core.c revision f7b8a47da17c9ee4998f2ca2018fcc424e953c0e
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#include <linux/compiler.h> 77 78#include <asm/switch_to.h> 79#include <asm/tlb.h> 80#include <asm/irq_regs.h> 81#include <asm/mutex.h> 82#ifdef CONFIG_PARAVIRT 83#include <asm/paravirt.h> 84#endif 85 86#include "sched.h" 87#include "../workqueue_internal.h" 88#include "../smpboot.h" 89 90#define CREATE_TRACE_POINTS 91#include <trace/events/sched.h> 92 93void start_bandwidth_timer(struct hrtimer *period_timer, ktime_t period) 94{ 95 unsigned long delta; 96 ktime_t soft, hard, now; 97 98 for (;;) { 99 if (hrtimer_active(period_timer)) 100 break; 101 102 now = hrtimer_cb_get_time(period_timer); 103 hrtimer_forward(period_timer, now, period); 104 105 soft = hrtimer_get_softexpires(period_timer); 106 hard = hrtimer_get_expires(period_timer); 107 delta = ktime_to_ns(ktime_sub(hard, soft)); 108 __hrtimer_start_range_ns(period_timer, soft, delta, 109 HRTIMER_MODE_ABS_PINNED, 0); 110 } 111} 112 113DEFINE_MUTEX(sched_domains_mutex); 114DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues); 115 116static void update_rq_clock_task(struct rq *rq, s64 delta); 117 118void update_rq_clock(struct rq *rq) 119{ 120 s64 delta; 121 122 if (rq->skip_clock_update > 0) 123 return; 124 125 delta = sched_clock_cpu(cpu_of(rq)) - rq->clock; 126 if (delta < 0) 127 return; 128 rq->clock += delta; 129 update_rq_clock_task(rq, delta); 130} 131 132/* 133 * Debugging: various feature bits 134 */ 135 136#define SCHED_FEAT(name, enabled) \ 137 (1UL << __SCHED_FEAT_##name) * enabled | 138 139const_debug unsigned int sysctl_sched_features = 140#include "features.h" 141 0; 142 143#undef SCHED_FEAT 144 145#ifdef CONFIG_SCHED_DEBUG 146#define SCHED_FEAT(name, enabled) \ 147 #name , 148 149static const char * const sched_feat_names[] = { 150#include "features.h" 151}; 152 153#undef SCHED_FEAT 154 155static int sched_feat_show(struct seq_file *m, void *v) 156{ 157 int i; 158 159 for (i = 0; i < __SCHED_FEAT_NR; i++) { 160 if (!(sysctl_sched_features & (1UL << i))) 161 seq_puts(m, "NO_"); 162 seq_printf(m, "%s ", sched_feat_names[i]); 163 } 164 seq_puts(m, "\n"); 165 166 return 0; 167} 168 169#ifdef HAVE_JUMP_LABEL 170 171#define jump_label_key__true STATIC_KEY_INIT_TRUE 172#define jump_label_key__false STATIC_KEY_INIT_FALSE 173 174#define SCHED_FEAT(name, enabled) \ 175 jump_label_key__##enabled , 176 177struct static_key sched_feat_keys[__SCHED_FEAT_NR] = { 178#include "features.h" 179}; 180 181#undef SCHED_FEAT 182 183static void sched_feat_disable(int i) 184{ 185 if (static_key_enabled(&sched_feat_keys[i])) 186 static_key_slow_dec(&sched_feat_keys[i]); 187} 188 189static void sched_feat_enable(int i) 190{ 191 if (!static_key_enabled(&sched_feat_keys[i])) 192 static_key_slow_inc(&sched_feat_keys[i]); 193} 194#else 195static void sched_feat_disable(int i) { }; 196static void sched_feat_enable(int i) { }; 197#endif /* HAVE_JUMP_LABEL */ 198 199static int sched_feat_set(char *cmp) 200{ 201 int i; 202 int neg = 0; 203 204 if (strncmp(cmp, "NO_", 3) == 0) { 205 neg = 1; 206 cmp += 3; 207 } 208 209 for (i = 0; i < __SCHED_FEAT_NR; i++) { 210 if (strcmp(cmp, sched_feat_names[i]) == 0) { 211 if (neg) { 212 sysctl_sched_features &= ~(1UL << i); 213 sched_feat_disable(i); 214 } else { 215 sysctl_sched_features |= (1UL << i); 216 sched_feat_enable(i); 217 } 218 break; 219 } 220 } 221 222 return i; 223} 224 225static ssize_t 226sched_feat_write(struct file *filp, const char __user *ubuf, 227 size_t cnt, loff_t *ppos) 228{ 229 char buf[64]; 230 char *cmp; 231 int i; 232 struct inode *inode; 233 234 if (cnt > 63) 235 cnt = 63; 236 237 if (copy_from_user(&buf, ubuf, cnt)) 238 return -EFAULT; 239 240 buf[cnt] = 0; 241 cmp = strstrip(buf); 242 243 /* Ensure the static_key remains in a consistent state */ 244 inode = file_inode(filp); 245 mutex_lock(&inode->i_mutex); 246 i = sched_feat_set(cmp); 247 mutex_unlock(&inode->i_mutex); 248 if (i == __SCHED_FEAT_NR) 249 return -EINVAL; 250 251 *ppos += cnt; 252 253 return cnt; 254} 255 256static int sched_feat_open(struct inode *inode, struct file *filp) 257{ 258 return single_open(filp, sched_feat_show, NULL); 259} 260 261static const struct file_operations sched_feat_fops = { 262 .open = sched_feat_open, 263 .write = sched_feat_write, 264 .read = seq_read, 265 .llseek = seq_lseek, 266 .release = single_release, 267}; 268 269static __init int sched_init_debug(void) 270{ 271 debugfs_create_file("sched_features", 0644, NULL, NULL, 272 &sched_feat_fops); 273 274 return 0; 275} 276late_initcall(sched_init_debug); 277#endif /* CONFIG_SCHED_DEBUG */ 278 279/* 280 * Number of tasks to iterate in a single balance run. 281 * Limited because this is done with IRQs disabled. 282 */ 283const_debug unsigned int sysctl_sched_nr_migrate = 32; 284 285/* 286 * period over which we average the RT time consumption, measured 287 * in ms. 288 * 289 * default: 1s 290 */ 291const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC; 292 293/* 294 * period over which we measure -rt task cpu usage in us. 295 * default: 1s 296 */ 297unsigned int sysctl_sched_rt_period = 1000000; 298 299__read_mostly int scheduler_running; 300 301/* 302 * part of the period that we allow rt tasks to run in us. 303 * default: 0.95s 304 */ 305int sysctl_sched_rt_runtime = 950000; 306 307/* 308 * __task_rq_lock - lock the rq @p resides on. 309 */ 310static inline struct rq *__task_rq_lock(struct task_struct *p) 311 __acquires(rq->lock) 312{ 313 struct rq *rq; 314 315 lockdep_assert_held(&p->pi_lock); 316 317 for (;;) { 318 rq = task_rq(p); 319 raw_spin_lock(&rq->lock); 320 if (likely(rq == task_rq(p) && !task_on_rq_migrating(p))) 321 return rq; 322 raw_spin_unlock(&rq->lock); 323 324 while (unlikely(task_on_rq_migrating(p))) 325 cpu_relax(); 326 } 327} 328 329/* 330 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on. 331 */ 332static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags) 333 __acquires(p->pi_lock) 334 __acquires(rq->lock) 335{ 336 struct rq *rq; 337 338 for (;;) { 339 raw_spin_lock_irqsave(&p->pi_lock, *flags); 340 rq = task_rq(p); 341 raw_spin_lock(&rq->lock); 342 if (likely(rq == task_rq(p) && !task_on_rq_migrating(p))) 343 return rq; 344 raw_spin_unlock(&rq->lock); 345 raw_spin_unlock_irqrestore(&p->pi_lock, *flags); 346 347 while (unlikely(task_on_rq_migrating(p))) 348 cpu_relax(); 349 } 350} 351 352static void __task_rq_unlock(struct rq *rq) 353 __releases(rq->lock) 354{ 355 raw_spin_unlock(&rq->lock); 356} 357 358static inline void 359task_rq_unlock(struct rq *rq, struct task_struct *p, unsigned long *flags) 360 __releases(rq->lock) 361 __releases(p->pi_lock) 362{ 363 raw_spin_unlock(&rq->lock); 364 raw_spin_unlock_irqrestore(&p->pi_lock, *flags); 365} 366 367/* 368 * this_rq_lock - lock this runqueue and disable interrupts. 369 */ 370static struct rq *this_rq_lock(void) 371 __acquires(rq->lock) 372{ 373 struct rq *rq; 374 375 local_irq_disable(); 376 rq = this_rq(); 377 raw_spin_lock(&rq->lock); 378 379 return rq; 380} 381 382#ifdef CONFIG_SCHED_HRTICK 383/* 384 * Use HR-timers to deliver accurate preemption points. 385 */ 386 387static void hrtick_clear(struct rq *rq) 388{ 389 if (hrtimer_active(&rq->hrtick_timer)) 390 hrtimer_cancel(&rq->hrtick_timer); 391} 392 393/* 394 * High-resolution timer tick. 395 * Runs from hardirq context with interrupts disabled. 396 */ 397static enum hrtimer_restart hrtick(struct hrtimer *timer) 398{ 399 struct rq *rq = container_of(timer, struct rq, hrtick_timer); 400 401 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id()); 402 403 raw_spin_lock(&rq->lock); 404 update_rq_clock(rq); 405 rq->curr->sched_class->task_tick(rq, rq->curr, 1); 406 raw_spin_unlock(&rq->lock); 407 408 return HRTIMER_NORESTART; 409} 410 411#ifdef CONFIG_SMP 412 413static int __hrtick_restart(struct rq *rq) 414{ 415 struct hrtimer *timer = &rq->hrtick_timer; 416 ktime_t time = hrtimer_get_softexpires(timer); 417 418 return __hrtimer_start_range_ns(timer, time, 0, HRTIMER_MODE_ABS_PINNED, 0); 419} 420 421/* 422 * called from hardirq (IPI) context 423 */ 424static void __hrtick_start(void *arg) 425{ 426 struct rq *rq = arg; 427 428 raw_spin_lock(&rq->lock); 429 __hrtick_restart(rq); 430 rq->hrtick_csd_pending = 0; 431 raw_spin_unlock(&rq->lock); 432} 433 434/* 435 * Called to set the hrtick timer state. 436 * 437 * called with rq->lock held and irqs disabled 438 */ 439void hrtick_start(struct rq *rq, u64 delay) 440{ 441 struct hrtimer *timer = &rq->hrtick_timer; 442 ktime_t time; 443 s64 delta; 444 445 /* 446 * Don't schedule slices shorter than 10000ns, that just 447 * doesn't make sense and can cause timer DoS. 448 */ 449 delta = max_t(s64, delay, 10000LL); 450 time = ktime_add_ns(timer->base->get_time(), delta); 451 452 hrtimer_set_expires(timer, time); 453 454 if (rq == this_rq()) { 455 __hrtick_restart(rq); 456 } else if (!rq->hrtick_csd_pending) { 457 smp_call_function_single_async(cpu_of(rq), &rq->hrtick_csd); 458 rq->hrtick_csd_pending = 1; 459 } 460} 461 462static int 463hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu) 464{ 465 int cpu = (int)(long)hcpu; 466 467 switch (action) { 468 case CPU_UP_CANCELED: 469 case CPU_UP_CANCELED_FROZEN: 470 case CPU_DOWN_PREPARE: 471 case CPU_DOWN_PREPARE_FROZEN: 472 case CPU_DEAD: 473 case CPU_DEAD_FROZEN: 474 hrtick_clear(cpu_rq(cpu)); 475 return NOTIFY_OK; 476 } 477 478 return NOTIFY_DONE; 479} 480 481static __init void init_hrtick(void) 482{ 483 hotcpu_notifier(hotplug_hrtick, 0); 484} 485#else 486/* 487 * Called to set the hrtick timer state. 488 * 489 * called with rq->lock held and irqs disabled 490 */ 491void hrtick_start(struct rq *rq, u64 delay) 492{ 493 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0, 494 HRTIMER_MODE_REL_PINNED, 0); 495} 496 497static inline void init_hrtick(void) 498{ 499} 500#endif /* CONFIG_SMP */ 501 502static void init_rq_hrtick(struct rq *rq) 503{ 504#ifdef CONFIG_SMP 505 rq->hrtick_csd_pending = 0; 506 507 rq->hrtick_csd.flags = 0; 508 rq->hrtick_csd.func = __hrtick_start; 509 rq->hrtick_csd.info = rq; 510#endif 511 512 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); 513 rq->hrtick_timer.function = hrtick; 514} 515#else /* CONFIG_SCHED_HRTICK */ 516static inline void hrtick_clear(struct rq *rq) 517{ 518} 519 520static inline void init_rq_hrtick(struct rq *rq) 521{ 522} 523 524static inline void init_hrtick(void) 525{ 526} 527#endif /* CONFIG_SCHED_HRTICK */ 528 529/* 530 * cmpxchg based fetch_or, macro so it works for different integer types 531 */ 532#define fetch_or(ptr, val) \ 533({ typeof(*(ptr)) __old, __val = *(ptr); \ 534 for (;;) { \ 535 __old = cmpxchg((ptr), __val, __val | (val)); \ 536 if (__old == __val) \ 537 break; \ 538 __val = __old; \ 539 } \ 540 __old; \ 541}) 542 543#if defined(CONFIG_SMP) && defined(TIF_POLLING_NRFLAG) 544/* 545 * Atomically set TIF_NEED_RESCHED and test for TIF_POLLING_NRFLAG, 546 * this avoids any races wrt polling state changes and thereby avoids 547 * spurious IPIs. 548 */ 549static bool set_nr_and_not_polling(struct task_struct *p) 550{ 551 struct thread_info *ti = task_thread_info(p); 552 return !(fetch_or(&ti->flags, _TIF_NEED_RESCHED) & _TIF_POLLING_NRFLAG); 553} 554 555/* 556 * Atomically set TIF_NEED_RESCHED if TIF_POLLING_NRFLAG is set. 557 * 558 * If this returns true, then the idle task promises to call 559 * sched_ttwu_pending() and reschedule soon. 560 */ 561static bool set_nr_if_polling(struct task_struct *p) 562{ 563 struct thread_info *ti = task_thread_info(p); 564 typeof(ti->flags) old, val = ACCESS_ONCE(ti->flags); 565 566 for (;;) { 567 if (!(val & _TIF_POLLING_NRFLAG)) 568 return false; 569 if (val & _TIF_NEED_RESCHED) 570 return true; 571 old = cmpxchg(&ti->flags, val, val | _TIF_NEED_RESCHED); 572 if (old == val) 573 break; 574 val = old; 575 } 576 return true; 577} 578 579#else 580static bool set_nr_and_not_polling(struct task_struct *p) 581{ 582 set_tsk_need_resched(p); 583 return true; 584} 585 586#ifdef CONFIG_SMP 587static bool set_nr_if_polling(struct task_struct *p) 588{ 589 return false; 590} 591#endif 592#endif 593 594/* 595 * resched_curr - mark rq's current task 'to be rescheduled now'. 596 * 597 * On UP this means the setting of the need_resched flag, on SMP it 598 * might also involve a cross-CPU call to trigger the scheduler on 599 * the target CPU. 600 */ 601void resched_curr(struct rq *rq) 602{ 603 struct task_struct *curr = rq->curr; 604 int cpu; 605 606 lockdep_assert_held(&rq->lock); 607 608 if (test_tsk_need_resched(curr)) 609 return; 610 611 cpu = cpu_of(rq); 612 613 if (cpu == smp_processor_id()) { 614 set_tsk_need_resched(curr); 615 set_preempt_need_resched(); 616 return; 617 } 618 619 if (set_nr_and_not_polling(curr)) 620 smp_send_reschedule(cpu); 621 else 622 trace_sched_wake_idle_without_ipi(cpu); 623} 624 625void resched_cpu(int cpu) 626{ 627 struct rq *rq = cpu_rq(cpu); 628 unsigned long flags; 629 630 if (!raw_spin_trylock_irqsave(&rq->lock, flags)) 631 return; 632 resched_curr(rq); 633 raw_spin_unlock_irqrestore(&rq->lock, flags); 634} 635 636#ifdef CONFIG_SMP 637#ifdef CONFIG_NO_HZ_COMMON 638/* 639 * In the semi idle case, use the nearest busy cpu for migrating timers 640 * from an idle cpu. This is good for power-savings. 641 * 642 * We don't do similar optimization for completely idle system, as 643 * selecting an idle cpu will add more delays to the timers than intended 644 * (as that cpu's timer base may not be uptodate wrt jiffies etc). 645 */ 646int get_nohz_timer_target(int pinned) 647{ 648 int cpu = smp_processor_id(); 649 int i; 650 struct sched_domain *sd; 651 652 if (pinned || !get_sysctl_timer_migration() || !idle_cpu(cpu)) 653 return cpu; 654 655 rcu_read_lock(); 656 for_each_domain(cpu, sd) { 657 for_each_cpu(i, sched_domain_span(sd)) { 658 if (!idle_cpu(i)) { 659 cpu = i; 660 goto unlock; 661 } 662 } 663 } 664unlock: 665 rcu_read_unlock(); 666 return cpu; 667} 668/* 669 * When add_timer_on() enqueues a timer into the timer wheel of an 670 * idle CPU then this timer might expire before the next timer event 671 * which is scheduled to wake up that CPU. In case of a completely 672 * idle system the next event might even be infinite time into the 673 * future. wake_up_idle_cpu() ensures that the CPU is woken up and 674 * leaves the inner idle loop so the newly added timer is taken into 675 * account when the CPU goes back to idle and evaluates the timer 676 * wheel for the next timer event. 677 */ 678static void wake_up_idle_cpu(int cpu) 679{ 680 struct rq *rq = cpu_rq(cpu); 681 682 if (cpu == smp_processor_id()) 683 return; 684 685 if (set_nr_and_not_polling(rq->idle)) 686 smp_send_reschedule(cpu); 687 else 688 trace_sched_wake_idle_without_ipi(cpu); 689} 690 691static bool wake_up_full_nohz_cpu(int cpu) 692{ 693 /* 694 * We just need the target to call irq_exit() and re-evaluate 695 * the next tick. The nohz full kick at least implies that. 696 * If needed we can still optimize that later with an 697 * empty IRQ. 698 */ 699 if (tick_nohz_full_cpu(cpu)) { 700 if (cpu != smp_processor_id() || 701 tick_nohz_tick_stopped()) 702 tick_nohz_full_kick_cpu(cpu); 703 return true; 704 } 705 706 return false; 707} 708 709void wake_up_nohz_cpu(int cpu) 710{ 711 if (!wake_up_full_nohz_cpu(cpu)) 712 wake_up_idle_cpu(cpu); 713} 714 715static inline bool got_nohz_idle_kick(void) 716{ 717 int cpu = smp_processor_id(); 718 719 if (!test_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu))) 720 return false; 721 722 if (idle_cpu(cpu) && !need_resched()) 723 return true; 724 725 /* 726 * We can't run Idle Load Balance on this CPU for this time so we 727 * cancel it and clear NOHZ_BALANCE_KICK 728 */ 729 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu)); 730 return false; 731} 732 733#else /* CONFIG_NO_HZ_COMMON */ 734 735static inline bool got_nohz_idle_kick(void) 736{ 737 return false; 738} 739 740#endif /* CONFIG_NO_HZ_COMMON */ 741 742#ifdef CONFIG_NO_HZ_FULL 743bool sched_can_stop_tick(void) 744{ 745 /* 746 * More than one running task need preemption. 747 * nr_running update is assumed to be visible 748 * after IPI is sent from wakers. 749 */ 750 if (this_rq()->nr_running > 1) 751 return false; 752 753 return true; 754} 755#endif /* CONFIG_NO_HZ_FULL */ 756 757void sched_avg_update(struct rq *rq) 758{ 759 s64 period = sched_avg_period(); 760 761 while ((s64)(rq_clock(rq) - rq->age_stamp) > period) { 762 /* 763 * Inline assembly required to prevent the compiler 764 * optimising this loop into a divmod call. 765 * See __iter_div_u64_rem() for another example of this. 766 */ 767 asm("" : "+rm" (rq->age_stamp)); 768 rq->age_stamp += period; 769 rq->rt_avg /= 2; 770 } 771} 772 773#endif /* CONFIG_SMP */ 774 775#if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \ 776 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH))) 777/* 778 * Iterate task_group tree rooted at *from, calling @down when first entering a 779 * node and @up when leaving it for the final time. 780 * 781 * Caller must hold rcu_lock or sufficient equivalent. 782 */ 783int walk_tg_tree_from(struct task_group *from, 784 tg_visitor down, tg_visitor up, void *data) 785{ 786 struct task_group *parent, *child; 787 int ret; 788 789 parent = from; 790 791down: 792 ret = (*down)(parent, data); 793 if (ret) 794 goto out; 795 list_for_each_entry_rcu(child, &parent->children, siblings) { 796 parent = child; 797 goto down; 798 799up: 800 continue; 801 } 802 ret = (*up)(parent, data); 803 if (ret || parent == from) 804 goto out; 805 806 child = parent; 807 parent = parent->parent; 808 if (parent) 809 goto up; 810out: 811 return ret; 812} 813 814int tg_nop(struct task_group *tg, void *data) 815{ 816 return 0; 817} 818#endif 819 820static void set_load_weight(struct task_struct *p) 821{ 822 int prio = p->static_prio - MAX_RT_PRIO; 823 struct load_weight *load = &p->se.load; 824 825 /* 826 * SCHED_IDLE tasks get minimal weight: 827 */ 828 if (p->policy == SCHED_IDLE) { 829 load->weight = scale_load(WEIGHT_IDLEPRIO); 830 load->inv_weight = WMULT_IDLEPRIO; 831 return; 832 } 833 834 load->weight = scale_load(prio_to_weight[prio]); 835 load->inv_weight = prio_to_wmult[prio]; 836} 837 838static void enqueue_task(struct rq *rq, struct task_struct *p, int flags) 839{ 840 update_rq_clock(rq); 841 sched_info_queued(rq, p); 842 p->sched_class->enqueue_task(rq, p, flags); 843} 844 845static void dequeue_task(struct rq *rq, struct task_struct *p, int flags) 846{ 847 update_rq_clock(rq); 848 sched_info_dequeued(rq, p); 849 p->sched_class->dequeue_task(rq, p, flags); 850} 851 852void activate_task(struct rq *rq, struct task_struct *p, int flags) 853{ 854 if (task_contributes_to_load(p)) 855 rq->nr_uninterruptible--; 856 857 enqueue_task(rq, p, flags); 858} 859 860void deactivate_task(struct rq *rq, struct task_struct *p, int flags) 861{ 862 if (task_contributes_to_load(p)) 863 rq->nr_uninterruptible++; 864 865 dequeue_task(rq, p, flags); 866} 867 868static void update_rq_clock_task(struct rq *rq, s64 delta) 869{ 870/* 871 * In theory, the compile should just see 0 here, and optimize out the call 872 * to sched_rt_avg_update. But I don't trust it... 873 */ 874#if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING) 875 s64 steal = 0, irq_delta = 0; 876#endif 877#ifdef CONFIG_IRQ_TIME_ACCOUNTING 878 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time; 879 880 /* 881 * Since irq_time is only updated on {soft,}irq_exit, we might run into 882 * this case when a previous update_rq_clock() happened inside a 883 * {soft,}irq region. 884 * 885 * When this happens, we stop ->clock_task and only update the 886 * prev_irq_time stamp to account for the part that fit, so that a next 887 * update will consume the rest. This ensures ->clock_task is 888 * monotonic. 889 * 890 * It does however cause some slight miss-attribution of {soft,}irq 891 * time, a more accurate solution would be to update the irq_time using 892 * the current rq->clock timestamp, except that would require using 893 * atomic ops. 894 */ 895 if (irq_delta > delta) 896 irq_delta = delta; 897 898 rq->prev_irq_time += irq_delta; 899 delta -= irq_delta; 900#endif 901#ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING 902 if (static_key_false((¶virt_steal_rq_enabled))) { 903 steal = paravirt_steal_clock(cpu_of(rq)); 904 steal -= rq->prev_steal_time_rq; 905 906 if (unlikely(steal > delta)) 907 steal = delta; 908 909 rq->prev_steal_time_rq += steal; 910 delta -= steal; 911 } 912#endif 913 914 rq->clock_task += delta; 915 916#if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING) 917 if ((irq_delta + steal) && sched_feat(NONTASK_CAPACITY)) 918 sched_rt_avg_update(rq, irq_delta + steal); 919#endif 920} 921 922void sched_set_stop_task(int cpu, struct task_struct *stop) 923{ 924 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 }; 925 struct task_struct *old_stop = cpu_rq(cpu)->stop; 926 927 if (stop) { 928 /* 929 * Make it appear like a SCHED_FIFO task, its something 930 * userspace knows about and won't get confused about. 931 * 932 * Also, it will make PI more or less work without too 933 * much confusion -- but then, stop work should not 934 * rely on PI working anyway. 935 */ 936 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m); 937 938 stop->sched_class = &stop_sched_class; 939 } 940 941 cpu_rq(cpu)->stop = stop; 942 943 if (old_stop) { 944 /* 945 * Reset it back to a normal scheduling class so that 946 * it can die in pieces. 947 */ 948 old_stop->sched_class = &rt_sched_class; 949 } 950} 951 952/* 953 * __normal_prio - return the priority that is based on the static prio 954 */ 955static inline int __normal_prio(struct task_struct *p) 956{ 957 return p->static_prio; 958} 959 960/* 961 * Calculate the expected normal priority: i.e. priority 962 * without taking RT-inheritance into account. Might be 963 * boosted by interactivity modifiers. Changes upon fork, 964 * setprio syscalls, and whenever the interactivity 965 * estimator recalculates. 966 */ 967static inline int normal_prio(struct task_struct *p) 968{ 969 int prio; 970 971 if (task_has_dl_policy(p)) 972 prio = MAX_DL_PRIO-1; 973 else if (task_has_rt_policy(p)) 974 prio = MAX_RT_PRIO-1 - p->rt_priority; 975 else 976 prio = __normal_prio(p); 977 return prio; 978} 979 980/* 981 * Calculate the current priority, i.e. the priority 982 * taken into account by the scheduler. This value might 983 * be boosted by RT tasks, or might be boosted by 984 * interactivity modifiers. Will be RT if the task got 985 * RT-boosted. If not then it returns p->normal_prio. 986 */ 987static int effective_prio(struct task_struct *p) 988{ 989 p->normal_prio = normal_prio(p); 990 /* 991 * If we are RT tasks or we were boosted to RT priority, 992 * keep the priority unchanged. Otherwise, update priority 993 * to the normal priority: 994 */ 995 if (!rt_prio(p->prio)) 996 return p->normal_prio; 997 return p->prio; 998} 999 1000/** 1001 * task_curr - is this task currently executing on a CPU? 1002 * @p: the task in question. 1003 * 1004 * Return: 1 if the task is currently executing. 0 otherwise. 1005 */ 1006inline int task_curr(const struct task_struct *p) 1007{ 1008 return cpu_curr(task_cpu(p)) == p; 1009} 1010 1011static inline void check_class_changed(struct rq *rq, struct task_struct *p, 1012 const struct sched_class *prev_class, 1013 int oldprio) 1014{ 1015 if (prev_class != p->sched_class) { 1016 if (prev_class->switched_from) 1017 prev_class->switched_from(rq, p); 1018 p->sched_class->switched_to(rq, p); 1019 } else if (oldprio != p->prio || dl_task(p)) 1020 p->sched_class->prio_changed(rq, p, oldprio); 1021} 1022 1023void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags) 1024{ 1025 const struct sched_class *class; 1026 1027 if (p->sched_class == rq->curr->sched_class) { 1028 rq->curr->sched_class->check_preempt_curr(rq, p, flags); 1029 } else { 1030 for_each_class(class) { 1031 if (class == rq->curr->sched_class) 1032 break; 1033 if (class == p->sched_class) { 1034 resched_curr(rq); 1035 break; 1036 } 1037 } 1038 } 1039 1040 /* 1041 * A queue event has occurred, and we're going to schedule. In 1042 * this case, we can save a useless back to back clock update. 1043 */ 1044 if (task_on_rq_queued(rq->curr) && test_tsk_need_resched(rq->curr)) 1045 rq->skip_clock_update = 1; 1046} 1047 1048#ifdef CONFIG_SMP 1049void set_task_cpu(struct task_struct *p, unsigned int new_cpu) 1050{ 1051#ifdef CONFIG_SCHED_DEBUG 1052 /* 1053 * We should never call set_task_cpu() on a blocked task, 1054 * ttwu() will sort out the placement. 1055 */ 1056 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING && 1057 !(task_preempt_count(p) & PREEMPT_ACTIVE)); 1058 1059#ifdef CONFIG_LOCKDEP 1060 /* 1061 * The caller should hold either p->pi_lock or rq->lock, when changing 1062 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks. 1063 * 1064 * sched_move_task() holds both and thus holding either pins the cgroup, 1065 * see task_group(). 1066 * 1067 * Furthermore, all task_rq users should acquire both locks, see 1068 * task_rq_lock(). 1069 */ 1070 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) || 1071 lockdep_is_held(&task_rq(p)->lock))); 1072#endif 1073#endif 1074 1075 trace_sched_migrate_task(p, new_cpu); 1076 1077 if (task_cpu(p) != new_cpu) { 1078 if (p->sched_class->migrate_task_rq) 1079 p->sched_class->migrate_task_rq(p, new_cpu); 1080 p->se.nr_migrations++; 1081 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, NULL, 0); 1082 } 1083 1084 __set_task_cpu(p, new_cpu); 1085} 1086 1087static void __migrate_swap_task(struct task_struct *p, int cpu) 1088{ 1089 if (task_on_rq_queued(p)) { 1090 struct rq *src_rq, *dst_rq; 1091 1092 src_rq = task_rq(p); 1093 dst_rq = cpu_rq(cpu); 1094 1095 deactivate_task(src_rq, p, 0); 1096 set_task_cpu(p, cpu); 1097 activate_task(dst_rq, p, 0); 1098 check_preempt_curr(dst_rq, p, 0); 1099 } else { 1100 /* 1101 * Task isn't running anymore; make it appear like we migrated 1102 * it before it went to sleep. This means on wakeup we make the 1103 * previous cpu our targer instead of where it really is. 1104 */ 1105 p->wake_cpu = cpu; 1106 } 1107} 1108 1109struct migration_swap_arg { 1110 struct task_struct *src_task, *dst_task; 1111 int src_cpu, dst_cpu; 1112}; 1113 1114static int migrate_swap_stop(void *data) 1115{ 1116 struct migration_swap_arg *arg = data; 1117 struct rq *src_rq, *dst_rq; 1118 int ret = -EAGAIN; 1119 1120 src_rq = cpu_rq(arg->src_cpu); 1121 dst_rq = cpu_rq(arg->dst_cpu); 1122 1123 double_raw_lock(&arg->src_task->pi_lock, 1124 &arg->dst_task->pi_lock); 1125 double_rq_lock(src_rq, dst_rq); 1126 if (task_cpu(arg->dst_task) != arg->dst_cpu) 1127 goto unlock; 1128 1129 if (task_cpu(arg->src_task) != arg->src_cpu) 1130 goto unlock; 1131 1132 if (!cpumask_test_cpu(arg->dst_cpu, tsk_cpus_allowed(arg->src_task))) 1133 goto unlock; 1134 1135 if (!cpumask_test_cpu(arg->src_cpu, tsk_cpus_allowed(arg->dst_task))) 1136 goto unlock; 1137 1138 __migrate_swap_task(arg->src_task, arg->dst_cpu); 1139 __migrate_swap_task(arg->dst_task, arg->src_cpu); 1140 1141 ret = 0; 1142 1143unlock: 1144 double_rq_unlock(src_rq, dst_rq); 1145 raw_spin_unlock(&arg->dst_task->pi_lock); 1146 raw_spin_unlock(&arg->src_task->pi_lock); 1147 1148 return ret; 1149} 1150 1151/* 1152 * Cross migrate two tasks 1153 */ 1154int migrate_swap(struct task_struct *cur, struct task_struct *p) 1155{ 1156 struct migration_swap_arg arg; 1157 int ret = -EINVAL; 1158 1159 arg = (struct migration_swap_arg){ 1160 .src_task = cur, 1161 .src_cpu = task_cpu(cur), 1162 .dst_task = p, 1163 .dst_cpu = task_cpu(p), 1164 }; 1165 1166 if (arg.src_cpu == arg.dst_cpu) 1167 goto out; 1168 1169 /* 1170 * These three tests are all lockless; this is OK since all of them 1171 * will be re-checked with proper locks held further down the line. 1172 */ 1173 if (!cpu_active(arg.src_cpu) || !cpu_active(arg.dst_cpu)) 1174 goto out; 1175 1176 if (!cpumask_test_cpu(arg.dst_cpu, tsk_cpus_allowed(arg.src_task))) 1177 goto out; 1178 1179 if (!cpumask_test_cpu(arg.src_cpu, tsk_cpus_allowed(arg.dst_task))) 1180 goto out; 1181 1182 trace_sched_swap_numa(cur, arg.src_cpu, p, arg.dst_cpu); 1183 ret = stop_two_cpus(arg.dst_cpu, arg.src_cpu, migrate_swap_stop, &arg); 1184 1185out: 1186 return ret; 1187} 1188 1189struct migration_arg { 1190 struct task_struct *task; 1191 int dest_cpu; 1192}; 1193 1194static int migration_cpu_stop(void *data); 1195 1196/* 1197 * wait_task_inactive - wait for a thread to unschedule. 1198 * 1199 * If @match_state is nonzero, it's the @p->state value just checked and 1200 * not expected to change. If it changes, i.e. @p might have woken up, 1201 * then return zero. When we succeed in waiting for @p to be off its CPU, 1202 * we return a positive number (its total switch count). If a second call 1203 * a short while later returns the same number, the caller can be sure that 1204 * @p has remained unscheduled the whole time. 1205 * 1206 * The caller must ensure that the task *will* unschedule sometime soon, 1207 * else this function might spin for a *long* time. This function can't 1208 * be called with interrupts off, or it may introduce deadlock with 1209 * smp_call_function() if an IPI is sent by the same process we are 1210 * waiting to become inactive. 1211 */ 1212unsigned long wait_task_inactive(struct task_struct *p, long match_state) 1213{ 1214 unsigned long flags; 1215 int running, queued; 1216 unsigned long ncsw; 1217 struct rq *rq; 1218 1219 for (;;) { 1220 /* 1221 * We do the initial early heuristics without holding 1222 * any task-queue locks at all. We'll only try to get 1223 * the runqueue lock when things look like they will 1224 * work out! 1225 */ 1226 rq = task_rq(p); 1227 1228 /* 1229 * If the task is actively running on another CPU 1230 * still, just relax and busy-wait without holding 1231 * any locks. 1232 * 1233 * NOTE! Since we don't hold any locks, it's not 1234 * even sure that "rq" stays as the right runqueue! 1235 * But we don't care, since "task_running()" will 1236 * return false if the runqueue has changed and p 1237 * is actually now running somewhere else! 1238 */ 1239 while (task_running(rq, p)) { 1240 if (match_state && unlikely(p->state != match_state)) 1241 return 0; 1242 cpu_relax(); 1243 } 1244 1245 /* 1246 * Ok, time to look more closely! We need the rq 1247 * lock now, to be *sure*. If we're wrong, we'll 1248 * just go back and repeat. 1249 */ 1250 rq = task_rq_lock(p, &flags); 1251 trace_sched_wait_task(p); 1252 running = task_running(rq, p); 1253 queued = task_on_rq_queued(p); 1254 ncsw = 0; 1255 if (!match_state || p->state == match_state) 1256 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */ 1257 task_rq_unlock(rq, p, &flags); 1258 1259 /* 1260 * If it changed from the expected state, bail out now. 1261 */ 1262 if (unlikely(!ncsw)) 1263 break; 1264 1265 /* 1266 * Was it really running after all now that we 1267 * checked with the proper locks actually held? 1268 * 1269 * Oops. Go back and try again.. 1270 */ 1271 if (unlikely(running)) { 1272 cpu_relax(); 1273 continue; 1274 } 1275 1276 /* 1277 * It's not enough that it's not actively running, 1278 * it must be off the runqueue _entirely_, and not 1279 * preempted! 1280 * 1281 * So if it was still runnable (but just not actively 1282 * running right now), it's preempted, and we should 1283 * yield - it could be a while. 1284 */ 1285 if (unlikely(queued)) { 1286 ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ); 1287 1288 set_current_state(TASK_UNINTERRUPTIBLE); 1289 schedule_hrtimeout(&to, HRTIMER_MODE_REL); 1290 continue; 1291 } 1292 1293 /* 1294 * Ahh, all good. It wasn't running, and it wasn't 1295 * runnable, which means that it will never become 1296 * running in the future either. We're all done! 1297 */ 1298 break; 1299 } 1300 1301 return ncsw; 1302} 1303 1304/*** 1305 * kick_process - kick a running thread to enter/exit the kernel 1306 * @p: the to-be-kicked thread 1307 * 1308 * Cause a process which is running on another CPU to enter 1309 * kernel-mode, without any delay. (to get signals handled.) 1310 * 1311 * NOTE: this function doesn't have to take the runqueue lock, 1312 * because all it wants to ensure is that the remote task enters 1313 * the kernel. If the IPI races and the task has been migrated 1314 * to another CPU then no harm is done and the purpose has been 1315 * achieved as well. 1316 */ 1317void kick_process(struct task_struct *p) 1318{ 1319 int cpu; 1320 1321 preempt_disable(); 1322 cpu = task_cpu(p); 1323 if ((cpu != smp_processor_id()) && task_curr(p)) 1324 smp_send_reschedule(cpu); 1325 preempt_enable(); 1326} 1327EXPORT_SYMBOL_GPL(kick_process); 1328#endif /* CONFIG_SMP */ 1329 1330#ifdef CONFIG_SMP 1331/* 1332 * ->cpus_allowed is protected by both rq->lock and p->pi_lock 1333 */ 1334static int select_fallback_rq(int cpu, struct task_struct *p) 1335{ 1336 int nid = cpu_to_node(cpu); 1337 const struct cpumask *nodemask = NULL; 1338 enum { cpuset, possible, fail } state = cpuset; 1339 int dest_cpu; 1340 1341 /* 1342 * If the node that the cpu is on has been offlined, cpu_to_node() 1343 * will return -1. There is no cpu on the node, and we should 1344 * select the cpu on the other node. 1345 */ 1346 if (nid != -1) { 1347 nodemask = cpumask_of_node(nid); 1348 1349 /* Look for allowed, online CPU in same node. */ 1350 for_each_cpu(dest_cpu, nodemask) { 1351 if (!cpu_online(dest_cpu)) 1352 continue; 1353 if (!cpu_active(dest_cpu)) 1354 continue; 1355 if (cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p))) 1356 return dest_cpu; 1357 } 1358 } 1359 1360 for (;;) { 1361 /* Any allowed, online CPU? */ 1362 for_each_cpu(dest_cpu, tsk_cpus_allowed(p)) { 1363 if (!cpu_online(dest_cpu)) 1364 continue; 1365 if (!cpu_active(dest_cpu)) 1366 continue; 1367 goto out; 1368 } 1369 1370 switch (state) { 1371 case cpuset: 1372 /* No more Mr. Nice Guy. */ 1373 cpuset_cpus_allowed_fallback(p); 1374 state = possible; 1375 break; 1376 1377 case possible: 1378 do_set_cpus_allowed(p, cpu_possible_mask); 1379 state = fail; 1380 break; 1381 1382 case fail: 1383 BUG(); 1384 break; 1385 } 1386 } 1387 1388out: 1389 if (state != cpuset) { 1390 /* 1391 * Don't tell them about moving exiting tasks or 1392 * kernel threads (both mm NULL), since they never 1393 * leave kernel. 1394 */ 1395 if (p->mm && printk_ratelimit()) { 1396 printk_deferred("process %d (%s) no longer affine to cpu%d\n", 1397 task_pid_nr(p), p->comm, cpu); 1398 } 1399 } 1400 1401 return dest_cpu; 1402} 1403 1404/* 1405 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable. 1406 */ 1407static inline 1408int select_task_rq(struct task_struct *p, int cpu, int sd_flags, int wake_flags) 1409{ 1410 cpu = p->sched_class->select_task_rq(p, cpu, sd_flags, wake_flags); 1411 1412 /* 1413 * In order not to call set_task_cpu() on a blocking task we need 1414 * to rely on ttwu() to place the task on a valid ->cpus_allowed 1415 * cpu. 1416 * 1417 * Since this is common to all placement strategies, this lives here. 1418 * 1419 * [ this allows ->select_task() to simply return task_cpu(p) and 1420 * not worry about this generic constraint ] 1421 */ 1422 if (unlikely(!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)) || 1423 !cpu_online(cpu))) 1424 cpu = select_fallback_rq(task_cpu(p), p); 1425 1426 return cpu; 1427} 1428 1429static void update_avg(u64 *avg, u64 sample) 1430{ 1431 s64 diff = sample - *avg; 1432 *avg += diff >> 3; 1433} 1434#endif 1435 1436static void 1437ttwu_stat(struct task_struct *p, int cpu, int wake_flags) 1438{ 1439#ifdef CONFIG_SCHEDSTATS 1440 struct rq *rq = this_rq(); 1441 1442#ifdef CONFIG_SMP 1443 int this_cpu = smp_processor_id(); 1444 1445 if (cpu == this_cpu) { 1446 schedstat_inc(rq, ttwu_local); 1447 schedstat_inc(p, se.statistics.nr_wakeups_local); 1448 } else { 1449 struct sched_domain *sd; 1450 1451 schedstat_inc(p, se.statistics.nr_wakeups_remote); 1452 rcu_read_lock(); 1453 for_each_domain(this_cpu, sd) { 1454 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) { 1455 schedstat_inc(sd, ttwu_wake_remote); 1456 break; 1457 } 1458 } 1459 rcu_read_unlock(); 1460 } 1461 1462 if (wake_flags & WF_MIGRATED) 1463 schedstat_inc(p, se.statistics.nr_wakeups_migrate); 1464 1465#endif /* CONFIG_SMP */ 1466 1467 schedstat_inc(rq, ttwu_count); 1468 schedstat_inc(p, se.statistics.nr_wakeups); 1469 1470 if (wake_flags & WF_SYNC) 1471 schedstat_inc(p, se.statistics.nr_wakeups_sync); 1472 1473#endif /* CONFIG_SCHEDSTATS */ 1474} 1475 1476static void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags) 1477{ 1478 activate_task(rq, p, en_flags); 1479 p->on_rq = TASK_ON_RQ_QUEUED; 1480 1481 /* if a worker is waking up, notify workqueue */ 1482 if (p->flags & PF_WQ_WORKER) 1483 wq_worker_waking_up(p, cpu_of(rq)); 1484} 1485 1486/* 1487 * Mark the task runnable and perform wakeup-preemption. 1488 */ 1489static void 1490ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags) 1491{ 1492 check_preempt_curr(rq, p, wake_flags); 1493 trace_sched_wakeup(p, true); 1494 1495 p->state = TASK_RUNNING; 1496#ifdef CONFIG_SMP 1497 if (p->sched_class->task_woken) 1498 p->sched_class->task_woken(rq, p); 1499 1500 if (rq->idle_stamp) { 1501 u64 delta = rq_clock(rq) - rq->idle_stamp; 1502 u64 max = 2*rq->max_idle_balance_cost; 1503 1504 update_avg(&rq->avg_idle, delta); 1505 1506 if (rq->avg_idle > max) 1507 rq->avg_idle = max; 1508 1509 rq->idle_stamp = 0; 1510 } 1511#endif 1512} 1513 1514static void 1515ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags) 1516{ 1517#ifdef CONFIG_SMP 1518 if (p->sched_contributes_to_load) 1519 rq->nr_uninterruptible--; 1520#endif 1521 1522 ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_WAKING); 1523 ttwu_do_wakeup(rq, p, wake_flags); 1524} 1525 1526/* 1527 * Called in case the task @p isn't fully descheduled from its runqueue, 1528 * in this case we must do a remote wakeup. Its a 'light' wakeup though, 1529 * since all we need to do is flip p->state to TASK_RUNNING, since 1530 * the task is still ->on_rq. 1531 */ 1532static int ttwu_remote(struct task_struct *p, int wake_flags) 1533{ 1534 struct rq *rq; 1535 int ret = 0; 1536 1537 rq = __task_rq_lock(p); 1538 if (task_on_rq_queued(p)) { 1539 /* check_preempt_curr() may use rq clock */ 1540 update_rq_clock(rq); 1541 ttwu_do_wakeup(rq, p, wake_flags); 1542 ret = 1; 1543 } 1544 __task_rq_unlock(rq); 1545 1546 return ret; 1547} 1548 1549#ifdef CONFIG_SMP 1550void sched_ttwu_pending(void) 1551{ 1552 struct rq *rq = this_rq(); 1553 struct llist_node *llist = llist_del_all(&rq->wake_list); 1554 struct task_struct *p; 1555 unsigned long flags; 1556 1557 if (!llist) 1558 return; 1559 1560 raw_spin_lock_irqsave(&rq->lock, flags); 1561 1562 while (llist) { 1563 p = llist_entry(llist, struct task_struct, wake_entry); 1564 llist = llist_next(llist); 1565 ttwu_do_activate(rq, p, 0); 1566 } 1567 1568 raw_spin_unlock_irqrestore(&rq->lock, flags); 1569} 1570 1571void scheduler_ipi(void) 1572{ 1573 /* 1574 * Fold TIF_NEED_RESCHED into the preempt_count; anybody setting 1575 * TIF_NEED_RESCHED remotely (for the first time) will also send 1576 * this IPI. 1577 */ 1578 preempt_fold_need_resched(); 1579 1580 if (llist_empty(&this_rq()->wake_list) && !got_nohz_idle_kick()) 1581 return; 1582 1583 /* 1584 * Not all reschedule IPI handlers call irq_enter/irq_exit, since 1585 * traditionally all their work was done from the interrupt return 1586 * path. Now that we actually do some work, we need to make sure 1587 * we do call them. 1588 * 1589 * Some archs already do call them, luckily irq_enter/exit nest 1590 * properly. 1591 * 1592 * Arguably we should visit all archs and update all handlers, 1593 * however a fair share of IPIs are still resched only so this would 1594 * somewhat pessimize the simple resched case. 1595 */ 1596 irq_enter(); 1597 sched_ttwu_pending(); 1598 1599 /* 1600 * Check if someone kicked us for doing the nohz idle load balance. 1601 */ 1602 if (unlikely(got_nohz_idle_kick())) { 1603 this_rq()->idle_balance = 1; 1604 raise_softirq_irqoff(SCHED_SOFTIRQ); 1605 } 1606 irq_exit(); 1607} 1608 1609static void ttwu_queue_remote(struct task_struct *p, int cpu) 1610{ 1611 struct rq *rq = cpu_rq(cpu); 1612 1613 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list)) { 1614 if (!set_nr_if_polling(rq->idle)) 1615 smp_send_reschedule(cpu); 1616 else 1617 trace_sched_wake_idle_without_ipi(cpu); 1618 } 1619} 1620 1621void wake_up_if_idle(int cpu) 1622{ 1623 struct rq *rq = cpu_rq(cpu); 1624 unsigned long flags; 1625 1626 if (!is_idle_task(rq->curr)) 1627 return; 1628 1629 if (set_nr_if_polling(rq->idle)) { 1630 trace_sched_wake_idle_without_ipi(cpu); 1631 } else { 1632 raw_spin_lock_irqsave(&rq->lock, flags); 1633 if (is_idle_task(rq->curr)) 1634 smp_send_reschedule(cpu); 1635 /* Else cpu is not in idle, do nothing here */ 1636 raw_spin_unlock_irqrestore(&rq->lock, flags); 1637 } 1638} 1639 1640bool cpus_share_cache(int this_cpu, int that_cpu) 1641{ 1642 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu); 1643} 1644#endif /* CONFIG_SMP */ 1645 1646static void ttwu_queue(struct task_struct *p, int cpu) 1647{ 1648 struct rq *rq = cpu_rq(cpu); 1649 1650#if defined(CONFIG_SMP) 1651 if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) { 1652 sched_clock_cpu(cpu); /* sync clocks x-cpu */ 1653 ttwu_queue_remote(p, cpu); 1654 return; 1655 } 1656#endif 1657 1658 raw_spin_lock(&rq->lock); 1659 ttwu_do_activate(rq, p, 0); 1660 raw_spin_unlock(&rq->lock); 1661} 1662 1663/** 1664 * try_to_wake_up - wake up a thread 1665 * @p: the thread to be awakened 1666 * @state: the mask of task states that can be woken 1667 * @wake_flags: wake modifier flags (WF_*) 1668 * 1669 * Put it on the run-queue if it's not already there. The "current" 1670 * thread is always on the run-queue (except when the actual 1671 * re-schedule is in progress), and as such you're allowed to do 1672 * the simpler "current->state = TASK_RUNNING" to mark yourself 1673 * runnable without the overhead of this. 1674 * 1675 * Return: %true if @p was woken up, %false if it was already running. 1676 * or @state didn't match @p's state. 1677 */ 1678static int 1679try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags) 1680{ 1681 unsigned long flags; 1682 int cpu, success = 0; 1683 1684 /* 1685 * If we are going to wake up a thread waiting for CONDITION we 1686 * need to ensure that CONDITION=1 done by the caller can not be 1687 * reordered with p->state check below. This pairs with mb() in 1688 * set_current_state() the waiting thread does. 1689 */ 1690 smp_mb__before_spinlock(); 1691 raw_spin_lock_irqsave(&p->pi_lock, flags); 1692 if (!(p->state & state)) 1693 goto out; 1694 1695 success = 1; /* we're going to change ->state */ 1696 cpu = task_cpu(p); 1697 1698 if (p->on_rq && ttwu_remote(p, wake_flags)) 1699 goto stat; 1700 1701#ifdef CONFIG_SMP 1702 /* 1703 * If the owning (remote) cpu is still in the middle of schedule() with 1704 * this task as prev, wait until its done referencing the task. 1705 */ 1706 while (p->on_cpu) 1707 cpu_relax(); 1708 /* 1709 * Pairs with the smp_wmb() in finish_lock_switch(). 1710 */ 1711 smp_rmb(); 1712 1713 p->sched_contributes_to_load = !!task_contributes_to_load(p); 1714 p->state = TASK_WAKING; 1715 1716 if (p->sched_class->task_waking) 1717 p->sched_class->task_waking(p); 1718 1719 cpu = select_task_rq(p, p->wake_cpu, SD_BALANCE_WAKE, wake_flags); 1720 if (task_cpu(p) != cpu) { 1721 wake_flags |= WF_MIGRATED; 1722 set_task_cpu(p, cpu); 1723 } 1724#endif /* CONFIG_SMP */ 1725 1726 ttwu_queue(p, cpu); 1727stat: 1728 ttwu_stat(p, cpu, wake_flags); 1729out: 1730 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 1731 1732 return success; 1733} 1734 1735/** 1736 * try_to_wake_up_local - try to wake up a local task with rq lock held 1737 * @p: the thread to be awakened 1738 * 1739 * Put @p on the run-queue if it's not already there. The caller must 1740 * ensure that this_rq() is locked, @p is bound to this_rq() and not 1741 * the current task. 1742 */ 1743static void try_to_wake_up_local(struct task_struct *p) 1744{ 1745 struct rq *rq = task_rq(p); 1746 1747 if (WARN_ON_ONCE(rq != this_rq()) || 1748 WARN_ON_ONCE(p == current)) 1749 return; 1750 1751 lockdep_assert_held(&rq->lock); 1752 1753 if (!raw_spin_trylock(&p->pi_lock)) { 1754 raw_spin_unlock(&rq->lock); 1755 raw_spin_lock(&p->pi_lock); 1756 raw_spin_lock(&rq->lock); 1757 } 1758 1759 if (!(p->state & TASK_NORMAL)) 1760 goto out; 1761 1762 if (!task_on_rq_queued(p)) 1763 ttwu_activate(rq, p, ENQUEUE_WAKEUP); 1764 1765 ttwu_do_wakeup(rq, p, 0); 1766 ttwu_stat(p, smp_processor_id(), 0); 1767out: 1768 raw_spin_unlock(&p->pi_lock); 1769} 1770 1771/** 1772 * wake_up_process - Wake up a specific process 1773 * @p: The process to be woken up. 1774 * 1775 * Attempt to wake up the nominated process and move it to the set of runnable 1776 * processes. 1777 * 1778 * Return: 1 if the process was woken up, 0 if it was already running. 1779 * 1780 * It may be assumed that this function implies a write memory barrier before 1781 * changing the task state if and only if any tasks are woken up. 1782 */ 1783int wake_up_process(struct task_struct *p) 1784{ 1785 WARN_ON(task_is_stopped_or_traced(p)); 1786 return try_to_wake_up(p, TASK_NORMAL, 0); 1787} 1788EXPORT_SYMBOL(wake_up_process); 1789 1790int wake_up_state(struct task_struct *p, unsigned int state) 1791{ 1792 return try_to_wake_up(p, state, 0); 1793} 1794 1795/* 1796 * This function clears the sched_dl_entity static params. 1797 */ 1798void __dl_clear_params(struct task_struct *p) 1799{ 1800 struct sched_dl_entity *dl_se = &p->dl; 1801 1802 dl_se->dl_runtime = 0; 1803 dl_se->dl_deadline = 0; 1804 dl_se->dl_period = 0; 1805 dl_se->flags = 0; 1806 dl_se->dl_bw = 0; 1807} 1808 1809/* 1810 * Perform scheduler related setup for a newly forked process p. 1811 * p is forked by current. 1812 * 1813 * __sched_fork() is basic setup used by init_idle() too: 1814 */ 1815static void __sched_fork(unsigned long clone_flags, struct task_struct *p) 1816{ 1817 p->on_rq = 0; 1818 1819 p->se.on_rq = 0; 1820 p->se.exec_start = 0; 1821 p->se.sum_exec_runtime = 0; 1822 p->se.prev_sum_exec_runtime = 0; 1823 p->se.nr_migrations = 0; 1824 p->se.vruntime = 0; 1825 INIT_LIST_HEAD(&p->se.group_node); 1826 1827#ifdef CONFIG_SCHEDSTATS 1828 memset(&p->se.statistics, 0, sizeof(p->se.statistics)); 1829#endif 1830 1831 RB_CLEAR_NODE(&p->dl.rb_node); 1832 hrtimer_init(&p->dl.dl_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); 1833 __dl_clear_params(p); 1834 1835 INIT_LIST_HEAD(&p->rt.run_list); 1836 1837#ifdef CONFIG_PREEMPT_NOTIFIERS 1838 INIT_HLIST_HEAD(&p->preempt_notifiers); 1839#endif 1840 1841#ifdef CONFIG_NUMA_BALANCING 1842 if (p->mm && atomic_read(&p->mm->mm_users) == 1) { 1843 p->mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay); 1844 p->mm->numa_scan_seq = 0; 1845 } 1846 1847 if (clone_flags & CLONE_VM) 1848 p->numa_preferred_nid = current->numa_preferred_nid; 1849 else 1850 p->numa_preferred_nid = -1; 1851 1852 p->node_stamp = 0ULL; 1853 p->numa_scan_seq = p->mm ? p->mm->numa_scan_seq : 0; 1854 p->numa_scan_period = sysctl_numa_balancing_scan_delay; 1855 p->numa_work.next = &p->numa_work; 1856 p->numa_faults_memory = NULL; 1857 p->numa_faults_buffer_memory = NULL; 1858 p->last_task_numa_placement = 0; 1859 p->last_sum_exec_runtime = 0; 1860 1861 INIT_LIST_HEAD(&p->numa_entry); 1862 p->numa_group = NULL; 1863#endif /* CONFIG_NUMA_BALANCING */ 1864} 1865 1866#ifdef CONFIG_NUMA_BALANCING 1867#ifdef CONFIG_SCHED_DEBUG 1868void set_numabalancing_state(bool enabled) 1869{ 1870 if (enabled) 1871 sched_feat_set("NUMA"); 1872 else 1873 sched_feat_set("NO_NUMA"); 1874} 1875#else 1876__read_mostly bool numabalancing_enabled; 1877 1878void set_numabalancing_state(bool enabled) 1879{ 1880 numabalancing_enabled = enabled; 1881} 1882#endif /* CONFIG_SCHED_DEBUG */ 1883 1884#ifdef CONFIG_PROC_SYSCTL 1885int sysctl_numa_balancing(struct ctl_table *table, int write, 1886 void __user *buffer, size_t *lenp, loff_t *ppos) 1887{ 1888 struct ctl_table t; 1889 int err; 1890 int state = numabalancing_enabled; 1891 1892 if (write && !capable(CAP_SYS_ADMIN)) 1893 return -EPERM; 1894 1895 t = *table; 1896 t.data = &state; 1897 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos); 1898 if (err < 0) 1899 return err; 1900 if (write) 1901 set_numabalancing_state(state); 1902 return err; 1903} 1904#endif 1905#endif 1906 1907/* 1908 * fork()/clone()-time setup: 1909 */ 1910int sched_fork(unsigned long clone_flags, struct task_struct *p) 1911{ 1912 unsigned long flags; 1913 int cpu = get_cpu(); 1914 1915 __sched_fork(clone_flags, p); 1916 /* 1917 * We mark the process as running here. This guarantees that 1918 * nobody will actually run it, and a signal or other external 1919 * event cannot wake it up and insert it on the runqueue either. 1920 */ 1921 p->state = TASK_RUNNING; 1922 1923 /* 1924 * Make sure we do not leak PI boosting priority to the child. 1925 */ 1926 p->prio = current->normal_prio; 1927 1928 /* 1929 * Revert to default priority/policy on fork if requested. 1930 */ 1931 if (unlikely(p->sched_reset_on_fork)) { 1932 if (task_has_dl_policy(p) || task_has_rt_policy(p)) { 1933 p->policy = SCHED_NORMAL; 1934 p->static_prio = NICE_TO_PRIO(0); 1935 p->rt_priority = 0; 1936 } else if (PRIO_TO_NICE(p->static_prio) < 0) 1937 p->static_prio = NICE_TO_PRIO(0); 1938 1939 p->prio = p->normal_prio = __normal_prio(p); 1940 set_load_weight(p); 1941 1942 /* 1943 * We don't need the reset flag anymore after the fork. It has 1944 * fulfilled its duty: 1945 */ 1946 p->sched_reset_on_fork = 0; 1947 } 1948 1949 if (dl_prio(p->prio)) { 1950 put_cpu(); 1951 return -EAGAIN; 1952 } else if (rt_prio(p->prio)) { 1953 p->sched_class = &rt_sched_class; 1954 } else { 1955 p->sched_class = &fair_sched_class; 1956 } 1957 1958 if (p->sched_class->task_fork) 1959 p->sched_class->task_fork(p); 1960 1961 /* 1962 * The child is not yet in the pid-hash so no cgroup attach races, 1963 * and the cgroup is pinned to this child due to cgroup_fork() 1964 * is ran before sched_fork(). 1965 * 1966 * Silence PROVE_RCU. 1967 */ 1968 raw_spin_lock_irqsave(&p->pi_lock, flags); 1969 set_task_cpu(p, cpu); 1970 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 1971 1972#if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT) 1973 if (likely(sched_info_on())) 1974 memset(&p->sched_info, 0, sizeof(p->sched_info)); 1975#endif 1976#if defined(CONFIG_SMP) 1977 p->on_cpu = 0; 1978#endif 1979 init_task_preempt_count(p); 1980#ifdef CONFIG_SMP 1981 plist_node_init(&p->pushable_tasks, MAX_PRIO); 1982 RB_CLEAR_NODE(&p->pushable_dl_tasks); 1983#endif 1984 1985 put_cpu(); 1986 return 0; 1987} 1988 1989unsigned long to_ratio(u64 period, u64 runtime) 1990{ 1991 if (runtime == RUNTIME_INF) 1992 return 1ULL << 20; 1993 1994 /* 1995 * Doing this here saves a lot of checks in all 1996 * the calling paths, and returning zero seems 1997 * safe for them anyway. 1998 */ 1999 if (period == 0) 2000 return 0; 2001 2002 return div64_u64(runtime << 20, period); 2003} 2004 2005#ifdef CONFIG_SMP 2006inline struct dl_bw *dl_bw_of(int i) 2007{ 2008 rcu_lockdep_assert(rcu_read_lock_sched_held(), 2009 "sched RCU must be held"); 2010 return &cpu_rq(i)->rd->dl_bw; 2011} 2012 2013static inline int dl_bw_cpus(int i) 2014{ 2015 struct root_domain *rd = cpu_rq(i)->rd; 2016 int cpus = 0; 2017 2018 rcu_lockdep_assert(rcu_read_lock_sched_held(), 2019 "sched RCU must be held"); 2020 for_each_cpu_and(i, rd->span, cpu_active_mask) 2021 cpus++; 2022 2023 return cpus; 2024} 2025#else 2026inline struct dl_bw *dl_bw_of(int i) 2027{ 2028 return &cpu_rq(i)->dl.dl_bw; 2029} 2030 2031static inline int dl_bw_cpus(int i) 2032{ 2033 return 1; 2034} 2035#endif 2036 2037static inline 2038void __dl_clear(struct dl_bw *dl_b, u64 tsk_bw) 2039{ 2040 dl_b->total_bw -= tsk_bw; 2041} 2042 2043static inline 2044void __dl_add(struct dl_bw *dl_b, u64 tsk_bw) 2045{ 2046 dl_b->total_bw += tsk_bw; 2047} 2048 2049static inline 2050bool __dl_overflow(struct dl_bw *dl_b, int cpus, u64 old_bw, u64 new_bw) 2051{ 2052 return dl_b->bw != -1 && 2053 dl_b->bw * cpus < dl_b->total_bw - old_bw + new_bw; 2054} 2055 2056/* 2057 * We must be sure that accepting a new task (or allowing changing the 2058 * parameters of an existing one) is consistent with the bandwidth 2059 * constraints. If yes, this function also accordingly updates the currently 2060 * allocated bandwidth to reflect the new situation. 2061 * 2062 * This function is called while holding p's rq->lock. 2063 */ 2064static int dl_overflow(struct task_struct *p, int policy, 2065 const struct sched_attr *attr) 2066{ 2067 2068 struct dl_bw *dl_b = dl_bw_of(task_cpu(p)); 2069 u64 period = attr->sched_period ?: attr->sched_deadline; 2070 u64 runtime = attr->sched_runtime; 2071 u64 new_bw = dl_policy(policy) ? to_ratio(period, runtime) : 0; 2072 int cpus, err = -1; 2073 2074 if (new_bw == p->dl.dl_bw) 2075 return 0; 2076 2077 /* 2078 * Either if a task, enters, leave, or stays -deadline but changes 2079 * its parameters, we may need to update accordingly the total 2080 * allocated bandwidth of the container. 2081 */ 2082 raw_spin_lock(&dl_b->lock); 2083 cpus = dl_bw_cpus(task_cpu(p)); 2084 if (dl_policy(policy) && !task_has_dl_policy(p) && 2085 !__dl_overflow(dl_b, cpus, 0, new_bw)) { 2086 __dl_add(dl_b, new_bw); 2087 err = 0; 2088 } else if (dl_policy(policy) && task_has_dl_policy(p) && 2089 !__dl_overflow(dl_b, cpus, p->dl.dl_bw, new_bw)) { 2090 __dl_clear(dl_b, p->dl.dl_bw); 2091 __dl_add(dl_b, new_bw); 2092 err = 0; 2093 } else if (!dl_policy(policy) && task_has_dl_policy(p)) { 2094 __dl_clear(dl_b, p->dl.dl_bw); 2095 err = 0; 2096 } 2097 raw_spin_unlock(&dl_b->lock); 2098 2099 return err; 2100} 2101 2102extern void init_dl_bw(struct dl_bw *dl_b); 2103 2104/* 2105 * wake_up_new_task - wake up a newly created task for the first time. 2106 * 2107 * This function will do some initial scheduler statistics housekeeping 2108 * that must be done for every newly created context, then puts the task 2109 * on the runqueue and wakes it. 2110 */ 2111void wake_up_new_task(struct task_struct *p) 2112{ 2113 unsigned long flags; 2114 struct rq *rq; 2115 2116 raw_spin_lock_irqsave(&p->pi_lock, flags); 2117#ifdef CONFIG_SMP 2118 /* 2119 * Fork balancing, do it here and not earlier because: 2120 * - cpus_allowed can change in the fork path 2121 * - any previously selected cpu might disappear through hotplug 2122 */ 2123 set_task_cpu(p, select_task_rq(p, task_cpu(p), SD_BALANCE_FORK, 0)); 2124#endif 2125 2126 /* Initialize new task's runnable average */ 2127 init_task_runnable_average(p); 2128 rq = __task_rq_lock(p); 2129 activate_task(rq, p, 0); 2130 p->on_rq = TASK_ON_RQ_QUEUED; 2131 trace_sched_wakeup_new(p, true); 2132 check_preempt_curr(rq, p, WF_FORK); 2133#ifdef CONFIG_SMP 2134 if (p->sched_class->task_woken) 2135 p->sched_class->task_woken(rq, p); 2136#endif 2137 task_rq_unlock(rq, p, &flags); 2138} 2139 2140#ifdef CONFIG_PREEMPT_NOTIFIERS 2141 2142/** 2143 * preempt_notifier_register - tell me when current is being preempted & rescheduled 2144 * @notifier: notifier struct to register 2145 */ 2146void preempt_notifier_register(struct preempt_notifier *notifier) 2147{ 2148 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers); 2149} 2150EXPORT_SYMBOL_GPL(preempt_notifier_register); 2151 2152/** 2153 * preempt_notifier_unregister - no longer interested in preemption notifications 2154 * @notifier: notifier struct to unregister 2155 * 2156 * This is safe to call from within a preemption notifier. 2157 */ 2158void preempt_notifier_unregister(struct preempt_notifier *notifier) 2159{ 2160 hlist_del(¬ifier->link); 2161} 2162EXPORT_SYMBOL_GPL(preempt_notifier_unregister); 2163 2164static void fire_sched_in_preempt_notifiers(struct task_struct *curr) 2165{ 2166 struct preempt_notifier *notifier; 2167 2168 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link) 2169 notifier->ops->sched_in(notifier, raw_smp_processor_id()); 2170} 2171 2172static void 2173fire_sched_out_preempt_notifiers(struct task_struct *curr, 2174 struct task_struct *next) 2175{ 2176 struct preempt_notifier *notifier; 2177 2178 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link) 2179 notifier->ops->sched_out(notifier, next); 2180} 2181 2182#else /* !CONFIG_PREEMPT_NOTIFIERS */ 2183 2184static void fire_sched_in_preempt_notifiers(struct task_struct *curr) 2185{ 2186} 2187 2188static void 2189fire_sched_out_preempt_notifiers(struct task_struct *curr, 2190 struct task_struct *next) 2191{ 2192} 2193 2194#endif /* CONFIG_PREEMPT_NOTIFIERS */ 2195 2196/** 2197 * prepare_task_switch - prepare to switch tasks 2198 * @rq: the runqueue preparing to switch 2199 * @prev: the current task that is being switched out 2200 * @next: the task we are going to switch to. 2201 * 2202 * This is called with the rq lock held and interrupts off. It must 2203 * be paired with a subsequent finish_task_switch after the context 2204 * switch. 2205 * 2206 * prepare_task_switch sets up locking and calls architecture specific 2207 * hooks. 2208 */ 2209static inline void 2210prepare_task_switch(struct rq *rq, struct task_struct *prev, 2211 struct task_struct *next) 2212{ 2213 trace_sched_switch(prev, next); 2214 sched_info_switch(rq, prev, next); 2215 perf_event_task_sched_out(prev, next); 2216 fire_sched_out_preempt_notifiers(prev, next); 2217 prepare_lock_switch(rq, next); 2218 prepare_arch_switch(next); 2219} 2220 2221/** 2222 * finish_task_switch - clean up after a task-switch 2223 * @rq: runqueue associated with task-switch 2224 * @prev: the thread we just switched away from. 2225 * 2226 * finish_task_switch must be called after the context switch, paired 2227 * with a prepare_task_switch call before the context switch. 2228 * finish_task_switch will reconcile locking set up by prepare_task_switch, 2229 * and do any other architecture-specific cleanup actions. 2230 * 2231 * Note that we may have delayed dropping an mm in context_switch(). If 2232 * so, we finish that here outside of the runqueue lock. (Doing it 2233 * with the lock held can cause deadlocks; see schedule() for 2234 * details.) 2235 */ 2236static void finish_task_switch(struct rq *rq, struct task_struct *prev) 2237 __releases(rq->lock) 2238{ 2239 struct mm_struct *mm = rq->prev_mm; 2240 long prev_state; 2241 2242 rq->prev_mm = NULL; 2243 2244 /* 2245 * A task struct has one reference for the use as "current". 2246 * If a task dies, then it sets TASK_DEAD in tsk->state and calls 2247 * schedule one last time. The schedule call will never return, and 2248 * the scheduled task must drop that reference. 2249 * The test for TASK_DEAD must occur while the runqueue locks are 2250 * still held, otherwise prev could be scheduled on another cpu, die 2251 * there before we look at prev->state, and then the reference would 2252 * be dropped twice. 2253 * Manfred Spraul <manfred@colorfullife.com> 2254 */ 2255 prev_state = prev->state; 2256 vtime_task_switch(prev); 2257 finish_arch_switch(prev); 2258 perf_event_task_sched_in(prev, current); 2259 finish_lock_switch(rq, prev); 2260 finish_arch_post_lock_switch(); 2261 2262 fire_sched_in_preempt_notifiers(current); 2263 if (mm) 2264 mmdrop(mm); 2265 if (unlikely(prev_state == TASK_DEAD)) { 2266 if (prev->sched_class->task_dead) 2267 prev->sched_class->task_dead(prev); 2268 2269 /* 2270 * Remove function-return probe instances associated with this 2271 * task and put them back on the free list. 2272 */ 2273 kprobe_flush_task(prev); 2274 put_task_struct(prev); 2275 } 2276 2277 tick_nohz_task_switch(current); 2278} 2279 2280#ifdef CONFIG_SMP 2281 2282/* rq->lock is NOT held, but preemption is disabled */ 2283static inline void post_schedule(struct rq *rq) 2284{ 2285 if (rq->post_schedule) { 2286 unsigned long flags; 2287 2288 raw_spin_lock_irqsave(&rq->lock, flags); 2289 if (rq->curr->sched_class->post_schedule) 2290 rq->curr->sched_class->post_schedule(rq); 2291 raw_spin_unlock_irqrestore(&rq->lock, flags); 2292 2293 rq->post_schedule = 0; 2294 } 2295} 2296 2297#else 2298 2299static inline void post_schedule(struct rq *rq) 2300{ 2301} 2302 2303#endif 2304 2305/** 2306 * schedule_tail - first thing a freshly forked thread must call. 2307 * @prev: the thread we just switched away from. 2308 */ 2309asmlinkage __visible void schedule_tail(struct task_struct *prev) 2310 __releases(rq->lock) 2311{ 2312 struct rq *rq = this_rq(); 2313 2314 finish_task_switch(rq, prev); 2315 2316 /* 2317 * FIXME: do we need to worry about rq being invalidated by the 2318 * task_switch? 2319 */ 2320 post_schedule(rq); 2321 2322 if (current->set_child_tid) 2323 put_user(task_pid_vnr(current), current->set_child_tid); 2324} 2325 2326/* 2327 * context_switch - switch to the new MM and the new 2328 * thread's register state. 2329 */ 2330static inline void 2331context_switch(struct rq *rq, struct task_struct *prev, 2332 struct task_struct *next) 2333{ 2334 struct mm_struct *mm, *oldmm; 2335 2336 prepare_task_switch(rq, prev, next); 2337 2338 mm = next->mm; 2339 oldmm = prev->active_mm; 2340 /* 2341 * For paravirt, this is coupled with an exit in switch_to to 2342 * combine the page table reload and the switch backend into 2343 * one hypercall. 2344 */ 2345 arch_start_context_switch(prev); 2346 2347 if (!mm) { 2348 next->active_mm = oldmm; 2349 atomic_inc(&oldmm->mm_count); 2350 enter_lazy_tlb(oldmm, next); 2351 } else 2352 switch_mm(oldmm, mm, next); 2353 2354 if (!prev->mm) { 2355 prev->active_mm = NULL; 2356 rq->prev_mm = oldmm; 2357 } 2358 /* 2359 * Since the runqueue lock will be released by the next 2360 * task (which is an invalid locking op but in the case 2361 * of the scheduler it's an obvious special-case), so we 2362 * do an early lockdep release here: 2363 */ 2364 spin_release(&rq->lock.dep_map, 1, _THIS_IP_); 2365 2366 context_tracking_task_switch(prev, next); 2367 /* Here we just switch the register state and the stack. */ 2368 switch_to(prev, next, prev); 2369 2370 barrier(); 2371 /* 2372 * this_rq must be evaluated again because prev may have moved 2373 * CPUs since it called schedule(), thus the 'rq' on its stack 2374 * frame will be invalid. 2375 */ 2376 finish_task_switch(this_rq(), prev); 2377} 2378 2379/* 2380 * nr_running and nr_context_switches: 2381 * 2382 * externally visible scheduler statistics: current number of runnable 2383 * threads, total number of context switches performed since bootup. 2384 */ 2385unsigned long nr_running(void) 2386{ 2387 unsigned long i, sum = 0; 2388 2389 for_each_online_cpu(i) 2390 sum += cpu_rq(i)->nr_running; 2391 2392 return sum; 2393} 2394 2395/* 2396 * Check if only the current task is running on the cpu. 2397 */ 2398bool single_task_running(void) 2399{ 2400 if (cpu_rq(smp_processor_id())->nr_running == 1) 2401 return true; 2402 else 2403 return false; 2404} 2405EXPORT_SYMBOL(single_task_running); 2406 2407unsigned long long nr_context_switches(void) 2408{ 2409 int i; 2410 unsigned long long sum = 0; 2411 2412 for_each_possible_cpu(i) 2413 sum += cpu_rq(i)->nr_switches; 2414 2415 return sum; 2416} 2417 2418unsigned long nr_iowait(void) 2419{ 2420 unsigned long i, sum = 0; 2421 2422 for_each_possible_cpu(i) 2423 sum += atomic_read(&cpu_rq(i)->nr_iowait); 2424 2425 return sum; 2426} 2427 2428unsigned long nr_iowait_cpu(int cpu) 2429{ 2430 struct rq *this = cpu_rq(cpu); 2431 return atomic_read(&this->nr_iowait); 2432} 2433 2434void get_iowait_load(unsigned long *nr_waiters, unsigned long *load) 2435{ 2436 struct rq *this = this_rq(); 2437 *nr_waiters = atomic_read(&this->nr_iowait); 2438 *load = this->cpu_load[0]; 2439} 2440 2441#ifdef CONFIG_SMP 2442 2443/* 2444 * sched_exec - execve() is a valuable balancing opportunity, because at 2445 * this point the task has the smallest effective memory and cache footprint. 2446 */ 2447void sched_exec(void) 2448{ 2449 struct task_struct *p = current; 2450 unsigned long flags; 2451 int dest_cpu; 2452 2453 raw_spin_lock_irqsave(&p->pi_lock, flags); 2454 dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), SD_BALANCE_EXEC, 0); 2455 if (dest_cpu == smp_processor_id()) 2456 goto unlock; 2457 2458 if (likely(cpu_active(dest_cpu))) { 2459 struct migration_arg arg = { p, dest_cpu }; 2460 2461 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 2462 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg); 2463 return; 2464 } 2465unlock: 2466 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 2467} 2468 2469#endif 2470 2471DEFINE_PER_CPU(struct kernel_stat, kstat); 2472DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat); 2473 2474EXPORT_PER_CPU_SYMBOL(kstat); 2475EXPORT_PER_CPU_SYMBOL(kernel_cpustat); 2476 2477/* 2478 * Return any ns on the sched_clock that have not yet been accounted in 2479 * @p in case that task is currently running. 2480 * 2481 * Called with task_rq_lock() held on @rq. 2482 */ 2483static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq) 2484{ 2485 u64 ns = 0; 2486 2487 /* 2488 * Must be ->curr _and_ ->on_rq. If dequeued, we would 2489 * project cycles that may never be accounted to this 2490 * thread, breaking clock_gettime(). 2491 */ 2492 if (task_current(rq, p) && task_on_rq_queued(p)) { 2493 update_rq_clock(rq); 2494 ns = rq_clock_task(rq) - p->se.exec_start; 2495 if ((s64)ns < 0) 2496 ns = 0; 2497 } 2498 2499 return ns; 2500} 2501 2502unsigned long long task_delta_exec(struct task_struct *p) 2503{ 2504 unsigned long flags; 2505 struct rq *rq; 2506 u64 ns = 0; 2507 2508 rq = task_rq_lock(p, &flags); 2509 ns = do_task_delta_exec(p, rq); 2510 task_rq_unlock(rq, p, &flags); 2511 2512 return ns; 2513} 2514 2515/* 2516 * Return accounted runtime for the task. 2517 * In case the task is currently running, return the runtime plus current's 2518 * pending runtime that have not been accounted yet. 2519 */ 2520unsigned long long task_sched_runtime(struct task_struct *p) 2521{ 2522 unsigned long flags; 2523 struct rq *rq; 2524 u64 ns = 0; 2525 2526#if defined(CONFIG_64BIT) && defined(CONFIG_SMP) 2527 /* 2528 * 64-bit doesn't need locks to atomically read a 64bit value. 2529 * So we have a optimization chance when the task's delta_exec is 0. 2530 * Reading ->on_cpu is racy, but this is ok. 2531 * 2532 * If we race with it leaving cpu, we'll take a lock. So we're correct. 2533 * If we race with it entering cpu, unaccounted time is 0. This is 2534 * indistinguishable from the read occurring a few cycles earlier. 2535 * If we see ->on_cpu without ->on_rq, the task is leaving, and has 2536 * been accounted, so we're correct here as well. 2537 */ 2538 if (!p->on_cpu || !task_on_rq_queued(p)) 2539 return p->se.sum_exec_runtime; 2540#endif 2541 2542 rq = task_rq_lock(p, &flags); 2543 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq); 2544 task_rq_unlock(rq, p, &flags); 2545 2546 return ns; 2547} 2548 2549/* 2550 * This function gets called by the timer code, with HZ frequency. 2551 * We call it with interrupts disabled. 2552 */ 2553void scheduler_tick(void) 2554{ 2555 int cpu = smp_processor_id(); 2556 struct rq *rq = cpu_rq(cpu); 2557 struct task_struct *curr = rq->curr; 2558 2559 sched_clock_tick(); 2560 2561 raw_spin_lock(&rq->lock); 2562 update_rq_clock(rq); 2563 curr->sched_class->task_tick(rq, curr, 0); 2564 update_cpu_load_active(rq); 2565 raw_spin_unlock(&rq->lock); 2566 2567 perf_event_task_tick(); 2568 2569#ifdef CONFIG_SMP 2570 rq->idle_balance = idle_cpu(cpu); 2571 trigger_load_balance(rq); 2572#endif 2573 rq_last_tick_reset(rq); 2574} 2575 2576#ifdef CONFIG_NO_HZ_FULL 2577/** 2578 * scheduler_tick_max_deferment 2579 * 2580 * Keep at least one tick per second when a single 2581 * active task is running because the scheduler doesn't 2582 * yet completely support full dynticks environment. 2583 * 2584 * This makes sure that uptime, CFS vruntime, load 2585 * balancing, etc... continue to move forward, even 2586 * with a very low granularity. 2587 * 2588 * Return: Maximum deferment in nanoseconds. 2589 */ 2590u64 scheduler_tick_max_deferment(void) 2591{ 2592 struct rq *rq = this_rq(); 2593 unsigned long next, now = ACCESS_ONCE(jiffies); 2594 2595 next = rq->last_sched_tick + HZ; 2596 2597 if (time_before_eq(next, now)) 2598 return 0; 2599 2600 return jiffies_to_nsecs(next - now); 2601} 2602#endif 2603 2604notrace unsigned long get_parent_ip(unsigned long addr) 2605{ 2606 if (in_lock_functions(addr)) { 2607 addr = CALLER_ADDR2; 2608 if (in_lock_functions(addr)) 2609 addr = CALLER_ADDR3; 2610 } 2611 return addr; 2612} 2613 2614#if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \ 2615 defined(CONFIG_PREEMPT_TRACER)) 2616 2617void preempt_count_add(int val) 2618{ 2619#ifdef CONFIG_DEBUG_PREEMPT 2620 /* 2621 * Underflow? 2622 */ 2623 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0))) 2624 return; 2625#endif 2626 __preempt_count_add(val); 2627#ifdef CONFIG_DEBUG_PREEMPT 2628 /* 2629 * Spinlock count overflowing soon? 2630 */ 2631 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >= 2632 PREEMPT_MASK - 10); 2633#endif 2634 if (preempt_count() == val) { 2635 unsigned long ip = get_parent_ip(CALLER_ADDR1); 2636#ifdef CONFIG_DEBUG_PREEMPT 2637 current->preempt_disable_ip = ip; 2638#endif 2639 trace_preempt_off(CALLER_ADDR0, ip); 2640 } 2641} 2642EXPORT_SYMBOL(preempt_count_add); 2643NOKPROBE_SYMBOL(preempt_count_add); 2644 2645void preempt_count_sub(int val) 2646{ 2647#ifdef CONFIG_DEBUG_PREEMPT 2648 /* 2649 * Underflow? 2650 */ 2651 if (DEBUG_LOCKS_WARN_ON(val > preempt_count())) 2652 return; 2653 /* 2654 * Is the spinlock portion underflowing? 2655 */ 2656 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) && 2657 !(preempt_count() & PREEMPT_MASK))) 2658 return; 2659#endif 2660 2661 if (preempt_count() == val) 2662 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1)); 2663 __preempt_count_sub(val); 2664} 2665EXPORT_SYMBOL(preempt_count_sub); 2666NOKPROBE_SYMBOL(preempt_count_sub); 2667 2668#endif 2669 2670/* 2671 * Print scheduling while atomic bug: 2672 */ 2673static noinline void __schedule_bug(struct task_struct *prev) 2674{ 2675 if (oops_in_progress) 2676 return; 2677 2678 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n", 2679 prev->comm, prev->pid, preempt_count()); 2680 2681 debug_show_held_locks(prev); 2682 print_modules(); 2683 if (irqs_disabled()) 2684 print_irqtrace_events(prev); 2685#ifdef CONFIG_DEBUG_PREEMPT 2686 if (in_atomic_preempt_off()) { 2687 pr_err("Preemption disabled at:"); 2688 print_ip_sym(current->preempt_disable_ip); 2689 pr_cont("\n"); 2690 } 2691#endif 2692 dump_stack(); 2693 add_taint(TAINT_WARN, LOCKDEP_STILL_OK); 2694} 2695 2696/* 2697 * Various schedule()-time debugging checks and statistics: 2698 */ 2699static inline void schedule_debug(struct task_struct *prev) 2700{ 2701#ifdef CONFIG_SCHED_STACK_END_CHECK 2702 BUG_ON(unlikely(task_stack_end_corrupted(prev))); 2703#endif 2704 /* 2705 * Test if we are atomic. Since do_exit() needs to call into 2706 * schedule() atomically, we ignore that path. Otherwise whine 2707 * if we are scheduling when we should not. 2708 */ 2709 if (unlikely(in_atomic_preempt_off() && prev->state != TASK_DEAD)) 2710 __schedule_bug(prev); 2711 rcu_sleep_check(); 2712 2713 profile_hit(SCHED_PROFILING, __builtin_return_address(0)); 2714 2715 schedstat_inc(this_rq(), sched_count); 2716} 2717 2718/* 2719 * Pick up the highest-prio task: 2720 */ 2721static inline struct task_struct * 2722pick_next_task(struct rq *rq, struct task_struct *prev) 2723{ 2724 const struct sched_class *class = &fair_sched_class; 2725 struct task_struct *p; 2726 2727 /* 2728 * Optimization: we know that if all tasks are in 2729 * the fair class we can call that function directly: 2730 */ 2731 if (likely(prev->sched_class == class && 2732 rq->nr_running == rq->cfs.h_nr_running)) { 2733 p = fair_sched_class.pick_next_task(rq, prev); 2734 if (unlikely(p == RETRY_TASK)) 2735 goto again; 2736 2737 /* assumes fair_sched_class->next == idle_sched_class */ 2738 if (unlikely(!p)) 2739 p = idle_sched_class.pick_next_task(rq, prev); 2740 2741 return p; 2742 } 2743 2744again: 2745 for_each_class(class) { 2746 p = class->pick_next_task(rq, prev); 2747 if (p) { 2748 if (unlikely(p == RETRY_TASK)) 2749 goto again; 2750 return p; 2751 } 2752 } 2753 2754 BUG(); /* the idle class will always have a runnable task */ 2755} 2756 2757/* 2758 * __schedule() is the main scheduler function. 2759 * 2760 * The main means of driving the scheduler and thus entering this function are: 2761 * 2762 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc. 2763 * 2764 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return 2765 * paths. For example, see arch/x86/entry_64.S. 2766 * 2767 * To drive preemption between tasks, the scheduler sets the flag in timer 2768 * interrupt handler scheduler_tick(). 2769 * 2770 * 3. Wakeups don't really cause entry into schedule(). They add a 2771 * task to the run-queue and that's it. 2772 * 2773 * Now, if the new task added to the run-queue preempts the current 2774 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets 2775 * called on the nearest possible occasion: 2776 * 2777 * - If the kernel is preemptible (CONFIG_PREEMPT=y): 2778 * 2779 * - in syscall or exception context, at the next outmost 2780 * preempt_enable(). (this might be as soon as the wake_up()'s 2781 * spin_unlock()!) 2782 * 2783 * - in IRQ context, return from interrupt-handler to 2784 * preemptible context 2785 * 2786 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set) 2787 * then at the next: 2788 * 2789 * - cond_resched() call 2790 * - explicit schedule() call 2791 * - return from syscall or exception to user-space 2792 * - return from interrupt-handler to user-space 2793 */ 2794static void __sched __schedule(void) 2795{ 2796 struct task_struct *prev, *next; 2797 unsigned long *switch_count; 2798 struct rq *rq; 2799 int cpu; 2800 2801need_resched: 2802 preempt_disable(); 2803 cpu = smp_processor_id(); 2804 rq = cpu_rq(cpu); 2805 rcu_note_context_switch(cpu); 2806 prev = rq->curr; 2807 2808 schedule_debug(prev); 2809 2810 if (sched_feat(HRTICK)) 2811 hrtick_clear(rq); 2812 2813 /* 2814 * Make sure that signal_pending_state()->signal_pending() below 2815 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE) 2816 * done by the caller to avoid the race with signal_wake_up(). 2817 */ 2818 smp_mb__before_spinlock(); 2819 raw_spin_lock_irq(&rq->lock); 2820 2821 switch_count = &prev->nivcsw; 2822 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) { 2823 if (unlikely(signal_pending_state(prev->state, prev))) { 2824 prev->state = TASK_RUNNING; 2825 } else { 2826 deactivate_task(rq, prev, DEQUEUE_SLEEP); 2827 prev->on_rq = 0; 2828 2829 /* 2830 * If a worker went to sleep, notify and ask workqueue 2831 * whether it wants to wake up a task to maintain 2832 * concurrency. 2833 */ 2834 if (prev->flags & PF_WQ_WORKER) { 2835 struct task_struct *to_wakeup; 2836 2837 to_wakeup = wq_worker_sleeping(prev, cpu); 2838 if (to_wakeup) 2839 try_to_wake_up_local(to_wakeup); 2840 } 2841 } 2842 switch_count = &prev->nvcsw; 2843 } 2844 2845 if (task_on_rq_queued(prev) || rq->skip_clock_update < 0) 2846 update_rq_clock(rq); 2847 2848 next = pick_next_task(rq, prev); 2849 clear_tsk_need_resched(prev); 2850 clear_preempt_need_resched(); 2851 rq->skip_clock_update = 0; 2852 2853 if (likely(prev != next)) { 2854 rq->nr_switches++; 2855 rq->curr = next; 2856 ++*switch_count; 2857 2858 context_switch(rq, prev, next); /* unlocks the rq */ 2859 /* 2860 * The context switch have flipped the stack from under us 2861 * and restored the local variables which were saved when 2862 * this task called schedule() in the past. prev == current 2863 * is still correct, but it can be moved to another cpu/rq. 2864 */ 2865 cpu = smp_processor_id(); 2866 rq = cpu_rq(cpu); 2867 } else 2868 raw_spin_unlock_irq(&rq->lock); 2869 2870 post_schedule(rq); 2871 2872 sched_preempt_enable_no_resched(); 2873 if (need_resched()) 2874 goto need_resched; 2875} 2876 2877static inline void sched_submit_work(struct task_struct *tsk) 2878{ 2879 if (!tsk->state || tsk_is_pi_blocked(tsk)) 2880 return; 2881 /* 2882 * If we are going to sleep and we have plugged IO queued, 2883 * make sure to submit it to avoid deadlocks. 2884 */ 2885 if (blk_needs_flush_plug(tsk)) 2886 blk_schedule_flush_plug(tsk); 2887} 2888 2889asmlinkage __visible void __sched schedule(void) 2890{ 2891 struct task_struct *tsk = current; 2892 2893 sched_submit_work(tsk); 2894 __schedule(); 2895} 2896EXPORT_SYMBOL(schedule); 2897 2898#ifdef CONFIG_CONTEXT_TRACKING 2899asmlinkage __visible void __sched schedule_user(void) 2900{ 2901 /* 2902 * If we come here after a random call to set_need_resched(), 2903 * or we have been woken up remotely but the IPI has not yet arrived, 2904 * we haven't yet exited the RCU idle mode. Do it here manually until 2905 * we find a better solution. 2906 */ 2907 user_exit(); 2908 schedule(); 2909 user_enter(); 2910} 2911#endif 2912 2913/** 2914 * schedule_preempt_disabled - called with preemption disabled 2915 * 2916 * Returns with preemption disabled. Note: preempt_count must be 1 2917 */ 2918void __sched schedule_preempt_disabled(void) 2919{ 2920 sched_preempt_enable_no_resched(); 2921 schedule(); 2922 preempt_disable(); 2923} 2924 2925#ifdef CONFIG_PREEMPT 2926/* 2927 * this is the entry point to schedule() from in-kernel preemption 2928 * off of preempt_enable. Kernel preemptions off return from interrupt 2929 * occur there and call schedule directly. 2930 */ 2931asmlinkage __visible void __sched notrace preempt_schedule(void) 2932{ 2933 /* 2934 * If there is a non-zero preempt_count or interrupts are disabled, 2935 * we do not want to preempt the current task. Just return.. 2936 */ 2937 if (likely(!preemptible())) 2938 return; 2939 2940 do { 2941 __preempt_count_add(PREEMPT_ACTIVE); 2942 __schedule(); 2943 __preempt_count_sub(PREEMPT_ACTIVE); 2944 2945 /* 2946 * Check again in case we missed a preemption opportunity 2947 * between schedule and now. 2948 */ 2949 barrier(); 2950 } while (need_resched()); 2951} 2952NOKPROBE_SYMBOL(preempt_schedule); 2953EXPORT_SYMBOL(preempt_schedule); 2954 2955#ifdef CONFIG_CONTEXT_TRACKING 2956/** 2957 * preempt_schedule_context - preempt_schedule called by tracing 2958 * 2959 * The tracing infrastructure uses preempt_enable_notrace to prevent 2960 * recursion and tracing preempt enabling caused by the tracing 2961 * infrastructure itself. But as tracing can happen in areas coming 2962 * from userspace or just about to enter userspace, a preempt enable 2963 * can occur before user_exit() is called. This will cause the scheduler 2964 * to be called when the system is still in usermode. 2965 * 2966 * To prevent this, the preempt_enable_notrace will use this function 2967 * instead of preempt_schedule() to exit user context if needed before 2968 * calling the scheduler. 2969 */ 2970asmlinkage __visible void __sched notrace preempt_schedule_context(void) 2971{ 2972 enum ctx_state prev_ctx; 2973 2974 if (likely(!preemptible())) 2975 return; 2976 2977 do { 2978 __preempt_count_add(PREEMPT_ACTIVE); 2979 /* 2980 * Needs preempt disabled in case user_exit() is traced 2981 * and the tracer calls preempt_enable_notrace() causing 2982 * an infinite recursion. 2983 */ 2984 prev_ctx = exception_enter(); 2985 __schedule(); 2986 exception_exit(prev_ctx); 2987 2988 __preempt_count_sub(PREEMPT_ACTIVE); 2989 barrier(); 2990 } while (need_resched()); 2991} 2992EXPORT_SYMBOL_GPL(preempt_schedule_context); 2993#endif /* CONFIG_CONTEXT_TRACKING */ 2994 2995#endif /* CONFIG_PREEMPT */ 2996 2997/* 2998 * this is the entry point to schedule() from kernel preemption 2999 * off of irq context. 3000 * Note, that this is called and return with irqs disabled. This will 3001 * protect us against recursive calling from irq. 3002 */ 3003asmlinkage __visible void __sched preempt_schedule_irq(void) 3004{ 3005 enum ctx_state prev_state; 3006 3007 /* Catch callers which need to be fixed */ 3008 BUG_ON(preempt_count() || !irqs_disabled()); 3009 3010 prev_state = exception_enter(); 3011 3012 do { 3013 __preempt_count_add(PREEMPT_ACTIVE); 3014 local_irq_enable(); 3015 __schedule(); 3016 local_irq_disable(); 3017 __preempt_count_sub(PREEMPT_ACTIVE); 3018 3019 /* 3020 * Check again in case we missed a preemption opportunity 3021 * between schedule and now. 3022 */ 3023 barrier(); 3024 } while (need_resched()); 3025 3026 exception_exit(prev_state); 3027} 3028 3029int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags, 3030 void *key) 3031{ 3032 return try_to_wake_up(curr->private, mode, wake_flags); 3033} 3034EXPORT_SYMBOL(default_wake_function); 3035 3036#ifdef CONFIG_RT_MUTEXES 3037 3038/* 3039 * rt_mutex_setprio - set the current priority of a task 3040 * @p: task 3041 * @prio: prio value (kernel-internal form) 3042 * 3043 * This function changes the 'effective' priority of a task. It does 3044 * not touch ->normal_prio like __setscheduler(). 3045 * 3046 * Used by the rt_mutex code to implement priority inheritance 3047 * logic. Call site only calls if the priority of the task changed. 3048 */ 3049void rt_mutex_setprio(struct task_struct *p, int prio) 3050{ 3051 int oldprio, queued, running, enqueue_flag = 0; 3052 struct rq *rq; 3053 const struct sched_class *prev_class; 3054 3055 BUG_ON(prio > MAX_PRIO); 3056 3057 rq = __task_rq_lock(p); 3058 3059 /* 3060 * Idle task boosting is a nono in general. There is one 3061 * exception, when PREEMPT_RT and NOHZ is active: 3062 * 3063 * The idle task calls get_next_timer_interrupt() and holds 3064 * the timer wheel base->lock on the CPU and another CPU wants 3065 * to access the timer (probably to cancel it). We can safely 3066 * ignore the boosting request, as the idle CPU runs this code 3067 * with interrupts disabled and will complete the lock 3068 * protected section without being interrupted. So there is no 3069 * real need to boost. 3070 */ 3071 if (unlikely(p == rq->idle)) { 3072 WARN_ON(p != rq->curr); 3073 WARN_ON(p->pi_blocked_on); 3074 goto out_unlock; 3075 } 3076 3077 trace_sched_pi_setprio(p, prio); 3078 oldprio = p->prio; 3079 prev_class = p->sched_class; 3080 queued = task_on_rq_queued(p); 3081 running = task_current(rq, p); 3082 if (queued) 3083 dequeue_task(rq, p, 0); 3084 if (running) 3085 put_prev_task(rq, p); 3086 3087 /* 3088 * Boosting condition are: 3089 * 1. -rt task is running and holds mutex A 3090 * --> -dl task blocks on mutex A 3091 * 3092 * 2. -dl task is running and holds mutex A 3093 * --> -dl task blocks on mutex A and could preempt the 3094 * running task 3095 */ 3096 if (dl_prio(prio)) { 3097 struct task_struct *pi_task = rt_mutex_get_top_task(p); 3098 if (!dl_prio(p->normal_prio) || 3099 (pi_task && dl_entity_preempt(&pi_task->dl, &p->dl))) { 3100 p->dl.dl_boosted = 1; 3101 p->dl.dl_throttled = 0; 3102 enqueue_flag = ENQUEUE_REPLENISH; 3103 } else 3104 p->dl.dl_boosted = 0; 3105 p->sched_class = &dl_sched_class; 3106 } else if (rt_prio(prio)) { 3107 if (dl_prio(oldprio)) 3108 p->dl.dl_boosted = 0; 3109 if (oldprio < prio) 3110 enqueue_flag = ENQUEUE_HEAD; 3111 p->sched_class = &rt_sched_class; 3112 } else { 3113 if (dl_prio(oldprio)) 3114 p->dl.dl_boosted = 0; 3115 p->sched_class = &fair_sched_class; 3116 } 3117 3118 p->prio = prio; 3119 3120 if (running) 3121 p->sched_class->set_curr_task(rq); 3122 if (queued) 3123 enqueue_task(rq, p, enqueue_flag); 3124 3125 check_class_changed(rq, p, prev_class, oldprio); 3126out_unlock: 3127 __task_rq_unlock(rq); 3128} 3129#endif 3130 3131void set_user_nice(struct task_struct *p, long nice) 3132{ 3133 int old_prio, delta, queued; 3134 unsigned long flags; 3135 struct rq *rq; 3136 3137 if (task_nice(p) == nice || nice < MIN_NICE || nice > MAX_NICE) 3138 return; 3139 /* 3140 * We have to be careful, if called from sys_setpriority(), 3141 * the task might be in the middle of scheduling on another CPU. 3142 */ 3143 rq = task_rq_lock(p, &flags); 3144 /* 3145 * The RT priorities are set via sched_setscheduler(), but we still 3146 * allow the 'normal' nice value to be set - but as expected 3147 * it wont have any effect on scheduling until the task is 3148 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR: 3149 */ 3150 if (task_has_dl_policy(p) || task_has_rt_policy(p)) { 3151 p->static_prio = NICE_TO_PRIO(nice); 3152 goto out_unlock; 3153 } 3154 queued = task_on_rq_queued(p); 3155 if (queued) 3156 dequeue_task(rq, p, 0); 3157 3158 p->static_prio = NICE_TO_PRIO(nice); 3159 set_load_weight(p); 3160 old_prio = p->prio; 3161 p->prio = effective_prio(p); 3162 delta = p->prio - old_prio; 3163 3164 if (queued) { 3165 enqueue_task(rq, p, 0); 3166 /* 3167 * If the task increased its priority or is running and 3168 * lowered its priority, then reschedule its CPU: 3169 */ 3170 if (delta < 0 || (delta > 0 && task_running(rq, p))) 3171 resched_curr(rq); 3172 } 3173out_unlock: 3174 task_rq_unlock(rq, p, &flags); 3175} 3176EXPORT_SYMBOL(set_user_nice); 3177 3178/* 3179 * can_nice - check if a task can reduce its nice value 3180 * @p: task 3181 * @nice: nice value 3182 */ 3183int can_nice(const struct task_struct *p, const int nice) 3184{ 3185 /* convert nice value [19,-20] to rlimit style value [1,40] */ 3186 int nice_rlim = nice_to_rlimit(nice); 3187 3188 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) || 3189 capable(CAP_SYS_NICE)); 3190} 3191 3192#ifdef __ARCH_WANT_SYS_NICE 3193 3194/* 3195 * sys_nice - change the priority of the current process. 3196 * @increment: priority increment 3197 * 3198 * sys_setpriority is a more generic, but much slower function that 3199 * does similar things. 3200 */ 3201SYSCALL_DEFINE1(nice, int, increment) 3202{ 3203 long nice, retval; 3204 3205 /* 3206 * Setpriority might change our priority at the same moment. 3207 * We don't have to worry. Conceptually one call occurs first 3208 * and we have a single winner. 3209 */ 3210 increment = clamp(increment, -NICE_WIDTH, NICE_WIDTH); 3211 nice = task_nice(current) + increment; 3212 3213 nice = clamp_val(nice, MIN_NICE, MAX_NICE); 3214 if (increment < 0 && !can_nice(current, nice)) 3215 return -EPERM; 3216 3217 retval = security_task_setnice(current, nice); 3218 if (retval) 3219 return retval; 3220 3221 set_user_nice(current, nice); 3222 return 0; 3223} 3224 3225#endif 3226 3227/** 3228 * task_prio - return the priority value of a given task. 3229 * @p: the task in question. 3230 * 3231 * Return: The priority value as seen by users in /proc. 3232 * RT tasks are offset by -200. Normal tasks are centered 3233 * around 0, value goes from -16 to +15. 3234 */ 3235int task_prio(const struct task_struct *p) 3236{ 3237 return p->prio - MAX_RT_PRIO; 3238} 3239 3240/** 3241 * idle_cpu - is a given cpu idle currently? 3242 * @cpu: the processor in question. 3243 * 3244 * Return: 1 if the CPU is currently idle. 0 otherwise. 3245 */ 3246int idle_cpu(int cpu) 3247{ 3248 struct rq *rq = cpu_rq(cpu); 3249 3250 if (rq->curr != rq->idle) 3251 return 0; 3252 3253 if (rq->nr_running) 3254 return 0; 3255 3256#ifdef CONFIG_SMP 3257 if (!llist_empty(&rq->wake_list)) 3258 return 0; 3259#endif 3260 3261 return 1; 3262} 3263 3264/** 3265 * idle_task - return the idle task for a given cpu. 3266 * @cpu: the processor in question. 3267 * 3268 * Return: The idle task for the cpu @cpu. 3269 */ 3270struct task_struct *idle_task(int cpu) 3271{ 3272 return cpu_rq(cpu)->idle; 3273} 3274 3275/** 3276 * find_process_by_pid - find a process with a matching PID value. 3277 * @pid: the pid in question. 3278 * 3279 * The task of @pid, if found. %NULL otherwise. 3280 */ 3281static struct task_struct *find_process_by_pid(pid_t pid) 3282{ 3283 return pid ? find_task_by_vpid(pid) : current; 3284} 3285 3286/* 3287 * This function initializes the sched_dl_entity of a newly becoming 3288 * SCHED_DEADLINE task. 3289 * 3290 * Only the static values are considered here, the actual runtime and the 3291 * absolute deadline will be properly calculated when the task is enqueued 3292 * for the first time with its new policy. 3293 */ 3294static void 3295__setparam_dl(struct task_struct *p, const struct sched_attr *attr) 3296{ 3297 struct sched_dl_entity *dl_se = &p->dl; 3298 3299 init_dl_task_timer(dl_se); 3300 dl_se->dl_runtime = attr->sched_runtime; 3301 dl_se->dl_deadline = attr->sched_deadline; 3302 dl_se->dl_period = attr->sched_period ?: dl_se->dl_deadline; 3303 dl_se->flags = attr->sched_flags; 3304 dl_se->dl_bw = to_ratio(dl_se->dl_period, dl_se->dl_runtime); 3305 dl_se->dl_throttled = 0; 3306 dl_se->dl_new = 1; 3307 dl_se->dl_yielded = 0; 3308} 3309 3310/* 3311 * sched_setparam() passes in -1 for its policy, to let the functions 3312 * it calls know not to change it. 3313 */ 3314#define SETPARAM_POLICY -1 3315 3316static void __setscheduler_params(struct task_struct *p, 3317 const struct sched_attr *attr) 3318{ 3319 int policy = attr->sched_policy; 3320 3321 if (policy == SETPARAM_POLICY) 3322 policy = p->policy; 3323 3324 p->policy = policy; 3325 3326 if (dl_policy(policy)) 3327 __setparam_dl(p, attr); 3328 else if (fair_policy(policy)) 3329 p->static_prio = NICE_TO_PRIO(attr->sched_nice); 3330 3331 /* 3332 * __sched_setscheduler() ensures attr->sched_priority == 0 when 3333 * !rt_policy. Always setting this ensures that things like 3334 * getparam()/getattr() don't report silly values for !rt tasks. 3335 */ 3336 p->rt_priority = attr->sched_priority; 3337 p->normal_prio = normal_prio(p); 3338 set_load_weight(p); 3339} 3340 3341/* Actually do priority change: must hold pi & rq lock. */ 3342static void __setscheduler(struct rq *rq, struct task_struct *p, 3343 const struct sched_attr *attr) 3344{ 3345 __setscheduler_params(p, attr); 3346 3347 /* 3348 * If we get here, there was no pi waiters boosting the 3349 * task. It is safe to use the normal prio. 3350 */ 3351 p->prio = normal_prio(p); 3352 3353 if (dl_prio(p->prio)) 3354 p->sched_class = &dl_sched_class; 3355 else if (rt_prio(p->prio)) 3356 p->sched_class = &rt_sched_class; 3357 else 3358 p->sched_class = &fair_sched_class; 3359} 3360 3361static void 3362__getparam_dl(struct task_struct *p, struct sched_attr *attr) 3363{ 3364 struct sched_dl_entity *dl_se = &p->dl; 3365 3366 attr->sched_priority = p->rt_priority; 3367 attr->sched_runtime = dl_se->dl_runtime; 3368 attr->sched_deadline = dl_se->dl_deadline; 3369 attr->sched_period = dl_se->dl_period; 3370 attr->sched_flags = dl_se->flags; 3371} 3372 3373/* 3374 * This function validates the new parameters of a -deadline task. 3375 * We ask for the deadline not being zero, and greater or equal 3376 * than the runtime, as well as the period of being zero or 3377 * greater than deadline. Furthermore, we have to be sure that 3378 * user parameters are above the internal resolution of 1us (we 3379 * check sched_runtime only since it is always the smaller one) and 3380 * below 2^63 ns (we have to check both sched_deadline and 3381 * sched_period, as the latter can be zero). 3382 */ 3383static bool 3384__checkparam_dl(const struct sched_attr *attr) 3385{ 3386 /* deadline != 0 */ 3387 if (attr->sched_deadline == 0) 3388 return false; 3389 3390 /* 3391 * Since we truncate DL_SCALE bits, make sure we're at least 3392 * that big. 3393 */ 3394 if (attr->sched_runtime < (1ULL << DL_SCALE)) 3395 return false; 3396 3397 /* 3398 * Since we use the MSB for wrap-around and sign issues, make 3399 * sure it's not set (mind that period can be equal to zero). 3400 */ 3401 if (attr->sched_deadline & (1ULL << 63) || 3402 attr->sched_period & (1ULL << 63)) 3403 return false; 3404 3405 /* runtime <= deadline <= period (if period != 0) */ 3406 if ((attr->sched_period != 0 && 3407 attr->sched_period < attr->sched_deadline) || 3408 attr->sched_deadline < attr->sched_runtime) 3409 return false; 3410 3411 return true; 3412} 3413 3414/* 3415 * check the target process has a UID that matches the current process's 3416 */ 3417static bool check_same_owner(struct task_struct *p) 3418{ 3419 const struct cred *cred = current_cred(), *pcred; 3420 bool match; 3421 3422 rcu_read_lock(); 3423 pcred = __task_cred(p); 3424 match = (uid_eq(cred->euid, pcred->euid) || 3425 uid_eq(cred->euid, pcred->uid)); 3426 rcu_read_unlock(); 3427 return match; 3428} 3429 3430static int __sched_setscheduler(struct task_struct *p, 3431 const struct sched_attr *attr, 3432 bool user) 3433{ 3434 int newprio = dl_policy(attr->sched_policy) ? MAX_DL_PRIO - 1 : 3435 MAX_RT_PRIO - 1 - attr->sched_priority; 3436 int retval, oldprio, oldpolicy = -1, queued, running; 3437 int policy = attr->sched_policy; 3438 unsigned long flags; 3439 const struct sched_class *prev_class; 3440 struct rq *rq; 3441 int reset_on_fork; 3442 3443 /* may grab non-irq protected spin_locks */ 3444 BUG_ON(in_interrupt()); 3445recheck: 3446 /* double check policy once rq lock held */ 3447 if (policy < 0) { 3448 reset_on_fork = p->sched_reset_on_fork; 3449 policy = oldpolicy = p->policy; 3450 } else { 3451 reset_on_fork = !!(attr->sched_flags & SCHED_FLAG_RESET_ON_FORK); 3452 3453 if (policy != SCHED_DEADLINE && 3454 policy != SCHED_FIFO && policy != SCHED_RR && 3455 policy != SCHED_NORMAL && policy != SCHED_BATCH && 3456 policy != SCHED_IDLE) 3457 return -EINVAL; 3458 } 3459 3460 if (attr->sched_flags & ~(SCHED_FLAG_RESET_ON_FORK)) 3461 return -EINVAL; 3462 3463 /* 3464 * Valid priorities for SCHED_FIFO and SCHED_RR are 3465 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL, 3466 * SCHED_BATCH and SCHED_IDLE is 0. 3467 */ 3468 if ((p->mm && attr->sched_priority > MAX_USER_RT_PRIO-1) || 3469 (!p->mm && attr->sched_priority > MAX_RT_PRIO-1)) 3470 return -EINVAL; 3471 if ((dl_policy(policy) && !__checkparam_dl(attr)) || 3472 (rt_policy(policy) != (attr->sched_priority != 0))) 3473 return -EINVAL; 3474 3475 /* 3476 * Allow unprivileged RT tasks to decrease priority: 3477 */ 3478 if (user && !capable(CAP_SYS_NICE)) { 3479 if (fair_policy(policy)) { 3480 if (attr->sched_nice < task_nice(p) && 3481 !can_nice(p, attr->sched_nice)) 3482 return -EPERM; 3483 } 3484 3485 if (rt_policy(policy)) { 3486 unsigned long rlim_rtprio = 3487 task_rlimit(p, RLIMIT_RTPRIO); 3488 3489 /* can't set/change the rt policy */ 3490 if (policy != p->policy && !rlim_rtprio) 3491 return -EPERM; 3492 3493 /* can't increase priority */ 3494 if (attr->sched_priority > p->rt_priority && 3495 attr->sched_priority > rlim_rtprio) 3496 return -EPERM; 3497 } 3498 3499 /* 3500 * Can't set/change SCHED_DEADLINE policy at all for now 3501 * (safest behavior); in the future we would like to allow 3502 * unprivileged DL tasks to increase their relative deadline 3503 * or reduce their runtime (both ways reducing utilization) 3504 */ 3505 if (dl_policy(policy)) 3506 return -EPERM; 3507 3508 /* 3509 * Treat SCHED_IDLE as nice 20. Only allow a switch to 3510 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it. 3511 */ 3512 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) { 3513 if (!can_nice(p, task_nice(p))) 3514 return -EPERM; 3515 } 3516 3517 /* can't change other user's priorities */ 3518 if (!check_same_owner(p)) 3519 return -EPERM; 3520 3521 /* Normal users shall not reset the sched_reset_on_fork flag */ 3522 if (p->sched_reset_on_fork && !reset_on_fork) 3523 return -EPERM; 3524 } 3525 3526 if (user) { 3527 retval = security_task_setscheduler(p); 3528 if (retval) 3529 return retval; 3530 } 3531 3532 /* 3533 * make sure no PI-waiters arrive (or leave) while we are 3534 * changing the priority of the task: 3535 * 3536 * To be able to change p->policy safely, the appropriate 3537 * runqueue lock must be held. 3538 */ 3539 rq = task_rq_lock(p, &flags); 3540 3541 /* 3542 * Changing the policy of the stop threads its a very bad idea 3543 */ 3544 if (p == rq->stop) { 3545 task_rq_unlock(rq, p, &flags); 3546 return -EINVAL; 3547 } 3548 3549 /* 3550 * If not changing anything there's no need to proceed further, 3551 * but store a possible modification of reset_on_fork. 3552 */ 3553 if (unlikely(policy == p->policy)) { 3554 if (fair_policy(policy) && attr->sched_nice != task_nice(p)) 3555 goto change; 3556 if (rt_policy(policy) && attr->sched_priority != p->rt_priority) 3557 goto change; 3558 if (dl_policy(policy)) 3559 goto change; 3560 3561 p->sched_reset_on_fork = reset_on_fork; 3562 task_rq_unlock(rq, p, &flags); 3563 return 0; 3564 } 3565change: 3566 3567 if (user) { 3568#ifdef CONFIG_RT_GROUP_SCHED 3569 /* 3570 * Do not allow realtime tasks into groups that have no runtime 3571 * assigned. 3572 */ 3573 if (rt_bandwidth_enabled() && rt_policy(policy) && 3574 task_group(p)->rt_bandwidth.rt_runtime == 0 && 3575 !task_group_is_autogroup(task_group(p))) { 3576 task_rq_unlock(rq, p, &flags); 3577 return -EPERM; 3578 } 3579#endif 3580#ifdef CONFIG_SMP 3581 if (dl_bandwidth_enabled() && dl_policy(policy)) { 3582 cpumask_t *span = rq->rd->span; 3583 3584 /* 3585 * Don't allow tasks with an affinity mask smaller than 3586 * the entire root_domain to become SCHED_DEADLINE. We 3587 * will also fail if there's no bandwidth available. 3588 */ 3589 if (!cpumask_subset(span, &p->cpus_allowed) || 3590 rq->rd->dl_bw.bw == 0) { 3591 task_rq_unlock(rq, p, &flags); 3592 return -EPERM; 3593 } 3594 } 3595#endif 3596 } 3597 3598 /* recheck policy now with rq lock held */ 3599 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) { 3600 policy = oldpolicy = -1; 3601 task_rq_unlock(rq, p, &flags); 3602 goto recheck; 3603 } 3604 3605 /* 3606 * If setscheduling to SCHED_DEADLINE (or changing the parameters 3607 * of a SCHED_DEADLINE task) we need to check if enough bandwidth 3608 * is available. 3609 */ 3610 if ((dl_policy(policy) || dl_task(p)) && dl_overflow(p, policy, attr)) { 3611 task_rq_unlock(rq, p, &flags); 3612 return -EBUSY; 3613 } 3614 3615 p->sched_reset_on_fork = reset_on_fork; 3616 oldprio = p->prio; 3617 3618 /* 3619 * Special case for priority boosted tasks. 3620 * 3621 * If the new priority is lower or equal (user space view) 3622 * than the current (boosted) priority, we just store the new 3623 * normal parameters and do not touch the scheduler class and 3624 * the runqueue. This will be done when the task deboost 3625 * itself. 3626 */ 3627 if (rt_mutex_check_prio(p, newprio)) { 3628 __setscheduler_params(p, attr); 3629 task_rq_unlock(rq, p, &flags); 3630 return 0; 3631 } 3632 3633 queued = task_on_rq_queued(p); 3634 running = task_current(rq, p); 3635 if (queued) 3636 dequeue_task(rq, p, 0); 3637 if (running) 3638 put_prev_task(rq, p); 3639 3640 prev_class = p->sched_class; 3641 __setscheduler(rq, p, attr); 3642 3643 if (running) 3644 p->sched_class->set_curr_task(rq); 3645 if (queued) { 3646 /* 3647 * We enqueue to tail when the priority of a task is 3648 * increased (user space view). 3649 */ 3650 enqueue_task(rq, p, oldprio <= p->prio ? ENQUEUE_HEAD : 0); 3651 } 3652 3653 check_class_changed(rq, p, prev_class, oldprio); 3654 task_rq_unlock(rq, p, &flags); 3655 3656 rt_mutex_adjust_pi(p); 3657 3658 return 0; 3659} 3660 3661static int _sched_setscheduler(struct task_struct *p, int policy, 3662 const struct sched_param *param, bool check) 3663{ 3664 struct sched_attr attr = { 3665 .sched_policy = policy, 3666 .sched_priority = param->sched_priority, 3667 .sched_nice = PRIO_TO_NICE(p->static_prio), 3668 }; 3669 3670 /* Fixup the legacy SCHED_RESET_ON_FORK hack. */ 3671 if ((policy != SETPARAM_POLICY) && (policy & SCHED_RESET_ON_FORK)) { 3672 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK; 3673 policy &= ~SCHED_RESET_ON_FORK; 3674 attr.sched_policy = policy; 3675 } 3676 3677 return __sched_setscheduler(p, &attr, check); 3678} 3679/** 3680 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread. 3681 * @p: the task in question. 3682 * @policy: new policy. 3683 * @param: structure containing the new RT priority. 3684 * 3685 * Return: 0 on success. An error code otherwise. 3686 * 3687 * NOTE that the task may be already dead. 3688 */ 3689int sched_setscheduler(struct task_struct *p, int policy, 3690 const struct sched_param *param) 3691{ 3692 return _sched_setscheduler(p, policy, param, true); 3693} 3694EXPORT_SYMBOL_GPL(sched_setscheduler); 3695 3696int sched_setattr(struct task_struct *p, const struct sched_attr *attr) 3697{ 3698 return __sched_setscheduler(p, attr, true); 3699} 3700EXPORT_SYMBOL_GPL(sched_setattr); 3701 3702/** 3703 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace. 3704 * @p: the task in question. 3705 * @policy: new policy. 3706 * @param: structure containing the new RT priority. 3707 * 3708 * Just like sched_setscheduler, only don't bother checking if the 3709 * current context has permission. For example, this is needed in 3710 * stop_machine(): we create temporary high priority worker threads, 3711 * but our caller might not have that capability. 3712 * 3713 * Return: 0 on success. An error code otherwise. 3714 */ 3715int sched_setscheduler_nocheck(struct task_struct *p, int policy, 3716 const struct sched_param *param) 3717{ 3718 return _sched_setscheduler(p, policy, param, false); 3719} 3720 3721static int 3722do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param) 3723{ 3724 struct sched_param lparam; 3725 struct task_struct *p; 3726 int retval; 3727 3728 if (!param || pid < 0) 3729 return -EINVAL; 3730 if (copy_from_user(&lparam, param, sizeof(struct sched_param))) 3731 return -EFAULT; 3732 3733 rcu_read_lock(); 3734 retval = -ESRCH; 3735 p = find_process_by_pid(pid); 3736 if (p != NULL) 3737 retval = sched_setscheduler(p, policy, &lparam); 3738 rcu_read_unlock(); 3739 3740 return retval; 3741} 3742 3743/* 3744 * Mimics kernel/events/core.c perf_copy_attr(). 3745 */ 3746static int sched_copy_attr(struct sched_attr __user *uattr, 3747 struct sched_attr *attr) 3748{ 3749 u32 size; 3750 int ret; 3751 3752 if (!access_ok(VERIFY_WRITE, uattr, SCHED_ATTR_SIZE_VER0)) 3753 return -EFAULT; 3754 3755 /* 3756 * zero the full structure, so that a short copy will be nice. 3757 */ 3758 memset(attr, 0, sizeof(*attr)); 3759 3760 ret = get_user(size, &uattr->size); 3761 if (ret) 3762 return ret; 3763 3764 if (size > PAGE_SIZE) /* silly large */ 3765 goto err_size; 3766 3767 if (!size) /* abi compat */ 3768 size = SCHED_ATTR_SIZE_VER0; 3769 3770 if (size < SCHED_ATTR_SIZE_VER0) 3771 goto err_size; 3772 3773 /* 3774 * If we're handed a bigger struct than we know of, 3775 * ensure all the unknown bits are 0 - i.e. new 3776 * user-space does not rely on any kernel feature 3777 * extensions we dont know about yet. 3778 */ 3779 if (size > sizeof(*attr)) { 3780 unsigned char __user *addr; 3781 unsigned char __user *end; 3782 unsigned char val; 3783 3784 addr = (void __user *)uattr + sizeof(*attr); 3785 end = (void __user *)uattr + size; 3786 3787 for (; addr < end; addr++) { 3788 ret = get_user(val, addr); 3789 if (ret) 3790 return ret; 3791 if (val) 3792 goto err_size; 3793 } 3794 size = sizeof(*attr); 3795 } 3796 3797 ret = copy_from_user(attr, uattr, size); 3798 if (ret) 3799 return -EFAULT; 3800 3801 /* 3802 * XXX: do we want to be lenient like existing syscalls; or do we want 3803 * to be strict and return an error on out-of-bounds values? 3804 */ 3805 attr->sched_nice = clamp(attr->sched_nice, MIN_NICE, MAX_NICE); 3806 3807 return 0; 3808 3809err_size: 3810 put_user(sizeof(*attr), &uattr->size); 3811 return -E2BIG; 3812} 3813 3814/** 3815 * sys_sched_setscheduler - set/change the scheduler policy and RT priority 3816 * @pid: the pid in question. 3817 * @policy: new policy. 3818 * @param: structure containing the new RT priority. 3819 * 3820 * Return: 0 on success. An error code otherwise. 3821 */ 3822SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy, 3823 struct sched_param __user *, param) 3824{ 3825 /* negative values for policy are not valid */ 3826 if (policy < 0) 3827 return -EINVAL; 3828 3829 return do_sched_setscheduler(pid, policy, param); 3830} 3831 3832/** 3833 * sys_sched_setparam - set/change the RT priority of a thread 3834 * @pid: the pid in question. 3835 * @param: structure containing the new RT priority. 3836 * 3837 * Return: 0 on success. An error code otherwise. 3838 */ 3839SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param) 3840{ 3841 return do_sched_setscheduler(pid, SETPARAM_POLICY, param); 3842} 3843 3844/** 3845 * sys_sched_setattr - same as above, but with extended sched_attr 3846 * @pid: the pid in question. 3847 * @uattr: structure containing the extended parameters. 3848 * @flags: for future extension. 3849 */ 3850SYSCALL_DEFINE3(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr, 3851 unsigned int, flags) 3852{ 3853 struct sched_attr attr; 3854 struct task_struct *p; 3855 int retval; 3856 3857 if (!uattr || pid < 0 || flags) 3858 return -EINVAL; 3859 3860 retval = sched_copy_attr(uattr, &attr); 3861 if (retval) 3862 return retval; 3863 3864 if ((int)attr.sched_policy < 0) 3865 return -EINVAL; 3866 3867 rcu_read_lock(); 3868 retval = -ESRCH; 3869 p = find_process_by_pid(pid); 3870 if (p != NULL) 3871 retval = sched_setattr(p, &attr); 3872 rcu_read_unlock(); 3873 3874 return retval; 3875} 3876 3877/** 3878 * sys_sched_getscheduler - get the policy (scheduling class) of a thread 3879 * @pid: the pid in question. 3880 * 3881 * Return: On success, the policy of the thread. Otherwise, a negative error 3882 * code. 3883 */ 3884SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid) 3885{ 3886 struct task_struct *p; 3887 int retval; 3888 3889 if (pid < 0) 3890 return -EINVAL; 3891 3892 retval = -ESRCH; 3893 rcu_read_lock(); 3894 p = find_process_by_pid(pid); 3895 if (p) { 3896 retval = security_task_getscheduler(p); 3897 if (!retval) 3898 retval = p->policy 3899 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0); 3900 } 3901 rcu_read_unlock(); 3902 return retval; 3903} 3904 3905/** 3906 * sys_sched_getparam - get the RT priority of a thread 3907 * @pid: the pid in question. 3908 * @param: structure containing the RT priority. 3909 * 3910 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error 3911 * code. 3912 */ 3913SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param) 3914{ 3915 struct sched_param lp = { .sched_priority = 0 }; 3916 struct task_struct *p; 3917 int retval; 3918 3919 if (!param || pid < 0) 3920 return -EINVAL; 3921 3922 rcu_read_lock(); 3923 p = find_process_by_pid(pid); 3924 retval = -ESRCH; 3925 if (!p) 3926 goto out_unlock; 3927 3928 retval = security_task_getscheduler(p); 3929 if (retval) 3930 goto out_unlock; 3931 3932 if (task_has_rt_policy(p)) 3933 lp.sched_priority = p->rt_priority; 3934 rcu_read_unlock(); 3935 3936 /* 3937 * This one might sleep, we cannot do it with a spinlock held ... 3938 */ 3939 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0; 3940 3941 return retval; 3942 3943out_unlock: 3944 rcu_read_unlock(); 3945 return retval; 3946} 3947 3948static int sched_read_attr(struct sched_attr __user *uattr, 3949 struct sched_attr *attr, 3950 unsigned int usize) 3951{ 3952 int ret; 3953 3954 if (!access_ok(VERIFY_WRITE, uattr, usize)) 3955 return -EFAULT; 3956 3957 /* 3958 * If we're handed a smaller struct than we know of, 3959 * ensure all the unknown bits are 0 - i.e. old 3960 * user-space does not get uncomplete information. 3961 */ 3962 if (usize < sizeof(*attr)) { 3963 unsigned char *addr; 3964 unsigned char *end; 3965 3966 addr = (void *)attr + usize; 3967 end = (void *)attr + sizeof(*attr); 3968 3969 for (; addr < end; addr++) { 3970 if (*addr) 3971 return -EFBIG; 3972 } 3973 3974 attr->size = usize; 3975 } 3976 3977 ret = copy_to_user(uattr, attr, attr->size); 3978 if (ret) 3979 return -EFAULT; 3980 3981 return 0; 3982} 3983 3984/** 3985 * sys_sched_getattr - similar to sched_getparam, but with sched_attr 3986 * @pid: the pid in question. 3987 * @uattr: structure containing the extended parameters. 3988 * @size: sizeof(attr) for fwd/bwd comp. 3989 * @flags: for future extension. 3990 */ 3991SYSCALL_DEFINE4(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr, 3992 unsigned int, size, unsigned int, flags) 3993{ 3994 struct sched_attr attr = { 3995 .size = sizeof(struct sched_attr), 3996 }; 3997 struct task_struct *p; 3998 int retval; 3999 4000 if (!uattr || pid < 0 || size > PAGE_SIZE || 4001 size < SCHED_ATTR_SIZE_VER0 || flags) 4002 return -EINVAL; 4003 4004 rcu_read_lock(); 4005 p = find_process_by_pid(pid); 4006 retval = -ESRCH; 4007 if (!p) 4008 goto out_unlock; 4009 4010 retval = security_task_getscheduler(p); 4011 if (retval) 4012 goto out_unlock; 4013 4014 attr.sched_policy = p->policy; 4015 if (p->sched_reset_on_fork) 4016 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK; 4017 if (task_has_dl_policy(p)) 4018 __getparam_dl(p, &attr); 4019 else if (task_has_rt_policy(p)) 4020 attr.sched_priority = p->rt_priority; 4021 else 4022 attr.sched_nice = task_nice(p); 4023 4024 rcu_read_unlock(); 4025 4026 retval = sched_read_attr(uattr, &attr, size); 4027 return retval; 4028 4029out_unlock: 4030 rcu_read_unlock(); 4031 return retval; 4032} 4033 4034long sched_setaffinity(pid_t pid, const struct cpumask *in_mask) 4035{ 4036 cpumask_var_t cpus_allowed, new_mask; 4037 struct task_struct *p; 4038 int retval; 4039 4040 rcu_read_lock(); 4041 4042 p = find_process_by_pid(pid); 4043 if (!p) { 4044 rcu_read_unlock(); 4045 return -ESRCH; 4046 } 4047 4048 /* Prevent p going away */ 4049 get_task_struct(p); 4050 rcu_read_unlock(); 4051 4052 if (p->flags & PF_NO_SETAFFINITY) { 4053 retval = -EINVAL; 4054 goto out_put_task; 4055 } 4056 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) { 4057 retval = -ENOMEM; 4058 goto out_put_task; 4059 } 4060 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) { 4061 retval = -ENOMEM; 4062 goto out_free_cpus_allowed; 4063 } 4064 retval = -EPERM; 4065 if (!check_same_owner(p)) { 4066 rcu_read_lock(); 4067 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) { 4068 rcu_read_unlock(); 4069 goto out_free_new_mask; 4070 } 4071 rcu_read_unlock(); 4072 } 4073 4074 retval = security_task_setscheduler(p); 4075 if (retval) 4076 goto out_free_new_mask; 4077 4078 4079 cpuset_cpus_allowed(p, cpus_allowed); 4080 cpumask_and(new_mask, in_mask, cpus_allowed); 4081 4082 /* 4083 * Since bandwidth control happens on root_domain basis, 4084 * if admission test is enabled, we only admit -deadline 4085 * tasks allowed to run on all the CPUs in the task's 4086 * root_domain. 4087 */ 4088#ifdef CONFIG_SMP 4089 if (task_has_dl_policy(p) && dl_bandwidth_enabled()) { 4090 rcu_read_lock(); 4091 if (!cpumask_subset(task_rq(p)->rd->span, new_mask)) { 4092 retval = -EBUSY; 4093 rcu_read_unlock(); 4094 goto out_free_new_mask; 4095 } 4096 rcu_read_unlock(); 4097 } 4098#endif 4099again: 4100 retval = set_cpus_allowed_ptr(p, new_mask); 4101 4102 if (!retval) { 4103 cpuset_cpus_allowed(p, cpus_allowed); 4104 if (!cpumask_subset(new_mask, cpus_allowed)) { 4105 /* 4106 * We must have raced with a concurrent cpuset 4107 * update. Just reset the cpus_allowed to the 4108 * cpuset's cpus_allowed 4109 */ 4110 cpumask_copy(new_mask, cpus_allowed); 4111 goto again; 4112 } 4113 } 4114out_free_new_mask: 4115 free_cpumask_var(new_mask); 4116out_free_cpus_allowed: 4117 free_cpumask_var(cpus_allowed); 4118out_put_task: 4119 put_task_struct(p); 4120 return retval; 4121} 4122 4123static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len, 4124 struct cpumask *new_mask) 4125{ 4126 if (len < cpumask_size()) 4127 cpumask_clear(new_mask); 4128 else if (len > cpumask_size()) 4129 len = cpumask_size(); 4130 4131 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0; 4132} 4133 4134/** 4135 * sys_sched_setaffinity - set the cpu affinity of a process 4136 * @pid: pid of the process 4137 * @len: length in bytes of the bitmask pointed to by user_mask_ptr 4138 * @user_mask_ptr: user-space pointer to the new cpu mask 4139 * 4140 * Return: 0 on success. An error code otherwise. 4141 */ 4142SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len, 4143 unsigned long __user *, user_mask_ptr) 4144{ 4145 cpumask_var_t new_mask; 4146 int retval; 4147 4148 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) 4149 return -ENOMEM; 4150 4151 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask); 4152 if (retval == 0) 4153 retval = sched_setaffinity(pid, new_mask); 4154 free_cpumask_var(new_mask); 4155 return retval; 4156} 4157 4158long sched_getaffinity(pid_t pid, struct cpumask *mask) 4159{ 4160 struct task_struct *p; 4161 unsigned long flags; 4162 int retval; 4163 4164 rcu_read_lock(); 4165 4166 retval = -ESRCH; 4167 p = find_process_by_pid(pid); 4168 if (!p) 4169 goto out_unlock; 4170 4171 retval = security_task_getscheduler(p); 4172 if (retval) 4173 goto out_unlock; 4174 4175 raw_spin_lock_irqsave(&p->pi_lock, flags); 4176 cpumask_and(mask, &p->cpus_allowed, cpu_active_mask); 4177 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 4178 4179out_unlock: 4180 rcu_read_unlock(); 4181 4182 return retval; 4183} 4184 4185/** 4186 * sys_sched_getaffinity - get the cpu affinity of a process 4187 * @pid: pid of the process 4188 * @len: length in bytes of the bitmask pointed to by user_mask_ptr 4189 * @user_mask_ptr: user-space pointer to hold the current cpu mask 4190 * 4191 * Return: 0 on success. An error code otherwise. 4192 */ 4193SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len, 4194 unsigned long __user *, user_mask_ptr) 4195{ 4196 int ret; 4197 cpumask_var_t mask; 4198 4199 if ((len * BITS_PER_BYTE) < nr_cpu_ids) 4200 return -EINVAL; 4201 if (len & (sizeof(unsigned long)-1)) 4202 return -EINVAL; 4203 4204 if (!alloc_cpumask_var(&mask, GFP_KERNEL)) 4205 return -ENOMEM; 4206 4207 ret = sched_getaffinity(pid, mask); 4208 if (ret == 0) { 4209 size_t retlen = min_t(size_t, len, cpumask_size()); 4210 4211 if (copy_to_user(user_mask_ptr, mask, retlen)) 4212 ret = -EFAULT; 4213 else 4214 ret = retlen; 4215 } 4216 free_cpumask_var(mask); 4217 4218 return ret; 4219} 4220 4221/** 4222 * sys_sched_yield - yield the current processor to other threads. 4223 * 4224 * This function yields the current CPU to other tasks. If there are no 4225 * other threads running on this CPU then this function will return. 4226 * 4227 * Return: 0. 4228 */ 4229SYSCALL_DEFINE0(sched_yield) 4230{ 4231 struct rq *rq = this_rq_lock(); 4232 4233 schedstat_inc(rq, yld_count); 4234 current->sched_class->yield_task(rq); 4235 4236 /* 4237 * Since we are going to call schedule() anyway, there's 4238 * no need to preempt or enable interrupts: 4239 */ 4240 __release(rq->lock); 4241 spin_release(&rq->lock.dep_map, 1, _THIS_IP_); 4242 do_raw_spin_unlock(&rq->lock); 4243 sched_preempt_enable_no_resched(); 4244 4245 schedule(); 4246 4247 return 0; 4248} 4249 4250static void __cond_resched(void) 4251{ 4252 __preempt_count_add(PREEMPT_ACTIVE); 4253 __schedule(); 4254 __preempt_count_sub(PREEMPT_ACTIVE); 4255} 4256 4257int __sched _cond_resched(void) 4258{ 4259 if (should_resched()) { 4260 __cond_resched(); 4261 return 1; 4262 } 4263 return 0; 4264} 4265EXPORT_SYMBOL(_cond_resched); 4266 4267/* 4268 * __cond_resched_lock() - if a reschedule is pending, drop the given lock, 4269 * call schedule, and on return reacquire the lock. 4270 * 4271 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level 4272 * operations here to prevent schedule() from being called twice (once via 4273 * spin_unlock(), once by hand). 4274 */ 4275int __cond_resched_lock(spinlock_t *lock) 4276{ 4277 int resched = should_resched(); 4278 int ret = 0; 4279 4280 lockdep_assert_held(lock); 4281 4282 if (spin_needbreak(lock) || resched) { 4283 spin_unlock(lock); 4284 if (resched) 4285 __cond_resched(); 4286 else 4287 cpu_relax(); 4288 ret = 1; 4289 spin_lock(lock); 4290 } 4291 return ret; 4292} 4293EXPORT_SYMBOL(__cond_resched_lock); 4294 4295int __sched __cond_resched_softirq(void) 4296{ 4297 BUG_ON(!in_softirq()); 4298 4299 if (should_resched()) { 4300 local_bh_enable(); 4301 __cond_resched(); 4302 local_bh_disable(); 4303 return 1; 4304 } 4305 return 0; 4306} 4307EXPORT_SYMBOL(__cond_resched_softirq); 4308 4309/** 4310 * yield - yield the current processor to other threads. 4311 * 4312 * Do not ever use this function, there's a 99% chance you're doing it wrong. 4313 * 4314 * The scheduler is at all times free to pick the calling task as the most 4315 * eligible task to run, if removing the yield() call from your code breaks 4316 * it, its already broken. 4317 * 4318 * Typical broken usage is: 4319 * 4320 * while (!event) 4321 * yield(); 4322 * 4323 * where one assumes that yield() will let 'the other' process run that will 4324 * make event true. If the current task is a SCHED_FIFO task that will never 4325 * happen. Never use yield() as a progress guarantee!! 4326 * 4327 * If you want to use yield() to wait for something, use wait_event(). 4328 * If you want to use yield() to be 'nice' for others, use cond_resched(). 4329 * If you still want to use yield(), do not! 4330 */ 4331void __sched yield(void) 4332{ 4333 set_current_state(TASK_RUNNING); 4334 sys_sched_yield(); 4335} 4336EXPORT_SYMBOL(yield); 4337 4338/** 4339 * yield_to - yield the current processor to another thread in 4340 * your thread group, or accelerate that thread toward the 4341 * processor it's on. 4342 * @p: target task 4343 * @preempt: whether task preemption is allowed or not 4344 * 4345 * It's the caller's job to ensure that the target task struct 4346 * can't go away on us before we can do any checks. 4347 * 4348 * Return: 4349 * true (>0) if we indeed boosted the target task. 4350 * false (0) if we failed to boost the target. 4351 * -ESRCH if there's no task to yield to. 4352 */ 4353int __sched yield_to(struct task_struct *p, bool preempt) 4354{ 4355 struct task_struct *curr = current; 4356 struct rq *rq, *p_rq; 4357 unsigned long flags; 4358 int yielded = 0; 4359 4360 local_irq_save(flags); 4361 rq = this_rq(); 4362 4363again: 4364 p_rq = task_rq(p); 4365 /* 4366 * If we're the only runnable task on the rq and target rq also 4367 * has only one task, there's absolutely no point in yielding. 4368 */ 4369 if (rq->nr_running == 1 && p_rq->nr_running == 1) { 4370 yielded = -ESRCH; 4371 goto out_irq; 4372 } 4373 4374 double_rq_lock(rq, p_rq); 4375 if (task_rq(p) != p_rq) { 4376 double_rq_unlock(rq, p_rq); 4377 goto again; 4378 } 4379 4380 if (!curr->sched_class->yield_to_task) 4381 goto out_unlock; 4382 4383 if (curr->sched_class != p->sched_class) 4384 goto out_unlock; 4385 4386 if (task_running(p_rq, p) || p->state) 4387 goto out_unlock; 4388 4389 yielded = curr->sched_class->yield_to_task(rq, p, preempt); 4390 if (yielded) { 4391 schedstat_inc(rq, yld_count); 4392 /* 4393 * Make p's CPU reschedule; pick_next_entity takes care of 4394 * fairness. 4395 */ 4396 if (preempt && rq != p_rq) 4397 resched_curr(p_rq); 4398 } 4399 4400out_unlock: 4401 double_rq_unlock(rq, p_rq); 4402out_irq: 4403 local_irq_restore(flags); 4404 4405 if (yielded > 0) 4406 schedule(); 4407 4408 return yielded; 4409} 4410EXPORT_SYMBOL_GPL(yield_to); 4411 4412/* 4413 * This task is about to go to sleep on IO. Increment rq->nr_iowait so 4414 * that process accounting knows that this is a task in IO wait state. 4415 */ 4416void __sched io_schedule(void) 4417{ 4418 struct rq *rq = raw_rq(); 4419 4420 delayacct_blkio_start(); 4421 atomic_inc(&rq->nr_iowait); 4422 blk_flush_plug(current); 4423 current->in_iowait = 1; 4424 schedule(); 4425 current->in_iowait = 0; 4426 atomic_dec(&rq->nr_iowait); 4427 delayacct_blkio_end(); 4428} 4429EXPORT_SYMBOL(io_schedule); 4430 4431long __sched io_schedule_timeout(long timeout) 4432{ 4433 struct rq *rq = raw_rq(); 4434 long ret; 4435 4436 delayacct_blkio_start(); 4437 atomic_inc(&rq->nr_iowait); 4438 blk_flush_plug(current); 4439 current->in_iowait = 1; 4440 ret = schedule_timeout(timeout); 4441 current->in_iowait = 0; 4442 atomic_dec(&rq->nr_iowait); 4443 delayacct_blkio_end(); 4444 return ret; 4445} 4446 4447/** 4448 * sys_sched_get_priority_max - return maximum RT priority. 4449 * @policy: scheduling class. 4450 * 4451 * Return: On success, this syscall returns the maximum 4452 * rt_priority that can be used by a given scheduling class. 4453 * On failure, a negative error code is returned. 4454 */ 4455SYSCALL_DEFINE1(sched_get_priority_max, int, policy) 4456{ 4457 int ret = -EINVAL; 4458 4459 switch (policy) { 4460 case SCHED_FIFO: 4461 case SCHED_RR: 4462 ret = MAX_USER_RT_PRIO-1; 4463 break; 4464 case SCHED_DEADLINE: 4465 case SCHED_NORMAL: 4466 case SCHED_BATCH: 4467 case SCHED_IDLE: 4468 ret = 0; 4469 break; 4470 } 4471 return ret; 4472} 4473 4474/** 4475 * sys_sched_get_priority_min - return minimum RT priority. 4476 * @policy: scheduling class. 4477 * 4478 * Return: On success, this syscall returns the minimum 4479 * rt_priority that can be used by a given scheduling class. 4480 * On failure, a negative error code is returned. 4481 */ 4482SYSCALL_DEFINE1(sched_get_priority_min, int, policy) 4483{ 4484 int ret = -EINVAL; 4485 4486 switch (policy) { 4487 case SCHED_FIFO: 4488 case SCHED_RR: 4489 ret = 1; 4490 break; 4491 case SCHED_DEADLINE: 4492 case SCHED_NORMAL: 4493 case SCHED_BATCH: 4494 case SCHED_IDLE: 4495 ret = 0; 4496 } 4497 return ret; 4498} 4499 4500/** 4501 * sys_sched_rr_get_interval - return the default timeslice of a process. 4502 * @pid: pid of the process. 4503 * @interval: userspace pointer to the timeslice value. 4504 * 4505 * this syscall writes the default timeslice value of a given process 4506 * into the user-space timespec buffer. A value of '0' means infinity. 4507 * 4508 * Return: On success, 0 and the timeslice is in @interval. Otherwise, 4509 * an error code. 4510 */ 4511SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid, 4512 struct timespec __user *, interval) 4513{ 4514 struct task_struct *p; 4515 unsigned int time_slice; 4516 unsigned long flags; 4517 struct rq *rq; 4518 int retval; 4519 struct timespec t; 4520 4521 if (pid < 0) 4522 return -EINVAL; 4523 4524 retval = -ESRCH; 4525 rcu_read_lock(); 4526 p = find_process_by_pid(pid); 4527 if (!p) 4528 goto out_unlock; 4529 4530 retval = security_task_getscheduler(p); 4531 if (retval) 4532 goto out_unlock; 4533 4534 rq = task_rq_lock(p, &flags); 4535 time_slice = 0; 4536 if (p->sched_class->get_rr_interval) 4537 time_slice = p->sched_class->get_rr_interval(rq, p); 4538 task_rq_unlock(rq, p, &flags); 4539 4540 rcu_read_unlock(); 4541 jiffies_to_timespec(time_slice, &t); 4542 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0; 4543 return retval; 4544 4545out_unlock: 4546 rcu_read_unlock(); 4547 return retval; 4548} 4549 4550static const char stat_nam[] = TASK_STATE_TO_CHAR_STR; 4551 4552void sched_show_task(struct task_struct *p) 4553{ 4554 unsigned long free = 0; 4555 int ppid; 4556 unsigned state; 4557 4558 state = p->state ? __ffs(p->state) + 1 : 0; 4559 printk(KERN_INFO "%-15.15s %c", p->comm, 4560 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?'); 4561#if BITS_PER_LONG == 32 4562 if (state == TASK_RUNNING) 4563 printk(KERN_CONT " running "); 4564 else 4565 printk(KERN_CONT " %08lx ", thread_saved_pc(p)); 4566#else 4567 if (state == TASK_RUNNING) 4568 printk(KERN_CONT " running task "); 4569 else 4570 printk(KERN_CONT " %016lx ", thread_saved_pc(p)); 4571#endif 4572#ifdef CONFIG_DEBUG_STACK_USAGE 4573 free = stack_not_used(p); 4574#endif 4575 rcu_read_lock(); 4576 ppid = task_pid_nr(rcu_dereference(p->real_parent)); 4577 rcu_read_unlock(); 4578 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free, 4579 task_pid_nr(p), ppid, 4580 (unsigned long)task_thread_info(p)->flags); 4581 4582 print_worker_info(KERN_INFO, p); 4583 show_stack(p, NULL); 4584} 4585 4586void show_state_filter(unsigned long state_filter) 4587{ 4588 struct task_struct *g, *p; 4589 4590#if BITS_PER_LONG == 32 4591 printk(KERN_INFO 4592 " task PC stack pid father\n"); 4593#else 4594 printk(KERN_INFO 4595 " task PC stack pid father\n"); 4596#endif 4597 rcu_read_lock(); 4598 for_each_process_thread(g, p) { 4599 /* 4600 * reset the NMI-timeout, listing all files on a slow 4601 * console might take a lot of time: 4602 */ 4603 touch_nmi_watchdog(); 4604 if (!state_filter || (p->state & state_filter)) 4605 sched_show_task(p); 4606 } 4607 4608 touch_all_softlockup_watchdogs(); 4609 4610#ifdef CONFIG_SCHED_DEBUG 4611 sysrq_sched_debug_show(); 4612#endif 4613 rcu_read_unlock(); 4614 /* 4615 * Only show locks if all tasks are dumped: 4616 */ 4617 if (!state_filter) 4618 debug_show_all_locks(); 4619} 4620 4621void init_idle_bootup_task(struct task_struct *idle) 4622{ 4623 idle->sched_class = &idle_sched_class; 4624} 4625 4626/** 4627 * init_idle - set up an idle thread for a given CPU 4628 * @idle: task in question 4629 * @cpu: cpu the idle task belongs to 4630 * 4631 * NOTE: this function does not set the idle thread's NEED_RESCHED 4632 * flag, to make booting more robust. 4633 */ 4634void init_idle(struct task_struct *idle, int cpu) 4635{ 4636 struct rq *rq = cpu_rq(cpu); 4637 unsigned long flags; 4638 4639 raw_spin_lock_irqsave(&rq->lock, flags); 4640 4641 __sched_fork(0, idle); 4642 idle->state = TASK_RUNNING; 4643 idle->se.exec_start = sched_clock(); 4644 4645 do_set_cpus_allowed(idle, cpumask_of(cpu)); 4646 /* 4647 * We're having a chicken and egg problem, even though we are 4648 * holding rq->lock, the cpu isn't yet set to this cpu so the 4649 * lockdep check in task_group() will fail. 4650 * 4651 * Similar case to sched_fork(). / Alternatively we could 4652 * use task_rq_lock() here and obtain the other rq->lock. 4653 * 4654 * Silence PROVE_RCU 4655 */ 4656 rcu_read_lock(); 4657 __set_task_cpu(idle, cpu); 4658 rcu_read_unlock(); 4659 4660 rq->curr = rq->idle = idle; 4661 idle->on_rq = TASK_ON_RQ_QUEUED; 4662#if defined(CONFIG_SMP) 4663 idle->on_cpu = 1; 4664#endif 4665 raw_spin_unlock_irqrestore(&rq->lock, flags); 4666 4667 /* Set the preempt count _outside_ the spinlocks! */ 4668 init_idle_preempt_count(idle, cpu); 4669 4670 /* 4671 * The idle tasks have their own, simple scheduling class: 4672 */ 4673 idle->sched_class = &idle_sched_class; 4674 ftrace_graph_init_idle_task(idle, cpu); 4675 vtime_init_idle(idle, cpu); 4676#if defined(CONFIG_SMP) 4677 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu); 4678#endif 4679} 4680 4681#ifdef CONFIG_SMP 4682/* 4683 * move_queued_task - move a queued task to new rq. 4684 * 4685 * Returns (locked) new rq. Old rq's lock is released. 4686 */ 4687static struct rq *move_queued_task(struct task_struct *p, int new_cpu) 4688{ 4689 struct rq *rq = task_rq(p); 4690 4691 lockdep_assert_held(&rq->lock); 4692 4693 dequeue_task(rq, p, 0); 4694 p->on_rq = TASK_ON_RQ_MIGRATING; 4695 set_task_cpu(p, new_cpu); 4696 raw_spin_unlock(&rq->lock); 4697 4698 rq = cpu_rq(new_cpu); 4699 4700 raw_spin_lock(&rq->lock); 4701 BUG_ON(task_cpu(p) != new_cpu); 4702 p->on_rq = TASK_ON_RQ_QUEUED; 4703 enqueue_task(rq, p, 0); 4704 check_preempt_curr(rq, p, 0); 4705 4706 return rq; 4707} 4708 4709void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask) 4710{ 4711 if (p->sched_class && p->sched_class->set_cpus_allowed) 4712 p->sched_class->set_cpus_allowed(p, new_mask); 4713 4714 cpumask_copy(&p->cpus_allowed, new_mask); 4715 p->nr_cpus_allowed = cpumask_weight(new_mask); 4716} 4717 4718/* 4719 * This is how migration works: 4720 * 4721 * 1) we invoke migration_cpu_stop() on the target CPU using 4722 * stop_one_cpu(). 4723 * 2) stopper starts to run (implicitly forcing the migrated thread 4724 * off the CPU) 4725 * 3) it checks whether the migrated task is still in the wrong runqueue. 4726 * 4) if it's in the wrong runqueue then the migration thread removes 4727 * it and puts it into the right queue. 4728 * 5) stopper completes and stop_one_cpu() returns and the migration 4729 * is done. 4730 */ 4731 4732/* 4733 * Change a given task's CPU affinity. Migrate the thread to a 4734 * proper CPU and schedule it away if the CPU it's executing on 4735 * is removed from the allowed bitmask. 4736 * 4737 * NOTE: the caller must have a valid reference to the task, the 4738 * task must not exit() & deallocate itself prematurely. The 4739 * call is not atomic; no spinlocks may be held. 4740 */ 4741int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask) 4742{ 4743 unsigned long flags; 4744 struct rq *rq; 4745 unsigned int dest_cpu; 4746 int ret = 0; 4747 4748 rq = task_rq_lock(p, &flags); 4749 4750 if (cpumask_equal(&p->cpus_allowed, new_mask)) 4751 goto out; 4752 4753 if (!cpumask_intersects(new_mask, cpu_active_mask)) { 4754 ret = -EINVAL; 4755 goto out; 4756 } 4757 4758 do_set_cpus_allowed(p, new_mask); 4759 4760 /* Can the task run on the task's current CPU? If so, we're done */ 4761 if (cpumask_test_cpu(task_cpu(p), new_mask)) 4762 goto out; 4763 4764 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask); 4765 if (task_running(rq, p) || p->state == TASK_WAKING) { 4766 struct migration_arg arg = { p, dest_cpu }; 4767 /* Need help from migration thread: drop lock and wait. */ 4768 task_rq_unlock(rq, p, &flags); 4769 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg); 4770 tlb_migrate_finish(p->mm); 4771 return 0; 4772 } else if (task_on_rq_queued(p)) 4773 rq = move_queued_task(p, dest_cpu); 4774out: 4775 task_rq_unlock(rq, p, &flags); 4776 4777 return ret; 4778} 4779EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr); 4780 4781/* 4782 * Move (not current) task off this cpu, onto dest cpu. We're doing 4783 * this because either it can't run here any more (set_cpus_allowed() 4784 * away from this CPU, or CPU going down), or because we're 4785 * attempting to rebalance this task on exec (sched_exec). 4786 * 4787 * So we race with normal scheduler movements, but that's OK, as long 4788 * as the task is no longer on this CPU. 4789 * 4790 * Returns non-zero if task was successfully migrated. 4791 */ 4792static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu) 4793{ 4794 struct rq *rq; 4795 int ret = 0; 4796 4797 if (unlikely(!cpu_active(dest_cpu))) 4798 return ret; 4799 4800 rq = cpu_rq(src_cpu); 4801 4802 raw_spin_lock(&p->pi_lock); 4803 raw_spin_lock(&rq->lock); 4804 /* Already moved. */ 4805 if (task_cpu(p) != src_cpu) 4806 goto done; 4807 4808 /* Affinity changed (again). */ 4809 if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p))) 4810 goto fail; 4811 4812 /* 4813 * If we're not on a rq, the next wake-up will ensure we're 4814 * placed properly. 4815 */ 4816 if (task_on_rq_queued(p)) 4817 rq = move_queued_task(p, dest_cpu); 4818done: 4819 ret = 1; 4820fail: 4821 raw_spin_unlock(&rq->lock); 4822 raw_spin_unlock(&p->pi_lock); 4823 return ret; 4824} 4825 4826#ifdef CONFIG_NUMA_BALANCING 4827/* Migrate current task p to target_cpu */ 4828int migrate_task_to(struct task_struct *p, int target_cpu) 4829{ 4830 struct migration_arg arg = { p, target_cpu }; 4831 int curr_cpu = task_cpu(p); 4832 4833 if (curr_cpu == target_cpu) 4834 return 0; 4835 4836 if (!cpumask_test_cpu(target_cpu, tsk_cpus_allowed(p))) 4837 return -EINVAL; 4838 4839 /* TODO: This is not properly updating schedstats */ 4840 4841 trace_sched_move_numa(p, curr_cpu, target_cpu); 4842 return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg); 4843} 4844 4845/* 4846 * Requeue a task on a given node and accurately track the number of NUMA 4847 * tasks on the runqueues 4848 */ 4849void sched_setnuma(struct task_struct *p, int nid) 4850{ 4851 struct rq *rq; 4852 unsigned long flags; 4853 bool queued, running; 4854 4855 rq = task_rq_lock(p, &flags); 4856 queued = task_on_rq_queued(p); 4857 running = task_current(rq, p); 4858 4859 if (queued) 4860 dequeue_task(rq, p, 0); 4861 if (running) 4862 put_prev_task(rq, p); 4863 4864 p->numa_preferred_nid = nid; 4865 4866 if (running) 4867 p->sched_class->set_curr_task(rq); 4868 if (queued) 4869 enqueue_task(rq, p, 0); 4870 task_rq_unlock(rq, p, &flags); 4871} 4872#endif 4873 4874/* 4875 * migration_cpu_stop - this will be executed by a highprio stopper thread 4876 * and performs thread migration by bumping thread off CPU then 4877 * 'pushing' onto another runqueue. 4878 */ 4879static int migration_cpu_stop(void *data) 4880{ 4881 struct migration_arg *arg = data; 4882 4883 /* 4884 * The original target cpu might have gone down and we might 4885 * be on another cpu but it doesn't matter. 4886 */ 4887 local_irq_disable(); 4888 /* 4889 * We need to explicitly wake pending tasks before running 4890 * __migrate_task() such that we will not miss enforcing cpus_allowed 4891 * during wakeups, see set_cpus_allowed_ptr()'s TASK_WAKING test. 4892 */ 4893 sched_ttwu_pending(); 4894 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu); 4895 local_irq_enable(); 4896 return 0; 4897} 4898 4899#ifdef CONFIG_HOTPLUG_CPU 4900 4901/* 4902 * Ensures that the idle task is using init_mm right before its cpu goes 4903 * offline. 4904 */ 4905void idle_task_exit(void) 4906{ 4907 struct mm_struct *mm = current->active_mm; 4908 4909 BUG_ON(cpu_online(smp_processor_id())); 4910 4911 if (mm != &init_mm) { 4912 switch_mm(mm, &init_mm, current); 4913 finish_arch_post_lock_switch(); 4914 } 4915 mmdrop(mm); 4916} 4917 4918/* 4919 * Since this CPU is going 'away' for a while, fold any nr_active delta 4920 * we might have. Assumes we're called after migrate_tasks() so that the 4921 * nr_active count is stable. 4922 * 4923 * Also see the comment "Global load-average calculations". 4924 */ 4925static void calc_load_migrate(struct rq *rq) 4926{ 4927 long delta = calc_load_fold_active(rq); 4928 if (delta) 4929 atomic_long_add(delta, &calc_load_tasks); 4930} 4931 4932static void put_prev_task_fake(struct rq *rq, struct task_struct *prev) 4933{ 4934} 4935 4936static const struct sched_class fake_sched_class = { 4937 .put_prev_task = put_prev_task_fake, 4938}; 4939 4940static struct task_struct fake_task = { 4941 /* 4942 * Avoid pull_{rt,dl}_task() 4943 */ 4944 .prio = MAX_PRIO + 1, 4945 .sched_class = &fake_sched_class, 4946}; 4947 4948/* 4949 * Migrate all tasks from the rq, sleeping tasks will be migrated by 4950 * try_to_wake_up()->select_task_rq(). 4951 * 4952 * Called with rq->lock held even though we'er in stop_machine() and 4953 * there's no concurrency possible, we hold the required locks anyway 4954 * because of lock validation efforts. 4955 */ 4956static void migrate_tasks(unsigned int dead_cpu) 4957{ 4958 struct rq *rq = cpu_rq(dead_cpu); 4959 struct task_struct *next, *stop = rq->stop; 4960 int dest_cpu; 4961 4962 /* 4963 * Fudge the rq selection such that the below task selection loop 4964 * doesn't get stuck on the currently eligible stop task. 4965 * 4966 * We're currently inside stop_machine() and the rq is either stuck 4967 * in the stop_machine_cpu_stop() loop, or we're executing this code, 4968 * either way we should never end up calling schedule() until we're 4969 * done here. 4970 */ 4971 rq->stop = NULL; 4972 4973 /* 4974 * put_prev_task() and pick_next_task() sched 4975 * class method both need to have an up-to-date 4976 * value of rq->clock[_task] 4977 */ 4978 update_rq_clock(rq); 4979 4980 for ( ; ; ) { 4981 /* 4982 * There's this thread running, bail when that's the only 4983 * remaining thread. 4984 */ 4985 if (rq->nr_running == 1) 4986 break; 4987 4988 next = pick_next_task(rq, &fake_task); 4989 BUG_ON(!next); 4990 next->sched_class->put_prev_task(rq, next); 4991 4992 /* Find suitable destination for @next, with force if needed. */ 4993 dest_cpu = select_fallback_rq(dead_cpu, next); 4994 raw_spin_unlock(&rq->lock); 4995 4996 __migrate_task(next, dead_cpu, dest_cpu); 4997 4998 raw_spin_lock(&rq->lock); 4999 } 5000 5001 rq->stop = stop; 5002} 5003 5004#endif /* CONFIG_HOTPLUG_CPU */ 5005 5006#if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL) 5007 5008static struct ctl_table sd_ctl_dir[] = { 5009 { 5010 .procname = "sched_domain", 5011 .mode = 0555, 5012 }, 5013 {} 5014}; 5015 5016static struct ctl_table sd_ctl_root[] = { 5017 { 5018 .procname = "kernel", 5019 .mode = 0555, 5020 .child = sd_ctl_dir, 5021 }, 5022 {} 5023}; 5024 5025static struct ctl_table *sd_alloc_ctl_entry(int n) 5026{ 5027 struct ctl_table *entry = 5028 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL); 5029 5030 return entry; 5031} 5032 5033static void sd_free_ctl_entry(struct ctl_table **tablep) 5034{ 5035 struct ctl_table *entry; 5036 5037 /* 5038 * In the intermediate directories, both the child directory and 5039 * procname are dynamically allocated and could fail but the mode 5040 * will always be set. In the lowest directory the names are 5041 * static strings and all have proc handlers. 5042 */ 5043 for (entry = *tablep; entry->mode; entry++) { 5044 if (entry->child) 5045 sd_free_ctl_entry(&entry->child); 5046 if (entry->proc_handler == NULL) 5047 kfree(entry->procname); 5048 } 5049 5050 kfree(*tablep); 5051 *tablep = NULL; 5052} 5053 5054static int min_load_idx = 0; 5055static int max_load_idx = CPU_LOAD_IDX_MAX-1; 5056 5057static void 5058set_table_entry(struct ctl_table *entry, 5059 const char *procname, void *data, int maxlen, 5060 umode_t mode, proc_handler *proc_handler, 5061 bool load_idx) 5062{ 5063 entry->procname = procname; 5064 entry->data = data; 5065 entry->maxlen = maxlen; 5066 entry->mode = mode; 5067 entry->proc_handler = proc_handler; 5068 5069 if (load_idx) { 5070 entry->extra1 = &min_load_idx; 5071 entry->extra2 = &max_load_idx; 5072 } 5073} 5074 5075static struct ctl_table * 5076sd_alloc_ctl_domain_table(struct sched_domain *sd) 5077{ 5078 struct ctl_table *table = sd_alloc_ctl_entry(14); 5079 5080 if (table == NULL) 5081 return NULL; 5082 5083 set_table_entry(&table[0], "min_interval", &sd->min_interval, 5084 sizeof(long), 0644, proc_doulongvec_minmax, false); 5085 set_table_entry(&table[1], "max_interval", &sd->max_interval, 5086 sizeof(long), 0644, proc_doulongvec_minmax, false); 5087 set_table_entry(&table[2], "busy_idx", &sd->busy_idx, 5088 sizeof(int), 0644, proc_dointvec_minmax, true); 5089 set_table_entry(&table[3], "idle_idx", &sd->idle_idx, 5090 sizeof(int), 0644, proc_dointvec_minmax, true); 5091 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx, 5092 sizeof(int), 0644, proc_dointvec_minmax, true); 5093 set_table_entry(&table[5], "wake_idx", &sd->wake_idx, 5094 sizeof(int), 0644, proc_dointvec_minmax, true); 5095 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx, 5096 sizeof(int), 0644, proc_dointvec_minmax, true); 5097 set_table_entry(&table[7], "busy_factor", &sd->busy_factor, 5098 sizeof(int), 0644, proc_dointvec_minmax, false); 5099 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct, 5100 sizeof(int), 0644, proc_dointvec_minmax, false); 5101 set_table_entry(&table[9], "cache_nice_tries", 5102 &sd->cache_nice_tries, 5103 sizeof(int), 0644, proc_dointvec_minmax, false); 5104 set_table_entry(&table[10], "flags", &sd->flags, 5105 sizeof(int), 0644, proc_dointvec_minmax, false); 5106 set_table_entry(&table[11], "max_newidle_lb_cost", 5107 &sd->max_newidle_lb_cost, 5108 sizeof(long), 0644, proc_doulongvec_minmax, false); 5109 set_table_entry(&table[12], "name", sd->name, 5110 CORENAME_MAX_SIZE, 0444, proc_dostring, false); 5111 /* &table[13] is terminator */ 5112 5113 return table; 5114} 5115 5116static struct ctl_table *sd_alloc_ctl_cpu_table(int cpu) 5117{ 5118 struct ctl_table *entry, *table; 5119 struct sched_domain *sd; 5120 int domain_num = 0, i; 5121 char buf[32]; 5122 5123 for_each_domain(cpu, sd) 5124 domain_num++; 5125 entry = table = sd_alloc_ctl_entry(domain_num + 1); 5126 if (table == NULL) 5127 return NULL; 5128 5129 i = 0; 5130 for_each_domain(cpu, sd) { 5131 snprintf(buf, 32, "domain%d", i); 5132 entry->procname = kstrdup(buf, GFP_KERNEL); 5133 entry->mode = 0555; 5134 entry->child = sd_alloc_ctl_domain_table(sd); 5135 entry++; 5136 i++; 5137 } 5138 return table; 5139} 5140 5141static struct ctl_table_header *sd_sysctl_header; 5142static void register_sched_domain_sysctl(void) 5143{ 5144 int i, cpu_num = num_possible_cpus(); 5145 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1); 5146 char buf[32]; 5147 5148 WARN_ON(sd_ctl_dir[0].child); 5149 sd_ctl_dir[0].child = entry; 5150 5151 if (entry == NULL) 5152 return; 5153 5154 for_each_possible_cpu(i) { 5155 snprintf(buf, 32, "cpu%d", i); 5156 entry->procname = kstrdup(buf, GFP_KERNEL); 5157 entry->mode = 0555; 5158 entry->child = sd_alloc_ctl_cpu_table(i); 5159 entry++; 5160 } 5161 5162 WARN_ON(sd_sysctl_header); 5163 sd_sysctl_header = register_sysctl_table(sd_ctl_root); 5164} 5165 5166/* may be called multiple times per register */ 5167static void unregister_sched_domain_sysctl(void) 5168{ 5169 if (sd_sysctl_header) 5170 unregister_sysctl_table(sd_sysctl_header); 5171 sd_sysctl_header = NULL; 5172 if (sd_ctl_dir[0].child) 5173 sd_free_ctl_entry(&sd_ctl_dir[0].child); 5174} 5175#else 5176static void register_sched_domain_sysctl(void) 5177{ 5178} 5179static void unregister_sched_domain_sysctl(void) 5180{ 5181} 5182#endif 5183 5184static void set_rq_online(struct rq *rq) 5185{ 5186 if (!rq->online) { 5187 const struct sched_class *class; 5188 5189 cpumask_set_cpu(rq->cpu, rq->rd->online); 5190 rq->online = 1; 5191 5192 for_each_class(class) { 5193 if (class->rq_online) 5194 class->rq_online(rq); 5195 } 5196 } 5197} 5198 5199static void set_rq_offline(struct rq *rq) 5200{ 5201 if (rq->online) { 5202 const struct sched_class *class; 5203 5204 for_each_class(class) { 5205 if (class->rq_offline) 5206 class->rq_offline(rq); 5207 } 5208 5209 cpumask_clear_cpu(rq->cpu, rq->rd->online); 5210 rq->online = 0; 5211 } 5212} 5213 5214/* 5215 * migration_call - callback that gets triggered when a CPU is added. 5216 * Here we can start up the necessary migration thread for the new CPU. 5217 */ 5218static int 5219migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu) 5220{ 5221 int cpu = (long)hcpu; 5222 unsigned long flags; 5223 struct rq *rq = cpu_rq(cpu); 5224 5225 switch (action & ~CPU_TASKS_FROZEN) { 5226 5227 case CPU_UP_PREPARE: 5228 rq->calc_load_update = calc_load_update; 5229 break; 5230 5231 case CPU_ONLINE: 5232 /* Update our root-domain */ 5233 raw_spin_lock_irqsave(&rq->lock, flags); 5234 if (rq->rd) { 5235 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span)); 5236 5237 set_rq_online(rq); 5238 } 5239 raw_spin_unlock_irqrestore(&rq->lock, flags); 5240 break; 5241 5242#ifdef CONFIG_HOTPLUG_CPU 5243 case CPU_DYING: 5244 sched_ttwu_pending(); 5245 /* Update our root-domain */ 5246 raw_spin_lock_irqsave(&rq->lock, flags); 5247 if (rq->rd) { 5248 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span)); 5249 set_rq_offline(rq); 5250 } 5251 migrate_tasks(cpu); 5252 BUG_ON(rq->nr_running != 1); /* the migration thread */ 5253 raw_spin_unlock_irqrestore(&rq->lock, flags); 5254 break; 5255 5256 case CPU_DEAD: 5257 calc_load_migrate(rq); 5258 break; 5259#endif 5260 } 5261 5262 update_max_interval(); 5263 5264 return NOTIFY_OK; 5265} 5266 5267/* 5268 * Register at high priority so that task migration (migrate_all_tasks) 5269 * happens before everything else. This has to be lower priority than 5270 * the notifier in the perf_event subsystem, though. 5271 */ 5272static struct notifier_block migration_notifier = { 5273 .notifier_call = migration_call, 5274 .priority = CPU_PRI_MIGRATION, 5275}; 5276 5277static void __cpuinit set_cpu_rq_start_time(void) 5278{ 5279 int cpu = smp_processor_id(); 5280 struct rq *rq = cpu_rq(cpu); 5281 rq->age_stamp = sched_clock_cpu(cpu); 5282} 5283 5284static int sched_cpu_active(struct notifier_block *nfb, 5285 unsigned long action, void *hcpu) 5286{ 5287 switch (action & ~CPU_TASKS_FROZEN) { 5288 case CPU_STARTING: 5289 set_cpu_rq_start_time(); 5290 return NOTIFY_OK; 5291 case CPU_DOWN_FAILED: 5292 set_cpu_active((long)hcpu, true); 5293 return NOTIFY_OK; 5294 default: 5295 return NOTIFY_DONE; 5296 } 5297} 5298 5299static int sched_cpu_inactive(struct notifier_block *nfb, 5300 unsigned long action, void *hcpu) 5301{ 5302 unsigned long flags; 5303 long cpu = (long)hcpu; 5304 struct dl_bw *dl_b; 5305 5306 switch (action & ~CPU_TASKS_FROZEN) { 5307 case CPU_DOWN_PREPARE: 5308 set_cpu_active(cpu, false); 5309 5310 /* explicitly allow suspend */ 5311 if (!(action & CPU_TASKS_FROZEN)) { 5312 bool overflow; 5313 int cpus; 5314 5315 rcu_read_lock_sched(); 5316 dl_b = dl_bw_of(cpu); 5317 5318 raw_spin_lock_irqsave(&dl_b->lock, flags); 5319 cpus = dl_bw_cpus(cpu); 5320 overflow = __dl_overflow(dl_b, cpus, 0, 0); 5321 raw_spin_unlock_irqrestore(&dl_b->lock, flags); 5322 5323 rcu_read_unlock_sched(); 5324 5325 if (overflow) 5326 return notifier_from_errno(-EBUSY); 5327 } 5328 return NOTIFY_OK; 5329 } 5330 5331 return NOTIFY_DONE; 5332} 5333 5334static int __init migration_init(void) 5335{ 5336 void *cpu = (void *)(long)smp_processor_id(); 5337 int err; 5338 5339 /* Initialize migration for the boot CPU */ 5340 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu); 5341 BUG_ON(err == NOTIFY_BAD); 5342 migration_call(&migration_notifier, CPU_ONLINE, cpu); 5343 register_cpu_notifier(&migration_notifier); 5344 5345 /* Register cpu active notifiers */ 5346 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE); 5347 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE); 5348 5349 return 0; 5350} 5351early_initcall(migration_init); 5352#endif 5353 5354#ifdef CONFIG_SMP 5355 5356static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */ 5357 5358#ifdef CONFIG_SCHED_DEBUG 5359 5360static __read_mostly int sched_debug_enabled; 5361 5362static int __init sched_debug_setup(char *str) 5363{ 5364 sched_debug_enabled = 1; 5365 5366 return 0; 5367} 5368early_param("sched_debug", sched_debug_setup); 5369 5370static inline bool sched_debug(void) 5371{ 5372 return sched_debug_enabled; 5373} 5374 5375static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level, 5376 struct cpumask *groupmask) 5377{ 5378 struct sched_group *group = sd->groups; 5379 char str[256]; 5380 5381 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd)); 5382 cpumask_clear(groupmask); 5383 5384 printk(KERN_DEBUG "%*s domain %d: ", level, "", level); 5385 5386 if (!(sd->flags & SD_LOAD_BALANCE)) { 5387 printk("does not load-balance\n"); 5388 if (sd->parent) 5389 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain" 5390 " has parent"); 5391 return -1; 5392 } 5393 5394 printk(KERN_CONT "span %s level %s\n", str, sd->name); 5395 5396 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) { 5397 printk(KERN_ERR "ERROR: domain->span does not contain " 5398 "CPU%d\n", cpu); 5399 } 5400 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) { 5401 printk(KERN_ERR "ERROR: domain->groups does not contain" 5402 " CPU%d\n", cpu); 5403 } 5404 5405 printk(KERN_DEBUG "%*s groups:", level + 1, ""); 5406 do { 5407 if (!group) { 5408 printk("\n"); 5409 printk(KERN_ERR "ERROR: group is NULL\n"); 5410 break; 5411 } 5412 5413 /* 5414 * Even though we initialize ->capacity to something semi-sane, 5415 * we leave capacity_orig unset. This allows us to detect if 5416 * domain iteration is still funny without causing /0 traps. 5417 */ 5418 if (!group->sgc->capacity_orig) { 5419 printk(KERN_CONT "\n"); 5420 printk(KERN_ERR "ERROR: domain->cpu_capacity not set\n"); 5421 break; 5422 } 5423 5424 if (!cpumask_weight(sched_group_cpus(group))) { 5425 printk(KERN_CONT "\n"); 5426 printk(KERN_ERR "ERROR: empty group\n"); 5427 break; 5428 } 5429 5430 if (!(sd->flags & SD_OVERLAP) && 5431 cpumask_intersects(groupmask, sched_group_cpus(group))) { 5432 printk(KERN_CONT "\n"); 5433 printk(KERN_ERR "ERROR: repeated CPUs\n"); 5434 break; 5435 } 5436 5437 cpumask_or(groupmask, groupmask, sched_group_cpus(group)); 5438 5439 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group)); 5440 5441 printk(KERN_CONT " %s", str); 5442 if (group->sgc->capacity != SCHED_CAPACITY_SCALE) { 5443 printk(KERN_CONT " (cpu_capacity = %d)", 5444 group->sgc->capacity); 5445 } 5446 5447 group = group->next; 5448 } while (group != sd->groups); 5449 printk(KERN_CONT "\n"); 5450 5451 if (!cpumask_equal(sched_domain_span(sd), groupmask)) 5452 printk(KERN_ERR "ERROR: groups don't span domain->span\n"); 5453 5454 if (sd->parent && 5455 !cpumask_subset(groupmask, sched_domain_span(sd->parent))) 5456 printk(KERN_ERR "ERROR: parent span is not a superset " 5457 "of domain->span\n"); 5458 return 0; 5459} 5460 5461static void sched_domain_debug(struct sched_domain *sd, int cpu) 5462{ 5463 int level = 0; 5464 5465 if (!sched_debug_enabled) 5466 return; 5467 5468 if (!sd) { 5469 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu); 5470 return; 5471 } 5472 5473 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu); 5474 5475 for (;;) { 5476 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask)) 5477 break; 5478 level++; 5479 sd = sd->parent; 5480 if (!sd) 5481 break; 5482 } 5483} 5484#else /* !CONFIG_SCHED_DEBUG */ 5485# define sched_domain_debug(sd, cpu) do { } while (0) 5486static inline bool sched_debug(void) 5487{ 5488 return false; 5489} 5490#endif /* CONFIG_SCHED_DEBUG */ 5491 5492static int sd_degenerate(struct sched_domain *sd) 5493{ 5494 if (cpumask_weight(sched_domain_span(sd)) == 1) 5495 return 1; 5496 5497 /* Following flags need at least 2 groups */ 5498 if (sd->flags & (SD_LOAD_BALANCE | 5499 SD_BALANCE_NEWIDLE | 5500 SD_BALANCE_FORK | 5501 SD_BALANCE_EXEC | 5502 SD_SHARE_CPUCAPACITY | 5503 SD_SHARE_PKG_RESOURCES | 5504 SD_SHARE_POWERDOMAIN)) { 5505 if (sd->groups != sd->groups->next) 5506 return 0; 5507 } 5508 5509 /* Following flags don't use groups */ 5510 if (sd->flags & (SD_WAKE_AFFINE)) 5511 return 0; 5512 5513 return 1; 5514} 5515 5516static int 5517sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent) 5518{ 5519 unsigned long cflags = sd->flags, pflags = parent->flags; 5520 5521 if (sd_degenerate(parent)) 5522 return 1; 5523 5524 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent))) 5525 return 0; 5526 5527 /* Flags needing groups don't count if only 1 group in parent */ 5528 if (parent->groups == parent->groups->next) { 5529 pflags &= ~(SD_LOAD_BALANCE | 5530 SD_BALANCE_NEWIDLE | 5531 SD_BALANCE_FORK | 5532 SD_BALANCE_EXEC | 5533 SD_SHARE_CPUCAPACITY | 5534 SD_SHARE_PKG_RESOURCES | 5535 SD_PREFER_SIBLING | 5536 SD_SHARE_POWERDOMAIN); 5537 if (nr_node_ids == 1) 5538 pflags &= ~SD_SERIALIZE; 5539 } 5540 if (~cflags & pflags) 5541 return 0; 5542 5543 return 1; 5544} 5545 5546static void free_rootdomain(struct rcu_head *rcu) 5547{ 5548 struct root_domain *rd = container_of(rcu, struct root_domain, rcu); 5549 5550 cpupri_cleanup(&rd->cpupri); 5551 cpudl_cleanup(&rd->cpudl); 5552 free_cpumask_var(rd->dlo_mask); 5553 free_cpumask_var(rd->rto_mask); 5554 free_cpumask_var(rd->online); 5555 free_cpumask_var(rd->span); 5556 kfree(rd); 5557} 5558 5559static void rq_attach_root(struct rq *rq, struct root_domain *rd) 5560{ 5561 struct root_domain *old_rd = NULL; 5562 unsigned long flags; 5563 5564 raw_spin_lock_irqsave(&rq->lock, flags); 5565 5566 if (rq->rd) { 5567 old_rd = rq->rd; 5568 5569 if (cpumask_test_cpu(rq->cpu, old_rd->online)) 5570 set_rq_offline(rq); 5571 5572 cpumask_clear_cpu(rq->cpu, old_rd->span); 5573 5574 /* 5575 * If we dont want to free the old_rd yet then 5576 * set old_rd to NULL to skip the freeing later 5577 * in this function: 5578 */ 5579 if (!atomic_dec_and_test(&old_rd->refcount)) 5580 old_rd = NULL; 5581 } 5582 5583 atomic_inc(&rd->refcount); 5584 rq->rd = rd; 5585 5586 cpumask_set_cpu(rq->cpu, rd->span); 5587 if (cpumask_test_cpu(rq->cpu, cpu_active_mask)) 5588 set_rq_online(rq); 5589 5590 raw_spin_unlock_irqrestore(&rq->lock, flags); 5591 5592 if (old_rd) 5593 call_rcu_sched(&old_rd->rcu, free_rootdomain); 5594} 5595 5596static int init_rootdomain(struct root_domain *rd) 5597{ 5598 memset(rd, 0, sizeof(*rd)); 5599 5600 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL)) 5601 goto out; 5602 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL)) 5603 goto free_span; 5604 if (!alloc_cpumask_var(&rd->dlo_mask, GFP_KERNEL)) 5605 goto free_online; 5606 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL)) 5607 goto free_dlo_mask; 5608 5609 init_dl_bw(&rd->dl_bw); 5610 if (cpudl_init(&rd->cpudl) != 0) 5611 goto free_dlo_mask; 5612 5613 if (cpupri_init(&rd->cpupri) != 0) 5614 goto free_rto_mask; 5615 return 0; 5616 5617free_rto_mask: 5618 free_cpumask_var(rd->rto_mask); 5619free_dlo_mask: 5620 free_cpumask_var(rd->dlo_mask); 5621free_online: 5622 free_cpumask_var(rd->online); 5623free_span: 5624 free_cpumask_var(rd->span); 5625out: 5626 return -ENOMEM; 5627} 5628 5629/* 5630 * By default the system creates a single root-domain with all cpus as 5631 * members (mimicking the global state we have today). 5632 */ 5633struct root_domain def_root_domain; 5634 5635static void init_defrootdomain(void) 5636{ 5637 init_rootdomain(&def_root_domain); 5638 5639 atomic_set(&def_root_domain.refcount, 1); 5640} 5641 5642static struct root_domain *alloc_rootdomain(void) 5643{ 5644 struct root_domain *rd; 5645 5646 rd = kmalloc(sizeof(*rd), GFP_KERNEL); 5647 if (!rd) 5648 return NULL; 5649 5650 if (init_rootdomain(rd) != 0) { 5651 kfree(rd); 5652 return NULL; 5653 } 5654 5655 return rd; 5656} 5657 5658static void free_sched_groups(struct sched_group *sg, int free_sgc) 5659{ 5660 struct sched_group *tmp, *first; 5661 5662 if (!sg) 5663 return; 5664 5665 first = sg; 5666 do { 5667 tmp = sg->next; 5668 5669 if (free_sgc && atomic_dec_and_test(&sg->sgc->ref)) 5670 kfree(sg->sgc); 5671 5672 kfree(sg); 5673 sg = tmp; 5674 } while (sg != first); 5675} 5676 5677static void free_sched_domain(struct rcu_head *rcu) 5678{ 5679 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu); 5680 5681 /* 5682 * If its an overlapping domain it has private groups, iterate and 5683 * nuke them all. 5684 */ 5685 if (sd->flags & SD_OVERLAP) { 5686 free_sched_groups(sd->groups, 1); 5687 } else if (atomic_dec_and_test(&sd->groups->ref)) { 5688 kfree(sd->groups->sgc); 5689 kfree(sd->groups); 5690 } 5691 kfree(sd); 5692} 5693 5694static void destroy_sched_domain(struct sched_domain *sd, int cpu) 5695{ 5696 call_rcu(&sd->rcu, free_sched_domain); 5697} 5698 5699static void destroy_sched_domains(struct sched_domain *sd, int cpu) 5700{ 5701 for (; sd; sd = sd->parent) 5702 destroy_sched_domain(sd, cpu); 5703} 5704 5705/* 5706 * Keep a special pointer to the highest sched_domain that has 5707 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this 5708 * allows us to avoid some pointer chasing select_idle_sibling(). 5709 * 5710 * Also keep a unique ID per domain (we use the first cpu number in 5711 * the cpumask of the domain), this allows us to quickly tell if 5712 * two cpus are in the same cache domain, see cpus_share_cache(). 5713 */ 5714DEFINE_PER_CPU(struct sched_domain *, sd_llc); 5715DEFINE_PER_CPU(int, sd_llc_size); 5716DEFINE_PER_CPU(int, sd_llc_id); 5717DEFINE_PER_CPU(struct sched_domain *, sd_numa); 5718DEFINE_PER_CPU(struct sched_domain *, sd_busy); 5719DEFINE_PER_CPU(struct sched_domain *, sd_asym); 5720 5721static void update_top_cache_domain(int cpu) 5722{ 5723 struct sched_domain *sd; 5724 struct sched_domain *busy_sd = NULL; 5725 int id = cpu; 5726 int size = 1; 5727 5728 sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES); 5729 if (sd) { 5730 id = cpumask_first(sched_domain_span(sd)); 5731 size = cpumask_weight(sched_domain_span(sd)); 5732 busy_sd = sd->parent; /* sd_busy */ 5733 } 5734 rcu_assign_pointer(per_cpu(sd_busy, cpu), busy_sd); 5735 5736 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd); 5737 per_cpu(sd_llc_size, cpu) = size; 5738 per_cpu(sd_llc_id, cpu) = id; 5739 5740 sd = lowest_flag_domain(cpu, SD_NUMA); 5741 rcu_assign_pointer(per_cpu(sd_numa, cpu), sd); 5742 5743 sd = highest_flag_domain(cpu, SD_ASYM_PACKING); 5744 rcu_assign_pointer(per_cpu(sd_asym, cpu), sd); 5745} 5746 5747/* 5748 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must 5749 * hold the hotplug lock. 5750 */ 5751static void 5752cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu) 5753{ 5754 struct rq *rq = cpu_rq(cpu); 5755 struct sched_domain *tmp; 5756 5757 /* Remove the sched domains which do not contribute to scheduling. */ 5758 for (tmp = sd; tmp; ) { 5759 struct sched_domain *parent = tmp->parent; 5760 if (!parent) 5761 break; 5762 5763 if (sd_parent_degenerate(tmp, parent)) { 5764 tmp->parent = parent->parent; 5765 if (parent->parent) 5766 parent->parent->child = tmp; 5767 /* 5768 * Transfer SD_PREFER_SIBLING down in case of a 5769 * degenerate parent; the spans match for this 5770 * so the property transfers. 5771 */ 5772 if (parent->flags & SD_PREFER_SIBLING) 5773 tmp->flags |= SD_PREFER_SIBLING; 5774 destroy_sched_domain(parent, cpu); 5775 } else 5776 tmp = tmp->parent; 5777 } 5778 5779 if (sd && sd_degenerate(sd)) { 5780 tmp = sd; 5781 sd = sd->parent; 5782 destroy_sched_domain(tmp, cpu); 5783 if (sd) 5784 sd->child = NULL; 5785 } 5786 5787 sched_domain_debug(sd, cpu); 5788 5789 rq_attach_root(rq, rd); 5790 tmp = rq->sd; 5791 rcu_assign_pointer(rq->sd, sd); 5792 destroy_sched_domains(tmp, cpu); 5793 5794 update_top_cache_domain(cpu); 5795} 5796 5797/* cpus with isolated domains */ 5798static cpumask_var_t cpu_isolated_map; 5799 5800/* Setup the mask of cpus configured for isolated domains */ 5801static int __init isolated_cpu_setup(char *str) 5802{ 5803 alloc_bootmem_cpumask_var(&cpu_isolated_map); 5804 cpulist_parse(str, cpu_isolated_map); 5805 return 1; 5806} 5807 5808__setup("isolcpus=", isolated_cpu_setup); 5809 5810struct s_data { 5811 struct sched_domain ** __percpu sd; 5812 struct root_domain *rd; 5813}; 5814 5815enum s_alloc { 5816 sa_rootdomain, 5817 sa_sd, 5818 sa_sd_storage, 5819 sa_none, 5820}; 5821 5822/* 5823 * Build an iteration mask that can exclude certain CPUs from the upwards 5824 * domain traversal. 5825 * 5826 * Asymmetric node setups can result in situations where the domain tree is of 5827 * unequal depth, make sure to skip domains that already cover the entire 5828 * range. 5829 * 5830 * In that case build_sched_domains() will have terminated the iteration early 5831 * and our sibling sd spans will be empty. Domains should always include the 5832 * cpu they're built on, so check that. 5833 * 5834 */ 5835static void build_group_mask(struct sched_domain *sd, struct sched_group *sg) 5836{ 5837 const struct cpumask *span = sched_domain_span(sd); 5838 struct sd_data *sdd = sd->private; 5839 struct sched_domain *sibling; 5840 int i; 5841 5842 for_each_cpu(i, span) { 5843 sibling = *per_cpu_ptr(sdd->sd, i); 5844 if (!cpumask_test_cpu(i, sched_domain_span(sibling))) 5845 continue; 5846 5847 cpumask_set_cpu(i, sched_group_mask(sg)); 5848 } 5849} 5850 5851/* 5852 * Return the canonical balance cpu for this group, this is the first cpu 5853 * of this group that's also in the iteration mask. 5854 */ 5855int group_balance_cpu(struct sched_group *sg) 5856{ 5857 return cpumask_first_and(sched_group_cpus(sg), sched_group_mask(sg)); 5858} 5859 5860static int 5861build_overlap_sched_groups(struct sched_domain *sd, int cpu) 5862{ 5863 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg; 5864 const struct cpumask *span = sched_domain_span(sd); 5865 struct cpumask *covered = sched_domains_tmpmask; 5866 struct sd_data *sdd = sd->private; 5867 struct sched_domain *sibling; 5868 int i; 5869 5870 cpumask_clear(covered); 5871 5872 for_each_cpu(i, span) { 5873 struct cpumask *sg_span; 5874 5875 if (cpumask_test_cpu(i, covered)) 5876 continue; 5877 5878 sibling = *per_cpu_ptr(sdd->sd, i); 5879 5880 /* See the comment near build_group_mask(). */ 5881 if (!cpumask_test_cpu(i, sched_domain_span(sibling))) 5882 continue; 5883 5884 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(), 5885 GFP_KERNEL, cpu_to_node(cpu)); 5886 5887 if (!sg) 5888 goto fail; 5889 5890 sg_span = sched_group_cpus(sg); 5891 if (sibling->child) 5892 cpumask_copy(sg_span, sched_domain_span(sibling->child)); 5893 else 5894 cpumask_set_cpu(i, sg_span); 5895 5896 cpumask_or(covered, covered, sg_span); 5897 5898 sg->sgc = *per_cpu_ptr(sdd->sgc, i); 5899 if (atomic_inc_return(&sg->sgc->ref) == 1) 5900 build_group_mask(sd, sg); 5901 5902 /* 5903 * Initialize sgc->capacity such that even if we mess up the 5904 * domains and no possible iteration will get us here, we won't 5905 * die on a /0 trap. 5906 */ 5907 sg->sgc->capacity = SCHED_CAPACITY_SCALE * cpumask_weight(sg_span); 5908 sg->sgc->capacity_orig = sg->sgc->capacity; 5909 5910 /* 5911 * Make sure the first group of this domain contains the 5912 * canonical balance cpu. Otherwise the sched_domain iteration 5913 * breaks. See update_sg_lb_stats(). 5914 */ 5915 if ((!groups && cpumask_test_cpu(cpu, sg_span)) || 5916 group_balance_cpu(sg) == cpu) 5917 groups = sg; 5918 5919 if (!first) 5920 first = sg; 5921 if (last) 5922 last->next = sg; 5923 last = sg; 5924 last->next = first; 5925 } 5926 sd->groups = groups; 5927 5928 return 0; 5929 5930fail: 5931 free_sched_groups(first, 0); 5932 5933 return -ENOMEM; 5934} 5935 5936static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg) 5937{ 5938 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu); 5939 struct sched_domain *child = sd->child; 5940 5941 if (child) 5942 cpu = cpumask_first(sched_domain_span(child)); 5943 5944 if (sg) { 5945 *sg = *per_cpu_ptr(sdd->sg, cpu); 5946 (*sg)->sgc = *per_cpu_ptr(sdd->sgc, cpu); 5947 atomic_set(&(*sg)->sgc->ref, 1); /* for claim_allocations */ 5948 } 5949 5950 return cpu; 5951} 5952 5953/* 5954 * build_sched_groups will build a circular linked list of the groups 5955 * covered by the given span, and will set each group's ->cpumask correctly, 5956 * and ->cpu_capacity to 0. 5957 * 5958 * Assumes the sched_domain tree is fully constructed 5959 */ 5960static int 5961build_sched_groups(struct sched_domain *sd, int cpu) 5962{ 5963 struct sched_group *first = NULL, *last = NULL; 5964 struct sd_data *sdd = sd->private; 5965 const struct cpumask *span = sched_domain_span(sd); 5966 struct cpumask *covered; 5967 int i; 5968 5969 get_group(cpu, sdd, &sd->groups); 5970 atomic_inc(&sd->groups->ref); 5971 5972 if (cpu != cpumask_first(span)) 5973 return 0; 5974 5975 lockdep_assert_held(&sched_domains_mutex); 5976 covered = sched_domains_tmpmask; 5977 5978 cpumask_clear(covered); 5979 5980 for_each_cpu(i, span) { 5981 struct sched_group *sg; 5982 int group, j; 5983 5984 if (cpumask_test_cpu(i, covered)) 5985 continue; 5986 5987 group = get_group(i, sdd, &sg); 5988 cpumask_setall(sched_group_mask(sg)); 5989 5990 for_each_cpu(j, span) { 5991 if (get_group(j, sdd, NULL) != group) 5992 continue; 5993 5994 cpumask_set_cpu(j, covered); 5995 cpumask_set_cpu(j, sched_group_cpus(sg)); 5996 } 5997 5998 if (!first) 5999 first = sg; 6000 if (last) 6001 last->next = sg; 6002 last = sg; 6003 } 6004 last->next = first; 6005 6006 return 0; 6007} 6008 6009/* 6010 * Initialize sched groups cpu_capacity. 6011 * 6012 * cpu_capacity indicates the capacity of sched group, which is used while 6013 * distributing the load between different sched groups in a sched domain. 6014 * Typically cpu_capacity for all the groups in a sched domain will be same 6015 * unless there are asymmetries in the topology. If there are asymmetries, 6016 * group having more cpu_capacity will pickup more load compared to the 6017 * group having less cpu_capacity. 6018 */ 6019static void init_sched_groups_capacity(int cpu, struct sched_domain *sd) 6020{ 6021 struct sched_group *sg = sd->groups; 6022 6023 WARN_ON(!sg); 6024 6025 do { 6026 sg->group_weight = cpumask_weight(sched_group_cpus(sg)); 6027 sg = sg->next; 6028 } while (sg != sd->groups); 6029 6030 if (cpu != group_balance_cpu(sg)) 6031 return; 6032 6033 update_group_capacity(sd, cpu); 6034 atomic_set(&sg->sgc->nr_busy_cpus, sg->group_weight); 6035} 6036 6037/* 6038 * Initializers for schedule domains 6039 * Non-inlined to reduce accumulated stack pressure in build_sched_domains() 6040 */ 6041 6042static int default_relax_domain_level = -1; 6043int sched_domain_level_max; 6044 6045static int __init setup_relax_domain_level(char *str) 6046{ 6047 if (kstrtoint(str, 0, &default_relax_domain_level)) 6048 pr_warn("Unable to set relax_domain_level\n"); 6049 6050 return 1; 6051} 6052__setup("relax_domain_level=", setup_relax_domain_level); 6053 6054static void set_domain_attribute(struct sched_domain *sd, 6055 struct sched_domain_attr *attr) 6056{ 6057 int request; 6058 6059 if (!attr || attr->relax_domain_level < 0) { 6060 if (default_relax_domain_level < 0) 6061 return; 6062 else 6063 request = default_relax_domain_level; 6064 } else 6065 request = attr->relax_domain_level; 6066 if (request < sd->level) { 6067 /* turn off idle balance on this domain */ 6068 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE); 6069 } else { 6070 /* turn on idle balance on this domain */ 6071 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE); 6072 } 6073} 6074 6075static void __sdt_free(const struct cpumask *cpu_map); 6076static int __sdt_alloc(const struct cpumask *cpu_map); 6077 6078static void __free_domain_allocs(struct s_data *d, enum s_alloc what, 6079 const struct cpumask *cpu_map) 6080{ 6081 switch (what) { 6082 case sa_rootdomain: 6083 if (!atomic_read(&d->rd->refcount)) 6084 free_rootdomain(&d->rd->rcu); /* fall through */ 6085 case sa_sd: 6086 free_percpu(d->sd); /* fall through */ 6087 case sa_sd_storage: 6088 __sdt_free(cpu_map); /* fall through */ 6089 case sa_none: 6090 break; 6091 } 6092} 6093 6094static enum s_alloc __visit_domain_allocation_hell(struct s_data *d, 6095 const struct cpumask *cpu_map) 6096{ 6097 memset(d, 0, sizeof(*d)); 6098 6099 if (__sdt_alloc(cpu_map)) 6100 return sa_sd_storage; 6101 d->sd = alloc_percpu(struct sched_domain *); 6102 if (!d->sd) 6103 return sa_sd_storage; 6104 d->rd = alloc_rootdomain(); 6105 if (!d->rd) 6106 return sa_sd; 6107 return sa_rootdomain; 6108} 6109 6110/* 6111 * NULL the sd_data elements we've used to build the sched_domain and 6112 * sched_group structure so that the subsequent __free_domain_allocs() 6113 * will not free the data we're using. 6114 */ 6115static void claim_allocations(int cpu, struct sched_domain *sd) 6116{ 6117 struct sd_data *sdd = sd->private; 6118 6119 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd); 6120 *per_cpu_ptr(sdd->sd, cpu) = NULL; 6121 6122 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref)) 6123 *per_cpu_ptr(sdd->sg, cpu) = NULL; 6124 6125 if (atomic_read(&(*per_cpu_ptr(sdd->sgc, cpu))->ref)) 6126 *per_cpu_ptr(sdd->sgc, cpu) = NULL; 6127} 6128 6129#ifdef CONFIG_NUMA 6130static int sched_domains_numa_levels; 6131static int *sched_domains_numa_distance; 6132static struct cpumask ***sched_domains_numa_masks; 6133static int sched_domains_curr_level; 6134#endif 6135 6136/* 6137 * SD_flags allowed in topology descriptions. 6138 * 6139 * SD_SHARE_CPUCAPACITY - describes SMT topologies 6140 * SD_SHARE_PKG_RESOURCES - describes shared caches 6141 * SD_NUMA - describes NUMA topologies 6142 * SD_SHARE_POWERDOMAIN - describes shared power domain 6143 * 6144 * Odd one out: 6145 * SD_ASYM_PACKING - describes SMT quirks 6146 */ 6147#define TOPOLOGY_SD_FLAGS \ 6148 (SD_SHARE_CPUCAPACITY | \ 6149 SD_SHARE_PKG_RESOURCES | \ 6150 SD_NUMA | \ 6151 SD_ASYM_PACKING | \ 6152 SD_SHARE_POWERDOMAIN) 6153 6154static struct sched_domain * 6155sd_init(struct sched_domain_topology_level *tl, int cpu) 6156{ 6157 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); 6158 int sd_weight, sd_flags = 0; 6159 6160#ifdef CONFIG_NUMA 6161 /* 6162 * Ugly hack to pass state to sd_numa_mask()... 6163 */ 6164 sched_domains_curr_level = tl->numa_level; 6165#endif 6166 6167 sd_weight = cpumask_weight(tl->mask(cpu)); 6168 6169 if (tl->sd_flags) 6170 sd_flags = (*tl->sd_flags)(); 6171 if (WARN_ONCE(sd_flags & ~TOPOLOGY_SD_FLAGS, 6172 "wrong sd_flags in topology description\n")) 6173 sd_flags &= ~TOPOLOGY_SD_FLAGS; 6174 6175 *sd = (struct sched_domain){ 6176 .min_interval = sd_weight, 6177 .max_interval = 2*sd_weight, 6178 .busy_factor = 32, 6179 .imbalance_pct = 125, 6180 6181 .cache_nice_tries = 0, 6182 .busy_idx = 0, 6183 .idle_idx = 0, 6184 .newidle_idx = 0, 6185 .wake_idx = 0, 6186 .forkexec_idx = 0, 6187 6188 .flags = 1*SD_LOAD_BALANCE 6189 | 1*SD_BALANCE_NEWIDLE 6190 | 1*SD_BALANCE_EXEC 6191 | 1*SD_BALANCE_FORK 6192 | 0*SD_BALANCE_WAKE 6193 | 1*SD_WAKE_AFFINE 6194 | 0*SD_SHARE_CPUCAPACITY 6195 | 0*SD_SHARE_PKG_RESOURCES 6196 | 0*SD_SERIALIZE 6197 | 0*SD_PREFER_SIBLING 6198 | 0*SD_NUMA 6199 | sd_flags 6200 , 6201 6202 .last_balance = jiffies, 6203 .balance_interval = sd_weight, 6204 .smt_gain = 0, 6205 .max_newidle_lb_cost = 0, 6206 .next_decay_max_lb_cost = jiffies, 6207#ifdef CONFIG_SCHED_DEBUG 6208 .name = tl->name, 6209#endif 6210 }; 6211 6212 /* 6213 * Convert topological properties into behaviour. 6214 */ 6215 6216 if (sd->flags & SD_SHARE_CPUCAPACITY) { 6217 sd->imbalance_pct = 110; 6218 sd->smt_gain = 1178; /* ~15% */ 6219 6220 } else if (sd->flags & SD_SHARE_PKG_RESOURCES) { 6221 sd->imbalance_pct = 117; 6222 sd->cache_nice_tries = 1; 6223 sd->busy_idx = 2; 6224 6225#ifdef CONFIG_NUMA 6226 } else if (sd->flags & SD_NUMA) { 6227 sd->cache_nice_tries = 2; 6228 sd->busy_idx = 3; 6229 sd->idle_idx = 2; 6230 6231 sd->flags |= SD_SERIALIZE; 6232 if (sched_domains_numa_distance[tl->numa_level] > RECLAIM_DISTANCE) { 6233 sd->flags &= ~(SD_BALANCE_EXEC | 6234 SD_BALANCE_FORK | 6235 SD_WAKE_AFFINE); 6236 } 6237 6238#endif 6239 } else { 6240 sd->flags |= SD_PREFER_SIBLING; 6241 sd->cache_nice_tries = 1; 6242 sd->busy_idx = 2; 6243 sd->idle_idx = 1; 6244 } 6245 6246 sd->private = &tl->data; 6247 6248 return sd; 6249} 6250 6251/* 6252 * Topology list, bottom-up. 6253 */ 6254static struct sched_domain_topology_level default_topology[] = { 6255#ifdef CONFIG_SCHED_SMT 6256 { cpu_smt_mask, cpu_smt_flags, SD_INIT_NAME(SMT) }, 6257#endif 6258#ifdef CONFIG_SCHED_MC 6259 { cpu_coregroup_mask, cpu_core_flags, SD_INIT_NAME(MC) }, 6260#endif 6261 { cpu_cpu_mask, SD_INIT_NAME(DIE) }, 6262 { NULL, }, 6263}; 6264 6265struct sched_domain_topology_level *sched_domain_topology = default_topology; 6266 6267#define for_each_sd_topology(tl) \ 6268 for (tl = sched_domain_topology; tl->mask; tl++) 6269 6270void set_sched_topology(struct sched_domain_topology_level *tl) 6271{ 6272 sched_domain_topology = tl; 6273} 6274 6275#ifdef CONFIG_NUMA 6276 6277static const struct cpumask *sd_numa_mask(int cpu) 6278{ 6279 return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)]; 6280} 6281 6282static void sched_numa_warn(const char *str) 6283{ 6284 static int done = false; 6285 int i,j; 6286 6287 if (done) 6288 return; 6289 6290 done = true; 6291 6292 printk(KERN_WARNING "ERROR: %s\n\n", str); 6293 6294 for (i = 0; i < nr_node_ids; i++) { 6295 printk(KERN_WARNING " "); 6296 for (j = 0; j < nr_node_ids; j++) 6297 printk(KERN_CONT "%02d ", node_distance(i,j)); 6298 printk(KERN_CONT "\n"); 6299 } 6300 printk(KERN_WARNING "\n"); 6301} 6302 6303static bool find_numa_distance(int distance) 6304{ 6305 int i; 6306 6307 if (distance == node_distance(0, 0)) 6308 return true; 6309 6310 for (i = 0; i < sched_domains_numa_levels; i++) { 6311 if (sched_domains_numa_distance[i] == distance) 6312 return true; 6313 } 6314 6315 return false; 6316} 6317 6318static void sched_init_numa(void) 6319{ 6320 int next_distance, curr_distance = node_distance(0, 0); 6321 struct sched_domain_topology_level *tl; 6322 int level = 0; 6323 int i, j, k; 6324 6325 sched_domains_numa_distance = kzalloc(sizeof(int) * nr_node_ids, GFP_KERNEL); 6326 if (!sched_domains_numa_distance) 6327 return; 6328 6329 /* 6330 * O(nr_nodes^2) deduplicating selection sort -- in order to find the 6331 * unique distances in the node_distance() table. 6332 * 6333 * Assumes node_distance(0,j) includes all distances in 6334 * node_distance(i,j) in order to avoid cubic time. 6335 */ 6336 next_distance = curr_distance; 6337 for (i = 0; i < nr_node_ids; i++) { 6338 for (j = 0; j < nr_node_ids; j++) { 6339 for (k = 0; k < nr_node_ids; k++) { 6340 int distance = node_distance(i, k); 6341 6342 if (distance > curr_distance && 6343 (distance < next_distance || 6344 next_distance == curr_distance)) 6345 next_distance = distance; 6346 6347 /* 6348 * While not a strong assumption it would be nice to know 6349 * about cases where if node A is connected to B, B is not 6350 * equally connected to A. 6351 */ 6352 if (sched_debug() && node_distance(k, i) != distance) 6353 sched_numa_warn("Node-distance not symmetric"); 6354 6355 if (sched_debug() && i && !find_numa_distance(distance)) 6356 sched_numa_warn("Node-0 not representative"); 6357 } 6358 if (next_distance != curr_distance) { 6359 sched_domains_numa_distance[level++] = next_distance; 6360 sched_domains_numa_levels = level; 6361 curr_distance = next_distance; 6362 } else break; 6363 } 6364 6365 /* 6366 * In case of sched_debug() we verify the above assumption. 6367 */ 6368 if (!sched_debug()) 6369 break; 6370 } 6371 /* 6372 * 'level' contains the number of unique distances, excluding the 6373 * identity distance node_distance(i,i). 6374 * 6375 * The sched_domains_numa_distance[] array includes the actual distance 6376 * numbers. 6377 */ 6378 6379 /* 6380 * Here, we should temporarily reset sched_domains_numa_levels to 0. 6381 * If it fails to allocate memory for array sched_domains_numa_masks[][], 6382 * the array will contain less then 'level' members. This could be 6383 * dangerous when we use it to iterate array sched_domains_numa_masks[][] 6384 * in other functions. 6385 * 6386 * We reset it to 'level' at the end of this function. 6387 */ 6388 sched_domains_numa_levels = 0; 6389 6390 sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL); 6391 if (!sched_domains_numa_masks) 6392 return; 6393 6394 /* 6395 * Now for each level, construct a mask per node which contains all 6396 * cpus of nodes that are that many hops away from us. 6397 */ 6398 for (i = 0; i < level; i++) { 6399 sched_domains_numa_masks[i] = 6400 kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL); 6401 if (!sched_domains_numa_masks[i]) 6402 return; 6403 6404 for (j = 0; j < nr_node_ids; j++) { 6405 struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL); 6406 if (!mask) 6407 return; 6408 6409 sched_domains_numa_masks[i][j] = mask; 6410 6411 for (k = 0; k < nr_node_ids; k++) { 6412 if (node_distance(j, k) > sched_domains_numa_distance[i]) 6413 continue; 6414 6415 cpumask_or(mask, mask, cpumask_of_node(k)); 6416 } 6417 } 6418 } 6419 6420 /* Compute default topology size */ 6421 for (i = 0; sched_domain_topology[i].mask; i++); 6422 6423 tl = kzalloc((i + level + 1) * 6424 sizeof(struct sched_domain_topology_level), GFP_KERNEL); 6425 if (!tl) 6426 return; 6427 6428 /* 6429 * Copy the default topology bits.. 6430 */ 6431 for (i = 0; sched_domain_topology[i].mask; i++) 6432 tl[i] = sched_domain_topology[i]; 6433 6434 /* 6435 * .. and append 'j' levels of NUMA goodness. 6436 */ 6437 for (j = 0; j < level; i++, j++) { 6438 tl[i] = (struct sched_domain_topology_level){ 6439 .mask = sd_numa_mask, 6440 .sd_flags = cpu_numa_flags, 6441 .flags = SDTL_OVERLAP, 6442 .numa_level = j, 6443 SD_INIT_NAME(NUMA) 6444 }; 6445 } 6446 6447 sched_domain_topology = tl; 6448 6449 sched_domains_numa_levels = level; 6450} 6451 6452static void sched_domains_numa_masks_set(int cpu) 6453{ 6454 int i, j; 6455 int node = cpu_to_node(cpu); 6456 6457 for (i = 0; i < sched_domains_numa_levels; i++) { 6458 for (j = 0; j < nr_node_ids; j++) { 6459 if (node_distance(j, node) <= sched_domains_numa_distance[i]) 6460 cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]); 6461 } 6462 } 6463} 6464 6465static void sched_domains_numa_masks_clear(int cpu) 6466{ 6467 int i, j; 6468 for (i = 0; i < sched_domains_numa_levels; i++) { 6469 for (j = 0; j < nr_node_ids; j++) 6470 cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]); 6471 } 6472} 6473 6474/* 6475 * Update sched_domains_numa_masks[level][node] array when new cpus 6476 * are onlined. 6477 */ 6478static int sched_domains_numa_masks_update(struct notifier_block *nfb, 6479 unsigned long action, 6480 void *hcpu) 6481{ 6482 int cpu = (long)hcpu; 6483 6484 switch (action & ~CPU_TASKS_FROZEN) { 6485 case CPU_ONLINE: 6486 sched_domains_numa_masks_set(cpu); 6487 break; 6488 6489 case CPU_DEAD: 6490 sched_domains_numa_masks_clear(cpu); 6491 break; 6492 6493 default: 6494 return NOTIFY_DONE; 6495 } 6496 6497 return NOTIFY_OK; 6498} 6499#else 6500static inline void sched_init_numa(void) 6501{ 6502} 6503 6504static int sched_domains_numa_masks_update(struct notifier_block *nfb, 6505 unsigned long action, 6506 void *hcpu) 6507{ 6508 return 0; 6509} 6510#endif /* CONFIG_NUMA */ 6511 6512static int __sdt_alloc(const struct cpumask *cpu_map) 6513{ 6514 struct sched_domain_topology_level *tl; 6515 int j; 6516 6517 for_each_sd_topology(tl) { 6518 struct sd_data *sdd = &tl->data; 6519 6520 sdd->sd = alloc_percpu(struct sched_domain *); 6521 if (!sdd->sd) 6522 return -ENOMEM; 6523 6524 sdd->sg = alloc_percpu(struct sched_group *); 6525 if (!sdd->sg) 6526 return -ENOMEM; 6527 6528 sdd->sgc = alloc_percpu(struct sched_group_capacity *); 6529 if (!sdd->sgc) 6530 return -ENOMEM; 6531 6532 for_each_cpu(j, cpu_map) { 6533 struct sched_domain *sd; 6534 struct sched_group *sg; 6535 struct sched_group_capacity *sgc; 6536 6537 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(), 6538 GFP_KERNEL, cpu_to_node(j)); 6539 if (!sd) 6540 return -ENOMEM; 6541 6542 *per_cpu_ptr(sdd->sd, j) = sd; 6543 6544 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(), 6545 GFP_KERNEL, cpu_to_node(j)); 6546 if (!sg) 6547 return -ENOMEM; 6548 6549 sg->next = sg; 6550 6551 *per_cpu_ptr(sdd->sg, j) = sg; 6552 6553 sgc = kzalloc_node(sizeof(struct sched_group_capacity) + cpumask_size(), 6554 GFP_KERNEL, cpu_to_node(j)); 6555 if (!sgc) 6556 return -ENOMEM; 6557 6558 *per_cpu_ptr(sdd->sgc, j) = sgc; 6559 } 6560 } 6561 6562 return 0; 6563} 6564 6565static void __sdt_free(const struct cpumask *cpu_map) 6566{ 6567 struct sched_domain_topology_level *tl; 6568 int j; 6569 6570 for_each_sd_topology(tl) { 6571 struct sd_data *sdd = &tl->data; 6572 6573 for_each_cpu(j, cpu_map) { 6574 struct sched_domain *sd; 6575 6576 if (sdd->sd) { 6577 sd = *per_cpu_ptr(sdd->sd, j); 6578 if (sd && (sd->flags & SD_OVERLAP)) 6579 free_sched_groups(sd->groups, 0); 6580 kfree(*per_cpu_ptr(sdd->sd, j)); 6581 } 6582 6583 if (sdd->sg) 6584 kfree(*per_cpu_ptr(sdd->sg, j)); 6585 if (sdd->sgc) 6586 kfree(*per_cpu_ptr(sdd->sgc, j)); 6587 } 6588 free_percpu(sdd->sd); 6589 sdd->sd = NULL; 6590 free_percpu(sdd->sg); 6591 sdd->sg = NULL; 6592 free_percpu(sdd->sgc); 6593 sdd->sgc = NULL; 6594 } 6595} 6596 6597struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl, 6598 const struct cpumask *cpu_map, struct sched_domain_attr *attr, 6599 struct sched_domain *child, int cpu) 6600{ 6601 struct sched_domain *sd = sd_init(tl, cpu); 6602 if (!sd) 6603 return child; 6604 6605 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu)); 6606 if (child) { 6607 sd->level = child->level + 1; 6608 sched_domain_level_max = max(sched_domain_level_max, sd->level); 6609 child->parent = sd; 6610 sd->child = child; 6611 6612 if (!cpumask_subset(sched_domain_span(child), 6613 sched_domain_span(sd))) { 6614 pr_err("BUG: arch topology borken\n"); 6615#ifdef CONFIG_SCHED_DEBUG 6616 pr_err(" the %s domain not a subset of the %s domain\n", 6617 child->name, sd->name); 6618#endif 6619 /* Fixup, ensure @sd has at least @child cpus. */ 6620 cpumask_or(sched_domain_span(sd), 6621 sched_domain_span(sd), 6622 sched_domain_span(child)); 6623 } 6624 6625 } 6626 set_domain_attribute(sd, attr); 6627 6628 return sd; 6629} 6630 6631/* 6632 * Build sched domains for a given set of cpus and attach the sched domains 6633 * to the individual cpus 6634 */ 6635static int build_sched_domains(const struct cpumask *cpu_map, 6636 struct sched_domain_attr *attr) 6637{ 6638 enum s_alloc alloc_state; 6639 struct sched_domain *sd; 6640 struct s_data d; 6641 int i, ret = -ENOMEM; 6642 6643 alloc_state = __visit_domain_allocation_hell(&d, cpu_map); 6644 if (alloc_state != sa_rootdomain) 6645 goto error; 6646 6647 /* Set up domains for cpus specified by the cpu_map. */ 6648 for_each_cpu(i, cpu_map) { 6649 struct sched_domain_topology_level *tl; 6650 6651 sd = NULL; 6652 for_each_sd_topology(tl) { 6653 sd = build_sched_domain(tl, cpu_map, attr, sd, i); 6654 if (tl == sched_domain_topology) 6655 *per_cpu_ptr(d.sd, i) = sd; 6656 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP)) 6657 sd->flags |= SD_OVERLAP; 6658 if (cpumask_equal(cpu_map, sched_domain_span(sd))) 6659 break; 6660 } 6661 } 6662 6663 /* Build the groups for the domains */ 6664 for_each_cpu(i, cpu_map) { 6665 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) { 6666 sd->span_weight = cpumask_weight(sched_domain_span(sd)); 6667 if (sd->flags & SD_OVERLAP) { 6668 if (build_overlap_sched_groups(sd, i)) 6669 goto error; 6670 } else { 6671 if (build_sched_groups(sd, i)) 6672 goto error; 6673 } 6674 } 6675 } 6676 6677 /* Calculate CPU capacity for physical packages and nodes */ 6678 for (i = nr_cpumask_bits-1; i >= 0; i--) { 6679 if (!cpumask_test_cpu(i, cpu_map)) 6680 continue; 6681 6682 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) { 6683 claim_allocations(i, sd); 6684 init_sched_groups_capacity(i, sd); 6685 } 6686 } 6687 6688 /* Attach the domains */ 6689 rcu_read_lock(); 6690 for_each_cpu(i, cpu_map) { 6691 sd = *per_cpu_ptr(d.sd, i); 6692 cpu_attach_domain(sd, d.rd, i); 6693 } 6694 rcu_read_unlock(); 6695 6696 ret = 0; 6697error: 6698 __free_domain_allocs(&d, alloc_state, cpu_map); 6699 return ret; 6700} 6701 6702static cpumask_var_t *doms_cur; /* current sched domains */ 6703static int ndoms_cur; /* number of sched domains in 'doms_cur' */ 6704static struct sched_domain_attr *dattr_cur; 6705 /* attribues of custom domains in 'doms_cur' */ 6706 6707/* 6708 * Special case: If a kmalloc of a doms_cur partition (array of 6709 * cpumask) fails, then fallback to a single sched domain, 6710 * as determined by the single cpumask fallback_doms. 6711 */ 6712static cpumask_var_t fallback_doms; 6713 6714/* 6715 * arch_update_cpu_topology lets virtualized architectures update the 6716 * cpu core maps. It is supposed to return 1 if the topology changed 6717 * or 0 if it stayed the same. 6718 */ 6719int __weak arch_update_cpu_topology(void) 6720{ 6721 return 0; 6722} 6723 6724cpumask_var_t *alloc_sched_domains(unsigned int ndoms) 6725{ 6726 int i; 6727 cpumask_var_t *doms; 6728 6729 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL); 6730 if (!doms) 6731 return NULL; 6732 for (i = 0; i < ndoms; i++) { 6733 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) { 6734 free_sched_domains(doms, i); 6735 return NULL; 6736 } 6737 } 6738 return doms; 6739} 6740 6741void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms) 6742{ 6743 unsigned int i; 6744 for (i = 0; i < ndoms; i++) 6745 free_cpumask_var(doms[i]); 6746 kfree(doms); 6747} 6748 6749/* 6750 * Set up scheduler domains and groups. Callers must hold the hotplug lock. 6751 * For now this just excludes isolated cpus, but could be used to 6752 * exclude other special cases in the future. 6753 */ 6754static int init_sched_domains(const struct cpumask *cpu_map) 6755{ 6756 int err; 6757 6758 arch_update_cpu_topology(); 6759 ndoms_cur = 1; 6760 doms_cur = alloc_sched_domains(ndoms_cur); 6761 if (!doms_cur) 6762 doms_cur = &fallback_doms; 6763 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map); 6764 err = build_sched_domains(doms_cur[0], NULL); 6765 register_sched_domain_sysctl(); 6766 6767 return err; 6768} 6769 6770/* 6771 * Detach sched domains from a group of cpus specified in cpu_map 6772 * These cpus will now be attached to the NULL domain 6773 */ 6774static void detach_destroy_domains(const struct cpumask *cpu_map) 6775{ 6776 int i; 6777 6778 rcu_read_lock(); 6779 for_each_cpu(i, cpu_map) 6780 cpu_attach_domain(NULL, &def_root_domain, i); 6781 rcu_read_unlock(); 6782} 6783 6784/* handle null as "default" */ 6785static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur, 6786 struct sched_domain_attr *new, int idx_new) 6787{ 6788 struct sched_domain_attr tmp; 6789 6790 /* fast path */ 6791 if (!new && !cur) 6792 return 1; 6793 6794 tmp = SD_ATTR_INIT; 6795 return !memcmp(cur ? (cur + idx_cur) : &tmp, 6796 new ? (new + idx_new) : &tmp, 6797 sizeof(struct sched_domain_attr)); 6798} 6799 6800/* 6801 * Partition sched domains as specified by the 'ndoms_new' 6802 * cpumasks in the array doms_new[] of cpumasks. This compares 6803 * doms_new[] to the current sched domain partitioning, doms_cur[]. 6804 * It destroys each deleted domain and builds each new domain. 6805 * 6806 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'. 6807 * The masks don't intersect (don't overlap.) We should setup one 6808 * sched domain for each mask. CPUs not in any of the cpumasks will 6809 * not be load balanced. If the same cpumask appears both in the 6810 * current 'doms_cur' domains and in the new 'doms_new', we can leave 6811 * it as it is. 6812 * 6813 * The passed in 'doms_new' should be allocated using 6814 * alloc_sched_domains. This routine takes ownership of it and will 6815 * free_sched_domains it when done with it. If the caller failed the 6816 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1, 6817 * and partition_sched_domains() will fallback to the single partition 6818 * 'fallback_doms', it also forces the domains to be rebuilt. 6819 * 6820 * If doms_new == NULL it will be replaced with cpu_online_mask. 6821 * ndoms_new == 0 is a special case for destroying existing domains, 6822 * and it will not create the default domain. 6823 * 6824 * Call with hotplug lock held 6825 */ 6826void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[], 6827 struct sched_domain_attr *dattr_new) 6828{ 6829 int i, j, n; 6830 int new_topology; 6831 6832 mutex_lock(&sched_domains_mutex); 6833 6834 /* always unregister in case we don't destroy any domains */ 6835 unregister_sched_domain_sysctl(); 6836 6837 /* Let architecture update cpu core mappings. */ 6838 new_topology = arch_update_cpu_topology(); 6839 6840 n = doms_new ? ndoms_new : 0; 6841 6842 /* Destroy deleted domains */ 6843 for (i = 0; i < ndoms_cur; i++) { 6844 for (j = 0; j < n && !new_topology; j++) { 6845 if (cpumask_equal(doms_cur[i], doms_new[j]) 6846 && dattrs_equal(dattr_cur, i, dattr_new, j)) 6847 goto match1; 6848 } 6849 /* no match - a current sched domain not in new doms_new[] */ 6850 detach_destroy_domains(doms_cur[i]); 6851match1: 6852 ; 6853 } 6854 6855 n = ndoms_cur; 6856 if (doms_new == NULL) { 6857 n = 0; 6858 doms_new = &fallback_doms; 6859 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map); 6860 WARN_ON_ONCE(dattr_new); 6861 } 6862 6863 /* Build new domains */ 6864 for (i = 0; i < ndoms_new; i++) { 6865 for (j = 0; j < n && !new_topology; j++) { 6866 if (cpumask_equal(doms_new[i], doms_cur[j]) 6867 && dattrs_equal(dattr_new, i, dattr_cur, j)) 6868 goto match2; 6869 } 6870 /* no match - add a new doms_new */ 6871 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL); 6872match2: 6873 ; 6874 } 6875 6876 /* Remember the new sched domains */ 6877 if (doms_cur != &fallback_doms) 6878 free_sched_domains(doms_cur, ndoms_cur); 6879 kfree(dattr_cur); /* kfree(NULL) is safe */ 6880 doms_cur = doms_new; 6881 dattr_cur = dattr_new; 6882 ndoms_cur = ndoms_new; 6883 6884 register_sched_domain_sysctl(); 6885 6886 mutex_unlock(&sched_domains_mutex); 6887} 6888 6889static int num_cpus_frozen; /* used to mark begin/end of suspend/resume */ 6890 6891/* 6892 * Update cpusets according to cpu_active mask. If cpusets are 6893 * disabled, cpuset_update_active_cpus() becomes a simple wrapper 6894 * around partition_sched_domains(). 6895 * 6896 * If we come here as part of a suspend/resume, don't touch cpusets because we 6897 * want to restore it back to its original state upon resume anyway. 6898 */ 6899static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action, 6900 void *hcpu) 6901{ 6902 switch (action) { 6903 case CPU_ONLINE_FROZEN: 6904 case CPU_DOWN_FAILED_FROZEN: 6905 6906 /* 6907 * num_cpus_frozen tracks how many CPUs are involved in suspend 6908 * resume sequence. As long as this is not the last online 6909 * operation in the resume sequence, just build a single sched 6910 * domain, ignoring cpusets. 6911 */ 6912 num_cpus_frozen--; 6913 if (likely(num_cpus_frozen)) { 6914 partition_sched_domains(1, NULL, NULL); 6915 break; 6916 } 6917 6918 /* 6919 * This is the last CPU online operation. So fall through and 6920 * restore the original sched domains by considering the 6921 * cpuset configurations. 6922 */ 6923 6924 case CPU_ONLINE: 6925 case CPU_DOWN_FAILED: 6926 cpuset_update_active_cpus(true); 6927 break; 6928 default: 6929 return NOTIFY_DONE; 6930 } 6931 return NOTIFY_OK; 6932} 6933 6934static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action, 6935 void *hcpu) 6936{ 6937 switch (action) { 6938 case CPU_DOWN_PREPARE: 6939 cpuset_update_active_cpus(false); 6940 break; 6941 case CPU_DOWN_PREPARE_FROZEN: 6942 num_cpus_frozen++; 6943 partition_sched_domains(1, NULL, NULL); 6944 break; 6945 default: 6946 return NOTIFY_DONE; 6947 } 6948 return NOTIFY_OK; 6949} 6950 6951void __init sched_init_smp(void) 6952{ 6953 cpumask_var_t non_isolated_cpus; 6954 6955 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL); 6956 alloc_cpumask_var(&fallback_doms, GFP_KERNEL); 6957 6958 sched_init_numa(); 6959 6960 /* 6961 * There's no userspace yet to cause hotplug operations; hence all the 6962 * cpu masks are stable and all blatant races in the below code cannot 6963 * happen. 6964 */ 6965 mutex_lock(&sched_domains_mutex); 6966 init_sched_domains(cpu_active_mask); 6967 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map); 6968 if (cpumask_empty(non_isolated_cpus)) 6969 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus); 6970 mutex_unlock(&sched_domains_mutex); 6971 6972 hotcpu_notifier(sched_domains_numa_masks_update, CPU_PRI_SCHED_ACTIVE); 6973 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE); 6974 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE); 6975 6976 init_hrtick(); 6977 6978 /* Move init over to a non-isolated CPU */ 6979 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0) 6980 BUG(); 6981 sched_init_granularity(); 6982 free_cpumask_var(non_isolated_cpus); 6983 6984 init_sched_rt_class(); 6985 init_sched_dl_class(); 6986} 6987#else 6988void __init sched_init_smp(void) 6989{ 6990 sched_init_granularity(); 6991} 6992#endif /* CONFIG_SMP */ 6993 6994const_debug unsigned int sysctl_timer_migration = 1; 6995 6996int in_sched_functions(unsigned long addr) 6997{ 6998 return in_lock_functions(addr) || 6999 (addr >= (unsigned long)__sched_text_start 7000 && addr < (unsigned long)__sched_text_end); 7001} 7002 7003#ifdef CONFIG_CGROUP_SCHED 7004/* 7005 * Default task group. 7006 * Every task in system belongs to this group at bootup. 7007 */ 7008struct task_group root_task_group; 7009LIST_HEAD(task_groups); 7010#endif 7011 7012DECLARE_PER_CPU(cpumask_var_t, load_balance_mask); 7013 7014void __init sched_init(void) 7015{ 7016 int i, j; 7017 unsigned long alloc_size = 0, ptr; 7018 7019#ifdef CONFIG_FAIR_GROUP_SCHED 7020 alloc_size += 2 * nr_cpu_ids * sizeof(void **); 7021#endif 7022#ifdef CONFIG_RT_GROUP_SCHED 7023 alloc_size += 2 * nr_cpu_ids * sizeof(void **); 7024#endif 7025#ifdef CONFIG_CPUMASK_OFFSTACK 7026 alloc_size += num_possible_cpus() * cpumask_size(); 7027#endif 7028 if (alloc_size) { 7029 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT); 7030 7031#ifdef CONFIG_FAIR_GROUP_SCHED 7032 root_task_group.se = (struct sched_entity **)ptr; 7033 ptr += nr_cpu_ids * sizeof(void **); 7034 7035 root_task_group.cfs_rq = (struct cfs_rq **)ptr; 7036 ptr += nr_cpu_ids * sizeof(void **); 7037 7038#endif /* CONFIG_FAIR_GROUP_SCHED */ 7039#ifdef CONFIG_RT_GROUP_SCHED 7040 root_task_group.rt_se = (struct sched_rt_entity **)ptr; 7041 ptr += nr_cpu_ids * sizeof(void **); 7042 7043 root_task_group.rt_rq = (struct rt_rq **)ptr; 7044 ptr += nr_cpu_ids * sizeof(void **); 7045 7046#endif /* CONFIG_RT_GROUP_SCHED */ 7047#ifdef CONFIG_CPUMASK_OFFSTACK 7048 for_each_possible_cpu(i) { 7049 per_cpu(load_balance_mask, i) = (void *)ptr; 7050 ptr += cpumask_size(); 7051 } 7052#endif /* CONFIG_CPUMASK_OFFSTACK */ 7053 } 7054 7055 init_rt_bandwidth(&def_rt_bandwidth, 7056 global_rt_period(), global_rt_runtime()); 7057 init_dl_bandwidth(&def_dl_bandwidth, 7058 global_rt_period(), global_rt_runtime()); 7059 7060#ifdef CONFIG_SMP 7061 init_defrootdomain(); 7062#endif 7063 7064#ifdef CONFIG_RT_GROUP_SCHED 7065 init_rt_bandwidth(&root_task_group.rt_bandwidth, 7066 global_rt_period(), global_rt_runtime()); 7067#endif /* CONFIG_RT_GROUP_SCHED */ 7068 7069#ifdef CONFIG_CGROUP_SCHED 7070 list_add(&root_task_group.list, &task_groups); 7071 INIT_LIST_HEAD(&root_task_group.children); 7072 INIT_LIST_HEAD(&root_task_group.siblings); 7073 autogroup_init(&init_task); 7074 7075#endif /* CONFIG_CGROUP_SCHED */ 7076 7077 for_each_possible_cpu(i) { 7078 struct rq *rq; 7079 7080 rq = cpu_rq(i); 7081 raw_spin_lock_init(&rq->lock); 7082 rq->nr_running = 0; 7083 rq->calc_load_active = 0; 7084 rq->calc_load_update = jiffies + LOAD_FREQ; 7085 init_cfs_rq(&rq->cfs); 7086 init_rt_rq(&rq->rt, rq); 7087 init_dl_rq(&rq->dl, rq); 7088#ifdef CONFIG_FAIR_GROUP_SCHED 7089 root_task_group.shares = ROOT_TASK_GROUP_LOAD; 7090 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list); 7091 /* 7092 * How much cpu bandwidth does root_task_group get? 7093 * 7094 * In case of task-groups formed thr' the cgroup filesystem, it 7095 * gets 100% of the cpu resources in the system. This overall 7096 * system cpu resource is divided among the tasks of 7097 * root_task_group and its child task-groups in a fair manner, 7098 * based on each entity's (task or task-group's) weight 7099 * (se->load.weight). 7100 * 7101 * In other words, if root_task_group has 10 tasks of weight 7102 * 1024) and two child groups A0 and A1 (of weight 1024 each), 7103 * then A0's share of the cpu resource is: 7104 * 7105 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33% 7106 * 7107 * We achieve this by letting root_task_group's tasks sit 7108 * directly in rq->cfs (i.e root_task_group->se[] = NULL). 7109 */ 7110 init_cfs_bandwidth(&root_task_group.cfs_bandwidth); 7111 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL); 7112#endif /* CONFIG_FAIR_GROUP_SCHED */ 7113 7114 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime; 7115#ifdef CONFIG_RT_GROUP_SCHED 7116 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL); 7117#endif 7118 7119 for (j = 0; j < CPU_LOAD_IDX_MAX; j++) 7120 rq->cpu_load[j] = 0; 7121 7122 rq->last_load_update_tick = jiffies; 7123 7124#ifdef CONFIG_SMP 7125 rq->sd = NULL; 7126 rq->rd = NULL; 7127 rq->cpu_capacity = SCHED_CAPACITY_SCALE; 7128 rq->post_schedule = 0; 7129 rq->active_balance = 0; 7130 rq->next_balance = jiffies; 7131 rq->push_cpu = 0; 7132 rq->cpu = i; 7133 rq->online = 0; 7134 rq->idle_stamp = 0; 7135 rq->avg_idle = 2*sysctl_sched_migration_cost; 7136 rq->max_idle_balance_cost = sysctl_sched_migration_cost; 7137 7138 INIT_LIST_HEAD(&rq->cfs_tasks); 7139 7140 rq_attach_root(rq, &def_root_domain); 7141#ifdef CONFIG_NO_HZ_COMMON 7142 rq->nohz_flags = 0; 7143#endif 7144#ifdef CONFIG_NO_HZ_FULL 7145 rq->last_sched_tick = 0; 7146#endif 7147#endif 7148 init_rq_hrtick(rq); 7149 atomic_set(&rq->nr_iowait, 0); 7150 } 7151 7152 set_load_weight(&init_task); 7153 7154#ifdef CONFIG_PREEMPT_NOTIFIERS 7155 INIT_HLIST_HEAD(&init_task.preempt_notifiers); 7156#endif 7157 7158 /* 7159 * The boot idle thread does lazy MMU switching as well: 7160 */ 7161 atomic_inc(&init_mm.mm_count); 7162 enter_lazy_tlb(&init_mm, current); 7163 7164 /* 7165 * Make us the idle thread. Technically, schedule() should not be 7166 * called from this thread, however somewhere below it might be, 7167 * but because we are the idle thread, we just pick up running again 7168 * when this runqueue becomes "idle". 7169 */ 7170 init_idle(current, smp_processor_id()); 7171 7172 calc_load_update = jiffies + LOAD_FREQ; 7173 7174 /* 7175 * During early bootup we pretend to be a normal task: 7176 */ 7177 current->sched_class = &fair_sched_class; 7178 7179#ifdef CONFIG_SMP 7180 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT); 7181 /* May be allocated at isolcpus cmdline parse time */ 7182 if (cpu_isolated_map == NULL) 7183 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT); 7184 idle_thread_set_boot_cpu(); 7185 set_cpu_rq_start_time(); 7186#endif 7187 init_sched_fair_class(); 7188 7189 scheduler_running = 1; 7190} 7191 7192#ifdef CONFIG_DEBUG_ATOMIC_SLEEP 7193static inline int preempt_count_equals(int preempt_offset) 7194{ 7195 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth(); 7196 7197 return (nested == preempt_offset); 7198} 7199 7200void __might_sleep(const char *file, int line, int preempt_offset) 7201{ 7202 static unsigned long prev_jiffy; /* ratelimiting */ 7203 7204 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */ 7205 if ((preempt_count_equals(preempt_offset) && !irqs_disabled() && 7206 !is_idle_task(current)) || 7207 system_state != SYSTEM_RUNNING || oops_in_progress) 7208 return; 7209 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy) 7210 return; 7211 prev_jiffy = jiffies; 7212 7213 printk(KERN_ERR 7214 "BUG: sleeping function called from invalid context at %s:%d\n", 7215 file, line); 7216 printk(KERN_ERR 7217 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n", 7218 in_atomic(), irqs_disabled(), 7219 current->pid, current->comm); 7220 7221 debug_show_held_locks(current); 7222 if (irqs_disabled()) 7223 print_irqtrace_events(current); 7224#ifdef CONFIG_DEBUG_PREEMPT 7225 if (!preempt_count_equals(preempt_offset)) { 7226 pr_err("Preemption disabled at:"); 7227 print_ip_sym(current->preempt_disable_ip); 7228 pr_cont("\n"); 7229 } 7230#endif 7231 dump_stack(); 7232} 7233EXPORT_SYMBOL(__might_sleep); 7234#endif 7235 7236#ifdef CONFIG_MAGIC_SYSRQ 7237static void normalize_task(struct rq *rq, struct task_struct *p) 7238{ 7239 const struct sched_class *prev_class = p->sched_class; 7240 struct sched_attr attr = { 7241 .sched_policy = SCHED_NORMAL, 7242 }; 7243 int old_prio = p->prio; 7244 int queued; 7245 7246 queued = task_on_rq_queued(p); 7247 if (queued) 7248 dequeue_task(rq, p, 0); 7249 __setscheduler(rq, p, &attr); 7250 if (queued) { 7251 enqueue_task(rq, p, 0); 7252 resched_curr(rq); 7253 } 7254 7255 check_class_changed(rq, p, prev_class, old_prio); 7256} 7257 7258void normalize_rt_tasks(void) 7259{ 7260 struct task_struct *g, *p; 7261 unsigned long flags; 7262 struct rq *rq; 7263 7264 read_lock(&tasklist_lock); 7265 for_each_process_thread(g, p) { 7266 /* 7267 * Only normalize user tasks: 7268 */ 7269 if (p->flags & PF_KTHREAD) 7270 continue; 7271 7272 p->se.exec_start = 0; 7273#ifdef CONFIG_SCHEDSTATS 7274 p->se.statistics.wait_start = 0; 7275 p->se.statistics.sleep_start = 0; 7276 p->se.statistics.block_start = 0; 7277#endif 7278 7279 if (!dl_task(p) && !rt_task(p)) { 7280 /* 7281 * Renice negative nice level userspace 7282 * tasks back to 0: 7283 */ 7284 if (task_nice(p) < 0) 7285 set_user_nice(p, 0); 7286 continue; 7287 } 7288 7289 rq = task_rq_lock(p, &flags); 7290 normalize_task(rq, p); 7291 task_rq_unlock(rq, p, &flags); 7292 } 7293 read_unlock(&tasklist_lock); 7294} 7295 7296#endif /* CONFIG_MAGIC_SYSRQ */ 7297 7298#if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) 7299/* 7300 * These functions are only useful for the IA64 MCA handling, or kdb. 7301 * 7302 * They can only be called when the whole system has been 7303 * stopped - every CPU needs to be quiescent, and no scheduling 7304 * activity can take place. Using them for anything else would 7305 * be a serious bug, and as a result, they aren't even visible 7306 * under any other configuration. 7307 */ 7308 7309/** 7310 * curr_task - return the current task for a given cpu. 7311 * @cpu: the processor in question. 7312 * 7313 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED! 7314 * 7315 * Return: The current task for @cpu. 7316 */ 7317struct task_struct *curr_task(int cpu) 7318{ 7319 return cpu_curr(cpu); 7320} 7321 7322#endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */ 7323 7324#ifdef CONFIG_IA64 7325/** 7326 * set_curr_task - set the current task for a given cpu. 7327 * @cpu: the processor in question. 7328 * @p: the task pointer to set. 7329 * 7330 * Description: This function must only be used when non-maskable interrupts 7331 * are serviced on a separate stack. It allows the architecture to switch the 7332 * notion of the current task on a cpu in a non-blocking manner. This function 7333 * must be called with all CPU's synchronized, and interrupts disabled, the 7334 * and caller must save the original value of the current task (see 7335 * curr_task() above) and restore that value before reenabling interrupts and 7336 * re-starting the system. 7337 * 7338 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED! 7339 */ 7340void set_curr_task(int cpu, struct task_struct *p) 7341{ 7342 cpu_curr(cpu) = p; 7343} 7344 7345#endif 7346 7347#ifdef CONFIG_CGROUP_SCHED 7348/* task_group_lock serializes the addition/removal of task groups */ 7349static DEFINE_SPINLOCK(task_group_lock); 7350 7351static void free_sched_group(struct task_group *tg) 7352{ 7353 free_fair_sched_group(tg); 7354 free_rt_sched_group(tg); 7355 autogroup_free(tg); 7356 kfree(tg); 7357} 7358 7359/* allocate runqueue etc for a new task group */ 7360struct task_group *sched_create_group(struct task_group *parent) 7361{ 7362 struct task_group *tg; 7363 7364 tg = kzalloc(sizeof(*tg), GFP_KERNEL); 7365 if (!tg) 7366 return ERR_PTR(-ENOMEM); 7367 7368 if (!alloc_fair_sched_group(tg, parent)) 7369 goto err; 7370 7371 if (!alloc_rt_sched_group(tg, parent)) 7372 goto err; 7373 7374 return tg; 7375 7376err: 7377 free_sched_group(tg); 7378 return ERR_PTR(-ENOMEM); 7379} 7380 7381void sched_online_group(struct task_group *tg, struct task_group *parent) 7382{ 7383 unsigned long flags; 7384 7385 spin_lock_irqsave(&task_group_lock, flags); 7386 list_add_rcu(&tg->list, &task_groups); 7387 7388 WARN_ON(!parent); /* root should already exist */ 7389 7390 tg->parent = parent; 7391 INIT_LIST_HEAD(&tg->children); 7392 list_add_rcu(&tg->siblings, &parent->children); 7393 spin_unlock_irqrestore(&task_group_lock, flags); 7394} 7395 7396/* rcu callback to free various structures associated with a task group */ 7397static void free_sched_group_rcu(struct rcu_head *rhp) 7398{ 7399 /* now it should be safe to free those cfs_rqs */ 7400 free_sched_group(container_of(rhp, struct task_group, rcu)); 7401} 7402 7403/* Destroy runqueue etc associated with a task group */ 7404void sched_destroy_group(struct task_group *tg) 7405{ 7406 /* wait for possible concurrent references to cfs_rqs complete */ 7407 call_rcu(&tg->rcu, free_sched_group_rcu); 7408} 7409 7410void sched_offline_group(struct task_group *tg) 7411{ 7412 unsigned long flags; 7413 int i; 7414 7415 /* end participation in shares distribution */ 7416 for_each_possible_cpu(i) 7417 unregister_fair_sched_group(tg, i); 7418 7419 spin_lock_irqsave(&task_group_lock, flags); 7420 list_del_rcu(&tg->list); 7421 list_del_rcu(&tg->siblings); 7422 spin_unlock_irqrestore(&task_group_lock, flags); 7423} 7424 7425/* change task's runqueue when it moves between groups. 7426 * The caller of this function should have put the task in its new group 7427 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to 7428 * reflect its new group. 7429 */ 7430void sched_move_task(struct task_struct *tsk) 7431{ 7432 struct task_group *tg; 7433 int queued, running; 7434 unsigned long flags; 7435 struct rq *rq; 7436 7437 rq = task_rq_lock(tsk, &flags); 7438 7439 running = task_current(rq, tsk); 7440 queued = task_on_rq_queued(tsk); 7441 7442 if (queued) 7443 dequeue_task(rq, tsk, 0); 7444 if (unlikely(running)) 7445 put_prev_task(rq, tsk); 7446 7447 /* 7448 * All callers are synchronized by task_rq_lock(); we do not use RCU 7449 * which is pointless here. Thus, we pass "true" to task_css_check() 7450 * to prevent lockdep warnings. 7451 */ 7452 tg = container_of(task_css_check(tsk, cpu_cgrp_id, true), 7453 struct task_group, css); 7454 tg = autogroup_task_group(tsk, tg); 7455 tsk->sched_task_group = tg; 7456 7457#ifdef CONFIG_FAIR_GROUP_SCHED 7458 if (tsk->sched_class->task_move_group) 7459 tsk->sched_class->task_move_group(tsk, queued); 7460 else 7461#endif 7462 set_task_rq(tsk, task_cpu(tsk)); 7463 7464 if (unlikely(running)) 7465 tsk->sched_class->set_curr_task(rq); 7466 if (queued) 7467 enqueue_task(rq, tsk, 0); 7468 7469 task_rq_unlock(rq, tsk, &flags); 7470} 7471#endif /* CONFIG_CGROUP_SCHED */ 7472 7473#ifdef CONFIG_RT_GROUP_SCHED 7474/* 7475 * Ensure that the real time constraints are schedulable. 7476 */ 7477static DEFINE_MUTEX(rt_constraints_mutex); 7478 7479/* Must be called with tasklist_lock held */ 7480static inline int tg_has_rt_tasks(struct task_group *tg) 7481{ 7482 struct task_struct *g, *p; 7483 7484 for_each_process_thread(g, p) { 7485 if (rt_task(p) && task_group(p) == tg) 7486 return 1; 7487 } 7488 7489 return 0; 7490} 7491 7492struct rt_schedulable_data { 7493 struct task_group *tg; 7494 u64 rt_period; 7495 u64 rt_runtime; 7496}; 7497 7498static int tg_rt_schedulable(struct task_group *tg, void *data) 7499{ 7500 struct rt_schedulable_data *d = data; 7501 struct task_group *child; 7502 unsigned long total, sum = 0; 7503 u64 period, runtime; 7504 7505 period = ktime_to_ns(tg->rt_bandwidth.rt_period); 7506 runtime = tg->rt_bandwidth.rt_runtime; 7507 7508 if (tg == d->tg) { 7509 period = d->rt_period; 7510 runtime = d->rt_runtime; 7511 } 7512 7513 /* 7514 * Cannot have more runtime than the period. 7515 */ 7516 if (runtime > period && runtime != RUNTIME_INF) 7517 return -EINVAL; 7518 7519 /* 7520 * Ensure we don't starve existing RT tasks. 7521 */ 7522 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg)) 7523 return -EBUSY; 7524 7525 total = to_ratio(period, runtime); 7526 7527 /* 7528 * Nobody can have more than the global setting allows. 7529 */ 7530 if (total > to_ratio(global_rt_period(), global_rt_runtime())) 7531 return -EINVAL; 7532 7533 /* 7534 * The sum of our children's runtime should not exceed our own. 7535 */ 7536 list_for_each_entry_rcu(child, &tg->children, siblings) { 7537 period = ktime_to_ns(child->rt_bandwidth.rt_period); 7538 runtime = child->rt_bandwidth.rt_runtime; 7539 7540 if (child == d->tg) { 7541 period = d->rt_period; 7542 runtime = d->rt_runtime; 7543 } 7544 7545 sum += to_ratio(period, runtime); 7546 } 7547 7548 if (sum > total) 7549 return -EINVAL; 7550 7551 return 0; 7552} 7553 7554static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime) 7555{ 7556 int ret; 7557 7558 struct rt_schedulable_data data = { 7559 .tg = tg, 7560 .rt_period = period, 7561 .rt_runtime = runtime, 7562 }; 7563 7564 rcu_read_lock(); 7565 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data); 7566 rcu_read_unlock(); 7567 7568 return ret; 7569} 7570 7571static int tg_set_rt_bandwidth(struct task_group *tg, 7572 u64 rt_period, u64 rt_runtime) 7573{ 7574 int i, err = 0; 7575 7576 mutex_lock(&rt_constraints_mutex); 7577 read_lock(&tasklist_lock); 7578 err = __rt_schedulable(tg, rt_period, rt_runtime); 7579 if (err) 7580 goto unlock; 7581 7582 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock); 7583 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period); 7584 tg->rt_bandwidth.rt_runtime = rt_runtime; 7585 7586 for_each_possible_cpu(i) { 7587 struct rt_rq *rt_rq = tg->rt_rq[i]; 7588 7589 raw_spin_lock(&rt_rq->rt_runtime_lock); 7590 rt_rq->rt_runtime = rt_runtime; 7591 raw_spin_unlock(&rt_rq->rt_runtime_lock); 7592 } 7593 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock); 7594unlock: 7595 read_unlock(&tasklist_lock); 7596 mutex_unlock(&rt_constraints_mutex); 7597 7598 return err; 7599} 7600 7601static int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us) 7602{ 7603 u64 rt_runtime, rt_period; 7604 7605 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period); 7606 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC; 7607 if (rt_runtime_us < 0) 7608 rt_runtime = RUNTIME_INF; 7609 7610 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime); 7611} 7612 7613static long sched_group_rt_runtime(struct task_group *tg) 7614{ 7615 u64 rt_runtime_us; 7616 7617 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF) 7618 return -1; 7619 7620 rt_runtime_us = tg->rt_bandwidth.rt_runtime; 7621 do_div(rt_runtime_us, NSEC_PER_USEC); 7622 return rt_runtime_us; 7623} 7624 7625static int sched_group_set_rt_period(struct task_group *tg, long rt_period_us) 7626{ 7627 u64 rt_runtime, rt_period; 7628 7629 rt_period = (u64)rt_period_us * NSEC_PER_USEC; 7630 rt_runtime = tg->rt_bandwidth.rt_runtime; 7631 7632 if (rt_period == 0) 7633 return -EINVAL; 7634 7635 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime); 7636} 7637 7638static long sched_group_rt_period(struct task_group *tg) 7639{ 7640 u64 rt_period_us; 7641 7642 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period); 7643 do_div(rt_period_us, NSEC_PER_USEC); 7644 return rt_period_us; 7645} 7646#endif /* CONFIG_RT_GROUP_SCHED */ 7647 7648#ifdef CONFIG_RT_GROUP_SCHED 7649static int sched_rt_global_constraints(void) 7650{ 7651 int ret = 0; 7652 7653 mutex_lock(&rt_constraints_mutex); 7654 read_lock(&tasklist_lock); 7655 ret = __rt_schedulable(NULL, 0, 0); 7656 read_unlock(&tasklist_lock); 7657 mutex_unlock(&rt_constraints_mutex); 7658 7659 return ret; 7660} 7661 7662static int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk) 7663{ 7664 /* Don't accept realtime tasks when there is no way for them to run */ 7665 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0) 7666 return 0; 7667 7668 return 1; 7669} 7670 7671#else /* !CONFIG_RT_GROUP_SCHED */ 7672static int sched_rt_global_constraints(void) 7673{ 7674 unsigned long flags; 7675 int i, ret = 0; 7676 7677 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags); 7678 for_each_possible_cpu(i) { 7679 struct rt_rq *rt_rq = &cpu_rq(i)->rt; 7680 7681 raw_spin_lock(&rt_rq->rt_runtime_lock); 7682 rt_rq->rt_runtime = global_rt_runtime(); 7683 raw_spin_unlock(&rt_rq->rt_runtime_lock); 7684 } 7685 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags); 7686 7687 return ret; 7688} 7689#endif /* CONFIG_RT_GROUP_SCHED */ 7690 7691static int sched_dl_global_constraints(void) 7692{ 7693 u64 runtime = global_rt_runtime(); 7694 u64 period = global_rt_period(); 7695 u64 new_bw = to_ratio(period, runtime); 7696 struct dl_bw *dl_b; 7697 int cpu, ret = 0; 7698 unsigned long flags; 7699 7700 /* 7701 * Here we want to check the bandwidth not being set to some 7702 * value smaller than the currently allocated bandwidth in 7703 * any of the root_domains. 7704 * 7705 * FIXME: Cycling on all the CPUs is overdoing, but simpler than 7706 * cycling on root_domains... Discussion on different/better 7707 * solutions is welcome! 7708 */ 7709 for_each_possible_cpu(cpu) { 7710 rcu_read_lock_sched(); 7711 dl_b = dl_bw_of(cpu); 7712 7713 raw_spin_lock_irqsave(&dl_b->lock, flags); 7714 if (new_bw < dl_b->total_bw) 7715 ret = -EBUSY; 7716 raw_spin_unlock_irqrestore(&dl_b->lock, flags); 7717 7718 rcu_read_unlock_sched(); 7719 7720 if (ret) 7721 break; 7722 } 7723 7724 return ret; 7725} 7726 7727static void sched_dl_do_global(void) 7728{ 7729 u64 new_bw = -1; 7730 struct dl_bw *dl_b; 7731 int cpu; 7732 unsigned long flags; 7733 7734 def_dl_bandwidth.dl_period = global_rt_period(); 7735 def_dl_bandwidth.dl_runtime = global_rt_runtime(); 7736 7737 if (global_rt_runtime() != RUNTIME_INF) 7738 new_bw = to_ratio(global_rt_period(), global_rt_runtime()); 7739 7740 /* 7741 * FIXME: As above... 7742 */ 7743 for_each_possible_cpu(cpu) { 7744 rcu_read_lock_sched(); 7745 dl_b = dl_bw_of(cpu); 7746 7747 raw_spin_lock_irqsave(&dl_b->lock, flags); 7748 dl_b->bw = new_bw; 7749 raw_spin_unlock_irqrestore(&dl_b->lock, flags); 7750 7751 rcu_read_unlock_sched(); 7752 } 7753} 7754 7755static int sched_rt_global_validate(void) 7756{ 7757 if (sysctl_sched_rt_period <= 0) 7758 return -EINVAL; 7759 7760 if ((sysctl_sched_rt_runtime != RUNTIME_INF) && 7761 (sysctl_sched_rt_runtime > sysctl_sched_rt_period)) 7762 return -EINVAL; 7763 7764 return 0; 7765} 7766 7767static void sched_rt_do_global(void) 7768{ 7769 def_rt_bandwidth.rt_runtime = global_rt_runtime(); 7770 def_rt_bandwidth.rt_period = ns_to_ktime(global_rt_period()); 7771} 7772 7773int sched_rt_handler(struct ctl_table *table, int write, 7774 void __user *buffer, size_t *lenp, 7775 loff_t *ppos) 7776{ 7777 int old_period, old_runtime; 7778 static DEFINE_MUTEX(mutex); 7779 int ret; 7780 7781 mutex_lock(&mutex); 7782 old_period = sysctl_sched_rt_period; 7783 old_runtime = sysctl_sched_rt_runtime; 7784 7785 ret = proc_dointvec(table, write, buffer, lenp, ppos); 7786 7787 if (!ret && write) { 7788 ret = sched_rt_global_validate(); 7789 if (ret) 7790 goto undo; 7791 7792 ret = sched_rt_global_constraints(); 7793 if (ret) 7794 goto undo; 7795 7796 ret = sched_dl_global_constraints(); 7797 if (ret) 7798 goto undo; 7799 7800 sched_rt_do_global(); 7801 sched_dl_do_global(); 7802 } 7803 if (0) { 7804undo: 7805 sysctl_sched_rt_period = old_period; 7806 sysctl_sched_rt_runtime = old_runtime; 7807 } 7808 mutex_unlock(&mutex); 7809 7810 return ret; 7811} 7812 7813int sched_rr_handler(struct ctl_table *table, int write, 7814 void __user *buffer, size_t *lenp, 7815 loff_t *ppos) 7816{ 7817 int ret; 7818 static DEFINE_MUTEX(mutex); 7819 7820 mutex_lock(&mutex); 7821 ret = proc_dointvec(table, write, buffer, lenp, ppos); 7822 /* make sure that internally we keep jiffies */ 7823 /* also, writing zero resets timeslice to default */ 7824 if (!ret && write) { 7825 sched_rr_timeslice = sched_rr_timeslice <= 0 ? 7826 RR_TIMESLICE : msecs_to_jiffies(sched_rr_timeslice); 7827 } 7828 mutex_unlock(&mutex); 7829 return ret; 7830} 7831 7832#ifdef CONFIG_CGROUP_SCHED 7833 7834static inline struct task_group *css_tg(struct cgroup_subsys_state *css) 7835{ 7836 return css ? container_of(css, struct task_group, css) : NULL; 7837} 7838 7839static struct cgroup_subsys_state * 7840cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css) 7841{ 7842 struct task_group *parent = css_tg(parent_css); 7843 struct task_group *tg; 7844 7845 if (!parent) { 7846 /* This is early initialization for the top cgroup */ 7847 return &root_task_group.css; 7848 } 7849 7850 tg = sched_create_group(parent); 7851 if (IS_ERR(tg)) 7852 return ERR_PTR(-ENOMEM); 7853 7854 return &tg->css; 7855} 7856 7857static int cpu_cgroup_css_online(struct cgroup_subsys_state *css) 7858{ 7859 struct task_group *tg = css_tg(css); 7860 struct task_group *parent = css_tg(css->parent); 7861 7862 if (parent) 7863 sched_online_group(tg, parent); 7864 return 0; 7865} 7866 7867static void cpu_cgroup_css_free(struct cgroup_subsys_state *css) 7868{ 7869 struct task_group *tg = css_tg(css); 7870 7871 sched_destroy_group(tg); 7872} 7873 7874static void cpu_cgroup_css_offline(struct cgroup_subsys_state *css) 7875{ 7876 struct task_group *tg = css_tg(css); 7877 7878 sched_offline_group(tg); 7879} 7880 7881static void cpu_cgroup_fork(struct task_struct *task) 7882{ 7883 sched_move_task(task); 7884} 7885 7886static int cpu_cgroup_can_attach(struct cgroup_subsys_state *css, 7887 struct cgroup_taskset *tset) 7888{ 7889 struct task_struct *task; 7890 7891 cgroup_taskset_for_each(task, tset) { 7892#ifdef CONFIG_RT_GROUP_SCHED 7893 if (!sched_rt_can_attach(css_tg(css), task)) 7894 return -EINVAL; 7895#else 7896 /* We don't support RT-tasks being in separate groups */ 7897 if (task->sched_class != &fair_sched_class) 7898 return -EINVAL; 7899#endif 7900 } 7901 return 0; 7902} 7903 7904static void cpu_cgroup_attach(struct cgroup_subsys_state *css, 7905 struct cgroup_taskset *tset) 7906{ 7907 struct task_struct *task; 7908 7909 cgroup_taskset_for_each(task, tset) 7910 sched_move_task(task); 7911} 7912 7913static void cpu_cgroup_exit(struct cgroup_subsys_state *css, 7914 struct cgroup_subsys_state *old_css, 7915 struct task_struct *task) 7916{ 7917 /* 7918 * cgroup_exit() is called in the copy_process() failure path. 7919 * Ignore this case since the task hasn't ran yet, this avoids 7920 * trying to poke a half freed task state from generic code. 7921 */ 7922 if (!(task->flags & PF_EXITING)) 7923 return; 7924 7925 sched_move_task(task); 7926} 7927 7928#ifdef CONFIG_FAIR_GROUP_SCHED 7929static int cpu_shares_write_u64(struct cgroup_subsys_state *css, 7930 struct cftype *cftype, u64 shareval) 7931{ 7932 return sched_group_set_shares(css_tg(css), scale_load(shareval)); 7933} 7934 7935static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css, 7936 struct cftype *cft) 7937{ 7938 struct task_group *tg = css_tg(css); 7939 7940 return (u64) scale_load_down(tg->shares); 7941} 7942 7943#ifdef CONFIG_CFS_BANDWIDTH 7944static DEFINE_MUTEX(cfs_constraints_mutex); 7945 7946const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */ 7947const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */ 7948 7949static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime); 7950 7951static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota) 7952{ 7953 int i, ret = 0, runtime_enabled, runtime_was_enabled; 7954 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth; 7955 7956 if (tg == &root_task_group) 7957 return -EINVAL; 7958 7959 /* 7960 * Ensure we have at some amount of bandwidth every period. This is 7961 * to prevent reaching a state of large arrears when throttled via 7962 * entity_tick() resulting in prolonged exit starvation. 7963 */ 7964 if (quota < min_cfs_quota_period || period < min_cfs_quota_period) 7965 return -EINVAL; 7966 7967 /* 7968 * Likewise, bound things on the otherside by preventing insane quota 7969 * periods. This also allows us to normalize in computing quota 7970 * feasibility. 7971 */ 7972 if (period > max_cfs_quota_period) 7973 return -EINVAL; 7974 7975 /* 7976 * Prevent race between setting of cfs_rq->runtime_enabled and 7977 * unthrottle_offline_cfs_rqs(). 7978 */ 7979 get_online_cpus(); 7980 mutex_lock(&cfs_constraints_mutex); 7981 ret = __cfs_schedulable(tg, period, quota); 7982 if (ret) 7983 goto out_unlock; 7984 7985 runtime_enabled = quota != RUNTIME_INF; 7986 runtime_was_enabled = cfs_b->quota != RUNTIME_INF; 7987 /* 7988 * If we need to toggle cfs_bandwidth_used, off->on must occur 7989 * before making related changes, and on->off must occur afterwards 7990 */ 7991 if (runtime_enabled && !runtime_was_enabled) 7992 cfs_bandwidth_usage_inc(); 7993 raw_spin_lock_irq(&cfs_b->lock); 7994 cfs_b->period = ns_to_ktime(period); 7995 cfs_b->quota = quota; 7996 7997 __refill_cfs_bandwidth_runtime(cfs_b); 7998 /* restart the period timer (if active) to handle new period expiry */ 7999 if (runtime_enabled && cfs_b->timer_active) { 8000 /* force a reprogram */ 8001 __start_cfs_bandwidth(cfs_b, true); 8002 } 8003 raw_spin_unlock_irq(&cfs_b->lock); 8004 8005 for_each_online_cpu(i) { 8006 struct cfs_rq *cfs_rq = tg->cfs_rq[i]; 8007 struct rq *rq = cfs_rq->rq; 8008 8009 raw_spin_lock_irq(&rq->lock); 8010 cfs_rq->runtime_enabled = runtime_enabled; 8011 cfs_rq->runtime_remaining = 0; 8012 8013 if (cfs_rq->throttled) 8014 unthrottle_cfs_rq(cfs_rq); 8015 raw_spin_unlock_irq(&rq->lock); 8016 } 8017 if (runtime_was_enabled && !runtime_enabled) 8018 cfs_bandwidth_usage_dec(); 8019out_unlock: 8020 mutex_unlock(&cfs_constraints_mutex); 8021 put_online_cpus(); 8022 8023 return ret; 8024} 8025 8026int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us) 8027{ 8028 u64 quota, period; 8029 8030 period = ktime_to_ns(tg->cfs_bandwidth.period); 8031 if (cfs_quota_us < 0) 8032 quota = RUNTIME_INF; 8033 else 8034 quota = (u64)cfs_quota_us * NSEC_PER_USEC; 8035 8036 return tg_set_cfs_bandwidth(tg, period, quota); 8037} 8038 8039long tg_get_cfs_quota(struct task_group *tg) 8040{ 8041 u64 quota_us; 8042 8043 if (tg->cfs_bandwidth.quota == RUNTIME_INF) 8044 return -1; 8045 8046 quota_us = tg->cfs_bandwidth.quota; 8047 do_div(quota_us, NSEC_PER_USEC); 8048 8049 return quota_us; 8050} 8051 8052int tg_set_cfs_period(struct task_group *tg, long cfs_period_us) 8053{ 8054 u64 quota, period; 8055 8056 period = (u64)cfs_period_us * NSEC_PER_USEC; 8057 quota = tg->cfs_bandwidth.quota; 8058 8059 return tg_set_cfs_bandwidth(tg, period, quota); 8060} 8061 8062long tg_get_cfs_period(struct task_group *tg) 8063{ 8064 u64 cfs_period_us; 8065 8066 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period); 8067 do_div(cfs_period_us, NSEC_PER_USEC); 8068 8069 return cfs_period_us; 8070} 8071 8072static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css, 8073 struct cftype *cft) 8074{ 8075 return tg_get_cfs_quota(css_tg(css)); 8076} 8077 8078static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css, 8079 struct cftype *cftype, s64 cfs_quota_us) 8080{ 8081 return tg_set_cfs_quota(css_tg(css), cfs_quota_us); 8082} 8083 8084static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css, 8085 struct cftype *cft) 8086{ 8087 return tg_get_cfs_period(css_tg(css)); 8088} 8089 8090static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css, 8091 struct cftype *cftype, u64 cfs_period_us) 8092{ 8093 return tg_set_cfs_period(css_tg(css), cfs_period_us); 8094} 8095 8096struct cfs_schedulable_data { 8097 struct task_group *tg; 8098 u64 period, quota; 8099}; 8100 8101/* 8102 * normalize group quota/period to be quota/max_period 8103 * note: units are usecs 8104 */ 8105static u64 normalize_cfs_quota(struct task_group *tg, 8106 struct cfs_schedulable_data *d) 8107{ 8108 u64 quota, period; 8109 8110 if (tg == d->tg) { 8111 period = d->period; 8112 quota = d->quota; 8113 } else { 8114 period = tg_get_cfs_period(tg); 8115 quota = tg_get_cfs_quota(tg); 8116 } 8117 8118 /* note: these should typically be equivalent */ 8119 if (quota == RUNTIME_INF || quota == -1) 8120 return RUNTIME_INF; 8121 8122 return to_ratio(period, quota); 8123} 8124 8125static int tg_cfs_schedulable_down(struct task_group *tg, void *data) 8126{ 8127 struct cfs_schedulable_data *d = data; 8128 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth; 8129 s64 quota = 0, parent_quota = -1; 8130 8131 if (!tg->parent) { 8132 quota = RUNTIME_INF; 8133 } else { 8134 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth; 8135 8136 quota = normalize_cfs_quota(tg, d); 8137 parent_quota = parent_b->hierarchical_quota; 8138 8139 /* 8140 * ensure max(child_quota) <= parent_quota, inherit when no 8141 * limit is set 8142 */ 8143 if (quota == RUNTIME_INF) 8144 quota = parent_quota; 8145 else if (parent_quota != RUNTIME_INF && quota > parent_quota) 8146 return -EINVAL; 8147 } 8148 cfs_b->hierarchical_quota = quota; 8149 8150 return 0; 8151} 8152 8153static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota) 8154{ 8155 int ret; 8156 struct cfs_schedulable_data data = { 8157 .tg = tg, 8158 .period = period, 8159 .quota = quota, 8160 }; 8161 8162 if (quota != RUNTIME_INF) { 8163 do_div(data.period, NSEC_PER_USEC); 8164 do_div(data.quota, NSEC_PER_USEC); 8165 } 8166 8167 rcu_read_lock(); 8168 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data); 8169 rcu_read_unlock(); 8170 8171 return ret; 8172} 8173 8174static int cpu_stats_show(struct seq_file *sf, void *v) 8175{ 8176 struct task_group *tg = css_tg(seq_css(sf)); 8177 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth; 8178 8179 seq_printf(sf, "nr_periods %d\n", cfs_b->nr_periods); 8180 seq_printf(sf, "nr_throttled %d\n", cfs_b->nr_throttled); 8181 seq_printf(sf, "throttled_time %llu\n", cfs_b->throttled_time); 8182 8183 return 0; 8184} 8185#endif /* CONFIG_CFS_BANDWIDTH */ 8186#endif /* CONFIG_FAIR_GROUP_SCHED */ 8187 8188#ifdef CONFIG_RT_GROUP_SCHED 8189static int cpu_rt_runtime_write(struct cgroup_subsys_state *css, 8190 struct cftype *cft, s64 val) 8191{ 8192 return sched_group_set_rt_runtime(css_tg(css), val); 8193} 8194 8195static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css, 8196 struct cftype *cft) 8197{ 8198 return sched_group_rt_runtime(css_tg(css)); 8199} 8200 8201static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css, 8202 struct cftype *cftype, u64 rt_period_us) 8203{ 8204 return sched_group_set_rt_period(css_tg(css), rt_period_us); 8205} 8206 8207static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css, 8208 struct cftype *cft) 8209{ 8210 return sched_group_rt_period(css_tg(css)); 8211} 8212#endif /* CONFIG_RT_GROUP_SCHED */ 8213 8214static struct cftype cpu_files[] = { 8215#ifdef CONFIG_FAIR_GROUP_SCHED 8216 { 8217 .name = "shares", 8218 .read_u64 = cpu_shares_read_u64, 8219 .write_u64 = cpu_shares_write_u64, 8220 }, 8221#endif 8222#ifdef CONFIG_CFS_BANDWIDTH 8223 { 8224 .name = "cfs_quota_us", 8225 .read_s64 = cpu_cfs_quota_read_s64, 8226 .write_s64 = cpu_cfs_quota_write_s64, 8227 }, 8228 { 8229 .name = "cfs_period_us", 8230 .read_u64 = cpu_cfs_period_read_u64, 8231 .write_u64 = cpu_cfs_period_write_u64, 8232 }, 8233 { 8234 .name = "stat", 8235 .seq_show = cpu_stats_show, 8236 }, 8237#endif 8238#ifdef CONFIG_RT_GROUP_SCHED 8239 { 8240 .name = "rt_runtime_us", 8241 .read_s64 = cpu_rt_runtime_read, 8242 .write_s64 = cpu_rt_runtime_write, 8243 }, 8244 { 8245 .name = "rt_period_us", 8246 .read_u64 = cpu_rt_period_read_uint, 8247 .write_u64 = cpu_rt_period_write_uint, 8248 }, 8249#endif 8250 { } /* terminate */ 8251}; 8252 8253struct cgroup_subsys cpu_cgrp_subsys = { 8254 .css_alloc = cpu_cgroup_css_alloc, 8255 .css_free = cpu_cgroup_css_free, 8256 .css_online = cpu_cgroup_css_online, 8257 .css_offline = cpu_cgroup_css_offline, 8258 .fork = cpu_cgroup_fork, 8259 .can_attach = cpu_cgroup_can_attach, 8260 .attach = cpu_cgroup_attach, 8261 .exit = cpu_cgroup_exit, 8262 .legacy_cftypes = cpu_files, 8263 .early_init = 1, 8264}; 8265 8266#endif /* CONFIG_CGROUP_SCHED */ 8267 8268void dump_cpu_task(int cpu) 8269{ 8270 pr_info("Task dump for CPU %d:\n", cpu); 8271 sched_show_task(cpu_curr(cpu)); 8272} 8273