core.c revision 3472eaa1f12e217e2b8b0ef658ff861b2308cbbd
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 93#ifdef smp_mb__before_atomic 94void __smp_mb__before_atomic(void) 95{ 96 smp_mb__before_atomic(); 97} 98EXPORT_SYMBOL(__smp_mb__before_atomic); 99#endif 100 101#ifdef smp_mb__after_atomic 102void __smp_mb__after_atomic(void) 103{ 104 smp_mb__after_atomic(); 105} 106EXPORT_SYMBOL(__smp_mb__after_atomic); 107#endif 108 109void start_bandwidth_timer(struct hrtimer *period_timer, ktime_t period) 110{ 111 unsigned long delta; 112 ktime_t soft, hard, now; 113 114 for (;;) { 115 if (hrtimer_active(period_timer)) 116 break; 117 118 now = hrtimer_cb_get_time(period_timer); 119 hrtimer_forward(period_timer, now, period); 120 121 soft = hrtimer_get_softexpires(period_timer); 122 hard = hrtimer_get_expires(period_timer); 123 delta = ktime_to_ns(ktime_sub(hard, soft)); 124 __hrtimer_start_range_ns(period_timer, soft, delta, 125 HRTIMER_MODE_ABS_PINNED, 0); 126 } 127} 128 129DEFINE_MUTEX(sched_domains_mutex); 130DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues); 131 132static void update_rq_clock_task(struct rq *rq, s64 delta); 133 134void update_rq_clock(struct rq *rq) 135{ 136 s64 delta; 137 138 if (rq->skip_clock_update > 0) 139 return; 140 141 delta = sched_clock_cpu(cpu_of(rq)) - rq->clock; 142 if (delta < 0) 143 return; 144 rq->clock += delta; 145 update_rq_clock_task(rq, delta); 146} 147 148/* 149 * Debugging: various feature bits 150 */ 151 152#define SCHED_FEAT(name, enabled) \ 153 (1UL << __SCHED_FEAT_##name) * enabled | 154 155const_debug unsigned int sysctl_sched_features = 156#include "features.h" 157 0; 158 159#undef SCHED_FEAT 160 161#ifdef CONFIG_SCHED_DEBUG 162#define SCHED_FEAT(name, enabled) \ 163 #name , 164 165static const char * const sched_feat_names[] = { 166#include "features.h" 167}; 168 169#undef SCHED_FEAT 170 171static int sched_feat_show(struct seq_file *m, void *v) 172{ 173 int i; 174 175 for (i = 0; i < __SCHED_FEAT_NR; i++) { 176 if (!(sysctl_sched_features & (1UL << i))) 177 seq_puts(m, "NO_"); 178 seq_printf(m, "%s ", sched_feat_names[i]); 179 } 180 seq_puts(m, "\n"); 181 182 return 0; 183} 184 185#ifdef HAVE_JUMP_LABEL 186 187#define jump_label_key__true STATIC_KEY_INIT_TRUE 188#define jump_label_key__false STATIC_KEY_INIT_FALSE 189 190#define SCHED_FEAT(name, enabled) \ 191 jump_label_key__##enabled , 192 193struct static_key sched_feat_keys[__SCHED_FEAT_NR] = { 194#include "features.h" 195}; 196 197#undef SCHED_FEAT 198 199static void sched_feat_disable(int i) 200{ 201 if (static_key_enabled(&sched_feat_keys[i])) 202 static_key_slow_dec(&sched_feat_keys[i]); 203} 204 205static void sched_feat_enable(int i) 206{ 207 if (!static_key_enabled(&sched_feat_keys[i])) 208 static_key_slow_inc(&sched_feat_keys[i]); 209} 210#else 211static void sched_feat_disable(int i) { }; 212static void sched_feat_enable(int i) { }; 213#endif /* HAVE_JUMP_LABEL */ 214 215static int sched_feat_set(char *cmp) 216{ 217 int i; 218 int neg = 0; 219 220 if (strncmp(cmp, "NO_", 3) == 0) { 221 neg = 1; 222 cmp += 3; 223 } 224 225 for (i = 0; i < __SCHED_FEAT_NR; i++) { 226 if (strcmp(cmp, sched_feat_names[i]) == 0) { 227 if (neg) { 228 sysctl_sched_features &= ~(1UL << i); 229 sched_feat_disable(i); 230 } else { 231 sysctl_sched_features |= (1UL << i); 232 sched_feat_enable(i); 233 } 234 break; 235 } 236 } 237 238 return i; 239} 240 241static ssize_t 242sched_feat_write(struct file *filp, const char __user *ubuf, 243 size_t cnt, loff_t *ppos) 244{ 245 char buf[64]; 246 char *cmp; 247 int i; 248 struct inode *inode; 249 250 if (cnt > 63) 251 cnt = 63; 252 253 if (copy_from_user(&buf, ubuf, cnt)) 254 return -EFAULT; 255 256 buf[cnt] = 0; 257 cmp = strstrip(buf); 258 259 /* Ensure the static_key remains in a consistent state */ 260 inode = file_inode(filp); 261 mutex_lock(&inode->i_mutex); 262 i = sched_feat_set(cmp); 263 mutex_unlock(&inode->i_mutex); 264 if (i == __SCHED_FEAT_NR) 265 return -EINVAL; 266 267 *ppos += cnt; 268 269 return cnt; 270} 271 272static int sched_feat_open(struct inode *inode, struct file *filp) 273{ 274 return single_open(filp, sched_feat_show, NULL); 275} 276 277static const struct file_operations sched_feat_fops = { 278 .open = sched_feat_open, 279 .write = sched_feat_write, 280 .read = seq_read, 281 .llseek = seq_lseek, 282 .release = single_release, 283}; 284 285static __init int sched_init_debug(void) 286{ 287 debugfs_create_file("sched_features", 0644, NULL, NULL, 288 &sched_feat_fops); 289 290 return 0; 291} 292late_initcall(sched_init_debug); 293#endif /* CONFIG_SCHED_DEBUG */ 294 295/* 296 * Number of tasks to iterate in a single balance run. 297 * Limited because this is done with IRQs disabled. 298 */ 299const_debug unsigned int sysctl_sched_nr_migrate = 32; 300 301/* 302 * period over which we average the RT time consumption, measured 303 * in ms. 304 * 305 * default: 1s 306 */ 307const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC; 308 309/* 310 * period over which we measure -rt task cpu usage in us. 311 * default: 1s 312 */ 313unsigned int sysctl_sched_rt_period = 1000000; 314 315__read_mostly int scheduler_running; 316 317/* 318 * part of the period that we allow rt tasks to run in us. 319 * default: 0.95s 320 */ 321int sysctl_sched_rt_runtime = 950000; 322 323/* 324 * __task_rq_lock - lock the rq @p resides on. 325 */ 326static inline struct rq *__task_rq_lock(struct task_struct *p) 327 __acquires(rq->lock) 328{ 329 struct rq *rq; 330 331 lockdep_assert_held(&p->pi_lock); 332 333 for (;;) { 334 rq = task_rq(p); 335 raw_spin_lock(&rq->lock); 336 if (likely(rq == task_rq(p) && !task_on_rq_migrating(p))) 337 return rq; 338 raw_spin_unlock(&rq->lock); 339 340 while (unlikely(task_on_rq_migrating(p))) 341 cpu_relax(); 342 } 343} 344 345/* 346 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on. 347 */ 348static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags) 349 __acquires(p->pi_lock) 350 __acquires(rq->lock) 351{ 352 struct rq *rq; 353 354 for (;;) { 355 raw_spin_lock_irqsave(&p->pi_lock, *flags); 356 rq = task_rq(p); 357 raw_spin_lock(&rq->lock); 358 if (likely(rq == task_rq(p) && !task_on_rq_migrating(p))) 359 return rq; 360 raw_spin_unlock(&rq->lock); 361 raw_spin_unlock_irqrestore(&p->pi_lock, *flags); 362 363 while (unlikely(task_on_rq_migrating(p))) 364 cpu_relax(); 365 } 366} 367 368static void __task_rq_unlock(struct rq *rq) 369 __releases(rq->lock) 370{ 371 raw_spin_unlock(&rq->lock); 372} 373 374static inline void 375task_rq_unlock(struct rq *rq, struct task_struct *p, unsigned long *flags) 376 __releases(rq->lock) 377 __releases(p->pi_lock) 378{ 379 raw_spin_unlock(&rq->lock); 380 raw_spin_unlock_irqrestore(&p->pi_lock, *flags); 381} 382 383/* 384 * this_rq_lock - lock this runqueue and disable interrupts. 385 */ 386static struct rq *this_rq_lock(void) 387 __acquires(rq->lock) 388{ 389 struct rq *rq; 390 391 local_irq_disable(); 392 rq = this_rq(); 393 raw_spin_lock(&rq->lock); 394 395 return rq; 396} 397 398#ifdef CONFIG_SCHED_HRTICK 399/* 400 * Use HR-timers to deliver accurate preemption points. 401 */ 402 403static void hrtick_clear(struct rq *rq) 404{ 405 if (hrtimer_active(&rq->hrtick_timer)) 406 hrtimer_cancel(&rq->hrtick_timer); 407} 408 409/* 410 * High-resolution timer tick. 411 * Runs from hardirq context with interrupts disabled. 412 */ 413static enum hrtimer_restart hrtick(struct hrtimer *timer) 414{ 415 struct rq *rq = container_of(timer, struct rq, hrtick_timer); 416 417 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id()); 418 419 raw_spin_lock(&rq->lock); 420 update_rq_clock(rq); 421 rq->curr->sched_class->task_tick(rq, rq->curr, 1); 422 raw_spin_unlock(&rq->lock); 423 424 return HRTIMER_NORESTART; 425} 426 427#ifdef CONFIG_SMP 428 429static int __hrtick_restart(struct rq *rq) 430{ 431 struct hrtimer *timer = &rq->hrtick_timer; 432 ktime_t time = hrtimer_get_softexpires(timer); 433 434 return __hrtimer_start_range_ns(timer, time, 0, HRTIMER_MODE_ABS_PINNED, 0); 435} 436 437/* 438 * called from hardirq (IPI) context 439 */ 440static void __hrtick_start(void *arg) 441{ 442 struct rq *rq = arg; 443 444 raw_spin_lock(&rq->lock); 445 __hrtick_restart(rq); 446 rq->hrtick_csd_pending = 0; 447 raw_spin_unlock(&rq->lock); 448} 449 450/* 451 * Called to set the hrtick timer state. 452 * 453 * called with rq->lock held and irqs disabled 454 */ 455void hrtick_start(struct rq *rq, u64 delay) 456{ 457 struct hrtimer *timer = &rq->hrtick_timer; 458 ktime_t time; 459 s64 delta; 460 461 /* 462 * Don't schedule slices shorter than 10000ns, that just 463 * doesn't make sense and can cause timer DoS. 464 */ 465 delta = max_t(s64, delay, 10000LL); 466 time = ktime_add_ns(timer->base->get_time(), delta); 467 468 hrtimer_set_expires(timer, time); 469 470 if (rq == this_rq()) { 471 __hrtick_restart(rq); 472 } else if (!rq->hrtick_csd_pending) { 473 smp_call_function_single_async(cpu_of(rq), &rq->hrtick_csd); 474 rq->hrtick_csd_pending = 1; 475 } 476} 477 478static int 479hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu) 480{ 481 int cpu = (int)(long)hcpu; 482 483 switch (action) { 484 case CPU_UP_CANCELED: 485 case CPU_UP_CANCELED_FROZEN: 486 case CPU_DOWN_PREPARE: 487 case CPU_DOWN_PREPARE_FROZEN: 488 case CPU_DEAD: 489 case CPU_DEAD_FROZEN: 490 hrtick_clear(cpu_rq(cpu)); 491 return NOTIFY_OK; 492 } 493 494 return NOTIFY_DONE; 495} 496 497static __init void init_hrtick(void) 498{ 499 hotcpu_notifier(hotplug_hrtick, 0); 500} 501#else 502/* 503 * Called to set the hrtick timer state. 504 * 505 * called with rq->lock held and irqs disabled 506 */ 507void hrtick_start(struct rq *rq, u64 delay) 508{ 509 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0, 510 HRTIMER_MODE_REL_PINNED, 0); 511} 512 513static inline void init_hrtick(void) 514{ 515} 516#endif /* CONFIG_SMP */ 517 518static void init_rq_hrtick(struct rq *rq) 519{ 520#ifdef CONFIG_SMP 521 rq->hrtick_csd_pending = 0; 522 523 rq->hrtick_csd.flags = 0; 524 rq->hrtick_csd.func = __hrtick_start; 525 rq->hrtick_csd.info = rq; 526#endif 527 528 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); 529 rq->hrtick_timer.function = hrtick; 530} 531#else /* CONFIG_SCHED_HRTICK */ 532static inline void hrtick_clear(struct rq *rq) 533{ 534} 535 536static inline void init_rq_hrtick(struct rq *rq) 537{ 538} 539 540static inline void init_hrtick(void) 541{ 542} 543#endif /* CONFIG_SCHED_HRTICK */ 544 545/* 546 * cmpxchg based fetch_or, macro so it works for different integer types 547 */ 548#define fetch_or(ptr, val) \ 549({ typeof(*(ptr)) __old, __val = *(ptr); \ 550 for (;;) { \ 551 __old = cmpxchg((ptr), __val, __val | (val)); \ 552 if (__old == __val) \ 553 break; \ 554 __val = __old; \ 555 } \ 556 __old; \ 557}) 558 559#if defined(CONFIG_SMP) && defined(TIF_POLLING_NRFLAG) 560/* 561 * Atomically set TIF_NEED_RESCHED and test for TIF_POLLING_NRFLAG, 562 * this avoids any races wrt polling state changes and thereby avoids 563 * spurious IPIs. 564 */ 565static bool set_nr_and_not_polling(struct task_struct *p) 566{ 567 struct thread_info *ti = task_thread_info(p); 568 return !(fetch_or(&ti->flags, _TIF_NEED_RESCHED) & _TIF_POLLING_NRFLAG); 569} 570 571/* 572 * Atomically set TIF_NEED_RESCHED if TIF_POLLING_NRFLAG is set. 573 * 574 * If this returns true, then the idle task promises to call 575 * sched_ttwu_pending() and reschedule soon. 576 */ 577static bool set_nr_if_polling(struct task_struct *p) 578{ 579 struct thread_info *ti = task_thread_info(p); 580 typeof(ti->flags) old, val = ACCESS_ONCE(ti->flags); 581 582 for (;;) { 583 if (!(val & _TIF_POLLING_NRFLAG)) 584 return false; 585 if (val & _TIF_NEED_RESCHED) 586 return true; 587 old = cmpxchg(&ti->flags, val, val | _TIF_NEED_RESCHED); 588 if (old == val) 589 break; 590 val = old; 591 } 592 return true; 593} 594 595#else 596static bool set_nr_and_not_polling(struct task_struct *p) 597{ 598 set_tsk_need_resched(p); 599 return true; 600} 601 602#ifdef CONFIG_SMP 603static bool set_nr_if_polling(struct task_struct *p) 604{ 605 return false; 606} 607#endif 608#endif 609 610/* 611 * resched_curr - mark rq's current task 'to be rescheduled now'. 612 * 613 * On UP this means the setting of the need_resched flag, on SMP it 614 * might also involve a cross-CPU call to trigger the scheduler on 615 * the target CPU. 616 */ 617void resched_curr(struct rq *rq) 618{ 619 struct task_struct *curr = rq->curr; 620 int cpu; 621 622 lockdep_assert_held(&rq->lock); 623 624 if (test_tsk_need_resched(curr)) 625 return; 626 627 cpu = cpu_of(rq); 628 629 if (cpu == smp_processor_id()) { 630 set_tsk_need_resched(curr); 631 set_preempt_need_resched(); 632 return; 633 } 634 635 if (set_nr_and_not_polling(curr)) 636 smp_send_reschedule(cpu); 637 else 638 trace_sched_wake_idle_without_ipi(cpu); 639} 640 641void resched_cpu(int cpu) 642{ 643 struct rq *rq = cpu_rq(cpu); 644 unsigned long flags; 645 646 if (!raw_spin_trylock_irqsave(&rq->lock, flags)) 647 return; 648 resched_curr(rq); 649 raw_spin_unlock_irqrestore(&rq->lock, flags); 650} 651 652#ifdef CONFIG_SMP 653#ifdef CONFIG_NO_HZ_COMMON 654/* 655 * In the semi idle case, use the nearest busy cpu for migrating timers 656 * from an idle cpu. This is good for power-savings. 657 * 658 * We don't do similar optimization for completely idle system, as 659 * selecting an idle cpu will add more delays to the timers than intended 660 * (as that cpu's timer base may not be uptodate wrt jiffies etc). 661 */ 662int get_nohz_timer_target(int pinned) 663{ 664 int cpu = smp_processor_id(); 665 int i; 666 struct sched_domain *sd; 667 668 if (pinned || !get_sysctl_timer_migration() || !idle_cpu(cpu)) 669 return cpu; 670 671 rcu_read_lock(); 672 for_each_domain(cpu, sd) { 673 for_each_cpu(i, sched_domain_span(sd)) { 674 if (!idle_cpu(i)) { 675 cpu = i; 676 goto unlock; 677 } 678 } 679 } 680unlock: 681 rcu_read_unlock(); 682 return cpu; 683} 684/* 685 * When add_timer_on() enqueues a timer into the timer wheel of an 686 * idle CPU then this timer might expire before the next timer event 687 * which is scheduled to wake up that CPU. In case of a completely 688 * idle system the next event might even be infinite time into the 689 * future. wake_up_idle_cpu() ensures that the CPU is woken up and 690 * leaves the inner idle loop so the newly added timer is taken into 691 * account when the CPU goes back to idle and evaluates the timer 692 * wheel for the next timer event. 693 */ 694static void wake_up_idle_cpu(int cpu) 695{ 696 struct rq *rq = cpu_rq(cpu); 697 698 if (cpu == smp_processor_id()) 699 return; 700 701 if (set_nr_and_not_polling(rq->idle)) 702 smp_send_reschedule(cpu); 703 else 704 trace_sched_wake_idle_without_ipi(cpu); 705} 706 707static bool wake_up_full_nohz_cpu(int cpu) 708{ 709 /* 710 * We just need the target to call irq_exit() and re-evaluate 711 * the next tick. The nohz full kick at least implies that. 712 * If needed we can still optimize that later with an 713 * empty IRQ. 714 */ 715 if (tick_nohz_full_cpu(cpu)) { 716 if (cpu != smp_processor_id() || 717 tick_nohz_tick_stopped()) 718 tick_nohz_full_kick_cpu(cpu); 719 return true; 720 } 721 722 return false; 723} 724 725void wake_up_nohz_cpu(int cpu) 726{ 727 if (!wake_up_full_nohz_cpu(cpu)) 728 wake_up_idle_cpu(cpu); 729} 730 731static inline bool got_nohz_idle_kick(void) 732{ 733 int cpu = smp_processor_id(); 734 735 if (!test_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu))) 736 return false; 737 738 if (idle_cpu(cpu) && !need_resched()) 739 return true; 740 741 /* 742 * We can't run Idle Load Balance on this CPU for this time so we 743 * cancel it and clear NOHZ_BALANCE_KICK 744 */ 745 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu)); 746 return false; 747} 748 749#else /* CONFIG_NO_HZ_COMMON */ 750 751static inline bool got_nohz_idle_kick(void) 752{ 753 return false; 754} 755 756#endif /* CONFIG_NO_HZ_COMMON */ 757 758#ifdef CONFIG_NO_HZ_FULL 759bool sched_can_stop_tick(void) 760{ 761 /* 762 * More than one running task need preemption. 763 * nr_running update is assumed to be visible 764 * after IPI is sent from wakers. 765 */ 766 if (this_rq()->nr_running > 1) 767 return false; 768 769 return true; 770} 771#endif /* CONFIG_NO_HZ_FULL */ 772 773void sched_avg_update(struct rq *rq) 774{ 775 s64 period = sched_avg_period(); 776 777 while ((s64)(rq_clock(rq) - rq->age_stamp) > period) { 778 /* 779 * Inline assembly required to prevent the compiler 780 * optimising this loop into a divmod call. 781 * See __iter_div_u64_rem() for another example of this. 782 */ 783 asm("" : "+rm" (rq->age_stamp)); 784 rq->age_stamp += period; 785 rq->rt_avg /= 2; 786 } 787} 788 789#endif /* CONFIG_SMP */ 790 791#if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \ 792 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH))) 793/* 794 * Iterate task_group tree rooted at *from, calling @down when first entering a 795 * node and @up when leaving it for the final time. 796 * 797 * Caller must hold rcu_lock or sufficient equivalent. 798 */ 799int walk_tg_tree_from(struct task_group *from, 800 tg_visitor down, tg_visitor up, void *data) 801{ 802 struct task_group *parent, *child; 803 int ret; 804 805 parent = from; 806 807down: 808 ret = (*down)(parent, data); 809 if (ret) 810 goto out; 811 list_for_each_entry_rcu(child, &parent->children, siblings) { 812 parent = child; 813 goto down; 814 815up: 816 continue; 817 } 818 ret = (*up)(parent, data); 819 if (ret || parent == from) 820 goto out; 821 822 child = parent; 823 parent = parent->parent; 824 if (parent) 825 goto up; 826out: 827 return ret; 828} 829 830int tg_nop(struct task_group *tg, void *data) 831{ 832 return 0; 833} 834#endif 835 836static void set_load_weight(struct task_struct *p) 837{ 838 int prio = p->static_prio - MAX_RT_PRIO; 839 struct load_weight *load = &p->se.load; 840 841 /* 842 * SCHED_IDLE tasks get minimal weight: 843 */ 844 if (p->policy == SCHED_IDLE) { 845 load->weight = scale_load(WEIGHT_IDLEPRIO); 846 load->inv_weight = WMULT_IDLEPRIO; 847 return; 848 } 849 850 load->weight = scale_load(prio_to_weight[prio]); 851 load->inv_weight = prio_to_wmult[prio]; 852} 853 854static void enqueue_task(struct rq *rq, struct task_struct *p, int flags) 855{ 856 update_rq_clock(rq); 857 sched_info_queued(rq, p); 858 p->sched_class->enqueue_task(rq, p, flags); 859} 860 861static void dequeue_task(struct rq *rq, struct task_struct *p, int flags) 862{ 863 update_rq_clock(rq); 864 sched_info_dequeued(rq, p); 865 p->sched_class->dequeue_task(rq, p, flags); 866} 867 868void activate_task(struct rq *rq, struct task_struct *p, int flags) 869{ 870 if (task_contributes_to_load(p)) 871 rq->nr_uninterruptible--; 872 873 enqueue_task(rq, p, flags); 874} 875 876void deactivate_task(struct rq *rq, struct task_struct *p, int flags) 877{ 878 if (task_contributes_to_load(p)) 879 rq->nr_uninterruptible++; 880 881 dequeue_task(rq, p, flags); 882} 883 884static void update_rq_clock_task(struct rq *rq, s64 delta) 885{ 886/* 887 * In theory, the compile should just see 0 here, and optimize out the call 888 * to sched_rt_avg_update. But I don't trust it... 889 */ 890#if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING) 891 s64 steal = 0, irq_delta = 0; 892#endif 893#ifdef CONFIG_IRQ_TIME_ACCOUNTING 894 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time; 895 896 /* 897 * Since irq_time is only updated on {soft,}irq_exit, we might run into 898 * this case when a previous update_rq_clock() happened inside a 899 * {soft,}irq region. 900 * 901 * When this happens, we stop ->clock_task and only update the 902 * prev_irq_time stamp to account for the part that fit, so that a next 903 * update will consume the rest. This ensures ->clock_task is 904 * monotonic. 905 * 906 * It does however cause some slight miss-attribution of {soft,}irq 907 * time, a more accurate solution would be to update the irq_time using 908 * the current rq->clock timestamp, except that would require using 909 * atomic ops. 910 */ 911 if (irq_delta > delta) 912 irq_delta = delta; 913 914 rq->prev_irq_time += irq_delta; 915 delta -= irq_delta; 916#endif 917#ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING 918 if (static_key_false((¶virt_steal_rq_enabled))) { 919 steal = paravirt_steal_clock(cpu_of(rq)); 920 steal -= rq->prev_steal_time_rq; 921 922 if (unlikely(steal > delta)) 923 steal = delta; 924 925 rq->prev_steal_time_rq += steal; 926 delta -= steal; 927 } 928#endif 929 930 rq->clock_task += delta; 931 932#if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING) 933 if ((irq_delta + steal) && sched_feat(NONTASK_CAPACITY)) 934 sched_rt_avg_update(rq, irq_delta + steal); 935#endif 936} 937 938void sched_set_stop_task(int cpu, struct task_struct *stop) 939{ 940 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 }; 941 struct task_struct *old_stop = cpu_rq(cpu)->stop; 942 943 if (stop) { 944 /* 945 * Make it appear like a SCHED_FIFO task, its something 946 * userspace knows about and won't get confused about. 947 * 948 * Also, it will make PI more or less work without too 949 * much confusion -- but then, stop work should not 950 * rely on PI working anyway. 951 */ 952 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m); 953 954 stop->sched_class = &stop_sched_class; 955 } 956 957 cpu_rq(cpu)->stop = stop; 958 959 if (old_stop) { 960 /* 961 * Reset it back to a normal scheduling class so that 962 * it can die in pieces. 963 */ 964 old_stop->sched_class = &rt_sched_class; 965 } 966} 967 968/* 969 * __normal_prio - return the priority that is based on the static prio 970 */ 971static inline int __normal_prio(struct task_struct *p) 972{ 973 return p->static_prio; 974} 975 976/* 977 * Calculate the expected normal priority: i.e. priority 978 * without taking RT-inheritance into account. Might be 979 * boosted by interactivity modifiers. Changes upon fork, 980 * setprio syscalls, and whenever the interactivity 981 * estimator recalculates. 982 */ 983static inline int normal_prio(struct task_struct *p) 984{ 985 int prio; 986 987 if (task_has_dl_policy(p)) 988 prio = MAX_DL_PRIO-1; 989 else if (task_has_rt_policy(p)) 990 prio = MAX_RT_PRIO-1 - p->rt_priority; 991 else 992 prio = __normal_prio(p); 993 return prio; 994} 995 996/* 997 * Calculate the current priority, i.e. the priority 998 * taken into account by the scheduler. This value might 999 * be boosted by RT tasks, or might be boosted by 1000 * interactivity modifiers. Will be RT if the task got 1001 * RT-boosted. If not then it returns p->normal_prio. 1002 */ 1003static int effective_prio(struct task_struct *p) 1004{ 1005 p->normal_prio = normal_prio(p); 1006 /* 1007 * If we are RT tasks or we were boosted to RT priority, 1008 * keep the priority unchanged. Otherwise, update priority 1009 * to the normal priority: 1010 */ 1011 if (!rt_prio(p->prio)) 1012 return p->normal_prio; 1013 return p->prio; 1014} 1015 1016/** 1017 * task_curr - is this task currently executing on a CPU? 1018 * @p: the task in question. 1019 * 1020 * Return: 1 if the task is currently executing. 0 otherwise. 1021 */ 1022inline int task_curr(const struct task_struct *p) 1023{ 1024 return cpu_curr(task_cpu(p)) == p; 1025} 1026 1027static inline void check_class_changed(struct rq *rq, struct task_struct *p, 1028 const struct sched_class *prev_class, 1029 int oldprio) 1030{ 1031 if (prev_class != p->sched_class) { 1032 if (prev_class->switched_from) 1033 prev_class->switched_from(rq, p); 1034 p->sched_class->switched_to(rq, p); 1035 } else if (oldprio != p->prio || dl_task(p)) 1036 p->sched_class->prio_changed(rq, p, oldprio); 1037} 1038 1039void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags) 1040{ 1041 const struct sched_class *class; 1042 1043 if (p->sched_class == rq->curr->sched_class) { 1044 rq->curr->sched_class->check_preempt_curr(rq, p, flags); 1045 } else { 1046 for_each_class(class) { 1047 if (class == rq->curr->sched_class) 1048 break; 1049 if (class == p->sched_class) { 1050 resched_curr(rq); 1051 break; 1052 } 1053 } 1054 } 1055 1056 /* 1057 * A queue event has occurred, and we're going to schedule. In 1058 * this case, we can save a useless back to back clock update. 1059 */ 1060 if (task_on_rq_queued(rq->curr) && test_tsk_need_resched(rq->curr)) 1061 rq->skip_clock_update = 1; 1062} 1063 1064#ifdef CONFIG_SMP 1065void set_task_cpu(struct task_struct *p, unsigned int new_cpu) 1066{ 1067#ifdef CONFIG_SCHED_DEBUG 1068 /* 1069 * We should never call set_task_cpu() on a blocked task, 1070 * ttwu() will sort out the placement. 1071 */ 1072 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING && 1073 !(task_preempt_count(p) & PREEMPT_ACTIVE)); 1074 1075#ifdef CONFIG_LOCKDEP 1076 /* 1077 * The caller should hold either p->pi_lock or rq->lock, when changing 1078 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks. 1079 * 1080 * sched_move_task() holds both and thus holding either pins the cgroup, 1081 * see task_group(). 1082 * 1083 * Furthermore, all task_rq users should acquire both locks, see 1084 * task_rq_lock(). 1085 */ 1086 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) || 1087 lockdep_is_held(&task_rq(p)->lock))); 1088#endif 1089#endif 1090 1091 trace_sched_migrate_task(p, new_cpu); 1092 1093 if (task_cpu(p) != new_cpu) { 1094 if (p->sched_class->migrate_task_rq) 1095 p->sched_class->migrate_task_rq(p, new_cpu); 1096 p->se.nr_migrations++; 1097 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, NULL, 0); 1098 } 1099 1100 __set_task_cpu(p, new_cpu); 1101} 1102 1103static void __migrate_swap_task(struct task_struct *p, int cpu) 1104{ 1105 if (task_on_rq_queued(p)) { 1106 struct rq *src_rq, *dst_rq; 1107 1108 src_rq = task_rq(p); 1109 dst_rq = cpu_rq(cpu); 1110 1111 deactivate_task(src_rq, p, 0); 1112 set_task_cpu(p, cpu); 1113 activate_task(dst_rq, p, 0); 1114 check_preempt_curr(dst_rq, p, 0); 1115 } else { 1116 /* 1117 * Task isn't running anymore; make it appear like we migrated 1118 * it before it went to sleep. This means on wakeup we make the 1119 * previous cpu our targer instead of where it really is. 1120 */ 1121 p->wake_cpu = cpu; 1122 } 1123} 1124 1125struct migration_swap_arg { 1126 struct task_struct *src_task, *dst_task; 1127 int src_cpu, dst_cpu; 1128}; 1129 1130static int migrate_swap_stop(void *data) 1131{ 1132 struct migration_swap_arg *arg = data; 1133 struct rq *src_rq, *dst_rq; 1134 int ret = -EAGAIN; 1135 1136 src_rq = cpu_rq(arg->src_cpu); 1137 dst_rq = cpu_rq(arg->dst_cpu); 1138 1139 double_raw_lock(&arg->src_task->pi_lock, 1140 &arg->dst_task->pi_lock); 1141 double_rq_lock(src_rq, dst_rq); 1142 if (task_cpu(arg->dst_task) != arg->dst_cpu) 1143 goto unlock; 1144 1145 if (task_cpu(arg->src_task) != arg->src_cpu) 1146 goto unlock; 1147 1148 if (!cpumask_test_cpu(arg->dst_cpu, tsk_cpus_allowed(arg->src_task))) 1149 goto unlock; 1150 1151 if (!cpumask_test_cpu(arg->src_cpu, tsk_cpus_allowed(arg->dst_task))) 1152 goto unlock; 1153 1154 __migrate_swap_task(arg->src_task, arg->dst_cpu); 1155 __migrate_swap_task(arg->dst_task, arg->src_cpu); 1156 1157 ret = 0; 1158 1159unlock: 1160 double_rq_unlock(src_rq, dst_rq); 1161 raw_spin_unlock(&arg->dst_task->pi_lock); 1162 raw_spin_unlock(&arg->src_task->pi_lock); 1163 1164 return ret; 1165} 1166 1167/* 1168 * Cross migrate two tasks 1169 */ 1170int migrate_swap(struct task_struct *cur, struct task_struct *p) 1171{ 1172 struct migration_swap_arg arg; 1173 int ret = -EINVAL; 1174 1175 arg = (struct migration_swap_arg){ 1176 .src_task = cur, 1177 .src_cpu = task_cpu(cur), 1178 .dst_task = p, 1179 .dst_cpu = task_cpu(p), 1180 }; 1181 1182 if (arg.src_cpu == arg.dst_cpu) 1183 goto out; 1184 1185 /* 1186 * These three tests are all lockless; this is OK since all of them 1187 * will be re-checked with proper locks held further down the line. 1188 */ 1189 if (!cpu_active(arg.src_cpu) || !cpu_active(arg.dst_cpu)) 1190 goto out; 1191 1192 if (!cpumask_test_cpu(arg.dst_cpu, tsk_cpus_allowed(arg.src_task))) 1193 goto out; 1194 1195 if (!cpumask_test_cpu(arg.src_cpu, tsk_cpus_allowed(arg.dst_task))) 1196 goto out; 1197 1198 trace_sched_swap_numa(cur, arg.src_cpu, p, arg.dst_cpu); 1199 ret = stop_two_cpus(arg.dst_cpu, arg.src_cpu, migrate_swap_stop, &arg); 1200 1201out: 1202 return ret; 1203} 1204 1205struct migration_arg { 1206 struct task_struct *task; 1207 int dest_cpu; 1208}; 1209 1210static int migration_cpu_stop(void *data); 1211 1212/* 1213 * wait_task_inactive - wait for a thread to unschedule. 1214 * 1215 * If @match_state is nonzero, it's the @p->state value just checked and 1216 * not expected to change. If it changes, i.e. @p might have woken up, 1217 * then return zero. When we succeed in waiting for @p to be off its CPU, 1218 * we return a positive number (its total switch count). If a second call 1219 * a short while later returns the same number, the caller can be sure that 1220 * @p has remained unscheduled the whole time. 1221 * 1222 * The caller must ensure that the task *will* unschedule sometime soon, 1223 * else this function might spin for a *long* time. This function can't 1224 * be called with interrupts off, or it may introduce deadlock with 1225 * smp_call_function() if an IPI is sent by the same process we are 1226 * waiting to become inactive. 1227 */ 1228unsigned long wait_task_inactive(struct task_struct *p, long match_state) 1229{ 1230 unsigned long flags; 1231 int running, queued; 1232 unsigned long ncsw; 1233 struct rq *rq; 1234 1235 for (;;) { 1236 /* 1237 * We do the initial early heuristics without holding 1238 * any task-queue locks at all. We'll only try to get 1239 * the runqueue lock when things look like they will 1240 * work out! 1241 */ 1242 rq = task_rq(p); 1243 1244 /* 1245 * If the task is actively running on another CPU 1246 * still, just relax and busy-wait without holding 1247 * any locks. 1248 * 1249 * NOTE! Since we don't hold any locks, it's not 1250 * even sure that "rq" stays as the right runqueue! 1251 * But we don't care, since "task_running()" will 1252 * return false if the runqueue has changed and p 1253 * is actually now running somewhere else! 1254 */ 1255 while (task_running(rq, p)) { 1256 if (match_state && unlikely(p->state != match_state)) 1257 return 0; 1258 cpu_relax(); 1259 } 1260 1261 /* 1262 * Ok, time to look more closely! We need the rq 1263 * lock now, to be *sure*. If we're wrong, we'll 1264 * just go back and repeat. 1265 */ 1266 rq = task_rq_lock(p, &flags); 1267 trace_sched_wait_task(p); 1268 running = task_running(rq, p); 1269 queued = task_on_rq_queued(p); 1270 ncsw = 0; 1271 if (!match_state || p->state == match_state) 1272 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */ 1273 task_rq_unlock(rq, p, &flags); 1274 1275 /* 1276 * If it changed from the expected state, bail out now. 1277 */ 1278 if (unlikely(!ncsw)) 1279 break; 1280 1281 /* 1282 * Was it really running after all now that we 1283 * checked with the proper locks actually held? 1284 * 1285 * Oops. Go back and try again.. 1286 */ 1287 if (unlikely(running)) { 1288 cpu_relax(); 1289 continue; 1290 } 1291 1292 /* 1293 * It's not enough that it's not actively running, 1294 * it must be off the runqueue _entirely_, and not 1295 * preempted! 1296 * 1297 * So if it was still runnable (but just not actively 1298 * running right now), it's preempted, and we should 1299 * yield - it could be a while. 1300 */ 1301 if (unlikely(queued)) { 1302 ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ); 1303 1304 set_current_state(TASK_UNINTERRUPTIBLE); 1305 schedule_hrtimeout(&to, HRTIMER_MODE_REL); 1306 continue; 1307 } 1308 1309 /* 1310 * Ahh, all good. It wasn't running, and it wasn't 1311 * runnable, which means that it will never become 1312 * running in the future either. We're all done! 1313 */ 1314 break; 1315 } 1316 1317 return ncsw; 1318} 1319 1320/*** 1321 * kick_process - kick a running thread to enter/exit the kernel 1322 * @p: the to-be-kicked thread 1323 * 1324 * Cause a process which is running on another CPU to enter 1325 * kernel-mode, without any delay. (to get signals handled.) 1326 * 1327 * NOTE: this function doesn't have to take the runqueue lock, 1328 * because all it wants to ensure is that the remote task enters 1329 * the kernel. If the IPI races and the task has been migrated 1330 * to another CPU then no harm is done and the purpose has been 1331 * achieved as well. 1332 */ 1333void kick_process(struct task_struct *p) 1334{ 1335 int cpu; 1336 1337 preempt_disable(); 1338 cpu = task_cpu(p); 1339 if ((cpu != smp_processor_id()) && task_curr(p)) 1340 smp_send_reschedule(cpu); 1341 preempt_enable(); 1342} 1343EXPORT_SYMBOL_GPL(kick_process); 1344#endif /* CONFIG_SMP */ 1345 1346#ifdef CONFIG_SMP 1347/* 1348 * ->cpus_allowed is protected by both rq->lock and p->pi_lock 1349 */ 1350static int select_fallback_rq(int cpu, struct task_struct *p) 1351{ 1352 int nid = cpu_to_node(cpu); 1353 const struct cpumask *nodemask = NULL; 1354 enum { cpuset, possible, fail } state = cpuset; 1355 int dest_cpu; 1356 1357 /* 1358 * If the node that the cpu is on has been offlined, cpu_to_node() 1359 * will return -1. There is no cpu on the node, and we should 1360 * select the cpu on the other node. 1361 */ 1362 if (nid != -1) { 1363 nodemask = cpumask_of_node(nid); 1364 1365 /* Look for allowed, online CPU in same node. */ 1366 for_each_cpu(dest_cpu, nodemask) { 1367 if (!cpu_online(dest_cpu)) 1368 continue; 1369 if (!cpu_active(dest_cpu)) 1370 continue; 1371 if (cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p))) 1372 return dest_cpu; 1373 } 1374 } 1375 1376 for (;;) { 1377 /* Any allowed, online CPU? */ 1378 for_each_cpu(dest_cpu, tsk_cpus_allowed(p)) { 1379 if (!cpu_online(dest_cpu)) 1380 continue; 1381 if (!cpu_active(dest_cpu)) 1382 continue; 1383 goto out; 1384 } 1385 1386 switch (state) { 1387 case cpuset: 1388 /* No more Mr. Nice Guy. */ 1389 cpuset_cpus_allowed_fallback(p); 1390 state = possible; 1391 break; 1392 1393 case possible: 1394 do_set_cpus_allowed(p, cpu_possible_mask); 1395 state = fail; 1396 break; 1397 1398 case fail: 1399 BUG(); 1400 break; 1401 } 1402 } 1403 1404out: 1405 if (state != cpuset) { 1406 /* 1407 * Don't tell them about moving exiting tasks or 1408 * kernel threads (both mm NULL), since they never 1409 * leave kernel. 1410 */ 1411 if (p->mm && printk_ratelimit()) { 1412 printk_deferred("process %d (%s) no longer affine to cpu%d\n", 1413 task_pid_nr(p), p->comm, cpu); 1414 } 1415 } 1416 1417 return dest_cpu; 1418} 1419 1420/* 1421 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable. 1422 */ 1423static inline 1424int select_task_rq(struct task_struct *p, int cpu, int sd_flags, int wake_flags) 1425{ 1426 cpu = p->sched_class->select_task_rq(p, cpu, sd_flags, wake_flags); 1427 1428 /* 1429 * In order not to call set_task_cpu() on a blocking task we need 1430 * to rely on ttwu() to place the task on a valid ->cpus_allowed 1431 * cpu. 1432 * 1433 * Since this is common to all placement strategies, this lives here. 1434 * 1435 * [ this allows ->select_task() to simply return task_cpu(p) and 1436 * not worry about this generic constraint ] 1437 */ 1438 if (unlikely(!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)) || 1439 !cpu_online(cpu))) 1440 cpu = select_fallback_rq(task_cpu(p), p); 1441 1442 return cpu; 1443} 1444 1445static void update_avg(u64 *avg, u64 sample) 1446{ 1447 s64 diff = sample - *avg; 1448 *avg += diff >> 3; 1449} 1450#endif 1451 1452static void 1453ttwu_stat(struct task_struct *p, int cpu, int wake_flags) 1454{ 1455#ifdef CONFIG_SCHEDSTATS 1456 struct rq *rq = this_rq(); 1457 1458#ifdef CONFIG_SMP 1459 int this_cpu = smp_processor_id(); 1460 1461 if (cpu == this_cpu) { 1462 schedstat_inc(rq, ttwu_local); 1463 schedstat_inc(p, se.statistics.nr_wakeups_local); 1464 } else { 1465 struct sched_domain *sd; 1466 1467 schedstat_inc(p, se.statistics.nr_wakeups_remote); 1468 rcu_read_lock(); 1469 for_each_domain(this_cpu, sd) { 1470 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) { 1471 schedstat_inc(sd, ttwu_wake_remote); 1472 break; 1473 } 1474 } 1475 rcu_read_unlock(); 1476 } 1477 1478 if (wake_flags & WF_MIGRATED) 1479 schedstat_inc(p, se.statistics.nr_wakeups_migrate); 1480 1481#endif /* CONFIG_SMP */ 1482 1483 schedstat_inc(rq, ttwu_count); 1484 schedstat_inc(p, se.statistics.nr_wakeups); 1485 1486 if (wake_flags & WF_SYNC) 1487 schedstat_inc(p, se.statistics.nr_wakeups_sync); 1488 1489#endif /* CONFIG_SCHEDSTATS */ 1490} 1491 1492static void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags) 1493{ 1494 activate_task(rq, p, en_flags); 1495 p->on_rq = TASK_ON_RQ_QUEUED; 1496 1497 /* if a worker is waking up, notify workqueue */ 1498 if (p->flags & PF_WQ_WORKER) 1499 wq_worker_waking_up(p, cpu_of(rq)); 1500} 1501 1502/* 1503 * Mark the task runnable and perform wakeup-preemption. 1504 */ 1505static void 1506ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags) 1507{ 1508 check_preempt_curr(rq, p, wake_flags); 1509 trace_sched_wakeup(p, true); 1510 1511 p->state = TASK_RUNNING; 1512#ifdef CONFIG_SMP 1513 if (p->sched_class->task_woken) 1514 p->sched_class->task_woken(rq, p); 1515 1516 if (rq->idle_stamp) { 1517 u64 delta = rq_clock(rq) - rq->idle_stamp; 1518 u64 max = 2*rq->max_idle_balance_cost; 1519 1520 update_avg(&rq->avg_idle, delta); 1521 1522 if (rq->avg_idle > max) 1523 rq->avg_idle = max; 1524 1525 rq->idle_stamp = 0; 1526 } 1527#endif 1528} 1529 1530static void 1531ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags) 1532{ 1533#ifdef CONFIG_SMP 1534 if (p->sched_contributes_to_load) 1535 rq->nr_uninterruptible--; 1536#endif 1537 1538 ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_WAKING); 1539 ttwu_do_wakeup(rq, p, wake_flags); 1540} 1541 1542/* 1543 * Called in case the task @p isn't fully descheduled from its runqueue, 1544 * in this case we must do a remote wakeup. Its a 'light' wakeup though, 1545 * since all we need to do is flip p->state to TASK_RUNNING, since 1546 * the task is still ->on_rq. 1547 */ 1548static int ttwu_remote(struct task_struct *p, int wake_flags) 1549{ 1550 struct rq *rq; 1551 int ret = 0; 1552 1553 rq = __task_rq_lock(p); 1554 if (task_on_rq_queued(p)) { 1555 /* check_preempt_curr() may use rq clock */ 1556 update_rq_clock(rq); 1557 ttwu_do_wakeup(rq, p, wake_flags); 1558 ret = 1; 1559 } 1560 __task_rq_unlock(rq); 1561 1562 return ret; 1563} 1564 1565#ifdef CONFIG_SMP 1566void sched_ttwu_pending(void) 1567{ 1568 struct rq *rq = this_rq(); 1569 struct llist_node *llist = llist_del_all(&rq->wake_list); 1570 struct task_struct *p; 1571 unsigned long flags; 1572 1573 if (!llist) 1574 return; 1575 1576 raw_spin_lock_irqsave(&rq->lock, flags); 1577 1578 while (llist) { 1579 p = llist_entry(llist, struct task_struct, wake_entry); 1580 llist = llist_next(llist); 1581 ttwu_do_activate(rq, p, 0); 1582 } 1583 1584 raw_spin_unlock_irqrestore(&rq->lock, flags); 1585} 1586 1587void scheduler_ipi(void) 1588{ 1589 /* 1590 * Fold TIF_NEED_RESCHED into the preempt_count; anybody setting 1591 * TIF_NEED_RESCHED remotely (for the first time) will also send 1592 * this IPI. 1593 */ 1594 preempt_fold_need_resched(); 1595 1596 if (llist_empty(&this_rq()->wake_list) && !got_nohz_idle_kick()) 1597 return; 1598 1599 /* 1600 * Not all reschedule IPI handlers call irq_enter/irq_exit, since 1601 * traditionally all their work was done from the interrupt return 1602 * path. Now that we actually do some work, we need to make sure 1603 * we do call them. 1604 * 1605 * Some archs already do call them, luckily irq_enter/exit nest 1606 * properly. 1607 * 1608 * Arguably we should visit all archs and update all handlers, 1609 * however a fair share of IPIs are still resched only so this would 1610 * somewhat pessimize the simple resched case. 1611 */ 1612 irq_enter(); 1613 sched_ttwu_pending(); 1614 1615 /* 1616 * Check if someone kicked us for doing the nohz idle load balance. 1617 */ 1618 if (unlikely(got_nohz_idle_kick())) { 1619 this_rq()->idle_balance = 1; 1620 raise_softirq_irqoff(SCHED_SOFTIRQ); 1621 } 1622 irq_exit(); 1623} 1624 1625static void ttwu_queue_remote(struct task_struct *p, int cpu) 1626{ 1627 struct rq *rq = cpu_rq(cpu); 1628 1629 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list)) { 1630 if (!set_nr_if_polling(rq->idle)) 1631 smp_send_reschedule(cpu); 1632 else 1633 trace_sched_wake_idle_without_ipi(cpu); 1634 } 1635} 1636 1637void wake_up_if_idle(int cpu) 1638{ 1639 struct rq *rq = cpu_rq(cpu); 1640 unsigned long flags; 1641 1642 if (!is_idle_task(rq->curr)) 1643 return; 1644 1645 if (set_nr_if_polling(rq->idle)) { 1646 trace_sched_wake_idle_without_ipi(cpu); 1647 } else { 1648 raw_spin_lock_irqsave(&rq->lock, flags); 1649 if (is_idle_task(rq->curr)) 1650 smp_send_reschedule(cpu); 1651 /* Else cpu is not in idle, do nothing here */ 1652 raw_spin_unlock_irqrestore(&rq->lock, flags); 1653 } 1654} 1655 1656bool cpus_share_cache(int this_cpu, int that_cpu) 1657{ 1658 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu); 1659} 1660#endif /* CONFIG_SMP */ 1661 1662static void ttwu_queue(struct task_struct *p, int cpu) 1663{ 1664 struct rq *rq = cpu_rq(cpu); 1665 1666#if defined(CONFIG_SMP) 1667 if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) { 1668 sched_clock_cpu(cpu); /* sync clocks x-cpu */ 1669 ttwu_queue_remote(p, cpu); 1670 return; 1671 } 1672#endif 1673 1674 raw_spin_lock(&rq->lock); 1675 ttwu_do_activate(rq, p, 0); 1676 raw_spin_unlock(&rq->lock); 1677} 1678 1679/** 1680 * try_to_wake_up - wake up a thread 1681 * @p: the thread to be awakened 1682 * @state: the mask of task states that can be woken 1683 * @wake_flags: wake modifier flags (WF_*) 1684 * 1685 * Put it on the run-queue if it's not already there. The "current" 1686 * thread is always on the run-queue (except when the actual 1687 * re-schedule is in progress), and as such you're allowed to do 1688 * the simpler "current->state = TASK_RUNNING" to mark yourself 1689 * runnable without the overhead of this. 1690 * 1691 * Return: %true if @p was woken up, %false if it was already running. 1692 * or @state didn't match @p's state. 1693 */ 1694static int 1695try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags) 1696{ 1697 unsigned long flags; 1698 int cpu, success = 0; 1699 1700 /* 1701 * If we are going to wake up a thread waiting for CONDITION we 1702 * need to ensure that CONDITION=1 done by the caller can not be 1703 * reordered with p->state check below. This pairs with mb() in 1704 * set_current_state() the waiting thread does. 1705 */ 1706 smp_mb__before_spinlock(); 1707 raw_spin_lock_irqsave(&p->pi_lock, flags); 1708 if (!(p->state & state)) 1709 goto out; 1710 1711 success = 1; /* we're going to change ->state */ 1712 cpu = task_cpu(p); 1713 1714 if (p->on_rq && ttwu_remote(p, wake_flags)) 1715 goto stat; 1716 1717#ifdef CONFIG_SMP 1718 /* 1719 * If the owning (remote) cpu is still in the middle of schedule() with 1720 * this task as prev, wait until its done referencing the task. 1721 */ 1722 while (p->on_cpu) 1723 cpu_relax(); 1724 /* 1725 * Pairs with the smp_wmb() in finish_lock_switch(). 1726 */ 1727 smp_rmb(); 1728 1729 p->sched_contributes_to_load = !!task_contributes_to_load(p); 1730 p->state = TASK_WAKING; 1731 1732 if (p->sched_class->task_waking) 1733 p->sched_class->task_waking(p); 1734 1735 cpu = select_task_rq(p, p->wake_cpu, SD_BALANCE_WAKE, wake_flags); 1736 if (task_cpu(p) != cpu) { 1737 wake_flags |= WF_MIGRATED; 1738 set_task_cpu(p, cpu); 1739 } 1740#endif /* CONFIG_SMP */ 1741 1742 ttwu_queue(p, cpu); 1743stat: 1744 ttwu_stat(p, cpu, wake_flags); 1745out: 1746 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 1747 1748 return success; 1749} 1750 1751/** 1752 * try_to_wake_up_local - try to wake up a local task with rq lock held 1753 * @p: the thread to be awakened 1754 * 1755 * Put @p on the run-queue if it's not already there. The caller must 1756 * ensure that this_rq() is locked, @p is bound to this_rq() and not 1757 * the current task. 1758 */ 1759static void try_to_wake_up_local(struct task_struct *p) 1760{ 1761 struct rq *rq = task_rq(p); 1762 1763 if (WARN_ON_ONCE(rq != this_rq()) || 1764 WARN_ON_ONCE(p == current)) 1765 return; 1766 1767 lockdep_assert_held(&rq->lock); 1768 1769 if (!raw_spin_trylock(&p->pi_lock)) { 1770 raw_spin_unlock(&rq->lock); 1771 raw_spin_lock(&p->pi_lock); 1772 raw_spin_lock(&rq->lock); 1773 } 1774 1775 if (!(p->state & TASK_NORMAL)) 1776 goto out; 1777 1778 if (!task_on_rq_queued(p)) 1779 ttwu_activate(rq, p, ENQUEUE_WAKEUP); 1780 1781 ttwu_do_wakeup(rq, p, 0); 1782 ttwu_stat(p, smp_processor_id(), 0); 1783out: 1784 raw_spin_unlock(&p->pi_lock); 1785} 1786 1787/** 1788 * wake_up_process - Wake up a specific process 1789 * @p: The process to be woken up. 1790 * 1791 * Attempt to wake up the nominated process and move it to the set of runnable 1792 * processes. 1793 * 1794 * Return: 1 if the process was woken up, 0 if it was already running. 1795 * 1796 * It may be assumed that this function implies a write memory barrier before 1797 * changing the task state if and only if any tasks are woken up. 1798 */ 1799int wake_up_process(struct task_struct *p) 1800{ 1801 WARN_ON(task_is_stopped_or_traced(p)); 1802 return try_to_wake_up(p, TASK_NORMAL, 0); 1803} 1804EXPORT_SYMBOL(wake_up_process); 1805 1806int wake_up_state(struct task_struct *p, unsigned int state) 1807{ 1808 return try_to_wake_up(p, state, 0); 1809} 1810 1811/* 1812 * This function clears the sched_dl_entity static params. 1813 */ 1814void __dl_clear_params(struct task_struct *p) 1815{ 1816 struct sched_dl_entity *dl_se = &p->dl; 1817 1818 dl_se->dl_runtime = 0; 1819 dl_se->dl_deadline = 0; 1820 dl_se->dl_period = 0; 1821 dl_se->flags = 0; 1822 dl_se->dl_bw = 0; 1823} 1824 1825/* 1826 * Perform scheduler related setup for a newly forked process p. 1827 * p is forked by current. 1828 * 1829 * __sched_fork() is basic setup used by init_idle() too: 1830 */ 1831static void __sched_fork(unsigned long clone_flags, struct task_struct *p) 1832{ 1833 p->on_rq = 0; 1834 1835 p->se.on_rq = 0; 1836 p->se.exec_start = 0; 1837 p->se.sum_exec_runtime = 0; 1838 p->se.prev_sum_exec_runtime = 0; 1839 p->se.nr_migrations = 0; 1840 p->se.vruntime = 0; 1841 INIT_LIST_HEAD(&p->se.group_node); 1842 1843#ifdef CONFIG_SCHEDSTATS 1844 memset(&p->se.statistics, 0, sizeof(p->se.statistics)); 1845#endif 1846 1847 RB_CLEAR_NODE(&p->dl.rb_node); 1848 hrtimer_init(&p->dl.dl_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); 1849 __dl_clear_params(p); 1850 1851 INIT_LIST_HEAD(&p->rt.run_list); 1852 1853#ifdef CONFIG_PREEMPT_NOTIFIERS 1854 INIT_HLIST_HEAD(&p->preempt_notifiers); 1855#endif 1856 1857#ifdef CONFIG_NUMA_BALANCING 1858 if (p->mm && atomic_read(&p->mm->mm_users) == 1) { 1859 p->mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay); 1860 p->mm->numa_scan_seq = 0; 1861 } 1862 1863 if (clone_flags & CLONE_VM) 1864 p->numa_preferred_nid = current->numa_preferred_nid; 1865 else 1866 p->numa_preferred_nid = -1; 1867 1868 p->node_stamp = 0ULL; 1869 p->numa_scan_seq = p->mm ? p->mm->numa_scan_seq : 0; 1870 p->numa_scan_period = sysctl_numa_balancing_scan_delay; 1871 p->numa_work.next = &p->numa_work; 1872 p->numa_faults_memory = NULL; 1873 p->numa_faults_buffer_memory = NULL; 1874 p->last_task_numa_placement = 0; 1875 p->last_sum_exec_runtime = 0; 1876 1877 INIT_LIST_HEAD(&p->numa_entry); 1878 p->numa_group = NULL; 1879#endif /* CONFIG_NUMA_BALANCING */ 1880} 1881 1882#ifdef CONFIG_NUMA_BALANCING 1883#ifdef CONFIG_SCHED_DEBUG 1884void set_numabalancing_state(bool enabled) 1885{ 1886 if (enabled) 1887 sched_feat_set("NUMA"); 1888 else 1889 sched_feat_set("NO_NUMA"); 1890} 1891#else 1892__read_mostly bool numabalancing_enabled; 1893 1894void set_numabalancing_state(bool enabled) 1895{ 1896 numabalancing_enabled = enabled; 1897} 1898#endif /* CONFIG_SCHED_DEBUG */ 1899 1900#ifdef CONFIG_PROC_SYSCTL 1901int sysctl_numa_balancing(struct ctl_table *table, int write, 1902 void __user *buffer, size_t *lenp, loff_t *ppos) 1903{ 1904 struct ctl_table t; 1905 int err; 1906 int state = numabalancing_enabled; 1907 1908 if (write && !capable(CAP_SYS_ADMIN)) 1909 return -EPERM; 1910 1911 t = *table; 1912 t.data = &state; 1913 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos); 1914 if (err < 0) 1915 return err; 1916 if (write) 1917 set_numabalancing_state(state); 1918 return err; 1919} 1920#endif 1921#endif 1922 1923/* 1924 * fork()/clone()-time setup: 1925 */ 1926int sched_fork(unsigned long clone_flags, struct task_struct *p) 1927{ 1928 unsigned long flags; 1929 int cpu = get_cpu(); 1930 1931 __sched_fork(clone_flags, p); 1932 /* 1933 * We mark the process as running here. This guarantees that 1934 * nobody will actually run it, and a signal or other external 1935 * event cannot wake it up and insert it on the runqueue either. 1936 */ 1937 p->state = TASK_RUNNING; 1938 1939 /* 1940 * Make sure we do not leak PI boosting priority to the child. 1941 */ 1942 p->prio = current->normal_prio; 1943 1944 /* 1945 * Revert to default priority/policy on fork if requested. 1946 */ 1947 if (unlikely(p->sched_reset_on_fork)) { 1948 if (task_has_dl_policy(p) || task_has_rt_policy(p)) { 1949 p->policy = SCHED_NORMAL; 1950 p->static_prio = NICE_TO_PRIO(0); 1951 p->rt_priority = 0; 1952 } else if (PRIO_TO_NICE(p->static_prio) < 0) 1953 p->static_prio = NICE_TO_PRIO(0); 1954 1955 p->prio = p->normal_prio = __normal_prio(p); 1956 set_load_weight(p); 1957 1958 /* 1959 * We don't need the reset flag anymore after the fork. It has 1960 * fulfilled its duty: 1961 */ 1962 p->sched_reset_on_fork = 0; 1963 } 1964 1965 if (dl_prio(p->prio)) { 1966 put_cpu(); 1967 return -EAGAIN; 1968 } else if (rt_prio(p->prio)) { 1969 p->sched_class = &rt_sched_class; 1970 } else { 1971 p->sched_class = &fair_sched_class; 1972 } 1973 1974 if (p->sched_class->task_fork) 1975 p->sched_class->task_fork(p); 1976 1977 /* 1978 * The child is not yet in the pid-hash so no cgroup attach races, 1979 * and the cgroup is pinned to this child due to cgroup_fork() 1980 * is ran before sched_fork(). 1981 * 1982 * Silence PROVE_RCU. 1983 */ 1984 raw_spin_lock_irqsave(&p->pi_lock, flags); 1985 set_task_cpu(p, cpu); 1986 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 1987 1988#if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT) 1989 if (likely(sched_info_on())) 1990 memset(&p->sched_info, 0, sizeof(p->sched_info)); 1991#endif 1992#if defined(CONFIG_SMP) 1993 p->on_cpu = 0; 1994#endif 1995 init_task_preempt_count(p); 1996#ifdef CONFIG_SMP 1997 plist_node_init(&p->pushable_tasks, MAX_PRIO); 1998 RB_CLEAR_NODE(&p->pushable_dl_tasks); 1999#endif 2000 2001 put_cpu(); 2002 return 0; 2003} 2004 2005unsigned long to_ratio(u64 period, u64 runtime) 2006{ 2007 if (runtime == RUNTIME_INF) 2008 return 1ULL << 20; 2009 2010 /* 2011 * Doing this here saves a lot of checks in all 2012 * the calling paths, and returning zero seems 2013 * safe for them anyway. 2014 */ 2015 if (period == 0) 2016 return 0; 2017 2018 return div64_u64(runtime << 20, period); 2019} 2020 2021#ifdef CONFIG_SMP 2022inline struct dl_bw *dl_bw_of(int i) 2023{ 2024 return &cpu_rq(i)->rd->dl_bw; 2025} 2026 2027static inline int dl_bw_cpus(int i) 2028{ 2029 struct root_domain *rd = cpu_rq(i)->rd; 2030 int cpus = 0; 2031 2032 for_each_cpu_and(i, rd->span, cpu_active_mask) 2033 cpus++; 2034 2035 return cpus; 2036} 2037#else 2038inline struct dl_bw *dl_bw_of(int i) 2039{ 2040 return &cpu_rq(i)->dl.dl_bw; 2041} 2042 2043static inline int dl_bw_cpus(int i) 2044{ 2045 return 1; 2046} 2047#endif 2048 2049static inline 2050void __dl_clear(struct dl_bw *dl_b, u64 tsk_bw) 2051{ 2052 dl_b->total_bw -= tsk_bw; 2053} 2054 2055static inline 2056void __dl_add(struct dl_bw *dl_b, u64 tsk_bw) 2057{ 2058 dl_b->total_bw += tsk_bw; 2059} 2060 2061static inline 2062bool __dl_overflow(struct dl_bw *dl_b, int cpus, u64 old_bw, u64 new_bw) 2063{ 2064 return dl_b->bw != -1 && 2065 dl_b->bw * cpus < dl_b->total_bw - old_bw + new_bw; 2066} 2067 2068/* 2069 * We must be sure that accepting a new task (or allowing changing the 2070 * parameters of an existing one) is consistent with the bandwidth 2071 * constraints. If yes, this function also accordingly updates the currently 2072 * allocated bandwidth to reflect the new situation. 2073 * 2074 * This function is called while holding p's rq->lock. 2075 */ 2076static int dl_overflow(struct task_struct *p, int policy, 2077 const struct sched_attr *attr) 2078{ 2079 2080 struct dl_bw *dl_b = dl_bw_of(task_cpu(p)); 2081 u64 period = attr->sched_period ?: attr->sched_deadline; 2082 u64 runtime = attr->sched_runtime; 2083 u64 new_bw = dl_policy(policy) ? to_ratio(period, runtime) : 0; 2084 int cpus, err = -1; 2085 2086 if (new_bw == p->dl.dl_bw) 2087 return 0; 2088 2089 /* 2090 * Either if a task, enters, leave, or stays -deadline but changes 2091 * its parameters, we may need to update accordingly the total 2092 * allocated bandwidth of the container. 2093 */ 2094 raw_spin_lock(&dl_b->lock); 2095 cpus = dl_bw_cpus(task_cpu(p)); 2096 if (dl_policy(policy) && !task_has_dl_policy(p) && 2097 !__dl_overflow(dl_b, cpus, 0, new_bw)) { 2098 __dl_add(dl_b, new_bw); 2099 err = 0; 2100 } else if (dl_policy(policy) && task_has_dl_policy(p) && 2101 !__dl_overflow(dl_b, cpus, p->dl.dl_bw, new_bw)) { 2102 __dl_clear(dl_b, p->dl.dl_bw); 2103 __dl_add(dl_b, new_bw); 2104 err = 0; 2105 } else if (!dl_policy(policy) && task_has_dl_policy(p)) { 2106 __dl_clear(dl_b, p->dl.dl_bw); 2107 err = 0; 2108 } 2109 raw_spin_unlock(&dl_b->lock); 2110 2111 return err; 2112} 2113 2114extern void init_dl_bw(struct dl_bw *dl_b); 2115 2116/* 2117 * wake_up_new_task - wake up a newly created task for the first time. 2118 * 2119 * This function will do some initial scheduler statistics housekeeping 2120 * that must be done for every newly created context, then puts the task 2121 * on the runqueue and wakes it. 2122 */ 2123void wake_up_new_task(struct task_struct *p) 2124{ 2125 unsigned long flags; 2126 struct rq *rq; 2127 2128 raw_spin_lock_irqsave(&p->pi_lock, flags); 2129#ifdef CONFIG_SMP 2130 /* 2131 * Fork balancing, do it here and not earlier because: 2132 * - cpus_allowed can change in the fork path 2133 * - any previously selected cpu might disappear through hotplug 2134 */ 2135 set_task_cpu(p, select_task_rq(p, task_cpu(p), SD_BALANCE_FORK, 0)); 2136#endif 2137 2138 /* Initialize new task's runnable average */ 2139 init_task_runnable_average(p); 2140 rq = __task_rq_lock(p); 2141 activate_task(rq, p, 0); 2142 p->on_rq = TASK_ON_RQ_QUEUED; 2143 trace_sched_wakeup_new(p, true); 2144 check_preempt_curr(rq, p, WF_FORK); 2145#ifdef CONFIG_SMP 2146 if (p->sched_class->task_woken) 2147 p->sched_class->task_woken(rq, p); 2148#endif 2149 task_rq_unlock(rq, p, &flags); 2150} 2151 2152#ifdef CONFIG_PREEMPT_NOTIFIERS 2153 2154/** 2155 * preempt_notifier_register - tell me when current is being preempted & rescheduled 2156 * @notifier: notifier struct to register 2157 */ 2158void preempt_notifier_register(struct preempt_notifier *notifier) 2159{ 2160 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers); 2161} 2162EXPORT_SYMBOL_GPL(preempt_notifier_register); 2163 2164/** 2165 * preempt_notifier_unregister - no longer interested in preemption notifications 2166 * @notifier: notifier struct to unregister 2167 * 2168 * This is safe to call from within a preemption notifier. 2169 */ 2170void preempt_notifier_unregister(struct preempt_notifier *notifier) 2171{ 2172 hlist_del(¬ifier->link); 2173} 2174EXPORT_SYMBOL_GPL(preempt_notifier_unregister); 2175 2176static void fire_sched_in_preempt_notifiers(struct task_struct *curr) 2177{ 2178 struct preempt_notifier *notifier; 2179 2180 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link) 2181 notifier->ops->sched_in(notifier, raw_smp_processor_id()); 2182} 2183 2184static void 2185fire_sched_out_preempt_notifiers(struct task_struct *curr, 2186 struct task_struct *next) 2187{ 2188 struct preempt_notifier *notifier; 2189 2190 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link) 2191 notifier->ops->sched_out(notifier, next); 2192} 2193 2194#else /* !CONFIG_PREEMPT_NOTIFIERS */ 2195 2196static void fire_sched_in_preempt_notifiers(struct task_struct *curr) 2197{ 2198} 2199 2200static void 2201fire_sched_out_preempt_notifiers(struct task_struct *curr, 2202 struct task_struct *next) 2203{ 2204} 2205 2206#endif /* CONFIG_PREEMPT_NOTIFIERS */ 2207 2208/** 2209 * prepare_task_switch - prepare to switch tasks 2210 * @rq: the runqueue preparing to switch 2211 * @prev: the current task that is being switched out 2212 * @next: the task we are going to switch to. 2213 * 2214 * This is called with the rq lock held and interrupts off. It must 2215 * be paired with a subsequent finish_task_switch after the context 2216 * switch. 2217 * 2218 * prepare_task_switch sets up locking and calls architecture specific 2219 * hooks. 2220 */ 2221static inline void 2222prepare_task_switch(struct rq *rq, struct task_struct *prev, 2223 struct task_struct *next) 2224{ 2225 trace_sched_switch(prev, next); 2226 sched_info_switch(rq, prev, next); 2227 perf_event_task_sched_out(prev, next); 2228 fire_sched_out_preempt_notifiers(prev, next); 2229 prepare_lock_switch(rq, next); 2230 prepare_arch_switch(next); 2231} 2232 2233/** 2234 * finish_task_switch - clean up after a task-switch 2235 * @rq: runqueue associated with task-switch 2236 * @prev: the thread we just switched away from. 2237 * 2238 * finish_task_switch must be called after the context switch, paired 2239 * with a prepare_task_switch call before the context switch. 2240 * finish_task_switch will reconcile locking set up by prepare_task_switch, 2241 * and do any other architecture-specific cleanup actions. 2242 * 2243 * Note that we may have delayed dropping an mm in context_switch(). If 2244 * so, we finish that here outside of the runqueue lock. (Doing it 2245 * with the lock held can cause deadlocks; see schedule() for 2246 * details.) 2247 */ 2248static void finish_task_switch(struct rq *rq, struct task_struct *prev) 2249 __releases(rq->lock) 2250{ 2251 struct mm_struct *mm = rq->prev_mm; 2252 long prev_state; 2253 2254 rq->prev_mm = NULL; 2255 2256 /* 2257 * A task struct has one reference for the use as "current". 2258 * If a task dies, then it sets TASK_DEAD in tsk->state and calls 2259 * schedule one last time. The schedule call will never return, and 2260 * the scheduled task must drop that reference. 2261 * The test for TASK_DEAD must occur while the runqueue locks are 2262 * still held, otherwise prev could be scheduled on another cpu, die 2263 * there before we look at prev->state, and then the reference would 2264 * be dropped twice. 2265 * Manfred Spraul <manfred@colorfullife.com> 2266 */ 2267 prev_state = prev->state; 2268 vtime_task_switch(prev); 2269 finish_arch_switch(prev); 2270 perf_event_task_sched_in(prev, current); 2271 finish_lock_switch(rq, prev); 2272 finish_arch_post_lock_switch(); 2273 2274 fire_sched_in_preempt_notifiers(current); 2275 if (mm) 2276 mmdrop(mm); 2277 if (unlikely(prev_state == TASK_DEAD)) { 2278 if (prev->sched_class->task_dead) 2279 prev->sched_class->task_dead(prev); 2280 2281 /* 2282 * Remove function-return probe instances associated with this 2283 * task and put them back on the free list. 2284 */ 2285 kprobe_flush_task(prev); 2286 put_task_struct(prev); 2287 } 2288 2289 tick_nohz_task_switch(current); 2290} 2291 2292#ifdef CONFIG_SMP 2293 2294/* rq->lock is NOT held, but preemption is disabled */ 2295static inline void post_schedule(struct rq *rq) 2296{ 2297 if (rq->post_schedule) { 2298 unsigned long flags; 2299 2300 raw_spin_lock_irqsave(&rq->lock, flags); 2301 if (rq->curr->sched_class->post_schedule) 2302 rq->curr->sched_class->post_schedule(rq); 2303 raw_spin_unlock_irqrestore(&rq->lock, flags); 2304 2305 rq->post_schedule = 0; 2306 } 2307} 2308 2309#else 2310 2311static inline void post_schedule(struct rq *rq) 2312{ 2313} 2314 2315#endif 2316 2317/** 2318 * schedule_tail - first thing a freshly forked thread must call. 2319 * @prev: the thread we just switched away from. 2320 */ 2321asmlinkage __visible void schedule_tail(struct task_struct *prev) 2322 __releases(rq->lock) 2323{ 2324 struct rq *rq = this_rq(); 2325 2326 finish_task_switch(rq, prev); 2327 2328 /* 2329 * FIXME: do we need to worry about rq being invalidated by the 2330 * task_switch? 2331 */ 2332 post_schedule(rq); 2333 2334#ifdef __ARCH_WANT_UNLOCKED_CTXSW 2335 /* In this case, finish_task_switch does not reenable preemption */ 2336 preempt_enable(); 2337#endif 2338 if (current->set_child_tid) 2339 put_user(task_pid_vnr(current), current->set_child_tid); 2340} 2341 2342/* 2343 * context_switch - switch to the new MM and the new 2344 * thread's register state. 2345 */ 2346static inline void 2347context_switch(struct rq *rq, struct task_struct *prev, 2348 struct task_struct *next) 2349{ 2350 struct mm_struct *mm, *oldmm; 2351 2352 prepare_task_switch(rq, prev, next); 2353 2354 mm = next->mm; 2355 oldmm = prev->active_mm; 2356 /* 2357 * For paravirt, this is coupled with an exit in switch_to to 2358 * combine the page table reload and the switch backend into 2359 * one hypercall. 2360 */ 2361 arch_start_context_switch(prev); 2362 2363 if (!mm) { 2364 next->active_mm = oldmm; 2365 atomic_inc(&oldmm->mm_count); 2366 enter_lazy_tlb(oldmm, next); 2367 } else 2368 switch_mm(oldmm, mm, next); 2369 2370 if (!prev->mm) { 2371 prev->active_mm = NULL; 2372 rq->prev_mm = oldmm; 2373 } 2374 /* 2375 * Since the runqueue lock will be released by the next 2376 * task (which is an invalid locking op but in the case 2377 * of the scheduler it's an obvious special-case), so we 2378 * do an early lockdep release here: 2379 */ 2380#ifndef __ARCH_WANT_UNLOCKED_CTXSW 2381 spin_release(&rq->lock.dep_map, 1, _THIS_IP_); 2382#endif 2383 2384 context_tracking_task_switch(prev, next); 2385 /* Here we just switch the register state and the stack. */ 2386 switch_to(prev, next, prev); 2387 2388 barrier(); 2389 /* 2390 * this_rq must be evaluated again because prev may have moved 2391 * CPUs since it called schedule(), thus the 'rq' on its stack 2392 * frame will be invalid. 2393 */ 2394 finish_task_switch(this_rq(), prev); 2395} 2396 2397/* 2398 * nr_running and nr_context_switches: 2399 * 2400 * externally visible scheduler statistics: current number of runnable 2401 * threads, total number of context switches performed since bootup. 2402 */ 2403unsigned long nr_running(void) 2404{ 2405 unsigned long i, sum = 0; 2406 2407 for_each_online_cpu(i) 2408 sum += cpu_rq(i)->nr_running; 2409 2410 return sum; 2411} 2412 2413unsigned long long nr_context_switches(void) 2414{ 2415 int i; 2416 unsigned long long sum = 0; 2417 2418 for_each_possible_cpu(i) 2419 sum += cpu_rq(i)->nr_switches; 2420 2421 return sum; 2422} 2423 2424unsigned long nr_iowait(void) 2425{ 2426 unsigned long i, sum = 0; 2427 2428 for_each_possible_cpu(i) 2429 sum += atomic_read(&cpu_rq(i)->nr_iowait); 2430 2431 return sum; 2432} 2433 2434unsigned long nr_iowait_cpu(int cpu) 2435{ 2436 struct rq *this = cpu_rq(cpu); 2437 return atomic_read(&this->nr_iowait); 2438} 2439 2440void get_iowait_load(unsigned long *nr_waiters, unsigned long *load) 2441{ 2442 struct rq *this = this_rq(); 2443 *nr_waiters = atomic_read(&this->nr_iowait); 2444 *load = this->cpu_load[0]; 2445} 2446 2447#ifdef CONFIG_SMP 2448 2449/* 2450 * sched_exec - execve() is a valuable balancing opportunity, because at 2451 * this point the task has the smallest effective memory and cache footprint. 2452 */ 2453void sched_exec(void) 2454{ 2455 struct task_struct *p = current; 2456 unsigned long flags; 2457 int dest_cpu; 2458 2459 raw_spin_lock_irqsave(&p->pi_lock, flags); 2460 dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), SD_BALANCE_EXEC, 0); 2461 if (dest_cpu == smp_processor_id()) 2462 goto unlock; 2463 2464 if (likely(cpu_active(dest_cpu))) { 2465 struct migration_arg arg = { p, dest_cpu }; 2466 2467 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 2468 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg); 2469 return; 2470 } 2471unlock: 2472 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 2473} 2474 2475#endif 2476 2477DEFINE_PER_CPU(struct kernel_stat, kstat); 2478DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat); 2479 2480EXPORT_PER_CPU_SYMBOL(kstat); 2481EXPORT_PER_CPU_SYMBOL(kernel_cpustat); 2482 2483/* 2484 * Return any ns on the sched_clock that have not yet been accounted in 2485 * @p in case that task is currently running. 2486 * 2487 * Called with task_rq_lock() held on @rq. 2488 */ 2489static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq) 2490{ 2491 u64 ns = 0; 2492 2493 /* 2494 * Must be ->curr _and_ ->on_rq. If dequeued, we would 2495 * project cycles that may never be accounted to this 2496 * thread, breaking clock_gettime(). 2497 */ 2498 if (task_current(rq, p) && task_on_rq_queued(p)) { 2499 update_rq_clock(rq); 2500 ns = rq_clock_task(rq) - p->se.exec_start; 2501 if ((s64)ns < 0) 2502 ns = 0; 2503 } 2504 2505 return ns; 2506} 2507 2508unsigned long long task_delta_exec(struct task_struct *p) 2509{ 2510 unsigned long flags; 2511 struct rq *rq; 2512 u64 ns = 0; 2513 2514 rq = task_rq_lock(p, &flags); 2515 ns = do_task_delta_exec(p, rq); 2516 task_rq_unlock(rq, p, &flags); 2517 2518 return ns; 2519} 2520 2521/* 2522 * Return accounted runtime for the task. 2523 * In case the task is currently running, return the runtime plus current's 2524 * pending runtime that have not been accounted yet. 2525 */ 2526unsigned long long task_sched_runtime(struct task_struct *p) 2527{ 2528 unsigned long flags; 2529 struct rq *rq; 2530 u64 ns = 0; 2531 2532#if defined(CONFIG_64BIT) && defined(CONFIG_SMP) 2533 /* 2534 * 64-bit doesn't need locks to atomically read a 64bit value. 2535 * So we have a optimization chance when the task's delta_exec is 0. 2536 * Reading ->on_cpu is racy, but this is ok. 2537 * 2538 * If we race with it leaving cpu, we'll take a lock. So we're correct. 2539 * If we race with it entering cpu, unaccounted time is 0. This is 2540 * indistinguishable from the read occurring a few cycles earlier. 2541 * If we see ->on_cpu without ->on_rq, the task is leaving, and has 2542 * been accounted, so we're correct here as well. 2543 */ 2544 if (!p->on_cpu || !task_on_rq_queued(p)) 2545 return p->se.sum_exec_runtime; 2546#endif 2547 2548 rq = task_rq_lock(p, &flags); 2549 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq); 2550 task_rq_unlock(rq, p, &flags); 2551 2552 return ns; 2553} 2554 2555/* 2556 * This function gets called by the timer code, with HZ frequency. 2557 * We call it with interrupts disabled. 2558 */ 2559void scheduler_tick(void) 2560{ 2561 int cpu = smp_processor_id(); 2562 struct rq *rq = cpu_rq(cpu); 2563 struct task_struct *curr = rq->curr; 2564 2565 sched_clock_tick(); 2566 2567 raw_spin_lock(&rq->lock); 2568 update_rq_clock(rq); 2569 curr->sched_class->task_tick(rq, curr, 0); 2570 update_cpu_load_active(rq); 2571 raw_spin_unlock(&rq->lock); 2572 2573 perf_event_task_tick(); 2574 2575#ifdef CONFIG_SMP 2576 rq->idle_balance = idle_cpu(cpu); 2577 trigger_load_balance(rq); 2578#endif 2579 rq_last_tick_reset(rq); 2580} 2581 2582#ifdef CONFIG_NO_HZ_FULL 2583/** 2584 * scheduler_tick_max_deferment 2585 * 2586 * Keep at least one tick per second when a single 2587 * active task is running because the scheduler doesn't 2588 * yet completely support full dynticks environment. 2589 * 2590 * This makes sure that uptime, CFS vruntime, load 2591 * balancing, etc... continue to move forward, even 2592 * with a very low granularity. 2593 * 2594 * Return: Maximum deferment in nanoseconds. 2595 */ 2596u64 scheduler_tick_max_deferment(void) 2597{ 2598 struct rq *rq = this_rq(); 2599 unsigned long next, now = ACCESS_ONCE(jiffies); 2600 2601 next = rq->last_sched_tick + HZ; 2602 2603 if (time_before_eq(next, now)) 2604 return 0; 2605 2606 return jiffies_to_nsecs(next - now); 2607} 2608#endif 2609 2610notrace unsigned long get_parent_ip(unsigned long addr) 2611{ 2612 if (in_lock_functions(addr)) { 2613 addr = CALLER_ADDR2; 2614 if (in_lock_functions(addr)) 2615 addr = CALLER_ADDR3; 2616 } 2617 return addr; 2618} 2619 2620#if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \ 2621 defined(CONFIG_PREEMPT_TRACER)) 2622 2623void preempt_count_add(int val) 2624{ 2625#ifdef CONFIG_DEBUG_PREEMPT 2626 /* 2627 * Underflow? 2628 */ 2629 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0))) 2630 return; 2631#endif 2632 __preempt_count_add(val); 2633#ifdef CONFIG_DEBUG_PREEMPT 2634 /* 2635 * Spinlock count overflowing soon? 2636 */ 2637 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >= 2638 PREEMPT_MASK - 10); 2639#endif 2640 if (preempt_count() == val) { 2641 unsigned long ip = get_parent_ip(CALLER_ADDR1); 2642#ifdef CONFIG_DEBUG_PREEMPT 2643 current->preempt_disable_ip = ip; 2644#endif 2645 trace_preempt_off(CALLER_ADDR0, ip); 2646 } 2647} 2648EXPORT_SYMBOL(preempt_count_add); 2649NOKPROBE_SYMBOL(preempt_count_add); 2650 2651void preempt_count_sub(int val) 2652{ 2653#ifdef CONFIG_DEBUG_PREEMPT 2654 /* 2655 * Underflow? 2656 */ 2657 if (DEBUG_LOCKS_WARN_ON(val > preempt_count())) 2658 return; 2659 /* 2660 * Is the spinlock portion underflowing? 2661 */ 2662 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) && 2663 !(preempt_count() & PREEMPT_MASK))) 2664 return; 2665#endif 2666 2667 if (preempt_count() == val) 2668 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1)); 2669 __preempt_count_sub(val); 2670} 2671EXPORT_SYMBOL(preempt_count_sub); 2672NOKPROBE_SYMBOL(preempt_count_sub); 2673 2674#endif 2675 2676/* 2677 * Print scheduling while atomic bug: 2678 */ 2679static noinline void __schedule_bug(struct task_struct *prev) 2680{ 2681 if (oops_in_progress) 2682 return; 2683 2684 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n", 2685 prev->comm, prev->pid, preempt_count()); 2686 2687 debug_show_held_locks(prev); 2688 print_modules(); 2689 if (irqs_disabled()) 2690 print_irqtrace_events(prev); 2691#ifdef CONFIG_DEBUG_PREEMPT 2692 if (in_atomic_preempt_off()) { 2693 pr_err("Preemption disabled at:"); 2694 print_ip_sym(current->preempt_disable_ip); 2695 pr_cont("\n"); 2696 } 2697#endif 2698 dump_stack(); 2699 add_taint(TAINT_WARN, LOCKDEP_STILL_OK); 2700} 2701 2702/* 2703 * Various schedule()-time debugging checks and statistics: 2704 */ 2705static inline void schedule_debug(struct task_struct *prev) 2706{ 2707#ifdef CONFIG_SCHED_STACK_END_CHECK 2708 BUG_ON(unlikely(task_stack_end_corrupted(prev))); 2709#endif 2710 /* 2711 * Test if we are atomic. Since do_exit() needs to call into 2712 * schedule() atomically, we ignore that path. Otherwise whine 2713 * if we are scheduling when we should not. 2714 */ 2715 if (unlikely(in_atomic_preempt_off() && prev->state != TASK_DEAD)) 2716 __schedule_bug(prev); 2717 rcu_sleep_check(); 2718 2719 profile_hit(SCHED_PROFILING, __builtin_return_address(0)); 2720 2721 schedstat_inc(this_rq(), sched_count); 2722} 2723 2724/* 2725 * Pick up the highest-prio task: 2726 */ 2727static inline struct task_struct * 2728pick_next_task(struct rq *rq, struct task_struct *prev) 2729{ 2730 const struct sched_class *class = &fair_sched_class; 2731 struct task_struct *p; 2732 2733 /* 2734 * Optimization: we know that if all tasks are in 2735 * the fair class we can call that function directly: 2736 */ 2737 if (likely(prev->sched_class == class && 2738 rq->nr_running == rq->cfs.h_nr_running)) { 2739 p = fair_sched_class.pick_next_task(rq, prev); 2740 if (unlikely(p == RETRY_TASK)) 2741 goto again; 2742 2743 /* assumes fair_sched_class->next == idle_sched_class */ 2744 if (unlikely(!p)) 2745 p = idle_sched_class.pick_next_task(rq, prev); 2746 2747 return p; 2748 } 2749 2750again: 2751 for_each_class(class) { 2752 p = class->pick_next_task(rq, prev); 2753 if (p) { 2754 if (unlikely(p == RETRY_TASK)) 2755 goto again; 2756 return p; 2757 } 2758 } 2759 2760 BUG(); /* the idle class will always have a runnable task */ 2761} 2762 2763/* 2764 * __schedule() is the main scheduler function. 2765 * 2766 * The main means of driving the scheduler and thus entering this function are: 2767 * 2768 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc. 2769 * 2770 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return 2771 * paths. For example, see arch/x86/entry_64.S. 2772 * 2773 * To drive preemption between tasks, the scheduler sets the flag in timer 2774 * interrupt handler scheduler_tick(). 2775 * 2776 * 3. Wakeups don't really cause entry into schedule(). They add a 2777 * task to the run-queue and that's it. 2778 * 2779 * Now, if the new task added to the run-queue preempts the current 2780 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets 2781 * called on the nearest possible occasion: 2782 * 2783 * - If the kernel is preemptible (CONFIG_PREEMPT=y): 2784 * 2785 * - in syscall or exception context, at the next outmost 2786 * preempt_enable(). (this might be as soon as the wake_up()'s 2787 * spin_unlock()!) 2788 * 2789 * - in IRQ context, return from interrupt-handler to 2790 * preemptible context 2791 * 2792 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set) 2793 * then at the next: 2794 * 2795 * - cond_resched() call 2796 * - explicit schedule() call 2797 * - return from syscall or exception to user-space 2798 * - return from interrupt-handler to user-space 2799 */ 2800static void __sched __schedule(void) 2801{ 2802 struct task_struct *prev, *next; 2803 unsigned long *switch_count; 2804 struct rq *rq; 2805 int cpu; 2806 2807need_resched: 2808 preempt_disable(); 2809 cpu = smp_processor_id(); 2810 rq = cpu_rq(cpu); 2811 rcu_note_context_switch(cpu); 2812 prev = rq->curr; 2813 2814 schedule_debug(prev); 2815 2816 if (sched_feat(HRTICK)) 2817 hrtick_clear(rq); 2818 2819 /* 2820 * Make sure that signal_pending_state()->signal_pending() below 2821 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE) 2822 * done by the caller to avoid the race with signal_wake_up(). 2823 */ 2824 smp_mb__before_spinlock(); 2825 raw_spin_lock_irq(&rq->lock); 2826 2827 switch_count = &prev->nivcsw; 2828 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) { 2829 if (unlikely(signal_pending_state(prev->state, prev))) { 2830 prev->state = TASK_RUNNING; 2831 } else { 2832 deactivate_task(rq, prev, DEQUEUE_SLEEP); 2833 prev->on_rq = 0; 2834 2835 /* 2836 * If a worker went to sleep, notify and ask workqueue 2837 * whether it wants to wake up a task to maintain 2838 * concurrency. 2839 */ 2840 if (prev->flags & PF_WQ_WORKER) { 2841 struct task_struct *to_wakeup; 2842 2843 to_wakeup = wq_worker_sleeping(prev, cpu); 2844 if (to_wakeup) 2845 try_to_wake_up_local(to_wakeup); 2846 } 2847 } 2848 switch_count = &prev->nvcsw; 2849 } 2850 2851 if (task_on_rq_queued(prev) || rq->skip_clock_update < 0) 2852 update_rq_clock(rq); 2853 2854 next = pick_next_task(rq, prev); 2855 clear_tsk_need_resched(prev); 2856 clear_preempt_need_resched(); 2857 rq->skip_clock_update = 0; 2858 2859 if (likely(prev != next)) { 2860 rq->nr_switches++; 2861 rq->curr = next; 2862 ++*switch_count; 2863 2864 context_switch(rq, prev, next); /* unlocks the rq */ 2865 /* 2866 * The context switch have flipped the stack from under us 2867 * and restored the local variables which were saved when 2868 * this task called schedule() in the past. prev == current 2869 * is still correct, but it can be moved to another cpu/rq. 2870 */ 2871 cpu = smp_processor_id(); 2872 rq = cpu_rq(cpu); 2873 } else 2874 raw_spin_unlock_irq(&rq->lock); 2875 2876 post_schedule(rq); 2877 2878 sched_preempt_enable_no_resched(); 2879 if (need_resched()) 2880 goto need_resched; 2881} 2882 2883static inline void sched_submit_work(struct task_struct *tsk) 2884{ 2885 if (!tsk->state || tsk_is_pi_blocked(tsk)) 2886 return; 2887 /* 2888 * If we are going to sleep and we have plugged IO queued, 2889 * make sure to submit it to avoid deadlocks. 2890 */ 2891 if (blk_needs_flush_plug(tsk)) 2892 blk_schedule_flush_plug(tsk); 2893} 2894 2895asmlinkage __visible void __sched schedule(void) 2896{ 2897 struct task_struct *tsk = current; 2898 2899 sched_submit_work(tsk); 2900 __schedule(); 2901} 2902EXPORT_SYMBOL(schedule); 2903 2904#ifdef CONFIG_CONTEXT_TRACKING 2905asmlinkage __visible void __sched schedule_user(void) 2906{ 2907 /* 2908 * If we come here after a random call to set_need_resched(), 2909 * or we have been woken up remotely but the IPI has not yet arrived, 2910 * we haven't yet exited the RCU idle mode. Do it here manually until 2911 * we find a better solution. 2912 */ 2913 user_exit(); 2914 schedule(); 2915 user_enter(); 2916} 2917#endif 2918 2919/** 2920 * schedule_preempt_disabled - called with preemption disabled 2921 * 2922 * Returns with preemption disabled. Note: preempt_count must be 1 2923 */ 2924void __sched schedule_preempt_disabled(void) 2925{ 2926 sched_preempt_enable_no_resched(); 2927 schedule(); 2928 preempt_disable(); 2929} 2930 2931#ifdef CONFIG_PREEMPT 2932/* 2933 * this is the entry point to schedule() from in-kernel preemption 2934 * off of preempt_enable. Kernel preemptions off return from interrupt 2935 * occur there and call schedule directly. 2936 */ 2937asmlinkage __visible void __sched notrace preempt_schedule(void) 2938{ 2939 /* 2940 * If there is a non-zero preempt_count or interrupts are disabled, 2941 * we do not want to preempt the current task. Just return.. 2942 */ 2943 if (likely(!preemptible())) 2944 return; 2945 2946 do { 2947 __preempt_count_add(PREEMPT_ACTIVE); 2948 __schedule(); 2949 __preempt_count_sub(PREEMPT_ACTIVE); 2950 2951 /* 2952 * Check again in case we missed a preemption opportunity 2953 * between schedule and now. 2954 */ 2955 barrier(); 2956 } while (need_resched()); 2957} 2958NOKPROBE_SYMBOL(preempt_schedule); 2959EXPORT_SYMBOL(preempt_schedule); 2960#endif /* CONFIG_PREEMPT */ 2961 2962/* 2963 * this is the entry point to schedule() from kernel preemption 2964 * off of irq context. 2965 * Note, that this is called and return with irqs disabled. This will 2966 * protect us against recursive calling from irq. 2967 */ 2968asmlinkage __visible void __sched preempt_schedule_irq(void) 2969{ 2970 enum ctx_state prev_state; 2971 2972 /* Catch callers which need to be fixed */ 2973 BUG_ON(preempt_count() || !irqs_disabled()); 2974 2975 prev_state = exception_enter(); 2976 2977 do { 2978 __preempt_count_add(PREEMPT_ACTIVE); 2979 local_irq_enable(); 2980 __schedule(); 2981 local_irq_disable(); 2982 __preempt_count_sub(PREEMPT_ACTIVE); 2983 2984 /* 2985 * Check again in case we missed a preemption opportunity 2986 * between schedule and now. 2987 */ 2988 barrier(); 2989 } while (need_resched()); 2990 2991 exception_exit(prev_state); 2992} 2993 2994int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags, 2995 void *key) 2996{ 2997 return try_to_wake_up(curr->private, mode, wake_flags); 2998} 2999EXPORT_SYMBOL(default_wake_function); 3000 3001#ifdef CONFIG_RT_MUTEXES 3002 3003/* 3004 * rt_mutex_setprio - set the current priority of a task 3005 * @p: task 3006 * @prio: prio value (kernel-internal form) 3007 * 3008 * This function changes the 'effective' priority of a task. It does 3009 * not touch ->normal_prio like __setscheduler(). 3010 * 3011 * Used by the rt_mutex code to implement priority inheritance 3012 * logic. Call site only calls if the priority of the task changed. 3013 */ 3014void rt_mutex_setprio(struct task_struct *p, int prio) 3015{ 3016 int oldprio, queued, running, enqueue_flag = 0; 3017 struct rq *rq; 3018 const struct sched_class *prev_class; 3019 3020 BUG_ON(prio > MAX_PRIO); 3021 3022 rq = __task_rq_lock(p); 3023 3024 /* 3025 * Idle task boosting is a nono in general. There is one 3026 * exception, when PREEMPT_RT and NOHZ is active: 3027 * 3028 * The idle task calls get_next_timer_interrupt() and holds 3029 * the timer wheel base->lock on the CPU and another CPU wants 3030 * to access the timer (probably to cancel it). We can safely 3031 * ignore the boosting request, as the idle CPU runs this code 3032 * with interrupts disabled and will complete the lock 3033 * protected section without being interrupted. So there is no 3034 * real need to boost. 3035 */ 3036 if (unlikely(p == rq->idle)) { 3037 WARN_ON(p != rq->curr); 3038 WARN_ON(p->pi_blocked_on); 3039 goto out_unlock; 3040 } 3041 3042 trace_sched_pi_setprio(p, prio); 3043 oldprio = p->prio; 3044 prev_class = p->sched_class; 3045 queued = task_on_rq_queued(p); 3046 running = task_current(rq, p); 3047 if (queued) 3048 dequeue_task(rq, p, 0); 3049 if (running) 3050 put_prev_task(rq, p); 3051 3052 /* 3053 * Boosting condition are: 3054 * 1. -rt task is running and holds mutex A 3055 * --> -dl task blocks on mutex A 3056 * 3057 * 2. -dl task is running and holds mutex A 3058 * --> -dl task blocks on mutex A and could preempt the 3059 * running task 3060 */ 3061 if (dl_prio(prio)) { 3062 struct task_struct *pi_task = rt_mutex_get_top_task(p); 3063 if (!dl_prio(p->normal_prio) || 3064 (pi_task && dl_entity_preempt(&pi_task->dl, &p->dl))) { 3065 p->dl.dl_boosted = 1; 3066 p->dl.dl_throttled = 0; 3067 enqueue_flag = ENQUEUE_REPLENISH; 3068 } else 3069 p->dl.dl_boosted = 0; 3070 p->sched_class = &dl_sched_class; 3071 } else if (rt_prio(prio)) { 3072 if (dl_prio(oldprio)) 3073 p->dl.dl_boosted = 0; 3074 if (oldprio < prio) 3075 enqueue_flag = ENQUEUE_HEAD; 3076 p->sched_class = &rt_sched_class; 3077 } else { 3078 if (dl_prio(oldprio)) 3079 p->dl.dl_boosted = 0; 3080 p->sched_class = &fair_sched_class; 3081 } 3082 3083 p->prio = prio; 3084 3085 if (running) 3086 p->sched_class->set_curr_task(rq); 3087 if (queued) 3088 enqueue_task(rq, p, enqueue_flag); 3089 3090 check_class_changed(rq, p, prev_class, oldprio); 3091out_unlock: 3092 __task_rq_unlock(rq); 3093} 3094#endif 3095 3096void set_user_nice(struct task_struct *p, long nice) 3097{ 3098 int old_prio, delta, queued; 3099 unsigned long flags; 3100 struct rq *rq; 3101 3102 if (task_nice(p) == nice || nice < MIN_NICE || nice > MAX_NICE) 3103 return; 3104 /* 3105 * We have to be careful, if called from sys_setpriority(), 3106 * the task might be in the middle of scheduling on another CPU. 3107 */ 3108 rq = task_rq_lock(p, &flags); 3109 /* 3110 * The RT priorities are set via sched_setscheduler(), but we still 3111 * allow the 'normal' nice value to be set - but as expected 3112 * it wont have any effect on scheduling until the task is 3113 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR: 3114 */ 3115 if (task_has_dl_policy(p) || task_has_rt_policy(p)) { 3116 p->static_prio = NICE_TO_PRIO(nice); 3117 goto out_unlock; 3118 } 3119 queued = task_on_rq_queued(p); 3120 if (queued) 3121 dequeue_task(rq, p, 0); 3122 3123 p->static_prio = NICE_TO_PRIO(nice); 3124 set_load_weight(p); 3125 old_prio = p->prio; 3126 p->prio = effective_prio(p); 3127 delta = p->prio - old_prio; 3128 3129 if (queued) { 3130 enqueue_task(rq, p, 0); 3131 /* 3132 * If the task increased its priority or is running and 3133 * lowered its priority, then reschedule its CPU: 3134 */ 3135 if (delta < 0 || (delta > 0 && task_running(rq, p))) 3136 resched_curr(rq); 3137 } 3138out_unlock: 3139 task_rq_unlock(rq, p, &flags); 3140} 3141EXPORT_SYMBOL(set_user_nice); 3142 3143/* 3144 * can_nice - check if a task can reduce its nice value 3145 * @p: task 3146 * @nice: nice value 3147 */ 3148int can_nice(const struct task_struct *p, const int nice) 3149{ 3150 /* convert nice value [19,-20] to rlimit style value [1,40] */ 3151 int nice_rlim = nice_to_rlimit(nice); 3152 3153 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) || 3154 capable(CAP_SYS_NICE)); 3155} 3156 3157#ifdef __ARCH_WANT_SYS_NICE 3158 3159/* 3160 * sys_nice - change the priority of the current process. 3161 * @increment: priority increment 3162 * 3163 * sys_setpriority is a more generic, but much slower function that 3164 * does similar things. 3165 */ 3166SYSCALL_DEFINE1(nice, int, increment) 3167{ 3168 long nice, retval; 3169 3170 /* 3171 * Setpriority might change our priority at the same moment. 3172 * We don't have to worry. Conceptually one call occurs first 3173 * and we have a single winner. 3174 */ 3175 increment = clamp(increment, -NICE_WIDTH, NICE_WIDTH); 3176 nice = task_nice(current) + increment; 3177 3178 nice = clamp_val(nice, MIN_NICE, MAX_NICE); 3179 if (increment < 0 && !can_nice(current, nice)) 3180 return -EPERM; 3181 3182 retval = security_task_setnice(current, nice); 3183 if (retval) 3184 return retval; 3185 3186 set_user_nice(current, nice); 3187 return 0; 3188} 3189 3190#endif 3191 3192/** 3193 * task_prio - return the priority value of a given task. 3194 * @p: the task in question. 3195 * 3196 * Return: The priority value as seen by users in /proc. 3197 * RT tasks are offset by -200. Normal tasks are centered 3198 * around 0, value goes from -16 to +15. 3199 */ 3200int task_prio(const struct task_struct *p) 3201{ 3202 return p->prio - MAX_RT_PRIO; 3203} 3204 3205/** 3206 * idle_cpu - is a given cpu idle currently? 3207 * @cpu: the processor in question. 3208 * 3209 * Return: 1 if the CPU is currently idle. 0 otherwise. 3210 */ 3211int idle_cpu(int cpu) 3212{ 3213 struct rq *rq = cpu_rq(cpu); 3214 3215 if (rq->curr != rq->idle) 3216 return 0; 3217 3218 if (rq->nr_running) 3219 return 0; 3220 3221#ifdef CONFIG_SMP 3222 if (!llist_empty(&rq->wake_list)) 3223 return 0; 3224#endif 3225 3226 return 1; 3227} 3228 3229/** 3230 * idle_task - return the idle task for a given cpu. 3231 * @cpu: the processor in question. 3232 * 3233 * Return: The idle task for the cpu @cpu. 3234 */ 3235struct task_struct *idle_task(int cpu) 3236{ 3237 return cpu_rq(cpu)->idle; 3238} 3239 3240/** 3241 * find_process_by_pid - find a process with a matching PID value. 3242 * @pid: the pid in question. 3243 * 3244 * The task of @pid, if found. %NULL otherwise. 3245 */ 3246static struct task_struct *find_process_by_pid(pid_t pid) 3247{ 3248 return pid ? find_task_by_vpid(pid) : current; 3249} 3250 3251/* 3252 * This function initializes the sched_dl_entity of a newly becoming 3253 * SCHED_DEADLINE task. 3254 * 3255 * Only the static values are considered here, the actual runtime and the 3256 * absolute deadline will be properly calculated when the task is enqueued 3257 * for the first time with its new policy. 3258 */ 3259static void 3260__setparam_dl(struct task_struct *p, const struct sched_attr *attr) 3261{ 3262 struct sched_dl_entity *dl_se = &p->dl; 3263 3264 init_dl_task_timer(dl_se); 3265 dl_se->dl_runtime = attr->sched_runtime; 3266 dl_se->dl_deadline = attr->sched_deadline; 3267 dl_se->dl_period = attr->sched_period ?: dl_se->dl_deadline; 3268 dl_se->flags = attr->sched_flags; 3269 dl_se->dl_bw = to_ratio(dl_se->dl_period, dl_se->dl_runtime); 3270 dl_se->dl_throttled = 0; 3271 dl_se->dl_new = 1; 3272 dl_se->dl_yielded = 0; 3273} 3274 3275/* 3276 * sched_setparam() passes in -1 for its policy, to let the functions 3277 * it calls know not to change it. 3278 */ 3279#define SETPARAM_POLICY -1 3280 3281static void __setscheduler_params(struct task_struct *p, 3282 const struct sched_attr *attr) 3283{ 3284 int policy = attr->sched_policy; 3285 3286 if (policy == SETPARAM_POLICY) 3287 policy = p->policy; 3288 3289 p->policy = policy; 3290 3291 if (dl_policy(policy)) 3292 __setparam_dl(p, attr); 3293 else if (fair_policy(policy)) 3294 p->static_prio = NICE_TO_PRIO(attr->sched_nice); 3295 3296 /* 3297 * __sched_setscheduler() ensures attr->sched_priority == 0 when 3298 * !rt_policy. Always setting this ensures that things like 3299 * getparam()/getattr() don't report silly values for !rt tasks. 3300 */ 3301 p->rt_priority = attr->sched_priority; 3302 p->normal_prio = normal_prio(p); 3303 set_load_weight(p); 3304} 3305 3306/* Actually do priority change: must hold pi & rq lock. */ 3307static void __setscheduler(struct rq *rq, struct task_struct *p, 3308 const struct sched_attr *attr) 3309{ 3310 __setscheduler_params(p, attr); 3311 3312 /* 3313 * If we get here, there was no pi waiters boosting the 3314 * task. It is safe to use the normal prio. 3315 */ 3316 p->prio = normal_prio(p); 3317 3318 if (dl_prio(p->prio)) 3319 p->sched_class = &dl_sched_class; 3320 else if (rt_prio(p->prio)) 3321 p->sched_class = &rt_sched_class; 3322 else 3323 p->sched_class = &fair_sched_class; 3324} 3325 3326static void 3327__getparam_dl(struct task_struct *p, struct sched_attr *attr) 3328{ 3329 struct sched_dl_entity *dl_se = &p->dl; 3330 3331 attr->sched_priority = p->rt_priority; 3332 attr->sched_runtime = dl_se->dl_runtime; 3333 attr->sched_deadline = dl_se->dl_deadline; 3334 attr->sched_period = dl_se->dl_period; 3335 attr->sched_flags = dl_se->flags; 3336} 3337 3338/* 3339 * This function validates the new parameters of a -deadline task. 3340 * We ask for the deadline not being zero, and greater or equal 3341 * than the runtime, as well as the period of being zero or 3342 * greater than deadline. Furthermore, we have to be sure that 3343 * user parameters are above the internal resolution of 1us (we 3344 * check sched_runtime only since it is always the smaller one) and 3345 * below 2^63 ns (we have to check both sched_deadline and 3346 * sched_period, as the latter can be zero). 3347 */ 3348static bool 3349__checkparam_dl(const struct sched_attr *attr) 3350{ 3351 /* deadline != 0 */ 3352 if (attr->sched_deadline == 0) 3353 return false; 3354 3355 /* 3356 * Since we truncate DL_SCALE bits, make sure we're at least 3357 * that big. 3358 */ 3359 if (attr->sched_runtime < (1ULL << DL_SCALE)) 3360 return false; 3361 3362 /* 3363 * Since we use the MSB for wrap-around and sign issues, make 3364 * sure it's not set (mind that period can be equal to zero). 3365 */ 3366 if (attr->sched_deadline & (1ULL << 63) || 3367 attr->sched_period & (1ULL << 63)) 3368 return false; 3369 3370 /* runtime <= deadline <= period (if period != 0) */ 3371 if ((attr->sched_period != 0 && 3372 attr->sched_period < attr->sched_deadline) || 3373 attr->sched_deadline < attr->sched_runtime) 3374 return false; 3375 3376 return true; 3377} 3378 3379/* 3380 * check the target process has a UID that matches the current process's 3381 */ 3382static bool check_same_owner(struct task_struct *p) 3383{ 3384 const struct cred *cred = current_cred(), *pcred; 3385 bool match; 3386 3387 rcu_read_lock(); 3388 pcred = __task_cred(p); 3389 match = (uid_eq(cred->euid, pcred->euid) || 3390 uid_eq(cred->euid, pcred->uid)); 3391 rcu_read_unlock(); 3392 return match; 3393} 3394 3395static int __sched_setscheduler(struct task_struct *p, 3396 const struct sched_attr *attr, 3397 bool user) 3398{ 3399 int newprio = dl_policy(attr->sched_policy) ? MAX_DL_PRIO - 1 : 3400 MAX_RT_PRIO - 1 - attr->sched_priority; 3401 int retval, oldprio, oldpolicy = -1, queued, running; 3402 int policy = attr->sched_policy; 3403 unsigned long flags; 3404 const struct sched_class *prev_class; 3405 struct rq *rq; 3406 int reset_on_fork; 3407 3408 /* may grab non-irq protected spin_locks */ 3409 BUG_ON(in_interrupt()); 3410recheck: 3411 /* double check policy once rq lock held */ 3412 if (policy < 0) { 3413 reset_on_fork = p->sched_reset_on_fork; 3414 policy = oldpolicy = p->policy; 3415 } else { 3416 reset_on_fork = !!(attr->sched_flags & SCHED_FLAG_RESET_ON_FORK); 3417 3418 if (policy != SCHED_DEADLINE && 3419 policy != SCHED_FIFO && policy != SCHED_RR && 3420 policy != SCHED_NORMAL && policy != SCHED_BATCH && 3421 policy != SCHED_IDLE) 3422 return -EINVAL; 3423 } 3424 3425 if (attr->sched_flags & ~(SCHED_FLAG_RESET_ON_FORK)) 3426 return -EINVAL; 3427 3428 /* 3429 * Valid priorities for SCHED_FIFO and SCHED_RR are 3430 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL, 3431 * SCHED_BATCH and SCHED_IDLE is 0. 3432 */ 3433 if ((p->mm && attr->sched_priority > MAX_USER_RT_PRIO-1) || 3434 (!p->mm && attr->sched_priority > MAX_RT_PRIO-1)) 3435 return -EINVAL; 3436 if ((dl_policy(policy) && !__checkparam_dl(attr)) || 3437 (rt_policy(policy) != (attr->sched_priority != 0))) 3438 return -EINVAL; 3439 3440 /* 3441 * Allow unprivileged RT tasks to decrease priority: 3442 */ 3443 if (user && !capable(CAP_SYS_NICE)) { 3444 if (fair_policy(policy)) { 3445 if (attr->sched_nice < task_nice(p) && 3446 !can_nice(p, attr->sched_nice)) 3447 return -EPERM; 3448 } 3449 3450 if (rt_policy(policy)) { 3451 unsigned long rlim_rtprio = 3452 task_rlimit(p, RLIMIT_RTPRIO); 3453 3454 /* can't set/change the rt policy */ 3455 if (policy != p->policy && !rlim_rtprio) 3456 return -EPERM; 3457 3458 /* can't increase priority */ 3459 if (attr->sched_priority > p->rt_priority && 3460 attr->sched_priority > rlim_rtprio) 3461 return -EPERM; 3462 } 3463 3464 /* 3465 * Can't set/change SCHED_DEADLINE policy at all for now 3466 * (safest behavior); in the future we would like to allow 3467 * unprivileged DL tasks to increase their relative deadline 3468 * or reduce their runtime (both ways reducing utilization) 3469 */ 3470 if (dl_policy(policy)) 3471 return -EPERM; 3472 3473 /* 3474 * Treat SCHED_IDLE as nice 20. Only allow a switch to 3475 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it. 3476 */ 3477 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) { 3478 if (!can_nice(p, task_nice(p))) 3479 return -EPERM; 3480 } 3481 3482 /* can't change other user's priorities */ 3483 if (!check_same_owner(p)) 3484 return -EPERM; 3485 3486 /* Normal users shall not reset the sched_reset_on_fork flag */ 3487 if (p->sched_reset_on_fork && !reset_on_fork) 3488 return -EPERM; 3489 } 3490 3491 if (user) { 3492 retval = security_task_setscheduler(p); 3493 if (retval) 3494 return retval; 3495 } 3496 3497 /* 3498 * make sure no PI-waiters arrive (or leave) while we are 3499 * changing the priority of the task: 3500 * 3501 * To be able to change p->policy safely, the appropriate 3502 * runqueue lock must be held. 3503 */ 3504 rq = task_rq_lock(p, &flags); 3505 3506 /* 3507 * Changing the policy of the stop threads its a very bad idea 3508 */ 3509 if (p == rq->stop) { 3510 task_rq_unlock(rq, p, &flags); 3511 return -EINVAL; 3512 } 3513 3514 /* 3515 * If not changing anything there's no need to proceed further, 3516 * but store a possible modification of reset_on_fork. 3517 */ 3518 if (unlikely(policy == p->policy)) { 3519 if (fair_policy(policy) && attr->sched_nice != task_nice(p)) 3520 goto change; 3521 if (rt_policy(policy) && attr->sched_priority != p->rt_priority) 3522 goto change; 3523 if (dl_policy(policy)) 3524 goto change; 3525 3526 p->sched_reset_on_fork = reset_on_fork; 3527 task_rq_unlock(rq, p, &flags); 3528 return 0; 3529 } 3530change: 3531 3532 if (user) { 3533#ifdef CONFIG_RT_GROUP_SCHED 3534 /* 3535 * Do not allow realtime tasks into groups that have no runtime 3536 * assigned. 3537 */ 3538 if (rt_bandwidth_enabled() && rt_policy(policy) && 3539 task_group(p)->rt_bandwidth.rt_runtime == 0 && 3540 !task_group_is_autogroup(task_group(p))) { 3541 task_rq_unlock(rq, p, &flags); 3542 return -EPERM; 3543 } 3544#endif 3545#ifdef CONFIG_SMP 3546 if (dl_bandwidth_enabled() && dl_policy(policy)) { 3547 cpumask_t *span = rq->rd->span; 3548 3549 /* 3550 * Don't allow tasks with an affinity mask smaller than 3551 * the entire root_domain to become SCHED_DEADLINE. We 3552 * will also fail if there's no bandwidth available. 3553 */ 3554 if (!cpumask_subset(span, &p->cpus_allowed) || 3555 rq->rd->dl_bw.bw == 0) { 3556 task_rq_unlock(rq, p, &flags); 3557 return -EPERM; 3558 } 3559 } 3560#endif 3561 } 3562 3563 /* recheck policy now with rq lock held */ 3564 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) { 3565 policy = oldpolicy = -1; 3566 task_rq_unlock(rq, p, &flags); 3567 goto recheck; 3568 } 3569 3570 /* 3571 * If setscheduling to SCHED_DEADLINE (or changing the parameters 3572 * of a SCHED_DEADLINE task) we need to check if enough bandwidth 3573 * is available. 3574 */ 3575 if ((dl_policy(policy) || dl_task(p)) && dl_overflow(p, policy, attr)) { 3576 task_rq_unlock(rq, p, &flags); 3577 return -EBUSY; 3578 } 3579 3580 p->sched_reset_on_fork = reset_on_fork; 3581 oldprio = p->prio; 3582 3583 /* 3584 * Special case for priority boosted tasks. 3585 * 3586 * If the new priority is lower or equal (user space view) 3587 * than the current (boosted) priority, we just store the new 3588 * normal parameters and do not touch the scheduler class and 3589 * the runqueue. This will be done when the task deboost 3590 * itself. 3591 */ 3592 if (rt_mutex_check_prio(p, newprio)) { 3593 __setscheduler_params(p, attr); 3594 task_rq_unlock(rq, p, &flags); 3595 return 0; 3596 } 3597 3598 queued = task_on_rq_queued(p); 3599 running = task_current(rq, p); 3600 if (queued) 3601 dequeue_task(rq, p, 0); 3602 if (running) 3603 put_prev_task(rq, p); 3604 3605 prev_class = p->sched_class; 3606 __setscheduler(rq, p, attr); 3607 3608 if (running) 3609 p->sched_class->set_curr_task(rq); 3610 if (queued) { 3611 /* 3612 * We enqueue to tail when the priority of a task is 3613 * increased (user space view). 3614 */ 3615 enqueue_task(rq, p, oldprio <= p->prio ? ENQUEUE_HEAD : 0); 3616 } 3617 3618 check_class_changed(rq, p, prev_class, oldprio); 3619 task_rq_unlock(rq, p, &flags); 3620 3621 rt_mutex_adjust_pi(p); 3622 3623 return 0; 3624} 3625 3626static int _sched_setscheduler(struct task_struct *p, int policy, 3627 const struct sched_param *param, bool check) 3628{ 3629 struct sched_attr attr = { 3630 .sched_policy = policy, 3631 .sched_priority = param->sched_priority, 3632 .sched_nice = PRIO_TO_NICE(p->static_prio), 3633 }; 3634 3635 /* Fixup the legacy SCHED_RESET_ON_FORK hack. */ 3636 if ((policy != SETPARAM_POLICY) && (policy & SCHED_RESET_ON_FORK)) { 3637 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK; 3638 policy &= ~SCHED_RESET_ON_FORK; 3639 attr.sched_policy = policy; 3640 } 3641 3642 return __sched_setscheduler(p, &attr, check); 3643} 3644/** 3645 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread. 3646 * @p: the task in question. 3647 * @policy: new policy. 3648 * @param: structure containing the new RT priority. 3649 * 3650 * Return: 0 on success. An error code otherwise. 3651 * 3652 * NOTE that the task may be already dead. 3653 */ 3654int sched_setscheduler(struct task_struct *p, int policy, 3655 const struct sched_param *param) 3656{ 3657 return _sched_setscheduler(p, policy, param, true); 3658} 3659EXPORT_SYMBOL_GPL(sched_setscheduler); 3660 3661int sched_setattr(struct task_struct *p, const struct sched_attr *attr) 3662{ 3663 return __sched_setscheduler(p, attr, true); 3664} 3665EXPORT_SYMBOL_GPL(sched_setattr); 3666 3667/** 3668 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace. 3669 * @p: the task in question. 3670 * @policy: new policy. 3671 * @param: structure containing the new RT priority. 3672 * 3673 * Just like sched_setscheduler, only don't bother checking if the 3674 * current context has permission. For example, this is needed in 3675 * stop_machine(): we create temporary high priority worker threads, 3676 * but our caller might not have that capability. 3677 * 3678 * Return: 0 on success. An error code otherwise. 3679 */ 3680int sched_setscheduler_nocheck(struct task_struct *p, int policy, 3681 const struct sched_param *param) 3682{ 3683 return _sched_setscheduler(p, policy, param, false); 3684} 3685 3686static int 3687do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param) 3688{ 3689 struct sched_param lparam; 3690 struct task_struct *p; 3691 int retval; 3692 3693 if (!param || pid < 0) 3694 return -EINVAL; 3695 if (copy_from_user(&lparam, param, sizeof(struct sched_param))) 3696 return -EFAULT; 3697 3698 rcu_read_lock(); 3699 retval = -ESRCH; 3700 p = find_process_by_pid(pid); 3701 if (p != NULL) 3702 retval = sched_setscheduler(p, policy, &lparam); 3703 rcu_read_unlock(); 3704 3705 return retval; 3706} 3707 3708/* 3709 * Mimics kernel/events/core.c perf_copy_attr(). 3710 */ 3711static int sched_copy_attr(struct sched_attr __user *uattr, 3712 struct sched_attr *attr) 3713{ 3714 u32 size; 3715 int ret; 3716 3717 if (!access_ok(VERIFY_WRITE, uattr, SCHED_ATTR_SIZE_VER0)) 3718 return -EFAULT; 3719 3720 /* 3721 * zero the full structure, so that a short copy will be nice. 3722 */ 3723 memset(attr, 0, sizeof(*attr)); 3724 3725 ret = get_user(size, &uattr->size); 3726 if (ret) 3727 return ret; 3728 3729 if (size > PAGE_SIZE) /* silly large */ 3730 goto err_size; 3731 3732 if (!size) /* abi compat */ 3733 size = SCHED_ATTR_SIZE_VER0; 3734 3735 if (size < SCHED_ATTR_SIZE_VER0) 3736 goto err_size; 3737 3738 /* 3739 * If we're handed a bigger struct than we know of, 3740 * ensure all the unknown bits are 0 - i.e. new 3741 * user-space does not rely on any kernel feature 3742 * extensions we dont know about yet. 3743 */ 3744 if (size > sizeof(*attr)) { 3745 unsigned char __user *addr; 3746 unsigned char __user *end; 3747 unsigned char val; 3748 3749 addr = (void __user *)uattr + sizeof(*attr); 3750 end = (void __user *)uattr + size; 3751 3752 for (; addr < end; addr++) { 3753 ret = get_user(val, addr); 3754 if (ret) 3755 return ret; 3756 if (val) 3757 goto err_size; 3758 } 3759 size = sizeof(*attr); 3760 } 3761 3762 ret = copy_from_user(attr, uattr, size); 3763 if (ret) 3764 return -EFAULT; 3765 3766 /* 3767 * XXX: do we want to be lenient like existing syscalls; or do we want 3768 * to be strict and return an error on out-of-bounds values? 3769 */ 3770 attr->sched_nice = clamp(attr->sched_nice, MIN_NICE, MAX_NICE); 3771 3772 return 0; 3773 3774err_size: 3775 put_user(sizeof(*attr), &uattr->size); 3776 return -E2BIG; 3777} 3778 3779/** 3780 * sys_sched_setscheduler - set/change the scheduler policy and RT priority 3781 * @pid: the pid in question. 3782 * @policy: new policy. 3783 * @param: structure containing the new RT priority. 3784 * 3785 * Return: 0 on success. An error code otherwise. 3786 */ 3787SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy, 3788 struct sched_param __user *, param) 3789{ 3790 /* negative values for policy are not valid */ 3791 if (policy < 0) 3792 return -EINVAL; 3793 3794 return do_sched_setscheduler(pid, policy, param); 3795} 3796 3797/** 3798 * sys_sched_setparam - set/change the RT priority of a thread 3799 * @pid: the pid in question. 3800 * @param: structure containing the new RT priority. 3801 * 3802 * Return: 0 on success. An error code otherwise. 3803 */ 3804SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param) 3805{ 3806 return do_sched_setscheduler(pid, SETPARAM_POLICY, param); 3807} 3808 3809/** 3810 * sys_sched_setattr - same as above, but with extended sched_attr 3811 * @pid: the pid in question. 3812 * @uattr: structure containing the extended parameters. 3813 * @flags: for future extension. 3814 */ 3815SYSCALL_DEFINE3(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr, 3816 unsigned int, flags) 3817{ 3818 struct sched_attr attr; 3819 struct task_struct *p; 3820 int retval; 3821 3822 if (!uattr || pid < 0 || flags) 3823 return -EINVAL; 3824 3825 retval = sched_copy_attr(uattr, &attr); 3826 if (retval) 3827 return retval; 3828 3829 if ((int)attr.sched_policy < 0) 3830 return -EINVAL; 3831 3832 rcu_read_lock(); 3833 retval = -ESRCH; 3834 p = find_process_by_pid(pid); 3835 if (p != NULL) 3836 retval = sched_setattr(p, &attr); 3837 rcu_read_unlock(); 3838 3839 return retval; 3840} 3841 3842/** 3843 * sys_sched_getscheduler - get the policy (scheduling class) of a thread 3844 * @pid: the pid in question. 3845 * 3846 * Return: On success, the policy of the thread. Otherwise, a negative error 3847 * code. 3848 */ 3849SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid) 3850{ 3851 struct task_struct *p; 3852 int retval; 3853 3854 if (pid < 0) 3855 return -EINVAL; 3856 3857 retval = -ESRCH; 3858 rcu_read_lock(); 3859 p = find_process_by_pid(pid); 3860 if (p) { 3861 retval = security_task_getscheduler(p); 3862 if (!retval) 3863 retval = p->policy 3864 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0); 3865 } 3866 rcu_read_unlock(); 3867 return retval; 3868} 3869 3870/** 3871 * sys_sched_getparam - get the RT priority of a thread 3872 * @pid: the pid in question. 3873 * @param: structure containing the RT priority. 3874 * 3875 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error 3876 * code. 3877 */ 3878SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param) 3879{ 3880 struct sched_param lp = { .sched_priority = 0 }; 3881 struct task_struct *p; 3882 int retval; 3883 3884 if (!param || pid < 0) 3885 return -EINVAL; 3886 3887 rcu_read_lock(); 3888 p = find_process_by_pid(pid); 3889 retval = -ESRCH; 3890 if (!p) 3891 goto out_unlock; 3892 3893 retval = security_task_getscheduler(p); 3894 if (retval) 3895 goto out_unlock; 3896 3897 if (task_has_rt_policy(p)) 3898 lp.sched_priority = p->rt_priority; 3899 rcu_read_unlock(); 3900 3901 /* 3902 * This one might sleep, we cannot do it with a spinlock held ... 3903 */ 3904 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0; 3905 3906 return retval; 3907 3908out_unlock: 3909 rcu_read_unlock(); 3910 return retval; 3911} 3912 3913static int sched_read_attr(struct sched_attr __user *uattr, 3914 struct sched_attr *attr, 3915 unsigned int usize) 3916{ 3917 int ret; 3918 3919 if (!access_ok(VERIFY_WRITE, uattr, usize)) 3920 return -EFAULT; 3921 3922 /* 3923 * If we're handed a smaller struct than we know of, 3924 * ensure all the unknown bits are 0 - i.e. old 3925 * user-space does not get uncomplete information. 3926 */ 3927 if (usize < sizeof(*attr)) { 3928 unsigned char *addr; 3929 unsigned char *end; 3930 3931 addr = (void *)attr + usize; 3932 end = (void *)attr + sizeof(*attr); 3933 3934 for (; addr < end; addr++) { 3935 if (*addr) 3936 return -EFBIG; 3937 } 3938 3939 attr->size = usize; 3940 } 3941 3942 ret = copy_to_user(uattr, attr, attr->size); 3943 if (ret) 3944 return -EFAULT; 3945 3946 return 0; 3947} 3948 3949/** 3950 * sys_sched_getattr - similar to sched_getparam, but with sched_attr 3951 * @pid: the pid in question. 3952 * @uattr: structure containing the extended parameters. 3953 * @size: sizeof(attr) for fwd/bwd comp. 3954 * @flags: for future extension. 3955 */ 3956SYSCALL_DEFINE4(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr, 3957 unsigned int, size, unsigned int, flags) 3958{ 3959 struct sched_attr attr = { 3960 .size = sizeof(struct sched_attr), 3961 }; 3962 struct task_struct *p; 3963 int retval; 3964 3965 if (!uattr || pid < 0 || size > PAGE_SIZE || 3966 size < SCHED_ATTR_SIZE_VER0 || flags) 3967 return -EINVAL; 3968 3969 rcu_read_lock(); 3970 p = find_process_by_pid(pid); 3971 retval = -ESRCH; 3972 if (!p) 3973 goto out_unlock; 3974 3975 retval = security_task_getscheduler(p); 3976 if (retval) 3977 goto out_unlock; 3978 3979 attr.sched_policy = p->policy; 3980 if (p->sched_reset_on_fork) 3981 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK; 3982 if (task_has_dl_policy(p)) 3983 __getparam_dl(p, &attr); 3984 else if (task_has_rt_policy(p)) 3985 attr.sched_priority = p->rt_priority; 3986 else 3987 attr.sched_nice = task_nice(p); 3988 3989 rcu_read_unlock(); 3990 3991 retval = sched_read_attr(uattr, &attr, size); 3992 return retval; 3993 3994out_unlock: 3995 rcu_read_unlock(); 3996 return retval; 3997} 3998 3999long sched_setaffinity(pid_t pid, const struct cpumask *in_mask) 4000{ 4001 cpumask_var_t cpus_allowed, new_mask; 4002 struct task_struct *p; 4003 int retval; 4004 4005 rcu_read_lock(); 4006 4007 p = find_process_by_pid(pid); 4008 if (!p) { 4009 rcu_read_unlock(); 4010 return -ESRCH; 4011 } 4012 4013 /* Prevent p going away */ 4014 get_task_struct(p); 4015 rcu_read_unlock(); 4016 4017 if (p->flags & PF_NO_SETAFFINITY) { 4018 retval = -EINVAL; 4019 goto out_put_task; 4020 } 4021 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) { 4022 retval = -ENOMEM; 4023 goto out_put_task; 4024 } 4025 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) { 4026 retval = -ENOMEM; 4027 goto out_free_cpus_allowed; 4028 } 4029 retval = -EPERM; 4030 if (!check_same_owner(p)) { 4031 rcu_read_lock(); 4032 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) { 4033 rcu_read_unlock(); 4034 goto out_unlock; 4035 } 4036 rcu_read_unlock(); 4037 } 4038 4039 retval = security_task_setscheduler(p); 4040 if (retval) 4041 goto out_unlock; 4042 4043 4044 cpuset_cpus_allowed(p, cpus_allowed); 4045 cpumask_and(new_mask, in_mask, cpus_allowed); 4046 4047 /* 4048 * Since bandwidth control happens on root_domain basis, 4049 * if admission test is enabled, we only admit -deadline 4050 * tasks allowed to run on all the CPUs in the task's 4051 * root_domain. 4052 */ 4053#ifdef CONFIG_SMP 4054 if (task_has_dl_policy(p)) { 4055 const struct cpumask *span = task_rq(p)->rd->span; 4056 4057 if (dl_bandwidth_enabled() && !cpumask_subset(span, new_mask)) { 4058 retval = -EBUSY; 4059 goto out_unlock; 4060 } 4061 } 4062#endif 4063again: 4064 retval = set_cpus_allowed_ptr(p, new_mask); 4065 4066 if (!retval) { 4067 cpuset_cpus_allowed(p, cpus_allowed); 4068 if (!cpumask_subset(new_mask, cpus_allowed)) { 4069 /* 4070 * We must have raced with a concurrent cpuset 4071 * update. Just reset the cpus_allowed to the 4072 * cpuset's cpus_allowed 4073 */ 4074 cpumask_copy(new_mask, cpus_allowed); 4075 goto again; 4076 } 4077 } 4078out_unlock: 4079 free_cpumask_var(new_mask); 4080out_free_cpus_allowed: 4081 free_cpumask_var(cpus_allowed); 4082out_put_task: 4083 put_task_struct(p); 4084 return retval; 4085} 4086 4087static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len, 4088 struct cpumask *new_mask) 4089{ 4090 if (len < cpumask_size()) 4091 cpumask_clear(new_mask); 4092 else if (len > cpumask_size()) 4093 len = cpumask_size(); 4094 4095 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0; 4096} 4097 4098/** 4099 * sys_sched_setaffinity - set the cpu affinity of a process 4100 * @pid: pid of the process 4101 * @len: length in bytes of the bitmask pointed to by user_mask_ptr 4102 * @user_mask_ptr: user-space pointer to the new cpu mask 4103 * 4104 * Return: 0 on success. An error code otherwise. 4105 */ 4106SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len, 4107 unsigned long __user *, user_mask_ptr) 4108{ 4109 cpumask_var_t new_mask; 4110 int retval; 4111 4112 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) 4113 return -ENOMEM; 4114 4115 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask); 4116 if (retval == 0) 4117 retval = sched_setaffinity(pid, new_mask); 4118 free_cpumask_var(new_mask); 4119 return retval; 4120} 4121 4122long sched_getaffinity(pid_t pid, struct cpumask *mask) 4123{ 4124 struct task_struct *p; 4125 unsigned long flags; 4126 int retval; 4127 4128 rcu_read_lock(); 4129 4130 retval = -ESRCH; 4131 p = find_process_by_pid(pid); 4132 if (!p) 4133 goto out_unlock; 4134 4135 retval = security_task_getscheduler(p); 4136 if (retval) 4137 goto out_unlock; 4138 4139 raw_spin_lock_irqsave(&p->pi_lock, flags); 4140 cpumask_and(mask, &p->cpus_allowed, cpu_active_mask); 4141 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 4142 4143out_unlock: 4144 rcu_read_unlock(); 4145 4146 return retval; 4147} 4148 4149/** 4150 * sys_sched_getaffinity - get the cpu affinity of a process 4151 * @pid: pid of the process 4152 * @len: length in bytes of the bitmask pointed to by user_mask_ptr 4153 * @user_mask_ptr: user-space pointer to hold the current cpu mask 4154 * 4155 * Return: 0 on success. An error code otherwise. 4156 */ 4157SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len, 4158 unsigned long __user *, user_mask_ptr) 4159{ 4160 int ret; 4161 cpumask_var_t mask; 4162 4163 if ((len * BITS_PER_BYTE) < nr_cpu_ids) 4164 return -EINVAL; 4165 if (len & (sizeof(unsigned long)-1)) 4166 return -EINVAL; 4167 4168 if (!alloc_cpumask_var(&mask, GFP_KERNEL)) 4169 return -ENOMEM; 4170 4171 ret = sched_getaffinity(pid, mask); 4172 if (ret == 0) { 4173 size_t retlen = min_t(size_t, len, cpumask_size()); 4174 4175 if (copy_to_user(user_mask_ptr, mask, retlen)) 4176 ret = -EFAULT; 4177 else 4178 ret = retlen; 4179 } 4180 free_cpumask_var(mask); 4181 4182 return ret; 4183} 4184 4185/** 4186 * sys_sched_yield - yield the current processor to other threads. 4187 * 4188 * This function yields the current CPU to other tasks. If there are no 4189 * other threads running on this CPU then this function will return. 4190 * 4191 * Return: 0. 4192 */ 4193SYSCALL_DEFINE0(sched_yield) 4194{ 4195 struct rq *rq = this_rq_lock(); 4196 4197 schedstat_inc(rq, yld_count); 4198 current->sched_class->yield_task(rq); 4199 4200 /* 4201 * Since we are going to call schedule() anyway, there's 4202 * no need to preempt or enable interrupts: 4203 */ 4204 __release(rq->lock); 4205 spin_release(&rq->lock.dep_map, 1, _THIS_IP_); 4206 do_raw_spin_unlock(&rq->lock); 4207 sched_preempt_enable_no_resched(); 4208 4209 schedule(); 4210 4211 return 0; 4212} 4213 4214static void __cond_resched(void) 4215{ 4216 __preempt_count_add(PREEMPT_ACTIVE); 4217 __schedule(); 4218 __preempt_count_sub(PREEMPT_ACTIVE); 4219} 4220 4221int __sched _cond_resched(void) 4222{ 4223 if (should_resched()) { 4224 __cond_resched(); 4225 return 1; 4226 } 4227 return 0; 4228} 4229EXPORT_SYMBOL(_cond_resched); 4230 4231/* 4232 * __cond_resched_lock() - if a reschedule is pending, drop the given lock, 4233 * call schedule, and on return reacquire the lock. 4234 * 4235 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level 4236 * operations here to prevent schedule() from being called twice (once via 4237 * spin_unlock(), once by hand). 4238 */ 4239int __cond_resched_lock(spinlock_t *lock) 4240{ 4241 int resched = should_resched(); 4242 int ret = 0; 4243 4244 lockdep_assert_held(lock); 4245 4246 if (spin_needbreak(lock) || resched) { 4247 spin_unlock(lock); 4248 if (resched) 4249 __cond_resched(); 4250 else 4251 cpu_relax(); 4252 ret = 1; 4253 spin_lock(lock); 4254 } 4255 return ret; 4256} 4257EXPORT_SYMBOL(__cond_resched_lock); 4258 4259int __sched __cond_resched_softirq(void) 4260{ 4261 BUG_ON(!in_softirq()); 4262 4263 if (should_resched()) { 4264 local_bh_enable(); 4265 __cond_resched(); 4266 local_bh_disable(); 4267 return 1; 4268 } 4269 return 0; 4270} 4271EXPORT_SYMBOL(__cond_resched_softirq); 4272 4273/** 4274 * yield - yield the current processor to other threads. 4275 * 4276 * Do not ever use this function, there's a 99% chance you're doing it wrong. 4277 * 4278 * The scheduler is at all times free to pick the calling task as the most 4279 * eligible task to run, if removing the yield() call from your code breaks 4280 * it, its already broken. 4281 * 4282 * Typical broken usage is: 4283 * 4284 * while (!event) 4285 * yield(); 4286 * 4287 * where one assumes that yield() will let 'the other' process run that will 4288 * make event true. If the current task is a SCHED_FIFO task that will never 4289 * happen. Never use yield() as a progress guarantee!! 4290 * 4291 * If you want to use yield() to wait for something, use wait_event(). 4292 * If you want to use yield() to be 'nice' for others, use cond_resched(). 4293 * If you still want to use yield(), do not! 4294 */ 4295void __sched yield(void) 4296{ 4297 set_current_state(TASK_RUNNING); 4298 sys_sched_yield(); 4299} 4300EXPORT_SYMBOL(yield); 4301 4302/** 4303 * yield_to - yield the current processor to another thread in 4304 * your thread group, or accelerate that thread toward the 4305 * processor it's on. 4306 * @p: target task 4307 * @preempt: whether task preemption is allowed or not 4308 * 4309 * It's the caller's job to ensure that the target task struct 4310 * can't go away on us before we can do any checks. 4311 * 4312 * Return: 4313 * true (>0) if we indeed boosted the target task. 4314 * false (0) if we failed to boost the target. 4315 * -ESRCH if there's no task to yield to. 4316 */ 4317int __sched yield_to(struct task_struct *p, bool preempt) 4318{ 4319 struct task_struct *curr = current; 4320 struct rq *rq, *p_rq; 4321 unsigned long flags; 4322 int yielded = 0; 4323 4324 local_irq_save(flags); 4325 rq = this_rq(); 4326 4327again: 4328 p_rq = task_rq(p); 4329 /* 4330 * If we're the only runnable task on the rq and target rq also 4331 * has only one task, there's absolutely no point in yielding. 4332 */ 4333 if (rq->nr_running == 1 && p_rq->nr_running == 1) { 4334 yielded = -ESRCH; 4335 goto out_irq; 4336 } 4337 4338 double_rq_lock(rq, p_rq); 4339 if (task_rq(p) != p_rq) { 4340 double_rq_unlock(rq, p_rq); 4341 goto again; 4342 } 4343 4344 if (!curr->sched_class->yield_to_task) 4345 goto out_unlock; 4346 4347 if (curr->sched_class != p->sched_class) 4348 goto out_unlock; 4349 4350 if (task_running(p_rq, p) || p->state) 4351 goto out_unlock; 4352 4353 yielded = curr->sched_class->yield_to_task(rq, p, preempt); 4354 if (yielded) { 4355 schedstat_inc(rq, yld_count); 4356 /* 4357 * Make p's CPU reschedule; pick_next_entity takes care of 4358 * fairness. 4359 */ 4360 if (preempt && rq != p_rq) 4361 resched_curr(p_rq); 4362 } 4363 4364out_unlock: 4365 double_rq_unlock(rq, p_rq); 4366out_irq: 4367 local_irq_restore(flags); 4368 4369 if (yielded > 0) 4370 schedule(); 4371 4372 return yielded; 4373} 4374EXPORT_SYMBOL_GPL(yield_to); 4375 4376/* 4377 * This task is about to go to sleep on IO. Increment rq->nr_iowait so 4378 * that process accounting knows that this is a task in IO wait state. 4379 */ 4380void __sched io_schedule(void) 4381{ 4382 struct rq *rq = raw_rq(); 4383 4384 delayacct_blkio_start(); 4385 atomic_inc(&rq->nr_iowait); 4386 blk_flush_plug(current); 4387 current->in_iowait = 1; 4388 schedule(); 4389 current->in_iowait = 0; 4390 atomic_dec(&rq->nr_iowait); 4391 delayacct_blkio_end(); 4392} 4393EXPORT_SYMBOL(io_schedule); 4394 4395long __sched io_schedule_timeout(long timeout) 4396{ 4397 struct rq *rq = raw_rq(); 4398 long ret; 4399 4400 delayacct_blkio_start(); 4401 atomic_inc(&rq->nr_iowait); 4402 blk_flush_plug(current); 4403 current->in_iowait = 1; 4404 ret = schedule_timeout(timeout); 4405 current->in_iowait = 0; 4406 atomic_dec(&rq->nr_iowait); 4407 delayacct_blkio_end(); 4408 return ret; 4409} 4410 4411/** 4412 * sys_sched_get_priority_max - return maximum RT priority. 4413 * @policy: scheduling class. 4414 * 4415 * Return: On success, this syscall returns the maximum 4416 * rt_priority that can be used by a given scheduling class. 4417 * On failure, a negative error code is returned. 4418 */ 4419SYSCALL_DEFINE1(sched_get_priority_max, int, policy) 4420{ 4421 int ret = -EINVAL; 4422 4423 switch (policy) { 4424 case SCHED_FIFO: 4425 case SCHED_RR: 4426 ret = MAX_USER_RT_PRIO-1; 4427 break; 4428 case SCHED_DEADLINE: 4429 case SCHED_NORMAL: 4430 case SCHED_BATCH: 4431 case SCHED_IDLE: 4432 ret = 0; 4433 break; 4434 } 4435 return ret; 4436} 4437 4438/** 4439 * sys_sched_get_priority_min - return minimum RT priority. 4440 * @policy: scheduling class. 4441 * 4442 * Return: On success, this syscall returns the minimum 4443 * rt_priority that can be used by a given scheduling class. 4444 * On failure, a negative error code is returned. 4445 */ 4446SYSCALL_DEFINE1(sched_get_priority_min, int, policy) 4447{ 4448 int ret = -EINVAL; 4449 4450 switch (policy) { 4451 case SCHED_FIFO: 4452 case SCHED_RR: 4453 ret = 1; 4454 break; 4455 case SCHED_DEADLINE: 4456 case SCHED_NORMAL: 4457 case SCHED_BATCH: 4458 case SCHED_IDLE: 4459 ret = 0; 4460 } 4461 return ret; 4462} 4463 4464/** 4465 * sys_sched_rr_get_interval - return the default timeslice of a process. 4466 * @pid: pid of the process. 4467 * @interval: userspace pointer to the timeslice value. 4468 * 4469 * this syscall writes the default timeslice value of a given process 4470 * into the user-space timespec buffer. A value of '0' means infinity. 4471 * 4472 * Return: On success, 0 and the timeslice is in @interval. Otherwise, 4473 * an error code. 4474 */ 4475SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid, 4476 struct timespec __user *, interval) 4477{ 4478 struct task_struct *p; 4479 unsigned int time_slice; 4480 unsigned long flags; 4481 struct rq *rq; 4482 int retval; 4483 struct timespec t; 4484 4485 if (pid < 0) 4486 return -EINVAL; 4487 4488 retval = -ESRCH; 4489 rcu_read_lock(); 4490 p = find_process_by_pid(pid); 4491 if (!p) 4492 goto out_unlock; 4493 4494 retval = security_task_getscheduler(p); 4495 if (retval) 4496 goto out_unlock; 4497 4498 rq = task_rq_lock(p, &flags); 4499 time_slice = 0; 4500 if (p->sched_class->get_rr_interval) 4501 time_slice = p->sched_class->get_rr_interval(rq, p); 4502 task_rq_unlock(rq, p, &flags); 4503 4504 rcu_read_unlock(); 4505 jiffies_to_timespec(time_slice, &t); 4506 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0; 4507 return retval; 4508 4509out_unlock: 4510 rcu_read_unlock(); 4511 return retval; 4512} 4513 4514static const char stat_nam[] = TASK_STATE_TO_CHAR_STR; 4515 4516void sched_show_task(struct task_struct *p) 4517{ 4518 unsigned long free = 0; 4519 int ppid; 4520 unsigned state; 4521 4522 state = p->state ? __ffs(p->state) + 1 : 0; 4523 printk(KERN_INFO "%-15.15s %c", p->comm, 4524 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?'); 4525#if BITS_PER_LONG == 32 4526 if (state == TASK_RUNNING) 4527 printk(KERN_CONT " running "); 4528 else 4529 printk(KERN_CONT " %08lx ", thread_saved_pc(p)); 4530#else 4531 if (state == TASK_RUNNING) 4532 printk(KERN_CONT " running task "); 4533 else 4534 printk(KERN_CONT " %016lx ", thread_saved_pc(p)); 4535#endif 4536#ifdef CONFIG_DEBUG_STACK_USAGE 4537 free = stack_not_used(p); 4538#endif 4539 rcu_read_lock(); 4540 ppid = task_pid_nr(rcu_dereference(p->real_parent)); 4541 rcu_read_unlock(); 4542 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free, 4543 task_pid_nr(p), ppid, 4544 (unsigned long)task_thread_info(p)->flags); 4545 4546 print_worker_info(KERN_INFO, p); 4547 show_stack(p, NULL); 4548} 4549 4550void show_state_filter(unsigned long state_filter) 4551{ 4552 struct task_struct *g, *p; 4553 4554#if BITS_PER_LONG == 32 4555 printk(KERN_INFO 4556 " task PC stack pid father\n"); 4557#else 4558 printk(KERN_INFO 4559 " task PC stack pid father\n"); 4560#endif 4561 rcu_read_lock(); 4562 for_each_process_thread(g, p) { 4563 /* 4564 * reset the NMI-timeout, listing all files on a slow 4565 * console might take a lot of time: 4566 */ 4567 touch_nmi_watchdog(); 4568 if (!state_filter || (p->state & state_filter)) 4569 sched_show_task(p); 4570 } 4571 4572 touch_all_softlockup_watchdogs(); 4573 4574#ifdef CONFIG_SCHED_DEBUG 4575 sysrq_sched_debug_show(); 4576#endif 4577 rcu_read_unlock(); 4578 /* 4579 * Only show locks if all tasks are dumped: 4580 */ 4581 if (!state_filter) 4582 debug_show_all_locks(); 4583} 4584 4585void init_idle_bootup_task(struct task_struct *idle) 4586{ 4587 idle->sched_class = &idle_sched_class; 4588} 4589 4590/** 4591 * init_idle - set up an idle thread for a given CPU 4592 * @idle: task in question 4593 * @cpu: cpu the idle task belongs to 4594 * 4595 * NOTE: this function does not set the idle thread's NEED_RESCHED 4596 * flag, to make booting more robust. 4597 */ 4598void init_idle(struct task_struct *idle, int cpu) 4599{ 4600 struct rq *rq = cpu_rq(cpu); 4601 unsigned long flags; 4602 4603 raw_spin_lock_irqsave(&rq->lock, flags); 4604 4605 __sched_fork(0, idle); 4606 idle->state = TASK_RUNNING; 4607 idle->se.exec_start = sched_clock(); 4608 4609 do_set_cpus_allowed(idle, cpumask_of(cpu)); 4610 /* 4611 * We're having a chicken and egg problem, even though we are 4612 * holding rq->lock, the cpu isn't yet set to this cpu so the 4613 * lockdep check in task_group() will fail. 4614 * 4615 * Similar case to sched_fork(). / Alternatively we could 4616 * use task_rq_lock() here and obtain the other rq->lock. 4617 * 4618 * Silence PROVE_RCU 4619 */ 4620 rcu_read_lock(); 4621 __set_task_cpu(idle, cpu); 4622 rcu_read_unlock(); 4623 4624 rq->curr = rq->idle = idle; 4625 idle->on_rq = TASK_ON_RQ_QUEUED; 4626#if defined(CONFIG_SMP) 4627 idle->on_cpu = 1; 4628#endif 4629 raw_spin_unlock_irqrestore(&rq->lock, flags); 4630 4631 /* Set the preempt count _outside_ the spinlocks! */ 4632 init_idle_preempt_count(idle, cpu); 4633 4634 /* 4635 * The idle tasks have their own, simple scheduling class: 4636 */ 4637 idle->sched_class = &idle_sched_class; 4638 ftrace_graph_init_idle_task(idle, cpu); 4639 vtime_init_idle(idle, cpu); 4640#if defined(CONFIG_SMP) 4641 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu); 4642#endif 4643} 4644 4645#ifdef CONFIG_SMP 4646/* 4647 * move_queued_task - move a queued task to new rq. 4648 * 4649 * Returns (locked) new rq. Old rq's lock is released. 4650 */ 4651static struct rq *move_queued_task(struct task_struct *p, int new_cpu) 4652{ 4653 struct rq *rq = task_rq(p); 4654 4655 lockdep_assert_held(&rq->lock); 4656 4657 dequeue_task(rq, p, 0); 4658 p->on_rq = TASK_ON_RQ_MIGRATING; 4659 set_task_cpu(p, new_cpu); 4660 raw_spin_unlock(&rq->lock); 4661 4662 rq = cpu_rq(new_cpu); 4663 4664 raw_spin_lock(&rq->lock); 4665 BUG_ON(task_cpu(p) != new_cpu); 4666 p->on_rq = TASK_ON_RQ_QUEUED; 4667 enqueue_task(rq, p, 0); 4668 check_preempt_curr(rq, p, 0); 4669 4670 return rq; 4671} 4672 4673void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask) 4674{ 4675 if (p->sched_class && p->sched_class->set_cpus_allowed) 4676 p->sched_class->set_cpus_allowed(p, new_mask); 4677 4678 cpumask_copy(&p->cpus_allowed, new_mask); 4679 p->nr_cpus_allowed = cpumask_weight(new_mask); 4680} 4681 4682/* 4683 * This is how migration works: 4684 * 4685 * 1) we invoke migration_cpu_stop() on the target CPU using 4686 * stop_one_cpu(). 4687 * 2) stopper starts to run (implicitly forcing the migrated thread 4688 * off the CPU) 4689 * 3) it checks whether the migrated task is still in the wrong runqueue. 4690 * 4) if it's in the wrong runqueue then the migration thread removes 4691 * it and puts it into the right queue. 4692 * 5) stopper completes and stop_one_cpu() returns and the migration 4693 * is done. 4694 */ 4695 4696/* 4697 * Change a given task's CPU affinity. Migrate the thread to a 4698 * proper CPU and schedule it away if the CPU it's executing on 4699 * is removed from the allowed bitmask. 4700 * 4701 * NOTE: the caller must have a valid reference to the task, the 4702 * task must not exit() & deallocate itself prematurely. The 4703 * call is not atomic; no spinlocks may be held. 4704 */ 4705int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask) 4706{ 4707 unsigned long flags; 4708 struct rq *rq; 4709 unsigned int dest_cpu; 4710 int ret = 0; 4711 4712 rq = task_rq_lock(p, &flags); 4713 4714 if (cpumask_equal(&p->cpus_allowed, new_mask)) 4715 goto out; 4716 4717 if (!cpumask_intersects(new_mask, cpu_active_mask)) { 4718 ret = -EINVAL; 4719 goto out; 4720 } 4721 4722 do_set_cpus_allowed(p, new_mask); 4723 4724 /* Can the task run on the task's current CPU? If so, we're done */ 4725 if (cpumask_test_cpu(task_cpu(p), new_mask)) 4726 goto out; 4727 4728 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask); 4729 if (task_running(rq, p) || p->state == TASK_WAKING) { 4730 struct migration_arg arg = { p, dest_cpu }; 4731 /* Need help from migration thread: drop lock and wait. */ 4732 task_rq_unlock(rq, p, &flags); 4733 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg); 4734 tlb_migrate_finish(p->mm); 4735 return 0; 4736 } else if (task_on_rq_queued(p)) 4737 rq = move_queued_task(p, dest_cpu); 4738out: 4739 task_rq_unlock(rq, p, &flags); 4740 4741 return ret; 4742} 4743EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr); 4744 4745/* 4746 * Move (not current) task off this cpu, onto dest cpu. We're doing 4747 * this because either it can't run here any more (set_cpus_allowed() 4748 * away from this CPU, or CPU going down), or because we're 4749 * attempting to rebalance this task on exec (sched_exec). 4750 * 4751 * So we race with normal scheduler movements, but that's OK, as long 4752 * as the task is no longer on this CPU. 4753 * 4754 * Returns non-zero if task was successfully migrated. 4755 */ 4756static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu) 4757{ 4758 struct rq *rq; 4759 int ret = 0; 4760 4761 if (unlikely(!cpu_active(dest_cpu))) 4762 return ret; 4763 4764 rq = cpu_rq(src_cpu); 4765 4766 raw_spin_lock(&p->pi_lock); 4767 raw_spin_lock(&rq->lock); 4768 /* Already moved. */ 4769 if (task_cpu(p) != src_cpu) 4770 goto done; 4771 4772 /* Affinity changed (again). */ 4773 if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p))) 4774 goto fail; 4775 4776 /* 4777 * If we're not on a rq, the next wake-up will ensure we're 4778 * placed properly. 4779 */ 4780 if (task_on_rq_queued(p)) 4781 rq = move_queued_task(p, dest_cpu); 4782done: 4783 ret = 1; 4784fail: 4785 raw_spin_unlock(&rq->lock); 4786 raw_spin_unlock(&p->pi_lock); 4787 return ret; 4788} 4789 4790#ifdef CONFIG_NUMA_BALANCING 4791/* Migrate current task p to target_cpu */ 4792int migrate_task_to(struct task_struct *p, int target_cpu) 4793{ 4794 struct migration_arg arg = { p, target_cpu }; 4795 int curr_cpu = task_cpu(p); 4796 4797 if (curr_cpu == target_cpu) 4798 return 0; 4799 4800 if (!cpumask_test_cpu(target_cpu, tsk_cpus_allowed(p))) 4801 return -EINVAL; 4802 4803 /* TODO: This is not properly updating schedstats */ 4804 4805 trace_sched_move_numa(p, curr_cpu, target_cpu); 4806 return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg); 4807} 4808 4809/* 4810 * Requeue a task on a given node and accurately track the number of NUMA 4811 * tasks on the runqueues 4812 */ 4813void sched_setnuma(struct task_struct *p, int nid) 4814{ 4815 struct rq *rq; 4816 unsigned long flags; 4817 bool queued, running; 4818 4819 rq = task_rq_lock(p, &flags); 4820 queued = task_on_rq_queued(p); 4821 running = task_current(rq, p); 4822 4823 if (queued) 4824 dequeue_task(rq, p, 0); 4825 if (running) 4826 put_prev_task(rq, p); 4827 4828 p->numa_preferred_nid = nid; 4829 4830 if (running) 4831 p->sched_class->set_curr_task(rq); 4832 if (queued) 4833 enqueue_task(rq, p, 0); 4834 task_rq_unlock(rq, p, &flags); 4835} 4836#endif 4837 4838/* 4839 * migration_cpu_stop - this will be executed by a highprio stopper thread 4840 * and performs thread migration by bumping thread off CPU then 4841 * 'pushing' onto another runqueue. 4842 */ 4843static int migration_cpu_stop(void *data) 4844{ 4845 struct migration_arg *arg = data; 4846 4847 /* 4848 * The original target cpu might have gone down and we might 4849 * be on another cpu but it doesn't matter. 4850 */ 4851 local_irq_disable(); 4852 /* 4853 * We need to explicitly wake pending tasks before running 4854 * __migrate_task() such that we will not miss enforcing cpus_allowed 4855 * during wakeups, see set_cpus_allowed_ptr()'s TASK_WAKING test. 4856 */ 4857 sched_ttwu_pending(); 4858 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu); 4859 local_irq_enable(); 4860 return 0; 4861} 4862 4863#ifdef CONFIG_HOTPLUG_CPU 4864 4865/* 4866 * Ensures that the idle task is using init_mm right before its cpu goes 4867 * offline. 4868 */ 4869void idle_task_exit(void) 4870{ 4871 struct mm_struct *mm = current->active_mm; 4872 4873 BUG_ON(cpu_online(smp_processor_id())); 4874 4875 if (mm != &init_mm) { 4876 switch_mm(mm, &init_mm, current); 4877 finish_arch_post_lock_switch(); 4878 } 4879 mmdrop(mm); 4880} 4881 4882/* 4883 * Since this CPU is going 'away' for a while, fold any nr_active delta 4884 * we might have. Assumes we're called after migrate_tasks() so that the 4885 * nr_active count is stable. 4886 * 4887 * Also see the comment "Global load-average calculations". 4888 */ 4889static void calc_load_migrate(struct rq *rq) 4890{ 4891 long delta = calc_load_fold_active(rq); 4892 if (delta) 4893 atomic_long_add(delta, &calc_load_tasks); 4894} 4895 4896static void put_prev_task_fake(struct rq *rq, struct task_struct *prev) 4897{ 4898} 4899 4900static const struct sched_class fake_sched_class = { 4901 .put_prev_task = put_prev_task_fake, 4902}; 4903 4904static struct task_struct fake_task = { 4905 /* 4906 * Avoid pull_{rt,dl}_task() 4907 */ 4908 .prio = MAX_PRIO + 1, 4909 .sched_class = &fake_sched_class, 4910}; 4911 4912/* 4913 * Migrate all tasks from the rq, sleeping tasks will be migrated by 4914 * try_to_wake_up()->select_task_rq(). 4915 * 4916 * Called with rq->lock held even though we'er in stop_machine() and 4917 * there's no concurrency possible, we hold the required locks anyway 4918 * because of lock validation efforts. 4919 */ 4920static void migrate_tasks(unsigned int dead_cpu) 4921{ 4922 struct rq *rq = cpu_rq(dead_cpu); 4923 struct task_struct *next, *stop = rq->stop; 4924 int dest_cpu; 4925 4926 /* 4927 * Fudge the rq selection such that the below task selection loop 4928 * doesn't get stuck on the currently eligible stop task. 4929 * 4930 * We're currently inside stop_machine() and the rq is either stuck 4931 * in the stop_machine_cpu_stop() loop, or we're executing this code, 4932 * either way we should never end up calling schedule() until we're 4933 * done here. 4934 */ 4935 rq->stop = NULL; 4936 4937 /* 4938 * put_prev_task() and pick_next_task() sched 4939 * class method both need to have an up-to-date 4940 * value of rq->clock[_task] 4941 */ 4942 update_rq_clock(rq); 4943 4944 for ( ; ; ) { 4945 /* 4946 * There's this thread running, bail when that's the only 4947 * remaining thread. 4948 */ 4949 if (rq->nr_running == 1) 4950 break; 4951 4952 next = pick_next_task(rq, &fake_task); 4953 BUG_ON(!next); 4954 next->sched_class->put_prev_task(rq, next); 4955 4956 /* Find suitable destination for @next, with force if needed. */ 4957 dest_cpu = select_fallback_rq(dead_cpu, next); 4958 raw_spin_unlock(&rq->lock); 4959 4960 __migrate_task(next, dead_cpu, dest_cpu); 4961 4962 raw_spin_lock(&rq->lock); 4963 } 4964 4965 rq->stop = stop; 4966} 4967 4968#endif /* CONFIG_HOTPLUG_CPU */ 4969 4970#if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL) 4971 4972static struct ctl_table sd_ctl_dir[] = { 4973 { 4974 .procname = "sched_domain", 4975 .mode = 0555, 4976 }, 4977 {} 4978}; 4979 4980static struct ctl_table sd_ctl_root[] = { 4981 { 4982 .procname = "kernel", 4983 .mode = 0555, 4984 .child = sd_ctl_dir, 4985 }, 4986 {} 4987}; 4988 4989static struct ctl_table *sd_alloc_ctl_entry(int n) 4990{ 4991 struct ctl_table *entry = 4992 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL); 4993 4994 return entry; 4995} 4996 4997static void sd_free_ctl_entry(struct ctl_table **tablep) 4998{ 4999 struct ctl_table *entry; 5000 5001 /* 5002 * In the intermediate directories, both the child directory and 5003 * procname are dynamically allocated and could fail but the mode 5004 * will always be set. In the lowest directory the names are 5005 * static strings and all have proc handlers. 5006 */ 5007 for (entry = *tablep; entry->mode; entry++) { 5008 if (entry->child) 5009 sd_free_ctl_entry(&entry->child); 5010 if (entry->proc_handler == NULL) 5011 kfree(entry->procname); 5012 } 5013 5014 kfree(*tablep); 5015 *tablep = NULL; 5016} 5017 5018static int min_load_idx = 0; 5019static int max_load_idx = CPU_LOAD_IDX_MAX-1; 5020 5021static void 5022set_table_entry(struct ctl_table *entry, 5023 const char *procname, void *data, int maxlen, 5024 umode_t mode, proc_handler *proc_handler, 5025 bool load_idx) 5026{ 5027 entry->procname = procname; 5028 entry->data = data; 5029 entry->maxlen = maxlen; 5030 entry->mode = mode; 5031 entry->proc_handler = proc_handler; 5032 5033 if (load_idx) { 5034 entry->extra1 = &min_load_idx; 5035 entry->extra2 = &max_load_idx; 5036 } 5037} 5038 5039static struct ctl_table * 5040sd_alloc_ctl_domain_table(struct sched_domain *sd) 5041{ 5042 struct ctl_table *table = sd_alloc_ctl_entry(14); 5043 5044 if (table == NULL) 5045 return NULL; 5046 5047 set_table_entry(&table[0], "min_interval", &sd->min_interval, 5048 sizeof(long), 0644, proc_doulongvec_minmax, false); 5049 set_table_entry(&table[1], "max_interval", &sd->max_interval, 5050 sizeof(long), 0644, proc_doulongvec_minmax, false); 5051 set_table_entry(&table[2], "busy_idx", &sd->busy_idx, 5052 sizeof(int), 0644, proc_dointvec_minmax, true); 5053 set_table_entry(&table[3], "idle_idx", &sd->idle_idx, 5054 sizeof(int), 0644, proc_dointvec_minmax, true); 5055 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx, 5056 sizeof(int), 0644, proc_dointvec_minmax, true); 5057 set_table_entry(&table[5], "wake_idx", &sd->wake_idx, 5058 sizeof(int), 0644, proc_dointvec_minmax, true); 5059 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx, 5060 sizeof(int), 0644, proc_dointvec_minmax, true); 5061 set_table_entry(&table[7], "busy_factor", &sd->busy_factor, 5062 sizeof(int), 0644, proc_dointvec_minmax, false); 5063 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct, 5064 sizeof(int), 0644, proc_dointvec_minmax, false); 5065 set_table_entry(&table[9], "cache_nice_tries", 5066 &sd->cache_nice_tries, 5067 sizeof(int), 0644, proc_dointvec_minmax, false); 5068 set_table_entry(&table[10], "flags", &sd->flags, 5069 sizeof(int), 0644, proc_dointvec_minmax, false); 5070 set_table_entry(&table[11], "max_newidle_lb_cost", 5071 &sd->max_newidle_lb_cost, 5072 sizeof(long), 0644, proc_doulongvec_minmax, false); 5073 set_table_entry(&table[12], "name", sd->name, 5074 CORENAME_MAX_SIZE, 0444, proc_dostring, false); 5075 /* &table[13] is terminator */ 5076 5077 return table; 5078} 5079 5080static struct ctl_table *sd_alloc_ctl_cpu_table(int cpu) 5081{ 5082 struct ctl_table *entry, *table; 5083 struct sched_domain *sd; 5084 int domain_num = 0, i; 5085 char buf[32]; 5086 5087 for_each_domain(cpu, sd) 5088 domain_num++; 5089 entry = table = sd_alloc_ctl_entry(domain_num + 1); 5090 if (table == NULL) 5091 return NULL; 5092 5093 i = 0; 5094 for_each_domain(cpu, sd) { 5095 snprintf(buf, 32, "domain%d", i); 5096 entry->procname = kstrdup(buf, GFP_KERNEL); 5097 entry->mode = 0555; 5098 entry->child = sd_alloc_ctl_domain_table(sd); 5099 entry++; 5100 i++; 5101 } 5102 return table; 5103} 5104 5105static struct ctl_table_header *sd_sysctl_header; 5106static void register_sched_domain_sysctl(void) 5107{ 5108 int i, cpu_num = num_possible_cpus(); 5109 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1); 5110 char buf[32]; 5111 5112 WARN_ON(sd_ctl_dir[0].child); 5113 sd_ctl_dir[0].child = entry; 5114 5115 if (entry == NULL) 5116 return; 5117 5118 for_each_possible_cpu(i) { 5119 snprintf(buf, 32, "cpu%d", i); 5120 entry->procname = kstrdup(buf, GFP_KERNEL); 5121 entry->mode = 0555; 5122 entry->child = sd_alloc_ctl_cpu_table(i); 5123 entry++; 5124 } 5125 5126 WARN_ON(sd_sysctl_header); 5127 sd_sysctl_header = register_sysctl_table(sd_ctl_root); 5128} 5129 5130/* may be called multiple times per register */ 5131static void unregister_sched_domain_sysctl(void) 5132{ 5133 if (sd_sysctl_header) 5134 unregister_sysctl_table(sd_sysctl_header); 5135 sd_sysctl_header = NULL; 5136 if (sd_ctl_dir[0].child) 5137 sd_free_ctl_entry(&sd_ctl_dir[0].child); 5138} 5139#else 5140static void register_sched_domain_sysctl(void) 5141{ 5142} 5143static void unregister_sched_domain_sysctl(void) 5144{ 5145} 5146#endif 5147 5148static void set_rq_online(struct rq *rq) 5149{ 5150 if (!rq->online) { 5151 const struct sched_class *class; 5152 5153 cpumask_set_cpu(rq->cpu, rq->rd->online); 5154 rq->online = 1; 5155 5156 for_each_class(class) { 5157 if (class->rq_online) 5158 class->rq_online(rq); 5159 } 5160 } 5161} 5162 5163static void set_rq_offline(struct rq *rq) 5164{ 5165 if (rq->online) { 5166 const struct sched_class *class; 5167 5168 for_each_class(class) { 5169 if (class->rq_offline) 5170 class->rq_offline(rq); 5171 } 5172 5173 cpumask_clear_cpu(rq->cpu, rq->rd->online); 5174 rq->online = 0; 5175 } 5176} 5177 5178/* 5179 * migration_call - callback that gets triggered when a CPU is added. 5180 * Here we can start up the necessary migration thread for the new CPU. 5181 */ 5182static int 5183migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu) 5184{ 5185 int cpu = (long)hcpu; 5186 unsigned long flags; 5187 struct rq *rq = cpu_rq(cpu); 5188 5189 switch (action & ~CPU_TASKS_FROZEN) { 5190 5191 case CPU_UP_PREPARE: 5192 rq->calc_load_update = calc_load_update; 5193 break; 5194 5195 case CPU_ONLINE: 5196 /* Update our root-domain */ 5197 raw_spin_lock_irqsave(&rq->lock, flags); 5198 if (rq->rd) { 5199 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span)); 5200 5201 set_rq_online(rq); 5202 } 5203 raw_spin_unlock_irqrestore(&rq->lock, flags); 5204 break; 5205 5206#ifdef CONFIG_HOTPLUG_CPU 5207 case CPU_DYING: 5208 sched_ttwu_pending(); 5209 /* Update our root-domain */ 5210 raw_spin_lock_irqsave(&rq->lock, flags); 5211 if (rq->rd) { 5212 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span)); 5213 set_rq_offline(rq); 5214 } 5215 migrate_tasks(cpu); 5216 BUG_ON(rq->nr_running != 1); /* the migration thread */ 5217 raw_spin_unlock_irqrestore(&rq->lock, flags); 5218 break; 5219 5220 case CPU_DEAD: 5221 calc_load_migrate(rq); 5222 break; 5223#endif 5224 } 5225 5226 update_max_interval(); 5227 5228 return NOTIFY_OK; 5229} 5230 5231/* 5232 * Register at high priority so that task migration (migrate_all_tasks) 5233 * happens before everything else. This has to be lower priority than 5234 * the notifier in the perf_event subsystem, though. 5235 */ 5236static struct notifier_block migration_notifier = { 5237 .notifier_call = migration_call, 5238 .priority = CPU_PRI_MIGRATION, 5239}; 5240 5241static void __cpuinit set_cpu_rq_start_time(void) 5242{ 5243 int cpu = smp_processor_id(); 5244 struct rq *rq = cpu_rq(cpu); 5245 rq->age_stamp = sched_clock_cpu(cpu); 5246} 5247 5248static int sched_cpu_active(struct notifier_block *nfb, 5249 unsigned long action, void *hcpu) 5250{ 5251 switch (action & ~CPU_TASKS_FROZEN) { 5252 case CPU_STARTING: 5253 set_cpu_rq_start_time(); 5254 return NOTIFY_OK; 5255 case CPU_DOWN_FAILED: 5256 set_cpu_active((long)hcpu, true); 5257 return NOTIFY_OK; 5258 default: 5259 return NOTIFY_DONE; 5260 } 5261} 5262 5263static int sched_cpu_inactive(struct notifier_block *nfb, 5264 unsigned long action, void *hcpu) 5265{ 5266 unsigned long flags; 5267 long cpu = (long)hcpu; 5268 5269 switch (action & ~CPU_TASKS_FROZEN) { 5270 case CPU_DOWN_PREPARE: 5271 set_cpu_active(cpu, false); 5272 5273 /* explicitly allow suspend */ 5274 if (!(action & CPU_TASKS_FROZEN)) { 5275 struct dl_bw *dl_b = dl_bw_of(cpu); 5276 bool overflow; 5277 int cpus; 5278 5279 raw_spin_lock_irqsave(&dl_b->lock, flags); 5280 cpus = dl_bw_cpus(cpu); 5281 overflow = __dl_overflow(dl_b, cpus, 0, 0); 5282 raw_spin_unlock_irqrestore(&dl_b->lock, flags); 5283 5284 if (overflow) 5285 return notifier_from_errno(-EBUSY); 5286 } 5287 return NOTIFY_OK; 5288 } 5289 5290 return NOTIFY_DONE; 5291} 5292 5293static int __init migration_init(void) 5294{ 5295 void *cpu = (void *)(long)smp_processor_id(); 5296 int err; 5297 5298 /* Initialize migration for the boot CPU */ 5299 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu); 5300 BUG_ON(err == NOTIFY_BAD); 5301 migration_call(&migration_notifier, CPU_ONLINE, cpu); 5302 register_cpu_notifier(&migration_notifier); 5303 5304 /* Register cpu active notifiers */ 5305 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE); 5306 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE); 5307 5308 return 0; 5309} 5310early_initcall(migration_init); 5311#endif 5312 5313#ifdef CONFIG_SMP 5314 5315static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */ 5316 5317#ifdef CONFIG_SCHED_DEBUG 5318 5319static __read_mostly int sched_debug_enabled; 5320 5321static int __init sched_debug_setup(char *str) 5322{ 5323 sched_debug_enabled = 1; 5324 5325 return 0; 5326} 5327early_param("sched_debug", sched_debug_setup); 5328 5329static inline bool sched_debug(void) 5330{ 5331 return sched_debug_enabled; 5332} 5333 5334static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level, 5335 struct cpumask *groupmask) 5336{ 5337 struct sched_group *group = sd->groups; 5338 char str[256]; 5339 5340 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd)); 5341 cpumask_clear(groupmask); 5342 5343 printk(KERN_DEBUG "%*s domain %d: ", level, "", level); 5344 5345 if (!(sd->flags & SD_LOAD_BALANCE)) { 5346 printk("does not load-balance\n"); 5347 if (sd->parent) 5348 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain" 5349 " has parent"); 5350 return -1; 5351 } 5352 5353 printk(KERN_CONT "span %s level %s\n", str, sd->name); 5354 5355 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) { 5356 printk(KERN_ERR "ERROR: domain->span does not contain " 5357 "CPU%d\n", cpu); 5358 } 5359 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) { 5360 printk(KERN_ERR "ERROR: domain->groups does not contain" 5361 " CPU%d\n", cpu); 5362 } 5363 5364 printk(KERN_DEBUG "%*s groups:", level + 1, ""); 5365 do { 5366 if (!group) { 5367 printk("\n"); 5368 printk(KERN_ERR "ERROR: group is NULL\n"); 5369 break; 5370 } 5371 5372 /* 5373 * Even though we initialize ->capacity to something semi-sane, 5374 * we leave capacity_orig unset. This allows us to detect if 5375 * domain iteration is still funny without causing /0 traps. 5376 */ 5377 if (!group->sgc->capacity_orig) { 5378 printk(KERN_CONT "\n"); 5379 printk(KERN_ERR "ERROR: domain->cpu_capacity not set\n"); 5380 break; 5381 } 5382 5383 if (!cpumask_weight(sched_group_cpus(group))) { 5384 printk(KERN_CONT "\n"); 5385 printk(KERN_ERR "ERROR: empty group\n"); 5386 break; 5387 } 5388 5389 if (!(sd->flags & SD_OVERLAP) && 5390 cpumask_intersects(groupmask, sched_group_cpus(group))) { 5391 printk(KERN_CONT "\n"); 5392 printk(KERN_ERR "ERROR: repeated CPUs\n"); 5393 break; 5394 } 5395 5396 cpumask_or(groupmask, groupmask, sched_group_cpus(group)); 5397 5398 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group)); 5399 5400 printk(KERN_CONT " %s", str); 5401 if (group->sgc->capacity != SCHED_CAPACITY_SCALE) { 5402 printk(KERN_CONT " (cpu_capacity = %d)", 5403 group->sgc->capacity); 5404 } 5405 5406 group = group->next; 5407 } while (group != sd->groups); 5408 printk(KERN_CONT "\n"); 5409 5410 if (!cpumask_equal(sched_domain_span(sd), groupmask)) 5411 printk(KERN_ERR "ERROR: groups don't span domain->span\n"); 5412 5413 if (sd->parent && 5414 !cpumask_subset(groupmask, sched_domain_span(sd->parent))) 5415 printk(KERN_ERR "ERROR: parent span is not a superset " 5416 "of domain->span\n"); 5417 return 0; 5418} 5419 5420static void sched_domain_debug(struct sched_domain *sd, int cpu) 5421{ 5422 int level = 0; 5423 5424 if (!sched_debug_enabled) 5425 return; 5426 5427 if (!sd) { 5428 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu); 5429 return; 5430 } 5431 5432 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu); 5433 5434 for (;;) { 5435 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask)) 5436 break; 5437 level++; 5438 sd = sd->parent; 5439 if (!sd) 5440 break; 5441 } 5442} 5443#else /* !CONFIG_SCHED_DEBUG */ 5444# define sched_domain_debug(sd, cpu) do { } while (0) 5445static inline bool sched_debug(void) 5446{ 5447 return false; 5448} 5449#endif /* CONFIG_SCHED_DEBUG */ 5450 5451static int sd_degenerate(struct sched_domain *sd) 5452{ 5453 if (cpumask_weight(sched_domain_span(sd)) == 1) 5454 return 1; 5455 5456 /* Following flags need at least 2 groups */ 5457 if (sd->flags & (SD_LOAD_BALANCE | 5458 SD_BALANCE_NEWIDLE | 5459 SD_BALANCE_FORK | 5460 SD_BALANCE_EXEC | 5461 SD_SHARE_CPUCAPACITY | 5462 SD_SHARE_PKG_RESOURCES | 5463 SD_SHARE_POWERDOMAIN)) { 5464 if (sd->groups != sd->groups->next) 5465 return 0; 5466 } 5467 5468 /* Following flags don't use groups */ 5469 if (sd->flags & (SD_WAKE_AFFINE)) 5470 return 0; 5471 5472 return 1; 5473} 5474 5475static int 5476sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent) 5477{ 5478 unsigned long cflags = sd->flags, pflags = parent->flags; 5479 5480 if (sd_degenerate(parent)) 5481 return 1; 5482 5483 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent))) 5484 return 0; 5485 5486 /* Flags needing groups don't count if only 1 group in parent */ 5487 if (parent->groups == parent->groups->next) { 5488 pflags &= ~(SD_LOAD_BALANCE | 5489 SD_BALANCE_NEWIDLE | 5490 SD_BALANCE_FORK | 5491 SD_BALANCE_EXEC | 5492 SD_SHARE_CPUCAPACITY | 5493 SD_SHARE_PKG_RESOURCES | 5494 SD_PREFER_SIBLING | 5495 SD_SHARE_POWERDOMAIN); 5496 if (nr_node_ids == 1) 5497 pflags &= ~SD_SERIALIZE; 5498 } 5499 if (~cflags & pflags) 5500 return 0; 5501 5502 return 1; 5503} 5504 5505static void free_rootdomain(struct rcu_head *rcu) 5506{ 5507 struct root_domain *rd = container_of(rcu, struct root_domain, rcu); 5508 5509 cpupri_cleanup(&rd->cpupri); 5510 cpudl_cleanup(&rd->cpudl); 5511 free_cpumask_var(rd->dlo_mask); 5512 free_cpumask_var(rd->rto_mask); 5513 free_cpumask_var(rd->online); 5514 free_cpumask_var(rd->span); 5515 kfree(rd); 5516} 5517 5518static void rq_attach_root(struct rq *rq, struct root_domain *rd) 5519{ 5520 struct root_domain *old_rd = NULL; 5521 unsigned long flags; 5522 5523 raw_spin_lock_irqsave(&rq->lock, flags); 5524 5525 if (rq->rd) { 5526 old_rd = rq->rd; 5527 5528 if (cpumask_test_cpu(rq->cpu, old_rd->online)) 5529 set_rq_offline(rq); 5530 5531 cpumask_clear_cpu(rq->cpu, old_rd->span); 5532 5533 /* 5534 * If we dont want to free the old_rd yet then 5535 * set old_rd to NULL to skip the freeing later 5536 * in this function: 5537 */ 5538 if (!atomic_dec_and_test(&old_rd->refcount)) 5539 old_rd = NULL; 5540 } 5541 5542 atomic_inc(&rd->refcount); 5543 rq->rd = rd; 5544 5545 cpumask_set_cpu(rq->cpu, rd->span); 5546 if (cpumask_test_cpu(rq->cpu, cpu_active_mask)) 5547 set_rq_online(rq); 5548 5549 raw_spin_unlock_irqrestore(&rq->lock, flags); 5550 5551 if (old_rd) 5552 call_rcu_sched(&old_rd->rcu, free_rootdomain); 5553} 5554 5555static int init_rootdomain(struct root_domain *rd) 5556{ 5557 memset(rd, 0, sizeof(*rd)); 5558 5559 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL)) 5560 goto out; 5561 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL)) 5562 goto free_span; 5563 if (!alloc_cpumask_var(&rd->dlo_mask, GFP_KERNEL)) 5564 goto free_online; 5565 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL)) 5566 goto free_dlo_mask; 5567 5568 init_dl_bw(&rd->dl_bw); 5569 if (cpudl_init(&rd->cpudl) != 0) 5570 goto free_dlo_mask; 5571 5572 if (cpupri_init(&rd->cpupri) != 0) 5573 goto free_rto_mask; 5574 return 0; 5575 5576free_rto_mask: 5577 free_cpumask_var(rd->rto_mask); 5578free_dlo_mask: 5579 free_cpumask_var(rd->dlo_mask); 5580free_online: 5581 free_cpumask_var(rd->online); 5582free_span: 5583 free_cpumask_var(rd->span); 5584out: 5585 return -ENOMEM; 5586} 5587 5588/* 5589 * By default the system creates a single root-domain with all cpus as 5590 * members (mimicking the global state we have today). 5591 */ 5592struct root_domain def_root_domain; 5593 5594static void init_defrootdomain(void) 5595{ 5596 init_rootdomain(&def_root_domain); 5597 5598 atomic_set(&def_root_domain.refcount, 1); 5599} 5600 5601static struct root_domain *alloc_rootdomain(void) 5602{ 5603 struct root_domain *rd; 5604 5605 rd = kmalloc(sizeof(*rd), GFP_KERNEL); 5606 if (!rd) 5607 return NULL; 5608 5609 if (init_rootdomain(rd) != 0) { 5610 kfree(rd); 5611 return NULL; 5612 } 5613 5614 return rd; 5615} 5616 5617static void free_sched_groups(struct sched_group *sg, int free_sgc) 5618{ 5619 struct sched_group *tmp, *first; 5620 5621 if (!sg) 5622 return; 5623 5624 first = sg; 5625 do { 5626 tmp = sg->next; 5627 5628 if (free_sgc && atomic_dec_and_test(&sg->sgc->ref)) 5629 kfree(sg->sgc); 5630 5631 kfree(sg); 5632 sg = tmp; 5633 } while (sg != first); 5634} 5635 5636static void free_sched_domain(struct rcu_head *rcu) 5637{ 5638 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu); 5639 5640 /* 5641 * If its an overlapping domain it has private groups, iterate and 5642 * nuke them all. 5643 */ 5644 if (sd->flags & SD_OVERLAP) { 5645 free_sched_groups(sd->groups, 1); 5646 } else if (atomic_dec_and_test(&sd->groups->ref)) { 5647 kfree(sd->groups->sgc); 5648 kfree(sd->groups); 5649 } 5650 kfree(sd); 5651} 5652 5653static void destroy_sched_domain(struct sched_domain *sd, int cpu) 5654{ 5655 call_rcu(&sd->rcu, free_sched_domain); 5656} 5657 5658static void destroy_sched_domains(struct sched_domain *sd, int cpu) 5659{ 5660 for (; sd; sd = sd->parent) 5661 destroy_sched_domain(sd, cpu); 5662} 5663 5664/* 5665 * Keep a special pointer to the highest sched_domain that has 5666 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this 5667 * allows us to avoid some pointer chasing select_idle_sibling(). 5668 * 5669 * Also keep a unique ID per domain (we use the first cpu number in 5670 * the cpumask of the domain), this allows us to quickly tell if 5671 * two cpus are in the same cache domain, see cpus_share_cache(). 5672 */ 5673DEFINE_PER_CPU(struct sched_domain *, sd_llc); 5674DEFINE_PER_CPU(int, sd_llc_size); 5675DEFINE_PER_CPU(int, sd_llc_id); 5676DEFINE_PER_CPU(struct sched_domain *, sd_numa); 5677DEFINE_PER_CPU(struct sched_domain *, sd_busy); 5678DEFINE_PER_CPU(struct sched_domain *, sd_asym); 5679 5680static void update_top_cache_domain(int cpu) 5681{ 5682 struct sched_domain *sd; 5683 struct sched_domain *busy_sd = NULL; 5684 int id = cpu; 5685 int size = 1; 5686 5687 sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES); 5688 if (sd) { 5689 id = cpumask_first(sched_domain_span(sd)); 5690 size = cpumask_weight(sched_domain_span(sd)); 5691 busy_sd = sd->parent; /* sd_busy */ 5692 } 5693 rcu_assign_pointer(per_cpu(sd_busy, cpu), busy_sd); 5694 5695 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd); 5696 per_cpu(sd_llc_size, cpu) = size; 5697 per_cpu(sd_llc_id, cpu) = id; 5698 5699 sd = lowest_flag_domain(cpu, SD_NUMA); 5700 rcu_assign_pointer(per_cpu(sd_numa, cpu), sd); 5701 5702 sd = highest_flag_domain(cpu, SD_ASYM_PACKING); 5703 rcu_assign_pointer(per_cpu(sd_asym, cpu), sd); 5704} 5705 5706/* 5707 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must 5708 * hold the hotplug lock. 5709 */ 5710static void 5711cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu) 5712{ 5713 struct rq *rq = cpu_rq(cpu); 5714 struct sched_domain *tmp; 5715 5716 /* Remove the sched domains which do not contribute to scheduling. */ 5717 for (tmp = sd; tmp; ) { 5718 struct sched_domain *parent = tmp->parent; 5719 if (!parent) 5720 break; 5721 5722 if (sd_parent_degenerate(tmp, parent)) { 5723 tmp->parent = parent->parent; 5724 if (parent->parent) 5725 parent->parent->child = tmp; 5726 /* 5727 * Transfer SD_PREFER_SIBLING down in case of a 5728 * degenerate parent; the spans match for this 5729 * so the property transfers. 5730 */ 5731 if (parent->flags & SD_PREFER_SIBLING) 5732 tmp->flags |= SD_PREFER_SIBLING; 5733 destroy_sched_domain(parent, cpu); 5734 } else 5735 tmp = tmp->parent; 5736 } 5737 5738 if (sd && sd_degenerate(sd)) { 5739 tmp = sd; 5740 sd = sd->parent; 5741 destroy_sched_domain(tmp, cpu); 5742 if (sd) 5743 sd->child = NULL; 5744 } 5745 5746 sched_domain_debug(sd, cpu); 5747 5748 rq_attach_root(rq, rd); 5749 tmp = rq->sd; 5750 rcu_assign_pointer(rq->sd, sd); 5751 destroy_sched_domains(tmp, cpu); 5752 5753 update_top_cache_domain(cpu); 5754} 5755 5756/* cpus with isolated domains */ 5757static cpumask_var_t cpu_isolated_map; 5758 5759/* Setup the mask of cpus configured for isolated domains */ 5760static int __init isolated_cpu_setup(char *str) 5761{ 5762 alloc_bootmem_cpumask_var(&cpu_isolated_map); 5763 cpulist_parse(str, cpu_isolated_map); 5764 return 1; 5765} 5766 5767__setup("isolcpus=", isolated_cpu_setup); 5768 5769struct s_data { 5770 struct sched_domain ** __percpu sd; 5771 struct root_domain *rd; 5772}; 5773 5774enum s_alloc { 5775 sa_rootdomain, 5776 sa_sd, 5777 sa_sd_storage, 5778 sa_none, 5779}; 5780 5781/* 5782 * Build an iteration mask that can exclude certain CPUs from the upwards 5783 * domain traversal. 5784 * 5785 * Asymmetric node setups can result in situations where the domain tree is of 5786 * unequal depth, make sure to skip domains that already cover the entire 5787 * range. 5788 * 5789 * In that case build_sched_domains() will have terminated the iteration early 5790 * and our sibling sd spans will be empty. Domains should always include the 5791 * cpu they're built on, so check that. 5792 * 5793 */ 5794static void build_group_mask(struct sched_domain *sd, struct sched_group *sg) 5795{ 5796 const struct cpumask *span = sched_domain_span(sd); 5797 struct sd_data *sdd = sd->private; 5798 struct sched_domain *sibling; 5799 int i; 5800 5801 for_each_cpu(i, span) { 5802 sibling = *per_cpu_ptr(sdd->sd, i); 5803 if (!cpumask_test_cpu(i, sched_domain_span(sibling))) 5804 continue; 5805 5806 cpumask_set_cpu(i, sched_group_mask(sg)); 5807 } 5808} 5809 5810/* 5811 * Return the canonical balance cpu for this group, this is the first cpu 5812 * of this group that's also in the iteration mask. 5813 */ 5814int group_balance_cpu(struct sched_group *sg) 5815{ 5816 return cpumask_first_and(sched_group_cpus(sg), sched_group_mask(sg)); 5817} 5818 5819static int 5820build_overlap_sched_groups(struct sched_domain *sd, int cpu) 5821{ 5822 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg; 5823 const struct cpumask *span = sched_domain_span(sd); 5824 struct cpumask *covered = sched_domains_tmpmask; 5825 struct sd_data *sdd = sd->private; 5826 struct sched_domain *sibling; 5827 int i; 5828 5829 cpumask_clear(covered); 5830 5831 for_each_cpu(i, span) { 5832 struct cpumask *sg_span; 5833 5834 if (cpumask_test_cpu(i, covered)) 5835 continue; 5836 5837 sibling = *per_cpu_ptr(sdd->sd, i); 5838 5839 /* See the comment near build_group_mask(). */ 5840 if (!cpumask_test_cpu(i, sched_domain_span(sibling))) 5841 continue; 5842 5843 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(), 5844 GFP_KERNEL, cpu_to_node(cpu)); 5845 5846 if (!sg) 5847 goto fail; 5848 5849 sg_span = sched_group_cpus(sg); 5850 if (sibling->child) 5851 cpumask_copy(sg_span, sched_domain_span(sibling->child)); 5852 else 5853 cpumask_set_cpu(i, sg_span); 5854 5855 cpumask_or(covered, covered, sg_span); 5856 5857 sg->sgc = *per_cpu_ptr(sdd->sgc, i); 5858 if (atomic_inc_return(&sg->sgc->ref) == 1) 5859 build_group_mask(sd, sg); 5860 5861 /* 5862 * Initialize sgc->capacity such that even if we mess up the 5863 * domains and no possible iteration will get us here, we won't 5864 * die on a /0 trap. 5865 */ 5866 sg->sgc->capacity = SCHED_CAPACITY_SCALE * cpumask_weight(sg_span); 5867 sg->sgc->capacity_orig = sg->sgc->capacity; 5868 5869 /* 5870 * Make sure the first group of this domain contains the 5871 * canonical balance cpu. Otherwise the sched_domain iteration 5872 * breaks. See update_sg_lb_stats(). 5873 */ 5874 if ((!groups && cpumask_test_cpu(cpu, sg_span)) || 5875 group_balance_cpu(sg) == cpu) 5876 groups = sg; 5877 5878 if (!first) 5879 first = sg; 5880 if (last) 5881 last->next = sg; 5882 last = sg; 5883 last->next = first; 5884 } 5885 sd->groups = groups; 5886 5887 return 0; 5888 5889fail: 5890 free_sched_groups(first, 0); 5891 5892 return -ENOMEM; 5893} 5894 5895static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg) 5896{ 5897 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu); 5898 struct sched_domain *child = sd->child; 5899 5900 if (child) 5901 cpu = cpumask_first(sched_domain_span(child)); 5902 5903 if (sg) { 5904 *sg = *per_cpu_ptr(sdd->sg, cpu); 5905 (*sg)->sgc = *per_cpu_ptr(sdd->sgc, cpu); 5906 atomic_set(&(*sg)->sgc->ref, 1); /* for claim_allocations */ 5907 } 5908 5909 return cpu; 5910} 5911 5912/* 5913 * build_sched_groups will build a circular linked list of the groups 5914 * covered by the given span, and will set each group's ->cpumask correctly, 5915 * and ->cpu_capacity to 0. 5916 * 5917 * Assumes the sched_domain tree is fully constructed 5918 */ 5919static int 5920build_sched_groups(struct sched_domain *sd, int cpu) 5921{ 5922 struct sched_group *first = NULL, *last = NULL; 5923 struct sd_data *sdd = sd->private; 5924 const struct cpumask *span = sched_domain_span(sd); 5925 struct cpumask *covered; 5926 int i; 5927 5928 get_group(cpu, sdd, &sd->groups); 5929 atomic_inc(&sd->groups->ref); 5930 5931 if (cpu != cpumask_first(span)) 5932 return 0; 5933 5934 lockdep_assert_held(&sched_domains_mutex); 5935 covered = sched_domains_tmpmask; 5936 5937 cpumask_clear(covered); 5938 5939 for_each_cpu(i, span) { 5940 struct sched_group *sg; 5941 int group, j; 5942 5943 if (cpumask_test_cpu(i, covered)) 5944 continue; 5945 5946 group = get_group(i, sdd, &sg); 5947 cpumask_setall(sched_group_mask(sg)); 5948 5949 for_each_cpu(j, span) { 5950 if (get_group(j, sdd, NULL) != group) 5951 continue; 5952 5953 cpumask_set_cpu(j, covered); 5954 cpumask_set_cpu(j, sched_group_cpus(sg)); 5955 } 5956 5957 if (!first) 5958 first = sg; 5959 if (last) 5960 last->next = sg; 5961 last = sg; 5962 } 5963 last->next = first; 5964 5965 return 0; 5966} 5967 5968/* 5969 * Initialize sched groups cpu_capacity. 5970 * 5971 * cpu_capacity indicates the capacity of sched group, which is used while 5972 * distributing the load between different sched groups in a sched domain. 5973 * Typically cpu_capacity for all the groups in a sched domain will be same 5974 * unless there are asymmetries in the topology. If there are asymmetries, 5975 * group having more cpu_capacity will pickup more load compared to the 5976 * group having less cpu_capacity. 5977 */ 5978static void init_sched_groups_capacity(int cpu, struct sched_domain *sd) 5979{ 5980 struct sched_group *sg = sd->groups; 5981 5982 WARN_ON(!sg); 5983 5984 do { 5985 sg->group_weight = cpumask_weight(sched_group_cpus(sg)); 5986 sg = sg->next; 5987 } while (sg != sd->groups); 5988 5989 if (cpu != group_balance_cpu(sg)) 5990 return; 5991 5992 update_group_capacity(sd, cpu); 5993 atomic_set(&sg->sgc->nr_busy_cpus, sg->group_weight); 5994} 5995 5996/* 5997 * Initializers for schedule domains 5998 * Non-inlined to reduce accumulated stack pressure in build_sched_domains() 5999 */ 6000 6001static int default_relax_domain_level = -1; 6002int sched_domain_level_max; 6003 6004static int __init setup_relax_domain_level(char *str) 6005{ 6006 if (kstrtoint(str, 0, &default_relax_domain_level)) 6007 pr_warn("Unable to set relax_domain_level\n"); 6008 6009 return 1; 6010} 6011__setup("relax_domain_level=", setup_relax_domain_level); 6012 6013static void set_domain_attribute(struct sched_domain *sd, 6014 struct sched_domain_attr *attr) 6015{ 6016 int request; 6017 6018 if (!attr || attr->relax_domain_level < 0) { 6019 if (default_relax_domain_level < 0) 6020 return; 6021 else 6022 request = default_relax_domain_level; 6023 } else 6024 request = attr->relax_domain_level; 6025 if (request < sd->level) { 6026 /* turn off idle balance on this domain */ 6027 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE); 6028 } else { 6029 /* turn on idle balance on this domain */ 6030 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE); 6031 } 6032} 6033 6034static void __sdt_free(const struct cpumask *cpu_map); 6035static int __sdt_alloc(const struct cpumask *cpu_map); 6036 6037static void __free_domain_allocs(struct s_data *d, enum s_alloc what, 6038 const struct cpumask *cpu_map) 6039{ 6040 switch (what) { 6041 case sa_rootdomain: 6042 if (!atomic_read(&d->rd->refcount)) 6043 free_rootdomain(&d->rd->rcu); /* fall through */ 6044 case sa_sd: 6045 free_percpu(d->sd); /* fall through */ 6046 case sa_sd_storage: 6047 __sdt_free(cpu_map); /* fall through */ 6048 case sa_none: 6049 break; 6050 } 6051} 6052 6053static enum s_alloc __visit_domain_allocation_hell(struct s_data *d, 6054 const struct cpumask *cpu_map) 6055{ 6056 memset(d, 0, sizeof(*d)); 6057 6058 if (__sdt_alloc(cpu_map)) 6059 return sa_sd_storage; 6060 d->sd = alloc_percpu(struct sched_domain *); 6061 if (!d->sd) 6062 return sa_sd_storage; 6063 d->rd = alloc_rootdomain(); 6064 if (!d->rd) 6065 return sa_sd; 6066 return sa_rootdomain; 6067} 6068 6069/* 6070 * NULL the sd_data elements we've used to build the sched_domain and 6071 * sched_group structure so that the subsequent __free_domain_allocs() 6072 * will not free the data we're using. 6073 */ 6074static void claim_allocations(int cpu, struct sched_domain *sd) 6075{ 6076 struct sd_data *sdd = sd->private; 6077 6078 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd); 6079 *per_cpu_ptr(sdd->sd, cpu) = NULL; 6080 6081 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref)) 6082 *per_cpu_ptr(sdd->sg, cpu) = NULL; 6083 6084 if (atomic_read(&(*per_cpu_ptr(sdd->sgc, cpu))->ref)) 6085 *per_cpu_ptr(sdd->sgc, cpu) = NULL; 6086} 6087 6088#ifdef CONFIG_NUMA 6089static int sched_domains_numa_levels; 6090static int *sched_domains_numa_distance; 6091static struct cpumask ***sched_domains_numa_masks; 6092static int sched_domains_curr_level; 6093#endif 6094 6095/* 6096 * SD_flags allowed in topology descriptions. 6097 * 6098 * SD_SHARE_CPUCAPACITY - describes SMT topologies 6099 * SD_SHARE_PKG_RESOURCES - describes shared caches 6100 * SD_NUMA - describes NUMA topologies 6101 * SD_SHARE_POWERDOMAIN - describes shared power domain 6102 * 6103 * Odd one out: 6104 * SD_ASYM_PACKING - describes SMT quirks 6105 */ 6106#define TOPOLOGY_SD_FLAGS \ 6107 (SD_SHARE_CPUCAPACITY | \ 6108 SD_SHARE_PKG_RESOURCES | \ 6109 SD_NUMA | \ 6110 SD_ASYM_PACKING | \ 6111 SD_SHARE_POWERDOMAIN) 6112 6113static struct sched_domain * 6114sd_init(struct sched_domain_topology_level *tl, int cpu) 6115{ 6116 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); 6117 int sd_weight, sd_flags = 0; 6118 6119#ifdef CONFIG_NUMA 6120 /* 6121 * Ugly hack to pass state to sd_numa_mask()... 6122 */ 6123 sched_domains_curr_level = tl->numa_level; 6124#endif 6125 6126 sd_weight = cpumask_weight(tl->mask(cpu)); 6127 6128 if (tl->sd_flags) 6129 sd_flags = (*tl->sd_flags)(); 6130 if (WARN_ONCE(sd_flags & ~TOPOLOGY_SD_FLAGS, 6131 "wrong sd_flags in topology description\n")) 6132 sd_flags &= ~TOPOLOGY_SD_FLAGS; 6133 6134 *sd = (struct sched_domain){ 6135 .min_interval = sd_weight, 6136 .max_interval = 2*sd_weight, 6137 .busy_factor = 32, 6138 .imbalance_pct = 125, 6139 6140 .cache_nice_tries = 0, 6141 .busy_idx = 0, 6142 .idle_idx = 0, 6143 .newidle_idx = 0, 6144 .wake_idx = 0, 6145 .forkexec_idx = 0, 6146 6147 .flags = 1*SD_LOAD_BALANCE 6148 | 1*SD_BALANCE_NEWIDLE 6149 | 1*SD_BALANCE_EXEC 6150 | 1*SD_BALANCE_FORK 6151 | 0*SD_BALANCE_WAKE 6152 | 1*SD_WAKE_AFFINE 6153 | 0*SD_SHARE_CPUCAPACITY 6154 | 0*SD_SHARE_PKG_RESOURCES 6155 | 0*SD_SERIALIZE 6156 | 0*SD_PREFER_SIBLING 6157 | 0*SD_NUMA 6158 | sd_flags 6159 , 6160 6161 .last_balance = jiffies, 6162 .balance_interval = sd_weight, 6163 .smt_gain = 0, 6164 .max_newidle_lb_cost = 0, 6165 .next_decay_max_lb_cost = jiffies, 6166#ifdef CONFIG_SCHED_DEBUG 6167 .name = tl->name, 6168#endif 6169 }; 6170 6171 /* 6172 * Convert topological properties into behaviour. 6173 */ 6174 6175 if (sd->flags & SD_SHARE_CPUCAPACITY) { 6176 sd->imbalance_pct = 110; 6177 sd->smt_gain = 1178; /* ~15% */ 6178 6179 } else if (sd->flags & SD_SHARE_PKG_RESOURCES) { 6180 sd->imbalance_pct = 117; 6181 sd->cache_nice_tries = 1; 6182 sd->busy_idx = 2; 6183 6184#ifdef CONFIG_NUMA 6185 } else if (sd->flags & SD_NUMA) { 6186 sd->cache_nice_tries = 2; 6187 sd->busy_idx = 3; 6188 sd->idle_idx = 2; 6189 6190 sd->flags |= SD_SERIALIZE; 6191 if (sched_domains_numa_distance[tl->numa_level] > RECLAIM_DISTANCE) { 6192 sd->flags &= ~(SD_BALANCE_EXEC | 6193 SD_BALANCE_FORK | 6194 SD_WAKE_AFFINE); 6195 } 6196 6197#endif 6198 } else { 6199 sd->flags |= SD_PREFER_SIBLING; 6200 sd->cache_nice_tries = 1; 6201 sd->busy_idx = 2; 6202 sd->idle_idx = 1; 6203 } 6204 6205 sd->private = &tl->data; 6206 6207 return sd; 6208} 6209 6210/* 6211 * Topology list, bottom-up. 6212 */ 6213static struct sched_domain_topology_level default_topology[] = { 6214#ifdef CONFIG_SCHED_SMT 6215 { cpu_smt_mask, cpu_smt_flags, SD_INIT_NAME(SMT) }, 6216#endif 6217#ifdef CONFIG_SCHED_MC 6218 { cpu_coregroup_mask, cpu_core_flags, SD_INIT_NAME(MC) }, 6219#endif 6220 { cpu_cpu_mask, SD_INIT_NAME(DIE) }, 6221 { NULL, }, 6222}; 6223 6224struct sched_domain_topology_level *sched_domain_topology = default_topology; 6225 6226#define for_each_sd_topology(tl) \ 6227 for (tl = sched_domain_topology; tl->mask; tl++) 6228 6229void set_sched_topology(struct sched_domain_topology_level *tl) 6230{ 6231 sched_domain_topology = tl; 6232} 6233 6234#ifdef CONFIG_NUMA 6235 6236static const struct cpumask *sd_numa_mask(int cpu) 6237{ 6238 return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)]; 6239} 6240 6241static void sched_numa_warn(const char *str) 6242{ 6243 static int done = false; 6244 int i,j; 6245 6246 if (done) 6247 return; 6248 6249 done = true; 6250 6251 printk(KERN_WARNING "ERROR: %s\n\n", str); 6252 6253 for (i = 0; i < nr_node_ids; i++) { 6254 printk(KERN_WARNING " "); 6255 for (j = 0; j < nr_node_ids; j++) 6256 printk(KERN_CONT "%02d ", node_distance(i,j)); 6257 printk(KERN_CONT "\n"); 6258 } 6259 printk(KERN_WARNING "\n"); 6260} 6261 6262static bool find_numa_distance(int distance) 6263{ 6264 int i; 6265 6266 if (distance == node_distance(0, 0)) 6267 return true; 6268 6269 for (i = 0; i < sched_domains_numa_levels; i++) { 6270 if (sched_domains_numa_distance[i] == distance) 6271 return true; 6272 } 6273 6274 return false; 6275} 6276 6277static void sched_init_numa(void) 6278{ 6279 int next_distance, curr_distance = node_distance(0, 0); 6280 struct sched_domain_topology_level *tl; 6281 int level = 0; 6282 int i, j, k; 6283 6284 sched_domains_numa_distance = kzalloc(sizeof(int) * nr_node_ids, GFP_KERNEL); 6285 if (!sched_domains_numa_distance) 6286 return; 6287 6288 /* 6289 * O(nr_nodes^2) deduplicating selection sort -- in order to find the 6290 * unique distances in the node_distance() table. 6291 * 6292 * Assumes node_distance(0,j) includes all distances in 6293 * node_distance(i,j) in order to avoid cubic time. 6294 */ 6295 next_distance = curr_distance; 6296 for (i = 0; i < nr_node_ids; i++) { 6297 for (j = 0; j < nr_node_ids; j++) { 6298 for (k = 0; k < nr_node_ids; k++) { 6299 int distance = node_distance(i, k); 6300 6301 if (distance > curr_distance && 6302 (distance < next_distance || 6303 next_distance == curr_distance)) 6304 next_distance = distance; 6305 6306 /* 6307 * While not a strong assumption it would be nice to know 6308 * about cases where if node A is connected to B, B is not 6309 * equally connected to A. 6310 */ 6311 if (sched_debug() && node_distance(k, i) != distance) 6312 sched_numa_warn("Node-distance not symmetric"); 6313 6314 if (sched_debug() && i && !find_numa_distance(distance)) 6315 sched_numa_warn("Node-0 not representative"); 6316 } 6317 if (next_distance != curr_distance) { 6318 sched_domains_numa_distance[level++] = next_distance; 6319 sched_domains_numa_levels = level; 6320 curr_distance = next_distance; 6321 } else break; 6322 } 6323 6324 /* 6325 * In case of sched_debug() we verify the above assumption. 6326 */ 6327 if (!sched_debug()) 6328 break; 6329 } 6330 /* 6331 * 'level' contains the number of unique distances, excluding the 6332 * identity distance node_distance(i,i). 6333 * 6334 * The sched_domains_numa_distance[] array includes the actual distance 6335 * numbers. 6336 */ 6337 6338 /* 6339 * Here, we should temporarily reset sched_domains_numa_levels to 0. 6340 * If it fails to allocate memory for array sched_domains_numa_masks[][], 6341 * the array will contain less then 'level' members. This could be 6342 * dangerous when we use it to iterate array sched_domains_numa_masks[][] 6343 * in other functions. 6344 * 6345 * We reset it to 'level' at the end of this function. 6346 */ 6347 sched_domains_numa_levels = 0; 6348 6349 sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL); 6350 if (!sched_domains_numa_masks) 6351 return; 6352 6353 /* 6354 * Now for each level, construct a mask per node which contains all 6355 * cpus of nodes that are that many hops away from us. 6356 */ 6357 for (i = 0; i < level; i++) { 6358 sched_domains_numa_masks[i] = 6359 kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL); 6360 if (!sched_domains_numa_masks[i]) 6361 return; 6362 6363 for (j = 0; j < nr_node_ids; j++) { 6364 struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL); 6365 if (!mask) 6366 return; 6367 6368 sched_domains_numa_masks[i][j] = mask; 6369 6370 for (k = 0; k < nr_node_ids; k++) { 6371 if (node_distance(j, k) > sched_domains_numa_distance[i]) 6372 continue; 6373 6374 cpumask_or(mask, mask, cpumask_of_node(k)); 6375 } 6376 } 6377 } 6378 6379 /* Compute default topology size */ 6380 for (i = 0; sched_domain_topology[i].mask; i++); 6381 6382 tl = kzalloc((i + level + 1) * 6383 sizeof(struct sched_domain_topology_level), GFP_KERNEL); 6384 if (!tl) 6385 return; 6386 6387 /* 6388 * Copy the default topology bits.. 6389 */ 6390 for (i = 0; sched_domain_topology[i].mask; i++) 6391 tl[i] = sched_domain_topology[i]; 6392 6393 /* 6394 * .. and append 'j' levels of NUMA goodness. 6395 */ 6396 for (j = 0; j < level; i++, j++) { 6397 tl[i] = (struct sched_domain_topology_level){ 6398 .mask = sd_numa_mask, 6399 .sd_flags = cpu_numa_flags, 6400 .flags = SDTL_OVERLAP, 6401 .numa_level = j, 6402 SD_INIT_NAME(NUMA) 6403 }; 6404 } 6405 6406 sched_domain_topology = tl; 6407 6408 sched_domains_numa_levels = level; 6409} 6410 6411static void sched_domains_numa_masks_set(int cpu) 6412{ 6413 int i, j; 6414 int node = cpu_to_node(cpu); 6415 6416 for (i = 0; i < sched_domains_numa_levels; i++) { 6417 for (j = 0; j < nr_node_ids; j++) { 6418 if (node_distance(j, node) <= sched_domains_numa_distance[i]) 6419 cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]); 6420 } 6421 } 6422} 6423 6424static void sched_domains_numa_masks_clear(int cpu) 6425{ 6426 int i, j; 6427 for (i = 0; i < sched_domains_numa_levels; i++) { 6428 for (j = 0; j < nr_node_ids; j++) 6429 cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]); 6430 } 6431} 6432 6433/* 6434 * Update sched_domains_numa_masks[level][node] array when new cpus 6435 * are onlined. 6436 */ 6437static int sched_domains_numa_masks_update(struct notifier_block *nfb, 6438 unsigned long action, 6439 void *hcpu) 6440{ 6441 int cpu = (long)hcpu; 6442 6443 switch (action & ~CPU_TASKS_FROZEN) { 6444 case CPU_ONLINE: 6445 sched_domains_numa_masks_set(cpu); 6446 break; 6447 6448 case CPU_DEAD: 6449 sched_domains_numa_masks_clear(cpu); 6450 break; 6451 6452 default: 6453 return NOTIFY_DONE; 6454 } 6455 6456 return NOTIFY_OK; 6457} 6458#else 6459static inline void sched_init_numa(void) 6460{ 6461} 6462 6463static int sched_domains_numa_masks_update(struct notifier_block *nfb, 6464 unsigned long action, 6465 void *hcpu) 6466{ 6467 return 0; 6468} 6469#endif /* CONFIG_NUMA */ 6470 6471static int __sdt_alloc(const struct cpumask *cpu_map) 6472{ 6473 struct sched_domain_topology_level *tl; 6474 int j; 6475 6476 for_each_sd_topology(tl) { 6477 struct sd_data *sdd = &tl->data; 6478 6479 sdd->sd = alloc_percpu(struct sched_domain *); 6480 if (!sdd->sd) 6481 return -ENOMEM; 6482 6483 sdd->sg = alloc_percpu(struct sched_group *); 6484 if (!sdd->sg) 6485 return -ENOMEM; 6486 6487 sdd->sgc = alloc_percpu(struct sched_group_capacity *); 6488 if (!sdd->sgc) 6489 return -ENOMEM; 6490 6491 for_each_cpu(j, cpu_map) { 6492 struct sched_domain *sd; 6493 struct sched_group *sg; 6494 struct sched_group_capacity *sgc; 6495 6496 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(), 6497 GFP_KERNEL, cpu_to_node(j)); 6498 if (!sd) 6499 return -ENOMEM; 6500 6501 *per_cpu_ptr(sdd->sd, j) = sd; 6502 6503 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(), 6504 GFP_KERNEL, cpu_to_node(j)); 6505 if (!sg) 6506 return -ENOMEM; 6507 6508 sg->next = sg; 6509 6510 *per_cpu_ptr(sdd->sg, j) = sg; 6511 6512 sgc = kzalloc_node(sizeof(struct sched_group_capacity) + cpumask_size(), 6513 GFP_KERNEL, cpu_to_node(j)); 6514 if (!sgc) 6515 return -ENOMEM; 6516 6517 *per_cpu_ptr(sdd->sgc, j) = sgc; 6518 } 6519 } 6520 6521 return 0; 6522} 6523 6524static void __sdt_free(const struct cpumask *cpu_map) 6525{ 6526 struct sched_domain_topology_level *tl; 6527 int j; 6528 6529 for_each_sd_topology(tl) { 6530 struct sd_data *sdd = &tl->data; 6531 6532 for_each_cpu(j, cpu_map) { 6533 struct sched_domain *sd; 6534 6535 if (sdd->sd) { 6536 sd = *per_cpu_ptr(sdd->sd, j); 6537 if (sd && (sd->flags & SD_OVERLAP)) 6538 free_sched_groups(sd->groups, 0); 6539 kfree(*per_cpu_ptr(sdd->sd, j)); 6540 } 6541 6542 if (sdd->sg) 6543 kfree(*per_cpu_ptr(sdd->sg, j)); 6544 if (sdd->sgc) 6545 kfree(*per_cpu_ptr(sdd->sgc, j)); 6546 } 6547 free_percpu(sdd->sd); 6548 sdd->sd = NULL; 6549 free_percpu(sdd->sg); 6550 sdd->sg = NULL; 6551 free_percpu(sdd->sgc); 6552 sdd->sgc = NULL; 6553 } 6554} 6555 6556struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl, 6557 const struct cpumask *cpu_map, struct sched_domain_attr *attr, 6558 struct sched_domain *child, int cpu) 6559{ 6560 struct sched_domain *sd = sd_init(tl, cpu); 6561 if (!sd) 6562 return child; 6563 6564 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu)); 6565 if (child) { 6566 sd->level = child->level + 1; 6567 sched_domain_level_max = max(sched_domain_level_max, sd->level); 6568 child->parent = sd; 6569 sd->child = child; 6570 6571 if (!cpumask_subset(sched_domain_span(child), 6572 sched_domain_span(sd))) { 6573 pr_err("BUG: arch topology borken\n"); 6574#ifdef CONFIG_SCHED_DEBUG 6575 pr_err(" the %s domain not a subset of the %s domain\n", 6576 child->name, sd->name); 6577#endif 6578 /* Fixup, ensure @sd has at least @child cpus. */ 6579 cpumask_or(sched_domain_span(sd), 6580 sched_domain_span(sd), 6581 sched_domain_span(child)); 6582 } 6583 6584 } 6585 set_domain_attribute(sd, attr); 6586 6587 return sd; 6588} 6589 6590/* 6591 * Build sched domains for a given set of cpus and attach the sched domains 6592 * to the individual cpus 6593 */ 6594static int build_sched_domains(const struct cpumask *cpu_map, 6595 struct sched_domain_attr *attr) 6596{ 6597 enum s_alloc alloc_state; 6598 struct sched_domain *sd; 6599 struct s_data d; 6600 int i, ret = -ENOMEM; 6601 6602 alloc_state = __visit_domain_allocation_hell(&d, cpu_map); 6603 if (alloc_state != sa_rootdomain) 6604 goto error; 6605 6606 /* Set up domains for cpus specified by the cpu_map. */ 6607 for_each_cpu(i, cpu_map) { 6608 struct sched_domain_topology_level *tl; 6609 6610 sd = NULL; 6611 for_each_sd_topology(tl) { 6612 sd = build_sched_domain(tl, cpu_map, attr, sd, i); 6613 if (tl == sched_domain_topology) 6614 *per_cpu_ptr(d.sd, i) = sd; 6615 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP)) 6616 sd->flags |= SD_OVERLAP; 6617 if (cpumask_equal(cpu_map, sched_domain_span(sd))) 6618 break; 6619 } 6620 } 6621 6622 /* Build the groups for the domains */ 6623 for_each_cpu(i, cpu_map) { 6624 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) { 6625 sd->span_weight = cpumask_weight(sched_domain_span(sd)); 6626 if (sd->flags & SD_OVERLAP) { 6627 if (build_overlap_sched_groups(sd, i)) 6628 goto error; 6629 } else { 6630 if (build_sched_groups(sd, i)) 6631 goto error; 6632 } 6633 } 6634 } 6635 6636 /* Calculate CPU capacity for physical packages and nodes */ 6637 for (i = nr_cpumask_bits-1; i >= 0; i--) { 6638 if (!cpumask_test_cpu(i, cpu_map)) 6639 continue; 6640 6641 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) { 6642 claim_allocations(i, sd); 6643 init_sched_groups_capacity(i, sd); 6644 } 6645 } 6646 6647 /* Attach the domains */ 6648 rcu_read_lock(); 6649 for_each_cpu(i, cpu_map) { 6650 sd = *per_cpu_ptr(d.sd, i); 6651 cpu_attach_domain(sd, d.rd, i); 6652 } 6653 rcu_read_unlock(); 6654 6655 ret = 0; 6656error: 6657 __free_domain_allocs(&d, alloc_state, cpu_map); 6658 return ret; 6659} 6660 6661static cpumask_var_t *doms_cur; /* current sched domains */ 6662static int ndoms_cur; /* number of sched domains in 'doms_cur' */ 6663static struct sched_domain_attr *dattr_cur; 6664 /* attribues of custom domains in 'doms_cur' */ 6665 6666/* 6667 * Special case: If a kmalloc of a doms_cur partition (array of 6668 * cpumask) fails, then fallback to a single sched domain, 6669 * as determined by the single cpumask fallback_doms. 6670 */ 6671static cpumask_var_t fallback_doms; 6672 6673/* 6674 * arch_update_cpu_topology lets virtualized architectures update the 6675 * cpu core maps. It is supposed to return 1 if the topology changed 6676 * or 0 if it stayed the same. 6677 */ 6678int __weak arch_update_cpu_topology(void) 6679{ 6680 return 0; 6681} 6682 6683cpumask_var_t *alloc_sched_domains(unsigned int ndoms) 6684{ 6685 int i; 6686 cpumask_var_t *doms; 6687 6688 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL); 6689 if (!doms) 6690 return NULL; 6691 for (i = 0; i < ndoms; i++) { 6692 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) { 6693 free_sched_domains(doms, i); 6694 return NULL; 6695 } 6696 } 6697 return doms; 6698} 6699 6700void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms) 6701{ 6702 unsigned int i; 6703 for (i = 0; i < ndoms; i++) 6704 free_cpumask_var(doms[i]); 6705 kfree(doms); 6706} 6707 6708/* 6709 * Set up scheduler domains and groups. Callers must hold the hotplug lock. 6710 * For now this just excludes isolated cpus, but could be used to 6711 * exclude other special cases in the future. 6712 */ 6713static int init_sched_domains(const struct cpumask *cpu_map) 6714{ 6715 int err; 6716 6717 arch_update_cpu_topology(); 6718 ndoms_cur = 1; 6719 doms_cur = alloc_sched_domains(ndoms_cur); 6720 if (!doms_cur) 6721 doms_cur = &fallback_doms; 6722 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map); 6723 err = build_sched_domains(doms_cur[0], NULL); 6724 register_sched_domain_sysctl(); 6725 6726 return err; 6727} 6728 6729/* 6730 * Detach sched domains from a group of cpus specified in cpu_map 6731 * These cpus will now be attached to the NULL domain 6732 */ 6733static void detach_destroy_domains(const struct cpumask *cpu_map) 6734{ 6735 int i; 6736 6737 rcu_read_lock(); 6738 for_each_cpu(i, cpu_map) 6739 cpu_attach_domain(NULL, &def_root_domain, i); 6740 rcu_read_unlock(); 6741} 6742 6743/* handle null as "default" */ 6744static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur, 6745 struct sched_domain_attr *new, int idx_new) 6746{ 6747 struct sched_domain_attr tmp; 6748 6749 /* fast path */ 6750 if (!new && !cur) 6751 return 1; 6752 6753 tmp = SD_ATTR_INIT; 6754 return !memcmp(cur ? (cur + idx_cur) : &tmp, 6755 new ? (new + idx_new) : &tmp, 6756 sizeof(struct sched_domain_attr)); 6757} 6758 6759/* 6760 * Partition sched domains as specified by the 'ndoms_new' 6761 * cpumasks in the array doms_new[] of cpumasks. This compares 6762 * doms_new[] to the current sched domain partitioning, doms_cur[]. 6763 * It destroys each deleted domain and builds each new domain. 6764 * 6765 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'. 6766 * The masks don't intersect (don't overlap.) We should setup one 6767 * sched domain for each mask. CPUs not in any of the cpumasks will 6768 * not be load balanced. If the same cpumask appears both in the 6769 * current 'doms_cur' domains and in the new 'doms_new', we can leave 6770 * it as it is. 6771 * 6772 * The passed in 'doms_new' should be allocated using 6773 * alloc_sched_domains. This routine takes ownership of it and will 6774 * free_sched_domains it when done with it. If the caller failed the 6775 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1, 6776 * and partition_sched_domains() will fallback to the single partition 6777 * 'fallback_doms', it also forces the domains to be rebuilt. 6778 * 6779 * If doms_new == NULL it will be replaced with cpu_online_mask. 6780 * ndoms_new == 0 is a special case for destroying existing domains, 6781 * and it will not create the default domain. 6782 * 6783 * Call with hotplug lock held 6784 */ 6785void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[], 6786 struct sched_domain_attr *dattr_new) 6787{ 6788 int i, j, n; 6789 int new_topology; 6790 6791 mutex_lock(&sched_domains_mutex); 6792 6793 /* always unregister in case we don't destroy any domains */ 6794 unregister_sched_domain_sysctl(); 6795 6796 /* Let architecture update cpu core mappings. */ 6797 new_topology = arch_update_cpu_topology(); 6798 6799 n = doms_new ? ndoms_new : 0; 6800 6801 /* Destroy deleted domains */ 6802 for (i = 0; i < ndoms_cur; i++) { 6803 for (j = 0; j < n && !new_topology; j++) { 6804 if (cpumask_equal(doms_cur[i], doms_new[j]) 6805 && dattrs_equal(dattr_cur, i, dattr_new, j)) 6806 goto match1; 6807 } 6808 /* no match - a current sched domain not in new doms_new[] */ 6809 detach_destroy_domains(doms_cur[i]); 6810match1: 6811 ; 6812 } 6813 6814 n = ndoms_cur; 6815 if (doms_new == NULL) { 6816 n = 0; 6817 doms_new = &fallback_doms; 6818 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map); 6819 WARN_ON_ONCE(dattr_new); 6820 } 6821 6822 /* Build new domains */ 6823 for (i = 0; i < ndoms_new; i++) { 6824 for (j = 0; j < n && !new_topology; j++) { 6825 if (cpumask_equal(doms_new[i], doms_cur[j]) 6826 && dattrs_equal(dattr_new, i, dattr_cur, j)) 6827 goto match2; 6828 } 6829 /* no match - add a new doms_new */ 6830 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL); 6831match2: 6832 ; 6833 } 6834 6835 /* Remember the new sched domains */ 6836 if (doms_cur != &fallback_doms) 6837 free_sched_domains(doms_cur, ndoms_cur); 6838 kfree(dattr_cur); /* kfree(NULL) is safe */ 6839 doms_cur = doms_new; 6840 dattr_cur = dattr_new; 6841 ndoms_cur = ndoms_new; 6842 6843 register_sched_domain_sysctl(); 6844 6845 mutex_unlock(&sched_domains_mutex); 6846} 6847 6848static int num_cpus_frozen; /* used to mark begin/end of suspend/resume */ 6849 6850/* 6851 * Update cpusets according to cpu_active mask. If cpusets are 6852 * disabled, cpuset_update_active_cpus() becomes a simple wrapper 6853 * around partition_sched_domains(). 6854 * 6855 * If we come here as part of a suspend/resume, don't touch cpusets because we 6856 * want to restore it back to its original state upon resume anyway. 6857 */ 6858static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action, 6859 void *hcpu) 6860{ 6861 switch (action) { 6862 case CPU_ONLINE_FROZEN: 6863 case CPU_DOWN_FAILED_FROZEN: 6864 6865 /* 6866 * num_cpus_frozen tracks how many CPUs are involved in suspend 6867 * resume sequence. As long as this is not the last online 6868 * operation in the resume sequence, just build a single sched 6869 * domain, ignoring cpusets. 6870 */ 6871 num_cpus_frozen--; 6872 if (likely(num_cpus_frozen)) { 6873 partition_sched_domains(1, NULL, NULL); 6874 break; 6875 } 6876 6877 /* 6878 * This is the last CPU online operation. So fall through and 6879 * restore the original sched domains by considering the 6880 * cpuset configurations. 6881 */ 6882 6883 case CPU_ONLINE: 6884 case CPU_DOWN_FAILED: 6885 cpuset_update_active_cpus(true); 6886 break; 6887 default: 6888 return NOTIFY_DONE; 6889 } 6890 return NOTIFY_OK; 6891} 6892 6893static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action, 6894 void *hcpu) 6895{ 6896 switch (action) { 6897 case CPU_DOWN_PREPARE: 6898 cpuset_update_active_cpus(false); 6899 break; 6900 case CPU_DOWN_PREPARE_FROZEN: 6901 num_cpus_frozen++; 6902 partition_sched_domains(1, NULL, NULL); 6903 break; 6904 default: 6905 return NOTIFY_DONE; 6906 } 6907 return NOTIFY_OK; 6908} 6909 6910void __init sched_init_smp(void) 6911{ 6912 cpumask_var_t non_isolated_cpus; 6913 6914 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL); 6915 alloc_cpumask_var(&fallback_doms, GFP_KERNEL); 6916 6917 sched_init_numa(); 6918 6919 /* 6920 * There's no userspace yet to cause hotplug operations; hence all the 6921 * cpu masks are stable and all blatant races in the below code cannot 6922 * happen. 6923 */ 6924 mutex_lock(&sched_domains_mutex); 6925 init_sched_domains(cpu_active_mask); 6926 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map); 6927 if (cpumask_empty(non_isolated_cpus)) 6928 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus); 6929 mutex_unlock(&sched_domains_mutex); 6930 6931 hotcpu_notifier(sched_domains_numa_masks_update, CPU_PRI_SCHED_ACTIVE); 6932 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE); 6933 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE); 6934 6935 init_hrtick(); 6936 6937 /* Move init over to a non-isolated CPU */ 6938 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0) 6939 BUG(); 6940 sched_init_granularity(); 6941 free_cpumask_var(non_isolated_cpus); 6942 6943 init_sched_rt_class(); 6944 init_sched_dl_class(); 6945} 6946#else 6947void __init sched_init_smp(void) 6948{ 6949 sched_init_granularity(); 6950} 6951#endif /* CONFIG_SMP */ 6952 6953const_debug unsigned int sysctl_timer_migration = 1; 6954 6955int in_sched_functions(unsigned long addr) 6956{ 6957 return in_lock_functions(addr) || 6958 (addr >= (unsigned long)__sched_text_start 6959 && addr < (unsigned long)__sched_text_end); 6960} 6961 6962#ifdef CONFIG_CGROUP_SCHED 6963/* 6964 * Default task group. 6965 * Every task in system belongs to this group at bootup. 6966 */ 6967struct task_group root_task_group; 6968LIST_HEAD(task_groups); 6969#endif 6970 6971DECLARE_PER_CPU(cpumask_var_t, load_balance_mask); 6972 6973void __init sched_init(void) 6974{ 6975 int i, j; 6976 unsigned long alloc_size = 0, ptr; 6977 6978#ifdef CONFIG_FAIR_GROUP_SCHED 6979 alloc_size += 2 * nr_cpu_ids * sizeof(void **); 6980#endif 6981#ifdef CONFIG_RT_GROUP_SCHED 6982 alloc_size += 2 * nr_cpu_ids * sizeof(void **); 6983#endif 6984#ifdef CONFIG_CPUMASK_OFFSTACK 6985 alloc_size += num_possible_cpus() * cpumask_size(); 6986#endif 6987 if (alloc_size) { 6988 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT); 6989 6990#ifdef CONFIG_FAIR_GROUP_SCHED 6991 root_task_group.se = (struct sched_entity **)ptr; 6992 ptr += nr_cpu_ids * sizeof(void **); 6993 6994 root_task_group.cfs_rq = (struct cfs_rq **)ptr; 6995 ptr += nr_cpu_ids * sizeof(void **); 6996 6997#endif /* CONFIG_FAIR_GROUP_SCHED */ 6998#ifdef CONFIG_RT_GROUP_SCHED 6999 root_task_group.rt_se = (struct sched_rt_entity **)ptr; 7000 ptr += nr_cpu_ids * sizeof(void **); 7001 7002 root_task_group.rt_rq = (struct rt_rq **)ptr; 7003 ptr += nr_cpu_ids * sizeof(void **); 7004 7005#endif /* CONFIG_RT_GROUP_SCHED */ 7006#ifdef CONFIG_CPUMASK_OFFSTACK 7007 for_each_possible_cpu(i) { 7008 per_cpu(load_balance_mask, i) = (void *)ptr; 7009 ptr += cpumask_size(); 7010 } 7011#endif /* CONFIG_CPUMASK_OFFSTACK */ 7012 } 7013 7014 init_rt_bandwidth(&def_rt_bandwidth, 7015 global_rt_period(), global_rt_runtime()); 7016 init_dl_bandwidth(&def_dl_bandwidth, 7017 global_rt_period(), global_rt_runtime()); 7018 7019#ifdef CONFIG_SMP 7020 init_defrootdomain(); 7021#endif 7022 7023#ifdef CONFIG_RT_GROUP_SCHED 7024 init_rt_bandwidth(&root_task_group.rt_bandwidth, 7025 global_rt_period(), global_rt_runtime()); 7026#endif /* CONFIG_RT_GROUP_SCHED */ 7027 7028#ifdef CONFIG_CGROUP_SCHED 7029 list_add(&root_task_group.list, &task_groups); 7030 INIT_LIST_HEAD(&root_task_group.children); 7031 INIT_LIST_HEAD(&root_task_group.siblings); 7032 autogroup_init(&init_task); 7033 7034#endif /* CONFIG_CGROUP_SCHED */ 7035 7036 for_each_possible_cpu(i) { 7037 struct rq *rq; 7038 7039 rq = cpu_rq(i); 7040 raw_spin_lock_init(&rq->lock); 7041 rq->nr_running = 0; 7042 rq->calc_load_active = 0; 7043 rq->calc_load_update = jiffies + LOAD_FREQ; 7044 init_cfs_rq(&rq->cfs); 7045 init_rt_rq(&rq->rt, rq); 7046 init_dl_rq(&rq->dl, rq); 7047#ifdef CONFIG_FAIR_GROUP_SCHED 7048 root_task_group.shares = ROOT_TASK_GROUP_LOAD; 7049 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list); 7050 /* 7051 * How much cpu bandwidth does root_task_group get? 7052 * 7053 * In case of task-groups formed thr' the cgroup filesystem, it 7054 * gets 100% of the cpu resources in the system. This overall 7055 * system cpu resource is divided among the tasks of 7056 * root_task_group and its child task-groups in a fair manner, 7057 * based on each entity's (task or task-group's) weight 7058 * (se->load.weight). 7059 * 7060 * In other words, if root_task_group has 10 tasks of weight 7061 * 1024) and two child groups A0 and A1 (of weight 1024 each), 7062 * then A0's share of the cpu resource is: 7063 * 7064 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33% 7065 * 7066 * We achieve this by letting root_task_group's tasks sit 7067 * directly in rq->cfs (i.e root_task_group->se[] = NULL). 7068 */ 7069 init_cfs_bandwidth(&root_task_group.cfs_bandwidth); 7070 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL); 7071#endif /* CONFIG_FAIR_GROUP_SCHED */ 7072 7073 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime; 7074#ifdef CONFIG_RT_GROUP_SCHED 7075 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL); 7076#endif 7077 7078 for (j = 0; j < CPU_LOAD_IDX_MAX; j++) 7079 rq->cpu_load[j] = 0; 7080 7081 rq->last_load_update_tick = jiffies; 7082 7083#ifdef CONFIG_SMP 7084 rq->sd = NULL; 7085 rq->rd = NULL; 7086 rq->cpu_capacity = SCHED_CAPACITY_SCALE; 7087 rq->post_schedule = 0; 7088 rq->active_balance = 0; 7089 rq->next_balance = jiffies; 7090 rq->push_cpu = 0; 7091 rq->cpu = i; 7092 rq->online = 0; 7093 rq->idle_stamp = 0; 7094 rq->avg_idle = 2*sysctl_sched_migration_cost; 7095 rq->max_idle_balance_cost = sysctl_sched_migration_cost; 7096 7097 INIT_LIST_HEAD(&rq->cfs_tasks); 7098 7099 rq_attach_root(rq, &def_root_domain); 7100#ifdef CONFIG_NO_HZ_COMMON 7101 rq->nohz_flags = 0; 7102#endif 7103#ifdef CONFIG_NO_HZ_FULL 7104 rq->last_sched_tick = 0; 7105#endif 7106#endif 7107 init_rq_hrtick(rq); 7108 atomic_set(&rq->nr_iowait, 0); 7109 } 7110 7111 set_load_weight(&init_task); 7112 7113#ifdef CONFIG_PREEMPT_NOTIFIERS 7114 INIT_HLIST_HEAD(&init_task.preempt_notifiers); 7115#endif 7116 7117 /* 7118 * The boot idle thread does lazy MMU switching as well: 7119 */ 7120 atomic_inc(&init_mm.mm_count); 7121 enter_lazy_tlb(&init_mm, current); 7122 7123 /* 7124 * Make us the idle thread. Technically, schedule() should not be 7125 * called from this thread, however somewhere below it might be, 7126 * but because we are the idle thread, we just pick up running again 7127 * when this runqueue becomes "idle". 7128 */ 7129 init_idle(current, smp_processor_id()); 7130 7131 calc_load_update = jiffies + LOAD_FREQ; 7132 7133 /* 7134 * During early bootup we pretend to be a normal task: 7135 */ 7136 current->sched_class = &fair_sched_class; 7137 7138#ifdef CONFIG_SMP 7139 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT); 7140 /* May be allocated at isolcpus cmdline parse time */ 7141 if (cpu_isolated_map == NULL) 7142 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT); 7143 idle_thread_set_boot_cpu(); 7144 set_cpu_rq_start_time(); 7145#endif 7146 init_sched_fair_class(); 7147 7148 scheduler_running = 1; 7149} 7150 7151#ifdef CONFIG_DEBUG_ATOMIC_SLEEP 7152static inline int preempt_count_equals(int preempt_offset) 7153{ 7154 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth(); 7155 7156 return (nested == preempt_offset); 7157} 7158 7159void __might_sleep(const char *file, int line, int preempt_offset) 7160{ 7161 static unsigned long prev_jiffy; /* ratelimiting */ 7162 7163 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */ 7164 if ((preempt_count_equals(preempt_offset) && !irqs_disabled() && 7165 !is_idle_task(current)) || 7166 system_state != SYSTEM_RUNNING || oops_in_progress) 7167 return; 7168 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy) 7169 return; 7170 prev_jiffy = jiffies; 7171 7172 printk(KERN_ERR 7173 "BUG: sleeping function called from invalid context at %s:%d\n", 7174 file, line); 7175 printk(KERN_ERR 7176 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n", 7177 in_atomic(), irqs_disabled(), 7178 current->pid, current->comm); 7179 7180 debug_show_held_locks(current); 7181 if (irqs_disabled()) 7182 print_irqtrace_events(current); 7183#ifdef CONFIG_DEBUG_PREEMPT 7184 if (!preempt_count_equals(preempt_offset)) { 7185 pr_err("Preemption disabled at:"); 7186 print_ip_sym(current->preempt_disable_ip); 7187 pr_cont("\n"); 7188 } 7189#endif 7190 dump_stack(); 7191} 7192EXPORT_SYMBOL(__might_sleep); 7193#endif 7194 7195#ifdef CONFIG_MAGIC_SYSRQ 7196static void normalize_task(struct rq *rq, struct task_struct *p) 7197{ 7198 const struct sched_class *prev_class = p->sched_class; 7199 struct sched_attr attr = { 7200 .sched_policy = SCHED_NORMAL, 7201 }; 7202 int old_prio = p->prio; 7203 int queued; 7204 7205 queued = task_on_rq_queued(p); 7206 if (queued) 7207 dequeue_task(rq, p, 0); 7208 __setscheduler(rq, p, &attr); 7209 if (queued) { 7210 enqueue_task(rq, p, 0); 7211 resched_curr(rq); 7212 } 7213 7214 check_class_changed(rq, p, prev_class, old_prio); 7215} 7216 7217void normalize_rt_tasks(void) 7218{ 7219 struct task_struct *g, *p; 7220 unsigned long flags; 7221 struct rq *rq; 7222 7223 read_lock(&tasklist_lock); 7224 for_each_process_thread(g, p) { 7225 /* 7226 * Only normalize user tasks: 7227 */ 7228 if (p->flags & PF_KTHREAD) 7229 continue; 7230 7231 p->se.exec_start = 0; 7232#ifdef CONFIG_SCHEDSTATS 7233 p->se.statistics.wait_start = 0; 7234 p->se.statistics.sleep_start = 0; 7235 p->se.statistics.block_start = 0; 7236#endif 7237 7238 if (!dl_task(p) && !rt_task(p)) { 7239 /* 7240 * Renice negative nice level userspace 7241 * tasks back to 0: 7242 */ 7243 if (task_nice(p) < 0) 7244 set_user_nice(p, 0); 7245 continue; 7246 } 7247 7248 rq = task_rq_lock(p, &flags); 7249 normalize_task(rq, p); 7250 task_rq_unlock(rq, p, &flags); 7251 } 7252 read_unlock(&tasklist_lock); 7253} 7254 7255#endif /* CONFIG_MAGIC_SYSRQ */ 7256 7257#if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) 7258/* 7259 * These functions are only useful for the IA64 MCA handling, or kdb. 7260 * 7261 * They can only be called when the whole system has been 7262 * stopped - every CPU needs to be quiescent, and no scheduling 7263 * activity can take place. Using them for anything else would 7264 * be a serious bug, and as a result, they aren't even visible 7265 * under any other configuration. 7266 */ 7267 7268/** 7269 * curr_task - return the current task for a given cpu. 7270 * @cpu: the processor in question. 7271 * 7272 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED! 7273 * 7274 * Return: The current task for @cpu. 7275 */ 7276struct task_struct *curr_task(int cpu) 7277{ 7278 return cpu_curr(cpu); 7279} 7280 7281#endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */ 7282 7283#ifdef CONFIG_IA64 7284/** 7285 * set_curr_task - set the current task for a given cpu. 7286 * @cpu: the processor in question. 7287 * @p: the task pointer to set. 7288 * 7289 * Description: This function must only be used when non-maskable interrupts 7290 * are serviced on a separate stack. It allows the architecture to switch the 7291 * notion of the current task on a cpu in a non-blocking manner. This function 7292 * must be called with all CPU's synchronized, and interrupts disabled, the 7293 * and caller must save the original value of the current task (see 7294 * curr_task() above) and restore that value before reenabling interrupts and 7295 * re-starting the system. 7296 * 7297 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED! 7298 */ 7299void set_curr_task(int cpu, struct task_struct *p) 7300{ 7301 cpu_curr(cpu) = p; 7302} 7303 7304#endif 7305 7306#ifdef CONFIG_CGROUP_SCHED 7307/* task_group_lock serializes the addition/removal of task groups */ 7308static DEFINE_SPINLOCK(task_group_lock); 7309 7310static void free_sched_group(struct task_group *tg) 7311{ 7312 free_fair_sched_group(tg); 7313 free_rt_sched_group(tg); 7314 autogroup_free(tg); 7315 kfree(tg); 7316} 7317 7318/* allocate runqueue etc for a new task group */ 7319struct task_group *sched_create_group(struct task_group *parent) 7320{ 7321 struct task_group *tg; 7322 7323 tg = kzalloc(sizeof(*tg), GFP_KERNEL); 7324 if (!tg) 7325 return ERR_PTR(-ENOMEM); 7326 7327 if (!alloc_fair_sched_group(tg, parent)) 7328 goto err; 7329 7330 if (!alloc_rt_sched_group(tg, parent)) 7331 goto err; 7332 7333 return tg; 7334 7335err: 7336 free_sched_group(tg); 7337 return ERR_PTR(-ENOMEM); 7338} 7339 7340void sched_online_group(struct task_group *tg, struct task_group *parent) 7341{ 7342 unsigned long flags; 7343 7344 spin_lock_irqsave(&task_group_lock, flags); 7345 list_add_rcu(&tg->list, &task_groups); 7346 7347 WARN_ON(!parent); /* root should already exist */ 7348 7349 tg->parent = parent; 7350 INIT_LIST_HEAD(&tg->children); 7351 list_add_rcu(&tg->siblings, &parent->children); 7352 spin_unlock_irqrestore(&task_group_lock, flags); 7353} 7354 7355/* rcu callback to free various structures associated with a task group */ 7356static void free_sched_group_rcu(struct rcu_head *rhp) 7357{ 7358 /* now it should be safe to free those cfs_rqs */ 7359 free_sched_group(container_of(rhp, struct task_group, rcu)); 7360} 7361 7362/* Destroy runqueue etc associated with a task group */ 7363void sched_destroy_group(struct task_group *tg) 7364{ 7365 /* wait for possible concurrent references to cfs_rqs complete */ 7366 call_rcu(&tg->rcu, free_sched_group_rcu); 7367} 7368 7369void sched_offline_group(struct task_group *tg) 7370{ 7371 unsigned long flags; 7372 int i; 7373 7374 /* end participation in shares distribution */ 7375 for_each_possible_cpu(i) 7376 unregister_fair_sched_group(tg, i); 7377 7378 spin_lock_irqsave(&task_group_lock, flags); 7379 list_del_rcu(&tg->list); 7380 list_del_rcu(&tg->siblings); 7381 spin_unlock_irqrestore(&task_group_lock, flags); 7382} 7383 7384/* change task's runqueue when it moves between groups. 7385 * The caller of this function should have put the task in its new group 7386 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to 7387 * reflect its new group. 7388 */ 7389void sched_move_task(struct task_struct *tsk) 7390{ 7391 struct task_group *tg; 7392 int queued, running; 7393 unsigned long flags; 7394 struct rq *rq; 7395 7396 rq = task_rq_lock(tsk, &flags); 7397 7398 running = task_current(rq, tsk); 7399 queued = task_on_rq_queued(tsk); 7400 7401 if (queued) 7402 dequeue_task(rq, tsk, 0); 7403 if (unlikely(running)) 7404 put_prev_task(rq, tsk); 7405 7406 tg = container_of(task_css_check(tsk, cpu_cgrp_id, 7407 lockdep_is_held(&tsk->sighand->siglock)), 7408 struct task_group, css); 7409 tg = autogroup_task_group(tsk, tg); 7410 tsk->sched_task_group = tg; 7411 7412#ifdef CONFIG_FAIR_GROUP_SCHED 7413 if (tsk->sched_class->task_move_group) 7414 tsk->sched_class->task_move_group(tsk, queued); 7415 else 7416#endif 7417 set_task_rq(tsk, task_cpu(tsk)); 7418 7419 if (unlikely(running)) 7420 tsk->sched_class->set_curr_task(rq); 7421 if (queued) 7422 enqueue_task(rq, tsk, 0); 7423 7424 task_rq_unlock(rq, tsk, &flags); 7425} 7426#endif /* CONFIG_CGROUP_SCHED */ 7427 7428#ifdef CONFIG_RT_GROUP_SCHED 7429/* 7430 * Ensure that the real time constraints are schedulable. 7431 */ 7432static DEFINE_MUTEX(rt_constraints_mutex); 7433 7434/* Must be called with tasklist_lock held */ 7435static inline int tg_has_rt_tasks(struct task_group *tg) 7436{ 7437 struct task_struct *g, *p; 7438 7439 for_each_process_thread(g, p) { 7440 if (rt_task(p) && task_group(p) == tg) 7441 return 1; 7442 } 7443 7444 return 0; 7445} 7446 7447struct rt_schedulable_data { 7448 struct task_group *tg; 7449 u64 rt_period; 7450 u64 rt_runtime; 7451}; 7452 7453static int tg_rt_schedulable(struct task_group *tg, void *data) 7454{ 7455 struct rt_schedulable_data *d = data; 7456 struct task_group *child; 7457 unsigned long total, sum = 0; 7458 u64 period, runtime; 7459 7460 period = ktime_to_ns(tg->rt_bandwidth.rt_period); 7461 runtime = tg->rt_bandwidth.rt_runtime; 7462 7463 if (tg == d->tg) { 7464 period = d->rt_period; 7465 runtime = d->rt_runtime; 7466 } 7467 7468 /* 7469 * Cannot have more runtime than the period. 7470 */ 7471 if (runtime > period && runtime != RUNTIME_INF) 7472 return -EINVAL; 7473 7474 /* 7475 * Ensure we don't starve existing RT tasks. 7476 */ 7477 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg)) 7478 return -EBUSY; 7479 7480 total = to_ratio(period, runtime); 7481 7482 /* 7483 * Nobody can have more than the global setting allows. 7484 */ 7485 if (total > to_ratio(global_rt_period(), global_rt_runtime())) 7486 return -EINVAL; 7487 7488 /* 7489 * The sum of our children's runtime should not exceed our own. 7490 */ 7491 list_for_each_entry_rcu(child, &tg->children, siblings) { 7492 period = ktime_to_ns(child->rt_bandwidth.rt_period); 7493 runtime = child->rt_bandwidth.rt_runtime; 7494 7495 if (child == d->tg) { 7496 period = d->rt_period; 7497 runtime = d->rt_runtime; 7498 } 7499 7500 sum += to_ratio(period, runtime); 7501 } 7502 7503 if (sum > total) 7504 return -EINVAL; 7505 7506 return 0; 7507} 7508 7509static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime) 7510{ 7511 int ret; 7512 7513 struct rt_schedulable_data data = { 7514 .tg = tg, 7515 .rt_period = period, 7516 .rt_runtime = runtime, 7517 }; 7518 7519 rcu_read_lock(); 7520 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data); 7521 rcu_read_unlock(); 7522 7523 return ret; 7524} 7525 7526static int tg_set_rt_bandwidth(struct task_group *tg, 7527 u64 rt_period, u64 rt_runtime) 7528{ 7529 int i, err = 0; 7530 7531 mutex_lock(&rt_constraints_mutex); 7532 read_lock(&tasklist_lock); 7533 err = __rt_schedulable(tg, rt_period, rt_runtime); 7534 if (err) 7535 goto unlock; 7536 7537 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock); 7538 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period); 7539 tg->rt_bandwidth.rt_runtime = rt_runtime; 7540 7541 for_each_possible_cpu(i) { 7542 struct rt_rq *rt_rq = tg->rt_rq[i]; 7543 7544 raw_spin_lock(&rt_rq->rt_runtime_lock); 7545 rt_rq->rt_runtime = rt_runtime; 7546 raw_spin_unlock(&rt_rq->rt_runtime_lock); 7547 } 7548 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock); 7549unlock: 7550 read_unlock(&tasklist_lock); 7551 mutex_unlock(&rt_constraints_mutex); 7552 7553 return err; 7554} 7555 7556static int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us) 7557{ 7558 u64 rt_runtime, rt_period; 7559 7560 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period); 7561 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC; 7562 if (rt_runtime_us < 0) 7563 rt_runtime = RUNTIME_INF; 7564 7565 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime); 7566} 7567 7568static long sched_group_rt_runtime(struct task_group *tg) 7569{ 7570 u64 rt_runtime_us; 7571 7572 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF) 7573 return -1; 7574 7575 rt_runtime_us = tg->rt_bandwidth.rt_runtime; 7576 do_div(rt_runtime_us, NSEC_PER_USEC); 7577 return rt_runtime_us; 7578} 7579 7580static int sched_group_set_rt_period(struct task_group *tg, long rt_period_us) 7581{ 7582 u64 rt_runtime, rt_period; 7583 7584 rt_period = (u64)rt_period_us * NSEC_PER_USEC; 7585 rt_runtime = tg->rt_bandwidth.rt_runtime; 7586 7587 if (rt_period == 0) 7588 return -EINVAL; 7589 7590 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime); 7591} 7592 7593static long sched_group_rt_period(struct task_group *tg) 7594{ 7595 u64 rt_period_us; 7596 7597 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period); 7598 do_div(rt_period_us, NSEC_PER_USEC); 7599 return rt_period_us; 7600} 7601#endif /* CONFIG_RT_GROUP_SCHED */ 7602 7603#ifdef CONFIG_RT_GROUP_SCHED 7604static int sched_rt_global_constraints(void) 7605{ 7606 int ret = 0; 7607 7608 mutex_lock(&rt_constraints_mutex); 7609 read_lock(&tasklist_lock); 7610 ret = __rt_schedulable(NULL, 0, 0); 7611 read_unlock(&tasklist_lock); 7612 mutex_unlock(&rt_constraints_mutex); 7613 7614 return ret; 7615} 7616 7617static int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk) 7618{ 7619 /* Don't accept realtime tasks when there is no way for them to run */ 7620 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0) 7621 return 0; 7622 7623 return 1; 7624} 7625 7626#else /* !CONFIG_RT_GROUP_SCHED */ 7627static int sched_rt_global_constraints(void) 7628{ 7629 unsigned long flags; 7630 int i, ret = 0; 7631 7632 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags); 7633 for_each_possible_cpu(i) { 7634 struct rt_rq *rt_rq = &cpu_rq(i)->rt; 7635 7636 raw_spin_lock(&rt_rq->rt_runtime_lock); 7637 rt_rq->rt_runtime = global_rt_runtime(); 7638 raw_spin_unlock(&rt_rq->rt_runtime_lock); 7639 } 7640 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags); 7641 7642 return ret; 7643} 7644#endif /* CONFIG_RT_GROUP_SCHED */ 7645 7646static int sched_dl_global_constraints(void) 7647{ 7648 u64 runtime = global_rt_runtime(); 7649 u64 period = global_rt_period(); 7650 u64 new_bw = to_ratio(period, runtime); 7651 int cpu, ret = 0; 7652 unsigned long flags; 7653 7654 /* 7655 * Here we want to check the bandwidth not being set to some 7656 * value smaller than the currently allocated bandwidth in 7657 * any of the root_domains. 7658 * 7659 * FIXME: Cycling on all the CPUs is overdoing, but simpler than 7660 * cycling on root_domains... Discussion on different/better 7661 * solutions is welcome! 7662 */ 7663 for_each_possible_cpu(cpu) { 7664 struct dl_bw *dl_b = dl_bw_of(cpu); 7665 7666 raw_spin_lock_irqsave(&dl_b->lock, flags); 7667 if (new_bw < dl_b->total_bw) 7668 ret = -EBUSY; 7669 raw_spin_unlock_irqrestore(&dl_b->lock, flags); 7670 7671 if (ret) 7672 break; 7673 } 7674 7675 return ret; 7676} 7677 7678static void sched_dl_do_global(void) 7679{ 7680 u64 new_bw = -1; 7681 int cpu; 7682 unsigned long flags; 7683 7684 def_dl_bandwidth.dl_period = global_rt_period(); 7685 def_dl_bandwidth.dl_runtime = global_rt_runtime(); 7686 7687 if (global_rt_runtime() != RUNTIME_INF) 7688 new_bw = to_ratio(global_rt_period(), global_rt_runtime()); 7689 7690 /* 7691 * FIXME: As above... 7692 */ 7693 for_each_possible_cpu(cpu) { 7694 struct dl_bw *dl_b = dl_bw_of(cpu); 7695 7696 raw_spin_lock_irqsave(&dl_b->lock, flags); 7697 dl_b->bw = new_bw; 7698 raw_spin_unlock_irqrestore(&dl_b->lock, flags); 7699 } 7700} 7701 7702static int sched_rt_global_validate(void) 7703{ 7704 if (sysctl_sched_rt_period <= 0) 7705 return -EINVAL; 7706 7707 if ((sysctl_sched_rt_runtime != RUNTIME_INF) && 7708 (sysctl_sched_rt_runtime > sysctl_sched_rt_period)) 7709 return -EINVAL; 7710 7711 return 0; 7712} 7713 7714static void sched_rt_do_global(void) 7715{ 7716 def_rt_bandwidth.rt_runtime = global_rt_runtime(); 7717 def_rt_bandwidth.rt_period = ns_to_ktime(global_rt_period()); 7718} 7719 7720int sched_rt_handler(struct ctl_table *table, int write, 7721 void __user *buffer, size_t *lenp, 7722 loff_t *ppos) 7723{ 7724 int old_period, old_runtime; 7725 static DEFINE_MUTEX(mutex); 7726 int ret; 7727 7728 mutex_lock(&mutex); 7729 old_period = sysctl_sched_rt_period; 7730 old_runtime = sysctl_sched_rt_runtime; 7731 7732 ret = proc_dointvec(table, write, buffer, lenp, ppos); 7733 7734 if (!ret && write) { 7735 ret = sched_rt_global_validate(); 7736 if (ret) 7737 goto undo; 7738 7739 ret = sched_rt_global_constraints(); 7740 if (ret) 7741 goto undo; 7742 7743 ret = sched_dl_global_constraints(); 7744 if (ret) 7745 goto undo; 7746 7747 sched_rt_do_global(); 7748 sched_dl_do_global(); 7749 } 7750 if (0) { 7751undo: 7752 sysctl_sched_rt_period = old_period; 7753 sysctl_sched_rt_runtime = old_runtime; 7754 } 7755 mutex_unlock(&mutex); 7756 7757 return ret; 7758} 7759 7760int sched_rr_handler(struct ctl_table *table, int write, 7761 void __user *buffer, size_t *lenp, 7762 loff_t *ppos) 7763{ 7764 int ret; 7765 static DEFINE_MUTEX(mutex); 7766 7767 mutex_lock(&mutex); 7768 ret = proc_dointvec(table, write, buffer, lenp, ppos); 7769 /* make sure that internally we keep jiffies */ 7770 /* also, writing zero resets timeslice to default */ 7771 if (!ret && write) { 7772 sched_rr_timeslice = sched_rr_timeslice <= 0 ? 7773 RR_TIMESLICE : msecs_to_jiffies(sched_rr_timeslice); 7774 } 7775 mutex_unlock(&mutex); 7776 return ret; 7777} 7778 7779#ifdef CONFIG_CGROUP_SCHED 7780 7781static inline struct task_group *css_tg(struct cgroup_subsys_state *css) 7782{ 7783 return css ? container_of(css, struct task_group, css) : NULL; 7784} 7785 7786static struct cgroup_subsys_state * 7787cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css) 7788{ 7789 struct task_group *parent = css_tg(parent_css); 7790 struct task_group *tg; 7791 7792 if (!parent) { 7793 /* This is early initialization for the top cgroup */ 7794 return &root_task_group.css; 7795 } 7796 7797 tg = sched_create_group(parent); 7798 if (IS_ERR(tg)) 7799 return ERR_PTR(-ENOMEM); 7800 7801 return &tg->css; 7802} 7803 7804static int cpu_cgroup_css_online(struct cgroup_subsys_state *css) 7805{ 7806 struct task_group *tg = css_tg(css); 7807 struct task_group *parent = css_tg(css->parent); 7808 7809 if (parent) 7810 sched_online_group(tg, parent); 7811 return 0; 7812} 7813 7814static void cpu_cgroup_css_free(struct cgroup_subsys_state *css) 7815{ 7816 struct task_group *tg = css_tg(css); 7817 7818 sched_destroy_group(tg); 7819} 7820 7821static void cpu_cgroup_css_offline(struct cgroup_subsys_state *css) 7822{ 7823 struct task_group *tg = css_tg(css); 7824 7825 sched_offline_group(tg); 7826} 7827 7828static int cpu_cgroup_can_attach(struct cgroup_subsys_state *css, 7829 struct cgroup_taskset *tset) 7830{ 7831 struct task_struct *task; 7832 7833 cgroup_taskset_for_each(task, tset) { 7834#ifdef CONFIG_RT_GROUP_SCHED 7835 if (!sched_rt_can_attach(css_tg(css), task)) 7836 return -EINVAL; 7837#else 7838 /* We don't support RT-tasks being in separate groups */ 7839 if (task->sched_class != &fair_sched_class) 7840 return -EINVAL; 7841#endif 7842 } 7843 return 0; 7844} 7845 7846static void cpu_cgroup_attach(struct cgroup_subsys_state *css, 7847 struct cgroup_taskset *tset) 7848{ 7849 struct task_struct *task; 7850 7851 cgroup_taskset_for_each(task, tset) 7852 sched_move_task(task); 7853} 7854 7855static void cpu_cgroup_exit(struct cgroup_subsys_state *css, 7856 struct cgroup_subsys_state *old_css, 7857 struct task_struct *task) 7858{ 7859 /* 7860 * cgroup_exit() is called in the copy_process() failure path. 7861 * Ignore this case since the task hasn't ran yet, this avoids 7862 * trying to poke a half freed task state from generic code. 7863 */ 7864 if (!(task->flags & PF_EXITING)) 7865 return; 7866 7867 sched_move_task(task); 7868} 7869 7870#ifdef CONFIG_FAIR_GROUP_SCHED 7871static int cpu_shares_write_u64(struct cgroup_subsys_state *css, 7872 struct cftype *cftype, u64 shareval) 7873{ 7874 return sched_group_set_shares(css_tg(css), scale_load(shareval)); 7875} 7876 7877static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css, 7878 struct cftype *cft) 7879{ 7880 struct task_group *tg = css_tg(css); 7881 7882 return (u64) scale_load_down(tg->shares); 7883} 7884 7885#ifdef CONFIG_CFS_BANDWIDTH 7886static DEFINE_MUTEX(cfs_constraints_mutex); 7887 7888const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */ 7889const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */ 7890 7891static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime); 7892 7893static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota) 7894{ 7895 int i, ret = 0, runtime_enabled, runtime_was_enabled; 7896 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth; 7897 7898 if (tg == &root_task_group) 7899 return -EINVAL; 7900 7901 /* 7902 * Ensure we have at some amount of bandwidth every period. This is 7903 * to prevent reaching a state of large arrears when throttled via 7904 * entity_tick() resulting in prolonged exit starvation. 7905 */ 7906 if (quota < min_cfs_quota_period || period < min_cfs_quota_period) 7907 return -EINVAL; 7908 7909 /* 7910 * Likewise, bound things on the otherside by preventing insane quota 7911 * periods. This also allows us to normalize in computing quota 7912 * feasibility. 7913 */ 7914 if (period > max_cfs_quota_period) 7915 return -EINVAL; 7916 7917 /* 7918 * Prevent race between setting of cfs_rq->runtime_enabled and 7919 * unthrottle_offline_cfs_rqs(). 7920 */ 7921 get_online_cpus(); 7922 mutex_lock(&cfs_constraints_mutex); 7923 ret = __cfs_schedulable(tg, period, quota); 7924 if (ret) 7925 goto out_unlock; 7926 7927 runtime_enabled = quota != RUNTIME_INF; 7928 runtime_was_enabled = cfs_b->quota != RUNTIME_INF; 7929 /* 7930 * If we need to toggle cfs_bandwidth_used, off->on must occur 7931 * before making related changes, and on->off must occur afterwards 7932 */ 7933 if (runtime_enabled && !runtime_was_enabled) 7934 cfs_bandwidth_usage_inc(); 7935 raw_spin_lock_irq(&cfs_b->lock); 7936 cfs_b->period = ns_to_ktime(period); 7937 cfs_b->quota = quota; 7938 7939 __refill_cfs_bandwidth_runtime(cfs_b); 7940 /* restart the period timer (if active) to handle new period expiry */ 7941 if (runtime_enabled && cfs_b->timer_active) { 7942 /* force a reprogram */ 7943 __start_cfs_bandwidth(cfs_b, true); 7944 } 7945 raw_spin_unlock_irq(&cfs_b->lock); 7946 7947 for_each_online_cpu(i) { 7948 struct cfs_rq *cfs_rq = tg->cfs_rq[i]; 7949 struct rq *rq = cfs_rq->rq; 7950 7951 raw_spin_lock_irq(&rq->lock); 7952 cfs_rq->runtime_enabled = runtime_enabled; 7953 cfs_rq->runtime_remaining = 0; 7954 7955 if (cfs_rq->throttled) 7956 unthrottle_cfs_rq(cfs_rq); 7957 raw_spin_unlock_irq(&rq->lock); 7958 } 7959 if (runtime_was_enabled && !runtime_enabled) 7960 cfs_bandwidth_usage_dec(); 7961out_unlock: 7962 mutex_unlock(&cfs_constraints_mutex); 7963 put_online_cpus(); 7964 7965 return ret; 7966} 7967 7968int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us) 7969{ 7970 u64 quota, period; 7971 7972 period = ktime_to_ns(tg->cfs_bandwidth.period); 7973 if (cfs_quota_us < 0) 7974 quota = RUNTIME_INF; 7975 else 7976 quota = (u64)cfs_quota_us * NSEC_PER_USEC; 7977 7978 return tg_set_cfs_bandwidth(tg, period, quota); 7979} 7980 7981long tg_get_cfs_quota(struct task_group *tg) 7982{ 7983 u64 quota_us; 7984 7985 if (tg->cfs_bandwidth.quota == RUNTIME_INF) 7986 return -1; 7987 7988 quota_us = tg->cfs_bandwidth.quota; 7989 do_div(quota_us, NSEC_PER_USEC); 7990 7991 return quota_us; 7992} 7993 7994int tg_set_cfs_period(struct task_group *tg, long cfs_period_us) 7995{ 7996 u64 quota, period; 7997 7998 period = (u64)cfs_period_us * NSEC_PER_USEC; 7999 quota = tg->cfs_bandwidth.quota; 8000 8001 return tg_set_cfs_bandwidth(tg, period, quota); 8002} 8003 8004long tg_get_cfs_period(struct task_group *tg) 8005{ 8006 u64 cfs_period_us; 8007 8008 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period); 8009 do_div(cfs_period_us, NSEC_PER_USEC); 8010 8011 return cfs_period_us; 8012} 8013 8014static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css, 8015 struct cftype *cft) 8016{ 8017 return tg_get_cfs_quota(css_tg(css)); 8018} 8019 8020static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css, 8021 struct cftype *cftype, s64 cfs_quota_us) 8022{ 8023 return tg_set_cfs_quota(css_tg(css), cfs_quota_us); 8024} 8025 8026static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css, 8027 struct cftype *cft) 8028{ 8029 return tg_get_cfs_period(css_tg(css)); 8030} 8031 8032static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css, 8033 struct cftype *cftype, u64 cfs_period_us) 8034{ 8035 return tg_set_cfs_period(css_tg(css), cfs_period_us); 8036} 8037 8038struct cfs_schedulable_data { 8039 struct task_group *tg; 8040 u64 period, quota; 8041}; 8042 8043/* 8044 * normalize group quota/period to be quota/max_period 8045 * note: units are usecs 8046 */ 8047static u64 normalize_cfs_quota(struct task_group *tg, 8048 struct cfs_schedulable_data *d) 8049{ 8050 u64 quota, period; 8051 8052 if (tg == d->tg) { 8053 period = d->period; 8054 quota = d->quota; 8055 } else { 8056 period = tg_get_cfs_period(tg); 8057 quota = tg_get_cfs_quota(tg); 8058 } 8059 8060 /* note: these should typically be equivalent */ 8061 if (quota == RUNTIME_INF || quota == -1) 8062 return RUNTIME_INF; 8063 8064 return to_ratio(period, quota); 8065} 8066 8067static int tg_cfs_schedulable_down(struct task_group *tg, void *data) 8068{ 8069 struct cfs_schedulable_data *d = data; 8070 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth; 8071 s64 quota = 0, parent_quota = -1; 8072 8073 if (!tg->parent) { 8074 quota = RUNTIME_INF; 8075 } else { 8076 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth; 8077 8078 quota = normalize_cfs_quota(tg, d); 8079 parent_quota = parent_b->hierarchical_quota; 8080 8081 /* 8082 * ensure max(child_quota) <= parent_quota, inherit when no 8083 * limit is set 8084 */ 8085 if (quota == RUNTIME_INF) 8086 quota = parent_quota; 8087 else if (parent_quota != RUNTIME_INF && quota > parent_quota) 8088 return -EINVAL; 8089 } 8090 cfs_b->hierarchical_quota = quota; 8091 8092 return 0; 8093} 8094 8095static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota) 8096{ 8097 int ret; 8098 struct cfs_schedulable_data data = { 8099 .tg = tg, 8100 .period = period, 8101 .quota = quota, 8102 }; 8103 8104 if (quota != RUNTIME_INF) { 8105 do_div(data.period, NSEC_PER_USEC); 8106 do_div(data.quota, NSEC_PER_USEC); 8107 } 8108 8109 rcu_read_lock(); 8110 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data); 8111 rcu_read_unlock(); 8112 8113 return ret; 8114} 8115 8116static int cpu_stats_show(struct seq_file *sf, void *v) 8117{ 8118 struct task_group *tg = css_tg(seq_css(sf)); 8119 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth; 8120 8121 seq_printf(sf, "nr_periods %d\n", cfs_b->nr_periods); 8122 seq_printf(sf, "nr_throttled %d\n", cfs_b->nr_throttled); 8123 seq_printf(sf, "throttled_time %llu\n", cfs_b->throttled_time); 8124 8125 return 0; 8126} 8127#endif /* CONFIG_CFS_BANDWIDTH */ 8128#endif /* CONFIG_FAIR_GROUP_SCHED */ 8129 8130#ifdef CONFIG_RT_GROUP_SCHED 8131static int cpu_rt_runtime_write(struct cgroup_subsys_state *css, 8132 struct cftype *cft, s64 val) 8133{ 8134 return sched_group_set_rt_runtime(css_tg(css), val); 8135} 8136 8137static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css, 8138 struct cftype *cft) 8139{ 8140 return sched_group_rt_runtime(css_tg(css)); 8141} 8142 8143static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css, 8144 struct cftype *cftype, u64 rt_period_us) 8145{ 8146 return sched_group_set_rt_period(css_tg(css), rt_period_us); 8147} 8148 8149static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css, 8150 struct cftype *cft) 8151{ 8152 return sched_group_rt_period(css_tg(css)); 8153} 8154#endif /* CONFIG_RT_GROUP_SCHED */ 8155 8156static struct cftype cpu_files[] = { 8157#ifdef CONFIG_FAIR_GROUP_SCHED 8158 { 8159 .name = "shares", 8160 .read_u64 = cpu_shares_read_u64, 8161 .write_u64 = cpu_shares_write_u64, 8162 }, 8163#endif 8164#ifdef CONFIG_CFS_BANDWIDTH 8165 { 8166 .name = "cfs_quota_us", 8167 .read_s64 = cpu_cfs_quota_read_s64, 8168 .write_s64 = cpu_cfs_quota_write_s64, 8169 }, 8170 { 8171 .name = "cfs_period_us", 8172 .read_u64 = cpu_cfs_period_read_u64, 8173 .write_u64 = cpu_cfs_period_write_u64, 8174 }, 8175 { 8176 .name = "stat", 8177 .seq_show = cpu_stats_show, 8178 }, 8179#endif 8180#ifdef CONFIG_RT_GROUP_SCHED 8181 { 8182 .name = "rt_runtime_us", 8183 .read_s64 = cpu_rt_runtime_read, 8184 .write_s64 = cpu_rt_runtime_write, 8185 }, 8186 { 8187 .name = "rt_period_us", 8188 .read_u64 = cpu_rt_period_read_uint, 8189 .write_u64 = cpu_rt_period_write_uint, 8190 }, 8191#endif 8192 { } /* terminate */ 8193}; 8194 8195struct cgroup_subsys cpu_cgrp_subsys = { 8196 .css_alloc = cpu_cgroup_css_alloc, 8197 .css_free = cpu_cgroup_css_free, 8198 .css_online = cpu_cgroup_css_online, 8199 .css_offline = cpu_cgroup_css_offline, 8200 .can_attach = cpu_cgroup_can_attach, 8201 .attach = cpu_cgroup_attach, 8202 .exit = cpu_cgroup_exit, 8203 .legacy_cftypes = cpu_files, 8204 .early_init = 1, 8205}; 8206 8207#endif /* CONFIG_CGROUP_SCHED */ 8208 8209void dump_cpu_task(int cpu) 8210{ 8211 pr_info("Task dump for CPU %d:\n", cpu); 8212 sched_show_task(cpu_curr(cpu)); 8213} 8214