core.c revision 5e1576ed0e54d419286a8096133029062b6ad456
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 */ 516void resched_task(struct task_struct *p) 517{ 518 int cpu; 519 520 lockdep_assert_held(&task_rq(p)->lock); 521 522 if (test_tsk_need_resched(p)) 523 return; 524 525 set_tsk_need_resched(p); 526 527 cpu = task_cpu(p); 528 if (cpu == smp_processor_id()) { 529 set_preempt_need_resched(); 530 return; 531 } 532 533 /* NEED_RESCHED must be visible before we test polling */ 534 smp_mb(); 535 if (!tsk_is_polling(p)) 536 smp_send_reschedule(cpu); 537} 538 539void resched_cpu(int cpu) 540{ 541 struct rq *rq = cpu_rq(cpu); 542 unsigned long flags; 543 544 if (!raw_spin_trylock_irqsave(&rq->lock, flags)) 545 return; 546 resched_task(cpu_curr(cpu)); 547 raw_spin_unlock_irqrestore(&rq->lock, flags); 548} 549 550#ifdef CONFIG_SMP 551#ifdef CONFIG_NO_HZ_COMMON 552/* 553 * In the semi idle case, use the nearest busy cpu for migrating timers 554 * from an idle cpu. This is good for power-savings. 555 * 556 * We don't do similar optimization for completely idle system, as 557 * selecting an idle cpu will add more delays to the timers than intended 558 * (as that cpu's timer base may not be uptodate wrt jiffies etc). 559 */ 560int get_nohz_timer_target(void) 561{ 562 int cpu = smp_processor_id(); 563 int i; 564 struct sched_domain *sd; 565 566 rcu_read_lock(); 567 for_each_domain(cpu, sd) { 568 for_each_cpu(i, sched_domain_span(sd)) { 569 if (!idle_cpu(i)) { 570 cpu = i; 571 goto unlock; 572 } 573 } 574 } 575unlock: 576 rcu_read_unlock(); 577 return cpu; 578} 579/* 580 * When add_timer_on() enqueues a timer into the timer wheel of an 581 * idle CPU then this timer might expire before the next timer event 582 * which is scheduled to wake up that CPU. In case of a completely 583 * idle system the next event might even be infinite time into the 584 * future. wake_up_idle_cpu() ensures that the CPU is woken up and 585 * leaves the inner idle loop so the newly added timer is taken into 586 * account when the CPU goes back to idle and evaluates the timer 587 * wheel for the next timer event. 588 */ 589static void wake_up_idle_cpu(int cpu) 590{ 591 struct rq *rq = cpu_rq(cpu); 592 593 if (cpu == smp_processor_id()) 594 return; 595 596 /* 597 * This is safe, as this function is called with the timer 598 * wheel base lock of (cpu) held. When the CPU is on the way 599 * to idle and has not yet set rq->curr to idle then it will 600 * be serialized on the timer wheel base lock and take the new 601 * timer into account automatically. 602 */ 603 if (rq->curr != rq->idle) 604 return; 605 606 /* 607 * We can set TIF_RESCHED on the idle task of the other CPU 608 * lockless. The worst case is that the other CPU runs the 609 * idle task through an additional NOOP schedule() 610 */ 611 set_tsk_need_resched(rq->idle); 612 613 /* NEED_RESCHED must be visible before we test polling */ 614 smp_mb(); 615 if (!tsk_is_polling(rq->idle)) 616 smp_send_reschedule(cpu); 617} 618 619static bool wake_up_full_nohz_cpu(int cpu) 620{ 621 if (tick_nohz_full_cpu(cpu)) { 622 if (cpu != smp_processor_id() || 623 tick_nohz_tick_stopped()) 624 smp_send_reschedule(cpu); 625 return true; 626 } 627 628 return false; 629} 630 631void wake_up_nohz_cpu(int cpu) 632{ 633 if (!wake_up_full_nohz_cpu(cpu)) 634 wake_up_idle_cpu(cpu); 635} 636 637static inline bool got_nohz_idle_kick(void) 638{ 639 int cpu = smp_processor_id(); 640 641 if (!test_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu))) 642 return false; 643 644 if (idle_cpu(cpu) && !need_resched()) 645 return true; 646 647 /* 648 * We can't run Idle Load Balance on this CPU for this time so we 649 * cancel it and clear NOHZ_BALANCE_KICK 650 */ 651 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu)); 652 return false; 653} 654 655#else /* CONFIG_NO_HZ_COMMON */ 656 657static inline bool got_nohz_idle_kick(void) 658{ 659 return false; 660} 661 662#endif /* CONFIG_NO_HZ_COMMON */ 663 664#ifdef CONFIG_NO_HZ_FULL 665bool sched_can_stop_tick(void) 666{ 667 struct rq *rq; 668 669 rq = this_rq(); 670 671 /* Make sure rq->nr_running update is visible after the IPI */ 672 smp_rmb(); 673 674 /* More than one running task need preemption */ 675 if (rq->nr_running > 1) 676 return false; 677 678 return true; 679} 680#endif /* CONFIG_NO_HZ_FULL */ 681 682void sched_avg_update(struct rq *rq) 683{ 684 s64 period = sched_avg_period(); 685 686 while ((s64)(rq_clock(rq) - rq->age_stamp) > period) { 687 /* 688 * Inline assembly required to prevent the compiler 689 * optimising this loop into a divmod call. 690 * See __iter_div_u64_rem() for another example of this. 691 */ 692 asm("" : "+rm" (rq->age_stamp)); 693 rq->age_stamp += period; 694 rq->rt_avg /= 2; 695 } 696} 697 698#endif /* CONFIG_SMP */ 699 700#if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \ 701 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH))) 702/* 703 * Iterate task_group tree rooted at *from, calling @down when first entering a 704 * node and @up when leaving it for the final time. 705 * 706 * Caller must hold rcu_lock or sufficient equivalent. 707 */ 708int walk_tg_tree_from(struct task_group *from, 709 tg_visitor down, tg_visitor up, void *data) 710{ 711 struct task_group *parent, *child; 712 int ret; 713 714 parent = from; 715 716down: 717 ret = (*down)(parent, data); 718 if (ret) 719 goto out; 720 list_for_each_entry_rcu(child, &parent->children, siblings) { 721 parent = child; 722 goto down; 723 724up: 725 continue; 726 } 727 ret = (*up)(parent, data); 728 if (ret || parent == from) 729 goto out; 730 731 child = parent; 732 parent = parent->parent; 733 if (parent) 734 goto up; 735out: 736 return ret; 737} 738 739int tg_nop(struct task_group *tg, void *data) 740{ 741 return 0; 742} 743#endif 744 745static void set_load_weight(struct task_struct *p) 746{ 747 int prio = p->static_prio - MAX_RT_PRIO; 748 struct load_weight *load = &p->se.load; 749 750 /* 751 * SCHED_IDLE tasks get minimal weight: 752 */ 753 if (p->policy == SCHED_IDLE) { 754 load->weight = scale_load(WEIGHT_IDLEPRIO); 755 load->inv_weight = WMULT_IDLEPRIO; 756 return; 757 } 758 759 load->weight = scale_load(prio_to_weight[prio]); 760 load->inv_weight = prio_to_wmult[prio]; 761} 762 763static void enqueue_task(struct rq *rq, struct task_struct *p, int flags) 764{ 765 update_rq_clock(rq); 766 sched_info_queued(rq, p); 767 p->sched_class->enqueue_task(rq, p, flags); 768} 769 770static void dequeue_task(struct rq *rq, struct task_struct *p, int flags) 771{ 772 update_rq_clock(rq); 773 sched_info_dequeued(rq, p); 774 p->sched_class->dequeue_task(rq, p, flags); 775} 776 777void activate_task(struct rq *rq, struct task_struct *p, int flags) 778{ 779 if (task_contributes_to_load(p)) 780 rq->nr_uninterruptible--; 781 782 enqueue_task(rq, p, flags); 783} 784 785void deactivate_task(struct rq *rq, struct task_struct *p, int flags) 786{ 787 if (task_contributes_to_load(p)) 788 rq->nr_uninterruptible++; 789 790 dequeue_task(rq, p, flags); 791} 792 793static void update_rq_clock_task(struct rq *rq, s64 delta) 794{ 795/* 796 * In theory, the compile should just see 0 here, and optimize out the call 797 * to sched_rt_avg_update. But I don't trust it... 798 */ 799#if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING) 800 s64 steal = 0, irq_delta = 0; 801#endif 802#ifdef CONFIG_IRQ_TIME_ACCOUNTING 803 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time; 804 805 /* 806 * Since irq_time is only updated on {soft,}irq_exit, we might run into 807 * this case when a previous update_rq_clock() happened inside a 808 * {soft,}irq region. 809 * 810 * When this happens, we stop ->clock_task and only update the 811 * prev_irq_time stamp to account for the part that fit, so that a next 812 * update will consume the rest. This ensures ->clock_task is 813 * monotonic. 814 * 815 * It does however cause some slight miss-attribution of {soft,}irq 816 * time, a more accurate solution would be to update the irq_time using 817 * the current rq->clock timestamp, except that would require using 818 * atomic ops. 819 */ 820 if (irq_delta > delta) 821 irq_delta = delta; 822 823 rq->prev_irq_time += irq_delta; 824 delta -= irq_delta; 825#endif 826#ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING 827 if (static_key_false((¶virt_steal_rq_enabled))) { 828 u64 st; 829 830 steal = paravirt_steal_clock(cpu_of(rq)); 831 steal -= rq->prev_steal_time_rq; 832 833 if (unlikely(steal > delta)) 834 steal = delta; 835 836 st = steal_ticks(steal); 837 steal = st * TICK_NSEC; 838 839 rq->prev_steal_time_rq += steal; 840 841 delta -= steal; 842 } 843#endif 844 845 rq->clock_task += delta; 846 847#if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING) 848 if ((irq_delta + steal) && sched_feat(NONTASK_POWER)) 849 sched_rt_avg_update(rq, irq_delta + steal); 850#endif 851} 852 853void sched_set_stop_task(int cpu, struct task_struct *stop) 854{ 855 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 }; 856 struct task_struct *old_stop = cpu_rq(cpu)->stop; 857 858 if (stop) { 859 /* 860 * Make it appear like a SCHED_FIFO task, its something 861 * userspace knows about and won't get confused about. 862 * 863 * Also, it will make PI more or less work without too 864 * much confusion -- but then, stop work should not 865 * rely on PI working anyway. 866 */ 867 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m); 868 869 stop->sched_class = &stop_sched_class; 870 } 871 872 cpu_rq(cpu)->stop = stop; 873 874 if (old_stop) { 875 /* 876 * Reset it back to a normal scheduling class so that 877 * it can die in pieces. 878 */ 879 old_stop->sched_class = &rt_sched_class; 880 } 881} 882 883/* 884 * __normal_prio - return the priority that is based on the static prio 885 */ 886static inline int __normal_prio(struct task_struct *p) 887{ 888 return p->static_prio; 889} 890 891/* 892 * Calculate the expected normal priority: i.e. priority 893 * without taking RT-inheritance into account. Might be 894 * boosted by interactivity modifiers. Changes upon fork, 895 * setprio syscalls, and whenever the interactivity 896 * estimator recalculates. 897 */ 898static inline int normal_prio(struct task_struct *p) 899{ 900 int prio; 901 902 if (task_has_rt_policy(p)) 903 prio = MAX_RT_PRIO-1 - p->rt_priority; 904 else 905 prio = __normal_prio(p); 906 return prio; 907} 908 909/* 910 * Calculate the current priority, i.e. the priority 911 * taken into account by the scheduler. This value might 912 * be boosted by RT tasks, or might be boosted by 913 * interactivity modifiers. Will be RT if the task got 914 * RT-boosted. If not then it returns p->normal_prio. 915 */ 916static int effective_prio(struct task_struct *p) 917{ 918 p->normal_prio = normal_prio(p); 919 /* 920 * If we are RT tasks or we were boosted to RT priority, 921 * keep the priority unchanged. Otherwise, update priority 922 * to the normal priority: 923 */ 924 if (!rt_prio(p->prio)) 925 return p->normal_prio; 926 return p->prio; 927} 928 929/** 930 * task_curr - is this task currently executing on a CPU? 931 * @p: the task in question. 932 * 933 * Return: 1 if the task is currently executing. 0 otherwise. 934 */ 935inline int task_curr(const struct task_struct *p) 936{ 937 return cpu_curr(task_cpu(p)) == p; 938} 939 940static inline void check_class_changed(struct rq *rq, struct task_struct *p, 941 const struct sched_class *prev_class, 942 int oldprio) 943{ 944 if (prev_class != p->sched_class) { 945 if (prev_class->switched_from) 946 prev_class->switched_from(rq, p); 947 p->sched_class->switched_to(rq, p); 948 } else if (oldprio != p->prio) 949 p->sched_class->prio_changed(rq, p, oldprio); 950} 951 952void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags) 953{ 954 const struct sched_class *class; 955 956 if (p->sched_class == rq->curr->sched_class) { 957 rq->curr->sched_class->check_preempt_curr(rq, p, flags); 958 } else { 959 for_each_class(class) { 960 if (class == rq->curr->sched_class) 961 break; 962 if (class == p->sched_class) { 963 resched_task(rq->curr); 964 break; 965 } 966 } 967 } 968 969 /* 970 * A queue event has occurred, and we're going to schedule. In 971 * this case, we can save a useless back to back clock update. 972 */ 973 if (rq->curr->on_rq && test_tsk_need_resched(rq->curr)) 974 rq->skip_clock_update = 1; 975} 976 977#ifdef CONFIG_SMP 978void set_task_cpu(struct task_struct *p, unsigned int new_cpu) 979{ 980#ifdef CONFIG_SCHED_DEBUG 981 /* 982 * We should never call set_task_cpu() on a blocked task, 983 * ttwu() will sort out the placement. 984 */ 985 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING && 986 !(task_preempt_count(p) & PREEMPT_ACTIVE)); 987 988#ifdef CONFIG_LOCKDEP 989 /* 990 * The caller should hold either p->pi_lock or rq->lock, when changing 991 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks. 992 * 993 * sched_move_task() holds both and thus holding either pins the cgroup, 994 * see task_group(). 995 * 996 * Furthermore, all task_rq users should acquire both locks, see 997 * task_rq_lock(). 998 */ 999 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) || 1000 lockdep_is_held(&task_rq(p)->lock))); 1001#endif 1002#endif 1003 1004 trace_sched_migrate_task(p, new_cpu); 1005 1006 if (task_cpu(p) != new_cpu) { 1007 if (p->sched_class->migrate_task_rq) 1008 p->sched_class->migrate_task_rq(p, new_cpu); 1009 p->se.nr_migrations++; 1010 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, NULL, 0); 1011 } 1012 1013 __set_task_cpu(p, new_cpu); 1014} 1015 1016static void __migrate_swap_task(struct task_struct *p, int cpu) 1017{ 1018 if (p->on_rq) { 1019 struct rq *src_rq, *dst_rq; 1020 1021 src_rq = task_rq(p); 1022 dst_rq = cpu_rq(cpu); 1023 1024 deactivate_task(src_rq, p, 0); 1025 set_task_cpu(p, cpu); 1026 activate_task(dst_rq, p, 0); 1027 check_preempt_curr(dst_rq, p, 0); 1028 } else { 1029 /* 1030 * Task isn't running anymore; make it appear like we migrated 1031 * it before it went to sleep. This means on wakeup we make the 1032 * previous cpu our targer instead of where it really is. 1033 */ 1034 p->wake_cpu = cpu; 1035 } 1036} 1037 1038struct migration_swap_arg { 1039 struct task_struct *src_task, *dst_task; 1040 int src_cpu, dst_cpu; 1041}; 1042 1043static int migrate_swap_stop(void *data) 1044{ 1045 struct migration_swap_arg *arg = data; 1046 struct rq *src_rq, *dst_rq; 1047 int ret = -EAGAIN; 1048 1049 src_rq = cpu_rq(arg->src_cpu); 1050 dst_rq = cpu_rq(arg->dst_cpu); 1051 1052 double_rq_lock(src_rq, dst_rq); 1053 if (task_cpu(arg->dst_task) != arg->dst_cpu) 1054 goto unlock; 1055 1056 if (task_cpu(arg->src_task) != arg->src_cpu) 1057 goto unlock; 1058 1059 if (!cpumask_test_cpu(arg->dst_cpu, tsk_cpus_allowed(arg->src_task))) 1060 goto unlock; 1061 1062 if (!cpumask_test_cpu(arg->src_cpu, tsk_cpus_allowed(arg->dst_task))) 1063 goto unlock; 1064 1065 __migrate_swap_task(arg->src_task, arg->dst_cpu); 1066 __migrate_swap_task(arg->dst_task, arg->src_cpu); 1067 1068 ret = 0; 1069 1070unlock: 1071 double_rq_unlock(src_rq, dst_rq); 1072 1073 return ret; 1074} 1075 1076/* 1077 * Cross migrate two tasks 1078 */ 1079int migrate_swap(struct task_struct *cur, struct task_struct *p) 1080{ 1081 struct migration_swap_arg arg; 1082 int ret = -EINVAL; 1083 1084 get_online_cpus(); 1085 1086 arg = (struct migration_swap_arg){ 1087 .src_task = cur, 1088 .src_cpu = task_cpu(cur), 1089 .dst_task = p, 1090 .dst_cpu = task_cpu(p), 1091 }; 1092 1093 if (arg.src_cpu == arg.dst_cpu) 1094 goto out; 1095 1096 if (!cpu_active(arg.src_cpu) || !cpu_active(arg.dst_cpu)) 1097 goto out; 1098 1099 if (!cpumask_test_cpu(arg.dst_cpu, tsk_cpus_allowed(arg.src_task))) 1100 goto out; 1101 1102 if (!cpumask_test_cpu(arg.src_cpu, tsk_cpus_allowed(arg.dst_task))) 1103 goto out; 1104 1105 ret = stop_two_cpus(arg.dst_cpu, arg.src_cpu, migrate_swap_stop, &arg); 1106 1107out: 1108 put_online_cpus(); 1109 return ret; 1110} 1111 1112struct migration_arg { 1113 struct task_struct *task; 1114 int dest_cpu; 1115}; 1116 1117static int migration_cpu_stop(void *data); 1118 1119/* 1120 * wait_task_inactive - wait for a thread to unschedule. 1121 * 1122 * If @match_state is nonzero, it's the @p->state value just checked and 1123 * not expected to change. If it changes, i.e. @p might have woken up, 1124 * then return zero. When we succeed in waiting for @p to be off its CPU, 1125 * we return a positive number (its total switch count). If a second call 1126 * a short while later returns the same number, the caller can be sure that 1127 * @p has remained unscheduled the whole time. 1128 * 1129 * The caller must ensure that the task *will* unschedule sometime soon, 1130 * else this function might spin for a *long* time. This function can't 1131 * be called with interrupts off, or it may introduce deadlock with 1132 * smp_call_function() if an IPI is sent by the same process we are 1133 * waiting to become inactive. 1134 */ 1135unsigned long wait_task_inactive(struct task_struct *p, long match_state) 1136{ 1137 unsigned long flags; 1138 int running, on_rq; 1139 unsigned long ncsw; 1140 struct rq *rq; 1141 1142 for (;;) { 1143 /* 1144 * We do the initial early heuristics without holding 1145 * any task-queue locks at all. We'll only try to get 1146 * the runqueue lock when things look like they will 1147 * work out! 1148 */ 1149 rq = task_rq(p); 1150 1151 /* 1152 * If the task is actively running on another CPU 1153 * still, just relax and busy-wait without holding 1154 * any locks. 1155 * 1156 * NOTE! Since we don't hold any locks, it's not 1157 * even sure that "rq" stays as the right runqueue! 1158 * But we don't care, since "task_running()" will 1159 * return false if the runqueue has changed and p 1160 * is actually now running somewhere else! 1161 */ 1162 while (task_running(rq, p)) { 1163 if (match_state && unlikely(p->state != match_state)) 1164 return 0; 1165 cpu_relax(); 1166 } 1167 1168 /* 1169 * Ok, time to look more closely! We need the rq 1170 * lock now, to be *sure*. If we're wrong, we'll 1171 * just go back and repeat. 1172 */ 1173 rq = task_rq_lock(p, &flags); 1174 trace_sched_wait_task(p); 1175 running = task_running(rq, p); 1176 on_rq = p->on_rq; 1177 ncsw = 0; 1178 if (!match_state || p->state == match_state) 1179 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */ 1180 task_rq_unlock(rq, p, &flags); 1181 1182 /* 1183 * If it changed from the expected state, bail out now. 1184 */ 1185 if (unlikely(!ncsw)) 1186 break; 1187 1188 /* 1189 * Was it really running after all now that we 1190 * checked with the proper locks actually held? 1191 * 1192 * Oops. Go back and try again.. 1193 */ 1194 if (unlikely(running)) { 1195 cpu_relax(); 1196 continue; 1197 } 1198 1199 /* 1200 * It's not enough that it's not actively running, 1201 * it must be off the runqueue _entirely_, and not 1202 * preempted! 1203 * 1204 * So if it was still runnable (but just not actively 1205 * running right now), it's preempted, and we should 1206 * yield - it could be a while. 1207 */ 1208 if (unlikely(on_rq)) { 1209 ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ); 1210 1211 set_current_state(TASK_UNINTERRUPTIBLE); 1212 schedule_hrtimeout(&to, HRTIMER_MODE_REL); 1213 continue; 1214 } 1215 1216 /* 1217 * Ahh, all good. It wasn't running, and it wasn't 1218 * runnable, which means that it will never become 1219 * running in the future either. We're all done! 1220 */ 1221 break; 1222 } 1223 1224 return ncsw; 1225} 1226 1227/*** 1228 * kick_process - kick a running thread to enter/exit the kernel 1229 * @p: the to-be-kicked thread 1230 * 1231 * Cause a process which is running on another CPU to enter 1232 * kernel-mode, without any delay. (to get signals handled.) 1233 * 1234 * NOTE: this function doesn't have to take the runqueue lock, 1235 * because all it wants to ensure is that the remote task enters 1236 * the kernel. If the IPI races and the task has been migrated 1237 * to another CPU then no harm is done and the purpose has been 1238 * achieved as well. 1239 */ 1240void kick_process(struct task_struct *p) 1241{ 1242 int cpu; 1243 1244 preempt_disable(); 1245 cpu = task_cpu(p); 1246 if ((cpu != smp_processor_id()) && task_curr(p)) 1247 smp_send_reschedule(cpu); 1248 preempt_enable(); 1249} 1250EXPORT_SYMBOL_GPL(kick_process); 1251#endif /* CONFIG_SMP */ 1252 1253#ifdef CONFIG_SMP 1254/* 1255 * ->cpus_allowed is protected by both rq->lock and p->pi_lock 1256 */ 1257static int select_fallback_rq(int cpu, struct task_struct *p) 1258{ 1259 int nid = cpu_to_node(cpu); 1260 const struct cpumask *nodemask = NULL; 1261 enum { cpuset, possible, fail } state = cpuset; 1262 int dest_cpu; 1263 1264 /* 1265 * If the node that the cpu is on has been offlined, cpu_to_node() 1266 * will return -1. There is no cpu on the node, and we should 1267 * select the cpu on the other node. 1268 */ 1269 if (nid != -1) { 1270 nodemask = cpumask_of_node(nid); 1271 1272 /* Look for allowed, online CPU in same node. */ 1273 for_each_cpu(dest_cpu, nodemask) { 1274 if (!cpu_online(dest_cpu)) 1275 continue; 1276 if (!cpu_active(dest_cpu)) 1277 continue; 1278 if (cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p))) 1279 return dest_cpu; 1280 } 1281 } 1282 1283 for (;;) { 1284 /* Any allowed, online CPU? */ 1285 for_each_cpu(dest_cpu, tsk_cpus_allowed(p)) { 1286 if (!cpu_online(dest_cpu)) 1287 continue; 1288 if (!cpu_active(dest_cpu)) 1289 continue; 1290 goto out; 1291 } 1292 1293 switch (state) { 1294 case cpuset: 1295 /* No more Mr. Nice Guy. */ 1296 cpuset_cpus_allowed_fallback(p); 1297 state = possible; 1298 break; 1299 1300 case possible: 1301 do_set_cpus_allowed(p, cpu_possible_mask); 1302 state = fail; 1303 break; 1304 1305 case fail: 1306 BUG(); 1307 break; 1308 } 1309 } 1310 1311out: 1312 if (state != cpuset) { 1313 /* 1314 * Don't tell them about moving exiting tasks or 1315 * kernel threads (both mm NULL), since they never 1316 * leave kernel. 1317 */ 1318 if (p->mm && printk_ratelimit()) { 1319 printk_sched("process %d (%s) no longer affine to cpu%d\n", 1320 task_pid_nr(p), p->comm, cpu); 1321 } 1322 } 1323 1324 return dest_cpu; 1325} 1326 1327/* 1328 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable. 1329 */ 1330static inline 1331int select_task_rq(struct task_struct *p, int cpu, int sd_flags, int wake_flags) 1332{ 1333 cpu = p->sched_class->select_task_rq(p, cpu, sd_flags, wake_flags); 1334 1335 /* 1336 * In order not to call set_task_cpu() on a blocking task we need 1337 * to rely on ttwu() to place the task on a valid ->cpus_allowed 1338 * cpu. 1339 * 1340 * Since this is common to all placement strategies, this lives here. 1341 * 1342 * [ this allows ->select_task() to simply return task_cpu(p) and 1343 * not worry about this generic constraint ] 1344 */ 1345 if (unlikely(!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)) || 1346 !cpu_online(cpu))) 1347 cpu = select_fallback_rq(task_cpu(p), p); 1348 1349 return cpu; 1350} 1351 1352static void update_avg(u64 *avg, u64 sample) 1353{ 1354 s64 diff = sample - *avg; 1355 *avg += diff >> 3; 1356} 1357#endif 1358 1359static void 1360ttwu_stat(struct task_struct *p, int cpu, int wake_flags) 1361{ 1362#ifdef CONFIG_SCHEDSTATS 1363 struct rq *rq = this_rq(); 1364 1365#ifdef CONFIG_SMP 1366 int this_cpu = smp_processor_id(); 1367 1368 if (cpu == this_cpu) { 1369 schedstat_inc(rq, ttwu_local); 1370 schedstat_inc(p, se.statistics.nr_wakeups_local); 1371 } else { 1372 struct sched_domain *sd; 1373 1374 schedstat_inc(p, se.statistics.nr_wakeups_remote); 1375 rcu_read_lock(); 1376 for_each_domain(this_cpu, sd) { 1377 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) { 1378 schedstat_inc(sd, ttwu_wake_remote); 1379 break; 1380 } 1381 } 1382 rcu_read_unlock(); 1383 } 1384 1385 if (wake_flags & WF_MIGRATED) 1386 schedstat_inc(p, se.statistics.nr_wakeups_migrate); 1387 1388#endif /* CONFIG_SMP */ 1389 1390 schedstat_inc(rq, ttwu_count); 1391 schedstat_inc(p, se.statistics.nr_wakeups); 1392 1393 if (wake_flags & WF_SYNC) 1394 schedstat_inc(p, se.statistics.nr_wakeups_sync); 1395 1396#endif /* CONFIG_SCHEDSTATS */ 1397} 1398 1399static void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags) 1400{ 1401 activate_task(rq, p, en_flags); 1402 p->on_rq = 1; 1403 1404 /* if a worker is waking up, notify workqueue */ 1405 if (p->flags & PF_WQ_WORKER) 1406 wq_worker_waking_up(p, cpu_of(rq)); 1407} 1408 1409/* 1410 * Mark the task runnable and perform wakeup-preemption. 1411 */ 1412static void 1413ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags) 1414{ 1415 check_preempt_curr(rq, p, wake_flags); 1416 trace_sched_wakeup(p, true); 1417 1418 p->state = TASK_RUNNING; 1419#ifdef CONFIG_SMP 1420 if (p->sched_class->task_woken) 1421 p->sched_class->task_woken(rq, p); 1422 1423 if (rq->idle_stamp) { 1424 u64 delta = rq_clock(rq) - rq->idle_stamp; 1425 u64 max = 2*rq->max_idle_balance_cost; 1426 1427 update_avg(&rq->avg_idle, delta); 1428 1429 if (rq->avg_idle > max) 1430 rq->avg_idle = max; 1431 1432 rq->idle_stamp = 0; 1433 } 1434#endif 1435} 1436 1437static void 1438ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags) 1439{ 1440#ifdef CONFIG_SMP 1441 if (p->sched_contributes_to_load) 1442 rq->nr_uninterruptible--; 1443#endif 1444 1445 ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_WAKING); 1446 ttwu_do_wakeup(rq, p, wake_flags); 1447} 1448 1449/* 1450 * Called in case the task @p isn't fully descheduled from its runqueue, 1451 * in this case we must do a remote wakeup. Its a 'light' wakeup though, 1452 * since all we need to do is flip p->state to TASK_RUNNING, since 1453 * the task is still ->on_rq. 1454 */ 1455static int ttwu_remote(struct task_struct *p, int wake_flags) 1456{ 1457 struct rq *rq; 1458 int ret = 0; 1459 1460 rq = __task_rq_lock(p); 1461 if (p->on_rq) { 1462 /* check_preempt_curr() may use rq clock */ 1463 update_rq_clock(rq); 1464 ttwu_do_wakeup(rq, p, wake_flags); 1465 ret = 1; 1466 } 1467 __task_rq_unlock(rq); 1468 1469 return ret; 1470} 1471 1472#ifdef CONFIG_SMP 1473static void sched_ttwu_pending(void) 1474{ 1475 struct rq *rq = this_rq(); 1476 struct llist_node *llist = llist_del_all(&rq->wake_list); 1477 struct task_struct *p; 1478 1479 raw_spin_lock(&rq->lock); 1480 1481 while (llist) { 1482 p = llist_entry(llist, struct task_struct, wake_entry); 1483 llist = llist_next(llist); 1484 ttwu_do_activate(rq, p, 0); 1485 } 1486 1487 raw_spin_unlock(&rq->lock); 1488} 1489 1490void scheduler_ipi(void) 1491{ 1492 /* 1493 * Fold TIF_NEED_RESCHED into the preempt_count; anybody setting 1494 * TIF_NEED_RESCHED remotely (for the first time) will also send 1495 * this IPI. 1496 */ 1497 if (tif_need_resched()) 1498 set_preempt_need_resched(); 1499 1500 if (llist_empty(&this_rq()->wake_list) 1501 && !tick_nohz_full_cpu(smp_processor_id()) 1502 && !got_nohz_idle_kick()) 1503 return; 1504 1505 /* 1506 * Not all reschedule IPI handlers call irq_enter/irq_exit, since 1507 * traditionally all their work was done from the interrupt return 1508 * path. Now that we actually do some work, we need to make sure 1509 * we do call them. 1510 * 1511 * Some archs already do call them, luckily irq_enter/exit nest 1512 * properly. 1513 * 1514 * Arguably we should visit all archs and update all handlers, 1515 * however a fair share of IPIs are still resched only so this would 1516 * somewhat pessimize the simple resched case. 1517 */ 1518 irq_enter(); 1519 tick_nohz_full_check(); 1520 sched_ttwu_pending(); 1521 1522 /* 1523 * Check if someone kicked us for doing the nohz idle load balance. 1524 */ 1525 if (unlikely(got_nohz_idle_kick())) { 1526 this_rq()->idle_balance = 1; 1527 raise_softirq_irqoff(SCHED_SOFTIRQ); 1528 } 1529 irq_exit(); 1530} 1531 1532static void ttwu_queue_remote(struct task_struct *p, int cpu) 1533{ 1534 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list)) 1535 smp_send_reschedule(cpu); 1536} 1537 1538bool cpus_share_cache(int this_cpu, int that_cpu) 1539{ 1540 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu); 1541} 1542#endif /* CONFIG_SMP */ 1543 1544static void ttwu_queue(struct task_struct *p, int cpu) 1545{ 1546 struct rq *rq = cpu_rq(cpu); 1547 1548#if defined(CONFIG_SMP) 1549 if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) { 1550 sched_clock_cpu(cpu); /* sync clocks x-cpu */ 1551 ttwu_queue_remote(p, cpu); 1552 return; 1553 } 1554#endif 1555 1556 raw_spin_lock(&rq->lock); 1557 ttwu_do_activate(rq, p, 0); 1558 raw_spin_unlock(&rq->lock); 1559} 1560 1561/** 1562 * try_to_wake_up - wake up a thread 1563 * @p: the thread to be awakened 1564 * @state: the mask of task states that can be woken 1565 * @wake_flags: wake modifier flags (WF_*) 1566 * 1567 * Put it on the run-queue if it's not already there. The "current" 1568 * thread is always on the run-queue (except when the actual 1569 * re-schedule is in progress), and as such you're allowed to do 1570 * the simpler "current->state = TASK_RUNNING" to mark yourself 1571 * runnable without the overhead of this. 1572 * 1573 * Return: %true if @p was woken up, %false if it was already running. 1574 * or @state didn't match @p's state. 1575 */ 1576static int 1577try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags) 1578{ 1579 unsigned long flags; 1580 int cpu, success = 0; 1581 1582 /* 1583 * If we are going to wake up a thread waiting for CONDITION we 1584 * need to ensure that CONDITION=1 done by the caller can not be 1585 * reordered with p->state check below. This pairs with mb() in 1586 * set_current_state() the waiting thread does. 1587 */ 1588 smp_mb__before_spinlock(); 1589 raw_spin_lock_irqsave(&p->pi_lock, flags); 1590 if (!(p->state & state)) 1591 goto out; 1592 1593 success = 1; /* we're going to change ->state */ 1594 cpu = task_cpu(p); 1595 1596 if (p->on_rq && ttwu_remote(p, wake_flags)) 1597 goto stat; 1598 1599#ifdef CONFIG_SMP 1600 /* 1601 * If the owning (remote) cpu is still in the middle of schedule() with 1602 * this task as prev, wait until its done referencing the task. 1603 */ 1604 while (p->on_cpu) 1605 cpu_relax(); 1606 /* 1607 * Pairs with the smp_wmb() in finish_lock_switch(). 1608 */ 1609 smp_rmb(); 1610 1611 p->sched_contributes_to_load = !!task_contributes_to_load(p); 1612 p->state = TASK_WAKING; 1613 1614 if (p->sched_class->task_waking) 1615 p->sched_class->task_waking(p); 1616 1617 cpu = select_task_rq(p, p->wake_cpu, SD_BALANCE_WAKE, wake_flags); 1618 if (task_cpu(p) != cpu) { 1619 wake_flags |= WF_MIGRATED; 1620 set_task_cpu(p, cpu); 1621 } 1622#endif /* CONFIG_SMP */ 1623 1624 ttwu_queue(p, cpu); 1625stat: 1626 ttwu_stat(p, cpu, wake_flags); 1627out: 1628 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 1629 1630 return success; 1631} 1632 1633/** 1634 * try_to_wake_up_local - try to wake up a local task with rq lock held 1635 * @p: the thread to be awakened 1636 * 1637 * Put @p on the run-queue if it's not already there. The caller must 1638 * ensure that this_rq() is locked, @p is bound to this_rq() and not 1639 * the current task. 1640 */ 1641static void try_to_wake_up_local(struct task_struct *p) 1642{ 1643 struct rq *rq = task_rq(p); 1644 1645 if (WARN_ON_ONCE(rq != this_rq()) || 1646 WARN_ON_ONCE(p == current)) 1647 return; 1648 1649 lockdep_assert_held(&rq->lock); 1650 1651 if (!raw_spin_trylock(&p->pi_lock)) { 1652 raw_spin_unlock(&rq->lock); 1653 raw_spin_lock(&p->pi_lock); 1654 raw_spin_lock(&rq->lock); 1655 } 1656 1657 if (!(p->state & TASK_NORMAL)) 1658 goto out; 1659 1660 if (!p->on_rq) 1661 ttwu_activate(rq, p, ENQUEUE_WAKEUP); 1662 1663 ttwu_do_wakeup(rq, p, 0); 1664 ttwu_stat(p, smp_processor_id(), 0); 1665out: 1666 raw_spin_unlock(&p->pi_lock); 1667} 1668 1669/** 1670 * wake_up_process - Wake up a specific process 1671 * @p: The process to be woken up. 1672 * 1673 * Attempt to wake up the nominated process and move it to the set of runnable 1674 * processes. 1675 * 1676 * Return: 1 if the process was woken up, 0 if it was already running. 1677 * 1678 * It may be assumed that this function implies a write memory barrier before 1679 * changing the task state if and only if any tasks are woken up. 1680 */ 1681int wake_up_process(struct task_struct *p) 1682{ 1683 WARN_ON(task_is_stopped_or_traced(p)); 1684 return try_to_wake_up(p, TASK_NORMAL, 0); 1685} 1686EXPORT_SYMBOL(wake_up_process); 1687 1688int wake_up_state(struct task_struct *p, unsigned int state) 1689{ 1690 return try_to_wake_up(p, state, 0); 1691} 1692 1693/* 1694 * Perform scheduler related setup for a newly forked process p. 1695 * p is forked by current. 1696 * 1697 * __sched_fork() is basic setup used by init_idle() too: 1698 */ 1699static void __sched_fork(unsigned long clone_flags, struct task_struct *p) 1700{ 1701 p->on_rq = 0; 1702 1703 p->se.on_rq = 0; 1704 p->se.exec_start = 0; 1705 p->se.sum_exec_runtime = 0; 1706 p->se.prev_sum_exec_runtime = 0; 1707 p->se.nr_migrations = 0; 1708 p->se.vruntime = 0; 1709 INIT_LIST_HEAD(&p->se.group_node); 1710 1711#ifdef CONFIG_SCHEDSTATS 1712 memset(&p->se.statistics, 0, sizeof(p->se.statistics)); 1713#endif 1714 1715 INIT_LIST_HEAD(&p->rt.run_list); 1716 1717#ifdef CONFIG_PREEMPT_NOTIFIERS 1718 INIT_HLIST_HEAD(&p->preempt_notifiers); 1719#endif 1720 1721#ifdef CONFIG_NUMA_BALANCING 1722 if (p->mm && atomic_read(&p->mm->mm_users) == 1) { 1723 p->mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay); 1724 p->mm->numa_next_reset = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_period_reset); 1725 p->mm->numa_scan_seq = 0; 1726 } 1727 1728 if (clone_flags & CLONE_VM) 1729 p->numa_preferred_nid = current->numa_preferred_nid; 1730 else 1731 p->numa_preferred_nid = -1; 1732 1733 p->node_stamp = 0ULL; 1734 p->numa_scan_seq = p->mm ? p->mm->numa_scan_seq : 0; 1735 p->numa_migrate_seq = 1; 1736 p->numa_scan_period = sysctl_numa_balancing_scan_delay; 1737 p->numa_work.next = &p->numa_work; 1738 p->numa_faults = NULL; 1739 p->numa_faults_buffer = NULL; 1740 1741 INIT_LIST_HEAD(&p->numa_entry); 1742 p->numa_group = NULL; 1743#endif /* CONFIG_NUMA_BALANCING */ 1744} 1745 1746#ifdef CONFIG_NUMA_BALANCING 1747#ifdef CONFIG_SCHED_DEBUG 1748void set_numabalancing_state(bool enabled) 1749{ 1750 if (enabled) 1751 sched_feat_set("NUMA"); 1752 else 1753 sched_feat_set("NO_NUMA"); 1754} 1755#else 1756__read_mostly bool numabalancing_enabled; 1757 1758void set_numabalancing_state(bool enabled) 1759{ 1760 numabalancing_enabled = enabled; 1761} 1762#endif /* CONFIG_SCHED_DEBUG */ 1763#endif /* CONFIG_NUMA_BALANCING */ 1764 1765/* 1766 * fork()/clone()-time setup: 1767 */ 1768void sched_fork(unsigned long clone_flags, struct task_struct *p) 1769{ 1770 unsigned long flags; 1771 int cpu = get_cpu(); 1772 1773 __sched_fork(clone_flags, p); 1774 /* 1775 * We mark the process as running here. This guarantees that 1776 * nobody will actually run it, and a signal or other external 1777 * event cannot wake it up and insert it on the runqueue either. 1778 */ 1779 p->state = TASK_RUNNING; 1780 1781 /* 1782 * Make sure we do not leak PI boosting priority to the child. 1783 */ 1784 p->prio = current->normal_prio; 1785 1786 /* 1787 * Revert to default priority/policy on fork if requested. 1788 */ 1789 if (unlikely(p->sched_reset_on_fork)) { 1790 if (task_has_rt_policy(p)) { 1791 p->policy = SCHED_NORMAL; 1792 p->static_prio = NICE_TO_PRIO(0); 1793 p->rt_priority = 0; 1794 } else if (PRIO_TO_NICE(p->static_prio) < 0) 1795 p->static_prio = NICE_TO_PRIO(0); 1796 1797 p->prio = p->normal_prio = __normal_prio(p); 1798 set_load_weight(p); 1799 1800 /* 1801 * We don't need the reset flag anymore after the fork. It has 1802 * fulfilled its duty: 1803 */ 1804 p->sched_reset_on_fork = 0; 1805 } 1806 1807 if (!rt_prio(p->prio)) 1808 p->sched_class = &fair_sched_class; 1809 1810 if (p->sched_class->task_fork) 1811 p->sched_class->task_fork(p); 1812 1813 /* 1814 * The child is not yet in the pid-hash so no cgroup attach races, 1815 * and the cgroup is pinned to this child due to cgroup_fork() 1816 * is ran before sched_fork(). 1817 * 1818 * Silence PROVE_RCU. 1819 */ 1820 raw_spin_lock_irqsave(&p->pi_lock, flags); 1821 set_task_cpu(p, cpu); 1822 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 1823 1824#if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT) 1825 if (likely(sched_info_on())) 1826 memset(&p->sched_info, 0, sizeof(p->sched_info)); 1827#endif 1828#if defined(CONFIG_SMP) 1829 p->on_cpu = 0; 1830#endif 1831 init_task_preempt_count(p); 1832#ifdef CONFIG_SMP 1833 plist_node_init(&p->pushable_tasks, MAX_PRIO); 1834#endif 1835 1836 put_cpu(); 1837} 1838 1839/* 1840 * wake_up_new_task - wake up a newly created task for the first time. 1841 * 1842 * This function will do some initial scheduler statistics housekeeping 1843 * that must be done for every newly created context, then puts the task 1844 * on the runqueue and wakes it. 1845 */ 1846void wake_up_new_task(struct task_struct *p) 1847{ 1848 unsigned long flags; 1849 struct rq *rq; 1850 1851 raw_spin_lock_irqsave(&p->pi_lock, flags); 1852#ifdef CONFIG_SMP 1853 /* 1854 * Fork balancing, do it here and not earlier because: 1855 * - cpus_allowed can change in the fork path 1856 * - any previously selected cpu might disappear through hotplug 1857 */ 1858 set_task_cpu(p, select_task_rq(p, task_cpu(p), SD_BALANCE_FORK, 0)); 1859#endif 1860 1861 /* Initialize new task's runnable average */ 1862 init_task_runnable_average(p); 1863 rq = __task_rq_lock(p); 1864 activate_task(rq, p, 0); 1865 p->on_rq = 1; 1866 trace_sched_wakeup_new(p, true); 1867 check_preempt_curr(rq, p, WF_FORK); 1868#ifdef CONFIG_SMP 1869 if (p->sched_class->task_woken) 1870 p->sched_class->task_woken(rq, p); 1871#endif 1872 task_rq_unlock(rq, p, &flags); 1873} 1874 1875#ifdef CONFIG_PREEMPT_NOTIFIERS 1876 1877/** 1878 * preempt_notifier_register - tell me when current is being preempted & rescheduled 1879 * @notifier: notifier struct to register 1880 */ 1881void preempt_notifier_register(struct preempt_notifier *notifier) 1882{ 1883 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers); 1884} 1885EXPORT_SYMBOL_GPL(preempt_notifier_register); 1886 1887/** 1888 * preempt_notifier_unregister - no longer interested in preemption notifications 1889 * @notifier: notifier struct to unregister 1890 * 1891 * This is safe to call from within a preemption notifier. 1892 */ 1893void preempt_notifier_unregister(struct preempt_notifier *notifier) 1894{ 1895 hlist_del(¬ifier->link); 1896} 1897EXPORT_SYMBOL_GPL(preempt_notifier_unregister); 1898 1899static void fire_sched_in_preempt_notifiers(struct task_struct *curr) 1900{ 1901 struct preempt_notifier *notifier; 1902 1903 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link) 1904 notifier->ops->sched_in(notifier, raw_smp_processor_id()); 1905} 1906 1907static void 1908fire_sched_out_preempt_notifiers(struct task_struct *curr, 1909 struct task_struct *next) 1910{ 1911 struct preempt_notifier *notifier; 1912 1913 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link) 1914 notifier->ops->sched_out(notifier, next); 1915} 1916 1917#else /* !CONFIG_PREEMPT_NOTIFIERS */ 1918 1919static void fire_sched_in_preempt_notifiers(struct task_struct *curr) 1920{ 1921} 1922 1923static void 1924fire_sched_out_preempt_notifiers(struct task_struct *curr, 1925 struct task_struct *next) 1926{ 1927} 1928 1929#endif /* CONFIG_PREEMPT_NOTIFIERS */ 1930 1931/** 1932 * prepare_task_switch - prepare to switch tasks 1933 * @rq: the runqueue preparing to switch 1934 * @prev: the current task that is being switched out 1935 * @next: the task we are going to switch to. 1936 * 1937 * This is called with the rq lock held and interrupts off. It must 1938 * be paired with a subsequent finish_task_switch after the context 1939 * switch. 1940 * 1941 * prepare_task_switch sets up locking and calls architecture specific 1942 * hooks. 1943 */ 1944static inline void 1945prepare_task_switch(struct rq *rq, struct task_struct *prev, 1946 struct task_struct *next) 1947{ 1948 trace_sched_switch(prev, next); 1949 sched_info_switch(rq, prev, next); 1950 perf_event_task_sched_out(prev, next); 1951 fire_sched_out_preempt_notifiers(prev, next); 1952 prepare_lock_switch(rq, next); 1953 prepare_arch_switch(next); 1954} 1955 1956/** 1957 * finish_task_switch - clean up after a task-switch 1958 * @rq: runqueue associated with task-switch 1959 * @prev: the thread we just switched away from. 1960 * 1961 * finish_task_switch must be called after the context switch, paired 1962 * with a prepare_task_switch call before the context switch. 1963 * finish_task_switch will reconcile locking set up by prepare_task_switch, 1964 * and do any other architecture-specific cleanup actions. 1965 * 1966 * Note that we may have delayed dropping an mm in context_switch(). If 1967 * so, we finish that here outside of the runqueue lock. (Doing it 1968 * with the lock held can cause deadlocks; see schedule() for 1969 * details.) 1970 */ 1971static void finish_task_switch(struct rq *rq, struct task_struct *prev) 1972 __releases(rq->lock) 1973{ 1974 struct mm_struct *mm = rq->prev_mm; 1975 long prev_state; 1976 1977 rq->prev_mm = NULL; 1978 1979 /* 1980 * A task struct has one reference for the use as "current". 1981 * If a task dies, then it sets TASK_DEAD in tsk->state and calls 1982 * schedule one last time. The schedule call will never return, and 1983 * the scheduled task must drop that reference. 1984 * The test for TASK_DEAD must occur while the runqueue locks are 1985 * still held, otherwise prev could be scheduled on another cpu, die 1986 * there before we look at prev->state, and then the reference would 1987 * be dropped twice. 1988 * Manfred Spraul <manfred@colorfullife.com> 1989 */ 1990 prev_state = prev->state; 1991 vtime_task_switch(prev); 1992 finish_arch_switch(prev); 1993 perf_event_task_sched_in(prev, current); 1994 finish_lock_switch(rq, prev); 1995 finish_arch_post_lock_switch(); 1996 1997 fire_sched_in_preempt_notifiers(current); 1998 if (mm) 1999 mmdrop(mm); 2000 if (unlikely(prev_state == TASK_DEAD)) { 2001 task_numa_free(prev); 2002 2003 /* 2004 * Remove function-return probe instances associated with this 2005 * task and put them back on the free list. 2006 */ 2007 kprobe_flush_task(prev); 2008 put_task_struct(prev); 2009 } 2010 2011 tick_nohz_task_switch(current); 2012} 2013 2014#ifdef CONFIG_SMP 2015 2016/* assumes rq->lock is held */ 2017static inline void pre_schedule(struct rq *rq, struct task_struct *prev) 2018{ 2019 if (prev->sched_class->pre_schedule) 2020 prev->sched_class->pre_schedule(rq, prev); 2021} 2022 2023/* rq->lock is NOT held, but preemption is disabled */ 2024static inline void post_schedule(struct rq *rq) 2025{ 2026 if (rq->post_schedule) { 2027 unsigned long flags; 2028 2029 raw_spin_lock_irqsave(&rq->lock, flags); 2030 if (rq->curr->sched_class->post_schedule) 2031 rq->curr->sched_class->post_schedule(rq); 2032 raw_spin_unlock_irqrestore(&rq->lock, flags); 2033 2034 rq->post_schedule = 0; 2035 } 2036} 2037 2038#else 2039 2040static inline void pre_schedule(struct rq *rq, struct task_struct *p) 2041{ 2042} 2043 2044static inline void post_schedule(struct rq *rq) 2045{ 2046} 2047 2048#endif 2049 2050/** 2051 * schedule_tail - first thing a freshly forked thread must call. 2052 * @prev: the thread we just switched away from. 2053 */ 2054asmlinkage void schedule_tail(struct task_struct *prev) 2055 __releases(rq->lock) 2056{ 2057 struct rq *rq = this_rq(); 2058 2059 finish_task_switch(rq, prev); 2060 2061 /* 2062 * FIXME: do we need to worry about rq being invalidated by the 2063 * task_switch? 2064 */ 2065 post_schedule(rq); 2066 2067#ifdef __ARCH_WANT_UNLOCKED_CTXSW 2068 /* In this case, finish_task_switch does not reenable preemption */ 2069 preempt_enable(); 2070#endif 2071 if (current->set_child_tid) 2072 put_user(task_pid_vnr(current), current->set_child_tid); 2073} 2074 2075/* 2076 * context_switch - switch to the new MM and the new 2077 * thread's register state. 2078 */ 2079static inline void 2080context_switch(struct rq *rq, struct task_struct *prev, 2081 struct task_struct *next) 2082{ 2083 struct mm_struct *mm, *oldmm; 2084 2085 prepare_task_switch(rq, prev, next); 2086 2087 mm = next->mm; 2088 oldmm = prev->active_mm; 2089 /* 2090 * For paravirt, this is coupled with an exit in switch_to to 2091 * combine the page table reload and the switch backend into 2092 * one hypercall. 2093 */ 2094 arch_start_context_switch(prev); 2095 2096 if (!mm) { 2097 next->active_mm = oldmm; 2098 atomic_inc(&oldmm->mm_count); 2099 enter_lazy_tlb(oldmm, next); 2100 } else 2101 switch_mm(oldmm, mm, next); 2102 2103 if (!prev->mm) { 2104 prev->active_mm = NULL; 2105 rq->prev_mm = oldmm; 2106 } 2107 /* 2108 * Since the runqueue lock will be released by the next 2109 * task (which is an invalid locking op but in the case 2110 * of the scheduler it's an obvious special-case), so we 2111 * do an early lockdep release here: 2112 */ 2113#ifndef __ARCH_WANT_UNLOCKED_CTXSW 2114 spin_release(&rq->lock.dep_map, 1, _THIS_IP_); 2115#endif 2116 2117 context_tracking_task_switch(prev, next); 2118 /* Here we just switch the register state and the stack. */ 2119 switch_to(prev, next, prev); 2120 2121 barrier(); 2122 /* 2123 * this_rq must be evaluated again because prev may have moved 2124 * CPUs since it called schedule(), thus the 'rq' on its stack 2125 * frame will be invalid. 2126 */ 2127 finish_task_switch(this_rq(), prev); 2128} 2129 2130/* 2131 * nr_running and nr_context_switches: 2132 * 2133 * externally visible scheduler statistics: current number of runnable 2134 * threads, total number of context switches performed since bootup. 2135 */ 2136unsigned long nr_running(void) 2137{ 2138 unsigned long i, sum = 0; 2139 2140 for_each_online_cpu(i) 2141 sum += cpu_rq(i)->nr_running; 2142 2143 return sum; 2144} 2145 2146unsigned long long nr_context_switches(void) 2147{ 2148 int i; 2149 unsigned long long sum = 0; 2150 2151 for_each_possible_cpu(i) 2152 sum += cpu_rq(i)->nr_switches; 2153 2154 return sum; 2155} 2156 2157unsigned long nr_iowait(void) 2158{ 2159 unsigned long i, sum = 0; 2160 2161 for_each_possible_cpu(i) 2162 sum += atomic_read(&cpu_rq(i)->nr_iowait); 2163 2164 return sum; 2165} 2166 2167unsigned long nr_iowait_cpu(int cpu) 2168{ 2169 struct rq *this = cpu_rq(cpu); 2170 return atomic_read(&this->nr_iowait); 2171} 2172 2173#ifdef CONFIG_SMP 2174 2175/* 2176 * sched_exec - execve() is a valuable balancing opportunity, because at 2177 * this point the task has the smallest effective memory and cache footprint. 2178 */ 2179void sched_exec(void) 2180{ 2181 struct task_struct *p = current; 2182 unsigned long flags; 2183 int dest_cpu; 2184 2185 raw_spin_lock_irqsave(&p->pi_lock, flags); 2186 dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), SD_BALANCE_EXEC, 0); 2187 if (dest_cpu == smp_processor_id()) 2188 goto unlock; 2189 2190 if (likely(cpu_active(dest_cpu))) { 2191 struct migration_arg arg = { p, dest_cpu }; 2192 2193 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 2194 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg); 2195 return; 2196 } 2197unlock: 2198 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 2199} 2200 2201#endif 2202 2203DEFINE_PER_CPU(struct kernel_stat, kstat); 2204DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat); 2205 2206EXPORT_PER_CPU_SYMBOL(kstat); 2207EXPORT_PER_CPU_SYMBOL(kernel_cpustat); 2208 2209/* 2210 * Return any ns on the sched_clock that have not yet been accounted in 2211 * @p in case that task is currently running. 2212 * 2213 * Called with task_rq_lock() held on @rq. 2214 */ 2215static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq) 2216{ 2217 u64 ns = 0; 2218 2219 if (task_current(rq, p)) { 2220 update_rq_clock(rq); 2221 ns = rq_clock_task(rq) - p->se.exec_start; 2222 if ((s64)ns < 0) 2223 ns = 0; 2224 } 2225 2226 return ns; 2227} 2228 2229unsigned long long task_delta_exec(struct task_struct *p) 2230{ 2231 unsigned long flags; 2232 struct rq *rq; 2233 u64 ns = 0; 2234 2235 rq = task_rq_lock(p, &flags); 2236 ns = do_task_delta_exec(p, rq); 2237 task_rq_unlock(rq, p, &flags); 2238 2239 return ns; 2240} 2241 2242/* 2243 * Return accounted runtime for the task. 2244 * In case the task is currently running, return the runtime plus current's 2245 * pending runtime that have not been accounted yet. 2246 */ 2247unsigned long long task_sched_runtime(struct task_struct *p) 2248{ 2249 unsigned long flags; 2250 struct rq *rq; 2251 u64 ns = 0; 2252 2253 rq = task_rq_lock(p, &flags); 2254 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq); 2255 task_rq_unlock(rq, p, &flags); 2256 2257 return ns; 2258} 2259 2260/* 2261 * This function gets called by the timer code, with HZ frequency. 2262 * We call it with interrupts disabled. 2263 */ 2264void scheduler_tick(void) 2265{ 2266 int cpu = smp_processor_id(); 2267 struct rq *rq = cpu_rq(cpu); 2268 struct task_struct *curr = rq->curr; 2269 2270 sched_clock_tick(); 2271 2272 raw_spin_lock(&rq->lock); 2273 update_rq_clock(rq); 2274 curr->sched_class->task_tick(rq, curr, 0); 2275 update_cpu_load_active(rq); 2276 raw_spin_unlock(&rq->lock); 2277 2278 perf_event_task_tick(); 2279 2280#ifdef CONFIG_SMP 2281 rq->idle_balance = idle_cpu(cpu); 2282 trigger_load_balance(rq, cpu); 2283#endif 2284 rq_last_tick_reset(rq); 2285} 2286 2287#ifdef CONFIG_NO_HZ_FULL 2288/** 2289 * scheduler_tick_max_deferment 2290 * 2291 * Keep at least one tick per second when a single 2292 * active task is running because the scheduler doesn't 2293 * yet completely support full dynticks environment. 2294 * 2295 * This makes sure that uptime, CFS vruntime, load 2296 * balancing, etc... continue to move forward, even 2297 * with a very low granularity. 2298 * 2299 * Return: Maximum deferment in nanoseconds. 2300 */ 2301u64 scheduler_tick_max_deferment(void) 2302{ 2303 struct rq *rq = this_rq(); 2304 unsigned long next, now = ACCESS_ONCE(jiffies); 2305 2306 next = rq->last_sched_tick + HZ; 2307 2308 if (time_before_eq(next, now)) 2309 return 0; 2310 2311 return jiffies_to_usecs(next - now) * NSEC_PER_USEC; 2312} 2313#endif 2314 2315notrace unsigned long get_parent_ip(unsigned long addr) 2316{ 2317 if (in_lock_functions(addr)) { 2318 addr = CALLER_ADDR2; 2319 if (in_lock_functions(addr)) 2320 addr = CALLER_ADDR3; 2321 } 2322 return addr; 2323} 2324 2325#if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \ 2326 defined(CONFIG_PREEMPT_TRACER)) 2327 2328void __kprobes preempt_count_add(int val) 2329{ 2330#ifdef CONFIG_DEBUG_PREEMPT 2331 /* 2332 * Underflow? 2333 */ 2334 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0))) 2335 return; 2336#endif 2337 __preempt_count_add(val); 2338#ifdef CONFIG_DEBUG_PREEMPT 2339 /* 2340 * Spinlock count overflowing soon? 2341 */ 2342 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >= 2343 PREEMPT_MASK - 10); 2344#endif 2345 if (preempt_count() == val) 2346 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1)); 2347} 2348EXPORT_SYMBOL(preempt_count_add); 2349 2350void __kprobes preempt_count_sub(int val) 2351{ 2352#ifdef CONFIG_DEBUG_PREEMPT 2353 /* 2354 * Underflow? 2355 */ 2356 if (DEBUG_LOCKS_WARN_ON(val > preempt_count())) 2357 return; 2358 /* 2359 * Is the spinlock portion underflowing? 2360 */ 2361 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) && 2362 !(preempt_count() & PREEMPT_MASK))) 2363 return; 2364#endif 2365 2366 if (preempt_count() == val) 2367 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1)); 2368 __preempt_count_sub(val); 2369} 2370EXPORT_SYMBOL(preempt_count_sub); 2371 2372#endif 2373 2374/* 2375 * Print scheduling while atomic bug: 2376 */ 2377static noinline void __schedule_bug(struct task_struct *prev) 2378{ 2379 if (oops_in_progress) 2380 return; 2381 2382 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n", 2383 prev->comm, prev->pid, preempt_count()); 2384 2385 debug_show_held_locks(prev); 2386 print_modules(); 2387 if (irqs_disabled()) 2388 print_irqtrace_events(prev); 2389 dump_stack(); 2390 add_taint(TAINT_WARN, LOCKDEP_STILL_OK); 2391} 2392 2393/* 2394 * Various schedule()-time debugging checks and statistics: 2395 */ 2396static inline void schedule_debug(struct task_struct *prev) 2397{ 2398 /* 2399 * Test if we are atomic. Since do_exit() needs to call into 2400 * schedule() atomically, we ignore that path for now. 2401 * Otherwise, whine if we are scheduling when we should not be. 2402 */ 2403 if (unlikely(in_atomic_preempt_off() && !prev->exit_state)) 2404 __schedule_bug(prev); 2405 rcu_sleep_check(); 2406 2407 profile_hit(SCHED_PROFILING, __builtin_return_address(0)); 2408 2409 schedstat_inc(this_rq(), sched_count); 2410} 2411 2412static void put_prev_task(struct rq *rq, struct task_struct *prev) 2413{ 2414 if (prev->on_rq || rq->skip_clock_update < 0) 2415 update_rq_clock(rq); 2416 prev->sched_class->put_prev_task(rq, prev); 2417} 2418 2419/* 2420 * Pick up the highest-prio task: 2421 */ 2422static inline struct task_struct * 2423pick_next_task(struct rq *rq) 2424{ 2425 const struct sched_class *class; 2426 struct task_struct *p; 2427 2428 /* 2429 * Optimization: we know that if all tasks are in 2430 * the fair class we can call that function directly: 2431 */ 2432 if (likely(rq->nr_running == rq->cfs.h_nr_running)) { 2433 p = fair_sched_class.pick_next_task(rq); 2434 if (likely(p)) 2435 return p; 2436 } 2437 2438 for_each_class(class) { 2439 p = class->pick_next_task(rq); 2440 if (p) 2441 return p; 2442 } 2443 2444 BUG(); /* the idle class will always have a runnable task */ 2445} 2446 2447/* 2448 * __schedule() is the main scheduler function. 2449 * 2450 * The main means of driving the scheduler and thus entering this function are: 2451 * 2452 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc. 2453 * 2454 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return 2455 * paths. For example, see arch/x86/entry_64.S. 2456 * 2457 * To drive preemption between tasks, the scheduler sets the flag in timer 2458 * interrupt handler scheduler_tick(). 2459 * 2460 * 3. Wakeups don't really cause entry into schedule(). They add a 2461 * task to the run-queue and that's it. 2462 * 2463 * Now, if the new task added to the run-queue preempts the current 2464 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets 2465 * called on the nearest possible occasion: 2466 * 2467 * - If the kernel is preemptible (CONFIG_PREEMPT=y): 2468 * 2469 * - in syscall or exception context, at the next outmost 2470 * preempt_enable(). (this might be as soon as the wake_up()'s 2471 * spin_unlock()!) 2472 * 2473 * - in IRQ context, return from interrupt-handler to 2474 * preemptible context 2475 * 2476 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set) 2477 * then at the next: 2478 * 2479 * - cond_resched() call 2480 * - explicit schedule() call 2481 * - return from syscall or exception to user-space 2482 * - return from interrupt-handler to user-space 2483 */ 2484static void __sched __schedule(void) 2485{ 2486 struct task_struct *prev, *next; 2487 unsigned long *switch_count; 2488 struct rq *rq; 2489 int cpu; 2490 2491need_resched: 2492 preempt_disable(); 2493 cpu = smp_processor_id(); 2494 rq = cpu_rq(cpu); 2495 rcu_note_context_switch(cpu); 2496 prev = rq->curr; 2497 2498 schedule_debug(prev); 2499 2500 if (sched_feat(HRTICK)) 2501 hrtick_clear(rq); 2502 2503 /* 2504 * Make sure that signal_pending_state()->signal_pending() below 2505 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE) 2506 * done by the caller to avoid the race with signal_wake_up(). 2507 */ 2508 smp_mb__before_spinlock(); 2509 raw_spin_lock_irq(&rq->lock); 2510 2511 switch_count = &prev->nivcsw; 2512 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) { 2513 if (unlikely(signal_pending_state(prev->state, prev))) { 2514 prev->state = TASK_RUNNING; 2515 } else { 2516 deactivate_task(rq, prev, DEQUEUE_SLEEP); 2517 prev->on_rq = 0; 2518 2519 /* 2520 * If a worker went to sleep, notify and ask workqueue 2521 * whether it wants to wake up a task to maintain 2522 * concurrency. 2523 */ 2524 if (prev->flags & PF_WQ_WORKER) { 2525 struct task_struct *to_wakeup; 2526 2527 to_wakeup = wq_worker_sleeping(prev, cpu); 2528 if (to_wakeup) 2529 try_to_wake_up_local(to_wakeup); 2530 } 2531 } 2532 switch_count = &prev->nvcsw; 2533 } 2534 2535 pre_schedule(rq, prev); 2536 2537 if (unlikely(!rq->nr_running)) 2538 idle_balance(cpu, rq); 2539 2540 put_prev_task(rq, prev); 2541 next = pick_next_task(rq); 2542 clear_tsk_need_resched(prev); 2543 clear_preempt_need_resched(); 2544 rq->skip_clock_update = 0; 2545 2546 if (likely(prev != next)) { 2547 rq->nr_switches++; 2548 rq->curr = next; 2549 ++*switch_count; 2550 2551 context_switch(rq, prev, next); /* unlocks the rq */ 2552 /* 2553 * The context switch have flipped the stack from under us 2554 * and restored the local variables which were saved when 2555 * this task called schedule() in the past. prev == current 2556 * is still correct, but it can be moved to another cpu/rq. 2557 */ 2558 cpu = smp_processor_id(); 2559 rq = cpu_rq(cpu); 2560 } else 2561 raw_spin_unlock_irq(&rq->lock); 2562 2563 post_schedule(rq); 2564 2565 sched_preempt_enable_no_resched(); 2566 if (need_resched()) 2567 goto need_resched; 2568} 2569 2570static inline void sched_submit_work(struct task_struct *tsk) 2571{ 2572 if (!tsk->state || tsk_is_pi_blocked(tsk)) 2573 return; 2574 /* 2575 * If we are going to sleep and we have plugged IO queued, 2576 * make sure to submit it to avoid deadlocks. 2577 */ 2578 if (blk_needs_flush_plug(tsk)) 2579 blk_schedule_flush_plug(tsk); 2580} 2581 2582asmlinkage void __sched schedule(void) 2583{ 2584 struct task_struct *tsk = current; 2585 2586 sched_submit_work(tsk); 2587 __schedule(); 2588} 2589EXPORT_SYMBOL(schedule); 2590 2591#ifdef CONFIG_CONTEXT_TRACKING 2592asmlinkage void __sched schedule_user(void) 2593{ 2594 /* 2595 * If we come here after a random call to set_need_resched(), 2596 * or we have been woken up remotely but the IPI has not yet arrived, 2597 * we haven't yet exited the RCU idle mode. Do it here manually until 2598 * we find a better solution. 2599 */ 2600 user_exit(); 2601 schedule(); 2602 user_enter(); 2603} 2604#endif 2605 2606/** 2607 * schedule_preempt_disabled - called with preemption disabled 2608 * 2609 * Returns with preemption disabled. Note: preempt_count must be 1 2610 */ 2611void __sched schedule_preempt_disabled(void) 2612{ 2613 sched_preempt_enable_no_resched(); 2614 schedule(); 2615 preempt_disable(); 2616} 2617 2618#ifdef CONFIG_PREEMPT 2619/* 2620 * this is the entry point to schedule() from in-kernel preemption 2621 * off of preempt_enable. Kernel preemptions off return from interrupt 2622 * occur there and call schedule directly. 2623 */ 2624asmlinkage void __sched notrace preempt_schedule(void) 2625{ 2626 /* 2627 * If there is a non-zero preempt_count or interrupts are disabled, 2628 * we do not want to preempt the current task. Just return.. 2629 */ 2630 if (likely(!preemptible())) 2631 return; 2632 2633 do { 2634 __preempt_count_add(PREEMPT_ACTIVE); 2635 __schedule(); 2636 __preempt_count_sub(PREEMPT_ACTIVE); 2637 2638 /* 2639 * Check again in case we missed a preemption opportunity 2640 * between schedule and now. 2641 */ 2642 barrier(); 2643 } while (need_resched()); 2644} 2645EXPORT_SYMBOL(preempt_schedule); 2646 2647/* 2648 * this is the entry point to schedule() from kernel preemption 2649 * off of irq context. 2650 * Note, that this is called and return with irqs disabled. This will 2651 * protect us against recursive calling from irq. 2652 */ 2653asmlinkage void __sched preempt_schedule_irq(void) 2654{ 2655 enum ctx_state prev_state; 2656 2657 /* Catch callers which need to be fixed */ 2658 BUG_ON(preempt_count() || !irqs_disabled()); 2659 2660 prev_state = exception_enter(); 2661 2662 do { 2663 __preempt_count_add(PREEMPT_ACTIVE); 2664 local_irq_enable(); 2665 __schedule(); 2666 local_irq_disable(); 2667 __preempt_count_sub(PREEMPT_ACTIVE); 2668 2669 /* 2670 * Check again in case we missed a preemption opportunity 2671 * between schedule and now. 2672 */ 2673 barrier(); 2674 } while (need_resched()); 2675 2676 exception_exit(prev_state); 2677} 2678 2679#endif /* CONFIG_PREEMPT */ 2680 2681int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags, 2682 void *key) 2683{ 2684 return try_to_wake_up(curr->private, mode, wake_flags); 2685} 2686EXPORT_SYMBOL(default_wake_function); 2687 2688/* 2689 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just 2690 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve 2691 * number) then we wake all the non-exclusive tasks and one exclusive task. 2692 * 2693 * There are circumstances in which we can try to wake a task which has already 2694 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns 2695 * zero in this (rare) case, and we handle it by continuing to scan the queue. 2696 */ 2697static void __wake_up_common(wait_queue_head_t *q, unsigned int mode, 2698 int nr_exclusive, int wake_flags, void *key) 2699{ 2700 wait_queue_t *curr, *next; 2701 2702 list_for_each_entry_safe(curr, next, &q->task_list, task_list) { 2703 unsigned flags = curr->flags; 2704 2705 if (curr->func(curr, mode, wake_flags, key) && 2706 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive) 2707 break; 2708 } 2709} 2710 2711/** 2712 * __wake_up - wake up threads blocked on a waitqueue. 2713 * @q: the waitqueue 2714 * @mode: which threads 2715 * @nr_exclusive: how many wake-one or wake-many threads to wake up 2716 * @key: is directly passed to the wakeup function 2717 * 2718 * It may be assumed that this function implies a write memory barrier before 2719 * changing the task state if and only if any tasks are woken up. 2720 */ 2721void __wake_up(wait_queue_head_t *q, unsigned int mode, 2722 int nr_exclusive, void *key) 2723{ 2724 unsigned long flags; 2725 2726 spin_lock_irqsave(&q->lock, flags); 2727 __wake_up_common(q, mode, nr_exclusive, 0, key); 2728 spin_unlock_irqrestore(&q->lock, flags); 2729} 2730EXPORT_SYMBOL(__wake_up); 2731 2732/* 2733 * Same as __wake_up but called with the spinlock in wait_queue_head_t held. 2734 */ 2735void __wake_up_locked(wait_queue_head_t *q, unsigned int mode, int nr) 2736{ 2737 __wake_up_common(q, mode, nr, 0, NULL); 2738} 2739EXPORT_SYMBOL_GPL(__wake_up_locked); 2740 2741void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key) 2742{ 2743 __wake_up_common(q, mode, 1, 0, key); 2744} 2745EXPORT_SYMBOL_GPL(__wake_up_locked_key); 2746 2747/** 2748 * __wake_up_sync_key - wake up threads blocked on a waitqueue. 2749 * @q: the waitqueue 2750 * @mode: which threads 2751 * @nr_exclusive: how many wake-one or wake-many threads to wake up 2752 * @key: opaque value to be passed to wakeup targets 2753 * 2754 * The sync wakeup differs that the waker knows that it will schedule 2755 * away soon, so while the target thread will be woken up, it will not 2756 * be migrated to another CPU - ie. the two threads are 'synchronized' 2757 * with each other. This can prevent needless bouncing between CPUs. 2758 * 2759 * On UP it can prevent extra preemption. 2760 * 2761 * It may be assumed that this function implies a write memory barrier before 2762 * changing the task state if and only if any tasks are woken up. 2763 */ 2764void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode, 2765 int nr_exclusive, void *key) 2766{ 2767 unsigned long flags; 2768 int wake_flags = WF_SYNC; 2769 2770 if (unlikely(!q)) 2771 return; 2772 2773 if (unlikely(nr_exclusive != 1)) 2774 wake_flags = 0; 2775 2776 spin_lock_irqsave(&q->lock, flags); 2777 __wake_up_common(q, mode, nr_exclusive, wake_flags, key); 2778 spin_unlock_irqrestore(&q->lock, flags); 2779} 2780EXPORT_SYMBOL_GPL(__wake_up_sync_key); 2781 2782/* 2783 * __wake_up_sync - see __wake_up_sync_key() 2784 */ 2785void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive) 2786{ 2787 __wake_up_sync_key(q, mode, nr_exclusive, NULL); 2788} 2789EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */ 2790 2791/** 2792 * complete: - signals a single thread waiting on this completion 2793 * @x: holds the state of this particular completion 2794 * 2795 * This will wake up a single thread waiting on this completion. Threads will be 2796 * awakened in the same order in which they were queued. 2797 * 2798 * See also complete_all(), wait_for_completion() and related routines. 2799 * 2800 * It may be assumed that this function implies a write memory barrier before 2801 * changing the task state if and only if any tasks are woken up. 2802 */ 2803void complete(struct completion *x) 2804{ 2805 unsigned long flags; 2806 2807 spin_lock_irqsave(&x->wait.lock, flags); 2808 x->done++; 2809 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL); 2810 spin_unlock_irqrestore(&x->wait.lock, flags); 2811} 2812EXPORT_SYMBOL(complete); 2813 2814/** 2815 * complete_all: - signals all threads waiting on this completion 2816 * @x: holds the state of this particular completion 2817 * 2818 * This will wake up all threads waiting on this particular completion event. 2819 * 2820 * It may be assumed that this function implies a write memory barrier before 2821 * changing the task state if and only if any tasks are woken up. 2822 */ 2823void complete_all(struct completion *x) 2824{ 2825 unsigned long flags; 2826 2827 spin_lock_irqsave(&x->wait.lock, flags); 2828 x->done += UINT_MAX/2; 2829 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL); 2830 spin_unlock_irqrestore(&x->wait.lock, flags); 2831} 2832EXPORT_SYMBOL(complete_all); 2833 2834static inline long __sched 2835do_wait_for_common(struct completion *x, 2836 long (*action)(long), long timeout, int state) 2837{ 2838 if (!x->done) { 2839 DECLARE_WAITQUEUE(wait, current); 2840 2841 __add_wait_queue_tail_exclusive(&x->wait, &wait); 2842 do { 2843 if (signal_pending_state(state, current)) { 2844 timeout = -ERESTARTSYS; 2845 break; 2846 } 2847 __set_current_state(state); 2848 spin_unlock_irq(&x->wait.lock); 2849 timeout = action(timeout); 2850 spin_lock_irq(&x->wait.lock); 2851 } while (!x->done && timeout); 2852 __remove_wait_queue(&x->wait, &wait); 2853 if (!x->done) 2854 return timeout; 2855 } 2856 x->done--; 2857 return timeout ?: 1; 2858} 2859 2860static inline long __sched 2861__wait_for_common(struct completion *x, 2862 long (*action)(long), long timeout, int state) 2863{ 2864 might_sleep(); 2865 2866 spin_lock_irq(&x->wait.lock); 2867 timeout = do_wait_for_common(x, action, timeout, state); 2868 spin_unlock_irq(&x->wait.lock); 2869 return timeout; 2870} 2871 2872static long __sched 2873wait_for_common(struct completion *x, long timeout, int state) 2874{ 2875 return __wait_for_common(x, schedule_timeout, timeout, state); 2876} 2877 2878static long __sched 2879wait_for_common_io(struct completion *x, long timeout, int state) 2880{ 2881 return __wait_for_common(x, io_schedule_timeout, timeout, state); 2882} 2883 2884/** 2885 * wait_for_completion: - waits for completion of a task 2886 * @x: holds the state of this particular completion 2887 * 2888 * This waits to be signaled for completion of a specific task. It is NOT 2889 * interruptible and there is no timeout. 2890 * 2891 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout 2892 * and interrupt capability. Also see complete(). 2893 */ 2894void __sched wait_for_completion(struct completion *x) 2895{ 2896 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE); 2897} 2898EXPORT_SYMBOL(wait_for_completion); 2899 2900/** 2901 * wait_for_completion_timeout: - waits for completion of a task (w/timeout) 2902 * @x: holds the state of this particular completion 2903 * @timeout: timeout value in jiffies 2904 * 2905 * This waits for either a completion of a specific task to be signaled or for a 2906 * specified timeout to expire. The timeout is in jiffies. It is not 2907 * interruptible. 2908 * 2909 * Return: 0 if timed out, and positive (at least 1, or number of jiffies left 2910 * till timeout) if completed. 2911 */ 2912unsigned long __sched 2913wait_for_completion_timeout(struct completion *x, unsigned long timeout) 2914{ 2915 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE); 2916} 2917EXPORT_SYMBOL(wait_for_completion_timeout); 2918 2919/** 2920 * wait_for_completion_io: - waits for completion of a task 2921 * @x: holds the state of this particular completion 2922 * 2923 * This waits to be signaled for completion of a specific task. It is NOT 2924 * interruptible and there is no timeout. The caller is accounted as waiting 2925 * for IO. 2926 */ 2927void __sched wait_for_completion_io(struct completion *x) 2928{ 2929 wait_for_common_io(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE); 2930} 2931EXPORT_SYMBOL(wait_for_completion_io); 2932 2933/** 2934 * wait_for_completion_io_timeout: - waits for completion of a task (w/timeout) 2935 * @x: holds the state of this particular completion 2936 * @timeout: timeout value in jiffies 2937 * 2938 * This waits for either a completion of a specific task to be signaled or for a 2939 * specified timeout to expire. The timeout is in jiffies. It is not 2940 * interruptible. The caller is accounted as waiting for IO. 2941 * 2942 * Return: 0 if timed out, and positive (at least 1, or number of jiffies left 2943 * till timeout) if completed. 2944 */ 2945unsigned long __sched 2946wait_for_completion_io_timeout(struct completion *x, unsigned long timeout) 2947{ 2948 return wait_for_common_io(x, timeout, TASK_UNINTERRUPTIBLE); 2949} 2950EXPORT_SYMBOL(wait_for_completion_io_timeout); 2951 2952/** 2953 * wait_for_completion_interruptible: - waits for completion of a task (w/intr) 2954 * @x: holds the state of this particular completion 2955 * 2956 * This waits for completion of a specific task to be signaled. It is 2957 * interruptible. 2958 * 2959 * Return: -ERESTARTSYS if interrupted, 0 if completed. 2960 */ 2961int __sched wait_for_completion_interruptible(struct completion *x) 2962{ 2963 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE); 2964 if (t == -ERESTARTSYS) 2965 return t; 2966 return 0; 2967} 2968EXPORT_SYMBOL(wait_for_completion_interruptible); 2969 2970/** 2971 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr)) 2972 * @x: holds the state of this particular completion 2973 * @timeout: timeout value in jiffies 2974 * 2975 * This waits for either a completion of a specific task to be signaled or for a 2976 * specified timeout to expire. It is interruptible. The timeout is in jiffies. 2977 * 2978 * Return: -ERESTARTSYS if interrupted, 0 if timed out, positive (at least 1, 2979 * or number of jiffies left till timeout) if completed. 2980 */ 2981long __sched 2982wait_for_completion_interruptible_timeout(struct completion *x, 2983 unsigned long timeout) 2984{ 2985 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE); 2986} 2987EXPORT_SYMBOL(wait_for_completion_interruptible_timeout); 2988 2989/** 2990 * wait_for_completion_killable: - waits for completion of a task (killable) 2991 * @x: holds the state of this particular completion 2992 * 2993 * This waits to be signaled for completion of a specific task. It can be 2994 * interrupted by a kill signal. 2995 * 2996 * Return: -ERESTARTSYS if interrupted, 0 if completed. 2997 */ 2998int __sched wait_for_completion_killable(struct completion *x) 2999{ 3000 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE); 3001 if (t == -ERESTARTSYS) 3002 return t; 3003 return 0; 3004} 3005EXPORT_SYMBOL(wait_for_completion_killable); 3006 3007/** 3008 * wait_for_completion_killable_timeout: - waits for completion of a task (w/(to,killable)) 3009 * @x: holds the state of this particular completion 3010 * @timeout: timeout value in jiffies 3011 * 3012 * This waits for either a completion of a specific task to be 3013 * signaled or for a specified timeout to expire. It can be 3014 * interrupted by a kill signal. The timeout is in jiffies. 3015 * 3016 * Return: -ERESTARTSYS if interrupted, 0 if timed out, positive (at least 1, 3017 * or number of jiffies left till timeout) if completed. 3018 */ 3019long __sched 3020wait_for_completion_killable_timeout(struct completion *x, 3021 unsigned long timeout) 3022{ 3023 return wait_for_common(x, timeout, TASK_KILLABLE); 3024} 3025EXPORT_SYMBOL(wait_for_completion_killable_timeout); 3026 3027/** 3028 * try_wait_for_completion - try to decrement a completion without blocking 3029 * @x: completion structure 3030 * 3031 * Return: 0 if a decrement cannot be done without blocking 3032 * 1 if a decrement succeeded. 3033 * 3034 * If a completion is being used as a counting completion, 3035 * attempt to decrement the counter without blocking. This 3036 * enables us to avoid waiting if the resource the completion 3037 * is protecting is not available. 3038 */ 3039bool try_wait_for_completion(struct completion *x) 3040{ 3041 unsigned long flags; 3042 int ret = 1; 3043 3044 spin_lock_irqsave(&x->wait.lock, flags); 3045 if (!x->done) 3046 ret = 0; 3047 else 3048 x->done--; 3049 spin_unlock_irqrestore(&x->wait.lock, flags); 3050 return ret; 3051} 3052EXPORT_SYMBOL(try_wait_for_completion); 3053 3054/** 3055 * completion_done - Test to see if a completion has any waiters 3056 * @x: completion structure 3057 * 3058 * Return: 0 if there are waiters (wait_for_completion() in progress) 3059 * 1 if there are no waiters. 3060 * 3061 */ 3062bool completion_done(struct completion *x) 3063{ 3064 unsigned long flags; 3065 int ret = 1; 3066 3067 spin_lock_irqsave(&x->wait.lock, flags); 3068 if (!x->done) 3069 ret = 0; 3070 spin_unlock_irqrestore(&x->wait.lock, flags); 3071 return ret; 3072} 3073EXPORT_SYMBOL(completion_done); 3074 3075static long __sched 3076sleep_on_common(wait_queue_head_t *q, int state, long timeout) 3077{ 3078 unsigned long flags; 3079 wait_queue_t wait; 3080 3081 init_waitqueue_entry(&wait, current); 3082 3083 __set_current_state(state); 3084 3085 spin_lock_irqsave(&q->lock, flags); 3086 __add_wait_queue(q, &wait); 3087 spin_unlock(&q->lock); 3088 timeout = schedule_timeout(timeout); 3089 spin_lock_irq(&q->lock); 3090 __remove_wait_queue(q, &wait); 3091 spin_unlock_irqrestore(&q->lock, flags); 3092 3093 return timeout; 3094} 3095 3096void __sched interruptible_sleep_on(wait_queue_head_t *q) 3097{ 3098 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT); 3099} 3100EXPORT_SYMBOL(interruptible_sleep_on); 3101 3102long __sched 3103interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout) 3104{ 3105 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout); 3106} 3107EXPORT_SYMBOL(interruptible_sleep_on_timeout); 3108 3109void __sched sleep_on(wait_queue_head_t *q) 3110{ 3111 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT); 3112} 3113EXPORT_SYMBOL(sleep_on); 3114 3115long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout) 3116{ 3117 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout); 3118} 3119EXPORT_SYMBOL(sleep_on_timeout); 3120 3121#ifdef CONFIG_RT_MUTEXES 3122 3123/* 3124 * rt_mutex_setprio - set the current priority of a task 3125 * @p: task 3126 * @prio: prio value (kernel-internal form) 3127 * 3128 * This function changes the 'effective' priority of a task. It does 3129 * not touch ->normal_prio like __setscheduler(). 3130 * 3131 * Used by the rt_mutex code to implement priority inheritance logic. 3132 */ 3133void rt_mutex_setprio(struct task_struct *p, int prio) 3134{ 3135 int oldprio, on_rq, running; 3136 struct rq *rq; 3137 const struct sched_class *prev_class; 3138 3139 BUG_ON(prio < 0 || prio > MAX_PRIO); 3140 3141 rq = __task_rq_lock(p); 3142 3143 /* 3144 * Idle task boosting is a nono in general. There is one 3145 * exception, when PREEMPT_RT and NOHZ is active: 3146 * 3147 * The idle task calls get_next_timer_interrupt() and holds 3148 * the timer wheel base->lock on the CPU and another CPU wants 3149 * to access the timer (probably to cancel it). We can safely 3150 * ignore the boosting request, as the idle CPU runs this code 3151 * with interrupts disabled and will complete the lock 3152 * protected section without being interrupted. So there is no 3153 * real need to boost. 3154 */ 3155 if (unlikely(p == rq->idle)) { 3156 WARN_ON(p != rq->curr); 3157 WARN_ON(p->pi_blocked_on); 3158 goto out_unlock; 3159 } 3160 3161 trace_sched_pi_setprio(p, prio); 3162 oldprio = p->prio; 3163 prev_class = p->sched_class; 3164 on_rq = p->on_rq; 3165 running = task_current(rq, p); 3166 if (on_rq) 3167 dequeue_task(rq, p, 0); 3168 if (running) 3169 p->sched_class->put_prev_task(rq, p); 3170 3171 if (rt_prio(prio)) 3172 p->sched_class = &rt_sched_class; 3173 else 3174 p->sched_class = &fair_sched_class; 3175 3176 p->prio = prio; 3177 3178 if (running) 3179 p->sched_class->set_curr_task(rq); 3180 if (on_rq) 3181 enqueue_task(rq, p, oldprio < prio ? ENQUEUE_HEAD : 0); 3182 3183 check_class_changed(rq, p, prev_class, oldprio); 3184out_unlock: 3185 __task_rq_unlock(rq); 3186} 3187#endif 3188void set_user_nice(struct task_struct *p, long nice) 3189{ 3190 int old_prio, delta, on_rq; 3191 unsigned long flags; 3192 struct rq *rq; 3193 3194 if (TASK_NICE(p) == nice || nice < -20 || nice > 19) 3195 return; 3196 /* 3197 * We have to be careful, if called from sys_setpriority(), 3198 * the task might be in the middle of scheduling on another CPU. 3199 */ 3200 rq = task_rq_lock(p, &flags); 3201 /* 3202 * The RT priorities are set via sched_setscheduler(), but we still 3203 * allow the 'normal' nice value to be set - but as expected 3204 * it wont have any effect on scheduling until the task is 3205 * SCHED_FIFO/SCHED_RR: 3206 */ 3207 if (task_has_rt_policy(p)) { 3208 p->static_prio = NICE_TO_PRIO(nice); 3209 goto out_unlock; 3210 } 3211 on_rq = p->on_rq; 3212 if (on_rq) 3213 dequeue_task(rq, p, 0); 3214 3215 p->static_prio = NICE_TO_PRIO(nice); 3216 set_load_weight(p); 3217 old_prio = p->prio; 3218 p->prio = effective_prio(p); 3219 delta = p->prio - old_prio; 3220 3221 if (on_rq) { 3222 enqueue_task(rq, p, 0); 3223 /* 3224 * If the task increased its priority or is running and 3225 * lowered its priority, then reschedule its CPU: 3226 */ 3227 if (delta < 0 || (delta > 0 && task_running(rq, p))) 3228 resched_task(rq->curr); 3229 } 3230out_unlock: 3231 task_rq_unlock(rq, p, &flags); 3232} 3233EXPORT_SYMBOL(set_user_nice); 3234 3235/* 3236 * can_nice - check if a task can reduce its nice value 3237 * @p: task 3238 * @nice: nice value 3239 */ 3240int can_nice(const struct task_struct *p, const int nice) 3241{ 3242 /* convert nice value [19,-20] to rlimit style value [1,40] */ 3243 int nice_rlim = 20 - nice; 3244 3245 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) || 3246 capable(CAP_SYS_NICE)); 3247} 3248 3249#ifdef __ARCH_WANT_SYS_NICE 3250 3251/* 3252 * sys_nice - change the priority of the current process. 3253 * @increment: priority increment 3254 * 3255 * sys_setpriority is a more generic, but much slower function that 3256 * does similar things. 3257 */ 3258SYSCALL_DEFINE1(nice, int, increment) 3259{ 3260 long nice, retval; 3261 3262 /* 3263 * Setpriority might change our priority at the same moment. 3264 * We don't have to worry. Conceptually one call occurs first 3265 * and we have a single winner. 3266 */ 3267 if (increment < -40) 3268 increment = -40; 3269 if (increment > 40) 3270 increment = 40; 3271 3272 nice = TASK_NICE(current) + increment; 3273 if (nice < -20) 3274 nice = -20; 3275 if (nice > 19) 3276 nice = 19; 3277 3278 if (increment < 0 && !can_nice(current, nice)) 3279 return -EPERM; 3280 3281 retval = security_task_setnice(current, nice); 3282 if (retval) 3283 return retval; 3284 3285 set_user_nice(current, nice); 3286 return 0; 3287} 3288 3289#endif 3290 3291/** 3292 * task_prio - return the priority value of a given task. 3293 * @p: the task in question. 3294 * 3295 * Return: The priority value as seen by users in /proc. 3296 * RT tasks are offset by -200. Normal tasks are centered 3297 * around 0, value goes from -16 to +15. 3298 */ 3299int task_prio(const struct task_struct *p) 3300{ 3301 return p->prio - MAX_RT_PRIO; 3302} 3303 3304/** 3305 * task_nice - return the nice value of a given task. 3306 * @p: the task in question. 3307 * 3308 * Return: The nice value [ -20 ... 0 ... 19 ]. 3309 */ 3310int task_nice(const struct task_struct *p) 3311{ 3312 return TASK_NICE(p); 3313} 3314EXPORT_SYMBOL(task_nice); 3315 3316/** 3317 * idle_cpu - is a given cpu idle currently? 3318 * @cpu: the processor in question. 3319 * 3320 * Return: 1 if the CPU is currently idle. 0 otherwise. 3321 */ 3322int idle_cpu(int cpu) 3323{ 3324 struct rq *rq = cpu_rq(cpu); 3325 3326 if (rq->curr != rq->idle) 3327 return 0; 3328 3329 if (rq->nr_running) 3330 return 0; 3331 3332#ifdef CONFIG_SMP 3333 if (!llist_empty(&rq->wake_list)) 3334 return 0; 3335#endif 3336 3337 return 1; 3338} 3339 3340/** 3341 * idle_task - return the idle task for a given cpu. 3342 * @cpu: the processor in question. 3343 * 3344 * Return: The idle task for the cpu @cpu. 3345 */ 3346struct task_struct *idle_task(int cpu) 3347{ 3348 return cpu_rq(cpu)->idle; 3349} 3350 3351/** 3352 * find_process_by_pid - find a process with a matching PID value. 3353 * @pid: the pid in question. 3354 * 3355 * The task of @pid, if found. %NULL otherwise. 3356 */ 3357static struct task_struct *find_process_by_pid(pid_t pid) 3358{ 3359 return pid ? find_task_by_vpid(pid) : current; 3360} 3361 3362/* Actually do priority change: must hold rq lock. */ 3363static void 3364__setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio) 3365{ 3366 p->policy = policy; 3367 p->rt_priority = prio; 3368 p->normal_prio = normal_prio(p); 3369 /* we are holding p->pi_lock already */ 3370 p->prio = rt_mutex_getprio(p); 3371 if (rt_prio(p->prio)) 3372 p->sched_class = &rt_sched_class; 3373 else 3374 p->sched_class = &fair_sched_class; 3375 set_load_weight(p); 3376} 3377 3378/* 3379 * check the target process has a UID that matches the current process's 3380 */ 3381static bool check_same_owner(struct task_struct *p) 3382{ 3383 const struct cred *cred = current_cred(), *pcred; 3384 bool match; 3385 3386 rcu_read_lock(); 3387 pcred = __task_cred(p); 3388 match = (uid_eq(cred->euid, pcred->euid) || 3389 uid_eq(cred->euid, pcred->uid)); 3390 rcu_read_unlock(); 3391 return match; 3392} 3393 3394static int __sched_setscheduler(struct task_struct *p, int policy, 3395 const struct sched_param *param, bool user) 3396{ 3397 int retval, oldprio, oldpolicy = -1, on_rq, running; 3398 unsigned long flags; 3399 const struct sched_class *prev_class; 3400 struct rq *rq; 3401 int reset_on_fork; 3402 3403 /* may grab non-irq protected spin_locks */ 3404 BUG_ON(in_interrupt()); 3405recheck: 3406 /* double check policy once rq lock held */ 3407 if (policy < 0) { 3408 reset_on_fork = p->sched_reset_on_fork; 3409 policy = oldpolicy = p->policy; 3410 } else { 3411 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK); 3412 policy &= ~SCHED_RESET_ON_FORK; 3413 3414 if (policy != SCHED_FIFO && policy != SCHED_RR && 3415 policy != SCHED_NORMAL && policy != SCHED_BATCH && 3416 policy != SCHED_IDLE) 3417 return -EINVAL; 3418 } 3419 3420 /* 3421 * Valid priorities for SCHED_FIFO and SCHED_RR are 3422 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL, 3423 * SCHED_BATCH and SCHED_IDLE is 0. 3424 */ 3425 if (param->sched_priority < 0 || 3426 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) || 3427 (!p->mm && param->sched_priority > MAX_RT_PRIO-1)) 3428 return -EINVAL; 3429 if (rt_policy(policy) != (param->sched_priority != 0)) 3430 return -EINVAL; 3431 3432 /* 3433 * Allow unprivileged RT tasks to decrease priority: 3434 */ 3435 if (user && !capable(CAP_SYS_NICE)) { 3436 if (rt_policy(policy)) { 3437 unsigned long rlim_rtprio = 3438 task_rlimit(p, RLIMIT_RTPRIO); 3439 3440 /* can't set/change the rt policy */ 3441 if (policy != p->policy && !rlim_rtprio) 3442 return -EPERM; 3443 3444 /* can't increase priority */ 3445 if (param->sched_priority > p->rt_priority && 3446 param->sched_priority > rlim_rtprio) 3447 return -EPERM; 3448 } 3449 3450 /* 3451 * Treat SCHED_IDLE as nice 20. Only allow a switch to 3452 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it. 3453 */ 3454 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) { 3455 if (!can_nice(p, TASK_NICE(p))) 3456 return -EPERM; 3457 } 3458 3459 /* can't change other user's priorities */ 3460 if (!check_same_owner(p)) 3461 return -EPERM; 3462 3463 /* Normal users shall not reset the sched_reset_on_fork flag */ 3464 if (p->sched_reset_on_fork && !reset_on_fork) 3465 return -EPERM; 3466 } 3467 3468 if (user) { 3469 retval = security_task_setscheduler(p); 3470 if (retval) 3471 return retval; 3472 } 3473 3474 /* 3475 * make sure no PI-waiters arrive (or leave) while we are 3476 * changing the priority of the task: 3477 * 3478 * To be able to change p->policy safely, the appropriate 3479 * runqueue lock must be held. 3480 */ 3481 rq = task_rq_lock(p, &flags); 3482 3483 /* 3484 * Changing the policy of the stop threads its a very bad idea 3485 */ 3486 if (p == rq->stop) { 3487 task_rq_unlock(rq, p, &flags); 3488 return -EINVAL; 3489 } 3490 3491 /* 3492 * If not changing anything there's no need to proceed further: 3493 */ 3494 if (unlikely(policy == p->policy && (!rt_policy(policy) || 3495 param->sched_priority == p->rt_priority))) { 3496 task_rq_unlock(rq, p, &flags); 3497 return 0; 3498 } 3499 3500#ifdef CONFIG_RT_GROUP_SCHED 3501 if (user) { 3502 /* 3503 * Do not allow realtime tasks into groups that have no runtime 3504 * assigned. 3505 */ 3506 if (rt_bandwidth_enabled() && rt_policy(policy) && 3507 task_group(p)->rt_bandwidth.rt_runtime == 0 && 3508 !task_group_is_autogroup(task_group(p))) { 3509 task_rq_unlock(rq, p, &flags); 3510 return -EPERM; 3511 } 3512 } 3513#endif 3514 3515 /* recheck policy now with rq lock held */ 3516 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) { 3517 policy = oldpolicy = -1; 3518 task_rq_unlock(rq, p, &flags); 3519 goto recheck; 3520 } 3521 on_rq = p->on_rq; 3522 running = task_current(rq, p); 3523 if (on_rq) 3524 dequeue_task(rq, p, 0); 3525 if (running) 3526 p->sched_class->put_prev_task(rq, p); 3527 3528 p->sched_reset_on_fork = reset_on_fork; 3529 3530 oldprio = p->prio; 3531 prev_class = p->sched_class; 3532 __setscheduler(rq, p, policy, param->sched_priority); 3533 3534 if (running) 3535 p->sched_class->set_curr_task(rq); 3536 if (on_rq) 3537 enqueue_task(rq, p, 0); 3538 3539 check_class_changed(rq, p, prev_class, oldprio); 3540 task_rq_unlock(rq, p, &flags); 3541 3542 rt_mutex_adjust_pi(p); 3543 3544 return 0; 3545} 3546 3547/** 3548 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread. 3549 * @p: the task in question. 3550 * @policy: new policy. 3551 * @param: structure containing the new RT priority. 3552 * 3553 * Return: 0 on success. An error code otherwise. 3554 * 3555 * NOTE that the task may be already dead. 3556 */ 3557int sched_setscheduler(struct task_struct *p, int policy, 3558 const struct sched_param *param) 3559{ 3560 return __sched_setscheduler(p, policy, param, true); 3561} 3562EXPORT_SYMBOL_GPL(sched_setscheduler); 3563 3564/** 3565 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace. 3566 * @p: the task in question. 3567 * @policy: new policy. 3568 * @param: structure containing the new RT priority. 3569 * 3570 * Just like sched_setscheduler, only don't bother checking if the 3571 * current context has permission. For example, this is needed in 3572 * stop_machine(): we create temporary high priority worker threads, 3573 * but our caller might not have that capability. 3574 * 3575 * Return: 0 on success. An error code otherwise. 3576 */ 3577int sched_setscheduler_nocheck(struct task_struct *p, int policy, 3578 const struct sched_param *param) 3579{ 3580 return __sched_setscheduler(p, policy, param, false); 3581} 3582 3583static int 3584do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param) 3585{ 3586 struct sched_param lparam; 3587 struct task_struct *p; 3588 int retval; 3589 3590 if (!param || pid < 0) 3591 return -EINVAL; 3592 if (copy_from_user(&lparam, param, sizeof(struct sched_param))) 3593 return -EFAULT; 3594 3595 rcu_read_lock(); 3596 retval = -ESRCH; 3597 p = find_process_by_pid(pid); 3598 if (p != NULL) 3599 retval = sched_setscheduler(p, policy, &lparam); 3600 rcu_read_unlock(); 3601 3602 return retval; 3603} 3604 3605/** 3606 * sys_sched_setscheduler - set/change the scheduler policy and RT priority 3607 * @pid: the pid in question. 3608 * @policy: new policy. 3609 * @param: structure containing the new RT priority. 3610 * 3611 * Return: 0 on success. An error code otherwise. 3612 */ 3613SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy, 3614 struct sched_param __user *, param) 3615{ 3616 /* negative values for policy are not valid */ 3617 if (policy < 0) 3618 return -EINVAL; 3619 3620 return do_sched_setscheduler(pid, policy, param); 3621} 3622 3623/** 3624 * sys_sched_setparam - set/change the RT priority of a thread 3625 * @pid: the pid in question. 3626 * @param: structure containing the new RT priority. 3627 * 3628 * Return: 0 on success. An error code otherwise. 3629 */ 3630SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param) 3631{ 3632 return do_sched_setscheduler(pid, -1, param); 3633} 3634 3635/** 3636 * sys_sched_getscheduler - get the policy (scheduling class) of a thread 3637 * @pid: the pid in question. 3638 * 3639 * Return: On success, the policy of the thread. Otherwise, a negative error 3640 * code. 3641 */ 3642SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid) 3643{ 3644 struct task_struct *p; 3645 int retval; 3646 3647 if (pid < 0) 3648 return -EINVAL; 3649 3650 retval = -ESRCH; 3651 rcu_read_lock(); 3652 p = find_process_by_pid(pid); 3653 if (p) { 3654 retval = security_task_getscheduler(p); 3655 if (!retval) 3656 retval = p->policy 3657 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0); 3658 } 3659 rcu_read_unlock(); 3660 return retval; 3661} 3662 3663/** 3664 * sys_sched_getparam - get the RT priority of a thread 3665 * @pid: the pid in question. 3666 * @param: structure containing the RT priority. 3667 * 3668 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error 3669 * code. 3670 */ 3671SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param) 3672{ 3673 struct sched_param lp; 3674 struct task_struct *p; 3675 int retval; 3676 3677 if (!param || pid < 0) 3678 return -EINVAL; 3679 3680 rcu_read_lock(); 3681 p = find_process_by_pid(pid); 3682 retval = -ESRCH; 3683 if (!p) 3684 goto out_unlock; 3685 3686 retval = security_task_getscheduler(p); 3687 if (retval) 3688 goto out_unlock; 3689 3690 lp.sched_priority = p->rt_priority; 3691 rcu_read_unlock(); 3692 3693 /* 3694 * This one might sleep, we cannot do it with a spinlock held ... 3695 */ 3696 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0; 3697 3698 return retval; 3699 3700out_unlock: 3701 rcu_read_unlock(); 3702 return retval; 3703} 3704 3705long sched_setaffinity(pid_t pid, const struct cpumask *in_mask) 3706{ 3707 cpumask_var_t cpus_allowed, new_mask; 3708 struct task_struct *p; 3709 int retval; 3710 3711 get_online_cpus(); 3712 rcu_read_lock(); 3713 3714 p = find_process_by_pid(pid); 3715 if (!p) { 3716 rcu_read_unlock(); 3717 put_online_cpus(); 3718 return -ESRCH; 3719 } 3720 3721 /* Prevent p going away */ 3722 get_task_struct(p); 3723 rcu_read_unlock(); 3724 3725 if (p->flags & PF_NO_SETAFFINITY) { 3726 retval = -EINVAL; 3727 goto out_put_task; 3728 } 3729 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) { 3730 retval = -ENOMEM; 3731 goto out_put_task; 3732 } 3733 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) { 3734 retval = -ENOMEM; 3735 goto out_free_cpus_allowed; 3736 } 3737 retval = -EPERM; 3738 if (!check_same_owner(p)) { 3739 rcu_read_lock(); 3740 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) { 3741 rcu_read_unlock(); 3742 goto out_unlock; 3743 } 3744 rcu_read_unlock(); 3745 } 3746 3747 retval = security_task_setscheduler(p); 3748 if (retval) 3749 goto out_unlock; 3750 3751 cpuset_cpus_allowed(p, cpus_allowed); 3752 cpumask_and(new_mask, in_mask, cpus_allowed); 3753again: 3754 retval = set_cpus_allowed_ptr(p, new_mask); 3755 3756 if (!retval) { 3757 cpuset_cpus_allowed(p, cpus_allowed); 3758 if (!cpumask_subset(new_mask, cpus_allowed)) { 3759 /* 3760 * We must have raced with a concurrent cpuset 3761 * update. Just reset the cpus_allowed to the 3762 * cpuset's cpus_allowed 3763 */ 3764 cpumask_copy(new_mask, cpus_allowed); 3765 goto again; 3766 } 3767 } 3768out_unlock: 3769 free_cpumask_var(new_mask); 3770out_free_cpus_allowed: 3771 free_cpumask_var(cpus_allowed); 3772out_put_task: 3773 put_task_struct(p); 3774 put_online_cpus(); 3775 return retval; 3776} 3777 3778static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len, 3779 struct cpumask *new_mask) 3780{ 3781 if (len < cpumask_size()) 3782 cpumask_clear(new_mask); 3783 else if (len > cpumask_size()) 3784 len = cpumask_size(); 3785 3786 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0; 3787} 3788 3789/** 3790 * sys_sched_setaffinity - set the cpu affinity of a process 3791 * @pid: pid of the process 3792 * @len: length in bytes of the bitmask pointed to by user_mask_ptr 3793 * @user_mask_ptr: user-space pointer to the new cpu mask 3794 * 3795 * Return: 0 on success. An error code otherwise. 3796 */ 3797SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len, 3798 unsigned long __user *, user_mask_ptr) 3799{ 3800 cpumask_var_t new_mask; 3801 int retval; 3802 3803 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) 3804 return -ENOMEM; 3805 3806 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask); 3807 if (retval == 0) 3808 retval = sched_setaffinity(pid, new_mask); 3809 free_cpumask_var(new_mask); 3810 return retval; 3811} 3812 3813long sched_getaffinity(pid_t pid, struct cpumask *mask) 3814{ 3815 struct task_struct *p; 3816 unsigned long flags; 3817 int retval; 3818 3819 get_online_cpus(); 3820 rcu_read_lock(); 3821 3822 retval = -ESRCH; 3823 p = find_process_by_pid(pid); 3824 if (!p) 3825 goto out_unlock; 3826 3827 retval = security_task_getscheduler(p); 3828 if (retval) 3829 goto out_unlock; 3830 3831 raw_spin_lock_irqsave(&p->pi_lock, flags); 3832 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask); 3833 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 3834 3835out_unlock: 3836 rcu_read_unlock(); 3837 put_online_cpus(); 3838 3839 return retval; 3840} 3841 3842/** 3843 * sys_sched_getaffinity - get the cpu affinity of a process 3844 * @pid: pid of the process 3845 * @len: length in bytes of the bitmask pointed to by user_mask_ptr 3846 * @user_mask_ptr: user-space pointer to hold the current cpu mask 3847 * 3848 * Return: 0 on success. An error code otherwise. 3849 */ 3850SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len, 3851 unsigned long __user *, user_mask_ptr) 3852{ 3853 int ret; 3854 cpumask_var_t mask; 3855 3856 if ((len * BITS_PER_BYTE) < nr_cpu_ids) 3857 return -EINVAL; 3858 if (len & (sizeof(unsigned long)-1)) 3859 return -EINVAL; 3860 3861 if (!alloc_cpumask_var(&mask, GFP_KERNEL)) 3862 return -ENOMEM; 3863 3864 ret = sched_getaffinity(pid, mask); 3865 if (ret == 0) { 3866 size_t retlen = min_t(size_t, len, cpumask_size()); 3867 3868 if (copy_to_user(user_mask_ptr, mask, retlen)) 3869 ret = -EFAULT; 3870 else 3871 ret = retlen; 3872 } 3873 free_cpumask_var(mask); 3874 3875 return ret; 3876} 3877 3878/** 3879 * sys_sched_yield - yield the current processor to other threads. 3880 * 3881 * This function yields the current CPU to other tasks. If there are no 3882 * other threads running on this CPU then this function will return. 3883 * 3884 * Return: 0. 3885 */ 3886SYSCALL_DEFINE0(sched_yield) 3887{ 3888 struct rq *rq = this_rq_lock(); 3889 3890 schedstat_inc(rq, yld_count); 3891 current->sched_class->yield_task(rq); 3892 3893 /* 3894 * Since we are going to call schedule() anyway, there's 3895 * no need to preempt or enable interrupts: 3896 */ 3897 __release(rq->lock); 3898 spin_release(&rq->lock.dep_map, 1, _THIS_IP_); 3899 do_raw_spin_unlock(&rq->lock); 3900 sched_preempt_enable_no_resched(); 3901 3902 schedule(); 3903 3904 return 0; 3905} 3906 3907static void __cond_resched(void) 3908{ 3909 __preempt_count_add(PREEMPT_ACTIVE); 3910 __schedule(); 3911 __preempt_count_sub(PREEMPT_ACTIVE); 3912} 3913 3914int __sched _cond_resched(void) 3915{ 3916 if (should_resched()) { 3917 __cond_resched(); 3918 return 1; 3919 } 3920 return 0; 3921} 3922EXPORT_SYMBOL(_cond_resched); 3923 3924/* 3925 * __cond_resched_lock() - if a reschedule is pending, drop the given lock, 3926 * call schedule, and on return reacquire the lock. 3927 * 3928 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level 3929 * operations here to prevent schedule() from being called twice (once via 3930 * spin_unlock(), once by hand). 3931 */ 3932int __cond_resched_lock(spinlock_t *lock) 3933{ 3934 int resched = should_resched(); 3935 int ret = 0; 3936 3937 lockdep_assert_held(lock); 3938 3939 if (spin_needbreak(lock) || resched) { 3940 spin_unlock(lock); 3941 if (resched) 3942 __cond_resched(); 3943 else 3944 cpu_relax(); 3945 ret = 1; 3946 spin_lock(lock); 3947 } 3948 return ret; 3949} 3950EXPORT_SYMBOL(__cond_resched_lock); 3951 3952int __sched __cond_resched_softirq(void) 3953{ 3954 BUG_ON(!in_softirq()); 3955 3956 if (should_resched()) { 3957 local_bh_enable(); 3958 __cond_resched(); 3959 local_bh_disable(); 3960 return 1; 3961 } 3962 return 0; 3963} 3964EXPORT_SYMBOL(__cond_resched_softirq); 3965 3966/** 3967 * yield - yield the current processor to other threads. 3968 * 3969 * Do not ever use this function, there's a 99% chance you're doing it wrong. 3970 * 3971 * The scheduler is at all times free to pick the calling task as the most 3972 * eligible task to run, if removing the yield() call from your code breaks 3973 * it, its already broken. 3974 * 3975 * Typical broken usage is: 3976 * 3977 * while (!event) 3978 * yield(); 3979 * 3980 * where one assumes that yield() will let 'the other' process run that will 3981 * make event true. If the current task is a SCHED_FIFO task that will never 3982 * happen. Never use yield() as a progress guarantee!! 3983 * 3984 * If you want to use yield() to wait for something, use wait_event(). 3985 * If you want to use yield() to be 'nice' for others, use cond_resched(). 3986 * If you still want to use yield(), do not! 3987 */ 3988void __sched yield(void) 3989{ 3990 set_current_state(TASK_RUNNING); 3991 sys_sched_yield(); 3992} 3993EXPORT_SYMBOL(yield); 3994 3995/** 3996 * yield_to - yield the current processor to another thread in 3997 * your thread group, or accelerate that thread toward the 3998 * processor it's on. 3999 * @p: target task 4000 * @preempt: whether task preemption is allowed or not 4001 * 4002 * It's the caller's job to ensure that the target task struct 4003 * can't go away on us before we can do any checks. 4004 * 4005 * Return: 4006 * true (>0) if we indeed boosted the target task. 4007 * false (0) if we failed to boost the target. 4008 * -ESRCH if there's no task to yield to. 4009 */ 4010bool __sched yield_to(struct task_struct *p, bool preempt) 4011{ 4012 struct task_struct *curr = current; 4013 struct rq *rq, *p_rq; 4014 unsigned long flags; 4015 int yielded = 0; 4016 4017 local_irq_save(flags); 4018 rq = this_rq(); 4019 4020again: 4021 p_rq = task_rq(p); 4022 /* 4023 * If we're the only runnable task on the rq and target rq also 4024 * has only one task, there's absolutely no point in yielding. 4025 */ 4026 if (rq->nr_running == 1 && p_rq->nr_running == 1) { 4027 yielded = -ESRCH; 4028 goto out_irq; 4029 } 4030 4031 double_rq_lock(rq, p_rq); 4032 while (task_rq(p) != p_rq) { 4033 double_rq_unlock(rq, p_rq); 4034 goto again; 4035 } 4036 4037 if (!curr->sched_class->yield_to_task) 4038 goto out_unlock; 4039 4040 if (curr->sched_class != p->sched_class) 4041 goto out_unlock; 4042 4043 if (task_running(p_rq, p) || p->state) 4044 goto out_unlock; 4045 4046 yielded = curr->sched_class->yield_to_task(rq, p, preempt); 4047 if (yielded) { 4048 schedstat_inc(rq, yld_count); 4049 /* 4050 * Make p's CPU reschedule; pick_next_entity takes care of 4051 * fairness. 4052 */ 4053 if (preempt && rq != p_rq) 4054 resched_task(p_rq->curr); 4055 } 4056 4057out_unlock: 4058 double_rq_unlock(rq, p_rq); 4059out_irq: 4060 local_irq_restore(flags); 4061 4062 if (yielded > 0) 4063 schedule(); 4064 4065 return yielded; 4066} 4067EXPORT_SYMBOL_GPL(yield_to); 4068 4069/* 4070 * This task is about to go to sleep on IO. Increment rq->nr_iowait so 4071 * that process accounting knows that this is a task in IO wait state. 4072 */ 4073void __sched io_schedule(void) 4074{ 4075 struct rq *rq = raw_rq(); 4076 4077 delayacct_blkio_start(); 4078 atomic_inc(&rq->nr_iowait); 4079 blk_flush_plug(current); 4080 current->in_iowait = 1; 4081 schedule(); 4082 current->in_iowait = 0; 4083 atomic_dec(&rq->nr_iowait); 4084 delayacct_blkio_end(); 4085} 4086EXPORT_SYMBOL(io_schedule); 4087 4088long __sched io_schedule_timeout(long timeout) 4089{ 4090 struct rq *rq = raw_rq(); 4091 long ret; 4092 4093 delayacct_blkio_start(); 4094 atomic_inc(&rq->nr_iowait); 4095 blk_flush_plug(current); 4096 current->in_iowait = 1; 4097 ret = schedule_timeout(timeout); 4098 current->in_iowait = 0; 4099 atomic_dec(&rq->nr_iowait); 4100 delayacct_blkio_end(); 4101 return ret; 4102} 4103 4104/** 4105 * sys_sched_get_priority_max - return maximum RT priority. 4106 * @policy: scheduling class. 4107 * 4108 * Return: On success, this syscall returns the maximum 4109 * rt_priority that can be used by a given scheduling class. 4110 * On failure, a negative error code is returned. 4111 */ 4112SYSCALL_DEFINE1(sched_get_priority_max, int, policy) 4113{ 4114 int ret = -EINVAL; 4115 4116 switch (policy) { 4117 case SCHED_FIFO: 4118 case SCHED_RR: 4119 ret = MAX_USER_RT_PRIO-1; 4120 break; 4121 case SCHED_NORMAL: 4122 case SCHED_BATCH: 4123 case SCHED_IDLE: 4124 ret = 0; 4125 break; 4126 } 4127 return ret; 4128} 4129 4130/** 4131 * sys_sched_get_priority_min - return minimum RT priority. 4132 * @policy: scheduling class. 4133 * 4134 * Return: On success, this syscall returns the minimum 4135 * rt_priority that can be used by a given scheduling class. 4136 * On failure, a negative error code is returned. 4137 */ 4138SYSCALL_DEFINE1(sched_get_priority_min, int, policy) 4139{ 4140 int ret = -EINVAL; 4141 4142 switch (policy) { 4143 case SCHED_FIFO: 4144 case SCHED_RR: 4145 ret = 1; 4146 break; 4147 case SCHED_NORMAL: 4148 case SCHED_BATCH: 4149 case SCHED_IDLE: 4150 ret = 0; 4151 } 4152 return ret; 4153} 4154 4155/** 4156 * sys_sched_rr_get_interval - return the default timeslice of a process. 4157 * @pid: pid of the process. 4158 * @interval: userspace pointer to the timeslice value. 4159 * 4160 * this syscall writes the default timeslice value of a given process 4161 * into the user-space timespec buffer. A value of '0' means infinity. 4162 * 4163 * Return: On success, 0 and the timeslice is in @interval. Otherwise, 4164 * an error code. 4165 */ 4166SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid, 4167 struct timespec __user *, interval) 4168{ 4169 struct task_struct *p; 4170 unsigned int time_slice; 4171 unsigned long flags; 4172 struct rq *rq; 4173 int retval; 4174 struct timespec t; 4175 4176 if (pid < 0) 4177 return -EINVAL; 4178 4179 retval = -ESRCH; 4180 rcu_read_lock(); 4181 p = find_process_by_pid(pid); 4182 if (!p) 4183 goto out_unlock; 4184 4185 retval = security_task_getscheduler(p); 4186 if (retval) 4187 goto out_unlock; 4188 4189 rq = task_rq_lock(p, &flags); 4190 time_slice = p->sched_class->get_rr_interval(rq, p); 4191 task_rq_unlock(rq, p, &flags); 4192 4193 rcu_read_unlock(); 4194 jiffies_to_timespec(time_slice, &t); 4195 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0; 4196 return retval; 4197 4198out_unlock: 4199 rcu_read_unlock(); 4200 return retval; 4201} 4202 4203static const char stat_nam[] = TASK_STATE_TO_CHAR_STR; 4204 4205void sched_show_task(struct task_struct *p) 4206{ 4207 unsigned long free = 0; 4208 int ppid; 4209 unsigned state; 4210 4211 state = p->state ? __ffs(p->state) + 1 : 0; 4212 printk(KERN_INFO "%-15.15s %c", p->comm, 4213 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?'); 4214#if BITS_PER_LONG == 32 4215 if (state == TASK_RUNNING) 4216 printk(KERN_CONT " running "); 4217 else 4218 printk(KERN_CONT " %08lx ", thread_saved_pc(p)); 4219#else 4220 if (state == TASK_RUNNING) 4221 printk(KERN_CONT " running task "); 4222 else 4223 printk(KERN_CONT " %016lx ", thread_saved_pc(p)); 4224#endif 4225#ifdef CONFIG_DEBUG_STACK_USAGE 4226 free = stack_not_used(p); 4227#endif 4228 rcu_read_lock(); 4229 ppid = task_pid_nr(rcu_dereference(p->real_parent)); 4230 rcu_read_unlock(); 4231 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free, 4232 task_pid_nr(p), ppid, 4233 (unsigned long)task_thread_info(p)->flags); 4234 4235 print_worker_info(KERN_INFO, p); 4236 show_stack(p, NULL); 4237} 4238 4239void show_state_filter(unsigned long state_filter) 4240{ 4241 struct task_struct *g, *p; 4242 4243#if BITS_PER_LONG == 32 4244 printk(KERN_INFO 4245 " task PC stack pid father\n"); 4246#else 4247 printk(KERN_INFO 4248 " task PC stack pid father\n"); 4249#endif 4250 rcu_read_lock(); 4251 do_each_thread(g, p) { 4252 /* 4253 * reset the NMI-timeout, listing all files on a slow 4254 * console might take a lot of time: 4255 */ 4256 touch_nmi_watchdog(); 4257 if (!state_filter || (p->state & state_filter)) 4258 sched_show_task(p); 4259 } while_each_thread(g, p); 4260 4261 touch_all_softlockup_watchdogs(); 4262 4263#ifdef CONFIG_SCHED_DEBUG 4264 sysrq_sched_debug_show(); 4265#endif 4266 rcu_read_unlock(); 4267 /* 4268 * Only show locks if all tasks are dumped: 4269 */ 4270 if (!state_filter) 4271 debug_show_all_locks(); 4272} 4273 4274void init_idle_bootup_task(struct task_struct *idle) 4275{ 4276 idle->sched_class = &idle_sched_class; 4277} 4278 4279/** 4280 * init_idle - set up an idle thread for a given CPU 4281 * @idle: task in question 4282 * @cpu: cpu the idle task belongs to 4283 * 4284 * NOTE: this function does not set the idle thread's NEED_RESCHED 4285 * flag, to make booting more robust. 4286 */ 4287void init_idle(struct task_struct *idle, int cpu) 4288{ 4289 struct rq *rq = cpu_rq(cpu); 4290 unsigned long flags; 4291 4292 raw_spin_lock_irqsave(&rq->lock, flags); 4293 4294 __sched_fork(0, idle); 4295 idle->state = TASK_RUNNING; 4296 idle->se.exec_start = sched_clock(); 4297 4298 do_set_cpus_allowed(idle, cpumask_of(cpu)); 4299 /* 4300 * We're having a chicken and egg problem, even though we are 4301 * holding rq->lock, the cpu isn't yet set to this cpu so the 4302 * lockdep check in task_group() will fail. 4303 * 4304 * Similar case to sched_fork(). / Alternatively we could 4305 * use task_rq_lock() here and obtain the other rq->lock. 4306 * 4307 * Silence PROVE_RCU 4308 */ 4309 rcu_read_lock(); 4310 __set_task_cpu(idle, cpu); 4311 rcu_read_unlock(); 4312 4313 rq->curr = rq->idle = idle; 4314#if defined(CONFIG_SMP) 4315 idle->on_cpu = 1; 4316#endif 4317 raw_spin_unlock_irqrestore(&rq->lock, flags); 4318 4319 /* Set the preempt count _outside_ the spinlocks! */ 4320 init_idle_preempt_count(idle, cpu); 4321 4322 /* 4323 * The idle tasks have their own, simple scheduling class: 4324 */ 4325 idle->sched_class = &idle_sched_class; 4326 ftrace_graph_init_idle_task(idle, cpu); 4327 vtime_init_idle(idle, cpu); 4328#if defined(CONFIG_SMP) 4329 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu); 4330#endif 4331} 4332 4333#ifdef CONFIG_SMP 4334void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask) 4335{ 4336 if (p->sched_class && p->sched_class->set_cpus_allowed) 4337 p->sched_class->set_cpus_allowed(p, new_mask); 4338 4339 cpumask_copy(&p->cpus_allowed, new_mask); 4340 p->nr_cpus_allowed = cpumask_weight(new_mask); 4341} 4342 4343/* 4344 * This is how migration works: 4345 * 4346 * 1) we invoke migration_cpu_stop() on the target CPU using 4347 * stop_one_cpu(). 4348 * 2) stopper starts to run (implicitly forcing the migrated thread 4349 * off the CPU) 4350 * 3) it checks whether the migrated task is still in the wrong runqueue. 4351 * 4) if it's in the wrong runqueue then the migration thread removes 4352 * it and puts it into the right queue. 4353 * 5) stopper completes and stop_one_cpu() returns and the migration 4354 * is done. 4355 */ 4356 4357/* 4358 * Change a given task's CPU affinity. Migrate the thread to a 4359 * proper CPU and schedule it away if the CPU it's executing on 4360 * is removed from the allowed bitmask. 4361 * 4362 * NOTE: the caller must have a valid reference to the task, the 4363 * task must not exit() & deallocate itself prematurely. The 4364 * call is not atomic; no spinlocks may be held. 4365 */ 4366int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask) 4367{ 4368 unsigned long flags; 4369 struct rq *rq; 4370 unsigned int dest_cpu; 4371 int ret = 0; 4372 4373 rq = task_rq_lock(p, &flags); 4374 4375 if (cpumask_equal(&p->cpus_allowed, new_mask)) 4376 goto out; 4377 4378 if (!cpumask_intersects(new_mask, cpu_active_mask)) { 4379 ret = -EINVAL; 4380 goto out; 4381 } 4382 4383 do_set_cpus_allowed(p, new_mask); 4384 4385 /* Can the task run on the task's current CPU? If so, we're done */ 4386 if (cpumask_test_cpu(task_cpu(p), new_mask)) 4387 goto out; 4388 4389 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask); 4390 if (p->on_rq) { 4391 struct migration_arg arg = { p, dest_cpu }; 4392 /* Need help from migration thread: drop lock and wait. */ 4393 task_rq_unlock(rq, p, &flags); 4394 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg); 4395 tlb_migrate_finish(p->mm); 4396 return 0; 4397 } 4398out: 4399 task_rq_unlock(rq, p, &flags); 4400 4401 return ret; 4402} 4403EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr); 4404 4405/* 4406 * Move (not current) task off this cpu, onto dest cpu. We're doing 4407 * this because either it can't run here any more (set_cpus_allowed() 4408 * away from this CPU, or CPU going down), or because we're 4409 * attempting to rebalance this task on exec (sched_exec). 4410 * 4411 * So we race with normal scheduler movements, but that's OK, as long 4412 * as the task is no longer on this CPU. 4413 * 4414 * Returns non-zero if task was successfully migrated. 4415 */ 4416static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu) 4417{ 4418 struct rq *rq_dest, *rq_src; 4419 int ret = 0; 4420 4421 if (unlikely(!cpu_active(dest_cpu))) 4422 return ret; 4423 4424 rq_src = cpu_rq(src_cpu); 4425 rq_dest = cpu_rq(dest_cpu); 4426 4427 raw_spin_lock(&p->pi_lock); 4428 double_rq_lock(rq_src, rq_dest); 4429 /* Already moved. */ 4430 if (task_cpu(p) != src_cpu) 4431 goto done; 4432 /* Affinity changed (again). */ 4433 if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p))) 4434 goto fail; 4435 4436 /* 4437 * If we're not on a rq, the next wake-up will ensure we're 4438 * placed properly. 4439 */ 4440 if (p->on_rq) { 4441 dequeue_task(rq_src, p, 0); 4442 set_task_cpu(p, dest_cpu); 4443 enqueue_task(rq_dest, p, 0); 4444 check_preempt_curr(rq_dest, p, 0); 4445 } 4446done: 4447 ret = 1; 4448fail: 4449 double_rq_unlock(rq_src, rq_dest); 4450 raw_spin_unlock(&p->pi_lock); 4451 return ret; 4452} 4453 4454#ifdef CONFIG_NUMA_BALANCING 4455/* Migrate current task p to target_cpu */ 4456int migrate_task_to(struct task_struct *p, int target_cpu) 4457{ 4458 struct migration_arg arg = { p, target_cpu }; 4459 int curr_cpu = task_cpu(p); 4460 4461 if (curr_cpu == target_cpu) 4462 return 0; 4463 4464 if (!cpumask_test_cpu(target_cpu, tsk_cpus_allowed(p))) 4465 return -EINVAL; 4466 4467 /* TODO: This is not properly updating schedstats */ 4468 4469 return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg); 4470} 4471#endif 4472 4473/* 4474 * migration_cpu_stop - this will be executed by a highprio stopper thread 4475 * and performs thread migration by bumping thread off CPU then 4476 * 'pushing' onto another runqueue. 4477 */ 4478static int migration_cpu_stop(void *data) 4479{ 4480 struct migration_arg *arg = data; 4481 4482 /* 4483 * The original target cpu might have gone down and we might 4484 * be on another cpu but it doesn't matter. 4485 */ 4486 local_irq_disable(); 4487 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu); 4488 local_irq_enable(); 4489 return 0; 4490} 4491 4492#ifdef CONFIG_HOTPLUG_CPU 4493 4494/* 4495 * Ensures that the idle task is using init_mm right before its cpu goes 4496 * offline. 4497 */ 4498void idle_task_exit(void) 4499{ 4500 struct mm_struct *mm = current->active_mm; 4501 4502 BUG_ON(cpu_online(smp_processor_id())); 4503 4504 if (mm != &init_mm) 4505 switch_mm(mm, &init_mm, current); 4506 mmdrop(mm); 4507} 4508 4509/* 4510 * Since this CPU is going 'away' for a while, fold any nr_active delta 4511 * we might have. Assumes we're called after migrate_tasks() so that the 4512 * nr_active count is stable. 4513 * 4514 * Also see the comment "Global load-average calculations". 4515 */ 4516static void calc_load_migrate(struct rq *rq) 4517{ 4518 long delta = calc_load_fold_active(rq); 4519 if (delta) 4520 atomic_long_add(delta, &calc_load_tasks); 4521} 4522 4523/* 4524 * Migrate all tasks from the rq, sleeping tasks will be migrated by 4525 * try_to_wake_up()->select_task_rq(). 4526 * 4527 * Called with rq->lock held even though we'er in stop_machine() and 4528 * there's no concurrency possible, we hold the required locks anyway 4529 * because of lock validation efforts. 4530 */ 4531static void migrate_tasks(unsigned int dead_cpu) 4532{ 4533 struct rq *rq = cpu_rq(dead_cpu); 4534 struct task_struct *next, *stop = rq->stop; 4535 int dest_cpu; 4536 4537 /* 4538 * Fudge the rq selection such that the below task selection loop 4539 * doesn't get stuck on the currently eligible stop task. 4540 * 4541 * We're currently inside stop_machine() and the rq is either stuck 4542 * in the stop_machine_cpu_stop() loop, or we're executing this code, 4543 * either way we should never end up calling schedule() until we're 4544 * done here. 4545 */ 4546 rq->stop = NULL; 4547 4548 /* 4549 * put_prev_task() and pick_next_task() sched 4550 * class method both need to have an up-to-date 4551 * value of rq->clock[_task] 4552 */ 4553 update_rq_clock(rq); 4554 4555 for ( ; ; ) { 4556 /* 4557 * There's this thread running, bail when that's the only 4558 * remaining thread. 4559 */ 4560 if (rq->nr_running == 1) 4561 break; 4562 4563 next = pick_next_task(rq); 4564 BUG_ON(!next); 4565 next->sched_class->put_prev_task(rq, next); 4566 4567 /* Find suitable destination for @next, with force if needed. */ 4568 dest_cpu = select_fallback_rq(dead_cpu, next); 4569 raw_spin_unlock(&rq->lock); 4570 4571 __migrate_task(next, dead_cpu, dest_cpu); 4572 4573 raw_spin_lock(&rq->lock); 4574 } 4575 4576 rq->stop = stop; 4577} 4578 4579#endif /* CONFIG_HOTPLUG_CPU */ 4580 4581#if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL) 4582 4583static struct ctl_table sd_ctl_dir[] = { 4584 { 4585 .procname = "sched_domain", 4586 .mode = 0555, 4587 }, 4588 {} 4589}; 4590 4591static struct ctl_table sd_ctl_root[] = { 4592 { 4593 .procname = "kernel", 4594 .mode = 0555, 4595 .child = sd_ctl_dir, 4596 }, 4597 {} 4598}; 4599 4600static struct ctl_table *sd_alloc_ctl_entry(int n) 4601{ 4602 struct ctl_table *entry = 4603 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL); 4604 4605 return entry; 4606} 4607 4608static void sd_free_ctl_entry(struct ctl_table **tablep) 4609{ 4610 struct ctl_table *entry; 4611 4612 /* 4613 * In the intermediate directories, both the child directory and 4614 * procname are dynamically allocated and could fail but the mode 4615 * will always be set. In the lowest directory the names are 4616 * static strings and all have proc handlers. 4617 */ 4618 for (entry = *tablep; entry->mode; entry++) { 4619 if (entry->child) 4620 sd_free_ctl_entry(&entry->child); 4621 if (entry->proc_handler == NULL) 4622 kfree(entry->procname); 4623 } 4624 4625 kfree(*tablep); 4626 *tablep = NULL; 4627} 4628 4629static int min_load_idx = 0; 4630static int max_load_idx = CPU_LOAD_IDX_MAX-1; 4631 4632static void 4633set_table_entry(struct ctl_table *entry, 4634 const char *procname, void *data, int maxlen, 4635 umode_t mode, proc_handler *proc_handler, 4636 bool load_idx) 4637{ 4638 entry->procname = procname; 4639 entry->data = data; 4640 entry->maxlen = maxlen; 4641 entry->mode = mode; 4642 entry->proc_handler = proc_handler; 4643 4644 if (load_idx) { 4645 entry->extra1 = &min_load_idx; 4646 entry->extra2 = &max_load_idx; 4647 } 4648} 4649 4650static struct ctl_table * 4651sd_alloc_ctl_domain_table(struct sched_domain *sd) 4652{ 4653 struct ctl_table *table = sd_alloc_ctl_entry(13); 4654 4655 if (table == NULL) 4656 return NULL; 4657 4658 set_table_entry(&table[0], "min_interval", &sd->min_interval, 4659 sizeof(long), 0644, proc_doulongvec_minmax, false); 4660 set_table_entry(&table[1], "max_interval", &sd->max_interval, 4661 sizeof(long), 0644, proc_doulongvec_minmax, false); 4662 set_table_entry(&table[2], "busy_idx", &sd->busy_idx, 4663 sizeof(int), 0644, proc_dointvec_minmax, true); 4664 set_table_entry(&table[3], "idle_idx", &sd->idle_idx, 4665 sizeof(int), 0644, proc_dointvec_minmax, true); 4666 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx, 4667 sizeof(int), 0644, proc_dointvec_minmax, true); 4668 set_table_entry(&table[5], "wake_idx", &sd->wake_idx, 4669 sizeof(int), 0644, proc_dointvec_minmax, true); 4670 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx, 4671 sizeof(int), 0644, proc_dointvec_minmax, true); 4672 set_table_entry(&table[7], "busy_factor", &sd->busy_factor, 4673 sizeof(int), 0644, proc_dointvec_minmax, false); 4674 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct, 4675 sizeof(int), 0644, proc_dointvec_minmax, false); 4676 set_table_entry(&table[9], "cache_nice_tries", 4677 &sd->cache_nice_tries, 4678 sizeof(int), 0644, proc_dointvec_minmax, false); 4679 set_table_entry(&table[10], "flags", &sd->flags, 4680 sizeof(int), 0644, proc_dointvec_minmax, false); 4681 set_table_entry(&table[11], "name", sd->name, 4682 CORENAME_MAX_SIZE, 0444, proc_dostring, false); 4683 /* &table[12] is terminator */ 4684 4685 return table; 4686} 4687 4688static struct ctl_table *sd_alloc_ctl_cpu_table(int cpu) 4689{ 4690 struct ctl_table *entry, *table; 4691 struct sched_domain *sd; 4692 int domain_num = 0, i; 4693 char buf[32]; 4694 4695 for_each_domain(cpu, sd) 4696 domain_num++; 4697 entry = table = sd_alloc_ctl_entry(domain_num + 1); 4698 if (table == NULL) 4699 return NULL; 4700 4701 i = 0; 4702 for_each_domain(cpu, sd) { 4703 snprintf(buf, 32, "domain%d", i); 4704 entry->procname = kstrdup(buf, GFP_KERNEL); 4705 entry->mode = 0555; 4706 entry->child = sd_alloc_ctl_domain_table(sd); 4707 entry++; 4708 i++; 4709 } 4710 return table; 4711} 4712 4713static struct ctl_table_header *sd_sysctl_header; 4714static void register_sched_domain_sysctl(void) 4715{ 4716 int i, cpu_num = num_possible_cpus(); 4717 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1); 4718 char buf[32]; 4719 4720 WARN_ON(sd_ctl_dir[0].child); 4721 sd_ctl_dir[0].child = entry; 4722 4723 if (entry == NULL) 4724 return; 4725 4726 for_each_possible_cpu(i) { 4727 snprintf(buf, 32, "cpu%d", i); 4728 entry->procname = kstrdup(buf, GFP_KERNEL); 4729 entry->mode = 0555; 4730 entry->child = sd_alloc_ctl_cpu_table(i); 4731 entry++; 4732 } 4733 4734 WARN_ON(sd_sysctl_header); 4735 sd_sysctl_header = register_sysctl_table(sd_ctl_root); 4736} 4737 4738/* may be called multiple times per register */ 4739static void unregister_sched_domain_sysctl(void) 4740{ 4741 if (sd_sysctl_header) 4742 unregister_sysctl_table(sd_sysctl_header); 4743 sd_sysctl_header = NULL; 4744 if (sd_ctl_dir[0].child) 4745 sd_free_ctl_entry(&sd_ctl_dir[0].child); 4746} 4747#else 4748static void register_sched_domain_sysctl(void) 4749{ 4750} 4751static void unregister_sched_domain_sysctl(void) 4752{ 4753} 4754#endif 4755 4756static void set_rq_online(struct rq *rq) 4757{ 4758 if (!rq->online) { 4759 const struct sched_class *class; 4760 4761 cpumask_set_cpu(rq->cpu, rq->rd->online); 4762 rq->online = 1; 4763 4764 for_each_class(class) { 4765 if (class->rq_online) 4766 class->rq_online(rq); 4767 } 4768 } 4769} 4770 4771static void set_rq_offline(struct rq *rq) 4772{ 4773 if (rq->online) { 4774 const struct sched_class *class; 4775 4776 for_each_class(class) { 4777 if (class->rq_offline) 4778 class->rq_offline(rq); 4779 } 4780 4781 cpumask_clear_cpu(rq->cpu, rq->rd->online); 4782 rq->online = 0; 4783 } 4784} 4785 4786/* 4787 * migration_call - callback that gets triggered when a CPU is added. 4788 * Here we can start up the necessary migration thread for the new CPU. 4789 */ 4790static int 4791migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu) 4792{ 4793 int cpu = (long)hcpu; 4794 unsigned long flags; 4795 struct rq *rq = cpu_rq(cpu); 4796 4797 switch (action & ~CPU_TASKS_FROZEN) { 4798 4799 case CPU_UP_PREPARE: 4800 rq->calc_load_update = calc_load_update; 4801 break; 4802 4803 case CPU_ONLINE: 4804 /* Update our root-domain */ 4805 raw_spin_lock_irqsave(&rq->lock, flags); 4806 if (rq->rd) { 4807 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span)); 4808 4809 set_rq_online(rq); 4810 } 4811 raw_spin_unlock_irqrestore(&rq->lock, flags); 4812 break; 4813 4814#ifdef CONFIG_HOTPLUG_CPU 4815 case CPU_DYING: 4816 sched_ttwu_pending(); 4817 /* Update our root-domain */ 4818 raw_spin_lock_irqsave(&rq->lock, flags); 4819 if (rq->rd) { 4820 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span)); 4821 set_rq_offline(rq); 4822 } 4823 migrate_tasks(cpu); 4824 BUG_ON(rq->nr_running != 1); /* the migration thread */ 4825 raw_spin_unlock_irqrestore(&rq->lock, flags); 4826 break; 4827 4828 case CPU_DEAD: 4829 calc_load_migrate(rq); 4830 break; 4831#endif 4832 } 4833 4834 update_max_interval(); 4835 4836 return NOTIFY_OK; 4837} 4838 4839/* 4840 * Register at high priority so that task migration (migrate_all_tasks) 4841 * happens before everything else. This has to be lower priority than 4842 * the notifier in the perf_event subsystem, though. 4843 */ 4844static struct notifier_block migration_notifier = { 4845 .notifier_call = migration_call, 4846 .priority = CPU_PRI_MIGRATION, 4847}; 4848 4849static int sched_cpu_active(struct notifier_block *nfb, 4850 unsigned long action, void *hcpu) 4851{ 4852 switch (action & ~CPU_TASKS_FROZEN) { 4853 case CPU_STARTING: 4854 case CPU_DOWN_FAILED: 4855 set_cpu_active((long)hcpu, true); 4856 return NOTIFY_OK; 4857 default: 4858 return NOTIFY_DONE; 4859 } 4860} 4861 4862static int sched_cpu_inactive(struct notifier_block *nfb, 4863 unsigned long action, void *hcpu) 4864{ 4865 switch (action & ~CPU_TASKS_FROZEN) { 4866 case CPU_DOWN_PREPARE: 4867 set_cpu_active((long)hcpu, false); 4868 return NOTIFY_OK; 4869 default: 4870 return NOTIFY_DONE; 4871 } 4872} 4873 4874static int __init migration_init(void) 4875{ 4876 void *cpu = (void *)(long)smp_processor_id(); 4877 int err; 4878 4879 /* Initialize migration for the boot CPU */ 4880 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu); 4881 BUG_ON(err == NOTIFY_BAD); 4882 migration_call(&migration_notifier, CPU_ONLINE, cpu); 4883 register_cpu_notifier(&migration_notifier); 4884 4885 /* Register cpu active notifiers */ 4886 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE); 4887 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE); 4888 4889 return 0; 4890} 4891early_initcall(migration_init); 4892#endif 4893 4894#ifdef CONFIG_SMP 4895 4896static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */ 4897 4898#ifdef CONFIG_SCHED_DEBUG 4899 4900static __read_mostly int sched_debug_enabled; 4901 4902static int __init sched_debug_setup(char *str) 4903{ 4904 sched_debug_enabled = 1; 4905 4906 return 0; 4907} 4908early_param("sched_debug", sched_debug_setup); 4909 4910static inline bool sched_debug(void) 4911{ 4912 return sched_debug_enabled; 4913} 4914 4915static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level, 4916 struct cpumask *groupmask) 4917{ 4918 struct sched_group *group = sd->groups; 4919 char str[256]; 4920 4921 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd)); 4922 cpumask_clear(groupmask); 4923 4924 printk(KERN_DEBUG "%*s domain %d: ", level, "", level); 4925 4926 if (!(sd->flags & SD_LOAD_BALANCE)) { 4927 printk("does not load-balance\n"); 4928 if (sd->parent) 4929 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain" 4930 " has parent"); 4931 return -1; 4932 } 4933 4934 printk(KERN_CONT "span %s level %s\n", str, sd->name); 4935 4936 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) { 4937 printk(KERN_ERR "ERROR: domain->span does not contain " 4938 "CPU%d\n", cpu); 4939 } 4940 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) { 4941 printk(KERN_ERR "ERROR: domain->groups does not contain" 4942 " CPU%d\n", cpu); 4943 } 4944 4945 printk(KERN_DEBUG "%*s groups:", level + 1, ""); 4946 do { 4947 if (!group) { 4948 printk("\n"); 4949 printk(KERN_ERR "ERROR: group is NULL\n"); 4950 break; 4951 } 4952 4953 /* 4954 * Even though we initialize ->power to something semi-sane, 4955 * we leave power_orig unset. This allows us to detect if 4956 * domain iteration is still funny without causing /0 traps. 4957 */ 4958 if (!group->sgp->power_orig) { 4959 printk(KERN_CONT "\n"); 4960 printk(KERN_ERR "ERROR: domain->cpu_power not " 4961 "set\n"); 4962 break; 4963 } 4964 4965 if (!cpumask_weight(sched_group_cpus(group))) { 4966 printk(KERN_CONT "\n"); 4967 printk(KERN_ERR "ERROR: empty group\n"); 4968 break; 4969 } 4970 4971 if (!(sd->flags & SD_OVERLAP) && 4972 cpumask_intersects(groupmask, sched_group_cpus(group))) { 4973 printk(KERN_CONT "\n"); 4974 printk(KERN_ERR "ERROR: repeated CPUs\n"); 4975 break; 4976 } 4977 4978 cpumask_or(groupmask, groupmask, sched_group_cpus(group)); 4979 4980 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group)); 4981 4982 printk(KERN_CONT " %s", str); 4983 if (group->sgp->power != SCHED_POWER_SCALE) { 4984 printk(KERN_CONT " (cpu_power = %d)", 4985 group->sgp->power); 4986 } 4987 4988 group = group->next; 4989 } while (group != sd->groups); 4990 printk(KERN_CONT "\n"); 4991 4992 if (!cpumask_equal(sched_domain_span(sd), groupmask)) 4993 printk(KERN_ERR "ERROR: groups don't span domain->span\n"); 4994 4995 if (sd->parent && 4996 !cpumask_subset(groupmask, sched_domain_span(sd->parent))) 4997 printk(KERN_ERR "ERROR: parent span is not a superset " 4998 "of domain->span\n"); 4999 return 0; 5000} 5001 5002static void sched_domain_debug(struct sched_domain *sd, int cpu) 5003{ 5004 int level = 0; 5005 5006 if (!sched_debug_enabled) 5007 return; 5008 5009 if (!sd) { 5010 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu); 5011 return; 5012 } 5013 5014 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu); 5015 5016 for (;;) { 5017 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask)) 5018 break; 5019 level++; 5020 sd = sd->parent; 5021 if (!sd) 5022 break; 5023 } 5024} 5025#else /* !CONFIG_SCHED_DEBUG */ 5026# define sched_domain_debug(sd, cpu) do { } while (0) 5027static inline bool sched_debug(void) 5028{ 5029 return false; 5030} 5031#endif /* CONFIG_SCHED_DEBUG */ 5032 5033static int sd_degenerate(struct sched_domain *sd) 5034{ 5035 if (cpumask_weight(sched_domain_span(sd)) == 1) 5036 return 1; 5037 5038 /* Following flags need at least 2 groups */ 5039 if (sd->flags & (SD_LOAD_BALANCE | 5040 SD_BALANCE_NEWIDLE | 5041 SD_BALANCE_FORK | 5042 SD_BALANCE_EXEC | 5043 SD_SHARE_CPUPOWER | 5044 SD_SHARE_PKG_RESOURCES)) { 5045 if (sd->groups != sd->groups->next) 5046 return 0; 5047 } 5048 5049 /* Following flags don't use groups */ 5050 if (sd->flags & (SD_WAKE_AFFINE)) 5051 return 0; 5052 5053 return 1; 5054} 5055 5056static int 5057sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent) 5058{ 5059 unsigned long cflags = sd->flags, pflags = parent->flags; 5060 5061 if (sd_degenerate(parent)) 5062 return 1; 5063 5064 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent))) 5065 return 0; 5066 5067 /* Flags needing groups don't count if only 1 group in parent */ 5068 if (parent->groups == parent->groups->next) { 5069 pflags &= ~(SD_LOAD_BALANCE | 5070 SD_BALANCE_NEWIDLE | 5071 SD_BALANCE_FORK | 5072 SD_BALANCE_EXEC | 5073 SD_SHARE_CPUPOWER | 5074 SD_SHARE_PKG_RESOURCES | 5075 SD_PREFER_SIBLING); 5076 if (nr_node_ids == 1) 5077 pflags &= ~SD_SERIALIZE; 5078 } 5079 if (~cflags & pflags) 5080 return 0; 5081 5082 return 1; 5083} 5084 5085static void free_rootdomain(struct rcu_head *rcu) 5086{ 5087 struct root_domain *rd = container_of(rcu, struct root_domain, rcu); 5088 5089 cpupri_cleanup(&rd->cpupri); 5090 free_cpumask_var(rd->rto_mask); 5091 free_cpumask_var(rd->online); 5092 free_cpumask_var(rd->span); 5093 kfree(rd); 5094} 5095 5096static void rq_attach_root(struct rq *rq, struct root_domain *rd) 5097{ 5098 struct root_domain *old_rd = NULL; 5099 unsigned long flags; 5100 5101 raw_spin_lock_irqsave(&rq->lock, flags); 5102 5103 if (rq->rd) { 5104 old_rd = rq->rd; 5105 5106 if (cpumask_test_cpu(rq->cpu, old_rd->online)) 5107 set_rq_offline(rq); 5108 5109 cpumask_clear_cpu(rq->cpu, old_rd->span); 5110 5111 /* 5112 * If we dont want to free the old_rt yet then 5113 * set old_rd to NULL to skip the freeing later 5114 * in this function: 5115 */ 5116 if (!atomic_dec_and_test(&old_rd->refcount)) 5117 old_rd = NULL; 5118 } 5119 5120 atomic_inc(&rd->refcount); 5121 rq->rd = rd; 5122 5123 cpumask_set_cpu(rq->cpu, rd->span); 5124 if (cpumask_test_cpu(rq->cpu, cpu_active_mask)) 5125 set_rq_online(rq); 5126 5127 raw_spin_unlock_irqrestore(&rq->lock, flags); 5128 5129 if (old_rd) 5130 call_rcu_sched(&old_rd->rcu, free_rootdomain); 5131} 5132 5133static int init_rootdomain(struct root_domain *rd) 5134{ 5135 memset(rd, 0, sizeof(*rd)); 5136 5137 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL)) 5138 goto out; 5139 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL)) 5140 goto free_span; 5141 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL)) 5142 goto free_online; 5143 5144 if (cpupri_init(&rd->cpupri) != 0) 5145 goto free_rto_mask; 5146 return 0; 5147 5148free_rto_mask: 5149 free_cpumask_var(rd->rto_mask); 5150free_online: 5151 free_cpumask_var(rd->online); 5152free_span: 5153 free_cpumask_var(rd->span); 5154out: 5155 return -ENOMEM; 5156} 5157 5158/* 5159 * By default the system creates a single root-domain with all cpus as 5160 * members (mimicking the global state we have today). 5161 */ 5162struct root_domain def_root_domain; 5163 5164static void init_defrootdomain(void) 5165{ 5166 init_rootdomain(&def_root_domain); 5167 5168 atomic_set(&def_root_domain.refcount, 1); 5169} 5170 5171static struct root_domain *alloc_rootdomain(void) 5172{ 5173 struct root_domain *rd; 5174 5175 rd = kmalloc(sizeof(*rd), GFP_KERNEL); 5176 if (!rd) 5177 return NULL; 5178 5179 if (init_rootdomain(rd) != 0) { 5180 kfree(rd); 5181 return NULL; 5182 } 5183 5184 return rd; 5185} 5186 5187static void free_sched_groups(struct sched_group *sg, int free_sgp) 5188{ 5189 struct sched_group *tmp, *first; 5190 5191 if (!sg) 5192 return; 5193 5194 first = sg; 5195 do { 5196 tmp = sg->next; 5197 5198 if (free_sgp && atomic_dec_and_test(&sg->sgp->ref)) 5199 kfree(sg->sgp); 5200 5201 kfree(sg); 5202 sg = tmp; 5203 } while (sg != first); 5204} 5205 5206static void free_sched_domain(struct rcu_head *rcu) 5207{ 5208 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu); 5209 5210 /* 5211 * If its an overlapping domain it has private groups, iterate and 5212 * nuke them all. 5213 */ 5214 if (sd->flags & SD_OVERLAP) { 5215 free_sched_groups(sd->groups, 1); 5216 } else if (atomic_dec_and_test(&sd->groups->ref)) { 5217 kfree(sd->groups->sgp); 5218 kfree(sd->groups); 5219 } 5220 kfree(sd); 5221} 5222 5223static void destroy_sched_domain(struct sched_domain *sd, int cpu) 5224{ 5225 call_rcu(&sd->rcu, free_sched_domain); 5226} 5227 5228static void destroy_sched_domains(struct sched_domain *sd, int cpu) 5229{ 5230 for (; sd; sd = sd->parent) 5231 destroy_sched_domain(sd, cpu); 5232} 5233 5234/* 5235 * Keep a special pointer to the highest sched_domain that has 5236 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this 5237 * allows us to avoid some pointer chasing select_idle_sibling(). 5238 * 5239 * Also keep a unique ID per domain (we use the first cpu number in 5240 * the cpumask of the domain), this allows us to quickly tell if 5241 * two cpus are in the same cache domain, see cpus_share_cache(). 5242 */ 5243DEFINE_PER_CPU(struct sched_domain *, sd_llc); 5244DEFINE_PER_CPU(int, sd_llc_size); 5245DEFINE_PER_CPU(int, sd_llc_id); 5246DEFINE_PER_CPU(struct sched_domain *, sd_numa); 5247 5248static void update_top_cache_domain(int cpu) 5249{ 5250 struct sched_domain *sd; 5251 int id = cpu; 5252 int size = 1; 5253 5254 sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES); 5255 if (sd) { 5256 id = cpumask_first(sched_domain_span(sd)); 5257 size = cpumask_weight(sched_domain_span(sd)); 5258 } 5259 5260 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd); 5261 per_cpu(sd_llc_size, cpu) = size; 5262 per_cpu(sd_llc_id, cpu) = id; 5263 5264 sd = lowest_flag_domain(cpu, SD_NUMA); 5265 rcu_assign_pointer(per_cpu(sd_numa, cpu), sd); 5266} 5267 5268/* 5269 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must 5270 * hold the hotplug lock. 5271 */ 5272static void 5273cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu) 5274{ 5275 struct rq *rq = cpu_rq(cpu); 5276 struct sched_domain *tmp; 5277 5278 /* Remove the sched domains which do not contribute to scheduling. */ 5279 for (tmp = sd; tmp; ) { 5280 struct sched_domain *parent = tmp->parent; 5281 if (!parent) 5282 break; 5283 5284 if (sd_parent_degenerate(tmp, parent)) { 5285 tmp->parent = parent->parent; 5286 if (parent->parent) 5287 parent->parent->child = tmp; 5288 /* 5289 * Transfer SD_PREFER_SIBLING down in case of a 5290 * degenerate parent; the spans match for this 5291 * so the property transfers. 5292 */ 5293 if (parent->flags & SD_PREFER_SIBLING) 5294 tmp->flags |= SD_PREFER_SIBLING; 5295 destroy_sched_domain(parent, cpu); 5296 } else 5297 tmp = tmp->parent; 5298 } 5299 5300 if (sd && sd_degenerate(sd)) { 5301 tmp = sd; 5302 sd = sd->parent; 5303 destroy_sched_domain(tmp, cpu); 5304 if (sd) 5305 sd->child = NULL; 5306 } 5307 5308 sched_domain_debug(sd, cpu); 5309 5310 rq_attach_root(rq, rd); 5311 tmp = rq->sd; 5312 rcu_assign_pointer(rq->sd, sd); 5313 destroy_sched_domains(tmp, cpu); 5314 5315 update_top_cache_domain(cpu); 5316} 5317 5318/* cpus with isolated domains */ 5319static cpumask_var_t cpu_isolated_map; 5320 5321/* Setup the mask of cpus configured for isolated domains */ 5322static int __init isolated_cpu_setup(char *str) 5323{ 5324 alloc_bootmem_cpumask_var(&cpu_isolated_map); 5325 cpulist_parse(str, cpu_isolated_map); 5326 return 1; 5327} 5328 5329__setup("isolcpus=", isolated_cpu_setup); 5330 5331static const struct cpumask *cpu_cpu_mask(int cpu) 5332{ 5333 return cpumask_of_node(cpu_to_node(cpu)); 5334} 5335 5336struct sd_data { 5337 struct sched_domain **__percpu sd; 5338 struct sched_group **__percpu sg; 5339 struct sched_group_power **__percpu sgp; 5340}; 5341 5342struct s_data { 5343 struct sched_domain ** __percpu sd; 5344 struct root_domain *rd; 5345}; 5346 5347enum s_alloc { 5348 sa_rootdomain, 5349 sa_sd, 5350 sa_sd_storage, 5351 sa_none, 5352}; 5353 5354struct sched_domain_topology_level; 5355 5356typedef struct sched_domain *(*sched_domain_init_f)(struct sched_domain_topology_level *tl, int cpu); 5357typedef const struct cpumask *(*sched_domain_mask_f)(int cpu); 5358 5359#define SDTL_OVERLAP 0x01 5360 5361struct sched_domain_topology_level { 5362 sched_domain_init_f init; 5363 sched_domain_mask_f mask; 5364 int flags; 5365 int numa_level; 5366 struct sd_data data; 5367}; 5368 5369/* 5370 * Build an iteration mask that can exclude certain CPUs from the upwards 5371 * domain traversal. 5372 * 5373 * Asymmetric node setups can result in situations where the domain tree is of 5374 * unequal depth, make sure to skip domains that already cover the entire 5375 * range. 5376 * 5377 * In that case build_sched_domains() will have terminated the iteration early 5378 * and our sibling sd spans will be empty. Domains should always include the 5379 * cpu they're built on, so check that. 5380 * 5381 */ 5382static void build_group_mask(struct sched_domain *sd, struct sched_group *sg) 5383{ 5384 const struct cpumask *span = sched_domain_span(sd); 5385 struct sd_data *sdd = sd->private; 5386 struct sched_domain *sibling; 5387 int i; 5388 5389 for_each_cpu(i, span) { 5390 sibling = *per_cpu_ptr(sdd->sd, i); 5391 if (!cpumask_test_cpu(i, sched_domain_span(sibling))) 5392 continue; 5393 5394 cpumask_set_cpu(i, sched_group_mask(sg)); 5395 } 5396} 5397 5398/* 5399 * Return the canonical balance cpu for this group, this is the first cpu 5400 * of this group that's also in the iteration mask. 5401 */ 5402int group_balance_cpu(struct sched_group *sg) 5403{ 5404 return cpumask_first_and(sched_group_cpus(sg), sched_group_mask(sg)); 5405} 5406 5407static int 5408build_overlap_sched_groups(struct sched_domain *sd, int cpu) 5409{ 5410 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg; 5411 const struct cpumask *span = sched_domain_span(sd); 5412 struct cpumask *covered = sched_domains_tmpmask; 5413 struct sd_data *sdd = sd->private; 5414 struct sched_domain *child; 5415 int i; 5416 5417 cpumask_clear(covered); 5418 5419 for_each_cpu(i, span) { 5420 struct cpumask *sg_span; 5421 5422 if (cpumask_test_cpu(i, covered)) 5423 continue; 5424 5425 child = *per_cpu_ptr(sdd->sd, i); 5426 5427 /* See the comment near build_group_mask(). */ 5428 if (!cpumask_test_cpu(i, sched_domain_span(child))) 5429 continue; 5430 5431 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(), 5432 GFP_KERNEL, cpu_to_node(cpu)); 5433 5434 if (!sg) 5435 goto fail; 5436 5437 sg_span = sched_group_cpus(sg); 5438 if (child->child) { 5439 child = child->child; 5440 cpumask_copy(sg_span, sched_domain_span(child)); 5441 } else 5442 cpumask_set_cpu(i, sg_span); 5443 5444 cpumask_or(covered, covered, sg_span); 5445 5446 sg->sgp = *per_cpu_ptr(sdd->sgp, i); 5447 if (atomic_inc_return(&sg->sgp->ref) == 1) 5448 build_group_mask(sd, sg); 5449 5450 /* 5451 * Initialize sgp->power such that even if we mess up the 5452 * domains and no possible iteration will get us here, we won't 5453 * die on a /0 trap. 5454 */ 5455 sg->sgp->power = SCHED_POWER_SCALE * cpumask_weight(sg_span); 5456 5457 /* 5458 * Make sure the first group of this domain contains the 5459 * canonical balance cpu. Otherwise the sched_domain iteration 5460 * breaks. See update_sg_lb_stats(). 5461 */ 5462 if ((!groups && cpumask_test_cpu(cpu, sg_span)) || 5463 group_balance_cpu(sg) == cpu) 5464 groups = sg; 5465 5466 if (!first) 5467 first = sg; 5468 if (last) 5469 last->next = sg; 5470 last = sg; 5471 last->next = first; 5472 } 5473 sd->groups = groups; 5474 5475 return 0; 5476 5477fail: 5478 free_sched_groups(first, 0); 5479 5480 return -ENOMEM; 5481} 5482 5483static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg) 5484{ 5485 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu); 5486 struct sched_domain *child = sd->child; 5487 5488 if (child) 5489 cpu = cpumask_first(sched_domain_span(child)); 5490 5491 if (sg) { 5492 *sg = *per_cpu_ptr(sdd->sg, cpu); 5493 (*sg)->sgp = *per_cpu_ptr(sdd->sgp, cpu); 5494 atomic_set(&(*sg)->sgp->ref, 1); /* for claim_allocations */ 5495 } 5496 5497 return cpu; 5498} 5499 5500/* 5501 * build_sched_groups will build a circular linked list of the groups 5502 * covered by the given span, and will set each group's ->cpumask correctly, 5503 * and ->cpu_power to 0. 5504 * 5505 * Assumes the sched_domain tree is fully constructed 5506 */ 5507static int 5508build_sched_groups(struct sched_domain *sd, int cpu) 5509{ 5510 struct sched_group *first = NULL, *last = NULL; 5511 struct sd_data *sdd = sd->private; 5512 const struct cpumask *span = sched_domain_span(sd); 5513 struct cpumask *covered; 5514 int i; 5515 5516 get_group(cpu, sdd, &sd->groups); 5517 atomic_inc(&sd->groups->ref); 5518 5519 if (cpu != cpumask_first(span)) 5520 return 0; 5521 5522 lockdep_assert_held(&sched_domains_mutex); 5523 covered = sched_domains_tmpmask; 5524 5525 cpumask_clear(covered); 5526 5527 for_each_cpu(i, span) { 5528 struct sched_group *sg; 5529 int group, j; 5530 5531 if (cpumask_test_cpu(i, covered)) 5532 continue; 5533 5534 group = get_group(i, sdd, &sg); 5535 cpumask_clear(sched_group_cpus(sg)); 5536 sg->sgp->power = 0; 5537 cpumask_setall(sched_group_mask(sg)); 5538 5539 for_each_cpu(j, span) { 5540 if (get_group(j, sdd, NULL) != group) 5541 continue; 5542 5543 cpumask_set_cpu(j, covered); 5544 cpumask_set_cpu(j, sched_group_cpus(sg)); 5545 } 5546 5547 if (!first) 5548 first = sg; 5549 if (last) 5550 last->next = sg; 5551 last = sg; 5552 } 5553 last->next = first; 5554 5555 return 0; 5556} 5557 5558/* 5559 * Initialize sched groups cpu_power. 5560 * 5561 * cpu_power indicates the capacity of sched group, which is used while 5562 * distributing the load between different sched groups in a sched domain. 5563 * Typically cpu_power for all the groups in a sched domain will be same unless 5564 * there are asymmetries in the topology. If there are asymmetries, group 5565 * having more cpu_power will pickup more load compared to the group having 5566 * less cpu_power. 5567 */ 5568static void init_sched_groups_power(int cpu, struct sched_domain *sd) 5569{ 5570 struct sched_group *sg = sd->groups; 5571 5572 WARN_ON(!sg); 5573 5574 do { 5575 sg->group_weight = cpumask_weight(sched_group_cpus(sg)); 5576 sg = sg->next; 5577 } while (sg != sd->groups); 5578 5579 if (cpu != group_balance_cpu(sg)) 5580 return; 5581 5582 update_group_power(sd, cpu); 5583 atomic_set(&sg->sgp->nr_busy_cpus, sg->group_weight); 5584} 5585 5586int __weak arch_sd_sibling_asym_packing(void) 5587{ 5588 return 0*SD_ASYM_PACKING; 5589} 5590 5591/* 5592 * Initializers for schedule domains 5593 * Non-inlined to reduce accumulated stack pressure in build_sched_domains() 5594 */ 5595 5596#ifdef CONFIG_SCHED_DEBUG 5597# define SD_INIT_NAME(sd, type) sd->name = #type 5598#else 5599# define SD_INIT_NAME(sd, type) do { } while (0) 5600#endif 5601 5602#define SD_INIT_FUNC(type) \ 5603static noinline struct sched_domain * \ 5604sd_init_##type(struct sched_domain_topology_level *tl, int cpu) \ 5605{ \ 5606 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); \ 5607 *sd = SD_##type##_INIT; \ 5608 SD_INIT_NAME(sd, type); \ 5609 sd->private = &tl->data; \ 5610 return sd; \ 5611} 5612 5613SD_INIT_FUNC(CPU) 5614#ifdef CONFIG_SCHED_SMT 5615 SD_INIT_FUNC(SIBLING) 5616#endif 5617#ifdef CONFIG_SCHED_MC 5618 SD_INIT_FUNC(MC) 5619#endif 5620#ifdef CONFIG_SCHED_BOOK 5621 SD_INIT_FUNC(BOOK) 5622#endif 5623 5624static int default_relax_domain_level = -1; 5625int sched_domain_level_max; 5626 5627static int __init setup_relax_domain_level(char *str) 5628{ 5629 if (kstrtoint(str, 0, &default_relax_domain_level)) 5630 pr_warn("Unable to set relax_domain_level\n"); 5631 5632 return 1; 5633} 5634__setup("relax_domain_level=", setup_relax_domain_level); 5635 5636static void set_domain_attribute(struct sched_domain *sd, 5637 struct sched_domain_attr *attr) 5638{ 5639 int request; 5640 5641 if (!attr || attr->relax_domain_level < 0) { 5642 if (default_relax_domain_level < 0) 5643 return; 5644 else 5645 request = default_relax_domain_level; 5646 } else 5647 request = attr->relax_domain_level; 5648 if (request < sd->level) { 5649 /* turn off idle balance on this domain */ 5650 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE); 5651 } else { 5652 /* turn on idle balance on this domain */ 5653 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE); 5654 } 5655} 5656 5657static void __sdt_free(const struct cpumask *cpu_map); 5658static int __sdt_alloc(const struct cpumask *cpu_map); 5659 5660static void __free_domain_allocs(struct s_data *d, enum s_alloc what, 5661 const struct cpumask *cpu_map) 5662{ 5663 switch (what) { 5664 case sa_rootdomain: 5665 if (!atomic_read(&d->rd->refcount)) 5666 free_rootdomain(&d->rd->rcu); /* fall through */ 5667 case sa_sd: 5668 free_percpu(d->sd); /* fall through */ 5669 case sa_sd_storage: 5670 __sdt_free(cpu_map); /* fall through */ 5671 case sa_none: 5672 break; 5673 } 5674} 5675 5676static enum s_alloc __visit_domain_allocation_hell(struct s_data *d, 5677 const struct cpumask *cpu_map) 5678{ 5679 memset(d, 0, sizeof(*d)); 5680 5681 if (__sdt_alloc(cpu_map)) 5682 return sa_sd_storage; 5683 d->sd = alloc_percpu(struct sched_domain *); 5684 if (!d->sd) 5685 return sa_sd_storage; 5686 d->rd = alloc_rootdomain(); 5687 if (!d->rd) 5688 return sa_sd; 5689 return sa_rootdomain; 5690} 5691 5692/* 5693 * NULL the sd_data elements we've used to build the sched_domain and 5694 * sched_group structure so that the subsequent __free_domain_allocs() 5695 * will not free the data we're using. 5696 */ 5697static void claim_allocations(int cpu, struct sched_domain *sd) 5698{ 5699 struct sd_data *sdd = sd->private; 5700 5701 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd); 5702 *per_cpu_ptr(sdd->sd, cpu) = NULL; 5703 5704 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref)) 5705 *per_cpu_ptr(sdd->sg, cpu) = NULL; 5706 5707 if (atomic_read(&(*per_cpu_ptr(sdd->sgp, cpu))->ref)) 5708 *per_cpu_ptr(sdd->sgp, cpu) = NULL; 5709} 5710 5711#ifdef CONFIG_SCHED_SMT 5712static const struct cpumask *cpu_smt_mask(int cpu) 5713{ 5714 return topology_thread_cpumask(cpu); 5715} 5716#endif 5717 5718/* 5719 * Topology list, bottom-up. 5720 */ 5721static struct sched_domain_topology_level default_topology[] = { 5722#ifdef CONFIG_SCHED_SMT 5723 { sd_init_SIBLING, cpu_smt_mask, }, 5724#endif 5725#ifdef CONFIG_SCHED_MC 5726 { sd_init_MC, cpu_coregroup_mask, }, 5727#endif 5728#ifdef CONFIG_SCHED_BOOK 5729 { sd_init_BOOK, cpu_book_mask, }, 5730#endif 5731 { sd_init_CPU, cpu_cpu_mask, }, 5732 { NULL, }, 5733}; 5734 5735static struct sched_domain_topology_level *sched_domain_topology = default_topology; 5736 5737#define for_each_sd_topology(tl) \ 5738 for (tl = sched_domain_topology; tl->init; tl++) 5739 5740#ifdef CONFIG_NUMA 5741 5742static int sched_domains_numa_levels; 5743static int *sched_domains_numa_distance; 5744static struct cpumask ***sched_domains_numa_masks; 5745static int sched_domains_curr_level; 5746 5747static inline int sd_local_flags(int level) 5748{ 5749 if (sched_domains_numa_distance[level] > RECLAIM_DISTANCE) 5750 return 0; 5751 5752 return SD_BALANCE_EXEC | SD_BALANCE_FORK | SD_WAKE_AFFINE; 5753} 5754 5755static struct sched_domain * 5756sd_numa_init(struct sched_domain_topology_level *tl, int cpu) 5757{ 5758 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); 5759 int level = tl->numa_level; 5760 int sd_weight = cpumask_weight( 5761 sched_domains_numa_masks[level][cpu_to_node(cpu)]); 5762 5763 *sd = (struct sched_domain){ 5764 .min_interval = sd_weight, 5765 .max_interval = 2*sd_weight, 5766 .busy_factor = 32, 5767 .imbalance_pct = 125, 5768 .cache_nice_tries = 2, 5769 .busy_idx = 3, 5770 .idle_idx = 2, 5771 .newidle_idx = 0, 5772 .wake_idx = 0, 5773 .forkexec_idx = 0, 5774 5775 .flags = 1*SD_LOAD_BALANCE 5776 | 1*SD_BALANCE_NEWIDLE 5777 | 0*SD_BALANCE_EXEC 5778 | 0*SD_BALANCE_FORK 5779 | 0*SD_BALANCE_WAKE 5780 | 0*SD_WAKE_AFFINE 5781 | 0*SD_SHARE_CPUPOWER 5782 | 0*SD_SHARE_PKG_RESOURCES 5783 | 1*SD_SERIALIZE 5784 | 0*SD_PREFER_SIBLING 5785 | 1*SD_NUMA 5786 | sd_local_flags(level) 5787 , 5788 .last_balance = jiffies, 5789 .balance_interval = sd_weight, 5790 }; 5791 SD_INIT_NAME(sd, NUMA); 5792 sd->private = &tl->data; 5793 5794 /* 5795 * Ugly hack to pass state to sd_numa_mask()... 5796 */ 5797 sched_domains_curr_level = tl->numa_level; 5798 5799 return sd; 5800} 5801 5802static const struct cpumask *sd_numa_mask(int cpu) 5803{ 5804 return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)]; 5805} 5806 5807static void sched_numa_warn(const char *str) 5808{ 5809 static int done = false; 5810 int i,j; 5811 5812 if (done) 5813 return; 5814 5815 done = true; 5816 5817 printk(KERN_WARNING "ERROR: %s\n\n", str); 5818 5819 for (i = 0; i < nr_node_ids; i++) { 5820 printk(KERN_WARNING " "); 5821 for (j = 0; j < nr_node_ids; j++) 5822 printk(KERN_CONT "%02d ", node_distance(i,j)); 5823 printk(KERN_CONT "\n"); 5824 } 5825 printk(KERN_WARNING "\n"); 5826} 5827 5828static bool find_numa_distance(int distance) 5829{ 5830 int i; 5831 5832 if (distance == node_distance(0, 0)) 5833 return true; 5834 5835 for (i = 0; i < sched_domains_numa_levels; i++) { 5836 if (sched_domains_numa_distance[i] == distance) 5837 return true; 5838 } 5839 5840 return false; 5841} 5842 5843static void sched_init_numa(void) 5844{ 5845 int next_distance, curr_distance = node_distance(0, 0); 5846 struct sched_domain_topology_level *tl; 5847 int level = 0; 5848 int i, j, k; 5849 5850 sched_domains_numa_distance = kzalloc(sizeof(int) * nr_node_ids, GFP_KERNEL); 5851 if (!sched_domains_numa_distance) 5852 return; 5853 5854 /* 5855 * O(nr_nodes^2) deduplicating selection sort -- in order to find the 5856 * unique distances in the node_distance() table. 5857 * 5858 * Assumes node_distance(0,j) includes all distances in 5859 * node_distance(i,j) in order to avoid cubic time. 5860 */ 5861 next_distance = curr_distance; 5862 for (i = 0; i < nr_node_ids; i++) { 5863 for (j = 0; j < nr_node_ids; j++) { 5864 for (k = 0; k < nr_node_ids; k++) { 5865 int distance = node_distance(i, k); 5866 5867 if (distance > curr_distance && 5868 (distance < next_distance || 5869 next_distance == curr_distance)) 5870 next_distance = distance; 5871 5872 /* 5873 * While not a strong assumption it would be nice to know 5874 * about cases where if node A is connected to B, B is not 5875 * equally connected to A. 5876 */ 5877 if (sched_debug() && node_distance(k, i) != distance) 5878 sched_numa_warn("Node-distance not symmetric"); 5879 5880 if (sched_debug() && i && !find_numa_distance(distance)) 5881 sched_numa_warn("Node-0 not representative"); 5882 } 5883 if (next_distance != curr_distance) { 5884 sched_domains_numa_distance[level++] = next_distance; 5885 sched_domains_numa_levels = level; 5886 curr_distance = next_distance; 5887 } else break; 5888 } 5889 5890 /* 5891 * In case of sched_debug() we verify the above assumption. 5892 */ 5893 if (!sched_debug()) 5894 break; 5895 } 5896 /* 5897 * 'level' contains the number of unique distances, excluding the 5898 * identity distance node_distance(i,i). 5899 * 5900 * The sched_domains_numa_distance[] array includes the actual distance 5901 * numbers. 5902 */ 5903 5904 /* 5905 * Here, we should temporarily reset sched_domains_numa_levels to 0. 5906 * If it fails to allocate memory for array sched_domains_numa_masks[][], 5907 * the array will contain less then 'level' members. This could be 5908 * dangerous when we use it to iterate array sched_domains_numa_masks[][] 5909 * in other functions. 5910 * 5911 * We reset it to 'level' at the end of this function. 5912 */ 5913 sched_domains_numa_levels = 0; 5914 5915 sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL); 5916 if (!sched_domains_numa_masks) 5917 return; 5918 5919 /* 5920 * Now for each level, construct a mask per node which contains all 5921 * cpus of nodes that are that many hops away from us. 5922 */ 5923 for (i = 0; i < level; i++) { 5924 sched_domains_numa_masks[i] = 5925 kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL); 5926 if (!sched_domains_numa_masks[i]) 5927 return; 5928 5929 for (j = 0; j < nr_node_ids; j++) { 5930 struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL); 5931 if (!mask) 5932 return; 5933 5934 sched_domains_numa_masks[i][j] = mask; 5935 5936 for (k = 0; k < nr_node_ids; k++) { 5937 if (node_distance(j, k) > sched_domains_numa_distance[i]) 5938 continue; 5939 5940 cpumask_or(mask, mask, cpumask_of_node(k)); 5941 } 5942 } 5943 } 5944 5945 tl = kzalloc((ARRAY_SIZE(default_topology) + level) * 5946 sizeof(struct sched_domain_topology_level), GFP_KERNEL); 5947 if (!tl) 5948 return; 5949 5950 /* 5951 * Copy the default topology bits.. 5952 */ 5953 for (i = 0; default_topology[i].init; i++) 5954 tl[i] = default_topology[i]; 5955 5956 /* 5957 * .. and append 'j' levels of NUMA goodness. 5958 */ 5959 for (j = 0; j < level; i++, j++) { 5960 tl[i] = (struct sched_domain_topology_level){ 5961 .init = sd_numa_init, 5962 .mask = sd_numa_mask, 5963 .flags = SDTL_OVERLAP, 5964 .numa_level = j, 5965 }; 5966 } 5967 5968 sched_domain_topology = tl; 5969 5970 sched_domains_numa_levels = level; 5971} 5972 5973static void sched_domains_numa_masks_set(int cpu) 5974{ 5975 int i, j; 5976 int node = cpu_to_node(cpu); 5977 5978 for (i = 0; i < sched_domains_numa_levels; i++) { 5979 for (j = 0; j < nr_node_ids; j++) { 5980 if (node_distance(j, node) <= sched_domains_numa_distance[i]) 5981 cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]); 5982 } 5983 } 5984} 5985 5986static void sched_domains_numa_masks_clear(int cpu) 5987{ 5988 int i, j; 5989 for (i = 0; i < sched_domains_numa_levels; i++) { 5990 for (j = 0; j < nr_node_ids; j++) 5991 cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]); 5992 } 5993} 5994 5995/* 5996 * Update sched_domains_numa_masks[level][node] array when new cpus 5997 * are onlined. 5998 */ 5999static int sched_domains_numa_masks_update(struct notifier_block *nfb, 6000 unsigned long action, 6001 void *hcpu) 6002{ 6003 int cpu = (long)hcpu; 6004 6005 switch (action & ~CPU_TASKS_FROZEN) { 6006 case CPU_ONLINE: 6007 sched_domains_numa_masks_set(cpu); 6008 break; 6009 6010 case CPU_DEAD: 6011 sched_domains_numa_masks_clear(cpu); 6012 break; 6013 6014 default: 6015 return NOTIFY_DONE; 6016 } 6017 6018 return NOTIFY_OK; 6019} 6020#else 6021static inline void sched_init_numa(void) 6022{ 6023} 6024 6025static int sched_domains_numa_masks_update(struct notifier_block *nfb, 6026 unsigned long action, 6027 void *hcpu) 6028{ 6029 return 0; 6030} 6031#endif /* CONFIG_NUMA */ 6032 6033static int __sdt_alloc(const struct cpumask *cpu_map) 6034{ 6035 struct sched_domain_topology_level *tl; 6036 int j; 6037 6038 for_each_sd_topology(tl) { 6039 struct sd_data *sdd = &tl->data; 6040 6041 sdd->sd = alloc_percpu(struct sched_domain *); 6042 if (!sdd->sd) 6043 return -ENOMEM; 6044 6045 sdd->sg = alloc_percpu(struct sched_group *); 6046 if (!sdd->sg) 6047 return -ENOMEM; 6048 6049 sdd->sgp = alloc_percpu(struct sched_group_power *); 6050 if (!sdd->sgp) 6051 return -ENOMEM; 6052 6053 for_each_cpu(j, cpu_map) { 6054 struct sched_domain *sd; 6055 struct sched_group *sg; 6056 struct sched_group_power *sgp; 6057 6058 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(), 6059 GFP_KERNEL, cpu_to_node(j)); 6060 if (!sd) 6061 return -ENOMEM; 6062 6063 *per_cpu_ptr(sdd->sd, j) = sd; 6064 6065 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(), 6066 GFP_KERNEL, cpu_to_node(j)); 6067 if (!sg) 6068 return -ENOMEM; 6069 6070 sg->next = sg; 6071 6072 *per_cpu_ptr(sdd->sg, j) = sg; 6073 6074 sgp = kzalloc_node(sizeof(struct sched_group_power) + cpumask_size(), 6075 GFP_KERNEL, cpu_to_node(j)); 6076 if (!sgp) 6077 return -ENOMEM; 6078 6079 *per_cpu_ptr(sdd->sgp, j) = sgp; 6080 } 6081 } 6082 6083 return 0; 6084} 6085 6086static void __sdt_free(const struct cpumask *cpu_map) 6087{ 6088 struct sched_domain_topology_level *tl; 6089 int j; 6090 6091 for_each_sd_topology(tl) { 6092 struct sd_data *sdd = &tl->data; 6093 6094 for_each_cpu(j, cpu_map) { 6095 struct sched_domain *sd; 6096 6097 if (sdd->sd) { 6098 sd = *per_cpu_ptr(sdd->sd, j); 6099 if (sd && (sd->flags & SD_OVERLAP)) 6100 free_sched_groups(sd->groups, 0); 6101 kfree(*per_cpu_ptr(sdd->sd, j)); 6102 } 6103 6104 if (sdd->sg) 6105 kfree(*per_cpu_ptr(sdd->sg, j)); 6106 if (sdd->sgp) 6107 kfree(*per_cpu_ptr(sdd->sgp, j)); 6108 } 6109 free_percpu(sdd->sd); 6110 sdd->sd = NULL; 6111 free_percpu(sdd->sg); 6112 sdd->sg = NULL; 6113 free_percpu(sdd->sgp); 6114 sdd->sgp = NULL; 6115 } 6116} 6117 6118struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl, 6119 const struct cpumask *cpu_map, struct sched_domain_attr *attr, 6120 struct sched_domain *child, int cpu) 6121{ 6122 struct sched_domain *sd = tl->init(tl, cpu); 6123 if (!sd) 6124 return child; 6125 6126 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu)); 6127 if (child) { 6128 sd->level = child->level + 1; 6129 sched_domain_level_max = max(sched_domain_level_max, sd->level); 6130 child->parent = sd; 6131 sd->child = child; 6132 } 6133 set_domain_attribute(sd, attr); 6134 6135 return sd; 6136} 6137 6138/* 6139 * Build sched domains for a given set of cpus and attach the sched domains 6140 * to the individual cpus 6141 */ 6142static int build_sched_domains(const struct cpumask *cpu_map, 6143 struct sched_domain_attr *attr) 6144{ 6145 enum s_alloc alloc_state; 6146 struct sched_domain *sd; 6147 struct s_data d; 6148 int i, ret = -ENOMEM; 6149 6150 alloc_state = __visit_domain_allocation_hell(&d, cpu_map); 6151 if (alloc_state != sa_rootdomain) 6152 goto error; 6153 6154 /* Set up domains for cpus specified by the cpu_map. */ 6155 for_each_cpu(i, cpu_map) { 6156 struct sched_domain_topology_level *tl; 6157 6158 sd = NULL; 6159 for_each_sd_topology(tl) { 6160 sd = build_sched_domain(tl, cpu_map, attr, sd, i); 6161 if (tl == sched_domain_topology) 6162 *per_cpu_ptr(d.sd, i) = sd; 6163 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP)) 6164 sd->flags |= SD_OVERLAP; 6165 if (cpumask_equal(cpu_map, sched_domain_span(sd))) 6166 break; 6167 } 6168 } 6169 6170 /* Build the groups for the domains */ 6171 for_each_cpu(i, cpu_map) { 6172 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) { 6173 sd->span_weight = cpumask_weight(sched_domain_span(sd)); 6174 if (sd->flags & SD_OVERLAP) { 6175 if (build_overlap_sched_groups(sd, i)) 6176 goto error; 6177 } else { 6178 if (build_sched_groups(sd, i)) 6179 goto error; 6180 } 6181 } 6182 } 6183 6184 /* Calculate CPU power for physical packages and nodes */ 6185 for (i = nr_cpumask_bits-1; i >= 0; i--) { 6186 if (!cpumask_test_cpu(i, cpu_map)) 6187 continue; 6188 6189 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) { 6190 claim_allocations(i, sd); 6191 init_sched_groups_power(i, sd); 6192 } 6193 } 6194 6195 /* Attach the domains */ 6196 rcu_read_lock(); 6197 for_each_cpu(i, cpu_map) { 6198 sd = *per_cpu_ptr(d.sd, i); 6199 cpu_attach_domain(sd, d.rd, i); 6200 } 6201 rcu_read_unlock(); 6202 6203 ret = 0; 6204error: 6205 __free_domain_allocs(&d, alloc_state, cpu_map); 6206 return ret; 6207} 6208 6209static cpumask_var_t *doms_cur; /* current sched domains */ 6210static int ndoms_cur; /* number of sched domains in 'doms_cur' */ 6211static struct sched_domain_attr *dattr_cur; 6212 /* attribues of custom domains in 'doms_cur' */ 6213 6214/* 6215 * Special case: If a kmalloc of a doms_cur partition (array of 6216 * cpumask) fails, then fallback to a single sched domain, 6217 * as determined by the single cpumask fallback_doms. 6218 */ 6219static cpumask_var_t fallback_doms; 6220 6221/* 6222 * arch_update_cpu_topology lets virtualized architectures update the 6223 * cpu core maps. It is supposed to return 1 if the topology changed 6224 * or 0 if it stayed the same. 6225 */ 6226int __attribute__((weak)) arch_update_cpu_topology(void) 6227{ 6228 return 0; 6229} 6230 6231cpumask_var_t *alloc_sched_domains(unsigned int ndoms) 6232{ 6233 int i; 6234 cpumask_var_t *doms; 6235 6236 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL); 6237 if (!doms) 6238 return NULL; 6239 for (i = 0; i < ndoms; i++) { 6240 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) { 6241 free_sched_domains(doms, i); 6242 return NULL; 6243 } 6244 } 6245 return doms; 6246} 6247 6248void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms) 6249{ 6250 unsigned int i; 6251 for (i = 0; i < ndoms; i++) 6252 free_cpumask_var(doms[i]); 6253 kfree(doms); 6254} 6255 6256/* 6257 * Set up scheduler domains and groups. Callers must hold the hotplug lock. 6258 * For now this just excludes isolated cpus, but could be used to 6259 * exclude other special cases in the future. 6260 */ 6261static int init_sched_domains(const struct cpumask *cpu_map) 6262{ 6263 int err; 6264 6265 arch_update_cpu_topology(); 6266 ndoms_cur = 1; 6267 doms_cur = alloc_sched_domains(ndoms_cur); 6268 if (!doms_cur) 6269 doms_cur = &fallback_doms; 6270 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map); 6271 err = build_sched_domains(doms_cur[0], NULL); 6272 register_sched_domain_sysctl(); 6273 6274 return err; 6275} 6276 6277/* 6278 * Detach sched domains from a group of cpus specified in cpu_map 6279 * These cpus will now be attached to the NULL domain 6280 */ 6281static void detach_destroy_domains(const struct cpumask *cpu_map) 6282{ 6283 int i; 6284 6285 rcu_read_lock(); 6286 for_each_cpu(i, cpu_map) 6287 cpu_attach_domain(NULL, &def_root_domain, i); 6288 rcu_read_unlock(); 6289} 6290 6291/* handle null as "default" */ 6292static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur, 6293 struct sched_domain_attr *new, int idx_new) 6294{ 6295 struct sched_domain_attr tmp; 6296 6297 /* fast path */ 6298 if (!new && !cur) 6299 return 1; 6300 6301 tmp = SD_ATTR_INIT; 6302 return !memcmp(cur ? (cur + idx_cur) : &tmp, 6303 new ? (new + idx_new) : &tmp, 6304 sizeof(struct sched_domain_attr)); 6305} 6306 6307/* 6308 * Partition sched domains as specified by the 'ndoms_new' 6309 * cpumasks in the array doms_new[] of cpumasks. This compares 6310 * doms_new[] to the current sched domain partitioning, doms_cur[]. 6311 * It destroys each deleted domain and builds each new domain. 6312 * 6313 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'. 6314 * The masks don't intersect (don't overlap.) We should setup one 6315 * sched domain for each mask. CPUs not in any of the cpumasks will 6316 * not be load balanced. If the same cpumask appears both in the 6317 * current 'doms_cur' domains and in the new 'doms_new', we can leave 6318 * it as it is. 6319 * 6320 * The passed in 'doms_new' should be allocated using 6321 * alloc_sched_domains. This routine takes ownership of it and will 6322 * free_sched_domains it when done with it. If the caller failed the 6323 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1, 6324 * and partition_sched_domains() will fallback to the single partition 6325 * 'fallback_doms', it also forces the domains to be rebuilt. 6326 * 6327 * If doms_new == NULL it will be replaced with cpu_online_mask. 6328 * ndoms_new == 0 is a special case for destroying existing domains, 6329 * and it will not create the default domain. 6330 * 6331 * Call with hotplug lock held 6332 */ 6333void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[], 6334 struct sched_domain_attr *dattr_new) 6335{ 6336 int i, j, n; 6337 int new_topology; 6338 6339 mutex_lock(&sched_domains_mutex); 6340 6341 /* always unregister in case we don't destroy any domains */ 6342 unregister_sched_domain_sysctl(); 6343 6344 /* Let architecture update cpu core mappings. */ 6345 new_topology = arch_update_cpu_topology(); 6346 6347 n = doms_new ? ndoms_new : 0; 6348 6349 /* Destroy deleted domains */ 6350 for (i = 0; i < ndoms_cur; i++) { 6351 for (j = 0; j < n && !new_topology; j++) { 6352 if (cpumask_equal(doms_cur[i], doms_new[j]) 6353 && dattrs_equal(dattr_cur, i, dattr_new, j)) 6354 goto match1; 6355 } 6356 /* no match - a current sched domain not in new doms_new[] */ 6357 detach_destroy_domains(doms_cur[i]); 6358match1: 6359 ; 6360 } 6361 6362 n = ndoms_cur; 6363 if (doms_new == NULL) { 6364 n = 0; 6365 doms_new = &fallback_doms; 6366 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map); 6367 WARN_ON_ONCE(dattr_new); 6368 } 6369 6370 /* Build new domains */ 6371 for (i = 0; i < ndoms_new; i++) { 6372 for (j = 0; j < n && !new_topology; j++) { 6373 if (cpumask_equal(doms_new[i], doms_cur[j]) 6374 && dattrs_equal(dattr_new, i, dattr_cur, j)) 6375 goto match2; 6376 } 6377 /* no match - add a new doms_new */ 6378 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL); 6379match2: 6380 ; 6381 } 6382 6383 /* Remember the new sched domains */ 6384 if (doms_cur != &fallback_doms) 6385 free_sched_domains(doms_cur, ndoms_cur); 6386 kfree(dattr_cur); /* kfree(NULL) is safe */ 6387 doms_cur = doms_new; 6388 dattr_cur = dattr_new; 6389 ndoms_cur = ndoms_new; 6390 6391 register_sched_domain_sysctl(); 6392 6393 mutex_unlock(&sched_domains_mutex); 6394} 6395 6396static int num_cpus_frozen; /* used to mark begin/end of suspend/resume */ 6397 6398/* 6399 * Update cpusets according to cpu_active mask. If cpusets are 6400 * disabled, cpuset_update_active_cpus() becomes a simple wrapper 6401 * around partition_sched_domains(). 6402 * 6403 * If we come here as part of a suspend/resume, don't touch cpusets because we 6404 * want to restore it back to its original state upon resume anyway. 6405 */ 6406static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action, 6407 void *hcpu) 6408{ 6409 switch (action) { 6410 case CPU_ONLINE_FROZEN: 6411 case CPU_DOWN_FAILED_FROZEN: 6412 6413 /* 6414 * num_cpus_frozen tracks how many CPUs are involved in suspend 6415 * resume sequence. As long as this is not the last online 6416 * operation in the resume sequence, just build a single sched 6417 * domain, ignoring cpusets. 6418 */ 6419 num_cpus_frozen--; 6420 if (likely(num_cpus_frozen)) { 6421 partition_sched_domains(1, NULL, NULL); 6422 break; 6423 } 6424 6425 /* 6426 * This is the last CPU online operation. So fall through and 6427 * restore the original sched domains by considering the 6428 * cpuset configurations. 6429 */ 6430 6431 case CPU_ONLINE: 6432 case CPU_DOWN_FAILED: 6433 cpuset_update_active_cpus(true); 6434 break; 6435 default: 6436 return NOTIFY_DONE; 6437 } 6438 return NOTIFY_OK; 6439} 6440 6441static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action, 6442 void *hcpu) 6443{ 6444 switch (action) { 6445 case CPU_DOWN_PREPARE: 6446 cpuset_update_active_cpus(false); 6447 break; 6448 case CPU_DOWN_PREPARE_FROZEN: 6449 num_cpus_frozen++; 6450 partition_sched_domains(1, NULL, NULL); 6451 break; 6452 default: 6453 return NOTIFY_DONE; 6454 } 6455 return NOTIFY_OK; 6456} 6457 6458void __init sched_init_smp(void) 6459{ 6460 cpumask_var_t non_isolated_cpus; 6461 6462 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL); 6463 alloc_cpumask_var(&fallback_doms, GFP_KERNEL); 6464 6465 sched_init_numa(); 6466 6467 get_online_cpus(); 6468 mutex_lock(&sched_domains_mutex); 6469 init_sched_domains(cpu_active_mask); 6470 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map); 6471 if (cpumask_empty(non_isolated_cpus)) 6472 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus); 6473 mutex_unlock(&sched_domains_mutex); 6474 put_online_cpus(); 6475 6476 hotcpu_notifier(sched_domains_numa_masks_update, CPU_PRI_SCHED_ACTIVE); 6477 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE); 6478 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE); 6479 6480 init_hrtick(); 6481 6482 /* Move init over to a non-isolated CPU */ 6483 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0) 6484 BUG(); 6485 sched_init_granularity(); 6486 free_cpumask_var(non_isolated_cpus); 6487 6488 init_sched_rt_class(); 6489} 6490#else 6491void __init sched_init_smp(void) 6492{ 6493 sched_init_granularity(); 6494} 6495#endif /* CONFIG_SMP */ 6496 6497const_debug unsigned int sysctl_timer_migration = 1; 6498 6499int in_sched_functions(unsigned long addr) 6500{ 6501 return in_lock_functions(addr) || 6502 (addr >= (unsigned long)__sched_text_start 6503 && addr < (unsigned long)__sched_text_end); 6504} 6505 6506#ifdef CONFIG_CGROUP_SCHED 6507/* 6508 * Default task group. 6509 * Every task in system belongs to this group at bootup. 6510 */ 6511struct task_group root_task_group; 6512LIST_HEAD(task_groups); 6513#endif 6514 6515DECLARE_PER_CPU(cpumask_var_t, load_balance_mask); 6516 6517void __init sched_init(void) 6518{ 6519 int i, j; 6520 unsigned long alloc_size = 0, ptr; 6521 6522#ifdef CONFIG_FAIR_GROUP_SCHED 6523 alloc_size += 2 * nr_cpu_ids * sizeof(void **); 6524#endif 6525#ifdef CONFIG_RT_GROUP_SCHED 6526 alloc_size += 2 * nr_cpu_ids * sizeof(void **); 6527#endif 6528#ifdef CONFIG_CPUMASK_OFFSTACK 6529 alloc_size += num_possible_cpus() * cpumask_size(); 6530#endif 6531 if (alloc_size) { 6532 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT); 6533 6534#ifdef CONFIG_FAIR_GROUP_SCHED 6535 root_task_group.se = (struct sched_entity **)ptr; 6536 ptr += nr_cpu_ids * sizeof(void **); 6537 6538 root_task_group.cfs_rq = (struct cfs_rq **)ptr; 6539 ptr += nr_cpu_ids * sizeof(void **); 6540 6541#endif /* CONFIG_FAIR_GROUP_SCHED */ 6542#ifdef CONFIG_RT_GROUP_SCHED 6543 root_task_group.rt_se = (struct sched_rt_entity **)ptr; 6544 ptr += nr_cpu_ids * sizeof(void **); 6545 6546 root_task_group.rt_rq = (struct rt_rq **)ptr; 6547 ptr += nr_cpu_ids * sizeof(void **); 6548 6549#endif /* CONFIG_RT_GROUP_SCHED */ 6550#ifdef CONFIG_CPUMASK_OFFSTACK 6551 for_each_possible_cpu(i) { 6552 per_cpu(load_balance_mask, i) = (void *)ptr; 6553 ptr += cpumask_size(); 6554 } 6555#endif /* CONFIG_CPUMASK_OFFSTACK */ 6556 } 6557 6558#ifdef CONFIG_SMP 6559 init_defrootdomain(); 6560#endif 6561 6562 init_rt_bandwidth(&def_rt_bandwidth, 6563 global_rt_period(), global_rt_runtime()); 6564 6565#ifdef CONFIG_RT_GROUP_SCHED 6566 init_rt_bandwidth(&root_task_group.rt_bandwidth, 6567 global_rt_period(), global_rt_runtime()); 6568#endif /* CONFIG_RT_GROUP_SCHED */ 6569 6570#ifdef CONFIG_CGROUP_SCHED 6571 list_add(&root_task_group.list, &task_groups); 6572 INIT_LIST_HEAD(&root_task_group.children); 6573 INIT_LIST_HEAD(&root_task_group.siblings); 6574 autogroup_init(&init_task); 6575 6576#endif /* CONFIG_CGROUP_SCHED */ 6577 6578 for_each_possible_cpu(i) { 6579 struct rq *rq; 6580 6581 rq = cpu_rq(i); 6582 raw_spin_lock_init(&rq->lock); 6583 rq->nr_running = 0; 6584 rq->calc_load_active = 0; 6585 rq->calc_load_update = jiffies + LOAD_FREQ; 6586 init_cfs_rq(&rq->cfs); 6587 init_rt_rq(&rq->rt, rq); 6588#ifdef CONFIG_FAIR_GROUP_SCHED 6589 root_task_group.shares = ROOT_TASK_GROUP_LOAD; 6590 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list); 6591 /* 6592 * How much cpu bandwidth does root_task_group get? 6593 * 6594 * In case of task-groups formed thr' the cgroup filesystem, it 6595 * gets 100% of the cpu resources in the system. This overall 6596 * system cpu resource is divided among the tasks of 6597 * root_task_group and its child task-groups in a fair manner, 6598 * based on each entity's (task or task-group's) weight 6599 * (se->load.weight). 6600 * 6601 * In other words, if root_task_group has 10 tasks of weight 6602 * 1024) and two child groups A0 and A1 (of weight 1024 each), 6603 * then A0's share of the cpu resource is: 6604 * 6605 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33% 6606 * 6607 * We achieve this by letting root_task_group's tasks sit 6608 * directly in rq->cfs (i.e root_task_group->se[] = NULL). 6609 */ 6610 init_cfs_bandwidth(&root_task_group.cfs_bandwidth); 6611 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL); 6612#endif /* CONFIG_FAIR_GROUP_SCHED */ 6613 6614 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime; 6615#ifdef CONFIG_RT_GROUP_SCHED 6616 INIT_LIST_HEAD(&rq->leaf_rt_rq_list); 6617 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL); 6618#endif 6619 6620 for (j = 0; j < CPU_LOAD_IDX_MAX; j++) 6621 rq->cpu_load[j] = 0; 6622 6623 rq->last_load_update_tick = jiffies; 6624 6625#ifdef CONFIG_SMP 6626 rq->sd = NULL; 6627 rq->rd = NULL; 6628 rq->cpu_power = SCHED_POWER_SCALE; 6629 rq->post_schedule = 0; 6630 rq->active_balance = 0; 6631 rq->next_balance = jiffies; 6632 rq->push_cpu = 0; 6633 rq->cpu = i; 6634 rq->online = 0; 6635 rq->idle_stamp = 0; 6636 rq->avg_idle = 2*sysctl_sched_migration_cost; 6637 rq->max_idle_balance_cost = sysctl_sched_migration_cost; 6638 6639 INIT_LIST_HEAD(&rq->cfs_tasks); 6640 6641 rq_attach_root(rq, &def_root_domain); 6642#ifdef CONFIG_NO_HZ_COMMON 6643 rq->nohz_flags = 0; 6644#endif 6645#ifdef CONFIG_NO_HZ_FULL 6646 rq->last_sched_tick = 0; 6647#endif 6648#endif 6649 init_rq_hrtick(rq); 6650 atomic_set(&rq->nr_iowait, 0); 6651 } 6652 6653 set_load_weight(&init_task); 6654 6655#ifdef CONFIG_PREEMPT_NOTIFIERS 6656 INIT_HLIST_HEAD(&init_task.preempt_notifiers); 6657#endif 6658 6659#ifdef CONFIG_RT_MUTEXES 6660 plist_head_init(&init_task.pi_waiters); 6661#endif 6662 6663 /* 6664 * The boot idle thread does lazy MMU switching as well: 6665 */ 6666 atomic_inc(&init_mm.mm_count); 6667 enter_lazy_tlb(&init_mm, current); 6668 6669 /* 6670 * Make us the idle thread. Technically, schedule() should not be 6671 * called from this thread, however somewhere below it might be, 6672 * but because we are the idle thread, we just pick up running again 6673 * when this runqueue becomes "idle". 6674 */ 6675 init_idle(current, smp_processor_id()); 6676 6677 calc_load_update = jiffies + LOAD_FREQ; 6678 6679 /* 6680 * During early bootup we pretend to be a normal task: 6681 */ 6682 current->sched_class = &fair_sched_class; 6683 6684#ifdef CONFIG_SMP 6685 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT); 6686 /* May be allocated at isolcpus cmdline parse time */ 6687 if (cpu_isolated_map == NULL) 6688 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT); 6689 idle_thread_set_boot_cpu(); 6690#endif 6691 init_sched_fair_class(); 6692 6693 scheduler_running = 1; 6694} 6695 6696#ifdef CONFIG_DEBUG_ATOMIC_SLEEP 6697static inline int preempt_count_equals(int preempt_offset) 6698{ 6699 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth(); 6700 6701 return (nested == preempt_offset); 6702} 6703 6704void __might_sleep(const char *file, int line, int preempt_offset) 6705{ 6706 static unsigned long prev_jiffy; /* ratelimiting */ 6707 6708 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */ 6709 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) || 6710 system_state != SYSTEM_RUNNING || oops_in_progress) 6711 return; 6712 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy) 6713 return; 6714 prev_jiffy = jiffies; 6715 6716 printk(KERN_ERR 6717 "BUG: sleeping function called from invalid context at %s:%d\n", 6718 file, line); 6719 printk(KERN_ERR 6720 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n", 6721 in_atomic(), irqs_disabled(), 6722 current->pid, current->comm); 6723 6724 debug_show_held_locks(current); 6725 if (irqs_disabled()) 6726 print_irqtrace_events(current); 6727 dump_stack(); 6728} 6729EXPORT_SYMBOL(__might_sleep); 6730#endif 6731 6732#ifdef CONFIG_MAGIC_SYSRQ 6733static void normalize_task(struct rq *rq, struct task_struct *p) 6734{ 6735 const struct sched_class *prev_class = p->sched_class; 6736 int old_prio = p->prio; 6737 int on_rq; 6738 6739 on_rq = p->on_rq; 6740 if (on_rq) 6741 dequeue_task(rq, p, 0); 6742 __setscheduler(rq, p, SCHED_NORMAL, 0); 6743 if (on_rq) { 6744 enqueue_task(rq, p, 0); 6745 resched_task(rq->curr); 6746 } 6747 6748 check_class_changed(rq, p, prev_class, old_prio); 6749} 6750 6751void normalize_rt_tasks(void) 6752{ 6753 struct task_struct *g, *p; 6754 unsigned long flags; 6755 struct rq *rq; 6756 6757 read_lock_irqsave(&tasklist_lock, flags); 6758 do_each_thread(g, p) { 6759 /* 6760 * Only normalize user tasks: 6761 */ 6762 if (!p->mm) 6763 continue; 6764 6765 p->se.exec_start = 0; 6766#ifdef CONFIG_SCHEDSTATS 6767 p->se.statistics.wait_start = 0; 6768 p->se.statistics.sleep_start = 0; 6769 p->se.statistics.block_start = 0; 6770#endif 6771 6772 if (!rt_task(p)) { 6773 /* 6774 * Renice negative nice level userspace 6775 * tasks back to 0: 6776 */ 6777 if (TASK_NICE(p) < 0 && p->mm) 6778 set_user_nice(p, 0); 6779 continue; 6780 } 6781 6782 raw_spin_lock(&p->pi_lock); 6783 rq = __task_rq_lock(p); 6784 6785 normalize_task(rq, p); 6786 6787 __task_rq_unlock(rq); 6788 raw_spin_unlock(&p->pi_lock); 6789 } while_each_thread(g, p); 6790 6791 read_unlock_irqrestore(&tasklist_lock, flags); 6792} 6793 6794#endif /* CONFIG_MAGIC_SYSRQ */ 6795 6796#if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) 6797/* 6798 * These functions are only useful for the IA64 MCA handling, or kdb. 6799 * 6800 * They can only be called when the whole system has been 6801 * stopped - every CPU needs to be quiescent, and no scheduling 6802 * activity can take place. Using them for anything else would 6803 * be a serious bug, and as a result, they aren't even visible 6804 * under any other configuration. 6805 */ 6806 6807/** 6808 * curr_task - return the current task for a given cpu. 6809 * @cpu: the processor in question. 6810 * 6811 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED! 6812 * 6813 * Return: The current task for @cpu. 6814 */ 6815struct task_struct *curr_task(int cpu) 6816{ 6817 return cpu_curr(cpu); 6818} 6819 6820#endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */ 6821 6822#ifdef CONFIG_IA64 6823/** 6824 * set_curr_task - set the current task for a given cpu. 6825 * @cpu: the processor in question. 6826 * @p: the task pointer to set. 6827 * 6828 * Description: This function must only be used when non-maskable interrupts 6829 * are serviced on a separate stack. It allows the architecture to switch the 6830 * notion of the current task on a cpu in a non-blocking manner. This function 6831 * must be called with all CPU's synchronized, and interrupts disabled, the 6832 * and caller must save the original value of the current task (see 6833 * curr_task() above) and restore that value before reenabling interrupts and 6834 * re-starting the system. 6835 * 6836 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED! 6837 */ 6838void set_curr_task(int cpu, struct task_struct *p) 6839{ 6840 cpu_curr(cpu) = p; 6841} 6842 6843#endif 6844 6845#ifdef CONFIG_CGROUP_SCHED 6846/* task_group_lock serializes the addition/removal of task groups */ 6847static DEFINE_SPINLOCK(task_group_lock); 6848 6849static void free_sched_group(struct task_group *tg) 6850{ 6851 free_fair_sched_group(tg); 6852 free_rt_sched_group(tg); 6853 autogroup_free(tg); 6854 kfree(tg); 6855} 6856 6857/* allocate runqueue etc for a new task group */ 6858struct task_group *sched_create_group(struct task_group *parent) 6859{ 6860 struct task_group *tg; 6861 6862 tg = kzalloc(sizeof(*tg), GFP_KERNEL); 6863 if (!tg) 6864 return ERR_PTR(-ENOMEM); 6865 6866 if (!alloc_fair_sched_group(tg, parent)) 6867 goto err; 6868 6869 if (!alloc_rt_sched_group(tg, parent)) 6870 goto err; 6871 6872 return tg; 6873 6874err: 6875 free_sched_group(tg); 6876 return ERR_PTR(-ENOMEM); 6877} 6878 6879void sched_online_group(struct task_group *tg, struct task_group *parent) 6880{ 6881 unsigned long flags; 6882 6883 spin_lock_irqsave(&task_group_lock, flags); 6884 list_add_rcu(&tg->list, &task_groups); 6885 6886 WARN_ON(!parent); /* root should already exist */ 6887 6888 tg->parent = parent; 6889 INIT_LIST_HEAD(&tg->children); 6890 list_add_rcu(&tg->siblings, &parent->children); 6891 spin_unlock_irqrestore(&task_group_lock, flags); 6892} 6893 6894/* rcu callback to free various structures associated with a task group */ 6895static void free_sched_group_rcu(struct rcu_head *rhp) 6896{ 6897 /* now it should be safe to free those cfs_rqs */ 6898 free_sched_group(container_of(rhp, struct task_group, rcu)); 6899} 6900 6901/* Destroy runqueue etc associated with a task group */ 6902void sched_destroy_group(struct task_group *tg) 6903{ 6904 /* wait for possible concurrent references to cfs_rqs complete */ 6905 call_rcu(&tg->rcu, free_sched_group_rcu); 6906} 6907 6908void sched_offline_group(struct task_group *tg) 6909{ 6910 unsigned long flags; 6911 int i; 6912 6913 /* end participation in shares distribution */ 6914 for_each_possible_cpu(i) 6915 unregister_fair_sched_group(tg, i); 6916 6917 spin_lock_irqsave(&task_group_lock, flags); 6918 list_del_rcu(&tg->list); 6919 list_del_rcu(&tg->siblings); 6920 spin_unlock_irqrestore(&task_group_lock, flags); 6921} 6922 6923/* change task's runqueue when it moves between groups. 6924 * The caller of this function should have put the task in its new group 6925 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to 6926 * reflect its new group. 6927 */ 6928void sched_move_task(struct task_struct *tsk) 6929{ 6930 struct task_group *tg; 6931 int on_rq, running; 6932 unsigned long flags; 6933 struct rq *rq; 6934 6935 rq = task_rq_lock(tsk, &flags); 6936 6937 running = task_current(rq, tsk); 6938 on_rq = tsk->on_rq; 6939 6940 if (on_rq) 6941 dequeue_task(rq, tsk, 0); 6942 if (unlikely(running)) 6943 tsk->sched_class->put_prev_task(rq, tsk); 6944 6945 tg = container_of(task_css_check(tsk, cpu_cgroup_subsys_id, 6946 lockdep_is_held(&tsk->sighand->siglock)), 6947 struct task_group, css); 6948 tg = autogroup_task_group(tsk, tg); 6949 tsk->sched_task_group = tg; 6950 6951#ifdef CONFIG_FAIR_GROUP_SCHED 6952 if (tsk->sched_class->task_move_group) 6953 tsk->sched_class->task_move_group(tsk, on_rq); 6954 else 6955#endif 6956 set_task_rq(tsk, task_cpu(tsk)); 6957 6958 if (unlikely(running)) 6959 tsk->sched_class->set_curr_task(rq); 6960 if (on_rq) 6961 enqueue_task(rq, tsk, 0); 6962 6963 task_rq_unlock(rq, tsk, &flags); 6964} 6965#endif /* CONFIG_CGROUP_SCHED */ 6966 6967#if defined(CONFIG_RT_GROUP_SCHED) || defined(CONFIG_CFS_BANDWIDTH) 6968static unsigned long to_ratio(u64 period, u64 runtime) 6969{ 6970 if (runtime == RUNTIME_INF) 6971 return 1ULL << 20; 6972 6973 return div64_u64(runtime << 20, period); 6974} 6975#endif 6976 6977#ifdef CONFIG_RT_GROUP_SCHED 6978/* 6979 * Ensure that the real time constraints are schedulable. 6980 */ 6981static DEFINE_MUTEX(rt_constraints_mutex); 6982 6983/* Must be called with tasklist_lock held */ 6984static inline int tg_has_rt_tasks(struct task_group *tg) 6985{ 6986 struct task_struct *g, *p; 6987 6988 do_each_thread(g, p) { 6989 if (rt_task(p) && task_rq(p)->rt.tg == tg) 6990 return 1; 6991 } while_each_thread(g, p); 6992 6993 return 0; 6994} 6995 6996struct rt_schedulable_data { 6997 struct task_group *tg; 6998 u64 rt_period; 6999 u64 rt_runtime; 7000}; 7001 7002static int tg_rt_schedulable(struct task_group *tg, void *data) 7003{ 7004 struct rt_schedulable_data *d = data; 7005 struct task_group *child; 7006 unsigned long total, sum = 0; 7007 u64 period, runtime; 7008 7009 period = ktime_to_ns(tg->rt_bandwidth.rt_period); 7010 runtime = tg->rt_bandwidth.rt_runtime; 7011 7012 if (tg == d->tg) { 7013 period = d->rt_period; 7014 runtime = d->rt_runtime; 7015 } 7016 7017 /* 7018 * Cannot have more runtime than the period. 7019 */ 7020 if (runtime > period && runtime != RUNTIME_INF) 7021 return -EINVAL; 7022 7023 /* 7024 * Ensure we don't starve existing RT tasks. 7025 */ 7026 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg)) 7027 return -EBUSY; 7028 7029 total = to_ratio(period, runtime); 7030 7031 /* 7032 * Nobody can have more than the global setting allows. 7033 */ 7034 if (total > to_ratio(global_rt_period(), global_rt_runtime())) 7035 return -EINVAL; 7036 7037 /* 7038 * The sum of our children's runtime should not exceed our own. 7039 */ 7040 list_for_each_entry_rcu(child, &tg->children, siblings) { 7041 period = ktime_to_ns(child->rt_bandwidth.rt_period); 7042 runtime = child->rt_bandwidth.rt_runtime; 7043 7044 if (child == d->tg) { 7045 period = d->rt_period; 7046 runtime = d->rt_runtime; 7047 } 7048 7049 sum += to_ratio(period, runtime); 7050 } 7051 7052 if (sum > total) 7053 return -EINVAL; 7054 7055 return 0; 7056} 7057 7058static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime) 7059{ 7060 int ret; 7061 7062 struct rt_schedulable_data data = { 7063 .tg = tg, 7064 .rt_period = period, 7065 .rt_runtime = runtime, 7066 }; 7067 7068 rcu_read_lock(); 7069 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data); 7070 rcu_read_unlock(); 7071 7072 return ret; 7073} 7074 7075static int tg_set_rt_bandwidth(struct task_group *tg, 7076 u64 rt_period, u64 rt_runtime) 7077{ 7078 int i, err = 0; 7079 7080 mutex_lock(&rt_constraints_mutex); 7081 read_lock(&tasklist_lock); 7082 err = __rt_schedulable(tg, rt_period, rt_runtime); 7083 if (err) 7084 goto unlock; 7085 7086 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock); 7087 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period); 7088 tg->rt_bandwidth.rt_runtime = rt_runtime; 7089 7090 for_each_possible_cpu(i) { 7091 struct rt_rq *rt_rq = tg->rt_rq[i]; 7092 7093 raw_spin_lock(&rt_rq->rt_runtime_lock); 7094 rt_rq->rt_runtime = rt_runtime; 7095 raw_spin_unlock(&rt_rq->rt_runtime_lock); 7096 } 7097 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock); 7098unlock: 7099 read_unlock(&tasklist_lock); 7100 mutex_unlock(&rt_constraints_mutex); 7101 7102 return err; 7103} 7104 7105static int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us) 7106{ 7107 u64 rt_runtime, rt_period; 7108 7109 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period); 7110 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC; 7111 if (rt_runtime_us < 0) 7112 rt_runtime = RUNTIME_INF; 7113 7114 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime); 7115} 7116 7117static long sched_group_rt_runtime(struct task_group *tg) 7118{ 7119 u64 rt_runtime_us; 7120 7121 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF) 7122 return -1; 7123 7124 rt_runtime_us = tg->rt_bandwidth.rt_runtime; 7125 do_div(rt_runtime_us, NSEC_PER_USEC); 7126 return rt_runtime_us; 7127} 7128 7129static int sched_group_set_rt_period(struct task_group *tg, long rt_period_us) 7130{ 7131 u64 rt_runtime, rt_period; 7132 7133 rt_period = (u64)rt_period_us * NSEC_PER_USEC; 7134 rt_runtime = tg->rt_bandwidth.rt_runtime; 7135 7136 if (rt_period == 0) 7137 return -EINVAL; 7138 7139 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime); 7140} 7141 7142static long sched_group_rt_period(struct task_group *tg) 7143{ 7144 u64 rt_period_us; 7145 7146 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period); 7147 do_div(rt_period_us, NSEC_PER_USEC); 7148 return rt_period_us; 7149} 7150 7151static int sched_rt_global_constraints(void) 7152{ 7153 u64 runtime, period; 7154 int ret = 0; 7155 7156 if (sysctl_sched_rt_period <= 0) 7157 return -EINVAL; 7158 7159 runtime = global_rt_runtime(); 7160 period = global_rt_period(); 7161 7162 /* 7163 * Sanity check on the sysctl variables. 7164 */ 7165 if (runtime > period && runtime != RUNTIME_INF) 7166 return -EINVAL; 7167 7168 mutex_lock(&rt_constraints_mutex); 7169 read_lock(&tasklist_lock); 7170 ret = __rt_schedulable(NULL, 0, 0); 7171 read_unlock(&tasklist_lock); 7172 mutex_unlock(&rt_constraints_mutex); 7173 7174 return ret; 7175} 7176 7177static int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk) 7178{ 7179 /* Don't accept realtime tasks when there is no way for them to run */ 7180 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0) 7181 return 0; 7182 7183 return 1; 7184} 7185 7186#else /* !CONFIG_RT_GROUP_SCHED */ 7187static int sched_rt_global_constraints(void) 7188{ 7189 unsigned long flags; 7190 int i; 7191 7192 if (sysctl_sched_rt_period <= 0) 7193 return -EINVAL; 7194 7195 /* 7196 * There's always some RT tasks in the root group 7197 * -- migration, kstopmachine etc.. 7198 */ 7199 if (sysctl_sched_rt_runtime == 0) 7200 return -EBUSY; 7201 7202 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags); 7203 for_each_possible_cpu(i) { 7204 struct rt_rq *rt_rq = &cpu_rq(i)->rt; 7205 7206 raw_spin_lock(&rt_rq->rt_runtime_lock); 7207 rt_rq->rt_runtime = global_rt_runtime(); 7208 raw_spin_unlock(&rt_rq->rt_runtime_lock); 7209 } 7210 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags); 7211 7212 return 0; 7213} 7214#endif /* CONFIG_RT_GROUP_SCHED */ 7215 7216int sched_rr_handler(struct ctl_table *table, int write, 7217 void __user *buffer, size_t *lenp, 7218 loff_t *ppos) 7219{ 7220 int ret; 7221 static DEFINE_MUTEX(mutex); 7222 7223 mutex_lock(&mutex); 7224 ret = proc_dointvec(table, write, buffer, lenp, ppos); 7225 /* make sure that internally we keep jiffies */ 7226 /* also, writing zero resets timeslice to default */ 7227 if (!ret && write) { 7228 sched_rr_timeslice = sched_rr_timeslice <= 0 ? 7229 RR_TIMESLICE : msecs_to_jiffies(sched_rr_timeslice); 7230 } 7231 mutex_unlock(&mutex); 7232 return ret; 7233} 7234 7235int sched_rt_handler(struct ctl_table *table, int write, 7236 void __user *buffer, size_t *lenp, 7237 loff_t *ppos) 7238{ 7239 int ret; 7240 int old_period, old_runtime; 7241 static DEFINE_MUTEX(mutex); 7242 7243 mutex_lock(&mutex); 7244 old_period = sysctl_sched_rt_period; 7245 old_runtime = sysctl_sched_rt_runtime; 7246 7247 ret = proc_dointvec(table, write, buffer, lenp, ppos); 7248 7249 if (!ret && write) { 7250 ret = sched_rt_global_constraints(); 7251 if (ret) { 7252 sysctl_sched_rt_period = old_period; 7253 sysctl_sched_rt_runtime = old_runtime; 7254 } else { 7255 def_rt_bandwidth.rt_runtime = global_rt_runtime(); 7256 def_rt_bandwidth.rt_period = 7257 ns_to_ktime(global_rt_period()); 7258 } 7259 } 7260 mutex_unlock(&mutex); 7261 7262 return ret; 7263} 7264 7265#ifdef CONFIG_CGROUP_SCHED 7266 7267static inline struct task_group *css_tg(struct cgroup_subsys_state *css) 7268{ 7269 return css ? container_of(css, struct task_group, css) : NULL; 7270} 7271 7272static struct cgroup_subsys_state * 7273cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css) 7274{ 7275 struct task_group *parent = css_tg(parent_css); 7276 struct task_group *tg; 7277 7278 if (!parent) { 7279 /* This is early initialization for the top cgroup */ 7280 return &root_task_group.css; 7281 } 7282 7283 tg = sched_create_group(parent); 7284 if (IS_ERR(tg)) 7285 return ERR_PTR(-ENOMEM); 7286 7287 return &tg->css; 7288} 7289 7290static int cpu_cgroup_css_online(struct cgroup_subsys_state *css) 7291{ 7292 struct task_group *tg = css_tg(css); 7293 struct task_group *parent = css_tg(css_parent(css)); 7294 7295 if (parent) 7296 sched_online_group(tg, parent); 7297 return 0; 7298} 7299 7300static void cpu_cgroup_css_free(struct cgroup_subsys_state *css) 7301{ 7302 struct task_group *tg = css_tg(css); 7303 7304 sched_destroy_group(tg); 7305} 7306 7307static void cpu_cgroup_css_offline(struct cgroup_subsys_state *css) 7308{ 7309 struct task_group *tg = css_tg(css); 7310 7311 sched_offline_group(tg); 7312} 7313 7314static int cpu_cgroup_can_attach(struct cgroup_subsys_state *css, 7315 struct cgroup_taskset *tset) 7316{ 7317 struct task_struct *task; 7318 7319 cgroup_taskset_for_each(task, css, tset) { 7320#ifdef CONFIG_RT_GROUP_SCHED 7321 if (!sched_rt_can_attach(css_tg(css), task)) 7322 return -EINVAL; 7323#else 7324 /* We don't support RT-tasks being in separate groups */ 7325 if (task->sched_class != &fair_sched_class) 7326 return -EINVAL; 7327#endif 7328 } 7329 return 0; 7330} 7331 7332static void cpu_cgroup_attach(struct cgroup_subsys_state *css, 7333 struct cgroup_taskset *tset) 7334{ 7335 struct task_struct *task; 7336 7337 cgroup_taskset_for_each(task, css, tset) 7338 sched_move_task(task); 7339} 7340 7341static void cpu_cgroup_exit(struct cgroup_subsys_state *css, 7342 struct cgroup_subsys_state *old_css, 7343 struct task_struct *task) 7344{ 7345 /* 7346 * cgroup_exit() is called in the copy_process() failure path. 7347 * Ignore this case since the task hasn't ran yet, this avoids 7348 * trying to poke a half freed task state from generic code. 7349 */ 7350 if (!(task->flags & PF_EXITING)) 7351 return; 7352 7353 sched_move_task(task); 7354} 7355 7356#ifdef CONFIG_FAIR_GROUP_SCHED 7357static int cpu_shares_write_u64(struct cgroup_subsys_state *css, 7358 struct cftype *cftype, u64 shareval) 7359{ 7360 return sched_group_set_shares(css_tg(css), scale_load(shareval)); 7361} 7362 7363static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css, 7364 struct cftype *cft) 7365{ 7366 struct task_group *tg = css_tg(css); 7367 7368 return (u64) scale_load_down(tg->shares); 7369} 7370 7371#ifdef CONFIG_CFS_BANDWIDTH 7372static DEFINE_MUTEX(cfs_constraints_mutex); 7373 7374const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */ 7375const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */ 7376 7377static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime); 7378 7379static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota) 7380{ 7381 int i, ret = 0, runtime_enabled, runtime_was_enabled; 7382 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth; 7383 7384 if (tg == &root_task_group) 7385 return -EINVAL; 7386 7387 /* 7388 * Ensure we have at some amount of bandwidth every period. This is 7389 * to prevent reaching a state of large arrears when throttled via 7390 * entity_tick() resulting in prolonged exit starvation. 7391 */ 7392 if (quota < min_cfs_quota_period || period < min_cfs_quota_period) 7393 return -EINVAL; 7394 7395 /* 7396 * Likewise, bound things on the otherside by preventing insane quota 7397 * periods. This also allows us to normalize in computing quota 7398 * feasibility. 7399 */ 7400 if (period > max_cfs_quota_period) 7401 return -EINVAL; 7402 7403 mutex_lock(&cfs_constraints_mutex); 7404 ret = __cfs_schedulable(tg, period, quota); 7405 if (ret) 7406 goto out_unlock; 7407 7408 runtime_enabled = quota != RUNTIME_INF; 7409 runtime_was_enabled = cfs_b->quota != RUNTIME_INF; 7410 account_cfs_bandwidth_used(runtime_enabled, runtime_was_enabled); 7411 raw_spin_lock_irq(&cfs_b->lock); 7412 cfs_b->period = ns_to_ktime(period); 7413 cfs_b->quota = quota; 7414 7415 __refill_cfs_bandwidth_runtime(cfs_b); 7416 /* restart the period timer (if active) to handle new period expiry */ 7417 if (runtime_enabled && cfs_b->timer_active) { 7418 /* force a reprogram */ 7419 cfs_b->timer_active = 0; 7420 __start_cfs_bandwidth(cfs_b); 7421 } 7422 raw_spin_unlock_irq(&cfs_b->lock); 7423 7424 for_each_possible_cpu(i) { 7425 struct cfs_rq *cfs_rq = tg->cfs_rq[i]; 7426 struct rq *rq = cfs_rq->rq; 7427 7428 raw_spin_lock_irq(&rq->lock); 7429 cfs_rq->runtime_enabled = runtime_enabled; 7430 cfs_rq->runtime_remaining = 0; 7431 7432 if (cfs_rq->throttled) 7433 unthrottle_cfs_rq(cfs_rq); 7434 raw_spin_unlock_irq(&rq->lock); 7435 } 7436out_unlock: 7437 mutex_unlock(&cfs_constraints_mutex); 7438 7439 return ret; 7440} 7441 7442int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us) 7443{ 7444 u64 quota, period; 7445 7446 period = ktime_to_ns(tg->cfs_bandwidth.period); 7447 if (cfs_quota_us < 0) 7448 quota = RUNTIME_INF; 7449 else 7450 quota = (u64)cfs_quota_us * NSEC_PER_USEC; 7451 7452 return tg_set_cfs_bandwidth(tg, period, quota); 7453} 7454 7455long tg_get_cfs_quota(struct task_group *tg) 7456{ 7457 u64 quota_us; 7458 7459 if (tg->cfs_bandwidth.quota == RUNTIME_INF) 7460 return -1; 7461 7462 quota_us = tg->cfs_bandwidth.quota; 7463 do_div(quota_us, NSEC_PER_USEC); 7464 7465 return quota_us; 7466} 7467 7468int tg_set_cfs_period(struct task_group *tg, long cfs_period_us) 7469{ 7470 u64 quota, period; 7471 7472 period = (u64)cfs_period_us * NSEC_PER_USEC; 7473 quota = tg->cfs_bandwidth.quota; 7474 7475 return tg_set_cfs_bandwidth(tg, period, quota); 7476} 7477 7478long tg_get_cfs_period(struct task_group *tg) 7479{ 7480 u64 cfs_period_us; 7481 7482 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period); 7483 do_div(cfs_period_us, NSEC_PER_USEC); 7484 7485 return cfs_period_us; 7486} 7487 7488static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css, 7489 struct cftype *cft) 7490{ 7491 return tg_get_cfs_quota(css_tg(css)); 7492} 7493 7494static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css, 7495 struct cftype *cftype, s64 cfs_quota_us) 7496{ 7497 return tg_set_cfs_quota(css_tg(css), cfs_quota_us); 7498} 7499 7500static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css, 7501 struct cftype *cft) 7502{ 7503 return tg_get_cfs_period(css_tg(css)); 7504} 7505 7506static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css, 7507 struct cftype *cftype, u64 cfs_period_us) 7508{ 7509 return tg_set_cfs_period(css_tg(css), cfs_period_us); 7510} 7511 7512struct cfs_schedulable_data { 7513 struct task_group *tg; 7514 u64 period, quota; 7515}; 7516 7517/* 7518 * normalize group quota/period to be quota/max_period 7519 * note: units are usecs 7520 */ 7521static u64 normalize_cfs_quota(struct task_group *tg, 7522 struct cfs_schedulable_data *d) 7523{ 7524 u64 quota, period; 7525 7526 if (tg == d->tg) { 7527 period = d->period; 7528 quota = d->quota; 7529 } else { 7530 period = tg_get_cfs_period(tg); 7531 quota = tg_get_cfs_quota(tg); 7532 } 7533 7534 /* note: these should typically be equivalent */ 7535 if (quota == RUNTIME_INF || quota == -1) 7536 return RUNTIME_INF; 7537 7538 return to_ratio(period, quota); 7539} 7540 7541static int tg_cfs_schedulable_down(struct task_group *tg, void *data) 7542{ 7543 struct cfs_schedulable_data *d = data; 7544 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth; 7545 s64 quota = 0, parent_quota = -1; 7546 7547 if (!tg->parent) { 7548 quota = RUNTIME_INF; 7549 } else { 7550 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth; 7551 7552 quota = normalize_cfs_quota(tg, d); 7553 parent_quota = parent_b->hierarchal_quota; 7554 7555 /* 7556 * ensure max(child_quota) <= parent_quota, inherit when no 7557 * limit is set 7558 */ 7559 if (quota == RUNTIME_INF) 7560 quota = parent_quota; 7561 else if (parent_quota != RUNTIME_INF && quota > parent_quota) 7562 return -EINVAL; 7563 } 7564 cfs_b->hierarchal_quota = quota; 7565 7566 return 0; 7567} 7568 7569static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota) 7570{ 7571 int ret; 7572 struct cfs_schedulable_data data = { 7573 .tg = tg, 7574 .period = period, 7575 .quota = quota, 7576 }; 7577 7578 if (quota != RUNTIME_INF) { 7579 do_div(data.period, NSEC_PER_USEC); 7580 do_div(data.quota, NSEC_PER_USEC); 7581 } 7582 7583 rcu_read_lock(); 7584 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data); 7585 rcu_read_unlock(); 7586 7587 return ret; 7588} 7589 7590static int cpu_stats_show(struct cgroup_subsys_state *css, struct cftype *cft, 7591 struct cgroup_map_cb *cb) 7592{ 7593 struct task_group *tg = css_tg(css); 7594 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth; 7595 7596 cb->fill(cb, "nr_periods", cfs_b->nr_periods); 7597 cb->fill(cb, "nr_throttled", cfs_b->nr_throttled); 7598 cb->fill(cb, "throttled_time", cfs_b->throttled_time); 7599 7600 return 0; 7601} 7602#endif /* CONFIG_CFS_BANDWIDTH */ 7603#endif /* CONFIG_FAIR_GROUP_SCHED */ 7604 7605#ifdef CONFIG_RT_GROUP_SCHED 7606static int cpu_rt_runtime_write(struct cgroup_subsys_state *css, 7607 struct cftype *cft, s64 val) 7608{ 7609 return sched_group_set_rt_runtime(css_tg(css), val); 7610} 7611 7612static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css, 7613 struct cftype *cft) 7614{ 7615 return sched_group_rt_runtime(css_tg(css)); 7616} 7617 7618static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css, 7619 struct cftype *cftype, u64 rt_period_us) 7620{ 7621 return sched_group_set_rt_period(css_tg(css), rt_period_us); 7622} 7623 7624static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css, 7625 struct cftype *cft) 7626{ 7627 return sched_group_rt_period(css_tg(css)); 7628} 7629#endif /* CONFIG_RT_GROUP_SCHED */ 7630 7631static struct cftype cpu_files[] = { 7632#ifdef CONFIG_FAIR_GROUP_SCHED 7633 { 7634 .name = "shares", 7635 .read_u64 = cpu_shares_read_u64, 7636 .write_u64 = cpu_shares_write_u64, 7637 }, 7638#endif 7639#ifdef CONFIG_CFS_BANDWIDTH 7640 { 7641 .name = "cfs_quota_us", 7642 .read_s64 = cpu_cfs_quota_read_s64, 7643 .write_s64 = cpu_cfs_quota_write_s64, 7644 }, 7645 { 7646 .name = "cfs_period_us", 7647 .read_u64 = cpu_cfs_period_read_u64, 7648 .write_u64 = cpu_cfs_period_write_u64, 7649 }, 7650 { 7651 .name = "stat", 7652 .read_map = cpu_stats_show, 7653 }, 7654#endif 7655#ifdef CONFIG_RT_GROUP_SCHED 7656 { 7657 .name = "rt_runtime_us", 7658 .read_s64 = cpu_rt_runtime_read, 7659 .write_s64 = cpu_rt_runtime_write, 7660 }, 7661 { 7662 .name = "rt_period_us", 7663 .read_u64 = cpu_rt_period_read_uint, 7664 .write_u64 = cpu_rt_period_write_uint, 7665 }, 7666#endif 7667 { } /* terminate */ 7668}; 7669 7670struct cgroup_subsys cpu_cgroup_subsys = { 7671 .name = "cpu", 7672 .css_alloc = cpu_cgroup_css_alloc, 7673 .css_free = cpu_cgroup_css_free, 7674 .css_online = cpu_cgroup_css_online, 7675 .css_offline = cpu_cgroup_css_offline, 7676 .can_attach = cpu_cgroup_can_attach, 7677 .attach = cpu_cgroup_attach, 7678 .exit = cpu_cgroup_exit, 7679 .subsys_id = cpu_cgroup_subsys_id, 7680 .base_cftypes = cpu_files, 7681 .early_init = 1, 7682}; 7683 7684#endif /* CONFIG_CGROUP_SCHED */ 7685 7686void dump_cpu_task(int cpu) 7687{ 7688 pr_info("Task dump for CPU %d:\n", cpu); 7689 sched_show_task(cpu_curr(cpu)); 7690} 7691