom-up. 5949 */ 5950static struct sched_domain_topology_level default_topology[] = { 5951#ifdef CONFIG_SCHED_SMT 5952 { sd_init_SIBLING, cpu_smt_mask, }, 5953#endif 5954#ifdef CONFIG_SCHED_MC 5955 { sd_init_MC, cpu_coregroup_mask, }, 5956#endif 5957#ifdef CONFIG_SCHED_BOOK 5958 { sd_init_BOOK, cpu_book_mask, }, 5959#endif 5960 { sd_init_CPU, cpu_cpu_mask, }, 5961 { NULL, }, 5962}; 5963 5964static struct sched_domain_topology_level *sched_domain_topology = default_topology; 5965 5966#ifdef CONFIG_NUMA 5967 5968static int sched_domains_numa_levels; 5969static int *sched_domains_numa_distance; 5970static struct cpumask ***sched_domains_numa_masks; 5971static int sched_domains_curr_level; 5972 5973static inline int sd_local_flags(int level) 5974{ 5975 if (sched_domains_numa_distance[level] > RECLAIM_DISTANCE) 5976 return 0; 5977 5978 return SD_BALANCE_EXEC | SD_BALANCE_FORK | SD_WAKE_AFFINE; 5979} 5980 5981static struct sched_domain * 5982sd_numa_init(struct sched_domain_topology_level *tl, int cpu) 5983{ 5984 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); 5985 int level = tl->numa_level; 5986 int sd_weight = cpumask_weight( 5987 sched_domains_numa_masks[level][cpu_to_node(cpu)]); 5988 5989 *sd = (struct sched_domain){ 5990 .min_interval = sd_weight, 5991 .max_interval = 2*sd_weight, 5992 .busy_factor = 32, 5993 .imbalance_pct = 125, 5994 .cache_nice_tries = 2, 5995 .busy_idx = 3, 5996 .idle_idx = 2, 5997 .newidle_idx = 0, 5998 .wake_idx = 0, 5999 .forkexec_idx = 0, 6000 6001 .flags = 1*SD_LOAD_BALANCE 6002 | 1*SD_BALANCE_NEWIDLE 6003 | 0*SD_BALANCE_EXEC 6004 | 0*SD_BALANCE_FORK 6005 | 0*SD_BALANCE_WAKE 6006 | 0*SD_WAKE_AFFINE 6007 | 0*SD_SHARE_CPUPOWER 6008 | 0*SD_SHARE_PKG_RESOURCES 6009 | 1*SD_SERIALIZE 6010 | 0*SD_PREFER_SIBLING 6011 | sd_local_flags(level) 6012 , 6013 .last_balance = jiffies, 6014 .balance_interval = sd_weight, 6015 }; 6016 SD_INIT_NAME(sd, NUMA); 6017 sd->private = &tl->data; 6018 6019 /* 6020 * Ugly hack to pass state to sd_numa_mask()... 6021 */ 6022 sched_domains_curr_level = tl->numa_level; 6023 6024 return sd; 6025} 6026 6027static const struct cpumask *sd_numa_mask(int cpu) 6028{ 6029 return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)]; 6030} 6031 6032static void sched_numa_warn(const char *str) 6033{ 6034 static int done = false; 6035 int i,j; 6036 6037 if (done) 6038 return; 6039 6040 done = true; 6041 6042 printk(KERN_WARNING "ERROR: %s\n\n", str); 6043 6044 for (i = 0; i < nr_node_ids; i++) { 6045 printk(KERN_WARNING " "); 6046 for (j = 0; j < nr_node_ids; j++) 6047 printk(KERN_CONT "%02d ", node_distance(i,j)); 6048 printk(KERN_CONT "\n"); 6049 } 6050 printk(KERN_WARNING "\n"); 6051} 6052 6053static bool find_numa_distance(int distance) 6054{ 6055 int i; 6056 6057 if (distance == node_distance(0, 0)) 6058 return true; 6059 6060 for (i = 0; i < sched_domains_numa_levels; i++) { 6061 if (sched_domains_numa_distance[i] == distance) 6062 return true; 6063 } 6064 6065 return false; 6066} 6067 6068static void sched_init_numa(void) 6069{ 6070 int next_distance, curr_distance = node_distance(0, 0); 6071 struct sched_domain_topology_level *tl; 6072 int level = 0; 6073 int i, j, k; 6074 6075 sched_domains_numa_distance = kzalloc(sizeof(int) * nr_node_ids, GFP_KERNEL); 6076 if (!sched_domains_numa_distance) 6077 return; 6078 6079 /* 6080 * O(nr_nodes^2) deduplicating selection sort -- in order to find the 6081 * unique distances in the node_distance() table. 6082 * 6083 * Assumes node_distance(0,j) includes all distances in 6084 * node_distance(i,j) in order to avoid cubic time. 6085 */ 6086 next_distance = curr_distance; 6087 for (i = 0; i < nr_node_ids; i++) { 6088 for (j = 0; j < nr_node_ids; j++) { 6089 for (k = 0; k < nr_node_ids; k++) { 6090 int distance = node_distance(i, k); 6091 6092 if (distance > curr_distance && 6093 (distance < next_distance || 6094 next_distance == curr_distance)) 6095 next_distance = distance; 6096 6097 /* 6098 * While not a strong assumption it would be nice to know 6099 * about cases where if node A is connected to B, B is not 6100 * equally connected to A. 6101 */ 6102 if (sched_debug() && node_distance(k, i) != distance) 6103 sched_numa_warn("Node-distance not symmetric"); 6104 6105 if (sched_debug() && i && !find_numa_distance(distance)) 6106 sched_numa_warn("Node-0 not representative"); 6107 } 6108 if (next_distance != curr_distance) { 6109 sched_domains_numa_distance[level++] = next_distance; 6110 sched_domains_numa_levels = level; 6111 curr_distance = next_distance; 6112 } else break; 6113 } 6114 6115 /* 6116 * In case of sched_debug() we verify the above assumption. 6117 */ 6118 if (!sched_debug()) 6119 break; 6120 } 6121 /* 6122 * 'level' contains the number of unique distances, excluding the 6123 * identity distance node_distance(i,i). 6124 * 6125 * The sched_domains_nume_distance[] array includes the actual distance 6126 * numbers. 6127 */ 6128 6129 /* 6130 * Here, we should temporarily reset sched_domains_numa_levels to 0. 6131 * If it fails to allocate memory for array sched_domains_numa_masks[][], 6132 * the array will contain less then 'level' members. This could be 6133 * dangerous when we use it to iterate array sched_domains_numa_masks[][] 6134 * in other functions. 6135 * 6136 * We reset it to 'level' at the end of this function. 6137 */ 6138 sched_domains_numa_levels = 0; 6139 6140 sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL); 6141 if (!sched_domains_numa_masks) 6142 return; 6143 6144 /* 6145 * Now for each level, construct a mask per node which contains all 6146 * cpus of nodes that are that many hops away from us. 6147 */ 6148 for (i = 0; i < level; i++) { 6149 sched_domains_numa_masks[i] = 6150 kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL); 6151 if (!sched_domains_numa_masks[i]) 6152 return; 6153 6154 for (j = 0; j < nr_node_ids; j++) { 6155 struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL); 6156 if (!mask) 6157 return; 6158 6159 sched_domains_numa_masks[i][j] = mask; 6160 6161 for (k = 0; k < nr_node_ids; k++) { 6162 if (node_distance(j, k) > sched_domains_numa_distance[i]) 6163 continue; 6164 6165 cpumask_or(mask, mask, cpumask_of_node(k)); 6166 } 6167 } 6168 } 6169 6170 tl = kzalloc((ARRAY_SIZE(default_topology) + level) * 6171 sizeof(struct sched_domain_topology_level), GFP_KERNEL); 6172 if (!tl) 6173 return; 6174 6175 /* 6176 * Copy the default topology bits.. 6177 */ 6178 for (i = 0; default_topology[i].init; i++) 6179 tl[i] = default_topology[i]; 6180 6181 /* 6182 * .. and append 'j' levels of NUMA goodness. 6183 */ 6184 for (j = 0; j < level; i++, j++) { 6185 tl[i] = (struct sched_domain_topology_level){ 6186 .init = sd_numa_init, 6187 .mask = sd_numa_mask, 6188 .flags = SDTL_OVERLAP, 6189 .numa_level = j, 6190 }; 6191 } 6192 6193 sched_domain_topology = tl; 6194 6195 sched_domains_numa_levels = level; 6196} 6197 6198static void sched_domains_numa_masks_set(int cpu) 6199{ 6200 int i, j; 6201 int node = cpu_to_node(cpu); 6202 6203 for (i = 0; i < sched_domains_numa_levels; i++) { 6204 for (j = 0; j < nr_node_ids; j++) { 6205 if (node_distance(j, node) <= sched_domains_numa_distance[i]) 6206 cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]); 6207 } 6208 } 6209} 6210 6211static void sched_domains_numa_masks_clear(int cpu) 6212{ 6213 int i, j; 6214 for (i = 0; i < sched_domains_numa_levels; i++) { 6215 for (j = 0; j < nr_node_ids; j++) 6216 cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]); 6217 } 6218} 6219 6220/* 6221 * Update sched_domains_numa_masks[level][node] array when new cpus 6222 * are onlined. 6223 */ 6224static int sched_domains_numa_masks_update(struct notifier_block *nfb, 6225 unsigned long action, 6226 void *hcpu) 6227{ 6228 int cpu = (long)hcpu; 6229 6230 switch (action & ~CPU_TASKS_FROZEN) { 6231 case CPU_ONLINE: 6232 sched_domains_numa_masks_set(cpu); 6233 break; 6234 6235 case CPU_DEAD: 6236 sched_domains_numa_masks_clear(cpu); 6237 break; 6238 6239 default: 6240 return NOTIFY_DONE; 6241 } 6242 6243 return NOTIFY_OK; 6244} 6245#else 6246static inline void sched_init_numa(void) 6247{ 6248} 6249 6250static int sched_domains_numa_masks_update(struct notifier_block *nfb, 6251 unsigned long action, 6252 void *hcpu) 6253{ 6254 return 0; 6255} 6256#endif /* CONFIG_NUMA */ 6257 6258static int __sdt_alloc(const struct cpumask *cpu_map) 6259{ 6260 struct sched_domain_topology_level *tl; 6261 int j; 6262 6263 for (tl = sched_domain_topology; tl->init; tl++) { 6264 struct sd_data *sdd = &tl->data; 6265 6266 sdd->sd = alloc_percpu(struct sched_domain *); 6267 if (!sdd->sd) 6268 return -ENOMEM; 6269 6270 sdd->sg = alloc_percpu(struct sched_group *); 6271 if (!sdd->sg) 6272 return -ENOMEM; 6273 6274 sdd->sgp = alloc_percpu(struct sched_group_power *); 6275 if (!sdd->sgp) 6276 return -ENOMEM; 6277 6278 for_each_cpu(j, cpu_map) { 6279 struct sched_domain *sd; 6280 struct sched_group *sg; 6281 struct sched_group_power *sgp; 6282 6283 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(), 6284 GFP_KERNEL, cpu_to_node(j)); 6285 if (!sd) 6286 return -ENOMEM; 6287 6288 *per_cpu_ptr(sdd->sd, j) = sd; 6289 6290 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(), 6291 GFP_KERNEL, cpu_to_node(j)); 6292 if (!sg) 6293 return -ENOMEM; 6294 6295 sg->next = sg; 6296 6297 *per_cpu_ptr(sdd->sg, j) = sg; 6298 6299 sgp = kzalloc_node(sizeof(struct sched_group_power) + cpumask_size(), 6300 GFP_KERNEL, cpu_to_node(j)); 6301 if (!sgp) 6302 return -ENOMEM; 6303 6304 *per_cpu_ptr(sdd->sgp, j) = sgp; 6305 } 6306 } 6307 6308 return 0; 6309} 6310 6311static void __sdt_free(const struct cpumask *cpu_map) 6312{ 6313 struct sched_domain_topology_level *tl; 6314 int j; 6315 6316 for (tl = sched_domain_topology; tl->init; tl++) { 6317 struct sd_data *sdd = &tl->data; 6318 6319 for_each_cpu(j, cpu_map) { 6320 struct sched_domain *sd; 6321 6322 if (sdd->sd) { 6323 sd = *per_cpu_ptr(sdd->sd, j); 6324 if (sd && (sd->flags & SD_OVERLAP)) 6325 free_sched_groups(sd->groups, 0); 6326 kfree(*per_cpu_ptr(sdd->sd, j)); 6327 } 6328 6329 if (sdd->sg) 6330 kfree(*per_cpu_ptr(sdd->sg, j)); 6331 if (sdd->sgp) 6332 kfree(*per_cpu_ptr(sdd->sgp, j)); 6333 } 6334 free_percpu(sdd->sd); 6335 sdd->sd = NULL; 6336 free_percpu(sdd->sg); 6337 sdd->sg = NULL; 6338 free_percpu(sdd->sgp); 6339 sdd->sgp = NULL; 6340 } 6341} 6342 6343struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl, 6344 struct s_data *d, const struct cpumask *cpu_map, 6345 struct sched_domain_attr *attr, struct sched_domain *child, 6346 int cpu) 6347{ 6348 struct sched_domain *sd = tl->init(tl, cpu); 6349 if (!sd) 6350 return child; 6351 6352 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu)); 6353 if (child) { 6354 sd->level = child->level + 1; 6355 sched_domain_level_max = max(sched_domain_level_max, sd->level); 6356 child->parent = sd; 6357 } 6358 sd->child = child; 6359 set_domain_attribute(sd, attr); 6360 6361 return sd; 6362} 6363 6364/* 6365 * Build sched domains for a given set of cpus and attach the sched domains 6366 * to the individual cpus 6367 */ 6368static int build_sched_domains(const struct cpumask *cpu_map, 6369 struct sched_domain_attr *attr) 6370{ 6371 enum s_alloc alloc_state = sa_none; 6372 struct sched_domain *sd; 6373 struct s_data d; 6374 int i, ret = -ENOMEM; 6375 6376 alloc_state = __visit_domain_allocation_hell(&d, cpu_map); 6377 if (alloc_state != sa_rootdomain) 6378 goto error; 6379 6380 /* Set up domains for cpus specified by the cpu_map. */ 6381 for_each_cpu(i, cpu_map) { 6382 struct sched_domain_topology_level *tl; 6383 6384 sd = NULL; 6385 for (tl = sched_domain_topology; tl->init; tl++) { 6386 sd = build_sched_domain(tl, &d, cpu_map, attr, sd, i); 6387 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP)) 6388 sd->flags |= SD_OVERLAP; 6389 if (cpumask_equal(cpu_map, sched_domain_span(sd))) 6390 break; 6391 } 6392 6393 while (sd->child) 6394 sd = sd->child; 6395 6396 *per_cpu_ptr(d.sd, i) = sd; 6397 } 6398 6399 /* Build the groups for the domains */ 6400 for_each_cpu(i, cpu_map) { 6401 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) { 6402 sd->span_weight = cpumask_weight(sched_domain_span(sd)); 6403 if (sd->flags & SD_OVERLAP) { 6404 if (build_overlap_sched_groups(sd, i)) 6405 goto error; 6406 } else { 6407 if (build_sched_groups(sd, i)) 6408 goto error; 6409 } 6410 } 6411 } 6412 6413 /* Calculate CPU power for physical packages and nodes */ 6414 for (i = nr_cpumask_bits-1; i >= 0; i--) { 6415 if (!cpumask_test_cpu(i, cpu_map)) 6416 continue; 6417 6418 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) { 6419 claim_allocations(i, sd); 6420 init_sched_groups_power(i, sd); 6421 } 6422 } 6423 6424 /* Attach the domains */ 6425 rcu_read_lock(); 6426 for_each_cpu(i, cpu_map) { 6427 sd = *per_cpu_ptr(d.sd, i); 6428 cpu_attach_domain(sd, d.rd, i); 6429 } 6430 rcu_read_unlock(); 6431 6432 ret = 0; 6433error: 6434 __free_domain_allocs(&d, alloc_state, cpu_map); 6435 return ret; 6436} 6437 6438static cpumask_var_t *doms_cur; /* current sched domains */ 6439static int ndoms_cur; /* number of sched domains in 'doms_cur' */ 6440static struct sched_domain_attr *dattr_cur; 6441 /* attribues of custom domains in 'doms_cur' */ 6442 6443/* 6444 * Special case: If a kmalloc of a doms_cur partition (array of 6445 * cpumask) fails, then fallback to a single sched domain, 6446 * as determined by the single cpumask fallback_doms. 6447 */ 6448static cpumask_var_t fallback_doms; 6449 6450/* 6451 * arch_update_cpu_topology lets virtualized architectures update the 6452 * cpu core maps. It is supposed to return 1 if the topology changed 6453 * or 0 if it stayed the same. 6454 */ 6455int __attribute__((weak)) arch_update_cpu_topology(void) 6456{ 6457 return 0; 6458} 6459 6460cpumask_var_t *alloc_sched_domains(unsigned int ndoms) 6461{ 6462 int i; 6463 cpumask_var_t *doms; 6464 6465 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL); 6466 if (!doms) 6467 return NULL; 6468 for (i = 0; i < ndoms; i++) { 6469 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) { 6470 free_sched_domains(doms, i); 6471 return NULL; 6472 } 6473 } 6474 return doms; 6475} 6476 6477void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms) 6478{ 6479 unsigned int i; 6480 for (i = 0; i < ndoms; i++) 6481 free_cpumask_var(doms[i]); 6482 kfree(doms); 6483} 6484 6485/* 6486 * Set up scheduler domains and groups. Callers must hold the hotplug lock. 6487 * For now this just excludes isolated cpus, but could be used to 6488 * exclude other special cases in the future. 6489 */ 6490static int init_sched_domains(const struct cpumask *cpu_map) 6491{ 6492 int err; 6493 6494 arch_update_cpu_topology(); 6495 ndoms_cur = 1; 6496 doms_cur = alloc_sched_domains(ndoms_cur); 6497 if (!doms_cur) 6498 doms_cur = &fallback_doms; 6499 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map); 6500 err = build_sched_domains(doms_cur[0], NULL); 6501 register_sched_domain_sysctl(); 6502 6503 return err; 6504} 6505 6506/* 6507 * Detach sched domains from a group of cpus specified in cpu_map 6508 * These cpus will now be attached to the NULL domain 6509 */ 6510static void detach_destroy_domains(const struct cpumask *cpu_map) 6511{ 6512 int i; 6513 6514 rcu_read_lock(); 6515 for_each_cpu(i, cpu_map) 6516 cpu_attach_domain(NULL, &def_root_domain, i); 6517 rcu_read_unlock(); 6518} 6519 6520/* handle null as "default" */ 6521static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur, 6522 struct sched_domain_attr *new, int idx_new) 6523{ 6524 struct sched_domain_attr tmp; 6525 6526 /* fast path */ 6527 if (!new && !cur) 6528 return 1; 6529 6530 tmp = SD_ATTR_INIT; 6531 return !memcmp(cur ? (cur + idx_cur) : &tmp, 6532 new ? (new + idx_new) : &tmp, 6533 sizeof(struct sched_domain_attr)); 6534} 6535 6536/* 6537 * Partition sched domains as specified by the 'ndoms_new' 6538 * cpumasks in the array doms_new[] of cpumasks. This compares 6539 * doms_new[] to the current sched domain partitioning, doms_cur[]. 6540 * It destroys each deleted domain and builds each new domain. 6541 * 6542 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'. 6543 * The masks don't intersect (don't overlap.) We should setup one 6544 * sched domain for each mask. CPUs not in any of the cpumasks will 6545 * not be load balanced. If the same cpumask appears both in the 6546 * current 'doms_cur' domains and in the new 'doms_new', we can leave 6547 * it as it is. 6548 * 6549 * The passed in 'doms_new' should be allocated using 6550 * alloc_sched_domains. This routine takes ownership of it and will 6551 * free_sched_domains it when done with it. If the caller failed the 6552 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1, 6553 * and partition_sched_domains() will fallback to the single partition 6554 * 'fallback_doms', it also forces the domains to be rebuilt. 6555 * 6556 * If doms_new == NULL it will be replaced with cpu_online_mask. 6557 * ndoms_new == 0 is a special case for destroying existing domains, 6558 * and it will not create the default domain. 6559 * 6560 * Call with hotplug lock held 6561 */ 6562void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[], 6563 struct sched_domain_attr *dattr_new) 6564{ 6565 int i, j, n; 6566 int new_topology; 6567 6568 mutex_lock(&sched_domains_mutex); 6569 6570 /* always unregister in case we don't destroy any domains */ 6571 unregister_sched_domain_sysctl(); 6572 6573 /* Let architecture update cpu core mappings. */ 6574 new_topology = arch_update_cpu_topology(); 6575 6576 n = doms_new ? ndoms_new : 0; 6577 6578 /* Destroy deleted domains */ 6579 for (i = 0; i < ndoms_cur; i++) { 6580 for (j = 0; j < n && !new_topology; j++) { 6581 if (cpumask_equal(doms_cur[i], doms_new[j]) 6582 && dattrs_equal(dattr_cur, i, dattr_new, j)) 6583 goto match1; 6584 } 6585 /* no match - a current sched domain not in new doms_new[] */ 6586 detach_destroy_domains(doms_cur[i]); 6587match1: 6588 ; 6589 } 6590 6591 if (doms_new == NULL) { 6592 ndoms_cur = 0; 6593 doms_new = &fallback_doms; 6594 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map); 6595 WARN_ON_ONCE(dattr_new); 6596 } 6597 6598 /* Build new domains */ 6599 for (i = 0; i < ndoms_new; i++) { 6600 for (j = 0; j < ndoms_cur && !new_topology; j++) { 6601 if (cpumask_equal(doms_new[i], doms_cur[j]) 6602 && dattrs_equal(dattr_new, i, dattr_cur, j)) 6603 goto match2; 6604 } 6605 /* no match - add a new doms_new */ 6606 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL); 6607match2: 6608 ; 6609 } 6610 6611 /* Remember the new sched domains */ 6612 if (doms_cur != &fallback_doms) 6613 free_sched_domains(doms_cur, ndoms_cur); 6614 kfree(dattr_cur); /* kfree(NULL) is safe */ 6615 doms_cur = doms_new; 6616 dattr_cur = dattr_new; 6617 ndoms_cur = ndoms_new; 6618 6619 register_sched_domain_sysctl(); 6620 6621 mutex_unlock(&sched_domains_mutex); 6622} 6623 6624static int num_cpus_frozen; /* used to mark begin/end of suspend/resume */ 6625 6626/* 6627 * Update cpusets according to cpu_active mask. If cpusets are 6628 * disabled, cpuset_update_active_cpus() becomes a simple wrapper 6629 * around partition_sched_domains(). 6630 * 6631 * If we come here as part of a suspend/resume, don't touch cpusets because we 6632 * want to restore it back to its original state upon resume anyway. 6633 */ 6634static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action, 6635 void *hcpu) 6636{ 6637 switch (action) { 6638 case CPU_ONLINE_FROZEN: 6639 case CPU_DOWN_FAILED_FROZEN: 6640 6641 /* 6642 * num_cpus_frozen tracks how many CPUs are involved in suspend 6643 * resume sequence. As long as this is not the last online 6644 * operation in the resume sequence, just build a single sched 6645 * domain, ignoring cpusets. 6646 */ 6647 num_cpus_frozen--; 6648 if (likely(num_cpus_frozen)) { 6649 partition_sched_domains(1, NULL, NULL); 6650 break; 6651 } 6652 6653 /* 6654 * This is the last CPU online operation. So fall through and 6655 * restore the original sched domains by considering the 6656 * cpuset configurations. 6657 */ 6658 6659 case CPU_ONLINE: 6660 case CPU_DOWN_FAILED: 6661 cpuset_update_active_cpus(true); 6662 break; 6663 default: 6664 return NOTIFY_DONE; 6665 } 6666 return NOTIFY_OK; 6667} 6668 6669static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action, 6670 void *hcpu) 6671{ 6672 switch (action) { 6673 case CPU_DOWN_PREPARE: 6674 cpuset_update_active_cpus(false); 6675 break; 6676 case CPU_DOWN_PREPARE_FROZEN: 6677 num_cpus_frozen++; 6678 partition_sched_domains(1, NULL, NULL); 6679 break; 6680 default: 6681 return NOTIFY_DONE; 6682 } 6683 return NOTIFY_OK; 6684} 6685 6686void __init sched_init_smp(void) 6687{ 6688 cpumask_var_t non_isolated_cpus; 6689 6690 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL); 6691 alloc_cpumask_var(&fallback_doms, GFP_KERNEL); 6692 6693 sched_init_numa(); 6694 6695 get_online_cpus(); 6696 mutex_lock(&sched_domains_mutex); 6697 init_sched_domains(cpu_active_mask); 6698 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map); 6699 if (cpumask_empty(non_isolated_cpus)) 6700 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus); 6701 mutex_unlock(&sched_domains_mutex); 6702 put_online_cpus(); 6703 6704 hotcpu_notifier(sched_domains_numa_masks_update, CPU_PRI_SCHED_ACTIVE); 6705 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE); 6706 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE); 6707 6708 /* RT runtime code needs to handle some hotplug events */ 6709 hotcpu_notifier(update_runtime, 0); 6710 6711 init_hrtick(); 6712 6713 /* Move init over to a non-isolated CPU */ 6714 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0) 6715 BUG(); 6716 sched_init_granularity(); 6717 free_cpumask_var(non_isolated_cpus); 6718 6719 init_sched_rt_class(); 6720} 6721#else 6722void __init sched_init_smp(void) 6723{ 6724 sched_init_granularity(); 6725} 6726#endif /* CONFIG_SMP */ 6727 6728const_debug unsigned int sysctl_timer_migration = 1; 6729 6730int in_sched_functions(unsigned long addr) 6731{ 6732 return in_lock_functions(addr) || 6733 (addr >= (unsigned long)__sched_text_start 6734 && addr < (unsigned long)__sched_text_end); 6735} 6736 6737#ifdef CONFIG_CGROUP_SCHED 6738struct task_group root_task_group; 6739LIST_HEAD(task_groups); 6740#endif 6741 6742DECLARE_PER_CPU(cpumask_var_t, load_balance_tmpmask); 6743 6744void __init sched_init(void) 6745{ 6746 int i, j; 6747 unsigned long alloc_size = 0, ptr; 6748 6749#ifdef CONFIG_FAIR_GROUP_SCHED 6750 alloc_size += 2 * nr_cpu_ids * sizeof(void **); 6751#endif 6752#ifdef CONFIG_RT_GROUP_SCHED 6753 alloc_size += 2 * nr_cpu_ids * sizeof(void **); 6754#endif 6755#ifdef CONFIG_CPUMASK_OFFSTACK 6756 alloc_size += num_possible_cpus() * cpumask_size(); 6757#endif 6758 if (alloc_size) { 6759 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT); 6760 6761#ifdef CONFIG_FAIR_GROUP_SCHED 6762 root_task_group.se = (struct sched_entity **)ptr; 6763 ptr += nr_cpu_ids * sizeof(void **); 6764 6765 root_task_group.cfs_rq = (struct cfs_rq **)ptr; 6766 ptr += nr_cpu_ids * sizeof(void **); 6767 6768#endif /* CONFIG_FAIR_GROUP_SCHED */ 6769#ifdef CONFIG_RT_GROUP_SCHED 6770 root_task_group.rt_se = (struct sched_rt_entity **)ptr; 6771 ptr += nr_cpu_ids * sizeof(void **); 6772 6773 root_task_group.rt_rq = (struct rt_rq **)ptr; 6774 ptr += nr_cpu_ids * sizeof(void **); 6775 6776#endif /* CONFIG_RT_GROUP_SCHED */ 6777#ifdef CONFIG_CPUMASK_OFFSTACK 6778 for_each_possible_cpu(i) { 6779 per_cpu(load_balance_tmpmask, i) = (void *)ptr; 6780 ptr += cpumask_size(); 6781 } 6782#endif /* CONFIG_CPUMASK_OFFSTACK */ 6783 } 6784 6785#ifdef CONFIG_SMP 6786 init_defrootdomain(); 6787#endif 6788 6789 init_rt_bandwidth(&def_rt_bandwidth, 6790 global_rt_period(), global_rt_runtime()); 6791 6792#ifdef CONFIG_RT_GROUP_SCHED 6793 init_rt_bandwidth(&root_task_group.rt_bandwidth, 6794 global_rt_period(), global_rt_runtime()); 6795#endif /* CONFIG_RT_GROUP_SCHED */ 6796 6797#ifdef CONFIG_CGROUP_SCHED 6798 list_add(&root_task_group.list, &task_groups); 6799 INIT_LIST_HEAD(&root_task_group.children); 6800 INIT_LIST_HEAD(&root_task_group.siblings); 6801 autogroup_init(&init_task); 6802 6803#endif /* CONFIG_CGROUP_SCHED */ 6804 6805#ifdef CONFIG_CGROUP_CPUACCT 6806 root_cpuacct.cpustat = &kernel_cpustat; 6807 root_cpuacct.cpuusage = alloc_percpu(u64); 6808 /* Too early, not expected to fail */ 6809 BUG_ON(!root_cpuacct.cpuusage); 6810#endif 6811 for_each_possible_cpu(i) { 6812 struct rq *rq; 6813 6814 rq = cpu_rq(i); 6815 raw_spin_lock_init(&rq->lock); 6816 rq->nr_running = 0; 6817 rq->calc_load_active = 0; 6818 rq->calc_load_update = jiffies + LOAD_FREQ; 6819 init_cfs_rq(&rq->cfs); 6820 init_rt_rq(&rq->rt, rq); 6821#ifdef CONFIG_FAIR_GROUP_SCHED 6822 root_task_group.shares = ROOT_TASK_GROUP_LOAD; 6823 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list); 6824 /* 6825 * How much cpu bandwidth does root_task_group get? 6826 * 6827 * In case of task-groups formed thr' the cgroup filesystem, it 6828 * gets 100% of the cpu resources in the system. This overall 6829 * system cpu resource is divided among the tasks of 6830 * root_task_group and its child task-groups in a fair manner, 6831 * based on each entity's (task or task-group's) weight 6832 * (se->load.weight). 6833 * 6834 * In other words, if root_task_group has 10 tasks of weight 6835 * 1024) and two child groups A0 and A1 (of weight 1024 each), 6836 * then A0's share of the cpu resource is: 6837 * 6838 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33% 6839 * 6840 * We achieve this by letting root_task_group's tasks sit 6841 * directly in rq->cfs (i.e root_task_group->se[] = NULL). 6842 */ 6843 init_cfs_bandwidth(&root_task_group.cfs_bandwidth); 6844 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL); 6845#endif /* CONFIG_FAIR_GROUP_SCHED */ 6846 6847 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime; 6848#ifdef CONFIG_RT_GROUP_SCHED 6849 INIT_LIST_HEAD(&rq->leaf_rt_rq_list); 6850 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL); 6851#endif 6852 6853 for (j = 0; j < CPU_LOAD_IDX_MAX; j++) 6854 rq->cpu_load[j] = 0; 6855 6856 rq->last_load_update_tick = jiffies; 6857 6858#ifdef CONFIG_SMP 6859 rq->sd = NULL; 6860 rq->rd = NULL; 6861 rq->cpu_power = SCHED_POWER_SCALE; 6862 rq->post_schedule = 0; 6863 rq->active_balance = 0; 6864 rq->next_balance = jiffies; 6865 rq->push_cpu = 0; 6866 rq->cpu = i; 6867 rq->online = 0; 6868 rq->idle_stamp = 0; 6869 rq->avg_idle = 2*sysctl_sched_migration_cost; 6870 6871 INIT_LIST_HEAD(&rq->cfs_tasks); 6872 6873 rq_attach_root(rq, &def_root_domain); 6874#ifdef CONFIG_NO_HZ 6875 rq->nohz_flags = 0; 6876#endif 6877#endif 6878 init_rq_hrtick(rq); 6879 atomic_set(&rq->nr_iowait, 0); 6880 } 6881 6882 set_load_weight(&init_task); 6883 6884#ifdef CONFIG_PREEMPT_NOTIFIERS 6885 INIT_HLIST_HEAD(&init_task.preempt_notifiers); 6886#endif 6887 6888#ifdef CONFIG_RT_MUTEXES 6889 plist_head_init(&init_task.pi_waiters); 6890#endif 6891 6892 /* 6893 * The boot idle thread does lazy MMU switching as well: 6894 */ 6895 atomic_inc(&init_mm.mm_count); 6896 enter_lazy_tlb(&init_mm, current); 6897 6898 /* 6899 * Make us the idle thread. Technically, schedule() should not be 6900 * called from this thread, however somewhere below it might be, 6901 * but because we are the idle thread, we just pick up running again 6902 * when this runqueue becomes "idle". 6903 */ 6904 init_idle(current, smp_processor_id()); 6905 6906 calc_load_update = jiffies + LOAD_FREQ; 6907 6908 /* 6909 * During early bootup we pretend to be a normal task: 6910 */ 6911 current->sched_class = &fair_sched_class; 6912 6913#ifdef CONFIG_SMP 6914 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT); 6915 /* May be allocated at isolcpus cmdline parse time */ 6916 if (cpu_isolated_map == NULL) 6917 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT); 6918 idle_thread_set_boot_cpu(); 6919#endif 6920 init_sched_fair_class(); 6921 6922 scheduler_running = 1; 6923} 6924 6925#ifdef CONFIG_DEBUG_ATOMIC_SLEEP 6926static inline int preempt_count_equals(int preempt_offset) 6927{ 6928 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth(); 6929 6930 return (nested == preempt_offset); 6931} 6932 6933void __might_sleep(const char *file, int line, int preempt_offset) 6934{ 6935 static unsigned long prev_jiffy; /* ratelimiting */ 6936 6937 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */ 6938 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) || 6939 system_state != SYSTEM_RUNNING || oops_in_progress) 6940 return; 6941 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy) 6942 return; 6943 prev_jiffy = jiffies; 6944 6945 printk(KERN_ERR 6946 "BUG: sleeping function called from invalid context at %s:%d\n", 6947 file, line); 6948 printk(KERN_ERR 6949 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n", 6950 in_atomic(), irqs_disabled(), 6951 current->pid, current->comm); 6952 6953 debug_show_held_locks(current); 6954 if (irqs_disabled()) 6955 print_irqtrace_events(current); 6956 dump_stack(); 6957} 6958EXPORT_SYMBOL(__might_sleep); 6959#endif 6960 6961#ifdef CONFIG_MAGIC_SYSRQ 6962static void normalize_task(struct rq *rq, struct task_struct *p) 6963{ 6964 const struct sched_class *prev_class = p->sched_class; 6965 int old_prio = p->prio; 6966 int on_rq; 6967 6968 on_rq = p->on_rq; 6969 if (on_rq) 6970 dequeue_task(rq, p, 0); 6971 __setscheduler(rq, p, SCHED_NORMAL, 0); 6972 if (on_rq) { 6973 enqueue_task(rq, p, 0); 6974 resched_task(rq->curr); 6975 } 6976 6977 check_class_changed(rq, p, prev_class, old_prio); 6978} 6979 6980void normalize_rt_tasks(void) 6981{ 6982 struct task_struct *g, *p; 6983 unsigned long flags; 6984 struct rq *rq; 6985 6986 read_lock_irqsave(&tasklist_lock, flags); 6987 do_each_thread(g, p) { 6988 /* 6989 * Only normalize user tasks: 6990 */ 6991 if (!p->mm) 6992 continue; 6993 6994 p->se.exec_start = 0; 6995#ifdef CONFIG_SCHEDSTATS 6996 p->se.statistics.wait_start = 0; 6997 p->se.statistics.sleep_start = 0; 6998 p->se.statistics.block_start = 0; 6999#endif 7000 7001 if (!rt_task(p)) { 7002 /* 7003 * Renice negative nice level userspace 7004 * tasks back to 0: 7005 */ 7006 if (TASK_NICE(p) < 0 && p->mm) 7007 set_user_nice(p, 0); 7008 continue; 7009 } 7010 7011 raw_spin_lock(&p->pi_lock); 7012 rq = __task_rq_lock(p); 7013 7014 normalize_task(rq, p); 7015 7016 __task_rq_unlock(rq); 7017 raw_spin_unlock(&p->pi_lock); 7018 } while_each_thread(g, p); 7019 7020 read_unlock_irqrestore(&tasklist_lock, flags); 7021} 7022 7023#endif /* CONFIG_MAGIC_SYSRQ */ 7024 7025#if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) 7026/* 7027 * These functions are only useful for the IA64 MCA handling, or kdb. 7028 * 7029 * They can only be called when the whole system has been 7030 * stopped - every CPU needs to be quiescent, and no scheduling 7031 * activity can take place. Using them for anything else would 7032 * be a serious bug, and as a result, they aren't even visible 7033 * under any other configuration. 7034 */ 7035 7036/** 7037 * curr_task - return the current task for a given cpu. 7038 * @cpu: the processor in question. 7039 * 7040 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED! 7041 */ 7042struct task_struct *curr_task(int cpu) 7043{ 7044 return cpu_curr(cpu); 7045} 7046 7047#endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */ 7048 7049#ifdef CONFIG_IA64 7050/** 7051 * set_curr_task - set the current task for a given cpu. 7052 * @cpu: the processor in question. 7053 * @p: the task pointer to set. 7054 * 7055 * Description: This function must only be used when non-maskable interrupts 7056 * are serviced on a separate stack. It allows the architecture to switch the 7057 * notion of the current task on a cpu in a non-blocking manner. This function 7058 * must be called with all CPU's synchronized, and interrupts disabled, the 7059 * and caller must save the original value of the current task (see 7060 * curr_task() above) and restore that value before reenabling interrupts and 7061 * re-starting the system. 7062 * 7063 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED! 7064 */ 7065void set_curr_task(int cpu, struct task_struct *p) 7066{ 7067 cpu_curr(cpu) = p; 7068} 7069 7070#endif 7071 7072#ifdef CONFIG_CGROUP_SCHED 7073/* task_group_lock serializes the addition/removal of task groups */ 7074static DEFINE_SPINLOCK(task_group_lock); 7075 7076static void free_sched_group(struct task_group *tg) 7077{ 7078 free_fair_sched_group(tg); 7079 free_rt_sched_group(tg); 7080 autogroup_free(tg); 7081 kfree(tg); 7082} 7083 7084/* allocate runqueue etc for a new task group */ 7085struct task_group *sched_create_group(struct task_group *parent) 7086{ 7087 struct task_group *tg; 7088 unsigned long flags; 7089 7090 tg = kzalloc(sizeof(*tg), GFP_KERNEL); 7091 if (!tg) 7092 return ERR_PTR(-ENOMEM); 7093 7094 if (!alloc_fair_sched_group(tg, parent)) 7095 goto err; 7096 7097 if (!alloc_rt_sched_group(tg, parent)) 7098 goto err; 7099 7100 spin_lock_irqsave(&task_group_lock, flags); 7101 list_add_rcu(&tg->list, &task_groups); 7102 7103 WARN_ON(!parent); /* root should already exist */ 7104 7105 tg->parent = parent; 7106 INIT_LIST_HEAD(&tg->children); 7107 list_add_rcu(&tg->siblings, &parent->children); 7108 spin_unlock_irqrestore(&task_group_lock, flags); 7109 7110 return tg; 7111 7112err: 7113 free_sched_group(tg); 7114 return ERR_PTR(-ENOMEM); 7115} 7116 7117/* rcu callback to free various structures associated with a task group */ 7118static void free_sched_group_rcu(struct rcu_head *rhp) 7119{ 7120 /* now it should be safe to free those cfs_rqs */ 7121 free_sched_group(container_of(rhp, struct task_group, rcu)); 7122} 7123 7124/* Destroy runqueue etc associated with a task group */ 7125void sched_destroy_group(struct task_group *tg) 7126{ 7127 unsigned long flags; 7128 int i; 7129 7130 /* end participation in shares distribution */ 7131 for_each_possible_cpu(i) 7132 unregister_fair_sched_group(tg, i); 7133 7134 spin_lock_irqsave(&task_group_lock, flags); 7135 list_del_rcu(&tg->list); 7136 list_del_rcu(&tg->siblings); 7137 spin_unlock_irqrestore(&task_group_lock, flags); 7138 7139 /* wait for possible concurrent references to cfs_rqs complete */ 7140 call_rcu(&tg->rcu, free_sched_group_rcu); 7141} 7142 7143/* change task's runqueue when it moves between groups. 7144 * The caller of this function should have put the task in its new group 7145 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to 7146 * reflect its new group. 7147 */ 7148void sched_move_task(struct task_struct *tsk) 7149{ 7150 struct task_group *tg; 7151 int on_rq, running; 7152 unsigned long flags; 7153 struct rq *rq; 7154 7155 rq = task_rq_lock(tsk, &flags); 7156 7157 running = task_current(rq, tsk); 7158 on_rq = tsk->on_rq; 7159 7160 if (on_rq) 7161 dequeue_task(rq, tsk, 0); 7162 if (unlikely(running)) 7163 tsk->sched_class->put_prev_task(rq, tsk); 7164 7165 tg = container_of(task_subsys_state_check(tsk, cpu_cgroup_subsys_id, 7166 lockdep_is_held(&tsk->sighand->siglock)), 7167 struct task_group, css); 7168 tg = autogroup_task_group(tsk, tg); 7169 tsk->sched_task_group = tg; 7170 7171#ifdef CONFIG_FAIR_GROUP_SCHED 7172 if (tsk->sched_class->task_move_group) 7173 tsk->sched_class->task_move_group(tsk, on_rq); 7174 else 7175#endif 7176 set_task_rq(tsk, task_cpu(tsk)); 7177 7178 if (unlikely(running)) 7179 tsk->sched_class->set_curr_task(rq); 7180 if (on_rq) 7181 enqueue_task(rq, tsk, 0); 7182 7183 task_rq_unlock(rq, tsk, &flags); 7184} 7185#endif /* CONFIG_CGROUP_SCHED */ 7186 7187#if defined(CONFIG_RT_GROUP_SCHED) || defined(CONFIG_CFS_BANDWIDTH) 7188static unsigned long to_ratio(u64 period, u64 runtime) 7189{ 7190 if (runtime == RUNTIME_INF) 7191 return 1ULL << 20; 7192 7193 return div64_u64(runtime << 20, period); 7194} 7195#endif 7196 7197#ifdef CONFIG_RT_GROUP_SCHED 7198/* 7199 * Ensure that the real time constraints are schedulable. 7200 */ 7201static DEFINE_MUTEX(rt_constraints_mutex); 7202 7203/* Must be called with tasklist_lock held */ 7204static inline int tg_has_rt_tasks(struct task_group *tg) 7205{ 7206 struct task_struct *g, *p; 7207 7208 do_each_thread(g, p) { 7209 if (rt_task(p) && task_rq(p)->rt.tg == tg) 7210 return 1; 7211 } while_each_thread(g, p); 7212 7213 return 0; 7214} 7215 7216struct rt_schedulable_data { 7217 struct task_group *tg; 7218 u64 rt_period; 7219 u64 rt_runtime; 7220}; 7221 7222static int tg_rt_schedulable(struct task_group *tg, void *data) 7223{ 7224 struct rt_schedulable_data *d = data; 7225 struct task_group *child; 7226 unsigned long total, sum = 0; 7227 u64 period, runtime; 7228 7229 period = ktime_to_ns(tg->rt_bandwidth.rt_period); 7230 runtime = tg->rt_bandwidth.rt_runtime; 7231 7232 if (tg == d->tg) { 7233 period = d->rt_period; 7234 runtime = d->rt_runtime; 7235 } 7236 7237 /* 7238 * Cannot have more runtime than the period. 7239 */ 7240 if (runtime > period && runtime != RUNTIME_INF) 7241 return -EINVAL; 7242 7243 /* 7244 * Ensure we don't starve existing RT tasks. 7245 */ 7246 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg)) 7247 return -EBUSY; 7248 7249 total = to_ratio(period, runtime); 7250 7251 /* 7252 * Nobody can have more than the global setting allows. 7253 */ 7254 if (total > to_ratio(global_rt_period(), global_rt_runtime())) 7255 return -EINVAL; 7256 7257 /* 7258 * The sum of our children's runtime should not exceed our own. 7259 */ 7260 list_for_each_entry_rcu(child, &tg->children, siblings) { 7261 period = ktime_to_ns(child->rt_bandwidth.rt_period); 7262 runtime = child->rt_bandwidth.rt_runtime; 7263 7264 if (child == d->tg) { 7265 period = d->rt_period; 7266 runtime = d->rt_runtime; 7267 } 7268 7269 sum += to_ratio(period, runtime); 7270 } 7271 7272 if (sum > total) 7273 return -EINVAL; 7274 7275 return 0; 7276} 7277 7278static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime) 7279{ 7280 int ret; 7281 7282 struct rt_schedulable_data data = { 7283 .tg = tg, 7284 .rt_period = period, 7285 .rt_runtime = runtime, 7286 }; 7287 7288 rcu_read_lock(); 7289 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data); 7290 rcu_read_unlock(); 7291 7292 return ret; 7293} 7294 7295static int tg_set_rt_bandwidth(struct task_group *tg, 7296 u64 rt_period, u64 rt_runtime) 7297{ 7298 int i, err = 0; 7299 7300 mutex_lock(&rt_constraints_mutex); 7301 read_lock(&tasklist_lock); 7302 err = __rt_schedulable(tg, rt_period, rt_runtime); 7303 if (err) 7304 goto unlock; 7305 7306 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock); 7307 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period); 7308 tg->rt_bandwidth.rt_runtime = rt_runtime; 7309 7310 for_each_possible_cpu(i) { 7311 struct rt_rq *rt_rq = tg->rt_rq[i]; 7312 7313 raw_spin_lock(&rt_rq->rt_runtime_lock); 7314 rt_rq->rt_runtime = rt_runtime; 7315 raw_spin_unlock(&rt_rq->rt_runtime_lock); 7316 } 7317 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock); 7318unlock: 7319 read_unlock(&tasklist_lock); 7320 mutex_unlock(&rt_constraints_mutex); 7321 7322 return err; 7323} 7324 7325int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us) 7326{ 7327 u64 rt_runtime, rt_period; 7328 7329 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period); 7330 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC; 7331 if (rt_runtime_us < 0) 7332 rt_runtime = RUNTIME_INF; 7333 7334 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime); 7335} 7336 7337long sched_group_rt_runtime(struct task_group *tg) 7338{ 7339 u64 rt_runtime_us; 7340 7341 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF) 7342 return -1; 7343 7344 rt_runtime_us = tg->rt_bandwidth.rt_runtime; 7345 do_div(rt_runtime_us, NSEC_PER_USEC); 7346 return rt_runtime_us; 7347} 7348 7349int sched_group_set_rt_period(struct task_group *tg, long rt_period_us) 7350{ 7351 u64 rt_runtime, rt_period; 7352 7353 rt_period = (u64)rt_period_us * NSEC_PER_USEC; 7354 rt_runtime = tg->rt_bandwidth.rt_runtime; 7355 7356 if (rt_period == 0) 7357 return -EINVAL; 7358 7359 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime); 7360} 7361 7362long sched_group_rt_period(struct task_group *tg) 7363{ 7364 u64 rt_period_us; 7365 7366 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period); 7367 do_div(rt_period_us, NSEC_PER_USEC); 7368 return rt_period_us; 7369} 7370 7371static int sched_rt_global_constraints(void) 7372{ 7373 u64 runtime, period; 7374 int ret = 0; 7375 7376 if (sysctl_sched_rt_period <= 0) 7377 return -EINVAL; 7378 7379 runtime = global_rt_runtime(); 7380 period = global_rt_period(); 7381 7382 /* 7383 * Sanity check on the sysctl variables. 7384 */ 7385 if (runtime > period && runtime != RUNTIME_INF) 7386 return -EINVAL; 7387 7388 mutex_lock(&rt_constraints_mutex); 7389 read_lock(&tasklist_lock); 7390 ret = __rt_schedulable(NULL, 0, 0); 7391 read_unlock(&tasklist_lock); 7392 mutex_unlock(&rt_constraints_mutex); 7393 7394 return ret; 7395} 7396 7397int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk) 7398{ 7399 /* Don't accept realtime tasks when there is no way for them to run */ 7400 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0) 7401 return 0; 7402 7403 return 1; 7404} 7405 7406#else /* !CONFIG_RT_GROUP_SCHED */ 7407static int sched_rt_global_constraints(void) 7408{ 7409 unsigned long flags; 7410 int i; 7411 7412 if (sysctl_sched_rt_period <= 0) 7413 return -EINVAL; 7414 7415 /* 7416 * There's always some RT tasks in the root group 7417 * -- migration, kstopmachine etc.. 7418 */ 7419 if (sysctl_sched_rt_runtime == 0) 7420 return -EBUSY; 7421 7422 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags); 7423 for_each_possible_cpu(i) { 7424 struct rt_rq *rt_rq = &cpu_rq(i)->rt; 7425 7426 raw_spin_lock(&rt_rq->rt_runtime_lock); 7427 rt_rq->rt_runtime = global_rt_runtime(); 7428 raw_spin_unlock(&rt_rq->rt_runtime_lock); 7429 } 7430 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags); 7431 7432 return 0; 7433} 7434#endif /* CONFIG_RT_GROUP_SCHED */ 7435 7436int sched_rt_handler(struct ctl_table *table, int write, 7437 void __user *buffer, size_t *lenp, 7438 loff_t *ppos) 7439{ 7440 int ret; 7441 int old_period, old_runtime; 7442 static DEFINE_MUTEX(mutex); 7443 7444 mutex_lock(&mutex); 7445 old_period = sysctl_sched_rt_period; 7446 old_runtime = sysctl_sched_rt_runtime; 7447 7448 ret = proc_dointvec(table, write, buffer, lenp, ppos); 7449 7450 if (!ret && write) { 7451 ret = sched_rt_global_constraints(); 7452 if (ret) { 7453 sysctl_sched_rt_period = old_period; 7454 sysctl_sched_rt_runtime = old_runtime; 7455 } else { 7456 def_rt_bandwidth.rt_runtime = global_rt_runtime(); 7457 def_rt_bandwidth.rt_period = 7458 ns_to_ktime(global_rt_period()); 7459 } 7460 } 7461 mutex_unlock(&mutex); 7462 7463 return ret; 7464} 7465 7466#ifdef CONFIG_CGROUP_SCHED 7467 7468/* return corresponding task_group object of a cgroup */ 7469static inline struct task_group *cgroup_tg(struct cgroup *cgrp) 7470{ 7471 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id), 7472 struct task_group, css); 7473} 7474 7475static struct cgroup_subsys_state *cpu_cgroup_create(struct cgroup *cgrp) 7476{ 7477 struct task_group *tg, *parent; 7478 7479 if (!cgrp->parent) { 7480 /* This is early initialization for the top cgroup */ 7481 return &root_task_group.css; 7482 } 7483 7484 parent = cgroup_tg(cgrp->parent); 7485 tg = sched_create_group(parent); 7486 if (IS_ERR(tg)) 7487 return ERR_PTR(-ENOMEM); 7488 7489 return &tg->css; 7490} 7491 7492static void cpu_cgroup_destroy(struct cgroup *cgrp) 7493{ 7494 struct task_group *tg = cgroup_tg(cgrp); 7495 7496 sched_destroy_group(tg); 7497} 7498 7499static int cpu_cgroup_can_attach(struct cgroup *cgrp, 7500 struct cgroup_taskset *tset) 7501{ 7502 struct task_struct *task; 7503 7504 cgroup_taskset_for_each(task, cgrp, tset) { 7505#ifdef CONFIG_RT_GROUP_SCHED 7506 if (!sched_rt_can_attach(cgroup_tg(cgrp), task)) 7507 return -EINVAL; 7508#else 7509 /* We don't support RT-tasks being in separate groups */ 7510 if (task->sched_class != &fair_sched_class) 7511 return -EINVAL; 7512#endif 7513 } 7514 return 0; 7515} 7516 7517static void cpu_cgroup_attach(struct cgroup *cgrp, 7518 struct cgroup_taskset *tset) 7519{ 7520 struct task_struct *task; 7521 7522 cgroup_taskset_for_each(task, cgrp, tset) 7523 sched_move_task(task); 7524} 7525 7526static void 7527cpu_cgroup_exit(struct cgroup *cgrp, struct cgroup *old_cgrp, 7528 struct task_struct *task) 7529{ 7530 /* 7531 * cgroup_exit() is called in the copy_process() failure path. 7532 * Ignore this case since the task hasn't ran yet, this avoids 7533 * trying to poke a half freed task state from generic code. 7534 */ 7535 if (!(task->flags & PF_EXITING)) 7536 return; 7537 7538 sched_move_task(task); 7539} 7540 7541#ifdef CONFIG_FAIR_GROUP_SCHED 7542static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype, 7543 u64 shareval) 7544{ 7545 return sched_group_set_shares(cgroup_tg(cgrp), scale_load(shareval)); 7546} 7547 7548static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft) 7549{ 7550 struct task_group *tg = cgroup_tg(cgrp); 7551 7552 return (u64) scale_load_down(tg->shares); 7553} 7554 7555#ifdef CONFIG_CFS_BANDWIDTH 7556static DEFINE_MUTEX(cfs_constraints_mutex); 7557 7558const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */ 7559const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */ 7560 7561static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime); 7562 7563static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota) 7564{ 7565 int i, ret = 0, runtime_enabled, runtime_was_enabled; 7566 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth; 7567 7568 if (tg == &root_task_group) 7569 return -EINVAL; 7570 7571 /* 7572 * Ensure we have at some amount of bandwidth every period. This is 7573 * to prevent reaching a state of large arrears when throttled via 7574 * entity_tick() resulting in prolonged exit starvation. 7575 */ 7576 if (quota < min_cfs_quota_period || period < min_cfs_quota_period) 7577 return -EINVAL; 7578 7579 /* 7580 * Likewise, bound things on the otherside by preventing insane quota 7581 * periods. This also allows us to normalize in computing quota 7582 * feasibility. 7583 */ 7584 if (period > max_cfs_quota_period) 7585 return -EINVAL; 7586 7587 mutex_lock(&cfs_constraints_mutex); 7588 ret = __cfs_schedulable(tg, period, quota); 7589 if (ret) 7590 goto out_unlock; 7591 7592 runtime_enabled = quota != RUNTIME_INF; 7593 runtime_was_enabled = cfs_b->quota != RUNTIME_INF; 7594 account_cfs_bandwidth_used(runtime_enabled, runtime_was_enabled); 7595 raw_spin_lock_irq(&cfs_b->lock); 7596 cfs_b->period = ns_to_ktime(period); 7597 cfs_b->quota = quota; 7598 7599 __refill_cfs_bandwidth_runtime(cfs_b); 7600 /* restart the period timer (if active) to handle new period expiry */ 7601 if (runtime_enabled && cfs_b->timer_active) { 7602 /* force a reprogram */ 7603 cfs_b->timer_active = 0; 7604 __start_cfs_bandwidth(cfs_b); 7605 } 7606 raw_spin_unlock_irq(&cfs_b->lock); 7607 7608 for_each_possible_cpu(i) { 7609 struct cfs_rq *cfs_rq = tg->cfs_rq[i]; 7610 struct rq *rq = cfs_rq->rq; 7611 7612 raw_spin_lock_irq(&rq->lock); 7613 cfs_rq->runtime_enabled = runtime_enabled; 7614 cfs_rq->runtime_remaining = 0; 7615 7616 if (cfs_rq->throttled) 7617 unthrottle_cfs_rq(cfs_rq); 7618 raw_spin_unlock_irq(&rq->lock); 7619 } 7620out_unlock: 7621 mutex_unlock(&cfs_constraints_mutex); 7622 7623 return ret; 7624} 7625 7626int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us) 7627{ 7628 u64 quota, period; 7629 7630 period = ktime_to_ns(tg->cfs_bandwidth.period); 7631 if (cfs_quota_us < 0) 7632 quota = RUNTIME_INF; 7633 else 7634 quota = (u64)cfs_quota_us * NSEC_PER_USEC; 7635 7636 return tg_set_cfs_bandwidth(tg, period, quota); 7637} 7638 7639long tg_get_cfs_quota(struct task_group *tg) 7640{ 7641 u64 quota_us; 7642 7643 if (tg->cfs_bandwidth.quota == RUNTIME_INF) 7644 return -1; 7645 7646 quota_us = tg->cfs_bandwidth.quota; 7647 do_div(quota_us, NSEC_PER_USEC); 7648 7649 return quota_us; 7650} 7651 7652int tg_set_cfs_period(struct task_group *tg, long cfs_period_us) 7653{ 7654 u64 quota, period; 7655 7656 period = (u64)cfs_period_us * NSEC_PER_USEC; 7657 quota = tg->cfs_bandwidth.quota; 7658 7659 return tg_set_cfs_bandwidth(tg, period, quota); 7660} 7661 7662long tg_get_cfs_period(struct task_group *tg) 7663{ 7664 u64 cfs_period_us; 7665 7666 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period); 7667 do_div(cfs_period_us, NSEC_PER_USEC); 7668 7669 return cfs_period_us; 7670} 7671 7672static s64 cpu_cfs_quota_read_s64(struct cgroup *cgrp, struct cftype *cft) 7673{ 7674 return tg_get_cfs_quota(cgroup_tg(cgrp)); 7675} 7676 7677static int cpu_cfs_quota_write_s64(struct cgroup *cgrp, struct cftype *cftype, 7678 s64 cfs_quota_us) 7679{ 7680 return tg_set_cfs_quota(cgroup_tg(cgrp), cfs_quota_us); 7681} 7682 7683static u64 cpu_cfs_period_read_u64(struct cgroup *cgrp, struct cftype *cft) 7684{ 7685 return tg_get_cfs_period(cgroup_tg(cgrp)); 7686} 7687 7688static int cpu_cfs_period_write_u64(struct cgroup *cgrp, struct cftype *cftype, 7689 u64 cfs_period_us) 7690{ 7691 return tg_set_cfs_period(cgroup_tg(cgrp), cfs_period_us); 7692} 7693 7694struct cfs_schedulable_data { 7695 struct task_group *tg; 7696 u64 period, quota; 7697}; 7698 7699/* 7700 * normalize group quota/period to be quota/max_period 7701 * note: units are usecs 7702 */ 7703static u64 normalize_cfs_quota(struct task_group *tg, 7704 struct cfs_schedulable_data *d) 7705{ 7706 u64 quota, period; 7707 7708 if (tg == d->tg) { 7709 period = d->period; 7710 quota = d->quota; 7711 } else { 7712 period = tg_get_cfs_period(tg); 7713 quota = tg_get_cfs_quota(tg); 7714 } 7715 7716 /* note: these should typically be equivalent */ 7717 if (quota == RUNTIME_INF || quota == -1) 7718 return RUNTIME_INF; 7719 7720 return to_ratio(period, quota); 7721} 7722 7723static int tg_cfs_schedulable_down(struct task_group *tg, void *data) 7724{ 7725 struct cfs_schedulable_data *d = data; 7726 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth; 7727 s64 quota = 0, parent_quota = -1; 7728 7729 if (!tg->parent) { 7730 quota = RUNTIME_INF; 7731 } else { 7732 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth; 7733 7734 quota = normalize_cfs_quota(tg, d); 7735 parent_quota = parent_b->hierarchal_quota; 7736 7737 /* 7738 * ensure max(child_quota) <= parent_quota, inherit when no 7739 * limit is set 7740 */ 7741 if (quota == RUNTIME_INF) 7742 quota = parent_quota; 7743 else if (parent_quota != RUNTIME_INF && quota > parent_quota) 7744 return -EINVAL; 7745 } 7746 cfs_b->hierarchal_quota = quota; 7747 7748 return 0; 7749} 7750 7751static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota) 7752{ 7753 int ret; 7754 struct cfs_schedulable_data data = { 7755 .tg = tg, 7756 .period = period, 7757 .quota = quota, 7758 }; 7759 7760 if (quota != RUNTIME_INF) { 7761 do_div(data.period, NSEC_PER_USEC); 7762 do_div(data.quota, NSEC_PER_USEC); 7763 } 7764 7765 rcu_read_lock(); 7766 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data); 7767 rcu_read_unlock(); 7768 7769 return ret; 7770} 7771 7772static int cpu_stats_show(struct cgroup *cgrp, struct cftype *cft, 7773 struct cgroup_map_cb *cb) 7774{ 7775 struct task_group *tg = cgroup_tg(cgrp); 7776 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth; 7777 7778 cb->fill(cb, "nr_periods", cfs_b->nr_periods); 7779 cb->fill(cb, "nr_throttled", cfs_b->nr_throttled); 7780 cb->fill(cb, "throttled_time", cfs_b->throttled_time); 7781 7782 return 0; 7783} 7784#endif /* CONFIG_CFS_BANDWIDTH */ 7785#endif /* CONFIG_FAIR_GROUP_SCHED */ 7786 7787#ifdef CONFIG_RT_GROUP_SCHED 7788static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft, 7789 s64 val) 7790{ 7791 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val); 7792} 7793 7794static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft) 7795{ 7796 return sched_group_rt_runtime(cgroup_tg(cgrp)); 7797} 7798 7799static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype, 7800 u64 rt_period_us) 7801{ 7802 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us); 7803} 7804 7805static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft) 7806{ 7807 return sched_group_rt_period(cgroup_tg(cgrp)); 7808} 7809#endif /* CONFIG_RT_GROUP_SCHED */ 7810 7811static struct cftype cpu_files[] = { 7812#ifdef CONFIG_FAIR_GROUP_SCHED 7813 { 7814 .name = "shares", 7815 .read_u64 = cpu_shares_read_u64, 7816 .write_u64 = cpu_shares_write_u64, 7817 }, 7818#endif 7819#ifdef CONFIG_CFS_BANDWIDTH 7820 { 7821 .name = "cfs_quota_us", 7822 .read_s64 = cpu_cfs_quota_read_s64, 7823 .write_s64 = cpu_cfs_quota_write_s64, 7824 }, 7825 { 7826 .name = "cfs_period_us", 7827 .read_u64 = cpu_cfs_period_read_u64, 7828 .write_u64 = cpu_cfs_period_write_u64, 7829 }, 7830 { 7831 .name = "stat", 7832 .read_map = cpu_stats_show, 7833 }, 7834#endif 7835#ifdef CONFIG_RT_GROUP_SCHED 7836 { 7837 .name = "rt_runtime_us", 7838 .read_s64 = cpu_rt_runtime_read, 7839 .write_s64 = cpu_rt_runtime_write, 7840 }, 7841 { 7842 .name = "rt_period_us", 7843 .read_u64 = cpu_rt_period_read_uint, 7844 .write_u64 = cpu_rt_period_write_uint, 7845 }, 7846#endif 7847 { } /* terminate */ 7848}; 7849 7850struct cgroup_subsys cpu_cgroup_subsys = { 7851 .name = "cpu", 7852 .create = cpu_cgroup_create, 7853 .destroy = cpu_cgroup_destroy, 7854 .can_attach = cpu_cgroup_can_attach, 7855 .attach = cpu_cgroup_attach, 7856 .exit = cpu_cgroup_exit, 7857 .subsys_id = cpu_cgroup_subsys_id, 7858 .base_cftypes = cpu_files, 7859 .early_init = 1, 7860}; 7861 7862#endif /* CONFIG_CGROUP_SCHED */ 7863 7864#ifdef CONFIG_CGROUP_CPUACCT 7865 7866/* 7867 * CPU accounting code for task groups. 7868 * 7869 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh 7870 * (balbir@in.ibm.com). 7871 */ 7872 7873struct cpuacct root_cpuacct; 7874 7875/* create a new cpu accounting group */ 7876static struct cgroup_subsys_state *cpuacct_create(struct cgroup *cgrp) 7877{ 7878 struct cpuacct *ca; 7879 7880 if (!cgrp->parent) 7881 return &root_cpuacct.css; 7882 7883 ca = kzalloc(sizeof(*ca), GFP_KERNEL); 7884 if (!ca) 7885 goto out; 7886 7887 ca->cpuusage = alloc_percpu(u64); 7888 if (!ca->cpuusage) 7889 goto out_free_ca; 7890 7891 ca->cpustat = alloc_percpu(struct kernel_cpustat); 7892 if (!ca->cpustat) 7893 goto out_free_cpuusage; 7894 7895 return &ca->css; 7896 7897out_free_cpuusage: 7898 free_percpu(ca->cpuusage); 7899out_free_ca: 7900 kfree(ca); 7901out: 7902 return ERR_PTR(-ENOMEM); 7903} 7904 7905/* destroy an existing cpu accounting group */ 7906static void cpuacct_destroy(struct cgroup *cgrp) 7907{ 7908 struct cpuacct *ca = cgroup_ca(cgrp); 7909 7910 free_percpu(ca->cpustat); 7911 free_percpu(ca->cpuusage); 7912 kfree(ca); 7913} 7914 7915static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu) 7916{ 7917 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu); 7918 u64 data; 7919 7920#ifndef CONFIG_64BIT 7921 /* 7922 * Take rq->lock to make 64-bit read safe on 32-bit platforms. 7923 */ 7924 raw_spin_lock_irq(&cpu_rq(cpu)->lock); 7925 data = *cpuusage; 7926 raw_spin_unlock_irq(&cpu_rq(cpu)->lock); 7927#else 7928 data = *cpuusage; 7929#endif 7930 7931 return data; 7932} 7933 7934static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val) 7935{ 7936 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu); 7937 7938#ifndef CONFIG_64BIT 7939 /* 7940 * Take rq->lock to make 64-bit write safe on 32-bit platforms. 7941 */ 7942 raw_spin_lock_irq(&cpu_rq(cpu)->lock); 7943 *cpuusage = val; 7944 raw_spin_unlock_irq(&cpu_rq(cpu)->lock); 7945#else 7946 *cpuusage = val; 7947#endif 7948} 7949 7950/* return total cpu usage (in nanoseconds) of a group */ 7951static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft) 7952{ 7953 struct cpuacct *ca = cgroup_ca(cgrp); 7954 u64 totalcpuusage = 0; 7955 int i; 7956 7957 for_each_present_cpu(i) 7958 totalcpuusage += cpuacct_cpuusage_read(ca, i); 7959 7960 return totalcpuusage; 7961} 7962 7963static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype, 7964 u64 reset) 7965{ 7966 struct cpuacct *ca = cgroup_ca(cgrp); 7967 int err = 0; 7968 int i; 7969 7970 if (reset) { 7971 err = -EINVAL; 7972 goto out; 7973 } 7974 7975 for_each_present_cpu(i) 7976 cpuacct_cpuusage_write(ca, i, 0); 7977 7978out: 7979 return err; 7980} 7981 7982static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft, 7983 struct seq_file *m) 7984{ 7985 struct cpuacct *ca = cgroup_ca(cgroup); 7986 u64 percpu; 7987 int i; 7988 7989 for_each_present_cpu(i) { 7990 percpu = cpuacct_cpuusage_read(ca, i); 7991 seq_printf(m, "%llu ", (unsigned long long) percpu); 7992 } 7993 seq_printf(m, "\n"); 7994 return 0; 7995} 7996 7997static const char *cpuacct_stat_desc[] = { 7998 [CPUACCT_STAT_USER] = "user", 7999 [CPUACCT_STAT_SYSTEM] = "system", 8000}; 8001 8002static int cpuacct_stats_show(struct cgroup *cgrp, struct cftype *cft, 8003 struct cgroup_map_cb *cb) 8004{ 8005 struct cpuacct *ca = cgroup_ca(cgrp); 8006 int cpu; 8007 s64 val = 0; 8008 8009 for_each_online_cpu(cpu) { 8010 struct kernel_cpustat *kcpustat = per_cpu_ptr(ca->cpustat, cpu); 8011 val += kcpustat->cpustat[CPUTIME_USER]; 8012 val += kcpustat->cpustat[CPUTIME_NICE]; 8013 } 8014 val = cputime64_to_clock_t(val); 8015 cb->fill(cb, cpuacct_stat_desc[CPUACCT_STAT_USER], val); 8016 8017 val = 0; 8018 for_each_online_cpu(cpu) { 8019 struct kernel_cpustat *kcpustat = per_cpu_ptr(ca->cpustat, cpu); 8020 val += kcpustat->cpustat[CPUTIME_SYSTEM]; 8021 val += kcpustat->cpustat[CPUTIME_IRQ]; 8022 val += kcpustat->cpustat[CPUTIME_SOFTIRQ]; 8023 } 8024 8025 val = cputime64_to_clock_t(val); 8026 cb->fill(cb, cpuacct_stat_desc[CPUACCT_STAT_SYSTEM], val); 8027 8028 return 0; 8029} 8030 8031static struct cftype files[] = { 8032 { 8033 .name = "usage", 8034 .read_u64 = cpuusage_read, 8035 .write_u64 = cpuusage_write, 8036 }, 8037 { 8038 .name = "usage_percpu", 8039 .read_seq_string = cpuacct_percpu_seq_read, 8040 }, 8041 { 8042 .name = "stat", 8043 .read_map = cpuacct_stats_show, 8044 }, 8045 { } /* terminate */ 8046}; 8047 8048/* 8049 * charge this task's execution time to its accounting group. 8050 * 8051 * called with rq->lock held. 8052 */ 8053void cpuacct_charge(struct task_struct *tsk, u64 cputime) 8054{ 8055 struct cpuacct *ca; 8056 int cpu; 8057 8058 if (unlikely(!cpuacct_subsys.active)) 8059 return; 8060 8061 cpu = task_cpu(tsk); 8062 8063 rcu_read_lock(); 8064 8065 ca = task_ca(tsk); 8066 8067 for (; ca; ca = parent_ca(ca)) { 8068 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu); 8069 *cpuusage += cputime; 8070 } 8071 8072 rcu_read_unlock(); 8073} 8074 8075struct cgroup_subsys cpuacct_subsys = { 8076 .name = "cpuacct", 8077 .create = cpuacct_create, 8078 .destroy = cpuacct_destroy, 8079 .subsys_id = cpuacct_subsys_id, 8080 .base_cftypes = files, 8081}; 8082#endif /* CONFIG_CGROUP_CPUACCT */ 8083