core.c revision 686855f5d833178e518d79e7912cdb3268a9fa69
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_sched.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 const struct cpumask *nodemask = cpumask_of_node(cpu_to_node(cpu)); 1136 enum { cpuset, possible, fail } state = cpuset; 1137 int dest_cpu; 1138 1139 /* Look for allowed, online CPU in same node. */ 1140 for_each_cpu(dest_cpu, nodemask) { 1141 if (!cpu_online(dest_cpu)) 1142 continue; 1143 if (!cpu_active(dest_cpu)) 1144 continue; 1145 if (cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p))) 1146 return dest_cpu; 1147 } 1148 1149 for (;;) { 1150 /* Any allowed, online CPU? */ 1151 for_each_cpu(dest_cpu, tsk_cpus_allowed(p)) { 1152 if (!cpu_online(dest_cpu)) 1153 continue; 1154 if (!cpu_active(dest_cpu)) 1155 continue; 1156 goto out; 1157 } 1158 1159 switch (state) { 1160 case cpuset: 1161 /* No more Mr. Nice Guy. */ 1162 cpuset_cpus_allowed_fallback(p); 1163 state = possible; 1164 break; 1165 1166 case possible: 1167 do_set_cpus_allowed(p, cpu_possible_mask); 1168 state = fail; 1169 break; 1170 1171 case fail: 1172 BUG(); 1173 break; 1174 } 1175 } 1176 1177out: 1178 if (state != cpuset) { 1179 /* 1180 * Don't tell them about moving exiting tasks or 1181 * kernel threads (both mm NULL), since they never 1182 * leave kernel. 1183 */ 1184 if (p->mm && printk_ratelimit()) { 1185 printk_sched("process %d (%s) no longer affine to cpu%d\n", 1186 task_pid_nr(p), p->comm, cpu); 1187 } 1188 } 1189 1190 return dest_cpu; 1191} 1192 1193/* 1194 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable. 1195 */ 1196static inline 1197int select_task_rq(struct task_struct *p, int sd_flags, int wake_flags) 1198{ 1199 int cpu = p->sched_class->select_task_rq(p, sd_flags, wake_flags); 1200 1201 /* 1202 * In order not to call set_task_cpu() on a blocking task we need 1203 * to rely on ttwu() to place the task on a valid ->cpus_allowed 1204 * cpu. 1205 * 1206 * Since this is common to all placement strategies, this lives here. 1207 * 1208 * [ this allows ->select_task() to simply return task_cpu(p) and 1209 * not worry about this generic constraint ] 1210 */ 1211 if (unlikely(!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)) || 1212 !cpu_online(cpu))) 1213 cpu = select_fallback_rq(task_cpu(p), p); 1214 1215 return cpu; 1216} 1217 1218static void update_avg(u64 *avg, u64 sample) 1219{ 1220 s64 diff = sample - *avg; 1221 *avg += diff >> 3; 1222} 1223#endif 1224 1225static void 1226ttwu_stat(struct task_struct *p, int cpu, int wake_flags) 1227{ 1228#ifdef CONFIG_SCHEDSTATS 1229 struct rq *rq = this_rq(); 1230 1231#ifdef CONFIG_SMP 1232 int this_cpu = smp_processor_id(); 1233 1234 if (cpu == this_cpu) { 1235 schedstat_inc(rq, ttwu_local); 1236 schedstat_inc(p, se.statistics.nr_wakeups_local); 1237 } else { 1238 struct sched_domain *sd; 1239 1240 schedstat_inc(p, se.statistics.nr_wakeups_remote); 1241 rcu_read_lock(); 1242 for_each_domain(this_cpu, sd) { 1243 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) { 1244 schedstat_inc(sd, ttwu_wake_remote); 1245 break; 1246 } 1247 } 1248 rcu_read_unlock(); 1249 } 1250 1251 if (wake_flags & WF_MIGRATED) 1252 schedstat_inc(p, se.statistics.nr_wakeups_migrate); 1253 1254#endif /* CONFIG_SMP */ 1255 1256 schedstat_inc(rq, ttwu_count); 1257 schedstat_inc(p, se.statistics.nr_wakeups); 1258 1259 if (wake_flags & WF_SYNC) 1260 schedstat_inc(p, se.statistics.nr_wakeups_sync); 1261 1262#endif /* CONFIG_SCHEDSTATS */ 1263} 1264 1265static void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags) 1266{ 1267 activate_task(rq, p, en_flags); 1268 p->on_rq = 1; 1269 1270 /* if a worker is waking up, notify workqueue */ 1271 if (p->flags & PF_WQ_WORKER) 1272 wq_worker_waking_up(p, cpu_of(rq)); 1273} 1274 1275/* 1276 * Mark the task runnable and perform wakeup-preemption. 1277 */ 1278static void 1279ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags) 1280{ 1281 trace_sched_wakeup(p, true); 1282 check_preempt_curr(rq, p, wake_flags); 1283 1284 p->state = TASK_RUNNING; 1285#ifdef CONFIG_SMP 1286 if (p->sched_class->task_woken) 1287 p->sched_class->task_woken(rq, p); 1288 1289 if (rq->idle_stamp) { 1290 u64 delta = rq->clock - rq->idle_stamp; 1291 u64 max = 2*sysctl_sched_migration_cost; 1292 1293 if (delta > max) 1294 rq->avg_idle = max; 1295 else 1296 update_avg(&rq->avg_idle, delta); 1297 rq->idle_stamp = 0; 1298 } 1299#endif 1300} 1301 1302static void 1303ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags) 1304{ 1305#ifdef CONFIG_SMP 1306 if (p->sched_contributes_to_load) 1307 rq->nr_uninterruptible--; 1308#endif 1309 1310 ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_WAKING); 1311 ttwu_do_wakeup(rq, p, wake_flags); 1312} 1313 1314/* 1315 * Called in case the task @p isn't fully descheduled from its runqueue, 1316 * in this case we must do a remote wakeup. Its a 'light' wakeup though, 1317 * since all we need to do is flip p->state to TASK_RUNNING, since 1318 * the task is still ->on_rq. 1319 */ 1320static int ttwu_remote(struct task_struct *p, int wake_flags) 1321{ 1322 struct rq *rq; 1323 int ret = 0; 1324 1325 rq = __task_rq_lock(p); 1326 if (p->on_rq) { 1327 ttwu_do_wakeup(rq, p, wake_flags); 1328 ret = 1; 1329 } 1330 __task_rq_unlock(rq); 1331 1332 return ret; 1333} 1334 1335#ifdef CONFIG_SMP 1336static void sched_ttwu_pending(void) 1337{ 1338 struct rq *rq = this_rq(); 1339 struct llist_node *llist = llist_del_all(&rq->wake_list); 1340 struct task_struct *p; 1341 1342 raw_spin_lock(&rq->lock); 1343 1344 while (llist) { 1345 p = llist_entry(llist, struct task_struct, wake_entry); 1346 llist = llist_next(llist); 1347 ttwu_do_activate(rq, p, 0); 1348 } 1349 1350 raw_spin_unlock(&rq->lock); 1351} 1352 1353void scheduler_ipi(void) 1354{ 1355 if (llist_empty(&this_rq()->wake_list) && !got_nohz_idle_kick()) 1356 return; 1357 1358 /* 1359 * Not all reschedule IPI handlers call irq_enter/irq_exit, since 1360 * traditionally all their work was done from the interrupt return 1361 * path. Now that we actually do some work, we need to make sure 1362 * we do call them. 1363 * 1364 * Some archs already do call them, luckily irq_enter/exit nest 1365 * properly. 1366 * 1367 * Arguably we should visit all archs and update all handlers, 1368 * however a fair share of IPIs are still resched only so this would 1369 * somewhat pessimize the simple resched case. 1370 */ 1371 irq_enter(); 1372 sched_ttwu_pending(); 1373 1374 /* 1375 * Check if someone kicked us for doing the nohz idle load balance. 1376 */ 1377 if (unlikely(got_nohz_idle_kick() && !need_resched())) { 1378 this_rq()->idle_balance = 1; 1379 raise_softirq_irqoff(SCHED_SOFTIRQ); 1380 } 1381 irq_exit(); 1382} 1383 1384static void ttwu_queue_remote(struct task_struct *p, int cpu) 1385{ 1386 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list)) 1387 smp_send_reschedule(cpu); 1388} 1389 1390bool cpus_share_cache(int this_cpu, int that_cpu) 1391{ 1392 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu); 1393} 1394#endif /* CONFIG_SMP */ 1395 1396static void ttwu_queue(struct task_struct *p, int cpu) 1397{ 1398 struct rq *rq = cpu_rq(cpu); 1399 1400#if defined(CONFIG_SMP) 1401 if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) { 1402 sched_clock_cpu(cpu); /* sync clocks x-cpu */ 1403 ttwu_queue_remote(p, cpu); 1404 return; 1405 } 1406#endif 1407 1408 raw_spin_lock(&rq->lock); 1409 ttwu_do_activate(rq, p, 0); 1410 raw_spin_unlock(&rq->lock); 1411} 1412 1413/** 1414 * try_to_wake_up - wake up a thread 1415 * @p: the thread to be awakened 1416 * @state: the mask of task states that can be woken 1417 * @wake_flags: wake modifier flags (WF_*) 1418 * 1419 * Put it on the run-queue if it's not already there. The "current" 1420 * thread is always on the run-queue (except when the actual 1421 * re-schedule is in progress), and as such you're allowed to do 1422 * the simpler "current->state = TASK_RUNNING" to mark yourself 1423 * runnable without the overhead of this. 1424 * 1425 * Returns %true if @p was woken up, %false if it was already running 1426 * or @state didn't match @p's state. 1427 */ 1428static int 1429try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags) 1430{ 1431 unsigned long flags; 1432 int cpu, success = 0; 1433 1434 smp_wmb(); 1435 raw_spin_lock_irqsave(&p->pi_lock, flags); 1436 if (!(p->state & state)) 1437 goto out; 1438 1439 success = 1; /* we're going to change ->state */ 1440 cpu = task_cpu(p); 1441 1442 if (p->on_rq && ttwu_remote(p, wake_flags)) 1443 goto stat; 1444 1445#ifdef CONFIG_SMP 1446 /* 1447 * If the owning (remote) cpu is still in the middle of schedule() with 1448 * this task as prev, wait until its done referencing the task. 1449 */ 1450 while (p->on_cpu) 1451 cpu_relax(); 1452 /* 1453 * Pairs with the smp_wmb() in finish_lock_switch(). 1454 */ 1455 smp_rmb(); 1456 1457 p->sched_contributes_to_load = !!task_contributes_to_load(p); 1458 p->state = TASK_WAKING; 1459 1460 if (p->sched_class->task_waking) 1461 p->sched_class->task_waking(p); 1462 1463 cpu = select_task_rq(p, SD_BALANCE_WAKE, wake_flags); 1464 if (task_cpu(p) != cpu) { 1465 wake_flags |= WF_MIGRATED; 1466 set_task_cpu(p, cpu); 1467 } 1468#endif /* CONFIG_SMP */ 1469 1470 ttwu_queue(p, cpu); 1471stat: 1472 ttwu_stat(p, cpu, wake_flags); 1473out: 1474 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 1475 1476 return success; 1477} 1478 1479/** 1480 * try_to_wake_up_local - try to wake up a local task with rq lock held 1481 * @p: the thread to be awakened 1482 * 1483 * Put @p on the run-queue if it's not already there. The caller must 1484 * ensure that this_rq() is locked, @p is bound to this_rq() and not 1485 * the current task. 1486 */ 1487static void try_to_wake_up_local(struct task_struct *p) 1488{ 1489 struct rq *rq = task_rq(p); 1490 1491 BUG_ON(rq != this_rq()); 1492 BUG_ON(p == current); 1493 lockdep_assert_held(&rq->lock); 1494 1495 if (!raw_spin_trylock(&p->pi_lock)) { 1496 raw_spin_unlock(&rq->lock); 1497 raw_spin_lock(&p->pi_lock); 1498 raw_spin_lock(&rq->lock); 1499 } 1500 1501 if (!(p->state & TASK_NORMAL)) 1502 goto out; 1503 1504 if (!p->on_rq) 1505 ttwu_activate(rq, p, ENQUEUE_WAKEUP); 1506 1507 ttwu_do_wakeup(rq, p, 0); 1508 ttwu_stat(p, smp_processor_id(), 0); 1509out: 1510 raw_spin_unlock(&p->pi_lock); 1511} 1512 1513/** 1514 * wake_up_process - Wake up a specific process 1515 * @p: The process to be woken up. 1516 * 1517 * Attempt to wake up the nominated process and move it to the set of runnable 1518 * processes. Returns 1 if the process was woken up, 0 if it was already 1519 * running. 1520 * 1521 * It may be assumed that this function implies a write memory barrier before 1522 * changing the task state if and only if any tasks are woken up. 1523 */ 1524int wake_up_process(struct task_struct *p) 1525{ 1526 return try_to_wake_up(p, TASK_ALL, 0); 1527} 1528EXPORT_SYMBOL(wake_up_process); 1529 1530int wake_up_state(struct task_struct *p, unsigned int state) 1531{ 1532 return try_to_wake_up(p, state, 0); 1533} 1534 1535/* 1536 * Perform scheduler related setup for a newly forked process p. 1537 * p is forked by current. 1538 * 1539 * __sched_fork() is basic setup used by init_idle() too: 1540 */ 1541static void __sched_fork(struct task_struct *p) 1542{ 1543 p->on_rq = 0; 1544 1545 p->se.on_rq = 0; 1546 p->se.exec_start = 0; 1547 p->se.sum_exec_runtime = 0; 1548 p->se.prev_sum_exec_runtime = 0; 1549 p->se.nr_migrations = 0; 1550 p->se.vruntime = 0; 1551 INIT_LIST_HEAD(&p->se.group_node); 1552 1553/* 1554 * Load-tracking only depends on SMP, FAIR_GROUP_SCHED dependency below may be 1555 * removed when useful for applications beyond shares distribution (e.g. 1556 * load-balance). 1557 */ 1558#if defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED) 1559 p->se.avg.runnable_avg_period = 0; 1560 p->se.avg.runnable_avg_sum = 0; 1561#endif 1562#ifdef CONFIG_SCHEDSTATS 1563 memset(&p->se.statistics, 0, sizeof(p->se.statistics)); 1564#endif 1565 1566 INIT_LIST_HEAD(&p->rt.run_list); 1567 1568#ifdef CONFIG_PREEMPT_NOTIFIERS 1569 INIT_HLIST_HEAD(&p->preempt_notifiers); 1570#endif 1571 1572#ifdef CONFIG_NUMA_BALANCING 1573 if (p->mm && atomic_read(&p->mm->mm_users) == 1) { 1574 p->mm->numa_next_scan = jiffies; 1575 p->mm->numa_next_reset = jiffies; 1576 p->mm->numa_scan_seq = 0; 1577 } 1578 1579 p->node_stamp = 0ULL; 1580 p->numa_scan_seq = p->mm ? p->mm->numa_scan_seq : 0; 1581 p->numa_migrate_seq = p->mm ? p->mm->numa_scan_seq - 1 : 0; 1582 p->numa_scan_period = sysctl_numa_balancing_scan_delay; 1583 p->numa_work.next = &p->numa_work; 1584#endif /* CONFIG_NUMA_BALANCING */ 1585} 1586 1587#ifdef CONFIG_NUMA_BALANCING 1588#ifdef CONFIG_SCHED_DEBUG 1589void set_numabalancing_state(bool enabled) 1590{ 1591 if (enabled) 1592 sched_feat_set("NUMA"); 1593 else 1594 sched_feat_set("NO_NUMA"); 1595} 1596#else 1597__read_mostly bool numabalancing_enabled; 1598 1599void set_numabalancing_state(bool enabled) 1600{ 1601 numabalancing_enabled = enabled; 1602} 1603#endif /* CONFIG_SCHED_DEBUG */ 1604#endif /* CONFIG_NUMA_BALANCING */ 1605 1606/* 1607 * fork()/clone()-time setup: 1608 */ 1609void sched_fork(struct task_struct *p) 1610{ 1611 unsigned long flags; 1612 int cpu = get_cpu(); 1613 1614 __sched_fork(p); 1615 /* 1616 * We mark the process as running here. This guarantees that 1617 * nobody will actually run it, and a signal or other external 1618 * event cannot wake it up and insert it on the runqueue either. 1619 */ 1620 p->state = TASK_RUNNING; 1621 1622 /* 1623 * Make sure we do not leak PI boosting priority to the child. 1624 */ 1625 p->prio = current->normal_prio; 1626 1627 /* 1628 * Revert to default priority/policy on fork if requested. 1629 */ 1630 if (unlikely(p->sched_reset_on_fork)) { 1631 if (task_has_rt_policy(p)) { 1632 p->policy = SCHED_NORMAL; 1633 p->static_prio = NICE_TO_PRIO(0); 1634 p->rt_priority = 0; 1635 } else if (PRIO_TO_NICE(p->static_prio) < 0) 1636 p->static_prio = NICE_TO_PRIO(0); 1637 1638 p->prio = p->normal_prio = __normal_prio(p); 1639 set_load_weight(p); 1640 1641 /* 1642 * We don't need the reset flag anymore after the fork. It has 1643 * fulfilled its duty: 1644 */ 1645 p->sched_reset_on_fork = 0; 1646 } 1647 1648 if (!rt_prio(p->prio)) 1649 p->sched_class = &fair_sched_class; 1650 1651 if (p->sched_class->task_fork) 1652 p->sched_class->task_fork(p); 1653 1654 /* 1655 * The child is not yet in the pid-hash so no cgroup attach races, 1656 * and the cgroup is pinned to this child due to cgroup_fork() 1657 * is ran before sched_fork(). 1658 * 1659 * Silence PROVE_RCU. 1660 */ 1661 raw_spin_lock_irqsave(&p->pi_lock, flags); 1662 set_task_cpu(p, cpu); 1663 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 1664 1665#if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT) 1666 if (likely(sched_info_on())) 1667 memset(&p->sched_info, 0, sizeof(p->sched_info)); 1668#endif 1669#if defined(CONFIG_SMP) 1670 p->on_cpu = 0; 1671#endif 1672#ifdef CONFIG_PREEMPT_COUNT 1673 /* Want to start with kernel preemption disabled. */ 1674 task_thread_info(p)->preempt_count = 1; 1675#endif 1676#ifdef CONFIG_SMP 1677 plist_node_init(&p->pushable_tasks, MAX_PRIO); 1678#endif 1679 1680 put_cpu(); 1681} 1682 1683/* 1684 * wake_up_new_task - wake up a newly created task for the first time. 1685 * 1686 * This function will do some initial scheduler statistics housekeeping 1687 * that must be done for every newly created context, then puts the task 1688 * on the runqueue and wakes it. 1689 */ 1690void wake_up_new_task(struct task_struct *p) 1691{ 1692 unsigned long flags; 1693 struct rq *rq; 1694 1695 raw_spin_lock_irqsave(&p->pi_lock, flags); 1696#ifdef CONFIG_SMP 1697 /* 1698 * Fork balancing, do it here and not earlier because: 1699 * - cpus_allowed can change in the fork path 1700 * - any previously selected cpu might disappear through hotplug 1701 */ 1702 set_task_cpu(p, select_task_rq(p, SD_BALANCE_FORK, 0)); 1703#endif 1704 1705 rq = __task_rq_lock(p); 1706 activate_task(rq, p, 0); 1707 p->on_rq = 1; 1708 trace_sched_wakeup_new(p, true); 1709 check_preempt_curr(rq, p, WF_FORK); 1710#ifdef CONFIG_SMP 1711 if (p->sched_class->task_woken) 1712 p->sched_class->task_woken(rq, p); 1713#endif 1714 task_rq_unlock(rq, p, &flags); 1715} 1716 1717#ifdef CONFIG_PREEMPT_NOTIFIERS 1718 1719/** 1720 * preempt_notifier_register - tell me when current is being preempted & rescheduled 1721 * @notifier: notifier struct to register 1722 */ 1723void preempt_notifier_register(struct preempt_notifier *notifier) 1724{ 1725 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers); 1726} 1727EXPORT_SYMBOL_GPL(preempt_notifier_register); 1728 1729/** 1730 * preempt_notifier_unregister - no longer interested in preemption notifications 1731 * @notifier: notifier struct to unregister 1732 * 1733 * This is safe to call from within a preemption notifier. 1734 */ 1735void preempt_notifier_unregister(struct preempt_notifier *notifier) 1736{ 1737 hlist_del(¬ifier->link); 1738} 1739EXPORT_SYMBOL_GPL(preempt_notifier_unregister); 1740 1741static void fire_sched_in_preempt_notifiers(struct task_struct *curr) 1742{ 1743 struct preempt_notifier *notifier; 1744 struct hlist_node *node; 1745 1746 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link) 1747 notifier->ops->sched_in(notifier, raw_smp_processor_id()); 1748} 1749 1750static void 1751fire_sched_out_preempt_notifiers(struct task_struct *curr, 1752 struct task_struct *next) 1753{ 1754 struct preempt_notifier *notifier; 1755 struct hlist_node *node; 1756 1757 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link) 1758 notifier->ops->sched_out(notifier, next); 1759} 1760 1761#else /* !CONFIG_PREEMPT_NOTIFIERS */ 1762 1763static void fire_sched_in_preempt_notifiers(struct task_struct *curr) 1764{ 1765} 1766 1767static void 1768fire_sched_out_preempt_notifiers(struct task_struct *curr, 1769 struct task_struct *next) 1770{ 1771} 1772 1773#endif /* CONFIG_PREEMPT_NOTIFIERS */ 1774 1775/** 1776 * prepare_task_switch - prepare to switch tasks 1777 * @rq: the runqueue preparing to switch 1778 * @prev: the current task that is being switched out 1779 * @next: the task we are going to switch to. 1780 * 1781 * This is called with the rq lock held and interrupts off. It must 1782 * be paired with a subsequent finish_task_switch after the context 1783 * switch. 1784 * 1785 * prepare_task_switch sets up locking and calls architecture specific 1786 * hooks. 1787 */ 1788static inline void 1789prepare_task_switch(struct rq *rq, struct task_struct *prev, 1790 struct task_struct *next) 1791{ 1792 trace_sched_switch(prev, next); 1793 sched_info_switch(prev, next); 1794 perf_event_task_sched_out(prev, next); 1795 fire_sched_out_preempt_notifiers(prev, next); 1796 prepare_lock_switch(rq, next); 1797 prepare_arch_switch(next); 1798} 1799 1800/** 1801 * finish_task_switch - clean up after a task-switch 1802 * @rq: runqueue associated with task-switch 1803 * @prev: the thread we just switched away from. 1804 * 1805 * finish_task_switch must be called after the context switch, paired 1806 * with a prepare_task_switch call before the context switch. 1807 * finish_task_switch will reconcile locking set up by prepare_task_switch, 1808 * and do any other architecture-specific cleanup actions. 1809 * 1810 * Note that we may have delayed dropping an mm in context_switch(). If 1811 * so, we finish that here outside of the runqueue lock. (Doing it 1812 * with the lock held can cause deadlocks; see schedule() for 1813 * details.) 1814 */ 1815static void finish_task_switch(struct rq *rq, struct task_struct *prev) 1816 __releases(rq->lock) 1817{ 1818 struct mm_struct *mm = rq->prev_mm; 1819 long prev_state; 1820 1821 rq->prev_mm = NULL; 1822 1823 /* 1824 * A task struct has one reference for the use as "current". 1825 * If a task dies, then it sets TASK_DEAD in tsk->state and calls 1826 * schedule one last time. The schedule call will never return, and 1827 * the scheduled task must drop that reference. 1828 * The test for TASK_DEAD must occur while the runqueue locks are 1829 * still held, otherwise prev could be scheduled on another cpu, die 1830 * there before we look at prev->state, and then the reference would 1831 * be dropped twice. 1832 * Manfred Spraul <manfred@colorfullife.com> 1833 */ 1834 prev_state = prev->state; 1835 vtime_task_switch(prev); 1836 finish_arch_switch(prev); 1837 perf_event_task_sched_in(prev, current); 1838 finish_lock_switch(rq, prev); 1839 finish_arch_post_lock_switch(); 1840 1841 fire_sched_in_preempt_notifiers(current); 1842 if (mm) 1843 mmdrop(mm); 1844 if (unlikely(prev_state == TASK_DEAD)) { 1845 /* 1846 * Remove function-return probe instances associated with this 1847 * task and put them back on the free list. 1848 */ 1849 kprobe_flush_task(prev); 1850 put_task_struct(prev); 1851 } 1852} 1853 1854#ifdef CONFIG_SMP 1855 1856/* assumes rq->lock is held */ 1857static inline void pre_schedule(struct rq *rq, struct task_struct *prev) 1858{ 1859 if (prev->sched_class->pre_schedule) 1860 prev->sched_class->pre_schedule(rq, prev); 1861} 1862 1863/* rq->lock is NOT held, but preemption is disabled */ 1864static inline void post_schedule(struct rq *rq) 1865{ 1866 if (rq->post_schedule) { 1867 unsigned long flags; 1868 1869 raw_spin_lock_irqsave(&rq->lock, flags); 1870 if (rq->curr->sched_class->post_schedule) 1871 rq->curr->sched_class->post_schedule(rq); 1872 raw_spin_unlock_irqrestore(&rq->lock, flags); 1873 1874 rq->post_schedule = 0; 1875 } 1876} 1877 1878#else 1879 1880static inline void pre_schedule(struct rq *rq, struct task_struct *p) 1881{ 1882} 1883 1884static inline void post_schedule(struct rq *rq) 1885{ 1886} 1887 1888#endif 1889 1890/** 1891 * schedule_tail - first thing a freshly forked thread must call. 1892 * @prev: the thread we just switched away from. 1893 */ 1894asmlinkage void schedule_tail(struct task_struct *prev) 1895 __releases(rq->lock) 1896{ 1897 struct rq *rq = this_rq(); 1898 1899 finish_task_switch(rq, prev); 1900 1901 /* 1902 * FIXME: do we need to worry about rq being invalidated by the 1903 * task_switch? 1904 */ 1905 post_schedule(rq); 1906 1907#ifdef __ARCH_WANT_UNLOCKED_CTXSW 1908 /* In this case, finish_task_switch does not reenable preemption */ 1909 preempt_enable(); 1910#endif 1911 if (current->set_child_tid) 1912 put_user(task_pid_vnr(current), current->set_child_tid); 1913} 1914 1915/* 1916 * context_switch - switch to the new MM and the new 1917 * thread's register state. 1918 */ 1919static inline void 1920context_switch(struct rq *rq, struct task_struct *prev, 1921 struct task_struct *next) 1922{ 1923 struct mm_struct *mm, *oldmm; 1924 1925 prepare_task_switch(rq, prev, next); 1926 1927 mm = next->mm; 1928 oldmm = prev->active_mm; 1929 /* 1930 * For paravirt, this is coupled with an exit in switch_to to 1931 * combine the page table reload and the switch backend into 1932 * one hypercall. 1933 */ 1934 arch_start_context_switch(prev); 1935 1936 if (!mm) { 1937 next->active_mm = oldmm; 1938 atomic_inc(&oldmm->mm_count); 1939 enter_lazy_tlb(oldmm, next); 1940 } else 1941 switch_mm(oldmm, mm, next); 1942 1943 if (!prev->mm) { 1944 prev->active_mm = NULL; 1945 rq->prev_mm = oldmm; 1946 } 1947 /* 1948 * Since the runqueue lock will be released by the next 1949 * task (which is an invalid locking op but in the case 1950 * of the scheduler it's an obvious special-case), so we 1951 * do an early lockdep release here: 1952 */ 1953#ifndef __ARCH_WANT_UNLOCKED_CTXSW 1954 spin_release(&rq->lock.dep_map, 1, _THIS_IP_); 1955#endif 1956 1957 context_tracking_task_switch(prev, next); 1958 /* Here we just switch the register state and the stack. */ 1959 switch_to(prev, next, prev); 1960 1961 barrier(); 1962 /* 1963 * this_rq must be evaluated again because prev may have moved 1964 * CPUs since it called schedule(), thus the 'rq' on its stack 1965 * frame will be invalid. 1966 */ 1967 finish_task_switch(this_rq(), prev); 1968} 1969 1970/* 1971 * nr_running, nr_uninterruptible and nr_context_switches: 1972 * 1973 * externally visible scheduler statistics: current number of runnable 1974 * threads, current number of uninterruptible-sleeping threads, total 1975 * number of context switches performed since bootup. 1976 */ 1977unsigned long nr_running(void) 1978{ 1979 unsigned long i, sum = 0; 1980 1981 for_each_online_cpu(i) 1982 sum += cpu_rq(i)->nr_running; 1983 1984 return sum; 1985} 1986 1987unsigned long nr_uninterruptible(void) 1988{ 1989 unsigned long i, sum = 0; 1990 1991 for_each_possible_cpu(i) 1992 sum += cpu_rq(i)->nr_uninterruptible; 1993 1994 /* 1995 * Since we read the counters lockless, it might be slightly 1996 * inaccurate. Do not allow it to go below zero though: 1997 */ 1998 if (unlikely((long)sum < 0)) 1999 sum = 0; 2000 2001 return sum; 2002} 2003 2004unsigned long long nr_context_switches(void) 2005{ 2006 int i; 2007 unsigned long long sum = 0; 2008 2009 for_each_possible_cpu(i) 2010 sum += cpu_rq(i)->nr_switches; 2011 2012 return sum; 2013} 2014 2015unsigned long nr_iowait(void) 2016{ 2017 unsigned long i, sum = 0; 2018 2019 for_each_possible_cpu(i) 2020 sum += atomic_read(&cpu_rq(i)->nr_iowait); 2021 2022 return sum; 2023} 2024 2025unsigned long nr_iowait_cpu(int cpu) 2026{ 2027 struct rq *this = cpu_rq(cpu); 2028 return atomic_read(&this->nr_iowait); 2029} 2030 2031unsigned long this_cpu_load(void) 2032{ 2033 struct rq *this = this_rq(); 2034 return this->cpu_load[0]; 2035} 2036 2037 2038/* 2039 * Global load-average calculations 2040 * 2041 * We take a distributed and async approach to calculating the global load-avg 2042 * in order to minimize overhead. 2043 * 2044 * The global load average is an exponentially decaying average of nr_running + 2045 * nr_uninterruptible. 2046 * 2047 * Once every LOAD_FREQ: 2048 * 2049 * nr_active = 0; 2050 * for_each_possible_cpu(cpu) 2051 * nr_active += cpu_of(cpu)->nr_running + cpu_of(cpu)->nr_uninterruptible; 2052 * 2053 * avenrun[n] = avenrun[0] * exp_n + nr_active * (1 - exp_n) 2054 * 2055 * Due to a number of reasons the above turns in the mess below: 2056 * 2057 * - for_each_possible_cpu() is prohibitively expensive on machines with 2058 * serious number of cpus, therefore we need to take a distributed approach 2059 * to calculating nr_active. 2060 * 2061 * \Sum_i x_i(t) = \Sum_i x_i(t) - x_i(t_0) | x_i(t_0) := 0 2062 * = \Sum_i { \Sum_j=1 x_i(t_j) - x_i(t_j-1) } 2063 * 2064 * So assuming nr_active := 0 when we start out -- true per definition, we 2065 * can simply take per-cpu deltas and fold those into a global accumulate 2066 * to obtain the same result. See calc_load_fold_active(). 2067 * 2068 * Furthermore, in order to avoid synchronizing all per-cpu delta folding 2069 * across the machine, we assume 10 ticks is sufficient time for every 2070 * cpu to have completed this task. 2071 * 2072 * This places an upper-bound on the IRQ-off latency of the machine. Then 2073 * again, being late doesn't loose the delta, just wrecks the sample. 2074 * 2075 * - cpu_rq()->nr_uninterruptible isn't accurately tracked per-cpu because 2076 * this would add another cross-cpu cacheline miss and atomic operation 2077 * to the wakeup path. Instead we increment on whatever cpu the task ran 2078 * when it went into uninterruptible state and decrement on whatever cpu 2079 * did the wakeup. This means that only the sum of nr_uninterruptible over 2080 * all cpus yields the correct result. 2081 * 2082 * This covers the NO_HZ=n code, for extra head-aches, see the comment below. 2083 */ 2084 2085/* Variables and functions for calc_load */ 2086static atomic_long_t calc_load_tasks; 2087static unsigned long calc_load_update; 2088unsigned long avenrun[3]; 2089EXPORT_SYMBOL(avenrun); /* should be removed */ 2090 2091/** 2092 * get_avenrun - get the load average array 2093 * @loads: pointer to dest load array 2094 * @offset: offset to add 2095 * @shift: shift count to shift the result left 2096 * 2097 * These values are estimates at best, so no need for locking. 2098 */ 2099void get_avenrun(unsigned long *loads, unsigned long offset, int shift) 2100{ 2101 loads[0] = (avenrun[0] + offset) << shift; 2102 loads[1] = (avenrun[1] + offset) << shift; 2103 loads[2] = (avenrun[2] + offset) << shift; 2104} 2105 2106static long calc_load_fold_active(struct rq *this_rq) 2107{ 2108 long nr_active, delta = 0; 2109 2110 nr_active = this_rq->nr_running; 2111 nr_active += (long) this_rq->nr_uninterruptible; 2112 2113 if (nr_active != this_rq->calc_load_active) { 2114 delta = nr_active - this_rq->calc_load_active; 2115 this_rq->calc_load_active = nr_active; 2116 } 2117 2118 return delta; 2119} 2120 2121/* 2122 * a1 = a0 * e + a * (1 - e) 2123 */ 2124static unsigned long 2125calc_load(unsigned long load, unsigned long exp, unsigned long active) 2126{ 2127 load *= exp; 2128 load += active * (FIXED_1 - exp); 2129 load += 1UL << (FSHIFT - 1); 2130 return load >> FSHIFT; 2131} 2132 2133#ifdef CONFIG_NO_HZ 2134/* 2135 * Handle NO_HZ for the global load-average. 2136 * 2137 * Since the above described distributed algorithm to compute the global 2138 * load-average relies on per-cpu sampling from the tick, it is affected by 2139 * NO_HZ. 2140 * 2141 * The basic idea is to fold the nr_active delta into a global idle-delta upon 2142 * entering NO_HZ state such that we can include this as an 'extra' cpu delta 2143 * when we read the global state. 2144 * 2145 * Obviously reality has to ruin such a delightfully simple scheme: 2146 * 2147 * - When we go NO_HZ idle during the window, we can negate our sample 2148 * contribution, causing under-accounting. 2149 * 2150 * We avoid this by keeping two idle-delta counters and flipping them 2151 * when the window starts, thus separating old and new NO_HZ load. 2152 * 2153 * The only trick is the slight shift in index flip for read vs write. 2154 * 2155 * 0s 5s 10s 15s 2156 * +10 +10 +10 +10 2157 * |-|-----------|-|-----------|-|-----------|-| 2158 * r:0 0 1 1 0 0 1 1 0 2159 * w:0 1 1 0 0 1 1 0 0 2160 * 2161 * This ensures we'll fold the old idle contribution in this window while 2162 * accumlating the new one. 2163 * 2164 * - When we wake up from NO_HZ idle during the window, we push up our 2165 * contribution, since we effectively move our sample point to a known 2166 * busy state. 2167 * 2168 * This is solved by pushing the window forward, and thus skipping the 2169 * sample, for this cpu (effectively using the idle-delta for this cpu which 2170 * was in effect at the time the window opened). This also solves the issue 2171 * of having to deal with a cpu having been in NOHZ idle for multiple 2172 * LOAD_FREQ intervals. 2173 * 2174 * When making the ILB scale, we should try to pull this in as well. 2175 */ 2176static atomic_long_t calc_load_idle[2]; 2177static int calc_load_idx; 2178 2179static inline int calc_load_write_idx(void) 2180{ 2181 int idx = calc_load_idx; 2182 2183 /* 2184 * See calc_global_nohz(), if we observe the new index, we also 2185 * need to observe the new update time. 2186 */ 2187 smp_rmb(); 2188 2189 /* 2190 * If the folding window started, make sure we start writing in the 2191 * next idle-delta. 2192 */ 2193 if (!time_before(jiffies, calc_load_update)) 2194 idx++; 2195 2196 return idx & 1; 2197} 2198 2199static inline int calc_load_read_idx(void) 2200{ 2201 return calc_load_idx & 1; 2202} 2203 2204void calc_load_enter_idle(void) 2205{ 2206 struct rq *this_rq = this_rq(); 2207 long delta; 2208 2209 /* 2210 * We're going into NOHZ mode, if there's any pending delta, fold it 2211 * into the pending idle delta. 2212 */ 2213 delta = calc_load_fold_active(this_rq); 2214 if (delta) { 2215 int idx = calc_load_write_idx(); 2216 atomic_long_add(delta, &calc_load_idle[idx]); 2217 } 2218} 2219 2220void calc_load_exit_idle(void) 2221{ 2222 struct rq *this_rq = this_rq(); 2223 2224 /* 2225 * If we're still before the sample window, we're done. 2226 */ 2227 if (time_before(jiffies, this_rq->calc_load_update)) 2228 return; 2229 2230 /* 2231 * We woke inside or after the sample window, this means we're already 2232 * accounted through the nohz accounting, so skip the entire deal and 2233 * sync up for the next window. 2234 */ 2235 this_rq->calc_load_update = calc_load_update; 2236 if (time_before(jiffies, this_rq->calc_load_update + 10)) 2237 this_rq->calc_load_update += LOAD_FREQ; 2238} 2239 2240static long calc_load_fold_idle(void) 2241{ 2242 int idx = calc_load_read_idx(); 2243 long delta = 0; 2244 2245 if (atomic_long_read(&calc_load_idle[idx])) 2246 delta = atomic_long_xchg(&calc_load_idle[idx], 0); 2247 2248 return delta; 2249} 2250 2251/** 2252 * fixed_power_int - compute: x^n, in O(log n) time 2253 * 2254 * @x: base of the power 2255 * @frac_bits: fractional bits of @x 2256 * @n: power to raise @x to. 2257 * 2258 * By exploiting the relation between the definition of the natural power 2259 * function: x^n := x*x*...*x (x multiplied by itself for n times), and 2260 * the binary encoding of numbers used by computers: n := \Sum n_i * 2^i, 2261 * (where: n_i \elem {0, 1}, the binary vector representing n), 2262 * we find: x^n := x^(\Sum n_i * 2^i) := \Prod x^(n_i * 2^i), which is 2263 * of course trivially computable in O(log_2 n), the length of our binary 2264 * vector. 2265 */ 2266static unsigned long 2267fixed_power_int(unsigned long x, unsigned int frac_bits, unsigned int n) 2268{ 2269 unsigned long result = 1UL << frac_bits; 2270 2271 if (n) for (;;) { 2272 if (n & 1) { 2273 result *= x; 2274 result += 1UL << (frac_bits - 1); 2275 result >>= frac_bits; 2276 } 2277 n >>= 1; 2278 if (!n) 2279 break; 2280 x *= x; 2281 x += 1UL << (frac_bits - 1); 2282 x >>= frac_bits; 2283 } 2284 2285 return result; 2286} 2287 2288/* 2289 * a1 = a0 * e + a * (1 - e) 2290 * 2291 * a2 = a1 * e + a * (1 - e) 2292 * = (a0 * e + a * (1 - e)) * e + a * (1 - e) 2293 * = a0 * e^2 + a * (1 - e) * (1 + e) 2294 * 2295 * a3 = a2 * e + a * (1 - e) 2296 * = (a0 * e^2 + a * (1 - e) * (1 + e)) * e + a * (1 - e) 2297 * = a0 * e^3 + a * (1 - e) * (1 + e + e^2) 2298 * 2299 * ... 2300 * 2301 * an = a0 * e^n + a * (1 - e) * (1 + e + ... + e^n-1) [1] 2302 * = a0 * e^n + a * (1 - e) * (1 - e^n)/(1 - e) 2303 * = a0 * e^n + a * (1 - e^n) 2304 * 2305 * [1] application of the geometric series: 2306 * 2307 * n 1 - x^(n+1) 2308 * S_n := \Sum x^i = ------------- 2309 * i=0 1 - x 2310 */ 2311static unsigned long 2312calc_load_n(unsigned long load, unsigned long exp, 2313 unsigned long active, unsigned int n) 2314{ 2315 2316 return calc_load(load, fixed_power_int(exp, FSHIFT, n), active); 2317} 2318 2319/* 2320 * NO_HZ can leave us missing all per-cpu ticks calling 2321 * calc_load_account_active(), but since an idle CPU folds its delta into 2322 * calc_load_tasks_idle per calc_load_account_idle(), all we need to do is fold 2323 * in the pending idle delta if our idle period crossed a load cycle boundary. 2324 * 2325 * Once we've updated the global active value, we need to apply the exponential 2326 * weights adjusted to the number of cycles missed. 2327 */ 2328static void calc_global_nohz(void) 2329{ 2330 long delta, active, n; 2331 2332 if (!time_before(jiffies, calc_load_update + 10)) { 2333 /* 2334 * Catch-up, fold however many we are behind still 2335 */ 2336 delta = jiffies - calc_load_update - 10; 2337 n = 1 + (delta / LOAD_FREQ); 2338 2339 active = atomic_long_read(&calc_load_tasks); 2340 active = active > 0 ? active * FIXED_1 : 0; 2341 2342 avenrun[0] = calc_load_n(avenrun[0], EXP_1, active, n); 2343 avenrun[1] = calc_load_n(avenrun[1], EXP_5, active, n); 2344 avenrun[2] = calc_load_n(avenrun[2], EXP_15, active, n); 2345 2346 calc_load_update += n * LOAD_FREQ; 2347 } 2348 2349 /* 2350 * Flip the idle index... 2351 * 2352 * Make sure we first write the new time then flip the index, so that 2353 * calc_load_write_idx() will see the new time when it reads the new 2354 * index, this avoids a double flip messing things up. 2355 */ 2356 smp_wmb(); 2357 calc_load_idx++; 2358} 2359#else /* !CONFIG_NO_HZ */ 2360 2361static inline long calc_load_fold_idle(void) { return 0; } 2362static inline void calc_global_nohz(void) { } 2363 2364#endif /* CONFIG_NO_HZ */ 2365 2366/* 2367 * calc_load - update the avenrun load estimates 10 ticks after the 2368 * CPUs have updated calc_load_tasks. 2369 */ 2370void calc_global_load(unsigned long ticks) 2371{ 2372 long active, delta; 2373 2374 if (time_before(jiffies, calc_load_update + 10)) 2375 return; 2376 2377 /* 2378 * Fold the 'old' idle-delta to include all NO_HZ cpus. 2379 */ 2380 delta = calc_load_fold_idle(); 2381 if (delta) 2382 atomic_long_add(delta, &calc_load_tasks); 2383 2384 active = atomic_long_read(&calc_load_tasks); 2385 active = active > 0 ? active * FIXED_1 : 0; 2386 2387 avenrun[0] = calc_load(avenrun[0], EXP_1, active); 2388 avenrun[1] = calc_load(avenrun[1], EXP_5, active); 2389 avenrun[2] = calc_load(avenrun[2], EXP_15, active); 2390 2391 calc_load_update += LOAD_FREQ; 2392 2393 /* 2394 * In case we idled for multiple LOAD_FREQ intervals, catch up in bulk. 2395 */ 2396 calc_global_nohz(); 2397} 2398 2399/* 2400 * Called from update_cpu_load() to periodically update this CPU's 2401 * active count. 2402 */ 2403static void calc_load_account_active(struct rq *this_rq) 2404{ 2405 long delta; 2406 2407 if (time_before(jiffies, this_rq->calc_load_update)) 2408 return; 2409 2410 delta = calc_load_fold_active(this_rq); 2411 if (delta) 2412 atomic_long_add(delta, &calc_load_tasks); 2413 2414 this_rq->calc_load_update += LOAD_FREQ; 2415} 2416 2417/* 2418 * End of global load-average stuff 2419 */ 2420 2421/* 2422 * The exact cpuload at various idx values, calculated at every tick would be 2423 * load = (2^idx - 1) / 2^idx * load + 1 / 2^idx * cur_load 2424 * 2425 * If a cpu misses updates for n-1 ticks (as it was idle) and update gets called 2426 * on nth tick when cpu may be busy, then we have: 2427 * load = ((2^idx - 1) / 2^idx)^(n-1) * load 2428 * load = (2^idx - 1) / 2^idx) * load + 1 / 2^idx * cur_load 2429 * 2430 * decay_load_missed() below does efficient calculation of 2431 * load = ((2^idx - 1) / 2^idx)^(n-1) * load 2432 * avoiding 0..n-1 loop doing load = ((2^idx - 1) / 2^idx) * load 2433 * 2434 * The calculation is approximated on a 128 point scale. 2435 * degrade_zero_ticks is the number of ticks after which load at any 2436 * particular idx is approximated to be zero. 2437 * degrade_factor is a precomputed table, a row for each load idx. 2438 * Each column corresponds to degradation factor for a power of two ticks, 2439 * based on 128 point scale. 2440 * Example: 2441 * row 2, col 3 (=12) says that the degradation at load idx 2 after 2442 * 8 ticks is 12/128 (which is an approximation of exact factor 3^8/4^8). 2443 * 2444 * With this power of 2 load factors, we can degrade the load n times 2445 * by looking at 1 bits in n and doing as many mult/shift instead of 2446 * n mult/shifts needed by the exact degradation. 2447 */ 2448#define DEGRADE_SHIFT 7 2449static const unsigned char 2450 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128}; 2451static const unsigned char 2452 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = { 2453 {0, 0, 0, 0, 0, 0, 0, 0}, 2454 {64, 32, 8, 0, 0, 0, 0, 0}, 2455 {96, 72, 40, 12, 1, 0, 0}, 2456 {112, 98, 75, 43, 15, 1, 0}, 2457 {120, 112, 98, 76, 45, 16, 2} }; 2458 2459/* 2460 * Update cpu_load for any missed ticks, due to tickless idle. The backlog 2461 * would be when CPU is idle and so we just decay the old load without 2462 * adding any new load. 2463 */ 2464static unsigned long 2465decay_load_missed(unsigned long load, unsigned long missed_updates, int idx) 2466{ 2467 int j = 0; 2468 2469 if (!missed_updates) 2470 return load; 2471 2472 if (missed_updates >= degrade_zero_ticks[idx]) 2473 return 0; 2474 2475 if (idx == 1) 2476 return load >> missed_updates; 2477 2478 while (missed_updates) { 2479 if (missed_updates % 2) 2480 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT; 2481 2482 missed_updates >>= 1; 2483 j++; 2484 } 2485 return load; 2486} 2487 2488/* 2489 * Update rq->cpu_load[] statistics. This function is usually called every 2490 * scheduler tick (TICK_NSEC). With tickless idle this will not be called 2491 * every tick. We fix it up based on jiffies. 2492 */ 2493static void __update_cpu_load(struct rq *this_rq, unsigned long this_load, 2494 unsigned long pending_updates) 2495{ 2496 int i, scale; 2497 2498 this_rq->nr_load_updates++; 2499 2500 /* Update our load: */ 2501 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */ 2502 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) { 2503 unsigned long old_load, new_load; 2504 2505 /* scale is effectively 1 << i now, and >> i divides by scale */ 2506 2507 old_load = this_rq->cpu_load[i]; 2508 old_load = decay_load_missed(old_load, pending_updates - 1, i); 2509 new_load = this_load; 2510 /* 2511 * Round up the averaging division if load is increasing. This 2512 * prevents us from getting stuck on 9 if the load is 10, for 2513 * example. 2514 */ 2515 if (new_load > old_load) 2516 new_load += scale - 1; 2517 2518 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i; 2519 } 2520 2521 sched_avg_update(this_rq); 2522} 2523 2524#ifdef CONFIG_NO_HZ 2525/* 2526 * There is no sane way to deal with nohz on smp when using jiffies because the 2527 * cpu doing the jiffies update might drift wrt the cpu doing the jiffy reading 2528 * causing off-by-one errors in observed deltas; {0,2} instead of {1,1}. 2529 * 2530 * Therefore we cannot use the delta approach from the regular tick since that 2531 * would seriously skew the load calculation. However we'll make do for those 2532 * updates happening while idle (nohz_idle_balance) or coming out of idle 2533 * (tick_nohz_idle_exit). 2534 * 2535 * This means we might still be one tick off for nohz periods. 2536 */ 2537 2538/* 2539 * Called from nohz_idle_balance() to update the load ratings before doing the 2540 * idle balance. 2541 */ 2542void update_idle_cpu_load(struct rq *this_rq) 2543{ 2544 unsigned long curr_jiffies = ACCESS_ONCE(jiffies); 2545 unsigned long load = this_rq->load.weight; 2546 unsigned long pending_updates; 2547 2548 /* 2549 * bail if there's load or we're actually up-to-date. 2550 */ 2551 if (load || curr_jiffies == this_rq->last_load_update_tick) 2552 return; 2553 2554 pending_updates = curr_jiffies - this_rq->last_load_update_tick; 2555 this_rq->last_load_update_tick = curr_jiffies; 2556 2557 __update_cpu_load(this_rq, load, pending_updates); 2558} 2559 2560/* 2561 * Called from tick_nohz_idle_exit() -- try and fix up the ticks we missed. 2562 */ 2563void update_cpu_load_nohz(void) 2564{ 2565 struct rq *this_rq = this_rq(); 2566 unsigned long curr_jiffies = ACCESS_ONCE(jiffies); 2567 unsigned long pending_updates; 2568 2569 if (curr_jiffies == this_rq->last_load_update_tick) 2570 return; 2571 2572 raw_spin_lock(&this_rq->lock); 2573 pending_updates = curr_jiffies - this_rq->last_load_update_tick; 2574 if (pending_updates) { 2575 this_rq->last_load_update_tick = curr_jiffies; 2576 /* 2577 * We were idle, this means load 0, the current load might be 2578 * !0 due to remote wakeups and the sort. 2579 */ 2580 __update_cpu_load(this_rq, 0, pending_updates); 2581 } 2582 raw_spin_unlock(&this_rq->lock); 2583} 2584#endif /* CONFIG_NO_HZ */ 2585 2586/* 2587 * Called from scheduler_tick() 2588 */ 2589static void update_cpu_load_active(struct rq *this_rq) 2590{ 2591 /* 2592 * See the mess around update_idle_cpu_load() / update_cpu_load_nohz(). 2593 */ 2594 this_rq->last_load_update_tick = jiffies; 2595 __update_cpu_load(this_rq, this_rq->load.weight, 1); 2596 2597 calc_load_account_active(this_rq); 2598} 2599 2600#ifdef CONFIG_SMP 2601 2602/* 2603 * sched_exec - execve() is a valuable balancing opportunity, because at 2604 * this point the task has the smallest effective memory and cache footprint. 2605 */ 2606void sched_exec(void) 2607{ 2608 struct task_struct *p = current; 2609 unsigned long flags; 2610 int dest_cpu; 2611 2612 raw_spin_lock_irqsave(&p->pi_lock, flags); 2613 dest_cpu = p->sched_class->select_task_rq(p, SD_BALANCE_EXEC, 0); 2614 if (dest_cpu == smp_processor_id()) 2615 goto unlock; 2616 2617 if (likely(cpu_active(dest_cpu))) { 2618 struct migration_arg arg = { p, dest_cpu }; 2619 2620 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 2621 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg); 2622 return; 2623 } 2624unlock: 2625 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 2626} 2627 2628#endif 2629 2630DEFINE_PER_CPU(struct kernel_stat, kstat); 2631DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat); 2632 2633EXPORT_PER_CPU_SYMBOL(kstat); 2634EXPORT_PER_CPU_SYMBOL(kernel_cpustat); 2635 2636/* 2637 * Return any ns on the sched_clock that have not yet been accounted in 2638 * @p in case that task is currently running. 2639 * 2640 * Called with task_rq_lock() held on @rq. 2641 */ 2642static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq) 2643{ 2644 u64 ns = 0; 2645 2646 if (task_current(rq, p)) { 2647 update_rq_clock(rq); 2648 ns = rq->clock_task - p->se.exec_start; 2649 if ((s64)ns < 0) 2650 ns = 0; 2651 } 2652 2653 return ns; 2654} 2655 2656unsigned long long task_delta_exec(struct task_struct *p) 2657{ 2658 unsigned long flags; 2659 struct rq *rq; 2660 u64 ns = 0; 2661 2662 rq = task_rq_lock(p, &flags); 2663 ns = do_task_delta_exec(p, rq); 2664 task_rq_unlock(rq, p, &flags); 2665 2666 return ns; 2667} 2668 2669/* 2670 * Return accounted runtime for the task. 2671 * In case the task is currently running, return the runtime plus current's 2672 * pending runtime that have not been accounted yet. 2673 */ 2674unsigned long long task_sched_runtime(struct task_struct *p) 2675{ 2676 unsigned long flags; 2677 struct rq *rq; 2678 u64 ns = 0; 2679 2680 rq = task_rq_lock(p, &flags); 2681 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq); 2682 task_rq_unlock(rq, p, &flags); 2683 2684 return ns; 2685} 2686 2687/* 2688 * This function gets called by the timer code, with HZ frequency. 2689 * We call it with interrupts disabled. 2690 */ 2691void scheduler_tick(void) 2692{ 2693 int cpu = smp_processor_id(); 2694 struct rq *rq = cpu_rq(cpu); 2695 struct task_struct *curr = rq->curr; 2696 2697 sched_clock_tick(); 2698 2699 raw_spin_lock(&rq->lock); 2700 update_rq_clock(rq); 2701 update_cpu_load_active(rq); 2702 curr->sched_class->task_tick(rq, curr, 0); 2703 raw_spin_unlock(&rq->lock); 2704 2705 perf_event_task_tick(); 2706 2707#ifdef CONFIG_SMP 2708 rq->idle_balance = idle_cpu(cpu); 2709 trigger_load_balance(rq, cpu); 2710#endif 2711} 2712 2713notrace unsigned long get_parent_ip(unsigned long addr) 2714{ 2715 if (in_lock_functions(addr)) { 2716 addr = CALLER_ADDR2; 2717 if (in_lock_functions(addr)) 2718 addr = CALLER_ADDR3; 2719 } 2720 return addr; 2721} 2722 2723#if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \ 2724 defined(CONFIG_PREEMPT_TRACER)) 2725 2726void __kprobes add_preempt_count(int val) 2727{ 2728#ifdef CONFIG_DEBUG_PREEMPT 2729 /* 2730 * Underflow? 2731 */ 2732 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0))) 2733 return; 2734#endif 2735 preempt_count() += val; 2736#ifdef CONFIG_DEBUG_PREEMPT 2737 /* 2738 * Spinlock count overflowing soon? 2739 */ 2740 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >= 2741 PREEMPT_MASK - 10); 2742#endif 2743 if (preempt_count() == val) 2744 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1)); 2745} 2746EXPORT_SYMBOL(add_preempt_count); 2747 2748void __kprobes sub_preempt_count(int val) 2749{ 2750#ifdef CONFIG_DEBUG_PREEMPT 2751 /* 2752 * Underflow? 2753 */ 2754 if (DEBUG_LOCKS_WARN_ON(val > preempt_count())) 2755 return; 2756 /* 2757 * Is the spinlock portion underflowing? 2758 */ 2759 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) && 2760 !(preempt_count() & PREEMPT_MASK))) 2761 return; 2762#endif 2763 2764 if (preempt_count() == val) 2765 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1)); 2766 preempt_count() -= val; 2767} 2768EXPORT_SYMBOL(sub_preempt_count); 2769 2770#endif 2771 2772/* 2773 * Print scheduling while atomic bug: 2774 */ 2775static noinline void __schedule_bug(struct task_struct *prev) 2776{ 2777 if (oops_in_progress) 2778 return; 2779 2780 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n", 2781 prev->comm, prev->pid, preempt_count()); 2782 2783 debug_show_held_locks(prev); 2784 print_modules(); 2785 if (irqs_disabled()) 2786 print_irqtrace_events(prev); 2787 dump_stack(); 2788 add_taint(TAINT_WARN); 2789} 2790 2791/* 2792 * Various schedule()-time debugging checks and statistics: 2793 */ 2794static inline void schedule_debug(struct task_struct *prev) 2795{ 2796 /* 2797 * Test if we are atomic. Since do_exit() needs to call into 2798 * schedule() atomically, we ignore that path for now. 2799 * Otherwise, whine if we are scheduling when we should not be. 2800 */ 2801 if (unlikely(in_atomic_preempt_off() && !prev->exit_state)) 2802 __schedule_bug(prev); 2803 rcu_sleep_check(); 2804 2805 profile_hit(SCHED_PROFILING, __builtin_return_address(0)); 2806 2807 schedstat_inc(this_rq(), sched_count); 2808} 2809 2810static void put_prev_task(struct rq *rq, struct task_struct *prev) 2811{ 2812 if (prev->on_rq || rq->skip_clock_update < 0) 2813 update_rq_clock(rq); 2814 prev->sched_class->put_prev_task(rq, prev); 2815} 2816 2817/* 2818 * Pick up the highest-prio task: 2819 */ 2820static inline struct task_struct * 2821pick_next_task(struct rq *rq) 2822{ 2823 const struct sched_class *class; 2824 struct task_struct *p; 2825 2826 /* 2827 * Optimization: we know that if all tasks are in 2828 * the fair class we can call that function directly: 2829 */ 2830 if (likely(rq->nr_running == rq->cfs.h_nr_running)) { 2831 p = fair_sched_class.pick_next_task(rq); 2832 if (likely(p)) 2833 return p; 2834 } 2835 2836 for_each_class(class) { 2837 p = class->pick_next_task(rq); 2838 if (p) 2839 return p; 2840 } 2841 2842 BUG(); /* the idle class will always have a runnable task */ 2843} 2844 2845/* 2846 * __schedule() is the main scheduler function. 2847 * 2848 * The main means of driving the scheduler and thus entering this function are: 2849 * 2850 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc. 2851 * 2852 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return 2853 * paths. For example, see arch/x86/entry_64.S. 2854 * 2855 * To drive preemption between tasks, the scheduler sets the flag in timer 2856 * interrupt handler scheduler_tick(). 2857 * 2858 * 3. Wakeups don't really cause entry into schedule(). They add a 2859 * task to the run-queue and that's it. 2860 * 2861 * Now, if the new task added to the run-queue preempts the current 2862 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets 2863 * called on the nearest possible occasion: 2864 * 2865 * - If the kernel is preemptible (CONFIG_PREEMPT=y): 2866 * 2867 * - in syscall or exception context, at the next outmost 2868 * preempt_enable(). (this might be as soon as the wake_up()'s 2869 * spin_unlock()!) 2870 * 2871 * - in IRQ context, return from interrupt-handler to 2872 * preemptible context 2873 * 2874 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set) 2875 * then at the next: 2876 * 2877 * - cond_resched() call 2878 * - explicit schedule() call 2879 * - return from syscall or exception to user-space 2880 * - return from interrupt-handler to user-space 2881 */ 2882static void __sched __schedule(void) 2883{ 2884 struct task_struct *prev, *next; 2885 unsigned long *switch_count; 2886 struct rq *rq; 2887 int cpu; 2888 2889need_resched: 2890 preempt_disable(); 2891 cpu = smp_processor_id(); 2892 rq = cpu_rq(cpu); 2893 rcu_note_context_switch(cpu); 2894 prev = rq->curr; 2895 2896 schedule_debug(prev); 2897 2898 if (sched_feat(HRTICK)) 2899 hrtick_clear(rq); 2900 2901 raw_spin_lock_irq(&rq->lock); 2902 2903 switch_count = &prev->nivcsw; 2904 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) { 2905 if (unlikely(signal_pending_state(prev->state, prev))) { 2906 prev->state = TASK_RUNNING; 2907 } else { 2908 deactivate_task(rq, prev, DEQUEUE_SLEEP); 2909 prev->on_rq = 0; 2910 2911 /* 2912 * If a worker went to sleep, notify and ask workqueue 2913 * whether it wants to wake up a task to maintain 2914 * concurrency. 2915 */ 2916 if (prev->flags & PF_WQ_WORKER) { 2917 struct task_struct *to_wakeup; 2918 2919 to_wakeup = wq_worker_sleeping(prev, cpu); 2920 if (to_wakeup) 2921 try_to_wake_up_local(to_wakeup); 2922 } 2923 } 2924 switch_count = &prev->nvcsw; 2925 } 2926 2927 pre_schedule(rq, prev); 2928 2929 if (unlikely(!rq->nr_running)) 2930 idle_balance(cpu, rq); 2931 2932 put_prev_task(rq, prev); 2933 next = pick_next_task(rq); 2934 clear_tsk_need_resched(prev); 2935 rq->skip_clock_update = 0; 2936 2937 if (likely(prev != next)) { 2938 rq->nr_switches++; 2939 rq->curr = next; 2940 ++*switch_count; 2941 2942 context_switch(rq, prev, next); /* unlocks the rq */ 2943 /* 2944 * The context switch have flipped the stack from under us 2945 * and restored the local variables which were saved when 2946 * this task called schedule() in the past. prev == current 2947 * is still correct, but it can be moved to another cpu/rq. 2948 */ 2949 cpu = smp_processor_id(); 2950 rq = cpu_rq(cpu); 2951 } else 2952 raw_spin_unlock_irq(&rq->lock); 2953 2954 post_schedule(rq); 2955 2956 sched_preempt_enable_no_resched(); 2957 if (need_resched()) 2958 goto need_resched; 2959} 2960 2961static inline void sched_submit_work(struct task_struct *tsk) 2962{ 2963 if (!tsk->state || tsk_is_pi_blocked(tsk)) 2964 return; 2965 /* 2966 * If we are going to sleep and we have plugged IO queued, 2967 * make sure to submit it to avoid deadlocks. 2968 */ 2969 if (blk_needs_flush_plug(tsk)) 2970 blk_schedule_flush_plug(tsk); 2971} 2972 2973asmlinkage void __sched schedule(void) 2974{ 2975 struct task_struct *tsk = current; 2976 2977 sched_submit_work(tsk); 2978 __schedule(); 2979} 2980EXPORT_SYMBOL(schedule); 2981 2982#ifdef CONFIG_CONTEXT_TRACKING 2983asmlinkage void __sched schedule_user(void) 2984{ 2985 /* 2986 * If we come here after a random call to set_need_resched(), 2987 * or we have been woken up remotely but the IPI has not yet arrived, 2988 * we haven't yet exited the RCU idle mode. Do it here manually until 2989 * we find a better solution. 2990 */ 2991 user_exit(); 2992 schedule(); 2993 user_enter(); 2994} 2995#endif 2996 2997/** 2998 * schedule_preempt_disabled - called with preemption disabled 2999 * 3000 * Returns with preemption disabled. Note: preempt_count must be 1 3001 */ 3002void __sched schedule_preempt_disabled(void) 3003{ 3004 sched_preempt_enable_no_resched(); 3005 schedule(); 3006 preempt_disable(); 3007} 3008 3009#ifdef CONFIG_MUTEX_SPIN_ON_OWNER 3010 3011static inline bool owner_running(struct mutex *lock, struct task_struct *owner) 3012{ 3013 if (lock->owner != owner) 3014 return false; 3015 3016 /* 3017 * Ensure we emit the owner->on_cpu, dereference _after_ checking 3018 * lock->owner still matches owner, if that fails, owner might 3019 * point to free()d memory, if it still matches, the rcu_read_lock() 3020 * ensures the memory stays valid. 3021 */ 3022 barrier(); 3023 3024 return owner->on_cpu; 3025} 3026 3027/* 3028 * Look out! "owner" is an entirely speculative pointer 3029 * access and not reliable. 3030 */ 3031int mutex_spin_on_owner(struct mutex *lock, struct task_struct *owner) 3032{ 3033 if (!sched_feat(OWNER_SPIN)) 3034 return 0; 3035 3036 rcu_read_lock(); 3037 while (owner_running(lock, owner)) { 3038 if (need_resched()) 3039 break; 3040 3041 arch_mutex_cpu_relax(); 3042 } 3043 rcu_read_unlock(); 3044 3045 /* 3046 * We break out the loop above on need_resched() and when the 3047 * owner changed, which is a sign for heavy contention. Return 3048 * success only when lock->owner is NULL. 3049 */ 3050 return lock->owner == NULL; 3051} 3052#endif 3053 3054#ifdef CONFIG_PREEMPT 3055/* 3056 * this is the entry point to schedule() from in-kernel preemption 3057 * off of preempt_enable. Kernel preemptions off return from interrupt 3058 * occur there and call schedule directly. 3059 */ 3060asmlinkage void __sched notrace preempt_schedule(void) 3061{ 3062 struct thread_info *ti = current_thread_info(); 3063 3064 /* 3065 * If there is a non-zero preempt_count or interrupts are disabled, 3066 * we do not want to preempt the current task. Just return.. 3067 */ 3068 if (likely(ti->preempt_count || irqs_disabled())) 3069 return; 3070 3071 do { 3072 add_preempt_count_notrace(PREEMPT_ACTIVE); 3073 __schedule(); 3074 sub_preempt_count_notrace(PREEMPT_ACTIVE); 3075 3076 /* 3077 * Check again in case we missed a preemption opportunity 3078 * between schedule and now. 3079 */ 3080 barrier(); 3081 } while (need_resched()); 3082} 3083EXPORT_SYMBOL(preempt_schedule); 3084 3085/* 3086 * this is the entry point to schedule() from kernel preemption 3087 * off of irq context. 3088 * Note, that this is called and return with irqs disabled. This will 3089 * protect us against recursive calling from irq. 3090 */ 3091asmlinkage void __sched preempt_schedule_irq(void) 3092{ 3093 struct thread_info *ti = current_thread_info(); 3094 3095 /* Catch callers which need to be fixed */ 3096 BUG_ON(ti->preempt_count || !irqs_disabled()); 3097 3098 user_exit(); 3099 do { 3100 add_preempt_count(PREEMPT_ACTIVE); 3101 local_irq_enable(); 3102 __schedule(); 3103 local_irq_disable(); 3104 sub_preempt_count(PREEMPT_ACTIVE); 3105 3106 /* 3107 * Check again in case we missed a preemption opportunity 3108 * between schedule and now. 3109 */ 3110 barrier(); 3111 } while (need_resched()); 3112} 3113 3114#endif /* CONFIG_PREEMPT */ 3115 3116int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags, 3117 void *key) 3118{ 3119 return try_to_wake_up(curr->private, mode, wake_flags); 3120} 3121EXPORT_SYMBOL(default_wake_function); 3122 3123/* 3124 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just 3125 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve 3126 * number) then we wake all the non-exclusive tasks and one exclusive task. 3127 * 3128 * There are circumstances in which we can try to wake a task which has already 3129 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns 3130 * zero in this (rare) case, and we handle it by continuing to scan the queue. 3131 */ 3132static void __wake_up_common(wait_queue_head_t *q, unsigned int mode, 3133 int nr_exclusive, int wake_flags, void *key) 3134{ 3135 wait_queue_t *curr, *next; 3136 3137 list_for_each_entry_safe(curr, next, &q->task_list, task_list) { 3138 unsigned flags = curr->flags; 3139 3140 if (curr->func(curr, mode, wake_flags, key) && 3141 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive) 3142 break; 3143 } 3144} 3145 3146/** 3147 * __wake_up - wake up threads blocked on a waitqueue. 3148 * @q: the waitqueue 3149 * @mode: which threads 3150 * @nr_exclusive: how many wake-one or wake-many threads to wake up 3151 * @key: is directly passed to the wakeup function 3152 * 3153 * It may be assumed that this function implies a write memory barrier before 3154 * changing the task state if and only if any tasks are woken up. 3155 */ 3156void __wake_up(wait_queue_head_t *q, unsigned int mode, 3157 int nr_exclusive, void *key) 3158{ 3159 unsigned long flags; 3160 3161 spin_lock_irqsave(&q->lock, flags); 3162 __wake_up_common(q, mode, nr_exclusive, 0, key); 3163 spin_unlock_irqrestore(&q->lock, flags); 3164} 3165EXPORT_SYMBOL(__wake_up); 3166 3167/* 3168 * Same as __wake_up but called with the spinlock in wait_queue_head_t held. 3169 */ 3170void __wake_up_locked(wait_queue_head_t *q, unsigned int mode, int nr) 3171{ 3172 __wake_up_common(q, mode, nr, 0, NULL); 3173} 3174EXPORT_SYMBOL_GPL(__wake_up_locked); 3175 3176void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key) 3177{ 3178 __wake_up_common(q, mode, 1, 0, key); 3179} 3180EXPORT_SYMBOL_GPL(__wake_up_locked_key); 3181 3182/** 3183 * __wake_up_sync_key - wake up threads blocked on a waitqueue. 3184 * @q: the waitqueue 3185 * @mode: which threads 3186 * @nr_exclusive: how many wake-one or wake-many threads to wake up 3187 * @key: opaque value to be passed to wakeup targets 3188 * 3189 * The sync wakeup differs that the waker knows that it will schedule 3190 * away soon, so while the target thread will be woken up, it will not 3191 * be migrated to another CPU - ie. the two threads are 'synchronized' 3192 * with each other. This can prevent needless bouncing between CPUs. 3193 * 3194 * On UP it can prevent extra preemption. 3195 * 3196 * It may be assumed that this function implies a write memory barrier before 3197 * changing the task state if and only if any tasks are woken up. 3198 */ 3199void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode, 3200 int nr_exclusive, void *key) 3201{ 3202 unsigned long flags; 3203 int wake_flags = WF_SYNC; 3204 3205 if (unlikely(!q)) 3206 return; 3207 3208 if (unlikely(!nr_exclusive)) 3209 wake_flags = 0; 3210 3211 spin_lock_irqsave(&q->lock, flags); 3212 __wake_up_common(q, mode, nr_exclusive, wake_flags, key); 3213 spin_unlock_irqrestore(&q->lock, flags); 3214} 3215EXPORT_SYMBOL_GPL(__wake_up_sync_key); 3216 3217/* 3218 * __wake_up_sync - see __wake_up_sync_key() 3219 */ 3220void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive) 3221{ 3222 __wake_up_sync_key(q, mode, nr_exclusive, NULL); 3223} 3224EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */ 3225 3226/** 3227 * complete: - signals a single thread waiting on this completion 3228 * @x: holds the state of this particular completion 3229 * 3230 * This will wake up a single thread waiting on this completion. Threads will be 3231 * awakened in the same order in which they were queued. 3232 * 3233 * See also complete_all(), wait_for_completion() and related routines. 3234 * 3235 * It may be assumed that this function implies a write memory barrier before 3236 * changing the task state if and only if any tasks are woken up. 3237 */ 3238void complete(struct completion *x) 3239{ 3240 unsigned long flags; 3241 3242 spin_lock_irqsave(&x->wait.lock, flags); 3243 x->done++; 3244 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL); 3245 spin_unlock_irqrestore(&x->wait.lock, flags); 3246} 3247EXPORT_SYMBOL(complete); 3248 3249/** 3250 * complete_all: - signals all threads waiting on this completion 3251 * @x: holds the state of this particular completion 3252 * 3253 * This will wake up all threads waiting on this particular completion event. 3254 * 3255 * It may be assumed that this function implies a write memory barrier before 3256 * changing the task state if and only if any tasks are woken up. 3257 */ 3258void complete_all(struct completion *x) 3259{ 3260 unsigned long flags; 3261 3262 spin_lock_irqsave(&x->wait.lock, flags); 3263 x->done += UINT_MAX/2; 3264 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL); 3265 spin_unlock_irqrestore(&x->wait.lock, flags); 3266} 3267EXPORT_SYMBOL(complete_all); 3268 3269static inline long __sched 3270do_wait_for_common(struct completion *x, 3271 long (*action)(long), long timeout, int state) 3272{ 3273 if (!x->done) { 3274 DECLARE_WAITQUEUE(wait, current); 3275 3276 __add_wait_queue_tail_exclusive(&x->wait, &wait); 3277 do { 3278 if (signal_pending_state(state, current)) { 3279 timeout = -ERESTARTSYS; 3280 break; 3281 } 3282 __set_current_state(state); 3283 spin_unlock_irq(&x->wait.lock); 3284 timeout = action(timeout); 3285 spin_lock_irq(&x->wait.lock); 3286 } while (!x->done && timeout); 3287 __remove_wait_queue(&x->wait, &wait); 3288 if (!x->done) 3289 return timeout; 3290 } 3291 x->done--; 3292 return timeout ?: 1; 3293} 3294 3295static inline long __sched 3296__wait_for_common(struct completion *x, 3297 long (*action)(long), long timeout, int state) 3298{ 3299 might_sleep(); 3300 3301 spin_lock_irq(&x->wait.lock); 3302 timeout = do_wait_for_common(x, action, timeout, state); 3303 spin_unlock_irq(&x->wait.lock); 3304 return timeout; 3305} 3306 3307static long __sched 3308wait_for_common(struct completion *x, long timeout, int state) 3309{ 3310 return __wait_for_common(x, schedule_timeout, timeout, state); 3311} 3312 3313static long __sched 3314wait_for_common_io(struct completion *x, long timeout, int state) 3315{ 3316 return __wait_for_common(x, io_schedule_timeout, timeout, state); 3317} 3318 3319/** 3320 * wait_for_completion: - waits for completion of a task 3321 * @x: holds the state of this particular completion 3322 * 3323 * This waits to be signaled for completion of a specific task. It is NOT 3324 * interruptible and there is no timeout. 3325 * 3326 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout 3327 * and interrupt capability. Also see complete(). 3328 */ 3329void __sched wait_for_completion(struct completion *x) 3330{ 3331 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE); 3332} 3333EXPORT_SYMBOL(wait_for_completion); 3334 3335/** 3336 * wait_for_completion_timeout: - waits for completion of a task (w/timeout) 3337 * @x: holds the state of this particular completion 3338 * @timeout: timeout value in jiffies 3339 * 3340 * This waits for either a completion of a specific task to be signaled or for a 3341 * specified timeout to expire. The timeout is in jiffies. It is not 3342 * interruptible. 3343 * 3344 * The return value is 0 if timed out, and positive (at least 1, or number of 3345 * jiffies left till timeout) if completed. 3346 */ 3347unsigned long __sched 3348wait_for_completion_timeout(struct completion *x, unsigned long timeout) 3349{ 3350 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE); 3351} 3352EXPORT_SYMBOL(wait_for_completion_timeout); 3353 3354/** 3355 * wait_for_completion_io: - waits for completion of a task 3356 * @x: holds the state of this particular completion 3357 * 3358 * This waits to be signaled for completion of a specific task. It is NOT 3359 * interruptible and there is no timeout. The caller is accounted as waiting 3360 * for IO. 3361 */ 3362void __sched wait_for_completion_io(struct completion *x) 3363{ 3364 wait_for_common_io(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE); 3365} 3366EXPORT_SYMBOL(wait_for_completion_io); 3367 3368/** 3369 * wait_for_completion_io_timeout: - waits for completion of a task (w/timeout) 3370 * @x: holds the state of this particular completion 3371 * @timeout: timeout value in jiffies 3372 * 3373 * This waits for either a completion of a specific task to be signaled or for a 3374 * specified timeout to expire. The timeout is in jiffies. It is not 3375 * interruptible. The caller is accounted as waiting for IO. 3376 * 3377 * The return value is 0 if timed out, and positive (at least 1, or number of 3378 * jiffies left till timeout) if completed. 3379 */ 3380unsigned long __sched 3381wait_for_completion_io_timeout(struct completion *x, unsigned long timeout) 3382{ 3383 return wait_for_common_io(x, timeout, TASK_UNINTERRUPTIBLE); 3384} 3385EXPORT_SYMBOL(wait_for_completion_io_timeout); 3386 3387/** 3388 * wait_for_completion_interruptible: - waits for completion of a task (w/intr) 3389 * @x: holds the state of this particular completion 3390 * 3391 * This waits for completion of a specific task to be signaled. It is 3392 * interruptible. 3393 * 3394 * The return value is -ERESTARTSYS if interrupted, 0 if completed. 3395 */ 3396int __sched wait_for_completion_interruptible(struct completion *x) 3397{ 3398 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE); 3399 if (t == -ERESTARTSYS) 3400 return t; 3401 return 0; 3402} 3403EXPORT_SYMBOL(wait_for_completion_interruptible); 3404 3405/** 3406 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr)) 3407 * @x: holds the state of this particular completion 3408 * @timeout: timeout value in jiffies 3409 * 3410 * This waits for either a completion of a specific task to be signaled or for a 3411 * specified timeout to expire. It is interruptible. The timeout is in jiffies. 3412 * 3413 * The return value is -ERESTARTSYS if interrupted, 0 if timed out, 3414 * positive (at least 1, or number of jiffies left till timeout) if completed. 3415 */ 3416long __sched 3417wait_for_completion_interruptible_timeout(struct completion *x, 3418 unsigned long timeout) 3419{ 3420 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE); 3421} 3422EXPORT_SYMBOL(wait_for_completion_interruptible_timeout); 3423 3424/** 3425 * wait_for_completion_killable: - waits for completion of a task (killable) 3426 * @x: holds the state of this particular completion 3427 * 3428 * This waits to be signaled for completion of a specific task. It can be 3429 * interrupted by a kill signal. 3430 * 3431 * The return value is -ERESTARTSYS if interrupted, 0 if completed. 3432 */ 3433int __sched wait_for_completion_killable(struct completion *x) 3434{ 3435 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE); 3436 if (t == -ERESTARTSYS) 3437 return t; 3438 return 0; 3439} 3440EXPORT_SYMBOL(wait_for_completion_killable); 3441 3442/** 3443 * wait_for_completion_killable_timeout: - waits for completion of a task (w/(to,killable)) 3444 * @x: holds the state of this particular completion 3445 * @timeout: timeout value in jiffies 3446 * 3447 * This waits for either a completion of a specific task to be 3448 * signaled or for a specified timeout to expire. It can be 3449 * interrupted by a kill signal. The timeout is in jiffies. 3450 * 3451 * The return value is -ERESTARTSYS if interrupted, 0 if timed out, 3452 * positive (at least 1, or number of jiffies left till timeout) if completed. 3453 */ 3454long __sched 3455wait_for_completion_killable_timeout(struct completion *x, 3456 unsigned long timeout) 3457{ 3458 return wait_for_common(x, timeout, TASK_KILLABLE); 3459} 3460EXPORT_SYMBOL(wait_for_completion_killable_timeout); 3461 3462/** 3463 * try_wait_for_completion - try to decrement a completion without blocking 3464 * @x: completion structure 3465 * 3466 * Returns: 0 if a decrement cannot be done without blocking 3467 * 1 if a decrement succeeded. 3468 * 3469 * If a completion is being used as a counting completion, 3470 * attempt to decrement the counter without blocking. This 3471 * enables us to avoid waiting if the resource the completion 3472 * is protecting is not available. 3473 */ 3474bool try_wait_for_completion(struct completion *x) 3475{ 3476 unsigned long flags; 3477 int ret = 1; 3478 3479 spin_lock_irqsave(&x->wait.lock, flags); 3480 if (!x->done) 3481 ret = 0; 3482 else 3483 x->done--; 3484 spin_unlock_irqrestore(&x->wait.lock, flags); 3485 return ret; 3486} 3487EXPORT_SYMBOL(try_wait_for_completion); 3488 3489/** 3490 * completion_done - Test to see if a completion has any waiters 3491 * @x: completion structure 3492 * 3493 * Returns: 0 if there are waiters (wait_for_completion() in progress) 3494 * 1 if there are no waiters. 3495 * 3496 */ 3497bool completion_done(struct completion *x) 3498{ 3499 unsigned long flags; 3500 int ret = 1; 3501 3502 spin_lock_irqsave(&x->wait.lock, flags); 3503 if (!x->done) 3504 ret = 0; 3505 spin_unlock_irqrestore(&x->wait.lock, flags); 3506 return ret; 3507} 3508EXPORT_SYMBOL(completion_done); 3509 3510static long __sched 3511sleep_on_common(wait_queue_head_t *q, int state, long timeout) 3512{ 3513 unsigned long flags; 3514 wait_queue_t wait; 3515 3516 init_waitqueue_entry(&wait, current); 3517 3518 __set_current_state(state); 3519 3520 spin_lock_irqsave(&q->lock, flags); 3521 __add_wait_queue(q, &wait); 3522 spin_unlock(&q->lock); 3523 timeout = schedule_timeout(timeout); 3524 spin_lock_irq(&q->lock); 3525 __remove_wait_queue(q, &wait); 3526 spin_unlock_irqrestore(&q->lock, flags); 3527 3528 return timeout; 3529} 3530 3531void __sched interruptible_sleep_on(wait_queue_head_t *q) 3532{ 3533 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT); 3534} 3535EXPORT_SYMBOL(interruptible_sleep_on); 3536 3537long __sched 3538interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout) 3539{ 3540 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout); 3541} 3542EXPORT_SYMBOL(interruptible_sleep_on_timeout); 3543 3544void __sched sleep_on(wait_queue_head_t *q) 3545{ 3546 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT); 3547} 3548EXPORT_SYMBOL(sleep_on); 3549 3550long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout) 3551{ 3552 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout); 3553} 3554EXPORT_SYMBOL(sleep_on_timeout); 3555 3556#ifdef CONFIG_RT_MUTEXES 3557 3558/* 3559 * rt_mutex_setprio - set the current priority of a task 3560 * @p: task 3561 * @prio: prio value (kernel-internal form) 3562 * 3563 * This function changes the 'effective' priority of a task. It does 3564 * not touch ->normal_prio like __setscheduler(). 3565 * 3566 * Used by the rt_mutex code to implement priority inheritance logic. 3567 */ 3568void rt_mutex_setprio(struct task_struct *p, int prio) 3569{ 3570 int oldprio, on_rq, running; 3571 struct rq *rq; 3572 const struct sched_class *prev_class; 3573 3574 BUG_ON(prio < 0 || prio > MAX_PRIO); 3575 3576 rq = __task_rq_lock(p); 3577 3578 /* 3579 * Idle task boosting is a nono in general. There is one 3580 * exception, when PREEMPT_RT and NOHZ is active: 3581 * 3582 * The idle task calls get_next_timer_interrupt() and holds 3583 * the timer wheel base->lock on the CPU and another CPU wants 3584 * to access the timer (probably to cancel it). We can safely 3585 * ignore the boosting request, as the idle CPU runs this code 3586 * with interrupts disabled and will complete the lock 3587 * protected section without being interrupted. So there is no 3588 * real need to boost. 3589 */ 3590 if (unlikely(p == rq->idle)) { 3591 WARN_ON(p != rq->curr); 3592 WARN_ON(p->pi_blocked_on); 3593 goto out_unlock; 3594 } 3595 3596 trace_sched_pi_setprio(p, prio); 3597 oldprio = p->prio; 3598 prev_class = p->sched_class; 3599 on_rq = p->on_rq; 3600 running = task_current(rq, p); 3601 if (on_rq) 3602 dequeue_task(rq, p, 0); 3603 if (running) 3604 p->sched_class->put_prev_task(rq, p); 3605 3606 if (rt_prio(prio)) 3607 p->sched_class = &rt_sched_class; 3608 else 3609 p->sched_class = &fair_sched_class; 3610 3611 p->prio = prio; 3612 3613 if (running) 3614 p->sched_class->set_curr_task(rq); 3615 if (on_rq) 3616 enqueue_task(rq, p, oldprio < prio ? ENQUEUE_HEAD : 0); 3617 3618 check_class_changed(rq, p, prev_class, oldprio); 3619out_unlock: 3620 __task_rq_unlock(rq); 3621} 3622#endif 3623void set_user_nice(struct task_struct *p, long nice) 3624{ 3625 int old_prio, delta, on_rq; 3626 unsigned long flags; 3627 struct rq *rq; 3628 3629 if (TASK_NICE(p) == nice || nice < -20 || nice > 19) 3630 return; 3631 /* 3632 * We have to be careful, if called from sys_setpriority(), 3633 * the task might be in the middle of scheduling on another CPU. 3634 */ 3635 rq = task_rq_lock(p, &flags); 3636 /* 3637 * The RT priorities are set via sched_setscheduler(), but we still 3638 * allow the 'normal' nice value to be set - but as expected 3639 * it wont have any effect on scheduling until the task is 3640 * SCHED_FIFO/SCHED_RR: 3641 */ 3642 if (task_has_rt_policy(p)) { 3643 p->static_prio = NICE_TO_PRIO(nice); 3644 goto out_unlock; 3645 } 3646 on_rq = p->on_rq; 3647 if (on_rq) 3648 dequeue_task(rq, p, 0); 3649 3650 p->static_prio = NICE_TO_PRIO(nice); 3651 set_load_weight(p); 3652 old_prio = p->prio; 3653 p->prio = effective_prio(p); 3654 delta = p->prio - old_prio; 3655 3656 if (on_rq) { 3657 enqueue_task(rq, p, 0); 3658 /* 3659 * If the task increased its priority or is running and 3660 * lowered its priority, then reschedule its CPU: 3661 */ 3662 if (delta < 0 || (delta > 0 && task_running(rq, p))) 3663 resched_task(rq->curr); 3664 } 3665out_unlock: 3666 task_rq_unlock(rq, p, &flags); 3667} 3668EXPORT_SYMBOL(set_user_nice); 3669 3670/* 3671 * can_nice - check if a task can reduce its nice value 3672 * @p: task 3673 * @nice: nice value 3674 */ 3675int can_nice(const struct task_struct *p, const int nice) 3676{ 3677 /* convert nice value [19,-20] to rlimit style value [1,40] */ 3678 int nice_rlim = 20 - nice; 3679 3680 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) || 3681 capable(CAP_SYS_NICE)); 3682} 3683 3684#ifdef __ARCH_WANT_SYS_NICE 3685 3686/* 3687 * sys_nice - change the priority of the current process. 3688 * @increment: priority increment 3689 * 3690 * sys_setpriority is a more generic, but much slower function that 3691 * does similar things. 3692 */ 3693SYSCALL_DEFINE1(nice, int, increment) 3694{ 3695 long nice, retval; 3696 3697 /* 3698 * Setpriority might change our priority at the same moment. 3699 * We don't have to worry. Conceptually one call occurs first 3700 * and we have a single winner. 3701 */ 3702 if (increment < -40) 3703 increment = -40; 3704 if (increment > 40) 3705 increment = 40; 3706 3707 nice = TASK_NICE(current) + increment; 3708 if (nice < -20) 3709 nice = -20; 3710 if (nice > 19) 3711 nice = 19; 3712 3713 if (increment < 0 && !can_nice(current, nice)) 3714 return -EPERM; 3715 3716 retval = security_task_setnice(current, nice); 3717 if (retval) 3718 return retval; 3719 3720 set_user_nice(current, nice); 3721 return 0; 3722} 3723 3724#endif 3725 3726/** 3727 * task_prio - return the priority value of a given task. 3728 * @p: the task in question. 3729 * 3730 * This is the priority value as seen by users in /proc. 3731 * RT tasks are offset by -200. Normal tasks are centered 3732 * around 0, value goes from -16 to +15. 3733 */ 3734int task_prio(const struct task_struct *p) 3735{ 3736 return p->prio - MAX_RT_PRIO; 3737} 3738 3739/** 3740 * task_nice - return the nice value of a given task. 3741 * @p: the task in question. 3742 */ 3743int task_nice(const struct task_struct *p) 3744{ 3745 return TASK_NICE(p); 3746} 3747EXPORT_SYMBOL(task_nice); 3748 3749/** 3750 * idle_cpu - is a given cpu idle currently? 3751 * @cpu: the processor in question. 3752 */ 3753int idle_cpu(int cpu) 3754{ 3755 struct rq *rq = cpu_rq(cpu); 3756 3757 if (rq->curr != rq->idle) 3758 return 0; 3759 3760 if (rq->nr_running) 3761 return 0; 3762 3763#ifdef CONFIG_SMP 3764 if (!llist_empty(&rq->wake_list)) 3765 return 0; 3766#endif 3767 3768 return 1; 3769} 3770 3771/** 3772 * idle_task - return the idle task for a given cpu. 3773 * @cpu: the processor in question. 3774 */ 3775struct task_struct *idle_task(int cpu) 3776{ 3777 return cpu_rq(cpu)->idle; 3778} 3779 3780/** 3781 * find_process_by_pid - find a process with a matching PID value. 3782 * @pid: the pid in question. 3783 */ 3784static struct task_struct *find_process_by_pid(pid_t pid) 3785{ 3786 return pid ? find_task_by_vpid(pid) : current; 3787} 3788 3789/* Actually do priority change: must hold rq lock. */ 3790static void 3791__setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio) 3792{ 3793 p->policy = policy; 3794 p->rt_priority = prio; 3795 p->normal_prio = normal_prio(p); 3796 /* we are holding p->pi_lock already */ 3797 p->prio = rt_mutex_getprio(p); 3798 if (rt_prio(p->prio)) 3799 p->sched_class = &rt_sched_class; 3800 else 3801 p->sched_class = &fair_sched_class; 3802 set_load_weight(p); 3803} 3804 3805/* 3806 * check the target process has a UID that matches the current process's 3807 */ 3808static bool check_same_owner(struct task_struct *p) 3809{ 3810 const struct cred *cred = current_cred(), *pcred; 3811 bool match; 3812 3813 rcu_read_lock(); 3814 pcred = __task_cred(p); 3815 match = (uid_eq(cred->euid, pcred->euid) || 3816 uid_eq(cred->euid, pcred->uid)); 3817 rcu_read_unlock(); 3818 return match; 3819} 3820 3821static int __sched_setscheduler(struct task_struct *p, int policy, 3822 const struct sched_param *param, bool user) 3823{ 3824 int retval, oldprio, oldpolicy = -1, on_rq, running; 3825 unsigned long flags; 3826 const struct sched_class *prev_class; 3827 struct rq *rq; 3828 int reset_on_fork; 3829 3830 /* may grab non-irq protected spin_locks */ 3831 BUG_ON(in_interrupt()); 3832recheck: 3833 /* double check policy once rq lock held */ 3834 if (policy < 0) { 3835 reset_on_fork = p->sched_reset_on_fork; 3836 policy = oldpolicy = p->policy; 3837 } else { 3838 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK); 3839 policy &= ~SCHED_RESET_ON_FORK; 3840 3841 if (policy != SCHED_FIFO && policy != SCHED_RR && 3842 policy != SCHED_NORMAL && policy != SCHED_BATCH && 3843 policy != SCHED_IDLE) 3844 return -EINVAL; 3845 } 3846 3847 /* 3848 * Valid priorities for SCHED_FIFO and SCHED_RR are 3849 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL, 3850 * SCHED_BATCH and SCHED_IDLE is 0. 3851 */ 3852 if (param->sched_priority < 0 || 3853 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) || 3854 (!p->mm && param->sched_priority > MAX_RT_PRIO-1)) 3855 return -EINVAL; 3856 if (rt_policy(policy) != (param->sched_priority != 0)) 3857 return -EINVAL; 3858 3859 /* 3860 * Allow unprivileged RT tasks to decrease priority: 3861 */ 3862 if (user && !capable(CAP_SYS_NICE)) { 3863 if (rt_policy(policy)) { 3864 unsigned long rlim_rtprio = 3865 task_rlimit(p, RLIMIT_RTPRIO); 3866 3867 /* can't set/change the rt policy */ 3868 if (policy != p->policy && !rlim_rtprio) 3869 return -EPERM; 3870 3871 /* can't increase priority */ 3872 if (param->sched_priority > p->rt_priority && 3873 param->sched_priority > rlim_rtprio) 3874 return -EPERM; 3875 } 3876 3877 /* 3878 * Treat SCHED_IDLE as nice 20. Only allow a switch to 3879 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it. 3880 */ 3881 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) { 3882 if (!can_nice(p, TASK_NICE(p))) 3883 return -EPERM; 3884 } 3885 3886 /* can't change other user's priorities */ 3887 if (!check_same_owner(p)) 3888 return -EPERM; 3889 3890 /* Normal users shall not reset the sched_reset_on_fork flag */ 3891 if (p->sched_reset_on_fork && !reset_on_fork) 3892 return -EPERM; 3893 } 3894 3895 if (user) { 3896 retval = security_task_setscheduler(p); 3897 if (retval) 3898 return retval; 3899 } 3900 3901 /* 3902 * make sure no PI-waiters arrive (or leave) while we are 3903 * changing the priority of the task: 3904 * 3905 * To be able to change p->policy safely, the appropriate 3906 * runqueue lock must be held. 3907 */ 3908 rq = task_rq_lock(p, &flags); 3909 3910 /* 3911 * Changing the policy of the stop threads its a very bad idea 3912 */ 3913 if (p == rq->stop) { 3914 task_rq_unlock(rq, p, &flags); 3915 return -EINVAL; 3916 } 3917 3918 /* 3919 * If not changing anything there's no need to proceed further: 3920 */ 3921 if (unlikely(policy == p->policy && (!rt_policy(policy) || 3922 param->sched_priority == p->rt_priority))) { 3923 task_rq_unlock(rq, p, &flags); 3924 return 0; 3925 } 3926 3927#ifdef CONFIG_RT_GROUP_SCHED 3928 if (user) { 3929 /* 3930 * Do not allow realtime tasks into groups that have no runtime 3931 * assigned. 3932 */ 3933 if (rt_bandwidth_enabled() && rt_policy(policy) && 3934 task_group(p)->rt_bandwidth.rt_runtime == 0 && 3935 !task_group_is_autogroup(task_group(p))) { 3936 task_rq_unlock(rq, p, &flags); 3937 return -EPERM; 3938 } 3939 } 3940#endif 3941 3942 /* recheck policy now with rq lock held */ 3943 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) { 3944 policy = oldpolicy = -1; 3945 task_rq_unlock(rq, p, &flags); 3946 goto recheck; 3947 } 3948 on_rq = p->on_rq; 3949 running = task_current(rq, p); 3950 if (on_rq) 3951 dequeue_task(rq, p, 0); 3952 if (running) 3953 p->sched_class->put_prev_task(rq, p); 3954 3955 p->sched_reset_on_fork = reset_on_fork; 3956 3957 oldprio = p->prio; 3958 prev_class = p->sched_class; 3959 __setscheduler(rq, p, policy, param->sched_priority); 3960 3961 if (running) 3962 p->sched_class->set_curr_task(rq); 3963 if (on_rq) 3964 enqueue_task(rq, p, 0); 3965 3966 check_class_changed(rq, p, prev_class, oldprio); 3967 task_rq_unlock(rq, p, &flags); 3968 3969 rt_mutex_adjust_pi(p); 3970 3971 return 0; 3972} 3973 3974/** 3975 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread. 3976 * @p: the task in question. 3977 * @policy: new policy. 3978 * @param: structure containing the new RT priority. 3979 * 3980 * NOTE that the task may be already dead. 3981 */ 3982int sched_setscheduler(struct task_struct *p, int policy, 3983 const struct sched_param *param) 3984{ 3985 return __sched_setscheduler(p, policy, param, true); 3986} 3987EXPORT_SYMBOL_GPL(sched_setscheduler); 3988 3989/** 3990 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace. 3991 * @p: the task in question. 3992 * @policy: new policy. 3993 * @param: structure containing the new RT priority. 3994 * 3995 * Just like sched_setscheduler, only don't bother checking if the 3996 * current context has permission. For example, this is needed in 3997 * stop_machine(): we create temporary high priority worker threads, 3998 * but our caller might not have that capability. 3999 */ 4000int sched_setscheduler_nocheck(struct task_struct *p, int policy, 4001 const struct sched_param *param) 4002{ 4003 return __sched_setscheduler(p, policy, param, false); 4004} 4005 4006static int 4007do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param) 4008{ 4009 struct sched_param lparam; 4010 struct task_struct *p; 4011 int retval; 4012 4013 if (!param || pid < 0) 4014 return -EINVAL; 4015 if (copy_from_user(&lparam, param, sizeof(struct sched_param))) 4016 return -EFAULT; 4017 4018 rcu_read_lock(); 4019 retval = -ESRCH; 4020 p = find_process_by_pid(pid); 4021 if (p != NULL) 4022 retval = sched_setscheduler(p, policy, &lparam); 4023 rcu_read_unlock(); 4024 4025 return retval; 4026} 4027 4028/** 4029 * sys_sched_setscheduler - set/change the scheduler policy and RT priority 4030 * @pid: the pid in question. 4031 * @policy: new policy. 4032 * @param: structure containing the new RT priority. 4033 */ 4034SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy, 4035 struct sched_param __user *, param) 4036{ 4037 /* negative values for policy are not valid */ 4038 if (policy < 0) 4039 return -EINVAL; 4040 4041 return do_sched_setscheduler(pid, policy, param); 4042} 4043 4044/** 4045 * sys_sched_setparam - set/change the RT priority of a thread 4046 * @pid: the pid in question. 4047 * @param: structure containing the new RT priority. 4048 */ 4049SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param) 4050{ 4051 return do_sched_setscheduler(pid, -1, param); 4052} 4053 4054/** 4055 * sys_sched_getscheduler - get the policy (scheduling class) of a thread 4056 * @pid: the pid in question. 4057 */ 4058SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid) 4059{ 4060 struct task_struct *p; 4061 int retval; 4062 4063 if (pid < 0) 4064 return -EINVAL; 4065 4066 retval = -ESRCH; 4067 rcu_read_lock(); 4068 p = find_process_by_pid(pid); 4069 if (p) { 4070 retval = security_task_getscheduler(p); 4071 if (!retval) 4072 retval = p->policy 4073 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0); 4074 } 4075 rcu_read_unlock(); 4076 return retval; 4077} 4078 4079/** 4080 * sys_sched_getparam - get the RT priority of a thread 4081 * @pid: the pid in question. 4082 * @param: structure containing the RT priority. 4083 */ 4084SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param) 4085{ 4086 struct sched_param lp; 4087 struct task_struct *p; 4088 int retval; 4089 4090 if (!param || pid < 0) 4091 return -EINVAL; 4092 4093 rcu_read_lock(); 4094 p = find_process_by_pid(pid); 4095 retval = -ESRCH; 4096 if (!p) 4097 goto out_unlock; 4098 4099 retval = security_task_getscheduler(p); 4100 if (retval) 4101 goto out_unlock; 4102 4103 lp.sched_priority = p->rt_priority; 4104 rcu_read_unlock(); 4105 4106 /* 4107 * This one might sleep, we cannot do it with a spinlock held ... 4108 */ 4109 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0; 4110 4111 return retval; 4112 4113out_unlock: 4114 rcu_read_unlock(); 4115 return retval; 4116} 4117 4118long sched_setaffinity(pid_t pid, const struct cpumask *in_mask) 4119{ 4120 cpumask_var_t cpus_allowed, new_mask; 4121 struct task_struct *p; 4122 int retval; 4123 4124 get_online_cpus(); 4125 rcu_read_lock(); 4126 4127 p = find_process_by_pid(pid); 4128 if (!p) { 4129 rcu_read_unlock(); 4130 put_online_cpus(); 4131 return -ESRCH; 4132 } 4133 4134 /* Prevent p going away */ 4135 get_task_struct(p); 4136 rcu_read_unlock(); 4137 4138 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) { 4139 retval = -ENOMEM; 4140 goto out_put_task; 4141 } 4142 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) { 4143 retval = -ENOMEM; 4144 goto out_free_cpus_allowed; 4145 } 4146 retval = -EPERM; 4147 if (!check_same_owner(p)) { 4148 rcu_read_lock(); 4149 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) { 4150 rcu_read_unlock(); 4151 goto out_unlock; 4152 } 4153 rcu_read_unlock(); 4154 } 4155 4156 retval = security_task_setscheduler(p); 4157 if (retval) 4158 goto out_unlock; 4159 4160 cpuset_cpus_allowed(p, cpus_allowed); 4161 cpumask_and(new_mask, in_mask, cpus_allowed); 4162again: 4163 retval = set_cpus_allowed_ptr(p, new_mask); 4164 4165 if (!retval) { 4166 cpuset_cpus_allowed(p, cpus_allowed); 4167 if (!cpumask_subset(new_mask, cpus_allowed)) { 4168 /* 4169 * We must have raced with a concurrent cpuset 4170 * update. Just reset the cpus_allowed to the 4171 * cpuset's cpus_allowed 4172 */ 4173 cpumask_copy(new_mask, cpus_allowed); 4174 goto again; 4175 } 4176 } 4177out_unlock: 4178 free_cpumask_var(new_mask); 4179out_free_cpus_allowed: 4180 free_cpumask_var(cpus_allowed); 4181out_put_task: 4182 put_task_struct(p); 4183 put_online_cpus(); 4184 return retval; 4185} 4186 4187static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len, 4188 struct cpumask *new_mask) 4189{ 4190 if (len < cpumask_size()) 4191 cpumask_clear(new_mask); 4192 else if (len > cpumask_size()) 4193 len = cpumask_size(); 4194 4195 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0; 4196} 4197 4198/** 4199 * sys_sched_setaffinity - set the cpu affinity of a process 4200 * @pid: pid of the process 4201 * @len: length in bytes of the bitmask pointed to by user_mask_ptr 4202 * @user_mask_ptr: user-space pointer to the new cpu mask 4203 */ 4204SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len, 4205 unsigned long __user *, user_mask_ptr) 4206{ 4207 cpumask_var_t new_mask; 4208 int retval; 4209 4210 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) 4211 return -ENOMEM; 4212 4213 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask); 4214 if (retval == 0) 4215 retval = sched_setaffinity(pid, new_mask); 4216 free_cpumask_var(new_mask); 4217 return retval; 4218} 4219 4220long sched_getaffinity(pid_t pid, struct cpumask *mask) 4221{ 4222 struct task_struct *p; 4223 unsigned long flags; 4224 int retval; 4225 4226 get_online_cpus(); 4227 rcu_read_lock(); 4228 4229 retval = -ESRCH; 4230 p = find_process_by_pid(pid); 4231 if (!p) 4232 goto out_unlock; 4233 4234 retval = security_task_getscheduler(p); 4235 if (retval) 4236 goto out_unlock; 4237 4238 raw_spin_lock_irqsave(&p->pi_lock, flags); 4239 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask); 4240 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 4241 4242out_unlock: 4243 rcu_read_unlock(); 4244 put_online_cpus(); 4245 4246 return retval; 4247} 4248 4249/** 4250 * sys_sched_getaffinity - get the cpu affinity of a process 4251 * @pid: pid of the process 4252 * @len: length in bytes of the bitmask pointed to by user_mask_ptr 4253 * @user_mask_ptr: user-space pointer to hold the current cpu mask 4254 */ 4255SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len, 4256 unsigned long __user *, user_mask_ptr) 4257{ 4258 int ret; 4259 cpumask_var_t mask; 4260 4261 if ((len * BITS_PER_BYTE) < nr_cpu_ids) 4262 return -EINVAL; 4263 if (len & (sizeof(unsigned long)-1)) 4264 return -EINVAL; 4265 4266 if (!alloc_cpumask_var(&mask, GFP_KERNEL)) 4267 return -ENOMEM; 4268 4269 ret = sched_getaffinity(pid, mask); 4270 if (ret == 0) { 4271 size_t retlen = min_t(size_t, len, cpumask_size()); 4272 4273 if (copy_to_user(user_mask_ptr, mask, retlen)) 4274 ret = -EFAULT; 4275 else 4276 ret = retlen; 4277 } 4278 free_cpumask_var(mask); 4279 4280 return ret; 4281} 4282 4283/** 4284 * sys_sched_yield - yield the current processor to other threads. 4285 * 4286 * This function yields the current CPU to other tasks. If there are no 4287 * other threads running on this CPU then this function will return. 4288 */ 4289SYSCALL_DEFINE0(sched_yield) 4290{ 4291 struct rq *rq = this_rq_lock(); 4292 4293 schedstat_inc(rq, yld_count); 4294 current->sched_class->yield_task(rq); 4295 4296 /* 4297 * Since we are going to call schedule() anyway, there's 4298 * no need to preempt or enable interrupts: 4299 */ 4300 __release(rq->lock); 4301 spin_release(&rq->lock.dep_map, 1, _THIS_IP_); 4302 do_raw_spin_unlock(&rq->lock); 4303 sched_preempt_enable_no_resched(); 4304 4305 schedule(); 4306 4307 return 0; 4308} 4309 4310static inline int should_resched(void) 4311{ 4312 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE); 4313} 4314 4315static void __cond_resched(void) 4316{ 4317 add_preempt_count(PREEMPT_ACTIVE); 4318 __schedule(); 4319 sub_preempt_count(PREEMPT_ACTIVE); 4320} 4321 4322int __sched _cond_resched(void) 4323{ 4324 if (should_resched()) { 4325 __cond_resched(); 4326 return 1; 4327 } 4328 return 0; 4329} 4330EXPORT_SYMBOL(_cond_resched); 4331 4332/* 4333 * __cond_resched_lock() - if a reschedule is pending, drop the given lock, 4334 * call schedule, and on return reacquire the lock. 4335 * 4336 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level 4337 * operations here to prevent schedule() from being called twice (once via 4338 * spin_unlock(), once by hand). 4339 */ 4340int __cond_resched_lock(spinlock_t *lock) 4341{ 4342 int resched = should_resched(); 4343 int ret = 0; 4344 4345 lockdep_assert_held(lock); 4346 4347 if (spin_needbreak(lock) || resched) { 4348 spin_unlock(lock); 4349 if (resched) 4350 __cond_resched(); 4351 else 4352 cpu_relax(); 4353 ret = 1; 4354 spin_lock(lock); 4355 } 4356 return ret; 4357} 4358EXPORT_SYMBOL(__cond_resched_lock); 4359 4360int __sched __cond_resched_softirq(void) 4361{ 4362 BUG_ON(!in_softirq()); 4363 4364 if (should_resched()) { 4365 local_bh_enable(); 4366 __cond_resched(); 4367 local_bh_disable(); 4368 return 1; 4369 } 4370 return 0; 4371} 4372EXPORT_SYMBOL(__cond_resched_softirq); 4373 4374/** 4375 * yield - yield the current processor to other threads. 4376 * 4377 * Do not ever use this function, there's a 99% chance you're doing it wrong. 4378 * 4379 * The scheduler is at all times free to pick the calling task as the most 4380 * eligible task to run, if removing the yield() call from your code breaks 4381 * it, its already broken. 4382 * 4383 * Typical broken usage is: 4384 * 4385 * while (!event) 4386 * yield(); 4387 * 4388 * where one assumes that yield() will let 'the other' process run that will 4389 * make event true. If the current task is a SCHED_FIFO task that will never 4390 * happen. Never use yield() as a progress guarantee!! 4391 * 4392 * If you want to use yield() to wait for something, use wait_event(). 4393 * If you want to use yield() to be 'nice' for others, use cond_resched(). 4394 * If you still want to use yield(), do not! 4395 */ 4396void __sched yield(void) 4397{ 4398 set_current_state(TASK_RUNNING); 4399 sys_sched_yield(); 4400} 4401EXPORT_SYMBOL(yield); 4402 4403/** 4404 * yield_to - yield the current processor to another thread in 4405 * your thread group, or accelerate that thread toward the 4406 * processor it's on. 4407 * @p: target task 4408 * @preempt: whether task preemption is allowed or not 4409 * 4410 * It's the caller's job to ensure that the target task struct 4411 * can't go away on us before we can do any checks. 4412 * 4413 * Returns true if we indeed boosted the target task. 4414 */ 4415bool __sched yield_to(struct task_struct *p, bool preempt) 4416{ 4417 struct task_struct *curr = current; 4418 struct rq *rq, *p_rq; 4419 unsigned long flags; 4420 bool yielded = 0; 4421 4422 local_irq_save(flags); 4423 rq = this_rq(); 4424 4425again: 4426 p_rq = task_rq(p); 4427 double_rq_lock(rq, p_rq); 4428 while (task_rq(p) != p_rq) { 4429 double_rq_unlock(rq, p_rq); 4430 goto again; 4431 } 4432 4433 if (!curr->sched_class->yield_to_task) 4434 goto out; 4435 4436 if (curr->sched_class != p->sched_class) 4437 goto out; 4438 4439 if (task_running(p_rq, p) || p->state) 4440 goto out; 4441 4442 yielded = curr->sched_class->yield_to_task(rq, p, preempt); 4443 if (yielded) { 4444 schedstat_inc(rq, yld_count); 4445 /* 4446 * Make p's CPU reschedule; pick_next_entity takes care of 4447 * fairness. 4448 */ 4449 if (preempt && rq != p_rq) 4450 resched_task(p_rq->curr); 4451 } 4452 4453out: 4454 double_rq_unlock(rq, p_rq); 4455 local_irq_restore(flags); 4456 4457 if (yielded) 4458 schedule(); 4459 4460 return yielded; 4461} 4462EXPORT_SYMBOL_GPL(yield_to); 4463 4464/* 4465 * This task is about to go to sleep on IO. Increment rq->nr_iowait so 4466 * that process accounting knows that this is a task in IO wait state. 4467 */ 4468void __sched io_schedule(void) 4469{ 4470 struct rq *rq = raw_rq(); 4471 4472 delayacct_blkio_start(); 4473 atomic_inc(&rq->nr_iowait); 4474 blk_flush_plug(current); 4475 current->in_iowait = 1; 4476 schedule(); 4477 current->in_iowait = 0; 4478 atomic_dec(&rq->nr_iowait); 4479 delayacct_blkio_end(); 4480} 4481EXPORT_SYMBOL(io_schedule); 4482 4483long __sched io_schedule_timeout(long timeout) 4484{ 4485 struct rq *rq = raw_rq(); 4486 long ret; 4487 4488 delayacct_blkio_start(); 4489 atomic_inc(&rq->nr_iowait); 4490 blk_flush_plug(current); 4491 current->in_iowait = 1; 4492 ret = schedule_timeout(timeout); 4493 current->in_iowait = 0; 4494 atomic_dec(&rq->nr_iowait); 4495 delayacct_blkio_end(); 4496 return ret; 4497} 4498 4499/** 4500 * sys_sched_get_priority_max - return maximum RT priority. 4501 * @policy: scheduling class. 4502 * 4503 * this syscall returns the maximum rt_priority that can be used 4504 * by a given scheduling class. 4505 */ 4506SYSCALL_DEFINE1(sched_get_priority_max, int, policy) 4507{ 4508 int ret = -EINVAL; 4509 4510 switch (policy) { 4511 case SCHED_FIFO: 4512 case SCHED_RR: 4513 ret = MAX_USER_RT_PRIO-1; 4514 break; 4515 case SCHED_NORMAL: 4516 case SCHED_BATCH: 4517 case SCHED_IDLE: 4518 ret = 0; 4519 break; 4520 } 4521 return ret; 4522} 4523 4524/** 4525 * sys_sched_get_priority_min - return minimum RT priority. 4526 * @policy: scheduling class. 4527 * 4528 * this syscall returns the minimum rt_priority that can be used 4529 * by a given scheduling class. 4530 */ 4531SYSCALL_DEFINE1(sched_get_priority_min, int, policy) 4532{ 4533 int ret = -EINVAL; 4534 4535 switch (policy) { 4536 case SCHED_FIFO: 4537 case SCHED_RR: 4538 ret = 1; 4539 break; 4540 case SCHED_NORMAL: 4541 case SCHED_BATCH: 4542 case SCHED_IDLE: 4543 ret = 0; 4544 } 4545 return ret; 4546} 4547 4548/** 4549 * sys_sched_rr_get_interval - return the default timeslice of a process. 4550 * @pid: pid of the process. 4551 * @interval: userspace pointer to the timeslice value. 4552 * 4553 * this syscall writes the default timeslice value of a given process 4554 * into the user-space timespec buffer. A value of '0' means infinity. 4555 */ 4556SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid, 4557 struct timespec __user *, interval) 4558{ 4559 struct task_struct *p; 4560 unsigned int time_slice; 4561 unsigned long flags; 4562 struct rq *rq; 4563 int retval; 4564 struct timespec t; 4565 4566 if (pid < 0) 4567 return -EINVAL; 4568 4569 retval = -ESRCH; 4570 rcu_read_lock(); 4571 p = find_process_by_pid(pid); 4572 if (!p) 4573 goto out_unlock; 4574 4575 retval = security_task_getscheduler(p); 4576 if (retval) 4577 goto out_unlock; 4578 4579 rq = task_rq_lock(p, &flags); 4580 time_slice = p->sched_class->get_rr_interval(rq, p); 4581 task_rq_unlock(rq, p, &flags); 4582 4583 rcu_read_unlock(); 4584 jiffies_to_timespec(time_slice, &t); 4585 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0; 4586 return retval; 4587 4588out_unlock: 4589 rcu_read_unlock(); 4590 return retval; 4591} 4592 4593static const char stat_nam[] = TASK_STATE_TO_CHAR_STR; 4594 4595void sched_show_task(struct task_struct *p) 4596{ 4597 unsigned long free = 0; 4598 int ppid; 4599 unsigned state; 4600 4601 state = p->state ? __ffs(p->state) + 1 : 0; 4602 printk(KERN_INFO "%-15.15s %c", p->comm, 4603 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?'); 4604#if BITS_PER_LONG == 32 4605 if (state == TASK_RUNNING) 4606 printk(KERN_CONT " running "); 4607 else 4608 printk(KERN_CONT " %08lx ", thread_saved_pc(p)); 4609#else 4610 if (state == TASK_RUNNING) 4611 printk(KERN_CONT " running task "); 4612 else 4613 printk(KERN_CONT " %016lx ", thread_saved_pc(p)); 4614#endif 4615#ifdef CONFIG_DEBUG_STACK_USAGE 4616 free = stack_not_used(p); 4617#endif 4618 rcu_read_lock(); 4619 ppid = task_pid_nr(rcu_dereference(p->real_parent)); 4620 rcu_read_unlock(); 4621 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free, 4622 task_pid_nr(p), ppid, 4623 (unsigned long)task_thread_info(p)->flags); 4624 4625 show_stack(p, NULL); 4626} 4627 4628void show_state_filter(unsigned long state_filter) 4629{ 4630 struct task_struct *g, *p; 4631 4632#if BITS_PER_LONG == 32 4633 printk(KERN_INFO 4634 " task PC stack pid father\n"); 4635#else 4636 printk(KERN_INFO 4637 " task PC stack pid father\n"); 4638#endif 4639 rcu_read_lock(); 4640 do_each_thread(g, p) { 4641 /* 4642 * reset the NMI-timeout, listing all files on a slow 4643 * console might take a lot of time: 4644 */ 4645 touch_nmi_watchdog(); 4646 if (!state_filter || (p->state & state_filter)) 4647 sched_show_task(p); 4648 } while_each_thread(g, p); 4649 4650 touch_all_softlockup_watchdogs(); 4651 4652#ifdef CONFIG_SCHED_DEBUG 4653 sysrq_sched_debug_show(); 4654#endif 4655 rcu_read_unlock(); 4656 /* 4657 * Only show locks if all tasks are dumped: 4658 */ 4659 if (!state_filter) 4660 debug_show_all_locks(); 4661} 4662 4663void __cpuinit init_idle_bootup_task(struct task_struct *idle) 4664{ 4665 idle->sched_class = &idle_sched_class; 4666} 4667 4668/** 4669 * init_idle - set up an idle thread for a given CPU 4670 * @idle: task in question 4671 * @cpu: cpu the idle task belongs to 4672 * 4673 * NOTE: this function does not set the idle thread's NEED_RESCHED 4674 * flag, to make booting more robust. 4675 */ 4676void __cpuinit init_idle(struct task_struct *idle, int cpu) 4677{ 4678 struct rq *rq = cpu_rq(cpu); 4679 unsigned long flags; 4680 4681 raw_spin_lock_irqsave(&rq->lock, flags); 4682 4683 __sched_fork(idle); 4684 idle->state = TASK_RUNNING; 4685 idle->se.exec_start = sched_clock(); 4686 4687 do_set_cpus_allowed(idle, cpumask_of(cpu)); 4688 /* 4689 * We're having a chicken and egg problem, even though we are 4690 * holding rq->lock, the cpu isn't yet set to this cpu so the 4691 * lockdep check in task_group() will fail. 4692 * 4693 * Similar case to sched_fork(). / Alternatively we could 4694 * use task_rq_lock() here and obtain the other rq->lock. 4695 * 4696 * Silence PROVE_RCU 4697 */ 4698 rcu_read_lock(); 4699 __set_task_cpu(idle, cpu); 4700 rcu_read_unlock(); 4701 4702 rq->curr = rq->idle = idle; 4703#if defined(CONFIG_SMP) 4704 idle->on_cpu = 1; 4705#endif 4706 raw_spin_unlock_irqrestore(&rq->lock, flags); 4707 4708 /* Set the preempt count _outside_ the spinlocks! */ 4709 task_thread_info(idle)->preempt_count = 0; 4710 4711 /* 4712 * The idle tasks have their own, simple scheduling class: 4713 */ 4714 idle->sched_class = &idle_sched_class; 4715 ftrace_graph_init_idle_task(idle, cpu); 4716#if defined(CONFIG_SMP) 4717 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu); 4718#endif 4719} 4720 4721#ifdef CONFIG_SMP 4722void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask) 4723{ 4724 if (p->sched_class && p->sched_class->set_cpus_allowed) 4725 p->sched_class->set_cpus_allowed(p, new_mask); 4726 4727 cpumask_copy(&p->cpus_allowed, new_mask); 4728 p->nr_cpus_allowed = cpumask_weight(new_mask); 4729} 4730 4731/* 4732 * This is how migration works: 4733 * 4734 * 1) we invoke migration_cpu_stop() on the target CPU using 4735 * stop_one_cpu(). 4736 * 2) stopper starts to run (implicitly forcing the migrated thread 4737 * off the CPU) 4738 * 3) it checks whether the migrated task is still in the wrong runqueue. 4739 * 4) if it's in the wrong runqueue then the migration thread removes 4740 * it and puts it into the right queue. 4741 * 5) stopper completes and stop_one_cpu() returns and the migration 4742 * is done. 4743 */ 4744 4745/* 4746 * Change a given task's CPU affinity. Migrate the thread to a 4747 * proper CPU and schedule it away if the CPU it's executing on 4748 * is removed from the allowed bitmask. 4749 * 4750 * NOTE: the caller must have a valid reference to the task, the 4751 * task must not exit() & deallocate itself prematurely. The 4752 * call is not atomic; no spinlocks may be held. 4753 */ 4754int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask) 4755{ 4756 unsigned long flags; 4757 struct rq *rq; 4758 unsigned int dest_cpu; 4759 int ret = 0; 4760 4761 rq = task_rq_lock(p, &flags); 4762 4763 if (cpumask_equal(&p->cpus_allowed, new_mask)) 4764 goto out; 4765 4766 if (!cpumask_intersects(new_mask, cpu_active_mask)) { 4767 ret = -EINVAL; 4768 goto out; 4769 } 4770 4771 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current)) { 4772 ret = -EINVAL; 4773 goto out; 4774 } 4775 4776 do_set_cpus_allowed(p, new_mask); 4777 4778 /* Can the task run on the task's current CPU? If so, we're done */ 4779 if (cpumask_test_cpu(task_cpu(p), new_mask)) 4780 goto out; 4781 4782 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask); 4783 if (p->on_rq) { 4784 struct migration_arg arg = { p, dest_cpu }; 4785 /* Need help from migration thread: drop lock and wait. */ 4786 task_rq_unlock(rq, p, &flags); 4787 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg); 4788 tlb_migrate_finish(p->mm); 4789 return 0; 4790 } 4791out: 4792 task_rq_unlock(rq, p, &flags); 4793 4794 return ret; 4795} 4796EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr); 4797 4798/* 4799 * Move (not current) task off this cpu, onto dest cpu. We're doing 4800 * this because either it can't run here any more (set_cpus_allowed() 4801 * away from this CPU, or CPU going down), or because we're 4802 * attempting to rebalance this task on exec (sched_exec). 4803 * 4804 * So we race with normal scheduler movements, but that's OK, as long 4805 * as the task is no longer on this CPU. 4806 * 4807 * Returns non-zero if task was successfully migrated. 4808 */ 4809static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu) 4810{ 4811 struct rq *rq_dest, *rq_src; 4812 int ret = 0; 4813 4814 if (unlikely(!cpu_active(dest_cpu))) 4815 return ret; 4816 4817 rq_src = cpu_rq(src_cpu); 4818 rq_dest = cpu_rq(dest_cpu); 4819 4820 raw_spin_lock(&p->pi_lock); 4821 double_rq_lock(rq_src, rq_dest); 4822 /* Already moved. */ 4823 if (task_cpu(p) != src_cpu) 4824 goto done; 4825 /* Affinity changed (again). */ 4826 if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p))) 4827 goto fail; 4828 4829 /* 4830 * If we're not on a rq, the next wake-up will ensure we're 4831 * placed properly. 4832 */ 4833 if (p->on_rq) { 4834 dequeue_task(rq_src, p, 0); 4835 set_task_cpu(p, dest_cpu); 4836 enqueue_task(rq_dest, p, 0); 4837 check_preempt_curr(rq_dest, p, 0); 4838 } 4839done: 4840 ret = 1; 4841fail: 4842 double_rq_unlock(rq_src, rq_dest); 4843 raw_spin_unlock(&p->pi_lock); 4844 return ret; 4845} 4846 4847/* 4848 * migration_cpu_stop - this will be executed by a highprio stopper thread 4849 * and performs thread migration by bumping thread off CPU then 4850 * 'pushing' onto another runqueue. 4851 */ 4852static int migration_cpu_stop(void *data) 4853{ 4854 struct migration_arg *arg = data; 4855 4856 /* 4857 * The original target cpu might have gone down and we might 4858 * be on another cpu but it doesn't matter. 4859 */ 4860 local_irq_disable(); 4861 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu); 4862 local_irq_enable(); 4863 return 0; 4864} 4865 4866#ifdef CONFIG_HOTPLUG_CPU 4867 4868/* 4869 * Ensures that the idle task is using init_mm right before its cpu goes 4870 * offline. 4871 */ 4872void idle_task_exit(void) 4873{ 4874 struct mm_struct *mm = current->active_mm; 4875 4876 BUG_ON(cpu_online(smp_processor_id())); 4877 4878 if (mm != &init_mm) 4879 switch_mm(mm, &init_mm, current); 4880 mmdrop(mm); 4881} 4882 4883/* 4884 * Since this CPU is going 'away' for a while, fold any nr_active delta 4885 * we might have. Assumes we're called after migrate_tasks() so that the 4886 * nr_active count is stable. 4887 * 4888 * Also see the comment "Global load-average calculations". 4889 */ 4890static void calc_load_migrate(struct rq *rq) 4891{ 4892 long delta = calc_load_fold_active(rq); 4893 if (delta) 4894 atomic_long_add(delta, &calc_load_tasks); 4895} 4896 4897/* 4898 * Migrate all tasks from the rq, sleeping tasks will be migrated by 4899 * try_to_wake_up()->select_task_rq(). 4900 * 4901 * Called with rq->lock held even though we'er in stop_machine() and 4902 * there's no concurrency possible, we hold the required locks anyway 4903 * because of lock validation efforts. 4904 */ 4905static void migrate_tasks(unsigned int dead_cpu) 4906{ 4907 struct rq *rq = cpu_rq(dead_cpu); 4908 struct task_struct *next, *stop = rq->stop; 4909 int dest_cpu; 4910 4911 /* 4912 * Fudge the rq selection such that the below task selection loop 4913 * doesn't get stuck on the currently eligible stop task. 4914 * 4915 * We're currently inside stop_machine() and the rq is either stuck 4916 * in the stop_machine_cpu_stop() loop, or we're executing this code, 4917 * either way we should never end up calling schedule() until we're 4918 * done here. 4919 */ 4920 rq->stop = NULL; 4921 4922 for ( ; ; ) { 4923 /* 4924 * There's this thread running, bail when that's the only 4925 * remaining thread. 4926 */ 4927 if (rq->nr_running == 1) 4928 break; 4929 4930 next = pick_next_task(rq); 4931 BUG_ON(!next); 4932 next->sched_class->put_prev_task(rq, next); 4933 4934 /* Find suitable destination for @next, with force if needed. */ 4935 dest_cpu = select_fallback_rq(dead_cpu, next); 4936 raw_spin_unlock(&rq->lock); 4937 4938 __migrate_task(next, dead_cpu, dest_cpu); 4939 4940 raw_spin_lock(&rq->lock); 4941 } 4942 4943 rq->stop = stop; 4944} 4945 4946#endif /* CONFIG_HOTPLUG_CPU */ 4947 4948#if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL) 4949 4950static struct ctl_table sd_ctl_dir[] = { 4951 { 4952 .procname = "sched_domain", 4953 .mode = 0555, 4954 }, 4955 {} 4956}; 4957 4958static struct ctl_table sd_ctl_root[] = { 4959 { 4960 .procname = "kernel", 4961 .mode = 0555, 4962 .child = sd_ctl_dir, 4963 }, 4964 {} 4965}; 4966 4967static struct ctl_table *sd_alloc_ctl_entry(int n) 4968{ 4969 struct ctl_table *entry = 4970 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL); 4971 4972 return entry; 4973} 4974 4975static void sd_free_ctl_entry(struct ctl_table **tablep) 4976{ 4977 struct ctl_table *entry; 4978 4979 /* 4980 * In the intermediate directories, both the child directory and 4981 * procname are dynamically allocated and could fail but the mode 4982 * will always be set. In the lowest directory the names are 4983 * static strings and all have proc handlers. 4984 */ 4985 for (entry = *tablep; entry->mode; entry++) { 4986 if (entry->child) 4987 sd_free_ctl_entry(&entry->child); 4988 if (entry->proc_handler == NULL) 4989 kfree(entry->procname); 4990 } 4991 4992 kfree(*tablep); 4993 *tablep = NULL; 4994} 4995 4996static int min_load_idx = 0; 4997static int max_load_idx = CPU_LOAD_IDX_MAX; 4998 4999static void 5000set_table_entry(struct ctl_table *entry, 5001 const char *procname, void *data, int maxlen, 5002 umode_t mode, proc_handler *proc_handler, 5003 bool load_idx) 5004{ 5005 entry->procname = procname; 5006 entry->data = data; 5007 entry->maxlen = maxlen; 5008 entry->mode = mode; 5009 entry->proc_handler = proc_handler; 5010 5011 if (load_idx) { 5012 entry->extra1 = &min_load_idx; 5013 entry->extra2 = &max_load_idx; 5014 } 5015} 5016 5017static struct ctl_table * 5018sd_alloc_ctl_domain_table(struct sched_domain *sd) 5019{ 5020 struct ctl_table *table = sd_alloc_ctl_entry(13); 5021 5022 if (table == NULL) 5023 return NULL; 5024 5025 set_table_entry(&table[0], "min_interval", &sd->min_interval, 5026 sizeof(long), 0644, proc_doulongvec_minmax, false); 5027 set_table_entry(&table[1], "max_interval", &sd->max_interval, 5028 sizeof(long), 0644, proc_doulongvec_minmax, false); 5029 set_table_entry(&table[2], "busy_idx", &sd->busy_idx, 5030 sizeof(int), 0644, proc_dointvec_minmax, true); 5031 set_table_entry(&table[3], "idle_idx", &sd->idle_idx, 5032 sizeof(int), 0644, proc_dointvec_minmax, true); 5033 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx, 5034 sizeof(int), 0644, proc_dointvec_minmax, true); 5035 set_table_entry(&table[5], "wake_idx", &sd->wake_idx, 5036 sizeof(int), 0644, proc_dointvec_minmax, true); 5037 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx, 5038 sizeof(int), 0644, proc_dointvec_minmax, true); 5039 set_table_entry(&table[7], "busy_factor", &sd->busy_factor, 5040 sizeof(int), 0644, proc_dointvec_minmax, false); 5041 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct, 5042 sizeof(int), 0644, proc_dointvec_minmax, false); 5043 set_table_entry(&table[9], "cache_nice_tries", 5044 &sd->cache_nice_tries, 5045 sizeof(int), 0644, proc_dointvec_minmax, false); 5046 set_table_entry(&table[10], "flags", &sd->flags, 5047 sizeof(int), 0644, proc_dointvec_minmax, false); 5048 set_table_entry(&table[11], "name", sd->name, 5049 CORENAME_MAX_SIZE, 0444, proc_dostring, false); 5050 /* &table[12] is terminator */ 5051 5052 return table; 5053} 5054 5055static ctl_table *sd_alloc_ctl_cpu_table(int cpu) 5056{ 5057 struct ctl_table *entry, *table; 5058 struct sched_domain *sd; 5059 int domain_num = 0, i; 5060 char buf[32]; 5061 5062 for_each_domain(cpu, sd) 5063 domain_num++; 5064 entry = table = sd_alloc_ctl_entry(domain_num + 1); 5065 if (table == NULL) 5066 return NULL; 5067 5068 i = 0; 5069 for_each_domain(cpu, sd) { 5070 snprintf(buf, 32, "domain%d", i); 5071 entry->procname = kstrdup(buf, GFP_KERNEL); 5072 entry->mode = 0555; 5073 entry->child = sd_alloc_ctl_domain_table(sd); 5074 entry++; 5075 i++; 5076 } 5077 return table; 5078} 5079 5080static struct ctl_table_header *sd_sysctl_header; 5081static void register_sched_domain_sysctl(void) 5082{ 5083 int i, cpu_num = num_possible_cpus(); 5084 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1); 5085 char buf[32]; 5086 5087 WARN_ON(sd_ctl_dir[0].child); 5088 sd_ctl_dir[0].child = entry; 5089 5090 if (entry == NULL) 5091 return; 5092 5093 for_each_possible_cpu(i) { 5094 snprintf(buf, 32, "cpu%d", i); 5095 entry->procname = kstrdup(buf, GFP_KERNEL); 5096 entry->mode = 0555; 5097 entry->child = sd_alloc_ctl_cpu_table(i); 5098 entry++; 5099 } 5100 5101 WARN_ON(sd_sysctl_header); 5102 sd_sysctl_header = register_sysctl_table(sd_ctl_root); 5103} 5104 5105/* may be called multiple times per register */ 5106static void unregister_sched_domain_sysctl(void) 5107{ 5108 if (sd_sysctl_header) 5109 unregister_sysctl_table(sd_sysctl_header); 5110 sd_sysctl_header = NULL; 5111 if (sd_ctl_dir[0].child) 5112 sd_free_ctl_entry(&sd_ctl_dir[0].child); 5113} 5114#else 5115static void register_sched_domain_sysctl(void) 5116{ 5117} 5118static void unregister_sched_domain_sysctl(void) 5119{ 5120} 5121#endif 5122 5123static void set_rq_online(struct rq *rq) 5124{ 5125 if (!rq->online) { 5126 const struct sched_class *class; 5127 5128 cpumask_set_cpu(rq->cpu, rq->rd->online); 5129 rq->online = 1; 5130 5131 for_each_class(class) { 5132 if (class->rq_online) 5133 class->rq_online(rq); 5134 } 5135 } 5136} 5137 5138static void set_rq_offline(struct rq *rq) 5139{ 5140 if (rq->online) { 5141 const struct sched_class *class; 5142 5143 for_each_class(class) { 5144 if (class->rq_offline) 5145 class->rq_offline(rq); 5146 } 5147 5148 cpumask_clear_cpu(rq->cpu, rq->rd->online); 5149 rq->online = 0; 5150 } 5151} 5152 5153/* 5154 * migration_call - callback that gets triggered when a CPU is added. 5155 * Here we can start up the necessary migration thread for the new CPU. 5156 */ 5157static int __cpuinit 5158migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu) 5159{ 5160 int cpu = (long)hcpu; 5161 unsigned long flags; 5162 struct rq *rq = cpu_rq(cpu); 5163 5164 switch (action & ~CPU_TASKS_FROZEN) { 5165 5166 case CPU_UP_PREPARE: 5167 rq->calc_load_update = calc_load_update; 5168 break; 5169 5170 case CPU_ONLINE: 5171 /* Update our root-domain */ 5172 raw_spin_lock_irqsave(&rq->lock, flags); 5173 if (rq->rd) { 5174 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span)); 5175 5176 set_rq_online(rq); 5177 } 5178 raw_spin_unlock_irqrestore(&rq->lock, flags); 5179 break; 5180 5181#ifdef CONFIG_HOTPLUG_CPU 5182 case CPU_DYING: 5183 sched_ttwu_pending(); 5184 /* Update our root-domain */ 5185 raw_spin_lock_irqsave(&rq->lock, flags); 5186 if (rq->rd) { 5187 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span)); 5188 set_rq_offline(rq); 5189 } 5190 migrate_tasks(cpu); 5191 BUG_ON(rq->nr_running != 1); /* the migration thread */ 5192 raw_spin_unlock_irqrestore(&rq->lock, flags); 5193 break; 5194 5195 case CPU_DEAD: 5196 calc_load_migrate(rq); 5197 break; 5198#endif 5199 } 5200 5201 update_max_interval(); 5202 5203 return NOTIFY_OK; 5204} 5205 5206/* 5207 * Register at high priority so that task migration (migrate_all_tasks) 5208 * happens before everything else. This has to be lower priority than 5209 * the notifier in the perf_event subsystem, though. 5210 */ 5211static struct notifier_block __cpuinitdata migration_notifier = { 5212 .notifier_call = migration_call, 5213 .priority = CPU_PRI_MIGRATION, 5214}; 5215 5216static int __cpuinit sched_cpu_active(struct notifier_block *nfb, 5217 unsigned long action, void *hcpu) 5218{ 5219 switch (action & ~CPU_TASKS_FROZEN) { 5220 case CPU_STARTING: 5221 case CPU_DOWN_FAILED: 5222 set_cpu_active((long)hcpu, true); 5223 return NOTIFY_OK; 5224 default: 5225 return NOTIFY_DONE; 5226 } 5227} 5228 5229static int __cpuinit sched_cpu_inactive(struct notifier_block *nfb, 5230 unsigned long action, void *hcpu) 5231{ 5232 switch (action & ~CPU_TASKS_FROZEN) { 5233 case CPU_DOWN_PREPARE: 5234 set_cpu_active((long)hcpu, false); 5235 return NOTIFY_OK; 5236 default: 5237 return NOTIFY_DONE; 5238 } 5239} 5240 5241static int __init migration_init(void) 5242{ 5243 void *cpu = (void *)(long)smp_processor_id(); 5244 int err; 5245 5246 /* Initialize migration for the boot CPU */ 5247 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu); 5248 BUG_ON(err == NOTIFY_BAD); 5249 migration_call(&migration_notifier, CPU_ONLINE, cpu); 5250 register_cpu_notifier(&migration_notifier); 5251 5252 /* Register cpu active notifiers */ 5253 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE); 5254 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE); 5255 5256 return 0; 5257} 5258early_initcall(migration_init); 5259#endif 5260 5261#ifdef CONFIG_SMP 5262 5263static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */ 5264 5265#ifdef CONFIG_SCHED_DEBUG 5266 5267static __read_mostly int sched_debug_enabled; 5268 5269static int __init sched_debug_setup(char *str) 5270{ 5271 sched_debug_enabled = 1; 5272 5273 return 0; 5274} 5275early_param("sched_debug", sched_debug_setup); 5276 5277static inline bool sched_debug(void) 5278{ 5279 return sched_debug_enabled; 5280} 5281 5282static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level, 5283 struct cpumask *groupmask) 5284{ 5285 struct sched_group *group = sd->groups; 5286 char str[256]; 5287 5288 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd)); 5289 cpumask_clear(groupmask); 5290 5291 printk(KERN_DEBUG "%*s domain %d: ", level, "", level); 5292 5293 if (!(sd->flags & SD_LOAD_BALANCE)) { 5294 printk("does not load-balance\n"); 5295 if (sd->parent) 5296 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain" 5297 " has parent"); 5298 return -1; 5299 } 5300 5301 printk(KERN_CONT "span %s level %s\n", str, sd->name); 5302 5303 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) { 5304 printk(KERN_ERR "ERROR: domain->span does not contain " 5305 "CPU%d\n", cpu); 5306 } 5307 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) { 5308 printk(KERN_ERR "ERROR: domain->groups does not contain" 5309 " CPU%d\n", cpu); 5310 } 5311 5312 printk(KERN_DEBUG "%*s groups:", level + 1, ""); 5313 do { 5314 if (!group) { 5315 printk("\n"); 5316 printk(KERN_ERR "ERROR: group is NULL\n"); 5317 break; 5318 } 5319 5320 /* 5321 * Even though we initialize ->power to something semi-sane, 5322 * we leave power_orig unset. This allows us to detect if 5323 * domain iteration is still funny without causing /0 traps. 5324 */ 5325 if (!group->sgp->power_orig) { 5326 printk(KERN_CONT "\n"); 5327 printk(KERN_ERR "ERROR: domain->cpu_power not " 5328 "set\n"); 5329 break; 5330 } 5331 5332 if (!cpumask_weight(sched_group_cpus(group))) { 5333 printk(KERN_CONT "\n"); 5334 printk(KERN_ERR "ERROR: empty group\n"); 5335 break; 5336 } 5337 5338 if (!(sd->flags & SD_OVERLAP) && 5339 cpumask_intersects(groupmask, sched_group_cpus(group))) { 5340 printk(KERN_CONT "\n"); 5341 printk(KERN_ERR "ERROR: repeated CPUs\n"); 5342 break; 5343 } 5344 5345 cpumask_or(groupmask, groupmask, sched_group_cpus(group)); 5346 5347 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group)); 5348 5349 printk(KERN_CONT " %s", str); 5350 if (group->sgp->power != SCHED_POWER_SCALE) { 5351 printk(KERN_CONT " (cpu_power = %d)", 5352 group->sgp->power); 5353 } 5354 5355 group = group->next; 5356 } while (group != sd->groups); 5357 printk(KERN_CONT "\n"); 5358 5359 if (!cpumask_equal(sched_domain_span(sd), groupmask)) 5360 printk(KERN_ERR "ERROR: groups don't span domain->span\n"); 5361 5362 if (sd->parent && 5363 !cpumask_subset(groupmask, sched_domain_span(sd->parent))) 5364 printk(KERN_ERR "ERROR: parent span is not a superset " 5365 "of domain->span\n"); 5366 return 0; 5367} 5368 5369static void sched_domain_debug(struct sched_domain *sd, int cpu) 5370{ 5371 int level = 0; 5372 5373 if (!sched_debug_enabled) 5374 return; 5375 5376 if (!sd) { 5377 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu); 5378 return; 5379 } 5380 5381 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu); 5382 5383 for (;;) { 5384 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask)) 5385 break; 5386 level++; 5387 sd = sd->parent; 5388 if (!sd) 5389 break; 5390 } 5391} 5392#else /* !CONFIG_SCHED_DEBUG */ 5393# define sched_domain_debug(sd, cpu) do { } while (0) 5394static inline bool sched_debug(void) 5395{ 5396 return false; 5397} 5398#endif /* CONFIG_SCHED_DEBUG */ 5399 5400static int sd_degenerate(struct sched_domain *sd) 5401{ 5402 if (cpumask_weight(sched_domain_span(sd)) == 1) 5403 return 1; 5404 5405 /* Following flags need at least 2 groups */ 5406 if (sd->flags & (SD_LOAD_BALANCE | 5407 SD_BALANCE_NEWIDLE | 5408 SD_BALANCE_FORK | 5409 SD_BALANCE_EXEC | 5410 SD_SHARE_CPUPOWER | 5411 SD_SHARE_PKG_RESOURCES)) { 5412 if (sd->groups != sd->groups->next) 5413 return 0; 5414 } 5415 5416 /* Following flags don't use groups */ 5417 if (sd->flags & (SD_WAKE_AFFINE)) 5418 return 0; 5419 5420 return 1; 5421} 5422 5423static int 5424sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent) 5425{ 5426 unsigned long cflags = sd->flags, pflags = parent->flags; 5427 5428 if (sd_degenerate(parent)) 5429 return 1; 5430 5431 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent))) 5432 return 0; 5433 5434 /* Flags needing groups don't count if only 1 group in parent */ 5435 if (parent->groups == parent->groups->next) { 5436 pflags &= ~(SD_LOAD_BALANCE | 5437 SD_BALANCE_NEWIDLE | 5438 SD_BALANCE_FORK | 5439 SD_BALANCE_EXEC | 5440 SD_SHARE_CPUPOWER | 5441 SD_SHARE_PKG_RESOURCES); 5442 if (nr_node_ids == 1) 5443 pflags &= ~SD_SERIALIZE; 5444 } 5445 if (~cflags & pflags) 5446 return 0; 5447 5448 return 1; 5449} 5450 5451static void free_rootdomain(struct rcu_head *rcu) 5452{ 5453 struct root_domain *rd = container_of(rcu, struct root_domain, rcu); 5454 5455 cpupri_cleanup(&rd->cpupri); 5456 free_cpumask_var(rd->rto_mask); 5457 free_cpumask_var(rd->online); 5458 free_cpumask_var(rd->span); 5459 kfree(rd); 5460} 5461 5462static void rq_attach_root(struct rq *rq, struct root_domain *rd) 5463{ 5464 struct root_domain *old_rd = NULL; 5465 unsigned long flags; 5466 5467 raw_spin_lock_irqsave(&rq->lock, flags); 5468 5469 if (rq->rd) { 5470 old_rd = rq->rd; 5471 5472 if (cpumask_test_cpu(rq->cpu, old_rd->online)) 5473 set_rq_offline(rq); 5474 5475 cpumask_clear_cpu(rq->cpu, old_rd->span); 5476 5477 /* 5478 * If we dont want to free the old_rt yet then 5479 * set old_rd to NULL to skip the freeing later 5480 * in this function: 5481 */ 5482 if (!atomic_dec_and_test(&old_rd->refcount)) 5483 old_rd = NULL; 5484 } 5485 5486 atomic_inc(&rd->refcount); 5487 rq->rd = rd; 5488 5489 cpumask_set_cpu(rq->cpu, rd->span); 5490 if (cpumask_test_cpu(rq->cpu, cpu_active_mask)) 5491 set_rq_online(rq); 5492 5493 raw_spin_unlock_irqrestore(&rq->lock, flags); 5494 5495 if (old_rd) 5496 call_rcu_sched(&old_rd->rcu, free_rootdomain); 5497} 5498 5499static int init_rootdomain(struct root_domain *rd) 5500{ 5501 memset(rd, 0, sizeof(*rd)); 5502 5503 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL)) 5504 goto out; 5505 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL)) 5506 goto free_span; 5507 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL)) 5508 goto free_online; 5509 5510 if (cpupri_init(&rd->cpupri) != 0) 5511 goto free_rto_mask; 5512 return 0; 5513 5514free_rto_mask: 5515 free_cpumask_var(rd->rto_mask); 5516free_online: 5517 free_cpumask_var(rd->online); 5518free_span: 5519 free_cpumask_var(rd->span); 5520out: 5521 return -ENOMEM; 5522} 5523 5524/* 5525 * By default the system creates a single root-domain with all cpus as 5526 * members (mimicking the global state we have today). 5527 */ 5528struct root_domain def_root_domain; 5529 5530static void init_defrootdomain(void) 5531{ 5532 init_rootdomain(&def_root_domain); 5533 5534 atomic_set(&def_root_domain.refcount, 1); 5535} 5536 5537static struct root_domain *alloc_rootdomain(void) 5538{ 5539 struct root_domain *rd; 5540 5541 rd = kmalloc(sizeof(*rd), GFP_KERNEL); 5542 if (!rd) 5543 return NULL; 5544 5545 if (init_rootdomain(rd) != 0) { 5546 kfree(rd); 5547 return NULL; 5548 } 5549 5550 return rd; 5551} 5552 5553static void free_sched_groups(struct sched_group *sg, int free_sgp) 5554{ 5555 struct sched_group *tmp, *first; 5556 5557 if (!sg) 5558 return; 5559 5560 first = sg; 5561 do { 5562 tmp = sg->next; 5563 5564 if (free_sgp && atomic_dec_and_test(&sg->sgp->ref)) 5565 kfree(sg->sgp); 5566 5567 kfree(sg); 5568 sg = tmp; 5569 } while (sg != first); 5570} 5571 5572static void free_sched_domain(struct rcu_head *rcu) 5573{ 5574 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu); 5575 5576 /* 5577 * If its an overlapping domain it has private groups, iterate and 5578 * nuke them all. 5579 */ 5580 if (sd->flags & SD_OVERLAP) { 5581 free_sched_groups(sd->groups, 1); 5582 } else if (atomic_dec_and_test(&sd->groups->ref)) { 5583 kfree(sd->groups->sgp); 5584 kfree(sd->groups); 5585 } 5586 kfree(sd); 5587} 5588 5589static void destroy_sched_domain(struct sched_domain *sd, int cpu) 5590{ 5591 call_rcu(&sd->rcu, free_sched_domain); 5592} 5593 5594static void destroy_sched_domains(struct sched_domain *sd, int cpu) 5595{ 5596 for (; sd; sd = sd->parent) 5597 destroy_sched_domain(sd, cpu); 5598} 5599 5600/* 5601 * Keep a special pointer to the highest sched_domain that has 5602 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this 5603 * allows us to avoid some pointer chasing select_idle_sibling(). 5604 * 5605 * Also keep a unique ID per domain (we use the first cpu number in 5606 * the cpumask of the domain), this allows us to quickly tell if 5607 * two cpus are in the same cache domain, see cpus_share_cache(). 5608 */ 5609DEFINE_PER_CPU(struct sched_domain *, sd_llc); 5610DEFINE_PER_CPU(int, sd_llc_id); 5611 5612static void update_top_cache_domain(int cpu) 5613{ 5614 struct sched_domain *sd; 5615 int id = cpu; 5616 5617 sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES); 5618 if (sd) 5619 id = cpumask_first(sched_domain_span(sd)); 5620 5621 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd); 5622 per_cpu(sd_llc_id, cpu) = id; 5623} 5624 5625/* 5626 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must 5627 * hold the hotplug lock. 5628 */ 5629static void 5630cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu) 5631{ 5632 struct rq *rq = cpu_rq(cpu); 5633 struct sched_domain *tmp; 5634 5635 /* Remove the sched domains which do not contribute to scheduling. */ 5636 for (tmp = sd; tmp; ) { 5637 struct sched_domain *parent = tmp->parent; 5638 if (!parent) 5639 break; 5640 5641 if (sd_parent_degenerate(tmp, parent)) { 5642 tmp->parent = parent->parent; 5643 if (parent->parent) 5644 parent->parent->child = tmp; 5645 destroy_sched_domain(parent, cpu); 5646 } else 5647 tmp = tmp->parent; 5648 } 5649 5650 if (sd && sd_degenerate(sd)) { 5651 tmp = sd; 5652 sd = sd->parent; 5653 destroy_sched_domain(tmp, cpu); 5654 if (sd) 5655 sd->child = NULL; 5656 } 5657 5658 sched_domain_debug(sd, cpu); 5659 5660 rq_attach_root(rq, rd); 5661 tmp = rq->sd; 5662 rcu_assign_pointer(rq->sd, sd); 5663 destroy_sched_domains(tmp, cpu); 5664 5665 update_top_cache_domain(cpu); 5666} 5667 5668/* cpus with isolated domains */ 5669static cpumask_var_t cpu_isolated_map; 5670 5671/* Setup the mask of cpus configured for isolated domains */ 5672static int __init isolated_cpu_setup(char *str) 5673{ 5674 alloc_bootmem_cpumask_var(&cpu_isolated_map); 5675 cpulist_parse(str, cpu_isolated_map); 5676 return 1; 5677} 5678 5679__setup("isolcpus=", isolated_cpu_setup); 5680 5681static const struct cpumask *cpu_cpu_mask(int cpu) 5682{ 5683 return cpumask_of_node(cpu_to_node(cpu)); 5684} 5685 5686struct sd_data { 5687 struct sched_domain **__percpu sd; 5688 struct sched_group **__percpu sg; 5689 struct sched_group_power **__percpu sgp; 5690}; 5691 5692struct s_data { 5693 struct sched_domain ** __percpu sd; 5694 struct root_domain *rd; 5695}; 5696 5697enum s_alloc { 5698 sa_rootdomain, 5699 sa_sd, 5700 sa_sd_storage, 5701 sa_none, 5702}; 5703 5704struct sched_domain_topology_level; 5705 5706typedef struct sched_domain *(*sched_domain_init_f)(struct sched_domain_topology_level *tl, int cpu); 5707typedef const struct cpumask *(*sched_domain_mask_f)(int cpu); 5708 5709#define SDTL_OVERLAP 0x01 5710 5711struct sched_domain_topology_level { 5712 sched_domain_init_f init; 5713 sched_domain_mask_f mask; 5714 int flags; 5715 int numa_level; 5716 struct sd_data data; 5717}; 5718 5719/* 5720 * Build an iteration mask that can exclude certain CPUs from the upwards 5721 * domain traversal. 5722 * 5723 * Asymmetric node setups can result in situations where the domain tree is of 5724 * unequal depth, make sure to skip domains that already cover the entire 5725 * range. 5726 * 5727 * In that case build_sched_domains() will have terminated the iteration early 5728 * and our sibling sd spans will be empty. Domains should always include the 5729 * cpu they're built on, so check that. 5730 * 5731 */ 5732static void build_group_mask(struct sched_domain *sd, struct sched_group *sg) 5733{ 5734 const struct cpumask *span = sched_domain_span(sd); 5735 struct sd_data *sdd = sd->private; 5736 struct sched_domain *sibling; 5737 int i; 5738 5739 for_each_cpu(i, span) { 5740 sibling = *per_cpu_ptr(sdd->sd, i); 5741 if (!cpumask_test_cpu(i, sched_domain_span(sibling))) 5742 continue; 5743 5744 cpumask_set_cpu(i, sched_group_mask(sg)); 5745 } 5746} 5747 5748/* 5749 * Return the canonical balance cpu for this group, this is the first cpu 5750 * of this group that's also in the iteration mask. 5751 */ 5752int group_balance_cpu(struct sched_group *sg) 5753{ 5754 return cpumask_first_and(sched_group_cpus(sg), sched_group_mask(sg)); 5755} 5756 5757static int 5758build_overlap_sched_groups(struct sched_domain *sd, int cpu) 5759{ 5760 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg; 5761 const struct cpumask *span = sched_domain_span(sd); 5762 struct cpumask *covered = sched_domains_tmpmask; 5763 struct sd_data *sdd = sd->private; 5764 struct sched_domain *child; 5765 int i; 5766 5767 cpumask_clear(covered); 5768 5769 for_each_cpu(i, span) { 5770 struct cpumask *sg_span; 5771 5772 if (cpumask_test_cpu(i, covered)) 5773 continue; 5774 5775 child = *per_cpu_ptr(sdd->sd, i); 5776 5777 /* See the comment near build_group_mask(). */ 5778 if (!cpumask_test_cpu(i, sched_domain_span(child))) 5779 continue; 5780 5781 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(), 5782 GFP_KERNEL, cpu_to_node(cpu)); 5783 5784 if (!sg) 5785 goto fail; 5786 5787 sg_span = sched_group_cpus(sg); 5788 if (child->child) { 5789 child = child->child; 5790 cpumask_copy(sg_span, sched_domain_span(child)); 5791 } else 5792 cpumask_set_cpu(i, sg_span); 5793 5794 cpumask_or(covered, covered, sg_span); 5795 5796 sg->sgp = *per_cpu_ptr(sdd->sgp, i); 5797 if (atomic_inc_return(&sg->sgp->ref) == 1) 5798 build_group_mask(sd, sg); 5799 5800 /* 5801 * Initialize sgp->power such that even if we mess up the 5802 * domains and no possible iteration will get us here, we won't 5803 * die on a /0 trap. 5804 */ 5805 sg->sgp->power = SCHED_POWER_SCALE * cpumask_weight(sg_span); 5806 5807 /* 5808 * Make sure the first group of this domain contains the 5809 * canonical balance cpu. Otherwise the sched_domain iteration 5810 * breaks. See update_sg_lb_stats(). 5811 */ 5812 if ((!groups && cpumask_test_cpu(cpu, sg_span)) || 5813 group_balance_cpu(sg) == cpu) 5814 groups = sg; 5815 5816 if (!first) 5817 first = sg; 5818 if (last) 5819 last->next = sg; 5820 last = sg; 5821 last->next = first; 5822 } 5823 sd->groups = groups; 5824 5825 return 0; 5826 5827fail: 5828 free_sched_groups(first, 0); 5829 5830 return -ENOMEM; 5831} 5832 5833static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg) 5834{ 5835 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu); 5836 struct sched_domain *child = sd->child; 5837 5838 if (child) 5839 cpu = cpumask_first(sched_domain_span(child)); 5840 5841 if (sg) { 5842 *sg = *per_cpu_ptr(sdd->sg, cpu); 5843 (*sg)->sgp = *per_cpu_ptr(sdd->sgp, cpu); 5844 atomic_set(&(*sg)->sgp->ref, 1); /* for claim_allocations */ 5845 } 5846 5847 return cpu; 5848} 5849 5850/* 5851 * build_sched_groups will build a circular linked list of the groups 5852 * covered by the given span, and will set each group's ->cpumask correctly, 5853 * and ->cpu_power to 0. 5854 * 5855 * Assumes the sched_domain tree is fully constructed 5856 */ 5857static int 5858build_sched_groups(struct sched_domain *sd, int cpu) 5859{ 5860 struct sched_group *first = NULL, *last = NULL; 5861 struct sd_data *sdd = sd->private; 5862 const struct cpumask *span = sched_domain_span(sd); 5863 struct cpumask *covered; 5864 int i; 5865 5866 get_group(cpu, sdd, &sd->groups); 5867 atomic_inc(&sd->groups->ref); 5868 5869 if (cpu != cpumask_first(sched_domain_span(sd))) 5870 return 0; 5871 5872 lockdep_assert_held(&sched_domains_mutex); 5873 covered = sched_domains_tmpmask; 5874 5875 cpumask_clear(covered); 5876 5877 for_each_cpu(i, span) { 5878 struct sched_group *sg; 5879 int group = get_group(i, sdd, &sg); 5880 int j; 5881 5882 if (cpumask_test_cpu(i, covered)) 5883 continue; 5884 5885 cpumask_clear(sched_group_cpus(sg)); 5886 sg->sgp->power = 0; 5887 cpumask_setall(sched_group_mask(sg)); 5888 5889 for_each_cpu(j, span) { 5890 if (get_group(j, sdd, NULL) != group) 5891 continue; 5892 5893 cpumask_set_cpu(j, covered); 5894 cpumask_set_cpu(j, sched_group_cpus(sg)); 5895 } 5896 5897 if (!first) 5898 first = sg; 5899 if (last) 5900 last->next = sg; 5901 last = sg; 5902 } 5903 last->next = first; 5904 5905 return 0; 5906} 5907 5908/* 5909 * Initialize sched groups cpu_power. 5910 * 5911 * cpu_power indicates the capacity of sched group, which is used while 5912 * distributing the load between different sched groups in a sched domain. 5913 * Typically cpu_power for all the groups in a sched domain will be same unless 5914 * there are asymmetries in the topology. If there are asymmetries, group 5915 * having more cpu_power will pickup more load compared to the group having 5916 * less cpu_power. 5917 */ 5918static void init_sched_groups_power(int cpu, struct sched_domain *sd) 5919{ 5920 struct sched_group *sg = sd->groups; 5921 5922 WARN_ON(!sd || !sg); 5923 5924 do { 5925 sg->group_weight = cpumask_weight(sched_group_cpus(sg)); 5926 sg = sg->next; 5927 } while (sg != sd->groups); 5928 5929 if (cpu != group_balance_cpu(sg)) 5930 return; 5931 5932 update_group_power(sd, cpu); 5933 atomic_set(&sg->sgp->nr_busy_cpus, sg->group_weight); 5934} 5935 5936int __weak arch_sd_sibling_asym_packing(void) 5937{ 5938 return 0*SD_ASYM_PACKING; 5939} 5940 5941/* 5942 * Initializers for schedule domains 5943 * Non-inlined to reduce accumulated stack pressure in build_sched_domains() 5944 */ 5945 5946#ifdef CONFIG_SCHED_DEBUG 5947# define SD_INIT_NAME(sd, type) sd->name = #type 5948#else 5949# define SD_INIT_NAME(sd, type) do { } while (0) 5950#endif 5951 5952#define SD_INIT_FUNC(type) \ 5953static noinline struct sched_domain * \ 5954sd_init_##type(struct sched_domain_topology_level *tl, int cpu) \ 5955{ \ 5956 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); \ 5957 *sd = SD_##type##_INIT; \ 5958 SD_INIT_NAME(sd, type); \ 5959 sd->private = &tl->data; \ 5960 return sd; \ 5961} 5962 5963SD_INIT_FUNC(CPU) 5964#ifdef CONFIG_SCHED_SMT 5965 SD_INIT_FUNC(SIBLING) 5966#endif 5967#ifdef CONFIG_SCHED_MC 5968 SD_INIT_FUNC(MC) 5969#endif 5970#ifdef CONFIG_SCHED_BOOK 5971 SD_INIT_FUNC(BOOK) 5972#endif 5973 5974static int default_relax_domain_level = -1; 5975int sched_domain_level_max; 5976 5977static int __init setup_relax_domain_level(char *str) 5978{ 5979 if (kstrtoint(str, 0, &default_relax_domain_level)) 5980 pr_warn("Unable to set relax_domain_level\n"); 5981 5982 return 1; 5983} 5984__setup("relax_domain_level=", setup_relax_domain_level); 5985 5986static void set_domain_attribute(struct sched_domain *sd, 5987 struct sched_domain_attr *attr) 5988{ 5989 int request; 5990 5991 if (!attr || attr->relax_domain_level < 0) { 5992 if (default_relax_domain_level < 0) 5993 return; 5994 else 5995 request = default_relax_domain_level; 5996 } else 5997 request = attr->relax_domain_level; 5998 if (request < sd->level) { 5999 /* turn off idle balance on this domain */ 6000 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE); 6001 } else { 6002 /* turn on idle balance on this domain */ 6003 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE); 6004 } 6005} 6006 6007static void __sdt_free(const struct cpumask *cpu_map); 6008static int __sdt_alloc(const struct cpumask *cpu_map); 6009 6010static void __free_domain_allocs(struct s_data *d, enum s_alloc what, 6011 const struct cpumask *cpu_map) 6012{ 6013 switch (what) { 6014 case sa_rootdomain: 6015 if (!atomic_read(&d->rd->refcount)) 6016 free_rootdomain(&d->rd->rcu); /* fall through */ 6017 case sa_sd: 6018 free_percpu(d->sd); /* fall through */ 6019 case sa_sd_storage: 6020 __sdt_free(cpu_map); /* fall through */ 6021 case sa_none: 6022 break; 6023 } 6024} 6025 6026static enum s_alloc __visit_domain_allocation_hell(struct s_data *d, 6027 const struct cpumask *cpu_map) 6028{ 6029 memset(d, 0, sizeof(*d)); 6030 6031 if (__sdt_alloc(cpu_map)) 6032 return sa_sd_storage; 6033 d->sd = alloc_percpu(struct sched_domain *); 6034 if (!d->sd) 6035 return sa_sd_storage; 6036 d->rd = alloc_rootdomain(); 6037 if (!d->rd) 6038 return sa_sd; 6039 return sa_rootdomain; 6040} 6041 6042/* 6043 * NULL the sd_data elements we've used to build the sched_domain and 6044 * sched_group structure so that the subsequent __free_domain_allocs() 6045 * will not free the data we're using. 6046 */ 6047static void claim_allocations(int cpu, struct sched_domain *sd) 6048{ 6049 struct sd_data *sdd = sd->private; 6050 6051 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd); 6052 *per_cpu_ptr(sdd->sd, cpu) = NULL; 6053 6054 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref)) 6055 *per_cpu_ptr(sdd->sg, cpu) = NULL; 6056 6057 if (atomic_read(&(*per_cpu_ptr(sdd->sgp, cpu))->ref)) 6058 *per_cpu_ptr(sdd->sgp, cpu) = NULL; 6059} 6060 6061#ifdef CONFIG_SCHED_SMT 6062static const struct cpumask *cpu_smt_mask(int cpu) 6063{ 6064 return topology_thread_cpumask(cpu); 6065} 6066#endif 6067 6068/* 6069 * Topology list, bottom-up. 6070 */ 6071static struct sched_domain_topology_level default_topology[] = { 6072#ifdef CONFIG_SCHED_SMT 6073 { sd_init_SIBLING, cpu_smt_mask, }, 6074#endif 6075#ifdef CONFIG_SCHED_MC 6076 { sd_init_MC, cpu_coregroup_mask, }, 6077#endif 6078#ifdef CONFIG_SCHED_BOOK 6079 { sd_init_BOOK, cpu_book_mask, }, 6080#endif 6081 { sd_init_CPU, cpu_cpu_mask, }, 6082 { NULL, }, 6083}; 6084 6085static struct sched_domain_topology_level *sched_domain_topology = default_topology; 6086 6087#ifdef CONFIG_NUMA 6088 6089static int sched_domains_numa_levels; 6090static int *sched_domains_numa_distance; 6091static struct cpumask ***sched_domains_numa_masks; 6092static int sched_domains_curr_level; 6093 6094static inline int sd_local_flags(int level) 6095{ 6096 if (sched_domains_numa_distance[level] > RECLAIM_DISTANCE) 6097 return 0; 6098 6099 return SD_BALANCE_EXEC | SD_BALANCE_FORK | SD_WAKE_AFFINE; 6100} 6101 6102static struct sched_domain * 6103sd_numa_init(struct sched_domain_topology_level *tl, int cpu) 6104{ 6105 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); 6106 int level = tl->numa_level; 6107 int sd_weight = cpumask_weight( 6108 sched_domains_numa_masks[level][cpu_to_node(cpu)]); 6109 6110 *sd = (struct sched_domain){ 6111 .min_interval = sd_weight, 6112 .max_interval = 2*sd_weight, 6113 .busy_factor = 32, 6114 .imbalance_pct = 125, 6115 .cache_nice_tries = 2, 6116 .busy_idx = 3, 6117 .idle_idx = 2, 6118 .newidle_idx = 0, 6119 .wake_idx = 0, 6120 .forkexec_idx = 0, 6121 6122 .flags = 1*SD_LOAD_BALANCE 6123 | 1*SD_BALANCE_NEWIDLE 6124 | 0*SD_BALANCE_EXEC 6125 | 0*SD_BALANCE_FORK 6126 | 0*SD_BALANCE_WAKE 6127 | 0*SD_WAKE_AFFINE 6128 | 0*SD_SHARE_CPUPOWER 6129 | 0*SD_SHARE_PKG_RESOURCES 6130 | 1*SD_SERIALIZE 6131 | 0*SD_PREFER_SIBLING 6132 | sd_local_flags(level) 6133 , 6134 .last_balance = jiffies, 6135 .balance_interval = sd_weight, 6136 }; 6137 SD_INIT_NAME(sd, NUMA); 6138 sd->private = &tl->data; 6139 6140 /* 6141 * Ugly hack to pass state to sd_numa_mask()... 6142 */ 6143 sched_domains_curr_level = tl->numa_level; 6144 6145 return sd; 6146} 6147 6148static const struct cpumask *sd_numa_mask(int cpu) 6149{ 6150 return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)]; 6151} 6152 6153static void sched_numa_warn(const char *str) 6154{ 6155 static int done = false; 6156 int i,j; 6157 6158 if (done) 6159 return; 6160 6161 done = true; 6162 6163 printk(KERN_WARNING "ERROR: %s\n\n", str); 6164 6165 for (i = 0; i < nr_node_ids; i++) { 6166 printk(KERN_WARNING " "); 6167 for (j = 0; j < nr_node_ids; j++) 6168 printk(KERN_CONT "%02d ", node_distance(i,j)); 6169 printk(KERN_CONT "\n"); 6170 } 6171 printk(KERN_WARNING "\n"); 6172} 6173 6174static bool find_numa_distance(int distance) 6175{ 6176 int i; 6177 6178 if (distance == node_distance(0, 0)) 6179 return true; 6180 6181 for (i = 0; i < sched_domains_numa_levels; i++) { 6182 if (sched_domains_numa_distance[i] == distance) 6183 return true; 6184 } 6185 6186 return false; 6187} 6188 6189static void sched_init_numa(void) 6190{ 6191 int next_distance, curr_distance = node_distance(0, 0); 6192 struct sched_domain_topology_level *tl; 6193 int level = 0; 6194 int i, j, k; 6195 6196 sched_domains_numa_distance = kzalloc(sizeof(int) * nr_node_ids, GFP_KERNEL); 6197 if (!sched_domains_numa_distance) 6198 return; 6199 6200 /* 6201 * O(nr_nodes^2) deduplicating selection sort -- in order to find the 6202 * unique distances in the node_distance() table. 6203 * 6204 * Assumes node_distance(0,j) includes all distances in 6205 * node_distance(i,j) in order to avoid cubic time. 6206 */ 6207 next_distance = curr_distance; 6208 for (i = 0; i < nr_node_ids; i++) { 6209 for (j = 0; j < nr_node_ids; j++) { 6210 for (k = 0; k < nr_node_ids; k++) { 6211 int distance = node_distance(i, k); 6212 6213 if (distance > curr_distance && 6214 (distance < next_distance || 6215 next_distance == curr_distance)) 6216 next_distance = distance; 6217 6218 /* 6219 * While not a strong assumption it would be nice to know 6220 * about cases where if node A is connected to B, B is not 6221 * equally connected to A. 6222 */ 6223 if (sched_debug() && node_distance(k, i) != distance) 6224 sched_numa_warn("Node-distance not symmetric"); 6225 6226 if (sched_debug() && i && !find_numa_distance(distance)) 6227 sched_numa_warn("Node-0 not representative"); 6228 } 6229 if (next_distance != curr_distance) { 6230 sched_domains_numa_distance[level++] = next_distance; 6231 sched_domains_numa_levels = level; 6232 curr_distance = next_distance; 6233 } else break; 6234 } 6235 6236 /* 6237 * In case of sched_debug() we verify the above assumption. 6238 */ 6239 if (!sched_debug()) 6240 break; 6241 } 6242 /* 6243 * 'level' contains the number of unique distances, excluding the 6244 * identity distance node_distance(i,i). 6245 * 6246 * The sched_domains_nume_distance[] array includes the actual distance 6247 * numbers. 6248 */ 6249 6250 /* 6251 * Here, we should temporarily reset sched_domains_numa_levels to 0. 6252 * If it fails to allocate memory for array sched_domains_numa_masks[][], 6253 * the array will contain less then 'level' members. This could be 6254 * dangerous when we use it to iterate array sched_domains_numa_masks[][] 6255 * in other functions. 6256 * 6257 * We reset it to 'level' at the end of this function. 6258 */ 6259 sched_domains_numa_levels = 0; 6260 6261 sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL); 6262 if (!sched_domains_numa_masks) 6263 return; 6264 6265 /* 6266 * Now for each level, construct a mask per node which contains all 6267 * cpus of nodes that are that many hops away from us. 6268 */ 6269 for (i = 0; i < level; i++) { 6270 sched_domains_numa_masks[i] = 6271 kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL); 6272 if (!sched_domains_numa_masks[i]) 6273 return; 6274 6275 for (j = 0; j < nr_node_ids; j++) { 6276 struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL); 6277 if (!mask) 6278 return; 6279 6280 sched_domains_numa_masks[i][j] = mask; 6281 6282 for (k = 0; k < nr_node_ids; k++) { 6283 if (node_distance(j, k) > sched_domains_numa_distance[i]) 6284 continue; 6285 6286 cpumask_or(mask, mask, cpumask_of_node(k)); 6287 } 6288 } 6289 } 6290 6291 tl = kzalloc((ARRAY_SIZE(default_topology) + level) * 6292 sizeof(struct sched_domain_topology_level), GFP_KERNEL); 6293 if (!tl) 6294 return; 6295 6296 /* 6297 * Copy the default topology bits.. 6298 */ 6299 for (i = 0; default_topology[i].init; i++) 6300 tl[i] = default_topology[i]; 6301 6302 /* 6303 * .. and append 'j' levels of NUMA goodness. 6304 */ 6305 for (j = 0; j < level; i++, j++) { 6306 tl[i] = (struct sched_domain_topology_level){ 6307 .init = sd_numa_init, 6308 .mask = sd_numa_mask, 6309 .flags = SDTL_OVERLAP, 6310 .numa_level = j, 6311 }; 6312 } 6313 6314 sched_domain_topology = tl; 6315 6316 sched_domains_numa_levels = level; 6317} 6318 6319static void sched_domains_numa_masks_set(int cpu) 6320{ 6321 int i, j; 6322 int node = cpu_to_node(cpu); 6323 6324 for (i = 0; i < sched_domains_numa_levels; i++) { 6325 for (j = 0; j < nr_node_ids; j++) { 6326 if (node_distance(j, node) <= sched_domains_numa_distance[i]) 6327 cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]); 6328 } 6329 } 6330} 6331 6332static void sched_domains_numa_masks_clear(int cpu) 6333{ 6334 int i, j; 6335 for (i = 0; i < sched_domains_numa_levels; i++) { 6336 for (j = 0; j < nr_node_ids; j++) 6337 cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]); 6338 } 6339} 6340 6341/* 6342 * Update sched_domains_numa_masks[level][node] array when new cpus 6343 * are onlined. 6344 */ 6345static int sched_domains_numa_masks_update(struct notifier_block *nfb, 6346 unsigned long action, 6347 void *hcpu) 6348{ 6349 int cpu = (long)hcpu; 6350 6351 switch (action & ~CPU_TASKS_FROZEN) { 6352 case CPU_ONLINE: 6353 sched_domains_numa_masks_set(cpu); 6354 break; 6355 6356 case CPU_DEAD: 6357 sched_domains_numa_masks_clear(cpu); 6358 break; 6359 6360 default: 6361 return NOTIFY_DONE; 6362 } 6363 6364 return NOTIFY_OK; 6365} 6366#else 6367static inline void sched_init_numa(void) 6368{ 6369} 6370 6371static int sched_domains_numa_masks_update(struct notifier_block *nfb, 6372 unsigned long action, 6373 void *hcpu) 6374{ 6375 return 0; 6376} 6377#endif /* CONFIG_NUMA */ 6378 6379static int __sdt_alloc(const struct cpumask *cpu_map) 6380{ 6381 struct sched_domain_topology_level *tl; 6382 int j; 6383 6384 for (tl = sched_domain_topology; tl->init; tl++) { 6385 struct sd_data *sdd = &tl->data; 6386 6387 sdd->sd = alloc_percpu(struct sched_domain *); 6388 if (!sdd->sd) 6389 return -ENOMEM; 6390 6391 sdd->sg = alloc_percpu(struct sched_group *); 6392 if (!sdd->sg) 6393 return -ENOMEM; 6394 6395 sdd->sgp = alloc_percpu(struct sched_group_power *); 6396 if (!sdd->sgp) 6397 return -ENOMEM; 6398 6399 for_each_cpu(j, cpu_map) { 6400 struct sched_domain *sd; 6401 struct sched_group *sg; 6402 struct sched_group_power *sgp; 6403 6404 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(), 6405 GFP_KERNEL, cpu_to_node(j)); 6406 if (!sd) 6407 return -ENOMEM; 6408 6409 *per_cpu_ptr(sdd->sd, j) = sd; 6410 6411 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(), 6412 GFP_KERNEL, cpu_to_node(j)); 6413 if (!sg) 6414 return -ENOMEM; 6415 6416 sg->next = sg; 6417 6418 *per_cpu_ptr(sdd->sg, j) = sg; 6419 6420 sgp = kzalloc_node(sizeof(struct sched_group_power) + cpumask_size(), 6421 GFP_KERNEL, cpu_to_node(j)); 6422 if (!sgp) 6423 return -ENOMEM; 6424 6425 *per_cpu_ptr(sdd->sgp, j) = sgp; 6426 } 6427 } 6428 6429 return 0; 6430} 6431 6432static void __sdt_free(const struct cpumask *cpu_map) 6433{ 6434 struct sched_domain_topology_level *tl; 6435 int j; 6436 6437 for (tl = sched_domain_topology; tl->init; tl++) { 6438 struct sd_data *sdd = &tl->data; 6439 6440 for_each_cpu(j, cpu_map) { 6441 struct sched_domain *sd; 6442 6443 if (sdd->sd) { 6444 sd = *per_cpu_ptr(sdd->sd, j); 6445 if (sd && (sd->flags & SD_OVERLAP)) 6446 free_sched_groups(sd->groups, 0); 6447 kfree(*per_cpu_ptr(sdd->sd, j)); 6448 } 6449 6450 if (sdd->sg) 6451 kfree(*per_cpu_ptr(sdd->sg, j)); 6452 if (sdd->sgp) 6453 kfree(*per_cpu_ptr(sdd->sgp, j)); 6454 } 6455 free_percpu(sdd->sd); 6456 sdd->sd = NULL; 6457 free_percpu(sdd->sg); 6458 sdd->sg = NULL; 6459 free_percpu(sdd->sgp); 6460 sdd->sgp = NULL; 6461 } 6462} 6463 6464struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl, 6465 struct s_data *d, const struct cpumask *cpu_map, 6466 struct sched_domain_attr *attr, struct sched_domain *child, 6467 int cpu) 6468{ 6469 struct sched_domain *sd = tl->init(tl, cpu); 6470 if (!sd) 6471 return child; 6472 6473 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu)); 6474 if (child) { 6475 sd->level = child->level + 1; 6476 sched_domain_level_max = max(sched_domain_level_max, sd->level); 6477 child->parent = sd; 6478 } 6479 sd->child = child; 6480 set_domain_attribute(sd, attr); 6481 6482 return sd; 6483} 6484 6485/* 6486 * Build sched domains for a given set of cpus and attach the sched domains 6487 * to the individual cpus 6488 */ 6489static int build_sched_domains(const struct cpumask *cpu_map, 6490 struct sched_domain_attr *attr) 6491{ 6492 enum s_alloc alloc_state = sa_none; 6493 struct sched_domain *sd; 6494 struct s_data d; 6495 int i, ret = -ENOMEM; 6496 6497 alloc_state = __visit_domain_allocation_hell(&d, cpu_map); 6498 if (alloc_state != sa_rootdomain) 6499 goto error; 6500 6501 /* Set up domains for cpus specified by the cpu_map. */ 6502 for_each_cpu(i, cpu_map) { 6503 struct sched_domain_topology_level *tl; 6504 6505 sd = NULL; 6506 for (tl = sched_domain_topology; tl->init; tl++) { 6507 sd = build_sched_domain(tl, &d, cpu_map, attr, sd, i); 6508 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP)) 6509 sd->flags |= SD_OVERLAP; 6510 if (cpumask_equal(cpu_map, sched_domain_span(sd))) 6511 break; 6512 } 6513 6514 while (sd->child) 6515 sd = sd->child; 6516 6517 *per_cpu_ptr(d.sd, i) = sd; 6518 } 6519 6520 /* Build the groups for the domains */ 6521 for_each_cpu(i, cpu_map) { 6522 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) { 6523 sd->span_weight = cpumask_weight(sched_domain_span(sd)); 6524 if (sd->flags & SD_OVERLAP) { 6525 if (build_overlap_sched_groups(sd, i)) 6526 goto error; 6527 } else { 6528 if (build_sched_groups(sd, i)) 6529 goto error; 6530 } 6531 } 6532 } 6533 6534 /* Calculate CPU power for physical packages and nodes */ 6535 for (i = nr_cpumask_bits-1; i >= 0; i--) { 6536 if (!cpumask_test_cpu(i, cpu_map)) 6537 continue; 6538 6539 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) { 6540 claim_allocations(i, sd); 6541 init_sched_groups_power(i, sd); 6542 } 6543 } 6544 6545 /* Attach the domains */ 6546 rcu_read_lock(); 6547 for_each_cpu(i, cpu_map) { 6548 sd = *per_cpu_ptr(d.sd, i); 6549 cpu_attach_domain(sd, d.rd, i); 6550 } 6551 rcu_read_unlock(); 6552 6553 ret = 0; 6554error: 6555 __free_domain_allocs(&d, alloc_state, cpu_map); 6556 return ret; 6557} 6558 6559static cpumask_var_t *doms_cur; /* current sched domains */ 6560static int ndoms_cur; /* number of sched domains in 'doms_cur' */ 6561static struct sched_domain_attr *dattr_cur; 6562 /* attribues of custom domains in 'doms_cur' */ 6563 6564/* 6565 * Special case: If a kmalloc of a doms_cur partition (array of 6566 * cpumask) fails, then fallback to a single sched domain, 6567 * as determined by the single cpumask fallback_doms. 6568 */ 6569static cpumask_var_t fallback_doms; 6570 6571/* 6572 * arch_update_cpu_topology lets virtualized architectures update the 6573 * cpu core maps. It is supposed to return 1 if the topology changed 6574 * or 0 if it stayed the same. 6575 */ 6576int __attribute__((weak)) arch_update_cpu_topology(void) 6577{ 6578 return 0; 6579} 6580 6581cpumask_var_t *alloc_sched_domains(unsigned int ndoms) 6582{ 6583 int i; 6584 cpumask_var_t *doms; 6585 6586 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL); 6587 if (!doms) 6588 return NULL; 6589 for (i = 0; i < ndoms; i++) { 6590 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) { 6591 free_sched_domains(doms, i); 6592 return NULL; 6593 } 6594 } 6595 return doms; 6596} 6597 6598void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms) 6599{ 6600 unsigned int i; 6601 for (i = 0; i < ndoms; i++) 6602 free_cpumask_var(doms[i]); 6603 kfree(doms); 6604} 6605 6606/* 6607 * Set up scheduler domains and groups. Callers must hold the hotplug lock. 6608 * For now this just excludes isolated cpus, but could be used to 6609 * exclude other special cases in the future. 6610 */ 6611static int init_sched_domains(const struct cpumask *cpu_map) 6612{ 6613 int err; 6614 6615 arch_update_cpu_topology(); 6616 ndoms_cur = 1; 6617 doms_cur = alloc_sched_domains(ndoms_cur); 6618 if (!doms_cur) 6619 doms_cur = &fallback_doms; 6620 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map); 6621 err = build_sched_domains(doms_cur[0], NULL); 6622 register_sched_domain_sysctl(); 6623 6624 return err; 6625} 6626 6627/* 6628 * Detach sched domains from a group of cpus specified in cpu_map 6629 * These cpus will now be attached to the NULL domain 6630 */ 6631static void detach_destroy_domains(const struct cpumask *cpu_map) 6632{ 6633 int i; 6634 6635 rcu_read_lock(); 6636 for_each_cpu(i, cpu_map) 6637 cpu_attach_domain(NULL, &def_root_domain, i); 6638 rcu_read_unlock(); 6639} 6640 6641/* handle null as "default" */ 6642static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur, 6643 struct sched_domain_attr *new, int idx_new) 6644{ 6645 struct sched_domain_attr tmp; 6646 6647 /* fast path */ 6648 if (!new && !cur) 6649 return 1; 6650 6651 tmp = SD_ATTR_INIT; 6652 return !memcmp(cur ? (cur + idx_cur) : &tmp, 6653 new ? (new + idx_new) : &tmp, 6654 sizeof(struct sched_domain_attr)); 6655} 6656 6657/* 6658 * Partition sched domains as specified by the 'ndoms_new' 6659 * cpumasks in the array doms_new[] of cpumasks. This compares 6660 * doms_new[] to the current sched domain partitioning, doms_cur[]. 6661 * It destroys each deleted domain and builds each new domain. 6662 * 6663 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'. 6664 * The masks don't intersect (don't overlap.) We should setup one 6665 * sched domain for each mask. CPUs not in any of the cpumasks will 6666 * not be load balanced. If the same cpumask appears both in the 6667 * current 'doms_cur' domains and in the new 'doms_new', we can leave 6668 * it as it is. 6669 * 6670 * The passed in 'doms_new' should be allocated using 6671 * alloc_sched_domains. This routine takes ownership of it and will 6672 * free_sched_domains it when done with it. If the caller failed the 6673 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1, 6674 * and partition_sched_domains() will fallback to the single partition 6675 * 'fallback_doms', it also forces the domains to be rebuilt. 6676 * 6677 * If doms_new == NULL it will be replaced with cpu_online_mask. 6678 * ndoms_new == 0 is a special case for destroying existing domains, 6679 * and it will not create the default domain. 6680 * 6681 * Call with hotplug lock held 6682 */ 6683void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[], 6684 struct sched_domain_attr *dattr_new) 6685{ 6686 int i, j, n; 6687 int new_topology; 6688 6689 mutex_lock(&sched_domains_mutex); 6690 6691 /* always unregister in case we don't destroy any domains */ 6692 unregister_sched_domain_sysctl(); 6693 6694 /* Let architecture update cpu core mappings. */ 6695 new_topology = arch_update_cpu_topology(); 6696 6697 n = doms_new ? ndoms_new : 0; 6698 6699 /* Destroy deleted domains */ 6700 for (i = 0; i < ndoms_cur; i++) { 6701 for (j = 0; j < n && !new_topology; j++) { 6702 if (cpumask_equal(doms_cur[i], doms_new[j]) 6703 && dattrs_equal(dattr_cur, i, dattr_new, j)) 6704 goto match1; 6705 } 6706 /* no match - a current sched domain not in new doms_new[] */ 6707 detach_destroy_domains(doms_cur[i]); 6708match1: 6709 ; 6710 } 6711 6712 if (doms_new == NULL) { 6713 ndoms_cur = 0; 6714 doms_new = &fallback_doms; 6715 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map); 6716 WARN_ON_ONCE(dattr_new); 6717 } 6718 6719 /* Build new domains */ 6720 for (i = 0; i < ndoms_new; i++) { 6721 for (j = 0; j < ndoms_cur && !new_topology; j++) { 6722 if (cpumask_equal(doms_new[i], doms_cur[j]) 6723 && dattrs_equal(dattr_new, i, dattr_cur, j)) 6724 goto match2; 6725 } 6726 /* no match - add a new doms_new */ 6727 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL); 6728match2: 6729 ; 6730 } 6731 6732 /* Remember the new sched domains */ 6733 if (doms_cur != &fallback_doms) 6734 free_sched_domains(doms_cur, ndoms_cur); 6735 kfree(dattr_cur); /* kfree(NULL) is safe */ 6736 doms_cur = doms_new; 6737 dattr_cur = dattr_new; 6738 ndoms_cur = ndoms_new; 6739 6740 register_sched_domain_sysctl(); 6741 6742 mutex_unlock(&sched_domains_mutex); 6743} 6744 6745static int num_cpus_frozen; /* used to mark begin/end of suspend/resume */ 6746 6747/* 6748 * Update cpusets according to cpu_active mask. If cpusets are 6749 * disabled, cpuset_update_active_cpus() becomes a simple wrapper 6750 * around partition_sched_domains(). 6751 * 6752 * If we come here as part of a suspend/resume, don't touch cpusets because we 6753 * want to restore it back to its original state upon resume anyway. 6754 */ 6755static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action, 6756 void *hcpu) 6757{ 6758 switch (action) { 6759 case CPU_ONLINE_FROZEN: 6760 case CPU_DOWN_FAILED_FROZEN: 6761 6762 /* 6763 * num_cpus_frozen tracks how many CPUs are involved in suspend 6764 * resume sequence. As long as this is not the last online 6765 * operation in the resume sequence, just build a single sched 6766 * domain, ignoring cpusets. 6767 */ 6768 num_cpus_frozen--; 6769 if (likely(num_cpus_frozen)) { 6770 partition_sched_domains(1, NULL, NULL); 6771 break; 6772 } 6773 6774 /* 6775 * This is the last CPU online operation. So fall through and 6776 * restore the original sched domains by considering the 6777 * cpuset configurations. 6778 */ 6779 6780 case CPU_ONLINE: 6781 case CPU_DOWN_FAILED: 6782 cpuset_update_active_cpus(true); 6783 break; 6784 default: 6785 return NOTIFY_DONE; 6786 } 6787 return NOTIFY_OK; 6788} 6789 6790static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action, 6791 void *hcpu) 6792{ 6793 switch (action) { 6794 case CPU_DOWN_PREPARE: 6795 cpuset_update_active_cpus(false); 6796 break; 6797 case CPU_DOWN_PREPARE_FROZEN: 6798 num_cpus_frozen++; 6799 partition_sched_domains(1, NULL, NULL); 6800 break; 6801 default: 6802 return NOTIFY_DONE; 6803 } 6804 return NOTIFY_OK; 6805} 6806 6807void __init sched_init_smp(void) 6808{ 6809 cpumask_var_t non_isolated_cpus; 6810 6811 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL); 6812 alloc_cpumask_var(&fallback_doms, GFP_KERNEL); 6813 6814 sched_init_numa(); 6815 6816 get_online_cpus(); 6817 mutex_lock(&sched_domains_mutex); 6818 init_sched_domains(cpu_active_mask); 6819 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map); 6820 if (cpumask_empty(non_isolated_cpus)) 6821 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus); 6822 mutex_unlock(&sched_domains_mutex); 6823 put_online_cpus(); 6824 6825 hotcpu_notifier(sched_domains_numa_masks_update, CPU_PRI_SCHED_ACTIVE); 6826 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE); 6827 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE); 6828 6829 /* RT runtime code needs to handle some hotplug events */ 6830 hotcpu_notifier(update_runtime, 0); 6831 6832 init_hrtick(); 6833 6834 /* Move init over to a non-isolated CPU */ 6835 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0) 6836 BUG(); 6837 sched_init_granularity(); 6838 free_cpumask_var(non_isolated_cpus); 6839 6840 init_sched_rt_class(); 6841} 6842#else 6843void __init sched_init_smp(void) 6844{ 6845 sched_init_granularity(); 6846} 6847#endif /* CONFIG_SMP */ 6848 6849const_debug unsigned int sysctl_timer_migration = 1; 6850 6851int in_sched_functions(unsigned long addr) 6852{ 6853 return in_lock_functions(addr) || 6854 (addr >= (unsigned long)__sched_text_start 6855 && addr < (unsigned long)__sched_text_end); 6856} 6857 6858#ifdef CONFIG_CGROUP_SCHED 6859struct task_group root_task_group; 6860LIST_HEAD(task_groups); 6861#endif 6862 6863DECLARE_PER_CPU(cpumask_var_t, load_balance_tmpmask); 6864 6865void __init sched_init(void) 6866{ 6867 int i, j; 6868 unsigned long alloc_size = 0, ptr; 6869 6870#ifdef CONFIG_FAIR_GROUP_SCHED 6871 alloc_size += 2 * nr_cpu_ids * sizeof(void **); 6872#endif 6873#ifdef CONFIG_RT_GROUP_SCHED 6874 alloc_size += 2 * nr_cpu_ids * sizeof(void **); 6875#endif 6876#ifdef CONFIG_CPUMASK_OFFSTACK 6877 alloc_size += num_possible_cpus() * cpumask_size(); 6878#endif 6879 if (alloc_size) { 6880 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT); 6881 6882#ifdef CONFIG_FAIR_GROUP_SCHED 6883 root_task_group.se = (struct sched_entity **)ptr; 6884 ptr += nr_cpu_ids * sizeof(void **); 6885 6886 root_task_group.cfs_rq = (struct cfs_rq **)ptr; 6887 ptr += nr_cpu_ids * sizeof(void **); 6888 6889#endif /* CONFIG_FAIR_GROUP_SCHED */ 6890#ifdef CONFIG_RT_GROUP_SCHED 6891 root_task_group.rt_se = (struct sched_rt_entity **)ptr; 6892 ptr += nr_cpu_ids * sizeof(void **); 6893 6894 root_task_group.rt_rq = (struct rt_rq **)ptr; 6895 ptr += nr_cpu_ids * sizeof(void **); 6896 6897#endif /* CONFIG_RT_GROUP_SCHED */ 6898#ifdef CONFIG_CPUMASK_OFFSTACK 6899 for_each_possible_cpu(i) { 6900 per_cpu(load_balance_tmpmask, i) = (void *)ptr; 6901 ptr += cpumask_size(); 6902 } 6903#endif /* CONFIG_CPUMASK_OFFSTACK */ 6904 } 6905 6906#ifdef CONFIG_SMP 6907 init_defrootdomain(); 6908#endif 6909 6910 init_rt_bandwidth(&def_rt_bandwidth, 6911 global_rt_period(), global_rt_runtime()); 6912 6913#ifdef CONFIG_RT_GROUP_SCHED 6914 init_rt_bandwidth(&root_task_group.rt_bandwidth, 6915 global_rt_period(), global_rt_runtime()); 6916#endif /* CONFIG_RT_GROUP_SCHED */ 6917 6918#ifdef CONFIG_CGROUP_SCHED 6919 list_add(&root_task_group.list, &task_groups); 6920 INIT_LIST_HEAD(&root_task_group.children); 6921 INIT_LIST_HEAD(&root_task_group.siblings); 6922 autogroup_init(&init_task); 6923 6924#endif /* CONFIG_CGROUP_SCHED */ 6925 6926#ifdef CONFIG_CGROUP_CPUACCT 6927 root_cpuacct.cpustat = &kernel_cpustat; 6928 root_cpuacct.cpuusage = alloc_percpu(u64); 6929 /* Too early, not expected to fail */ 6930 BUG_ON(!root_cpuacct.cpuusage); 6931#endif 6932 for_each_possible_cpu(i) { 6933 struct rq *rq; 6934 6935 rq = cpu_rq(i); 6936 raw_spin_lock_init(&rq->lock); 6937 rq->nr_running = 0; 6938 rq->calc_load_active = 0; 6939 rq->calc_load_update = jiffies + LOAD_FREQ; 6940 init_cfs_rq(&rq->cfs); 6941 init_rt_rq(&rq->rt, rq); 6942#ifdef CONFIG_FAIR_GROUP_SCHED 6943 root_task_group.shares = ROOT_TASK_GROUP_LOAD; 6944 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list); 6945 /* 6946 * How much cpu bandwidth does root_task_group get? 6947 * 6948 * In case of task-groups formed thr' the cgroup filesystem, it 6949 * gets 100% of the cpu resources in the system. This overall 6950 * system cpu resource is divided among the tasks of 6951 * root_task_group and its child task-groups in a fair manner, 6952 * based on each entity's (task or task-group's) weight 6953 * (se->load.weight). 6954 * 6955 * In other words, if root_task_group has 10 tasks of weight 6956 * 1024) and two child groups A0 and A1 (of weight 1024 each), 6957 * then A0's share of the cpu resource is: 6958 * 6959 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33% 6960 * 6961 * We achieve this by letting root_task_group's tasks sit 6962 * directly in rq->cfs (i.e root_task_group->se[] = NULL). 6963 */ 6964 init_cfs_bandwidth(&root_task_group.cfs_bandwidth); 6965 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL); 6966#endif /* CONFIG_FAIR_GROUP_SCHED */ 6967 6968 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime; 6969#ifdef CONFIG_RT_GROUP_SCHED 6970 INIT_LIST_HEAD(&rq->leaf_rt_rq_list); 6971 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL); 6972#endif 6973 6974 for (j = 0; j < CPU_LOAD_IDX_MAX; j++) 6975 rq->cpu_load[j] = 0; 6976 6977 rq->last_load_update_tick = jiffies; 6978 6979#ifdef CONFIG_SMP 6980 rq->sd = NULL; 6981 rq->rd = NULL; 6982 rq->cpu_power = SCHED_POWER_SCALE; 6983 rq->post_schedule = 0; 6984 rq->active_balance = 0; 6985 rq->next_balance = jiffies; 6986 rq->push_cpu = 0; 6987 rq->cpu = i; 6988 rq->online = 0; 6989 rq->idle_stamp = 0; 6990 rq->avg_idle = 2*sysctl_sched_migration_cost; 6991 6992 INIT_LIST_HEAD(&rq->cfs_tasks); 6993 6994 rq_attach_root(rq, &def_root_domain); 6995#ifdef CONFIG_NO_HZ 6996 rq->nohz_flags = 0; 6997#endif 6998#endif 6999 init_rq_hrtick(rq); 7000 atomic_set(&rq->nr_iowait, 0); 7001 } 7002 7003 set_load_weight(&init_task); 7004 7005#ifdef CONFIG_PREEMPT_NOTIFIERS 7006 INIT_HLIST_HEAD(&init_task.preempt_notifiers); 7007#endif 7008 7009#ifdef CONFIG_RT_MUTEXES 7010 plist_head_init(&init_task.pi_waiters); 7011#endif 7012 7013 /* 7014 * The boot idle thread does lazy MMU switching as well: 7015 */ 7016 atomic_inc(&init_mm.mm_count); 7017 enter_lazy_tlb(&init_mm, current); 7018 7019 /* 7020 * Make us the idle thread. Technically, schedule() should not be 7021 * called from this thread, however somewhere below it might be, 7022 * but because we are the idle thread, we just pick up running again 7023 * when this runqueue becomes "idle". 7024 */ 7025 init_idle(current, smp_processor_id()); 7026 7027 calc_load_update = jiffies + LOAD_FREQ; 7028 7029 /* 7030 * During early bootup we pretend to be a normal task: 7031 */ 7032 current->sched_class = &fair_sched_class; 7033 7034#ifdef CONFIG_SMP 7035 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT); 7036 /* May be allocated at isolcpus cmdline parse time */ 7037 if (cpu_isolated_map == NULL) 7038 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT); 7039 idle_thread_set_boot_cpu(); 7040#endif 7041 init_sched_fair_class(); 7042 7043 scheduler_running = 1; 7044} 7045 7046#ifdef CONFIG_DEBUG_ATOMIC_SLEEP 7047static inline int preempt_count_equals(int preempt_offset) 7048{ 7049 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth(); 7050 7051 return (nested == preempt_offset); 7052} 7053 7054void __might_sleep(const char *file, int line, int preempt_offset) 7055{ 7056 static unsigned long prev_jiffy; /* ratelimiting */ 7057 7058 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */ 7059 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) || 7060 system_state != SYSTEM_RUNNING || oops_in_progress) 7061 return; 7062 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy) 7063 return; 7064 prev_jiffy = jiffies; 7065 7066 printk(KERN_ERR 7067 "BUG: sleeping function called from invalid context at %s:%d\n", 7068 file, line); 7069 printk(KERN_ERR 7070 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n", 7071 in_atomic(), irqs_disabled(), 7072 current->pid, current->comm); 7073 7074 debug_show_held_locks(current); 7075 if (irqs_disabled()) 7076 print_irqtrace_events(current); 7077 dump_stack(); 7078} 7079EXPORT_SYMBOL(__might_sleep); 7080#endif 7081 7082#ifdef CONFIG_MAGIC_SYSRQ 7083static void normalize_task(struct rq *rq, struct task_struct *p) 7084{ 7085 const struct sched_class *prev_class = p->sched_class; 7086 int old_prio = p->prio; 7087 int on_rq; 7088 7089 on_rq = p->on_rq; 7090 if (on_rq) 7091 dequeue_task(rq, p, 0); 7092 __setscheduler(rq, p, SCHED_NORMAL, 0); 7093 if (on_rq) { 7094 enqueue_task(rq, p, 0); 7095 resched_task(rq->curr); 7096 } 7097 7098 check_class_changed(rq, p, prev_class, old_prio); 7099} 7100 7101void normalize_rt_tasks(void) 7102{ 7103 struct task_struct *g, *p; 7104 unsigned long flags; 7105 struct rq *rq; 7106 7107 read_lock_irqsave(&tasklist_lock, flags); 7108 do_each_thread(g, p) { 7109 /* 7110 * Only normalize user tasks: 7111 */ 7112 if (!p->mm) 7113 continue; 7114 7115 p->se.exec_start = 0; 7116#ifdef CONFIG_SCHEDSTATS 7117 p->se.statistics.wait_start = 0; 7118 p->se.statistics.sleep_start = 0; 7119 p->se.statistics.block_start = 0; 7120#endif 7121 7122 if (!rt_task(p)) { 7123 /* 7124 * Renice negative nice level userspace 7125 * tasks back to 0: 7126 */ 7127 if (TASK_NICE(p) < 0 && p->mm) 7128 set_user_nice(p, 0); 7129 continue; 7130 } 7131 7132 raw_spin_lock(&p->pi_lock); 7133 rq = __task_rq_lock(p); 7134 7135 normalize_task(rq, p); 7136 7137 __task_rq_unlock(rq); 7138 raw_spin_unlock(&p->pi_lock); 7139 } while_each_thread(g, p); 7140 7141 read_unlock_irqrestore(&tasklist_lock, flags); 7142} 7143 7144#endif /* CONFIG_MAGIC_SYSRQ */ 7145 7146#if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) 7147/* 7148 * These functions are only useful for the IA64 MCA handling, or kdb. 7149 * 7150 * They can only be called when the whole system has been 7151 * stopped - every CPU needs to be quiescent, and no scheduling 7152 * activity can take place. Using them for anything else would 7153 * be a serious bug, and as a result, they aren't even visible 7154 * under any other configuration. 7155 */ 7156 7157/** 7158 * curr_task - return the current task for a given cpu. 7159 * @cpu: the processor in question. 7160 * 7161 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED! 7162 */ 7163struct task_struct *curr_task(int cpu) 7164{ 7165 return cpu_curr(cpu); 7166} 7167 7168#endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */ 7169 7170#ifdef CONFIG_IA64 7171/** 7172 * set_curr_task - set the current task for a given cpu. 7173 * @cpu: the processor in question. 7174 * @p: the task pointer to set. 7175 * 7176 * Description: This function must only be used when non-maskable interrupts 7177 * are serviced on a separate stack. It allows the architecture to switch the 7178 * notion of the current task on a cpu in a non-blocking manner. This function 7179 * must be called with all CPU's synchronized, and interrupts disabled, the 7180 * and caller must save the original value of the current task (see 7181 * curr_task() above) and restore that value before reenabling interrupts and 7182 * re-starting the system. 7183 * 7184 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED! 7185 */ 7186void set_curr_task(int cpu, struct task_struct *p) 7187{ 7188 cpu_curr(cpu) = p; 7189} 7190 7191#endif 7192 7193#ifdef CONFIG_CGROUP_SCHED 7194/* task_group_lock serializes the addition/removal of task groups */ 7195static DEFINE_SPINLOCK(task_group_lock); 7196 7197static void free_sched_group(struct task_group *tg) 7198{ 7199 free_fair_sched_group(tg); 7200 free_rt_sched_group(tg); 7201 autogroup_free(tg); 7202 kfree(tg); 7203} 7204 7205/* allocate runqueue etc for a new task group */ 7206struct task_group *sched_create_group(struct task_group *parent) 7207{ 7208 struct task_group *tg; 7209 unsigned long flags; 7210 7211 tg = kzalloc(sizeof(*tg), GFP_KERNEL); 7212 if (!tg) 7213 return ERR_PTR(-ENOMEM); 7214 7215 if (!alloc_fair_sched_group(tg, parent)) 7216 goto err; 7217 7218 if (!alloc_rt_sched_group(tg, parent)) 7219 goto err; 7220 7221 spin_lock_irqsave(&task_group_lock, flags); 7222 list_add_rcu(&tg->list, &task_groups); 7223 7224 WARN_ON(!parent); /* root should already exist */ 7225 7226 tg->parent = parent; 7227 INIT_LIST_HEAD(&tg->children); 7228 list_add_rcu(&tg->siblings, &parent->children); 7229 spin_unlock_irqrestore(&task_group_lock, flags); 7230 7231 return tg; 7232 7233err: 7234 free_sched_group(tg); 7235 return ERR_PTR(-ENOMEM); 7236} 7237 7238/* rcu callback to free various structures associated with a task group */ 7239static void free_sched_group_rcu(struct rcu_head *rhp) 7240{ 7241 /* now it should be safe to free those cfs_rqs */ 7242 free_sched_group(container_of(rhp, struct task_group, rcu)); 7243} 7244 7245/* Destroy runqueue etc associated with a task group */ 7246void sched_destroy_group(struct task_group *tg) 7247{ 7248 unsigned long flags; 7249 int i; 7250 7251 /* end participation in shares distribution */ 7252 for_each_possible_cpu(i) 7253 unregister_fair_sched_group(tg, i); 7254 7255 spin_lock_irqsave(&task_group_lock, flags); 7256 list_del_rcu(&tg->list); 7257 list_del_rcu(&tg->siblings); 7258 spin_unlock_irqrestore(&task_group_lock, flags); 7259 7260 /* wait for possible concurrent references to cfs_rqs complete */ 7261 call_rcu(&tg->rcu, free_sched_group_rcu); 7262} 7263 7264/* change task's runqueue when it moves between groups. 7265 * The caller of this function should have put the task in its new group 7266 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to 7267 * reflect its new group. 7268 */ 7269void sched_move_task(struct task_struct *tsk) 7270{ 7271 struct task_group *tg; 7272 int on_rq, running; 7273 unsigned long flags; 7274 struct rq *rq; 7275 7276 rq = task_rq_lock(tsk, &flags); 7277 7278 running = task_current(rq, tsk); 7279 on_rq = tsk->on_rq; 7280 7281 if (on_rq) 7282 dequeue_task(rq, tsk, 0); 7283 if (unlikely(running)) 7284 tsk->sched_class->put_prev_task(rq, tsk); 7285 7286 tg = container_of(task_subsys_state_check(tsk, cpu_cgroup_subsys_id, 7287 lockdep_is_held(&tsk->sighand->siglock)), 7288 struct task_group, css); 7289 tg = autogroup_task_group(tsk, tg); 7290 tsk->sched_task_group = tg; 7291 7292#ifdef CONFIG_FAIR_GROUP_SCHED 7293 if (tsk->sched_class->task_move_group) 7294 tsk->sched_class->task_move_group(tsk, on_rq); 7295 else 7296#endif 7297 set_task_rq(tsk, task_cpu(tsk)); 7298 7299 if (unlikely(running)) 7300 tsk->sched_class->set_curr_task(rq); 7301 if (on_rq) 7302 enqueue_task(rq, tsk, 0); 7303 7304 task_rq_unlock(rq, tsk, &flags); 7305} 7306#endif /* CONFIG_CGROUP_SCHED */ 7307 7308#if defined(CONFIG_RT_GROUP_SCHED) || defined(CONFIG_CFS_BANDWIDTH) 7309static unsigned long to_ratio(u64 period, u64 runtime) 7310{ 7311 if (runtime == RUNTIME_INF) 7312 return 1ULL << 20; 7313 7314 return div64_u64(runtime << 20, period); 7315} 7316#endif 7317 7318#ifdef CONFIG_RT_GROUP_SCHED 7319/* 7320 * Ensure that the real time constraints are schedulable. 7321 */ 7322static DEFINE_MUTEX(rt_constraints_mutex); 7323 7324/* Must be called with tasklist_lock held */ 7325static inline int tg_has_rt_tasks(struct task_group *tg) 7326{ 7327 struct task_struct *g, *p; 7328 7329 do_each_thread(g, p) { 7330 if (rt_task(p) && task_rq(p)->rt.tg == tg) 7331 return 1; 7332 } while_each_thread(g, p); 7333 7334 return 0; 7335} 7336 7337struct rt_schedulable_data { 7338 struct task_group *tg; 7339 u64 rt_period; 7340 u64 rt_runtime; 7341}; 7342 7343static int tg_rt_schedulable(struct task_group *tg, void *data) 7344{ 7345 struct rt_schedulable_data *d = data; 7346 struct task_group *child; 7347 unsigned long total, sum = 0; 7348 u64 period, runtime; 7349 7350 period = ktime_to_ns(tg->rt_bandwidth.rt_period); 7351 runtime = tg->rt_bandwidth.rt_runtime; 7352 7353 if (tg == d->tg) { 7354 period = d->rt_period; 7355 runtime = d->rt_runtime; 7356 } 7357 7358 /* 7359 * Cannot have more runtime than the period. 7360 */ 7361 if (runtime > period && runtime != RUNTIME_INF) 7362 return -EINVAL; 7363 7364 /* 7365 * Ensure we don't starve existing RT tasks. 7366 */ 7367 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg)) 7368 return -EBUSY; 7369 7370 total = to_ratio(period, runtime); 7371 7372 /* 7373 * Nobody can have more than the global setting allows. 7374 */ 7375 if (total > to_ratio(global_rt_period(), global_rt_runtime())) 7376 return -EINVAL; 7377 7378 /* 7379 * The sum of our children's runtime should not exceed our own. 7380 */ 7381 list_for_each_entry_rcu(child, &tg->children, siblings) { 7382 period = ktime_to_ns(child->rt_bandwidth.rt_period); 7383 runtime = child->rt_bandwidth.rt_runtime; 7384 7385 if (child == d->tg) { 7386 period = d->rt_period; 7387 runtime = d->rt_runtime; 7388 } 7389 7390 sum += to_ratio(period, runtime); 7391 } 7392 7393 if (sum > total) 7394 return -EINVAL; 7395 7396 return 0; 7397} 7398 7399static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime) 7400{ 7401 int ret; 7402 7403 struct rt_schedulable_data data = { 7404 .tg = tg, 7405 .rt_period = period, 7406 .rt_runtime = runtime, 7407 }; 7408 7409 rcu_read_lock(); 7410 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data); 7411 rcu_read_unlock(); 7412 7413 return ret; 7414} 7415 7416static int tg_set_rt_bandwidth(struct task_group *tg, 7417 u64 rt_period, u64 rt_runtime) 7418{ 7419 int i, err = 0; 7420 7421 mutex_lock(&rt_constraints_mutex); 7422 read_lock(&tasklist_lock); 7423 err = __rt_schedulable(tg, rt_period, rt_runtime); 7424 if (err) 7425 goto unlock; 7426 7427 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock); 7428 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period); 7429 tg->rt_bandwidth.rt_runtime = rt_runtime; 7430 7431 for_each_possible_cpu(i) { 7432 struct rt_rq *rt_rq = tg->rt_rq[i]; 7433 7434 raw_spin_lock(&rt_rq->rt_runtime_lock); 7435 rt_rq->rt_runtime = rt_runtime; 7436 raw_spin_unlock(&rt_rq->rt_runtime_lock); 7437 } 7438 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock); 7439unlock: 7440 read_unlock(&tasklist_lock); 7441 mutex_unlock(&rt_constraints_mutex); 7442 7443 return err; 7444} 7445 7446int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us) 7447{ 7448 u64 rt_runtime, rt_period; 7449 7450 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period); 7451 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC; 7452 if (rt_runtime_us < 0) 7453 rt_runtime = RUNTIME_INF; 7454 7455 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime); 7456} 7457 7458long sched_group_rt_runtime(struct task_group *tg) 7459{ 7460 u64 rt_runtime_us; 7461 7462 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF) 7463 return -1; 7464 7465 rt_runtime_us = tg->rt_bandwidth.rt_runtime; 7466 do_div(rt_runtime_us, NSEC_PER_USEC); 7467 return rt_runtime_us; 7468} 7469 7470int sched_group_set_rt_period(struct task_group *tg, long rt_period_us) 7471{ 7472 u64 rt_runtime, rt_period; 7473 7474 rt_period = (u64)rt_period_us * NSEC_PER_USEC; 7475 rt_runtime = tg->rt_bandwidth.rt_runtime; 7476 7477 if (rt_period == 0) 7478 return -EINVAL; 7479 7480 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime); 7481} 7482 7483long sched_group_rt_period(struct task_group *tg) 7484{ 7485 u64 rt_period_us; 7486 7487 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period); 7488 do_div(rt_period_us, NSEC_PER_USEC); 7489 return rt_period_us; 7490} 7491 7492static int sched_rt_global_constraints(void) 7493{ 7494 u64 runtime, period; 7495 int ret = 0; 7496 7497 if (sysctl_sched_rt_period <= 0) 7498 return -EINVAL; 7499 7500 runtime = global_rt_runtime(); 7501 period = global_rt_period(); 7502 7503 /* 7504 * Sanity check on the sysctl variables. 7505 */ 7506 if (runtime > period && runtime != RUNTIME_INF) 7507 return -EINVAL; 7508 7509 mutex_lock(&rt_constraints_mutex); 7510 read_lock(&tasklist_lock); 7511 ret = __rt_schedulable(NULL, 0, 0); 7512 read_unlock(&tasklist_lock); 7513 mutex_unlock(&rt_constraints_mutex); 7514 7515 return ret; 7516} 7517 7518int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk) 7519{ 7520 /* Don't accept realtime tasks when there is no way for them to run */ 7521 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0) 7522 return 0; 7523 7524 return 1; 7525} 7526 7527#else /* !CONFIG_RT_GROUP_SCHED */ 7528static int sched_rt_global_constraints(void) 7529{ 7530 unsigned long flags; 7531 int i; 7532 7533 if (sysctl_sched_rt_period <= 0) 7534 return -EINVAL; 7535 7536 /* 7537 * There's always some RT tasks in the root group 7538 * -- migration, kstopmachine etc.. 7539 */ 7540 if (sysctl_sched_rt_runtime == 0) 7541 return -EBUSY; 7542 7543 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags); 7544 for_each_possible_cpu(i) { 7545 struct rt_rq *rt_rq = &cpu_rq(i)->rt; 7546 7547 raw_spin_lock(&rt_rq->rt_runtime_lock); 7548 rt_rq->rt_runtime = global_rt_runtime(); 7549 raw_spin_unlock(&rt_rq->rt_runtime_lock); 7550 } 7551 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags); 7552 7553 return 0; 7554} 7555#endif /* CONFIG_RT_GROUP_SCHED */ 7556 7557int sched_rt_handler(struct ctl_table *table, int write, 7558 void __user *buffer, size_t *lenp, 7559 loff_t *ppos) 7560{ 7561 int ret; 7562 int old_period, old_runtime; 7563 static DEFINE_MUTEX(mutex); 7564 7565 mutex_lock(&mutex); 7566 old_period = sysctl_sched_rt_period; 7567 old_runtime = sysctl_sched_rt_runtime; 7568 7569 ret = proc_dointvec(table, write, buffer, lenp, ppos); 7570 7571 if (!ret && write) { 7572 ret = sched_rt_global_constraints(); 7573 if (ret) { 7574 sysctl_sched_rt_period = old_period; 7575 sysctl_sched_rt_runtime = old_runtime; 7576 } else { 7577 def_rt_bandwidth.rt_runtime = global_rt_runtime(); 7578 def_rt_bandwidth.rt_period = 7579 ns_to_ktime(global_rt_period()); 7580 } 7581 } 7582 mutex_unlock(&mutex); 7583 7584 return ret; 7585} 7586 7587#ifdef CONFIG_CGROUP_SCHED 7588 7589/* return corresponding task_group object of a cgroup */ 7590static inline struct task_group *cgroup_tg(struct cgroup *cgrp) 7591{ 7592 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id), 7593 struct task_group, css); 7594} 7595 7596static struct cgroup_subsys_state *cpu_cgroup_css_alloc(struct cgroup *cgrp) 7597{ 7598 struct task_group *tg, *parent; 7599 7600 if (!cgrp->parent) { 7601 /* This is early initialization for the top cgroup */ 7602 return &root_task_group.css; 7603 } 7604 7605 parent = cgroup_tg(cgrp->parent); 7606 tg = sched_create_group(parent); 7607 if (IS_ERR(tg)) 7608 return ERR_PTR(-ENOMEM); 7609 7610 return &tg->css; 7611} 7612 7613static void cpu_cgroup_css_free(struct cgroup *cgrp) 7614{ 7615 struct task_group *tg = cgroup_tg(cgrp); 7616 7617 sched_destroy_group(tg); 7618} 7619 7620static int cpu_cgroup_can_attach(struct cgroup *cgrp, 7621 struct cgroup_taskset *tset) 7622{ 7623 struct task_struct *task; 7624 7625 cgroup_taskset_for_each(task, cgrp, tset) { 7626#ifdef CONFIG_RT_GROUP_SCHED 7627 if (!sched_rt_can_attach(cgroup_tg(cgrp), task)) 7628 return -EINVAL; 7629#else 7630 /* We don't support RT-tasks being in separate groups */ 7631 if (task->sched_class != &fair_sched_class) 7632 return -EINVAL; 7633#endif 7634 } 7635 return 0; 7636} 7637 7638static void cpu_cgroup_attach(struct cgroup *cgrp, 7639 struct cgroup_taskset *tset) 7640{ 7641 struct task_struct *task; 7642 7643 cgroup_taskset_for_each(task, cgrp, tset) 7644 sched_move_task(task); 7645} 7646 7647static void 7648cpu_cgroup_exit(struct cgroup *cgrp, struct cgroup *old_cgrp, 7649 struct task_struct *task) 7650{ 7651 /* 7652 * cgroup_exit() is called in the copy_process() failure path. 7653 * Ignore this case since the task hasn't ran yet, this avoids 7654 * trying to poke a half freed task state from generic code. 7655 */ 7656 if (!(task->flags & PF_EXITING)) 7657 return; 7658 7659 sched_move_task(task); 7660} 7661 7662#ifdef CONFIG_FAIR_GROUP_SCHED 7663static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype, 7664 u64 shareval) 7665{ 7666 return sched_group_set_shares(cgroup_tg(cgrp), scale_load(shareval)); 7667} 7668 7669static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft) 7670{ 7671 struct task_group *tg = cgroup_tg(cgrp); 7672 7673 return (u64) scale_load_down(tg->shares); 7674} 7675 7676#ifdef CONFIG_CFS_BANDWIDTH 7677static DEFINE_MUTEX(cfs_constraints_mutex); 7678 7679const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */ 7680const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */ 7681 7682static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime); 7683 7684static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota) 7685{ 7686 int i, ret = 0, runtime_enabled, runtime_was_enabled; 7687 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth; 7688 7689 if (tg == &root_task_group) 7690 return -EINVAL; 7691 7692 /* 7693 * Ensure we have at some amount of bandwidth every period. This is 7694 * to prevent reaching a state of large arrears when throttled via 7695 * entity_tick() resulting in prolonged exit starvation. 7696 */ 7697 if (quota < min_cfs_quota_period || period < min_cfs_quota_period) 7698 return -EINVAL; 7699 7700 /* 7701 * Likewise, bound things on the otherside by preventing insane quota 7702 * periods. This also allows us to normalize in computing quota 7703 * feasibility. 7704 */ 7705 if (period > max_cfs_quota_period) 7706 return -EINVAL; 7707 7708 mutex_lock(&cfs_constraints_mutex); 7709 ret = __cfs_schedulable(tg, period, quota); 7710 if (ret) 7711 goto out_unlock; 7712 7713 runtime_enabled = quota != RUNTIME_INF; 7714 runtime_was_enabled = cfs_b->quota != RUNTIME_INF; 7715 account_cfs_bandwidth_used(runtime_enabled, runtime_was_enabled); 7716 raw_spin_lock_irq(&cfs_b->lock); 7717 cfs_b->period = ns_to_ktime(period); 7718 cfs_b->quota = quota; 7719 7720 __refill_cfs_bandwidth_runtime(cfs_b); 7721 /* restart the period timer (if active) to handle new period expiry */ 7722 if (runtime_enabled && cfs_b->timer_active) { 7723 /* force a reprogram */ 7724 cfs_b->timer_active = 0; 7725 __start_cfs_bandwidth(cfs_b); 7726 } 7727 raw_spin_unlock_irq(&cfs_b->lock); 7728 7729 for_each_possible_cpu(i) { 7730 struct cfs_rq *cfs_rq = tg->cfs_rq[i]; 7731 struct rq *rq = cfs_rq->rq; 7732 7733 raw_spin_lock_irq(&rq->lock); 7734 cfs_rq->runtime_enabled = runtime_enabled; 7735 cfs_rq->runtime_remaining = 0; 7736 7737 if (cfs_rq->throttled) 7738 unthrottle_cfs_rq(cfs_rq); 7739 raw_spin_unlock_irq(&rq->lock); 7740 } 7741out_unlock: 7742 mutex_unlock(&cfs_constraints_mutex); 7743 7744 return ret; 7745} 7746 7747int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us) 7748{ 7749 u64 quota, period; 7750 7751 period = ktime_to_ns(tg->cfs_bandwidth.period); 7752 if (cfs_quota_us < 0) 7753 quota = RUNTIME_INF; 7754 else 7755 quota = (u64)cfs_quota_us * NSEC_PER_USEC; 7756 7757 return tg_set_cfs_bandwidth(tg, period, quota); 7758} 7759 7760long tg_get_cfs_quota(struct task_group *tg) 7761{ 7762 u64 quota_us; 7763 7764 if (tg->cfs_bandwidth.quota == RUNTIME_INF) 7765 return -1; 7766 7767 quota_us = tg->cfs_bandwidth.quota; 7768 do_div(quota_us, NSEC_PER_USEC); 7769 7770 return quota_us; 7771} 7772 7773int tg_set_cfs_period(struct task_group *tg, long cfs_period_us) 7774{ 7775 u64 quota, period; 7776 7777 period = (u64)cfs_period_us * NSEC_PER_USEC; 7778 quota = tg->cfs_bandwidth.quota; 7779 7780 return tg_set_cfs_bandwidth(tg, period, quota); 7781} 7782 7783long tg_get_cfs_period(struct task_group *tg) 7784{ 7785 u64 cfs_period_us; 7786 7787 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period); 7788 do_div(cfs_period_us, NSEC_PER_USEC); 7789 7790 return cfs_period_us; 7791} 7792 7793static s64 cpu_cfs_quota_read_s64(struct cgroup *cgrp, struct cftype *cft) 7794{ 7795 return tg_get_cfs_quota(cgroup_tg(cgrp)); 7796} 7797 7798static int cpu_cfs_quota_write_s64(struct cgroup *cgrp, struct cftype *cftype, 7799 s64 cfs_quota_us) 7800{ 7801 return tg_set_cfs_quota(cgroup_tg(cgrp), cfs_quota_us); 7802} 7803 7804static u64 cpu_cfs_period_read_u64(struct cgroup *cgrp, struct cftype *cft) 7805{ 7806 return tg_get_cfs_period(cgroup_tg(cgrp)); 7807} 7808 7809static int cpu_cfs_period_write_u64(struct cgroup *cgrp, struct cftype *cftype, 7810 u64 cfs_period_us) 7811{ 7812 return tg_set_cfs_period(cgroup_tg(cgrp), cfs_period_us); 7813} 7814 7815struct cfs_schedulable_data { 7816 struct task_group *tg; 7817 u64 period, quota; 7818}; 7819 7820/* 7821 * normalize group quota/period to be quota/max_period 7822 * note: units are usecs 7823 */ 7824static u64 normalize_cfs_quota(struct task_group *tg, 7825 struct cfs_schedulable_data *d) 7826{ 7827 u64 quota, period; 7828 7829 if (tg == d->tg) { 7830 period = d->period; 7831 quota = d->quota; 7832 } else { 7833 period = tg_get_cfs_period(tg); 7834 quota = tg_get_cfs_quota(tg); 7835 } 7836 7837 /* note: these should typically be equivalent */ 7838 if (quota == RUNTIME_INF || quota == -1) 7839 return RUNTIME_INF; 7840 7841 return to_ratio(period, quota); 7842} 7843 7844static int tg_cfs_schedulable_down(struct task_group *tg, void *data) 7845{ 7846 struct cfs_schedulable_data *d = data; 7847 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth; 7848 s64 quota = 0, parent_quota = -1; 7849 7850 if (!tg->parent) { 7851 quota = RUNTIME_INF; 7852 } else { 7853 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth; 7854 7855 quota = normalize_cfs_quota(tg, d); 7856 parent_quota = parent_b->hierarchal_quota; 7857 7858 /* 7859 * ensure max(child_quota) <= parent_quota, inherit when no 7860 * limit is set 7861 */ 7862 if (quota == RUNTIME_INF) 7863 quota = parent_quota; 7864 else if (parent_quota != RUNTIME_INF && quota > parent_quota) 7865 return -EINVAL; 7866 } 7867 cfs_b->hierarchal_quota = quota; 7868 7869 return 0; 7870} 7871 7872static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota) 7873{ 7874 int ret; 7875 struct cfs_schedulable_data data = { 7876 .tg = tg, 7877 .period = period, 7878 .quota = quota, 7879 }; 7880 7881 if (quota != RUNTIME_INF) { 7882 do_div(data.period, NSEC_PER_USEC); 7883 do_div(data.quota, NSEC_PER_USEC); 7884 } 7885 7886 rcu_read_lock(); 7887 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data); 7888 rcu_read_unlock(); 7889 7890 return ret; 7891} 7892 7893static int cpu_stats_show(struct cgroup *cgrp, struct cftype *cft, 7894 struct cgroup_map_cb *cb) 7895{ 7896 struct task_group *tg = cgroup_tg(cgrp); 7897 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth; 7898 7899 cb->fill(cb, "nr_periods", cfs_b->nr_periods); 7900 cb->fill(cb, "nr_throttled", cfs_b->nr_throttled); 7901 cb->fill(cb, "throttled_time", cfs_b->throttled_time); 7902 7903 return 0; 7904} 7905#endif /* CONFIG_CFS_BANDWIDTH */ 7906#endif /* CONFIG_FAIR_GROUP_SCHED */ 7907 7908#ifdef CONFIG_RT_GROUP_SCHED 7909static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft, 7910 s64 val) 7911{ 7912 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val); 7913} 7914 7915static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft) 7916{ 7917 return sched_group_rt_runtime(cgroup_tg(cgrp)); 7918} 7919 7920static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype, 7921 u64 rt_period_us) 7922{ 7923 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us); 7924} 7925 7926static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft) 7927{ 7928 return sched_group_rt_period(cgroup_tg(cgrp)); 7929} 7930#endif /* CONFIG_RT_GROUP_SCHED */ 7931 7932static struct cftype cpu_files[] = { 7933#ifdef CONFIG_FAIR_GROUP_SCHED 7934 { 7935 .name = "shares", 7936 .read_u64 = cpu_shares_read_u64, 7937 .write_u64 = cpu_shares_write_u64, 7938 }, 7939#endif 7940#ifdef CONFIG_CFS_BANDWIDTH 7941 { 7942 .name = "cfs_quota_us", 7943 .read_s64 = cpu_cfs_quota_read_s64, 7944 .write_s64 = cpu_cfs_quota_write_s64, 7945 }, 7946 { 7947 .name = "cfs_period_us", 7948 .read_u64 = cpu_cfs_period_read_u64, 7949 .write_u64 = cpu_cfs_period_write_u64, 7950 }, 7951 { 7952 .name = "stat", 7953 .read_map = cpu_stats_show, 7954 }, 7955#endif 7956#ifdef CONFIG_RT_GROUP_SCHED 7957 { 7958 .name = "rt_runtime_us", 7959 .read_s64 = cpu_rt_runtime_read, 7960 .write_s64 = cpu_rt_runtime_write, 7961 }, 7962 { 7963 .name = "rt_period_us", 7964 .read_u64 = cpu_rt_period_read_uint, 7965 .write_u64 = cpu_rt_period_write_uint, 7966 }, 7967#endif 7968 { } /* terminate */ 7969}; 7970 7971struct cgroup_subsys cpu_cgroup_subsys = { 7972 .name = "cpu", 7973 .css_alloc = cpu_cgroup_css_alloc, 7974 .css_free = cpu_cgroup_css_free, 7975 .can_attach = cpu_cgroup_can_attach, 7976 .attach = cpu_cgroup_attach, 7977 .exit = cpu_cgroup_exit, 7978 .subsys_id = cpu_cgroup_subsys_id, 7979 .base_cftypes = cpu_files, 7980 .early_init = 1, 7981}; 7982 7983#endif /* CONFIG_CGROUP_SCHED */ 7984 7985#ifdef CONFIG_CGROUP_CPUACCT 7986 7987/* 7988 * CPU accounting code for task groups. 7989 * 7990 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh 7991 * (balbir@in.ibm.com). 7992 */ 7993 7994struct cpuacct root_cpuacct; 7995 7996/* create a new cpu accounting group */ 7997static struct cgroup_subsys_state *cpuacct_css_alloc(struct cgroup *cgrp) 7998{ 7999 struct cpuacct *ca; 8000 8001 if (!cgrp->parent) 8002 return &root_cpuacct.css; 8003 8004 ca = kzalloc(sizeof(*ca), GFP_KERNEL); 8005 if (!ca) 8006 goto out; 8007 8008 ca->cpuusage = alloc_percpu(u64); 8009 if (!ca->cpuusage) 8010 goto out_free_ca; 8011 8012 ca->cpustat = alloc_percpu(struct kernel_cpustat); 8013 if (!ca->cpustat) 8014 goto out_free_cpuusage; 8015 8016 return &ca->css; 8017 8018out_free_cpuusage: 8019 free_percpu(ca->cpuusage); 8020out_free_ca: 8021 kfree(ca); 8022out: 8023 return ERR_PTR(-ENOMEM); 8024} 8025 8026/* destroy an existing cpu accounting group */ 8027static void cpuacct_css_free(struct cgroup *cgrp) 8028{ 8029 struct cpuacct *ca = cgroup_ca(cgrp); 8030 8031 free_percpu(ca->cpustat); 8032 free_percpu(ca->cpuusage); 8033 kfree(ca); 8034} 8035 8036static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu) 8037{ 8038 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu); 8039 u64 data; 8040 8041#ifndef CONFIG_64BIT 8042 /* 8043 * Take rq->lock to make 64-bit read safe on 32-bit platforms. 8044 */ 8045 raw_spin_lock_irq(&cpu_rq(cpu)->lock); 8046 data = *cpuusage; 8047 raw_spin_unlock_irq(&cpu_rq(cpu)->lock); 8048#else 8049 data = *cpuusage; 8050#endif 8051 8052 return data; 8053} 8054 8055static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val) 8056{ 8057 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu); 8058 8059#ifndef CONFIG_64BIT 8060 /* 8061 * Take rq->lock to make 64-bit write safe on 32-bit platforms. 8062 */ 8063 raw_spin_lock_irq(&cpu_rq(cpu)->lock); 8064 *cpuusage = val; 8065 raw_spin_unlock_irq(&cpu_rq(cpu)->lock); 8066#else 8067 *cpuusage = val; 8068#endif 8069} 8070 8071/* return total cpu usage (in nanoseconds) of a group */ 8072static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft) 8073{ 8074 struct cpuacct *ca = cgroup_ca(cgrp); 8075 u64 totalcpuusage = 0; 8076 int i; 8077 8078 for_each_present_cpu(i) 8079 totalcpuusage += cpuacct_cpuusage_read(ca, i); 8080 8081 return totalcpuusage; 8082} 8083 8084static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype, 8085 u64 reset) 8086{ 8087 struct cpuacct *ca = cgroup_ca(cgrp); 8088 int err = 0; 8089 int i; 8090 8091 if (reset) { 8092 err = -EINVAL; 8093 goto out; 8094 } 8095 8096 for_each_present_cpu(i) 8097 cpuacct_cpuusage_write(ca, i, 0); 8098 8099out: 8100 return err; 8101} 8102 8103static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft, 8104 struct seq_file *m) 8105{ 8106 struct cpuacct *ca = cgroup_ca(cgroup); 8107 u64 percpu; 8108 int i; 8109 8110 for_each_present_cpu(i) { 8111 percpu = cpuacct_cpuusage_read(ca, i); 8112 seq_printf(m, "%llu ", (unsigned long long) percpu); 8113 } 8114 seq_printf(m, "\n"); 8115 return 0; 8116} 8117 8118static const char *cpuacct_stat_desc[] = { 8119 [CPUACCT_STAT_USER] = "user", 8120 [CPUACCT_STAT_SYSTEM] = "system", 8121}; 8122 8123static int cpuacct_stats_show(struct cgroup *cgrp, struct cftype *cft, 8124 struct cgroup_map_cb *cb) 8125{ 8126 struct cpuacct *ca = cgroup_ca(cgrp); 8127 int cpu; 8128 s64 val = 0; 8129 8130 for_each_online_cpu(cpu) { 8131 struct kernel_cpustat *kcpustat = per_cpu_ptr(ca->cpustat, cpu); 8132 val += kcpustat->cpustat[CPUTIME_USER]; 8133 val += kcpustat->cpustat[CPUTIME_NICE]; 8134 } 8135 val = cputime64_to_clock_t(val); 8136 cb->fill(cb, cpuacct_stat_desc[CPUACCT_STAT_USER], val); 8137 8138 val = 0; 8139 for_each_online_cpu(cpu) { 8140 struct kernel_cpustat *kcpustat = per_cpu_ptr(ca->cpustat, cpu); 8141 val += kcpustat->cpustat[CPUTIME_SYSTEM]; 8142 val += kcpustat->cpustat[CPUTIME_IRQ]; 8143 val += kcpustat->cpustat[CPUTIME_SOFTIRQ]; 8144 } 8145 8146 val = cputime64_to_clock_t(val); 8147 cb->fill(cb, cpuacct_stat_desc[CPUACCT_STAT_SYSTEM], val); 8148 8149 return 0; 8150} 8151 8152static struct cftype files[] = { 8153 { 8154 .name = "usage", 8155 .read_u64 = cpuusage_read, 8156 .write_u64 = cpuusage_write, 8157 }, 8158 { 8159 .name = "usage_percpu", 8160 .read_seq_string = cpuacct_percpu_seq_read, 8161 }, 8162 { 8163 .name = "stat", 8164 .read_map = cpuacct_stats_show, 8165 }, 8166 { } /* terminate */ 8167}; 8168 8169/* 8170 * charge this task's execution time to its accounting group. 8171 * 8172 * called with rq->lock held. 8173 */ 8174void cpuacct_charge(struct task_struct *tsk, u64 cputime) 8175{ 8176 struct cpuacct *ca; 8177 int cpu; 8178 8179 if (unlikely(!cpuacct_subsys.active)) 8180 return; 8181 8182 cpu = task_cpu(tsk); 8183 8184 rcu_read_lock(); 8185 8186 ca = task_ca(tsk); 8187 8188 for (; ca; ca = parent_ca(ca)) { 8189 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu); 8190 *cpuusage += cputime; 8191 } 8192 8193 rcu_read_unlock(); 8194} 8195 8196struct cgroup_subsys cpuacct_subsys = { 8197 .name = "cpuacct", 8198 .css_alloc = cpuacct_css_alloc, 8199 .css_free = cpuacct_css_free, 8200 .subsys_id = cpuacct_subsys_id, 8201 .base_cftypes = files, 8202}; 8203#endif /* CONFIG_CGROUP_CPUACCT */ 8204 8205void dump_cpu_task(int cpu) 8206{ 8207 pr_info("Task dump for CPU %d:\n", cpu); 8208 sched_show_task(cpu_curr(cpu)); 8209} 8210