core.c revision 466af29bf4270e84261712428a1304c28e3743fa
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; 248 249 if (cnt > 63) 250 cnt = 63; 251 252 if (copy_from_user(&buf, ubuf, cnt)) 253 return -EFAULT; 254 255 buf[cnt] = 0; 256 cmp = strstrip(buf); 257 258 i = sched_feat_set(cmp); 259 if (i == __SCHED_FEAT_NR) 260 return -EINVAL; 261 262 *ppos += cnt; 263 264 return cnt; 265} 266 267static int sched_feat_open(struct inode *inode, struct file *filp) 268{ 269 return single_open(filp, sched_feat_show, NULL); 270} 271 272static const struct file_operations sched_feat_fops = { 273 .open = sched_feat_open, 274 .write = sched_feat_write, 275 .read = seq_read, 276 .llseek = seq_lseek, 277 .release = single_release, 278}; 279 280static __init int sched_init_debug(void) 281{ 282 debugfs_create_file("sched_features", 0644, NULL, NULL, 283 &sched_feat_fops); 284 285 return 0; 286} 287late_initcall(sched_init_debug); 288#endif /* CONFIG_SCHED_DEBUG */ 289 290/* 291 * Number of tasks to iterate in a single balance run. 292 * Limited because this is done with IRQs disabled. 293 */ 294const_debug unsigned int sysctl_sched_nr_migrate = 32; 295 296/* 297 * period over which we average the RT time consumption, measured 298 * in ms. 299 * 300 * default: 1s 301 */ 302const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC; 303 304/* 305 * period over which we measure -rt task cpu usage in us. 306 * default: 1s 307 */ 308unsigned int sysctl_sched_rt_period = 1000000; 309 310__read_mostly int scheduler_running; 311 312/* 313 * part of the period that we allow rt tasks to run in us. 314 * default: 0.95s 315 */ 316int sysctl_sched_rt_runtime = 950000; 317 318/* 319 * __task_rq_lock - lock the rq @p resides on. 320 */ 321static inline struct rq *__task_rq_lock(struct task_struct *p) 322 __acquires(rq->lock) 323{ 324 struct rq *rq; 325 326 lockdep_assert_held(&p->pi_lock); 327 328 for (;;) { 329 rq = task_rq(p); 330 raw_spin_lock(&rq->lock); 331 if (likely(rq == task_rq(p))) 332 return rq; 333 raw_spin_unlock(&rq->lock); 334 } 335} 336 337/* 338 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on. 339 */ 340static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags) 341 __acquires(p->pi_lock) 342 __acquires(rq->lock) 343{ 344 struct rq *rq; 345 346 for (;;) { 347 raw_spin_lock_irqsave(&p->pi_lock, *flags); 348 rq = task_rq(p); 349 raw_spin_lock(&rq->lock); 350 if (likely(rq == task_rq(p))) 351 return rq; 352 raw_spin_unlock(&rq->lock); 353 raw_spin_unlock_irqrestore(&p->pi_lock, *flags); 354 } 355} 356 357static void __task_rq_unlock(struct rq *rq) 358 __releases(rq->lock) 359{ 360 raw_spin_unlock(&rq->lock); 361} 362 363static inline void 364task_rq_unlock(struct rq *rq, struct task_struct *p, unsigned long *flags) 365 __releases(rq->lock) 366 __releases(p->pi_lock) 367{ 368 raw_spin_unlock(&rq->lock); 369 raw_spin_unlock_irqrestore(&p->pi_lock, *flags); 370} 371 372/* 373 * this_rq_lock - lock this runqueue and disable interrupts. 374 */ 375static struct rq *this_rq_lock(void) 376 __acquires(rq->lock) 377{ 378 struct rq *rq; 379 380 local_irq_disable(); 381 rq = this_rq(); 382 raw_spin_lock(&rq->lock); 383 384 return rq; 385} 386 387#ifdef CONFIG_SCHED_HRTICK 388/* 389 * Use HR-timers to deliver accurate preemption points. 390 */ 391 392static void hrtick_clear(struct rq *rq) 393{ 394 if (hrtimer_active(&rq->hrtick_timer)) 395 hrtimer_cancel(&rq->hrtick_timer); 396} 397 398/* 399 * High-resolution timer tick. 400 * Runs from hardirq context with interrupts disabled. 401 */ 402static enum hrtimer_restart hrtick(struct hrtimer *timer) 403{ 404 struct rq *rq = container_of(timer, struct rq, hrtick_timer); 405 406 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id()); 407 408 raw_spin_lock(&rq->lock); 409 update_rq_clock(rq); 410 rq->curr->sched_class->task_tick(rq, rq->curr, 1); 411 raw_spin_unlock(&rq->lock); 412 413 return HRTIMER_NORESTART; 414} 415 416#ifdef CONFIG_SMP 417 418static int __hrtick_restart(struct rq *rq) 419{ 420 struct hrtimer *timer = &rq->hrtick_timer; 421 ktime_t time = hrtimer_get_softexpires(timer); 422 423 return __hrtimer_start_range_ns(timer, time, 0, HRTIMER_MODE_ABS_PINNED, 0); 424} 425 426/* 427 * called from hardirq (IPI) context 428 */ 429static void __hrtick_start(void *arg) 430{ 431 struct rq *rq = arg; 432 433 raw_spin_lock(&rq->lock); 434 __hrtick_restart(rq); 435 rq->hrtick_csd_pending = 0; 436 raw_spin_unlock(&rq->lock); 437} 438 439/* 440 * Called to set the hrtick timer state. 441 * 442 * called with rq->lock held and irqs disabled 443 */ 444void hrtick_start(struct rq *rq, u64 delay) 445{ 446 struct hrtimer *timer = &rq->hrtick_timer; 447 ktime_t time = ktime_add_ns(timer->base->get_time(), delay); 448 449 hrtimer_set_expires(timer, time); 450 451 if (rq == this_rq()) { 452 __hrtick_restart(rq); 453 } else if (!rq->hrtick_csd_pending) { 454 smp_call_function_single_async(cpu_of(rq), &rq->hrtick_csd); 455 rq->hrtick_csd_pending = 1; 456 } 457} 458 459static int 460hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu) 461{ 462 int cpu = (int)(long)hcpu; 463 464 switch (action) { 465 case CPU_UP_CANCELED: 466 case CPU_UP_CANCELED_FROZEN: 467 case CPU_DOWN_PREPARE: 468 case CPU_DOWN_PREPARE_FROZEN: 469 case CPU_DEAD: 470 case CPU_DEAD_FROZEN: 471 hrtick_clear(cpu_rq(cpu)); 472 return NOTIFY_OK; 473 } 474 475 return NOTIFY_DONE; 476} 477 478static __init void init_hrtick(void) 479{ 480 hotcpu_notifier(hotplug_hrtick, 0); 481} 482#else 483/* 484 * Called to set the hrtick timer state. 485 * 486 * called with rq->lock held and irqs disabled 487 */ 488void hrtick_start(struct rq *rq, u64 delay) 489{ 490 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0, 491 HRTIMER_MODE_REL_PINNED, 0); 492} 493 494static inline void init_hrtick(void) 495{ 496} 497#endif /* CONFIG_SMP */ 498 499static void init_rq_hrtick(struct rq *rq) 500{ 501#ifdef CONFIG_SMP 502 rq->hrtick_csd_pending = 0; 503 504 rq->hrtick_csd.flags = 0; 505 rq->hrtick_csd.func = __hrtick_start; 506 rq->hrtick_csd.info = rq; 507#endif 508 509 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); 510 rq->hrtick_timer.function = hrtick; 511} 512#else /* CONFIG_SCHED_HRTICK */ 513static inline void hrtick_clear(struct rq *rq) 514{ 515} 516 517static inline void init_rq_hrtick(struct rq *rq) 518{ 519} 520 521static inline void init_hrtick(void) 522{ 523} 524#endif /* CONFIG_SCHED_HRTICK */ 525 526/* 527 * cmpxchg based fetch_or, macro so it works for different integer types 528 */ 529#define fetch_or(ptr, val) \ 530({ typeof(*(ptr)) __old, __val = *(ptr); \ 531 for (;;) { \ 532 __old = cmpxchg((ptr), __val, __val | (val)); \ 533 if (__old == __val) \ 534 break; \ 535 __val = __old; \ 536 } \ 537 __old; \ 538}) 539 540#if defined(CONFIG_SMP) && defined(TIF_POLLING_NRFLAG) 541/* 542 * Atomically set TIF_NEED_RESCHED and test for TIF_POLLING_NRFLAG, 543 * this avoids any races wrt polling state changes and thereby avoids 544 * spurious IPIs. 545 */ 546static bool set_nr_and_not_polling(struct task_struct *p) 547{ 548 struct thread_info *ti = task_thread_info(p); 549 return !(fetch_or(&ti->flags, _TIF_NEED_RESCHED) & _TIF_POLLING_NRFLAG); 550} 551 552/* 553 * Atomically set TIF_NEED_RESCHED if TIF_POLLING_NRFLAG is set. 554 * 555 * If this returns true, then the idle task promises to call 556 * sched_ttwu_pending() and reschedule soon. 557 */ 558static bool set_nr_if_polling(struct task_struct *p) 559{ 560 struct thread_info *ti = task_thread_info(p); 561 typeof(ti->flags) old, val = ACCESS_ONCE(ti->flags); 562 563 for (;;) { 564 if (!(val & _TIF_POLLING_NRFLAG)) 565 return false; 566 if (val & _TIF_NEED_RESCHED) 567 return true; 568 old = cmpxchg(&ti->flags, val, val | _TIF_NEED_RESCHED); 569 if (old == val) 570 break; 571 val = old; 572 } 573 return true; 574} 575 576#else 577static bool set_nr_and_not_polling(struct task_struct *p) 578{ 579 set_tsk_need_resched(p); 580 return true; 581} 582 583#ifdef CONFIG_SMP 584static bool set_nr_if_polling(struct task_struct *p) 585{ 586 return false; 587} 588#endif 589#endif 590 591/* 592 * resched_task - mark a task 'to be rescheduled now'. 593 * 594 * On UP this means the setting of the need_resched flag, on SMP it 595 * might also involve a cross-CPU call to trigger the scheduler on 596 * the target CPU. 597 */ 598void resched_task(struct task_struct *p) 599{ 600 int cpu; 601 602 lockdep_assert_held(&task_rq(p)->lock); 603 604 if (test_tsk_need_resched(p)) 605 return; 606 607 cpu = task_cpu(p); 608 609 if (cpu == smp_processor_id()) { 610 set_tsk_need_resched(p); 611 set_preempt_need_resched(); 612 return; 613 } 614 615 if (set_nr_and_not_polling(p)) 616 smp_send_reschedule(cpu); 617 else 618 trace_sched_wake_idle_without_ipi(cpu); 619} 620 621void resched_cpu(int cpu) 622{ 623 struct rq *rq = cpu_rq(cpu); 624 unsigned long flags; 625 626 if (!raw_spin_trylock_irqsave(&rq->lock, flags)) 627 return; 628 resched_task(cpu_curr(cpu)); 629 raw_spin_unlock_irqrestore(&rq->lock, flags); 630} 631 632#ifdef CONFIG_SMP 633#ifdef CONFIG_NO_HZ_COMMON 634/* 635 * In the semi idle case, use the nearest busy cpu for migrating timers 636 * from an idle cpu. This is good for power-savings. 637 * 638 * We don't do similar optimization for completely idle system, as 639 * selecting an idle cpu will add more delays to the timers than intended 640 * (as that cpu's timer base may not be uptodate wrt jiffies etc). 641 */ 642int get_nohz_timer_target(int pinned) 643{ 644 int cpu = smp_processor_id(); 645 int i; 646 struct sched_domain *sd; 647 648 if (pinned || !get_sysctl_timer_migration() || !idle_cpu(cpu)) 649 return cpu; 650 651 rcu_read_lock(); 652 for_each_domain(cpu, sd) { 653 for_each_cpu(i, sched_domain_span(sd)) { 654 if (!idle_cpu(i)) { 655 cpu = i; 656 goto unlock; 657 } 658 } 659 } 660unlock: 661 rcu_read_unlock(); 662 return cpu; 663} 664/* 665 * When add_timer_on() enqueues a timer into the timer wheel of an 666 * idle CPU then this timer might expire before the next timer event 667 * which is scheduled to wake up that CPU. In case of a completely 668 * idle system the next event might even be infinite time into the 669 * future. wake_up_idle_cpu() ensures that the CPU is woken up and 670 * leaves the inner idle loop so the newly added timer is taken into 671 * account when the CPU goes back to idle and evaluates the timer 672 * wheel for the next timer event. 673 */ 674static void wake_up_idle_cpu(int cpu) 675{ 676 struct rq *rq = cpu_rq(cpu); 677 678 if (cpu == smp_processor_id()) 679 return; 680 681 if (set_nr_and_not_polling(rq->idle)) 682 smp_send_reschedule(cpu); 683 else 684 trace_sched_wake_idle_without_ipi(cpu); 685} 686 687static bool wake_up_full_nohz_cpu(int cpu) 688{ 689 /* 690 * We just need the target to call irq_exit() and re-evaluate 691 * the next tick. The nohz full kick at least implies that. 692 * If needed we can still optimize that later with an 693 * empty IRQ. 694 */ 695 if (tick_nohz_full_cpu(cpu)) { 696 if (cpu != smp_processor_id() || 697 tick_nohz_tick_stopped()) 698 tick_nohz_full_kick_cpu(cpu); 699 return true; 700 } 701 702 return false; 703} 704 705void wake_up_nohz_cpu(int cpu) 706{ 707 if (!wake_up_full_nohz_cpu(cpu)) 708 wake_up_idle_cpu(cpu); 709} 710 711static inline bool got_nohz_idle_kick(void) 712{ 713 int cpu = smp_processor_id(); 714 715 if (!test_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu))) 716 return false; 717 718 if (idle_cpu(cpu) && !need_resched()) 719 return true; 720 721 /* 722 * We can't run Idle Load Balance on this CPU for this time so we 723 * cancel it and clear NOHZ_BALANCE_KICK 724 */ 725 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu)); 726 return false; 727} 728 729#else /* CONFIG_NO_HZ_COMMON */ 730 731static inline bool got_nohz_idle_kick(void) 732{ 733 return false; 734} 735 736#endif /* CONFIG_NO_HZ_COMMON */ 737 738#ifdef CONFIG_NO_HZ_FULL 739bool sched_can_stop_tick(void) 740{ 741 /* 742 * More than one running task need preemption. 743 * nr_running update is assumed to be visible 744 * after IPI is sent from wakers. 745 */ 746 if (this_rq()->nr_running > 1) 747 return false; 748 749 return true; 750} 751#endif /* CONFIG_NO_HZ_FULL */ 752 753void sched_avg_update(struct rq *rq) 754{ 755 s64 period = sched_avg_period(); 756 757 while ((s64)(rq_clock(rq) - rq->age_stamp) > period) { 758 /* 759 * Inline assembly required to prevent the compiler 760 * optimising this loop into a divmod call. 761 * See __iter_div_u64_rem() for another example of this. 762 */ 763 asm("" : "+rm" (rq->age_stamp)); 764 rq->age_stamp += period; 765 rq->rt_avg /= 2; 766 } 767} 768 769#endif /* CONFIG_SMP */ 770 771#if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \ 772 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH))) 773/* 774 * Iterate task_group tree rooted at *from, calling @down when first entering a 775 * node and @up when leaving it for the final time. 776 * 777 * Caller must hold rcu_lock or sufficient equivalent. 778 */ 779int walk_tg_tree_from(struct task_group *from, 780 tg_visitor down, tg_visitor up, void *data) 781{ 782 struct task_group *parent, *child; 783 int ret; 784 785 parent = from; 786 787down: 788 ret = (*down)(parent, data); 789 if (ret) 790 goto out; 791 list_for_each_entry_rcu(child, &parent->children, siblings) { 792 parent = child; 793 goto down; 794 795up: 796 continue; 797 } 798 ret = (*up)(parent, data); 799 if (ret || parent == from) 800 goto out; 801 802 child = parent; 803 parent = parent->parent; 804 if (parent) 805 goto up; 806out: 807 return ret; 808} 809 810int tg_nop(struct task_group *tg, void *data) 811{ 812 return 0; 813} 814#endif 815 816static void set_load_weight(struct task_struct *p) 817{ 818 int prio = p->static_prio - MAX_RT_PRIO; 819 struct load_weight *load = &p->se.load; 820 821 /* 822 * SCHED_IDLE tasks get minimal weight: 823 */ 824 if (p->policy == SCHED_IDLE) { 825 load->weight = scale_load(WEIGHT_IDLEPRIO); 826 load->inv_weight = WMULT_IDLEPRIO; 827 return; 828 } 829 830 load->weight = scale_load(prio_to_weight[prio]); 831 load->inv_weight = prio_to_wmult[prio]; 832} 833 834static void enqueue_task(struct rq *rq, struct task_struct *p, int flags) 835{ 836 update_rq_clock(rq); 837 sched_info_queued(rq, p); 838 p->sched_class->enqueue_task(rq, p, flags); 839} 840 841static void dequeue_task(struct rq *rq, struct task_struct *p, int flags) 842{ 843 update_rq_clock(rq); 844 sched_info_dequeued(rq, p); 845 p->sched_class->dequeue_task(rq, p, flags); 846} 847 848void activate_task(struct rq *rq, struct task_struct *p, int flags) 849{ 850 if (task_contributes_to_load(p)) 851 rq->nr_uninterruptible--; 852 853 enqueue_task(rq, p, flags); 854} 855 856void deactivate_task(struct rq *rq, struct task_struct *p, int flags) 857{ 858 if (task_contributes_to_load(p)) 859 rq->nr_uninterruptible++; 860 861 dequeue_task(rq, p, flags); 862} 863 864static void update_rq_clock_task(struct rq *rq, s64 delta) 865{ 866/* 867 * In theory, the compile should just see 0 here, and optimize out the call 868 * to sched_rt_avg_update. But I don't trust it... 869 */ 870#if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING) 871 s64 steal = 0, irq_delta = 0; 872#endif 873#ifdef CONFIG_IRQ_TIME_ACCOUNTING 874 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time; 875 876 /* 877 * Since irq_time is only updated on {soft,}irq_exit, we might run into 878 * this case when a previous update_rq_clock() happened inside a 879 * {soft,}irq region. 880 * 881 * When this happens, we stop ->clock_task and only update the 882 * prev_irq_time stamp to account for the part that fit, so that a next 883 * update will consume the rest. This ensures ->clock_task is 884 * monotonic. 885 * 886 * It does however cause some slight miss-attribution of {soft,}irq 887 * time, a more accurate solution would be to update the irq_time using 888 * the current rq->clock timestamp, except that would require using 889 * atomic ops. 890 */ 891 if (irq_delta > delta) 892 irq_delta = delta; 893 894 rq->prev_irq_time += irq_delta; 895 delta -= irq_delta; 896#endif 897#ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING 898 if (static_key_false((¶virt_steal_rq_enabled))) { 899 steal = paravirt_steal_clock(cpu_of(rq)); 900 steal -= rq->prev_steal_time_rq; 901 902 if (unlikely(steal > delta)) 903 steal = delta; 904 905 rq->prev_steal_time_rq += steal; 906 delta -= steal; 907 } 908#endif 909 910 rq->clock_task += delta; 911 912#if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING) 913 if ((irq_delta + steal) && sched_feat(NONTASK_CAPACITY)) 914 sched_rt_avg_update(rq, irq_delta + steal); 915#endif 916} 917 918void sched_set_stop_task(int cpu, struct task_struct *stop) 919{ 920 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 }; 921 struct task_struct *old_stop = cpu_rq(cpu)->stop; 922 923 if (stop) { 924 /* 925 * Make it appear like a SCHED_FIFO task, its something 926 * userspace knows about and won't get confused about. 927 * 928 * Also, it will make PI more or less work without too 929 * much confusion -- but then, stop work should not 930 * rely on PI working anyway. 931 */ 932 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m); 933 934 stop->sched_class = &stop_sched_class; 935 } 936 937 cpu_rq(cpu)->stop = stop; 938 939 if (old_stop) { 940 /* 941 * Reset it back to a normal scheduling class so that 942 * it can die in pieces. 943 */ 944 old_stop->sched_class = &rt_sched_class; 945 } 946} 947 948/* 949 * __normal_prio - return the priority that is based on the static prio 950 */ 951static inline int __normal_prio(struct task_struct *p) 952{ 953 return p->static_prio; 954} 955 956/* 957 * Calculate the expected normal priority: i.e. priority 958 * without taking RT-inheritance into account. Might be 959 * boosted by interactivity modifiers. Changes upon fork, 960 * setprio syscalls, and whenever the interactivity 961 * estimator recalculates. 962 */ 963static inline int normal_prio(struct task_struct *p) 964{ 965 int prio; 966 967 if (task_has_dl_policy(p)) 968 prio = MAX_DL_PRIO-1; 969 else if (task_has_rt_policy(p)) 970 prio = MAX_RT_PRIO-1 - p->rt_priority; 971 else 972 prio = __normal_prio(p); 973 return prio; 974} 975 976/* 977 * Calculate the current priority, i.e. the priority 978 * taken into account by the scheduler. This value might 979 * be boosted by RT tasks, or might be boosted by 980 * interactivity modifiers. Will be RT if the task got 981 * RT-boosted. If not then it returns p->normal_prio. 982 */ 983static int effective_prio(struct task_struct *p) 984{ 985 p->normal_prio = normal_prio(p); 986 /* 987 * If we are RT tasks or we were boosted to RT priority, 988 * keep the priority unchanged. Otherwise, update priority 989 * to the normal priority: 990 */ 991 if (!rt_prio(p->prio)) 992 return p->normal_prio; 993 return p->prio; 994} 995 996/** 997 * task_curr - is this task currently executing on a CPU? 998 * @p: the task in question. 999 * 1000 * Return: 1 if the task is currently executing. 0 otherwise. 1001 */ 1002inline int task_curr(const struct task_struct *p) 1003{ 1004 return cpu_curr(task_cpu(p)) == p; 1005} 1006 1007static inline void check_class_changed(struct rq *rq, struct task_struct *p, 1008 const struct sched_class *prev_class, 1009 int oldprio) 1010{ 1011 if (prev_class != p->sched_class) { 1012 if (prev_class->switched_from) 1013 prev_class->switched_from(rq, p); 1014 p->sched_class->switched_to(rq, p); 1015 } else if (oldprio != p->prio || dl_task(p)) 1016 p->sched_class->prio_changed(rq, p, oldprio); 1017} 1018 1019void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags) 1020{ 1021 const struct sched_class *class; 1022 1023 if (p->sched_class == rq->curr->sched_class) { 1024 rq->curr->sched_class->check_preempt_curr(rq, p, flags); 1025 } else { 1026 for_each_class(class) { 1027 if (class == rq->curr->sched_class) 1028 break; 1029 if (class == p->sched_class) { 1030 resched_task(rq->curr); 1031 break; 1032 } 1033 } 1034 } 1035 1036 /* 1037 * A queue event has occurred, and we're going to schedule. In 1038 * this case, we can save a useless back to back clock update. 1039 */ 1040 if (rq->curr->on_rq && test_tsk_need_resched(rq->curr)) 1041 rq->skip_clock_update = 1; 1042} 1043 1044#ifdef CONFIG_SMP 1045void set_task_cpu(struct task_struct *p, unsigned int new_cpu) 1046{ 1047#ifdef CONFIG_SCHED_DEBUG 1048 /* 1049 * We should never call set_task_cpu() on a blocked task, 1050 * ttwu() will sort out the placement. 1051 */ 1052 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING && 1053 !(task_preempt_count(p) & PREEMPT_ACTIVE)); 1054 1055#ifdef CONFIG_LOCKDEP 1056 /* 1057 * The caller should hold either p->pi_lock or rq->lock, when changing 1058 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks. 1059 * 1060 * sched_move_task() holds both and thus holding either pins the cgroup, 1061 * see task_group(). 1062 * 1063 * Furthermore, all task_rq users should acquire both locks, see 1064 * task_rq_lock(). 1065 */ 1066 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) || 1067 lockdep_is_held(&task_rq(p)->lock))); 1068#endif 1069#endif 1070 1071 trace_sched_migrate_task(p, new_cpu); 1072 1073 if (task_cpu(p) != new_cpu) { 1074 if (p->sched_class->migrate_task_rq) 1075 p->sched_class->migrate_task_rq(p, new_cpu); 1076 p->se.nr_migrations++; 1077 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, NULL, 0); 1078 } 1079 1080 __set_task_cpu(p, new_cpu); 1081} 1082 1083static void __migrate_swap_task(struct task_struct *p, int cpu) 1084{ 1085 if (p->on_rq) { 1086 struct rq *src_rq, *dst_rq; 1087 1088 src_rq = task_rq(p); 1089 dst_rq = cpu_rq(cpu); 1090 1091 deactivate_task(src_rq, p, 0); 1092 set_task_cpu(p, cpu); 1093 activate_task(dst_rq, p, 0); 1094 check_preempt_curr(dst_rq, p, 0); 1095 } else { 1096 /* 1097 * Task isn't running anymore; make it appear like we migrated 1098 * it before it went to sleep. This means on wakeup we make the 1099 * previous cpu our targer instead of where it really is. 1100 */ 1101 p->wake_cpu = cpu; 1102 } 1103} 1104 1105struct migration_swap_arg { 1106 struct task_struct *src_task, *dst_task; 1107 int src_cpu, dst_cpu; 1108}; 1109 1110static int migrate_swap_stop(void *data) 1111{ 1112 struct migration_swap_arg *arg = data; 1113 struct rq *src_rq, *dst_rq; 1114 int ret = -EAGAIN; 1115 1116 src_rq = cpu_rq(arg->src_cpu); 1117 dst_rq = cpu_rq(arg->dst_cpu); 1118 1119 double_raw_lock(&arg->src_task->pi_lock, 1120 &arg->dst_task->pi_lock); 1121 double_rq_lock(src_rq, dst_rq); 1122 if (task_cpu(arg->dst_task) != arg->dst_cpu) 1123 goto unlock; 1124 1125 if (task_cpu(arg->src_task) != arg->src_cpu) 1126 goto unlock; 1127 1128 if (!cpumask_test_cpu(arg->dst_cpu, tsk_cpus_allowed(arg->src_task))) 1129 goto unlock; 1130 1131 if (!cpumask_test_cpu(arg->src_cpu, tsk_cpus_allowed(arg->dst_task))) 1132 goto unlock; 1133 1134 __migrate_swap_task(arg->src_task, arg->dst_cpu); 1135 __migrate_swap_task(arg->dst_task, arg->src_cpu); 1136 1137 ret = 0; 1138 1139unlock: 1140 double_rq_unlock(src_rq, dst_rq); 1141 raw_spin_unlock(&arg->dst_task->pi_lock); 1142 raw_spin_unlock(&arg->src_task->pi_lock); 1143 1144 return ret; 1145} 1146 1147/* 1148 * Cross migrate two tasks 1149 */ 1150int migrate_swap(struct task_struct *cur, struct task_struct *p) 1151{ 1152 struct migration_swap_arg arg; 1153 int ret = -EINVAL; 1154 1155 arg = (struct migration_swap_arg){ 1156 .src_task = cur, 1157 .src_cpu = task_cpu(cur), 1158 .dst_task = p, 1159 .dst_cpu = task_cpu(p), 1160 }; 1161 1162 if (arg.src_cpu == arg.dst_cpu) 1163 goto out; 1164 1165 /* 1166 * These three tests are all lockless; this is OK since all of them 1167 * will be re-checked with proper locks held further down the line. 1168 */ 1169 if (!cpu_active(arg.src_cpu) || !cpu_active(arg.dst_cpu)) 1170 goto out; 1171 1172 if (!cpumask_test_cpu(arg.dst_cpu, tsk_cpus_allowed(arg.src_task))) 1173 goto out; 1174 1175 if (!cpumask_test_cpu(arg.src_cpu, tsk_cpus_allowed(arg.dst_task))) 1176 goto out; 1177 1178 trace_sched_swap_numa(cur, arg.src_cpu, p, arg.dst_cpu); 1179 ret = stop_two_cpus(arg.dst_cpu, arg.src_cpu, migrate_swap_stop, &arg); 1180 1181out: 1182 return ret; 1183} 1184 1185struct migration_arg { 1186 struct task_struct *task; 1187 int dest_cpu; 1188}; 1189 1190static int migration_cpu_stop(void *data); 1191 1192/* 1193 * wait_task_inactive - wait for a thread to unschedule. 1194 * 1195 * If @match_state is nonzero, it's the @p->state value just checked and 1196 * not expected to change. If it changes, i.e. @p might have woken up, 1197 * then return zero. When we succeed in waiting for @p to be off its CPU, 1198 * we return a positive number (its total switch count). If a second call 1199 * a short while later returns the same number, the caller can be sure that 1200 * @p has remained unscheduled the whole time. 1201 * 1202 * The caller must ensure that the task *will* unschedule sometime soon, 1203 * else this function might spin for a *long* time. This function can't 1204 * be called with interrupts off, or it may introduce deadlock with 1205 * smp_call_function() if an IPI is sent by the same process we are 1206 * waiting to become inactive. 1207 */ 1208unsigned long wait_task_inactive(struct task_struct *p, long match_state) 1209{ 1210 unsigned long flags; 1211 int running, on_rq; 1212 unsigned long ncsw; 1213 struct rq *rq; 1214 1215 for (;;) { 1216 /* 1217 * We do the initial early heuristics without holding 1218 * any task-queue locks at all. We'll only try to get 1219 * the runqueue lock when things look like they will 1220 * work out! 1221 */ 1222 rq = task_rq(p); 1223 1224 /* 1225 * If the task is actively running on another CPU 1226 * still, just relax and busy-wait without holding 1227 * any locks. 1228 * 1229 * NOTE! Since we don't hold any locks, it's not 1230 * even sure that "rq" stays as the right runqueue! 1231 * But we don't care, since "task_running()" will 1232 * return false if the runqueue has changed and p 1233 * is actually now running somewhere else! 1234 */ 1235 while (task_running(rq, p)) { 1236 if (match_state && unlikely(p->state != match_state)) 1237 return 0; 1238 cpu_relax(); 1239 } 1240 1241 /* 1242 * Ok, time to look more closely! We need the rq 1243 * lock now, to be *sure*. If we're wrong, we'll 1244 * just go back and repeat. 1245 */ 1246 rq = task_rq_lock(p, &flags); 1247 trace_sched_wait_task(p); 1248 running = task_running(rq, p); 1249 on_rq = p->on_rq; 1250 ncsw = 0; 1251 if (!match_state || p->state == match_state) 1252 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */ 1253 task_rq_unlock(rq, p, &flags); 1254 1255 /* 1256 * If it changed from the expected state, bail out now. 1257 */ 1258 if (unlikely(!ncsw)) 1259 break; 1260 1261 /* 1262 * Was it really running after all now that we 1263 * checked with the proper locks actually held? 1264 * 1265 * Oops. Go back and try again.. 1266 */ 1267 if (unlikely(running)) { 1268 cpu_relax(); 1269 continue; 1270 } 1271 1272 /* 1273 * It's not enough that it's not actively running, 1274 * it must be off the runqueue _entirely_, and not 1275 * preempted! 1276 * 1277 * So if it was still runnable (but just not actively 1278 * running right now), it's preempted, and we should 1279 * yield - it could be a while. 1280 */ 1281 if (unlikely(on_rq)) { 1282 ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ); 1283 1284 set_current_state(TASK_UNINTERRUPTIBLE); 1285 schedule_hrtimeout(&to, HRTIMER_MODE_REL); 1286 continue; 1287 } 1288 1289 /* 1290 * Ahh, all good. It wasn't running, and it wasn't 1291 * runnable, which means that it will never become 1292 * running in the future either. We're all done! 1293 */ 1294 break; 1295 } 1296 1297 return ncsw; 1298} 1299 1300/*** 1301 * kick_process - kick a running thread to enter/exit the kernel 1302 * @p: the to-be-kicked thread 1303 * 1304 * Cause a process which is running on another CPU to enter 1305 * kernel-mode, without any delay. (to get signals handled.) 1306 * 1307 * NOTE: this function doesn't have to take the runqueue lock, 1308 * because all it wants to ensure is that the remote task enters 1309 * the kernel. If the IPI races and the task has been migrated 1310 * to another CPU then no harm is done and the purpose has been 1311 * achieved as well. 1312 */ 1313void kick_process(struct task_struct *p) 1314{ 1315 int cpu; 1316 1317 preempt_disable(); 1318 cpu = task_cpu(p); 1319 if ((cpu != smp_processor_id()) && task_curr(p)) 1320 smp_send_reschedule(cpu); 1321 preempt_enable(); 1322} 1323EXPORT_SYMBOL_GPL(kick_process); 1324#endif /* CONFIG_SMP */ 1325 1326#ifdef CONFIG_SMP 1327/* 1328 * ->cpus_allowed is protected by both rq->lock and p->pi_lock 1329 */ 1330static int select_fallback_rq(int cpu, struct task_struct *p) 1331{ 1332 int nid = cpu_to_node(cpu); 1333 const struct cpumask *nodemask = NULL; 1334 enum { cpuset, possible, fail } state = cpuset; 1335 int dest_cpu; 1336 1337 /* 1338 * If the node that the cpu is on has been offlined, cpu_to_node() 1339 * will return -1. There is no cpu on the node, and we should 1340 * select the cpu on the other node. 1341 */ 1342 if (nid != -1) { 1343 nodemask = cpumask_of_node(nid); 1344 1345 /* Look for allowed, online CPU in same node. */ 1346 for_each_cpu(dest_cpu, nodemask) { 1347 if (!cpu_online(dest_cpu)) 1348 continue; 1349 if (!cpu_active(dest_cpu)) 1350 continue; 1351 if (cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p))) 1352 return dest_cpu; 1353 } 1354 } 1355 1356 for (;;) { 1357 /* Any allowed, online CPU? */ 1358 for_each_cpu(dest_cpu, tsk_cpus_allowed(p)) { 1359 if (!cpu_online(dest_cpu)) 1360 continue; 1361 if (!cpu_active(dest_cpu)) 1362 continue; 1363 goto out; 1364 } 1365 1366 switch (state) { 1367 case cpuset: 1368 /* No more Mr. Nice Guy. */ 1369 cpuset_cpus_allowed_fallback(p); 1370 state = possible; 1371 break; 1372 1373 case possible: 1374 do_set_cpus_allowed(p, cpu_possible_mask); 1375 state = fail; 1376 break; 1377 1378 case fail: 1379 BUG(); 1380 break; 1381 } 1382 } 1383 1384out: 1385 if (state != cpuset) { 1386 /* 1387 * Don't tell them about moving exiting tasks or 1388 * kernel threads (both mm NULL), since they never 1389 * leave kernel. 1390 */ 1391 if (p->mm && printk_ratelimit()) { 1392 printk_deferred("process %d (%s) no longer affine to cpu%d\n", 1393 task_pid_nr(p), p->comm, cpu); 1394 } 1395 } 1396 1397 return dest_cpu; 1398} 1399 1400/* 1401 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable. 1402 */ 1403static inline 1404int select_task_rq(struct task_struct *p, int cpu, int sd_flags, int wake_flags) 1405{ 1406 cpu = p->sched_class->select_task_rq(p, cpu, sd_flags, wake_flags); 1407 1408 /* 1409 * In order not to call set_task_cpu() on a blocking task we need 1410 * to rely on ttwu() to place the task on a valid ->cpus_allowed 1411 * cpu. 1412 * 1413 * Since this is common to all placement strategies, this lives here. 1414 * 1415 * [ this allows ->select_task() to simply return task_cpu(p) and 1416 * not worry about this generic constraint ] 1417 */ 1418 if (unlikely(!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)) || 1419 !cpu_online(cpu))) 1420 cpu = select_fallback_rq(task_cpu(p), p); 1421 1422 return cpu; 1423} 1424 1425static void update_avg(u64 *avg, u64 sample) 1426{ 1427 s64 diff = sample - *avg; 1428 *avg += diff >> 3; 1429} 1430#endif 1431 1432static void 1433ttwu_stat(struct task_struct *p, int cpu, int wake_flags) 1434{ 1435#ifdef CONFIG_SCHEDSTATS 1436 struct rq *rq = this_rq(); 1437 1438#ifdef CONFIG_SMP 1439 int this_cpu = smp_processor_id(); 1440 1441 if (cpu == this_cpu) { 1442 schedstat_inc(rq, ttwu_local); 1443 schedstat_inc(p, se.statistics.nr_wakeups_local); 1444 } else { 1445 struct sched_domain *sd; 1446 1447 schedstat_inc(p, se.statistics.nr_wakeups_remote); 1448 rcu_read_lock(); 1449 for_each_domain(this_cpu, sd) { 1450 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) { 1451 schedstat_inc(sd, ttwu_wake_remote); 1452 break; 1453 } 1454 } 1455 rcu_read_unlock(); 1456 } 1457 1458 if (wake_flags & WF_MIGRATED) 1459 schedstat_inc(p, se.statistics.nr_wakeups_migrate); 1460 1461#endif /* CONFIG_SMP */ 1462 1463 schedstat_inc(rq, ttwu_count); 1464 schedstat_inc(p, se.statistics.nr_wakeups); 1465 1466 if (wake_flags & WF_SYNC) 1467 schedstat_inc(p, se.statistics.nr_wakeups_sync); 1468 1469#endif /* CONFIG_SCHEDSTATS */ 1470} 1471 1472static void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags) 1473{ 1474 activate_task(rq, p, en_flags); 1475 p->on_rq = 1; 1476 1477 /* if a worker is waking up, notify workqueue */ 1478 if (p->flags & PF_WQ_WORKER) 1479 wq_worker_waking_up(p, cpu_of(rq)); 1480} 1481 1482/* 1483 * Mark the task runnable and perform wakeup-preemption. 1484 */ 1485static void 1486ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags) 1487{ 1488 check_preempt_curr(rq, p, wake_flags); 1489 trace_sched_wakeup(p, true); 1490 1491 p->state = TASK_RUNNING; 1492#ifdef CONFIG_SMP 1493 if (p->sched_class->task_woken) 1494 p->sched_class->task_woken(rq, p); 1495 1496 if (rq->idle_stamp) { 1497 u64 delta = rq_clock(rq) - rq->idle_stamp; 1498 u64 max = 2*rq->max_idle_balance_cost; 1499 1500 update_avg(&rq->avg_idle, delta); 1501 1502 if (rq->avg_idle > max) 1503 rq->avg_idle = max; 1504 1505 rq->idle_stamp = 0; 1506 } 1507#endif 1508} 1509 1510static void 1511ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags) 1512{ 1513#ifdef CONFIG_SMP 1514 if (p->sched_contributes_to_load) 1515 rq->nr_uninterruptible--; 1516#endif 1517 1518 ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_WAKING); 1519 ttwu_do_wakeup(rq, p, wake_flags); 1520} 1521 1522/* 1523 * Called in case the task @p isn't fully descheduled from its runqueue, 1524 * in this case we must do a remote wakeup. Its a 'light' wakeup though, 1525 * since all we need to do is flip p->state to TASK_RUNNING, since 1526 * the task is still ->on_rq. 1527 */ 1528static int ttwu_remote(struct task_struct *p, int wake_flags) 1529{ 1530 struct rq *rq; 1531 int ret = 0; 1532 1533 rq = __task_rq_lock(p); 1534 if (p->on_rq) { 1535 /* check_preempt_curr() may use rq clock */ 1536 update_rq_clock(rq); 1537 ttwu_do_wakeup(rq, p, wake_flags); 1538 ret = 1; 1539 } 1540 __task_rq_unlock(rq); 1541 1542 return ret; 1543} 1544 1545#ifdef CONFIG_SMP 1546void sched_ttwu_pending(void) 1547{ 1548 struct rq *rq = this_rq(); 1549 struct llist_node *llist = llist_del_all(&rq->wake_list); 1550 struct task_struct *p; 1551 unsigned long flags; 1552 1553 if (!llist) 1554 return; 1555 1556 raw_spin_lock_irqsave(&rq->lock, flags); 1557 1558 while (llist) { 1559 p = llist_entry(llist, struct task_struct, wake_entry); 1560 llist = llist_next(llist); 1561 ttwu_do_activate(rq, p, 0); 1562 } 1563 1564 raw_spin_unlock_irqrestore(&rq->lock, flags); 1565} 1566 1567void scheduler_ipi(void) 1568{ 1569 /* 1570 * Fold TIF_NEED_RESCHED into the preempt_count; anybody setting 1571 * TIF_NEED_RESCHED remotely (for the first time) will also send 1572 * this IPI. 1573 */ 1574 preempt_fold_need_resched(); 1575 1576 if (llist_empty(&this_rq()->wake_list) && !got_nohz_idle_kick()) 1577 return; 1578 1579 /* 1580 * Not all reschedule IPI handlers call irq_enter/irq_exit, since 1581 * traditionally all their work was done from the interrupt return 1582 * path. Now that we actually do some work, we need to make sure 1583 * we do call them. 1584 * 1585 * Some archs already do call them, luckily irq_enter/exit nest 1586 * properly. 1587 * 1588 * Arguably we should visit all archs and update all handlers, 1589 * however a fair share of IPIs are still resched only so this would 1590 * somewhat pessimize the simple resched case. 1591 */ 1592 irq_enter(); 1593 sched_ttwu_pending(); 1594 1595 /* 1596 * Check if someone kicked us for doing the nohz idle load balance. 1597 */ 1598 if (unlikely(got_nohz_idle_kick())) { 1599 this_rq()->idle_balance = 1; 1600 raise_softirq_irqoff(SCHED_SOFTIRQ); 1601 } 1602 irq_exit(); 1603} 1604 1605static void ttwu_queue_remote(struct task_struct *p, int cpu) 1606{ 1607 struct rq *rq = cpu_rq(cpu); 1608 1609 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list)) { 1610 if (!set_nr_if_polling(rq->idle)) 1611 smp_send_reschedule(cpu); 1612 else 1613 trace_sched_wake_idle_without_ipi(cpu); 1614 } 1615} 1616 1617bool cpus_share_cache(int this_cpu, int that_cpu) 1618{ 1619 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu); 1620} 1621#endif /* CONFIG_SMP */ 1622 1623static void ttwu_queue(struct task_struct *p, int cpu) 1624{ 1625 struct rq *rq = cpu_rq(cpu); 1626 1627#if defined(CONFIG_SMP) 1628 if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) { 1629 sched_clock_cpu(cpu); /* sync clocks x-cpu */ 1630 ttwu_queue_remote(p, cpu); 1631 return; 1632 } 1633#endif 1634 1635 raw_spin_lock(&rq->lock); 1636 ttwu_do_activate(rq, p, 0); 1637 raw_spin_unlock(&rq->lock); 1638} 1639 1640/** 1641 * try_to_wake_up - wake up a thread 1642 * @p: the thread to be awakened 1643 * @state: the mask of task states that can be woken 1644 * @wake_flags: wake modifier flags (WF_*) 1645 * 1646 * Put it on the run-queue if it's not already there. The "current" 1647 * thread is always on the run-queue (except when the actual 1648 * re-schedule is in progress), and as such you're allowed to do 1649 * the simpler "current->state = TASK_RUNNING" to mark yourself 1650 * runnable without the overhead of this. 1651 * 1652 * Return: %true if @p was woken up, %false if it was already running. 1653 * or @state didn't match @p's state. 1654 */ 1655static int 1656try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags) 1657{ 1658 unsigned long flags; 1659 int cpu, success = 0; 1660 1661 /* 1662 * If we are going to wake up a thread waiting for CONDITION we 1663 * need to ensure that CONDITION=1 done by the caller can not be 1664 * reordered with p->state check below. This pairs with mb() in 1665 * set_current_state() the waiting thread does. 1666 */ 1667 smp_mb__before_spinlock(); 1668 raw_spin_lock_irqsave(&p->pi_lock, flags); 1669 if (!(p->state & state)) 1670 goto out; 1671 1672 success = 1; /* we're going to change ->state */ 1673 cpu = task_cpu(p); 1674 1675 if (p->on_rq && ttwu_remote(p, wake_flags)) 1676 goto stat; 1677 1678#ifdef CONFIG_SMP 1679 /* 1680 * If the owning (remote) cpu is still in the middle of schedule() with 1681 * this task as prev, wait until its done referencing the task. 1682 */ 1683 while (p->on_cpu) 1684 cpu_relax(); 1685 /* 1686 * Pairs with the smp_wmb() in finish_lock_switch(). 1687 */ 1688 smp_rmb(); 1689 1690 p->sched_contributes_to_load = !!task_contributes_to_load(p); 1691 p->state = TASK_WAKING; 1692 1693 if (p->sched_class->task_waking) 1694 p->sched_class->task_waking(p); 1695 1696 cpu = select_task_rq(p, p->wake_cpu, SD_BALANCE_WAKE, wake_flags); 1697 if (task_cpu(p) != cpu) { 1698 wake_flags |= WF_MIGRATED; 1699 set_task_cpu(p, cpu); 1700 } 1701#endif /* CONFIG_SMP */ 1702 1703 ttwu_queue(p, cpu); 1704stat: 1705 ttwu_stat(p, cpu, wake_flags); 1706out: 1707 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 1708 1709 return success; 1710} 1711 1712/** 1713 * try_to_wake_up_local - try to wake up a local task with rq lock held 1714 * @p: the thread to be awakened 1715 * 1716 * Put @p on the run-queue if it's not already there. The caller must 1717 * ensure that this_rq() is locked, @p is bound to this_rq() and not 1718 * the current task. 1719 */ 1720static void try_to_wake_up_local(struct task_struct *p) 1721{ 1722 struct rq *rq = task_rq(p); 1723 1724 if (WARN_ON_ONCE(rq != this_rq()) || 1725 WARN_ON_ONCE(p == current)) 1726 return; 1727 1728 lockdep_assert_held(&rq->lock); 1729 1730 if (!raw_spin_trylock(&p->pi_lock)) { 1731 raw_spin_unlock(&rq->lock); 1732 raw_spin_lock(&p->pi_lock); 1733 raw_spin_lock(&rq->lock); 1734 } 1735 1736 if (!(p->state & TASK_NORMAL)) 1737 goto out; 1738 1739 if (!p->on_rq) 1740 ttwu_activate(rq, p, ENQUEUE_WAKEUP); 1741 1742 ttwu_do_wakeup(rq, p, 0); 1743 ttwu_stat(p, smp_processor_id(), 0); 1744out: 1745 raw_spin_unlock(&p->pi_lock); 1746} 1747 1748/** 1749 * wake_up_process - Wake up a specific process 1750 * @p: The process to be woken up. 1751 * 1752 * Attempt to wake up the nominated process and move it to the set of runnable 1753 * processes. 1754 * 1755 * Return: 1 if the process was woken up, 0 if it was already running. 1756 * 1757 * It may be assumed that this function implies a write memory barrier before 1758 * changing the task state if and only if any tasks are woken up. 1759 */ 1760int wake_up_process(struct task_struct *p) 1761{ 1762 WARN_ON(task_is_stopped_or_traced(p)); 1763 return try_to_wake_up(p, TASK_NORMAL, 0); 1764} 1765EXPORT_SYMBOL(wake_up_process); 1766 1767int wake_up_state(struct task_struct *p, unsigned int state) 1768{ 1769 return try_to_wake_up(p, state, 0); 1770} 1771 1772/* 1773 * Perform scheduler related setup for a newly forked process p. 1774 * p is forked by current. 1775 * 1776 * __sched_fork() is basic setup used by init_idle() too: 1777 */ 1778static void __sched_fork(unsigned long clone_flags, struct task_struct *p) 1779{ 1780 p->on_rq = 0; 1781 1782 p->se.on_rq = 0; 1783 p->se.exec_start = 0; 1784 p->se.sum_exec_runtime = 0; 1785 p->se.prev_sum_exec_runtime = 0; 1786 p->se.nr_migrations = 0; 1787 p->se.vruntime = 0; 1788 INIT_LIST_HEAD(&p->se.group_node); 1789 1790#ifdef CONFIG_SCHEDSTATS 1791 memset(&p->se.statistics, 0, sizeof(p->se.statistics)); 1792#endif 1793 1794 RB_CLEAR_NODE(&p->dl.rb_node); 1795 hrtimer_init(&p->dl.dl_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); 1796 p->dl.dl_runtime = p->dl.runtime = 0; 1797 p->dl.dl_deadline = p->dl.deadline = 0; 1798 p->dl.dl_period = 0; 1799 p->dl.flags = 0; 1800 1801 INIT_LIST_HEAD(&p->rt.run_list); 1802 1803#ifdef CONFIG_PREEMPT_NOTIFIERS 1804 INIT_HLIST_HEAD(&p->preempt_notifiers); 1805#endif 1806 1807#ifdef CONFIG_NUMA_BALANCING 1808 if (p->mm && atomic_read(&p->mm->mm_users) == 1) { 1809 p->mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay); 1810 p->mm->numa_scan_seq = 0; 1811 } 1812 1813 if (clone_flags & CLONE_VM) 1814 p->numa_preferred_nid = current->numa_preferred_nid; 1815 else 1816 p->numa_preferred_nid = -1; 1817 1818 p->node_stamp = 0ULL; 1819 p->numa_scan_seq = p->mm ? p->mm->numa_scan_seq : 0; 1820 p->numa_scan_period = sysctl_numa_balancing_scan_delay; 1821 p->numa_work.next = &p->numa_work; 1822 p->numa_faults_memory = NULL; 1823 p->numa_faults_buffer_memory = NULL; 1824 p->last_task_numa_placement = 0; 1825 p->last_sum_exec_runtime = 0; 1826 1827 INIT_LIST_HEAD(&p->numa_entry); 1828 p->numa_group = NULL; 1829#endif /* CONFIG_NUMA_BALANCING */ 1830} 1831 1832#ifdef CONFIG_NUMA_BALANCING 1833#ifdef CONFIG_SCHED_DEBUG 1834void set_numabalancing_state(bool enabled) 1835{ 1836 if (enabled) 1837 sched_feat_set("NUMA"); 1838 else 1839 sched_feat_set("NO_NUMA"); 1840} 1841#else 1842__read_mostly bool numabalancing_enabled; 1843 1844void set_numabalancing_state(bool enabled) 1845{ 1846 numabalancing_enabled = enabled; 1847} 1848#endif /* CONFIG_SCHED_DEBUG */ 1849 1850#ifdef CONFIG_PROC_SYSCTL 1851int sysctl_numa_balancing(struct ctl_table *table, int write, 1852 void __user *buffer, size_t *lenp, loff_t *ppos) 1853{ 1854 struct ctl_table t; 1855 int err; 1856 int state = numabalancing_enabled; 1857 1858 if (write && !capable(CAP_SYS_ADMIN)) 1859 return -EPERM; 1860 1861 t = *table; 1862 t.data = &state; 1863 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos); 1864 if (err < 0) 1865 return err; 1866 if (write) 1867 set_numabalancing_state(state); 1868 return err; 1869} 1870#endif 1871#endif 1872 1873/* 1874 * fork()/clone()-time setup: 1875 */ 1876int sched_fork(unsigned long clone_flags, struct task_struct *p) 1877{ 1878 unsigned long flags; 1879 int cpu = get_cpu(); 1880 1881 __sched_fork(clone_flags, p); 1882 /* 1883 * We mark the process as running here. This guarantees that 1884 * nobody will actually run it, and a signal or other external 1885 * event cannot wake it up and insert it on the runqueue either. 1886 */ 1887 p->state = TASK_RUNNING; 1888 1889 /* 1890 * Make sure we do not leak PI boosting priority to the child. 1891 */ 1892 p->prio = current->normal_prio; 1893 1894 /* 1895 * Revert to default priority/policy on fork if requested. 1896 */ 1897 if (unlikely(p->sched_reset_on_fork)) { 1898 if (task_has_dl_policy(p) || task_has_rt_policy(p)) { 1899 p->policy = SCHED_NORMAL; 1900 p->static_prio = NICE_TO_PRIO(0); 1901 p->rt_priority = 0; 1902 } else if (PRIO_TO_NICE(p->static_prio) < 0) 1903 p->static_prio = NICE_TO_PRIO(0); 1904 1905 p->prio = p->normal_prio = __normal_prio(p); 1906 set_load_weight(p); 1907 1908 /* 1909 * We don't need the reset flag anymore after the fork. It has 1910 * fulfilled its duty: 1911 */ 1912 p->sched_reset_on_fork = 0; 1913 } 1914 1915 if (dl_prio(p->prio)) { 1916 put_cpu(); 1917 return -EAGAIN; 1918 } else if (rt_prio(p->prio)) { 1919 p->sched_class = &rt_sched_class; 1920 } else { 1921 p->sched_class = &fair_sched_class; 1922 } 1923 1924 if (p->sched_class->task_fork) 1925 p->sched_class->task_fork(p); 1926 1927 /* 1928 * The child is not yet in the pid-hash so no cgroup attach races, 1929 * and the cgroup is pinned to this child due to cgroup_fork() 1930 * is ran before sched_fork(). 1931 * 1932 * Silence PROVE_RCU. 1933 */ 1934 raw_spin_lock_irqsave(&p->pi_lock, flags); 1935 set_task_cpu(p, cpu); 1936 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 1937 1938#if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT) 1939 if (likely(sched_info_on())) 1940 memset(&p->sched_info, 0, sizeof(p->sched_info)); 1941#endif 1942#if defined(CONFIG_SMP) 1943 p->on_cpu = 0; 1944#endif 1945 init_task_preempt_count(p); 1946#ifdef CONFIG_SMP 1947 plist_node_init(&p->pushable_tasks, MAX_PRIO); 1948 RB_CLEAR_NODE(&p->pushable_dl_tasks); 1949#endif 1950 1951 put_cpu(); 1952 return 0; 1953} 1954 1955unsigned long to_ratio(u64 period, u64 runtime) 1956{ 1957 if (runtime == RUNTIME_INF) 1958 return 1ULL << 20; 1959 1960 /* 1961 * Doing this here saves a lot of checks in all 1962 * the calling paths, and returning zero seems 1963 * safe for them anyway. 1964 */ 1965 if (period == 0) 1966 return 0; 1967 1968 return div64_u64(runtime << 20, period); 1969} 1970 1971#ifdef CONFIG_SMP 1972inline struct dl_bw *dl_bw_of(int i) 1973{ 1974 return &cpu_rq(i)->rd->dl_bw; 1975} 1976 1977static inline int dl_bw_cpus(int i) 1978{ 1979 struct root_domain *rd = cpu_rq(i)->rd; 1980 int cpus = 0; 1981 1982 for_each_cpu_and(i, rd->span, cpu_active_mask) 1983 cpus++; 1984 1985 return cpus; 1986} 1987#else 1988inline struct dl_bw *dl_bw_of(int i) 1989{ 1990 return &cpu_rq(i)->dl.dl_bw; 1991} 1992 1993static inline int dl_bw_cpus(int i) 1994{ 1995 return 1; 1996} 1997#endif 1998 1999static inline 2000void __dl_clear(struct dl_bw *dl_b, u64 tsk_bw) 2001{ 2002 dl_b->total_bw -= tsk_bw; 2003} 2004 2005static inline 2006void __dl_add(struct dl_bw *dl_b, u64 tsk_bw) 2007{ 2008 dl_b->total_bw += tsk_bw; 2009} 2010 2011static inline 2012bool __dl_overflow(struct dl_bw *dl_b, int cpus, u64 old_bw, u64 new_bw) 2013{ 2014 return dl_b->bw != -1 && 2015 dl_b->bw * cpus < dl_b->total_bw - old_bw + new_bw; 2016} 2017 2018/* 2019 * We must be sure that accepting a new task (or allowing changing the 2020 * parameters of an existing one) is consistent with the bandwidth 2021 * constraints. If yes, this function also accordingly updates the currently 2022 * allocated bandwidth to reflect the new situation. 2023 * 2024 * This function is called while holding p's rq->lock. 2025 */ 2026static int dl_overflow(struct task_struct *p, int policy, 2027 const struct sched_attr *attr) 2028{ 2029 2030 struct dl_bw *dl_b = dl_bw_of(task_cpu(p)); 2031 u64 period = attr->sched_period ?: attr->sched_deadline; 2032 u64 runtime = attr->sched_runtime; 2033 u64 new_bw = dl_policy(policy) ? to_ratio(period, runtime) : 0; 2034 int cpus, err = -1; 2035 2036 if (new_bw == p->dl.dl_bw) 2037 return 0; 2038 2039 /* 2040 * Either if a task, enters, leave, or stays -deadline but changes 2041 * its parameters, we may need to update accordingly the total 2042 * allocated bandwidth of the container. 2043 */ 2044 raw_spin_lock(&dl_b->lock); 2045 cpus = dl_bw_cpus(task_cpu(p)); 2046 if (dl_policy(policy) && !task_has_dl_policy(p) && 2047 !__dl_overflow(dl_b, cpus, 0, new_bw)) { 2048 __dl_add(dl_b, new_bw); 2049 err = 0; 2050 } else if (dl_policy(policy) && task_has_dl_policy(p) && 2051 !__dl_overflow(dl_b, cpus, p->dl.dl_bw, new_bw)) { 2052 __dl_clear(dl_b, p->dl.dl_bw); 2053 __dl_add(dl_b, new_bw); 2054 err = 0; 2055 } else if (!dl_policy(policy) && task_has_dl_policy(p)) { 2056 __dl_clear(dl_b, p->dl.dl_bw); 2057 err = 0; 2058 } 2059 raw_spin_unlock(&dl_b->lock); 2060 2061 return err; 2062} 2063 2064extern void init_dl_bw(struct dl_bw *dl_b); 2065 2066/* 2067 * wake_up_new_task - wake up a newly created task for the first time. 2068 * 2069 * This function will do some initial scheduler statistics housekeeping 2070 * that must be done for every newly created context, then puts the task 2071 * on the runqueue and wakes it. 2072 */ 2073void wake_up_new_task(struct task_struct *p) 2074{ 2075 unsigned long flags; 2076 struct rq *rq; 2077 2078 raw_spin_lock_irqsave(&p->pi_lock, flags); 2079#ifdef CONFIG_SMP 2080 /* 2081 * Fork balancing, do it here and not earlier because: 2082 * - cpus_allowed can change in the fork path 2083 * - any previously selected cpu might disappear through hotplug 2084 */ 2085 set_task_cpu(p, select_task_rq(p, task_cpu(p), SD_BALANCE_FORK, 0)); 2086#endif 2087 2088 /* Initialize new task's runnable average */ 2089 init_task_runnable_average(p); 2090 rq = __task_rq_lock(p); 2091 activate_task(rq, p, 0); 2092 p->on_rq = 1; 2093 trace_sched_wakeup_new(p, true); 2094 check_preempt_curr(rq, p, WF_FORK); 2095#ifdef CONFIG_SMP 2096 if (p->sched_class->task_woken) 2097 p->sched_class->task_woken(rq, p); 2098#endif 2099 task_rq_unlock(rq, p, &flags); 2100} 2101 2102#ifdef CONFIG_PREEMPT_NOTIFIERS 2103 2104/** 2105 * preempt_notifier_register - tell me when current is being preempted & rescheduled 2106 * @notifier: notifier struct to register 2107 */ 2108void preempt_notifier_register(struct preempt_notifier *notifier) 2109{ 2110 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers); 2111} 2112EXPORT_SYMBOL_GPL(preempt_notifier_register); 2113 2114/** 2115 * preempt_notifier_unregister - no longer interested in preemption notifications 2116 * @notifier: notifier struct to unregister 2117 * 2118 * This is safe to call from within a preemption notifier. 2119 */ 2120void preempt_notifier_unregister(struct preempt_notifier *notifier) 2121{ 2122 hlist_del(¬ifier->link); 2123} 2124EXPORT_SYMBOL_GPL(preempt_notifier_unregister); 2125 2126static void fire_sched_in_preempt_notifiers(struct task_struct *curr) 2127{ 2128 struct preempt_notifier *notifier; 2129 2130 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link) 2131 notifier->ops->sched_in(notifier, raw_smp_processor_id()); 2132} 2133 2134static void 2135fire_sched_out_preempt_notifiers(struct task_struct *curr, 2136 struct task_struct *next) 2137{ 2138 struct preempt_notifier *notifier; 2139 2140 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link) 2141 notifier->ops->sched_out(notifier, next); 2142} 2143 2144#else /* !CONFIG_PREEMPT_NOTIFIERS */ 2145 2146static void fire_sched_in_preempt_notifiers(struct task_struct *curr) 2147{ 2148} 2149 2150static void 2151fire_sched_out_preempt_notifiers(struct task_struct *curr, 2152 struct task_struct *next) 2153{ 2154} 2155 2156#endif /* CONFIG_PREEMPT_NOTIFIERS */ 2157 2158/** 2159 * prepare_task_switch - prepare to switch tasks 2160 * @rq: the runqueue preparing to switch 2161 * @prev: the current task that is being switched out 2162 * @next: the task we are going to switch to. 2163 * 2164 * This is called with the rq lock held and interrupts off. It must 2165 * be paired with a subsequent finish_task_switch after the context 2166 * switch. 2167 * 2168 * prepare_task_switch sets up locking and calls architecture specific 2169 * hooks. 2170 */ 2171static inline void 2172prepare_task_switch(struct rq *rq, struct task_struct *prev, 2173 struct task_struct *next) 2174{ 2175 trace_sched_switch(prev, next); 2176 sched_info_switch(rq, prev, next); 2177 perf_event_task_sched_out(prev, next); 2178 fire_sched_out_preempt_notifiers(prev, next); 2179 prepare_lock_switch(rq, next); 2180 prepare_arch_switch(next); 2181} 2182 2183/** 2184 * finish_task_switch - clean up after a task-switch 2185 * @rq: runqueue associated with task-switch 2186 * @prev: the thread we just switched away from. 2187 * 2188 * finish_task_switch must be called after the context switch, paired 2189 * with a prepare_task_switch call before the context switch. 2190 * finish_task_switch will reconcile locking set up by prepare_task_switch, 2191 * and do any other architecture-specific cleanup actions. 2192 * 2193 * Note that we may have delayed dropping an mm in context_switch(). If 2194 * so, we finish that here outside of the runqueue lock. (Doing it 2195 * with the lock held can cause deadlocks; see schedule() for 2196 * details.) 2197 */ 2198static void finish_task_switch(struct rq *rq, struct task_struct *prev) 2199 __releases(rq->lock) 2200{ 2201 struct mm_struct *mm = rq->prev_mm; 2202 long prev_state; 2203 2204 rq->prev_mm = NULL; 2205 2206 /* 2207 * A task struct has one reference for the use as "current". 2208 * If a task dies, then it sets TASK_DEAD in tsk->state and calls 2209 * schedule one last time. The schedule call will never return, and 2210 * the scheduled task must drop that reference. 2211 * The test for TASK_DEAD must occur while the runqueue locks are 2212 * still held, otherwise prev could be scheduled on another cpu, die 2213 * there before we look at prev->state, and then the reference would 2214 * be dropped twice. 2215 * Manfred Spraul <manfred@colorfullife.com> 2216 */ 2217 prev_state = prev->state; 2218 vtime_task_switch(prev); 2219 finish_arch_switch(prev); 2220 perf_event_task_sched_in(prev, current); 2221 finish_lock_switch(rq, prev); 2222 finish_arch_post_lock_switch(); 2223 2224 fire_sched_in_preempt_notifiers(current); 2225 if (mm) 2226 mmdrop(mm); 2227 if (unlikely(prev_state == TASK_DEAD)) { 2228 if (prev->sched_class->task_dead) 2229 prev->sched_class->task_dead(prev); 2230 2231 /* 2232 * Remove function-return probe instances associated with this 2233 * task and put them back on the free list. 2234 */ 2235 kprobe_flush_task(prev); 2236 put_task_struct(prev); 2237 } 2238 2239 tick_nohz_task_switch(current); 2240} 2241 2242#ifdef CONFIG_SMP 2243 2244/* rq->lock is NOT held, but preemption is disabled */ 2245static inline void post_schedule(struct rq *rq) 2246{ 2247 if (rq->post_schedule) { 2248 unsigned long flags; 2249 2250 raw_spin_lock_irqsave(&rq->lock, flags); 2251 if (rq->curr->sched_class->post_schedule) 2252 rq->curr->sched_class->post_schedule(rq); 2253 raw_spin_unlock_irqrestore(&rq->lock, flags); 2254 2255 rq->post_schedule = 0; 2256 } 2257} 2258 2259#else 2260 2261static inline void post_schedule(struct rq *rq) 2262{ 2263} 2264 2265#endif 2266 2267/** 2268 * schedule_tail - first thing a freshly forked thread must call. 2269 * @prev: the thread we just switched away from. 2270 */ 2271asmlinkage __visible void schedule_tail(struct task_struct *prev) 2272 __releases(rq->lock) 2273{ 2274 struct rq *rq = this_rq(); 2275 2276 finish_task_switch(rq, prev); 2277 2278 /* 2279 * FIXME: do we need to worry about rq being invalidated by the 2280 * task_switch? 2281 */ 2282 post_schedule(rq); 2283 2284#ifdef __ARCH_WANT_UNLOCKED_CTXSW 2285 /* In this case, finish_task_switch does not reenable preemption */ 2286 preempt_enable(); 2287#endif 2288 if (current->set_child_tid) 2289 put_user(task_pid_vnr(current), current->set_child_tid); 2290} 2291 2292/* 2293 * context_switch - switch to the new MM and the new 2294 * thread's register state. 2295 */ 2296static inline void 2297context_switch(struct rq *rq, struct task_struct *prev, 2298 struct task_struct *next) 2299{ 2300 struct mm_struct *mm, *oldmm; 2301 2302 prepare_task_switch(rq, prev, next); 2303 2304 mm = next->mm; 2305 oldmm = prev->active_mm; 2306 /* 2307 * For paravirt, this is coupled with an exit in switch_to to 2308 * combine the page table reload and the switch backend into 2309 * one hypercall. 2310 */ 2311 arch_start_context_switch(prev); 2312 2313 if (!mm) { 2314 next->active_mm = oldmm; 2315 atomic_inc(&oldmm->mm_count); 2316 enter_lazy_tlb(oldmm, next); 2317 } else 2318 switch_mm(oldmm, mm, next); 2319 2320 if (!prev->mm) { 2321 prev->active_mm = NULL; 2322 rq->prev_mm = oldmm; 2323 } 2324 /* 2325 * Since the runqueue lock will be released by the next 2326 * task (which is an invalid locking op but in the case 2327 * of the scheduler it's an obvious special-case), so we 2328 * do an early lockdep release here: 2329 */ 2330#ifndef __ARCH_WANT_UNLOCKED_CTXSW 2331 spin_release(&rq->lock.dep_map, 1, _THIS_IP_); 2332#endif 2333 2334 context_tracking_task_switch(prev, next); 2335 /* Here we just switch the register state and the stack. */ 2336 switch_to(prev, next, prev); 2337 2338 barrier(); 2339 /* 2340 * this_rq must be evaluated again because prev may have moved 2341 * CPUs since it called schedule(), thus the 'rq' on its stack 2342 * frame will be invalid. 2343 */ 2344 finish_task_switch(this_rq(), prev); 2345} 2346 2347/* 2348 * nr_running and nr_context_switches: 2349 * 2350 * externally visible scheduler statistics: current number of runnable 2351 * threads, total number of context switches performed since bootup. 2352 */ 2353unsigned long nr_running(void) 2354{ 2355 unsigned long i, sum = 0; 2356 2357 for_each_online_cpu(i) 2358 sum += cpu_rq(i)->nr_running; 2359 2360 return sum; 2361} 2362 2363unsigned long long nr_context_switches(void) 2364{ 2365 int i; 2366 unsigned long long sum = 0; 2367 2368 for_each_possible_cpu(i) 2369 sum += cpu_rq(i)->nr_switches; 2370 2371 return sum; 2372} 2373 2374unsigned long nr_iowait(void) 2375{ 2376 unsigned long i, sum = 0; 2377 2378 for_each_possible_cpu(i) 2379 sum += atomic_read(&cpu_rq(i)->nr_iowait); 2380 2381 return sum; 2382} 2383 2384unsigned long nr_iowait_cpu(int cpu) 2385{ 2386 struct rq *this = cpu_rq(cpu); 2387 return atomic_read(&this->nr_iowait); 2388} 2389 2390#ifdef CONFIG_SMP 2391 2392/* 2393 * sched_exec - execve() is a valuable balancing opportunity, because at 2394 * this point the task has the smallest effective memory and cache footprint. 2395 */ 2396void sched_exec(void) 2397{ 2398 struct task_struct *p = current; 2399 unsigned long flags; 2400 int dest_cpu; 2401 2402 raw_spin_lock_irqsave(&p->pi_lock, flags); 2403 dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), SD_BALANCE_EXEC, 0); 2404 if (dest_cpu == smp_processor_id()) 2405 goto unlock; 2406 2407 if (likely(cpu_active(dest_cpu))) { 2408 struct migration_arg arg = { p, dest_cpu }; 2409 2410 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 2411 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg); 2412 return; 2413 } 2414unlock: 2415 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 2416} 2417 2418#endif 2419 2420DEFINE_PER_CPU(struct kernel_stat, kstat); 2421DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat); 2422 2423EXPORT_PER_CPU_SYMBOL(kstat); 2424EXPORT_PER_CPU_SYMBOL(kernel_cpustat); 2425 2426/* 2427 * Return any ns on the sched_clock that have not yet been accounted in 2428 * @p in case that task is currently running. 2429 * 2430 * Called with task_rq_lock() held on @rq. 2431 */ 2432static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq) 2433{ 2434 u64 ns = 0; 2435 2436 /* 2437 * Must be ->curr _and_ ->on_rq. If dequeued, we would 2438 * project cycles that may never be accounted to this 2439 * thread, breaking clock_gettime(). 2440 */ 2441 if (task_current(rq, p) && p->on_rq) { 2442 update_rq_clock(rq); 2443 ns = rq_clock_task(rq) - p->se.exec_start; 2444 if ((s64)ns < 0) 2445 ns = 0; 2446 } 2447 2448 return ns; 2449} 2450 2451unsigned long long task_delta_exec(struct task_struct *p) 2452{ 2453 unsigned long flags; 2454 struct rq *rq; 2455 u64 ns = 0; 2456 2457 rq = task_rq_lock(p, &flags); 2458 ns = do_task_delta_exec(p, rq); 2459 task_rq_unlock(rq, p, &flags); 2460 2461 return ns; 2462} 2463 2464/* 2465 * Return accounted runtime for the task. 2466 * In case the task is currently running, return the runtime plus current's 2467 * pending runtime that have not been accounted yet. 2468 */ 2469unsigned long long task_sched_runtime(struct task_struct *p) 2470{ 2471 unsigned long flags; 2472 struct rq *rq; 2473 u64 ns = 0; 2474 2475#if defined(CONFIG_64BIT) && defined(CONFIG_SMP) 2476 /* 2477 * 64-bit doesn't need locks to atomically read a 64bit value. 2478 * So we have a optimization chance when the task's delta_exec is 0. 2479 * Reading ->on_cpu is racy, but this is ok. 2480 * 2481 * If we race with it leaving cpu, we'll take a lock. So we're correct. 2482 * If we race with it entering cpu, unaccounted time is 0. This is 2483 * indistinguishable from the read occurring a few cycles earlier. 2484 * If we see ->on_cpu without ->on_rq, the task is leaving, and has 2485 * been accounted, so we're correct here as well. 2486 */ 2487 if (!p->on_cpu || !p->on_rq) 2488 return p->se.sum_exec_runtime; 2489#endif 2490 2491 rq = task_rq_lock(p, &flags); 2492 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq); 2493 task_rq_unlock(rq, p, &flags); 2494 2495 return ns; 2496} 2497 2498/* 2499 * This function gets called by the timer code, with HZ frequency. 2500 * We call it with interrupts disabled. 2501 */ 2502void scheduler_tick(void) 2503{ 2504 int cpu = smp_processor_id(); 2505 struct rq *rq = cpu_rq(cpu); 2506 struct task_struct *curr = rq->curr; 2507 2508 sched_clock_tick(); 2509 2510 raw_spin_lock(&rq->lock); 2511 update_rq_clock(rq); 2512 curr->sched_class->task_tick(rq, curr, 0); 2513 update_cpu_load_active(rq); 2514 raw_spin_unlock(&rq->lock); 2515 2516 perf_event_task_tick(); 2517 2518#ifdef CONFIG_SMP 2519 rq->idle_balance = idle_cpu(cpu); 2520 trigger_load_balance(rq); 2521#endif 2522 rq_last_tick_reset(rq); 2523} 2524 2525#ifdef CONFIG_NO_HZ_FULL 2526/** 2527 * scheduler_tick_max_deferment 2528 * 2529 * Keep at least one tick per second when a single 2530 * active task is running because the scheduler doesn't 2531 * yet completely support full dynticks environment. 2532 * 2533 * This makes sure that uptime, CFS vruntime, load 2534 * balancing, etc... continue to move forward, even 2535 * with a very low granularity. 2536 * 2537 * Return: Maximum deferment in nanoseconds. 2538 */ 2539u64 scheduler_tick_max_deferment(void) 2540{ 2541 struct rq *rq = this_rq(); 2542 unsigned long next, now = ACCESS_ONCE(jiffies); 2543 2544 next = rq->last_sched_tick + HZ; 2545 2546 if (time_before_eq(next, now)) 2547 return 0; 2548 2549 return jiffies_to_nsecs(next - now); 2550} 2551#endif 2552 2553notrace unsigned long get_parent_ip(unsigned long addr) 2554{ 2555 if (in_lock_functions(addr)) { 2556 addr = CALLER_ADDR2; 2557 if (in_lock_functions(addr)) 2558 addr = CALLER_ADDR3; 2559 } 2560 return addr; 2561} 2562 2563#if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \ 2564 defined(CONFIG_PREEMPT_TRACER)) 2565 2566void preempt_count_add(int val) 2567{ 2568#ifdef CONFIG_DEBUG_PREEMPT 2569 /* 2570 * Underflow? 2571 */ 2572 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0))) 2573 return; 2574#endif 2575 __preempt_count_add(val); 2576#ifdef CONFIG_DEBUG_PREEMPT 2577 /* 2578 * Spinlock count overflowing soon? 2579 */ 2580 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >= 2581 PREEMPT_MASK - 10); 2582#endif 2583 if (preempt_count() == val) { 2584 unsigned long ip = get_parent_ip(CALLER_ADDR1); 2585#ifdef CONFIG_DEBUG_PREEMPT 2586 current->preempt_disable_ip = ip; 2587#endif 2588 trace_preempt_off(CALLER_ADDR0, ip); 2589 } 2590} 2591EXPORT_SYMBOL(preempt_count_add); 2592NOKPROBE_SYMBOL(preempt_count_add); 2593 2594void preempt_count_sub(int val) 2595{ 2596#ifdef CONFIG_DEBUG_PREEMPT 2597 /* 2598 * Underflow? 2599 */ 2600 if (DEBUG_LOCKS_WARN_ON(val > preempt_count())) 2601 return; 2602 /* 2603 * Is the spinlock portion underflowing? 2604 */ 2605 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) && 2606 !(preempt_count() & PREEMPT_MASK))) 2607 return; 2608#endif 2609 2610 if (preempt_count() == val) 2611 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1)); 2612 __preempt_count_sub(val); 2613} 2614EXPORT_SYMBOL(preempt_count_sub); 2615NOKPROBE_SYMBOL(preempt_count_sub); 2616 2617#endif 2618 2619/* 2620 * Print scheduling while atomic bug: 2621 */ 2622static noinline void __schedule_bug(struct task_struct *prev) 2623{ 2624 if (oops_in_progress) 2625 return; 2626 2627 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n", 2628 prev->comm, prev->pid, preempt_count()); 2629 2630 debug_show_held_locks(prev); 2631 print_modules(); 2632 if (irqs_disabled()) 2633 print_irqtrace_events(prev); 2634#ifdef CONFIG_DEBUG_PREEMPT 2635 if (in_atomic_preempt_off()) { 2636 pr_err("Preemption disabled at:"); 2637 print_ip_sym(current->preempt_disable_ip); 2638 pr_cont("\n"); 2639 } 2640#endif 2641 dump_stack(); 2642 add_taint(TAINT_WARN, LOCKDEP_STILL_OK); 2643} 2644 2645/* 2646 * Various schedule()-time debugging checks and statistics: 2647 */ 2648static inline void schedule_debug(struct task_struct *prev) 2649{ 2650 /* 2651 * Test if we are atomic. Since do_exit() needs to call into 2652 * schedule() atomically, we ignore that path. Otherwise whine 2653 * if we are scheduling when we should not. 2654 */ 2655 if (unlikely(in_atomic_preempt_off() && prev->state != TASK_DEAD)) 2656 __schedule_bug(prev); 2657 rcu_sleep_check(); 2658 2659 profile_hit(SCHED_PROFILING, __builtin_return_address(0)); 2660 2661 schedstat_inc(this_rq(), sched_count); 2662} 2663 2664/* 2665 * Pick up the highest-prio task: 2666 */ 2667static inline struct task_struct * 2668pick_next_task(struct rq *rq, struct task_struct *prev) 2669{ 2670 const struct sched_class *class = &fair_sched_class; 2671 struct task_struct *p; 2672 2673 /* 2674 * Optimization: we know that if all tasks are in 2675 * the fair class we can call that function directly: 2676 */ 2677 if (likely(prev->sched_class == class && 2678 rq->nr_running == rq->cfs.h_nr_running)) { 2679 p = fair_sched_class.pick_next_task(rq, prev); 2680 if (unlikely(p == RETRY_TASK)) 2681 goto again; 2682 2683 /* assumes fair_sched_class->next == idle_sched_class */ 2684 if (unlikely(!p)) 2685 p = idle_sched_class.pick_next_task(rq, prev); 2686 2687 return p; 2688 } 2689 2690again: 2691 for_each_class(class) { 2692 p = class->pick_next_task(rq, prev); 2693 if (p) { 2694 if (unlikely(p == RETRY_TASK)) 2695 goto again; 2696 return p; 2697 } 2698 } 2699 2700 BUG(); /* the idle class will always have a runnable task */ 2701} 2702 2703/* 2704 * __schedule() is the main scheduler function. 2705 * 2706 * The main means of driving the scheduler and thus entering this function are: 2707 * 2708 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc. 2709 * 2710 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return 2711 * paths. For example, see arch/x86/entry_64.S. 2712 * 2713 * To drive preemption between tasks, the scheduler sets the flag in timer 2714 * interrupt handler scheduler_tick(). 2715 * 2716 * 3. Wakeups don't really cause entry into schedule(). They add a 2717 * task to the run-queue and that's it. 2718 * 2719 * Now, if the new task added to the run-queue preempts the current 2720 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets 2721 * called on the nearest possible occasion: 2722 * 2723 * - If the kernel is preemptible (CONFIG_PREEMPT=y): 2724 * 2725 * - in syscall or exception context, at the next outmost 2726 * preempt_enable(). (this might be as soon as the wake_up()'s 2727 * spin_unlock()!) 2728 * 2729 * - in IRQ context, return from interrupt-handler to 2730 * preemptible context 2731 * 2732 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set) 2733 * then at the next: 2734 * 2735 * - cond_resched() call 2736 * - explicit schedule() call 2737 * - return from syscall or exception to user-space 2738 * - return from interrupt-handler to user-space 2739 */ 2740static void __sched __schedule(void) 2741{ 2742 struct task_struct *prev, *next; 2743 unsigned long *switch_count; 2744 struct rq *rq; 2745 int cpu; 2746 2747need_resched: 2748 preempt_disable(); 2749 cpu = smp_processor_id(); 2750 rq = cpu_rq(cpu); 2751 rcu_note_context_switch(cpu); 2752 prev = rq->curr; 2753 2754 schedule_debug(prev); 2755 2756 if (sched_feat(HRTICK)) 2757 hrtick_clear(rq); 2758 2759 /* 2760 * Make sure that signal_pending_state()->signal_pending() below 2761 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE) 2762 * done by the caller to avoid the race with signal_wake_up(). 2763 */ 2764 smp_mb__before_spinlock(); 2765 raw_spin_lock_irq(&rq->lock); 2766 2767 switch_count = &prev->nivcsw; 2768 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) { 2769 if (unlikely(signal_pending_state(prev->state, prev))) { 2770 prev->state = TASK_RUNNING; 2771 } else { 2772 deactivate_task(rq, prev, DEQUEUE_SLEEP); 2773 prev->on_rq = 0; 2774 2775 /* 2776 * If a worker went to sleep, notify and ask workqueue 2777 * whether it wants to wake up a task to maintain 2778 * concurrency. 2779 */ 2780 if (prev->flags & PF_WQ_WORKER) { 2781 struct task_struct *to_wakeup; 2782 2783 to_wakeup = wq_worker_sleeping(prev, cpu); 2784 if (to_wakeup) 2785 try_to_wake_up_local(to_wakeup); 2786 } 2787 } 2788 switch_count = &prev->nvcsw; 2789 } 2790 2791 if (prev->on_rq || rq->skip_clock_update < 0) 2792 update_rq_clock(rq); 2793 2794 next = pick_next_task(rq, prev); 2795 clear_tsk_need_resched(prev); 2796 clear_preempt_need_resched(); 2797 rq->skip_clock_update = 0; 2798 2799 if (likely(prev != next)) { 2800 rq->nr_switches++; 2801 rq->curr = next; 2802 ++*switch_count; 2803 2804 context_switch(rq, prev, next); /* unlocks the rq */ 2805 /* 2806 * The context switch have flipped the stack from under us 2807 * and restored the local variables which were saved when 2808 * this task called schedule() in the past. prev == current 2809 * is still correct, but it can be moved to another cpu/rq. 2810 */ 2811 cpu = smp_processor_id(); 2812 rq = cpu_rq(cpu); 2813 } else 2814 raw_spin_unlock_irq(&rq->lock); 2815 2816 post_schedule(rq); 2817 2818 sched_preempt_enable_no_resched(); 2819 if (need_resched()) 2820 goto need_resched; 2821} 2822 2823static inline void sched_submit_work(struct task_struct *tsk) 2824{ 2825 if (!tsk->state || tsk_is_pi_blocked(tsk)) 2826 return; 2827 /* 2828 * If we are going to sleep and we have plugged IO queued, 2829 * make sure to submit it to avoid deadlocks. 2830 */ 2831 if (blk_needs_flush_plug(tsk)) 2832 blk_schedule_flush_plug(tsk); 2833} 2834 2835asmlinkage __visible void __sched schedule(void) 2836{ 2837 struct task_struct *tsk = current; 2838 2839 sched_submit_work(tsk); 2840 __schedule(); 2841} 2842EXPORT_SYMBOL(schedule); 2843 2844#ifdef CONFIG_CONTEXT_TRACKING 2845asmlinkage __visible void __sched schedule_user(void) 2846{ 2847 /* 2848 * If we come here after a random call to set_need_resched(), 2849 * or we have been woken up remotely but the IPI has not yet arrived, 2850 * we haven't yet exited the RCU idle mode. Do it here manually until 2851 * we find a better solution. 2852 */ 2853 user_exit(); 2854 schedule(); 2855 user_enter(); 2856} 2857#endif 2858 2859/** 2860 * schedule_preempt_disabled - called with preemption disabled 2861 * 2862 * Returns with preemption disabled. Note: preempt_count must be 1 2863 */ 2864void __sched schedule_preempt_disabled(void) 2865{ 2866 sched_preempt_enable_no_resched(); 2867 schedule(); 2868 preempt_disable(); 2869} 2870 2871#ifdef CONFIG_PREEMPT 2872/* 2873 * this is the entry point to schedule() from in-kernel preemption 2874 * off of preempt_enable. Kernel preemptions off return from interrupt 2875 * occur there and call schedule directly. 2876 */ 2877asmlinkage __visible void __sched notrace preempt_schedule(void) 2878{ 2879 /* 2880 * If there is a non-zero preempt_count or interrupts are disabled, 2881 * we do not want to preempt the current task. Just return.. 2882 */ 2883 if (likely(!preemptible())) 2884 return; 2885 2886 do { 2887 __preempt_count_add(PREEMPT_ACTIVE); 2888 __schedule(); 2889 __preempt_count_sub(PREEMPT_ACTIVE); 2890 2891 /* 2892 * Check again in case we missed a preemption opportunity 2893 * between schedule and now. 2894 */ 2895 barrier(); 2896 } while (need_resched()); 2897} 2898NOKPROBE_SYMBOL(preempt_schedule); 2899EXPORT_SYMBOL(preempt_schedule); 2900#endif /* CONFIG_PREEMPT */ 2901 2902/* 2903 * this is the entry point to schedule() from kernel preemption 2904 * off of irq context. 2905 * Note, that this is called and return with irqs disabled. This will 2906 * protect us against recursive calling from irq. 2907 */ 2908asmlinkage __visible void __sched preempt_schedule_irq(void) 2909{ 2910 enum ctx_state prev_state; 2911 2912 /* Catch callers which need to be fixed */ 2913 BUG_ON(preempt_count() || !irqs_disabled()); 2914 2915 prev_state = exception_enter(); 2916 2917 do { 2918 __preempt_count_add(PREEMPT_ACTIVE); 2919 local_irq_enable(); 2920 __schedule(); 2921 local_irq_disable(); 2922 __preempt_count_sub(PREEMPT_ACTIVE); 2923 2924 /* 2925 * Check again in case we missed a preemption opportunity 2926 * between schedule and now. 2927 */ 2928 barrier(); 2929 } while (need_resched()); 2930 2931 exception_exit(prev_state); 2932} 2933 2934int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags, 2935 void *key) 2936{ 2937 return try_to_wake_up(curr->private, mode, wake_flags); 2938} 2939EXPORT_SYMBOL(default_wake_function); 2940 2941#ifdef CONFIG_RT_MUTEXES 2942 2943/* 2944 * rt_mutex_setprio - set the current priority of a task 2945 * @p: task 2946 * @prio: prio value (kernel-internal form) 2947 * 2948 * This function changes the 'effective' priority of a task. It does 2949 * not touch ->normal_prio like __setscheduler(). 2950 * 2951 * Used by the rt_mutex code to implement priority inheritance 2952 * logic. Call site only calls if the priority of the task changed. 2953 */ 2954void rt_mutex_setprio(struct task_struct *p, int prio) 2955{ 2956 int oldprio, on_rq, running, enqueue_flag = 0; 2957 struct rq *rq; 2958 const struct sched_class *prev_class; 2959 2960 BUG_ON(prio > MAX_PRIO); 2961 2962 rq = __task_rq_lock(p); 2963 2964 /* 2965 * Idle task boosting is a nono in general. There is one 2966 * exception, when PREEMPT_RT and NOHZ is active: 2967 * 2968 * The idle task calls get_next_timer_interrupt() and holds 2969 * the timer wheel base->lock on the CPU and another CPU wants 2970 * to access the timer (probably to cancel it). We can safely 2971 * ignore the boosting request, as the idle CPU runs this code 2972 * with interrupts disabled and will complete the lock 2973 * protected section without being interrupted. So there is no 2974 * real need to boost. 2975 */ 2976 if (unlikely(p == rq->idle)) { 2977 WARN_ON(p != rq->curr); 2978 WARN_ON(p->pi_blocked_on); 2979 goto out_unlock; 2980 } 2981 2982 trace_sched_pi_setprio(p, prio); 2983 oldprio = p->prio; 2984 prev_class = p->sched_class; 2985 on_rq = p->on_rq; 2986 running = task_current(rq, p); 2987 if (on_rq) 2988 dequeue_task(rq, p, 0); 2989 if (running) 2990 p->sched_class->put_prev_task(rq, p); 2991 2992 /* 2993 * Boosting condition are: 2994 * 1. -rt task is running and holds mutex A 2995 * --> -dl task blocks on mutex A 2996 * 2997 * 2. -dl task is running and holds mutex A 2998 * --> -dl task blocks on mutex A and could preempt the 2999 * running task 3000 */ 3001 if (dl_prio(prio)) { 3002 struct task_struct *pi_task = rt_mutex_get_top_task(p); 3003 if (!dl_prio(p->normal_prio) || 3004 (pi_task && dl_entity_preempt(&pi_task->dl, &p->dl))) { 3005 p->dl.dl_boosted = 1; 3006 p->dl.dl_throttled = 0; 3007 enqueue_flag = ENQUEUE_REPLENISH; 3008 } else 3009 p->dl.dl_boosted = 0; 3010 p->sched_class = &dl_sched_class; 3011 } else if (rt_prio(prio)) { 3012 if (dl_prio(oldprio)) 3013 p->dl.dl_boosted = 0; 3014 if (oldprio < prio) 3015 enqueue_flag = ENQUEUE_HEAD; 3016 p->sched_class = &rt_sched_class; 3017 } else { 3018 if (dl_prio(oldprio)) 3019 p->dl.dl_boosted = 0; 3020 p->sched_class = &fair_sched_class; 3021 } 3022 3023 p->prio = prio; 3024 3025 if (running) 3026 p->sched_class->set_curr_task(rq); 3027 if (on_rq) 3028 enqueue_task(rq, p, enqueue_flag); 3029 3030 check_class_changed(rq, p, prev_class, oldprio); 3031out_unlock: 3032 __task_rq_unlock(rq); 3033} 3034#endif 3035 3036void set_user_nice(struct task_struct *p, long nice) 3037{ 3038 int old_prio, delta, on_rq; 3039 unsigned long flags; 3040 struct rq *rq; 3041 3042 if (task_nice(p) == nice || nice < MIN_NICE || nice > MAX_NICE) 3043 return; 3044 /* 3045 * We have to be careful, if called from sys_setpriority(), 3046 * the task might be in the middle of scheduling on another CPU. 3047 */ 3048 rq = task_rq_lock(p, &flags); 3049 /* 3050 * The RT priorities are set via sched_setscheduler(), but we still 3051 * allow the 'normal' nice value to be set - but as expected 3052 * it wont have any effect on scheduling until the task is 3053 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR: 3054 */ 3055 if (task_has_dl_policy(p) || task_has_rt_policy(p)) { 3056 p->static_prio = NICE_TO_PRIO(nice); 3057 goto out_unlock; 3058 } 3059 on_rq = p->on_rq; 3060 if (on_rq) 3061 dequeue_task(rq, p, 0); 3062 3063 p->static_prio = NICE_TO_PRIO(nice); 3064 set_load_weight(p); 3065 old_prio = p->prio; 3066 p->prio = effective_prio(p); 3067 delta = p->prio - old_prio; 3068 3069 if (on_rq) { 3070 enqueue_task(rq, p, 0); 3071 /* 3072 * If the task increased its priority or is running and 3073 * lowered its priority, then reschedule its CPU: 3074 */ 3075 if (delta < 0 || (delta > 0 && task_running(rq, p))) 3076 resched_task(rq->curr); 3077 } 3078out_unlock: 3079 task_rq_unlock(rq, p, &flags); 3080} 3081EXPORT_SYMBOL(set_user_nice); 3082 3083/* 3084 * can_nice - check if a task can reduce its nice value 3085 * @p: task 3086 * @nice: nice value 3087 */ 3088int can_nice(const struct task_struct *p, const int nice) 3089{ 3090 /* convert nice value [19,-20] to rlimit style value [1,40] */ 3091 int nice_rlim = nice_to_rlimit(nice); 3092 3093 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) || 3094 capable(CAP_SYS_NICE)); 3095} 3096 3097#ifdef __ARCH_WANT_SYS_NICE 3098 3099/* 3100 * sys_nice - change the priority of the current process. 3101 * @increment: priority increment 3102 * 3103 * sys_setpriority is a more generic, but much slower function that 3104 * does similar things. 3105 */ 3106SYSCALL_DEFINE1(nice, int, increment) 3107{ 3108 long nice, retval; 3109 3110 /* 3111 * Setpriority might change our priority at the same moment. 3112 * We don't have to worry. Conceptually one call occurs first 3113 * and we have a single winner. 3114 */ 3115 increment = clamp(increment, -NICE_WIDTH, NICE_WIDTH); 3116 nice = task_nice(current) + increment; 3117 3118 nice = clamp_val(nice, MIN_NICE, MAX_NICE); 3119 if (increment < 0 && !can_nice(current, nice)) 3120 return -EPERM; 3121 3122 retval = security_task_setnice(current, nice); 3123 if (retval) 3124 return retval; 3125 3126 set_user_nice(current, nice); 3127 return 0; 3128} 3129 3130#endif 3131 3132/** 3133 * task_prio - return the priority value of a given task. 3134 * @p: the task in question. 3135 * 3136 * Return: The priority value as seen by users in /proc. 3137 * RT tasks are offset by -200. Normal tasks are centered 3138 * around 0, value goes from -16 to +15. 3139 */ 3140int task_prio(const struct task_struct *p) 3141{ 3142 return p->prio - MAX_RT_PRIO; 3143} 3144 3145/** 3146 * idle_cpu - is a given cpu idle currently? 3147 * @cpu: the processor in question. 3148 * 3149 * Return: 1 if the CPU is currently idle. 0 otherwise. 3150 */ 3151int idle_cpu(int cpu) 3152{ 3153 struct rq *rq = cpu_rq(cpu); 3154 3155 if (rq->curr != rq->idle) 3156 return 0; 3157 3158 if (rq->nr_running) 3159 return 0; 3160 3161#ifdef CONFIG_SMP 3162 if (!llist_empty(&rq->wake_list)) 3163 return 0; 3164#endif 3165 3166 return 1; 3167} 3168 3169/** 3170 * idle_task - return the idle task for a given cpu. 3171 * @cpu: the processor in question. 3172 * 3173 * Return: The idle task for the cpu @cpu. 3174 */ 3175struct task_struct *idle_task(int cpu) 3176{ 3177 return cpu_rq(cpu)->idle; 3178} 3179 3180/** 3181 * find_process_by_pid - find a process with a matching PID value. 3182 * @pid: the pid in question. 3183 * 3184 * The task of @pid, if found. %NULL otherwise. 3185 */ 3186static struct task_struct *find_process_by_pid(pid_t pid) 3187{ 3188 return pid ? find_task_by_vpid(pid) : current; 3189} 3190 3191/* 3192 * This function initializes the sched_dl_entity of a newly becoming 3193 * SCHED_DEADLINE task. 3194 * 3195 * Only the static values are considered here, the actual runtime and the 3196 * absolute deadline will be properly calculated when the task is enqueued 3197 * for the first time with its new policy. 3198 */ 3199static void 3200__setparam_dl(struct task_struct *p, const struct sched_attr *attr) 3201{ 3202 struct sched_dl_entity *dl_se = &p->dl; 3203 3204 init_dl_task_timer(dl_se); 3205 dl_se->dl_runtime = attr->sched_runtime; 3206 dl_se->dl_deadline = attr->sched_deadline; 3207 dl_se->dl_period = attr->sched_period ?: dl_se->dl_deadline; 3208 dl_se->flags = attr->sched_flags; 3209 dl_se->dl_bw = to_ratio(dl_se->dl_period, dl_se->dl_runtime); 3210 dl_se->dl_throttled = 0; 3211 dl_se->dl_new = 1; 3212 dl_se->dl_yielded = 0; 3213} 3214 3215static void __setscheduler_params(struct task_struct *p, 3216 const struct sched_attr *attr) 3217{ 3218 int policy = attr->sched_policy; 3219 3220 if (policy == -1) /* setparam */ 3221 policy = p->policy; 3222 3223 p->policy = policy; 3224 3225 if (dl_policy(policy)) 3226 __setparam_dl(p, attr); 3227 else if (fair_policy(policy)) 3228 p->static_prio = NICE_TO_PRIO(attr->sched_nice); 3229 3230 /* 3231 * __sched_setscheduler() ensures attr->sched_priority == 0 when 3232 * !rt_policy. Always setting this ensures that things like 3233 * getparam()/getattr() don't report silly values for !rt tasks. 3234 */ 3235 p->rt_priority = attr->sched_priority; 3236 p->normal_prio = normal_prio(p); 3237 set_load_weight(p); 3238} 3239 3240/* Actually do priority change: must hold pi & rq lock. */ 3241static void __setscheduler(struct rq *rq, struct task_struct *p, 3242 const struct sched_attr *attr) 3243{ 3244 __setscheduler_params(p, attr); 3245 3246 /* 3247 * If we get here, there was no pi waiters boosting the 3248 * task. It is safe to use the normal prio. 3249 */ 3250 p->prio = normal_prio(p); 3251 3252 if (dl_prio(p->prio)) 3253 p->sched_class = &dl_sched_class; 3254 else if (rt_prio(p->prio)) 3255 p->sched_class = &rt_sched_class; 3256 else 3257 p->sched_class = &fair_sched_class; 3258} 3259 3260static void 3261__getparam_dl(struct task_struct *p, struct sched_attr *attr) 3262{ 3263 struct sched_dl_entity *dl_se = &p->dl; 3264 3265 attr->sched_priority = p->rt_priority; 3266 attr->sched_runtime = dl_se->dl_runtime; 3267 attr->sched_deadline = dl_se->dl_deadline; 3268 attr->sched_period = dl_se->dl_period; 3269 attr->sched_flags = dl_se->flags; 3270} 3271 3272/* 3273 * This function validates the new parameters of a -deadline task. 3274 * We ask for the deadline not being zero, and greater or equal 3275 * than the runtime, as well as the period of being zero or 3276 * greater than deadline. Furthermore, we have to be sure that 3277 * user parameters are above the internal resolution of 1us (we 3278 * check sched_runtime only since it is always the smaller one) and 3279 * below 2^63 ns (we have to check both sched_deadline and 3280 * sched_period, as the latter can be zero). 3281 */ 3282static bool 3283__checkparam_dl(const struct sched_attr *attr) 3284{ 3285 /* deadline != 0 */ 3286 if (attr->sched_deadline == 0) 3287 return false; 3288 3289 /* 3290 * Since we truncate DL_SCALE bits, make sure we're at least 3291 * that big. 3292 */ 3293 if (attr->sched_runtime < (1ULL << DL_SCALE)) 3294 return false; 3295 3296 /* 3297 * Since we use the MSB for wrap-around and sign issues, make 3298 * sure it's not set (mind that period can be equal to zero). 3299 */ 3300 if (attr->sched_deadline & (1ULL << 63) || 3301 attr->sched_period & (1ULL << 63)) 3302 return false; 3303 3304 /* runtime <= deadline <= period (if period != 0) */ 3305 if ((attr->sched_period != 0 && 3306 attr->sched_period < attr->sched_deadline) || 3307 attr->sched_deadline < attr->sched_runtime) 3308 return false; 3309 3310 return true; 3311} 3312 3313/* 3314 * check the target process has a UID that matches the current process's 3315 */ 3316static bool check_same_owner(struct task_struct *p) 3317{ 3318 const struct cred *cred = current_cred(), *pcred; 3319 bool match; 3320 3321 rcu_read_lock(); 3322 pcred = __task_cred(p); 3323 match = (uid_eq(cred->euid, pcred->euid) || 3324 uid_eq(cred->euid, pcred->uid)); 3325 rcu_read_unlock(); 3326 return match; 3327} 3328 3329static int __sched_setscheduler(struct task_struct *p, 3330 const struct sched_attr *attr, 3331 bool user) 3332{ 3333 int newprio = dl_policy(attr->sched_policy) ? MAX_DL_PRIO - 1 : 3334 MAX_RT_PRIO - 1 - attr->sched_priority; 3335 int retval, oldprio, oldpolicy = -1, on_rq, running; 3336 int policy = attr->sched_policy; 3337 unsigned long flags; 3338 const struct sched_class *prev_class; 3339 struct rq *rq; 3340 int reset_on_fork; 3341 3342 /* may grab non-irq protected spin_locks */ 3343 BUG_ON(in_interrupt()); 3344recheck: 3345 /* double check policy once rq lock held */ 3346 if (policy < 0) { 3347 reset_on_fork = p->sched_reset_on_fork; 3348 policy = oldpolicy = p->policy; 3349 } else { 3350 reset_on_fork = !!(attr->sched_flags & SCHED_FLAG_RESET_ON_FORK); 3351 3352 if (policy != SCHED_DEADLINE && 3353 policy != SCHED_FIFO && policy != SCHED_RR && 3354 policy != SCHED_NORMAL && policy != SCHED_BATCH && 3355 policy != SCHED_IDLE) 3356 return -EINVAL; 3357 } 3358 3359 if (attr->sched_flags & ~(SCHED_FLAG_RESET_ON_FORK)) 3360 return -EINVAL; 3361 3362 /* 3363 * Valid priorities for SCHED_FIFO and SCHED_RR are 3364 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL, 3365 * SCHED_BATCH and SCHED_IDLE is 0. 3366 */ 3367 if ((p->mm && attr->sched_priority > MAX_USER_RT_PRIO-1) || 3368 (!p->mm && attr->sched_priority > MAX_RT_PRIO-1)) 3369 return -EINVAL; 3370 if ((dl_policy(policy) && !__checkparam_dl(attr)) || 3371 (rt_policy(policy) != (attr->sched_priority != 0))) 3372 return -EINVAL; 3373 3374 /* 3375 * Allow unprivileged RT tasks to decrease priority: 3376 */ 3377 if (user && !capable(CAP_SYS_NICE)) { 3378 if (fair_policy(policy)) { 3379 if (attr->sched_nice < task_nice(p) && 3380 !can_nice(p, attr->sched_nice)) 3381 return -EPERM; 3382 } 3383 3384 if (rt_policy(policy)) { 3385 unsigned long rlim_rtprio = 3386 task_rlimit(p, RLIMIT_RTPRIO); 3387 3388 /* can't set/change the rt policy */ 3389 if (policy != p->policy && !rlim_rtprio) 3390 return -EPERM; 3391 3392 /* can't increase priority */ 3393 if (attr->sched_priority > p->rt_priority && 3394 attr->sched_priority > rlim_rtprio) 3395 return -EPERM; 3396 } 3397 3398 /* 3399 * Can't set/change SCHED_DEADLINE policy at all for now 3400 * (safest behavior); in the future we would like to allow 3401 * unprivileged DL tasks to increase their relative deadline 3402 * or reduce their runtime (both ways reducing utilization) 3403 */ 3404 if (dl_policy(policy)) 3405 return -EPERM; 3406 3407 /* 3408 * Treat SCHED_IDLE as nice 20. Only allow a switch to 3409 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it. 3410 */ 3411 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) { 3412 if (!can_nice(p, task_nice(p))) 3413 return -EPERM; 3414 } 3415 3416 /* can't change other user's priorities */ 3417 if (!check_same_owner(p)) 3418 return -EPERM; 3419 3420 /* Normal users shall not reset the sched_reset_on_fork flag */ 3421 if (p->sched_reset_on_fork && !reset_on_fork) 3422 return -EPERM; 3423 } 3424 3425 if (user) { 3426 retval = security_task_setscheduler(p); 3427 if (retval) 3428 return retval; 3429 } 3430 3431 /* 3432 * make sure no PI-waiters arrive (or leave) while we are 3433 * changing the priority of the task: 3434 * 3435 * To be able to change p->policy safely, the appropriate 3436 * runqueue lock must be held. 3437 */ 3438 rq = task_rq_lock(p, &flags); 3439 3440 /* 3441 * Changing the policy of the stop threads its a very bad idea 3442 */ 3443 if (p == rq->stop) { 3444 task_rq_unlock(rq, p, &flags); 3445 return -EINVAL; 3446 } 3447 3448 /* 3449 * If not changing anything there's no need to proceed further, 3450 * but store a possible modification of reset_on_fork. 3451 */ 3452 if (unlikely(policy == p->policy)) { 3453 if (fair_policy(policy) && attr->sched_nice != task_nice(p)) 3454 goto change; 3455 if (rt_policy(policy) && attr->sched_priority != p->rt_priority) 3456 goto change; 3457 if (dl_policy(policy)) 3458 goto change; 3459 3460 p->sched_reset_on_fork = reset_on_fork; 3461 task_rq_unlock(rq, p, &flags); 3462 return 0; 3463 } 3464change: 3465 3466 if (user) { 3467#ifdef CONFIG_RT_GROUP_SCHED 3468 /* 3469 * Do not allow realtime tasks into groups that have no runtime 3470 * assigned. 3471 */ 3472 if (rt_bandwidth_enabled() && rt_policy(policy) && 3473 task_group(p)->rt_bandwidth.rt_runtime == 0 && 3474 !task_group_is_autogroup(task_group(p))) { 3475 task_rq_unlock(rq, p, &flags); 3476 return -EPERM; 3477 } 3478#endif 3479#ifdef CONFIG_SMP 3480 if (dl_bandwidth_enabled() && dl_policy(policy)) { 3481 cpumask_t *span = rq->rd->span; 3482 3483 /* 3484 * Don't allow tasks with an affinity mask smaller than 3485 * the entire root_domain to become SCHED_DEADLINE. We 3486 * will also fail if there's no bandwidth available. 3487 */ 3488 if (!cpumask_subset(span, &p->cpus_allowed) || 3489 rq->rd->dl_bw.bw == 0) { 3490 task_rq_unlock(rq, p, &flags); 3491 return -EPERM; 3492 } 3493 } 3494#endif 3495 } 3496 3497 /* recheck policy now with rq lock held */ 3498 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) { 3499 policy = oldpolicy = -1; 3500 task_rq_unlock(rq, p, &flags); 3501 goto recheck; 3502 } 3503 3504 /* 3505 * If setscheduling to SCHED_DEADLINE (or changing the parameters 3506 * of a SCHED_DEADLINE task) we need to check if enough bandwidth 3507 * is available. 3508 */ 3509 if ((dl_policy(policy) || dl_task(p)) && dl_overflow(p, policy, attr)) { 3510 task_rq_unlock(rq, p, &flags); 3511 return -EBUSY; 3512 } 3513 3514 p->sched_reset_on_fork = reset_on_fork; 3515 oldprio = p->prio; 3516 3517 /* 3518 * Special case for priority boosted tasks. 3519 * 3520 * If the new priority is lower or equal (user space view) 3521 * than the current (boosted) priority, we just store the new 3522 * normal parameters and do not touch the scheduler class and 3523 * the runqueue. This will be done when the task deboost 3524 * itself. 3525 */ 3526 if (rt_mutex_check_prio(p, newprio)) { 3527 __setscheduler_params(p, attr); 3528 task_rq_unlock(rq, p, &flags); 3529 return 0; 3530 } 3531 3532 on_rq = p->on_rq; 3533 running = task_current(rq, p); 3534 if (on_rq) 3535 dequeue_task(rq, p, 0); 3536 if (running) 3537 p->sched_class->put_prev_task(rq, p); 3538 3539 prev_class = p->sched_class; 3540 __setscheduler(rq, p, attr); 3541 3542 if (running) 3543 p->sched_class->set_curr_task(rq); 3544 if (on_rq) { 3545 /* 3546 * We enqueue to tail when the priority of a task is 3547 * increased (user space view). 3548 */ 3549 enqueue_task(rq, p, oldprio <= p->prio ? ENQUEUE_HEAD : 0); 3550 } 3551 3552 check_class_changed(rq, p, prev_class, oldprio); 3553 task_rq_unlock(rq, p, &flags); 3554 3555 rt_mutex_adjust_pi(p); 3556 3557 return 0; 3558} 3559 3560static int _sched_setscheduler(struct task_struct *p, int policy, 3561 const struct sched_param *param, bool check) 3562{ 3563 struct sched_attr attr = { 3564 .sched_policy = policy, 3565 .sched_priority = param->sched_priority, 3566 .sched_nice = PRIO_TO_NICE(p->static_prio), 3567 }; 3568 3569 /* 3570 * Fixup the legacy SCHED_RESET_ON_FORK hack 3571 */ 3572 if (policy & SCHED_RESET_ON_FORK) { 3573 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK; 3574 policy &= ~SCHED_RESET_ON_FORK; 3575 attr.sched_policy = policy; 3576 } 3577 3578 return __sched_setscheduler(p, &attr, check); 3579} 3580/** 3581 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread. 3582 * @p: the task in question. 3583 * @policy: new policy. 3584 * @param: structure containing the new RT priority. 3585 * 3586 * Return: 0 on success. An error code otherwise. 3587 * 3588 * NOTE that the task may be already dead. 3589 */ 3590int sched_setscheduler(struct task_struct *p, int policy, 3591 const struct sched_param *param) 3592{ 3593 return _sched_setscheduler(p, policy, param, true); 3594} 3595EXPORT_SYMBOL_GPL(sched_setscheduler); 3596 3597int sched_setattr(struct task_struct *p, const struct sched_attr *attr) 3598{ 3599 return __sched_setscheduler(p, attr, true); 3600} 3601EXPORT_SYMBOL_GPL(sched_setattr); 3602 3603/** 3604 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace. 3605 * @p: the task in question. 3606 * @policy: new policy. 3607 * @param: structure containing the new RT priority. 3608 * 3609 * Just like sched_setscheduler, only don't bother checking if the 3610 * current context has permission. For example, this is needed in 3611 * stop_machine(): we create temporary high priority worker threads, 3612 * but our caller might not have that capability. 3613 * 3614 * Return: 0 on success. An error code otherwise. 3615 */ 3616int sched_setscheduler_nocheck(struct task_struct *p, int policy, 3617 const struct sched_param *param) 3618{ 3619 return _sched_setscheduler(p, policy, param, false); 3620} 3621 3622static int 3623do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param) 3624{ 3625 struct sched_param lparam; 3626 struct task_struct *p; 3627 int retval; 3628 3629 if (!param || pid < 0) 3630 return -EINVAL; 3631 if (copy_from_user(&lparam, param, sizeof(struct sched_param))) 3632 return -EFAULT; 3633 3634 rcu_read_lock(); 3635 retval = -ESRCH; 3636 p = find_process_by_pid(pid); 3637 if (p != NULL) 3638 retval = sched_setscheduler(p, policy, &lparam); 3639 rcu_read_unlock(); 3640 3641 return retval; 3642} 3643 3644/* 3645 * Mimics kernel/events/core.c perf_copy_attr(). 3646 */ 3647static int sched_copy_attr(struct sched_attr __user *uattr, 3648 struct sched_attr *attr) 3649{ 3650 u32 size; 3651 int ret; 3652 3653 if (!access_ok(VERIFY_WRITE, uattr, SCHED_ATTR_SIZE_VER0)) 3654 return -EFAULT; 3655 3656 /* 3657 * zero the full structure, so that a short copy will be nice. 3658 */ 3659 memset(attr, 0, sizeof(*attr)); 3660 3661 ret = get_user(size, &uattr->size); 3662 if (ret) 3663 return ret; 3664 3665 if (size > PAGE_SIZE) /* silly large */ 3666 goto err_size; 3667 3668 if (!size) /* abi compat */ 3669 size = SCHED_ATTR_SIZE_VER0; 3670 3671 if (size < SCHED_ATTR_SIZE_VER0) 3672 goto err_size; 3673 3674 /* 3675 * If we're handed a bigger struct than we know of, 3676 * ensure all the unknown bits are 0 - i.e. new 3677 * user-space does not rely on any kernel feature 3678 * extensions we dont know about yet. 3679 */ 3680 if (size > sizeof(*attr)) { 3681 unsigned char __user *addr; 3682 unsigned char __user *end; 3683 unsigned char val; 3684 3685 addr = (void __user *)uattr + sizeof(*attr); 3686 end = (void __user *)uattr + size; 3687 3688 for (; addr < end; addr++) { 3689 ret = get_user(val, addr); 3690 if (ret) 3691 return ret; 3692 if (val) 3693 goto err_size; 3694 } 3695 size = sizeof(*attr); 3696 } 3697 3698 ret = copy_from_user(attr, uattr, size); 3699 if (ret) 3700 return -EFAULT; 3701 3702 /* 3703 * XXX: do we want to be lenient like existing syscalls; or do we want 3704 * to be strict and return an error on out-of-bounds values? 3705 */ 3706 attr->sched_nice = clamp(attr->sched_nice, MIN_NICE, MAX_NICE); 3707 3708 return 0; 3709 3710err_size: 3711 put_user(sizeof(*attr), &uattr->size); 3712 return -E2BIG; 3713} 3714 3715/** 3716 * sys_sched_setscheduler - set/change the scheduler policy and RT priority 3717 * @pid: the pid in question. 3718 * @policy: new policy. 3719 * @param: structure containing the new RT priority. 3720 * 3721 * Return: 0 on success. An error code otherwise. 3722 */ 3723SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy, 3724 struct sched_param __user *, param) 3725{ 3726 /* negative values for policy are not valid */ 3727 if (policy < 0) 3728 return -EINVAL; 3729 3730 return do_sched_setscheduler(pid, policy, param); 3731} 3732 3733/** 3734 * sys_sched_setparam - set/change the RT priority of a thread 3735 * @pid: the pid in question. 3736 * @param: structure containing the new RT priority. 3737 * 3738 * Return: 0 on success. An error code otherwise. 3739 */ 3740SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param) 3741{ 3742 return do_sched_setscheduler(pid, -1, param); 3743} 3744 3745/** 3746 * sys_sched_setattr - same as above, but with extended sched_attr 3747 * @pid: the pid in question. 3748 * @uattr: structure containing the extended parameters. 3749 * @flags: for future extension. 3750 */ 3751SYSCALL_DEFINE3(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr, 3752 unsigned int, flags) 3753{ 3754 struct sched_attr attr; 3755 struct task_struct *p; 3756 int retval; 3757 3758 if (!uattr || pid < 0 || flags) 3759 return -EINVAL; 3760 3761 retval = sched_copy_attr(uattr, &attr); 3762 if (retval) 3763 return retval; 3764 3765 if ((int)attr.sched_policy < 0) 3766 return -EINVAL; 3767 3768 rcu_read_lock(); 3769 retval = -ESRCH; 3770 p = find_process_by_pid(pid); 3771 if (p != NULL) 3772 retval = sched_setattr(p, &attr); 3773 rcu_read_unlock(); 3774 3775 return retval; 3776} 3777 3778/** 3779 * sys_sched_getscheduler - get the policy (scheduling class) of a thread 3780 * @pid: the pid in question. 3781 * 3782 * Return: On success, the policy of the thread. Otherwise, a negative error 3783 * code. 3784 */ 3785SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid) 3786{ 3787 struct task_struct *p; 3788 int retval; 3789 3790 if (pid < 0) 3791 return -EINVAL; 3792 3793 retval = -ESRCH; 3794 rcu_read_lock(); 3795 p = find_process_by_pid(pid); 3796 if (p) { 3797 retval = security_task_getscheduler(p); 3798 if (!retval) 3799 retval = p->policy 3800 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0); 3801 } 3802 rcu_read_unlock(); 3803 return retval; 3804} 3805 3806/** 3807 * sys_sched_getparam - get the RT priority of a thread 3808 * @pid: the pid in question. 3809 * @param: structure containing the RT priority. 3810 * 3811 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error 3812 * code. 3813 */ 3814SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param) 3815{ 3816 struct sched_param lp = { .sched_priority = 0 }; 3817 struct task_struct *p; 3818 int retval; 3819 3820 if (!param || pid < 0) 3821 return -EINVAL; 3822 3823 rcu_read_lock(); 3824 p = find_process_by_pid(pid); 3825 retval = -ESRCH; 3826 if (!p) 3827 goto out_unlock; 3828 3829 retval = security_task_getscheduler(p); 3830 if (retval) 3831 goto out_unlock; 3832 3833 if (task_has_rt_policy(p)) 3834 lp.sched_priority = p->rt_priority; 3835 rcu_read_unlock(); 3836 3837 /* 3838 * This one might sleep, we cannot do it with a spinlock held ... 3839 */ 3840 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0; 3841 3842 return retval; 3843 3844out_unlock: 3845 rcu_read_unlock(); 3846 return retval; 3847} 3848 3849static int sched_read_attr(struct sched_attr __user *uattr, 3850 struct sched_attr *attr, 3851 unsigned int usize) 3852{ 3853 int ret; 3854 3855 if (!access_ok(VERIFY_WRITE, uattr, usize)) 3856 return -EFAULT; 3857 3858 /* 3859 * If we're handed a smaller struct than we know of, 3860 * ensure all the unknown bits are 0 - i.e. old 3861 * user-space does not get uncomplete information. 3862 */ 3863 if (usize < sizeof(*attr)) { 3864 unsigned char *addr; 3865 unsigned char *end; 3866 3867 addr = (void *)attr + usize; 3868 end = (void *)attr + sizeof(*attr); 3869 3870 for (; addr < end; addr++) { 3871 if (*addr) 3872 return -EFBIG; 3873 } 3874 3875 attr->size = usize; 3876 } 3877 3878 ret = copy_to_user(uattr, attr, attr->size); 3879 if (ret) 3880 return -EFAULT; 3881 3882 return 0; 3883} 3884 3885/** 3886 * sys_sched_getattr - similar to sched_getparam, but with sched_attr 3887 * @pid: the pid in question. 3888 * @uattr: structure containing the extended parameters. 3889 * @size: sizeof(attr) for fwd/bwd comp. 3890 * @flags: for future extension. 3891 */ 3892SYSCALL_DEFINE4(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr, 3893 unsigned int, size, unsigned int, flags) 3894{ 3895 struct sched_attr attr = { 3896 .size = sizeof(struct sched_attr), 3897 }; 3898 struct task_struct *p; 3899 int retval; 3900 3901 if (!uattr || pid < 0 || size > PAGE_SIZE || 3902 size < SCHED_ATTR_SIZE_VER0 || flags) 3903 return -EINVAL; 3904 3905 rcu_read_lock(); 3906 p = find_process_by_pid(pid); 3907 retval = -ESRCH; 3908 if (!p) 3909 goto out_unlock; 3910 3911 retval = security_task_getscheduler(p); 3912 if (retval) 3913 goto out_unlock; 3914 3915 attr.sched_policy = p->policy; 3916 if (p->sched_reset_on_fork) 3917 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK; 3918 if (task_has_dl_policy(p)) 3919 __getparam_dl(p, &attr); 3920 else if (task_has_rt_policy(p)) 3921 attr.sched_priority = p->rt_priority; 3922 else 3923 attr.sched_nice = task_nice(p); 3924 3925 rcu_read_unlock(); 3926 3927 retval = sched_read_attr(uattr, &attr, size); 3928 return retval; 3929 3930out_unlock: 3931 rcu_read_unlock(); 3932 return retval; 3933} 3934 3935long sched_setaffinity(pid_t pid, const struct cpumask *in_mask) 3936{ 3937 cpumask_var_t cpus_allowed, new_mask; 3938 struct task_struct *p; 3939 int retval; 3940 3941 rcu_read_lock(); 3942 3943 p = find_process_by_pid(pid); 3944 if (!p) { 3945 rcu_read_unlock(); 3946 return -ESRCH; 3947 } 3948 3949 /* Prevent p going away */ 3950 get_task_struct(p); 3951 rcu_read_unlock(); 3952 3953 if (p->flags & PF_NO_SETAFFINITY) { 3954 retval = -EINVAL; 3955 goto out_put_task; 3956 } 3957 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) { 3958 retval = -ENOMEM; 3959 goto out_put_task; 3960 } 3961 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) { 3962 retval = -ENOMEM; 3963 goto out_free_cpus_allowed; 3964 } 3965 retval = -EPERM; 3966 if (!check_same_owner(p)) { 3967 rcu_read_lock(); 3968 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) { 3969 rcu_read_unlock(); 3970 goto out_unlock; 3971 } 3972 rcu_read_unlock(); 3973 } 3974 3975 retval = security_task_setscheduler(p); 3976 if (retval) 3977 goto out_unlock; 3978 3979 3980 cpuset_cpus_allowed(p, cpus_allowed); 3981 cpumask_and(new_mask, in_mask, cpus_allowed); 3982 3983 /* 3984 * Since bandwidth control happens on root_domain basis, 3985 * if admission test is enabled, we only admit -deadline 3986 * tasks allowed to run on all the CPUs in the task's 3987 * root_domain. 3988 */ 3989#ifdef CONFIG_SMP 3990 if (task_has_dl_policy(p)) { 3991 const struct cpumask *span = task_rq(p)->rd->span; 3992 3993 if (dl_bandwidth_enabled() && !cpumask_subset(span, new_mask)) { 3994 retval = -EBUSY; 3995 goto out_unlock; 3996 } 3997 } 3998#endif 3999again: 4000 retval = set_cpus_allowed_ptr(p, new_mask); 4001 4002 if (!retval) { 4003 cpuset_cpus_allowed(p, cpus_allowed); 4004 if (!cpumask_subset(new_mask, cpus_allowed)) { 4005 /* 4006 * We must have raced with a concurrent cpuset 4007 * update. Just reset the cpus_allowed to the 4008 * cpuset's cpus_allowed 4009 */ 4010 cpumask_copy(new_mask, cpus_allowed); 4011 goto again; 4012 } 4013 } 4014out_unlock: 4015 free_cpumask_var(new_mask); 4016out_free_cpus_allowed: 4017 free_cpumask_var(cpus_allowed); 4018out_put_task: 4019 put_task_struct(p); 4020 return retval; 4021} 4022 4023static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len, 4024 struct cpumask *new_mask) 4025{ 4026 if (len < cpumask_size()) 4027 cpumask_clear(new_mask); 4028 else if (len > cpumask_size()) 4029 len = cpumask_size(); 4030 4031 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0; 4032} 4033 4034/** 4035 * sys_sched_setaffinity - set the cpu affinity of a process 4036 * @pid: pid of the process 4037 * @len: length in bytes of the bitmask pointed to by user_mask_ptr 4038 * @user_mask_ptr: user-space pointer to the new cpu mask 4039 * 4040 * Return: 0 on success. An error code otherwise. 4041 */ 4042SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len, 4043 unsigned long __user *, user_mask_ptr) 4044{ 4045 cpumask_var_t new_mask; 4046 int retval; 4047 4048 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) 4049 return -ENOMEM; 4050 4051 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask); 4052 if (retval == 0) 4053 retval = sched_setaffinity(pid, new_mask); 4054 free_cpumask_var(new_mask); 4055 return retval; 4056} 4057 4058long sched_getaffinity(pid_t pid, struct cpumask *mask) 4059{ 4060 struct task_struct *p; 4061 unsigned long flags; 4062 int retval; 4063 4064 rcu_read_lock(); 4065 4066 retval = -ESRCH; 4067 p = find_process_by_pid(pid); 4068 if (!p) 4069 goto out_unlock; 4070 4071 retval = security_task_getscheduler(p); 4072 if (retval) 4073 goto out_unlock; 4074 4075 raw_spin_lock_irqsave(&p->pi_lock, flags); 4076 cpumask_and(mask, &p->cpus_allowed, cpu_active_mask); 4077 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 4078 4079out_unlock: 4080 rcu_read_unlock(); 4081 4082 return retval; 4083} 4084 4085/** 4086 * sys_sched_getaffinity - get the cpu affinity of a process 4087 * @pid: pid of the process 4088 * @len: length in bytes of the bitmask pointed to by user_mask_ptr 4089 * @user_mask_ptr: user-space pointer to hold the current cpu mask 4090 * 4091 * Return: 0 on success. An error code otherwise. 4092 */ 4093SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len, 4094 unsigned long __user *, user_mask_ptr) 4095{ 4096 int ret; 4097 cpumask_var_t mask; 4098 4099 if ((len * BITS_PER_BYTE) < nr_cpu_ids) 4100 return -EINVAL; 4101 if (len & (sizeof(unsigned long)-1)) 4102 return -EINVAL; 4103 4104 if (!alloc_cpumask_var(&mask, GFP_KERNEL)) 4105 return -ENOMEM; 4106 4107 ret = sched_getaffinity(pid, mask); 4108 if (ret == 0) { 4109 size_t retlen = min_t(size_t, len, cpumask_size()); 4110 4111 if (copy_to_user(user_mask_ptr, mask, retlen)) 4112 ret = -EFAULT; 4113 else 4114 ret = retlen; 4115 } 4116 free_cpumask_var(mask); 4117 4118 return ret; 4119} 4120 4121/** 4122 * sys_sched_yield - yield the current processor to other threads. 4123 * 4124 * This function yields the current CPU to other tasks. If there are no 4125 * other threads running on this CPU then this function will return. 4126 * 4127 * Return: 0. 4128 */ 4129SYSCALL_DEFINE0(sched_yield) 4130{ 4131 struct rq *rq = this_rq_lock(); 4132 4133 schedstat_inc(rq, yld_count); 4134 current->sched_class->yield_task(rq); 4135 4136 /* 4137 * Since we are going to call schedule() anyway, there's 4138 * no need to preempt or enable interrupts: 4139 */ 4140 __release(rq->lock); 4141 spin_release(&rq->lock.dep_map, 1, _THIS_IP_); 4142 do_raw_spin_unlock(&rq->lock); 4143 sched_preempt_enable_no_resched(); 4144 4145 schedule(); 4146 4147 return 0; 4148} 4149 4150static void __cond_resched(void) 4151{ 4152 __preempt_count_add(PREEMPT_ACTIVE); 4153 __schedule(); 4154 __preempt_count_sub(PREEMPT_ACTIVE); 4155} 4156 4157int __sched _cond_resched(void) 4158{ 4159 rcu_cond_resched(); 4160 if (should_resched()) { 4161 __cond_resched(); 4162 return 1; 4163 } 4164 return 0; 4165} 4166EXPORT_SYMBOL(_cond_resched); 4167 4168/* 4169 * __cond_resched_lock() - if a reschedule is pending, drop the given lock, 4170 * call schedule, and on return reacquire the lock. 4171 * 4172 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level 4173 * operations here to prevent schedule() from being called twice (once via 4174 * spin_unlock(), once by hand). 4175 */ 4176int __cond_resched_lock(spinlock_t *lock) 4177{ 4178 bool need_rcu_resched = rcu_should_resched(); 4179 int resched = should_resched(); 4180 int ret = 0; 4181 4182 lockdep_assert_held(lock); 4183 4184 if (spin_needbreak(lock) || resched || need_rcu_resched) { 4185 spin_unlock(lock); 4186 if (resched) 4187 __cond_resched(); 4188 else if (unlikely(need_rcu_resched)) 4189 rcu_resched(); 4190 else 4191 cpu_relax(); 4192 ret = 1; 4193 spin_lock(lock); 4194 } 4195 return ret; 4196} 4197EXPORT_SYMBOL(__cond_resched_lock); 4198 4199int __sched __cond_resched_softirq(void) 4200{ 4201 BUG_ON(!in_softirq()); 4202 4203 rcu_cond_resched(); /* BH disabled OK, just recording QSes. */ 4204 if (should_resched()) { 4205 local_bh_enable(); 4206 __cond_resched(); 4207 local_bh_disable(); 4208 return 1; 4209 } 4210 return 0; 4211} 4212EXPORT_SYMBOL(__cond_resched_softirq); 4213 4214/** 4215 * yield - yield the current processor to other threads. 4216 * 4217 * Do not ever use this function, there's a 99% chance you're doing it wrong. 4218 * 4219 * The scheduler is at all times free to pick the calling task as the most 4220 * eligible task to run, if removing the yield() call from your code breaks 4221 * it, its already broken. 4222 * 4223 * Typical broken usage is: 4224 * 4225 * while (!event) 4226 * yield(); 4227 * 4228 * where one assumes that yield() will let 'the other' process run that will 4229 * make event true. If the current task is a SCHED_FIFO task that will never 4230 * happen. Never use yield() as a progress guarantee!! 4231 * 4232 * If you want to use yield() to wait for something, use wait_event(). 4233 * If you want to use yield() to be 'nice' for others, use cond_resched(). 4234 * If you still want to use yield(), do not! 4235 */ 4236void __sched yield(void) 4237{ 4238 set_current_state(TASK_RUNNING); 4239 sys_sched_yield(); 4240} 4241EXPORT_SYMBOL(yield); 4242 4243/** 4244 * yield_to - yield the current processor to another thread in 4245 * your thread group, or accelerate that thread toward the 4246 * processor it's on. 4247 * @p: target task 4248 * @preempt: whether task preemption is allowed or not 4249 * 4250 * It's the caller's job to ensure that the target task struct 4251 * can't go away on us before we can do any checks. 4252 * 4253 * Return: 4254 * true (>0) if we indeed boosted the target task. 4255 * false (0) if we failed to boost the target. 4256 * -ESRCH if there's no task to yield to. 4257 */ 4258int __sched yield_to(struct task_struct *p, bool preempt) 4259{ 4260 struct task_struct *curr = current; 4261 struct rq *rq, *p_rq; 4262 unsigned long flags; 4263 int yielded = 0; 4264 4265 local_irq_save(flags); 4266 rq = this_rq(); 4267 4268again: 4269 p_rq = task_rq(p); 4270 /* 4271 * If we're the only runnable task on the rq and target rq also 4272 * has only one task, there's absolutely no point in yielding. 4273 */ 4274 if (rq->nr_running == 1 && p_rq->nr_running == 1) { 4275 yielded = -ESRCH; 4276 goto out_irq; 4277 } 4278 4279 double_rq_lock(rq, p_rq); 4280 if (task_rq(p) != p_rq) { 4281 double_rq_unlock(rq, p_rq); 4282 goto again; 4283 } 4284 4285 if (!curr->sched_class->yield_to_task) 4286 goto out_unlock; 4287 4288 if (curr->sched_class != p->sched_class) 4289 goto out_unlock; 4290 4291 if (task_running(p_rq, p) || p->state) 4292 goto out_unlock; 4293 4294 yielded = curr->sched_class->yield_to_task(rq, p, preempt); 4295 if (yielded) { 4296 schedstat_inc(rq, yld_count); 4297 /* 4298 * Make p's CPU reschedule; pick_next_entity takes care of 4299 * fairness. 4300 */ 4301 if (preempt && rq != p_rq) 4302 resched_task(p_rq->curr); 4303 } 4304 4305out_unlock: 4306 double_rq_unlock(rq, p_rq); 4307out_irq: 4308 local_irq_restore(flags); 4309 4310 if (yielded > 0) 4311 schedule(); 4312 4313 return yielded; 4314} 4315EXPORT_SYMBOL_GPL(yield_to); 4316 4317/* 4318 * This task is about to go to sleep on IO. Increment rq->nr_iowait so 4319 * that process accounting knows that this is a task in IO wait state. 4320 */ 4321void __sched io_schedule(void) 4322{ 4323 struct rq *rq = raw_rq(); 4324 4325 delayacct_blkio_start(); 4326 atomic_inc(&rq->nr_iowait); 4327 blk_flush_plug(current); 4328 current->in_iowait = 1; 4329 schedule(); 4330 current->in_iowait = 0; 4331 atomic_dec(&rq->nr_iowait); 4332 delayacct_blkio_end(); 4333} 4334EXPORT_SYMBOL(io_schedule); 4335 4336long __sched io_schedule_timeout(long timeout) 4337{ 4338 struct rq *rq = raw_rq(); 4339 long ret; 4340 4341 delayacct_blkio_start(); 4342 atomic_inc(&rq->nr_iowait); 4343 blk_flush_plug(current); 4344 current->in_iowait = 1; 4345 ret = schedule_timeout(timeout); 4346 current->in_iowait = 0; 4347 atomic_dec(&rq->nr_iowait); 4348 delayacct_blkio_end(); 4349 return ret; 4350} 4351 4352/** 4353 * sys_sched_get_priority_max - return maximum RT priority. 4354 * @policy: scheduling class. 4355 * 4356 * Return: On success, this syscall returns the maximum 4357 * rt_priority that can be used by a given scheduling class. 4358 * On failure, a negative error code is returned. 4359 */ 4360SYSCALL_DEFINE1(sched_get_priority_max, int, policy) 4361{ 4362 int ret = -EINVAL; 4363 4364 switch (policy) { 4365 case SCHED_FIFO: 4366 case SCHED_RR: 4367 ret = MAX_USER_RT_PRIO-1; 4368 break; 4369 case SCHED_DEADLINE: 4370 case SCHED_NORMAL: 4371 case SCHED_BATCH: 4372 case SCHED_IDLE: 4373 ret = 0; 4374 break; 4375 } 4376 return ret; 4377} 4378 4379/** 4380 * sys_sched_get_priority_min - return minimum RT priority. 4381 * @policy: scheduling class. 4382 * 4383 * Return: On success, this syscall returns the minimum 4384 * rt_priority that can be used by a given scheduling class. 4385 * On failure, a negative error code is returned. 4386 */ 4387SYSCALL_DEFINE1(sched_get_priority_min, int, policy) 4388{ 4389 int ret = -EINVAL; 4390 4391 switch (policy) { 4392 case SCHED_FIFO: 4393 case SCHED_RR: 4394 ret = 1; 4395 break; 4396 case SCHED_DEADLINE: 4397 case SCHED_NORMAL: 4398 case SCHED_BATCH: 4399 case SCHED_IDLE: 4400 ret = 0; 4401 } 4402 return ret; 4403} 4404 4405/** 4406 * sys_sched_rr_get_interval - return the default timeslice of a process. 4407 * @pid: pid of the process. 4408 * @interval: userspace pointer to the timeslice value. 4409 * 4410 * this syscall writes the default timeslice value of a given process 4411 * into the user-space timespec buffer. A value of '0' means infinity. 4412 * 4413 * Return: On success, 0 and the timeslice is in @interval. Otherwise, 4414 * an error code. 4415 */ 4416SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid, 4417 struct timespec __user *, interval) 4418{ 4419 struct task_struct *p; 4420 unsigned int time_slice; 4421 unsigned long flags; 4422 struct rq *rq; 4423 int retval; 4424 struct timespec t; 4425 4426 if (pid < 0) 4427 return -EINVAL; 4428 4429 retval = -ESRCH; 4430 rcu_read_lock(); 4431 p = find_process_by_pid(pid); 4432 if (!p) 4433 goto out_unlock; 4434 4435 retval = security_task_getscheduler(p); 4436 if (retval) 4437 goto out_unlock; 4438 4439 rq = task_rq_lock(p, &flags); 4440 time_slice = 0; 4441 if (p->sched_class->get_rr_interval) 4442 time_slice = p->sched_class->get_rr_interval(rq, p); 4443 task_rq_unlock(rq, p, &flags); 4444 4445 rcu_read_unlock(); 4446 jiffies_to_timespec(time_slice, &t); 4447 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0; 4448 return retval; 4449 4450out_unlock: 4451 rcu_read_unlock(); 4452 return retval; 4453} 4454 4455static const char stat_nam[] = TASK_STATE_TO_CHAR_STR; 4456 4457void sched_show_task(struct task_struct *p) 4458{ 4459 unsigned long free = 0; 4460 int ppid; 4461 unsigned state; 4462 4463 state = p->state ? __ffs(p->state) + 1 : 0; 4464 printk(KERN_INFO "%-15.15s %c", p->comm, 4465 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?'); 4466#if BITS_PER_LONG == 32 4467 if (state == TASK_RUNNING) 4468 printk(KERN_CONT " running "); 4469 else 4470 printk(KERN_CONT " %08lx ", thread_saved_pc(p)); 4471#else 4472 if (state == TASK_RUNNING) 4473 printk(KERN_CONT " running task "); 4474 else 4475 printk(KERN_CONT " %016lx ", thread_saved_pc(p)); 4476#endif 4477#ifdef CONFIG_DEBUG_STACK_USAGE 4478 free = stack_not_used(p); 4479#endif 4480 rcu_read_lock(); 4481 ppid = task_pid_nr(rcu_dereference(p->real_parent)); 4482 rcu_read_unlock(); 4483 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free, 4484 task_pid_nr(p), ppid, 4485 (unsigned long)task_thread_info(p)->flags); 4486 4487 print_worker_info(KERN_INFO, p); 4488 show_stack(p, NULL); 4489} 4490 4491void show_state_filter(unsigned long state_filter) 4492{ 4493 struct task_struct *g, *p; 4494 4495#if BITS_PER_LONG == 32 4496 printk(KERN_INFO 4497 " task PC stack pid father\n"); 4498#else 4499 printk(KERN_INFO 4500 " task PC stack pid father\n"); 4501#endif 4502 rcu_read_lock(); 4503 do_each_thread(g, p) { 4504 /* 4505 * reset the NMI-timeout, listing all files on a slow 4506 * console might take a lot of time: 4507 */ 4508 touch_nmi_watchdog(); 4509 if (!state_filter || (p->state & state_filter)) 4510 sched_show_task(p); 4511 } while_each_thread(g, p); 4512 4513 touch_all_softlockup_watchdogs(); 4514 4515#ifdef CONFIG_SCHED_DEBUG 4516 sysrq_sched_debug_show(); 4517#endif 4518 rcu_read_unlock(); 4519 /* 4520 * Only show locks if all tasks are dumped: 4521 */ 4522 if (!state_filter) 4523 debug_show_all_locks(); 4524} 4525 4526void init_idle_bootup_task(struct task_struct *idle) 4527{ 4528 idle->sched_class = &idle_sched_class; 4529} 4530 4531/** 4532 * init_idle - set up an idle thread for a given CPU 4533 * @idle: task in question 4534 * @cpu: cpu the idle task belongs to 4535 * 4536 * NOTE: this function does not set the idle thread's NEED_RESCHED 4537 * flag, to make booting more robust. 4538 */ 4539void init_idle(struct task_struct *idle, int cpu) 4540{ 4541 struct rq *rq = cpu_rq(cpu); 4542 unsigned long flags; 4543 4544 raw_spin_lock_irqsave(&rq->lock, flags); 4545 4546 __sched_fork(0, idle); 4547 idle->state = TASK_RUNNING; 4548 idle->se.exec_start = sched_clock(); 4549 4550 do_set_cpus_allowed(idle, cpumask_of(cpu)); 4551 /* 4552 * We're having a chicken and egg problem, even though we are 4553 * holding rq->lock, the cpu isn't yet set to this cpu so the 4554 * lockdep check in task_group() will fail. 4555 * 4556 * Similar case to sched_fork(). / Alternatively we could 4557 * use task_rq_lock() here and obtain the other rq->lock. 4558 * 4559 * Silence PROVE_RCU 4560 */ 4561 rcu_read_lock(); 4562 __set_task_cpu(idle, cpu); 4563 rcu_read_unlock(); 4564 4565 rq->curr = rq->idle = idle; 4566 idle->on_rq = 1; 4567#if defined(CONFIG_SMP) 4568 idle->on_cpu = 1; 4569#endif 4570 raw_spin_unlock_irqrestore(&rq->lock, flags); 4571 4572 /* Set the preempt count _outside_ the spinlocks! */ 4573 init_idle_preempt_count(idle, cpu); 4574 4575 /* 4576 * The idle tasks have their own, simple scheduling class: 4577 */ 4578 idle->sched_class = &idle_sched_class; 4579 ftrace_graph_init_idle_task(idle, cpu); 4580 vtime_init_idle(idle, cpu); 4581#if defined(CONFIG_SMP) 4582 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu); 4583#endif 4584} 4585 4586#ifdef CONFIG_SMP 4587void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask) 4588{ 4589 if (p->sched_class && p->sched_class->set_cpus_allowed) 4590 p->sched_class->set_cpus_allowed(p, new_mask); 4591 4592 cpumask_copy(&p->cpus_allowed, new_mask); 4593 p->nr_cpus_allowed = cpumask_weight(new_mask); 4594} 4595 4596/* 4597 * This is how migration works: 4598 * 4599 * 1) we invoke migration_cpu_stop() on the target CPU using 4600 * stop_one_cpu(). 4601 * 2) stopper starts to run (implicitly forcing the migrated thread 4602 * off the CPU) 4603 * 3) it checks whether the migrated task is still in the wrong runqueue. 4604 * 4) if it's in the wrong runqueue then the migration thread removes 4605 * it and puts it into the right queue. 4606 * 5) stopper completes and stop_one_cpu() returns and the migration 4607 * is done. 4608 */ 4609 4610/* 4611 * Change a given task's CPU affinity. Migrate the thread to a 4612 * proper CPU and schedule it away if the CPU it's executing on 4613 * is removed from the allowed bitmask. 4614 * 4615 * NOTE: the caller must have a valid reference to the task, the 4616 * task must not exit() & deallocate itself prematurely. The 4617 * call is not atomic; no spinlocks may be held. 4618 */ 4619int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask) 4620{ 4621 unsigned long flags; 4622 struct rq *rq; 4623 unsigned int dest_cpu; 4624 int ret = 0; 4625 4626 rq = task_rq_lock(p, &flags); 4627 4628 if (cpumask_equal(&p->cpus_allowed, new_mask)) 4629 goto out; 4630 4631 if (!cpumask_intersects(new_mask, cpu_active_mask)) { 4632 ret = -EINVAL; 4633 goto out; 4634 } 4635 4636 do_set_cpus_allowed(p, new_mask); 4637 4638 /* Can the task run on the task's current CPU? If so, we're done */ 4639 if (cpumask_test_cpu(task_cpu(p), new_mask)) 4640 goto out; 4641 4642 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask); 4643 if (p->on_rq) { 4644 struct migration_arg arg = { p, dest_cpu }; 4645 /* Need help from migration thread: drop lock and wait. */ 4646 task_rq_unlock(rq, p, &flags); 4647 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg); 4648 tlb_migrate_finish(p->mm); 4649 return 0; 4650 } 4651out: 4652 task_rq_unlock(rq, p, &flags); 4653 4654 return ret; 4655} 4656EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr); 4657 4658/* 4659 * Move (not current) task off this cpu, onto dest cpu. We're doing 4660 * this because either it can't run here any more (set_cpus_allowed() 4661 * away from this CPU, or CPU going down), or because we're 4662 * attempting to rebalance this task on exec (sched_exec). 4663 * 4664 * So we race with normal scheduler movements, but that's OK, as long 4665 * as the task is no longer on this CPU. 4666 * 4667 * Returns non-zero if task was successfully migrated. 4668 */ 4669static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu) 4670{ 4671 struct rq *rq_dest, *rq_src; 4672 int ret = 0; 4673 4674 if (unlikely(!cpu_active(dest_cpu))) 4675 return ret; 4676 4677 rq_src = cpu_rq(src_cpu); 4678 rq_dest = cpu_rq(dest_cpu); 4679 4680 raw_spin_lock(&p->pi_lock); 4681 double_rq_lock(rq_src, rq_dest); 4682 /* Already moved. */ 4683 if (task_cpu(p) != src_cpu) 4684 goto done; 4685 /* Affinity changed (again). */ 4686 if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p))) 4687 goto fail; 4688 4689 /* 4690 * If we're not on a rq, the next wake-up will ensure we're 4691 * placed properly. 4692 */ 4693 if (p->on_rq) { 4694 dequeue_task(rq_src, p, 0); 4695 set_task_cpu(p, dest_cpu); 4696 enqueue_task(rq_dest, p, 0); 4697 check_preempt_curr(rq_dest, p, 0); 4698 } 4699done: 4700 ret = 1; 4701fail: 4702 double_rq_unlock(rq_src, rq_dest); 4703 raw_spin_unlock(&p->pi_lock); 4704 return ret; 4705} 4706 4707#ifdef CONFIG_NUMA_BALANCING 4708/* Migrate current task p to target_cpu */ 4709int migrate_task_to(struct task_struct *p, int target_cpu) 4710{ 4711 struct migration_arg arg = { p, target_cpu }; 4712 int curr_cpu = task_cpu(p); 4713 4714 if (curr_cpu == target_cpu) 4715 return 0; 4716 4717 if (!cpumask_test_cpu(target_cpu, tsk_cpus_allowed(p))) 4718 return -EINVAL; 4719 4720 /* TODO: This is not properly updating schedstats */ 4721 4722 trace_sched_move_numa(p, curr_cpu, target_cpu); 4723 return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg); 4724} 4725 4726/* 4727 * Requeue a task on a given node and accurately track the number of NUMA 4728 * tasks on the runqueues 4729 */ 4730void sched_setnuma(struct task_struct *p, int nid) 4731{ 4732 struct rq *rq; 4733 unsigned long flags; 4734 bool on_rq, running; 4735 4736 rq = task_rq_lock(p, &flags); 4737 on_rq = p->on_rq; 4738 running = task_current(rq, p); 4739 4740 if (on_rq) 4741 dequeue_task(rq, p, 0); 4742 if (running) 4743 p->sched_class->put_prev_task(rq, p); 4744 4745 p->numa_preferred_nid = nid; 4746 4747 if (running) 4748 p->sched_class->set_curr_task(rq); 4749 if (on_rq) 4750 enqueue_task(rq, p, 0); 4751 task_rq_unlock(rq, p, &flags); 4752} 4753#endif 4754 4755/* 4756 * migration_cpu_stop - this will be executed by a highprio stopper thread 4757 * and performs thread migration by bumping thread off CPU then 4758 * 'pushing' onto another runqueue. 4759 */ 4760static int migration_cpu_stop(void *data) 4761{ 4762 struct migration_arg *arg = data; 4763 4764 /* 4765 * The original target cpu might have gone down and we might 4766 * be on another cpu but it doesn't matter. 4767 */ 4768 local_irq_disable(); 4769 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu); 4770 local_irq_enable(); 4771 return 0; 4772} 4773 4774#ifdef CONFIG_HOTPLUG_CPU 4775 4776/* 4777 * Ensures that the idle task is using init_mm right before its cpu goes 4778 * offline. 4779 */ 4780void idle_task_exit(void) 4781{ 4782 struct mm_struct *mm = current->active_mm; 4783 4784 BUG_ON(cpu_online(smp_processor_id())); 4785 4786 if (mm != &init_mm) { 4787 switch_mm(mm, &init_mm, current); 4788 finish_arch_post_lock_switch(); 4789 } 4790 mmdrop(mm); 4791} 4792 4793/* 4794 * Since this CPU is going 'away' for a while, fold any nr_active delta 4795 * we might have. Assumes we're called after migrate_tasks() so that the 4796 * nr_active count is stable. 4797 * 4798 * Also see the comment "Global load-average calculations". 4799 */ 4800static void calc_load_migrate(struct rq *rq) 4801{ 4802 long delta = calc_load_fold_active(rq); 4803 if (delta) 4804 atomic_long_add(delta, &calc_load_tasks); 4805} 4806 4807static void put_prev_task_fake(struct rq *rq, struct task_struct *prev) 4808{ 4809} 4810 4811static const struct sched_class fake_sched_class = { 4812 .put_prev_task = put_prev_task_fake, 4813}; 4814 4815static struct task_struct fake_task = { 4816 /* 4817 * Avoid pull_{rt,dl}_task() 4818 */ 4819 .prio = MAX_PRIO + 1, 4820 .sched_class = &fake_sched_class, 4821}; 4822 4823/* 4824 * Migrate all tasks from the rq, sleeping tasks will be migrated by 4825 * try_to_wake_up()->select_task_rq(). 4826 * 4827 * Called with rq->lock held even though we'er in stop_machine() and 4828 * there's no concurrency possible, we hold the required locks anyway 4829 * because of lock validation efforts. 4830 */ 4831static void migrate_tasks(unsigned int dead_cpu) 4832{ 4833 struct rq *rq = cpu_rq(dead_cpu); 4834 struct task_struct *next, *stop = rq->stop; 4835 int dest_cpu; 4836 4837 /* 4838 * Fudge the rq selection such that the below task selection loop 4839 * doesn't get stuck on the currently eligible stop task. 4840 * 4841 * We're currently inside stop_machine() and the rq is either stuck 4842 * in the stop_machine_cpu_stop() loop, or we're executing this code, 4843 * either way we should never end up calling schedule() until we're 4844 * done here. 4845 */ 4846 rq->stop = NULL; 4847 4848 /* 4849 * put_prev_task() and pick_next_task() sched 4850 * class method both need to have an up-to-date 4851 * value of rq->clock[_task] 4852 */ 4853 update_rq_clock(rq); 4854 4855 for ( ; ; ) { 4856 /* 4857 * There's this thread running, bail when that's the only 4858 * remaining thread. 4859 */ 4860 if (rq->nr_running == 1) 4861 break; 4862 4863 next = pick_next_task(rq, &fake_task); 4864 BUG_ON(!next); 4865 next->sched_class->put_prev_task(rq, next); 4866 4867 /* Find suitable destination for @next, with force if needed. */ 4868 dest_cpu = select_fallback_rq(dead_cpu, next); 4869 raw_spin_unlock(&rq->lock); 4870 4871 __migrate_task(next, dead_cpu, dest_cpu); 4872 4873 raw_spin_lock(&rq->lock); 4874 } 4875 4876 rq->stop = stop; 4877} 4878 4879#endif /* CONFIG_HOTPLUG_CPU */ 4880 4881#if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL) 4882 4883static struct ctl_table sd_ctl_dir[] = { 4884 { 4885 .procname = "sched_domain", 4886 .mode = 0555, 4887 }, 4888 {} 4889}; 4890 4891static struct ctl_table sd_ctl_root[] = { 4892 { 4893 .procname = "kernel", 4894 .mode = 0555, 4895 .child = sd_ctl_dir, 4896 }, 4897 {} 4898}; 4899 4900static struct ctl_table *sd_alloc_ctl_entry(int n) 4901{ 4902 struct ctl_table *entry = 4903 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL); 4904 4905 return entry; 4906} 4907 4908static void sd_free_ctl_entry(struct ctl_table **tablep) 4909{ 4910 struct ctl_table *entry; 4911 4912 /* 4913 * In the intermediate directories, both the child directory and 4914 * procname are dynamically allocated and could fail but the mode 4915 * will always be set. In the lowest directory the names are 4916 * static strings and all have proc handlers. 4917 */ 4918 for (entry = *tablep; entry->mode; entry++) { 4919 if (entry->child) 4920 sd_free_ctl_entry(&entry->child); 4921 if (entry->proc_handler == NULL) 4922 kfree(entry->procname); 4923 } 4924 4925 kfree(*tablep); 4926 *tablep = NULL; 4927} 4928 4929static int min_load_idx = 0; 4930static int max_load_idx = CPU_LOAD_IDX_MAX-1; 4931 4932static void 4933set_table_entry(struct ctl_table *entry, 4934 const char *procname, void *data, int maxlen, 4935 umode_t mode, proc_handler *proc_handler, 4936 bool load_idx) 4937{ 4938 entry->procname = procname; 4939 entry->data = data; 4940 entry->maxlen = maxlen; 4941 entry->mode = mode; 4942 entry->proc_handler = proc_handler; 4943 4944 if (load_idx) { 4945 entry->extra1 = &min_load_idx; 4946 entry->extra2 = &max_load_idx; 4947 } 4948} 4949 4950static struct ctl_table * 4951sd_alloc_ctl_domain_table(struct sched_domain *sd) 4952{ 4953 struct ctl_table *table = sd_alloc_ctl_entry(14); 4954 4955 if (table == NULL) 4956 return NULL; 4957 4958 set_table_entry(&table[0], "min_interval", &sd->min_interval, 4959 sizeof(long), 0644, proc_doulongvec_minmax, false); 4960 set_table_entry(&table[1], "max_interval", &sd->max_interval, 4961 sizeof(long), 0644, proc_doulongvec_minmax, false); 4962 set_table_entry(&table[2], "busy_idx", &sd->busy_idx, 4963 sizeof(int), 0644, proc_dointvec_minmax, true); 4964 set_table_entry(&table[3], "idle_idx", &sd->idle_idx, 4965 sizeof(int), 0644, proc_dointvec_minmax, true); 4966 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx, 4967 sizeof(int), 0644, proc_dointvec_minmax, true); 4968 set_table_entry(&table[5], "wake_idx", &sd->wake_idx, 4969 sizeof(int), 0644, proc_dointvec_minmax, true); 4970 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx, 4971 sizeof(int), 0644, proc_dointvec_minmax, true); 4972 set_table_entry(&table[7], "busy_factor", &sd->busy_factor, 4973 sizeof(int), 0644, proc_dointvec_minmax, false); 4974 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct, 4975 sizeof(int), 0644, proc_dointvec_minmax, false); 4976 set_table_entry(&table[9], "cache_nice_tries", 4977 &sd->cache_nice_tries, 4978 sizeof(int), 0644, proc_dointvec_minmax, false); 4979 set_table_entry(&table[10], "flags", &sd->flags, 4980 sizeof(int), 0644, proc_dointvec_minmax, false); 4981 set_table_entry(&table[11], "max_newidle_lb_cost", 4982 &sd->max_newidle_lb_cost, 4983 sizeof(long), 0644, proc_doulongvec_minmax, false); 4984 set_table_entry(&table[12], "name", sd->name, 4985 CORENAME_MAX_SIZE, 0444, proc_dostring, false); 4986 /* &table[13] is terminator */ 4987 4988 return table; 4989} 4990 4991static struct ctl_table *sd_alloc_ctl_cpu_table(int cpu) 4992{ 4993 struct ctl_table *entry, *table; 4994 struct sched_domain *sd; 4995 int domain_num = 0, i; 4996 char buf[32]; 4997 4998 for_each_domain(cpu, sd) 4999 domain_num++; 5000 entry = table = sd_alloc_ctl_entry(domain_num + 1); 5001 if (table == NULL) 5002 return NULL; 5003 5004 i = 0; 5005 for_each_domain(cpu, sd) { 5006 snprintf(buf, 32, "domain%d", i); 5007 entry->procname = kstrdup(buf, GFP_KERNEL); 5008 entry->mode = 0555; 5009 entry->child = sd_alloc_ctl_domain_table(sd); 5010 entry++; 5011 i++; 5012 } 5013 return table; 5014} 5015 5016static struct ctl_table_header *sd_sysctl_header; 5017static void register_sched_domain_sysctl(void) 5018{ 5019 int i, cpu_num = num_possible_cpus(); 5020 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1); 5021 char buf[32]; 5022 5023 WARN_ON(sd_ctl_dir[0].child); 5024 sd_ctl_dir[0].child = entry; 5025 5026 if (entry == NULL) 5027 return; 5028 5029 for_each_possible_cpu(i) { 5030 snprintf(buf, 32, "cpu%d", i); 5031 entry->procname = kstrdup(buf, GFP_KERNEL); 5032 entry->mode = 0555; 5033 entry->child = sd_alloc_ctl_cpu_table(i); 5034 entry++; 5035 } 5036 5037 WARN_ON(sd_sysctl_header); 5038 sd_sysctl_header = register_sysctl_table(sd_ctl_root); 5039} 5040 5041/* may be called multiple times per register */ 5042static void unregister_sched_domain_sysctl(void) 5043{ 5044 if (sd_sysctl_header) 5045 unregister_sysctl_table(sd_sysctl_header); 5046 sd_sysctl_header = NULL; 5047 if (sd_ctl_dir[0].child) 5048 sd_free_ctl_entry(&sd_ctl_dir[0].child); 5049} 5050#else 5051static void register_sched_domain_sysctl(void) 5052{ 5053} 5054static void unregister_sched_domain_sysctl(void) 5055{ 5056} 5057#endif 5058 5059static void set_rq_online(struct rq *rq) 5060{ 5061 if (!rq->online) { 5062 const struct sched_class *class; 5063 5064 cpumask_set_cpu(rq->cpu, rq->rd->online); 5065 rq->online = 1; 5066 5067 for_each_class(class) { 5068 if (class->rq_online) 5069 class->rq_online(rq); 5070 } 5071 } 5072} 5073 5074static void set_rq_offline(struct rq *rq) 5075{ 5076 if (rq->online) { 5077 const struct sched_class *class; 5078 5079 for_each_class(class) { 5080 if (class->rq_offline) 5081 class->rq_offline(rq); 5082 } 5083 5084 cpumask_clear_cpu(rq->cpu, rq->rd->online); 5085 rq->online = 0; 5086 } 5087} 5088 5089/* 5090 * migration_call - callback that gets triggered when a CPU is added. 5091 * Here we can start up the necessary migration thread for the new CPU. 5092 */ 5093static int 5094migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu) 5095{ 5096 int cpu = (long)hcpu; 5097 unsigned long flags; 5098 struct rq *rq = cpu_rq(cpu); 5099 5100 switch (action & ~CPU_TASKS_FROZEN) { 5101 5102 case CPU_UP_PREPARE: 5103 rq->calc_load_update = calc_load_update; 5104 break; 5105 5106 case CPU_ONLINE: 5107 /* Update our root-domain */ 5108 raw_spin_lock_irqsave(&rq->lock, flags); 5109 if (rq->rd) { 5110 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span)); 5111 5112 set_rq_online(rq); 5113 } 5114 raw_spin_unlock_irqrestore(&rq->lock, flags); 5115 break; 5116 5117#ifdef CONFIG_HOTPLUG_CPU 5118 case CPU_DYING: 5119 sched_ttwu_pending(); 5120 /* Update our root-domain */ 5121 raw_spin_lock_irqsave(&rq->lock, flags); 5122 if (rq->rd) { 5123 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span)); 5124 set_rq_offline(rq); 5125 } 5126 migrate_tasks(cpu); 5127 BUG_ON(rq->nr_running != 1); /* the migration thread */ 5128 raw_spin_unlock_irqrestore(&rq->lock, flags); 5129 break; 5130 5131 case CPU_DEAD: 5132 calc_load_migrate(rq); 5133 break; 5134#endif 5135 } 5136 5137 update_max_interval(); 5138 5139 return NOTIFY_OK; 5140} 5141 5142/* 5143 * Register at high priority so that task migration (migrate_all_tasks) 5144 * happens before everything else. This has to be lower priority than 5145 * the notifier in the perf_event subsystem, though. 5146 */ 5147static struct notifier_block migration_notifier = { 5148 .notifier_call = migration_call, 5149 .priority = CPU_PRI_MIGRATION, 5150}; 5151 5152static void __cpuinit set_cpu_rq_start_time(void) 5153{ 5154 int cpu = smp_processor_id(); 5155 struct rq *rq = cpu_rq(cpu); 5156 rq->age_stamp = sched_clock_cpu(cpu); 5157} 5158 5159static int sched_cpu_active(struct notifier_block *nfb, 5160 unsigned long action, void *hcpu) 5161{ 5162 switch (action & ~CPU_TASKS_FROZEN) { 5163 case CPU_STARTING: 5164 set_cpu_rq_start_time(); 5165 return NOTIFY_OK; 5166 case CPU_DOWN_FAILED: 5167 set_cpu_active((long)hcpu, true); 5168 return NOTIFY_OK; 5169 default: 5170 return NOTIFY_DONE; 5171 } 5172} 5173 5174static int sched_cpu_inactive(struct notifier_block *nfb, 5175 unsigned long action, void *hcpu) 5176{ 5177 unsigned long flags; 5178 long cpu = (long)hcpu; 5179 5180 switch (action & ~CPU_TASKS_FROZEN) { 5181 case CPU_DOWN_PREPARE: 5182 set_cpu_active(cpu, false); 5183 5184 /* explicitly allow suspend */ 5185 if (!(action & CPU_TASKS_FROZEN)) { 5186 struct dl_bw *dl_b = dl_bw_of(cpu); 5187 bool overflow; 5188 int cpus; 5189 5190 raw_spin_lock_irqsave(&dl_b->lock, flags); 5191 cpus = dl_bw_cpus(cpu); 5192 overflow = __dl_overflow(dl_b, cpus, 0, 0); 5193 raw_spin_unlock_irqrestore(&dl_b->lock, flags); 5194 5195 if (overflow) 5196 return notifier_from_errno(-EBUSY); 5197 } 5198 return NOTIFY_OK; 5199 } 5200 5201 return NOTIFY_DONE; 5202} 5203 5204static int __init migration_init(void) 5205{ 5206 void *cpu = (void *)(long)smp_processor_id(); 5207 int err; 5208 5209 /* Initialize migration for the boot CPU */ 5210 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu); 5211 BUG_ON(err == NOTIFY_BAD); 5212 migration_call(&migration_notifier, CPU_ONLINE, cpu); 5213 register_cpu_notifier(&migration_notifier); 5214 5215 /* Register cpu active notifiers */ 5216 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE); 5217 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE); 5218 5219 return 0; 5220} 5221early_initcall(migration_init); 5222#endif 5223 5224#ifdef CONFIG_SMP 5225 5226static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */ 5227 5228#ifdef CONFIG_SCHED_DEBUG 5229 5230static __read_mostly int sched_debug_enabled; 5231 5232static int __init sched_debug_setup(char *str) 5233{ 5234 sched_debug_enabled = 1; 5235 5236 return 0; 5237} 5238early_param("sched_debug", sched_debug_setup); 5239 5240static inline bool sched_debug(void) 5241{ 5242 return sched_debug_enabled; 5243} 5244 5245static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level, 5246 struct cpumask *groupmask) 5247{ 5248 struct sched_group *group = sd->groups; 5249 char str[256]; 5250 5251 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd)); 5252 cpumask_clear(groupmask); 5253 5254 printk(KERN_DEBUG "%*s domain %d: ", level, "", level); 5255 5256 if (!(sd->flags & SD_LOAD_BALANCE)) { 5257 printk("does not load-balance\n"); 5258 if (sd->parent) 5259 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain" 5260 " has parent"); 5261 return -1; 5262 } 5263 5264 printk(KERN_CONT "span %s level %s\n", str, sd->name); 5265 5266 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) { 5267 printk(KERN_ERR "ERROR: domain->span does not contain " 5268 "CPU%d\n", cpu); 5269 } 5270 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) { 5271 printk(KERN_ERR "ERROR: domain->groups does not contain" 5272 " CPU%d\n", cpu); 5273 } 5274 5275 printk(KERN_DEBUG "%*s groups:", level + 1, ""); 5276 do { 5277 if (!group) { 5278 printk("\n"); 5279 printk(KERN_ERR "ERROR: group is NULL\n"); 5280 break; 5281 } 5282 5283 /* 5284 * Even though we initialize ->capacity to something semi-sane, 5285 * we leave capacity_orig unset. This allows us to detect if 5286 * domain iteration is still funny without causing /0 traps. 5287 */ 5288 if (!group->sgc->capacity_orig) { 5289 printk(KERN_CONT "\n"); 5290 printk(KERN_ERR "ERROR: domain->cpu_capacity not set\n"); 5291 break; 5292 } 5293 5294 if (!cpumask_weight(sched_group_cpus(group))) { 5295 printk(KERN_CONT "\n"); 5296 printk(KERN_ERR "ERROR: empty group\n"); 5297 break; 5298 } 5299 5300 if (!(sd->flags & SD_OVERLAP) && 5301 cpumask_intersects(groupmask, sched_group_cpus(group))) { 5302 printk(KERN_CONT "\n"); 5303 printk(KERN_ERR "ERROR: repeated CPUs\n"); 5304 break; 5305 } 5306 5307 cpumask_or(groupmask, groupmask, sched_group_cpus(group)); 5308 5309 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group)); 5310 5311 printk(KERN_CONT " %s", str); 5312 if (group->sgc->capacity != SCHED_CAPACITY_SCALE) { 5313 printk(KERN_CONT " (cpu_capacity = %d)", 5314 group->sgc->capacity); 5315 } 5316 5317 group = group->next; 5318 } while (group != sd->groups); 5319 printk(KERN_CONT "\n"); 5320 5321 if (!cpumask_equal(sched_domain_span(sd), groupmask)) 5322 printk(KERN_ERR "ERROR: groups don't span domain->span\n"); 5323 5324 if (sd->parent && 5325 !cpumask_subset(groupmask, sched_domain_span(sd->parent))) 5326 printk(KERN_ERR "ERROR: parent span is not a superset " 5327 "of domain->span\n"); 5328 return 0; 5329} 5330 5331static void sched_domain_debug(struct sched_domain *sd, int cpu) 5332{ 5333 int level = 0; 5334 5335 if (!sched_debug_enabled) 5336 return; 5337 5338 if (!sd) { 5339 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu); 5340 return; 5341 } 5342 5343 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu); 5344 5345 for (;;) { 5346 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask)) 5347 break; 5348 level++; 5349 sd = sd->parent; 5350 if (!sd) 5351 break; 5352 } 5353} 5354#else /* !CONFIG_SCHED_DEBUG */ 5355# define sched_domain_debug(sd, cpu) do { } while (0) 5356static inline bool sched_debug(void) 5357{ 5358 return false; 5359} 5360#endif /* CONFIG_SCHED_DEBUG */ 5361 5362static int sd_degenerate(struct sched_domain *sd) 5363{ 5364 if (cpumask_weight(sched_domain_span(sd)) == 1) 5365 return 1; 5366 5367 /* Following flags need at least 2 groups */ 5368 if (sd->flags & (SD_LOAD_BALANCE | 5369 SD_BALANCE_NEWIDLE | 5370 SD_BALANCE_FORK | 5371 SD_BALANCE_EXEC | 5372 SD_SHARE_CPUCAPACITY | 5373 SD_SHARE_PKG_RESOURCES | 5374 SD_SHARE_POWERDOMAIN)) { 5375 if (sd->groups != sd->groups->next) 5376 return 0; 5377 } 5378 5379 /* Following flags don't use groups */ 5380 if (sd->flags & (SD_WAKE_AFFINE)) 5381 return 0; 5382 5383 return 1; 5384} 5385 5386static int 5387sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent) 5388{ 5389 unsigned long cflags = sd->flags, pflags = parent->flags; 5390 5391 if (sd_degenerate(parent)) 5392 return 1; 5393 5394 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent))) 5395 return 0; 5396 5397 /* Flags needing groups don't count if only 1 group in parent */ 5398 if (parent->groups == parent->groups->next) { 5399 pflags &= ~(SD_LOAD_BALANCE | 5400 SD_BALANCE_NEWIDLE | 5401 SD_BALANCE_FORK | 5402 SD_BALANCE_EXEC | 5403 SD_SHARE_CPUCAPACITY | 5404 SD_SHARE_PKG_RESOURCES | 5405 SD_PREFER_SIBLING | 5406 SD_SHARE_POWERDOMAIN); 5407 if (nr_node_ids == 1) 5408 pflags &= ~SD_SERIALIZE; 5409 } 5410 if (~cflags & pflags) 5411 return 0; 5412 5413 return 1; 5414} 5415 5416static void free_rootdomain(struct rcu_head *rcu) 5417{ 5418 struct root_domain *rd = container_of(rcu, struct root_domain, rcu); 5419 5420 cpupri_cleanup(&rd->cpupri); 5421 cpudl_cleanup(&rd->cpudl); 5422 free_cpumask_var(rd->dlo_mask); 5423 free_cpumask_var(rd->rto_mask); 5424 free_cpumask_var(rd->online); 5425 free_cpumask_var(rd->span); 5426 kfree(rd); 5427} 5428 5429static void rq_attach_root(struct rq *rq, struct root_domain *rd) 5430{ 5431 struct root_domain *old_rd = NULL; 5432 unsigned long flags; 5433 5434 raw_spin_lock_irqsave(&rq->lock, flags); 5435 5436 if (rq->rd) { 5437 old_rd = rq->rd; 5438 5439 if (cpumask_test_cpu(rq->cpu, old_rd->online)) 5440 set_rq_offline(rq); 5441 5442 cpumask_clear_cpu(rq->cpu, old_rd->span); 5443 5444 /* 5445 * If we dont want to free the old_rd yet then 5446 * set old_rd to NULL to skip the freeing later 5447 * in this function: 5448 */ 5449 if (!atomic_dec_and_test(&old_rd->refcount)) 5450 old_rd = NULL; 5451 } 5452 5453 atomic_inc(&rd->refcount); 5454 rq->rd = rd; 5455 5456 cpumask_set_cpu(rq->cpu, rd->span); 5457 if (cpumask_test_cpu(rq->cpu, cpu_active_mask)) 5458 set_rq_online(rq); 5459 5460 raw_spin_unlock_irqrestore(&rq->lock, flags); 5461 5462 if (old_rd) 5463 call_rcu_sched(&old_rd->rcu, free_rootdomain); 5464} 5465 5466static int init_rootdomain(struct root_domain *rd) 5467{ 5468 memset(rd, 0, sizeof(*rd)); 5469 5470 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL)) 5471 goto out; 5472 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL)) 5473 goto free_span; 5474 if (!alloc_cpumask_var(&rd->dlo_mask, GFP_KERNEL)) 5475 goto free_online; 5476 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL)) 5477 goto free_dlo_mask; 5478 5479 init_dl_bw(&rd->dl_bw); 5480 if (cpudl_init(&rd->cpudl) != 0) 5481 goto free_dlo_mask; 5482 5483 if (cpupri_init(&rd->cpupri) != 0) 5484 goto free_rto_mask; 5485 return 0; 5486 5487free_rto_mask: 5488 free_cpumask_var(rd->rto_mask); 5489free_dlo_mask: 5490 free_cpumask_var(rd->dlo_mask); 5491free_online: 5492 free_cpumask_var(rd->online); 5493free_span: 5494 free_cpumask_var(rd->span); 5495out: 5496 return -ENOMEM; 5497} 5498 5499/* 5500 * By default the system creates a single root-domain with all cpus as 5501 * members (mimicking the global state we have today). 5502 */ 5503struct root_domain def_root_domain; 5504 5505static void init_defrootdomain(void) 5506{ 5507 init_rootdomain(&def_root_domain); 5508 5509 atomic_set(&def_root_domain.refcount, 1); 5510} 5511 5512static struct root_domain *alloc_rootdomain(void) 5513{ 5514 struct root_domain *rd; 5515 5516 rd = kmalloc(sizeof(*rd), GFP_KERNEL); 5517 if (!rd) 5518 return NULL; 5519 5520 if (init_rootdomain(rd) != 0) { 5521 kfree(rd); 5522 return NULL; 5523 } 5524 5525 return rd; 5526} 5527 5528static void free_sched_groups(struct sched_group *sg, int free_sgc) 5529{ 5530 struct sched_group *tmp, *first; 5531 5532 if (!sg) 5533 return; 5534 5535 first = sg; 5536 do { 5537 tmp = sg->next; 5538 5539 if (free_sgc && atomic_dec_and_test(&sg->sgc->ref)) 5540 kfree(sg->sgc); 5541 5542 kfree(sg); 5543 sg = tmp; 5544 } while (sg != first); 5545} 5546 5547static void free_sched_domain(struct rcu_head *rcu) 5548{ 5549 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu); 5550 5551 /* 5552 * If its an overlapping domain it has private groups, iterate and 5553 * nuke them all. 5554 */ 5555 if (sd->flags & SD_OVERLAP) { 5556 free_sched_groups(sd->groups, 1); 5557 } else if (atomic_dec_and_test(&sd->groups->ref)) { 5558 kfree(sd->groups->sgc); 5559 kfree(sd->groups); 5560 } 5561 kfree(sd); 5562} 5563 5564static void destroy_sched_domain(struct sched_domain *sd, int cpu) 5565{ 5566 call_rcu(&sd->rcu, free_sched_domain); 5567} 5568 5569static void destroy_sched_domains(struct sched_domain *sd, int cpu) 5570{ 5571 for (; sd; sd = sd->parent) 5572 destroy_sched_domain(sd, cpu); 5573} 5574 5575/* 5576 * Keep a special pointer to the highest sched_domain that has 5577 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this 5578 * allows us to avoid some pointer chasing select_idle_sibling(). 5579 * 5580 * Also keep a unique ID per domain (we use the first cpu number in 5581 * the cpumask of the domain), this allows us to quickly tell if 5582 * two cpus are in the same cache domain, see cpus_share_cache(). 5583 */ 5584DEFINE_PER_CPU(struct sched_domain *, sd_llc); 5585DEFINE_PER_CPU(int, sd_llc_size); 5586DEFINE_PER_CPU(int, sd_llc_id); 5587DEFINE_PER_CPU(struct sched_domain *, sd_numa); 5588DEFINE_PER_CPU(struct sched_domain *, sd_busy); 5589DEFINE_PER_CPU(struct sched_domain *, sd_asym); 5590 5591static void update_top_cache_domain(int cpu) 5592{ 5593 struct sched_domain *sd; 5594 struct sched_domain *busy_sd = NULL; 5595 int id = cpu; 5596 int size = 1; 5597 5598 sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES); 5599 if (sd) { 5600 id = cpumask_first(sched_domain_span(sd)); 5601 size = cpumask_weight(sched_domain_span(sd)); 5602 busy_sd = sd->parent; /* sd_busy */ 5603 } 5604 rcu_assign_pointer(per_cpu(sd_busy, cpu), busy_sd); 5605 5606 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd); 5607 per_cpu(sd_llc_size, cpu) = size; 5608 per_cpu(sd_llc_id, cpu) = id; 5609 5610 sd = lowest_flag_domain(cpu, SD_NUMA); 5611 rcu_assign_pointer(per_cpu(sd_numa, cpu), sd); 5612 5613 sd = highest_flag_domain(cpu, SD_ASYM_PACKING); 5614 rcu_assign_pointer(per_cpu(sd_asym, cpu), sd); 5615} 5616 5617/* 5618 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must 5619 * hold the hotplug lock. 5620 */ 5621static void 5622cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu) 5623{ 5624 struct rq *rq = cpu_rq(cpu); 5625 struct sched_domain *tmp; 5626 5627 /* Remove the sched domains which do not contribute to scheduling. */ 5628 for (tmp = sd; tmp; ) { 5629 struct sched_domain *parent = tmp->parent; 5630 if (!parent) 5631 break; 5632 5633 if (sd_parent_degenerate(tmp, parent)) { 5634 tmp->parent = parent->parent; 5635 if (parent->parent) 5636 parent->parent->child = tmp; 5637 /* 5638 * Transfer SD_PREFER_SIBLING down in case of a 5639 * degenerate parent; the spans match for this 5640 * so the property transfers. 5641 */ 5642 if (parent->flags & SD_PREFER_SIBLING) 5643 tmp->flags |= SD_PREFER_SIBLING; 5644 destroy_sched_domain(parent, cpu); 5645 } else 5646 tmp = tmp->parent; 5647 } 5648 5649 if (sd && sd_degenerate(sd)) { 5650 tmp = sd; 5651 sd = sd->parent; 5652 destroy_sched_domain(tmp, cpu); 5653 if (sd) 5654 sd->child = NULL; 5655 } 5656 5657 sched_domain_debug(sd, cpu); 5658 5659 rq_attach_root(rq, rd); 5660 tmp = rq->sd; 5661 rcu_assign_pointer(rq->sd, sd); 5662 destroy_sched_domains(tmp, cpu); 5663 5664 update_top_cache_domain(cpu); 5665} 5666 5667/* cpus with isolated domains */ 5668static cpumask_var_t cpu_isolated_map; 5669 5670/* Setup the mask of cpus configured for isolated domains */ 5671static int __init isolated_cpu_setup(char *str) 5672{ 5673 alloc_bootmem_cpumask_var(&cpu_isolated_map); 5674 cpulist_parse(str, cpu_isolated_map); 5675 return 1; 5676} 5677 5678__setup("isolcpus=", isolated_cpu_setup); 5679 5680struct s_data { 5681 struct sched_domain ** __percpu sd; 5682 struct root_domain *rd; 5683}; 5684 5685enum s_alloc { 5686 sa_rootdomain, 5687 sa_sd, 5688 sa_sd_storage, 5689 sa_none, 5690}; 5691 5692/* 5693 * Build an iteration mask that can exclude certain CPUs from the upwards 5694 * domain traversal. 5695 * 5696 * Asymmetric node setups can result in situations where the domain tree is of 5697 * unequal depth, make sure to skip domains that already cover the entire 5698 * range. 5699 * 5700 * In that case build_sched_domains() will have terminated the iteration early 5701 * and our sibling sd spans will be empty. Domains should always include the 5702 * cpu they're built on, so check that. 5703 * 5704 */ 5705static void build_group_mask(struct sched_domain *sd, struct sched_group *sg) 5706{ 5707 const struct cpumask *span = sched_domain_span(sd); 5708 struct sd_data *sdd = sd->private; 5709 struct sched_domain *sibling; 5710 int i; 5711 5712 for_each_cpu(i, span) { 5713 sibling = *per_cpu_ptr(sdd->sd, i); 5714 if (!cpumask_test_cpu(i, sched_domain_span(sibling))) 5715 continue; 5716 5717 cpumask_set_cpu(i, sched_group_mask(sg)); 5718 } 5719} 5720 5721/* 5722 * Return the canonical balance cpu for this group, this is the first cpu 5723 * of this group that's also in the iteration mask. 5724 */ 5725int group_balance_cpu(struct sched_group *sg) 5726{ 5727 return cpumask_first_and(sched_group_cpus(sg), sched_group_mask(sg)); 5728} 5729 5730static int 5731build_overlap_sched_groups(struct sched_domain *sd, int cpu) 5732{ 5733 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg; 5734 const struct cpumask *span = sched_domain_span(sd); 5735 struct cpumask *covered = sched_domains_tmpmask; 5736 struct sd_data *sdd = sd->private; 5737 struct sched_domain *child; 5738 int i; 5739 5740 cpumask_clear(covered); 5741 5742 for_each_cpu(i, span) { 5743 struct cpumask *sg_span; 5744 5745 if (cpumask_test_cpu(i, covered)) 5746 continue; 5747 5748 child = *per_cpu_ptr(sdd->sd, i); 5749 5750 /* See the comment near build_group_mask(). */ 5751 if (!cpumask_test_cpu(i, sched_domain_span(child))) 5752 continue; 5753 5754 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(), 5755 GFP_KERNEL, cpu_to_node(cpu)); 5756 5757 if (!sg) 5758 goto fail; 5759 5760 sg_span = sched_group_cpus(sg); 5761 if (child->child) { 5762 child = child->child; 5763 cpumask_copy(sg_span, sched_domain_span(child)); 5764 } else 5765 cpumask_set_cpu(i, sg_span); 5766 5767 cpumask_or(covered, covered, sg_span); 5768 5769 sg->sgc = *per_cpu_ptr(sdd->sgc, i); 5770 if (atomic_inc_return(&sg->sgc->ref) == 1) 5771 build_group_mask(sd, sg); 5772 5773 /* 5774 * Initialize sgc->capacity such that even if we mess up the 5775 * domains and no possible iteration will get us here, we won't 5776 * die on a /0 trap. 5777 */ 5778 sg->sgc->capacity = SCHED_CAPACITY_SCALE * cpumask_weight(sg_span); 5779 sg->sgc->capacity_orig = sg->sgc->capacity; 5780 5781 /* 5782 * Make sure the first group of this domain contains the 5783 * canonical balance cpu. Otherwise the sched_domain iteration 5784 * breaks. See update_sg_lb_stats(). 5785 */ 5786 if ((!groups && cpumask_test_cpu(cpu, sg_span)) || 5787 group_balance_cpu(sg) == cpu) 5788 groups = sg; 5789 5790 if (!first) 5791 first = sg; 5792 if (last) 5793 last->next = sg; 5794 last = sg; 5795 last->next = first; 5796 } 5797 sd->groups = groups; 5798 5799 return 0; 5800 5801fail: 5802 free_sched_groups(first, 0); 5803 5804 return -ENOMEM; 5805} 5806 5807static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg) 5808{ 5809 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu); 5810 struct sched_domain *child = sd->child; 5811 5812 if (child) 5813 cpu = cpumask_first(sched_domain_span(child)); 5814 5815 if (sg) { 5816 *sg = *per_cpu_ptr(sdd->sg, cpu); 5817 (*sg)->sgc = *per_cpu_ptr(sdd->sgc, cpu); 5818 atomic_set(&(*sg)->sgc->ref, 1); /* for claim_allocations */ 5819 } 5820 5821 return cpu; 5822} 5823 5824/* 5825 * build_sched_groups will build a circular linked list of the groups 5826 * covered by the given span, and will set each group's ->cpumask correctly, 5827 * and ->cpu_capacity to 0. 5828 * 5829 * Assumes the sched_domain tree is fully constructed 5830 */ 5831static int 5832build_sched_groups(struct sched_domain *sd, int cpu) 5833{ 5834 struct sched_group *first = NULL, *last = NULL; 5835 struct sd_data *sdd = sd->private; 5836 const struct cpumask *span = sched_domain_span(sd); 5837 struct cpumask *covered; 5838 int i; 5839 5840 get_group(cpu, sdd, &sd->groups); 5841 atomic_inc(&sd->groups->ref); 5842 5843 if (cpu != cpumask_first(span)) 5844 return 0; 5845 5846 lockdep_assert_held(&sched_domains_mutex); 5847 covered = sched_domains_tmpmask; 5848 5849 cpumask_clear(covered); 5850 5851 for_each_cpu(i, span) { 5852 struct sched_group *sg; 5853 int group, j; 5854 5855 if (cpumask_test_cpu(i, covered)) 5856 continue; 5857 5858 group = get_group(i, sdd, &sg); 5859 cpumask_setall(sched_group_mask(sg)); 5860 5861 for_each_cpu(j, span) { 5862 if (get_group(j, sdd, NULL) != group) 5863 continue; 5864 5865 cpumask_set_cpu(j, covered); 5866 cpumask_set_cpu(j, sched_group_cpus(sg)); 5867 } 5868 5869 if (!first) 5870 first = sg; 5871 if (last) 5872 last->next = sg; 5873 last = sg; 5874 } 5875 last->next = first; 5876 5877 return 0; 5878} 5879 5880/* 5881 * Initialize sched groups cpu_capacity. 5882 * 5883 * cpu_capacity indicates the capacity of sched group, which is used while 5884 * distributing the load between different sched groups in a sched domain. 5885 * Typically cpu_capacity for all the groups in a sched domain will be same 5886 * unless there are asymmetries in the topology. If there are asymmetries, 5887 * group having more cpu_capacity will pickup more load compared to the 5888 * group having less cpu_capacity. 5889 */ 5890static void init_sched_groups_capacity(int cpu, struct sched_domain *sd) 5891{ 5892 struct sched_group *sg = sd->groups; 5893 5894 WARN_ON(!sg); 5895 5896 do { 5897 sg->group_weight = cpumask_weight(sched_group_cpus(sg)); 5898 sg = sg->next; 5899 } while (sg != sd->groups); 5900 5901 if (cpu != group_balance_cpu(sg)) 5902 return; 5903 5904 update_group_capacity(sd, cpu); 5905 atomic_set(&sg->sgc->nr_busy_cpus, sg->group_weight); 5906} 5907 5908/* 5909 * Initializers for schedule domains 5910 * Non-inlined to reduce accumulated stack pressure in build_sched_domains() 5911 */ 5912 5913static int default_relax_domain_level = -1; 5914int sched_domain_level_max; 5915 5916static int __init setup_relax_domain_level(char *str) 5917{ 5918 if (kstrtoint(str, 0, &default_relax_domain_level)) 5919 pr_warn("Unable to set relax_domain_level\n"); 5920 5921 return 1; 5922} 5923__setup("relax_domain_level=", setup_relax_domain_level); 5924 5925static void set_domain_attribute(struct sched_domain *sd, 5926 struct sched_domain_attr *attr) 5927{ 5928 int request; 5929 5930 if (!attr || attr->relax_domain_level < 0) { 5931 if (default_relax_domain_level < 0) 5932 return; 5933 else 5934 request = default_relax_domain_level; 5935 } else 5936 request = attr->relax_domain_level; 5937 if (request < sd->level) { 5938 /* turn off idle balance on this domain */ 5939 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE); 5940 } else { 5941 /* turn on idle balance on this domain */ 5942 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE); 5943 } 5944} 5945 5946static void __sdt_free(const struct cpumask *cpu_map); 5947static int __sdt_alloc(const struct cpumask *cpu_map); 5948 5949static void __free_domain_allocs(struct s_data *d, enum s_alloc what, 5950 const struct cpumask *cpu_map) 5951{ 5952 switch (what) { 5953 case sa_rootdomain: 5954 if (!atomic_read(&d->rd->refcount)) 5955 free_rootdomain(&d->rd->rcu); /* fall through */ 5956 case sa_sd: 5957 free_percpu(d->sd); /* fall through */ 5958 case sa_sd_storage: 5959 __sdt_free(cpu_map); /* fall through */ 5960 case sa_none: 5961 break; 5962 } 5963} 5964 5965static enum s_alloc __visit_domain_allocation_hell(struct s_data *d, 5966 const struct cpumask *cpu_map) 5967{ 5968 memset(d, 0, sizeof(*d)); 5969 5970 if (__sdt_alloc(cpu_map)) 5971 return sa_sd_storage; 5972 d->sd = alloc_percpu(struct sched_domain *); 5973 if (!d->sd) 5974 return sa_sd_storage; 5975 d->rd = alloc_rootdomain(); 5976 if (!d->rd) 5977 return sa_sd; 5978 return sa_rootdomain; 5979} 5980 5981/* 5982 * NULL the sd_data elements we've used to build the sched_domain and 5983 * sched_group structure so that the subsequent __free_domain_allocs() 5984 * will not free the data we're using. 5985 */ 5986static void claim_allocations(int cpu, struct sched_domain *sd) 5987{ 5988 struct sd_data *sdd = sd->private; 5989 5990 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd); 5991 *per_cpu_ptr(sdd->sd, cpu) = NULL; 5992 5993 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref)) 5994 *per_cpu_ptr(sdd->sg, cpu) = NULL; 5995 5996 if (atomic_read(&(*per_cpu_ptr(sdd->sgc, cpu))->ref)) 5997 *per_cpu_ptr(sdd->sgc, cpu) = NULL; 5998} 5999 6000#ifdef CONFIG_NUMA 6001static int sched_domains_numa_levels; 6002static int *sched_domains_numa_distance; 6003static struct cpumask ***sched_domains_numa_masks; 6004static int sched_domains_curr_level; 6005#endif 6006 6007/* 6008 * SD_flags allowed in topology descriptions. 6009 * 6010 * SD_SHARE_CPUCAPACITY - describes SMT topologies 6011 * SD_SHARE_PKG_RESOURCES - describes shared caches 6012 * SD_NUMA - describes NUMA topologies 6013 * SD_SHARE_POWERDOMAIN - describes shared power domain 6014 * 6015 * Odd one out: 6016 * SD_ASYM_PACKING - describes SMT quirks 6017 */ 6018#define TOPOLOGY_SD_FLAGS \ 6019 (SD_SHARE_CPUCAPACITY | \ 6020 SD_SHARE_PKG_RESOURCES | \ 6021 SD_NUMA | \ 6022 SD_ASYM_PACKING | \ 6023 SD_SHARE_POWERDOMAIN) 6024 6025static struct sched_domain * 6026sd_init(struct sched_domain_topology_level *tl, int cpu) 6027{ 6028 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); 6029 int sd_weight, sd_flags = 0; 6030 6031#ifdef CONFIG_NUMA 6032 /* 6033 * Ugly hack to pass state to sd_numa_mask()... 6034 */ 6035 sched_domains_curr_level = tl->numa_level; 6036#endif 6037 6038 sd_weight = cpumask_weight(tl->mask(cpu)); 6039 6040 if (tl->sd_flags) 6041 sd_flags = (*tl->sd_flags)(); 6042 if (WARN_ONCE(sd_flags & ~TOPOLOGY_SD_FLAGS, 6043 "wrong sd_flags in topology description\n")) 6044 sd_flags &= ~TOPOLOGY_SD_FLAGS; 6045 6046 *sd = (struct sched_domain){ 6047 .min_interval = sd_weight, 6048 .max_interval = 2*sd_weight, 6049 .busy_factor = 32, 6050 .imbalance_pct = 125, 6051 6052 .cache_nice_tries = 0, 6053 .busy_idx = 0, 6054 .idle_idx = 0, 6055 .newidle_idx = 0, 6056 .wake_idx = 0, 6057 .forkexec_idx = 0, 6058 6059 .flags = 1*SD_LOAD_BALANCE 6060 | 1*SD_BALANCE_NEWIDLE 6061 | 1*SD_BALANCE_EXEC 6062 | 1*SD_BALANCE_FORK 6063 | 0*SD_BALANCE_WAKE 6064 | 1*SD_WAKE_AFFINE 6065 | 0*SD_SHARE_CPUCAPACITY 6066 | 0*SD_SHARE_PKG_RESOURCES 6067 | 0*SD_SERIALIZE 6068 | 0*SD_PREFER_SIBLING 6069 | 0*SD_NUMA 6070 | sd_flags 6071 , 6072 6073 .last_balance = jiffies, 6074 .balance_interval = sd_weight, 6075 .smt_gain = 0, 6076 .max_newidle_lb_cost = 0, 6077 .next_decay_max_lb_cost = jiffies, 6078#ifdef CONFIG_SCHED_DEBUG 6079 .name = tl->name, 6080#endif 6081 }; 6082 6083 /* 6084 * Convert topological properties into behaviour. 6085 */ 6086 6087 if (sd->flags & SD_SHARE_CPUCAPACITY) { 6088 sd->imbalance_pct = 110; 6089 sd->smt_gain = 1178; /* ~15% */ 6090 6091 } else if (sd->flags & SD_SHARE_PKG_RESOURCES) { 6092 sd->imbalance_pct = 117; 6093 sd->cache_nice_tries = 1; 6094 sd->busy_idx = 2; 6095 6096#ifdef CONFIG_NUMA 6097 } else if (sd->flags & SD_NUMA) { 6098 sd->cache_nice_tries = 2; 6099 sd->busy_idx = 3; 6100 sd->idle_idx = 2; 6101 6102 sd->flags |= SD_SERIALIZE; 6103 if (sched_domains_numa_distance[tl->numa_level] > RECLAIM_DISTANCE) { 6104 sd->flags &= ~(SD_BALANCE_EXEC | 6105 SD_BALANCE_FORK | 6106 SD_WAKE_AFFINE); 6107 } 6108 6109#endif 6110 } else { 6111 sd->flags |= SD_PREFER_SIBLING; 6112 sd->cache_nice_tries = 1; 6113 sd->busy_idx = 2; 6114 sd->idle_idx = 1; 6115 } 6116 6117 sd->private = &tl->data; 6118 6119 return sd; 6120} 6121 6122/* 6123 * Topology list, bottom-up. 6124 */ 6125static struct sched_domain_topology_level default_topology[] = { 6126#ifdef CONFIG_SCHED_SMT 6127 { cpu_smt_mask, cpu_smt_flags, SD_INIT_NAME(SMT) }, 6128#endif 6129#ifdef CONFIG_SCHED_MC 6130 { cpu_coregroup_mask, cpu_core_flags, SD_INIT_NAME(MC) }, 6131#endif 6132 { cpu_cpu_mask, SD_INIT_NAME(DIE) }, 6133 { NULL, }, 6134}; 6135 6136struct sched_domain_topology_level *sched_domain_topology = default_topology; 6137 6138#define for_each_sd_topology(tl) \ 6139 for (tl = sched_domain_topology; tl->mask; tl++) 6140 6141void set_sched_topology(struct sched_domain_topology_level *tl) 6142{ 6143 sched_domain_topology = tl; 6144} 6145 6146#ifdef CONFIG_NUMA 6147 6148static const struct cpumask *sd_numa_mask(int cpu) 6149{ 6150 return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)]; 6151} 6152 6153static void sched_numa_warn(const char *str) 6154{ 6155 static int done = false; 6156 int i,j; 6157 6158 if (done) 6159 return; 6160 6161 done = true; 6162 6163 printk(KERN_WARNING "ERROR: %s\n\n", str); 6164 6165 for (i = 0; i < nr_node_ids; i++) { 6166 printk(KERN_WARNING " "); 6167 for (j = 0; j < nr_node_ids; j++) 6168 printk(KERN_CONT "%02d ", node_distance(i,j)); 6169 printk(KERN_CONT "\n"); 6170 } 6171 printk(KERN_WARNING "\n"); 6172} 6173 6174static bool find_numa_distance(int distance) 6175{ 6176 int i; 6177 6178 if (distance == node_distance(0, 0)) 6179 return true; 6180 6181 for (i = 0; i < sched_domains_numa_levels; i++) { 6182 if (sched_domains_numa_distance[i] == distance) 6183 return true; 6184 } 6185 6186 return false; 6187} 6188 6189static void sched_init_numa(void) 6190{ 6191 int next_distance, curr_distance = node_distance(0, 0); 6192 struct sched_domain_topology_level *tl; 6193 int level = 0; 6194 int i, j, k; 6195 6196 sched_domains_numa_distance = kzalloc(sizeof(int) * nr_node_ids, GFP_KERNEL); 6197 if (!sched_domains_numa_distance) 6198 return; 6199 6200 /* 6201 * O(nr_nodes^2) deduplicating selection sort -- in order to find the 6202 * unique distances in the node_distance() table. 6203 * 6204 * Assumes node_distance(0,j) includes all distances in 6205 * node_distance(i,j) in order to avoid cubic time. 6206 */ 6207 next_distance = curr_distance; 6208 for (i = 0; i < nr_node_ids; i++) { 6209 for (j = 0; j < nr_node_ids; j++) { 6210 for (k = 0; k < nr_node_ids; k++) { 6211 int distance = node_distance(i, k); 6212 6213 if (distance > curr_distance && 6214 (distance < next_distance || 6215 next_distance == curr_distance)) 6216 next_distance = distance; 6217 6218 /* 6219 * While not a strong assumption it would be nice to know 6220 * about cases where if node A is connected to B, B is not 6221 * equally connected to A. 6222 */ 6223 if (sched_debug() && node_distance(k, i) != distance) 6224 sched_numa_warn("Node-distance not symmetric"); 6225 6226 if (sched_debug() && i && !find_numa_distance(distance)) 6227 sched_numa_warn("Node-0 not representative"); 6228 } 6229 if (next_distance != curr_distance) { 6230 sched_domains_numa_distance[level++] = next_distance; 6231 sched_domains_numa_levels = level; 6232 curr_distance = next_distance; 6233 } else break; 6234 } 6235 6236 /* 6237 * In case of sched_debug() we verify the above assumption. 6238 */ 6239 if (!sched_debug()) 6240 break; 6241 } 6242 /* 6243 * 'level' contains the number of unique distances, excluding the 6244 * identity distance node_distance(i,i). 6245 * 6246 * The sched_domains_numa_distance[] array includes the actual distance 6247 * numbers. 6248 */ 6249 6250 /* 6251 * Here, we should temporarily reset sched_domains_numa_levels to 0. 6252 * If it fails to allocate memory for array sched_domains_numa_masks[][], 6253 * the array will contain less then 'level' members. This could be 6254 * dangerous when we use it to iterate array sched_domains_numa_masks[][] 6255 * in other functions. 6256 * 6257 * We reset it to 'level' at the end of this function. 6258 */ 6259 sched_domains_numa_levels = 0; 6260 6261 sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL); 6262 if (!sched_domains_numa_masks) 6263 return; 6264 6265 /* 6266 * Now for each level, construct a mask per node which contains all 6267 * cpus of nodes that are that many hops away from us. 6268 */ 6269 for (i = 0; i < level; i++) { 6270 sched_domains_numa_masks[i] = 6271 kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL); 6272 if (!sched_domains_numa_masks[i]) 6273 return; 6274 6275 for (j = 0; j < nr_node_ids; j++) { 6276 struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL); 6277 if (!mask) 6278 return; 6279 6280 sched_domains_numa_masks[i][j] = mask; 6281 6282 for (k = 0; k < nr_node_ids; k++) { 6283 if (node_distance(j, k) > sched_domains_numa_distance[i]) 6284 continue; 6285 6286 cpumask_or(mask, mask, cpumask_of_node(k)); 6287 } 6288 } 6289 } 6290 6291 /* Compute default topology size */ 6292 for (i = 0; sched_domain_topology[i].mask; i++); 6293 6294 tl = kzalloc((i + level + 1) * 6295 sizeof(struct sched_domain_topology_level), GFP_KERNEL); 6296 if (!tl) 6297 return; 6298 6299 /* 6300 * Copy the default topology bits.. 6301 */ 6302 for (i = 0; sched_domain_topology[i].mask; i++) 6303 tl[i] = sched_domain_topology[i]; 6304 6305 /* 6306 * .. and append 'j' levels of NUMA goodness. 6307 */ 6308 for (j = 0; j < level; i++, j++) { 6309 tl[i] = (struct sched_domain_topology_level){ 6310 .mask = sd_numa_mask, 6311 .sd_flags = cpu_numa_flags, 6312 .flags = SDTL_OVERLAP, 6313 .numa_level = j, 6314 SD_INIT_NAME(NUMA) 6315 }; 6316 } 6317 6318 sched_domain_topology = tl; 6319 6320 sched_domains_numa_levels = level; 6321} 6322 6323static void sched_domains_numa_masks_set(int cpu) 6324{ 6325 int i, j; 6326 int node = cpu_to_node(cpu); 6327 6328 for (i = 0; i < sched_domains_numa_levels; i++) { 6329 for (j = 0; j < nr_node_ids; j++) { 6330 if (node_distance(j, node) <= sched_domains_numa_distance[i]) 6331 cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]); 6332 } 6333 } 6334} 6335 6336static void sched_domains_numa_masks_clear(int cpu) 6337{ 6338 int i, j; 6339 for (i = 0; i < sched_domains_numa_levels; i++) { 6340 for (j = 0; j < nr_node_ids; j++) 6341 cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]); 6342 } 6343} 6344 6345/* 6346 * Update sched_domains_numa_masks[level][node] array when new cpus 6347 * are onlined. 6348 */ 6349static int sched_domains_numa_masks_update(struct notifier_block *nfb, 6350 unsigned long action, 6351 void *hcpu) 6352{ 6353 int cpu = (long)hcpu; 6354 6355 switch (action & ~CPU_TASKS_FROZEN) { 6356 case CPU_ONLINE: 6357 sched_domains_numa_masks_set(cpu); 6358 break; 6359 6360 case CPU_DEAD: 6361 sched_domains_numa_masks_clear(cpu); 6362 break; 6363 6364 default: 6365 return NOTIFY_DONE; 6366 } 6367 6368 return NOTIFY_OK; 6369} 6370#else 6371static inline void sched_init_numa(void) 6372{ 6373} 6374 6375static int sched_domains_numa_masks_update(struct notifier_block *nfb, 6376 unsigned long action, 6377 void *hcpu) 6378{ 6379 return 0; 6380} 6381#endif /* CONFIG_NUMA */ 6382 6383static int __sdt_alloc(const struct cpumask *cpu_map) 6384{ 6385 struct sched_domain_topology_level *tl; 6386 int j; 6387 6388 for_each_sd_topology(tl) { 6389 struct sd_data *sdd = &tl->data; 6390 6391 sdd->sd = alloc_percpu(struct sched_domain *); 6392 if (!sdd->sd) 6393 return -ENOMEM; 6394 6395 sdd->sg = alloc_percpu(struct sched_group *); 6396 if (!sdd->sg) 6397 return -ENOMEM; 6398 6399 sdd->sgc = alloc_percpu(struct sched_group_capacity *); 6400 if (!sdd->sgc) 6401 return -ENOMEM; 6402 6403 for_each_cpu(j, cpu_map) { 6404 struct sched_domain *sd; 6405 struct sched_group *sg; 6406 struct sched_group_capacity *sgc; 6407 6408 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(), 6409 GFP_KERNEL, cpu_to_node(j)); 6410 if (!sd) 6411 return -ENOMEM; 6412 6413 *per_cpu_ptr(sdd->sd, j) = sd; 6414 6415 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(), 6416 GFP_KERNEL, cpu_to_node(j)); 6417 if (!sg) 6418 return -ENOMEM; 6419 6420 sg->next = sg; 6421 6422 *per_cpu_ptr(sdd->sg, j) = sg; 6423 6424 sgc = kzalloc_node(sizeof(struct sched_group_capacity) + cpumask_size(), 6425 GFP_KERNEL, cpu_to_node(j)); 6426 if (!sgc) 6427 return -ENOMEM; 6428 6429 *per_cpu_ptr(sdd->sgc, j) = sgc; 6430 } 6431 } 6432 6433 return 0; 6434} 6435 6436static void __sdt_free(const struct cpumask *cpu_map) 6437{ 6438 struct sched_domain_topology_level *tl; 6439 int j; 6440 6441 for_each_sd_topology(tl) { 6442 struct sd_data *sdd = &tl->data; 6443 6444 for_each_cpu(j, cpu_map) { 6445 struct sched_domain *sd; 6446 6447 if (sdd->sd) { 6448 sd = *per_cpu_ptr(sdd->sd, j); 6449 if (sd && (sd->flags & SD_OVERLAP)) 6450 free_sched_groups(sd->groups, 0); 6451 kfree(*per_cpu_ptr(sdd->sd, j)); 6452 } 6453 6454 if (sdd->sg) 6455 kfree(*per_cpu_ptr(sdd->sg, j)); 6456 if (sdd->sgc) 6457 kfree(*per_cpu_ptr(sdd->sgc, j)); 6458 } 6459 free_percpu(sdd->sd); 6460 sdd->sd = NULL; 6461 free_percpu(sdd->sg); 6462 sdd->sg = NULL; 6463 free_percpu(sdd->sgc); 6464 sdd->sgc = NULL; 6465 } 6466} 6467 6468struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl, 6469 const struct cpumask *cpu_map, struct sched_domain_attr *attr, 6470 struct sched_domain *child, int cpu) 6471{ 6472 struct sched_domain *sd = sd_init(tl, cpu); 6473 if (!sd) 6474 return child; 6475 6476 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu)); 6477 if (child) { 6478 sd->level = child->level + 1; 6479 sched_domain_level_max = max(sched_domain_level_max, sd->level); 6480 child->parent = sd; 6481 sd->child = child; 6482 } 6483 set_domain_attribute(sd, attr); 6484 6485 return sd; 6486} 6487 6488/* 6489 * Build sched domains for a given set of cpus and attach the sched domains 6490 * to the individual cpus 6491 */ 6492static int build_sched_domains(const struct cpumask *cpu_map, 6493 struct sched_domain_attr *attr) 6494{ 6495 enum s_alloc alloc_state; 6496 struct sched_domain *sd; 6497 struct s_data d; 6498 int i, ret = -ENOMEM; 6499 6500 alloc_state = __visit_domain_allocation_hell(&d, cpu_map); 6501 if (alloc_state != sa_rootdomain) 6502 goto error; 6503 6504 /* Set up domains for cpus specified by the cpu_map. */ 6505 for_each_cpu(i, cpu_map) { 6506 struct sched_domain_topology_level *tl; 6507 6508 sd = NULL; 6509 for_each_sd_topology(tl) { 6510 sd = build_sched_domain(tl, cpu_map, attr, sd, i); 6511 if (tl == sched_domain_topology) 6512 *per_cpu_ptr(d.sd, i) = sd; 6513 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP)) 6514 sd->flags |= SD_OVERLAP; 6515 if (cpumask_equal(cpu_map, sched_domain_span(sd))) 6516 break; 6517 } 6518 } 6519 6520 /* Build the groups for the domains */ 6521 for_each_cpu(i, cpu_map) { 6522 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) { 6523 sd->span_weight = cpumask_weight(sched_domain_span(sd)); 6524 if (sd->flags & SD_OVERLAP) { 6525 if (build_overlap_sched_groups(sd, i)) 6526 goto error; 6527 } else { 6528 if (build_sched_groups(sd, i)) 6529 goto error; 6530 } 6531 } 6532 } 6533 6534 /* Calculate CPU capacity for physical packages and nodes */ 6535 for (i = nr_cpumask_bits-1; i >= 0; i--) { 6536 if (!cpumask_test_cpu(i, cpu_map)) 6537 continue; 6538 6539 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) { 6540 claim_allocations(i, sd); 6541 init_sched_groups_capacity(i, sd); 6542 } 6543 } 6544 6545 /* Attach the domains */ 6546 rcu_read_lock(); 6547 for_each_cpu(i, cpu_map) { 6548 sd = *per_cpu_ptr(d.sd, i); 6549 cpu_attach_domain(sd, d.rd, i); 6550 } 6551 rcu_read_unlock(); 6552 6553 ret = 0; 6554error: 6555 __free_domain_allocs(&d, alloc_state, cpu_map); 6556 return ret; 6557} 6558 6559static cpumask_var_t *doms_cur; /* current sched domains */ 6560static int ndoms_cur; /* number of sched domains in 'doms_cur' */ 6561static struct sched_domain_attr *dattr_cur; 6562 /* attribues of custom domains in 'doms_cur' */ 6563 6564/* 6565 * Special case: If a kmalloc of a doms_cur partition (array of 6566 * cpumask) fails, then fallback to a single sched domain, 6567 * as determined by the single cpumask fallback_doms. 6568 */ 6569static cpumask_var_t fallback_doms; 6570 6571/* 6572 * arch_update_cpu_topology lets virtualized architectures update the 6573 * cpu core maps. It is supposed to return 1 if the topology changed 6574 * or 0 if it stayed the same. 6575 */ 6576int __weak arch_update_cpu_topology(void) 6577{ 6578 return 0; 6579} 6580 6581cpumask_var_t *alloc_sched_domains(unsigned int ndoms) 6582{ 6583 int i; 6584 cpumask_var_t *doms; 6585 6586 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL); 6587 if (!doms) 6588 return NULL; 6589 for (i = 0; i < ndoms; i++) { 6590 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) { 6591 free_sched_domains(doms, i); 6592 return NULL; 6593 } 6594 } 6595 return doms; 6596} 6597 6598void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms) 6599{ 6600 unsigned int i; 6601 for (i = 0; i < ndoms; i++) 6602 free_cpumask_var(doms[i]); 6603 kfree(doms); 6604} 6605 6606/* 6607 * Set up scheduler domains and groups. Callers must hold the hotplug lock. 6608 * For now this just excludes isolated cpus, but could be used to 6609 * exclude other special cases in the future. 6610 */ 6611static int init_sched_domains(const struct cpumask *cpu_map) 6612{ 6613 int err; 6614 6615 arch_update_cpu_topology(); 6616 ndoms_cur = 1; 6617 doms_cur = alloc_sched_domains(ndoms_cur); 6618 if (!doms_cur) 6619 doms_cur = &fallback_doms; 6620 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map); 6621 err = build_sched_domains(doms_cur[0], NULL); 6622 register_sched_domain_sysctl(); 6623 6624 return err; 6625} 6626 6627/* 6628 * Detach sched domains from a group of cpus specified in cpu_map 6629 * These cpus will now be attached to the NULL domain 6630 */ 6631static void detach_destroy_domains(const struct cpumask *cpu_map) 6632{ 6633 int i; 6634 6635 rcu_read_lock(); 6636 for_each_cpu(i, cpu_map) 6637 cpu_attach_domain(NULL, &def_root_domain, i); 6638 rcu_read_unlock(); 6639} 6640 6641/* handle null as "default" */ 6642static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur, 6643 struct sched_domain_attr *new, int idx_new) 6644{ 6645 struct sched_domain_attr tmp; 6646 6647 /* fast path */ 6648 if (!new && !cur) 6649 return 1; 6650 6651 tmp = SD_ATTR_INIT; 6652 return !memcmp(cur ? (cur + idx_cur) : &tmp, 6653 new ? (new + idx_new) : &tmp, 6654 sizeof(struct sched_domain_attr)); 6655} 6656 6657/* 6658 * Partition sched domains as specified by the 'ndoms_new' 6659 * cpumasks in the array doms_new[] of cpumasks. This compares 6660 * doms_new[] to the current sched domain partitioning, doms_cur[]. 6661 * It destroys each deleted domain and builds each new domain. 6662 * 6663 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'. 6664 * The masks don't intersect (don't overlap.) We should setup one 6665 * sched domain for each mask. CPUs not in any of the cpumasks will 6666 * not be load balanced. If the same cpumask appears both in the 6667 * current 'doms_cur' domains and in the new 'doms_new', we can leave 6668 * it as it is. 6669 * 6670 * The passed in 'doms_new' should be allocated using 6671 * alloc_sched_domains. This routine takes ownership of it and will 6672 * free_sched_domains it when done with it. If the caller failed the 6673 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1, 6674 * and partition_sched_domains() will fallback to the single partition 6675 * 'fallback_doms', it also forces the domains to be rebuilt. 6676 * 6677 * If doms_new == NULL it will be replaced with cpu_online_mask. 6678 * ndoms_new == 0 is a special case for destroying existing domains, 6679 * and it will not create the default domain. 6680 * 6681 * Call with hotplug lock held 6682 */ 6683void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[], 6684 struct sched_domain_attr *dattr_new) 6685{ 6686 int i, j, n; 6687 int new_topology; 6688 6689 mutex_lock(&sched_domains_mutex); 6690 6691 /* always unregister in case we don't destroy any domains */ 6692 unregister_sched_domain_sysctl(); 6693 6694 /* Let architecture update cpu core mappings. */ 6695 new_topology = arch_update_cpu_topology(); 6696 6697 n = doms_new ? ndoms_new : 0; 6698 6699 /* Destroy deleted domains */ 6700 for (i = 0; i < ndoms_cur; i++) { 6701 for (j = 0; j < n && !new_topology; j++) { 6702 if (cpumask_equal(doms_cur[i], doms_new[j]) 6703 && dattrs_equal(dattr_cur, i, dattr_new, j)) 6704 goto match1; 6705 } 6706 /* no match - a current sched domain not in new doms_new[] */ 6707 detach_destroy_domains(doms_cur[i]); 6708match1: 6709 ; 6710 } 6711 6712 n = ndoms_cur; 6713 if (doms_new == NULL) { 6714 n = 0; 6715 doms_new = &fallback_doms; 6716 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map); 6717 WARN_ON_ONCE(dattr_new); 6718 } 6719 6720 /* Build new domains */ 6721 for (i = 0; i < ndoms_new; i++) { 6722 for (j = 0; j < n && !new_topology; j++) { 6723 if (cpumask_equal(doms_new[i], doms_cur[j]) 6724 && dattrs_equal(dattr_new, i, dattr_cur, j)) 6725 goto match2; 6726 } 6727 /* no match - add a new doms_new */ 6728 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL); 6729match2: 6730 ; 6731 } 6732 6733 /* Remember the new sched domains */ 6734 if (doms_cur != &fallback_doms) 6735 free_sched_domains(doms_cur, ndoms_cur); 6736 kfree(dattr_cur); /* kfree(NULL) is safe */ 6737 doms_cur = doms_new; 6738 dattr_cur = dattr_new; 6739 ndoms_cur = ndoms_new; 6740 6741 register_sched_domain_sysctl(); 6742 6743 mutex_unlock(&sched_domains_mutex); 6744} 6745 6746static int num_cpus_frozen; /* used to mark begin/end of suspend/resume */ 6747 6748/* 6749 * Update cpusets according to cpu_active mask. If cpusets are 6750 * disabled, cpuset_update_active_cpus() becomes a simple wrapper 6751 * around partition_sched_domains(). 6752 * 6753 * If we come here as part of a suspend/resume, don't touch cpusets because we 6754 * want to restore it back to its original state upon resume anyway. 6755 */ 6756static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action, 6757 void *hcpu) 6758{ 6759 switch (action) { 6760 case CPU_ONLINE_FROZEN: 6761 case CPU_DOWN_FAILED_FROZEN: 6762 6763 /* 6764 * num_cpus_frozen tracks how many CPUs are involved in suspend 6765 * resume sequence. As long as this is not the last online 6766 * operation in the resume sequence, just build a single sched 6767 * domain, ignoring cpusets. 6768 */ 6769 num_cpus_frozen--; 6770 if (likely(num_cpus_frozen)) { 6771 partition_sched_domains(1, NULL, NULL); 6772 break; 6773 } 6774 6775 /* 6776 * This is the last CPU online operation. So fall through and 6777 * restore the original sched domains by considering the 6778 * cpuset configurations. 6779 */ 6780 6781 case CPU_ONLINE: 6782 case CPU_DOWN_FAILED: 6783 cpuset_update_active_cpus(true); 6784 break; 6785 default: 6786 return NOTIFY_DONE; 6787 } 6788 return NOTIFY_OK; 6789} 6790 6791static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action, 6792 void *hcpu) 6793{ 6794 switch (action) { 6795 case CPU_DOWN_PREPARE: 6796 cpuset_update_active_cpus(false); 6797 break; 6798 case CPU_DOWN_PREPARE_FROZEN: 6799 num_cpus_frozen++; 6800 partition_sched_domains(1, NULL, NULL); 6801 break; 6802 default: 6803 return NOTIFY_DONE; 6804 } 6805 return NOTIFY_OK; 6806} 6807 6808void __init sched_init_smp(void) 6809{ 6810 cpumask_var_t non_isolated_cpus; 6811 6812 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL); 6813 alloc_cpumask_var(&fallback_doms, GFP_KERNEL); 6814 6815 sched_init_numa(); 6816 6817 /* 6818 * There's no userspace yet to cause hotplug operations; hence all the 6819 * cpu masks are stable and all blatant races in the below code cannot 6820 * happen. 6821 */ 6822 mutex_lock(&sched_domains_mutex); 6823 init_sched_domains(cpu_active_mask); 6824 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map); 6825 if (cpumask_empty(non_isolated_cpus)) 6826 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus); 6827 mutex_unlock(&sched_domains_mutex); 6828 6829 hotcpu_notifier(sched_domains_numa_masks_update, CPU_PRI_SCHED_ACTIVE); 6830 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE); 6831 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE); 6832 6833 init_hrtick(); 6834 6835 /* Move init over to a non-isolated CPU */ 6836 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0) 6837 BUG(); 6838 sched_init_granularity(); 6839 free_cpumask_var(non_isolated_cpus); 6840 6841 init_sched_rt_class(); 6842 init_sched_dl_class(); 6843} 6844#else 6845void __init sched_init_smp(void) 6846{ 6847 sched_init_granularity(); 6848} 6849#endif /* CONFIG_SMP */ 6850 6851const_debug unsigned int sysctl_timer_migration = 1; 6852 6853int in_sched_functions(unsigned long addr) 6854{ 6855 return in_lock_functions(addr) || 6856 (addr >= (unsigned long)__sched_text_start 6857 && addr < (unsigned long)__sched_text_end); 6858} 6859 6860#ifdef CONFIG_CGROUP_SCHED 6861/* 6862 * Default task group. 6863 * Every task in system belongs to this group at bootup. 6864 */ 6865struct task_group root_task_group; 6866LIST_HEAD(task_groups); 6867#endif 6868 6869DECLARE_PER_CPU(cpumask_var_t, load_balance_mask); 6870 6871void __init sched_init(void) 6872{ 6873 int i, j; 6874 unsigned long alloc_size = 0, ptr; 6875 6876#ifdef CONFIG_FAIR_GROUP_SCHED 6877 alloc_size += 2 * nr_cpu_ids * sizeof(void **); 6878#endif 6879#ifdef CONFIG_RT_GROUP_SCHED 6880 alloc_size += 2 * nr_cpu_ids * sizeof(void **); 6881#endif 6882#ifdef CONFIG_CPUMASK_OFFSTACK 6883 alloc_size += num_possible_cpus() * cpumask_size(); 6884#endif 6885 if (alloc_size) { 6886 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT); 6887 6888#ifdef CONFIG_FAIR_GROUP_SCHED 6889 root_task_group.se = (struct sched_entity **)ptr; 6890 ptr += nr_cpu_ids * sizeof(void **); 6891 6892 root_task_group.cfs_rq = (struct cfs_rq **)ptr; 6893 ptr += nr_cpu_ids * sizeof(void **); 6894 6895#endif /* CONFIG_FAIR_GROUP_SCHED */ 6896#ifdef CONFIG_RT_GROUP_SCHED 6897 root_task_group.rt_se = (struct sched_rt_entity **)ptr; 6898 ptr += nr_cpu_ids * sizeof(void **); 6899 6900 root_task_group.rt_rq = (struct rt_rq **)ptr; 6901 ptr += nr_cpu_ids * sizeof(void **); 6902 6903#endif /* CONFIG_RT_GROUP_SCHED */ 6904#ifdef CONFIG_CPUMASK_OFFSTACK 6905 for_each_possible_cpu(i) { 6906 per_cpu(load_balance_mask, i) = (void *)ptr; 6907 ptr += cpumask_size(); 6908 } 6909#endif /* CONFIG_CPUMASK_OFFSTACK */ 6910 } 6911 6912 init_rt_bandwidth(&def_rt_bandwidth, 6913 global_rt_period(), global_rt_runtime()); 6914 init_dl_bandwidth(&def_dl_bandwidth, 6915 global_rt_period(), global_rt_runtime()); 6916 6917#ifdef CONFIG_SMP 6918 init_defrootdomain(); 6919#endif 6920 6921#ifdef CONFIG_RT_GROUP_SCHED 6922 init_rt_bandwidth(&root_task_group.rt_bandwidth, 6923 global_rt_period(), global_rt_runtime()); 6924#endif /* CONFIG_RT_GROUP_SCHED */ 6925 6926#ifdef CONFIG_CGROUP_SCHED 6927 list_add(&root_task_group.list, &task_groups); 6928 INIT_LIST_HEAD(&root_task_group.children); 6929 INIT_LIST_HEAD(&root_task_group.siblings); 6930 autogroup_init(&init_task); 6931 6932#endif /* CONFIG_CGROUP_SCHED */ 6933 6934 for_each_possible_cpu(i) { 6935 struct rq *rq; 6936 6937 rq = cpu_rq(i); 6938 raw_spin_lock_init(&rq->lock); 6939 rq->nr_running = 0; 6940 rq->calc_load_active = 0; 6941 rq->calc_load_update = jiffies + LOAD_FREQ; 6942 init_cfs_rq(&rq->cfs); 6943 init_rt_rq(&rq->rt, rq); 6944 init_dl_rq(&rq->dl, rq); 6945#ifdef CONFIG_FAIR_GROUP_SCHED 6946 root_task_group.shares = ROOT_TASK_GROUP_LOAD; 6947 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list); 6948 /* 6949 * How much cpu bandwidth does root_task_group get? 6950 * 6951 * In case of task-groups formed thr' the cgroup filesystem, it 6952 * gets 100% of the cpu resources in the system. This overall 6953 * system cpu resource is divided among the tasks of 6954 * root_task_group and its child task-groups in a fair manner, 6955 * based on each entity's (task or task-group's) weight 6956 * (se->load.weight). 6957 * 6958 * In other words, if root_task_group has 10 tasks of weight 6959 * 1024) and two child groups A0 and A1 (of weight 1024 each), 6960 * then A0's share of the cpu resource is: 6961 * 6962 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33% 6963 * 6964 * We achieve this by letting root_task_group's tasks sit 6965 * directly in rq->cfs (i.e root_task_group->se[] = NULL). 6966 */ 6967 init_cfs_bandwidth(&root_task_group.cfs_bandwidth); 6968 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL); 6969#endif /* CONFIG_FAIR_GROUP_SCHED */ 6970 6971 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime; 6972#ifdef CONFIG_RT_GROUP_SCHED 6973 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL); 6974#endif 6975 6976 for (j = 0; j < CPU_LOAD_IDX_MAX; j++) 6977 rq->cpu_load[j] = 0; 6978 6979 rq->last_load_update_tick = jiffies; 6980 6981#ifdef CONFIG_SMP 6982 rq->sd = NULL; 6983 rq->rd = NULL; 6984 rq->cpu_capacity = SCHED_CAPACITY_SCALE; 6985 rq->post_schedule = 0; 6986 rq->active_balance = 0; 6987 rq->next_balance = jiffies; 6988 rq->push_cpu = 0; 6989 rq->cpu = i; 6990 rq->online = 0; 6991 rq->idle_stamp = 0; 6992 rq->avg_idle = 2*sysctl_sched_migration_cost; 6993 rq->max_idle_balance_cost = sysctl_sched_migration_cost; 6994 6995 INIT_LIST_HEAD(&rq->cfs_tasks); 6996 6997 rq_attach_root(rq, &def_root_domain); 6998#ifdef CONFIG_NO_HZ_COMMON 6999 rq->nohz_flags = 0; 7000#endif 7001#ifdef CONFIG_NO_HZ_FULL 7002 rq->last_sched_tick = 0; 7003#endif 7004#endif 7005 init_rq_hrtick(rq); 7006 atomic_set(&rq->nr_iowait, 0); 7007 } 7008 7009 set_load_weight(&init_task); 7010 7011#ifdef CONFIG_PREEMPT_NOTIFIERS 7012 INIT_HLIST_HEAD(&init_task.preempt_notifiers); 7013#endif 7014 7015 /* 7016 * The boot idle thread does lazy MMU switching as well: 7017 */ 7018 atomic_inc(&init_mm.mm_count); 7019 enter_lazy_tlb(&init_mm, current); 7020 7021 /* 7022 * Make us the idle thread. Technically, schedule() should not be 7023 * called from this thread, however somewhere below it might be, 7024 * but because we are the idle thread, we just pick up running again 7025 * when this runqueue becomes "idle". 7026 */ 7027 init_idle(current, smp_processor_id()); 7028 7029 calc_load_update = jiffies + LOAD_FREQ; 7030 7031 /* 7032 * During early bootup we pretend to be a normal task: 7033 */ 7034 current->sched_class = &fair_sched_class; 7035 7036#ifdef CONFIG_SMP 7037 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT); 7038 /* May be allocated at isolcpus cmdline parse time */ 7039 if (cpu_isolated_map == NULL) 7040 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT); 7041 idle_thread_set_boot_cpu(); 7042 set_cpu_rq_start_time(); 7043#endif 7044 init_sched_fair_class(); 7045 7046 scheduler_running = 1; 7047} 7048 7049#ifdef CONFIG_DEBUG_ATOMIC_SLEEP 7050static inline int preempt_count_equals(int preempt_offset) 7051{ 7052 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth(); 7053 7054 return (nested == preempt_offset); 7055} 7056 7057void __might_sleep(const char *file, int line, int preempt_offset) 7058{ 7059 static unsigned long prev_jiffy; /* ratelimiting */ 7060 7061 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */ 7062 if ((preempt_count_equals(preempt_offset) && !irqs_disabled() && 7063 !is_idle_task(current)) || 7064 system_state != SYSTEM_RUNNING || oops_in_progress) 7065 return; 7066 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy) 7067 return; 7068 prev_jiffy = jiffies; 7069 7070 printk(KERN_ERR 7071 "BUG: sleeping function called from invalid context at %s:%d\n", 7072 file, line); 7073 printk(KERN_ERR 7074 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n", 7075 in_atomic(), irqs_disabled(), 7076 current->pid, current->comm); 7077 7078 debug_show_held_locks(current); 7079 if (irqs_disabled()) 7080 print_irqtrace_events(current); 7081#ifdef CONFIG_DEBUG_PREEMPT 7082 if (!preempt_count_equals(preempt_offset)) { 7083 pr_err("Preemption disabled at:"); 7084 print_ip_sym(current->preempt_disable_ip); 7085 pr_cont("\n"); 7086 } 7087#endif 7088 dump_stack(); 7089} 7090EXPORT_SYMBOL(__might_sleep); 7091#endif 7092 7093#ifdef CONFIG_MAGIC_SYSRQ 7094static void normalize_task(struct rq *rq, struct task_struct *p) 7095{ 7096 const struct sched_class *prev_class = p->sched_class; 7097 struct sched_attr attr = { 7098 .sched_policy = SCHED_NORMAL, 7099 }; 7100 int old_prio = p->prio; 7101 int on_rq; 7102 7103 on_rq = p->on_rq; 7104 if (on_rq) 7105 dequeue_task(rq, p, 0); 7106 __setscheduler(rq, p, &attr); 7107 if (on_rq) { 7108 enqueue_task(rq, p, 0); 7109 resched_task(rq->curr); 7110 } 7111 7112 check_class_changed(rq, p, prev_class, old_prio); 7113} 7114 7115void normalize_rt_tasks(void) 7116{ 7117 struct task_struct *g, *p; 7118 unsigned long flags; 7119 struct rq *rq; 7120 7121 read_lock_irqsave(&tasklist_lock, flags); 7122 do_each_thread(g, p) { 7123 /* 7124 * Only normalize user tasks: 7125 */ 7126 if (!p->mm) 7127 continue; 7128 7129 p->se.exec_start = 0; 7130#ifdef CONFIG_SCHEDSTATS 7131 p->se.statistics.wait_start = 0; 7132 p->se.statistics.sleep_start = 0; 7133 p->se.statistics.block_start = 0; 7134#endif 7135 7136 if (!dl_task(p) && !rt_task(p)) { 7137 /* 7138 * Renice negative nice level userspace 7139 * tasks back to 0: 7140 */ 7141 if (task_nice(p) < 0 && p->mm) 7142 set_user_nice(p, 0); 7143 continue; 7144 } 7145 7146 raw_spin_lock(&p->pi_lock); 7147 rq = __task_rq_lock(p); 7148 7149 normalize_task(rq, p); 7150 7151 __task_rq_unlock(rq); 7152 raw_spin_unlock(&p->pi_lock); 7153 } while_each_thread(g, p); 7154 7155 read_unlock_irqrestore(&tasklist_lock, flags); 7156} 7157 7158#endif /* CONFIG_MAGIC_SYSRQ */ 7159 7160#if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) 7161/* 7162 * These functions are only useful for the IA64 MCA handling, or kdb. 7163 * 7164 * They can only be called when the whole system has been 7165 * stopped - every CPU needs to be quiescent, and no scheduling 7166 * activity can take place. Using them for anything else would 7167 * be a serious bug, and as a result, they aren't even visible 7168 * under any other configuration. 7169 */ 7170 7171/** 7172 * curr_task - return the current task for a given cpu. 7173 * @cpu: the processor in question. 7174 * 7175 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED! 7176 * 7177 * Return: The current task for @cpu. 7178 */ 7179struct task_struct *curr_task(int cpu) 7180{ 7181 return cpu_curr(cpu); 7182} 7183 7184#endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */ 7185 7186#ifdef CONFIG_IA64 7187/** 7188 * set_curr_task - set the current task for a given cpu. 7189 * @cpu: the processor in question. 7190 * @p: the task pointer to set. 7191 * 7192 * Description: This function must only be used when non-maskable interrupts 7193 * are serviced on a separate stack. It allows the architecture to switch the 7194 * notion of the current task on a cpu in a non-blocking manner. This function 7195 * must be called with all CPU's synchronized, and interrupts disabled, the 7196 * and caller must save the original value of the current task (see 7197 * curr_task() above) and restore that value before reenabling interrupts and 7198 * re-starting the system. 7199 * 7200 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED! 7201 */ 7202void set_curr_task(int cpu, struct task_struct *p) 7203{ 7204 cpu_curr(cpu) = p; 7205} 7206 7207#endif 7208 7209#ifdef CONFIG_CGROUP_SCHED 7210/* task_group_lock serializes the addition/removal of task groups */ 7211static DEFINE_SPINLOCK(task_group_lock); 7212 7213static void free_sched_group(struct task_group *tg) 7214{ 7215 free_fair_sched_group(tg); 7216 free_rt_sched_group(tg); 7217 autogroup_free(tg); 7218 kfree(tg); 7219} 7220 7221/* allocate runqueue etc for a new task group */ 7222struct task_group *sched_create_group(struct task_group *parent) 7223{ 7224 struct task_group *tg; 7225 7226 tg = kzalloc(sizeof(*tg), GFP_KERNEL); 7227 if (!tg) 7228 return ERR_PTR(-ENOMEM); 7229 7230 if (!alloc_fair_sched_group(tg, parent)) 7231 goto err; 7232 7233 if (!alloc_rt_sched_group(tg, parent)) 7234 goto err; 7235 7236 return tg; 7237 7238err: 7239 free_sched_group(tg); 7240 return ERR_PTR(-ENOMEM); 7241} 7242 7243void sched_online_group(struct task_group *tg, struct task_group *parent) 7244{ 7245 unsigned long flags; 7246 7247 spin_lock_irqsave(&task_group_lock, flags); 7248 list_add_rcu(&tg->list, &task_groups); 7249 7250 WARN_ON(!parent); /* root should already exist */ 7251 7252 tg->parent = parent; 7253 INIT_LIST_HEAD(&tg->children); 7254 list_add_rcu(&tg->siblings, &parent->children); 7255 spin_unlock_irqrestore(&task_group_lock, flags); 7256} 7257 7258/* rcu callback to free various structures associated with a task group */ 7259static void free_sched_group_rcu(struct rcu_head *rhp) 7260{ 7261 /* now it should be safe to free those cfs_rqs */ 7262 free_sched_group(container_of(rhp, struct task_group, rcu)); 7263} 7264 7265/* Destroy runqueue etc associated with a task group */ 7266void sched_destroy_group(struct task_group *tg) 7267{ 7268 /* wait for possible concurrent references to cfs_rqs complete */ 7269 call_rcu(&tg->rcu, free_sched_group_rcu); 7270} 7271 7272void sched_offline_group(struct task_group *tg) 7273{ 7274 unsigned long flags; 7275 int i; 7276 7277 /* end participation in shares distribution */ 7278 for_each_possible_cpu(i) 7279 unregister_fair_sched_group(tg, i); 7280 7281 spin_lock_irqsave(&task_group_lock, flags); 7282 list_del_rcu(&tg->list); 7283 list_del_rcu(&tg->siblings); 7284 spin_unlock_irqrestore(&task_group_lock, flags); 7285} 7286 7287/* change task's runqueue when it moves between groups. 7288 * The caller of this function should have put the task in its new group 7289 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to 7290 * reflect its new group. 7291 */ 7292void sched_move_task(struct task_struct *tsk) 7293{ 7294 struct task_group *tg; 7295 int on_rq, running; 7296 unsigned long flags; 7297 struct rq *rq; 7298 7299 rq = task_rq_lock(tsk, &flags); 7300 7301 running = task_current(rq, tsk); 7302 on_rq = tsk->on_rq; 7303 7304 if (on_rq) 7305 dequeue_task(rq, tsk, 0); 7306 if (unlikely(running)) 7307 tsk->sched_class->put_prev_task(rq, tsk); 7308 7309 tg = container_of(task_css_check(tsk, cpu_cgrp_id, 7310 lockdep_is_held(&tsk->sighand->siglock)), 7311 struct task_group, css); 7312 tg = autogroup_task_group(tsk, tg); 7313 tsk->sched_task_group = tg; 7314 7315#ifdef CONFIG_FAIR_GROUP_SCHED 7316 if (tsk->sched_class->task_move_group) 7317 tsk->sched_class->task_move_group(tsk, on_rq); 7318 else 7319#endif 7320 set_task_rq(tsk, task_cpu(tsk)); 7321 7322 if (unlikely(running)) 7323 tsk->sched_class->set_curr_task(rq); 7324 if (on_rq) 7325 enqueue_task(rq, tsk, 0); 7326 7327 task_rq_unlock(rq, tsk, &flags); 7328} 7329#endif /* CONFIG_CGROUP_SCHED */ 7330 7331#ifdef CONFIG_RT_GROUP_SCHED 7332/* 7333 * Ensure that the real time constraints are schedulable. 7334 */ 7335static DEFINE_MUTEX(rt_constraints_mutex); 7336 7337/* Must be called with tasklist_lock held */ 7338static inline int tg_has_rt_tasks(struct task_group *tg) 7339{ 7340 struct task_struct *g, *p; 7341 7342 do_each_thread(g, p) { 7343 if (rt_task(p) && task_rq(p)->rt.tg == tg) 7344 return 1; 7345 } while_each_thread(g, p); 7346 7347 return 0; 7348} 7349 7350struct rt_schedulable_data { 7351 struct task_group *tg; 7352 u64 rt_period; 7353 u64 rt_runtime; 7354}; 7355 7356static int tg_rt_schedulable(struct task_group *tg, void *data) 7357{ 7358 struct rt_schedulable_data *d = data; 7359 struct task_group *child; 7360 unsigned long total, sum = 0; 7361 u64 period, runtime; 7362 7363 period = ktime_to_ns(tg->rt_bandwidth.rt_period); 7364 runtime = tg->rt_bandwidth.rt_runtime; 7365 7366 if (tg == d->tg) { 7367 period = d->rt_period; 7368 runtime = d->rt_runtime; 7369 } 7370 7371 /* 7372 * Cannot have more runtime than the period. 7373 */ 7374 if (runtime > period && runtime != RUNTIME_INF) 7375 return -EINVAL; 7376 7377 /* 7378 * Ensure we don't starve existing RT tasks. 7379 */ 7380 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg)) 7381 return -EBUSY; 7382 7383 total = to_ratio(period, runtime); 7384 7385 /* 7386 * Nobody can have more than the global setting allows. 7387 */ 7388 if (total > to_ratio(global_rt_period(), global_rt_runtime())) 7389 return -EINVAL; 7390 7391 /* 7392 * The sum of our children's runtime should not exceed our own. 7393 */ 7394 list_for_each_entry_rcu(child, &tg->children, siblings) { 7395 period = ktime_to_ns(child->rt_bandwidth.rt_period); 7396 runtime = child->rt_bandwidth.rt_runtime; 7397 7398 if (child == d->tg) { 7399 period = d->rt_period; 7400 runtime = d->rt_runtime; 7401 } 7402 7403 sum += to_ratio(period, runtime); 7404 } 7405 7406 if (sum > total) 7407 return -EINVAL; 7408 7409 return 0; 7410} 7411 7412static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime) 7413{ 7414 int ret; 7415 7416 struct rt_schedulable_data data = { 7417 .tg = tg, 7418 .rt_period = period, 7419 .rt_runtime = runtime, 7420 }; 7421 7422 rcu_read_lock(); 7423 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data); 7424 rcu_read_unlock(); 7425 7426 return ret; 7427} 7428 7429static int tg_set_rt_bandwidth(struct task_group *tg, 7430 u64 rt_period, u64 rt_runtime) 7431{ 7432 int i, err = 0; 7433 7434 mutex_lock(&rt_constraints_mutex); 7435 read_lock(&tasklist_lock); 7436 err = __rt_schedulable(tg, rt_period, rt_runtime); 7437 if (err) 7438 goto unlock; 7439 7440 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock); 7441 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period); 7442 tg->rt_bandwidth.rt_runtime = rt_runtime; 7443 7444 for_each_possible_cpu(i) { 7445 struct rt_rq *rt_rq = tg->rt_rq[i]; 7446 7447 raw_spin_lock(&rt_rq->rt_runtime_lock); 7448 rt_rq->rt_runtime = rt_runtime; 7449 raw_spin_unlock(&rt_rq->rt_runtime_lock); 7450 } 7451 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock); 7452unlock: 7453 read_unlock(&tasklist_lock); 7454 mutex_unlock(&rt_constraints_mutex); 7455 7456 return err; 7457} 7458 7459static int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us) 7460{ 7461 u64 rt_runtime, rt_period; 7462 7463 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period); 7464 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC; 7465 if (rt_runtime_us < 0) 7466 rt_runtime = RUNTIME_INF; 7467 7468 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime); 7469} 7470 7471static long sched_group_rt_runtime(struct task_group *tg) 7472{ 7473 u64 rt_runtime_us; 7474 7475 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF) 7476 return -1; 7477 7478 rt_runtime_us = tg->rt_bandwidth.rt_runtime; 7479 do_div(rt_runtime_us, NSEC_PER_USEC); 7480 return rt_runtime_us; 7481} 7482 7483static int sched_group_set_rt_period(struct task_group *tg, long rt_period_us) 7484{ 7485 u64 rt_runtime, rt_period; 7486 7487 rt_period = (u64)rt_period_us * NSEC_PER_USEC; 7488 rt_runtime = tg->rt_bandwidth.rt_runtime; 7489 7490 if (rt_period == 0) 7491 return -EINVAL; 7492 7493 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime); 7494} 7495 7496static long sched_group_rt_period(struct task_group *tg) 7497{ 7498 u64 rt_period_us; 7499 7500 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period); 7501 do_div(rt_period_us, NSEC_PER_USEC); 7502 return rt_period_us; 7503} 7504#endif /* CONFIG_RT_GROUP_SCHED */ 7505 7506#ifdef CONFIG_RT_GROUP_SCHED 7507static int sched_rt_global_constraints(void) 7508{ 7509 int ret = 0; 7510 7511 mutex_lock(&rt_constraints_mutex); 7512 read_lock(&tasklist_lock); 7513 ret = __rt_schedulable(NULL, 0, 0); 7514 read_unlock(&tasklist_lock); 7515 mutex_unlock(&rt_constraints_mutex); 7516 7517 return ret; 7518} 7519 7520static int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk) 7521{ 7522 /* Don't accept realtime tasks when there is no way for them to run */ 7523 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0) 7524 return 0; 7525 7526 return 1; 7527} 7528 7529#else /* !CONFIG_RT_GROUP_SCHED */ 7530static int sched_rt_global_constraints(void) 7531{ 7532 unsigned long flags; 7533 int i, ret = 0; 7534 7535 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags); 7536 for_each_possible_cpu(i) { 7537 struct rt_rq *rt_rq = &cpu_rq(i)->rt; 7538 7539 raw_spin_lock(&rt_rq->rt_runtime_lock); 7540 rt_rq->rt_runtime = global_rt_runtime(); 7541 raw_spin_unlock(&rt_rq->rt_runtime_lock); 7542 } 7543 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags); 7544 7545 return ret; 7546} 7547#endif /* CONFIG_RT_GROUP_SCHED */ 7548 7549static int sched_dl_global_constraints(void) 7550{ 7551 u64 runtime = global_rt_runtime(); 7552 u64 period = global_rt_period(); 7553 u64 new_bw = to_ratio(period, runtime); 7554 int cpu, ret = 0; 7555 unsigned long flags; 7556 7557 /* 7558 * Here we want to check the bandwidth not being set to some 7559 * value smaller than the currently allocated bandwidth in 7560 * any of the root_domains. 7561 * 7562 * FIXME: Cycling on all the CPUs is overdoing, but simpler than 7563 * cycling on root_domains... Discussion on different/better 7564 * solutions is welcome! 7565 */ 7566 for_each_possible_cpu(cpu) { 7567 struct dl_bw *dl_b = dl_bw_of(cpu); 7568 7569 raw_spin_lock_irqsave(&dl_b->lock, flags); 7570 if (new_bw < dl_b->total_bw) 7571 ret = -EBUSY; 7572 raw_spin_unlock_irqrestore(&dl_b->lock, flags); 7573 7574 if (ret) 7575 break; 7576 } 7577 7578 return ret; 7579} 7580 7581static void sched_dl_do_global(void) 7582{ 7583 u64 new_bw = -1; 7584 int cpu; 7585 unsigned long flags; 7586 7587 def_dl_bandwidth.dl_period = global_rt_period(); 7588 def_dl_bandwidth.dl_runtime = global_rt_runtime(); 7589 7590 if (global_rt_runtime() != RUNTIME_INF) 7591 new_bw = to_ratio(global_rt_period(), global_rt_runtime()); 7592 7593 /* 7594 * FIXME: As above... 7595 */ 7596 for_each_possible_cpu(cpu) { 7597 struct dl_bw *dl_b = dl_bw_of(cpu); 7598 7599 raw_spin_lock_irqsave(&dl_b->lock, flags); 7600 dl_b->bw = new_bw; 7601 raw_spin_unlock_irqrestore(&dl_b->lock, flags); 7602 } 7603} 7604 7605static int sched_rt_global_validate(void) 7606{ 7607 if (sysctl_sched_rt_period <= 0) 7608 return -EINVAL; 7609 7610 if ((sysctl_sched_rt_runtime != RUNTIME_INF) && 7611 (sysctl_sched_rt_runtime > sysctl_sched_rt_period)) 7612 return -EINVAL; 7613 7614 return 0; 7615} 7616 7617static void sched_rt_do_global(void) 7618{ 7619 def_rt_bandwidth.rt_runtime = global_rt_runtime(); 7620 def_rt_bandwidth.rt_period = ns_to_ktime(global_rt_period()); 7621} 7622 7623int sched_rt_handler(struct ctl_table *table, int write, 7624 void __user *buffer, size_t *lenp, 7625 loff_t *ppos) 7626{ 7627 int old_period, old_runtime; 7628 static DEFINE_MUTEX(mutex); 7629 int ret; 7630 7631 mutex_lock(&mutex); 7632 old_period = sysctl_sched_rt_period; 7633 old_runtime = sysctl_sched_rt_runtime; 7634 7635 ret = proc_dointvec(table, write, buffer, lenp, ppos); 7636 7637 if (!ret && write) { 7638 ret = sched_rt_global_validate(); 7639 if (ret) 7640 goto undo; 7641 7642 ret = sched_rt_global_constraints(); 7643 if (ret) 7644 goto undo; 7645 7646 ret = sched_dl_global_constraints(); 7647 if (ret) 7648 goto undo; 7649 7650 sched_rt_do_global(); 7651 sched_dl_do_global(); 7652 } 7653 if (0) { 7654undo: 7655 sysctl_sched_rt_period = old_period; 7656 sysctl_sched_rt_runtime = old_runtime; 7657 } 7658 mutex_unlock(&mutex); 7659 7660 return ret; 7661} 7662 7663int sched_rr_handler(struct ctl_table *table, int write, 7664 void __user *buffer, size_t *lenp, 7665 loff_t *ppos) 7666{ 7667 int ret; 7668 static DEFINE_MUTEX(mutex); 7669 7670 mutex_lock(&mutex); 7671 ret = proc_dointvec(table, write, buffer, lenp, ppos); 7672 /* make sure that internally we keep jiffies */ 7673 /* also, writing zero resets timeslice to default */ 7674 if (!ret && write) { 7675 sched_rr_timeslice = sched_rr_timeslice <= 0 ? 7676 RR_TIMESLICE : msecs_to_jiffies(sched_rr_timeslice); 7677 } 7678 mutex_unlock(&mutex); 7679 return ret; 7680} 7681 7682#ifdef CONFIG_CGROUP_SCHED 7683 7684static inline struct task_group *css_tg(struct cgroup_subsys_state *css) 7685{ 7686 return css ? container_of(css, struct task_group, css) : NULL; 7687} 7688 7689static struct cgroup_subsys_state * 7690cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css) 7691{ 7692 struct task_group *parent = css_tg(parent_css); 7693 struct task_group *tg; 7694 7695 if (!parent) { 7696 /* This is early initialization for the top cgroup */ 7697 return &root_task_group.css; 7698 } 7699 7700 tg = sched_create_group(parent); 7701 if (IS_ERR(tg)) 7702 return ERR_PTR(-ENOMEM); 7703 7704 return &tg->css; 7705} 7706 7707static int cpu_cgroup_css_online(struct cgroup_subsys_state *css) 7708{ 7709 struct task_group *tg = css_tg(css); 7710 struct task_group *parent = css_tg(css->parent); 7711 7712 if (parent) 7713 sched_online_group(tg, parent); 7714 return 0; 7715} 7716 7717static void cpu_cgroup_css_free(struct cgroup_subsys_state *css) 7718{ 7719 struct task_group *tg = css_tg(css); 7720 7721 sched_destroy_group(tg); 7722} 7723 7724static void cpu_cgroup_css_offline(struct cgroup_subsys_state *css) 7725{ 7726 struct task_group *tg = css_tg(css); 7727 7728 sched_offline_group(tg); 7729} 7730 7731static int cpu_cgroup_can_attach(struct cgroup_subsys_state *css, 7732 struct cgroup_taskset *tset) 7733{ 7734 struct task_struct *task; 7735 7736 cgroup_taskset_for_each(task, tset) { 7737#ifdef CONFIG_RT_GROUP_SCHED 7738 if (!sched_rt_can_attach(css_tg(css), task)) 7739 return -EINVAL; 7740#else 7741 /* We don't support RT-tasks being in separate groups */ 7742 if (task->sched_class != &fair_sched_class) 7743 return -EINVAL; 7744#endif 7745 } 7746 return 0; 7747} 7748 7749static void cpu_cgroup_attach(struct cgroup_subsys_state *css, 7750 struct cgroup_taskset *tset) 7751{ 7752 struct task_struct *task; 7753 7754 cgroup_taskset_for_each(task, tset) 7755 sched_move_task(task); 7756} 7757 7758static void cpu_cgroup_exit(struct cgroup_subsys_state *css, 7759 struct cgroup_subsys_state *old_css, 7760 struct task_struct *task) 7761{ 7762 /* 7763 * cgroup_exit() is called in the copy_process() failure path. 7764 * Ignore this case since the task hasn't ran yet, this avoids 7765 * trying to poke a half freed task state from generic code. 7766 */ 7767 if (!(task->flags & PF_EXITING)) 7768 return; 7769 7770 sched_move_task(task); 7771} 7772 7773#ifdef CONFIG_FAIR_GROUP_SCHED 7774static int cpu_shares_write_u64(struct cgroup_subsys_state *css, 7775 struct cftype *cftype, u64 shareval) 7776{ 7777 return sched_group_set_shares(css_tg(css), scale_load(shareval)); 7778} 7779 7780static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css, 7781 struct cftype *cft) 7782{ 7783 struct task_group *tg = css_tg(css); 7784 7785 return (u64) scale_load_down(tg->shares); 7786} 7787 7788#ifdef CONFIG_CFS_BANDWIDTH 7789static DEFINE_MUTEX(cfs_constraints_mutex); 7790 7791const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */ 7792const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */ 7793 7794static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime); 7795 7796static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota) 7797{ 7798 int i, ret = 0, runtime_enabled, runtime_was_enabled; 7799 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth; 7800 7801 if (tg == &root_task_group) 7802 return -EINVAL; 7803 7804 /* 7805 * Ensure we have at some amount of bandwidth every period. This is 7806 * to prevent reaching a state of large arrears when throttled via 7807 * entity_tick() resulting in prolonged exit starvation. 7808 */ 7809 if (quota < min_cfs_quota_period || period < min_cfs_quota_period) 7810 return -EINVAL; 7811 7812 /* 7813 * Likewise, bound things on the otherside by preventing insane quota 7814 * periods. This also allows us to normalize in computing quota 7815 * feasibility. 7816 */ 7817 if (period > max_cfs_quota_period) 7818 return -EINVAL; 7819 7820 /* 7821 * Prevent race between setting of cfs_rq->runtime_enabled and 7822 * unthrottle_offline_cfs_rqs(). 7823 */ 7824 get_online_cpus(); 7825 mutex_lock(&cfs_constraints_mutex); 7826 ret = __cfs_schedulable(tg, period, quota); 7827 if (ret) 7828 goto out_unlock; 7829 7830 runtime_enabled = quota != RUNTIME_INF; 7831 runtime_was_enabled = cfs_b->quota != RUNTIME_INF; 7832 /* 7833 * If we need to toggle cfs_bandwidth_used, off->on must occur 7834 * before making related changes, and on->off must occur afterwards 7835 */ 7836 if (runtime_enabled && !runtime_was_enabled) 7837 cfs_bandwidth_usage_inc(); 7838 raw_spin_lock_irq(&cfs_b->lock); 7839 cfs_b->period = ns_to_ktime(period); 7840 cfs_b->quota = quota; 7841 7842 __refill_cfs_bandwidth_runtime(cfs_b); 7843 /* restart the period timer (if active) to handle new period expiry */ 7844 if (runtime_enabled && cfs_b->timer_active) { 7845 /* force a reprogram */ 7846 __start_cfs_bandwidth(cfs_b, true); 7847 } 7848 raw_spin_unlock_irq(&cfs_b->lock); 7849 7850 for_each_online_cpu(i) { 7851 struct cfs_rq *cfs_rq = tg->cfs_rq[i]; 7852 struct rq *rq = cfs_rq->rq; 7853 7854 raw_spin_lock_irq(&rq->lock); 7855 cfs_rq->runtime_enabled = runtime_enabled; 7856 cfs_rq->runtime_remaining = 0; 7857 7858 if (cfs_rq->throttled) 7859 unthrottle_cfs_rq(cfs_rq); 7860 raw_spin_unlock_irq(&rq->lock); 7861 } 7862 if (runtime_was_enabled && !runtime_enabled) 7863 cfs_bandwidth_usage_dec(); 7864out_unlock: 7865 mutex_unlock(&cfs_constraints_mutex); 7866 put_online_cpus(); 7867 7868 return ret; 7869} 7870 7871int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us) 7872{ 7873 u64 quota, period; 7874 7875 period = ktime_to_ns(tg->cfs_bandwidth.period); 7876 if (cfs_quota_us < 0) 7877 quota = RUNTIME_INF; 7878 else 7879 quota = (u64)cfs_quota_us * NSEC_PER_USEC; 7880 7881 return tg_set_cfs_bandwidth(tg, period, quota); 7882} 7883 7884long tg_get_cfs_quota(struct task_group *tg) 7885{ 7886 u64 quota_us; 7887 7888 if (tg->cfs_bandwidth.quota == RUNTIME_INF) 7889 return -1; 7890 7891 quota_us = tg->cfs_bandwidth.quota; 7892 do_div(quota_us, NSEC_PER_USEC); 7893 7894 return quota_us; 7895} 7896 7897int tg_set_cfs_period(struct task_group *tg, long cfs_period_us) 7898{ 7899 u64 quota, period; 7900 7901 period = (u64)cfs_period_us * NSEC_PER_USEC; 7902 quota = tg->cfs_bandwidth.quota; 7903 7904 return tg_set_cfs_bandwidth(tg, period, quota); 7905} 7906 7907long tg_get_cfs_period(struct task_group *tg) 7908{ 7909 u64 cfs_period_us; 7910 7911 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period); 7912 do_div(cfs_period_us, NSEC_PER_USEC); 7913 7914 return cfs_period_us; 7915} 7916 7917static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css, 7918 struct cftype *cft) 7919{ 7920 return tg_get_cfs_quota(css_tg(css)); 7921} 7922 7923static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css, 7924 struct cftype *cftype, s64 cfs_quota_us) 7925{ 7926 return tg_set_cfs_quota(css_tg(css), cfs_quota_us); 7927} 7928 7929static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css, 7930 struct cftype *cft) 7931{ 7932 return tg_get_cfs_period(css_tg(css)); 7933} 7934 7935static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css, 7936 struct cftype *cftype, u64 cfs_period_us) 7937{ 7938 return tg_set_cfs_period(css_tg(css), cfs_period_us); 7939} 7940 7941struct cfs_schedulable_data { 7942 struct task_group *tg; 7943 u64 period, quota; 7944}; 7945 7946/* 7947 * normalize group quota/period to be quota/max_period 7948 * note: units are usecs 7949 */ 7950static u64 normalize_cfs_quota(struct task_group *tg, 7951 struct cfs_schedulable_data *d) 7952{ 7953 u64 quota, period; 7954 7955 if (tg == d->tg) { 7956 period = d->period; 7957 quota = d->quota; 7958 } else { 7959 period = tg_get_cfs_period(tg); 7960 quota = tg_get_cfs_quota(tg); 7961 } 7962 7963 /* note: these should typically be equivalent */ 7964 if (quota == RUNTIME_INF || quota == -1) 7965 return RUNTIME_INF; 7966 7967 return to_ratio(period, quota); 7968} 7969 7970static int tg_cfs_schedulable_down(struct task_group *tg, void *data) 7971{ 7972 struct cfs_schedulable_data *d = data; 7973 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth; 7974 s64 quota = 0, parent_quota = -1; 7975 7976 if (!tg->parent) { 7977 quota = RUNTIME_INF; 7978 } else { 7979 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth; 7980 7981 quota = normalize_cfs_quota(tg, d); 7982 parent_quota = parent_b->hierarchal_quota; 7983 7984 /* 7985 * ensure max(child_quota) <= parent_quota, inherit when no 7986 * limit is set 7987 */ 7988 if (quota == RUNTIME_INF) 7989 quota = parent_quota; 7990 else if (parent_quota != RUNTIME_INF && quota > parent_quota) 7991 return -EINVAL; 7992 } 7993 cfs_b->hierarchal_quota = quota; 7994 7995 return 0; 7996} 7997 7998static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota) 7999{ 8000 int ret; 8001 struct cfs_schedulable_data data = { 8002 .tg = tg, 8003 .period = period, 8004 .quota = quota, 8005 }; 8006 8007 if (quota != RUNTIME_INF) { 8008 do_div(data.period, NSEC_PER_USEC); 8009 do_div(data.quota, NSEC_PER_USEC); 8010 } 8011 8012 rcu_read_lock(); 8013 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data); 8014 rcu_read_unlock(); 8015 8016 return ret; 8017} 8018 8019static int cpu_stats_show(struct seq_file *sf, void *v) 8020{ 8021 struct task_group *tg = css_tg(seq_css(sf)); 8022 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth; 8023 8024 seq_printf(sf, "nr_periods %d\n", cfs_b->nr_periods); 8025 seq_printf(sf, "nr_throttled %d\n", cfs_b->nr_throttled); 8026 seq_printf(sf, "throttled_time %llu\n", cfs_b->throttled_time); 8027 8028 return 0; 8029} 8030#endif /* CONFIG_CFS_BANDWIDTH */ 8031#endif /* CONFIG_FAIR_GROUP_SCHED */ 8032 8033#ifdef CONFIG_RT_GROUP_SCHED 8034static int cpu_rt_runtime_write(struct cgroup_subsys_state *css, 8035 struct cftype *cft, s64 val) 8036{ 8037 return sched_group_set_rt_runtime(css_tg(css), val); 8038} 8039 8040static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css, 8041 struct cftype *cft) 8042{ 8043 return sched_group_rt_runtime(css_tg(css)); 8044} 8045 8046static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css, 8047 struct cftype *cftype, u64 rt_period_us) 8048{ 8049 return sched_group_set_rt_period(css_tg(css), rt_period_us); 8050} 8051 8052static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css, 8053 struct cftype *cft) 8054{ 8055 return sched_group_rt_period(css_tg(css)); 8056} 8057#endif /* CONFIG_RT_GROUP_SCHED */ 8058 8059static struct cftype cpu_files[] = { 8060#ifdef CONFIG_FAIR_GROUP_SCHED 8061 { 8062 .name = "shares", 8063 .read_u64 = cpu_shares_read_u64, 8064 .write_u64 = cpu_shares_write_u64, 8065 }, 8066#endif 8067#ifdef CONFIG_CFS_BANDWIDTH 8068 { 8069 .name = "cfs_quota_us", 8070 .read_s64 = cpu_cfs_quota_read_s64, 8071 .write_s64 = cpu_cfs_quota_write_s64, 8072 }, 8073 { 8074 .name = "cfs_period_us", 8075 .read_u64 = cpu_cfs_period_read_u64, 8076 .write_u64 = cpu_cfs_period_write_u64, 8077 }, 8078 { 8079 .name = "stat", 8080 .seq_show = cpu_stats_show, 8081 }, 8082#endif 8083#ifdef CONFIG_RT_GROUP_SCHED 8084 { 8085 .name = "rt_runtime_us", 8086 .read_s64 = cpu_rt_runtime_read, 8087 .write_s64 = cpu_rt_runtime_write, 8088 }, 8089 { 8090 .name = "rt_period_us", 8091 .read_u64 = cpu_rt_period_read_uint, 8092 .write_u64 = cpu_rt_period_write_uint, 8093 }, 8094#endif 8095 { } /* terminate */ 8096}; 8097 8098struct cgroup_subsys cpu_cgrp_subsys = { 8099 .css_alloc = cpu_cgroup_css_alloc, 8100 .css_free = cpu_cgroup_css_free, 8101 .css_online = cpu_cgroup_css_online, 8102 .css_offline = cpu_cgroup_css_offline, 8103 .can_attach = cpu_cgroup_can_attach, 8104 .attach = cpu_cgroup_attach, 8105 .exit = cpu_cgroup_exit, 8106 .base_cftypes = cpu_files, 8107 .early_init = 1, 8108}; 8109 8110#endif /* CONFIG_CGROUP_SCHED */ 8111 8112void dump_cpu_task(int cpu) 8113{ 8114 pr_info("Task dump for CPU %d:\n", cpu); 8115 sched_show_task(cpu_curr(cpu)); 8116} 8117