core.c revision 7d9ffa8961482232d964173cccba6e14d2d543b2
1/* 2 * kernel/sched/core.c 3 * 4 * Kernel scheduler and related syscalls 5 * 6 * Copyright (C) 1991-2002 Linus Torvalds 7 * 8 * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and 9 * make semaphores SMP safe 10 * 1998-11-19 Implemented schedule_timeout() and related stuff 11 * by Andrea Arcangeli 12 * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar: 13 * hybrid priority-list and round-robin design with 14 * an array-switch method of distributing timeslices 15 * and per-CPU runqueues. Cleanups and useful suggestions 16 * by Davide Libenzi, preemptible kernel bits by Robert Love. 17 * 2003-09-03 Interactivity tuning by Con Kolivas. 18 * 2004-04-02 Scheduler domains code by Nick Piggin 19 * 2007-04-15 Work begun on replacing all interactivity tuning with a 20 * fair scheduling design by Con Kolivas. 21 * 2007-05-05 Load balancing (smp-nice) and other improvements 22 * by Peter Williams 23 * 2007-05-06 Interactivity improvements to CFS by Mike Galbraith 24 * 2007-07-01 Group scheduling enhancements by Srivatsa Vaddagiri 25 * 2007-11-29 RT balancing improvements by Steven Rostedt, Gregory Haskins, 26 * Thomas Gleixner, Mike Kravetz 27 */ 28 29#include <linux/mm.h> 30#include <linux/module.h> 31#include <linux/nmi.h> 32#include <linux/init.h> 33#include <linux/uaccess.h> 34#include <linux/highmem.h> 35#include <asm/mmu_context.h> 36#include <linux/interrupt.h> 37#include <linux/capability.h> 38#include <linux/completion.h> 39#include <linux/kernel_stat.h> 40#include <linux/debug_locks.h> 41#include <linux/perf_event.h> 42#include <linux/security.h> 43#include <linux/notifier.h> 44#include <linux/profile.h> 45#include <linux/freezer.h> 46#include <linux/vmalloc.h> 47#include <linux/blkdev.h> 48#include <linux/delay.h> 49#include <linux/pid_namespace.h> 50#include <linux/smp.h> 51#include <linux/threads.h> 52#include <linux/timer.h> 53#include <linux/rcupdate.h> 54#include <linux/cpu.h> 55#include <linux/cpuset.h> 56#include <linux/percpu.h> 57#include <linux/proc_fs.h> 58#include <linux/seq_file.h> 59#include <linux/sysctl.h> 60#include <linux/syscalls.h> 61#include <linux/times.h> 62#include <linux/tsacct_kern.h> 63#include <linux/kprobes.h> 64#include <linux/delayacct.h> 65#include <linux/unistd.h> 66#include <linux/pagemap.h> 67#include <linux/hrtimer.h> 68#include <linux/tick.h> 69#include <linux/debugfs.h> 70#include <linux/ctype.h> 71#include <linux/ftrace.h> 72#include <linux/slab.h> 73#include <linux/init_task.h> 74#include <linux/binfmts.h> 75#include <linux/context_tracking.h> 76 77#include <asm/switch_to.h> 78#include <asm/tlb.h> 79#include <asm/irq_regs.h> 80#include <asm/mutex.h> 81#ifdef CONFIG_PARAVIRT 82#include <asm/paravirt.h> 83#endif 84 85#include "sched.h" 86#include "../workqueue_internal.h" 87#include "../smpboot.h" 88 89#define CREATE_TRACE_POINTS 90#include <trace/events/sched.h> 91 92void start_bandwidth_timer(struct hrtimer *period_timer, ktime_t period) 93{ 94 unsigned long delta; 95 ktime_t soft, hard, now; 96 97 for (;;) { 98 if (hrtimer_active(period_timer)) 99 break; 100 101 now = hrtimer_cb_get_time(period_timer); 102 hrtimer_forward(period_timer, now, period); 103 104 soft = hrtimer_get_softexpires(period_timer); 105 hard = hrtimer_get_expires(period_timer); 106 delta = ktime_to_ns(ktime_sub(hard, soft)); 107 __hrtimer_start_range_ns(period_timer, soft, delta, 108 HRTIMER_MODE_ABS_PINNED, 0); 109 } 110} 111 112DEFINE_MUTEX(sched_domains_mutex); 113DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues); 114 115static void update_rq_clock_task(struct rq *rq, s64 delta); 116 117void update_rq_clock(struct rq *rq) 118{ 119 s64 delta; 120 121 if (rq->skip_clock_update > 0) 122 return; 123 124 delta = sched_clock_cpu(cpu_of(rq)) - rq->clock; 125 rq->clock += delta; 126 update_rq_clock_task(rq, delta); 127} 128 129/* 130 * Debugging: various feature bits 131 */ 132 133#define SCHED_FEAT(name, enabled) \ 134 (1UL << __SCHED_FEAT_##name) * enabled | 135 136const_debug unsigned int sysctl_sched_features = 137#include "features.h" 138 0; 139 140#undef SCHED_FEAT 141 142#ifdef CONFIG_SCHED_DEBUG 143#define SCHED_FEAT(name, enabled) \ 144 #name , 145 146static const char * const sched_feat_names[] = { 147#include "features.h" 148}; 149 150#undef SCHED_FEAT 151 152static int sched_feat_show(struct seq_file *m, void *v) 153{ 154 int i; 155 156 for (i = 0; i < __SCHED_FEAT_NR; i++) { 157 if (!(sysctl_sched_features & (1UL << i))) 158 seq_puts(m, "NO_"); 159 seq_printf(m, "%s ", sched_feat_names[i]); 160 } 161 seq_puts(m, "\n"); 162 163 return 0; 164} 165 166#ifdef HAVE_JUMP_LABEL 167 168#define jump_label_key__true STATIC_KEY_INIT_TRUE 169#define jump_label_key__false STATIC_KEY_INIT_FALSE 170 171#define SCHED_FEAT(name, enabled) \ 172 jump_label_key__##enabled , 173 174struct static_key sched_feat_keys[__SCHED_FEAT_NR] = { 175#include "features.h" 176}; 177 178#undef SCHED_FEAT 179 180static void sched_feat_disable(int i) 181{ 182 if (static_key_enabled(&sched_feat_keys[i])) 183 static_key_slow_dec(&sched_feat_keys[i]); 184} 185 186static void sched_feat_enable(int i) 187{ 188 if (!static_key_enabled(&sched_feat_keys[i])) 189 static_key_slow_inc(&sched_feat_keys[i]); 190} 191#else 192static void sched_feat_disable(int i) { }; 193static void sched_feat_enable(int i) { }; 194#endif /* HAVE_JUMP_LABEL */ 195 196static int sched_feat_set(char *cmp) 197{ 198 int i; 199 int neg = 0; 200 201 if (strncmp(cmp, "NO_", 3) == 0) { 202 neg = 1; 203 cmp += 3; 204 } 205 206 for (i = 0; i < __SCHED_FEAT_NR; i++) { 207 if (strcmp(cmp, sched_feat_names[i]) == 0) { 208 if (neg) { 209 sysctl_sched_features &= ~(1UL << i); 210 sched_feat_disable(i); 211 } else { 212 sysctl_sched_features |= (1UL << i); 213 sched_feat_enable(i); 214 } 215 break; 216 } 217 } 218 219 return i; 220} 221 222static ssize_t 223sched_feat_write(struct file *filp, const char __user *ubuf, 224 size_t cnt, loff_t *ppos) 225{ 226 char buf[64]; 227 char *cmp; 228 int i; 229 230 if (cnt > 63) 231 cnt = 63; 232 233 if (copy_from_user(&buf, ubuf, cnt)) 234 return -EFAULT; 235 236 buf[cnt] = 0; 237 cmp = strstrip(buf); 238 239 i = sched_feat_set(cmp); 240 if (i == __SCHED_FEAT_NR) 241 return -EINVAL; 242 243 *ppos += cnt; 244 245 return cnt; 246} 247 248static int sched_feat_open(struct inode *inode, struct file *filp) 249{ 250 return single_open(filp, sched_feat_show, NULL); 251} 252 253static const struct file_operations sched_feat_fops = { 254 .open = sched_feat_open, 255 .write = sched_feat_write, 256 .read = seq_read, 257 .llseek = seq_lseek, 258 .release = single_release, 259}; 260 261static __init int sched_init_debug(void) 262{ 263 debugfs_create_file("sched_features", 0644, NULL, NULL, 264 &sched_feat_fops); 265 266 return 0; 267} 268late_initcall(sched_init_debug); 269#endif /* CONFIG_SCHED_DEBUG */ 270 271/* 272 * Number of tasks to iterate in a single balance run. 273 * Limited because this is done with IRQs disabled. 274 */ 275const_debug unsigned int sysctl_sched_nr_migrate = 32; 276 277/* 278 * period over which we average the RT time consumption, measured 279 * in ms. 280 * 281 * default: 1s 282 */ 283const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC; 284 285/* 286 * period over which we measure -rt task cpu usage in us. 287 * default: 1s 288 */ 289unsigned int sysctl_sched_rt_period = 1000000; 290 291__read_mostly int scheduler_running; 292 293/* 294 * part of the period that we allow rt tasks to run in us. 295 * default: 0.95s 296 */ 297int sysctl_sched_rt_runtime = 950000; 298 299 300 301/* 302 * __task_rq_lock - lock the rq @p resides on. 303 */ 304static inline struct rq *__task_rq_lock(struct task_struct *p) 305 __acquires(rq->lock) 306{ 307 struct rq *rq; 308 309 lockdep_assert_held(&p->pi_lock); 310 311 for (;;) { 312 rq = task_rq(p); 313 raw_spin_lock(&rq->lock); 314 if (likely(rq == task_rq(p))) 315 return rq; 316 raw_spin_unlock(&rq->lock); 317 } 318} 319 320/* 321 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on. 322 */ 323static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags) 324 __acquires(p->pi_lock) 325 __acquires(rq->lock) 326{ 327 struct rq *rq; 328 329 for (;;) { 330 raw_spin_lock_irqsave(&p->pi_lock, *flags); 331 rq = task_rq(p); 332 raw_spin_lock(&rq->lock); 333 if (likely(rq == task_rq(p))) 334 return rq; 335 raw_spin_unlock(&rq->lock); 336 raw_spin_unlock_irqrestore(&p->pi_lock, *flags); 337 } 338} 339 340static void __task_rq_unlock(struct rq *rq) 341 __releases(rq->lock) 342{ 343 raw_spin_unlock(&rq->lock); 344} 345 346static inline void 347task_rq_unlock(struct rq *rq, struct task_struct *p, unsigned long *flags) 348 __releases(rq->lock) 349 __releases(p->pi_lock) 350{ 351 raw_spin_unlock(&rq->lock); 352 raw_spin_unlock_irqrestore(&p->pi_lock, *flags); 353} 354 355/* 356 * this_rq_lock - lock this runqueue and disable interrupts. 357 */ 358static struct rq *this_rq_lock(void) 359 __acquires(rq->lock) 360{ 361 struct rq *rq; 362 363 local_irq_disable(); 364 rq = this_rq(); 365 raw_spin_lock(&rq->lock); 366 367 return rq; 368} 369 370#ifdef CONFIG_SCHED_HRTICK 371/* 372 * Use HR-timers to deliver accurate preemption points. 373 */ 374 375static void hrtick_clear(struct rq *rq) 376{ 377 if (hrtimer_active(&rq->hrtick_timer)) 378 hrtimer_cancel(&rq->hrtick_timer); 379} 380 381/* 382 * High-resolution timer tick. 383 * Runs from hardirq context with interrupts disabled. 384 */ 385static enum hrtimer_restart hrtick(struct hrtimer *timer) 386{ 387 struct rq *rq = container_of(timer, struct rq, hrtick_timer); 388 389 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id()); 390 391 raw_spin_lock(&rq->lock); 392 update_rq_clock(rq); 393 rq->curr->sched_class->task_tick(rq, rq->curr, 1); 394 raw_spin_unlock(&rq->lock); 395 396 return HRTIMER_NORESTART; 397} 398 399#ifdef CONFIG_SMP 400 401static int __hrtick_restart(struct rq *rq) 402{ 403 struct hrtimer *timer = &rq->hrtick_timer; 404 ktime_t time = hrtimer_get_softexpires(timer); 405 406 return __hrtimer_start_range_ns(timer, time, 0, HRTIMER_MODE_ABS_PINNED, 0); 407} 408 409/* 410 * called from hardirq (IPI) context 411 */ 412static void __hrtick_start(void *arg) 413{ 414 struct rq *rq = arg; 415 416 raw_spin_lock(&rq->lock); 417 __hrtick_restart(rq); 418 rq->hrtick_csd_pending = 0; 419 raw_spin_unlock(&rq->lock); 420} 421 422/* 423 * Called to set the hrtick timer state. 424 * 425 * called with rq->lock held and irqs disabled 426 */ 427void hrtick_start(struct rq *rq, u64 delay) 428{ 429 struct hrtimer *timer = &rq->hrtick_timer; 430 ktime_t time = ktime_add_ns(timer->base->get_time(), delay); 431 432 hrtimer_set_expires(timer, time); 433 434 if (rq == this_rq()) { 435 __hrtick_restart(rq); 436 } else if (!rq->hrtick_csd_pending) { 437 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd, 0); 438 rq->hrtick_csd_pending = 1; 439 } 440} 441 442static int 443hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu) 444{ 445 int cpu = (int)(long)hcpu; 446 447 switch (action) { 448 case CPU_UP_CANCELED: 449 case CPU_UP_CANCELED_FROZEN: 450 case CPU_DOWN_PREPARE: 451 case CPU_DOWN_PREPARE_FROZEN: 452 case CPU_DEAD: 453 case CPU_DEAD_FROZEN: 454 hrtick_clear(cpu_rq(cpu)); 455 return NOTIFY_OK; 456 } 457 458 return NOTIFY_DONE; 459} 460 461static __init void init_hrtick(void) 462{ 463 hotcpu_notifier(hotplug_hrtick, 0); 464} 465#else 466/* 467 * Called to set the hrtick timer state. 468 * 469 * called with rq->lock held and irqs disabled 470 */ 471void hrtick_start(struct rq *rq, u64 delay) 472{ 473 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0, 474 HRTIMER_MODE_REL_PINNED, 0); 475} 476 477static inline void init_hrtick(void) 478{ 479} 480#endif /* CONFIG_SMP */ 481 482static void init_rq_hrtick(struct rq *rq) 483{ 484#ifdef CONFIG_SMP 485 rq->hrtick_csd_pending = 0; 486 487 rq->hrtick_csd.flags = 0; 488 rq->hrtick_csd.func = __hrtick_start; 489 rq->hrtick_csd.info = rq; 490#endif 491 492 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); 493 rq->hrtick_timer.function = hrtick; 494} 495#else /* CONFIG_SCHED_HRTICK */ 496static inline void hrtick_clear(struct rq *rq) 497{ 498} 499 500static inline void init_rq_hrtick(struct rq *rq) 501{ 502} 503 504static inline void init_hrtick(void) 505{ 506} 507#endif /* CONFIG_SCHED_HRTICK */ 508 509/* 510 * resched_task - mark a task 'to be rescheduled now'. 511 * 512 * On UP this means the setting of the need_resched flag, on SMP it 513 * might also involve a cross-CPU call to trigger the scheduler on 514 * the target CPU. 515 */ 516#ifdef CONFIG_SMP 517void resched_task(struct task_struct *p) 518{ 519 int cpu; 520 521 assert_raw_spin_locked(&task_rq(p)->lock); 522 523 if (test_tsk_need_resched(p)) 524 return; 525 526 set_tsk_need_resched(p); 527 528 cpu = task_cpu(p); 529 if (cpu == smp_processor_id()) 530 return; 531 532 /* NEED_RESCHED must be visible before we test polling */ 533 smp_mb(); 534 if (!tsk_is_polling(p)) 535 smp_send_reschedule(cpu); 536} 537 538void resched_cpu(int cpu) 539{ 540 struct rq *rq = cpu_rq(cpu); 541 unsigned long flags; 542 543 if (!raw_spin_trylock_irqsave(&rq->lock, flags)) 544 return; 545 resched_task(cpu_curr(cpu)); 546 raw_spin_unlock_irqrestore(&rq->lock, flags); 547} 548 549#ifdef CONFIG_NO_HZ_COMMON 550/* 551 * In the semi idle case, use the nearest busy cpu for migrating timers 552 * from an idle cpu. This is good for power-savings. 553 * 554 * We don't do similar optimization for completely idle system, as 555 * selecting an idle cpu will add more delays to the timers than intended 556 * (as that cpu's timer base may not be uptodate wrt jiffies etc). 557 */ 558int get_nohz_timer_target(void) 559{ 560 int cpu = smp_processor_id(); 561 int i; 562 struct sched_domain *sd; 563 564 rcu_read_lock(); 565 for_each_domain(cpu, sd) { 566 for_each_cpu(i, sched_domain_span(sd)) { 567 if (!idle_cpu(i)) { 568 cpu = i; 569 goto unlock; 570 } 571 } 572 } 573unlock: 574 rcu_read_unlock(); 575 return cpu; 576} 577/* 578 * When add_timer_on() enqueues a timer into the timer wheel of an 579 * idle CPU then this timer might expire before the next timer event 580 * which is scheduled to wake up that CPU. In case of a completely 581 * idle system the next event might even be infinite time into the 582 * future. wake_up_idle_cpu() ensures that the CPU is woken up and 583 * leaves the inner idle loop so the newly added timer is taken into 584 * account when the CPU goes back to idle and evaluates the timer 585 * wheel for the next timer event. 586 */ 587static void wake_up_idle_cpu(int cpu) 588{ 589 struct rq *rq = cpu_rq(cpu); 590 591 if (cpu == smp_processor_id()) 592 return; 593 594 /* 595 * This is safe, as this function is called with the timer 596 * wheel base lock of (cpu) held. When the CPU is on the way 597 * to idle and has not yet set rq->curr to idle then it will 598 * be serialized on the timer wheel base lock and take the new 599 * timer into account automatically. 600 */ 601 if (rq->curr != rq->idle) 602 return; 603 604 /* 605 * We can set TIF_RESCHED on the idle task of the other CPU 606 * lockless. The worst case is that the other CPU runs the 607 * idle task through an additional NOOP schedule() 608 */ 609 set_tsk_need_resched(rq->idle); 610 611 /* NEED_RESCHED must be visible before we test polling */ 612 smp_mb(); 613 if (!tsk_is_polling(rq->idle)) 614 smp_send_reschedule(cpu); 615} 616 617static bool wake_up_full_nohz_cpu(int cpu) 618{ 619 if (tick_nohz_full_cpu(cpu)) { 620 if (cpu != smp_processor_id() || 621 tick_nohz_tick_stopped()) 622 smp_send_reschedule(cpu); 623 return true; 624 } 625 626 return false; 627} 628 629void wake_up_nohz_cpu(int cpu) 630{ 631 if (!wake_up_full_nohz_cpu(cpu)) 632 wake_up_idle_cpu(cpu); 633} 634 635static inline bool got_nohz_idle_kick(void) 636{ 637 int cpu = smp_processor_id(); 638 639 if (!test_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu))) 640 return false; 641 642 if (idle_cpu(cpu) && !need_resched()) 643 return true; 644 645 /* 646 * We can't run Idle Load Balance on this CPU for this time so we 647 * cancel it and clear NOHZ_BALANCE_KICK 648 */ 649 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu)); 650 return false; 651} 652 653#else /* CONFIG_NO_HZ_COMMON */ 654 655static inline bool got_nohz_idle_kick(void) 656{ 657 return false; 658} 659 660#endif /* CONFIG_NO_HZ_COMMON */ 661 662#ifdef CONFIG_NO_HZ_FULL 663bool sched_can_stop_tick(void) 664{ 665 struct rq *rq; 666 667 rq = this_rq(); 668 669 /* Make sure rq->nr_running update is visible after the IPI */ 670 smp_rmb(); 671 672 /* More than one running task need preemption */ 673 if (rq->nr_running > 1) 674 return false; 675 676 return true; 677} 678#endif /* CONFIG_NO_HZ_FULL */ 679 680void sched_avg_update(struct rq *rq) 681{ 682 s64 period = sched_avg_period(); 683 684 while ((s64)(rq_clock(rq) - rq->age_stamp) > period) { 685 /* 686 * Inline assembly required to prevent the compiler 687 * optimising this loop into a divmod call. 688 * See __iter_div_u64_rem() for another example of this. 689 */ 690 asm("" : "+rm" (rq->age_stamp)); 691 rq->age_stamp += period; 692 rq->rt_avg /= 2; 693 } 694} 695 696#else /* !CONFIG_SMP */ 697void resched_task(struct task_struct *p) 698{ 699 assert_raw_spin_locked(&task_rq(p)->lock); 700 set_tsk_need_resched(p); 701} 702#endif /* CONFIG_SMP */ 703 704#if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \ 705 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH))) 706/* 707 * Iterate task_group tree rooted at *from, calling @down when first entering a 708 * node and @up when leaving it for the final time. 709 * 710 * Caller must hold rcu_lock or sufficient equivalent. 711 */ 712int walk_tg_tree_from(struct task_group *from, 713 tg_visitor down, tg_visitor up, void *data) 714{ 715 struct task_group *parent, *child; 716 int ret; 717 718 parent = from; 719 720down: 721 ret = (*down)(parent, data); 722 if (ret) 723 goto out; 724 list_for_each_entry_rcu(child, &parent->children, siblings) { 725 parent = child; 726 goto down; 727 728up: 729 continue; 730 } 731 ret = (*up)(parent, data); 732 if (ret || parent == from) 733 goto out; 734 735 child = parent; 736 parent = parent->parent; 737 if (parent) 738 goto up; 739out: 740 return ret; 741} 742 743int tg_nop(struct task_group *tg, void *data) 744{ 745 return 0; 746} 747#endif 748 749static void set_load_weight(struct task_struct *p) 750{ 751 int prio = p->static_prio - MAX_RT_PRIO; 752 struct load_weight *load = &p->se.load; 753 754 /* 755 * SCHED_IDLE tasks get minimal weight: 756 */ 757 if (p->policy == SCHED_IDLE) { 758 load->weight = scale_load(WEIGHT_IDLEPRIO); 759 load->inv_weight = WMULT_IDLEPRIO; 760 return; 761 } 762 763 load->weight = scale_load(prio_to_weight[prio]); 764 load->inv_weight = prio_to_wmult[prio]; 765} 766 767static void enqueue_task(struct rq *rq, struct task_struct *p, int flags) 768{ 769 update_rq_clock(rq); 770 sched_info_queued(p); 771 p->sched_class->enqueue_task(rq, p, flags); 772} 773 774static void dequeue_task(struct rq *rq, struct task_struct *p, int flags) 775{ 776 update_rq_clock(rq); 777 sched_info_dequeued(p); 778 p->sched_class->dequeue_task(rq, p, flags); 779} 780 781void activate_task(struct rq *rq, struct task_struct *p, int flags) 782{ 783 if (task_contributes_to_load(p)) 784 rq->nr_uninterruptible--; 785 786 enqueue_task(rq, p, flags); 787} 788 789void deactivate_task(struct rq *rq, struct task_struct *p, int flags) 790{ 791 if (task_contributes_to_load(p)) 792 rq->nr_uninterruptible++; 793 794 dequeue_task(rq, p, flags); 795} 796 797static void update_rq_clock_task(struct rq *rq, s64 delta) 798{ 799/* 800 * In theory, the compile should just see 0 here, and optimize out the call 801 * to sched_rt_avg_update. But I don't trust it... 802 */ 803#if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING) 804 s64 steal = 0, irq_delta = 0; 805#endif 806#ifdef CONFIG_IRQ_TIME_ACCOUNTING 807 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time; 808 809 /* 810 * Since irq_time is only updated on {soft,}irq_exit, we might run into 811 * this case when a previous update_rq_clock() happened inside a 812 * {soft,}irq region. 813 * 814 * When this happens, we stop ->clock_task and only update the 815 * prev_irq_time stamp to account for the part that fit, so that a next 816 * update will consume the rest. This ensures ->clock_task is 817 * monotonic. 818 * 819 * It does however cause some slight miss-attribution of {soft,}irq 820 * time, a more accurate solution would be to update the irq_time using 821 * the current rq->clock timestamp, except that would require using 822 * atomic ops. 823 */ 824 if (irq_delta > delta) 825 irq_delta = delta; 826 827 rq->prev_irq_time += irq_delta; 828 delta -= irq_delta; 829#endif 830#ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING 831 if (static_key_false((¶virt_steal_rq_enabled))) { 832 u64 st; 833 834 steal = paravirt_steal_clock(cpu_of(rq)); 835 steal -= rq->prev_steal_time_rq; 836 837 if (unlikely(steal > delta)) 838 steal = delta; 839 840 st = steal_ticks(steal); 841 steal = st * TICK_NSEC; 842 843 rq->prev_steal_time_rq += steal; 844 845 delta -= steal; 846 } 847#endif 848 849 rq->clock_task += delta; 850 851#if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING) 852 if ((irq_delta + steal) && sched_feat(NONTASK_POWER)) 853 sched_rt_avg_update(rq, irq_delta + steal); 854#endif 855} 856 857void sched_set_stop_task(int cpu, struct task_struct *stop) 858{ 859 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 }; 860 struct task_struct *old_stop = cpu_rq(cpu)->stop; 861 862 if (stop) { 863 /* 864 * Make it appear like a SCHED_FIFO task, its something 865 * userspace knows about and won't get confused about. 866 * 867 * Also, it will make PI more or less work without too 868 * much confusion -- but then, stop work should not 869 * rely on PI working anyway. 870 */ 871 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m); 872 873 stop->sched_class = &stop_sched_class; 874 } 875 876 cpu_rq(cpu)->stop = stop; 877 878 if (old_stop) { 879 /* 880 * Reset it back to a normal scheduling class so that 881 * it can die in pieces. 882 */ 883 old_stop->sched_class = &rt_sched_class; 884 } 885} 886 887/* 888 * __normal_prio - return the priority that is based on the static prio 889 */ 890static inline int __normal_prio(struct task_struct *p) 891{ 892 return p->static_prio; 893} 894 895/* 896 * Calculate the expected normal priority: i.e. priority 897 * without taking RT-inheritance into account. Might be 898 * boosted by interactivity modifiers. Changes upon fork, 899 * setprio syscalls, and whenever the interactivity 900 * estimator recalculates. 901 */ 902static inline int normal_prio(struct task_struct *p) 903{ 904 int prio; 905 906 if (task_has_rt_policy(p)) 907 prio = MAX_RT_PRIO-1 - p->rt_priority; 908 else 909 prio = __normal_prio(p); 910 return prio; 911} 912 913/* 914 * Calculate the current priority, i.e. the priority 915 * taken into account by the scheduler. This value might 916 * be boosted by RT tasks, or might be boosted by 917 * interactivity modifiers. Will be RT if the task got 918 * RT-boosted. If not then it returns p->normal_prio. 919 */ 920static int effective_prio(struct task_struct *p) 921{ 922 p->normal_prio = normal_prio(p); 923 /* 924 * If we are RT tasks or we were boosted to RT priority, 925 * keep the priority unchanged. Otherwise, update priority 926 * to the normal priority: 927 */ 928 if (!rt_prio(p->prio)) 929 return p->normal_prio; 930 return p->prio; 931} 932 933/** 934 * task_curr - is this task currently executing on a CPU? 935 * @p: the task in question. 936 */ 937inline int task_curr(const struct task_struct *p) 938{ 939 return cpu_curr(task_cpu(p)) == p; 940} 941 942static inline void check_class_changed(struct rq *rq, struct task_struct *p, 943 const struct sched_class *prev_class, 944 int oldprio) 945{ 946 if (prev_class != p->sched_class) { 947 if (prev_class->switched_from) 948 prev_class->switched_from(rq, p); 949 p->sched_class->switched_to(rq, p); 950 } else if (oldprio != p->prio) 951 p->sched_class->prio_changed(rq, p, oldprio); 952} 953 954void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags) 955{ 956 const struct sched_class *class; 957 958 if (p->sched_class == rq->curr->sched_class) { 959 rq->curr->sched_class->check_preempt_curr(rq, p, flags); 960 } else { 961 for_each_class(class) { 962 if (class == rq->curr->sched_class) 963 break; 964 if (class == p->sched_class) { 965 resched_task(rq->curr); 966 break; 967 } 968 } 969 } 970 971 /* 972 * A queue event has occurred, and we're going to schedule. In 973 * this case, we can save a useless back to back clock update. 974 */ 975 if (rq->curr->on_rq && test_tsk_need_resched(rq->curr)) 976 rq->skip_clock_update = 1; 977} 978 979static ATOMIC_NOTIFIER_HEAD(task_migration_notifier); 980 981void register_task_migration_notifier(struct notifier_block *n) 982{ 983 atomic_notifier_chain_register(&task_migration_notifier, n); 984} 985 986#ifdef CONFIG_SMP 987void set_task_cpu(struct task_struct *p, unsigned int new_cpu) 988{ 989#ifdef CONFIG_SCHED_DEBUG 990 /* 991 * We should never call set_task_cpu() on a blocked task, 992 * ttwu() will sort out the placement. 993 */ 994 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING && 995 !(task_thread_info(p)->preempt_count & PREEMPT_ACTIVE)); 996 997#ifdef CONFIG_LOCKDEP 998 /* 999 * The caller should hold either p->pi_lock or rq->lock, when changing 1000 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks. 1001 * 1002 * sched_move_task() holds both and thus holding either pins the cgroup, 1003 * see task_group(). 1004 * 1005 * Furthermore, all task_rq users should acquire both locks, see 1006 * task_rq_lock(). 1007 */ 1008 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) || 1009 lockdep_is_held(&task_rq(p)->lock))); 1010#endif 1011#endif 1012 1013 trace_sched_migrate_task(p, new_cpu); 1014 1015 if (task_cpu(p) != new_cpu) { 1016 struct task_migration_notifier tmn; 1017 1018 if (p->sched_class->migrate_task_rq) 1019 p->sched_class->migrate_task_rq(p, new_cpu); 1020 p->se.nr_migrations++; 1021 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, NULL, 0); 1022 1023 tmn.task = p; 1024 tmn.from_cpu = task_cpu(p); 1025 tmn.to_cpu = new_cpu; 1026 1027 atomic_notifier_call_chain(&task_migration_notifier, 0, &tmn); 1028 } 1029 1030 __set_task_cpu(p, new_cpu); 1031} 1032 1033struct migration_arg { 1034 struct task_struct *task; 1035 int dest_cpu; 1036}; 1037 1038static int migration_cpu_stop(void *data); 1039 1040/* 1041 * wait_task_inactive - wait for a thread to unschedule. 1042 * 1043 * If @match_state is nonzero, it's the @p->state value just checked and 1044 * not expected to change. If it changes, i.e. @p might have woken up, 1045 * then return zero. When we succeed in waiting for @p to be off its CPU, 1046 * we return a positive number (its total switch count). If a second call 1047 * a short while later returns the same number, the caller can be sure that 1048 * @p has remained unscheduled the whole time. 1049 * 1050 * The caller must ensure that the task *will* unschedule sometime soon, 1051 * else this function might spin for a *long* time. This function can't 1052 * be called with interrupts off, or it may introduce deadlock with 1053 * smp_call_function() if an IPI is sent by the same process we are 1054 * waiting to become inactive. 1055 */ 1056unsigned long wait_task_inactive(struct task_struct *p, long match_state) 1057{ 1058 unsigned long flags; 1059 int running, on_rq; 1060 unsigned long ncsw; 1061 struct rq *rq; 1062 1063 for (;;) { 1064 /* 1065 * We do the initial early heuristics without holding 1066 * any task-queue locks at all. We'll only try to get 1067 * the runqueue lock when things look like they will 1068 * work out! 1069 */ 1070 rq = task_rq(p); 1071 1072 /* 1073 * If the task is actively running on another CPU 1074 * still, just relax and busy-wait without holding 1075 * any locks. 1076 * 1077 * NOTE! Since we don't hold any locks, it's not 1078 * even sure that "rq" stays as the right runqueue! 1079 * But we don't care, since "task_running()" will 1080 * return false if the runqueue has changed and p 1081 * is actually now running somewhere else! 1082 */ 1083 while (task_running(rq, p)) { 1084 if (match_state && unlikely(p->state != match_state)) 1085 return 0; 1086 cpu_relax(); 1087 } 1088 1089 /* 1090 * Ok, time to look more closely! We need the rq 1091 * lock now, to be *sure*. If we're wrong, we'll 1092 * just go back and repeat. 1093 */ 1094 rq = task_rq_lock(p, &flags); 1095 trace_sched_wait_task(p); 1096 running = task_running(rq, p); 1097 on_rq = p->on_rq; 1098 ncsw = 0; 1099 if (!match_state || p->state == match_state) 1100 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */ 1101 task_rq_unlock(rq, p, &flags); 1102 1103 /* 1104 * If it changed from the expected state, bail out now. 1105 */ 1106 if (unlikely(!ncsw)) 1107 break; 1108 1109 /* 1110 * Was it really running after all now that we 1111 * checked with the proper locks actually held? 1112 * 1113 * Oops. Go back and try again.. 1114 */ 1115 if (unlikely(running)) { 1116 cpu_relax(); 1117 continue; 1118 } 1119 1120 /* 1121 * It's not enough that it's not actively running, 1122 * it must be off the runqueue _entirely_, and not 1123 * preempted! 1124 * 1125 * So if it was still runnable (but just not actively 1126 * running right now), it's preempted, and we should 1127 * yield - it could be a while. 1128 */ 1129 if (unlikely(on_rq)) { 1130 ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ); 1131 1132 set_current_state(TASK_UNINTERRUPTIBLE); 1133 schedule_hrtimeout(&to, HRTIMER_MODE_REL); 1134 continue; 1135 } 1136 1137 /* 1138 * Ahh, all good. It wasn't running, and it wasn't 1139 * runnable, which means that it will never become 1140 * running in the future either. We're all done! 1141 */ 1142 break; 1143 } 1144 1145 return ncsw; 1146} 1147 1148/*** 1149 * kick_process - kick a running thread to enter/exit the kernel 1150 * @p: the to-be-kicked thread 1151 * 1152 * Cause a process which is running on another CPU to enter 1153 * kernel-mode, without any delay. (to get signals handled.) 1154 * 1155 * NOTE: this function doesn't have to take the runqueue lock, 1156 * because all it wants to ensure is that the remote task enters 1157 * the kernel. If the IPI races and the task has been migrated 1158 * to another CPU then no harm is done and the purpose has been 1159 * achieved as well. 1160 */ 1161void kick_process(struct task_struct *p) 1162{ 1163 int cpu; 1164 1165 preempt_disable(); 1166 cpu = task_cpu(p); 1167 if ((cpu != smp_processor_id()) && task_curr(p)) 1168 smp_send_reschedule(cpu); 1169 preempt_enable(); 1170} 1171EXPORT_SYMBOL_GPL(kick_process); 1172#endif /* CONFIG_SMP */ 1173 1174#ifdef CONFIG_SMP 1175/* 1176 * ->cpus_allowed is protected by both rq->lock and p->pi_lock 1177 */ 1178static int select_fallback_rq(int cpu, struct task_struct *p) 1179{ 1180 int nid = cpu_to_node(cpu); 1181 const struct cpumask *nodemask = NULL; 1182 enum { cpuset, possible, fail } state = cpuset; 1183 int dest_cpu; 1184 1185 /* 1186 * If the node that the cpu is on has been offlined, cpu_to_node() 1187 * will return -1. There is no cpu on the node, and we should 1188 * select the cpu on the other node. 1189 */ 1190 if (nid != -1) { 1191 nodemask = cpumask_of_node(nid); 1192 1193 /* Look for allowed, online CPU in same node. */ 1194 for_each_cpu(dest_cpu, nodemask) { 1195 if (!cpu_online(dest_cpu)) 1196 continue; 1197 if (!cpu_active(dest_cpu)) 1198 continue; 1199 if (cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p))) 1200 return dest_cpu; 1201 } 1202 } 1203 1204 for (;;) { 1205 /* Any allowed, online CPU? */ 1206 for_each_cpu(dest_cpu, tsk_cpus_allowed(p)) { 1207 if (!cpu_online(dest_cpu)) 1208 continue; 1209 if (!cpu_active(dest_cpu)) 1210 continue; 1211 goto out; 1212 } 1213 1214 switch (state) { 1215 case cpuset: 1216 /* No more Mr. Nice Guy. */ 1217 cpuset_cpus_allowed_fallback(p); 1218 state = possible; 1219 break; 1220 1221 case possible: 1222 do_set_cpus_allowed(p, cpu_possible_mask); 1223 state = fail; 1224 break; 1225 1226 case fail: 1227 BUG(); 1228 break; 1229 } 1230 } 1231 1232out: 1233 if (state != cpuset) { 1234 /* 1235 * Don't tell them about moving exiting tasks or 1236 * kernel threads (both mm NULL), since they never 1237 * leave kernel. 1238 */ 1239 if (p->mm && printk_ratelimit()) { 1240 printk_sched("process %d (%s) no longer affine to cpu%d\n", 1241 task_pid_nr(p), p->comm, cpu); 1242 } 1243 } 1244 1245 return dest_cpu; 1246} 1247 1248/* 1249 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable. 1250 */ 1251static inline 1252int select_task_rq(struct task_struct *p, int sd_flags, int wake_flags) 1253{ 1254 int cpu = p->sched_class->select_task_rq(p, sd_flags, wake_flags); 1255 1256 /* 1257 * In order not to call set_task_cpu() on a blocking task we need 1258 * to rely on ttwu() to place the task on a valid ->cpus_allowed 1259 * cpu. 1260 * 1261 * Since this is common to all placement strategies, this lives here. 1262 * 1263 * [ this allows ->select_task() to simply return task_cpu(p) and 1264 * not worry about this generic constraint ] 1265 */ 1266 if (unlikely(!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)) || 1267 !cpu_online(cpu))) 1268 cpu = select_fallback_rq(task_cpu(p), p); 1269 1270 return cpu; 1271} 1272 1273static void update_avg(u64 *avg, u64 sample) 1274{ 1275 s64 diff = sample - *avg; 1276 *avg += diff >> 3; 1277} 1278#endif 1279 1280static void 1281ttwu_stat(struct task_struct *p, int cpu, int wake_flags) 1282{ 1283#ifdef CONFIG_SCHEDSTATS 1284 struct rq *rq = this_rq(); 1285 1286#ifdef CONFIG_SMP 1287 int this_cpu = smp_processor_id(); 1288 1289 if (cpu == this_cpu) { 1290 schedstat_inc(rq, ttwu_local); 1291 schedstat_inc(p, se.statistics.nr_wakeups_local); 1292 } else { 1293 struct sched_domain *sd; 1294 1295 schedstat_inc(p, se.statistics.nr_wakeups_remote); 1296 rcu_read_lock(); 1297 for_each_domain(this_cpu, sd) { 1298 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) { 1299 schedstat_inc(sd, ttwu_wake_remote); 1300 break; 1301 } 1302 } 1303 rcu_read_unlock(); 1304 } 1305 1306 if (wake_flags & WF_MIGRATED) 1307 schedstat_inc(p, se.statistics.nr_wakeups_migrate); 1308 1309#endif /* CONFIG_SMP */ 1310 1311 schedstat_inc(rq, ttwu_count); 1312 schedstat_inc(p, se.statistics.nr_wakeups); 1313 1314 if (wake_flags & WF_SYNC) 1315 schedstat_inc(p, se.statistics.nr_wakeups_sync); 1316 1317#endif /* CONFIG_SCHEDSTATS */ 1318} 1319 1320static void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags) 1321{ 1322 activate_task(rq, p, en_flags); 1323 p->on_rq = 1; 1324 1325 /* if a worker is waking up, notify workqueue */ 1326 if (p->flags & PF_WQ_WORKER) 1327 wq_worker_waking_up(p, cpu_of(rq)); 1328} 1329 1330/* 1331 * Mark the task runnable and perform wakeup-preemption. 1332 */ 1333static void 1334ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags) 1335{ 1336 check_preempt_curr(rq, p, wake_flags); 1337 trace_sched_wakeup(p, true); 1338 1339 p->state = TASK_RUNNING; 1340#ifdef CONFIG_SMP 1341 if (p->sched_class->task_woken) 1342 p->sched_class->task_woken(rq, p); 1343 1344 if (rq->idle_stamp) { 1345 u64 delta = rq_clock(rq) - rq->idle_stamp; 1346 u64 max = 2*sysctl_sched_migration_cost; 1347 1348 if (delta > max) 1349 rq->avg_idle = max; 1350 else 1351 update_avg(&rq->avg_idle, delta); 1352 rq->idle_stamp = 0; 1353 } 1354#endif 1355} 1356 1357static void 1358ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags) 1359{ 1360#ifdef CONFIG_SMP 1361 if (p->sched_contributes_to_load) 1362 rq->nr_uninterruptible--; 1363#endif 1364 1365 ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_WAKING); 1366 ttwu_do_wakeup(rq, p, wake_flags); 1367} 1368 1369/* 1370 * Called in case the task @p isn't fully descheduled from its runqueue, 1371 * in this case we must do a remote wakeup. Its a 'light' wakeup though, 1372 * since all we need to do is flip p->state to TASK_RUNNING, since 1373 * the task is still ->on_rq. 1374 */ 1375static int ttwu_remote(struct task_struct *p, int wake_flags) 1376{ 1377 struct rq *rq; 1378 int ret = 0; 1379 1380 rq = __task_rq_lock(p); 1381 if (p->on_rq) { 1382 /* check_preempt_curr() may use rq clock */ 1383 update_rq_clock(rq); 1384 ttwu_do_wakeup(rq, p, wake_flags); 1385 ret = 1; 1386 } 1387 __task_rq_unlock(rq); 1388 1389 return ret; 1390} 1391 1392#ifdef CONFIG_SMP 1393static void sched_ttwu_pending(void) 1394{ 1395 struct rq *rq = this_rq(); 1396 struct llist_node *llist = llist_del_all(&rq->wake_list); 1397 struct task_struct *p; 1398 1399 raw_spin_lock(&rq->lock); 1400 1401 while (llist) { 1402 p = llist_entry(llist, struct task_struct, wake_entry); 1403 llist = llist_next(llist); 1404 ttwu_do_activate(rq, p, 0); 1405 } 1406 1407 raw_spin_unlock(&rq->lock); 1408} 1409 1410void scheduler_ipi(void) 1411{ 1412 if (llist_empty(&this_rq()->wake_list) 1413 && !tick_nohz_full_cpu(smp_processor_id()) 1414 && !got_nohz_idle_kick()) 1415 return; 1416 1417 /* 1418 * Not all reschedule IPI handlers call irq_enter/irq_exit, since 1419 * traditionally all their work was done from the interrupt return 1420 * path. Now that we actually do some work, we need to make sure 1421 * we do call them. 1422 * 1423 * Some archs already do call them, luckily irq_enter/exit nest 1424 * properly. 1425 * 1426 * Arguably we should visit all archs and update all handlers, 1427 * however a fair share of IPIs are still resched only so this would 1428 * somewhat pessimize the simple resched case. 1429 */ 1430 irq_enter(); 1431 tick_nohz_full_check(); 1432 sched_ttwu_pending(); 1433 1434 /* 1435 * Check if someone kicked us for doing the nohz idle load balance. 1436 */ 1437 if (unlikely(got_nohz_idle_kick())) { 1438 this_rq()->idle_balance = 1; 1439 raise_softirq_irqoff(SCHED_SOFTIRQ); 1440 } 1441 irq_exit(); 1442} 1443 1444static void ttwu_queue_remote(struct task_struct *p, int cpu) 1445{ 1446 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list)) 1447 smp_send_reschedule(cpu); 1448} 1449 1450bool cpus_share_cache(int this_cpu, int that_cpu) 1451{ 1452 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu); 1453} 1454#endif /* CONFIG_SMP */ 1455 1456static void ttwu_queue(struct task_struct *p, int cpu) 1457{ 1458 struct rq *rq = cpu_rq(cpu); 1459 1460#if defined(CONFIG_SMP) 1461 if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) { 1462 sched_clock_cpu(cpu); /* sync clocks x-cpu */ 1463 ttwu_queue_remote(p, cpu); 1464 return; 1465 } 1466#endif 1467 1468 raw_spin_lock(&rq->lock); 1469 ttwu_do_activate(rq, p, 0); 1470 raw_spin_unlock(&rq->lock); 1471} 1472 1473/** 1474 * try_to_wake_up - wake up a thread 1475 * @p: the thread to be awakened 1476 * @state: the mask of task states that can be woken 1477 * @wake_flags: wake modifier flags (WF_*) 1478 * 1479 * Put it on the run-queue if it's not already there. The "current" 1480 * thread is always on the run-queue (except when the actual 1481 * re-schedule is in progress), and as such you're allowed to do 1482 * the simpler "current->state = TASK_RUNNING" to mark yourself 1483 * runnable without the overhead of this. 1484 * 1485 * Returns %true if @p was woken up, %false if it was already running 1486 * or @state didn't match @p's state. 1487 */ 1488static int 1489try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags) 1490{ 1491 unsigned long flags; 1492 int cpu, success = 0; 1493 1494 smp_wmb(); 1495 raw_spin_lock_irqsave(&p->pi_lock, flags); 1496 if (!(p->state & state)) 1497 goto out; 1498 1499 success = 1; /* we're going to change ->state */ 1500 cpu = task_cpu(p); 1501 1502 if (p->on_rq && ttwu_remote(p, wake_flags)) 1503 goto stat; 1504 1505#ifdef CONFIG_SMP 1506 /* 1507 * If the owning (remote) cpu is still in the middle of schedule() with 1508 * this task as prev, wait until its done referencing the task. 1509 */ 1510 while (p->on_cpu) 1511 cpu_relax(); 1512 /* 1513 * Pairs with the smp_wmb() in finish_lock_switch(). 1514 */ 1515 smp_rmb(); 1516 1517 p->sched_contributes_to_load = !!task_contributes_to_load(p); 1518 p->state = TASK_WAKING; 1519 1520 if (p->sched_class->task_waking) 1521 p->sched_class->task_waking(p); 1522 1523 cpu = select_task_rq(p, SD_BALANCE_WAKE, wake_flags); 1524 if (task_cpu(p) != cpu) { 1525 wake_flags |= WF_MIGRATED; 1526 set_task_cpu(p, cpu); 1527 } 1528#endif /* CONFIG_SMP */ 1529 1530 ttwu_queue(p, cpu); 1531stat: 1532 ttwu_stat(p, cpu, wake_flags); 1533out: 1534 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 1535 1536 return success; 1537} 1538 1539/** 1540 * try_to_wake_up_local - try to wake up a local task with rq lock held 1541 * @p: the thread to be awakened 1542 * 1543 * Put @p on the run-queue if it's not already there. The caller must 1544 * ensure that this_rq() is locked, @p is bound to this_rq() and not 1545 * the current task. 1546 */ 1547static void try_to_wake_up_local(struct task_struct *p) 1548{ 1549 struct rq *rq = task_rq(p); 1550 1551 if (WARN_ON_ONCE(rq != this_rq()) || 1552 WARN_ON_ONCE(p == current)) 1553 return; 1554 1555 lockdep_assert_held(&rq->lock); 1556 1557 if (!raw_spin_trylock(&p->pi_lock)) { 1558 raw_spin_unlock(&rq->lock); 1559 raw_spin_lock(&p->pi_lock); 1560 raw_spin_lock(&rq->lock); 1561 } 1562 1563 if (!(p->state & TASK_NORMAL)) 1564 goto out; 1565 1566 if (!p->on_rq) 1567 ttwu_activate(rq, p, ENQUEUE_WAKEUP); 1568 1569 ttwu_do_wakeup(rq, p, 0); 1570 ttwu_stat(p, smp_processor_id(), 0); 1571out: 1572 raw_spin_unlock(&p->pi_lock); 1573} 1574 1575/** 1576 * wake_up_process - Wake up a specific process 1577 * @p: The process to be woken up. 1578 * 1579 * Attempt to wake up the nominated process and move it to the set of runnable 1580 * processes. Returns 1 if the process was woken up, 0 if it was already 1581 * running. 1582 * 1583 * It may be assumed that this function implies a write memory barrier before 1584 * changing the task state if and only if any tasks are woken up. 1585 */ 1586int wake_up_process(struct task_struct *p) 1587{ 1588 WARN_ON(task_is_stopped_or_traced(p)); 1589 return try_to_wake_up(p, TASK_NORMAL, 0); 1590} 1591EXPORT_SYMBOL(wake_up_process); 1592 1593int wake_up_state(struct task_struct *p, unsigned int state) 1594{ 1595 return try_to_wake_up(p, state, 0); 1596} 1597 1598/* 1599 * Perform scheduler related setup for a newly forked process p. 1600 * p is forked by current. 1601 * 1602 * __sched_fork() is basic setup used by init_idle() too: 1603 */ 1604static void __sched_fork(struct task_struct *p) 1605{ 1606 p->on_rq = 0; 1607 1608 p->se.on_rq = 0; 1609 p->se.exec_start = 0; 1610 p->se.sum_exec_runtime = 0; 1611 p->se.prev_sum_exec_runtime = 0; 1612 p->se.nr_migrations = 0; 1613 p->se.vruntime = 0; 1614 INIT_LIST_HEAD(&p->se.group_node); 1615 1616#ifdef CONFIG_SCHEDSTATS 1617 memset(&p->se.statistics, 0, sizeof(p->se.statistics)); 1618#endif 1619 1620 INIT_LIST_HEAD(&p->rt.run_list); 1621 1622#ifdef CONFIG_PREEMPT_NOTIFIERS 1623 INIT_HLIST_HEAD(&p->preempt_notifiers); 1624#endif 1625 1626#ifdef CONFIG_NUMA_BALANCING 1627 if (p->mm && atomic_read(&p->mm->mm_users) == 1) { 1628 p->mm->numa_next_scan = jiffies; 1629 p->mm->numa_next_reset = jiffies; 1630 p->mm->numa_scan_seq = 0; 1631 } 1632 1633 p->node_stamp = 0ULL; 1634 p->numa_scan_seq = p->mm ? p->mm->numa_scan_seq : 0; 1635 p->numa_migrate_seq = p->mm ? p->mm->numa_scan_seq - 1 : 0; 1636 p->numa_scan_period = sysctl_numa_balancing_scan_delay; 1637 p->numa_work.next = &p->numa_work; 1638#endif /* CONFIG_NUMA_BALANCING */ 1639} 1640 1641#ifdef CONFIG_NUMA_BALANCING 1642#ifdef CONFIG_SCHED_DEBUG 1643void set_numabalancing_state(bool enabled) 1644{ 1645 if (enabled) 1646 sched_feat_set("NUMA"); 1647 else 1648 sched_feat_set("NO_NUMA"); 1649} 1650#else 1651__read_mostly bool numabalancing_enabled; 1652 1653void set_numabalancing_state(bool enabled) 1654{ 1655 numabalancing_enabled = enabled; 1656} 1657#endif /* CONFIG_SCHED_DEBUG */ 1658#endif /* CONFIG_NUMA_BALANCING */ 1659 1660/* 1661 * fork()/clone()-time setup: 1662 */ 1663void sched_fork(struct task_struct *p) 1664{ 1665 unsigned long flags; 1666 int cpu = get_cpu(); 1667 1668 __sched_fork(p); 1669 /* 1670 * We mark the process as running here. This guarantees that 1671 * nobody will actually run it, and a signal or other external 1672 * event cannot wake it up and insert it on the runqueue either. 1673 */ 1674 p->state = TASK_RUNNING; 1675 1676 /* 1677 * Make sure we do not leak PI boosting priority to the child. 1678 */ 1679 p->prio = current->normal_prio; 1680 1681 /* 1682 * Revert to default priority/policy on fork if requested. 1683 */ 1684 if (unlikely(p->sched_reset_on_fork)) { 1685 if (task_has_rt_policy(p)) { 1686 p->policy = SCHED_NORMAL; 1687 p->static_prio = NICE_TO_PRIO(0); 1688 p->rt_priority = 0; 1689 } else if (PRIO_TO_NICE(p->static_prio) < 0) 1690 p->static_prio = NICE_TO_PRIO(0); 1691 1692 p->prio = p->normal_prio = __normal_prio(p); 1693 set_load_weight(p); 1694 1695 /* 1696 * We don't need the reset flag anymore after the fork. It has 1697 * fulfilled its duty: 1698 */ 1699 p->sched_reset_on_fork = 0; 1700 } 1701 1702 if (!rt_prio(p->prio)) 1703 p->sched_class = &fair_sched_class; 1704 1705 if (p->sched_class->task_fork) 1706 p->sched_class->task_fork(p); 1707 1708 /* 1709 * The child is not yet in the pid-hash so no cgroup attach races, 1710 * and the cgroup is pinned to this child due to cgroup_fork() 1711 * is ran before sched_fork(). 1712 * 1713 * Silence PROVE_RCU. 1714 */ 1715 raw_spin_lock_irqsave(&p->pi_lock, flags); 1716 set_task_cpu(p, cpu); 1717 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 1718 1719#if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT) 1720 if (likely(sched_info_on())) 1721 memset(&p->sched_info, 0, sizeof(p->sched_info)); 1722#endif 1723#if defined(CONFIG_SMP) 1724 p->on_cpu = 0; 1725#endif 1726#ifdef CONFIG_PREEMPT_COUNT 1727 /* Want to start with kernel preemption disabled. */ 1728 task_thread_info(p)->preempt_count = 1; 1729#endif 1730#ifdef CONFIG_SMP 1731 plist_node_init(&p->pushable_tasks, MAX_PRIO); 1732#endif 1733 1734 put_cpu(); 1735} 1736 1737/* 1738 * wake_up_new_task - wake up a newly created task for the first time. 1739 * 1740 * This function will do some initial scheduler statistics housekeeping 1741 * that must be done for every newly created context, then puts the task 1742 * on the runqueue and wakes it. 1743 */ 1744void wake_up_new_task(struct task_struct *p) 1745{ 1746 unsigned long flags; 1747 struct rq *rq; 1748 1749 raw_spin_lock_irqsave(&p->pi_lock, flags); 1750#ifdef CONFIG_SMP 1751 /* 1752 * Fork balancing, do it here and not earlier because: 1753 * - cpus_allowed can change in the fork path 1754 * - any previously selected cpu might disappear through hotplug 1755 */ 1756 set_task_cpu(p, select_task_rq(p, SD_BALANCE_FORK, 0)); 1757#endif 1758 1759 /* Initialize new task's runnable average */ 1760 init_task_runnable_average(p); 1761 rq = __task_rq_lock(p); 1762 activate_task(rq, p, 0); 1763 p->on_rq = 1; 1764 trace_sched_wakeup_new(p, true); 1765 check_preempt_curr(rq, p, WF_FORK); 1766#ifdef CONFIG_SMP 1767 if (p->sched_class->task_woken) 1768 p->sched_class->task_woken(rq, p); 1769#endif 1770 task_rq_unlock(rq, p, &flags); 1771} 1772 1773#ifdef CONFIG_PREEMPT_NOTIFIERS 1774 1775/** 1776 * preempt_notifier_register - tell me when current is being preempted & rescheduled 1777 * @notifier: notifier struct to register 1778 */ 1779void preempt_notifier_register(struct preempt_notifier *notifier) 1780{ 1781 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers); 1782} 1783EXPORT_SYMBOL_GPL(preempt_notifier_register); 1784 1785/** 1786 * preempt_notifier_unregister - no longer interested in preemption notifications 1787 * @notifier: notifier struct to unregister 1788 * 1789 * This is safe to call from within a preemption notifier. 1790 */ 1791void preempt_notifier_unregister(struct preempt_notifier *notifier) 1792{ 1793 hlist_del(¬ifier->link); 1794} 1795EXPORT_SYMBOL_GPL(preempt_notifier_unregister); 1796 1797static void fire_sched_in_preempt_notifiers(struct task_struct *curr) 1798{ 1799 struct preempt_notifier *notifier; 1800 1801 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link) 1802 notifier->ops->sched_in(notifier, raw_smp_processor_id()); 1803} 1804 1805static void 1806fire_sched_out_preempt_notifiers(struct task_struct *curr, 1807 struct task_struct *next) 1808{ 1809 struct preempt_notifier *notifier; 1810 1811 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link) 1812 notifier->ops->sched_out(notifier, next); 1813} 1814 1815#else /* !CONFIG_PREEMPT_NOTIFIERS */ 1816 1817static void fire_sched_in_preempt_notifiers(struct task_struct *curr) 1818{ 1819} 1820 1821static void 1822fire_sched_out_preempt_notifiers(struct task_struct *curr, 1823 struct task_struct *next) 1824{ 1825} 1826 1827#endif /* CONFIG_PREEMPT_NOTIFIERS */ 1828 1829/** 1830 * prepare_task_switch - prepare to switch tasks 1831 * @rq: the runqueue preparing to switch 1832 * @prev: the current task that is being switched out 1833 * @next: the task we are going to switch to. 1834 * 1835 * This is called with the rq lock held and interrupts off. It must 1836 * be paired with a subsequent finish_task_switch after the context 1837 * switch. 1838 * 1839 * prepare_task_switch sets up locking and calls architecture specific 1840 * hooks. 1841 */ 1842static inline void 1843prepare_task_switch(struct rq *rq, struct task_struct *prev, 1844 struct task_struct *next) 1845{ 1846 trace_sched_switch(prev, next); 1847 sched_info_switch(prev, next); 1848 perf_event_task_sched_out(prev, next); 1849 fire_sched_out_preempt_notifiers(prev, next); 1850 prepare_lock_switch(rq, next); 1851 prepare_arch_switch(next); 1852} 1853 1854/** 1855 * finish_task_switch - clean up after a task-switch 1856 * @rq: runqueue associated with task-switch 1857 * @prev: the thread we just switched away from. 1858 * 1859 * finish_task_switch must be called after the context switch, paired 1860 * with a prepare_task_switch call before the context switch. 1861 * finish_task_switch will reconcile locking set up by prepare_task_switch, 1862 * and do any other architecture-specific cleanup actions. 1863 * 1864 * Note that we may have delayed dropping an mm in context_switch(). If 1865 * so, we finish that here outside of the runqueue lock. (Doing it 1866 * with the lock held can cause deadlocks; see schedule() for 1867 * details.) 1868 */ 1869static void finish_task_switch(struct rq *rq, struct task_struct *prev) 1870 __releases(rq->lock) 1871{ 1872 struct mm_struct *mm = rq->prev_mm; 1873 long prev_state; 1874 1875 rq->prev_mm = NULL; 1876 1877 /* 1878 * A task struct has one reference for the use as "current". 1879 * If a task dies, then it sets TASK_DEAD in tsk->state and calls 1880 * schedule one last time. The schedule call will never return, and 1881 * the scheduled task must drop that reference. 1882 * The test for TASK_DEAD must occur while the runqueue locks are 1883 * still held, otherwise prev could be scheduled on another cpu, die 1884 * there before we look at prev->state, and then the reference would 1885 * be dropped twice. 1886 * Manfred Spraul <manfred@colorfullife.com> 1887 */ 1888 prev_state = prev->state; 1889 vtime_task_switch(prev); 1890 finish_arch_switch(prev); 1891 perf_event_task_sched_in(prev, current); 1892 finish_lock_switch(rq, prev); 1893 finish_arch_post_lock_switch(); 1894 1895 fire_sched_in_preempt_notifiers(current); 1896 if (mm) 1897 mmdrop(mm); 1898 if (unlikely(prev_state == TASK_DEAD)) { 1899 /* 1900 * Remove function-return probe instances associated with this 1901 * task and put them back on the free list. 1902 */ 1903 kprobe_flush_task(prev); 1904 put_task_struct(prev); 1905 } 1906 1907 tick_nohz_task_switch(current); 1908} 1909 1910#ifdef CONFIG_SMP 1911 1912/* assumes rq->lock is held */ 1913static inline void pre_schedule(struct rq *rq, struct task_struct *prev) 1914{ 1915 if (prev->sched_class->pre_schedule) 1916 prev->sched_class->pre_schedule(rq, prev); 1917} 1918 1919/* rq->lock is NOT held, but preemption is disabled */ 1920static inline void post_schedule(struct rq *rq) 1921{ 1922 if (rq->post_schedule) { 1923 unsigned long flags; 1924 1925 raw_spin_lock_irqsave(&rq->lock, flags); 1926 if (rq->curr->sched_class->post_schedule) 1927 rq->curr->sched_class->post_schedule(rq); 1928 raw_spin_unlock_irqrestore(&rq->lock, flags); 1929 1930 rq->post_schedule = 0; 1931 } 1932} 1933 1934#else 1935 1936static inline void pre_schedule(struct rq *rq, struct task_struct *p) 1937{ 1938} 1939 1940static inline void post_schedule(struct rq *rq) 1941{ 1942} 1943 1944#endif 1945 1946/** 1947 * schedule_tail - first thing a freshly forked thread must call. 1948 * @prev: the thread we just switched away from. 1949 */ 1950asmlinkage void schedule_tail(struct task_struct *prev) 1951 __releases(rq->lock) 1952{ 1953 struct rq *rq = this_rq(); 1954 1955 finish_task_switch(rq, prev); 1956 1957 /* 1958 * FIXME: do we need to worry about rq being invalidated by the 1959 * task_switch? 1960 */ 1961 post_schedule(rq); 1962 1963#ifdef __ARCH_WANT_UNLOCKED_CTXSW 1964 /* In this case, finish_task_switch does not reenable preemption */ 1965 preempt_enable(); 1966#endif 1967 if (current->set_child_tid) 1968 put_user(task_pid_vnr(current), current->set_child_tid); 1969} 1970 1971/* 1972 * context_switch - switch to the new MM and the new 1973 * thread's register state. 1974 */ 1975static inline void 1976context_switch(struct rq *rq, struct task_struct *prev, 1977 struct task_struct *next) 1978{ 1979 struct mm_struct *mm, *oldmm; 1980 1981 prepare_task_switch(rq, prev, next); 1982 1983 mm = next->mm; 1984 oldmm = prev->active_mm; 1985 /* 1986 * For paravirt, this is coupled with an exit in switch_to to 1987 * combine the page table reload and the switch backend into 1988 * one hypercall. 1989 */ 1990 arch_start_context_switch(prev); 1991 1992 if (!mm) { 1993 next->active_mm = oldmm; 1994 atomic_inc(&oldmm->mm_count); 1995 enter_lazy_tlb(oldmm, next); 1996 } else 1997 switch_mm(oldmm, mm, next); 1998 1999 if (!prev->mm) { 2000 prev->active_mm = NULL; 2001 rq->prev_mm = oldmm; 2002 } 2003 /* 2004 * Since the runqueue lock will be released by the next 2005 * task (which is an invalid locking op but in the case 2006 * of the scheduler it's an obvious special-case), so we 2007 * do an early lockdep release here: 2008 */ 2009#ifndef __ARCH_WANT_UNLOCKED_CTXSW 2010 spin_release(&rq->lock.dep_map, 1, _THIS_IP_); 2011#endif 2012 2013 context_tracking_task_switch(prev, next); 2014 /* Here we just switch the register state and the stack. */ 2015 switch_to(prev, next, prev); 2016 2017 barrier(); 2018 /* 2019 * this_rq must be evaluated again because prev may have moved 2020 * CPUs since it called schedule(), thus the 'rq' on its stack 2021 * frame will be invalid. 2022 */ 2023 finish_task_switch(this_rq(), prev); 2024} 2025 2026/* 2027 * nr_running and nr_context_switches: 2028 * 2029 * externally visible scheduler statistics: current number of runnable 2030 * threads, total number of context switches performed since bootup. 2031 */ 2032unsigned long nr_running(void) 2033{ 2034 unsigned long i, sum = 0; 2035 2036 for_each_online_cpu(i) 2037 sum += cpu_rq(i)->nr_running; 2038 2039 return sum; 2040} 2041 2042unsigned long long nr_context_switches(void) 2043{ 2044 int i; 2045 unsigned long long sum = 0; 2046 2047 for_each_possible_cpu(i) 2048 sum += cpu_rq(i)->nr_switches; 2049 2050 return sum; 2051} 2052 2053unsigned long nr_iowait(void) 2054{ 2055 unsigned long i, sum = 0; 2056 2057 for_each_possible_cpu(i) 2058 sum += atomic_read(&cpu_rq(i)->nr_iowait); 2059 2060 return sum; 2061} 2062 2063unsigned long nr_iowait_cpu(int cpu) 2064{ 2065 struct rq *this = cpu_rq(cpu); 2066 return atomic_read(&this->nr_iowait); 2067} 2068 2069#ifdef CONFIG_SMP 2070 2071/* 2072 * sched_exec - execve() is a valuable balancing opportunity, because at 2073 * this point the task has the smallest effective memory and cache footprint. 2074 */ 2075void sched_exec(void) 2076{ 2077 struct task_struct *p = current; 2078 unsigned long flags; 2079 int dest_cpu; 2080 2081 raw_spin_lock_irqsave(&p->pi_lock, flags); 2082 dest_cpu = p->sched_class->select_task_rq(p, SD_BALANCE_EXEC, 0); 2083 if (dest_cpu == smp_processor_id()) 2084 goto unlock; 2085 2086 if (likely(cpu_active(dest_cpu))) { 2087 struct migration_arg arg = { p, dest_cpu }; 2088 2089 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 2090 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg); 2091 return; 2092 } 2093unlock: 2094 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 2095} 2096 2097#endif 2098 2099DEFINE_PER_CPU(struct kernel_stat, kstat); 2100DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat); 2101 2102EXPORT_PER_CPU_SYMBOL(kstat); 2103EXPORT_PER_CPU_SYMBOL(kernel_cpustat); 2104 2105/* 2106 * Return any ns on the sched_clock that have not yet been accounted in 2107 * @p in case that task is currently running. 2108 * 2109 * Called with task_rq_lock() held on @rq. 2110 */ 2111static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq) 2112{ 2113 u64 ns = 0; 2114 2115 if (task_current(rq, p)) { 2116 update_rq_clock(rq); 2117 ns = rq_clock_task(rq) - p->se.exec_start; 2118 if ((s64)ns < 0) 2119 ns = 0; 2120 } 2121 2122 return ns; 2123} 2124 2125unsigned long long task_delta_exec(struct task_struct *p) 2126{ 2127 unsigned long flags; 2128 struct rq *rq; 2129 u64 ns = 0; 2130 2131 rq = task_rq_lock(p, &flags); 2132 ns = do_task_delta_exec(p, rq); 2133 task_rq_unlock(rq, p, &flags); 2134 2135 return ns; 2136} 2137 2138/* 2139 * Return accounted runtime for the task. 2140 * In case the task is currently running, return the runtime plus current's 2141 * pending runtime that have not been accounted yet. 2142 */ 2143unsigned long long task_sched_runtime(struct task_struct *p) 2144{ 2145 unsigned long flags; 2146 struct rq *rq; 2147 u64 ns = 0; 2148 2149 rq = task_rq_lock(p, &flags); 2150 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq); 2151 task_rq_unlock(rq, p, &flags); 2152 2153 return ns; 2154} 2155 2156/* 2157 * This function gets called by the timer code, with HZ frequency. 2158 * We call it with interrupts disabled. 2159 */ 2160void scheduler_tick(void) 2161{ 2162 int cpu = smp_processor_id(); 2163 struct rq *rq = cpu_rq(cpu); 2164 struct task_struct *curr = rq->curr; 2165 2166 sched_clock_tick(); 2167 2168 raw_spin_lock(&rq->lock); 2169 update_rq_clock(rq); 2170 curr->sched_class->task_tick(rq, curr, 0); 2171 update_cpu_load_active(rq); 2172 raw_spin_unlock(&rq->lock); 2173 2174 perf_event_task_tick(); 2175 2176#ifdef CONFIG_SMP 2177 rq->idle_balance = idle_cpu(cpu); 2178 trigger_load_balance(rq, cpu); 2179#endif 2180 rq_last_tick_reset(rq); 2181} 2182 2183#ifdef CONFIG_NO_HZ_FULL 2184/** 2185 * scheduler_tick_max_deferment 2186 * 2187 * Keep at least one tick per second when a single 2188 * active task is running because the scheduler doesn't 2189 * yet completely support full dynticks environment. 2190 * 2191 * This makes sure that uptime, CFS vruntime, load 2192 * balancing, etc... continue to move forward, even 2193 * with a very low granularity. 2194 */ 2195u64 scheduler_tick_max_deferment(void) 2196{ 2197 struct rq *rq = this_rq(); 2198 unsigned long next, now = ACCESS_ONCE(jiffies); 2199 2200 next = rq->last_sched_tick + HZ; 2201 2202 if (time_before_eq(next, now)) 2203 return 0; 2204 2205 return jiffies_to_usecs(next - now) * NSEC_PER_USEC; 2206} 2207#endif 2208 2209notrace unsigned long get_parent_ip(unsigned long addr) 2210{ 2211 if (in_lock_functions(addr)) { 2212 addr = CALLER_ADDR2; 2213 if (in_lock_functions(addr)) 2214 addr = CALLER_ADDR3; 2215 } 2216 return addr; 2217} 2218 2219#if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \ 2220 defined(CONFIG_PREEMPT_TRACER)) 2221 2222void __kprobes add_preempt_count(int val) 2223{ 2224#ifdef CONFIG_DEBUG_PREEMPT 2225 /* 2226 * Underflow? 2227 */ 2228 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0))) 2229 return; 2230#endif 2231 preempt_count() += val; 2232#ifdef CONFIG_DEBUG_PREEMPT 2233 /* 2234 * Spinlock count overflowing soon? 2235 */ 2236 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >= 2237 PREEMPT_MASK - 10); 2238#endif 2239 if (preempt_count() == val) 2240 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1)); 2241} 2242EXPORT_SYMBOL(add_preempt_count); 2243 2244void __kprobes sub_preempt_count(int val) 2245{ 2246#ifdef CONFIG_DEBUG_PREEMPT 2247 /* 2248 * Underflow? 2249 */ 2250 if (DEBUG_LOCKS_WARN_ON(val > preempt_count())) 2251 return; 2252 /* 2253 * Is the spinlock portion underflowing? 2254 */ 2255 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) && 2256 !(preempt_count() & PREEMPT_MASK))) 2257 return; 2258#endif 2259 2260 if (preempt_count() == val) 2261 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1)); 2262 preempt_count() -= val; 2263} 2264EXPORT_SYMBOL(sub_preempt_count); 2265 2266#endif 2267 2268/* 2269 * Print scheduling while atomic bug: 2270 */ 2271static noinline void __schedule_bug(struct task_struct *prev) 2272{ 2273 if (oops_in_progress) 2274 return; 2275 2276 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n", 2277 prev->comm, prev->pid, preempt_count()); 2278 2279 debug_show_held_locks(prev); 2280 print_modules(); 2281 if (irqs_disabled()) 2282 print_irqtrace_events(prev); 2283 dump_stack(); 2284 add_taint(TAINT_WARN, LOCKDEP_STILL_OK); 2285} 2286 2287/* 2288 * Various schedule()-time debugging checks and statistics: 2289 */ 2290static inline void schedule_debug(struct task_struct *prev) 2291{ 2292 /* 2293 * Test if we are atomic. Since do_exit() needs to call into 2294 * schedule() atomically, we ignore that path for now. 2295 * Otherwise, whine if we are scheduling when we should not be. 2296 */ 2297 if (unlikely(in_atomic_preempt_off() && !prev->exit_state)) 2298 __schedule_bug(prev); 2299 rcu_sleep_check(); 2300 2301 profile_hit(SCHED_PROFILING, __builtin_return_address(0)); 2302 2303 schedstat_inc(this_rq(), sched_count); 2304} 2305 2306static void put_prev_task(struct rq *rq, struct task_struct *prev) 2307{ 2308 if (prev->on_rq || rq->skip_clock_update < 0) 2309 update_rq_clock(rq); 2310 prev->sched_class->put_prev_task(rq, prev); 2311} 2312 2313/* 2314 * Pick up the highest-prio task: 2315 */ 2316static inline struct task_struct * 2317pick_next_task(struct rq *rq) 2318{ 2319 const struct sched_class *class; 2320 struct task_struct *p; 2321 2322 /* 2323 * Optimization: we know that if all tasks are in 2324 * the fair class we can call that function directly: 2325 */ 2326 if (likely(rq->nr_running == rq->cfs.h_nr_running)) { 2327 p = fair_sched_class.pick_next_task(rq); 2328 if (likely(p)) 2329 return p; 2330 } 2331 2332 for_each_class(class) { 2333 p = class->pick_next_task(rq); 2334 if (p) 2335 return p; 2336 } 2337 2338 BUG(); /* the idle class will always have a runnable task */ 2339} 2340 2341/* 2342 * __schedule() is the main scheduler function. 2343 * 2344 * The main means of driving the scheduler and thus entering this function are: 2345 * 2346 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc. 2347 * 2348 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return 2349 * paths. For example, see arch/x86/entry_64.S. 2350 * 2351 * To drive preemption between tasks, the scheduler sets the flag in timer 2352 * interrupt handler scheduler_tick(). 2353 * 2354 * 3. Wakeups don't really cause entry into schedule(). They add a 2355 * task to the run-queue and that's it. 2356 * 2357 * Now, if the new task added to the run-queue preempts the current 2358 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets 2359 * called on the nearest possible occasion: 2360 * 2361 * - If the kernel is preemptible (CONFIG_PREEMPT=y): 2362 * 2363 * - in syscall or exception context, at the next outmost 2364 * preempt_enable(). (this might be as soon as the wake_up()'s 2365 * spin_unlock()!) 2366 * 2367 * - in IRQ context, return from interrupt-handler to 2368 * preemptible context 2369 * 2370 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set) 2371 * then at the next: 2372 * 2373 * - cond_resched() call 2374 * - explicit schedule() call 2375 * - return from syscall or exception to user-space 2376 * - return from interrupt-handler to user-space 2377 */ 2378static void __sched __schedule(void) 2379{ 2380 struct task_struct *prev, *next; 2381 unsigned long *switch_count; 2382 struct rq *rq; 2383 int cpu; 2384 2385need_resched: 2386 preempt_disable(); 2387 cpu = smp_processor_id(); 2388 rq = cpu_rq(cpu); 2389 rcu_note_context_switch(cpu); 2390 prev = rq->curr; 2391 2392 schedule_debug(prev); 2393 2394 if (sched_feat(HRTICK)) 2395 hrtick_clear(rq); 2396 2397 raw_spin_lock_irq(&rq->lock); 2398 2399 switch_count = &prev->nivcsw; 2400 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) { 2401 if (unlikely(signal_pending_state(prev->state, prev))) { 2402 prev->state = TASK_RUNNING; 2403 } else { 2404 deactivate_task(rq, prev, DEQUEUE_SLEEP); 2405 prev->on_rq = 0; 2406 2407 /* 2408 * If a worker went to sleep, notify and ask workqueue 2409 * whether it wants to wake up a task to maintain 2410 * concurrency. 2411 */ 2412 if (prev->flags & PF_WQ_WORKER) { 2413 struct task_struct *to_wakeup; 2414 2415 to_wakeup = wq_worker_sleeping(prev, cpu); 2416 if (to_wakeup) 2417 try_to_wake_up_local(to_wakeup); 2418 } 2419 } 2420 switch_count = &prev->nvcsw; 2421 } 2422 2423 pre_schedule(rq, prev); 2424 2425 if (unlikely(!rq->nr_running)) 2426 idle_balance(cpu, rq); 2427 2428 put_prev_task(rq, prev); 2429 next = pick_next_task(rq); 2430 clear_tsk_need_resched(prev); 2431 rq->skip_clock_update = 0; 2432 2433 if (likely(prev != next)) { 2434 rq->nr_switches++; 2435 rq->curr = next; 2436 ++*switch_count; 2437 2438 context_switch(rq, prev, next); /* unlocks the rq */ 2439 /* 2440 * The context switch have flipped the stack from under us 2441 * and restored the local variables which were saved when 2442 * this task called schedule() in the past. prev == current 2443 * is still correct, but it can be moved to another cpu/rq. 2444 */ 2445 cpu = smp_processor_id(); 2446 rq = cpu_rq(cpu); 2447 } else 2448 raw_spin_unlock_irq(&rq->lock); 2449 2450 post_schedule(rq); 2451 2452 sched_preempt_enable_no_resched(); 2453 if (need_resched()) 2454 goto need_resched; 2455} 2456 2457static inline void sched_submit_work(struct task_struct *tsk) 2458{ 2459 if (!tsk->state || tsk_is_pi_blocked(tsk)) 2460 return; 2461 /* 2462 * If we are going to sleep and we have plugged IO queued, 2463 * make sure to submit it to avoid deadlocks. 2464 */ 2465 if (blk_needs_flush_plug(tsk)) 2466 blk_schedule_flush_plug(tsk); 2467} 2468 2469asmlinkage void __sched schedule(void) 2470{ 2471 struct task_struct *tsk = current; 2472 2473 sched_submit_work(tsk); 2474 __schedule(); 2475} 2476EXPORT_SYMBOL(schedule); 2477 2478#ifdef CONFIG_CONTEXT_TRACKING 2479asmlinkage void __sched schedule_user(void) 2480{ 2481 /* 2482 * If we come here after a random call to set_need_resched(), 2483 * or we have been woken up remotely but the IPI has not yet arrived, 2484 * we haven't yet exited the RCU idle mode. Do it here manually until 2485 * we find a better solution. 2486 */ 2487 user_exit(); 2488 schedule(); 2489 user_enter(); 2490} 2491#endif 2492 2493/** 2494 * schedule_preempt_disabled - called with preemption disabled 2495 * 2496 * Returns with preemption disabled. Note: preempt_count must be 1 2497 */ 2498void __sched schedule_preempt_disabled(void) 2499{ 2500 sched_preempt_enable_no_resched(); 2501 schedule(); 2502 preempt_disable(); 2503} 2504 2505#ifdef CONFIG_PREEMPT 2506/* 2507 * this is the entry point to schedule() from in-kernel preemption 2508 * off of preempt_enable. Kernel preemptions off return from interrupt 2509 * occur there and call schedule directly. 2510 */ 2511asmlinkage void __sched notrace preempt_schedule(void) 2512{ 2513 struct thread_info *ti = current_thread_info(); 2514 2515 /* 2516 * If there is a non-zero preempt_count or interrupts are disabled, 2517 * we do not want to preempt the current task. Just return.. 2518 */ 2519 if (likely(ti->preempt_count || irqs_disabled())) 2520 return; 2521 2522 do { 2523 add_preempt_count_notrace(PREEMPT_ACTIVE); 2524 __schedule(); 2525 sub_preempt_count_notrace(PREEMPT_ACTIVE); 2526 2527 /* 2528 * Check again in case we missed a preemption opportunity 2529 * between schedule and now. 2530 */ 2531 barrier(); 2532 } while (need_resched()); 2533} 2534EXPORT_SYMBOL(preempt_schedule); 2535 2536/* 2537 * this is the entry point to schedule() from kernel preemption 2538 * off of irq context. 2539 * Note, that this is called and return with irqs disabled. This will 2540 * protect us against recursive calling from irq. 2541 */ 2542asmlinkage void __sched preempt_schedule_irq(void) 2543{ 2544 struct thread_info *ti = current_thread_info(); 2545 enum ctx_state prev_state; 2546 2547 /* Catch callers which need to be fixed */ 2548 BUG_ON(ti->preempt_count || !irqs_disabled()); 2549 2550 prev_state = exception_enter(); 2551 2552 do { 2553 add_preempt_count(PREEMPT_ACTIVE); 2554 local_irq_enable(); 2555 __schedule(); 2556 local_irq_disable(); 2557 sub_preempt_count(PREEMPT_ACTIVE); 2558 2559 /* 2560 * Check again in case we missed a preemption opportunity 2561 * between schedule and now. 2562 */ 2563 barrier(); 2564 } while (need_resched()); 2565 2566 exception_exit(prev_state); 2567} 2568 2569#endif /* CONFIG_PREEMPT */ 2570 2571int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags, 2572 void *key) 2573{ 2574 return try_to_wake_up(curr->private, mode, wake_flags); 2575} 2576EXPORT_SYMBOL(default_wake_function); 2577 2578/* 2579 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just 2580 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve 2581 * number) then we wake all the non-exclusive tasks and one exclusive task. 2582 * 2583 * There are circumstances in which we can try to wake a task which has already 2584 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns 2585 * zero in this (rare) case, and we handle it by continuing to scan the queue. 2586 */ 2587static void __wake_up_common(wait_queue_head_t *q, unsigned int mode, 2588 int nr_exclusive, int wake_flags, void *key) 2589{ 2590 wait_queue_t *curr, *next; 2591 2592 list_for_each_entry_safe(curr, next, &q->task_list, task_list) { 2593 unsigned flags = curr->flags; 2594 2595 if (curr->func(curr, mode, wake_flags, key) && 2596 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive) 2597 break; 2598 } 2599} 2600 2601/** 2602 * __wake_up - wake up threads blocked on a waitqueue. 2603 * @q: the waitqueue 2604 * @mode: which threads 2605 * @nr_exclusive: how many wake-one or wake-many threads to wake up 2606 * @key: is directly passed to the wakeup function 2607 * 2608 * It may be assumed that this function implies a write memory barrier before 2609 * changing the task state if and only if any tasks are woken up. 2610 */ 2611void __wake_up(wait_queue_head_t *q, unsigned int mode, 2612 int nr_exclusive, void *key) 2613{ 2614 unsigned long flags; 2615 2616 spin_lock_irqsave(&q->lock, flags); 2617 __wake_up_common(q, mode, nr_exclusive, 0, key); 2618 spin_unlock_irqrestore(&q->lock, flags); 2619} 2620EXPORT_SYMBOL(__wake_up); 2621 2622/* 2623 * Same as __wake_up but called with the spinlock in wait_queue_head_t held. 2624 */ 2625void __wake_up_locked(wait_queue_head_t *q, unsigned int mode, int nr) 2626{ 2627 __wake_up_common(q, mode, nr, 0, NULL); 2628} 2629EXPORT_SYMBOL_GPL(__wake_up_locked); 2630 2631void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key) 2632{ 2633 __wake_up_common(q, mode, 1, 0, key); 2634} 2635EXPORT_SYMBOL_GPL(__wake_up_locked_key); 2636 2637/** 2638 * __wake_up_sync_key - wake up threads blocked on a waitqueue. 2639 * @q: the waitqueue 2640 * @mode: which threads 2641 * @nr_exclusive: how many wake-one or wake-many threads to wake up 2642 * @key: opaque value to be passed to wakeup targets 2643 * 2644 * The sync wakeup differs that the waker knows that it will schedule 2645 * away soon, so while the target thread will be woken up, it will not 2646 * be migrated to another CPU - ie. the two threads are 'synchronized' 2647 * with each other. This can prevent needless bouncing between CPUs. 2648 * 2649 * On UP it can prevent extra preemption. 2650 * 2651 * It may be assumed that this function implies a write memory barrier before 2652 * changing the task state if and only if any tasks are woken up. 2653 */ 2654void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode, 2655 int nr_exclusive, void *key) 2656{ 2657 unsigned long flags; 2658 int wake_flags = WF_SYNC; 2659 2660 if (unlikely(!q)) 2661 return; 2662 2663 if (unlikely(!nr_exclusive)) 2664 wake_flags = 0; 2665 2666 spin_lock_irqsave(&q->lock, flags); 2667 __wake_up_common(q, mode, nr_exclusive, wake_flags, key); 2668 spin_unlock_irqrestore(&q->lock, flags); 2669} 2670EXPORT_SYMBOL_GPL(__wake_up_sync_key); 2671 2672/* 2673 * __wake_up_sync - see __wake_up_sync_key() 2674 */ 2675void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive) 2676{ 2677 __wake_up_sync_key(q, mode, nr_exclusive, NULL); 2678} 2679EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */ 2680 2681/** 2682 * complete: - signals a single thread waiting on this completion 2683 * @x: holds the state of this particular completion 2684 * 2685 * This will wake up a single thread waiting on this completion. Threads will be 2686 * awakened in the same order in which they were queued. 2687 * 2688 * See also complete_all(), wait_for_completion() and related routines. 2689 * 2690 * It may be assumed that this function implies a write memory barrier before 2691 * changing the task state if and only if any tasks are woken up. 2692 */ 2693void complete(struct completion *x) 2694{ 2695 unsigned long flags; 2696 2697 spin_lock_irqsave(&x->wait.lock, flags); 2698 x->done++; 2699 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL); 2700 spin_unlock_irqrestore(&x->wait.lock, flags); 2701} 2702EXPORT_SYMBOL(complete); 2703 2704/** 2705 * complete_all: - signals all threads waiting on this completion 2706 * @x: holds the state of this particular completion 2707 * 2708 * This will wake up all threads waiting on this particular completion event. 2709 * 2710 * It may be assumed that this function implies a write memory barrier before 2711 * changing the task state if and only if any tasks are woken up. 2712 */ 2713void complete_all(struct completion *x) 2714{ 2715 unsigned long flags; 2716 2717 spin_lock_irqsave(&x->wait.lock, flags); 2718 x->done += UINT_MAX/2; 2719 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL); 2720 spin_unlock_irqrestore(&x->wait.lock, flags); 2721} 2722EXPORT_SYMBOL(complete_all); 2723 2724static inline long __sched 2725do_wait_for_common(struct completion *x, 2726 long (*action)(long), long timeout, int state) 2727{ 2728 if (!x->done) { 2729 DECLARE_WAITQUEUE(wait, current); 2730 2731 __add_wait_queue_tail_exclusive(&x->wait, &wait); 2732 do { 2733 if (signal_pending_state(state, current)) { 2734 timeout = -ERESTARTSYS; 2735 break; 2736 } 2737 __set_current_state(state); 2738 spin_unlock_irq(&x->wait.lock); 2739 timeout = action(timeout); 2740 spin_lock_irq(&x->wait.lock); 2741 } while (!x->done && timeout); 2742 __remove_wait_queue(&x->wait, &wait); 2743 if (!x->done) 2744 return timeout; 2745 } 2746 x->done--; 2747 return timeout ?: 1; 2748} 2749 2750static inline long __sched 2751__wait_for_common(struct completion *x, 2752 long (*action)(long), long timeout, int state) 2753{ 2754 might_sleep(); 2755 2756 spin_lock_irq(&x->wait.lock); 2757 timeout = do_wait_for_common(x, action, timeout, state); 2758 spin_unlock_irq(&x->wait.lock); 2759 return timeout; 2760} 2761 2762static long __sched 2763wait_for_common(struct completion *x, long timeout, int state) 2764{ 2765 return __wait_for_common(x, schedule_timeout, timeout, state); 2766} 2767 2768static long __sched 2769wait_for_common_io(struct completion *x, long timeout, int state) 2770{ 2771 return __wait_for_common(x, io_schedule_timeout, timeout, state); 2772} 2773 2774/** 2775 * wait_for_completion: - waits for completion of a task 2776 * @x: holds the state of this particular completion 2777 * 2778 * This waits to be signaled for completion of a specific task. It is NOT 2779 * interruptible and there is no timeout. 2780 * 2781 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout 2782 * and interrupt capability. Also see complete(). 2783 */ 2784void __sched wait_for_completion(struct completion *x) 2785{ 2786 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE); 2787} 2788EXPORT_SYMBOL(wait_for_completion); 2789 2790/** 2791 * wait_for_completion_timeout: - waits for completion of a task (w/timeout) 2792 * @x: holds the state of this particular completion 2793 * @timeout: timeout value in jiffies 2794 * 2795 * This waits for either a completion of a specific task to be signaled or for a 2796 * specified timeout to expire. The timeout is in jiffies. It is not 2797 * interruptible. 2798 * 2799 * The return value is 0 if timed out, and positive (at least 1, or number of 2800 * jiffies left till timeout) if completed. 2801 */ 2802unsigned long __sched 2803wait_for_completion_timeout(struct completion *x, unsigned long timeout) 2804{ 2805 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE); 2806} 2807EXPORT_SYMBOL(wait_for_completion_timeout); 2808 2809/** 2810 * wait_for_completion_io: - waits for completion of a task 2811 * @x: holds the state of this particular completion 2812 * 2813 * This waits to be signaled for completion of a specific task. It is NOT 2814 * interruptible and there is no timeout. The caller is accounted as waiting 2815 * for IO. 2816 */ 2817void __sched wait_for_completion_io(struct completion *x) 2818{ 2819 wait_for_common_io(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE); 2820} 2821EXPORT_SYMBOL(wait_for_completion_io); 2822 2823/** 2824 * wait_for_completion_io_timeout: - waits for completion of a task (w/timeout) 2825 * @x: holds the state of this particular completion 2826 * @timeout: timeout value in jiffies 2827 * 2828 * This waits for either a completion of a specific task to be signaled or for a 2829 * specified timeout to expire. The timeout is in jiffies. It is not 2830 * interruptible. The caller is accounted as waiting for IO. 2831 * 2832 * The return value is 0 if timed out, and positive (at least 1, or number of 2833 * jiffies left till timeout) if completed. 2834 */ 2835unsigned long __sched 2836wait_for_completion_io_timeout(struct completion *x, unsigned long timeout) 2837{ 2838 return wait_for_common_io(x, timeout, TASK_UNINTERRUPTIBLE); 2839} 2840EXPORT_SYMBOL(wait_for_completion_io_timeout); 2841 2842/** 2843 * wait_for_completion_interruptible: - waits for completion of a task (w/intr) 2844 * @x: holds the state of this particular completion 2845 * 2846 * This waits for completion of a specific task to be signaled. It is 2847 * interruptible. 2848 * 2849 * The return value is -ERESTARTSYS if interrupted, 0 if completed. 2850 */ 2851int __sched wait_for_completion_interruptible(struct completion *x) 2852{ 2853 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE); 2854 if (t == -ERESTARTSYS) 2855 return t; 2856 return 0; 2857} 2858EXPORT_SYMBOL(wait_for_completion_interruptible); 2859 2860/** 2861 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr)) 2862 * @x: holds the state of this particular completion 2863 * @timeout: timeout value in jiffies 2864 * 2865 * This waits for either a completion of a specific task to be signaled or for a 2866 * specified timeout to expire. It is interruptible. The timeout is in jiffies. 2867 * 2868 * The return value is -ERESTARTSYS if interrupted, 0 if timed out, 2869 * positive (at least 1, or number of jiffies left till timeout) if completed. 2870 */ 2871long __sched 2872wait_for_completion_interruptible_timeout(struct completion *x, 2873 unsigned long timeout) 2874{ 2875 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE); 2876} 2877EXPORT_SYMBOL(wait_for_completion_interruptible_timeout); 2878 2879/** 2880 * wait_for_completion_killable: - waits for completion of a task (killable) 2881 * @x: holds the state of this particular completion 2882 * 2883 * This waits to be signaled for completion of a specific task. It can be 2884 * interrupted by a kill signal. 2885 * 2886 * The return value is -ERESTARTSYS if interrupted, 0 if completed. 2887 */ 2888int __sched wait_for_completion_killable(struct completion *x) 2889{ 2890 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE); 2891 if (t == -ERESTARTSYS) 2892 return t; 2893 return 0; 2894} 2895EXPORT_SYMBOL(wait_for_completion_killable); 2896 2897/** 2898 * wait_for_completion_killable_timeout: - waits for completion of a task (w/(to,killable)) 2899 * @x: holds the state of this particular completion 2900 * @timeout: timeout value in jiffies 2901 * 2902 * This waits for either a completion of a specific task to be 2903 * signaled or for a specified timeout to expire. It can be 2904 * interrupted by a kill signal. The timeout is in jiffies. 2905 * 2906 * The return value is -ERESTARTSYS if interrupted, 0 if timed out, 2907 * positive (at least 1, or number of jiffies left till timeout) if completed. 2908 */ 2909long __sched 2910wait_for_completion_killable_timeout(struct completion *x, 2911 unsigned long timeout) 2912{ 2913 return wait_for_common(x, timeout, TASK_KILLABLE); 2914} 2915EXPORT_SYMBOL(wait_for_completion_killable_timeout); 2916 2917/** 2918 * try_wait_for_completion - try to decrement a completion without blocking 2919 * @x: completion structure 2920 * 2921 * Returns: 0 if a decrement cannot be done without blocking 2922 * 1 if a decrement succeeded. 2923 * 2924 * If a completion is being used as a counting completion, 2925 * attempt to decrement the counter without blocking. This 2926 * enables us to avoid waiting if the resource the completion 2927 * is protecting is not available. 2928 */ 2929bool try_wait_for_completion(struct completion *x) 2930{ 2931 unsigned long flags; 2932 int ret = 1; 2933 2934 spin_lock_irqsave(&x->wait.lock, flags); 2935 if (!x->done) 2936 ret = 0; 2937 else 2938 x->done--; 2939 spin_unlock_irqrestore(&x->wait.lock, flags); 2940 return ret; 2941} 2942EXPORT_SYMBOL(try_wait_for_completion); 2943 2944/** 2945 * completion_done - Test to see if a completion has any waiters 2946 * @x: completion structure 2947 * 2948 * Returns: 0 if there are waiters (wait_for_completion() in progress) 2949 * 1 if there are no waiters. 2950 * 2951 */ 2952bool completion_done(struct completion *x) 2953{ 2954 unsigned long flags; 2955 int ret = 1; 2956 2957 spin_lock_irqsave(&x->wait.lock, flags); 2958 if (!x->done) 2959 ret = 0; 2960 spin_unlock_irqrestore(&x->wait.lock, flags); 2961 return ret; 2962} 2963EXPORT_SYMBOL(completion_done); 2964 2965static long __sched 2966sleep_on_common(wait_queue_head_t *q, int state, long timeout) 2967{ 2968 unsigned long flags; 2969 wait_queue_t wait; 2970 2971 init_waitqueue_entry(&wait, current); 2972 2973 __set_current_state(state); 2974 2975 spin_lock_irqsave(&q->lock, flags); 2976 __add_wait_queue(q, &wait); 2977 spin_unlock(&q->lock); 2978 timeout = schedule_timeout(timeout); 2979 spin_lock_irq(&q->lock); 2980 __remove_wait_queue(q, &wait); 2981 spin_unlock_irqrestore(&q->lock, flags); 2982 2983 return timeout; 2984} 2985 2986void __sched interruptible_sleep_on(wait_queue_head_t *q) 2987{ 2988 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT); 2989} 2990EXPORT_SYMBOL(interruptible_sleep_on); 2991 2992long __sched 2993interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout) 2994{ 2995 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout); 2996} 2997EXPORT_SYMBOL(interruptible_sleep_on_timeout); 2998 2999void __sched sleep_on(wait_queue_head_t *q) 3000{ 3001 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT); 3002} 3003EXPORT_SYMBOL(sleep_on); 3004 3005long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout) 3006{ 3007 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout); 3008} 3009EXPORT_SYMBOL(sleep_on_timeout); 3010 3011#ifdef CONFIG_RT_MUTEXES 3012 3013/* 3014 * rt_mutex_setprio - set the current priority of a task 3015 * @p: task 3016 * @prio: prio value (kernel-internal form) 3017 * 3018 * This function changes the 'effective' priority of a task. It does 3019 * not touch ->normal_prio like __setscheduler(). 3020 * 3021 * Used by the rt_mutex code to implement priority inheritance logic. 3022 */ 3023void rt_mutex_setprio(struct task_struct *p, int prio) 3024{ 3025 int oldprio, on_rq, running; 3026 struct rq *rq; 3027 const struct sched_class *prev_class; 3028 3029 BUG_ON(prio < 0 || prio > MAX_PRIO); 3030 3031 rq = __task_rq_lock(p); 3032 3033 /* 3034 * Idle task boosting is a nono in general. There is one 3035 * exception, when PREEMPT_RT and NOHZ is active: 3036 * 3037 * The idle task calls get_next_timer_interrupt() and holds 3038 * the timer wheel base->lock on the CPU and another CPU wants 3039 * to access the timer (probably to cancel it). We can safely 3040 * ignore the boosting request, as the idle CPU runs this code 3041 * with interrupts disabled and will complete the lock 3042 * protected section without being interrupted. So there is no 3043 * real need to boost. 3044 */ 3045 if (unlikely(p == rq->idle)) { 3046 WARN_ON(p != rq->curr); 3047 WARN_ON(p->pi_blocked_on); 3048 goto out_unlock; 3049 } 3050 3051 trace_sched_pi_setprio(p, prio); 3052 oldprio = p->prio; 3053 prev_class = p->sched_class; 3054 on_rq = p->on_rq; 3055 running = task_current(rq, p); 3056 if (on_rq) 3057 dequeue_task(rq, p, 0); 3058 if (running) 3059 p->sched_class->put_prev_task(rq, p); 3060 3061 if (rt_prio(prio)) 3062 p->sched_class = &rt_sched_class; 3063 else 3064 p->sched_class = &fair_sched_class; 3065 3066 p->prio = prio; 3067 3068 if (running) 3069 p->sched_class->set_curr_task(rq); 3070 if (on_rq) 3071 enqueue_task(rq, p, oldprio < prio ? ENQUEUE_HEAD : 0); 3072 3073 check_class_changed(rq, p, prev_class, oldprio); 3074out_unlock: 3075 __task_rq_unlock(rq); 3076} 3077#endif 3078void set_user_nice(struct task_struct *p, long nice) 3079{ 3080 int old_prio, delta, on_rq; 3081 unsigned long flags; 3082 struct rq *rq; 3083 3084 if (TASK_NICE(p) == nice || nice < -20 || nice > 19) 3085 return; 3086 /* 3087 * We have to be careful, if called from sys_setpriority(), 3088 * the task might be in the middle of scheduling on another CPU. 3089 */ 3090 rq = task_rq_lock(p, &flags); 3091 /* 3092 * The RT priorities are set via sched_setscheduler(), but we still 3093 * allow the 'normal' nice value to be set - but as expected 3094 * it wont have any effect on scheduling until the task is 3095 * SCHED_FIFO/SCHED_RR: 3096 */ 3097 if (task_has_rt_policy(p)) { 3098 p->static_prio = NICE_TO_PRIO(nice); 3099 goto out_unlock; 3100 } 3101 on_rq = p->on_rq; 3102 if (on_rq) 3103 dequeue_task(rq, p, 0); 3104 3105 p->static_prio = NICE_TO_PRIO(nice); 3106 set_load_weight(p); 3107 old_prio = p->prio; 3108 p->prio = effective_prio(p); 3109 delta = p->prio - old_prio; 3110 3111 if (on_rq) { 3112 enqueue_task(rq, p, 0); 3113 /* 3114 * If the task increased its priority or is running and 3115 * lowered its priority, then reschedule its CPU: 3116 */ 3117 if (delta < 0 || (delta > 0 && task_running(rq, p))) 3118 resched_task(rq->curr); 3119 } 3120out_unlock: 3121 task_rq_unlock(rq, p, &flags); 3122} 3123EXPORT_SYMBOL(set_user_nice); 3124 3125/* 3126 * can_nice - check if a task can reduce its nice value 3127 * @p: task 3128 * @nice: nice value 3129 */ 3130int can_nice(const struct task_struct *p, const int nice) 3131{ 3132 /* convert nice value [19,-20] to rlimit style value [1,40] */ 3133 int nice_rlim = 20 - nice; 3134 3135 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) || 3136 capable(CAP_SYS_NICE)); 3137} 3138 3139#ifdef __ARCH_WANT_SYS_NICE 3140 3141/* 3142 * sys_nice - change the priority of the current process. 3143 * @increment: priority increment 3144 * 3145 * sys_setpriority is a more generic, but much slower function that 3146 * does similar things. 3147 */ 3148SYSCALL_DEFINE1(nice, int, increment) 3149{ 3150 long nice, retval; 3151 3152 /* 3153 * Setpriority might change our priority at the same moment. 3154 * We don't have to worry. Conceptually one call occurs first 3155 * and we have a single winner. 3156 */ 3157 if (increment < -40) 3158 increment = -40; 3159 if (increment > 40) 3160 increment = 40; 3161 3162 nice = TASK_NICE(current) + increment; 3163 if (nice < -20) 3164 nice = -20; 3165 if (nice > 19) 3166 nice = 19; 3167 3168 if (increment < 0 && !can_nice(current, nice)) 3169 return -EPERM; 3170 3171 retval = security_task_setnice(current, nice); 3172 if (retval) 3173 return retval; 3174 3175 set_user_nice(current, nice); 3176 return 0; 3177} 3178 3179#endif 3180 3181/** 3182 * task_prio - return the priority value of a given task. 3183 * @p: the task in question. 3184 * 3185 * This is the priority value as seen by users in /proc. 3186 * RT tasks are offset by -200. Normal tasks are centered 3187 * around 0, value goes from -16 to +15. 3188 */ 3189int task_prio(const struct task_struct *p) 3190{ 3191 return p->prio - MAX_RT_PRIO; 3192} 3193 3194/** 3195 * task_nice - return the nice value of a given task. 3196 * @p: the task in question. 3197 */ 3198int task_nice(const struct task_struct *p) 3199{ 3200 return TASK_NICE(p); 3201} 3202EXPORT_SYMBOL(task_nice); 3203 3204/** 3205 * idle_cpu - is a given cpu idle currently? 3206 * @cpu: the processor in question. 3207 */ 3208int idle_cpu(int cpu) 3209{ 3210 struct rq *rq = cpu_rq(cpu); 3211 3212 if (rq->curr != rq->idle) 3213 return 0; 3214 3215 if (rq->nr_running) 3216 return 0; 3217 3218#ifdef CONFIG_SMP 3219 if (!llist_empty(&rq->wake_list)) 3220 return 0; 3221#endif 3222 3223 return 1; 3224} 3225 3226/** 3227 * idle_task - return the idle task for a given cpu. 3228 * @cpu: the processor in question. 3229 */ 3230struct task_struct *idle_task(int cpu) 3231{ 3232 return cpu_rq(cpu)->idle; 3233} 3234 3235/** 3236 * find_process_by_pid - find a process with a matching PID value. 3237 * @pid: the pid in question. 3238 */ 3239static struct task_struct *find_process_by_pid(pid_t pid) 3240{ 3241 return pid ? find_task_by_vpid(pid) : current; 3242} 3243 3244/* Actually do priority change: must hold rq lock. */ 3245static void 3246__setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio) 3247{ 3248 p->policy = policy; 3249 p->rt_priority = prio; 3250 p->normal_prio = normal_prio(p); 3251 /* we are holding p->pi_lock already */ 3252 p->prio = rt_mutex_getprio(p); 3253 if (rt_prio(p->prio)) 3254 p->sched_class = &rt_sched_class; 3255 else 3256 p->sched_class = &fair_sched_class; 3257 set_load_weight(p); 3258} 3259 3260/* 3261 * check the target process has a UID that matches the current process's 3262 */ 3263static bool check_same_owner(struct task_struct *p) 3264{ 3265 const struct cred *cred = current_cred(), *pcred; 3266 bool match; 3267 3268 rcu_read_lock(); 3269 pcred = __task_cred(p); 3270 match = (uid_eq(cred->euid, pcred->euid) || 3271 uid_eq(cred->euid, pcred->uid)); 3272 rcu_read_unlock(); 3273 return match; 3274} 3275 3276static int __sched_setscheduler(struct task_struct *p, int policy, 3277 const struct sched_param *param, bool user) 3278{ 3279 int retval, oldprio, oldpolicy = -1, on_rq, running; 3280 unsigned long flags; 3281 const struct sched_class *prev_class; 3282 struct rq *rq; 3283 int reset_on_fork; 3284 3285 /* may grab non-irq protected spin_locks */ 3286 BUG_ON(in_interrupt()); 3287recheck: 3288 /* double check policy once rq lock held */ 3289 if (policy < 0) { 3290 reset_on_fork = p->sched_reset_on_fork; 3291 policy = oldpolicy = p->policy; 3292 } else { 3293 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK); 3294 policy &= ~SCHED_RESET_ON_FORK; 3295 3296 if (policy != SCHED_FIFO && policy != SCHED_RR && 3297 policy != SCHED_NORMAL && policy != SCHED_BATCH && 3298 policy != SCHED_IDLE) 3299 return -EINVAL; 3300 } 3301 3302 /* 3303 * Valid priorities for SCHED_FIFO and SCHED_RR are 3304 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL, 3305 * SCHED_BATCH and SCHED_IDLE is 0. 3306 */ 3307 if (param->sched_priority < 0 || 3308 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) || 3309 (!p->mm && param->sched_priority > MAX_RT_PRIO-1)) 3310 return -EINVAL; 3311 if (rt_policy(policy) != (param->sched_priority != 0)) 3312 return -EINVAL; 3313 3314 /* 3315 * Allow unprivileged RT tasks to decrease priority: 3316 */ 3317 if (user && !capable(CAP_SYS_NICE)) { 3318 if (rt_policy(policy)) { 3319 unsigned long rlim_rtprio = 3320 task_rlimit(p, RLIMIT_RTPRIO); 3321 3322 /* can't set/change the rt policy */ 3323 if (policy != p->policy && !rlim_rtprio) 3324 return -EPERM; 3325 3326 /* can't increase priority */ 3327 if (param->sched_priority > p->rt_priority && 3328 param->sched_priority > rlim_rtprio) 3329 return -EPERM; 3330 } 3331 3332 /* 3333 * Treat SCHED_IDLE as nice 20. Only allow a switch to 3334 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it. 3335 */ 3336 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) { 3337 if (!can_nice(p, TASK_NICE(p))) 3338 return -EPERM; 3339 } 3340 3341 /* can't change other user's priorities */ 3342 if (!check_same_owner(p)) 3343 return -EPERM; 3344 3345 /* Normal users shall not reset the sched_reset_on_fork flag */ 3346 if (p->sched_reset_on_fork && !reset_on_fork) 3347 return -EPERM; 3348 } 3349 3350 if (user) { 3351 retval = security_task_setscheduler(p); 3352 if (retval) 3353 return retval; 3354 } 3355 3356 /* 3357 * make sure no PI-waiters arrive (or leave) while we are 3358 * changing the priority of the task: 3359 * 3360 * To be able to change p->policy safely, the appropriate 3361 * runqueue lock must be held. 3362 */ 3363 rq = task_rq_lock(p, &flags); 3364 3365 /* 3366 * Changing the policy of the stop threads its a very bad idea 3367 */ 3368 if (p == rq->stop) { 3369 task_rq_unlock(rq, p, &flags); 3370 return -EINVAL; 3371 } 3372 3373 /* 3374 * If not changing anything there's no need to proceed further: 3375 */ 3376 if (unlikely(policy == p->policy && (!rt_policy(policy) || 3377 param->sched_priority == p->rt_priority))) { 3378 task_rq_unlock(rq, p, &flags); 3379 return 0; 3380 } 3381 3382#ifdef CONFIG_RT_GROUP_SCHED 3383 if (user) { 3384 /* 3385 * Do not allow realtime tasks into groups that have no runtime 3386 * assigned. 3387 */ 3388 if (rt_bandwidth_enabled() && rt_policy(policy) && 3389 task_group(p)->rt_bandwidth.rt_runtime == 0 && 3390 !task_group_is_autogroup(task_group(p))) { 3391 task_rq_unlock(rq, p, &flags); 3392 return -EPERM; 3393 } 3394 } 3395#endif 3396 3397 /* recheck policy now with rq lock held */ 3398 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) { 3399 policy = oldpolicy = -1; 3400 task_rq_unlock(rq, p, &flags); 3401 goto recheck; 3402 } 3403 on_rq = p->on_rq; 3404 running = task_current(rq, p); 3405 if (on_rq) 3406 dequeue_task(rq, p, 0); 3407 if (running) 3408 p->sched_class->put_prev_task(rq, p); 3409 3410 p->sched_reset_on_fork = reset_on_fork; 3411 3412 oldprio = p->prio; 3413 prev_class = p->sched_class; 3414 __setscheduler(rq, p, policy, param->sched_priority); 3415 3416 if (running) 3417 p->sched_class->set_curr_task(rq); 3418 if (on_rq) 3419 enqueue_task(rq, p, 0); 3420 3421 check_class_changed(rq, p, prev_class, oldprio); 3422 task_rq_unlock(rq, p, &flags); 3423 3424 rt_mutex_adjust_pi(p); 3425 3426 return 0; 3427} 3428 3429/** 3430 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread. 3431 * @p: the task in question. 3432 * @policy: new policy. 3433 * @param: structure containing the new RT priority. 3434 * 3435 * NOTE that the task may be already dead. 3436 */ 3437int sched_setscheduler(struct task_struct *p, int policy, 3438 const struct sched_param *param) 3439{ 3440 return __sched_setscheduler(p, policy, param, true); 3441} 3442EXPORT_SYMBOL_GPL(sched_setscheduler); 3443 3444/** 3445 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace. 3446 * @p: the task in question. 3447 * @policy: new policy. 3448 * @param: structure containing the new RT priority. 3449 * 3450 * Just like sched_setscheduler, only don't bother checking if the 3451 * current context has permission. For example, this is needed in 3452 * stop_machine(): we create temporary high priority worker threads, 3453 * but our caller might not have that capability. 3454 */ 3455int sched_setscheduler_nocheck(struct task_struct *p, int policy, 3456 const struct sched_param *param) 3457{ 3458 return __sched_setscheduler(p, policy, param, false); 3459} 3460 3461static int 3462do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param) 3463{ 3464 struct sched_param lparam; 3465 struct task_struct *p; 3466 int retval; 3467 3468 if (!param || pid < 0) 3469 return -EINVAL; 3470 if (copy_from_user(&lparam, param, sizeof(struct sched_param))) 3471 return -EFAULT; 3472 3473 rcu_read_lock(); 3474 retval = -ESRCH; 3475 p = find_process_by_pid(pid); 3476 if (p != NULL) 3477 retval = sched_setscheduler(p, policy, &lparam); 3478 rcu_read_unlock(); 3479 3480 return retval; 3481} 3482 3483/** 3484 * sys_sched_setscheduler - set/change the scheduler policy and RT priority 3485 * @pid: the pid in question. 3486 * @policy: new policy. 3487 * @param: structure containing the new RT priority. 3488 */ 3489SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy, 3490 struct sched_param __user *, param) 3491{ 3492 /* negative values for policy are not valid */ 3493 if (policy < 0) 3494 return -EINVAL; 3495 3496 return do_sched_setscheduler(pid, policy, param); 3497} 3498 3499/** 3500 * sys_sched_setparam - set/change the RT priority of a thread 3501 * @pid: the pid in question. 3502 * @param: structure containing the new RT priority. 3503 */ 3504SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param) 3505{ 3506 return do_sched_setscheduler(pid, -1, param); 3507} 3508 3509/** 3510 * sys_sched_getscheduler - get the policy (scheduling class) of a thread 3511 * @pid: the pid in question. 3512 */ 3513SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid) 3514{ 3515 struct task_struct *p; 3516 int retval; 3517 3518 if (pid < 0) 3519 return -EINVAL; 3520 3521 retval = -ESRCH; 3522 rcu_read_lock(); 3523 p = find_process_by_pid(pid); 3524 if (p) { 3525 retval = security_task_getscheduler(p); 3526 if (!retval) 3527 retval = p->policy 3528 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0); 3529 } 3530 rcu_read_unlock(); 3531 return retval; 3532} 3533 3534/** 3535 * sys_sched_getparam - get the RT priority of a thread 3536 * @pid: the pid in question. 3537 * @param: structure containing the RT priority. 3538 */ 3539SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param) 3540{ 3541 struct sched_param lp; 3542 struct task_struct *p; 3543 int retval; 3544 3545 if (!param || pid < 0) 3546 return -EINVAL; 3547 3548 rcu_read_lock(); 3549 p = find_process_by_pid(pid); 3550 retval = -ESRCH; 3551 if (!p) 3552 goto out_unlock; 3553 3554 retval = security_task_getscheduler(p); 3555 if (retval) 3556 goto out_unlock; 3557 3558 lp.sched_priority = p->rt_priority; 3559 rcu_read_unlock(); 3560 3561 /* 3562 * This one might sleep, we cannot do it with a spinlock held ... 3563 */ 3564 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0; 3565 3566 return retval; 3567 3568out_unlock: 3569 rcu_read_unlock(); 3570 return retval; 3571} 3572 3573long sched_setaffinity(pid_t pid, const struct cpumask *in_mask) 3574{ 3575 cpumask_var_t cpus_allowed, new_mask; 3576 struct task_struct *p; 3577 int retval; 3578 3579 get_online_cpus(); 3580 rcu_read_lock(); 3581 3582 p = find_process_by_pid(pid); 3583 if (!p) { 3584 rcu_read_unlock(); 3585 put_online_cpus(); 3586 return -ESRCH; 3587 } 3588 3589 /* Prevent p going away */ 3590 get_task_struct(p); 3591 rcu_read_unlock(); 3592 3593 if (p->flags & PF_NO_SETAFFINITY) { 3594 retval = -EINVAL; 3595 goto out_put_task; 3596 } 3597 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) { 3598 retval = -ENOMEM; 3599 goto out_put_task; 3600 } 3601 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) { 3602 retval = -ENOMEM; 3603 goto out_free_cpus_allowed; 3604 } 3605 retval = -EPERM; 3606 if (!check_same_owner(p)) { 3607 rcu_read_lock(); 3608 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) { 3609 rcu_read_unlock(); 3610 goto out_unlock; 3611 } 3612 rcu_read_unlock(); 3613 } 3614 3615 retval = security_task_setscheduler(p); 3616 if (retval) 3617 goto out_unlock; 3618 3619 cpuset_cpus_allowed(p, cpus_allowed); 3620 cpumask_and(new_mask, in_mask, cpus_allowed); 3621again: 3622 retval = set_cpus_allowed_ptr(p, new_mask); 3623 3624 if (!retval) { 3625 cpuset_cpus_allowed(p, cpus_allowed); 3626 if (!cpumask_subset(new_mask, cpus_allowed)) { 3627 /* 3628 * We must have raced with a concurrent cpuset 3629 * update. Just reset the cpus_allowed to the 3630 * cpuset's cpus_allowed 3631 */ 3632 cpumask_copy(new_mask, cpus_allowed); 3633 goto again; 3634 } 3635 } 3636out_unlock: 3637 free_cpumask_var(new_mask); 3638out_free_cpus_allowed: 3639 free_cpumask_var(cpus_allowed); 3640out_put_task: 3641 put_task_struct(p); 3642 put_online_cpus(); 3643 return retval; 3644} 3645 3646static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len, 3647 struct cpumask *new_mask) 3648{ 3649 if (len < cpumask_size()) 3650 cpumask_clear(new_mask); 3651 else if (len > cpumask_size()) 3652 len = cpumask_size(); 3653 3654 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0; 3655} 3656 3657/** 3658 * sys_sched_setaffinity - set the cpu affinity of a process 3659 * @pid: pid of the process 3660 * @len: length in bytes of the bitmask pointed to by user_mask_ptr 3661 * @user_mask_ptr: user-space pointer to the new cpu mask 3662 */ 3663SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len, 3664 unsigned long __user *, user_mask_ptr) 3665{ 3666 cpumask_var_t new_mask; 3667 int retval; 3668 3669 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) 3670 return -ENOMEM; 3671 3672 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask); 3673 if (retval == 0) 3674 retval = sched_setaffinity(pid, new_mask); 3675 free_cpumask_var(new_mask); 3676 return retval; 3677} 3678 3679long sched_getaffinity(pid_t pid, struct cpumask *mask) 3680{ 3681 struct task_struct *p; 3682 unsigned long flags; 3683 int retval; 3684 3685 get_online_cpus(); 3686 rcu_read_lock(); 3687 3688 retval = -ESRCH; 3689 p = find_process_by_pid(pid); 3690 if (!p) 3691 goto out_unlock; 3692 3693 retval = security_task_getscheduler(p); 3694 if (retval) 3695 goto out_unlock; 3696 3697 raw_spin_lock_irqsave(&p->pi_lock, flags); 3698 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask); 3699 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 3700 3701out_unlock: 3702 rcu_read_unlock(); 3703 put_online_cpus(); 3704 3705 return retval; 3706} 3707 3708/** 3709 * sys_sched_getaffinity - get the cpu affinity of a process 3710 * @pid: pid of the process 3711 * @len: length in bytes of the bitmask pointed to by user_mask_ptr 3712 * @user_mask_ptr: user-space pointer to hold the current cpu mask 3713 */ 3714SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len, 3715 unsigned long __user *, user_mask_ptr) 3716{ 3717 int ret; 3718 cpumask_var_t mask; 3719 3720 if ((len * BITS_PER_BYTE) < nr_cpu_ids) 3721 return -EINVAL; 3722 if (len & (sizeof(unsigned long)-1)) 3723 return -EINVAL; 3724 3725 if (!alloc_cpumask_var(&mask, GFP_KERNEL)) 3726 return -ENOMEM; 3727 3728 ret = sched_getaffinity(pid, mask); 3729 if (ret == 0) { 3730 size_t retlen = min_t(size_t, len, cpumask_size()); 3731 3732 if (copy_to_user(user_mask_ptr, mask, retlen)) 3733 ret = -EFAULT; 3734 else 3735 ret = retlen; 3736 } 3737 free_cpumask_var(mask); 3738 3739 return ret; 3740} 3741 3742/** 3743 * sys_sched_yield - yield the current processor to other threads. 3744 * 3745 * This function yields the current CPU to other tasks. If there are no 3746 * other threads running on this CPU then this function will return. 3747 */ 3748SYSCALL_DEFINE0(sched_yield) 3749{ 3750 struct rq *rq = this_rq_lock(); 3751 3752 schedstat_inc(rq, yld_count); 3753 current->sched_class->yield_task(rq); 3754 3755 /* 3756 * Since we are going to call schedule() anyway, there's 3757 * no need to preempt or enable interrupts: 3758 */ 3759 __release(rq->lock); 3760 spin_release(&rq->lock.dep_map, 1, _THIS_IP_); 3761 do_raw_spin_unlock(&rq->lock); 3762 sched_preempt_enable_no_resched(); 3763 3764 schedule(); 3765 3766 return 0; 3767} 3768 3769static inline int should_resched(void) 3770{ 3771 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE); 3772} 3773 3774static void __cond_resched(void) 3775{ 3776 add_preempt_count(PREEMPT_ACTIVE); 3777 __schedule(); 3778 sub_preempt_count(PREEMPT_ACTIVE); 3779} 3780 3781int __sched _cond_resched(void) 3782{ 3783 if (should_resched()) { 3784 __cond_resched(); 3785 return 1; 3786 } 3787 return 0; 3788} 3789EXPORT_SYMBOL(_cond_resched); 3790 3791/* 3792 * __cond_resched_lock() - if a reschedule is pending, drop the given lock, 3793 * call schedule, and on return reacquire the lock. 3794 * 3795 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level 3796 * operations here to prevent schedule() from being called twice (once via 3797 * spin_unlock(), once by hand). 3798 */ 3799int __cond_resched_lock(spinlock_t *lock) 3800{ 3801 int resched = should_resched(); 3802 int ret = 0; 3803 3804 lockdep_assert_held(lock); 3805 3806 if (spin_needbreak(lock) || resched) { 3807 spin_unlock(lock); 3808 if (resched) 3809 __cond_resched(); 3810 else 3811 cpu_relax(); 3812 ret = 1; 3813 spin_lock(lock); 3814 } 3815 return ret; 3816} 3817EXPORT_SYMBOL(__cond_resched_lock); 3818 3819int __sched __cond_resched_softirq(void) 3820{ 3821 BUG_ON(!in_softirq()); 3822 3823 if (should_resched()) { 3824 local_bh_enable(); 3825 __cond_resched(); 3826 local_bh_disable(); 3827 return 1; 3828 } 3829 return 0; 3830} 3831EXPORT_SYMBOL(__cond_resched_softirq); 3832 3833/** 3834 * yield - yield the current processor to other threads. 3835 * 3836 * Do not ever use this function, there's a 99% chance you're doing it wrong. 3837 * 3838 * The scheduler is at all times free to pick the calling task as the most 3839 * eligible task to run, if removing the yield() call from your code breaks 3840 * it, its already broken. 3841 * 3842 * Typical broken usage is: 3843 * 3844 * while (!event) 3845 * yield(); 3846 * 3847 * where one assumes that yield() will let 'the other' process run that will 3848 * make event true. If the current task is a SCHED_FIFO task that will never 3849 * happen. Never use yield() as a progress guarantee!! 3850 * 3851 * If you want to use yield() to wait for something, use wait_event(). 3852 * If you want to use yield() to be 'nice' for others, use cond_resched(). 3853 * If you still want to use yield(), do not! 3854 */ 3855void __sched yield(void) 3856{ 3857 set_current_state(TASK_RUNNING); 3858 sys_sched_yield(); 3859} 3860EXPORT_SYMBOL(yield); 3861 3862/** 3863 * yield_to - yield the current processor to another thread in 3864 * your thread group, or accelerate that thread toward the 3865 * processor it's on. 3866 * @p: target task 3867 * @preempt: whether task preemption is allowed or not 3868 * 3869 * It's the caller's job to ensure that the target task struct 3870 * can't go away on us before we can do any checks. 3871 * 3872 * Returns: 3873 * true (>0) if we indeed boosted the target task. 3874 * false (0) if we failed to boost the target. 3875 * -ESRCH if there's no task to yield to. 3876 */ 3877bool __sched yield_to(struct task_struct *p, bool preempt) 3878{ 3879 struct task_struct *curr = current; 3880 struct rq *rq, *p_rq; 3881 unsigned long flags; 3882 int yielded = 0; 3883 3884 local_irq_save(flags); 3885 rq = this_rq(); 3886 3887again: 3888 p_rq = task_rq(p); 3889 /* 3890 * If we're the only runnable task on the rq and target rq also 3891 * has only one task, there's absolutely no point in yielding. 3892 */ 3893 if (rq->nr_running == 1 && p_rq->nr_running == 1) { 3894 yielded = -ESRCH; 3895 goto out_irq; 3896 } 3897 3898 double_rq_lock(rq, p_rq); 3899 while (task_rq(p) != p_rq) { 3900 double_rq_unlock(rq, p_rq); 3901 goto again; 3902 } 3903 3904 if (!curr->sched_class->yield_to_task) 3905 goto out_unlock; 3906 3907 if (curr->sched_class != p->sched_class) 3908 goto out_unlock; 3909 3910 if (task_running(p_rq, p) || p->state) 3911 goto out_unlock; 3912 3913 yielded = curr->sched_class->yield_to_task(rq, p, preempt); 3914 if (yielded) { 3915 schedstat_inc(rq, yld_count); 3916 /* 3917 * Make p's CPU reschedule; pick_next_entity takes care of 3918 * fairness. 3919 */ 3920 if (preempt && rq != p_rq) 3921 resched_task(p_rq->curr); 3922 } 3923 3924out_unlock: 3925 double_rq_unlock(rq, p_rq); 3926out_irq: 3927 local_irq_restore(flags); 3928 3929 if (yielded > 0) 3930 schedule(); 3931 3932 return yielded; 3933} 3934EXPORT_SYMBOL_GPL(yield_to); 3935 3936/* 3937 * This task is about to go to sleep on IO. Increment rq->nr_iowait so 3938 * that process accounting knows that this is a task in IO wait state. 3939 */ 3940void __sched io_schedule(void) 3941{ 3942 struct rq *rq = raw_rq(); 3943 3944 delayacct_blkio_start(); 3945 atomic_inc(&rq->nr_iowait); 3946 blk_flush_plug(current); 3947 current->in_iowait = 1; 3948 schedule(); 3949 current->in_iowait = 0; 3950 atomic_dec(&rq->nr_iowait); 3951 delayacct_blkio_end(); 3952} 3953EXPORT_SYMBOL(io_schedule); 3954 3955long __sched io_schedule_timeout(long timeout) 3956{ 3957 struct rq *rq = raw_rq(); 3958 long ret; 3959 3960 delayacct_blkio_start(); 3961 atomic_inc(&rq->nr_iowait); 3962 blk_flush_plug(current); 3963 current->in_iowait = 1; 3964 ret = schedule_timeout(timeout); 3965 current->in_iowait = 0; 3966 atomic_dec(&rq->nr_iowait); 3967 delayacct_blkio_end(); 3968 return ret; 3969} 3970 3971/** 3972 * sys_sched_get_priority_max - return maximum RT priority. 3973 * @policy: scheduling class. 3974 * 3975 * this syscall returns the maximum rt_priority that can be used 3976 * by a given scheduling class. 3977 */ 3978SYSCALL_DEFINE1(sched_get_priority_max, int, policy) 3979{ 3980 int ret = -EINVAL; 3981 3982 switch (policy) { 3983 case SCHED_FIFO: 3984 case SCHED_RR: 3985 ret = MAX_USER_RT_PRIO-1; 3986 break; 3987 case SCHED_NORMAL: 3988 case SCHED_BATCH: 3989 case SCHED_IDLE: 3990 ret = 0; 3991 break; 3992 } 3993 return ret; 3994} 3995 3996/** 3997 * sys_sched_get_priority_min - return minimum RT priority. 3998 * @policy: scheduling class. 3999 * 4000 * this syscall returns the minimum rt_priority that can be used 4001 * by a given scheduling class. 4002 */ 4003SYSCALL_DEFINE1(sched_get_priority_min, int, policy) 4004{ 4005 int ret = -EINVAL; 4006 4007 switch (policy) { 4008 case SCHED_FIFO: 4009 case SCHED_RR: 4010 ret = 1; 4011 break; 4012 case SCHED_NORMAL: 4013 case SCHED_BATCH: 4014 case SCHED_IDLE: 4015 ret = 0; 4016 } 4017 return ret; 4018} 4019 4020/** 4021 * sys_sched_rr_get_interval - return the default timeslice of a process. 4022 * @pid: pid of the process. 4023 * @interval: userspace pointer to the timeslice value. 4024 * 4025 * this syscall writes the default timeslice value of a given process 4026 * into the user-space timespec buffer. A value of '0' means infinity. 4027 */ 4028SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid, 4029 struct timespec __user *, interval) 4030{ 4031 struct task_struct *p; 4032 unsigned int time_slice; 4033 unsigned long flags; 4034 struct rq *rq; 4035 int retval; 4036 struct timespec t; 4037 4038 if (pid < 0) 4039 return -EINVAL; 4040 4041 retval = -ESRCH; 4042 rcu_read_lock(); 4043 p = find_process_by_pid(pid); 4044 if (!p) 4045 goto out_unlock; 4046 4047 retval = security_task_getscheduler(p); 4048 if (retval) 4049 goto out_unlock; 4050 4051 rq = task_rq_lock(p, &flags); 4052 time_slice = p->sched_class->get_rr_interval(rq, p); 4053 task_rq_unlock(rq, p, &flags); 4054 4055 rcu_read_unlock(); 4056 jiffies_to_timespec(time_slice, &t); 4057 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0; 4058 return retval; 4059 4060out_unlock: 4061 rcu_read_unlock(); 4062 return retval; 4063} 4064 4065static const char stat_nam[] = TASK_STATE_TO_CHAR_STR; 4066 4067void sched_show_task(struct task_struct *p) 4068{ 4069 unsigned long free = 0; 4070 int ppid; 4071 unsigned state; 4072 4073 state = p->state ? __ffs(p->state) + 1 : 0; 4074 printk(KERN_INFO "%-15.15s %c", p->comm, 4075 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?'); 4076#if BITS_PER_LONG == 32 4077 if (state == TASK_RUNNING) 4078 printk(KERN_CONT " running "); 4079 else 4080 printk(KERN_CONT " %08lx ", thread_saved_pc(p)); 4081#else 4082 if (state == TASK_RUNNING) 4083 printk(KERN_CONT " running task "); 4084 else 4085 printk(KERN_CONT " %016lx ", thread_saved_pc(p)); 4086#endif 4087#ifdef CONFIG_DEBUG_STACK_USAGE 4088 free = stack_not_used(p); 4089#endif 4090 rcu_read_lock(); 4091 ppid = task_pid_nr(rcu_dereference(p->real_parent)); 4092 rcu_read_unlock(); 4093 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free, 4094 task_pid_nr(p), ppid, 4095 (unsigned long)task_thread_info(p)->flags); 4096 4097 print_worker_info(KERN_INFO, p); 4098 show_stack(p, NULL); 4099} 4100 4101void show_state_filter(unsigned long state_filter) 4102{ 4103 struct task_struct *g, *p; 4104 4105#if BITS_PER_LONG == 32 4106 printk(KERN_INFO 4107 " task PC stack pid father\n"); 4108#else 4109 printk(KERN_INFO 4110 " task PC stack pid father\n"); 4111#endif 4112 rcu_read_lock(); 4113 do_each_thread(g, p) { 4114 /* 4115 * reset the NMI-timeout, listing all files on a slow 4116 * console might take a lot of time: 4117 */ 4118 touch_nmi_watchdog(); 4119 if (!state_filter || (p->state & state_filter)) 4120 sched_show_task(p); 4121 } while_each_thread(g, p); 4122 4123 touch_all_softlockup_watchdogs(); 4124 4125#ifdef CONFIG_SCHED_DEBUG 4126 sysrq_sched_debug_show(); 4127#endif 4128 rcu_read_unlock(); 4129 /* 4130 * Only show locks if all tasks are dumped: 4131 */ 4132 if (!state_filter) 4133 debug_show_all_locks(); 4134} 4135 4136void init_idle_bootup_task(struct task_struct *idle) 4137{ 4138 idle->sched_class = &idle_sched_class; 4139} 4140 4141/** 4142 * init_idle - set up an idle thread for a given CPU 4143 * @idle: task in question 4144 * @cpu: cpu the idle task belongs to 4145 * 4146 * NOTE: this function does not set the idle thread's NEED_RESCHED 4147 * flag, to make booting more robust. 4148 */ 4149void init_idle(struct task_struct *idle, int cpu) 4150{ 4151 struct rq *rq = cpu_rq(cpu); 4152 unsigned long flags; 4153 4154 raw_spin_lock_irqsave(&rq->lock, flags); 4155 4156 __sched_fork(idle); 4157 idle->state = TASK_RUNNING; 4158 idle->se.exec_start = sched_clock(); 4159 4160 do_set_cpus_allowed(idle, cpumask_of(cpu)); 4161 /* 4162 * We're having a chicken and egg problem, even though we are 4163 * holding rq->lock, the cpu isn't yet set to this cpu so the 4164 * lockdep check in task_group() will fail. 4165 * 4166 * Similar case to sched_fork(). / Alternatively we could 4167 * use task_rq_lock() here and obtain the other rq->lock. 4168 * 4169 * Silence PROVE_RCU 4170 */ 4171 rcu_read_lock(); 4172 __set_task_cpu(idle, cpu); 4173 rcu_read_unlock(); 4174 4175 rq->curr = rq->idle = idle; 4176#if defined(CONFIG_SMP) 4177 idle->on_cpu = 1; 4178#endif 4179 raw_spin_unlock_irqrestore(&rq->lock, flags); 4180 4181 /* Set the preempt count _outside_ the spinlocks! */ 4182 task_thread_info(idle)->preempt_count = 0; 4183 4184 /* 4185 * The idle tasks have their own, simple scheduling class: 4186 */ 4187 idle->sched_class = &idle_sched_class; 4188 ftrace_graph_init_idle_task(idle, cpu); 4189 vtime_init_idle(idle, cpu); 4190#if defined(CONFIG_SMP) 4191 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu); 4192#endif 4193} 4194 4195#ifdef CONFIG_SMP 4196void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask) 4197{ 4198 if (p->sched_class && p->sched_class->set_cpus_allowed) 4199 p->sched_class->set_cpus_allowed(p, new_mask); 4200 4201 cpumask_copy(&p->cpus_allowed, new_mask); 4202 p->nr_cpus_allowed = cpumask_weight(new_mask); 4203} 4204 4205/* 4206 * This is how migration works: 4207 * 4208 * 1) we invoke migration_cpu_stop() on the target CPU using 4209 * stop_one_cpu(). 4210 * 2) stopper starts to run (implicitly forcing the migrated thread 4211 * off the CPU) 4212 * 3) it checks whether the migrated task is still in the wrong runqueue. 4213 * 4) if it's in the wrong runqueue then the migration thread removes 4214 * it and puts it into the right queue. 4215 * 5) stopper completes and stop_one_cpu() returns and the migration 4216 * is done. 4217 */ 4218 4219/* 4220 * Change a given task's CPU affinity. Migrate the thread to a 4221 * proper CPU and schedule it away if the CPU it's executing on 4222 * is removed from the allowed bitmask. 4223 * 4224 * NOTE: the caller must have a valid reference to the task, the 4225 * task must not exit() & deallocate itself prematurely. The 4226 * call is not atomic; no spinlocks may be held. 4227 */ 4228int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask) 4229{ 4230 unsigned long flags; 4231 struct rq *rq; 4232 unsigned int dest_cpu; 4233 int ret = 0; 4234 4235 rq = task_rq_lock(p, &flags); 4236 4237 if (cpumask_equal(&p->cpus_allowed, new_mask)) 4238 goto out; 4239 4240 if (!cpumask_intersects(new_mask, cpu_active_mask)) { 4241 ret = -EINVAL; 4242 goto out; 4243 } 4244 4245 do_set_cpus_allowed(p, new_mask); 4246 4247 /* Can the task run on the task's current CPU? If so, we're done */ 4248 if (cpumask_test_cpu(task_cpu(p), new_mask)) 4249 goto out; 4250 4251 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask); 4252 if (p->on_rq) { 4253 struct migration_arg arg = { p, dest_cpu }; 4254 /* Need help from migration thread: drop lock and wait. */ 4255 task_rq_unlock(rq, p, &flags); 4256 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg); 4257 tlb_migrate_finish(p->mm); 4258 return 0; 4259 } 4260out: 4261 task_rq_unlock(rq, p, &flags); 4262 4263 return ret; 4264} 4265EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr); 4266 4267/* 4268 * Move (not current) task off this cpu, onto dest cpu. We're doing 4269 * this because either it can't run here any more (set_cpus_allowed() 4270 * away from this CPU, or CPU going down), or because we're 4271 * attempting to rebalance this task on exec (sched_exec). 4272 * 4273 * So we race with normal scheduler movements, but that's OK, as long 4274 * as the task is no longer on this CPU. 4275 * 4276 * Returns non-zero if task was successfully migrated. 4277 */ 4278static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu) 4279{ 4280 struct rq *rq_dest, *rq_src; 4281 int ret = 0; 4282 4283 if (unlikely(!cpu_active(dest_cpu))) 4284 return ret; 4285 4286 rq_src = cpu_rq(src_cpu); 4287 rq_dest = cpu_rq(dest_cpu); 4288 4289 raw_spin_lock(&p->pi_lock); 4290 double_rq_lock(rq_src, rq_dest); 4291 /* Already moved. */ 4292 if (task_cpu(p) != src_cpu) 4293 goto done; 4294 /* Affinity changed (again). */ 4295 if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p))) 4296 goto fail; 4297 4298 /* 4299 * If we're not on a rq, the next wake-up will ensure we're 4300 * placed properly. 4301 */ 4302 if (p->on_rq) { 4303 dequeue_task(rq_src, p, 0); 4304 set_task_cpu(p, dest_cpu); 4305 enqueue_task(rq_dest, p, 0); 4306 check_preempt_curr(rq_dest, p, 0); 4307 } 4308done: 4309 ret = 1; 4310fail: 4311 double_rq_unlock(rq_src, rq_dest); 4312 raw_spin_unlock(&p->pi_lock); 4313 return ret; 4314} 4315 4316/* 4317 * migration_cpu_stop - this will be executed by a highprio stopper thread 4318 * and performs thread migration by bumping thread off CPU then 4319 * 'pushing' onto another runqueue. 4320 */ 4321static int migration_cpu_stop(void *data) 4322{ 4323 struct migration_arg *arg = data; 4324 4325 /* 4326 * The original target cpu might have gone down and we might 4327 * be on another cpu but it doesn't matter. 4328 */ 4329 local_irq_disable(); 4330 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu); 4331 local_irq_enable(); 4332 return 0; 4333} 4334 4335#ifdef CONFIG_HOTPLUG_CPU 4336 4337/* 4338 * Ensures that the idle task is using init_mm right before its cpu goes 4339 * offline. 4340 */ 4341void idle_task_exit(void) 4342{ 4343 struct mm_struct *mm = current->active_mm; 4344 4345 BUG_ON(cpu_online(smp_processor_id())); 4346 4347 if (mm != &init_mm) 4348 switch_mm(mm, &init_mm, current); 4349 mmdrop(mm); 4350} 4351 4352/* 4353 * Since this CPU is going 'away' for a while, fold any nr_active delta 4354 * we might have. Assumes we're called after migrate_tasks() so that the 4355 * nr_active count is stable. 4356 * 4357 * Also see the comment "Global load-average calculations". 4358 */ 4359static void calc_load_migrate(struct rq *rq) 4360{ 4361 long delta = calc_load_fold_active(rq); 4362 if (delta) 4363 atomic_long_add(delta, &calc_load_tasks); 4364} 4365 4366/* 4367 * Migrate all tasks from the rq, sleeping tasks will be migrated by 4368 * try_to_wake_up()->select_task_rq(). 4369 * 4370 * Called with rq->lock held even though we'er in stop_machine() and 4371 * there's no concurrency possible, we hold the required locks anyway 4372 * because of lock validation efforts. 4373 */ 4374static void migrate_tasks(unsigned int dead_cpu) 4375{ 4376 struct rq *rq = cpu_rq(dead_cpu); 4377 struct task_struct *next, *stop = rq->stop; 4378 int dest_cpu; 4379 4380 /* 4381 * Fudge the rq selection such that the below task selection loop 4382 * doesn't get stuck on the currently eligible stop task. 4383 * 4384 * We're currently inside stop_machine() and the rq is either stuck 4385 * in the stop_machine_cpu_stop() loop, or we're executing this code, 4386 * either way we should never end up calling schedule() until we're 4387 * done here. 4388 */ 4389 rq->stop = NULL; 4390 4391 /* 4392 * put_prev_task() and pick_next_task() sched 4393 * class method both need to have an up-to-date 4394 * value of rq->clock[_task] 4395 */ 4396 update_rq_clock(rq); 4397 4398 for ( ; ; ) { 4399 /* 4400 * There's this thread running, bail when that's the only 4401 * remaining thread. 4402 */ 4403 if (rq->nr_running == 1) 4404 break; 4405 4406 next = pick_next_task(rq); 4407 BUG_ON(!next); 4408 next->sched_class->put_prev_task(rq, next); 4409 4410 /* Find suitable destination for @next, with force if needed. */ 4411 dest_cpu = select_fallback_rq(dead_cpu, next); 4412 raw_spin_unlock(&rq->lock); 4413 4414 __migrate_task(next, dead_cpu, dest_cpu); 4415 4416 raw_spin_lock(&rq->lock); 4417 } 4418 4419 rq->stop = stop; 4420} 4421 4422#endif /* CONFIG_HOTPLUG_CPU */ 4423 4424#if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL) 4425 4426static struct ctl_table sd_ctl_dir[] = { 4427 { 4428 .procname = "sched_domain", 4429 .mode = 0555, 4430 }, 4431 {} 4432}; 4433 4434static struct ctl_table sd_ctl_root[] = { 4435 { 4436 .procname = "kernel", 4437 .mode = 0555, 4438 .child = sd_ctl_dir, 4439 }, 4440 {} 4441}; 4442 4443static struct ctl_table *sd_alloc_ctl_entry(int n) 4444{ 4445 struct ctl_table *entry = 4446 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL); 4447 4448 return entry; 4449} 4450 4451static void sd_free_ctl_entry(struct ctl_table **tablep) 4452{ 4453 struct ctl_table *entry; 4454 4455 /* 4456 * In the intermediate directories, both the child directory and 4457 * procname are dynamically allocated and could fail but the mode 4458 * will always be set. In the lowest directory the names are 4459 * static strings and all have proc handlers. 4460 */ 4461 for (entry = *tablep; entry->mode; entry++) { 4462 if (entry->child) 4463 sd_free_ctl_entry(&entry->child); 4464 if (entry->proc_handler == NULL) 4465 kfree(entry->procname); 4466 } 4467 4468 kfree(*tablep); 4469 *tablep = NULL; 4470} 4471 4472static int min_load_idx = 0; 4473static int max_load_idx = CPU_LOAD_IDX_MAX-1; 4474 4475static void 4476set_table_entry(struct ctl_table *entry, 4477 const char *procname, void *data, int maxlen, 4478 umode_t mode, proc_handler *proc_handler, 4479 bool load_idx) 4480{ 4481 entry->procname = procname; 4482 entry->data = data; 4483 entry->maxlen = maxlen; 4484 entry->mode = mode; 4485 entry->proc_handler = proc_handler; 4486 4487 if (load_idx) { 4488 entry->extra1 = &min_load_idx; 4489 entry->extra2 = &max_load_idx; 4490 } 4491} 4492 4493static struct ctl_table * 4494sd_alloc_ctl_domain_table(struct sched_domain *sd) 4495{ 4496 struct ctl_table *table = sd_alloc_ctl_entry(13); 4497 4498 if (table == NULL) 4499 return NULL; 4500 4501 set_table_entry(&table[0], "min_interval", &sd->min_interval, 4502 sizeof(long), 0644, proc_doulongvec_minmax, false); 4503 set_table_entry(&table[1], "max_interval", &sd->max_interval, 4504 sizeof(long), 0644, proc_doulongvec_minmax, false); 4505 set_table_entry(&table[2], "busy_idx", &sd->busy_idx, 4506 sizeof(int), 0644, proc_dointvec_minmax, true); 4507 set_table_entry(&table[3], "idle_idx", &sd->idle_idx, 4508 sizeof(int), 0644, proc_dointvec_minmax, true); 4509 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx, 4510 sizeof(int), 0644, proc_dointvec_minmax, true); 4511 set_table_entry(&table[5], "wake_idx", &sd->wake_idx, 4512 sizeof(int), 0644, proc_dointvec_minmax, true); 4513 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx, 4514 sizeof(int), 0644, proc_dointvec_minmax, true); 4515 set_table_entry(&table[7], "busy_factor", &sd->busy_factor, 4516 sizeof(int), 0644, proc_dointvec_minmax, false); 4517 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct, 4518 sizeof(int), 0644, proc_dointvec_minmax, false); 4519 set_table_entry(&table[9], "cache_nice_tries", 4520 &sd->cache_nice_tries, 4521 sizeof(int), 0644, proc_dointvec_minmax, false); 4522 set_table_entry(&table[10], "flags", &sd->flags, 4523 sizeof(int), 0644, proc_dointvec_minmax, false); 4524 set_table_entry(&table[11], "name", sd->name, 4525 CORENAME_MAX_SIZE, 0444, proc_dostring, false); 4526 /* &table[12] is terminator */ 4527 4528 return table; 4529} 4530 4531static struct ctl_table *sd_alloc_ctl_cpu_table(int cpu) 4532{ 4533 struct ctl_table *entry, *table; 4534 struct sched_domain *sd; 4535 int domain_num = 0, i; 4536 char buf[32]; 4537 4538 for_each_domain(cpu, sd) 4539 domain_num++; 4540 entry = table = sd_alloc_ctl_entry(domain_num + 1); 4541 if (table == NULL) 4542 return NULL; 4543 4544 i = 0; 4545 for_each_domain(cpu, sd) { 4546 snprintf(buf, 32, "domain%d", i); 4547 entry->procname = kstrdup(buf, GFP_KERNEL); 4548 entry->mode = 0555; 4549 entry->child = sd_alloc_ctl_domain_table(sd); 4550 entry++; 4551 i++; 4552 } 4553 return table; 4554} 4555 4556static struct ctl_table_header *sd_sysctl_header; 4557static void register_sched_domain_sysctl(void) 4558{ 4559 int i, cpu_num = num_possible_cpus(); 4560 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1); 4561 char buf[32]; 4562 4563 WARN_ON(sd_ctl_dir[0].child); 4564 sd_ctl_dir[0].child = entry; 4565 4566 if (entry == NULL) 4567 return; 4568 4569 for_each_possible_cpu(i) { 4570 snprintf(buf, 32, "cpu%d", i); 4571 entry->procname = kstrdup(buf, GFP_KERNEL); 4572 entry->mode = 0555; 4573 entry->child = sd_alloc_ctl_cpu_table(i); 4574 entry++; 4575 } 4576 4577 WARN_ON(sd_sysctl_header); 4578 sd_sysctl_header = register_sysctl_table(sd_ctl_root); 4579} 4580 4581/* may be called multiple times per register */ 4582static void unregister_sched_domain_sysctl(void) 4583{ 4584 if (sd_sysctl_header) 4585 unregister_sysctl_table(sd_sysctl_header); 4586 sd_sysctl_header = NULL; 4587 if (sd_ctl_dir[0].child) 4588 sd_free_ctl_entry(&sd_ctl_dir[0].child); 4589} 4590#else 4591static void register_sched_domain_sysctl(void) 4592{ 4593} 4594static void unregister_sched_domain_sysctl(void) 4595{ 4596} 4597#endif 4598 4599static void set_rq_online(struct rq *rq) 4600{ 4601 if (!rq->online) { 4602 const struct sched_class *class; 4603 4604 cpumask_set_cpu(rq->cpu, rq->rd->online); 4605 rq->online = 1; 4606 4607 for_each_class(class) { 4608 if (class->rq_online) 4609 class->rq_online(rq); 4610 } 4611 } 4612} 4613 4614static void set_rq_offline(struct rq *rq) 4615{ 4616 if (rq->online) { 4617 const struct sched_class *class; 4618 4619 for_each_class(class) { 4620 if (class->rq_offline) 4621 class->rq_offline(rq); 4622 } 4623 4624 cpumask_clear_cpu(rq->cpu, rq->rd->online); 4625 rq->online = 0; 4626 } 4627} 4628 4629/* 4630 * migration_call - callback that gets triggered when a CPU is added. 4631 * Here we can start up the necessary migration thread for the new CPU. 4632 */ 4633static int 4634migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu) 4635{ 4636 int cpu = (long)hcpu; 4637 unsigned long flags; 4638 struct rq *rq = cpu_rq(cpu); 4639 4640 switch (action & ~CPU_TASKS_FROZEN) { 4641 4642 case CPU_UP_PREPARE: 4643 rq->calc_load_update = calc_load_update; 4644 break; 4645 4646 case CPU_ONLINE: 4647 /* Update our root-domain */ 4648 raw_spin_lock_irqsave(&rq->lock, flags); 4649 if (rq->rd) { 4650 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span)); 4651 4652 set_rq_online(rq); 4653 } 4654 raw_spin_unlock_irqrestore(&rq->lock, flags); 4655 break; 4656 4657#ifdef CONFIG_HOTPLUG_CPU 4658 case CPU_DYING: 4659 sched_ttwu_pending(); 4660 /* Update our root-domain */ 4661 raw_spin_lock_irqsave(&rq->lock, flags); 4662 if (rq->rd) { 4663 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span)); 4664 set_rq_offline(rq); 4665 } 4666 migrate_tasks(cpu); 4667 BUG_ON(rq->nr_running != 1); /* the migration thread */ 4668 raw_spin_unlock_irqrestore(&rq->lock, flags); 4669 break; 4670 4671 case CPU_DEAD: 4672 calc_load_migrate(rq); 4673 break; 4674#endif 4675 } 4676 4677 update_max_interval(); 4678 4679 return NOTIFY_OK; 4680} 4681 4682/* 4683 * Register at high priority so that task migration (migrate_all_tasks) 4684 * happens before everything else. This has to be lower priority than 4685 * the notifier in the perf_event subsystem, though. 4686 */ 4687static struct notifier_block migration_notifier = { 4688 .notifier_call = migration_call, 4689 .priority = CPU_PRI_MIGRATION, 4690}; 4691 4692static int sched_cpu_active(struct notifier_block *nfb, 4693 unsigned long action, void *hcpu) 4694{ 4695 switch (action & ~CPU_TASKS_FROZEN) { 4696 case CPU_STARTING: 4697 case CPU_DOWN_FAILED: 4698 set_cpu_active((long)hcpu, true); 4699 return NOTIFY_OK; 4700 default: 4701 return NOTIFY_DONE; 4702 } 4703} 4704 4705static int sched_cpu_inactive(struct notifier_block *nfb, 4706 unsigned long action, void *hcpu) 4707{ 4708 switch (action & ~CPU_TASKS_FROZEN) { 4709 case CPU_DOWN_PREPARE: 4710 set_cpu_active((long)hcpu, false); 4711 return NOTIFY_OK; 4712 default: 4713 return NOTIFY_DONE; 4714 } 4715} 4716 4717static int __init migration_init(void) 4718{ 4719 void *cpu = (void *)(long)smp_processor_id(); 4720 int err; 4721 4722 /* Initialize migration for the boot CPU */ 4723 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu); 4724 BUG_ON(err == NOTIFY_BAD); 4725 migration_call(&migration_notifier, CPU_ONLINE, cpu); 4726 register_cpu_notifier(&migration_notifier); 4727 4728 /* Register cpu active notifiers */ 4729 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE); 4730 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE); 4731 4732 return 0; 4733} 4734early_initcall(migration_init); 4735#endif 4736 4737#ifdef CONFIG_SMP 4738 4739static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */ 4740 4741#ifdef CONFIG_SCHED_DEBUG 4742 4743static __read_mostly int sched_debug_enabled; 4744 4745static int __init sched_debug_setup(char *str) 4746{ 4747 sched_debug_enabled = 1; 4748 4749 return 0; 4750} 4751early_param("sched_debug", sched_debug_setup); 4752 4753static inline bool sched_debug(void) 4754{ 4755 return sched_debug_enabled; 4756} 4757 4758static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level, 4759 struct cpumask *groupmask) 4760{ 4761 struct sched_group *group = sd->groups; 4762 char str[256]; 4763 4764 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd)); 4765 cpumask_clear(groupmask); 4766 4767 printk(KERN_DEBUG "%*s domain %d: ", level, "", level); 4768 4769 if (!(sd->flags & SD_LOAD_BALANCE)) { 4770 printk("does not load-balance\n"); 4771 if (sd->parent) 4772 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain" 4773 " has parent"); 4774 return -1; 4775 } 4776 4777 printk(KERN_CONT "span %s level %s\n", str, sd->name); 4778 4779 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) { 4780 printk(KERN_ERR "ERROR: domain->span does not contain " 4781 "CPU%d\n", cpu); 4782 } 4783 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) { 4784 printk(KERN_ERR "ERROR: domain->groups does not contain" 4785 " CPU%d\n", cpu); 4786 } 4787 4788 printk(KERN_DEBUG "%*s groups:", level + 1, ""); 4789 do { 4790 if (!group) { 4791 printk("\n"); 4792 printk(KERN_ERR "ERROR: group is NULL\n"); 4793 break; 4794 } 4795 4796 /* 4797 * Even though we initialize ->power to something semi-sane, 4798 * we leave power_orig unset. This allows us to detect if 4799 * domain iteration is still funny without causing /0 traps. 4800 */ 4801 if (!group->sgp->power_orig) { 4802 printk(KERN_CONT "\n"); 4803 printk(KERN_ERR "ERROR: domain->cpu_power not " 4804 "set\n"); 4805 break; 4806 } 4807 4808 if (!cpumask_weight(sched_group_cpus(group))) { 4809 printk(KERN_CONT "\n"); 4810 printk(KERN_ERR "ERROR: empty group\n"); 4811 break; 4812 } 4813 4814 if (!(sd->flags & SD_OVERLAP) && 4815 cpumask_intersects(groupmask, sched_group_cpus(group))) { 4816 printk(KERN_CONT "\n"); 4817 printk(KERN_ERR "ERROR: repeated CPUs\n"); 4818 break; 4819 } 4820 4821 cpumask_or(groupmask, groupmask, sched_group_cpus(group)); 4822 4823 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group)); 4824 4825 printk(KERN_CONT " %s", str); 4826 if (group->sgp->power != SCHED_POWER_SCALE) { 4827 printk(KERN_CONT " (cpu_power = %d)", 4828 group->sgp->power); 4829 } 4830 4831 group = group->next; 4832 } while (group != sd->groups); 4833 printk(KERN_CONT "\n"); 4834 4835 if (!cpumask_equal(sched_domain_span(sd), groupmask)) 4836 printk(KERN_ERR "ERROR: groups don't span domain->span\n"); 4837 4838 if (sd->parent && 4839 !cpumask_subset(groupmask, sched_domain_span(sd->parent))) 4840 printk(KERN_ERR "ERROR: parent span is not a superset " 4841 "of domain->span\n"); 4842 return 0; 4843} 4844 4845static void sched_domain_debug(struct sched_domain *sd, int cpu) 4846{ 4847 int level = 0; 4848 4849 if (!sched_debug_enabled) 4850 return; 4851 4852 if (!sd) { 4853 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu); 4854 return; 4855 } 4856 4857 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu); 4858 4859 for (;;) { 4860 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask)) 4861 break; 4862 level++; 4863 sd = sd->parent; 4864 if (!sd) 4865 break; 4866 } 4867} 4868#else /* !CONFIG_SCHED_DEBUG */ 4869# define sched_domain_debug(sd, cpu) do { } while (0) 4870static inline bool sched_debug(void) 4871{ 4872 return false; 4873} 4874#endif /* CONFIG_SCHED_DEBUG */ 4875 4876static int sd_degenerate(struct sched_domain *sd) 4877{ 4878 if (cpumask_weight(sched_domain_span(sd)) == 1) 4879 return 1; 4880 4881 /* Following flags need at least 2 groups */ 4882 if (sd->flags & (SD_LOAD_BALANCE | 4883 SD_BALANCE_NEWIDLE | 4884 SD_BALANCE_FORK | 4885 SD_BALANCE_EXEC | 4886 SD_SHARE_CPUPOWER | 4887 SD_SHARE_PKG_RESOURCES)) { 4888 if (sd->groups != sd->groups->next) 4889 return 0; 4890 } 4891 4892 /* Following flags don't use groups */ 4893 if (sd->flags & (SD_WAKE_AFFINE)) 4894 return 0; 4895 4896 return 1; 4897} 4898 4899static int 4900sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent) 4901{ 4902 unsigned long cflags = sd->flags, pflags = parent->flags; 4903 4904 if (sd_degenerate(parent)) 4905 return 1; 4906 4907 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent))) 4908 return 0; 4909 4910 /* Flags needing groups don't count if only 1 group in parent */ 4911 if (parent->groups == parent->groups->next) { 4912 pflags &= ~(SD_LOAD_BALANCE | 4913 SD_BALANCE_NEWIDLE | 4914 SD_BALANCE_FORK | 4915 SD_BALANCE_EXEC | 4916 SD_SHARE_CPUPOWER | 4917 SD_SHARE_PKG_RESOURCES); 4918 if (nr_node_ids == 1) 4919 pflags &= ~SD_SERIALIZE; 4920 } 4921 if (~cflags & pflags) 4922 return 0; 4923 4924 return 1; 4925} 4926 4927static void free_rootdomain(struct rcu_head *rcu) 4928{ 4929 struct root_domain *rd = container_of(rcu, struct root_domain, rcu); 4930 4931 cpupri_cleanup(&rd->cpupri); 4932 free_cpumask_var(rd->rto_mask); 4933 free_cpumask_var(rd->online); 4934 free_cpumask_var(rd->span); 4935 kfree(rd); 4936} 4937 4938static void rq_attach_root(struct rq *rq, struct root_domain *rd) 4939{ 4940 struct root_domain *old_rd = NULL; 4941 unsigned long flags; 4942 4943 raw_spin_lock_irqsave(&rq->lock, flags); 4944 4945 if (rq->rd) { 4946 old_rd = rq->rd; 4947 4948 if (cpumask_test_cpu(rq->cpu, old_rd->online)) 4949 set_rq_offline(rq); 4950 4951 cpumask_clear_cpu(rq->cpu, old_rd->span); 4952 4953 /* 4954 * If we dont want to free the old_rt yet then 4955 * set old_rd to NULL to skip the freeing later 4956 * in this function: 4957 */ 4958 if (!atomic_dec_and_test(&old_rd->refcount)) 4959 old_rd = NULL; 4960 } 4961 4962 atomic_inc(&rd->refcount); 4963 rq->rd = rd; 4964 4965 cpumask_set_cpu(rq->cpu, rd->span); 4966 if (cpumask_test_cpu(rq->cpu, cpu_active_mask)) 4967 set_rq_online(rq); 4968 4969 raw_spin_unlock_irqrestore(&rq->lock, flags); 4970 4971 if (old_rd) 4972 call_rcu_sched(&old_rd->rcu, free_rootdomain); 4973} 4974 4975static int init_rootdomain(struct root_domain *rd) 4976{ 4977 memset(rd, 0, sizeof(*rd)); 4978 4979 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL)) 4980 goto out; 4981 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL)) 4982 goto free_span; 4983 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL)) 4984 goto free_online; 4985 4986 if (cpupri_init(&rd->cpupri) != 0) 4987 goto free_rto_mask; 4988 return 0; 4989 4990free_rto_mask: 4991 free_cpumask_var(rd->rto_mask); 4992free_online: 4993 free_cpumask_var(rd->online); 4994free_span: 4995 free_cpumask_var(rd->span); 4996out: 4997 return -ENOMEM; 4998} 4999 5000/* 5001 * By default the system creates a single root-domain with all cpus as 5002 * members (mimicking the global state we have today). 5003 */ 5004struct root_domain def_root_domain; 5005 5006static void init_defrootdomain(void) 5007{ 5008 init_rootdomain(&def_root_domain); 5009 5010 atomic_set(&def_root_domain.refcount, 1); 5011} 5012 5013static struct root_domain *alloc_rootdomain(void) 5014{ 5015 struct root_domain *rd; 5016 5017 rd = kmalloc(sizeof(*rd), GFP_KERNEL); 5018 if (!rd) 5019 return NULL; 5020 5021 if (init_rootdomain(rd) != 0) { 5022 kfree(rd); 5023 return NULL; 5024 } 5025 5026 return rd; 5027} 5028 5029static void free_sched_groups(struct sched_group *sg, int free_sgp) 5030{ 5031 struct sched_group *tmp, *first; 5032 5033 if (!sg) 5034 return; 5035 5036 first = sg; 5037 do { 5038 tmp = sg->next; 5039 5040 if (free_sgp && atomic_dec_and_test(&sg->sgp->ref)) 5041 kfree(sg->sgp); 5042 5043 kfree(sg); 5044 sg = tmp; 5045 } while (sg != first); 5046} 5047 5048static void free_sched_domain(struct rcu_head *rcu) 5049{ 5050 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu); 5051 5052 /* 5053 * If its an overlapping domain it has private groups, iterate and 5054 * nuke them all. 5055 */ 5056 if (sd->flags & SD_OVERLAP) { 5057 free_sched_groups(sd->groups, 1); 5058 } else if (atomic_dec_and_test(&sd->groups->ref)) { 5059 kfree(sd->groups->sgp); 5060 kfree(sd->groups); 5061 } 5062 kfree(sd); 5063} 5064 5065static void destroy_sched_domain(struct sched_domain *sd, int cpu) 5066{ 5067 call_rcu(&sd->rcu, free_sched_domain); 5068} 5069 5070static void destroy_sched_domains(struct sched_domain *sd, int cpu) 5071{ 5072 for (; sd; sd = sd->parent) 5073 destroy_sched_domain(sd, cpu); 5074} 5075 5076/* 5077 * Keep a special pointer to the highest sched_domain that has 5078 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this 5079 * allows us to avoid some pointer chasing select_idle_sibling(). 5080 * 5081 * Also keep a unique ID per domain (we use the first cpu number in 5082 * the cpumask of the domain), this allows us to quickly tell if 5083 * two cpus are in the same cache domain, see cpus_share_cache(). 5084 */ 5085DEFINE_PER_CPU(struct sched_domain *, sd_llc); 5086DEFINE_PER_CPU(int, sd_llc_size); 5087DEFINE_PER_CPU(int, sd_llc_id); 5088 5089static void update_top_cache_domain(int cpu) 5090{ 5091 struct sched_domain *sd; 5092 int id = cpu; 5093 int size = 1; 5094 5095 sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES); 5096 if (sd) { 5097 id = cpumask_first(sched_domain_span(sd)); 5098 size = cpumask_weight(sched_domain_span(sd)); 5099 } 5100 5101 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd); 5102 per_cpu(sd_llc_size, cpu) = size; 5103 per_cpu(sd_llc_id, cpu) = id; 5104} 5105 5106/* 5107 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must 5108 * hold the hotplug lock. 5109 */ 5110static void 5111cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu) 5112{ 5113 struct rq *rq = cpu_rq(cpu); 5114 struct sched_domain *tmp; 5115 5116 /* Remove the sched domains which do not contribute to scheduling. */ 5117 for (tmp = sd; tmp; ) { 5118 struct sched_domain *parent = tmp->parent; 5119 if (!parent) 5120 break; 5121 5122 if (sd_parent_degenerate(tmp, parent)) { 5123 tmp->parent = parent->parent; 5124 if (parent->parent) 5125 parent->parent->child = tmp; 5126 destroy_sched_domain(parent, cpu); 5127 } else 5128 tmp = tmp->parent; 5129 } 5130 5131 if (sd && sd_degenerate(sd)) { 5132 tmp = sd; 5133 sd = sd->parent; 5134 destroy_sched_domain(tmp, cpu); 5135 if (sd) 5136 sd->child = NULL; 5137 } 5138 5139 sched_domain_debug(sd, cpu); 5140 5141 rq_attach_root(rq, rd); 5142 tmp = rq->sd; 5143 rcu_assign_pointer(rq->sd, sd); 5144 destroy_sched_domains(tmp, cpu); 5145 5146 update_top_cache_domain(cpu); 5147} 5148 5149/* cpus with isolated domains */ 5150static cpumask_var_t cpu_isolated_map; 5151 5152/* Setup the mask of cpus configured for isolated domains */ 5153static int __init isolated_cpu_setup(char *str) 5154{ 5155 alloc_bootmem_cpumask_var(&cpu_isolated_map); 5156 cpulist_parse(str, cpu_isolated_map); 5157 return 1; 5158} 5159 5160__setup("isolcpus=", isolated_cpu_setup); 5161 5162static const struct cpumask *cpu_cpu_mask(int cpu) 5163{ 5164 return cpumask_of_node(cpu_to_node(cpu)); 5165} 5166 5167struct sd_data { 5168 struct sched_domain **__percpu sd; 5169 struct sched_group **__percpu sg; 5170 struct sched_group_power **__percpu sgp; 5171}; 5172 5173struct s_data { 5174 struct sched_domain ** __percpu sd; 5175 struct root_domain *rd; 5176}; 5177 5178enum s_alloc { 5179 sa_rootdomain, 5180 sa_sd, 5181 sa_sd_storage, 5182 sa_none, 5183}; 5184 5185struct sched_domain_topology_level; 5186 5187typedef struct sched_domain *(*sched_domain_init_f)(struct sched_domain_topology_level *tl, int cpu); 5188typedef const struct cpumask *(*sched_domain_mask_f)(int cpu); 5189 5190#define SDTL_OVERLAP 0x01 5191 5192struct sched_domain_topology_level { 5193 sched_domain_init_f init; 5194 sched_domain_mask_f mask; 5195 int flags; 5196 int numa_level; 5197 struct sd_data data; 5198}; 5199 5200/* 5201 * Build an iteration mask that can exclude certain CPUs from the upwards 5202 * domain traversal. 5203 * 5204 * Asymmetric node setups can result in situations where the domain tree is of 5205 * unequal depth, make sure to skip domains that already cover the entire 5206 * range. 5207 * 5208 * In that case build_sched_domains() will have terminated the iteration early 5209 * and our sibling sd spans will be empty. Domains should always include the 5210 * cpu they're built on, so check that. 5211 * 5212 */ 5213static void build_group_mask(struct sched_domain *sd, struct sched_group *sg) 5214{ 5215 const struct cpumask *span = sched_domain_span(sd); 5216 struct sd_data *sdd = sd->private; 5217 struct sched_domain *sibling; 5218 int i; 5219 5220 for_each_cpu(i, span) { 5221 sibling = *per_cpu_ptr(sdd->sd, i); 5222 if (!cpumask_test_cpu(i, sched_domain_span(sibling))) 5223 continue; 5224 5225 cpumask_set_cpu(i, sched_group_mask(sg)); 5226 } 5227} 5228 5229/* 5230 * Return the canonical balance cpu for this group, this is the first cpu 5231 * of this group that's also in the iteration mask. 5232 */ 5233int group_balance_cpu(struct sched_group *sg) 5234{ 5235 return cpumask_first_and(sched_group_cpus(sg), sched_group_mask(sg)); 5236} 5237 5238static int 5239build_overlap_sched_groups(struct sched_domain *sd, int cpu) 5240{ 5241 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg; 5242 const struct cpumask *span = sched_domain_span(sd); 5243 struct cpumask *covered = sched_domains_tmpmask; 5244 struct sd_data *sdd = sd->private; 5245 struct sched_domain *child; 5246 int i; 5247 5248 cpumask_clear(covered); 5249 5250 for_each_cpu(i, span) { 5251 struct cpumask *sg_span; 5252 5253 if (cpumask_test_cpu(i, covered)) 5254 continue; 5255 5256 child = *per_cpu_ptr(sdd->sd, i); 5257 5258 /* See the comment near build_group_mask(). */ 5259 if (!cpumask_test_cpu(i, sched_domain_span(child))) 5260 continue; 5261 5262 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(), 5263 GFP_KERNEL, cpu_to_node(cpu)); 5264 5265 if (!sg) 5266 goto fail; 5267 5268 sg_span = sched_group_cpus(sg); 5269 if (child->child) { 5270 child = child->child; 5271 cpumask_copy(sg_span, sched_domain_span(child)); 5272 } else 5273 cpumask_set_cpu(i, sg_span); 5274 5275 cpumask_or(covered, covered, sg_span); 5276 5277 sg->sgp = *per_cpu_ptr(sdd->sgp, i); 5278 if (atomic_inc_return(&sg->sgp->ref) == 1) 5279 build_group_mask(sd, sg); 5280 5281 /* 5282 * Initialize sgp->power such that even if we mess up the 5283 * domains and no possible iteration will get us here, we won't 5284 * die on a /0 trap. 5285 */ 5286 sg->sgp->power = SCHED_POWER_SCALE * cpumask_weight(sg_span); 5287 5288 /* 5289 * Make sure the first group of this domain contains the 5290 * canonical balance cpu. Otherwise the sched_domain iteration 5291 * breaks. See update_sg_lb_stats(). 5292 */ 5293 if ((!groups && cpumask_test_cpu(cpu, sg_span)) || 5294 group_balance_cpu(sg) == cpu) 5295 groups = sg; 5296 5297 if (!first) 5298 first = sg; 5299 if (last) 5300 last->next = sg; 5301 last = sg; 5302 last->next = first; 5303 } 5304 sd->groups = groups; 5305 5306 return 0; 5307 5308fail: 5309 free_sched_groups(first, 0); 5310 5311 return -ENOMEM; 5312} 5313 5314static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg) 5315{ 5316 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu); 5317 struct sched_domain *child = sd->child; 5318 5319 if (child) 5320 cpu = cpumask_first(sched_domain_span(child)); 5321 5322 if (sg) { 5323 *sg = *per_cpu_ptr(sdd->sg, cpu); 5324 (*sg)->sgp = *per_cpu_ptr(sdd->sgp, cpu); 5325 atomic_set(&(*sg)->sgp->ref, 1); /* for claim_allocations */ 5326 } 5327 5328 return cpu; 5329} 5330 5331/* 5332 * build_sched_groups will build a circular linked list of the groups 5333 * covered by the given span, and will set each group's ->cpumask correctly, 5334 * and ->cpu_power to 0. 5335 * 5336 * Assumes the sched_domain tree is fully constructed 5337 */ 5338static int 5339build_sched_groups(struct sched_domain *sd, int cpu) 5340{ 5341 struct sched_group *first = NULL, *last = NULL; 5342 struct sd_data *sdd = sd->private; 5343 const struct cpumask *span = sched_domain_span(sd); 5344 struct cpumask *covered; 5345 int i; 5346 5347 get_group(cpu, sdd, &sd->groups); 5348 atomic_inc(&sd->groups->ref); 5349 5350 if (cpu != cpumask_first(span)) 5351 return 0; 5352 5353 lockdep_assert_held(&sched_domains_mutex); 5354 covered = sched_domains_tmpmask; 5355 5356 cpumask_clear(covered); 5357 5358 for_each_cpu(i, span) { 5359 struct sched_group *sg; 5360 int group, j; 5361 5362 if (cpumask_test_cpu(i, covered)) 5363 continue; 5364 5365 group = get_group(i, sdd, &sg); 5366 cpumask_clear(sched_group_cpus(sg)); 5367 sg->sgp->power = 0; 5368 cpumask_setall(sched_group_mask(sg)); 5369 5370 for_each_cpu(j, span) { 5371 if (get_group(j, sdd, NULL) != group) 5372 continue; 5373 5374 cpumask_set_cpu(j, covered); 5375 cpumask_set_cpu(j, sched_group_cpus(sg)); 5376 } 5377 5378 if (!first) 5379 first = sg; 5380 if (last) 5381 last->next = sg; 5382 last = sg; 5383 } 5384 last->next = first; 5385 5386 return 0; 5387} 5388 5389/* 5390 * Initialize sched groups cpu_power. 5391 * 5392 * cpu_power indicates the capacity of sched group, which is used while 5393 * distributing the load between different sched groups in a sched domain. 5394 * Typically cpu_power for all the groups in a sched domain will be same unless 5395 * there are asymmetries in the topology. If there are asymmetries, group 5396 * having more cpu_power will pickup more load compared to the group having 5397 * less cpu_power. 5398 */ 5399static void init_sched_groups_power(int cpu, struct sched_domain *sd) 5400{ 5401 struct sched_group *sg = sd->groups; 5402 5403 WARN_ON(!sg); 5404 5405 do { 5406 sg->group_weight = cpumask_weight(sched_group_cpus(sg)); 5407 sg = sg->next; 5408 } while (sg != sd->groups); 5409 5410 if (cpu != group_balance_cpu(sg)) 5411 return; 5412 5413 update_group_power(sd, cpu); 5414 atomic_set(&sg->sgp->nr_busy_cpus, sg->group_weight); 5415} 5416 5417int __weak arch_sd_sibling_asym_packing(void) 5418{ 5419 return 0*SD_ASYM_PACKING; 5420} 5421 5422/* 5423 * Initializers for schedule domains 5424 * Non-inlined to reduce accumulated stack pressure in build_sched_domains() 5425 */ 5426 5427#ifdef CONFIG_SCHED_DEBUG 5428# define SD_INIT_NAME(sd, type) sd->name = #type 5429#else 5430# define SD_INIT_NAME(sd, type) do { } while (0) 5431#endif 5432 5433#define SD_INIT_FUNC(type) \ 5434static noinline struct sched_domain * \ 5435sd_init_##type(struct sched_domain_topology_level *tl, int cpu) \ 5436{ \ 5437 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); \ 5438 *sd = SD_##type##_INIT; \ 5439 SD_INIT_NAME(sd, type); \ 5440 sd->private = &tl->data; \ 5441 return sd; \ 5442} 5443 5444SD_INIT_FUNC(CPU) 5445#ifdef CONFIG_SCHED_SMT 5446 SD_INIT_FUNC(SIBLING) 5447#endif 5448#ifdef CONFIG_SCHED_MC 5449 SD_INIT_FUNC(MC) 5450#endif 5451#ifdef CONFIG_SCHED_BOOK 5452 SD_INIT_FUNC(BOOK) 5453#endif 5454 5455static int default_relax_domain_level = -1; 5456int sched_domain_level_max; 5457 5458static int __init setup_relax_domain_level(char *str) 5459{ 5460 if (kstrtoint(str, 0, &default_relax_domain_level)) 5461 pr_warn("Unable to set relax_domain_level\n"); 5462 5463 return 1; 5464} 5465__setup("relax_domain_level=", setup_relax_domain_level); 5466 5467static void set_domain_attribute(struct sched_domain *sd, 5468 struct sched_domain_attr *attr) 5469{ 5470 int request; 5471 5472 if (!attr || attr->relax_domain_level < 0) { 5473 if (default_relax_domain_level < 0) 5474 return; 5475 else 5476 request = default_relax_domain_level; 5477 } else 5478 request = attr->relax_domain_level; 5479 if (request < sd->level) { 5480 /* turn off idle balance on this domain */ 5481 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE); 5482 } else { 5483 /* turn on idle balance on this domain */ 5484 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE); 5485 } 5486} 5487 5488static void __sdt_free(const struct cpumask *cpu_map); 5489static int __sdt_alloc(const struct cpumask *cpu_map); 5490 5491static void __free_domain_allocs(struct s_data *d, enum s_alloc what, 5492 const struct cpumask *cpu_map) 5493{ 5494 switch (what) { 5495 case sa_rootdomain: 5496 if (!atomic_read(&d->rd->refcount)) 5497 free_rootdomain(&d->rd->rcu); /* fall through */ 5498 case sa_sd: 5499 free_percpu(d->sd); /* fall through */ 5500 case sa_sd_storage: 5501 __sdt_free(cpu_map); /* fall through */ 5502 case sa_none: 5503 break; 5504 } 5505} 5506 5507static enum s_alloc __visit_domain_allocation_hell(struct s_data *d, 5508 const struct cpumask *cpu_map) 5509{ 5510 memset(d, 0, sizeof(*d)); 5511 5512 if (__sdt_alloc(cpu_map)) 5513 return sa_sd_storage; 5514 d->sd = alloc_percpu(struct sched_domain *); 5515 if (!d->sd) 5516 return sa_sd_storage; 5517 d->rd = alloc_rootdomain(); 5518 if (!d->rd) 5519 return sa_sd; 5520 return sa_rootdomain; 5521} 5522 5523/* 5524 * NULL the sd_data elements we've used to build the sched_domain and 5525 * sched_group structure so that the subsequent __free_domain_allocs() 5526 * will not free the data we're using. 5527 */ 5528static void claim_allocations(int cpu, struct sched_domain *sd) 5529{ 5530 struct sd_data *sdd = sd->private; 5531 5532 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd); 5533 *per_cpu_ptr(sdd->sd, cpu) = NULL; 5534 5535 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref)) 5536 *per_cpu_ptr(sdd->sg, cpu) = NULL; 5537 5538 if (atomic_read(&(*per_cpu_ptr(sdd->sgp, cpu))->ref)) 5539 *per_cpu_ptr(sdd->sgp, cpu) = NULL; 5540} 5541 5542#ifdef CONFIG_SCHED_SMT 5543static const struct cpumask *cpu_smt_mask(int cpu) 5544{ 5545 return topology_thread_cpumask(cpu); 5546} 5547#endif 5548 5549/* 5550 * Topology list, bottom-up. 5551 */ 5552static struct sched_domain_topology_level default_topology[] = { 5553#ifdef CONFIG_SCHED_SMT 5554 { sd_init_SIBLING, cpu_smt_mask, }, 5555#endif 5556#ifdef CONFIG_SCHED_MC 5557 { sd_init_MC, cpu_coregroup_mask, }, 5558#endif 5559#ifdef CONFIG_SCHED_BOOK 5560 { sd_init_BOOK, cpu_book_mask, }, 5561#endif 5562 { sd_init_CPU, cpu_cpu_mask, }, 5563 { NULL, }, 5564}; 5565 5566static struct sched_domain_topology_level *sched_domain_topology = default_topology; 5567 5568#define for_each_sd_topology(tl) \ 5569 for (tl = sched_domain_topology; tl->init; tl++) 5570 5571#ifdef CONFIG_NUMA 5572 5573static int sched_domains_numa_levels; 5574static int *sched_domains_numa_distance; 5575static struct cpumask ***sched_domains_numa_masks; 5576static int sched_domains_curr_level; 5577 5578static inline int sd_local_flags(int level) 5579{ 5580 if (sched_domains_numa_distance[level] > RECLAIM_DISTANCE) 5581 return 0; 5582 5583 return SD_BALANCE_EXEC | SD_BALANCE_FORK | SD_WAKE_AFFINE; 5584} 5585 5586static struct sched_domain * 5587sd_numa_init(struct sched_domain_topology_level *tl, int cpu) 5588{ 5589 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); 5590 int level = tl->numa_level; 5591 int sd_weight = cpumask_weight( 5592 sched_domains_numa_masks[level][cpu_to_node(cpu)]); 5593 5594 *sd = (struct sched_domain){ 5595 .min_interval = sd_weight, 5596 .max_interval = 2*sd_weight, 5597 .busy_factor = 32, 5598 .imbalance_pct = 125, 5599 .cache_nice_tries = 2, 5600 .busy_idx = 3, 5601 .idle_idx = 2, 5602 .newidle_idx = 0, 5603 .wake_idx = 0, 5604 .forkexec_idx = 0, 5605 5606 .flags = 1*SD_LOAD_BALANCE 5607 | 1*SD_BALANCE_NEWIDLE 5608 | 0*SD_BALANCE_EXEC 5609 | 0*SD_BALANCE_FORK 5610 | 0*SD_BALANCE_WAKE 5611 | 0*SD_WAKE_AFFINE 5612 | 0*SD_SHARE_CPUPOWER 5613 | 0*SD_SHARE_PKG_RESOURCES 5614 | 1*SD_SERIALIZE 5615 | 0*SD_PREFER_SIBLING 5616 | sd_local_flags(level) 5617 , 5618 .last_balance = jiffies, 5619 .balance_interval = sd_weight, 5620 }; 5621 SD_INIT_NAME(sd, NUMA); 5622 sd->private = &tl->data; 5623 5624 /* 5625 * Ugly hack to pass state to sd_numa_mask()... 5626 */ 5627 sched_domains_curr_level = tl->numa_level; 5628 5629 return sd; 5630} 5631 5632static const struct cpumask *sd_numa_mask(int cpu) 5633{ 5634 return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)]; 5635} 5636 5637static void sched_numa_warn(const char *str) 5638{ 5639 static int done = false; 5640 int i,j; 5641 5642 if (done) 5643 return; 5644 5645 done = true; 5646 5647 printk(KERN_WARNING "ERROR: %s\n\n", str); 5648 5649 for (i = 0; i < nr_node_ids; i++) { 5650 printk(KERN_WARNING " "); 5651 for (j = 0; j < nr_node_ids; j++) 5652 printk(KERN_CONT "%02d ", node_distance(i,j)); 5653 printk(KERN_CONT "\n"); 5654 } 5655 printk(KERN_WARNING "\n"); 5656} 5657 5658static bool find_numa_distance(int distance) 5659{ 5660 int i; 5661 5662 if (distance == node_distance(0, 0)) 5663 return true; 5664 5665 for (i = 0; i < sched_domains_numa_levels; i++) { 5666 if (sched_domains_numa_distance[i] == distance) 5667 return true; 5668 } 5669 5670 return false; 5671} 5672 5673static void sched_init_numa(void) 5674{ 5675 int next_distance, curr_distance = node_distance(0, 0); 5676 struct sched_domain_topology_level *tl; 5677 int level = 0; 5678 int i, j, k; 5679 5680 sched_domains_numa_distance = kzalloc(sizeof(int) * nr_node_ids, GFP_KERNEL); 5681 if (!sched_domains_numa_distance) 5682 return; 5683 5684 /* 5685 * O(nr_nodes^2) deduplicating selection sort -- in order to find the 5686 * unique distances in the node_distance() table. 5687 * 5688 * Assumes node_distance(0,j) includes all distances in 5689 * node_distance(i,j) in order to avoid cubic time. 5690 */ 5691 next_distance = curr_distance; 5692 for (i = 0; i < nr_node_ids; i++) { 5693 for (j = 0; j < nr_node_ids; j++) { 5694 for (k = 0; k < nr_node_ids; k++) { 5695 int distance = node_distance(i, k); 5696 5697 if (distance > curr_distance && 5698 (distance < next_distance || 5699 next_distance == curr_distance)) 5700 next_distance = distance; 5701 5702 /* 5703 * While not a strong assumption it would be nice to know 5704 * about cases where if node A is connected to B, B is not 5705 * equally connected to A. 5706 */ 5707 if (sched_debug() && node_distance(k, i) != distance) 5708 sched_numa_warn("Node-distance not symmetric"); 5709 5710 if (sched_debug() && i && !find_numa_distance(distance)) 5711 sched_numa_warn("Node-0 not representative"); 5712 } 5713 if (next_distance != curr_distance) { 5714 sched_domains_numa_distance[level++] = next_distance; 5715 sched_domains_numa_levels = level; 5716 curr_distance = next_distance; 5717 } else break; 5718 } 5719 5720 /* 5721 * In case of sched_debug() we verify the above assumption. 5722 */ 5723 if (!sched_debug()) 5724 break; 5725 } 5726 /* 5727 * 'level' contains the number of unique distances, excluding the 5728 * identity distance node_distance(i,i). 5729 * 5730 * The sched_domains_numa_distance[] array includes the actual distance 5731 * numbers. 5732 */ 5733 5734 /* 5735 * Here, we should temporarily reset sched_domains_numa_levels to 0. 5736 * If it fails to allocate memory for array sched_domains_numa_masks[][], 5737 * the array will contain less then 'level' members. This could be 5738 * dangerous when we use it to iterate array sched_domains_numa_masks[][] 5739 * in other functions. 5740 * 5741 * We reset it to 'level' at the end of this function. 5742 */ 5743 sched_domains_numa_levels = 0; 5744 5745 sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL); 5746 if (!sched_domains_numa_masks) 5747 return; 5748 5749 /* 5750 * Now for each level, construct a mask per node which contains all 5751 * cpus of nodes that are that many hops away from us. 5752 */ 5753 for (i = 0; i < level; i++) { 5754 sched_domains_numa_masks[i] = 5755 kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL); 5756 if (!sched_domains_numa_masks[i]) 5757 return; 5758 5759 for (j = 0; j < nr_node_ids; j++) { 5760 struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL); 5761 if (!mask) 5762 return; 5763 5764 sched_domains_numa_masks[i][j] = mask; 5765 5766 for (k = 0; k < nr_node_ids; k++) { 5767 if (node_distance(j, k) > sched_domains_numa_distance[i]) 5768 continue; 5769 5770 cpumask_or(mask, mask, cpumask_of_node(k)); 5771 } 5772 } 5773 } 5774 5775 tl = kzalloc((ARRAY_SIZE(default_topology) + level) * 5776 sizeof(struct sched_domain_topology_level), GFP_KERNEL); 5777 if (!tl) 5778 return; 5779 5780 /* 5781 * Copy the default topology bits.. 5782 */ 5783 for (i = 0; default_topology[i].init; i++) 5784 tl[i] = default_topology[i]; 5785 5786 /* 5787 * .. and append 'j' levels of NUMA goodness. 5788 */ 5789 for (j = 0; j < level; i++, j++) { 5790 tl[i] = (struct sched_domain_topology_level){ 5791 .init = sd_numa_init, 5792 .mask = sd_numa_mask, 5793 .flags = SDTL_OVERLAP, 5794 .numa_level = j, 5795 }; 5796 } 5797 5798 sched_domain_topology = tl; 5799 5800 sched_domains_numa_levels = level; 5801} 5802 5803static void sched_domains_numa_masks_set(int cpu) 5804{ 5805 int i, j; 5806 int node = cpu_to_node(cpu); 5807 5808 for (i = 0; i < sched_domains_numa_levels; i++) { 5809 for (j = 0; j < nr_node_ids; j++) { 5810 if (node_distance(j, node) <= sched_domains_numa_distance[i]) 5811 cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]); 5812 } 5813 } 5814} 5815 5816static void sched_domains_numa_masks_clear(int cpu) 5817{ 5818 int i, j; 5819 for (i = 0; i < sched_domains_numa_levels; i++) { 5820 for (j = 0; j < nr_node_ids; j++) 5821 cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]); 5822 } 5823} 5824 5825/* 5826 * Update sched_domains_numa_masks[level][node] array when new cpus 5827 * are onlined. 5828 */ 5829static int sched_domains_numa_masks_update(struct notifier_block *nfb, 5830 unsigned long action, 5831 void *hcpu) 5832{ 5833 int cpu = (long)hcpu; 5834 5835 switch (action & ~CPU_TASKS_FROZEN) { 5836 case CPU_ONLINE: 5837 sched_domains_numa_masks_set(cpu); 5838 break; 5839 5840 case CPU_DEAD: 5841 sched_domains_numa_masks_clear(cpu); 5842 break; 5843 5844 default: 5845 return NOTIFY_DONE; 5846 } 5847 5848 return NOTIFY_OK; 5849} 5850#else 5851static inline void sched_init_numa(void) 5852{ 5853} 5854 5855static int sched_domains_numa_masks_update(struct notifier_block *nfb, 5856 unsigned long action, 5857 void *hcpu) 5858{ 5859 return 0; 5860} 5861#endif /* CONFIG_NUMA */ 5862 5863static int __sdt_alloc(const struct cpumask *cpu_map) 5864{ 5865 struct sched_domain_topology_level *tl; 5866 int j; 5867 5868 for_each_sd_topology(tl) { 5869 struct sd_data *sdd = &tl->data; 5870 5871 sdd->sd = alloc_percpu(struct sched_domain *); 5872 if (!sdd->sd) 5873 return -ENOMEM; 5874 5875 sdd->sg = alloc_percpu(struct sched_group *); 5876 if (!sdd->sg) 5877 return -ENOMEM; 5878 5879 sdd->sgp = alloc_percpu(struct sched_group_power *); 5880 if (!sdd->sgp) 5881 return -ENOMEM; 5882 5883 for_each_cpu(j, cpu_map) { 5884 struct sched_domain *sd; 5885 struct sched_group *sg; 5886 struct sched_group_power *sgp; 5887 5888 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(), 5889 GFP_KERNEL, cpu_to_node(j)); 5890 if (!sd) 5891 return -ENOMEM; 5892 5893 *per_cpu_ptr(sdd->sd, j) = sd; 5894 5895 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(), 5896 GFP_KERNEL, cpu_to_node(j)); 5897 if (!sg) 5898 return -ENOMEM; 5899 5900 sg->next = sg; 5901 5902 *per_cpu_ptr(sdd->sg, j) = sg; 5903 5904 sgp = kzalloc_node(sizeof(struct sched_group_power) + cpumask_size(), 5905 GFP_KERNEL, cpu_to_node(j)); 5906 if (!sgp) 5907 return -ENOMEM; 5908 5909 *per_cpu_ptr(sdd->sgp, j) = sgp; 5910 } 5911 } 5912 5913 return 0; 5914} 5915 5916static void __sdt_free(const struct cpumask *cpu_map) 5917{ 5918 struct sched_domain_topology_level *tl; 5919 int j; 5920 5921 for_each_sd_topology(tl) { 5922 struct sd_data *sdd = &tl->data; 5923 5924 for_each_cpu(j, cpu_map) { 5925 struct sched_domain *sd; 5926 5927 if (sdd->sd) { 5928 sd = *per_cpu_ptr(sdd->sd, j); 5929 if (sd && (sd->flags & SD_OVERLAP)) 5930 free_sched_groups(sd->groups, 0); 5931 kfree(*per_cpu_ptr(sdd->sd, j)); 5932 } 5933 5934 if (sdd->sg) 5935 kfree(*per_cpu_ptr(sdd->sg, j)); 5936 if (sdd->sgp) 5937 kfree(*per_cpu_ptr(sdd->sgp, j)); 5938 } 5939 free_percpu(sdd->sd); 5940 sdd->sd = NULL; 5941 free_percpu(sdd->sg); 5942 sdd->sg = NULL; 5943 free_percpu(sdd->sgp); 5944 sdd->sgp = NULL; 5945 } 5946} 5947 5948struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl, 5949 const struct cpumask *cpu_map, struct sched_domain_attr *attr, 5950 struct sched_domain *child, int cpu) 5951{ 5952 struct sched_domain *sd = tl->init(tl, cpu); 5953 if (!sd) 5954 return child; 5955 5956 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu)); 5957 if (child) { 5958 sd->level = child->level + 1; 5959 sched_domain_level_max = max(sched_domain_level_max, sd->level); 5960 child->parent = sd; 5961 sd->child = child; 5962 } 5963 set_domain_attribute(sd, attr); 5964 5965 return sd; 5966} 5967 5968/* 5969 * Build sched domains for a given set of cpus and attach the sched domains 5970 * to the individual cpus 5971 */ 5972static int build_sched_domains(const struct cpumask *cpu_map, 5973 struct sched_domain_attr *attr) 5974{ 5975 enum s_alloc alloc_state; 5976 struct sched_domain *sd; 5977 struct s_data d; 5978 int i, ret = -ENOMEM; 5979 5980 alloc_state = __visit_domain_allocation_hell(&d, cpu_map); 5981 if (alloc_state != sa_rootdomain) 5982 goto error; 5983 5984 /* Set up domains for cpus specified by the cpu_map. */ 5985 for_each_cpu(i, cpu_map) { 5986 struct sched_domain_topology_level *tl; 5987 5988 sd = NULL; 5989 for_each_sd_topology(tl) { 5990 sd = build_sched_domain(tl, cpu_map, attr, sd, i); 5991 if (tl == sched_domain_topology) 5992 *per_cpu_ptr(d.sd, i) = sd; 5993 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP)) 5994 sd->flags |= SD_OVERLAP; 5995 if (cpumask_equal(cpu_map, sched_domain_span(sd))) 5996 break; 5997 } 5998 } 5999 6000 /* Build the groups for the domains */ 6001 for_each_cpu(i, cpu_map) { 6002 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) { 6003 sd->span_weight = cpumask_weight(sched_domain_span(sd)); 6004 if (sd->flags & SD_OVERLAP) { 6005 if (build_overlap_sched_groups(sd, i)) 6006 goto error; 6007 } else { 6008 if (build_sched_groups(sd, i)) 6009 goto error; 6010 } 6011 } 6012 } 6013 6014 /* Calculate CPU power for physical packages and nodes */ 6015 for (i = nr_cpumask_bits-1; i >= 0; i--) { 6016 if (!cpumask_test_cpu(i, cpu_map)) 6017 continue; 6018 6019 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) { 6020 claim_allocations(i, sd); 6021 init_sched_groups_power(i, sd); 6022 } 6023 } 6024 6025 /* Attach the domains */ 6026 rcu_read_lock(); 6027 for_each_cpu(i, cpu_map) { 6028 sd = *per_cpu_ptr(d.sd, i); 6029 cpu_attach_domain(sd, d.rd, i); 6030 } 6031 rcu_read_unlock(); 6032 6033 ret = 0; 6034error: 6035 __free_domain_allocs(&d, alloc_state, cpu_map); 6036 return ret; 6037} 6038 6039static cpumask_var_t *doms_cur; /* current sched domains */ 6040static int ndoms_cur; /* number of sched domains in 'doms_cur' */ 6041static struct sched_domain_attr *dattr_cur; 6042 /* attribues of custom domains in 'doms_cur' */ 6043 6044/* 6045 * Special case: If a kmalloc of a doms_cur partition (array of 6046 * cpumask) fails, then fallback to a single sched domain, 6047 * as determined by the single cpumask fallback_doms. 6048 */ 6049static cpumask_var_t fallback_doms; 6050 6051/* 6052 * arch_update_cpu_topology lets virtualized architectures update the 6053 * cpu core maps. It is supposed to return 1 if the topology changed 6054 * or 0 if it stayed the same. 6055 */ 6056int __attribute__((weak)) arch_update_cpu_topology(void) 6057{ 6058 return 0; 6059} 6060 6061cpumask_var_t *alloc_sched_domains(unsigned int ndoms) 6062{ 6063 int i; 6064 cpumask_var_t *doms; 6065 6066 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL); 6067 if (!doms) 6068 return NULL; 6069 for (i = 0; i < ndoms; i++) { 6070 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) { 6071 free_sched_domains(doms, i); 6072 return NULL; 6073 } 6074 } 6075 return doms; 6076} 6077 6078void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms) 6079{ 6080 unsigned int i; 6081 for (i = 0; i < ndoms; i++) 6082 free_cpumask_var(doms[i]); 6083 kfree(doms); 6084} 6085 6086/* 6087 * Set up scheduler domains and groups. Callers must hold the hotplug lock. 6088 * For now this just excludes isolated cpus, but could be used to 6089 * exclude other special cases in the future. 6090 */ 6091static int init_sched_domains(const struct cpumask *cpu_map) 6092{ 6093 int err; 6094 6095 arch_update_cpu_topology(); 6096 ndoms_cur = 1; 6097 doms_cur = alloc_sched_domains(ndoms_cur); 6098 if (!doms_cur) 6099 doms_cur = &fallback_doms; 6100 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map); 6101 err = build_sched_domains(doms_cur[0], NULL); 6102 register_sched_domain_sysctl(); 6103 6104 return err; 6105} 6106 6107/* 6108 * Detach sched domains from a group of cpus specified in cpu_map 6109 * These cpus will now be attached to the NULL domain 6110 */ 6111static void detach_destroy_domains(const struct cpumask *cpu_map) 6112{ 6113 int i; 6114 6115 rcu_read_lock(); 6116 for_each_cpu(i, cpu_map) 6117 cpu_attach_domain(NULL, &def_root_domain, i); 6118 rcu_read_unlock(); 6119} 6120 6121/* handle null as "default" */ 6122static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur, 6123 struct sched_domain_attr *new, int idx_new) 6124{ 6125 struct sched_domain_attr tmp; 6126 6127 /* fast path */ 6128 if (!new && !cur) 6129 return 1; 6130 6131 tmp = SD_ATTR_INIT; 6132 return !memcmp(cur ? (cur + idx_cur) : &tmp, 6133 new ? (new + idx_new) : &tmp, 6134 sizeof(struct sched_domain_attr)); 6135} 6136 6137/* 6138 * Partition sched domains as specified by the 'ndoms_new' 6139 * cpumasks in the array doms_new[] of cpumasks. This compares 6140 * doms_new[] to the current sched domain partitioning, doms_cur[]. 6141 * It destroys each deleted domain and builds each new domain. 6142 * 6143 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'. 6144 * The masks don't intersect (don't overlap.) We should setup one 6145 * sched domain for each mask. CPUs not in any of the cpumasks will 6146 * not be load balanced. If the same cpumask appears both in the 6147 * current 'doms_cur' domains and in the new 'doms_new', we can leave 6148 * it as it is. 6149 * 6150 * The passed in 'doms_new' should be allocated using 6151 * alloc_sched_domains. This routine takes ownership of it and will 6152 * free_sched_domains it when done with it. If the caller failed the 6153 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1, 6154 * and partition_sched_domains() will fallback to the single partition 6155 * 'fallback_doms', it also forces the domains to be rebuilt. 6156 * 6157 * If doms_new == NULL it will be replaced with cpu_online_mask. 6158 * ndoms_new == 0 is a special case for destroying existing domains, 6159 * and it will not create the default domain. 6160 * 6161 * Call with hotplug lock held 6162 */ 6163void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[], 6164 struct sched_domain_attr *dattr_new) 6165{ 6166 int i, j, n; 6167 int new_topology; 6168 6169 mutex_lock(&sched_domains_mutex); 6170 6171 /* always unregister in case we don't destroy any domains */ 6172 unregister_sched_domain_sysctl(); 6173 6174 /* Let architecture update cpu core mappings. */ 6175 new_topology = arch_update_cpu_topology(); 6176 6177 n = doms_new ? ndoms_new : 0; 6178 6179 /* Destroy deleted domains */ 6180 for (i = 0; i < ndoms_cur; i++) { 6181 for (j = 0; j < n && !new_topology; j++) { 6182 if (cpumask_equal(doms_cur[i], doms_new[j]) 6183 && dattrs_equal(dattr_cur, i, dattr_new, j)) 6184 goto match1; 6185 } 6186 /* no match - a current sched domain not in new doms_new[] */ 6187 detach_destroy_domains(doms_cur[i]); 6188match1: 6189 ; 6190 } 6191 6192 if (doms_new == NULL) { 6193 ndoms_cur = 0; 6194 doms_new = &fallback_doms; 6195 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map); 6196 WARN_ON_ONCE(dattr_new); 6197 } 6198 6199 /* Build new domains */ 6200 for (i = 0; i < ndoms_new; i++) { 6201 for (j = 0; j < ndoms_cur && !new_topology; j++) { 6202 if (cpumask_equal(doms_new[i], doms_cur[j]) 6203 && dattrs_equal(dattr_new, i, dattr_cur, j)) 6204 goto match2; 6205 } 6206 /* no match - add a new doms_new */ 6207 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL); 6208match2: 6209 ; 6210 } 6211 6212 /* Remember the new sched domains */ 6213 if (doms_cur != &fallback_doms) 6214 free_sched_domains(doms_cur, ndoms_cur); 6215 kfree(dattr_cur); /* kfree(NULL) is safe */ 6216 doms_cur = doms_new; 6217 dattr_cur = dattr_new; 6218 ndoms_cur = ndoms_new; 6219 6220 register_sched_domain_sysctl(); 6221 6222 mutex_unlock(&sched_domains_mutex); 6223} 6224 6225static int num_cpus_frozen; /* used to mark begin/end of suspend/resume */ 6226 6227/* 6228 * Update cpusets according to cpu_active mask. If cpusets are 6229 * disabled, cpuset_update_active_cpus() becomes a simple wrapper 6230 * around partition_sched_domains(). 6231 * 6232 * If we come here as part of a suspend/resume, don't touch cpusets because we 6233 * want to restore it back to its original state upon resume anyway. 6234 */ 6235static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action, 6236 void *hcpu) 6237{ 6238 switch (action) { 6239 case CPU_ONLINE_FROZEN: 6240 case CPU_DOWN_FAILED_FROZEN: 6241 6242 /* 6243 * num_cpus_frozen tracks how many CPUs are involved in suspend 6244 * resume sequence. As long as this is not the last online 6245 * operation in the resume sequence, just build a single sched 6246 * domain, ignoring cpusets. 6247 */ 6248 num_cpus_frozen--; 6249 if (likely(num_cpus_frozen)) { 6250 partition_sched_domains(1, NULL, NULL); 6251 break; 6252 } 6253 6254 /* 6255 * This is the last CPU online operation. So fall through and 6256 * restore the original sched domains by considering the 6257 * cpuset configurations. 6258 */ 6259 6260 case CPU_ONLINE: 6261 case CPU_DOWN_FAILED: 6262 cpuset_update_active_cpus(true); 6263 break; 6264 default: 6265 return NOTIFY_DONE; 6266 } 6267 return NOTIFY_OK; 6268} 6269 6270static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action, 6271 void *hcpu) 6272{ 6273 switch (action) { 6274 case CPU_DOWN_PREPARE: 6275 cpuset_update_active_cpus(false); 6276 break; 6277 case CPU_DOWN_PREPARE_FROZEN: 6278 num_cpus_frozen++; 6279 partition_sched_domains(1, NULL, NULL); 6280 break; 6281 default: 6282 return NOTIFY_DONE; 6283 } 6284 return NOTIFY_OK; 6285} 6286 6287void __init sched_init_smp(void) 6288{ 6289 cpumask_var_t non_isolated_cpus; 6290 6291 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL); 6292 alloc_cpumask_var(&fallback_doms, GFP_KERNEL); 6293 6294 sched_init_numa(); 6295 6296 get_online_cpus(); 6297 mutex_lock(&sched_domains_mutex); 6298 init_sched_domains(cpu_active_mask); 6299 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map); 6300 if (cpumask_empty(non_isolated_cpus)) 6301 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus); 6302 mutex_unlock(&sched_domains_mutex); 6303 put_online_cpus(); 6304 6305 hotcpu_notifier(sched_domains_numa_masks_update, CPU_PRI_SCHED_ACTIVE); 6306 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE); 6307 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE); 6308 6309 init_hrtick(); 6310 6311 /* Move init over to a non-isolated CPU */ 6312 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0) 6313 BUG(); 6314 sched_init_granularity(); 6315 free_cpumask_var(non_isolated_cpus); 6316 6317 init_sched_rt_class(); 6318} 6319#else 6320void __init sched_init_smp(void) 6321{ 6322 sched_init_granularity(); 6323} 6324#endif /* CONFIG_SMP */ 6325 6326const_debug unsigned int sysctl_timer_migration = 1; 6327 6328int in_sched_functions(unsigned long addr) 6329{ 6330 return in_lock_functions(addr) || 6331 (addr >= (unsigned long)__sched_text_start 6332 && addr < (unsigned long)__sched_text_end); 6333} 6334 6335#ifdef CONFIG_CGROUP_SCHED 6336/* 6337 * Default task group. 6338 * Every task in system belongs to this group at bootup. 6339 */ 6340struct task_group root_task_group; 6341LIST_HEAD(task_groups); 6342#endif 6343 6344DECLARE_PER_CPU(cpumask_var_t, load_balance_mask); 6345 6346void __init sched_init(void) 6347{ 6348 int i, j; 6349 unsigned long alloc_size = 0, ptr; 6350 6351#ifdef CONFIG_FAIR_GROUP_SCHED 6352 alloc_size += 2 * nr_cpu_ids * sizeof(void **); 6353#endif 6354#ifdef CONFIG_RT_GROUP_SCHED 6355 alloc_size += 2 * nr_cpu_ids * sizeof(void **); 6356#endif 6357#ifdef CONFIG_CPUMASK_OFFSTACK 6358 alloc_size += num_possible_cpus() * cpumask_size(); 6359#endif 6360 if (alloc_size) { 6361 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT); 6362 6363#ifdef CONFIG_FAIR_GROUP_SCHED 6364 root_task_group.se = (struct sched_entity **)ptr; 6365 ptr += nr_cpu_ids * sizeof(void **); 6366 6367 root_task_group.cfs_rq = (struct cfs_rq **)ptr; 6368 ptr += nr_cpu_ids * sizeof(void **); 6369 6370#endif /* CONFIG_FAIR_GROUP_SCHED */ 6371#ifdef CONFIG_RT_GROUP_SCHED 6372 root_task_group.rt_se = (struct sched_rt_entity **)ptr; 6373 ptr += nr_cpu_ids * sizeof(void **); 6374 6375 root_task_group.rt_rq = (struct rt_rq **)ptr; 6376 ptr += nr_cpu_ids * sizeof(void **); 6377 6378#endif /* CONFIG_RT_GROUP_SCHED */ 6379#ifdef CONFIG_CPUMASK_OFFSTACK 6380 for_each_possible_cpu(i) { 6381 per_cpu(load_balance_mask, i) = (void *)ptr; 6382 ptr += cpumask_size(); 6383 } 6384#endif /* CONFIG_CPUMASK_OFFSTACK */ 6385 } 6386 6387#ifdef CONFIG_SMP 6388 init_defrootdomain(); 6389#endif 6390 6391 init_rt_bandwidth(&def_rt_bandwidth, 6392 global_rt_period(), global_rt_runtime()); 6393 6394#ifdef CONFIG_RT_GROUP_SCHED 6395 init_rt_bandwidth(&root_task_group.rt_bandwidth, 6396 global_rt_period(), global_rt_runtime()); 6397#endif /* CONFIG_RT_GROUP_SCHED */ 6398 6399#ifdef CONFIG_CGROUP_SCHED 6400 list_add(&root_task_group.list, &task_groups); 6401 INIT_LIST_HEAD(&root_task_group.children); 6402 INIT_LIST_HEAD(&root_task_group.siblings); 6403 autogroup_init(&init_task); 6404 6405#endif /* CONFIG_CGROUP_SCHED */ 6406 6407 for_each_possible_cpu(i) { 6408 struct rq *rq; 6409 6410 rq = cpu_rq(i); 6411 raw_spin_lock_init(&rq->lock); 6412 rq->nr_running = 0; 6413 rq->calc_load_active = 0; 6414 rq->calc_load_update = jiffies + LOAD_FREQ; 6415 init_cfs_rq(&rq->cfs); 6416 init_rt_rq(&rq->rt, rq); 6417#ifdef CONFIG_FAIR_GROUP_SCHED 6418 root_task_group.shares = ROOT_TASK_GROUP_LOAD; 6419 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list); 6420 /* 6421 * How much cpu bandwidth does root_task_group get? 6422 * 6423 * In case of task-groups formed thr' the cgroup filesystem, it 6424 * gets 100% of the cpu resources in the system. This overall 6425 * system cpu resource is divided among the tasks of 6426 * root_task_group and its child task-groups in a fair manner, 6427 * based on each entity's (task or task-group's) weight 6428 * (se->load.weight). 6429 * 6430 * In other words, if root_task_group has 10 tasks of weight 6431 * 1024) and two child groups A0 and A1 (of weight 1024 each), 6432 * then A0's share of the cpu resource is: 6433 * 6434 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33% 6435 * 6436 * We achieve this by letting root_task_group's tasks sit 6437 * directly in rq->cfs (i.e root_task_group->se[] = NULL). 6438 */ 6439 init_cfs_bandwidth(&root_task_group.cfs_bandwidth); 6440 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL); 6441#endif /* CONFIG_FAIR_GROUP_SCHED */ 6442 6443 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime; 6444#ifdef CONFIG_RT_GROUP_SCHED 6445 INIT_LIST_HEAD(&rq->leaf_rt_rq_list); 6446 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL); 6447#endif 6448 6449 for (j = 0; j < CPU_LOAD_IDX_MAX; j++) 6450 rq->cpu_load[j] = 0; 6451 6452 rq->last_load_update_tick = jiffies; 6453 6454#ifdef CONFIG_SMP 6455 rq->sd = NULL; 6456 rq->rd = NULL; 6457 rq->cpu_power = SCHED_POWER_SCALE; 6458 rq->post_schedule = 0; 6459 rq->active_balance = 0; 6460 rq->next_balance = jiffies; 6461 rq->push_cpu = 0; 6462 rq->cpu = i; 6463 rq->online = 0; 6464 rq->idle_stamp = 0; 6465 rq->avg_idle = 2*sysctl_sched_migration_cost; 6466 6467 INIT_LIST_HEAD(&rq->cfs_tasks); 6468 6469 rq_attach_root(rq, &def_root_domain); 6470#ifdef CONFIG_NO_HZ_COMMON 6471 rq->nohz_flags = 0; 6472#endif 6473#ifdef CONFIG_NO_HZ_FULL 6474 rq->last_sched_tick = 0; 6475#endif 6476#endif 6477 init_rq_hrtick(rq); 6478 atomic_set(&rq->nr_iowait, 0); 6479 } 6480 6481 set_load_weight(&init_task); 6482 6483#ifdef CONFIG_PREEMPT_NOTIFIERS 6484 INIT_HLIST_HEAD(&init_task.preempt_notifiers); 6485#endif 6486 6487#ifdef CONFIG_RT_MUTEXES 6488 plist_head_init(&init_task.pi_waiters); 6489#endif 6490 6491 /* 6492 * The boot idle thread does lazy MMU switching as well: 6493 */ 6494 atomic_inc(&init_mm.mm_count); 6495 enter_lazy_tlb(&init_mm, current); 6496 6497 /* 6498 * Make us the idle thread. Technically, schedule() should not be 6499 * called from this thread, however somewhere below it might be, 6500 * but because we are the idle thread, we just pick up running again 6501 * when this runqueue becomes "idle". 6502 */ 6503 init_idle(current, smp_processor_id()); 6504 6505 calc_load_update = jiffies + LOAD_FREQ; 6506 6507 /* 6508 * During early bootup we pretend to be a normal task: 6509 */ 6510 current->sched_class = &fair_sched_class; 6511 6512#ifdef CONFIG_SMP 6513 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT); 6514 /* May be allocated at isolcpus cmdline parse time */ 6515 if (cpu_isolated_map == NULL) 6516 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT); 6517 idle_thread_set_boot_cpu(); 6518#endif 6519 init_sched_fair_class(); 6520 6521 scheduler_running = 1; 6522} 6523 6524#ifdef CONFIG_DEBUG_ATOMIC_SLEEP 6525static inline int preempt_count_equals(int preempt_offset) 6526{ 6527 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth(); 6528 6529 return (nested == preempt_offset); 6530} 6531 6532void __might_sleep(const char *file, int line, int preempt_offset) 6533{ 6534 static unsigned long prev_jiffy; /* ratelimiting */ 6535 6536 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */ 6537 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) || 6538 system_state != SYSTEM_RUNNING || oops_in_progress) 6539 return; 6540 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy) 6541 return; 6542 prev_jiffy = jiffies; 6543 6544 printk(KERN_ERR 6545 "BUG: sleeping function called from invalid context at %s:%d\n", 6546 file, line); 6547 printk(KERN_ERR 6548 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n", 6549 in_atomic(), irqs_disabled(), 6550 current->pid, current->comm); 6551 6552 debug_show_held_locks(current); 6553 if (irqs_disabled()) 6554 print_irqtrace_events(current); 6555 dump_stack(); 6556} 6557EXPORT_SYMBOL(__might_sleep); 6558#endif 6559 6560#ifdef CONFIG_MAGIC_SYSRQ 6561static void normalize_task(struct rq *rq, struct task_struct *p) 6562{ 6563 const struct sched_class *prev_class = p->sched_class; 6564 int old_prio = p->prio; 6565 int on_rq; 6566 6567 on_rq = p->on_rq; 6568 if (on_rq) 6569 dequeue_task(rq, p, 0); 6570 __setscheduler(rq, p, SCHED_NORMAL, 0); 6571 if (on_rq) { 6572 enqueue_task(rq, p, 0); 6573 resched_task(rq->curr); 6574 } 6575 6576 check_class_changed(rq, p, prev_class, old_prio); 6577} 6578 6579void normalize_rt_tasks(void) 6580{ 6581 struct task_struct *g, *p; 6582 unsigned long flags; 6583 struct rq *rq; 6584 6585 read_lock_irqsave(&tasklist_lock, flags); 6586 do_each_thread(g, p) { 6587 /* 6588 * Only normalize user tasks: 6589 */ 6590 if (!p->mm) 6591 continue; 6592 6593 p->se.exec_start = 0; 6594#ifdef CONFIG_SCHEDSTATS 6595 p->se.statistics.wait_start = 0; 6596 p->se.statistics.sleep_start = 0; 6597 p->se.statistics.block_start = 0; 6598#endif 6599 6600 if (!rt_task(p)) { 6601 /* 6602 * Renice negative nice level userspace 6603 * tasks back to 0: 6604 */ 6605 if (TASK_NICE(p) < 0 && p->mm) 6606 set_user_nice(p, 0); 6607 continue; 6608 } 6609 6610 raw_spin_lock(&p->pi_lock); 6611 rq = __task_rq_lock(p); 6612 6613 normalize_task(rq, p); 6614 6615 __task_rq_unlock(rq); 6616 raw_spin_unlock(&p->pi_lock); 6617 } while_each_thread(g, p); 6618 6619 read_unlock_irqrestore(&tasklist_lock, flags); 6620} 6621 6622#endif /* CONFIG_MAGIC_SYSRQ */ 6623 6624#if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) 6625/* 6626 * These functions are only useful for the IA64 MCA handling, or kdb. 6627 * 6628 * They can only be called when the whole system has been 6629 * stopped - every CPU needs to be quiescent, and no scheduling 6630 * activity can take place. Using them for anything else would 6631 * be a serious bug, and as a result, they aren't even visible 6632 * under any other configuration. 6633 */ 6634 6635/** 6636 * curr_task - return the current task for a given cpu. 6637 * @cpu: the processor in question. 6638 * 6639 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED! 6640 */ 6641struct task_struct *curr_task(int cpu) 6642{ 6643 return cpu_curr(cpu); 6644} 6645 6646#endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */ 6647 6648#ifdef CONFIG_IA64 6649/** 6650 * set_curr_task - set the current task for a given cpu. 6651 * @cpu: the processor in question. 6652 * @p: the task pointer to set. 6653 * 6654 * Description: This function must only be used when non-maskable interrupts 6655 * are serviced on a separate stack. It allows the architecture to switch the 6656 * notion of the current task on a cpu in a non-blocking manner. This function 6657 * must be called with all CPU's synchronized, and interrupts disabled, the 6658 * and caller must save the original value of the current task (see 6659 * curr_task() above) and restore that value before reenabling interrupts and 6660 * re-starting the system. 6661 * 6662 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED! 6663 */ 6664void set_curr_task(int cpu, struct task_struct *p) 6665{ 6666 cpu_curr(cpu) = p; 6667} 6668 6669#endif 6670 6671#ifdef CONFIG_CGROUP_SCHED 6672/* task_group_lock serializes the addition/removal of task groups */ 6673static DEFINE_SPINLOCK(task_group_lock); 6674 6675static void free_sched_group(struct task_group *tg) 6676{ 6677 free_fair_sched_group(tg); 6678 free_rt_sched_group(tg); 6679 autogroup_free(tg); 6680 kfree(tg); 6681} 6682 6683/* allocate runqueue etc for a new task group */ 6684struct task_group *sched_create_group(struct task_group *parent) 6685{ 6686 struct task_group *tg; 6687 6688 tg = kzalloc(sizeof(*tg), GFP_KERNEL); 6689 if (!tg) 6690 return ERR_PTR(-ENOMEM); 6691 6692 if (!alloc_fair_sched_group(tg, parent)) 6693 goto err; 6694 6695 if (!alloc_rt_sched_group(tg, parent)) 6696 goto err; 6697 6698 return tg; 6699 6700err: 6701 free_sched_group(tg); 6702 return ERR_PTR(-ENOMEM); 6703} 6704 6705void sched_online_group(struct task_group *tg, struct task_group *parent) 6706{ 6707 unsigned long flags; 6708 6709 spin_lock_irqsave(&task_group_lock, flags); 6710 list_add_rcu(&tg->list, &task_groups); 6711 6712 WARN_ON(!parent); /* root should already exist */ 6713 6714 tg->parent = parent; 6715 INIT_LIST_HEAD(&tg->children); 6716 list_add_rcu(&tg->siblings, &parent->children); 6717 spin_unlock_irqrestore(&task_group_lock, flags); 6718} 6719 6720/* rcu callback to free various structures associated with a task group */ 6721static void free_sched_group_rcu(struct rcu_head *rhp) 6722{ 6723 /* now it should be safe to free those cfs_rqs */ 6724 free_sched_group(container_of(rhp, struct task_group, rcu)); 6725} 6726 6727/* Destroy runqueue etc associated with a task group */ 6728void sched_destroy_group(struct task_group *tg) 6729{ 6730 /* wait for possible concurrent references to cfs_rqs complete */ 6731 call_rcu(&tg->rcu, free_sched_group_rcu); 6732} 6733 6734void sched_offline_group(struct task_group *tg) 6735{ 6736 unsigned long flags; 6737 int i; 6738 6739 /* end participation in shares distribution */ 6740 for_each_possible_cpu(i) 6741 unregister_fair_sched_group(tg, i); 6742 6743 spin_lock_irqsave(&task_group_lock, flags); 6744 list_del_rcu(&tg->list); 6745 list_del_rcu(&tg->siblings); 6746 spin_unlock_irqrestore(&task_group_lock, flags); 6747} 6748 6749/* change task's runqueue when it moves between groups. 6750 * The caller of this function should have put the task in its new group 6751 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to 6752 * reflect its new group. 6753 */ 6754void sched_move_task(struct task_struct *tsk) 6755{ 6756 struct task_group *tg; 6757 int on_rq, running; 6758 unsigned long flags; 6759 struct rq *rq; 6760 6761 rq = task_rq_lock(tsk, &flags); 6762 6763 running = task_current(rq, tsk); 6764 on_rq = tsk->on_rq; 6765 6766 if (on_rq) 6767 dequeue_task(rq, tsk, 0); 6768 if (unlikely(running)) 6769 tsk->sched_class->put_prev_task(rq, tsk); 6770 6771 tg = container_of(task_subsys_state_check(tsk, cpu_cgroup_subsys_id, 6772 lockdep_is_held(&tsk->sighand->siglock)), 6773 struct task_group, css); 6774 tg = autogroup_task_group(tsk, tg); 6775 tsk->sched_task_group = tg; 6776 6777#ifdef CONFIG_FAIR_GROUP_SCHED 6778 if (tsk->sched_class->task_move_group) 6779 tsk->sched_class->task_move_group(tsk, on_rq); 6780 else 6781#endif 6782 set_task_rq(tsk, task_cpu(tsk)); 6783 6784 if (unlikely(running)) 6785 tsk->sched_class->set_curr_task(rq); 6786 if (on_rq) 6787 enqueue_task(rq, tsk, 0); 6788 6789 task_rq_unlock(rq, tsk, &flags); 6790} 6791#endif /* CONFIG_CGROUP_SCHED */ 6792 6793#if defined(CONFIG_RT_GROUP_SCHED) || defined(CONFIG_CFS_BANDWIDTH) 6794static unsigned long to_ratio(u64 period, u64 runtime) 6795{ 6796 if (runtime == RUNTIME_INF) 6797 return 1ULL << 20; 6798 6799 return div64_u64(runtime << 20, period); 6800} 6801#endif 6802 6803#ifdef CONFIG_RT_GROUP_SCHED 6804/* 6805 * Ensure that the real time constraints are schedulable. 6806 */ 6807static DEFINE_MUTEX(rt_constraints_mutex); 6808 6809/* Must be called with tasklist_lock held */ 6810static inline int tg_has_rt_tasks(struct task_group *tg) 6811{ 6812 struct task_struct *g, *p; 6813 6814 do_each_thread(g, p) { 6815 if (rt_task(p) && task_rq(p)->rt.tg == tg) 6816 return 1; 6817 } while_each_thread(g, p); 6818 6819 return 0; 6820} 6821 6822struct rt_schedulable_data { 6823 struct task_group *tg; 6824 u64 rt_period; 6825 u64 rt_runtime; 6826}; 6827 6828static int tg_rt_schedulable(struct task_group *tg, void *data) 6829{ 6830 struct rt_schedulable_data *d = data; 6831 struct task_group *child; 6832 unsigned long total, sum = 0; 6833 u64 period, runtime; 6834 6835 period = ktime_to_ns(tg->rt_bandwidth.rt_period); 6836 runtime = tg->rt_bandwidth.rt_runtime; 6837 6838 if (tg == d->tg) { 6839 period = d->rt_period; 6840 runtime = d->rt_runtime; 6841 } 6842 6843 /* 6844 * Cannot have more runtime than the period. 6845 */ 6846 if (runtime > period && runtime != RUNTIME_INF) 6847 return -EINVAL; 6848 6849 /* 6850 * Ensure we don't starve existing RT tasks. 6851 */ 6852 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg)) 6853 return -EBUSY; 6854 6855 total = to_ratio(period, runtime); 6856 6857 /* 6858 * Nobody can have more than the global setting allows. 6859 */ 6860 if (total > to_ratio(global_rt_period(), global_rt_runtime())) 6861 return -EINVAL; 6862 6863 /* 6864 * The sum of our children's runtime should not exceed our own. 6865 */ 6866 list_for_each_entry_rcu(child, &tg->children, siblings) { 6867 period = ktime_to_ns(child->rt_bandwidth.rt_period); 6868 runtime = child->rt_bandwidth.rt_runtime; 6869 6870 if (child == d->tg) { 6871 period = d->rt_period; 6872 runtime = d->rt_runtime; 6873 } 6874 6875 sum += to_ratio(period, runtime); 6876 } 6877 6878 if (sum > total) 6879 return -EINVAL; 6880 6881 return 0; 6882} 6883 6884static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime) 6885{ 6886 int ret; 6887 6888 struct rt_schedulable_data data = { 6889 .tg = tg, 6890 .rt_period = period, 6891 .rt_runtime = runtime, 6892 }; 6893 6894 rcu_read_lock(); 6895 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data); 6896 rcu_read_unlock(); 6897 6898 return ret; 6899} 6900 6901static int tg_set_rt_bandwidth(struct task_group *tg, 6902 u64 rt_period, u64 rt_runtime) 6903{ 6904 int i, err = 0; 6905 6906 mutex_lock(&rt_constraints_mutex); 6907 read_lock(&tasklist_lock); 6908 err = __rt_schedulable(tg, rt_period, rt_runtime); 6909 if (err) 6910 goto unlock; 6911 6912 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock); 6913 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period); 6914 tg->rt_bandwidth.rt_runtime = rt_runtime; 6915 6916 for_each_possible_cpu(i) { 6917 struct rt_rq *rt_rq = tg->rt_rq[i]; 6918 6919 raw_spin_lock(&rt_rq->rt_runtime_lock); 6920 rt_rq->rt_runtime = rt_runtime; 6921 raw_spin_unlock(&rt_rq->rt_runtime_lock); 6922 } 6923 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock); 6924unlock: 6925 read_unlock(&tasklist_lock); 6926 mutex_unlock(&rt_constraints_mutex); 6927 6928 return err; 6929} 6930 6931static int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us) 6932{ 6933 u64 rt_runtime, rt_period; 6934 6935 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period); 6936 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC; 6937 if (rt_runtime_us < 0) 6938 rt_runtime = RUNTIME_INF; 6939 6940 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime); 6941} 6942 6943static long sched_group_rt_runtime(struct task_group *tg) 6944{ 6945 u64 rt_runtime_us; 6946 6947 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF) 6948 return -1; 6949 6950 rt_runtime_us = tg->rt_bandwidth.rt_runtime; 6951 do_div(rt_runtime_us, NSEC_PER_USEC); 6952 return rt_runtime_us; 6953} 6954 6955static int sched_group_set_rt_period(struct task_group *tg, long rt_period_us) 6956{ 6957 u64 rt_runtime, rt_period; 6958 6959 rt_period = (u64)rt_period_us * NSEC_PER_USEC; 6960 rt_runtime = tg->rt_bandwidth.rt_runtime; 6961 6962 if (rt_period == 0) 6963 return -EINVAL; 6964 6965 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime); 6966} 6967 6968static long sched_group_rt_period(struct task_group *tg) 6969{ 6970 u64 rt_period_us; 6971 6972 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period); 6973 do_div(rt_period_us, NSEC_PER_USEC); 6974 return rt_period_us; 6975} 6976 6977static int sched_rt_global_constraints(void) 6978{ 6979 u64 runtime, period; 6980 int ret = 0; 6981 6982 if (sysctl_sched_rt_period <= 0) 6983 return -EINVAL; 6984 6985 runtime = global_rt_runtime(); 6986 period = global_rt_period(); 6987 6988 /* 6989 * Sanity check on the sysctl variables. 6990 */ 6991 if (runtime > period && runtime != RUNTIME_INF) 6992 return -EINVAL; 6993 6994 mutex_lock(&rt_constraints_mutex); 6995 read_lock(&tasklist_lock); 6996 ret = __rt_schedulable(NULL, 0, 0); 6997 read_unlock(&tasklist_lock); 6998 mutex_unlock(&rt_constraints_mutex); 6999 7000 return ret; 7001} 7002 7003static int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk) 7004{ 7005 /* Don't accept realtime tasks when there is no way for them to run */ 7006 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0) 7007 return 0; 7008 7009 return 1; 7010} 7011 7012#else /* !CONFIG_RT_GROUP_SCHED */ 7013static int sched_rt_global_constraints(void) 7014{ 7015 unsigned long flags; 7016 int i; 7017 7018 if (sysctl_sched_rt_period <= 0) 7019 return -EINVAL; 7020 7021 /* 7022 * There's always some RT tasks in the root group 7023 * -- migration, kstopmachine etc.. 7024 */ 7025 if (sysctl_sched_rt_runtime == 0) 7026 return -EBUSY; 7027 7028 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags); 7029 for_each_possible_cpu(i) { 7030 struct rt_rq *rt_rq = &cpu_rq(i)->rt; 7031 7032 raw_spin_lock(&rt_rq->rt_runtime_lock); 7033 rt_rq->rt_runtime = global_rt_runtime(); 7034 raw_spin_unlock(&rt_rq->rt_runtime_lock); 7035 } 7036 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags); 7037 7038 return 0; 7039} 7040#endif /* CONFIG_RT_GROUP_SCHED */ 7041 7042int sched_rr_handler(struct ctl_table *table, int write, 7043 void __user *buffer, size_t *lenp, 7044 loff_t *ppos) 7045{ 7046 int ret; 7047 static DEFINE_MUTEX(mutex); 7048 7049 mutex_lock(&mutex); 7050 ret = proc_dointvec(table, write, buffer, lenp, ppos); 7051 /* make sure that internally we keep jiffies */ 7052 /* also, writing zero resets timeslice to default */ 7053 if (!ret && write) { 7054 sched_rr_timeslice = sched_rr_timeslice <= 0 ? 7055 RR_TIMESLICE : msecs_to_jiffies(sched_rr_timeslice); 7056 } 7057 mutex_unlock(&mutex); 7058 return ret; 7059} 7060 7061int sched_rt_handler(struct ctl_table *table, int write, 7062 void __user *buffer, size_t *lenp, 7063 loff_t *ppos) 7064{ 7065 int ret; 7066 int old_period, old_runtime; 7067 static DEFINE_MUTEX(mutex); 7068 7069 mutex_lock(&mutex); 7070 old_period = sysctl_sched_rt_period; 7071 old_runtime = sysctl_sched_rt_runtime; 7072 7073 ret = proc_dointvec(table, write, buffer, lenp, ppos); 7074 7075 if (!ret && write) { 7076 ret = sched_rt_global_constraints(); 7077 if (ret) { 7078 sysctl_sched_rt_period = old_period; 7079 sysctl_sched_rt_runtime = old_runtime; 7080 } else { 7081 def_rt_bandwidth.rt_runtime = global_rt_runtime(); 7082 def_rt_bandwidth.rt_period = 7083 ns_to_ktime(global_rt_period()); 7084 } 7085 } 7086 mutex_unlock(&mutex); 7087 7088 return ret; 7089} 7090 7091#ifdef CONFIG_CGROUP_SCHED 7092 7093/* return corresponding task_group object of a cgroup */ 7094static inline struct task_group *cgroup_tg(struct cgroup *cgrp) 7095{ 7096 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id), 7097 struct task_group, css); 7098} 7099 7100static struct cgroup_subsys_state *cpu_cgroup_css_alloc(struct cgroup *cgrp) 7101{ 7102 struct task_group *tg, *parent; 7103 7104 if (!cgrp->parent) { 7105 /* This is early initialization for the top cgroup */ 7106 return &root_task_group.css; 7107 } 7108 7109 parent = cgroup_tg(cgrp->parent); 7110 tg = sched_create_group(parent); 7111 if (IS_ERR(tg)) 7112 return ERR_PTR(-ENOMEM); 7113 7114 return &tg->css; 7115} 7116 7117static int cpu_cgroup_css_online(struct cgroup *cgrp) 7118{ 7119 struct task_group *tg = cgroup_tg(cgrp); 7120 struct task_group *parent; 7121 7122 if (!cgrp->parent) 7123 return 0; 7124 7125 parent = cgroup_tg(cgrp->parent); 7126 sched_online_group(tg, parent); 7127 return 0; 7128} 7129 7130static void cpu_cgroup_css_free(struct cgroup *cgrp) 7131{ 7132 struct task_group *tg = cgroup_tg(cgrp); 7133 7134 sched_destroy_group(tg); 7135} 7136 7137static void cpu_cgroup_css_offline(struct cgroup *cgrp) 7138{ 7139 struct task_group *tg = cgroup_tg(cgrp); 7140 7141 sched_offline_group(tg); 7142} 7143 7144static int cpu_cgroup_can_attach(struct cgroup *cgrp, 7145 struct cgroup_taskset *tset) 7146{ 7147 struct task_struct *task; 7148 7149 cgroup_taskset_for_each(task, cgrp, tset) { 7150#ifdef CONFIG_RT_GROUP_SCHED 7151 if (!sched_rt_can_attach(cgroup_tg(cgrp), task)) 7152 return -EINVAL; 7153#else 7154 /* We don't support RT-tasks being in separate groups */ 7155 if (task->sched_class != &fair_sched_class) 7156 return -EINVAL; 7157#endif 7158 } 7159 return 0; 7160} 7161 7162static void cpu_cgroup_attach(struct cgroup *cgrp, 7163 struct cgroup_taskset *tset) 7164{ 7165 struct task_struct *task; 7166 7167 cgroup_taskset_for_each(task, cgrp, tset) 7168 sched_move_task(task); 7169} 7170 7171static void 7172cpu_cgroup_exit(struct cgroup *cgrp, struct cgroup *old_cgrp, 7173 struct task_struct *task) 7174{ 7175 /* 7176 * cgroup_exit() is called in the copy_process() failure path. 7177 * Ignore this case since the task hasn't ran yet, this avoids 7178 * trying to poke a half freed task state from generic code. 7179 */ 7180 if (!(task->flags & PF_EXITING)) 7181 return; 7182 7183 sched_move_task(task); 7184} 7185 7186#ifdef CONFIG_FAIR_GROUP_SCHED 7187static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype, 7188 u64 shareval) 7189{ 7190 return sched_group_set_shares(cgroup_tg(cgrp), scale_load(shareval)); 7191} 7192 7193static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft) 7194{ 7195 struct task_group *tg = cgroup_tg(cgrp); 7196 7197 return (u64) scale_load_down(tg->shares); 7198} 7199 7200#ifdef CONFIG_CFS_BANDWIDTH 7201static DEFINE_MUTEX(cfs_constraints_mutex); 7202 7203const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */ 7204const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */ 7205 7206static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime); 7207 7208static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota) 7209{ 7210 int i, ret = 0, runtime_enabled, runtime_was_enabled; 7211 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth; 7212 7213 if (tg == &root_task_group) 7214 return -EINVAL; 7215 7216 /* 7217 * Ensure we have at some amount of bandwidth every period. This is 7218 * to prevent reaching a state of large arrears when throttled via 7219 * entity_tick() resulting in prolonged exit starvation. 7220 */ 7221 if (quota < min_cfs_quota_period || period < min_cfs_quota_period) 7222 return -EINVAL; 7223 7224 /* 7225 * Likewise, bound things on the otherside by preventing insane quota 7226 * periods. This also allows us to normalize in computing quota 7227 * feasibility. 7228 */ 7229 if (period > max_cfs_quota_period) 7230 return -EINVAL; 7231 7232 mutex_lock(&cfs_constraints_mutex); 7233 ret = __cfs_schedulable(tg, period, quota); 7234 if (ret) 7235 goto out_unlock; 7236 7237 runtime_enabled = quota != RUNTIME_INF; 7238 runtime_was_enabled = cfs_b->quota != RUNTIME_INF; 7239 account_cfs_bandwidth_used(runtime_enabled, runtime_was_enabled); 7240 raw_spin_lock_irq(&cfs_b->lock); 7241 cfs_b->period = ns_to_ktime(period); 7242 cfs_b->quota = quota; 7243 7244 __refill_cfs_bandwidth_runtime(cfs_b); 7245 /* restart the period timer (if active) to handle new period expiry */ 7246 if (runtime_enabled && cfs_b->timer_active) { 7247 /* force a reprogram */ 7248 cfs_b->timer_active = 0; 7249 __start_cfs_bandwidth(cfs_b); 7250 } 7251 raw_spin_unlock_irq(&cfs_b->lock); 7252 7253 for_each_possible_cpu(i) { 7254 struct cfs_rq *cfs_rq = tg->cfs_rq[i]; 7255 struct rq *rq = cfs_rq->rq; 7256 7257 raw_spin_lock_irq(&rq->lock); 7258 cfs_rq->runtime_enabled = runtime_enabled; 7259 cfs_rq->runtime_remaining = 0; 7260 7261 if (cfs_rq->throttled) 7262 unthrottle_cfs_rq(cfs_rq); 7263 raw_spin_unlock_irq(&rq->lock); 7264 } 7265out_unlock: 7266 mutex_unlock(&cfs_constraints_mutex); 7267 7268 return ret; 7269} 7270 7271int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us) 7272{ 7273 u64 quota, period; 7274 7275 period = ktime_to_ns(tg->cfs_bandwidth.period); 7276 if (cfs_quota_us < 0) 7277 quota = RUNTIME_INF; 7278 else 7279 quota = (u64)cfs_quota_us * NSEC_PER_USEC; 7280 7281 return tg_set_cfs_bandwidth(tg, period, quota); 7282} 7283 7284long tg_get_cfs_quota(struct task_group *tg) 7285{ 7286 u64 quota_us; 7287 7288 if (tg->cfs_bandwidth.quota == RUNTIME_INF) 7289 return -1; 7290 7291 quota_us = tg->cfs_bandwidth.quota; 7292 do_div(quota_us, NSEC_PER_USEC); 7293 7294 return quota_us; 7295} 7296 7297int tg_set_cfs_period(struct task_group *tg, long cfs_period_us) 7298{ 7299 u64 quota, period; 7300 7301 period = (u64)cfs_period_us * NSEC_PER_USEC; 7302 quota = tg->cfs_bandwidth.quota; 7303 7304 return tg_set_cfs_bandwidth(tg, period, quota); 7305} 7306 7307long tg_get_cfs_period(struct task_group *tg) 7308{ 7309 u64 cfs_period_us; 7310 7311 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period); 7312 do_div(cfs_period_us, NSEC_PER_USEC); 7313 7314 return cfs_period_us; 7315} 7316 7317static s64 cpu_cfs_quota_read_s64(struct cgroup *cgrp, struct cftype *cft) 7318{ 7319 return tg_get_cfs_quota(cgroup_tg(cgrp)); 7320} 7321 7322static int cpu_cfs_quota_write_s64(struct cgroup *cgrp, struct cftype *cftype, 7323 s64 cfs_quota_us) 7324{ 7325 return tg_set_cfs_quota(cgroup_tg(cgrp), cfs_quota_us); 7326} 7327 7328static u64 cpu_cfs_period_read_u64(struct cgroup *cgrp, struct cftype *cft) 7329{ 7330 return tg_get_cfs_period(cgroup_tg(cgrp)); 7331} 7332 7333static int cpu_cfs_period_write_u64(struct cgroup *cgrp, struct cftype *cftype, 7334 u64 cfs_period_us) 7335{ 7336 return tg_set_cfs_period(cgroup_tg(cgrp), cfs_period_us); 7337} 7338 7339struct cfs_schedulable_data { 7340 struct task_group *tg; 7341 u64 period, quota; 7342}; 7343 7344/* 7345 * normalize group quota/period to be quota/max_period 7346 * note: units are usecs 7347 */ 7348static u64 normalize_cfs_quota(struct task_group *tg, 7349 struct cfs_schedulable_data *d) 7350{ 7351 u64 quota, period; 7352 7353 if (tg == d->tg) { 7354 period = d->period; 7355 quota = d->quota; 7356 } else { 7357 period = tg_get_cfs_period(tg); 7358 quota = tg_get_cfs_quota(tg); 7359 } 7360 7361 /* note: these should typically be equivalent */ 7362 if (quota == RUNTIME_INF || quota == -1) 7363 return RUNTIME_INF; 7364 7365 return to_ratio(period, quota); 7366} 7367 7368static int tg_cfs_schedulable_down(struct task_group *tg, void *data) 7369{ 7370 struct cfs_schedulable_data *d = data; 7371 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth; 7372 s64 quota = 0, parent_quota = -1; 7373 7374 if (!tg->parent) { 7375 quota = RUNTIME_INF; 7376 } else { 7377 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth; 7378 7379 quota = normalize_cfs_quota(tg, d); 7380 parent_quota = parent_b->hierarchal_quota; 7381 7382 /* 7383 * ensure max(child_quota) <= parent_quota, inherit when no 7384 * limit is set 7385 */ 7386 if (quota == RUNTIME_INF) 7387 quota = parent_quota; 7388 else if (parent_quota != RUNTIME_INF && quota > parent_quota) 7389 return -EINVAL; 7390 } 7391 cfs_b->hierarchal_quota = quota; 7392 7393 return 0; 7394} 7395 7396static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota) 7397{ 7398 int ret; 7399 struct cfs_schedulable_data data = { 7400 .tg = tg, 7401 .period = period, 7402 .quota = quota, 7403 }; 7404 7405 if (quota != RUNTIME_INF) { 7406 do_div(data.period, NSEC_PER_USEC); 7407 do_div(data.quota, NSEC_PER_USEC); 7408 } 7409 7410 rcu_read_lock(); 7411 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data); 7412 rcu_read_unlock(); 7413 7414 return ret; 7415} 7416 7417static int cpu_stats_show(struct cgroup *cgrp, struct cftype *cft, 7418 struct cgroup_map_cb *cb) 7419{ 7420 struct task_group *tg = cgroup_tg(cgrp); 7421 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth; 7422 7423 cb->fill(cb, "nr_periods", cfs_b->nr_periods); 7424 cb->fill(cb, "nr_throttled", cfs_b->nr_throttled); 7425 cb->fill(cb, "throttled_time", cfs_b->throttled_time); 7426 7427 return 0; 7428} 7429#endif /* CONFIG_CFS_BANDWIDTH */ 7430#endif /* CONFIG_FAIR_GROUP_SCHED */ 7431 7432#ifdef CONFIG_RT_GROUP_SCHED 7433static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft, 7434 s64 val) 7435{ 7436 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val); 7437} 7438 7439static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft) 7440{ 7441 return sched_group_rt_runtime(cgroup_tg(cgrp)); 7442} 7443 7444static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype, 7445 u64 rt_period_us) 7446{ 7447 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us); 7448} 7449 7450static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft) 7451{ 7452 return sched_group_rt_period(cgroup_tg(cgrp)); 7453} 7454#endif /* CONFIG_RT_GROUP_SCHED */ 7455 7456static struct cftype cpu_files[] = { 7457#ifdef CONFIG_FAIR_GROUP_SCHED 7458 { 7459 .name = "shares", 7460 .read_u64 = cpu_shares_read_u64, 7461 .write_u64 = cpu_shares_write_u64, 7462 }, 7463#endif 7464#ifdef CONFIG_CFS_BANDWIDTH 7465 { 7466 .name = "cfs_quota_us", 7467 .read_s64 = cpu_cfs_quota_read_s64, 7468 .write_s64 = cpu_cfs_quota_write_s64, 7469 }, 7470 { 7471 .name = "cfs_period_us", 7472 .read_u64 = cpu_cfs_period_read_u64, 7473 .write_u64 = cpu_cfs_period_write_u64, 7474 }, 7475 { 7476 .name = "stat", 7477 .read_map = cpu_stats_show, 7478 }, 7479#endif 7480#ifdef CONFIG_RT_GROUP_SCHED 7481 { 7482 .name = "rt_runtime_us", 7483 .read_s64 = cpu_rt_runtime_read, 7484 .write_s64 = cpu_rt_runtime_write, 7485 }, 7486 { 7487 .name = "rt_period_us", 7488 .read_u64 = cpu_rt_period_read_uint, 7489 .write_u64 = cpu_rt_period_write_uint, 7490 }, 7491#endif 7492 { } /* terminate */ 7493}; 7494 7495struct cgroup_subsys cpu_cgroup_subsys = { 7496 .name = "cpu", 7497 .css_alloc = cpu_cgroup_css_alloc, 7498 .css_free = cpu_cgroup_css_free, 7499 .css_online = cpu_cgroup_css_online, 7500 .css_offline = cpu_cgroup_css_offline, 7501 .can_attach = cpu_cgroup_can_attach, 7502 .attach = cpu_cgroup_attach, 7503 .exit = cpu_cgroup_exit, 7504 .subsys_id = cpu_cgroup_subsys_id, 7505 .base_cftypes = cpu_files, 7506 .early_init = 1, 7507}; 7508 7509#endif /* CONFIG_CGROUP_SCHED */ 7510 7511void dump_cpu_task(int cpu) 7512{ 7513 pr_info("Task dump for CPU %d:\n", cpu); 7514 sched_show_task(cpu_curr(cpu)); 7515} 7516