core.c revision 28b4a521f618d9722bc780ea38b44718ce0fe283
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#include <linux/context_tracking.h> 76 77#include <asm/switch_to.h> 78#include <asm/tlb.h> 79#include <asm/irq_regs.h> 80#include <asm/mutex.h> 81#ifdef CONFIG_PARAVIRT 82#include <asm/paravirt.h> 83#endif 84 85#include "sched.h" 86#include "../workqueue_internal.h" 87#include "../smpboot.h" 88 89#define CREATE_TRACE_POINTS 90#include <trace/events/sched.h> 91 92void start_bandwidth_timer(struct hrtimer *period_timer, ktime_t period) 93{ 94 unsigned long delta; 95 ktime_t soft, hard, now; 96 97 for (;;) { 98 if (hrtimer_active(period_timer)) 99 break; 100 101 now = hrtimer_cb_get_time(period_timer); 102 hrtimer_forward(period_timer, now, period); 103 104 soft = hrtimer_get_softexpires(period_timer); 105 hard = hrtimer_get_expires(period_timer); 106 delta = ktime_to_ns(ktime_sub(hard, soft)); 107 __hrtimer_start_range_ns(period_timer, soft, delta, 108 HRTIMER_MODE_ABS_PINNED, 0); 109 } 110} 111 112DEFINE_MUTEX(sched_domains_mutex); 113DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues); 114 115static void update_rq_clock_task(struct rq *rq, s64 delta); 116 117void update_rq_clock(struct rq *rq) 118{ 119 s64 delta; 120 121 if (rq->skip_clock_update > 0) 122 return; 123 124 delta = sched_clock_cpu(cpu_of(rq)) - rq->clock; 125 rq->clock += delta; 126 update_rq_clock_task(rq, delta); 127} 128 129/* 130 * Debugging: various feature bits 131 */ 132 133#define SCHED_FEAT(name, enabled) \ 134 (1UL << __SCHED_FEAT_##name) * enabled | 135 136const_debug unsigned int sysctl_sched_features = 137#include "features.h" 138 0; 139 140#undef SCHED_FEAT 141 142#ifdef CONFIG_SCHED_DEBUG 143#define SCHED_FEAT(name, enabled) \ 144 #name , 145 146static const char * const sched_feat_names[] = { 147#include "features.h" 148}; 149 150#undef SCHED_FEAT 151 152static int sched_feat_show(struct seq_file *m, void *v) 153{ 154 int i; 155 156 for (i = 0; i < __SCHED_FEAT_NR; i++) { 157 if (!(sysctl_sched_features & (1UL << i))) 158 seq_puts(m, "NO_"); 159 seq_printf(m, "%s ", sched_feat_names[i]); 160 } 161 seq_puts(m, "\n"); 162 163 return 0; 164} 165 166#ifdef HAVE_JUMP_LABEL 167 168#define jump_label_key__true STATIC_KEY_INIT_TRUE 169#define jump_label_key__false STATIC_KEY_INIT_FALSE 170 171#define SCHED_FEAT(name, enabled) \ 172 jump_label_key__##enabled , 173 174struct static_key sched_feat_keys[__SCHED_FEAT_NR] = { 175#include "features.h" 176}; 177 178#undef SCHED_FEAT 179 180static void sched_feat_disable(int i) 181{ 182 if (static_key_enabled(&sched_feat_keys[i])) 183 static_key_slow_dec(&sched_feat_keys[i]); 184} 185 186static void sched_feat_enable(int i) 187{ 188 if (!static_key_enabled(&sched_feat_keys[i])) 189 static_key_slow_inc(&sched_feat_keys[i]); 190} 191#else 192static void sched_feat_disable(int i) { }; 193static void sched_feat_enable(int i) { }; 194#endif /* HAVE_JUMP_LABEL */ 195 196static int sched_feat_set(char *cmp) 197{ 198 int i; 199 int neg = 0; 200 201 if (strncmp(cmp, "NO_", 3) == 0) { 202 neg = 1; 203 cmp += 3; 204 } 205 206 for (i = 0; i < __SCHED_FEAT_NR; i++) { 207 if (strcmp(cmp, sched_feat_names[i]) == 0) { 208 if (neg) { 209 sysctl_sched_features &= ~(1UL << i); 210 sched_feat_disable(i); 211 } else { 212 sysctl_sched_features |= (1UL << i); 213 sched_feat_enable(i); 214 } 215 break; 216 } 217 } 218 219 return i; 220} 221 222static ssize_t 223sched_feat_write(struct file *filp, const char __user *ubuf, 224 size_t cnt, loff_t *ppos) 225{ 226 char buf[64]; 227 char *cmp; 228 int i; 229 230 if (cnt > 63) 231 cnt = 63; 232 233 if (copy_from_user(&buf, ubuf, cnt)) 234 return -EFAULT; 235 236 buf[cnt] = 0; 237 cmp = strstrip(buf); 238 239 i = sched_feat_set(cmp); 240 if (i == __SCHED_FEAT_NR) 241 return -EINVAL; 242 243 *ppos += cnt; 244 245 return cnt; 246} 247 248static int sched_feat_open(struct inode *inode, struct file *filp) 249{ 250 return single_open(filp, sched_feat_show, NULL); 251} 252 253static const struct file_operations sched_feat_fops = { 254 .open = sched_feat_open, 255 .write = sched_feat_write, 256 .read = seq_read, 257 .llseek = seq_lseek, 258 .release = single_release, 259}; 260 261static __init int sched_init_debug(void) 262{ 263 debugfs_create_file("sched_features", 0644, NULL, NULL, 264 &sched_feat_fops); 265 266 return 0; 267} 268late_initcall(sched_init_debug); 269#endif /* CONFIG_SCHED_DEBUG */ 270 271/* 272 * Number of tasks to iterate in a single balance run. 273 * Limited because this is done with IRQs disabled. 274 */ 275const_debug unsigned int sysctl_sched_nr_migrate = 32; 276 277/* 278 * period over which we average the RT time consumption, measured 279 * in ms. 280 * 281 * default: 1s 282 */ 283const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC; 284 285/* 286 * period over which we measure -rt task cpu usage in us. 287 * default: 1s 288 */ 289unsigned int sysctl_sched_rt_period = 1000000; 290 291__read_mostly int scheduler_running; 292 293/* 294 * part of the period that we allow rt tasks to run in us. 295 * default: 0.95s 296 */ 297int sysctl_sched_rt_runtime = 950000; 298 299 300 301/* 302 * __task_rq_lock - lock the rq @p resides on. 303 */ 304static inline struct rq *__task_rq_lock(struct task_struct *p) 305 __acquires(rq->lock) 306{ 307 struct rq *rq; 308 309 lockdep_assert_held(&p->pi_lock); 310 311 for (;;) { 312 rq = task_rq(p); 313 raw_spin_lock(&rq->lock); 314 if (likely(rq == task_rq(p))) 315 return rq; 316 raw_spin_unlock(&rq->lock); 317 } 318} 319 320/* 321 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on. 322 */ 323static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags) 324 __acquires(p->pi_lock) 325 __acquires(rq->lock) 326{ 327 struct rq *rq; 328 329 for (;;) { 330 raw_spin_lock_irqsave(&p->pi_lock, *flags); 331 rq = task_rq(p); 332 raw_spin_lock(&rq->lock); 333 if (likely(rq == task_rq(p))) 334 return rq; 335 raw_spin_unlock(&rq->lock); 336 raw_spin_unlock_irqrestore(&p->pi_lock, *flags); 337 } 338} 339 340static void __task_rq_unlock(struct rq *rq) 341 __releases(rq->lock) 342{ 343 raw_spin_unlock(&rq->lock); 344} 345 346static inline void 347task_rq_unlock(struct rq *rq, struct task_struct *p, unsigned long *flags) 348 __releases(rq->lock) 349 __releases(p->pi_lock) 350{ 351 raw_spin_unlock(&rq->lock); 352 raw_spin_unlock_irqrestore(&p->pi_lock, *flags); 353} 354 355/* 356 * this_rq_lock - lock this runqueue and disable interrupts. 357 */ 358static struct rq *this_rq_lock(void) 359 __acquires(rq->lock) 360{ 361 struct rq *rq; 362 363 local_irq_disable(); 364 rq = this_rq(); 365 raw_spin_lock(&rq->lock); 366 367 return rq; 368} 369 370#ifdef CONFIG_SCHED_HRTICK 371/* 372 * Use HR-timers to deliver accurate preemption points. 373 * 374 * Its all a bit involved since we cannot program an hrt while holding the 375 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a 376 * reschedule event. 377 * 378 * When we get rescheduled we reprogram the hrtick_timer outside of the 379 * rq->lock. 380 */ 381 382static void hrtick_clear(struct rq *rq) 383{ 384 if (hrtimer_active(&rq->hrtick_timer)) 385 hrtimer_cancel(&rq->hrtick_timer); 386} 387 388/* 389 * High-resolution timer tick. 390 * Runs from hardirq context with interrupts disabled. 391 */ 392static enum hrtimer_restart hrtick(struct hrtimer *timer) 393{ 394 struct rq *rq = container_of(timer, struct rq, hrtick_timer); 395 396 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id()); 397 398 raw_spin_lock(&rq->lock); 399 update_rq_clock(rq); 400 rq->curr->sched_class->task_tick(rq, rq->curr, 1); 401 raw_spin_unlock(&rq->lock); 402 403 return HRTIMER_NORESTART; 404} 405 406#ifdef CONFIG_SMP 407/* 408 * called from hardirq (IPI) context 409 */ 410static void __hrtick_start(void *arg) 411{ 412 struct rq *rq = arg; 413 414 raw_spin_lock(&rq->lock); 415 hrtimer_restart(&rq->hrtick_timer); 416 rq->hrtick_csd_pending = 0; 417 raw_spin_unlock(&rq->lock); 418} 419 420/* 421 * Called to set the hrtick timer state. 422 * 423 * called with rq->lock held and irqs disabled 424 */ 425void hrtick_start(struct rq *rq, u64 delay) 426{ 427 struct hrtimer *timer = &rq->hrtick_timer; 428 ktime_t time = ktime_add_ns(timer->base->get_time(), delay); 429 430 hrtimer_set_expires(timer, time); 431 432 if (rq == this_rq()) { 433 hrtimer_restart(timer); 434 } else if (!rq->hrtick_csd_pending) { 435 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd, 0); 436 rq->hrtick_csd_pending = 1; 437 } 438} 439 440static int 441hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu) 442{ 443 int cpu = (int)(long)hcpu; 444 445 switch (action) { 446 case CPU_UP_CANCELED: 447 case CPU_UP_CANCELED_FROZEN: 448 case CPU_DOWN_PREPARE: 449 case CPU_DOWN_PREPARE_FROZEN: 450 case CPU_DEAD: 451 case CPU_DEAD_FROZEN: 452 hrtick_clear(cpu_rq(cpu)); 453 return NOTIFY_OK; 454 } 455 456 return NOTIFY_DONE; 457} 458 459static __init void init_hrtick(void) 460{ 461 hotcpu_notifier(hotplug_hrtick, 0); 462} 463#else 464/* 465 * Called to set the hrtick timer state. 466 * 467 * called with rq->lock held and irqs disabled 468 */ 469void hrtick_start(struct rq *rq, u64 delay) 470{ 471 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0, 472 HRTIMER_MODE_REL_PINNED, 0); 473} 474 475static inline void init_hrtick(void) 476{ 477} 478#endif /* CONFIG_SMP */ 479 480static void init_rq_hrtick(struct rq *rq) 481{ 482#ifdef CONFIG_SMP 483 rq->hrtick_csd_pending = 0; 484 485 rq->hrtick_csd.flags = 0; 486 rq->hrtick_csd.func = __hrtick_start; 487 rq->hrtick_csd.info = rq; 488#endif 489 490 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); 491 rq->hrtick_timer.function = hrtick; 492} 493#else /* CONFIG_SCHED_HRTICK */ 494static inline void hrtick_clear(struct rq *rq) 495{ 496} 497 498static inline void init_rq_hrtick(struct rq *rq) 499{ 500} 501 502static inline void init_hrtick(void) 503{ 504} 505#endif /* CONFIG_SCHED_HRTICK */ 506 507/* 508 * resched_task - mark a task 'to be rescheduled now'. 509 * 510 * On UP this means the setting of the need_resched flag, on SMP it 511 * might also involve a cross-CPU call to trigger the scheduler on 512 * the target CPU. 513 */ 514#ifdef CONFIG_SMP 515 516#ifndef tsk_is_polling 517#define tsk_is_polling(t) 0 518#endif 519 520void resched_task(struct task_struct *p) 521{ 522 int cpu; 523 524 assert_raw_spin_locked(&task_rq(p)->lock); 525 526 if (test_tsk_need_resched(p)) 527 return; 528 529 set_tsk_need_resched(p); 530 531 cpu = task_cpu(p); 532 if (cpu == smp_processor_id()) 533 return; 534 535 /* NEED_RESCHED must be visible before we test polling */ 536 smp_mb(); 537 if (!tsk_is_polling(p)) 538 smp_send_reschedule(cpu); 539} 540 541void resched_cpu(int cpu) 542{ 543 struct rq *rq = cpu_rq(cpu); 544 unsigned long flags; 545 546 if (!raw_spin_trylock_irqsave(&rq->lock, flags)) 547 return; 548 resched_task(cpu_curr(cpu)); 549 raw_spin_unlock_irqrestore(&rq->lock, flags); 550} 551 552#ifdef CONFIG_NO_HZ 553/* 554 * In the semi idle case, use the nearest busy cpu for migrating timers 555 * from an idle cpu. This is good for power-savings. 556 * 557 * We don't do similar optimization for completely idle system, as 558 * selecting an idle cpu will add more delays to the timers than intended 559 * (as that cpu's timer base may not be uptodate wrt jiffies etc). 560 */ 561int get_nohz_timer_target(void) 562{ 563 int cpu = smp_processor_id(); 564 int i; 565 struct sched_domain *sd; 566 567 rcu_read_lock(); 568 for_each_domain(cpu, sd) { 569 for_each_cpu(i, sched_domain_span(sd)) { 570 if (!idle_cpu(i)) { 571 cpu = i; 572 goto unlock; 573 } 574 } 575 } 576unlock: 577 rcu_read_unlock(); 578 return cpu; 579} 580/* 581 * When add_timer_on() enqueues a timer into the timer wheel of an 582 * idle CPU then this timer might expire before the next timer event 583 * which is scheduled to wake up that CPU. In case of a completely 584 * idle system the next event might even be infinite time into the 585 * future. wake_up_idle_cpu() ensures that the CPU is woken up and 586 * leaves the inner idle loop so the newly added timer is taken into 587 * account when the CPU goes back to idle and evaluates the timer 588 * wheel for the next timer event. 589 */ 590void wake_up_idle_cpu(int cpu) 591{ 592 struct rq *rq = cpu_rq(cpu); 593 594 if (cpu == smp_processor_id()) 595 return; 596 597 /* 598 * This is safe, as this function is called with the timer 599 * wheel base lock of (cpu) held. When the CPU is on the way 600 * to idle and has not yet set rq->curr to idle then it will 601 * be serialized on the timer wheel base lock and take the new 602 * timer into account automatically. 603 */ 604 if (rq->curr != rq->idle) 605 return; 606 607 /* 608 * We can set TIF_RESCHED on the idle task of the other CPU 609 * lockless. The worst case is that the other CPU runs the 610 * idle task through an additional NOOP schedule() 611 */ 612 set_tsk_need_resched(rq->idle); 613 614 /* NEED_RESCHED must be visible before we test polling */ 615 smp_mb(); 616 if (!tsk_is_polling(rq->idle)) 617 smp_send_reschedule(cpu); 618} 619 620static inline bool got_nohz_idle_kick(void) 621{ 622 int cpu = smp_processor_id(); 623 return idle_cpu(cpu) && test_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu)); 624} 625 626#else /* CONFIG_NO_HZ */ 627 628static inline bool got_nohz_idle_kick(void) 629{ 630 return false; 631} 632 633#endif /* CONFIG_NO_HZ */ 634 635void sched_avg_update(struct rq *rq) 636{ 637 s64 period = sched_avg_period(); 638 639 while ((s64)(rq->clock - rq->age_stamp) > period) { 640 /* 641 * Inline assembly required to prevent the compiler 642 * optimising this loop into a divmod call. 643 * See __iter_div_u64_rem() for another example of this. 644 */ 645 asm("" : "+rm" (rq->age_stamp)); 646 rq->age_stamp += period; 647 rq->rt_avg /= 2; 648 } 649} 650 651#else /* !CONFIG_SMP */ 652void resched_task(struct task_struct *p) 653{ 654 assert_raw_spin_locked(&task_rq(p)->lock); 655 set_tsk_need_resched(p); 656} 657#endif /* CONFIG_SMP */ 658 659#if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \ 660 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH))) 661/* 662 * Iterate task_group tree rooted at *from, calling @down when first entering a 663 * node and @up when leaving it for the final time. 664 * 665 * Caller must hold rcu_lock or sufficient equivalent. 666 */ 667int walk_tg_tree_from(struct task_group *from, 668 tg_visitor down, tg_visitor up, void *data) 669{ 670 struct task_group *parent, *child; 671 int ret; 672 673 parent = from; 674 675down: 676 ret = (*down)(parent, data); 677 if (ret) 678 goto out; 679 list_for_each_entry_rcu(child, &parent->children, siblings) { 680 parent = child; 681 goto down; 682 683up: 684 continue; 685 } 686 ret = (*up)(parent, data); 687 if (ret || parent == from) 688 goto out; 689 690 child = parent; 691 parent = parent->parent; 692 if (parent) 693 goto up; 694out: 695 return ret; 696} 697 698int tg_nop(struct task_group *tg, void *data) 699{ 700 return 0; 701} 702#endif 703 704static void set_load_weight(struct task_struct *p) 705{ 706 int prio = p->static_prio - MAX_RT_PRIO; 707 struct load_weight *load = &p->se.load; 708 709 /* 710 * SCHED_IDLE tasks get minimal weight: 711 */ 712 if (p->policy == SCHED_IDLE) { 713 load->weight = scale_load(WEIGHT_IDLEPRIO); 714 load->inv_weight = WMULT_IDLEPRIO; 715 return; 716 } 717 718 load->weight = scale_load(prio_to_weight[prio]); 719 load->inv_weight = prio_to_wmult[prio]; 720} 721 722static void enqueue_task(struct rq *rq, struct task_struct *p, int flags) 723{ 724 update_rq_clock(rq); 725 sched_info_queued(p); 726 p->sched_class->enqueue_task(rq, p, flags); 727} 728 729static void dequeue_task(struct rq *rq, struct task_struct *p, int flags) 730{ 731 update_rq_clock(rq); 732 sched_info_dequeued(p); 733 p->sched_class->dequeue_task(rq, p, flags); 734} 735 736void activate_task(struct rq *rq, struct task_struct *p, int flags) 737{ 738 if (task_contributes_to_load(p)) 739 rq->nr_uninterruptible--; 740 741 enqueue_task(rq, p, flags); 742} 743 744void deactivate_task(struct rq *rq, struct task_struct *p, int flags) 745{ 746 if (task_contributes_to_load(p)) 747 rq->nr_uninterruptible++; 748 749 dequeue_task(rq, p, flags); 750} 751 752static void update_rq_clock_task(struct rq *rq, s64 delta) 753{ 754/* 755 * In theory, the compile should just see 0 here, and optimize out the call 756 * to sched_rt_avg_update. But I don't trust it... 757 */ 758#if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING) 759 s64 steal = 0, irq_delta = 0; 760#endif 761#ifdef CONFIG_IRQ_TIME_ACCOUNTING 762 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time; 763 764 /* 765 * Since irq_time is only updated on {soft,}irq_exit, we might run into 766 * this case when a previous update_rq_clock() happened inside a 767 * {soft,}irq region. 768 * 769 * When this happens, we stop ->clock_task and only update the 770 * prev_irq_time stamp to account for the part that fit, so that a next 771 * update will consume the rest. This ensures ->clock_task is 772 * monotonic. 773 * 774 * It does however cause some slight miss-attribution of {soft,}irq 775 * time, a more accurate solution would be to update the irq_time using 776 * the current rq->clock timestamp, except that would require using 777 * atomic ops. 778 */ 779 if (irq_delta > delta) 780 irq_delta = delta; 781 782 rq->prev_irq_time += irq_delta; 783 delta -= irq_delta; 784#endif 785#ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING 786 if (static_key_false((¶virt_steal_rq_enabled))) { 787 u64 st; 788 789 steal = paravirt_steal_clock(cpu_of(rq)); 790 steal -= rq->prev_steal_time_rq; 791 792 if (unlikely(steal > delta)) 793 steal = delta; 794 795 st = steal_ticks(steal); 796 steal = st * TICK_NSEC; 797 798 rq->prev_steal_time_rq += steal; 799 800 delta -= steal; 801 } 802#endif 803 804 rq->clock_task += delta; 805 806#if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING) 807 if ((irq_delta + steal) && sched_feat(NONTASK_POWER)) 808 sched_rt_avg_update(rq, irq_delta + steal); 809#endif 810} 811 812void sched_set_stop_task(int cpu, struct task_struct *stop) 813{ 814 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 }; 815 struct task_struct *old_stop = cpu_rq(cpu)->stop; 816 817 if (stop) { 818 /* 819 * Make it appear like a SCHED_FIFO task, its something 820 * userspace knows about and won't get confused about. 821 * 822 * Also, it will make PI more or less work without too 823 * much confusion -- but then, stop work should not 824 * rely on PI working anyway. 825 */ 826 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m); 827 828 stop->sched_class = &stop_sched_class; 829 } 830 831 cpu_rq(cpu)->stop = stop; 832 833 if (old_stop) { 834 /* 835 * Reset it back to a normal scheduling class so that 836 * it can die in pieces. 837 */ 838 old_stop->sched_class = &rt_sched_class; 839 } 840} 841 842/* 843 * __normal_prio - return the priority that is based on the static prio 844 */ 845static inline int __normal_prio(struct task_struct *p) 846{ 847 return p->static_prio; 848} 849 850/* 851 * Calculate the expected normal priority: i.e. priority 852 * without taking RT-inheritance into account. Might be 853 * boosted by interactivity modifiers. Changes upon fork, 854 * setprio syscalls, and whenever the interactivity 855 * estimator recalculates. 856 */ 857static inline int normal_prio(struct task_struct *p) 858{ 859 int prio; 860 861 if (task_has_rt_policy(p)) 862 prio = MAX_RT_PRIO-1 - p->rt_priority; 863 else 864 prio = __normal_prio(p); 865 return prio; 866} 867 868/* 869 * Calculate the current priority, i.e. the priority 870 * taken into account by the scheduler. This value might 871 * be boosted by RT tasks, or might be boosted by 872 * interactivity modifiers. Will be RT if the task got 873 * RT-boosted. If not then it returns p->normal_prio. 874 */ 875static int effective_prio(struct task_struct *p) 876{ 877 p->normal_prio = normal_prio(p); 878 /* 879 * If we are RT tasks or we were boosted to RT priority, 880 * keep the priority unchanged. Otherwise, update priority 881 * to the normal priority: 882 */ 883 if (!rt_prio(p->prio)) 884 return p->normal_prio; 885 return p->prio; 886} 887 888/** 889 * task_curr - is this task currently executing on a CPU? 890 * @p: the task in question. 891 */ 892inline int task_curr(const struct task_struct *p) 893{ 894 return cpu_curr(task_cpu(p)) == p; 895} 896 897static inline void check_class_changed(struct rq *rq, struct task_struct *p, 898 const struct sched_class *prev_class, 899 int oldprio) 900{ 901 if (prev_class != p->sched_class) { 902 if (prev_class->switched_from) 903 prev_class->switched_from(rq, p); 904 p->sched_class->switched_to(rq, p); 905 } else if (oldprio != p->prio) 906 p->sched_class->prio_changed(rq, p, oldprio); 907} 908 909void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags) 910{ 911 const struct sched_class *class; 912 913 if (p->sched_class == rq->curr->sched_class) { 914 rq->curr->sched_class->check_preempt_curr(rq, p, flags); 915 } else { 916 for_each_class(class) { 917 if (class == rq->curr->sched_class) 918 break; 919 if (class == p->sched_class) { 920 resched_task(rq->curr); 921 break; 922 } 923 } 924 } 925 926 /* 927 * A queue event has occurred, and we're going to schedule. In 928 * this case, we can save a useless back to back clock update. 929 */ 930 if (rq->curr->on_rq && test_tsk_need_resched(rq->curr)) 931 rq->skip_clock_update = 1; 932} 933 934static ATOMIC_NOTIFIER_HEAD(task_migration_notifier); 935 936void register_task_migration_notifier(struct notifier_block *n) 937{ 938 atomic_notifier_chain_register(&task_migration_notifier, n); 939} 940 941#ifdef CONFIG_SMP 942void set_task_cpu(struct task_struct *p, unsigned int new_cpu) 943{ 944#ifdef CONFIG_SCHED_DEBUG 945 /* 946 * We should never call set_task_cpu() on a blocked task, 947 * ttwu() will sort out the placement. 948 */ 949 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING && 950 !(task_thread_info(p)->preempt_count & PREEMPT_ACTIVE)); 951 952#ifdef CONFIG_LOCKDEP 953 /* 954 * The caller should hold either p->pi_lock or rq->lock, when changing 955 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks. 956 * 957 * sched_move_task() holds both and thus holding either pins the cgroup, 958 * see task_group(). 959 * 960 * Furthermore, all task_rq users should acquire both locks, see 961 * task_rq_lock(). 962 */ 963 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) || 964 lockdep_is_held(&task_rq(p)->lock))); 965#endif 966#endif 967 968 trace_sched_migrate_task(p, new_cpu); 969 970 if (task_cpu(p) != new_cpu) { 971 struct task_migration_notifier tmn; 972 973 if (p->sched_class->migrate_task_rq) 974 p->sched_class->migrate_task_rq(p, new_cpu); 975 p->se.nr_migrations++; 976 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, NULL, 0); 977 978 tmn.task = p; 979 tmn.from_cpu = task_cpu(p); 980 tmn.to_cpu = new_cpu; 981 982 atomic_notifier_call_chain(&task_migration_notifier, 0, &tmn); 983 } 984 985 __set_task_cpu(p, new_cpu); 986} 987 988struct migration_arg { 989 struct task_struct *task; 990 int dest_cpu; 991}; 992 993static int migration_cpu_stop(void *data); 994 995/* 996 * wait_task_inactive - wait for a thread to unschedule. 997 * 998 * If @match_state is nonzero, it's the @p->state value just checked and 999 * not expected to change. If it changes, i.e. @p might have woken up, 1000 * then return zero. When we succeed in waiting for @p to be off its CPU, 1001 * we return a positive number (its total switch count). If a second call 1002 * a short while later returns the same number, the caller can be sure that 1003 * @p has remained unscheduled the whole time. 1004 * 1005 * The caller must ensure that the task *will* unschedule sometime soon, 1006 * else this function might spin for a *long* time. This function can't 1007 * be called with interrupts off, or it may introduce deadlock with 1008 * smp_call_function() if an IPI is sent by the same process we are 1009 * waiting to become inactive. 1010 */ 1011unsigned long wait_task_inactive(struct task_struct *p, long match_state) 1012{ 1013 unsigned long flags; 1014 int running, on_rq; 1015 unsigned long ncsw; 1016 struct rq *rq; 1017 1018 for (;;) { 1019 /* 1020 * We do the initial early heuristics without holding 1021 * any task-queue locks at all. We'll only try to get 1022 * the runqueue lock when things look like they will 1023 * work out! 1024 */ 1025 rq = task_rq(p); 1026 1027 /* 1028 * If the task is actively running on another CPU 1029 * still, just relax and busy-wait without holding 1030 * any locks. 1031 * 1032 * NOTE! Since we don't hold any locks, it's not 1033 * even sure that "rq" stays as the right runqueue! 1034 * But we don't care, since "task_running()" will 1035 * return false if the runqueue has changed and p 1036 * is actually now running somewhere else! 1037 */ 1038 while (task_running(rq, p)) { 1039 if (match_state && unlikely(p->state != match_state)) 1040 return 0; 1041 cpu_relax(); 1042 } 1043 1044 /* 1045 * Ok, time to look more closely! We need the rq 1046 * lock now, to be *sure*. If we're wrong, we'll 1047 * just go back and repeat. 1048 */ 1049 rq = task_rq_lock(p, &flags); 1050 trace_sched_wait_task(p); 1051 running = task_running(rq, p); 1052 on_rq = p->on_rq; 1053 ncsw = 0; 1054 if (!match_state || p->state == match_state) 1055 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */ 1056 task_rq_unlock(rq, p, &flags); 1057 1058 /* 1059 * If it changed from the expected state, bail out now. 1060 */ 1061 if (unlikely(!ncsw)) 1062 break; 1063 1064 /* 1065 * Was it really running after all now that we 1066 * checked with the proper locks actually held? 1067 * 1068 * Oops. Go back and try again.. 1069 */ 1070 if (unlikely(running)) { 1071 cpu_relax(); 1072 continue; 1073 } 1074 1075 /* 1076 * It's not enough that it's not actively running, 1077 * it must be off the runqueue _entirely_, and not 1078 * preempted! 1079 * 1080 * So if it was still runnable (but just not actively 1081 * running right now), it's preempted, and we should 1082 * yield - it could be a while. 1083 */ 1084 if (unlikely(on_rq)) { 1085 ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ); 1086 1087 set_current_state(TASK_UNINTERRUPTIBLE); 1088 schedule_hrtimeout(&to, HRTIMER_MODE_REL); 1089 continue; 1090 } 1091 1092 /* 1093 * Ahh, all good. It wasn't running, and it wasn't 1094 * runnable, which means that it will never become 1095 * running in the future either. We're all done! 1096 */ 1097 break; 1098 } 1099 1100 return ncsw; 1101} 1102 1103/*** 1104 * kick_process - kick a running thread to enter/exit the kernel 1105 * @p: the to-be-kicked thread 1106 * 1107 * Cause a process which is running on another CPU to enter 1108 * kernel-mode, without any delay. (to get signals handled.) 1109 * 1110 * NOTE: this function doesn't have to take the runqueue lock, 1111 * because all it wants to ensure is that the remote task enters 1112 * the kernel. If the IPI races and the task has been migrated 1113 * to another CPU then no harm is done and the purpose has been 1114 * achieved as well. 1115 */ 1116void kick_process(struct task_struct *p) 1117{ 1118 int cpu; 1119 1120 preempt_disable(); 1121 cpu = task_cpu(p); 1122 if ((cpu != smp_processor_id()) && task_curr(p)) 1123 smp_send_reschedule(cpu); 1124 preempt_enable(); 1125} 1126EXPORT_SYMBOL_GPL(kick_process); 1127#endif /* CONFIG_SMP */ 1128 1129#ifdef CONFIG_SMP 1130/* 1131 * ->cpus_allowed is protected by both rq->lock and p->pi_lock 1132 */ 1133static int select_fallback_rq(int cpu, struct task_struct *p) 1134{ 1135 int nid = cpu_to_node(cpu); 1136 const struct cpumask *nodemask = NULL; 1137 enum { cpuset, possible, fail } state = cpuset; 1138 int dest_cpu; 1139 1140 /* 1141 * If the node that the cpu is on has been offlined, cpu_to_node() 1142 * will return -1. There is no cpu on the node, and we should 1143 * select the cpu on the other node. 1144 */ 1145 if (nid != -1) { 1146 nodemask = cpumask_of_node(nid); 1147 1148 /* Look for allowed, online CPU in same node. */ 1149 for_each_cpu(dest_cpu, nodemask) { 1150 if (!cpu_online(dest_cpu)) 1151 continue; 1152 if (!cpu_active(dest_cpu)) 1153 continue; 1154 if (cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p))) 1155 return dest_cpu; 1156 } 1157 } 1158 1159 for (;;) { 1160 /* Any allowed, online CPU? */ 1161 for_each_cpu(dest_cpu, tsk_cpus_allowed(p)) { 1162 if (!cpu_online(dest_cpu)) 1163 continue; 1164 if (!cpu_active(dest_cpu)) 1165 continue; 1166 goto out; 1167 } 1168 1169 switch (state) { 1170 case cpuset: 1171 /* No more Mr. Nice Guy. */ 1172 cpuset_cpus_allowed_fallback(p); 1173 state = possible; 1174 break; 1175 1176 case possible: 1177 do_set_cpus_allowed(p, cpu_possible_mask); 1178 state = fail; 1179 break; 1180 1181 case fail: 1182 BUG(); 1183 break; 1184 } 1185 } 1186 1187out: 1188 if (state != cpuset) { 1189 /* 1190 * Don't tell them about moving exiting tasks or 1191 * kernel threads (both mm NULL), since they never 1192 * leave kernel. 1193 */ 1194 if (p->mm && printk_ratelimit()) { 1195 printk_sched("process %d (%s) no longer affine to cpu%d\n", 1196 task_pid_nr(p), p->comm, cpu); 1197 } 1198 } 1199 1200 return dest_cpu; 1201} 1202 1203/* 1204 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable. 1205 */ 1206static inline 1207int select_task_rq(struct task_struct *p, int sd_flags, int wake_flags) 1208{ 1209 int cpu = p->sched_class->select_task_rq(p, sd_flags, wake_flags); 1210 1211 /* 1212 * In order not to call set_task_cpu() on a blocking task we need 1213 * to rely on ttwu() to place the task on a valid ->cpus_allowed 1214 * cpu. 1215 * 1216 * Since this is common to all placement strategies, this lives here. 1217 * 1218 * [ this allows ->select_task() to simply return task_cpu(p) and 1219 * not worry about this generic constraint ] 1220 */ 1221 if (unlikely(!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)) || 1222 !cpu_online(cpu))) 1223 cpu = select_fallback_rq(task_cpu(p), p); 1224 1225 return cpu; 1226} 1227 1228static void update_avg(u64 *avg, u64 sample) 1229{ 1230 s64 diff = sample - *avg; 1231 *avg += diff >> 3; 1232} 1233#endif 1234 1235static void 1236ttwu_stat(struct task_struct *p, int cpu, int wake_flags) 1237{ 1238#ifdef CONFIG_SCHEDSTATS 1239 struct rq *rq = this_rq(); 1240 1241#ifdef CONFIG_SMP 1242 int this_cpu = smp_processor_id(); 1243 1244 if (cpu == this_cpu) { 1245 schedstat_inc(rq, ttwu_local); 1246 schedstat_inc(p, se.statistics.nr_wakeups_local); 1247 } else { 1248 struct sched_domain *sd; 1249 1250 schedstat_inc(p, se.statistics.nr_wakeups_remote); 1251 rcu_read_lock(); 1252 for_each_domain(this_cpu, sd) { 1253 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) { 1254 schedstat_inc(sd, ttwu_wake_remote); 1255 break; 1256 } 1257 } 1258 rcu_read_unlock(); 1259 } 1260 1261 if (wake_flags & WF_MIGRATED) 1262 schedstat_inc(p, se.statistics.nr_wakeups_migrate); 1263 1264#endif /* CONFIG_SMP */ 1265 1266 schedstat_inc(rq, ttwu_count); 1267 schedstat_inc(p, se.statistics.nr_wakeups); 1268 1269 if (wake_flags & WF_SYNC) 1270 schedstat_inc(p, se.statistics.nr_wakeups_sync); 1271 1272#endif /* CONFIG_SCHEDSTATS */ 1273} 1274 1275static void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags) 1276{ 1277 activate_task(rq, p, en_flags); 1278 p->on_rq = 1; 1279 1280 /* if a worker is waking up, notify workqueue */ 1281 if (p->flags & PF_WQ_WORKER) 1282 wq_worker_waking_up(p, cpu_of(rq)); 1283} 1284 1285/* 1286 * Mark the task runnable and perform wakeup-preemption. 1287 */ 1288static void 1289ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags) 1290{ 1291 check_preempt_curr(rq, p, wake_flags); 1292 trace_sched_wakeup(p, true); 1293 1294 p->state = TASK_RUNNING; 1295#ifdef CONFIG_SMP 1296 if (p->sched_class->task_woken) 1297 p->sched_class->task_woken(rq, p); 1298 1299 if (rq->idle_stamp) { 1300 u64 delta = rq->clock - rq->idle_stamp; 1301 u64 max = 2*sysctl_sched_migration_cost; 1302 1303 if (delta > max) 1304 rq->avg_idle = max; 1305 else 1306 update_avg(&rq->avg_idle, delta); 1307 rq->idle_stamp = 0; 1308 } 1309#endif 1310} 1311 1312static void 1313ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags) 1314{ 1315#ifdef CONFIG_SMP 1316 if (p->sched_contributes_to_load) 1317 rq->nr_uninterruptible--; 1318#endif 1319 1320 ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_WAKING); 1321 ttwu_do_wakeup(rq, p, wake_flags); 1322} 1323 1324/* 1325 * Called in case the task @p isn't fully descheduled from its runqueue, 1326 * in this case we must do a remote wakeup. Its a 'light' wakeup though, 1327 * since all we need to do is flip p->state to TASK_RUNNING, since 1328 * the task is still ->on_rq. 1329 */ 1330static int ttwu_remote(struct task_struct *p, int wake_flags) 1331{ 1332 struct rq *rq; 1333 int ret = 0; 1334 1335 rq = __task_rq_lock(p); 1336 if (p->on_rq) { 1337 ttwu_do_wakeup(rq, p, wake_flags); 1338 ret = 1; 1339 } 1340 __task_rq_unlock(rq); 1341 1342 return ret; 1343} 1344 1345#ifdef CONFIG_SMP 1346static void sched_ttwu_pending(void) 1347{ 1348 struct rq *rq = this_rq(); 1349 struct llist_node *llist = llist_del_all(&rq->wake_list); 1350 struct task_struct *p; 1351 1352 raw_spin_lock(&rq->lock); 1353 1354 while (llist) { 1355 p = llist_entry(llist, struct task_struct, wake_entry); 1356 llist = llist_next(llist); 1357 ttwu_do_activate(rq, p, 0); 1358 } 1359 1360 raw_spin_unlock(&rq->lock); 1361} 1362 1363void scheduler_ipi(void) 1364{ 1365 if (llist_empty(&this_rq()->wake_list) && !got_nohz_idle_kick()) 1366 return; 1367 1368 /* 1369 * Not all reschedule IPI handlers call irq_enter/irq_exit, since 1370 * traditionally all their work was done from the interrupt return 1371 * path. Now that we actually do some work, we need to make sure 1372 * we do call them. 1373 * 1374 * Some archs already do call them, luckily irq_enter/exit nest 1375 * properly. 1376 * 1377 * Arguably we should visit all archs and update all handlers, 1378 * however a fair share of IPIs are still resched only so this would 1379 * somewhat pessimize the simple resched case. 1380 */ 1381 irq_enter(); 1382 sched_ttwu_pending(); 1383 1384 /* 1385 * Check if someone kicked us for doing the nohz idle load balance. 1386 */ 1387 if (unlikely(got_nohz_idle_kick() && !need_resched())) { 1388 this_rq()->idle_balance = 1; 1389 raise_softirq_irqoff(SCHED_SOFTIRQ); 1390 } 1391 irq_exit(); 1392} 1393 1394static void ttwu_queue_remote(struct task_struct *p, int cpu) 1395{ 1396 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list)) 1397 smp_send_reschedule(cpu); 1398} 1399 1400bool cpus_share_cache(int this_cpu, int that_cpu) 1401{ 1402 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu); 1403} 1404#endif /* CONFIG_SMP */ 1405 1406static void ttwu_queue(struct task_struct *p, int cpu) 1407{ 1408 struct rq *rq = cpu_rq(cpu); 1409 1410#if defined(CONFIG_SMP) 1411 if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) { 1412 sched_clock_cpu(cpu); /* sync clocks x-cpu */ 1413 ttwu_queue_remote(p, cpu); 1414 return; 1415 } 1416#endif 1417 1418 raw_spin_lock(&rq->lock); 1419 ttwu_do_activate(rq, p, 0); 1420 raw_spin_unlock(&rq->lock); 1421} 1422 1423/** 1424 * try_to_wake_up - wake up a thread 1425 * @p: the thread to be awakened 1426 * @state: the mask of task states that can be woken 1427 * @wake_flags: wake modifier flags (WF_*) 1428 * 1429 * Put it on the run-queue if it's not already there. The "current" 1430 * thread is always on the run-queue (except when the actual 1431 * re-schedule is in progress), and as such you're allowed to do 1432 * the simpler "current->state = TASK_RUNNING" to mark yourself 1433 * runnable without the overhead of this. 1434 * 1435 * Returns %true if @p was woken up, %false if it was already running 1436 * or @state didn't match @p's state. 1437 */ 1438static int 1439try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags) 1440{ 1441 unsigned long flags; 1442 int cpu, success = 0; 1443 1444 smp_wmb(); 1445 raw_spin_lock_irqsave(&p->pi_lock, flags); 1446 if (!(p->state & state)) 1447 goto out; 1448 1449 success = 1; /* we're going to change ->state */ 1450 cpu = task_cpu(p); 1451 1452 if (p->on_rq && ttwu_remote(p, wake_flags)) 1453 goto stat; 1454 1455#ifdef CONFIG_SMP 1456 /* 1457 * If the owning (remote) cpu is still in the middle of schedule() with 1458 * this task as prev, wait until its done referencing the task. 1459 */ 1460 while (p->on_cpu) 1461 cpu_relax(); 1462 /* 1463 * Pairs with the smp_wmb() in finish_lock_switch(). 1464 */ 1465 smp_rmb(); 1466 1467 p->sched_contributes_to_load = !!task_contributes_to_load(p); 1468 p->state = TASK_WAKING; 1469 1470 if (p->sched_class->task_waking) 1471 p->sched_class->task_waking(p); 1472 1473 cpu = select_task_rq(p, SD_BALANCE_WAKE, wake_flags); 1474 if (task_cpu(p) != cpu) { 1475 wake_flags |= WF_MIGRATED; 1476 set_task_cpu(p, cpu); 1477 } 1478#endif /* CONFIG_SMP */ 1479 1480 ttwu_queue(p, cpu); 1481stat: 1482 ttwu_stat(p, cpu, wake_flags); 1483out: 1484 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 1485 1486 return success; 1487} 1488 1489/** 1490 * try_to_wake_up_local - try to wake up a local task with rq lock held 1491 * @p: the thread to be awakened 1492 * 1493 * Put @p on the run-queue if it's not already there. The caller must 1494 * ensure that this_rq() is locked, @p is bound to this_rq() and not 1495 * the current task. 1496 */ 1497static void try_to_wake_up_local(struct task_struct *p) 1498{ 1499 struct rq *rq = task_rq(p); 1500 1501 BUG_ON(rq != this_rq()); 1502 BUG_ON(p == current); 1503 lockdep_assert_held(&rq->lock); 1504 1505 if (!raw_spin_trylock(&p->pi_lock)) { 1506 raw_spin_unlock(&rq->lock); 1507 raw_spin_lock(&p->pi_lock); 1508 raw_spin_lock(&rq->lock); 1509 } 1510 1511 if (!(p->state & TASK_NORMAL)) 1512 goto out; 1513 1514 if (!p->on_rq) 1515 ttwu_activate(rq, p, ENQUEUE_WAKEUP); 1516 1517 ttwu_do_wakeup(rq, p, 0); 1518 ttwu_stat(p, smp_processor_id(), 0); 1519out: 1520 raw_spin_unlock(&p->pi_lock); 1521} 1522 1523/** 1524 * wake_up_process - Wake up a specific process 1525 * @p: The process to be woken up. 1526 * 1527 * Attempt to wake up the nominated process and move it to the set of runnable 1528 * processes. Returns 1 if the process was woken up, 0 if it was already 1529 * running. 1530 * 1531 * It may be assumed that this function implies a write memory barrier before 1532 * changing the task state if and only if any tasks are woken up. 1533 */ 1534int wake_up_process(struct task_struct *p) 1535{ 1536 WARN_ON(task_is_stopped_or_traced(p)); 1537 return try_to_wake_up(p, TASK_NORMAL, 0); 1538} 1539EXPORT_SYMBOL(wake_up_process); 1540 1541int wake_up_state(struct task_struct *p, unsigned int state) 1542{ 1543 return try_to_wake_up(p, state, 0); 1544} 1545 1546/* 1547 * Perform scheduler related setup for a newly forked process p. 1548 * p is forked by current. 1549 * 1550 * __sched_fork() is basic setup used by init_idle() too: 1551 */ 1552static void __sched_fork(struct task_struct *p) 1553{ 1554 p->on_rq = 0; 1555 1556 p->se.on_rq = 0; 1557 p->se.exec_start = 0; 1558 p->se.sum_exec_runtime = 0; 1559 p->se.prev_sum_exec_runtime = 0; 1560 p->se.nr_migrations = 0; 1561 p->se.vruntime = 0; 1562 INIT_LIST_HEAD(&p->se.group_node); 1563 1564/* 1565 * Load-tracking only depends on SMP, FAIR_GROUP_SCHED dependency below may be 1566 * removed when useful for applications beyond shares distribution (e.g. 1567 * load-balance). 1568 */ 1569#if defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED) 1570 p->se.avg.runnable_avg_period = 0; 1571 p->se.avg.runnable_avg_sum = 0; 1572#endif 1573#ifdef CONFIG_SCHEDSTATS 1574 memset(&p->se.statistics, 0, sizeof(p->se.statistics)); 1575#endif 1576 1577 INIT_LIST_HEAD(&p->rt.run_list); 1578 1579#ifdef CONFIG_PREEMPT_NOTIFIERS 1580 INIT_HLIST_HEAD(&p->preempt_notifiers); 1581#endif 1582 1583#ifdef CONFIG_NUMA_BALANCING 1584 if (p->mm && atomic_read(&p->mm->mm_users) == 1) { 1585 p->mm->numa_next_scan = jiffies; 1586 p->mm->numa_next_reset = jiffies; 1587 p->mm->numa_scan_seq = 0; 1588 } 1589 1590 p->node_stamp = 0ULL; 1591 p->numa_scan_seq = p->mm ? p->mm->numa_scan_seq : 0; 1592 p->numa_migrate_seq = p->mm ? p->mm->numa_scan_seq - 1 : 0; 1593 p->numa_scan_period = sysctl_numa_balancing_scan_delay; 1594 p->numa_work.next = &p->numa_work; 1595#endif /* CONFIG_NUMA_BALANCING */ 1596} 1597 1598#ifdef CONFIG_NUMA_BALANCING 1599#ifdef CONFIG_SCHED_DEBUG 1600void set_numabalancing_state(bool enabled) 1601{ 1602 if (enabled) 1603 sched_feat_set("NUMA"); 1604 else 1605 sched_feat_set("NO_NUMA"); 1606} 1607#else 1608__read_mostly bool numabalancing_enabled; 1609 1610void set_numabalancing_state(bool enabled) 1611{ 1612 numabalancing_enabled = enabled; 1613} 1614#endif /* CONFIG_SCHED_DEBUG */ 1615#endif /* CONFIG_NUMA_BALANCING */ 1616 1617/* 1618 * fork()/clone()-time setup: 1619 */ 1620void sched_fork(struct task_struct *p) 1621{ 1622 unsigned long flags; 1623 int cpu = get_cpu(); 1624 1625 __sched_fork(p); 1626 /* 1627 * We mark the process as running here. This guarantees that 1628 * nobody will actually run it, and a signal or other external 1629 * event cannot wake it up and insert it on the runqueue either. 1630 */ 1631 p->state = TASK_RUNNING; 1632 1633 /* 1634 * Make sure we do not leak PI boosting priority to the child. 1635 */ 1636 p->prio = current->normal_prio; 1637 1638 /* 1639 * Revert to default priority/policy on fork if requested. 1640 */ 1641 if (unlikely(p->sched_reset_on_fork)) { 1642 if (task_has_rt_policy(p)) { 1643 p->policy = SCHED_NORMAL; 1644 p->static_prio = NICE_TO_PRIO(0); 1645 p->rt_priority = 0; 1646 } else if (PRIO_TO_NICE(p->static_prio) < 0) 1647 p->static_prio = NICE_TO_PRIO(0); 1648 1649 p->prio = p->normal_prio = __normal_prio(p); 1650 set_load_weight(p); 1651 1652 /* 1653 * We don't need the reset flag anymore after the fork. It has 1654 * fulfilled its duty: 1655 */ 1656 p->sched_reset_on_fork = 0; 1657 } 1658 1659 if (!rt_prio(p->prio)) 1660 p->sched_class = &fair_sched_class; 1661 1662 if (p->sched_class->task_fork) 1663 p->sched_class->task_fork(p); 1664 1665 /* 1666 * The child is not yet in the pid-hash so no cgroup attach races, 1667 * and the cgroup is pinned to this child due to cgroup_fork() 1668 * is ran before sched_fork(). 1669 * 1670 * Silence PROVE_RCU. 1671 */ 1672 raw_spin_lock_irqsave(&p->pi_lock, flags); 1673 set_task_cpu(p, cpu); 1674 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 1675 1676#if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT) 1677 if (likely(sched_info_on())) 1678 memset(&p->sched_info, 0, sizeof(p->sched_info)); 1679#endif 1680#if defined(CONFIG_SMP) 1681 p->on_cpu = 0; 1682#endif 1683#ifdef CONFIG_PREEMPT_COUNT 1684 /* Want to start with kernel preemption disabled. */ 1685 task_thread_info(p)->preempt_count = 1; 1686#endif 1687#ifdef CONFIG_SMP 1688 plist_node_init(&p->pushable_tasks, MAX_PRIO); 1689#endif 1690 1691 put_cpu(); 1692} 1693 1694/* 1695 * wake_up_new_task - wake up a newly created task for the first time. 1696 * 1697 * This function will do some initial scheduler statistics housekeeping 1698 * that must be done for every newly created context, then puts the task 1699 * on the runqueue and wakes it. 1700 */ 1701void wake_up_new_task(struct task_struct *p) 1702{ 1703 unsigned long flags; 1704 struct rq *rq; 1705 1706 raw_spin_lock_irqsave(&p->pi_lock, flags); 1707#ifdef CONFIG_SMP 1708 /* 1709 * Fork balancing, do it here and not earlier because: 1710 * - cpus_allowed can change in the fork path 1711 * - any previously selected cpu might disappear through hotplug 1712 */ 1713 set_task_cpu(p, select_task_rq(p, SD_BALANCE_FORK, 0)); 1714#endif 1715 1716 rq = __task_rq_lock(p); 1717 activate_task(rq, p, 0); 1718 p->on_rq = 1; 1719 trace_sched_wakeup_new(p, true); 1720 check_preempt_curr(rq, p, WF_FORK); 1721#ifdef CONFIG_SMP 1722 if (p->sched_class->task_woken) 1723 p->sched_class->task_woken(rq, p); 1724#endif 1725 task_rq_unlock(rq, p, &flags); 1726} 1727 1728#ifdef CONFIG_PREEMPT_NOTIFIERS 1729 1730/** 1731 * preempt_notifier_register - tell me when current is being preempted & rescheduled 1732 * @notifier: notifier struct to register 1733 */ 1734void preempt_notifier_register(struct preempt_notifier *notifier) 1735{ 1736 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers); 1737} 1738EXPORT_SYMBOL_GPL(preempt_notifier_register); 1739 1740/** 1741 * preempt_notifier_unregister - no longer interested in preemption notifications 1742 * @notifier: notifier struct to unregister 1743 * 1744 * This is safe to call from within a preemption notifier. 1745 */ 1746void preempt_notifier_unregister(struct preempt_notifier *notifier) 1747{ 1748 hlist_del(¬ifier->link); 1749} 1750EXPORT_SYMBOL_GPL(preempt_notifier_unregister); 1751 1752static void fire_sched_in_preempt_notifiers(struct task_struct *curr) 1753{ 1754 struct preempt_notifier *notifier; 1755 1756 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link) 1757 notifier->ops->sched_in(notifier, raw_smp_processor_id()); 1758} 1759 1760static void 1761fire_sched_out_preempt_notifiers(struct task_struct *curr, 1762 struct task_struct *next) 1763{ 1764 struct preempt_notifier *notifier; 1765 1766 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link) 1767 notifier->ops->sched_out(notifier, next); 1768} 1769 1770#else /* !CONFIG_PREEMPT_NOTIFIERS */ 1771 1772static void fire_sched_in_preempt_notifiers(struct task_struct *curr) 1773{ 1774} 1775 1776static void 1777fire_sched_out_preempt_notifiers(struct task_struct *curr, 1778 struct task_struct *next) 1779{ 1780} 1781 1782#endif /* CONFIG_PREEMPT_NOTIFIERS */ 1783 1784/** 1785 * prepare_task_switch - prepare to switch tasks 1786 * @rq: the runqueue preparing to switch 1787 * @prev: the current task that is being switched out 1788 * @next: the task we are going to switch to. 1789 * 1790 * This is called with the rq lock held and interrupts off. It must 1791 * be paired with a subsequent finish_task_switch after the context 1792 * switch. 1793 * 1794 * prepare_task_switch sets up locking and calls architecture specific 1795 * hooks. 1796 */ 1797static inline void 1798prepare_task_switch(struct rq *rq, struct task_struct *prev, 1799 struct task_struct *next) 1800{ 1801 trace_sched_switch(prev, next); 1802 sched_info_switch(prev, next); 1803 perf_event_task_sched_out(prev, next); 1804 fire_sched_out_preempt_notifiers(prev, next); 1805 prepare_lock_switch(rq, next); 1806 prepare_arch_switch(next); 1807} 1808 1809/** 1810 * finish_task_switch - clean up after a task-switch 1811 * @rq: runqueue associated with task-switch 1812 * @prev: the thread we just switched away from. 1813 * 1814 * finish_task_switch must be called after the context switch, paired 1815 * with a prepare_task_switch call before the context switch. 1816 * finish_task_switch will reconcile locking set up by prepare_task_switch, 1817 * and do any other architecture-specific cleanup actions. 1818 * 1819 * Note that we may have delayed dropping an mm in context_switch(). If 1820 * so, we finish that here outside of the runqueue lock. (Doing it 1821 * with the lock held can cause deadlocks; see schedule() for 1822 * details.) 1823 */ 1824static void finish_task_switch(struct rq *rq, struct task_struct *prev) 1825 __releases(rq->lock) 1826{ 1827 struct mm_struct *mm = rq->prev_mm; 1828 long prev_state; 1829 1830 rq->prev_mm = NULL; 1831 1832 /* 1833 * A task struct has one reference for the use as "current". 1834 * If a task dies, then it sets TASK_DEAD in tsk->state and calls 1835 * schedule one last time. The schedule call will never return, and 1836 * the scheduled task must drop that reference. 1837 * The test for TASK_DEAD must occur while the runqueue locks are 1838 * still held, otherwise prev could be scheduled on another cpu, die 1839 * there before we look at prev->state, and then the reference would 1840 * be dropped twice. 1841 * Manfred Spraul <manfred@colorfullife.com> 1842 */ 1843 prev_state = prev->state; 1844 vtime_task_switch(prev); 1845 finish_arch_switch(prev); 1846 perf_event_task_sched_in(prev, current); 1847 finish_lock_switch(rq, prev); 1848 finish_arch_post_lock_switch(); 1849 1850 fire_sched_in_preempt_notifiers(current); 1851 if (mm) 1852 mmdrop(mm); 1853 if (unlikely(prev_state == TASK_DEAD)) { 1854 /* 1855 * Remove function-return probe instances associated with this 1856 * task and put them back on the free list. 1857 */ 1858 kprobe_flush_task(prev); 1859 put_task_struct(prev); 1860 } 1861} 1862 1863#ifdef CONFIG_SMP 1864 1865/* assumes rq->lock is held */ 1866static inline void pre_schedule(struct rq *rq, struct task_struct *prev) 1867{ 1868 if (prev->sched_class->pre_schedule) 1869 prev->sched_class->pre_schedule(rq, prev); 1870} 1871 1872/* rq->lock is NOT held, but preemption is disabled */ 1873static inline void post_schedule(struct rq *rq) 1874{ 1875 if (rq->post_schedule) { 1876 unsigned long flags; 1877 1878 raw_spin_lock_irqsave(&rq->lock, flags); 1879 if (rq->curr->sched_class->post_schedule) 1880 rq->curr->sched_class->post_schedule(rq); 1881 raw_spin_unlock_irqrestore(&rq->lock, flags); 1882 1883 rq->post_schedule = 0; 1884 } 1885} 1886 1887#else 1888 1889static inline void pre_schedule(struct rq *rq, struct task_struct *p) 1890{ 1891} 1892 1893static inline void post_schedule(struct rq *rq) 1894{ 1895} 1896 1897#endif 1898 1899/** 1900 * schedule_tail - first thing a freshly forked thread must call. 1901 * @prev: the thread we just switched away from. 1902 */ 1903asmlinkage void schedule_tail(struct task_struct *prev) 1904 __releases(rq->lock) 1905{ 1906 struct rq *rq = this_rq(); 1907 1908 finish_task_switch(rq, prev); 1909 1910 /* 1911 * FIXME: do we need to worry about rq being invalidated by the 1912 * task_switch? 1913 */ 1914 post_schedule(rq); 1915 1916#ifdef __ARCH_WANT_UNLOCKED_CTXSW 1917 /* In this case, finish_task_switch does not reenable preemption */ 1918 preempt_enable(); 1919#endif 1920 if (current->set_child_tid) 1921 put_user(task_pid_vnr(current), current->set_child_tid); 1922} 1923 1924/* 1925 * context_switch - switch to the new MM and the new 1926 * thread's register state. 1927 */ 1928static inline void 1929context_switch(struct rq *rq, struct task_struct *prev, 1930 struct task_struct *next) 1931{ 1932 struct mm_struct *mm, *oldmm; 1933 1934 prepare_task_switch(rq, prev, next); 1935 1936 mm = next->mm; 1937 oldmm = prev->active_mm; 1938 /* 1939 * For paravirt, this is coupled with an exit in switch_to to 1940 * combine the page table reload and the switch backend into 1941 * one hypercall. 1942 */ 1943 arch_start_context_switch(prev); 1944 1945 if (!mm) { 1946 next->active_mm = oldmm; 1947 atomic_inc(&oldmm->mm_count); 1948 enter_lazy_tlb(oldmm, next); 1949 } else 1950 switch_mm(oldmm, mm, next); 1951 1952 if (!prev->mm) { 1953 prev->active_mm = NULL; 1954 rq->prev_mm = oldmm; 1955 } 1956 /* 1957 * Since the runqueue lock will be released by the next 1958 * task (which is an invalid locking op but in the case 1959 * of the scheduler it's an obvious special-case), so we 1960 * do an early lockdep release here: 1961 */ 1962#ifndef __ARCH_WANT_UNLOCKED_CTXSW 1963 spin_release(&rq->lock.dep_map, 1, _THIS_IP_); 1964#endif 1965 1966 context_tracking_task_switch(prev, next); 1967 /* Here we just switch the register state and the stack. */ 1968 switch_to(prev, next, prev); 1969 1970 barrier(); 1971 /* 1972 * this_rq must be evaluated again because prev may have moved 1973 * CPUs since it called schedule(), thus the 'rq' on its stack 1974 * frame will be invalid. 1975 */ 1976 finish_task_switch(this_rq(), prev); 1977} 1978 1979/* 1980 * nr_running and nr_context_switches: 1981 * 1982 * externally visible scheduler statistics: current number of runnable 1983 * threads, total number of context switches performed since bootup. 1984 */ 1985unsigned long nr_running(void) 1986{ 1987 unsigned long i, sum = 0; 1988 1989 for_each_online_cpu(i) 1990 sum += cpu_rq(i)->nr_running; 1991 1992 return sum; 1993} 1994 1995unsigned long long nr_context_switches(void) 1996{ 1997 int i; 1998 unsigned long long sum = 0; 1999 2000 for_each_possible_cpu(i) 2001 sum += cpu_rq(i)->nr_switches; 2002 2003 return sum; 2004} 2005 2006unsigned long nr_iowait(void) 2007{ 2008 unsigned long i, sum = 0; 2009 2010 for_each_possible_cpu(i) 2011 sum += atomic_read(&cpu_rq(i)->nr_iowait); 2012 2013 return sum; 2014} 2015 2016unsigned long nr_iowait_cpu(int cpu) 2017{ 2018 struct rq *this = cpu_rq(cpu); 2019 return atomic_read(&this->nr_iowait); 2020} 2021 2022unsigned long this_cpu_load(void) 2023{ 2024 struct rq *this = this_rq(); 2025 return this->cpu_load[0]; 2026} 2027 2028 2029/* 2030 * Global load-average calculations 2031 * 2032 * We take a distributed and async approach to calculating the global load-avg 2033 * in order to minimize overhead. 2034 * 2035 * The global load average is an exponentially decaying average of nr_running + 2036 * nr_uninterruptible. 2037 * 2038 * Once every LOAD_FREQ: 2039 * 2040 * nr_active = 0; 2041 * for_each_possible_cpu(cpu) 2042 * nr_active += cpu_of(cpu)->nr_running + cpu_of(cpu)->nr_uninterruptible; 2043 * 2044 * avenrun[n] = avenrun[0] * exp_n + nr_active * (1 - exp_n) 2045 * 2046 * Due to a number of reasons the above turns in the mess below: 2047 * 2048 * - for_each_possible_cpu() is prohibitively expensive on machines with 2049 * serious number of cpus, therefore we need to take a distributed approach 2050 * to calculating nr_active. 2051 * 2052 * \Sum_i x_i(t) = \Sum_i x_i(t) - x_i(t_0) | x_i(t_0) := 0 2053 * = \Sum_i { \Sum_j=1 x_i(t_j) - x_i(t_j-1) } 2054 * 2055 * So assuming nr_active := 0 when we start out -- true per definition, we 2056 * can simply take per-cpu deltas and fold those into a global accumulate 2057 * to obtain the same result. See calc_load_fold_active(). 2058 * 2059 * Furthermore, in order to avoid synchronizing all per-cpu delta folding 2060 * across the machine, we assume 10 ticks is sufficient time for every 2061 * cpu to have completed this task. 2062 * 2063 * This places an upper-bound on the IRQ-off latency of the machine. Then 2064 * again, being late doesn't loose the delta, just wrecks the sample. 2065 * 2066 * - cpu_rq()->nr_uninterruptible isn't accurately tracked per-cpu because 2067 * this would add another cross-cpu cacheline miss and atomic operation 2068 * to the wakeup path. Instead we increment on whatever cpu the task ran 2069 * when it went into uninterruptible state and decrement on whatever cpu 2070 * did the wakeup. This means that only the sum of nr_uninterruptible over 2071 * all cpus yields the correct result. 2072 * 2073 * This covers the NO_HZ=n code, for extra head-aches, see the comment below. 2074 */ 2075 2076/* Variables and functions for calc_load */ 2077static atomic_long_t calc_load_tasks; 2078static unsigned long calc_load_update; 2079unsigned long avenrun[3]; 2080EXPORT_SYMBOL(avenrun); /* should be removed */ 2081 2082/** 2083 * get_avenrun - get the load average array 2084 * @loads: pointer to dest load array 2085 * @offset: offset to add 2086 * @shift: shift count to shift the result left 2087 * 2088 * These values are estimates at best, so no need for locking. 2089 */ 2090void get_avenrun(unsigned long *loads, unsigned long offset, int shift) 2091{ 2092 loads[0] = (avenrun[0] + offset) << shift; 2093 loads[1] = (avenrun[1] + offset) << shift; 2094 loads[2] = (avenrun[2] + offset) << shift; 2095} 2096 2097static long calc_load_fold_active(struct rq *this_rq) 2098{ 2099 long nr_active, delta = 0; 2100 2101 nr_active = this_rq->nr_running; 2102 nr_active += (long) this_rq->nr_uninterruptible; 2103 2104 if (nr_active != this_rq->calc_load_active) { 2105 delta = nr_active - this_rq->calc_load_active; 2106 this_rq->calc_load_active = nr_active; 2107 } 2108 2109 return delta; 2110} 2111 2112/* 2113 * a1 = a0 * e + a * (1 - e) 2114 */ 2115static unsigned long 2116calc_load(unsigned long load, unsigned long exp, unsigned long active) 2117{ 2118 load *= exp; 2119 load += active * (FIXED_1 - exp); 2120 load += 1UL << (FSHIFT - 1); 2121 return load >> FSHIFT; 2122} 2123 2124#ifdef CONFIG_NO_HZ 2125/* 2126 * Handle NO_HZ for the global load-average. 2127 * 2128 * Since the above described distributed algorithm to compute the global 2129 * load-average relies on per-cpu sampling from the tick, it is affected by 2130 * NO_HZ. 2131 * 2132 * The basic idea is to fold the nr_active delta into a global idle-delta upon 2133 * entering NO_HZ state such that we can include this as an 'extra' cpu delta 2134 * when we read the global state. 2135 * 2136 * Obviously reality has to ruin such a delightfully simple scheme: 2137 * 2138 * - When we go NO_HZ idle during the window, we can negate our sample 2139 * contribution, causing under-accounting. 2140 * 2141 * We avoid this by keeping two idle-delta counters and flipping them 2142 * when the window starts, thus separating old and new NO_HZ load. 2143 * 2144 * The only trick is the slight shift in index flip for read vs write. 2145 * 2146 * 0s 5s 10s 15s 2147 * +10 +10 +10 +10 2148 * |-|-----------|-|-----------|-|-----------|-| 2149 * r:0 0 1 1 0 0 1 1 0 2150 * w:0 1 1 0 0 1 1 0 0 2151 * 2152 * This ensures we'll fold the old idle contribution in this window while 2153 * accumlating the new one. 2154 * 2155 * - When we wake up from NO_HZ idle during the window, we push up our 2156 * contribution, since we effectively move our sample point to a known 2157 * busy state. 2158 * 2159 * This is solved by pushing the window forward, and thus skipping the 2160 * sample, for this cpu (effectively using the idle-delta for this cpu which 2161 * was in effect at the time the window opened). This also solves the issue 2162 * of having to deal with a cpu having been in NOHZ idle for multiple 2163 * LOAD_FREQ intervals. 2164 * 2165 * When making the ILB scale, we should try to pull this in as well. 2166 */ 2167static atomic_long_t calc_load_idle[2]; 2168static int calc_load_idx; 2169 2170static inline int calc_load_write_idx(void) 2171{ 2172 int idx = calc_load_idx; 2173 2174 /* 2175 * See calc_global_nohz(), if we observe the new index, we also 2176 * need to observe the new update time. 2177 */ 2178 smp_rmb(); 2179 2180 /* 2181 * If the folding window started, make sure we start writing in the 2182 * next idle-delta. 2183 */ 2184 if (!time_before(jiffies, calc_load_update)) 2185 idx++; 2186 2187 return idx & 1; 2188} 2189 2190static inline int calc_load_read_idx(void) 2191{ 2192 return calc_load_idx & 1; 2193} 2194 2195void calc_load_enter_idle(void) 2196{ 2197 struct rq *this_rq = this_rq(); 2198 long delta; 2199 2200 /* 2201 * We're going into NOHZ mode, if there's any pending delta, fold it 2202 * into the pending idle delta. 2203 */ 2204 delta = calc_load_fold_active(this_rq); 2205 if (delta) { 2206 int idx = calc_load_write_idx(); 2207 atomic_long_add(delta, &calc_load_idle[idx]); 2208 } 2209} 2210 2211void calc_load_exit_idle(void) 2212{ 2213 struct rq *this_rq = this_rq(); 2214 2215 /* 2216 * If we're still before the sample window, we're done. 2217 */ 2218 if (time_before(jiffies, this_rq->calc_load_update)) 2219 return; 2220 2221 /* 2222 * We woke inside or after the sample window, this means we're already 2223 * accounted through the nohz accounting, so skip the entire deal and 2224 * sync up for the next window. 2225 */ 2226 this_rq->calc_load_update = calc_load_update; 2227 if (time_before(jiffies, this_rq->calc_load_update + 10)) 2228 this_rq->calc_load_update += LOAD_FREQ; 2229} 2230 2231static long calc_load_fold_idle(void) 2232{ 2233 int idx = calc_load_read_idx(); 2234 long delta = 0; 2235 2236 if (atomic_long_read(&calc_load_idle[idx])) 2237 delta = atomic_long_xchg(&calc_load_idle[idx], 0); 2238 2239 return delta; 2240} 2241 2242/** 2243 * fixed_power_int - compute: x^n, in O(log n) time 2244 * 2245 * @x: base of the power 2246 * @frac_bits: fractional bits of @x 2247 * @n: power to raise @x to. 2248 * 2249 * By exploiting the relation between the definition of the natural power 2250 * function: x^n := x*x*...*x (x multiplied by itself for n times), and 2251 * the binary encoding of numbers used by computers: n := \Sum n_i * 2^i, 2252 * (where: n_i \elem {0, 1}, the binary vector representing n), 2253 * we find: x^n := x^(\Sum n_i * 2^i) := \Prod x^(n_i * 2^i), which is 2254 * of course trivially computable in O(log_2 n), the length of our binary 2255 * vector. 2256 */ 2257static unsigned long 2258fixed_power_int(unsigned long x, unsigned int frac_bits, unsigned int n) 2259{ 2260 unsigned long result = 1UL << frac_bits; 2261 2262 if (n) for (;;) { 2263 if (n & 1) { 2264 result *= x; 2265 result += 1UL << (frac_bits - 1); 2266 result >>= frac_bits; 2267 } 2268 n >>= 1; 2269 if (!n) 2270 break; 2271 x *= x; 2272 x += 1UL << (frac_bits - 1); 2273 x >>= frac_bits; 2274 } 2275 2276 return result; 2277} 2278 2279/* 2280 * a1 = a0 * e + a * (1 - e) 2281 * 2282 * a2 = a1 * e + a * (1 - e) 2283 * = (a0 * e + a * (1 - e)) * e + a * (1 - e) 2284 * = a0 * e^2 + a * (1 - e) * (1 + e) 2285 * 2286 * a3 = a2 * e + a * (1 - e) 2287 * = (a0 * e^2 + a * (1 - e) * (1 + e)) * e + a * (1 - e) 2288 * = a0 * e^3 + a * (1 - e) * (1 + e + e^2) 2289 * 2290 * ... 2291 * 2292 * an = a0 * e^n + a * (1 - e) * (1 + e + ... + e^n-1) [1] 2293 * = a0 * e^n + a * (1 - e) * (1 - e^n)/(1 - e) 2294 * = a0 * e^n + a * (1 - e^n) 2295 * 2296 * [1] application of the geometric series: 2297 * 2298 * n 1 - x^(n+1) 2299 * S_n := \Sum x^i = ------------- 2300 * i=0 1 - x 2301 */ 2302static unsigned long 2303calc_load_n(unsigned long load, unsigned long exp, 2304 unsigned long active, unsigned int n) 2305{ 2306 2307 return calc_load(load, fixed_power_int(exp, FSHIFT, n), active); 2308} 2309 2310/* 2311 * NO_HZ can leave us missing all per-cpu ticks calling 2312 * calc_load_account_active(), but since an idle CPU folds its delta into 2313 * calc_load_tasks_idle per calc_load_account_idle(), all we need to do is fold 2314 * in the pending idle delta if our idle period crossed a load cycle boundary. 2315 * 2316 * Once we've updated the global active value, we need to apply the exponential 2317 * weights adjusted to the number of cycles missed. 2318 */ 2319static void calc_global_nohz(void) 2320{ 2321 long delta, active, n; 2322 2323 if (!time_before(jiffies, calc_load_update + 10)) { 2324 /* 2325 * Catch-up, fold however many we are behind still 2326 */ 2327 delta = jiffies - calc_load_update - 10; 2328 n = 1 + (delta / LOAD_FREQ); 2329 2330 active = atomic_long_read(&calc_load_tasks); 2331 active = active > 0 ? active * FIXED_1 : 0; 2332 2333 avenrun[0] = calc_load_n(avenrun[0], EXP_1, active, n); 2334 avenrun[1] = calc_load_n(avenrun[1], EXP_5, active, n); 2335 avenrun[2] = calc_load_n(avenrun[2], EXP_15, active, n); 2336 2337 calc_load_update += n * LOAD_FREQ; 2338 } 2339 2340 /* 2341 * Flip the idle index... 2342 * 2343 * Make sure we first write the new time then flip the index, so that 2344 * calc_load_write_idx() will see the new time when it reads the new 2345 * index, this avoids a double flip messing things up. 2346 */ 2347 smp_wmb(); 2348 calc_load_idx++; 2349} 2350#else /* !CONFIG_NO_HZ */ 2351 2352static inline long calc_load_fold_idle(void) { return 0; } 2353static inline void calc_global_nohz(void) { } 2354 2355#endif /* CONFIG_NO_HZ */ 2356 2357/* 2358 * calc_load - update the avenrun load estimates 10 ticks after the 2359 * CPUs have updated calc_load_tasks. 2360 */ 2361void calc_global_load(unsigned long ticks) 2362{ 2363 long active, delta; 2364 2365 if (time_before(jiffies, calc_load_update + 10)) 2366 return; 2367 2368 /* 2369 * Fold the 'old' idle-delta to include all NO_HZ cpus. 2370 */ 2371 delta = calc_load_fold_idle(); 2372 if (delta) 2373 atomic_long_add(delta, &calc_load_tasks); 2374 2375 active = atomic_long_read(&calc_load_tasks); 2376 active = active > 0 ? active * FIXED_1 : 0; 2377 2378 avenrun[0] = calc_load(avenrun[0], EXP_1, active); 2379 avenrun[1] = calc_load(avenrun[1], EXP_5, active); 2380 avenrun[2] = calc_load(avenrun[2], EXP_15, active); 2381 2382 calc_load_update += LOAD_FREQ; 2383 2384 /* 2385 * In case we idled for multiple LOAD_FREQ intervals, catch up in bulk. 2386 */ 2387 calc_global_nohz(); 2388} 2389 2390/* 2391 * Called from update_cpu_load() to periodically update this CPU's 2392 * active count. 2393 */ 2394static void calc_load_account_active(struct rq *this_rq) 2395{ 2396 long delta; 2397 2398 if (time_before(jiffies, this_rq->calc_load_update)) 2399 return; 2400 2401 delta = calc_load_fold_active(this_rq); 2402 if (delta) 2403 atomic_long_add(delta, &calc_load_tasks); 2404 2405 this_rq->calc_load_update += LOAD_FREQ; 2406} 2407 2408/* 2409 * End of global load-average stuff 2410 */ 2411 2412/* 2413 * The exact cpuload at various idx values, calculated at every tick would be 2414 * load = (2^idx - 1) / 2^idx * load + 1 / 2^idx * cur_load 2415 * 2416 * If a cpu misses updates for n-1 ticks (as it was idle) and update gets called 2417 * on nth tick when cpu may be busy, then we have: 2418 * load = ((2^idx - 1) / 2^idx)^(n-1) * load 2419 * load = (2^idx - 1) / 2^idx) * load + 1 / 2^idx * cur_load 2420 * 2421 * decay_load_missed() below does efficient calculation of 2422 * load = ((2^idx - 1) / 2^idx)^(n-1) * load 2423 * avoiding 0..n-1 loop doing load = ((2^idx - 1) / 2^idx) * load 2424 * 2425 * The calculation is approximated on a 128 point scale. 2426 * degrade_zero_ticks is the number of ticks after which load at any 2427 * particular idx is approximated to be zero. 2428 * degrade_factor is a precomputed table, a row for each load idx. 2429 * Each column corresponds to degradation factor for a power of two ticks, 2430 * based on 128 point scale. 2431 * Example: 2432 * row 2, col 3 (=12) says that the degradation at load idx 2 after 2433 * 8 ticks is 12/128 (which is an approximation of exact factor 3^8/4^8). 2434 * 2435 * With this power of 2 load factors, we can degrade the load n times 2436 * by looking at 1 bits in n and doing as many mult/shift instead of 2437 * n mult/shifts needed by the exact degradation. 2438 */ 2439#define DEGRADE_SHIFT 7 2440static const unsigned char 2441 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128}; 2442static const unsigned char 2443 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = { 2444 {0, 0, 0, 0, 0, 0, 0, 0}, 2445 {64, 32, 8, 0, 0, 0, 0, 0}, 2446 {96, 72, 40, 12, 1, 0, 0}, 2447 {112, 98, 75, 43, 15, 1, 0}, 2448 {120, 112, 98, 76, 45, 16, 2} }; 2449 2450/* 2451 * Update cpu_load for any missed ticks, due to tickless idle. The backlog 2452 * would be when CPU is idle and so we just decay the old load without 2453 * adding any new load. 2454 */ 2455static unsigned long 2456decay_load_missed(unsigned long load, unsigned long missed_updates, int idx) 2457{ 2458 int j = 0; 2459 2460 if (!missed_updates) 2461 return load; 2462 2463 if (missed_updates >= degrade_zero_ticks[idx]) 2464 return 0; 2465 2466 if (idx == 1) 2467 return load >> missed_updates; 2468 2469 while (missed_updates) { 2470 if (missed_updates % 2) 2471 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT; 2472 2473 missed_updates >>= 1; 2474 j++; 2475 } 2476 return load; 2477} 2478 2479/* 2480 * Update rq->cpu_load[] statistics. This function is usually called every 2481 * scheduler tick (TICK_NSEC). With tickless idle this will not be called 2482 * every tick. We fix it up based on jiffies. 2483 */ 2484static void __update_cpu_load(struct rq *this_rq, unsigned long this_load, 2485 unsigned long pending_updates) 2486{ 2487 int i, scale; 2488 2489 this_rq->nr_load_updates++; 2490 2491 /* Update our load: */ 2492 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */ 2493 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) { 2494 unsigned long old_load, new_load; 2495 2496 /* scale is effectively 1 << i now, and >> i divides by scale */ 2497 2498 old_load = this_rq->cpu_load[i]; 2499 old_load = decay_load_missed(old_load, pending_updates - 1, i); 2500 new_load = this_load; 2501 /* 2502 * Round up the averaging division if load is increasing. This 2503 * prevents us from getting stuck on 9 if the load is 10, for 2504 * example. 2505 */ 2506 if (new_load > old_load) 2507 new_load += scale - 1; 2508 2509 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i; 2510 } 2511 2512 sched_avg_update(this_rq); 2513} 2514 2515#ifdef CONFIG_NO_HZ 2516/* 2517 * There is no sane way to deal with nohz on smp when using jiffies because the 2518 * cpu doing the jiffies update might drift wrt the cpu doing the jiffy reading 2519 * causing off-by-one errors in observed deltas; {0,2} instead of {1,1}. 2520 * 2521 * Therefore we cannot use the delta approach from the regular tick since that 2522 * would seriously skew the load calculation. However we'll make do for those 2523 * updates happening while idle (nohz_idle_balance) or coming out of idle 2524 * (tick_nohz_idle_exit). 2525 * 2526 * This means we might still be one tick off for nohz periods. 2527 */ 2528 2529/* 2530 * Called from nohz_idle_balance() to update the load ratings before doing the 2531 * idle balance. 2532 */ 2533void update_idle_cpu_load(struct rq *this_rq) 2534{ 2535 unsigned long curr_jiffies = ACCESS_ONCE(jiffies); 2536 unsigned long load = this_rq->load.weight; 2537 unsigned long pending_updates; 2538 2539 /* 2540 * bail if there's load or we're actually up-to-date. 2541 */ 2542 if (load || curr_jiffies == this_rq->last_load_update_tick) 2543 return; 2544 2545 pending_updates = curr_jiffies - this_rq->last_load_update_tick; 2546 this_rq->last_load_update_tick = curr_jiffies; 2547 2548 __update_cpu_load(this_rq, load, pending_updates); 2549} 2550 2551/* 2552 * Called from tick_nohz_idle_exit() -- try and fix up the ticks we missed. 2553 */ 2554void update_cpu_load_nohz(void) 2555{ 2556 struct rq *this_rq = this_rq(); 2557 unsigned long curr_jiffies = ACCESS_ONCE(jiffies); 2558 unsigned long pending_updates; 2559 2560 if (curr_jiffies == this_rq->last_load_update_tick) 2561 return; 2562 2563 raw_spin_lock(&this_rq->lock); 2564 pending_updates = curr_jiffies - this_rq->last_load_update_tick; 2565 if (pending_updates) { 2566 this_rq->last_load_update_tick = curr_jiffies; 2567 /* 2568 * We were idle, this means load 0, the current load might be 2569 * !0 due to remote wakeups and the sort. 2570 */ 2571 __update_cpu_load(this_rq, 0, pending_updates); 2572 } 2573 raw_spin_unlock(&this_rq->lock); 2574} 2575#endif /* CONFIG_NO_HZ */ 2576 2577/* 2578 * Called from scheduler_tick() 2579 */ 2580static void update_cpu_load_active(struct rq *this_rq) 2581{ 2582 /* 2583 * See the mess around update_idle_cpu_load() / update_cpu_load_nohz(). 2584 */ 2585 this_rq->last_load_update_tick = jiffies; 2586 __update_cpu_load(this_rq, this_rq->load.weight, 1); 2587 2588 calc_load_account_active(this_rq); 2589} 2590 2591#ifdef CONFIG_SMP 2592 2593/* 2594 * sched_exec - execve() is a valuable balancing opportunity, because at 2595 * this point the task has the smallest effective memory and cache footprint. 2596 */ 2597void sched_exec(void) 2598{ 2599 struct task_struct *p = current; 2600 unsigned long flags; 2601 int dest_cpu; 2602 2603 raw_spin_lock_irqsave(&p->pi_lock, flags); 2604 dest_cpu = p->sched_class->select_task_rq(p, SD_BALANCE_EXEC, 0); 2605 if (dest_cpu == smp_processor_id()) 2606 goto unlock; 2607 2608 if (likely(cpu_active(dest_cpu))) { 2609 struct migration_arg arg = { p, dest_cpu }; 2610 2611 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 2612 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg); 2613 return; 2614 } 2615unlock: 2616 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 2617} 2618 2619#endif 2620 2621DEFINE_PER_CPU(struct kernel_stat, kstat); 2622DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat); 2623 2624EXPORT_PER_CPU_SYMBOL(kstat); 2625EXPORT_PER_CPU_SYMBOL(kernel_cpustat); 2626 2627/* 2628 * Return any ns on the sched_clock that have not yet been accounted in 2629 * @p in case that task is currently running. 2630 * 2631 * Called with task_rq_lock() held on @rq. 2632 */ 2633static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq) 2634{ 2635 u64 ns = 0; 2636 2637 if (task_current(rq, p)) { 2638 update_rq_clock(rq); 2639 ns = rq->clock_task - p->se.exec_start; 2640 if ((s64)ns < 0) 2641 ns = 0; 2642 } 2643 2644 return ns; 2645} 2646 2647unsigned long long task_delta_exec(struct task_struct *p) 2648{ 2649 unsigned long flags; 2650 struct rq *rq; 2651 u64 ns = 0; 2652 2653 rq = task_rq_lock(p, &flags); 2654 ns = do_task_delta_exec(p, rq); 2655 task_rq_unlock(rq, p, &flags); 2656 2657 return ns; 2658} 2659 2660/* 2661 * Return accounted runtime for the task. 2662 * In case the task is currently running, return the runtime plus current's 2663 * pending runtime that have not been accounted yet. 2664 */ 2665unsigned long long task_sched_runtime(struct task_struct *p) 2666{ 2667 unsigned long flags; 2668 struct rq *rq; 2669 u64 ns = 0; 2670 2671 rq = task_rq_lock(p, &flags); 2672 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq); 2673 task_rq_unlock(rq, p, &flags); 2674 2675 return ns; 2676} 2677 2678/* 2679 * This function gets called by the timer code, with HZ frequency. 2680 * We call it with interrupts disabled. 2681 */ 2682void scheduler_tick(void) 2683{ 2684 int cpu = smp_processor_id(); 2685 struct rq *rq = cpu_rq(cpu); 2686 struct task_struct *curr = rq->curr; 2687 2688 sched_clock_tick(); 2689 2690 raw_spin_lock(&rq->lock); 2691 update_rq_clock(rq); 2692 update_cpu_load_active(rq); 2693 curr->sched_class->task_tick(rq, curr, 0); 2694 raw_spin_unlock(&rq->lock); 2695 2696 perf_event_task_tick(); 2697 2698#ifdef CONFIG_SMP 2699 rq->idle_balance = idle_cpu(cpu); 2700 trigger_load_balance(rq, cpu); 2701#endif 2702} 2703 2704notrace unsigned long get_parent_ip(unsigned long addr) 2705{ 2706 if (in_lock_functions(addr)) { 2707 addr = CALLER_ADDR2; 2708 if (in_lock_functions(addr)) 2709 addr = CALLER_ADDR3; 2710 } 2711 return addr; 2712} 2713 2714#if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \ 2715 defined(CONFIG_PREEMPT_TRACER)) 2716 2717void __kprobes add_preempt_count(int val) 2718{ 2719#ifdef CONFIG_DEBUG_PREEMPT 2720 /* 2721 * Underflow? 2722 */ 2723 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0))) 2724 return; 2725#endif 2726 preempt_count() += val; 2727#ifdef CONFIG_DEBUG_PREEMPT 2728 /* 2729 * Spinlock count overflowing soon? 2730 */ 2731 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >= 2732 PREEMPT_MASK - 10); 2733#endif 2734 if (preempt_count() == val) 2735 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1)); 2736} 2737EXPORT_SYMBOL(add_preempt_count); 2738 2739void __kprobes sub_preempt_count(int val) 2740{ 2741#ifdef CONFIG_DEBUG_PREEMPT 2742 /* 2743 * Underflow? 2744 */ 2745 if (DEBUG_LOCKS_WARN_ON(val > preempt_count())) 2746 return; 2747 /* 2748 * Is the spinlock portion underflowing? 2749 */ 2750 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) && 2751 !(preempt_count() & PREEMPT_MASK))) 2752 return; 2753#endif 2754 2755 if (preempt_count() == val) 2756 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1)); 2757 preempt_count() -= val; 2758} 2759EXPORT_SYMBOL(sub_preempt_count); 2760 2761#endif 2762 2763/* 2764 * Print scheduling while atomic bug: 2765 */ 2766static noinline void __schedule_bug(struct task_struct *prev) 2767{ 2768 if (oops_in_progress) 2769 return; 2770 2771 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n", 2772 prev->comm, prev->pid, preempt_count()); 2773 2774 debug_show_held_locks(prev); 2775 print_modules(); 2776 if (irqs_disabled()) 2777 print_irqtrace_events(prev); 2778 dump_stack(); 2779 add_taint(TAINT_WARN, LOCKDEP_STILL_OK); 2780} 2781 2782/* 2783 * Various schedule()-time debugging checks and statistics: 2784 */ 2785static inline void schedule_debug(struct task_struct *prev) 2786{ 2787 /* 2788 * Test if we are atomic. Since do_exit() needs to call into 2789 * schedule() atomically, we ignore that path for now. 2790 * Otherwise, whine if we are scheduling when we should not be. 2791 */ 2792 if (unlikely(in_atomic_preempt_off() && !prev->exit_state)) 2793 __schedule_bug(prev); 2794 rcu_sleep_check(); 2795 2796 profile_hit(SCHED_PROFILING, __builtin_return_address(0)); 2797 2798 schedstat_inc(this_rq(), sched_count); 2799} 2800 2801static void put_prev_task(struct rq *rq, struct task_struct *prev) 2802{ 2803 if (prev->on_rq || rq->skip_clock_update < 0) 2804 update_rq_clock(rq); 2805 prev->sched_class->put_prev_task(rq, prev); 2806} 2807 2808/* 2809 * Pick up the highest-prio task: 2810 */ 2811static inline struct task_struct * 2812pick_next_task(struct rq *rq) 2813{ 2814 const struct sched_class *class; 2815 struct task_struct *p; 2816 2817 /* 2818 * Optimization: we know that if all tasks are in 2819 * the fair class we can call that function directly: 2820 */ 2821 if (likely(rq->nr_running == rq->cfs.h_nr_running)) { 2822 p = fair_sched_class.pick_next_task(rq); 2823 if (likely(p)) 2824 return p; 2825 } 2826 2827 for_each_class(class) { 2828 p = class->pick_next_task(rq); 2829 if (p) 2830 return p; 2831 } 2832 2833 BUG(); /* the idle class will always have a runnable task */ 2834} 2835 2836/* 2837 * __schedule() is the main scheduler function. 2838 * 2839 * The main means of driving the scheduler and thus entering this function are: 2840 * 2841 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc. 2842 * 2843 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return 2844 * paths. For example, see arch/x86/entry_64.S. 2845 * 2846 * To drive preemption between tasks, the scheduler sets the flag in timer 2847 * interrupt handler scheduler_tick(). 2848 * 2849 * 3. Wakeups don't really cause entry into schedule(). They add a 2850 * task to the run-queue and that's it. 2851 * 2852 * Now, if the new task added to the run-queue preempts the current 2853 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets 2854 * called on the nearest possible occasion: 2855 * 2856 * - If the kernel is preemptible (CONFIG_PREEMPT=y): 2857 * 2858 * - in syscall or exception context, at the next outmost 2859 * preempt_enable(). (this might be as soon as the wake_up()'s 2860 * spin_unlock()!) 2861 * 2862 * - in IRQ context, return from interrupt-handler to 2863 * preemptible context 2864 * 2865 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set) 2866 * then at the next: 2867 * 2868 * - cond_resched() call 2869 * - explicit schedule() call 2870 * - return from syscall or exception to user-space 2871 * - return from interrupt-handler to user-space 2872 */ 2873static void __sched __schedule(void) 2874{ 2875 struct task_struct *prev, *next; 2876 unsigned long *switch_count; 2877 struct rq *rq; 2878 int cpu; 2879 2880need_resched: 2881 preempt_disable(); 2882 cpu = smp_processor_id(); 2883 rq = cpu_rq(cpu); 2884 rcu_note_context_switch(cpu); 2885 prev = rq->curr; 2886 2887 schedule_debug(prev); 2888 2889 if (sched_feat(HRTICK)) 2890 hrtick_clear(rq); 2891 2892 raw_spin_lock_irq(&rq->lock); 2893 2894 switch_count = &prev->nivcsw; 2895 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) { 2896 if (unlikely(signal_pending_state(prev->state, prev))) { 2897 prev->state = TASK_RUNNING; 2898 } else { 2899 deactivate_task(rq, prev, DEQUEUE_SLEEP); 2900 prev->on_rq = 0; 2901 2902 /* 2903 * If a worker went to sleep, notify and ask workqueue 2904 * whether it wants to wake up a task to maintain 2905 * concurrency. 2906 */ 2907 if (prev->flags & PF_WQ_WORKER) { 2908 struct task_struct *to_wakeup; 2909 2910 to_wakeup = wq_worker_sleeping(prev, cpu); 2911 if (to_wakeup) 2912 try_to_wake_up_local(to_wakeup); 2913 } 2914 } 2915 switch_count = &prev->nvcsw; 2916 } 2917 2918 pre_schedule(rq, prev); 2919 2920 if (unlikely(!rq->nr_running)) 2921 idle_balance(cpu, rq); 2922 2923 put_prev_task(rq, prev); 2924 next = pick_next_task(rq); 2925 clear_tsk_need_resched(prev); 2926 rq->skip_clock_update = 0; 2927 2928 if (likely(prev != next)) { 2929 rq->nr_switches++; 2930 rq->curr = next; 2931 ++*switch_count; 2932 2933 context_switch(rq, prev, next); /* unlocks the rq */ 2934 /* 2935 * The context switch have flipped the stack from under us 2936 * and restored the local variables which were saved when 2937 * this task called schedule() in the past. prev == current 2938 * is still correct, but it can be moved to another cpu/rq. 2939 */ 2940 cpu = smp_processor_id(); 2941 rq = cpu_rq(cpu); 2942 } else 2943 raw_spin_unlock_irq(&rq->lock); 2944 2945 post_schedule(rq); 2946 2947 sched_preempt_enable_no_resched(); 2948 if (need_resched()) 2949 goto need_resched; 2950} 2951 2952static inline void sched_submit_work(struct task_struct *tsk) 2953{ 2954 if (!tsk->state || tsk_is_pi_blocked(tsk)) 2955 return; 2956 /* 2957 * If we are going to sleep and we have plugged IO queued, 2958 * make sure to submit it to avoid deadlocks. 2959 */ 2960 if (blk_needs_flush_plug(tsk)) 2961 blk_schedule_flush_plug(tsk); 2962} 2963 2964asmlinkage void __sched schedule(void) 2965{ 2966 struct task_struct *tsk = current; 2967 2968 sched_submit_work(tsk); 2969 __schedule(); 2970} 2971EXPORT_SYMBOL(schedule); 2972 2973#ifdef CONFIG_CONTEXT_TRACKING 2974asmlinkage void __sched schedule_user(void) 2975{ 2976 /* 2977 * If we come here after a random call to set_need_resched(), 2978 * or we have been woken up remotely but the IPI has not yet arrived, 2979 * we haven't yet exited the RCU idle mode. Do it here manually until 2980 * we find a better solution. 2981 */ 2982 user_exit(); 2983 schedule(); 2984 user_enter(); 2985} 2986#endif 2987 2988/** 2989 * schedule_preempt_disabled - called with preemption disabled 2990 * 2991 * Returns with preemption disabled. Note: preempt_count must be 1 2992 */ 2993void __sched schedule_preempt_disabled(void) 2994{ 2995 sched_preempt_enable_no_resched(); 2996 schedule(); 2997 preempt_disable(); 2998} 2999 3000#ifdef CONFIG_MUTEX_SPIN_ON_OWNER 3001 3002static inline bool owner_running(struct mutex *lock, struct task_struct *owner) 3003{ 3004 if (lock->owner != owner) 3005 return false; 3006 3007 /* 3008 * Ensure we emit the owner->on_cpu, dereference _after_ checking 3009 * lock->owner still matches owner, if that fails, owner might 3010 * point to free()d memory, if it still matches, the rcu_read_lock() 3011 * ensures the memory stays valid. 3012 */ 3013 barrier(); 3014 3015 return owner->on_cpu; 3016} 3017 3018/* 3019 * Look out! "owner" is an entirely speculative pointer 3020 * access and not reliable. 3021 */ 3022int mutex_spin_on_owner(struct mutex *lock, struct task_struct *owner) 3023{ 3024 if (!sched_feat(OWNER_SPIN)) 3025 return 0; 3026 3027 rcu_read_lock(); 3028 while (owner_running(lock, owner)) { 3029 if (need_resched()) 3030 break; 3031 3032 arch_mutex_cpu_relax(); 3033 } 3034 rcu_read_unlock(); 3035 3036 /* 3037 * We break out the loop above on need_resched() and when the 3038 * owner changed, which is a sign for heavy contention. Return 3039 * success only when lock->owner is NULL. 3040 */ 3041 return lock->owner == NULL; 3042} 3043#endif 3044 3045#ifdef CONFIG_PREEMPT 3046/* 3047 * this is the entry point to schedule() from in-kernel preemption 3048 * off of preempt_enable. Kernel preemptions off return from interrupt 3049 * occur there and call schedule directly. 3050 */ 3051asmlinkage void __sched notrace preempt_schedule(void) 3052{ 3053 struct thread_info *ti = current_thread_info(); 3054 3055 /* 3056 * If there is a non-zero preempt_count or interrupts are disabled, 3057 * we do not want to preempt the current task. Just return.. 3058 */ 3059 if (likely(ti->preempt_count || irqs_disabled())) 3060 return; 3061 3062 do { 3063 add_preempt_count_notrace(PREEMPT_ACTIVE); 3064 __schedule(); 3065 sub_preempt_count_notrace(PREEMPT_ACTIVE); 3066 3067 /* 3068 * Check again in case we missed a preemption opportunity 3069 * between schedule and now. 3070 */ 3071 barrier(); 3072 } while (need_resched()); 3073} 3074EXPORT_SYMBOL(preempt_schedule); 3075 3076/* 3077 * this is the entry point to schedule() from kernel preemption 3078 * off of irq context. 3079 * Note, that this is called and return with irqs disabled. This will 3080 * protect us against recursive calling from irq. 3081 */ 3082asmlinkage void __sched preempt_schedule_irq(void) 3083{ 3084 struct thread_info *ti = current_thread_info(); 3085 enum ctx_state prev_state; 3086 3087 /* Catch callers which need to be fixed */ 3088 BUG_ON(ti->preempt_count || !irqs_disabled()); 3089 3090 prev_state = exception_enter(); 3091 3092 do { 3093 add_preempt_count(PREEMPT_ACTIVE); 3094 local_irq_enable(); 3095 __schedule(); 3096 local_irq_disable(); 3097 sub_preempt_count(PREEMPT_ACTIVE); 3098 3099 /* 3100 * Check again in case we missed a preemption opportunity 3101 * between schedule and now. 3102 */ 3103 barrier(); 3104 } while (need_resched()); 3105 3106 exception_exit(prev_state); 3107} 3108 3109#endif /* CONFIG_PREEMPT */ 3110 3111int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags, 3112 void *key) 3113{ 3114 return try_to_wake_up(curr->private, mode, wake_flags); 3115} 3116EXPORT_SYMBOL(default_wake_function); 3117 3118/* 3119 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just 3120 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve 3121 * number) then we wake all the non-exclusive tasks and one exclusive task. 3122 * 3123 * There are circumstances in which we can try to wake a task which has already 3124 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns 3125 * zero in this (rare) case, and we handle it by continuing to scan the queue. 3126 */ 3127static void __wake_up_common(wait_queue_head_t *q, unsigned int mode, 3128 int nr_exclusive, int wake_flags, void *key) 3129{ 3130 wait_queue_t *curr, *next; 3131 3132 list_for_each_entry_safe(curr, next, &q->task_list, task_list) { 3133 unsigned flags = curr->flags; 3134 3135 if (curr->func(curr, mode, wake_flags, key) && 3136 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive) 3137 break; 3138 } 3139} 3140 3141/** 3142 * __wake_up - wake up threads blocked on a waitqueue. 3143 * @q: the waitqueue 3144 * @mode: which threads 3145 * @nr_exclusive: how many wake-one or wake-many threads to wake up 3146 * @key: is directly passed to the wakeup function 3147 * 3148 * It may be assumed that this function implies a write memory barrier before 3149 * changing the task state if and only if any tasks are woken up. 3150 */ 3151void __wake_up(wait_queue_head_t *q, unsigned int mode, 3152 int nr_exclusive, void *key) 3153{ 3154 unsigned long flags; 3155 3156 spin_lock_irqsave(&q->lock, flags); 3157 __wake_up_common(q, mode, nr_exclusive, 0, key); 3158 spin_unlock_irqrestore(&q->lock, flags); 3159} 3160EXPORT_SYMBOL(__wake_up); 3161 3162/* 3163 * Same as __wake_up but called with the spinlock in wait_queue_head_t held. 3164 */ 3165void __wake_up_locked(wait_queue_head_t *q, unsigned int mode, int nr) 3166{ 3167 __wake_up_common(q, mode, nr, 0, NULL); 3168} 3169EXPORT_SYMBOL_GPL(__wake_up_locked); 3170 3171void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key) 3172{ 3173 __wake_up_common(q, mode, 1, 0, key); 3174} 3175EXPORT_SYMBOL_GPL(__wake_up_locked_key); 3176 3177/** 3178 * __wake_up_sync_key - wake up threads blocked on a waitqueue. 3179 * @q: the waitqueue 3180 * @mode: which threads 3181 * @nr_exclusive: how many wake-one or wake-many threads to wake up 3182 * @key: opaque value to be passed to wakeup targets 3183 * 3184 * The sync wakeup differs that the waker knows that it will schedule 3185 * away soon, so while the target thread will be woken up, it will not 3186 * be migrated to another CPU - ie. the two threads are 'synchronized' 3187 * with each other. This can prevent needless bouncing between CPUs. 3188 * 3189 * On UP it can prevent extra preemption. 3190 * 3191 * It may be assumed that this function implies a write memory barrier before 3192 * changing the task state if and only if any tasks are woken up. 3193 */ 3194void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode, 3195 int nr_exclusive, void *key) 3196{ 3197 unsigned long flags; 3198 int wake_flags = WF_SYNC; 3199 3200 if (unlikely(!q)) 3201 return; 3202 3203 if (unlikely(!nr_exclusive)) 3204 wake_flags = 0; 3205 3206 spin_lock_irqsave(&q->lock, flags); 3207 __wake_up_common(q, mode, nr_exclusive, wake_flags, key); 3208 spin_unlock_irqrestore(&q->lock, flags); 3209} 3210EXPORT_SYMBOL_GPL(__wake_up_sync_key); 3211 3212/* 3213 * __wake_up_sync - see __wake_up_sync_key() 3214 */ 3215void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive) 3216{ 3217 __wake_up_sync_key(q, mode, nr_exclusive, NULL); 3218} 3219EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */ 3220 3221/** 3222 * complete: - signals a single thread waiting on this completion 3223 * @x: holds the state of this particular completion 3224 * 3225 * This will wake up a single thread waiting on this completion. Threads will be 3226 * awakened in the same order in which they were queued. 3227 * 3228 * See also complete_all(), wait_for_completion() and related routines. 3229 * 3230 * It may be assumed that this function implies a write memory barrier before 3231 * changing the task state if and only if any tasks are woken up. 3232 */ 3233void complete(struct completion *x) 3234{ 3235 unsigned long flags; 3236 3237 spin_lock_irqsave(&x->wait.lock, flags); 3238 x->done++; 3239 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL); 3240 spin_unlock_irqrestore(&x->wait.lock, flags); 3241} 3242EXPORT_SYMBOL(complete); 3243 3244/** 3245 * complete_all: - signals all threads waiting on this completion 3246 * @x: holds the state of this particular completion 3247 * 3248 * This will wake up all threads waiting on this particular completion event. 3249 * 3250 * It may be assumed that this function implies a write memory barrier before 3251 * changing the task state if and only if any tasks are woken up. 3252 */ 3253void complete_all(struct completion *x) 3254{ 3255 unsigned long flags; 3256 3257 spin_lock_irqsave(&x->wait.lock, flags); 3258 x->done += UINT_MAX/2; 3259 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL); 3260 spin_unlock_irqrestore(&x->wait.lock, flags); 3261} 3262EXPORT_SYMBOL(complete_all); 3263 3264static inline long __sched 3265do_wait_for_common(struct completion *x, 3266 long (*action)(long), long timeout, int state) 3267{ 3268 if (!x->done) { 3269 DECLARE_WAITQUEUE(wait, current); 3270 3271 __add_wait_queue_tail_exclusive(&x->wait, &wait); 3272 do { 3273 if (signal_pending_state(state, current)) { 3274 timeout = -ERESTARTSYS; 3275 break; 3276 } 3277 __set_current_state(state); 3278 spin_unlock_irq(&x->wait.lock); 3279 timeout = action(timeout); 3280 spin_lock_irq(&x->wait.lock); 3281 } while (!x->done && timeout); 3282 __remove_wait_queue(&x->wait, &wait); 3283 if (!x->done) 3284 return timeout; 3285 } 3286 x->done--; 3287 return timeout ?: 1; 3288} 3289 3290static inline long __sched 3291__wait_for_common(struct completion *x, 3292 long (*action)(long), long timeout, int state) 3293{ 3294 might_sleep(); 3295 3296 spin_lock_irq(&x->wait.lock); 3297 timeout = do_wait_for_common(x, action, timeout, state); 3298 spin_unlock_irq(&x->wait.lock); 3299 return timeout; 3300} 3301 3302static long __sched 3303wait_for_common(struct completion *x, long timeout, int state) 3304{ 3305 return __wait_for_common(x, schedule_timeout, timeout, state); 3306} 3307 3308static long __sched 3309wait_for_common_io(struct completion *x, long timeout, int state) 3310{ 3311 return __wait_for_common(x, io_schedule_timeout, timeout, state); 3312} 3313 3314/** 3315 * wait_for_completion: - waits for completion of a task 3316 * @x: holds the state of this particular completion 3317 * 3318 * This waits to be signaled for completion of a specific task. It is NOT 3319 * interruptible and there is no timeout. 3320 * 3321 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout 3322 * and interrupt capability. Also see complete(). 3323 */ 3324void __sched wait_for_completion(struct completion *x) 3325{ 3326 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE); 3327} 3328EXPORT_SYMBOL(wait_for_completion); 3329 3330/** 3331 * wait_for_completion_timeout: - waits for completion of a task (w/timeout) 3332 * @x: holds the state of this particular completion 3333 * @timeout: timeout value in jiffies 3334 * 3335 * This waits for either a completion of a specific task to be signaled or for a 3336 * specified timeout to expire. The timeout is in jiffies. It is not 3337 * interruptible. 3338 * 3339 * The return value is 0 if timed out, and positive (at least 1, or number of 3340 * jiffies left till timeout) if completed. 3341 */ 3342unsigned long __sched 3343wait_for_completion_timeout(struct completion *x, unsigned long timeout) 3344{ 3345 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE); 3346} 3347EXPORT_SYMBOL(wait_for_completion_timeout); 3348 3349/** 3350 * wait_for_completion_io: - waits for completion of a task 3351 * @x: holds the state of this particular completion 3352 * 3353 * This waits to be signaled for completion of a specific task. It is NOT 3354 * interruptible and there is no timeout. The caller is accounted as waiting 3355 * for IO. 3356 */ 3357void __sched wait_for_completion_io(struct completion *x) 3358{ 3359 wait_for_common_io(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE); 3360} 3361EXPORT_SYMBOL(wait_for_completion_io); 3362 3363/** 3364 * wait_for_completion_io_timeout: - waits for completion of a task (w/timeout) 3365 * @x: holds the state of this particular completion 3366 * @timeout: timeout value in jiffies 3367 * 3368 * This waits for either a completion of a specific task to be signaled or for a 3369 * specified timeout to expire. The timeout is in jiffies. It is not 3370 * interruptible. The caller is accounted as waiting for IO. 3371 * 3372 * The return value is 0 if timed out, and positive (at least 1, or number of 3373 * jiffies left till timeout) if completed. 3374 */ 3375unsigned long __sched 3376wait_for_completion_io_timeout(struct completion *x, unsigned long timeout) 3377{ 3378 return wait_for_common_io(x, timeout, TASK_UNINTERRUPTIBLE); 3379} 3380EXPORT_SYMBOL(wait_for_completion_io_timeout); 3381 3382/** 3383 * wait_for_completion_interruptible: - waits for completion of a task (w/intr) 3384 * @x: holds the state of this particular completion 3385 * 3386 * This waits for completion of a specific task to be signaled. It is 3387 * interruptible. 3388 * 3389 * The return value is -ERESTARTSYS if interrupted, 0 if completed. 3390 */ 3391int __sched wait_for_completion_interruptible(struct completion *x) 3392{ 3393 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE); 3394 if (t == -ERESTARTSYS) 3395 return t; 3396 return 0; 3397} 3398EXPORT_SYMBOL(wait_for_completion_interruptible); 3399 3400/** 3401 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr)) 3402 * @x: holds the state of this particular completion 3403 * @timeout: timeout value in jiffies 3404 * 3405 * This waits for either a completion of a specific task to be signaled or for a 3406 * specified timeout to expire. It is interruptible. The timeout is in jiffies. 3407 * 3408 * The return value is -ERESTARTSYS if interrupted, 0 if timed out, 3409 * positive (at least 1, or number of jiffies left till timeout) if completed. 3410 */ 3411long __sched 3412wait_for_completion_interruptible_timeout(struct completion *x, 3413 unsigned long timeout) 3414{ 3415 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE); 3416} 3417EXPORT_SYMBOL(wait_for_completion_interruptible_timeout); 3418 3419/** 3420 * wait_for_completion_killable: - waits for completion of a task (killable) 3421 * @x: holds the state of this particular completion 3422 * 3423 * This waits to be signaled for completion of a specific task. It can be 3424 * interrupted by a kill signal. 3425 * 3426 * The return value is -ERESTARTSYS if interrupted, 0 if completed. 3427 */ 3428int __sched wait_for_completion_killable(struct completion *x) 3429{ 3430 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE); 3431 if (t == -ERESTARTSYS) 3432 return t; 3433 return 0; 3434} 3435EXPORT_SYMBOL(wait_for_completion_killable); 3436 3437/** 3438 * wait_for_completion_killable_timeout: - waits for completion of a task (w/(to,killable)) 3439 * @x: holds the state of this particular completion 3440 * @timeout: timeout value in jiffies 3441 * 3442 * This waits for either a completion of a specific task to be 3443 * signaled or for a specified timeout to expire. It can be 3444 * interrupted by a kill signal. The timeout is in jiffies. 3445 * 3446 * The return value is -ERESTARTSYS if interrupted, 0 if timed out, 3447 * positive (at least 1, or number of jiffies left till timeout) if completed. 3448 */ 3449long __sched 3450wait_for_completion_killable_timeout(struct completion *x, 3451 unsigned long timeout) 3452{ 3453 return wait_for_common(x, timeout, TASK_KILLABLE); 3454} 3455EXPORT_SYMBOL(wait_for_completion_killable_timeout); 3456 3457/** 3458 * try_wait_for_completion - try to decrement a completion without blocking 3459 * @x: completion structure 3460 * 3461 * Returns: 0 if a decrement cannot be done without blocking 3462 * 1 if a decrement succeeded. 3463 * 3464 * If a completion is being used as a counting completion, 3465 * attempt to decrement the counter without blocking. This 3466 * enables us to avoid waiting if the resource the completion 3467 * is protecting is not available. 3468 */ 3469bool try_wait_for_completion(struct completion *x) 3470{ 3471 unsigned long flags; 3472 int ret = 1; 3473 3474 spin_lock_irqsave(&x->wait.lock, flags); 3475 if (!x->done) 3476 ret = 0; 3477 else 3478 x->done--; 3479 spin_unlock_irqrestore(&x->wait.lock, flags); 3480 return ret; 3481} 3482EXPORT_SYMBOL(try_wait_for_completion); 3483 3484/** 3485 * completion_done - Test to see if a completion has any waiters 3486 * @x: completion structure 3487 * 3488 * Returns: 0 if there are waiters (wait_for_completion() in progress) 3489 * 1 if there are no waiters. 3490 * 3491 */ 3492bool completion_done(struct completion *x) 3493{ 3494 unsigned long flags; 3495 int ret = 1; 3496 3497 spin_lock_irqsave(&x->wait.lock, flags); 3498 if (!x->done) 3499 ret = 0; 3500 spin_unlock_irqrestore(&x->wait.lock, flags); 3501 return ret; 3502} 3503EXPORT_SYMBOL(completion_done); 3504 3505static long __sched 3506sleep_on_common(wait_queue_head_t *q, int state, long timeout) 3507{ 3508 unsigned long flags; 3509 wait_queue_t wait; 3510 3511 init_waitqueue_entry(&wait, current); 3512 3513 __set_current_state(state); 3514 3515 spin_lock_irqsave(&q->lock, flags); 3516 __add_wait_queue(q, &wait); 3517 spin_unlock(&q->lock); 3518 timeout = schedule_timeout(timeout); 3519 spin_lock_irq(&q->lock); 3520 __remove_wait_queue(q, &wait); 3521 spin_unlock_irqrestore(&q->lock, flags); 3522 3523 return timeout; 3524} 3525 3526void __sched interruptible_sleep_on(wait_queue_head_t *q) 3527{ 3528 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT); 3529} 3530EXPORT_SYMBOL(interruptible_sleep_on); 3531 3532long __sched 3533interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout) 3534{ 3535 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout); 3536} 3537EXPORT_SYMBOL(interruptible_sleep_on_timeout); 3538 3539void __sched sleep_on(wait_queue_head_t *q) 3540{ 3541 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT); 3542} 3543EXPORT_SYMBOL(sleep_on); 3544 3545long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout) 3546{ 3547 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout); 3548} 3549EXPORT_SYMBOL(sleep_on_timeout); 3550 3551#ifdef CONFIG_RT_MUTEXES 3552 3553/* 3554 * rt_mutex_setprio - set the current priority of a task 3555 * @p: task 3556 * @prio: prio value (kernel-internal form) 3557 * 3558 * This function changes the 'effective' priority of a task. It does 3559 * not touch ->normal_prio like __setscheduler(). 3560 * 3561 * Used by the rt_mutex code to implement priority inheritance logic. 3562 */ 3563void rt_mutex_setprio(struct task_struct *p, int prio) 3564{ 3565 int oldprio, on_rq, running; 3566 struct rq *rq; 3567 const struct sched_class *prev_class; 3568 3569 BUG_ON(prio < 0 || prio > MAX_PRIO); 3570 3571 rq = __task_rq_lock(p); 3572 3573 /* 3574 * Idle task boosting is a nono in general. There is one 3575 * exception, when PREEMPT_RT and NOHZ is active: 3576 * 3577 * The idle task calls get_next_timer_interrupt() and holds 3578 * the timer wheel base->lock on the CPU and another CPU wants 3579 * to access the timer (probably to cancel it). We can safely 3580 * ignore the boosting request, as the idle CPU runs this code 3581 * with interrupts disabled and will complete the lock 3582 * protected section without being interrupted. So there is no 3583 * real need to boost. 3584 */ 3585 if (unlikely(p == rq->idle)) { 3586 WARN_ON(p != rq->curr); 3587 WARN_ON(p->pi_blocked_on); 3588 goto out_unlock; 3589 } 3590 3591 trace_sched_pi_setprio(p, prio); 3592 oldprio = p->prio; 3593 prev_class = p->sched_class; 3594 on_rq = p->on_rq; 3595 running = task_current(rq, p); 3596 if (on_rq) 3597 dequeue_task(rq, p, 0); 3598 if (running) 3599 p->sched_class->put_prev_task(rq, p); 3600 3601 if (rt_prio(prio)) 3602 p->sched_class = &rt_sched_class; 3603 else 3604 p->sched_class = &fair_sched_class; 3605 3606 p->prio = prio; 3607 3608 if (running) 3609 p->sched_class->set_curr_task(rq); 3610 if (on_rq) 3611 enqueue_task(rq, p, oldprio < prio ? ENQUEUE_HEAD : 0); 3612 3613 check_class_changed(rq, p, prev_class, oldprio); 3614out_unlock: 3615 __task_rq_unlock(rq); 3616} 3617#endif 3618void set_user_nice(struct task_struct *p, long nice) 3619{ 3620 int old_prio, delta, on_rq; 3621 unsigned long flags; 3622 struct rq *rq; 3623 3624 if (TASK_NICE(p) == nice || nice < -20 || nice > 19) 3625 return; 3626 /* 3627 * We have to be careful, if called from sys_setpriority(), 3628 * the task might be in the middle of scheduling on another CPU. 3629 */ 3630 rq = task_rq_lock(p, &flags); 3631 /* 3632 * The RT priorities are set via sched_setscheduler(), but we still 3633 * allow the 'normal' nice value to be set - but as expected 3634 * it wont have any effect on scheduling until the task is 3635 * SCHED_FIFO/SCHED_RR: 3636 */ 3637 if (task_has_rt_policy(p)) { 3638 p->static_prio = NICE_TO_PRIO(nice); 3639 goto out_unlock; 3640 } 3641 on_rq = p->on_rq; 3642 if (on_rq) 3643 dequeue_task(rq, p, 0); 3644 3645 p->static_prio = NICE_TO_PRIO(nice); 3646 set_load_weight(p); 3647 old_prio = p->prio; 3648 p->prio = effective_prio(p); 3649 delta = p->prio - old_prio; 3650 3651 if (on_rq) { 3652 enqueue_task(rq, p, 0); 3653 /* 3654 * If the task increased its priority or is running and 3655 * lowered its priority, then reschedule its CPU: 3656 */ 3657 if (delta < 0 || (delta > 0 && task_running(rq, p))) 3658 resched_task(rq->curr); 3659 } 3660out_unlock: 3661 task_rq_unlock(rq, p, &flags); 3662} 3663EXPORT_SYMBOL(set_user_nice); 3664 3665/* 3666 * can_nice - check if a task can reduce its nice value 3667 * @p: task 3668 * @nice: nice value 3669 */ 3670int can_nice(const struct task_struct *p, const int nice) 3671{ 3672 /* convert nice value [19,-20] to rlimit style value [1,40] */ 3673 int nice_rlim = 20 - nice; 3674 3675 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) || 3676 capable(CAP_SYS_NICE)); 3677} 3678 3679#ifdef __ARCH_WANT_SYS_NICE 3680 3681/* 3682 * sys_nice - change the priority of the current process. 3683 * @increment: priority increment 3684 * 3685 * sys_setpriority is a more generic, but much slower function that 3686 * does similar things. 3687 */ 3688SYSCALL_DEFINE1(nice, int, increment) 3689{ 3690 long nice, retval; 3691 3692 /* 3693 * Setpriority might change our priority at the same moment. 3694 * We don't have to worry. Conceptually one call occurs first 3695 * and we have a single winner. 3696 */ 3697 if (increment < -40) 3698 increment = -40; 3699 if (increment > 40) 3700 increment = 40; 3701 3702 nice = TASK_NICE(current) + increment; 3703 if (nice < -20) 3704 nice = -20; 3705 if (nice > 19) 3706 nice = 19; 3707 3708 if (increment < 0 && !can_nice(current, nice)) 3709 return -EPERM; 3710 3711 retval = security_task_setnice(current, nice); 3712 if (retval) 3713 return retval; 3714 3715 set_user_nice(current, nice); 3716 return 0; 3717} 3718 3719#endif 3720 3721/** 3722 * task_prio - return the priority value of a given task. 3723 * @p: the task in question. 3724 * 3725 * This is the priority value as seen by users in /proc. 3726 * RT tasks are offset by -200. Normal tasks are centered 3727 * around 0, value goes from -16 to +15. 3728 */ 3729int task_prio(const struct task_struct *p) 3730{ 3731 return p->prio - MAX_RT_PRIO; 3732} 3733 3734/** 3735 * task_nice - return the nice value of a given task. 3736 * @p: the task in question. 3737 */ 3738int task_nice(const struct task_struct *p) 3739{ 3740 return TASK_NICE(p); 3741} 3742EXPORT_SYMBOL(task_nice); 3743 3744/** 3745 * idle_cpu - is a given cpu idle currently? 3746 * @cpu: the processor in question. 3747 */ 3748int idle_cpu(int cpu) 3749{ 3750 struct rq *rq = cpu_rq(cpu); 3751 3752 if (rq->curr != rq->idle) 3753 return 0; 3754 3755 if (rq->nr_running) 3756 return 0; 3757 3758#ifdef CONFIG_SMP 3759 if (!llist_empty(&rq->wake_list)) 3760 return 0; 3761#endif 3762 3763 return 1; 3764} 3765 3766/** 3767 * idle_task - return the idle task for a given cpu. 3768 * @cpu: the processor in question. 3769 */ 3770struct task_struct *idle_task(int cpu) 3771{ 3772 return cpu_rq(cpu)->idle; 3773} 3774 3775/** 3776 * find_process_by_pid - find a process with a matching PID value. 3777 * @pid: the pid in question. 3778 */ 3779static struct task_struct *find_process_by_pid(pid_t pid) 3780{ 3781 return pid ? find_task_by_vpid(pid) : current; 3782} 3783 3784/* Actually do priority change: must hold rq lock. */ 3785static void 3786__setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio) 3787{ 3788 p->policy = policy; 3789 p->rt_priority = prio; 3790 p->normal_prio = normal_prio(p); 3791 /* we are holding p->pi_lock already */ 3792 p->prio = rt_mutex_getprio(p); 3793 if (rt_prio(p->prio)) 3794 p->sched_class = &rt_sched_class; 3795 else 3796 p->sched_class = &fair_sched_class; 3797 set_load_weight(p); 3798} 3799 3800/* 3801 * check the target process has a UID that matches the current process's 3802 */ 3803static bool check_same_owner(struct task_struct *p) 3804{ 3805 const struct cred *cred = current_cred(), *pcred; 3806 bool match; 3807 3808 rcu_read_lock(); 3809 pcred = __task_cred(p); 3810 match = (uid_eq(cred->euid, pcred->euid) || 3811 uid_eq(cred->euid, pcred->uid)); 3812 rcu_read_unlock(); 3813 return match; 3814} 3815 3816static int __sched_setscheduler(struct task_struct *p, int policy, 3817 const struct sched_param *param, bool user) 3818{ 3819 int retval, oldprio, oldpolicy = -1, on_rq, running; 3820 unsigned long flags; 3821 const struct sched_class *prev_class; 3822 struct rq *rq; 3823 int reset_on_fork; 3824 3825 /* may grab non-irq protected spin_locks */ 3826 BUG_ON(in_interrupt()); 3827recheck: 3828 /* double check policy once rq lock held */ 3829 if (policy < 0) { 3830 reset_on_fork = p->sched_reset_on_fork; 3831 policy = oldpolicy = p->policy; 3832 } else { 3833 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK); 3834 policy &= ~SCHED_RESET_ON_FORK; 3835 3836 if (policy != SCHED_FIFO && policy != SCHED_RR && 3837 policy != SCHED_NORMAL && policy != SCHED_BATCH && 3838 policy != SCHED_IDLE) 3839 return -EINVAL; 3840 } 3841 3842 /* 3843 * Valid priorities for SCHED_FIFO and SCHED_RR are 3844 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL, 3845 * SCHED_BATCH and SCHED_IDLE is 0. 3846 */ 3847 if (param->sched_priority < 0 || 3848 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) || 3849 (!p->mm && param->sched_priority > MAX_RT_PRIO-1)) 3850 return -EINVAL; 3851 if (rt_policy(policy) != (param->sched_priority != 0)) 3852 return -EINVAL; 3853 3854 /* 3855 * Allow unprivileged RT tasks to decrease priority: 3856 */ 3857 if (user && !capable(CAP_SYS_NICE)) { 3858 if (rt_policy(policy)) { 3859 unsigned long rlim_rtprio = 3860 task_rlimit(p, RLIMIT_RTPRIO); 3861 3862 /* can't set/change the rt policy */ 3863 if (policy != p->policy && !rlim_rtprio) 3864 return -EPERM; 3865 3866 /* can't increase priority */ 3867 if (param->sched_priority > p->rt_priority && 3868 param->sched_priority > rlim_rtprio) 3869 return -EPERM; 3870 } 3871 3872 /* 3873 * Treat SCHED_IDLE as nice 20. Only allow a switch to 3874 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it. 3875 */ 3876 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) { 3877 if (!can_nice(p, TASK_NICE(p))) 3878 return -EPERM; 3879 } 3880 3881 /* can't change other user's priorities */ 3882 if (!check_same_owner(p)) 3883 return -EPERM; 3884 3885 /* Normal users shall not reset the sched_reset_on_fork flag */ 3886 if (p->sched_reset_on_fork && !reset_on_fork) 3887 return -EPERM; 3888 } 3889 3890 if (user) { 3891 retval = security_task_setscheduler(p); 3892 if (retval) 3893 return retval; 3894 } 3895 3896 /* 3897 * make sure no PI-waiters arrive (or leave) while we are 3898 * changing the priority of the task: 3899 * 3900 * To be able to change p->policy safely, the appropriate 3901 * runqueue lock must be held. 3902 */ 3903 rq = task_rq_lock(p, &flags); 3904 3905 /* 3906 * Changing the policy of the stop threads its a very bad idea 3907 */ 3908 if (p == rq->stop) { 3909 task_rq_unlock(rq, p, &flags); 3910 return -EINVAL; 3911 } 3912 3913 /* 3914 * If not changing anything there's no need to proceed further: 3915 */ 3916 if (unlikely(policy == p->policy && (!rt_policy(policy) || 3917 param->sched_priority == p->rt_priority))) { 3918 task_rq_unlock(rq, p, &flags); 3919 return 0; 3920 } 3921 3922#ifdef CONFIG_RT_GROUP_SCHED 3923 if (user) { 3924 /* 3925 * Do not allow realtime tasks into groups that have no runtime 3926 * assigned. 3927 */ 3928 if (rt_bandwidth_enabled() && rt_policy(policy) && 3929 task_group(p)->rt_bandwidth.rt_runtime == 0 && 3930 !task_group_is_autogroup(task_group(p))) { 3931 task_rq_unlock(rq, p, &flags); 3932 return -EPERM; 3933 } 3934 } 3935#endif 3936 3937 /* recheck policy now with rq lock held */ 3938 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) { 3939 policy = oldpolicy = -1; 3940 task_rq_unlock(rq, p, &flags); 3941 goto recheck; 3942 } 3943 on_rq = p->on_rq; 3944 running = task_current(rq, p); 3945 if (on_rq) 3946 dequeue_task(rq, p, 0); 3947 if (running) 3948 p->sched_class->put_prev_task(rq, p); 3949 3950 p->sched_reset_on_fork = reset_on_fork; 3951 3952 oldprio = p->prio; 3953 prev_class = p->sched_class; 3954 __setscheduler(rq, p, policy, param->sched_priority); 3955 3956 if (running) 3957 p->sched_class->set_curr_task(rq); 3958 if (on_rq) 3959 enqueue_task(rq, p, 0); 3960 3961 check_class_changed(rq, p, prev_class, oldprio); 3962 task_rq_unlock(rq, p, &flags); 3963 3964 rt_mutex_adjust_pi(p); 3965 3966 return 0; 3967} 3968 3969/** 3970 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread. 3971 * @p: the task in question. 3972 * @policy: new policy. 3973 * @param: structure containing the new RT priority. 3974 * 3975 * NOTE that the task may be already dead. 3976 */ 3977int sched_setscheduler(struct task_struct *p, int policy, 3978 const struct sched_param *param) 3979{ 3980 return __sched_setscheduler(p, policy, param, true); 3981} 3982EXPORT_SYMBOL_GPL(sched_setscheduler); 3983 3984/** 3985 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace. 3986 * @p: the task in question. 3987 * @policy: new policy. 3988 * @param: structure containing the new RT priority. 3989 * 3990 * Just like sched_setscheduler, only don't bother checking if the 3991 * current context has permission. For example, this is needed in 3992 * stop_machine(): we create temporary high priority worker threads, 3993 * but our caller might not have that capability. 3994 */ 3995int sched_setscheduler_nocheck(struct task_struct *p, int policy, 3996 const struct sched_param *param) 3997{ 3998 return __sched_setscheduler(p, policy, param, false); 3999} 4000 4001static int 4002do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param) 4003{ 4004 struct sched_param lparam; 4005 struct task_struct *p; 4006 int retval; 4007 4008 if (!param || pid < 0) 4009 return -EINVAL; 4010 if (copy_from_user(&lparam, param, sizeof(struct sched_param))) 4011 return -EFAULT; 4012 4013 rcu_read_lock(); 4014 retval = -ESRCH; 4015 p = find_process_by_pid(pid); 4016 if (p != NULL) 4017 retval = sched_setscheduler(p, policy, &lparam); 4018 rcu_read_unlock(); 4019 4020 return retval; 4021} 4022 4023/** 4024 * sys_sched_setscheduler - set/change the scheduler policy and RT priority 4025 * @pid: the pid in question. 4026 * @policy: new policy. 4027 * @param: structure containing the new RT priority. 4028 */ 4029SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy, 4030 struct sched_param __user *, param) 4031{ 4032 /* negative values for policy are not valid */ 4033 if (policy < 0) 4034 return -EINVAL; 4035 4036 return do_sched_setscheduler(pid, policy, param); 4037} 4038 4039/** 4040 * sys_sched_setparam - set/change the RT priority of a thread 4041 * @pid: the pid in question. 4042 * @param: structure containing the new RT priority. 4043 */ 4044SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param) 4045{ 4046 return do_sched_setscheduler(pid, -1, param); 4047} 4048 4049/** 4050 * sys_sched_getscheduler - get the policy (scheduling class) of a thread 4051 * @pid: the pid in question. 4052 */ 4053SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid) 4054{ 4055 struct task_struct *p; 4056 int retval; 4057 4058 if (pid < 0) 4059 return -EINVAL; 4060 4061 retval = -ESRCH; 4062 rcu_read_lock(); 4063 p = find_process_by_pid(pid); 4064 if (p) { 4065 retval = security_task_getscheduler(p); 4066 if (!retval) 4067 retval = p->policy 4068 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0); 4069 } 4070 rcu_read_unlock(); 4071 return retval; 4072} 4073 4074/** 4075 * sys_sched_getparam - get the RT priority of a thread 4076 * @pid: the pid in question. 4077 * @param: structure containing the RT priority. 4078 */ 4079SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param) 4080{ 4081 struct sched_param lp; 4082 struct task_struct *p; 4083 int retval; 4084 4085 if (!param || pid < 0) 4086 return -EINVAL; 4087 4088 rcu_read_lock(); 4089 p = find_process_by_pid(pid); 4090 retval = -ESRCH; 4091 if (!p) 4092 goto out_unlock; 4093 4094 retval = security_task_getscheduler(p); 4095 if (retval) 4096 goto out_unlock; 4097 4098 lp.sched_priority = p->rt_priority; 4099 rcu_read_unlock(); 4100 4101 /* 4102 * This one might sleep, we cannot do it with a spinlock held ... 4103 */ 4104 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0; 4105 4106 return retval; 4107 4108out_unlock: 4109 rcu_read_unlock(); 4110 return retval; 4111} 4112 4113long sched_setaffinity(pid_t pid, const struct cpumask *in_mask) 4114{ 4115 cpumask_var_t cpus_allowed, new_mask; 4116 struct task_struct *p; 4117 int retval; 4118 4119 get_online_cpus(); 4120 rcu_read_lock(); 4121 4122 p = find_process_by_pid(pid); 4123 if (!p) { 4124 rcu_read_unlock(); 4125 put_online_cpus(); 4126 return -ESRCH; 4127 } 4128 4129 /* Prevent p going away */ 4130 get_task_struct(p); 4131 rcu_read_unlock(); 4132 4133 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) { 4134 retval = -ENOMEM; 4135 goto out_put_task; 4136 } 4137 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) { 4138 retval = -ENOMEM; 4139 goto out_free_cpus_allowed; 4140 } 4141 retval = -EPERM; 4142 if (!check_same_owner(p)) { 4143 rcu_read_lock(); 4144 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) { 4145 rcu_read_unlock(); 4146 goto out_unlock; 4147 } 4148 rcu_read_unlock(); 4149 } 4150 4151 retval = security_task_setscheduler(p); 4152 if (retval) 4153 goto out_unlock; 4154 4155 cpuset_cpus_allowed(p, cpus_allowed); 4156 cpumask_and(new_mask, in_mask, cpus_allowed); 4157again: 4158 retval = set_cpus_allowed_ptr(p, new_mask); 4159 4160 if (!retval) { 4161 cpuset_cpus_allowed(p, cpus_allowed); 4162 if (!cpumask_subset(new_mask, cpus_allowed)) { 4163 /* 4164 * We must have raced with a concurrent cpuset 4165 * update. Just reset the cpus_allowed to the 4166 * cpuset's cpus_allowed 4167 */ 4168 cpumask_copy(new_mask, cpus_allowed); 4169 goto again; 4170 } 4171 } 4172out_unlock: 4173 free_cpumask_var(new_mask); 4174out_free_cpus_allowed: 4175 free_cpumask_var(cpus_allowed); 4176out_put_task: 4177 put_task_struct(p); 4178 put_online_cpus(); 4179 return retval; 4180} 4181 4182static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len, 4183 struct cpumask *new_mask) 4184{ 4185 if (len < cpumask_size()) 4186 cpumask_clear(new_mask); 4187 else if (len > cpumask_size()) 4188 len = cpumask_size(); 4189 4190 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0; 4191} 4192 4193/** 4194 * sys_sched_setaffinity - set the cpu affinity of a process 4195 * @pid: pid of the process 4196 * @len: length in bytes of the bitmask pointed to by user_mask_ptr 4197 * @user_mask_ptr: user-space pointer to the new cpu mask 4198 */ 4199SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len, 4200 unsigned long __user *, user_mask_ptr) 4201{ 4202 cpumask_var_t new_mask; 4203 int retval; 4204 4205 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) 4206 return -ENOMEM; 4207 4208 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask); 4209 if (retval == 0) 4210 retval = sched_setaffinity(pid, new_mask); 4211 free_cpumask_var(new_mask); 4212 return retval; 4213} 4214 4215long sched_getaffinity(pid_t pid, struct cpumask *mask) 4216{ 4217 struct task_struct *p; 4218 unsigned long flags; 4219 int retval; 4220 4221 get_online_cpus(); 4222 rcu_read_lock(); 4223 4224 retval = -ESRCH; 4225 p = find_process_by_pid(pid); 4226 if (!p) 4227 goto out_unlock; 4228 4229 retval = security_task_getscheduler(p); 4230 if (retval) 4231 goto out_unlock; 4232 4233 raw_spin_lock_irqsave(&p->pi_lock, flags); 4234 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask); 4235 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 4236 4237out_unlock: 4238 rcu_read_unlock(); 4239 put_online_cpus(); 4240 4241 return retval; 4242} 4243 4244/** 4245 * sys_sched_getaffinity - get the cpu affinity of a process 4246 * @pid: pid of the process 4247 * @len: length in bytes of the bitmask pointed to by user_mask_ptr 4248 * @user_mask_ptr: user-space pointer to hold the current cpu mask 4249 */ 4250SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len, 4251 unsigned long __user *, user_mask_ptr) 4252{ 4253 int ret; 4254 cpumask_var_t mask; 4255 4256 if ((len * BITS_PER_BYTE) < nr_cpu_ids) 4257 return -EINVAL; 4258 if (len & (sizeof(unsigned long)-1)) 4259 return -EINVAL; 4260 4261 if (!alloc_cpumask_var(&mask, GFP_KERNEL)) 4262 return -ENOMEM; 4263 4264 ret = sched_getaffinity(pid, mask); 4265 if (ret == 0) { 4266 size_t retlen = min_t(size_t, len, cpumask_size()); 4267 4268 if (copy_to_user(user_mask_ptr, mask, retlen)) 4269 ret = -EFAULT; 4270 else 4271 ret = retlen; 4272 } 4273 free_cpumask_var(mask); 4274 4275 return ret; 4276} 4277 4278/** 4279 * sys_sched_yield - yield the current processor to other threads. 4280 * 4281 * This function yields the current CPU to other tasks. If there are no 4282 * other threads running on this CPU then this function will return. 4283 */ 4284SYSCALL_DEFINE0(sched_yield) 4285{ 4286 struct rq *rq = this_rq_lock(); 4287 4288 schedstat_inc(rq, yld_count); 4289 current->sched_class->yield_task(rq); 4290 4291 /* 4292 * Since we are going to call schedule() anyway, there's 4293 * no need to preempt or enable interrupts: 4294 */ 4295 __release(rq->lock); 4296 spin_release(&rq->lock.dep_map, 1, _THIS_IP_); 4297 do_raw_spin_unlock(&rq->lock); 4298 sched_preempt_enable_no_resched(); 4299 4300 schedule(); 4301 4302 return 0; 4303} 4304 4305static inline int should_resched(void) 4306{ 4307 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE); 4308} 4309 4310static void __cond_resched(void) 4311{ 4312 add_preempt_count(PREEMPT_ACTIVE); 4313 __schedule(); 4314 sub_preempt_count(PREEMPT_ACTIVE); 4315} 4316 4317int __sched _cond_resched(void) 4318{ 4319 if (should_resched()) { 4320 __cond_resched(); 4321 return 1; 4322 } 4323 return 0; 4324} 4325EXPORT_SYMBOL(_cond_resched); 4326 4327/* 4328 * __cond_resched_lock() - if a reschedule is pending, drop the given lock, 4329 * call schedule, and on return reacquire the lock. 4330 * 4331 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level 4332 * operations here to prevent schedule() from being called twice (once via 4333 * spin_unlock(), once by hand). 4334 */ 4335int __cond_resched_lock(spinlock_t *lock) 4336{ 4337 int resched = should_resched(); 4338 int ret = 0; 4339 4340 lockdep_assert_held(lock); 4341 4342 if (spin_needbreak(lock) || resched) { 4343 spin_unlock(lock); 4344 if (resched) 4345 __cond_resched(); 4346 else 4347 cpu_relax(); 4348 ret = 1; 4349 spin_lock(lock); 4350 } 4351 return ret; 4352} 4353EXPORT_SYMBOL(__cond_resched_lock); 4354 4355int __sched __cond_resched_softirq(void) 4356{ 4357 BUG_ON(!in_softirq()); 4358 4359 if (should_resched()) { 4360 local_bh_enable(); 4361 __cond_resched(); 4362 local_bh_disable(); 4363 return 1; 4364 } 4365 return 0; 4366} 4367EXPORT_SYMBOL(__cond_resched_softirq); 4368 4369/** 4370 * yield - yield the current processor to other threads. 4371 * 4372 * Do not ever use this function, there's a 99% chance you're doing it wrong. 4373 * 4374 * The scheduler is at all times free to pick the calling task as the most 4375 * eligible task to run, if removing the yield() call from your code breaks 4376 * it, its already broken. 4377 * 4378 * Typical broken usage is: 4379 * 4380 * while (!event) 4381 * yield(); 4382 * 4383 * where one assumes that yield() will let 'the other' process run that will 4384 * make event true. If the current task is a SCHED_FIFO task that will never 4385 * happen. Never use yield() as a progress guarantee!! 4386 * 4387 * If you want to use yield() to wait for something, use wait_event(). 4388 * If you want to use yield() to be 'nice' for others, use cond_resched(). 4389 * If you still want to use yield(), do not! 4390 */ 4391void __sched yield(void) 4392{ 4393 set_current_state(TASK_RUNNING); 4394 sys_sched_yield(); 4395} 4396EXPORT_SYMBOL(yield); 4397 4398/** 4399 * yield_to - yield the current processor to another thread in 4400 * your thread group, or accelerate that thread toward the 4401 * processor it's on. 4402 * @p: target task 4403 * @preempt: whether task preemption is allowed or not 4404 * 4405 * It's the caller's job to ensure that the target task struct 4406 * can't go away on us before we can do any checks. 4407 * 4408 * Returns: 4409 * true (>0) if we indeed boosted the target task. 4410 * false (0) if we failed to boost the target. 4411 * -ESRCH if there's no task to yield to. 4412 */ 4413bool __sched yield_to(struct task_struct *p, bool preempt) 4414{ 4415 struct task_struct *curr = current; 4416 struct rq *rq, *p_rq; 4417 unsigned long flags; 4418 int yielded = 0; 4419 4420 local_irq_save(flags); 4421 rq = this_rq(); 4422 4423again: 4424 p_rq = task_rq(p); 4425 /* 4426 * If we're the only runnable task on the rq and target rq also 4427 * has only one task, there's absolutely no point in yielding. 4428 */ 4429 if (rq->nr_running == 1 && p_rq->nr_running == 1) { 4430 yielded = -ESRCH; 4431 goto out_irq; 4432 } 4433 4434 double_rq_lock(rq, p_rq); 4435 while (task_rq(p) != p_rq) { 4436 double_rq_unlock(rq, p_rq); 4437 goto again; 4438 } 4439 4440 if (!curr->sched_class->yield_to_task) 4441 goto out_unlock; 4442 4443 if (curr->sched_class != p->sched_class) 4444 goto out_unlock; 4445 4446 if (task_running(p_rq, p) || p->state) 4447 goto out_unlock; 4448 4449 yielded = curr->sched_class->yield_to_task(rq, p, preempt); 4450 if (yielded) { 4451 schedstat_inc(rq, yld_count); 4452 /* 4453 * Make p's CPU reschedule; pick_next_entity takes care of 4454 * fairness. 4455 */ 4456 if (preempt && rq != p_rq) 4457 resched_task(p_rq->curr); 4458 } 4459 4460out_unlock: 4461 double_rq_unlock(rq, p_rq); 4462out_irq: 4463 local_irq_restore(flags); 4464 4465 if (yielded > 0) 4466 schedule(); 4467 4468 return yielded; 4469} 4470EXPORT_SYMBOL_GPL(yield_to); 4471 4472/* 4473 * This task is about to go to sleep on IO. Increment rq->nr_iowait so 4474 * that process accounting knows that this is a task in IO wait state. 4475 */ 4476void __sched io_schedule(void) 4477{ 4478 struct rq *rq = raw_rq(); 4479 4480 delayacct_blkio_start(); 4481 atomic_inc(&rq->nr_iowait); 4482 blk_flush_plug(current); 4483 current->in_iowait = 1; 4484 schedule(); 4485 current->in_iowait = 0; 4486 atomic_dec(&rq->nr_iowait); 4487 delayacct_blkio_end(); 4488} 4489EXPORT_SYMBOL(io_schedule); 4490 4491long __sched io_schedule_timeout(long timeout) 4492{ 4493 struct rq *rq = raw_rq(); 4494 long ret; 4495 4496 delayacct_blkio_start(); 4497 atomic_inc(&rq->nr_iowait); 4498 blk_flush_plug(current); 4499 current->in_iowait = 1; 4500 ret = schedule_timeout(timeout); 4501 current->in_iowait = 0; 4502 atomic_dec(&rq->nr_iowait); 4503 delayacct_blkio_end(); 4504 return ret; 4505} 4506 4507/** 4508 * sys_sched_get_priority_max - return maximum RT priority. 4509 * @policy: scheduling class. 4510 * 4511 * this syscall returns the maximum rt_priority that can be used 4512 * by a given scheduling class. 4513 */ 4514SYSCALL_DEFINE1(sched_get_priority_max, int, policy) 4515{ 4516 int ret = -EINVAL; 4517 4518 switch (policy) { 4519 case SCHED_FIFO: 4520 case SCHED_RR: 4521 ret = MAX_USER_RT_PRIO-1; 4522 break; 4523 case SCHED_NORMAL: 4524 case SCHED_BATCH: 4525 case SCHED_IDLE: 4526 ret = 0; 4527 break; 4528 } 4529 return ret; 4530} 4531 4532/** 4533 * sys_sched_get_priority_min - return minimum RT priority. 4534 * @policy: scheduling class. 4535 * 4536 * this syscall returns the minimum rt_priority that can be used 4537 * by a given scheduling class. 4538 */ 4539SYSCALL_DEFINE1(sched_get_priority_min, int, policy) 4540{ 4541 int ret = -EINVAL; 4542 4543 switch (policy) { 4544 case SCHED_FIFO: 4545 case SCHED_RR: 4546 ret = 1; 4547 break; 4548 case SCHED_NORMAL: 4549 case SCHED_BATCH: 4550 case SCHED_IDLE: 4551 ret = 0; 4552 } 4553 return ret; 4554} 4555 4556/** 4557 * sys_sched_rr_get_interval - return the default timeslice of a process. 4558 * @pid: pid of the process. 4559 * @interval: userspace pointer to the timeslice value. 4560 * 4561 * this syscall writes the default timeslice value of a given process 4562 * into the user-space timespec buffer. A value of '0' means infinity. 4563 */ 4564SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid, 4565 struct timespec __user *, interval) 4566{ 4567 struct task_struct *p; 4568 unsigned int time_slice; 4569 unsigned long flags; 4570 struct rq *rq; 4571 int retval; 4572 struct timespec t; 4573 4574 if (pid < 0) 4575 return -EINVAL; 4576 4577 retval = -ESRCH; 4578 rcu_read_lock(); 4579 p = find_process_by_pid(pid); 4580 if (!p) 4581 goto out_unlock; 4582 4583 retval = security_task_getscheduler(p); 4584 if (retval) 4585 goto out_unlock; 4586 4587 rq = task_rq_lock(p, &flags); 4588 time_slice = p->sched_class->get_rr_interval(rq, p); 4589 task_rq_unlock(rq, p, &flags); 4590 4591 rcu_read_unlock(); 4592 jiffies_to_timespec(time_slice, &t); 4593 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0; 4594 return retval; 4595 4596out_unlock: 4597 rcu_read_unlock(); 4598 return retval; 4599} 4600 4601static const char stat_nam[] = TASK_STATE_TO_CHAR_STR; 4602 4603void sched_show_task(struct task_struct *p) 4604{ 4605 unsigned long free = 0; 4606 int ppid; 4607 unsigned state; 4608 4609 state = p->state ? __ffs(p->state) + 1 : 0; 4610 printk(KERN_INFO "%-15.15s %c", p->comm, 4611 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?'); 4612#if BITS_PER_LONG == 32 4613 if (state == TASK_RUNNING) 4614 printk(KERN_CONT " running "); 4615 else 4616 printk(KERN_CONT " %08lx ", thread_saved_pc(p)); 4617#else 4618 if (state == TASK_RUNNING) 4619 printk(KERN_CONT " running task "); 4620 else 4621 printk(KERN_CONT " %016lx ", thread_saved_pc(p)); 4622#endif 4623#ifdef CONFIG_DEBUG_STACK_USAGE 4624 free = stack_not_used(p); 4625#endif 4626 rcu_read_lock(); 4627 ppid = task_pid_nr(rcu_dereference(p->real_parent)); 4628 rcu_read_unlock(); 4629 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free, 4630 task_pid_nr(p), ppid, 4631 (unsigned long)task_thread_info(p)->flags); 4632 4633 show_stack(p, NULL); 4634} 4635 4636void show_state_filter(unsigned long state_filter) 4637{ 4638 struct task_struct *g, *p; 4639 4640#if BITS_PER_LONG == 32 4641 printk(KERN_INFO 4642 " task PC stack pid father\n"); 4643#else 4644 printk(KERN_INFO 4645 " task PC stack pid father\n"); 4646#endif 4647 rcu_read_lock(); 4648 do_each_thread(g, p) { 4649 /* 4650 * reset the NMI-timeout, listing all files on a slow 4651 * console might take a lot of time: 4652 */ 4653 touch_nmi_watchdog(); 4654 if (!state_filter || (p->state & state_filter)) 4655 sched_show_task(p); 4656 } while_each_thread(g, p); 4657 4658 touch_all_softlockup_watchdogs(); 4659 4660#ifdef CONFIG_SCHED_DEBUG 4661 sysrq_sched_debug_show(); 4662#endif 4663 rcu_read_unlock(); 4664 /* 4665 * Only show locks if all tasks are dumped: 4666 */ 4667 if (!state_filter) 4668 debug_show_all_locks(); 4669} 4670 4671void __cpuinit init_idle_bootup_task(struct task_struct *idle) 4672{ 4673 idle->sched_class = &idle_sched_class; 4674} 4675 4676/** 4677 * init_idle - set up an idle thread for a given CPU 4678 * @idle: task in question 4679 * @cpu: cpu the idle task belongs to 4680 * 4681 * NOTE: this function does not set the idle thread's NEED_RESCHED 4682 * flag, to make booting more robust. 4683 */ 4684void __cpuinit init_idle(struct task_struct *idle, int cpu) 4685{ 4686 struct rq *rq = cpu_rq(cpu); 4687 unsigned long flags; 4688 4689 raw_spin_lock_irqsave(&rq->lock, flags); 4690 4691 __sched_fork(idle); 4692 idle->state = TASK_RUNNING; 4693 idle->se.exec_start = sched_clock(); 4694 4695 do_set_cpus_allowed(idle, cpumask_of(cpu)); 4696 /* 4697 * We're having a chicken and egg problem, even though we are 4698 * holding rq->lock, the cpu isn't yet set to this cpu so the 4699 * lockdep check in task_group() will fail. 4700 * 4701 * Similar case to sched_fork(). / Alternatively we could 4702 * use task_rq_lock() here and obtain the other rq->lock. 4703 * 4704 * Silence PROVE_RCU 4705 */ 4706 rcu_read_lock(); 4707 __set_task_cpu(idle, cpu); 4708 rcu_read_unlock(); 4709 4710 rq->curr = rq->idle = idle; 4711#if defined(CONFIG_SMP) 4712 idle->on_cpu = 1; 4713#endif 4714 raw_spin_unlock_irqrestore(&rq->lock, flags); 4715 4716 /* Set the preempt count _outside_ the spinlocks! */ 4717 task_thread_info(idle)->preempt_count = 0; 4718 4719 /* 4720 * The idle tasks have their own, simple scheduling class: 4721 */ 4722 idle->sched_class = &idle_sched_class; 4723 ftrace_graph_init_idle_task(idle, cpu); 4724 vtime_init_idle(idle); 4725#if defined(CONFIG_SMP) 4726 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu); 4727#endif 4728} 4729 4730#ifdef CONFIG_SMP 4731void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask) 4732{ 4733 if (p->sched_class && p->sched_class->set_cpus_allowed) 4734 p->sched_class->set_cpus_allowed(p, new_mask); 4735 4736 cpumask_copy(&p->cpus_allowed, new_mask); 4737 p->nr_cpus_allowed = cpumask_weight(new_mask); 4738} 4739 4740/* 4741 * This is how migration works: 4742 * 4743 * 1) we invoke migration_cpu_stop() on the target CPU using 4744 * stop_one_cpu(). 4745 * 2) stopper starts to run (implicitly forcing the migrated thread 4746 * off the CPU) 4747 * 3) it checks whether the migrated task is still in the wrong runqueue. 4748 * 4) if it's in the wrong runqueue then the migration thread removes 4749 * it and puts it into the right queue. 4750 * 5) stopper completes and stop_one_cpu() returns and the migration 4751 * is done. 4752 */ 4753 4754/* 4755 * Change a given task's CPU affinity. Migrate the thread to a 4756 * proper CPU and schedule it away if the CPU it's executing on 4757 * is removed from the allowed bitmask. 4758 * 4759 * NOTE: the caller must have a valid reference to the task, the 4760 * task must not exit() & deallocate itself prematurely. The 4761 * call is not atomic; no spinlocks may be held. 4762 */ 4763int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask) 4764{ 4765 unsigned long flags; 4766 struct rq *rq; 4767 unsigned int dest_cpu; 4768 int ret = 0; 4769 4770 rq = task_rq_lock(p, &flags); 4771 4772 if (cpumask_equal(&p->cpus_allowed, new_mask)) 4773 goto out; 4774 4775 if (!cpumask_intersects(new_mask, cpu_active_mask)) { 4776 ret = -EINVAL; 4777 goto out; 4778 } 4779 4780 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current)) { 4781 ret = -EINVAL; 4782 goto out; 4783 } 4784 4785 do_set_cpus_allowed(p, new_mask); 4786 4787 /* Can the task run on the task's current CPU? If so, we're done */ 4788 if (cpumask_test_cpu(task_cpu(p), new_mask)) 4789 goto out; 4790 4791 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask); 4792 if (p->on_rq) { 4793 struct migration_arg arg = { p, dest_cpu }; 4794 /* Need help from migration thread: drop lock and wait. */ 4795 task_rq_unlock(rq, p, &flags); 4796 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg); 4797 tlb_migrate_finish(p->mm); 4798 return 0; 4799 } 4800out: 4801 task_rq_unlock(rq, p, &flags); 4802 4803 return ret; 4804} 4805EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr); 4806 4807/* 4808 * Move (not current) task off this cpu, onto dest cpu. We're doing 4809 * this because either it can't run here any more (set_cpus_allowed() 4810 * away from this CPU, or CPU going down), or because we're 4811 * attempting to rebalance this task on exec (sched_exec). 4812 * 4813 * So we race with normal scheduler movements, but that's OK, as long 4814 * as the task is no longer on this CPU. 4815 * 4816 * Returns non-zero if task was successfully migrated. 4817 */ 4818static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu) 4819{ 4820 struct rq *rq_dest, *rq_src; 4821 int ret = 0; 4822 4823 if (unlikely(!cpu_active(dest_cpu))) 4824 return ret; 4825 4826 rq_src = cpu_rq(src_cpu); 4827 rq_dest = cpu_rq(dest_cpu); 4828 4829 raw_spin_lock(&p->pi_lock); 4830 double_rq_lock(rq_src, rq_dest); 4831 /* Already moved. */ 4832 if (task_cpu(p) != src_cpu) 4833 goto done; 4834 /* Affinity changed (again). */ 4835 if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p))) 4836 goto fail; 4837 4838 /* 4839 * If we're not on a rq, the next wake-up will ensure we're 4840 * placed properly. 4841 */ 4842 if (p->on_rq) { 4843 dequeue_task(rq_src, p, 0); 4844 set_task_cpu(p, dest_cpu); 4845 enqueue_task(rq_dest, p, 0); 4846 check_preempt_curr(rq_dest, p, 0); 4847 } 4848done: 4849 ret = 1; 4850fail: 4851 double_rq_unlock(rq_src, rq_dest); 4852 raw_spin_unlock(&p->pi_lock); 4853 return ret; 4854} 4855 4856/* 4857 * migration_cpu_stop - this will be executed by a highprio stopper thread 4858 * and performs thread migration by bumping thread off CPU then 4859 * 'pushing' onto another runqueue. 4860 */ 4861static int migration_cpu_stop(void *data) 4862{ 4863 struct migration_arg *arg = data; 4864 4865 /* 4866 * The original target cpu might have gone down and we might 4867 * be on another cpu but it doesn't matter. 4868 */ 4869 local_irq_disable(); 4870 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu); 4871 local_irq_enable(); 4872 return 0; 4873} 4874 4875#ifdef CONFIG_HOTPLUG_CPU 4876 4877/* 4878 * Ensures that the idle task is using init_mm right before its cpu goes 4879 * offline. 4880 */ 4881void idle_task_exit(void) 4882{ 4883 struct mm_struct *mm = current->active_mm; 4884 4885 BUG_ON(cpu_online(smp_processor_id())); 4886 4887 if (mm != &init_mm) 4888 switch_mm(mm, &init_mm, current); 4889 mmdrop(mm); 4890} 4891 4892/* 4893 * Since this CPU is going 'away' for a while, fold any nr_active delta 4894 * we might have. Assumes we're called after migrate_tasks() so that the 4895 * nr_active count is stable. 4896 * 4897 * Also see the comment "Global load-average calculations". 4898 */ 4899static void calc_load_migrate(struct rq *rq) 4900{ 4901 long delta = calc_load_fold_active(rq); 4902 if (delta) 4903 atomic_long_add(delta, &calc_load_tasks); 4904} 4905 4906/* 4907 * Migrate all tasks from the rq, sleeping tasks will be migrated by 4908 * try_to_wake_up()->select_task_rq(). 4909 * 4910 * Called with rq->lock held even though we'er in stop_machine() and 4911 * there's no concurrency possible, we hold the required locks anyway 4912 * because of lock validation efforts. 4913 */ 4914static void migrate_tasks(unsigned int dead_cpu) 4915{ 4916 struct rq *rq = cpu_rq(dead_cpu); 4917 struct task_struct *next, *stop = rq->stop; 4918 int dest_cpu; 4919 4920 /* 4921 * Fudge the rq selection such that the below task selection loop 4922 * doesn't get stuck on the currently eligible stop task. 4923 * 4924 * We're currently inside stop_machine() and the rq is either stuck 4925 * in the stop_machine_cpu_stop() loop, or we're executing this code, 4926 * either way we should never end up calling schedule() until we're 4927 * done here. 4928 */ 4929 rq->stop = NULL; 4930 4931 for ( ; ; ) { 4932 /* 4933 * There's this thread running, bail when that's the only 4934 * remaining thread. 4935 */ 4936 if (rq->nr_running == 1) 4937 break; 4938 4939 next = pick_next_task(rq); 4940 BUG_ON(!next); 4941 next->sched_class->put_prev_task(rq, next); 4942 4943 /* Find suitable destination for @next, with force if needed. */ 4944 dest_cpu = select_fallback_rq(dead_cpu, next); 4945 raw_spin_unlock(&rq->lock); 4946 4947 __migrate_task(next, dead_cpu, dest_cpu); 4948 4949 raw_spin_lock(&rq->lock); 4950 } 4951 4952 rq->stop = stop; 4953} 4954 4955#endif /* CONFIG_HOTPLUG_CPU */ 4956 4957#if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL) 4958 4959static struct ctl_table sd_ctl_dir[] = { 4960 { 4961 .procname = "sched_domain", 4962 .mode = 0555, 4963 }, 4964 {} 4965}; 4966 4967static struct ctl_table sd_ctl_root[] = { 4968 { 4969 .procname = "kernel", 4970 .mode = 0555, 4971 .child = sd_ctl_dir, 4972 }, 4973 {} 4974}; 4975 4976static struct ctl_table *sd_alloc_ctl_entry(int n) 4977{ 4978 struct ctl_table *entry = 4979 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL); 4980 4981 return entry; 4982} 4983 4984static void sd_free_ctl_entry(struct ctl_table **tablep) 4985{ 4986 struct ctl_table *entry; 4987 4988 /* 4989 * In the intermediate directories, both the child directory and 4990 * procname are dynamically allocated and could fail but the mode 4991 * will always be set. In the lowest directory the names are 4992 * static strings and all have proc handlers. 4993 */ 4994 for (entry = *tablep; entry->mode; entry++) { 4995 if (entry->child) 4996 sd_free_ctl_entry(&entry->child); 4997 if (entry->proc_handler == NULL) 4998 kfree(entry->procname); 4999 } 5000 5001 kfree(*tablep); 5002 *tablep = NULL; 5003} 5004 5005static int min_load_idx = 0; 5006static int max_load_idx = CPU_LOAD_IDX_MAX; 5007 5008static void 5009set_table_entry(struct ctl_table *entry, 5010 const char *procname, void *data, int maxlen, 5011 umode_t mode, proc_handler *proc_handler, 5012 bool load_idx) 5013{ 5014 entry->procname = procname; 5015 entry->data = data; 5016 entry->maxlen = maxlen; 5017 entry->mode = mode; 5018 entry->proc_handler = proc_handler; 5019 5020 if (load_idx) { 5021 entry->extra1 = &min_load_idx; 5022 entry->extra2 = &max_load_idx; 5023 } 5024} 5025 5026static struct ctl_table * 5027sd_alloc_ctl_domain_table(struct sched_domain *sd) 5028{ 5029 struct ctl_table *table = sd_alloc_ctl_entry(13); 5030 5031 if (table == NULL) 5032 return NULL; 5033 5034 set_table_entry(&table[0], "min_interval", &sd->min_interval, 5035 sizeof(long), 0644, proc_doulongvec_minmax, false); 5036 set_table_entry(&table[1], "max_interval", &sd->max_interval, 5037 sizeof(long), 0644, proc_doulongvec_minmax, false); 5038 set_table_entry(&table[2], "busy_idx", &sd->busy_idx, 5039 sizeof(int), 0644, proc_dointvec_minmax, true); 5040 set_table_entry(&table[3], "idle_idx", &sd->idle_idx, 5041 sizeof(int), 0644, proc_dointvec_minmax, true); 5042 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx, 5043 sizeof(int), 0644, proc_dointvec_minmax, true); 5044 set_table_entry(&table[5], "wake_idx", &sd->wake_idx, 5045 sizeof(int), 0644, proc_dointvec_minmax, true); 5046 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx, 5047 sizeof(int), 0644, proc_dointvec_minmax, true); 5048 set_table_entry(&table[7], "busy_factor", &sd->busy_factor, 5049 sizeof(int), 0644, proc_dointvec_minmax, false); 5050 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct, 5051 sizeof(int), 0644, proc_dointvec_minmax, false); 5052 set_table_entry(&table[9], "cache_nice_tries", 5053 &sd->cache_nice_tries, 5054 sizeof(int), 0644, proc_dointvec_minmax, false); 5055 set_table_entry(&table[10], "flags", &sd->flags, 5056 sizeof(int), 0644, proc_dointvec_minmax, false); 5057 set_table_entry(&table[11], "name", sd->name, 5058 CORENAME_MAX_SIZE, 0444, proc_dostring, false); 5059 /* &table[12] is terminator */ 5060 5061 return table; 5062} 5063 5064static ctl_table *sd_alloc_ctl_cpu_table(int cpu) 5065{ 5066 struct ctl_table *entry, *table; 5067 struct sched_domain *sd; 5068 int domain_num = 0, i; 5069 char buf[32]; 5070 5071 for_each_domain(cpu, sd) 5072 domain_num++; 5073 entry = table = sd_alloc_ctl_entry(domain_num + 1); 5074 if (table == NULL) 5075 return NULL; 5076 5077 i = 0; 5078 for_each_domain(cpu, sd) { 5079 snprintf(buf, 32, "domain%d", i); 5080 entry->procname = kstrdup(buf, GFP_KERNEL); 5081 entry->mode = 0555; 5082 entry->child = sd_alloc_ctl_domain_table(sd); 5083 entry++; 5084 i++; 5085 } 5086 return table; 5087} 5088 5089static struct ctl_table_header *sd_sysctl_header; 5090static void register_sched_domain_sysctl(void) 5091{ 5092 int i, cpu_num = num_possible_cpus(); 5093 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1); 5094 char buf[32]; 5095 5096 WARN_ON(sd_ctl_dir[0].child); 5097 sd_ctl_dir[0].child = entry; 5098 5099 if (entry == NULL) 5100 return; 5101 5102 for_each_possible_cpu(i) { 5103 snprintf(buf, 32, "cpu%d", i); 5104 entry->procname = kstrdup(buf, GFP_KERNEL); 5105 entry->mode = 0555; 5106 entry->child = sd_alloc_ctl_cpu_table(i); 5107 entry++; 5108 } 5109 5110 WARN_ON(sd_sysctl_header); 5111 sd_sysctl_header = register_sysctl_table(sd_ctl_root); 5112} 5113 5114/* may be called multiple times per register */ 5115static void unregister_sched_domain_sysctl(void) 5116{ 5117 if (sd_sysctl_header) 5118 unregister_sysctl_table(sd_sysctl_header); 5119 sd_sysctl_header = NULL; 5120 if (sd_ctl_dir[0].child) 5121 sd_free_ctl_entry(&sd_ctl_dir[0].child); 5122} 5123#else 5124static void register_sched_domain_sysctl(void) 5125{ 5126} 5127static void unregister_sched_domain_sysctl(void) 5128{ 5129} 5130#endif 5131 5132static void set_rq_online(struct rq *rq) 5133{ 5134 if (!rq->online) { 5135 const struct sched_class *class; 5136 5137 cpumask_set_cpu(rq->cpu, rq->rd->online); 5138 rq->online = 1; 5139 5140 for_each_class(class) { 5141 if (class->rq_online) 5142 class->rq_online(rq); 5143 } 5144 } 5145} 5146 5147static void set_rq_offline(struct rq *rq) 5148{ 5149 if (rq->online) { 5150 const struct sched_class *class; 5151 5152 for_each_class(class) { 5153 if (class->rq_offline) 5154 class->rq_offline(rq); 5155 } 5156 5157 cpumask_clear_cpu(rq->cpu, rq->rd->online); 5158 rq->online = 0; 5159 } 5160} 5161 5162/* 5163 * migration_call - callback that gets triggered when a CPU is added. 5164 * Here we can start up the necessary migration thread for the new CPU. 5165 */ 5166static int __cpuinit 5167migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu) 5168{ 5169 int cpu = (long)hcpu; 5170 unsigned long flags; 5171 struct rq *rq = cpu_rq(cpu); 5172 5173 switch (action & ~CPU_TASKS_FROZEN) { 5174 5175 case CPU_UP_PREPARE: 5176 rq->calc_load_update = calc_load_update; 5177 break; 5178 5179 case CPU_ONLINE: 5180 /* Update our root-domain */ 5181 raw_spin_lock_irqsave(&rq->lock, flags); 5182 if (rq->rd) { 5183 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span)); 5184 5185 set_rq_online(rq); 5186 } 5187 raw_spin_unlock_irqrestore(&rq->lock, flags); 5188 break; 5189 5190#ifdef CONFIG_HOTPLUG_CPU 5191 case CPU_DYING: 5192 sched_ttwu_pending(); 5193 /* Update our root-domain */ 5194 raw_spin_lock_irqsave(&rq->lock, flags); 5195 if (rq->rd) { 5196 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span)); 5197 set_rq_offline(rq); 5198 } 5199 migrate_tasks(cpu); 5200 BUG_ON(rq->nr_running != 1); /* the migration thread */ 5201 raw_spin_unlock_irqrestore(&rq->lock, flags); 5202 break; 5203 5204 case CPU_DEAD: 5205 calc_load_migrate(rq); 5206 break; 5207#endif 5208 } 5209 5210 update_max_interval(); 5211 5212 return NOTIFY_OK; 5213} 5214 5215/* 5216 * Register at high priority so that task migration (migrate_all_tasks) 5217 * happens before everything else. This has to be lower priority than 5218 * the notifier in the perf_event subsystem, though. 5219 */ 5220static struct notifier_block __cpuinitdata migration_notifier = { 5221 .notifier_call = migration_call, 5222 .priority = CPU_PRI_MIGRATION, 5223}; 5224 5225static int __cpuinit sched_cpu_active(struct notifier_block *nfb, 5226 unsigned long action, void *hcpu) 5227{ 5228 switch (action & ~CPU_TASKS_FROZEN) { 5229 case CPU_STARTING: 5230 case CPU_DOWN_FAILED: 5231 set_cpu_active((long)hcpu, true); 5232 return NOTIFY_OK; 5233 default: 5234 return NOTIFY_DONE; 5235 } 5236} 5237 5238static int __cpuinit sched_cpu_inactive(struct notifier_block *nfb, 5239 unsigned long action, void *hcpu) 5240{ 5241 switch (action & ~CPU_TASKS_FROZEN) { 5242 case CPU_DOWN_PREPARE: 5243 set_cpu_active((long)hcpu, false); 5244 return NOTIFY_OK; 5245 default: 5246 return NOTIFY_DONE; 5247 } 5248} 5249 5250static int __init migration_init(void) 5251{ 5252 void *cpu = (void *)(long)smp_processor_id(); 5253 int err; 5254 5255 /* Initialize migration for the boot CPU */ 5256 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu); 5257 BUG_ON(err == NOTIFY_BAD); 5258 migration_call(&migration_notifier, CPU_ONLINE, cpu); 5259 register_cpu_notifier(&migration_notifier); 5260 5261 /* Register cpu active notifiers */ 5262 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE); 5263 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE); 5264 5265 return 0; 5266} 5267early_initcall(migration_init); 5268#endif 5269 5270#ifdef CONFIG_SMP 5271 5272static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */ 5273 5274#ifdef CONFIG_SCHED_DEBUG 5275 5276static __read_mostly int sched_debug_enabled; 5277 5278static int __init sched_debug_setup(char *str) 5279{ 5280 sched_debug_enabled = 1; 5281 5282 return 0; 5283} 5284early_param("sched_debug", sched_debug_setup); 5285 5286static inline bool sched_debug(void) 5287{ 5288 return sched_debug_enabled; 5289} 5290 5291static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level, 5292 struct cpumask *groupmask) 5293{ 5294 struct sched_group *group = sd->groups; 5295 char str[256]; 5296 5297 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd)); 5298 cpumask_clear(groupmask); 5299 5300 printk(KERN_DEBUG "%*s domain %d: ", level, "", level); 5301 5302 if (!(sd->flags & SD_LOAD_BALANCE)) { 5303 printk("does not load-balance\n"); 5304 if (sd->parent) 5305 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain" 5306 " has parent"); 5307 return -1; 5308 } 5309 5310 printk(KERN_CONT "span %s level %s\n", str, sd->name); 5311 5312 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) { 5313 printk(KERN_ERR "ERROR: domain->span does not contain " 5314 "CPU%d\n", cpu); 5315 } 5316 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) { 5317 printk(KERN_ERR "ERROR: domain->groups does not contain" 5318 " CPU%d\n", cpu); 5319 } 5320 5321 printk(KERN_DEBUG "%*s groups:", level + 1, ""); 5322 do { 5323 if (!group) { 5324 printk("\n"); 5325 printk(KERN_ERR "ERROR: group is NULL\n"); 5326 break; 5327 } 5328 5329 /* 5330 * Even though we initialize ->power to something semi-sane, 5331 * we leave power_orig unset. This allows us to detect if 5332 * domain iteration is still funny without causing /0 traps. 5333 */ 5334 if (!group->sgp->power_orig) { 5335 printk(KERN_CONT "\n"); 5336 printk(KERN_ERR "ERROR: domain->cpu_power not " 5337 "set\n"); 5338 break; 5339 } 5340 5341 if (!cpumask_weight(sched_group_cpus(group))) { 5342 printk(KERN_CONT "\n"); 5343 printk(KERN_ERR "ERROR: empty group\n"); 5344 break; 5345 } 5346 5347 if (!(sd->flags & SD_OVERLAP) && 5348 cpumask_intersects(groupmask, sched_group_cpus(group))) { 5349 printk(KERN_CONT "\n"); 5350 printk(KERN_ERR "ERROR: repeated CPUs\n"); 5351 break; 5352 } 5353 5354 cpumask_or(groupmask, groupmask, sched_group_cpus(group)); 5355 5356 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group)); 5357 5358 printk(KERN_CONT " %s", str); 5359 if (group->sgp->power != SCHED_POWER_SCALE) { 5360 printk(KERN_CONT " (cpu_power = %d)", 5361 group->sgp->power); 5362 } 5363 5364 group = group->next; 5365 } while (group != sd->groups); 5366 printk(KERN_CONT "\n"); 5367 5368 if (!cpumask_equal(sched_domain_span(sd), groupmask)) 5369 printk(KERN_ERR "ERROR: groups don't span domain->span\n"); 5370 5371 if (sd->parent && 5372 !cpumask_subset(groupmask, sched_domain_span(sd->parent))) 5373 printk(KERN_ERR "ERROR: parent span is not a superset " 5374 "of domain->span\n"); 5375 return 0; 5376} 5377 5378static void sched_domain_debug(struct sched_domain *sd, int cpu) 5379{ 5380 int level = 0; 5381 5382 if (!sched_debug_enabled) 5383 return; 5384 5385 if (!sd) { 5386 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu); 5387 return; 5388 } 5389 5390 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu); 5391 5392 for (;;) { 5393 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask)) 5394 break; 5395 level++; 5396 sd = sd->parent; 5397 if (!sd) 5398 break; 5399 } 5400} 5401#else /* !CONFIG_SCHED_DEBUG */ 5402# define sched_domain_debug(sd, cpu) do { } while (0) 5403static inline bool sched_debug(void) 5404{ 5405 return false; 5406} 5407#endif /* CONFIG_SCHED_DEBUG */ 5408 5409static int sd_degenerate(struct sched_domain *sd) 5410{ 5411 if (cpumask_weight(sched_domain_span(sd)) == 1) 5412 return 1; 5413 5414 /* Following flags need at least 2 groups */ 5415 if (sd->flags & (SD_LOAD_BALANCE | 5416 SD_BALANCE_NEWIDLE | 5417 SD_BALANCE_FORK | 5418 SD_BALANCE_EXEC | 5419 SD_SHARE_CPUPOWER | 5420 SD_SHARE_PKG_RESOURCES)) { 5421 if (sd->groups != sd->groups->next) 5422 return 0; 5423 } 5424 5425 /* Following flags don't use groups */ 5426 if (sd->flags & (SD_WAKE_AFFINE)) 5427 return 0; 5428 5429 return 1; 5430} 5431 5432static int 5433sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent) 5434{ 5435 unsigned long cflags = sd->flags, pflags = parent->flags; 5436 5437 if (sd_degenerate(parent)) 5438 return 1; 5439 5440 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent))) 5441 return 0; 5442 5443 /* Flags needing groups don't count if only 1 group in parent */ 5444 if (parent->groups == parent->groups->next) { 5445 pflags &= ~(SD_LOAD_BALANCE | 5446 SD_BALANCE_NEWIDLE | 5447 SD_BALANCE_FORK | 5448 SD_BALANCE_EXEC | 5449 SD_SHARE_CPUPOWER | 5450 SD_SHARE_PKG_RESOURCES); 5451 if (nr_node_ids == 1) 5452 pflags &= ~SD_SERIALIZE; 5453 } 5454 if (~cflags & pflags) 5455 return 0; 5456 5457 return 1; 5458} 5459 5460static void free_rootdomain(struct rcu_head *rcu) 5461{ 5462 struct root_domain *rd = container_of(rcu, struct root_domain, rcu); 5463 5464 cpupri_cleanup(&rd->cpupri); 5465 free_cpumask_var(rd->rto_mask); 5466 free_cpumask_var(rd->online); 5467 free_cpumask_var(rd->span); 5468 kfree(rd); 5469} 5470 5471static void rq_attach_root(struct rq *rq, struct root_domain *rd) 5472{ 5473 struct root_domain *old_rd = NULL; 5474 unsigned long flags; 5475 5476 raw_spin_lock_irqsave(&rq->lock, flags); 5477 5478 if (rq->rd) { 5479 old_rd = rq->rd; 5480 5481 if (cpumask_test_cpu(rq->cpu, old_rd->online)) 5482 set_rq_offline(rq); 5483 5484 cpumask_clear_cpu(rq->cpu, old_rd->span); 5485 5486 /* 5487 * If we dont want to free the old_rt yet then 5488 * set old_rd to NULL to skip the freeing later 5489 * in this function: 5490 */ 5491 if (!atomic_dec_and_test(&old_rd->refcount)) 5492 old_rd = NULL; 5493 } 5494 5495 atomic_inc(&rd->refcount); 5496 rq->rd = rd; 5497 5498 cpumask_set_cpu(rq->cpu, rd->span); 5499 if (cpumask_test_cpu(rq->cpu, cpu_active_mask)) 5500 set_rq_online(rq); 5501 5502 raw_spin_unlock_irqrestore(&rq->lock, flags); 5503 5504 if (old_rd) 5505 call_rcu_sched(&old_rd->rcu, free_rootdomain); 5506} 5507 5508static int init_rootdomain(struct root_domain *rd) 5509{ 5510 memset(rd, 0, sizeof(*rd)); 5511 5512 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL)) 5513 goto out; 5514 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL)) 5515 goto free_span; 5516 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL)) 5517 goto free_online; 5518 5519 if (cpupri_init(&rd->cpupri) != 0) 5520 goto free_rto_mask; 5521 return 0; 5522 5523free_rto_mask: 5524 free_cpumask_var(rd->rto_mask); 5525free_online: 5526 free_cpumask_var(rd->online); 5527free_span: 5528 free_cpumask_var(rd->span); 5529out: 5530 return -ENOMEM; 5531} 5532 5533/* 5534 * By default the system creates a single root-domain with all cpus as 5535 * members (mimicking the global state we have today). 5536 */ 5537struct root_domain def_root_domain; 5538 5539static void init_defrootdomain(void) 5540{ 5541 init_rootdomain(&def_root_domain); 5542 5543 atomic_set(&def_root_domain.refcount, 1); 5544} 5545 5546static struct root_domain *alloc_rootdomain(void) 5547{ 5548 struct root_domain *rd; 5549 5550 rd = kmalloc(sizeof(*rd), GFP_KERNEL); 5551 if (!rd) 5552 return NULL; 5553 5554 if (init_rootdomain(rd) != 0) { 5555 kfree(rd); 5556 return NULL; 5557 } 5558 5559 return rd; 5560} 5561 5562static void free_sched_groups(struct sched_group *sg, int free_sgp) 5563{ 5564 struct sched_group *tmp, *first; 5565 5566 if (!sg) 5567 return; 5568 5569 first = sg; 5570 do { 5571 tmp = sg->next; 5572 5573 if (free_sgp && atomic_dec_and_test(&sg->sgp->ref)) 5574 kfree(sg->sgp); 5575 5576 kfree(sg); 5577 sg = tmp; 5578 } while (sg != first); 5579} 5580 5581static void free_sched_domain(struct rcu_head *rcu) 5582{ 5583 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu); 5584 5585 /* 5586 * If its an overlapping domain it has private groups, iterate and 5587 * nuke them all. 5588 */ 5589 if (sd->flags & SD_OVERLAP) { 5590 free_sched_groups(sd->groups, 1); 5591 } else if (atomic_dec_and_test(&sd->groups->ref)) { 5592 kfree(sd->groups->sgp); 5593 kfree(sd->groups); 5594 } 5595 kfree(sd); 5596} 5597 5598static void destroy_sched_domain(struct sched_domain *sd, int cpu) 5599{ 5600 call_rcu(&sd->rcu, free_sched_domain); 5601} 5602 5603static void destroy_sched_domains(struct sched_domain *sd, int cpu) 5604{ 5605 for (; sd; sd = sd->parent) 5606 destroy_sched_domain(sd, cpu); 5607} 5608 5609/* 5610 * Keep a special pointer to the highest sched_domain that has 5611 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this 5612 * allows us to avoid some pointer chasing select_idle_sibling(). 5613 * 5614 * Also keep a unique ID per domain (we use the first cpu number in 5615 * the cpumask of the domain), this allows us to quickly tell if 5616 * two cpus are in the same cache domain, see cpus_share_cache(). 5617 */ 5618DEFINE_PER_CPU(struct sched_domain *, sd_llc); 5619DEFINE_PER_CPU(int, sd_llc_id); 5620 5621static void update_top_cache_domain(int cpu) 5622{ 5623 struct sched_domain *sd; 5624 int id = cpu; 5625 5626 sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES); 5627 if (sd) 5628 id = cpumask_first(sched_domain_span(sd)); 5629 5630 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd); 5631 per_cpu(sd_llc_id, cpu) = id; 5632} 5633 5634/* 5635 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must 5636 * hold the hotplug lock. 5637 */ 5638static void 5639cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu) 5640{ 5641 struct rq *rq = cpu_rq(cpu); 5642 struct sched_domain *tmp; 5643 5644 /* Remove the sched domains which do not contribute to scheduling. */ 5645 for (tmp = sd; tmp; ) { 5646 struct sched_domain *parent = tmp->parent; 5647 if (!parent) 5648 break; 5649 5650 if (sd_parent_degenerate(tmp, parent)) { 5651 tmp->parent = parent->parent; 5652 if (parent->parent) 5653 parent->parent->child = tmp; 5654 destroy_sched_domain(parent, cpu); 5655 } else 5656 tmp = tmp->parent; 5657 } 5658 5659 if (sd && sd_degenerate(sd)) { 5660 tmp = sd; 5661 sd = sd->parent; 5662 destroy_sched_domain(tmp, cpu); 5663 if (sd) 5664 sd->child = NULL; 5665 } 5666 5667 sched_domain_debug(sd, cpu); 5668 5669 rq_attach_root(rq, rd); 5670 tmp = rq->sd; 5671 rcu_assign_pointer(rq->sd, sd); 5672 destroy_sched_domains(tmp, cpu); 5673 5674 update_top_cache_domain(cpu); 5675} 5676 5677/* cpus with isolated domains */ 5678static cpumask_var_t cpu_isolated_map; 5679 5680/* Setup the mask of cpus configured for isolated domains */ 5681static int __init isolated_cpu_setup(char *str) 5682{ 5683 alloc_bootmem_cpumask_var(&cpu_isolated_map); 5684 cpulist_parse(str, cpu_isolated_map); 5685 return 1; 5686} 5687 5688__setup("isolcpus=", isolated_cpu_setup); 5689 5690static const struct cpumask *cpu_cpu_mask(int cpu) 5691{ 5692 return cpumask_of_node(cpu_to_node(cpu)); 5693} 5694 5695struct sd_data { 5696 struct sched_domain **__percpu sd; 5697 struct sched_group **__percpu sg; 5698 struct sched_group_power **__percpu sgp; 5699}; 5700 5701struct s_data { 5702 struct sched_domain ** __percpu sd; 5703 struct root_domain *rd; 5704}; 5705 5706enum s_alloc { 5707 sa_rootdomain, 5708 sa_sd, 5709 sa_sd_storage, 5710 sa_none, 5711}; 5712 5713struct sched_domain_topology_level; 5714 5715typedef struct sched_domain *(*sched_domain_init_f)(struct sched_domain_topology_level *tl, int cpu); 5716typedef const struct cpumask *(*sched_domain_mask_f)(int cpu); 5717 5718#define SDTL_OVERLAP 0x01 5719 5720struct sched_domain_topology_level { 5721 sched_domain_init_f init; 5722 sched_domain_mask_f mask; 5723 int flags; 5724 int numa_level; 5725 struct sd_data data; 5726}; 5727 5728/* 5729 * Build an iteration mask that can exclude certain CPUs from the upwards 5730 * domain traversal. 5731 * 5732 * Asymmetric node setups can result in situations where the domain tree is of 5733 * unequal depth, make sure to skip domains that already cover the entire 5734 * range. 5735 * 5736 * In that case build_sched_domains() will have terminated the iteration early 5737 * and our sibling sd spans will be empty. Domains should always include the 5738 * cpu they're built on, so check that. 5739 * 5740 */ 5741static void build_group_mask(struct sched_domain *sd, struct sched_group *sg) 5742{ 5743 const struct cpumask *span = sched_domain_span(sd); 5744 struct sd_data *sdd = sd->private; 5745 struct sched_domain *sibling; 5746 int i; 5747 5748 for_each_cpu(i, span) { 5749 sibling = *per_cpu_ptr(sdd->sd, i); 5750 if (!cpumask_test_cpu(i, sched_domain_span(sibling))) 5751 continue; 5752 5753 cpumask_set_cpu(i, sched_group_mask(sg)); 5754 } 5755} 5756 5757/* 5758 * Return the canonical balance cpu for this group, this is the first cpu 5759 * of this group that's also in the iteration mask. 5760 */ 5761int group_balance_cpu(struct sched_group *sg) 5762{ 5763 return cpumask_first_and(sched_group_cpus(sg), sched_group_mask(sg)); 5764} 5765 5766static int 5767build_overlap_sched_groups(struct sched_domain *sd, int cpu) 5768{ 5769 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg; 5770 const struct cpumask *span = sched_domain_span(sd); 5771 struct cpumask *covered = sched_domains_tmpmask; 5772 struct sd_data *sdd = sd->private; 5773 struct sched_domain *child; 5774 int i; 5775 5776 cpumask_clear(covered); 5777 5778 for_each_cpu(i, span) { 5779 struct cpumask *sg_span; 5780 5781 if (cpumask_test_cpu(i, covered)) 5782 continue; 5783 5784 child = *per_cpu_ptr(sdd->sd, i); 5785 5786 /* See the comment near build_group_mask(). */ 5787 if (!cpumask_test_cpu(i, sched_domain_span(child))) 5788 continue; 5789 5790 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(), 5791 GFP_KERNEL, cpu_to_node(cpu)); 5792 5793 if (!sg) 5794 goto fail; 5795 5796 sg_span = sched_group_cpus(sg); 5797 if (child->child) { 5798 child = child->child; 5799 cpumask_copy(sg_span, sched_domain_span(child)); 5800 } else 5801 cpumask_set_cpu(i, sg_span); 5802 5803 cpumask_or(covered, covered, sg_span); 5804 5805 sg->sgp = *per_cpu_ptr(sdd->sgp, i); 5806 if (atomic_inc_return(&sg->sgp->ref) == 1) 5807 build_group_mask(sd, sg); 5808 5809 /* 5810 * Initialize sgp->power such that even if we mess up the 5811 * domains and no possible iteration will get us here, we won't 5812 * die on a /0 trap. 5813 */ 5814 sg->sgp->power = SCHED_POWER_SCALE * cpumask_weight(sg_span); 5815 5816 /* 5817 * Make sure the first group of this domain contains the 5818 * canonical balance cpu. Otherwise the sched_domain iteration 5819 * breaks. See update_sg_lb_stats(). 5820 */ 5821 if ((!groups && cpumask_test_cpu(cpu, sg_span)) || 5822 group_balance_cpu(sg) == cpu) 5823 groups = sg; 5824 5825 if (!first) 5826 first = sg; 5827 if (last) 5828 last->next = sg; 5829 last = sg; 5830 last->next = first; 5831 } 5832 sd->groups = groups; 5833 5834 return 0; 5835 5836fail: 5837 free_sched_groups(first, 0); 5838 5839 return -ENOMEM; 5840} 5841 5842static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg) 5843{ 5844 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu); 5845 struct sched_domain *child = sd->child; 5846 5847 if (child) 5848 cpu = cpumask_first(sched_domain_span(child)); 5849 5850 if (sg) { 5851 *sg = *per_cpu_ptr(sdd->sg, cpu); 5852 (*sg)->sgp = *per_cpu_ptr(sdd->sgp, cpu); 5853 atomic_set(&(*sg)->sgp->ref, 1); /* for claim_allocations */ 5854 } 5855 5856 return cpu; 5857} 5858 5859/* 5860 * build_sched_groups will build a circular linked list of the groups 5861 * covered by the given span, and will set each group's ->cpumask correctly, 5862 * and ->cpu_power to 0. 5863 * 5864 * Assumes the sched_domain tree is fully constructed 5865 */ 5866static int 5867build_sched_groups(struct sched_domain *sd, int cpu) 5868{ 5869 struct sched_group *first = NULL, *last = NULL; 5870 struct sd_data *sdd = sd->private; 5871 const struct cpumask *span = sched_domain_span(sd); 5872 struct cpumask *covered; 5873 int i; 5874 5875 get_group(cpu, sdd, &sd->groups); 5876 atomic_inc(&sd->groups->ref); 5877 5878 if (cpu != cpumask_first(sched_domain_span(sd))) 5879 return 0; 5880 5881 lockdep_assert_held(&sched_domains_mutex); 5882 covered = sched_domains_tmpmask; 5883 5884 cpumask_clear(covered); 5885 5886 for_each_cpu(i, span) { 5887 struct sched_group *sg; 5888 int group = get_group(i, sdd, &sg); 5889 int j; 5890 5891 if (cpumask_test_cpu(i, covered)) 5892 continue; 5893 5894 cpumask_clear(sched_group_cpus(sg)); 5895 sg->sgp->power = 0; 5896 cpumask_setall(sched_group_mask(sg)); 5897 5898 for_each_cpu(j, span) { 5899 if (get_group(j, sdd, NULL) != group) 5900 continue; 5901 5902 cpumask_set_cpu(j, covered); 5903 cpumask_set_cpu(j, sched_group_cpus(sg)); 5904 } 5905 5906 if (!first) 5907 first = sg; 5908 if (last) 5909 last->next = sg; 5910 last = sg; 5911 } 5912 last->next = first; 5913 5914 return 0; 5915} 5916 5917/* 5918 * Initialize sched groups cpu_power. 5919 * 5920 * cpu_power indicates the capacity of sched group, which is used while 5921 * distributing the load between different sched groups in a sched domain. 5922 * Typically cpu_power for all the groups in a sched domain will be same unless 5923 * there are asymmetries in the topology. If there are asymmetries, group 5924 * having more cpu_power will pickup more load compared to the group having 5925 * less cpu_power. 5926 */ 5927static void init_sched_groups_power(int cpu, struct sched_domain *sd) 5928{ 5929 struct sched_group *sg = sd->groups; 5930 5931 WARN_ON(!sd || !sg); 5932 5933 do { 5934 sg->group_weight = cpumask_weight(sched_group_cpus(sg)); 5935 sg = sg->next; 5936 } while (sg != sd->groups); 5937 5938 if (cpu != group_balance_cpu(sg)) 5939 return; 5940 5941 update_group_power(sd, cpu); 5942 atomic_set(&sg->sgp->nr_busy_cpus, sg->group_weight); 5943} 5944 5945int __weak arch_sd_sibling_asym_packing(void) 5946{ 5947 return 0*SD_ASYM_PACKING; 5948} 5949 5950/* 5951 * Initializers for schedule domains 5952 * Non-inlined to reduce accumulated stack pressure in build_sched_domains() 5953 */ 5954 5955#ifdef CONFIG_SCHED_DEBUG 5956# define SD_INIT_NAME(sd, type) sd->name = #type 5957#else 5958# define SD_INIT_NAME(sd, type) do { } while (0) 5959#endif 5960 5961#define SD_INIT_FUNC(type) \ 5962static noinline struct sched_domain * \ 5963sd_init_##type(struct sched_domain_topology_level *tl, int cpu) \ 5964{ \ 5965 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); \ 5966 *sd = SD_##type##_INIT; \ 5967 SD_INIT_NAME(sd, type); \ 5968 sd->private = &tl->data; \ 5969 return sd; \ 5970} 5971 5972SD_INIT_FUNC(CPU) 5973#ifdef CONFIG_SCHED_SMT 5974 SD_INIT_FUNC(SIBLING) 5975#endif 5976#ifdef CONFIG_SCHED_MC 5977 SD_INIT_FUNC(MC) 5978#endif 5979#ifdef CONFIG_SCHED_BOOK 5980 SD_INIT_FUNC(BOOK) 5981#endif 5982 5983static int default_relax_domain_level = -1; 5984int sched_domain_level_max; 5985 5986static int __init setup_relax_domain_level(char *str) 5987{ 5988 if (kstrtoint(str, 0, &default_relax_domain_level)) 5989 pr_warn("Unable to set relax_domain_level\n"); 5990 5991 return 1; 5992} 5993__setup("relax_domain_level=", setup_relax_domain_level); 5994 5995static void set_domain_attribute(struct sched_domain *sd, 5996 struct sched_domain_attr *attr) 5997{ 5998 int request; 5999 6000 if (!attr || attr->relax_domain_level < 0) { 6001 if (default_relax_domain_level < 0) 6002 return; 6003 else 6004 request = default_relax_domain_level; 6005 } else 6006 request = attr->relax_domain_level; 6007 if (request < sd->level) { 6008 /* turn off idle balance on this domain */ 6009 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE); 6010 } else { 6011 /* turn on idle balance on this domain */ 6012 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE); 6013 } 6014} 6015 6016static void __sdt_free(const struct cpumask *cpu_map); 6017static int __sdt_alloc(const struct cpumask *cpu_map); 6018 6019static void __free_domain_allocs(struct s_data *d, enum s_alloc what, 6020 const struct cpumask *cpu_map) 6021{ 6022 switch (what) { 6023 case sa_rootdomain: 6024 if (!atomic_read(&d->rd->refcount)) 6025 free_rootdomain(&d->rd->rcu); /* fall through */ 6026 case sa_sd: 6027 free_percpu(d->sd); /* fall through */ 6028 case sa_sd_storage: 6029 __sdt_free(cpu_map); /* fall through */ 6030 case sa_none: 6031 break; 6032 } 6033} 6034 6035static enum s_alloc __visit_domain_allocation_hell(struct s_data *d, 6036 const struct cpumask *cpu_map) 6037{ 6038 memset(d, 0, sizeof(*d)); 6039 6040 if (__sdt_alloc(cpu_map)) 6041 return sa_sd_storage; 6042 d->sd = alloc_percpu(struct sched_domain *); 6043 if (!d->sd) 6044 return sa_sd_storage; 6045 d->rd = alloc_rootdomain(); 6046 if (!d->rd) 6047 return sa_sd; 6048 return sa_rootdomain; 6049} 6050 6051/* 6052 * NULL the sd_data elements we've used to build the sched_domain and 6053 * sched_group structure so that the subsequent __free_domain_allocs() 6054 * will not free the data we're using. 6055 */ 6056static void claim_allocations(int cpu, struct sched_domain *sd) 6057{ 6058 struct sd_data *sdd = sd->private; 6059 6060 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd); 6061 *per_cpu_ptr(sdd->sd, cpu) = NULL; 6062 6063 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref)) 6064 *per_cpu_ptr(sdd->sg, cpu) = NULL; 6065 6066 if (atomic_read(&(*per_cpu_ptr(sdd->sgp, cpu))->ref)) 6067 *per_cpu_ptr(sdd->sgp, cpu) = NULL; 6068} 6069 6070#ifdef CONFIG_SCHED_SMT 6071static const struct cpumask *cpu_smt_mask(int cpu) 6072{ 6073 return topology_thread_cpumask(cpu); 6074} 6075#endif 6076 6077/* 6078 * Topology list, bottom-up. 6079 */ 6080static struct sched_domain_topology_level default_topology[] = { 6081#ifdef CONFIG_SCHED_SMT 6082 { sd_init_SIBLING, cpu_smt_mask, }, 6083#endif 6084#ifdef CONFIG_SCHED_MC 6085 { sd_init_MC, cpu_coregroup_mask, }, 6086#endif 6087#ifdef CONFIG_SCHED_BOOK 6088 { sd_init_BOOK, cpu_book_mask, }, 6089#endif 6090 { sd_init_CPU, cpu_cpu_mask, }, 6091 { NULL, }, 6092}; 6093 6094static struct sched_domain_topology_level *sched_domain_topology = default_topology; 6095 6096#ifdef CONFIG_NUMA 6097 6098static int sched_domains_numa_levels; 6099static int *sched_domains_numa_distance; 6100static struct cpumask ***sched_domains_numa_masks; 6101static int sched_domains_curr_level; 6102 6103static inline int sd_local_flags(int level) 6104{ 6105 if (sched_domains_numa_distance[level] > RECLAIM_DISTANCE) 6106 return 0; 6107 6108 return SD_BALANCE_EXEC | SD_BALANCE_FORK | SD_WAKE_AFFINE; 6109} 6110 6111static struct sched_domain * 6112sd_numa_init(struct sched_domain_topology_level *tl, int cpu) 6113{ 6114 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); 6115 int level = tl->numa_level; 6116 int sd_weight = cpumask_weight( 6117 sched_domains_numa_masks[level][cpu_to_node(cpu)]); 6118 6119 *sd = (struct sched_domain){ 6120 .min_interval = sd_weight, 6121 .max_interval = 2*sd_weight, 6122 .busy_factor = 32, 6123 .imbalance_pct = 125, 6124 .cache_nice_tries = 2, 6125 .busy_idx = 3, 6126 .idle_idx = 2, 6127 .newidle_idx = 0, 6128 .wake_idx = 0, 6129 .forkexec_idx = 0, 6130 6131 .flags = 1*SD_LOAD_BALANCE 6132 | 1*SD_BALANCE_NEWIDLE 6133 | 0*SD_BALANCE_EXEC 6134 | 0*SD_BALANCE_FORK 6135 | 0*SD_BALANCE_WAKE 6136 | 0*SD_WAKE_AFFINE 6137 | 0*SD_SHARE_CPUPOWER 6138 | 0*SD_SHARE_PKG_RESOURCES 6139 | 1*SD_SERIALIZE 6140 | 0*SD_PREFER_SIBLING 6141 | sd_local_flags(level) 6142 , 6143 .last_balance = jiffies, 6144 .balance_interval = sd_weight, 6145 }; 6146 SD_INIT_NAME(sd, NUMA); 6147 sd->private = &tl->data; 6148 6149 /* 6150 * Ugly hack to pass state to sd_numa_mask()... 6151 */ 6152 sched_domains_curr_level = tl->numa_level; 6153 6154 return sd; 6155} 6156 6157static const struct cpumask *sd_numa_mask(int cpu) 6158{ 6159 return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)]; 6160} 6161 6162static void sched_numa_warn(const char *str) 6163{ 6164 static int done = false; 6165 int i,j; 6166 6167 if (done) 6168 return; 6169 6170 done = true; 6171 6172 printk(KERN_WARNING "ERROR: %s\n\n", str); 6173 6174 for (i = 0; i < nr_node_ids; i++) { 6175 printk(KERN_WARNING " "); 6176 for (j = 0; j < nr_node_ids; j++) 6177 printk(KERN_CONT "%02d ", node_distance(i,j)); 6178 printk(KERN_CONT "\n"); 6179 } 6180 printk(KERN_WARNING "\n"); 6181} 6182 6183static bool find_numa_distance(int distance) 6184{ 6185 int i; 6186 6187 if (distance == node_distance(0, 0)) 6188 return true; 6189 6190 for (i = 0; i < sched_domains_numa_levels; i++) { 6191 if (sched_domains_numa_distance[i] == distance) 6192 return true; 6193 } 6194 6195 return false; 6196} 6197 6198static void sched_init_numa(void) 6199{ 6200 int next_distance, curr_distance = node_distance(0, 0); 6201 struct sched_domain_topology_level *tl; 6202 int level = 0; 6203 int i, j, k; 6204 6205 sched_domains_numa_distance = kzalloc(sizeof(int) * nr_node_ids, GFP_KERNEL); 6206 if (!sched_domains_numa_distance) 6207 return; 6208 6209 /* 6210 * O(nr_nodes^2) deduplicating selection sort -- in order to find the 6211 * unique distances in the node_distance() table. 6212 * 6213 * Assumes node_distance(0,j) includes all distances in 6214 * node_distance(i,j) in order to avoid cubic time. 6215 */ 6216 next_distance = curr_distance; 6217 for (i = 0; i < nr_node_ids; i++) { 6218 for (j = 0; j < nr_node_ids; j++) { 6219 for (k = 0; k < nr_node_ids; k++) { 6220 int distance = node_distance(i, k); 6221 6222 if (distance > curr_distance && 6223 (distance < next_distance || 6224 next_distance == curr_distance)) 6225 next_distance = distance; 6226 6227 /* 6228 * While not a strong assumption it would be nice to know 6229 * about cases where if node A is connected to B, B is not 6230 * equally connected to A. 6231 */ 6232 if (sched_debug() && node_distance(k, i) != distance) 6233 sched_numa_warn("Node-distance not symmetric"); 6234 6235 if (sched_debug() && i && !find_numa_distance(distance)) 6236 sched_numa_warn("Node-0 not representative"); 6237 } 6238 if (next_distance != curr_distance) { 6239 sched_domains_numa_distance[level++] = next_distance; 6240 sched_domains_numa_levels = level; 6241 curr_distance = next_distance; 6242 } else break; 6243 } 6244 6245 /* 6246 * In case of sched_debug() we verify the above assumption. 6247 */ 6248 if (!sched_debug()) 6249 break; 6250 } 6251 /* 6252 * 'level' contains the number of unique distances, excluding the 6253 * identity distance node_distance(i,i). 6254 * 6255 * The sched_domains_numa_distance[] array includes the actual distance 6256 * numbers. 6257 */ 6258 6259 /* 6260 * Here, we should temporarily reset sched_domains_numa_levels to 0. 6261 * If it fails to allocate memory for array sched_domains_numa_masks[][], 6262 * the array will contain less then 'level' members. This could be 6263 * dangerous when we use it to iterate array sched_domains_numa_masks[][] 6264 * in other functions. 6265 * 6266 * We reset it to 'level' at the end of this function. 6267 */ 6268 sched_domains_numa_levels = 0; 6269 6270 sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL); 6271 if (!sched_domains_numa_masks) 6272 return; 6273 6274 /* 6275 * Now for each level, construct a mask per node which contains all 6276 * cpus of nodes that are that many hops away from us. 6277 */ 6278 for (i = 0; i < level; i++) { 6279 sched_domains_numa_masks[i] = 6280 kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL); 6281 if (!sched_domains_numa_masks[i]) 6282 return; 6283 6284 for (j = 0; j < nr_node_ids; j++) { 6285 struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL); 6286 if (!mask) 6287 return; 6288 6289 sched_domains_numa_masks[i][j] = mask; 6290 6291 for (k = 0; k < nr_node_ids; k++) { 6292 if (node_distance(j, k) > sched_domains_numa_distance[i]) 6293 continue; 6294 6295 cpumask_or(mask, mask, cpumask_of_node(k)); 6296 } 6297 } 6298 } 6299 6300 tl = kzalloc((ARRAY_SIZE(default_topology) + level) * 6301 sizeof(struct sched_domain_topology_level), GFP_KERNEL); 6302 if (!tl) 6303 return; 6304 6305 /* 6306 * Copy the default topology bits.. 6307 */ 6308 for (i = 0; default_topology[i].init; i++) 6309 tl[i] = default_topology[i]; 6310 6311 /* 6312 * .. and append 'j' levels of NUMA goodness. 6313 */ 6314 for (j = 0; j < level; i++, j++) { 6315 tl[i] = (struct sched_domain_topology_level){ 6316 .init = sd_numa_init, 6317 .mask = sd_numa_mask, 6318 .flags = SDTL_OVERLAP, 6319 .numa_level = j, 6320 }; 6321 } 6322 6323 sched_domain_topology = tl; 6324 6325 sched_domains_numa_levels = level; 6326} 6327 6328static void sched_domains_numa_masks_set(int cpu) 6329{ 6330 int i, j; 6331 int node = cpu_to_node(cpu); 6332 6333 for (i = 0; i < sched_domains_numa_levels; i++) { 6334 for (j = 0; j < nr_node_ids; j++) { 6335 if (node_distance(j, node) <= sched_domains_numa_distance[i]) 6336 cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]); 6337 } 6338 } 6339} 6340 6341static void sched_domains_numa_masks_clear(int cpu) 6342{ 6343 int i, j; 6344 for (i = 0; i < sched_domains_numa_levels; i++) { 6345 for (j = 0; j < nr_node_ids; j++) 6346 cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]); 6347 } 6348} 6349 6350/* 6351 * Update sched_domains_numa_masks[level][node] array when new cpus 6352 * are onlined. 6353 */ 6354static int sched_domains_numa_masks_update(struct notifier_block *nfb, 6355 unsigned long action, 6356 void *hcpu) 6357{ 6358 int cpu = (long)hcpu; 6359 6360 switch (action & ~CPU_TASKS_FROZEN) { 6361 case CPU_ONLINE: 6362 sched_domains_numa_masks_set(cpu); 6363 break; 6364 6365 case CPU_DEAD: 6366 sched_domains_numa_masks_clear(cpu); 6367 break; 6368 6369 default: 6370 return NOTIFY_DONE; 6371 } 6372 6373 return NOTIFY_OK; 6374} 6375#else 6376static inline void sched_init_numa(void) 6377{ 6378} 6379 6380static int sched_domains_numa_masks_update(struct notifier_block *nfb, 6381 unsigned long action, 6382 void *hcpu) 6383{ 6384 return 0; 6385} 6386#endif /* CONFIG_NUMA */ 6387 6388static int __sdt_alloc(const struct cpumask *cpu_map) 6389{ 6390 struct sched_domain_topology_level *tl; 6391 int j; 6392 6393 for (tl = sched_domain_topology; tl->init; tl++) { 6394 struct sd_data *sdd = &tl->data; 6395 6396 sdd->sd = alloc_percpu(struct sched_domain *); 6397 if (!sdd->sd) 6398 return -ENOMEM; 6399 6400 sdd->sg = alloc_percpu(struct sched_group *); 6401 if (!sdd->sg) 6402 return -ENOMEM; 6403 6404 sdd->sgp = alloc_percpu(struct sched_group_power *); 6405 if (!sdd->sgp) 6406 return -ENOMEM; 6407 6408 for_each_cpu(j, cpu_map) { 6409 struct sched_domain *sd; 6410 struct sched_group *sg; 6411 struct sched_group_power *sgp; 6412 6413 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(), 6414 GFP_KERNEL, cpu_to_node(j)); 6415 if (!sd) 6416 return -ENOMEM; 6417 6418 *per_cpu_ptr(sdd->sd, j) = sd; 6419 6420 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(), 6421 GFP_KERNEL, cpu_to_node(j)); 6422 if (!sg) 6423 return -ENOMEM; 6424 6425 sg->next = sg; 6426 6427 *per_cpu_ptr(sdd->sg, j) = sg; 6428 6429 sgp = kzalloc_node(sizeof(struct sched_group_power) + cpumask_size(), 6430 GFP_KERNEL, cpu_to_node(j)); 6431 if (!sgp) 6432 return -ENOMEM; 6433 6434 *per_cpu_ptr(sdd->sgp, j) = sgp; 6435 } 6436 } 6437 6438 return 0; 6439} 6440 6441static void __sdt_free(const struct cpumask *cpu_map) 6442{ 6443 struct sched_domain_topology_level *tl; 6444 int j; 6445 6446 for (tl = sched_domain_topology; tl->init; tl++) { 6447 struct sd_data *sdd = &tl->data; 6448 6449 for_each_cpu(j, cpu_map) { 6450 struct sched_domain *sd; 6451 6452 if (sdd->sd) { 6453 sd = *per_cpu_ptr(sdd->sd, j); 6454 if (sd && (sd->flags & SD_OVERLAP)) 6455 free_sched_groups(sd->groups, 0); 6456 kfree(*per_cpu_ptr(sdd->sd, j)); 6457 } 6458 6459 if (sdd->sg) 6460 kfree(*per_cpu_ptr(sdd->sg, j)); 6461 if (sdd->sgp) 6462 kfree(*per_cpu_ptr(sdd->sgp, j)); 6463 } 6464 free_percpu(sdd->sd); 6465 sdd->sd = NULL; 6466 free_percpu(sdd->sg); 6467 sdd->sg = NULL; 6468 free_percpu(sdd->sgp); 6469 sdd->sgp = NULL; 6470 } 6471} 6472 6473struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl, 6474 struct s_data *d, const struct cpumask *cpu_map, 6475 struct sched_domain_attr *attr, struct sched_domain *child, 6476 int cpu) 6477{ 6478 struct sched_domain *sd = tl->init(tl, cpu); 6479 if (!sd) 6480 return child; 6481 6482 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu)); 6483 if (child) { 6484 sd->level = child->level + 1; 6485 sched_domain_level_max = max(sched_domain_level_max, sd->level); 6486 child->parent = sd; 6487 } 6488 sd->child = child; 6489 set_domain_attribute(sd, attr); 6490 6491 return sd; 6492} 6493 6494/* 6495 * Build sched domains for a given set of cpus and attach the sched domains 6496 * to the individual cpus 6497 */ 6498static int build_sched_domains(const struct cpumask *cpu_map, 6499 struct sched_domain_attr *attr) 6500{ 6501 enum s_alloc alloc_state = sa_none; 6502 struct sched_domain *sd; 6503 struct s_data d; 6504 int i, ret = -ENOMEM; 6505 6506 alloc_state = __visit_domain_allocation_hell(&d, cpu_map); 6507 if (alloc_state != sa_rootdomain) 6508 goto error; 6509 6510 /* Set up domains for cpus specified by the cpu_map. */ 6511 for_each_cpu(i, cpu_map) { 6512 struct sched_domain_topology_level *tl; 6513 6514 sd = NULL; 6515 for (tl = sched_domain_topology; tl->init; tl++) { 6516 sd = build_sched_domain(tl, &d, cpu_map, attr, sd, i); 6517 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP)) 6518 sd->flags |= SD_OVERLAP; 6519 if (cpumask_equal(cpu_map, sched_domain_span(sd))) 6520 break; 6521 } 6522 6523 while (sd->child) 6524 sd = sd->child; 6525 6526 *per_cpu_ptr(d.sd, i) = sd; 6527 } 6528 6529 /* Build the groups for the domains */ 6530 for_each_cpu(i, cpu_map) { 6531 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) { 6532 sd->span_weight = cpumask_weight(sched_domain_span(sd)); 6533 if (sd->flags & SD_OVERLAP) { 6534 if (build_overlap_sched_groups(sd, i)) 6535 goto error; 6536 } else { 6537 if (build_sched_groups(sd, i)) 6538 goto error; 6539 } 6540 } 6541 } 6542 6543 /* Calculate CPU power for physical packages and nodes */ 6544 for (i = nr_cpumask_bits-1; i >= 0; i--) { 6545 if (!cpumask_test_cpu(i, cpu_map)) 6546 continue; 6547 6548 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) { 6549 claim_allocations(i, sd); 6550 init_sched_groups_power(i, sd); 6551 } 6552 } 6553 6554 /* Attach the domains */ 6555 rcu_read_lock(); 6556 for_each_cpu(i, cpu_map) { 6557 sd = *per_cpu_ptr(d.sd, i); 6558 cpu_attach_domain(sd, d.rd, i); 6559 } 6560 rcu_read_unlock(); 6561 6562 ret = 0; 6563error: 6564 __free_domain_allocs(&d, alloc_state, cpu_map); 6565 return ret; 6566} 6567 6568static cpumask_var_t *doms_cur; /* current sched domains */ 6569static int ndoms_cur; /* number of sched domains in 'doms_cur' */ 6570static struct sched_domain_attr *dattr_cur; 6571 /* attribues of custom domains in 'doms_cur' */ 6572 6573/* 6574 * Special case: If a kmalloc of a doms_cur partition (array of 6575 * cpumask) fails, then fallback to a single sched domain, 6576 * as determined by the single cpumask fallback_doms. 6577 */ 6578static cpumask_var_t fallback_doms; 6579 6580/* 6581 * arch_update_cpu_topology lets virtualized architectures update the 6582 * cpu core maps. It is supposed to return 1 if the topology changed 6583 * or 0 if it stayed the same. 6584 */ 6585int __attribute__((weak)) arch_update_cpu_topology(void) 6586{ 6587 return 0; 6588} 6589 6590cpumask_var_t *alloc_sched_domains(unsigned int ndoms) 6591{ 6592 int i; 6593 cpumask_var_t *doms; 6594 6595 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL); 6596 if (!doms) 6597 return NULL; 6598 for (i = 0; i < ndoms; i++) { 6599 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) { 6600 free_sched_domains(doms, i); 6601 return NULL; 6602 } 6603 } 6604 return doms; 6605} 6606 6607void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms) 6608{ 6609 unsigned int i; 6610 for (i = 0; i < ndoms; i++) 6611 free_cpumask_var(doms[i]); 6612 kfree(doms); 6613} 6614 6615/* 6616 * Set up scheduler domains and groups. Callers must hold the hotplug lock. 6617 * For now this just excludes isolated cpus, but could be used to 6618 * exclude other special cases in the future. 6619 */ 6620static int init_sched_domains(const struct cpumask *cpu_map) 6621{ 6622 int err; 6623 6624 arch_update_cpu_topology(); 6625 ndoms_cur = 1; 6626 doms_cur = alloc_sched_domains(ndoms_cur); 6627 if (!doms_cur) 6628 doms_cur = &fallback_doms; 6629 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map); 6630 err = build_sched_domains(doms_cur[0], NULL); 6631 register_sched_domain_sysctl(); 6632 6633 return err; 6634} 6635 6636/* 6637 * Detach sched domains from a group of cpus specified in cpu_map 6638 * These cpus will now be attached to the NULL domain 6639 */ 6640static void detach_destroy_domains(const struct cpumask *cpu_map) 6641{ 6642 int i; 6643 6644 rcu_read_lock(); 6645 for_each_cpu(i, cpu_map) 6646 cpu_attach_domain(NULL, &def_root_domain, i); 6647 rcu_read_unlock(); 6648} 6649 6650/* handle null as "default" */ 6651static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur, 6652 struct sched_domain_attr *new, int idx_new) 6653{ 6654 struct sched_domain_attr tmp; 6655 6656 /* fast path */ 6657 if (!new && !cur) 6658 return 1; 6659 6660 tmp = SD_ATTR_INIT; 6661 return !memcmp(cur ? (cur + idx_cur) : &tmp, 6662 new ? (new + idx_new) : &tmp, 6663 sizeof(struct sched_domain_attr)); 6664} 6665 6666/* 6667 * Partition sched domains as specified by the 'ndoms_new' 6668 * cpumasks in the array doms_new[] of cpumasks. This compares 6669 * doms_new[] to the current sched domain partitioning, doms_cur[]. 6670 * It destroys each deleted domain and builds each new domain. 6671 * 6672 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'. 6673 * The masks don't intersect (don't overlap.) We should setup one 6674 * sched domain for each mask. CPUs not in any of the cpumasks will 6675 * not be load balanced. If the same cpumask appears both in the 6676 * current 'doms_cur' domains and in the new 'doms_new', we can leave 6677 * it as it is. 6678 * 6679 * The passed in 'doms_new' should be allocated using 6680 * alloc_sched_domains. This routine takes ownership of it and will 6681 * free_sched_domains it when done with it. If the caller failed the 6682 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1, 6683 * and partition_sched_domains() will fallback to the single partition 6684 * 'fallback_doms', it also forces the domains to be rebuilt. 6685 * 6686 * If doms_new == NULL it will be replaced with cpu_online_mask. 6687 * ndoms_new == 0 is a special case for destroying existing domains, 6688 * and it will not create the default domain. 6689 * 6690 * Call with hotplug lock held 6691 */ 6692void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[], 6693 struct sched_domain_attr *dattr_new) 6694{ 6695 int i, j, n; 6696 int new_topology; 6697 6698 mutex_lock(&sched_domains_mutex); 6699 6700 /* always unregister in case we don't destroy any domains */ 6701 unregister_sched_domain_sysctl(); 6702 6703 /* Let architecture update cpu core mappings. */ 6704 new_topology = arch_update_cpu_topology(); 6705 6706 n = doms_new ? ndoms_new : 0; 6707 6708 /* Destroy deleted domains */ 6709 for (i = 0; i < ndoms_cur; i++) { 6710 for (j = 0; j < n && !new_topology; j++) { 6711 if (cpumask_equal(doms_cur[i], doms_new[j]) 6712 && dattrs_equal(dattr_cur, i, dattr_new, j)) 6713 goto match1; 6714 } 6715 /* no match - a current sched domain not in new doms_new[] */ 6716 detach_destroy_domains(doms_cur[i]); 6717match1: 6718 ; 6719 } 6720 6721 if (doms_new == NULL) { 6722 ndoms_cur = 0; 6723 doms_new = &fallback_doms; 6724 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map); 6725 WARN_ON_ONCE(dattr_new); 6726 } 6727 6728 /* Build new domains */ 6729 for (i = 0; i < ndoms_new; i++) { 6730 for (j = 0; j < ndoms_cur && !new_topology; j++) { 6731 if (cpumask_equal(doms_new[i], doms_cur[j]) 6732 && dattrs_equal(dattr_new, i, dattr_cur, j)) 6733 goto match2; 6734 } 6735 /* no match - add a new doms_new */ 6736 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL); 6737match2: 6738 ; 6739 } 6740 6741 /* Remember the new sched domains */ 6742 if (doms_cur != &fallback_doms) 6743 free_sched_domains(doms_cur, ndoms_cur); 6744 kfree(dattr_cur); /* kfree(NULL) is safe */ 6745 doms_cur = doms_new; 6746 dattr_cur = dattr_new; 6747 ndoms_cur = ndoms_new; 6748 6749 register_sched_domain_sysctl(); 6750 6751 mutex_unlock(&sched_domains_mutex); 6752} 6753 6754static int num_cpus_frozen; /* used to mark begin/end of suspend/resume */ 6755 6756/* 6757 * Update cpusets according to cpu_active mask. If cpusets are 6758 * disabled, cpuset_update_active_cpus() becomes a simple wrapper 6759 * around partition_sched_domains(). 6760 * 6761 * If we come here as part of a suspend/resume, don't touch cpusets because we 6762 * want to restore it back to its original state upon resume anyway. 6763 */ 6764static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action, 6765 void *hcpu) 6766{ 6767 switch (action) { 6768 case CPU_ONLINE_FROZEN: 6769 case CPU_DOWN_FAILED_FROZEN: 6770 6771 /* 6772 * num_cpus_frozen tracks how many CPUs are involved in suspend 6773 * resume sequence. As long as this is not the last online 6774 * operation in the resume sequence, just build a single sched 6775 * domain, ignoring cpusets. 6776 */ 6777 num_cpus_frozen--; 6778 if (likely(num_cpus_frozen)) { 6779 partition_sched_domains(1, NULL, NULL); 6780 break; 6781 } 6782 6783 /* 6784 * This is the last CPU online operation. So fall through and 6785 * restore the original sched domains by considering the 6786 * cpuset configurations. 6787 */ 6788 6789 case CPU_ONLINE: 6790 case CPU_DOWN_FAILED: 6791 cpuset_update_active_cpus(true); 6792 break; 6793 default: 6794 return NOTIFY_DONE; 6795 } 6796 return NOTIFY_OK; 6797} 6798 6799static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action, 6800 void *hcpu) 6801{ 6802 switch (action) { 6803 case CPU_DOWN_PREPARE: 6804 cpuset_update_active_cpus(false); 6805 break; 6806 case CPU_DOWN_PREPARE_FROZEN: 6807 num_cpus_frozen++; 6808 partition_sched_domains(1, NULL, NULL); 6809 break; 6810 default: 6811 return NOTIFY_DONE; 6812 } 6813 return NOTIFY_OK; 6814} 6815 6816void __init sched_init_smp(void) 6817{ 6818 cpumask_var_t non_isolated_cpus; 6819 6820 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL); 6821 alloc_cpumask_var(&fallback_doms, GFP_KERNEL); 6822 6823 sched_init_numa(); 6824 6825 get_online_cpus(); 6826 mutex_lock(&sched_domains_mutex); 6827 init_sched_domains(cpu_active_mask); 6828 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map); 6829 if (cpumask_empty(non_isolated_cpus)) 6830 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus); 6831 mutex_unlock(&sched_domains_mutex); 6832 put_online_cpus(); 6833 6834 hotcpu_notifier(sched_domains_numa_masks_update, CPU_PRI_SCHED_ACTIVE); 6835 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE); 6836 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE); 6837 6838 /* RT runtime code needs to handle some hotplug events */ 6839 hotcpu_notifier(update_runtime, 0); 6840 6841 init_hrtick(); 6842 6843 /* Move init over to a non-isolated CPU */ 6844 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0) 6845 BUG(); 6846 sched_init_granularity(); 6847 free_cpumask_var(non_isolated_cpus); 6848 6849 init_sched_rt_class(); 6850} 6851#else 6852void __init sched_init_smp(void) 6853{ 6854 sched_init_granularity(); 6855} 6856#endif /* CONFIG_SMP */ 6857 6858const_debug unsigned int sysctl_timer_migration = 1; 6859 6860int in_sched_functions(unsigned long addr) 6861{ 6862 return in_lock_functions(addr) || 6863 (addr >= (unsigned long)__sched_text_start 6864 && addr < (unsigned long)__sched_text_end); 6865} 6866 6867#ifdef CONFIG_CGROUP_SCHED 6868/* 6869 * Default task group. 6870 * Every task in system belongs to this group at bootup. 6871 */ 6872struct task_group root_task_group; 6873LIST_HEAD(task_groups); 6874#endif 6875 6876DECLARE_PER_CPU(cpumask_var_t, load_balance_tmpmask); 6877 6878void __init sched_init(void) 6879{ 6880 int i, j; 6881 unsigned long alloc_size = 0, ptr; 6882 6883#ifdef CONFIG_FAIR_GROUP_SCHED 6884 alloc_size += 2 * nr_cpu_ids * sizeof(void **); 6885#endif 6886#ifdef CONFIG_RT_GROUP_SCHED 6887 alloc_size += 2 * nr_cpu_ids * sizeof(void **); 6888#endif 6889#ifdef CONFIG_CPUMASK_OFFSTACK 6890 alloc_size += num_possible_cpus() * cpumask_size(); 6891#endif 6892 if (alloc_size) { 6893 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT); 6894 6895#ifdef CONFIG_FAIR_GROUP_SCHED 6896 root_task_group.se = (struct sched_entity **)ptr; 6897 ptr += nr_cpu_ids * sizeof(void **); 6898 6899 root_task_group.cfs_rq = (struct cfs_rq **)ptr; 6900 ptr += nr_cpu_ids * sizeof(void **); 6901 6902#endif /* CONFIG_FAIR_GROUP_SCHED */ 6903#ifdef CONFIG_RT_GROUP_SCHED 6904 root_task_group.rt_se = (struct sched_rt_entity **)ptr; 6905 ptr += nr_cpu_ids * sizeof(void **); 6906 6907 root_task_group.rt_rq = (struct rt_rq **)ptr; 6908 ptr += nr_cpu_ids * sizeof(void **); 6909 6910#endif /* CONFIG_RT_GROUP_SCHED */ 6911#ifdef CONFIG_CPUMASK_OFFSTACK 6912 for_each_possible_cpu(i) { 6913 per_cpu(load_balance_tmpmask, i) = (void *)ptr; 6914 ptr += cpumask_size(); 6915 } 6916#endif /* CONFIG_CPUMASK_OFFSTACK */ 6917 } 6918 6919#ifdef CONFIG_SMP 6920 init_defrootdomain(); 6921#endif 6922 6923 init_rt_bandwidth(&def_rt_bandwidth, 6924 global_rt_period(), global_rt_runtime()); 6925 6926#ifdef CONFIG_RT_GROUP_SCHED 6927 init_rt_bandwidth(&root_task_group.rt_bandwidth, 6928 global_rt_period(), global_rt_runtime()); 6929#endif /* CONFIG_RT_GROUP_SCHED */ 6930 6931#ifdef CONFIG_CGROUP_SCHED 6932 list_add(&root_task_group.list, &task_groups); 6933 INIT_LIST_HEAD(&root_task_group.children); 6934 INIT_LIST_HEAD(&root_task_group.siblings); 6935 autogroup_init(&init_task); 6936 6937#endif /* CONFIG_CGROUP_SCHED */ 6938 6939#ifdef CONFIG_CGROUP_CPUACCT 6940 root_cpuacct.cpustat = &kernel_cpustat; 6941 root_cpuacct.cpuusage = alloc_percpu(u64); 6942 /* Too early, not expected to fail */ 6943 BUG_ON(!root_cpuacct.cpuusage); 6944#endif 6945 for_each_possible_cpu(i) { 6946 struct rq *rq; 6947 6948 rq = cpu_rq(i); 6949 raw_spin_lock_init(&rq->lock); 6950 rq->nr_running = 0; 6951 rq->calc_load_active = 0; 6952 rq->calc_load_update = jiffies + LOAD_FREQ; 6953 init_cfs_rq(&rq->cfs); 6954 init_rt_rq(&rq->rt, rq); 6955#ifdef CONFIG_FAIR_GROUP_SCHED 6956 root_task_group.shares = ROOT_TASK_GROUP_LOAD; 6957 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list); 6958 /* 6959 * How much cpu bandwidth does root_task_group get? 6960 * 6961 * In case of task-groups formed thr' the cgroup filesystem, it 6962 * gets 100% of the cpu resources in the system. This overall 6963 * system cpu resource is divided among the tasks of 6964 * root_task_group and its child task-groups in a fair manner, 6965 * based on each entity's (task or task-group's) weight 6966 * (se->load.weight). 6967 * 6968 * In other words, if root_task_group has 10 tasks of weight 6969 * 1024) and two child groups A0 and A1 (of weight 1024 each), 6970 * then A0's share of the cpu resource is: 6971 * 6972 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33% 6973 * 6974 * We achieve this by letting root_task_group's tasks sit 6975 * directly in rq->cfs (i.e root_task_group->se[] = NULL). 6976 */ 6977 init_cfs_bandwidth(&root_task_group.cfs_bandwidth); 6978 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL); 6979#endif /* CONFIG_FAIR_GROUP_SCHED */ 6980 6981 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime; 6982#ifdef CONFIG_RT_GROUP_SCHED 6983 INIT_LIST_HEAD(&rq->leaf_rt_rq_list); 6984 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL); 6985#endif 6986 6987 for (j = 0; j < CPU_LOAD_IDX_MAX; j++) 6988 rq->cpu_load[j] = 0; 6989 6990 rq->last_load_update_tick = jiffies; 6991 6992#ifdef CONFIG_SMP 6993 rq->sd = NULL; 6994 rq->rd = NULL; 6995 rq->cpu_power = SCHED_POWER_SCALE; 6996 rq->post_schedule = 0; 6997 rq->active_balance = 0; 6998 rq->next_balance = jiffies; 6999 rq->push_cpu = 0; 7000 rq->cpu = i; 7001 rq->online = 0; 7002 rq->idle_stamp = 0; 7003 rq->avg_idle = 2*sysctl_sched_migration_cost; 7004 7005 INIT_LIST_HEAD(&rq->cfs_tasks); 7006 7007 rq_attach_root(rq, &def_root_domain); 7008#ifdef CONFIG_NO_HZ 7009 rq->nohz_flags = 0; 7010#endif 7011#endif 7012 init_rq_hrtick(rq); 7013 atomic_set(&rq->nr_iowait, 0); 7014 } 7015 7016 set_load_weight(&init_task); 7017 7018#ifdef CONFIG_PREEMPT_NOTIFIERS 7019 INIT_HLIST_HEAD(&init_task.preempt_notifiers); 7020#endif 7021 7022#ifdef CONFIG_RT_MUTEXES 7023 plist_head_init(&init_task.pi_waiters); 7024#endif 7025 7026 /* 7027 * The boot idle thread does lazy MMU switching as well: 7028 */ 7029 atomic_inc(&init_mm.mm_count); 7030 enter_lazy_tlb(&init_mm, current); 7031 7032 /* 7033 * Make us the idle thread. Technically, schedule() should not be 7034 * called from this thread, however somewhere below it might be, 7035 * but because we are the idle thread, we just pick up running again 7036 * when this runqueue becomes "idle". 7037 */ 7038 init_idle(current, smp_processor_id()); 7039 7040 calc_load_update = jiffies + LOAD_FREQ; 7041 7042 /* 7043 * During early bootup we pretend to be a normal task: 7044 */ 7045 current->sched_class = &fair_sched_class; 7046 7047#ifdef CONFIG_SMP 7048 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT); 7049 /* May be allocated at isolcpus cmdline parse time */ 7050 if (cpu_isolated_map == NULL) 7051 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT); 7052 idle_thread_set_boot_cpu(); 7053#endif 7054 init_sched_fair_class(); 7055 7056 scheduler_running = 1; 7057} 7058 7059#ifdef CONFIG_DEBUG_ATOMIC_SLEEP 7060static inline int preempt_count_equals(int preempt_offset) 7061{ 7062 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth(); 7063 7064 return (nested == preempt_offset); 7065} 7066 7067void __might_sleep(const char *file, int line, int preempt_offset) 7068{ 7069 static unsigned long prev_jiffy; /* ratelimiting */ 7070 7071 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */ 7072 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) || 7073 system_state != SYSTEM_RUNNING || oops_in_progress) 7074 return; 7075 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy) 7076 return; 7077 prev_jiffy = jiffies; 7078 7079 printk(KERN_ERR 7080 "BUG: sleeping function called from invalid context at %s:%d\n", 7081 file, line); 7082 printk(KERN_ERR 7083 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n", 7084 in_atomic(), irqs_disabled(), 7085 current->pid, current->comm); 7086 7087 debug_show_held_locks(current); 7088 if (irqs_disabled()) 7089 print_irqtrace_events(current); 7090 dump_stack(); 7091} 7092EXPORT_SYMBOL(__might_sleep); 7093#endif 7094 7095#ifdef CONFIG_MAGIC_SYSRQ 7096static void normalize_task(struct rq *rq, struct task_struct *p) 7097{ 7098 const struct sched_class *prev_class = p->sched_class; 7099 int old_prio = p->prio; 7100 int on_rq; 7101 7102 on_rq = p->on_rq; 7103 if (on_rq) 7104 dequeue_task(rq, p, 0); 7105 __setscheduler(rq, p, SCHED_NORMAL, 0); 7106 if (on_rq) { 7107 enqueue_task(rq, p, 0); 7108 resched_task(rq->curr); 7109 } 7110 7111 check_class_changed(rq, p, prev_class, old_prio); 7112} 7113 7114void normalize_rt_tasks(void) 7115{ 7116 struct task_struct *g, *p; 7117 unsigned long flags; 7118 struct rq *rq; 7119 7120 read_lock_irqsave(&tasklist_lock, flags); 7121 do_each_thread(g, p) { 7122 /* 7123 * Only normalize user tasks: 7124 */ 7125 if (!p->mm) 7126 continue; 7127 7128 p->se.exec_start = 0; 7129#ifdef CONFIG_SCHEDSTATS 7130 p->se.statistics.wait_start = 0; 7131 p->se.statistics.sleep_start = 0; 7132 p->se.statistics.block_start = 0; 7133#endif 7134 7135 if (!rt_task(p)) { 7136 /* 7137 * Renice negative nice level userspace 7138 * tasks back to 0: 7139 */ 7140 if (TASK_NICE(p) < 0 && p->mm) 7141 set_user_nice(p, 0); 7142 continue; 7143 } 7144 7145 raw_spin_lock(&p->pi_lock); 7146 rq = __task_rq_lock(p); 7147 7148 normalize_task(rq, p); 7149 7150 __task_rq_unlock(rq); 7151 raw_spin_unlock(&p->pi_lock); 7152 } while_each_thread(g, p); 7153 7154 read_unlock_irqrestore(&tasklist_lock, flags); 7155} 7156 7157#endif /* CONFIG_MAGIC_SYSRQ */ 7158 7159#if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) 7160/* 7161 * These functions are only useful for the IA64 MCA handling, or kdb. 7162 * 7163 * They can only be called when the whole system has been 7164 * stopped - every CPU needs to be quiescent, and no scheduling 7165 * activity can take place. Using them for anything else would 7166 * be a serious bug, and as a result, they aren't even visible 7167 * under any other configuration. 7168 */ 7169 7170/** 7171 * curr_task - return the current task for a given cpu. 7172 * @cpu: the processor in question. 7173 * 7174 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED! 7175 */ 7176struct task_struct *curr_task(int cpu) 7177{ 7178 return cpu_curr(cpu); 7179} 7180 7181#endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */ 7182 7183#ifdef CONFIG_IA64 7184/** 7185 * set_curr_task - set the current task for a given cpu. 7186 * @cpu: the processor in question. 7187 * @p: the task pointer to set. 7188 * 7189 * Description: This function must only be used when non-maskable interrupts 7190 * are serviced on a separate stack. It allows the architecture to switch the 7191 * notion of the current task on a cpu in a non-blocking manner. This function 7192 * must be called with all CPU's synchronized, and interrupts disabled, the 7193 * and caller must save the original value of the current task (see 7194 * curr_task() above) and restore that value before reenabling interrupts and 7195 * re-starting the system. 7196 * 7197 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED! 7198 */ 7199void set_curr_task(int cpu, struct task_struct *p) 7200{ 7201 cpu_curr(cpu) = p; 7202} 7203 7204#endif 7205 7206#ifdef CONFIG_CGROUP_SCHED 7207/* task_group_lock serializes the addition/removal of task groups */ 7208static DEFINE_SPINLOCK(task_group_lock); 7209 7210static void free_sched_group(struct task_group *tg) 7211{ 7212 free_fair_sched_group(tg); 7213 free_rt_sched_group(tg); 7214 autogroup_free(tg); 7215 kfree(tg); 7216} 7217 7218/* allocate runqueue etc for a new task group */ 7219struct task_group *sched_create_group(struct task_group *parent) 7220{ 7221 struct task_group *tg; 7222 7223 tg = kzalloc(sizeof(*tg), GFP_KERNEL); 7224 if (!tg) 7225 return ERR_PTR(-ENOMEM); 7226 7227 if (!alloc_fair_sched_group(tg, parent)) 7228 goto err; 7229 7230 if (!alloc_rt_sched_group(tg, parent)) 7231 goto err; 7232 7233 return tg; 7234 7235err: 7236 free_sched_group(tg); 7237 return ERR_PTR(-ENOMEM); 7238} 7239 7240void sched_online_group(struct task_group *tg, struct task_group *parent) 7241{ 7242 unsigned long flags; 7243 7244 spin_lock_irqsave(&task_group_lock, flags); 7245 list_add_rcu(&tg->list, &task_groups); 7246 7247 WARN_ON(!parent); /* root should already exist */ 7248 7249 tg->parent = parent; 7250 INIT_LIST_HEAD(&tg->children); 7251 list_add_rcu(&tg->siblings, &parent->children); 7252 spin_unlock_irqrestore(&task_group_lock, flags); 7253} 7254 7255/* rcu callback to free various structures associated with a task group */ 7256static void free_sched_group_rcu(struct rcu_head *rhp) 7257{ 7258 /* now it should be safe to free those cfs_rqs */ 7259 free_sched_group(container_of(rhp, struct task_group, rcu)); 7260} 7261 7262/* Destroy runqueue etc associated with a task group */ 7263void sched_destroy_group(struct task_group *tg) 7264{ 7265 /* wait for possible concurrent references to cfs_rqs complete */ 7266 call_rcu(&tg->rcu, free_sched_group_rcu); 7267} 7268 7269void sched_offline_group(struct task_group *tg) 7270{ 7271 unsigned long flags; 7272 int i; 7273 7274 /* end participation in shares distribution */ 7275 for_each_possible_cpu(i) 7276 unregister_fair_sched_group(tg, i); 7277 7278 spin_lock_irqsave(&task_group_lock, flags); 7279 list_del_rcu(&tg->list); 7280 list_del_rcu(&tg->siblings); 7281 spin_unlock_irqrestore(&task_group_lock, flags); 7282} 7283 7284/* change task's runqueue when it moves between groups. 7285 * The caller of this function should have put the task in its new group 7286 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to 7287 * reflect its new group. 7288 */ 7289void sched_move_task(struct task_struct *tsk) 7290{ 7291 struct task_group *tg; 7292 int on_rq, running; 7293 unsigned long flags; 7294 struct rq *rq; 7295 7296 rq = task_rq_lock(tsk, &flags); 7297 7298 running = task_current(rq, tsk); 7299 on_rq = tsk->on_rq; 7300 7301 if (on_rq) 7302 dequeue_task(rq, tsk, 0); 7303 if (unlikely(running)) 7304 tsk->sched_class->put_prev_task(rq, tsk); 7305 7306 tg = container_of(task_subsys_state_check(tsk, cpu_cgroup_subsys_id, 7307 lockdep_is_held(&tsk->sighand->siglock)), 7308 struct task_group, css); 7309 tg = autogroup_task_group(tsk, tg); 7310 tsk->sched_task_group = tg; 7311 7312#ifdef CONFIG_FAIR_GROUP_SCHED 7313 if (tsk->sched_class->task_move_group) 7314 tsk->sched_class->task_move_group(tsk, on_rq); 7315 else 7316#endif 7317 set_task_rq(tsk, task_cpu(tsk)); 7318 7319 if (unlikely(running)) 7320 tsk->sched_class->set_curr_task(rq); 7321 if (on_rq) 7322 enqueue_task(rq, tsk, 0); 7323 7324 task_rq_unlock(rq, tsk, &flags); 7325} 7326#endif /* CONFIG_CGROUP_SCHED */ 7327 7328#if defined(CONFIG_RT_GROUP_SCHED) || defined(CONFIG_CFS_BANDWIDTH) 7329static unsigned long to_ratio(u64 period, u64 runtime) 7330{ 7331 if (runtime == RUNTIME_INF) 7332 return 1ULL << 20; 7333 7334 return div64_u64(runtime << 20, period); 7335} 7336#endif 7337 7338#ifdef CONFIG_RT_GROUP_SCHED 7339/* 7340 * Ensure that the real time constraints are schedulable. 7341 */ 7342static DEFINE_MUTEX(rt_constraints_mutex); 7343 7344/* Must be called with tasklist_lock held */ 7345static inline int tg_has_rt_tasks(struct task_group *tg) 7346{ 7347 struct task_struct *g, *p; 7348 7349 do_each_thread(g, p) { 7350 if (rt_task(p) && task_rq(p)->rt.tg == tg) 7351 return 1; 7352 } while_each_thread(g, p); 7353 7354 return 0; 7355} 7356 7357struct rt_schedulable_data { 7358 struct task_group *tg; 7359 u64 rt_period; 7360 u64 rt_runtime; 7361}; 7362 7363static int tg_rt_schedulable(struct task_group *tg, void *data) 7364{ 7365 struct rt_schedulable_data *d = data; 7366 struct task_group *child; 7367 unsigned long total, sum = 0; 7368 u64 period, runtime; 7369 7370 period = ktime_to_ns(tg->rt_bandwidth.rt_period); 7371 runtime = tg->rt_bandwidth.rt_runtime; 7372 7373 if (tg == d->tg) { 7374 period = d->rt_period; 7375 runtime = d->rt_runtime; 7376 } 7377 7378 /* 7379 * Cannot have more runtime than the period. 7380 */ 7381 if (runtime > period && runtime != RUNTIME_INF) 7382 return -EINVAL; 7383 7384 /* 7385 * Ensure we don't starve existing RT tasks. 7386 */ 7387 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg)) 7388 return -EBUSY; 7389 7390 total = to_ratio(period, runtime); 7391 7392 /* 7393 * Nobody can have more than the global setting allows. 7394 */ 7395 if (total > to_ratio(global_rt_period(), global_rt_runtime())) 7396 return -EINVAL; 7397 7398 /* 7399 * The sum of our children's runtime should not exceed our own. 7400 */ 7401 list_for_each_entry_rcu(child, &tg->children, siblings) { 7402 period = ktime_to_ns(child->rt_bandwidth.rt_period); 7403 runtime = child->rt_bandwidth.rt_runtime; 7404 7405 if (child == d->tg) { 7406 period = d->rt_period; 7407 runtime = d->rt_runtime; 7408 } 7409 7410 sum += to_ratio(period, runtime); 7411 } 7412 7413 if (sum > total) 7414 return -EINVAL; 7415 7416 return 0; 7417} 7418 7419static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime) 7420{ 7421 int ret; 7422 7423 struct rt_schedulable_data data = { 7424 .tg = tg, 7425 .rt_period = period, 7426 .rt_runtime = runtime, 7427 }; 7428 7429 rcu_read_lock(); 7430 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data); 7431 rcu_read_unlock(); 7432 7433 return ret; 7434} 7435 7436static int tg_set_rt_bandwidth(struct task_group *tg, 7437 u64 rt_period, u64 rt_runtime) 7438{ 7439 int i, err = 0; 7440 7441 mutex_lock(&rt_constraints_mutex); 7442 read_lock(&tasklist_lock); 7443 err = __rt_schedulable(tg, rt_period, rt_runtime); 7444 if (err) 7445 goto unlock; 7446 7447 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock); 7448 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period); 7449 tg->rt_bandwidth.rt_runtime = rt_runtime; 7450 7451 for_each_possible_cpu(i) { 7452 struct rt_rq *rt_rq = tg->rt_rq[i]; 7453 7454 raw_spin_lock(&rt_rq->rt_runtime_lock); 7455 rt_rq->rt_runtime = rt_runtime; 7456 raw_spin_unlock(&rt_rq->rt_runtime_lock); 7457 } 7458 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock); 7459unlock: 7460 read_unlock(&tasklist_lock); 7461 mutex_unlock(&rt_constraints_mutex); 7462 7463 return err; 7464} 7465 7466static int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us) 7467{ 7468 u64 rt_runtime, rt_period; 7469 7470 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period); 7471 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC; 7472 if (rt_runtime_us < 0) 7473 rt_runtime = RUNTIME_INF; 7474 7475 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime); 7476} 7477 7478static long sched_group_rt_runtime(struct task_group *tg) 7479{ 7480 u64 rt_runtime_us; 7481 7482 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF) 7483 return -1; 7484 7485 rt_runtime_us = tg->rt_bandwidth.rt_runtime; 7486 do_div(rt_runtime_us, NSEC_PER_USEC); 7487 return rt_runtime_us; 7488} 7489 7490static int sched_group_set_rt_period(struct task_group *tg, long rt_period_us) 7491{ 7492 u64 rt_runtime, rt_period; 7493 7494 rt_period = (u64)rt_period_us * NSEC_PER_USEC; 7495 rt_runtime = tg->rt_bandwidth.rt_runtime; 7496 7497 if (rt_period == 0) 7498 return -EINVAL; 7499 7500 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime); 7501} 7502 7503static long sched_group_rt_period(struct task_group *tg) 7504{ 7505 u64 rt_period_us; 7506 7507 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period); 7508 do_div(rt_period_us, NSEC_PER_USEC); 7509 return rt_period_us; 7510} 7511 7512static int sched_rt_global_constraints(void) 7513{ 7514 u64 runtime, period; 7515 int ret = 0; 7516 7517 if (sysctl_sched_rt_period <= 0) 7518 return -EINVAL; 7519 7520 runtime = global_rt_runtime(); 7521 period = global_rt_period(); 7522 7523 /* 7524 * Sanity check on the sysctl variables. 7525 */ 7526 if (runtime > period && runtime != RUNTIME_INF) 7527 return -EINVAL; 7528 7529 mutex_lock(&rt_constraints_mutex); 7530 read_lock(&tasklist_lock); 7531 ret = __rt_schedulable(NULL, 0, 0); 7532 read_unlock(&tasklist_lock); 7533 mutex_unlock(&rt_constraints_mutex); 7534 7535 return ret; 7536} 7537 7538static int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk) 7539{ 7540 /* Don't accept realtime tasks when there is no way for them to run */ 7541 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0) 7542 return 0; 7543 7544 return 1; 7545} 7546 7547#else /* !CONFIG_RT_GROUP_SCHED */ 7548static int sched_rt_global_constraints(void) 7549{ 7550 unsigned long flags; 7551 int i; 7552 7553 if (sysctl_sched_rt_period <= 0) 7554 return -EINVAL; 7555 7556 /* 7557 * There's always some RT tasks in the root group 7558 * -- migration, kstopmachine etc.. 7559 */ 7560 if (sysctl_sched_rt_runtime == 0) 7561 return -EBUSY; 7562 7563 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags); 7564 for_each_possible_cpu(i) { 7565 struct rt_rq *rt_rq = &cpu_rq(i)->rt; 7566 7567 raw_spin_lock(&rt_rq->rt_runtime_lock); 7568 rt_rq->rt_runtime = global_rt_runtime(); 7569 raw_spin_unlock(&rt_rq->rt_runtime_lock); 7570 } 7571 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags); 7572 7573 return 0; 7574} 7575#endif /* CONFIG_RT_GROUP_SCHED */ 7576 7577int sched_rr_handler(struct ctl_table *table, int write, 7578 void __user *buffer, size_t *lenp, 7579 loff_t *ppos) 7580{ 7581 int ret; 7582 static DEFINE_MUTEX(mutex); 7583 7584 mutex_lock(&mutex); 7585 ret = proc_dointvec(table, write, buffer, lenp, ppos); 7586 /* make sure that internally we keep jiffies */ 7587 /* also, writing zero resets timeslice to default */ 7588 if (!ret && write) { 7589 sched_rr_timeslice = sched_rr_timeslice <= 0 ? 7590 RR_TIMESLICE : msecs_to_jiffies(sched_rr_timeslice); 7591 } 7592 mutex_unlock(&mutex); 7593 return ret; 7594} 7595 7596int sched_rt_handler(struct ctl_table *table, int write, 7597 void __user *buffer, size_t *lenp, 7598 loff_t *ppos) 7599{ 7600 int ret; 7601 int old_period, old_runtime; 7602 static DEFINE_MUTEX(mutex); 7603 7604 mutex_lock(&mutex); 7605 old_period = sysctl_sched_rt_period; 7606 old_runtime = sysctl_sched_rt_runtime; 7607 7608 ret = proc_dointvec(table, write, buffer, lenp, ppos); 7609 7610 if (!ret && write) { 7611 ret = sched_rt_global_constraints(); 7612 if (ret) { 7613 sysctl_sched_rt_period = old_period; 7614 sysctl_sched_rt_runtime = old_runtime; 7615 } else { 7616 def_rt_bandwidth.rt_runtime = global_rt_runtime(); 7617 def_rt_bandwidth.rt_period = 7618 ns_to_ktime(global_rt_period()); 7619 } 7620 } 7621 mutex_unlock(&mutex); 7622 7623 return ret; 7624} 7625 7626#ifdef CONFIG_CGROUP_SCHED 7627 7628/* return corresponding task_group object of a cgroup */ 7629static inline struct task_group *cgroup_tg(struct cgroup *cgrp) 7630{ 7631 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id), 7632 struct task_group, css); 7633} 7634 7635static struct cgroup_subsys_state *cpu_cgroup_css_alloc(struct cgroup *cgrp) 7636{ 7637 struct task_group *tg, *parent; 7638 7639 if (!cgrp->parent) { 7640 /* This is early initialization for the top cgroup */ 7641 return &root_task_group.css; 7642 } 7643 7644 parent = cgroup_tg(cgrp->parent); 7645 tg = sched_create_group(parent); 7646 if (IS_ERR(tg)) 7647 return ERR_PTR(-ENOMEM); 7648 7649 return &tg->css; 7650} 7651 7652static int cpu_cgroup_css_online(struct cgroup *cgrp) 7653{ 7654 struct task_group *tg = cgroup_tg(cgrp); 7655 struct task_group *parent; 7656 7657 if (!cgrp->parent) 7658 return 0; 7659 7660 parent = cgroup_tg(cgrp->parent); 7661 sched_online_group(tg, parent); 7662 return 0; 7663} 7664 7665static void cpu_cgroup_css_free(struct cgroup *cgrp) 7666{ 7667 struct task_group *tg = cgroup_tg(cgrp); 7668 7669 sched_destroy_group(tg); 7670} 7671 7672static void cpu_cgroup_css_offline(struct cgroup *cgrp) 7673{ 7674 struct task_group *tg = cgroup_tg(cgrp); 7675 7676 sched_offline_group(tg); 7677} 7678 7679static int cpu_cgroup_can_attach(struct cgroup *cgrp, 7680 struct cgroup_taskset *tset) 7681{ 7682 struct task_struct *task; 7683 7684 cgroup_taskset_for_each(task, cgrp, tset) { 7685#ifdef CONFIG_RT_GROUP_SCHED 7686 if (!sched_rt_can_attach(cgroup_tg(cgrp), task)) 7687 return -EINVAL; 7688#else 7689 /* We don't support RT-tasks being in separate groups */ 7690 if (task->sched_class != &fair_sched_class) 7691 return -EINVAL; 7692#endif 7693 } 7694 return 0; 7695} 7696 7697static void cpu_cgroup_attach(struct cgroup *cgrp, 7698 struct cgroup_taskset *tset) 7699{ 7700 struct task_struct *task; 7701 7702 cgroup_taskset_for_each(task, cgrp, tset) 7703 sched_move_task(task); 7704} 7705 7706static void 7707cpu_cgroup_exit(struct cgroup *cgrp, struct cgroup *old_cgrp, 7708 struct task_struct *task) 7709{ 7710 /* 7711 * cgroup_exit() is called in the copy_process() failure path. 7712 * Ignore this case since the task hasn't ran yet, this avoids 7713 * trying to poke a half freed task state from generic code. 7714 */ 7715 if (!(task->flags & PF_EXITING)) 7716 return; 7717 7718 sched_move_task(task); 7719} 7720 7721#ifdef CONFIG_FAIR_GROUP_SCHED 7722static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype, 7723 u64 shareval) 7724{ 7725 return sched_group_set_shares(cgroup_tg(cgrp), scale_load(shareval)); 7726} 7727 7728static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft) 7729{ 7730 struct task_group *tg = cgroup_tg(cgrp); 7731 7732 return (u64) scale_load_down(tg->shares); 7733} 7734 7735#ifdef CONFIG_CFS_BANDWIDTH 7736static DEFINE_MUTEX(cfs_constraints_mutex); 7737 7738const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */ 7739const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */ 7740 7741static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime); 7742 7743static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota) 7744{ 7745 int i, ret = 0, runtime_enabled, runtime_was_enabled; 7746 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth; 7747 7748 if (tg == &root_task_group) 7749 return -EINVAL; 7750 7751 /* 7752 * Ensure we have at some amount of bandwidth every period. This is 7753 * to prevent reaching a state of large arrears when throttled via 7754 * entity_tick() resulting in prolonged exit starvation. 7755 */ 7756 if (quota < min_cfs_quota_period || period < min_cfs_quota_period) 7757 return -EINVAL; 7758 7759 /* 7760 * Likewise, bound things on the otherside by preventing insane quota 7761 * periods. This also allows us to normalize in computing quota 7762 * feasibility. 7763 */ 7764 if (period > max_cfs_quota_period) 7765 return -EINVAL; 7766 7767 mutex_lock(&cfs_constraints_mutex); 7768 ret = __cfs_schedulable(tg, period, quota); 7769 if (ret) 7770 goto out_unlock; 7771 7772 runtime_enabled = quota != RUNTIME_INF; 7773 runtime_was_enabled = cfs_b->quota != RUNTIME_INF; 7774 account_cfs_bandwidth_used(runtime_enabled, runtime_was_enabled); 7775 raw_spin_lock_irq(&cfs_b->lock); 7776 cfs_b->period = ns_to_ktime(period); 7777 cfs_b->quota = quota; 7778 7779 __refill_cfs_bandwidth_runtime(cfs_b); 7780 /* restart the period timer (if active) to handle new period expiry */ 7781 if (runtime_enabled && cfs_b->timer_active) { 7782 /* force a reprogram */ 7783 cfs_b->timer_active = 0; 7784 __start_cfs_bandwidth(cfs_b); 7785 } 7786 raw_spin_unlock_irq(&cfs_b->lock); 7787 7788 for_each_possible_cpu(i) { 7789 struct cfs_rq *cfs_rq = tg->cfs_rq[i]; 7790 struct rq *rq = cfs_rq->rq; 7791 7792 raw_spin_lock_irq(&rq->lock); 7793 cfs_rq->runtime_enabled = runtime_enabled; 7794 cfs_rq->runtime_remaining = 0; 7795 7796 if (cfs_rq->throttled) 7797 unthrottle_cfs_rq(cfs_rq); 7798 raw_spin_unlock_irq(&rq->lock); 7799 } 7800out_unlock: 7801 mutex_unlock(&cfs_constraints_mutex); 7802 7803 return ret; 7804} 7805 7806int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us) 7807{ 7808 u64 quota, period; 7809 7810 period = ktime_to_ns(tg->cfs_bandwidth.period); 7811 if (cfs_quota_us < 0) 7812 quota = RUNTIME_INF; 7813 else 7814 quota = (u64)cfs_quota_us * NSEC_PER_USEC; 7815 7816 return tg_set_cfs_bandwidth(tg, period, quota); 7817} 7818 7819long tg_get_cfs_quota(struct task_group *tg) 7820{ 7821 u64 quota_us; 7822 7823 if (tg->cfs_bandwidth.quota == RUNTIME_INF) 7824 return -1; 7825 7826 quota_us = tg->cfs_bandwidth.quota; 7827 do_div(quota_us, NSEC_PER_USEC); 7828 7829 return quota_us; 7830} 7831 7832int tg_set_cfs_period(struct task_group *tg, long cfs_period_us) 7833{ 7834 u64 quota, period; 7835 7836 period = (u64)cfs_period_us * NSEC_PER_USEC; 7837 quota = tg->cfs_bandwidth.quota; 7838 7839 return tg_set_cfs_bandwidth(tg, period, quota); 7840} 7841 7842long tg_get_cfs_period(struct task_group *tg) 7843{ 7844 u64 cfs_period_us; 7845 7846 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period); 7847 do_div(cfs_period_us, NSEC_PER_USEC); 7848 7849 return cfs_period_us; 7850} 7851 7852static s64 cpu_cfs_quota_read_s64(struct cgroup *cgrp, struct cftype *cft) 7853{ 7854 return tg_get_cfs_quota(cgroup_tg(cgrp)); 7855} 7856 7857static int cpu_cfs_quota_write_s64(struct cgroup *cgrp, struct cftype *cftype, 7858 s64 cfs_quota_us) 7859{ 7860 return tg_set_cfs_quota(cgroup_tg(cgrp), cfs_quota_us); 7861} 7862 7863static u64 cpu_cfs_period_read_u64(struct cgroup *cgrp, struct cftype *cft) 7864{ 7865 return tg_get_cfs_period(cgroup_tg(cgrp)); 7866} 7867 7868static int cpu_cfs_period_write_u64(struct cgroup *cgrp, struct cftype *cftype, 7869 u64 cfs_period_us) 7870{ 7871 return tg_set_cfs_period(cgroup_tg(cgrp), cfs_period_us); 7872} 7873 7874struct cfs_schedulable_data { 7875 struct task_group *tg; 7876 u64 period, quota; 7877}; 7878 7879/* 7880 * normalize group quota/period to be quota/max_period 7881 * note: units are usecs 7882 */ 7883static u64 normalize_cfs_quota(struct task_group *tg, 7884 struct cfs_schedulable_data *d) 7885{ 7886 u64 quota, period; 7887 7888 if (tg == d->tg) { 7889 period = d->period; 7890 quota = d->quota; 7891 } else { 7892 period = tg_get_cfs_period(tg); 7893 quota = tg_get_cfs_quota(tg); 7894 } 7895 7896 /* note: these should typically be equivalent */ 7897 if (quota == RUNTIME_INF || quota == -1) 7898 return RUNTIME_INF; 7899 7900 return to_ratio(period, quota); 7901} 7902 7903static int tg_cfs_schedulable_down(struct task_group *tg, void *data) 7904{ 7905 struct cfs_schedulable_data *d = data; 7906 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth; 7907 s64 quota = 0, parent_quota = -1; 7908 7909 if (!tg->parent) { 7910 quota = RUNTIME_INF; 7911 } else { 7912 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth; 7913 7914 quota = normalize_cfs_quota(tg, d); 7915 parent_quota = parent_b->hierarchal_quota; 7916 7917 /* 7918 * ensure max(child_quota) <= parent_quota, inherit when no 7919 * limit is set 7920 */ 7921 if (quota == RUNTIME_INF) 7922 quota = parent_quota; 7923 else if (parent_quota != RUNTIME_INF && quota > parent_quota) 7924 return -EINVAL; 7925 } 7926 cfs_b->hierarchal_quota = quota; 7927 7928 return 0; 7929} 7930 7931static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota) 7932{ 7933 int ret; 7934 struct cfs_schedulable_data data = { 7935 .tg = tg, 7936 .period = period, 7937 .quota = quota, 7938 }; 7939 7940 if (quota != RUNTIME_INF) { 7941 do_div(data.period, NSEC_PER_USEC); 7942 do_div(data.quota, NSEC_PER_USEC); 7943 } 7944 7945 rcu_read_lock(); 7946 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data); 7947 rcu_read_unlock(); 7948 7949 return ret; 7950} 7951 7952static int cpu_stats_show(struct cgroup *cgrp, struct cftype *cft, 7953 struct cgroup_map_cb *cb) 7954{ 7955 struct task_group *tg = cgroup_tg(cgrp); 7956 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth; 7957 7958 cb->fill(cb, "nr_periods", cfs_b->nr_periods); 7959 cb->fill(cb, "nr_throttled", cfs_b->nr_throttled); 7960 cb->fill(cb, "throttled_time", cfs_b->throttled_time); 7961 7962 return 0; 7963} 7964#endif /* CONFIG_CFS_BANDWIDTH */ 7965#endif /* CONFIG_FAIR_GROUP_SCHED */ 7966 7967#ifdef CONFIG_RT_GROUP_SCHED 7968static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft, 7969 s64 val) 7970{ 7971 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val); 7972} 7973 7974static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft) 7975{ 7976 return sched_group_rt_runtime(cgroup_tg(cgrp)); 7977} 7978 7979static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype, 7980 u64 rt_period_us) 7981{ 7982 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us); 7983} 7984 7985static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft) 7986{ 7987 return sched_group_rt_period(cgroup_tg(cgrp)); 7988} 7989#endif /* CONFIG_RT_GROUP_SCHED */ 7990 7991static struct cftype cpu_files[] = { 7992#ifdef CONFIG_FAIR_GROUP_SCHED 7993 { 7994 .name = "shares", 7995 .read_u64 = cpu_shares_read_u64, 7996 .write_u64 = cpu_shares_write_u64, 7997 }, 7998#endif 7999#ifdef CONFIG_CFS_BANDWIDTH 8000 { 8001 .name = "cfs_quota_us", 8002 .read_s64 = cpu_cfs_quota_read_s64, 8003 .write_s64 = cpu_cfs_quota_write_s64, 8004 }, 8005 { 8006 .name = "cfs_period_us", 8007 .read_u64 = cpu_cfs_period_read_u64, 8008 .write_u64 = cpu_cfs_period_write_u64, 8009 }, 8010 { 8011 .name = "stat", 8012 .read_map = cpu_stats_show, 8013 }, 8014#endif 8015#ifdef CONFIG_RT_GROUP_SCHED 8016 { 8017 .name = "rt_runtime_us", 8018 .read_s64 = cpu_rt_runtime_read, 8019 .write_s64 = cpu_rt_runtime_write, 8020 }, 8021 { 8022 .name = "rt_period_us", 8023 .read_u64 = cpu_rt_period_read_uint, 8024 .write_u64 = cpu_rt_period_write_uint, 8025 }, 8026#endif 8027 { } /* terminate */ 8028}; 8029 8030struct cgroup_subsys cpu_cgroup_subsys = { 8031 .name = "cpu", 8032 .css_alloc = cpu_cgroup_css_alloc, 8033 .css_free = cpu_cgroup_css_free, 8034 .css_online = cpu_cgroup_css_online, 8035 .css_offline = cpu_cgroup_css_offline, 8036 .can_attach = cpu_cgroup_can_attach, 8037 .attach = cpu_cgroup_attach, 8038 .exit = cpu_cgroup_exit, 8039 .subsys_id = cpu_cgroup_subsys_id, 8040 .base_cftypes = cpu_files, 8041 .early_init = 1, 8042}; 8043 8044#endif /* CONFIG_CGROUP_SCHED */ 8045 8046#ifdef CONFIG_CGROUP_CPUACCT 8047 8048/* 8049 * CPU accounting code for task groups. 8050 * 8051 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh 8052 * (balbir@in.ibm.com). 8053 */ 8054 8055struct cpuacct root_cpuacct; 8056 8057/* create a new cpu accounting group */ 8058static struct cgroup_subsys_state *cpuacct_css_alloc(struct cgroup *cgrp) 8059{ 8060 struct cpuacct *ca; 8061 8062 if (!cgrp->parent) 8063 return &root_cpuacct.css; 8064 8065 ca = kzalloc(sizeof(*ca), GFP_KERNEL); 8066 if (!ca) 8067 goto out; 8068 8069 ca->cpuusage = alloc_percpu(u64); 8070 if (!ca->cpuusage) 8071 goto out_free_ca; 8072 8073 ca->cpustat = alloc_percpu(struct kernel_cpustat); 8074 if (!ca->cpustat) 8075 goto out_free_cpuusage; 8076 8077 return &ca->css; 8078 8079out_free_cpuusage: 8080 free_percpu(ca->cpuusage); 8081out_free_ca: 8082 kfree(ca); 8083out: 8084 return ERR_PTR(-ENOMEM); 8085} 8086 8087/* destroy an existing cpu accounting group */ 8088static void cpuacct_css_free(struct cgroup *cgrp) 8089{ 8090 struct cpuacct *ca = cgroup_ca(cgrp); 8091 8092 free_percpu(ca->cpustat); 8093 free_percpu(ca->cpuusage); 8094 kfree(ca); 8095} 8096 8097static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu) 8098{ 8099 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu); 8100 u64 data; 8101 8102#ifndef CONFIG_64BIT 8103 /* 8104 * Take rq->lock to make 64-bit read safe on 32-bit platforms. 8105 */ 8106 raw_spin_lock_irq(&cpu_rq(cpu)->lock); 8107 data = *cpuusage; 8108 raw_spin_unlock_irq(&cpu_rq(cpu)->lock); 8109#else 8110 data = *cpuusage; 8111#endif 8112 8113 return data; 8114} 8115 8116static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val) 8117{ 8118 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu); 8119 8120#ifndef CONFIG_64BIT 8121 /* 8122 * Take rq->lock to make 64-bit write safe on 32-bit platforms. 8123 */ 8124 raw_spin_lock_irq(&cpu_rq(cpu)->lock); 8125 *cpuusage = val; 8126 raw_spin_unlock_irq(&cpu_rq(cpu)->lock); 8127#else 8128 *cpuusage = val; 8129#endif 8130} 8131 8132/* return total cpu usage (in nanoseconds) of a group */ 8133static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft) 8134{ 8135 struct cpuacct *ca = cgroup_ca(cgrp); 8136 u64 totalcpuusage = 0; 8137 int i; 8138 8139 for_each_present_cpu(i) 8140 totalcpuusage += cpuacct_cpuusage_read(ca, i); 8141 8142 return totalcpuusage; 8143} 8144 8145static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype, 8146 u64 reset) 8147{ 8148 struct cpuacct *ca = cgroup_ca(cgrp); 8149 int err = 0; 8150 int i; 8151 8152 if (reset) { 8153 err = -EINVAL; 8154 goto out; 8155 } 8156 8157 for_each_present_cpu(i) 8158 cpuacct_cpuusage_write(ca, i, 0); 8159 8160out: 8161 return err; 8162} 8163 8164static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft, 8165 struct seq_file *m) 8166{ 8167 struct cpuacct *ca = cgroup_ca(cgroup); 8168 u64 percpu; 8169 int i; 8170 8171 for_each_present_cpu(i) { 8172 percpu = cpuacct_cpuusage_read(ca, i); 8173 seq_printf(m, "%llu ", (unsigned long long) percpu); 8174 } 8175 seq_printf(m, "\n"); 8176 return 0; 8177} 8178 8179static const char *cpuacct_stat_desc[] = { 8180 [CPUACCT_STAT_USER] = "user", 8181 [CPUACCT_STAT_SYSTEM] = "system", 8182}; 8183 8184static int cpuacct_stats_show(struct cgroup *cgrp, struct cftype *cft, 8185 struct cgroup_map_cb *cb) 8186{ 8187 struct cpuacct *ca = cgroup_ca(cgrp); 8188 int cpu; 8189 s64 val = 0; 8190 8191 for_each_online_cpu(cpu) { 8192 struct kernel_cpustat *kcpustat = per_cpu_ptr(ca->cpustat, cpu); 8193 val += kcpustat->cpustat[CPUTIME_USER]; 8194 val += kcpustat->cpustat[CPUTIME_NICE]; 8195 } 8196 val = cputime64_to_clock_t(val); 8197 cb->fill(cb, cpuacct_stat_desc[CPUACCT_STAT_USER], val); 8198 8199 val = 0; 8200 for_each_online_cpu(cpu) { 8201 struct kernel_cpustat *kcpustat = per_cpu_ptr(ca->cpustat, cpu); 8202 val += kcpustat->cpustat[CPUTIME_SYSTEM]; 8203 val += kcpustat->cpustat[CPUTIME_IRQ]; 8204 val += kcpustat->cpustat[CPUTIME_SOFTIRQ]; 8205 } 8206 8207 val = cputime64_to_clock_t(val); 8208 cb->fill(cb, cpuacct_stat_desc[CPUACCT_STAT_SYSTEM], val); 8209 8210 return 0; 8211} 8212 8213static struct cftype files[] = { 8214 { 8215 .name = "usage", 8216 .read_u64 = cpuusage_read, 8217 .write_u64 = cpuusage_write, 8218 }, 8219 { 8220 .name = "usage_percpu", 8221 .read_seq_string = cpuacct_percpu_seq_read, 8222 }, 8223 { 8224 .name = "stat", 8225 .read_map = cpuacct_stats_show, 8226 }, 8227 { } /* terminate */ 8228}; 8229 8230/* 8231 * charge this task's execution time to its accounting group. 8232 * 8233 * called with rq->lock held. 8234 */ 8235void cpuacct_charge(struct task_struct *tsk, u64 cputime) 8236{ 8237 struct cpuacct *ca; 8238 int cpu; 8239 8240 if (unlikely(!cpuacct_subsys.active)) 8241 return; 8242 8243 cpu = task_cpu(tsk); 8244 8245 rcu_read_lock(); 8246 8247 ca = task_ca(tsk); 8248 8249 for (; ca; ca = parent_ca(ca)) { 8250 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu); 8251 *cpuusage += cputime; 8252 } 8253 8254 rcu_read_unlock(); 8255} 8256 8257struct cgroup_subsys cpuacct_subsys = { 8258 .name = "cpuacct", 8259 .css_alloc = cpuacct_css_alloc, 8260 .css_free = cpuacct_css_free, 8261 .subsys_id = cpuacct_subsys_id, 8262 .base_cftypes = files, 8263}; 8264#endif /* CONFIG_CGROUP_CPUACCT */ 8265 8266void dump_cpu_task(int cpu) 8267{ 8268 pr_info("Task dump for CPU %d:\n", cpu); 8269 sched_show_task(cpu_curr(cpu)); 8270} 8271