core.c revision 6a61671bb2f3a1bd12cd17b8fca811a624782632
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, long timeout, int state) 3271{ 3272 if (!x->done) { 3273 DECLARE_WAITQUEUE(wait, current); 3274 3275 __add_wait_queue_tail_exclusive(&x->wait, &wait); 3276 do { 3277 if (signal_pending_state(state, current)) { 3278 timeout = -ERESTARTSYS; 3279 break; 3280 } 3281 __set_current_state(state); 3282 spin_unlock_irq(&x->wait.lock); 3283 timeout = schedule_timeout(timeout); 3284 spin_lock_irq(&x->wait.lock); 3285 } while (!x->done && timeout); 3286 __remove_wait_queue(&x->wait, &wait); 3287 if (!x->done) 3288 return timeout; 3289 } 3290 x->done--; 3291 return timeout ?: 1; 3292} 3293 3294static long __sched 3295wait_for_common(struct completion *x, long timeout, int state) 3296{ 3297 might_sleep(); 3298 3299 spin_lock_irq(&x->wait.lock); 3300 timeout = do_wait_for_common(x, timeout, state); 3301 spin_unlock_irq(&x->wait.lock); 3302 return timeout; 3303} 3304 3305/** 3306 * wait_for_completion: - waits for completion of a task 3307 * @x: holds the state of this particular completion 3308 * 3309 * This waits to be signaled for completion of a specific task. It is NOT 3310 * interruptible and there is no timeout. 3311 * 3312 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout 3313 * and interrupt capability. Also see complete(). 3314 */ 3315void __sched wait_for_completion(struct completion *x) 3316{ 3317 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE); 3318} 3319EXPORT_SYMBOL(wait_for_completion); 3320 3321/** 3322 * wait_for_completion_timeout: - waits for completion of a task (w/timeout) 3323 * @x: holds the state of this particular completion 3324 * @timeout: timeout value in jiffies 3325 * 3326 * This waits for either a completion of a specific task to be signaled or for a 3327 * specified timeout to expire. The timeout is in jiffies. It is not 3328 * interruptible. 3329 * 3330 * The return value is 0 if timed out, and positive (at least 1, or number of 3331 * jiffies left till timeout) if completed. 3332 */ 3333unsigned long __sched 3334wait_for_completion_timeout(struct completion *x, unsigned long timeout) 3335{ 3336 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE); 3337} 3338EXPORT_SYMBOL(wait_for_completion_timeout); 3339 3340/** 3341 * wait_for_completion_interruptible: - waits for completion of a task (w/intr) 3342 * @x: holds the state of this particular completion 3343 * 3344 * This waits for completion of a specific task to be signaled. It is 3345 * interruptible. 3346 * 3347 * The return value is -ERESTARTSYS if interrupted, 0 if completed. 3348 */ 3349int __sched wait_for_completion_interruptible(struct completion *x) 3350{ 3351 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE); 3352 if (t == -ERESTARTSYS) 3353 return t; 3354 return 0; 3355} 3356EXPORT_SYMBOL(wait_for_completion_interruptible); 3357 3358/** 3359 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr)) 3360 * @x: holds the state of this particular completion 3361 * @timeout: timeout value in jiffies 3362 * 3363 * This waits for either a completion of a specific task to be signaled or for a 3364 * specified timeout to expire. It is interruptible. The timeout is in jiffies. 3365 * 3366 * The return value is -ERESTARTSYS if interrupted, 0 if timed out, 3367 * positive (at least 1, or number of jiffies left till timeout) if completed. 3368 */ 3369long __sched 3370wait_for_completion_interruptible_timeout(struct completion *x, 3371 unsigned long timeout) 3372{ 3373 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE); 3374} 3375EXPORT_SYMBOL(wait_for_completion_interruptible_timeout); 3376 3377/** 3378 * wait_for_completion_killable: - waits for completion of a task (killable) 3379 * @x: holds the state of this particular completion 3380 * 3381 * This waits to be signaled for completion of a specific task. It can be 3382 * interrupted by a kill signal. 3383 * 3384 * The return value is -ERESTARTSYS if interrupted, 0 if completed. 3385 */ 3386int __sched wait_for_completion_killable(struct completion *x) 3387{ 3388 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE); 3389 if (t == -ERESTARTSYS) 3390 return t; 3391 return 0; 3392} 3393EXPORT_SYMBOL(wait_for_completion_killable); 3394 3395/** 3396 * wait_for_completion_killable_timeout: - waits for completion of a task (w/(to,killable)) 3397 * @x: holds the state of this particular completion 3398 * @timeout: timeout value in jiffies 3399 * 3400 * This waits for either a completion of a specific task to be 3401 * signaled or for a specified timeout to expire. It can be 3402 * interrupted by a kill signal. The timeout is in jiffies. 3403 * 3404 * The return value is -ERESTARTSYS if interrupted, 0 if timed out, 3405 * positive (at least 1, or number of jiffies left till timeout) if completed. 3406 */ 3407long __sched 3408wait_for_completion_killable_timeout(struct completion *x, 3409 unsigned long timeout) 3410{ 3411 return wait_for_common(x, timeout, TASK_KILLABLE); 3412} 3413EXPORT_SYMBOL(wait_for_completion_killable_timeout); 3414 3415/** 3416 * try_wait_for_completion - try to decrement a completion without blocking 3417 * @x: completion structure 3418 * 3419 * Returns: 0 if a decrement cannot be done without blocking 3420 * 1 if a decrement succeeded. 3421 * 3422 * If a completion is being used as a counting completion, 3423 * attempt to decrement the counter without blocking. This 3424 * enables us to avoid waiting if the resource the completion 3425 * is protecting is not available. 3426 */ 3427bool try_wait_for_completion(struct completion *x) 3428{ 3429 unsigned long flags; 3430 int ret = 1; 3431 3432 spin_lock_irqsave(&x->wait.lock, flags); 3433 if (!x->done) 3434 ret = 0; 3435 else 3436 x->done--; 3437 spin_unlock_irqrestore(&x->wait.lock, flags); 3438 return ret; 3439} 3440EXPORT_SYMBOL(try_wait_for_completion); 3441 3442/** 3443 * completion_done - Test to see if a completion has any waiters 3444 * @x: completion structure 3445 * 3446 * Returns: 0 if there are waiters (wait_for_completion() in progress) 3447 * 1 if there are no waiters. 3448 * 3449 */ 3450bool completion_done(struct completion *x) 3451{ 3452 unsigned long flags; 3453 int ret = 1; 3454 3455 spin_lock_irqsave(&x->wait.lock, flags); 3456 if (!x->done) 3457 ret = 0; 3458 spin_unlock_irqrestore(&x->wait.lock, flags); 3459 return ret; 3460} 3461EXPORT_SYMBOL(completion_done); 3462 3463static long __sched 3464sleep_on_common(wait_queue_head_t *q, int state, long timeout) 3465{ 3466 unsigned long flags; 3467 wait_queue_t wait; 3468 3469 init_waitqueue_entry(&wait, current); 3470 3471 __set_current_state(state); 3472 3473 spin_lock_irqsave(&q->lock, flags); 3474 __add_wait_queue(q, &wait); 3475 spin_unlock(&q->lock); 3476 timeout = schedule_timeout(timeout); 3477 spin_lock_irq(&q->lock); 3478 __remove_wait_queue(q, &wait); 3479 spin_unlock_irqrestore(&q->lock, flags); 3480 3481 return timeout; 3482} 3483 3484void __sched interruptible_sleep_on(wait_queue_head_t *q) 3485{ 3486 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT); 3487} 3488EXPORT_SYMBOL(interruptible_sleep_on); 3489 3490long __sched 3491interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout) 3492{ 3493 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout); 3494} 3495EXPORT_SYMBOL(interruptible_sleep_on_timeout); 3496 3497void __sched sleep_on(wait_queue_head_t *q) 3498{ 3499 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT); 3500} 3501EXPORT_SYMBOL(sleep_on); 3502 3503long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout) 3504{ 3505 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout); 3506} 3507EXPORT_SYMBOL(sleep_on_timeout); 3508 3509#ifdef CONFIG_RT_MUTEXES 3510 3511/* 3512 * rt_mutex_setprio - set the current priority of a task 3513 * @p: task 3514 * @prio: prio value (kernel-internal form) 3515 * 3516 * This function changes the 'effective' priority of a task. It does 3517 * not touch ->normal_prio like __setscheduler(). 3518 * 3519 * Used by the rt_mutex code to implement priority inheritance logic. 3520 */ 3521void rt_mutex_setprio(struct task_struct *p, int prio) 3522{ 3523 int oldprio, on_rq, running; 3524 struct rq *rq; 3525 const struct sched_class *prev_class; 3526 3527 BUG_ON(prio < 0 || prio > MAX_PRIO); 3528 3529 rq = __task_rq_lock(p); 3530 3531 /* 3532 * Idle task boosting is a nono in general. There is one 3533 * exception, when PREEMPT_RT and NOHZ is active: 3534 * 3535 * The idle task calls get_next_timer_interrupt() and holds 3536 * the timer wheel base->lock on the CPU and another CPU wants 3537 * to access the timer (probably to cancel it). We can safely 3538 * ignore the boosting request, as the idle CPU runs this code 3539 * with interrupts disabled and will complete the lock 3540 * protected section without being interrupted. So there is no 3541 * real need to boost. 3542 */ 3543 if (unlikely(p == rq->idle)) { 3544 WARN_ON(p != rq->curr); 3545 WARN_ON(p->pi_blocked_on); 3546 goto out_unlock; 3547 } 3548 3549 trace_sched_pi_setprio(p, prio); 3550 oldprio = p->prio; 3551 prev_class = p->sched_class; 3552 on_rq = p->on_rq; 3553 running = task_current(rq, p); 3554 if (on_rq) 3555 dequeue_task(rq, p, 0); 3556 if (running) 3557 p->sched_class->put_prev_task(rq, p); 3558 3559 if (rt_prio(prio)) 3560 p->sched_class = &rt_sched_class; 3561 else 3562 p->sched_class = &fair_sched_class; 3563 3564 p->prio = prio; 3565 3566 if (running) 3567 p->sched_class->set_curr_task(rq); 3568 if (on_rq) 3569 enqueue_task(rq, p, oldprio < prio ? ENQUEUE_HEAD : 0); 3570 3571 check_class_changed(rq, p, prev_class, oldprio); 3572out_unlock: 3573 __task_rq_unlock(rq); 3574} 3575#endif 3576void set_user_nice(struct task_struct *p, long nice) 3577{ 3578 int old_prio, delta, on_rq; 3579 unsigned long flags; 3580 struct rq *rq; 3581 3582 if (TASK_NICE(p) == nice || nice < -20 || nice > 19) 3583 return; 3584 /* 3585 * We have to be careful, if called from sys_setpriority(), 3586 * the task might be in the middle of scheduling on another CPU. 3587 */ 3588 rq = task_rq_lock(p, &flags); 3589 /* 3590 * The RT priorities are set via sched_setscheduler(), but we still 3591 * allow the 'normal' nice value to be set - but as expected 3592 * it wont have any effect on scheduling until the task is 3593 * SCHED_FIFO/SCHED_RR: 3594 */ 3595 if (task_has_rt_policy(p)) { 3596 p->static_prio = NICE_TO_PRIO(nice); 3597 goto out_unlock; 3598 } 3599 on_rq = p->on_rq; 3600 if (on_rq) 3601 dequeue_task(rq, p, 0); 3602 3603 p->static_prio = NICE_TO_PRIO(nice); 3604 set_load_weight(p); 3605 old_prio = p->prio; 3606 p->prio = effective_prio(p); 3607 delta = p->prio - old_prio; 3608 3609 if (on_rq) { 3610 enqueue_task(rq, p, 0); 3611 /* 3612 * If the task increased its priority or is running and 3613 * lowered its priority, then reschedule its CPU: 3614 */ 3615 if (delta < 0 || (delta > 0 && task_running(rq, p))) 3616 resched_task(rq->curr); 3617 } 3618out_unlock: 3619 task_rq_unlock(rq, p, &flags); 3620} 3621EXPORT_SYMBOL(set_user_nice); 3622 3623/* 3624 * can_nice - check if a task can reduce its nice value 3625 * @p: task 3626 * @nice: nice value 3627 */ 3628int can_nice(const struct task_struct *p, const int nice) 3629{ 3630 /* convert nice value [19,-20] to rlimit style value [1,40] */ 3631 int nice_rlim = 20 - nice; 3632 3633 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) || 3634 capable(CAP_SYS_NICE)); 3635} 3636 3637#ifdef __ARCH_WANT_SYS_NICE 3638 3639/* 3640 * sys_nice - change the priority of the current process. 3641 * @increment: priority increment 3642 * 3643 * sys_setpriority is a more generic, but much slower function that 3644 * does similar things. 3645 */ 3646SYSCALL_DEFINE1(nice, int, increment) 3647{ 3648 long nice, retval; 3649 3650 /* 3651 * Setpriority might change our priority at the same moment. 3652 * We don't have to worry. Conceptually one call occurs first 3653 * and we have a single winner. 3654 */ 3655 if (increment < -40) 3656 increment = -40; 3657 if (increment > 40) 3658 increment = 40; 3659 3660 nice = TASK_NICE(current) + increment; 3661 if (nice < -20) 3662 nice = -20; 3663 if (nice > 19) 3664 nice = 19; 3665 3666 if (increment < 0 && !can_nice(current, nice)) 3667 return -EPERM; 3668 3669 retval = security_task_setnice(current, nice); 3670 if (retval) 3671 return retval; 3672 3673 set_user_nice(current, nice); 3674 return 0; 3675} 3676 3677#endif 3678 3679/** 3680 * task_prio - return the priority value of a given task. 3681 * @p: the task in question. 3682 * 3683 * This is the priority value as seen by users in /proc. 3684 * RT tasks are offset by -200. Normal tasks are centered 3685 * around 0, value goes from -16 to +15. 3686 */ 3687int task_prio(const struct task_struct *p) 3688{ 3689 return p->prio - MAX_RT_PRIO; 3690} 3691 3692/** 3693 * task_nice - return the nice value of a given task. 3694 * @p: the task in question. 3695 */ 3696int task_nice(const struct task_struct *p) 3697{ 3698 return TASK_NICE(p); 3699} 3700EXPORT_SYMBOL(task_nice); 3701 3702/** 3703 * idle_cpu - is a given cpu idle currently? 3704 * @cpu: the processor in question. 3705 */ 3706int idle_cpu(int cpu) 3707{ 3708 struct rq *rq = cpu_rq(cpu); 3709 3710 if (rq->curr != rq->idle) 3711 return 0; 3712 3713 if (rq->nr_running) 3714 return 0; 3715 3716#ifdef CONFIG_SMP 3717 if (!llist_empty(&rq->wake_list)) 3718 return 0; 3719#endif 3720 3721 return 1; 3722} 3723 3724/** 3725 * idle_task - return the idle task for a given cpu. 3726 * @cpu: the processor in question. 3727 */ 3728struct task_struct *idle_task(int cpu) 3729{ 3730 return cpu_rq(cpu)->idle; 3731} 3732 3733/** 3734 * find_process_by_pid - find a process with a matching PID value. 3735 * @pid: the pid in question. 3736 */ 3737static struct task_struct *find_process_by_pid(pid_t pid) 3738{ 3739 return pid ? find_task_by_vpid(pid) : current; 3740} 3741 3742/* Actually do priority change: must hold rq lock. */ 3743static void 3744__setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio) 3745{ 3746 p->policy = policy; 3747 p->rt_priority = prio; 3748 p->normal_prio = normal_prio(p); 3749 /* we are holding p->pi_lock already */ 3750 p->prio = rt_mutex_getprio(p); 3751 if (rt_prio(p->prio)) 3752 p->sched_class = &rt_sched_class; 3753 else 3754 p->sched_class = &fair_sched_class; 3755 set_load_weight(p); 3756} 3757 3758/* 3759 * check the target process has a UID that matches the current process's 3760 */ 3761static bool check_same_owner(struct task_struct *p) 3762{ 3763 const struct cred *cred = current_cred(), *pcred; 3764 bool match; 3765 3766 rcu_read_lock(); 3767 pcred = __task_cred(p); 3768 match = (uid_eq(cred->euid, pcred->euid) || 3769 uid_eq(cred->euid, pcred->uid)); 3770 rcu_read_unlock(); 3771 return match; 3772} 3773 3774static int __sched_setscheduler(struct task_struct *p, int policy, 3775 const struct sched_param *param, bool user) 3776{ 3777 int retval, oldprio, oldpolicy = -1, on_rq, running; 3778 unsigned long flags; 3779 const struct sched_class *prev_class; 3780 struct rq *rq; 3781 int reset_on_fork; 3782 3783 /* may grab non-irq protected spin_locks */ 3784 BUG_ON(in_interrupt()); 3785recheck: 3786 /* double check policy once rq lock held */ 3787 if (policy < 0) { 3788 reset_on_fork = p->sched_reset_on_fork; 3789 policy = oldpolicy = p->policy; 3790 } else { 3791 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK); 3792 policy &= ~SCHED_RESET_ON_FORK; 3793 3794 if (policy != SCHED_FIFO && policy != SCHED_RR && 3795 policy != SCHED_NORMAL && policy != SCHED_BATCH && 3796 policy != SCHED_IDLE) 3797 return -EINVAL; 3798 } 3799 3800 /* 3801 * Valid priorities for SCHED_FIFO and SCHED_RR are 3802 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL, 3803 * SCHED_BATCH and SCHED_IDLE is 0. 3804 */ 3805 if (param->sched_priority < 0 || 3806 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) || 3807 (!p->mm && param->sched_priority > MAX_RT_PRIO-1)) 3808 return -EINVAL; 3809 if (rt_policy(policy) != (param->sched_priority != 0)) 3810 return -EINVAL; 3811 3812 /* 3813 * Allow unprivileged RT tasks to decrease priority: 3814 */ 3815 if (user && !capable(CAP_SYS_NICE)) { 3816 if (rt_policy(policy)) { 3817 unsigned long rlim_rtprio = 3818 task_rlimit(p, RLIMIT_RTPRIO); 3819 3820 /* can't set/change the rt policy */ 3821 if (policy != p->policy && !rlim_rtprio) 3822 return -EPERM; 3823 3824 /* can't increase priority */ 3825 if (param->sched_priority > p->rt_priority && 3826 param->sched_priority > rlim_rtprio) 3827 return -EPERM; 3828 } 3829 3830 /* 3831 * Treat SCHED_IDLE as nice 20. Only allow a switch to 3832 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it. 3833 */ 3834 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) { 3835 if (!can_nice(p, TASK_NICE(p))) 3836 return -EPERM; 3837 } 3838 3839 /* can't change other user's priorities */ 3840 if (!check_same_owner(p)) 3841 return -EPERM; 3842 3843 /* Normal users shall not reset the sched_reset_on_fork flag */ 3844 if (p->sched_reset_on_fork && !reset_on_fork) 3845 return -EPERM; 3846 } 3847 3848 if (user) { 3849 retval = security_task_setscheduler(p); 3850 if (retval) 3851 return retval; 3852 } 3853 3854 /* 3855 * make sure no PI-waiters arrive (or leave) while we are 3856 * changing the priority of the task: 3857 * 3858 * To be able to change p->policy safely, the appropriate 3859 * runqueue lock must be held. 3860 */ 3861 rq = task_rq_lock(p, &flags); 3862 3863 /* 3864 * Changing the policy of the stop threads its a very bad idea 3865 */ 3866 if (p == rq->stop) { 3867 task_rq_unlock(rq, p, &flags); 3868 return -EINVAL; 3869 } 3870 3871 /* 3872 * If not changing anything there's no need to proceed further: 3873 */ 3874 if (unlikely(policy == p->policy && (!rt_policy(policy) || 3875 param->sched_priority == p->rt_priority))) { 3876 task_rq_unlock(rq, p, &flags); 3877 return 0; 3878 } 3879 3880#ifdef CONFIG_RT_GROUP_SCHED 3881 if (user) { 3882 /* 3883 * Do not allow realtime tasks into groups that have no runtime 3884 * assigned. 3885 */ 3886 if (rt_bandwidth_enabled() && rt_policy(policy) && 3887 task_group(p)->rt_bandwidth.rt_runtime == 0 && 3888 !task_group_is_autogroup(task_group(p))) { 3889 task_rq_unlock(rq, p, &flags); 3890 return -EPERM; 3891 } 3892 } 3893#endif 3894 3895 /* recheck policy now with rq lock held */ 3896 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) { 3897 policy = oldpolicy = -1; 3898 task_rq_unlock(rq, p, &flags); 3899 goto recheck; 3900 } 3901 on_rq = p->on_rq; 3902 running = task_current(rq, p); 3903 if (on_rq) 3904 dequeue_task(rq, p, 0); 3905 if (running) 3906 p->sched_class->put_prev_task(rq, p); 3907 3908 p->sched_reset_on_fork = reset_on_fork; 3909 3910 oldprio = p->prio; 3911 prev_class = p->sched_class; 3912 __setscheduler(rq, p, policy, param->sched_priority); 3913 3914 if (running) 3915 p->sched_class->set_curr_task(rq); 3916 if (on_rq) 3917 enqueue_task(rq, p, 0); 3918 3919 check_class_changed(rq, p, prev_class, oldprio); 3920 task_rq_unlock(rq, p, &flags); 3921 3922 rt_mutex_adjust_pi(p); 3923 3924 return 0; 3925} 3926 3927/** 3928 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread. 3929 * @p: the task in question. 3930 * @policy: new policy. 3931 * @param: structure containing the new RT priority. 3932 * 3933 * NOTE that the task may be already dead. 3934 */ 3935int sched_setscheduler(struct task_struct *p, int policy, 3936 const struct sched_param *param) 3937{ 3938 return __sched_setscheduler(p, policy, param, true); 3939} 3940EXPORT_SYMBOL_GPL(sched_setscheduler); 3941 3942/** 3943 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace. 3944 * @p: the task in question. 3945 * @policy: new policy. 3946 * @param: structure containing the new RT priority. 3947 * 3948 * Just like sched_setscheduler, only don't bother checking if the 3949 * current context has permission. For example, this is needed in 3950 * stop_machine(): we create temporary high priority worker threads, 3951 * but our caller might not have that capability. 3952 */ 3953int sched_setscheduler_nocheck(struct task_struct *p, int policy, 3954 const struct sched_param *param) 3955{ 3956 return __sched_setscheduler(p, policy, param, false); 3957} 3958 3959static int 3960do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param) 3961{ 3962 struct sched_param lparam; 3963 struct task_struct *p; 3964 int retval; 3965 3966 if (!param || pid < 0) 3967 return -EINVAL; 3968 if (copy_from_user(&lparam, param, sizeof(struct sched_param))) 3969 return -EFAULT; 3970 3971 rcu_read_lock(); 3972 retval = -ESRCH; 3973 p = find_process_by_pid(pid); 3974 if (p != NULL) 3975 retval = sched_setscheduler(p, policy, &lparam); 3976 rcu_read_unlock(); 3977 3978 return retval; 3979} 3980 3981/** 3982 * sys_sched_setscheduler - set/change the scheduler policy and RT priority 3983 * @pid: the pid in question. 3984 * @policy: new policy. 3985 * @param: structure containing the new RT priority. 3986 */ 3987SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy, 3988 struct sched_param __user *, param) 3989{ 3990 /* negative values for policy are not valid */ 3991 if (policy < 0) 3992 return -EINVAL; 3993 3994 return do_sched_setscheduler(pid, policy, param); 3995} 3996 3997/** 3998 * sys_sched_setparam - set/change the RT priority of a thread 3999 * @pid: the pid in question. 4000 * @param: structure containing the new RT priority. 4001 */ 4002SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param) 4003{ 4004 return do_sched_setscheduler(pid, -1, param); 4005} 4006 4007/** 4008 * sys_sched_getscheduler - get the policy (scheduling class) of a thread 4009 * @pid: the pid in question. 4010 */ 4011SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid) 4012{ 4013 struct task_struct *p; 4014 int retval; 4015 4016 if (pid < 0) 4017 return -EINVAL; 4018 4019 retval = -ESRCH; 4020 rcu_read_lock(); 4021 p = find_process_by_pid(pid); 4022 if (p) { 4023 retval = security_task_getscheduler(p); 4024 if (!retval) 4025 retval = p->policy 4026 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0); 4027 } 4028 rcu_read_unlock(); 4029 return retval; 4030} 4031 4032/** 4033 * sys_sched_getparam - get the RT priority of a thread 4034 * @pid: the pid in question. 4035 * @param: structure containing the RT priority. 4036 */ 4037SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param) 4038{ 4039 struct sched_param lp; 4040 struct task_struct *p; 4041 int retval; 4042 4043 if (!param || pid < 0) 4044 return -EINVAL; 4045 4046 rcu_read_lock(); 4047 p = find_process_by_pid(pid); 4048 retval = -ESRCH; 4049 if (!p) 4050 goto out_unlock; 4051 4052 retval = security_task_getscheduler(p); 4053 if (retval) 4054 goto out_unlock; 4055 4056 lp.sched_priority = p->rt_priority; 4057 rcu_read_unlock(); 4058 4059 /* 4060 * This one might sleep, we cannot do it with a spinlock held ... 4061 */ 4062 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0; 4063 4064 return retval; 4065 4066out_unlock: 4067 rcu_read_unlock(); 4068 return retval; 4069} 4070 4071long sched_setaffinity(pid_t pid, const struct cpumask *in_mask) 4072{ 4073 cpumask_var_t cpus_allowed, new_mask; 4074 struct task_struct *p; 4075 int retval; 4076 4077 get_online_cpus(); 4078 rcu_read_lock(); 4079 4080 p = find_process_by_pid(pid); 4081 if (!p) { 4082 rcu_read_unlock(); 4083 put_online_cpus(); 4084 return -ESRCH; 4085 } 4086 4087 /* Prevent p going away */ 4088 get_task_struct(p); 4089 rcu_read_unlock(); 4090 4091 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) { 4092 retval = -ENOMEM; 4093 goto out_put_task; 4094 } 4095 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) { 4096 retval = -ENOMEM; 4097 goto out_free_cpus_allowed; 4098 } 4099 retval = -EPERM; 4100 if (!check_same_owner(p)) { 4101 rcu_read_lock(); 4102 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) { 4103 rcu_read_unlock(); 4104 goto out_unlock; 4105 } 4106 rcu_read_unlock(); 4107 } 4108 4109 retval = security_task_setscheduler(p); 4110 if (retval) 4111 goto out_unlock; 4112 4113 cpuset_cpus_allowed(p, cpus_allowed); 4114 cpumask_and(new_mask, in_mask, cpus_allowed); 4115again: 4116 retval = set_cpus_allowed_ptr(p, new_mask); 4117 4118 if (!retval) { 4119 cpuset_cpus_allowed(p, cpus_allowed); 4120 if (!cpumask_subset(new_mask, cpus_allowed)) { 4121 /* 4122 * We must have raced with a concurrent cpuset 4123 * update. Just reset the cpus_allowed to the 4124 * cpuset's cpus_allowed 4125 */ 4126 cpumask_copy(new_mask, cpus_allowed); 4127 goto again; 4128 } 4129 } 4130out_unlock: 4131 free_cpumask_var(new_mask); 4132out_free_cpus_allowed: 4133 free_cpumask_var(cpus_allowed); 4134out_put_task: 4135 put_task_struct(p); 4136 put_online_cpus(); 4137 return retval; 4138} 4139 4140static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len, 4141 struct cpumask *new_mask) 4142{ 4143 if (len < cpumask_size()) 4144 cpumask_clear(new_mask); 4145 else if (len > cpumask_size()) 4146 len = cpumask_size(); 4147 4148 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0; 4149} 4150 4151/** 4152 * sys_sched_setaffinity - set the cpu affinity of a process 4153 * @pid: pid of the process 4154 * @len: length in bytes of the bitmask pointed to by user_mask_ptr 4155 * @user_mask_ptr: user-space pointer to the new cpu mask 4156 */ 4157SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len, 4158 unsigned long __user *, user_mask_ptr) 4159{ 4160 cpumask_var_t new_mask; 4161 int retval; 4162 4163 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) 4164 return -ENOMEM; 4165 4166 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask); 4167 if (retval == 0) 4168 retval = sched_setaffinity(pid, new_mask); 4169 free_cpumask_var(new_mask); 4170 return retval; 4171} 4172 4173long sched_getaffinity(pid_t pid, struct cpumask *mask) 4174{ 4175 struct task_struct *p; 4176 unsigned long flags; 4177 int retval; 4178 4179 get_online_cpus(); 4180 rcu_read_lock(); 4181 4182 retval = -ESRCH; 4183 p = find_process_by_pid(pid); 4184 if (!p) 4185 goto out_unlock; 4186 4187 retval = security_task_getscheduler(p); 4188 if (retval) 4189 goto out_unlock; 4190 4191 raw_spin_lock_irqsave(&p->pi_lock, flags); 4192 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask); 4193 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 4194 4195out_unlock: 4196 rcu_read_unlock(); 4197 put_online_cpus(); 4198 4199 return retval; 4200} 4201 4202/** 4203 * sys_sched_getaffinity - get the cpu affinity of a process 4204 * @pid: pid of the process 4205 * @len: length in bytes of the bitmask pointed to by user_mask_ptr 4206 * @user_mask_ptr: user-space pointer to hold the current cpu mask 4207 */ 4208SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len, 4209 unsigned long __user *, user_mask_ptr) 4210{ 4211 int ret; 4212 cpumask_var_t mask; 4213 4214 if ((len * BITS_PER_BYTE) < nr_cpu_ids) 4215 return -EINVAL; 4216 if (len & (sizeof(unsigned long)-1)) 4217 return -EINVAL; 4218 4219 if (!alloc_cpumask_var(&mask, GFP_KERNEL)) 4220 return -ENOMEM; 4221 4222 ret = sched_getaffinity(pid, mask); 4223 if (ret == 0) { 4224 size_t retlen = min_t(size_t, len, cpumask_size()); 4225 4226 if (copy_to_user(user_mask_ptr, mask, retlen)) 4227 ret = -EFAULT; 4228 else 4229 ret = retlen; 4230 } 4231 free_cpumask_var(mask); 4232 4233 return ret; 4234} 4235 4236/** 4237 * sys_sched_yield - yield the current processor to other threads. 4238 * 4239 * This function yields the current CPU to other tasks. If there are no 4240 * other threads running on this CPU then this function will return. 4241 */ 4242SYSCALL_DEFINE0(sched_yield) 4243{ 4244 struct rq *rq = this_rq_lock(); 4245 4246 schedstat_inc(rq, yld_count); 4247 current->sched_class->yield_task(rq); 4248 4249 /* 4250 * Since we are going to call schedule() anyway, there's 4251 * no need to preempt or enable interrupts: 4252 */ 4253 __release(rq->lock); 4254 spin_release(&rq->lock.dep_map, 1, _THIS_IP_); 4255 do_raw_spin_unlock(&rq->lock); 4256 sched_preempt_enable_no_resched(); 4257 4258 schedule(); 4259 4260 return 0; 4261} 4262 4263static inline int should_resched(void) 4264{ 4265 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE); 4266} 4267 4268static void __cond_resched(void) 4269{ 4270 add_preempt_count(PREEMPT_ACTIVE); 4271 __schedule(); 4272 sub_preempt_count(PREEMPT_ACTIVE); 4273} 4274 4275int __sched _cond_resched(void) 4276{ 4277 if (should_resched()) { 4278 __cond_resched(); 4279 return 1; 4280 } 4281 return 0; 4282} 4283EXPORT_SYMBOL(_cond_resched); 4284 4285/* 4286 * __cond_resched_lock() - if a reschedule is pending, drop the given lock, 4287 * call schedule, and on return reacquire the lock. 4288 * 4289 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level 4290 * operations here to prevent schedule() from being called twice (once via 4291 * spin_unlock(), once by hand). 4292 */ 4293int __cond_resched_lock(spinlock_t *lock) 4294{ 4295 int resched = should_resched(); 4296 int ret = 0; 4297 4298 lockdep_assert_held(lock); 4299 4300 if (spin_needbreak(lock) || resched) { 4301 spin_unlock(lock); 4302 if (resched) 4303 __cond_resched(); 4304 else 4305 cpu_relax(); 4306 ret = 1; 4307 spin_lock(lock); 4308 } 4309 return ret; 4310} 4311EXPORT_SYMBOL(__cond_resched_lock); 4312 4313int __sched __cond_resched_softirq(void) 4314{ 4315 BUG_ON(!in_softirq()); 4316 4317 if (should_resched()) { 4318 local_bh_enable(); 4319 __cond_resched(); 4320 local_bh_disable(); 4321 return 1; 4322 } 4323 return 0; 4324} 4325EXPORT_SYMBOL(__cond_resched_softirq); 4326 4327/** 4328 * yield - yield the current processor to other threads. 4329 * 4330 * Do not ever use this function, there's a 99% chance you're doing it wrong. 4331 * 4332 * The scheduler is at all times free to pick the calling task as the most 4333 * eligible task to run, if removing the yield() call from your code breaks 4334 * it, its already broken. 4335 * 4336 * Typical broken usage is: 4337 * 4338 * while (!event) 4339 * yield(); 4340 * 4341 * where one assumes that yield() will let 'the other' process run that will 4342 * make event true. If the current task is a SCHED_FIFO task that will never 4343 * happen. Never use yield() as a progress guarantee!! 4344 * 4345 * If you want to use yield() to wait for something, use wait_event(). 4346 * If you want to use yield() to be 'nice' for others, use cond_resched(). 4347 * If you still want to use yield(), do not! 4348 */ 4349void __sched yield(void) 4350{ 4351 set_current_state(TASK_RUNNING); 4352 sys_sched_yield(); 4353} 4354EXPORT_SYMBOL(yield); 4355 4356/** 4357 * yield_to - yield the current processor to another thread in 4358 * your thread group, or accelerate that thread toward the 4359 * processor it's on. 4360 * @p: target task 4361 * @preempt: whether task preemption is allowed or not 4362 * 4363 * It's the caller's job to ensure that the target task struct 4364 * can't go away on us before we can do any checks. 4365 * 4366 * Returns true if we indeed boosted the target task. 4367 */ 4368bool __sched yield_to(struct task_struct *p, bool preempt) 4369{ 4370 struct task_struct *curr = current; 4371 struct rq *rq, *p_rq; 4372 unsigned long flags; 4373 bool yielded = 0; 4374 4375 local_irq_save(flags); 4376 rq = this_rq(); 4377 4378again: 4379 p_rq = task_rq(p); 4380 double_rq_lock(rq, p_rq); 4381 while (task_rq(p) != p_rq) { 4382 double_rq_unlock(rq, p_rq); 4383 goto again; 4384 } 4385 4386 if (!curr->sched_class->yield_to_task) 4387 goto out; 4388 4389 if (curr->sched_class != p->sched_class) 4390 goto out; 4391 4392 if (task_running(p_rq, p) || p->state) 4393 goto out; 4394 4395 yielded = curr->sched_class->yield_to_task(rq, p, preempt); 4396 if (yielded) { 4397 schedstat_inc(rq, yld_count); 4398 /* 4399 * Make p's CPU reschedule; pick_next_entity takes care of 4400 * fairness. 4401 */ 4402 if (preempt && rq != p_rq) 4403 resched_task(p_rq->curr); 4404 } 4405 4406out: 4407 double_rq_unlock(rq, p_rq); 4408 local_irq_restore(flags); 4409 4410 if (yielded) 4411 schedule(); 4412 4413 return yielded; 4414} 4415EXPORT_SYMBOL_GPL(yield_to); 4416 4417/* 4418 * This task is about to go to sleep on IO. Increment rq->nr_iowait so 4419 * that process accounting knows that this is a task in IO wait state. 4420 */ 4421void __sched io_schedule(void) 4422{ 4423 struct rq *rq = raw_rq(); 4424 4425 delayacct_blkio_start(); 4426 atomic_inc(&rq->nr_iowait); 4427 blk_flush_plug(current); 4428 current->in_iowait = 1; 4429 schedule(); 4430 current->in_iowait = 0; 4431 atomic_dec(&rq->nr_iowait); 4432 delayacct_blkio_end(); 4433} 4434EXPORT_SYMBOL(io_schedule); 4435 4436long __sched io_schedule_timeout(long timeout) 4437{ 4438 struct rq *rq = raw_rq(); 4439 long ret; 4440 4441 delayacct_blkio_start(); 4442 atomic_inc(&rq->nr_iowait); 4443 blk_flush_plug(current); 4444 current->in_iowait = 1; 4445 ret = schedule_timeout(timeout); 4446 current->in_iowait = 0; 4447 atomic_dec(&rq->nr_iowait); 4448 delayacct_blkio_end(); 4449 return ret; 4450} 4451 4452/** 4453 * sys_sched_get_priority_max - return maximum RT priority. 4454 * @policy: scheduling class. 4455 * 4456 * this syscall returns the maximum rt_priority that can be used 4457 * by a given scheduling class. 4458 */ 4459SYSCALL_DEFINE1(sched_get_priority_max, int, policy) 4460{ 4461 int ret = -EINVAL; 4462 4463 switch (policy) { 4464 case SCHED_FIFO: 4465 case SCHED_RR: 4466 ret = MAX_USER_RT_PRIO-1; 4467 break; 4468 case SCHED_NORMAL: 4469 case SCHED_BATCH: 4470 case SCHED_IDLE: 4471 ret = 0; 4472 break; 4473 } 4474 return ret; 4475} 4476 4477/** 4478 * sys_sched_get_priority_min - return minimum RT priority. 4479 * @policy: scheduling class. 4480 * 4481 * this syscall returns the minimum rt_priority that can be used 4482 * by a given scheduling class. 4483 */ 4484SYSCALL_DEFINE1(sched_get_priority_min, int, policy) 4485{ 4486 int ret = -EINVAL; 4487 4488 switch (policy) { 4489 case SCHED_FIFO: 4490 case SCHED_RR: 4491 ret = 1; 4492 break; 4493 case SCHED_NORMAL: 4494 case SCHED_BATCH: 4495 case SCHED_IDLE: 4496 ret = 0; 4497 } 4498 return ret; 4499} 4500 4501/** 4502 * sys_sched_rr_get_interval - return the default timeslice of a process. 4503 * @pid: pid of the process. 4504 * @interval: userspace pointer to the timeslice value. 4505 * 4506 * this syscall writes the default timeslice value of a given process 4507 * into the user-space timespec buffer. A value of '0' means infinity. 4508 */ 4509SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid, 4510 struct timespec __user *, interval) 4511{ 4512 struct task_struct *p; 4513 unsigned int time_slice; 4514 unsigned long flags; 4515 struct rq *rq; 4516 int retval; 4517 struct timespec t; 4518 4519 if (pid < 0) 4520 return -EINVAL; 4521 4522 retval = -ESRCH; 4523 rcu_read_lock(); 4524 p = find_process_by_pid(pid); 4525 if (!p) 4526 goto out_unlock; 4527 4528 retval = security_task_getscheduler(p); 4529 if (retval) 4530 goto out_unlock; 4531 4532 rq = task_rq_lock(p, &flags); 4533 time_slice = p->sched_class->get_rr_interval(rq, p); 4534 task_rq_unlock(rq, p, &flags); 4535 4536 rcu_read_unlock(); 4537 jiffies_to_timespec(time_slice, &t); 4538 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0; 4539 return retval; 4540 4541out_unlock: 4542 rcu_read_unlock(); 4543 return retval; 4544} 4545 4546static const char stat_nam[] = TASK_STATE_TO_CHAR_STR; 4547 4548void sched_show_task(struct task_struct *p) 4549{ 4550 unsigned long free = 0; 4551 int ppid; 4552 unsigned state; 4553 4554 state = p->state ? __ffs(p->state) + 1 : 0; 4555 printk(KERN_INFO "%-15.15s %c", p->comm, 4556 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?'); 4557#if BITS_PER_LONG == 32 4558 if (state == TASK_RUNNING) 4559 printk(KERN_CONT " running "); 4560 else 4561 printk(KERN_CONT " %08lx ", thread_saved_pc(p)); 4562#else 4563 if (state == TASK_RUNNING) 4564 printk(KERN_CONT " running task "); 4565 else 4566 printk(KERN_CONT " %016lx ", thread_saved_pc(p)); 4567#endif 4568#ifdef CONFIG_DEBUG_STACK_USAGE 4569 free = stack_not_used(p); 4570#endif 4571 rcu_read_lock(); 4572 ppid = task_pid_nr(rcu_dereference(p->real_parent)); 4573 rcu_read_unlock(); 4574 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free, 4575 task_pid_nr(p), ppid, 4576 (unsigned long)task_thread_info(p)->flags); 4577 4578 show_stack(p, NULL); 4579} 4580 4581void show_state_filter(unsigned long state_filter) 4582{ 4583 struct task_struct *g, *p; 4584 4585#if BITS_PER_LONG == 32 4586 printk(KERN_INFO 4587 " task PC stack pid father\n"); 4588#else 4589 printk(KERN_INFO 4590 " task PC stack pid father\n"); 4591#endif 4592 rcu_read_lock(); 4593 do_each_thread(g, p) { 4594 /* 4595 * reset the NMI-timeout, listing all files on a slow 4596 * console might take a lot of time: 4597 */ 4598 touch_nmi_watchdog(); 4599 if (!state_filter || (p->state & state_filter)) 4600 sched_show_task(p); 4601 } while_each_thread(g, p); 4602 4603 touch_all_softlockup_watchdogs(); 4604 4605#ifdef CONFIG_SCHED_DEBUG 4606 sysrq_sched_debug_show(); 4607#endif 4608 rcu_read_unlock(); 4609 /* 4610 * Only show locks if all tasks are dumped: 4611 */ 4612 if (!state_filter) 4613 debug_show_all_locks(); 4614} 4615 4616void __cpuinit init_idle_bootup_task(struct task_struct *idle) 4617{ 4618 idle->sched_class = &idle_sched_class; 4619} 4620 4621/** 4622 * init_idle - set up an idle thread for a given CPU 4623 * @idle: task in question 4624 * @cpu: cpu the idle task belongs to 4625 * 4626 * NOTE: this function does not set the idle thread's NEED_RESCHED 4627 * flag, to make booting more robust. 4628 */ 4629void __cpuinit init_idle(struct task_struct *idle, int cpu) 4630{ 4631 struct rq *rq = cpu_rq(cpu); 4632 unsigned long flags; 4633 4634 raw_spin_lock_irqsave(&rq->lock, flags); 4635 4636 __sched_fork(idle); 4637 idle->state = TASK_RUNNING; 4638 idle->se.exec_start = sched_clock(); 4639 4640 do_set_cpus_allowed(idle, cpumask_of(cpu)); 4641 /* 4642 * We're having a chicken and egg problem, even though we are 4643 * holding rq->lock, the cpu isn't yet set to this cpu so the 4644 * lockdep check in task_group() will fail. 4645 * 4646 * Similar case to sched_fork(). / Alternatively we could 4647 * use task_rq_lock() here and obtain the other rq->lock. 4648 * 4649 * Silence PROVE_RCU 4650 */ 4651 rcu_read_lock(); 4652 __set_task_cpu(idle, cpu); 4653 rcu_read_unlock(); 4654 4655 rq->curr = rq->idle = idle; 4656#if defined(CONFIG_SMP) 4657 idle->on_cpu = 1; 4658#endif 4659 raw_spin_unlock_irqrestore(&rq->lock, flags); 4660 4661 /* Set the preempt count _outside_ the spinlocks! */ 4662 task_thread_info(idle)->preempt_count = 0; 4663 4664 /* 4665 * The idle tasks have their own, simple scheduling class: 4666 */ 4667 idle->sched_class = &idle_sched_class; 4668 ftrace_graph_init_idle_task(idle, cpu); 4669 vtime_init_idle(idle); 4670#if defined(CONFIG_SMP) 4671 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu); 4672#endif 4673} 4674 4675#ifdef CONFIG_SMP 4676void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask) 4677{ 4678 if (p->sched_class && p->sched_class->set_cpus_allowed) 4679 p->sched_class->set_cpus_allowed(p, new_mask); 4680 4681 cpumask_copy(&p->cpus_allowed, new_mask); 4682 p->nr_cpus_allowed = cpumask_weight(new_mask); 4683} 4684 4685/* 4686 * This is how migration works: 4687 * 4688 * 1) we invoke migration_cpu_stop() on the target CPU using 4689 * stop_one_cpu(). 4690 * 2) stopper starts to run (implicitly forcing the migrated thread 4691 * off the CPU) 4692 * 3) it checks whether the migrated task is still in the wrong runqueue. 4693 * 4) if it's in the wrong runqueue then the migration thread removes 4694 * it and puts it into the right queue. 4695 * 5) stopper completes and stop_one_cpu() returns and the migration 4696 * is done. 4697 */ 4698 4699/* 4700 * Change a given task's CPU affinity. Migrate the thread to a 4701 * proper CPU and schedule it away if the CPU it's executing on 4702 * is removed from the allowed bitmask. 4703 * 4704 * NOTE: the caller must have a valid reference to the task, the 4705 * task must not exit() & deallocate itself prematurely. The 4706 * call is not atomic; no spinlocks may be held. 4707 */ 4708int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask) 4709{ 4710 unsigned long flags; 4711 struct rq *rq; 4712 unsigned int dest_cpu; 4713 int ret = 0; 4714 4715 rq = task_rq_lock(p, &flags); 4716 4717 if (cpumask_equal(&p->cpus_allowed, new_mask)) 4718 goto out; 4719 4720 if (!cpumask_intersects(new_mask, cpu_active_mask)) { 4721 ret = -EINVAL; 4722 goto out; 4723 } 4724 4725 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current)) { 4726 ret = -EINVAL; 4727 goto out; 4728 } 4729 4730 do_set_cpus_allowed(p, new_mask); 4731 4732 /* Can the task run on the task's current CPU? If so, we're done */ 4733 if (cpumask_test_cpu(task_cpu(p), new_mask)) 4734 goto out; 4735 4736 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask); 4737 if (p->on_rq) { 4738 struct migration_arg arg = { p, dest_cpu }; 4739 /* Need help from migration thread: drop lock and wait. */ 4740 task_rq_unlock(rq, p, &flags); 4741 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg); 4742 tlb_migrate_finish(p->mm); 4743 return 0; 4744 } 4745out: 4746 task_rq_unlock(rq, p, &flags); 4747 4748 return ret; 4749} 4750EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr); 4751 4752/* 4753 * Move (not current) task off this cpu, onto dest cpu. We're doing 4754 * this because either it can't run here any more (set_cpus_allowed() 4755 * away from this CPU, or CPU going down), or because we're 4756 * attempting to rebalance this task on exec (sched_exec). 4757 * 4758 * So we race with normal scheduler movements, but that's OK, as long 4759 * as the task is no longer on this CPU. 4760 * 4761 * Returns non-zero if task was successfully migrated. 4762 */ 4763static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu) 4764{ 4765 struct rq *rq_dest, *rq_src; 4766 int ret = 0; 4767 4768 if (unlikely(!cpu_active(dest_cpu))) 4769 return ret; 4770 4771 rq_src = cpu_rq(src_cpu); 4772 rq_dest = cpu_rq(dest_cpu); 4773 4774 raw_spin_lock(&p->pi_lock); 4775 double_rq_lock(rq_src, rq_dest); 4776 /* Already moved. */ 4777 if (task_cpu(p) != src_cpu) 4778 goto done; 4779 /* Affinity changed (again). */ 4780 if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p))) 4781 goto fail; 4782 4783 /* 4784 * If we're not on a rq, the next wake-up will ensure we're 4785 * placed properly. 4786 */ 4787 if (p->on_rq) { 4788 dequeue_task(rq_src, p, 0); 4789 set_task_cpu(p, dest_cpu); 4790 enqueue_task(rq_dest, p, 0); 4791 check_preempt_curr(rq_dest, p, 0); 4792 } 4793done: 4794 ret = 1; 4795fail: 4796 double_rq_unlock(rq_src, rq_dest); 4797 raw_spin_unlock(&p->pi_lock); 4798 return ret; 4799} 4800 4801/* 4802 * migration_cpu_stop - this will be executed by a highprio stopper thread 4803 * and performs thread migration by bumping thread off CPU then 4804 * 'pushing' onto another runqueue. 4805 */ 4806static int migration_cpu_stop(void *data) 4807{ 4808 struct migration_arg *arg = data; 4809 4810 /* 4811 * The original target cpu might have gone down and we might 4812 * be on another cpu but it doesn't matter. 4813 */ 4814 local_irq_disable(); 4815 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu); 4816 local_irq_enable(); 4817 return 0; 4818} 4819 4820#ifdef CONFIG_HOTPLUG_CPU 4821 4822/* 4823 * Ensures that the idle task is using init_mm right before its cpu goes 4824 * offline. 4825 */ 4826void idle_task_exit(void) 4827{ 4828 struct mm_struct *mm = current->active_mm; 4829 4830 BUG_ON(cpu_online(smp_processor_id())); 4831 4832 if (mm != &init_mm) 4833 switch_mm(mm, &init_mm, current); 4834 mmdrop(mm); 4835} 4836 4837/* 4838 * Since this CPU is going 'away' for a while, fold any nr_active delta 4839 * we might have. Assumes we're called after migrate_tasks() so that the 4840 * nr_active count is stable. 4841 * 4842 * Also see the comment "Global load-average calculations". 4843 */ 4844static void calc_load_migrate(struct rq *rq) 4845{ 4846 long delta = calc_load_fold_active(rq); 4847 if (delta) 4848 atomic_long_add(delta, &calc_load_tasks); 4849} 4850 4851/* 4852 * Migrate all tasks from the rq, sleeping tasks will be migrated by 4853 * try_to_wake_up()->select_task_rq(). 4854 * 4855 * Called with rq->lock held even though we'er in stop_machine() and 4856 * there's no concurrency possible, we hold the required locks anyway 4857 * because of lock validation efforts. 4858 */ 4859static void migrate_tasks(unsigned int dead_cpu) 4860{ 4861 struct rq *rq = cpu_rq(dead_cpu); 4862 struct task_struct *next, *stop = rq->stop; 4863 int dest_cpu; 4864 4865 /* 4866 * Fudge the rq selection such that the below task selection loop 4867 * doesn't get stuck on the currently eligible stop task. 4868 * 4869 * We're currently inside stop_machine() and the rq is either stuck 4870 * in the stop_machine_cpu_stop() loop, or we're executing this code, 4871 * either way we should never end up calling schedule() until we're 4872 * done here. 4873 */ 4874 rq->stop = NULL; 4875 4876 for ( ; ; ) { 4877 /* 4878 * There's this thread running, bail when that's the only 4879 * remaining thread. 4880 */ 4881 if (rq->nr_running == 1) 4882 break; 4883 4884 next = pick_next_task(rq); 4885 BUG_ON(!next); 4886 next->sched_class->put_prev_task(rq, next); 4887 4888 /* Find suitable destination for @next, with force if needed. */ 4889 dest_cpu = select_fallback_rq(dead_cpu, next); 4890 raw_spin_unlock(&rq->lock); 4891 4892 __migrate_task(next, dead_cpu, dest_cpu); 4893 4894 raw_spin_lock(&rq->lock); 4895 } 4896 4897 rq->stop = stop; 4898} 4899 4900#endif /* CONFIG_HOTPLUG_CPU */ 4901 4902#if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL) 4903 4904static struct ctl_table sd_ctl_dir[] = { 4905 { 4906 .procname = "sched_domain", 4907 .mode = 0555, 4908 }, 4909 {} 4910}; 4911 4912static struct ctl_table sd_ctl_root[] = { 4913 { 4914 .procname = "kernel", 4915 .mode = 0555, 4916 .child = sd_ctl_dir, 4917 }, 4918 {} 4919}; 4920 4921static struct ctl_table *sd_alloc_ctl_entry(int n) 4922{ 4923 struct ctl_table *entry = 4924 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL); 4925 4926 return entry; 4927} 4928 4929static void sd_free_ctl_entry(struct ctl_table **tablep) 4930{ 4931 struct ctl_table *entry; 4932 4933 /* 4934 * In the intermediate directories, both the child directory and 4935 * procname are dynamically allocated and could fail but the mode 4936 * will always be set. In the lowest directory the names are 4937 * static strings and all have proc handlers. 4938 */ 4939 for (entry = *tablep; entry->mode; entry++) { 4940 if (entry->child) 4941 sd_free_ctl_entry(&entry->child); 4942 if (entry->proc_handler == NULL) 4943 kfree(entry->procname); 4944 } 4945 4946 kfree(*tablep); 4947 *tablep = NULL; 4948} 4949 4950static int min_load_idx = 0; 4951static int max_load_idx = CPU_LOAD_IDX_MAX; 4952 4953static void 4954set_table_entry(struct ctl_table *entry, 4955 const char *procname, void *data, int maxlen, 4956 umode_t mode, proc_handler *proc_handler, 4957 bool load_idx) 4958{ 4959 entry->procname = procname; 4960 entry->data = data; 4961 entry->maxlen = maxlen; 4962 entry->mode = mode; 4963 entry->proc_handler = proc_handler; 4964 4965 if (load_idx) { 4966 entry->extra1 = &min_load_idx; 4967 entry->extra2 = &max_load_idx; 4968 } 4969} 4970 4971static struct ctl_table * 4972sd_alloc_ctl_domain_table(struct sched_domain *sd) 4973{ 4974 struct ctl_table *table = sd_alloc_ctl_entry(13); 4975 4976 if (table == NULL) 4977 return NULL; 4978 4979 set_table_entry(&table[0], "min_interval", &sd->min_interval, 4980 sizeof(long), 0644, proc_doulongvec_minmax, false); 4981 set_table_entry(&table[1], "max_interval", &sd->max_interval, 4982 sizeof(long), 0644, proc_doulongvec_minmax, false); 4983 set_table_entry(&table[2], "busy_idx", &sd->busy_idx, 4984 sizeof(int), 0644, proc_dointvec_minmax, true); 4985 set_table_entry(&table[3], "idle_idx", &sd->idle_idx, 4986 sizeof(int), 0644, proc_dointvec_minmax, true); 4987 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx, 4988 sizeof(int), 0644, proc_dointvec_minmax, true); 4989 set_table_entry(&table[5], "wake_idx", &sd->wake_idx, 4990 sizeof(int), 0644, proc_dointvec_minmax, true); 4991 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx, 4992 sizeof(int), 0644, proc_dointvec_minmax, true); 4993 set_table_entry(&table[7], "busy_factor", &sd->busy_factor, 4994 sizeof(int), 0644, proc_dointvec_minmax, false); 4995 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct, 4996 sizeof(int), 0644, proc_dointvec_minmax, false); 4997 set_table_entry(&table[9], "cache_nice_tries", 4998 &sd->cache_nice_tries, 4999 sizeof(int), 0644, proc_dointvec_minmax, false); 5000 set_table_entry(&table[10], "flags", &sd->flags, 5001 sizeof(int), 0644, proc_dointvec_minmax, false); 5002 set_table_entry(&table[11], "name", sd->name, 5003 CORENAME_MAX_SIZE, 0444, proc_dostring, false); 5004 /* &table[12] is terminator */ 5005 5006 return table; 5007} 5008 5009static ctl_table *sd_alloc_ctl_cpu_table(int cpu) 5010{ 5011 struct ctl_table *entry, *table; 5012 struct sched_domain *sd; 5013 int domain_num = 0, i; 5014 char buf[32]; 5015 5016 for_each_domain(cpu, sd) 5017 domain_num++; 5018 entry = table = sd_alloc_ctl_entry(domain_num + 1); 5019 if (table == NULL) 5020 return NULL; 5021 5022 i = 0; 5023 for_each_domain(cpu, sd) { 5024 snprintf(buf, 32, "domain%d", i); 5025 entry->procname = kstrdup(buf, GFP_KERNEL); 5026 entry->mode = 0555; 5027 entry->child = sd_alloc_ctl_domain_table(sd); 5028 entry++; 5029 i++; 5030 } 5031 return table; 5032} 5033 5034static struct ctl_table_header *sd_sysctl_header; 5035static void register_sched_domain_sysctl(void) 5036{ 5037 int i, cpu_num = num_possible_cpus(); 5038 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1); 5039 char buf[32]; 5040 5041 WARN_ON(sd_ctl_dir[0].child); 5042 sd_ctl_dir[0].child = entry; 5043 5044 if (entry == NULL) 5045 return; 5046 5047 for_each_possible_cpu(i) { 5048 snprintf(buf, 32, "cpu%d", i); 5049 entry->procname = kstrdup(buf, GFP_KERNEL); 5050 entry->mode = 0555; 5051 entry->child = sd_alloc_ctl_cpu_table(i); 5052 entry++; 5053 } 5054 5055 WARN_ON(sd_sysctl_header); 5056 sd_sysctl_header = register_sysctl_table(sd_ctl_root); 5057} 5058 5059/* may be called multiple times per register */ 5060static void unregister_sched_domain_sysctl(void) 5061{ 5062 if (sd_sysctl_header) 5063 unregister_sysctl_table(sd_sysctl_header); 5064 sd_sysctl_header = NULL; 5065 if (sd_ctl_dir[0].child) 5066 sd_free_ctl_entry(&sd_ctl_dir[0].child); 5067} 5068#else 5069static void register_sched_domain_sysctl(void) 5070{ 5071} 5072static void unregister_sched_domain_sysctl(void) 5073{ 5074} 5075#endif 5076 5077static void set_rq_online(struct rq *rq) 5078{ 5079 if (!rq->online) { 5080 const struct sched_class *class; 5081 5082 cpumask_set_cpu(rq->cpu, rq->rd->online); 5083 rq->online = 1; 5084 5085 for_each_class(class) { 5086 if (class->rq_online) 5087 class->rq_online(rq); 5088 } 5089 } 5090} 5091 5092static void set_rq_offline(struct rq *rq) 5093{ 5094 if (rq->online) { 5095 const struct sched_class *class; 5096 5097 for_each_class(class) { 5098 if (class->rq_offline) 5099 class->rq_offline(rq); 5100 } 5101 5102 cpumask_clear_cpu(rq->cpu, rq->rd->online); 5103 rq->online = 0; 5104 } 5105} 5106 5107/* 5108 * migration_call - callback that gets triggered when a CPU is added. 5109 * Here we can start up the necessary migration thread for the new CPU. 5110 */ 5111static int __cpuinit 5112migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu) 5113{ 5114 int cpu = (long)hcpu; 5115 unsigned long flags; 5116 struct rq *rq = cpu_rq(cpu); 5117 5118 switch (action & ~CPU_TASKS_FROZEN) { 5119 5120 case CPU_UP_PREPARE: 5121 rq->calc_load_update = calc_load_update; 5122 break; 5123 5124 case CPU_ONLINE: 5125 /* Update our root-domain */ 5126 raw_spin_lock_irqsave(&rq->lock, flags); 5127 if (rq->rd) { 5128 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span)); 5129 5130 set_rq_online(rq); 5131 } 5132 raw_spin_unlock_irqrestore(&rq->lock, flags); 5133 break; 5134 5135#ifdef CONFIG_HOTPLUG_CPU 5136 case CPU_DYING: 5137 sched_ttwu_pending(); 5138 /* Update our root-domain */ 5139 raw_spin_lock_irqsave(&rq->lock, flags); 5140 if (rq->rd) { 5141 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span)); 5142 set_rq_offline(rq); 5143 } 5144 migrate_tasks(cpu); 5145 BUG_ON(rq->nr_running != 1); /* the migration thread */ 5146 raw_spin_unlock_irqrestore(&rq->lock, flags); 5147 break; 5148 5149 case CPU_DEAD: 5150 calc_load_migrate(rq); 5151 break; 5152#endif 5153 } 5154 5155 update_max_interval(); 5156 5157 return NOTIFY_OK; 5158} 5159 5160/* 5161 * Register at high priority so that task migration (migrate_all_tasks) 5162 * happens before everything else. This has to be lower priority than 5163 * the notifier in the perf_event subsystem, though. 5164 */ 5165static struct notifier_block __cpuinitdata migration_notifier = { 5166 .notifier_call = migration_call, 5167 .priority = CPU_PRI_MIGRATION, 5168}; 5169 5170static int __cpuinit sched_cpu_active(struct notifier_block *nfb, 5171 unsigned long action, void *hcpu) 5172{ 5173 switch (action & ~CPU_TASKS_FROZEN) { 5174 case CPU_STARTING: 5175 case CPU_DOWN_FAILED: 5176 set_cpu_active((long)hcpu, true); 5177 return NOTIFY_OK; 5178 default: 5179 return NOTIFY_DONE; 5180 } 5181} 5182 5183static int __cpuinit sched_cpu_inactive(struct notifier_block *nfb, 5184 unsigned long action, void *hcpu) 5185{ 5186 switch (action & ~CPU_TASKS_FROZEN) { 5187 case CPU_DOWN_PREPARE: 5188 set_cpu_active((long)hcpu, false); 5189 return NOTIFY_OK; 5190 default: 5191 return NOTIFY_DONE; 5192 } 5193} 5194 5195static int __init migration_init(void) 5196{ 5197 void *cpu = (void *)(long)smp_processor_id(); 5198 int err; 5199 5200 /* Initialize migration for the boot CPU */ 5201 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu); 5202 BUG_ON(err == NOTIFY_BAD); 5203 migration_call(&migration_notifier, CPU_ONLINE, cpu); 5204 register_cpu_notifier(&migration_notifier); 5205 5206 /* Register cpu active notifiers */ 5207 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE); 5208 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE); 5209 5210 return 0; 5211} 5212early_initcall(migration_init); 5213#endif 5214 5215#ifdef CONFIG_SMP 5216 5217static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */ 5218 5219#ifdef CONFIG_SCHED_DEBUG 5220 5221static __read_mostly int sched_debug_enabled; 5222 5223static int __init sched_debug_setup(char *str) 5224{ 5225 sched_debug_enabled = 1; 5226 5227 return 0; 5228} 5229early_param("sched_debug", sched_debug_setup); 5230 5231static inline bool sched_debug(void) 5232{ 5233 return sched_debug_enabled; 5234} 5235 5236static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level, 5237 struct cpumask *groupmask) 5238{ 5239 struct sched_group *group = sd->groups; 5240 char str[256]; 5241 5242 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd)); 5243 cpumask_clear(groupmask); 5244 5245 printk(KERN_DEBUG "%*s domain %d: ", level, "", level); 5246 5247 if (!(sd->flags & SD_LOAD_BALANCE)) { 5248 printk("does not load-balance\n"); 5249 if (sd->parent) 5250 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain" 5251 " has parent"); 5252 return -1; 5253 } 5254 5255 printk(KERN_CONT "span %s level %s\n", str, sd->name); 5256 5257 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) { 5258 printk(KERN_ERR "ERROR: domain->span does not contain " 5259 "CPU%d\n", cpu); 5260 } 5261 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) { 5262 printk(KERN_ERR "ERROR: domain->groups does not contain" 5263 " CPU%d\n", cpu); 5264 } 5265 5266 printk(KERN_DEBUG "%*s groups:", level + 1, ""); 5267 do { 5268 if (!group) { 5269 printk("\n"); 5270 printk(KERN_ERR "ERROR: group is NULL\n"); 5271 break; 5272 } 5273 5274 /* 5275 * Even though we initialize ->power to something semi-sane, 5276 * we leave power_orig unset. This allows us to detect if 5277 * domain iteration is still funny without causing /0 traps. 5278 */ 5279 if (!group->sgp->power_orig) { 5280 printk(KERN_CONT "\n"); 5281 printk(KERN_ERR "ERROR: domain->cpu_power not " 5282 "set\n"); 5283 break; 5284 } 5285 5286 if (!cpumask_weight(sched_group_cpus(group))) { 5287 printk(KERN_CONT "\n"); 5288 printk(KERN_ERR "ERROR: empty group\n"); 5289 break; 5290 } 5291 5292 if (!(sd->flags & SD_OVERLAP) && 5293 cpumask_intersects(groupmask, sched_group_cpus(group))) { 5294 printk(KERN_CONT "\n"); 5295 printk(KERN_ERR "ERROR: repeated CPUs\n"); 5296 break; 5297 } 5298 5299 cpumask_or(groupmask, groupmask, sched_group_cpus(group)); 5300 5301 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group)); 5302 5303 printk(KERN_CONT " %s", str); 5304 if (group->sgp->power != SCHED_POWER_SCALE) { 5305 printk(KERN_CONT " (cpu_power = %d)", 5306 group->sgp->power); 5307 } 5308 5309 group = group->next; 5310 } while (group != sd->groups); 5311 printk(KERN_CONT "\n"); 5312 5313 if (!cpumask_equal(sched_domain_span(sd), groupmask)) 5314 printk(KERN_ERR "ERROR: groups don't span domain->span\n"); 5315 5316 if (sd->parent && 5317 !cpumask_subset(groupmask, sched_domain_span(sd->parent))) 5318 printk(KERN_ERR "ERROR: parent span is not a superset " 5319 "of domain->span\n"); 5320 return 0; 5321} 5322 5323static void sched_domain_debug(struct sched_domain *sd, int cpu) 5324{ 5325 int level = 0; 5326 5327 if (!sched_debug_enabled) 5328 return; 5329 5330 if (!sd) { 5331 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu); 5332 return; 5333 } 5334 5335 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu); 5336 5337 for (;;) { 5338 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask)) 5339 break; 5340 level++; 5341 sd = sd->parent; 5342 if (!sd) 5343 break; 5344 } 5345} 5346#else /* !CONFIG_SCHED_DEBUG */ 5347# define sched_domain_debug(sd, cpu) do { } while (0) 5348static inline bool sched_debug(void) 5349{ 5350 return false; 5351} 5352#endif /* CONFIG_SCHED_DEBUG */ 5353 5354static int sd_degenerate(struct sched_domain *sd) 5355{ 5356 if (cpumask_weight(sched_domain_span(sd)) == 1) 5357 return 1; 5358 5359 /* Following flags need at least 2 groups */ 5360 if (sd->flags & (SD_LOAD_BALANCE | 5361 SD_BALANCE_NEWIDLE | 5362 SD_BALANCE_FORK | 5363 SD_BALANCE_EXEC | 5364 SD_SHARE_CPUPOWER | 5365 SD_SHARE_PKG_RESOURCES)) { 5366 if (sd->groups != sd->groups->next) 5367 return 0; 5368 } 5369 5370 /* Following flags don't use groups */ 5371 if (sd->flags & (SD_WAKE_AFFINE)) 5372 return 0; 5373 5374 return 1; 5375} 5376 5377static int 5378sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent) 5379{ 5380 unsigned long cflags = sd->flags, pflags = parent->flags; 5381 5382 if (sd_degenerate(parent)) 5383 return 1; 5384 5385 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent))) 5386 return 0; 5387 5388 /* Flags needing groups don't count if only 1 group in parent */ 5389 if (parent->groups == parent->groups->next) { 5390 pflags &= ~(SD_LOAD_BALANCE | 5391 SD_BALANCE_NEWIDLE | 5392 SD_BALANCE_FORK | 5393 SD_BALANCE_EXEC | 5394 SD_SHARE_CPUPOWER | 5395 SD_SHARE_PKG_RESOURCES); 5396 if (nr_node_ids == 1) 5397 pflags &= ~SD_SERIALIZE; 5398 } 5399 if (~cflags & pflags) 5400 return 0; 5401 5402 return 1; 5403} 5404 5405static void free_rootdomain(struct rcu_head *rcu) 5406{ 5407 struct root_domain *rd = container_of(rcu, struct root_domain, rcu); 5408 5409 cpupri_cleanup(&rd->cpupri); 5410 free_cpumask_var(rd->rto_mask); 5411 free_cpumask_var(rd->online); 5412 free_cpumask_var(rd->span); 5413 kfree(rd); 5414} 5415 5416static void rq_attach_root(struct rq *rq, struct root_domain *rd) 5417{ 5418 struct root_domain *old_rd = NULL; 5419 unsigned long flags; 5420 5421 raw_spin_lock_irqsave(&rq->lock, flags); 5422 5423 if (rq->rd) { 5424 old_rd = rq->rd; 5425 5426 if (cpumask_test_cpu(rq->cpu, old_rd->online)) 5427 set_rq_offline(rq); 5428 5429 cpumask_clear_cpu(rq->cpu, old_rd->span); 5430 5431 /* 5432 * If we dont want to free the old_rt yet then 5433 * set old_rd to NULL to skip the freeing later 5434 * in this function: 5435 */ 5436 if (!atomic_dec_and_test(&old_rd->refcount)) 5437 old_rd = NULL; 5438 } 5439 5440 atomic_inc(&rd->refcount); 5441 rq->rd = rd; 5442 5443 cpumask_set_cpu(rq->cpu, rd->span); 5444 if (cpumask_test_cpu(rq->cpu, cpu_active_mask)) 5445 set_rq_online(rq); 5446 5447 raw_spin_unlock_irqrestore(&rq->lock, flags); 5448 5449 if (old_rd) 5450 call_rcu_sched(&old_rd->rcu, free_rootdomain); 5451} 5452 5453static int init_rootdomain(struct root_domain *rd) 5454{ 5455 memset(rd, 0, sizeof(*rd)); 5456 5457 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL)) 5458 goto out; 5459 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL)) 5460 goto free_span; 5461 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL)) 5462 goto free_online; 5463 5464 if (cpupri_init(&rd->cpupri) != 0) 5465 goto free_rto_mask; 5466 return 0; 5467 5468free_rto_mask: 5469 free_cpumask_var(rd->rto_mask); 5470free_online: 5471 free_cpumask_var(rd->online); 5472free_span: 5473 free_cpumask_var(rd->span); 5474out: 5475 return -ENOMEM; 5476} 5477 5478/* 5479 * By default the system creates a single root-domain with all cpus as 5480 * members (mimicking the global state we have today). 5481 */ 5482struct root_domain def_root_domain; 5483 5484static void init_defrootdomain(void) 5485{ 5486 init_rootdomain(&def_root_domain); 5487 5488 atomic_set(&def_root_domain.refcount, 1); 5489} 5490 5491static struct root_domain *alloc_rootdomain(void) 5492{ 5493 struct root_domain *rd; 5494 5495 rd = kmalloc(sizeof(*rd), GFP_KERNEL); 5496 if (!rd) 5497 return NULL; 5498 5499 if (init_rootdomain(rd) != 0) { 5500 kfree(rd); 5501 return NULL; 5502 } 5503 5504 return rd; 5505} 5506 5507static void free_sched_groups(struct sched_group *sg, int free_sgp) 5508{ 5509 struct sched_group *tmp, *first; 5510 5511 if (!sg) 5512 return; 5513 5514 first = sg; 5515 do { 5516 tmp = sg->next; 5517 5518 if (free_sgp && atomic_dec_and_test(&sg->sgp->ref)) 5519 kfree(sg->sgp); 5520 5521 kfree(sg); 5522 sg = tmp; 5523 } while (sg != first); 5524} 5525 5526static void free_sched_domain(struct rcu_head *rcu) 5527{ 5528 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu); 5529 5530 /* 5531 * If its an overlapping domain it has private groups, iterate and 5532 * nuke them all. 5533 */ 5534 if (sd->flags & SD_OVERLAP) { 5535 free_sched_groups(sd->groups, 1); 5536 } else if (atomic_dec_and_test(&sd->groups->ref)) { 5537 kfree(sd->groups->sgp); 5538 kfree(sd->groups); 5539 } 5540 kfree(sd); 5541} 5542 5543static void destroy_sched_domain(struct sched_domain *sd, int cpu) 5544{ 5545 call_rcu(&sd->rcu, free_sched_domain); 5546} 5547 5548static void destroy_sched_domains(struct sched_domain *sd, int cpu) 5549{ 5550 for (; sd; sd = sd->parent) 5551 destroy_sched_domain(sd, cpu); 5552} 5553 5554/* 5555 * Keep a special pointer to the highest sched_domain that has 5556 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this 5557 * allows us to avoid some pointer chasing select_idle_sibling(). 5558 * 5559 * Also keep a unique ID per domain (we use the first cpu number in 5560 * the cpumask of the domain), this allows us to quickly tell if 5561 * two cpus are in the same cache domain, see cpus_share_cache(). 5562 */ 5563DEFINE_PER_CPU(struct sched_domain *, sd_llc); 5564DEFINE_PER_CPU(int, sd_llc_id); 5565 5566static void update_top_cache_domain(int cpu) 5567{ 5568 struct sched_domain *sd; 5569 int id = cpu; 5570 5571 sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES); 5572 if (sd) 5573 id = cpumask_first(sched_domain_span(sd)); 5574 5575 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd); 5576 per_cpu(sd_llc_id, cpu) = id; 5577} 5578 5579/* 5580 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must 5581 * hold the hotplug lock. 5582 */ 5583static void 5584cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu) 5585{ 5586 struct rq *rq = cpu_rq(cpu); 5587 struct sched_domain *tmp; 5588 5589 /* Remove the sched domains which do not contribute to scheduling. */ 5590 for (tmp = sd; tmp; ) { 5591 struct sched_domain *parent = tmp->parent; 5592 if (!parent) 5593 break; 5594 5595 if (sd_parent_degenerate(tmp, parent)) { 5596 tmp->parent = parent->parent; 5597 if (parent->parent) 5598 parent->parent->child = tmp; 5599 destroy_sched_domain(parent, cpu); 5600 } else 5601 tmp = tmp->parent; 5602 } 5603 5604 if (sd && sd_degenerate(sd)) { 5605 tmp = sd; 5606 sd = sd->parent; 5607 destroy_sched_domain(tmp, cpu); 5608 if (sd) 5609 sd->child = NULL; 5610 } 5611 5612 sched_domain_debug(sd, cpu); 5613 5614 rq_attach_root(rq, rd); 5615 tmp = rq->sd; 5616 rcu_assign_pointer(rq->sd, sd); 5617 destroy_sched_domains(tmp, cpu); 5618 5619 update_top_cache_domain(cpu); 5620} 5621 5622/* cpus with isolated domains */ 5623static cpumask_var_t cpu_isolated_map; 5624 5625/* Setup the mask of cpus configured for isolated domains */ 5626static int __init isolated_cpu_setup(char *str) 5627{ 5628 alloc_bootmem_cpumask_var(&cpu_isolated_map); 5629 cpulist_parse(str, cpu_isolated_map); 5630 return 1; 5631} 5632 5633__setup("isolcpus=", isolated_cpu_setup); 5634 5635static const struct cpumask *cpu_cpu_mask(int cpu) 5636{ 5637 return cpumask_of_node(cpu_to_node(cpu)); 5638} 5639 5640struct sd_data { 5641 struct sched_domain **__percpu sd; 5642 struct sched_group **__percpu sg; 5643 struct sched_group_power **__percpu sgp; 5644}; 5645 5646struct s_data { 5647 struct sched_domain ** __percpu sd; 5648 struct root_domain *rd; 5649}; 5650 5651enum s_alloc { 5652 sa_rootdomain, 5653 sa_sd, 5654 sa_sd_storage, 5655 sa_none, 5656}; 5657 5658struct sched_domain_topology_level; 5659 5660typedef struct sched_domain *(*sched_domain_init_f)(struct sched_domain_topology_level *tl, int cpu); 5661typedef const struct cpumask *(*sched_domain_mask_f)(int cpu); 5662 5663#define SDTL_OVERLAP 0x01 5664 5665struct sched_domain_topology_level { 5666 sched_domain_init_f init; 5667 sched_domain_mask_f mask; 5668 int flags; 5669 int numa_level; 5670 struct sd_data data; 5671}; 5672 5673/* 5674 * Build an iteration mask that can exclude certain CPUs from the upwards 5675 * domain traversal. 5676 * 5677 * Asymmetric node setups can result in situations where the domain tree is of 5678 * unequal depth, make sure to skip domains that already cover the entire 5679 * range. 5680 * 5681 * In that case build_sched_domains() will have terminated the iteration early 5682 * and our sibling sd spans will be empty. Domains should always include the 5683 * cpu they're built on, so check that. 5684 * 5685 */ 5686static void build_group_mask(struct sched_domain *sd, struct sched_group *sg) 5687{ 5688 const struct cpumask *span = sched_domain_span(sd); 5689 struct sd_data *sdd = sd->private; 5690 struct sched_domain *sibling; 5691 int i; 5692 5693 for_each_cpu(i, span) { 5694 sibling = *per_cpu_ptr(sdd->sd, i); 5695 if (!cpumask_test_cpu(i, sched_domain_span(sibling))) 5696 continue; 5697 5698 cpumask_set_cpu(i, sched_group_mask(sg)); 5699 } 5700} 5701 5702/* 5703 * Return the canonical balance cpu for this group, this is the first cpu 5704 * of this group that's also in the iteration mask. 5705 */ 5706int group_balance_cpu(struct sched_group *sg) 5707{ 5708 return cpumask_first_and(sched_group_cpus(sg), sched_group_mask(sg)); 5709} 5710 5711static int 5712build_overlap_sched_groups(struct sched_domain *sd, int cpu) 5713{ 5714 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg; 5715 const struct cpumask *span = sched_domain_span(sd); 5716 struct cpumask *covered = sched_domains_tmpmask; 5717 struct sd_data *sdd = sd->private; 5718 struct sched_domain *child; 5719 int i; 5720 5721 cpumask_clear(covered); 5722 5723 for_each_cpu(i, span) { 5724 struct cpumask *sg_span; 5725 5726 if (cpumask_test_cpu(i, covered)) 5727 continue; 5728 5729 child = *per_cpu_ptr(sdd->sd, i); 5730 5731 /* See the comment near build_group_mask(). */ 5732 if (!cpumask_test_cpu(i, sched_domain_span(child))) 5733 continue; 5734 5735 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(), 5736 GFP_KERNEL, cpu_to_node(cpu)); 5737 5738 if (!sg) 5739 goto fail; 5740 5741 sg_span = sched_group_cpus(sg); 5742 if (child->child) { 5743 child = child->child; 5744 cpumask_copy(sg_span, sched_domain_span(child)); 5745 } else 5746 cpumask_set_cpu(i, sg_span); 5747 5748 cpumask_or(covered, covered, sg_span); 5749 5750 sg->sgp = *per_cpu_ptr(sdd->sgp, i); 5751 if (atomic_inc_return(&sg->sgp->ref) == 1) 5752 build_group_mask(sd, sg); 5753 5754 /* 5755 * Initialize sgp->power such that even if we mess up the 5756 * domains and no possible iteration will get us here, we won't 5757 * die on a /0 trap. 5758 */ 5759 sg->sgp->power = SCHED_POWER_SCALE * cpumask_weight(sg_span); 5760 5761 /* 5762 * Make sure the first group of this domain contains the 5763 * canonical balance cpu. Otherwise the sched_domain iteration 5764 * breaks. See update_sg_lb_stats(). 5765 */ 5766 if ((!groups && cpumask_test_cpu(cpu, sg_span)) || 5767 group_balance_cpu(sg) == cpu) 5768 groups = sg; 5769 5770 if (!first) 5771 first = sg; 5772 if (last) 5773 last->next = sg; 5774 last = sg; 5775 last->next = first; 5776 } 5777 sd->groups = groups; 5778 5779 return 0; 5780 5781fail: 5782 free_sched_groups(first, 0); 5783 5784 return -ENOMEM; 5785} 5786 5787static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg) 5788{ 5789 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu); 5790 struct sched_domain *child = sd->child; 5791 5792 if (child) 5793 cpu = cpumask_first(sched_domain_span(child)); 5794 5795 if (sg) { 5796 *sg = *per_cpu_ptr(sdd->sg, cpu); 5797 (*sg)->sgp = *per_cpu_ptr(sdd->sgp, cpu); 5798 atomic_set(&(*sg)->sgp->ref, 1); /* for claim_allocations */ 5799 } 5800 5801 return cpu; 5802} 5803 5804/* 5805 * build_sched_groups will build a circular linked list of the groups 5806 * covered by the given span, and will set each group's ->cpumask correctly, 5807 * and ->cpu_power to 0. 5808 * 5809 * Assumes the sched_domain tree is fully constructed 5810 */ 5811static int 5812build_sched_groups(struct sched_domain *sd, int cpu) 5813{ 5814 struct sched_group *first = NULL, *last = NULL; 5815 struct sd_data *sdd = sd->private; 5816 const struct cpumask *span = sched_domain_span(sd); 5817 struct cpumask *covered; 5818 int i; 5819 5820 get_group(cpu, sdd, &sd->groups); 5821 atomic_inc(&sd->groups->ref); 5822 5823 if (cpu != cpumask_first(sched_domain_span(sd))) 5824 return 0; 5825 5826 lockdep_assert_held(&sched_domains_mutex); 5827 covered = sched_domains_tmpmask; 5828 5829 cpumask_clear(covered); 5830 5831 for_each_cpu(i, span) { 5832 struct sched_group *sg; 5833 int group = get_group(i, sdd, &sg); 5834 int j; 5835 5836 if (cpumask_test_cpu(i, covered)) 5837 continue; 5838 5839 cpumask_clear(sched_group_cpus(sg)); 5840 sg->sgp->power = 0; 5841 cpumask_setall(sched_group_mask(sg)); 5842 5843 for_each_cpu(j, span) { 5844 if (get_group(j, sdd, NULL) != group) 5845 continue; 5846 5847 cpumask_set_cpu(j, covered); 5848 cpumask_set_cpu(j, sched_group_cpus(sg)); 5849 } 5850 5851 if (!first) 5852 first = sg; 5853 if (last) 5854 last->next = sg; 5855 last = sg; 5856 } 5857 last->next = first; 5858 5859 return 0; 5860} 5861 5862/* 5863 * Initialize sched groups cpu_power. 5864 * 5865 * cpu_power indicates the capacity of sched group, which is used while 5866 * distributing the load between different sched groups in a sched domain. 5867 * Typically cpu_power for all the groups in a sched domain will be same unless 5868 * there are asymmetries in the topology. If there are asymmetries, group 5869 * having more cpu_power will pickup more load compared to the group having 5870 * less cpu_power. 5871 */ 5872static void init_sched_groups_power(int cpu, struct sched_domain *sd) 5873{ 5874 struct sched_group *sg = sd->groups; 5875 5876 WARN_ON(!sd || !sg); 5877 5878 do { 5879 sg->group_weight = cpumask_weight(sched_group_cpus(sg)); 5880 sg = sg->next; 5881 } while (sg != sd->groups); 5882 5883 if (cpu != group_balance_cpu(sg)) 5884 return; 5885 5886 update_group_power(sd, cpu); 5887 atomic_set(&sg->sgp->nr_busy_cpus, sg->group_weight); 5888} 5889 5890int __weak arch_sd_sibling_asym_packing(void) 5891{ 5892 return 0*SD_ASYM_PACKING; 5893} 5894 5895/* 5896 * Initializers for schedule domains 5897 * Non-inlined to reduce accumulated stack pressure in build_sched_domains() 5898 */ 5899 5900#ifdef CONFIG_SCHED_DEBUG 5901# define SD_INIT_NAME(sd, type) sd->name = #type 5902#else 5903# define SD_INIT_NAME(sd, type) do { } while (0) 5904#endif 5905 5906#define SD_INIT_FUNC(type) \ 5907static noinline struct sched_domain * \ 5908sd_init_##type(struct sched_domain_topology_level *tl, int cpu) \ 5909{ \ 5910 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); \ 5911 *sd = SD_##type##_INIT; \ 5912 SD_INIT_NAME(sd, type); \ 5913 sd->private = &tl->data; \ 5914 return sd; \ 5915} 5916 5917SD_INIT_FUNC(CPU) 5918#ifdef CONFIG_SCHED_SMT 5919 SD_INIT_FUNC(SIBLING) 5920#endif 5921#ifdef CONFIG_SCHED_MC 5922 SD_INIT_FUNC(MC) 5923#endif 5924#ifdef CONFIG_SCHED_BOOK 5925 SD_INIT_FUNC(BOOK) 5926#endif 5927 5928static int default_relax_domain_level = -1; 5929int sched_domain_level_max; 5930 5931static int __init setup_relax_domain_level(char *str) 5932{ 5933 if (kstrtoint(str, 0, &default_relax_domain_level)) 5934 pr_warn("Unable to set relax_domain_level\n"); 5935 5936 return 1; 5937} 5938__setup("relax_domain_level=", setup_relax_domain_level); 5939 5940static void set_domain_attribute(struct sched_domain *sd, 5941 struct sched_domain_attr *attr) 5942{ 5943 int request; 5944 5945 if (!attr || attr->relax_domain_level < 0) { 5946 if (default_relax_domain_level < 0) 5947 return; 5948 else 5949 request = default_relax_domain_level; 5950 } else 5951 request = attr->relax_domain_level; 5952 if (request < sd->level) { 5953 /* turn off idle balance on this domain */ 5954 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE); 5955 } else { 5956 /* turn on idle balance on this domain */ 5957 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE); 5958 } 5959} 5960 5961static void __sdt_free(const struct cpumask *cpu_map); 5962static int __sdt_alloc(const struct cpumask *cpu_map); 5963 5964static void __free_domain_allocs(struct s_data *d, enum s_alloc what, 5965 const struct cpumask *cpu_map) 5966{ 5967 switch (what) { 5968 case sa_rootdomain: 5969 if (!atomic_read(&d->rd->refcount)) 5970 free_rootdomain(&d->rd->rcu); /* fall through */ 5971 case sa_sd: 5972 free_percpu(d->sd); /* fall through */ 5973 case sa_sd_storage: 5974 __sdt_free(cpu_map); /* fall through */ 5975 case sa_none: 5976 break; 5977 } 5978} 5979 5980static enum s_alloc __visit_domain_allocation_hell(struct s_data *d, 5981 const struct cpumask *cpu_map) 5982{ 5983 memset(d, 0, sizeof(*d)); 5984 5985 if (__sdt_alloc(cpu_map)) 5986 return sa_sd_storage; 5987 d->sd = alloc_percpu(struct sched_domain *); 5988 if (!d->sd) 5989 return sa_sd_storage; 5990 d->rd = alloc_rootdomain(); 5991 if (!d->rd) 5992 return sa_sd; 5993 return sa_rootdomain; 5994} 5995 5996/* 5997 * NULL the sd_data elements we've used to build the sched_domain and 5998 * sched_group structure so that the subsequent __free_domain_allocs() 5999 * will not free the data we're using. 6000 */ 6001static void claim_allocations(int cpu, struct sched_domain *sd) 6002{ 6003 struct sd_data *sdd = sd->private; 6004 6005 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd); 6006 *per_cpu_ptr(sdd->sd, cpu) = NULL; 6007 6008 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref)) 6009 *per_cpu_ptr(sdd->sg, cpu) = NULL; 6010 6011 if (atomic_read(&(*per_cpu_ptr(sdd->sgp, cpu))->ref)) 6012 *per_cpu_ptr(sdd->sgp, cpu) = NULL; 6013} 6014 6015#ifdef CONFIG_SCHED_SMT 6016static const struct cpumask *cpu_smt_mask(int cpu) 6017{ 6018 return topology_thread_cpumask(cpu); 6019} 6020#endif 6021 6022/* 6023 * Topology list, bottom-up. 6024 */ 6025static struct sched_domain_topology_level default_topology[] = { 6026#ifdef CONFIG_SCHED_SMT 6027 { sd_init_SIBLING, cpu_smt_mask, }, 6028#endif 6029#ifdef CONFIG_SCHED_MC 6030 { sd_init_MC, cpu_coregroup_mask, }, 6031#endif 6032#ifdef CONFIG_SCHED_BOOK 6033 { sd_init_BOOK, cpu_book_mask, }, 6034#endif 6035 { sd_init_CPU, cpu_cpu_mask, }, 6036 { NULL, }, 6037}; 6038 6039static struct sched_domain_topology_level *sched_domain_topology = default_topology; 6040 6041#ifdef CONFIG_NUMA 6042 6043static int sched_domains_numa_levels; 6044static int *sched_domains_numa_distance; 6045static struct cpumask ***sched_domains_numa_masks; 6046static int sched_domains_curr_level; 6047 6048static inline int sd_local_flags(int level) 6049{ 6050 if (sched_domains_numa_distance[level] > RECLAIM_DISTANCE) 6051 return 0; 6052 6053 return SD_BALANCE_EXEC | SD_BALANCE_FORK | SD_WAKE_AFFINE; 6054} 6055 6056static struct sched_domain * 6057sd_numa_init(struct sched_domain_topology_level *tl, int cpu) 6058{ 6059 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); 6060 int level = tl->numa_level; 6061 int sd_weight = cpumask_weight( 6062 sched_domains_numa_masks[level][cpu_to_node(cpu)]); 6063 6064 *sd = (struct sched_domain){ 6065 .min_interval = sd_weight, 6066 .max_interval = 2*sd_weight, 6067 .busy_factor = 32, 6068 .imbalance_pct = 125, 6069 .cache_nice_tries = 2, 6070 .busy_idx = 3, 6071 .idle_idx = 2, 6072 .newidle_idx = 0, 6073 .wake_idx = 0, 6074 .forkexec_idx = 0, 6075 6076 .flags = 1*SD_LOAD_BALANCE 6077 | 1*SD_BALANCE_NEWIDLE 6078 | 0*SD_BALANCE_EXEC 6079 | 0*SD_BALANCE_FORK 6080 | 0*SD_BALANCE_WAKE 6081 | 0*SD_WAKE_AFFINE 6082 | 0*SD_SHARE_CPUPOWER 6083 | 0*SD_SHARE_PKG_RESOURCES 6084 | 1*SD_SERIALIZE 6085 | 0*SD_PREFER_SIBLING 6086 | sd_local_flags(level) 6087 , 6088 .last_balance = jiffies, 6089 .balance_interval = sd_weight, 6090 }; 6091 SD_INIT_NAME(sd, NUMA); 6092 sd->private = &tl->data; 6093 6094 /* 6095 * Ugly hack to pass state to sd_numa_mask()... 6096 */ 6097 sched_domains_curr_level = tl->numa_level; 6098 6099 return sd; 6100} 6101 6102static const struct cpumask *sd_numa_mask(int cpu) 6103{ 6104 return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)]; 6105} 6106 6107static void sched_numa_warn(const char *str) 6108{ 6109 static int done = false; 6110 int i,j; 6111 6112 if (done) 6113 return; 6114 6115 done = true; 6116 6117 printk(KERN_WARNING "ERROR: %s\n\n", str); 6118 6119 for (i = 0; i < nr_node_ids; i++) { 6120 printk(KERN_WARNING " "); 6121 for (j = 0; j < nr_node_ids; j++) 6122 printk(KERN_CONT "%02d ", node_distance(i,j)); 6123 printk(KERN_CONT "\n"); 6124 } 6125 printk(KERN_WARNING "\n"); 6126} 6127 6128static bool find_numa_distance(int distance) 6129{ 6130 int i; 6131 6132 if (distance == node_distance(0, 0)) 6133 return true; 6134 6135 for (i = 0; i < sched_domains_numa_levels; i++) { 6136 if (sched_domains_numa_distance[i] == distance) 6137 return true; 6138 } 6139 6140 return false; 6141} 6142 6143static void sched_init_numa(void) 6144{ 6145 int next_distance, curr_distance = node_distance(0, 0); 6146 struct sched_domain_topology_level *tl; 6147 int level = 0; 6148 int i, j, k; 6149 6150 sched_domains_numa_distance = kzalloc(sizeof(int) * nr_node_ids, GFP_KERNEL); 6151 if (!sched_domains_numa_distance) 6152 return; 6153 6154 /* 6155 * O(nr_nodes^2) deduplicating selection sort -- in order to find the 6156 * unique distances in the node_distance() table. 6157 * 6158 * Assumes node_distance(0,j) includes all distances in 6159 * node_distance(i,j) in order to avoid cubic time. 6160 */ 6161 next_distance = curr_distance; 6162 for (i = 0; i < nr_node_ids; i++) { 6163 for (j = 0; j < nr_node_ids; j++) { 6164 for (k = 0; k < nr_node_ids; k++) { 6165 int distance = node_distance(i, k); 6166 6167 if (distance > curr_distance && 6168 (distance < next_distance || 6169 next_distance == curr_distance)) 6170 next_distance = distance; 6171 6172 /* 6173 * While not a strong assumption it would be nice to know 6174 * about cases where if node A is connected to B, B is not 6175 * equally connected to A. 6176 */ 6177 if (sched_debug() && node_distance(k, i) != distance) 6178 sched_numa_warn("Node-distance not symmetric"); 6179 6180 if (sched_debug() && i && !find_numa_distance(distance)) 6181 sched_numa_warn("Node-0 not representative"); 6182 } 6183 if (next_distance != curr_distance) { 6184 sched_domains_numa_distance[level++] = next_distance; 6185 sched_domains_numa_levels = level; 6186 curr_distance = next_distance; 6187 } else break; 6188 } 6189 6190 /* 6191 * In case of sched_debug() we verify the above assumption. 6192 */ 6193 if (!sched_debug()) 6194 break; 6195 } 6196 /* 6197 * 'level' contains the number of unique distances, excluding the 6198 * identity distance node_distance(i,i). 6199 * 6200 * The sched_domains_nume_distance[] array includes the actual distance 6201 * numbers. 6202 */ 6203 6204 /* 6205 * Here, we should temporarily reset sched_domains_numa_levels to 0. 6206 * If it fails to allocate memory for array sched_domains_numa_masks[][], 6207 * the array will contain less then 'level' members. This could be 6208 * dangerous when we use it to iterate array sched_domains_numa_masks[][] 6209 * in other functions. 6210 * 6211 * We reset it to 'level' at the end of this function. 6212 */ 6213 sched_domains_numa_levels = 0; 6214 6215 sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL); 6216 if (!sched_domains_numa_masks) 6217 return; 6218 6219 /* 6220 * Now for each level, construct a mask per node which contains all 6221 * cpus of nodes that are that many hops away from us. 6222 */ 6223 for (i = 0; i < level; i++) { 6224 sched_domains_numa_masks[i] = 6225 kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL); 6226 if (!sched_domains_numa_masks[i]) 6227 return; 6228 6229 for (j = 0; j < nr_node_ids; j++) { 6230 struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL); 6231 if (!mask) 6232 return; 6233 6234 sched_domains_numa_masks[i][j] = mask; 6235 6236 for (k = 0; k < nr_node_ids; k++) { 6237 if (node_distance(j, k) > sched_domains_numa_distance[i]) 6238 continue; 6239 6240 cpumask_or(mask, mask, cpumask_of_node(k)); 6241 } 6242 } 6243 } 6244 6245 tl = kzalloc((ARRAY_SIZE(default_topology) + level) * 6246 sizeof(struct sched_domain_topology_level), GFP_KERNEL); 6247 if (!tl) 6248 return; 6249 6250 /* 6251 * Copy the default topology bits.. 6252 */ 6253 for (i = 0; default_topology[i].init; i++) 6254 tl[i] = default_topology[i]; 6255 6256 /* 6257 * .. and append 'j' levels of NUMA goodness. 6258 */ 6259 for (j = 0; j < level; i++, j++) { 6260 tl[i] = (struct sched_domain_topology_level){ 6261 .init = sd_numa_init, 6262 .mask = sd_numa_mask, 6263 .flags = SDTL_OVERLAP, 6264 .numa_level = j, 6265 }; 6266 } 6267 6268 sched_domain_topology = tl; 6269 6270 sched_domains_numa_levels = level; 6271} 6272 6273static void sched_domains_numa_masks_set(int cpu) 6274{ 6275 int i, j; 6276 int node = cpu_to_node(cpu); 6277 6278 for (i = 0; i < sched_domains_numa_levels; i++) { 6279 for (j = 0; j < nr_node_ids; j++) { 6280 if (node_distance(j, node) <= sched_domains_numa_distance[i]) 6281 cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]); 6282 } 6283 } 6284} 6285 6286static void sched_domains_numa_masks_clear(int cpu) 6287{ 6288 int i, j; 6289 for (i = 0; i < sched_domains_numa_levels; i++) { 6290 for (j = 0; j < nr_node_ids; j++) 6291 cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]); 6292 } 6293} 6294 6295/* 6296 * Update sched_domains_numa_masks[level][node] array when new cpus 6297 * are onlined. 6298 */ 6299static int sched_domains_numa_masks_update(struct notifier_block *nfb, 6300 unsigned long action, 6301 void *hcpu) 6302{ 6303 int cpu = (long)hcpu; 6304 6305 switch (action & ~CPU_TASKS_FROZEN) { 6306 case CPU_ONLINE: 6307 sched_domains_numa_masks_set(cpu); 6308 break; 6309 6310 case CPU_DEAD: 6311 sched_domains_numa_masks_clear(cpu); 6312 break; 6313 6314 default: 6315 return NOTIFY_DONE; 6316 } 6317 6318 return NOTIFY_OK; 6319} 6320#else 6321static inline void sched_init_numa(void) 6322{ 6323} 6324 6325static int sched_domains_numa_masks_update(struct notifier_block *nfb, 6326 unsigned long action, 6327 void *hcpu) 6328{ 6329 return 0; 6330} 6331#endif /* CONFIG_NUMA */ 6332 6333static int __sdt_alloc(const struct cpumask *cpu_map) 6334{ 6335 struct sched_domain_topology_level *tl; 6336 int j; 6337 6338 for (tl = sched_domain_topology; tl->init; tl++) { 6339 struct sd_data *sdd = &tl->data; 6340 6341 sdd->sd = alloc_percpu(struct sched_domain *); 6342 if (!sdd->sd) 6343 return -ENOMEM; 6344 6345 sdd->sg = alloc_percpu(struct sched_group *); 6346 if (!sdd->sg) 6347 return -ENOMEM; 6348 6349 sdd->sgp = alloc_percpu(struct sched_group_power *); 6350 if (!sdd->sgp) 6351 return -ENOMEM; 6352 6353 for_each_cpu(j, cpu_map) { 6354 struct sched_domain *sd; 6355 struct sched_group *sg; 6356 struct sched_group_power *sgp; 6357 6358 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(), 6359 GFP_KERNEL, cpu_to_node(j)); 6360 if (!sd) 6361 return -ENOMEM; 6362 6363 *per_cpu_ptr(sdd->sd, j) = sd; 6364 6365 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(), 6366 GFP_KERNEL, cpu_to_node(j)); 6367 if (!sg) 6368 return -ENOMEM; 6369 6370 sg->next = sg; 6371 6372 *per_cpu_ptr(sdd->sg, j) = sg; 6373 6374 sgp = kzalloc_node(sizeof(struct sched_group_power) + cpumask_size(), 6375 GFP_KERNEL, cpu_to_node(j)); 6376 if (!sgp) 6377 return -ENOMEM; 6378 6379 *per_cpu_ptr(sdd->sgp, j) = sgp; 6380 } 6381 } 6382 6383 return 0; 6384} 6385 6386static void __sdt_free(const struct cpumask *cpu_map) 6387{ 6388 struct sched_domain_topology_level *tl; 6389 int j; 6390 6391 for (tl = sched_domain_topology; tl->init; tl++) { 6392 struct sd_data *sdd = &tl->data; 6393 6394 for_each_cpu(j, cpu_map) { 6395 struct sched_domain *sd; 6396 6397 if (sdd->sd) { 6398 sd = *per_cpu_ptr(sdd->sd, j); 6399 if (sd && (sd->flags & SD_OVERLAP)) 6400 free_sched_groups(sd->groups, 0); 6401 kfree(*per_cpu_ptr(sdd->sd, j)); 6402 } 6403 6404 if (sdd->sg) 6405 kfree(*per_cpu_ptr(sdd->sg, j)); 6406 if (sdd->sgp) 6407 kfree(*per_cpu_ptr(sdd->sgp, j)); 6408 } 6409 free_percpu(sdd->sd); 6410 sdd->sd = NULL; 6411 free_percpu(sdd->sg); 6412 sdd->sg = NULL; 6413 free_percpu(sdd->sgp); 6414 sdd->sgp = NULL; 6415 } 6416} 6417 6418struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl, 6419 struct s_data *d, const struct cpumask *cpu_map, 6420 struct sched_domain_attr *attr, struct sched_domain *child, 6421 int cpu) 6422{ 6423 struct sched_domain *sd = tl->init(tl, cpu); 6424 if (!sd) 6425 return child; 6426 6427 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu)); 6428 if (child) { 6429 sd->level = child->level + 1; 6430 sched_domain_level_max = max(sched_domain_level_max, sd->level); 6431 child->parent = sd; 6432 } 6433 sd->child = child; 6434 set_domain_attribute(sd, attr); 6435 6436 return sd; 6437} 6438 6439/* 6440 * Build sched domains for a given set of cpus and attach the sched domains 6441 * to the individual cpus 6442 */ 6443static int build_sched_domains(const struct cpumask *cpu_map, 6444 struct sched_domain_attr *attr) 6445{ 6446 enum s_alloc alloc_state = sa_none; 6447 struct sched_domain *sd; 6448 struct s_data d; 6449 int i, ret = -ENOMEM; 6450 6451 alloc_state = __visit_domain_allocation_hell(&d, cpu_map); 6452 if (alloc_state != sa_rootdomain) 6453 goto error; 6454 6455 /* Set up domains for cpus specified by the cpu_map. */ 6456 for_each_cpu(i, cpu_map) { 6457 struct sched_domain_topology_level *tl; 6458 6459 sd = NULL; 6460 for (tl = sched_domain_topology; tl->init; tl++) { 6461 sd = build_sched_domain(tl, &d, cpu_map, attr, sd, i); 6462 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP)) 6463 sd->flags |= SD_OVERLAP; 6464 if (cpumask_equal(cpu_map, sched_domain_span(sd))) 6465 break; 6466 } 6467 6468 while (sd->child) 6469 sd = sd->child; 6470 6471 *per_cpu_ptr(d.sd, i) = sd; 6472 } 6473 6474 /* Build the groups for the domains */ 6475 for_each_cpu(i, cpu_map) { 6476 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) { 6477 sd->span_weight = cpumask_weight(sched_domain_span(sd)); 6478 if (sd->flags & SD_OVERLAP) { 6479 if (build_overlap_sched_groups(sd, i)) 6480 goto error; 6481 } else { 6482 if (build_sched_groups(sd, i)) 6483 goto error; 6484 } 6485 } 6486 } 6487 6488 /* Calculate CPU power for physical packages and nodes */ 6489 for (i = nr_cpumask_bits-1; i >= 0; i--) { 6490 if (!cpumask_test_cpu(i, cpu_map)) 6491 continue; 6492 6493 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) { 6494 claim_allocations(i, sd); 6495 init_sched_groups_power(i, sd); 6496 } 6497 } 6498 6499 /* Attach the domains */ 6500 rcu_read_lock(); 6501 for_each_cpu(i, cpu_map) { 6502 sd = *per_cpu_ptr(d.sd, i); 6503 cpu_attach_domain(sd, d.rd, i); 6504 } 6505 rcu_read_unlock(); 6506 6507 ret = 0; 6508error: 6509 __free_domain_allocs(&d, alloc_state, cpu_map); 6510 return ret; 6511} 6512 6513static cpumask_var_t *doms_cur; /* current sched domains */ 6514static int ndoms_cur; /* number of sched domains in 'doms_cur' */ 6515static struct sched_domain_attr *dattr_cur; 6516 /* attribues of custom domains in 'doms_cur' */ 6517 6518/* 6519 * Special case: If a kmalloc of a doms_cur partition (array of 6520 * cpumask) fails, then fallback to a single sched domain, 6521 * as determined by the single cpumask fallback_doms. 6522 */ 6523static cpumask_var_t fallback_doms; 6524 6525/* 6526 * arch_update_cpu_topology lets virtualized architectures update the 6527 * cpu core maps. It is supposed to return 1 if the topology changed 6528 * or 0 if it stayed the same. 6529 */ 6530int __attribute__((weak)) arch_update_cpu_topology(void) 6531{ 6532 return 0; 6533} 6534 6535cpumask_var_t *alloc_sched_domains(unsigned int ndoms) 6536{ 6537 int i; 6538 cpumask_var_t *doms; 6539 6540 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL); 6541 if (!doms) 6542 return NULL; 6543 for (i = 0; i < ndoms; i++) { 6544 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) { 6545 free_sched_domains(doms, i); 6546 return NULL; 6547 } 6548 } 6549 return doms; 6550} 6551 6552void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms) 6553{ 6554 unsigned int i; 6555 for (i = 0; i < ndoms; i++) 6556 free_cpumask_var(doms[i]); 6557 kfree(doms); 6558} 6559 6560/* 6561 * Set up scheduler domains and groups. Callers must hold the hotplug lock. 6562 * For now this just excludes isolated cpus, but could be used to 6563 * exclude other special cases in the future. 6564 */ 6565static int init_sched_domains(const struct cpumask *cpu_map) 6566{ 6567 int err; 6568 6569 arch_update_cpu_topology(); 6570 ndoms_cur = 1; 6571 doms_cur = alloc_sched_domains(ndoms_cur); 6572 if (!doms_cur) 6573 doms_cur = &fallback_doms; 6574 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map); 6575 err = build_sched_domains(doms_cur[0], NULL); 6576 register_sched_domain_sysctl(); 6577 6578 return err; 6579} 6580 6581/* 6582 * Detach sched domains from a group of cpus specified in cpu_map 6583 * These cpus will now be attached to the NULL domain 6584 */ 6585static void detach_destroy_domains(const struct cpumask *cpu_map) 6586{ 6587 int i; 6588 6589 rcu_read_lock(); 6590 for_each_cpu(i, cpu_map) 6591 cpu_attach_domain(NULL, &def_root_domain, i); 6592 rcu_read_unlock(); 6593} 6594 6595/* handle null as "default" */ 6596static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur, 6597 struct sched_domain_attr *new, int idx_new) 6598{ 6599 struct sched_domain_attr tmp; 6600 6601 /* fast path */ 6602 if (!new && !cur) 6603 return 1; 6604 6605 tmp = SD_ATTR_INIT; 6606 return !memcmp(cur ? (cur + idx_cur) : &tmp, 6607 new ? (new + idx_new) : &tmp, 6608 sizeof(struct sched_domain_attr)); 6609} 6610 6611/* 6612 * Partition sched domains as specified by the 'ndoms_new' 6613 * cpumasks in the array doms_new[] of cpumasks. This compares 6614 * doms_new[] to the current sched domain partitioning, doms_cur[]. 6615 * It destroys each deleted domain and builds each new domain. 6616 * 6617 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'. 6618 * The masks don't intersect (don't overlap.) We should setup one 6619 * sched domain for each mask. CPUs not in any of the cpumasks will 6620 * not be load balanced. If the same cpumask appears both in the 6621 * current 'doms_cur' domains and in the new 'doms_new', we can leave 6622 * it as it is. 6623 * 6624 * The passed in 'doms_new' should be allocated using 6625 * alloc_sched_domains. This routine takes ownership of it and will 6626 * free_sched_domains it when done with it. If the caller failed the 6627 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1, 6628 * and partition_sched_domains() will fallback to the single partition 6629 * 'fallback_doms', it also forces the domains to be rebuilt. 6630 * 6631 * If doms_new == NULL it will be replaced with cpu_online_mask. 6632 * ndoms_new == 0 is a special case for destroying existing domains, 6633 * and it will not create the default domain. 6634 * 6635 * Call with hotplug lock held 6636 */ 6637void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[], 6638 struct sched_domain_attr *dattr_new) 6639{ 6640 int i, j, n; 6641 int new_topology; 6642 6643 mutex_lock(&sched_domains_mutex); 6644 6645 /* always unregister in case we don't destroy any domains */ 6646 unregister_sched_domain_sysctl(); 6647 6648 /* Let architecture update cpu core mappings. */ 6649 new_topology = arch_update_cpu_topology(); 6650 6651 n = doms_new ? ndoms_new : 0; 6652 6653 /* Destroy deleted domains */ 6654 for (i = 0; i < ndoms_cur; i++) { 6655 for (j = 0; j < n && !new_topology; j++) { 6656 if (cpumask_equal(doms_cur[i], doms_new[j]) 6657 && dattrs_equal(dattr_cur, i, dattr_new, j)) 6658 goto match1; 6659 } 6660 /* no match - a current sched domain not in new doms_new[] */ 6661 detach_destroy_domains(doms_cur[i]); 6662match1: 6663 ; 6664 } 6665 6666 if (doms_new == NULL) { 6667 ndoms_cur = 0; 6668 doms_new = &fallback_doms; 6669 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map); 6670 WARN_ON_ONCE(dattr_new); 6671 } 6672 6673 /* Build new domains */ 6674 for (i = 0; i < ndoms_new; i++) { 6675 for (j = 0; j < ndoms_cur && !new_topology; j++) { 6676 if (cpumask_equal(doms_new[i], doms_cur[j]) 6677 && dattrs_equal(dattr_new, i, dattr_cur, j)) 6678 goto match2; 6679 } 6680 /* no match - add a new doms_new */ 6681 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL); 6682match2: 6683 ; 6684 } 6685 6686 /* Remember the new sched domains */ 6687 if (doms_cur != &fallback_doms) 6688 free_sched_domains(doms_cur, ndoms_cur); 6689 kfree(dattr_cur); /* kfree(NULL) is safe */ 6690 doms_cur = doms_new; 6691 dattr_cur = dattr_new; 6692 ndoms_cur = ndoms_new; 6693 6694 register_sched_domain_sysctl(); 6695 6696 mutex_unlock(&sched_domains_mutex); 6697} 6698 6699static int num_cpus_frozen; /* used to mark begin/end of suspend/resume */ 6700 6701/* 6702 * Update cpusets according to cpu_active mask. If cpusets are 6703 * disabled, cpuset_update_active_cpus() becomes a simple wrapper 6704 * around partition_sched_domains(). 6705 * 6706 * If we come here as part of a suspend/resume, don't touch cpusets because we 6707 * want to restore it back to its original state upon resume anyway. 6708 */ 6709static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action, 6710 void *hcpu) 6711{ 6712 switch (action) { 6713 case CPU_ONLINE_FROZEN: 6714 case CPU_DOWN_FAILED_FROZEN: 6715 6716 /* 6717 * num_cpus_frozen tracks how many CPUs are involved in suspend 6718 * resume sequence. As long as this is not the last online 6719 * operation in the resume sequence, just build a single sched 6720 * domain, ignoring cpusets. 6721 */ 6722 num_cpus_frozen--; 6723 if (likely(num_cpus_frozen)) { 6724 partition_sched_domains(1, NULL, NULL); 6725 break; 6726 } 6727 6728 /* 6729 * This is the last CPU online operation. So fall through and 6730 * restore the original sched domains by considering the 6731 * cpuset configurations. 6732 */ 6733 6734 case CPU_ONLINE: 6735 case CPU_DOWN_FAILED: 6736 cpuset_update_active_cpus(true); 6737 break; 6738 default: 6739 return NOTIFY_DONE; 6740 } 6741 return NOTIFY_OK; 6742} 6743 6744static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action, 6745 void *hcpu) 6746{ 6747 switch (action) { 6748 case CPU_DOWN_PREPARE: 6749 cpuset_update_active_cpus(false); 6750 break; 6751 case CPU_DOWN_PREPARE_FROZEN: 6752 num_cpus_frozen++; 6753 partition_sched_domains(1, NULL, NULL); 6754 break; 6755 default: 6756 return NOTIFY_DONE; 6757 } 6758 return NOTIFY_OK; 6759} 6760 6761void __init sched_init_smp(void) 6762{ 6763 cpumask_var_t non_isolated_cpus; 6764 6765 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL); 6766 alloc_cpumask_var(&fallback_doms, GFP_KERNEL); 6767 6768 sched_init_numa(); 6769 6770 get_online_cpus(); 6771 mutex_lock(&sched_domains_mutex); 6772 init_sched_domains(cpu_active_mask); 6773 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map); 6774 if (cpumask_empty(non_isolated_cpus)) 6775 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus); 6776 mutex_unlock(&sched_domains_mutex); 6777 put_online_cpus(); 6778 6779 hotcpu_notifier(sched_domains_numa_masks_update, CPU_PRI_SCHED_ACTIVE); 6780 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE); 6781 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE); 6782 6783 /* RT runtime code needs to handle some hotplug events */ 6784 hotcpu_notifier(update_runtime, 0); 6785 6786 init_hrtick(); 6787 6788 /* Move init over to a non-isolated CPU */ 6789 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0) 6790 BUG(); 6791 sched_init_granularity(); 6792 free_cpumask_var(non_isolated_cpus); 6793 6794 init_sched_rt_class(); 6795} 6796#else 6797void __init sched_init_smp(void) 6798{ 6799 sched_init_granularity(); 6800} 6801#endif /* CONFIG_SMP */ 6802 6803const_debug unsigned int sysctl_timer_migration = 1; 6804 6805int in_sched_functions(unsigned long addr) 6806{ 6807 return in_lock_functions(addr) || 6808 (addr >= (unsigned long)__sched_text_start 6809 && addr < (unsigned long)__sched_text_end); 6810} 6811 6812#ifdef CONFIG_CGROUP_SCHED 6813struct task_group root_task_group; 6814LIST_HEAD(task_groups); 6815#endif 6816 6817DECLARE_PER_CPU(cpumask_var_t, load_balance_tmpmask); 6818 6819void __init sched_init(void) 6820{ 6821 int i, j; 6822 unsigned long alloc_size = 0, ptr; 6823 6824#ifdef CONFIG_FAIR_GROUP_SCHED 6825 alloc_size += 2 * nr_cpu_ids * sizeof(void **); 6826#endif 6827#ifdef CONFIG_RT_GROUP_SCHED 6828 alloc_size += 2 * nr_cpu_ids * sizeof(void **); 6829#endif 6830#ifdef CONFIG_CPUMASK_OFFSTACK 6831 alloc_size += num_possible_cpus() * cpumask_size(); 6832#endif 6833 if (alloc_size) { 6834 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT); 6835 6836#ifdef CONFIG_FAIR_GROUP_SCHED 6837 root_task_group.se = (struct sched_entity **)ptr; 6838 ptr += nr_cpu_ids * sizeof(void **); 6839 6840 root_task_group.cfs_rq = (struct cfs_rq **)ptr; 6841 ptr += nr_cpu_ids * sizeof(void **); 6842 6843#endif /* CONFIG_FAIR_GROUP_SCHED */ 6844#ifdef CONFIG_RT_GROUP_SCHED 6845 root_task_group.rt_se = (struct sched_rt_entity **)ptr; 6846 ptr += nr_cpu_ids * sizeof(void **); 6847 6848 root_task_group.rt_rq = (struct rt_rq **)ptr; 6849 ptr += nr_cpu_ids * sizeof(void **); 6850 6851#endif /* CONFIG_RT_GROUP_SCHED */ 6852#ifdef CONFIG_CPUMASK_OFFSTACK 6853 for_each_possible_cpu(i) { 6854 per_cpu(load_balance_tmpmask, i) = (void *)ptr; 6855 ptr += cpumask_size(); 6856 } 6857#endif /* CONFIG_CPUMASK_OFFSTACK */ 6858 } 6859 6860#ifdef CONFIG_SMP 6861 init_defrootdomain(); 6862#endif 6863 6864 init_rt_bandwidth(&def_rt_bandwidth, 6865 global_rt_period(), global_rt_runtime()); 6866 6867#ifdef CONFIG_RT_GROUP_SCHED 6868 init_rt_bandwidth(&root_task_group.rt_bandwidth, 6869 global_rt_period(), global_rt_runtime()); 6870#endif /* CONFIG_RT_GROUP_SCHED */ 6871 6872#ifdef CONFIG_CGROUP_SCHED 6873 list_add(&root_task_group.list, &task_groups); 6874 INIT_LIST_HEAD(&root_task_group.children); 6875 INIT_LIST_HEAD(&root_task_group.siblings); 6876 autogroup_init(&init_task); 6877 6878#endif /* CONFIG_CGROUP_SCHED */ 6879 6880#ifdef CONFIG_CGROUP_CPUACCT 6881 root_cpuacct.cpustat = &kernel_cpustat; 6882 root_cpuacct.cpuusage = alloc_percpu(u64); 6883 /* Too early, not expected to fail */ 6884 BUG_ON(!root_cpuacct.cpuusage); 6885#endif 6886 for_each_possible_cpu(i) { 6887 struct rq *rq; 6888 6889 rq = cpu_rq(i); 6890 raw_spin_lock_init(&rq->lock); 6891 rq->nr_running = 0; 6892 rq->calc_load_active = 0; 6893 rq->calc_load_update = jiffies + LOAD_FREQ; 6894 init_cfs_rq(&rq->cfs); 6895 init_rt_rq(&rq->rt, rq); 6896#ifdef CONFIG_FAIR_GROUP_SCHED 6897 root_task_group.shares = ROOT_TASK_GROUP_LOAD; 6898 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list); 6899 /* 6900 * How much cpu bandwidth does root_task_group get? 6901 * 6902 * In case of task-groups formed thr' the cgroup filesystem, it 6903 * gets 100% of the cpu resources in the system. This overall 6904 * system cpu resource is divided among the tasks of 6905 * root_task_group and its child task-groups in a fair manner, 6906 * based on each entity's (task or task-group's) weight 6907 * (se->load.weight). 6908 * 6909 * In other words, if root_task_group has 10 tasks of weight 6910 * 1024) and two child groups A0 and A1 (of weight 1024 each), 6911 * then A0's share of the cpu resource is: 6912 * 6913 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33% 6914 * 6915 * We achieve this by letting root_task_group's tasks sit 6916 * directly in rq->cfs (i.e root_task_group->se[] = NULL). 6917 */ 6918 init_cfs_bandwidth(&root_task_group.cfs_bandwidth); 6919 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL); 6920#endif /* CONFIG_FAIR_GROUP_SCHED */ 6921 6922 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime; 6923#ifdef CONFIG_RT_GROUP_SCHED 6924 INIT_LIST_HEAD(&rq->leaf_rt_rq_list); 6925 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL); 6926#endif 6927 6928 for (j = 0; j < CPU_LOAD_IDX_MAX; j++) 6929 rq->cpu_load[j] = 0; 6930 6931 rq->last_load_update_tick = jiffies; 6932 6933#ifdef CONFIG_SMP 6934 rq->sd = NULL; 6935 rq->rd = NULL; 6936 rq->cpu_power = SCHED_POWER_SCALE; 6937 rq->post_schedule = 0; 6938 rq->active_balance = 0; 6939 rq->next_balance = jiffies; 6940 rq->push_cpu = 0; 6941 rq->cpu = i; 6942 rq->online = 0; 6943 rq->idle_stamp = 0; 6944 rq->avg_idle = 2*sysctl_sched_migration_cost; 6945 6946 INIT_LIST_HEAD(&rq->cfs_tasks); 6947 6948 rq_attach_root(rq, &def_root_domain); 6949#ifdef CONFIG_NO_HZ 6950 rq->nohz_flags = 0; 6951#endif 6952#endif 6953 init_rq_hrtick(rq); 6954 atomic_set(&rq->nr_iowait, 0); 6955 } 6956 6957 set_load_weight(&init_task); 6958 6959#ifdef CONFIG_PREEMPT_NOTIFIERS 6960 INIT_HLIST_HEAD(&init_task.preempt_notifiers); 6961#endif 6962 6963#ifdef CONFIG_RT_MUTEXES 6964 plist_head_init(&init_task.pi_waiters); 6965#endif 6966 6967 /* 6968 * The boot idle thread does lazy MMU switching as well: 6969 */ 6970 atomic_inc(&init_mm.mm_count); 6971 enter_lazy_tlb(&init_mm, current); 6972 6973 /* 6974 * Make us the idle thread. Technically, schedule() should not be 6975 * called from this thread, however somewhere below it might be, 6976 * but because we are the idle thread, we just pick up running again 6977 * when this runqueue becomes "idle". 6978 */ 6979 init_idle(current, smp_processor_id()); 6980 6981 calc_load_update = jiffies + LOAD_FREQ; 6982 6983 /* 6984 * During early bootup we pretend to be a normal task: 6985 */ 6986 current->sched_class = &fair_sched_class; 6987 6988#ifdef CONFIG_SMP 6989 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT); 6990 /* May be allocated at isolcpus cmdline parse time */ 6991 if (cpu_isolated_map == NULL) 6992 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT); 6993 idle_thread_set_boot_cpu(); 6994#endif 6995 init_sched_fair_class(); 6996 6997 scheduler_running = 1; 6998} 6999 7000#ifdef CONFIG_DEBUG_ATOMIC_SLEEP 7001static inline int preempt_count_equals(int preempt_offset) 7002{ 7003 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth(); 7004 7005 return (nested == preempt_offset); 7006} 7007 7008void __might_sleep(const char *file, int line, int preempt_offset) 7009{ 7010 static unsigned long prev_jiffy; /* ratelimiting */ 7011 7012 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */ 7013 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) || 7014 system_state != SYSTEM_RUNNING || oops_in_progress) 7015 return; 7016 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy) 7017 return; 7018 prev_jiffy = jiffies; 7019 7020 printk(KERN_ERR 7021 "BUG: sleeping function called from invalid context at %s:%d\n", 7022 file, line); 7023 printk(KERN_ERR 7024 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n", 7025 in_atomic(), irqs_disabled(), 7026 current->pid, current->comm); 7027 7028 debug_show_held_locks(current); 7029 if (irqs_disabled()) 7030 print_irqtrace_events(current); 7031 dump_stack(); 7032} 7033EXPORT_SYMBOL(__might_sleep); 7034#endif 7035 7036#ifdef CONFIG_MAGIC_SYSRQ 7037static void normalize_task(struct rq *rq, struct task_struct *p) 7038{ 7039 const struct sched_class *prev_class = p->sched_class; 7040 int old_prio = p->prio; 7041 int on_rq; 7042 7043 on_rq = p->on_rq; 7044 if (on_rq) 7045 dequeue_task(rq, p, 0); 7046 __setscheduler(rq, p, SCHED_NORMAL, 0); 7047 if (on_rq) { 7048 enqueue_task(rq, p, 0); 7049 resched_task(rq->curr); 7050 } 7051 7052 check_class_changed(rq, p, prev_class, old_prio); 7053} 7054 7055void normalize_rt_tasks(void) 7056{ 7057 struct task_struct *g, *p; 7058 unsigned long flags; 7059 struct rq *rq; 7060 7061 read_lock_irqsave(&tasklist_lock, flags); 7062 do_each_thread(g, p) { 7063 /* 7064 * Only normalize user tasks: 7065 */ 7066 if (!p->mm) 7067 continue; 7068 7069 p->se.exec_start = 0; 7070#ifdef CONFIG_SCHEDSTATS 7071 p->se.statistics.wait_start = 0; 7072 p->se.statistics.sleep_start = 0; 7073 p->se.statistics.block_start = 0; 7074#endif 7075 7076 if (!rt_task(p)) { 7077 /* 7078 * Renice negative nice level userspace 7079 * tasks back to 0: 7080 */ 7081 if (TASK_NICE(p) < 0 && p->mm) 7082 set_user_nice(p, 0); 7083 continue; 7084 } 7085 7086 raw_spin_lock(&p->pi_lock); 7087 rq = __task_rq_lock(p); 7088 7089 normalize_task(rq, p); 7090 7091 __task_rq_unlock(rq); 7092 raw_spin_unlock(&p->pi_lock); 7093 } while_each_thread(g, p); 7094 7095 read_unlock_irqrestore(&tasklist_lock, flags); 7096} 7097 7098#endif /* CONFIG_MAGIC_SYSRQ */ 7099 7100#if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) 7101/* 7102 * These functions are only useful for the IA64 MCA handling, or kdb. 7103 * 7104 * They can only be called when the whole system has been 7105 * stopped - every CPU needs to be quiescent, and no scheduling 7106 * activity can take place. Using them for anything else would 7107 * be a serious bug, and as a result, they aren't even visible 7108 * under any other configuration. 7109 */ 7110 7111/** 7112 * curr_task - return the current task for a given cpu. 7113 * @cpu: the processor in question. 7114 * 7115 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED! 7116 */ 7117struct task_struct *curr_task(int cpu) 7118{ 7119 return cpu_curr(cpu); 7120} 7121 7122#endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */ 7123 7124#ifdef CONFIG_IA64 7125/** 7126 * set_curr_task - set the current task for a given cpu. 7127 * @cpu: the processor in question. 7128 * @p: the task pointer to set. 7129 * 7130 * Description: This function must only be used when non-maskable interrupts 7131 * are serviced on a separate stack. It allows the architecture to switch the 7132 * notion of the current task on a cpu in a non-blocking manner. This function 7133 * must be called with all CPU's synchronized, and interrupts disabled, the 7134 * and caller must save the original value of the current task (see 7135 * curr_task() above) and restore that value before reenabling interrupts and 7136 * re-starting the system. 7137 * 7138 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED! 7139 */ 7140void set_curr_task(int cpu, struct task_struct *p) 7141{ 7142 cpu_curr(cpu) = p; 7143} 7144 7145#endif 7146 7147#ifdef CONFIG_CGROUP_SCHED 7148/* task_group_lock serializes the addition/removal of task groups */ 7149static DEFINE_SPINLOCK(task_group_lock); 7150 7151static void free_sched_group(struct task_group *tg) 7152{ 7153 free_fair_sched_group(tg); 7154 free_rt_sched_group(tg); 7155 autogroup_free(tg); 7156 kfree(tg); 7157} 7158 7159/* allocate runqueue etc for a new task group */ 7160struct task_group *sched_create_group(struct task_group *parent) 7161{ 7162 struct task_group *tg; 7163 unsigned long flags; 7164 7165 tg = kzalloc(sizeof(*tg), GFP_KERNEL); 7166 if (!tg) 7167 return ERR_PTR(-ENOMEM); 7168 7169 if (!alloc_fair_sched_group(tg, parent)) 7170 goto err; 7171 7172 if (!alloc_rt_sched_group(tg, parent)) 7173 goto err; 7174 7175 spin_lock_irqsave(&task_group_lock, flags); 7176 list_add_rcu(&tg->list, &task_groups); 7177 7178 WARN_ON(!parent); /* root should already exist */ 7179 7180 tg->parent = parent; 7181 INIT_LIST_HEAD(&tg->children); 7182 list_add_rcu(&tg->siblings, &parent->children); 7183 spin_unlock_irqrestore(&task_group_lock, flags); 7184 7185 return tg; 7186 7187err: 7188 free_sched_group(tg); 7189 return ERR_PTR(-ENOMEM); 7190} 7191 7192/* rcu callback to free various structures associated with a task group */ 7193static void free_sched_group_rcu(struct rcu_head *rhp) 7194{ 7195 /* now it should be safe to free those cfs_rqs */ 7196 free_sched_group(container_of(rhp, struct task_group, rcu)); 7197} 7198 7199/* Destroy runqueue etc associated with a task group */ 7200void sched_destroy_group(struct task_group *tg) 7201{ 7202 unsigned long flags; 7203 int i; 7204 7205 /* end participation in shares distribution */ 7206 for_each_possible_cpu(i) 7207 unregister_fair_sched_group(tg, i); 7208 7209 spin_lock_irqsave(&task_group_lock, flags); 7210 list_del_rcu(&tg->list); 7211 list_del_rcu(&tg->siblings); 7212 spin_unlock_irqrestore(&task_group_lock, flags); 7213 7214 /* wait for possible concurrent references to cfs_rqs complete */ 7215 call_rcu(&tg->rcu, free_sched_group_rcu); 7216} 7217 7218/* change task's runqueue when it moves between groups. 7219 * The caller of this function should have put the task in its new group 7220 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to 7221 * reflect its new group. 7222 */ 7223void sched_move_task(struct task_struct *tsk) 7224{ 7225 struct task_group *tg; 7226 int on_rq, running; 7227 unsigned long flags; 7228 struct rq *rq; 7229 7230 rq = task_rq_lock(tsk, &flags); 7231 7232 running = task_current(rq, tsk); 7233 on_rq = tsk->on_rq; 7234 7235 if (on_rq) 7236 dequeue_task(rq, tsk, 0); 7237 if (unlikely(running)) 7238 tsk->sched_class->put_prev_task(rq, tsk); 7239 7240 tg = container_of(task_subsys_state_check(tsk, cpu_cgroup_subsys_id, 7241 lockdep_is_held(&tsk->sighand->siglock)), 7242 struct task_group, css); 7243 tg = autogroup_task_group(tsk, tg); 7244 tsk->sched_task_group = tg; 7245 7246#ifdef CONFIG_FAIR_GROUP_SCHED 7247 if (tsk->sched_class->task_move_group) 7248 tsk->sched_class->task_move_group(tsk, on_rq); 7249 else 7250#endif 7251 set_task_rq(tsk, task_cpu(tsk)); 7252 7253 if (unlikely(running)) 7254 tsk->sched_class->set_curr_task(rq); 7255 if (on_rq) 7256 enqueue_task(rq, tsk, 0); 7257 7258 task_rq_unlock(rq, tsk, &flags); 7259} 7260#endif /* CONFIG_CGROUP_SCHED */ 7261 7262#if defined(CONFIG_RT_GROUP_SCHED) || defined(CONFIG_CFS_BANDWIDTH) 7263static unsigned long to_ratio(u64 period, u64 runtime) 7264{ 7265 if (runtime == RUNTIME_INF) 7266 return 1ULL << 20; 7267 7268 return div64_u64(runtime << 20, period); 7269} 7270#endif 7271 7272#ifdef CONFIG_RT_GROUP_SCHED 7273/* 7274 * Ensure that the real time constraints are schedulable. 7275 */ 7276static DEFINE_MUTEX(rt_constraints_mutex); 7277 7278/* Must be called with tasklist_lock held */ 7279static inline int tg_has_rt_tasks(struct task_group *tg) 7280{ 7281 struct task_struct *g, *p; 7282 7283 do_each_thread(g, p) { 7284 if (rt_task(p) && task_rq(p)->rt.tg == tg) 7285 return 1; 7286 } while_each_thread(g, p); 7287 7288 return 0; 7289} 7290 7291struct rt_schedulable_data { 7292 struct task_group *tg; 7293 u64 rt_period; 7294 u64 rt_runtime; 7295}; 7296 7297static int tg_rt_schedulable(struct task_group *tg, void *data) 7298{ 7299 struct rt_schedulable_data *d = data; 7300 struct task_group *child; 7301 unsigned long total, sum = 0; 7302 u64 period, runtime; 7303 7304 period = ktime_to_ns(tg->rt_bandwidth.rt_period); 7305 runtime = tg->rt_bandwidth.rt_runtime; 7306 7307 if (tg == d->tg) { 7308 period = d->rt_period; 7309 runtime = d->rt_runtime; 7310 } 7311 7312 /* 7313 * Cannot have more runtime than the period. 7314 */ 7315 if (runtime > period && runtime != RUNTIME_INF) 7316 return -EINVAL; 7317 7318 /* 7319 * Ensure we don't starve existing RT tasks. 7320 */ 7321 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg)) 7322 return -EBUSY; 7323 7324 total = to_ratio(period, runtime); 7325 7326 /* 7327 * Nobody can have more than the global setting allows. 7328 */ 7329 if (total > to_ratio(global_rt_period(), global_rt_runtime())) 7330 return -EINVAL; 7331 7332 /* 7333 * The sum of our children's runtime should not exceed our own. 7334 */ 7335 list_for_each_entry_rcu(child, &tg->children, siblings) { 7336 period = ktime_to_ns(child->rt_bandwidth.rt_period); 7337 runtime = child->rt_bandwidth.rt_runtime; 7338 7339 if (child == d->tg) { 7340 period = d->rt_period; 7341 runtime = d->rt_runtime; 7342 } 7343 7344 sum += to_ratio(period, runtime); 7345 } 7346 7347 if (sum > total) 7348 return -EINVAL; 7349 7350 return 0; 7351} 7352 7353static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime) 7354{ 7355 int ret; 7356 7357 struct rt_schedulable_data data = { 7358 .tg = tg, 7359 .rt_period = period, 7360 .rt_runtime = runtime, 7361 }; 7362 7363 rcu_read_lock(); 7364 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data); 7365 rcu_read_unlock(); 7366 7367 return ret; 7368} 7369 7370static int tg_set_rt_bandwidth(struct task_group *tg, 7371 u64 rt_period, u64 rt_runtime) 7372{ 7373 int i, err = 0; 7374 7375 mutex_lock(&rt_constraints_mutex); 7376 read_lock(&tasklist_lock); 7377 err = __rt_schedulable(tg, rt_period, rt_runtime); 7378 if (err) 7379 goto unlock; 7380 7381 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock); 7382 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period); 7383 tg->rt_bandwidth.rt_runtime = rt_runtime; 7384 7385 for_each_possible_cpu(i) { 7386 struct rt_rq *rt_rq = tg->rt_rq[i]; 7387 7388 raw_spin_lock(&rt_rq->rt_runtime_lock); 7389 rt_rq->rt_runtime = rt_runtime; 7390 raw_spin_unlock(&rt_rq->rt_runtime_lock); 7391 } 7392 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock); 7393unlock: 7394 read_unlock(&tasklist_lock); 7395 mutex_unlock(&rt_constraints_mutex); 7396 7397 return err; 7398} 7399 7400int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us) 7401{ 7402 u64 rt_runtime, rt_period; 7403 7404 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period); 7405 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC; 7406 if (rt_runtime_us < 0) 7407 rt_runtime = RUNTIME_INF; 7408 7409 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime); 7410} 7411 7412long sched_group_rt_runtime(struct task_group *tg) 7413{ 7414 u64 rt_runtime_us; 7415 7416 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF) 7417 return -1; 7418 7419 rt_runtime_us = tg->rt_bandwidth.rt_runtime; 7420 do_div(rt_runtime_us, NSEC_PER_USEC); 7421 return rt_runtime_us; 7422} 7423 7424int sched_group_set_rt_period(struct task_group *tg, long rt_period_us) 7425{ 7426 u64 rt_runtime, rt_period; 7427 7428 rt_period = (u64)rt_period_us * NSEC_PER_USEC; 7429 rt_runtime = tg->rt_bandwidth.rt_runtime; 7430 7431 if (rt_period == 0) 7432 return -EINVAL; 7433 7434 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime); 7435} 7436 7437long sched_group_rt_period(struct task_group *tg) 7438{ 7439 u64 rt_period_us; 7440 7441 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period); 7442 do_div(rt_period_us, NSEC_PER_USEC); 7443 return rt_period_us; 7444} 7445 7446static int sched_rt_global_constraints(void) 7447{ 7448 u64 runtime, period; 7449 int ret = 0; 7450 7451 if (sysctl_sched_rt_period <= 0) 7452 return -EINVAL; 7453 7454 runtime = global_rt_runtime(); 7455 period = global_rt_period(); 7456 7457 /* 7458 * Sanity check on the sysctl variables. 7459 */ 7460 if (runtime > period && runtime != RUNTIME_INF) 7461 return -EINVAL; 7462 7463 mutex_lock(&rt_constraints_mutex); 7464 read_lock(&tasklist_lock); 7465 ret = __rt_schedulable(NULL, 0, 0); 7466 read_unlock(&tasklist_lock); 7467 mutex_unlock(&rt_constraints_mutex); 7468 7469 return ret; 7470} 7471 7472int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk) 7473{ 7474 /* Don't accept realtime tasks when there is no way for them to run */ 7475 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0) 7476 return 0; 7477 7478 return 1; 7479} 7480 7481#else /* !CONFIG_RT_GROUP_SCHED */ 7482static int sched_rt_global_constraints(void) 7483{ 7484 unsigned long flags; 7485 int i; 7486 7487 if (sysctl_sched_rt_period <= 0) 7488 return -EINVAL; 7489 7490 /* 7491 * There's always some RT tasks in the root group 7492 * -- migration, kstopmachine etc.. 7493 */ 7494 if (sysctl_sched_rt_runtime == 0) 7495 return -EBUSY; 7496 7497 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags); 7498 for_each_possible_cpu(i) { 7499 struct rt_rq *rt_rq = &cpu_rq(i)->rt; 7500 7501 raw_spin_lock(&rt_rq->rt_runtime_lock); 7502 rt_rq->rt_runtime = global_rt_runtime(); 7503 raw_spin_unlock(&rt_rq->rt_runtime_lock); 7504 } 7505 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags); 7506 7507 return 0; 7508} 7509#endif /* CONFIG_RT_GROUP_SCHED */ 7510 7511int sched_rt_handler(struct ctl_table *table, int write, 7512 void __user *buffer, size_t *lenp, 7513 loff_t *ppos) 7514{ 7515 int ret; 7516 int old_period, old_runtime; 7517 static DEFINE_MUTEX(mutex); 7518 7519 mutex_lock(&mutex); 7520 old_period = sysctl_sched_rt_period; 7521 old_runtime = sysctl_sched_rt_runtime; 7522 7523 ret = proc_dointvec(table, write, buffer, lenp, ppos); 7524 7525 if (!ret && write) { 7526 ret = sched_rt_global_constraints(); 7527 if (ret) { 7528 sysctl_sched_rt_period = old_period; 7529 sysctl_sched_rt_runtime = old_runtime; 7530 } else { 7531 def_rt_bandwidth.rt_runtime = global_rt_runtime(); 7532 def_rt_bandwidth.rt_period = 7533 ns_to_ktime(global_rt_period()); 7534 } 7535 } 7536 mutex_unlock(&mutex); 7537 7538 return ret; 7539} 7540 7541#ifdef CONFIG_CGROUP_SCHED 7542 7543/* return corresponding task_group object of a cgroup */ 7544static inline struct task_group *cgroup_tg(struct cgroup *cgrp) 7545{ 7546 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id), 7547 struct task_group, css); 7548} 7549 7550static struct cgroup_subsys_state *cpu_cgroup_css_alloc(struct cgroup *cgrp) 7551{ 7552 struct task_group *tg, *parent; 7553 7554 if (!cgrp->parent) { 7555 /* This is early initialization for the top cgroup */ 7556 return &root_task_group.css; 7557 } 7558 7559 parent = cgroup_tg(cgrp->parent); 7560 tg = sched_create_group(parent); 7561 if (IS_ERR(tg)) 7562 return ERR_PTR(-ENOMEM); 7563 7564 return &tg->css; 7565} 7566 7567static void cpu_cgroup_css_free(struct cgroup *cgrp) 7568{ 7569 struct task_group *tg = cgroup_tg(cgrp); 7570 7571 sched_destroy_group(tg); 7572} 7573 7574static int cpu_cgroup_can_attach(struct cgroup *cgrp, 7575 struct cgroup_taskset *tset) 7576{ 7577 struct task_struct *task; 7578 7579 cgroup_taskset_for_each(task, cgrp, tset) { 7580#ifdef CONFIG_RT_GROUP_SCHED 7581 if (!sched_rt_can_attach(cgroup_tg(cgrp), task)) 7582 return -EINVAL; 7583#else 7584 /* We don't support RT-tasks being in separate groups */ 7585 if (task->sched_class != &fair_sched_class) 7586 return -EINVAL; 7587#endif 7588 } 7589 return 0; 7590} 7591 7592static void cpu_cgroup_attach(struct cgroup *cgrp, 7593 struct cgroup_taskset *tset) 7594{ 7595 struct task_struct *task; 7596 7597 cgroup_taskset_for_each(task, cgrp, tset) 7598 sched_move_task(task); 7599} 7600 7601static void 7602cpu_cgroup_exit(struct cgroup *cgrp, struct cgroup *old_cgrp, 7603 struct task_struct *task) 7604{ 7605 /* 7606 * cgroup_exit() is called in the copy_process() failure path. 7607 * Ignore this case since the task hasn't ran yet, this avoids 7608 * trying to poke a half freed task state from generic code. 7609 */ 7610 if (!(task->flags & PF_EXITING)) 7611 return; 7612 7613 sched_move_task(task); 7614} 7615 7616#ifdef CONFIG_FAIR_GROUP_SCHED 7617static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype, 7618 u64 shareval) 7619{ 7620 return sched_group_set_shares(cgroup_tg(cgrp), scale_load(shareval)); 7621} 7622 7623static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft) 7624{ 7625 struct task_group *tg = cgroup_tg(cgrp); 7626 7627 return (u64) scale_load_down(tg->shares); 7628} 7629 7630#ifdef CONFIG_CFS_BANDWIDTH 7631static DEFINE_MUTEX(cfs_constraints_mutex); 7632 7633const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */ 7634const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */ 7635 7636static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime); 7637 7638static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota) 7639{ 7640 int i, ret = 0, runtime_enabled, runtime_was_enabled; 7641 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth; 7642 7643 if (tg == &root_task_group) 7644 return -EINVAL; 7645 7646 /* 7647 * Ensure we have at some amount of bandwidth every period. This is 7648 * to prevent reaching a state of large arrears when throttled via 7649 * entity_tick() resulting in prolonged exit starvation. 7650 */ 7651 if (quota < min_cfs_quota_period || period < min_cfs_quota_period) 7652 return -EINVAL; 7653 7654 /* 7655 * Likewise, bound things on the otherside by preventing insane quota 7656 * periods. This also allows us to normalize in computing quota 7657 * feasibility. 7658 */ 7659 if (period > max_cfs_quota_period) 7660 return -EINVAL; 7661 7662 mutex_lock(&cfs_constraints_mutex); 7663 ret = __cfs_schedulable(tg, period, quota); 7664 if (ret) 7665 goto out_unlock; 7666 7667 runtime_enabled = quota != RUNTIME_INF; 7668 runtime_was_enabled = cfs_b->quota != RUNTIME_INF; 7669 account_cfs_bandwidth_used(runtime_enabled, runtime_was_enabled); 7670 raw_spin_lock_irq(&cfs_b->lock); 7671 cfs_b->period = ns_to_ktime(period); 7672 cfs_b->quota = quota; 7673 7674 __refill_cfs_bandwidth_runtime(cfs_b); 7675 /* restart the period timer (if active) to handle new period expiry */ 7676 if (runtime_enabled && cfs_b->timer_active) { 7677 /* force a reprogram */ 7678 cfs_b->timer_active = 0; 7679 __start_cfs_bandwidth(cfs_b); 7680 } 7681 raw_spin_unlock_irq(&cfs_b->lock); 7682 7683 for_each_possible_cpu(i) { 7684 struct cfs_rq *cfs_rq = tg->cfs_rq[i]; 7685 struct rq *rq = cfs_rq->rq; 7686 7687 raw_spin_lock_irq(&rq->lock); 7688 cfs_rq->runtime_enabled = runtime_enabled; 7689 cfs_rq->runtime_remaining = 0; 7690 7691 if (cfs_rq->throttled) 7692 unthrottle_cfs_rq(cfs_rq); 7693 raw_spin_unlock_irq(&rq->lock); 7694 } 7695out_unlock: 7696 mutex_unlock(&cfs_constraints_mutex); 7697 7698 return ret; 7699} 7700 7701int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us) 7702{ 7703 u64 quota, period; 7704 7705 period = ktime_to_ns(tg->cfs_bandwidth.period); 7706 if (cfs_quota_us < 0) 7707 quota = RUNTIME_INF; 7708 else 7709 quota = (u64)cfs_quota_us * NSEC_PER_USEC; 7710 7711 return tg_set_cfs_bandwidth(tg, period, quota); 7712} 7713 7714long tg_get_cfs_quota(struct task_group *tg) 7715{ 7716 u64 quota_us; 7717 7718 if (tg->cfs_bandwidth.quota == RUNTIME_INF) 7719 return -1; 7720 7721 quota_us = tg->cfs_bandwidth.quota; 7722 do_div(quota_us, NSEC_PER_USEC); 7723 7724 return quota_us; 7725} 7726 7727int tg_set_cfs_period(struct task_group *tg, long cfs_period_us) 7728{ 7729 u64 quota, period; 7730 7731 period = (u64)cfs_period_us * NSEC_PER_USEC; 7732 quota = tg->cfs_bandwidth.quota; 7733 7734 return tg_set_cfs_bandwidth(tg, period, quota); 7735} 7736 7737long tg_get_cfs_period(struct task_group *tg) 7738{ 7739 u64 cfs_period_us; 7740 7741 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period); 7742 do_div(cfs_period_us, NSEC_PER_USEC); 7743 7744 return cfs_period_us; 7745} 7746 7747static s64 cpu_cfs_quota_read_s64(struct cgroup *cgrp, struct cftype *cft) 7748{ 7749 return tg_get_cfs_quota(cgroup_tg(cgrp)); 7750} 7751 7752static int cpu_cfs_quota_write_s64(struct cgroup *cgrp, struct cftype *cftype, 7753 s64 cfs_quota_us) 7754{ 7755 return tg_set_cfs_quota(cgroup_tg(cgrp), cfs_quota_us); 7756} 7757 7758static u64 cpu_cfs_period_read_u64(struct cgroup *cgrp, struct cftype *cft) 7759{ 7760 return tg_get_cfs_period(cgroup_tg(cgrp)); 7761} 7762 7763static int cpu_cfs_period_write_u64(struct cgroup *cgrp, struct cftype *cftype, 7764 u64 cfs_period_us) 7765{ 7766 return tg_set_cfs_period(cgroup_tg(cgrp), cfs_period_us); 7767} 7768 7769struct cfs_schedulable_data { 7770 struct task_group *tg; 7771 u64 period, quota; 7772}; 7773 7774/* 7775 * normalize group quota/period to be quota/max_period 7776 * note: units are usecs 7777 */ 7778static u64 normalize_cfs_quota(struct task_group *tg, 7779 struct cfs_schedulable_data *d) 7780{ 7781 u64 quota, period; 7782 7783 if (tg == d->tg) { 7784 period = d->period; 7785 quota = d->quota; 7786 } else { 7787 period = tg_get_cfs_period(tg); 7788 quota = tg_get_cfs_quota(tg); 7789 } 7790 7791 /* note: these should typically be equivalent */ 7792 if (quota == RUNTIME_INF || quota == -1) 7793 return RUNTIME_INF; 7794 7795 return to_ratio(period, quota); 7796} 7797 7798static int tg_cfs_schedulable_down(struct task_group *tg, void *data) 7799{ 7800 struct cfs_schedulable_data *d = data; 7801 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth; 7802 s64 quota = 0, parent_quota = -1; 7803 7804 if (!tg->parent) { 7805 quota = RUNTIME_INF; 7806 } else { 7807 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth; 7808 7809 quota = normalize_cfs_quota(tg, d); 7810 parent_quota = parent_b->hierarchal_quota; 7811 7812 /* 7813 * ensure max(child_quota) <= parent_quota, inherit when no 7814 * limit is set 7815 */ 7816 if (quota == RUNTIME_INF) 7817 quota = parent_quota; 7818 else if (parent_quota != RUNTIME_INF && quota > parent_quota) 7819 return -EINVAL; 7820 } 7821 cfs_b->hierarchal_quota = quota; 7822 7823 return 0; 7824} 7825 7826static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota) 7827{ 7828 int ret; 7829 struct cfs_schedulable_data data = { 7830 .tg = tg, 7831 .period = period, 7832 .quota = quota, 7833 }; 7834 7835 if (quota != RUNTIME_INF) { 7836 do_div(data.period, NSEC_PER_USEC); 7837 do_div(data.quota, NSEC_PER_USEC); 7838 } 7839 7840 rcu_read_lock(); 7841 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data); 7842 rcu_read_unlock(); 7843 7844 return ret; 7845} 7846 7847static int cpu_stats_show(struct cgroup *cgrp, struct cftype *cft, 7848 struct cgroup_map_cb *cb) 7849{ 7850 struct task_group *tg = cgroup_tg(cgrp); 7851 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth; 7852 7853 cb->fill(cb, "nr_periods", cfs_b->nr_periods); 7854 cb->fill(cb, "nr_throttled", cfs_b->nr_throttled); 7855 cb->fill(cb, "throttled_time", cfs_b->throttled_time); 7856 7857 return 0; 7858} 7859#endif /* CONFIG_CFS_BANDWIDTH */ 7860#endif /* CONFIG_FAIR_GROUP_SCHED */ 7861 7862#ifdef CONFIG_RT_GROUP_SCHED 7863static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft, 7864 s64 val) 7865{ 7866 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val); 7867} 7868 7869static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft) 7870{ 7871 return sched_group_rt_runtime(cgroup_tg(cgrp)); 7872} 7873 7874static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype, 7875 u64 rt_period_us) 7876{ 7877 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us); 7878} 7879 7880static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft) 7881{ 7882 return sched_group_rt_period(cgroup_tg(cgrp)); 7883} 7884#endif /* CONFIG_RT_GROUP_SCHED */ 7885 7886static struct cftype cpu_files[] = { 7887#ifdef CONFIG_FAIR_GROUP_SCHED 7888 { 7889 .name = "shares", 7890 .read_u64 = cpu_shares_read_u64, 7891 .write_u64 = cpu_shares_write_u64, 7892 }, 7893#endif 7894#ifdef CONFIG_CFS_BANDWIDTH 7895 { 7896 .name = "cfs_quota_us", 7897 .read_s64 = cpu_cfs_quota_read_s64, 7898 .write_s64 = cpu_cfs_quota_write_s64, 7899 }, 7900 { 7901 .name = "cfs_period_us", 7902 .read_u64 = cpu_cfs_period_read_u64, 7903 .write_u64 = cpu_cfs_period_write_u64, 7904 }, 7905 { 7906 .name = "stat", 7907 .read_map = cpu_stats_show, 7908 }, 7909#endif 7910#ifdef CONFIG_RT_GROUP_SCHED 7911 { 7912 .name = "rt_runtime_us", 7913 .read_s64 = cpu_rt_runtime_read, 7914 .write_s64 = cpu_rt_runtime_write, 7915 }, 7916 { 7917 .name = "rt_period_us", 7918 .read_u64 = cpu_rt_period_read_uint, 7919 .write_u64 = cpu_rt_period_write_uint, 7920 }, 7921#endif 7922 { } /* terminate */ 7923}; 7924 7925struct cgroup_subsys cpu_cgroup_subsys = { 7926 .name = "cpu", 7927 .css_alloc = cpu_cgroup_css_alloc, 7928 .css_free = cpu_cgroup_css_free, 7929 .can_attach = cpu_cgroup_can_attach, 7930 .attach = cpu_cgroup_attach, 7931 .exit = cpu_cgroup_exit, 7932 .subsys_id = cpu_cgroup_subsys_id, 7933 .base_cftypes = cpu_files, 7934 .early_init = 1, 7935}; 7936 7937#endif /* CONFIG_CGROUP_SCHED */ 7938 7939#ifdef CONFIG_CGROUP_CPUACCT 7940 7941/* 7942 * CPU accounting code for task groups. 7943 * 7944 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh 7945 * (balbir@in.ibm.com). 7946 */ 7947 7948struct cpuacct root_cpuacct; 7949 7950/* create a new cpu accounting group */ 7951static struct cgroup_subsys_state *cpuacct_css_alloc(struct cgroup *cgrp) 7952{ 7953 struct cpuacct *ca; 7954 7955 if (!cgrp->parent) 7956 return &root_cpuacct.css; 7957 7958 ca = kzalloc(sizeof(*ca), GFP_KERNEL); 7959 if (!ca) 7960 goto out; 7961 7962 ca->cpuusage = alloc_percpu(u64); 7963 if (!ca->cpuusage) 7964 goto out_free_ca; 7965 7966 ca->cpustat = alloc_percpu(struct kernel_cpustat); 7967 if (!ca->cpustat) 7968 goto out_free_cpuusage; 7969 7970 return &ca->css; 7971 7972out_free_cpuusage: 7973 free_percpu(ca->cpuusage); 7974out_free_ca: 7975 kfree(ca); 7976out: 7977 return ERR_PTR(-ENOMEM); 7978} 7979 7980/* destroy an existing cpu accounting group */ 7981static void cpuacct_css_free(struct cgroup *cgrp) 7982{ 7983 struct cpuacct *ca = cgroup_ca(cgrp); 7984 7985 free_percpu(ca->cpustat); 7986 free_percpu(ca->cpuusage); 7987 kfree(ca); 7988} 7989 7990static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu) 7991{ 7992 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu); 7993 u64 data; 7994 7995#ifndef CONFIG_64BIT 7996 /* 7997 * Take rq->lock to make 64-bit read safe on 32-bit platforms. 7998 */ 7999 raw_spin_lock_irq(&cpu_rq(cpu)->lock); 8000 data = *cpuusage; 8001 raw_spin_unlock_irq(&cpu_rq(cpu)->lock); 8002#else 8003 data = *cpuusage; 8004#endif 8005 8006 return data; 8007} 8008 8009static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val) 8010{ 8011 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu); 8012 8013#ifndef CONFIG_64BIT 8014 /* 8015 * Take rq->lock to make 64-bit write safe on 32-bit platforms. 8016 */ 8017 raw_spin_lock_irq(&cpu_rq(cpu)->lock); 8018 *cpuusage = val; 8019 raw_spin_unlock_irq(&cpu_rq(cpu)->lock); 8020#else 8021 *cpuusage = val; 8022#endif 8023} 8024 8025/* return total cpu usage (in nanoseconds) of a group */ 8026static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft) 8027{ 8028 struct cpuacct *ca = cgroup_ca(cgrp); 8029 u64 totalcpuusage = 0; 8030 int i; 8031 8032 for_each_present_cpu(i) 8033 totalcpuusage += cpuacct_cpuusage_read(ca, i); 8034 8035 return totalcpuusage; 8036} 8037 8038static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype, 8039 u64 reset) 8040{ 8041 struct cpuacct *ca = cgroup_ca(cgrp); 8042 int err = 0; 8043 int i; 8044 8045 if (reset) { 8046 err = -EINVAL; 8047 goto out; 8048 } 8049 8050 for_each_present_cpu(i) 8051 cpuacct_cpuusage_write(ca, i, 0); 8052 8053out: 8054 return err; 8055} 8056 8057static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft, 8058 struct seq_file *m) 8059{ 8060 struct cpuacct *ca = cgroup_ca(cgroup); 8061 u64 percpu; 8062 int i; 8063 8064 for_each_present_cpu(i) { 8065 percpu = cpuacct_cpuusage_read(ca, i); 8066 seq_printf(m, "%llu ", (unsigned long long) percpu); 8067 } 8068 seq_printf(m, "\n"); 8069 return 0; 8070} 8071 8072static const char *cpuacct_stat_desc[] = { 8073 [CPUACCT_STAT_USER] = "user", 8074 [CPUACCT_STAT_SYSTEM] = "system", 8075}; 8076 8077static int cpuacct_stats_show(struct cgroup *cgrp, struct cftype *cft, 8078 struct cgroup_map_cb *cb) 8079{ 8080 struct cpuacct *ca = cgroup_ca(cgrp); 8081 int cpu; 8082 s64 val = 0; 8083 8084 for_each_online_cpu(cpu) { 8085 struct kernel_cpustat *kcpustat = per_cpu_ptr(ca->cpustat, cpu); 8086 val += kcpustat->cpustat[CPUTIME_USER]; 8087 val += kcpustat->cpustat[CPUTIME_NICE]; 8088 } 8089 val = cputime64_to_clock_t(val); 8090 cb->fill(cb, cpuacct_stat_desc[CPUACCT_STAT_USER], val); 8091 8092 val = 0; 8093 for_each_online_cpu(cpu) { 8094 struct kernel_cpustat *kcpustat = per_cpu_ptr(ca->cpustat, cpu); 8095 val += kcpustat->cpustat[CPUTIME_SYSTEM]; 8096 val += kcpustat->cpustat[CPUTIME_IRQ]; 8097 val += kcpustat->cpustat[CPUTIME_SOFTIRQ]; 8098 } 8099 8100 val = cputime64_to_clock_t(val); 8101 cb->fill(cb, cpuacct_stat_desc[CPUACCT_STAT_SYSTEM], val); 8102 8103 return 0; 8104} 8105 8106static struct cftype files[] = { 8107 { 8108 .name = "usage", 8109 .read_u64 = cpuusage_read, 8110 .write_u64 = cpuusage_write, 8111 }, 8112 { 8113 .name = "usage_percpu", 8114 .read_seq_string = cpuacct_percpu_seq_read, 8115 }, 8116 { 8117 .name = "stat", 8118 .read_map = cpuacct_stats_show, 8119 }, 8120 { } /* terminate */ 8121}; 8122 8123/* 8124 * charge this task's execution time to its accounting group. 8125 * 8126 * called with rq->lock held. 8127 */ 8128void cpuacct_charge(struct task_struct *tsk, u64 cputime) 8129{ 8130 struct cpuacct *ca; 8131 int cpu; 8132 8133 if (unlikely(!cpuacct_subsys.active)) 8134 return; 8135 8136 cpu = task_cpu(tsk); 8137 8138 rcu_read_lock(); 8139 8140 ca = task_ca(tsk); 8141 8142 for (; ca; ca = parent_ca(ca)) { 8143 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu); 8144 *cpuusage += cputime; 8145 } 8146 8147 rcu_read_unlock(); 8148} 8149 8150struct cgroup_subsys cpuacct_subsys = { 8151 .name = "cpuacct", 8152 .css_alloc = cpuacct_css_alloc, 8153 .css_free = cpuacct_css_free, 8154 .subsys_id = cpuacct_subsys_id, 8155 .base_cftypes = files, 8156}; 8157#endif /* CONFIG_CGROUP_CPUACCT */ 8158 8159void dump_cpu_task(int cpu) 8160{ 8161 pr_info("Task dump for CPU %d:\n", cpu); 8162 sched_show_task(cpu_curr(cpu)); 8163} 8164