core.c revision 201c373e8e4823700d3160d5c28e1ab18fd1193e
1/* 2 * kernel/sched/core.c 3 * 4 * Kernel scheduler and related syscalls 5 * 6 * Copyright (C) 1991-2002 Linus Torvalds 7 * 8 * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and 9 * make semaphores SMP safe 10 * 1998-11-19 Implemented schedule_timeout() and related stuff 11 * by Andrea Arcangeli 12 * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar: 13 * hybrid priority-list and round-robin design with 14 * an array-switch method of distributing timeslices 15 * and per-CPU runqueues. Cleanups and useful suggestions 16 * by Davide Libenzi, preemptible kernel bits by Robert Love. 17 * 2003-09-03 Interactivity tuning by Con Kolivas. 18 * 2004-04-02 Scheduler domains code by Nick Piggin 19 * 2007-04-15 Work begun on replacing all interactivity tuning with a 20 * fair scheduling design by Con Kolivas. 21 * 2007-05-05 Load balancing (smp-nice) and other improvements 22 * by Peter Williams 23 * 2007-05-06 Interactivity improvements to CFS by Mike Galbraith 24 * 2007-07-01 Group scheduling enhancements by Srivatsa Vaddagiri 25 * 2007-11-29 RT balancing improvements by Steven Rostedt, Gregory Haskins, 26 * Thomas Gleixner, Mike Kravetz 27 */ 28 29#include <linux/mm.h> 30#include <linux/module.h> 31#include <linux/nmi.h> 32#include <linux/init.h> 33#include <linux/uaccess.h> 34#include <linux/highmem.h> 35#include <asm/mmu_context.h> 36#include <linux/interrupt.h> 37#include <linux/capability.h> 38#include <linux/completion.h> 39#include <linux/kernel_stat.h> 40#include <linux/debug_locks.h> 41#include <linux/perf_event.h> 42#include <linux/security.h> 43#include <linux/notifier.h> 44#include <linux/profile.h> 45#include <linux/freezer.h> 46#include <linux/vmalloc.h> 47#include <linux/blkdev.h> 48#include <linux/delay.h> 49#include <linux/pid_namespace.h> 50#include <linux/smp.h> 51#include <linux/threads.h> 52#include <linux/timer.h> 53#include <linux/rcupdate.h> 54#include <linux/cpu.h> 55#include <linux/cpuset.h> 56#include <linux/percpu.h> 57#include <linux/proc_fs.h> 58#include <linux/seq_file.h> 59#include <linux/sysctl.h> 60#include <linux/syscalls.h> 61#include <linux/times.h> 62#include <linux/tsacct_kern.h> 63#include <linux/kprobes.h> 64#include <linux/delayacct.h> 65#include <linux/unistd.h> 66#include <linux/pagemap.h> 67#include <linux/hrtimer.h> 68#include <linux/tick.h> 69#include <linux/debugfs.h> 70#include <linux/ctype.h> 71#include <linux/ftrace.h> 72#include <linux/slab.h> 73#include <linux/init_task.h> 74#include <linux/binfmts.h> 75 76#include <asm/switch_to.h> 77#include <asm/tlb.h> 78#include <asm/irq_regs.h> 79#include <asm/mutex.h> 80#ifdef CONFIG_PARAVIRT 81#include <asm/paravirt.h> 82#endif 83 84#include "sched.h" 85#include "../workqueue_sched.h" 86#include "../smpboot.h" 87 88#define CREATE_TRACE_POINTS 89#include <trace/events/sched.h> 90 91void start_bandwidth_timer(struct hrtimer *period_timer, ktime_t period) 92{ 93 unsigned long delta; 94 ktime_t soft, hard, now; 95 96 for (;;) { 97 if (hrtimer_active(period_timer)) 98 break; 99 100 now = hrtimer_cb_get_time(period_timer); 101 hrtimer_forward(period_timer, now, period); 102 103 soft = hrtimer_get_softexpires(period_timer); 104 hard = hrtimer_get_expires(period_timer); 105 delta = ktime_to_ns(ktime_sub(hard, soft)); 106 __hrtimer_start_range_ns(period_timer, soft, delta, 107 HRTIMER_MODE_ABS_PINNED, 0); 108 } 109} 110 111DEFINE_MUTEX(sched_domains_mutex); 112DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues); 113 114static void update_rq_clock_task(struct rq *rq, s64 delta); 115 116void update_rq_clock(struct rq *rq) 117{ 118 s64 delta; 119 120 if (rq->skip_clock_update > 0) 121 return; 122 123 delta = sched_clock_cpu(cpu_of(rq)) - rq->clock; 124 rq->clock += delta; 125 update_rq_clock_task(rq, delta); 126} 127 128/* 129 * Debugging: various feature bits 130 */ 131 132#define SCHED_FEAT(name, enabled) \ 133 (1UL << __SCHED_FEAT_##name) * enabled | 134 135const_debug unsigned int sysctl_sched_features = 136#include "features.h" 137 0; 138 139#undef SCHED_FEAT 140 141#ifdef CONFIG_SCHED_DEBUG 142#define SCHED_FEAT(name, enabled) \ 143 #name , 144 145static const char * const sched_feat_names[] = { 146#include "features.h" 147}; 148 149#undef SCHED_FEAT 150 151static int sched_feat_show(struct seq_file *m, void *v) 152{ 153 int i; 154 155 for (i = 0; i < __SCHED_FEAT_NR; i++) { 156 if (!(sysctl_sched_features & (1UL << i))) 157 seq_puts(m, "NO_"); 158 seq_printf(m, "%s ", sched_feat_names[i]); 159 } 160 seq_puts(m, "\n"); 161 162 return 0; 163} 164 165#ifdef HAVE_JUMP_LABEL 166 167#define jump_label_key__true STATIC_KEY_INIT_TRUE 168#define jump_label_key__false STATIC_KEY_INIT_FALSE 169 170#define SCHED_FEAT(name, enabled) \ 171 jump_label_key__##enabled , 172 173struct static_key sched_feat_keys[__SCHED_FEAT_NR] = { 174#include "features.h" 175}; 176 177#undef SCHED_FEAT 178 179static void sched_feat_disable(int i) 180{ 181 if (static_key_enabled(&sched_feat_keys[i])) 182 static_key_slow_dec(&sched_feat_keys[i]); 183} 184 185static void sched_feat_enable(int i) 186{ 187 if (!static_key_enabled(&sched_feat_keys[i])) 188 static_key_slow_inc(&sched_feat_keys[i]); 189} 190#else 191static void sched_feat_disable(int i) { }; 192static void sched_feat_enable(int i) { }; 193#endif /* HAVE_JUMP_LABEL */ 194 195static ssize_t 196sched_feat_write(struct file *filp, const char __user *ubuf, 197 size_t cnt, loff_t *ppos) 198{ 199 char buf[64]; 200 char *cmp; 201 int neg = 0; 202 int i; 203 204 if (cnt > 63) 205 cnt = 63; 206 207 if (copy_from_user(&buf, ubuf, cnt)) 208 return -EFAULT; 209 210 buf[cnt] = 0; 211 cmp = strstrip(buf); 212 213 if (strncmp(cmp, "NO_", 3) == 0) { 214 neg = 1; 215 cmp += 3; 216 } 217 218 for (i = 0; i < __SCHED_FEAT_NR; i++) { 219 if (strcmp(cmp, sched_feat_names[i]) == 0) { 220 if (neg) { 221 sysctl_sched_features &= ~(1UL << i); 222 sched_feat_disable(i); 223 } else { 224 sysctl_sched_features |= (1UL << i); 225 sched_feat_enable(i); 226 } 227 break; 228 } 229 } 230 231 if (i == __SCHED_FEAT_NR) 232 return -EINVAL; 233 234 *ppos += cnt; 235 236 return cnt; 237} 238 239static int sched_feat_open(struct inode *inode, struct file *filp) 240{ 241 return single_open(filp, sched_feat_show, NULL); 242} 243 244static const struct file_operations sched_feat_fops = { 245 .open = sched_feat_open, 246 .write = sched_feat_write, 247 .read = seq_read, 248 .llseek = seq_lseek, 249 .release = single_release, 250}; 251 252static __init int sched_init_debug(void) 253{ 254 debugfs_create_file("sched_features", 0644, NULL, NULL, 255 &sched_feat_fops); 256 257 return 0; 258} 259late_initcall(sched_init_debug); 260#endif /* CONFIG_SCHED_DEBUG */ 261 262/* 263 * Number of tasks to iterate in a single balance run. 264 * Limited because this is done with IRQs disabled. 265 */ 266const_debug unsigned int sysctl_sched_nr_migrate = 32; 267 268/* 269 * period over which we average the RT time consumption, measured 270 * in ms. 271 * 272 * default: 1s 273 */ 274const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC; 275 276/* 277 * period over which we measure -rt task cpu usage in us. 278 * default: 1s 279 */ 280unsigned int sysctl_sched_rt_period = 1000000; 281 282__read_mostly int scheduler_running; 283 284/* 285 * part of the period that we allow rt tasks to run in us. 286 * default: 0.95s 287 */ 288int sysctl_sched_rt_runtime = 950000; 289 290 291 292/* 293 * __task_rq_lock - lock the rq @p resides on. 294 */ 295static inline struct rq *__task_rq_lock(struct task_struct *p) 296 __acquires(rq->lock) 297{ 298 struct rq *rq; 299 300 lockdep_assert_held(&p->pi_lock); 301 302 for (;;) { 303 rq = task_rq(p); 304 raw_spin_lock(&rq->lock); 305 if (likely(rq == task_rq(p))) 306 return rq; 307 raw_spin_unlock(&rq->lock); 308 } 309} 310 311/* 312 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on. 313 */ 314static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags) 315 __acquires(p->pi_lock) 316 __acquires(rq->lock) 317{ 318 struct rq *rq; 319 320 for (;;) { 321 raw_spin_lock_irqsave(&p->pi_lock, *flags); 322 rq = task_rq(p); 323 raw_spin_lock(&rq->lock); 324 if (likely(rq == task_rq(p))) 325 return rq; 326 raw_spin_unlock(&rq->lock); 327 raw_spin_unlock_irqrestore(&p->pi_lock, *flags); 328 } 329} 330 331static void __task_rq_unlock(struct rq *rq) 332 __releases(rq->lock) 333{ 334 raw_spin_unlock(&rq->lock); 335} 336 337static inline void 338task_rq_unlock(struct rq *rq, struct task_struct *p, unsigned long *flags) 339 __releases(rq->lock) 340 __releases(p->pi_lock) 341{ 342 raw_spin_unlock(&rq->lock); 343 raw_spin_unlock_irqrestore(&p->pi_lock, *flags); 344} 345 346/* 347 * this_rq_lock - lock this runqueue and disable interrupts. 348 */ 349static struct rq *this_rq_lock(void) 350 __acquires(rq->lock) 351{ 352 struct rq *rq; 353 354 local_irq_disable(); 355 rq = this_rq(); 356 raw_spin_lock(&rq->lock); 357 358 return rq; 359} 360 361#ifdef CONFIG_SCHED_HRTICK 362/* 363 * Use HR-timers to deliver accurate preemption points. 364 * 365 * Its all a bit involved since we cannot program an hrt while holding the 366 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a 367 * reschedule event. 368 * 369 * When we get rescheduled we reprogram the hrtick_timer outside of the 370 * rq->lock. 371 */ 372 373static void hrtick_clear(struct rq *rq) 374{ 375 if (hrtimer_active(&rq->hrtick_timer)) 376 hrtimer_cancel(&rq->hrtick_timer); 377} 378 379/* 380 * High-resolution timer tick. 381 * Runs from hardirq context with interrupts disabled. 382 */ 383static enum hrtimer_restart hrtick(struct hrtimer *timer) 384{ 385 struct rq *rq = container_of(timer, struct rq, hrtick_timer); 386 387 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id()); 388 389 raw_spin_lock(&rq->lock); 390 update_rq_clock(rq); 391 rq->curr->sched_class->task_tick(rq, rq->curr, 1); 392 raw_spin_unlock(&rq->lock); 393 394 return HRTIMER_NORESTART; 395} 396 397#ifdef CONFIG_SMP 398/* 399 * called from hardirq (IPI) context 400 */ 401static void __hrtick_start(void *arg) 402{ 403 struct rq *rq = arg; 404 405 raw_spin_lock(&rq->lock); 406 hrtimer_restart(&rq->hrtick_timer); 407 rq->hrtick_csd_pending = 0; 408 raw_spin_unlock(&rq->lock); 409} 410 411/* 412 * Called to set the hrtick timer state. 413 * 414 * called with rq->lock held and irqs disabled 415 */ 416void hrtick_start(struct rq *rq, u64 delay) 417{ 418 struct hrtimer *timer = &rq->hrtick_timer; 419 ktime_t time = ktime_add_ns(timer->base->get_time(), delay); 420 421 hrtimer_set_expires(timer, time); 422 423 if (rq == this_rq()) { 424 hrtimer_restart(timer); 425 } else if (!rq->hrtick_csd_pending) { 426 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd, 0); 427 rq->hrtick_csd_pending = 1; 428 } 429} 430 431static int 432hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu) 433{ 434 int cpu = (int)(long)hcpu; 435 436 switch (action) { 437 case CPU_UP_CANCELED: 438 case CPU_UP_CANCELED_FROZEN: 439 case CPU_DOWN_PREPARE: 440 case CPU_DOWN_PREPARE_FROZEN: 441 case CPU_DEAD: 442 case CPU_DEAD_FROZEN: 443 hrtick_clear(cpu_rq(cpu)); 444 return NOTIFY_OK; 445 } 446 447 return NOTIFY_DONE; 448} 449 450static __init void init_hrtick(void) 451{ 452 hotcpu_notifier(hotplug_hrtick, 0); 453} 454#else 455/* 456 * Called to set the hrtick timer state. 457 * 458 * called with rq->lock held and irqs disabled 459 */ 460void hrtick_start(struct rq *rq, u64 delay) 461{ 462 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0, 463 HRTIMER_MODE_REL_PINNED, 0); 464} 465 466static inline void init_hrtick(void) 467{ 468} 469#endif /* CONFIG_SMP */ 470 471static void init_rq_hrtick(struct rq *rq) 472{ 473#ifdef CONFIG_SMP 474 rq->hrtick_csd_pending = 0; 475 476 rq->hrtick_csd.flags = 0; 477 rq->hrtick_csd.func = __hrtick_start; 478 rq->hrtick_csd.info = rq; 479#endif 480 481 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); 482 rq->hrtick_timer.function = hrtick; 483} 484#else /* CONFIG_SCHED_HRTICK */ 485static inline void hrtick_clear(struct rq *rq) 486{ 487} 488 489static inline void init_rq_hrtick(struct rq *rq) 490{ 491} 492 493static inline void init_hrtick(void) 494{ 495} 496#endif /* CONFIG_SCHED_HRTICK */ 497 498/* 499 * resched_task - mark a task 'to be rescheduled now'. 500 * 501 * On UP this means the setting of the need_resched flag, on SMP it 502 * might also involve a cross-CPU call to trigger the scheduler on 503 * the target CPU. 504 */ 505#ifdef CONFIG_SMP 506 507#ifndef tsk_is_polling 508#define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG) 509#endif 510 511void resched_task(struct task_struct *p) 512{ 513 int cpu; 514 515 assert_raw_spin_locked(&task_rq(p)->lock); 516 517 if (test_tsk_need_resched(p)) 518 return; 519 520 set_tsk_need_resched(p); 521 522 cpu = task_cpu(p); 523 if (cpu == smp_processor_id()) 524 return; 525 526 /* NEED_RESCHED must be visible before we test polling */ 527 smp_mb(); 528 if (!tsk_is_polling(p)) 529 smp_send_reschedule(cpu); 530} 531 532void resched_cpu(int cpu) 533{ 534 struct rq *rq = cpu_rq(cpu); 535 unsigned long flags; 536 537 if (!raw_spin_trylock_irqsave(&rq->lock, flags)) 538 return; 539 resched_task(cpu_curr(cpu)); 540 raw_spin_unlock_irqrestore(&rq->lock, flags); 541} 542 543#ifdef CONFIG_NO_HZ 544/* 545 * In the semi idle case, use the nearest busy cpu for migrating timers 546 * from an idle cpu. This is good for power-savings. 547 * 548 * We don't do similar optimization for completely idle system, as 549 * selecting an idle cpu will add more delays to the timers than intended 550 * (as that cpu's timer base may not be uptodate wrt jiffies etc). 551 */ 552int get_nohz_timer_target(void) 553{ 554 int cpu = smp_processor_id(); 555 int i; 556 struct sched_domain *sd; 557 558 rcu_read_lock(); 559 for_each_domain(cpu, sd) { 560 for_each_cpu(i, sched_domain_span(sd)) { 561 if (!idle_cpu(i)) { 562 cpu = i; 563 goto unlock; 564 } 565 } 566 } 567unlock: 568 rcu_read_unlock(); 569 return cpu; 570} 571/* 572 * When add_timer_on() enqueues a timer into the timer wheel of an 573 * idle CPU then this timer might expire before the next timer event 574 * which is scheduled to wake up that CPU. In case of a completely 575 * idle system the next event might even be infinite time into the 576 * future. wake_up_idle_cpu() ensures that the CPU is woken up and 577 * leaves the inner idle loop so the newly added timer is taken into 578 * account when the CPU goes back to idle and evaluates the timer 579 * wheel for the next timer event. 580 */ 581void wake_up_idle_cpu(int cpu) 582{ 583 struct rq *rq = cpu_rq(cpu); 584 585 if (cpu == smp_processor_id()) 586 return; 587 588 /* 589 * This is safe, as this function is called with the timer 590 * wheel base lock of (cpu) held. When the CPU is on the way 591 * to idle and has not yet set rq->curr to idle then it will 592 * be serialized on the timer wheel base lock and take the new 593 * timer into account automatically. 594 */ 595 if (rq->curr != rq->idle) 596 return; 597 598 /* 599 * We can set TIF_RESCHED on the idle task of the other CPU 600 * lockless. The worst case is that the other CPU runs the 601 * idle task through an additional NOOP schedule() 602 */ 603 set_tsk_need_resched(rq->idle); 604 605 /* NEED_RESCHED must be visible before we test polling */ 606 smp_mb(); 607 if (!tsk_is_polling(rq->idle)) 608 smp_send_reschedule(cpu); 609} 610 611static inline bool got_nohz_idle_kick(void) 612{ 613 int cpu = smp_processor_id(); 614 return idle_cpu(cpu) && test_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu)); 615} 616 617#else /* CONFIG_NO_HZ */ 618 619static inline bool got_nohz_idle_kick(void) 620{ 621 return false; 622} 623 624#endif /* CONFIG_NO_HZ */ 625 626void sched_avg_update(struct rq *rq) 627{ 628 s64 period = sched_avg_period(); 629 630 while ((s64)(rq->clock - rq->age_stamp) > period) { 631 /* 632 * Inline assembly required to prevent the compiler 633 * optimising this loop into a divmod call. 634 * See __iter_div_u64_rem() for another example of this. 635 */ 636 asm("" : "+rm" (rq->age_stamp)); 637 rq->age_stamp += period; 638 rq->rt_avg /= 2; 639 } 640} 641 642#else /* !CONFIG_SMP */ 643void resched_task(struct task_struct *p) 644{ 645 assert_raw_spin_locked(&task_rq(p)->lock); 646 set_tsk_need_resched(p); 647} 648#endif /* CONFIG_SMP */ 649 650#if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \ 651 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH))) 652/* 653 * Iterate task_group tree rooted at *from, calling @down when first entering a 654 * node and @up when leaving it for the final time. 655 * 656 * Caller must hold rcu_lock or sufficient equivalent. 657 */ 658int walk_tg_tree_from(struct task_group *from, 659 tg_visitor down, tg_visitor up, void *data) 660{ 661 struct task_group *parent, *child; 662 int ret; 663 664 parent = from; 665 666down: 667 ret = (*down)(parent, data); 668 if (ret) 669 goto out; 670 list_for_each_entry_rcu(child, &parent->children, siblings) { 671 parent = child; 672 goto down; 673 674up: 675 continue; 676 } 677 ret = (*up)(parent, data); 678 if (ret || parent == from) 679 goto out; 680 681 child = parent; 682 parent = parent->parent; 683 if (parent) 684 goto up; 685out: 686 return ret; 687} 688 689int tg_nop(struct task_group *tg, void *data) 690{ 691 return 0; 692} 693#endif 694 695static void set_load_weight(struct task_struct *p) 696{ 697 int prio = p->static_prio - MAX_RT_PRIO; 698 struct load_weight *load = &p->se.load; 699 700 /* 701 * SCHED_IDLE tasks get minimal weight: 702 */ 703 if (p->policy == SCHED_IDLE) { 704 load->weight = scale_load(WEIGHT_IDLEPRIO); 705 load->inv_weight = WMULT_IDLEPRIO; 706 return; 707 } 708 709 load->weight = scale_load(prio_to_weight[prio]); 710 load->inv_weight = prio_to_wmult[prio]; 711} 712 713static void enqueue_task(struct rq *rq, struct task_struct *p, int flags) 714{ 715 update_rq_clock(rq); 716 sched_info_queued(p); 717 p->sched_class->enqueue_task(rq, p, flags); 718} 719 720static void dequeue_task(struct rq *rq, struct task_struct *p, int flags) 721{ 722 update_rq_clock(rq); 723 sched_info_dequeued(p); 724 p->sched_class->dequeue_task(rq, p, flags); 725} 726 727void activate_task(struct rq *rq, struct task_struct *p, int flags) 728{ 729 if (task_contributes_to_load(p)) 730 rq->nr_uninterruptible--; 731 732 enqueue_task(rq, p, flags); 733} 734 735void deactivate_task(struct rq *rq, struct task_struct *p, int flags) 736{ 737 if (task_contributes_to_load(p)) 738 rq->nr_uninterruptible++; 739 740 dequeue_task(rq, p, flags); 741} 742 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 1364#ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW 1365static int ttwu_activate_remote(struct task_struct *p, int wake_flags) 1366{ 1367 struct rq *rq; 1368 int ret = 0; 1369 1370 rq = __task_rq_lock(p); 1371 if (p->on_cpu) { 1372 ttwu_activate(rq, p, ENQUEUE_WAKEUP); 1373 ttwu_do_wakeup(rq, p, wake_flags); 1374 ret = 1; 1375 } 1376 __task_rq_unlock(rq); 1377 1378 return ret; 1379 1380} 1381#endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */ 1382 1383bool cpus_share_cache(int this_cpu, int that_cpu) 1384{ 1385 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu); 1386} 1387#endif /* CONFIG_SMP */ 1388 1389static void ttwu_queue(struct task_struct *p, int cpu) 1390{ 1391 struct rq *rq = cpu_rq(cpu); 1392 1393#if defined(CONFIG_SMP) 1394 if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) { 1395 sched_clock_cpu(cpu); /* sync clocks x-cpu */ 1396 ttwu_queue_remote(p, cpu); 1397 return; 1398 } 1399#endif 1400 1401 raw_spin_lock(&rq->lock); 1402 ttwu_do_activate(rq, p, 0); 1403 raw_spin_unlock(&rq->lock); 1404} 1405 1406/** 1407 * try_to_wake_up - wake up a thread 1408 * @p: the thread to be awakened 1409 * @state: the mask of task states that can be woken 1410 * @wake_flags: wake modifier flags (WF_*) 1411 * 1412 * Put it on the run-queue if it's not already there. The "current" 1413 * thread is always on the run-queue (except when the actual 1414 * re-schedule is in progress), and as such you're allowed to do 1415 * the simpler "current->state = TASK_RUNNING" to mark yourself 1416 * runnable without the overhead of this. 1417 * 1418 * Returns %true if @p was woken up, %false if it was already running 1419 * or @state didn't match @p's state. 1420 */ 1421static int 1422try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags) 1423{ 1424 unsigned long flags; 1425 int cpu, success = 0; 1426 1427 smp_wmb(); 1428 raw_spin_lock_irqsave(&p->pi_lock, flags); 1429 if (!(p->state & state)) 1430 goto out; 1431 1432 success = 1; /* we're going to change ->state */ 1433 cpu = task_cpu(p); 1434 1435 if (p->on_rq && ttwu_remote(p, wake_flags)) 1436 goto stat; 1437 1438#ifdef CONFIG_SMP 1439 /* 1440 * If the owning (remote) cpu is still in the middle of schedule() with 1441 * this task as prev, wait until its done referencing the task. 1442 */ 1443 while (p->on_cpu) { 1444#ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW 1445 /* 1446 * In case the architecture enables interrupts in 1447 * context_switch(), we cannot busy wait, since that 1448 * would lead to deadlocks when an interrupt hits and 1449 * tries to wake up @prev. So bail and do a complete 1450 * remote wakeup. 1451 */ 1452 if (ttwu_activate_remote(p, wake_flags)) 1453 goto stat; 1454#else 1455 cpu_relax(); 1456#endif 1457 } 1458 /* 1459 * Pairs with the smp_wmb() in finish_lock_switch(). 1460 */ 1461 smp_rmb(); 1462 1463 p->sched_contributes_to_load = !!task_contributes_to_load(p); 1464 p->state = TASK_WAKING; 1465 1466 if (p->sched_class->task_waking) 1467 p->sched_class->task_waking(p); 1468 1469 cpu = select_task_rq(p, SD_BALANCE_WAKE, wake_flags); 1470 if (task_cpu(p) != cpu) { 1471 wake_flags |= WF_MIGRATED; 1472 set_task_cpu(p, cpu); 1473 } 1474#endif /* CONFIG_SMP */ 1475 1476 ttwu_queue(p, cpu); 1477stat: 1478 ttwu_stat(p, cpu, wake_flags); 1479out: 1480 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 1481 1482 return success; 1483} 1484 1485/** 1486 * try_to_wake_up_local - try to wake up a local task with rq lock held 1487 * @p: the thread to be awakened 1488 * 1489 * Put @p on the run-queue if it's not already there. The caller must 1490 * ensure that this_rq() is locked, @p is bound to this_rq() and not 1491 * the current task. 1492 */ 1493static void try_to_wake_up_local(struct task_struct *p) 1494{ 1495 struct rq *rq = task_rq(p); 1496 1497 BUG_ON(rq != this_rq()); 1498 BUG_ON(p == current); 1499 lockdep_assert_held(&rq->lock); 1500 1501 if (!raw_spin_trylock(&p->pi_lock)) { 1502 raw_spin_unlock(&rq->lock); 1503 raw_spin_lock(&p->pi_lock); 1504 raw_spin_lock(&rq->lock); 1505 } 1506 1507 if (!(p->state & TASK_NORMAL)) 1508 goto out; 1509 1510 if (!p->on_rq) 1511 ttwu_activate(rq, p, ENQUEUE_WAKEUP); 1512 1513 ttwu_do_wakeup(rq, p, 0); 1514 ttwu_stat(p, smp_processor_id(), 0); 1515out: 1516 raw_spin_unlock(&p->pi_lock); 1517} 1518 1519/** 1520 * wake_up_process - Wake up a specific process 1521 * @p: The process to be woken up. 1522 * 1523 * Attempt to wake up the nominated process and move it to the set of runnable 1524 * processes. Returns 1 if the process was woken up, 0 if it was already 1525 * running. 1526 * 1527 * It may be assumed that this function implies a write memory barrier before 1528 * changing the task state if and only if any tasks are woken up. 1529 */ 1530int wake_up_process(struct task_struct *p) 1531{ 1532 return try_to_wake_up(p, TASK_ALL, 0); 1533} 1534EXPORT_SYMBOL(wake_up_process); 1535 1536int wake_up_state(struct task_struct *p, unsigned int state) 1537{ 1538 return try_to_wake_up(p, state, 0); 1539} 1540 1541/* 1542 * Perform scheduler related setup for a newly forked process p. 1543 * p is forked by current. 1544 * 1545 * __sched_fork() is basic setup used by init_idle() too: 1546 */ 1547static void __sched_fork(struct task_struct *p) 1548{ 1549 p->on_rq = 0; 1550 1551 p->se.on_rq = 0; 1552 p->se.exec_start = 0; 1553 p->se.sum_exec_runtime = 0; 1554 p->se.prev_sum_exec_runtime = 0; 1555 p->se.nr_migrations = 0; 1556 p->se.vruntime = 0; 1557 INIT_LIST_HEAD(&p->se.group_node); 1558 1559#ifdef CONFIG_SCHEDSTATS 1560 memset(&p->se.statistics, 0, sizeof(p->se.statistics)); 1561#endif 1562 1563 INIT_LIST_HEAD(&p->rt.run_list); 1564 1565#ifdef CONFIG_PREEMPT_NOTIFIERS 1566 INIT_HLIST_HEAD(&p->preempt_notifiers); 1567#endif 1568} 1569 1570/* 1571 * fork()/clone()-time setup: 1572 */ 1573void sched_fork(struct task_struct *p) 1574{ 1575 unsigned long flags; 1576 int cpu = get_cpu(); 1577 1578 __sched_fork(p); 1579 /* 1580 * We mark the process as running here. This guarantees that 1581 * nobody will actually run it, and a signal or other external 1582 * event cannot wake it up and insert it on the runqueue either. 1583 */ 1584 p->state = TASK_RUNNING; 1585 1586 /* 1587 * Make sure we do not leak PI boosting priority to the child. 1588 */ 1589 p->prio = current->normal_prio; 1590 1591 /* 1592 * Revert to default priority/policy on fork if requested. 1593 */ 1594 if (unlikely(p->sched_reset_on_fork)) { 1595 if (task_has_rt_policy(p)) { 1596 p->policy = SCHED_NORMAL; 1597 p->static_prio = NICE_TO_PRIO(0); 1598 p->rt_priority = 0; 1599 } else if (PRIO_TO_NICE(p->static_prio) < 0) 1600 p->static_prio = NICE_TO_PRIO(0); 1601 1602 p->prio = p->normal_prio = __normal_prio(p); 1603 set_load_weight(p); 1604 1605 /* 1606 * We don't need the reset flag anymore after the fork. It has 1607 * fulfilled its duty: 1608 */ 1609 p->sched_reset_on_fork = 0; 1610 } 1611 1612 if (!rt_prio(p->prio)) 1613 p->sched_class = &fair_sched_class; 1614 1615 if (p->sched_class->task_fork) 1616 p->sched_class->task_fork(p); 1617 1618 /* 1619 * The child is not yet in the pid-hash so no cgroup attach races, 1620 * and the cgroup is pinned to this child due to cgroup_fork() 1621 * is ran before sched_fork(). 1622 * 1623 * Silence PROVE_RCU. 1624 */ 1625 raw_spin_lock_irqsave(&p->pi_lock, flags); 1626 set_task_cpu(p, cpu); 1627 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 1628 1629#if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT) 1630 if (likely(sched_info_on())) 1631 memset(&p->sched_info, 0, sizeof(p->sched_info)); 1632#endif 1633#if defined(CONFIG_SMP) 1634 p->on_cpu = 0; 1635#endif 1636#ifdef CONFIG_PREEMPT_COUNT 1637 /* Want to start with kernel preemption disabled. */ 1638 task_thread_info(p)->preempt_count = 1; 1639#endif 1640#ifdef CONFIG_SMP 1641 plist_node_init(&p->pushable_tasks, MAX_PRIO); 1642#endif 1643 1644 put_cpu(); 1645} 1646 1647/* 1648 * wake_up_new_task - wake up a newly created task for the first time. 1649 * 1650 * This function will do some initial scheduler statistics housekeeping 1651 * that must be done for every newly created context, then puts the task 1652 * on the runqueue and wakes it. 1653 */ 1654void wake_up_new_task(struct task_struct *p) 1655{ 1656 unsigned long flags; 1657 struct rq *rq; 1658 1659 raw_spin_lock_irqsave(&p->pi_lock, flags); 1660#ifdef CONFIG_SMP 1661 /* 1662 * Fork balancing, do it here and not earlier because: 1663 * - cpus_allowed can change in the fork path 1664 * - any previously selected cpu might disappear through hotplug 1665 */ 1666 set_task_cpu(p, select_task_rq(p, SD_BALANCE_FORK, 0)); 1667#endif 1668 1669 rq = __task_rq_lock(p); 1670 activate_task(rq, p, 0); 1671 p->on_rq = 1; 1672 trace_sched_wakeup_new(p, true); 1673 check_preempt_curr(rq, p, WF_FORK); 1674#ifdef CONFIG_SMP 1675 if (p->sched_class->task_woken) 1676 p->sched_class->task_woken(rq, p); 1677#endif 1678 task_rq_unlock(rq, p, &flags); 1679} 1680 1681#ifdef CONFIG_PREEMPT_NOTIFIERS 1682 1683/** 1684 * preempt_notifier_register - tell me when current is being preempted & rescheduled 1685 * @notifier: notifier struct to register 1686 */ 1687void preempt_notifier_register(struct preempt_notifier *notifier) 1688{ 1689 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers); 1690} 1691EXPORT_SYMBOL_GPL(preempt_notifier_register); 1692 1693/** 1694 * preempt_notifier_unregister - no longer interested in preemption notifications 1695 * @notifier: notifier struct to unregister 1696 * 1697 * This is safe to call from within a preemption notifier. 1698 */ 1699void preempt_notifier_unregister(struct preempt_notifier *notifier) 1700{ 1701 hlist_del(¬ifier->link); 1702} 1703EXPORT_SYMBOL_GPL(preempt_notifier_unregister); 1704 1705static void fire_sched_in_preempt_notifiers(struct task_struct *curr) 1706{ 1707 struct preempt_notifier *notifier; 1708 struct hlist_node *node; 1709 1710 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link) 1711 notifier->ops->sched_in(notifier, raw_smp_processor_id()); 1712} 1713 1714static void 1715fire_sched_out_preempt_notifiers(struct task_struct *curr, 1716 struct task_struct *next) 1717{ 1718 struct preempt_notifier *notifier; 1719 struct hlist_node *node; 1720 1721 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link) 1722 notifier->ops->sched_out(notifier, next); 1723} 1724 1725#else /* !CONFIG_PREEMPT_NOTIFIERS */ 1726 1727static void fire_sched_in_preempt_notifiers(struct task_struct *curr) 1728{ 1729} 1730 1731static void 1732fire_sched_out_preempt_notifiers(struct task_struct *curr, 1733 struct task_struct *next) 1734{ 1735} 1736 1737#endif /* CONFIG_PREEMPT_NOTIFIERS */ 1738 1739/** 1740 * prepare_task_switch - prepare to switch tasks 1741 * @rq: the runqueue preparing to switch 1742 * @prev: the current task that is being switched out 1743 * @next: the task we are going to switch to. 1744 * 1745 * This is called with the rq lock held and interrupts off. It must 1746 * be paired with a subsequent finish_task_switch after the context 1747 * switch. 1748 * 1749 * prepare_task_switch sets up locking and calls architecture specific 1750 * hooks. 1751 */ 1752static inline void 1753prepare_task_switch(struct rq *rq, struct task_struct *prev, 1754 struct task_struct *next) 1755{ 1756 trace_sched_switch(prev, next); 1757 sched_info_switch(prev, next); 1758 perf_event_task_sched_out(prev, next); 1759 fire_sched_out_preempt_notifiers(prev, next); 1760 prepare_lock_switch(rq, next); 1761 prepare_arch_switch(next); 1762} 1763 1764/** 1765 * finish_task_switch - clean up after a task-switch 1766 * @rq: runqueue associated with task-switch 1767 * @prev: the thread we just switched away from. 1768 * 1769 * finish_task_switch must be called after the context switch, paired 1770 * with a prepare_task_switch call before the context switch. 1771 * finish_task_switch will reconcile locking set up by prepare_task_switch, 1772 * and do any other architecture-specific cleanup actions. 1773 * 1774 * Note that we may have delayed dropping an mm in context_switch(). If 1775 * so, we finish that here outside of the runqueue lock. (Doing it 1776 * with the lock held can cause deadlocks; see schedule() for 1777 * details.) 1778 */ 1779static void finish_task_switch(struct rq *rq, struct task_struct *prev) 1780 __releases(rq->lock) 1781{ 1782 struct mm_struct *mm = rq->prev_mm; 1783 long prev_state; 1784 1785 rq->prev_mm = NULL; 1786 1787 /* 1788 * A task struct has one reference for the use as "current". 1789 * If a task dies, then it sets TASK_DEAD in tsk->state and calls 1790 * schedule one last time. The schedule call will never return, and 1791 * the scheduled task must drop that reference. 1792 * The test for TASK_DEAD must occur while the runqueue locks are 1793 * still held, otherwise prev could be scheduled on another cpu, die 1794 * there before we look at prev->state, and then the reference would 1795 * be dropped twice. 1796 * Manfred Spraul <manfred@colorfullife.com> 1797 */ 1798 prev_state = prev->state; 1799 account_switch_vtime(prev); 1800 finish_arch_switch(prev); 1801#ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW 1802 local_irq_disable(); 1803#endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */ 1804 perf_event_task_sched_in(prev, current); 1805#ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW 1806 local_irq_enable(); 1807#endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */ 1808 finish_lock_switch(rq, prev); 1809 finish_arch_post_lock_switch(); 1810 1811 fire_sched_in_preempt_notifiers(current); 1812 if (mm) 1813 mmdrop(mm); 1814 if (unlikely(prev_state == TASK_DEAD)) { 1815 /* 1816 * Remove function-return probe instances associated with this 1817 * task and put them back on the free list. 1818 */ 1819 kprobe_flush_task(prev); 1820 put_task_struct(prev); 1821 } 1822} 1823 1824#ifdef CONFIG_SMP 1825 1826/* assumes rq->lock is held */ 1827static inline void pre_schedule(struct rq *rq, struct task_struct *prev) 1828{ 1829 if (prev->sched_class->pre_schedule) 1830 prev->sched_class->pre_schedule(rq, prev); 1831} 1832 1833/* rq->lock is NOT held, but preemption is disabled */ 1834static inline void post_schedule(struct rq *rq) 1835{ 1836 if (rq->post_schedule) { 1837 unsigned long flags; 1838 1839 raw_spin_lock_irqsave(&rq->lock, flags); 1840 if (rq->curr->sched_class->post_schedule) 1841 rq->curr->sched_class->post_schedule(rq); 1842 raw_spin_unlock_irqrestore(&rq->lock, flags); 1843 1844 rq->post_schedule = 0; 1845 } 1846} 1847 1848#else 1849 1850static inline void pre_schedule(struct rq *rq, struct task_struct *p) 1851{ 1852} 1853 1854static inline void post_schedule(struct rq *rq) 1855{ 1856} 1857 1858#endif 1859 1860/** 1861 * schedule_tail - first thing a freshly forked thread must call. 1862 * @prev: the thread we just switched away from. 1863 */ 1864asmlinkage void schedule_tail(struct task_struct *prev) 1865 __releases(rq->lock) 1866{ 1867 struct rq *rq = this_rq(); 1868 1869 finish_task_switch(rq, prev); 1870 1871 /* 1872 * FIXME: do we need to worry about rq being invalidated by the 1873 * task_switch? 1874 */ 1875 post_schedule(rq); 1876 1877#ifdef __ARCH_WANT_UNLOCKED_CTXSW 1878 /* In this case, finish_task_switch does not reenable preemption */ 1879 preempt_enable(); 1880#endif 1881 if (current->set_child_tid) 1882 put_user(task_pid_vnr(current), current->set_child_tid); 1883} 1884 1885/* 1886 * context_switch - switch to the new MM and the new 1887 * thread's register state. 1888 */ 1889static inline void 1890context_switch(struct rq *rq, struct task_struct *prev, 1891 struct task_struct *next) 1892{ 1893 struct mm_struct *mm, *oldmm; 1894 1895 prepare_task_switch(rq, prev, next); 1896 1897 mm = next->mm; 1898 oldmm = prev->active_mm; 1899 /* 1900 * For paravirt, this is coupled with an exit in switch_to to 1901 * combine the page table reload and the switch backend into 1902 * one hypercall. 1903 */ 1904 arch_start_context_switch(prev); 1905 1906 if (!mm) { 1907 next->active_mm = oldmm; 1908 atomic_inc(&oldmm->mm_count); 1909 enter_lazy_tlb(oldmm, next); 1910 } else 1911 switch_mm(oldmm, mm, next); 1912 1913 if (!prev->mm) { 1914 prev->active_mm = NULL; 1915 rq->prev_mm = oldmm; 1916 } 1917 /* 1918 * Since the runqueue lock will be released by the next 1919 * task (which is an invalid locking op but in the case 1920 * of the scheduler it's an obvious special-case), so we 1921 * do an early lockdep release here: 1922 */ 1923#ifndef __ARCH_WANT_UNLOCKED_CTXSW 1924 spin_release(&rq->lock.dep_map, 1, _THIS_IP_); 1925#endif 1926 1927 /* Here we just switch the register state and the stack. */ 1928 switch_to(prev, next, prev); 1929 1930 barrier(); 1931 /* 1932 * this_rq must be evaluated again because prev may have moved 1933 * CPUs since it called schedule(), thus the 'rq' on its stack 1934 * frame will be invalid. 1935 */ 1936 finish_task_switch(this_rq(), prev); 1937} 1938 1939/* 1940 * nr_running, nr_uninterruptible and nr_context_switches: 1941 * 1942 * externally visible scheduler statistics: current number of runnable 1943 * threads, current number of uninterruptible-sleeping threads, total 1944 * number of context switches performed since bootup. 1945 */ 1946unsigned long nr_running(void) 1947{ 1948 unsigned long i, sum = 0; 1949 1950 for_each_online_cpu(i) 1951 sum += cpu_rq(i)->nr_running; 1952 1953 return sum; 1954} 1955 1956unsigned long nr_uninterruptible(void) 1957{ 1958 unsigned long i, sum = 0; 1959 1960 for_each_possible_cpu(i) 1961 sum += cpu_rq(i)->nr_uninterruptible; 1962 1963 /* 1964 * Since we read the counters lockless, it might be slightly 1965 * inaccurate. Do not allow it to go below zero though: 1966 */ 1967 if (unlikely((long)sum < 0)) 1968 sum = 0; 1969 1970 return sum; 1971} 1972 1973unsigned long long nr_context_switches(void) 1974{ 1975 int i; 1976 unsigned long long sum = 0; 1977 1978 for_each_possible_cpu(i) 1979 sum += cpu_rq(i)->nr_switches; 1980 1981 return sum; 1982} 1983 1984unsigned long nr_iowait(void) 1985{ 1986 unsigned long i, sum = 0; 1987 1988 for_each_possible_cpu(i) 1989 sum += atomic_read(&cpu_rq(i)->nr_iowait); 1990 1991 return sum; 1992} 1993 1994unsigned long nr_iowait_cpu(int cpu) 1995{ 1996 struct rq *this = cpu_rq(cpu); 1997 return atomic_read(&this->nr_iowait); 1998} 1999 2000unsigned long this_cpu_load(void) 2001{ 2002 struct rq *this = this_rq(); 2003 return this->cpu_load[0]; 2004} 2005 2006 2007/* 2008 * Global load-average calculations 2009 * 2010 * We take a distributed and async approach to calculating the global load-avg 2011 * in order to minimize overhead. 2012 * 2013 * The global load average is an exponentially decaying average of nr_running + 2014 * nr_uninterruptible. 2015 * 2016 * Once every LOAD_FREQ: 2017 * 2018 * nr_active = 0; 2019 * for_each_possible_cpu(cpu) 2020 * nr_active += cpu_of(cpu)->nr_running + cpu_of(cpu)->nr_uninterruptible; 2021 * 2022 * avenrun[n] = avenrun[0] * exp_n + nr_active * (1 - exp_n) 2023 * 2024 * Due to a number of reasons the above turns in the mess below: 2025 * 2026 * - for_each_possible_cpu() is prohibitively expensive on machines with 2027 * serious number of cpus, therefore we need to take a distributed approach 2028 * to calculating nr_active. 2029 * 2030 * \Sum_i x_i(t) = \Sum_i x_i(t) - x_i(t_0) | x_i(t_0) := 0 2031 * = \Sum_i { \Sum_j=1 x_i(t_j) - x_i(t_j-1) } 2032 * 2033 * So assuming nr_active := 0 when we start out -- true per definition, we 2034 * can simply take per-cpu deltas and fold those into a global accumulate 2035 * to obtain the same result. See calc_load_fold_active(). 2036 * 2037 * Furthermore, in order to avoid synchronizing all per-cpu delta folding 2038 * across the machine, we assume 10 ticks is sufficient time for every 2039 * cpu to have completed this task. 2040 * 2041 * This places an upper-bound on the IRQ-off latency of the machine. Then 2042 * again, being late doesn't loose the delta, just wrecks the sample. 2043 * 2044 * - cpu_rq()->nr_uninterruptible isn't accurately tracked per-cpu because 2045 * this would add another cross-cpu cacheline miss and atomic operation 2046 * to the wakeup path. Instead we increment on whatever cpu the task ran 2047 * when it went into uninterruptible state and decrement on whatever cpu 2048 * did the wakeup. This means that only the sum of nr_uninterruptible over 2049 * all cpus yields the correct result. 2050 * 2051 * This covers the NO_HZ=n code, for extra head-aches, see the comment below. 2052 */ 2053 2054/* Variables and functions for calc_load */ 2055static atomic_long_t calc_load_tasks; 2056static unsigned long calc_load_update; 2057unsigned long avenrun[3]; 2058EXPORT_SYMBOL(avenrun); /* should be removed */ 2059 2060/** 2061 * get_avenrun - get the load average array 2062 * @loads: pointer to dest load array 2063 * @offset: offset to add 2064 * @shift: shift count to shift the result left 2065 * 2066 * These values are estimates at best, so no need for locking. 2067 */ 2068void get_avenrun(unsigned long *loads, unsigned long offset, int shift) 2069{ 2070 loads[0] = (avenrun[0] + offset) << shift; 2071 loads[1] = (avenrun[1] + offset) << shift; 2072 loads[2] = (avenrun[2] + offset) << shift; 2073} 2074 2075static long calc_load_fold_active(struct rq *this_rq) 2076{ 2077 long nr_active, delta = 0; 2078 2079 nr_active = this_rq->nr_running; 2080 nr_active += (long) this_rq->nr_uninterruptible; 2081 2082 if (nr_active != this_rq->calc_load_active) { 2083 delta = nr_active - this_rq->calc_load_active; 2084 this_rq->calc_load_active = nr_active; 2085 } 2086 2087 return delta; 2088} 2089 2090/* 2091 * a1 = a0 * e + a * (1 - e) 2092 */ 2093static unsigned long 2094calc_load(unsigned long load, unsigned long exp, unsigned long active) 2095{ 2096 load *= exp; 2097 load += active * (FIXED_1 - exp); 2098 load += 1UL << (FSHIFT - 1); 2099 return load >> FSHIFT; 2100} 2101 2102#ifdef CONFIG_NO_HZ 2103/* 2104 * Handle NO_HZ for the global load-average. 2105 * 2106 * Since the above described distributed algorithm to compute the global 2107 * load-average relies on per-cpu sampling from the tick, it is affected by 2108 * NO_HZ. 2109 * 2110 * The basic idea is to fold the nr_active delta into a global idle-delta upon 2111 * entering NO_HZ state such that we can include this as an 'extra' cpu delta 2112 * when we read the global state. 2113 * 2114 * Obviously reality has to ruin such a delightfully simple scheme: 2115 * 2116 * - When we go NO_HZ idle during the window, we can negate our sample 2117 * contribution, causing under-accounting. 2118 * 2119 * We avoid this by keeping two idle-delta counters and flipping them 2120 * when the window starts, thus separating old and new NO_HZ load. 2121 * 2122 * The only trick is the slight shift in index flip for read vs write. 2123 * 2124 * 0s 5s 10s 15s 2125 * +10 +10 +10 +10 2126 * |-|-----------|-|-----------|-|-----------|-| 2127 * r:0 0 1 1 0 0 1 1 0 2128 * w:0 1 1 0 0 1 1 0 0 2129 * 2130 * This ensures we'll fold the old idle contribution in this window while 2131 * accumlating the new one. 2132 * 2133 * - When we wake up from NO_HZ idle during the window, we push up our 2134 * contribution, since we effectively move our sample point to a known 2135 * busy state. 2136 * 2137 * This is solved by pushing the window forward, and thus skipping the 2138 * sample, for this cpu (effectively using the idle-delta for this cpu which 2139 * was in effect at the time the window opened). This also solves the issue 2140 * of having to deal with a cpu having been in NOHZ idle for multiple 2141 * LOAD_FREQ intervals. 2142 * 2143 * When making the ILB scale, we should try to pull this in as well. 2144 */ 2145static atomic_long_t calc_load_idle[2]; 2146static int calc_load_idx; 2147 2148static inline int calc_load_write_idx(void) 2149{ 2150 int idx = calc_load_idx; 2151 2152 /* 2153 * See calc_global_nohz(), if we observe the new index, we also 2154 * need to observe the new update time. 2155 */ 2156 smp_rmb(); 2157 2158 /* 2159 * If the folding window started, make sure we start writing in the 2160 * next idle-delta. 2161 */ 2162 if (!time_before(jiffies, calc_load_update)) 2163 idx++; 2164 2165 return idx & 1; 2166} 2167 2168static inline int calc_load_read_idx(void) 2169{ 2170 return calc_load_idx & 1; 2171} 2172 2173void calc_load_enter_idle(void) 2174{ 2175 struct rq *this_rq = this_rq(); 2176 long delta; 2177 2178 /* 2179 * We're going into NOHZ mode, if there's any pending delta, fold it 2180 * into the pending idle delta. 2181 */ 2182 delta = calc_load_fold_active(this_rq); 2183 if (delta) { 2184 int idx = calc_load_write_idx(); 2185 atomic_long_add(delta, &calc_load_idle[idx]); 2186 } 2187} 2188 2189void calc_load_exit_idle(void) 2190{ 2191 struct rq *this_rq = this_rq(); 2192 2193 /* 2194 * If we're still before the sample window, we're done. 2195 */ 2196 if (time_before(jiffies, this_rq->calc_load_update)) 2197 return; 2198 2199 /* 2200 * We woke inside or after the sample window, this means we're already 2201 * accounted through the nohz accounting, so skip the entire deal and 2202 * sync up for the next window. 2203 */ 2204 this_rq->calc_load_update = calc_load_update; 2205 if (time_before(jiffies, this_rq->calc_load_update + 10)) 2206 this_rq->calc_load_update += LOAD_FREQ; 2207} 2208 2209static long calc_load_fold_idle(void) 2210{ 2211 int idx = calc_load_read_idx(); 2212 long delta = 0; 2213 2214 if (atomic_long_read(&calc_load_idle[idx])) 2215 delta = atomic_long_xchg(&calc_load_idle[idx], 0); 2216 2217 return delta; 2218} 2219 2220/** 2221 * fixed_power_int - compute: x^n, in O(log n) time 2222 * 2223 * @x: base of the power 2224 * @frac_bits: fractional bits of @x 2225 * @n: power to raise @x to. 2226 * 2227 * By exploiting the relation between the definition of the natural power 2228 * function: x^n := x*x*...*x (x multiplied by itself for n times), and 2229 * the binary encoding of numbers used by computers: n := \Sum n_i * 2^i, 2230 * (where: n_i \elem {0, 1}, the binary vector representing n), 2231 * we find: x^n := x^(\Sum n_i * 2^i) := \Prod x^(n_i * 2^i), which is 2232 * of course trivially computable in O(log_2 n), the length of our binary 2233 * vector. 2234 */ 2235static unsigned long 2236fixed_power_int(unsigned long x, unsigned int frac_bits, unsigned int n) 2237{ 2238 unsigned long result = 1UL << frac_bits; 2239 2240 if (n) for (;;) { 2241 if (n & 1) { 2242 result *= x; 2243 result += 1UL << (frac_bits - 1); 2244 result >>= frac_bits; 2245 } 2246 n >>= 1; 2247 if (!n) 2248 break; 2249 x *= x; 2250 x += 1UL << (frac_bits - 1); 2251 x >>= frac_bits; 2252 } 2253 2254 return result; 2255} 2256 2257/* 2258 * a1 = a0 * e + a * (1 - e) 2259 * 2260 * a2 = a1 * e + a * (1 - e) 2261 * = (a0 * e + a * (1 - e)) * e + a * (1 - e) 2262 * = a0 * e^2 + a * (1 - e) * (1 + e) 2263 * 2264 * a3 = a2 * e + a * (1 - e) 2265 * = (a0 * e^2 + a * (1 - e) * (1 + e)) * e + a * (1 - e) 2266 * = a0 * e^3 + a * (1 - e) * (1 + e + e^2) 2267 * 2268 * ... 2269 * 2270 * an = a0 * e^n + a * (1 - e) * (1 + e + ... + e^n-1) [1] 2271 * = a0 * e^n + a * (1 - e) * (1 - e^n)/(1 - e) 2272 * = a0 * e^n + a * (1 - e^n) 2273 * 2274 * [1] application of the geometric series: 2275 * 2276 * n 1 - x^(n+1) 2277 * S_n := \Sum x^i = ------------- 2278 * i=0 1 - x 2279 */ 2280static unsigned long 2281calc_load_n(unsigned long load, unsigned long exp, 2282 unsigned long active, unsigned int n) 2283{ 2284 2285 return calc_load(load, fixed_power_int(exp, FSHIFT, n), active); 2286} 2287 2288/* 2289 * NO_HZ can leave us missing all per-cpu ticks calling 2290 * calc_load_account_active(), but since an idle CPU folds its delta into 2291 * calc_load_tasks_idle per calc_load_account_idle(), all we need to do is fold 2292 * in the pending idle delta if our idle period crossed a load cycle boundary. 2293 * 2294 * Once we've updated the global active value, we need to apply the exponential 2295 * weights adjusted to the number of cycles missed. 2296 */ 2297static void calc_global_nohz(void) 2298{ 2299 long delta, active, n; 2300 2301 if (!time_before(jiffies, calc_load_update + 10)) { 2302 /* 2303 * Catch-up, fold however many we are behind still 2304 */ 2305 delta = jiffies - calc_load_update - 10; 2306 n = 1 + (delta / LOAD_FREQ); 2307 2308 active = atomic_long_read(&calc_load_tasks); 2309 active = active > 0 ? active * FIXED_1 : 0; 2310 2311 avenrun[0] = calc_load_n(avenrun[0], EXP_1, active, n); 2312 avenrun[1] = calc_load_n(avenrun[1], EXP_5, active, n); 2313 avenrun[2] = calc_load_n(avenrun[2], EXP_15, active, n); 2314 2315 calc_load_update += n * LOAD_FREQ; 2316 } 2317 2318 /* 2319 * Flip the idle index... 2320 * 2321 * Make sure we first write the new time then flip the index, so that 2322 * calc_load_write_idx() will see the new time when it reads the new 2323 * index, this avoids a double flip messing things up. 2324 */ 2325 smp_wmb(); 2326 calc_load_idx++; 2327} 2328#else /* !CONFIG_NO_HZ */ 2329 2330static inline long calc_load_fold_idle(void) { return 0; } 2331static inline void calc_global_nohz(void) { } 2332 2333#endif /* CONFIG_NO_HZ */ 2334 2335/* 2336 * calc_load - update the avenrun load estimates 10 ticks after the 2337 * CPUs have updated calc_load_tasks. 2338 */ 2339void calc_global_load(unsigned long ticks) 2340{ 2341 long active, delta; 2342 2343 if (time_before(jiffies, calc_load_update + 10)) 2344 return; 2345 2346 /* 2347 * Fold the 'old' idle-delta to include all NO_HZ cpus. 2348 */ 2349 delta = calc_load_fold_idle(); 2350 if (delta) 2351 atomic_long_add(delta, &calc_load_tasks); 2352 2353 active = atomic_long_read(&calc_load_tasks); 2354 active = active > 0 ? active * FIXED_1 : 0; 2355 2356 avenrun[0] = calc_load(avenrun[0], EXP_1, active); 2357 avenrun[1] = calc_load(avenrun[1], EXP_5, active); 2358 avenrun[2] = calc_load(avenrun[2], EXP_15, active); 2359 2360 calc_load_update += LOAD_FREQ; 2361 2362 /* 2363 * In case we idled for multiple LOAD_FREQ intervals, catch up in bulk. 2364 */ 2365 calc_global_nohz(); 2366} 2367 2368/* 2369 * Called from update_cpu_load() to periodically update this CPU's 2370 * active count. 2371 */ 2372static void calc_load_account_active(struct rq *this_rq) 2373{ 2374 long delta; 2375 2376 if (time_before(jiffies, this_rq->calc_load_update)) 2377 return; 2378 2379 delta = calc_load_fold_active(this_rq); 2380 if (delta) 2381 atomic_long_add(delta, &calc_load_tasks); 2382 2383 this_rq->calc_load_update += LOAD_FREQ; 2384} 2385 2386/* 2387 * End of global load-average stuff 2388 */ 2389 2390/* 2391 * The exact cpuload at various idx values, calculated at every tick would be 2392 * load = (2^idx - 1) / 2^idx * load + 1 / 2^idx * cur_load 2393 * 2394 * If a cpu misses updates for n-1 ticks (as it was idle) and update gets called 2395 * on nth tick when cpu may be busy, then we have: 2396 * load = ((2^idx - 1) / 2^idx)^(n-1) * load 2397 * load = (2^idx - 1) / 2^idx) * load + 1 / 2^idx * cur_load 2398 * 2399 * decay_load_missed() below does efficient calculation of 2400 * load = ((2^idx - 1) / 2^idx)^(n-1) * load 2401 * avoiding 0..n-1 loop doing load = ((2^idx - 1) / 2^idx) * load 2402 * 2403 * The calculation is approximated on a 128 point scale. 2404 * degrade_zero_ticks is the number of ticks after which load at any 2405 * particular idx is approximated to be zero. 2406 * degrade_factor is a precomputed table, a row for each load idx. 2407 * Each column corresponds to degradation factor for a power of two ticks, 2408 * based on 128 point scale. 2409 * Example: 2410 * row 2, col 3 (=12) says that the degradation at load idx 2 after 2411 * 8 ticks is 12/128 (which is an approximation of exact factor 3^8/4^8). 2412 * 2413 * With this power of 2 load factors, we can degrade the load n times 2414 * by looking at 1 bits in n and doing as many mult/shift instead of 2415 * n mult/shifts needed by the exact degradation. 2416 */ 2417#define DEGRADE_SHIFT 7 2418static const unsigned char 2419 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128}; 2420static const unsigned char 2421 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = { 2422 {0, 0, 0, 0, 0, 0, 0, 0}, 2423 {64, 32, 8, 0, 0, 0, 0, 0}, 2424 {96, 72, 40, 12, 1, 0, 0}, 2425 {112, 98, 75, 43, 15, 1, 0}, 2426 {120, 112, 98, 76, 45, 16, 2} }; 2427 2428/* 2429 * Update cpu_load for any missed ticks, due to tickless idle. The backlog 2430 * would be when CPU is idle and so we just decay the old load without 2431 * adding any new load. 2432 */ 2433static unsigned long 2434decay_load_missed(unsigned long load, unsigned long missed_updates, int idx) 2435{ 2436 int j = 0; 2437 2438 if (!missed_updates) 2439 return load; 2440 2441 if (missed_updates >= degrade_zero_ticks[idx]) 2442 return 0; 2443 2444 if (idx == 1) 2445 return load >> missed_updates; 2446 2447 while (missed_updates) { 2448 if (missed_updates % 2) 2449 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT; 2450 2451 missed_updates >>= 1; 2452 j++; 2453 } 2454 return load; 2455} 2456 2457/* 2458 * Update rq->cpu_load[] statistics. This function is usually called every 2459 * scheduler tick (TICK_NSEC). With tickless idle this will not be called 2460 * every tick. We fix it up based on jiffies. 2461 */ 2462static void __update_cpu_load(struct rq *this_rq, unsigned long this_load, 2463 unsigned long pending_updates) 2464{ 2465 int i, scale; 2466 2467 this_rq->nr_load_updates++; 2468 2469 /* Update our load: */ 2470 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */ 2471 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) { 2472 unsigned long old_load, new_load; 2473 2474 /* scale is effectively 1 << i now, and >> i divides by scale */ 2475 2476 old_load = this_rq->cpu_load[i]; 2477 old_load = decay_load_missed(old_load, pending_updates - 1, i); 2478 new_load = this_load; 2479 /* 2480 * Round up the averaging division if load is increasing. This 2481 * prevents us from getting stuck on 9 if the load is 10, for 2482 * example. 2483 */ 2484 if (new_load > old_load) 2485 new_load += scale - 1; 2486 2487 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i; 2488 } 2489 2490 sched_avg_update(this_rq); 2491} 2492 2493#ifdef CONFIG_NO_HZ 2494/* 2495 * There is no sane way to deal with nohz on smp when using jiffies because the 2496 * cpu doing the jiffies update might drift wrt the cpu doing the jiffy reading 2497 * causing off-by-one errors in observed deltas; {0,2} instead of {1,1}. 2498 * 2499 * Therefore we cannot use the delta approach from the regular tick since that 2500 * would seriously skew the load calculation. However we'll make do for those 2501 * updates happening while idle (nohz_idle_balance) or coming out of idle 2502 * (tick_nohz_idle_exit). 2503 * 2504 * This means we might still be one tick off for nohz periods. 2505 */ 2506 2507/* 2508 * Called from nohz_idle_balance() to update the load ratings before doing the 2509 * idle balance. 2510 */ 2511void update_idle_cpu_load(struct rq *this_rq) 2512{ 2513 unsigned long curr_jiffies = ACCESS_ONCE(jiffies); 2514 unsigned long load = this_rq->load.weight; 2515 unsigned long pending_updates; 2516 2517 /* 2518 * bail if there's load or we're actually up-to-date. 2519 */ 2520 if (load || curr_jiffies == this_rq->last_load_update_tick) 2521 return; 2522 2523 pending_updates = curr_jiffies - this_rq->last_load_update_tick; 2524 this_rq->last_load_update_tick = curr_jiffies; 2525 2526 __update_cpu_load(this_rq, load, pending_updates); 2527} 2528 2529/* 2530 * Called from tick_nohz_idle_exit() -- try and fix up the ticks we missed. 2531 */ 2532void update_cpu_load_nohz(void) 2533{ 2534 struct rq *this_rq = this_rq(); 2535 unsigned long curr_jiffies = ACCESS_ONCE(jiffies); 2536 unsigned long pending_updates; 2537 2538 if (curr_jiffies == this_rq->last_load_update_tick) 2539 return; 2540 2541 raw_spin_lock(&this_rq->lock); 2542 pending_updates = curr_jiffies - this_rq->last_load_update_tick; 2543 if (pending_updates) { 2544 this_rq->last_load_update_tick = curr_jiffies; 2545 /* 2546 * We were idle, this means load 0, the current load might be 2547 * !0 due to remote wakeups and the sort. 2548 */ 2549 __update_cpu_load(this_rq, 0, pending_updates); 2550 } 2551 raw_spin_unlock(&this_rq->lock); 2552} 2553#endif /* CONFIG_NO_HZ */ 2554 2555/* 2556 * Called from scheduler_tick() 2557 */ 2558static void update_cpu_load_active(struct rq *this_rq) 2559{ 2560 /* 2561 * See the mess around update_idle_cpu_load() / update_cpu_load_nohz(). 2562 */ 2563 this_rq->last_load_update_tick = jiffies; 2564 __update_cpu_load(this_rq, this_rq->load.weight, 1); 2565 2566 calc_load_account_active(this_rq); 2567} 2568 2569#ifdef CONFIG_SMP 2570 2571/* 2572 * sched_exec - execve() is a valuable balancing opportunity, because at 2573 * this point the task has the smallest effective memory and cache footprint. 2574 */ 2575void sched_exec(void) 2576{ 2577 struct task_struct *p = current; 2578 unsigned long flags; 2579 int dest_cpu; 2580 2581 raw_spin_lock_irqsave(&p->pi_lock, flags); 2582 dest_cpu = p->sched_class->select_task_rq(p, SD_BALANCE_EXEC, 0); 2583 if (dest_cpu == smp_processor_id()) 2584 goto unlock; 2585 2586 if (likely(cpu_active(dest_cpu))) { 2587 struct migration_arg arg = { p, dest_cpu }; 2588 2589 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 2590 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg); 2591 return; 2592 } 2593unlock: 2594 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 2595} 2596 2597#endif 2598 2599DEFINE_PER_CPU(struct kernel_stat, kstat); 2600DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat); 2601 2602EXPORT_PER_CPU_SYMBOL(kstat); 2603EXPORT_PER_CPU_SYMBOL(kernel_cpustat); 2604 2605/* 2606 * Return any ns on the sched_clock that have not yet been accounted in 2607 * @p in case that task is currently running. 2608 * 2609 * Called with task_rq_lock() held on @rq. 2610 */ 2611static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq) 2612{ 2613 u64 ns = 0; 2614 2615 if (task_current(rq, p)) { 2616 update_rq_clock(rq); 2617 ns = rq->clock_task - p->se.exec_start; 2618 if ((s64)ns < 0) 2619 ns = 0; 2620 } 2621 2622 return ns; 2623} 2624 2625unsigned long long task_delta_exec(struct task_struct *p) 2626{ 2627 unsigned long flags; 2628 struct rq *rq; 2629 u64 ns = 0; 2630 2631 rq = task_rq_lock(p, &flags); 2632 ns = do_task_delta_exec(p, rq); 2633 task_rq_unlock(rq, p, &flags); 2634 2635 return ns; 2636} 2637 2638/* 2639 * Return accounted runtime for the task. 2640 * In case the task is currently running, return the runtime plus current's 2641 * pending runtime that have not been accounted yet. 2642 */ 2643unsigned long long task_sched_runtime(struct task_struct *p) 2644{ 2645 unsigned long flags; 2646 struct rq *rq; 2647 u64 ns = 0; 2648 2649 rq = task_rq_lock(p, &flags); 2650 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq); 2651 task_rq_unlock(rq, p, &flags); 2652 2653 return ns; 2654} 2655 2656/* 2657 * This function gets called by the timer code, with HZ frequency. 2658 * We call it with interrupts disabled. 2659 */ 2660void scheduler_tick(void) 2661{ 2662 int cpu = smp_processor_id(); 2663 struct rq *rq = cpu_rq(cpu); 2664 struct task_struct *curr = rq->curr; 2665 2666 sched_clock_tick(); 2667 2668 raw_spin_lock(&rq->lock); 2669 update_rq_clock(rq); 2670 update_cpu_load_active(rq); 2671 curr->sched_class->task_tick(rq, curr, 0); 2672 raw_spin_unlock(&rq->lock); 2673 2674 perf_event_task_tick(); 2675 2676#ifdef CONFIG_SMP 2677 rq->idle_balance = idle_cpu(cpu); 2678 trigger_load_balance(rq, cpu); 2679#endif 2680} 2681 2682notrace unsigned long get_parent_ip(unsigned long addr) 2683{ 2684 if (in_lock_functions(addr)) { 2685 addr = CALLER_ADDR2; 2686 if (in_lock_functions(addr)) 2687 addr = CALLER_ADDR3; 2688 } 2689 return addr; 2690} 2691 2692#if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \ 2693 defined(CONFIG_PREEMPT_TRACER)) 2694 2695void __kprobes add_preempt_count(int val) 2696{ 2697#ifdef CONFIG_DEBUG_PREEMPT 2698 /* 2699 * Underflow? 2700 */ 2701 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0))) 2702 return; 2703#endif 2704 preempt_count() += val; 2705#ifdef CONFIG_DEBUG_PREEMPT 2706 /* 2707 * Spinlock count overflowing soon? 2708 */ 2709 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >= 2710 PREEMPT_MASK - 10); 2711#endif 2712 if (preempt_count() == val) 2713 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1)); 2714} 2715EXPORT_SYMBOL(add_preempt_count); 2716 2717void __kprobes sub_preempt_count(int val) 2718{ 2719#ifdef CONFIG_DEBUG_PREEMPT 2720 /* 2721 * Underflow? 2722 */ 2723 if (DEBUG_LOCKS_WARN_ON(val > preempt_count())) 2724 return; 2725 /* 2726 * Is the spinlock portion underflowing? 2727 */ 2728 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) && 2729 !(preempt_count() & PREEMPT_MASK))) 2730 return; 2731#endif 2732 2733 if (preempt_count() == val) 2734 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1)); 2735 preempt_count() -= val; 2736} 2737EXPORT_SYMBOL(sub_preempt_count); 2738 2739#endif 2740 2741/* 2742 * Print scheduling while atomic bug: 2743 */ 2744static noinline void __schedule_bug(struct task_struct *prev) 2745{ 2746 if (oops_in_progress) 2747 return; 2748 2749 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n", 2750 prev->comm, prev->pid, preempt_count()); 2751 2752 debug_show_held_locks(prev); 2753 print_modules(); 2754 if (irqs_disabled()) 2755 print_irqtrace_events(prev); 2756 dump_stack(); 2757 add_taint(TAINT_WARN); 2758} 2759 2760/* 2761 * Various schedule()-time debugging checks and statistics: 2762 */ 2763static inline void schedule_debug(struct task_struct *prev) 2764{ 2765 /* 2766 * Test if we are atomic. Since do_exit() needs to call into 2767 * schedule() atomically, we ignore that path for now. 2768 * Otherwise, whine if we are scheduling when we should not be. 2769 */ 2770 if (unlikely(in_atomic_preempt_off() && !prev->exit_state)) 2771 __schedule_bug(prev); 2772 rcu_sleep_check(); 2773 2774 profile_hit(SCHED_PROFILING, __builtin_return_address(0)); 2775 2776 schedstat_inc(this_rq(), sched_count); 2777} 2778 2779static void put_prev_task(struct rq *rq, struct task_struct *prev) 2780{ 2781 if (prev->on_rq || rq->skip_clock_update < 0) 2782 update_rq_clock(rq); 2783 prev->sched_class->put_prev_task(rq, prev); 2784} 2785 2786/* 2787 * Pick up the highest-prio task: 2788 */ 2789static inline struct task_struct * 2790pick_next_task(struct rq *rq) 2791{ 2792 const struct sched_class *class; 2793 struct task_struct *p; 2794 2795 /* 2796 * Optimization: we know that if all tasks are in 2797 * the fair class we can call that function directly: 2798 */ 2799 if (likely(rq->nr_running == rq->cfs.h_nr_running)) { 2800 p = fair_sched_class.pick_next_task(rq); 2801 if (likely(p)) 2802 return p; 2803 } 2804 2805 for_each_class(class) { 2806 p = class->pick_next_task(rq); 2807 if (p) 2808 return p; 2809 } 2810 2811 BUG(); /* the idle class will always have a runnable task */ 2812} 2813 2814/* 2815 * __schedule() is the main scheduler function. 2816 * 2817 * The main means of driving the scheduler and thus entering this function are: 2818 * 2819 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc. 2820 * 2821 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return 2822 * paths. For example, see arch/x86/entry_64.S. 2823 * 2824 * To drive preemption between tasks, the scheduler sets the flag in timer 2825 * interrupt handler scheduler_tick(). 2826 * 2827 * 3. Wakeups don't really cause entry into schedule(). They add a 2828 * task to the run-queue and that's it. 2829 * 2830 * Now, if the new task added to the run-queue preempts the current 2831 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets 2832 * called on the nearest possible occasion: 2833 * 2834 * - If the kernel is preemptible (CONFIG_PREEMPT=y): 2835 * 2836 * - in syscall or exception context, at the next outmost 2837 * preempt_enable(). (this might be as soon as the wake_up()'s 2838 * spin_unlock()!) 2839 * 2840 * - in IRQ context, return from interrupt-handler to 2841 * preemptible context 2842 * 2843 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set) 2844 * then at the next: 2845 * 2846 * - cond_resched() call 2847 * - explicit schedule() call 2848 * - return from syscall or exception to user-space 2849 * - return from interrupt-handler to user-space 2850 */ 2851static void __sched __schedule(void) 2852{ 2853 struct task_struct *prev, *next; 2854 unsigned long *switch_count; 2855 struct rq *rq; 2856 int cpu; 2857 2858need_resched: 2859 preempt_disable(); 2860 cpu = smp_processor_id(); 2861 rq = cpu_rq(cpu); 2862 rcu_note_context_switch(cpu); 2863 prev = rq->curr; 2864 2865 schedule_debug(prev); 2866 2867 if (sched_feat(HRTICK)) 2868 hrtick_clear(rq); 2869 2870 raw_spin_lock_irq(&rq->lock); 2871 2872 switch_count = &prev->nivcsw; 2873 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) { 2874 if (unlikely(signal_pending_state(prev->state, prev))) { 2875 prev->state = TASK_RUNNING; 2876 } else { 2877 deactivate_task(rq, prev, DEQUEUE_SLEEP); 2878 prev->on_rq = 0; 2879 2880 /* 2881 * If a worker went to sleep, notify and ask workqueue 2882 * whether it wants to wake up a task to maintain 2883 * concurrency. 2884 */ 2885 if (prev->flags & PF_WQ_WORKER) { 2886 struct task_struct *to_wakeup; 2887 2888 to_wakeup = wq_worker_sleeping(prev, cpu); 2889 if (to_wakeup) 2890 try_to_wake_up_local(to_wakeup); 2891 } 2892 } 2893 switch_count = &prev->nvcsw; 2894 } 2895 2896 pre_schedule(rq, prev); 2897 2898 if (unlikely(!rq->nr_running)) 2899 idle_balance(cpu, rq); 2900 2901 put_prev_task(rq, prev); 2902 next = pick_next_task(rq); 2903 clear_tsk_need_resched(prev); 2904 rq->skip_clock_update = 0; 2905 2906 if (likely(prev != next)) { 2907 rq->nr_switches++; 2908 rq->curr = next; 2909 ++*switch_count; 2910 2911 context_switch(rq, prev, next); /* unlocks the rq */ 2912 /* 2913 * The context switch have flipped the stack from under us 2914 * and restored the local variables which were saved when 2915 * this task called schedule() in the past. prev == current 2916 * is still correct, but it can be moved to another cpu/rq. 2917 */ 2918 cpu = smp_processor_id(); 2919 rq = cpu_rq(cpu); 2920 } else 2921 raw_spin_unlock_irq(&rq->lock); 2922 2923 post_schedule(rq); 2924 2925 sched_preempt_enable_no_resched(); 2926 if (need_resched()) 2927 goto need_resched; 2928} 2929 2930static inline void sched_submit_work(struct task_struct *tsk) 2931{ 2932 if (!tsk->state || tsk_is_pi_blocked(tsk)) 2933 return; 2934 /* 2935 * If we are going to sleep and we have plugged IO queued, 2936 * make sure to submit it to avoid deadlocks. 2937 */ 2938 if (blk_needs_flush_plug(tsk)) 2939 blk_schedule_flush_plug(tsk); 2940} 2941 2942asmlinkage void __sched schedule(void) 2943{ 2944 struct task_struct *tsk = current; 2945 2946 sched_submit_work(tsk); 2947 __schedule(); 2948} 2949EXPORT_SYMBOL(schedule); 2950 2951/** 2952 * schedule_preempt_disabled - called with preemption disabled 2953 * 2954 * Returns with preemption disabled. Note: preempt_count must be 1 2955 */ 2956void __sched schedule_preempt_disabled(void) 2957{ 2958 sched_preempt_enable_no_resched(); 2959 schedule(); 2960 preempt_disable(); 2961} 2962 2963#ifdef CONFIG_MUTEX_SPIN_ON_OWNER 2964 2965static inline bool owner_running(struct mutex *lock, struct task_struct *owner) 2966{ 2967 if (lock->owner != owner) 2968 return false; 2969 2970 /* 2971 * Ensure we emit the owner->on_cpu, dereference _after_ checking 2972 * lock->owner still matches owner, if that fails, owner might 2973 * point to free()d memory, if it still matches, the rcu_read_lock() 2974 * ensures the memory stays valid. 2975 */ 2976 barrier(); 2977 2978 return owner->on_cpu; 2979} 2980 2981/* 2982 * Look out! "owner" is an entirely speculative pointer 2983 * access and not reliable. 2984 */ 2985int mutex_spin_on_owner(struct mutex *lock, struct task_struct *owner) 2986{ 2987 if (!sched_feat(OWNER_SPIN)) 2988 return 0; 2989 2990 rcu_read_lock(); 2991 while (owner_running(lock, owner)) { 2992 if (need_resched()) 2993 break; 2994 2995 arch_mutex_cpu_relax(); 2996 } 2997 rcu_read_unlock(); 2998 2999 /* 3000 * We break out the loop above on need_resched() and when the 3001 * owner changed, which is a sign for heavy contention. Return 3002 * success only when lock->owner is NULL. 3003 */ 3004 return lock->owner == NULL; 3005} 3006#endif 3007 3008#ifdef CONFIG_PREEMPT 3009/* 3010 * this is the entry point to schedule() from in-kernel preemption 3011 * off of preempt_enable. Kernel preemptions off return from interrupt 3012 * occur there and call schedule directly. 3013 */ 3014asmlinkage void __sched notrace preempt_schedule(void) 3015{ 3016 struct thread_info *ti = current_thread_info(); 3017 3018 /* 3019 * If there is a non-zero preempt_count or interrupts are disabled, 3020 * we do not want to preempt the current task. Just return.. 3021 */ 3022 if (likely(ti->preempt_count || irqs_disabled())) 3023 return; 3024 3025 do { 3026 add_preempt_count_notrace(PREEMPT_ACTIVE); 3027 __schedule(); 3028 sub_preempt_count_notrace(PREEMPT_ACTIVE); 3029 3030 /* 3031 * Check again in case we missed a preemption opportunity 3032 * between schedule and now. 3033 */ 3034 barrier(); 3035 } while (need_resched()); 3036} 3037EXPORT_SYMBOL(preempt_schedule); 3038 3039/* 3040 * this is the entry point to schedule() from kernel preemption 3041 * off of irq context. 3042 * Note, that this is called and return with irqs disabled. This will 3043 * protect us against recursive calling from irq. 3044 */ 3045asmlinkage void __sched preempt_schedule_irq(void) 3046{ 3047 struct thread_info *ti = current_thread_info(); 3048 3049 /* Catch callers which need to be fixed */ 3050 BUG_ON(ti->preempt_count || !irqs_disabled()); 3051 3052 do { 3053 add_preempt_count(PREEMPT_ACTIVE); 3054 local_irq_enable(); 3055 __schedule(); 3056 local_irq_disable(); 3057 sub_preempt_count(PREEMPT_ACTIVE); 3058 3059 /* 3060 * Check again in case we missed a preemption opportunity 3061 * between schedule and now. 3062 */ 3063 barrier(); 3064 } while (need_resched()); 3065} 3066 3067#endif /* CONFIG_PREEMPT */ 3068 3069int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags, 3070 void *key) 3071{ 3072 return try_to_wake_up(curr->private, mode, wake_flags); 3073} 3074EXPORT_SYMBOL(default_wake_function); 3075 3076/* 3077 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just 3078 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve 3079 * number) then we wake all the non-exclusive tasks and one exclusive task. 3080 * 3081 * There are circumstances in which we can try to wake a task which has already 3082 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns 3083 * zero in this (rare) case, and we handle it by continuing to scan the queue. 3084 */ 3085static void __wake_up_common(wait_queue_head_t *q, unsigned int mode, 3086 int nr_exclusive, int wake_flags, void *key) 3087{ 3088 wait_queue_t *curr, *next; 3089 3090 list_for_each_entry_safe(curr, next, &q->task_list, task_list) { 3091 unsigned flags = curr->flags; 3092 3093 if (curr->func(curr, mode, wake_flags, key) && 3094 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive) 3095 break; 3096 } 3097} 3098 3099/** 3100 * __wake_up - wake up threads blocked on a waitqueue. 3101 * @q: the waitqueue 3102 * @mode: which threads 3103 * @nr_exclusive: how many wake-one or wake-many threads to wake up 3104 * @key: is directly passed to the wakeup function 3105 * 3106 * It may be assumed that this function implies a write memory barrier before 3107 * changing the task state if and only if any tasks are woken up. 3108 */ 3109void __wake_up(wait_queue_head_t *q, unsigned int mode, 3110 int nr_exclusive, void *key) 3111{ 3112 unsigned long flags; 3113 3114 spin_lock_irqsave(&q->lock, flags); 3115 __wake_up_common(q, mode, nr_exclusive, 0, key); 3116 spin_unlock_irqrestore(&q->lock, flags); 3117} 3118EXPORT_SYMBOL(__wake_up); 3119 3120/* 3121 * Same as __wake_up but called with the spinlock in wait_queue_head_t held. 3122 */ 3123void __wake_up_locked(wait_queue_head_t *q, unsigned int mode, int nr) 3124{ 3125 __wake_up_common(q, mode, nr, 0, NULL); 3126} 3127EXPORT_SYMBOL_GPL(__wake_up_locked); 3128 3129void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key) 3130{ 3131 __wake_up_common(q, mode, 1, 0, key); 3132} 3133EXPORT_SYMBOL_GPL(__wake_up_locked_key); 3134 3135/** 3136 * __wake_up_sync_key - wake up threads blocked on a waitqueue. 3137 * @q: the waitqueue 3138 * @mode: which threads 3139 * @nr_exclusive: how many wake-one or wake-many threads to wake up 3140 * @key: opaque value to be passed to wakeup targets 3141 * 3142 * The sync wakeup differs that the waker knows that it will schedule 3143 * away soon, so while the target thread will be woken up, it will not 3144 * be migrated to another CPU - ie. the two threads are 'synchronized' 3145 * with each other. This can prevent needless bouncing between CPUs. 3146 * 3147 * On UP it can prevent extra preemption. 3148 * 3149 * It may be assumed that this function implies a write memory barrier before 3150 * changing the task state if and only if any tasks are woken up. 3151 */ 3152void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode, 3153 int nr_exclusive, void *key) 3154{ 3155 unsigned long flags; 3156 int wake_flags = WF_SYNC; 3157 3158 if (unlikely(!q)) 3159 return; 3160 3161 if (unlikely(!nr_exclusive)) 3162 wake_flags = 0; 3163 3164 spin_lock_irqsave(&q->lock, flags); 3165 __wake_up_common(q, mode, nr_exclusive, wake_flags, key); 3166 spin_unlock_irqrestore(&q->lock, flags); 3167} 3168EXPORT_SYMBOL_GPL(__wake_up_sync_key); 3169 3170/* 3171 * __wake_up_sync - see __wake_up_sync_key() 3172 */ 3173void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive) 3174{ 3175 __wake_up_sync_key(q, mode, nr_exclusive, NULL); 3176} 3177EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */ 3178 3179/** 3180 * complete: - signals a single thread waiting on this completion 3181 * @x: holds the state of this particular completion 3182 * 3183 * This will wake up a single thread waiting on this completion. Threads will be 3184 * awakened in the same order in which they were queued. 3185 * 3186 * See also complete_all(), wait_for_completion() and related routines. 3187 * 3188 * It may be assumed that this function implies a write memory barrier before 3189 * changing the task state if and only if any tasks are woken up. 3190 */ 3191void complete(struct completion *x) 3192{ 3193 unsigned long flags; 3194 3195 spin_lock_irqsave(&x->wait.lock, flags); 3196 x->done++; 3197 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL); 3198 spin_unlock_irqrestore(&x->wait.lock, flags); 3199} 3200EXPORT_SYMBOL(complete); 3201 3202/** 3203 * complete_all: - signals all threads waiting on this completion 3204 * @x: holds the state of this particular completion 3205 * 3206 * This will wake up all threads waiting on this particular completion event. 3207 * 3208 * It may be assumed that this function implies a write memory barrier before 3209 * changing the task state if and only if any tasks are woken up. 3210 */ 3211void complete_all(struct completion *x) 3212{ 3213 unsigned long flags; 3214 3215 spin_lock_irqsave(&x->wait.lock, flags); 3216 x->done += UINT_MAX/2; 3217 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL); 3218 spin_unlock_irqrestore(&x->wait.lock, flags); 3219} 3220EXPORT_SYMBOL(complete_all); 3221 3222static inline long __sched 3223do_wait_for_common(struct completion *x, long timeout, int state) 3224{ 3225 if (!x->done) { 3226 DECLARE_WAITQUEUE(wait, current); 3227 3228 __add_wait_queue_tail_exclusive(&x->wait, &wait); 3229 do { 3230 if (signal_pending_state(state, current)) { 3231 timeout = -ERESTARTSYS; 3232 break; 3233 } 3234 __set_current_state(state); 3235 spin_unlock_irq(&x->wait.lock); 3236 timeout = schedule_timeout(timeout); 3237 spin_lock_irq(&x->wait.lock); 3238 } while (!x->done && timeout); 3239 __remove_wait_queue(&x->wait, &wait); 3240 if (!x->done) 3241 return timeout; 3242 } 3243 x->done--; 3244 return timeout ?: 1; 3245} 3246 3247static long __sched 3248wait_for_common(struct completion *x, long timeout, int state) 3249{ 3250 might_sleep(); 3251 3252 spin_lock_irq(&x->wait.lock); 3253 timeout = do_wait_for_common(x, timeout, state); 3254 spin_unlock_irq(&x->wait.lock); 3255 return timeout; 3256} 3257 3258/** 3259 * wait_for_completion: - waits for completion of a task 3260 * @x: holds the state of this particular completion 3261 * 3262 * This waits to be signaled for completion of a specific task. It is NOT 3263 * interruptible and there is no timeout. 3264 * 3265 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout 3266 * and interrupt capability. Also see complete(). 3267 */ 3268void __sched wait_for_completion(struct completion *x) 3269{ 3270 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE); 3271} 3272EXPORT_SYMBOL(wait_for_completion); 3273 3274/** 3275 * wait_for_completion_timeout: - waits for completion of a task (w/timeout) 3276 * @x: holds the state of this particular completion 3277 * @timeout: timeout value in jiffies 3278 * 3279 * This waits for either a completion of a specific task to be signaled or for a 3280 * specified timeout to expire. The timeout is in jiffies. It is not 3281 * interruptible. 3282 * 3283 * The return value is 0 if timed out, and positive (at least 1, or number of 3284 * jiffies left till timeout) if completed. 3285 */ 3286unsigned long __sched 3287wait_for_completion_timeout(struct completion *x, unsigned long timeout) 3288{ 3289 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE); 3290} 3291EXPORT_SYMBOL(wait_for_completion_timeout); 3292 3293/** 3294 * wait_for_completion_interruptible: - waits for completion of a task (w/intr) 3295 * @x: holds the state of this particular completion 3296 * 3297 * This waits for completion of a specific task to be signaled. It is 3298 * interruptible. 3299 * 3300 * The return value is -ERESTARTSYS if interrupted, 0 if completed. 3301 */ 3302int __sched wait_for_completion_interruptible(struct completion *x) 3303{ 3304 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE); 3305 if (t == -ERESTARTSYS) 3306 return t; 3307 return 0; 3308} 3309EXPORT_SYMBOL(wait_for_completion_interruptible); 3310 3311/** 3312 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr)) 3313 * @x: holds the state of this particular completion 3314 * @timeout: timeout value in jiffies 3315 * 3316 * This waits for either a completion of a specific task to be signaled or for a 3317 * specified timeout to expire. It is interruptible. The timeout is in jiffies. 3318 * 3319 * The return value is -ERESTARTSYS if interrupted, 0 if timed out, 3320 * positive (at least 1, or number of jiffies left till timeout) if completed. 3321 */ 3322long __sched 3323wait_for_completion_interruptible_timeout(struct completion *x, 3324 unsigned long timeout) 3325{ 3326 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE); 3327} 3328EXPORT_SYMBOL(wait_for_completion_interruptible_timeout); 3329 3330/** 3331 * wait_for_completion_killable: - waits for completion of a task (killable) 3332 * @x: holds the state of this particular completion 3333 * 3334 * This waits to be signaled for completion of a specific task. It can be 3335 * interrupted by a kill signal. 3336 * 3337 * The return value is -ERESTARTSYS if interrupted, 0 if completed. 3338 */ 3339int __sched wait_for_completion_killable(struct completion *x) 3340{ 3341 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE); 3342 if (t == -ERESTARTSYS) 3343 return t; 3344 return 0; 3345} 3346EXPORT_SYMBOL(wait_for_completion_killable); 3347 3348/** 3349 * wait_for_completion_killable_timeout: - waits for completion of a task (w/(to,killable)) 3350 * @x: holds the state of this particular completion 3351 * @timeout: timeout value in jiffies 3352 * 3353 * This waits for either a completion of a specific task to be 3354 * signaled or for a specified timeout to expire. It can be 3355 * interrupted by a kill signal. The timeout is in jiffies. 3356 * 3357 * The return value is -ERESTARTSYS if interrupted, 0 if timed out, 3358 * positive (at least 1, or number of jiffies left till timeout) if completed. 3359 */ 3360long __sched 3361wait_for_completion_killable_timeout(struct completion *x, 3362 unsigned long timeout) 3363{ 3364 return wait_for_common(x, timeout, TASK_KILLABLE); 3365} 3366EXPORT_SYMBOL(wait_for_completion_killable_timeout); 3367 3368/** 3369 * try_wait_for_completion - try to decrement a completion without blocking 3370 * @x: completion structure 3371 * 3372 * Returns: 0 if a decrement cannot be done without blocking 3373 * 1 if a decrement succeeded. 3374 * 3375 * If a completion is being used as a counting completion, 3376 * attempt to decrement the counter without blocking. This 3377 * enables us to avoid waiting if the resource the completion 3378 * is protecting is not available. 3379 */ 3380bool try_wait_for_completion(struct completion *x) 3381{ 3382 unsigned long flags; 3383 int ret = 1; 3384 3385 spin_lock_irqsave(&x->wait.lock, flags); 3386 if (!x->done) 3387 ret = 0; 3388 else 3389 x->done--; 3390 spin_unlock_irqrestore(&x->wait.lock, flags); 3391 return ret; 3392} 3393EXPORT_SYMBOL(try_wait_for_completion); 3394 3395/** 3396 * completion_done - Test to see if a completion has any waiters 3397 * @x: completion structure 3398 * 3399 * Returns: 0 if there are waiters (wait_for_completion() in progress) 3400 * 1 if there are no waiters. 3401 * 3402 */ 3403bool completion_done(struct completion *x) 3404{ 3405 unsigned long flags; 3406 int ret = 1; 3407 3408 spin_lock_irqsave(&x->wait.lock, flags); 3409 if (!x->done) 3410 ret = 0; 3411 spin_unlock_irqrestore(&x->wait.lock, flags); 3412 return ret; 3413} 3414EXPORT_SYMBOL(completion_done); 3415 3416static long __sched 3417sleep_on_common(wait_queue_head_t *q, int state, long timeout) 3418{ 3419 unsigned long flags; 3420 wait_queue_t wait; 3421 3422 init_waitqueue_entry(&wait, current); 3423 3424 __set_current_state(state); 3425 3426 spin_lock_irqsave(&q->lock, flags); 3427 __add_wait_queue(q, &wait); 3428 spin_unlock(&q->lock); 3429 timeout = schedule_timeout(timeout); 3430 spin_lock_irq(&q->lock); 3431 __remove_wait_queue(q, &wait); 3432 spin_unlock_irqrestore(&q->lock, flags); 3433 3434 return timeout; 3435} 3436 3437void __sched interruptible_sleep_on(wait_queue_head_t *q) 3438{ 3439 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT); 3440} 3441EXPORT_SYMBOL(interruptible_sleep_on); 3442 3443long __sched 3444interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout) 3445{ 3446 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout); 3447} 3448EXPORT_SYMBOL(interruptible_sleep_on_timeout); 3449 3450void __sched sleep_on(wait_queue_head_t *q) 3451{ 3452 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT); 3453} 3454EXPORT_SYMBOL(sleep_on); 3455 3456long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout) 3457{ 3458 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout); 3459} 3460EXPORT_SYMBOL(sleep_on_timeout); 3461 3462#ifdef CONFIG_RT_MUTEXES 3463 3464/* 3465 * rt_mutex_setprio - set the current priority of a task 3466 * @p: task 3467 * @prio: prio value (kernel-internal form) 3468 * 3469 * This function changes the 'effective' priority of a task. It does 3470 * not touch ->normal_prio like __setscheduler(). 3471 * 3472 * Used by the rt_mutex code to implement priority inheritance logic. 3473 */ 3474void rt_mutex_setprio(struct task_struct *p, int prio) 3475{ 3476 int oldprio, on_rq, running; 3477 struct rq *rq; 3478 const struct sched_class *prev_class; 3479 3480 BUG_ON(prio < 0 || prio > MAX_PRIO); 3481 3482 rq = __task_rq_lock(p); 3483 3484 /* 3485 * Idle task boosting is a nono in general. There is one 3486 * exception, when PREEMPT_RT and NOHZ is active: 3487 * 3488 * The idle task calls get_next_timer_interrupt() and holds 3489 * the timer wheel base->lock on the CPU and another CPU wants 3490 * to access the timer (probably to cancel it). We can safely 3491 * ignore the boosting request, as the idle CPU runs this code 3492 * with interrupts disabled and will complete the lock 3493 * protected section without being interrupted. So there is no 3494 * real need to boost. 3495 */ 3496 if (unlikely(p == rq->idle)) { 3497 WARN_ON(p != rq->curr); 3498 WARN_ON(p->pi_blocked_on); 3499 goto out_unlock; 3500 } 3501 3502 trace_sched_pi_setprio(p, prio); 3503 oldprio = p->prio; 3504 prev_class = p->sched_class; 3505 on_rq = p->on_rq; 3506 running = task_current(rq, p); 3507 if (on_rq) 3508 dequeue_task(rq, p, 0); 3509 if (running) 3510 p->sched_class->put_prev_task(rq, p); 3511 3512 if (rt_prio(prio)) 3513 p->sched_class = &rt_sched_class; 3514 else 3515 p->sched_class = &fair_sched_class; 3516 3517 p->prio = prio; 3518 3519 if (running) 3520 p->sched_class->set_curr_task(rq); 3521 if (on_rq) 3522 enqueue_task(rq, p, oldprio < prio ? ENQUEUE_HEAD : 0); 3523 3524 check_class_changed(rq, p, prev_class, oldprio); 3525out_unlock: 3526 __task_rq_unlock(rq); 3527} 3528#endif 3529void set_user_nice(struct task_struct *p, long nice) 3530{ 3531 int old_prio, delta, on_rq; 3532 unsigned long flags; 3533 struct rq *rq; 3534 3535 if (TASK_NICE(p) == nice || nice < -20 || nice > 19) 3536 return; 3537 /* 3538 * We have to be careful, if called from sys_setpriority(), 3539 * the task might be in the middle of scheduling on another CPU. 3540 */ 3541 rq = task_rq_lock(p, &flags); 3542 /* 3543 * The RT priorities are set via sched_setscheduler(), but we still 3544 * allow the 'normal' nice value to be set - but as expected 3545 * it wont have any effect on scheduling until the task is 3546 * SCHED_FIFO/SCHED_RR: 3547 */ 3548 if (task_has_rt_policy(p)) { 3549 p->static_prio = NICE_TO_PRIO(nice); 3550 goto out_unlock; 3551 } 3552 on_rq = p->on_rq; 3553 if (on_rq) 3554 dequeue_task(rq, p, 0); 3555 3556 p->static_prio = NICE_TO_PRIO(nice); 3557 set_load_weight(p); 3558 old_prio = p->prio; 3559 p->prio = effective_prio(p); 3560 delta = p->prio - old_prio; 3561 3562 if (on_rq) { 3563 enqueue_task(rq, p, 0); 3564 /* 3565 * If the task increased its priority or is running and 3566 * lowered its priority, then reschedule its CPU: 3567 */ 3568 if (delta < 0 || (delta > 0 && task_running(rq, p))) 3569 resched_task(rq->curr); 3570 } 3571out_unlock: 3572 task_rq_unlock(rq, p, &flags); 3573} 3574EXPORT_SYMBOL(set_user_nice); 3575 3576/* 3577 * can_nice - check if a task can reduce its nice value 3578 * @p: task 3579 * @nice: nice value 3580 */ 3581int can_nice(const struct task_struct *p, const int nice) 3582{ 3583 /* convert nice value [19,-20] to rlimit style value [1,40] */ 3584 int nice_rlim = 20 - nice; 3585 3586 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) || 3587 capable(CAP_SYS_NICE)); 3588} 3589 3590#ifdef __ARCH_WANT_SYS_NICE 3591 3592/* 3593 * sys_nice - change the priority of the current process. 3594 * @increment: priority increment 3595 * 3596 * sys_setpriority is a more generic, but much slower function that 3597 * does similar things. 3598 */ 3599SYSCALL_DEFINE1(nice, int, increment) 3600{ 3601 long nice, retval; 3602 3603 /* 3604 * Setpriority might change our priority at the same moment. 3605 * We don't have to worry. Conceptually one call occurs first 3606 * and we have a single winner. 3607 */ 3608 if (increment < -40) 3609 increment = -40; 3610 if (increment > 40) 3611 increment = 40; 3612 3613 nice = TASK_NICE(current) + increment; 3614 if (nice < -20) 3615 nice = -20; 3616 if (nice > 19) 3617 nice = 19; 3618 3619 if (increment < 0 && !can_nice(current, nice)) 3620 return -EPERM; 3621 3622 retval = security_task_setnice(current, nice); 3623 if (retval) 3624 return retval; 3625 3626 set_user_nice(current, nice); 3627 return 0; 3628} 3629 3630#endif 3631 3632/** 3633 * task_prio - return the priority value of a given task. 3634 * @p: the task in question. 3635 * 3636 * This is the priority value as seen by users in /proc. 3637 * RT tasks are offset by -200. Normal tasks are centered 3638 * around 0, value goes from -16 to +15. 3639 */ 3640int task_prio(const struct task_struct *p) 3641{ 3642 return p->prio - MAX_RT_PRIO; 3643} 3644 3645/** 3646 * task_nice - return the nice value of a given task. 3647 * @p: the task in question. 3648 */ 3649int task_nice(const struct task_struct *p) 3650{ 3651 return TASK_NICE(p); 3652} 3653EXPORT_SYMBOL(task_nice); 3654 3655/** 3656 * idle_cpu - is a given cpu idle currently? 3657 * @cpu: the processor in question. 3658 */ 3659int idle_cpu(int cpu) 3660{ 3661 struct rq *rq = cpu_rq(cpu); 3662 3663 if (rq->curr != rq->idle) 3664 return 0; 3665 3666 if (rq->nr_running) 3667 return 0; 3668 3669#ifdef CONFIG_SMP 3670 if (!llist_empty(&rq->wake_list)) 3671 return 0; 3672#endif 3673 3674 return 1; 3675} 3676 3677/** 3678 * idle_task - return the idle task for a given cpu. 3679 * @cpu: the processor in question. 3680 */ 3681struct task_struct *idle_task(int cpu) 3682{ 3683 return cpu_rq(cpu)->idle; 3684} 3685 3686/** 3687 * find_process_by_pid - find a process with a matching PID value. 3688 * @pid: the pid in question. 3689 */ 3690static struct task_struct *find_process_by_pid(pid_t pid) 3691{ 3692 return pid ? find_task_by_vpid(pid) : current; 3693} 3694 3695/* Actually do priority change: must hold rq lock. */ 3696static void 3697__setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio) 3698{ 3699 p->policy = policy; 3700 p->rt_priority = prio; 3701 p->normal_prio = normal_prio(p); 3702 /* we are holding p->pi_lock already */ 3703 p->prio = rt_mutex_getprio(p); 3704 if (rt_prio(p->prio)) 3705 p->sched_class = &rt_sched_class; 3706 else 3707 p->sched_class = &fair_sched_class; 3708 set_load_weight(p); 3709} 3710 3711/* 3712 * check the target process has a UID that matches the current process's 3713 */ 3714static bool check_same_owner(struct task_struct *p) 3715{ 3716 const struct cred *cred = current_cred(), *pcred; 3717 bool match; 3718 3719 rcu_read_lock(); 3720 pcred = __task_cred(p); 3721 match = (uid_eq(cred->euid, pcred->euid) || 3722 uid_eq(cred->euid, pcred->uid)); 3723 rcu_read_unlock(); 3724 return match; 3725} 3726 3727static int __sched_setscheduler(struct task_struct *p, int policy, 3728 const struct sched_param *param, bool user) 3729{ 3730 int retval, oldprio, oldpolicy = -1, on_rq, running; 3731 unsigned long flags; 3732 const struct sched_class *prev_class; 3733 struct rq *rq; 3734 int reset_on_fork; 3735 3736 /* may grab non-irq protected spin_locks */ 3737 BUG_ON(in_interrupt()); 3738recheck: 3739 /* double check policy once rq lock held */ 3740 if (policy < 0) { 3741 reset_on_fork = p->sched_reset_on_fork; 3742 policy = oldpolicy = p->policy; 3743 } else { 3744 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK); 3745 policy &= ~SCHED_RESET_ON_FORK; 3746 3747 if (policy != SCHED_FIFO && policy != SCHED_RR && 3748 policy != SCHED_NORMAL && policy != SCHED_BATCH && 3749 policy != SCHED_IDLE) 3750 return -EINVAL; 3751 } 3752 3753 /* 3754 * Valid priorities for SCHED_FIFO and SCHED_RR are 3755 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL, 3756 * SCHED_BATCH and SCHED_IDLE is 0. 3757 */ 3758 if (param->sched_priority < 0 || 3759 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) || 3760 (!p->mm && param->sched_priority > MAX_RT_PRIO-1)) 3761 return -EINVAL; 3762 if (rt_policy(policy) != (param->sched_priority != 0)) 3763 return -EINVAL; 3764 3765 /* 3766 * Allow unprivileged RT tasks to decrease priority: 3767 */ 3768 if (user && !capable(CAP_SYS_NICE)) { 3769 if (rt_policy(policy)) { 3770 unsigned long rlim_rtprio = 3771 task_rlimit(p, RLIMIT_RTPRIO); 3772 3773 /* can't set/change the rt policy */ 3774 if (policy != p->policy && !rlim_rtprio) 3775 return -EPERM; 3776 3777 /* can't increase priority */ 3778 if (param->sched_priority > p->rt_priority && 3779 param->sched_priority > rlim_rtprio) 3780 return -EPERM; 3781 } 3782 3783 /* 3784 * Treat SCHED_IDLE as nice 20. Only allow a switch to 3785 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it. 3786 */ 3787 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) { 3788 if (!can_nice(p, TASK_NICE(p))) 3789 return -EPERM; 3790 } 3791 3792 /* can't change other user's priorities */ 3793 if (!check_same_owner(p)) 3794 return -EPERM; 3795 3796 /* Normal users shall not reset the sched_reset_on_fork flag */ 3797 if (p->sched_reset_on_fork && !reset_on_fork) 3798 return -EPERM; 3799 } 3800 3801 if (user) { 3802 retval = security_task_setscheduler(p); 3803 if (retval) 3804 return retval; 3805 } 3806 3807 /* 3808 * make sure no PI-waiters arrive (or leave) while we are 3809 * changing the priority of the task: 3810 * 3811 * To be able to change p->policy safely, the appropriate 3812 * runqueue lock must be held. 3813 */ 3814 rq = task_rq_lock(p, &flags); 3815 3816 /* 3817 * Changing the policy of the stop threads its a very bad idea 3818 */ 3819 if (p == rq->stop) { 3820 task_rq_unlock(rq, p, &flags); 3821 return -EINVAL; 3822 } 3823 3824 /* 3825 * If not changing anything there's no need to proceed further: 3826 */ 3827 if (unlikely(policy == p->policy && (!rt_policy(policy) || 3828 param->sched_priority == p->rt_priority))) { 3829 task_rq_unlock(rq, p, &flags); 3830 return 0; 3831 } 3832 3833#ifdef CONFIG_RT_GROUP_SCHED 3834 if (user) { 3835 /* 3836 * Do not allow realtime tasks into groups that have no runtime 3837 * assigned. 3838 */ 3839 if (rt_bandwidth_enabled() && rt_policy(policy) && 3840 task_group(p)->rt_bandwidth.rt_runtime == 0 && 3841 !task_group_is_autogroup(task_group(p))) { 3842 task_rq_unlock(rq, p, &flags); 3843 return -EPERM; 3844 } 3845 } 3846#endif 3847 3848 /* recheck policy now with rq lock held */ 3849 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) { 3850 policy = oldpolicy = -1; 3851 task_rq_unlock(rq, p, &flags); 3852 goto recheck; 3853 } 3854 on_rq = p->on_rq; 3855 running = task_current(rq, p); 3856 if (on_rq) 3857 dequeue_task(rq, p, 0); 3858 if (running) 3859 p->sched_class->put_prev_task(rq, p); 3860 3861 p->sched_reset_on_fork = reset_on_fork; 3862 3863 oldprio = p->prio; 3864 prev_class = p->sched_class; 3865 __setscheduler(rq, p, policy, param->sched_priority); 3866 3867 if (running) 3868 p->sched_class->set_curr_task(rq); 3869 if (on_rq) 3870 enqueue_task(rq, p, 0); 3871 3872 check_class_changed(rq, p, prev_class, oldprio); 3873 task_rq_unlock(rq, p, &flags); 3874 3875 rt_mutex_adjust_pi(p); 3876 3877 return 0; 3878} 3879 3880/** 3881 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread. 3882 * @p: the task in question. 3883 * @policy: new policy. 3884 * @param: structure containing the new RT priority. 3885 * 3886 * NOTE that the task may be already dead. 3887 */ 3888int sched_setscheduler(struct task_struct *p, int policy, 3889 const struct sched_param *param) 3890{ 3891 return __sched_setscheduler(p, policy, param, true); 3892} 3893EXPORT_SYMBOL_GPL(sched_setscheduler); 3894 3895/** 3896 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace. 3897 * @p: the task in question. 3898 * @policy: new policy. 3899 * @param: structure containing the new RT priority. 3900 * 3901 * Just like sched_setscheduler, only don't bother checking if the 3902 * current context has permission. For example, this is needed in 3903 * stop_machine(): we create temporary high priority worker threads, 3904 * but our caller might not have that capability. 3905 */ 3906int sched_setscheduler_nocheck(struct task_struct *p, int policy, 3907 const struct sched_param *param) 3908{ 3909 return __sched_setscheduler(p, policy, param, false); 3910} 3911 3912static int 3913do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param) 3914{ 3915 struct sched_param lparam; 3916 struct task_struct *p; 3917 int retval; 3918 3919 if (!param || pid < 0) 3920 return -EINVAL; 3921 if (copy_from_user(&lparam, param, sizeof(struct sched_param))) 3922 return -EFAULT; 3923 3924 rcu_read_lock(); 3925 retval = -ESRCH; 3926 p = find_process_by_pid(pid); 3927 if (p != NULL) 3928 retval = sched_setscheduler(p, policy, &lparam); 3929 rcu_read_unlock(); 3930 3931 return retval; 3932} 3933 3934/** 3935 * sys_sched_setscheduler - set/change the scheduler policy and RT priority 3936 * @pid: the pid in question. 3937 * @policy: new policy. 3938 * @param: structure containing the new RT priority. 3939 */ 3940SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy, 3941 struct sched_param __user *, param) 3942{ 3943 /* negative values for policy are not valid */ 3944 if (policy < 0) 3945 return -EINVAL; 3946 3947 return do_sched_setscheduler(pid, policy, param); 3948} 3949 3950/** 3951 * sys_sched_setparam - set/change the RT priority of a thread 3952 * @pid: the pid in question. 3953 * @param: structure containing the new RT priority. 3954 */ 3955SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param) 3956{ 3957 return do_sched_setscheduler(pid, -1, param); 3958} 3959 3960/** 3961 * sys_sched_getscheduler - get the policy (scheduling class) of a thread 3962 * @pid: the pid in question. 3963 */ 3964SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid) 3965{ 3966 struct task_struct *p; 3967 int retval; 3968 3969 if (pid < 0) 3970 return -EINVAL; 3971 3972 retval = -ESRCH; 3973 rcu_read_lock(); 3974 p = find_process_by_pid(pid); 3975 if (p) { 3976 retval = security_task_getscheduler(p); 3977 if (!retval) 3978 retval = p->policy 3979 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0); 3980 } 3981 rcu_read_unlock(); 3982 return retval; 3983} 3984 3985/** 3986 * sys_sched_getparam - get the RT priority of a thread 3987 * @pid: the pid in question. 3988 * @param: structure containing the RT priority. 3989 */ 3990SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param) 3991{ 3992 struct sched_param lp; 3993 struct task_struct *p; 3994 int retval; 3995 3996 if (!param || pid < 0) 3997 return -EINVAL; 3998 3999 rcu_read_lock(); 4000 p = find_process_by_pid(pid); 4001 retval = -ESRCH; 4002 if (!p) 4003 goto out_unlock; 4004 4005 retval = security_task_getscheduler(p); 4006 if (retval) 4007 goto out_unlock; 4008 4009 lp.sched_priority = p->rt_priority; 4010 rcu_read_unlock(); 4011 4012 /* 4013 * This one might sleep, we cannot do it with a spinlock held ... 4014 */ 4015 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0; 4016 4017 return retval; 4018 4019out_unlock: 4020 rcu_read_unlock(); 4021 return retval; 4022} 4023 4024long sched_setaffinity(pid_t pid, const struct cpumask *in_mask) 4025{ 4026 cpumask_var_t cpus_allowed, new_mask; 4027 struct task_struct *p; 4028 int retval; 4029 4030 get_online_cpus(); 4031 rcu_read_lock(); 4032 4033 p = find_process_by_pid(pid); 4034 if (!p) { 4035 rcu_read_unlock(); 4036 put_online_cpus(); 4037 return -ESRCH; 4038 } 4039 4040 /* Prevent p going away */ 4041 get_task_struct(p); 4042 rcu_read_unlock(); 4043 4044 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) { 4045 retval = -ENOMEM; 4046 goto out_put_task; 4047 } 4048 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) { 4049 retval = -ENOMEM; 4050 goto out_free_cpus_allowed; 4051 } 4052 retval = -EPERM; 4053 if (!check_same_owner(p) && !ns_capable(task_user_ns(p), CAP_SYS_NICE)) 4054 goto out_unlock; 4055 4056 retval = security_task_setscheduler(p); 4057 if (retval) 4058 goto out_unlock; 4059 4060 cpuset_cpus_allowed(p, cpus_allowed); 4061 cpumask_and(new_mask, in_mask, cpus_allowed); 4062again: 4063 retval = set_cpus_allowed_ptr(p, new_mask); 4064 4065 if (!retval) { 4066 cpuset_cpus_allowed(p, cpus_allowed); 4067 if (!cpumask_subset(new_mask, cpus_allowed)) { 4068 /* 4069 * We must have raced with a concurrent cpuset 4070 * update. Just reset the cpus_allowed to the 4071 * cpuset's cpus_allowed 4072 */ 4073 cpumask_copy(new_mask, cpus_allowed); 4074 goto again; 4075 } 4076 } 4077out_unlock: 4078 free_cpumask_var(new_mask); 4079out_free_cpus_allowed: 4080 free_cpumask_var(cpus_allowed); 4081out_put_task: 4082 put_task_struct(p); 4083 put_online_cpus(); 4084 return retval; 4085} 4086 4087static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len, 4088 struct cpumask *new_mask) 4089{ 4090 if (len < cpumask_size()) 4091 cpumask_clear(new_mask); 4092 else if (len > cpumask_size()) 4093 len = cpumask_size(); 4094 4095 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0; 4096} 4097 4098/** 4099 * sys_sched_setaffinity - set the cpu affinity of a process 4100 * @pid: pid of the process 4101 * @len: length in bytes of the bitmask pointed to by user_mask_ptr 4102 * @user_mask_ptr: user-space pointer to the new cpu mask 4103 */ 4104SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len, 4105 unsigned long __user *, user_mask_ptr) 4106{ 4107 cpumask_var_t new_mask; 4108 int retval; 4109 4110 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) 4111 return -ENOMEM; 4112 4113 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask); 4114 if (retval == 0) 4115 retval = sched_setaffinity(pid, new_mask); 4116 free_cpumask_var(new_mask); 4117 return retval; 4118} 4119 4120long sched_getaffinity(pid_t pid, struct cpumask *mask) 4121{ 4122 struct task_struct *p; 4123 unsigned long flags; 4124 int retval; 4125 4126 get_online_cpus(); 4127 rcu_read_lock(); 4128 4129 retval = -ESRCH; 4130 p = find_process_by_pid(pid); 4131 if (!p) 4132 goto out_unlock; 4133 4134 retval = security_task_getscheduler(p); 4135 if (retval) 4136 goto out_unlock; 4137 4138 raw_spin_lock_irqsave(&p->pi_lock, flags); 4139 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask); 4140 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 4141 4142out_unlock: 4143 rcu_read_unlock(); 4144 put_online_cpus(); 4145 4146 return retval; 4147} 4148 4149/** 4150 * sys_sched_getaffinity - get the cpu affinity of a process 4151 * @pid: pid of the process 4152 * @len: length in bytes of the bitmask pointed to by user_mask_ptr 4153 * @user_mask_ptr: user-space pointer to hold the current cpu mask 4154 */ 4155SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len, 4156 unsigned long __user *, user_mask_ptr) 4157{ 4158 int ret; 4159 cpumask_var_t mask; 4160 4161 if ((len * BITS_PER_BYTE) < nr_cpu_ids) 4162 return -EINVAL; 4163 if (len & (sizeof(unsigned long)-1)) 4164 return -EINVAL; 4165 4166 if (!alloc_cpumask_var(&mask, GFP_KERNEL)) 4167 return -ENOMEM; 4168 4169 ret = sched_getaffinity(pid, mask); 4170 if (ret == 0) { 4171 size_t retlen = min_t(size_t, len, cpumask_size()); 4172 4173 if (copy_to_user(user_mask_ptr, mask, retlen)) 4174 ret = -EFAULT; 4175 else 4176 ret = retlen; 4177 } 4178 free_cpumask_var(mask); 4179 4180 return ret; 4181} 4182 4183/** 4184 * sys_sched_yield - yield the current processor to other threads. 4185 * 4186 * This function yields the current CPU to other tasks. If there are no 4187 * other threads running on this CPU then this function will return. 4188 */ 4189SYSCALL_DEFINE0(sched_yield) 4190{ 4191 struct rq *rq = this_rq_lock(); 4192 4193 schedstat_inc(rq, yld_count); 4194 current->sched_class->yield_task(rq); 4195 4196 /* 4197 * Since we are going to call schedule() anyway, there's 4198 * no need to preempt or enable interrupts: 4199 */ 4200 __release(rq->lock); 4201 spin_release(&rq->lock.dep_map, 1, _THIS_IP_); 4202 do_raw_spin_unlock(&rq->lock); 4203 sched_preempt_enable_no_resched(); 4204 4205 schedule(); 4206 4207 return 0; 4208} 4209 4210static inline int should_resched(void) 4211{ 4212 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE); 4213} 4214 4215static void __cond_resched(void) 4216{ 4217 add_preempt_count(PREEMPT_ACTIVE); 4218 __schedule(); 4219 sub_preempt_count(PREEMPT_ACTIVE); 4220} 4221 4222int __sched _cond_resched(void) 4223{ 4224 if (should_resched()) { 4225 __cond_resched(); 4226 return 1; 4227 } 4228 return 0; 4229} 4230EXPORT_SYMBOL(_cond_resched); 4231 4232/* 4233 * __cond_resched_lock() - if a reschedule is pending, drop the given lock, 4234 * call schedule, and on return reacquire the lock. 4235 * 4236 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level 4237 * operations here to prevent schedule() from being called twice (once via 4238 * spin_unlock(), once by hand). 4239 */ 4240int __cond_resched_lock(spinlock_t *lock) 4241{ 4242 int resched = should_resched(); 4243 int ret = 0; 4244 4245 lockdep_assert_held(lock); 4246 4247 if (spin_needbreak(lock) || resched) { 4248 spin_unlock(lock); 4249 if (resched) 4250 __cond_resched(); 4251 else 4252 cpu_relax(); 4253 ret = 1; 4254 spin_lock(lock); 4255 } 4256 return ret; 4257} 4258EXPORT_SYMBOL(__cond_resched_lock); 4259 4260int __sched __cond_resched_softirq(void) 4261{ 4262 BUG_ON(!in_softirq()); 4263 4264 if (should_resched()) { 4265 local_bh_enable(); 4266 __cond_resched(); 4267 local_bh_disable(); 4268 return 1; 4269 } 4270 return 0; 4271} 4272EXPORT_SYMBOL(__cond_resched_softirq); 4273 4274/** 4275 * yield - yield the current processor to other threads. 4276 * 4277 * Do not ever use this function, there's a 99% chance you're doing it wrong. 4278 * 4279 * The scheduler is at all times free to pick the calling task as the most 4280 * eligible task to run, if removing the yield() call from your code breaks 4281 * it, its already broken. 4282 * 4283 * Typical broken usage is: 4284 * 4285 * while (!event) 4286 * yield(); 4287 * 4288 * where one assumes that yield() will let 'the other' process run that will 4289 * make event true. If the current task is a SCHED_FIFO task that will never 4290 * happen. Never use yield() as a progress guarantee!! 4291 * 4292 * If you want to use yield() to wait for something, use wait_event(). 4293 * If you want to use yield() to be 'nice' for others, use cond_resched(). 4294 * If you still want to use yield(), do not! 4295 */ 4296void __sched yield(void) 4297{ 4298 set_current_state(TASK_RUNNING); 4299 sys_sched_yield(); 4300} 4301EXPORT_SYMBOL(yield); 4302 4303/** 4304 * yield_to - yield the current processor to another thread in 4305 * your thread group, or accelerate that thread toward the 4306 * processor it's on. 4307 * @p: target task 4308 * @preempt: whether task preemption is allowed or not 4309 * 4310 * It's the caller's job to ensure that the target task struct 4311 * can't go away on us before we can do any checks. 4312 * 4313 * Returns true if we indeed boosted the target task. 4314 */ 4315bool __sched yield_to(struct task_struct *p, bool preempt) 4316{ 4317 struct task_struct *curr = current; 4318 struct rq *rq, *p_rq; 4319 unsigned long flags; 4320 bool yielded = 0; 4321 4322 local_irq_save(flags); 4323 rq = this_rq(); 4324 4325again: 4326 p_rq = task_rq(p); 4327 double_rq_lock(rq, p_rq); 4328 while (task_rq(p) != p_rq) { 4329 double_rq_unlock(rq, p_rq); 4330 goto again; 4331 } 4332 4333 if (!curr->sched_class->yield_to_task) 4334 goto out; 4335 4336 if (curr->sched_class != p->sched_class) 4337 goto out; 4338 4339 if (task_running(p_rq, p) || p->state) 4340 goto out; 4341 4342 yielded = curr->sched_class->yield_to_task(rq, p, preempt); 4343 if (yielded) { 4344 schedstat_inc(rq, yld_count); 4345 /* 4346 * Make p's CPU reschedule; pick_next_entity takes care of 4347 * fairness. 4348 */ 4349 if (preempt && rq != p_rq) 4350 resched_task(p_rq->curr); 4351 } else { 4352 /* 4353 * We might have set it in task_yield_fair(), but are 4354 * not going to schedule(), so don't want to skip 4355 * the next update. 4356 */ 4357 rq->skip_clock_update = 0; 4358 } 4359 4360out: 4361 double_rq_unlock(rq, p_rq); 4362 local_irq_restore(flags); 4363 4364 if (yielded) 4365 schedule(); 4366 4367 return yielded; 4368} 4369EXPORT_SYMBOL_GPL(yield_to); 4370 4371/* 4372 * This task is about to go to sleep on IO. Increment rq->nr_iowait so 4373 * that process accounting knows that this is a task in IO wait state. 4374 */ 4375void __sched io_schedule(void) 4376{ 4377 struct rq *rq = raw_rq(); 4378 4379 delayacct_blkio_start(); 4380 atomic_inc(&rq->nr_iowait); 4381 blk_flush_plug(current); 4382 current->in_iowait = 1; 4383 schedule(); 4384 current->in_iowait = 0; 4385 atomic_dec(&rq->nr_iowait); 4386 delayacct_blkio_end(); 4387} 4388EXPORT_SYMBOL(io_schedule); 4389 4390long __sched io_schedule_timeout(long timeout) 4391{ 4392 struct rq *rq = raw_rq(); 4393 long ret; 4394 4395 delayacct_blkio_start(); 4396 atomic_inc(&rq->nr_iowait); 4397 blk_flush_plug(current); 4398 current->in_iowait = 1; 4399 ret = schedule_timeout(timeout); 4400 current->in_iowait = 0; 4401 atomic_dec(&rq->nr_iowait); 4402 delayacct_blkio_end(); 4403 return ret; 4404} 4405 4406/** 4407 * sys_sched_get_priority_max - return maximum RT priority. 4408 * @policy: scheduling class. 4409 * 4410 * this syscall returns the maximum rt_priority that can be used 4411 * by a given scheduling class. 4412 */ 4413SYSCALL_DEFINE1(sched_get_priority_max, int, policy) 4414{ 4415 int ret = -EINVAL; 4416 4417 switch (policy) { 4418 case SCHED_FIFO: 4419 case SCHED_RR: 4420 ret = MAX_USER_RT_PRIO-1; 4421 break; 4422 case SCHED_NORMAL: 4423 case SCHED_BATCH: 4424 case SCHED_IDLE: 4425 ret = 0; 4426 break; 4427 } 4428 return ret; 4429} 4430 4431/** 4432 * sys_sched_get_priority_min - return minimum RT priority. 4433 * @policy: scheduling class. 4434 * 4435 * this syscall returns the minimum rt_priority that can be used 4436 * by a given scheduling class. 4437 */ 4438SYSCALL_DEFINE1(sched_get_priority_min, int, policy) 4439{ 4440 int ret = -EINVAL; 4441 4442 switch (policy) { 4443 case SCHED_FIFO: 4444 case SCHED_RR: 4445 ret = 1; 4446 break; 4447 case SCHED_NORMAL: 4448 case SCHED_BATCH: 4449 case SCHED_IDLE: 4450 ret = 0; 4451 } 4452 return ret; 4453} 4454 4455/** 4456 * sys_sched_rr_get_interval - return the default timeslice of a process. 4457 * @pid: pid of the process. 4458 * @interval: userspace pointer to the timeslice value. 4459 * 4460 * this syscall writes the default timeslice value of a given process 4461 * into the user-space timespec buffer. A value of '0' means infinity. 4462 */ 4463SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid, 4464 struct timespec __user *, interval) 4465{ 4466 struct task_struct *p; 4467 unsigned int time_slice; 4468 unsigned long flags; 4469 struct rq *rq; 4470 int retval; 4471 struct timespec t; 4472 4473 if (pid < 0) 4474 return -EINVAL; 4475 4476 retval = -ESRCH; 4477 rcu_read_lock(); 4478 p = find_process_by_pid(pid); 4479 if (!p) 4480 goto out_unlock; 4481 4482 retval = security_task_getscheduler(p); 4483 if (retval) 4484 goto out_unlock; 4485 4486 rq = task_rq_lock(p, &flags); 4487 time_slice = p->sched_class->get_rr_interval(rq, p); 4488 task_rq_unlock(rq, p, &flags); 4489 4490 rcu_read_unlock(); 4491 jiffies_to_timespec(time_slice, &t); 4492 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0; 4493 return retval; 4494 4495out_unlock: 4496 rcu_read_unlock(); 4497 return retval; 4498} 4499 4500static const char stat_nam[] = TASK_STATE_TO_CHAR_STR; 4501 4502void sched_show_task(struct task_struct *p) 4503{ 4504 unsigned long free = 0; 4505 unsigned state; 4506 4507 state = p->state ? __ffs(p->state) + 1 : 0; 4508 printk(KERN_INFO "%-15.15s %c", p->comm, 4509 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?'); 4510#if BITS_PER_LONG == 32 4511 if (state == TASK_RUNNING) 4512 printk(KERN_CONT " running "); 4513 else 4514 printk(KERN_CONT " %08lx ", thread_saved_pc(p)); 4515#else 4516 if (state == TASK_RUNNING) 4517 printk(KERN_CONT " running task "); 4518 else 4519 printk(KERN_CONT " %016lx ", thread_saved_pc(p)); 4520#endif 4521#ifdef CONFIG_DEBUG_STACK_USAGE 4522 free = stack_not_used(p); 4523#endif 4524 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free, 4525 task_pid_nr(p), task_pid_nr(rcu_dereference(p->real_parent)), 4526 (unsigned long)task_thread_info(p)->flags); 4527 4528 show_stack(p, NULL); 4529} 4530 4531void show_state_filter(unsigned long state_filter) 4532{ 4533 struct task_struct *g, *p; 4534 4535#if BITS_PER_LONG == 32 4536 printk(KERN_INFO 4537 " task PC stack pid father\n"); 4538#else 4539 printk(KERN_INFO 4540 " task PC stack pid father\n"); 4541#endif 4542 rcu_read_lock(); 4543 do_each_thread(g, p) { 4544 /* 4545 * reset the NMI-timeout, listing all files on a slow 4546 * console might take a lot of time: 4547 */ 4548 touch_nmi_watchdog(); 4549 if (!state_filter || (p->state & state_filter)) 4550 sched_show_task(p); 4551 } while_each_thread(g, p); 4552 4553 touch_all_softlockup_watchdogs(); 4554 4555#ifdef CONFIG_SCHED_DEBUG 4556 sysrq_sched_debug_show(); 4557#endif 4558 rcu_read_unlock(); 4559 /* 4560 * Only show locks if all tasks are dumped: 4561 */ 4562 if (!state_filter) 4563 debug_show_all_locks(); 4564} 4565 4566void __cpuinit init_idle_bootup_task(struct task_struct *idle) 4567{ 4568 idle->sched_class = &idle_sched_class; 4569} 4570 4571/** 4572 * init_idle - set up an idle thread for a given CPU 4573 * @idle: task in question 4574 * @cpu: cpu the idle task belongs to 4575 * 4576 * NOTE: this function does not set the idle thread's NEED_RESCHED 4577 * flag, to make booting more robust. 4578 */ 4579void __cpuinit init_idle(struct task_struct *idle, int cpu) 4580{ 4581 struct rq *rq = cpu_rq(cpu); 4582 unsigned long flags; 4583 4584 raw_spin_lock_irqsave(&rq->lock, flags); 4585 4586 __sched_fork(idle); 4587 idle->state = TASK_RUNNING; 4588 idle->se.exec_start = sched_clock(); 4589 4590 do_set_cpus_allowed(idle, cpumask_of(cpu)); 4591 /* 4592 * We're having a chicken and egg problem, even though we are 4593 * holding rq->lock, the cpu isn't yet set to this cpu so the 4594 * lockdep check in task_group() will fail. 4595 * 4596 * Similar case to sched_fork(). / Alternatively we could 4597 * use task_rq_lock() here and obtain the other rq->lock. 4598 * 4599 * Silence PROVE_RCU 4600 */ 4601 rcu_read_lock(); 4602 __set_task_cpu(idle, cpu); 4603 rcu_read_unlock(); 4604 4605 rq->curr = rq->idle = idle; 4606#if defined(CONFIG_SMP) 4607 idle->on_cpu = 1; 4608#endif 4609 raw_spin_unlock_irqrestore(&rq->lock, flags); 4610 4611 /* Set the preempt count _outside_ the spinlocks! */ 4612 task_thread_info(idle)->preempt_count = 0; 4613 4614 /* 4615 * The idle tasks have their own, simple scheduling class: 4616 */ 4617 idle->sched_class = &idle_sched_class; 4618 ftrace_graph_init_idle_task(idle, cpu); 4619#if defined(CONFIG_SMP) 4620 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu); 4621#endif 4622} 4623 4624#ifdef CONFIG_SMP 4625void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask) 4626{ 4627 if (p->sched_class && p->sched_class->set_cpus_allowed) 4628 p->sched_class->set_cpus_allowed(p, new_mask); 4629 4630 cpumask_copy(&p->cpus_allowed, new_mask); 4631 p->nr_cpus_allowed = cpumask_weight(new_mask); 4632} 4633 4634/* 4635 * This is how migration works: 4636 * 4637 * 1) we invoke migration_cpu_stop() on the target CPU using 4638 * stop_one_cpu(). 4639 * 2) stopper starts to run (implicitly forcing the migrated thread 4640 * off the CPU) 4641 * 3) it checks whether the migrated task is still in the wrong runqueue. 4642 * 4) if it's in the wrong runqueue then the migration thread removes 4643 * it and puts it into the right queue. 4644 * 5) stopper completes and stop_one_cpu() returns and the migration 4645 * is done. 4646 */ 4647 4648/* 4649 * Change a given task's CPU affinity. Migrate the thread to a 4650 * proper CPU and schedule it away if the CPU it's executing on 4651 * is removed from the allowed bitmask. 4652 * 4653 * NOTE: the caller must have a valid reference to the task, the 4654 * task must not exit() & deallocate itself prematurely. The 4655 * call is not atomic; no spinlocks may be held. 4656 */ 4657int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask) 4658{ 4659 unsigned long flags; 4660 struct rq *rq; 4661 unsigned int dest_cpu; 4662 int ret = 0; 4663 4664 rq = task_rq_lock(p, &flags); 4665 4666 if (cpumask_equal(&p->cpus_allowed, new_mask)) 4667 goto out; 4668 4669 if (!cpumask_intersects(new_mask, cpu_active_mask)) { 4670 ret = -EINVAL; 4671 goto out; 4672 } 4673 4674 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current)) { 4675 ret = -EINVAL; 4676 goto out; 4677 } 4678 4679 do_set_cpus_allowed(p, new_mask); 4680 4681 /* Can the task run on the task's current CPU? If so, we're done */ 4682 if (cpumask_test_cpu(task_cpu(p), new_mask)) 4683 goto out; 4684 4685 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask); 4686 if (p->on_rq) { 4687 struct migration_arg arg = { p, dest_cpu }; 4688 /* Need help from migration thread: drop lock and wait. */ 4689 task_rq_unlock(rq, p, &flags); 4690 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg); 4691 tlb_migrate_finish(p->mm); 4692 return 0; 4693 } 4694out: 4695 task_rq_unlock(rq, p, &flags); 4696 4697 return ret; 4698} 4699EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr); 4700 4701/* 4702 * Move (not current) task off this cpu, onto dest cpu. We're doing 4703 * this because either it can't run here any more (set_cpus_allowed() 4704 * away from this CPU, or CPU going down), or because we're 4705 * attempting to rebalance this task on exec (sched_exec). 4706 * 4707 * So we race with normal scheduler movements, but that's OK, as long 4708 * as the task is no longer on this CPU. 4709 * 4710 * Returns non-zero if task was successfully migrated. 4711 */ 4712static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu) 4713{ 4714 struct rq *rq_dest, *rq_src; 4715 int ret = 0; 4716 4717 if (unlikely(!cpu_active(dest_cpu))) 4718 return ret; 4719 4720 rq_src = cpu_rq(src_cpu); 4721 rq_dest = cpu_rq(dest_cpu); 4722 4723 raw_spin_lock(&p->pi_lock); 4724 double_rq_lock(rq_src, rq_dest); 4725 /* Already moved. */ 4726 if (task_cpu(p) != src_cpu) 4727 goto done; 4728 /* Affinity changed (again). */ 4729 if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p))) 4730 goto fail; 4731 4732 /* 4733 * If we're not on a rq, the next wake-up will ensure we're 4734 * placed properly. 4735 */ 4736 if (p->on_rq) { 4737 dequeue_task(rq_src, p, 0); 4738 set_task_cpu(p, dest_cpu); 4739 enqueue_task(rq_dest, p, 0); 4740 check_preempt_curr(rq_dest, p, 0); 4741 } 4742done: 4743 ret = 1; 4744fail: 4745 double_rq_unlock(rq_src, rq_dest); 4746 raw_spin_unlock(&p->pi_lock); 4747 return ret; 4748} 4749 4750/* 4751 * migration_cpu_stop - this will be executed by a highprio stopper thread 4752 * and performs thread migration by bumping thread off CPU then 4753 * 'pushing' onto another runqueue. 4754 */ 4755static int migration_cpu_stop(void *data) 4756{ 4757 struct migration_arg *arg = data; 4758 4759 /* 4760 * The original target cpu might have gone down and we might 4761 * be on another cpu but it doesn't matter. 4762 */ 4763 local_irq_disable(); 4764 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu); 4765 local_irq_enable(); 4766 return 0; 4767} 4768 4769#ifdef CONFIG_HOTPLUG_CPU 4770 4771/* 4772 * Ensures that the idle task is using init_mm right before its cpu goes 4773 * offline. 4774 */ 4775void idle_task_exit(void) 4776{ 4777 struct mm_struct *mm = current->active_mm; 4778 4779 BUG_ON(cpu_online(smp_processor_id())); 4780 4781 if (mm != &init_mm) 4782 switch_mm(mm, &init_mm, current); 4783 mmdrop(mm); 4784} 4785 4786/* 4787 * Since this CPU is going 'away' for a while, fold any nr_active delta 4788 * we might have. Assumes we're called after migrate_tasks() so that the 4789 * nr_active count is stable. 4790 * 4791 * Also see the comment "Global load-average calculations". 4792 */ 4793static void calc_load_migrate(struct rq *rq) 4794{ 4795 long delta = calc_load_fold_active(rq); 4796 if (delta) 4797 atomic_long_add(delta, &calc_load_tasks); 4798} 4799 4800/* 4801 * Migrate all tasks from the rq, sleeping tasks will be migrated by 4802 * try_to_wake_up()->select_task_rq(). 4803 * 4804 * Called with rq->lock held even though we'er in stop_machine() and 4805 * there's no concurrency possible, we hold the required locks anyway 4806 * because of lock validation efforts. 4807 */ 4808static void migrate_tasks(unsigned int dead_cpu) 4809{ 4810 struct rq *rq = cpu_rq(dead_cpu); 4811 struct task_struct *next, *stop = rq->stop; 4812 int dest_cpu; 4813 4814 /* 4815 * Fudge the rq selection such that the below task selection loop 4816 * doesn't get stuck on the currently eligible stop task. 4817 * 4818 * We're currently inside stop_machine() and the rq is either stuck 4819 * in the stop_machine_cpu_stop() loop, or we're executing this code, 4820 * either way we should never end up calling schedule() until we're 4821 * done here. 4822 */ 4823 rq->stop = NULL; 4824 4825 for ( ; ; ) { 4826 /* 4827 * There's this thread running, bail when that's the only 4828 * remaining thread. 4829 */ 4830 if (rq->nr_running == 1) 4831 break; 4832 4833 next = pick_next_task(rq); 4834 BUG_ON(!next); 4835 next->sched_class->put_prev_task(rq, next); 4836 4837 /* Find suitable destination for @next, with force if needed. */ 4838 dest_cpu = select_fallback_rq(dead_cpu, next); 4839 raw_spin_unlock(&rq->lock); 4840 4841 __migrate_task(next, dead_cpu, dest_cpu); 4842 4843 raw_spin_lock(&rq->lock); 4844 } 4845 4846 rq->stop = stop; 4847} 4848 4849#endif /* CONFIG_HOTPLUG_CPU */ 4850 4851#if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL) 4852 4853static struct ctl_table sd_ctl_dir[] = { 4854 { 4855 .procname = "sched_domain", 4856 .mode = 0555, 4857 }, 4858 {} 4859}; 4860 4861static struct ctl_table sd_ctl_root[] = { 4862 { 4863 .procname = "kernel", 4864 .mode = 0555, 4865 .child = sd_ctl_dir, 4866 }, 4867 {} 4868}; 4869 4870static struct ctl_table *sd_alloc_ctl_entry(int n) 4871{ 4872 struct ctl_table *entry = 4873 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL); 4874 4875 return entry; 4876} 4877 4878static void sd_free_ctl_entry(struct ctl_table **tablep) 4879{ 4880 struct ctl_table *entry; 4881 4882 /* 4883 * In the intermediate directories, both the child directory and 4884 * procname are dynamically allocated and could fail but the mode 4885 * will always be set. In the lowest directory the names are 4886 * static strings and all have proc handlers. 4887 */ 4888 for (entry = *tablep; entry->mode; entry++) { 4889 if (entry->child) 4890 sd_free_ctl_entry(&entry->child); 4891 if (entry->proc_handler == NULL) 4892 kfree(entry->procname); 4893 } 4894 4895 kfree(*tablep); 4896 *tablep = NULL; 4897} 4898 4899static int min_load_idx = 0; 4900static int max_load_idx = CPU_LOAD_IDX_MAX; 4901 4902static void 4903set_table_entry(struct ctl_table *entry, 4904 const char *procname, void *data, int maxlen, 4905 umode_t mode, proc_handler *proc_handler, 4906 bool load_idx) 4907{ 4908 entry->procname = procname; 4909 entry->data = data; 4910 entry->maxlen = maxlen; 4911 entry->mode = mode; 4912 entry->proc_handler = proc_handler; 4913 4914 if (load_idx) { 4915 entry->extra1 = &min_load_idx; 4916 entry->extra2 = &max_load_idx; 4917 } 4918} 4919 4920static struct ctl_table * 4921sd_alloc_ctl_domain_table(struct sched_domain *sd) 4922{ 4923 struct ctl_table *table = sd_alloc_ctl_entry(13); 4924 4925 if (table == NULL) 4926 return NULL; 4927 4928 set_table_entry(&table[0], "min_interval", &sd->min_interval, 4929 sizeof(long), 0644, proc_doulongvec_minmax, false); 4930 set_table_entry(&table[1], "max_interval", &sd->max_interval, 4931 sizeof(long), 0644, proc_doulongvec_minmax, false); 4932 set_table_entry(&table[2], "busy_idx", &sd->busy_idx, 4933 sizeof(int), 0644, proc_dointvec_minmax, true); 4934 set_table_entry(&table[3], "idle_idx", &sd->idle_idx, 4935 sizeof(int), 0644, proc_dointvec_minmax, true); 4936 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx, 4937 sizeof(int), 0644, proc_dointvec_minmax, true); 4938 set_table_entry(&table[5], "wake_idx", &sd->wake_idx, 4939 sizeof(int), 0644, proc_dointvec_minmax, true); 4940 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx, 4941 sizeof(int), 0644, proc_dointvec_minmax, true); 4942 set_table_entry(&table[7], "busy_factor", &sd->busy_factor, 4943 sizeof(int), 0644, proc_dointvec_minmax, false); 4944 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct, 4945 sizeof(int), 0644, proc_dointvec_minmax, false); 4946 set_table_entry(&table[9], "cache_nice_tries", 4947 &sd->cache_nice_tries, 4948 sizeof(int), 0644, proc_dointvec_minmax, false); 4949 set_table_entry(&table[10], "flags", &sd->flags, 4950 sizeof(int), 0644, proc_dointvec_minmax, false); 4951 set_table_entry(&table[11], "name", sd->name, 4952 CORENAME_MAX_SIZE, 0444, proc_dostring, false); 4953 /* &table[12] is terminator */ 4954 4955 return table; 4956} 4957 4958static ctl_table *sd_alloc_ctl_cpu_table(int cpu) 4959{ 4960 struct ctl_table *entry, *table; 4961 struct sched_domain *sd; 4962 int domain_num = 0, i; 4963 char buf[32]; 4964 4965 for_each_domain(cpu, sd) 4966 domain_num++; 4967 entry = table = sd_alloc_ctl_entry(domain_num + 1); 4968 if (table == NULL) 4969 return NULL; 4970 4971 i = 0; 4972 for_each_domain(cpu, sd) { 4973 snprintf(buf, 32, "domain%d", i); 4974 entry->procname = kstrdup(buf, GFP_KERNEL); 4975 entry->mode = 0555; 4976 entry->child = sd_alloc_ctl_domain_table(sd); 4977 entry++; 4978 i++; 4979 } 4980 return table; 4981} 4982 4983static struct ctl_table_header *sd_sysctl_header; 4984static void register_sched_domain_sysctl(void) 4985{ 4986 int i, cpu_num = num_possible_cpus(); 4987 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1); 4988 char buf[32]; 4989 4990 WARN_ON(sd_ctl_dir[0].child); 4991 sd_ctl_dir[0].child = entry; 4992 4993 if (entry == NULL) 4994 return; 4995 4996 for_each_possible_cpu(i) { 4997 snprintf(buf, 32, "cpu%d", i); 4998 entry->procname = kstrdup(buf, GFP_KERNEL); 4999 entry->mode = 0555; 5000 entry->child = sd_alloc_ctl_cpu_table(i); 5001 entry++; 5002 } 5003 5004 WARN_ON(sd_sysctl_header); 5005 sd_sysctl_header = register_sysctl_table(sd_ctl_root); 5006} 5007 5008/* may be called multiple times per register */ 5009static void unregister_sched_domain_sysctl(void) 5010{ 5011 if (sd_sysctl_header) 5012 unregister_sysctl_table(sd_sysctl_header); 5013 sd_sysctl_header = NULL; 5014 if (sd_ctl_dir[0].child) 5015 sd_free_ctl_entry(&sd_ctl_dir[0].child); 5016} 5017#else 5018static void register_sched_domain_sysctl(void) 5019{ 5020} 5021static void unregister_sched_domain_sysctl(void) 5022{ 5023} 5024#endif 5025 5026static void set_rq_online(struct rq *rq) 5027{ 5028 if (!rq->online) { 5029 const struct sched_class *class; 5030 5031 cpumask_set_cpu(rq->cpu, rq->rd->online); 5032 rq->online = 1; 5033 5034 for_each_class(class) { 5035 if (class->rq_online) 5036 class->rq_online(rq); 5037 } 5038 } 5039} 5040 5041static void set_rq_offline(struct rq *rq) 5042{ 5043 if (rq->online) { 5044 const struct sched_class *class; 5045 5046 for_each_class(class) { 5047 if (class->rq_offline) 5048 class->rq_offline(rq); 5049 } 5050 5051 cpumask_clear_cpu(rq->cpu, rq->rd->online); 5052 rq->online = 0; 5053 } 5054} 5055 5056/* 5057 * migration_call - callback that gets triggered when a CPU is added. 5058 * Here we can start up the necessary migration thread for the new CPU. 5059 */ 5060static int __cpuinit 5061migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu) 5062{ 5063 int cpu = (long)hcpu; 5064 unsigned long flags; 5065 struct rq *rq = cpu_rq(cpu); 5066 5067 switch (action & ~CPU_TASKS_FROZEN) { 5068 5069 case CPU_UP_PREPARE: 5070 rq->calc_load_update = calc_load_update; 5071 break; 5072 5073 case CPU_ONLINE: 5074 /* Update our root-domain */ 5075 raw_spin_lock_irqsave(&rq->lock, flags); 5076 if (rq->rd) { 5077 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span)); 5078 5079 set_rq_online(rq); 5080 } 5081 raw_spin_unlock_irqrestore(&rq->lock, flags); 5082 break; 5083 5084#ifdef CONFIG_HOTPLUG_CPU 5085 case CPU_DYING: 5086 sched_ttwu_pending(); 5087 /* Update our root-domain */ 5088 raw_spin_lock_irqsave(&rq->lock, flags); 5089 if (rq->rd) { 5090 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span)); 5091 set_rq_offline(rq); 5092 } 5093 migrate_tasks(cpu); 5094 BUG_ON(rq->nr_running != 1); /* the migration thread */ 5095 raw_spin_unlock_irqrestore(&rq->lock, flags); 5096 5097 calc_load_migrate(rq); 5098 break; 5099#endif 5100 } 5101 5102 update_max_interval(); 5103 5104 return NOTIFY_OK; 5105} 5106 5107/* 5108 * Register at high priority so that task migration (migrate_all_tasks) 5109 * happens before everything else. This has to be lower priority than 5110 * the notifier in the perf_event subsystem, though. 5111 */ 5112static struct notifier_block __cpuinitdata migration_notifier = { 5113 .notifier_call = migration_call, 5114 .priority = CPU_PRI_MIGRATION, 5115}; 5116 5117static int __cpuinit sched_cpu_active(struct notifier_block *nfb, 5118 unsigned long action, void *hcpu) 5119{ 5120 switch (action & ~CPU_TASKS_FROZEN) { 5121 case CPU_STARTING: 5122 case CPU_DOWN_FAILED: 5123 set_cpu_active((long)hcpu, true); 5124 return NOTIFY_OK; 5125 default: 5126 return NOTIFY_DONE; 5127 } 5128} 5129 5130static int __cpuinit sched_cpu_inactive(struct notifier_block *nfb, 5131 unsigned long action, void *hcpu) 5132{ 5133 switch (action & ~CPU_TASKS_FROZEN) { 5134 case CPU_DOWN_PREPARE: 5135 set_cpu_active((long)hcpu, false); 5136 return NOTIFY_OK; 5137 default: 5138 return NOTIFY_DONE; 5139 } 5140} 5141 5142static int __init migration_init(void) 5143{ 5144 void *cpu = (void *)(long)smp_processor_id(); 5145 int err; 5146 5147 /* Initialize migration for the boot CPU */ 5148 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu); 5149 BUG_ON(err == NOTIFY_BAD); 5150 migration_call(&migration_notifier, CPU_ONLINE, cpu); 5151 register_cpu_notifier(&migration_notifier); 5152 5153 /* Register cpu active notifiers */ 5154 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE); 5155 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE); 5156 5157 return 0; 5158} 5159early_initcall(migration_init); 5160#endif 5161 5162#ifdef CONFIG_SMP 5163 5164static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */ 5165 5166#ifdef CONFIG_SCHED_DEBUG 5167 5168static __read_mostly int sched_debug_enabled; 5169 5170static int __init sched_debug_setup(char *str) 5171{ 5172 sched_debug_enabled = 1; 5173 5174 return 0; 5175} 5176early_param("sched_debug", sched_debug_setup); 5177 5178static inline bool sched_debug(void) 5179{ 5180 return sched_debug_enabled; 5181} 5182 5183static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level, 5184 struct cpumask *groupmask) 5185{ 5186 struct sched_group *group = sd->groups; 5187 char str[256]; 5188 5189 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd)); 5190 cpumask_clear(groupmask); 5191 5192 printk(KERN_DEBUG "%*s domain %d: ", level, "", level); 5193 5194 if (!(sd->flags & SD_LOAD_BALANCE)) { 5195 printk("does not load-balance\n"); 5196 if (sd->parent) 5197 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain" 5198 " has parent"); 5199 return -1; 5200 } 5201 5202 printk(KERN_CONT "span %s level %s\n", str, sd->name); 5203 5204 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) { 5205 printk(KERN_ERR "ERROR: domain->span does not contain " 5206 "CPU%d\n", cpu); 5207 } 5208 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) { 5209 printk(KERN_ERR "ERROR: domain->groups does not contain" 5210 " CPU%d\n", cpu); 5211 } 5212 5213 printk(KERN_DEBUG "%*s groups:", level + 1, ""); 5214 do { 5215 if (!group) { 5216 printk("\n"); 5217 printk(KERN_ERR "ERROR: group is NULL\n"); 5218 break; 5219 } 5220 5221 /* 5222 * Even though we initialize ->power to something semi-sane, 5223 * we leave power_orig unset. This allows us to detect if 5224 * domain iteration is still funny without causing /0 traps. 5225 */ 5226 if (!group->sgp->power_orig) { 5227 printk(KERN_CONT "\n"); 5228 printk(KERN_ERR "ERROR: domain->cpu_power not " 5229 "set\n"); 5230 break; 5231 } 5232 5233 if (!cpumask_weight(sched_group_cpus(group))) { 5234 printk(KERN_CONT "\n"); 5235 printk(KERN_ERR "ERROR: empty group\n"); 5236 break; 5237 } 5238 5239 if (!(sd->flags & SD_OVERLAP) && 5240 cpumask_intersects(groupmask, sched_group_cpus(group))) { 5241 printk(KERN_CONT "\n"); 5242 printk(KERN_ERR "ERROR: repeated CPUs\n"); 5243 break; 5244 } 5245 5246 cpumask_or(groupmask, groupmask, sched_group_cpus(group)); 5247 5248 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group)); 5249 5250 printk(KERN_CONT " %s", str); 5251 if (group->sgp->power != SCHED_POWER_SCALE) { 5252 printk(KERN_CONT " (cpu_power = %d)", 5253 group->sgp->power); 5254 } 5255 5256 group = group->next; 5257 } while (group != sd->groups); 5258 printk(KERN_CONT "\n"); 5259 5260 if (!cpumask_equal(sched_domain_span(sd), groupmask)) 5261 printk(KERN_ERR "ERROR: groups don't span domain->span\n"); 5262 5263 if (sd->parent && 5264 !cpumask_subset(groupmask, sched_domain_span(sd->parent))) 5265 printk(KERN_ERR "ERROR: parent span is not a superset " 5266 "of domain->span\n"); 5267 return 0; 5268} 5269 5270static void sched_domain_debug(struct sched_domain *sd, int cpu) 5271{ 5272 int level = 0; 5273 5274 if (!sched_debug_enabled) 5275 return; 5276 5277 if (!sd) { 5278 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu); 5279 return; 5280 } 5281 5282 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu); 5283 5284 for (;;) { 5285 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask)) 5286 break; 5287 level++; 5288 sd = sd->parent; 5289 if (!sd) 5290 break; 5291 } 5292} 5293#else /* !CONFIG_SCHED_DEBUG */ 5294# define sched_domain_debug(sd, cpu) do { } while (0) 5295static inline bool sched_debug(void) 5296{ 5297 return false; 5298} 5299#endif /* CONFIG_SCHED_DEBUG */ 5300 5301static int sd_degenerate(struct sched_domain *sd) 5302{ 5303 if (cpumask_weight(sched_domain_span(sd)) == 1) 5304 return 1; 5305 5306 /* Following flags need at least 2 groups */ 5307 if (sd->flags & (SD_LOAD_BALANCE | 5308 SD_BALANCE_NEWIDLE | 5309 SD_BALANCE_FORK | 5310 SD_BALANCE_EXEC | 5311 SD_SHARE_CPUPOWER | 5312 SD_SHARE_PKG_RESOURCES)) { 5313 if (sd->groups != sd->groups->next) 5314 return 0; 5315 } 5316 5317 /* Following flags don't use groups */ 5318 if (sd->flags & (SD_WAKE_AFFINE)) 5319 return 0; 5320 5321 return 1; 5322} 5323 5324static int 5325sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent) 5326{ 5327 unsigned long cflags = sd->flags, pflags = parent->flags; 5328 5329 if (sd_degenerate(parent)) 5330 return 1; 5331 5332 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent))) 5333 return 0; 5334 5335 /* Flags needing groups don't count if only 1 group in parent */ 5336 if (parent->groups == parent->groups->next) { 5337 pflags &= ~(SD_LOAD_BALANCE | 5338 SD_BALANCE_NEWIDLE | 5339 SD_BALANCE_FORK | 5340 SD_BALANCE_EXEC | 5341 SD_SHARE_CPUPOWER | 5342 SD_SHARE_PKG_RESOURCES); 5343 if (nr_node_ids == 1) 5344 pflags &= ~SD_SERIALIZE; 5345 } 5346 if (~cflags & pflags) 5347 return 0; 5348 5349 return 1; 5350} 5351 5352static void free_rootdomain(struct rcu_head *rcu) 5353{ 5354 struct root_domain *rd = container_of(rcu, struct root_domain, rcu); 5355 5356 cpupri_cleanup(&rd->cpupri); 5357 free_cpumask_var(rd->rto_mask); 5358 free_cpumask_var(rd->online); 5359 free_cpumask_var(rd->span); 5360 kfree(rd); 5361} 5362 5363static void rq_attach_root(struct rq *rq, struct root_domain *rd) 5364{ 5365 struct root_domain *old_rd = NULL; 5366 unsigned long flags; 5367 5368 raw_spin_lock_irqsave(&rq->lock, flags); 5369 5370 if (rq->rd) { 5371 old_rd = rq->rd; 5372 5373 if (cpumask_test_cpu(rq->cpu, old_rd->online)) 5374 set_rq_offline(rq); 5375 5376 cpumask_clear_cpu(rq->cpu, old_rd->span); 5377 5378 /* 5379 * If we dont want to free the old_rt yet then 5380 * set old_rd to NULL to skip the freeing later 5381 * in this function: 5382 */ 5383 if (!atomic_dec_and_test(&old_rd->refcount)) 5384 old_rd = NULL; 5385 } 5386 5387 atomic_inc(&rd->refcount); 5388 rq->rd = rd; 5389 5390 cpumask_set_cpu(rq->cpu, rd->span); 5391 if (cpumask_test_cpu(rq->cpu, cpu_active_mask)) 5392 set_rq_online(rq); 5393 5394 raw_spin_unlock_irqrestore(&rq->lock, flags); 5395 5396 if (old_rd) 5397 call_rcu_sched(&old_rd->rcu, free_rootdomain); 5398} 5399 5400static int init_rootdomain(struct root_domain *rd) 5401{ 5402 memset(rd, 0, sizeof(*rd)); 5403 5404 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL)) 5405 goto out; 5406 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL)) 5407 goto free_span; 5408 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL)) 5409 goto free_online; 5410 5411 if (cpupri_init(&rd->cpupri) != 0) 5412 goto free_rto_mask; 5413 return 0; 5414 5415free_rto_mask: 5416 free_cpumask_var(rd->rto_mask); 5417free_online: 5418 free_cpumask_var(rd->online); 5419free_span: 5420 free_cpumask_var(rd->span); 5421out: 5422 return -ENOMEM; 5423} 5424 5425/* 5426 * By default the system creates a single root-domain with all cpus as 5427 * members (mimicking the global state we have today). 5428 */ 5429struct root_domain def_root_domain; 5430 5431static void init_defrootdomain(void) 5432{ 5433 init_rootdomain(&def_root_domain); 5434 5435 atomic_set(&def_root_domain.refcount, 1); 5436} 5437 5438static struct root_domain *alloc_rootdomain(void) 5439{ 5440 struct root_domain *rd; 5441 5442 rd = kmalloc(sizeof(*rd), GFP_KERNEL); 5443 if (!rd) 5444 return NULL; 5445 5446 if (init_rootdomain(rd) != 0) { 5447 kfree(rd); 5448 return NULL; 5449 } 5450 5451 return rd; 5452} 5453 5454static void free_sched_groups(struct sched_group *sg, int free_sgp) 5455{ 5456 struct sched_group *tmp, *first; 5457 5458 if (!sg) 5459 return; 5460 5461 first = sg; 5462 do { 5463 tmp = sg->next; 5464 5465 if (free_sgp && atomic_dec_and_test(&sg->sgp->ref)) 5466 kfree(sg->sgp); 5467 5468 kfree(sg); 5469 sg = tmp; 5470 } while (sg != first); 5471} 5472 5473static void free_sched_domain(struct rcu_head *rcu) 5474{ 5475 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu); 5476 5477 /* 5478 * If its an overlapping domain it has private groups, iterate and 5479 * nuke them all. 5480 */ 5481 if (sd->flags & SD_OVERLAP) { 5482 free_sched_groups(sd->groups, 1); 5483 } else if (atomic_dec_and_test(&sd->groups->ref)) { 5484 kfree(sd->groups->sgp); 5485 kfree(sd->groups); 5486 } 5487 kfree(sd); 5488} 5489 5490static void destroy_sched_domain(struct sched_domain *sd, int cpu) 5491{ 5492 call_rcu(&sd->rcu, free_sched_domain); 5493} 5494 5495static void destroy_sched_domains(struct sched_domain *sd, int cpu) 5496{ 5497 for (; sd; sd = sd->parent) 5498 destroy_sched_domain(sd, cpu); 5499} 5500 5501/* 5502 * Keep a special pointer to the highest sched_domain that has 5503 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this 5504 * allows us to avoid some pointer chasing select_idle_sibling(). 5505 * 5506 * Iterate domains and sched_groups downward, assigning CPUs to be 5507 * select_idle_sibling() hw buddy. Cross-wiring hw makes bouncing 5508 * due to random perturbation self canceling, ie sw buddies pull 5509 * their counterpart to their CPU's hw counterpart. 5510 * 5511 * Also keep a unique ID per domain (we use the first cpu number in 5512 * the cpumask of the domain), this allows us to quickly tell if 5513 * two cpus are in the same cache domain, see cpus_share_cache(). 5514 */ 5515DEFINE_PER_CPU(struct sched_domain *, sd_llc); 5516DEFINE_PER_CPU(int, sd_llc_id); 5517 5518static void update_top_cache_domain(int cpu) 5519{ 5520 struct sched_domain *sd; 5521 int id = cpu; 5522 5523 sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES); 5524 if (sd) { 5525 struct sched_domain *tmp = sd; 5526 struct sched_group *sg, *prev; 5527 bool right; 5528 5529 /* 5530 * Traverse to first CPU in group, and count hops 5531 * to cpu from there, switching direction on each 5532 * hop, never ever pointing the last CPU rightward. 5533 */ 5534 do { 5535 id = cpumask_first(sched_domain_span(tmp)); 5536 prev = sg = tmp->groups; 5537 right = 1; 5538 5539 while (cpumask_first(sched_group_cpus(sg)) != id) 5540 sg = sg->next; 5541 5542 while (!cpumask_test_cpu(cpu, sched_group_cpus(sg))) { 5543 prev = sg; 5544 sg = sg->next; 5545 right = !right; 5546 } 5547 5548 /* A CPU went down, never point back to domain start. */ 5549 if (right && cpumask_first(sched_group_cpus(sg->next)) == id) 5550 right = false; 5551 5552 sg = right ? sg->next : prev; 5553 tmp->idle_buddy = cpumask_first(sched_group_cpus(sg)); 5554 } while ((tmp = tmp->child)); 5555 5556 id = cpumask_first(sched_domain_span(sd)); 5557 } 5558 5559 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd); 5560 per_cpu(sd_llc_id, cpu) = id; 5561} 5562 5563/* 5564 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must 5565 * hold the hotplug lock. 5566 */ 5567static void 5568cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu) 5569{ 5570 struct rq *rq = cpu_rq(cpu); 5571 struct sched_domain *tmp; 5572 5573 /* Remove the sched domains which do not contribute to scheduling. */ 5574 for (tmp = sd; tmp; ) { 5575 struct sched_domain *parent = tmp->parent; 5576 if (!parent) 5577 break; 5578 5579 if (sd_parent_degenerate(tmp, parent)) { 5580 tmp->parent = parent->parent; 5581 if (parent->parent) 5582 parent->parent->child = tmp; 5583 destroy_sched_domain(parent, cpu); 5584 } else 5585 tmp = tmp->parent; 5586 } 5587 5588 if (sd && sd_degenerate(sd)) { 5589 tmp = sd; 5590 sd = sd->parent; 5591 destroy_sched_domain(tmp, cpu); 5592 if (sd) 5593 sd->child = NULL; 5594 } 5595 5596 sched_domain_debug(sd, cpu); 5597 5598 rq_attach_root(rq, rd); 5599 tmp = rq->sd; 5600 rcu_assign_pointer(rq->sd, sd); 5601 destroy_sched_domains(tmp, cpu); 5602 5603 update_top_cache_domain(cpu); 5604} 5605 5606/* cpus with isolated domains */ 5607static cpumask_var_t cpu_isolated_map; 5608 5609/* Setup the mask of cpus configured for isolated domains */ 5610static int __init isolated_cpu_setup(char *str) 5611{ 5612 alloc_bootmem_cpumask_var(&cpu_isolated_map); 5613 cpulist_parse(str, cpu_isolated_map); 5614 return 1; 5615} 5616 5617__setup("isolcpus=", isolated_cpu_setup); 5618 5619static const struct cpumask *cpu_cpu_mask(int cpu) 5620{ 5621 return cpumask_of_node(cpu_to_node(cpu)); 5622} 5623 5624struct sd_data { 5625 struct sched_domain **__percpu sd; 5626 struct sched_group **__percpu sg; 5627 struct sched_group_power **__percpu sgp; 5628}; 5629 5630struct s_data { 5631 struct sched_domain ** __percpu sd; 5632 struct root_domain *rd; 5633}; 5634 5635enum s_alloc { 5636 sa_rootdomain, 5637 sa_sd, 5638 sa_sd_storage, 5639 sa_none, 5640}; 5641 5642struct sched_domain_topology_level; 5643 5644typedef struct sched_domain *(*sched_domain_init_f)(struct sched_domain_topology_level *tl, int cpu); 5645typedef const struct cpumask *(*sched_domain_mask_f)(int cpu); 5646 5647#define SDTL_OVERLAP 0x01 5648 5649struct sched_domain_topology_level { 5650 sched_domain_init_f init; 5651 sched_domain_mask_f mask; 5652 int flags; 5653 int numa_level; 5654 struct sd_data data; 5655}; 5656 5657/* 5658 * Build an iteration mask that can exclude certain CPUs from the upwards 5659 * domain traversal. 5660 * 5661 * Asymmetric node setups can result in situations where the domain tree is of 5662 * unequal depth, make sure to skip domains that already cover the entire 5663 * range. 5664 * 5665 * In that case build_sched_domains() will have terminated the iteration early 5666 * and our sibling sd spans will be empty. Domains should always include the 5667 * cpu they're built on, so check that. 5668 * 5669 */ 5670static void build_group_mask(struct sched_domain *sd, struct sched_group *sg) 5671{ 5672 const struct cpumask *span = sched_domain_span(sd); 5673 struct sd_data *sdd = sd->private; 5674 struct sched_domain *sibling; 5675 int i; 5676 5677 for_each_cpu(i, span) { 5678 sibling = *per_cpu_ptr(sdd->sd, i); 5679 if (!cpumask_test_cpu(i, sched_domain_span(sibling))) 5680 continue; 5681 5682 cpumask_set_cpu(i, sched_group_mask(sg)); 5683 } 5684} 5685 5686/* 5687 * Return the canonical balance cpu for this group, this is the first cpu 5688 * of this group that's also in the iteration mask. 5689 */ 5690int group_balance_cpu(struct sched_group *sg) 5691{ 5692 return cpumask_first_and(sched_group_cpus(sg), sched_group_mask(sg)); 5693} 5694 5695static int 5696build_overlap_sched_groups(struct sched_domain *sd, int cpu) 5697{ 5698 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg; 5699 const struct cpumask *span = sched_domain_span(sd); 5700 struct cpumask *covered = sched_domains_tmpmask; 5701 struct sd_data *sdd = sd->private; 5702 struct sched_domain *child; 5703 int i; 5704 5705 cpumask_clear(covered); 5706 5707 for_each_cpu(i, span) { 5708 struct cpumask *sg_span; 5709 5710 if (cpumask_test_cpu(i, covered)) 5711 continue; 5712 5713 child = *per_cpu_ptr(sdd->sd, i); 5714 5715 /* See the comment near build_group_mask(). */ 5716 if (!cpumask_test_cpu(i, sched_domain_span(child))) 5717 continue; 5718 5719 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(), 5720 GFP_KERNEL, cpu_to_node(cpu)); 5721 5722 if (!sg) 5723 goto fail; 5724 5725 sg_span = sched_group_cpus(sg); 5726 if (child->child) { 5727 child = child->child; 5728 cpumask_copy(sg_span, sched_domain_span(child)); 5729 } else 5730 cpumask_set_cpu(i, sg_span); 5731 5732 cpumask_or(covered, covered, sg_span); 5733 5734 sg->sgp = *per_cpu_ptr(sdd->sgp, i); 5735 if (atomic_inc_return(&sg->sgp->ref) == 1) 5736 build_group_mask(sd, sg); 5737 5738 /* 5739 * Initialize sgp->power such that even if we mess up the 5740 * domains and no possible iteration will get us here, we won't 5741 * die on a /0 trap. 5742 */ 5743 sg->sgp->power = SCHED_POWER_SCALE * cpumask_weight(sg_span); 5744 5745 /* 5746 * Make sure the first group of this domain contains the 5747 * canonical balance cpu. Otherwise the sched_domain iteration 5748 * breaks. See update_sg_lb_stats(). 5749 */ 5750 if ((!groups && cpumask_test_cpu(cpu, sg_span)) || 5751 group_balance_cpu(sg) == cpu) 5752 groups = sg; 5753 5754 if (!first) 5755 first = sg; 5756 if (last) 5757 last->next = sg; 5758 last = sg; 5759 last->next = first; 5760 } 5761 sd->groups = groups; 5762 5763 return 0; 5764 5765fail: 5766 free_sched_groups(first, 0); 5767 5768 return -ENOMEM; 5769} 5770 5771static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg) 5772{ 5773 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu); 5774 struct sched_domain *child = sd->child; 5775 5776 if (child) 5777 cpu = cpumask_first(sched_domain_span(child)); 5778 5779 if (sg) { 5780 *sg = *per_cpu_ptr(sdd->sg, cpu); 5781 (*sg)->sgp = *per_cpu_ptr(sdd->sgp, cpu); 5782 atomic_set(&(*sg)->sgp->ref, 1); /* for claim_allocations */ 5783 } 5784 5785 return cpu; 5786} 5787 5788/* 5789 * build_sched_groups will build a circular linked list of the groups 5790 * covered by the given span, and will set each group's ->cpumask correctly, 5791 * and ->cpu_power to 0. 5792 * 5793 * Assumes the sched_domain tree is fully constructed 5794 */ 5795static int 5796build_sched_groups(struct sched_domain *sd, int cpu) 5797{ 5798 struct sched_group *first = NULL, *last = NULL; 5799 struct sd_data *sdd = sd->private; 5800 const struct cpumask *span = sched_domain_span(sd); 5801 struct cpumask *covered; 5802 int i; 5803 5804 get_group(cpu, sdd, &sd->groups); 5805 atomic_inc(&sd->groups->ref); 5806 5807 if (cpu != cpumask_first(sched_domain_span(sd))) 5808 return 0; 5809 5810 lockdep_assert_held(&sched_domains_mutex); 5811 covered = sched_domains_tmpmask; 5812 5813 cpumask_clear(covered); 5814 5815 for_each_cpu(i, span) { 5816 struct sched_group *sg; 5817 int group = get_group(i, sdd, &sg); 5818 int j; 5819 5820 if (cpumask_test_cpu(i, covered)) 5821 continue; 5822 5823 cpumask_clear(sched_group_cpus(sg)); 5824 sg->sgp->power = 0; 5825 cpumask_setall(sched_group_mask(sg)); 5826 5827 for_each_cpu(j, span) { 5828 if (get_group(j, sdd, NULL) != group) 5829 continue; 5830 5831 cpumask_set_cpu(j, covered); 5832 cpumask_set_cpu(j, sched_group_cpus(sg)); 5833 } 5834 5835 if (!first) 5836 first = sg; 5837 if (last) 5838 last->next = sg; 5839 last = sg; 5840 } 5841 last->next = first; 5842 5843 return 0; 5844} 5845 5846/* 5847 * Initialize sched groups cpu_power. 5848 * 5849 * cpu_power indicates the capacity of sched group, which is used while 5850 * distributing the load between different sched groups in a sched domain. 5851 * Typically cpu_power for all the groups in a sched domain will be same unless 5852 * there are asymmetries in the topology. If there are asymmetries, group 5853 * having more cpu_power will pickup more load compared to the group having 5854 * less cpu_power. 5855 */ 5856static void init_sched_groups_power(int cpu, struct sched_domain *sd) 5857{ 5858 struct sched_group *sg = sd->groups; 5859 5860 WARN_ON(!sd || !sg); 5861 5862 do { 5863 sg->group_weight = cpumask_weight(sched_group_cpus(sg)); 5864 sg = sg->next; 5865 } while (sg != sd->groups); 5866 5867 if (cpu != group_balance_cpu(sg)) 5868 return; 5869 5870 update_group_power(sd, cpu); 5871 atomic_set(&sg->sgp->nr_busy_cpus, sg->group_weight); 5872} 5873 5874int __weak arch_sd_sibling_asym_packing(void) 5875{ 5876 return 0*SD_ASYM_PACKING; 5877} 5878 5879/* 5880 * Initializers for schedule domains 5881 * Non-inlined to reduce accumulated stack pressure in build_sched_domains() 5882 */ 5883 5884#ifdef CONFIG_SCHED_DEBUG 5885# define SD_INIT_NAME(sd, type) sd->name = #type 5886#else 5887# define SD_INIT_NAME(sd, type) do { } while (0) 5888#endif 5889 5890#define SD_INIT_FUNC(type) \ 5891static noinline struct sched_domain * \ 5892sd_init_##type(struct sched_domain_topology_level *tl, int cpu) \ 5893{ \ 5894 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); \ 5895 *sd = SD_##type##_INIT; \ 5896 SD_INIT_NAME(sd, type); \ 5897 sd->private = &tl->data; \ 5898 return sd; \ 5899} 5900 5901SD_INIT_FUNC(CPU) 5902#ifdef CONFIG_SCHED_SMT 5903 SD_INIT_FUNC(SIBLING) 5904#endif 5905#ifdef CONFIG_SCHED_MC 5906 SD_INIT_FUNC(MC) 5907#endif 5908#ifdef CONFIG_SCHED_BOOK 5909 SD_INIT_FUNC(BOOK) 5910#endif 5911 5912static int default_relax_domain_level = -1; 5913int sched_domain_level_max; 5914 5915static int __init setup_relax_domain_level(char *str) 5916{ 5917 if (kstrtoint(str, 0, &default_relax_domain_level)) 5918 pr_warn("Unable to set relax_domain_level\n"); 5919 5920 return 1; 5921} 5922__setup("relax_domain_level=", setup_relax_domain_level); 5923 5924static void set_domain_attribute(struct sched_domain *sd, 5925 struct sched_domain_attr *attr) 5926{ 5927 int request; 5928 5929 if (!attr || attr->relax_domain_level < 0) { 5930 if (default_relax_domain_level < 0) 5931 return; 5932 else 5933 request = default_relax_domain_level; 5934 } else 5935 request = attr->relax_domain_level; 5936 if (request < sd->level) { 5937 /* turn off idle balance on this domain */ 5938 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE); 5939 } else { 5940 /* turn on idle balance on this domain */ 5941 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE); 5942 } 5943} 5944 5945static void __sdt_free(const struct cpumask *cpu_map); 5946static int __sdt_alloc(const struct cpumask *cpu_map); 5947 5948static void __free_domain_allocs(struct s_data *d, enum s_alloc what, 5949 const struct cpumask *cpu_map) 5950{ 5951 switch (what) { 5952 case sa_rootdomain: 5953 if (!atomic_read(&d->rd->refcount)) 5954 free_rootdomain(&d->rd->rcu); /* fall through */ 5955 case sa_sd: 5956 free_percpu(d->sd); /* fall through */ 5957 case sa_sd_storage: 5958 __sdt_free(cpu_map); /* fall through */ 5959 case sa_none: 5960 break; 5961 } 5962} 5963 5964static enum s_alloc __visit_domain_allocation_hell(struct s_data *d, 5965 const struct cpumask *cpu_map) 5966{ 5967 memset(d, 0, sizeof(*d)); 5968 5969 if (__sdt_alloc(cpu_map)) 5970 return sa_sd_storage; 5971 d->sd = alloc_percpu(struct sched_domain *); 5972 if (!d->sd) 5973 return sa_sd_storage; 5974 d->rd = alloc_rootdomain(); 5975 if (!d->rd) 5976 return sa_sd; 5977 return sa_rootdomain; 5978} 5979 5980/* 5981 * NULL the sd_data elements we've used to build the sched_domain and 5982 * sched_group structure so that the subsequent __free_domain_allocs() 5983 * will not free the data we're using. 5984 */ 5985static void claim_allocations(int cpu, struct sched_domain *sd) 5986{ 5987 struct sd_data *sdd = sd->private; 5988 5989 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd); 5990 *per_cpu_ptr(sdd->sd, cpu) = NULL; 5991 5992 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref)) 5993 *per_cpu_ptr(sdd->sg, cpu) = NULL; 5994 5995 if (atomic_read(&(*per_cpu_ptr(sdd->sgp, cpu))->ref)) 5996 *per_cpu_ptr(sdd->sgp, cpu) = NULL; 5997} 5998 5999#ifdef CONFIG_SCHED_SMT 6000static const struct cpumask *cpu_smt_mask(int cpu) 6001{ 6002 return topology_thread_cpumask(cpu); 6003} 6004#endif 6005 6006/* 6007 * Topology list, bottom-up. 6008 */ 6009static struct sched_domain_topology_level default_topology[] = { 6010#ifdef CONFIG_SCHED_SMT 6011 { sd_init_SIBLING, cpu_smt_mask, }, 6012#endif 6013#ifdef CONFIG_SCHED_MC 6014 { sd_init_MC, cpu_coregroup_mask, }, 6015#endif 6016#ifdef CONFIG_SCHED_BOOK 6017 { sd_init_BOOK, cpu_book_mask, }, 6018#endif 6019 { sd_init_CPU, cpu_cpu_mask, }, 6020 { NULL, }, 6021}; 6022 6023static struct sched_domain_topology_level *sched_domain_topology = default_topology; 6024 6025#ifdef CONFIG_NUMA 6026 6027static int sched_domains_numa_levels; 6028static int *sched_domains_numa_distance; 6029static struct cpumask ***sched_domains_numa_masks; 6030static int sched_domains_curr_level; 6031 6032static inline int sd_local_flags(int level) 6033{ 6034 if (sched_domains_numa_distance[level] > RECLAIM_DISTANCE) 6035 return 0; 6036 6037 return SD_BALANCE_EXEC | SD_BALANCE_FORK | SD_WAKE_AFFINE; 6038} 6039 6040static struct sched_domain * 6041sd_numa_init(struct sched_domain_topology_level *tl, int cpu) 6042{ 6043 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); 6044 int level = tl->numa_level; 6045 int sd_weight = cpumask_weight( 6046 sched_domains_numa_masks[level][cpu_to_node(cpu)]); 6047 6048 *sd = (struct sched_domain){ 6049 .min_interval = sd_weight, 6050 .max_interval = 2*sd_weight, 6051 .busy_factor = 32, 6052 .imbalance_pct = 125, 6053 .cache_nice_tries = 2, 6054 .busy_idx = 3, 6055 .idle_idx = 2, 6056 .newidle_idx = 0, 6057 .wake_idx = 0, 6058 .forkexec_idx = 0, 6059 6060 .flags = 1*SD_LOAD_BALANCE 6061 | 1*SD_BALANCE_NEWIDLE 6062 | 0*SD_BALANCE_EXEC 6063 | 0*SD_BALANCE_FORK 6064 | 0*SD_BALANCE_WAKE 6065 | 0*SD_WAKE_AFFINE 6066 | 0*SD_SHARE_CPUPOWER 6067 | 0*SD_SHARE_PKG_RESOURCES 6068 | 1*SD_SERIALIZE 6069 | 0*SD_PREFER_SIBLING 6070 | sd_local_flags(level) 6071 , 6072 .last_balance = jiffies, 6073 .balance_interval = sd_weight, 6074 }; 6075 SD_INIT_NAME(sd, NUMA); 6076 sd->private = &tl->data; 6077 6078 /* 6079 * Ugly hack to pass state to sd_numa_mask()... 6080 */ 6081 sched_domains_curr_level = tl->numa_level; 6082 6083 return sd; 6084} 6085 6086static const struct cpumask *sd_numa_mask(int cpu) 6087{ 6088 return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)]; 6089} 6090 6091static void sched_numa_warn(const char *str) 6092{ 6093 static int done = false; 6094 int i,j; 6095 6096 if (done) 6097 return; 6098 6099 done = true; 6100 6101 printk(KERN_WARNING "ERROR: %s\n\n", str); 6102 6103 for (i = 0; i < nr_node_ids; i++) { 6104 printk(KERN_WARNING " "); 6105 for (j = 0; j < nr_node_ids; j++) 6106 printk(KERN_CONT "%02d ", node_distance(i,j)); 6107 printk(KERN_CONT "\n"); 6108 } 6109 printk(KERN_WARNING "\n"); 6110} 6111 6112static bool find_numa_distance(int distance) 6113{ 6114 int i; 6115 6116 if (distance == node_distance(0, 0)) 6117 return true; 6118 6119 for (i = 0; i < sched_domains_numa_levels; i++) { 6120 if (sched_domains_numa_distance[i] == distance) 6121 return true; 6122 } 6123 6124 return false; 6125} 6126 6127static void sched_init_numa(void) 6128{ 6129 int next_distance, curr_distance = node_distance(0, 0); 6130 struct sched_domain_topology_level *tl; 6131 int level = 0; 6132 int i, j, k; 6133 6134 sched_domains_numa_distance = kzalloc(sizeof(int) * nr_node_ids, GFP_KERNEL); 6135 if (!sched_domains_numa_distance) 6136 return; 6137 6138 /* 6139 * O(nr_nodes^2) deduplicating selection sort -- in order to find the 6140 * unique distances in the node_distance() table. 6141 * 6142 * Assumes node_distance(0,j) includes all distances in 6143 * node_distance(i,j) in order to avoid cubic time. 6144 */ 6145 next_distance = curr_distance; 6146 for (i = 0; i < nr_node_ids; i++) { 6147 for (j = 0; j < nr_node_ids; j++) { 6148 for (k = 0; k < nr_node_ids; k++) { 6149 int distance = node_distance(i, k); 6150 6151 if (distance > curr_distance && 6152 (distance < next_distance || 6153 next_distance == curr_distance)) 6154 next_distance = distance; 6155 6156 /* 6157 * While not a strong assumption it would be nice to know 6158 * about cases where if node A is connected to B, B is not 6159 * equally connected to A. 6160 */ 6161 if (sched_debug() && node_distance(k, i) != distance) 6162 sched_numa_warn("Node-distance not symmetric"); 6163 6164 if (sched_debug() && i && !find_numa_distance(distance)) 6165 sched_numa_warn("Node-0 not representative"); 6166 } 6167 if (next_distance != curr_distance) { 6168 sched_domains_numa_distance[level++] = next_distance; 6169 sched_domains_numa_levels = level; 6170 curr_distance = next_distance; 6171 } else break; 6172 } 6173 6174 /* 6175 * In case of sched_debug() we verify the above assumption. 6176 */ 6177 if (!sched_debug()) 6178 break; 6179 } 6180 /* 6181 * 'level' contains the number of unique distances, excluding the 6182 * identity distance node_distance(i,i). 6183 * 6184 * The sched_domains_nume_distance[] array includes the actual distance 6185 * numbers. 6186 */ 6187 6188 sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL); 6189 if (!sched_domains_numa_masks) 6190 return; 6191 6192 /* 6193 * Now for each level, construct a mask per node which contains all 6194 * cpus of nodes that are that many hops away from us. 6195 */ 6196 for (i = 0; i < level; i++) { 6197 sched_domains_numa_masks[i] = 6198 kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL); 6199 if (!sched_domains_numa_masks[i]) 6200 return; 6201 6202 for (j = 0; j < nr_node_ids; j++) { 6203 struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL); 6204 if (!mask) 6205 return; 6206 6207 sched_domains_numa_masks[i][j] = mask; 6208 6209 for (k = 0; k < nr_node_ids; k++) { 6210 if (node_distance(j, k) > sched_domains_numa_distance[i]) 6211 continue; 6212 6213 cpumask_or(mask, mask, cpumask_of_node(k)); 6214 } 6215 } 6216 } 6217 6218 tl = kzalloc((ARRAY_SIZE(default_topology) + level) * 6219 sizeof(struct sched_domain_topology_level), GFP_KERNEL); 6220 if (!tl) 6221 return; 6222 6223 /* 6224 * Copy the default topology bits.. 6225 */ 6226 for (i = 0; default_topology[i].init; i++) 6227 tl[i] = default_topology[i]; 6228 6229 /* 6230 * .. and append 'j' levels of NUMA goodness. 6231 */ 6232 for (j = 0; j < level; i++, j++) { 6233 tl[i] = (struct sched_domain_topology_level){ 6234 .init = sd_numa_init, 6235 .mask = sd_numa_mask, 6236 .flags = SDTL_OVERLAP, 6237 .numa_level = j, 6238 }; 6239 } 6240 6241 sched_domain_topology = tl; 6242} 6243#else 6244static inline void sched_init_numa(void) 6245{ 6246} 6247#endif /* CONFIG_NUMA */ 6248 6249static int __sdt_alloc(const struct cpumask *cpu_map) 6250{ 6251 struct sched_domain_topology_level *tl; 6252 int j; 6253 6254 for (tl = sched_domain_topology; tl->init; tl++) { 6255 struct sd_data *sdd = &tl->data; 6256 6257 sdd->sd = alloc_percpu(struct sched_domain *); 6258 if (!sdd->sd) 6259 return -ENOMEM; 6260 6261 sdd->sg = alloc_percpu(struct sched_group *); 6262 if (!sdd->sg) 6263 return -ENOMEM; 6264 6265 sdd->sgp = alloc_percpu(struct sched_group_power *); 6266 if (!sdd->sgp) 6267 return -ENOMEM; 6268 6269 for_each_cpu(j, cpu_map) { 6270 struct sched_domain *sd; 6271 struct sched_group *sg; 6272 struct sched_group_power *sgp; 6273 6274 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(), 6275 GFP_KERNEL, cpu_to_node(j)); 6276 if (!sd) 6277 return -ENOMEM; 6278 6279 *per_cpu_ptr(sdd->sd, j) = sd; 6280 6281 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(), 6282 GFP_KERNEL, cpu_to_node(j)); 6283 if (!sg) 6284 return -ENOMEM; 6285 6286 sg->next = sg; 6287 6288 *per_cpu_ptr(sdd->sg, j) = sg; 6289 6290 sgp = kzalloc_node(sizeof(struct sched_group_power) + cpumask_size(), 6291 GFP_KERNEL, cpu_to_node(j)); 6292 if (!sgp) 6293 return -ENOMEM; 6294 6295 *per_cpu_ptr(sdd->sgp, j) = sgp; 6296 } 6297 } 6298 6299 return 0; 6300} 6301 6302static void __sdt_free(const struct cpumask *cpu_map) 6303{ 6304 struct sched_domain_topology_level *tl; 6305 int j; 6306 6307 for (tl = sched_domain_topology; tl->init; tl++) { 6308 struct sd_data *sdd = &tl->data; 6309 6310 for_each_cpu(j, cpu_map) { 6311 struct sched_domain *sd; 6312 6313 if (sdd->sd) { 6314 sd = *per_cpu_ptr(sdd->sd, j); 6315 if (sd && (sd->flags & SD_OVERLAP)) 6316 free_sched_groups(sd->groups, 0); 6317 kfree(*per_cpu_ptr(sdd->sd, j)); 6318 } 6319 6320 if (sdd->sg) 6321 kfree(*per_cpu_ptr(sdd->sg, j)); 6322 if (sdd->sgp) 6323 kfree(*per_cpu_ptr(sdd->sgp, j)); 6324 } 6325 free_percpu(sdd->sd); 6326 sdd->sd = NULL; 6327 free_percpu(sdd->sg); 6328 sdd->sg = NULL; 6329 free_percpu(sdd->sgp); 6330 sdd->sgp = NULL; 6331 } 6332} 6333 6334struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl, 6335 struct s_data *d, const struct cpumask *cpu_map, 6336 struct sched_domain_attr *attr, struct sched_domain *child, 6337 int cpu) 6338{ 6339 struct sched_domain *sd = tl->init(tl, cpu); 6340 if (!sd) 6341 return child; 6342 6343 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu)); 6344 if (child) { 6345 sd->level = child->level + 1; 6346 sched_domain_level_max = max(sched_domain_level_max, sd->level); 6347 child->parent = sd; 6348 } 6349 sd->child = child; 6350 set_domain_attribute(sd, attr); 6351 6352 return sd; 6353} 6354 6355/* 6356 * Build sched domains for a given set of cpus and attach the sched domains 6357 * to the individual cpus 6358 */ 6359static int build_sched_domains(const struct cpumask *cpu_map, 6360 struct sched_domain_attr *attr) 6361{ 6362 enum s_alloc alloc_state = sa_none; 6363 struct sched_domain *sd; 6364 struct s_data d; 6365 int i, ret = -ENOMEM; 6366 6367 alloc_state = __visit_domain_allocation_hell(&d, cpu_map); 6368 if (alloc_state != sa_rootdomain) 6369 goto error; 6370 6371 /* Set up domains for cpus specified by the cpu_map. */ 6372 for_each_cpu(i, cpu_map) { 6373 struct sched_domain_topology_level *tl; 6374 6375 sd = NULL; 6376 for (tl = sched_domain_topology; tl->init; tl++) { 6377 sd = build_sched_domain(tl, &d, cpu_map, attr, sd, i); 6378 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP)) 6379 sd->flags |= SD_OVERLAP; 6380 if (cpumask_equal(cpu_map, sched_domain_span(sd))) 6381 break; 6382 } 6383 6384 while (sd->child) 6385 sd = sd->child; 6386 6387 *per_cpu_ptr(d.sd, i) = sd; 6388 } 6389 6390 /* Build the groups for the domains */ 6391 for_each_cpu(i, cpu_map) { 6392 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) { 6393 sd->span_weight = cpumask_weight(sched_domain_span(sd)); 6394 if (sd->flags & SD_OVERLAP) { 6395 if (build_overlap_sched_groups(sd, i)) 6396 goto error; 6397 } else { 6398 if (build_sched_groups(sd, i)) 6399 goto error; 6400 } 6401 } 6402 } 6403 6404 /* Calculate CPU power for physical packages and nodes */ 6405 for (i = nr_cpumask_bits-1; i >= 0; i--) { 6406 if (!cpumask_test_cpu(i, cpu_map)) 6407 continue; 6408 6409 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) { 6410 claim_allocations(i, sd); 6411 init_sched_groups_power(i, sd); 6412 } 6413 } 6414 6415 /* Attach the domains */ 6416 rcu_read_lock(); 6417 for_each_cpu(i, cpu_map) { 6418 sd = *per_cpu_ptr(d.sd, i); 6419 cpu_attach_domain(sd, d.rd, i); 6420 } 6421 rcu_read_unlock(); 6422 6423 ret = 0; 6424error: 6425 __free_domain_allocs(&d, alloc_state, cpu_map); 6426 return ret; 6427} 6428 6429static cpumask_var_t *doms_cur; /* current sched domains */ 6430static int ndoms_cur; /* number of sched domains in 'doms_cur' */ 6431static struct sched_domain_attr *dattr_cur; 6432 /* attribues of custom domains in 'doms_cur' */ 6433 6434/* 6435 * Special case: If a kmalloc of a doms_cur partition (array of 6436 * cpumask) fails, then fallback to a single sched domain, 6437 * as determined by the single cpumask fallback_doms. 6438 */ 6439static cpumask_var_t fallback_doms; 6440 6441/* 6442 * arch_update_cpu_topology lets virtualized architectures update the 6443 * cpu core maps. It is supposed to return 1 if the topology changed 6444 * or 0 if it stayed the same. 6445 */ 6446int __attribute__((weak)) arch_update_cpu_topology(void) 6447{ 6448 return 0; 6449} 6450 6451cpumask_var_t *alloc_sched_domains(unsigned int ndoms) 6452{ 6453 int i; 6454 cpumask_var_t *doms; 6455 6456 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL); 6457 if (!doms) 6458 return NULL; 6459 for (i = 0; i < ndoms; i++) { 6460 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) { 6461 free_sched_domains(doms, i); 6462 return NULL; 6463 } 6464 } 6465 return doms; 6466} 6467 6468void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms) 6469{ 6470 unsigned int i; 6471 for (i = 0; i < ndoms; i++) 6472 free_cpumask_var(doms[i]); 6473 kfree(doms); 6474} 6475 6476/* 6477 * Set up scheduler domains and groups. Callers must hold the hotplug lock. 6478 * For now this just excludes isolated cpus, but could be used to 6479 * exclude other special cases in the future. 6480 */ 6481static int init_sched_domains(const struct cpumask *cpu_map) 6482{ 6483 int err; 6484 6485 arch_update_cpu_topology(); 6486 ndoms_cur = 1; 6487 doms_cur = alloc_sched_domains(ndoms_cur); 6488 if (!doms_cur) 6489 doms_cur = &fallback_doms; 6490 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map); 6491 err = build_sched_domains(doms_cur[0], NULL); 6492 register_sched_domain_sysctl(); 6493 6494 return err; 6495} 6496 6497/* 6498 * Detach sched domains from a group of cpus specified in cpu_map 6499 * These cpus will now be attached to the NULL domain 6500 */ 6501static void detach_destroy_domains(const struct cpumask *cpu_map) 6502{ 6503 int i; 6504 6505 rcu_read_lock(); 6506 for_each_cpu(i, cpu_map) 6507 cpu_attach_domain(NULL, &def_root_domain, i); 6508 rcu_read_unlock(); 6509} 6510 6511/* handle null as "default" */ 6512static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur, 6513 struct sched_domain_attr *new, int idx_new) 6514{ 6515 struct sched_domain_attr tmp; 6516 6517 /* fast path */ 6518 if (!new && !cur) 6519 return 1; 6520 6521 tmp = SD_ATTR_INIT; 6522 return !memcmp(cur ? (cur + idx_cur) : &tmp, 6523 new ? (new + idx_new) : &tmp, 6524 sizeof(struct sched_domain_attr)); 6525} 6526 6527/* 6528 * Partition sched domains as specified by the 'ndoms_new' 6529 * cpumasks in the array doms_new[] of cpumasks. This compares 6530 * doms_new[] to the current sched domain partitioning, doms_cur[]. 6531 * It destroys each deleted domain and builds each new domain. 6532 * 6533 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'. 6534 * The masks don't intersect (don't overlap.) We should setup one 6535 * sched domain for each mask. CPUs not in any of the cpumasks will 6536 * not be load balanced. If the same cpumask appears both in the 6537 * current 'doms_cur' domains and in the new 'doms_new', we can leave 6538 * it as it is. 6539 * 6540 * The passed in 'doms_new' should be allocated using 6541 * alloc_sched_domains. This routine takes ownership of it and will 6542 * free_sched_domains it when done with it. If the caller failed the 6543 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1, 6544 * and partition_sched_domains() will fallback to the single partition 6545 * 'fallback_doms', it also forces the domains to be rebuilt. 6546 * 6547 * If doms_new == NULL it will be replaced with cpu_online_mask. 6548 * ndoms_new == 0 is a special case for destroying existing domains, 6549 * and it will not create the default domain. 6550 * 6551 * Call with hotplug lock held 6552 */ 6553void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[], 6554 struct sched_domain_attr *dattr_new) 6555{ 6556 int i, j, n; 6557 int new_topology; 6558 6559 mutex_lock(&sched_domains_mutex); 6560 6561 /* always unregister in case we don't destroy any domains */ 6562 unregister_sched_domain_sysctl(); 6563 6564 /* Let architecture update cpu core mappings. */ 6565 new_topology = arch_update_cpu_topology(); 6566 6567 n = doms_new ? ndoms_new : 0; 6568 6569 /* Destroy deleted domains */ 6570 for (i = 0; i < ndoms_cur; i++) { 6571 for (j = 0; j < n && !new_topology; j++) { 6572 if (cpumask_equal(doms_cur[i], doms_new[j]) 6573 && dattrs_equal(dattr_cur, i, dattr_new, j)) 6574 goto match1; 6575 } 6576 /* no match - a current sched domain not in new doms_new[] */ 6577 detach_destroy_domains(doms_cur[i]); 6578match1: 6579 ; 6580 } 6581 6582 if (doms_new == NULL) { 6583 ndoms_cur = 0; 6584 doms_new = &fallback_doms; 6585 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map); 6586 WARN_ON_ONCE(dattr_new); 6587 } 6588 6589 /* Build new domains */ 6590 for (i = 0; i < ndoms_new; i++) { 6591 for (j = 0; j < ndoms_cur && !new_topology; j++) { 6592 if (cpumask_equal(doms_new[i], doms_cur[j]) 6593 && dattrs_equal(dattr_new, i, dattr_cur, j)) 6594 goto match2; 6595 } 6596 /* no match - add a new doms_new */ 6597 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL); 6598match2: 6599 ; 6600 } 6601 6602 /* Remember the new sched domains */ 6603 if (doms_cur != &fallback_doms) 6604 free_sched_domains(doms_cur, ndoms_cur); 6605 kfree(dattr_cur); /* kfree(NULL) is safe */ 6606 doms_cur = doms_new; 6607 dattr_cur = dattr_new; 6608 ndoms_cur = ndoms_new; 6609 6610 register_sched_domain_sysctl(); 6611 6612 mutex_unlock(&sched_domains_mutex); 6613} 6614 6615static int num_cpus_frozen; /* used to mark begin/end of suspend/resume */ 6616 6617/* 6618 * Update cpusets according to cpu_active mask. If cpusets are 6619 * disabled, cpuset_update_active_cpus() becomes a simple wrapper 6620 * around partition_sched_domains(). 6621 * 6622 * If we come here as part of a suspend/resume, don't touch cpusets because we 6623 * want to restore it back to its original state upon resume anyway. 6624 */ 6625static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action, 6626 void *hcpu) 6627{ 6628 switch (action) { 6629 case CPU_ONLINE_FROZEN: 6630 case CPU_DOWN_FAILED_FROZEN: 6631 6632 /* 6633 * num_cpus_frozen tracks how many CPUs are involved in suspend 6634 * resume sequence. As long as this is not the last online 6635 * operation in the resume sequence, just build a single sched 6636 * domain, ignoring cpusets. 6637 */ 6638 num_cpus_frozen--; 6639 if (likely(num_cpus_frozen)) { 6640 partition_sched_domains(1, NULL, NULL); 6641 break; 6642 } 6643 6644 /* 6645 * This is the last CPU online operation. So fall through and 6646 * restore the original sched domains by considering the 6647 * cpuset configurations. 6648 */ 6649 6650 case CPU_ONLINE: 6651 case CPU_DOWN_FAILED: 6652 cpuset_update_active_cpus(true); 6653 break; 6654 default: 6655 return NOTIFY_DONE; 6656 } 6657 return NOTIFY_OK; 6658} 6659 6660static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action, 6661 void *hcpu) 6662{ 6663 switch (action) { 6664 case CPU_DOWN_PREPARE: 6665 cpuset_update_active_cpus(false); 6666 break; 6667 case CPU_DOWN_PREPARE_FROZEN: 6668 num_cpus_frozen++; 6669 partition_sched_domains(1, NULL, NULL); 6670 break; 6671 default: 6672 return NOTIFY_DONE; 6673 } 6674 return NOTIFY_OK; 6675} 6676 6677void __init sched_init_smp(void) 6678{ 6679 cpumask_var_t non_isolated_cpus; 6680 6681 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL); 6682 alloc_cpumask_var(&fallback_doms, GFP_KERNEL); 6683 6684 sched_init_numa(); 6685 6686 get_online_cpus(); 6687 mutex_lock(&sched_domains_mutex); 6688 init_sched_domains(cpu_active_mask); 6689 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map); 6690 if (cpumask_empty(non_isolated_cpus)) 6691 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus); 6692 mutex_unlock(&sched_domains_mutex); 6693 put_online_cpus(); 6694 6695 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE); 6696 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE); 6697 6698 /* RT runtime code needs to handle some hotplug events */ 6699 hotcpu_notifier(update_runtime, 0); 6700 6701 init_hrtick(); 6702 6703 /* Move init over to a non-isolated CPU */ 6704 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0) 6705 BUG(); 6706 sched_init_granularity(); 6707 free_cpumask_var(non_isolated_cpus); 6708 6709 init_sched_rt_class(); 6710} 6711#else 6712void __init sched_init_smp(void) 6713{ 6714 sched_init_granularity(); 6715} 6716#endif /* CONFIG_SMP */ 6717 6718const_debug unsigned int sysctl_timer_migration = 1; 6719 6720int in_sched_functions(unsigned long addr) 6721{ 6722 return in_lock_functions(addr) || 6723 (addr >= (unsigned long)__sched_text_start 6724 && addr < (unsigned long)__sched_text_end); 6725} 6726 6727#ifdef CONFIG_CGROUP_SCHED 6728struct task_group root_task_group; 6729LIST_HEAD(task_groups); 6730#endif 6731 6732DECLARE_PER_CPU(cpumask_var_t, load_balance_tmpmask); 6733 6734void __init sched_init(void) 6735{ 6736 int i, j; 6737 unsigned long alloc_size = 0, ptr; 6738 6739#ifdef CONFIG_FAIR_GROUP_SCHED 6740 alloc_size += 2 * nr_cpu_ids * sizeof(void **); 6741#endif 6742#ifdef CONFIG_RT_GROUP_SCHED 6743 alloc_size += 2 * nr_cpu_ids * sizeof(void **); 6744#endif 6745#ifdef CONFIG_CPUMASK_OFFSTACK 6746 alloc_size += num_possible_cpus() * cpumask_size(); 6747#endif 6748 if (alloc_size) { 6749 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT); 6750 6751#ifdef CONFIG_FAIR_GROUP_SCHED 6752 root_task_group.se = (struct sched_entity **)ptr; 6753 ptr += nr_cpu_ids * sizeof(void **); 6754 6755 root_task_group.cfs_rq = (struct cfs_rq **)ptr; 6756 ptr += nr_cpu_ids * sizeof(void **); 6757 6758#endif /* CONFIG_FAIR_GROUP_SCHED */ 6759#ifdef CONFIG_RT_GROUP_SCHED 6760 root_task_group.rt_se = (struct sched_rt_entity **)ptr; 6761 ptr += nr_cpu_ids * sizeof(void **); 6762 6763 root_task_group.rt_rq = (struct rt_rq **)ptr; 6764 ptr += nr_cpu_ids * sizeof(void **); 6765 6766#endif /* CONFIG_RT_GROUP_SCHED */ 6767#ifdef CONFIG_CPUMASK_OFFSTACK 6768 for_each_possible_cpu(i) { 6769 per_cpu(load_balance_tmpmask, i) = (void *)ptr; 6770 ptr += cpumask_size(); 6771 } 6772#endif /* CONFIG_CPUMASK_OFFSTACK */ 6773 } 6774 6775#ifdef CONFIG_SMP 6776 init_defrootdomain(); 6777#endif 6778 6779 init_rt_bandwidth(&def_rt_bandwidth, 6780 global_rt_period(), global_rt_runtime()); 6781 6782#ifdef CONFIG_RT_GROUP_SCHED 6783 init_rt_bandwidth(&root_task_group.rt_bandwidth, 6784 global_rt_period(), global_rt_runtime()); 6785#endif /* CONFIG_RT_GROUP_SCHED */ 6786 6787#ifdef CONFIG_CGROUP_SCHED 6788 list_add(&root_task_group.list, &task_groups); 6789 INIT_LIST_HEAD(&root_task_group.children); 6790 INIT_LIST_HEAD(&root_task_group.siblings); 6791 autogroup_init(&init_task); 6792 6793#endif /* CONFIG_CGROUP_SCHED */ 6794 6795#ifdef CONFIG_CGROUP_CPUACCT 6796 root_cpuacct.cpustat = &kernel_cpustat; 6797 root_cpuacct.cpuusage = alloc_percpu(u64); 6798 /* Too early, not expected to fail */ 6799 BUG_ON(!root_cpuacct.cpuusage); 6800#endif 6801 for_each_possible_cpu(i) { 6802 struct rq *rq; 6803 6804 rq = cpu_rq(i); 6805 raw_spin_lock_init(&rq->lock); 6806 rq->nr_running = 0; 6807 rq->calc_load_active = 0; 6808 rq->calc_load_update = jiffies + LOAD_FREQ; 6809 init_cfs_rq(&rq->cfs); 6810 init_rt_rq(&rq->rt, rq); 6811#ifdef CONFIG_FAIR_GROUP_SCHED 6812 root_task_group.shares = ROOT_TASK_GROUP_LOAD; 6813 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list); 6814 /* 6815 * How much cpu bandwidth does root_task_group get? 6816 * 6817 * In case of task-groups formed thr' the cgroup filesystem, it 6818 * gets 100% of the cpu resources in the system. This overall 6819 * system cpu resource is divided among the tasks of 6820 * root_task_group and its child task-groups in a fair manner, 6821 * based on each entity's (task or task-group's) weight 6822 * (se->load.weight). 6823 * 6824 * In other words, if root_task_group has 10 tasks of weight 6825 * 1024) and two child groups A0 and A1 (of weight 1024 each), 6826 * then A0's share of the cpu resource is: 6827 * 6828 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33% 6829 * 6830 * We achieve this by letting root_task_group's tasks sit 6831 * directly in rq->cfs (i.e root_task_group->se[] = NULL). 6832 */ 6833 init_cfs_bandwidth(&root_task_group.cfs_bandwidth); 6834 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL); 6835#endif /* CONFIG_FAIR_GROUP_SCHED */ 6836 6837 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime; 6838#ifdef CONFIG_RT_GROUP_SCHED 6839 INIT_LIST_HEAD(&rq->leaf_rt_rq_list); 6840 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL); 6841#endif 6842 6843 for (j = 0; j < CPU_LOAD_IDX_MAX; j++) 6844 rq->cpu_load[j] = 0; 6845 6846 rq->last_load_update_tick = jiffies; 6847 6848#ifdef CONFIG_SMP 6849 rq->sd = NULL; 6850 rq->rd = NULL; 6851 rq->cpu_power = SCHED_POWER_SCALE; 6852 rq->post_schedule = 0; 6853 rq->active_balance = 0; 6854 rq->next_balance = jiffies; 6855 rq->push_cpu = 0; 6856 rq->cpu = i; 6857 rq->online = 0; 6858 rq->idle_stamp = 0; 6859 rq->avg_idle = 2*sysctl_sched_migration_cost; 6860 6861 INIT_LIST_HEAD(&rq->cfs_tasks); 6862 6863 rq_attach_root(rq, &def_root_domain); 6864#ifdef CONFIG_NO_HZ 6865 rq->nohz_flags = 0; 6866#endif 6867#endif 6868 init_rq_hrtick(rq); 6869 atomic_set(&rq->nr_iowait, 0); 6870 } 6871 6872 set_load_weight(&init_task); 6873 6874#ifdef CONFIG_PREEMPT_NOTIFIERS 6875 INIT_HLIST_HEAD(&init_task.preempt_notifiers); 6876#endif 6877 6878#ifdef CONFIG_RT_MUTEXES 6879 plist_head_init(&init_task.pi_waiters); 6880#endif 6881 6882 /* 6883 * The boot idle thread does lazy MMU switching as well: 6884 */ 6885 atomic_inc(&init_mm.mm_count); 6886 enter_lazy_tlb(&init_mm, current); 6887 6888 /* 6889 * Make us the idle thread. Technically, schedule() should not be 6890 * called from this thread, however somewhere below it might be, 6891 * but because we are the idle thread, we just pick up running again 6892 * when this runqueue becomes "idle". 6893 */ 6894 init_idle(current, smp_processor_id()); 6895 6896 calc_load_update = jiffies + LOAD_FREQ; 6897 6898 /* 6899 * During early bootup we pretend to be a normal task: 6900 */ 6901 current->sched_class = &fair_sched_class; 6902 6903#ifdef CONFIG_SMP 6904 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT); 6905 /* May be allocated at isolcpus cmdline parse time */ 6906 if (cpu_isolated_map == NULL) 6907 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT); 6908 idle_thread_set_boot_cpu(); 6909#endif 6910 init_sched_fair_class(); 6911 6912 scheduler_running = 1; 6913} 6914 6915#ifdef CONFIG_DEBUG_ATOMIC_SLEEP 6916static inline int preempt_count_equals(int preempt_offset) 6917{ 6918 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth(); 6919 6920 return (nested == preempt_offset); 6921} 6922 6923void __might_sleep(const char *file, int line, int preempt_offset) 6924{ 6925 static unsigned long prev_jiffy; /* ratelimiting */ 6926 6927 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */ 6928 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) || 6929 system_state != SYSTEM_RUNNING || oops_in_progress) 6930 return; 6931 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy) 6932 return; 6933 prev_jiffy = jiffies; 6934 6935 printk(KERN_ERR 6936 "BUG: sleeping function called from invalid context at %s:%d\n", 6937 file, line); 6938 printk(KERN_ERR 6939 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n", 6940 in_atomic(), irqs_disabled(), 6941 current->pid, current->comm); 6942 6943 debug_show_held_locks(current); 6944 if (irqs_disabled()) 6945 print_irqtrace_events(current); 6946 dump_stack(); 6947} 6948EXPORT_SYMBOL(__might_sleep); 6949#endif 6950 6951#ifdef CONFIG_MAGIC_SYSRQ 6952static void normalize_task(struct rq *rq, struct task_struct *p) 6953{ 6954 const struct sched_class *prev_class = p->sched_class; 6955 int old_prio = p->prio; 6956 int on_rq; 6957 6958 on_rq = p->on_rq; 6959 if (on_rq) 6960 dequeue_task(rq, p, 0); 6961 __setscheduler(rq, p, SCHED_NORMAL, 0); 6962 if (on_rq) { 6963 enqueue_task(rq, p, 0); 6964 resched_task(rq->curr); 6965 } 6966 6967 check_class_changed(rq, p, prev_class, old_prio); 6968} 6969 6970void normalize_rt_tasks(void) 6971{ 6972 struct task_struct *g, *p; 6973 unsigned long flags; 6974 struct rq *rq; 6975 6976 read_lock_irqsave(&tasklist_lock, flags); 6977 do_each_thread(g, p) { 6978 /* 6979 * Only normalize user tasks: 6980 */ 6981 if (!p->mm) 6982 continue; 6983 6984 p->se.exec_start = 0; 6985#ifdef CONFIG_SCHEDSTATS 6986 p->se.statistics.wait_start = 0; 6987 p->se.statistics.sleep_start = 0; 6988 p->se.statistics.block_start = 0; 6989#endif 6990 6991 if (!rt_task(p)) { 6992 /* 6993 * Renice negative nice level userspace 6994 * tasks back to 0: 6995 */ 6996 if (TASK_NICE(p) < 0 && p->mm) 6997 set_user_nice(p, 0); 6998 continue; 6999 } 7000 7001 raw_spin_lock(&p->pi_lock); 7002 rq = __task_rq_lock(p); 7003 7004 normalize_task(rq, p); 7005 7006 __task_rq_unlock(rq); 7007 raw_spin_unlock(&p->pi_lock); 7008 } while_each_thread(g, p); 7009 7010 read_unlock_irqrestore(&tasklist_lock, flags); 7011} 7012 7013#endif /* CONFIG_MAGIC_SYSRQ */ 7014 7015#if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) 7016/* 7017 * These functions are only useful for the IA64 MCA handling, or kdb. 7018 * 7019 * They can only be called when the whole system has been 7020 * stopped - every CPU needs to be quiescent, and no scheduling 7021 * activity can take place. Using them for anything else would 7022 * be a serious bug, and as a result, they aren't even visible 7023 * under any other configuration. 7024 */ 7025 7026/** 7027 * curr_task - return the current task for a given cpu. 7028 * @cpu: the processor in question. 7029 * 7030 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED! 7031 */ 7032struct task_struct *curr_task(int cpu) 7033{ 7034 return cpu_curr(cpu); 7035} 7036 7037#endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */ 7038 7039#ifdef CONFIG_IA64 7040/** 7041 * set_curr_task - set the current task for a given cpu. 7042 * @cpu: the processor in question. 7043 * @p: the task pointer to set. 7044 * 7045 * Description: This function must only be used when non-maskable interrupts 7046 * are serviced on a separate stack. It allows the architecture to switch the 7047 * notion of the current task on a cpu in a non-blocking manner. This function 7048 * must be called with all CPU's synchronized, and interrupts disabled, the 7049 * and caller must save the original value of the current task (see 7050 * curr_task() above) and restore that value before reenabling interrupts and 7051 * re-starting the system. 7052 * 7053 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED! 7054 */ 7055void set_curr_task(int cpu, struct task_struct *p) 7056{ 7057 cpu_curr(cpu) = p; 7058} 7059 7060#endif 7061 7062#ifdef CONFIG_CGROUP_SCHED 7063/* task_group_lock serializes the addition/removal of task groups */ 7064static DEFINE_SPINLOCK(task_group_lock); 7065 7066static void free_sched_group(struct task_group *tg) 7067{ 7068 free_fair_sched_group(tg); 7069 free_rt_sched_group(tg); 7070 autogroup_free(tg); 7071 kfree(tg); 7072} 7073 7074/* allocate runqueue etc for a new task group */ 7075struct task_group *sched_create_group(struct task_group *parent) 7076{ 7077 struct task_group *tg; 7078 unsigned long flags; 7079 7080 tg = kzalloc(sizeof(*tg), GFP_KERNEL); 7081 if (!tg) 7082 return ERR_PTR(-ENOMEM); 7083 7084 if (!alloc_fair_sched_group(tg, parent)) 7085 goto err; 7086 7087 if (!alloc_rt_sched_group(tg, parent)) 7088 goto err; 7089 7090 spin_lock_irqsave(&task_group_lock, flags); 7091 list_add_rcu(&tg->list, &task_groups); 7092 7093 WARN_ON(!parent); /* root should already exist */ 7094 7095 tg->parent = parent; 7096 INIT_LIST_HEAD(&tg->children); 7097 list_add_rcu(&tg->siblings, &parent->children); 7098 spin_unlock_irqrestore(&task_group_lock, flags); 7099 7100 return tg; 7101 7102err: 7103 free_sched_group(tg); 7104 return ERR_PTR(-ENOMEM); 7105} 7106 7107/* rcu callback to free various structures associated with a task group */ 7108static void free_sched_group_rcu(struct rcu_head *rhp) 7109{ 7110 /* now it should be safe to free those cfs_rqs */ 7111 free_sched_group(container_of(rhp, struct task_group, rcu)); 7112} 7113 7114/* Destroy runqueue etc associated with a task group */ 7115void sched_destroy_group(struct task_group *tg) 7116{ 7117 unsigned long flags; 7118 int i; 7119 7120 /* end participation in shares distribution */ 7121 for_each_possible_cpu(i) 7122 unregister_fair_sched_group(tg, i); 7123 7124 spin_lock_irqsave(&task_group_lock, flags); 7125 list_del_rcu(&tg->list); 7126 list_del_rcu(&tg->siblings); 7127 spin_unlock_irqrestore(&task_group_lock, flags); 7128 7129 /* wait for possible concurrent references to cfs_rqs complete */ 7130 call_rcu(&tg->rcu, free_sched_group_rcu); 7131} 7132 7133/* change task's runqueue when it moves between groups. 7134 * The caller of this function should have put the task in its new group 7135 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to 7136 * reflect its new group. 7137 */ 7138void sched_move_task(struct task_struct *tsk) 7139{ 7140 struct task_group *tg; 7141 int on_rq, running; 7142 unsigned long flags; 7143 struct rq *rq; 7144 7145 rq = task_rq_lock(tsk, &flags); 7146 7147 running = task_current(rq, tsk); 7148 on_rq = tsk->on_rq; 7149 7150 if (on_rq) 7151 dequeue_task(rq, tsk, 0); 7152 if (unlikely(running)) 7153 tsk->sched_class->put_prev_task(rq, tsk); 7154 7155 tg = container_of(task_subsys_state_check(tsk, cpu_cgroup_subsys_id, 7156 lockdep_is_held(&tsk->sighand->siglock)), 7157 struct task_group, css); 7158 tg = autogroup_task_group(tsk, tg); 7159 tsk->sched_task_group = tg; 7160 7161#ifdef CONFIG_FAIR_GROUP_SCHED 7162 if (tsk->sched_class->task_move_group) 7163 tsk->sched_class->task_move_group(tsk, on_rq); 7164 else 7165#endif 7166 set_task_rq(tsk, task_cpu(tsk)); 7167 7168 if (unlikely(running)) 7169 tsk->sched_class->set_curr_task(rq); 7170 if (on_rq) 7171 enqueue_task(rq, tsk, 0); 7172 7173 task_rq_unlock(rq, tsk, &flags); 7174} 7175#endif /* CONFIG_CGROUP_SCHED */ 7176 7177#if defined(CONFIG_RT_GROUP_SCHED) || defined(CONFIG_CFS_BANDWIDTH) 7178static unsigned long to_ratio(u64 period, u64 runtime) 7179{ 7180 if (runtime == RUNTIME_INF) 7181 return 1ULL << 20; 7182 7183 return div64_u64(runtime << 20, period); 7184} 7185#endif 7186 7187#ifdef CONFIG_RT_GROUP_SCHED 7188/* 7189 * Ensure that the real time constraints are schedulable. 7190 */ 7191static DEFINE_MUTEX(rt_constraints_mutex); 7192 7193/* Must be called with tasklist_lock held */ 7194static inline int tg_has_rt_tasks(struct task_group *tg) 7195{ 7196 struct task_struct *g, *p; 7197 7198 do_each_thread(g, p) { 7199 if (rt_task(p) && task_rq(p)->rt.tg == tg) 7200 return 1; 7201 } while_each_thread(g, p); 7202 7203 return 0; 7204} 7205 7206struct rt_schedulable_data { 7207 struct task_group *tg; 7208 u64 rt_period; 7209 u64 rt_runtime; 7210}; 7211 7212static int tg_rt_schedulable(struct task_group *tg, void *data) 7213{ 7214 struct rt_schedulable_data *d = data; 7215 struct task_group *child; 7216 unsigned long total, sum = 0; 7217 u64 period, runtime; 7218 7219 period = ktime_to_ns(tg->rt_bandwidth.rt_period); 7220 runtime = tg->rt_bandwidth.rt_runtime; 7221 7222 if (tg == d->tg) { 7223 period = d->rt_period; 7224 runtime = d->rt_runtime; 7225 } 7226 7227 /* 7228 * Cannot have more runtime than the period. 7229 */ 7230 if (runtime > period && runtime != RUNTIME_INF) 7231 return -EINVAL; 7232 7233 /* 7234 * Ensure we don't starve existing RT tasks. 7235 */ 7236 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg)) 7237 return -EBUSY; 7238 7239 total = to_ratio(period, runtime); 7240 7241 /* 7242 * Nobody can have more than the global setting allows. 7243 */ 7244 if (total > to_ratio(global_rt_period(), global_rt_runtime())) 7245 return -EINVAL; 7246 7247 /* 7248 * The sum of our children's runtime should not exceed our own. 7249 */ 7250 list_for_each_entry_rcu(child, &tg->children, siblings) { 7251 period = ktime_to_ns(child->rt_bandwidth.rt_period); 7252 runtime = child->rt_bandwidth.rt_runtime; 7253 7254 if (child == d->tg) { 7255 period = d->rt_period; 7256 runtime = d->rt_runtime; 7257 } 7258 7259 sum += to_ratio(period, runtime); 7260 } 7261 7262 if (sum > total) 7263 return -EINVAL; 7264 7265 return 0; 7266} 7267 7268static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime) 7269{ 7270 int ret; 7271 7272 struct rt_schedulable_data data = { 7273 .tg = tg, 7274 .rt_period = period, 7275 .rt_runtime = runtime, 7276 }; 7277 7278 rcu_read_lock(); 7279 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data); 7280 rcu_read_unlock(); 7281 7282 return ret; 7283} 7284 7285static int tg_set_rt_bandwidth(struct task_group *tg, 7286 u64 rt_period, u64 rt_runtime) 7287{ 7288 int i, err = 0; 7289 7290 mutex_lock(&rt_constraints_mutex); 7291 read_lock(&tasklist_lock); 7292 err = __rt_schedulable(tg, rt_period, rt_runtime); 7293 if (err) 7294 goto unlock; 7295 7296 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock); 7297 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period); 7298 tg->rt_bandwidth.rt_runtime = rt_runtime; 7299 7300 for_each_possible_cpu(i) { 7301 struct rt_rq *rt_rq = tg->rt_rq[i]; 7302 7303 raw_spin_lock(&rt_rq->rt_runtime_lock); 7304 rt_rq->rt_runtime = rt_runtime; 7305 raw_spin_unlock(&rt_rq->rt_runtime_lock); 7306 } 7307 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock); 7308unlock: 7309 read_unlock(&tasklist_lock); 7310 mutex_unlock(&rt_constraints_mutex); 7311 7312 return err; 7313} 7314 7315int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us) 7316{ 7317 u64 rt_runtime, rt_period; 7318 7319 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period); 7320 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC; 7321 if (rt_runtime_us < 0) 7322 rt_runtime = RUNTIME_INF; 7323 7324 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime); 7325} 7326 7327long sched_group_rt_runtime(struct task_group *tg) 7328{ 7329 u64 rt_runtime_us; 7330 7331 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF) 7332 return -1; 7333 7334 rt_runtime_us = tg->rt_bandwidth.rt_runtime; 7335 do_div(rt_runtime_us, NSEC_PER_USEC); 7336 return rt_runtime_us; 7337} 7338 7339int sched_group_set_rt_period(struct task_group *tg, long rt_period_us) 7340{ 7341 u64 rt_runtime, rt_period; 7342 7343 rt_period = (u64)rt_period_us * NSEC_PER_USEC; 7344 rt_runtime = tg->rt_bandwidth.rt_runtime; 7345 7346 if (rt_period == 0) 7347 return -EINVAL; 7348 7349 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime); 7350} 7351 7352long sched_group_rt_period(struct task_group *tg) 7353{ 7354 u64 rt_period_us; 7355 7356 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period); 7357 do_div(rt_period_us, NSEC_PER_USEC); 7358 return rt_period_us; 7359} 7360 7361static int sched_rt_global_constraints(void) 7362{ 7363 u64 runtime, period; 7364 int ret = 0; 7365 7366 if (sysctl_sched_rt_period <= 0) 7367 return -EINVAL; 7368 7369 runtime = global_rt_runtime(); 7370 period = global_rt_period(); 7371 7372 /* 7373 * Sanity check on the sysctl variables. 7374 */ 7375 if (runtime > period && runtime != RUNTIME_INF) 7376 return -EINVAL; 7377 7378 mutex_lock(&rt_constraints_mutex); 7379 read_lock(&tasklist_lock); 7380 ret = __rt_schedulable(NULL, 0, 0); 7381 read_unlock(&tasklist_lock); 7382 mutex_unlock(&rt_constraints_mutex); 7383 7384 return ret; 7385} 7386 7387int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk) 7388{ 7389 /* Don't accept realtime tasks when there is no way for them to run */ 7390 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0) 7391 return 0; 7392 7393 return 1; 7394} 7395 7396#else /* !CONFIG_RT_GROUP_SCHED */ 7397static int sched_rt_global_constraints(void) 7398{ 7399 unsigned long flags; 7400 int i; 7401 7402 if (sysctl_sched_rt_period <= 0) 7403 return -EINVAL; 7404 7405 /* 7406 * There's always some RT tasks in the root group 7407 * -- migration, kstopmachine etc.. 7408 */ 7409 if (sysctl_sched_rt_runtime == 0) 7410 return -EBUSY; 7411 7412 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags); 7413 for_each_possible_cpu(i) { 7414 struct rt_rq *rt_rq = &cpu_rq(i)->rt; 7415 7416 raw_spin_lock(&rt_rq->rt_runtime_lock); 7417 rt_rq->rt_runtime = global_rt_runtime(); 7418 raw_spin_unlock(&rt_rq->rt_runtime_lock); 7419 } 7420 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags); 7421 7422 return 0; 7423} 7424#endif /* CONFIG_RT_GROUP_SCHED */ 7425 7426int sched_rt_handler(struct ctl_table *table, int write, 7427 void __user *buffer, size_t *lenp, 7428 loff_t *ppos) 7429{ 7430 int ret; 7431 int old_period, old_runtime; 7432 static DEFINE_MUTEX(mutex); 7433 7434 mutex_lock(&mutex); 7435 old_period = sysctl_sched_rt_period; 7436 old_runtime = sysctl_sched_rt_runtime; 7437 7438 ret = proc_dointvec(table, write, buffer, lenp, ppos); 7439 7440 if (!ret && write) { 7441 ret = sched_rt_global_constraints(); 7442 if (ret) { 7443 sysctl_sched_rt_period = old_period; 7444 sysctl_sched_rt_runtime = old_runtime; 7445 } else { 7446 def_rt_bandwidth.rt_runtime = global_rt_runtime(); 7447 def_rt_bandwidth.rt_period = 7448 ns_to_ktime(global_rt_period()); 7449 } 7450 } 7451 mutex_unlock(&mutex); 7452 7453 return ret; 7454} 7455 7456#ifdef CONFIG_CGROUP_SCHED 7457 7458/* return corresponding task_group object of a cgroup */ 7459static inline struct task_group *cgroup_tg(struct cgroup *cgrp) 7460{ 7461 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id), 7462 struct task_group, css); 7463} 7464 7465static struct cgroup_subsys_state *cpu_cgroup_create(struct cgroup *cgrp) 7466{ 7467 struct task_group *tg, *parent; 7468 7469 if (!cgrp->parent) { 7470 /* This is early initialization for the top cgroup */ 7471 return &root_task_group.css; 7472 } 7473 7474 parent = cgroup_tg(cgrp->parent); 7475 tg = sched_create_group(parent); 7476 if (IS_ERR(tg)) 7477 return ERR_PTR(-ENOMEM); 7478 7479 return &tg->css; 7480} 7481 7482static void cpu_cgroup_destroy(struct cgroup *cgrp) 7483{ 7484 struct task_group *tg = cgroup_tg(cgrp); 7485 7486 sched_destroy_group(tg); 7487} 7488 7489static int cpu_cgroup_can_attach(struct cgroup *cgrp, 7490 struct cgroup_taskset *tset) 7491{ 7492 struct task_struct *task; 7493 7494 cgroup_taskset_for_each(task, cgrp, tset) { 7495#ifdef CONFIG_RT_GROUP_SCHED 7496 if (!sched_rt_can_attach(cgroup_tg(cgrp), task)) 7497 return -EINVAL; 7498#else 7499 /* We don't support RT-tasks being in separate groups */ 7500 if (task->sched_class != &fair_sched_class) 7501 return -EINVAL; 7502#endif 7503 } 7504 return 0; 7505} 7506 7507static void cpu_cgroup_attach(struct cgroup *cgrp, 7508 struct cgroup_taskset *tset) 7509{ 7510 struct task_struct *task; 7511 7512 cgroup_taskset_for_each(task, cgrp, tset) 7513 sched_move_task(task); 7514} 7515 7516static void 7517cpu_cgroup_exit(struct cgroup *cgrp, struct cgroup *old_cgrp, 7518 struct task_struct *task) 7519{ 7520 /* 7521 * cgroup_exit() is called in the copy_process() failure path. 7522 * Ignore this case since the task hasn't ran yet, this avoids 7523 * trying to poke a half freed task state from generic code. 7524 */ 7525 if (!(task->flags & PF_EXITING)) 7526 return; 7527 7528 sched_move_task(task); 7529} 7530 7531#ifdef CONFIG_FAIR_GROUP_SCHED 7532static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype, 7533 u64 shareval) 7534{ 7535 return sched_group_set_shares(cgroup_tg(cgrp), scale_load(shareval)); 7536} 7537 7538static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft) 7539{ 7540 struct task_group *tg = cgroup_tg(cgrp); 7541 7542 return (u64) scale_load_down(tg->shares); 7543} 7544 7545#ifdef CONFIG_CFS_BANDWIDTH 7546static DEFINE_MUTEX(cfs_constraints_mutex); 7547 7548const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */ 7549const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */ 7550 7551static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime); 7552 7553static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota) 7554{ 7555 int i, ret = 0, runtime_enabled, runtime_was_enabled; 7556 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth; 7557 7558 if (tg == &root_task_group) 7559 return -EINVAL; 7560 7561 /* 7562 * Ensure we have at some amount of bandwidth every period. This is 7563 * to prevent reaching a state of large arrears when throttled via 7564 * entity_tick() resulting in prolonged exit starvation. 7565 */ 7566 if (quota < min_cfs_quota_period || period < min_cfs_quota_period) 7567 return -EINVAL; 7568 7569 /* 7570 * Likewise, bound things on the otherside by preventing insane quota 7571 * periods. This also allows us to normalize in computing quota 7572 * feasibility. 7573 */ 7574 if (period > max_cfs_quota_period) 7575 return -EINVAL; 7576 7577 mutex_lock(&cfs_constraints_mutex); 7578 ret = __cfs_schedulable(tg, period, quota); 7579 if (ret) 7580 goto out_unlock; 7581 7582 runtime_enabled = quota != RUNTIME_INF; 7583 runtime_was_enabled = cfs_b->quota != RUNTIME_INF; 7584 account_cfs_bandwidth_used(runtime_enabled, runtime_was_enabled); 7585 raw_spin_lock_irq(&cfs_b->lock); 7586 cfs_b->period = ns_to_ktime(period); 7587 cfs_b->quota = quota; 7588 7589 __refill_cfs_bandwidth_runtime(cfs_b); 7590 /* restart the period timer (if active) to handle new period expiry */ 7591 if (runtime_enabled && cfs_b->timer_active) { 7592 /* force a reprogram */ 7593 cfs_b->timer_active = 0; 7594 __start_cfs_bandwidth(cfs_b); 7595 } 7596 raw_spin_unlock_irq(&cfs_b->lock); 7597 7598 for_each_possible_cpu(i) { 7599 struct cfs_rq *cfs_rq = tg->cfs_rq[i]; 7600 struct rq *rq = cfs_rq->rq; 7601 7602 raw_spin_lock_irq(&rq->lock); 7603 cfs_rq->runtime_enabled = runtime_enabled; 7604 cfs_rq->runtime_remaining = 0; 7605 7606 if (cfs_rq->throttled) 7607 unthrottle_cfs_rq(cfs_rq); 7608 raw_spin_unlock_irq(&rq->lock); 7609 } 7610out_unlock: 7611 mutex_unlock(&cfs_constraints_mutex); 7612 7613 return ret; 7614} 7615 7616int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us) 7617{ 7618 u64 quota, period; 7619 7620 period = ktime_to_ns(tg->cfs_bandwidth.period); 7621 if (cfs_quota_us < 0) 7622 quota = RUNTIME_INF; 7623 else 7624 quota = (u64)cfs_quota_us * NSEC_PER_USEC; 7625 7626 return tg_set_cfs_bandwidth(tg, period, quota); 7627} 7628 7629long tg_get_cfs_quota(struct task_group *tg) 7630{ 7631 u64 quota_us; 7632 7633 if (tg->cfs_bandwidth.quota == RUNTIME_INF) 7634 return -1; 7635 7636 quota_us = tg->cfs_bandwidth.quota; 7637 do_div(quota_us, NSEC_PER_USEC); 7638 7639 return quota_us; 7640} 7641 7642int tg_set_cfs_period(struct task_group *tg, long cfs_period_us) 7643{ 7644 u64 quota, period; 7645 7646 period = (u64)cfs_period_us * NSEC_PER_USEC; 7647 quota = tg->cfs_bandwidth.quota; 7648 7649 return tg_set_cfs_bandwidth(tg, period, quota); 7650} 7651 7652long tg_get_cfs_period(struct task_group *tg) 7653{ 7654 u64 cfs_period_us; 7655 7656 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period); 7657 do_div(cfs_period_us, NSEC_PER_USEC); 7658 7659 return cfs_period_us; 7660} 7661 7662static s64 cpu_cfs_quota_read_s64(struct cgroup *cgrp, struct cftype *cft) 7663{ 7664 return tg_get_cfs_quota(cgroup_tg(cgrp)); 7665} 7666 7667static int cpu_cfs_quota_write_s64(struct cgroup *cgrp, struct cftype *cftype, 7668 s64 cfs_quota_us) 7669{ 7670 return tg_set_cfs_quota(cgroup_tg(cgrp), cfs_quota_us); 7671} 7672 7673static u64 cpu_cfs_period_read_u64(struct cgroup *cgrp, struct cftype *cft) 7674{ 7675 return tg_get_cfs_period(cgroup_tg(cgrp)); 7676} 7677 7678static int cpu_cfs_period_write_u64(struct cgroup *cgrp, struct cftype *cftype, 7679 u64 cfs_period_us) 7680{ 7681 return tg_set_cfs_period(cgroup_tg(cgrp), cfs_period_us); 7682} 7683 7684struct cfs_schedulable_data { 7685 struct task_group *tg; 7686 u64 period, quota; 7687}; 7688 7689/* 7690 * normalize group quota/period to be quota/max_period 7691 * note: units are usecs 7692 */ 7693static u64 normalize_cfs_quota(struct task_group *tg, 7694 struct cfs_schedulable_data *d) 7695{ 7696 u64 quota, period; 7697 7698 if (tg == d->tg) { 7699 period = d->period; 7700 quota = d->quota; 7701 } else { 7702 period = tg_get_cfs_period(tg); 7703 quota = tg_get_cfs_quota(tg); 7704 } 7705 7706 /* note: these should typically be equivalent */ 7707 if (quota == RUNTIME_INF || quota == -1) 7708 return RUNTIME_INF; 7709 7710 return to_ratio(period, quota); 7711} 7712 7713static int tg_cfs_schedulable_down(struct task_group *tg, void *data) 7714{ 7715 struct cfs_schedulable_data *d = data; 7716 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth; 7717 s64 quota = 0, parent_quota = -1; 7718 7719 if (!tg->parent) { 7720 quota = RUNTIME_INF; 7721 } else { 7722 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth; 7723 7724 quota = normalize_cfs_quota(tg, d); 7725 parent_quota = parent_b->hierarchal_quota; 7726 7727 /* 7728 * ensure max(child_quota) <= parent_quota, inherit when no 7729 * limit is set 7730 */ 7731 if (quota == RUNTIME_INF) 7732 quota = parent_quota; 7733 else if (parent_quota != RUNTIME_INF && quota > parent_quota) 7734 return -EINVAL; 7735 } 7736 cfs_b->hierarchal_quota = quota; 7737 7738 return 0; 7739} 7740 7741static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota) 7742{ 7743 int ret; 7744 struct cfs_schedulable_data data = { 7745 .tg = tg, 7746 .period = period, 7747 .quota = quota, 7748 }; 7749 7750 if (quota != RUNTIME_INF) { 7751 do_div(data.period, NSEC_PER_USEC); 7752 do_div(data.quota, NSEC_PER_USEC); 7753 } 7754 7755 rcu_read_lock(); 7756 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data); 7757 rcu_read_unlock(); 7758 7759 return ret; 7760} 7761 7762static int cpu_stats_show(struct cgroup *cgrp, struct cftype *cft, 7763 struct cgroup_map_cb *cb) 7764{ 7765 struct task_group *tg = cgroup_tg(cgrp); 7766 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth; 7767 7768 cb->fill(cb, "nr_periods", cfs_b->nr_periods); 7769 cb->fill(cb, "nr_throttled", cfs_b->nr_throttled); 7770 cb->fill(cb, "throttled_time", cfs_b->throttled_time); 7771 7772 return 0; 7773} 7774#endif /* CONFIG_CFS_BANDWIDTH */ 7775#endif /* CONFIG_FAIR_GROUP_SCHED */ 7776 7777#ifdef CONFIG_RT_GROUP_SCHED 7778static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft, 7779 s64 val) 7780{ 7781 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val); 7782} 7783 7784static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft) 7785{ 7786 return sched_group_rt_runtime(cgroup_tg(cgrp)); 7787} 7788 7789static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype, 7790 u64 rt_period_us) 7791{ 7792 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us); 7793} 7794 7795static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft) 7796{ 7797 return sched_group_rt_period(cgroup_tg(cgrp)); 7798} 7799#endif /* CONFIG_RT_GROUP_SCHED */ 7800 7801static struct cftype cpu_files[] = { 7802#ifdef CONFIG_FAIR_GROUP_SCHED 7803 { 7804 .name = "shares", 7805 .read_u64 = cpu_shares_read_u64, 7806 .write_u64 = cpu_shares_write_u64, 7807 }, 7808#endif 7809#ifdef CONFIG_CFS_BANDWIDTH 7810 { 7811 .name = "cfs_quota_us", 7812 .read_s64 = cpu_cfs_quota_read_s64, 7813 .write_s64 = cpu_cfs_quota_write_s64, 7814 }, 7815 { 7816 .name = "cfs_period_us", 7817 .read_u64 = cpu_cfs_period_read_u64, 7818 .write_u64 = cpu_cfs_period_write_u64, 7819 }, 7820 { 7821 .name = "stat", 7822 .read_map = cpu_stats_show, 7823 }, 7824#endif 7825#ifdef CONFIG_RT_GROUP_SCHED 7826 { 7827 .name = "rt_runtime_us", 7828 .read_s64 = cpu_rt_runtime_read, 7829 .write_s64 = cpu_rt_runtime_write, 7830 }, 7831 { 7832 .name = "rt_period_us", 7833 .read_u64 = cpu_rt_period_read_uint, 7834 .write_u64 = cpu_rt_period_write_uint, 7835 }, 7836#endif 7837 { } /* terminate */ 7838}; 7839 7840struct cgroup_subsys cpu_cgroup_subsys = { 7841 .name = "cpu", 7842 .create = cpu_cgroup_create, 7843 .destroy = cpu_cgroup_destroy, 7844 .can_attach = cpu_cgroup_can_attach, 7845 .attach = cpu_cgroup_attach, 7846 .exit = cpu_cgroup_exit, 7847 .subsys_id = cpu_cgroup_subsys_id, 7848 .base_cftypes = cpu_files, 7849 .early_init = 1, 7850}; 7851 7852#endif /* CONFIG_CGROUP_SCHED */ 7853 7854#ifdef CONFIG_CGROUP_CPUACCT 7855 7856/* 7857 * CPU accounting code for task groups. 7858 * 7859 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh 7860 * (balbir@in.ibm.com). 7861 */ 7862 7863struct cpuacct root_cpuacct; 7864 7865/* create a new cpu accounting group */ 7866static struct cgroup_subsys_state *cpuacct_create(struct cgroup *cgrp) 7867{ 7868 struct cpuacct *ca; 7869 7870 if (!cgrp->parent) 7871 return &root_cpuacct.css; 7872 7873 ca = kzalloc(sizeof(*ca), GFP_KERNEL); 7874 if (!ca) 7875 goto out; 7876 7877 ca->cpuusage = alloc_percpu(u64); 7878 if (!ca->cpuusage) 7879 goto out_free_ca; 7880 7881 ca->cpustat = alloc_percpu(struct kernel_cpustat); 7882 if (!ca->cpustat) 7883 goto out_free_cpuusage; 7884 7885 return &ca->css; 7886 7887out_free_cpuusage: 7888 free_percpu(ca->cpuusage); 7889out_free_ca: 7890 kfree(ca); 7891out: 7892 return ERR_PTR(-ENOMEM); 7893} 7894 7895/* destroy an existing cpu accounting group */ 7896static void cpuacct_destroy(struct cgroup *cgrp) 7897{ 7898 struct cpuacct *ca = cgroup_ca(cgrp); 7899 7900 free_percpu(ca->cpustat); 7901 free_percpu(ca->cpuusage); 7902 kfree(ca); 7903} 7904 7905static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu) 7906{ 7907 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu); 7908 u64 data; 7909 7910#ifndef CONFIG_64BIT 7911 /* 7912 * Take rq->lock to make 64-bit read safe on 32-bit platforms. 7913 */ 7914 raw_spin_lock_irq(&cpu_rq(cpu)->lock); 7915 data = *cpuusage; 7916 raw_spin_unlock_irq(&cpu_rq(cpu)->lock); 7917#else 7918 data = *cpuusage; 7919#endif 7920 7921 return data; 7922} 7923 7924static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val) 7925{ 7926 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu); 7927 7928#ifndef CONFIG_64BIT 7929 /* 7930 * Take rq->lock to make 64-bit write safe on 32-bit platforms. 7931 */ 7932 raw_spin_lock_irq(&cpu_rq(cpu)->lock); 7933 *cpuusage = val; 7934 raw_spin_unlock_irq(&cpu_rq(cpu)->lock); 7935#else 7936 *cpuusage = val; 7937#endif 7938} 7939 7940/* return total cpu usage (in nanoseconds) of a group */ 7941static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft) 7942{ 7943 struct cpuacct *ca = cgroup_ca(cgrp); 7944 u64 totalcpuusage = 0; 7945 int i; 7946 7947 for_each_present_cpu(i) 7948 totalcpuusage += cpuacct_cpuusage_read(ca, i); 7949 7950 return totalcpuusage; 7951} 7952 7953static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype, 7954 u64 reset) 7955{ 7956 struct cpuacct *ca = cgroup_ca(cgrp); 7957 int err = 0; 7958 int i; 7959 7960 if (reset) { 7961 err = -EINVAL; 7962 goto out; 7963 } 7964 7965 for_each_present_cpu(i) 7966 cpuacct_cpuusage_write(ca, i, 0); 7967 7968out: 7969 return err; 7970} 7971 7972static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft, 7973 struct seq_file *m) 7974{ 7975 struct cpuacct *ca = cgroup_ca(cgroup); 7976 u64 percpu; 7977 int i; 7978 7979 for_each_present_cpu(i) { 7980 percpu = cpuacct_cpuusage_read(ca, i); 7981 seq_printf(m, "%llu ", (unsigned long long) percpu); 7982 } 7983 seq_printf(m, "\n"); 7984 return 0; 7985} 7986 7987static const char *cpuacct_stat_desc[] = { 7988 [CPUACCT_STAT_USER] = "user", 7989 [CPUACCT_STAT_SYSTEM] = "system", 7990}; 7991 7992static int cpuacct_stats_show(struct cgroup *cgrp, struct cftype *cft, 7993 struct cgroup_map_cb *cb) 7994{ 7995 struct cpuacct *ca = cgroup_ca(cgrp); 7996 int cpu; 7997 s64 val = 0; 7998 7999 for_each_online_cpu(cpu) { 8000 struct kernel_cpustat *kcpustat = per_cpu_ptr(ca->cpustat, cpu); 8001 val += kcpustat->cpustat[CPUTIME_USER]; 8002 val += kcpustat->cpustat[CPUTIME_NICE]; 8003 } 8004 val = cputime64_to_clock_t(val); 8005 cb->fill(cb, cpuacct_stat_desc[CPUACCT_STAT_USER], val); 8006 8007 val = 0; 8008 for_each_online_cpu(cpu) { 8009 struct kernel_cpustat *kcpustat = per_cpu_ptr(ca->cpustat, cpu); 8010 val += kcpustat->cpustat[CPUTIME_SYSTEM]; 8011 val += kcpustat->cpustat[CPUTIME_IRQ]; 8012 val += kcpustat->cpustat[CPUTIME_SOFTIRQ]; 8013 } 8014 8015 val = cputime64_to_clock_t(val); 8016 cb->fill(cb, cpuacct_stat_desc[CPUACCT_STAT_SYSTEM], val); 8017 8018 return 0; 8019} 8020 8021static struct cftype files[] = { 8022 { 8023 .name = "usage", 8024 .read_u64 = cpuusage_read, 8025 .write_u64 = cpuusage_write, 8026 }, 8027 { 8028 .name = "usage_percpu", 8029 .read_seq_string = cpuacct_percpu_seq_read, 8030 }, 8031 { 8032 .name = "stat", 8033 .read_map = cpuacct_stats_show, 8034 }, 8035 { } /* terminate */ 8036}; 8037 8038/* 8039 * charge this task's execution time to its accounting group. 8040 * 8041 * called with rq->lock held. 8042 */ 8043void cpuacct_charge(struct task_struct *tsk, u64 cputime) 8044{ 8045 struct cpuacct *ca; 8046 int cpu; 8047 8048 if (unlikely(!cpuacct_subsys.active)) 8049 return; 8050 8051 cpu = task_cpu(tsk); 8052 8053 rcu_read_lock(); 8054 8055 ca = task_ca(tsk); 8056 8057 for (; ca; ca = parent_ca(ca)) { 8058 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu); 8059 *cpuusage += cputime; 8060 } 8061 8062 rcu_read_unlock(); 8063} 8064 8065struct cgroup_subsys cpuacct_subsys = { 8066 .name = "cpuacct", 8067 .create = cpuacct_create, 8068 .destroy = cpuacct_destroy, 8069 .subsys_id = cpuacct_subsys_id, 8070 .base_cftypes = files, 8071}; 8072#endif /* CONFIG_CGROUP_CPUACCT */ 8073