core.c revision 4a2b4b222743bb07fedf985b884550f2ca067ea9
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 return; 530 531 /* NEED_RESCHED must be visible before we test polling */ 532 smp_mb(); 533 if (!tsk_is_polling(p)) 534 smp_send_reschedule(cpu); 535} 536 537void resched_cpu(int cpu) 538{ 539 struct rq *rq = cpu_rq(cpu); 540 unsigned long flags; 541 542 if (!raw_spin_trylock_irqsave(&rq->lock, flags)) 543 return; 544 resched_task(cpu_curr(cpu)); 545 raw_spin_unlock_irqrestore(&rq->lock, flags); 546} 547 548#ifdef CONFIG_SMP 549#ifdef CONFIG_NO_HZ_COMMON 550/* 551 * In the semi idle case, use the nearest busy cpu for migrating timers 552 * from an idle cpu. This is good for power-savings. 553 * 554 * We don't do similar optimization for completely idle system, as 555 * selecting an idle cpu will add more delays to the timers than intended 556 * (as that cpu's timer base may not be uptodate wrt jiffies etc). 557 */ 558int get_nohz_timer_target(void) 559{ 560 int cpu = smp_processor_id(); 561 int i; 562 struct sched_domain *sd; 563 564 rcu_read_lock(); 565 for_each_domain(cpu, sd) { 566 for_each_cpu(i, sched_domain_span(sd)) { 567 if (!idle_cpu(i)) { 568 cpu = i; 569 goto unlock; 570 } 571 } 572 } 573unlock: 574 rcu_read_unlock(); 575 return cpu; 576} 577/* 578 * When add_timer_on() enqueues a timer into the timer wheel of an 579 * idle CPU then this timer might expire before the next timer event 580 * which is scheduled to wake up that CPU. In case of a completely 581 * idle system the next event might even be infinite time into the 582 * future. wake_up_idle_cpu() ensures that the CPU is woken up and 583 * leaves the inner idle loop so the newly added timer is taken into 584 * account when the CPU goes back to idle and evaluates the timer 585 * wheel for the next timer event. 586 */ 587static void wake_up_idle_cpu(int cpu) 588{ 589 struct rq *rq = cpu_rq(cpu); 590 591 if (cpu == smp_processor_id()) 592 return; 593 594 /* 595 * This is safe, as this function is called with the timer 596 * wheel base lock of (cpu) held. When the CPU is on the way 597 * to idle and has not yet set rq->curr to idle then it will 598 * be serialized on the timer wheel base lock and take the new 599 * timer into account automatically. 600 */ 601 if (rq->curr != rq->idle) 602 return; 603 604 /* 605 * We can set TIF_RESCHED on the idle task of the other CPU 606 * lockless. The worst case is that the other CPU runs the 607 * idle task through an additional NOOP schedule() 608 */ 609 set_tsk_need_resched(rq->idle); 610 611 /* NEED_RESCHED must be visible before we test polling */ 612 smp_mb(); 613 if (!tsk_is_polling(rq->idle)) 614 smp_send_reschedule(cpu); 615} 616 617static bool wake_up_full_nohz_cpu(int cpu) 618{ 619 if (tick_nohz_full_cpu(cpu)) { 620 if (cpu != smp_processor_id() || 621 tick_nohz_tick_stopped()) 622 smp_send_reschedule(cpu); 623 return true; 624 } 625 626 return false; 627} 628 629void wake_up_nohz_cpu(int cpu) 630{ 631 if (!wake_up_full_nohz_cpu(cpu)) 632 wake_up_idle_cpu(cpu); 633} 634 635static inline bool got_nohz_idle_kick(void) 636{ 637 int cpu = smp_processor_id(); 638 639 if (!test_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu))) 640 return false; 641 642 if (idle_cpu(cpu) && !need_resched()) 643 return true; 644 645 /* 646 * We can't run Idle Load Balance on this CPU for this time so we 647 * cancel it and clear NOHZ_BALANCE_KICK 648 */ 649 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu)); 650 return false; 651} 652 653#else /* CONFIG_NO_HZ_COMMON */ 654 655static inline bool got_nohz_idle_kick(void) 656{ 657 return false; 658} 659 660#endif /* CONFIG_NO_HZ_COMMON */ 661 662#ifdef CONFIG_NO_HZ_FULL 663bool sched_can_stop_tick(void) 664{ 665 struct rq *rq; 666 667 rq = this_rq(); 668 669 /* Make sure rq->nr_running update is visible after the IPI */ 670 smp_rmb(); 671 672 /* More than one running task need preemption */ 673 if (rq->nr_running > 1) 674 return false; 675 676 return true; 677} 678#endif /* CONFIG_NO_HZ_FULL */ 679 680void sched_avg_update(struct rq *rq) 681{ 682 s64 period = sched_avg_period(); 683 684 while ((s64)(rq_clock(rq) - rq->age_stamp) > period) { 685 /* 686 * Inline assembly required to prevent the compiler 687 * optimising this loop into a divmod call. 688 * See __iter_div_u64_rem() for another example of this. 689 */ 690 asm("" : "+rm" (rq->age_stamp)); 691 rq->age_stamp += period; 692 rq->rt_avg /= 2; 693 } 694} 695 696#endif /* CONFIG_SMP */ 697 698#if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \ 699 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH))) 700/* 701 * Iterate task_group tree rooted at *from, calling @down when first entering a 702 * node and @up when leaving it for the final time. 703 * 704 * Caller must hold rcu_lock or sufficient equivalent. 705 */ 706int walk_tg_tree_from(struct task_group *from, 707 tg_visitor down, tg_visitor up, void *data) 708{ 709 struct task_group *parent, *child; 710 int ret; 711 712 parent = from; 713 714down: 715 ret = (*down)(parent, data); 716 if (ret) 717 goto out; 718 list_for_each_entry_rcu(child, &parent->children, siblings) { 719 parent = child; 720 goto down; 721 722up: 723 continue; 724 } 725 ret = (*up)(parent, data); 726 if (ret || parent == from) 727 goto out; 728 729 child = parent; 730 parent = parent->parent; 731 if (parent) 732 goto up; 733out: 734 return ret; 735} 736 737int tg_nop(struct task_group *tg, void *data) 738{ 739 return 0; 740} 741#endif 742 743static void set_load_weight(struct task_struct *p) 744{ 745 int prio = p->static_prio - MAX_RT_PRIO; 746 struct load_weight *load = &p->se.load; 747 748 /* 749 * SCHED_IDLE tasks get minimal weight: 750 */ 751 if (p->policy == SCHED_IDLE) { 752 load->weight = scale_load(WEIGHT_IDLEPRIO); 753 load->inv_weight = WMULT_IDLEPRIO; 754 return; 755 } 756 757 load->weight = scale_load(prio_to_weight[prio]); 758 load->inv_weight = prio_to_wmult[prio]; 759} 760 761static void enqueue_task(struct rq *rq, struct task_struct *p, int flags) 762{ 763 update_rq_clock(rq); 764 sched_info_queued(rq, p); 765 p->sched_class->enqueue_task(rq, p, flags); 766} 767 768static void dequeue_task(struct rq *rq, struct task_struct *p, int flags) 769{ 770 update_rq_clock(rq); 771 sched_info_dequeued(rq, p); 772 p->sched_class->dequeue_task(rq, p, flags); 773} 774 775void activate_task(struct rq *rq, struct task_struct *p, int flags) 776{ 777 if (task_contributes_to_load(p)) 778 rq->nr_uninterruptible--; 779 780 enqueue_task(rq, p, flags); 781} 782 783void deactivate_task(struct rq *rq, struct task_struct *p, int flags) 784{ 785 if (task_contributes_to_load(p)) 786 rq->nr_uninterruptible++; 787 788 dequeue_task(rq, p, flags); 789} 790 791static void update_rq_clock_task(struct rq *rq, s64 delta) 792{ 793/* 794 * In theory, the compile should just see 0 here, and optimize out the call 795 * to sched_rt_avg_update. But I don't trust it... 796 */ 797#if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING) 798 s64 steal = 0, irq_delta = 0; 799#endif 800#ifdef CONFIG_IRQ_TIME_ACCOUNTING 801 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time; 802 803 /* 804 * Since irq_time is only updated on {soft,}irq_exit, we might run into 805 * this case when a previous update_rq_clock() happened inside a 806 * {soft,}irq region. 807 * 808 * When this happens, we stop ->clock_task and only update the 809 * prev_irq_time stamp to account for the part that fit, so that a next 810 * update will consume the rest. This ensures ->clock_task is 811 * monotonic. 812 * 813 * It does however cause some slight miss-attribution of {soft,}irq 814 * time, a more accurate solution would be to update the irq_time using 815 * the current rq->clock timestamp, except that would require using 816 * atomic ops. 817 */ 818 if (irq_delta > delta) 819 irq_delta = delta; 820 821 rq->prev_irq_time += irq_delta; 822 delta -= irq_delta; 823#endif 824#ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING 825 if (static_key_false((¶virt_steal_rq_enabled))) { 826 u64 st; 827 828 steal = paravirt_steal_clock(cpu_of(rq)); 829 steal -= rq->prev_steal_time_rq; 830 831 if (unlikely(steal > delta)) 832 steal = delta; 833 834 st = steal_ticks(steal); 835 steal = st * TICK_NSEC; 836 837 rq->prev_steal_time_rq += steal; 838 839 delta -= steal; 840 } 841#endif 842 843 rq->clock_task += delta; 844 845#if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING) 846 if ((irq_delta + steal) && sched_feat(NONTASK_POWER)) 847 sched_rt_avg_update(rq, irq_delta + steal); 848#endif 849} 850 851void sched_set_stop_task(int cpu, struct task_struct *stop) 852{ 853 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 }; 854 struct task_struct *old_stop = cpu_rq(cpu)->stop; 855 856 if (stop) { 857 /* 858 * Make it appear like a SCHED_FIFO task, its something 859 * userspace knows about and won't get confused about. 860 * 861 * Also, it will make PI more or less work without too 862 * much confusion -- but then, stop work should not 863 * rely on PI working anyway. 864 */ 865 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m); 866 867 stop->sched_class = &stop_sched_class; 868 } 869 870 cpu_rq(cpu)->stop = stop; 871 872 if (old_stop) { 873 /* 874 * Reset it back to a normal scheduling class so that 875 * it can die in pieces. 876 */ 877 old_stop->sched_class = &rt_sched_class; 878 } 879} 880 881/* 882 * __normal_prio - return the priority that is based on the static prio 883 */ 884static inline int __normal_prio(struct task_struct *p) 885{ 886 return p->static_prio; 887} 888 889/* 890 * Calculate the expected normal priority: i.e. priority 891 * without taking RT-inheritance into account. Might be 892 * boosted by interactivity modifiers. Changes upon fork, 893 * setprio syscalls, and whenever the interactivity 894 * estimator recalculates. 895 */ 896static inline int normal_prio(struct task_struct *p) 897{ 898 int prio; 899 900 if (task_has_rt_policy(p)) 901 prio = MAX_RT_PRIO-1 - p->rt_priority; 902 else 903 prio = __normal_prio(p); 904 return prio; 905} 906 907/* 908 * Calculate the current priority, i.e. the priority 909 * taken into account by the scheduler. This value might 910 * be boosted by RT tasks, or might be boosted by 911 * interactivity modifiers. Will be RT if the task got 912 * RT-boosted. If not then it returns p->normal_prio. 913 */ 914static int effective_prio(struct task_struct *p) 915{ 916 p->normal_prio = normal_prio(p); 917 /* 918 * If we are RT tasks or we were boosted to RT priority, 919 * keep the priority unchanged. Otherwise, update priority 920 * to the normal priority: 921 */ 922 if (!rt_prio(p->prio)) 923 return p->normal_prio; 924 return p->prio; 925} 926 927/** 928 * task_curr - is this task currently executing on a CPU? 929 * @p: the task in question. 930 * 931 * Return: 1 if the task is currently executing. 0 otherwise. 932 */ 933inline int task_curr(const struct task_struct *p) 934{ 935 return cpu_curr(task_cpu(p)) == p; 936} 937 938static inline void check_class_changed(struct rq *rq, struct task_struct *p, 939 const struct sched_class *prev_class, 940 int oldprio) 941{ 942 if (prev_class != p->sched_class) { 943 if (prev_class->switched_from) 944 prev_class->switched_from(rq, p); 945 p->sched_class->switched_to(rq, p); 946 } else if (oldprio != p->prio) 947 p->sched_class->prio_changed(rq, p, oldprio); 948} 949 950void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags) 951{ 952 const struct sched_class *class; 953 954 if (p->sched_class == rq->curr->sched_class) { 955 rq->curr->sched_class->check_preempt_curr(rq, p, flags); 956 } else { 957 for_each_class(class) { 958 if (class == rq->curr->sched_class) 959 break; 960 if (class == p->sched_class) { 961 resched_task(rq->curr); 962 break; 963 } 964 } 965 } 966 967 /* 968 * A queue event has occurred, and we're going to schedule. In 969 * this case, we can save a useless back to back clock update. 970 */ 971 if (rq->curr->on_rq && test_tsk_need_resched(rq->curr)) 972 rq->skip_clock_update = 1; 973} 974 975#ifdef CONFIG_SMP 976void set_task_cpu(struct task_struct *p, unsigned int new_cpu) 977{ 978#ifdef CONFIG_SCHED_DEBUG 979 /* 980 * We should never call set_task_cpu() on a blocked task, 981 * ttwu() will sort out the placement. 982 */ 983 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING && 984 !(task_thread_info(p)->preempt_count & PREEMPT_ACTIVE)); 985 986#ifdef CONFIG_LOCKDEP 987 /* 988 * The caller should hold either p->pi_lock or rq->lock, when changing 989 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks. 990 * 991 * sched_move_task() holds both and thus holding either pins the cgroup, 992 * see task_group(). 993 * 994 * Furthermore, all task_rq users should acquire both locks, see 995 * task_rq_lock(). 996 */ 997 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) || 998 lockdep_is_held(&task_rq(p)->lock))); 999#endif 1000#endif 1001 1002 trace_sched_migrate_task(p, new_cpu); 1003 1004 if (task_cpu(p) != new_cpu) { 1005 if (p->sched_class->migrate_task_rq) 1006 p->sched_class->migrate_task_rq(p, new_cpu); 1007 p->se.nr_migrations++; 1008 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, NULL, 0); 1009 } 1010 1011 __set_task_cpu(p, new_cpu); 1012} 1013 1014struct migration_arg { 1015 struct task_struct *task; 1016 int dest_cpu; 1017}; 1018 1019static int migration_cpu_stop(void *data); 1020 1021/* 1022 * wait_task_inactive - wait for a thread to unschedule. 1023 * 1024 * If @match_state is nonzero, it's the @p->state value just checked and 1025 * not expected to change. If it changes, i.e. @p might have woken up, 1026 * then return zero. When we succeed in waiting for @p to be off its CPU, 1027 * we return a positive number (its total switch count). If a second call 1028 * a short while later returns the same number, the caller can be sure that 1029 * @p has remained unscheduled the whole time. 1030 * 1031 * The caller must ensure that the task *will* unschedule sometime soon, 1032 * else this function might spin for a *long* time. This function can't 1033 * be called with interrupts off, or it may introduce deadlock with 1034 * smp_call_function() if an IPI is sent by the same process we are 1035 * waiting to become inactive. 1036 */ 1037unsigned long wait_task_inactive(struct task_struct *p, long match_state) 1038{ 1039 unsigned long flags; 1040 int running, on_rq; 1041 unsigned long ncsw; 1042 struct rq *rq; 1043 1044 for (;;) { 1045 /* 1046 * We do the initial early heuristics without holding 1047 * any task-queue locks at all. We'll only try to get 1048 * the runqueue lock when things look like they will 1049 * work out! 1050 */ 1051 rq = task_rq(p); 1052 1053 /* 1054 * If the task is actively running on another CPU 1055 * still, just relax and busy-wait without holding 1056 * any locks. 1057 * 1058 * NOTE! Since we don't hold any locks, it's not 1059 * even sure that "rq" stays as the right runqueue! 1060 * But we don't care, since "task_running()" will 1061 * return false if the runqueue has changed and p 1062 * is actually now running somewhere else! 1063 */ 1064 while (task_running(rq, p)) { 1065 if (match_state && unlikely(p->state != match_state)) 1066 return 0; 1067 cpu_relax(); 1068 } 1069 1070 /* 1071 * Ok, time to look more closely! We need the rq 1072 * lock now, to be *sure*. If we're wrong, we'll 1073 * just go back and repeat. 1074 */ 1075 rq = task_rq_lock(p, &flags); 1076 trace_sched_wait_task(p); 1077 running = task_running(rq, p); 1078 on_rq = p->on_rq; 1079 ncsw = 0; 1080 if (!match_state || p->state == match_state) 1081 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */ 1082 task_rq_unlock(rq, p, &flags); 1083 1084 /* 1085 * If it changed from the expected state, bail out now. 1086 */ 1087 if (unlikely(!ncsw)) 1088 break; 1089 1090 /* 1091 * Was it really running after all now that we 1092 * checked with the proper locks actually held? 1093 * 1094 * Oops. Go back and try again.. 1095 */ 1096 if (unlikely(running)) { 1097 cpu_relax(); 1098 continue; 1099 } 1100 1101 /* 1102 * It's not enough that it's not actively running, 1103 * it must be off the runqueue _entirely_, and not 1104 * preempted! 1105 * 1106 * So if it was still runnable (but just not actively 1107 * running right now), it's preempted, and we should 1108 * yield - it could be a while. 1109 */ 1110 if (unlikely(on_rq)) { 1111 ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ); 1112 1113 set_current_state(TASK_UNINTERRUPTIBLE); 1114 schedule_hrtimeout(&to, HRTIMER_MODE_REL); 1115 continue; 1116 } 1117 1118 /* 1119 * Ahh, all good. It wasn't running, and it wasn't 1120 * runnable, which means that it will never become 1121 * running in the future either. We're all done! 1122 */ 1123 break; 1124 } 1125 1126 return ncsw; 1127} 1128 1129/*** 1130 * kick_process - kick a running thread to enter/exit the kernel 1131 * @p: the to-be-kicked thread 1132 * 1133 * Cause a process which is running on another CPU to enter 1134 * kernel-mode, without any delay. (to get signals handled.) 1135 * 1136 * NOTE: this function doesn't have to take the runqueue lock, 1137 * because all it wants to ensure is that the remote task enters 1138 * the kernel. If the IPI races and the task has been migrated 1139 * to another CPU then no harm is done and the purpose has been 1140 * achieved as well. 1141 */ 1142void kick_process(struct task_struct *p) 1143{ 1144 int cpu; 1145 1146 preempt_disable(); 1147 cpu = task_cpu(p); 1148 if ((cpu != smp_processor_id()) && task_curr(p)) 1149 smp_send_reschedule(cpu); 1150 preempt_enable(); 1151} 1152EXPORT_SYMBOL_GPL(kick_process); 1153#endif /* CONFIG_SMP */ 1154 1155#ifdef CONFIG_SMP 1156/* 1157 * ->cpus_allowed is protected by both rq->lock and p->pi_lock 1158 */ 1159static int select_fallback_rq(int cpu, struct task_struct *p) 1160{ 1161 int nid = cpu_to_node(cpu); 1162 const struct cpumask *nodemask = NULL; 1163 enum { cpuset, possible, fail } state = cpuset; 1164 int dest_cpu; 1165 1166 /* 1167 * If the node that the cpu is on has been offlined, cpu_to_node() 1168 * will return -1. There is no cpu on the node, and we should 1169 * select the cpu on the other node. 1170 */ 1171 if (nid != -1) { 1172 nodemask = cpumask_of_node(nid); 1173 1174 /* Look for allowed, online CPU in same node. */ 1175 for_each_cpu(dest_cpu, nodemask) { 1176 if (!cpu_online(dest_cpu)) 1177 continue; 1178 if (!cpu_active(dest_cpu)) 1179 continue; 1180 if (cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p))) 1181 return dest_cpu; 1182 } 1183 } 1184 1185 for (;;) { 1186 /* Any allowed, online CPU? */ 1187 for_each_cpu(dest_cpu, tsk_cpus_allowed(p)) { 1188 if (!cpu_online(dest_cpu)) 1189 continue; 1190 if (!cpu_active(dest_cpu)) 1191 continue; 1192 goto out; 1193 } 1194 1195 switch (state) { 1196 case cpuset: 1197 /* No more Mr. Nice Guy. */ 1198 cpuset_cpus_allowed_fallback(p); 1199 state = possible; 1200 break; 1201 1202 case possible: 1203 do_set_cpus_allowed(p, cpu_possible_mask); 1204 state = fail; 1205 break; 1206 1207 case fail: 1208 BUG(); 1209 break; 1210 } 1211 } 1212 1213out: 1214 if (state != cpuset) { 1215 /* 1216 * Don't tell them about moving exiting tasks or 1217 * kernel threads (both mm NULL), since they never 1218 * leave kernel. 1219 */ 1220 if (p->mm && printk_ratelimit()) { 1221 printk_sched("process %d (%s) no longer affine to cpu%d\n", 1222 task_pid_nr(p), p->comm, cpu); 1223 } 1224 } 1225 1226 return dest_cpu; 1227} 1228 1229/* 1230 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable. 1231 */ 1232static inline 1233int select_task_rq(struct task_struct *p, int sd_flags, int wake_flags) 1234{ 1235 int cpu = p->sched_class->select_task_rq(p, sd_flags, wake_flags); 1236 1237 /* 1238 * In order not to call set_task_cpu() on a blocking task we need 1239 * to rely on ttwu() to place the task on a valid ->cpus_allowed 1240 * cpu. 1241 * 1242 * Since this is common to all placement strategies, this lives here. 1243 * 1244 * [ this allows ->select_task() to simply return task_cpu(p) and 1245 * not worry about this generic constraint ] 1246 */ 1247 if (unlikely(!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)) || 1248 !cpu_online(cpu))) 1249 cpu = select_fallback_rq(task_cpu(p), p); 1250 1251 return cpu; 1252} 1253 1254static void update_avg(u64 *avg, u64 sample) 1255{ 1256 s64 diff = sample - *avg; 1257 *avg += diff >> 3; 1258} 1259#endif 1260 1261static void 1262ttwu_stat(struct task_struct *p, int cpu, int wake_flags) 1263{ 1264#ifdef CONFIG_SCHEDSTATS 1265 struct rq *rq = this_rq(); 1266 1267#ifdef CONFIG_SMP 1268 int this_cpu = smp_processor_id(); 1269 1270 if (cpu == this_cpu) { 1271 schedstat_inc(rq, ttwu_local); 1272 schedstat_inc(p, se.statistics.nr_wakeups_local); 1273 } else { 1274 struct sched_domain *sd; 1275 1276 schedstat_inc(p, se.statistics.nr_wakeups_remote); 1277 rcu_read_lock(); 1278 for_each_domain(this_cpu, sd) { 1279 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) { 1280 schedstat_inc(sd, ttwu_wake_remote); 1281 break; 1282 } 1283 } 1284 rcu_read_unlock(); 1285 } 1286 1287 if (wake_flags & WF_MIGRATED) 1288 schedstat_inc(p, se.statistics.nr_wakeups_migrate); 1289 1290#endif /* CONFIG_SMP */ 1291 1292 schedstat_inc(rq, ttwu_count); 1293 schedstat_inc(p, se.statistics.nr_wakeups); 1294 1295 if (wake_flags & WF_SYNC) 1296 schedstat_inc(p, se.statistics.nr_wakeups_sync); 1297 1298#endif /* CONFIG_SCHEDSTATS */ 1299} 1300 1301static void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags) 1302{ 1303 activate_task(rq, p, en_flags); 1304 p->on_rq = 1; 1305 1306 /* if a worker is waking up, notify workqueue */ 1307 if (p->flags & PF_WQ_WORKER) 1308 wq_worker_waking_up(p, cpu_of(rq)); 1309} 1310 1311/* 1312 * Mark the task runnable and perform wakeup-preemption. 1313 */ 1314static void 1315ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags) 1316{ 1317 check_preempt_curr(rq, p, wake_flags); 1318 trace_sched_wakeup(p, true); 1319 1320 p->state = TASK_RUNNING; 1321#ifdef CONFIG_SMP 1322 if (p->sched_class->task_woken) 1323 p->sched_class->task_woken(rq, p); 1324 1325 if (rq->idle_stamp) { 1326 u64 delta = rq_clock(rq) - rq->idle_stamp; 1327 u64 max = 2*rq->max_idle_balance_cost; 1328 1329 update_avg(&rq->avg_idle, delta); 1330 1331 if (rq->avg_idle > max) 1332 rq->avg_idle = max; 1333 1334 rq->idle_stamp = 0; 1335 } 1336#endif 1337} 1338 1339static void 1340ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags) 1341{ 1342#ifdef CONFIG_SMP 1343 if (p->sched_contributes_to_load) 1344 rq->nr_uninterruptible--; 1345#endif 1346 1347 ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_WAKING); 1348 ttwu_do_wakeup(rq, p, wake_flags); 1349} 1350 1351/* 1352 * Called in case the task @p isn't fully descheduled from its runqueue, 1353 * in this case we must do a remote wakeup. Its a 'light' wakeup though, 1354 * since all we need to do is flip p->state to TASK_RUNNING, since 1355 * the task is still ->on_rq. 1356 */ 1357static int ttwu_remote(struct task_struct *p, int wake_flags) 1358{ 1359 struct rq *rq; 1360 int ret = 0; 1361 1362 rq = __task_rq_lock(p); 1363 if (p->on_rq) { 1364 /* check_preempt_curr() may use rq clock */ 1365 update_rq_clock(rq); 1366 ttwu_do_wakeup(rq, p, wake_flags); 1367 ret = 1; 1368 } 1369 __task_rq_unlock(rq); 1370 1371 return ret; 1372} 1373 1374#ifdef CONFIG_SMP 1375static void sched_ttwu_pending(void) 1376{ 1377 struct rq *rq = this_rq(); 1378 struct llist_node *llist = llist_del_all(&rq->wake_list); 1379 struct task_struct *p; 1380 1381 raw_spin_lock(&rq->lock); 1382 1383 while (llist) { 1384 p = llist_entry(llist, struct task_struct, wake_entry); 1385 llist = llist_next(llist); 1386 ttwu_do_activate(rq, p, 0); 1387 } 1388 1389 raw_spin_unlock(&rq->lock); 1390} 1391 1392void scheduler_ipi(void) 1393{ 1394 if (llist_empty(&this_rq()->wake_list) 1395 && !tick_nohz_full_cpu(smp_processor_id()) 1396 && !got_nohz_idle_kick()) 1397 return; 1398 1399 /* 1400 * Not all reschedule IPI handlers call irq_enter/irq_exit, since 1401 * traditionally all their work was done from the interrupt return 1402 * path. Now that we actually do some work, we need to make sure 1403 * we do call them. 1404 * 1405 * Some archs already do call them, luckily irq_enter/exit nest 1406 * properly. 1407 * 1408 * Arguably we should visit all archs and update all handlers, 1409 * however a fair share of IPIs are still resched only so this would 1410 * somewhat pessimize the simple resched case. 1411 */ 1412 irq_enter(); 1413 tick_nohz_full_check(); 1414 sched_ttwu_pending(); 1415 1416 /* 1417 * Check if someone kicked us for doing the nohz idle load balance. 1418 */ 1419 if (unlikely(got_nohz_idle_kick())) { 1420 this_rq()->idle_balance = 1; 1421 raise_softirq_irqoff(SCHED_SOFTIRQ); 1422 } 1423 irq_exit(); 1424} 1425 1426static void ttwu_queue_remote(struct task_struct *p, int cpu) 1427{ 1428 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list)) 1429 smp_send_reschedule(cpu); 1430} 1431 1432bool cpus_share_cache(int this_cpu, int that_cpu) 1433{ 1434 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu); 1435} 1436#endif /* CONFIG_SMP */ 1437 1438static void ttwu_queue(struct task_struct *p, int cpu) 1439{ 1440 struct rq *rq = cpu_rq(cpu); 1441 1442#if defined(CONFIG_SMP) 1443 if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) { 1444 sched_clock_cpu(cpu); /* sync clocks x-cpu */ 1445 ttwu_queue_remote(p, cpu); 1446 return; 1447 } 1448#endif 1449 1450 raw_spin_lock(&rq->lock); 1451 ttwu_do_activate(rq, p, 0); 1452 raw_spin_unlock(&rq->lock); 1453} 1454 1455/** 1456 * try_to_wake_up - wake up a thread 1457 * @p: the thread to be awakened 1458 * @state: the mask of task states that can be woken 1459 * @wake_flags: wake modifier flags (WF_*) 1460 * 1461 * Put it on the run-queue if it's not already there. The "current" 1462 * thread is always on the run-queue (except when the actual 1463 * re-schedule is in progress), and as such you're allowed to do 1464 * the simpler "current->state = TASK_RUNNING" to mark yourself 1465 * runnable without the overhead of this. 1466 * 1467 * Return: %true if @p was woken up, %false if it was already running. 1468 * or @state didn't match @p's state. 1469 */ 1470static int 1471try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags) 1472{ 1473 unsigned long flags; 1474 int cpu, success = 0; 1475 1476 /* 1477 * If we are going to wake up a thread waiting for CONDITION we 1478 * need to ensure that CONDITION=1 done by the caller can not be 1479 * reordered with p->state check below. This pairs with mb() in 1480 * set_current_state() the waiting thread does. 1481 */ 1482 smp_mb__before_spinlock(); 1483 raw_spin_lock_irqsave(&p->pi_lock, flags); 1484 if (!(p->state & state)) 1485 goto out; 1486 1487 success = 1; /* we're going to change ->state */ 1488 cpu = task_cpu(p); 1489 1490 if (p->on_rq && ttwu_remote(p, wake_flags)) 1491 goto stat; 1492 1493#ifdef CONFIG_SMP 1494 /* 1495 * If the owning (remote) cpu is still in the middle of schedule() with 1496 * this task as prev, wait until its done referencing the task. 1497 */ 1498 while (p->on_cpu) 1499 cpu_relax(); 1500 /* 1501 * Pairs with the smp_wmb() in finish_lock_switch(). 1502 */ 1503 smp_rmb(); 1504 1505 p->sched_contributes_to_load = !!task_contributes_to_load(p); 1506 p->state = TASK_WAKING; 1507 1508 if (p->sched_class->task_waking) 1509 p->sched_class->task_waking(p); 1510 1511 cpu = select_task_rq(p, SD_BALANCE_WAKE, wake_flags); 1512 if (task_cpu(p) != cpu) { 1513 wake_flags |= WF_MIGRATED; 1514 set_task_cpu(p, cpu); 1515 } 1516#endif /* CONFIG_SMP */ 1517 1518 ttwu_queue(p, cpu); 1519stat: 1520 ttwu_stat(p, cpu, wake_flags); 1521out: 1522 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 1523 1524 return success; 1525} 1526 1527/** 1528 * try_to_wake_up_local - try to wake up a local task with rq lock held 1529 * @p: the thread to be awakened 1530 * 1531 * Put @p on the run-queue if it's not already there. The caller must 1532 * ensure that this_rq() is locked, @p is bound to this_rq() and not 1533 * the current task. 1534 */ 1535static void try_to_wake_up_local(struct task_struct *p) 1536{ 1537 struct rq *rq = task_rq(p); 1538 1539 if (WARN_ON_ONCE(rq != this_rq()) || 1540 WARN_ON_ONCE(p == current)) 1541 return; 1542 1543 lockdep_assert_held(&rq->lock); 1544 1545 if (!raw_spin_trylock(&p->pi_lock)) { 1546 raw_spin_unlock(&rq->lock); 1547 raw_spin_lock(&p->pi_lock); 1548 raw_spin_lock(&rq->lock); 1549 } 1550 1551 if (!(p->state & TASK_NORMAL)) 1552 goto out; 1553 1554 if (!p->on_rq) 1555 ttwu_activate(rq, p, ENQUEUE_WAKEUP); 1556 1557 ttwu_do_wakeup(rq, p, 0); 1558 ttwu_stat(p, smp_processor_id(), 0); 1559out: 1560 raw_spin_unlock(&p->pi_lock); 1561} 1562 1563/** 1564 * wake_up_process - Wake up a specific process 1565 * @p: The process to be woken up. 1566 * 1567 * Attempt to wake up the nominated process and move it to the set of runnable 1568 * processes. 1569 * 1570 * Return: 1 if the process was woken up, 0 if it was already running. 1571 * 1572 * It may be assumed that this function implies a write memory barrier before 1573 * changing the task state if and only if any tasks are woken up. 1574 */ 1575int wake_up_process(struct task_struct *p) 1576{ 1577 WARN_ON(task_is_stopped_or_traced(p)); 1578 return try_to_wake_up(p, TASK_NORMAL, 0); 1579} 1580EXPORT_SYMBOL(wake_up_process); 1581 1582int wake_up_state(struct task_struct *p, unsigned int state) 1583{ 1584 return try_to_wake_up(p, state, 0); 1585} 1586 1587/* 1588 * Perform scheduler related setup for a newly forked process p. 1589 * p is forked by current. 1590 * 1591 * __sched_fork() is basic setup used by init_idle() too: 1592 */ 1593static void __sched_fork(struct task_struct *p) 1594{ 1595 p->on_rq = 0; 1596 1597 p->se.on_rq = 0; 1598 p->se.exec_start = 0; 1599 p->se.sum_exec_runtime = 0; 1600 p->se.prev_sum_exec_runtime = 0; 1601 p->se.nr_migrations = 0; 1602 p->se.vruntime = 0; 1603 INIT_LIST_HEAD(&p->se.group_node); 1604 1605#ifdef CONFIG_SCHEDSTATS 1606 memset(&p->se.statistics, 0, sizeof(p->se.statistics)); 1607#endif 1608 1609 INIT_LIST_HEAD(&p->rt.run_list); 1610 1611#ifdef CONFIG_PREEMPT_NOTIFIERS 1612 INIT_HLIST_HEAD(&p->preempt_notifiers); 1613#endif 1614 1615#ifdef CONFIG_NUMA_BALANCING 1616 if (p->mm && atomic_read(&p->mm->mm_users) == 1) { 1617 p->mm->numa_next_scan = jiffies; 1618 p->mm->numa_next_reset = jiffies; 1619 p->mm->numa_scan_seq = 0; 1620 } 1621 1622 p->node_stamp = 0ULL; 1623 p->numa_scan_seq = p->mm ? p->mm->numa_scan_seq : 0; 1624 p->numa_migrate_seq = p->mm ? p->mm->numa_scan_seq - 1 : 0; 1625 p->numa_scan_period = sysctl_numa_balancing_scan_delay; 1626 p->numa_work.next = &p->numa_work; 1627#endif /* CONFIG_NUMA_BALANCING */ 1628} 1629 1630#ifdef CONFIG_NUMA_BALANCING 1631#ifdef CONFIG_SCHED_DEBUG 1632void set_numabalancing_state(bool enabled) 1633{ 1634 if (enabled) 1635 sched_feat_set("NUMA"); 1636 else 1637 sched_feat_set("NO_NUMA"); 1638} 1639#else 1640__read_mostly bool numabalancing_enabled; 1641 1642void set_numabalancing_state(bool enabled) 1643{ 1644 numabalancing_enabled = enabled; 1645} 1646#endif /* CONFIG_SCHED_DEBUG */ 1647#endif /* CONFIG_NUMA_BALANCING */ 1648 1649/* 1650 * fork()/clone()-time setup: 1651 */ 1652void sched_fork(struct task_struct *p) 1653{ 1654 unsigned long flags; 1655 int cpu = get_cpu(); 1656 1657 __sched_fork(p); 1658 /* 1659 * We mark the process as running here. This guarantees that 1660 * nobody will actually run it, and a signal or other external 1661 * event cannot wake it up and insert it on the runqueue either. 1662 */ 1663 p->state = TASK_RUNNING; 1664 1665 /* 1666 * Make sure we do not leak PI boosting priority to the child. 1667 */ 1668 p->prio = current->normal_prio; 1669 1670 /* 1671 * Revert to default priority/policy on fork if requested. 1672 */ 1673 if (unlikely(p->sched_reset_on_fork)) { 1674 if (task_has_rt_policy(p)) { 1675 p->policy = SCHED_NORMAL; 1676 p->static_prio = NICE_TO_PRIO(0); 1677 p->rt_priority = 0; 1678 } else if (PRIO_TO_NICE(p->static_prio) < 0) 1679 p->static_prio = NICE_TO_PRIO(0); 1680 1681 p->prio = p->normal_prio = __normal_prio(p); 1682 set_load_weight(p); 1683 1684 /* 1685 * We don't need the reset flag anymore after the fork. It has 1686 * fulfilled its duty: 1687 */ 1688 p->sched_reset_on_fork = 0; 1689 } 1690 1691 if (!rt_prio(p->prio)) 1692 p->sched_class = &fair_sched_class; 1693 1694 if (p->sched_class->task_fork) 1695 p->sched_class->task_fork(p); 1696 1697 /* 1698 * The child is not yet in the pid-hash so no cgroup attach races, 1699 * and the cgroup is pinned to this child due to cgroup_fork() 1700 * is ran before sched_fork(). 1701 * 1702 * Silence PROVE_RCU. 1703 */ 1704 raw_spin_lock_irqsave(&p->pi_lock, flags); 1705 set_task_cpu(p, cpu); 1706 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 1707 1708#if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT) 1709 if (likely(sched_info_on())) 1710 memset(&p->sched_info, 0, sizeof(p->sched_info)); 1711#endif 1712#if defined(CONFIG_SMP) 1713 p->on_cpu = 0; 1714#endif 1715#ifdef CONFIG_PREEMPT_COUNT 1716 /* Want to start with kernel preemption disabled. */ 1717 task_thread_info(p)->preempt_count = 1; 1718#endif 1719#ifdef CONFIG_SMP 1720 plist_node_init(&p->pushable_tasks, MAX_PRIO); 1721#endif 1722 1723 put_cpu(); 1724} 1725 1726/* 1727 * wake_up_new_task - wake up a newly created task for the first time. 1728 * 1729 * This function will do some initial scheduler statistics housekeeping 1730 * that must be done for every newly created context, then puts the task 1731 * on the runqueue and wakes it. 1732 */ 1733void wake_up_new_task(struct task_struct *p) 1734{ 1735 unsigned long flags; 1736 struct rq *rq; 1737 1738 raw_spin_lock_irqsave(&p->pi_lock, flags); 1739#ifdef CONFIG_SMP 1740 /* 1741 * Fork balancing, do it here and not earlier because: 1742 * - cpus_allowed can change in the fork path 1743 * - any previously selected cpu might disappear through hotplug 1744 */ 1745 set_task_cpu(p, select_task_rq(p, SD_BALANCE_FORK, 0)); 1746#endif 1747 1748 /* Initialize new task's runnable average */ 1749 init_task_runnable_average(p); 1750 rq = __task_rq_lock(p); 1751 activate_task(rq, p, 0); 1752 p->on_rq = 1; 1753 trace_sched_wakeup_new(p, true); 1754 check_preempt_curr(rq, p, WF_FORK); 1755#ifdef CONFIG_SMP 1756 if (p->sched_class->task_woken) 1757 p->sched_class->task_woken(rq, p); 1758#endif 1759 task_rq_unlock(rq, p, &flags); 1760} 1761 1762#ifdef CONFIG_PREEMPT_NOTIFIERS 1763 1764/** 1765 * preempt_notifier_register - tell me when current is being preempted & rescheduled 1766 * @notifier: notifier struct to register 1767 */ 1768void preempt_notifier_register(struct preempt_notifier *notifier) 1769{ 1770 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers); 1771} 1772EXPORT_SYMBOL_GPL(preempt_notifier_register); 1773 1774/** 1775 * preempt_notifier_unregister - no longer interested in preemption notifications 1776 * @notifier: notifier struct to unregister 1777 * 1778 * This is safe to call from within a preemption notifier. 1779 */ 1780void preempt_notifier_unregister(struct preempt_notifier *notifier) 1781{ 1782 hlist_del(¬ifier->link); 1783} 1784EXPORT_SYMBOL_GPL(preempt_notifier_unregister); 1785 1786static void fire_sched_in_preempt_notifiers(struct task_struct *curr) 1787{ 1788 struct preempt_notifier *notifier; 1789 1790 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link) 1791 notifier->ops->sched_in(notifier, raw_smp_processor_id()); 1792} 1793 1794static void 1795fire_sched_out_preempt_notifiers(struct task_struct *curr, 1796 struct task_struct *next) 1797{ 1798 struct preempt_notifier *notifier; 1799 1800 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link) 1801 notifier->ops->sched_out(notifier, next); 1802} 1803 1804#else /* !CONFIG_PREEMPT_NOTIFIERS */ 1805 1806static void fire_sched_in_preempt_notifiers(struct task_struct *curr) 1807{ 1808} 1809 1810static void 1811fire_sched_out_preempt_notifiers(struct task_struct *curr, 1812 struct task_struct *next) 1813{ 1814} 1815 1816#endif /* CONFIG_PREEMPT_NOTIFIERS */ 1817 1818/** 1819 * prepare_task_switch - prepare to switch tasks 1820 * @rq: the runqueue preparing to switch 1821 * @prev: the current task that is being switched out 1822 * @next: the task we are going to switch to. 1823 * 1824 * This is called with the rq lock held and interrupts off. It must 1825 * be paired with a subsequent finish_task_switch after the context 1826 * switch. 1827 * 1828 * prepare_task_switch sets up locking and calls architecture specific 1829 * hooks. 1830 */ 1831static inline void 1832prepare_task_switch(struct rq *rq, struct task_struct *prev, 1833 struct task_struct *next) 1834{ 1835 trace_sched_switch(prev, next); 1836 sched_info_switch(rq, prev, next); 1837 perf_event_task_sched_out(prev, next); 1838 fire_sched_out_preempt_notifiers(prev, next); 1839 prepare_lock_switch(rq, next); 1840 prepare_arch_switch(next); 1841} 1842 1843/** 1844 * finish_task_switch - clean up after a task-switch 1845 * @rq: runqueue associated with task-switch 1846 * @prev: the thread we just switched away from. 1847 * 1848 * finish_task_switch must be called after the context switch, paired 1849 * with a prepare_task_switch call before the context switch. 1850 * finish_task_switch will reconcile locking set up by prepare_task_switch, 1851 * and do any other architecture-specific cleanup actions. 1852 * 1853 * Note that we may have delayed dropping an mm in context_switch(). If 1854 * so, we finish that here outside of the runqueue lock. (Doing it 1855 * with the lock held can cause deadlocks; see schedule() for 1856 * details.) 1857 */ 1858static void finish_task_switch(struct rq *rq, struct task_struct *prev) 1859 __releases(rq->lock) 1860{ 1861 struct mm_struct *mm = rq->prev_mm; 1862 long prev_state; 1863 1864 rq->prev_mm = NULL; 1865 1866 /* 1867 * A task struct has one reference for the use as "current". 1868 * If a task dies, then it sets TASK_DEAD in tsk->state and calls 1869 * schedule one last time. The schedule call will never return, and 1870 * the scheduled task must drop that reference. 1871 * The test for TASK_DEAD must occur while the runqueue locks are 1872 * still held, otherwise prev could be scheduled on another cpu, die 1873 * there before we look at prev->state, and then the reference would 1874 * be dropped twice. 1875 * Manfred Spraul <manfred@colorfullife.com> 1876 */ 1877 prev_state = prev->state; 1878 vtime_task_switch(prev); 1879 finish_arch_switch(prev); 1880 perf_event_task_sched_in(prev, current); 1881 finish_lock_switch(rq, prev); 1882 finish_arch_post_lock_switch(); 1883 1884 fire_sched_in_preempt_notifiers(current); 1885 if (mm) 1886 mmdrop(mm); 1887 if (unlikely(prev_state == TASK_DEAD)) { 1888 /* 1889 * Remove function-return probe instances associated with this 1890 * task and put them back on the free list. 1891 */ 1892 kprobe_flush_task(prev); 1893 put_task_struct(prev); 1894 } 1895 1896 tick_nohz_task_switch(current); 1897} 1898 1899#ifdef CONFIG_SMP 1900 1901/* assumes rq->lock is held */ 1902static inline void pre_schedule(struct rq *rq, struct task_struct *prev) 1903{ 1904 if (prev->sched_class->pre_schedule) 1905 prev->sched_class->pre_schedule(rq, prev); 1906} 1907 1908/* rq->lock is NOT held, but preemption is disabled */ 1909static inline void post_schedule(struct rq *rq) 1910{ 1911 if (rq->post_schedule) { 1912 unsigned long flags; 1913 1914 raw_spin_lock_irqsave(&rq->lock, flags); 1915 if (rq->curr->sched_class->post_schedule) 1916 rq->curr->sched_class->post_schedule(rq); 1917 raw_spin_unlock_irqrestore(&rq->lock, flags); 1918 1919 rq->post_schedule = 0; 1920 } 1921} 1922 1923#else 1924 1925static inline void pre_schedule(struct rq *rq, struct task_struct *p) 1926{ 1927} 1928 1929static inline void post_schedule(struct rq *rq) 1930{ 1931} 1932 1933#endif 1934 1935/** 1936 * schedule_tail - first thing a freshly forked thread must call. 1937 * @prev: the thread we just switched away from. 1938 */ 1939asmlinkage void schedule_tail(struct task_struct *prev) 1940 __releases(rq->lock) 1941{ 1942 struct rq *rq = this_rq(); 1943 1944 finish_task_switch(rq, prev); 1945 1946 /* 1947 * FIXME: do we need to worry about rq being invalidated by the 1948 * task_switch? 1949 */ 1950 post_schedule(rq); 1951 1952#ifdef __ARCH_WANT_UNLOCKED_CTXSW 1953 /* In this case, finish_task_switch does not reenable preemption */ 1954 preempt_enable(); 1955#endif 1956 if (current->set_child_tid) 1957 put_user(task_pid_vnr(current), current->set_child_tid); 1958} 1959 1960/* 1961 * context_switch - switch to the new MM and the new 1962 * thread's register state. 1963 */ 1964static inline void 1965context_switch(struct rq *rq, struct task_struct *prev, 1966 struct task_struct *next) 1967{ 1968 struct mm_struct *mm, *oldmm; 1969 1970 prepare_task_switch(rq, prev, next); 1971 1972 mm = next->mm; 1973 oldmm = prev->active_mm; 1974 /* 1975 * For paravirt, this is coupled with an exit in switch_to to 1976 * combine the page table reload and the switch backend into 1977 * one hypercall. 1978 */ 1979 arch_start_context_switch(prev); 1980 1981 if (!mm) { 1982 next->active_mm = oldmm; 1983 atomic_inc(&oldmm->mm_count); 1984 enter_lazy_tlb(oldmm, next); 1985 } else 1986 switch_mm(oldmm, mm, next); 1987 1988 if (!prev->mm) { 1989 prev->active_mm = NULL; 1990 rq->prev_mm = oldmm; 1991 } 1992 /* 1993 * Since the runqueue lock will be released by the next 1994 * task (which is an invalid locking op but in the case 1995 * of the scheduler it's an obvious special-case), so we 1996 * do an early lockdep release here: 1997 */ 1998#ifndef __ARCH_WANT_UNLOCKED_CTXSW 1999 spin_release(&rq->lock.dep_map, 1, _THIS_IP_); 2000#endif 2001 2002 context_tracking_task_switch(prev, next); 2003 /* Here we just switch the register state and the stack. */ 2004 switch_to(prev, next, prev); 2005 2006 barrier(); 2007 /* 2008 * this_rq must be evaluated again because prev may have moved 2009 * CPUs since it called schedule(), thus the 'rq' on its stack 2010 * frame will be invalid. 2011 */ 2012 finish_task_switch(this_rq(), prev); 2013} 2014 2015/* 2016 * nr_running and nr_context_switches: 2017 * 2018 * externally visible scheduler statistics: current number of runnable 2019 * threads, total number of context switches performed since bootup. 2020 */ 2021unsigned long nr_running(void) 2022{ 2023 unsigned long i, sum = 0; 2024 2025 for_each_online_cpu(i) 2026 sum += cpu_rq(i)->nr_running; 2027 2028 return sum; 2029} 2030 2031unsigned long long nr_context_switches(void) 2032{ 2033 int i; 2034 unsigned long long sum = 0; 2035 2036 for_each_possible_cpu(i) 2037 sum += cpu_rq(i)->nr_switches; 2038 2039 return sum; 2040} 2041 2042unsigned long nr_iowait(void) 2043{ 2044 unsigned long i, sum = 0; 2045 2046 for_each_possible_cpu(i) 2047 sum += atomic_read(&cpu_rq(i)->nr_iowait); 2048 2049 return sum; 2050} 2051 2052unsigned long nr_iowait_cpu(int cpu) 2053{ 2054 struct rq *this = cpu_rq(cpu); 2055 return atomic_read(&this->nr_iowait); 2056} 2057 2058#ifdef CONFIG_SMP 2059 2060/* 2061 * sched_exec - execve() is a valuable balancing opportunity, because at 2062 * this point the task has the smallest effective memory and cache footprint. 2063 */ 2064void sched_exec(void) 2065{ 2066 struct task_struct *p = current; 2067 unsigned long flags; 2068 int dest_cpu; 2069 2070 raw_spin_lock_irqsave(&p->pi_lock, flags); 2071 dest_cpu = p->sched_class->select_task_rq(p, SD_BALANCE_EXEC, 0); 2072 if (dest_cpu == smp_processor_id()) 2073 goto unlock; 2074 2075 if (likely(cpu_active(dest_cpu))) { 2076 struct migration_arg arg = { p, dest_cpu }; 2077 2078 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 2079 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg); 2080 return; 2081 } 2082unlock: 2083 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 2084} 2085 2086#endif 2087 2088DEFINE_PER_CPU(struct kernel_stat, kstat); 2089DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat); 2090 2091EXPORT_PER_CPU_SYMBOL(kstat); 2092EXPORT_PER_CPU_SYMBOL(kernel_cpustat); 2093 2094/* 2095 * Return any ns on the sched_clock that have not yet been accounted in 2096 * @p in case that task is currently running. 2097 * 2098 * Called with task_rq_lock() held on @rq. 2099 */ 2100static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq) 2101{ 2102 u64 ns = 0; 2103 2104 if (task_current(rq, p)) { 2105 update_rq_clock(rq); 2106 ns = rq_clock_task(rq) - p->se.exec_start; 2107 if ((s64)ns < 0) 2108 ns = 0; 2109 } 2110 2111 return ns; 2112} 2113 2114unsigned long long task_delta_exec(struct task_struct *p) 2115{ 2116 unsigned long flags; 2117 struct rq *rq; 2118 u64 ns = 0; 2119 2120 rq = task_rq_lock(p, &flags); 2121 ns = do_task_delta_exec(p, rq); 2122 task_rq_unlock(rq, p, &flags); 2123 2124 return ns; 2125} 2126 2127/* 2128 * Return accounted runtime for the task. 2129 * In case the task is currently running, return the runtime plus current's 2130 * pending runtime that have not been accounted yet. 2131 */ 2132unsigned long long task_sched_runtime(struct task_struct *p) 2133{ 2134 unsigned long flags; 2135 struct rq *rq; 2136 u64 ns = 0; 2137 2138 rq = task_rq_lock(p, &flags); 2139 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq); 2140 task_rq_unlock(rq, p, &flags); 2141 2142 return ns; 2143} 2144 2145/* 2146 * This function gets called by the timer code, with HZ frequency. 2147 * We call it with interrupts disabled. 2148 */ 2149void scheduler_tick(void) 2150{ 2151 int cpu = smp_processor_id(); 2152 struct rq *rq = cpu_rq(cpu); 2153 struct task_struct *curr = rq->curr; 2154 2155 sched_clock_tick(); 2156 2157 raw_spin_lock(&rq->lock); 2158 update_rq_clock(rq); 2159 curr->sched_class->task_tick(rq, curr, 0); 2160 update_cpu_load_active(rq); 2161 raw_spin_unlock(&rq->lock); 2162 2163 perf_event_task_tick(); 2164 2165#ifdef CONFIG_SMP 2166 rq->idle_balance = idle_cpu(cpu); 2167 trigger_load_balance(rq, cpu); 2168#endif 2169 rq_last_tick_reset(rq); 2170} 2171 2172#ifdef CONFIG_NO_HZ_FULL 2173/** 2174 * scheduler_tick_max_deferment 2175 * 2176 * Keep at least one tick per second when a single 2177 * active task is running because the scheduler doesn't 2178 * yet completely support full dynticks environment. 2179 * 2180 * This makes sure that uptime, CFS vruntime, load 2181 * balancing, etc... continue to move forward, even 2182 * with a very low granularity. 2183 * 2184 * Return: Maximum deferment in nanoseconds. 2185 */ 2186u64 scheduler_tick_max_deferment(void) 2187{ 2188 struct rq *rq = this_rq(); 2189 unsigned long next, now = ACCESS_ONCE(jiffies); 2190 2191 next = rq->last_sched_tick + HZ; 2192 2193 if (time_before_eq(next, now)) 2194 return 0; 2195 2196 return jiffies_to_usecs(next - now) * NSEC_PER_USEC; 2197} 2198#endif 2199 2200notrace unsigned long get_parent_ip(unsigned long addr) 2201{ 2202 if (in_lock_functions(addr)) { 2203 addr = CALLER_ADDR2; 2204 if (in_lock_functions(addr)) 2205 addr = CALLER_ADDR3; 2206 } 2207 return addr; 2208} 2209 2210#if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \ 2211 defined(CONFIG_PREEMPT_TRACER)) 2212 2213void __kprobes add_preempt_count(int val) 2214{ 2215#ifdef CONFIG_DEBUG_PREEMPT 2216 /* 2217 * Underflow? 2218 */ 2219 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0))) 2220 return; 2221#endif 2222 add_preempt_count_notrace(val); 2223#ifdef CONFIG_DEBUG_PREEMPT 2224 /* 2225 * Spinlock count overflowing soon? 2226 */ 2227 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >= 2228 PREEMPT_MASK - 10); 2229#endif 2230 if (preempt_count() == val) 2231 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1)); 2232} 2233EXPORT_SYMBOL(add_preempt_count); 2234 2235void __kprobes sub_preempt_count(int val) 2236{ 2237#ifdef CONFIG_DEBUG_PREEMPT 2238 /* 2239 * Underflow? 2240 */ 2241 if (DEBUG_LOCKS_WARN_ON(val > preempt_count())) 2242 return; 2243 /* 2244 * Is the spinlock portion underflowing? 2245 */ 2246 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) && 2247 !(preempt_count() & PREEMPT_MASK))) 2248 return; 2249#endif 2250 2251 if (preempt_count() == val) 2252 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1)); 2253 sub_preempt_count_notrace(val); 2254} 2255EXPORT_SYMBOL(sub_preempt_count); 2256 2257#endif 2258 2259/* 2260 * Print scheduling while atomic bug: 2261 */ 2262static noinline void __schedule_bug(struct task_struct *prev) 2263{ 2264 if (oops_in_progress) 2265 return; 2266 2267 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n", 2268 prev->comm, prev->pid, preempt_count()); 2269 2270 debug_show_held_locks(prev); 2271 print_modules(); 2272 if (irqs_disabled()) 2273 print_irqtrace_events(prev); 2274 dump_stack(); 2275 add_taint(TAINT_WARN, LOCKDEP_STILL_OK); 2276} 2277 2278/* 2279 * Various schedule()-time debugging checks and statistics: 2280 */ 2281static inline void schedule_debug(struct task_struct *prev) 2282{ 2283 /* 2284 * Test if we are atomic. Since do_exit() needs to call into 2285 * schedule() atomically, we ignore that path for now. 2286 * Otherwise, whine if we are scheduling when we should not be. 2287 */ 2288 if (unlikely(in_atomic_preempt_off() && !prev->exit_state)) 2289 __schedule_bug(prev); 2290 rcu_sleep_check(); 2291 2292 profile_hit(SCHED_PROFILING, __builtin_return_address(0)); 2293 2294 schedstat_inc(this_rq(), sched_count); 2295} 2296 2297static void put_prev_task(struct rq *rq, struct task_struct *prev) 2298{ 2299 if (prev->on_rq || rq->skip_clock_update < 0) 2300 update_rq_clock(rq); 2301 prev->sched_class->put_prev_task(rq, prev); 2302} 2303 2304/* 2305 * Pick up the highest-prio task: 2306 */ 2307static inline struct task_struct * 2308pick_next_task(struct rq *rq) 2309{ 2310 const struct sched_class *class; 2311 struct task_struct *p; 2312 2313 /* 2314 * Optimization: we know that if all tasks are in 2315 * the fair class we can call that function directly: 2316 */ 2317 if (likely(rq->nr_running == rq->cfs.h_nr_running)) { 2318 p = fair_sched_class.pick_next_task(rq); 2319 if (likely(p)) 2320 return p; 2321 } 2322 2323 for_each_class(class) { 2324 p = class->pick_next_task(rq); 2325 if (p) 2326 return p; 2327 } 2328 2329 BUG(); /* the idle class will always have a runnable task */ 2330} 2331 2332/* 2333 * __schedule() is the main scheduler function. 2334 * 2335 * The main means of driving the scheduler and thus entering this function are: 2336 * 2337 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc. 2338 * 2339 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return 2340 * paths. For example, see arch/x86/entry_64.S. 2341 * 2342 * To drive preemption between tasks, the scheduler sets the flag in timer 2343 * interrupt handler scheduler_tick(). 2344 * 2345 * 3. Wakeups don't really cause entry into schedule(). They add a 2346 * task to the run-queue and that's it. 2347 * 2348 * Now, if the new task added to the run-queue preempts the current 2349 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets 2350 * called on the nearest possible occasion: 2351 * 2352 * - If the kernel is preemptible (CONFIG_PREEMPT=y): 2353 * 2354 * - in syscall or exception context, at the next outmost 2355 * preempt_enable(). (this might be as soon as the wake_up()'s 2356 * spin_unlock()!) 2357 * 2358 * - in IRQ context, return from interrupt-handler to 2359 * preemptible context 2360 * 2361 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set) 2362 * then at the next: 2363 * 2364 * - cond_resched() call 2365 * - explicit schedule() call 2366 * - return from syscall or exception to user-space 2367 * - return from interrupt-handler to user-space 2368 */ 2369static void __sched __schedule(void) 2370{ 2371 struct task_struct *prev, *next; 2372 unsigned long *switch_count; 2373 struct rq *rq; 2374 int cpu; 2375 2376need_resched: 2377 preempt_disable(); 2378 cpu = smp_processor_id(); 2379 rq = cpu_rq(cpu); 2380 rcu_note_context_switch(cpu); 2381 prev = rq->curr; 2382 2383 schedule_debug(prev); 2384 2385 if (sched_feat(HRTICK)) 2386 hrtick_clear(rq); 2387 2388 /* 2389 * Make sure that signal_pending_state()->signal_pending() below 2390 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE) 2391 * done by the caller to avoid the race with signal_wake_up(). 2392 */ 2393 smp_mb__before_spinlock(); 2394 raw_spin_lock_irq(&rq->lock); 2395 2396 switch_count = &prev->nivcsw; 2397 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) { 2398 if (unlikely(signal_pending_state(prev->state, prev))) { 2399 prev->state = TASK_RUNNING; 2400 } else { 2401 deactivate_task(rq, prev, DEQUEUE_SLEEP); 2402 prev->on_rq = 0; 2403 2404 /* 2405 * If a worker went to sleep, notify and ask workqueue 2406 * whether it wants to wake up a task to maintain 2407 * concurrency. 2408 */ 2409 if (prev->flags & PF_WQ_WORKER) { 2410 struct task_struct *to_wakeup; 2411 2412 to_wakeup = wq_worker_sleeping(prev, cpu); 2413 if (to_wakeup) 2414 try_to_wake_up_local(to_wakeup); 2415 } 2416 } 2417 switch_count = &prev->nvcsw; 2418 } 2419 2420 pre_schedule(rq, prev); 2421 2422 if (unlikely(!rq->nr_running)) 2423 idle_balance(cpu, rq); 2424 2425 put_prev_task(rq, prev); 2426 next = pick_next_task(rq); 2427 clear_tsk_need_resched(prev); 2428 rq->skip_clock_update = 0; 2429 2430 if (likely(prev != next)) { 2431 rq->nr_switches++; 2432 rq->curr = next; 2433 ++*switch_count; 2434 2435 context_switch(rq, prev, next); /* unlocks the rq */ 2436 /* 2437 * The context switch have flipped the stack from under us 2438 * and restored the local variables which were saved when 2439 * this task called schedule() in the past. prev == current 2440 * is still correct, but it can be moved to another cpu/rq. 2441 */ 2442 cpu = smp_processor_id(); 2443 rq = cpu_rq(cpu); 2444 } else 2445 raw_spin_unlock_irq(&rq->lock); 2446 2447 post_schedule(rq); 2448 2449 sched_preempt_enable_no_resched(); 2450 if (need_resched()) 2451 goto need_resched; 2452} 2453 2454static inline void sched_submit_work(struct task_struct *tsk) 2455{ 2456 if (!tsk->state || tsk_is_pi_blocked(tsk)) 2457 return; 2458 /* 2459 * If we are going to sleep and we have plugged IO queued, 2460 * make sure to submit it to avoid deadlocks. 2461 */ 2462 if (blk_needs_flush_plug(tsk)) 2463 blk_schedule_flush_plug(tsk); 2464} 2465 2466asmlinkage void __sched schedule(void) 2467{ 2468 struct task_struct *tsk = current; 2469 2470 sched_submit_work(tsk); 2471 __schedule(); 2472} 2473EXPORT_SYMBOL(schedule); 2474 2475#ifdef CONFIG_CONTEXT_TRACKING 2476asmlinkage void __sched schedule_user(void) 2477{ 2478 /* 2479 * If we come here after a random call to set_need_resched(), 2480 * or we have been woken up remotely but the IPI has not yet arrived, 2481 * we haven't yet exited the RCU idle mode. Do it here manually until 2482 * we find a better solution. 2483 */ 2484 user_exit(); 2485 schedule(); 2486 user_enter(); 2487} 2488#endif 2489 2490/** 2491 * schedule_preempt_disabled - called with preemption disabled 2492 * 2493 * Returns with preemption disabled. Note: preempt_count must be 1 2494 */ 2495void __sched schedule_preempt_disabled(void) 2496{ 2497 sched_preempt_enable_no_resched(); 2498 schedule(); 2499 preempt_disable(); 2500} 2501 2502#ifdef CONFIG_PREEMPT 2503/* 2504 * this is the entry point to schedule() from in-kernel preemption 2505 * off of preempt_enable. Kernel preemptions off return from interrupt 2506 * occur there and call schedule directly. 2507 */ 2508asmlinkage void __sched notrace preempt_schedule(void) 2509{ 2510 /* 2511 * If there is a non-zero preempt_count or interrupts are disabled, 2512 * we do not want to preempt the current task. Just return.. 2513 */ 2514 if (likely(!preemptible())) 2515 return; 2516 2517 do { 2518 add_preempt_count_notrace(PREEMPT_ACTIVE); 2519 __schedule(); 2520 sub_preempt_count_notrace(PREEMPT_ACTIVE); 2521 2522 /* 2523 * Check again in case we missed a preemption opportunity 2524 * between schedule and now. 2525 */ 2526 barrier(); 2527 } while (need_resched()); 2528} 2529EXPORT_SYMBOL(preempt_schedule); 2530 2531/* 2532 * this is the entry point to schedule() from kernel preemption 2533 * off of irq context. 2534 * Note, that this is called and return with irqs disabled. This will 2535 * protect us against recursive calling from irq. 2536 */ 2537asmlinkage void __sched preempt_schedule_irq(void) 2538{ 2539 struct thread_info *ti = current_thread_info(); 2540 enum ctx_state prev_state; 2541 2542 /* Catch callers which need to be fixed */ 2543 BUG_ON(ti->preempt_count || !irqs_disabled()); 2544 2545 prev_state = exception_enter(); 2546 2547 do { 2548 add_preempt_count(PREEMPT_ACTIVE); 2549 local_irq_enable(); 2550 __schedule(); 2551 local_irq_disable(); 2552 sub_preempt_count(PREEMPT_ACTIVE); 2553 2554 /* 2555 * Check again in case we missed a preemption opportunity 2556 * between schedule and now. 2557 */ 2558 barrier(); 2559 } while (need_resched()); 2560 2561 exception_exit(prev_state); 2562} 2563 2564#endif /* CONFIG_PREEMPT */ 2565 2566int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags, 2567 void *key) 2568{ 2569 return try_to_wake_up(curr->private, mode, wake_flags); 2570} 2571EXPORT_SYMBOL(default_wake_function); 2572 2573/* 2574 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just 2575 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve 2576 * number) then we wake all the non-exclusive tasks and one exclusive task. 2577 * 2578 * There are circumstances in which we can try to wake a task which has already 2579 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns 2580 * zero in this (rare) case, and we handle it by continuing to scan the queue. 2581 */ 2582static void __wake_up_common(wait_queue_head_t *q, unsigned int mode, 2583 int nr_exclusive, int wake_flags, void *key) 2584{ 2585 wait_queue_t *curr, *next; 2586 2587 list_for_each_entry_safe(curr, next, &q->task_list, task_list) { 2588 unsigned flags = curr->flags; 2589 2590 if (curr->func(curr, mode, wake_flags, key) && 2591 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive) 2592 break; 2593 } 2594} 2595 2596/** 2597 * __wake_up - wake up threads blocked on a waitqueue. 2598 * @q: the waitqueue 2599 * @mode: which threads 2600 * @nr_exclusive: how many wake-one or wake-many threads to wake up 2601 * @key: is directly passed to the wakeup function 2602 * 2603 * It may be assumed that this function implies a write memory barrier before 2604 * changing the task state if and only if any tasks are woken up. 2605 */ 2606void __wake_up(wait_queue_head_t *q, unsigned int mode, 2607 int nr_exclusive, void *key) 2608{ 2609 unsigned long flags; 2610 2611 spin_lock_irqsave(&q->lock, flags); 2612 __wake_up_common(q, mode, nr_exclusive, 0, key); 2613 spin_unlock_irqrestore(&q->lock, flags); 2614} 2615EXPORT_SYMBOL(__wake_up); 2616 2617/* 2618 * Same as __wake_up but called with the spinlock in wait_queue_head_t held. 2619 */ 2620void __wake_up_locked(wait_queue_head_t *q, unsigned int mode, int nr) 2621{ 2622 __wake_up_common(q, mode, nr, 0, NULL); 2623} 2624EXPORT_SYMBOL_GPL(__wake_up_locked); 2625 2626void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key) 2627{ 2628 __wake_up_common(q, mode, 1, 0, key); 2629} 2630EXPORT_SYMBOL_GPL(__wake_up_locked_key); 2631 2632/** 2633 * __wake_up_sync_key - wake up threads blocked on a waitqueue. 2634 * @q: the waitqueue 2635 * @mode: which threads 2636 * @nr_exclusive: how many wake-one or wake-many threads to wake up 2637 * @key: opaque value to be passed to wakeup targets 2638 * 2639 * The sync wakeup differs that the waker knows that it will schedule 2640 * away soon, so while the target thread will be woken up, it will not 2641 * be migrated to another CPU - ie. the two threads are 'synchronized' 2642 * with each other. This can prevent needless bouncing between CPUs. 2643 * 2644 * On UP it can prevent extra preemption. 2645 * 2646 * It may be assumed that this function implies a write memory barrier before 2647 * changing the task state if and only if any tasks are woken up. 2648 */ 2649void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode, 2650 int nr_exclusive, void *key) 2651{ 2652 unsigned long flags; 2653 int wake_flags = WF_SYNC; 2654 2655 if (unlikely(!q)) 2656 return; 2657 2658 if (unlikely(nr_exclusive != 1)) 2659 wake_flags = 0; 2660 2661 spin_lock_irqsave(&q->lock, flags); 2662 __wake_up_common(q, mode, nr_exclusive, wake_flags, key); 2663 spin_unlock_irqrestore(&q->lock, flags); 2664} 2665EXPORT_SYMBOL_GPL(__wake_up_sync_key); 2666 2667/* 2668 * __wake_up_sync - see __wake_up_sync_key() 2669 */ 2670void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive) 2671{ 2672 __wake_up_sync_key(q, mode, nr_exclusive, NULL); 2673} 2674EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */ 2675 2676/** 2677 * complete: - signals a single thread waiting on this completion 2678 * @x: holds the state of this particular completion 2679 * 2680 * This will wake up a single thread waiting on this completion. Threads will be 2681 * awakened in the same order in which they were queued. 2682 * 2683 * See also complete_all(), wait_for_completion() and related routines. 2684 * 2685 * It may be assumed that this function implies a write memory barrier before 2686 * changing the task state if and only if any tasks are woken up. 2687 */ 2688void complete(struct completion *x) 2689{ 2690 unsigned long flags; 2691 2692 spin_lock_irqsave(&x->wait.lock, flags); 2693 x->done++; 2694 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL); 2695 spin_unlock_irqrestore(&x->wait.lock, flags); 2696} 2697EXPORT_SYMBOL(complete); 2698 2699/** 2700 * complete_all: - signals all threads waiting on this completion 2701 * @x: holds the state of this particular completion 2702 * 2703 * This will wake up all threads waiting on this particular completion event. 2704 * 2705 * It may be assumed that this function implies a write memory barrier before 2706 * changing the task state if and only if any tasks are woken up. 2707 */ 2708void complete_all(struct completion *x) 2709{ 2710 unsigned long flags; 2711 2712 spin_lock_irqsave(&x->wait.lock, flags); 2713 x->done += UINT_MAX/2; 2714 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL); 2715 spin_unlock_irqrestore(&x->wait.lock, flags); 2716} 2717EXPORT_SYMBOL(complete_all); 2718 2719static inline long __sched 2720do_wait_for_common(struct completion *x, 2721 long (*action)(long), long timeout, int state) 2722{ 2723 if (!x->done) { 2724 DECLARE_WAITQUEUE(wait, current); 2725 2726 __add_wait_queue_tail_exclusive(&x->wait, &wait); 2727 do { 2728 if (signal_pending_state(state, current)) { 2729 timeout = -ERESTARTSYS; 2730 break; 2731 } 2732 __set_current_state(state); 2733 spin_unlock_irq(&x->wait.lock); 2734 timeout = action(timeout); 2735 spin_lock_irq(&x->wait.lock); 2736 } while (!x->done && timeout); 2737 __remove_wait_queue(&x->wait, &wait); 2738 if (!x->done) 2739 return timeout; 2740 } 2741 x->done--; 2742 return timeout ?: 1; 2743} 2744 2745static inline long __sched 2746__wait_for_common(struct completion *x, 2747 long (*action)(long), long timeout, int state) 2748{ 2749 might_sleep(); 2750 2751 spin_lock_irq(&x->wait.lock); 2752 timeout = do_wait_for_common(x, action, timeout, state); 2753 spin_unlock_irq(&x->wait.lock); 2754 return timeout; 2755} 2756 2757static long __sched 2758wait_for_common(struct completion *x, long timeout, int state) 2759{ 2760 return __wait_for_common(x, schedule_timeout, timeout, state); 2761} 2762 2763static long __sched 2764wait_for_common_io(struct completion *x, long timeout, int state) 2765{ 2766 return __wait_for_common(x, io_schedule_timeout, timeout, state); 2767} 2768 2769/** 2770 * wait_for_completion: - waits for completion of a task 2771 * @x: holds the state of this particular completion 2772 * 2773 * This waits to be signaled for completion of a specific task. It is NOT 2774 * interruptible and there is no timeout. 2775 * 2776 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout 2777 * and interrupt capability. Also see complete(). 2778 */ 2779void __sched wait_for_completion(struct completion *x) 2780{ 2781 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE); 2782} 2783EXPORT_SYMBOL(wait_for_completion); 2784 2785/** 2786 * wait_for_completion_timeout: - waits for completion of a task (w/timeout) 2787 * @x: holds the state of this particular completion 2788 * @timeout: timeout value in jiffies 2789 * 2790 * This waits for either a completion of a specific task to be signaled or for a 2791 * specified timeout to expire. The timeout is in jiffies. It is not 2792 * interruptible. 2793 * 2794 * Return: 0 if timed out, and positive (at least 1, or number of jiffies left 2795 * till timeout) if completed. 2796 */ 2797unsigned long __sched 2798wait_for_completion_timeout(struct completion *x, unsigned long timeout) 2799{ 2800 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE); 2801} 2802EXPORT_SYMBOL(wait_for_completion_timeout); 2803 2804/** 2805 * wait_for_completion_io: - waits for completion of a task 2806 * @x: holds the state of this particular completion 2807 * 2808 * This waits to be signaled for completion of a specific task. It is NOT 2809 * interruptible and there is no timeout. The caller is accounted as waiting 2810 * for IO. 2811 */ 2812void __sched wait_for_completion_io(struct completion *x) 2813{ 2814 wait_for_common_io(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE); 2815} 2816EXPORT_SYMBOL(wait_for_completion_io); 2817 2818/** 2819 * wait_for_completion_io_timeout: - waits for completion of a task (w/timeout) 2820 * @x: holds the state of this particular completion 2821 * @timeout: timeout value in jiffies 2822 * 2823 * This waits for either a completion of a specific task to be signaled or for a 2824 * specified timeout to expire. The timeout is in jiffies. It is not 2825 * interruptible. The caller is accounted as waiting for IO. 2826 * 2827 * Return: 0 if timed out, and positive (at least 1, or number of jiffies left 2828 * till timeout) if completed. 2829 */ 2830unsigned long __sched 2831wait_for_completion_io_timeout(struct completion *x, unsigned long timeout) 2832{ 2833 return wait_for_common_io(x, timeout, TASK_UNINTERRUPTIBLE); 2834} 2835EXPORT_SYMBOL(wait_for_completion_io_timeout); 2836 2837/** 2838 * wait_for_completion_interruptible: - waits for completion of a task (w/intr) 2839 * @x: holds the state of this particular completion 2840 * 2841 * This waits for completion of a specific task to be signaled. It is 2842 * interruptible. 2843 * 2844 * Return: -ERESTARTSYS if interrupted, 0 if completed. 2845 */ 2846int __sched wait_for_completion_interruptible(struct completion *x) 2847{ 2848 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE); 2849 if (t == -ERESTARTSYS) 2850 return t; 2851 return 0; 2852} 2853EXPORT_SYMBOL(wait_for_completion_interruptible); 2854 2855/** 2856 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr)) 2857 * @x: holds the state of this particular completion 2858 * @timeout: timeout value in jiffies 2859 * 2860 * This waits for either a completion of a specific task to be signaled or for a 2861 * specified timeout to expire. It is interruptible. The timeout is in jiffies. 2862 * 2863 * Return: -ERESTARTSYS if interrupted, 0 if timed out, positive (at least 1, 2864 * or number of jiffies left till timeout) if completed. 2865 */ 2866long __sched 2867wait_for_completion_interruptible_timeout(struct completion *x, 2868 unsigned long timeout) 2869{ 2870 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE); 2871} 2872EXPORT_SYMBOL(wait_for_completion_interruptible_timeout); 2873 2874/** 2875 * wait_for_completion_killable: - waits for completion of a task (killable) 2876 * @x: holds the state of this particular completion 2877 * 2878 * This waits to be signaled for completion of a specific task. It can be 2879 * interrupted by a kill signal. 2880 * 2881 * Return: -ERESTARTSYS if interrupted, 0 if completed. 2882 */ 2883int __sched wait_for_completion_killable(struct completion *x) 2884{ 2885 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE); 2886 if (t == -ERESTARTSYS) 2887 return t; 2888 return 0; 2889} 2890EXPORT_SYMBOL(wait_for_completion_killable); 2891 2892/** 2893 * wait_for_completion_killable_timeout: - waits for completion of a task (w/(to,killable)) 2894 * @x: holds the state of this particular completion 2895 * @timeout: timeout value in jiffies 2896 * 2897 * This waits for either a completion of a specific task to be 2898 * signaled or for a specified timeout to expire. It can be 2899 * interrupted by a kill signal. The timeout is in jiffies. 2900 * 2901 * Return: -ERESTARTSYS if interrupted, 0 if timed out, positive (at least 1, 2902 * or number of jiffies left till timeout) if completed. 2903 */ 2904long __sched 2905wait_for_completion_killable_timeout(struct completion *x, 2906 unsigned long timeout) 2907{ 2908 return wait_for_common(x, timeout, TASK_KILLABLE); 2909} 2910EXPORT_SYMBOL(wait_for_completion_killable_timeout); 2911 2912/** 2913 * try_wait_for_completion - try to decrement a completion without blocking 2914 * @x: completion structure 2915 * 2916 * Return: 0 if a decrement cannot be done without blocking 2917 * 1 if a decrement succeeded. 2918 * 2919 * If a completion is being used as a counting completion, 2920 * attempt to decrement the counter without blocking. This 2921 * enables us to avoid waiting if the resource the completion 2922 * is protecting is not available. 2923 */ 2924bool try_wait_for_completion(struct completion *x) 2925{ 2926 unsigned long flags; 2927 int ret = 1; 2928 2929 spin_lock_irqsave(&x->wait.lock, flags); 2930 if (!x->done) 2931 ret = 0; 2932 else 2933 x->done--; 2934 spin_unlock_irqrestore(&x->wait.lock, flags); 2935 return ret; 2936} 2937EXPORT_SYMBOL(try_wait_for_completion); 2938 2939/** 2940 * completion_done - Test to see if a completion has any waiters 2941 * @x: completion structure 2942 * 2943 * Return: 0 if there are waiters (wait_for_completion() in progress) 2944 * 1 if there are no waiters. 2945 * 2946 */ 2947bool completion_done(struct completion *x) 2948{ 2949 unsigned long flags; 2950 int ret = 1; 2951 2952 spin_lock_irqsave(&x->wait.lock, flags); 2953 if (!x->done) 2954 ret = 0; 2955 spin_unlock_irqrestore(&x->wait.lock, flags); 2956 return ret; 2957} 2958EXPORT_SYMBOL(completion_done); 2959 2960static long __sched 2961sleep_on_common(wait_queue_head_t *q, int state, long timeout) 2962{ 2963 unsigned long flags; 2964 wait_queue_t wait; 2965 2966 init_waitqueue_entry(&wait, current); 2967 2968 __set_current_state(state); 2969 2970 spin_lock_irqsave(&q->lock, flags); 2971 __add_wait_queue(q, &wait); 2972 spin_unlock(&q->lock); 2973 timeout = schedule_timeout(timeout); 2974 spin_lock_irq(&q->lock); 2975 __remove_wait_queue(q, &wait); 2976 spin_unlock_irqrestore(&q->lock, flags); 2977 2978 return timeout; 2979} 2980 2981void __sched interruptible_sleep_on(wait_queue_head_t *q) 2982{ 2983 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT); 2984} 2985EXPORT_SYMBOL(interruptible_sleep_on); 2986 2987long __sched 2988interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout) 2989{ 2990 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout); 2991} 2992EXPORT_SYMBOL(interruptible_sleep_on_timeout); 2993 2994void __sched sleep_on(wait_queue_head_t *q) 2995{ 2996 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT); 2997} 2998EXPORT_SYMBOL(sleep_on); 2999 3000long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout) 3001{ 3002 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout); 3003} 3004EXPORT_SYMBOL(sleep_on_timeout); 3005 3006#ifdef CONFIG_RT_MUTEXES 3007 3008/* 3009 * rt_mutex_setprio - set the current priority of a task 3010 * @p: task 3011 * @prio: prio value (kernel-internal form) 3012 * 3013 * This function changes the 'effective' priority of a task. It does 3014 * not touch ->normal_prio like __setscheduler(). 3015 * 3016 * Used by the rt_mutex code to implement priority inheritance logic. 3017 */ 3018void rt_mutex_setprio(struct task_struct *p, int prio) 3019{ 3020 int oldprio, on_rq, running; 3021 struct rq *rq; 3022 const struct sched_class *prev_class; 3023 3024 BUG_ON(prio < 0 || prio > MAX_PRIO); 3025 3026 rq = __task_rq_lock(p); 3027 3028 /* 3029 * Idle task boosting is a nono in general. There is one 3030 * exception, when PREEMPT_RT and NOHZ is active: 3031 * 3032 * The idle task calls get_next_timer_interrupt() and holds 3033 * the timer wheel base->lock on the CPU and another CPU wants 3034 * to access the timer (probably to cancel it). We can safely 3035 * ignore the boosting request, as the idle CPU runs this code 3036 * with interrupts disabled and will complete the lock 3037 * protected section without being interrupted. So there is no 3038 * real need to boost. 3039 */ 3040 if (unlikely(p == rq->idle)) { 3041 WARN_ON(p != rq->curr); 3042 WARN_ON(p->pi_blocked_on); 3043 goto out_unlock; 3044 } 3045 3046 trace_sched_pi_setprio(p, prio); 3047 oldprio = p->prio; 3048 prev_class = p->sched_class; 3049 on_rq = p->on_rq; 3050 running = task_current(rq, p); 3051 if (on_rq) 3052 dequeue_task(rq, p, 0); 3053 if (running) 3054 p->sched_class->put_prev_task(rq, p); 3055 3056 if (rt_prio(prio)) 3057 p->sched_class = &rt_sched_class; 3058 else 3059 p->sched_class = &fair_sched_class; 3060 3061 p->prio = prio; 3062 3063 if (running) 3064 p->sched_class->set_curr_task(rq); 3065 if (on_rq) 3066 enqueue_task(rq, p, oldprio < prio ? ENQUEUE_HEAD : 0); 3067 3068 check_class_changed(rq, p, prev_class, oldprio); 3069out_unlock: 3070 __task_rq_unlock(rq); 3071} 3072#endif 3073void set_user_nice(struct task_struct *p, long nice) 3074{ 3075 int old_prio, delta, on_rq; 3076 unsigned long flags; 3077 struct rq *rq; 3078 3079 if (TASK_NICE(p) == nice || nice < -20 || nice > 19) 3080 return; 3081 /* 3082 * We have to be careful, if called from sys_setpriority(), 3083 * the task might be in the middle of scheduling on another CPU. 3084 */ 3085 rq = task_rq_lock(p, &flags); 3086 /* 3087 * The RT priorities are set via sched_setscheduler(), but we still 3088 * allow the 'normal' nice value to be set - but as expected 3089 * it wont have any effect on scheduling until the task is 3090 * SCHED_FIFO/SCHED_RR: 3091 */ 3092 if (task_has_rt_policy(p)) { 3093 p->static_prio = NICE_TO_PRIO(nice); 3094 goto out_unlock; 3095 } 3096 on_rq = p->on_rq; 3097 if (on_rq) 3098 dequeue_task(rq, p, 0); 3099 3100 p->static_prio = NICE_TO_PRIO(nice); 3101 set_load_weight(p); 3102 old_prio = p->prio; 3103 p->prio = effective_prio(p); 3104 delta = p->prio - old_prio; 3105 3106 if (on_rq) { 3107 enqueue_task(rq, p, 0); 3108 /* 3109 * If the task increased its priority or is running and 3110 * lowered its priority, then reschedule its CPU: 3111 */ 3112 if (delta < 0 || (delta > 0 && task_running(rq, p))) 3113 resched_task(rq->curr); 3114 } 3115out_unlock: 3116 task_rq_unlock(rq, p, &flags); 3117} 3118EXPORT_SYMBOL(set_user_nice); 3119 3120/* 3121 * can_nice - check if a task can reduce its nice value 3122 * @p: task 3123 * @nice: nice value 3124 */ 3125int can_nice(const struct task_struct *p, const int nice) 3126{ 3127 /* convert nice value [19,-20] to rlimit style value [1,40] */ 3128 int nice_rlim = 20 - nice; 3129 3130 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) || 3131 capable(CAP_SYS_NICE)); 3132} 3133 3134#ifdef __ARCH_WANT_SYS_NICE 3135 3136/* 3137 * sys_nice - change the priority of the current process. 3138 * @increment: priority increment 3139 * 3140 * sys_setpriority is a more generic, but much slower function that 3141 * does similar things. 3142 */ 3143SYSCALL_DEFINE1(nice, int, increment) 3144{ 3145 long nice, retval; 3146 3147 /* 3148 * Setpriority might change our priority at the same moment. 3149 * We don't have to worry. Conceptually one call occurs first 3150 * and we have a single winner. 3151 */ 3152 if (increment < -40) 3153 increment = -40; 3154 if (increment > 40) 3155 increment = 40; 3156 3157 nice = TASK_NICE(current) + increment; 3158 if (nice < -20) 3159 nice = -20; 3160 if (nice > 19) 3161 nice = 19; 3162 3163 if (increment < 0 && !can_nice(current, nice)) 3164 return -EPERM; 3165 3166 retval = security_task_setnice(current, nice); 3167 if (retval) 3168 return retval; 3169 3170 set_user_nice(current, nice); 3171 return 0; 3172} 3173 3174#endif 3175 3176/** 3177 * task_prio - return the priority value of a given task. 3178 * @p: the task in question. 3179 * 3180 * Return: The priority value as seen by users in /proc. 3181 * RT tasks are offset by -200. Normal tasks are centered 3182 * around 0, value goes from -16 to +15. 3183 */ 3184int task_prio(const struct task_struct *p) 3185{ 3186 return p->prio - MAX_RT_PRIO; 3187} 3188 3189/** 3190 * task_nice - return the nice value of a given task. 3191 * @p: the task in question. 3192 * 3193 * Return: The nice value [ -20 ... 0 ... 19 ]. 3194 */ 3195int task_nice(const struct task_struct *p) 3196{ 3197 return TASK_NICE(p); 3198} 3199EXPORT_SYMBOL(task_nice); 3200 3201/** 3202 * idle_cpu - is a given cpu idle currently? 3203 * @cpu: the processor in question. 3204 * 3205 * Return: 1 if the CPU is currently idle. 0 otherwise. 3206 */ 3207int idle_cpu(int cpu) 3208{ 3209 struct rq *rq = cpu_rq(cpu); 3210 3211 if (rq->curr != rq->idle) 3212 return 0; 3213 3214 if (rq->nr_running) 3215 return 0; 3216 3217#ifdef CONFIG_SMP 3218 if (!llist_empty(&rq->wake_list)) 3219 return 0; 3220#endif 3221 3222 return 1; 3223} 3224 3225/** 3226 * idle_task - return the idle task for a given cpu. 3227 * @cpu: the processor in question. 3228 * 3229 * Return: The idle task for the cpu @cpu. 3230 */ 3231struct task_struct *idle_task(int cpu) 3232{ 3233 return cpu_rq(cpu)->idle; 3234} 3235 3236/** 3237 * find_process_by_pid - find a process with a matching PID value. 3238 * @pid: the pid in question. 3239 * 3240 * The task of @pid, if found. %NULL otherwise. 3241 */ 3242static struct task_struct *find_process_by_pid(pid_t pid) 3243{ 3244 return pid ? find_task_by_vpid(pid) : current; 3245} 3246 3247/* Actually do priority change: must hold rq lock. */ 3248static void 3249__setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio) 3250{ 3251 p->policy = policy; 3252 p->rt_priority = prio; 3253 p->normal_prio = normal_prio(p); 3254 /* we are holding p->pi_lock already */ 3255 p->prio = rt_mutex_getprio(p); 3256 if (rt_prio(p->prio)) 3257 p->sched_class = &rt_sched_class; 3258 else 3259 p->sched_class = &fair_sched_class; 3260 set_load_weight(p); 3261} 3262 3263/* 3264 * check the target process has a UID that matches the current process's 3265 */ 3266static bool check_same_owner(struct task_struct *p) 3267{ 3268 const struct cred *cred = current_cred(), *pcred; 3269 bool match; 3270 3271 rcu_read_lock(); 3272 pcred = __task_cred(p); 3273 match = (uid_eq(cred->euid, pcred->euid) || 3274 uid_eq(cred->euid, pcred->uid)); 3275 rcu_read_unlock(); 3276 return match; 3277} 3278 3279static int __sched_setscheduler(struct task_struct *p, int policy, 3280 const struct sched_param *param, bool user) 3281{ 3282 int retval, oldprio, oldpolicy = -1, on_rq, running; 3283 unsigned long flags; 3284 const struct sched_class *prev_class; 3285 struct rq *rq; 3286 int reset_on_fork; 3287 3288 /* may grab non-irq protected spin_locks */ 3289 BUG_ON(in_interrupt()); 3290recheck: 3291 /* double check policy once rq lock held */ 3292 if (policy < 0) { 3293 reset_on_fork = p->sched_reset_on_fork; 3294 policy = oldpolicy = p->policy; 3295 } else { 3296 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK); 3297 policy &= ~SCHED_RESET_ON_FORK; 3298 3299 if (policy != SCHED_FIFO && policy != SCHED_RR && 3300 policy != SCHED_NORMAL && policy != SCHED_BATCH && 3301 policy != SCHED_IDLE) 3302 return -EINVAL; 3303 } 3304 3305 /* 3306 * Valid priorities for SCHED_FIFO and SCHED_RR are 3307 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL, 3308 * SCHED_BATCH and SCHED_IDLE is 0. 3309 */ 3310 if (param->sched_priority < 0 || 3311 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) || 3312 (!p->mm && param->sched_priority > MAX_RT_PRIO-1)) 3313 return -EINVAL; 3314 if (rt_policy(policy) != (param->sched_priority != 0)) 3315 return -EINVAL; 3316 3317 /* 3318 * Allow unprivileged RT tasks to decrease priority: 3319 */ 3320 if (user && !capable(CAP_SYS_NICE)) { 3321 if (rt_policy(policy)) { 3322 unsigned long rlim_rtprio = 3323 task_rlimit(p, RLIMIT_RTPRIO); 3324 3325 /* can't set/change the rt policy */ 3326 if (policy != p->policy && !rlim_rtprio) 3327 return -EPERM; 3328 3329 /* can't increase priority */ 3330 if (param->sched_priority > p->rt_priority && 3331 param->sched_priority > rlim_rtprio) 3332 return -EPERM; 3333 } 3334 3335 /* 3336 * Treat SCHED_IDLE as nice 20. Only allow a switch to 3337 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it. 3338 */ 3339 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) { 3340 if (!can_nice(p, TASK_NICE(p))) 3341 return -EPERM; 3342 } 3343 3344 /* can't change other user's priorities */ 3345 if (!check_same_owner(p)) 3346 return -EPERM; 3347 3348 /* Normal users shall not reset the sched_reset_on_fork flag */ 3349 if (p->sched_reset_on_fork && !reset_on_fork) 3350 return -EPERM; 3351 } 3352 3353 if (user) { 3354 retval = security_task_setscheduler(p); 3355 if (retval) 3356 return retval; 3357 } 3358 3359 /* 3360 * make sure no PI-waiters arrive (or leave) while we are 3361 * changing the priority of the task: 3362 * 3363 * To be able to change p->policy safely, the appropriate 3364 * runqueue lock must be held. 3365 */ 3366 rq = task_rq_lock(p, &flags); 3367 3368 /* 3369 * Changing the policy of the stop threads its a very bad idea 3370 */ 3371 if (p == rq->stop) { 3372 task_rq_unlock(rq, p, &flags); 3373 return -EINVAL; 3374 } 3375 3376 /* 3377 * If not changing anything there's no need to proceed further: 3378 */ 3379 if (unlikely(policy == p->policy && (!rt_policy(policy) || 3380 param->sched_priority == p->rt_priority))) { 3381 task_rq_unlock(rq, p, &flags); 3382 return 0; 3383 } 3384 3385#ifdef CONFIG_RT_GROUP_SCHED 3386 if (user) { 3387 /* 3388 * Do not allow realtime tasks into groups that have no runtime 3389 * assigned. 3390 */ 3391 if (rt_bandwidth_enabled() && rt_policy(policy) && 3392 task_group(p)->rt_bandwidth.rt_runtime == 0 && 3393 !task_group_is_autogroup(task_group(p))) { 3394 task_rq_unlock(rq, p, &flags); 3395 return -EPERM; 3396 } 3397 } 3398#endif 3399 3400 /* recheck policy now with rq lock held */ 3401 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) { 3402 policy = oldpolicy = -1; 3403 task_rq_unlock(rq, p, &flags); 3404 goto recheck; 3405 } 3406 on_rq = p->on_rq; 3407 running = task_current(rq, p); 3408 if (on_rq) 3409 dequeue_task(rq, p, 0); 3410 if (running) 3411 p->sched_class->put_prev_task(rq, p); 3412 3413 p->sched_reset_on_fork = reset_on_fork; 3414 3415 oldprio = p->prio; 3416 prev_class = p->sched_class; 3417 __setscheduler(rq, p, policy, param->sched_priority); 3418 3419 if (running) 3420 p->sched_class->set_curr_task(rq); 3421 if (on_rq) 3422 enqueue_task(rq, p, 0); 3423 3424 check_class_changed(rq, p, prev_class, oldprio); 3425 task_rq_unlock(rq, p, &flags); 3426 3427 rt_mutex_adjust_pi(p); 3428 3429 return 0; 3430} 3431 3432/** 3433 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread. 3434 * @p: the task in question. 3435 * @policy: new policy. 3436 * @param: structure containing the new RT priority. 3437 * 3438 * Return: 0 on success. An error code otherwise. 3439 * 3440 * NOTE that the task may be already dead. 3441 */ 3442int sched_setscheduler(struct task_struct *p, int policy, 3443 const struct sched_param *param) 3444{ 3445 return __sched_setscheduler(p, policy, param, true); 3446} 3447EXPORT_SYMBOL_GPL(sched_setscheduler); 3448 3449/** 3450 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace. 3451 * @p: the task in question. 3452 * @policy: new policy. 3453 * @param: structure containing the new RT priority. 3454 * 3455 * Just like sched_setscheduler, only don't bother checking if the 3456 * current context has permission. For example, this is needed in 3457 * stop_machine(): we create temporary high priority worker threads, 3458 * but our caller might not have that capability. 3459 * 3460 * Return: 0 on success. An error code otherwise. 3461 */ 3462int sched_setscheduler_nocheck(struct task_struct *p, int policy, 3463 const struct sched_param *param) 3464{ 3465 return __sched_setscheduler(p, policy, param, false); 3466} 3467 3468static int 3469do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param) 3470{ 3471 struct sched_param lparam; 3472 struct task_struct *p; 3473 int retval; 3474 3475 if (!param || pid < 0) 3476 return -EINVAL; 3477 if (copy_from_user(&lparam, param, sizeof(struct sched_param))) 3478 return -EFAULT; 3479 3480 rcu_read_lock(); 3481 retval = -ESRCH; 3482 p = find_process_by_pid(pid); 3483 if (p != NULL) 3484 retval = sched_setscheduler(p, policy, &lparam); 3485 rcu_read_unlock(); 3486 3487 return retval; 3488} 3489 3490/** 3491 * sys_sched_setscheduler - set/change the scheduler policy and RT priority 3492 * @pid: the pid in question. 3493 * @policy: new policy. 3494 * @param: structure containing the new RT priority. 3495 * 3496 * Return: 0 on success. An error code otherwise. 3497 */ 3498SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy, 3499 struct sched_param __user *, param) 3500{ 3501 /* negative values for policy are not valid */ 3502 if (policy < 0) 3503 return -EINVAL; 3504 3505 return do_sched_setscheduler(pid, policy, param); 3506} 3507 3508/** 3509 * sys_sched_setparam - set/change the RT priority of a thread 3510 * @pid: the pid in question. 3511 * @param: structure containing the new RT priority. 3512 * 3513 * Return: 0 on success. An error code otherwise. 3514 */ 3515SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param) 3516{ 3517 return do_sched_setscheduler(pid, -1, param); 3518} 3519 3520/** 3521 * sys_sched_getscheduler - get the policy (scheduling class) of a thread 3522 * @pid: the pid in question. 3523 * 3524 * Return: On success, the policy of the thread. Otherwise, a negative error 3525 * code. 3526 */ 3527SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid) 3528{ 3529 struct task_struct *p; 3530 int retval; 3531 3532 if (pid < 0) 3533 return -EINVAL; 3534 3535 retval = -ESRCH; 3536 rcu_read_lock(); 3537 p = find_process_by_pid(pid); 3538 if (p) { 3539 retval = security_task_getscheduler(p); 3540 if (!retval) 3541 retval = p->policy 3542 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0); 3543 } 3544 rcu_read_unlock(); 3545 return retval; 3546} 3547 3548/** 3549 * sys_sched_getparam - get the RT priority of a thread 3550 * @pid: the pid in question. 3551 * @param: structure containing the RT priority. 3552 * 3553 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error 3554 * code. 3555 */ 3556SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param) 3557{ 3558 struct sched_param lp; 3559 struct task_struct *p; 3560 int retval; 3561 3562 if (!param || pid < 0) 3563 return -EINVAL; 3564 3565 rcu_read_lock(); 3566 p = find_process_by_pid(pid); 3567 retval = -ESRCH; 3568 if (!p) 3569 goto out_unlock; 3570 3571 retval = security_task_getscheduler(p); 3572 if (retval) 3573 goto out_unlock; 3574 3575 lp.sched_priority = p->rt_priority; 3576 rcu_read_unlock(); 3577 3578 /* 3579 * This one might sleep, we cannot do it with a spinlock held ... 3580 */ 3581 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0; 3582 3583 return retval; 3584 3585out_unlock: 3586 rcu_read_unlock(); 3587 return retval; 3588} 3589 3590long sched_setaffinity(pid_t pid, const struct cpumask *in_mask) 3591{ 3592 cpumask_var_t cpus_allowed, new_mask; 3593 struct task_struct *p; 3594 int retval; 3595 3596 get_online_cpus(); 3597 rcu_read_lock(); 3598 3599 p = find_process_by_pid(pid); 3600 if (!p) { 3601 rcu_read_unlock(); 3602 put_online_cpus(); 3603 return -ESRCH; 3604 } 3605 3606 /* Prevent p going away */ 3607 get_task_struct(p); 3608 rcu_read_unlock(); 3609 3610 if (p->flags & PF_NO_SETAFFINITY) { 3611 retval = -EINVAL; 3612 goto out_put_task; 3613 } 3614 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) { 3615 retval = -ENOMEM; 3616 goto out_put_task; 3617 } 3618 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) { 3619 retval = -ENOMEM; 3620 goto out_free_cpus_allowed; 3621 } 3622 retval = -EPERM; 3623 if (!check_same_owner(p)) { 3624 rcu_read_lock(); 3625 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) { 3626 rcu_read_unlock(); 3627 goto out_unlock; 3628 } 3629 rcu_read_unlock(); 3630 } 3631 3632 retval = security_task_setscheduler(p); 3633 if (retval) 3634 goto out_unlock; 3635 3636 cpuset_cpus_allowed(p, cpus_allowed); 3637 cpumask_and(new_mask, in_mask, cpus_allowed); 3638again: 3639 retval = set_cpus_allowed_ptr(p, new_mask); 3640 3641 if (!retval) { 3642 cpuset_cpus_allowed(p, cpus_allowed); 3643 if (!cpumask_subset(new_mask, cpus_allowed)) { 3644 /* 3645 * We must have raced with a concurrent cpuset 3646 * update. Just reset the cpus_allowed to the 3647 * cpuset's cpus_allowed 3648 */ 3649 cpumask_copy(new_mask, cpus_allowed); 3650 goto again; 3651 } 3652 } 3653out_unlock: 3654 free_cpumask_var(new_mask); 3655out_free_cpus_allowed: 3656 free_cpumask_var(cpus_allowed); 3657out_put_task: 3658 put_task_struct(p); 3659 put_online_cpus(); 3660 return retval; 3661} 3662 3663static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len, 3664 struct cpumask *new_mask) 3665{ 3666 if (len < cpumask_size()) 3667 cpumask_clear(new_mask); 3668 else if (len > cpumask_size()) 3669 len = cpumask_size(); 3670 3671 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0; 3672} 3673 3674/** 3675 * sys_sched_setaffinity - set the cpu affinity of a process 3676 * @pid: pid of the process 3677 * @len: length in bytes of the bitmask pointed to by user_mask_ptr 3678 * @user_mask_ptr: user-space pointer to the new cpu mask 3679 * 3680 * Return: 0 on success. An error code otherwise. 3681 */ 3682SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len, 3683 unsigned long __user *, user_mask_ptr) 3684{ 3685 cpumask_var_t new_mask; 3686 int retval; 3687 3688 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) 3689 return -ENOMEM; 3690 3691 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask); 3692 if (retval == 0) 3693 retval = sched_setaffinity(pid, new_mask); 3694 free_cpumask_var(new_mask); 3695 return retval; 3696} 3697 3698long sched_getaffinity(pid_t pid, struct cpumask *mask) 3699{ 3700 struct task_struct *p; 3701 unsigned long flags; 3702 int retval; 3703 3704 get_online_cpus(); 3705 rcu_read_lock(); 3706 3707 retval = -ESRCH; 3708 p = find_process_by_pid(pid); 3709 if (!p) 3710 goto out_unlock; 3711 3712 retval = security_task_getscheduler(p); 3713 if (retval) 3714 goto out_unlock; 3715 3716 raw_spin_lock_irqsave(&p->pi_lock, flags); 3717 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask); 3718 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 3719 3720out_unlock: 3721 rcu_read_unlock(); 3722 put_online_cpus(); 3723 3724 return retval; 3725} 3726 3727/** 3728 * sys_sched_getaffinity - get the cpu affinity of a process 3729 * @pid: pid of the process 3730 * @len: length in bytes of the bitmask pointed to by user_mask_ptr 3731 * @user_mask_ptr: user-space pointer to hold the current cpu mask 3732 * 3733 * Return: 0 on success. An error code otherwise. 3734 */ 3735SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len, 3736 unsigned long __user *, user_mask_ptr) 3737{ 3738 int ret; 3739 cpumask_var_t mask; 3740 3741 if ((len * BITS_PER_BYTE) < nr_cpu_ids) 3742 return -EINVAL; 3743 if (len & (sizeof(unsigned long)-1)) 3744 return -EINVAL; 3745 3746 if (!alloc_cpumask_var(&mask, GFP_KERNEL)) 3747 return -ENOMEM; 3748 3749 ret = sched_getaffinity(pid, mask); 3750 if (ret == 0) { 3751 size_t retlen = min_t(size_t, len, cpumask_size()); 3752 3753 if (copy_to_user(user_mask_ptr, mask, retlen)) 3754 ret = -EFAULT; 3755 else 3756 ret = retlen; 3757 } 3758 free_cpumask_var(mask); 3759 3760 return ret; 3761} 3762 3763/** 3764 * sys_sched_yield - yield the current processor to other threads. 3765 * 3766 * This function yields the current CPU to other tasks. If there are no 3767 * other threads running on this CPU then this function will return. 3768 * 3769 * Return: 0. 3770 */ 3771SYSCALL_DEFINE0(sched_yield) 3772{ 3773 struct rq *rq = this_rq_lock(); 3774 3775 schedstat_inc(rq, yld_count); 3776 current->sched_class->yield_task(rq); 3777 3778 /* 3779 * Since we are going to call schedule() anyway, there's 3780 * no need to preempt or enable interrupts: 3781 */ 3782 __release(rq->lock); 3783 spin_release(&rq->lock.dep_map, 1, _THIS_IP_); 3784 do_raw_spin_unlock(&rq->lock); 3785 sched_preempt_enable_no_resched(); 3786 3787 schedule(); 3788 3789 return 0; 3790} 3791 3792static inline int should_resched(void) 3793{ 3794 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE); 3795} 3796 3797static void __cond_resched(void) 3798{ 3799 add_preempt_count(PREEMPT_ACTIVE); 3800 __schedule(); 3801 sub_preempt_count(PREEMPT_ACTIVE); 3802} 3803 3804int __sched _cond_resched(void) 3805{ 3806 if (should_resched()) { 3807 __cond_resched(); 3808 return 1; 3809 } 3810 return 0; 3811} 3812EXPORT_SYMBOL(_cond_resched); 3813 3814/* 3815 * __cond_resched_lock() - if a reschedule is pending, drop the given lock, 3816 * call schedule, and on return reacquire the lock. 3817 * 3818 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level 3819 * operations here to prevent schedule() from being called twice (once via 3820 * spin_unlock(), once by hand). 3821 */ 3822int __cond_resched_lock(spinlock_t *lock) 3823{ 3824 int resched = should_resched(); 3825 int ret = 0; 3826 3827 lockdep_assert_held(lock); 3828 3829 if (spin_needbreak(lock) || resched) { 3830 spin_unlock(lock); 3831 if (resched) 3832 __cond_resched(); 3833 else 3834 cpu_relax(); 3835 ret = 1; 3836 spin_lock(lock); 3837 } 3838 return ret; 3839} 3840EXPORT_SYMBOL(__cond_resched_lock); 3841 3842int __sched __cond_resched_softirq(void) 3843{ 3844 BUG_ON(!in_softirq()); 3845 3846 if (should_resched()) { 3847 local_bh_enable(); 3848 __cond_resched(); 3849 local_bh_disable(); 3850 return 1; 3851 } 3852 return 0; 3853} 3854EXPORT_SYMBOL(__cond_resched_softirq); 3855 3856/** 3857 * yield - yield the current processor to other threads. 3858 * 3859 * Do not ever use this function, there's a 99% chance you're doing it wrong. 3860 * 3861 * The scheduler is at all times free to pick the calling task as the most 3862 * eligible task to run, if removing the yield() call from your code breaks 3863 * it, its already broken. 3864 * 3865 * Typical broken usage is: 3866 * 3867 * while (!event) 3868 * yield(); 3869 * 3870 * where one assumes that yield() will let 'the other' process run that will 3871 * make event true. If the current task is a SCHED_FIFO task that will never 3872 * happen. Never use yield() as a progress guarantee!! 3873 * 3874 * If you want to use yield() to wait for something, use wait_event(). 3875 * If you want to use yield() to be 'nice' for others, use cond_resched(). 3876 * If you still want to use yield(), do not! 3877 */ 3878void __sched yield(void) 3879{ 3880 set_current_state(TASK_RUNNING); 3881 sys_sched_yield(); 3882} 3883EXPORT_SYMBOL(yield); 3884 3885/** 3886 * yield_to - yield the current processor to another thread in 3887 * your thread group, or accelerate that thread toward the 3888 * processor it's on. 3889 * @p: target task 3890 * @preempt: whether task preemption is allowed or not 3891 * 3892 * It's the caller's job to ensure that the target task struct 3893 * can't go away on us before we can do any checks. 3894 * 3895 * Return: 3896 * true (>0) if we indeed boosted the target task. 3897 * false (0) if we failed to boost the target. 3898 * -ESRCH if there's no task to yield to. 3899 */ 3900bool __sched yield_to(struct task_struct *p, bool preempt) 3901{ 3902 struct task_struct *curr = current; 3903 struct rq *rq, *p_rq; 3904 unsigned long flags; 3905 int yielded = 0; 3906 3907 local_irq_save(flags); 3908 rq = this_rq(); 3909 3910again: 3911 p_rq = task_rq(p); 3912 /* 3913 * If we're the only runnable task on the rq and target rq also 3914 * has only one task, there's absolutely no point in yielding. 3915 */ 3916 if (rq->nr_running == 1 && p_rq->nr_running == 1) { 3917 yielded = -ESRCH; 3918 goto out_irq; 3919 } 3920 3921 double_rq_lock(rq, p_rq); 3922 while (task_rq(p) != p_rq) { 3923 double_rq_unlock(rq, p_rq); 3924 goto again; 3925 } 3926 3927 if (!curr->sched_class->yield_to_task) 3928 goto out_unlock; 3929 3930 if (curr->sched_class != p->sched_class) 3931 goto out_unlock; 3932 3933 if (task_running(p_rq, p) || p->state) 3934 goto out_unlock; 3935 3936 yielded = curr->sched_class->yield_to_task(rq, p, preempt); 3937 if (yielded) { 3938 schedstat_inc(rq, yld_count); 3939 /* 3940 * Make p's CPU reschedule; pick_next_entity takes care of 3941 * fairness. 3942 */ 3943 if (preempt && rq != p_rq) 3944 resched_task(p_rq->curr); 3945 } 3946 3947out_unlock: 3948 double_rq_unlock(rq, p_rq); 3949out_irq: 3950 local_irq_restore(flags); 3951 3952 if (yielded > 0) 3953 schedule(); 3954 3955 return yielded; 3956} 3957EXPORT_SYMBOL_GPL(yield_to); 3958 3959/* 3960 * This task is about to go to sleep on IO. Increment rq->nr_iowait so 3961 * that process accounting knows that this is a task in IO wait state. 3962 */ 3963void __sched io_schedule(void) 3964{ 3965 struct rq *rq = raw_rq(); 3966 3967 delayacct_blkio_start(); 3968 atomic_inc(&rq->nr_iowait); 3969 blk_flush_plug(current); 3970 current->in_iowait = 1; 3971 schedule(); 3972 current->in_iowait = 0; 3973 atomic_dec(&rq->nr_iowait); 3974 delayacct_blkio_end(); 3975} 3976EXPORT_SYMBOL(io_schedule); 3977 3978long __sched io_schedule_timeout(long timeout) 3979{ 3980 struct rq *rq = raw_rq(); 3981 long ret; 3982 3983 delayacct_blkio_start(); 3984 atomic_inc(&rq->nr_iowait); 3985 blk_flush_plug(current); 3986 current->in_iowait = 1; 3987 ret = schedule_timeout(timeout); 3988 current->in_iowait = 0; 3989 atomic_dec(&rq->nr_iowait); 3990 delayacct_blkio_end(); 3991 return ret; 3992} 3993 3994/** 3995 * sys_sched_get_priority_max - return maximum RT priority. 3996 * @policy: scheduling class. 3997 * 3998 * Return: On success, this syscall returns the maximum 3999 * rt_priority that can be used by a given scheduling class. 4000 * On failure, a negative error code is returned. 4001 */ 4002SYSCALL_DEFINE1(sched_get_priority_max, int, policy) 4003{ 4004 int ret = -EINVAL; 4005 4006 switch (policy) { 4007 case SCHED_FIFO: 4008 case SCHED_RR: 4009 ret = MAX_USER_RT_PRIO-1; 4010 break; 4011 case SCHED_NORMAL: 4012 case SCHED_BATCH: 4013 case SCHED_IDLE: 4014 ret = 0; 4015 break; 4016 } 4017 return ret; 4018} 4019 4020/** 4021 * sys_sched_get_priority_min - return minimum RT priority. 4022 * @policy: scheduling class. 4023 * 4024 * Return: On success, this syscall returns the minimum 4025 * rt_priority that can be used by a given scheduling class. 4026 * On failure, a negative error code is returned. 4027 */ 4028SYSCALL_DEFINE1(sched_get_priority_min, int, policy) 4029{ 4030 int ret = -EINVAL; 4031 4032 switch (policy) { 4033 case SCHED_FIFO: 4034 case SCHED_RR: 4035 ret = 1; 4036 break; 4037 case SCHED_NORMAL: 4038 case SCHED_BATCH: 4039 case SCHED_IDLE: 4040 ret = 0; 4041 } 4042 return ret; 4043} 4044 4045/** 4046 * sys_sched_rr_get_interval - return the default timeslice of a process. 4047 * @pid: pid of the process. 4048 * @interval: userspace pointer to the timeslice value. 4049 * 4050 * this syscall writes the default timeslice value of a given process 4051 * into the user-space timespec buffer. A value of '0' means infinity. 4052 * 4053 * Return: On success, 0 and the timeslice is in @interval. Otherwise, 4054 * an error code. 4055 */ 4056SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid, 4057 struct timespec __user *, interval) 4058{ 4059 struct task_struct *p; 4060 unsigned int time_slice; 4061 unsigned long flags; 4062 struct rq *rq; 4063 int retval; 4064 struct timespec t; 4065 4066 if (pid < 0) 4067 return -EINVAL; 4068 4069 retval = -ESRCH; 4070 rcu_read_lock(); 4071 p = find_process_by_pid(pid); 4072 if (!p) 4073 goto out_unlock; 4074 4075 retval = security_task_getscheduler(p); 4076 if (retval) 4077 goto out_unlock; 4078 4079 rq = task_rq_lock(p, &flags); 4080 time_slice = p->sched_class->get_rr_interval(rq, p); 4081 task_rq_unlock(rq, p, &flags); 4082 4083 rcu_read_unlock(); 4084 jiffies_to_timespec(time_slice, &t); 4085 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0; 4086 return retval; 4087 4088out_unlock: 4089 rcu_read_unlock(); 4090 return retval; 4091} 4092 4093static const char stat_nam[] = TASK_STATE_TO_CHAR_STR; 4094 4095void sched_show_task(struct task_struct *p) 4096{ 4097 unsigned long free = 0; 4098 int ppid; 4099 unsigned state; 4100 4101 state = p->state ? __ffs(p->state) + 1 : 0; 4102 printk(KERN_INFO "%-15.15s %c", p->comm, 4103 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?'); 4104#if BITS_PER_LONG == 32 4105 if (state == TASK_RUNNING) 4106 printk(KERN_CONT " running "); 4107 else 4108 printk(KERN_CONT " %08lx ", thread_saved_pc(p)); 4109#else 4110 if (state == TASK_RUNNING) 4111 printk(KERN_CONT " running task "); 4112 else 4113 printk(KERN_CONT " %016lx ", thread_saved_pc(p)); 4114#endif 4115#ifdef CONFIG_DEBUG_STACK_USAGE 4116 free = stack_not_used(p); 4117#endif 4118 rcu_read_lock(); 4119 ppid = task_pid_nr(rcu_dereference(p->real_parent)); 4120 rcu_read_unlock(); 4121 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free, 4122 task_pid_nr(p), ppid, 4123 (unsigned long)task_thread_info(p)->flags); 4124 4125 print_worker_info(KERN_INFO, p); 4126 show_stack(p, NULL); 4127} 4128 4129void show_state_filter(unsigned long state_filter) 4130{ 4131 struct task_struct *g, *p; 4132 4133#if BITS_PER_LONG == 32 4134 printk(KERN_INFO 4135 " task PC stack pid father\n"); 4136#else 4137 printk(KERN_INFO 4138 " task PC stack pid father\n"); 4139#endif 4140 rcu_read_lock(); 4141 do_each_thread(g, p) { 4142 /* 4143 * reset the NMI-timeout, listing all files on a slow 4144 * console might take a lot of time: 4145 */ 4146 touch_nmi_watchdog(); 4147 if (!state_filter || (p->state & state_filter)) 4148 sched_show_task(p); 4149 } while_each_thread(g, p); 4150 4151 touch_all_softlockup_watchdogs(); 4152 4153#ifdef CONFIG_SCHED_DEBUG 4154 sysrq_sched_debug_show(); 4155#endif 4156 rcu_read_unlock(); 4157 /* 4158 * Only show locks if all tasks are dumped: 4159 */ 4160 if (!state_filter) 4161 debug_show_all_locks(); 4162} 4163 4164void init_idle_bootup_task(struct task_struct *idle) 4165{ 4166 idle->sched_class = &idle_sched_class; 4167} 4168 4169/** 4170 * init_idle - set up an idle thread for a given CPU 4171 * @idle: task in question 4172 * @cpu: cpu the idle task belongs to 4173 * 4174 * NOTE: this function does not set the idle thread's NEED_RESCHED 4175 * flag, to make booting more robust. 4176 */ 4177void init_idle(struct task_struct *idle, int cpu) 4178{ 4179 struct rq *rq = cpu_rq(cpu); 4180 unsigned long flags; 4181 4182 raw_spin_lock_irqsave(&rq->lock, flags); 4183 4184 __sched_fork(idle); 4185 idle->state = TASK_RUNNING; 4186 idle->se.exec_start = sched_clock(); 4187 4188 do_set_cpus_allowed(idle, cpumask_of(cpu)); 4189 /* 4190 * We're having a chicken and egg problem, even though we are 4191 * holding rq->lock, the cpu isn't yet set to this cpu so the 4192 * lockdep check in task_group() will fail. 4193 * 4194 * Similar case to sched_fork(). / Alternatively we could 4195 * use task_rq_lock() here and obtain the other rq->lock. 4196 * 4197 * Silence PROVE_RCU 4198 */ 4199 rcu_read_lock(); 4200 __set_task_cpu(idle, cpu); 4201 rcu_read_unlock(); 4202 4203 rq->curr = rq->idle = idle; 4204#if defined(CONFIG_SMP) 4205 idle->on_cpu = 1; 4206#endif 4207 raw_spin_unlock_irqrestore(&rq->lock, flags); 4208 4209 /* Set the preempt count _outside_ the spinlocks! */ 4210 task_thread_info(idle)->preempt_count = 0; 4211 4212 /* 4213 * The idle tasks have their own, simple scheduling class: 4214 */ 4215 idle->sched_class = &idle_sched_class; 4216 ftrace_graph_init_idle_task(idle, cpu); 4217 vtime_init_idle(idle, cpu); 4218#if defined(CONFIG_SMP) 4219 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu); 4220#endif 4221} 4222 4223#ifdef CONFIG_SMP 4224void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask) 4225{ 4226 if (p->sched_class && p->sched_class->set_cpus_allowed) 4227 p->sched_class->set_cpus_allowed(p, new_mask); 4228 4229 cpumask_copy(&p->cpus_allowed, new_mask); 4230 p->nr_cpus_allowed = cpumask_weight(new_mask); 4231} 4232 4233/* 4234 * This is how migration works: 4235 * 4236 * 1) we invoke migration_cpu_stop() on the target CPU using 4237 * stop_one_cpu(). 4238 * 2) stopper starts to run (implicitly forcing the migrated thread 4239 * off the CPU) 4240 * 3) it checks whether the migrated task is still in the wrong runqueue. 4241 * 4) if it's in the wrong runqueue then the migration thread removes 4242 * it and puts it into the right queue. 4243 * 5) stopper completes and stop_one_cpu() returns and the migration 4244 * is done. 4245 */ 4246 4247/* 4248 * Change a given task's CPU affinity. Migrate the thread to a 4249 * proper CPU and schedule it away if the CPU it's executing on 4250 * is removed from the allowed bitmask. 4251 * 4252 * NOTE: the caller must have a valid reference to the task, the 4253 * task must not exit() & deallocate itself prematurely. The 4254 * call is not atomic; no spinlocks may be held. 4255 */ 4256int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask) 4257{ 4258 unsigned long flags; 4259 struct rq *rq; 4260 unsigned int dest_cpu; 4261 int ret = 0; 4262 4263 rq = task_rq_lock(p, &flags); 4264 4265 if (cpumask_equal(&p->cpus_allowed, new_mask)) 4266 goto out; 4267 4268 if (!cpumask_intersects(new_mask, cpu_active_mask)) { 4269 ret = -EINVAL; 4270 goto out; 4271 } 4272 4273 do_set_cpus_allowed(p, new_mask); 4274 4275 /* Can the task run on the task's current CPU? If so, we're done */ 4276 if (cpumask_test_cpu(task_cpu(p), new_mask)) 4277 goto out; 4278 4279 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask); 4280 if (p->on_rq) { 4281 struct migration_arg arg = { p, dest_cpu }; 4282 /* Need help from migration thread: drop lock and wait. */ 4283 task_rq_unlock(rq, p, &flags); 4284 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg); 4285 tlb_migrate_finish(p->mm); 4286 return 0; 4287 } 4288out: 4289 task_rq_unlock(rq, p, &flags); 4290 4291 return ret; 4292} 4293EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr); 4294 4295/* 4296 * Move (not current) task off this cpu, onto dest cpu. We're doing 4297 * this because either it can't run here any more (set_cpus_allowed() 4298 * away from this CPU, or CPU going down), or because we're 4299 * attempting to rebalance this task on exec (sched_exec). 4300 * 4301 * So we race with normal scheduler movements, but that's OK, as long 4302 * as the task is no longer on this CPU. 4303 * 4304 * Returns non-zero if task was successfully migrated. 4305 */ 4306static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu) 4307{ 4308 struct rq *rq_dest, *rq_src; 4309 int ret = 0; 4310 4311 if (unlikely(!cpu_active(dest_cpu))) 4312 return ret; 4313 4314 rq_src = cpu_rq(src_cpu); 4315 rq_dest = cpu_rq(dest_cpu); 4316 4317 raw_spin_lock(&p->pi_lock); 4318 double_rq_lock(rq_src, rq_dest); 4319 /* Already moved. */ 4320 if (task_cpu(p) != src_cpu) 4321 goto done; 4322 /* Affinity changed (again). */ 4323 if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p))) 4324 goto fail; 4325 4326 /* 4327 * If we're not on a rq, the next wake-up will ensure we're 4328 * placed properly. 4329 */ 4330 if (p->on_rq) { 4331 dequeue_task(rq_src, p, 0); 4332 set_task_cpu(p, dest_cpu); 4333 enqueue_task(rq_dest, p, 0); 4334 check_preempt_curr(rq_dest, p, 0); 4335 } 4336done: 4337 ret = 1; 4338fail: 4339 double_rq_unlock(rq_src, rq_dest); 4340 raw_spin_unlock(&p->pi_lock); 4341 return ret; 4342} 4343 4344/* 4345 * migration_cpu_stop - this will be executed by a highprio stopper thread 4346 * and performs thread migration by bumping thread off CPU then 4347 * 'pushing' onto another runqueue. 4348 */ 4349static int migration_cpu_stop(void *data) 4350{ 4351 struct migration_arg *arg = data; 4352 4353 /* 4354 * The original target cpu might have gone down and we might 4355 * be on another cpu but it doesn't matter. 4356 */ 4357 local_irq_disable(); 4358 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu); 4359 local_irq_enable(); 4360 return 0; 4361} 4362 4363#ifdef CONFIG_HOTPLUG_CPU 4364 4365/* 4366 * Ensures that the idle task is using init_mm right before its cpu goes 4367 * offline. 4368 */ 4369void idle_task_exit(void) 4370{ 4371 struct mm_struct *mm = current->active_mm; 4372 4373 BUG_ON(cpu_online(smp_processor_id())); 4374 4375 if (mm != &init_mm) 4376 switch_mm(mm, &init_mm, current); 4377 mmdrop(mm); 4378} 4379 4380/* 4381 * Since this CPU is going 'away' for a while, fold any nr_active delta 4382 * we might have. Assumes we're called after migrate_tasks() so that the 4383 * nr_active count is stable. 4384 * 4385 * Also see the comment "Global load-average calculations". 4386 */ 4387static void calc_load_migrate(struct rq *rq) 4388{ 4389 long delta = calc_load_fold_active(rq); 4390 if (delta) 4391 atomic_long_add(delta, &calc_load_tasks); 4392} 4393 4394/* 4395 * Migrate all tasks from the rq, sleeping tasks will be migrated by 4396 * try_to_wake_up()->select_task_rq(). 4397 * 4398 * Called with rq->lock held even though we'er in stop_machine() and 4399 * there's no concurrency possible, we hold the required locks anyway 4400 * because of lock validation efforts. 4401 */ 4402static void migrate_tasks(unsigned int dead_cpu) 4403{ 4404 struct rq *rq = cpu_rq(dead_cpu); 4405 struct task_struct *next, *stop = rq->stop; 4406 int dest_cpu; 4407 4408 /* 4409 * Fudge the rq selection such that the below task selection loop 4410 * doesn't get stuck on the currently eligible stop task. 4411 * 4412 * We're currently inside stop_machine() and the rq is either stuck 4413 * in the stop_machine_cpu_stop() loop, or we're executing this code, 4414 * either way we should never end up calling schedule() until we're 4415 * done here. 4416 */ 4417 rq->stop = NULL; 4418 4419 /* 4420 * put_prev_task() and pick_next_task() sched 4421 * class method both need to have an up-to-date 4422 * value of rq->clock[_task] 4423 */ 4424 update_rq_clock(rq); 4425 4426 for ( ; ; ) { 4427 /* 4428 * There's this thread running, bail when that's the only 4429 * remaining thread. 4430 */ 4431 if (rq->nr_running == 1) 4432 break; 4433 4434 next = pick_next_task(rq); 4435 BUG_ON(!next); 4436 next->sched_class->put_prev_task(rq, next); 4437 4438 /* Find suitable destination for @next, with force if needed. */ 4439 dest_cpu = select_fallback_rq(dead_cpu, next); 4440 raw_spin_unlock(&rq->lock); 4441 4442 __migrate_task(next, dead_cpu, dest_cpu); 4443 4444 raw_spin_lock(&rq->lock); 4445 } 4446 4447 rq->stop = stop; 4448} 4449 4450#endif /* CONFIG_HOTPLUG_CPU */ 4451 4452#if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL) 4453 4454static struct ctl_table sd_ctl_dir[] = { 4455 { 4456 .procname = "sched_domain", 4457 .mode = 0555, 4458 }, 4459 {} 4460}; 4461 4462static struct ctl_table sd_ctl_root[] = { 4463 { 4464 .procname = "kernel", 4465 .mode = 0555, 4466 .child = sd_ctl_dir, 4467 }, 4468 {} 4469}; 4470 4471static struct ctl_table *sd_alloc_ctl_entry(int n) 4472{ 4473 struct ctl_table *entry = 4474 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL); 4475 4476 return entry; 4477} 4478 4479static void sd_free_ctl_entry(struct ctl_table **tablep) 4480{ 4481 struct ctl_table *entry; 4482 4483 /* 4484 * In the intermediate directories, both the child directory and 4485 * procname are dynamically allocated and could fail but the mode 4486 * will always be set. In the lowest directory the names are 4487 * static strings and all have proc handlers. 4488 */ 4489 for (entry = *tablep; entry->mode; entry++) { 4490 if (entry->child) 4491 sd_free_ctl_entry(&entry->child); 4492 if (entry->proc_handler == NULL) 4493 kfree(entry->procname); 4494 } 4495 4496 kfree(*tablep); 4497 *tablep = NULL; 4498} 4499 4500static int min_load_idx = 0; 4501static int max_load_idx = CPU_LOAD_IDX_MAX-1; 4502 4503static void 4504set_table_entry(struct ctl_table *entry, 4505 const char *procname, void *data, int maxlen, 4506 umode_t mode, proc_handler *proc_handler, 4507 bool load_idx) 4508{ 4509 entry->procname = procname; 4510 entry->data = data; 4511 entry->maxlen = maxlen; 4512 entry->mode = mode; 4513 entry->proc_handler = proc_handler; 4514 4515 if (load_idx) { 4516 entry->extra1 = &min_load_idx; 4517 entry->extra2 = &max_load_idx; 4518 } 4519} 4520 4521static struct ctl_table * 4522sd_alloc_ctl_domain_table(struct sched_domain *sd) 4523{ 4524 struct ctl_table *table = sd_alloc_ctl_entry(13); 4525 4526 if (table == NULL) 4527 return NULL; 4528 4529 set_table_entry(&table[0], "min_interval", &sd->min_interval, 4530 sizeof(long), 0644, proc_doulongvec_minmax, false); 4531 set_table_entry(&table[1], "max_interval", &sd->max_interval, 4532 sizeof(long), 0644, proc_doulongvec_minmax, false); 4533 set_table_entry(&table[2], "busy_idx", &sd->busy_idx, 4534 sizeof(int), 0644, proc_dointvec_minmax, true); 4535 set_table_entry(&table[3], "idle_idx", &sd->idle_idx, 4536 sizeof(int), 0644, proc_dointvec_minmax, true); 4537 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx, 4538 sizeof(int), 0644, proc_dointvec_minmax, true); 4539 set_table_entry(&table[5], "wake_idx", &sd->wake_idx, 4540 sizeof(int), 0644, proc_dointvec_minmax, true); 4541 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx, 4542 sizeof(int), 0644, proc_dointvec_minmax, true); 4543 set_table_entry(&table[7], "busy_factor", &sd->busy_factor, 4544 sizeof(int), 0644, proc_dointvec_minmax, false); 4545 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct, 4546 sizeof(int), 0644, proc_dointvec_minmax, false); 4547 set_table_entry(&table[9], "cache_nice_tries", 4548 &sd->cache_nice_tries, 4549 sizeof(int), 0644, proc_dointvec_minmax, false); 4550 set_table_entry(&table[10], "flags", &sd->flags, 4551 sizeof(int), 0644, proc_dointvec_minmax, false); 4552 set_table_entry(&table[11], "name", sd->name, 4553 CORENAME_MAX_SIZE, 0444, proc_dostring, false); 4554 /* &table[12] is terminator */ 4555 4556 return table; 4557} 4558 4559static struct ctl_table *sd_alloc_ctl_cpu_table(int cpu) 4560{ 4561 struct ctl_table *entry, *table; 4562 struct sched_domain *sd; 4563 int domain_num = 0, i; 4564 char buf[32]; 4565 4566 for_each_domain(cpu, sd) 4567 domain_num++; 4568 entry = table = sd_alloc_ctl_entry(domain_num + 1); 4569 if (table == NULL) 4570 return NULL; 4571 4572 i = 0; 4573 for_each_domain(cpu, sd) { 4574 snprintf(buf, 32, "domain%d", i); 4575 entry->procname = kstrdup(buf, GFP_KERNEL); 4576 entry->mode = 0555; 4577 entry->child = sd_alloc_ctl_domain_table(sd); 4578 entry++; 4579 i++; 4580 } 4581 return table; 4582} 4583 4584static struct ctl_table_header *sd_sysctl_header; 4585static void register_sched_domain_sysctl(void) 4586{ 4587 int i, cpu_num = num_possible_cpus(); 4588 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1); 4589 char buf[32]; 4590 4591 WARN_ON(sd_ctl_dir[0].child); 4592 sd_ctl_dir[0].child = entry; 4593 4594 if (entry == NULL) 4595 return; 4596 4597 for_each_possible_cpu(i) { 4598 snprintf(buf, 32, "cpu%d", i); 4599 entry->procname = kstrdup(buf, GFP_KERNEL); 4600 entry->mode = 0555; 4601 entry->child = sd_alloc_ctl_cpu_table(i); 4602 entry++; 4603 } 4604 4605 WARN_ON(sd_sysctl_header); 4606 sd_sysctl_header = register_sysctl_table(sd_ctl_root); 4607} 4608 4609/* may be called multiple times per register */ 4610static void unregister_sched_domain_sysctl(void) 4611{ 4612 if (sd_sysctl_header) 4613 unregister_sysctl_table(sd_sysctl_header); 4614 sd_sysctl_header = NULL; 4615 if (sd_ctl_dir[0].child) 4616 sd_free_ctl_entry(&sd_ctl_dir[0].child); 4617} 4618#else 4619static void register_sched_domain_sysctl(void) 4620{ 4621} 4622static void unregister_sched_domain_sysctl(void) 4623{ 4624} 4625#endif 4626 4627static void set_rq_online(struct rq *rq) 4628{ 4629 if (!rq->online) { 4630 const struct sched_class *class; 4631 4632 cpumask_set_cpu(rq->cpu, rq->rd->online); 4633 rq->online = 1; 4634 4635 for_each_class(class) { 4636 if (class->rq_online) 4637 class->rq_online(rq); 4638 } 4639 } 4640} 4641 4642static void set_rq_offline(struct rq *rq) 4643{ 4644 if (rq->online) { 4645 const struct sched_class *class; 4646 4647 for_each_class(class) { 4648 if (class->rq_offline) 4649 class->rq_offline(rq); 4650 } 4651 4652 cpumask_clear_cpu(rq->cpu, rq->rd->online); 4653 rq->online = 0; 4654 } 4655} 4656 4657/* 4658 * migration_call - callback that gets triggered when a CPU is added. 4659 * Here we can start up the necessary migration thread for the new CPU. 4660 */ 4661static int 4662migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu) 4663{ 4664 int cpu = (long)hcpu; 4665 unsigned long flags; 4666 struct rq *rq = cpu_rq(cpu); 4667 4668 switch (action & ~CPU_TASKS_FROZEN) { 4669 4670 case CPU_UP_PREPARE: 4671 rq->calc_load_update = calc_load_update; 4672 break; 4673 4674 case CPU_ONLINE: 4675 /* Update our root-domain */ 4676 raw_spin_lock_irqsave(&rq->lock, flags); 4677 if (rq->rd) { 4678 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span)); 4679 4680 set_rq_online(rq); 4681 } 4682 raw_spin_unlock_irqrestore(&rq->lock, flags); 4683 break; 4684 4685#ifdef CONFIG_HOTPLUG_CPU 4686 case CPU_DYING: 4687 sched_ttwu_pending(); 4688 /* Update our root-domain */ 4689 raw_spin_lock_irqsave(&rq->lock, flags); 4690 if (rq->rd) { 4691 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span)); 4692 set_rq_offline(rq); 4693 } 4694 migrate_tasks(cpu); 4695 BUG_ON(rq->nr_running != 1); /* the migration thread */ 4696 raw_spin_unlock_irqrestore(&rq->lock, flags); 4697 break; 4698 4699 case CPU_DEAD: 4700 calc_load_migrate(rq); 4701 break; 4702#endif 4703 } 4704 4705 update_max_interval(); 4706 4707 return NOTIFY_OK; 4708} 4709 4710/* 4711 * Register at high priority so that task migration (migrate_all_tasks) 4712 * happens before everything else. This has to be lower priority than 4713 * the notifier in the perf_event subsystem, though. 4714 */ 4715static struct notifier_block migration_notifier = { 4716 .notifier_call = migration_call, 4717 .priority = CPU_PRI_MIGRATION, 4718}; 4719 4720static int sched_cpu_active(struct notifier_block *nfb, 4721 unsigned long action, void *hcpu) 4722{ 4723 switch (action & ~CPU_TASKS_FROZEN) { 4724 case CPU_STARTING: 4725 case CPU_DOWN_FAILED: 4726 set_cpu_active((long)hcpu, true); 4727 return NOTIFY_OK; 4728 default: 4729 return NOTIFY_DONE; 4730 } 4731} 4732 4733static int sched_cpu_inactive(struct notifier_block *nfb, 4734 unsigned long action, void *hcpu) 4735{ 4736 switch (action & ~CPU_TASKS_FROZEN) { 4737 case CPU_DOWN_PREPARE: 4738 set_cpu_active((long)hcpu, false); 4739 return NOTIFY_OK; 4740 default: 4741 return NOTIFY_DONE; 4742 } 4743} 4744 4745static int __init migration_init(void) 4746{ 4747 void *cpu = (void *)(long)smp_processor_id(); 4748 int err; 4749 4750 /* Initialize migration for the boot CPU */ 4751 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu); 4752 BUG_ON(err == NOTIFY_BAD); 4753 migration_call(&migration_notifier, CPU_ONLINE, cpu); 4754 register_cpu_notifier(&migration_notifier); 4755 4756 /* Register cpu active notifiers */ 4757 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE); 4758 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE); 4759 4760 return 0; 4761} 4762early_initcall(migration_init); 4763#endif 4764 4765#ifdef CONFIG_SMP 4766 4767static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */ 4768 4769#ifdef CONFIG_SCHED_DEBUG 4770 4771static __read_mostly int sched_debug_enabled; 4772 4773static int __init sched_debug_setup(char *str) 4774{ 4775 sched_debug_enabled = 1; 4776 4777 return 0; 4778} 4779early_param("sched_debug", sched_debug_setup); 4780 4781static inline bool sched_debug(void) 4782{ 4783 return sched_debug_enabled; 4784} 4785 4786static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level, 4787 struct cpumask *groupmask) 4788{ 4789 struct sched_group *group = sd->groups; 4790 char str[256]; 4791 4792 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd)); 4793 cpumask_clear(groupmask); 4794 4795 printk(KERN_DEBUG "%*s domain %d: ", level, "", level); 4796 4797 if (!(sd->flags & SD_LOAD_BALANCE)) { 4798 printk("does not load-balance\n"); 4799 if (sd->parent) 4800 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain" 4801 " has parent"); 4802 return -1; 4803 } 4804 4805 printk(KERN_CONT "span %s level %s\n", str, sd->name); 4806 4807 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) { 4808 printk(KERN_ERR "ERROR: domain->span does not contain " 4809 "CPU%d\n", cpu); 4810 } 4811 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) { 4812 printk(KERN_ERR "ERROR: domain->groups does not contain" 4813 " CPU%d\n", cpu); 4814 } 4815 4816 printk(KERN_DEBUG "%*s groups:", level + 1, ""); 4817 do { 4818 if (!group) { 4819 printk("\n"); 4820 printk(KERN_ERR "ERROR: group is NULL\n"); 4821 break; 4822 } 4823 4824 /* 4825 * Even though we initialize ->power to something semi-sane, 4826 * we leave power_orig unset. This allows us to detect if 4827 * domain iteration is still funny without causing /0 traps. 4828 */ 4829 if (!group->sgp->power_orig) { 4830 printk(KERN_CONT "\n"); 4831 printk(KERN_ERR "ERROR: domain->cpu_power not " 4832 "set\n"); 4833 break; 4834 } 4835 4836 if (!cpumask_weight(sched_group_cpus(group))) { 4837 printk(KERN_CONT "\n"); 4838 printk(KERN_ERR "ERROR: empty group\n"); 4839 break; 4840 } 4841 4842 if (!(sd->flags & SD_OVERLAP) && 4843 cpumask_intersects(groupmask, sched_group_cpus(group))) { 4844 printk(KERN_CONT "\n"); 4845 printk(KERN_ERR "ERROR: repeated CPUs\n"); 4846 break; 4847 } 4848 4849 cpumask_or(groupmask, groupmask, sched_group_cpus(group)); 4850 4851 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group)); 4852 4853 printk(KERN_CONT " %s", str); 4854 if (group->sgp->power != SCHED_POWER_SCALE) { 4855 printk(KERN_CONT " (cpu_power = %d)", 4856 group->sgp->power); 4857 } 4858 4859 group = group->next; 4860 } while (group != sd->groups); 4861 printk(KERN_CONT "\n"); 4862 4863 if (!cpumask_equal(sched_domain_span(sd), groupmask)) 4864 printk(KERN_ERR "ERROR: groups don't span domain->span\n"); 4865 4866 if (sd->parent && 4867 !cpumask_subset(groupmask, sched_domain_span(sd->parent))) 4868 printk(KERN_ERR "ERROR: parent span is not a superset " 4869 "of domain->span\n"); 4870 return 0; 4871} 4872 4873static void sched_domain_debug(struct sched_domain *sd, int cpu) 4874{ 4875 int level = 0; 4876 4877 if (!sched_debug_enabled) 4878 return; 4879 4880 if (!sd) { 4881 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu); 4882 return; 4883 } 4884 4885 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu); 4886 4887 for (;;) { 4888 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask)) 4889 break; 4890 level++; 4891 sd = sd->parent; 4892 if (!sd) 4893 break; 4894 } 4895} 4896#else /* !CONFIG_SCHED_DEBUG */ 4897# define sched_domain_debug(sd, cpu) do { } while (0) 4898static inline bool sched_debug(void) 4899{ 4900 return false; 4901} 4902#endif /* CONFIG_SCHED_DEBUG */ 4903 4904static int sd_degenerate(struct sched_domain *sd) 4905{ 4906 if (cpumask_weight(sched_domain_span(sd)) == 1) 4907 return 1; 4908 4909 /* Following flags need at least 2 groups */ 4910 if (sd->flags & (SD_LOAD_BALANCE | 4911 SD_BALANCE_NEWIDLE | 4912 SD_BALANCE_FORK | 4913 SD_BALANCE_EXEC | 4914 SD_SHARE_CPUPOWER | 4915 SD_SHARE_PKG_RESOURCES)) { 4916 if (sd->groups != sd->groups->next) 4917 return 0; 4918 } 4919 4920 /* Following flags don't use groups */ 4921 if (sd->flags & (SD_WAKE_AFFINE)) 4922 return 0; 4923 4924 return 1; 4925} 4926 4927static int 4928sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent) 4929{ 4930 unsigned long cflags = sd->flags, pflags = parent->flags; 4931 4932 if (sd_degenerate(parent)) 4933 return 1; 4934 4935 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent))) 4936 return 0; 4937 4938 /* Flags needing groups don't count if only 1 group in parent */ 4939 if (parent->groups == parent->groups->next) { 4940 pflags &= ~(SD_LOAD_BALANCE | 4941 SD_BALANCE_NEWIDLE | 4942 SD_BALANCE_FORK | 4943 SD_BALANCE_EXEC | 4944 SD_SHARE_CPUPOWER | 4945 SD_SHARE_PKG_RESOURCES | 4946 SD_PREFER_SIBLING); 4947 if (nr_node_ids == 1) 4948 pflags &= ~SD_SERIALIZE; 4949 } 4950 if (~cflags & pflags) 4951 return 0; 4952 4953 return 1; 4954} 4955 4956static void free_rootdomain(struct rcu_head *rcu) 4957{ 4958 struct root_domain *rd = container_of(rcu, struct root_domain, rcu); 4959 4960 cpupri_cleanup(&rd->cpupri); 4961 free_cpumask_var(rd->rto_mask); 4962 free_cpumask_var(rd->online); 4963 free_cpumask_var(rd->span); 4964 kfree(rd); 4965} 4966 4967static void rq_attach_root(struct rq *rq, struct root_domain *rd) 4968{ 4969 struct root_domain *old_rd = NULL; 4970 unsigned long flags; 4971 4972 raw_spin_lock_irqsave(&rq->lock, flags); 4973 4974 if (rq->rd) { 4975 old_rd = rq->rd; 4976 4977 if (cpumask_test_cpu(rq->cpu, old_rd->online)) 4978 set_rq_offline(rq); 4979 4980 cpumask_clear_cpu(rq->cpu, old_rd->span); 4981 4982 /* 4983 * If we dont want to free the old_rt yet then 4984 * set old_rd to NULL to skip the freeing later 4985 * in this function: 4986 */ 4987 if (!atomic_dec_and_test(&old_rd->refcount)) 4988 old_rd = NULL; 4989 } 4990 4991 atomic_inc(&rd->refcount); 4992 rq->rd = rd; 4993 4994 cpumask_set_cpu(rq->cpu, rd->span); 4995 if (cpumask_test_cpu(rq->cpu, cpu_active_mask)) 4996 set_rq_online(rq); 4997 4998 raw_spin_unlock_irqrestore(&rq->lock, flags); 4999 5000 if (old_rd) 5001 call_rcu_sched(&old_rd->rcu, free_rootdomain); 5002} 5003 5004static int init_rootdomain(struct root_domain *rd) 5005{ 5006 memset(rd, 0, sizeof(*rd)); 5007 5008 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL)) 5009 goto out; 5010 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL)) 5011 goto free_span; 5012 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL)) 5013 goto free_online; 5014 5015 if (cpupri_init(&rd->cpupri) != 0) 5016 goto free_rto_mask; 5017 return 0; 5018 5019free_rto_mask: 5020 free_cpumask_var(rd->rto_mask); 5021free_online: 5022 free_cpumask_var(rd->online); 5023free_span: 5024 free_cpumask_var(rd->span); 5025out: 5026 return -ENOMEM; 5027} 5028 5029/* 5030 * By default the system creates a single root-domain with all cpus as 5031 * members (mimicking the global state we have today). 5032 */ 5033struct root_domain def_root_domain; 5034 5035static void init_defrootdomain(void) 5036{ 5037 init_rootdomain(&def_root_domain); 5038 5039 atomic_set(&def_root_domain.refcount, 1); 5040} 5041 5042static struct root_domain *alloc_rootdomain(void) 5043{ 5044 struct root_domain *rd; 5045 5046 rd = kmalloc(sizeof(*rd), GFP_KERNEL); 5047 if (!rd) 5048 return NULL; 5049 5050 if (init_rootdomain(rd) != 0) { 5051 kfree(rd); 5052 return NULL; 5053 } 5054 5055 return rd; 5056} 5057 5058static void free_sched_groups(struct sched_group *sg, int free_sgp) 5059{ 5060 struct sched_group *tmp, *first; 5061 5062 if (!sg) 5063 return; 5064 5065 first = sg; 5066 do { 5067 tmp = sg->next; 5068 5069 if (free_sgp && atomic_dec_and_test(&sg->sgp->ref)) 5070 kfree(sg->sgp); 5071 5072 kfree(sg); 5073 sg = tmp; 5074 } while (sg != first); 5075} 5076 5077static void free_sched_domain(struct rcu_head *rcu) 5078{ 5079 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu); 5080 5081 /* 5082 * If its an overlapping domain it has private groups, iterate and 5083 * nuke them all. 5084 */ 5085 if (sd->flags & SD_OVERLAP) { 5086 free_sched_groups(sd->groups, 1); 5087 } else if (atomic_dec_and_test(&sd->groups->ref)) { 5088 kfree(sd->groups->sgp); 5089 kfree(sd->groups); 5090 } 5091 kfree(sd); 5092} 5093 5094static void destroy_sched_domain(struct sched_domain *sd, int cpu) 5095{ 5096 call_rcu(&sd->rcu, free_sched_domain); 5097} 5098 5099static void destroy_sched_domains(struct sched_domain *sd, int cpu) 5100{ 5101 for (; sd; sd = sd->parent) 5102 destroy_sched_domain(sd, cpu); 5103} 5104 5105/* 5106 * Keep a special pointer to the highest sched_domain that has 5107 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this 5108 * allows us to avoid some pointer chasing select_idle_sibling(). 5109 * 5110 * Also keep a unique ID per domain (we use the first cpu number in 5111 * the cpumask of the domain), this allows us to quickly tell if 5112 * two cpus are in the same cache domain, see cpus_share_cache(). 5113 */ 5114DEFINE_PER_CPU(struct sched_domain *, sd_llc); 5115DEFINE_PER_CPU(int, sd_llc_size); 5116DEFINE_PER_CPU(int, sd_llc_id); 5117 5118static void update_top_cache_domain(int cpu) 5119{ 5120 struct sched_domain *sd; 5121 int id = cpu; 5122 int size = 1; 5123 5124 sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES); 5125 if (sd) { 5126 id = cpumask_first(sched_domain_span(sd)); 5127 size = cpumask_weight(sched_domain_span(sd)); 5128 } 5129 5130 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd); 5131 per_cpu(sd_llc_size, cpu) = size; 5132 per_cpu(sd_llc_id, cpu) = id; 5133} 5134 5135/* 5136 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must 5137 * hold the hotplug lock. 5138 */ 5139static void 5140cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu) 5141{ 5142 struct rq *rq = cpu_rq(cpu); 5143 struct sched_domain *tmp; 5144 5145 /* Remove the sched domains which do not contribute to scheduling. */ 5146 for (tmp = sd; tmp; ) { 5147 struct sched_domain *parent = tmp->parent; 5148 if (!parent) 5149 break; 5150 5151 if (sd_parent_degenerate(tmp, parent)) { 5152 tmp->parent = parent->parent; 5153 if (parent->parent) 5154 parent->parent->child = tmp; 5155 /* 5156 * Transfer SD_PREFER_SIBLING down in case of a 5157 * degenerate parent; the spans match for this 5158 * so the property transfers. 5159 */ 5160 if (parent->flags & SD_PREFER_SIBLING) 5161 tmp->flags |= SD_PREFER_SIBLING; 5162 destroy_sched_domain(parent, cpu); 5163 } else 5164 tmp = tmp->parent; 5165 } 5166 5167 if (sd && sd_degenerate(sd)) { 5168 tmp = sd; 5169 sd = sd->parent; 5170 destroy_sched_domain(tmp, cpu); 5171 if (sd) 5172 sd->child = NULL; 5173 } 5174 5175 sched_domain_debug(sd, cpu); 5176 5177 rq_attach_root(rq, rd); 5178 tmp = rq->sd; 5179 rcu_assign_pointer(rq->sd, sd); 5180 destroy_sched_domains(tmp, cpu); 5181 5182 update_top_cache_domain(cpu); 5183} 5184 5185/* cpus with isolated domains */ 5186static cpumask_var_t cpu_isolated_map; 5187 5188/* Setup the mask of cpus configured for isolated domains */ 5189static int __init isolated_cpu_setup(char *str) 5190{ 5191 alloc_bootmem_cpumask_var(&cpu_isolated_map); 5192 cpulist_parse(str, cpu_isolated_map); 5193 return 1; 5194} 5195 5196__setup("isolcpus=", isolated_cpu_setup); 5197 5198static const struct cpumask *cpu_cpu_mask(int cpu) 5199{ 5200 return cpumask_of_node(cpu_to_node(cpu)); 5201} 5202 5203struct sd_data { 5204 struct sched_domain **__percpu sd; 5205 struct sched_group **__percpu sg; 5206 struct sched_group_power **__percpu sgp; 5207}; 5208 5209struct s_data { 5210 struct sched_domain ** __percpu sd; 5211 struct root_domain *rd; 5212}; 5213 5214enum s_alloc { 5215 sa_rootdomain, 5216 sa_sd, 5217 sa_sd_storage, 5218 sa_none, 5219}; 5220 5221struct sched_domain_topology_level; 5222 5223typedef struct sched_domain *(*sched_domain_init_f)(struct sched_domain_topology_level *tl, int cpu); 5224typedef const struct cpumask *(*sched_domain_mask_f)(int cpu); 5225 5226#define SDTL_OVERLAP 0x01 5227 5228struct sched_domain_topology_level { 5229 sched_domain_init_f init; 5230 sched_domain_mask_f mask; 5231 int flags; 5232 int numa_level; 5233 struct sd_data data; 5234}; 5235 5236/* 5237 * Build an iteration mask that can exclude certain CPUs from the upwards 5238 * domain traversal. 5239 * 5240 * Asymmetric node setups can result in situations where the domain tree is of 5241 * unequal depth, make sure to skip domains that already cover the entire 5242 * range. 5243 * 5244 * In that case build_sched_domains() will have terminated the iteration early 5245 * and our sibling sd spans will be empty. Domains should always include the 5246 * cpu they're built on, so check that. 5247 * 5248 */ 5249static void build_group_mask(struct sched_domain *sd, struct sched_group *sg) 5250{ 5251 const struct cpumask *span = sched_domain_span(sd); 5252 struct sd_data *sdd = sd->private; 5253 struct sched_domain *sibling; 5254 int i; 5255 5256 for_each_cpu(i, span) { 5257 sibling = *per_cpu_ptr(sdd->sd, i); 5258 if (!cpumask_test_cpu(i, sched_domain_span(sibling))) 5259 continue; 5260 5261 cpumask_set_cpu(i, sched_group_mask(sg)); 5262 } 5263} 5264 5265/* 5266 * Return the canonical balance cpu for this group, this is the first cpu 5267 * of this group that's also in the iteration mask. 5268 */ 5269int group_balance_cpu(struct sched_group *sg) 5270{ 5271 return cpumask_first_and(sched_group_cpus(sg), sched_group_mask(sg)); 5272} 5273 5274static int 5275build_overlap_sched_groups(struct sched_domain *sd, int cpu) 5276{ 5277 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg; 5278 const struct cpumask *span = sched_domain_span(sd); 5279 struct cpumask *covered = sched_domains_tmpmask; 5280 struct sd_data *sdd = sd->private; 5281 struct sched_domain *child; 5282 int i; 5283 5284 cpumask_clear(covered); 5285 5286 for_each_cpu(i, span) { 5287 struct cpumask *sg_span; 5288 5289 if (cpumask_test_cpu(i, covered)) 5290 continue; 5291 5292 child = *per_cpu_ptr(sdd->sd, i); 5293 5294 /* See the comment near build_group_mask(). */ 5295 if (!cpumask_test_cpu(i, sched_domain_span(child))) 5296 continue; 5297 5298 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(), 5299 GFP_KERNEL, cpu_to_node(cpu)); 5300 5301 if (!sg) 5302 goto fail; 5303 5304 sg_span = sched_group_cpus(sg); 5305 if (child->child) { 5306 child = child->child; 5307 cpumask_copy(sg_span, sched_domain_span(child)); 5308 } else 5309 cpumask_set_cpu(i, sg_span); 5310 5311 cpumask_or(covered, covered, sg_span); 5312 5313 sg->sgp = *per_cpu_ptr(sdd->sgp, i); 5314 if (atomic_inc_return(&sg->sgp->ref) == 1) 5315 build_group_mask(sd, sg); 5316 5317 /* 5318 * Initialize sgp->power such that even if we mess up the 5319 * domains and no possible iteration will get us here, we won't 5320 * die on a /0 trap. 5321 */ 5322 sg->sgp->power = SCHED_POWER_SCALE * cpumask_weight(sg_span); 5323 5324 /* 5325 * Make sure the first group of this domain contains the 5326 * canonical balance cpu. Otherwise the sched_domain iteration 5327 * breaks. See update_sg_lb_stats(). 5328 */ 5329 if ((!groups && cpumask_test_cpu(cpu, sg_span)) || 5330 group_balance_cpu(sg) == cpu) 5331 groups = sg; 5332 5333 if (!first) 5334 first = sg; 5335 if (last) 5336 last->next = sg; 5337 last = sg; 5338 last->next = first; 5339 } 5340 sd->groups = groups; 5341 5342 return 0; 5343 5344fail: 5345 free_sched_groups(first, 0); 5346 5347 return -ENOMEM; 5348} 5349 5350static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg) 5351{ 5352 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu); 5353 struct sched_domain *child = sd->child; 5354 5355 if (child) 5356 cpu = cpumask_first(sched_domain_span(child)); 5357 5358 if (sg) { 5359 *sg = *per_cpu_ptr(sdd->sg, cpu); 5360 (*sg)->sgp = *per_cpu_ptr(sdd->sgp, cpu); 5361 atomic_set(&(*sg)->sgp->ref, 1); /* for claim_allocations */ 5362 } 5363 5364 return cpu; 5365} 5366 5367/* 5368 * build_sched_groups will build a circular linked list of the groups 5369 * covered by the given span, and will set each group's ->cpumask correctly, 5370 * and ->cpu_power to 0. 5371 * 5372 * Assumes the sched_domain tree is fully constructed 5373 */ 5374static int 5375build_sched_groups(struct sched_domain *sd, int cpu) 5376{ 5377 struct sched_group *first = NULL, *last = NULL; 5378 struct sd_data *sdd = sd->private; 5379 const struct cpumask *span = sched_domain_span(sd); 5380 struct cpumask *covered; 5381 int i; 5382 5383 get_group(cpu, sdd, &sd->groups); 5384 atomic_inc(&sd->groups->ref); 5385 5386 if (cpu != cpumask_first(span)) 5387 return 0; 5388 5389 lockdep_assert_held(&sched_domains_mutex); 5390 covered = sched_domains_tmpmask; 5391 5392 cpumask_clear(covered); 5393 5394 for_each_cpu(i, span) { 5395 struct sched_group *sg; 5396 int group, j; 5397 5398 if (cpumask_test_cpu(i, covered)) 5399 continue; 5400 5401 group = get_group(i, sdd, &sg); 5402 cpumask_clear(sched_group_cpus(sg)); 5403 sg->sgp->power = 0; 5404 cpumask_setall(sched_group_mask(sg)); 5405 5406 for_each_cpu(j, span) { 5407 if (get_group(j, sdd, NULL) != group) 5408 continue; 5409 5410 cpumask_set_cpu(j, covered); 5411 cpumask_set_cpu(j, sched_group_cpus(sg)); 5412 } 5413 5414 if (!first) 5415 first = sg; 5416 if (last) 5417 last->next = sg; 5418 last = sg; 5419 } 5420 last->next = first; 5421 5422 return 0; 5423} 5424 5425/* 5426 * Initialize sched groups cpu_power. 5427 * 5428 * cpu_power indicates the capacity of sched group, which is used while 5429 * distributing the load between different sched groups in a sched domain. 5430 * Typically cpu_power for all the groups in a sched domain will be same unless 5431 * there are asymmetries in the topology. If there are asymmetries, group 5432 * having more cpu_power will pickup more load compared to the group having 5433 * less cpu_power. 5434 */ 5435static void init_sched_groups_power(int cpu, struct sched_domain *sd) 5436{ 5437 struct sched_group *sg = sd->groups; 5438 5439 WARN_ON(!sg); 5440 5441 do { 5442 sg->group_weight = cpumask_weight(sched_group_cpus(sg)); 5443 sg = sg->next; 5444 } while (sg != sd->groups); 5445 5446 if (cpu != group_balance_cpu(sg)) 5447 return; 5448 5449 update_group_power(sd, cpu); 5450 atomic_set(&sg->sgp->nr_busy_cpus, sg->group_weight); 5451} 5452 5453int __weak arch_sd_sibling_asym_packing(void) 5454{ 5455 return 0*SD_ASYM_PACKING; 5456} 5457 5458/* 5459 * Initializers for schedule domains 5460 * Non-inlined to reduce accumulated stack pressure in build_sched_domains() 5461 */ 5462 5463#ifdef CONFIG_SCHED_DEBUG 5464# define SD_INIT_NAME(sd, type) sd->name = #type 5465#else 5466# define SD_INIT_NAME(sd, type) do { } while (0) 5467#endif 5468 5469#define SD_INIT_FUNC(type) \ 5470static noinline struct sched_domain * \ 5471sd_init_##type(struct sched_domain_topology_level *tl, int cpu) \ 5472{ \ 5473 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); \ 5474 *sd = SD_##type##_INIT; \ 5475 SD_INIT_NAME(sd, type); \ 5476 sd->private = &tl->data; \ 5477 return sd; \ 5478} 5479 5480SD_INIT_FUNC(CPU) 5481#ifdef CONFIG_SCHED_SMT 5482 SD_INIT_FUNC(SIBLING) 5483#endif 5484#ifdef CONFIG_SCHED_MC 5485 SD_INIT_FUNC(MC) 5486#endif 5487#ifdef CONFIG_SCHED_BOOK 5488 SD_INIT_FUNC(BOOK) 5489#endif 5490 5491static int default_relax_domain_level = -1; 5492int sched_domain_level_max; 5493 5494static int __init setup_relax_domain_level(char *str) 5495{ 5496 if (kstrtoint(str, 0, &default_relax_domain_level)) 5497 pr_warn("Unable to set relax_domain_level\n"); 5498 5499 return 1; 5500} 5501__setup("relax_domain_level=", setup_relax_domain_level); 5502 5503static void set_domain_attribute(struct sched_domain *sd, 5504 struct sched_domain_attr *attr) 5505{ 5506 int request; 5507 5508 if (!attr || attr->relax_domain_level < 0) { 5509 if (default_relax_domain_level < 0) 5510 return; 5511 else 5512 request = default_relax_domain_level; 5513 } else 5514 request = attr->relax_domain_level; 5515 if (request < sd->level) { 5516 /* turn off idle balance on this domain */ 5517 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE); 5518 } else { 5519 /* turn on idle balance on this domain */ 5520 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE); 5521 } 5522} 5523 5524static void __sdt_free(const struct cpumask *cpu_map); 5525static int __sdt_alloc(const struct cpumask *cpu_map); 5526 5527static void __free_domain_allocs(struct s_data *d, enum s_alloc what, 5528 const struct cpumask *cpu_map) 5529{ 5530 switch (what) { 5531 case sa_rootdomain: 5532 if (!atomic_read(&d->rd->refcount)) 5533 free_rootdomain(&d->rd->rcu); /* fall through */ 5534 case sa_sd: 5535 free_percpu(d->sd); /* fall through */ 5536 case sa_sd_storage: 5537 __sdt_free(cpu_map); /* fall through */ 5538 case sa_none: 5539 break; 5540 } 5541} 5542 5543static enum s_alloc __visit_domain_allocation_hell(struct s_data *d, 5544 const struct cpumask *cpu_map) 5545{ 5546 memset(d, 0, sizeof(*d)); 5547 5548 if (__sdt_alloc(cpu_map)) 5549 return sa_sd_storage; 5550 d->sd = alloc_percpu(struct sched_domain *); 5551 if (!d->sd) 5552 return sa_sd_storage; 5553 d->rd = alloc_rootdomain(); 5554 if (!d->rd) 5555 return sa_sd; 5556 return sa_rootdomain; 5557} 5558 5559/* 5560 * NULL the sd_data elements we've used to build the sched_domain and 5561 * sched_group structure so that the subsequent __free_domain_allocs() 5562 * will not free the data we're using. 5563 */ 5564static void claim_allocations(int cpu, struct sched_domain *sd) 5565{ 5566 struct sd_data *sdd = sd->private; 5567 5568 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd); 5569 *per_cpu_ptr(sdd->sd, cpu) = NULL; 5570 5571 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref)) 5572 *per_cpu_ptr(sdd->sg, cpu) = NULL; 5573 5574 if (atomic_read(&(*per_cpu_ptr(sdd->sgp, cpu))->ref)) 5575 *per_cpu_ptr(sdd->sgp, cpu) = NULL; 5576} 5577 5578#ifdef CONFIG_SCHED_SMT 5579static const struct cpumask *cpu_smt_mask(int cpu) 5580{ 5581 return topology_thread_cpumask(cpu); 5582} 5583#endif 5584 5585/* 5586 * Topology list, bottom-up. 5587 */ 5588static struct sched_domain_topology_level default_topology[] = { 5589#ifdef CONFIG_SCHED_SMT 5590 { sd_init_SIBLING, cpu_smt_mask, }, 5591#endif 5592#ifdef CONFIG_SCHED_MC 5593 { sd_init_MC, cpu_coregroup_mask, }, 5594#endif 5595#ifdef CONFIG_SCHED_BOOK 5596 { sd_init_BOOK, cpu_book_mask, }, 5597#endif 5598 { sd_init_CPU, cpu_cpu_mask, }, 5599 { NULL, }, 5600}; 5601 5602static struct sched_domain_topology_level *sched_domain_topology = default_topology; 5603 5604#define for_each_sd_topology(tl) \ 5605 for (tl = sched_domain_topology; tl->init; tl++) 5606 5607#ifdef CONFIG_NUMA 5608 5609static int sched_domains_numa_levels; 5610static int *sched_domains_numa_distance; 5611static struct cpumask ***sched_domains_numa_masks; 5612static int sched_domains_curr_level; 5613 5614static inline int sd_local_flags(int level) 5615{ 5616 if (sched_domains_numa_distance[level] > RECLAIM_DISTANCE) 5617 return 0; 5618 5619 return SD_BALANCE_EXEC | SD_BALANCE_FORK | SD_WAKE_AFFINE; 5620} 5621 5622static struct sched_domain * 5623sd_numa_init(struct sched_domain_topology_level *tl, int cpu) 5624{ 5625 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); 5626 int level = tl->numa_level; 5627 int sd_weight = cpumask_weight( 5628 sched_domains_numa_masks[level][cpu_to_node(cpu)]); 5629 5630 *sd = (struct sched_domain){ 5631 .min_interval = sd_weight, 5632 .max_interval = 2*sd_weight, 5633 .busy_factor = 32, 5634 .imbalance_pct = 125, 5635 .cache_nice_tries = 2, 5636 .busy_idx = 3, 5637 .idle_idx = 2, 5638 .newidle_idx = 0, 5639 .wake_idx = 0, 5640 .forkexec_idx = 0, 5641 5642 .flags = 1*SD_LOAD_BALANCE 5643 | 1*SD_BALANCE_NEWIDLE 5644 | 0*SD_BALANCE_EXEC 5645 | 0*SD_BALANCE_FORK 5646 | 0*SD_BALANCE_WAKE 5647 | 0*SD_WAKE_AFFINE 5648 | 0*SD_SHARE_CPUPOWER 5649 | 0*SD_SHARE_PKG_RESOURCES 5650 | 1*SD_SERIALIZE 5651 | 0*SD_PREFER_SIBLING 5652 | sd_local_flags(level) 5653 , 5654 .last_balance = jiffies, 5655 .balance_interval = sd_weight, 5656 }; 5657 SD_INIT_NAME(sd, NUMA); 5658 sd->private = &tl->data; 5659 5660 /* 5661 * Ugly hack to pass state to sd_numa_mask()... 5662 */ 5663 sched_domains_curr_level = tl->numa_level; 5664 5665 return sd; 5666} 5667 5668static const struct cpumask *sd_numa_mask(int cpu) 5669{ 5670 return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)]; 5671} 5672 5673static void sched_numa_warn(const char *str) 5674{ 5675 static int done = false; 5676 int i,j; 5677 5678 if (done) 5679 return; 5680 5681 done = true; 5682 5683 printk(KERN_WARNING "ERROR: %s\n\n", str); 5684 5685 for (i = 0; i < nr_node_ids; i++) { 5686 printk(KERN_WARNING " "); 5687 for (j = 0; j < nr_node_ids; j++) 5688 printk(KERN_CONT "%02d ", node_distance(i,j)); 5689 printk(KERN_CONT "\n"); 5690 } 5691 printk(KERN_WARNING "\n"); 5692} 5693 5694static bool find_numa_distance(int distance) 5695{ 5696 int i; 5697 5698 if (distance == node_distance(0, 0)) 5699 return true; 5700 5701 for (i = 0; i < sched_domains_numa_levels; i++) { 5702 if (sched_domains_numa_distance[i] == distance) 5703 return true; 5704 } 5705 5706 return false; 5707} 5708 5709static void sched_init_numa(void) 5710{ 5711 int next_distance, curr_distance = node_distance(0, 0); 5712 struct sched_domain_topology_level *tl; 5713 int level = 0; 5714 int i, j, k; 5715 5716 sched_domains_numa_distance = kzalloc(sizeof(int) * nr_node_ids, GFP_KERNEL); 5717 if (!sched_domains_numa_distance) 5718 return; 5719 5720 /* 5721 * O(nr_nodes^2) deduplicating selection sort -- in order to find the 5722 * unique distances in the node_distance() table. 5723 * 5724 * Assumes node_distance(0,j) includes all distances in 5725 * node_distance(i,j) in order to avoid cubic time. 5726 */ 5727 next_distance = curr_distance; 5728 for (i = 0; i < nr_node_ids; i++) { 5729 for (j = 0; j < nr_node_ids; j++) { 5730 for (k = 0; k < nr_node_ids; k++) { 5731 int distance = node_distance(i, k); 5732 5733 if (distance > curr_distance && 5734 (distance < next_distance || 5735 next_distance == curr_distance)) 5736 next_distance = distance; 5737 5738 /* 5739 * While not a strong assumption it would be nice to know 5740 * about cases where if node A is connected to B, B is not 5741 * equally connected to A. 5742 */ 5743 if (sched_debug() && node_distance(k, i) != distance) 5744 sched_numa_warn("Node-distance not symmetric"); 5745 5746 if (sched_debug() && i && !find_numa_distance(distance)) 5747 sched_numa_warn("Node-0 not representative"); 5748 } 5749 if (next_distance != curr_distance) { 5750 sched_domains_numa_distance[level++] = next_distance; 5751 sched_domains_numa_levels = level; 5752 curr_distance = next_distance; 5753 } else break; 5754 } 5755 5756 /* 5757 * In case of sched_debug() we verify the above assumption. 5758 */ 5759 if (!sched_debug()) 5760 break; 5761 } 5762 /* 5763 * 'level' contains the number of unique distances, excluding the 5764 * identity distance node_distance(i,i). 5765 * 5766 * The sched_domains_numa_distance[] array includes the actual distance 5767 * numbers. 5768 */ 5769 5770 /* 5771 * Here, we should temporarily reset sched_domains_numa_levels to 0. 5772 * If it fails to allocate memory for array sched_domains_numa_masks[][], 5773 * the array will contain less then 'level' members. This could be 5774 * dangerous when we use it to iterate array sched_domains_numa_masks[][] 5775 * in other functions. 5776 * 5777 * We reset it to 'level' at the end of this function. 5778 */ 5779 sched_domains_numa_levels = 0; 5780 5781 sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL); 5782 if (!sched_domains_numa_masks) 5783 return; 5784 5785 /* 5786 * Now for each level, construct a mask per node which contains all 5787 * cpus of nodes that are that many hops away from us. 5788 */ 5789 for (i = 0; i < level; i++) { 5790 sched_domains_numa_masks[i] = 5791 kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL); 5792 if (!sched_domains_numa_masks[i]) 5793 return; 5794 5795 for (j = 0; j < nr_node_ids; j++) { 5796 struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL); 5797 if (!mask) 5798 return; 5799 5800 sched_domains_numa_masks[i][j] = mask; 5801 5802 for (k = 0; k < nr_node_ids; k++) { 5803 if (node_distance(j, k) > sched_domains_numa_distance[i]) 5804 continue; 5805 5806 cpumask_or(mask, mask, cpumask_of_node(k)); 5807 } 5808 } 5809 } 5810 5811 tl = kzalloc((ARRAY_SIZE(default_topology) + level) * 5812 sizeof(struct sched_domain_topology_level), GFP_KERNEL); 5813 if (!tl) 5814 return; 5815 5816 /* 5817 * Copy the default topology bits.. 5818 */ 5819 for (i = 0; default_topology[i].init; i++) 5820 tl[i] = default_topology[i]; 5821 5822 /* 5823 * .. and append 'j' levels of NUMA goodness. 5824 */ 5825 for (j = 0; j < level; i++, j++) { 5826 tl[i] = (struct sched_domain_topology_level){ 5827 .init = sd_numa_init, 5828 .mask = sd_numa_mask, 5829 .flags = SDTL_OVERLAP, 5830 .numa_level = j, 5831 }; 5832 } 5833 5834 sched_domain_topology = tl; 5835 5836 sched_domains_numa_levels = level; 5837} 5838 5839static void sched_domains_numa_masks_set(int cpu) 5840{ 5841 int i, j; 5842 int node = cpu_to_node(cpu); 5843 5844 for (i = 0; i < sched_domains_numa_levels; i++) { 5845 for (j = 0; j < nr_node_ids; j++) { 5846 if (node_distance(j, node) <= sched_domains_numa_distance[i]) 5847 cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]); 5848 } 5849 } 5850} 5851 5852static void sched_domains_numa_masks_clear(int cpu) 5853{ 5854 int i, j; 5855 for (i = 0; i < sched_domains_numa_levels; i++) { 5856 for (j = 0; j < nr_node_ids; j++) 5857 cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]); 5858 } 5859} 5860 5861/* 5862 * Update sched_domains_numa_masks[level][node] array when new cpus 5863 * are onlined. 5864 */ 5865static int sched_domains_numa_masks_update(struct notifier_block *nfb, 5866 unsigned long action, 5867 void *hcpu) 5868{ 5869 int cpu = (long)hcpu; 5870 5871 switch (action & ~CPU_TASKS_FROZEN) { 5872 case CPU_ONLINE: 5873 sched_domains_numa_masks_set(cpu); 5874 break; 5875 5876 case CPU_DEAD: 5877 sched_domains_numa_masks_clear(cpu); 5878 break; 5879 5880 default: 5881 return NOTIFY_DONE; 5882 } 5883 5884 return NOTIFY_OK; 5885} 5886#else 5887static inline void sched_init_numa(void) 5888{ 5889} 5890 5891static int sched_domains_numa_masks_update(struct notifier_block *nfb, 5892 unsigned long action, 5893 void *hcpu) 5894{ 5895 return 0; 5896} 5897#endif /* CONFIG_NUMA */ 5898 5899static int __sdt_alloc(const struct cpumask *cpu_map) 5900{ 5901 struct sched_domain_topology_level *tl; 5902 int j; 5903 5904 for_each_sd_topology(tl) { 5905 struct sd_data *sdd = &tl->data; 5906 5907 sdd->sd = alloc_percpu(struct sched_domain *); 5908 if (!sdd->sd) 5909 return -ENOMEM; 5910 5911 sdd->sg = alloc_percpu(struct sched_group *); 5912 if (!sdd->sg) 5913 return -ENOMEM; 5914 5915 sdd->sgp = alloc_percpu(struct sched_group_power *); 5916 if (!sdd->sgp) 5917 return -ENOMEM; 5918 5919 for_each_cpu(j, cpu_map) { 5920 struct sched_domain *sd; 5921 struct sched_group *sg; 5922 struct sched_group_power *sgp; 5923 5924 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(), 5925 GFP_KERNEL, cpu_to_node(j)); 5926 if (!sd) 5927 return -ENOMEM; 5928 5929 *per_cpu_ptr(sdd->sd, j) = sd; 5930 5931 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(), 5932 GFP_KERNEL, cpu_to_node(j)); 5933 if (!sg) 5934 return -ENOMEM; 5935 5936 sg->next = sg; 5937 5938 *per_cpu_ptr(sdd->sg, j) = sg; 5939 5940 sgp = kzalloc_node(sizeof(struct sched_group_power) + cpumask_size(), 5941 GFP_KERNEL, cpu_to_node(j)); 5942 if (!sgp) 5943 return -ENOMEM; 5944 5945 *per_cpu_ptr(sdd->sgp, j) = sgp; 5946 } 5947 } 5948 5949 return 0; 5950} 5951 5952static void __sdt_free(const struct cpumask *cpu_map) 5953{ 5954 struct sched_domain_topology_level *tl; 5955 int j; 5956 5957 for_each_sd_topology(tl) { 5958 struct sd_data *sdd = &tl->data; 5959 5960 for_each_cpu(j, cpu_map) { 5961 struct sched_domain *sd; 5962 5963 if (sdd->sd) { 5964 sd = *per_cpu_ptr(sdd->sd, j); 5965 if (sd && (sd->flags & SD_OVERLAP)) 5966 free_sched_groups(sd->groups, 0); 5967 kfree(*per_cpu_ptr(sdd->sd, j)); 5968 } 5969 5970 if (sdd->sg) 5971 kfree(*per_cpu_ptr(sdd->sg, j)); 5972 if (sdd->sgp) 5973 kfree(*per_cpu_ptr(sdd->sgp, j)); 5974 } 5975 free_percpu(sdd->sd); 5976 sdd->sd = NULL; 5977 free_percpu(sdd->sg); 5978 sdd->sg = NULL; 5979 free_percpu(sdd->sgp); 5980 sdd->sgp = NULL; 5981 } 5982} 5983 5984struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl, 5985 const struct cpumask *cpu_map, struct sched_domain_attr *attr, 5986 struct sched_domain *child, int cpu) 5987{ 5988 struct sched_domain *sd = tl->init(tl, cpu); 5989 if (!sd) 5990 return child; 5991 5992 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu)); 5993 if (child) { 5994 sd->level = child->level + 1; 5995 sched_domain_level_max = max(sched_domain_level_max, sd->level); 5996 child->parent = sd; 5997 sd->child = child; 5998 } 5999 set_domain_attribute(sd, attr); 6000 6001 return sd; 6002} 6003 6004/* 6005 * Build sched domains for a given set of cpus and attach the sched domains 6006 * to the individual cpus 6007 */ 6008static int build_sched_domains(const struct cpumask *cpu_map, 6009 struct sched_domain_attr *attr) 6010{ 6011 enum s_alloc alloc_state; 6012 struct sched_domain *sd; 6013 struct s_data d; 6014 int i, ret = -ENOMEM; 6015 6016 alloc_state = __visit_domain_allocation_hell(&d, cpu_map); 6017 if (alloc_state != sa_rootdomain) 6018 goto error; 6019 6020 /* Set up domains for cpus specified by the cpu_map. */ 6021 for_each_cpu(i, cpu_map) { 6022 struct sched_domain_topology_level *tl; 6023 6024 sd = NULL; 6025 for_each_sd_topology(tl) { 6026 sd = build_sched_domain(tl, cpu_map, attr, sd, i); 6027 if (tl == sched_domain_topology) 6028 *per_cpu_ptr(d.sd, i) = sd; 6029 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP)) 6030 sd->flags |= SD_OVERLAP; 6031 if (cpumask_equal(cpu_map, sched_domain_span(sd))) 6032 break; 6033 } 6034 } 6035 6036 /* Build the groups for the domains */ 6037 for_each_cpu(i, cpu_map) { 6038 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) { 6039 sd->span_weight = cpumask_weight(sched_domain_span(sd)); 6040 if (sd->flags & SD_OVERLAP) { 6041 if (build_overlap_sched_groups(sd, i)) 6042 goto error; 6043 } else { 6044 if (build_sched_groups(sd, i)) 6045 goto error; 6046 } 6047 } 6048 } 6049 6050 /* Calculate CPU power for physical packages and nodes */ 6051 for (i = nr_cpumask_bits-1; i >= 0; i--) { 6052 if (!cpumask_test_cpu(i, cpu_map)) 6053 continue; 6054 6055 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) { 6056 claim_allocations(i, sd); 6057 init_sched_groups_power(i, sd); 6058 } 6059 } 6060 6061 /* Attach the domains */ 6062 rcu_read_lock(); 6063 for_each_cpu(i, cpu_map) { 6064 sd = *per_cpu_ptr(d.sd, i); 6065 cpu_attach_domain(sd, d.rd, i); 6066 } 6067 rcu_read_unlock(); 6068 6069 ret = 0; 6070error: 6071 __free_domain_allocs(&d, alloc_state, cpu_map); 6072 return ret; 6073} 6074 6075static cpumask_var_t *doms_cur; /* current sched domains */ 6076static int ndoms_cur; /* number of sched domains in 'doms_cur' */ 6077static struct sched_domain_attr *dattr_cur; 6078 /* attribues of custom domains in 'doms_cur' */ 6079 6080/* 6081 * Special case: If a kmalloc of a doms_cur partition (array of 6082 * cpumask) fails, then fallback to a single sched domain, 6083 * as determined by the single cpumask fallback_doms. 6084 */ 6085static cpumask_var_t fallback_doms; 6086 6087/* 6088 * arch_update_cpu_topology lets virtualized architectures update the 6089 * cpu core maps. It is supposed to return 1 if the topology changed 6090 * or 0 if it stayed the same. 6091 */ 6092int __attribute__((weak)) arch_update_cpu_topology(void) 6093{ 6094 return 0; 6095} 6096 6097cpumask_var_t *alloc_sched_domains(unsigned int ndoms) 6098{ 6099 int i; 6100 cpumask_var_t *doms; 6101 6102 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL); 6103 if (!doms) 6104 return NULL; 6105 for (i = 0; i < ndoms; i++) { 6106 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) { 6107 free_sched_domains(doms, i); 6108 return NULL; 6109 } 6110 } 6111 return doms; 6112} 6113 6114void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms) 6115{ 6116 unsigned int i; 6117 for (i = 0; i < ndoms; i++) 6118 free_cpumask_var(doms[i]); 6119 kfree(doms); 6120} 6121 6122/* 6123 * Set up scheduler domains and groups. Callers must hold the hotplug lock. 6124 * For now this just excludes isolated cpus, but could be used to 6125 * exclude other special cases in the future. 6126 */ 6127static int init_sched_domains(const struct cpumask *cpu_map) 6128{ 6129 int err; 6130 6131 arch_update_cpu_topology(); 6132 ndoms_cur = 1; 6133 doms_cur = alloc_sched_domains(ndoms_cur); 6134 if (!doms_cur) 6135 doms_cur = &fallback_doms; 6136 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map); 6137 err = build_sched_domains(doms_cur[0], NULL); 6138 register_sched_domain_sysctl(); 6139 6140 return err; 6141} 6142 6143/* 6144 * Detach sched domains from a group of cpus specified in cpu_map 6145 * These cpus will now be attached to the NULL domain 6146 */ 6147static void detach_destroy_domains(const struct cpumask *cpu_map) 6148{ 6149 int i; 6150 6151 rcu_read_lock(); 6152 for_each_cpu(i, cpu_map) 6153 cpu_attach_domain(NULL, &def_root_domain, i); 6154 rcu_read_unlock(); 6155} 6156 6157/* handle null as "default" */ 6158static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur, 6159 struct sched_domain_attr *new, int idx_new) 6160{ 6161 struct sched_domain_attr tmp; 6162 6163 /* fast path */ 6164 if (!new && !cur) 6165 return 1; 6166 6167 tmp = SD_ATTR_INIT; 6168 return !memcmp(cur ? (cur + idx_cur) : &tmp, 6169 new ? (new + idx_new) : &tmp, 6170 sizeof(struct sched_domain_attr)); 6171} 6172 6173/* 6174 * Partition sched domains as specified by the 'ndoms_new' 6175 * cpumasks in the array doms_new[] of cpumasks. This compares 6176 * doms_new[] to the current sched domain partitioning, doms_cur[]. 6177 * It destroys each deleted domain and builds each new domain. 6178 * 6179 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'. 6180 * The masks don't intersect (don't overlap.) We should setup one 6181 * sched domain for each mask. CPUs not in any of the cpumasks will 6182 * not be load balanced. If the same cpumask appears both in the 6183 * current 'doms_cur' domains and in the new 'doms_new', we can leave 6184 * it as it is. 6185 * 6186 * The passed in 'doms_new' should be allocated using 6187 * alloc_sched_domains. This routine takes ownership of it and will 6188 * free_sched_domains it when done with it. If the caller failed the 6189 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1, 6190 * and partition_sched_domains() will fallback to the single partition 6191 * 'fallback_doms', it also forces the domains to be rebuilt. 6192 * 6193 * If doms_new == NULL it will be replaced with cpu_online_mask. 6194 * ndoms_new == 0 is a special case for destroying existing domains, 6195 * and it will not create the default domain. 6196 * 6197 * Call with hotplug lock held 6198 */ 6199void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[], 6200 struct sched_domain_attr *dattr_new) 6201{ 6202 int i, j, n; 6203 int new_topology; 6204 6205 mutex_lock(&sched_domains_mutex); 6206 6207 /* always unregister in case we don't destroy any domains */ 6208 unregister_sched_domain_sysctl(); 6209 6210 /* Let architecture update cpu core mappings. */ 6211 new_topology = arch_update_cpu_topology(); 6212 6213 n = doms_new ? ndoms_new : 0; 6214 6215 /* Destroy deleted domains */ 6216 for (i = 0; i < ndoms_cur; i++) { 6217 for (j = 0; j < n && !new_topology; j++) { 6218 if (cpumask_equal(doms_cur[i], doms_new[j]) 6219 && dattrs_equal(dattr_cur, i, dattr_new, j)) 6220 goto match1; 6221 } 6222 /* no match - a current sched domain not in new doms_new[] */ 6223 detach_destroy_domains(doms_cur[i]); 6224match1: 6225 ; 6226 } 6227 6228 n = ndoms_cur; 6229 if (doms_new == NULL) { 6230 n = 0; 6231 doms_new = &fallback_doms; 6232 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map); 6233 WARN_ON_ONCE(dattr_new); 6234 } 6235 6236 /* Build new domains */ 6237 for (i = 0; i < ndoms_new; i++) { 6238 for (j = 0; j < n && !new_topology; j++) { 6239 if (cpumask_equal(doms_new[i], doms_cur[j]) 6240 && dattrs_equal(dattr_new, i, dattr_cur, j)) 6241 goto match2; 6242 } 6243 /* no match - add a new doms_new */ 6244 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL); 6245match2: 6246 ; 6247 } 6248 6249 /* Remember the new sched domains */ 6250 if (doms_cur != &fallback_doms) 6251 free_sched_domains(doms_cur, ndoms_cur); 6252 kfree(dattr_cur); /* kfree(NULL) is safe */ 6253 doms_cur = doms_new; 6254 dattr_cur = dattr_new; 6255 ndoms_cur = ndoms_new; 6256 6257 register_sched_domain_sysctl(); 6258 6259 mutex_unlock(&sched_domains_mutex); 6260} 6261 6262static int num_cpus_frozen; /* used to mark begin/end of suspend/resume */ 6263 6264/* 6265 * Update cpusets according to cpu_active mask. If cpusets are 6266 * disabled, cpuset_update_active_cpus() becomes a simple wrapper 6267 * around partition_sched_domains(). 6268 * 6269 * If we come here as part of a suspend/resume, don't touch cpusets because we 6270 * want to restore it back to its original state upon resume anyway. 6271 */ 6272static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action, 6273 void *hcpu) 6274{ 6275 switch (action) { 6276 case CPU_ONLINE_FROZEN: 6277 case CPU_DOWN_FAILED_FROZEN: 6278 6279 /* 6280 * num_cpus_frozen tracks how many CPUs are involved in suspend 6281 * resume sequence. As long as this is not the last online 6282 * operation in the resume sequence, just build a single sched 6283 * domain, ignoring cpusets. 6284 */ 6285 num_cpus_frozen--; 6286 if (likely(num_cpus_frozen)) { 6287 partition_sched_domains(1, NULL, NULL); 6288 break; 6289 } 6290 6291 /* 6292 * This is the last CPU online operation. So fall through and 6293 * restore the original sched domains by considering the 6294 * cpuset configurations. 6295 */ 6296 6297 case CPU_ONLINE: 6298 case CPU_DOWN_FAILED: 6299 cpuset_update_active_cpus(true); 6300 break; 6301 default: 6302 return NOTIFY_DONE; 6303 } 6304 return NOTIFY_OK; 6305} 6306 6307static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action, 6308 void *hcpu) 6309{ 6310 switch (action) { 6311 case CPU_DOWN_PREPARE: 6312 cpuset_update_active_cpus(false); 6313 break; 6314 case CPU_DOWN_PREPARE_FROZEN: 6315 num_cpus_frozen++; 6316 partition_sched_domains(1, NULL, NULL); 6317 break; 6318 default: 6319 return NOTIFY_DONE; 6320 } 6321 return NOTIFY_OK; 6322} 6323 6324void __init sched_init_smp(void) 6325{ 6326 cpumask_var_t non_isolated_cpus; 6327 6328 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL); 6329 alloc_cpumask_var(&fallback_doms, GFP_KERNEL); 6330 6331 sched_init_numa(); 6332 6333 get_online_cpus(); 6334 mutex_lock(&sched_domains_mutex); 6335 init_sched_domains(cpu_active_mask); 6336 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map); 6337 if (cpumask_empty(non_isolated_cpus)) 6338 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus); 6339 mutex_unlock(&sched_domains_mutex); 6340 put_online_cpus(); 6341 6342 hotcpu_notifier(sched_domains_numa_masks_update, CPU_PRI_SCHED_ACTIVE); 6343 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE); 6344 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE); 6345 6346 init_hrtick(); 6347 6348 /* Move init over to a non-isolated CPU */ 6349 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0) 6350 BUG(); 6351 sched_init_granularity(); 6352 free_cpumask_var(non_isolated_cpus); 6353 6354 init_sched_rt_class(); 6355} 6356#else 6357void __init sched_init_smp(void) 6358{ 6359 sched_init_granularity(); 6360} 6361#endif /* CONFIG_SMP */ 6362 6363const_debug unsigned int sysctl_timer_migration = 1; 6364 6365int in_sched_functions(unsigned long addr) 6366{ 6367 return in_lock_functions(addr) || 6368 (addr >= (unsigned long)__sched_text_start 6369 && addr < (unsigned long)__sched_text_end); 6370} 6371 6372#ifdef CONFIG_CGROUP_SCHED 6373/* 6374 * Default task group. 6375 * Every task in system belongs to this group at bootup. 6376 */ 6377struct task_group root_task_group; 6378LIST_HEAD(task_groups); 6379#endif 6380 6381DECLARE_PER_CPU(cpumask_var_t, load_balance_mask); 6382 6383void __init sched_init(void) 6384{ 6385 int i, j; 6386 unsigned long alloc_size = 0, ptr; 6387 6388#ifdef CONFIG_FAIR_GROUP_SCHED 6389 alloc_size += 2 * nr_cpu_ids * sizeof(void **); 6390#endif 6391#ifdef CONFIG_RT_GROUP_SCHED 6392 alloc_size += 2 * nr_cpu_ids * sizeof(void **); 6393#endif 6394#ifdef CONFIG_CPUMASK_OFFSTACK 6395 alloc_size += num_possible_cpus() * cpumask_size(); 6396#endif 6397 if (alloc_size) { 6398 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT); 6399 6400#ifdef CONFIG_FAIR_GROUP_SCHED 6401 root_task_group.se = (struct sched_entity **)ptr; 6402 ptr += nr_cpu_ids * sizeof(void **); 6403 6404 root_task_group.cfs_rq = (struct cfs_rq **)ptr; 6405 ptr += nr_cpu_ids * sizeof(void **); 6406 6407#endif /* CONFIG_FAIR_GROUP_SCHED */ 6408#ifdef CONFIG_RT_GROUP_SCHED 6409 root_task_group.rt_se = (struct sched_rt_entity **)ptr; 6410 ptr += nr_cpu_ids * sizeof(void **); 6411 6412 root_task_group.rt_rq = (struct rt_rq **)ptr; 6413 ptr += nr_cpu_ids * sizeof(void **); 6414 6415#endif /* CONFIG_RT_GROUP_SCHED */ 6416#ifdef CONFIG_CPUMASK_OFFSTACK 6417 for_each_possible_cpu(i) { 6418 per_cpu(load_balance_mask, i) = (void *)ptr; 6419 ptr += cpumask_size(); 6420 } 6421#endif /* CONFIG_CPUMASK_OFFSTACK */ 6422 } 6423 6424#ifdef CONFIG_SMP 6425 init_defrootdomain(); 6426#endif 6427 6428 init_rt_bandwidth(&def_rt_bandwidth, 6429 global_rt_period(), global_rt_runtime()); 6430 6431#ifdef CONFIG_RT_GROUP_SCHED 6432 init_rt_bandwidth(&root_task_group.rt_bandwidth, 6433 global_rt_period(), global_rt_runtime()); 6434#endif /* CONFIG_RT_GROUP_SCHED */ 6435 6436#ifdef CONFIG_CGROUP_SCHED 6437 list_add(&root_task_group.list, &task_groups); 6438 INIT_LIST_HEAD(&root_task_group.children); 6439 INIT_LIST_HEAD(&root_task_group.siblings); 6440 autogroup_init(&init_task); 6441 6442#endif /* CONFIG_CGROUP_SCHED */ 6443 6444 for_each_possible_cpu(i) { 6445 struct rq *rq; 6446 6447 rq = cpu_rq(i); 6448 raw_spin_lock_init(&rq->lock); 6449 rq->nr_running = 0; 6450 rq->calc_load_active = 0; 6451 rq->calc_load_update = jiffies + LOAD_FREQ; 6452 init_cfs_rq(&rq->cfs); 6453 init_rt_rq(&rq->rt, rq); 6454#ifdef CONFIG_FAIR_GROUP_SCHED 6455 root_task_group.shares = ROOT_TASK_GROUP_LOAD; 6456 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list); 6457 /* 6458 * How much cpu bandwidth does root_task_group get? 6459 * 6460 * In case of task-groups formed thr' the cgroup filesystem, it 6461 * gets 100% of the cpu resources in the system. This overall 6462 * system cpu resource is divided among the tasks of 6463 * root_task_group and its child task-groups in a fair manner, 6464 * based on each entity's (task or task-group's) weight 6465 * (se->load.weight). 6466 * 6467 * In other words, if root_task_group has 10 tasks of weight 6468 * 1024) and two child groups A0 and A1 (of weight 1024 each), 6469 * then A0's share of the cpu resource is: 6470 * 6471 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33% 6472 * 6473 * We achieve this by letting root_task_group's tasks sit 6474 * directly in rq->cfs (i.e root_task_group->se[] = NULL). 6475 */ 6476 init_cfs_bandwidth(&root_task_group.cfs_bandwidth); 6477 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL); 6478#endif /* CONFIG_FAIR_GROUP_SCHED */ 6479 6480 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime; 6481#ifdef CONFIG_RT_GROUP_SCHED 6482 INIT_LIST_HEAD(&rq->leaf_rt_rq_list); 6483 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL); 6484#endif 6485 6486 for (j = 0; j < CPU_LOAD_IDX_MAX; j++) 6487 rq->cpu_load[j] = 0; 6488 6489 rq->last_load_update_tick = jiffies; 6490 6491#ifdef CONFIG_SMP 6492 rq->sd = NULL; 6493 rq->rd = NULL; 6494 rq->cpu_power = SCHED_POWER_SCALE; 6495 rq->post_schedule = 0; 6496 rq->active_balance = 0; 6497 rq->next_balance = jiffies; 6498 rq->push_cpu = 0; 6499 rq->cpu = i; 6500 rq->online = 0; 6501 rq->idle_stamp = 0; 6502 rq->avg_idle = 2*sysctl_sched_migration_cost; 6503 rq->max_idle_balance_cost = sysctl_sched_migration_cost; 6504 6505 INIT_LIST_HEAD(&rq->cfs_tasks); 6506 6507 rq_attach_root(rq, &def_root_domain); 6508#ifdef CONFIG_NO_HZ_COMMON 6509 rq->nohz_flags = 0; 6510#endif 6511#ifdef CONFIG_NO_HZ_FULL 6512 rq->last_sched_tick = 0; 6513#endif 6514#endif 6515 init_rq_hrtick(rq); 6516 atomic_set(&rq->nr_iowait, 0); 6517 } 6518 6519 set_load_weight(&init_task); 6520 6521#ifdef CONFIG_PREEMPT_NOTIFIERS 6522 INIT_HLIST_HEAD(&init_task.preempt_notifiers); 6523#endif 6524 6525#ifdef CONFIG_RT_MUTEXES 6526 plist_head_init(&init_task.pi_waiters); 6527#endif 6528 6529 /* 6530 * The boot idle thread does lazy MMU switching as well: 6531 */ 6532 atomic_inc(&init_mm.mm_count); 6533 enter_lazy_tlb(&init_mm, current); 6534 6535 /* 6536 * Make us the idle thread. Technically, schedule() should not be 6537 * called from this thread, however somewhere below it might be, 6538 * but because we are the idle thread, we just pick up running again 6539 * when this runqueue becomes "idle". 6540 */ 6541 init_idle(current, smp_processor_id()); 6542 6543 calc_load_update = jiffies + LOAD_FREQ; 6544 6545 /* 6546 * During early bootup we pretend to be a normal task: 6547 */ 6548 current->sched_class = &fair_sched_class; 6549 6550#ifdef CONFIG_SMP 6551 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT); 6552 /* May be allocated at isolcpus cmdline parse time */ 6553 if (cpu_isolated_map == NULL) 6554 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT); 6555 idle_thread_set_boot_cpu(); 6556#endif 6557 init_sched_fair_class(); 6558 6559 scheduler_running = 1; 6560} 6561 6562#ifdef CONFIG_DEBUG_ATOMIC_SLEEP 6563static inline int preempt_count_equals(int preempt_offset) 6564{ 6565 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth(); 6566 6567 return (nested == preempt_offset); 6568} 6569 6570void __might_sleep(const char *file, int line, int preempt_offset) 6571{ 6572 static unsigned long prev_jiffy; /* ratelimiting */ 6573 6574 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */ 6575 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) || 6576 system_state != SYSTEM_RUNNING || oops_in_progress) 6577 return; 6578 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy) 6579 return; 6580 prev_jiffy = jiffies; 6581 6582 printk(KERN_ERR 6583 "BUG: sleeping function called from invalid context at %s:%d\n", 6584 file, line); 6585 printk(KERN_ERR 6586 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n", 6587 in_atomic(), irqs_disabled(), 6588 current->pid, current->comm); 6589 6590 debug_show_held_locks(current); 6591 if (irqs_disabled()) 6592 print_irqtrace_events(current); 6593 dump_stack(); 6594} 6595EXPORT_SYMBOL(__might_sleep); 6596#endif 6597 6598#ifdef CONFIG_MAGIC_SYSRQ 6599static void normalize_task(struct rq *rq, struct task_struct *p) 6600{ 6601 const struct sched_class *prev_class = p->sched_class; 6602 int old_prio = p->prio; 6603 int on_rq; 6604 6605 on_rq = p->on_rq; 6606 if (on_rq) 6607 dequeue_task(rq, p, 0); 6608 __setscheduler(rq, p, SCHED_NORMAL, 0); 6609 if (on_rq) { 6610 enqueue_task(rq, p, 0); 6611 resched_task(rq->curr); 6612 } 6613 6614 check_class_changed(rq, p, prev_class, old_prio); 6615} 6616 6617void normalize_rt_tasks(void) 6618{ 6619 struct task_struct *g, *p; 6620 unsigned long flags; 6621 struct rq *rq; 6622 6623 read_lock_irqsave(&tasklist_lock, flags); 6624 do_each_thread(g, p) { 6625 /* 6626 * Only normalize user tasks: 6627 */ 6628 if (!p->mm) 6629 continue; 6630 6631 p->se.exec_start = 0; 6632#ifdef CONFIG_SCHEDSTATS 6633 p->se.statistics.wait_start = 0; 6634 p->se.statistics.sleep_start = 0; 6635 p->se.statistics.block_start = 0; 6636#endif 6637 6638 if (!rt_task(p)) { 6639 /* 6640 * Renice negative nice level userspace 6641 * tasks back to 0: 6642 */ 6643 if (TASK_NICE(p) < 0 && p->mm) 6644 set_user_nice(p, 0); 6645 continue; 6646 } 6647 6648 raw_spin_lock(&p->pi_lock); 6649 rq = __task_rq_lock(p); 6650 6651 normalize_task(rq, p); 6652 6653 __task_rq_unlock(rq); 6654 raw_spin_unlock(&p->pi_lock); 6655 } while_each_thread(g, p); 6656 6657 read_unlock_irqrestore(&tasklist_lock, flags); 6658} 6659 6660#endif /* CONFIG_MAGIC_SYSRQ */ 6661 6662#if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) 6663/* 6664 * These functions are only useful for the IA64 MCA handling, or kdb. 6665 * 6666 * They can only be called when the whole system has been 6667 * stopped - every CPU needs to be quiescent, and no scheduling 6668 * activity can take place. Using them for anything else would 6669 * be a serious bug, and as a result, they aren't even visible 6670 * under any other configuration. 6671 */ 6672 6673/** 6674 * curr_task - return the current task for a given cpu. 6675 * @cpu: the processor in question. 6676 * 6677 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED! 6678 * 6679 * Return: The current task for @cpu. 6680 */ 6681struct task_struct *curr_task(int cpu) 6682{ 6683 return cpu_curr(cpu); 6684} 6685 6686#endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */ 6687 6688#ifdef CONFIG_IA64 6689/** 6690 * set_curr_task - set the current task for a given cpu. 6691 * @cpu: the processor in question. 6692 * @p: the task pointer to set. 6693 * 6694 * Description: This function must only be used when non-maskable interrupts 6695 * are serviced on a separate stack. It allows the architecture to switch the 6696 * notion of the current task on a cpu in a non-blocking manner. This function 6697 * must be called with all CPU's synchronized, and interrupts disabled, the 6698 * and caller must save the original value of the current task (see 6699 * curr_task() above) and restore that value before reenabling interrupts and 6700 * re-starting the system. 6701 * 6702 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED! 6703 */ 6704void set_curr_task(int cpu, struct task_struct *p) 6705{ 6706 cpu_curr(cpu) = p; 6707} 6708 6709#endif 6710 6711#ifdef CONFIG_CGROUP_SCHED 6712/* task_group_lock serializes the addition/removal of task groups */ 6713static DEFINE_SPINLOCK(task_group_lock); 6714 6715static void free_sched_group(struct task_group *tg) 6716{ 6717 free_fair_sched_group(tg); 6718 free_rt_sched_group(tg); 6719 autogroup_free(tg); 6720 kfree(tg); 6721} 6722 6723/* allocate runqueue etc for a new task group */ 6724struct task_group *sched_create_group(struct task_group *parent) 6725{ 6726 struct task_group *tg; 6727 6728 tg = kzalloc(sizeof(*tg), GFP_KERNEL); 6729 if (!tg) 6730 return ERR_PTR(-ENOMEM); 6731 6732 if (!alloc_fair_sched_group(tg, parent)) 6733 goto err; 6734 6735 if (!alloc_rt_sched_group(tg, parent)) 6736 goto err; 6737 6738 return tg; 6739 6740err: 6741 free_sched_group(tg); 6742 return ERR_PTR(-ENOMEM); 6743} 6744 6745void sched_online_group(struct task_group *tg, struct task_group *parent) 6746{ 6747 unsigned long flags; 6748 6749 spin_lock_irqsave(&task_group_lock, flags); 6750 list_add_rcu(&tg->list, &task_groups); 6751 6752 WARN_ON(!parent); /* root should already exist */ 6753 6754 tg->parent = parent; 6755 INIT_LIST_HEAD(&tg->children); 6756 list_add_rcu(&tg->siblings, &parent->children); 6757 spin_unlock_irqrestore(&task_group_lock, flags); 6758} 6759 6760/* rcu callback to free various structures associated with a task group */ 6761static void free_sched_group_rcu(struct rcu_head *rhp) 6762{ 6763 /* now it should be safe to free those cfs_rqs */ 6764 free_sched_group(container_of(rhp, struct task_group, rcu)); 6765} 6766 6767/* Destroy runqueue etc associated with a task group */ 6768void sched_destroy_group(struct task_group *tg) 6769{ 6770 /* wait for possible concurrent references to cfs_rqs complete */ 6771 call_rcu(&tg->rcu, free_sched_group_rcu); 6772} 6773 6774void sched_offline_group(struct task_group *tg) 6775{ 6776 unsigned long flags; 6777 int i; 6778 6779 /* end participation in shares distribution */ 6780 for_each_possible_cpu(i) 6781 unregister_fair_sched_group(tg, i); 6782 6783 spin_lock_irqsave(&task_group_lock, flags); 6784 list_del_rcu(&tg->list); 6785 list_del_rcu(&tg->siblings); 6786 spin_unlock_irqrestore(&task_group_lock, flags); 6787} 6788 6789/* change task's runqueue when it moves between groups. 6790 * The caller of this function should have put the task in its new group 6791 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to 6792 * reflect its new group. 6793 */ 6794void sched_move_task(struct task_struct *tsk) 6795{ 6796 struct task_group *tg; 6797 int on_rq, running; 6798 unsigned long flags; 6799 struct rq *rq; 6800 6801 rq = task_rq_lock(tsk, &flags); 6802 6803 running = task_current(rq, tsk); 6804 on_rq = tsk->on_rq; 6805 6806 if (on_rq) 6807 dequeue_task(rq, tsk, 0); 6808 if (unlikely(running)) 6809 tsk->sched_class->put_prev_task(rq, tsk); 6810 6811 tg = container_of(task_css_check(tsk, cpu_cgroup_subsys_id, 6812 lockdep_is_held(&tsk->sighand->siglock)), 6813 struct task_group, css); 6814 tg = autogroup_task_group(tsk, tg); 6815 tsk->sched_task_group = tg; 6816 6817#ifdef CONFIG_FAIR_GROUP_SCHED 6818 if (tsk->sched_class->task_move_group) 6819 tsk->sched_class->task_move_group(tsk, on_rq); 6820 else 6821#endif 6822 set_task_rq(tsk, task_cpu(tsk)); 6823 6824 if (unlikely(running)) 6825 tsk->sched_class->set_curr_task(rq); 6826 if (on_rq) 6827 enqueue_task(rq, tsk, 0); 6828 6829 task_rq_unlock(rq, tsk, &flags); 6830} 6831#endif /* CONFIG_CGROUP_SCHED */ 6832 6833#if defined(CONFIG_RT_GROUP_SCHED) || defined(CONFIG_CFS_BANDWIDTH) 6834static unsigned long to_ratio(u64 period, u64 runtime) 6835{ 6836 if (runtime == RUNTIME_INF) 6837 return 1ULL << 20; 6838 6839 return div64_u64(runtime << 20, period); 6840} 6841#endif 6842 6843#ifdef CONFIG_RT_GROUP_SCHED 6844/* 6845 * Ensure that the real time constraints are schedulable. 6846 */ 6847static DEFINE_MUTEX(rt_constraints_mutex); 6848 6849/* Must be called with tasklist_lock held */ 6850static inline int tg_has_rt_tasks(struct task_group *tg) 6851{ 6852 struct task_struct *g, *p; 6853 6854 do_each_thread(g, p) { 6855 if (rt_task(p) && task_rq(p)->rt.tg == tg) 6856 return 1; 6857 } while_each_thread(g, p); 6858 6859 return 0; 6860} 6861 6862struct rt_schedulable_data { 6863 struct task_group *tg; 6864 u64 rt_period; 6865 u64 rt_runtime; 6866}; 6867 6868static int tg_rt_schedulable(struct task_group *tg, void *data) 6869{ 6870 struct rt_schedulable_data *d = data; 6871 struct task_group *child; 6872 unsigned long total, sum = 0; 6873 u64 period, runtime; 6874 6875 period = ktime_to_ns(tg->rt_bandwidth.rt_period); 6876 runtime = tg->rt_bandwidth.rt_runtime; 6877 6878 if (tg == d->tg) { 6879 period = d->rt_period; 6880 runtime = d->rt_runtime; 6881 } 6882 6883 /* 6884 * Cannot have more runtime than the period. 6885 */ 6886 if (runtime > period && runtime != RUNTIME_INF) 6887 return -EINVAL; 6888 6889 /* 6890 * Ensure we don't starve existing RT tasks. 6891 */ 6892 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg)) 6893 return -EBUSY; 6894 6895 total = to_ratio(period, runtime); 6896 6897 /* 6898 * Nobody can have more than the global setting allows. 6899 */ 6900 if (total > to_ratio(global_rt_period(), global_rt_runtime())) 6901 return -EINVAL; 6902 6903 /* 6904 * The sum of our children's runtime should not exceed our own. 6905 */ 6906 list_for_each_entry_rcu(child, &tg->children, siblings) { 6907 period = ktime_to_ns(child->rt_bandwidth.rt_period); 6908 runtime = child->rt_bandwidth.rt_runtime; 6909 6910 if (child == d->tg) { 6911 period = d->rt_period; 6912 runtime = d->rt_runtime; 6913 } 6914 6915 sum += to_ratio(period, runtime); 6916 } 6917 6918 if (sum > total) 6919 return -EINVAL; 6920 6921 return 0; 6922} 6923 6924static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime) 6925{ 6926 int ret; 6927 6928 struct rt_schedulable_data data = { 6929 .tg = tg, 6930 .rt_period = period, 6931 .rt_runtime = runtime, 6932 }; 6933 6934 rcu_read_lock(); 6935 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data); 6936 rcu_read_unlock(); 6937 6938 return ret; 6939} 6940 6941static int tg_set_rt_bandwidth(struct task_group *tg, 6942 u64 rt_period, u64 rt_runtime) 6943{ 6944 int i, err = 0; 6945 6946 mutex_lock(&rt_constraints_mutex); 6947 read_lock(&tasklist_lock); 6948 err = __rt_schedulable(tg, rt_period, rt_runtime); 6949 if (err) 6950 goto unlock; 6951 6952 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock); 6953 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period); 6954 tg->rt_bandwidth.rt_runtime = rt_runtime; 6955 6956 for_each_possible_cpu(i) { 6957 struct rt_rq *rt_rq = tg->rt_rq[i]; 6958 6959 raw_spin_lock(&rt_rq->rt_runtime_lock); 6960 rt_rq->rt_runtime = rt_runtime; 6961 raw_spin_unlock(&rt_rq->rt_runtime_lock); 6962 } 6963 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock); 6964unlock: 6965 read_unlock(&tasklist_lock); 6966 mutex_unlock(&rt_constraints_mutex); 6967 6968 return err; 6969} 6970 6971static int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us) 6972{ 6973 u64 rt_runtime, rt_period; 6974 6975 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period); 6976 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC; 6977 if (rt_runtime_us < 0) 6978 rt_runtime = RUNTIME_INF; 6979 6980 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime); 6981} 6982 6983static long sched_group_rt_runtime(struct task_group *tg) 6984{ 6985 u64 rt_runtime_us; 6986 6987 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF) 6988 return -1; 6989 6990 rt_runtime_us = tg->rt_bandwidth.rt_runtime; 6991 do_div(rt_runtime_us, NSEC_PER_USEC); 6992 return rt_runtime_us; 6993} 6994 6995static int sched_group_set_rt_period(struct task_group *tg, long rt_period_us) 6996{ 6997 u64 rt_runtime, rt_period; 6998 6999 rt_period = (u64)rt_period_us * NSEC_PER_USEC; 7000 rt_runtime = tg->rt_bandwidth.rt_runtime; 7001 7002 if (rt_period == 0) 7003 return -EINVAL; 7004 7005 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime); 7006} 7007 7008static long sched_group_rt_period(struct task_group *tg) 7009{ 7010 u64 rt_period_us; 7011 7012 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period); 7013 do_div(rt_period_us, NSEC_PER_USEC); 7014 return rt_period_us; 7015} 7016 7017static int sched_rt_global_constraints(void) 7018{ 7019 u64 runtime, period; 7020 int ret = 0; 7021 7022 if (sysctl_sched_rt_period <= 0) 7023 return -EINVAL; 7024 7025 runtime = global_rt_runtime(); 7026 period = global_rt_period(); 7027 7028 /* 7029 * Sanity check on the sysctl variables. 7030 */ 7031 if (runtime > period && runtime != RUNTIME_INF) 7032 return -EINVAL; 7033 7034 mutex_lock(&rt_constraints_mutex); 7035 read_lock(&tasklist_lock); 7036 ret = __rt_schedulable(NULL, 0, 0); 7037 read_unlock(&tasklist_lock); 7038 mutex_unlock(&rt_constraints_mutex); 7039 7040 return ret; 7041} 7042 7043static int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk) 7044{ 7045 /* Don't accept realtime tasks when there is no way for them to run */ 7046 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0) 7047 return 0; 7048 7049 return 1; 7050} 7051 7052#else /* !CONFIG_RT_GROUP_SCHED */ 7053static int sched_rt_global_constraints(void) 7054{ 7055 unsigned long flags; 7056 int i; 7057 7058 if (sysctl_sched_rt_period <= 0) 7059 return -EINVAL; 7060 7061 /* 7062 * There's always some RT tasks in the root group 7063 * -- migration, kstopmachine etc.. 7064 */ 7065 if (sysctl_sched_rt_runtime == 0) 7066 return -EBUSY; 7067 7068 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags); 7069 for_each_possible_cpu(i) { 7070 struct rt_rq *rt_rq = &cpu_rq(i)->rt; 7071 7072 raw_spin_lock(&rt_rq->rt_runtime_lock); 7073 rt_rq->rt_runtime = global_rt_runtime(); 7074 raw_spin_unlock(&rt_rq->rt_runtime_lock); 7075 } 7076 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags); 7077 7078 return 0; 7079} 7080#endif /* CONFIG_RT_GROUP_SCHED */ 7081 7082int sched_rr_handler(struct ctl_table *table, int write, 7083 void __user *buffer, size_t *lenp, 7084 loff_t *ppos) 7085{ 7086 int ret; 7087 static DEFINE_MUTEX(mutex); 7088 7089 mutex_lock(&mutex); 7090 ret = proc_dointvec(table, write, buffer, lenp, ppos); 7091 /* make sure that internally we keep jiffies */ 7092 /* also, writing zero resets timeslice to default */ 7093 if (!ret && write) { 7094 sched_rr_timeslice = sched_rr_timeslice <= 0 ? 7095 RR_TIMESLICE : msecs_to_jiffies(sched_rr_timeslice); 7096 } 7097 mutex_unlock(&mutex); 7098 return ret; 7099} 7100 7101int sched_rt_handler(struct ctl_table *table, int write, 7102 void __user *buffer, size_t *lenp, 7103 loff_t *ppos) 7104{ 7105 int ret; 7106 int old_period, old_runtime; 7107 static DEFINE_MUTEX(mutex); 7108 7109 mutex_lock(&mutex); 7110 old_period = sysctl_sched_rt_period; 7111 old_runtime = sysctl_sched_rt_runtime; 7112 7113 ret = proc_dointvec(table, write, buffer, lenp, ppos); 7114 7115 if (!ret && write) { 7116 ret = sched_rt_global_constraints(); 7117 if (ret) { 7118 sysctl_sched_rt_period = old_period; 7119 sysctl_sched_rt_runtime = old_runtime; 7120 } else { 7121 def_rt_bandwidth.rt_runtime = global_rt_runtime(); 7122 def_rt_bandwidth.rt_period = 7123 ns_to_ktime(global_rt_period()); 7124 } 7125 } 7126 mutex_unlock(&mutex); 7127 7128 return ret; 7129} 7130 7131#ifdef CONFIG_CGROUP_SCHED 7132 7133static inline struct task_group *css_tg(struct cgroup_subsys_state *css) 7134{ 7135 return css ? container_of(css, struct task_group, css) : NULL; 7136} 7137 7138static struct cgroup_subsys_state * 7139cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css) 7140{ 7141 struct task_group *parent = css_tg(parent_css); 7142 struct task_group *tg; 7143 7144 if (!parent) { 7145 /* This is early initialization for the top cgroup */ 7146 return &root_task_group.css; 7147 } 7148 7149 tg = sched_create_group(parent); 7150 if (IS_ERR(tg)) 7151 return ERR_PTR(-ENOMEM); 7152 7153 return &tg->css; 7154} 7155 7156static int cpu_cgroup_css_online(struct cgroup_subsys_state *css) 7157{ 7158 struct task_group *tg = css_tg(css); 7159 struct task_group *parent = css_tg(css_parent(css)); 7160 7161 if (parent) 7162 sched_online_group(tg, parent); 7163 return 0; 7164} 7165 7166static void cpu_cgroup_css_free(struct cgroup_subsys_state *css) 7167{ 7168 struct task_group *tg = css_tg(css); 7169 7170 sched_destroy_group(tg); 7171} 7172 7173static void cpu_cgroup_css_offline(struct cgroup_subsys_state *css) 7174{ 7175 struct task_group *tg = css_tg(css); 7176 7177 sched_offline_group(tg); 7178} 7179 7180static int cpu_cgroup_can_attach(struct cgroup_subsys_state *css, 7181 struct cgroup_taskset *tset) 7182{ 7183 struct task_struct *task; 7184 7185 cgroup_taskset_for_each(task, css, tset) { 7186#ifdef CONFIG_RT_GROUP_SCHED 7187 if (!sched_rt_can_attach(css_tg(css), task)) 7188 return -EINVAL; 7189#else 7190 /* We don't support RT-tasks being in separate groups */ 7191 if (task->sched_class != &fair_sched_class) 7192 return -EINVAL; 7193#endif 7194 } 7195 return 0; 7196} 7197 7198static void cpu_cgroup_attach(struct cgroup_subsys_state *css, 7199 struct cgroup_taskset *tset) 7200{ 7201 struct task_struct *task; 7202 7203 cgroup_taskset_for_each(task, css, tset) 7204 sched_move_task(task); 7205} 7206 7207static void cpu_cgroup_exit(struct cgroup_subsys_state *css, 7208 struct cgroup_subsys_state *old_css, 7209 struct task_struct *task) 7210{ 7211 /* 7212 * cgroup_exit() is called in the copy_process() failure path. 7213 * Ignore this case since the task hasn't ran yet, this avoids 7214 * trying to poke a half freed task state from generic code. 7215 */ 7216 if (!(task->flags & PF_EXITING)) 7217 return; 7218 7219 sched_move_task(task); 7220} 7221 7222#ifdef CONFIG_FAIR_GROUP_SCHED 7223static int cpu_shares_write_u64(struct cgroup_subsys_state *css, 7224 struct cftype *cftype, u64 shareval) 7225{ 7226 return sched_group_set_shares(css_tg(css), scale_load(shareval)); 7227} 7228 7229static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css, 7230 struct cftype *cft) 7231{ 7232 struct task_group *tg = css_tg(css); 7233 7234 return (u64) scale_load_down(tg->shares); 7235} 7236 7237#ifdef CONFIG_CFS_BANDWIDTH 7238static DEFINE_MUTEX(cfs_constraints_mutex); 7239 7240const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */ 7241const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */ 7242 7243static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime); 7244 7245static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota) 7246{ 7247 int i, ret = 0, runtime_enabled, runtime_was_enabled; 7248 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth; 7249 7250 if (tg == &root_task_group) 7251 return -EINVAL; 7252 7253 /* 7254 * Ensure we have at some amount of bandwidth every period. This is 7255 * to prevent reaching a state of large arrears when throttled via 7256 * entity_tick() resulting in prolonged exit starvation. 7257 */ 7258 if (quota < min_cfs_quota_period || period < min_cfs_quota_period) 7259 return -EINVAL; 7260 7261 /* 7262 * Likewise, bound things on the otherside by preventing insane quota 7263 * periods. This also allows us to normalize in computing quota 7264 * feasibility. 7265 */ 7266 if (period > max_cfs_quota_period) 7267 return -EINVAL; 7268 7269 mutex_lock(&cfs_constraints_mutex); 7270 ret = __cfs_schedulable(tg, period, quota); 7271 if (ret) 7272 goto out_unlock; 7273 7274 runtime_enabled = quota != RUNTIME_INF; 7275 runtime_was_enabled = cfs_b->quota != RUNTIME_INF; 7276 account_cfs_bandwidth_used(runtime_enabled, runtime_was_enabled); 7277 raw_spin_lock_irq(&cfs_b->lock); 7278 cfs_b->period = ns_to_ktime(period); 7279 cfs_b->quota = quota; 7280 7281 __refill_cfs_bandwidth_runtime(cfs_b); 7282 /* restart the period timer (if active) to handle new period expiry */ 7283 if (runtime_enabled && cfs_b->timer_active) { 7284 /* force a reprogram */ 7285 cfs_b->timer_active = 0; 7286 __start_cfs_bandwidth(cfs_b); 7287 } 7288 raw_spin_unlock_irq(&cfs_b->lock); 7289 7290 for_each_possible_cpu(i) { 7291 struct cfs_rq *cfs_rq = tg->cfs_rq[i]; 7292 struct rq *rq = cfs_rq->rq; 7293 7294 raw_spin_lock_irq(&rq->lock); 7295 cfs_rq->runtime_enabled = runtime_enabled; 7296 cfs_rq->runtime_remaining = 0; 7297 7298 if (cfs_rq->throttled) 7299 unthrottle_cfs_rq(cfs_rq); 7300 raw_spin_unlock_irq(&rq->lock); 7301 } 7302out_unlock: 7303 mutex_unlock(&cfs_constraints_mutex); 7304 7305 return ret; 7306} 7307 7308int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us) 7309{ 7310 u64 quota, period; 7311 7312 period = ktime_to_ns(tg->cfs_bandwidth.period); 7313 if (cfs_quota_us < 0) 7314 quota = RUNTIME_INF; 7315 else 7316 quota = (u64)cfs_quota_us * NSEC_PER_USEC; 7317 7318 return tg_set_cfs_bandwidth(tg, period, quota); 7319} 7320 7321long tg_get_cfs_quota(struct task_group *tg) 7322{ 7323 u64 quota_us; 7324 7325 if (tg->cfs_bandwidth.quota == RUNTIME_INF) 7326 return -1; 7327 7328 quota_us = tg->cfs_bandwidth.quota; 7329 do_div(quota_us, NSEC_PER_USEC); 7330 7331 return quota_us; 7332} 7333 7334int tg_set_cfs_period(struct task_group *tg, long cfs_period_us) 7335{ 7336 u64 quota, period; 7337 7338 period = (u64)cfs_period_us * NSEC_PER_USEC; 7339 quota = tg->cfs_bandwidth.quota; 7340 7341 return tg_set_cfs_bandwidth(tg, period, quota); 7342} 7343 7344long tg_get_cfs_period(struct task_group *tg) 7345{ 7346 u64 cfs_period_us; 7347 7348 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period); 7349 do_div(cfs_period_us, NSEC_PER_USEC); 7350 7351 return cfs_period_us; 7352} 7353 7354static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css, 7355 struct cftype *cft) 7356{ 7357 return tg_get_cfs_quota(css_tg(css)); 7358} 7359 7360static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css, 7361 struct cftype *cftype, s64 cfs_quota_us) 7362{ 7363 return tg_set_cfs_quota(css_tg(css), cfs_quota_us); 7364} 7365 7366static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css, 7367 struct cftype *cft) 7368{ 7369 return tg_get_cfs_period(css_tg(css)); 7370} 7371 7372static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css, 7373 struct cftype *cftype, u64 cfs_period_us) 7374{ 7375 return tg_set_cfs_period(css_tg(css), cfs_period_us); 7376} 7377 7378struct cfs_schedulable_data { 7379 struct task_group *tg; 7380 u64 period, quota; 7381}; 7382 7383/* 7384 * normalize group quota/period to be quota/max_period 7385 * note: units are usecs 7386 */ 7387static u64 normalize_cfs_quota(struct task_group *tg, 7388 struct cfs_schedulable_data *d) 7389{ 7390 u64 quota, period; 7391 7392 if (tg == d->tg) { 7393 period = d->period; 7394 quota = d->quota; 7395 } else { 7396 period = tg_get_cfs_period(tg); 7397 quota = tg_get_cfs_quota(tg); 7398 } 7399 7400 /* note: these should typically be equivalent */ 7401 if (quota == RUNTIME_INF || quota == -1) 7402 return RUNTIME_INF; 7403 7404 return to_ratio(period, quota); 7405} 7406 7407static int tg_cfs_schedulable_down(struct task_group *tg, void *data) 7408{ 7409 struct cfs_schedulable_data *d = data; 7410 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth; 7411 s64 quota = 0, parent_quota = -1; 7412 7413 if (!tg->parent) { 7414 quota = RUNTIME_INF; 7415 } else { 7416 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth; 7417 7418 quota = normalize_cfs_quota(tg, d); 7419 parent_quota = parent_b->hierarchal_quota; 7420 7421 /* 7422 * ensure max(child_quota) <= parent_quota, inherit when no 7423 * limit is set 7424 */ 7425 if (quota == RUNTIME_INF) 7426 quota = parent_quota; 7427 else if (parent_quota != RUNTIME_INF && quota > parent_quota) 7428 return -EINVAL; 7429 } 7430 cfs_b->hierarchal_quota = quota; 7431 7432 return 0; 7433} 7434 7435static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota) 7436{ 7437 int ret; 7438 struct cfs_schedulable_data data = { 7439 .tg = tg, 7440 .period = period, 7441 .quota = quota, 7442 }; 7443 7444 if (quota != RUNTIME_INF) { 7445 do_div(data.period, NSEC_PER_USEC); 7446 do_div(data.quota, NSEC_PER_USEC); 7447 } 7448 7449 rcu_read_lock(); 7450 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data); 7451 rcu_read_unlock(); 7452 7453 return ret; 7454} 7455 7456static int cpu_stats_show(struct cgroup_subsys_state *css, struct cftype *cft, 7457 struct cgroup_map_cb *cb) 7458{ 7459 struct task_group *tg = css_tg(css); 7460 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth; 7461 7462 cb->fill(cb, "nr_periods", cfs_b->nr_periods); 7463 cb->fill(cb, "nr_throttled", cfs_b->nr_throttled); 7464 cb->fill(cb, "throttled_time", cfs_b->throttled_time); 7465 7466 return 0; 7467} 7468#endif /* CONFIG_CFS_BANDWIDTH */ 7469#endif /* CONFIG_FAIR_GROUP_SCHED */ 7470 7471#ifdef CONFIG_RT_GROUP_SCHED 7472static int cpu_rt_runtime_write(struct cgroup_subsys_state *css, 7473 struct cftype *cft, s64 val) 7474{ 7475 return sched_group_set_rt_runtime(css_tg(css), val); 7476} 7477 7478static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css, 7479 struct cftype *cft) 7480{ 7481 return sched_group_rt_runtime(css_tg(css)); 7482} 7483 7484static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css, 7485 struct cftype *cftype, u64 rt_period_us) 7486{ 7487 return sched_group_set_rt_period(css_tg(css), rt_period_us); 7488} 7489 7490static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css, 7491 struct cftype *cft) 7492{ 7493 return sched_group_rt_period(css_tg(css)); 7494} 7495#endif /* CONFIG_RT_GROUP_SCHED */ 7496 7497static struct cftype cpu_files[] = { 7498#ifdef CONFIG_FAIR_GROUP_SCHED 7499 { 7500 .name = "shares", 7501 .read_u64 = cpu_shares_read_u64, 7502 .write_u64 = cpu_shares_write_u64, 7503 }, 7504#endif 7505#ifdef CONFIG_CFS_BANDWIDTH 7506 { 7507 .name = "cfs_quota_us", 7508 .read_s64 = cpu_cfs_quota_read_s64, 7509 .write_s64 = cpu_cfs_quota_write_s64, 7510 }, 7511 { 7512 .name = "cfs_period_us", 7513 .read_u64 = cpu_cfs_period_read_u64, 7514 .write_u64 = cpu_cfs_period_write_u64, 7515 }, 7516 { 7517 .name = "stat", 7518 .read_map = cpu_stats_show, 7519 }, 7520#endif 7521#ifdef CONFIG_RT_GROUP_SCHED 7522 { 7523 .name = "rt_runtime_us", 7524 .read_s64 = cpu_rt_runtime_read, 7525 .write_s64 = cpu_rt_runtime_write, 7526 }, 7527 { 7528 .name = "rt_period_us", 7529 .read_u64 = cpu_rt_period_read_uint, 7530 .write_u64 = cpu_rt_period_write_uint, 7531 }, 7532#endif 7533 { } /* terminate */ 7534}; 7535 7536struct cgroup_subsys cpu_cgroup_subsys = { 7537 .name = "cpu", 7538 .css_alloc = cpu_cgroup_css_alloc, 7539 .css_free = cpu_cgroup_css_free, 7540 .css_online = cpu_cgroup_css_online, 7541 .css_offline = cpu_cgroup_css_offline, 7542 .can_attach = cpu_cgroup_can_attach, 7543 .attach = cpu_cgroup_attach, 7544 .exit = cpu_cgroup_exit, 7545 .subsys_id = cpu_cgroup_subsys_id, 7546 .base_cftypes = cpu_files, 7547 .early_init = 1, 7548}; 7549 7550#endif /* CONFIG_CGROUP_SCHED */ 7551 7552void dump_cpu_task(int cpu) 7553{ 7554 pr_info("Task dump for CPU %d:\n", cpu); 7555 sched_show_task(cpu_curr(cpu)); 7556} 7557