core.c revision 3882ec643997757824cd5f25180cd8a787b9dbe1
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 rq->clock += delta; 143 update_rq_clock_task(rq, delta); 144} 145 146/* 147 * Debugging: various feature bits 148 */ 149 150#define SCHED_FEAT(name, enabled) \ 151 (1UL << __SCHED_FEAT_##name) * enabled | 152 153const_debug unsigned int sysctl_sched_features = 154#include "features.h" 155 0; 156 157#undef SCHED_FEAT 158 159#ifdef CONFIG_SCHED_DEBUG 160#define SCHED_FEAT(name, enabled) \ 161 #name , 162 163static const char * const sched_feat_names[] = { 164#include "features.h" 165}; 166 167#undef SCHED_FEAT 168 169static int sched_feat_show(struct seq_file *m, void *v) 170{ 171 int i; 172 173 for (i = 0; i < __SCHED_FEAT_NR; i++) { 174 if (!(sysctl_sched_features & (1UL << i))) 175 seq_puts(m, "NO_"); 176 seq_printf(m, "%s ", sched_feat_names[i]); 177 } 178 seq_puts(m, "\n"); 179 180 return 0; 181} 182 183#ifdef HAVE_JUMP_LABEL 184 185#define jump_label_key__true STATIC_KEY_INIT_TRUE 186#define jump_label_key__false STATIC_KEY_INIT_FALSE 187 188#define SCHED_FEAT(name, enabled) \ 189 jump_label_key__##enabled , 190 191struct static_key sched_feat_keys[__SCHED_FEAT_NR] = { 192#include "features.h" 193}; 194 195#undef SCHED_FEAT 196 197static void sched_feat_disable(int i) 198{ 199 if (static_key_enabled(&sched_feat_keys[i])) 200 static_key_slow_dec(&sched_feat_keys[i]); 201} 202 203static void sched_feat_enable(int i) 204{ 205 if (!static_key_enabled(&sched_feat_keys[i])) 206 static_key_slow_inc(&sched_feat_keys[i]); 207} 208#else 209static void sched_feat_disable(int i) { }; 210static void sched_feat_enable(int i) { }; 211#endif /* HAVE_JUMP_LABEL */ 212 213static int sched_feat_set(char *cmp) 214{ 215 int i; 216 int neg = 0; 217 218 if (strncmp(cmp, "NO_", 3) == 0) { 219 neg = 1; 220 cmp += 3; 221 } 222 223 for (i = 0; i < __SCHED_FEAT_NR; i++) { 224 if (strcmp(cmp, sched_feat_names[i]) == 0) { 225 if (neg) { 226 sysctl_sched_features &= ~(1UL << i); 227 sched_feat_disable(i); 228 } else { 229 sysctl_sched_features |= (1UL << i); 230 sched_feat_enable(i); 231 } 232 break; 233 } 234 } 235 236 return i; 237} 238 239static ssize_t 240sched_feat_write(struct file *filp, const char __user *ubuf, 241 size_t cnt, loff_t *ppos) 242{ 243 char buf[64]; 244 char *cmp; 245 int i; 246 247 if (cnt > 63) 248 cnt = 63; 249 250 if (copy_from_user(&buf, ubuf, cnt)) 251 return -EFAULT; 252 253 buf[cnt] = 0; 254 cmp = strstrip(buf); 255 256 i = sched_feat_set(cmp); 257 if (i == __SCHED_FEAT_NR) 258 return -EINVAL; 259 260 *ppos += cnt; 261 262 return cnt; 263} 264 265static int sched_feat_open(struct inode *inode, struct file *filp) 266{ 267 return single_open(filp, sched_feat_show, NULL); 268} 269 270static const struct file_operations sched_feat_fops = { 271 .open = sched_feat_open, 272 .write = sched_feat_write, 273 .read = seq_read, 274 .llseek = seq_lseek, 275 .release = single_release, 276}; 277 278static __init int sched_init_debug(void) 279{ 280 debugfs_create_file("sched_features", 0644, NULL, NULL, 281 &sched_feat_fops); 282 283 return 0; 284} 285late_initcall(sched_init_debug); 286#endif /* CONFIG_SCHED_DEBUG */ 287 288/* 289 * Number of tasks to iterate in a single balance run. 290 * Limited because this is done with IRQs disabled. 291 */ 292const_debug unsigned int sysctl_sched_nr_migrate = 32; 293 294/* 295 * period over which we average the RT time consumption, measured 296 * in ms. 297 * 298 * default: 1s 299 */ 300const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC; 301 302/* 303 * period over which we measure -rt task cpu usage in us. 304 * default: 1s 305 */ 306unsigned int sysctl_sched_rt_period = 1000000; 307 308__read_mostly int scheduler_running; 309 310/* 311 * part of the period that we allow rt tasks to run in us. 312 * default: 0.95s 313 */ 314int sysctl_sched_rt_runtime = 950000; 315 316/* 317 * __task_rq_lock - lock the rq @p resides on. 318 */ 319static inline struct rq *__task_rq_lock(struct task_struct *p) 320 __acquires(rq->lock) 321{ 322 struct rq *rq; 323 324 lockdep_assert_held(&p->pi_lock); 325 326 for (;;) { 327 rq = task_rq(p); 328 raw_spin_lock(&rq->lock); 329 if (likely(rq == task_rq(p))) 330 return rq; 331 raw_spin_unlock(&rq->lock); 332 } 333} 334 335/* 336 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on. 337 */ 338static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags) 339 __acquires(p->pi_lock) 340 __acquires(rq->lock) 341{ 342 struct rq *rq; 343 344 for (;;) { 345 raw_spin_lock_irqsave(&p->pi_lock, *flags); 346 rq = task_rq(p); 347 raw_spin_lock(&rq->lock); 348 if (likely(rq == task_rq(p))) 349 return rq; 350 raw_spin_unlock(&rq->lock); 351 raw_spin_unlock_irqrestore(&p->pi_lock, *flags); 352 } 353} 354 355static void __task_rq_unlock(struct rq *rq) 356 __releases(rq->lock) 357{ 358 raw_spin_unlock(&rq->lock); 359} 360 361static inline void 362task_rq_unlock(struct rq *rq, struct task_struct *p, unsigned long *flags) 363 __releases(rq->lock) 364 __releases(p->pi_lock) 365{ 366 raw_spin_unlock(&rq->lock); 367 raw_spin_unlock_irqrestore(&p->pi_lock, *flags); 368} 369 370/* 371 * this_rq_lock - lock this runqueue and disable interrupts. 372 */ 373static struct rq *this_rq_lock(void) 374 __acquires(rq->lock) 375{ 376 struct rq *rq; 377 378 local_irq_disable(); 379 rq = this_rq(); 380 raw_spin_lock(&rq->lock); 381 382 return rq; 383} 384 385#ifdef CONFIG_SCHED_HRTICK 386/* 387 * Use HR-timers to deliver accurate preemption points. 388 */ 389 390static void hrtick_clear(struct rq *rq) 391{ 392 if (hrtimer_active(&rq->hrtick_timer)) 393 hrtimer_cancel(&rq->hrtick_timer); 394} 395 396/* 397 * High-resolution timer tick. 398 * Runs from hardirq context with interrupts disabled. 399 */ 400static enum hrtimer_restart hrtick(struct hrtimer *timer) 401{ 402 struct rq *rq = container_of(timer, struct rq, hrtick_timer); 403 404 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id()); 405 406 raw_spin_lock(&rq->lock); 407 update_rq_clock(rq); 408 rq->curr->sched_class->task_tick(rq, rq->curr, 1); 409 raw_spin_unlock(&rq->lock); 410 411 return HRTIMER_NORESTART; 412} 413 414#ifdef CONFIG_SMP 415 416static int __hrtick_restart(struct rq *rq) 417{ 418 struct hrtimer *timer = &rq->hrtick_timer; 419 ktime_t time = hrtimer_get_softexpires(timer); 420 421 return __hrtimer_start_range_ns(timer, time, 0, HRTIMER_MODE_ABS_PINNED, 0); 422} 423 424/* 425 * called from hardirq (IPI) context 426 */ 427static void __hrtick_start(void *arg) 428{ 429 struct rq *rq = arg; 430 431 raw_spin_lock(&rq->lock); 432 __hrtick_restart(rq); 433 rq->hrtick_csd_pending = 0; 434 raw_spin_unlock(&rq->lock); 435} 436 437/* 438 * Called to set the hrtick timer state. 439 * 440 * called with rq->lock held and irqs disabled 441 */ 442void hrtick_start(struct rq *rq, u64 delay) 443{ 444 struct hrtimer *timer = &rq->hrtick_timer; 445 ktime_t time = ktime_add_ns(timer->base->get_time(), delay); 446 447 hrtimer_set_expires(timer, time); 448 449 if (rq == this_rq()) { 450 __hrtick_restart(rq); 451 } else if (!rq->hrtick_csd_pending) { 452 smp_call_function_single_async(cpu_of(rq), &rq->hrtick_csd); 453 rq->hrtick_csd_pending = 1; 454 } 455} 456 457static int 458hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu) 459{ 460 int cpu = (int)(long)hcpu; 461 462 switch (action) { 463 case CPU_UP_CANCELED: 464 case CPU_UP_CANCELED_FROZEN: 465 case CPU_DOWN_PREPARE: 466 case CPU_DOWN_PREPARE_FROZEN: 467 case CPU_DEAD: 468 case CPU_DEAD_FROZEN: 469 hrtick_clear(cpu_rq(cpu)); 470 return NOTIFY_OK; 471 } 472 473 return NOTIFY_DONE; 474} 475 476static __init void init_hrtick(void) 477{ 478 hotcpu_notifier(hotplug_hrtick, 0); 479} 480#else 481/* 482 * Called to set the hrtick timer state. 483 * 484 * called with rq->lock held and irqs disabled 485 */ 486void hrtick_start(struct rq *rq, u64 delay) 487{ 488 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0, 489 HRTIMER_MODE_REL_PINNED, 0); 490} 491 492static inline void init_hrtick(void) 493{ 494} 495#endif /* CONFIG_SMP */ 496 497static void init_rq_hrtick(struct rq *rq) 498{ 499#ifdef CONFIG_SMP 500 rq->hrtick_csd_pending = 0; 501 502 rq->hrtick_csd.flags = 0; 503 rq->hrtick_csd.func = __hrtick_start; 504 rq->hrtick_csd.info = rq; 505#endif 506 507 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); 508 rq->hrtick_timer.function = hrtick; 509} 510#else /* CONFIG_SCHED_HRTICK */ 511static inline void hrtick_clear(struct rq *rq) 512{ 513} 514 515static inline void init_rq_hrtick(struct rq *rq) 516{ 517} 518 519static inline void init_hrtick(void) 520{ 521} 522#endif /* CONFIG_SCHED_HRTICK */ 523 524/* 525 * cmpxchg based fetch_or, macro so it works for different integer types 526 */ 527#define fetch_or(ptr, val) \ 528({ typeof(*(ptr)) __old, __val = *(ptr); \ 529 for (;;) { \ 530 __old = cmpxchg((ptr), __val, __val | (val)); \ 531 if (__old == __val) \ 532 break; \ 533 __val = __old; \ 534 } \ 535 __old; \ 536}) 537 538#if defined(CONFIG_SMP) && defined(TIF_POLLING_NRFLAG) 539/* 540 * Atomically set TIF_NEED_RESCHED and test for TIF_POLLING_NRFLAG, 541 * this avoids any races wrt polling state changes and thereby avoids 542 * spurious IPIs. 543 */ 544static bool set_nr_and_not_polling(struct task_struct *p) 545{ 546 struct thread_info *ti = task_thread_info(p); 547 return !(fetch_or(&ti->flags, _TIF_NEED_RESCHED) & _TIF_POLLING_NRFLAG); 548} 549 550/* 551 * Atomically set TIF_NEED_RESCHED if TIF_POLLING_NRFLAG is set. 552 * 553 * If this returns true, then the idle task promises to call 554 * sched_ttwu_pending() and reschedule soon. 555 */ 556static bool set_nr_if_polling(struct task_struct *p) 557{ 558 struct thread_info *ti = task_thread_info(p); 559 typeof(ti->flags) old, val = ACCESS_ONCE(ti->flags); 560 561 for (;;) { 562 if (!(val & _TIF_POLLING_NRFLAG)) 563 return false; 564 if (val & _TIF_NEED_RESCHED) 565 return true; 566 old = cmpxchg(&ti->flags, val, val | _TIF_NEED_RESCHED); 567 if (old == val) 568 break; 569 val = old; 570 } 571 return true; 572} 573 574#else 575static bool set_nr_and_not_polling(struct task_struct *p) 576{ 577 set_tsk_need_resched(p); 578 return true; 579} 580 581#ifdef CONFIG_SMP 582static bool set_nr_if_polling(struct task_struct *p) 583{ 584 return false; 585} 586#endif 587#endif 588 589/* 590 * resched_task - mark a task 'to be rescheduled now'. 591 * 592 * On UP this means the setting of the need_resched flag, on SMP it 593 * might also involve a cross-CPU call to trigger the scheduler on 594 * the target CPU. 595 */ 596void resched_task(struct task_struct *p) 597{ 598 int cpu; 599 600 lockdep_assert_held(&task_rq(p)->lock); 601 602 if (test_tsk_need_resched(p)) 603 return; 604 605 cpu = task_cpu(p); 606 607 if (cpu == smp_processor_id()) { 608 set_tsk_need_resched(p); 609 set_preempt_need_resched(); 610 return; 611 } 612 613 if (set_nr_and_not_polling(p)) 614 smp_send_reschedule(cpu); 615 else 616 trace_sched_wake_idle_without_ipi(cpu); 617} 618 619void resched_cpu(int cpu) 620{ 621 struct rq *rq = cpu_rq(cpu); 622 unsigned long flags; 623 624 if (!raw_spin_trylock_irqsave(&rq->lock, flags)) 625 return; 626 resched_task(cpu_curr(cpu)); 627 raw_spin_unlock_irqrestore(&rq->lock, flags); 628} 629 630#ifdef CONFIG_SMP 631#ifdef CONFIG_NO_HZ_COMMON 632/* 633 * In the semi idle case, use the nearest busy cpu for migrating timers 634 * from an idle cpu. This is good for power-savings. 635 * 636 * We don't do similar optimization for completely idle system, as 637 * selecting an idle cpu will add more delays to the timers than intended 638 * (as that cpu's timer base may not be uptodate wrt jiffies etc). 639 */ 640int get_nohz_timer_target(int pinned) 641{ 642 int cpu = smp_processor_id(); 643 int i; 644 struct sched_domain *sd; 645 646 if (pinned || !get_sysctl_timer_migration() || !idle_cpu(cpu)) 647 return cpu; 648 649 rcu_read_lock(); 650 for_each_domain(cpu, sd) { 651 for_each_cpu(i, sched_domain_span(sd)) { 652 if (!idle_cpu(i)) { 653 cpu = i; 654 goto unlock; 655 } 656 } 657 } 658unlock: 659 rcu_read_unlock(); 660 return cpu; 661} 662/* 663 * When add_timer_on() enqueues a timer into the timer wheel of an 664 * idle CPU then this timer might expire before the next timer event 665 * which is scheduled to wake up that CPU. In case of a completely 666 * idle system the next event might even be infinite time into the 667 * future. wake_up_idle_cpu() ensures that the CPU is woken up and 668 * leaves the inner idle loop so the newly added timer is taken into 669 * account when the CPU goes back to idle and evaluates the timer 670 * wheel for the next timer event. 671 */ 672static void wake_up_idle_cpu(int cpu) 673{ 674 struct rq *rq = cpu_rq(cpu); 675 676 if (cpu == smp_processor_id()) 677 return; 678 679 if (set_nr_and_not_polling(rq->idle)) 680 smp_send_reschedule(cpu); 681 else 682 trace_sched_wake_idle_without_ipi(cpu); 683} 684 685static bool wake_up_full_nohz_cpu(int cpu) 686{ 687 /* 688 * We just need the target to call irq_exit() and re-evaluate 689 * the next tick. The nohz full kick at least implies that. 690 * If needed we can still optimize that later with an 691 * empty IRQ. 692 */ 693 if (tick_nohz_full_cpu(cpu)) { 694 if (cpu != smp_processor_id() || 695 tick_nohz_tick_stopped()) 696 tick_nohz_full_kick_cpu(cpu); 697 return true; 698 } 699 700 return false; 701} 702 703void wake_up_nohz_cpu(int cpu) 704{ 705 if (!wake_up_full_nohz_cpu(cpu)) 706 wake_up_idle_cpu(cpu); 707} 708 709static inline bool got_nohz_idle_kick(void) 710{ 711 int cpu = smp_processor_id(); 712 713 if (!test_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu))) 714 return false; 715 716 if (idle_cpu(cpu) && !need_resched()) 717 return true; 718 719 /* 720 * We can't run Idle Load Balance on this CPU for this time so we 721 * cancel it and clear NOHZ_BALANCE_KICK 722 */ 723 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu)); 724 return false; 725} 726 727#else /* CONFIG_NO_HZ_COMMON */ 728 729static inline bool got_nohz_idle_kick(void) 730{ 731 return false; 732} 733 734#endif /* CONFIG_NO_HZ_COMMON */ 735 736#ifdef CONFIG_NO_HZ_FULL 737bool sched_can_stop_tick(void) 738{ 739 struct rq *rq; 740 741 rq = this_rq(); 742 743 /* 744 * More than one running task need preemption. 745 * nr_running update is assumed to be visible 746 * after IPI is sent from wakers. 747 */ 748 if (rq->nr_running > 1) 749 return false; 750 751 return true; 752} 753#endif /* CONFIG_NO_HZ_FULL */ 754 755void sched_avg_update(struct rq *rq) 756{ 757 s64 period = sched_avg_period(); 758 759 while ((s64)(rq_clock(rq) - rq->age_stamp) > period) { 760 /* 761 * Inline assembly required to prevent the compiler 762 * optimising this loop into a divmod call. 763 * See __iter_div_u64_rem() for another example of this. 764 */ 765 asm("" : "+rm" (rq->age_stamp)); 766 rq->age_stamp += period; 767 rq->rt_avg /= 2; 768 } 769} 770 771#endif /* CONFIG_SMP */ 772 773#if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \ 774 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH))) 775/* 776 * Iterate task_group tree rooted at *from, calling @down when first entering a 777 * node and @up when leaving it for the final time. 778 * 779 * Caller must hold rcu_lock or sufficient equivalent. 780 */ 781int walk_tg_tree_from(struct task_group *from, 782 tg_visitor down, tg_visitor up, void *data) 783{ 784 struct task_group *parent, *child; 785 int ret; 786 787 parent = from; 788 789down: 790 ret = (*down)(parent, data); 791 if (ret) 792 goto out; 793 list_for_each_entry_rcu(child, &parent->children, siblings) { 794 parent = child; 795 goto down; 796 797up: 798 continue; 799 } 800 ret = (*up)(parent, data); 801 if (ret || parent == from) 802 goto out; 803 804 child = parent; 805 parent = parent->parent; 806 if (parent) 807 goto up; 808out: 809 return ret; 810} 811 812int tg_nop(struct task_group *tg, void *data) 813{ 814 return 0; 815} 816#endif 817 818static void set_load_weight(struct task_struct *p) 819{ 820 int prio = p->static_prio - MAX_RT_PRIO; 821 struct load_weight *load = &p->se.load; 822 823 /* 824 * SCHED_IDLE tasks get minimal weight: 825 */ 826 if (p->policy == SCHED_IDLE) { 827 load->weight = scale_load(WEIGHT_IDLEPRIO); 828 load->inv_weight = WMULT_IDLEPRIO; 829 return; 830 } 831 832 load->weight = scale_load(prio_to_weight[prio]); 833 load->inv_weight = prio_to_wmult[prio]; 834} 835 836static void enqueue_task(struct rq *rq, struct task_struct *p, int flags) 837{ 838 update_rq_clock(rq); 839 sched_info_queued(rq, p); 840 p->sched_class->enqueue_task(rq, p, flags); 841} 842 843static void dequeue_task(struct rq *rq, struct task_struct *p, int flags) 844{ 845 update_rq_clock(rq); 846 sched_info_dequeued(rq, p); 847 p->sched_class->dequeue_task(rq, p, flags); 848} 849 850void activate_task(struct rq *rq, struct task_struct *p, int flags) 851{ 852 if (task_contributes_to_load(p)) 853 rq->nr_uninterruptible--; 854 855 enqueue_task(rq, p, flags); 856} 857 858void deactivate_task(struct rq *rq, struct task_struct *p, int flags) 859{ 860 if (task_contributes_to_load(p)) 861 rq->nr_uninterruptible++; 862 863 dequeue_task(rq, p, flags); 864} 865 866static void update_rq_clock_task(struct rq *rq, s64 delta) 867{ 868/* 869 * In theory, the compile should just see 0 here, and optimize out the call 870 * to sched_rt_avg_update. But I don't trust it... 871 */ 872#if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING) 873 s64 steal = 0, irq_delta = 0; 874#endif 875#ifdef CONFIG_IRQ_TIME_ACCOUNTING 876 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time; 877 878 /* 879 * Since irq_time is only updated on {soft,}irq_exit, we might run into 880 * this case when a previous update_rq_clock() happened inside a 881 * {soft,}irq region. 882 * 883 * When this happens, we stop ->clock_task and only update the 884 * prev_irq_time stamp to account for the part that fit, so that a next 885 * update will consume the rest. This ensures ->clock_task is 886 * monotonic. 887 * 888 * It does however cause some slight miss-attribution of {soft,}irq 889 * time, a more accurate solution would be to update the irq_time using 890 * the current rq->clock timestamp, except that would require using 891 * atomic ops. 892 */ 893 if (irq_delta > delta) 894 irq_delta = delta; 895 896 rq->prev_irq_time += irq_delta; 897 delta -= irq_delta; 898#endif 899#ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING 900 if (static_key_false((¶virt_steal_rq_enabled))) { 901 steal = paravirt_steal_clock(cpu_of(rq)); 902 steal -= rq->prev_steal_time_rq; 903 904 if (unlikely(steal > delta)) 905 steal = delta; 906 907 rq->prev_steal_time_rq += steal; 908 delta -= steal; 909 } 910#endif 911 912 rq->clock_task += delta; 913 914#if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING) 915 if ((irq_delta + steal) && sched_feat(NONTASK_CAPACITY)) 916 sched_rt_avg_update(rq, irq_delta + steal); 917#endif 918} 919 920void sched_set_stop_task(int cpu, struct task_struct *stop) 921{ 922 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 }; 923 struct task_struct *old_stop = cpu_rq(cpu)->stop; 924 925 if (stop) { 926 /* 927 * Make it appear like a SCHED_FIFO task, its something 928 * userspace knows about and won't get confused about. 929 * 930 * Also, it will make PI more or less work without too 931 * much confusion -- but then, stop work should not 932 * rely on PI working anyway. 933 */ 934 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m); 935 936 stop->sched_class = &stop_sched_class; 937 } 938 939 cpu_rq(cpu)->stop = stop; 940 941 if (old_stop) { 942 /* 943 * Reset it back to a normal scheduling class so that 944 * it can die in pieces. 945 */ 946 old_stop->sched_class = &rt_sched_class; 947 } 948} 949 950/* 951 * __normal_prio - return the priority that is based on the static prio 952 */ 953static inline int __normal_prio(struct task_struct *p) 954{ 955 return p->static_prio; 956} 957 958/* 959 * Calculate the expected normal priority: i.e. priority 960 * without taking RT-inheritance into account. Might be 961 * boosted by interactivity modifiers. Changes upon fork, 962 * setprio syscalls, and whenever the interactivity 963 * estimator recalculates. 964 */ 965static inline int normal_prio(struct task_struct *p) 966{ 967 int prio; 968 969 if (task_has_dl_policy(p)) 970 prio = MAX_DL_PRIO-1; 971 else if (task_has_rt_policy(p)) 972 prio = MAX_RT_PRIO-1 - p->rt_priority; 973 else 974 prio = __normal_prio(p); 975 return prio; 976} 977 978/* 979 * Calculate the current priority, i.e. the priority 980 * taken into account by the scheduler. This value might 981 * be boosted by RT tasks, or might be boosted by 982 * interactivity modifiers. Will be RT if the task got 983 * RT-boosted. If not then it returns p->normal_prio. 984 */ 985static int effective_prio(struct task_struct *p) 986{ 987 p->normal_prio = normal_prio(p); 988 /* 989 * If we are RT tasks or we were boosted to RT priority, 990 * keep the priority unchanged. Otherwise, update priority 991 * to the normal priority: 992 */ 993 if (!rt_prio(p->prio)) 994 return p->normal_prio; 995 return p->prio; 996} 997 998/** 999 * task_curr - is this task currently executing on a CPU? 1000 * @p: the task in question. 1001 * 1002 * Return: 1 if the task is currently executing. 0 otherwise. 1003 */ 1004inline int task_curr(const struct task_struct *p) 1005{ 1006 return cpu_curr(task_cpu(p)) == p; 1007} 1008 1009static inline void check_class_changed(struct rq *rq, struct task_struct *p, 1010 const struct sched_class *prev_class, 1011 int oldprio) 1012{ 1013 if (prev_class != p->sched_class) { 1014 if (prev_class->switched_from) 1015 prev_class->switched_from(rq, p); 1016 p->sched_class->switched_to(rq, p); 1017 } else if (oldprio != p->prio || dl_task(p)) 1018 p->sched_class->prio_changed(rq, p, oldprio); 1019} 1020 1021void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags) 1022{ 1023 const struct sched_class *class; 1024 1025 if (p->sched_class == rq->curr->sched_class) { 1026 rq->curr->sched_class->check_preempt_curr(rq, p, flags); 1027 } else { 1028 for_each_class(class) { 1029 if (class == rq->curr->sched_class) 1030 break; 1031 if (class == p->sched_class) { 1032 resched_task(rq->curr); 1033 break; 1034 } 1035 } 1036 } 1037 1038 /* 1039 * A queue event has occurred, and we're going to schedule. In 1040 * this case, we can save a useless back to back clock update. 1041 */ 1042 if (rq->curr->on_rq && test_tsk_need_resched(rq->curr)) 1043 rq->skip_clock_update = 1; 1044} 1045 1046#ifdef CONFIG_SMP 1047void set_task_cpu(struct task_struct *p, unsigned int new_cpu) 1048{ 1049#ifdef CONFIG_SCHED_DEBUG 1050 /* 1051 * We should never call set_task_cpu() on a blocked task, 1052 * ttwu() will sort out the placement. 1053 */ 1054 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING && 1055 !(task_preempt_count(p) & PREEMPT_ACTIVE)); 1056 1057#ifdef CONFIG_LOCKDEP 1058 /* 1059 * The caller should hold either p->pi_lock or rq->lock, when changing 1060 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks. 1061 * 1062 * sched_move_task() holds both and thus holding either pins the cgroup, 1063 * see task_group(). 1064 * 1065 * Furthermore, all task_rq users should acquire both locks, see 1066 * task_rq_lock(). 1067 */ 1068 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) || 1069 lockdep_is_held(&task_rq(p)->lock))); 1070#endif 1071#endif 1072 1073 trace_sched_migrate_task(p, new_cpu); 1074 1075 if (task_cpu(p) != new_cpu) { 1076 if (p->sched_class->migrate_task_rq) 1077 p->sched_class->migrate_task_rq(p, new_cpu); 1078 p->se.nr_migrations++; 1079 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, NULL, 0); 1080 } 1081 1082 __set_task_cpu(p, new_cpu); 1083} 1084 1085static void __migrate_swap_task(struct task_struct *p, int cpu) 1086{ 1087 if (p->on_rq) { 1088 struct rq *src_rq, *dst_rq; 1089 1090 src_rq = task_rq(p); 1091 dst_rq = cpu_rq(cpu); 1092 1093 deactivate_task(src_rq, p, 0); 1094 set_task_cpu(p, cpu); 1095 activate_task(dst_rq, p, 0); 1096 check_preempt_curr(dst_rq, p, 0); 1097 } else { 1098 /* 1099 * Task isn't running anymore; make it appear like we migrated 1100 * it before it went to sleep. This means on wakeup we make the 1101 * previous cpu our targer instead of where it really is. 1102 */ 1103 p->wake_cpu = cpu; 1104 } 1105} 1106 1107struct migration_swap_arg { 1108 struct task_struct *src_task, *dst_task; 1109 int src_cpu, dst_cpu; 1110}; 1111 1112static int migrate_swap_stop(void *data) 1113{ 1114 struct migration_swap_arg *arg = data; 1115 struct rq *src_rq, *dst_rq; 1116 int ret = -EAGAIN; 1117 1118 src_rq = cpu_rq(arg->src_cpu); 1119 dst_rq = cpu_rq(arg->dst_cpu); 1120 1121 double_raw_lock(&arg->src_task->pi_lock, 1122 &arg->dst_task->pi_lock); 1123 double_rq_lock(src_rq, dst_rq); 1124 if (task_cpu(arg->dst_task) != arg->dst_cpu) 1125 goto unlock; 1126 1127 if (task_cpu(arg->src_task) != arg->src_cpu) 1128 goto unlock; 1129 1130 if (!cpumask_test_cpu(arg->dst_cpu, tsk_cpus_allowed(arg->src_task))) 1131 goto unlock; 1132 1133 if (!cpumask_test_cpu(arg->src_cpu, tsk_cpus_allowed(arg->dst_task))) 1134 goto unlock; 1135 1136 __migrate_swap_task(arg->src_task, arg->dst_cpu); 1137 __migrate_swap_task(arg->dst_task, arg->src_cpu); 1138 1139 ret = 0; 1140 1141unlock: 1142 double_rq_unlock(src_rq, dst_rq); 1143 raw_spin_unlock(&arg->dst_task->pi_lock); 1144 raw_spin_unlock(&arg->src_task->pi_lock); 1145 1146 return ret; 1147} 1148 1149/* 1150 * Cross migrate two tasks 1151 */ 1152int migrate_swap(struct task_struct *cur, struct task_struct *p) 1153{ 1154 struct migration_swap_arg arg; 1155 int ret = -EINVAL; 1156 1157 arg = (struct migration_swap_arg){ 1158 .src_task = cur, 1159 .src_cpu = task_cpu(cur), 1160 .dst_task = p, 1161 .dst_cpu = task_cpu(p), 1162 }; 1163 1164 if (arg.src_cpu == arg.dst_cpu) 1165 goto out; 1166 1167 /* 1168 * These three tests are all lockless; this is OK since all of them 1169 * will be re-checked with proper locks held further down the line. 1170 */ 1171 if (!cpu_active(arg.src_cpu) || !cpu_active(arg.dst_cpu)) 1172 goto out; 1173 1174 if (!cpumask_test_cpu(arg.dst_cpu, tsk_cpus_allowed(arg.src_task))) 1175 goto out; 1176 1177 if (!cpumask_test_cpu(arg.src_cpu, tsk_cpus_allowed(arg.dst_task))) 1178 goto out; 1179 1180 trace_sched_swap_numa(cur, arg.src_cpu, p, arg.dst_cpu); 1181 ret = stop_two_cpus(arg.dst_cpu, arg.src_cpu, migrate_swap_stop, &arg); 1182 1183out: 1184 return ret; 1185} 1186 1187struct migration_arg { 1188 struct task_struct *task; 1189 int dest_cpu; 1190}; 1191 1192static int migration_cpu_stop(void *data); 1193 1194/* 1195 * wait_task_inactive - wait for a thread to unschedule. 1196 * 1197 * If @match_state is nonzero, it's the @p->state value just checked and 1198 * not expected to change. If it changes, i.e. @p might have woken up, 1199 * then return zero. When we succeed in waiting for @p to be off its CPU, 1200 * we return a positive number (its total switch count). If a second call 1201 * a short while later returns the same number, the caller can be sure that 1202 * @p has remained unscheduled the whole time. 1203 * 1204 * The caller must ensure that the task *will* unschedule sometime soon, 1205 * else this function might spin for a *long* time. This function can't 1206 * be called with interrupts off, or it may introduce deadlock with 1207 * smp_call_function() if an IPI is sent by the same process we are 1208 * waiting to become inactive. 1209 */ 1210unsigned long wait_task_inactive(struct task_struct *p, long match_state) 1211{ 1212 unsigned long flags; 1213 int running, on_rq; 1214 unsigned long ncsw; 1215 struct rq *rq; 1216 1217 for (;;) { 1218 /* 1219 * We do the initial early heuristics without holding 1220 * any task-queue locks at all. We'll only try to get 1221 * the runqueue lock when things look like they will 1222 * work out! 1223 */ 1224 rq = task_rq(p); 1225 1226 /* 1227 * If the task is actively running on another CPU 1228 * still, just relax and busy-wait without holding 1229 * any locks. 1230 * 1231 * NOTE! Since we don't hold any locks, it's not 1232 * even sure that "rq" stays as the right runqueue! 1233 * But we don't care, since "task_running()" will 1234 * return false if the runqueue has changed and p 1235 * is actually now running somewhere else! 1236 */ 1237 while (task_running(rq, p)) { 1238 if (match_state && unlikely(p->state != match_state)) 1239 return 0; 1240 cpu_relax(); 1241 } 1242 1243 /* 1244 * Ok, time to look more closely! We need the rq 1245 * lock now, to be *sure*. If we're wrong, we'll 1246 * just go back and repeat. 1247 */ 1248 rq = task_rq_lock(p, &flags); 1249 trace_sched_wait_task(p); 1250 running = task_running(rq, p); 1251 on_rq = p->on_rq; 1252 ncsw = 0; 1253 if (!match_state || p->state == match_state) 1254 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */ 1255 task_rq_unlock(rq, p, &flags); 1256 1257 /* 1258 * If it changed from the expected state, bail out now. 1259 */ 1260 if (unlikely(!ncsw)) 1261 break; 1262 1263 /* 1264 * Was it really running after all now that we 1265 * checked with the proper locks actually held? 1266 * 1267 * Oops. Go back and try again.. 1268 */ 1269 if (unlikely(running)) { 1270 cpu_relax(); 1271 continue; 1272 } 1273 1274 /* 1275 * It's not enough that it's not actively running, 1276 * it must be off the runqueue _entirely_, and not 1277 * preempted! 1278 * 1279 * So if it was still runnable (but just not actively 1280 * running right now), it's preempted, and we should 1281 * yield - it could be a while. 1282 */ 1283 if (unlikely(on_rq)) { 1284 ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ); 1285 1286 set_current_state(TASK_UNINTERRUPTIBLE); 1287 schedule_hrtimeout(&to, HRTIMER_MODE_REL); 1288 continue; 1289 } 1290 1291 /* 1292 * Ahh, all good. It wasn't running, and it wasn't 1293 * runnable, which means that it will never become 1294 * running in the future either. We're all done! 1295 */ 1296 break; 1297 } 1298 1299 return ncsw; 1300} 1301 1302/*** 1303 * kick_process - kick a running thread to enter/exit the kernel 1304 * @p: the to-be-kicked thread 1305 * 1306 * Cause a process which is running on another CPU to enter 1307 * kernel-mode, without any delay. (to get signals handled.) 1308 * 1309 * NOTE: this function doesn't have to take the runqueue lock, 1310 * because all it wants to ensure is that the remote task enters 1311 * the kernel. If the IPI races and the task has been migrated 1312 * to another CPU then no harm is done and the purpose has been 1313 * achieved as well. 1314 */ 1315void kick_process(struct task_struct *p) 1316{ 1317 int cpu; 1318 1319 preempt_disable(); 1320 cpu = task_cpu(p); 1321 if ((cpu != smp_processor_id()) && task_curr(p)) 1322 smp_send_reschedule(cpu); 1323 preempt_enable(); 1324} 1325EXPORT_SYMBOL_GPL(kick_process); 1326#endif /* CONFIG_SMP */ 1327 1328#ifdef CONFIG_SMP 1329/* 1330 * ->cpus_allowed is protected by both rq->lock and p->pi_lock 1331 */ 1332static int select_fallback_rq(int cpu, struct task_struct *p) 1333{ 1334 int nid = cpu_to_node(cpu); 1335 const struct cpumask *nodemask = NULL; 1336 enum { cpuset, possible, fail } state = cpuset; 1337 int dest_cpu; 1338 1339 /* 1340 * If the node that the cpu is on has been offlined, cpu_to_node() 1341 * will return -1. There is no cpu on the node, and we should 1342 * select the cpu on the other node. 1343 */ 1344 if (nid != -1) { 1345 nodemask = cpumask_of_node(nid); 1346 1347 /* Look for allowed, online CPU in same node. */ 1348 for_each_cpu(dest_cpu, nodemask) { 1349 if (!cpu_online(dest_cpu)) 1350 continue; 1351 if (!cpu_active(dest_cpu)) 1352 continue; 1353 if (cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p))) 1354 return dest_cpu; 1355 } 1356 } 1357 1358 for (;;) { 1359 /* Any allowed, online CPU? */ 1360 for_each_cpu(dest_cpu, tsk_cpus_allowed(p)) { 1361 if (!cpu_online(dest_cpu)) 1362 continue; 1363 if (!cpu_active(dest_cpu)) 1364 continue; 1365 goto out; 1366 } 1367 1368 switch (state) { 1369 case cpuset: 1370 /* No more Mr. Nice Guy. */ 1371 cpuset_cpus_allowed_fallback(p); 1372 state = possible; 1373 break; 1374 1375 case possible: 1376 do_set_cpus_allowed(p, cpu_possible_mask); 1377 state = fail; 1378 break; 1379 1380 case fail: 1381 BUG(); 1382 break; 1383 } 1384 } 1385 1386out: 1387 if (state != cpuset) { 1388 /* 1389 * Don't tell them about moving exiting tasks or 1390 * kernel threads (both mm NULL), since they never 1391 * leave kernel. 1392 */ 1393 if (p->mm && printk_ratelimit()) { 1394 printk_deferred("process %d (%s) no longer affine to cpu%d\n", 1395 task_pid_nr(p), p->comm, cpu); 1396 } 1397 } 1398 1399 return dest_cpu; 1400} 1401 1402/* 1403 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable. 1404 */ 1405static inline 1406int select_task_rq(struct task_struct *p, int cpu, int sd_flags, int wake_flags) 1407{ 1408 cpu = p->sched_class->select_task_rq(p, cpu, sd_flags, wake_flags); 1409 1410 /* 1411 * In order not to call set_task_cpu() on a blocking task we need 1412 * to rely on ttwu() to place the task on a valid ->cpus_allowed 1413 * cpu. 1414 * 1415 * Since this is common to all placement strategies, this lives here. 1416 * 1417 * [ this allows ->select_task() to simply return task_cpu(p) and 1418 * not worry about this generic constraint ] 1419 */ 1420 if (unlikely(!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)) || 1421 !cpu_online(cpu))) 1422 cpu = select_fallback_rq(task_cpu(p), p); 1423 1424 return cpu; 1425} 1426 1427static void update_avg(u64 *avg, u64 sample) 1428{ 1429 s64 diff = sample - *avg; 1430 *avg += diff >> 3; 1431} 1432#endif 1433 1434static void 1435ttwu_stat(struct task_struct *p, int cpu, int wake_flags) 1436{ 1437#ifdef CONFIG_SCHEDSTATS 1438 struct rq *rq = this_rq(); 1439 1440#ifdef CONFIG_SMP 1441 int this_cpu = smp_processor_id(); 1442 1443 if (cpu == this_cpu) { 1444 schedstat_inc(rq, ttwu_local); 1445 schedstat_inc(p, se.statistics.nr_wakeups_local); 1446 } else { 1447 struct sched_domain *sd; 1448 1449 schedstat_inc(p, se.statistics.nr_wakeups_remote); 1450 rcu_read_lock(); 1451 for_each_domain(this_cpu, sd) { 1452 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) { 1453 schedstat_inc(sd, ttwu_wake_remote); 1454 break; 1455 } 1456 } 1457 rcu_read_unlock(); 1458 } 1459 1460 if (wake_flags & WF_MIGRATED) 1461 schedstat_inc(p, se.statistics.nr_wakeups_migrate); 1462 1463#endif /* CONFIG_SMP */ 1464 1465 schedstat_inc(rq, ttwu_count); 1466 schedstat_inc(p, se.statistics.nr_wakeups); 1467 1468 if (wake_flags & WF_SYNC) 1469 schedstat_inc(p, se.statistics.nr_wakeups_sync); 1470 1471#endif /* CONFIG_SCHEDSTATS */ 1472} 1473 1474static void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags) 1475{ 1476 activate_task(rq, p, en_flags); 1477 p->on_rq = 1; 1478 1479 /* if a worker is waking up, notify workqueue */ 1480 if (p->flags & PF_WQ_WORKER) 1481 wq_worker_waking_up(p, cpu_of(rq)); 1482} 1483 1484/* 1485 * Mark the task runnable and perform wakeup-preemption. 1486 */ 1487static void 1488ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags) 1489{ 1490 check_preempt_curr(rq, p, wake_flags); 1491 trace_sched_wakeup(p, true); 1492 1493 p->state = TASK_RUNNING; 1494#ifdef CONFIG_SMP 1495 if (p->sched_class->task_woken) 1496 p->sched_class->task_woken(rq, p); 1497 1498 if (rq->idle_stamp) { 1499 u64 delta = rq_clock(rq) - rq->idle_stamp; 1500 u64 max = 2*rq->max_idle_balance_cost; 1501 1502 update_avg(&rq->avg_idle, delta); 1503 1504 if (rq->avg_idle > max) 1505 rq->avg_idle = max; 1506 1507 rq->idle_stamp = 0; 1508 } 1509#endif 1510} 1511 1512static void 1513ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags) 1514{ 1515#ifdef CONFIG_SMP 1516 if (p->sched_contributes_to_load) 1517 rq->nr_uninterruptible--; 1518#endif 1519 1520 ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_WAKING); 1521 ttwu_do_wakeup(rq, p, wake_flags); 1522} 1523 1524/* 1525 * Called in case the task @p isn't fully descheduled from its runqueue, 1526 * in this case we must do a remote wakeup. Its a 'light' wakeup though, 1527 * since all we need to do is flip p->state to TASK_RUNNING, since 1528 * the task is still ->on_rq. 1529 */ 1530static int ttwu_remote(struct task_struct *p, int wake_flags) 1531{ 1532 struct rq *rq; 1533 int ret = 0; 1534 1535 rq = __task_rq_lock(p); 1536 if (p->on_rq) { 1537 /* check_preempt_curr() may use rq clock */ 1538 update_rq_clock(rq); 1539 ttwu_do_wakeup(rq, p, wake_flags); 1540 ret = 1; 1541 } 1542 __task_rq_unlock(rq); 1543 1544 return ret; 1545} 1546 1547#ifdef CONFIG_SMP 1548void sched_ttwu_pending(void) 1549{ 1550 struct rq *rq = this_rq(); 1551 struct llist_node *llist = llist_del_all(&rq->wake_list); 1552 struct task_struct *p; 1553 unsigned long flags; 1554 1555 if (!llist) 1556 return; 1557 1558 raw_spin_lock_irqsave(&rq->lock, flags); 1559 1560 while (llist) { 1561 p = llist_entry(llist, struct task_struct, wake_entry); 1562 llist = llist_next(llist); 1563 ttwu_do_activate(rq, p, 0); 1564 } 1565 1566 raw_spin_unlock_irqrestore(&rq->lock, flags); 1567} 1568 1569void scheduler_ipi(void) 1570{ 1571 /* 1572 * Fold TIF_NEED_RESCHED into the preempt_count; anybody setting 1573 * TIF_NEED_RESCHED remotely (for the first time) will also send 1574 * this IPI. 1575 */ 1576 preempt_fold_need_resched(); 1577 1578 if (llist_empty(&this_rq()->wake_list) && !got_nohz_idle_kick()) 1579 return; 1580 1581 /* 1582 * Not all reschedule IPI handlers call irq_enter/irq_exit, since 1583 * traditionally all their work was done from the interrupt return 1584 * path. Now that we actually do some work, we need to make sure 1585 * we do call them. 1586 * 1587 * Some archs already do call them, luckily irq_enter/exit nest 1588 * properly. 1589 * 1590 * Arguably we should visit all archs and update all handlers, 1591 * however a fair share of IPIs are still resched only so this would 1592 * somewhat pessimize the simple resched case. 1593 */ 1594 irq_enter(); 1595 sched_ttwu_pending(); 1596 1597 /* 1598 * Check if someone kicked us for doing the nohz idle load balance. 1599 */ 1600 if (unlikely(got_nohz_idle_kick())) { 1601 this_rq()->idle_balance = 1; 1602 raise_softirq_irqoff(SCHED_SOFTIRQ); 1603 } 1604 irq_exit(); 1605} 1606 1607static void ttwu_queue_remote(struct task_struct *p, int cpu) 1608{ 1609 struct rq *rq = cpu_rq(cpu); 1610 1611 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list)) { 1612 if (!set_nr_if_polling(rq->idle)) 1613 smp_send_reschedule(cpu); 1614 else 1615 trace_sched_wake_idle_without_ipi(cpu); 1616 } 1617} 1618 1619bool cpus_share_cache(int this_cpu, int that_cpu) 1620{ 1621 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu); 1622} 1623#endif /* CONFIG_SMP */ 1624 1625static void ttwu_queue(struct task_struct *p, int cpu) 1626{ 1627 struct rq *rq = cpu_rq(cpu); 1628 1629#if defined(CONFIG_SMP) 1630 if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) { 1631 sched_clock_cpu(cpu); /* sync clocks x-cpu */ 1632 ttwu_queue_remote(p, cpu); 1633 return; 1634 } 1635#endif 1636 1637 raw_spin_lock(&rq->lock); 1638 ttwu_do_activate(rq, p, 0); 1639 raw_spin_unlock(&rq->lock); 1640} 1641 1642/** 1643 * try_to_wake_up - wake up a thread 1644 * @p: the thread to be awakened 1645 * @state: the mask of task states that can be woken 1646 * @wake_flags: wake modifier flags (WF_*) 1647 * 1648 * Put it on the run-queue if it's not already there. The "current" 1649 * thread is always on the run-queue (except when the actual 1650 * re-schedule is in progress), and as such you're allowed to do 1651 * the simpler "current->state = TASK_RUNNING" to mark yourself 1652 * runnable without the overhead of this. 1653 * 1654 * Return: %true if @p was woken up, %false if it was already running. 1655 * or @state didn't match @p's state. 1656 */ 1657static int 1658try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags) 1659{ 1660 unsigned long flags; 1661 int cpu, success = 0; 1662 1663 /* 1664 * If we are going to wake up a thread waiting for CONDITION we 1665 * need to ensure that CONDITION=1 done by the caller can not be 1666 * reordered with p->state check below. This pairs with mb() in 1667 * set_current_state() the waiting thread does. 1668 */ 1669 smp_mb__before_spinlock(); 1670 raw_spin_lock_irqsave(&p->pi_lock, flags); 1671 if (!(p->state & state)) 1672 goto out; 1673 1674 success = 1; /* we're going to change ->state */ 1675 cpu = task_cpu(p); 1676 1677 if (p->on_rq && ttwu_remote(p, wake_flags)) 1678 goto stat; 1679 1680#ifdef CONFIG_SMP 1681 /* 1682 * If the owning (remote) cpu is still in the middle of schedule() with 1683 * this task as prev, wait until its done referencing the task. 1684 */ 1685 while (p->on_cpu) 1686 cpu_relax(); 1687 /* 1688 * Pairs with the smp_wmb() in finish_lock_switch(). 1689 */ 1690 smp_rmb(); 1691 1692 p->sched_contributes_to_load = !!task_contributes_to_load(p); 1693 p->state = TASK_WAKING; 1694 1695 if (p->sched_class->task_waking) 1696 p->sched_class->task_waking(p); 1697 1698 cpu = select_task_rq(p, p->wake_cpu, SD_BALANCE_WAKE, wake_flags); 1699 if (task_cpu(p) != cpu) { 1700 wake_flags |= WF_MIGRATED; 1701 set_task_cpu(p, cpu); 1702 } 1703#endif /* CONFIG_SMP */ 1704 1705 ttwu_queue(p, cpu); 1706stat: 1707 ttwu_stat(p, cpu, wake_flags); 1708out: 1709 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 1710 1711 return success; 1712} 1713 1714/** 1715 * try_to_wake_up_local - try to wake up a local task with rq lock held 1716 * @p: the thread to be awakened 1717 * 1718 * Put @p on the run-queue if it's not already there. The caller must 1719 * ensure that this_rq() is locked, @p is bound to this_rq() and not 1720 * the current task. 1721 */ 1722static void try_to_wake_up_local(struct task_struct *p) 1723{ 1724 struct rq *rq = task_rq(p); 1725 1726 if (WARN_ON_ONCE(rq != this_rq()) || 1727 WARN_ON_ONCE(p == current)) 1728 return; 1729 1730 lockdep_assert_held(&rq->lock); 1731 1732 if (!raw_spin_trylock(&p->pi_lock)) { 1733 raw_spin_unlock(&rq->lock); 1734 raw_spin_lock(&p->pi_lock); 1735 raw_spin_lock(&rq->lock); 1736 } 1737 1738 if (!(p->state & TASK_NORMAL)) 1739 goto out; 1740 1741 if (!p->on_rq) 1742 ttwu_activate(rq, p, ENQUEUE_WAKEUP); 1743 1744 ttwu_do_wakeup(rq, p, 0); 1745 ttwu_stat(p, smp_processor_id(), 0); 1746out: 1747 raw_spin_unlock(&p->pi_lock); 1748} 1749 1750/** 1751 * wake_up_process - Wake up a specific process 1752 * @p: The process to be woken up. 1753 * 1754 * Attempt to wake up the nominated process and move it to the set of runnable 1755 * processes. 1756 * 1757 * Return: 1 if the process was woken up, 0 if it was already running. 1758 * 1759 * It may be assumed that this function implies a write memory barrier before 1760 * changing the task state if and only if any tasks are woken up. 1761 */ 1762int wake_up_process(struct task_struct *p) 1763{ 1764 WARN_ON(task_is_stopped_or_traced(p)); 1765 return try_to_wake_up(p, TASK_NORMAL, 0); 1766} 1767EXPORT_SYMBOL(wake_up_process); 1768 1769int wake_up_state(struct task_struct *p, unsigned int state) 1770{ 1771 return try_to_wake_up(p, state, 0); 1772} 1773 1774/* 1775 * Perform scheduler related setup for a newly forked process p. 1776 * p is forked by current. 1777 * 1778 * __sched_fork() is basic setup used by init_idle() too: 1779 */ 1780static void __sched_fork(unsigned long clone_flags, struct task_struct *p) 1781{ 1782 p->on_rq = 0; 1783 1784 p->se.on_rq = 0; 1785 p->se.exec_start = 0; 1786 p->se.sum_exec_runtime = 0; 1787 p->se.prev_sum_exec_runtime = 0; 1788 p->se.nr_migrations = 0; 1789 p->se.vruntime = 0; 1790 INIT_LIST_HEAD(&p->se.group_node); 1791 1792#ifdef CONFIG_SCHEDSTATS 1793 memset(&p->se.statistics, 0, sizeof(p->se.statistics)); 1794#endif 1795 1796 RB_CLEAR_NODE(&p->dl.rb_node); 1797 hrtimer_init(&p->dl.dl_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); 1798 p->dl.dl_runtime = p->dl.runtime = 0; 1799 p->dl.dl_deadline = p->dl.deadline = 0; 1800 p->dl.dl_period = 0; 1801 p->dl.flags = 0; 1802 1803 INIT_LIST_HEAD(&p->rt.run_list); 1804 1805#ifdef CONFIG_PREEMPT_NOTIFIERS 1806 INIT_HLIST_HEAD(&p->preempt_notifiers); 1807#endif 1808 1809#ifdef CONFIG_NUMA_BALANCING 1810 if (p->mm && atomic_read(&p->mm->mm_users) == 1) { 1811 p->mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay); 1812 p->mm->numa_scan_seq = 0; 1813 } 1814 1815 if (clone_flags & CLONE_VM) 1816 p->numa_preferred_nid = current->numa_preferred_nid; 1817 else 1818 p->numa_preferred_nid = -1; 1819 1820 p->node_stamp = 0ULL; 1821 p->numa_scan_seq = p->mm ? p->mm->numa_scan_seq : 0; 1822 p->numa_scan_period = sysctl_numa_balancing_scan_delay; 1823 p->numa_work.next = &p->numa_work; 1824 p->numa_faults_memory = NULL; 1825 p->numa_faults_buffer_memory = NULL; 1826 p->last_task_numa_placement = 0; 1827 p->last_sum_exec_runtime = 0; 1828 1829 INIT_LIST_HEAD(&p->numa_entry); 1830 p->numa_group = NULL; 1831#endif /* CONFIG_NUMA_BALANCING */ 1832} 1833 1834#ifdef CONFIG_NUMA_BALANCING 1835#ifdef CONFIG_SCHED_DEBUG 1836void set_numabalancing_state(bool enabled) 1837{ 1838 if (enabled) 1839 sched_feat_set("NUMA"); 1840 else 1841 sched_feat_set("NO_NUMA"); 1842} 1843#else 1844__read_mostly bool numabalancing_enabled; 1845 1846void set_numabalancing_state(bool enabled) 1847{ 1848 numabalancing_enabled = enabled; 1849} 1850#endif /* CONFIG_SCHED_DEBUG */ 1851 1852#ifdef CONFIG_PROC_SYSCTL 1853int sysctl_numa_balancing(struct ctl_table *table, int write, 1854 void __user *buffer, size_t *lenp, loff_t *ppos) 1855{ 1856 struct ctl_table t; 1857 int err; 1858 int state = numabalancing_enabled; 1859 1860 if (write && !capable(CAP_SYS_ADMIN)) 1861 return -EPERM; 1862 1863 t = *table; 1864 t.data = &state; 1865 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos); 1866 if (err < 0) 1867 return err; 1868 if (write) 1869 set_numabalancing_state(state); 1870 return err; 1871} 1872#endif 1873#endif 1874 1875/* 1876 * fork()/clone()-time setup: 1877 */ 1878int sched_fork(unsigned long clone_flags, struct task_struct *p) 1879{ 1880 unsigned long flags; 1881 int cpu = get_cpu(); 1882 1883 __sched_fork(clone_flags, p); 1884 /* 1885 * We mark the process as running here. This guarantees that 1886 * nobody will actually run it, and a signal or other external 1887 * event cannot wake it up and insert it on the runqueue either. 1888 */ 1889 p->state = TASK_RUNNING; 1890 1891 /* 1892 * Make sure we do not leak PI boosting priority to the child. 1893 */ 1894 p->prio = current->normal_prio; 1895 1896 /* 1897 * Revert to default priority/policy on fork if requested. 1898 */ 1899 if (unlikely(p->sched_reset_on_fork)) { 1900 if (task_has_dl_policy(p) || task_has_rt_policy(p)) { 1901 p->policy = SCHED_NORMAL; 1902 p->static_prio = NICE_TO_PRIO(0); 1903 p->rt_priority = 0; 1904 } else if (PRIO_TO_NICE(p->static_prio) < 0) 1905 p->static_prio = NICE_TO_PRIO(0); 1906 1907 p->prio = p->normal_prio = __normal_prio(p); 1908 set_load_weight(p); 1909 1910 /* 1911 * We don't need the reset flag anymore after the fork. It has 1912 * fulfilled its duty: 1913 */ 1914 p->sched_reset_on_fork = 0; 1915 } 1916 1917 if (dl_prio(p->prio)) { 1918 put_cpu(); 1919 return -EAGAIN; 1920 } else if (rt_prio(p->prio)) { 1921 p->sched_class = &rt_sched_class; 1922 } else { 1923 p->sched_class = &fair_sched_class; 1924 } 1925 1926 if (p->sched_class->task_fork) 1927 p->sched_class->task_fork(p); 1928 1929 /* 1930 * The child is not yet in the pid-hash so no cgroup attach races, 1931 * and the cgroup is pinned to this child due to cgroup_fork() 1932 * is ran before sched_fork(). 1933 * 1934 * Silence PROVE_RCU. 1935 */ 1936 raw_spin_lock_irqsave(&p->pi_lock, flags); 1937 set_task_cpu(p, cpu); 1938 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 1939 1940#if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT) 1941 if (likely(sched_info_on())) 1942 memset(&p->sched_info, 0, sizeof(p->sched_info)); 1943#endif 1944#if defined(CONFIG_SMP) 1945 p->on_cpu = 0; 1946#endif 1947 init_task_preempt_count(p); 1948#ifdef CONFIG_SMP 1949 plist_node_init(&p->pushable_tasks, MAX_PRIO); 1950 RB_CLEAR_NODE(&p->pushable_dl_tasks); 1951#endif 1952 1953 put_cpu(); 1954 return 0; 1955} 1956 1957unsigned long to_ratio(u64 period, u64 runtime) 1958{ 1959 if (runtime == RUNTIME_INF) 1960 return 1ULL << 20; 1961 1962 /* 1963 * Doing this here saves a lot of checks in all 1964 * the calling paths, and returning zero seems 1965 * safe for them anyway. 1966 */ 1967 if (period == 0) 1968 return 0; 1969 1970 return div64_u64(runtime << 20, period); 1971} 1972 1973#ifdef CONFIG_SMP 1974inline struct dl_bw *dl_bw_of(int i) 1975{ 1976 return &cpu_rq(i)->rd->dl_bw; 1977} 1978 1979static inline int dl_bw_cpus(int i) 1980{ 1981 struct root_domain *rd = cpu_rq(i)->rd; 1982 int cpus = 0; 1983 1984 for_each_cpu_and(i, rd->span, cpu_active_mask) 1985 cpus++; 1986 1987 return cpus; 1988} 1989#else 1990inline struct dl_bw *dl_bw_of(int i) 1991{ 1992 return &cpu_rq(i)->dl.dl_bw; 1993} 1994 1995static inline int dl_bw_cpus(int i) 1996{ 1997 return 1; 1998} 1999#endif 2000 2001static inline 2002void __dl_clear(struct dl_bw *dl_b, u64 tsk_bw) 2003{ 2004 dl_b->total_bw -= tsk_bw; 2005} 2006 2007static inline 2008void __dl_add(struct dl_bw *dl_b, u64 tsk_bw) 2009{ 2010 dl_b->total_bw += tsk_bw; 2011} 2012 2013static inline 2014bool __dl_overflow(struct dl_bw *dl_b, int cpus, u64 old_bw, u64 new_bw) 2015{ 2016 return dl_b->bw != -1 && 2017 dl_b->bw * cpus < dl_b->total_bw - old_bw + new_bw; 2018} 2019 2020/* 2021 * We must be sure that accepting a new task (or allowing changing the 2022 * parameters of an existing one) is consistent with the bandwidth 2023 * constraints. If yes, this function also accordingly updates the currently 2024 * allocated bandwidth to reflect the new situation. 2025 * 2026 * This function is called while holding p's rq->lock. 2027 */ 2028static int dl_overflow(struct task_struct *p, int policy, 2029 const struct sched_attr *attr) 2030{ 2031 2032 struct dl_bw *dl_b = dl_bw_of(task_cpu(p)); 2033 u64 period = attr->sched_period ?: attr->sched_deadline; 2034 u64 runtime = attr->sched_runtime; 2035 u64 new_bw = dl_policy(policy) ? to_ratio(period, runtime) : 0; 2036 int cpus, err = -1; 2037 2038 if (new_bw == p->dl.dl_bw) 2039 return 0; 2040 2041 /* 2042 * Either if a task, enters, leave, or stays -deadline but changes 2043 * its parameters, we may need to update accordingly the total 2044 * allocated bandwidth of the container. 2045 */ 2046 raw_spin_lock(&dl_b->lock); 2047 cpus = dl_bw_cpus(task_cpu(p)); 2048 if (dl_policy(policy) && !task_has_dl_policy(p) && 2049 !__dl_overflow(dl_b, cpus, 0, new_bw)) { 2050 __dl_add(dl_b, new_bw); 2051 err = 0; 2052 } else if (dl_policy(policy) && task_has_dl_policy(p) && 2053 !__dl_overflow(dl_b, cpus, p->dl.dl_bw, new_bw)) { 2054 __dl_clear(dl_b, p->dl.dl_bw); 2055 __dl_add(dl_b, new_bw); 2056 err = 0; 2057 } else if (!dl_policy(policy) && task_has_dl_policy(p)) { 2058 __dl_clear(dl_b, p->dl.dl_bw); 2059 err = 0; 2060 } 2061 raw_spin_unlock(&dl_b->lock); 2062 2063 return err; 2064} 2065 2066extern void init_dl_bw(struct dl_bw *dl_b); 2067 2068/* 2069 * wake_up_new_task - wake up a newly created task for the first time. 2070 * 2071 * This function will do some initial scheduler statistics housekeeping 2072 * that must be done for every newly created context, then puts the task 2073 * on the runqueue and wakes it. 2074 */ 2075void wake_up_new_task(struct task_struct *p) 2076{ 2077 unsigned long flags; 2078 struct rq *rq; 2079 2080 raw_spin_lock_irqsave(&p->pi_lock, flags); 2081#ifdef CONFIG_SMP 2082 /* 2083 * Fork balancing, do it here and not earlier because: 2084 * - cpus_allowed can change in the fork path 2085 * - any previously selected cpu might disappear through hotplug 2086 */ 2087 set_task_cpu(p, select_task_rq(p, task_cpu(p), SD_BALANCE_FORK, 0)); 2088#endif 2089 2090 /* Initialize new task's runnable average */ 2091 init_task_runnable_average(p); 2092 rq = __task_rq_lock(p); 2093 activate_task(rq, p, 0); 2094 p->on_rq = 1; 2095 trace_sched_wakeup_new(p, true); 2096 check_preempt_curr(rq, p, WF_FORK); 2097#ifdef CONFIG_SMP 2098 if (p->sched_class->task_woken) 2099 p->sched_class->task_woken(rq, p); 2100#endif 2101 task_rq_unlock(rq, p, &flags); 2102} 2103 2104#ifdef CONFIG_PREEMPT_NOTIFIERS 2105 2106/** 2107 * preempt_notifier_register - tell me when current is being preempted & rescheduled 2108 * @notifier: notifier struct to register 2109 */ 2110void preempt_notifier_register(struct preempt_notifier *notifier) 2111{ 2112 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers); 2113} 2114EXPORT_SYMBOL_GPL(preempt_notifier_register); 2115 2116/** 2117 * preempt_notifier_unregister - no longer interested in preemption notifications 2118 * @notifier: notifier struct to unregister 2119 * 2120 * This is safe to call from within a preemption notifier. 2121 */ 2122void preempt_notifier_unregister(struct preempt_notifier *notifier) 2123{ 2124 hlist_del(¬ifier->link); 2125} 2126EXPORT_SYMBOL_GPL(preempt_notifier_unregister); 2127 2128static void fire_sched_in_preempt_notifiers(struct task_struct *curr) 2129{ 2130 struct preempt_notifier *notifier; 2131 2132 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link) 2133 notifier->ops->sched_in(notifier, raw_smp_processor_id()); 2134} 2135 2136static void 2137fire_sched_out_preempt_notifiers(struct task_struct *curr, 2138 struct task_struct *next) 2139{ 2140 struct preempt_notifier *notifier; 2141 2142 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link) 2143 notifier->ops->sched_out(notifier, next); 2144} 2145 2146#else /* !CONFIG_PREEMPT_NOTIFIERS */ 2147 2148static void fire_sched_in_preempt_notifiers(struct task_struct *curr) 2149{ 2150} 2151 2152static void 2153fire_sched_out_preempt_notifiers(struct task_struct *curr, 2154 struct task_struct *next) 2155{ 2156} 2157 2158#endif /* CONFIG_PREEMPT_NOTIFIERS */ 2159 2160/** 2161 * prepare_task_switch - prepare to switch tasks 2162 * @rq: the runqueue preparing to switch 2163 * @prev: the current task that is being switched out 2164 * @next: the task we are going to switch to. 2165 * 2166 * This is called with the rq lock held and interrupts off. It must 2167 * be paired with a subsequent finish_task_switch after the context 2168 * switch. 2169 * 2170 * prepare_task_switch sets up locking and calls architecture specific 2171 * hooks. 2172 */ 2173static inline void 2174prepare_task_switch(struct rq *rq, struct task_struct *prev, 2175 struct task_struct *next) 2176{ 2177 trace_sched_switch(prev, next); 2178 sched_info_switch(rq, prev, next); 2179 perf_event_task_sched_out(prev, next); 2180 fire_sched_out_preempt_notifiers(prev, next); 2181 prepare_lock_switch(rq, next); 2182 prepare_arch_switch(next); 2183} 2184 2185/** 2186 * finish_task_switch - clean up after a task-switch 2187 * @rq: runqueue associated with task-switch 2188 * @prev: the thread we just switched away from. 2189 * 2190 * finish_task_switch must be called after the context switch, paired 2191 * with a prepare_task_switch call before the context switch. 2192 * finish_task_switch will reconcile locking set up by prepare_task_switch, 2193 * and do any other architecture-specific cleanup actions. 2194 * 2195 * Note that we may have delayed dropping an mm in context_switch(). If 2196 * so, we finish that here outside of the runqueue lock. (Doing it 2197 * with the lock held can cause deadlocks; see schedule() for 2198 * details.) 2199 */ 2200static void finish_task_switch(struct rq *rq, struct task_struct *prev) 2201 __releases(rq->lock) 2202{ 2203 struct mm_struct *mm = rq->prev_mm; 2204 long prev_state; 2205 2206 rq->prev_mm = NULL; 2207 2208 /* 2209 * A task struct has one reference for the use as "current". 2210 * If a task dies, then it sets TASK_DEAD in tsk->state and calls 2211 * schedule one last time. The schedule call will never return, and 2212 * the scheduled task must drop that reference. 2213 * The test for TASK_DEAD must occur while the runqueue locks are 2214 * still held, otherwise prev could be scheduled on another cpu, die 2215 * there before we look at prev->state, and then the reference would 2216 * be dropped twice. 2217 * Manfred Spraul <manfred@colorfullife.com> 2218 */ 2219 prev_state = prev->state; 2220 vtime_task_switch(prev); 2221 finish_arch_switch(prev); 2222 perf_event_task_sched_in(prev, current); 2223 finish_lock_switch(rq, prev); 2224 finish_arch_post_lock_switch(); 2225 2226 fire_sched_in_preempt_notifiers(current); 2227 if (mm) 2228 mmdrop(mm); 2229 if (unlikely(prev_state == TASK_DEAD)) { 2230 if (prev->sched_class->task_dead) 2231 prev->sched_class->task_dead(prev); 2232 2233 /* 2234 * Remove function-return probe instances associated with this 2235 * task and put them back on the free list. 2236 */ 2237 kprobe_flush_task(prev); 2238 put_task_struct(prev); 2239 } 2240 2241 tick_nohz_task_switch(current); 2242} 2243 2244#ifdef CONFIG_SMP 2245 2246/* rq->lock is NOT held, but preemption is disabled */ 2247static inline void post_schedule(struct rq *rq) 2248{ 2249 if (rq->post_schedule) { 2250 unsigned long flags; 2251 2252 raw_spin_lock_irqsave(&rq->lock, flags); 2253 if (rq->curr->sched_class->post_schedule) 2254 rq->curr->sched_class->post_schedule(rq); 2255 raw_spin_unlock_irqrestore(&rq->lock, flags); 2256 2257 rq->post_schedule = 0; 2258 } 2259} 2260 2261#else 2262 2263static inline void post_schedule(struct rq *rq) 2264{ 2265} 2266 2267#endif 2268 2269/** 2270 * schedule_tail - first thing a freshly forked thread must call. 2271 * @prev: the thread we just switched away from. 2272 */ 2273asmlinkage __visible void schedule_tail(struct task_struct *prev) 2274 __releases(rq->lock) 2275{ 2276 struct rq *rq = this_rq(); 2277 2278 finish_task_switch(rq, prev); 2279 2280 /* 2281 * FIXME: do we need to worry about rq being invalidated by the 2282 * task_switch? 2283 */ 2284 post_schedule(rq); 2285 2286#ifdef __ARCH_WANT_UNLOCKED_CTXSW 2287 /* In this case, finish_task_switch does not reenable preemption */ 2288 preempt_enable(); 2289#endif 2290 if (current->set_child_tid) 2291 put_user(task_pid_vnr(current), current->set_child_tid); 2292} 2293 2294/* 2295 * context_switch - switch to the new MM and the new 2296 * thread's register state. 2297 */ 2298static inline void 2299context_switch(struct rq *rq, struct task_struct *prev, 2300 struct task_struct *next) 2301{ 2302 struct mm_struct *mm, *oldmm; 2303 2304 prepare_task_switch(rq, prev, next); 2305 2306 mm = next->mm; 2307 oldmm = prev->active_mm; 2308 /* 2309 * For paravirt, this is coupled with an exit in switch_to to 2310 * combine the page table reload and the switch backend into 2311 * one hypercall. 2312 */ 2313 arch_start_context_switch(prev); 2314 2315 if (!mm) { 2316 next->active_mm = oldmm; 2317 atomic_inc(&oldmm->mm_count); 2318 enter_lazy_tlb(oldmm, next); 2319 } else 2320 switch_mm(oldmm, mm, next); 2321 2322 if (!prev->mm) { 2323 prev->active_mm = NULL; 2324 rq->prev_mm = oldmm; 2325 } 2326 /* 2327 * Since the runqueue lock will be released by the next 2328 * task (which is an invalid locking op but in the case 2329 * of the scheduler it's an obvious special-case), so we 2330 * do an early lockdep release here: 2331 */ 2332#ifndef __ARCH_WANT_UNLOCKED_CTXSW 2333 spin_release(&rq->lock.dep_map, 1, _THIS_IP_); 2334#endif 2335 2336 context_tracking_task_switch(prev, next); 2337 /* Here we just switch the register state and the stack. */ 2338 switch_to(prev, next, prev); 2339 2340 barrier(); 2341 /* 2342 * this_rq must be evaluated again because prev may have moved 2343 * CPUs since it called schedule(), thus the 'rq' on its stack 2344 * frame will be invalid. 2345 */ 2346 finish_task_switch(this_rq(), prev); 2347} 2348 2349/* 2350 * nr_running and nr_context_switches: 2351 * 2352 * externally visible scheduler statistics: current number of runnable 2353 * threads, total number of context switches performed since bootup. 2354 */ 2355unsigned long nr_running(void) 2356{ 2357 unsigned long i, sum = 0; 2358 2359 for_each_online_cpu(i) 2360 sum += cpu_rq(i)->nr_running; 2361 2362 return sum; 2363} 2364 2365unsigned long long nr_context_switches(void) 2366{ 2367 int i; 2368 unsigned long long sum = 0; 2369 2370 for_each_possible_cpu(i) 2371 sum += cpu_rq(i)->nr_switches; 2372 2373 return sum; 2374} 2375 2376unsigned long nr_iowait(void) 2377{ 2378 unsigned long i, sum = 0; 2379 2380 for_each_possible_cpu(i) 2381 sum += atomic_read(&cpu_rq(i)->nr_iowait); 2382 2383 return sum; 2384} 2385 2386unsigned long nr_iowait_cpu(int cpu) 2387{ 2388 struct rq *this = cpu_rq(cpu); 2389 return atomic_read(&this->nr_iowait); 2390} 2391 2392#ifdef CONFIG_SMP 2393 2394/* 2395 * sched_exec - execve() is a valuable balancing opportunity, because at 2396 * this point the task has the smallest effective memory and cache footprint. 2397 */ 2398void sched_exec(void) 2399{ 2400 struct task_struct *p = current; 2401 unsigned long flags; 2402 int dest_cpu; 2403 2404 raw_spin_lock_irqsave(&p->pi_lock, flags); 2405 dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), SD_BALANCE_EXEC, 0); 2406 if (dest_cpu == smp_processor_id()) 2407 goto unlock; 2408 2409 if (likely(cpu_active(dest_cpu))) { 2410 struct migration_arg arg = { p, dest_cpu }; 2411 2412 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 2413 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg); 2414 return; 2415 } 2416unlock: 2417 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 2418} 2419 2420#endif 2421 2422DEFINE_PER_CPU(struct kernel_stat, kstat); 2423DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat); 2424 2425EXPORT_PER_CPU_SYMBOL(kstat); 2426EXPORT_PER_CPU_SYMBOL(kernel_cpustat); 2427 2428/* 2429 * Return any ns on the sched_clock that have not yet been accounted in 2430 * @p in case that task is currently running. 2431 * 2432 * Called with task_rq_lock() held on @rq. 2433 */ 2434static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq) 2435{ 2436 u64 ns = 0; 2437 2438 if (task_current(rq, p)) { 2439 update_rq_clock(rq); 2440 ns = rq_clock_task(rq) - p->se.exec_start; 2441 if ((s64)ns < 0) 2442 ns = 0; 2443 } 2444 2445 return ns; 2446} 2447 2448unsigned long long task_delta_exec(struct task_struct *p) 2449{ 2450 unsigned long flags; 2451 struct rq *rq; 2452 u64 ns = 0; 2453 2454 rq = task_rq_lock(p, &flags); 2455 ns = do_task_delta_exec(p, rq); 2456 task_rq_unlock(rq, p, &flags); 2457 2458 return ns; 2459} 2460 2461/* 2462 * Return accounted runtime for the task. 2463 * In case the task is currently running, return the runtime plus current's 2464 * pending runtime that have not been accounted yet. 2465 */ 2466unsigned long long task_sched_runtime(struct task_struct *p) 2467{ 2468 unsigned long flags; 2469 struct rq *rq; 2470 u64 ns = 0; 2471 2472#if defined(CONFIG_64BIT) && defined(CONFIG_SMP) 2473 /* 2474 * 64-bit doesn't need locks to atomically read a 64bit value. 2475 * So we have a optimization chance when the task's delta_exec is 0. 2476 * Reading ->on_cpu is racy, but this is ok. 2477 * 2478 * If we race with it leaving cpu, we'll take a lock. So we're correct. 2479 * If we race with it entering cpu, unaccounted time is 0. This is 2480 * indistinguishable from the read occurring a few cycles earlier. 2481 */ 2482 if (!p->on_cpu) 2483 return p->se.sum_exec_runtime; 2484#endif 2485 2486 rq = task_rq_lock(p, &flags); 2487 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq); 2488 task_rq_unlock(rq, p, &flags); 2489 2490 return ns; 2491} 2492 2493/* 2494 * This function gets called by the timer code, with HZ frequency. 2495 * We call it with interrupts disabled. 2496 */ 2497void scheduler_tick(void) 2498{ 2499 int cpu = smp_processor_id(); 2500 struct rq *rq = cpu_rq(cpu); 2501 struct task_struct *curr = rq->curr; 2502 2503 sched_clock_tick(); 2504 2505 raw_spin_lock(&rq->lock); 2506 update_rq_clock(rq); 2507 curr->sched_class->task_tick(rq, curr, 0); 2508 update_cpu_load_active(rq); 2509 raw_spin_unlock(&rq->lock); 2510 2511 perf_event_task_tick(); 2512 2513#ifdef CONFIG_SMP 2514 rq->idle_balance = idle_cpu(cpu); 2515 trigger_load_balance(rq); 2516#endif 2517 rq_last_tick_reset(rq); 2518} 2519 2520#ifdef CONFIG_NO_HZ_FULL 2521/** 2522 * scheduler_tick_max_deferment 2523 * 2524 * Keep at least one tick per second when a single 2525 * active task is running because the scheduler doesn't 2526 * yet completely support full dynticks environment. 2527 * 2528 * This makes sure that uptime, CFS vruntime, load 2529 * balancing, etc... continue to move forward, even 2530 * with a very low granularity. 2531 * 2532 * Return: Maximum deferment in nanoseconds. 2533 */ 2534u64 scheduler_tick_max_deferment(void) 2535{ 2536 struct rq *rq = this_rq(); 2537 unsigned long next, now = ACCESS_ONCE(jiffies); 2538 2539 next = rq->last_sched_tick + HZ; 2540 2541 if (time_before_eq(next, now)) 2542 return 0; 2543 2544 return jiffies_to_nsecs(next - now); 2545} 2546#endif 2547 2548notrace unsigned long get_parent_ip(unsigned long addr) 2549{ 2550 if (in_lock_functions(addr)) { 2551 addr = CALLER_ADDR2; 2552 if (in_lock_functions(addr)) 2553 addr = CALLER_ADDR3; 2554 } 2555 return addr; 2556} 2557 2558#if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \ 2559 defined(CONFIG_PREEMPT_TRACER)) 2560 2561void preempt_count_add(int val) 2562{ 2563#ifdef CONFIG_DEBUG_PREEMPT 2564 /* 2565 * Underflow? 2566 */ 2567 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0))) 2568 return; 2569#endif 2570 __preempt_count_add(val); 2571#ifdef CONFIG_DEBUG_PREEMPT 2572 /* 2573 * Spinlock count overflowing soon? 2574 */ 2575 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >= 2576 PREEMPT_MASK - 10); 2577#endif 2578 if (preempt_count() == val) { 2579 unsigned long ip = get_parent_ip(CALLER_ADDR1); 2580#ifdef CONFIG_DEBUG_PREEMPT 2581 current->preempt_disable_ip = ip; 2582#endif 2583 trace_preempt_off(CALLER_ADDR0, ip); 2584 } 2585} 2586EXPORT_SYMBOL(preempt_count_add); 2587NOKPROBE_SYMBOL(preempt_count_add); 2588 2589void preempt_count_sub(int val) 2590{ 2591#ifdef CONFIG_DEBUG_PREEMPT 2592 /* 2593 * Underflow? 2594 */ 2595 if (DEBUG_LOCKS_WARN_ON(val > preempt_count())) 2596 return; 2597 /* 2598 * Is the spinlock portion underflowing? 2599 */ 2600 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) && 2601 !(preempt_count() & PREEMPT_MASK))) 2602 return; 2603#endif 2604 2605 if (preempt_count() == val) 2606 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1)); 2607 __preempt_count_sub(val); 2608} 2609EXPORT_SYMBOL(preempt_count_sub); 2610NOKPROBE_SYMBOL(preempt_count_sub); 2611 2612#endif 2613 2614/* 2615 * Print scheduling while atomic bug: 2616 */ 2617static noinline void __schedule_bug(struct task_struct *prev) 2618{ 2619 if (oops_in_progress) 2620 return; 2621 2622 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n", 2623 prev->comm, prev->pid, preempt_count()); 2624 2625 debug_show_held_locks(prev); 2626 print_modules(); 2627 if (irqs_disabled()) 2628 print_irqtrace_events(prev); 2629#ifdef CONFIG_DEBUG_PREEMPT 2630 if (in_atomic_preempt_off()) { 2631 pr_err("Preemption disabled at:"); 2632 print_ip_sym(current->preempt_disable_ip); 2633 pr_cont("\n"); 2634 } 2635#endif 2636 dump_stack(); 2637 add_taint(TAINT_WARN, LOCKDEP_STILL_OK); 2638} 2639 2640/* 2641 * Various schedule()-time debugging checks and statistics: 2642 */ 2643static inline void schedule_debug(struct task_struct *prev) 2644{ 2645 /* 2646 * Test if we are atomic. Since do_exit() needs to call into 2647 * schedule() atomically, we ignore that path. Otherwise whine 2648 * if we are scheduling when we should not. 2649 */ 2650 if (unlikely(in_atomic_preempt_off() && prev->state != TASK_DEAD)) 2651 __schedule_bug(prev); 2652 rcu_sleep_check(); 2653 2654 profile_hit(SCHED_PROFILING, __builtin_return_address(0)); 2655 2656 schedstat_inc(this_rq(), sched_count); 2657} 2658 2659/* 2660 * Pick up the highest-prio task: 2661 */ 2662static inline struct task_struct * 2663pick_next_task(struct rq *rq, struct task_struct *prev) 2664{ 2665 const struct sched_class *class = &fair_sched_class; 2666 struct task_struct *p; 2667 2668 /* 2669 * Optimization: we know that if all tasks are in 2670 * the fair class we can call that function directly: 2671 */ 2672 if (likely(prev->sched_class == class && 2673 rq->nr_running == rq->cfs.h_nr_running)) { 2674 p = fair_sched_class.pick_next_task(rq, prev); 2675 if (unlikely(p == RETRY_TASK)) 2676 goto again; 2677 2678 /* assumes fair_sched_class->next == idle_sched_class */ 2679 if (unlikely(!p)) 2680 p = idle_sched_class.pick_next_task(rq, prev); 2681 2682 return p; 2683 } 2684 2685again: 2686 for_each_class(class) { 2687 p = class->pick_next_task(rq, prev); 2688 if (p) { 2689 if (unlikely(p == RETRY_TASK)) 2690 goto again; 2691 return p; 2692 } 2693 } 2694 2695 BUG(); /* the idle class will always have a runnable task */ 2696} 2697 2698/* 2699 * __schedule() is the main scheduler function. 2700 * 2701 * The main means of driving the scheduler and thus entering this function are: 2702 * 2703 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc. 2704 * 2705 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return 2706 * paths. For example, see arch/x86/entry_64.S. 2707 * 2708 * To drive preemption between tasks, the scheduler sets the flag in timer 2709 * interrupt handler scheduler_tick(). 2710 * 2711 * 3. Wakeups don't really cause entry into schedule(). They add a 2712 * task to the run-queue and that's it. 2713 * 2714 * Now, if the new task added to the run-queue preempts the current 2715 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets 2716 * called on the nearest possible occasion: 2717 * 2718 * - If the kernel is preemptible (CONFIG_PREEMPT=y): 2719 * 2720 * - in syscall or exception context, at the next outmost 2721 * preempt_enable(). (this might be as soon as the wake_up()'s 2722 * spin_unlock()!) 2723 * 2724 * - in IRQ context, return from interrupt-handler to 2725 * preemptible context 2726 * 2727 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set) 2728 * then at the next: 2729 * 2730 * - cond_resched() call 2731 * - explicit schedule() call 2732 * - return from syscall or exception to user-space 2733 * - return from interrupt-handler to user-space 2734 */ 2735static void __sched __schedule(void) 2736{ 2737 struct task_struct *prev, *next; 2738 unsigned long *switch_count; 2739 struct rq *rq; 2740 int cpu; 2741 2742need_resched: 2743 preempt_disable(); 2744 cpu = smp_processor_id(); 2745 rq = cpu_rq(cpu); 2746 rcu_note_context_switch(cpu); 2747 prev = rq->curr; 2748 2749 schedule_debug(prev); 2750 2751 if (sched_feat(HRTICK)) 2752 hrtick_clear(rq); 2753 2754 /* 2755 * Make sure that signal_pending_state()->signal_pending() below 2756 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE) 2757 * done by the caller to avoid the race with signal_wake_up(). 2758 */ 2759 smp_mb__before_spinlock(); 2760 raw_spin_lock_irq(&rq->lock); 2761 2762 switch_count = &prev->nivcsw; 2763 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) { 2764 if (unlikely(signal_pending_state(prev->state, prev))) { 2765 prev->state = TASK_RUNNING; 2766 } else { 2767 deactivate_task(rq, prev, DEQUEUE_SLEEP); 2768 prev->on_rq = 0; 2769 2770 /* 2771 * If a worker went to sleep, notify and ask workqueue 2772 * whether it wants to wake up a task to maintain 2773 * concurrency. 2774 */ 2775 if (prev->flags & PF_WQ_WORKER) { 2776 struct task_struct *to_wakeup; 2777 2778 to_wakeup = wq_worker_sleeping(prev, cpu); 2779 if (to_wakeup) 2780 try_to_wake_up_local(to_wakeup); 2781 } 2782 } 2783 switch_count = &prev->nvcsw; 2784 } 2785 2786 if (prev->on_rq || rq->skip_clock_update < 0) 2787 update_rq_clock(rq); 2788 2789 next = pick_next_task(rq, prev); 2790 clear_tsk_need_resched(prev); 2791 clear_preempt_need_resched(); 2792 rq->skip_clock_update = 0; 2793 2794 if (likely(prev != next)) { 2795 rq->nr_switches++; 2796 rq->curr = next; 2797 ++*switch_count; 2798 2799 context_switch(rq, prev, next); /* unlocks the rq */ 2800 /* 2801 * The context switch have flipped the stack from under us 2802 * and restored the local variables which were saved when 2803 * this task called schedule() in the past. prev == current 2804 * is still correct, but it can be moved to another cpu/rq. 2805 */ 2806 cpu = smp_processor_id(); 2807 rq = cpu_rq(cpu); 2808 } else 2809 raw_spin_unlock_irq(&rq->lock); 2810 2811 post_schedule(rq); 2812 2813 sched_preempt_enable_no_resched(); 2814 if (need_resched()) 2815 goto need_resched; 2816} 2817 2818static inline void sched_submit_work(struct task_struct *tsk) 2819{ 2820 if (!tsk->state || tsk_is_pi_blocked(tsk)) 2821 return; 2822 /* 2823 * If we are going to sleep and we have plugged IO queued, 2824 * make sure to submit it to avoid deadlocks. 2825 */ 2826 if (blk_needs_flush_plug(tsk)) 2827 blk_schedule_flush_plug(tsk); 2828} 2829 2830asmlinkage __visible void __sched schedule(void) 2831{ 2832 struct task_struct *tsk = current; 2833 2834 sched_submit_work(tsk); 2835 __schedule(); 2836} 2837EXPORT_SYMBOL(schedule); 2838 2839#ifdef CONFIG_CONTEXT_TRACKING 2840asmlinkage __visible void __sched schedule_user(void) 2841{ 2842 /* 2843 * If we come here after a random call to set_need_resched(), 2844 * or we have been woken up remotely but the IPI has not yet arrived, 2845 * we haven't yet exited the RCU idle mode. Do it here manually until 2846 * we find a better solution. 2847 */ 2848 user_exit(); 2849 schedule(); 2850 user_enter(); 2851} 2852#endif 2853 2854/** 2855 * schedule_preempt_disabled - called with preemption disabled 2856 * 2857 * Returns with preemption disabled. Note: preempt_count must be 1 2858 */ 2859void __sched schedule_preempt_disabled(void) 2860{ 2861 sched_preempt_enable_no_resched(); 2862 schedule(); 2863 preempt_disable(); 2864} 2865 2866#ifdef CONFIG_PREEMPT 2867/* 2868 * this is the entry point to schedule() from in-kernel preemption 2869 * off of preempt_enable. Kernel preemptions off return from interrupt 2870 * occur there and call schedule directly. 2871 */ 2872asmlinkage __visible void __sched notrace preempt_schedule(void) 2873{ 2874 /* 2875 * If there is a non-zero preempt_count or interrupts are disabled, 2876 * we do not want to preempt the current task. Just return.. 2877 */ 2878 if (likely(!preemptible())) 2879 return; 2880 2881 do { 2882 __preempt_count_add(PREEMPT_ACTIVE); 2883 __schedule(); 2884 __preempt_count_sub(PREEMPT_ACTIVE); 2885 2886 /* 2887 * Check again in case we missed a preemption opportunity 2888 * between schedule and now. 2889 */ 2890 barrier(); 2891 } while (need_resched()); 2892} 2893NOKPROBE_SYMBOL(preempt_schedule); 2894EXPORT_SYMBOL(preempt_schedule); 2895#endif /* CONFIG_PREEMPT */ 2896 2897/* 2898 * this is the entry point to schedule() from kernel preemption 2899 * off of irq context. 2900 * Note, that this is called and return with irqs disabled. This will 2901 * protect us against recursive calling from irq. 2902 */ 2903asmlinkage __visible void __sched preempt_schedule_irq(void) 2904{ 2905 enum ctx_state prev_state; 2906 2907 /* Catch callers which need to be fixed */ 2908 BUG_ON(preempt_count() || !irqs_disabled()); 2909 2910 prev_state = exception_enter(); 2911 2912 do { 2913 __preempt_count_add(PREEMPT_ACTIVE); 2914 local_irq_enable(); 2915 __schedule(); 2916 local_irq_disable(); 2917 __preempt_count_sub(PREEMPT_ACTIVE); 2918 2919 /* 2920 * Check again in case we missed a preemption opportunity 2921 * between schedule and now. 2922 */ 2923 barrier(); 2924 } while (need_resched()); 2925 2926 exception_exit(prev_state); 2927} 2928 2929int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags, 2930 void *key) 2931{ 2932 return try_to_wake_up(curr->private, mode, wake_flags); 2933} 2934EXPORT_SYMBOL(default_wake_function); 2935 2936#ifdef CONFIG_RT_MUTEXES 2937 2938/* 2939 * rt_mutex_setprio - set the current priority of a task 2940 * @p: task 2941 * @prio: prio value (kernel-internal form) 2942 * 2943 * This function changes the 'effective' priority of a task. It does 2944 * not touch ->normal_prio like __setscheduler(). 2945 * 2946 * Used by the rt_mutex code to implement priority inheritance 2947 * logic. Call site only calls if the priority of the task changed. 2948 */ 2949void rt_mutex_setprio(struct task_struct *p, int prio) 2950{ 2951 int oldprio, on_rq, running, enqueue_flag = 0; 2952 struct rq *rq; 2953 const struct sched_class *prev_class; 2954 2955 BUG_ON(prio > MAX_PRIO); 2956 2957 rq = __task_rq_lock(p); 2958 2959 /* 2960 * Idle task boosting is a nono in general. There is one 2961 * exception, when PREEMPT_RT and NOHZ is active: 2962 * 2963 * The idle task calls get_next_timer_interrupt() and holds 2964 * the timer wheel base->lock on the CPU and another CPU wants 2965 * to access the timer (probably to cancel it). We can safely 2966 * ignore the boosting request, as the idle CPU runs this code 2967 * with interrupts disabled and will complete the lock 2968 * protected section without being interrupted. So there is no 2969 * real need to boost. 2970 */ 2971 if (unlikely(p == rq->idle)) { 2972 WARN_ON(p != rq->curr); 2973 WARN_ON(p->pi_blocked_on); 2974 goto out_unlock; 2975 } 2976 2977 trace_sched_pi_setprio(p, prio); 2978 p->pi_top_task = rt_mutex_get_top_task(p); 2979 oldprio = p->prio; 2980 prev_class = p->sched_class; 2981 on_rq = p->on_rq; 2982 running = task_current(rq, p); 2983 if (on_rq) 2984 dequeue_task(rq, p, 0); 2985 if (running) 2986 p->sched_class->put_prev_task(rq, p); 2987 2988 /* 2989 * Boosting condition are: 2990 * 1. -rt task is running and holds mutex A 2991 * --> -dl task blocks on mutex A 2992 * 2993 * 2. -dl task is running and holds mutex A 2994 * --> -dl task blocks on mutex A and could preempt the 2995 * running task 2996 */ 2997 if (dl_prio(prio)) { 2998 if (!dl_prio(p->normal_prio) || (p->pi_top_task && 2999 dl_entity_preempt(&p->pi_top_task->dl, &p->dl))) { 3000 p->dl.dl_boosted = 1; 3001 p->dl.dl_throttled = 0; 3002 enqueue_flag = ENQUEUE_REPLENISH; 3003 } else 3004 p->dl.dl_boosted = 0; 3005 p->sched_class = &dl_sched_class; 3006 } else if (rt_prio(prio)) { 3007 if (dl_prio(oldprio)) 3008 p->dl.dl_boosted = 0; 3009 if (oldprio < prio) 3010 enqueue_flag = ENQUEUE_HEAD; 3011 p->sched_class = &rt_sched_class; 3012 } else { 3013 if (dl_prio(oldprio)) 3014 p->dl.dl_boosted = 0; 3015 p->sched_class = &fair_sched_class; 3016 } 3017 3018 p->prio = prio; 3019 3020 if (running) 3021 p->sched_class->set_curr_task(rq); 3022 if (on_rq) 3023 enqueue_task(rq, p, enqueue_flag); 3024 3025 check_class_changed(rq, p, prev_class, oldprio); 3026out_unlock: 3027 __task_rq_unlock(rq); 3028} 3029#endif 3030 3031void set_user_nice(struct task_struct *p, long nice) 3032{ 3033 int old_prio, delta, on_rq; 3034 unsigned long flags; 3035 struct rq *rq; 3036 3037 if (task_nice(p) == nice || nice < MIN_NICE || nice > MAX_NICE) 3038 return; 3039 /* 3040 * We have to be careful, if called from sys_setpriority(), 3041 * the task might be in the middle of scheduling on another CPU. 3042 */ 3043 rq = task_rq_lock(p, &flags); 3044 /* 3045 * The RT priorities are set via sched_setscheduler(), but we still 3046 * allow the 'normal' nice value to be set - but as expected 3047 * it wont have any effect on scheduling until the task is 3048 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR: 3049 */ 3050 if (task_has_dl_policy(p) || task_has_rt_policy(p)) { 3051 p->static_prio = NICE_TO_PRIO(nice); 3052 goto out_unlock; 3053 } 3054 on_rq = p->on_rq; 3055 if (on_rq) 3056 dequeue_task(rq, p, 0); 3057 3058 p->static_prio = NICE_TO_PRIO(nice); 3059 set_load_weight(p); 3060 old_prio = p->prio; 3061 p->prio = effective_prio(p); 3062 delta = p->prio - old_prio; 3063 3064 if (on_rq) { 3065 enqueue_task(rq, p, 0); 3066 /* 3067 * If the task increased its priority or is running and 3068 * lowered its priority, then reschedule its CPU: 3069 */ 3070 if (delta < 0 || (delta > 0 && task_running(rq, p))) 3071 resched_task(rq->curr); 3072 } 3073out_unlock: 3074 task_rq_unlock(rq, p, &flags); 3075} 3076EXPORT_SYMBOL(set_user_nice); 3077 3078/* 3079 * can_nice - check if a task can reduce its nice value 3080 * @p: task 3081 * @nice: nice value 3082 */ 3083int can_nice(const struct task_struct *p, const int nice) 3084{ 3085 /* convert nice value [19,-20] to rlimit style value [1,40] */ 3086 int nice_rlim = nice_to_rlimit(nice); 3087 3088 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) || 3089 capable(CAP_SYS_NICE)); 3090} 3091 3092#ifdef __ARCH_WANT_SYS_NICE 3093 3094/* 3095 * sys_nice - change the priority of the current process. 3096 * @increment: priority increment 3097 * 3098 * sys_setpriority is a more generic, but much slower function that 3099 * does similar things. 3100 */ 3101SYSCALL_DEFINE1(nice, int, increment) 3102{ 3103 long nice, retval; 3104 3105 /* 3106 * Setpriority might change our priority at the same moment. 3107 * We don't have to worry. Conceptually one call occurs first 3108 * and we have a single winner. 3109 */ 3110 increment = clamp(increment, -NICE_WIDTH, NICE_WIDTH); 3111 nice = task_nice(current) + increment; 3112 3113 nice = clamp_val(nice, MIN_NICE, MAX_NICE); 3114 if (increment < 0 && !can_nice(current, nice)) 3115 return -EPERM; 3116 3117 retval = security_task_setnice(current, nice); 3118 if (retval) 3119 return retval; 3120 3121 set_user_nice(current, nice); 3122 return 0; 3123} 3124 3125#endif 3126 3127/** 3128 * task_prio - return the priority value of a given task. 3129 * @p: the task in question. 3130 * 3131 * Return: The priority value as seen by users in /proc. 3132 * RT tasks are offset by -200. Normal tasks are centered 3133 * around 0, value goes from -16 to +15. 3134 */ 3135int task_prio(const struct task_struct *p) 3136{ 3137 return p->prio - MAX_RT_PRIO; 3138} 3139 3140/** 3141 * idle_cpu - is a given cpu idle currently? 3142 * @cpu: the processor in question. 3143 * 3144 * Return: 1 if the CPU is currently idle. 0 otherwise. 3145 */ 3146int idle_cpu(int cpu) 3147{ 3148 struct rq *rq = cpu_rq(cpu); 3149 3150 if (rq->curr != rq->idle) 3151 return 0; 3152 3153 if (rq->nr_running) 3154 return 0; 3155 3156#ifdef CONFIG_SMP 3157 if (!llist_empty(&rq->wake_list)) 3158 return 0; 3159#endif 3160 3161 return 1; 3162} 3163 3164/** 3165 * idle_task - return the idle task for a given cpu. 3166 * @cpu: the processor in question. 3167 * 3168 * Return: The idle task for the cpu @cpu. 3169 */ 3170struct task_struct *idle_task(int cpu) 3171{ 3172 return cpu_rq(cpu)->idle; 3173} 3174 3175/** 3176 * find_process_by_pid - find a process with a matching PID value. 3177 * @pid: the pid in question. 3178 * 3179 * The task of @pid, if found. %NULL otherwise. 3180 */ 3181static struct task_struct *find_process_by_pid(pid_t pid) 3182{ 3183 return pid ? find_task_by_vpid(pid) : current; 3184} 3185 3186/* 3187 * This function initializes the sched_dl_entity of a newly becoming 3188 * SCHED_DEADLINE task. 3189 * 3190 * Only the static values are considered here, the actual runtime and the 3191 * absolute deadline will be properly calculated when the task is enqueued 3192 * for the first time with its new policy. 3193 */ 3194static void 3195__setparam_dl(struct task_struct *p, const struct sched_attr *attr) 3196{ 3197 struct sched_dl_entity *dl_se = &p->dl; 3198 3199 init_dl_task_timer(dl_se); 3200 dl_se->dl_runtime = attr->sched_runtime; 3201 dl_se->dl_deadline = attr->sched_deadline; 3202 dl_se->dl_period = attr->sched_period ?: dl_se->dl_deadline; 3203 dl_se->flags = attr->sched_flags; 3204 dl_se->dl_bw = to_ratio(dl_se->dl_period, dl_se->dl_runtime); 3205 dl_se->dl_throttled = 0; 3206 dl_se->dl_new = 1; 3207 dl_se->dl_yielded = 0; 3208} 3209 3210static void __setscheduler_params(struct task_struct *p, 3211 const struct sched_attr *attr) 3212{ 3213 int policy = attr->sched_policy; 3214 3215 if (policy == -1) /* setparam */ 3216 policy = p->policy; 3217 3218 p->policy = policy; 3219 3220 if (dl_policy(policy)) 3221 __setparam_dl(p, attr); 3222 else if (fair_policy(policy)) 3223 p->static_prio = NICE_TO_PRIO(attr->sched_nice); 3224 3225 /* 3226 * __sched_setscheduler() ensures attr->sched_priority == 0 when 3227 * !rt_policy. Always setting this ensures that things like 3228 * getparam()/getattr() don't report silly values for !rt tasks. 3229 */ 3230 p->rt_priority = attr->sched_priority; 3231 p->normal_prio = normal_prio(p); 3232 set_load_weight(p); 3233} 3234 3235/* Actually do priority change: must hold pi & rq lock. */ 3236static void __setscheduler(struct rq *rq, struct task_struct *p, 3237 const struct sched_attr *attr) 3238{ 3239 __setscheduler_params(p, attr); 3240 3241 /* 3242 * If we get here, there was no pi waiters boosting the 3243 * task. It is safe to use the normal prio. 3244 */ 3245 p->prio = normal_prio(p); 3246 3247 if (dl_prio(p->prio)) 3248 p->sched_class = &dl_sched_class; 3249 else if (rt_prio(p->prio)) 3250 p->sched_class = &rt_sched_class; 3251 else 3252 p->sched_class = &fair_sched_class; 3253} 3254 3255static void 3256__getparam_dl(struct task_struct *p, struct sched_attr *attr) 3257{ 3258 struct sched_dl_entity *dl_se = &p->dl; 3259 3260 attr->sched_priority = p->rt_priority; 3261 attr->sched_runtime = dl_se->dl_runtime; 3262 attr->sched_deadline = dl_se->dl_deadline; 3263 attr->sched_period = dl_se->dl_period; 3264 attr->sched_flags = dl_se->flags; 3265} 3266 3267/* 3268 * This function validates the new parameters of a -deadline task. 3269 * We ask for the deadline not being zero, and greater or equal 3270 * than the runtime, as well as the period of being zero or 3271 * greater than deadline. Furthermore, we have to be sure that 3272 * user parameters are above the internal resolution of 1us (we 3273 * check sched_runtime only since it is always the smaller one) and 3274 * below 2^63 ns (we have to check both sched_deadline and 3275 * sched_period, as the latter can be zero). 3276 */ 3277static bool 3278__checkparam_dl(const struct sched_attr *attr) 3279{ 3280 /* deadline != 0 */ 3281 if (attr->sched_deadline == 0) 3282 return false; 3283 3284 /* 3285 * Since we truncate DL_SCALE bits, make sure we're at least 3286 * that big. 3287 */ 3288 if (attr->sched_runtime < (1ULL << DL_SCALE)) 3289 return false; 3290 3291 /* 3292 * Since we use the MSB for wrap-around and sign issues, make 3293 * sure it's not set (mind that period can be equal to zero). 3294 */ 3295 if (attr->sched_deadline & (1ULL << 63) || 3296 attr->sched_period & (1ULL << 63)) 3297 return false; 3298 3299 /* runtime <= deadline <= period (if period != 0) */ 3300 if ((attr->sched_period != 0 && 3301 attr->sched_period < attr->sched_deadline) || 3302 attr->sched_deadline < attr->sched_runtime) 3303 return false; 3304 3305 return true; 3306} 3307 3308/* 3309 * check the target process has a UID that matches the current process's 3310 */ 3311static bool check_same_owner(struct task_struct *p) 3312{ 3313 const struct cred *cred = current_cred(), *pcred; 3314 bool match; 3315 3316 rcu_read_lock(); 3317 pcred = __task_cred(p); 3318 match = (uid_eq(cred->euid, pcred->euid) || 3319 uid_eq(cred->euid, pcred->uid)); 3320 rcu_read_unlock(); 3321 return match; 3322} 3323 3324static int __sched_setscheduler(struct task_struct *p, 3325 const struct sched_attr *attr, 3326 bool user) 3327{ 3328 int newprio = dl_policy(attr->sched_policy) ? MAX_DL_PRIO - 1 : 3329 MAX_RT_PRIO - 1 - attr->sched_priority; 3330 int retval, oldprio, oldpolicy = -1, on_rq, running; 3331 int policy = attr->sched_policy; 3332 unsigned long flags; 3333 const struct sched_class *prev_class; 3334 struct rq *rq; 3335 int reset_on_fork; 3336 3337 /* may grab non-irq protected spin_locks */ 3338 BUG_ON(in_interrupt()); 3339recheck: 3340 /* double check policy once rq lock held */ 3341 if (policy < 0) { 3342 reset_on_fork = p->sched_reset_on_fork; 3343 policy = oldpolicy = p->policy; 3344 } else { 3345 reset_on_fork = !!(attr->sched_flags & SCHED_FLAG_RESET_ON_FORK); 3346 3347 if (policy != SCHED_DEADLINE && 3348 policy != SCHED_FIFO && policy != SCHED_RR && 3349 policy != SCHED_NORMAL && policy != SCHED_BATCH && 3350 policy != SCHED_IDLE) 3351 return -EINVAL; 3352 } 3353 3354 if (attr->sched_flags & ~(SCHED_FLAG_RESET_ON_FORK)) 3355 return -EINVAL; 3356 3357 /* 3358 * Valid priorities for SCHED_FIFO and SCHED_RR are 3359 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL, 3360 * SCHED_BATCH and SCHED_IDLE is 0. 3361 */ 3362 if ((p->mm && attr->sched_priority > MAX_USER_RT_PRIO-1) || 3363 (!p->mm && attr->sched_priority > MAX_RT_PRIO-1)) 3364 return -EINVAL; 3365 if ((dl_policy(policy) && !__checkparam_dl(attr)) || 3366 (rt_policy(policy) != (attr->sched_priority != 0))) 3367 return -EINVAL; 3368 3369 /* 3370 * Allow unprivileged RT tasks to decrease priority: 3371 */ 3372 if (user && !capable(CAP_SYS_NICE)) { 3373 if (fair_policy(policy)) { 3374 if (attr->sched_nice < task_nice(p) && 3375 !can_nice(p, attr->sched_nice)) 3376 return -EPERM; 3377 } 3378 3379 if (rt_policy(policy)) { 3380 unsigned long rlim_rtprio = 3381 task_rlimit(p, RLIMIT_RTPRIO); 3382 3383 /* can't set/change the rt policy */ 3384 if (policy != p->policy && !rlim_rtprio) 3385 return -EPERM; 3386 3387 /* can't increase priority */ 3388 if (attr->sched_priority > p->rt_priority && 3389 attr->sched_priority > rlim_rtprio) 3390 return -EPERM; 3391 } 3392 3393 /* 3394 * Can't set/change SCHED_DEADLINE policy at all for now 3395 * (safest behavior); in the future we would like to allow 3396 * unprivileged DL tasks to increase their relative deadline 3397 * or reduce their runtime (both ways reducing utilization) 3398 */ 3399 if (dl_policy(policy)) 3400 return -EPERM; 3401 3402 /* 3403 * Treat SCHED_IDLE as nice 20. Only allow a switch to 3404 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it. 3405 */ 3406 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) { 3407 if (!can_nice(p, task_nice(p))) 3408 return -EPERM; 3409 } 3410 3411 /* can't change other user's priorities */ 3412 if (!check_same_owner(p)) 3413 return -EPERM; 3414 3415 /* Normal users shall not reset the sched_reset_on_fork flag */ 3416 if (p->sched_reset_on_fork && !reset_on_fork) 3417 return -EPERM; 3418 } 3419 3420 if (user) { 3421 retval = security_task_setscheduler(p); 3422 if (retval) 3423 return retval; 3424 } 3425 3426 /* 3427 * make sure no PI-waiters arrive (or leave) while we are 3428 * changing the priority of the task: 3429 * 3430 * To be able to change p->policy safely, the appropriate 3431 * runqueue lock must be held. 3432 */ 3433 rq = task_rq_lock(p, &flags); 3434 3435 /* 3436 * Changing the policy of the stop threads its a very bad idea 3437 */ 3438 if (p == rq->stop) { 3439 task_rq_unlock(rq, p, &flags); 3440 return -EINVAL; 3441 } 3442 3443 /* 3444 * If not changing anything there's no need to proceed further, 3445 * but store a possible modification of reset_on_fork. 3446 */ 3447 if (unlikely(policy == p->policy)) { 3448 if (fair_policy(policy) && attr->sched_nice != task_nice(p)) 3449 goto change; 3450 if (rt_policy(policy) && attr->sched_priority != p->rt_priority) 3451 goto change; 3452 if (dl_policy(policy)) 3453 goto change; 3454 3455 p->sched_reset_on_fork = reset_on_fork; 3456 task_rq_unlock(rq, p, &flags); 3457 return 0; 3458 } 3459change: 3460 3461 if (user) { 3462#ifdef CONFIG_RT_GROUP_SCHED 3463 /* 3464 * Do not allow realtime tasks into groups that have no runtime 3465 * assigned. 3466 */ 3467 if (rt_bandwidth_enabled() && rt_policy(policy) && 3468 task_group(p)->rt_bandwidth.rt_runtime == 0 && 3469 !task_group_is_autogroup(task_group(p))) { 3470 task_rq_unlock(rq, p, &flags); 3471 return -EPERM; 3472 } 3473#endif 3474#ifdef CONFIG_SMP 3475 if (dl_bandwidth_enabled() && dl_policy(policy)) { 3476 cpumask_t *span = rq->rd->span; 3477 3478 /* 3479 * Don't allow tasks with an affinity mask smaller than 3480 * the entire root_domain to become SCHED_DEADLINE. We 3481 * will also fail if there's no bandwidth available. 3482 */ 3483 if (!cpumask_subset(span, &p->cpus_allowed) || 3484 rq->rd->dl_bw.bw == 0) { 3485 task_rq_unlock(rq, p, &flags); 3486 return -EPERM; 3487 } 3488 } 3489#endif 3490 } 3491 3492 /* recheck policy now with rq lock held */ 3493 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) { 3494 policy = oldpolicy = -1; 3495 task_rq_unlock(rq, p, &flags); 3496 goto recheck; 3497 } 3498 3499 /* 3500 * If setscheduling to SCHED_DEADLINE (or changing the parameters 3501 * of a SCHED_DEADLINE task) we need to check if enough bandwidth 3502 * is available. 3503 */ 3504 if ((dl_policy(policy) || dl_task(p)) && dl_overflow(p, policy, attr)) { 3505 task_rq_unlock(rq, p, &flags); 3506 return -EBUSY; 3507 } 3508 3509 p->sched_reset_on_fork = reset_on_fork; 3510 oldprio = p->prio; 3511 3512 /* 3513 * Special case for priority boosted tasks. 3514 * 3515 * If the new priority is lower or equal (user space view) 3516 * than the current (boosted) priority, we just store the new 3517 * normal parameters and do not touch the scheduler class and 3518 * the runqueue. This will be done when the task deboost 3519 * itself. 3520 */ 3521 if (rt_mutex_check_prio(p, newprio)) { 3522 __setscheduler_params(p, attr); 3523 task_rq_unlock(rq, p, &flags); 3524 return 0; 3525 } 3526 3527 on_rq = p->on_rq; 3528 running = task_current(rq, p); 3529 if (on_rq) 3530 dequeue_task(rq, p, 0); 3531 if (running) 3532 p->sched_class->put_prev_task(rq, p); 3533 3534 prev_class = p->sched_class; 3535 __setscheduler(rq, p, attr); 3536 3537 if (running) 3538 p->sched_class->set_curr_task(rq); 3539 if (on_rq) { 3540 /* 3541 * We enqueue to tail when the priority of a task is 3542 * increased (user space view). 3543 */ 3544 enqueue_task(rq, p, oldprio <= p->prio ? ENQUEUE_HEAD : 0); 3545 } 3546 3547 check_class_changed(rq, p, prev_class, oldprio); 3548 task_rq_unlock(rq, p, &flags); 3549 3550 rt_mutex_adjust_pi(p); 3551 3552 return 0; 3553} 3554 3555static int _sched_setscheduler(struct task_struct *p, int policy, 3556 const struct sched_param *param, bool check) 3557{ 3558 struct sched_attr attr = { 3559 .sched_policy = policy, 3560 .sched_priority = param->sched_priority, 3561 .sched_nice = PRIO_TO_NICE(p->static_prio), 3562 }; 3563 3564 /* 3565 * Fixup the legacy SCHED_RESET_ON_FORK hack 3566 */ 3567 if (policy & SCHED_RESET_ON_FORK) { 3568 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK; 3569 policy &= ~SCHED_RESET_ON_FORK; 3570 attr.sched_policy = policy; 3571 } 3572 3573 return __sched_setscheduler(p, &attr, check); 3574} 3575/** 3576 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread. 3577 * @p: the task in question. 3578 * @policy: new policy. 3579 * @param: structure containing the new RT priority. 3580 * 3581 * Return: 0 on success. An error code otherwise. 3582 * 3583 * NOTE that the task may be already dead. 3584 */ 3585int sched_setscheduler(struct task_struct *p, int policy, 3586 const struct sched_param *param) 3587{ 3588 return _sched_setscheduler(p, policy, param, true); 3589} 3590EXPORT_SYMBOL_GPL(sched_setscheduler); 3591 3592int sched_setattr(struct task_struct *p, const struct sched_attr *attr) 3593{ 3594 return __sched_setscheduler(p, attr, true); 3595} 3596EXPORT_SYMBOL_GPL(sched_setattr); 3597 3598/** 3599 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace. 3600 * @p: the task in question. 3601 * @policy: new policy. 3602 * @param: structure containing the new RT priority. 3603 * 3604 * Just like sched_setscheduler, only don't bother checking if the 3605 * current context has permission. For example, this is needed in 3606 * stop_machine(): we create temporary high priority worker threads, 3607 * but our caller might not have that capability. 3608 * 3609 * Return: 0 on success. An error code otherwise. 3610 */ 3611int sched_setscheduler_nocheck(struct task_struct *p, int policy, 3612 const struct sched_param *param) 3613{ 3614 return _sched_setscheduler(p, policy, param, false); 3615} 3616 3617static int 3618do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param) 3619{ 3620 struct sched_param lparam; 3621 struct task_struct *p; 3622 int retval; 3623 3624 if (!param || pid < 0) 3625 return -EINVAL; 3626 if (copy_from_user(&lparam, param, sizeof(struct sched_param))) 3627 return -EFAULT; 3628 3629 rcu_read_lock(); 3630 retval = -ESRCH; 3631 p = find_process_by_pid(pid); 3632 if (p != NULL) 3633 retval = sched_setscheduler(p, policy, &lparam); 3634 rcu_read_unlock(); 3635 3636 return retval; 3637} 3638 3639/* 3640 * Mimics kernel/events/core.c perf_copy_attr(). 3641 */ 3642static int sched_copy_attr(struct sched_attr __user *uattr, 3643 struct sched_attr *attr) 3644{ 3645 u32 size; 3646 int ret; 3647 3648 if (!access_ok(VERIFY_WRITE, uattr, SCHED_ATTR_SIZE_VER0)) 3649 return -EFAULT; 3650 3651 /* 3652 * zero the full structure, so that a short copy will be nice. 3653 */ 3654 memset(attr, 0, sizeof(*attr)); 3655 3656 ret = get_user(size, &uattr->size); 3657 if (ret) 3658 return ret; 3659 3660 if (size > PAGE_SIZE) /* silly large */ 3661 goto err_size; 3662 3663 if (!size) /* abi compat */ 3664 size = SCHED_ATTR_SIZE_VER0; 3665 3666 if (size < SCHED_ATTR_SIZE_VER0) 3667 goto err_size; 3668 3669 /* 3670 * If we're handed a bigger struct than we know of, 3671 * ensure all the unknown bits are 0 - i.e. new 3672 * user-space does not rely on any kernel feature 3673 * extensions we dont know about yet. 3674 */ 3675 if (size > sizeof(*attr)) { 3676 unsigned char __user *addr; 3677 unsigned char __user *end; 3678 unsigned char val; 3679 3680 addr = (void __user *)uattr + sizeof(*attr); 3681 end = (void __user *)uattr + size; 3682 3683 for (; addr < end; addr++) { 3684 ret = get_user(val, addr); 3685 if (ret) 3686 return ret; 3687 if (val) 3688 goto err_size; 3689 } 3690 size = sizeof(*attr); 3691 } 3692 3693 ret = copy_from_user(attr, uattr, size); 3694 if (ret) 3695 return -EFAULT; 3696 3697 /* 3698 * XXX: do we want to be lenient like existing syscalls; or do we want 3699 * to be strict and return an error on out-of-bounds values? 3700 */ 3701 attr->sched_nice = clamp(attr->sched_nice, MIN_NICE, MAX_NICE); 3702 3703 return 0; 3704 3705err_size: 3706 put_user(sizeof(*attr), &uattr->size); 3707 return -E2BIG; 3708} 3709 3710/** 3711 * sys_sched_setscheduler - set/change the scheduler policy and RT priority 3712 * @pid: the pid in question. 3713 * @policy: new policy. 3714 * @param: structure containing the new RT priority. 3715 * 3716 * Return: 0 on success. An error code otherwise. 3717 */ 3718SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy, 3719 struct sched_param __user *, param) 3720{ 3721 /* negative values for policy are not valid */ 3722 if (policy < 0) 3723 return -EINVAL; 3724 3725 return do_sched_setscheduler(pid, policy, param); 3726} 3727 3728/** 3729 * sys_sched_setparam - set/change the RT priority of a thread 3730 * @pid: the pid in question. 3731 * @param: structure containing the new RT priority. 3732 * 3733 * Return: 0 on success. An error code otherwise. 3734 */ 3735SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param) 3736{ 3737 return do_sched_setscheduler(pid, -1, param); 3738} 3739 3740/** 3741 * sys_sched_setattr - same as above, but with extended sched_attr 3742 * @pid: the pid in question. 3743 * @uattr: structure containing the extended parameters. 3744 * @flags: for future extension. 3745 */ 3746SYSCALL_DEFINE3(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr, 3747 unsigned int, flags) 3748{ 3749 struct sched_attr attr; 3750 struct task_struct *p; 3751 int retval; 3752 3753 if (!uattr || pid < 0 || flags) 3754 return -EINVAL; 3755 3756 retval = sched_copy_attr(uattr, &attr); 3757 if (retval) 3758 return retval; 3759 3760 if ((int)attr.sched_policy < 0) 3761 return -EINVAL; 3762 3763 rcu_read_lock(); 3764 retval = -ESRCH; 3765 p = find_process_by_pid(pid); 3766 if (p != NULL) 3767 retval = sched_setattr(p, &attr); 3768 rcu_read_unlock(); 3769 3770 return retval; 3771} 3772 3773/** 3774 * sys_sched_getscheduler - get the policy (scheduling class) of a thread 3775 * @pid: the pid in question. 3776 * 3777 * Return: On success, the policy of the thread. Otherwise, a negative error 3778 * code. 3779 */ 3780SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid) 3781{ 3782 struct task_struct *p; 3783 int retval; 3784 3785 if (pid < 0) 3786 return -EINVAL; 3787 3788 retval = -ESRCH; 3789 rcu_read_lock(); 3790 p = find_process_by_pid(pid); 3791 if (p) { 3792 retval = security_task_getscheduler(p); 3793 if (!retval) 3794 retval = p->policy 3795 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0); 3796 } 3797 rcu_read_unlock(); 3798 return retval; 3799} 3800 3801/** 3802 * sys_sched_getparam - get the RT priority of a thread 3803 * @pid: the pid in question. 3804 * @param: structure containing the RT priority. 3805 * 3806 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error 3807 * code. 3808 */ 3809SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param) 3810{ 3811 struct sched_param lp = { .sched_priority = 0 }; 3812 struct task_struct *p; 3813 int retval; 3814 3815 if (!param || pid < 0) 3816 return -EINVAL; 3817 3818 rcu_read_lock(); 3819 p = find_process_by_pid(pid); 3820 retval = -ESRCH; 3821 if (!p) 3822 goto out_unlock; 3823 3824 retval = security_task_getscheduler(p); 3825 if (retval) 3826 goto out_unlock; 3827 3828 if (task_has_rt_policy(p)) 3829 lp.sched_priority = p->rt_priority; 3830 rcu_read_unlock(); 3831 3832 /* 3833 * This one might sleep, we cannot do it with a spinlock held ... 3834 */ 3835 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0; 3836 3837 return retval; 3838 3839out_unlock: 3840 rcu_read_unlock(); 3841 return retval; 3842} 3843 3844static int sched_read_attr(struct sched_attr __user *uattr, 3845 struct sched_attr *attr, 3846 unsigned int usize) 3847{ 3848 int ret; 3849 3850 if (!access_ok(VERIFY_WRITE, uattr, usize)) 3851 return -EFAULT; 3852 3853 /* 3854 * If we're handed a smaller struct than we know of, 3855 * ensure all the unknown bits are 0 - i.e. old 3856 * user-space does not get uncomplete information. 3857 */ 3858 if (usize < sizeof(*attr)) { 3859 unsigned char *addr; 3860 unsigned char *end; 3861 3862 addr = (void *)attr + usize; 3863 end = (void *)attr + sizeof(*attr); 3864 3865 for (; addr < end; addr++) { 3866 if (*addr) 3867 return -EFBIG; 3868 } 3869 3870 attr->size = usize; 3871 } 3872 3873 ret = copy_to_user(uattr, attr, attr->size); 3874 if (ret) 3875 return -EFAULT; 3876 3877 return 0; 3878} 3879 3880/** 3881 * sys_sched_getattr - similar to sched_getparam, but with sched_attr 3882 * @pid: the pid in question. 3883 * @uattr: structure containing the extended parameters. 3884 * @size: sizeof(attr) for fwd/bwd comp. 3885 * @flags: for future extension. 3886 */ 3887SYSCALL_DEFINE4(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr, 3888 unsigned int, size, unsigned int, flags) 3889{ 3890 struct sched_attr attr = { 3891 .size = sizeof(struct sched_attr), 3892 }; 3893 struct task_struct *p; 3894 int retval; 3895 3896 if (!uattr || pid < 0 || size > PAGE_SIZE || 3897 size < SCHED_ATTR_SIZE_VER0 || flags) 3898 return -EINVAL; 3899 3900 rcu_read_lock(); 3901 p = find_process_by_pid(pid); 3902 retval = -ESRCH; 3903 if (!p) 3904 goto out_unlock; 3905 3906 retval = security_task_getscheduler(p); 3907 if (retval) 3908 goto out_unlock; 3909 3910 attr.sched_policy = p->policy; 3911 if (p->sched_reset_on_fork) 3912 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK; 3913 if (task_has_dl_policy(p)) 3914 __getparam_dl(p, &attr); 3915 else if (task_has_rt_policy(p)) 3916 attr.sched_priority = p->rt_priority; 3917 else 3918 attr.sched_nice = task_nice(p); 3919 3920 rcu_read_unlock(); 3921 3922 retval = sched_read_attr(uattr, &attr, size); 3923 return retval; 3924 3925out_unlock: 3926 rcu_read_unlock(); 3927 return retval; 3928} 3929 3930long sched_setaffinity(pid_t pid, const struct cpumask *in_mask) 3931{ 3932 cpumask_var_t cpus_allowed, new_mask; 3933 struct task_struct *p; 3934 int retval; 3935 3936 rcu_read_lock(); 3937 3938 p = find_process_by_pid(pid); 3939 if (!p) { 3940 rcu_read_unlock(); 3941 return -ESRCH; 3942 } 3943 3944 /* Prevent p going away */ 3945 get_task_struct(p); 3946 rcu_read_unlock(); 3947 3948 if (p->flags & PF_NO_SETAFFINITY) { 3949 retval = -EINVAL; 3950 goto out_put_task; 3951 } 3952 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) { 3953 retval = -ENOMEM; 3954 goto out_put_task; 3955 } 3956 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) { 3957 retval = -ENOMEM; 3958 goto out_free_cpus_allowed; 3959 } 3960 retval = -EPERM; 3961 if (!check_same_owner(p)) { 3962 rcu_read_lock(); 3963 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) { 3964 rcu_read_unlock(); 3965 goto out_unlock; 3966 } 3967 rcu_read_unlock(); 3968 } 3969 3970 retval = security_task_setscheduler(p); 3971 if (retval) 3972 goto out_unlock; 3973 3974 3975 cpuset_cpus_allowed(p, cpus_allowed); 3976 cpumask_and(new_mask, in_mask, cpus_allowed); 3977 3978 /* 3979 * Since bandwidth control happens on root_domain basis, 3980 * if admission test is enabled, we only admit -deadline 3981 * tasks allowed to run on all the CPUs in the task's 3982 * root_domain. 3983 */ 3984#ifdef CONFIG_SMP 3985 if (task_has_dl_policy(p)) { 3986 const struct cpumask *span = task_rq(p)->rd->span; 3987 3988 if (dl_bandwidth_enabled() && !cpumask_subset(span, new_mask)) { 3989 retval = -EBUSY; 3990 goto out_unlock; 3991 } 3992 } 3993#endif 3994again: 3995 retval = set_cpus_allowed_ptr(p, new_mask); 3996 3997 if (!retval) { 3998 cpuset_cpus_allowed(p, cpus_allowed); 3999 if (!cpumask_subset(new_mask, cpus_allowed)) { 4000 /* 4001 * We must have raced with a concurrent cpuset 4002 * update. Just reset the cpus_allowed to the 4003 * cpuset's cpus_allowed 4004 */ 4005 cpumask_copy(new_mask, cpus_allowed); 4006 goto again; 4007 } 4008 } 4009out_unlock: 4010 free_cpumask_var(new_mask); 4011out_free_cpus_allowed: 4012 free_cpumask_var(cpus_allowed); 4013out_put_task: 4014 put_task_struct(p); 4015 return retval; 4016} 4017 4018static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len, 4019 struct cpumask *new_mask) 4020{ 4021 if (len < cpumask_size()) 4022 cpumask_clear(new_mask); 4023 else if (len > cpumask_size()) 4024 len = cpumask_size(); 4025 4026 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0; 4027} 4028 4029/** 4030 * sys_sched_setaffinity - set the cpu affinity of a process 4031 * @pid: pid of the process 4032 * @len: length in bytes of the bitmask pointed to by user_mask_ptr 4033 * @user_mask_ptr: user-space pointer to the new cpu mask 4034 * 4035 * Return: 0 on success. An error code otherwise. 4036 */ 4037SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len, 4038 unsigned long __user *, user_mask_ptr) 4039{ 4040 cpumask_var_t new_mask; 4041 int retval; 4042 4043 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) 4044 return -ENOMEM; 4045 4046 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask); 4047 if (retval == 0) 4048 retval = sched_setaffinity(pid, new_mask); 4049 free_cpumask_var(new_mask); 4050 return retval; 4051} 4052 4053long sched_getaffinity(pid_t pid, struct cpumask *mask) 4054{ 4055 struct task_struct *p; 4056 unsigned long flags; 4057 int retval; 4058 4059 rcu_read_lock(); 4060 4061 retval = -ESRCH; 4062 p = find_process_by_pid(pid); 4063 if (!p) 4064 goto out_unlock; 4065 4066 retval = security_task_getscheduler(p); 4067 if (retval) 4068 goto out_unlock; 4069 4070 raw_spin_lock_irqsave(&p->pi_lock, flags); 4071 cpumask_and(mask, &p->cpus_allowed, cpu_active_mask); 4072 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 4073 4074out_unlock: 4075 rcu_read_unlock(); 4076 4077 return retval; 4078} 4079 4080/** 4081 * sys_sched_getaffinity - get the cpu affinity of a process 4082 * @pid: pid of the process 4083 * @len: length in bytes of the bitmask pointed to by user_mask_ptr 4084 * @user_mask_ptr: user-space pointer to hold the current cpu mask 4085 * 4086 * Return: 0 on success. An error code otherwise. 4087 */ 4088SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len, 4089 unsigned long __user *, user_mask_ptr) 4090{ 4091 int ret; 4092 cpumask_var_t mask; 4093 4094 if ((len * BITS_PER_BYTE) < nr_cpu_ids) 4095 return -EINVAL; 4096 if (len & (sizeof(unsigned long)-1)) 4097 return -EINVAL; 4098 4099 if (!alloc_cpumask_var(&mask, GFP_KERNEL)) 4100 return -ENOMEM; 4101 4102 ret = sched_getaffinity(pid, mask); 4103 if (ret == 0) { 4104 size_t retlen = min_t(size_t, len, cpumask_size()); 4105 4106 if (copy_to_user(user_mask_ptr, mask, retlen)) 4107 ret = -EFAULT; 4108 else 4109 ret = retlen; 4110 } 4111 free_cpumask_var(mask); 4112 4113 return ret; 4114} 4115 4116/** 4117 * sys_sched_yield - yield the current processor to other threads. 4118 * 4119 * This function yields the current CPU to other tasks. If there are no 4120 * other threads running on this CPU then this function will return. 4121 * 4122 * Return: 0. 4123 */ 4124SYSCALL_DEFINE0(sched_yield) 4125{ 4126 struct rq *rq = this_rq_lock(); 4127 4128 schedstat_inc(rq, yld_count); 4129 current->sched_class->yield_task(rq); 4130 4131 /* 4132 * Since we are going to call schedule() anyway, there's 4133 * no need to preempt or enable interrupts: 4134 */ 4135 __release(rq->lock); 4136 spin_release(&rq->lock.dep_map, 1, _THIS_IP_); 4137 do_raw_spin_unlock(&rq->lock); 4138 sched_preempt_enable_no_resched(); 4139 4140 schedule(); 4141 4142 return 0; 4143} 4144 4145static void __cond_resched(void) 4146{ 4147 __preempt_count_add(PREEMPT_ACTIVE); 4148 __schedule(); 4149 __preempt_count_sub(PREEMPT_ACTIVE); 4150} 4151 4152int __sched _cond_resched(void) 4153{ 4154 rcu_cond_resched(); 4155 if (should_resched()) { 4156 __cond_resched(); 4157 return 1; 4158 } 4159 return 0; 4160} 4161EXPORT_SYMBOL(_cond_resched); 4162 4163/* 4164 * __cond_resched_lock() - if a reschedule is pending, drop the given lock, 4165 * call schedule, and on return reacquire the lock. 4166 * 4167 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level 4168 * operations here to prevent schedule() from being called twice (once via 4169 * spin_unlock(), once by hand). 4170 */ 4171int __cond_resched_lock(spinlock_t *lock) 4172{ 4173 bool need_rcu_resched = rcu_should_resched(); 4174 int resched = should_resched(); 4175 int ret = 0; 4176 4177 lockdep_assert_held(lock); 4178 4179 if (spin_needbreak(lock) || resched || need_rcu_resched) { 4180 spin_unlock(lock); 4181 if (resched) 4182 __cond_resched(); 4183 else if (unlikely(need_rcu_resched)) 4184 rcu_resched(); 4185 else 4186 cpu_relax(); 4187 ret = 1; 4188 spin_lock(lock); 4189 } 4190 return ret; 4191} 4192EXPORT_SYMBOL(__cond_resched_lock); 4193 4194int __sched __cond_resched_softirq(void) 4195{ 4196 BUG_ON(!in_softirq()); 4197 4198 rcu_cond_resched(); /* BH disabled OK, just recording QSes. */ 4199 if (should_resched()) { 4200 local_bh_enable(); 4201 __cond_resched(); 4202 local_bh_disable(); 4203 return 1; 4204 } 4205 return 0; 4206} 4207EXPORT_SYMBOL(__cond_resched_softirq); 4208 4209/** 4210 * yield - yield the current processor to other threads. 4211 * 4212 * Do not ever use this function, there's a 99% chance you're doing it wrong. 4213 * 4214 * The scheduler is at all times free to pick the calling task as the most 4215 * eligible task to run, if removing the yield() call from your code breaks 4216 * it, its already broken. 4217 * 4218 * Typical broken usage is: 4219 * 4220 * while (!event) 4221 * yield(); 4222 * 4223 * where one assumes that yield() will let 'the other' process run that will 4224 * make event true. If the current task is a SCHED_FIFO task that will never 4225 * happen. Never use yield() as a progress guarantee!! 4226 * 4227 * If you want to use yield() to wait for something, use wait_event(). 4228 * If you want to use yield() to be 'nice' for others, use cond_resched(). 4229 * If you still want to use yield(), do not! 4230 */ 4231void __sched yield(void) 4232{ 4233 set_current_state(TASK_RUNNING); 4234 sys_sched_yield(); 4235} 4236EXPORT_SYMBOL(yield); 4237 4238/** 4239 * yield_to - yield the current processor to another thread in 4240 * your thread group, or accelerate that thread toward the 4241 * processor it's on. 4242 * @p: target task 4243 * @preempt: whether task preemption is allowed or not 4244 * 4245 * It's the caller's job to ensure that the target task struct 4246 * can't go away on us before we can do any checks. 4247 * 4248 * Return: 4249 * true (>0) if we indeed boosted the target task. 4250 * false (0) if we failed to boost the target. 4251 * -ESRCH if there's no task to yield to. 4252 */ 4253int __sched yield_to(struct task_struct *p, bool preempt) 4254{ 4255 struct task_struct *curr = current; 4256 struct rq *rq, *p_rq; 4257 unsigned long flags; 4258 int yielded = 0; 4259 4260 local_irq_save(flags); 4261 rq = this_rq(); 4262 4263again: 4264 p_rq = task_rq(p); 4265 /* 4266 * If we're the only runnable task on the rq and target rq also 4267 * has only one task, there's absolutely no point in yielding. 4268 */ 4269 if (rq->nr_running == 1 && p_rq->nr_running == 1) { 4270 yielded = -ESRCH; 4271 goto out_irq; 4272 } 4273 4274 double_rq_lock(rq, p_rq); 4275 if (task_rq(p) != p_rq) { 4276 double_rq_unlock(rq, p_rq); 4277 goto again; 4278 } 4279 4280 if (!curr->sched_class->yield_to_task) 4281 goto out_unlock; 4282 4283 if (curr->sched_class != p->sched_class) 4284 goto out_unlock; 4285 4286 if (task_running(p_rq, p) || p->state) 4287 goto out_unlock; 4288 4289 yielded = curr->sched_class->yield_to_task(rq, p, preempt); 4290 if (yielded) { 4291 schedstat_inc(rq, yld_count); 4292 /* 4293 * Make p's CPU reschedule; pick_next_entity takes care of 4294 * fairness. 4295 */ 4296 if (preempt && rq != p_rq) 4297 resched_task(p_rq->curr); 4298 } 4299 4300out_unlock: 4301 double_rq_unlock(rq, p_rq); 4302out_irq: 4303 local_irq_restore(flags); 4304 4305 if (yielded > 0) 4306 schedule(); 4307 4308 return yielded; 4309} 4310EXPORT_SYMBOL_GPL(yield_to); 4311 4312/* 4313 * This task is about to go to sleep on IO. Increment rq->nr_iowait so 4314 * that process accounting knows that this is a task in IO wait state. 4315 */ 4316void __sched io_schedule(void) 4317{ 4318 struct rq *rq = raw_rq(); 4319 4320 delayacct_blkio_start(); 4321 atomic_inc(&rq->nr_iowait); 4322 blk_flush_plug(current); 4323 current->in_iowait = 1; 4324 schedule(); 4325 current->in_iowait = 0; 4326 atomic_dec(&rq->nr_iowait); 4327 delayacct_blkio_end(); 4328} 4329EXPORT_SYMBOL(io_schedule); 4330 4331long __sched io_schedule_timeout(long timeout) 4332{ 4333 struct rq *rq = raw_rq(); 4334 long ret; 4335 4336 delayacct_blkio_start(); 4337 atomic_inc(&rq->nr_iowait); 4338 blk_flush_plug(current); 4339 current->in_iowait = 1; 4340 ret = schedule_timeout(timeout); 4341 current->in_iowait = 0; 4342 atomic_dec(&rq->nr_iowait); 4343 delayacct_blkio_end(); 4344 return ret; 4345} 4346 4347/** 4348 * sys_sched_get_priority_max - return maximum RT priority. 4349 * @policy: scheduling class. 4350 * 4351 * Return: On success, this syscall returns the maximum 4352 * rt_priority that can be used by a given scheduling class. 4353 * On failure, a negative error code is returned. 4354 */ 4355SYSCALL_DEFINE1(sched_get_priority_max, int, policy) 4356{ 4357 int ret = -EINVAL; 4358 4359 switch (policy) { 4360 case SCHED_FIFO: 4361 case SCHED_RR: 4362 ret = MAX_USER_RT_PRIO-1; 4363 break; 4364 case SCHED_DEADLINE: 4365 case SCHED_NORMAL: 4366 case SCHED_BATCH: 4367 case SCHED_IDLE: 4368 ret = 0; 4369 break; 4370 } 4371 return ret; 4372} 4373 4374/** 4375 * sys_sched_get_priority_min - return minimum RT priority. 4376 * @policy: scheduling class. 4377 * 4378 * Return: On success, this syscall returns the minimum 4379 * rt_priority that can be used by a given scheduling class. 4380 * On failure, a negative error code is returned. 4381 */ 4382SYSCALL_DEFINE1(sched_get_priority_min, int, policy) 4383{ 4384 int ret = -EINVAL; 4385 4386 switch (policy) { 4387 case SCHED_FIFO: 4388 case SCHED_RR: 4389 ret = 1; 4390 break; 4391 case SCHED_DEADLINE: 4392 case SCHED_NORMAL: 4393 case SCHED_BATCH: 4394 case SCHED_IDLE: 4395 ret = 0; 4396 } 4397 return ret; 4398} 4399 4400/** 4401 * sys_sched_rr_get_interval - return the default timeslice of a process. 4402 * @pid: pid of the process. 4403 * @interval: userspace pointer to the timeslice value. 4404 * 4405 * this syscall writes the default timeslice value of a given process 4406 * into the user-space timespec buffer. A value of '0' means infinity. 4407 * 4408 * Return: On success, 0 and the timeslice is in @interval. Otherwise, 4409 * an error code. 4410 */ 4411SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid, 4412 struct timespec __user *, interval) 4413{ 4414 struct task_struct *p; 4415 unsigned int time_slice; 4416 unsigned long flags; 4417 struct rq *rq; 4418 int retval; 4419 struct timespec t; 4420 4421 if (pid < 0) 4422 return -EINVAL; 4423 4424 retval = -ESRCH; 4425 rcu_read_lock(); 4426 p = find_process_by_pid(pid); 4427 if (!p) 4428 goto out_unlock; 4429 4430 retval = security_task_getscheduler(p); 4431 if (retval) 4432 goto out_unlock; 4433 4434 rq = task_rq_lock(p, &flags); 4435 time_slice = 0; 4436 if (p->sched_class->get_rr_interval) 4437 time_slice = p->sched_class->get_rr_interval(rq, p); 4438 task_rq_unlock(rq, p, &flags); 4439 4440 rcu_read_unlock(); 4441 jiffies_to_timespec(time_slice, &t); 4442 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0; 4443 return retval; 4444 4445out_unlock: 4446 rcu_read_unlock(); 4447 return retval; 4448} 4449 4450static const char stat_nam[] = TASK_STATE_TO_CHAR_STR; 4451 4452void sched_show_task(struct task_struct *p) 4453{ 4454 unsigned long free = 0; 4455 int ppid; 4456 unsigned state; 4457 4458 state = p->state ? __ffs(p->state) + 1 : 0; 4459 printk(KERN_INFO "%-15.15s %c", p->comm, 4460 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?'); 4461#if BITS_PER_LONG == 32 4462 if (state == TASK_RUNNING) 4463 printk(KERN_CONT " running "); 4464 else 4465 printk(KERN_CONT " %08lx ", thread_saved_pc(p)); 4466#else 4467 if (state == TASK_RUNNING) 4468 printk(KERN_CONT " running task "); 4469 else 4470 printk(KERN_CONT " %016lx ", thread_saved_pc(p)); 4471#endif 4472#ifdef CONFIG_DEBUG_STACK_USAGE 4473 free = stack_not_used(p); 4474#endif 4475 rcu_read_lock(); 4476 ppid = task_pid_nr(rcu_dereference(p->real_parent)); 4477 rcu_read_unlock(); 4478 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free, 4479 task_pid_nr(p), ppid, 4480 (unsigned long)task_thread_info(p)->flags); 4481 4482 print_worker_info(KERN_INFO, p); 4483 show_stack(p, NULL); 4484} 4485 4486void show_state_filter(unsigned long state_filter) 4487{ 4488 struct task_struct *g, *p; 4489 4490#if BITS_PER_LONG == 32 4491 printk(KERN_INFO 4492 " task PC stack pid father\n"); 4493#else 4494 printk(KERN_INFO 4495 " task PC stack pid father\n"); 4496#endif 4497 rcu_read_lock(); 4498 do_each_thread(g, p) { 4499 /* 4500 * reset the NMI-timeout, listing all files on a slow 4501 * console might take a lot of time: 4502 */ 4503 touch_nmi_watchdog(); 4504 if (!state_filter || (p->state & state_filter)) 4505 sched_show_task(p); 4506 } while_each_thread(g, p); 4507 4508 touch_all_softlockup_watchdogs(); 4509 4510#ifdef CONFIG_SCHED_DEBUG 4511 sysrq_sched_debug_show(); 4512#endif 4513 rcu_read_unlock(); 4514 /* 4515 * Only show locks if all tasks are dumped: 4516 */ 4517 if (!state_filter) 4518 debug_show_all_locks(); 4519} 4520 4521void init_idle_bootup_task(struct task_struct *idle) 4522{ 4523 idle->sched_class = &idle_sched_class; 4524} 4525 4526/** 4527 * init_idle - set up an idle thread for a given CPU 4528 * @idle: task in question 4529 * @cpu: cpu the idle task belongs to 4530 * 4531 * NOTE: this function does not set the idle thread's NEED_RESCHED 4532 * flag, to make booting more robust. 4533 */ 4534void init_idle(struct task_struct *idle, int cpu) 4535{ 4536 struct rq *rq = cpu_rq(cpu); 4537 unsigned long flags; 4538 4539 raw_spin_lock_irqsave(&rq->lock, flags); 4540 4541 __sched_fork(0, idle); 4542 idle->state = TASK_RUNNING; 4543 idle->se.exec_start = sched_clock(); 4544 4545 do_set_cpus_allowed(idle, cpumask_of(cpu)); 4546 /* 4547 * We're having a chicken and egg problem, even though we are 4548 * holding rq->lock, the cpu isn't yet set to this cpu so the 4549 * lockdep check in task_group() will fail. 4550 * 4551 * Similar case to sched_fork(). / Alternatively we could 4552 * use task_rq_lock() here and obtain the other rq->lock. 4553 * 4554 * Silence PROVE_RCU 4555 */ 4556 rcu_read_lock(); 4557 __set_task_cpu(idle, cpu); 4558 rcu_read_unlock(); 4559 4560 rq->curr = rq->idle = idle; 4561 idle->on_rq = 1; 4562#if defined(CONFIG_SMP) 4563 idle->on_cpu = 1; 4564#endif 4565 raw_spin_unlock_irqrestore(&rq->lock, flags); 4566 4567 /* Set the preempt count _outside_ the spinlocks! */ 4568 init_idle_preempt_count(idle, cpu); 4569 4570 /* 4571 * The idle tasks have their own, simple scheduling class: 4572 */ 4573 idle->sched_class = &idle_sched_class; 4574 ftrace_graph_init_idle_task(idle, cpu); 4575 vtime_init_idle(idle, cpu); 4576#if defined(CONFIG_SMP) 4577 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu); 4578#endif 4579} 4580 4581#ifdef CONFIG_SMP 4582void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask) 4583{ 4584 if (p->sched_class && p->sched_class->set_cpus_allowed) 4585 p->sched_class->set_cpus_allowed(p, new_mask); 4586 4587 cpumask_copy(&p->cpus_allowed, new_mask); 4588 p->nr_cpus_allowed = cpumask_weight(new_mask); 4589} 4590 4591/* 4592 * This is how migration works: 4593 * 4594 * 1) we invoke migration_cpu_stop() on the target CPU using 4595 * stop_one_cpu(). 4596 * 2) stopper starts to run (implicitly forcing the migrated thread 4597 * off the CPU) 4598 * 3) it checks whether the migrated task is still in the wrong runqueue. 4599 * 4) if it's in the wrong runqueue then the migration thread removes 4600 * it and puts it into the right queue. 4601 * 5) stopper completes and stop_one_cpu() returns and the migration 4602 * is done. 4603 */ 4604 4605/* 4606 * Change a given task's CPU affinity. Migrate the thread to a 4607 * proper CPU and schedule it away if the CPU it's executing on 4608 * is removed from the allowed bitmask. 4609 * 4610 * NOTE: the caller must have a valid reference to the task, the 4611 * task must not exit() & deallocate itself prematurely. The 4612 * call is not atomic; no spinlocks may be held. 4613 */ 4614int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask) 4615{ 4616 unsigned long flags; 4617 struct rq *rq; 4618 unsigned int dest_cpu; 4619 int ret = 0; 4620 4621 rq = task_rq_lock(p, &flags); 4622 4623 if (cpumask_equal(&p->cpus_allowed, new_mask)) 4624 goto out; 4625 4626 if (!cpumask_intersects(new_mask, cpu_active_mask)) { 4627 ret = -EINVAL; 4628 goto out; 4629 } 4630 4631 do_set_cpus_allowed(p, new_mask); 4632 4633 /* Can the task run on the task's current CPU? If so, we're done */ 4634 if (cpumask_test_cpu(task_cpu(p), new_mask)) 4635 goto out; 4636 4637 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask); 4638 if (p->on_rq) { 4639 struct migration_arg arg = { p, dest_cpu }; 4640 /* Need help from migration thread: drop lock and wait. */ 4641 task_rq_unlock(rq, p, &flags); 4642 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg); 4643 tlb_migrate_finish(p->mm); 4644 return 0; 4645 } 4646out: 4647 task_rq_unlock(rq, p, &flags); 4648 4649 return ret; 4650} 4651EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr); 4652 4653/* 4654 * Move (not current) task off this cpu, onto dest cpu. We're doing 4655 * this because either it can't run here any more (set_cpus_allowed() 4656 * away from this CPU, or CPU going down), or because we're 4657 * attempting to rebalance this task on exec (sched_exec). 4658 * 4659 * So we race with normal scheduler movements, but that's OK, as long 4660 * as the task is no longer on this CPU. 4661 * 4662 * Returns non-zero if task was successfully migrated. 4663 */ 4664static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu) 4665{ 4666 struct rq *rq_dest, *rq_src; 4667 int ret = 0; 4668 4669 if (unlikely(!cpu_active(dest_cpu))) 4670 return ret; 4671 4672 rq_src = cpu_rq(src_cpu); 4673 rq_dest = cpu_rq(dest_cpu); 4674 4675 raw_spin_lock(&p->pi_lock); 4676 double_rq_lock(rq_src, rq_dest); 4677 /* Already moved. */ 4678 if (task_cpu(p) != src_cpu) 4679 goto done; 4680 /* Affinity changed (again). */ 4681 if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p))) 4682 goto fail; 4683 4684 /* 4685 * If we're not on a rq, the next wake-up will ensure we're 4686 * placed properly. 4687 */ 4688 if (p->on_rq) { 4689 dequeue_task(rq_src, p, 0); 4690 set_task_cpu(p, dest_cpu); 4691 enqueue_task(rq_dest, p, 0); 4692 check_preempt_curr(rq_dest, p, 0); 4693 } 4694done: 4695 ret = 1; 4696fail: 4697 double_rq_unlock(rq_src, rq_dest); 4698 raw_spin_unlock(&p->pi_lock); 4699 return ret; 4700} 4701 4702#ifdef CONFIG_NUMA_BALANCING 4703/* Migrate current task p to target_cpu */ 4704int migrate_task_to(struct task_struct *p, int target_cpu) 4705{ 4706 struct migration_arg arg = { p, target_cpu }; 4707 int curr_cpu = task_cpu(p); 4708 4709 if (curr_cpu == target_cpu) 4710 return 0; 4711 4712 if (!cpumask_test_cpu(target_cpu, tsk_cpus_allowed(p))) 4713 return -EINVAL; 4714 4715 /* TODO: This is not properly updating schedstats */ 4716 4717 trace_sched_move_numa(p, curr_cpu, target_cpu); 4718 return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg); 4719} 4720 4721/* 4722 * Requeue a task on a given node and accurately track the number of NUMA 4723 * tasks on the runqueues 4724 */ 4725void sched_setnuma(struct task_struct *p, int nid) 4726{ 4727 struct rq *rq; 4728 unsigned long flags; 4729 bool on_rq, running; 4730 4731 rq = task_rq_lock(p, &flags); 4732 on_rq = p->on_rq; 4733 running = task_current(rq, p); 4734 4735 if (on_rq) 4736 dequeue_task(rq, p, 0); 4737 if (running) 4738 p->sched_class->put_prev_task(rq, p); 4739 4740 p->numa_preferred_nid = nid; 4741 4742 if (running) 4743 p->sched_class->set_curr_task(rq); 4744 if (on_rq) 4745 enqueue_task(rq, p, 0); 4746 task_rq_unlock(rq, p, &flags); 4747} 4748#endif 4749 4750/* 4751 * migration_cpu_stop - this will be executed by a highprio stopper thread 4752 * and performs thread migration by bumping thread off CPU then 4753 * 'pushing' onto another runqueue. 4754 */ 4755static int migration_cpu_stop(void *data) 4756{ 4757 struct migration_arg *arg = data; 4758 4759 /* 4760 * The original target cpu might have gone down and we might 4761 * be on another cpu but it doesn't matter. 4762 */ 4763 local_irq_disable(); 4764 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu); 4765 local_irq_enable(); 4766 return 0; 4767} 4768 4769#ifdef CONFIG_HOTPLUG_CPU 4770 4771/* 4772 * Ensures that the idle task is using init_mm right before its cpu goes 4773 * offline. 4774 */ 4775void idle_task_exit(void) 4776{ 4777 struct mm_struct *mm = current->active_mm; 4778 4779 BUG_ON(cpu_online(smp_processor_id())); 4780 4781 if (mm != &init_mm) { 4782 switch_mm(mm, &init_mm, current); 4783 finish_arch_post_lock_switch(); 4784 } 4785 mmdrop(mm); 4786} 4787 4788/* 4789 * Since this CPU is going 'away' for a while, fold any nr_active delta 4790 * we might have. Assumes we're called after migrate_tasks() so that the 4791 * nr_active count is stable. 4792 * 4793 * Also see the comment "Global load-average calculations". 4794 */ 4795static void calc_load_migrate(struct rq *rq) 4796{ 4797 long delta = calc_load_fold_active(rq); 4798 if (delta) 4799 atomic_long_add(delta, &calc_load_tasks); 4800} 4801 4802static void put_prev_task_fake(struct rq *rq, struct task_struct *prev) 4803{ 4804} 4805 4806static const struct sched_class fake_sched_class = { 4807 .put_prev_task = put_prev_task_fake, 4808}; 4809 4810static struct task_struct fake_task = { 4811 /* 4812 * Avoid pull_{rt,dl}_task() 4813 */ 4814 .prio = MAX_PRIO + 1, 4815 .sched_class = &fake_sched_class, 4816}; 4817 4818/* 4819 * Migrate all tasks from the rq, sleeping tasks will be migrated by 4820 * try_to_wake_up()->select_task_rq(). 4821 * 4822 * Called with rq->lock held even though we'er in stop_machine() and 4823 * there's no concurrency possible, we hold the required locks anyway 4824 * because of lock validation efforts. 4825 */ 4826static void migrate_tasks(unsigned int dead_cpu) 4827{ 4828 struct rq *rq = cpu_rq(dead_cpu); 4829 struct task_struct *next, *stop = rq->stop; 4830 int dest_cpu; 4831 4832 /* 4833 * Fudge the rq selection such that the below task selection loop 4834 * doesn't get stuck on the currently eligible stop task. 4835 * 4836 * We're currently inside stop_machine() and the rq is either stuck 4837 * in the stop_machine_cpu_stop() loop, or we're executing this code, 4838 * either way we should never end up calling schedule() until we're 4839 * done here. 4840 */ 4841 rq->stop = NULL; 4842 4843 /* 4844 * put_prev_task() and pick_next_task() sched 4845 * class method both need to have an up-to-date 4846 * value of rq->clock[_task] 4847 */ 4848 update_rq_clock(rq); 4849 4850 for ( ; ; ) { 4851 /* 4852 * There's this thread running, bail when that's the only 4853 * remaining thread. 4854 */ 4855 if (rq->nr_running == 1) 4856 break; 4857 4858 next = pick_next_task(rq, &fake_task); 4859 BUG_ON(!next); 4860 next->sched_class->put_prev_task(rq, next); 4861 4862 /* Find suitable destination for @next, with force if needed. */ 4863 dest_cpu = select_fallback_rq(dead_cpu, next); 4864 raw_spin_unlock(&rq->lock); 4865 4866 __migrate_task(next, dead_cpu, dest_cpu); 4867 4868 raw_spin_lock(&rq->lock); 4869 } 4870 4871 rq->stop = stop; 4872} 4873 4874#endif /* CONFIG_HOTPLUG_CPU */ 4875 4876#if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL) 4877 4878static struct ctl_table sd_ctl_dir[] = { 4879 { 4880 .procname = "sched_domain", 4881 .mode = 0555, 4882 }, 4883 {} 4884}; 4885 4886static struct ctl_table sd_ctl_root[] = { 4887 { 4888 .procname = "kernel", 4889 .mode = 0555, 4890 .child = sd_ctl_dir, 4891 }, 4892 {} 4893}; 4894 4895static struct ctl_table *sd_alloc_ctl_entry(int n) 4896{ 4897 struct ctl_table *entry = 4898 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL); 4899 4900 return entry; 4901} 4902 4903static void sd_free_ctl_entry(struct ctl_table **tablep) 4904{ 4905 struct ctl_table *entry; 4906 4907 /* 4908 * In the intermediate directories, both the child directory and 4909 * procname are dynamically allocated and could fail but the mode 4910 * will always be set. In the lowest directory the names are 4911 * static strings and all have proc handlers. 4912 */ 4913 for (entry = *tablep; entry->mode; entry++) { 4914 if (entry->child) 4915 sd_free_ctl_entry(&entry->child); 4916 if (entry->proc_handler == NULL) 4917 kfree(entry->procname); 4918 } 4919 4920 kfree(*tablep); 4921 *tablep = NULL; 4922} 4923 4924static int min_load_idx = 0; 4925static int max_load_idx = CPU_LOAD_IDX_MAX-1; 4926 4927static void 4928set_table_entry(struct ctl_table *entry, 4929 const char *procname, void *data, int maxlen, 4930 umode_t mode, proc_handler *proc_handler, 4931 bool load_idx) 4932{ 4933 entry->procname = procname; 4934 entry->data = data; 4935 entry->maxlen = maxlen; 4936 entry->mode = mode; 4937 entry->proc_handler = proc_handler; 4938 4939 if (load_idx) { 4940 entry->extra1 = &min_load_idx; 4941 entry->extra2 = &max_load_idx; 4942 } 4943} 4944 4945static struct ctl_table * 4946sd_alloc_ctl_domain_table(struct sched_domain *sd) 4947{ 4948 struct ctl_table *table = sd_alloc_ctl_entry(14); 4949 4950 if (table == NULL) 4951 return NULL; 4952 4953 set_table_entry(&table[0], "min_interval", &sd->min_interval, 4954 sizeof(long), 0644, proc_doulongvec_minmax, false); 4955 set_table_entry(&table[1], "max_interval", &sd->max_interval, 4956 sizeof(long), 0644, proc_doulongvec_minmax, false); 4957 set_table_entry(&table[2], "busy_idx", &sd->busy_idx, 4958 sizeof(int), 0644, proc_dointvec_minmax, true); 4959 set_table_entry(&table[3], "idle_idx", &sd->idle_idx, 4960 sizeof(int), 0644, proc_dointvec_minmax, true); 4961 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx, 4962 sizeof(int), 0644, proc_dointvec_minmax, true); 4963 set_table_entry(&table[5], "wake_idx", &sd->wake_idx, 4964 sizeof(int), 0644, proc_dointvec_minmax, true); 4965 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx, 4966 sizeof(int), 0644, proc_dointvec_minmax, true); 4967 set_table_entry(&table[7], "busy_factor", &sd->busy_factor, 4968 sizeof(int), 0644, proc_dointvec_minmax, false); 4969 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct, 4970 sizeof(int), 0644, proc_dointvec_minmax, false); 4971 set_table_entry(&table[9], "cache_nice_tries", 4972 &sd->cache_nice_tries, 4973 sizeof(int), 0644, proc_dointvec_minmax, false); 4974 set_table_entry(&table[10], "flags", &sd->flags, 4975 sizeof(int), 0644, proc_dointvec_minmax, false); 4976 set_table_entry(&table[11], "max_newidle_lb_cost", 4977 &sd->max_newidle_lb_cost, 4978 sizeof(long), 0644, proc_doulongvec_minmax, false); 4979 set_table_entry(&table[12], "name", sd->name, 4980 CORENAME_MAX_SIZE, 0444, proc_dostring, false); 4981 /* &table[13] is terminator */ 4982 4983 return table; 4984} 4985 4986static struct ctl_table *sd_alloc_ctl_cpu_table(int cpu) 4987{ 4988 struct ctl_table *entry, *table; 4989 struct sched_domain *sd; 4990 int domain_num = 0, i; 4991 char buf[32]; 4992 4993 for_each_domain(cpu, sd) 4994 domain_num++; 4995 entry = table = sd_alloc_ctl_entry(domain_num + 1); 4996 if (table == NULL) 4997 return NULL; 4998 4999 i = 0; 5000 for_each_domain(cpu, sd) { 5001 snprintf(buf, 32, "domain%d", i); 5002 entry->procname = kstrdup(buf, GFP_KERNEL); 5003 entry->mode = 0555; 5004 entry->child = sd_alloc_ctl_domain_table(sd); 5005 entry++; 5006 i++; 5007 } 5008 return table; 5009} 5010 5011static struct ctl_table_header *sd_sysctl_header; 5012static void register_sched_domain_sysctl(void) 5013{ 5014 int i, cpu_num = num_possible_cpus(); 5015 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1); 5016 char buf[32]; 5017 5018 WARN_ON(sd_ctl_dir[0].child); 5019 sd_ctl_dir[0].child = entry; 5020 5021 if (entry == NULL) 5022 return; 5023 5024 for_each_possible_cpu(i) { 5025 snprintf(buf, 32, "cpu%d", i); 5026 entry->procname = kstrdup(buf, GFP_KERNEL); 5027 entry->mode = 0555; 5028 entry->child = sd_alloc_ctl_cpu_table(i); 5029 entry++; 5030 } 5031 5032 WARN_ON(sd_sysctl_header); 5033 sd_sysctl_header = register_sysctl_table(sd_ctl_root); 5034} 5035 5036/* may be called multiple times per register */ 5037static void unregister_sched_domain_sysctl(void) 5038{ 5039 if (sd_sysctl_header) 5040 unregister_sysctl_table(sd_sysctl_header); 5041 sd_sysctl_header = NULL; 5042 if (sd_ctl_dir[0].child) 5043 sd_free_ctl_entry(&sd_ctl_dir[0].child); 5044} 5045#else 5046static void register_sched_domain_sysctl(void) 5047{ 5048} 5049static void unregister_sched_domain_sysctl(void) 5050{ 5051} 5052#endif 5053 5054static void set_rq_online(struct rq *rq) 5055{ 5056 if (!rq->online) { 5057 const struct sched_class *class; 5058 5059 cpumask_set_cpu(rq->cpu, rq->rd->online); 5060 rq->online = 1; 5061 5062 for_each_class(class) { 5063 if (class->rq_online) 5064 class->rq_online(rq); 5065 } 5066 } 5067} 5068 5069static void set_rq_offline(struct rq *rq) 5070{ 5071 if (rq->online) { 5072 const struct sched_class *class; 5073 5074 for_each_class(class) { 5075 if (class->rq_offline) 5076 class->rq_offline(rq); 5077 } 5078 5079 cpumask_clear_cpu(rq->cpu, rq->rd->online); 5080 rq->online = 0; 5081 } 5082} 5083 5084/* 5085 * migration_call - callback that gets triggered when a CPU is added. 5086 * Here we can start up the necessary migration thread for the new CPU. 5087 */ 5088static int 5089migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu) 5090{ 5091 int cpu = (long)hcpu; 5092 unsigned long flags; 5093 struct rq *rq = cpu_rq(cpu); 5094 5095 switch (action & ~CPU_TASKS_FROZEN) { 5096 5097 case CPU_UP_PREPARE: 5098 rq->calc_load_update = calc_load_update; 5099 break; 5100 5101 case CPU_ONLINE: 5102 /* Update our root-domain */ 5103 raw_spin_lock_irqsave(&rq->lock, flags); 5104 if (rq->rd) { 5105 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span)); 5106 5107 set_rq_online(rq); 5108 } 5109 raw_spin_unlock_irqrestore(&rq->lock, flags); 5110 break; 5111 5112#ifdef CONFIG_HOTPLUG_CPU 5113 case CPU_DYING: 5114 sched_ttwu_pending(); 5115 /* Update our root-domain */ 5116 raw_spin_lock_irqsave(&rq->lock, flags); 5117 if (rq->rd) { 5118 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span)); 5119 set_rq_offline(rq); 5120 } 5121 migrate_tasks(cpu); 5122 BUG_ON(rq->nr_running != 1); /* the migration thread */ 5123 raw_spin_unlock_irqrestore(&rq->lock, flags); 5124 break; 5125 5126 case CPU_DEAD: 5127 calc_load_migrate(rq); 5128 break; 5129#endif 5130 } 5131 5132 update_max_interval(); 5133 5134 return NOTIFY_OK; 5135} 5136 5137/* 5138 * Register at high priority so that task migration (migrate_all_tasks) 5139 * happens before everything else. This has to be lower priority than 5140 * the notifier in the perf_event subsystem, though. 5141 */ 5142static struct notifier_block migration_notifier = { 5143 .notifier_call = migration_call, 5144 .priority = CPU_PRI_MIGRATION, 5145}; 5146 5147static void __cpuinit set_cpu_rq_start_time(void) 5148{ 5149 int cpu = smp_processor_id(); 5150 struct rq *rq = cpu_rq(cpu); 5151 rq->age_stamp = sched_clock_cpu(cpu); 5152} 5153 5154static int sched_cpu_active(struct notifier_block *nfb, 5155 unsigned long action, void *hcpu) 5156{ 5157 switch (action & ~CPU_TASKS_FROZEN) { 5158 case CPU_STARTING: 5159 set_cpu_rq_start_time(); 5160 return NOTIFY_OK; 5161 case CPU_DOWN_FAILED: 5162 set_cpu_active((long)hcpu, true); 5163 return NOTIFY_OK; 5164 default: 5165 return NOTIFY_DONE; 5166 } 5167} 5168 5169static int sched_cpu_inactive(struct notifier_block *nfb, 5170 unsigned long action, void *hcpu) 5171{ 5172 unsigned long flags; 5173 long cpu = (long)hcpu; 5174 5175 switch (action & ~CPU_TASKS_FROZEN) { 5176 case CPU_DOWN_PREPARE: 5177 set_cpu_active(cpu, false); 5178 5179 /* explicitly allow suspend */ 5180 if (!(action & CPU_TASKS_FROZEN)) { 5181 struct dl_bw *dl_b = dl_bw_of(cpu); 5182 bool overflow; 5183 int cpus; 5184 5185 raw_spin_lock_irqsave(&dl_b->lock, flags); 5186 cpus = dl_bw_cpus(cpu); 5187 overflow = __dl_overflow(dl_b, cpus, 0, 0); 5188 raw_spin_unlock_irqrestore(&dl_b->lock, flags); 5189 5190 if (overflow) 5191 return notifier_from_errno(-EBUSY); 5192 } 5193 return NOTIFY_OK; 5194 } 5195 5196 return NOTIFY_DONE; 5197} 5198 5199static int __init migration_init(void) 5200{ 5201 void *cpu = (void *)(long)smp_processor_id(); 5202 int err; 5203 5204 /* Initialize migration for the boot CPU */ 5205 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu); 5206 BUG_ON(err == NOTIFY_BAD); 5207 migration_call(&migration_notifier, CPU_ONLINE, cpu); 5208 register_cpu_notifier(&migration_notifier); 5209 5210 /* Register cpu active notifiers */ 5211 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE); 5212 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE); 5213 5214 return 0; 5215} 5216early_initcall(migration_init); 5217#endif 5218 5219#ifdef CONFIG_SMP 5220 5221static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */ 5222 5223#ifdef CONFIG_SCHED_DEBUG 5224 5225static __read_mostly int sched_debug_enabled; 5226 5227static int __init sched_debug_setup(char *str) 5228{ 5229 sched_debug_enabled = 1; 5230 5231 return 0; 5232} 5233early_param("sched_debug", sched_debug_setup); 5234 5235static inline bool sched_debug(void) 5236{ 5237 return sched_debug_enabled; 5238} 5239 5240static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level, 5241 struct cpumask *groupmask) 5242{ 5243 struct sched_group *group = sd->groups; 5244 char str[256]; 5245 5246 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd)); 5247 cpumask_clear(groupmask); 5248 5249 printk(KERN_DEBUG "%*s domain %d: ", level, "", level); 5250 5251 if (!(sd->flags & SD_LOAD_BALANCE)) { 5252 printk("does not load-balance\n"); 5253 if (sd->parent) 5254 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain" 5255 " has parent"); 5256 return -1; 5257 } 5258 5259 printk(KERN_CONT "span %s level %s\n", str, sd->name); 5260 5261 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) { 5262 printk(KERN_ERR "ERROR: domain->span does not contain " 5263 "CPU%d\n", cpu); 5264 } 5265 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) { 5266 printk(KERN_ERR "ERROR: domain->groups does not contain" 5267 " CPU%d\n", cpu); 5268 } 5269 5270 printk(KERN_DEBUG "%*s groups:", level + 1, ""); 5271 do { 5272 if (!group) { 5273 printk("\n"); 5274 printk(KERN_ERR "ERROR: group is NULL\n"); 5275 break; 5276 } 5277 5278 /* 5279 * Even though we initialize ->capacity to something semi-sane, 5280 * we leave capacity_orig unset. This allows us to detect if 5281 * domain iteration is still funny without causing /0 traps. 5282 */ 5283 if (!group->sgc->capacity_orig) { 5284 printk(KERN_CONT "\n"); 5285 printk(KERN_ERR "ERROR: domain->cpu_capacity not set\n"); 5286 break; 5287 } 5288 5289 if (!cpumask_weight(sched_group_cpus(group))) { 5290 printk(KERN_CONT "\n"); 5291 printk(KERN_ERR "ERROR: empty group\n"); 5292 break; 5293 } 5294 5295 if (!(sd->flags & SD_OVERLAP) && 5296 cpumask_intersects(groupmask, sched_group_cpus(group))) { 5297 printk(KERN_CONT "\n"); 5298 printk(KERN_ERR "ERROR: repeated CPUs\n"); 5299 break; 5300 } 5301 5302 cpumask_or(groupmask, groupmask, sched_group_cpus(group)); 5303 5304 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group)); 5305 5306 printk(KERN_CONT " %s", str); 5307 if (group->sgc->capacity != SCHED_CAPACITY_SCALE) { 5308 printk(KERN_CONT " (cpu_capacity = %d)", 5309 group->sgc->capacity); 5310 } 5311 5312 group = group->next; 5313 } while (group != sd->groups); 5314 printk(KERN_CONT "\n"); 5315 5316 if (!cpumask_equal(sched_domain_span(sd), groupmask)) 5317 printk(KERN_ERR "ERROR: groups don't span domain->span\n"); 5318 5319 if (sd->parent && 5320 !cpumask_subset(groupmask, sched_domain_span(sd->parent))) 5321 printk(KERN_ERR "ERROR: parent span is not a superset " 5322 "of domain->span\n"); 5323 return 0; 5324} 5325 5326static void sched_domain_debug(struct sched_domain *sd, int cpu) 5327{ 5328 int level = 0; 5329 5330 if (!sched_debug_enabled) 5331 return; 5332 5333 if (!sd) { 5334 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu); 5335 return; 5336 } 5337 5338 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu); 5339 5340 for (;;) { 5341 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask)) 5342 break; 5343 level++; 5344 sd = sd->parent; 5345 if (!sd) 5346 break; 5347 } 5348} 5349#else /* !CONFIG_SCHED_DEBUG */ 5350# define sched_domain_debug(sd, cpu) do { } while (0) 5351static inline bool sched_debug(void) 5352{ 5353 return false; 5354} 5355#endif /* CONFIG_SCHED_DEBUG */ 5356 5357static int sd_degenerate(struct sched_domain *sd) 5358{ 5359 if (cpumask_weight(sched_domain_span(sd)) == 1) 5360 return 1; 5361 5362 /* Following flags need at least 2 groups */ 5363 if (sd->flags & (SD_LOAD_BALANCE | 5364 SD_BALANCE_NEWIDLE | 5365 SD_BALANCE_FORK | 5366 SD_BALANCE_EXEC | 5367 SD_SHARE_CPUCAPACITY | 5368 SD_SHARE_PKG_RESOURCES | 5369 SD_SHARE_POWERDOMAIN)) { 5370 if (sd->groups != sd->groups->next) 5371 return 0; 5372 } 5373 5374 /* Following flags don't use groups */ 5375 if (sd->flags & (SD_WAKE_AFFINE)) 5376 return 0; 5377 5378 return 1; 5379} 5380 5381static int 5382sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent) 5383{ 5384 unsigned long cflags = sd->flags, pflags = parent->flags; 5385 5386 if (sd_degenerate(parent)) 5387 return 1; 5388 5389 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent))) 5390 return 0; 5391 5392 /* Flags needing groups don't count if only 1 group in parent */ 5393 if (parent->groups == parent->groups->next) { 5394 pflags &= ~(SD_LOAD_BALANCE | 5395 SD_BALANCE_NEWIDLE | 5396 SD_BALANCE_FORK | 5397 SD_BALANCE_EXEC | 5398 SD_SHARE_CPUCAPACITY | 5399 SD_SHARE_PKG_RESOURCES | 5400 SD_PREFER_SIBLING | 5401 SD_SHARE_POWERDOMAIN); 5402 if (nr_node_ids == 1) 5403 pflags &= ~SD_SERIALIZE; 5404 } 5405 if (~cflags & pflags) 5406 return 0; 5407 5408 return 1; 5409} 5410 5411static void free_rootdomain(struct rcu_head *rcu) 5412{ 5413 struct root_domain *rd = container_of(rcu, struct root_domain, rcu); 5414 5415 cpupri_cleanup(&rd->cpupri); 5416 cpudl_cleanup(&rd->cpudl); 5417 free_cpumask_var(rd->dlo_mask); 5418 free_cpumask_var(rd->rto_mask); 5419 free_cpumask_var(rd->online); 5420 free_cpumask_var(rd->span); 5421 kfree(rd); 5422} 5423 5424static void rq_attach_root(struct rq *rq, struct root_domain *rd) 5425{ 5426 struct root_domain *old_rd = NULL; 5427 unsigned long flags; 5428 5429 raw_spin_lock_irqsave(&rq->lock, flags); 5430 5431 if (rq->rd) { 5432 old_rd = rq->rd; 5433 5434 if (cpumask_test_cpu(rq->cpu, old_rd->online)) 5435 set_rq_offline(rq); 5436 5437 cpumask_clear_cpu(rq->cpu, old_rd->span); 5438 5439 /* 5440 * If we dont want to free the old_rd yet then 5441 * set old_rd to NULL to skip the freeing later 5442 * in this function: 5443 */ 5444 if (!atomic_dec_and_test(&old_rd->refcount)) 5445 old_rd = NULL; 5446 } 5447 5448 atomic_inc(&rd->refcount); 5449 rq->rd = rd; 5450 5451 cpumask_set_cpu(rq->cpu, rd->span); 5452 if (cpumask_test_cpu(rq->cpu, cpu_active_mask)) 5453 set_rq_online(rq); 5454 5455 raw_spin_unlock_irqrestore(&rq->lock, flags); 5456 5457 if (old_rd) 5458 call_rcu_sched(&old_rd->rcu, free_rootdomain); 5459} 5460 5461static int init_rootdomain(struct root_domain *rd) 5462{ 5463 memset(rd, 0, sizeof(*rd)); 5464 5465 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL)) 5466 goto out; 5467 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL)) 5468 goto free_span; 5469 if (!alloc_cpumask_var(&rd->dlo_mask, GFP_KERNEL)) 5470 goto free_online; 5471 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL)) 5472 goto free_dlo_mask; 5473 5474 init_dl_bw(&rd->dl_bw); 5475 if (cpudl_init(&rd->cpudl) != 0) 5476 goto free_dlo_mask; 5477 5478 if (cpupri_init(&rd->cpupri) != 0) 5479 goto free_rto_mask; 5480 return 0; 5481 5482free_rto_mask: 5483 free_cpumask_var(rd->rto_mask); 5484free_dlo_mask: 5485 free_cpumask_var(rd->dlo_mask); 5486free_online: 5487 free_cpumask_var(rd->online); 5488free_span: 5489 free_cpumask_var(rd->span); 5490out: 5491 return -ENOMEM; 5492} 5493 5494/* 5495 * By default the system creates a single root-domain with all cpus as 5496 * members (mimicking the global state we have today). 5497 */ 5498struct root_domain def_root_domain; 5499 5500static void init_defrootdomain(void) 5501{ 5502 init_rootdomain(&def_root_domain); 5503 5504 atomic_set(&def_root_domain.refcount, 1); 5505} 5506 5507static struct root_domain *alloc_rootdomain(void) 5508{ 5509 struct root_domain *rd; 5510 5511 rd = kmalloc(sizeof(*rd), GFP_KERNEL); 5512 if (!rd) 5513 return NULL; 5514 5515 if (init_rootdomain(rd) != 0) { 5516 kfree(rd); 5517 return NULL; 5518 } 5519 5520 return rd; 5521} 5522 5523static void free_sched_groups(struct sched_group *sg, int free_sgc) 5524{ 5525 struct sched_group *tmp, *first; 5526 5527 if (!sg) 5528 return; 5529 5530 first = sg; 5531 do { 5532 tmp = sg->next; 5533 5534 if (free_sgc && atomic_dec_and_test(&sg->sgc->ref)) 5535 kfree(sg->sgc); 5536 5537 kfree(sg); 5538 sg = tmp; 5539 } while (sg != first); 5540} 5541 5542static void free_sched_domain(struct rcu_head *rcu) 5543{ 5544 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu); 5545 5546 /* 5547 * If its an overlapping domain it has private groups, iterate and 5548 * nuke them all. 5549 */ 5550 if (sd->flags & SD_OVERLAP) { 5551 free_sched_groups(sd->groups, 1); 5552 } else if (atomic_dec_and_test(&sd->groups->ref)) { 5553 kfree(sd->groups->sgc); 5554 kfree(sd->groups); 5555 } 5556 kfree(sd); 5557} 5558 5559static void destroy_sched_domain(struct sched_domain *sd, int cpu) 5560{ 5561 call_rcu(&sd->rcu, free_sched_domain); 5562} 5563 5564static void destroy_sched_domains(struct sched_domain *sd, int cpu) 5565{ 5566 for (; sd; sd = sd->parent) 5567 destroy_sched_domain(sd, cpu); 5568} 5569 5570/* 5571 * Keep a special pointer to the highest sched_domain that has 5572 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this 5573 * allows us to avoid some pointer chasing select_idle_sibling(). 5574 * 5575 * Also keep a unique ID per domain (we use the first cpu number in 5576 * the cpumask of the domain), this allows us to quickly tell if 5577 * two cpus are in the same cache domain, see cpus_share_cache(). 5578 */ 5579DEFINE_PER_CPU(struct sched_domain *, sd_llc); 5580DEFINE_PER_CPU(int, sd_llc_size); 5581DEFINE_PER_CPU(int, sd_llc_id); 5582DEFINE_PER_CPU(struct sched_domain *, sd_numa); 5583DEFINE_PER_CPU(struct sched_domain *, sd_busy); 5584DEFINE_PER_CPU(struct sched_domain *, sd_asym); 5585 5586static void update_top_cache_domain(int cpu) 5587{ 5588 struct sched_domain *sd; 5589 struct sched_domain *busy_sd = NULL; 5590 int id = cpu; 5591 int size = 1; 5592 5593 sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES); 5594 if (sd) { 5595 id = cpumask_first(sched_domain_span(sd)); 5596 size = cpumask_weight(sched_domain_span(sd)); 5597 busy_sd = sd->parent; /* sd_busy */ 5598 } 5599 rcu_assign_pointer(per_cpu(sd_busy, cpu), busy_sd); 5600 5601 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd); 5602 per_cpu(sd_llc_size, cpu) = size; 5603 per_cpu(sd_llc_id, cpu) = id; 5604 5605 sd = lowest_flag_domain(cpu, SD_NUMA); 5606 rcu_assign_pointer(per_cpu(sd_numa, cpu), sd); 5607 5608 sd = highest_flag_domain(cpu, SD_ASYM_PACKING); 5609 rcu_assign_pointer(per_cpu(sd_asym, cpu), sd); 5610} 5611 5612/* 5613 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must 5614 * hold the hotplug lock. 5615 */ 5616static void 5617cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu) 5618{ 5619 struct rq *rq = cpu_rq(cpu); 5620 struct sched_domain *tmp; 5621 5622 /* Remove the sched domains which do not contribute to scheduling. */ 5623 for (tmp = sd; tmp; ) { 5624 struct sched_domain *parent = tmp->parent; 5625 if (!parent) 5626 break; 5627 5628 if (sd_parent_degenerate(tmp, parent)) { 5629 tmp->parent = parent->parent; 5630 if (parent->parent) 5631 parent->parent->child = tmp; 5632 /* 5633 * Transfer SD_PREFER_SIBLING down in case of a 5634 * degenerate parent; the spans match for this 5635 * so the property transfers. 5636 */ 5637 if (parent->flags & SD_PREFER_SIBLING) 5638 tmp->flags |= SD_PREFER_SIBLING; 5639 destroy_sched_domain(parent, cpu); 5640 } else 5641 tmp = tmp->parent; 5642 } 5643 5644 if (sd && sd_degenerate(sd)) { 5645 tmp = sd; 5646 sd = sd->parent; 5647 destroy_sched_domain(tmp, cpu); 5648 if (sd) 5649 sd->child = NULL; 5650 } 5651 5652 sched_domain_debug(sd, cpu); 5653 5654 rq_attach_root(rq, rd); 5655 tmp = rq->sd; 5656 rcu_assign_pointer(rq->sd, sd); 5657 destroy_sched_domains(tmp, cpu); 5658 5659 update_top_cache_domain(cpu); 5660} 5661 5662/* cpus with isolated domains */ 5663static cpumask_var_t cpu_isolated_map; 5664 5665/* Setup the mask of cpus configured for isolated domains */ 5666static int __init isolated_cpu_setup(char *str) 5667{ 5668 alloc_bootmem_cpumask_var(&cpu_isolated_map); 5669 cpulist_parse(str, cpu_isolated_map); 5670 return 1; 5671} 5672 5673__setup("isolcpus=", isolated_cpu_setup); 5674 5675struct s_data { 5676 struct sched_domain ** __percpu sd; 5677 struct root_domain *rd; 5678}; 5679 5680enum s_alloc { 5681 sa_rootdomain, 5682 sa_sd, 5683 sa_sd_storage, 5684 sa_none, 5685}; 5686 5687/* 5688 * Build an iteration mask that can exclude certain CPUs from the upwards 5689 * domain traversal. 5690 * 5691 * Asymmetric node setups can result in situations where the domain tree is of 5692 * unequal depth, make sure to skip domains that already cover the entire 5693 * range. 5694 * 5695 * In that case build_sched_domains() will have terminated the iteration early 5696 * and our sibling sd spans will be empty. Domains should always include the 5697 * cpu they're built on, so check that. 5698 * 5699 */ 5700static void build_group_mask(struct sched_domain *sd, struct sched_group *sg) 5701{ 5702 const struct cpumask *span = sched_domain_span(sd); 5703 struct sd_data *sdd = sd->private; 5704 struct sched_domain *sibling; 5705 int i; 5706 5707 for_each_cpu(i, span) { 5708 sibling = *per_cpu_ptr(sdd->sd, i); 5709 if (!cpumask_test_cpu(i, sched_domain_span(sibling))) 5710 continue; 5711 5712 cpumask_set_cpu(i, sched_group_mask(sg)); 5713 } 5714} 5715 5716/* 5717 * Return the canonical balance cpu for this group, this is the first cpu 5718 * of this group that's also in the iteration mask. 5719 */ 5720int group_balance_cpu(struct sched_group *sg) 5721{ 5722 return cpumask_first_and(sched_group_cpus(sg), sched_group_mask(sg)); 5723} 5724 5725static int 5726build_overlap_sched_groups(struct sched_domain *sd, int cpu) 5727{ 5728 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg; 5729 const struct cpumask *span = sched_domain_span(sd); 5730 struct cpumask *covered = sched_domains_tmpmask; 5731 struct sd_data *sdd = sd->private; 5732 struct sched_domain *child; 5733 int i; 5734 5735 cpumask_clear(covered); 5736 5737 for_each_cpu(i, span) { 5738 struct cpumask *sg_span; 5739 5740 if (cpumask_test_cpu(i, covered)) 5741 continue; 5742 5743 child = *per_cpu_ptr(sdd->sd, i); 5744 5745 /* See the comment near build_group_mask(). */ 5746 if (!cpumask_test_cpu(i, sched_domain_span(child))) 5747 continue; 5748 5749 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(), 5750 GFP_KERNEL, cpu_to_node(cpu)); 5751 5752 if (!sg) 5753 goto fail; 5754 5755 sg_span = sched_group_cpus(sg); 5756 if (child->child) { 5757 child = child->child; 5758 cpumask_copy(sg_span, sched_domain_span(child)); 5759 } else 5760 cpumask_set_cpu(i, sg_span); 5761 5762 cpumask_or(covered, covered, sg_span); 5763 5764 sg->sgc = *per_cpu_ptr(sdd->sgc, i); 5765 if (atomic_inc_return(&sg->sgc->ref) == 1) 5766 build_group_mask(sd, sg); 5767 5768 /* 5769 * Initialize sgc->capacity such that even if we mess up the 5770 * domains and no possible iteration will get us here, we won't 5771 * die on a /0 trap. 5772 */ 5773 sg->sgc->capacity = SCHED_CAPACITY_SCALE * cpumask_weight(sg_span); 5774 sg->sgc->capacity_orig = sg->sgc->capacity; 5775 5776 /* 5777 * Make sure the first group of this domain contains the 5778 * canonical balance cpu. Otherwise the sched_domain iteration 5779 * breaks. See update_sg_lb_stats(). 5780 */ 5781 if ((!groups && cpumask_test_cpu(cpu, sg_span)) || 5782 group_balance_cpu(sg) == cpu) 5783 groups = sg; 5784 5785 if (!first) 5786 first = sg; 5787 if (last) 5788 last->next = sg; 5789 last = sg; 5790 last->next = first; 5791 } 5792 sd->groups = groups; 5793 5794 return 0; 5795 5796fail: 5797 free_sched_groups(first, 0); 5798 5799 return -ENOMEM; 5800} 5801 5802static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg) 5803{ 5804 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu); 5805 struct sched_domain *child = sd->child; 5806 5807 if (child) 5808 cpu = cpumask_first(sched_domain_span(child)); 5809 5810 if (sg) { 5811 *sg = *per_cpu_ptr(sdd->sg, cpu); 5812 (*sg)->sgc = *per_cpu_ptr(sdd->sgc, cpu); 5813 atomic_set(&(*sg)->sgc->ref, 1); /* for claim_allocations */ 5814 } 5815 5816 return cpu; 5817} 5818 5819/* 5820 * build_sched_groups will build a circular linked list of the groups 5821 * covered by the given span, and will set each group's ->cpumask correctly, 5822 * and ->cpu_capacity to 0. 5823 * 5824 * Assumes the sched_domain tree is fully constructed 5825 */ 5826static int 5827build_sched_groups(struct sched_domain *sd, int cpu) 5828{ 5829 struct sched_group *first = NULL, *last = NULL; 5830 struct sd_data *sdd = sd->private; 5831 const struct cpumask *span = sched_domain_span(sd); 5832 struct cpumask *covered; 5833 int i; 5834 5835 get_group(cpu, sdd, &sd->groups); 5836 atomic_inc(&sd->groups->ref); 5837 5838 if (cpu != cpumask_first(span)) 5839 return 0; 5840 5841 lockdep_assert_held(&sched_domains_mutex); 5842 covered = sched_domains_tmpmask; 5843 5844 cpumask_clear(covered); 5845 5846 for_each_cpu(i, span) { 5847 struct sched_group *sg; 5848 int group, j; 5849 5850 if (cpumask_test_cpu(i, covered)) 5851 continue; 5852 5853 group = get_group(i, sdd, &sg); 5854 cpumask_setall(sched_group_mask(sg)); 5855 5856 for_each_cpu(j, span) { 5857 if (get_group(j, sdd, NULL) != group) 5858 continue; 5859 5860 cpumask_set_cpu(j, covered); 5861 cpumask_set_cpu(j, sched_group_cpus(sg)); 5862 } 5863 5864 if (!first) 5865 first = sg; 5866 if (last) 5867 last->next = sg; 5868 last = sg; 5869 } 5870 last->next = first; 5871 5872 return 0; 5873} 5874 5875/* 5876 * Initialize sched groups cpu_capacity. 5877 * 5878 * cpu_capacity indicates the capacity of sched group, which is used while 5879 * distributing the load between different sched groups in a sched domain. 5880 * Typically cpu_capacity for all the groups in a sched domain will be same 5881 * unless there are asymmetries in the topology. If there are asymmetries, 5882 * group having more cpu_capacity will pickup more load compared to the 5883 * group having less cpu_capacity. 5884 */ 5885static void init_sched_groups_capacity(int cpu, struct sched_domain *sd) 5886{ 5887 struct sched_group *sg = sd->groups; 5888 5889 WARN_ON(!sg); 5890 5891 do { 5892 sg->group_weight = cpumask_weight(sched_group_cpus(sg)); 5893 sg = sg->next; 5894 } while (sg != sd->groups); 5895 5896 if (cpu != group_balance_cpu(sg)) 5897 return; 5898 5899 update_group_capacity(sd, cpu); 5900 atomic_set(&sg->sgc->nr_busy_cpus, sg->group_weight); 5901} 5902 5903/* 5904 * Initializers for schedule domains 5905 * Non-inlined to reduce accumulated stack pressure in build_sched_domains() 5906 */ 5907 5908static int default_relax_domain_level = -1; 5909int sched_domain_level_max; 5910 5911static int __init setup_relax_domain_level(char *str) 5912{ 5913 if (kstrtoint(str, 0, &default_relax_domain_level)) 5914 pr_warn("Unable to set relax_domain_level\n"); 5915 5916 return 1; 5917} 5918__setup("relax_domain_level=", setup_relax_domain_level); 5919 5920static void set_domain_attribute(struct sched_domain *sd, 5921 struct sched_domain_attr *attr) 5922{ 5923 int request; 5924 5925 if (!attr || attr->relax_domain_level < 0) { 5926 if (default_relax_domain_level < 0) 5927 return; 5928 else 5929 request = default_relax_domain_level; 5930 } else 5931 request = attr->relax_domain_level; 5932 if (request < sd->level) { 5933 /* turn off idle balance on this domain */ 5934 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE); 5935 } else { 5936 /* turn on idle balance on this domain */ 5937 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE); 5938 } 5939} 5940 5941static void __sdt_free(const struct cpumask *cpu_map); 5942static int __sdt_alloc(const struct cpumask *cpu_map); 5943 5944static void __free_domain_allocs(struct s_data *d, enum s_alloc what, 5945 const struct cpumask *cpu_map) 5946{ 5947 switch (what) { 5948 case sa_rootdomain: 5949 if (!atomic_read(&d->rd->refcount)) 5950 free_rootdomain(&d->rd->rcu); /* fall through */ 5951 case sa_sd: 5952 free_percpu(d->sd); /* fall through */ 5953 case sa_sd_storage: 5954 __sdt_free(cpu_map); /* fall through */ 5955 case sa_none: 5956 break; 5957 } 5958} 5959 5960static enum s_alloc __visit_domain_allocation_hell(struct s_data *d, 5961 const struct cpumask *cpu_map) 5962{ 5963 memset(d, 0, sizeof(*d)); 5964 5965 if (__sdt_alloc(cpu_map)) 5966 return sa_sd_storage; 5967 d->sd = alloc_percpu(struct sched_domain *); 5968 if (!d->sd) 5969 return sa_sd_storage; 5970 d->rd = alloc_rootdomain(); 5971 if (!d->rd) 5972 return sa_sd; 5973 return sa_rootdomain; 5974} 5975 5976/* 5977 * NULL the sd_data elements we've used to build the sched_domain and 5978 * sched_group structure so that the subsequent __free_domain_allocs() 5979 * will not free the data we're using. 5980 */ 5981static void claim_allocations(int cpu, struct sched_domain *sd) 5982{ 5983 struct sd_data *sdd = sd->private; 5984 5985 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd); 5986 *per_cpu_ptr(sdd->sd, cpu) = NULL; 5987 5988 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref)) 5989 *per_cpu_ptr(sdd->sg, cpu) = NULL; 5990 5991 if (atomic_read(&(*per_cpu_ptr(sdd->sgc, cpu))->ref)) 5992 *per_cpu_ptr(sdd->sgc, cpu) = NULL; 5993} 5994 5995#ifdef CONFIG_NUMA 5996static int sched_domains_numa_levels; 5997static int *sched_domains_numa_distance; 5998static struct cpumask ***sched_domains_numa_masks; 5999static int sched_domains_curr_level; 6000#endif 6001 6002/* 6003 * SD_flags allowed in topology descriptions. 6004 * 6005 * SD_SHARE_CPUCAPACITY - describes SMT topologies 6006 * SD_SHARE_PKG_RESOURCES - describes shared caches 6007 * SD_NUMA - describes NUMA topologies 6008 * SD_SHARE_POWERDOMAIN - describes shared power domain 6009 * 6010 * Odd one out: 6011 * SD_ASYM_PACKING - describes SMT quirks 6012 */ 6013#define TOPOLOGY_SD_FLAGS \ 6014 (SD_SHARE_CPUCAPACITY | \ 6015 SD_SHARE_PKG_RESOURCES | \ 6016 SD_NUMA | \ 6017 SD_ASYM_PACKING | \ 6018 SD_SHARE_POWERDOMAIN) 6019 6020static struct sched_domain * 6021sd_init(struct sched_domain_topology_level *tl, int cpu) 6022{ 6023 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); 6024 int sd_weight, sd_flags = 0; 6025 6026#ifdef CONFIG_NUMA 6027 /* 6028 * Ugly hack to pass state to sd_numa_mask()... 6029 */ 6030 sched_domains_curr_level = tl->numa_level; 6031#endif 6032 6033 sd_weight = cpumask_weight(tl->mask(cpu)); 6034 6035 if (tl->sd_flags) 6036 sd_flags = (*tl->sd_flags)(); 6037 if (WARN_ONCE(sd_flags & ~TOPOLOGY_SD_FLAGS, 6038 "wrong sd_flags in topology description\n")) 6039 sd_flags &= ~TOPOLOGY_SD_FLAGS; 6040 6041 *sd = (struct sched_domain){ 6042 .min_interval = sd_weight, 6043 .max_interval = 2*sd_weight, 6044 .busy_factor = 32, 6045 .imbalance_pct = 125, 6046 6047 .cache_nice_tries = 0, 6048 .busy_idx = 0, 6049 .idle_idx = 0, 6050 .newidle_idx = 0, 6051 .wake_idx = 0, 6052 .forkexec_idx = 0, 6053 6054 .flags = 1*SD_LOAD_BALANCE 6055 | 1*SD_BALANCE_NEWIDLE 6056 | 1*SD_BALANCE_EXEC 6057 | 1*SD_BALANCE_FORK 6058 | 0*SD_BALANCE_WAKE 6059 | 1*SD_WAKE_AFFINE 6060 | 0*SD_SHARE_CPUCAPACITY 6061 | 0*SD_SHARE_PKG_RESOURCES 6062 | 0*SD_SERIALIZE 6063 | 0*SD_PREFER_SIBLING 6064 | 0*SD_NUMA 6065 | sd_flags 6066 , 6067 6068 .last_balance = jiffies, 6069 .balance_interval = sd_weight, 6070 .smt_gain = 0, 6071 .max_newidle_lb_cost = 0, 6072 .next_decay_max_lb_cost = jiffies, 6073#ifdef CONFIG_SCHED_DEBUG 6074 .name = tl->name, 6075#endif 6076 }; 6077 6078 /* 6079 * Convert topological properties into behaviour. 6080 */ 6081 6082 if (sd->flags & SD_SHARE_CPUCAPACITY) { 6083 sd->imbalance_pct = 110; 6084 sd->smt_gain = 1178; /* ~15% */ 6085 6086 } else if (sd->flags & SD_SHARE_PKG_RESOURCES) { 6087 sd->imbalance_pct = 117; 6088 sd->cache_nice_tries = 1; 6089 sd->busy_idx = 2; 6090 6091#ifdef CONFIG_NUMA 6092 } else if (sd->flags & SD_NUMA) { 6093 sd->cache_nice_tries = 2; 6094 sd->busy_idx = 3; 6095 sd->idle_idx = 2; 6096 6097 sd->flags |= SD_SERIALIZE; 6098 if (sched_domains_numa_distance[tl->numa_level] > RECLAIM_DISTANCE) { 6099 sd->flags &= ~(SD_BALANCE_EXEC | 6100 SD_BALANCE_FORK | 6101 SD_WAKE_AFFINE); 6102 } 6103 6104#endif 6105 } else { 6106 sd->flags |= SD_PREFER_SIBLING; 6107 sd->cache_nice_tries = 1; 6108 sd->busy_idx = 2; 6109 sd->idle_idx = 1; 6110 } 6111 6112 sd->private = &tl->data; 6113 6114 return sd; 6115} 6116 6117/* 6118 * Topology list, bottom-up. 6119 */ 6120static struct sched_domain_topology_level default_topology[] = { 6121#ifdef CONFIG_SCHED_SMT 6122 { cpu_smt_mask, cpu_smt_flags, SD_INIT_NAME(SMT) }, 6123#endif 6124#ifdef CONFIG_SCHED_MC 6125 { cpu_coregroup_mask, cpu_core_flags, SD_INIT_NAME(MC) }, 6126#endif 6127 { cpu_cpu_mask, SD_INIT_NAME(DIE) }, 6128 { NULL, }, 6129}; 6130 6131struct sched_domain_topology_level *sched_domain_topology = default_topology; 6132 6133#define for_each_sd_topology(tl) \ 6134 for (tl = sched_domain_topology; tl->mask; tl++) 6135 6136void set_sched_topology(struct sched_domain_topology_level *tl) 6137{ 6138 sched_domain_topology = tl; 6139} 6140 6141#ifdef CONFIG_NUMA 6142 6143static const struct cpumask *sd_numa_mask(int cpu) 6144{ 6145 return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)]; 6146} 6147 6148static void sched_numa_warn(const char *str) 6149{ 6150 static int done = false; 6151 int i,j; 6152 6153 if (done) 6154 return; 6155 6156 done = true; 6157 6158 printk(KERN_WARNING "ERROR: %s\n\n", str); 6159 6160 for (i = 0; i < nr_node_ids; i++) { 6161 printk(KERN_WARNING " "); 6162 for (j = 0; j < nr_node_ids; j++) 6163 printk(KERN_CONT "%02d ", node_distance(i,j)); 6164 printk(KERN_CONT "\n"); 6165 } 6166 printk(KERN_WARNING "\n"); 6167} 6168 6169static bool find_numa_distance(int distance) 6170{ 6171 int i; 6172 6173 if (distance == node_distance(0, 0)) 6174 return true; 6175 6176 for (i = 0; i < sched_domains_numa_levels; i++) { 6177 if (sched_domains_numa_distance[i] == distance) 6178 return true; 6179 } 6180 6181 return false; 6182} 6183 6184static void sched_init_numa(void) 6185{ 6186 int next_distance, curr_distance = node_distance(0, 0); 6187 struct sched_domain_topology_level *tl; 6188 int level = 0; 6189 int i, j, k; 6190 6191 sched_domains_numa_distance = kzalloc(sizeof(int) * nr_node_ids, GFP_KERNEL); 6192 if (!sched_domains_numa_distance) 6193 return; 6194 6195 /* 6196 * O(nr_nodes^2) deduplicating selection sort -- in order to find the 6197 * unique distances in the node_distance() table. 6198 * 6199 * Assumes node_distance(0,j) includes all distances in 6200 * node_distance(i,j) in order to avoid cubic time. 6201 */ 6202 next_distance = curr_distance; 6203 for (i = 0; i < nr_node_ids; i++) { 6204 for (j = 0; j < nr_node_ids; j++) { 6205 for (k = 0; k < nr_node_ids; k++) { 6206 int distance = node_distance(i, k); 6207 6208 if (distance > curr_distance && 6209 (distance < next_distance || 6210 next_distance == curr_distance)) 6211 next_distance = distance; 6212 6213 /* 6214 * While not a strong assumption it would be nice to know 6215 * about cases where if node A is connected to B, B is not 6216 * equally connected to A. 6217 */ 6218 if (sched_debug() && node_distance(k, i) != distance) 6219 sched_numa_warn("Node-distance not symmetric"); 6220 6221 if (sched_debug() && i && !find_numa_distance(distance)) 6222 sched_numa_warn("Node-0 not representative"); 6223 } 6224 if (next_distance != curr_distance) { 6225 sched_domains_numa_distance[level++] = next_distance; 6226 sched_domains_numa_levels = level; 6227 curr_distance = next_distance; 6228 } else break; 6229 } 6230 6231 /* 6232 * In case of sched_debug() we verify the above assumption. 6233 */ 6234 if (!sched_debug()) 6235 break; 6236 } 6237 /* 6238 * 'level' contains the number of unique distances, excluding the 6239 * identity distance node_distance(i,i). 6240 * 6241 * The sched_domains_numa_distance[] array includes the actual distance 6242 * numbers. 6243 */ 6244 6245 /* 6246 * Here, we should temporarily reset sched_domains_numa_levels to 0. 6247 * If it fails to allocate memory for array sched_domains_numa_masks[][], 6248 * the array will contain less then 'level' members. This could be 6249 * dangerous when we use it to iterate array sched_domains_numa_masks[][] 6250 * in other functions. 6251 * 6252 * We reset it to 'level' at the end of this function. 6253 */ 6254 sched_domains_numa_levels = 0; 6255 6256 sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL); 6257 if (!sched_domains_numa_masks) 6258 return; 6259 6260 /* 6261 * Now for each level, construct a mask per node which contains all 6262 * cpus of nodes that are that many hops away from us. 6263 */ 6264 for (i = 0; i < level; i++) { 6265 sched_domains_numa_masks[i] = 6266 kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL); 6267 if (!sched_domains_numa_masks[i]) 6268 return; 6269 6270 for (j = 0; j < nr_node_ids; j++) { 6271 struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL); 6272 if (!mask) 6273 return; 6274 6275 sched_domains_numa_masks[i][j] = mask; 6276 6277 for (k = 0; k < nr_node_ids; k++) { 6278 if (node_distance(j, k) > sched_domains_numa_distance[i]) 6279 continue; 6280 6281 cpumask_or(mask, mask, cpumask_of_node(k)); 6282 } 6283 } 6284 } 6285 6286 /* Compute default topology size */ 6287 for (i = 0; sched_domain_topology[i].mask; i++); 6288 6289 tl = kzalloc((i + level + 1) * 6290 sizeof(struct sched_domain_topology_level), GFP_KERNEL); 6291 if (!tl) 6292 return; 6293 6294 /* 6295 * Copy the default topology bits.. 6296 */ 6297 for (i = 0; sched_domain_topology[i].mask; i++) 6298 tl[i] = sched_domain_topology[i]; 6299 6300 /* 6301 * .. and append 'j' levels of NUMA goodness. 6302 */ 6303 for (j = 0; j < level; i++, j++) { 6304 tl[i] = (struct sched_domain_topology_level){ 6305 .mask = sd_numa_mask, 6306 .sd_flags = cpu_numa_flags, 6307 .flags = SDTL_OVERLAP, 6308 .numa_level = j, 6309 SD_INIT_NAME(NUMA) 6310 }; 6311 } 6312 6313 sched_domain_topology = tl; 6314 6315 sched_domains_numa_levels = level; 6316} 6317 6318static void sched_domains_numa_masks_set(int cpu) 6319{ 6320 int i, j; 6321 int node = cpu_to_node(cpu); 6322 6323 for (i = 0; i < sched_domains_numa_levels; i++) { 6324 for (j = 0; j < nr_node_ids; j++) { 6325 if (node_distance(j, node) <= sched_domains_numa_distance[i]) 6326 cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]); 6327 } 6328 } 6329} 6330 6331static void sched_domains_numa_masks_clear(int cpu) 6332{ 6333 int i, j; 6334 for (i = 0; i < sched_domains_numa_levels; i++) { 6335 for (j = 0; j < nr_node_ids; j++) 6336 cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]); 6337 } 6338} 6339 6340/* 6341 * Update sched_domains_numa_masks[level][node] array when new cpus 6342 * are onlined. 6343 */ 6344static int sched_domains_numa_masks_update(struct notifier_block *nfb, 6345 unsigned long action, 6346 void *hcpu) 6347{ 6348 int cpu = (long)hcpu; 6349 6350 switch (action & ~CPU_TASKS_FROZEN) { 6351 case CPU_ONLINE: 6352 sched_domains_numa_masks_set(cpu); 6353 break; 6354 6355 case CPU_DEAD: 6356 sched_domains_numa_masks_clear(cpu); 6357 break; 6358 6359 default: 6360 return NOTIFY_DONE; 6361 } 6362 6363 return NOTIFY_OK; 6364} 6365#else 6366static inline void sched_init_numa(void) 6367{ 6368} 6369 6370static int sched_domains_numa_masks_update(struct notifier_block *nfb, 6371 unsigned long action, 6372 void *hcpu) 6373{ 6374 return 0; 6375} 6376#endif /* CONFIG_NUMA */ 6377 6378static int __sdt_alloc(const struct cpumask *cpu_map) 6379{ 6380 struct sched_domain_topology_level *tl; 6381 int j; 6382 6383 for_each_sd_topology(tl) { 6384 struct sd_data *sdd = &tl->data; 6385 6386 sdd->sd = alloc_percpu(struct sched_domain *); 6387 if (!sdd->sd) 6388 return -ENOMEM; 6389 6390 sdd->sg = alloc_percpu(struct sched_group *); 6391 if (!sdd->sg) 6392 return -ENOMEM; 6393 6394 sdd->sgc = alloc_percpu(struct sched_group_capacity *); 6395 if (!sdd->sgc) 6396 return -ENOMEM; 6397 6398 for_each_cpu(j, cpu_map) { 6399 struct sched_domain *sd; 6400 struct sched_group *sg; 6401 struct sched_group_capacity *sgc; 6402 6403 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(), 6404 GFP_KERNEL, cpu_to_node(j)); 6405 if (!sd) 6406 return -ENOMEM; 6407 6408 *per_cpu_ptr(sdd->sd, j) = sd; 6409 6410 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(), 6411 GFP_KERNEL, cpu_to_node(j)); 6412 if (!sg) 6413 return -ENOMEM; 6414 6415 sg->next = sg; 6416 6417 *per_cpu_ptr(sdd->sg, j) = sg; 6418 6419 sgc = kzalloc_node(sizeof(struct sched_group_capacity) + cpumask_size(), 6420 GFP_KERNEL, cpu_to_node(j)); 6421 if (!sgc) 6422 return -ENOMEM; 6423 6424 *per_cpu_ptr(sdd->sgc, j) = sgc; 6425 } 6426 } 6427 6428 return 0; 6429} 6430 6431static void __sdt_free(const struct cpumask *cpu_map) 6432{ 6433 struct sched_domain_topology_level *tl; 6434 int j; 6435 6436 for_each_sd_topology(tl) { 6437 struct sd_data *sdd = &tl->data; 6438 6439 for_each_cpu(j, cpu_map) { 6440 struct sched_domain *sd; 6441 6442 if (sdd->sd) { 6443 sd = *per_cpu_ptr(sdd->sd, j); 6444 if (sd && (sd->flags & SD_OVERLAP)) 6445 free_sched_groups(sd->groups, 0); 6446 kfree(*per_cpu_ptr(sdd->sd, j)); 6447 } 6448 6449 if (sdd->sg) 6450 kfree(*per_cpu_ptr(sdd->sg, j)); 6451 if (sdd->sgc) 6452 kfree(*per_cpu_ptr(sdd->sgc, j)); 6453 } 6454 free_percpu(sdd->sd); 6455 sdd->sd = NULL; 6456 free_percpu(sdd->sg); 6457 sdd->sg = NULL; 6458 free_percpu(sdd->sgc); 6459 sdd->sgc = NULL; 6460 } 6461} 6462 6463struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl, 6464 const struct cpumask *cpu_map, struct sched_domain_attr *attr, 6465 struct sched_domain *child, int cpu) 6466{ 6467 struct sched_domain *sd = sd_init(tl, cpu); 6468 if (!sd) 6469 return child; 6470 6471 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu)); 6472 if (child) { 6473 sd->level = child->level + 1; 6474 sched_domain_level_max = max(sched_domain_level_max, sd->level); 6475 child->parent = sd; 6476 sd->child = child; 6477 } 6478 set_domain_attribute(sd, attr); 6479 6480 return sd; 6481} 6482 6483/* 6484 * Build sched domains for a given set of cpus and attach the sched domains 6485 * to the individual cpus 6486 */ 6487static int build_sched_domains(const struct cpumask *cpu_map, 6488 struct sched_domain_attr *attr) 6489{ 6490 enum s_alloc alloc_state; 6491 struct sched_domain *sd; 6492 struct s_data d; 6493 int i, ret = -ENOMEM; 6494 6495 alloc_state = __visit_domain_allocation_hell(&d, cpu_map); 6496 if (alloc_state != sa_rootdomain) 6497 goto error; 6498 6499 /* Set up domains for cpus specified by the cpu_map. */ 6500 for_each_cpu(i, cpu_map) { 6501 struct sched_domain_topology_level *tl; 6502 6503 sd = NULL; 6504 for_each_sd_topology(tl) { 6505 sd = build_sched_domain(tl, cpu_map, attr, sd, i); 6506 if (tl == sched_domain_topology) 6507 *per_cpu_ptr(d.sd, i) = sd; 6508 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP)) 6509 sd->flags |= SD_OVERLAP; 6510 if (cpumask_equal(cpu_map, sched_domain_span(sd))) 6511 break; 6512 } 6513 } 6514 6515 /* Build the groups for the domains */ 6516 for_each_cpu(i, cpu_map) { 6517 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) { 6518 sd->span_weight = cpumask_weight(sched_domain_span(sd)); 6519 if (sd->flags & SD_OVERLAP) { 6520 if (build_overlap_sched_groups(sd, i)) 6521 goto error; 6522 } else { 6523 if (build_sched_groups(sd, i)) 6524 goto error; 6525 } 6526 } 6527 } 6528 6529 /* Calculate CPU capacity for physical packages and nodes */ 6530 for (i = nr_cpumask_bits-1; i >= 0; i--) { 6531 if (!cpumask_test_cpu(i, cpu_map)) 6532 continue; 6533 6534 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) { 6535 claim_allocations(i, sd); 6536 init_sched_groups_capacity(i, sd); 6537 } 6538 } 6539 6540 /* Attach the domains */ 6541 rcu_read_lock(); 6542 for_each_cpu(i, cpu_map) { 6543 sd = *per_cpu_ptr(d.sd, i); 6544 cpu_attach_domain(sd, d.rd, i); 6545 } 6546 rcu_read_unlock(); 6547 6548 ret = 0; 6549error: 6550 __free_domain_allocs(&d, alloc_state, cpu_map); 6551 return ret; 6552} 6553 6554static cpumask_var_t *doms_cur; /* current sched domains */ 6555static int ndoms_cur; /* number of sched domains in 'doms_cur' */ 6556static struct sched_domain_attr *dattr_cur; 6557 /* attribues of custom domains in 'doms_cur' */ 6558 6559/* 6560 * Special case: If a kmalloc of a doms_cur partition (array of 6561 * cpumask) fails, then fallback to a single sched domain, 6562 * as determined by the single cpumask fallback_doms. 6563 */ 6564static cpumask_var_t fallback_doms; 6565 6566/* 6567 * arch_update_cpu_topology lets virtualized architectures update the 6568 * cpu core maps. It is supposed to return 1 if the topology changed 6569 * or 0 if it stayed the same. 6570 */ 6571int __weak arch_update_cpu_topology(void) 6572{ 6573 return 0; 6574} 6575 6576cpumask_var_t *alloc_sched_domains(unsigned int ndoms) 6577{ 6578 int i; 6579 cpumask_var_t *doms; 6580 6581 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL); 6582 if (!doms) 6583 return NULL; 6584 for (i = 0; i < ndoms; i++) { 6585 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) { 6586 free_sched_domains(doms, i); 6587 return NULL; 6588 } 6589 } 6590 return doms; 6591} 6592 6593void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms) 6594{ 6595 unsigned int i; 6596 for (i = 0; i < ndoms; i++) 6597 free_cpumask_var(doms[i]); 6598 kfree(doms); 6599} 6600 6601/* 6602 * Set up scheduler domains and groups. Callers must hold the hotplug lock. 6603 * For now this just excludes isolated cpus, but could be used to 6604 * exclude other special cases in the future. 6605 */ 6606static int init_sched_domains(const struct cpumask *cpu_map) 6607{ 6608 int err; 6609 6610 arch_update_cpu_topology(); 6611 ndoms_cur = 1; 6612 doms_cur = alloc_sched_domains(ndoms_cur); 6613 if (!doms_cur) 6614 doms_cur = &fallback_doms; 6615 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map); 6616 err = build_sched_domains(doms_cur[0], NULL); 6617 register_sched_domain_sysctl(); 6618 6619 return err; 6620} 6621 6622/* 6623 * Detach sched domains from a group of cpus specified in cpu_map 6624 * These cpus will now be attached to the NULL domain 6625 */ 6626static void detach_destroy_domains(const struct cpumask *cpu_map) 6627{ 6628 int i; 6629 6630 rcu_read_lock(); 6631 for_each_cpu(i, cpu_map) 6632 cpu_attach_domain(NULL, &def_root_domain, i); 6633 rcu_read_unlock(); 6634} 6635 6636/* handle null as "default" */ 6637static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur, 6638 struct sched_domain_attr *new, int idx_new) 6639{ 6640 struct sched_domain_attr tmp; 6641 6642 /* fast path */ 6643 if (!new && !cur) 6644 return 1; 6645 6646 tmp = SD_ATTR_INIT; 6647 return !memcmp(cur ? (cur + idx_cur) : &tmp, 6648 new ? (new + idx_new) : &tmp, 6649 sizeof(struct sched_domain_attr)); 6650} 6651 6652/* 6653 * Partition sched domains as specified by the 'ndoms_new' 6654 * cpumasks in the array doms_new[] of cpumasks. This compares 6655 * doms_new[] to the current sched domain partitioning, doms_cur[]. 6656 * It destroys each deleted domain and builds each new domain. 6657 * 6658 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'. 6659 * The masks don't intersect (don't overlap.) We should setup one 6660 * sched domain for each mask. CPUs not in any of the cpumasks will 6661 * not be load balanced. If the same cpumask appears both in the 6662 * current 'doms_cur' domains and in the new 'doms_new', we can leave 6663 * it as it is. 6664 * 6665 * The passed in 'doms_new' should be allocated using 6666 * alloc_sched_domains. This routine takes ownership of it and will 6667 * free_sched_domains it when done with it. If the caller failed the 6668 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1, 6669 * and partition_sched_domains() will fallback to the single partition 6670 * 'fallback_doms', it also forces the domains to be rebuilt. 6671 * 6672 * If doms_new == NULL it will be replaced with cpu_online_mask. 6673 * ndoms_new == 0 is a special case for destroying existing domains, 6674 * and it will not create the default domain. 6675 * 6676 * Call with hotplug lock held 6677 */ 6678void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[], 6679 struct sched_domain_attr *dattr_new) 6680{ 6681 int i, j, n; 6682 int new_topology; 6683 6684 mutex_lock(&sched_domains_mutex); 6685 6686 /* always unregister in case we don't destroy any domains */ 6687 unregister_sched_domain_sysctl(); 6688 6689 /* Let architecture update cpu core mappings. */ 6690 new_topology = arch_update_cpu_topology(); 6691 6692 n = doms_new ? ndoms_new : 0; 6693 6694 /* Destroy deleted domains */ 6695 for (i = 0; i < ndoms_cur; i++) { 6696 for (j = 0; j < n && !new_topology; j++) { 6697 if (cpumask_equal(doms_cur[i], doms_new[j]) 6698 && dattrs_equal(dattr_cur, i, dattr_new, j)) 6699 goto match1; 6700 } 6701 /* no match - a current sched domain not in new doms_new[] */ 6702 detach_destroy_domains(doms_cur[i]); 6703match1: 6704 ; 6705 } 6706 6707 n = ndoms_cur; 6708 if (doms_new == NULL) { 6709 n = 0; 6710 doms_new = &fallback_doms; 6711 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map); 6712 WARN_ON_ONCE(dattr_new); 6713 } 6714 6715 /* Build new domains */ 6716 for (i = 0; i < ndoms_new; i++) { 6717 for (j = 0; j < n && !new_topology; j++) { 6718 if (cpumask_equal(doms_new[i], doms_cur[j]) 6719 && dattrs_equal(dattr_new, i, dattr_cur, j)) 6720 goto match2; 6721 } 6722 /* no match - add a new doms_new */ 6723 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL); 6724match2: 6725 ; 6726 } 6727 6728 /* Remember the new sched domains */ 6729 if (doms_cur != &fallback_doms) 6730 free_sched_domains(doms_cur, ndoms_cur); 6731 kfree(dattr_cur); /* kfree(NULL) is safe */ 6732 doms_cur = doms_new; 6733 dattr_cur = dattr_new; 6734 ndoms_cur = ndoms_new; 6735 6736 register_sched_domain_sysctl(); 6737 6738 mutex_unlock(&sched_domains_mutex); 6739} 6740 6741static int num_cpus_frozen; /* used to mark begin/end of suspend/resume */ 6742 6743/* 6744 * Update cpusets according to cpu_active mask. If cpusets are 6745 * disabled, cpuset_update_active_cpus() becomes a simple wrapper 6746 * around partition_sched_domains(). 6747 * 6748 * If we come here as part of a suspend/resume, don't touch cpusets because we 6749 * want to restore it back to its original state upon resume anyway. 6750 */ 6751static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action, 6752 void *hcpu) 6753{ 6754 switch (action) { 6755 case CPU_ONLINE_FROZEN: 6756 case CPU_DOWN_FAILED_FROZEN: 6757 6758 /* 6759 * num_cpus_frozen tracks how many CPUs are involved in suspend 6760 * resume sequence. As long as this is not the last online 6761 * operation in the resume sequence, just build a single sched 6762 * domain, ignoring cpusets. 6763 */ 6764 num_cpus_frozen--; 6765 if (likely(num_cpus_frozen)) { 6766 partition_sched_domains(1, NULL, NULL); 6767 break; 6768 } 6769 6770 /* 6771 * This is the last CPU online operation. So fall through and 6772 * restore the original sched domains by considering the 6773 * cpuset configurations. 6774 */ 6775 6776 case CPU_ONLINE: 6777 case CPU_DOWN_FAILED: 6778 cpuset_update_active_cpus(true); 6779 break; 6780 default: 6781 return NOTIFY_DONE; 6782 } 6783 return NOTIFY_OK; 6784} 6785 6786static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action, 6787 void *hcpu) 6788{ 6789 switch (action) { 6790 case CPU_DOWN_PREPARE: 6791 cpuset_update_active_cpus(false); 6792 break; 6793 case CPU_DOWN_PREPARE_FROZEN: 6794 num_cpus_frozen++; 6795 partition_sched_domains(1, NULL, NULL); 6796 break; 6797 default: 6798 return NOTIFY_DONE; 6799 } 6800 return NOTIFY_OK; 6801} 6802 6803void __init sched_init_smp(void) 6804{ 6805 cpumask_var_t non_isolated_cpus; 6806 6807 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL); 6808 alloc_cpumask_var(&fallback_doms, GFP_KERNEL); 6809 6810 sched_init_numa(); 6811 6812 /* 6813 * There's no userspace yet to cause hotplug operations; hence all the 6814 * cpu masks are stable and all blatant races in the below code cannot 6815 * happen. 6816 */ 6817 mutex_lock(&sched_domains_mutex); 6818 init_sched_domains(cpu_active_mask); 6819 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map); 6820 if (cpumask_empty(non_isolated_cpus)) 6821 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus); 6822 mutex_unlock(&sched_domains_mutex); 6823 6824 hotcpu_notifier(sched_domains_numa_masks_update, CPU_PRI_SCHED_ACTIVE); 6825 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE); 6826 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE); 6827 6828 init_hrtick(); 6829 6830 /* Move init over to a non-isolated CPU */ 6831 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0) 6832 BUG(); 6833 sched_init_granularity(); 6834 free_cpumask_var(non_isolated_cpus); 6835 6836 init_sched_rt_class(); 6837 init_sched_dl_class(); 6838} 6839#else 6840void __init sched_init_smp(void) 6841{ 6842 sched_init_granularity(); 6843} 6844#endif /* CONFIG_SMP */ 6845 6846const_debug unsigned int sysctl_timer_migration = 1; 6847 6848int in_sched_functions(unsigned long addr) 6849{ 6850 return in_lock_functions(addr) || 6851 (addr >= (unsigned long)__sched_text_start 6852 && addr < (unsigned long)__sched_text_end); 6853} 6854 6855#ifdef CONFIG_CGROUP_SCHED 6856/* 6857 * Default task group. 6858 * Every task in system belongs to this group at bootup. 6859 */ 6860struct task_group root_task_group; 6861LIST_HEAD(task_groups); 6862#endif 6863 6864DECLARE_PER_CPU(cpumask_var_t, load_balance_mask); 6865 6866void __init sched_init(void) 6867{ 6868 int i, j; 6869 unsigned long alloc_size = 0, ptr; 6870 6871#ifdef CONFIG_FAIR_GROUP_SCHED 6872 alloc_size += 2 * nr_cpu_ids * sizeof(void **); 6873#endif 6874#ifdef CONFIG_RT_GROUP_SCHED 6875 alloc_size += 2 * nr_cpu_ids * sizeof(void **); 6876#endif 6877#ifdef CONFIG_CPUMASK_OFFSTACK 6878 alloc_size += num_possible_cpus() * cpumask_size(); 6879#endif 6880 if (alloc_size) { 6881 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT); 6882 6883#ifdef CONFIG_FAIR_GROUP_SCHED 6884 root_task_group.se = (struct sched_entity **)ptr; 6885 ptr += nr_cpu_ids * sizeof(void **); 6886 6887 root_task_group.cfs_rq = (struct cfs_rq **)ptr; 6888 ptr += nr_cpu_ids * sizeof(void **); 6889 6890#endif /* CONFIG_FAIR_GROUP_SCHED */ 6891#ifdef CONFIG_RT_GROUP_SCHED 6892 root_task_group.rt_se = (struct sched_rt_entity **)ptr; 6893 ptr += nr_cpu_ids * sizeof(void **); 6894 6895 root_task_group.rt_rq = (struct rt_rq **)ptr; 6896 ptr += nr_cpu_ids * sizeof(void **); 6897 6898#endif /* CONFIG_RT_GROUP_SCHED */ 6899#ifdef CONFIG_CPUMASK_OFFSTACK 6900 for_each_possible_cpu(i) { 6901 per_cpu(load_balance_mask, i) = (void *)ptr; 6902 ptr += cpumask_size(); 6903 } 6904#endif /* CONFIG_CPUMASK_OFFSTACK */ 6905 } 6906 6907 init_rt_bandwidth(&def_rt_bandwidth, 6908 global_rt_period(), global_rt_runtime()); 6909 init_dl_bandwidth(&def_dl_bandwidth, 6910 global_rt_period(), global_rt_runtime()); 6911 6912#ifdef CONFIG_SMP 6913 init_defrootdomain(); 6914#endif 6915 6916#ifdef CONFIG_RT_GROUP_SCHED 6917 init_rt_bandwidth(&root_task_group.rt_bandwidth, 6918 global_rt_period(), global_rt_runtime()); 6919#endif /* CONFIG_RT_GROUP_SCHED */ 6920 6921#ifdef CONFIG_CGROUP_SCHED 6922 list_add(&root_task_group.list, &task_groups); 6923 INIT_LIST_HEAD(&root_task_group.children); 6924 INIT_LIST_HEAD(&root_task_group.siblings); 6925 autogroup_init(&init_task); 6926 6927#endif /* CONFIG_CGROUP_SCHED */ 6928 6929 for_each_possible_cpu(i) { 6930 struct rq *rq; 6931 6932 rq = cpu_rq(i); 6933 raw_spin_lock_init(&rq->lock); 6934 rq->nr_running = 0; 6935 rq->calc_load_active = 0; 6936 rq->calc_load_update = jiffies + LOAD_FREQ; 6937 init_cfs_rq(&rq->cfs); 6938 init_rt_rq(&rq->rt, rq); 6939 init_dl_rq(&rq->dl, rq); 6940#ifdef CONFIG_FAIR_GROUP_SCHED 6941 root_task_group.shares = ROOT_TASK_GROUP_LOAD; 6942 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list); 6943 /* 6944 * How much cpu bandwidth does root_task_group get? 6945 * 6946 * In case of task-groups formed thr' the cgroup filesystem, it 6947 * gets 100% of the cpu resources in the system. This overall 6948 * system cpu resource is divided among the tasks of 6949 * root_task_group and its child task-groups in a fair manner, 6950 * based on each entity's (task or task-group's) weight 6951 * (se->load.weight). 6952 * 6953 * In other words, if root_task_group has 10 tasks of weight 6954 * 1024) and two child groups A0 and A1 (of weight 1024 each), 6955 * then A0's share of the cpu resource is: 6956 * 6957 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33% 6958 * 6959 * We achieve this by letting root_task_group's tasks sit 6960 * directly in rq->cfs (i.e root_task_group->se[] = NULL). 6961 */ 6962 init_cfs_bandwidth(&root_task_group.cfs_bandwidth); 6963 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL); 6964#endif /* CONFIG_FAIR_GROUP_SCHED */ 6965 6966 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime; 6967#ifdef CONFIG_RT_GROUP_SCHED 6968 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL); 6969#endif 6970 6971 for (j = 0; j < CPU_LOAD_IDX_MAX; j++) 6972 rq->cpu_load[j] = 0; 6973 6974 rq->last_load_update_tick = jiffies; 6975 6976#ifdef CONFIG_SMP 6977 rq->sd = NULL; 6978 rq->rd = NULL; 6979 rq->cpu_capacity = SCHED_CAPACITY_SCALE; 6980 rq->post_schedule = 0; 6981 rq->active_balance = 0; 6982 rq->next_balance = jiffies; 6983 rq->push_cpu = 0; 6984 rq->cpu = i; 6985 rq->online = 0; 6986 rq->idle_stamp = 0; 6987 rq->avg_idle = 2*sysctl_sched_migration_cost; 6988 rq->max_idle_balance_cost = sysctl_sched_migration_cost; 6989 6990 INIT_LIST_HEAD(&rq->cfs_tasks); 6991 6992 rq_attach_root(rq, &def_root_domain); 6993#ifdef CONFIG_NO_HZ_COMMON 6994 rq->nohz_flags = 0; 6995#endif 6996#ifdef CONFIG_NO_HZ_FULL 6997 rq->last_sched_tick = 0; 6998#endif 6999#endif 7000 init_rq_hrtick(rq); 7001 atomic_set(&rq->nr_iowait, 0); 7002 } 7003 7004 set_load_weight(&init_task); 7005 7006#ifdef CONFIG_PREEMPT_NOTIFIERS 7007 INIT_HLIST_HEAD(&init_task.preempt_notifiers); 7008#endif 7009 7010 /* 7011 * The boot idle thread does lazy MMU switching as well: 7012 */ 7013 atomic_inc(&init_mm.mm_count); 7014 enter_lazy_tlb(&init_mm, current); 7015 7016 /* 7017 * Make us the idle thread. Technically, schedule() should not be 7018 * called from this thread, however somewhere below it might be, 7019 * but because we are the idle thread, we just pick up running again 7020 * when this runqueue becomes "idle". 7021 */ 7022 init_idle(current, smp_processor_id()); 7023 7024 calc_load_update = jiffies + LOAD_FREQ; 7025 7026 /* 7027 * During early bootup we pretend to be a normal task: 7028 */ 7029 current->sched_class = &fair_sched_class; 7030 7031#ifdef CONFIG_SMP 7032 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT); 7033 /* May be allocated at isolcpus cmdline parse time */ 7034 if (cpu_isolated_map == NULL) 7035 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT); 7036 idle_thread_set_boot_cpu(); 7037 set_cpu_rq_start_time(); 7038#endif 7039 init_sched_fair_class(); 7040 7041 scheduler_running = 1; 7042} 7043 7044#ifdef CONFIG_DEBUG_ATOMIC_SLEEP 7045static inline int preempt_count_equals(int preempt_offset) 7046{ 7047 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth(); 7048 7049 return (nested == preempt_offset); 7050} 7051 7052void __might_sleep(const char *file, int line, int preempt_offset) 7053{ 7054 static unsigned long prev_jiffy; /* ratelimiting */ 7055 7056 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */ 7057 if ((preempt_count_equals(preempt_offset) && !irqs_disabled() && 7058 !is_idle_task(current)) || 7059 system_state != SYSTEM_RUNNING || oops_in_progress) 7060 return; 7061 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy) 7062 return; 7063 prev_jiffy = jiffies; 7064 7065 printk(KERN_ERR 7066 "BUG: sleeping function called from invalid context at %s:%d\n", 7067 file, line); 7068 printk(KERN_ERR 7069 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n", 7070 in_atomic(), irqs_disabled(), 7071 current->pid, current->comm); 7072 7073 debug_show_held_locks(current); 7074 if (irqs_disabled()) 7075 print_irqtrace_events(current); 7076#ifdef CONFIG_DEBUG_PREEMPT 7077 if (!preempt_count_equals(preempt_offset)) { 7078 pr_err("Preemption disabled at:"); 7079 print_ip_sym(current->preempt_disable_ip); 7080 pr_cont("\n"); 7081 } 7082#endif 7083 dump_stack(); 7084} 7085EXPORT_SYMBOL(__might_sleep); 7086#endif 7087 7088#ifdef CONFIG_MAGIC_SYSRQ 7089static void normalize_task(struct rq *rq, struct task_struct *p) 7090{ 7091 const struct sched_class *prev_class = p->sched_class; 7092 struct sched_attr attr = { 7093 .sched_policy = SCHED_NORMAL, 7094 }; 7095 int old_prio = p->prio; 7096 int on_rq; 7097 7098 on_rq = p->on_rq; 7099 if (on_rq) 7100 dequeue_task(rq, p, 0); 7101 __setscheduler(rq, p, &attr); 7102 if (on_rq) { 7103 enqueue_task(rq, p, 0); 7104 resched_task(rq->curr); 7105 } 7106 7107 check_class_changed(rq, p, prev_class, old_prio); 7108} 7109 7110void normalize_rt_tasks(void) 7111{ 7112 struct task_struct *g, *p; 7113 unsigned long flags; 7114 struct rq *rq; 7115 7116 read_lock_irqsave(&tasklist_lock, flags); 7117 do_each_thread(g, p) { 7118 /* 7119 * Only normalize user tasks: 7120 */ 7121 if (!p->mm) 7122 continue; 7123 7124 p->se.exec_start = 0; 7125#ifdef CONFIG_SCHEDSTATS 7126 p->se.statistics.wait_start = 0; 7127 p->se.statistics.sleep_start = 0; 7128 p->se.statistics.block_start = 0; 7129#endif 7130 7131 if (!dl_task(p) && !rt_task(p)) { 7132 /* 7133 * Renice negative nice level userspace 7134 * tasks back to 0: 7135 */ 7136 if (task_nice(p) < 0 && p->mm) 7137 set_user_nice(p, 0); 7138 continue; 7139 } 7140 7141 raw_spin_lock(&p->pi_lock); 7142 rq = __task_rq_lock(p); 7143 7144 normalize_task(rq, p); 7145 7146 __task_rq_unlock(rq); 7147 raw_spin_unlock(&p->pi_lock); 7148 } while_each_thread(g, p); 7149 7150 read_unlock_irqrestore(&tasklist_lock, flags); 7151} 7152 7153#endif /* CONFIG_MAGIC_SYSRQ */ 7154 7155#if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) 7156/* 7157 * These functions are only useful for the IA64 MCA handling, or kdb. 7158 * 7159 * They can only be called when the whole system has been 7160 * stopped - every CPU needs to be quiescent, and no scheduling 7161 * activity can take place. Using them for anything else would 7162 * be a serious bug, and as a result, they aren't even visible 7163 * under any other configuration. 7164 */ 7165 7166/** 7167 * curr_task - return the current task for a given cpu. 7168 * @cpu: the processor in question. 7169 * 7170 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED! 7171 * 7172 * Return: The current task for @cpu. 7173 */ 7174struct task_struct *curr_task(int cpu) 7175{ 7176 return cpu_curr(cpu); 7177} 7178 7179#endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */ 7180 7181#ifdef CONFIG_IA64 7182/** 7183 * set_curr_task - set the current task for a given cpu. 7184 * @cpu: the processor in question. 7185 * @p: the task pointer to set. 7186 * 7187 * Description: This function must only be used when non-maskable interrupts 7188 * are serviced on a separate stack. It allows the architecture to switch the 7189 * notion of the current task on a cpu in a non-blocking manner. This function 7190 * must be called with all CPU's synchronized, and interrupts disabled, the 7191 * and caller must save the original value of the current task (see 7192 * curr_task() above) and restore that value before reenabling interrupts and 7193 * re-starting the system. 7194 * 7195 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED! 7196 */ 7197void set_curr_task(int cpu, struct task_struct *p) 7198{ 7199 cpu_curr(cpu) = p; 7200} 7201 7202#endif 7203 7204#ifdef CONFIG_CGROUP_SCHED 7205/* task_group_lock serializes the addition/removal of task groups */ 7206static DEFINE_SPINLOCK(task_group_lock); 7207 7208static void free_sched_group(struct task_group *tg) 7209{ 7210 free_fair_sched_group(tg); 7211 free_rt_sched_group(tg); 7212 autogroup_free(tg); 7213 kfree(tg); 7214} 7215 7216/* allocate runqueue etc for a new task group */ 7217struct task_group *sched_create_group(struct task_group *parent) 7218{ 7219 struct task_group *tg; 7220 7221 tg = kzalloc(sizeof(*tg), GFP_KERNEL); 7222 if (!tg) 7223 return ERR_PTR(-ENOMEM); 7224 7225 if (!alloc_fair_sched_group(tg, parent)) 7226 goto err; 7227 7228 if (!alloc_rt_sched_group(tg, parent)) 7229 goto err; 7230 7231 return tg; 7232 7233err: 7234 free_sched_group(tg); 7235 return ERR_PTR(-ENOMEM); 7236} 7237 7238void sched_online_group(struct task_group *tg, struct task_group *parent) 7239{ 7240 unsigned long flags; 7241 7242 spin_lock_irqsave(&task_group_lock, flags); 7243 list_add_rcu(&tg->list, &task_groups); 7244 7245 WARN_ON(!parent); /* root should already exist */ 7246 7247 tg->parent = parent; 7248 INIT_LIST_HEAD(&tg->children); 7249 list_add_rcu(&tg->siblings, &parent->children); 7250 spin_unlock_irqrestore(&task_group_lock, flags); 7251} 7252 7253/* rcu callback to free various structures associated with a task group */ 7254static void free_sched_group_rcu(struct rcu_head *rhp) 7255{ 7256 /* now it should be safe to free those cfs_rqs */ 7257 free_sched_group(container_of(rhp, struct task_group, rcu)); 7258} 7259 7260/* Destroy runqueue etc associated with a task group */ 7261void sched_destroy_group(struct task_group *tg) 7262{ 7263 /* wait for possible concurrent references to cfs_rqs complete */ 7264 call_rcu(&tg->rcu, free_sched_group_rcu); 7265} 7266 7267void sched_offline_group(struct task_group *tg) 7268{ 7269 unsigned long flags; 7270 int i; 7271 7272 /* end participation in shares distribution */ 7273 for_each_possible_cpu(i) 7274 unregister_fair_sched_group(tg, i); 7275 7276 spin_lock_irqsave(&task_group_lock, flags); 7277 list_del_rcu(&tg->list); 7278 list_del_rcu(&tg->siblings); 7279 spin_unlock_irqrestore(&task_group_lock, flags); 7280} 7281 7282/* change task's runqueue when it moves between groups. 7283 * The caller of this function should have put the task in its new group 7284 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to 7285 * reflect its new group. 7286 */ 7287void sched_move_task(struct task_struct *tsk) 7288{ 7289 struct task_group *tg; 7290 int on_rq, running; 7291 unsigned long flags; 7292 struct rq *rq; 7293 7294 rq = task_rq_lock(tsk, &flags); 7295 7296 running = task_current(rq, tsk); 7297 on_rq = tsk->on_rq; 7298 7299 if (on_rq) 7300 dequeue_task(rq, tsk, 0); 7301 if (unlikely(running)) 7302 tsk->sched_class->put_prev_task(rq, tsk); 7303 7304 tg = container_of(task_css_check(tsk, cpu_cgrp_id, 7305 lockdep_is_held(&tsk->sighand->siglock)), 7306 struct task_group, css); 7307 tg = autogroup_task_group(tsk, tg); 7308 tsk->sched_task_group = tg; 7309 7310#ifdef CONFIG_FAIR_GROUP_SCHED 7311 if (tsk->sched_class->task_move_group) 7312 tsk->sched_class->task_move_group(tsk, on_rq); 7313 else 7314#endif 7315 set_task_rq(tsk, task_cpu(tsk)); 7316 7317 if (unlikely(running)) 7318 tsk->sched_class->set_curr_task(rq); 7319 if (on_rq) 7320 enqueue_task(rq, tsk, 0); 7321 7322 task_rq_unlock(rq, tsk, &flags); 7323} 7324#endif /* CONFIG_CGROUP_SCHED */ 7325 7326#ifdef CONFIG_RT_GROUP_SCHED 7327/* 7328 * Ensure that the real time constraints are schedulable. 7329 */ 7330static DEFINE_MUTEX(rt_constraints_mutex); 7331 7332/* Must be called with tasklist_lock held */ 7333static inline int tg_has_rt_tasks(struct task_group *tg) 7334{ 7335 struct task_struct *g, *p; 7336 7337 do_each_thread(g, p) { 7338 if (rt_task(p) && task_rq(p)->rt.tg == tg) 7339 return 1; 7340 } while_each_thread(g, p); 7341 7342 return 0; 7343} 7344 7345struct rt_schedulable_data { 7346 struct task_group *tg; 7347 u64 rt_period; 7348 u64 rt_runtime; 7349}; 7350 7351static int tg_rt_schedulable(struct task_group *tg, void *data) 7352{ 7353 struct rt_schedulable_data *d = data; 7354 struct task_group *child; 7355 unsigned long total, sum = 0; 7356 u64 period, runtime; 7357 7358 period = ktime_to_ns(tg->rt_bandwidth.rt_period); 7359 runtime = tg->rt_bandwidth.rt_runtime; 7360 7361 if (tg == d->tg) { 7362 period = d->rt_period; 7363 runtime = d->rt_runtime; 7364 } 7365 7366 /* 7367 * Cannot have more runtime than the period. 7368 */ 7369 if (runtime > period && runtime != RUNTIME_INF) 7370 return -EINVAL; 7371 7372 /* 7373 * Ensure we don't starve existing RT tasks. 7374 */ 7375 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg)) 7376 return -EBUSY; 7377 7378 total = to_ratio(period, runtime); 7379 7380 /* 7381 * Nobody can have more than the global setting allows. 7382 */ 7383 if (total > to_ratio(global_rt_period(), global_rt_runtime())) 7384 return -EINVAL; 7385 7386 /* 7387 * The sum of our children's runtime should not exceed our own. 7388 */ 7389 list_for_each_entry_rcu(child, &tg->children, siblings) { 7390 period = ktime_to_ns(child->rt_bandwidth.rt_period); 7391 runtime = child->rt_bandwidth.rt_runtime; 7392 7393 if (child == d->tg) { 7394 period = d->rt_period; 7395 runtime = d->rt_runtime; 7396 } 7397 7398 sum += to_ratio(period, runtime); 7399 } 7400 7401 if (sum > total) 7402 return -EINVAL; 7403 7404 return 0; 7405} 7406 7407static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime) 7408{ 7409 int ret; 7410 7411 struct rt_schedulable_data data = { 7412 .tg = tg, 7413 .rt_period = period, 7414 .rt_runtime = runtime, 7415 }; 7416 7417 rcu_read_lock(); 7418 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data); 7419 rcu_read_unlock(); 7420 7421 return ret; 7422} 7423 7424static int tg_set_rt_bandwidth(struct task_group *tg, 7425 u64 rt_period, u64 rt_runtime) 7426{ 7427 int i, err = 0; 7428 7429 mutex_lock(&rt_constraints_mutex); 7430 read_lock(&tasklist_lock); 7431 err = __rt_schedulable(tg, rt_period, rt_runtime); 7432 if (err) 7433 goto unlock; 7434 7435 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock); 7436 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period); 7437 tg->rt_bandwidth.rt_runtime = rt_runtime; 7438 7439 for_each_possible_cpu(i) { 7440 struct rt_rq *rt_rq = tg->rt_rq[i]; 7441 7442 raw_spin_lock(&rt_rq->rt_runtime_lock); 7443 rt_rq->rt_runtime = rt_runtime; 7444 raw_spin_unlock(&rt_rq->rt_runtime_lock); 7445 } 7446 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock); 7447unlock: 7448 read_unlock(&tasklist_lock); 7449 mutex_unlock(&rt_constraints_mutex); 7450 7451 return err; 7452} 7453 7454static int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us) 7455{ 7456 u64 rt_runtime, rt_period; 7457 7458 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period); 7459 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC; 7460 if (rt_runtime_us < 0) 7461 rt_runtime = RUNTIME_INF; 7462 7463 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime); 7464} 7465 7466static long sched_group_rt_runtime(struct task_group *tg) 7467{ 7468 u64 rt_runtime_us; 7469 7470 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF) 7471 return -1; 7472 7473 rt_runtime_us = tg->rt_bandwidth.rt_runtime; 7474 do_div(rt_runtime_us, NSEC_PER_USEC); 7475 return rt_runtime_us; 7476} 7477 7478static int sched_group_set_rt_period(struct task_group *tg, long rt_period_us) 7479{ 7480 u64 rt_runtime, rt_period; 7481 7482 rt_period = (u64)rt_period_us * NSEC_PER_USEC; 7483 rt_runtime = tg->rt_bandwidth.rt_runtime; 7484 7485 if (rt_period == 0) 7486 return -EINVAL; 7487 7488 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime); 7489} 7490 7491static long sched_group_rt_period(struct task_group *tg) 7492{ 7493 u64 rt_period_us; 7494 7495 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period); 7496 do_div(rt_period_us, NSEC_PER_USEC); 7497 return rt_period_us; 7498} 7499#endif /* CONFIG_RT_GROUP_SCHED */ 7500 7501#ifdef CONFIG_RT_GROUP_SCHED 7502static int sched_rt_global_constraints(void) 7503{ 7504 int ret = 0; 7505 7506 mutex_lock(&rt_constraints_mutex); 7507 read_lock(&tasklist_lock); 7508 ret = __rt_schedulable(NULL, 0, 0); 7509 read_unlock(&tasklist_lock); 7510 mutex_unlock(&rt_constraints_mutex); 7511 7512 return ret; 7513} 7514 7515static int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk) 7516{ 7517 /* Don't accept realtime tasks when there is no way for them to run */ 7518 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0) 7519 return 0; 7520 7521 return 1; 7522} 7523 7524#else /* !CONFIG_RT_GROUP_SCHED */ 7525static int sched_rt_global_constraints(void) 7526{ 7527 unsigned long flags; 7528 int i, ret = 0; 7529 7530 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags); 7531 for_each_possible_cpu(i) { 7532 struct rt_rq *rt_rq = &cpu_rq(i)->rt; 7533 7534 raw_spin_lock(&rt_rq->rt_runtime_lock); 7535 rt_rq->rt_runtime = global_rt_runtime(); 7536 raw_spin_unlock(&rt_rq->rt_runtime_lock); 7537 } 7538 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags); 7539 7540 return ret; 7541} 7542#endif /* CONFIG_RT_GROUP_SCHED */ 7543 7544static int sched_dl_global_constraints(void) 7545{ 7546 u64 runtime = global_rt_runtime(); 7547 u64 period = global_rt_period(); 7548 u64 new_bw = to_ratio(period, runtime); 7549 int cpu, ret = 0; 7550 unsigned long flags; 7551 7552 /* 7553 * Here we want to check the bandwidth not being set to some 7554 * value smaller than the currently allocated bandwidth in 7555 * any of the root_domains. 7556 * 7557 * FIXME: Cycling on all the CPUs is overdoing, but simpler than 7558 * cycling on root_domains... Discussion on different/better 7559 * solutions is welcome! 7560 */ 7561 for_each_possible_cpu(cpu) { 7562 struct dl_bw *dl_b = dl_bw_of(cpu); 7563 7564 raw_spin_lock_irqsave(&dl_b->lock, flags); 7565 if (new_bw < dl_b->total_bw) 7566 ret = -EBUSY; 7567 raw_spin_unlock_irqrestore(&dl_b->lock, flags); 7568 7569 if (ret) 7570 break; 7571 } 7572 7573 return ret; 7574} 7575 7576static void sched_dl_do_global(void) 7577{ 7578 u64 new_bw = -1; 7579 int cpu; 7580 unsigned long flags; 7581 7582 def_dl_bandwidth.dl_period = global_rt_period(); 7583 def_dl_bandwidth.dl_runtime = global_rt_runtime(); 7584 7585 if (global_rt_runtime() != RUNTIME_INF) 7586 new_bw = to_ratio(global_rt_period(), global_rt_runtime()); 7587 7588 /* 7589 * FIXME: As above... 7590 */ 7591 for_each_possible_cpu(cpu) { 7592 struct dl_bw *dl_b = dl_bw_of(cpu); 7593 7594 raw_spin_lock_irqsave(&dl_b->lock, flags); 7595 dl_b->bw = new_bw; 7596 raw_spin_unlock_irqrestore(&dl_b->lock, flags); 7597 } 7598} 7599 7600static int sched_rt_global_validate(void) 7601{ 7602 if (sysctl_sched_rt_period <= 0) 7603 return -EINVAL; 7604 7605 if ((sysctl_sched_rt_runtime != RUNTIME_INF) && 7606 (sysctl_sched_rt_runtime > sysctl_sched_rt_period)) 7607 return -EINVAL; 7608 7609 return 0; 7610} 7611 7612static void sched_rt_do_global(void) 7613{ 7614 def_rt_bandwidth.rt_runtime = global_rt_runtime(); 7615 def_rt_bandwidth.rt_period = ns_to_ktime(global_rt_period()); 7616} 7617 7618int sched_rt_handler(struct ctl_table *table, int write, 7619 void __user *buffer, size_t *lenp, 7620 loff_t *ppos) 7621{ 7622 int old_period, old_runtime; 7623 static DEFINE_MUTEX(mutex); 7624 int ret; 7625 7626 mutex_lock(&mutex); 7627 old_period = sysctl_sched_rt_period; 7628 old_runtime = sysctl_sched_rt_runtime; 7629 7630 ret = proc_dointvec(table, write, buffer, lenp, ppos); 7631 7632 if (!ret && write) { 7633 ret = sched_rt_global_validate(); 7634 if (ret) 7635 goto undo; 7636 7637 ret = sched_rt_global_constraints(); 7638 if (ret) 7639 goto undo; 7640 7641 ret = sched_dl_global_constraints(); 7642 if (ret) 7643 goto undo; 7644 7645 sched_rt_do_global(); 7646 sched_dl_do_global(); 7647 } 7648 if (0) { 7649undo: 7650 sysctl_sched_rt_period = old_period; 7651 sysctl_sched_rt_runtime = old_runtime; 7652 } 7653 mutex_unlock(&mutex); 7654 7655 return ret; 7656} 7657 7658int sched_rr_handler(struct ctl_table *table, int write, 7659 void __user *buffer, size_t *lenp, 7660 loff_t *ppos) 7661{ 7662 int ret; 7663 static DEFINE_MUTEX(mutex); 7664 7665 mutex_lock(&mutex); 7666 ret = proc_dointvec(table, write, buffer, lenp, ppos); 7667 /* make sure that internally we keep jiffies */ 7668 /* also, writing zero resets timeslice to default */ 7669 if (!ret && write) { 7670 sched_rr_timeslice = sched_rr_timeslice <= 0 ? 7671 RR_TIMESLICE : msecs_to_jiffies(sched_rr_timeslice); 7672 } 7673 mutex_unlock(&mutex); 7674 return ret; 7675} 7676 7677#ifdef CONFIG_CGROUP_SCHED 7678 7679static inline struct task_group *css_tg(struct cgroup_subsys_state *css) 7680{ 7681 return css ? container_of(css, struct task_group, css) : NULL; 7682} 7683 7684static struct cgroup_subsys_state * 7685cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css) 7686{ 7687 struct task_group *parent = css_tg(parent_css); 7688 struct task_group *tg; 7689 7690 if (!parent) { 7691 /* This is early initialization for the top cgroup */ 7692 return &root_task_group.css; 7693 } 7694 7695 tg = sched_create_group(parent); 7696 if (IS_ERR(tg)) 7697 return ERR_PTR(-ENOMEM); 7698 7699 return &tg->css; 7700} 7701 7702static int cpu_cgroup_css_online(struct cgroup_subsys_state *css) 7703{ 7704 struct task_group *tg = css_tg(css); 7705 struct task_group *parent = css_tg(css->parent); 7706 7707 if (parent) 7708 sched_online_group(tg, parent); 7709 return 0; 7710} 7711 7712static void cpu_cgroup_css_free(struct cgroup_subsys_state *css) 7713{ 7714 struct task_group *tg = css_tg(css); 7715 7716 sched_destroy_group(tg); 7717} 7718 7719static void cpu_cgroup_css_offline(struct cgroup_subsys_state *css) 7720{ 7721 struct task_group *tg = css_tg(css); 7722 7723 sched_offline_group(tg); 7724} 7725 7726static int cpu_cgroup_can_attach(struct cgroup_subsys_state *css, 7727 struct cgroup_taskset *tset) 7728{ 7729 struct task_struct *task; 7730 7731 cgroup_taskset_for_each(task, tset) { 7732#ifdef CONFIG_RT_GROUP_SCHED 7733 if (!sched_rt_can_attach(css_tg(css), task)) 7734 return -EINVAL; 7735#else 7736 /* We don't support RT-tasks being in separate groups */ 7737 if (task->sched_class != &fair_sched_class) 7738 return -EINVAL; 7739#endif 7740 } 7741 return 0; 7742} 7743 7744static void cpu_cgroup_attach(struct cgroup_subsys_state *css, 7745 struct cgroup_taskset *tset) 7746{ 7747 struct task_struct *task; 7748 7749 cgroup_taskset_for_each(task, tset) 7750 sched_move_task(task); 7751} 7752 7753static void cpu_cgroup_exit(struct cgroup_subsys_state *css, 7754 struct cgroup_subsys_state *old_css, 7755 struct task_struct *task) 7756{ 7757 /* 7758 * cgroup_exit() is called in the copy_process() failure path. 7759 * Ignore this case since the task hasn't ran yet, this avoids 7760 * trying to poke a half freed task state from generic code. 7761 */ 7762 if (!(task->flags & PF_EXITING)) 7763 return; 7764 7765 sched_move_task(task); 7766} 7767 7768#ifdef CONFIG_FAIR_GROUP_SCHED 7769static int cpu_shares_write_u64(struct cgroup_subsys_state *css, 7770 struct cftype *cftype, u64 shareval) 7771{ 7772 return sched_group_set_shares(css_tg(css), scale_load(shareval)); 7773} 7774 7775static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css, 7776 struct cftype *cft) 7777{ 7778 struct task_group *tg = css_tg(css); 7779 7780 return (u64) scale_load_down(tg->shares); 7781} 7782 7783#ifdef CONFIG_CFS_BANDWIDTH 7784static DEFINE_MUTEX(cfs_constraints_mutex); 7785 7786const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */ 7787const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */ 7788 7789static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime); 7790 7791static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota) 7792{ 7793 int i, ret = 0, runtime_enabled, runtime_was_enabled; 7794 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth; 7795 7796 if (tg == &root_task_group) 7797 return -EINVAL; 7798 7799 /* 7800 * Ensure we have at some amount of bandwidth every period. This is 7801 * to prevent reaching a state of large arrears when throttled via 7802 * entity_tick() resulting in prolonged exit starvation. 7803 */ 7804 if (quota < min_cfs_quota_period || period < min_cfs_quota_period) 7805 return -EINVAL; 7806 7807 /* 7808 * Likewise, bound things on the otherside by preventing insane quota 7809 * periods. This also allows us to normalize in computing quota 7810 * feasibility. 7811 */ 7812 if (period > max_cfs_quota_period) 7813 return -EINVAL; 7814 7815 mutex_lock(&cfs_constraints_mutex); 7816 ret = __cfs_schedulable(tg, period, quota); 7817 if (ret) 7818 goto out_unlock; 7819 7820 runtime_enabled = quota != RUNTIME_INF; 7821 runtime_was_enabled = cfs_b->quota != RUNTIME_INF; 7822 /* 7823 * If we need to toggle cfs_bandwidth_used, off->on must occur 7824 * before making related changes, and on->off must occur afterwards 7825 */ 7826 if (runtime_enabled && !runtime_was_enabled) 7827 cfs_bandwidth_usage_inc(); 7828 raw_spin_lock_irq(&cfs_b->lock); 7829 cfs_b->period = ns_to_ktime(period); 7830 cfs_b->quota = quota; 7831 7832 __refill_cfs_bandwidth_runtime(cfs_b); 7833 /* restart the period timer (if active) to handle new period expiry */ 7834 if (runtime_enabled && cfs_b->timer_active) { 7835 /* force a reprogram */ 7836 __start_cfs_bandwidth(cfs_b, true); 7837 } 7838 raw_spin_unlock_irq(&cfs_b->lock); 7839 7840 for_each_possible_cpu(i) { 7841 struct cfs_rq *cfs_rq = tg->cfs_rq[i]; 7842 struct rq *rq = cfs_rq->rq; 7843 7844 raw_spin_lock_irq(&rq->lock); 7845 cfs_rq->runtime_enabled = runtime_enabled; 7846 cfs_rq->runtime_remaining = 0; 7847 7848 if (cfs_rq->throttled) 7849 unthrottle_cfs_rq(cfs_rq); 7850 raw_spin_unlock_irq(&rq->lock); 7851 } 7852 if (runtime_was_enabled && !runtime_enabled) 7853 cfs_bandwidth_usage_dec(); 7854out_unlock: 7855 mutex_unlock(&cfs_constraints_mutex); 7856 7857 return ret; 7858} 7859 7860int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us) 7861{ 7862 u64 quota, period; 7863 7864 period = ktime_to_ns(tg->cfs_bandwidth.period); 7865 if (cfs_quota_us < 0) 7866 quota = RUNTIME_INF; 7867 else 7868 quota = (u64)cfs_quota_us * NSEC_PER_USEC; 7869 7870 return tg_set_cfs_bandwidth(tg, period, quota); 7871} 7872 7873long tg_get_cfs_quota(struct task_group *tg) 7874{ 7875 u64 quota_us; 7876 7877 if (tg->cfs_bandwidth.quota == RUNTIME_INF) 7878 return -1; 7879 7880 quota_us = tg->cfs_bandwidth.quota; 7881 do_div(quota_us, NSEC_PER_USEC); 7882 7883 return quota_us; 7884} 7885 7886int tg_set_cfs_period(struct task_group *tg, long cfs_period_us) 7887{ 7888 u64 quota, period; 7889 7890 period = (u64)cfs_period_us * NSEC_PER_USEC; 7891 quota = tg->cfs_bandwidth.quota; 7892 7893 return tg_set_cfs_bandwidth(tg, period, quota); 7894} 7895 7896long tg_get_cfs_period(struct task_group *tg) 7897{ 7898 u64 cfs_period_us; 7899 7900 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period); 7901 do_div(cfs_period_us, NSEC_PER_USEC); 7902 7903 return cfs_period_us; 7904} 7905 7906static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css, 7907 struct cftype *cft) 7908{ 7909 return tg_get_cfs_quota(css_tg(css)); 7910} 7911 7912static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css, 7913 struct cftype *cftype, s64 cfs_quota_us) 7914{ 7915 return tg_set_cfs_quota(css_tg(css), cfs_quota_us); 7916} 7917 7918static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css, 7919 struct cftype *cft) 7920{ 7921 return tg_get_cfs_period(css_tg(css)); 7922} 7923 7924static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css, 7925 struct cftype *cftype, u64 cfs_period_us) 7926{ 7927 return tg_set_cfs_period(css_tg(css), cfs_period_us); 7928} 7929 7930struct cfs_schedulable_data { 7931 struct task_group *tg; 7932 u64 period, quota; 7933}; 7934 7935/* 7936 * normalize group quota/period to be quota/max_period 7937 * note: units are usecs 7938 */ 7939static u64 normalize_cfs_quota(struct task_group *tg, 7940 struct cfs_schedulable_data *d) 7941{ 7942 u64 quota, period; 7943 7944 if (tg == d->tg) { 7945 period = d->period; 7946 quota = d->quota; 7947 } else { 7948 period = tg_get_cfs_period(tg); 7949 quota = tg_get_cfs_quota(tg); 7950 } 7951 7952 /* note: these should typically be equivalent */ 7953 if (quota == RUNTIME_INF || quota == -1) 7954 return RUNTIME_INF; 7955 7956 return to_ratio(period, quota); 7957} 7958 7959static int tg_cfs_schedulable_down(struct task_group *tg, void *data) 7960{ 7961 struct cfs_schedulable_data *d = data; 7962 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth; 7963 s64 quota = 0, parent_quota = -1; 7964 7965 if (!tg->parent) { 7966 quota = RUNTIME_INF; 7967 } else { 7968 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth; 7969 7970 quota = normalize_cfs_quota(tg, d); 7971 parent_quota = parent_b->hierarchal_quota; 7972 7973 /* 7974 * ensure max(child_quota) <= parent_quota, inherit when no 7975 * limit is set 7976 */ 7977 if (quota == RUNTIME_INF) 7978 quota = parent_quota; 7979 else if (parent_quota != RUNTIME_INF && quota > parent_quota) 7980 return -EINVAL; 7981 } 7982 cfs_b->hierarchal_quota = quota; 7983 7984 return 0; 7985} 7986 7987static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota) 7988{ 7989 int ret; 7990 struct cfs_schedulable_data data = { 7991 .tg = tg, 7992 .period = period, 7993 .quota = quota, 7994 }; 7995 7996 if (quota != RUNTIME_INF) { 7997 do_div(data.period, NSEC_PER_USEC); 7998 do_div(data.quota, NSEC_PER_USEC); 7999 } 8000 8001 rcu_read_lock(); 8002 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data); 8003 rcu_read_unlock(); 8004 8005 return ret; 8006} 8007 8008static int cpu_stats_show(struct seq_file *sf, void *v) 8009{ 8010 struct task_group *tg = css_tg(seq_css(sf)); 8011 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth; 8012 8013 seq_printf(sf, "nr_periods %d\n", cfs_b->nr_periods); 8014 seq_printf(sf, "nr_throttled %d\n", cfs_b->nr_throttled); 8015 seq_printf(sf, "throttled_time %llu\n", cfs_b->throttled_time); 8016 8017 return 0; 8018} 8019#endif /* CONFIG_CFS_BANDWIDTH */ 8020#endif /* CONFIG_FAIR_GROUP_SCHED */ 8021 8022#ifdef CONFIG_RT_GROUP_SCHED 8023static int cpu_rt_runtime_write(struct cgroup_subsys_state *css, 8024 struct cftype *cft, s64 val) 8025{ 8026 return sched_group_set_rt_runtime(css_tg(css), val); 8027} 8028 8029static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css, 8030 struct cftype *cft) 8031{ 8032 return sched_group_rt_runtime(css_tg(css)); 8033} 8034 8035static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css, 8036 struct cftype *cftype, u64 rt_period_us) 8037{ 8038 return sched_group_set_rt_period(css_tg(css), rt_period_us); 8039} 8040 8041static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css, 8042 struct cftype *cft) 8043{ 8044 return sched_group_rt_period(css_tg(css)); 8045} 8046#endif /* CONFIG_RT_GROUP_SCHED */ 8047 8048static struct cftype cpu_files[] = { 8049#ifdef CONFIG_FAIR_GROUP_SCHED 8050 { 8051 .name = "shares", 8052 .read_u64 = cpu_shares_read_u64, 8053 .write_u64 = cpu_shares_write_u64, 8054 }, 8055#endif 8056#ifdef CONFIG_CFS_BANDWIDTH 8057 { 8058 .name = "cfs_quota_us", 8059 .read_s64 = cpu_cfs_quota_read_s64, 8060 .write_s64 = cpu_cfs_quota_write_s64, 8061 }, 8062 { 8063 .name = "cfs_period_us", 8064 .read_u64 = cpu_cfs_period_read_u64, 8065 .write_u64 = cpu_cfs_period_write_u64, 8066 }, 8067 { 8068 .name = "stat", 8069 .seq_show = cpu_stats_show, 8070 }, 8071#endif 8072#ifdef CONFIG_RT_GROUP_SCHED 8073 { 8074 .name = "rt_runtime_us", 8075 .read_s64 = cpu_rt_runtime_read, 8076 .write_s64 = cpu_rt_runtime_write, 8077 }, 8078 { 8079 .name = "rt_period_us", 8080 .read_u64 = cpu_rt_period_read_uint, 8081 .write_u64 = cpu_rt_period_write_uint, 8082 }, 8083#endif 8084 { } /* terminate */ 8085}; 8086 8087struct cgroup_subsys cpu_cgrp_subsys = { 8088 .css_alloc = cpu_cgroup_css_alloc, 8089 .css_free = cpu_cgroup_css_free, 8090 .css_online = cpu_cgroup_css_online, 8091 .css_offline = cpu_cgroup_css_offline, 8092 .can_attach = cpu_cgroup_can_attach, 8093 .attach = cpu_cgroup_attach, 8094 .exit = cpu_cgroup_exit, 8095 .base_cftypes = cpu_files, 8096 .early_init = 1, 8097}; 8098 8099#endif /* CONFIG_CGROUP_SCHED */ 8100 8101void dump_cpu_task(int cpu) 8102{ 8103 pr_info("Task dump for CPU %d:\n", cpu); 8104 sched_show_task(cpu_curr(cpu)); 8105} 8106