core.c revision e69f61862ab833e9b8d3c15b6ce07fd69f3bfecc
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 * 937 * Return: 1 if the task is currently executing. 0 otherwise. 938 */ 939inline int task_curr(const struct task_struct *p) 940{ 941 return cpu_curr(task_cpu(p)) == p; 942} 943 944static inline void check_class_changed(struct rq *rq, struct task_struct *p, 945 const struct sched_class *prev_class, 946 int oldprio) 947{ 948 if (prev_class != p->sched_class) { 949 if (prev_class->switched_from) 950 prev_class->switched_from(rq, p); 951 p->sched_class->switched_to(rq, p); 952 } else if (oldprio != p->prio) 953 p->sched_class->prio_changed(rq, p, oldprio); 954} 955 956void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags) 957{ 958 const struct sched_class *class; 959 960 if (p->sched_class == rq->curr->sched_class) { 961 rq->curr->sched_class->check_preempt_curr(rq, p, flags); 962 } else { 963 for_each_class(class) { 964 if (class == rq->curr->sched_class) 965 break; 966 if (class == p->sched_class) { 967 resched_task(rq->curr); 968 break; 969 } 970 } 971 } 972 973 /* 974 * A queue event has occurred, and we're going to schedule. In 975 * this case, we can save a useless back to back clock update. 976 */ 977 if (rq->curr->on_rq && test_tsk_need_resched(rq->curr)) 978 rq->skip_clock_update = 1; 979} 980 981static ATOMIC_NOTIFIER_HEAD(task_migration_notifier); 982 983void register_task_migration_notifier(struct notifier_block *n) 984{ 985 atomic_notifier_chain_register(&task_migration_notifier, n); 986} 987 988#ifdef CONFIG_SMP 989void set_task_cpu(struct task_struct *p, unsigned int new_cpu) 990{ 991#ifdef CONFIG_SCHED_DEBUG 992 /* 993 * We should never call set_task_cpu() on a blocked task, 994 * ttwu() will sort out the placement. 995 */ 996 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING && 997 !(task_thread_info(p)->preempt_count & PREEMPT_ACTIVE)); 998 999#ifdef CONFIG_LOCKDEP 1000 /* 1001 * The caller should hold either p->pi_lock or rq->lock, when changing 1002 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks. 1003 * 1004 * sched_move_task() holds both and thus holding either pins the cgroup, 1005 * see task_group(). 1006 * 1007 * Furthermore, all task_rq users should acquire both locks, see 1008 * task_rq_lock(). 1009 */ 1010 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) || 1011 lockdep_is_held(&task_rq(p)->lock))); 1012#endif 1013#endif 1014 1015 trace_sched_migrate_task(p, new_cpu); 1016 1017 if (task_cpu(p) != new_cpu) { 1018 struct task_migration_notifier tmn; 1019 1020 if (p->sched_class->migrate_task_rq) 1021 p->sched_class->migrate_task_rq(p, new_cpu); 1022 p->se.nr_migrations++; 1023 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, NULL, 0); 1024 1025 tmn.task = p; 1026 tmn.from_cpu = task_cpu(p); 1027 tmn.to_cpu = new_cpu; 1028 1029 atomic_notifier_call_chain(&task_migration_notifier, 0, &tmn); 1030 } 1031 1032 __set_task_cpu(p, new_cpu); 1033} 1034 1035struct migration_arg { 1036 struct task_struct *task; 1037 int dest_cpu; 1038}; 1039 1040static int migration_cpu_stop(void *data); 1041 1042/* 1043 * wait_task_inactive - wait for a thread to unschedule. 1044 * 1045 * If @match_state is nonzero, it's the @p->state value just checked and 1046 * not expected to change. If it changes, i.e. @p might have woken up, 1047 * then return zero. When we succeed in waiting for @p to be off its CPU, 1048 * we return a positive number (its total switch count). If a second call 1049 * a short while later returns the same number, the caller can be sure that 1050 * @p has remained unscheduled the whole time. 1051 * 1052 * The caller must ensure that the task *will* unschedule sometime soon, 1053 * else this function might spin for a *long* time. This function can't 1054 * be called with interrupts off, or it may introduce deadlock with 1055 * smp_call_function() if an IPI is sent by the same process we are 1056 * waiting to become inactive. 1057 */ 1058unsigned long wait_task_inactive(struct task_struct *p, long match_state) 1059{ 1060 unsigned long flags; 1061 int running, on_rq; 1062 unsigned long ncsw; 1063 struct rq *rq; 1064 1065 for (;;) { 1066 /* 1067 * We do the initial early heuristics without holding 1068 * any task-queue locks at all. We'll only try to get 1069 * the runqueue lock when things look like they will 1070 * work out! 1071 */ 1072 rq = task_rq(p); 1073 1074 /* 1075 * If the task is actively running on another CPU 1076 * still, just relax and busy-wait without holding 1077 * any locks. 1078 * 1079 * NOTE! Since we don't hold any locks, it's not 1080 * even sure that "rq" stays as the right runqueue! 1081 * But we don't care, since "task_running()" will 1082 * return false if the runqueue has changed and p 1083 * is actually now running somewhere else! 1084 */ 1085 while (task_running(rq, p)) { 1086 if (match_state && unlikely(p->state != match_state)) 1087 return 0; 1088 cpu_relax(); 1089 } 1090 1091 /* 1092 * Ok, time to look more closely! We need the rq 1093 * lock now, to be *sure*. If we're wrong, we'll 1094 * just go back and repeat. 1095 */ 1096 rq = task_rq_lock(p, &flags); 1097 trace_sched_wait_task(p); 1098 running = task_running(rq, p); 1099 on_rq = p->on_rq; 1100 ncsw = 0; 1101 if (!match_state || p->state == match_state) 1102 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */ 1103 task_rq_unlock(rq, p, &flags); 1104 1105 /* 1106 * If it changed from the expected state, bail out now. 1107 */ 1108 if (unlikely(!ncsw)) 1109 break; 1110 1111 /* 1112 * Was it really running after all now that we 1113 * checked with the proper locks actually held? 1114 * 1115 * Oops. Go back and try again.. 1116 */ 1117 if (unlikely(running)) { 1118 cpu_relax(); 1119 continue; 1120 } 1121 1122 /* 1123 * It's not enough that it's not actively running, 1124 * it must be off the runqueue _entirely_, and not 1125 * preempted! 1126 * 1127 * So if it was still runnable (but just not actively 1128 * running right now), it's preempted, and we should 1129 * yield - it could be a while. 1130 */ 1131 if (unlikely(on_rq)) { 1132 ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ); 1133 1134 set_current_state(TASK_UNINTERRUPTIBLE); 1135 schedule_hrtimeout(&to, HRTIMER_MODE_REL); 1136 continue; 1137 } 1138 1139 /* 1140 * Ahh, all good. It wasn't running, and it wasn't 1141 * runnable, which means that it will never become 1142 * running in the future either. We're all done! 1143 */ 1144 break; 1145 } 1146 1147 return ncsw; 1148} 1149 1150/*** 1151 * kick_process - kick a running thread to enter/exit the kernel 1152 * @p: the to-be-kicked thread 1153 * 1154 * Cause a process which is running on another CPU to enter 1155 * kernel-mode, without any delay. (to get signals handled.) 1156 * 1157 * NOTE: this function doesn't have to take the runqueue lock, 1158 * because all it wants to ensure is that the remote task enters 1159 * the kernel. If the IPI races and the task has been migrated 1160 * to another CPU then no harm is done and the purpose has been 1161 * achieved as well. 1162 */ 1163void kick_process(struct task_struct *p) 1164{ 1165 int cpu; 1166 1167 preempt_disable(); 1168 cpu = task_cpu(p); 1169 if ((cpu != smp_processor_id()) && task_curr(p)) 1170 smp_send_reschedule(cpu); 1171 preempt_enable(); 1172} 1173EXPORT_SYMBOL_GPL(kick_process); 1174#endif /* CONFIG_SMP */ 1175 1176#ifdef CONFIG_SMP 1177/* 1178 * ->cpus_allowed is protected by both rq->lock and p->pi_lock 1179 */ 1180static int select_fallback_rq(int cpu, struct task_struct *p) 1181{ 1182 int nid = cpu_to_node(cpu); 1183 const struct cpumask *nodemask = NULL; 1184 enum { cpuset, possible, fail } state = cpuset; 1185 int dest_cpu; 1186 1187 /* 1188 * If the node that the cpu is on has been offlined, cpu_to_node() 1189 * will return -1. There is no cpu on the node, and we should 1190 * select the cpu on the other node. 1191 */ 1192 if (nid != -1) { 1193 nodemask = cpumask_of_node(nid); 1194 1195 /* Look for allowed, online CPU in same node. */ 1196 for_each_cpu(dest_cpu, nodemask) { 1197 if (!cpu_online(dest_cpu)) 1198 continue; 1199 if (!cpu_active(dest_cpu)) 1200 continue; 1201 if (cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p))) 1202 return dest_cpu; 1203 } 1204 } 1205 1206 for (;;) { 1207 /* Any allowed, online CPU? */ 1208 for_each_cpu(dest_cpu, tsk_cpus_allowed(p)) { 1209 if (!cpu_online(dest_cpu)) 1210 continue; 1211 if (!cpu_active(dest_cpu)) 1212 continue; 1213 goto out; 1214 } 1215 1216 switch (state) { 1217 case cpuset: 1218 /* No more Mr. Nice Guy. */ 1219 cpuset_cpus_allowed_fallback(p); 1220 state = possible; 1221 break; 1222 1223 case possible: 1224 do_set_cpus_allowed(p, cpu_possible_mask); 1225 state = fail; 1226 break; 1227 1228 case fail: 1229 BUG(); 1230 break; 1231 } 1232 } 1233 1234out: 1235 if (state != cpuset) { 1236 /* 1237 * Don't tell them about moving exiting tasks or 1238 * kernel threads (both mm NULL), since they never 1239 * leave kernel. 1240 */ 1241 if (p->mm && printk_ratelimit()) { 1242 printk_sched("process %d (%s) no longer affine to cpu%d\n", 1243 task_pid_nr(p), p->comm, cpu); 1244 } 1245 } 1246 1247 return dest_cpu; 1248} 1249 1250/* 1251 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable. 1252 */ 1253static inline 1254int select_task_rq(struct task_struct *p, int sd_flags, int wake_flags) 1255{ 1256 int cpu = p->sched_class->select_task_rq(p, sd_flags, wake_flags); 1257 1258 /* 1259 * In order not to call set_task_cpu() on a blocking task we need 1260 * to rely on ttwu() to place the task on a valid ->cpus_allowed 1261 * cpu. 1262 * 1263 * Since this is common to all placement strategies, this lives here. 1264 * 1265 * [ this allows ->select_task() to simply return task_cpu(p) and 1266 * not worry about this generic constraint ] 1267 */ 1268 if (unlikely(!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)) || 1269 !cpu_online(cpu))) 1270 cpu = select_fallback_rq(task_cpu(p), p); 1271 1272 return cpu; 1273} 1274 1275static void update_avg(u64 *avg, u64 sample) 1276{ 1277 s64 diff = sample - *avg; 1278 *avg += diff >> 3; 1279} 1280#endif 1281 1282static void 1283ttwu_stat(struct task_struct *p, int cpu, int wake_flags) 1284{ 1285#ifdef CONFIG_SCHEDSTATS 1286 struct rq *rq = this_rq(); 1287 1288#ifdef CONFIG_SMP 1289 int this_cpu = smp_processor_id(); 1290 1291 if (cpu == this_cpu) { 1292 schedstat_inc(rq, ttwu_local); 1293 schedstat_inc(p, se.statistics.nr_wakeups_local); 1294 } else { 1295 struct sched_domain *sd; 1296 1297 schedstat_inc(p, se.statistics.nr_wakeups_remote); 1298 rcu_read_lock(); 1299 for_each_domain(this_cpu, sd) { 1300 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) { 1301 schedstat_inc(sd, ttwu_wake_remote); 1302 break; 1303 } 1304 } 1305 rcu_read_unlock(); 1306 } 1307 1308 if (wake_flags & WF_MIGRATED) 1309 schedstat_inc(p, se.statistics.nr_wakeups_migrate); 1310 1311#endif /* CONFIG_SMP */ 1312 1313 schedstat_inc(rq, ttwu_count); 1314 schedstat_inc(p, se.statistics.nr_wakeups); 1315 1316 if (wake_flags & WF_SYNC) 1317 schedstat_inc(p, se.statistics.nr_wakeups_sync); 1318 1319#endif /* CONFIG_SCHEDSTATS */ 1320} 1321 1322static void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags) 1323{ 1324 activate_task(rq, p, en_flags); 1325 p->on_rq = 1; 1326 1327 /* if a worker is waking up, notify workqueue */ 1328 if (p->flags & PF_WQ_WORKER) 1329 wq_worker_waking_up(p, cpu_of(rq)); 1330} 1331 1332/* 1333 * Mark the task runnable and perform wakeup-preemption. 1334 */ 1335static void 1336ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags) 1337{ 1338 check_preempt_curr(rq, p, wake_flags); 1339 trace_sched_wakeup(p, true); 1340 1341 p->state = TASK_RUNNING; 1342#ifdef CONFIG_SMP 1343 if (p->sched_class->task_woken) 1344 p->sched_class->task_woken(rq, p); 1345 1346 if (rq->idle_stamp) { 1347 u64 delta = rq_clock(rq) - rq->idle_stamp; 1348 u64 max = 2*sysctl_sched_migration_cost; 1349 1350 if (delta > max) 1351 rq->avg_idle = max; 1352 else 1353 update_avg(&rq->avg_idle, delta); 1354 rq->idle_stamp = 0; 1355 } 1356#endif 1357} 1358 1359static void 1360ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags) 1361{ 1362#ifdef CONFIG_SMP 1363 if (p->sched_contributes_to_load) 1364 rq->nr_uninterruptible--; 1365#endif 1366 1367 ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_WAKING); 1368 ttwu_do_wakeup(rq, p, wake_flags); 1369} 1370 1371/* 1372 * Called in case the task @p isn't fully descheduled from its runqueue, 1373 * in this case we must do a remote wakeup. Its a 'light' wakeup though, 1374 * since all we need to do is flip p->state to TASK_RUNNING, since 1375 * the task is still ->on_rq. 1376 */ 1377static int ttwu_remote(struct task_struct *p, int wake_flags) 1378{ 1379 struct rq *rq; 1380 int ret = 0; 1381 1382 rq = __task_rq_lock(p); 1383 if (p->on_rq) { 1384 /* check_preempt_curr() may use rq clock */ 1385 update_rq_clock(rq); 1386 ttwu_do_wakeup(rq, p, wake_flags); 1387 ret = 1; 1388 } 1389 __task_rq_unlock(rq); 1390 1391 return ret; 1392} 1393 1394#ifdef CONFIG_SMP 1395static void sched_ttwu_pending(void) 1396{ 1397 struct rq *rq = this_rq(); 1398 struct llist_node *llist = llist_del_all(&rq->wake_list); 1399 struct task_struct *p; 1400 1401 raw_spin_lock(&rq->lock); 1402 1403 while (llist) { 1404 p = llist_entry(llist, struct task_struct, wake_entry); 1405 llist = llist_next(llist); 1406 ttwu_do_activate(rq, p, 0); 1407 } 1408 1409 raw_spin_unlock(&rq->lock); 1410} 1411 1412void scheduler_ipi(void) 1413{ 1414 if (llist_empty(&this_rq()->wake_list) 1415 && !tick_nohz_full_cpu(smp_processor_id()) 1416 && !got_nohz_idle_kick()) 1417 return; 1418 1419 /* 1420 * Not all reschedule IPI handlers call irq_enter/irq_exit, since 1421 * traditionally all their work was done from the interrupt return 1422 * path. Now that we actually do some work, we need to make sure 1423 * we do call them. 1424 * 1425 * Some archs already do call them, luckily irq_enter/exit nest 1426 * properly. 1427 * 1428 * Arguably we should visit all archs and update all handlers, 1429 * however a fair share of IPIs are still resched only so this would 1430 * somewhat pessimize the simple resched case. 1431 */ 1432 irq_enter(); 1433 tick_nohz_full_check(); 1434 sched_ttwu_pending(); 1435 1436 /* 1437 * Check if someone kicked us for doing the nohz idle load balance. 1438 */ 1439 if (unlikely(got_nohz_idle_kick())) { 1440 this_rq()->idle_balance = 1; 1441 raise_softirq_irqoff(SCHED_SOFTIRQ); 1442 } 1443 irq_exit(); 1444} 1445 1446static void ttwu_queue_remote(struct task_struct *p, int cpu) 1447{ 1448 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list)) 1449 smp_send_reschedule(cpu); 1450} 1451 1452bool cpus_share_cache(int this_cpu, int that_cpu) 1453{ 1454 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu); 1455} 1456#endif /* CONFIG_SMP */ 1457 1458static void ttwu_queue(struct task_struct *p, int cpu) 1459{ 1460 struct rq *rq = cpu_rq(cpu); 1461 1462#if defined(CONFIG_SMP) 1463 if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) { 1464 sched_clock_cpu(cpu); /* sync clocks x-cpu */ 1465 ttwu_queue_remote(p, cpu); 1466 return; 1467 } 1468#endif 1469 1470 raw_spin_lock(&rq->lock); 1471 ttwu_do_activate(rq, p, 0); 1472 raw_spin_unlock(&rq->lock); 1473} 1474 1475/** 1476 * try_to_wake_up - wake up a thread 1477 * @p: the thread to be awakened 1478 * @state: the mask of task states that can be woken 1479 * @wake_flags: wake modifier flags (WF_*) 1480 * 1481 * Put it on the run-queue if it's not already there. The "current" 1482 * thread is always on the run-queue (except when the actual 1483 * re-schedule is in progress), and as such you're allowed to do 1484 * the simpler "current->state = TASK_RUNNING" to mark yourself 1485 * runnable without the overhead of this. 1486 * 1487 * Return: %true if @p was woken up, %false if it was already running. 1488 * or @state didn't match @p's state. 1489 */ 1490static int 1491try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags) 1492{ 1493 unsigned long flags; 1494 int cpu, success = 0; 1495 1496 smp_wmb(); 1497 raw_spin_lock_irqsave(&p->pi_lock, flags); 1498 if (!(p->state & state)) 1499 goto out; 1500 1501 success = 1; /* we're going to change ->state */ 1502 cpu = task_cpu(p); 1503 1504 if (p->on_rq && ttwu_remote(p, wake_flags)) 1505 goto stat; 1506 1507#ifdef CONFIG_SMP 1508 /* 1509 * If the owning (remote) cpu is still in the middle of schedule() with 1510 * this task as prev, wait until its done referencing the task. 1511 */ 1512 while (p->on_cpu) 1513 cpu_relax(); 1514 /* 1515 * Pairs with the smp_wmb() in finish_lock_switch(). 1516 */ 1517 smp_rmb(); 1518 1519 p->sched_contributes_to_load = !!task_contributes_to_load(p); 1520 p->state = TASK_WAKING; 1521 1522 if (p->sched_class->task_waking) 1523 p->sched_class->task_waking(p); 1524 1525 cpu = select_task_rq(p, SD_BALANCE_WAKE, wake_flags); 1526 if (task_cpu(p) != cpu) { 1527 wake_flags |= WF_MIGRATED; 1528 set_task_cpu(p, cpu); 1529 } 1530#endif /* CONFIG_SMP */ 1531 1532 ttwu_queue(p, cpu); 1533stat: 1534 ttwu_stat(p, cpu, wake_flags); 1535out: 1536 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 1537 1538 return success; 1539} 1540 1541/** 1542 * try_to_wake_up_local - try to wake up a local task with rq lock held 1543 * @p: the thread to be awakened 1544 * 1545 * Put @p on the run-queue if it's not already there. The caller must 1546 * ensure that this_rq() is locked, @p is bound to this_rq() and not 1547 * the current task. 1548 */ 1549static void try_to_wake_up_local(struct task_struct *p) 1550{ 1551 struct rq *rq = task_rq(p); 1552 1553 if (WARN_ON_ONCE(rq != this_rq()) || 1554 WARN_ON_ONCE(p == current)) 1555 return; 1556 1557 lockdep_assert_held(&rq->lock); 1558 1559 if (!raw_spin_trylock(&p->pi_lock)) { 1560 raw_spin_unlock(&rq->lock); 1561 raw_spin_lock(&p->pi_lock); 1562 raw_spin_lock(&rq->lock); 1563 } 1564 1565 if (!(p->state & TASK_NORMAL)) 1566 goto out; 1567 1568 if (!p->on_rq) 1569 ttwu_activate(rq, p, ENQUEUE_WAKEUP); 1570 1571 ttwu_do_wakeup(rq, p, 0); 1572 ttwu_stat(p, smp_processor_id(), 0); 1573out: 1574 raw_spin_unlock(&p->pi_lock); 1575} 1576 1577/** 1578 * wake_up_process - Wake up a specific process 1579 * @p: The process to be woken up. 1580 * 1581 * Attempt to wake up the nominated process and move it to the set of runnable 1582 * processes. 1583 * 1584 * Return: 1 if the process was woken up, 0 if it was already running. 1585 * 1586 * It may be assumed that this function implies a write memory barrier before 1587 * changing the task state if and only if any tasks are woken up. 1588 */ 1589int wake_up_process(struct task_struct *p) 1590{ 1591 WARN_ON(task_is_stopped_or_traced(p)); 1592 return try_to_wake_up(p, TASK_NORMAL, 0); 1593} 1594EXPORT_SYMBOL(wake_up_process); 1595 1596int wake_up_state(struct task_struct *p, unsigned int state) 1597{ 1598 return try_to_wake_up(p, state, 0); 1599} 1600 1601/* 1602 * Perform scheduler related setup for a newly forked process p. 1603 * p is forked by current. 1604 * 1605 * __sched_fork() is basic setup used by init_idle() too: 1606 */ 1607static void __sched_fork(struct task_struct *p) 1608{ 1609 p->on_rq = 0; 1610 1611 p->se.on_rq = 0; 1612 p->se.exec_start = 0; 1613 p->se.sum_exec_runtime = 0; 1614 p->se.prev_sum_exec_runtime = 0; 1615 p->se.nr_migrations = 0; 1616 p->se.vruntime = 0; 1617 INIT_LIST_HEAD(&p->se.group_node); 1618 1619#ifdef CONFIG_SCHEDSTATS 1620 memset(&p->se.statistics, 0, sizeof(p->se.statistics)); 1621#endif 1622 1623 INIT_LIST_HEAD(&p->rt.run_list); 1624 1625#ifdef CONFIG_PREEMPT_NOTIFIERS 1626 INIT_HLIST_HEAD(&p->preempt_notifiers); 1627#endif 1628 1629#ifdef CONFIG_NUMA_BALANCING 1630 if (p->mm && atomic_read(&p->mm->mm_users) == 1) { 1631 p->mm->numa_next_scan = jiffies; 1632 p->mm->numa_next_reset = jiffies; 1633 p->mm->numa_scan_seq = 0; 1634 } 1635 1636 p->node_stamp = 0ULL; 1637 p->numa_scan_seq = p->mm ? p->mm->numa_scan_seq : 0; 1638 p->numa_migrate_seq = p->mm ? p->mm->numa_scan_seq - 1 : 0; 1639 p->numa_scan_period = sysctl_numa_balancing_scan_delay; 1640 p->numa_work.next = &p->numa_work; 1641#endif /* CONFIG_NUMA_BALANCING */ 1642} 1643 1644#ifdef CONFIG_NUMA_BALANCING 1645#ifdef CONFIG_SCHED_DEBUG 1646void set_numabalancing_state(bool enabled) 1647{ 1648 if (enabled) 1649 sched_feat_set("NUMA"); 1650 else 1651 sched_feat_set("NO_NUMA"); 1652} 1653#else 1654__read_mostly bool numabalancing_enabled; 1655 1656void set_numabalancing_state(bool enabled) 1657{ 1658 numabalancing_enabled = enabled; 1659} 1660#endif /* CONFIG_SCHED_DEBUG */ 1661#endif /* CONFIG_NUMA_BALANCING */ 1662 1663/* 1664 * fork()/clone()-time setup: 1665 */ 1666void sched_fork(struct task_struct *p) 1667{ 1668 unsigned long flags; 1669 int cpu = get_cpu(); 1670 1671 __sched_fork(p); 1672 /* 1673 * We mark the process as running here. This guarantees that 1674 * nobody will actually run it, and a signal or other external 1675 * event cannot wake it up and insert it on the runqueue either. 1676 */ 1677 p->state = TASK_RUNNING; 1678 1679 /* 1680 * Make sure we do not leak PI boosting priority to the child. 1681 */ 1682 p->prio = current->normal_prio; 1683 1684 /* 1685 * Revert to default priority/policy on fork if requested. 1686 */ 1687 if (unlikely(p->sched_reset_on_fork)) { 1688 if (task_has_rt_policy(p)) { 1689 p->policy = SCHED_NORMAL; 1690 p->static_prio = NICE_TO_PRIO(0); 1691 p->rt_priority = 0; 1692 } else if (PRIO_TO_NICE(p->static_prio) < 0) 1693 p->static_prio = NICE_TO_PRIO(0); 1694 1695 p->prio = p->normal_prio = __normal_prio(p); 1696 set_load_weight(p); 1697 1698 /* 1699 * We don't need the reset flag anymore after the fork. It has 1700 * fulfilled its duty: 1701 */ 1702 p->sched_reset_on_fork = 0; 1703 } 1704 1705 if (!rt_prio(p->prio)) 1706 p->sched_class = &fair_sched_class; 1707 1708 if (p->sched_class->task_fork) 1709 p->sched_class->task_fork(p); 1710 1711 /* 1712 * The child is not yet in the pid-hash so no cgroup attach races, 1713 * and the cgroup is pinned to this child due to cgroup_fork() 1714 * is ran before sched_fork(). 1715 * 1716 * Silence PROVE_RCU. 1717 */ 1718 raw_spin_lock_irqsave(&p->pi_lock, flags); 1719 set_task_cpu(p, cpu); 1720 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 1721 1722#if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT) 1723 if (likely(sched_info_on())) 1724 memset(&p->sched_info, 0, sizeof(p->sched_info)); 1725#endif 1726#if defined(CONFIG_SMP) 1727 p->on_cpu = 0; 1728#endif 1729#ifdef CONFIG_PREEMPT_COUNT 1730 /* Want to start with kernel preemption disabled. */ 1731 task_thread_info(p)->preempt_count = 1; 1732#endif 1733#ifdef CONFIG_SMP 1734 plist_node_init(&p->pushable_tasks, MAX_PRIO); 1735#endif 1736 1737 put_cpu(); 1738} 1739 1740/* 1741 * wake_up_new_task - wake up a newly created task for the first time. 1742 * 1743 * This function will do some initial scheduler statistics housekeeping 1744 * that must be done for every newly created context, then puts the task 1745 * on the runqueue and wakes it. 1746 */ 1747void wake_up_new_task(struct task_struct *p) 1748{ 1749 unsigned long flags; 1750 struct rq *rq; 1751 1752 raw_spin_lock_irqsave(&p->pi_lock, flags); 1753#ifdef CONFIG_SMP 1754 /* 1755 * Fork balancing, do it here and not earlier because: 1756 * - cpus_allowed can change in the fork path 1757 * - any previously selected cpu might disappear through hotplug 1758 */ 1759 set_task_cpu(p, select_task_rq(p, SD_BALANCE_FORK, 0)); 1760#endif 1761 1762 /* Initialize new task's runnable average */ 1763 init_task_runnable_average(p); 1764 rq = __task_rq_lock(p); 1765 activate_task(rq, p, 0); 1766 p->on_rq = 1; 1767 trace_sched_wakeup_new(p, true); 1768 check_preempt_curr(rq, p, WF_FORK); 1769#ifdef CONFIG_SMP 1770 if (p->sched_class->task_woken) 1771 p->sched_class->task_woken(rq, p); 1772#endif 1773 task_rq_unlock(rq, p, &flags); 1774} 1775 1776#ifdef CONFIG_PREEMPT_NOTIFIERS 1777 1778/** 1779 * preempt_notifier_register - tell me when current is being preempted & rescheduled 1780 * @notifier: notifier struct to register 1781 */ 1782void preempt_notifier_register(struct preempt_notifier *notifier) 1783{ 1784 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers); 1785} 1786EXPORT_SYMBOL_GPL(preempt_notifier_register); 1787 1788/** 1789 * preempt_notifier_unregister - no longer interested in preemption notifications 1790 * @notifier: notifier struct to unregister 1791 * 1792 * This is safe to call from within a preemption notifier. 1793 */ 1794void preempt_notifier_unregister(struct preempt_notifier *notifier) 1795{ 1796 hlist_del(¬ifier->link); 1797} 1798EXPORT_SYMBOL_GPL(preempt_notifier_unregister); 1799 1800static void fire_sched_in_preempt_notifiers(struct task_struct *curr) 1801{ 1802 struct preempt_notifier *notifier; 1803 1804 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link) 1805 notifier->ops->sched_in(notifier, raw_smp_processor_id()); 1806} 1807 1808static void 1809fire_sched_out_preempt_notifiers(struct task_struct *curr, 1810 struct task_struct *next) 1811{ 1812 struct preempt_notifier *notifier; 1813 1814 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link) 1815 notifier->ops->sched_out(notifier, next); 1816} 1817 1818#else /* !CONFIG_PREEMPT_NOTIFIERS */ 1819 1820static void fire_sched_in_preempt_notifiers(struct task_struct *curr) 1821{ 1822} 1823 1824static void 1825fire_sched_out_preempt_notifiers(struct task_struct *curr, 1826 struct task_struct *next) 1827{ 1828} 1829 1830#endif /* CONFIG_PREEMPT_NOTIFIERS */ 1831 1832/** 1833 * prepare_task_switch - prepare to switch tasks 1834 * @rq: the runqueue preparing to switch 1835 * @prev: the current task that is being switched out 1836 * @next: the task we are going to switch to. 1837 * 1838 * This is called with the rq lock held and interrupts off. It must 1839 * be paired with a subsequent finish_task_switch after the context 1840 * switch. 1841 * 1842 * prepare_task_switch sets up locking and calls architecture specific 1843 * hooks. 1844 */ 1845static inline void 1846prepare_task_switch(struct rq *rq, struct task_struct *prev, 1847 struct task_struct *next) 1848{ 1849 trace_sched_switch(prev, next); 1850 sched_info_switch(prev, next); 1851 perf_event_task_sched_out(prev, next); 1852 fire_sched_out_preempt_notifiers(prev, next); 1853 prepare_lock_switch(rq, next); 1854 prepare_arch_switch(next); 1855} 1856 1857/** 1858 * finish_task_switch - clean up after a task-switch 1859 * @rq: runqueue associated with task-switch 1860 * @prev: the thread we just switched away from. 1861 * 1862 * finish_task_switch must be called after the context switch, paired 1863 * with a prepare_task_switch call before the context switch. 1864 * finish_task_switch will reconcile locking set up by prepare_task_switch, 1865 * and do any other architecture-specific cleanup actions. 1866 * 1867 * Note that we may have delayed dropping an mm in context_switch(). If 1868 * so, we finish that here outside of the runqueue lock. (Doing it 1869 * with the lock held can cause deadlocks; see schedule() for 1870 * details.) 1871 */ 1872static void finish_task_switch(struct rq *rq, struct task_struct *prev) 1873 __releases(rq->lock) 1874{ 1875 struct mm_struct *mm = rq->prev_mm; 1876 long prev_state; 1877 1878 rq->prev_mm = NULL; 1879 1880 /* 1881 * A task struct has one reference for the use as "current". 1882 * If a task dies, then it sets TASK_DEAD in tsk->state and calls 1883 * schedule one last time. The schedule call will never return, and 1884 * the scheduled task must drop that reference. 1885 * The test for TASK_DEAD must occur while the runqueue locks are 1886 * still held, otherwise prev could be scheduled on another cpu, die 1887 * there before we look at prev->state, and then the reference would 1888 * be dropped twice. 1889 * Manfred Spraul <manfred@colorfullife.com> 1890 */ 1891 prev_state = prev->state; 1892 vtime_task_switch(prev); 1893 finish_arch_switch(prev); 1894 perf_event_task_sched_in(prev, current); 1895 finish_lock_switch(rq, prev); 1896 finish_arch_post_lock_switch(); 1897 1898 fire_sched_in_preempt_notifiers(current); 1899 if (mm) 1900 mmdrop(mm); 1901 if (unlikely(prev_state == TASK_DEAD)) { 1902 /* 1903 * Remove function-return probe instances associated with this 1904 * task and put them back on the free list. 1905 */ 1906 kprobe_flush_task(prev); 1907 put_task_struct(prev); 1908 } 1909 1910 tick_nohz_task_switch(current); 1911} 1912 1913#ifdef CONFIG_SMP 1914 1915/* assumes rq->lock is held */ 1916static inline void pre_schedule(struct rq *rq, struct task_struct *prev) 1917{ 1918 if (prev->sched_class->pre_schedule) 1919 prev->sched_class->pre_schedule(rq, prev); 1920} 1921 1922/* rq->lock is NOT held, but preemption is disabled */ 1923static inline void post_schedule(struct rq *rq) 1924{ 1925 if (rq->post_schedule) { 1926 unsigned long flags; 1927 1928 raw_spin_lock_irqsave(&rq->lock, flags); 1929 if (rq->curr->sched_class->post_schedule) 1930 rq->curr->sched_class->post_schedule(rq); 1931 raw_spin_unlock_irqrestore(&rq->lock, flags); 1932 1933 rq->post_schedule = 0; 1934 } 1935} 1936 1937#else 1938 1939static inline void pre_schedule(struct rq *rq, struct task_struct *p) 1940{ 1941} 1942 1943static inline void post_schedule(struct rq *rq) 1944{ 1945} 1946 1947#endif 1948 1949/** 1950 * schedule_tail - first thing a freshly forked thread must call. 1951 * @prev: the thread we just switched away from. 1952 */ 1953asmlinkage void schedule_tail(struct task_struct *prev) 1954 __releases(rq->lock) 1955{ 1956 struct rq *rq = this_rq(); 1957 1958 finish_task_switch(rq, prev); 1959 1960 /* 1961 * FIXME: do we need to worry about rq being invalidated by the 1962 * task_switch? 1963 */ 1964 post_schedule(rq); 1965 1966#ifdef __ARCH_WANT_UNLOCKED_CTXSW 1967 /* In this case, finish_task_switch does not reenable preemption */ 1968 preempt_enable(); 1969#endif 1970 if (current->set_child_tid) 1971 put_user(task_pid_vnr(current), current->set_child_tid); 1972} 1973 1974/* 1975 * context_switch - switch to the new MM and the new 1976 * thread's register state. 1977 */ 1978static inline void 1979context_switch(struct rq *rq, struct task_struct *prev, 1980 struct task_struct *next) 1981{ 1982 struct mm_struct *mm, *oldmm; 1983 1984 prepare_task_switch(rq, prev, next); 1985 1986 mm = next->mm; 1987 oldmm = prev->active_mm; 1988 /* 1989 * For paravirt, this is coupled with an exit in switch_to to 1990 * combine the page table reload and the switch backend into 1991 * one hypercall. 1992 */ 1993 arch_start_context_switch(prev); 1994 1995 if (!mm) { 1996 next->active_mm = oldmm; 1997 atomic_inc(&oldmm->mm_count); 1998 enter_lazy_tlb(oldmm, next); 1999 } else 2000 switch_mm(oldmm, mm, next); 2001 2002 if (!prev->mm) { 2003 prev->active_mm = NULL; 2004 rq->prev_mm = oldmm; 2005 } 2006 /* 2007 * Since the runqueue lock will be released by the next 2008 * task (which is an invalid locking op but in the case 2009 * of the scheduler it's an obvious special-case), so we 2010 * do an early lockdep release here: 2011 */ 2012#ifndef __ARCH_WANT_UNLOCKED_CTXSW 2013 spin_release(&rq->lock.dep_map, 1, _THIS_IP_); 2014#endif 2015 2016 context_tracking_task_switch(prev, next); 2017 /* Here we just switch the register state and the stack. */ 2018 switch_to(prev, next, prev); 2019 2020 barrier(); 2021 /* 2022 * this_rq must be evaluated again because prev may have moved 2023 * CPUs since it called schedule(), thus the 'rq' on its stack 2024 * frame will be invalid. 2025 */ 2026 finish_task_switch(this_rq(), prev); 2027} 2028 2029/* 2030 * nr_running and nr_context_switches: 2031 * 2032 * externally visible scheduler statistics: current number of runnable 2033 * threads, total number of context switches performed since bootup. 2034 */ 2035unsigned long nr_running(void) 2036{ 2037 unsigned long i, sum = 0; 2038 2039 for_each_online_cpu(i) 2040 sum += cpu_rq(i)->nr_running; 2041 2042 return sum; 2043} 2044 2045unsigned long long nr_context_switches(void) 2046{ 2047 int i; 2048 unsigned long long sum = 0; 2049 2050 for_each_possible_cpu(i) 2051 sum += cpu_rq(i)->nr_switches; 2052 2053 return sum; 2054} 2055 2056unsigned long nr_iowait(void) 2057{ 2058 unsigned long i, sum = 0; 2059 2060 for_each_possible_cpu(i) 2061 sum += atomic_read(&cpu_rq(i)->nr_iowait); 2062 2063 return sum; 2064} 2065 2066unsigned long nr_iowait_cpu(int cpu) 2067{ 2068 struct rq *this = cpu_rq(cpu); 2069 return atomic_read(&this->nr_iowait); 2070} 2071 2072#ifdef CONFIG_SMP 2073 2074/* 2075 * sched_exec - execve() is a valuable balancing opportunity, because at 2076 * this point the task has the smallest effective memory and cache footprint. 2077 */ 2078void sched_exec(void) 2079{ 2080 struct task_struct *p = current; 2081 unsigned long flags; 2082 int dest_cpu; 2083 2084 raw_spin_lock_irqsave(&p->pi_lock, flags); 2085 dest_cpu = p->sched_class->select_task_rq(p, SD_BALANCE_EXEC, 0); 2086 if (dest_cpu == smp_processor_id()) 2087 goto unlock; 2088 2089 if (likely(cpu_active(dest_cpu))) { 2090 struct migration_arg arg = { p, dest_cpu }; 2091 2092 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 2093 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg); 2094 return; 2095 } 2096unlock: 2097 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 2098} 2099 2100#endif 2101 2102DEFINE_PER_CPU(struct kernel_stat, kstat); 2103DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat); 2104 2105EXPORT_PER_CPU_SYMBOL(kstat); 2106EXPORT_PER_CPU_SYMBOL(kernel_cpustat); 2107 2108/* 2109 * Return any ns on the sched_clock that have not yet been accounted in 2110 * @p in case that task is currently running. 2111 * 2112 * Called with task_rq_lock() held on @rq. 2113 */ 2114static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq) 2115{ 2116 u64 ns = 0; 2117 2118 if (task_current(rq, p)) { 2119 update_rq_clock(rq); 2120 ns = rq_clock_task(rq) - p->se.exec_start; 2121 if ((s64)ns < 0) 2122 ns = 0; 2123 } 2124 2125 return ns; 2126} 2127 2128unsigned long long task_delta_exec(struct task_struct *p) 2129{ 2130 unsigned long flags; 2131 struct rq *rq; 2132 u64 ns = 0; 2133 2134 rq = task_rq_lock(p, &flags); 2135 ns = do_task_delta_exec(p, rq); 2136 task_rq_unlock(rq, p, &flags); 2137 2138 return ns; 2139} 2140 2141/* 2142 * Return accounted runtime for the task. 2143 * In case the task is currently running, return the runtime plus current's 2144 * pending runtime that have not been accounted yet. 2145 */ 2146unsigned long long task_sched_runtime(struct task_struct *p) 2147{ 2148 unsigned long flags; 2149 struct rq *rq; 2150 u64 ns = 0; 2151 2152 rq = task_rq_lock(p, &flags); 2153 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq); 2154 task_rq_unlock(rq, p, &flags); 2155 2156 return ns; 2157} 2158 2159/* 2160 * This function gets called by the timer code, with HZ frequency. 2161 * We call it with interrupts disabled. 2162 */ 2163void scheduler_tick(void) 2164{ 2165 int cpu = smp_processor_id(); 2166 struct rq *rq = cpu_rq(cpu); 2167 struct task_struct *curr = rq->curr; 2168 2169 sched_clock_tick(); 2170 2171 raw_spin_lock(&rq->lock); 2172 update_rq_clock(rq); 2173 curr->sched_class->task_tick(rq, curr, 0); 2174 update_cpu_load_active(rq); 2175 raw_spin_unlock(&rq->lock); 2176 2177 perf_event_task_tick(); 2178 2179#ifdef CONFIG_SMP 2180 rq->idle_balance = idle_cpu(cpu); 2181 trigger_load_balance(rq, cpu); 2182#endif 2183 rq_last_tick_reset(rq); 2184} 2185 2186#ifdef CONFIG_NO_HZ_FULL 2187/** 2188 * scheduler_tick_max_deferment 2189 * 2190 * Keep at least one tick per second when a single 2191 * active task is running because the scheduler doesn't 2192 * yet completely support full dynticks environment. 2193 * 2194 * This makes sure that uptime, CFS vruntime, load 2195 * balancing, etc... continue to move forward, even 2196 * with a very low granularity. 2197 * 2198 * Return: Maximum deferment in nanoseconds. 2199 */ 2200u64 scheduler_tick_max_deferment(void) 2201{ 2202 struct rq *rq = this_rq(); 2203 unsigned long next, now = ACCESS_ONCE(jiffies); 2204 2205 next = rq->last_sched_tick + HZ; 2206 2207 if (time_before_eq(next, now)) 2208 return 0; 2209 2210 return jiffies_to_usecs(next - now) * NSEC_PER_USEC; 2211} 2212#endif 2213 2214notrace unsigned long get_parent_ip(unsigned long addr) 2215{ 2216 if (in_lock_functions(addr)) { 2217 addr = CALLER_ADDR2; 2218 if (in_lock_functions(addr)) 2219 addr = CALLER_ADDR3; 2220 } 2221 return addr; 2222} 2223 2224#if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \ 2225 defined(CONFIG_PREEMPT_TRACER)) 2226 2227void __kprobes add_preempt_count(int val) 2228{ 2229#ifdef CONFIG_DEBUG_PREEMPT 2230 /* 2231 * Underflow? 2232 */ 2233 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0))) 2234 return; 2235#endif 2236 preempt_count() += val; 2237#ifdef CONFIG_DEBUG_PREEMPT 2238 /* 2239 * Spinlock count overflowing soon? 2240 */ 2241 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >= 2242 PREEMPT_MASK - 10); 2243#endif 2244 if (preempt_count() == val) 2245 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1)); 2246} 2247EXPORT_SYMBOL(add_preempt_count); 2248 2249void __kprobes sub_preempt_count(int val) 2250{ 2251#ifdef CONFIG_DEBUG_PREEMPT 2252 /* 2253 * Underflow? 2254 */ 2255 if (DEBUG_LOCKS_WARN_ON(val > preempt_count())) 2256 return; 2257 /* 2258 * Is the spinlock portion underflowing? 2259 */ 2260 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) && 2261 !(preempt_count() & PREEMPT_MASK))) 2262 return; 2263#endif 2264 2265 if (preempt_count() == val) 2266 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1)); 2267 preempt_count() -= val; 2268} 2269EXPORT_SYMBOL(sub_preempt_count); 2270 2271#endif 2272 2273/* 2274 * Print scheduling while atomic bug: 2275 */ 2276static noinline void __schedule_bug(struct task_struct *prev) 2277{ 2278 if (oops_in_progress) 2279 return; 2280 2281 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n", 2282 prev->comm, prev->pid, preempt_count()); 2283 2284 debug_show_held_locks(prev); 2285 print_modules(); 2286 if (irqs_disabled()) 2287 print_irqtrace_events(prev); 2288 dump_stack(); 2289 add_taint(TAINT_WARN, LOCKDEP_STILL_OK); 2290} 2291 2292/* 2293 * Various schedule()-time debugging checks and statistics: 2294 */ 2295static inline void schedule_debug(struct task_struct *prev) 2296{ 2297 /* 2298 * Test if we are atomic. Since do_exit() needs to call into 2299 * schedule() atomically, we ignore that path for now. 2300 * Otherwise, whine if we are scheduling when we should not be. 2301 */ 2302 if (unlikely(in_atomic_preempt_off() && !prev->exit_state)) 2303 __schedule_bug(prev); 2304 rcu_sleep_check(); 2305 2306 profile_hit(SCHED_PROFILING, __builtin_return_address(0)); 2307 2308 schedstat_inc(this_rq(), sched_count); 2309} 2310 2311static void put_prev_task(struct rq *rq, struct task_struct *prev) 2312{ 2313 if (prev->on_rq || rq->skip_clock_update < 0) 2314 update_rq_clock(rq); 2315 prev->sched_class->put_prev_task(rq, prev); 2316} 2317 2318/* 2319 * Pick up the highest-prio task: 2320 */ 2321static inline struct task_struct * 2322pick_next_task(struct rq *rq) 2323{ 2324 const struct sched_class *class; 2325 struct task_struct *p; 2326 2327 /* 2328 * Optimization: we know that if all tasks are in 2329 * the fair class we can call that function directly: 2330 */ 2331 if (likely(rq->nr_running == rq->cfs.h_nr_running)) { 2332 p = fair_sched_class.pick_next_task(rq); 2333 if (likely(p)) 2334 return p; 2335 } 2336 2337 for_each_class(class) { 2338 p = class->pick_next_task(rq); 2339 if (p) 2340 return p; 2341 } 2342 2343 BUG(); /* the idle class will always have a runnable task */ 2344} 2345 2346/* 2347 * __schedule() is the main scheduler function. 2348 * 2349 * The main means of driving the scheduler and thus entering this function are: 2350 * 2351 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc. 2352 * 2353 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return 2354 * paths. For example, see arch/x86/entry_64.S. 2355 * 2356 * To drive preemption between tasks, the scheduler sets the flag in timer 2357 * interrupt handler scheduler_tick(). 2358 * 2359 * 3. Wakeups don't really cause entry into schedule(). They add a 2360 * task to the run-queue and that's it. 2361 * 2362 * Now, if the new task added to the run-queue preempts the current 2363 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets 2364 * called on the nearest possible occasion: 2365 * 2366 * - If the kernel is preemptible (CONFIG_PREEMPT=y): 2367 * 2368 * - in syscall or exception context, at the next outmost 2369 * preempt_enable(). (this might be as soon as the wake_up()'s 2370 * spin_unlock()!) 2371 * 2372 * - in IRQ context, return from interrupt-handler to 2373 * preemptible context 2374 * 2375 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set) 2376 * then at the next: 2377 * 2378 * - cond_resched() call 2379 * - explicit schedule() call 2380 * - return from syscall or exception to user-space 2381 * - return from interrupt-handler to user-space 2382 */ 2383static void __sched __schedule(void) 2384{ 2385 struct task_struct *prev, *next; 2386 unsigned long *switch_count; 2387 struct rq *rq; 2388 int cpu; 2389 2390need_resched: 2391 preempt_disable(); 2392 cpu = smp_processor_id(); 2393 rq = cpu_rq(cpu); 2394 rcu_note_context_switch(cpu); 2395 prev = rq->curr; 2396 2397 schedule_debug(prev); 2398 2399 if (sched_feat(HRTICK)) 2400 hrtick_clear(rq); 2401 2402 raw_spin_lock_irq(&rq->lock); 2403 2404 switch_count = &prev->nivcsw; 2405 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) { 2406 if (unlikely(signal_pending_state(prev->state, prev))) { 2407 prev->state = TASK_RUNNING; 2408 } else { 2409 deactivate_task(rq, prev, DEQUEUE_SLEEP); 2410 prev->on_rq = 0; 2411 2412 /* 2413 * If a worker went to sleep, notify and ask workqueue 2414 * whether it wants to wake up a task to maintain 2415 * concurrency. 2416 */ 2417 if (prev->flags & PF_WQ_WORKER) { 2418 struct task_struct *to_wakeup; 2419 2420 to_wakeup = wq_worker_sleeping(prev, cpu); 2421 if (to_wakeup) 2422 try_to_wake_up_local(to_wakeup); 2423 } 2424 } 2425 switch_count = &prev->nvcsw; 2426 } 2427 2428 pre_schedule(rq, prev); 2429 2430 if (unlikely(!rq->nr_running)) 2431 idle_balance(cpu, rq); 2432 2433 put_prev_task(rq, prev); 2434 next = pick_next_task(rq); 2435 clear_tsk_need_resched(prev); 2436 rq->skip_clock_update = 0; 2437 2438 if (likely(prev != next)) { 2439 rq->nr_switches++; 2440 rq->curr = next; 2441 ++*switch_count; 2442 2443 context_switch(rq, prev, next); /* unlocks the rq */ 2444 /* 2445 * The context switch have flipped the stack from under us 2446 * and restored the local variables which were saved when 2447 * this task called schedule() in the past. prev == current 2448 * is still correct, but it can be moved to another cpu/rq. 2449 */ 2450 cpu = smp_processor_id(); 2451 rq = cpu_rq(cpu); 2452 } else 2453 raw_spin_unlock_irq(&rq->lock); 2454 2455 post_schedule(rq); 2456 2457 sched_preempt_enable_no_resched(); 2458 if (need_resched()) 2459 goto need_resched; 2460} 2461 2462static inline void sched_submit_work(struct task_struct *tsk) 2463{ 2464 if (!tsk->state || tsk_is_pi_blocked(tsk)) 2465 return; 2466 /* 2467 * If we are going to sleep and we have plugged IO queued, 2468 * make sure to submit it to avoid deadlocks. 2469 */ 2470 if (blk_needs_flush_plug(tsk)) 2471 blk_schedule_flush_plug(tsk); 2472} 2473 2474asmlinkage void __sched schedule(void) 2475{ 2476 struct task_struct *tsk = current; 2477 2478 sched_submit_work(tsk); 2479 __schedule(); 2480} 2481EXPORT_SYMBOL(schedule); 2482 2483#ifdef CONFIG_CONTEXT_TRACKING 2484asmlinkage void __sched schedule_user(void) 2485{ 2486 /* 2487 * If we come here after a random call to set_need_resched(), 2488 * or we have been woken up remotely but the IPI has not yet arrived, 2489 * we haven't yet exited the RCU idle mode. Do it here manually until 2490 * we find a better solution. 2491 */ 2492 user_exit(); 2493 schedule(); 2494 user_enter(); 2495} 2496#endif 2497 2498/** 2499 * schedule_preempt_disabled - called with preemption disabled 2500 * 2501 * Returns with preemption disabled. Note: preempt_count must be 1 2502 */ 2503void __sched schedule_preempt_disabled(void) 2504{ 2505 sched_preempt_enable_no_resched(); 2506 schedule(); 2507 preempt_disable(); 2508} 2509 2510#ifdef CONFIG_PREEMPT 2511/* 2512 * this is the entry point to schedule() from in-kernel preemption 2513 * off of preempt_enable. Kernel preemptions off return from interrupt 2514 * occur there and call schedule directly. 2515 */ 2516asmlinkage void __sched notrace preempt_schedule(void) 2517{ 2518 struct thread_info *ti = current_thread_info(); 2519 2520 /* 2521 * If there is a non-zero preempt_count or interrupts are disabled, 2522 * we do not want to preempt the current task. Just return.. 2523 */ 2524 if (likely(ti->preempt_count || irqs_disabled())) 2525 return; 2526 2527 do { 2528 add_preempt_count_notrace(PREEMPT_ACTIVE); 2529 __schedule(); 2530 sub_preempt_count_notrace(PREEMPT_ACTIVE); 2531 2532 /* 2533 * Check again in case we missed a preemption opportunity 2534 * between schedule and now. 2535 */ 2536 barrier(); 2537 } while (need_resched()); 2538} 2539EXPORT_SYMBOL(preempt_schedule); 2540 2541/* 2542 * this is the entry point to schedule() from kernel preemption 2543 * off of irq context. 2544 * Note, that this is called and return with irqs disabled. This will 2545 * protect us against recursive calling from irq. 2546 */ 2547asmlinkage void __sched preempt_schedule_irq(void) 2548{ 2549 struct thread_info *ti = current_thread_info(); 2550 enum ctx_state prev_state; 2551 2552 /* Catch callers which need to be fixed */ 2553 BUG_ON(ti->preempt_count || !irqs_disabled()); 2554 2555 prev_state = exception_enter(); 2556 2557 do { 2558 add_preempt_count(PREEMPT_ACTIVE); 2559 local_irq_enable(); 2560 __schedule(); 2561 local_irq_disable(); 2562 sub_preempt_count(PREEMPT_ACTIVE); 2563 2564 /* 2565 * Check again in case we missed a preemption opportunity 2566 * between schedule and now. 2567 */ 2568 barrier(); 2569 } while (need_resched()); 2570 2571 exception_exit(prev_state); 2572} 2573 2574#endif /* CONFIG_PREEMPT */ 2575 2576int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags, 2577 void *key) 2578{ 2579 return try_to_wake_up(curr->private, mode, wake_flags); 2580} 2581EXPORT_SYMBOL(default_wake_function); 2582 2583/* 2584 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just 2585 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve 2586 * number) then we wake all the non-exclusive tasks and one exclusive task. 2587 * 2588 * There are circumstances in which we can try to wake a task which has already 2589 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns 2590 * zero in this (rare) case, and we handle it by continuing to scan the queue. 2591 */ 2592static void __wake_up_common(wait_queue_head_t *q, unsigned int mode, 2593 int nr_exclusive, int wake_flags, void *key) 2594{ 2595 wait_queue_t *curr, *next; 2596 2597 list_for_each_entry_safe(curr, next, &q->task_list, task_list) { 2598 unsigned flags = curr->flags; 2599 2600 if (curr->func(curr, mode, wake_flags, key) && 2601 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive) 2602 break; 2603 } 2604} 2605 2606/** 2607 * __wake_up - wake up threads blocked on a waitqueue. 2608 * @q: the waitqueue 2609 * @mode: which threads 2610 * @nr_exclusive: how many wake-one or wake-many threads to wake up 2611 * @key: is directly passed to the wakeup function 2612 * 2613 * It may be assumed that this function implies a write memory barrier before 2614 * changing the task state if and only if any tasks are woken up. 2615 */ 2616void __wake_up(wait_queue_head_t *q, unsigned int mode, 2617 int nr_exclusive, void *key) 2618{ 2619 unsigned long flags; 2620 2621 spin_lock_irqsave(&q->lock, flags); 2622 __wake_up_common(q, mode, nr_exclusive, 0, key); 2623 spin_unlock_irqrestore(&q->lock, flags); 2624} 2625EXPORT_SYMBOL(__wake_up); 2626 2627/* 2628 * Same as __wake_up but called with the spinlock in wait_queue_head_t held. 2629 */ 2630void __wake_up_locked(wait_queue_head_t *q, unsigned int mode, int nr) 2631{ 2632 __wake_up_common(q, mode, nr, 0, NULL); 2633} 2634EXPORT_SYMBOL_GPL(__wake_up_locked); 2635 2636void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key) 2637{ 2638 __wake_up_common(q, mode, 1, 0, key); 2639} 2640EXPORT_SYMBOL_GPL(__wake_up_locked_key); 2641 2642/** 2643 * __wake_up_sync_key - wake up threads blocked on a waitqueue. 2644 * @q: the waitqueue 2645 * @mode: which threads 2646 * @nr_exclusive: how many wake-one or wake-many threads to wake up 2647 * @key: opaque value to be passed to wakeup targets 2648 * 2649 * The sync wakeup differs that the waker knows that it will schedule 2650 * away soon, so while the target thread will be woken up, it will not 2651 * be migrated to another CPU - ie. the two threads are 'synchronized' 2652 * with each other. This can prevent needless bouncing between CPUs. 2653 * 2654 * On UP it can prevent extra preemption. 2655 * 2656 * It may be assumed that this function implies a write memory barrier before 2657 * changing the task state if and only if any tasks are woken up. 2658 */ 2659void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode, 2660 int nr_exclusive, void *key) 2661{ 2662 unsigned long flags; 2663 int wake_flags = WF_SYNC; 2664 2665 if (unlikely(!q)) 2666 return; 2667 2668 if (unlikely(!nr_exclusive)) 2669 wake_flags = 0; 2670 2671 spin_lock_irqsave(&q->lock, flags); 2672 __wake_up_common(q, mode, nr_exclusive, wake_flags, key); 2673 spin_unlock_irqrestore(&q->lock, flags); 2674} 2675EXPORT_SYMBOL_GPL(__wake_up_sync_key); 2676 2677/* 2678 * __wake_up_sync - see __wake_up_sync_key() 2679 */ 2680void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive) 2681{ 2682 __wake_up_sync_key(q, mode, nr_exclusive, NULL); 2683} 2684EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */ 2685 2686/** 2687 * complete: - signals a single thread waiting on this completion 2688 * @x: holds the state of this particular completion 2689 * 2690 * This will wake up a single thread waiting on this completion. Threads will be 2691 * awakened in the same order in which they were queued. 2692 * 2693 * See also complete_all(), wait_for_completion() and related routines. 2694 * 2695 * It may be assumed that this function implies a write memory barrier before 2696 * changing the task state if and only if any tasks are woken up. 2697 */ 2698void complete(struct completion *x) 2699{ 2700 unsigned long flags; 2701 2702 spin_lock_irqsave(&x->wait.lock, flags); 2703 x->done++; 2704 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL); 2705 spin_unlock_irqrestore(&x->wait.lock, flags); 2706} 2707EXPORT_SYMBOL(complete); 2708 2709/** 2710 * complete_all: - signals all threads waiting on this completion 2711 * @x: holds the state of this particular completion 2712 * 2713 * This will wake up all threads waiting on this particular completion event. 2714 * 2715 * It may be assumed that this function implies a write memory barrier before 2716 * changing the task state if and only if any tasks are woken up. 2717 */ 2718void complete_all(struct completion *x) 2719{ 2720 unsigned long flags; 2721 2722 spin_lock_irqsave(&x->wait.lock, flags); 2723 x->done += UINT_MAX/2; 2724 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL); 2725 spin_unlock_irqrestore(&x->wait.lock, flags); 2726} 2727EXPORT_SYMBOL(complete_all); 2728 2729static inline long __sched 2730do_wait_for_common(struct completion *x, 2731 long (*action)(long), long timeout, int state) 2732{ 2733 if (!x->done) { 2734 DECLARE_WAITQUEUE(wait, current); 2735 2736 __add_wait_queue_tail_exclusive(&x->wait, &wait); 2737 do { 2738 if (signal_pending_state(state, current)) { 2739 timeout = -ERESTARTSYS; 2740 break; 2741 } 2742 __set_current_state(state); 2743 spin_unlock_irq(&x->wait.lock); 2744 timeout = action(timeout); 2745 spin_lock_irq(&x->wait.lock); 2746 } while (!x->done && timeout); 2747 __remove_wait_queue(&x->wait, &wait); 2748 if (!x->done) 2749 return timeout; 2750 } 2751 x->done--; 2752 return timeout ?: 1; 2753} 2754 2755static inline long __sched 2756__wait_for_common(struct completion *x, 2757 long (*action)(long), long timeout, int state) 2758{ 2759 might_sleep(); 2760 2761 spin_lock_irq(&x->wait.lock); 2762 timeout = do_wait_for_common(x, action, timeout, state); 2763 spin_unlock_irq(&x->wait.lock); 2764 return timeout; 2765} 2766 2767static long __sched 2768wait_for_common(struct completion *x, long timeout, int state) 2769{ 2770 return __wait_for_common(x, schedule_timeout, timeout, state); 2771} 2772 2773static long __sched 2774wait_for_common_io(struct completion *x, long timeout, int state) 2775{ 2776 return __wait_for_common(x, io_schedule_timeout, timeout, state); 2777} 2778 2779/** 2780 * wait_for_completion: - waits for completion of a task 2781 * @x: holds the state of this particular completion 2782 * 2783 * This waits to be signaled for completion of a specific task. It is NOT 2784 * interruptible and there is no timeout. 2785 * 2786 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout 2787 * and interrupt capability. Also see complete(). 2788 */ 2789void __sched wait_for_completion(struct completion *x) 2790{ 2791 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE); 2792} 2793EXPORT_SYMBOL(wait_for_completion); 2794 2795/** 2796 * wait_for_completion_timeout: - waits for completion of a task (w/timeout) 2797 * @x: holds the state of this particular completion 2798 * @timeout: timeout value in jiffies 2799 * 2800 * This waits for either a completion of a specific task to be signaled or for a 2801 * specified timeout to expire. The timeout is in jiffies. It is not 2802 * interruptible. 2803 * 2804 * Return: 0 if timed out, and positive (at least 1, or number of jiffies left 2805 * till timeout) if completed. 2806 */ 2807unsigned long __sched 2808wait_for_completion_timeout(struct completion *x, unsigned long timeout) 2809{ 2810 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE); 2811} 2812EXPORT_SYMBOL(wait_for_completion_timeout); 2813 2814/** 2815 * wait_for_completion_io: - waits for completion of a task 2816 * @x: holds the state of this particular completion 2817 * 2818 * This waits to be signaled for completion of a specific task. It is NOT 2819 * interruptible and there is no timeout. The caller is accounted as waiting 2820 * for IO. 2821 */ 2822void __sched wait_for_completion_io(struct completion *x) 2823{ 2824 wait_for_common_io(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE); 2825} 2826EXPORT_SYMBOL(wait_for_completion_io); 2827 2828/** 2829 * wait_for_completion_io_timeout: - waits for completion of a task (w/timeout) 2830 * @x: holds the state of this particular completion 2831 * @timeout: timeout value in jiffies 2832 * 2833 * This waits for either a completion of a specific task to be signaled or for a 2834 * specified timeout to expire. The timeout is in jiffies. It is not 2835 * interruptible. The caller is accounted as waiting for IO. 2836 * 2837 * Return: 0 if timed out, and positive (at least 1, or number of jiffies left 2838 * till timeout) if completed. 2839 */ 2840unsigned long __sched 2841wait_for_completion_io_timeout(struct completion *x, unsigned long timeout) 2842{ 2843 return wait_for_common_io(x, timeout, TASK_UNINTERRUPTIBLE); 2844} 2845EXPORT_SYMBOL(wait_for_completion_io_timeout); 2846 2847/** 2848 * wait_for_completion_interruptible: - waits for completion of a task (w/intr) 2849 * @x: holds the state of this particular completion 2850 * 2851 * This waits for completion of a specific task to be signaled. It is 2852 * interruptible. 2853 * 2854 * Return: -ERESTARTSYS if interrupted, 0 if completed. 2855 */ 2856int __sched wait_for_completion_interruptible(struct completion *x) 2857{ 2858 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE); 2859 if (t == -ERESTARTSYS) 2860 return t; 2861 return 0; 2862} 2863EXPORT_SYMBOL(wait_for_completion_interruptible); 2864 2865/** 2866 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr)) 2867 * @x: holds the state of this particular completion 2868 * @timeout: timeout value in jiffies 2869 * 2870 * This waits for either a completion of a specific task to be signaled or for a 2871 * specified timeout to expire. It is interruptible. The timeout is in jiffies. 2872 * 2873 * Return: -ERESTARTSYS if interrupted, 0 if timed out, positive (at least 1, 2874 * or number of jiffies left till timeout) if completed. 2875 */ 2876long __sched 2877wait_for_completion_interruptible_timeout(struct completion *x, 2878 unsigned long timeout) 2879{ 2880 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE); 2881} 2882EXPORT_SYMBOL(wait_for_completion_interruptible_timeout); 2883 2884/** 2885 * wait_for_completion_killable: - waits for completion of a task (killable) 2886 * @x: holds the state of this particular completion 2887 * 2888 * This waits to be signaled for completion of a specific task. It can be 2889 * interrupted by a kill signal. 2890 * 2891 * Return: -ERESTARTSYS if interrupted, 0 if completed. 2892 */ 2893int __sched wait_for_completion_killable(struct completion *x) 2894{ 2895 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE); 2896 if (t == -ERESTARTSYS) 2897 return t; 2898 return 0; 2899} 2900EXPORT_SYMBOL(wait_for_completion_killable); 2901 2902/** 2903 * wait_for_completion_killable_timeout: - waits for completion of a task (w/(to,killable)) 2904 * @x: holds the state of this particular completion 2905 * @timeout: timeout value in jiffies 2906 * 2907 * This waits for either a completion of a specific task to be 2908 * signaled or for a specified timeout to expire. It can be 2909 * interrupted by a kill signal. The timeout is in jiffies. 2910 * 2911 * Return: -ERESTARTSYS if interrupted, 0 if timed out, positive (at least 1, 2912 * or number of jiffies left till timeout) if completed. 2913 */ 2914long __sched 2915wait_for_completion_killable_timeout(struct completion *x, 2916 unsigned long timeout) 2917{ 2918 return wait_for_common(x, timeout, TASK_KILLABLE); 2919} 2920EXPORT_SYMBOL(wait_for_completion_killable_timeout); 2921 2922/** 2923 * try_wait_for_completion - try to decrement a completion without blocking 2924 * @x: completion structure 2925 * 2926 * Return: 0 if a decrement cannot be done without blocking 2927 * 1 if a decrement succeeded. 2928 * 2929 * If a completion is being used as a counting completion, 2930 * attempt to decrement the counter without blocking. This 2931 * enables us to avoid waiting if the resource the completion 2932 * is protecting is not available. 2933 */ 2934bool try_wait_for_completion(struct completion *x) 2935{ 2936 unsigned long flags; 2937 int ret = 1; 2938 2939 spin_lock_irqsave(&x->wait.lock, flags); 2940 if (!x->done) 2941 ret = 0; 2942 else 2943 x->done--; 2944 spin_unlock_irqrestore(&x->wait.lock, flags); 2945 return ret; 2946} 2947EXPORT_SYMBOL(try_wait_for_completion); 2948 2949/** 2950 * completion_done - Test to see if a completion has any waiters 2951 * @x: completion structure 2952 * 2953 * Return: 0 if there are waiters (wait_for_completion() in progress) 2954 * 1 if there are no waiters. 2955 * 2956 */ 2957bool completion_done(struct completion *x) 2958{ 2959 unsigned long flags; 2960 int ret = 1; 2961 2962 spin_lock_irqsave(&x->wait.lock, flags); 2963 if (!x->done) 2964 ret = 0; 2965 spin_unlock_irqrestore(&x->wait.lock, flags); 2966 return ret; 2967} 2968EXPORT_SYMBOL(completion_done); 2969 2970static long __sched 2971sleep_on_common(wait_queue_head_t *q, int state, long timeout) 2972{ 2973 unsigned long flags; 2974 wait_queue_t wait; 2975 2976 init_waitqueue_entry(&wait, current); 2977 2978 __set_current_state(state); 2979 2980 spin_lock_irqsave(&q->lock, flags); 2981 __add_wait_queue(q, &wait); 2982 spin_unlock(&q->lock); 2983 timeout = schedule_timeout(timeout); 2984 spin_lock_irq(&q->lock); 2985 __remove_wait_queue(q, &wait); 2986 spin_unlock_irqrestore(&q->lock, flags); 2987 2988 return timeout; 2989} 2990 2991void __sched interruptible_sleep_on(wait_queue_head_t *q) 2992{ 2993 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT); 2994} 2995EXPORT_SYMBOL(interruptible_sleep_on); 2996 2997long __sched 2998interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout) 2999{ 3000 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout); 3001} 3002EXPORT_SYMBOL(interruptible_sleep_on_timeout); 3003 3004void __sched sleep_on(wait_queue_head_t *q) 3005{ 3006 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT); 3007} 3008EXPORT_SYMBOL(sleep_on); 3009 3010long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout) 3011{ 3012 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout); 3013} 3014EXPORT_SYMBOL(sleep_on_timeout); 3015 3016#ifdef CONFIG_RT_MUTEXES 3017 3018/* 3019 * rt_mutex_setprio - set the current priority of a task 3020 * @p: task 3021 * @prio: prio value (kernel-internal form) 3022 * 3023 * This function changes the 'effective' priority of a task. It does 3024 * not touch ->normal_prio like __setscheduler(). 3025 * 3026 * Used by the rt_mutex code to implement priority inheritance logic. 3027 */ 3028void rt_mutex_setprio(struct task_struct *p, int prio) 3029{ 3030 int oldprio, on_rq, running; 3031 struct rq *rq; 3032 const struct sched_class *prev_class; 3033 3034 BUG_ON(prio < 0 || prio > MAX_PRIO); 3035 3036 rq = __task_rq_lock(p); 3037 3038 /* 3039 * Idle task boosting is a nono in general. There is one 3040 * exception, when PREEMPT_RT and NOHZ is active: 3041 * 3042 * The idle task calls get_next_timer_interrupt() and holds 3043 * the timer wheel base->lock on the CPU and another CPU wants 3044 * to access the timer (probably to cancel it). We can safely 3045 * ignore the boosting request, as the idle CPU runs this code 3046 * with interrupts disabled and will complete the lock 3047 * protected section without being interrupted. So there is no 3048 * real need to boost. 3049 */ 3050 if (unlikely(p == rq->idle)) { 3051 WARN_ON(p != rq->curr); 3052 WARN_ON(p->pi_blocked_on); 3053 goto out_unlock; 3054 } 3055 3056 trace_sched_pi_setprio(p, prio); 3057 oldprio = p->prio; 3058 prev_class = p->sched_class; 3059 on_rq = p->on_rq; 3060 running = task_current(rq, p); 3061 if (on_rq) 3062 dequeue_task(rq, p, 0); 3063 if (running) 3064 p->sched_class->put_prev_task(rq, p); 3065 3066 if (rt_prio(prio)) 3067 p->sched_class = &rt_sched_class; 3068 else 3069 p->sched_class = &fair_sched_class; 3070 3071 p->prio = prio; 3072 3073 if (running) 3074 p->sched_class->set_curr_task(rq); 3075 if (on_rq) 3076 enqueue_task(rq, p, oldprio < prio ? ENQUEUE_HEAD : 0); 3077 3078 check_class_changed(rq, p, prev_class, oldprio); 3079out_unlock: 3080 __task_rq_unlock(rq); 3081} 3082#endif 3083void set_user_nice(struct task_struct *p, long nice) 3084{ 3085 int old_prio, delta, on_rq; 3086 unsigned long flags; 3087 struct rq *rq; 3088 3089 if (TASK_NICE(p) == nice || nice < -20 || nice > 19) 3090 return; 3091 /* 3092 * We have to be careful, if called from sys_setpriority(), 3093 * the task might be in the middle of scheduling on another CPU. 3094 */ 3095 rq = task_rq_lock(p, &flags); 3096 /* 3097 * The RT priorities are set via sched_setscheduler(), but we still 3098 * allow the 'normal' nice value to be set - but as expected 3099 * it wont have any effect on scheduling until the task is 3100 * SCHED_FIFO/SCHED_RR: 3101 */ 3102 if (task_has_rt_policy(p)) { 3103 p->static_prio = NICE_TO_PRIO(nice); 3104 goto out_unlock; 3105 } 3106 on_rq = p->on_rq; 3107 if (on_rq) 3108 dequeue_task(rq, p, 0); 3109 3110 p->static_prio = NICE_TO_PRIO(nice); 3111 set_load_weight(p); 3112 old_prio = p->prio; 3113 p->prio = effective_prio(p); 3114 delta = p->prio - old_prio; 3115 3116 if (on_rq) { 3117 enqueue_task(rq, p, 0); 3118 /* 3119 * If the task increased its priority or is running and 3120 * lowered its priority, then reschedule its CPU: 3121 */ 3122 if (delta < 0 || (delta > 0 && task_running(rq, p))) 3123 resched_task(rq->curr); 3124 } 3125out_unlock: 3126 task_rq_unlock(rq, p, &flags); 3127} 3128EXPORT_SYMBOL(set_user_nice); 3129 3130/* 3131 * can_nice - check if a task can reduce its nice value 3132 * @p: task 3133 * @nice: nice value 3134 */ 3135int can_nice(const struct task_struct *p, const int nice) 3136{ 3137 /* convert nice value [19,-20] to rlimit style value [1,40] */ 3138 int nice_rlim = 20 - nice; 3139 3140 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) || 3141 capable(CAP_SYS_NICE)); 3142} 3143 3144#ifdef __ARCH_WANT_SYS_NICE 3145 3146/* 3147 * sys_nice - change the priority of the current process. 3148 * @increment: priority increment 3149 * 3150 * sys_setpriority is a more generic, but much slower function that 3151 * does similar things. 3152 */ 3153SYSCALL_DEFINE1(nice, int, increment) 3154{ 3155 long nice, retval; 3156 3157 /* 3158 * Setpriority might change our priority at the same moment. 3159 * We don't have to worry. Conceptually one call occurs first 3160 * and we have a single winner. 3161 */ 3162 if (increment < -40) 3163 increment = -40; 3164 if (increment > 40) 3165 increment = 40; 3166 3167 nice = TASK_NICE(current) + increment; 3168 if (nice < -20) 3169 nice = -20; 3170 if (nice > 19) 3171 nice = 19; 3172 3173 if (increment < 0 && !can_nice(current, nice)) 3174 return -EPERM; 3175 3176 retval = security_task_setnice(current, nice); 3177 if (retval) 3178 return retval; 3179 3180 set_user_nice(current, nice); 3181 return 0; 3182} 3183 3184#endif 3185 3186/** 3187 * task_prio - return the priority value of a given task. 3188 * @p: the task in question. 3189 * 3190 * Return: The priority value as seen by users in /proc. 3191 * RT tasks are offset by -200. Normal tasks are centered 3192 * around 0, value goes from -16 to +15. 3193 */ 3194int task_prio(const struct task_struct *p) 3195{ 3196 return p->prio - MAX_RT_PRIO; 3197} 3198 3199/** 3200 * task_nice - return the nice value of a given task. 3201 * @p: the task in question. 3202 * 3203 * Return: The nice value [ -20 ... 0 ... 19 ]. 3204 */ 3205int task_nice(const struct task_struct *p) 3206{ 3207 return TASK_NICE(p); 3208} 3209EXPORT_SYMBOL(task_nice); 3210 3211/** 3212 * idle_cpu - is a given cpu idle currently? 3213 * @cpu: the processor in question. 3214 * 3215 * Return: 1 if the CPU is currently idle. 0 otherwise. 3216 */ 3217int idle_cpu(int cpu) 3218{ 3219 struct rq *rq = cpu_rq(cpu); 3220 3221 if (rq->curr != rq->idle) 3222 return 0; 3223 3224 if (rq->nr_running) 3225 return 0; 3226 3227#ifdef CONFIG_SMP 3228 if (!llist_empty(&rq->wake_list)) 3229 return 0; 3230#endif 3231 3232 return 1; 3233} 3234 3235/** 3236 * idle_task - return the idle task for a given cpu. 3237 * @cpu: the processor in question. 3238 * 3239 * Return: The idle task for the cpu @cpu. 3240 */ 3241struct task_struct *idle_task(int cpu) 3242{ 3243 return cpu_rq(cpu)->idle; 3244} 3245 3246/** 3247 * find_process_by_pid - find a process with a matching PID value. 3248 * @pid: the pid in question. 3249 * 3250 * The task of @pid, if found. %NULL otherwise. 3251 */ 3252static struct task_struct *find_process_by_pid(pid_t pid) 3253{ 3254 return pid ? find_task_by_vpid(pid) : current; 3255} 3256 3257/* Actually do priority change: must hold rq lock. */ 3258static void 3259__setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio) 3260{ 3261 p->policy = policy; 3262 p->rt_priority = prio; 3263 p->normal_prio = normal_prio(p); 3264 /* we are holding p->pi_lock already */ 3265 p->prio = rt_mutex_getprio(p); 3266 if (rt_prio(p->prio)) 3267 p->sched_class = &rt_sched_class; 3268 else 3269 p->sched_class = &fair_sched_class; 3270 set_load_weight(p); 3271} 3272 3273/* 3274 * check the target process has a UID that matches the current process's 3275 */ 3276static bool check_same_owner(struct task_struct *p) 3277{ 3278 const struct cred *cred = current_cred(), *pcred; 3279 bool match; 3280 3281 rcu_read_lock(); 3282 pcred = __task_cred(p); 3283 match = (uid_eq(cred->euid, pcred->euid) || 3284 uid_eq(cred->euid, pcred->uid)); 3285 rcu_read_unlock(); 3286 return match; 3287} 3288 3289static int __sched_setscheduler(struct task_struct *p, int policy, 3290 const struct sched_param *param, bool user) 3291{ 3292 int retval, oldprio, oldpolicy = -1, on_rq, running; 3293 unsigned long flags; 3294 const struct sched_class *prev_class; 3295 struct rq *rq; 3296 int reset_on_fork; 3297 3298 /* may grab non-irq protected spin_locks */ 3299 BUG_ON(in_interrupt()); 3300recheck: 3301 /* double check policy once rq lock held */ 3302 if (policy < 0) { 3303 reset_on_fork = p->sched_reset_on_fork; 3304 policy = oldpolicy = p->policy; 3305 } else { 3306 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK); 3307 policy &= ~SCHED_RESET_ON_FORK; 3308 3309 if (policy != SCHED_FIFO && policy != SCHED_RR && 3310 policy != SCHED_NORMAL && policy != SCHED_BATCH && 3311 policy != SCHED_IDLE) 3312 return -EINVAL; 3313 } 3314 3315 /* 3316 * Valid priorities for SCHED_FIFO and SCHED_RR are 3317 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL, 3318 * SCHED_BATCH and SCHED_IDLE is 0. 3319 */ 3320 if (param->sched_priority < 0 || 3321 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) || 3322 (!p->mm && param->sched_priority > MAX_RT_PRIO-1)) 3323 return -EINVAL; 3324 if (rt_policy(policy) != (param->sched_priority != 0)) 3325 return -EINVAL; 3326 3327 /* 3328 * Allow unprivileged RT tasks to decrease priority: 3329 */ 3330 if (user && !capable(CAP_SYS_NICE)) { 3331 if (rt_policy(policy)) { 3332 unsigned long rlim_rtprio = 3333 task_rlimit(p, RLIMIT_RTPRIO); 3334 3335 /* can't set/change the rt policy */ 3336 if (policy != p->policy && !rlim_rtprio) 3337 return -EPERM; 3338 3339 /* can't increase priority */ 3340 if (param->sched_priority > p->rt_priority && 3341 param->sched_priority > rlim_rtprio) 3342 return -EPERM; 3343 } 3344 3345 /* 3346 * Treat SCHED_IDLE as nice 20. Only allow a switch to 3347 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it. 3348 */ 3349 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) { 3350 if (!can_nice(p, TASK_NICE(p))) 3351 return -EPERM; 3352 } 3353 3354 /* can't change other user's priorities */ 3355 if (!check_same_owner(p)) 3356 return -EPERM; 3357 3358 /* Normal users shall not reset the sched_reset_on_fork flag */ 3359 if (p->sched_reset_on_fork && !reset_on_fork) 3360 return -EPERM; 3361 } 3362 3363 if (user) { 3364 retval = security_task_setscheduler(p); 3365 if (retval) 3366 return retval; 3367 } 3368 3369 /* 3370 * make sure no PI-waiters arrive (or leave) while we are 3371 * changing the priority of the task: 3372 * 3373 * To be able to change p->policy safely, the appropriate 3374 * runqueue lock must be held. 3375 */ 3376 rq = task_rq_lock(p, &flags); 3377 3378 /* 3379 * Changing the policy of the stop threads its a very bad idea 3380 */ 3381 if (p == rq->stop) { 3382 task_rq_unlock(rq, p, &flags); 3383 return -EINVAL; 3384 } 3385 3386 /* 3387 * If not changing anything there's no need to proceed further: 3388 */ 3389 if (unlikely(policy == p->policy && (!rt_policy(policy) || 3390 param->sched_priority == p->rt_priority))) { 3391 task_rq_unlock(rq, p, &flags); 3392 return 0; 3393 } 3394 3395#ifdef CONFIG_RT_GROUP_SCHED 3396 if (user) { 3397 /* 3398 * Do not allow realtime tasks into groups that have no runtime 3399 * assigned. 3400 */ 3401 if (rt_bandwidth_enabled() && rt_policy(policy) && 3402 task_group(p)->rt_bandwidth.rt_runtime == 0 && 3403 !task_group_is_autogroup(task_group(p))) { 3404 task_rq_unlock(rq, p, &flags); 3405 return -EPERM; 3406 } 3407 } 3408#endif 3409 3410 /* recheck policy now with rq lock held */ 3411 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) { 3412 policy = oldpolicy = -1; 3413 task_rq_unlock(rq, p, &flags); 3414 goto recheck; 3415 } 3416 on_rq = p->on_rq; 3417 running = task_current(rq, p); 3418 if (on_rq) 3419 dequeue_task(rq, p, 0); 3420 if (running) 3421 p->sched_class->put_prev_task(rq, p); 3422 3423 p->sched_reset_on_fork = reset_on_fork; 3424 3425 oldprio = p->prio; 3426 prev_class = p->sched_class; 3427 __setscheduler(rq, p, policy, param->sched_priority); 3428 3429 if (running) 3430 p->sched_class->set_curr_task(rq); 3431 if (on_rq) 3432 enqueue_task(rq, p, 0); 3433 3434 check_class_changed(rq, p, prev_class, oldprio); 3435 task_rq_unlock(rq, p, &flags); 3436 3437 rt_mutex_adjust_pi(p); 3438 3439 return 0; 3440} 3441 3442/** 3443 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread. 3444 * @p: the task in question. 3445 * @policy: new policy. 3446 * @param: structure containing the new RT priority. 3447 * 3448 * Return: 0 on success. An error code otherwise. 3449 * 3450 * NOTE that the task may be already dead. 3451 */ 3452int sched_setscheduler(struct task_struct *p, int policy, 3453 const struct sched_param *param) 3454{ 3455 return __sched_setscheduler(p, policy, param, true); 3456} 3457EXPORT_SYMBOL_GPL(sched_setscheduler); 3458 3459/** 3460 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace. 3461 * @p: the task in question. 3462 * @policy: new policy. 3463 * @param: structure containing the new RT priority. 3464 * 3465 * Just like sched_setscheduler, only don't bother checking if the 3466 * current context has permission. For example, this is needed in 3467 * stop_machine(): we create temporary high priority worker threads, 3468 * but our caller might not have that capability. 3469 * 3470 * Return: 0 on success. An error code otherwise. 3471 */ 3472int sched_setscheduler_nocheck(struct task_struct *p, int policy, 3473 const struct sched_param *param) 3474{ 3475 return __sched_setscheduler(p, policy, param, false); 3476} 3477 3478static int 3479do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param) 3480{ 3481 struct sched_param lparam; 3482 struct task_struct *p; 3483 int retval; 3484 3485 if (!param || pid < 0) 3486 return -EINVAL; 3487 if (copy_from_user(&lparam, param, sizeof(struct sched_param))) 3488 return -EFAULT; 3489 3490 rcu_read_lock(); 3491 retval = -ESRCH; 3492 p = find_process_by_pid(pid); 3493 if (p != NULL) 3494 retval = sched_setscheduler(p, policy, &lparam); 3495 rcu_read_unlock(); 3496 3497 return retval; 3498} 3499 3500/** 3501 * sys_sched_setscheduler - set/change the scheduler policy and RT priority 3502 * @pid: the pid in question. 3503 * @policy: new policy. 3504 * @param: structure containing the new RT priority. 3505 * 3506 * Return: 0 on success. An error code otherwise. 3507 */ 3508SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy, 3509 struct sched_param __user *, param) 3510{ 3511 /* negative values for policy are not valid */ 3512 if (policy < 0) 3513 return -EINVAL; 3514 3515 return do_sched_setscheduler(pid, policy, param); 3516} 3517 3518/** 3519 * sys_sched_setparam - set/change the RT priority of a thread 3520 * @pid: the pid in question. 3521 * @param: structure containing the new RT priority. 3522 * 3523 * Return: 0 on success. An error code otherwise. 3524 */ 3525SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param) 3526{ 3527 return do_sched_setscheduler(pid, -1, param); 3528} 3529 3530/** 3531 * sys_sched_getscheduler - get the policy (scheduling class) of a thread 3532 * @pid: the pid in question. 3533 * 3534 * Return: On success, the policy of the thread. Otherwise, a negative error 3535 * code. 3536 */ 3537SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid) 3538{ 3539 struct task_struct *p; 3540 int retval; 3541 3542 if (pid < 0) 3543 return -EINVAL; 3544 3545 retval = -ESRCH; 3546 rcu_read_lock(); 3547 p = find_process_by_pid(pid); 3548 if (p) { 3549 retval = security_task_getscheduler(p); 3550 if (!retval) 3551 retval = p->policy 3552 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0); 3553 } 3554 rcu_read_unlock(); 3555 return retval; 3556} 3557 3558/** 3559 * sys_sched_getparam - get the RT priority of a thread 3560 * @pid: the pid in question. 3561 * @param: structure containing the RT priority. 3562 * 3563 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error 3564 * code. 3565 */ 3566SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param) 3567{ 3568 struct sched_param lp; 3569 struct task_struct *p; 3570 int retval; 3571 3572 if (!param || pid < 0) 3573 return -EINVAL; 3574 3575 rcu_read_lock(); 3576 p = find_process_by_pid(pid); 3577 retval = -ESRCH; 3578 if (!p) 3579 goto out_unlock; 3580 3581 retval = security_task_getscheduler(p); 3582 if (retval) 3583 goto out_unlock; 3584 3585 lp.sched_priority = p->rt_priority; 3586 rcu_read_unlock(); 3587 3588 /* 3589 * This one might sleep, we cannot do it with a spinlock held ... 3590 */ 3591 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0; 3592 3593 return retval; 3594 3595out_unlock: 3596 rcu_read_unlock(); 3597 return retval; 3598} 3599 3600long sched_setaffinity(pid_t pid, const struct cpumask *in_mask) 3601{ 3602 cpumask_var_t cpus_allowed, new_mask; 3603 struct task_struct *p; 3604 int retval; 3605 3606 get_online_cpus(); 3607 rcu_read_lock(); 3608 3609 p = find_process_by_pid(pid); 3610 if (!p) { 3611 rcu_read_unlock(); 3612 put_online_cpus(); 3613 return -ESRCH; 3614 } 3615 3616 /* Prevent p going away */ 3617 get_task_struct(p); 3618 rcu_read_unlock(); 3619 3620 if (p->flags & PF_NO_SETAFFINITY) { 3621 retval = -EINVAL; 3622 goto out_put_task; 3623 } 3624 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) { 3625 retval = -ENOMEM; 3626 goto out_put_task; 3627 } 3628 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) { 3629 retval = -ENOMEM; 3630 goto out_free_cpus_allowed; 3631 } 3632 retval = -EPERM; 3633 if (!check_same_owner(p)) { 3634 rcu_read_lock(); 3635 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) { 3636 rcu_read_unlock(); 3637 goto out_unlock; 3638 } 3639 rcu_read_unlock(); 3640 } 3641 3642 retval = security_task_setscheduler(p); 3643 if (retval) 3644 goto out_unlock; 3645 3646 cpuset_cpus_allowed(p, cpus_allowed); 3647 cpumask_and(new_mask, in_mask, cpus_allowed); 3648again: 3649 retval = set_cpus_allowed_ptr(p, new_mask); 3650 3651 if (!retval) { 3652 cpuset_cpus_allowed(p, cpus_allowed); 3653 if (!cpumask_subset(new_mask, cpus_allowed)) { 3654 /* 3655 * We must have raced with a concurrent cpuset 3656 * update. Just reset the cpus_allowed to the 3657 * cpuset's cpus_allowed 3658 */ 3659 cpumask_copy(new_mask, cpus_allowed); 3660 goto again; 3661 } 3662 } 3663out_unlock: 3664 free_cpumask_var(new_mask); 3665out_free_cpus_allowed: 3666 free_cpumask_var(cpus_allowed); 3667out_put_task: 3668 put_task_struct(p); 3669 put_online_cpus(); 3670 return retval; 3671} 3672 3673static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len, 3674 struct cpumask *new_mask) 3675{ 3676 if (len < cpumask_size()) 3677 cpumask_clear(new_mask); 3678 else if (len > cpumask_size()) 3679 len = cpumask_size(); 3680 3681 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0; 3682} 3683 3684/** 3685 * sys_sched_setaffinity - set the cpu affinity of a process 3686 * @pid: pid of the process 3687 * @len: length in bytes of the bitmask pointed to by user_mask_ptr 3688 * @user_mask_ptr: user-space pointer to the new cpu mask 3689 * 3690 * Return: 0 on success. An error code otherwise. 3691 */ 3692SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len, 3693 unsigned long __user *, user_mask_ptr) 3694{ 3695 cpumask_var_t new_mask; 3696 int retval; 3697 3698 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) 3699 return -ENOMEM; 3700 3701 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask); 3702 if (retval == 0) 3703 retval = sched_setaffinity(pid, new_mask); 3704 free_cpumask_var(new_mask); 3705 return retval; 3706} 3707 3708long sched_getaffinity(pid_t pid, struct cpumask *mask) 3709{ 3710 struct task_struct *p; 3711 unsigned long flags; 3712 int retval; 3713 3714 get_online_cpus(); 3715 rcu_read_lock(); 3716 3717 retval = -ESRCH; 3718 p = find_process_by_pid(pid); 3719 if (!p) 3720 goto out_unlock; 3721 3722 retval = security_task_getscheduler(p); 3723 if (retval) 3724 goto out_unlock; 3725 3726 raw_spin_lock_irqsave(&p->pi_lock, flags); 3727 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask); 3728 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 3729 3730out_unlock: 3731 rcu_read_unlock(); 3732 put_online_cpus(); 3733 3734 return retval; 3735} 3736 3737/** 3738 * sys_sched_getaffinity - get the cpu affinity of a process 3739 * @pid: pid of the process 3740 * @len: length in bytes of the bitmask pointed to by user_mask_ptr 3741 * @user_mask_ptr: user-space pointer to hold the current cpu mask 3742 * 3743 * Return: 0 on success. An error code otherwise. 3744 */ 3745SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len, 3746 unsigned long __user *, user_mask_ptr) 3747{ 3748 int ret; 3749 cpumask_var_t mask; 3750 3751 if ((len * BITS_PER_BYTE) < nr_cpu_ids) 3752 return -EINVAL; 3753 if (len & (sizeof(unsigned long)-1)) 3754 return -EINVAL; 3755 3756 if (!alloc_cpumask_var(&mask, GFP_KERNEL)) 3757 return -ENOMEM; 3758 3759 ret = sched_getaffinity(pid, mask); 3760 if (ret == 0) { 3761 size_t retlen = min_t(size_t, len, cpumask_size()); 3762 3763 if (copy_to_user(user_mask_ptr, mask, retlen)) 3764 ret = -EFAULT; 3765 else 3766 ret = retlen; 3767 } 3768 free_cpumask_var(mask); 3769 3770 return ret; 3771} 3772 3773/** 3774 * sys_sched_yield - yield the current processor to other threads. 3775 * 3776 * This function yields the current CPU to other tasks. If there are no 3777 * other threads running on this CPU then this function will return. 3778 * 3779 * Return: 0. 3780 */ 3781SYSCALL_DEFINE0(sched_yield) 3782{ 3783 struct rq *rq = this_rq_lock(); 3784 3785 schedstat_inc(rq, yld_count); 3786 current->sched_class->yield_task(rq); 3787 3788 /* 3789 * Since we are going to call schedule() anyway, there's 3790 * no need to preempt or enable interrupts: 3791 */ 3792 __release(rq->lock); 3793 spin_release(&rq->lock.dep_map, 1, _THIS_IP_); 3794 do_raw_spin_unlock(&rq->lock); 3795 sched_preempt_enable_no_resched(); 3796 3797 schedule(); 3798 3799 return 0; 3800} 3801 3802static inline int should_resched(void) 3803{ 3804 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE); 3805} 3806 3807static void __cond_resched(void) 3808{ 3809 add_preempt_count(PREEMPT_ACTIVE); 3810 __schedule(); 3811 sub_preempt_count(PREEMPT_ACTIVE); 3812} 3813 3814int __sched _cond_resched(void) 3815{ 3816 if (should_resched()) { 3817 __cond_resched(); 3818 return 1; 3819 } 3820 return 0; 3821} 3822EXPORT_SYMBOL(_cond_resched); 3823 3824/* 3825 * __cond_resched_lock() - if a reschedule is pending, drop the given lock, 3826 * call schedule, and on return reacquire the lock. 3827 * 3828 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level 3829 * operations here to prevent schedule() from being called twice (once via 3830 * spin_unlock(), once by hand). 3831 */ 3832int __cond_resched_lock(spinlock_t *lock) 3833{ 3834 int resched = should_resched(); 3835 int ret = 0; 3836 3837 lockdep_assert_held(lock); 3838 3839 if (spin_needbreak(lock) || resched) { 3840 spin_unlock(lock); 3841 if (resched) 3842 __cond_resched(); 3843 else 3844 cpu_relax(); 3845 ret = 1; 3846 spin_lock(lock); 3847 } 3848 return ret; 3849} 3850EXPORT_SYMBOL(__cond_resched_lock); 3851 3852int __sched __cond_resched_softirq(void) 3853{ 3854 BUG_ON(!in_softirq()); 3855 3856 if (should_resched()) { 3857 local_bh_enable(); 3858 __cond_resched(); 3859 local_bh_disable(); 3860 return 1; 3861 } 3862 return 0; 3863} 3864EXPORT_SYMBOL(__cond_resched_softirq); 3865 3866/** 3867 * yield - yield the current processor to other threads. 3868 * 3869 * Do not ever use this function, there's a 99% chance you're doing it wrong. 3870 * 3871 * The scheduler is at all times free to pick the calling task as the most 3872 * eligible task to run, if removing the yield() call from your code breaks 3873 * it, its already broken. 3874 * 3875 * Typical broken usage is: 3876 * 3877 * while (!event) 3878 * yield(); 3879 * 3880 * where one assumes that yield() will let 'the other' process run that will 3881 * make event true. If the current task is a SCHED_FIFO task that will never 3882 * happen. Never use yield() as a progress guarantee!! 3883 * 3884 * If you want to use yield() to wait for something, use wait_event(). 3885 * If you want to use yield() to be 'nice' for others, use cond_resched(). 3886 * If you still want to use yield(), do not! 3887 */ 3888void __sched yield(void) 3889{ 3890 set_current_state(TASK_RUNNING); 3891 sys_sched_yield(); 3892} 3893EXPORT_SYMBOL(yield); 3894 3895/** 3896 * yield_to - yield the current processor to another thread in 3897 * your thread group, or accelerate that thread toward the 3898 * processor it's on. 3899 * @p: target task 3900 * @preempt: whether task preemption is allowed or not 3901 * 3902 * It's the caller's job to ensure that the target task struct 3903 * can't go away on us before we can do any checks. 3904 * 3905 * Return: 3906 * true (>0) if we indeed boosted the target task. 3907 * false (0) if we failed to boost the target. 3908 * -ESRCH if there's no task to yield to. 3909 */ 3910bool __sched yield_to(struct task_struct *p, bool preempt) 3911{ 3912 struct task_struct *curr = current; 3913 struct rq *rq, *p_rq; 3914 unsigned long flags; 3915 int yielded = 0; 3916 3917 local_irq_save(flags); 3918 rq = this_rq(); 3919 3920again: 3921 p_rq = task_rq(p); 3922 /* 3923 * If we're the only runnable task on the rq and target rq also 3924 * has only one task, there's absolutely no point in yielding. 3925 */ 3926 if (rq->nr_running == 1 && p_rq->nr_running == 1) { 3927 yielded = -ESRCH; 3928 goto out_irq; 3929 } 3930 3931 double_rq_lock(rq, p_rq); 3932 while (task_rq(p) != p_rq) { 3933 double_rq_unlock(rq, p_rq); 3934 goto again; 3935 } 3936 3937 if (!curr->sched_class->yield_to_task) 3938 goto out_unlock; 3939 3940 if (curr->sched_class != p->sched_class) 3941 goto out_unlock; 3942 3943 if (task_running(p_rq, p) || p->state) 3944 goto out_unlock; 3945 3946 yielded = curr->sched_class->yield_to_task(rq, p, preempt); 3947 if (yielded) { 3948 schedstat_inc(rq, yld_count); 3949 /* 3950 * Make p's CPU reschedule; pick_next_entity takes care of 3951 * fairness. 3952 */ 3953 if (preempt && rq != p_rq) 3954 resched_task(p_rq->curr); 3955 } 3956 3957out_unlock: 3958 double_rq_unlock(rq, p_rq); 3959out_irq: 3960 local_irq_restore(flags); 3961 3962 if (yielded > 0) 3963 schedule(); 3964 3965 return yielded; 3966} 3967EXPORT_SYMBOL_GPL(yield_to); 3968 3969/* 3970 * This task is about to go to sleep on IO. Increment rq->nr_iowait so 3971 * that process accounting knows that this is a task in IO wait state. 3972 */ 3973void __sched io_schedule(void) 3974{ 3975 struct rq *rq = raw_rq(); 3976 3977 delayacct_blkio_start(); 3978 atomic_inc(&rq->nr_iowait); 3979 blk_flush_plug(current); 3980 current->in_iowait = 1; 3981 schedule(); 3982 current->in_iowait = 0; 3983 atomic_dec(&rq->nr_iowait); 3984 delayacct_blkio_end(); 3985} 3986EXPORT_SYMBOL(io_schedule); 3987 3988long __sched io_schedule_timeout(long timeout) 3989{ 3990 struct rq *rq = raw_rq(); 3991 long ret; 3992 3993 delayacct_blkio_start(); 3994 atomic_inc(&rq->nr_iowait); 3995 blk_flush_plug(current); 3996 current->in_iowait = 1; 3997 ret = schedule_timeout(timeout); 3998 current->in_iowait = 0; 3999 atomic_dec(&rq->nr_iowait); 4000 delayacct_blkio_end(); 4001 return ret; 4002} 4003 4004/** 4005 * sys_sched_get_priority_max - return maximum RT priority. 4006 * @policy: scheduling class. 4007 * 4008 * Return: On success, this syscall returns the maximum 4009 * rt_priority that can be used by a given scheduling class. 4010 * On failure, a negative error code is returned. 4011 */ 4012SYSCALL_DEFINE1(sched_get_priority_max, int, policy) 4013{ 4014 int ret = -EINVAL; 4015 4016 switch (policy) { 4017 case SCHED_FIFO: 4018 case SCHED_RR: 4019 ret = MAX_USER_RT_PRIO-1; 4020 break; 4021 case SCHED_NORMAL: 4022 case SCHED_BATCH: 4023 case SCHED_IDLE: 4024 ret = 0; 4025 break; 4026 } 4027 return ret; 4028} 4029 4030/** 4031 * sys_sched_get_priority_min - return minimum RT priority. 4032 * @policy: scheduling class. 4033 * 4034 * Return: On success, this syscall returns the minimum 4035 * rt_priority that can be used by a given scheduling class. 4036 * On failure, a negative error code is returned. 4037 */ 4038SYSCALL_DEFINE1(sched_get_priority_min, int, policy) 4039{ 4040 int ret = -EINVAL; 4041 4042 switch (policy) { 4043 case SCHED_FIFO: 4044 case SCHED_RR: 4045 ret = 1; 4046 break; 4047 case SCHED_NORMAL: 4048 case SCHED_BATCH: 4049 case SCHED_IDLE: 4050 ret = 0; 4051 } 4052 return ret; 4053} 4054 4055/** 4056 * sys_sched_rr_get_interval - return the default timeslice of a process. 4057 * @pid: pid of the process. 4058 * @interval: userspace pointer to the timeslice value. 4059 * 4060 * this syscall writes the default timeslice value of a given process 4061 * into the user-space timespec buffer. A value of '0' means infinity. 4062 * 4063 * Return: On success, 0 and the timeslice is in @interval. Otherwise, 4064 * an error code. 4065 */ 4066SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid, 4067 struct timespec __user *, interval) 4068{ 4069 struct task_struct *p; 4070 unsigned int time_slice; 4071 unsigned long flags; 4072 struct rq *rq; 4073 int retval; 4074 struct timespec t; 4075 4076 if (pid < 0) 4077 return -EINVAL; 4078 4079 retval = -ESRCH; 4080 rcu_read_lock(); 4081 p = find_process_by_pid(pid); 4082 if (!p) 4083 goto out_unlock; 4084 4085 retval = security_task_getscheduler(p); 4086 if (retval) 4087 goto out_unlock; 4088 4089 rq = task_rq_lock(p, &flags); 4090 time_slice = p->sched_class->get_rr_interval(rq, p); 4091 task_rq_unlock(rq, p, &flags); 4092 4093 rcu_read_unlock(); 4094 jiffies_to_timespec(time_slice, &t); 4095 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0; 4096 return retval; 4097 4098out_unlock: 4099 rcu_read_unlock(); 4100 return retval; 4101} 4102 4103static const char stat_nam[] = TASK_STATE_TO_CHAR_STR; 4104 4105void sched_show_task(struct task_struct *p) 4106{ 4107 unsigned long free = 0; 4108 int ppid; 4109 unsigned state; 4110 4111 state = p->state ? __ffs(p->state) + 1 : 0; 4112 printk(KERN_INFO "%-15.15s %c", p->comm, 4113 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?'); 4114#if BITS_PER_LONG == 32 4115 if (state == TASK_RUNNING) 4116 printk(KERN_CONT " running "); 4117 else 4118 printk(KERN_CONT " %08lx ", thread_saved_pc(p)); 4119#else 4120 if (state == TASK_RUNNING) 4121 printk(KERN_CONT " running task "); 4122 else 4123 printk(KERN_CONT " %016lx ", thread_saved_pc(p)); 4124#endif 4125#ifdef CONFIG_DEBUG_STACK_USAGE 4126 free = stack_not_used(p); 4127#endif 4128 rcu_read_lock(); 4129 ppid = task_pid_nr(rcu_dereference(p->real_parent)); 4130 rcu_read_unlock(); 4131 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free, 4132 task_pid_nr(p), ppid, 4133 (unsigned long)task_thread_info(p)->flags); 4134 4135 print_worker_info(KERN_INFO, p); 4136 show_stack(p, NULL); 4137} 4138 4139void show_state_filter(unsigned long state_filter) 4140{ 4141 struct task_struct *g, *p; 4142 4143#if BITS_PER_LONG == 32 4144 printk(KERN_INFO 4145 " task PC stack pid father\n"); 4146#else 4147 printk(KERN_INFO 4148 " task PC stack pid father\n"); 4149#endif 4150 rcu_read_lock(); 4151 do_each_thread(g, p) { 4152 /* 4153 * reset the NMI-timeout, listing all files on a slow 4154 * console might take a lot of time: 4155 */ 4156 touch_nmi_watchdog(); 4157 if (!state_filter || (p->state & state_filter)) 4158 sched_show_task(p); 4159 } while_each_thread(g, p); 4160 4161 touch_all_softlockup_watchdogs(); 4162 4163#ifdef CONFIG_SCHED_DEBUG 4164 sysrq_sched_debug_show(); 4165#endif 4166 rcu_read_unlock(); 4167 /* 4168 * Only show locks if all tasks are dumped: 4169 */ 4170 if (!state_filter) 4171 debug_show_all_locks(); 4172} 4173 4174void __cpuinit init_idle_bootup_task(struct task_struct *idle) 4175{ 4176 idle->sched_class = &idle_sched_class; 4177} 4178 4179/** 4180 * init_idle - set up an idle thread for a given CPU 4181 * @idle: task in question 4182 * @cpu: cpu the idle task belongs to 4183 * 4184 * NOTE: this function does not set the idle thread's NEED_RESCHED 4185 * flag, to make booting more robust. 4186 */ 4187void __cpuinit init_idle(struct task_struct *idle, int cpu) 4188{ 4189 struct rq *rq = cpu_rq(cpu); 4190 unsigned long flags; 4191 4192 raw_spin_lock_irqsave(&rq->lock, flags); 4193 4194 __sched_fork(idle); 4195 idle->state = TASK_RUNNING; 4196 idle->se.exec_start = sched_clock(); 4197 4198 do_set_cpus_allowed(idle, cpumask_of(cpu)); 4199 /* 4200 * We're having a chicken and egg problem, even though we are 4201 * holding rq->lock, the cpu isn't yet set to this cpu so the 4202 * lockdep check in task_group() will fail. 4203 * 4204 * Similar case to sched_fork(). / Alternatively we could 4205 * use task_rq_lock() here and obtain the other rq->lock. 4206 * 4207 * Silence PROVE_RCU 4208 */ 4209 rcu_read_lock(); 4210 __set_task_cpu(idle, cpu); 4211 rcu_read_unlock(); 4212 4213 rq->curr = rq->idle = idle; 4214#if defined(CONFIG_SMP) 4215 idle->on_cpu = 1; 4216#endif 4217 raw_spin_unlock_irqrestore(&rq->lock, flags); 4218 4219 /* Set the preempt count _outside_ the spinlocks! */ 4220 task_thread_info(idle)->preempt_count = 0; 4221 4222 /* 4223 * The idle tasks have their own, simple scheduling class: 4224 */ 4225 idle->sched_class = &idle_sched_class; 4226 ftrace_graph_init_idle_task(idle, cpu); 4227 vtime_init_idle(idle, cpu); 4228#if defined(CONFIG_SMP) 4229 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu); 4230#endif 4231} 4232 4233#ifdef CONFIG_SMP 4234void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask) 4235{ 4236 if (p->sched_class && p->sched_class->set_cpus_allowed) 4237 p->sched_class->set_cpus_allowed(p, new_mask); 4238 4239 cpumask_copy(&p->cpus_allowed, new_mask); 4240 p->nr_cpus_allowed = cpumask_weight(new_mask); 4241} 4242 4243/* 4244 * This is how migration works: 4245 * 4246 * 1) we invoke migration_cpu_stop() on the target CPU using 4247 * stop_one_cpu(). 4248 * 2) stopper starts to run (implicitly forcing the migrated thread 4249 * off the CPU) 4250 * 3) it checks whether the migrated task is still in the wrong runqueue. 4251 * 4) if it's in the wrong runqueue then the migration thread removes 4252 * it and puts it into the right queue. 4253 * 5) stopper completes and stop_one_cpu() returns and the migration 4254 * is done. 4255 */ 4256 4257/* 4258 * Change a given task's CPU affinity. Migrate the thread to a 4259 * proper CPU and schedule it away if the CPU it's executing on 4260 * is removed from the allowed bitmask. 4261 * 4262 * NOTE: the caller must have a valid reference to the task, the 4263 * task must not exit() & deallocate itself prematurely. The 4264 * call is not atomic; no spinlocks may be held. 4265 */ 4266int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask) 4267{ 4268 unsigned long flags; 4269 struct rq *rq; 4270 unsigned int dest_cpu; 4271 int ret = 0; 4272 4273 rq = task_rq_lock(p, &flags); 4274 4275 if (cpumask_equal(&p->cpus_allowed, new_mask)) 4276 goto out; 4277 4278 if (!cpumask_intersects(new_mask, cpu_active_mask)) { 4279 ret = -EINVAL; 4280 goto out; 4281 } 4282 4283 do_set_cpus_allowed(p, new_mask); 4284 4285 /* Can the task run on the task's current CPU? If so, we're done */ 4286 if (cpumask_test_cpu(task_cpu(p), new_mask)) 4287 goto out; 4288 4289 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask); 4290 if (p->on_rq) { 4291 struct migration_arg arg = { p, dest_cpu }; 4292 /* Need help from migration thread: drop lock and wait. */ 4293 task_rq_unlock(rq, p, &flags); 4294 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg); 4295 tlb_migrate_finish(p->mm); 4296 return 0; 4297 } 4298out: 4299 task_rq_unlock(rq, p, &flags); 4300 4301 return ret; 4302} 4303EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr); 4304 4305/* 4306 * Move (not current) task off this cpu, onto dest cpu. We're doing 4307 * this because either it can't run here any more (set_cpus_allowed() 4308 * away from this CPU, or CPU going down), or because we're 4309 * attempting to rebalance this task on exec (sched_exec). 4310 * 4311 * So we race with normal scheduler movements, but that's OK, as long 4312 * as the task is no longer on this CPU. 4313 * 4314 * Returns non-zero if task was successfully migrated. 4315 */ 4316static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu) 4317{ 4318 struct rq *rq_dest, *rq_src; 4319 int ret = 0; 4320 4321 if (unlikely(!cpu_active(dest_cpu))) 4322 return ret; 4323 4324 rq_src = cpu_rq(src_cpu); 4325 rq_dest = cpu_rq(dest_cpu); 4326 4327 raw_spin_lock(&p->pi_lock); 4328 double_rq_lock(rq_src, rq_dest); 4329 /* Already moved. */ 4330 if (task_cpu(p) != src_cpu) 4331 goto done; 4332 /* Affinity changed (again). */ 4333 if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p))) 4334 goto fail; 4335 4336 /* 4337 * If we're not on a rq, the next wake-up will ensure we're 4338 * placed properly. 4339 */ 4340 if (p->on_rq) { 4341 dequeue_task(rq_src, p, 0); 4342 set_task_cpu(p, dest_cpu); 4343 enqueue_task(rq_dest, p, 0); 4344 check_preempt_curr(rq_dest, p, 0); 4345 } 4346done: 4347 ret = 1; 4348fail: 4349 double_rq_unlock(rq_src, rq_dest); 4350 raw_spin_unlock(&p->pi_lock); 4351 return ret; 4352} 4353 4354/* 4355 * migration_cpu_stop - this will be executed by a highprio stopper thread 4356 * and performs thread migration by bumping thread off CPU then 4357 * 'pushing' onto another runqueue. 4358 */ 4359static int migration_cpu_stop(void *data) 4360{ 4361 struct migration_arg *arg = data; 4362 4363 /* 4364 * The original target cpu might have gone down and we might 4365 * be on another cpu but it doesn't matter. 4366 */ 4367 local_irq_disable(); 4368 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu); 4369 local_irq_enable(); 4370 return 0; 4371} 4372 4373#ifdef CONFIG_HOTPLUG_CPU 4374 4375/* 4376 * Ensures that the idle task is using init_mm right before its cpu goes 4377 * offline. 4378 */ 4379void idle_task_exit(void) 4380{ 4381 struct mm_struct *mm = current->active_mm; 4382 4383 BUG_ON(cpu_online(smp_processor_id())); 4384 4385 if (mm != &init_mm) 4386 switch_mm(mm, &init_mm, current); 4387 mmdrop(mm); 4388} 4389 4390/* 4391 * Since this CPU is going 'away' for a while, fold any nr_active delta 4392 * we might have. Assumes we're called after migrate_tasks() so that the 4393 * nr_active count is stable. 4394 * 4395 * Also see the comment "Global load-average calculations". 4396 */ 4397static void calc_load_migrate(struct rq *rq) 4398{ 4399 long delta = calc_load_fold_active(rq); 4400 if (delta) 4401 atomic_long_add(delta, &calc_load_tasks); 4402} 4403 4404/* 4405 * Migrate all tasks from the rq, sleeping tasks will be migrated by 4406 * try_to_wake_up()->select_task_rq(). 4407 * 4408 * Called with rq->lock held even though we'er in stop_machine() and 4409 * there's no concurrency possible, we hold the required locks anyway 4410 * because of lock validation efforts. 4411 */ 4412static void migrate_tasks(unsigned int dead_cpu) 4413{ 4414 struct rq *rq = cpu_rq(dead_cpu); 4415 struct task_struct *next, *stop = rq->stop; 4416 int dest_cpu; 4417 4418 /* 4419 * Fudge the rq selection such that the below task selection loop 4420 * doesn't get stuck on the currently eligible stop task. 4421 * 4422 * We're currently inside stop_machine() and the rq is either stuck 4423 * in the stop_machine_cpu_stop() loop, or we're executing this code, 4424 * either way we should never end up calling schedule() until we're 4425 * done here. 4426 */ 4427 rq->stop = NULL; 4428 4429 /* 4430 * put_prev_task() and pick_next_task() sched 4431 * class method both need to have an up-to-date 4432 * value of rq->clock[_task] 4433 */ 4434 update_rq_clock(rq); 4435 4436 for ( ; ; ) { 4437 /* 4438 * There's this thread running, bail when that's the only 4439 * remaining thread. 4440 */ 4441 if (rq->nr_running == 1) 4442 break; 4443 4444 next = pick_next_task(rq); 4445 BUG_ON(!next); 4446 next->sched_class->put_prev_task(rq, next); 4447 4448 /* Find suitable destination for @next, with force if needed. */ 4449 dest_cpu = select_fallback_rq(dead_cpu, next); 4450 raw_spin_unlock(&rq->lock); 4451 4452 __migrate_task(next, dead_cpu, dest_cpu); 4453 4454 raw_spin_lock(&rq->lock); 4455 } 4456 4457 rq->stop = stop; 4458} 4459 4460#endif /* CONFIG_HOTPLUG_CPU */ 4461 4462#if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL) 4463 4464static struct ctl_table sd_ctl_dir[] = { 4465 { 4466 .procname = "sched_domain", 4467 .mode = 0555, 4468 }, 4469 {} 4470}; 4471 4472static struct ctl_table sd_ctl_root[] = { 4473 { 4474 .procname = "kernel", 4475 .mode = 0555, 4476 .child = sd_ctl_dir, 4477 }, 4478 {} 4479}; 4480 4481static struct ctl_table *sd_alloc_ctl_entry(int n) 4482{ 4483 struct ctl_table *entry = 4484 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL); 4485 4486 return entry; 4487} 4488 4489static void sd_free_ctl_entry(struct ctl_table **tablep) 4490{ 4491 struct ctl_table *entry; 4492 4493 /* 4494 * In the intermediate directories, both the child directory and 4495 * procname are dynamically allocated and could fail but the mode 4496 * will always be set. In the lowest directory the names are 4497 * static strings and all have proc handlers. 4498 */ 4499 for (entry = *tablep; entry->mode; entry++) { 4500 if (entry->child) 4501 sd_free_ctl_entry(&entry->child); 4502 if (entry->proc_handler == NULL) 4503 kfree(entry->procname); 4504 } 4505 4506 kfree(*tablep); 4507 *tablep = NULL; 4508} 4509 4510static int min_load_idx = 0; 4511static int max_load_idx = CPU_LOAD_IDX_MAX-1; 4512 4513static void 4514set_table_entry(struct ctl_table *entry, 4515 const char *procname, void *data, int maxlen, 4516 umode_t mode, proc_handler *proc_handler, 4517 bool load_idx) 4518{ 4519 entry->procname = procname; 4520 entry->data = data; 4521 entry->maxlen = maxlen; 4522 entry->mode = mode; 4523 entry->proc_handler = proc_handler; 4524 4525 if (load_idx) { 4526 entry->extra1 = &min_load_idx; 4527 entry->extra2 = &max_load_idx; 4528 } 4529} 4530 4531static struct ctl_table * 4532sd_alloc_ctl_domain_table(struct sched_domain *sd) 4533{ 4534 struct ctl_table *table = sd_alloc_ctl_entry(13); 4535 4536 if (table == NULL) 4537 return NULL; 4538 4539 set_table_entry(&table[0], "min_interval", &sd->min_interval, 4540 sizeof(long), 0644, proc_doulongvec_minmax, false); 4541 set_table_entry(&table[1], "max_interval", &sd->max_interval, 4542 sizeof(long), 0644, proc_doulongvec_minmax, false); 4543 set_table_entry(&table[2], "busy_idx", &sd->busy_idx, 4544 sizeof(int), 0644, proc_dointvec_minmax, true); 4545 set_table_entry(&table[3], "idle_idx", &sd->idle_idx, 4546 sizeof(int), 0644, proc_dointvec_minmax, true); 4547 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx, 4548 sizeof(int), 0644, proc_dointvec_minmax, true); 4549 set_table_entry(&table[5], "wake_idx", &sd->wake_idx, 4550 sizeof(int), 0644, proc_dointvec_minmax, true); 4551 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx, 4552 sizeof(int), 0644, proc_dointvec_minmax, true); 4553 set_table_entry(&table[7], "busy_factor", &sd->busy_factor, 4554 sizeof(int), 0644, proc_dointvec_minmax, false); 4555 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct, 4556 sizeof(int), 0644, proc_dointvec_minmax, false); 4557 set_table_entry(&table[9], "cache_nice_tries", 4558 &sd->cache_nice_tries, 4559 sizeof(int), 0644, proc_dointvec_minmax, false); 4560 set_table_entry(&table[10], "flags", &sd->flags, 4561 sizeof(int), 0644, proc_dointvec_minmax, false); 4562 set_table_entry(&table[11], "name", sd->name, 4563 CORENAME_MAX_SIZE, 0444, proc_dostring, false); 4564 /* &table[12] is terminator */ 4565 4566 return table; 4567} 4568 4569static struct ctl_table *sd_alloc_ctl_cpu_table(int cpu) 4570{ 4571 struct ctl_table *entry, *table; 4572 struct sched_domain *sd; 4573 int domain_num = 0, i; 4574 char buf[32]; 4575 4576 for_each_domain(cpu, sd) 4577 domain_num++; 4578 entry = table = sd_alloc_ctl_entry(domain_num + 1); 4579 if (table == NULL) 4580 return NULL; 4581 4582 i = 0; 4583 for_each_domain(cpu, sd) { 4584 snprintf(buf, 32, "domain%d", i); 4585 entry->procname = kstrdup(buf, GFP_KERNEL); 4586 entry->mode = 0555; 4587 entry->child = sd_alloc_ctl_domain_table(sd); 4588 entry++; 4589 i++; 4590 } 4591 return table; 4592} 4593 4594static struct ctl_table_header *sd_sysctl_header; 4595static void register_sched_domain_sysctl(void) 4596{ 4597 int i, cpu_num = num_possible_cpus(); 4598 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1); 4599 char buf[32]; 4600 4601 WARN_ON(sd_ctl_dir[0].child); 4602 sd_ctl_dir[0].child = entry; 4603 4604 if (entry == NULL) 4605 return; 4606 4607 for_each_possible_cpu(i) { 4608 snprintf(buf, 32, "cpu%d", i); 4609 entry->procname = kstrdup(buf, GFP_KERNEL); 4610 entry->mode = 0555; 4611 entry->child = sd_alloc_ctl_cpu_table(i); 4612 entry++; 4613 } 4614 4615 WARN_ON(sd_sysctl_header); 4616 sd_sysctl_header = register_sysctl_table(sd_ctl_root); 4617} 4618 4619/* may be called multiple times per register */ 4620static void unregister_sched_domain_sysctl(void) 4621{ 4622 if (sd_sysctl_header) 4623 unregister_sysctl_table(sd_sysctl_header); 4624 sd_sysctl_header = NULL; 4625 if (sd_ctl_dir[0].child) 4626 sd_free_ctl_entry(&sd_ctl_dir[0].child); 4627} 4628#else 4629static void register_sched_domain_sysctl(void) 4630{ 4631} 4632static void unregister_sched_domain_sysctl(void) 4633{ 4634} 4635#endif 4636 4637static void set_rq_online(struct rq *rq) 4638{ 4639 if (!rq->online) { 4640 const struct sched_class *class; 4641 4642 cpumask_set_cpu(rq->cpu, rq->rd->online); 4643 rq->online = 1; 4644 4645 for_each_class(class) { 4646 if (class->rq_online) 4647 class->rq_online(rq); 4648 } 4649 } 4650} 4651 4652static void set_rq_offline(struct rq *rq) 4653{ 4654 if (rq->online) { 4655 const struct sched_class *class; 4656 4657 for_each_class(class) { 4658 if (class->rq_offline) 4659 class->rq_offline(rq); 4660 } 4661 4662 cpumask_clear_cpu(rq->cpu, rq->rd->online); 4663 rq->online = 0; 4664 } 4665} 4666 4667/* 4668 * migration_call - callback that gets triggered when a CPU is added. 4669 * Here we can start up the necessary migration thread for the new CPU. 4670 */ 4671static int __cpuinit 4672migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu) 4673{ 4674 int cpu = (long)hcpu; 4675 unsigned long flags; 4676 struct rq *rq = cpu_rq(cpu); 4677 4678 switch (action & ~CPU_TASKS_FROZEN) { 4679 4680 case CPU_UP_PREPARE: 4681 rq->calc_load_update = calc_load_update; 4682 break; 4683 4684 case CPU_ONLINE: 4685 /* Update our root-domain */ 4686 raw_spin_lock_irqsave(&rq->lock, flags); 4687 if (rq->rd) { 4688 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span)); 4689 4690 set_rq_online(rq); 4691 } 4692 raw_spin_unlock_irqrestore(&rq->lock, flags); 4693 break; 4694 4695#ifdef CONFIG_HOTPLUG_CPU 4696 case CPU_DYING: 4697 sched_ttwu_pending(); 4698 /* Update our root-domain */ 4699 raw_spin_lock_irqsave(&rq->lock, flags); 4700 if (rq->rd) { 4701 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span)); 4702 set_rq_offline(rq); 4703 } 4704 migrate_tasks(cpu); 4705 BUG_ON(rq->nr_running != 1); /* the migration thread */ 4706 raw_spin_unlock_irqrestore(&rq->lock, flags); 4707 break; 4708 4709 case CPU_DEAD: 4710 calc_load_migrate(rq); 4711 break; 4712#endif 4713 } 4714 4715 update_max_interval(); 4716 4717 return NOTIFY_OK; 4718} 4719 4720/* 4721 * Register at high priority so that task migration (migrate_all_tasks) 4722 * happens before everything else. This has to be lower priority than 4723 * the notifier in the perf_event subsystem, though. 4724 */ 4725static struct notifier_block __cpuinitdata migration_notifier = { 4726 .notifier_call = migration_call, 4727 .priority = CPU_PRI_MIGRATION, 4728}; 4729 4730static int __cpuinit sched_cpu_active(struct notifier_block *nfb, 4731 unsigned long action, void *hcpu) 4732{ 4733 switch (action & ~CPU_TASKS_FROZEN) { 4734 case CPU_STARTING: 4735 case CPU_DOWN_FAILED: 4736 set_cpu_active((long)hcpu, true); 4737 return NOTIFY_OK; 4738 default: 4739 return NOTIFY_DONE; 4740 } 4741} 4742 4743static int __cpuinit sched_cpu_inactive(struct notifier_block *nfb, 4744 unsigned long action, void *hcpu) 4745{ 4746 switch (action & ~CPU_TASKS_FROZEN) { 4747 case CPU_DOWN_PREPARE: 4748 set_cpu_active((long)hcpu, false); 4749 return NOTIFY_OK; 4750 default: 4751 return NOTIFY_DONE; 4752 } 4753} 4754 4755static int __init migration_init(void) 4756{ 4757 void *cpu = (void *)(long)smp_processor_id(); 4758 int err; 4759 4760 /* Initialize migration for the boot CPU */ 4761 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu); 4762 BUG_ON(err == NOTIFY_BAD); 4763 migration_call(&migration_notifier, CPU_ONLINE, cpu); 4764 register_cpu_notifier(&migration_notifier); 4765 4766 /* Register cpu active notifiers */ 4767 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE); 4768 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE); 4769 4770 return 0; 4771} 4772early_initcall(migration_init); 4773#endif 4774 4775#ifdef CONFIG_SMP 4776 4777static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */ 4778 4779#ifdef CONFIG_SCHED_DEBUG 4780 4781static __read_mostly int sched_debug_enabled; 4782 4783static int __init sched_debug_setup(char *str) 4784{ 4785 sched_debug_enabled = 1; 4786 4787 return 0; 4788} 4789early_param("sched_debug", sched_debug_setup); 4790 4791static inline bool sched_debug(void) 4792{ 4793 return sched_debug_enabled; 4794} 4795 4796static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level, 4797 struct cpumask *groupmask) 4798{ 4799 struct sched_group *group = sd->groups; 4800 char str[256]; 4801 4802 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd)); 4803 cpumask_clear(groupmask); 4804 4805 printk(KERN_DEBUG "%*s domain %d: ", level, "", level); 4806 4807 if (!(sd->flags & SD_LOAD_BALANCE)) { 4808 printk("does not load-balance\n"); 4809 if (sd->parent) 4810 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain" 4811 " has parent"); 4812 return -1; 4813 } 4814 4815 printk(KERN_CONT "span %s level %s\n", str, sd->name); 4816 4817 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) { 4818 printk(KERN_ERR "ERROR: domain->span does not contain " 4819 "CPU%d\n", cpu); 4820 } 4821 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) { 4822 printk(KERN_ERR "ERROR: domain->groups does not contain" 4823 " CPU%d\n", cpu); 4824 } 4825 4826 printk(KERN_DEBUG "%*s groups:", level + 1, ""); 4827 do { 4828 if (!group) { 4829 printk("\n"); 4830 printk(KERN_ERR "ERROR: group is NULL\n"); 4831 break; 4832 } 4833 4834 /* 4835 * Even though we initialize ->power to something semi-sane, 4836 * we leave power_orig unset. This allows us to detect if 4837 * domain iteration is still funny without causing /0 traps. 4838 */ 4839 if (!group->sgp->power_orig) { 4840 printk(KERN_CONT "\n"); 4841 printk(KERN_ERR "ERROR: domain->cpu_power not " 4842 "set\n"); 4843 break; 4844 } 4845 4846 if (!cpumask_weight(sched_group_cpus(group))) { 4847 printk(KERN_CONT "\n"); 4848 printk(KERN_ERR "ERROR: empty group\n"); 4849 break; 4850 } 4851 4852 if (!(sd->flags & SD_OVERLAP) && 4853 cpumask_intersects(groupmask, sched_group_cpus(group))) { 4854 printk(KERN_CONT "\n"); 4855 printk(KERN_ERR "ERROR: repeated CPUs\n"); 4856 break; 4857 } 4858 4859 cpumask_or(groupmask, groupmask, sched_group_cpus(group)); 4860 4861 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group)); 4862 4863 printk(KERN_CONT " %s", str); 4864 if (group->sgp->power != SCHED_POWER_SCALE) { 4865 printk(KERN_CONT " (cpu_power = %d)", 4866 group->sgp->power); 4867 } 4868 4869 group = group->next; 4870 } while (group != sd->groups); 4871 printk(KERN_CONT "\n"); 4872 4873 if (!cpumask_equal(sched_domain_span(sd), groupmask)) 4874 printk(KERN_ERR "ERROR: groups don't span domain->span\n"); 4875 4876 if (sd->parent && 4877 !cpumask_subset(groupmask, sched_domain_span(sd->parent))) 4878 printk(KERN_ERR "ERROR: parent span is not a superset " 4879 "of domain->span\n"); 4880 return 0; 4881} 4882 4883static void sched_domain_debug(struct sched_domain *sd, int cpu) 4884{ 4885 int level = 0; 4886 4887 if (!sched_debug_enabled) 4888 return; 4889 4890 if (!sd) { 4891 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu); 4892 return; 4893 } 4894 4895 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu); 4896 4897 for (;;) { 4898 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask)) 4899 break; 4900 level++; 4901 sd = sd->parent; 4902 if (!sd) 4903 break; 4904 } 4905} 4906#else /* !CONFIG_SCHED_DEBUG */ 4907# define sched_domain_debug(sd, cpu) do { } while (0) 4908static inline bool sched_debug(void) 4909{ 4910 return false; 4911} 4912#endif /* CONFIG_SCHED_DEBUG */ 4913 4914static int sd_degenerate(struct sched_domain *sd) 4915{ 4916 if (cpumask_weight(sched_domain_span(sd)) == 1) 4917 return 1; 4918 4919 /* Following flags need at least 2 groups */ 4920 if (sd->flags & (SD_LOAD_BALANCE | 4921 SD_BALANCE_NEWIDLE | 4922 SD_BALANCE_FORK | 4923 SD_BALANCE_EXEC | 4924 SD_SHARE_CPUPOWER | 4925 SD_SHARE_PKG_RESOURCES)) { 4926 if (sd->groups != sd->groups->next) 4927 return 0; 4928 } 4929 4930 /* Following flags don't use groups */ 4931 if (sd->flags & (SD_WAKE_AFFINE)) 4932 return 0; 4933 4934 return 1; 4935} 4936 4937static int 4938sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent) 4939{ 4940 unsigned long cflags = sd->flags, pflags = parent->flags; 4941 4942 if (sd_degenerate(parent)) 4943 return 1; 4944 4945 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent))) 4946 return 0; 4947 4948 /* Flags needing groups don't count if only 1 group in parent */ 4949 if (parent->groups == parent->groups->next) { 4950 pflags &= ~(SD_LOAD_BALANCE | 4951 SD_BALANCE_NEWIDLE | 4952 SD_BALANCE_FORK | 4953 SD_BALANCE_EXEC | 4954 SD_SHARE_CPUPOWER | 4955 SD_SHARE_PKG_RESOURCES); 4956 if (nr_node_ids == 1) 4957 pflags &= ~SD_SERIALIZE; 4958 } 4959 if (~cflags & pflags) 4960 return 0; 4961 4962 return 1; 4963} 4964 4965static void free_rootdomain(struct rcu_head *rcu) 4966{ 4967 struct root_domain *rd = container_of(rcu, struct root_domain, rcu); 4968 4969 cpupri_cleanup(&rd->cpupri); 4970 free_cpumask_var(rd->rto_mask); 4971 free_cpumask_var(rd->online); 4972 free_cpumask_var(rd->span); 4973 kfree(rd); 4974} 4975 4976static void rq_attach_root(struct rq *rq, struct root_domain *rd) 4977{ 4978 struct root_domain *old_rd = NULL; 4979 unsigned long flags; 4980 4981 raw_spin_lock_irqsave(&rq->lock, flags); 4982 4983 if (rq->rd) { 4984 old_rd = rq->rd; 4985 4986 if (cpumask_test_cpu(rq->cpu, old_rd->online)) 4987 set_rq_offline(rq); 4988 4989 cpumask_clear_cpu(rq->cpu, old_rd->span); 4990 4991 /* 4992 * If we dont want to free the old_rt yet then 4993 * set old_rd to NULL to skip the freeing later 4994 * in this function: 4995 */ 4996 if (!atomic_dec_and_test(&old_rd->refcount)) 4997 old_rd = NULL; 4998 } 4999 5000 atomic_inc(&rd->refcount); 5001 rq->rd = rd; 5002 5003 cpumask_set_cpu(rq->cpu, rd->span); 5004 if (cpumask_test_cpu(rq->cpu, cpu_active_mask)) 5005 set_rq_online(rq); 5006 5007 raw_spin_unlock_irqrestore(&rq->lock, flags); 5008 5009 if (old_rd) 5010 call_rcu_sched(&old_rd->rcu, free_rootdomain); 5011} 5012 5013static int init_rootdomain(struct root_domain *rd) 5014{ 5015 memset(rd, 0, sizeof(*rd)); 5016 5017 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL)) 5018 goto out; 5019 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL)) 5020 goto free_span; 5021 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL)) 5022 goto free_online; 5023 5024 if (cpupri_init(&rd->cpupri) != 0) 5025 goto free_rto_mask; 5026 return 0; 5027 5028free_rto_mask: 5029 free_cpumask_var(rd->rto_mask); 5030free_online: 5031 free_cpumask_var(rd->online); 5032free_span: 5033 free_cpumask_var(rd->span); 5034out: 5035 return -ENOMEM; 5036} 5037 5038/* 5039 * By default the system creates a single root-domain with all cpus as 5040 * members (mimicking the global state we have today). 5041 */ 5042struct root_domain def_root_domain; 5043 5044static void init_defrootdomain(void) 5045{ 5046 init_rootdomain(&def_root_domain); 5047 5048 atomic_set(&def_root_domain.refcount, 1); 5049} 5050 5051static struct root_domain *alloc_rootdomain(void) 5052{ 5053 struct root_domain *rd; 5054 5055 rd = kmalloc(sizeof(*rd), GFP_KERNEL); 5056 if (!rd) 5057 return NULL; 5058 5059 if (init_rootdomain(rd) != 0) { 5060 kfree(rd); 5061 return NULL; 5062 } 5063 5064 return rd; 5065} 5066 5067static void free_sched_groups(struct sched_group *sg, int free_sgp) 5068{ 5069 struct sched_group *tmp, *first; 5070 5071 if (!sg) 5072 return; 5073 5074 first = sg; 5075 do { 5076 tmp = sg->next; 5077 5078 if (free_sgp && atomic_dec_and_test(&sg->sgp->ref)) 5079 kfree(sg->sgp); 5080 5081 kfree(sg); 5082 sg = tmp; 5083 } while (sg != first); 5084} 5085 5086static void free_sched_domain(struct rcu_head *rcu) 5087{ 5088 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu); 5089 5090 /* 5091 * If its an overlapping domain it has private groups, iterate and 5092 * nuke them all. 5093 */ 5094 if (sd->flags & SD_OVERLAP) { 5095 free_sched_groups(sd->groups, 1); 5096 } else if (atomic_dec_and_test(&sd->groups->ref)) { 5097 kfree(sd->groups->sgp); 5098 kfree(sd->groups); 5099 } 5100 kfree(sd); 5101} 5102 5103static void destroy_sched_domain(struct sched_domain *sd, int cpu) 5104{ 5105 call_rcu(&sd->rcu, free_sched_domain); 5106} 5107 5108static void destroy_sched_domains(struct sched_domain *sd, int cpu) 5109{ 5110 for (; sd; sd = sd->parent) 5111 destroy_sched_domain(sd, cpu); 5112} 5113 5114/* 5115 * Keep a special pointer to the highest sched_domain that has 5116 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this 5117 * allows us to avoid some pointer chasing select_idle_sibling(). 5118 * 5119 * Also keep a unique ID per domain (we use the first cpu number in 5120 * the cpumask of the domain), this allows us to quickly tell if 5121 * two cpus are in the same cache domain, see cpus_share_cache(). 5122 */ 5123DEFINE_PER_CPU(struct sched_domain *, sd_llc); 5124DEFINE_PER_CPU(int, sd_llc_id); 5125 5126static void update_top_cache_domain(int cpu) 5127{ 5128 struct sched_domain *sd; 5129 int id = cpu; 5130 5131 sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES); 5132 if (sd) 5133 id = cpumask_first(sched_domain_span(sd)); 5134 5135 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd); 5136 per_cpu(sd_llc_id, cpu) = id; 5137} 5138 5139/* 5140 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must 5141 * hold the hotplug lock. 5142 */ 5143static void 5144cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu) 5145{ 5146 struct rq *rq = cpu_rq(cpu); 5147 struct sched_domain *tmp; 5148 5149 /* Remove the sched domains which do not contribute to scheduling. */ 5150 for (tmp = sd; tmp; ) { 5151 struct sched_domain *parent = tmp->parent; 5152 if (!parent) 5153 break; 5154 5155 if (sd_parent_degenerate(tmp, parent)) { 5156 tmp->parent = parent->parent; 5157 if (parent->parent) 5158 parent->parent->child = tmp; 5159 destroy_sched_domain(parent, cpu); 5160 } else 5161 tmp = tmp->parent; 5162 } 5163 5164 if (sd && sd_degenerate(sd)) { 5165 tmp = sd; 5166 sd = sd->parent; 5167 destroy_sched_domain(tmp, cpu); 5168 if (sd) 5169 sd->child = NULL; 5170 } 5171 5172 sched_domain_debug(sd, cpu); 5173 5174 rq_attach_root(rq, rd); 5175 tmp = rq->sd; 5176 rcu_assign_pointer(rq->sd, sd); 5177 destroy_sched_domains(tmp, cpu); 5178 5179 update_top_cache_domain(cpu); 5180} 5181 5182/* cpus with isolated domains */ 5183static cpumask_var_t cpu_isolated_map; 5184 5185/* Setup the mask of cpus configured for isolated domains */ 5186static int __init isolated_cpu_setup(char *str) 5187{ 5188 alloc_bootmem_cpumask_var(&cpu_isolated_map); 5189 cpulist_parse(str, cpu_isolated_map); 5190 return 1; 5191} 5192 5193__setup("isolcpus=", isolated_cpu_setup); 5194 5195static const struct cpumask *cpu_cpu_mask(int cpu) 5196{ 5197 return cpumask_of_node(cpu_to_node(cpu)); 5198} 5199 5200struct sd_data { 5201 struct sched_domain **__percpu sd; 5202 struct sched_group **__percpu sg; 5203 struct sched_group_power **__percpu sgp; 5204}; 5205 5206struct s_data { 5207 struct sched_domain ** __percpu sd; 5208 struct root_domain *rd; 5209}; 5210 5211enum s_alloc { 5212 sa_rootdomain, 5213 sa_sd, 5214 sa_sd_storage, 5215 sa_none, 5216}; 5217 5218struct sched_domain_topology_level; 5219 5220typedef struct sched_domain *(*sched_domain_init_f)(struct sched_domain_topology_level *tl, int cpu); 5221typedef const struct cpumask *(*sched_domain_mask_f)(int cpu); 5222 5223#define SDTL_OVERLAP 0x01 5224 5225struct sched_domain_topology_level { 5226 sched_domain_init_f init; 5227 sched_domain_mask_f mask; 5228 int flags; 5229 int numa_level; 5230 struct sd_data data; 5231}; 5232 5233/* 5234 * Build an iteration mask that can exclude certain CPUs from the upwards 5235 * domain traversal. 5236 * 5237 * Asymmetric node setups can result in situations where the domain tree is of 5238 * unequal depth, make sure to skip domains that already cover the entire 5239 * range. 5240 * 5241 * In that case build_sched_domains() will have terminated the iteration early 5242 * and our sibling sd spans will be empty. Domains should always include the 5243 * cpu they're built on, so check that. 5244 * 5245 */ 5246static void build_group_mask(struct sched_domain *sd, struct sched_group *sg) 5247{ 5248 const struct cpumask *span = sched_domain_span(sd); 5249 struct sd_data *sdd = sd->private; 5250 struct sched_domain *sibling; 5251 int i; 5252 5253 for_each_cpu(i, span) { 5254 sibling = *per_cpu_ptr(sdd->sd, i); 5255 if (!cpumask_test_cpu(i, sched_domain_span(sibling))) 5256 continue; 5257 5258 cpumask_set_cpu(i, sched_group_mask(sg)); 5259 } 5260} 5261 5262/* 5263 * Return the canonical balance cpu for this group, this is the first cpu 5264 * of this group that's also in the iteration mask. 5265 */ 5266int group_balance_cpu(struct sched_group *sg) 5267{ 5268 return cpumask_first_and(sched_group_cpus(sg), sched_group_mask(sg)); 5269} 5270 5271static int 5272build_overlap_sched_groups(struct sched_domain *sd, int cpu) 5273{ 5274 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg; 5275 const struct cpumask *span = sched_domain_span(sd); 5276 struct cpumask *covered = sched_domains_tmpmask; 5277 struct sd_data *sdd = sd->private; 5278 struct sched_domain *child; 5279 int i; 5280 5281 cpumask_clear(covered); 5282 5283 for_each_cpu(i, span) { 5284 struct cpumask *sg_span; 5285 5286 if (cpumask_test_cpu(i, covered)) 5287 continue; 5288 5289 child = *per_cpu_ptr(sdd->sd, i); 5290 5291 /* See the comment near build_group_mask(). */ 5292 if (!cpumask_test_cpu(i, sched_domain_span(child))) 5293 continue; 5294 5295 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(), 5296 GFP_KERNEL, cpu_to_node(cpu)); 5297 5298 if (!sg) 5299 goto fail; 5300 5301 sg_span = sched_group_cpus(sg); 5302 if (child->child) { 5303 child = child->child; 5304 cpumask_copy(sg_span, sched_domain_span(child)); 5305 } else 5306 cpumask_set_cpu(i, sg_span); 5307 5308 cpumask_or(covered, covered, sg_span); 5309 5310 sg->sgp = *per_cpu_ptr(sdd->sgp, i); 5311 if (atomic_inc_return(&sg->sgp->ref) == 1) 5312 build_group_mask(sd, sg); 5313 5314 /* 5315 * Initialize sgp->power such that even if we mess up the 5316 * domains and no possible iteration will get us here, we won't 5317 * die on a /0 trap. 5318 */ 5319 sg->sgp->power = SCHED_POWER_SCALE * cpumask_weight(sg_span); 5320 5321 /* 5322 * Make sure the first group of this domain contains the 5323 * canonical balance cpu. Otherwise the sched_domain iteration 5324 * breaks. See update_sg_lb_stats(). 5325 */ 5326 if ((!groups && cpumask_test_cpu(cpu, sg_span)) || 5327 group_balance_cpu(sg) == cpu) 5328 groups = sg; 5329 5330 if (!first) 5331 first = sg; 5332 if (last) 5333 last->next = sg; 5334 last = sg; 5335 last->next = first; 5336 } 5337 sd->groups = groups; 5338 5339 return 0; 5340 5341fail: 5342 free_sched_groups(first, 0); 5343 5344 return -ENOMEM; 5345} 5346 5347static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg) 5348{ 5349 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu); 5350 struct sched_domain *child = sd->child; 5351 5352 if (child) 5353 cpu = cpumask_first(sched_domain_span(child)); 5354 5355 if (sg) { 5356 *sg = *per_cpu_ptr(sdd->sg, cpu); 5357 (*sg)->sgp = *per_cpu_ptr(sdd->sgp, cpu); 5358 atomic_set(&(*sg)->sgp->ref, 1); /* for claim_allocations */ 5359 } 5360 5361 return cpu; 5362} 5363 5364/* 5365 * build_sched_groups will build a circular linked list of the groups 5366 * covered by the given span, and will set each group's ->cpumask correctly, 5367 * and ->cpu_power to 0. 5368 * 5369 * Assumes the sched_domain tree is fully constructed 5370 */ 5371static int 5372build_sched_groups(struct sched_domain *sd, int cpu) 5373{ 5374 struct sched_group *first = NULL, *last = NULL; 5375 struct sd_data *sdd = sd->private; 5376 const struct cpumask *span = sched_domain_span(sd); 5377 struct cpumask *covered; 5378 int i; 5379 5380 get_group(cpu, sdd, &sd->groups); 5381 atomic_inc(&sd->groups->ref); 5382 5383 if (cpu != cpumask_first(span)) 5384 return 0; 5385 5386 lockdep_assert_held(&sched_domains_mutex); 5387 covered = sched_domains_tmpmask; 5388 5389 cpumask_clear(covered); 5390 5391 for_each_cpu(i, span) { 5392 struct sched_group *sg; 5393 int group, j; 5394 5395 if (cpumask_test_cpu(i, covered)) 5396 continue; 5397 5398 group = get_group(i, sdd, &sg); 5399 cpumask_clear(sched_group_cpus(sg)); 5400 sg->sgp->power = 0; 5401 cpumask_setall(sched_group_mask(sg)); 5402 5403 for_each_cpu(j, span) { 5404 if (get_group(j, sdd, NULL) != group) 5405 continue; 5406 5407 cpumask_set_cpu(j, covered); 5408 cpumask_set_cpu(j, sched_group_cpus(sg)); 5409 } 5410 5411 if (!first) 5412 first = sg; 5413 if (last) 5414 last->next = sg; 5415 last = sg; 5416 } 5417 last->next = first; 5418 5419 return 0; 5420} 5421 5422/* 5423 * Initialize sched groups cpu_power. 5424 * 5425 * cpu_power indicates the capacity of sched group, which is used while 5426 * distributing the load between different sched groups in a sched domain. 5427 * Typically cpu_power for all the groups in a sched domain will be same unless 5428 * there are asymmetries in the topology. If there are asymmetries, group 5429 * having more cpu_power will pickup more load compared to the group having 5430 * less cpu_power. 5431 */ 5432static void init_sched_groups_power(int cpu, struct sched_domain *sd) 5433{ 5434 struct sched_group *sg = sd->groups; 5435 5436 WARN_ON(!sg); 5437 5438 do { 5439 sg->group_weight = cpumask_weight(sched_group_cpus(sg)); 5440 sg = sg->next; 5441 } while (sg != sd->groups); 5442 5443 if (cpu != group_balance_cpu(sg)) 5444 return; 5445 5446 update_group_power(sd, cpu); 5447 atomic_set(&sg->sgp->nr_busy_cpus, sg->group_weight); 5448} 5449 5450int __weak arch_sd_sibling_asym_packing(void) 5451{ 5452 return 0*SD_ASYM_PACKING; 5453} 5454 5455/* 5456 * Initializers for schedule domains 5457 * Non-inlined to reduce accumulated stack pressure in build_sched_domains() 5458 */ 5459 5460#ifdef CONFIG_SCHED_DEBUG 5461# define SD_INIT_NAME(sd, type) sd->name = #type 5462#else 5463# define SD_INIT_NAME(sd, type) do { } while (0) 5464#endif 5465 5466#define SD_INIT_FUNC(type) \ 5467static noinline struct sched_domain * \ 5468sd_init_##type(struct sched_domain_topology_level *tl, int cpu) \ 5469{ \ 5470 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); \ 5471 *sd = SD_##type##_INIT; \ 5472 SD_INIT_NAME(sd, type); \ 5473 sd->private = &tl->data; \ 5474 return sd; \ 5475} 5476 5477SD_INIT_FUNC(CPU) 5478#ifdef CONFIG_SCHED_SMT 5479 SD_INIT_FUNC(SIBLING) 5480#endif 5481#ifdef CONFIG_SCHED_MC 5482 SD_INIT_FUNC(MC) 5483#endif 5484#ifdef CONFIG_SCHED_BOOK 5485 SD_INIT_FUNC(BOOK) 5486#endif 5487 5488static int default_relax_domain_level = -1; 5489int sched_domain_level_max; 5490 5491static int __init setup_relax_domain_level(char *str) 5492{ 5493 if (kstrtoint(str, 0, &default_relax_domain_level)) 5494 pr_warn("Unable to set relax_domain_level\n"); 5495 5496 return 1; 5497} 5498__setup("relax_domain_level=", setup_relax_domain_level); 5499 5500static void set_domain_attribute(struct sched_domain *sd, 5501 struct sched_domain_attr *attr) 5502{ 5503 int request; 5504 5505 if (!attr || attr->relax_domain_level < 0) { 5506 if (default_relax_domain_level < 0) 5507 return; 5508 else 5509 request = default_relax_domain_level; 5510 } else 5511 request = attr->relax_domain_level; 5512 if (request < sd->level) { 5513 /* turn off idle balance on this domain */ 5514 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE); 5515 } else { 5516 /* turn on idle balance on this domain */ 5517 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE); 5518 } 5519} 5520 5521static void __sdt_free(const struct cpumask *cpu_map); 5522static int __sdt_alloc(const struct cpumask *cpu_map); 5523 5524static void __free_domain_allocs(struct s_data *d, enum s_alloc what, 5525 const struct cpumask *cpu_map) 5526{ 5527 switch (what) { 5528 case sa_rootdomain: 5529 if (!atomic_read(&d->rd->refcount)) 5530 free_rootdomain(&d->rd->rcu); /* fall through */ 5531 case sa_sd: 5532 free_percpu(d->sd); /* fall through */ 5533 case sa_sd_storage: 5534 __sdt_free(cpu_map); /* fall through */ 5535 case sa_none: 5536 break; 5537 } 5538} 5539 5540static enum s_alloc __visit_domain_allocation_hell(struct s_data *d, 5541 const struct cpumask *cpu_map) 5542{ 5543 memset(d, 0, sizeof(*d)); 5544 5545 if (__sdt_alloc(cpu_map)) 5546 return sa_sd_storage; 5547 d->sd = alloc_percpu(struct sched_domain *); 5548 if (!d->sd) 5549 return sa_sd_storage; 5550 d->rd = alloc_rootdomain(); 5551 if (!d->rd) 5552 return sa_sd; 5553 return sa_rootdomain; 5554} 5555 5556/* 5557 * NULL the sd_data elements we've used to build the sched_domain and 5558 * sched_group structure so that the subsequent __free_domain_allocs() 5559 * will not free the data we're using. 5560 */ 5561static void claim_allocations(int cpu, struct sched_domain *sd) 5562{ 5563 struct sd_data *sdd = sd->private; 5564 5565 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd); 5566 *per_cpu_ptr(sdd->sd, cpu) = NULL; 5567 5568 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref)) 5569 *per_cpu_ptr(sdd->sg, cpu) = NULL; 5570 5571 if (atomic_read(&(*per_cpu_ptr(sdd->sgp, cpu))->ref)) 5572 *per_cpu_ptr(sdd->sgp, cpu) = NULL; 5573} 5574 5575#ifdef CONFIG_SCHED_SMT 5576static const struct cpumask *cpu_smt_mask(int cpu) 5577{ 5578 return topology_thread_cpumask(cpu); 5579} 5580#endif 5581 5582/* 5583 * Topology list, bottom-up. 5584 */ 5585static struct sched_domain_topology_level default_topology[] = { 5586#ifdef CONFIG_SCHED_SMT 5587 { sd_init_SIBLING, cpu_smt_mask, }, 5588#endif 5589#ifdef CONFIG_SCHED_MC 5590 { sd_init_MC, cpu_coregroup_mask, }, 5591#endif 5592#ifdef CONFIG_SCHED_BOOK 5593 { sd_init_BOOK, cpu_book_mask, }, 5594#endif 5595 { sd_init_CPU, cpu_cpu_mask, }, 5596 { NULL, }, 5597}; 5598 5599static struct sched_domain_topology_level *sched_domain_topology = default_topology; 5600 5601#define for_each_sd_topology(tl) \ 5602 for (tl = sched_domain_topology; tl->init; tl++) 5603 5604#ifdef CONFIG_NUMA 5605 5606static int sched_domains_numa_levels; 5607static int *sched_domains_numa_distance; 5608static struct cpumask ***sched_domains_numa_masks; 5609static int sched_domains_curr_level; 5610 5611static inline int sd_local_flags(int level) 5612{ 5613 if (sched_domains_numa_distance[level] > RECLAIM_DISTANCE) 5614 return 0; 5615 5616 return SD_BALANCE_EXEC | SD_BALANCE_FORK | SD_WAKE_AFFINE; 5617} 5618 5619static struct sched_domain * 5620sd_numa_init(struct sched_domain_topology_level *tl, int cpu) 5621{ 5622 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); 5623 int level = tl->numa_level; 5624 int sd_weight = cpumask_weight( 5625 sched_domains_numa_masks[level][cpu_to_node(cpu)]); 5626 5627 *sd = (struct sched_domain){ 5628 .min_interval = sd_weight, 5629 .max_interval = 2*sd_weight, 5630 .busy_factor = 32, 5631 .imbalance_pct = 125, 5632 .cache_nice_tries = 2, 5633 .busy_idx = 3, 5634 .idle_idx = 2, 5635 .newidle_idx = 0, 5636 .wake_idx = 0, 5637 .forkexec_idx = 0, 5638 5639 .flags = 1*SD_LOAD_BALANCE 5640 | 1*SD_BALANCE_NEWIDLE 5641 | 0*SD_BALANCE_EXEC 5642 | 0*SD_BALANCE_FORK 5643 | 0*SD_BALANCE_WAKE 5644 | 0*SD_WAKE_AFFINE 5645 | 0*SD_SHARE_CPUPOWER 5646 | 0*SD_SHARE_PKG_RESOURCES 5647 | 1*SD_SERIALIZE 5648 | 0*SD_PREFER_SIBLING 5649 | sd_local_flags(level) 5650 , 5651 .last_balance = jiffies, 5652 .balance_interval = sd_weight, 5653 }; 5654 SD_INIT_NAME(sd, NUMA); 5655 sd->private = &tl->data; 5656 5657 /* 5658 * Ugly hack to pass state to sd_numa_mask()... 5659 */ 5660 sched_domains_curr_level = tl->numa_level; 5661 5662 return sd; 5663} 5664 5665static const struct cpumask *sd_numa_mask(int cpu) 5666{ 5667 return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)]; 5668} 5669 5670static void sched_numa_warn(const char *str) 5671{ 5672 static int done = false; 5673 int i,j; 5674 5675 if (done) 5676 return; 5677 5678 done = true; 5679 5680 printk(KERN_WARNING "ERROR: %s\n\n", str); 5681 5682 for (i = 0; i < nr_node_ids; i++) { 5683 printk(KERN_WARNING " "); 5684 for (j = 0; j < nr_node_ids; j++) 5685 printk(KERN_CONT "%02d ", node_distance(i,j)); 5686 printk(KERN_CONT "\n"); 5687 } 5688 printk(KERN_WARNING "\n"); 5689} 5690 5691static bool find_numa_distance(int distance) 5692{ 5693 int i; 5694 5695 if (distance == node_distance(0, 0)) 5696 return true; 5697 5698 for (i = 0; i < sched_domains_numa_levels; i++) { 5699 if (sched_domains_numa_distance[i] == distance) 5700 return true; 5701 } 5702 5703 return false; 5704} 5705 5706static void sched_init_numa(void) 5707{ 5708 int next_distance, curr_distance = node_distance(0, 0); 5709 struct sched_domain_topology_level *tl; 5710 int level = 0; 5711 int i, j, k; 5712 5713 sched_domains_numa_distance = kzalloc(sizeof(int) * nr_node_ids, GFP_KERNEL); 5714 if (!sched_domains_numa_distance) 5715 return; 5716 5717 /* 5718 * O(nr_nodes^2) deduplicating selection sort -- in order to find the 5719 * unique distances in the node_distance() table. 5720 * 5721 * Assumes node_distance(0,j) includes all distances in 5722 * node_distance(i,j) in order to avoid cubic time. 5723 */ 5724 next_distance = curr_distance; 5725 for (i = 0; i < nr_node_ids; i++) { 5726 for (j = 0; j < nr_node_ids; j++) { 5727 for (k = 0; k < nr_node_ids; k++) { 5728 int distance = node_distance(i, k); 5729 5730 if (distance > curr_distance && 5731 (distance < next_distance || 5732 next_distance == curr_distance)) 5733 next_distance = distance; 5734 5735 /* 5736 * While not a strong assumption it would be nice to know 5737 * about cases where if node A is connected to B, B is not 5738 * equally connected to A. 5739 */ 5740 if (sched_debug() && node_distance(k, i) != distance) 5741 sched_numa_warn("Node-distance not symmetric"); 5742 5743 if (sched_debug() && i && !find_numa_distance(distance)) 5744 sched_numa_warn("Node-0 not representative"); 5745 } 5746 if (next_distance != curr_distance) { 5747 sched_domains_numa_distance[level++] = next_distance; 5748 sched_domains_numa_levels = level; 5749 curr_distance = next_distance; 5750 } else break; 5751 } 5752 5753 /* 5754 * In case of sched_debug() we verify the above assumption. 5755 */ 5756 if (!sched_debug()) 5757 break; 5758 } 5759 /* 5760 * 'level' contains the number of unique distances, excluding the 5761 * identity distance node_distance(i,i). 5762 * 5763 * The sched_domains_numa_distance[] array includes the actual distance 5764 * numbers. 5765 */ 5766 5767 /* 5768 * Here, we should temporarily reset sched_domains_numa_levels to 0. 5769 * If it fails to allocate memory for array sched_domains_numa_masks[][], 5770 * the array will contain less then 'level' members. This could be 5771 * dangerous when we use it to iterate array sched_domains_numa_masks[][] 5772 * in other functions. 5773 * 5774 * We reset it to 'level' at the end of this function. 5775 */ 5776 sched_domains_numa_levels = 0; 5777 5778 sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL); 5779 if (!sched_domains_numa_masks) 5780 return; 5781 5782 /* 5783 * Now for each level, construct a mask per node which contains all 5784 * cpus of nodes that are that many hops away from us. 5785 */ 5786 for (i = 0; i < level; i++) { 5787 sched_domains_numa_masks[i] = 5788 kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL); 5789 if (!sched_domains_numa_masks[i]) 5790 return; 5791 5792 for (j = 0; j < nr_node_ids; j++) { 5793 struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL); 5794 if (!mask) 5795 return; 5796 5797 sched_domains_numa_masks[i][j] = mask; 5798 5799 for (k = 0; k < nr_node_ids; k++) { 5800 if (node_distance(j, k) > sched_domains_numa_distance[i]) 5801 continue; 5802 5803 cpumask_or(mask, mask, cpumask_of_node(k)); 5804 } 5805 } 5806 } 5807 5808 tl = kzalloc((ARRAY_SIZE(default_topology) + level) * 5809 sizeof(struct sched_domain_topology_level), GFP_KERNEL); 5810 if (!tl) 5811 return; 5812 5813 /* 5814 * Copy the default topology bits.. 5815 */ 5816 for (i = 0; default_topology[i].init; i++) 5817 tl[i] = default_topology[i]; 5818 5819 /* 5820 * .. and append 'j' levels of NUMA goodness. 5821 */ 5822 for (j = 0; j < level; i++, j++) { 5823 tl[i] = (struct sched_domain_topology_level){ 5824 .init = sd_numa_init, 5825 .mask = sd_numa_mask, 5826 .flags = SDTL_OVERLAP, 5827 .numa_level = j, 5828 }; 5829 } 5830 5831 sched_domain_topology = tl; 5832 5833 sched_domains_numa_levels = level; 5834} 5835 5836static void sched_domains_numa_masks_set(int cpu) 5837{ 5838 int i, j; 5839 int node = cpu_to_node(cpu); 5840 5841 for (i = 0; i < sched_domains_numa_levels; i++) { 5842 for (j = 0; j < nr_node_ids; j++) { 5843 if (node_distance(j, node) <= sched_domains_numa_distance[i]) 5844 cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]); 5845 } 5846 } 5847} 5848 5849static void sched_domains_numa_masks_clear(int cpu) 5850{ 5851 int i, j; 5852 for (i = 0; i < sched_domains_numa_levels; i++) { 5853 for (j = 0; j < nr_node_ids; j++) 5854 cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]); 5855 } 5856} 5857 5858/* 5859 * Update sched_domains_numa_masks[level][node] array when new cpus 5860 * are onlined. 5861 */ 5862static int sched_domains_numa_masks_update(struct notifier_block *nfb, 5863 unsigned long action, 5864 void *hcpu) 5865{ 5866 int cpu = (long)hcpu; 5867 5868 switch (action & ~CPU_TASKS_FROZEN) { 5869 case CPU_ONLINE: 5870 sched_domains_numa_masks_set(cpu); 5871 break; 5872 5873 case CPU_DEAD: 5874 sched_domains_numa_masks_clear(cpu); 5875 break; 5876 5877 default: 5878 return NOTIFY_DONE; 5879 } 5880 5881 return NOTIFY_OK; 5882} 5883#else 5884static inline void sched_init_numa(void) 5885{ 5886} 5887 5888static int sched_domains_numa_masks_update(struct notifier_block *nfb, 5889 unsigned long action, 5890 void *hcpu) 5891{ 5892 return 0; 5893} 5894#endif /* CONFIG_NUMA */ 5895 5896static int __sdt_alloc(const struct cpumask *cpu_map) 5897{ 5898 struct sched_domain_topology_level *tl; 5899 int j; 5900 5901 for_each_sd_topology(tl) { 5902 struct sd_data *sdd = &tl->data; 5903 5904 sdd->sd = alloc_percpu(struct sched_domain *); 5905 if (!sdd->sd) 5906 return -ENOMEM; 5907 5908 sdd->sg = alloc_percpu(struct sched_group *); 5909 if (!sdd->sg) 5910 return -ENOMEM; 5911 5912 sdd->sgp = alloc_percpu(struct sched_group_power *); 5913 if (!sdd->sgp) 5914 return -ENOMEM; 5915 5916 for_each_cpu(j, cpu_map) { 5917 struct sched_domain *sd; 5918 struct sched_group *sg; 5919 struct sched_group_power *sgp; 5920 5921 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(), 5922 GFP_KERNEL, cpu_to_node(j)); 5923 if (!sd) 5924 return -ENOMEM; 5925 5926 *per_cpu_ptr(sdd->sd, j) = sd; 5927 5928 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(), 5929 GFP_KERNEL, cpu_to_node(j)); 5930 if (!sg) 5931 return -ENOMEM; 5932 5933 sg->next = sg; 5934 5935 *per_cpu_ptr(sdd->sg, j) = sg; 5936 5937 sgp = kzalloc_node(sizeof(struct sched_group_power) + cpumask_size(), 5938 GFP_KERNEL, cpu_to_node(j)); 5939 if (!sgp) 5940 return -ENOMEM; 5941 5942 *per_cpu_ptr(sdd->sgp, j) = sgp; 5943 } 5944 } 5945 5946 return 0; 5947} 5948 5949static void __sdt_free(const struct cpumask *cpu_map) 5950{ 5951 struct sched_domain_topology_level *tl; 5952 int j; 5953 5954 for_each_sd_topology(tl) { 5955 struct sd_data *sdd = &tl->data; 5956 5957 for_each_cpu(j, cpu_map) { 5958 struct sched_domain *sd; 5959 5960 if (sdd->sd) { 5961 sd = *per_cpu_ptr(sdd->sd, j); 5962 if (sd && (sd->flags & SD_OVERLAP)) 5963 free_sched_groups(sd->groups, 0); 5964 kfree(*per_cpu_ptr(sdd->sd, j)); 5965 } 5966 5967 if (sdd->sg) 5968 kfree(*per_cpu_ptr(sdd->sg, j)); 5969 if (sdd->sgp) 5970 kfree(*per_cpu_ptr(sdd->sgp, j)); 5971 } 5972 free_percpu(sdd->sd); 5973 sdd->sd = NULL; 5974 free_percpu(sdd->sg); 5975 sdd->sg = NULL; 5976 free_percpu(sdd->sgp); 5977 sdd->sgp = NULL; 5978 } 5979} 5980 5981struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl, 5982 const struct cpumask *cpu_map, struct sched_domain_attr *attr, 5983 struct sched_domain *child, int cpu) 5984{ 5985 struct sched_domain *sd = tl->init(tl, cpu); 5986 if (!sd) 5987 return child; 5988 5989 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu)); 5990 if (child) { 5991 sd->level = child->level + 1; 5992 sched_domain_level_max = max(sched_domain_level_max, sd->level); 5993 child->parent = sd; 5994 sd->child = child; 5995 } 5996 set_domain_attribute(sd, attr); 5997 5998 return sd; 5999} 6000 6001/* 6002 * Build sched domains for a given set of cpus and attach the sched domains 6003 * to the individual cpus 6004 */ 6005static int build_sched_domains(const struct cpumask *cpu_map, 6006 struct sched_domain_attr *attr) 6007{ 6008 enum s_alloc alloc_state; 6009 struct sched_domain *sd; 6010 struct s_data d; 6011 int i, ret = -ENOMEM; 6012 6013 alloc_state = __visit_domain_allocation_hell(&d, cpu_map); 6014 if (alloc_state != sa_rootdomain) 6015 goto error; 6016 6017 /* Set up domains for cpus specified by the cpu_map. */ 6018 for_each_cpu(i, cpu_map) { 6019 struct sched_domain_topology_level *tl; 6020 6021 sd = NULL; 6022 for_each_sd_topology(tl) { 6023 sd = build_sched_domain(tl, cpu_map, attr, sd, i); 6024 if (tl == sched_domain_topology) 6025 *per_cpu_ptr(d.sd, i) = sd; 6026 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP)) 6027 sd->flags |= SD_OVERLAP; 6028 if (cpumask_equal(cpu_map, sched_domain_span(sd))) 6029 break; 6030 } 6031 } 6032 6033 /* Build the groups for the domains */ 6034 for_each_cpu(i, cpu_map) { 6035 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) { 6036 sd->span_weight = cpumask_weight(sched_domain_span(sd)); 6037 if (sd->flags & SD_OVERLAP) { 6038 if (build_overlap_sched_groups(sd, i)) 6039 goto error; 6040 } else { 6041 if (build_sched_groups(sd, i)) 6042 goto error; 6043 } 6044 } 6045 } 6046 6047 /* Calculate CPU power for physical packages and nodes */ 6048 for (i = nr_cpumask_bits-1; i >= 0; i--) { 6049 if (!cpumask_test_cpu(i, cpu_map)) 6050 continue; 6051 6052 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) { 6053 claim_allocations(i, sd); 6054 init_sched_groups_power(i, sd); 6055 } 6056 } 6057 6058 /* Attach the domains */ 6059 rcu_read_lock(); 6060 for_each_cpu(i, cpu_map) { 6061 sd = *per_cpu_ptr(d.sd, i); 6062 cpu_attach_domain(sd, d.rd, i); 6063 } 6064 rcu_read_unlock(); 6065 6066 ret = 0; 6067error: 6068 __free_domain_allocs(&d, alloc_state, cpu_map); 6069 return ret; 6070} 6071 6072static cpumask_var_t *doms_cur; /* current sched domains */ 6073static int ndoms_cur; /* number of sched domains in 'doms_cur' */ 6074static struct sched_domain_attr *dattr_cur; 6075 /* attribues of custom domains in 'doms_cur' */ 6076 6077/* 6078 * Special case: If a kmalloc of a doms_cur partition (array of 6079 * cpumask) fails, then fallback to a single sched domain, 6080 * as determined by the single cpumask fallback_doms. 6081 */ 6082static cpumask_var_t fallback_doms; 6083 6084/* 6085 * arch_update_cpu_topology lets virtualized architectures update the 6086 * cpu core maps. It is supposed to return 1 if the topology changed 6087 * or 0 if it stayed the same. 6088 */ 6089int __attribute__((weak)) arch_update_cpu_topology(void) 6090{ 6091 return 0; 6092} 6093 6094cpumask_var_t *alloc_sched_domains(unsigned int ndoms) 6095{ 6096 int i; 6097 cpumask_var_t *doms; 6098 6099 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL); 6100 if (!doms) 6101 return NULL; 6102 for (i = 0; i < ndoms; i++) { 6103 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) { 6104 free_sched_domains(doms, i); 6105 return NULL; 6106 } 6107 } 6108 return doms; 6109} 6110 6111void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms) 6112{ 6113 unsigned int i; 6114 for (i = 0; i < ndoms; i++) 6115 free_cpumask_var(doms[i]); 6116 kfree(doms); 6117} 6118 6119/* 6120 * Set up scheduler domains and groups. Callers must hold the hotplug lock. 6121 * For now this just excludes isolated cpus, but could be used to 6122 * exclude other special cases in the future. 6123 */ 6124static int init_sched_domains(const struct cpumask *cpu_map) 6125{ 6126 int err; 6127 6128 arch_update_cpu_topology(); 6129 ndoms_cur = 1; 6130 doms_cur = alloc_sched_domains(ndoms_cur); 6131 if (!doms_cur) 6132 doms_cur = &fallback_doms; 6133 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map); 6134 err = build_sched_domains(doms_cur[0], NULL); 6135 register_sched_domain_sysctl(); 6136 6137 return err; 6138} 6139 6140/* 6141 * Detach sched domains from a group of cpus specified in cpu_map 6142 * These cpus will now be attached to the NULL domain 6143 */ 6144static void detach_destroy_domains(const struct cpumask *cpu_map) 6145{ 6146 int i; 6147 6148 rcu_read_lock(); 6149 for_each_cpu(i, cpu_map) 6150 cpu_attach_domain(NULL, &def_root_domain, i); 6151 rcu_read_unlock(); 6152} 6153 6154/* handle null as "default" */ 6155static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur, 6156 struct sched_domain_attr *new, int idx_new) 6157{ 6158 struct sched_domain_attr tmp; 6159 6160 /* fast path */ 6161 if (!new && !cur) 6162 return 1; 6163 6164 tmp = SD_ATTR_INIT; 6165 return !memcmp(cur ? (cur + idx_cur) : &tmp, 6166 new ? (new + idx_new) : &tmp, 6167 sizeof(struct sched_domain_attr)); 6168} 6169 6170/* 6171 * Partition sched domains as specified by the 'ndoms_new' 6172 * cpumasks in the array doms_new[] of cpumasks. This compares 6173 * doms_new[] to the current sched domain partitioning, doms_cur[]. 6174 * It destroys each deleted domain and builds each new domain. 6175 * 6176 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'. 6177 * The masks don't intersect (don't overlap.) We should setup one 6178 * sched domain for each mask. CPUs not in any of the cpumasks will 6179 * not be load balanced. If the same cpumask appears both in the 6180 * current 'doms_cur' domains and in the new 'doms_new', we can leave 6181 * it as it is. 6182 * 6183 * The passed in 'doms_new' should be allocated using 6184 * alloc_sched_domains. This routine takes ownership of it and will 6185 * free_sched_domains it when done with it. If the caller failed the 6186 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1, 6187 * and partition_sched_domains() will fallback to the single partition 6188 * 'fallback_doms', it also forces the domains to be rebuilt. 6189 * 6190 * If doms_new == NULL it will be replaced with cpu_online_mask. 6191 * ndoms_new == 0 is a special case for destroying existing domains, 6192 * and it will not create the default domain. 6193 * 6194 * Call with hotplug lock held 6195 */ 6196void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[], 6197 struct sched_domain_attr *dattr_new) 6198{ 6199 int i, j, n; 6200 int new_topology; 6201 6202 mutex_lock(&sched_domains_mutex); 6203 6204 /* always unregister in case we don't destroy any domains */ 6205 unregister_sched_domain_sysctl(); 6206 6207 /* Let architecture update cpu core mappings. */ 6208 new_topology = arch_update_cpu_topology(); 6209 6210 n = doms_new ? ndoms_new : 0; 6211 6212 /* Destroy deleted domains */ 6213 for (i = 0; i < ndoms_cur; i++) { 6214 for (j = 0; j < n && !new_topology; j++) { 6215 if (cpumask_equal(doms_cur[i], doms_new[j]) 6216 && dattrs_equal(dattr_cur, i, dattr_new, j)) 6217 goto match1; 6218 } 6219 /* no match - a current sched domain not in new doms_new[] */ 6220 detach_destroy_domains(doms_cur[i]); 6221match1: 6222 ; 6223 } 6224 6225 if (doms_new == NULL) { 6226 ndoms_cur = 0; 6227 doms_new = &fallback_doms; 6228 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map); 6229 WARN_ON_ONCE(dattr_new); 6230 } 6231 6232 /* Build new domains */ 6233 for (i = 0; i < ndoms_new; i++) { 6234 for (j = 0; j < ndoms_cur && !new_topology; j++) { 6235 if (cpumask_equal(doms_new[i], doms_cur[j]) 6236 && dattrs_equal(dattr_new, i, dattr_cur, j)) 6237 goto match2; 6238 } 6239 /* no match - add a new doms_new */ 6240 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL); 6241match2: 6242 ; 6243 } 6244 6245 /* Remember the new sched domains */ 6246 if (doms_cur != &fallback_doms) 6247 free_sched_domains(doms_cur, ndoms_cur); 6248 kfree(dattr_cur); /* kfree(NULL) is safe */ 6249 doms_cur = doms_new; 6250 dattr_cur = dattr_new; 6251 ndoms_cur = ndoms_new; 6252 6253 register_sched_domain_sysctl(); 6254 6255 mutex_unlock(&sched_domains_mutex); 6256} 6257 6258static int num_cpus_frozen; /* used to mark begin/end of suspend/resume */ 6259 6260/* 6261 * Update cpusets according to cpu_active mask. If cpusets are 6262 * disabled, cpuset_update_active_cpus() becomes a simple wrapper 6263 * around partition_sched_domains(). 6264 * 6265 * If we come here as part of a suspend/resume, don't touch cpusets because we 6266 * want to restore it back to its original state upon resume anyway. 6267 */ 6268static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action, 6269 void *hcpu) 6270{ 6271 switch (action) { 6272 case CPU_ONLINE_FROZEN: 6273 case CPU_DOWN_FAILED_FROZEN: 6274 6275 /* 6276 * num_cpus_frozen tracks how many CPUs are involved in suspend 6277 * resume sequence. As long as this is not the last online 6278 * operation in the resume sequence, just build a single sched 6279 * domain, ignoring cpusets. 6280 */ 6281 num_cpus_frozen--; 6282 if (likely(num_cpus_frozen)) { 6283 partition_sched_domains(1, NULL, NULL); 6284 break; 6285 } 6286 6287 /* 6288 * This is the last CPU online operation. So fall through and 6289 * restore the original sched domains by considering the 6290 * cpuset configurations. 6291 */ 6292 6293 case CPU_ONLINE: 6294 case CPU_DOWN_FAILED: 6295 cpuset_update_active_cpus(true); 6296 break; 6297 default: 6298 return NOTIFY_DONE; 6299 } 6300 return NOTIFY_OK; 6301} 6302 6303static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action, 6304 void *hcpu) 6305{ 6306 switch (action) { 6307 case CPU_DOWN_PREPARE: 6308 cpuset_update_active_cpus(false); 6309 break; 6310 case CPU_DOWN_PREPARE_FROZEN: 6311 num_cpus_frozen++; 6312 partition_sched_domains(1, NULL, NULL); 6313 break; 6314 default: 6315 return NOTIFY_DONE; 6316 } 6317 return NOTIFY_OK; 6318} 6319 6320void __init sched_init_smp(void) 6321{ 6322 cpumask_var_t non_isolated_cpus; 6323 6324 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL); 6325 alloc_cpumask_var(&fallback_doms, GFP_KERNEL); 6326 6327 sched_init_numa(); 6328 6329 get_online_cpus(); 6330 mutex_lock(&sched_domains_mutex); 6331 init_sched_domains(cpu_active_mask); 6332 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map); 6333 if (cpumask_empty(non_isolated_cpus)) 6334 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus); 6335 mutex_unlock(&sched_domains_mutex); 6336 put_online_cpus(); 6337 6338 hotcpu_notifier(sched_domains_numa_masks_update, CPU_PRI_SCHED_ACTIVE); 6339 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE); 6340 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE); 6341 6342 init_hrtick(); 6343 6344 /* Move init over to a non-isolated CPU */ 6345 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0) 6346 BUG(); 6347 sched_init_granularity(); 6348 free_cpumask_var(non_isolated_cpus); 6349 6350 init_sched_rt_class(); 6351} 6352#else 6353void __init sched_init_smp(void) 6354{ 6355 sched_init_granularity(); 6356} 6357#endif /* CONFIG_SMP */ 6358 6359const_debug unsigned int sysctl_timer_migration = 1; 6360 6361int in_sched_functions(unsigned long addr) 6362{ 6363 return in_lock_functions(addr) || 6364 (addr >= (unsigned long)__sched_text_start 6365 && addr < (unsigned long)__sched_text_end); 6366} 6367 6368#ifdef CONFIG_CGROUP_SCHED 6369/* 6370 * Default task group. 6371 * Every task in system belongs to this group at bootup. 6372 */ 6373struct task_group root_task_group; 6374LIST_HEAD(task_groups); 6375#endif 6376 6377DECLARE_PER_CPU(cpumask_var_t, load_balance_mask); 6378 6379void __init sched_init(void) 6380{ 6381 int i, j; 6382 unsigned long alloc_size = 0, ptr; 6383 6384#ifdef CONFIG_FAIR_GROUP_SCHED 6385 alloc_size += 2 * nr_cpu_ids * sizeof(void **); 6386#endif 6387#ifdef CONFIG_RT_GROUP_SCHED 6388 alloc_size += 2 * nr_cpu_ids * sizeof(void **); 6389#endif 6390#ifdef CONFIG_CPUMASK_OFFSTACK 6391 alloc_size += num_possible_cpus() * cpumask_size(); 6392#endif 6393 if (alloc_size) { 6394 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT); 6395 6396#ifdef CONFIG_FAIR_GROUP_SCHED 6397 root_task_group.se = (struct sched_entity **)ptr; 6398 ptr += nr_cpu_ids * sizeof(void **); 6399 6400 root_task_group.cfs_rq = (struct cfs_rq **)ptr; 6401 ptr += nr_cpu_ids * sizeof(void **); 6402 6403#endif /* CONFIG_FAIR_GROUP_SCHED */ 6404#ifdef CONFIG_RT_GROUP_SCHED 6405 root_task_group.rt_se = (struct sched_rt_entity **)ptr; 6406 ptr += nr_cpu_ids * sizeof(void **); 6407 6408 root_task_group.rt_rq = (struct rt_rq **)ptr; 6409 ptr += nr_cpu_ids * sizeof(void **); 6410 6411#endif /* CONFIG_RT_GROUP_SCHED */ 6412#ifdef CONFIG_CPUMASK_OFFSTACK 6413 for_each_possible_cpu(i) { 6414 per_cpu(load_balance_mask, i) = (void *)ptr; 6415 ptr += cpumask_size(); 6416 } 6417#endif /* CONFIG_CPUMASK_OFFSTACK */ 6418 } 6419 6420#ifdef CONFIG_SMP 6421 init_defrootdomain(); 6422#endif 6423 6424 init_rt_bandwidth(&def_rt_bandwidth, 6425 global_rt_period(), global_rt_runtime()); 6426 6427#ifdef CONFIG_RT_GROUP_SCHED 6428 init_rt_bandwidth(&root_task_group.rt_bandwidth, 6429 global_rt_period(), global_rt_runtime()); 6430#endif /* CONFIG_RT_GROUP_SCHED */ 6431 6432#ifdef CONFIG_CGROUP_SCHED 6433 list_add(&root_task_group.list, &task_groups); 6434 INIT_LIST_HEAD(&root_task_group.children); 6435 INIT_LIST_HEAD(&root_task_group.siblings); 6436 autogroup_init(&init_task); 6437 6438#endif /* CONFIG_CGROUP_SCHED */ 6439 6440 for_each_possible_cpu(i) { 6441 struct rq *rq; 6442 6443 rq = cpu_rq(i); 6444 raw_spin_lock_init(&rq->lock); 6445 rq->nr_running = 0; 6446 rq->calc_load_active = 0; 6447 rq->calc_load_update = jiffies + LOAD_FREQ; 6448 init_cfs_rq(&rq->cfs); 6449 init_rt_rq(&rq->rt, rq); 6450#ifdef CONFIG_FAIR_GROUP_SCHED 6451 root_task_group.shares = ROOT_TASK_GROUP_LOAD; 6452 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list); 6453 /* 6454 * How much cpu bandwidth does root_task_group get? 6455 * 6456 * In case of task-groups formed thr' the cgroup filesystem, it 6457 * gets 100% of the cpu resources in the system. This overall 6458 * system cpu resource is divided among the tasks of 6459 * root_task_group and its child task-groups in a fair manner, 6460 * based on each entity's (task or task-group's) weight 6461 * (se->load.weight). 6462 * 6463 * In other words, if root_task_group has 10 tasks of weight 6464 * 1024) and two child groups A0 and A1 (of weight 1024 each), 6465 * then A0's share of the cpu resource is: 6466 * 6467 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33% 6468 * 6469 * We achieve this by letting root_task_group's tasks sit 6470 * directly in rq->cfs (i.e root_task_group->se[] = NULL). 6471 */ 6472 init_cfs_bandwidth(&root_task_group.cfs_bandwidth); 6473 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL); 6474#endif /* CONFIG_FAIR_GROUP_SCHED */ 6475 6476 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime; 6477#ifdef CONFIG_RT_GROUP_SCHED 6478 INIT_LIST_HEAD(&rq->leaf_rt_rq_list); 6479 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL); 6480#endif 6481 6482 for (j = 0; j < CPU_LOAD_IDX_MAX; j++) 6483 rq->cpu_load[j] = 0; 6484 6485 rq->last_load_update_tick = jiffies; 6486 6487#ifdef CONFIG_SMP 6488 rq->sd = NULL; 6489 rq->rd = NULL; 6490 rq->cpu_power = SCHED_POWER_SCALE; 6491 rq->post_schedule = 0; 6492 rq->active_balance = 0; 6493 rq->next_balance = jiffies; 6494 rq->push_cpu = 0; 6495 rq->cpu = i; 6496 rq->online = 0; 6497 rq->idle_stamp = 0; 6498 rq->avg_idle = 2*sysctl_sched_migration_cost; 6499 6500 INIT_LIST_HEAD(&rq->cfs_tasks); 6501 6502 rq_attach_root(rq, &def_root_domain); 6503#ifdef CONFIG_NO_HZ_COMMON 6504 rq->nohz_flags = 0; 6505#endif 6506#ifdef CONFIG_NO_HZ_FULL 6507 rq->last_sched_tick = 0; 6508#endif 6509#endif 6510 init_rq_hrtick(rq); 6511 atomic_set(&rq->nr_iowait, 0); 6512 } 6513 6514 set_load_weight(&init_task); 6515 6516#ifdef CONFIG_PREEMPT_NOTIFIERS 6517 INIT_HLIST_HEAD(&init_task.preempt_notifiers); 6518#endif 6519 6520#ifdef CONFIG_RT_MUTEXES 6521 plist_head_init(&init_task.pi_waiters); 6522#endif 6523 6524 /* 6525 * The boot idle thread does lazy MMU switching as well: 6526 */ 6527 atomic_inc(&init_mm.mm_count); 6528 enter_lazy_tlb(&init_mm, current); 6529 6530 /* 6531 * Make us the idle thread. Technically, schedule() should not be 6532 * called from this thread, however somewhere below it might be, 6533 * but because we are the idle thread, we just pick up running again 6534 * when this runqueue becomes "idle". 6535 */ 6536 init_idle(current, smp_processor_id()); 6537 6538 calc_load_update = jiffies + LOAD_FREQ; 6539 6540 /* 6541 * During early bootup we pretend to be a normal task: 6542 */ 6543 current->sched_class = &fair_sched_class; 6544 6545#ifdef CONFIG_SMP 6546 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT); 6547 /* May be allocated at isolcpus cmdline parse time */ 6548 if (cpu_isolated_map == NULL) 6549 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT); 6550 idle_thread_set_boot_cpu(); 6551#endif 6552 init_sched_fair_class(); 6553 6554 scheduler_running = 1; 6555} 6556 6557#ifdef CONFIG_DEBUG_ATOMIC_SLEEP 6558static inline int preempt_count_equals(int preempt_offset) 6559{ 6560 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth(); 6561 6562 return (nested == preempt_offset); 6563} 6564 6565void __might_sleep(const char *file, int line, int preempt_offset) 6566{ 6567 static unsigned long prev_jiffy; /* ratelimiting */ 6568 6569 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */ 6570 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) || 6571 system_state != SYSTEM_RUNNING || oops_in_progress) 6572 return; 6573 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy) 6574 return; 6575 prev_jiffy = jiffies; 6576 6577 printk(KERN_ERR 6578 "BUG: sleeping function called from invalid context at %s:%d\n", 6579 file, line); 6580 printk(KERN_ERR 6581 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n", 6582 in_atomic(), irqs_disabled(), 6583 current->pid, current->comm); 6584 6585 debug_show_held_locks(current); 6586 if (irqs_disabled()) 6587 print_irqtrace_events(current); 6588 dump_stack(); 6589} 6590EXPORT_SYMBOL(__might_sleep); 6591#endif 6592 6593#ifdef CONFIG_MAGIC_SYSRQ 6594static void normalize_task(struct rq *rq, struct task_struct *p) 6595{ 6596 const struct sched_class *prev_class = p->sched_class; 6597 int old_prio = p->prio; 6598 int on_rq; 6599 6600 on_rq = p->on_rq; 6601 if (on_rq) 6602 dequeue_task(rq, p, 0); 6603 __setscheduler(rq, p, SCHED_NORMAL, 0); 6604 if (on_rq) { 6605 enqueue_task(rq, p, 0); 6606 resched_task(rq->curr); 6607 } 6608 6609 check_class_changed(rq, p, prev_class, old_prio); 6610} 6611 6612void normalize_rt_tasks(void) 6613{ 6614 struct task_struct *g, *p; 6615 unsigned long flags; 6616 struct rq *rq; 6617 6618 read_lock_irqsave(&tasklist_lock, flags); 6619 do_each_thread(g, p) { 6620 /* 6621 * Only normalize user tasks: 6622 */ 6623 if (!p->mm) 6624 continue; 6625 6626 p->se.exec_start = 0; 6627#ifdef CONFIG_SCHEDSTATS 6628 p->se.statistics.wait_start = 0; 6629 p->se.statistics.sleep_start = 0; 6630 p->se.statistics.block_start = 0; 6631#endif 6632 6633 if (!rt_task(p)) { 6634 /* 6635 * Renice negative nice level userspace 6636 * tasks back to 0: 6637 */ 6638 if (TASK_NICE(p) < 0 && p->mm) 6639 set_user_nice(p, 0); 6640 continue; 6641 } 6642 6643 raw_spin_lock(&p->pi_lock); 6644 rq = __task_rq_lock(p); 6645 6646 normalize_task(rq, p); 6647 6648 __task_rq_unlock(rq); 6649 raw_spin_unlock(&p->pi_lock); 6650 } while_each_thread(g, p); 6651 6652 read_unlock_irqrestore(&tasklist_lock, flags); 6653} 6654 6655#endif /* CONFIG_MAGIC_SYSRQ */ 6656 6657#if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) 6658/* 6659 * These functions are only useful for the IA64 MCA handling, or kdb. 6660 * 6661 * They can only be called when the whole system has been 6662 * stopped - every CPU needs to be quiescent, and no scheduling 6663 * activity can take place. Using them for anything else would 6664 * be a serious bug, and as a result, they aren't even visible 6665 * under any other configuration. 6666 */ 6667 6668/** 6669 * curr_task - return the current task for a given cpu. 6670 * @cpu: the processor in question. 6671 * 6672 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED! 6673 * 6674 * Return: The current task for @cpu. 6675 */ 6676struct task_struct *curr_task(int cpu) 6677{ 6678 return cpu_curr(cpu); 6679} 6680 6681#endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */ 6682 6683#ifdef CONFIG_IA64 6684/** 6685 * set_curr_task - set the current task for a given cpu. 6686 * @cpu: the processor in question. 6687 * @p: the task pointer to set. 6688 * 6689 * Description: This function must only be used when non-maskable interrupts 6690 * are serviced on a separate stack. It allows the architecture to switch the 6691 * notion of the current task on a cpu in a non-blocking manner. This function 6692 * must be called with all CPU's synchronized, and interrupts disabled, the 6693 * and caller must save the original value of the current task (see 6694 * curr_task() above) and restore that value before reenabling interrupts and 6695 * re-starting the system. 6696 * 6697 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED! 6698 */ 6699void set_curr_task(int cpu, struct task_struct *p) 6700{ 6701 cpu_curr(cpu) = p; 6702} 6703 6704#endif 6705 6706#ifdef CONFIG_CGROUP_SCHED 6707/* task_group_lock serializes the addition/removal of task groups */ 6708static DEFINE_SPINLOCK(task_group_lock); 6709 6710static void free_sched_group(struct task_group *tg) 6711{ 6712 free_fair_sched_group(tg); 6713 free_rt_sched_group(tg); 6714 autogroup_free(tg); 6715 kfree(tg); 6716} 6717 6718/* allocate runqueue etc for a new task group */ 6719struct task_group *sched_create_group(struct task_group *parent) 6720{ 6721 struct task_group *tg; 6722 6723 tg = kzalloc(sizeof(*tg), GFP_KERNEL); 6724 if (!tg) 6725 return ERR_PTR(-ENOMEM); 6726 6727 if (!alloc_fair_sched_group(tg, parent)) 6728 goto err; 6729 6730 if (!alloc_rt_sched_group(tg, parent)) 6731 goto err; 6732 6733 return tg; 6734 6735err: 6736 free_sched_group(tg); 6737 return ERR_PTR(-ENOMEM); 6738} 6739 6740void sched_online_group(struct task_group *tg, struct task_group *parent) 6741{ 6742 unsigned long flags; 6743 6744 spin_lock_irqsave(&task_group_lock, flags); 6745 list_add_rcu(&tg->list, &task_groups); 6746 6747 WARN_ON(!parent); /* root should already exist */ 6748 6749 tg->parent = parent; 6750 INIT_LIST_HEAD(&tg->children); 6751 list_add_rcu(&tg->siblings, &parent->children); 6752 spin_unlock_irqrestore(&task_group_lock, flags); 6753} 6754 6755/* rcu callback to free various structures associated with a task group */ 6756static void free_sched_group_rcu(struct rcu_head *rhp) 6757{ 6758 /* now it should be safe to free those cfs_rqs */ 6759 free_sched_group(container_of(rhp, struct task_group, rcu)); 6760} 6761 6762/* Destroy runqueue etc associated with a task group */ 6763void sched_destroy_group(struct task_group *tg) 6764{ 6765 /* wait for possible concurrent references to cfs_rqs complete */ 6766 call_rcu(&tg->rcu, free_sched_group_rcu); 6767} 6768 6769void sched_offline_group(struct task_group *tg) 6770{ 6771 unsigned long flags; 6772 int i; 6773 6774 /* end participation in shares distribution */ 6775 for_each_possible_cpu(i) 6776 unregister_fair_sched_group(tg, i); 6777 6778 spin_lock_irqsave(&task_group_lock, flags); 6779 list_del_rcu(&tg->list); 6780 list_del_rcu(&tg->siblings); 6781 spin_unlock_irqrestore(&task_group_lock, flags); 6782} 6783 6784/* change task's runqueue when it moves between groups. 6785 * The caller of this function should have put the task in its new group 6786 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to 6787 * reflect its new group. 6788 */ 6789void sched_move_task(struct task_struct *tsk) 6790{ 6791 struct task_group *tg; 6792 int on_rq, running; 6793 unsigned long flags; 6794 struct rq *rq; 6795 6796 rq = task_rq_lock(tsk, &flags); 6797 6798 running = task_current(rq, tsk); 6799 on_rq = tsk->on_rq; 6800 6801 if (on_rq) 6802 dequeue_task(rq, tsk, 0); 6803 if (unlikely(running)) 6804 tsk->sched_class->put_prev_task(rq, tsk); 6805 6806 tg = container_of(task_subsys_state_check(tsk, cpu_cgroup_subsys_id, 6807 lockdep_is_held(&tsk->sighand->siglock)), 6808 struct task_group, css); 6809 tg = autogroup_task_group(tsk, tg); 6810 tsk->sched_task_group = tg; 6811 6812#ifdef CONFIG_FAIR_GROUP_SCHED 6813 if (tsk->sched_class->task_move_group) 6814 tsk->sched_class->task_move_group(tsk, on_rq); 6815 else 6816#endif 6817 set_task_rq(tsk, task_cpu(tsk)); 6818 6819 if (unlikely(running)) 6820 tsk->sched_class->set_curr_task(rq); 6821 if (on_rq) 6822 enqueue_task(rq, tsk, 0); 6823 6824 task_rq_unlock(rq, tsk, &flags); 6825} 6826#endif /* CONFIG_CGROUP_SCHED */ 6827 6828#if defined(CONFIG_RT_GROUP_SCHED) || defined(CONFIG_CFS_BANDWIDTH) 6829static unsigned long to_ratio(u64 period, u64 runtime) 6830{ 6831 if (runtime == RUNTIME_INF) 6832 return 1ULL << 20; 6833 6834 return div64_u64(runtime << 20, period); 6835} 6836#endif 6837 6838#ifdef CONFIG_RT_GROUP_SCHED 6839/* 6840 * Ensure that the real time constraints are schedulable. 6841 */ 6842static DEFINE_MUTEX(rt_constraints_mutex); 6843 6844/* Must be called with tasklist_lock held */ 6845static inline int tg_has_rt_tasks(struct task_group *tg) 6846{ 6847 struct task_struct *g, *p; 6848 6849 do_each_thread(g, p) { 6850 if (rt_task(p) && task_rq(p)->rt.tg == tg) 6851 return 1; 6852 } while_each_thread(g, p); 6853 6854 return 0; 6855} 6856 6857struct rt_schedulable_data { 6858 struct task_group *tg; 6859 u64 rt_period; 6860 u64 rt_runtime; 6861}; 6862 6863static int tg_rt_schedulable(struct task_group *tg, void *data) 6864{ 6865 struct rt_schedulable_data *d = data; 6866 struct task_group *child; 6867 unsigned long total, sum = 0; 6868 u64 period, runtime; 6869 6870 period = ktime_to_ns(tg->rt_bandwidth.rt_period); 6871 runtime = tg->rt_bandwidth.rt_runtime; 6872 6873 if (tg == d->tg) { 6874 period = d->rt_period; 6875 runtime = d->rt_runtime; 6876 } 6877 6878 /* 6879 * Cannot have more runtime than the period. 6880 */ 6881 if (runtime > period && runtime != RUNTIME_INF) 6882 return -EINVAL; 6883 6884 /* 6885 * Ensure we don't starve existing RT tasks. 6886 */ 6887 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg)) 6888 return -EBUSY; 6889 6890 total = to_ratio(period, runtime); 6891 6892 /* 6893 * Nobody can have more than the global setting allows. 6894 */ 6895 if (total > to_ratio(global_rt_period(), global_rt_runtime())) 6896 return -EINVAL; 6897 6898 /* 6899 * The sum of our children's runtime should not exceed our own. 6900 */ 6901 list_for_each_entry_rcu(child, &tg->children, siblings) { 6902 period = ktime_to_ns(child->rt_bandwidth.rt_period); 6903 runtime = child->rt_bandwidth.rt_runtime; 6904 6905 if (child == d->tg) { 6906 period = d->rt_period; 6907 runtime = d->rt_runtime; 6908 } 6909 6910 sum += to_ratio(period, runtime); 6911 } 6912 6913 if (sum > total) 6914 return -EINVAL; 6915 6916 return 0; 6917} 6918 6919static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime) 6920{ 6921 int ret; 6922 6923 struct rt_schedulable_data data = { 6924 .tg = tg, 6925 .rt_period = period, 6926 .rt_runtime = runtime, 6927 }; 6928 6929 rcu_read_lock(); 6930 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data); 6931 rcu_read_unlock(); 6932 6933 return ret; 6934} 6935 6936static int tg_set_rt_bandwidth(struct task_group *tg, 6937 u64 rt_period, u64 rt_runtime) 6938{ 6939 int i, err = 0; 6940 6941 mutex_lock(&rt_constraints_mutex); 6942 read_lock(&tasklist_lock); 6943 err = __rt_schedulable(tg, rt_period, rt_runtime); 6944 if (err) 6945 goto unlock; 6946 6947 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock); 6948 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period); 6949 tg->rt_bandwidth.rt_runtime = rt_runtime; 6950 6951 for_each_possible_cpu(i) { 6952 struct rt_rq *rt_rq = tg->rt_rq[i]; 6953 6954 raw_spin_lock(&rt_rq->rt_runtime_lock); 6955 rt_rq->rt_runtime = rt_runtime; 6956 raw_spin_unlock(&rt_rq->rt_runtime_lock); 6957 } 6958 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock); 6959unlock: 6960 read_unlock(&tasklist_lock); 6961 mutex_unlock(&rt_constraints_mutex); 6962 6963 return err; 6964} 6965 6966static int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us) 6967{ 6968 u64 rt_runtime, rt_period; 6969 6970 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period); 6971 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC; 6972 if (rt_runtime_us < 0) 6973 rt_runtime = RUNTIME_INF; 6974 6975 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime); 6976} 6977 6978static long sched_group_rt_runtime(struct task_group *tg) 6979{ 6980 u64 rt_runtime_us; 6981 6982 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF) 6983 return -1; 6984 6985 rt_runtime_us = tg->rt_bandwidth.rt_runtime; 6986 do_div(rt_runtime_us, NSEC_PER_USEC); 6987 return rt_runtime_us; 6988} 6989 6990static int sched_group_set_rt_period(struct task_group *tg, long rt_period_us) 6991{ 6992 u64 rt_runtime, rt_period; 6993 6994 rt_period = (u64)rt_period_us * NSEC_PER_USEC; 6995 rt_runtime = tg->rt_bandwidth.rt_runtime; 6996 6997 if (rt_period == 0) 6998 return -EINVAL; 6999 7000 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime); 7001} 7002 7003static long sched_group_rt_period(struct task_group *tg) 7004{ 7005 u64 rt_period_us; 7006 7007 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period); 7008 do_div(rt_period_us, NSEC_PER_USEC); 7009 return rt_period_us; 7010} 7011 7012static int sched_rt_global_constraints(void) 7013{ 7014 u64 runtime, period; 7015 int ret = 0; 7016 7017 if (sysctl_sched_rt_period <= 0) 7018 return -EINVAL; 7019 7020 runtime = global_rt_runtime(); 7021 period = global_rt_period(); 7022 7023 /* 7024 * Sanity check on the sysctl variables. 7025 */ 7026 if (runtime > period && runtime != RUNTIME_INF) 7027 return -EINVAL; 7028 7029 mutex_lock(&rt_constraints_mutex); 7030 read_lock(&tasklist_lock); 7031 ret = __rt_schedulable(NULL, 0, 0); 7032 read_unlock(&tasklist_lock); 7033 mutex_unlock(&rt_constraints_mutex); 7034 7035 return ret; 7036} 7037 7038static int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk) 7039{ 7040 /* Don't accept realtime tasks when there is no way for them to run */ 7041 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0) 7042 return 0; 7043 7044 return 1; 7045} 7046 7047#else /* !CONFIG_RT_GROUP_SCHED */ 7048static int sched_rt_global_constraints(void) 7049{ 7050 unsigned long flags; 7051 int i; 7052 7053 if (sysctl_sched_rt_period <= 0) 7054 return -EINVAL; 7055 7056 /* 7057 * There's always some RT tasks in the root group 7058 * -- migration, kstopmachine etc.. 7059 */ 7060 if (sysctl_sched_rt_runtime == 0) 7061 return -EBUSY; 7062 7063 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags); 7064 for_each_possible_cpu(i) { 7065 struct rt_rq *rt_rq = &cpu_rq(i)->rt; 7066 7067 raw_spin_lock(&rt_rq->rt_runtime_lock); 7068 rt_rq->rt_runtime = global_rt_runtime(); 7069 raw_spin_unlock(&rt_rq->rt_runtime_lock); 7070 } 7071 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags); 7072 7073 return 0; 7074} 7075#endif /* CONFIG_RT_GROUP_SCHED */ 7076 7077int sched_rr_handler(struct ctl_table *table, int write, 7078 void __user *buffer, size_t *lenp, 7079 loff_t *ppos) 7080{ 7081 int ret; 7082 static DEFINE_MUTEX(mutex); 7083 7084 mutex_lock(&mutex); 7085 ret = proc_dointvec(table, write, buffer, lenp, ppos); 7086 /* make sure that internally we keep jiffies */ 7087 /* also, writing zero resets timeslice to default */ 7088 if (!ret && write) { 7089 sched_rr_timeslice = sched_rr_timeslice <= 0 ? 7090 RR_TIMESLICE : msecs_to_jiffies(sched_rr_timeslice); 7091 } 7092 mutex_unlock(&mutex); 7093 return ret; 7094} 7095 7096int sched_rt_handler(struct ctl_table *table, int write, 7097 void __user *buffer, size_t *lenp, 7098 loff_t *ppos) 7099{ 7100 int ret; 7101 int old_period, old_runtime; 7102 static DEFINE_MUTEX(mutex); 7103 7104 mutex_lock(&mutex); 7105 old_period = sysctl_sched_rt_period; 7106 old_runtime = sysctl_sched_rt_runtime; 7107 7108 ret = proc_dointvec(table, write, buffer, lenp, ppos); 7109 7110 if (!ret && write) { 7111 ret = sched_rt_global_constraints(); 7112 if (ret) { 7113 sysctl_sched_rt_period = old_period; 7114 sysctl_sched_rt_runtime = old_runtime; 7115 } else { 7116 def_rt_bandwidth.rt_runtime = global_rt_runtime(); 7117 def_rt_bandwidth.rt_period = 7118 ns_to_ktime(global_rt_period()); 7119 } 7120 } 7121 mutex_unlock(&mutex); 7122 7123 return ret; 7124} 7125 7126#ifdef CONFIG_CGROUP_SCHED 7127 7128/* return corresponding task_group object of a cgroup */ 7129static inline struct task_group *cgroup_tg(struct cgroup *cgrp) 7130{ 7131 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id), 7132 struct task_group, css); 7133} 7134 7135static struct cgroup_subsys_state *cpu_cgroup_css_alloc(struct cgroup *cgrp) 7136{ 7137 struct task_group *tg, *parent; 7138 7139 if (!cgrp->parent) { 7140 /* This is early initialization for the top cgroup */ 7141 return &root_task_group.css; 7142 } 7143 7144 parent = cgroup_tg(cgrp->parent); 7145 tg = sched_create_group(parent); 7146 if (IS_ERR(tg)) 7147 return ERR_PTR(-ENOMEM); 7148 7149 return &tg->css; 7150} 7151 7152static int cpu_cgroup_css_online(struct cgroup *cgrp) 7153{ 7154 struct task_group *tg = cgroup_tg(cgrp); 7155 struct task_group *parent; 7156 7157 if (!cgrp->parent) 7158 return 0; 7159 7160 parent = cgroup_tg(cgrp->parent); 7161 sched_online_group(tg, parent); 7162 return 0; 7163} 7164 7165static void cpu_cgroup_css_free(struct cgroup *cgrp) 7166{ 7167 struct task_group *tg = cgroup_tg(cgrp); 7168 7169 sched_destroy_group(tg); 7170} 7171 7172static void cpu_cgroup_css_offline(struct cgroup *cgrp) 7173{ 7174 struct task_group *tg = cgroup_tg(cgrp); 7175 7176 sched_offline_group(tg); 7177} 7178 7179static int cpu_cgroup_can_attach(struct cgroup *cgrp, 7180 struct cgroup_taskset *tset) 7181{ 7182 struct task_struct *task; 7183 7184 cgroup_taskset_for_each(task, cgrp, tset) { 7185#ifdef CONFIG_RT_GROUP_SCHED 7186 if (!sched_rt_can_attach(cgroup_tg(cgrp), task)) 7187 return -EINVAL; 7188#else 7189 /* We don't support RT-tasks being in separate groups */ 7190 if (task->sched_class != &fair_sched_class) 7191 return -EINVAL; 7192#endif 7193 } 7194 return 0; 7195} 7196 7197static void cpu_cgroup_attach(struct cgroup *cgrp, 7198 struct cgroup_taskset *tset) 7199{ 7200 struct task_struct *task; 7201 7202 cgroup_taskset_for_each(task, cgrp, tset) 7203 sched_move_task(task); 7204} 7205 7206static void 7207cpu_cgroup_exit(struct cgroup *cgrp, struct cgroup *old_cgrp, 7208 struct task_struct *task) 7209{ 7210 /* 7211 * cgroup_exit() is called in the copy_process() failure path. 7212 * Ignore this case since the task hasn't ran yet, this avoids 7213 * trying to poke a half freed task state from generic code. 7214 */ 7215 if (!(task->flags & PF_EXITING)) 7216 return; 7217 7218 sched_move_task(task); 7219} 7220 7221#ifdef CONFIG_FAIR_GROUP_SCHED 7222static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype, 7223 u64 shareval) 7224{ 7225 return sched_group_set_shares(cgroup_tg(cgrp), scale_load(shareval)); 7226} 7227 7228static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft) 7229{ 7230 struct task_group *tg = cgroup_tg(cgrp); 7231 7232 return (u64) scale_load_down(tg->shares); 7233} 7234 7235#ifdef CONFIG_CFS_BANDWIDTH 7236static DEFINE_MUTEX(cfs_constraints_mutex); 7237 7238const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */ 7239const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */ 7240 7241static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime); 7242 7243static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota) 7244{ 7245 int i, ret = 0, runtime_enabled, runtime_was_enabled; 7246 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth; 7247 7248 if (tg == &root_task_group) 7249 return -EINVAL; 7250 7251 /* 7252 * Ensure we have at some amount of bandwidth every period. This is 7253 * to prevent reaching a state of large arrears when throttled via 7254 * entity_tick() resulting in prolonged exit starvation. 7255 */ 7256 if (quota < min_cfs_quota_period || period < min_cfs_quota_period) 7257 return -EINVAL; 7258 7259 /* 7260 * Likewise, bound things on the otherside by preventing insane quota 7261 * periods. This also allows us to normalize in computing quota 7262 * feasibility. 7263 */ 7264 if (period > max_cfs_quota_period) 7265 return -EINVAL; 7266 7267 mutex_lock(&cfs_constraints_mutex); 7268 ret = __cfs_schedulable(tg, period, quota); 7269 if (ret) 7270 goto out_unlock; 7271 7272 runtime_enabled = quota != RUNTIME_INF; 7273 runtime_was_enabled = cfs_b->quota != RUNTIME_INF; 7274 account_cfs_bandwidth_used(runtime_enabled, runtime_was_enabled); 7275 raw_spin_lock_irq(&cfs_b->lock); 7276 cfs_b->period = ns_to_ktime(period); 7277 cfs_b->quota = quota; 7278 7279 __refill_cfs_bandwidth_runtime(cfs_b); 7280 /* restart the period timer (if active) to handle new period expiry */ 7281 if (runtime_enabled && cfs_b->timer_active) { 7282 /* force a reprogram */ 7283 cfs_b->timer_active = 0; 7284 __start_cfs_bandwidth(cfs_b); 7285 } 7286 raw_spin_unlock_irq(&cfs_b->lock); 7287 7288 for_each_possible_cpu(i) { 7289 struct cfs_rq *cfs_rq = tg->cfs_rq[i]; 7290 struct rq *rq = cfs_rq->rq; 7291 7292 raw_spin_lock_irq(&rq->lock); 7293 cfs_rq->runtime_enabled = runtime_enabled; 7294 cfs_rq->runtime_remaining = 0; 7295 7296 if (cfs_rq->throttled) 7297 unthrottle_cfs_rq(cfs_rq); 7298 raw_spin_unlock_irq(&rq->lock); 7299 } 7300out_unlock: 7301 mutex_unlock(&cfs_constraints_mutex); 7302 7303 return ret; 7304} 7305 7306int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us) 7307{ 7308 u64 quota, period; 7309 7310 period = ktime_to_ns(tg->cfs_bandwidth.period); 7311 if (cfs_quota_us < 0) 7312 quota = RUNTIME_INF; 7313 else 7314 quota = (u64)cfs_quota_us * NSEC_PER_USEC; 7315 7316 return tg_set_cfs_bandwidth(tg, period, quota); 7317} 7318 7319long tg_get_cfs_quota(struct task_group *tg) 7320{ 7321 u64 quota_us; 7322 7323 if (tg->cfs_bandwidth.quota == RUNTIME_INF) 7324 return -1; 7325 7326 quota_us = tg->cfs_bandwidth.quota; 7327 do_div(quota_us, NSEC_PER_USEC); 7328 7329 return quota_us; 7330} 7331 7332int tg_set_cfs_period(struct task_group *tg, long cfs_period_us) 7333{ 7334 u64 quota, period; 7335 7336 period = (u64)cfs_period_us * NSEC_PER_USEC; 7337 quota = tg->cfs_bandwidth.quota; 7338 7339 return tg_set_cfs_bandwidth(tg, period, quota); 7340} 7341 7342long tg_get_cfs_period(struct task_group *tg) 7343{ 7344 u64 cfs_period_us; 7345 7346 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period); 7347 do_div(cfs_period_us, NSEC_PER_USEC); 7348 7349 return cfs_period_us; 7350} 7351 7352static s64 cpu_cfs_quota_read_s64(struct cgroup *cgrp, struct cftype *cft) 7353{ 7354 return tg_get_cfs_quota(cgroup_tg(cgrp)); 7355} 7356 7357static int cpu_cfs_quota_write_s64(struct cgroup *cgrp, struct cftype *cftype, 7358 s64 cfs_quota_us) 7359{ 7360 return tg_set_cfs_quota(cgroup_tg(cgrp), cfs_quota_us); 7361} 7362 7363static u64 cpu_cfs_period_read_u64(struct cgroup *cgrp, struct cftype *cft) 7364{ 7365 return tg_get_cfs_period(cgroup_tg(cgrp)); 7366} 7367 7368static int cpu_cfs_period_write_u64(struct cgroup *cgrp, struct cftype *cftype, 7369 u64 cfs_period_us) 7370{ 7371 return tg_set_cfs_period(cgroup_tg(cgrp), cfs_period_us); 7372} 7373 7374struct cfs_schedulable_data { 7375 struct task_group *tg; 7376 u64 period, quota; 7377}; 7378 7379/* 7380 * normalize group quota/period to be quota/max_period 7381 * note: units are usecs 7382 */ 7383static u64 normalize_cfs_quota(struct task_group *tg, 7384 struct cfs_schedulable_data *d) 7385{ 7386 u64 quota, period; 7387 7388 if (tg == d->tg) { 7389 period = d->period; 7390 quota = d->quota; 7391 } else { 7392 period = tg_get_cfs_period(tg); 7393 quota = tg_get_cfs_quota(tg); 7394 } 7395 7396 /* note: these should typically be equivalent */ 7397 if (quota == RUNTIME_INF || quota == -1) 7398 return RUNTIME_INF; 7399 7400 return to_ratio(period, quota); 7401} 7402 7403static int tg_cfs_schedulable_down(struct task_group *tg, void *data) 7404{ 7405 struct cfs_schedulable_data *d = data; 7406 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth; 7407 s64 quota = 0, parent_quota = -1; 7408 7409 if (!tg->parent) { 7410 quota = RUNTIME_INF; 7411 } else { 7412 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth; 7413 7414 quota = normalize_cfs_quota(tg, d); 7415 parent_quota = parent_b->hierarchal_quota; 7416 7417 /* 7418 * ensure max(child_quota) <= parent_quota, inherit when no 7419 * limit is set 7420 */ 7421 if (quota == RUNTIME_INF) 7422 quota = parent_quota; 7423 else if (parent_quota != RUNTIME_INF && quota > parent_quota) 7424 return -EINVAL; 7425 } 7426 cfs_b->hierarchal_quota = quota; 7427 7428 return 0; 7429} 7430 7431static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota) 7432{ 7433 int ret; 7434 struct cfs_schedulable_data data = { 7435 .tg = tg, 7436 .period = period, 7437 .quota = quota, 7438 }; 7439 7440 if (quota != RUNTIME_INF) { 7441 do_div(data.period, NSEC_PER_USEC); 7442 do_div(data.quota, NSEC_PER_USEC); 7443 } 7444 7445 rcu_read_lock(); 7446 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data); 7447 rcu_read_unlock(); 7448 7449 return ret; 7450} 7451 7452static int cpu_stats_show(struct cgroup *cgrp, struct cftype *cft, 7453 struct cgroup_map_cb *cb) 7454{ 7455 struct task_group *tg = cgroup_tg(cgrp); 7456 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth; 7457 7458 cb->fill(cb, "nr_periods", cfs_b->nr_periods); 7459 cb->fill(cb, "nr_throttled", cfs_b->nr_throttled); 7460 cb->fill(cb, "throttled_time", cfs_b->throttled_time); 7461 7462 return 0; 7463} 7464#endif /* CONFIG_CFS_BANDWIDTH */ 7465#endif /* CONFIG_FAIR_GROUP_SCHED */ 7466 7467#ifdef CONFIG_RT_GROUP_SCHED 7468static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft, 7469 s64 val) 7470{ 7471 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val); 7472} 7473 7474static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft) 7475{ 7476 return sched_group_rt_runtime(cgroup_tg(cgrp)); 7477} 7478 7479static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype, 7480 u64 rt_period_us) 7481{ 7482 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us); 7483} 7484 7485static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft) 7486{ 7487 return sched_group_rt_period(cgroup_tg(cgrp)); 7488} 7489#endif /* CONFIG_RT_GROUP_SCHED */ 7490 7491static struct cftype cpu_files[] = { 7492#ifdef CONFIG_FAIR_GROUP_SCHED 7493 { 7494 .name = "shares", 7495 .read_u64 = cpu_shares_read_u64, 7496 .write_u64 = cpu_shares_write_u64, 7497 }, 7498#endif 7499#ifdef CONFIG_CFS_BANDWIDTH 7500 { 7501 .name = "cfs_quota_us", 7502 .read_s64 = cpu_cfs_quota_read_s64, 7503 .write_s64 = cpu_cfs_quota_write_s64, 7504 }, 7505 { 7506 .name = "cfs_period_us", 7507 .read_u64 = cpu_cfs_period_read_u64, 7508 .write_u64 = cpu_cfs_period_write_u64, 7509 }, 7510 { 7511 .name = "stat", 7512 .read_map = cpu_stats_show, 7513 }, 7514#endif 7515#ifdef CONFIG_RT_GROUP_SCHED 7516 { 7517 .name = "rt_runtime_us", 7518 .read_s64 = cpu_rt_runtime_read, 7519 .write_s64 = cpu_rt_runtime_write, 7520 }, 7521 { 7522 .name = "rt_period_us", 7523 .read_u64 = cpu_rt_period_read_uint, 7524 .write_u64 = cpu_rt_period_write_uint, 7525 }, 7526#endif 7527 { } /* terminate */ 7528}; 7529 7530struct cgroup_subsys cpu_cgroup_subsys = { 7531 .name = "cpu", 7532 .css_alloc = cpu_cgroup_css_alloc, 7533 .css_free = cpu_cgroup_css_free, 7534 .css_online = cpu_cgroup_css_online, 7535 .css_offline = cpu_cgroup_css_offline, 7536 .can_attach = cpu_cgroup_can_attach, 7537 .attach = cpu_cgroup_attach, 7538 .exit = cpu_cgroup_exit, 7539 .subsys_id = cpu_cgroup_subsys_id, 7540 .base_cftypes = cpu_files, 7541 .early_init = 1, 7542}; 7543 7544#endif /* CONFIG_CGROUP_SCHED */ 7545 7546void dump_cpu_task(int cpu) 7547{ 7548 pr_info("Task dump for CPU %d:\n", cpu); 7549 sched_show_task(cpu_curr(cpu)); 7550} 7551