core.c revision fd2ac4f4a65a7f34b0bc6433fcca1192d7ba8b8e
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 /* Make sure rq->nr_running update is visible after the IPI */ 744 smp_rmb(); 745 746 /* More than one running task need preemption */ 747 if (rq->nr_running > 1) 748 return false; 749 750 return true; 751} 752#endif /* CONFIG_NO_HZ_FULL */ 753 754void sched_avg_update(struct rq *rq) 755{ 756 s64 period = sched_avg_period(); 757 758 while ((s64)(rq_clock(rq) - rq->age_stamp) > period) { 759 /* 760 * Inline assembly required to prevent the compiler 761 * optimising this loop into a divmod call. 762 * See __iter_div_u64_rem() for another example of this. 763 */ 764 asm("" : "+rm" (rq->age_stamp)); 765 rq->age_stamp += period; 766 rq->rt_avg /= 2; 767 } 768} 769 770#endif /* CONFIG_SMP */ 771 772#if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \ 773 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH))) 774/* 775 * Iterate task_group tree rooted at *from, calling @down when first entering a 776 * node and @up when leaving it for the final time. 777 * 778 * Caller must hold rcu_lock or sufficient equivalent. 779 */ 780int walk_tg_tree_from(struct task_group *from, 781 tg_visitor down, tg_visitor up, void *data) 782{ 783 struct task_group *parent, *child; 784 int ret; 785 786 parent = from; 787 788down: 789 ret = (*down)(parent, data); 790 if (ret) 791 goto out; 792 list_for_each_entry_rcu(child, &parent->children, siblings) { 793 parent = child; 794 goto down; 795 796up: 797 continue; 798 } 799 ret = (*up)(parent, data); 800 if (ret || parent == from) 801 goto out; 802 803 child = parent; 804 parent = parent->parent; 805 if (parent) 806 goto up; 807out: 808 return ret; 809} 810 811int tg_nop(struct task_group *tg, void *data) 812{ 813 return 0; 814} 815#endif 816 817static void set_load_weight(struct task_struct *p) 818{ 819 int prio = p->static_prio - MAX_RT_PRIO; 820 struct load_weight *load = &p->se.load; 821 822 /* 823 * SCHED_IDLE tasks get minimal weight: 824 */ 825 if (p->policy == SCHED_IDLE) { 826 load->weight = scale_load(WEIGHT_IDLEPRIO); 827 load->inv_weight = WMULT_IDLEPRIO; 828 return; 829 } 830 831 load->weight = scale_load(prio_to_weight[prio]); 832 load->inv_weight = prio_to_wmult[prio]; 833} 834 835static void enqueue_task(struct rq *rq, struct task_struct *p, int flags) 836{ 837 update_rq_clock(rq); 838 sched_info_queued(rq, p); 839 p->sched_class->enqueue_task(rq, p, flags); 840} 841 842static void dequeue_task(struct rq *rq, struct task_struct *p, int flags) 843{ 844 update_rq_clock(rq); 845 sched_info_dequeued(rq, p); 846 p->sched_class->dequeue_task(rq, p, flags); 847} 848 849void activate_task(struct rq *rq, struct task_struct *p, int flags) 850{ 851 if (task_contributes_to_load(p)) 852 rq->nr_uninterruptible--; 853 854 enqueue_task(rq, p, flags); 855} 856 857void deactivate_task(struct rq *rq, struct task_struct *p, int flags) 858{ 859 if (task_contributes_to_load(p)) 860 rq->nr_uninterruptible++; 861 862 dequeue_task(rq, p, flags); 863} 864 865static void update_rq_clock_task(struct rq *rq, s64 delta) 866{ 867/* 868 * In theory, the compile should just see 0 here, and optimize out the call 869 * to sched_rt_avg_update. But I don't trust it... 870 */ 871#if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING) 872 s64 steal = 0, irq_delta = 0; 873#endif 874#ifdef CONFIG_IRQ_TIME_ACCOUNTING 875 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time; 876 877 /* 878 * Since irq_time is only updated on {soft,}irq_exit, we might run into 879 * this case when a previous update_rq_clock() happened inside a 880 * {soft,}irq region. 881 * 882 * When this happens, we stop ->clock_task and only update the 883 * prev_irq_time stamp to account for the part that fit, so that a next 884 * update will consume the rest. This ensures ->clock_task is 885 * monotonic. 886 * 887 * It does however cause some slight miss-attribution of {soft,}irq 888 * time, a more accurate solution would be to update the irq_time using 889 * the current rq->clock timestamp, except that would require using 890 * atomic ops. 891 */ 892 if (irq_delta > delta) 893 irq_delta = delta; 894 895 rq->prev_irq_time += irq_delta; 896 delta -= irq_delta; 897#endif 898#ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING 899 if (static_key_false((¶virt_steal_rq_enabled))) { 900 steal = paravirt_steal_clock(cpu_of(rq)); 901 steal -= rq->prev_steal_time_rq; 902 903 if (unlikely(steal > delta)) 904 steal = delta; 905 906 rq->prev_steal_time_rq += steal; 907 delta -= steal; 908 } 909#endif 910 911 rq->clock_task += delta; 912 913#if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING) 914 if ((irq_delta + steal) && sched_feat(NONTASK_CAPACITY)) 915 sched_rt_avg_update(rq, irq_delta + steal); 916#endif 917} 918 919void sched_set_stop_task(int cpu, struct task_struct *stop) 920{ 921 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 }; 922 struct task_struct *old_stop = cpu_rq(cpu)->stop; 923 924 if (stop) { 925 /* 926 * Make it appear like a SCHED_FIFO task, its something 927 * userspace knows about and won't get confused about. 928 * 929 * Also, it will make PI more or less work without too 930 * much confusion -- but then, stop work should not 931 * rely on PI working anyway. 932 */ 933 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m); 934 935 stop->sched_class = &stop_sched_class; 936 } 937 938 cpu_rq(cpu)->stop = stop; 939 940 if (old_stop) { 941 /* 942 * Reset it back to a normal scheduling class so that 943 * it can die in pieces. 944 */ 945 old_stop->sched_class = &rt_sched_class; 946 } 947} 948 949/* 950 * __normal_prio - return the priority that is based on the static prio 951 */ 952static inline int __normal_prio(struct task_struct *p) 953{ 954 return p->static_prio; 955} 956 957/* 958 * Calculate the expected normal priority: i.e. priority 959 * without taking RT-inheritance into account. Might be 960 * boosted by interactivity modifiers. Changes upon fork, 961 * setprio syscalls, and whenever the interactivity 962 * estimator recalculates. 963 */ 964static inline int normal_prio(struct task_struct *p) 965{ 966 int prio; 967 968 if (task_has_dl_policy(p)) 969 prio = MAX_DL_PRIO-1; 970 else if (task_has_rt_policy(p)) 971 prio = MAX_RT_PRIO-1 - p->rt_priority; 972 else 973 prio = __normal_prio(p); 974 return prio; 975} 976 977/* 978 * Calculate the current priority, i.e. the priority 979 * taken into account by the scheduler. This value might 980 * be boosted by RT tasks, or might be boosted by 981 * interactivity modifiers. Will be RT if the task got 982 * RT-boosted. If not then it returns p->normal_prio. 983 */ 984static int effective_prio(struct task_struct *p) 985{ 986 p->normal_prio = normal_prio(p); 987 /* 988 * If we are RT tasks or we were boosted to RT priority, 989 * keep the priority unchanged. Otherwise, update priority 990 * to the normal priority: 991 */ 992 if (!rt_prio(p->prio)) 993 return p->normal_prio; 994 return p->prio; 995} 996 997/** 998 * task_curr - is this task currently executing on a CPU? 999 * @p: the task in question. 1000 * 1001 * Return: 1 if the task is currently executing. 0 otherwise. 1002 */ 1003inline int task_curr(const struct task_struct *p) 1004{ 1005 return cpu_curr(task_cpu(p)) == p; 1006} 1007 1008static inline void check_class_changed(struct rq *rq, struct task_struct *p, 1009 const struct sched_class *prev_class, 1010 int oldprio) 1011{ 1012 if (prev_class != p->sched_class) { 1013 if (prev_class->switched_from) 1014 prev_class->switched_from(rq, p); 1015 p->sched_class->switched_to(rq, p); 1016 } else if (oldprio != p->prio || dl_task(p)) 1017 p->sched_class->prio_changed(rq, p, oldprio); 1018} 1019 1020void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags) 1021{ 1022 const struct sched_class *class; 1023 1024 if (p->sched_class == rq->curr->sched_class) { 1025 rq->curr->sched_class->check_preempt_curr(rq, p, flags); 1026 } else { 1027 for_each_class(class) { 1028 if (class == rq->curr->sched_class) 1029 break; 1030 if (class == p->sched_class) { 1031 resched_task(rq->curr); 1032 break; 1033 } 1034 } 1035 } 1036 1037 /* 1038 * A queue event has occurred, and we're going to schedule. In 1039 * this case, we can save a useless back to back clock update. 1040 */ 1041 if (rq->curr->on_rq && test_tsk_need_resched(rq->curr)) 1042 rq->skip_clock_update = 1; 1043} 1044 1045#ifdef CONFIG_SMP 1046void set_task_cpu(struct task_struct *p, unsigned int new_cpu) 1047{ 1048#ifdef CONFIG_SCHED_DEBUG 1049 /* 1050 * We should never call set_task_cpu() on a blocked task, 1051 * ttwu() will sort out the placement. 1052 */ 1053 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING && 1054 !(task_preempt_count(p) & PREEMPT_ACTIVE)); 1055 1056#ifdef CONFIG_LOCKDEP 1057 /* 1058 * The caller should hold either p->pi_lock or rq->lock, when changing 1059 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks. 1060 * 1061 * sched_move_task() holds both and thus holding either pins the cgroup, 1062 * see task_group(). 1063 * 1064 * Furthermore, all task_rq users should acquire both locks, see 1065 * task_rq_lock(). 1066 */ 1067 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) || 1068 lockdep_is_held(&task_rq(p)->lock))); 1069#endif 1070#endif 1071 1072 trace_sched_migrate_task(p, new_cpu); 1073 1074 if (task_cpu(p) != new_cpu) { 1075 if (p->sched_class->migrate_task_rq) 1076 p->sched_class->migrate_task_rq(p, new_cpu); 1077 p->se.nr_migrations++; 1078 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, NULL, 0); 1079 } 1080 1081 __set_task_cpu(p, new_cpu); 1082} 1083 1084static void __migrate_swap_task(struct task_struct *p, int cpu) 1085{ 1086 if (p->on_rq) { 1087 struct rq *src_rq, *dst_rq; 1088 1089 src_rq = task_rq(p); 1090 dst_rq = cpu_rq(cpu); 1091 1092 deactivate_task(src_rq, p, 0); 1093 set_task_cpu(p, cpu); 1094 activate_task(dst_rq, p, 0); 1095 check_preempt_curr(dst_rq, p, 0); 1096 } else { 1097 /* 1098 * Task isn't running anymore; make it appear like we migrated 1099 * it before it went to sleep. This means on wakeup we make the 1100 * previous cpu our targer instead of where it really is. 1101 */ 1102 p->wake_cpu = cpu; 1103 } 1104} 1105 1106struct migration_swap_arg { 1107 struct task_struct *src_task, *dst_task; 1108 int src_cpu, dst_cpu; 1109}; 1110 1111static int migrate_swap_stop(void *data) 1112{ 1113 struct migration_swap_arg *arg = data; 1114 struct rq *src_rq, *dst_rq; 1115 int ret = -EAGAIN; 1116 1117 src_rq = cpu_rq(arg->src_cpu); 1118 dst_rq = cpu_rq(arg->dst_cpu); 1119 1120 double_raw_lock(&arg->src_task->pi_lock, 1121 &arg->dst_task->pi_lock); 1122 double_rq_lock(src_rq, dst_rq); 1123 if (task_cpu(arg->dst_task) != arg->dst_cpu) 1124 goto unlock; 1125 1126 if (task_cpu(arg->src_task) != arg->src_cpu) 1127 goto unlock; 1128 1129 if (!cpumask_test_cpu(arg->dst_cpu, tsk_cpus_allowed(arg->src_task))) 1130 goto unlock; 1131 1132 if (!cpumask_test_cpu(arg->src_cpu, tsk_cpus_allowed(arg->dst_task))) 1133 goto unlock; 1134 1135 __migrate_swap_task(arg->src_task, arg->dst_cpu); 1136 __migrate_swap_task(arg->dst_task, arg->src_cpu); 1137 1138 ret = 0; 1139 1140unlock: 1141 double_rq_unlock(src_rq, dst_rq); 1142 raw_spin_unlock(&arg->dst_task->pi_lock); 1143 raw_spin_unlock(&arg->src_task->pi_lock); 1144 1145 return ret; 1146} 1147 1148/* 1149 * Cross migrate two tasks 1150 */ 1151int migrate_swap(struct task_struct *cur, struct task_struct *p) 1152{ 1153 struct migration_swap_arg arg; 1154 int ret = -EINVAL; 1155 1156 arg = (struct migration_swap_arg){ 1157 .src_task = cur, 1158 .src_cpu = task_cpu(cur), 1159 .dst_task = p, 1160 .dst_cpu = task_cpu(p), 1161 }; 1162 1163 if (arg.src_cpu == arg.dst_cpu) 1164 goto out; 1165 1166 /* 1167 * These three tests are all lockless; this is OK since all of them 1168 * will be re-checked with proper locks held further down the line. 1169 */ 1170 if (!cpu_active(arg.src_cpu) || !cpu_active(arg.dst_cpu)) 1171 goto out; 1172 1173 if (!cpumask_test_cpu(arg.dst_cpu, tsk_cpus_allowed(arg.src_task))) 1174 goto out; 1175 1176 if (!cpumask_test_cpu(arg.src_cpu, tsk_cpus_allowed(arg.dst_task))) 1177 goto out; 1178 1179 trace_sched_swap_numa(cur, arg.src_cpu, p, arg.dst_cpu); 1180 ret = stop_two_cpus(arg.dst_cpu, arg.src_cpu, migrate_swap_stop, &arg); 1181 1182out: 1183 return ret; 1184} 1185 1186struct migration_arg { 1187 struct task_struct *task; 1188 int dest_cpu; 1189}; 1190 1191static int migration_cpu_stop(void *data); 1192 1193/* 1194 * wait_task_inactive - wait for a thread to unschedule. 1195 * 1196 * If @match_state is nonzero, it's the @p->state value just checked and 1197 * not expected to change. If it changes, i.e. @p might have woken up, 1198 * then return zero. When we succeed in waiting for @p to be off its CPU, 1199 * we return a positive number (its total switch count). If a second call 1200 * a short while later returns the same number, the caller can be sure that 1201 * @p has remained unscheduled the whole time. 1202 * 1203 * The caller must ensure that the task *will* unschedule sometime soon, 1204 * else this function might spin for a *long* time. This function can't 1205 * be called with interrupts off, or it may introduce deadlock with 1206 * smp_call_function() if an IPI is sent by the same process we are 1207 * waiting to become inactive. 1208 */ 1209unsigned long wait_task_inactive(struct task_struct *p, long match_state) 1210{ 1211 unsigned long flags; 1212 int running, on_rq; 1213 unsigned long ncsw; 1214 struct rq *rq; 1215 1216 for (;;) { 1217 /* 1218 * We do the initial early heuristics without holding 1219 * any task-queue locks at all. We'll only try to get 1220 * the runqueue lock when things look like they will 1221 * work out! 1222 */ 1223 rq = task_rq(p); 1224 1225 /* 1226 * If the task is actively running on another CPU 1227 * still, just relax and busy-wait without holding 1228 * any locks. 1229 * 1230 * NOTE! Since we don't hold any locks, it's not 1231 * even sure that "rq" stays as the right runqueue! 1232 * But we don't care, since "task_running()" will 1233 * return false if the runqueue has changed and p 1234 * is actually now running somewhere else! 1235 */ 1236 while (task_running(rq, p)) { 1237 if (match_state && unlikely(p->state != match_state)) 1238 return 0; 1239 cpu_relax(); 1240 } 1241 1242 /* 1243 * Ok, time to look more closely! We need the rq 1244 * lock now, to be *sure*. If we're wrong, we'll 1245 * just go back and repeat. 1246 */ 1247 rq = task_rq_lock(p, &flags); 1248 trace_sched_wait_task(p); 1249 running = task_running(rq, p); 1250 on_rq = p->on_rq; 1251 ncsw = 0; 1252 if (!match_state || p->state == match_state) 1253 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */ 1254 task_rq_unlock(rq, p, &flags); 1255 1256 /* 1257 * If it changed from the expected state, bail out now. 1258 */ 1259 if (unlikely(!ncsw)) 1260 break; 1261 1262 /* 1263 * Was it really running after all now that we 1264 * checked with the proper locks actually held? 1265 * 1266 * Oops. Go back and try again.. 1267 */ 1268 if (unlikely(running)) { 1269 cpu_relax(); 1270 continue; 1271 } 1272 1273 /* 1274 * It's not enough that it's not actively running, 1275 * it must be off the runqueue _entirely_, and not 1276 * preempted! 1277 * 1278 * So if it was still runnable (but just not actively 1279 * running right now), it's preempted, and we should 1280 * yield - it could be a while. 1281 */ 1282 if (unlikely(on_rq)) { 1283 ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ); 1284 1285 set_current_state(TASK_UNINTERRUPTIBLE); 1286 schedule_hrtimeout(&to, HRTIMER_MODE_REL); 1287 continue; 1288 } 1289 1290 /* 1291 * Ahh, all good. It wasn't running, and it wasn't 1292 * runnable, which means that it will never become 1293 * running in the future either. We're all done! 1294 */ 1295 break; 1296 } 1297 1298 return ncsw; 1299} 1300 1301/*** 1302 * kick_process - kick a running thread to enter/exit the kernel 1303 * @p: the to-be-kicked thread 1304 * 1305 * Cause a process which is running on another CPU to enter 1306 * kernel-mode, without any delay. (to get signals handled.) 1307 * 1308 * NOTE: this function doesn't have to take the runqueue lock, 1309 * because all it wants to ensure is that the remote task enters 1310 * the kernel. If the IPI races and the task has been migrated 1311 * to another CPU then no harm is done and the purpose has been 1312 * achieved as well. 1313 */ 1314void kick_process(struct task_struct *p) 1315{ 1316 int cpu; 1317 1318 preempt_disable(); 1319 cpu = task_cpu(p); 1320 if ((cpu != smp_processor_id()) && task_curr(p)) 1321 smp_send_reschedule(cpu); 1322 preempt_enable(); 1323} 1324EXPORT_SYMBOL_GPL(kick_process); 1325#endif /* CONFIG_SMP */ 1326 1327#ifdef CONFIG_SMP 1328/* 1329 * ->cpus_allowed is protected by both rq->lock and p->pi_lock 1330 */ 1331static int select_fallback_rq(int cpu, struct task_struct *p) 1332{ 1333 int nid = cpu_to_node(cpu); 1334 const struct cpumask *nodemask = NULL; 1335 enum { cpuset, possible, fail } state = cpuset; 1336 int dest_cpu; 1337 1338 /* 1339 * If the node that the cpu is on has been offlined, cpu_to_node() 1340 * will return -1. There is no cpu on the node, and we should 1341 * select the cpu on the other node. 1342 */ 1343 if (nid != -1) { 1344 nodemask = cpumask_of_node(nid); 1345 1346 /* Look for allowed, online CPU in same node. */ 1347 for_each_cpu(dest_cpu, nodemask) { 1348 if (!cpu_online(dest_cpu)) 1349 continue; 1350 if (!cpu_active(dest_cpu)) 1351 continue; 1352 if (cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p))) 1353 return dest_cpu; 1354 } 1355 } 1356 1357 for (;;) { 1358 /* Any allowed, online CPU? */ 1359 for_each_cpu(dest_cpu, tsk_cpus_allowed(p)) { 1360 if (!cpu_online(dest_cpu)) 1361 continue; 1362 if (!cpu_active(dest_cpu)) 1363 continue; 1364 goto out; 1365 } 1366 1367 switch (state) { 1368 case cpuset: 1369 /* No more Mr. Nice Guy. */ 1370 cpuset_cpus_allowed_fallback(p); 1371 state = possible; 1372 break; 1373 1374 case possible: 1375 do_set_cpus_allowed(p, cpu_possible_mask); 1376 state = fail; 1377 break; 1378 1379 case fail: 1380 BUG(); 1381 break; 1382 } 1383 } 1384 1385out: 1386 if (state != cpuset) { 1387 /* 1388 * Don't tell them about moving exiting tasks or 1389 * kernel threads (both mm NULL), since they never 1390 * leave kernel. 1391 */ 1392 if (p->mm && printk_ratelimit()) { 1393 printk_deferred("process %d (%s) no longer affine to cpu%d\n", 1394 task_pid_nr(p), p->comm, cpu); 1395 } 1396 } 1397 1398 return dest_cpu; 1399} 1400 1401/* 1402 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable. 1403 */ 1404static inline 1405int select_task_rq(struct task_struct *p, int cpu, int sd_flags, int wake_flags) 1406{ 1407 cpu = p->sched_class->select_task_rq(p, cpu, sd_flags, wake_flags); 1408 1409 /* 1410 * In order not to call set_task_cpu() on a blocking task we need 1411 * to rely on ttwu() to place the task on a valid ->cpus_allowed 1412 * cpu. 1413 * 1414 * Since this is common to all placement strategies, this lives here. 1415 * 1416 * [ this allows ->select_task() to simply return task_cpu(p) and 1417 * not worry about this generic constraint ] 1418 */ 1419 if (unlikely(!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)) || 1420 !cpu_online(cpu))) 1421 cpu = select_fallback_rq(task_cpu(p), p); 1422 1423 return cpu; 1424} 1425 1426static void update_avg(u64 *avg, u64 sample) 1427{ 1428 s64 diff = sample - *avg; 1429 *avg += diff >> 3; 1430} 1431#endif 1432 1433static void 1434ttwu_stat(struct task_struct *p, int cpu, int wake_flags) 1435{ 1436#ifdef CONFIG_SCHEDSTATS 1437 struct rq *rq = this_rq(); 1438 1439#ifdef CONFIG_SMP 1440 int this_cpu = smp_processor_id(); 1441 1442 if (cpu == this_cpu) { 1443 schedstat_inc(rq, ttwu_local); 1444 schedstat_inc(p, se.statistics.nr_wakeups_local); 1445 } else { 1446 struct sched_domain *sd; 1447 1448 schedstat_inc(p, se.statistics.nr_wakeups_remote); 1449 rcu_read_lock(); 1450 for_each_domain(this_cpu, sd) { 1451 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) { 1452 schedstat_inc(sd, ttwu_wake_remote); 1453 break; 1454 } 1455 } 1456 rcu_read_unlock(); 1457 } 1458 1459 if (wake_flags & WF_MIGRATED) 1460 schedstat_inc(p, se.statistics.nr_wakeups_migrate); 1461 1462#endif /* CONFIG_SMP */ 1463 1464 schedstat_inc(rq, ttwu_count); 1465 schedstat_inc(p, se.statistics.nr_wakeups); 1466 1467 if (wake_flags & WF_SYNC) 1468 schedstat_inc(p, se.statistics.nr_wakeups_sync); 1469 1470#endif /* CONFIG_SCHEDSTATS */ 1471} 1472 1473static void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags) 1474{ 1475 activate_task(rq, p, en_flags); 1476 p->on_rq = 1; 1477 1478 /* if a worker is waking up, notify workqueue */ 1479 if (p->flags & PF_WQ_WORKER) 1480 wq_worker_waking_up(p, cpu_of(rq)); 1481} 1482 1483/* 1484 * Mark the task runnable and perform wakeup-preemption. 1485 */ 1486static void 1487ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags) 1488{ 1489 check_preempt_curr(rq, p, wake_flags); 1490 trace_sched_wakeup(p, true); 1491 1492 p->state = TASK_RUNNING; 1493#ifdef CONFIG_SMP 1494 if (p->sched_class->task_woken) 1495 p->sched_class->task_woken(rq, p); 1496 1497 if (rq->idle_stamp) { 1498 u64 delta = rq_clock(rq) - rq->idle_stamp; 1499 u64 max = 2*rq->max_idle_balance_cost; 1500 1501 update_avg(&rq->avg_idle, delta); 1502 1503 if (rq->avg_idle > max) 1504 rq->avg_idle = max; 1505 1506 rq->idle_stamp = 0; 1507 } 1508#endif 1509} 1510 1511static void 1512ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags) 1513{ 1514#ifdef CONFIG_SMP 1515 if (p->sched_contributes_to_load) 1516 rq->nr_uninterruptible--; 1517#endif 1518 1519 ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_WAKING); 1520 ttwu_do_wakeup(rq, p, wake_flags); 1521} 1522 1523/* 1524 * Called in case the task @p isn't fully descheduled from its runqueue, 1525 * in this case we must do a remote wakeup. Its a 'light' wakeup though, 1526 * since all we need to do is flip p->state to TASK_RUNNING, since 1527 * the task is still ->on_rq. 1528 */ 1529static int ttwu_remote(struct task_struct *p, int wake_flags) 1530{ 1531 struct rq *rq; 1532 int ret = 0; 1533 1534 rq = __task_rq_lock(p); 1535 if (p->on_rq) { 1536 /* check_preempt_curr() may use rq clock */ 1537 update_rq_clock(rq); 1538 ttwu_do_wakeup(rq, p, wake_flags); 1539 ret = 1; 1540 } 1541 __task_rq_unlock(rq); 1542 1543 return ret; 1544} 1545 1546#ifdef CONFIG_SMP 1547void sched_ttwu_pending(void) 1548{ 1549 struct rq *rq = this_rq(); 1550 struct llist_node *llist = llist_del_all(&rq->wake_list); 1551 struct task_struct *p; 1552 unsigned long flags; 1553 1554 if (!llist) 1555 return; 1556 1557 raw_spin_lock_irqsave(&rq->lock, flags); 1558 1559 while (llist) { 1560 p = llist_entry(llist, struct task_struct, wake_entry); 1561 llist = llist_next(llist); 1562 ttwu_do_activate(rq, p, 0); 1563 } 1564 1565 raw_spin_unlock_irqrestore(&rq->lock, flags); 1566} 1567 1568void scheduler_ipi(void) 1569{ 1570 /* 1571 * Fold TIF_NEED_RESCHED into the preempt_count; anybody setting 1572 * TIF_NEED_RESCHED remotely (for the first time) will also send 1573 * this IPI. 1574 */ 1575 preempt_fold_need_resched(); 1576 1577 if (llist_empty(&this_rq()->wake_list) && !got_nohz_idle_kick()) 1578 return; 1579 1580 /* 1581 * Not all reschedule IPI handlers call irq_enter/irq_exit, since 1582 * traditionally all their work was done from the interrupt return 1583 * path. Now that we actually do some work, we need to make sure 1584 * we do call them. 1585 * 1586 * Some archs already do call them, luckily irq_enter/exit nest 1587 * properly. 1588 * 1589 * Arguably we should visit all archs and update all handlers, 1590 * however a fair share of IPIs are still resched only so this would 1591 * somewhat pessimize the simple resched case. 1592 */ 1593 irq_enter(); 1594 sched_ttwu_pending(); 1595 1596 /* 1597 * Check if someone kicked us for doing the nohz idle load balance. 1598 */ 1599 if (unlikely(got_nohz_idle_kick())) { 1600 this_rq()->idle_balance = 1; 1601 raise_softirq_irqoff(SCHED_SOFTIRQ); 1602 } 1603 irq_exit(); 1604} 1605 1606static void ttwu_queue_remote(struct task_struct *p, int cpu) 1607{ 1608 struct rq *rq = cpu_rq(cpu); 1609 1610 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list)) { 1611 if (!set_nr_if_polling(rq->idle)) 1612 smp_send_reschedule(cpu); 1613 else 1614 trace_sched_wake_idle_without_ipi(cpu); 1615 } 1616} 1617 1618bool cpus_share_cache(int this_cpu, int that_cpu) 1619{ 1620 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu); 1621} 1622#endif /* CONFIG_SMP */ 1623 1624static void ttwu_queue(struct task_struct *p, int cpu) 1625{ 1626 struct rq *rq = cpu_rq(cpu); 1627 1628#if defined(CONFIG_SMP) 1629 if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) { 1630 sched_clock_cpu(cpu); /* sync clocks x-cpu */ 1631 ttwu_queue_remote(p, cpu); 1632 return; 1633 } 1634#endif 1635 1636 raw_spin_lock(&rq->lock); 1637 ttwu_do_activate(rq, p, 0); 1638 raw_spin_unlock(&rq->lock); 1639} 1640 1641/** 1642 * try_to_wake_up - wake up a thread 1643 * @p: the thread to be awakened 1644 * @state: the mask of task states that can be woken 1645 * @wake_flags: wake modifier flags (WF_*) 1646 * 1647 * Put it on the run-queue if it's not already there. The "current" 1648 * thread is always on the run-queue (except when the actual 1649 * re-schedule is in progress), and as such you're allowed to do 1650 * the simpler "current->state = TASK_RUNNING" to mark yourself 1651 * runnable without the overhead of this. 1652 * 1653 * Return: %true if @p was woken up, %false if it was already running. 1654 * or @state didn't match @p's state. 1655 */ 1656static int 1657try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags) 1658{ 1659 unsigned long flags; 1660 int cpu, success = 0; 1661 1662 /* 1663 * If we are going to wake up a thread waiting for CONDITION we 1664 * need to ensure that CONDITION=1 done by the caller can not be 1665 * reordered with p->state check below. This pairs with mb() in 1666 * set_current_state() the waiting thread does. 1667 */ 1668 smp_mb__before_spinlock(); 1669 raw_spin_lock_irqsave(&p->pi_lock, flags); 1670 if (!(p->state & state)) 1671 goto out; 1672 1673 success = 1; /* we're going to change ->state */ 1674 cpu = task_cpu(p); 1675 1676 if (p->on_rq && ttwu_remote(p, wake_flags)) 1677 goto stat; 1678 1679#ifdef CONFIG_SMP 1680 /* 1681 * If the owning (remote) cpu is still in the middle of schedule() with 1682 * this task as prev, wait until its done referencing the task. 1683 */ 1684 while (p->on_cpu) 1685 cpu_relax(); 1686 /* 1687 * Pairs with the smp_wmb() in finish_lock_switch(). 1688 */ 1689 smp_rmb(); 1690 1691 p->sched_contributes_to_load = !!task_contributes_to_load(p); 1692 p->state = TASK_WAKING; 1693 1694 if (p->sched_class->task_waking) 1695 p->sched_class->task_waking(p); 1696 1697 cpu = select_task_rq(p, p->wake_cpu, SD_BALANCE_WAKE, wake_flags); 1698 if (task_cpu(p) != cpu) { 1699 wake_flags |= WF_MIGRATED; 1700 set_task_cpu(p, cpu); 1701 } 1702#endif /* CONFIG_SMP */ 1703 1704 ttwu_queue(p, cpu); 1705stat: 1706 ttwu_stat(p, cpu, wake_flags); 1707out: 1708 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 1709 1710 return success; 1711} 1712 1713/** 1714 * try_to_wake_up_local - try to wake up a local task with rq lock held 1715 * @p: the thread to be awakened 1716 * 1717 * Put @p on the run-queue if it's not already there. The caller must 1718 * ensure that this_rq() is locked, @p is bound to this_rq() and not 1719 * the current task. 1720 */ 1721static void try_to_wake_up_local(struct task_struct *p) 1722{ 1723 struct rq *rq = task_rq(p); 1724 1725 if (WARN_ON_ONCE(rq != this_rq()) || 1726 WARN_ON_ONCE(p == current)) 1727 return; 1728 1729 lockdep_assert_held(&rq->lock); 1730 1731 if (!raw_spin_trylock(&p->pi_lock)) { 1732 raw_spin_unlock(&rq->lock); 1733 raw_spin_lock(&p->pi_lock); 1734 raw_spin_lock(&rq->lock); 1735 } 1736 1737 if (!(p->state & TASK_NORMAL)) 1738 goto out; 1739 1740 if (!p->on_rq) 1741 ttwu_activate(rq, p, ENQUEUE_WAKEUP); 1742 1743 ttwu_do_wakeup(rq, p, 0); 1744 ttwu_stat(p, smp_processor_id(), 0); 1745out: 1746 raw_spin_unlock(&p->pi_lock); 1747} 1748 1749/** 1750 * wake_up_process - Wake up a specific process 1751 * @p: The process to be woken up. 1752 * 1753 * Attempt to wake up the nominated process and move it to the set of runnable 1754 * processes. 1755 * 1756 * Return: 1 if the process was woken up, 0 if it was already running. 1757 * 1758 * It may be assumed that this function implies a write memory barrier before 1759 * changing the task state if and only if any tasks are woken up. 1760 */ 1761int wake_up_process(struct task_struct *p) 1762{ 1763 WARN_ON(task_is_stopped_or_traced(p)); 1764 return try_to_wake_up(p, TASK_NORMAL, 0); 1765} 1766EXPORT_SYMBOL(wake_up_process); 1767 1768int wake_up_state(struct task_struct *p, unsigned int state) 1769{ 1770 return try_to_wake_up(p, state, 0); 1771} 1772 1773/* 1774 * Perform scheduler related setup for a newly forked process p. 1775 * p is forked by current. 1776 * 1777 * __sched_fork() is basic setup used by init_idle() too: 1778 */ 1779static void __sched_fork(unsigned long clone_flags, struct task_struct *p) 1780{ 1781 p->on_rq = 0; 1782 1783 p->se.on_rq = 0; 1784 p->se.exec_start = 0; 1785 p->se.sum_exec_runtime = 0; 1786 p->se.prev_sum_exec_runtime = 0; 1787 p->se.nr_migrations = 0; 1788 p->se.vruntime = 0; 1789 INIT_LIST_HEAD(&p->se.group_node); 1790 1791#ifdef CONFIG_SCHEDSTATS 1792 memset(&p->se.statistics, 0, sizeof(p->se.statistics)); 1793#endif 1794 1795 RB_CLEAR_NODE(&p->dl.rb_node); 1796 hrtimer_init(&p->dl.dl_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); 1797 p->dl.dl_runtime = p->dl.runtime = 0; 1798 p->dl.dl_deadline = p->dl.deadline = 0; 1799 p->dl.dl_period = 0; 1800 p->dl.flags = 0; 1801 1802 INIT_LIST_HEAD(&p->rt.run_list); 1803 1804#ifdef CONFIG_PREEMPT_NOTIFIERS 1805 INIT_HLIST_HEAD(&p->preempt_notifiers); 1806#endif 1807 1808#ifdef CONFIG_NUMA_BALANCING 1809 if (p->mm && atomic_read(&p->mm->mm_users) == 1) { 1810 p->mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay); 1811 p->mm->numa_scan_seq = 0; 1812 } 1813 1814 if (clone_flags & CLONE_VM) 1815 p->numa_preferred_nid = current->numa_preferred_nid; 1816 else 1817 p->numa_preferred_nid = -1; 1818 1819 p->node_stamp = 0ULL; 1820 p->numa_scan_seq = p->mm ? p->mm->numa_scan_seq : 0; 1821 p->numa_scan_period = sysctl_numa_balancing_scan_delay; 1822 p->numa_work.next = &p->numa_work; 1823 p->numa_faults_memory = NULL; 1824 p->numa_faults_buffer_memory = NULL; 1825 p->last_task_numa_placement = 0; 1826 p->last_sum_exec_runtime = 0; 1827 1828 INIT_LIST_HEAD(&p->numa_entry); 1829 p->numa_group = NULL; 1830#endif /* CONFIG_NUMA_BALANCING */ 1831} 1832 1833#ifdef CONFIG_NUMA_BALANCING 1834#ifdef CONFIG_SCHED_DEBUG 1835void set_numabalancing_state(bool enabled) 1836{ 1837 if (enabled) 1838 sched_feat_set("NUMA"); 1839 else 1840 sched_feat_set("NO_NUMA"); 1841} 1842#else 1843__read_mostly bool numabalancing_enabled; 1844 1845void set_numabalancing_state(bool enabled) 1846{ 1847 numabalancing_enabled = enabled; 1848} 1849#endif /* CONFIG_SCHED_DEBUG */ 1850 1851#ifdef CONFIG_PROC_SYSCTL 1852int sysctl_numa_balancing(struct ctl_table *table, int write, 1853 void __user *buffer, size_t *lenp, loff_t *ppos) 1854{ 1855 struct ctl_table t; 1856 int err; 1857 int state = numabalancing_enabled; 1858 1859 if (write && !capable(CAP_SYS_ADMIN)) 1860 return -EPERM; 1861 1862 t = *table; 1863 t.data = &state; 1864 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos); 1865 if (err < 0) 1866 return err; 1867 if (write) 1868 set_numabalancing_state(state); 1869 return err; 1870} 1871#endif 1872#endif 1873 1874/* 1875 * fork()/clone()-time setup: 1876 */ 1877int sched_fork(unsigned long clone_flags, struct task_struct *p) 1878{ 1879 unsigned long flags; 1880 int cpu = get_cpu(); 1881 1882 __sched_fork(clone_flags, p); 1883 /* 1884 * We mark the process as running here. This guarantees that 1885 * nobody will actually run it, and a signal or other external 1886 * event cannot wake it up and insert it on the runqueue either. 1887 */ 1888 p->state = TASK_RUNNING; 1889 1890 /* 1891 * Make sure we do not leak PI boosting priority to the child. 1892 */ 1893 p->prio = current->normal_prio; 1894 1895 /* 1896 * Revert to default priority/policy on fork if requested. 1897 */ 1898 if (unlikely(p->sched_reset_on_fork)) { 1899 if (task_has_dl_policy(p) || task_has_rt_policy(p)) { 1900 p->policy = SCHED_NORMAL; 1901 p->static_prio = NICE_TO_PRIO(0); 1902 p->rt_priority = 0; 1903 } else if (PRIO_TO_NICE(p->static_prio) < 0) 1904 p->static_prio = NICE_TO_PRIO(0); 1905 1906 p->prio = p->normal_prio = __normal_prio(p); 1907 set_load_weight(p); 1908 1909 /* 1910 * We don't need the reset flag anymore after the fork. It has 1911 * fulfilled its duty: 1912 */ 1913 p->sched_reset_on_fork = 0; 1914 } 1915 1916 if (dl_prio(p->prio)) { 1917 put_cpu(); 1918 return -EAGAIN; 1919 } else if (rt_prio(p->prio)) { 1920 p->sched_class = &rt_sched_class; 1921 } else { 1922 p->sched_class = &fair_sched_class; 1923 } 1924 1925 if (p->sched_class->task_fork) 1926 p->sched_class->task_fork(p); 1927 1928 /* 1929 * The child is not yet in the pid-hash so no cgroup attach races, 1930 * and the cgroup is pinned to this child due to cgroup_fork() 1931 * is ran before sched_fork(). 1932 * 1933 * Silence PROVE_RCU. 1934 */ 1935 raw_spin_lock_irqsave(&p->pi_lock, flags); 1936 set_task_cpu(p, cpu); 1937 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 1938 1939#if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT) 1940 if (likely(sched_info_on())) 1941 memset(&p->sched_info, 0, sizeof(p->sched_info)); 1942#endif 1943#if defined(CONFIG_SMP) 1944 p->on_cpu = 0; 1945#endif 1946 init_task_preempt_count(p); 1947#ifdef CONFIG_SMP 1948 plist_node_init(&p->pushable_tasks, MAX_PRIO); 1949 RB_CLEAR_NODE(&p->pushable_dl_tasks); 1950#endif 1951 1952 put_cpu(); 1953 return 0; 1954} 1955 1956unsigned long to_ratio(u64 period, u64 runtime) 1957{ 1958 if (runtime == RUNTIME_INF) 1959 return 1ULL << 20; 1960 1961 /* 1962 * Doing this here saves a lot of checks in all 1963 * the calling paths, and returning zero seems 1964 * safe for them anyway. 1965 */ 1966 if (period == 0) 1967 return 0; 1968 1969 return div64_u64(runtime << 20, period); 1970} 1971 1972#ifdef CONFIG_SMP 1973inline struct dl_bw *dl_bw_of(int i) 1974{ 1975 return &cpu_rq(i)->rd->dl_bw; 1976} 1977 1978static inline int dl_bw_cpus(int i) 1979{ 1980 struct root_domain *rd = cpu_rq(i)->rd; 1981 int cpus = 0; 1982 1983 for_each_cpu_and(i, rd->span, cpu_active_mask) 1984 cpus++; 1985 1986 return cpus; 1987} 1988#else 1989inline struct dl_bw *dl_bw_of(int i) 1990{ 1991 return &cpu_rq(i)->dl.dl_bw; 1992} 1993 1994static inline int dl_bw_cpus(int i) 1995{ 1996 return 1; 1997} 1998#endif 1999 2000static inline 2001void __dl_clear(struct dl_bw *dl_b, u64 tsk_bw) 2002{ 2003 dl_b->total_bw -= tsk_bw; 2004} 2005 2006static inline 2007void __dl_add(struct dl_bw *dl_b, u64 tsk_bw) 2008{ 2009 dl_b->total_bw += tsk_bw; 2010} 2011 2012static inline 2013bool __dl_overflow(struct dl_bw *dl_b, int cpus, u64 old_bw, u64 new_bw) 2014{ 2015 return dl_b->bw != -1 && 2016 dl_b->bw * cpus < dl_b->total_bw - old_bw + new_bw; 2017} 2018 2019/* 2020 * We must be sure that accepting a new task (or allowing changing the 2021 * parameters of an existing one) is consistent with the bandwidth 2022 * constraints. If yes, this function also accordingly updates the currently 2023 * allocated bandwidth to reflect the new situation. 2024 * 2025 * This function is called while holding p's rq->lock. 2026 */ 2027static int dl_overflow(struct task_struct *p, int policy, 2028 const struct sched_attr *attr) 2029{ 2030 2031 struct dl_bw *dl_b = dl_bw_of(task_cpu(p)); 2032 u64 period = attr->sched_period ?: attr->sched_deadline; 2033 u64 runtime = attr->sched_runtime; 2034 u64 new_bw = dl_policy(policy) ? to_ratio(period, runtime) : 0; 2035 int cpus, err = -1; 2036 2037 if (new_bw == p->dl.dl_bw) 2038 return 0; 2039 2040 /* 2041 * Either if a task, enters, leave, or stays -deadline but changes 2042 * its parameters, we may need to update accordingly the total 2043 * allocated bandwidth of the container. 2044 */ 2045 raw_spin_lock(&dl_b->lock); 2046 cpus = dl_bw_cpus(task_cpu(p)); 2047 if (dl_policy(policy) && !task_has_dl_policy(p) && 2048 !__dl_overflow(dl_b, cpus, 0, new_bw)) { 2049 __dl_add(dl_b, new_bw); 2050 err = 0; 2051 } else if (dl_policy(policy) && task_has_dl_policy(p) && 2052 !__dl_overflow(dl_b, cpus, p->dl.dl_bw, new_bw)) { 2053 __dl_clear(dl_b, p->dl.dl_bw); 2054 __dl_add(dl_b, new_bw); 2055 err = 0; 2056 } else if (!dl_policy(policy) && task_has_dl_policy(p)) { 2057 __dl_clear(dl_b, p->dl.dl_bw); 2058 err = 0; 2059 } 2060 raw_spin_unlock(&dl_b->lock); 2061 2062 return err; 2063} 2064 2065extern void init_dl_bw(struct dl_bw *dl_b); 2066 2067/* 2068 * wake_up_new_task - wake up a newly created task for the first time. 2069 * 2070 * This function will do some initial scheduler statistics housekeeping 2071 * that must be done for every newly created context, then puts the task 2072 * on the runqueue and wakes it. 2073 */ 2074void wake_up_new_task(struct task_struct *p) 2075{ 2076 unsigned long flags; 2077 struct rq *rq; 2078 2079 raw_spin_lock_irqsave(&p->pi_lock, flags); 2080#ifdef CONFIG_SMP 2081 /* 2082 * Fork balancing, do it here and not earlier because: 2083 * - cpus_allowed can change in the fork path 2084 * - any previously selected cpu might disappear through hotplug 2085 */ 2086 set_task_cpu(p, select_task_rq(p, task_cpu(p), SD_BALANCE_FORK, 0)); 2087#endif 2088 2089 /* Initialize new task's runnable average */ 2090 init_task_runnable_average(p); 2091 rq = __task_rq_lock(p); 2092 activate_task(rq, p, 0); 2093 p->on_rq = 1; 2094 trace_sched_wakeup_new(p, true); 2095 check_preempt_curr(rq, p, WF_FORK); 2096#ifdef CONFIG_SMP 2097 if (p->sched_class->task_woken) 2098 p->sched_class->task_woken(rq, p); 2099#endif 2100 task_rq_unlock(rq, p, &flags); 2101} 2102 2103#ifdef CONFIG_PREEMPT_NOTIFIERS 2104 2105/** 2106 * preempt_notifier_register - tell me when current is being preempted & rescheduled 2107 * @notifier: notifier struct to register 2108 */ 2109void preempt_notifier_register(struct preempt_notifier *notifier) 2110{ 2111 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers); 2112} 2113EXPORT_SYMBOL_GPL(preempt_notifier_register); 2114 2115/** 2116 * preempt_notifier_unregister - no longer interested in preemption notifications 2117 * @notifier: notifier struct to unregister 2118 * 2119 * This is safe to call from within a preemption notifier. 2120 */ 2121void preempt_notifier_unregister(struct preempt_notifier *notifier) 2122{ 2123 hlist_del(¬ifier->link); 2124} 2125EXPORT_SYMBOL_GPL(preempt_notifier_unregister); 2126 2127static void fire_sched_in_preempt_notifiers(struct task_struct *curr) 2128{ 2129 struct preempt_notifier *notifier; 2130 2131 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link) 2132 notifier->ops->sched_in(notifier, raw_smp_processor_id()); 2133} 2134 2135static void 2136fire_sched_out_preempt_notifiers(struct task_struct *curr, 2137 struct task_struct *next) 2138{ 2139 struct preempt_notifier *notifier; 2140 2141 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link) 2142 notifier->ops->sched_out(notifier, next); 2143} 2144 2145#else /* !CONFIG_PREEMPT_NOTIFIERS */ 2146 2147static void fire_sched_in_preempt_notifiers(struct task_struct *curr) 2148{ 2149} 2150 2151static void 2152fire_sched_out_preempt_notifiers(struct task_struct *curr, 2153 struct task_struct *next) 2154{ 2155} 2156 2157#endif /* CONFIG_PREEMPT_NOTIFIERS */ 2158 2159/** 2160 * prepare_task_switch - prepare to switch tasks 2161 * @rq: the runqueue preparing to switch 2162 * @prev: the current task that is being switched out 2163 * @next: the task we are going to switch to. 2164 * 2165 * This is called with the rq lock held and interrupts off. It must 2166 * be paired with a subsequent finish_task_switch after the context 2167 * switch. 2168 * 2169 * prepare_task_switch sets up locking and calls architecture specific 2170 * hooks. 2171 */ 2172static inline void 2173prepare_task_switch(struct rq *rq, struct task_struct *prev, 2174 struct task_struct *next) 2175{ 2176 trace_sched_switch(prev, next); 2177 sched_info_switch(rq, prev, next); 2178 perf_event_task_sched_out(prev, next); 2179 fire_sched_out_preempt_notifiers(prev, next); 2180 prepare_lock_switch(rq, next); 2181 prepare_arch_switch(next); 2182} 2183 2184/** 2185 * finish_task_switch - clean up after a task-switch 2186 * @rq: runqueue associated with task-switch 2187 * @prev: the thread we just switched away from. 2188 * 2189 * finish_task_switch must be called after the context switch, paired 2190 * with a prepare_task_switch call before the context switch. 2191 * finish_task_switch will reconcile locking set up by prepare_task_switch, 2192 * and do any other architecture-specific cleanup actions. 2193 * 2194 * Note that we may have delayed dropping an mm in context_switch(). If 2195 * so, we finish that here outside of the runqueue lock. (Doing it 2196 * with the lock held can cause deadlocks; see schedule() for 2197 * details.) 2198 */ 2199static void finish_task_switch(struct rq *rq, struct task_struct *prev) 2200 __releases(rq->lock) 2201{ 2202 struct mm_struct *mm = rq->prev_mm; 2203 long prev_state; 2204 2205 rq->prev_mm = NULL; 2206 2207 /* 2208 * A task struct has one reference for the use as "current". 2209 * If a task dies, then it sets TASK_DEAD in tsk->state and calls 2210 * schedule one last time. The schedule call will never return, and 2211 * the scheduled task must drop that reference. 2212 * The test for TASK_DEAD must occur while the runqueue locks are 2213 * still held, otherwise prev could be scheduled on another cpu, die 2214 * there before we look at prev->state, and then the reference would 2215 * be dropped twice. 2216 * Manfred Spraul <manfred@colorfullife.com> 2217 */ 2218 prev_state = prev->state; 2219 vtime_task_switch(prev); 2220 finish_arch_switch(prev); 2221 perf_event_task_sched_in(prev, current); 2222 finish_lock_switch(rq, prev); 2223 finish_arch_post_lock_switch(); 2224 2225 fire_sched_in_preempt_notifiers(current); 2226 if (mm) 2227 mmdrop(mm); 2228 if (unlikely(prev_state == TASK_DEAD)) { 2229 if (prev->sched_class->task_dead) 2230 prev->sched_class->task_dead(prev); 2231 2232 /* 2233 * Remove function-return probe instances associated with this 2234 * task and put them back on the free list. 2235 */ 2236 kprobe_flush_task(prev); 2237 put_task_struct(prev); 2238 } 2239 2240 tick_nohz_task_switch(current); 2241} 2242 2243#ifdef CONFIG_SMP 2244 2245/* rq->lock is NOT held, but preemption is disabled */ 2246static inline void post_schedule(struct rq *rq) 2247{ 2248 if (rq->post_schedule) { 2249 unsigned long flags; 2250 2251 raw_spin_lock_irqsave(&rq->lock, flags); 2252 if (rq->curr->sched_class->post_schedule) 2253 rq->curr->sched_class->post_schedule(rq); 2254 raw_spin_unlock_irqrestore(&rq->lock, flags); 2255 2256 rq->post_schedule = 0; 2257 } 2258} 2259 2260#else 2261 2262static inline void post_schedule(struct rq *rq) 2263{ 2264} 2265 2266#endif 2267 2268/** 2269 * schedule_tail - first thing a freshly forked thread must call. 2270 * @prev: the thread we just switched away from. 2271 */ 2272asmlinkage __visible void schedule_tail(struct task_struct *prev) 2273 __releases(rq->lock) 2274{ 2275 struct rq *rq = this_rq(); 2276 2277 finish_task_switch(rq, prev); 2278 2279 /* 2280 * FIXME: do we need to worry about rq being invalidated by the 2281 * task_switch? 2282 */ 2283 post_schedule(rq); 2284 2285#ifdef __ARCH_WANT_UNLOCKED_CTXSW 2286 /* In this case, finish_task_switch does not reenable preemption */ 2287 preempt_enable(); 2288#endif 2289 if (current->set_child_tid) 2290 put_user(task_pid_vnr(current), current->set_child_tid); 2291} 2292 2293/* 2294 * context_switch - switch to the new MM and the new 2295 * thread's register state. 2296 */ 2297static inline void 2298context_switch(struct rq *rq, struct task_struct *prev, 2299 struct task_struct *next) 2300{ 2301 struct mm_struct *mm, *oldmm; 2302 2303 prepare_task_switch(rq, prev, next); 2304 2305 mm = next->mm; 2306 oldmm = prev->active_mm; 2307 /* 2308 * For paravirt, this is coupled with an exit in switch_to to 2309 * combine the page table reload and the switch backend into 2310 * one hypercall. 2311 */ 2312 arch_start_context_switch(prev); 2313 2314 if (!mm) { 2315 next->active_mm = oldmm; 2316 atomic_inc(&oldmm->mm_count); 2317 enter_lazy_tlb(oldmm, next); 2318 } else 2319 switch_mm(oldmm, mm, next); 2320 2321 if (!prev->mm) { 2322 prev->active_mm = NULL; 2323 rq->prev_mm = oldmm; 2324 } 2325 /* 2326 * Since the runqueue lock will be released by the next 2327 * task (which is an invalid locking op but in the case 2328 * of the scheduler it's an obvious special-case), so we 2329 * do an early lockdep release here: 2330 */ 2331#ifndef __ARCH_WANT_UNLOCKED_CTXSW 2332 spin_release(&rq->lock.dep_map, 1, _THIS_IP_); 2333#endif 2334 2335 context_tracking_task_switch(prev, next); 2336 /* Here we just switch the register state and the stack. */ 2337 switch_to(prev, next, prev); 2338 2339 barrier(); 2340 /* 2341 * this_rq must be evaluated again because prev may have moved 2342 * CPUs since it called schedule(), thus the 'rq' on its stack 2343 * frame will be invalid. 2344 */ 2345 finish_task_switch(this_rq(), prev); 2346} 2347 2348/* 2349 * nr_running and nr_context_switches: 2350 * 2351 * externally visible scheduler statistics: current number of runnable 2352 * threads, total number of context switches performed since bootup. 2353 */ 2354unsigned long nr_running(void) 2355{ 2356 unsigned long i, sum = 0; 2357 2358 for_each_online_cpu(i) 2359 sum += cpu_rq(i)->nr_running; 2360 2361 return sum; 2362} 2363 2364unsigned long long nr_context_switches(void) 2365{ 2366 int i; 2367 unsigned long long sum = 0; 2368 2369 for_each_possible_cpu(i) 2370 sum += cpu_rq(i)->nr_switches; 2371 2372 return sum; 2373} 2374 2375unsigned long nr_iowait(void) 2376{ 2377 unsigned long i, sum = 0; 2378 2379 for_each_possible_cpu(i) 2380 sum += atomic_read(&cpu_rq(i)->nr_iowait); 2381 2382 return sum; 2383} 2384 2385unsigned long nr_iowait_cpu(int cpu) 2386{ 2387 struct rq *this = cpu_rq(cpu); 2388 return atomic_read(&this->nr_iowait); 2389} 2390 2391#ifdef CONFIG_SMP 2392 2393/* 2394 * sched_exec - execve() is a valuable balancing opportunity, because at 2395 * this point the task has the smallest effective memory and cache footprint. 2396 */ 2397void sched_exec(void) 2398{ 2399 struct task_struct *p = current; 2400 unsigned long flags; 2401 int dest_cpu; 2402 2403 raw_spin_lock_irqsave(&p->pi_lock, flags); 2404 dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), SD_BALANCE_EXEC, 0); 2405 if (dest_cpu == smp_processor_id()) 2406 goto unlock; 2407 2408 if (likely(cpu_active(dest_cpu))) { 2409 struct migration_arg arg = { p, dest_cpu }; 2410 2411 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 2412 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg); 2413 return; 2414 } 2415unlock: 2416 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 2417} 2418 2419#endif 2420 2421DEFINE_PER_CPU(struct kernel_stat, kstat); 2422DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat); 2423 2424EXPORT_PER_CPU_SYMBOL(kstat); 2425EXPORT_PER_CPU_SYMBOL(kernel_cpustat); 2426 2427/* 2428 * Return any ns on the sched_clock that have not yet been accounted in 2429 * @p in case that task is currently running. 2430 * 2431 * Called with task_rq_lock() held on @rq. 2432 */ 2433static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq) 2434{ 2435 u64 ns = 0; 2436 2437 if (task_current(rq, p)) { 2438 update_rq_clock(rq); 2439 ns = rq_clock_task(rq) - p->se.exec_start; 2440 if ((s64)ns < 0) 2441 ns = 0; 2442 } 2443 2444 return ns; 2445} 2446 2447unsigned long long task_delta_exec(struct task_struct *p) 2448{ 2449 unsigned long flags; 2450 struct rq *rq; 2451 u64 ns = 0; 2452 2453 rq = task_rq_lock(p, &flags); 2454 ns = do_task_delta_exec(p, rq); 2455 task_rq_unlock(rq, p, &flags); 2456 2457 return ns; 2458} 2459 2460/* 2461 * Return accounted runtime for the task. 2462 * In case the task is currently running, return the runtime plus current's 2463 * pending runtime that have not been accounted yet. 2464 */ 2465unsigned long long task_sched_runtime(struct task_struct *p) 2466{ 2467 unsigned long flags; 2468 struct rq *rq; 2469 u64 ns = 0; 2470 2471#if defined(CONFIG_64BIT) && defined(CONFIG_SMP) 2472 /* 2473 * 64-bit doesn't need locks to atomically read a 64bit value. 2474 * So we have a optimization chance when the task's delta_exec is 0. 2475 * Reading ->on_cpu is racy, but this is ok. 2476 * 2477 * If we race with it leaving cpu, we'll take a lock. So we're correct. 2478 * If we race with it entering cpu, unaccounted time is 0. This is 2479 * indistinguishable from the read occurring a few cycles earlier. 2480 */ 2481 if (!p->on_cpu) 2482 return p->se.sum_exec_runtime; 2483#endif 2484 2485 rq = task_rq_lock(p, &flags); 2486 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq); 2487 task_rq_unlock(rq, p, &flags); 2488 2489 return ns; 2490} 2491 2492/* 2493 * This function gets called by the timer code, with HZ frequency. 2494 * We call it with interrupts disabled. 2495 */ 2496void scheduler_tick(void) 2497{ 2498 int cpu = smp_processor_id(); 2499 struct rq *rq = cpu_rq(cpu); 2500 struct task_struct *curr = rq->curr; 2501 2502 sched_clock_tick(); 2503 2504 raw_spin_lock(&rq->lock); 2505 update_rq_clock(rq); 2506 curr->sched_class->task_tick(rq, curr, 0); 2507 update_cpu_load_active(rq); 2508 raw_spin_unlock(&rq->lock); 2509 2510 perf_event_task_tick(); 2511 2512#ifdef CONFIG_SMP 2513 rq->idle_balance = idle_cpu(cpu); 2514 trigger_load_balance(rq); 2515#endif 2516 rq_last_tick_reset(rq); 2517} 2518 2519#ifdef CONFIG_NO_HZ_FULL 2520/** 2521 * scheduler_tick_max_deferment 2522 * 2523 * Keep at least one tick per second when a single 2524 * active task is running because the scheduler doesn't 2525 * yet completely support full dynticks environment. 2526 * 2527 * This makes sure that uptime, CFS vruntime, load 2528 * balancing, etc... continue to move forward, even 2529 * with a very low granularity. 2530 * 2531 * Return: Maximum deferment in nanoseconds. 2532 */ 2533u64 scheduler_tick_max_deferment(void) 2534{ 2535 struct rq *rq = this_rq(); 2536 unsigned long next, now = ACCESS_ONCE(jiffies); 2537 2538 next = rq->last_sched_tick + HZ; 2539 2540 if (time_before_eq(next, now)) 2541 return 0; 2542 2543 return jiffies_to_nsecs(next - now); 2544} 2545#endif 2546 2547notrace unsigned long get_parent_ip(unsigned long addr) 2548{ 2549 if (in_lock_functions(addr)) { 2550 addr = CALLER_ADDR2; 2551 if (in_lock_functions(addr)) 2552 addr = CALLER_ADDR3; 2553 } 2554 return addr; 2555} 2556 2557#if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \ 2558 defined(CONFIG_PREEMPT_TRACER)) 2559 2560void preempt_count_add(int val) 2561{ 2562#ifdef CONFIG_DEBUG_PREEMPT 2563 /* 2564 * Underflow? 2565 */ 2566 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0))) 2567 return; 2568#endif 2569 __preempt_count_add(val); 2570#ifdef CONFIG_DEBUG_PREEMPT 2571 /* 2572 * Spinlock count overflowing soon? 2573 */ 2574 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >= 2575 PREEMPT_MASK - 10); 2576#endif 2577 if (preempt_count() == val) { 2578 unsigned long ip = get_parent_ip(CALLER_ADDR1); 2579#ifdef CONFIG_DEBUG_PREEMPT 2580 current->preempt_disable_ip = ip; 2581#endif 2582 trace_preempt_off(CALLER_ADDR0, ip); 2583 } 2584} 2585EXPORT_SYMBOL(preempt_count_add); 2586NOKPROBE_SYMBOL(preempt_count_add); 2587 2588void preempt_count_sub(int val) 2589{ 2590#ifdef CONFIG_DEBUG_PREEMPT 2591 /* 2592 * Underflow? 2593 */ 2594 if (DEBUG_LOCKS_WARN_ON(val > preempt_count())) 2595 return; 2596 /* 2597 * Is the spinlock portion underflowing? 2598 */ 2599 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) && 2600 !(preempt_count() & PREEMPT_MASK))) 2601 return; 2602#endif 2603 2604 if (preempt_count() == val) 2605 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1)); 2606 __preempt_count_sub(val); 2607} 2608EXPORT_SYMBOL(preempt_count_sub); 2609NOKPROBE_SYMBOL(preempt_count_sub); 2610 2611#endif 2612 2613/* 2614 * Print scheduling while atomic bug: 2615 */ 2616static noinline void __schedule_bug(struct task_struct *prev) 2617{ 2618 if (oops_in_progress) 2619 return; 2620 2621 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n", 2622 prev->comm, prev->pid, preempt_count()); 2623 2624 debug_show_held_locks(prev); 2625 print_modules(); 2626 if (irqs_disabled()) 2627 print_irqtrace_events(prev); 2628#ifdef CONFIG_DEBUG_PREEMPT 2629 if (in_atomic_preempt_off()) { 2630 pr_err("Preemption disabled at:"); 2631 print_ip_sym(current->preempt_disable_ip); 2632 pr_cont("\n"); 2633 } 2634#endif 2635 dump_stack(); 2636 add_taint(TAINT_WARN, LOCKDEP_STILL_OK); 2637} 2638 2639/* 2640 * Various schedule()-time debugging checks and statistics: 2641 */ 2642static inline void schedule_debug(struct task_struct *prev) 2643{ 2644 /* 2645 * Test if we are atomic. Since do_exit() needs to call into 2646 * schedule() atomically, we ignore that path. Otherwise whine 2647 * if we are scheduling when we should not. 2648 */ 2649 if (unlikely(in_atomic_preempt_off() && prev->state != TASK_DEAD)) 2650 __schedule_bug(prev); 2651 rcu_sleep_check(); 2652 2653 profile_hit(SCHED_PROFILING, __builtin_return_address(0)); 2654 2655 schedstat_inc(this_rq(), sched_count); 2656} 2657 2658/* 2659 * Pick up the highest-prio task: 2660 */ 2661static inline struct task_struct * 2662pick_next_task(struct rq *rq, struct task_struct *prev) 2663{ 2664 const struct sched_class *class = &fair_sched_class; 2665 struct task_struct *p; 2666 2667 /* 2668 * Optimization: we know that if all tasks are in 2669 * the fair class we can call that function directly: 2670 */ 2671 if (likely(prev->sched_class == class && 2672 rq->nr_running == rq->cfs.h_nr_running)) { 2673 p = fair_sched_class.pick_next_task(rq, prev); 2674 if (unlikely(p == RETRY_TASK)) 2675 goto again; 2676 2677 /* assumes fair_sched_class->next == idle_sched_class */ 2678 if (unlikely(!p)) 2679 p = idle_sched_class.pick_next_task(rq, prev); 2680 2681 return p; 2682 } 2683 2684again: 2685 for_each_class(class) { 2686 p = class->pick_next_task(rq, prev); 2687 if (p) { 2688 if (unlikely(p == RETRY_TASK)) 2689 goto again; 2690 return p; 2691 } 2692 } 2693 2694 BUG(); /* the idle class will always have a runnable task */ 2695} 2696 2697/* 2698 * __schedule() is the main scheduler function. 2699 * 2700 * The main means of driving the scheduler and thus entering this function are: 2701 * 2702 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc. 2703 * 2704 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return 2705 * paths. For example, see arch/x86/entry_64.S. 2706 * 2707 * To drive preemption between tasks, the scheduler sets the flag in timer 2708 * interrupt handler scheduler_tick(). 2709 * 2710 * 3. Wakeups don't really cause entry into schedule(). They add a 2711 * task to the run-queue and that's it. 2712 * 2713 * Now, if the new task added to the run-queue preempts the current 2714 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets 2715 * called on the nearest possible occasion: 2716 * 2717 * - If the kernel is preemptible (CONFIG_PREEMPT=y): 2718 * 2719 * - in syscall or exception context, at the next outmost 2720 * preempt_enable(). (this might be as soon as the wake_up()'s 2721 * spin_unlock()!) 2722 * 2723 * - in IRQ context, return from interrupt-handler to 2724 * preemptible context 2725 * 2726 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set) 2727 * then at the next: 2728 * 2729 * - cond_resched() call 2730 * - explicit schedule() call 2731 * - return from syscall or exception to user-space 2732 * - return from interrupt-handler to user-space 2733 */ 2734static void __sched __schedule(void) 2735{ 2736 struct task_struct *prev, *next; 2737 unsigned long *switch_count; 2738 struct rq *rq; 2739 int cpu; 2740 2741need_resched: 2742 preempt_disable(); 2743 cpu = smp_processor_id(); 2744 rq = cpu_rq(cpu); 2745 rcu_note_context_switch(cpu); 2746 prev = rq->curr; 2747 2748 schedule_debug(prev); 2749 2750 if (sched_feat(HRTICK)) 2751 hrtick_clear(rq); 2752 2753 /* 2754 * Make sure that signal_pending_state()->signal_pending() below 2755 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE) 2756 * done by the caller to avoid the race with signal_wake_up(). 2757 */ 2758 smp_mb__before_spinlock(); 2759 raw_spin_lock_irq(&rq->lock); 2760 2761 switch_count = &prev->nivcsw; 2762 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) { 2763 if (unlikely(signal_pending_state(prev->state, prev))) { 2764 prev->state = TASK_RUNNING; 2765 } else { 2766 deactivate_task(rq, prev, DEQUEUE_SLEEP); 2767 prev->on_rq = 0; 2768 2769 /* 2770 * If a worker went to sleep, notify and ask workqueue 2771 * whether it wants to wake up a task to maintain 2772 * concurrency. 2773 */ 2774 if (prev->flags & PF_WQ_WORKER) { 2775 struct task_struct *to_wakeup; 2776 2777 to_wakeup = wq_worker_sleeping(prev, cpu); 2778 if (to_wakeup) 2779 try_to_wake_up_local(to_wakeup); 2780 } 2781 } 2782 switch_count = &prev->nvcsw; 2783 } 2784 2785 if (prev->on_rq || rq->skip_clock_update < 0) 2786 update_rq_clock(rq); 2787 2788 next = pick_next_task(rq, prev); 2789 clear_tsk_need_resched(prev); 2790 clear_preempt_need_resched(); 2791 rq->skip_clock_update = 0; 2792 2793 if (likely(prev != next)) { 2794 rq->nr_switches++; 2795 rq->curr = next; 2796 ++*switch_count; 2797 2798 context_switch(rq, prev, next); /* unlocks the rq */ 2799 /* 2800 * The context switch have flipped the stack from under us 2801 * and restored the local variables which were saved when 2802 * this task called schedule() in the past. prev == current 2803 * is still correct, but it can be moved to another cpu/rq. 2804 */ 2805 cpu = smp_processor_id(); 2806 rq = cpu_rq(cpu); 2807 } else 2808 raw_spin_unlock_irq(&rq->lock); 2809 2810 post_schedule(rq); 2811 2812 sched_preempt_enable_no_resched(); 2813 if (need_resched()) 2814 goto need_resched; 2815} 2816 2817static inline void sched_submit_work(struct task_struct *tsk) 2818{ 2819 if (!tsk->state || tsk_is_pi_blocked(tsk)) 2820 return; 2821 /* 2822 * If we are going to sleep and we have plugged IO queued, 2823 * make sure to submit it to avoid deadlocks. 2824 */ 2825 if (blk_needs_flush_plug(tsk)) 2826 blk_schedule_flush_plug(tsk); 2827} 2828 2829asmlinkage __visible void __sched schedule(void) 2830{ 2831 struct task_struct *tsk = current; 2832 2833 sched_submit_work(tsk); 2834 __schedule(); 2835} 2836EXPORT_SYMBOL(schedule); 2837 2838#ifdef CONFIG_CONTEXT_TRACKING 2839asmlinkage __visible void __sched schedule_user(void) 2840{ 2841 /* 2842 * If we come here after a random call to set_need_resched(), 2843 * or we have been woken up remotely but the IPI has not yet arrived, 2844 * we haven't yet exited the RCU idle mode. Do it here manually until 2845 * we find a better solution. 2846 */ 2847 user_exit(); 2848 schedule(); 2849 user_enter(); 2850} 2851#endif 2852 2853/** 2854 * schedule_preempt_disabled - called with preemption disabled 2855 * 2856 * Returns with preemption disabled. Note: preempt_count must be 1 2857 */ 2858void __sched schedule_preempt_disabled(void) 2859{ 2860 sched_preempt_enable_no_resched(); 2861 schedule(); 2862 preempt_disable(); 2863} 2864 2865#ifdef CONFIG_PREEMPT 2866/* 2867 * this is the entry point to schedule() from in-kernel preemption 2868 * off of preempt_enable. Kernel preemptions off return from interrupt 2869 * occur there and call schedule directly. 2870 */ 2871asmlinkage __visible void __sched notrace preempt_schedule(void) 2872{ 2873 /* 2874 * If there is a non-zero preempt_count or interrupts are disabled, 2875 * we do not want to preempt the current task. Just return.. 2876 */ 2877 if (likely(!preemptible())) 2878 return; 2879 2880 do { 2881 __preempt_count_add(PREEMPT_ACTIVE); 2882 __schedule(); 2883 __preempt_count_sub(PREEMPT_ACTIVE); 2884 2885 /* 2886 * Check again in case we missed a preemption opportunity 2887 * between schedule and now. 2888 */ 2889 barrier(); 2890 } while (need_resched()); 2891} 2892NOKPROBE_SYMBOL(preempt_schedule); 2893EXPORT_SYMBOL(preempt_schedule); 2894#endif /* CONFIG_PREEMPT */ 2895 2896/* 2897 * this is the entry point to schedule() from kernel preemption 2898 * off of irq context. 2899 * Note, that this is called and return with irqs disabled. This will 2900 * protect us against recursive calling from irq. 2901 */ 2902asmlinkage __visible void __sched preempt_schedule_irq(void) 2903{ 2904 enum ctx_state prev_state; 2905 2906 /* Catch callers which need to be fixed */ 2907 BUG_ON(preempt_count() || !irqs_disabled()); 2908 2909 prev_state = exception_enter(); 2910 2911 do { 2912 __preempt_count_add(PREEMPT_ACTIVE); 2913 local_irq_enable(); 2914 __schedule(); 2915 local_irq_disable(); 2916 __preempt_count_sub(PREEMPT_ACTIVE); 2917 2918 /* 2919 * Check again in case we missed a preemption opportunity 2920 * between schedule and now. 2921 */ 2922 barrier(); 2923 } while (need_resched()); 2924 2925 exception_exit(prev_state); 2926} 2927 2928int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags, 2929 void *key) 2930{ 2931 return try_to_wake_up(curr->private, mode, wake_flags); 2932} 2933EXPORT_SYMBOL(default_wake_function); 2934 2935#ifdef CONFIG_RT_MUTEXES 2936 2937/* 2938 * rt_mutex_setprio - set the current priority of a task 2939 * @p: task 2940 * @prio: prio value (kernel-internal form) 2941 * 2942 * This function changes the 'effective' priority of a task. It does 2943 * not touch ->normal_prio like __setscheduler(). 2944 * 2945 * Used by the rt_mutex code to implement priority inheritance 2946 * logic. Call site only calls if the priority of the task changed. 2947 */ 2948void rt_mutex_setprio(struct task_struct *p, int prio) 2949{ 2950 int oldprio, on_rq, running, enqueue_flag = 0; 2951 struct rq *rq; 2952 const struct sched_class *prev_class; 2953 2954 BUG_ON(prio > MAX_PRIO); 2955 2956 rq = __task_rq_lock(p); 2957 2958 /* 2959 * Idle task boosting is a nono in general. There is one 2960 * exception, when PREEMPT_RT and NOHZ is active: 2961 * 2962 * The idle task calls get_next_timer_interrupt() and holds 2963 * the timer wheel base->lock on the CPU and another CPU wants 2964 * to access the timer (probably to cancel it). We can safely 2965 * ignore the boosting request, as the idle CPU runs this code 2966 * with interrupts disabled and will complete the lock 2967 * protected section without being interrupted. So there is no 2968 * real need to boost. 2969 */ 2970 if (unlikely(p == rq->idle)) { 2971 WARN_ON(p != rq->curr); 2972 WARN_ON(p->pi_blocked_on); 2973 goto out_unlock; 2974 } 2975 2976 trace_sched_pi_setprio(p, prio); 2977 p->pi_top_task = rt_mutex_get_top_task(p); 2978 oldprio = p->prio; 2979 prev_class = p->sched_class; 2980 on_rq = p->on_rq; 2981 running = task_current(rq, p); 2982 if (on_rq) 2983 dequeue_task(rq, p, 0); 2984 if (running) 2985 p->sched_class->put_prev_task(rq, p); 2986 2987 /* 2988 * Boosting condition are: 2989 * 1. -rt task is running and holds mutex A 2990 * --> -dl task blocks on mutex A 2991 * 2992 * 2. -dl task is running and holds mutex A 2993 * --> -dl task blocks on mutex A and could preempt the 2994 * running task 2995 */ 2996 if (dl_prio(prio)) { 2997 if (!dl_prio(p->normal_prio) || (p->pi_top_task && 2998 dl_entity_preempt(&p->pi_top_task->dl, &p->dl))) { 2999 p->dl.dl_boosted = 1; 3000 p->dl.dl_throttled = 0; 3001 enqueue_flag = ENQUEUE_REPLENISH; 3002 } else 3003 p->dl.dl_boosted = 0; 3004 p->sched_class = &dl_sched_class; 3005 } else if (rt_prio(prio)) { 3006 if (dl_prio(oldprio)) 3007 p->dl.dl_boosted = 0; 3008 if (oldprio < prio) 3009 enqueue_flag = ENQUEUE_HEAD; 3010 p->sched_class = &rt_sched_class; 3011 } else { 3012 if (dl_prio(oldprio)) 3013 p->dl.dl_boosted = 0; 3014 p->sched_class = &fair_sched_class; 3015 } 3016 3017 p->prio = prio; 3018 3019 if (running) 3020 p->sched_class->set_curr_task(rq); 3021 if (on_rq) 3022 enqueue_task(rq, p, enqueue_flag); 3023 3024 check_class_changed(rq, p, prev_class, oldprio); 3025out_unlock: 3026 __task_rq_unlock(rq); 3027} 3028#endif 3029 3030void set_user_nice(struct task_struct *p, long nice) 3031{ 3032 int old_prio, delta, on_rq; 3033 unsigned long flags; 3034 struct rq *rq; 3035 3036 if (task_nice(p) == nice || nice < MIN_NICE || nice > MAX_NICE) 3037 return; 3038 /* 3039 * We have to be careful, if called from sys_setpriority(), 3040 * the task might be in the middle of scheduling on another CPU. 3041 */ 3042 rq = task_rq_lock(p, &flags); 3043 /* 3044 * The RT priorities are set via sched_setscheduler(), but we still 3045 * allow the 'normal' nice value to be set - but as expected 3046 * it wont have any effect on scheduling until the task is 3047 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR: 3048 */ 3049 if (task_has_dl_policy(p) || task_has_rt_policy(p)) { 3050 p->static_prio = NICE_TO_PRIO(nice); 3051 goto out_unlock; 3052 } 3053 on_rq = p->on_rq; 3054 if (on_rq) 3055 dequeue_task(rq, p, 0); 3056 3057 p->static_prio = NICE_TO_PRIO(nice); 3058 set_load_weight(p); 3059 old_prio = p->prio; 3060 p->prio = effective_prio(p); 3061 delta = p->prio - old_prio; 3062 3063 if (on_rq) { 3064 enqueue_task(rq, p, 0); 3065 /* 3066 * If the task increased its priority or is running and 3067 * lowered its priority, then reschedule its CPU: 3068 */ 3069 if (delta < 0 || (delta > 0 && task_running(rq, p))) 3070 resched_task(rq->curr); 3071 } 3072out_unlock: 3073 task_rq_unlock(rq, p, &flags); 3074} 3075EXPORT_SYMBOL(set_user_nice); 3076 3077/* 3078 * can_nice - check if a task can reduce its nice value 3079 * @p: task 3080 * @nice: nice value 3081 */ 3082int can_nice(const struct task_struct *p, const int nice) 3083{ 3084 /* convert nice value [19,-20] to rlimit style value [1,40] */ 3085 int nice_rlim = nice_to_rlimit(nice); 3086 3087 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) || 3088 capable(CAP_SYS_NICE)); 3089} 3090 3091#ifdef __ARCH_WANT_SYS_NICE 3092 3093/* 3094 * sys_nice - change the priority of the current process. 3095 * @increment: priority increment 3096 * 3097 * sys_setpriority is a more generic, but much slower function that 3098 * does similar things. 3099 */ 3100SYSCALL_DEFINE1(nice, int, increment) 3101{ 3102 long nice, retval; 3103 3104 /* 3105 * Setpriority might change our priority at the same moment. 3106 * We don't have to worry. Conceptually one call occurs first 3107 * and we have a single winner. 3108 */ 3109 increment = clamp(increment, -NICE_WIDTH, NICE_WIDTH); 3110 nice = task_nice(current) + increment; 3111 3112 nice = clamp_val(nice, MIN_NICE, MAX_NICE); 3113 if (increment < 0 && !can_nice(current, nice)) 3114 return -EPERM; 3115 3116 retval = security_task_setnice(current, nice); 3117 if (retval) 3118 return retval; 3119 3120 set_user_nice(current, nice); 3121 return 0; 3122} 3123 3124#endif 3125 3126/** 3127 * task_prio - return the priority value of a given task. 3128 * @p: the task in question. 3129 * 3130 * Return: The priority value as seen by users in /proc. 3131 * RT tasks are offset by -200. Normal tasks are centered 3132 * around 0, value goes from -16 to +15. 3133 */ 3134int task_prio(const struct task_struct *p) 3135{ 3136 return p->prio - MAX_RT_PRIO; 3137} 3138 3139/** 3140 * idle_cpu - is a given cpu idle currently? 3141 * @cpu: the processor in question. 3142 * 3143 * Return: 1 if the CPU is currently idle. 0 otherwise. 3144 */ 3145int idle_cpu(int cpu) 3146{ 3147 struct rq *rq = cpu_rq(cpu); 3148 3149 if (rq->curr != rq->idle) 3150 return 0; 3151 3152 if (rq->nr_running) 3153 return 0; 3154 3155#ifdef CONFIG_SMP 3156 if (!llist_empty(&rq->wake_list)) 3157 return 0; 3158#endif 3159 3160 return 1; 3161} 3162 3163/** 3164 * idle_task - return the idle task for a given cpu. 3165 * @cpu: the processor in question. 3166 * 3167 * Return: The idle task for the cpu @cpu. 3168 */ 3169struct task_struct *idle_task(int cpu) 3170{ 3171 return cpu_rq(cpu)->idle; 3172} 3173 3174/** 3175 * find_process_by_pid - find a process with a matching PID value. 3176 * @pid: the pid in question. 3177 * 3178 * The task of @pid, if found. %NULL otherwise. 3179 */ 3180static struct task_struct *find_process_by_pid(pid_t pid) 3181{ 3182 return pid ? find_task_by_vpid(pid) : current; 3183} 3184 3185/* 3186 * This function initializes the sched_dl_entity of a newly becoming 3187 * SCHED_DEADLINE task. 3188 * 3189 * Only the static values are considered here, the actual runtime and the 3190 * absolute deadline will be properly calculated when the task is enqueued 3191 * for the first time with its new policy. 3192 */ 3193static void 3194__setparam_dl(struct task_struct *p, const struct sched_attr *attr) 3195{ 3196 struct sched_dl_entity *dl_se = &p->dl; 3197 3198 init_dl_task_timer(dl_se); 3199 dl_se->dl_runtime = attr->sched_runtime; 3200 dl_se->dl_deadline = attr->sched_deadline; 3201 dl_se->dl_period = attr->sched_period ?: dl_se->dl_deadline; 3202 dl_se->flags = attr->sched_flags; 3203 dl_se->dl_bw = to_ratio(dl_se->dl_period, dl_se->dl_runtime); 3204 dl_se->dl_throttled = 0; 3205 dl_se->dl_new = 1; 3206 dl_se->dl_yielded = 0; 3207} 3208 3209static void __setscheduler_params(struct task_struct *p, 3210 const struct sched_attr *attr) 3211{ 3212 int policy = attr->sched_policy; 3213 3214 if (policy == -1) /* setparam */ 3215 policy = p->policy; 3216 3217 p->policy = policy; 3218 3219 if (dl_policy(policy)) 3220 __setparam_dl(p, attr); 3221 else if (fair_policy(policy)) 3222 p->static_prio = NICE_TO_PRIO(attr->sched_nice); 3223 3224 /* 3225 * __sched_setscheduler() ensures attr->sched_priority == 0 when 3226 * !rt_policy. Always setting this ensures that things like 3227 * getparam()/getattr() don't report silly values for !rt tasks. 3228 */ 3229 p->rt_priority = attr->sched_priority; 3230 p->normal_prio = normal_prio(p); 3231 set_load_weight(p); 3232} 3233 3234/* Actually do priority change: must hold pi & rq lock. */ 3235static void __setscheduler(struct rq *rq, struct task_struct *p, 3236 const struct sched_attr *attr) 3237{ 3238 __setscheduler_params(p, attr); 3239 3240 /* 3241 * If we get here, there was no pi waiters boosting the 3242 * task. It is safe to use the normal prio. 3243 */ 3244 p->prio = normal_prio(p); 3245 3246 if (dl_prio(p->prio)) 3247 p->sched_class = &dl_sched_class; 3248 else if (rt_prio(p->prio)) 3249 p->sched_class = &rt_sched_class; 3250 else 3251 p->sched_class = &fair_sched_class; 3252} 3253 3254static void 3255__getparam_dl(struct task_struct *p, struct sched_attr *attr) 3256{ 3257 struct sched_dl_entity *dl_se = &p->dl; 3258 3259 attr->sched_priority = p->rt_priority; 3260 attr->sched_runtime = dl_se->dl_runtime; 3261 attr->sched_deadline = dl_se->dl_deadline; 3262 attr->sched_period = dl_se->dl_period; 3263 attr->sched_flags = dl_se->flags; 3264} 3265 3266/* 3267 * This function validates the new parameters of a -deadline task. 3268 * We ask for the deadline not being zero, and greater or equal 3269 * than the runtime, as well as the period of being zero or 3270 * greater than deadline. Furthermore, we have to be sure that 3271 * user parameters are above the internal resolution of 1us (we 3272 * check sched_runtime only since it is always the smaller one) and 3273 * below 2^63 ns (we have to check both sched_deadline and 3274 * sched_period, as the latter can be zero). 3275 */ 3276static bool 3277__checkparam_dl(const struct sched_attr *attr) 3278{ 3279 /* deadline != 0 */ 3280 if (attr->sched_deadline == 0) 3281 return false; 3282 3283 /* 3284 * Since we truncate DL_SCALE bits, make sure we're at least 3285 * that big. 3286 */ 3287 if (attr->sched_runtime < (1ULL << DL_SCALE)) 3288 return false; 3289 3290 /* 3291 * Since we use the MSB for wrap-around and sign issues, make 3292 * sure it's not set (mind that period can be equal to zero). 3293 */ 3294 if (attr->sched_deadline & (1ULL << 63) || 3295 attr->sched_period & (1ULL << 63)) 3296 return false; 3297 3298 /* runtime <= deadline <= period (if period != 0) */ 3299 if ((attr->sched_period != 0 && 3300 attr->sched_period < attr->sched_deadline) || 3301 attr->sched_deadline < attr->sched_runtime) 3302 return false; 3303 3304 return true; 3305} 3306 3307/* 3308 * check the target process has a UID that matches the current process's 3309 */ 3310static bool check_same_owner(struct task_struct *p) 3311{ 3312 const struct cred *cred = current_cred(), *pcred; 3313 bool match; 3314 3315 rcu_read_lock(); 3316 pcred = __task_cred(p); 3317 match = (uid_eq(cred->euid, pcred->euid) || 3318 uid_eq(cred->euid, pcred->uid)); 3319 rcu_read_unlock(); 3320 return match; 3321} 3322 3323static int __sched_setscheduler(struct task_struct *p, 3324 const struct sched_attr *attr, 3325 bool user) 3326{ 3327 int newprio = dl_policy(attr->sched_policy) ? MAX_DL_PRIO - 1 : 3328 MAX_RT_PRIO - 1 - attr->sched_priority; 3329 int retval, oldprio, oldpolicy = -1, on_rq, running; 3330 int policy = attr->sched_policy; 3331 unsigned long flags; 3332 const struct sched_class *prev_class; 3333 struct rq *rq; 3334 int reset_on_fork; 3335 3336 /* may grab non-irq protected spin_locks */ 3337 BUG_ON(in_interrupt()); 3338recheck: 3339 /* double check policy once rq lock held */ 3340 if (policy < 0) { 3341 reset_on_fork = p->sched_reset_on_fork; 3342 policy = oldpolicy = p->policy; 3343 } else { 3344 reset_on_fork = !!(attr->sched_flags & SCHED_FLAG_RESET_ON_FORK); 3345 3346 if (policy != SCHED_DEADLINE && 3347 policy != SCHED_FIFO && policy != SCHED_RR && 3348 policy != SCHED_NORMAL && policy != SCHED_BATCH && 3349 policy != SCHED_IDLE) 3350 return -EINVAL; 3351 } 3352 3353 if (attr->sched_flags & ~(SCHED_FLAG_RESET_ON_FORK)) 3354 return -EINVAL; 3355 3356 /* 3357 * Valid priorities for SCHED_FIFO and SCHED_RR are 3358 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL, 3359 * SCHED_BATCH and SCHED_IDLE is 0. 3360 */ 3361 if ((p->mm && attr->sched_priority > MAX_USER_RT_PRIO-1) || 3362 (!p->mm && attr->sched_priority > MAX_RT_PRIO-1)) 3363 return -EINVAL; 3364 if ((dl_policy(policy) && !__checkparam_dl(attr)) || 3365 (rt_policy(policy) != (attr->sched_priority != 0))) 3366 return -EINVAL; 3367 3368 /* 3369 * Allow unprivileged RT tasks to decrease priority: 3370 */ 3371 if (user && !capable(CAP_SYS_NICE)) { 3372 if (fair_policy(policy)) { 3373 if (attr->sched_nice < task_nice(p) && 3374 !can_nice(p, attr->sched_nice)) 3375 return -EPERM; 3376 } 3377 3378 if (rt_policy(policy)) { 3379 unsigned long rlim_rtprio = 3380 task_rlimit(p, RLIMIT_RTPRIO); 3381 3382 /* can't set/change the rt policy */ 3383 if (policy != p->policy && !rlim_rtprio) 3384 return -EPERM; 3385 3386 /* can't increase priority */ 3387 if (attr->sched_priority > p->rt_priority && 3388 attr->sched_priority > rlim_rtprio) 3389 return -EPERM; 3390 } 3391 3392 /* 3393 * Can't set/change SCHED_DEADLINE policy at all for now 3394 * (safest behavior); in the future we would like to allow 3395 * unprivileged DL tasks to increase their relative deadline 3396 * or reduce their runtime (both ways reducing utilization) 3397 */ 3398 if (dl_policy(policy)) 3399 return -EPERM; 3400 3401 /* 3402 * Treat SCHED_IDLE as nice 20. Only allow a switch to 3403 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it. 3404 */ 3405 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) { 3406 if (!can_nice(p, task_nice(p))) 3407 return -EPERM; 3408 } 3409 3410 /* can't change other user's priorities */ 3411 if (!check_same_owner(p)) 3412 return -EPERM; 3413 3414 /* Normal users shall not reset the sched_reset_on_fork flag */ 3415 if (p->sched_reset_on_fork && !reset_on_fork) 3416 return -EPERM; 3417 } 3418 3419 if (user) { 3420 retval = security_task_setscheduler(p); 3421 if (retval) 3422 return retval; 3423 } 3424 3425 /* 3426 * make sure no PI-waiters arrive (or leave) while we are 3427 * changing the priority of the task: 3428 * 3429 * To be able to change p->policy safely, the appropriate 3430 * runqueue lock must be held. 3431 */ 3432 rq = task_rq_lock(p, &flags); 3433 3434 /* 3435 * Changing the policy of the stop threads its a very bad idea 3436 */ 3437 if (p == rq->stop) { 3438 task_rq_unlock(rq, p, &flags); 3439 return -EINVAL; 3440 } 3441 3442 /* 3443 * If not changing anything there's no need to proceed further, 3444 * but store a possible modification of reset_on_fork. 3445 */ 3446 if (unlikely(policy == p->policy)) { 3447 if (fair_policy(policy) && attr->sched_nice != task_nice(p)) 3448 goto change; 3449 if (rt_policy(policy) && attr->sched_priority != p->rt_priority) 3450 goto change; 3451 if (dl_policy(policy)) 3452 goto change; 3453 3454 p->sched_reset_on_fork = reset_on_fork; 3455 task_rq_unlock(rq, p, &flags); 3456 return 0; 3457 } 3458change: 3459 3460 if (user) { 3461#ifdef CONFIG_RT_GROUP_SCHED 3462 /* 3463 * Do not allow realtime tasks into groups that have no runtime 3464 * assigned. 3465 */ 3466 if (rt_bandwidth_enabled() && rt_policy(policy) && 3467 task_group(p)->rt_bandwidth.rt_runtime == 0 && 3468 !task_group_is_autogroup(task_group(p))) { 3469 task_rq_unlock(rq, p, &flags); 3470 return -EPERM; 3471 } 3472#endif 3473#ifdef CONFIG_SMP 3474 if (dl_bandwidth_enabled() && dl_policy(policy)) { 3475 cpumask_t *span = rq->rd->span; 3476 3477 /* 3478 * Don't allow tasks with an affinity mask smaller than 3479 * the entire root_domain to become SCHED_DEADLINE. We 3480 * will also fail if there's no bandwidth available. 3481 */ 3482 if (!cpumask_subset(span, &p->cpus_allowed) || 3483 rq->rd->dl_bw.bw == 0) { 3484 task_rq_unlock(rq, p, &flags); 3485 return -EPERM; 3486 } 3487 } 3488#endif 3489 } 3490 3491 /* recheck policy now with rq lock held */ 3492 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) { 3493 policy = oldpolicy = -1; 3494 task_rq_unlock(rq, p, &flags); 3495 goto recheck; 3496 } 3497 3498 /* 3499 * If setscheduling to SCHED_DEADLINE (or changing the parameters 3500 * of a SCHED_DEADLINE task) we need to check if enough bandwidth 3501 * is available. 3502 */ 3503 if ((dl_policy(policy) || dl_task(p)) && dl_overflow(p, policy, attr)) { 3504 task_rq_unlock(rq, p, &flags); 3505 return -EBUSY; 3506 } 3507 3508 p->sched_reset_on_fork = reset_on_fork; 3509 oldprio = p->prio; 3510 3511 /* 3512 * Special case for priority boosted tasks. 3513 * 3514 * If the new priority is lower or equal (user space view) 3515 * than the current (boosted) priority, we just store the new 3516 * normal parameters and do not touch the scheduler class and 3517 * the runqueue. This will be done when the task deboost 3518 * itself. 3519 */ 3520 if (rt_mutex_check_prio(p, newprio)) { 3521 __setscheduler_params(p, attr); 3522 task_rq_unlock(rq, p, &flags); 3523 return 0; 3524 } 3525 3526 on_rq = p->on_rq; 3527 running = task_current(rq, p); 3528 if (on_rq) 3529 dequeue_task(rq, p, 0); 3530 if (running) 3531 p->sched_class->put_prev_task(rq, p); 3532 3533 prev_class = p->sched_class; 3534 __setscheduler(rq, p, attr); 3535 3536 if (running) 3537 p->sched_class->set_curr_task(rq); 3538 if (on_rq) { 3539 /* 3540 * We enqueue to tail when the priority of a task is 3541 * increased (user space view). 3542 */ 3543 enqueue_task(rq, p, oldprio <= p->prio ? ENQUEUE_HEAD : 0); 3544 } 3545 3546 check_class_changed(rq, p, prev_class, oldprio); 3547 task_rq_unlock(rq, p, &flags); 3548 3549 rt_mutex_adjust_pi(p); 3550 3551 return 0; 3552} 3553 3554static int _sched_setscheduler(struct task_struct *p, int policy, 3555 const struct sched_param *param, bool check) 3556{ 3557 struct sched_attr attr = { 3558 .sched_policy = policy, 3559 .sched_priority = param->sched_priority, 3560 .sched_nice = PRIO_TO_NICE(p->static_prio), 3561 }; 3562 3563 /* 3564 * Fixup the legacy SCHED_RESET_ON_FORK hack 3565 */ 3566 if (policy & SCHED_RESET_ON_FORK) { 3567 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK; 3568 policy &= ~SCHED_RESET_ON_FORK; 3569 attr.sched_policy = policy; 3570 } 3571 3572 return __sched_setscheduler(p, &attr, check); 3573} 3574/** 3575 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread. 3576 * @p: the task in question. 3577 * @policy: new policy. 3578 * @param: structure containing the new RT priority. 3579 * 3580 * Return: 0 on success. An error code otherwise. 3581 * 3582 * NOTE that the task may be already dead. 3583 */ 3584int sched_setscheduler(struct task_struct *p, int policy, 3585 const struct sched_param *param) 3586{ 3587 return _sched_setscheduler(p, policy, param, true); 3588} 3589EXPORT_SYMBOL_GPL(sched_setscheduler); 3590 3591int sched_setattr(struct task_struct *p, const struct sched_attr *attr) 3592{ 3593 return __sched_setscheduler(p, attr, true); 3594} 3595EXPORT_SYMBOL_GPL(sched_setattr); 3596 3597/** 3598 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace. 3599 * @p: the task in question. 3600 * @policy: new policy. 3601 * @param: structure containing the new RT priority. 3602 * 3603 * Just like sched_setscheduler, only don't bother checking if the 3604 * current context has permission. For example, this is needed in 3605 * stop_machine(): we create temporary high priority worker threads, 3606 * but our caller might not have that capability. 3607 * 3608 * Return: 0 on success. An error code otherwise. 3609 */ 3610int sched_setscheduler_nocheck(struct task_struct *p, int policy, 3611 const struct sched_param *param) 3612{ 3613 return _sched_setscheduler(p, policy, param, false); 3614} 3615 3616static int 3617do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param) 3618{ 3619 struct sched_param lparam; 3620 struct task_struct *p; 3621 int retval; 3622 3623 if (!param || pid < 0) 3624 return -EINVAL; 3625 if (copy_from_user(&lparam, param, sizeof(struct sched_param))) 3626 return -EFAULT; 3627 3628 rcu_read_lock(); 3629 retval = -ESRCH; 3630 p = find_process_by_pid(pid); 3631 if (p != NULL) 3632 retval = sched_setscheduler(p, policy, &lparam); 3633 rcu_read_unlock(); 3634 3635 return retval; 3636} 3637 3638/* 3639 * Mimics kernel/events/core.c perf_copy_attr(). 3640 */ 3641static int sched_copy_attr(struct sched_attr __user *uattr, 3642 struct sched_attr *attr) 3643{ 3644 u32 size; 3645 int ret; 3646 3647 if (!access_ok(VERIFY_WRITE, uattr, SCHED_ATTR_SIZE_VER0)) 3648 return -EFAULT; 3649 3650 /* 3651 * zero the full structure, so that a short copy will be nice. 3652 */ 3653 memset(attr, 0, sizeof(*attr)); 3654 3655 ret = get_user(size, &uattr->size); 3656 if (ret) 3657 return ret; 3658 3659 if (size > PAGE_SIZE) /* silly large */ 3660 goto err_size; 3661 3662 if (!size) /* abi compat */ 3663 size = SCHED_ATTR_SIZE_VER0; 3664 3665 if (size < SCHED_ATTR_SIZE_VER0) 3666 goto err_size; 3667 3668 /* 3669 * If we're handed a bigger struct than we know of, 3670 * ensure all the unknown bits are 0 - i.e. new 3671 * user-space does not rely on any kernel feature 3672 * extensions we dont know about yet. 3673 */ 3674 if (size > sizeof(*attr)) { 3675 unsigned char __user *addr; 3676 unsigned char __user *end; 3677 unsigned char val; 3678 3679 addr = (void __user *)uattr + sizeof(*attr); 3680 end = (void __user *)uattr + size; 3681 3682 for (; addr < end; addr++) { 3683 ret = get_user(val, addr); 3684 if (ret) 3685 return ret; 3686 if (val) 3687 goto err_size; 3688 } 3689 size = sizeof(*attr); 3690 } 3691 3692 ret = copy_from_user(attr, uattr, size); 3693 if (ret) 3694 return -EFAULT; 3695 3696 /* 3697 * XXX: do we want to be lenient like existing syscalls; or do we want 3698 * to be strict and return an error on out-of-bounds values? 3699 */ 3700 attr->sched_nice = clamp(attr->sched_nice, MIN_NICE, MAX_NICE); 3701 3702 return 0; 3703 3704err_size: 3705 put_user(sizeof(*attr), &uattr->size); 3706 return -E2BIG; 3707} 3708 3709/** 3710 * sys_sched_setscheduler - set/change the scheduler policy and RT priority 3711 * @pid: the pid in question. 3712 * @policy: new policy. 3713 * @param: structure containing the new RT priority. 3714 * 3715 * Return: 0 on success. An error code otherwise. 3716 */ 3717SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy, 3718 struct sched_param __user *, param) 3719{ 3720 /* negative values for policy are not valid */ 3721 if (policy < 0) 3722 return -EINVAL; 3723 3724 return do_sched_setscheduler(pid, policy, param); 3725} 3726 3727/** 3728 * sys_sched_setparam - set/change the RT priority of a thread 3729 * @pid: the pid in question. 3730 * @param: structure containing the new RT priority. 3731 * 3732 * Return: 0 on success. An error code otherwise. 3733 */ 3734SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param) 3735{ 3736 return do_sched_setscheduler(pid, -1, param); 3737} 3738 3739/** 3740 * sys_sched_setattr - same as above, but with extended sched_attr 3741 * @pid: the pid in question. 3742 * @uattr: structure containing the extended parameters. 3743 * @flags: for future extension. 3744 */ 3745SYSCALL_DEFINE3(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr, 3746 unsigned int, flags) 3747{ 3748 struct sched_attr attr; 3749 struct task_struct *p; 3750 int retval; 3751 3752 if (!uattr || pid < 0 || flags) 3753 return -EINVAL; 3754 3755 retval = sched_copy_attr(uattr, &attr); 3756 if (retval) 3757 return retval; 3758 3759 if ((int)attr.sched_policy < 0) 3760 return -EINVAL; 3761 3762 rcu_read_lock(); 3763 retval = -ESRCH; 3764 p = find_process_by_pid(pid); 3765 if (p != NULL) 3766 retval = sched_setattr(p, &attr); 3767 rcu_read_unlock(); 3768 3769 return retval; 3770} 3771 3772/** 3773 * sys_sched_getscheduler - get the policy (scheduling class) of a thread 3774 * @pid: the pid in question. 3775 * 3776 * Return: On success, the policy of the thread. Otherwise, a negative error 3777 * code. 3778 */ 3779SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid) 3780{ 3781 struct task_struct *p; 3782 int retval; 3783 3784 if (pid < 0) 3785 return -EINVAL; 3786 3787 retval = -ESRCH; 3788 rcu_read_lock(); 3789 p = find_process_by_pid(pid); 3790 if (p) { 3791 retval = security_task_getscheduler(p); 3792 if (!retval) 3793 retval = p->policy 3794 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0); 3795 } 3796 rcu_read_unlock(); 3797 return retval; 3798} 3799 3800/** 3801 * sys_sched_getparam - get the RT priority of a thread 3802 * @pid: the pid in question. 3803 * @param: structure containing the RT priority. 3804 * 3805 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error 3806 * code. 3807 */ 3808SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param) 3809{ 3810 struct sched_param lp = { .sched_priority = 0 }; 3811 struct task_struct *p; 3812 int retval; 3813 3814 if (!param || pid < 0) 3815 return -EINVAL; 3816 3817 rcu_read_lock(); 3818 p = find_process_by_pid(pid); 3819 retval = -ESRCH; 3820 if (!p) 3821 goto out_unlock; 3822 3823 retval = security_task_getscheduler(p); 3824 if (retval) 3825 goto out_unlock; 3826 3827 if (task_has_rt_policy(p)) 3828 lp.sched_priority = p->rt_priority; 3829 rcu_read_unlock(); 3830 3831 /* 3832 * This one might sleep, we cannot do it with a spinlock held ... 3833 */ 3834 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0; 3835 3836 return retval; 3837 3838out_unlock: 3839 rcu_read_unlock(); 3840 return retval; 3841} 3842 3843static int sched_read_attr(struct sched_attr __user *uattr, 3844 struct sched_attr *attr, 3845 unsigned int usize) 3846{ 3847 int ret; 3848 3849 if (!access_ok(VERIFY_WRITE, uattr, usize)) 3850 return -EFAULT; 3851 3852 /* 3853 * If we're handed a smaller struct than we know of, 3854 * ensure all the unknown bits are 0 - i.e. old 3855 * user-space does not get uncomplete information. 3856 */ 3857 if (usize < sizeof(*attr)) { 3858 unsigned char *addr; 3859 unsigned char *end; 3860 3861 addr = (void *)attr + usize; 3862 end = (void *)attr + sizeof(*attr); 3863 3864 for (; addr < end; addr++) { 3865 if (*addr) 3866 return -EFBIG; 3867 } 3868 3869 attr->size = usize; 3870 } 3871 3872 ret = copy_to_user(uattr, attr, attr->size); 3873 if (ret) 3874 return -EFAULT; 3875 3876 return 0; 3877} 3878 3879/** 3880 * sys_sched_getattr - similar to sched_getparam, but with sched_attr 3881 * @pid: the pid in question. 3882 * @uattr: structure containing the extended parameters. 3883 * @size: sizeof(attr) for fwd/bwd comp. 3884 * @flags: for future extension. 3885 */ 3886SYSCALL_DEFINE4(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr, 3887 unsigned int, size, unsigned int, flags) 3888{ 3889 struct sched_attr attr = { 3890 .size = sizeof(struct sched_attr), 3891 }; 3892 struct task_struct *p; 3893 int retval; 3894 3895 if (!uattr || pid < 0 || size > PAGE_SIZE || 3896 size < SCHED_ATTR_SIZE_VER0 || flags) 3897 return -EINVAL; 3898 3899 rcu_read_lock(); 3900 p = find_process_by_pid(pid); 3901 retval = -ESRCH; 3902 if (!p) 3903 goto out_unlock; 3904 3905 retval = security_task_getscheduler(p); 3906 if (retval) 3907 goto out_unlock; 3908 3909 attr.sched_policy = p->policy; 3910 if (p->sched_reset_on_fork) 3911 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK; 3912 if (task_has_dl_policy(p)) 3913 __getparam_dl(p, &attr); 3914 else if (task_has_rt_policy(p)) 3915 attr.sched_priority = p->rt_priority; 3916 else 3917 attr.sched_nice = task_nice(p); 3918 3919 rcu_read_unlock(); 3920 3921 retval = sched_read_attr(uattr, &attr, size); 3922 return retval; 3923 3924out_unlock: 3925 rcu_read_unlock(); 3926 return retval; 3927} 3928 3929long sched_setaffinity(pid_t pid, const struct cpumask *in_mask) 3930{ 3931 cpumask_var_t cpus_allowed, new_mask; 3932 struct task_struct *p; 3933 int retval; 3934 3935 rcu_read_lock(); 3936 3937 p = find_process_by_pid(pid); 3938 if (!p) { 3939 rcu_read_unlock(); 3940 return -ESRCH; 3941 } 3942 3943 /* Prevent p going away */ 3944 get_task_struct(p); 3945 rcu_read_unlock(); 3946 3947 if (p->flags & PF_NO_SETAFFINITY) { 3948 retval = -EINVAL; 3949 goto out_put_task; 3950 } 3951 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) { 3952 retval = -ENOMEM; 3953 goto out_put_task; 3954 } 3955 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) { 3956 retval = -ENOMEM; 3957 goto out_free_cpus_allowed; 3958 } 3959 retval = -EPERM; 3960 if (!check_same_owner(p)) { 3961 rcu_read_lock(); 3962 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) { 3963 rcu_read_unlock(); 3964 goto out_unlock; 3965 } 3966 rcu_read_unlock(); 3967 } 3968 3969 retval = security_task_setscheduler(p); 3970 if (retval) 3971 goto out_unlock; 3972 3973 3974 cpuset_cpus_allowed(p, cpus_allowed); 3975 cpumask_and(new_mask, in_mask, cpus_allowed); 3976 3977 /* 3978 * Since bandwidth control happens on root_domain basis, 3979 * if admission test is enabled, we only admit -deadline 3980 * tasks allowed to run on all the CPUs in the task's 3981 * root_domain. 3982 */ 3983#ifdef CONFIG_SMP 3984 if (task_has_dl_policy(p)) { 3985 const struct cpumask *span = task_rq(p)->rd->span; 3986 3987 if (dl_bandwidth_enabled() && !cpumask_subset(span, new_mask)) { 3988 retval = -EBUSY; 3989 goto out_unlock; 3990 } 3991 } 3992#endif 3993again: 3994 retval = set_cpus_allowed_ptr(p, new_mask); 3995 3996 if (!retval) { 3997 cpuset_cpus_allowed(p, cpus_allowed); 3998 if (!cpumask_subset(new_mask, cpus_allowed)) { 3999 /* 4000 * We must have raced with a concurrent cpuset 4001 * update. Just reset the cpus_allowed to the 4002 * cpuset's cpus_allowed 4003 */ 4004 cpumask_copy(new_mask, cpus_allowed); 4005 goto again; 4006 } 4007 } 4008out_unlock: 4009 free_cpumask_var(new_mask); 4010out_free_cpus_allowed: 4011 free_cpumask_var(cpus_allowed); 4012out_put_task: 4013 put_task_struct(p); 4014 return retval; 4015} 4016 4017static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len, 4018 struct cpumask *new_mask) 4019{ 4020 if (len < cpumask_size()) 4021 cpumask_clear(new_mask); 4022 else if (len > cpumask_size()) 4023 len = cpumask_size(); 4024 4025 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0; 4026} 4027 4028/** 4029 * sys_sched_setaffinity - set the cpu affinity of a process 4030 * @pid: pid of the process 4031 * @len: length in bytes of the bitmask pointed to by user_mask_ptr 4032 * @user_mask_ptr: user-space pointer to the new cpu mask 4033 * 4034 * Return: 0 on success. An error code otherwise. 4035 */ 4036SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len, 4037 unsigned long __user *, user_mask_ptr) 4038{ 4039 cpumask_var_t new_mask; 4040 int retval; 4041 4042 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) 4043 return -ENOMEM; 4044 4045 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask); 4046 if (retval == 0) 4047 retval = sched_setaffinity(pid, new_mask); 4048 free_cpumask_var(new_mask); 4049 return retval; 4050} 4051 4052long sched_getaffinity(pid_t pid, struct cpumask *mask) 4053{ 4054 struct task_struct *p; 4055 unsigned long flags; 4056 int retval; 4057 4058 rcu_read_lock(); 4059 4060 retval = -ESRCH; 4061 p = find_process_by_pid(pid); 4062 if (!p) 4063 goto out_unlock; 4064 4065 retval = security_task_getscheduler(p); 4066 if (retval) 4067 goto out_unlock; 4068 4069 raw_spin_lock_irqsave(&p->pi_lock, flags); 4070 cpumask_and(mask, &p->cpus_allowed, cpu_active_mask); 4071 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 4072 4073out_unlock: 4074 rcu_read_unlock(); 4075 4076 return retval; 4077} 4078 4079/** 4080 * sys_sched_getaffinity - get the cpu affinity of a process 4081 * @pid: pid of the process 4082 * @len: length in bytes of the bitmask pointed to by user_mask_ptr 4083 * @user_mask_ptr: user-space pointer to hold the current cpu mask 4084 * 4085 * Return: 0 on success. An error code otherwise. 4086 */ 4087SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len, 4088 unsigned long __user *, user_mask_ptr) 4089{ 4090 int ret; 4091 cpumask_var_t mask; 4092 4093 if ((len * BITS_PER_BYTE) < nr_cpu_ids) 4094 return -EINVAL; 4095 if (len & (sizeof(unsigned long)-1)) 4096 return -EINVAL; 4097 4098 if (!alloc_cpumask_var(&mask, GFP_KERNEL)) 4099 return -ENOMEM; 4100 4101 ret = sched_getaffinity(pid, mask); 4102 if (ret == 0) { 4103 size_t retlen = min_t(size_t, len, cpumask_size()); 4104 4105 if (copy_to_user(user_mask_ptr, mask, retlen)) 4106 ret = -EFAULT; 4107 else 4108 ret = retlen; 4109 } 4110 free_cpumask_var(mask); 4111 4112 return ret; 4113} 4114 4115/** 4116 * sys_sched_yield - yield the current processor to other threads. 4117 * 4118 * This function yields the current CPU to other tasks. If there are no 4119 * other threads running on this CPU then this function will return. 4120 * 4121 * Return: 0. 4122 */ 4123SYSCALL_DEFINE0(sched_yield) 4124{ 4125 struct rq *rq = this_rq_lock(); 4126 4127 schedstat_inc(rq, yld_count); 4128 current->sched_class->yield_task(rq); 4129 4130 /* 4131 * Since we are going to call schedule() anyway, there's 4132 * no need to preempt or enable interrupts: 4133 */ 4134 __release(rq->lock); 4135 spin_release(&rq->lock.dep_map, 1, _THIS_IP_); 4136 do_raw_spin_unlock(&rq->lock); 4137 sched_preempt_enable_no_resched(); 4138 4139 schedule(); 4140 4141 return 0; 4142} 4143 4144static void __cond_resched(void) 4145{ 4146 __preempt_count_add(PREEMPT_ACTIVE); 4147 __schedule(); 4148 __preempt_count_sub(PREEMPT_ACTIVE); 4149} 4150 4151int __sched _cond_resched(void) 4152{ 4153 rcu_cond_resched(); 4154 if (should_resched()) { 4155 __cond_resched(); 4156 return 1; 4157 } 4158 return 0; 4159} 4160EXPORT_SYMBOL(_cond_resched); 4161 4162/* 4163 * __cond_resched_lock() - if a reschedule is pending, drop the given lock, 4164 * call schedule, and on return reacquire the lock. 4165 * 4166 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level 4167 * operations here to prevent schedule() from being called twice (once via 4168 * spin_unlock(), once by hand). 4169 */ 4170int __cond_resched_lock(spinlock_t *lock) 4171{ 4172 bool need_rcu_resched = rcu_should_resched(); 4173 int resched = should_resched(); 4174 int ret = 0; 4175 4176 lockdep_assert_held(lock); 4177 4178 if (spin_needbreak(lock) || resched || need_rcu_resched) { 4179 spin_unlock(lock); 4180 if (resched) 4181 __cond_resched(); 4182 else if (unlikely(need_rcu_resched)) 4183 rcu_resched(); 4184 else 4185 cpu_relax(); 4186 ret = 1; 4187 spin_lock(lock); 4188 } 4189 return ret; 4190} 4191EXPORT_SYMBOL(__cond_resched_lock); 4192 4193int __sched __cond_resched_softirq(void) 4194{ 4195 BUG_ON(!in_softirq()); 4196 4197 rcu_cond_resched(); /* BH disabled OK, just recording QSes. */ 4198 if (should_resched()) { 4199 local_bh_enable(); 4200 __cond_resched(); 4201 local_bh_disable(); 4202 return 1; 4203 } 4204 return 0; 4205} 4206EXPORT_SYMBOL(__cond_resched_softirq); 4207 4208/** 4209 * yield - yield the current processor to other threads. 4210 * 4211 * Do not ever use this function, there's a 99% chance you're doing it wrong. 4212 * 4213 * The scheduler is at all times free to pick the calling task as the most 4214 * eligible task to run, if removing the yield() call from your code breaks 4215 * it, its already broken. 4216 * 4217 * Typical broken usage is: 4218 * 4219 * while (!event) 4220 * yield(); 4221 * 4222 * where one assumes that yield() will let 'the other' process run that will 4223 * make event true. If the current task is a SCHED_FIFO task that will never 4224 * happen. Never use yield() as a progress guarantee!! 4225 * 4226 * If you want to use yield() to wait for something, use wait_event(). 4227 * If you want to use yield() to be 'nice' for others, use cond_resched(). 4228 * If you still want to use yield(), do not! 4229 */ 4230void __sched yield(void) 4231{ 4232 set_current_state(TASK_RUNNING); 4233 sys_sched_yield(); 4234} 4235EXPORT_SYMBOL(yield); 4236 4237/** 4238 * yield_to - yield the current processor to another thread in 4239 * your thread group, or accelerate that thread toward the 4240 * processor it's on. 4241 * @p: target task 4242 * @preempt: whether task preemption is allowed or not 4243 * 4244 * It's the caller's job to ensure that the target task struct 4245 * can't go away on us before we can do any checks. 4246 * 4247 * Return: 4248 * true (>0) if we indeed boosted the target task. 4249 * false (0) if we failed to boost the target. 4250 * -ESRCH if there's no task to yield to. 4251 */ 4252int __sched yield_to(struct task_struct *p, bool preempt) 4253{ 4254 struct task_struct *curr = current; 4255 struct rq *rq, *p_rq; 4256 unsigned long flags; 4257 int yielded = 0; 4258 4259 local_irq_save(flags); 4260 rq = this_rq(); 4261 4262again: 4263 p_rq = task_rq(p); 4264 /* 4265 * If we're the only runnable task on the rq and target rq also 4266 * has only one task, there's absolutely no point in yielding. 4267 */ 4268 if (rq->nr_running == 1 && p_rq->nr_running == 1) { 4269 yielded = -ESRCH; 4270 goto out_irq; 4271 } 4272 4273 double_rq_lock(rq, p_rq); 4274 if (task_rq(p) != p_rq) { 4275 double_rq_unlock(rq, p_rq); 4276 goto again; 4277 } 4278 4279 if (!curr->sched_class->yield_to_task) 4280 goto out_unlock; 4281 4282 if (curr->sched_class != p->sched_class) 4283 goto out_unlock; 4284 4285 if (task_running(p_rq, p) || p->state) 4286 goto out_unlock; 4287 4288 yielded = curr->sched_class->yield_to_task(rq, p, preempt); 4289 if (yielded) { 4290 schedstat_inc(rq, yld_count); 4291 /* 4292 * Make p's CPU reschedule; pick_next_entity takes care of 4293 * fairness. 4294 */ 4295 if (preempt && rq != p_rq) 4296 resched_task(p_rq->curr); 4297 } 4298 4299out_unlock: 4300 double_rq_unlock(rq, p_rq); 4301out_irq: 4302 local_irq_restore(flags); 4303 4304 if (yielded > 0) 4305 schedule(); 4306 4307 return yielded; 4308} 4309EXPORT_SYMBOL_GPL(yield_to); 4310 4311/* 4312 * This task is about to go to sleep on IO. Increment rq->nr_iowait so 4313 * that process accounting knows that this is a task in IO wait state. 4314 */ 4315void __sched io_schedule(void) 4316{ 4317 struct rq *rq = raw_rq(); 4318 4319 delayacct_blkio_start(); 4320 atomic_inc(&rq->nr_iowait); 4321 blk_flush_plug(current); 4322 current->in_iowait = 1; 4323 schedule(); 4324 current->in_iowait = 0; 4325 atomic_dec(&rq->nr_iowait); 4326 delayacct_blkio_end(); 4327} 4328EXPORT_SYMBOL(io_schedule); 4329 4330long __sched io_schedule_timeout(long timeout) 4331{ 4332 struct rq *rq = raw_rq(); 4333 long ret; 4334 4335 delayacct_blkio_start(); 4336 atomic_inc(&rq->nr_iowait); 4337 blk_flush_plug(current); 4338 current->in_iowait = 1; 4339 ret = schedule_timeout(timeout); 4340 current->in_iowait = 0; 4341 atomic_dec(&rq->nr_iowait); 4342 delayacct_blkio_end(); 4343 return ret; 4344} 4345 4346/** 4347 * sys_sched_get_priority_max - return maximum RT priority. 4348 * @policy: scheduling class. 4349 * 4350 * Return: On success, this syscall returns the maximum 4351 * rt_priority that can be used by a given scheduling class. 4352 * On failure, a negative error code is returned. 4353 */ 4354SYSCALL_DEFINE1(sched_get_priority_max, int, policy) 4355{ 4356 int ret = -EINVAL; 4357 4358 switch (policy) { 4359 case SCHED_FIFO: 4360 case SCHED_RR: 4361 ret = MAX_USER_RT_PRIO-1; 4362 break; 4363 case SCHED_DEADLINE: 4364 case SCHED_NORMAL: 4365 case SCHED_BATCH: 4366 case SCHED_IDLE: 4367 ret = 0; 4368 break; 4369 } 4370 return ret; 4371} 4372 4373/** 4374 * sys_sched_get_priority_min - return minimum RT priority. 4375 * @policy: scheduling class. 4376 * 4377 * Return: On success, this syscall returns the minimum 4378 * rt_priority that can be used by a given scheduling class. 4379 * On failure, a negative error code is returned. 4380 */ 4381SYSCALL_DEFINE1(sched_get_priority_min, int, policy) 4382{ 4383 int ret = -EINVAL; 4384 4385 switch (policy) { 4386 case SCHED_FIFO: 4387 case SCHED_RR: 4388 ret = 1; 4389 break; 4390 case SCHED_DEADLINE: 4391 case SCHED_NORMAL: 4392 case SCHED_BATCH: 4393 case SCHED_IDLE: 4394 ret = 0; 4395 } 4396 return ret; 4397} 4398 4399/** 4400 * sys_sched_rr_get_interval - return the default timeslice of a process. 4401 * @pid: pid of the process. 4402 * @interval: userspace pointer to the timeslice value. 4403 * 4404 * this syscall writes the default timeslice value of a given process 4405 * into the user-space timespec buffer. A value of '0' means infinity. 4406 * 4407 * Return: On success, 0 and the timeslice is in @interval. Otherwise, 4408 * an error code. 4409 */ 4410SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid, 4411 struct timespec __user *, interval) 4412{ 4413 struct task_struct *p; 4414 unsigned int time_slice; 4415 unsigned long flags; 4416 struct rq *rq; 4417 int retval; 4418 struct timespec t; 4419 4420 if (pid < 0) 4421 return -EINVAL; 4422 4423 retval = -ESRCH; 4424 rcu_read_lock(); 4425 p = find_process_by_pid(pid); 4426 if (!p) 4427 goto out_unlock; 4428 4429 retval = security_task_getscheduler(p); 4430 if (retval) 4431 goto out_unlock; 4432 4433 rq = task_rq_lock(p, &flags); 4434 time_slice = 0; 4435 if (p->sched_class->get_rr_interval) 4436 time_slice = p->sched_class->get_rr_interval(rq, p); 4437 task_rq_unlock(rq, p, &flags); 4438 4439 rcu_read_unlock(); 4440 jiffies_to_timespec(time_slice, &t); 4441 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0; 4442 return retval; 4443 4444out_unlock: 4445 rcu_read_unlock(); 4446 return retval; 4447} 4448 4449static const char stat_nam[] = TASK_STATE_TO_CHAR_STR; 4450 4451void sched_show_task(struct task_struct *p) 4452{ 4453 unsigned long free = 0; 4454 int ppid; 4455 unsigned state; 4456 4457 state = p->state ? __ffs(p->state) + 1 : 0; 4458 printk(KERN_INFO "%-15.15s %c", p->comm, 4459 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?'); 4460#if BITS_PER_LONG == 32 4461 if (state == TASK_RUNNING) 4462 printk(KERN_CONT " running "); 4463 else 4464 printk(KERN_CONT " %08lx ", thread_saved_pc(p)); 4465#else 4466 if (state == TASK_RUNNING) 4467 printk(KERN_CONT " running task "); 4468 else 4469 printk(KERN_CONT " %016lx ", thread_saved_pc(p)); 4470#endif 4471#ifdef CONFIG_DEBUG_STACK_USAGE 4472 free = stack_not_used(p); 4473#endif 4474 rcu_read_lock(); 4475 ppid = task_pid_nr(rcu_dereference(p->real_parent)); 4476 rcu_read_unlock(); 4477 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free, 4478 task_pid_nr(p), ppid, 4479 (unsigned long)task_thread_info(p)->flags); 4480 4481 print_worker_info(KERN_INFO, p); 4482 show_stack(p, NULL); 4483} 4484 4485void show_state_filter(unsigned long state_filter) 4486{ 4487 struct task_struct *g, *p; 4488 4489#if BITS_PER_LONG == 32 4490 printk(KERN_INFO 4491 " task PC stack pid father\n"); 4492#else 4493 printk(KERN_INFO 4494 " task PC stack pid father\n"); 4495#endif 4496 rcu_read_lock(); 4497 do_each_thread(g, p) { 4498 /* 4499 * reset the NMI-timeout, listing all files on a slow 4500 * console might take a lot of time: 4501 */ 4502 touch_nmi_watchdog(); 4503 if (!state_filter || (p->state & state_filter)) 4504 sched_show_task(p); 4505 } while_each_thread(g, p); 4506 4507 touch_all_softlockup_watchdogs(); 4508 4509#ifdef CONFIG_SCHED_DEBUG 4510 sysrq_sched_debug_show(); 4511#endif 4512 rcu_read_unlock(); 4513 /* 4514 * Only show locks if all tasks are dumped: 4515 */ 4516 if (!state_filter) 4517 debug_show_all_locks(); 4518} 4519 4520void init_idle_bootup_task(struct task_struct *idle) 4521{ 4522 idle->sched_class = &idle_sched_class; 4523} 4524 4525/** 4526 * init_idle - set up an idle thread for a given CPU 4527 * @idle: task in question 4528 * @cpu: cpu the idle task belongs to 4529 * 4530 * NOTE: this function does not set the idle thread's NEED_RESCHED 4531 * flag, to make booting more robust. 4532 */ 4533void init_idle(struct task_struct *idle, int cpu) 4534{ 4535 struct rq *rq = cpu_rq(cpu); 4536 unsigned long flags; 4537 4538 raw_spin_lock_irqsave(&rq->lock, flags); 4539 4540 __sched_fork(0, idle); 4541 idle->state = TASK_RUNNING; 4542 idle->se.exec_start = sched_clock(); 4543 4544 do_set_cpus_allowed(idle, cpumask_of(cpu)); 4545 /* 4546 * We're having a chicken and egg problem, even though we are 4547 * holding rq->lock, the cpu isn't yet set to this cpu so the 4548 * lockdep check in task_group() will fail. 4549 * 4550 * Similar case to sched_fork(). / Alternatively we could 4551 * use task_rq_lock() here and obtain the other rq->lock. 4552 * 4553 * Silence PROVE_RCU 4554 */ 4555 rcu_read_lock(); 4556 __set_task_cpu(idle, cpu); 4557 rcu_read_unlock(); 4558 4559 rq->curr = rq->idle = idle; 4560 idle->on_rq = 1; 4561#if defined(CONFIG_SMP) 4562 idle->on_cpu = 1; 4563#endif 4564 raw_spin_unlock_irqrestore(&rq->lock, flags); 4565 4566 /* Set the preempt count _outside_ the spinlocks! */ 4567 init_idle_preempt_count(idle, cpu); 4568 4569 /* 4570 * The idle tasks have their own, simple scheduling class: 4571 */ 4572 idle->sched_class = &idle_sched_class; 4573 ftrace_graph_init_idle_task(idle, cpu); 4574 vtime_init_idle(idle, cpu); 4575#if defined(CONFIG_SMP) 4576 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu); 4577#endif 4578} 4579 4580#ifdef CONFIG_SMP 4581void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask) 4582{ 4583 if (p->sched_class && p->sched_class->set_cpus_allowed) 4584 p->sched_class->set_cpus_allowed(p, new_mask); 4585 4586 cpumask_copy(&p->cpus_allowed, new_mask); 4587 p->nr_cpus_allowed = cpumask_weight(new_mask); 4588} 4589 4590/* 4591 * This is how migration works: 4592 * 4593 * 1) we invoke migration_cpu_stop() on the target CPU using 4594 * stop_one_cpu(). 4595 * 2) stopper starts to run (implicitly forcing the migrated thread 4596 * off the CPU) 4597 * 3) it checks whether the migrated task is still in the wrong runqueue. 4598 * 4) if it's in the wrong runqueue then the migration thread removes 4599 * it and puts it into the right queue. 4600 * 5) stopper completes and stop_one_cpu() returns and the migration 4601 * is done. 4602 */ 4603 4604/* 4605 * Change a given task's CPU affinity. Migrate the thread to a 4606 * proper CPU and schedule it away if the CPU it's executing on 4607 * is removed from the allowed bitmask. 4608 * 4609 * NOTE: the caller must have a valid reference to the task, the 4610 * task must not exit() & deallocate itself prematurely. The 4611 * call is not atomic; no spinlocks may be held. 4612 */ 4613int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask) 4614{ 4615 unsigned long flags; 4616 struct rq *rq; 4617 unsigned int dest_cpu; 4618 int ret = 0; 4619 4620 rq = task_rq_lock(p, &flags); 4621 4622 if (cpumask_equal(&p->cpus_allowed, new_mask)) 4623 goto out; 4624 4625 if (!cpumask_intersects(new_mask, cpu_active_mask)) { 4626 ret = -EINVAL; 4627 goto out; 4628 } 4629 4630 do_set_cpus_allowed(p, new_mask); 4631 4632 /* Can the task run on the task's current CPU? If so, we're done */ 4633 if (cpumask_test_cpu(task_cpu(p), new_mask)) 4634 goto out; 4635 4636 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask); 4637 if (p->on_rq) { 4638 struct migration_arg arg = { p, dest_cpu }; 4639 /* Need help from migration thread: drop lock and wait. */ 4640 task_rq_unlock(rq, p, &flags); 4641 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg); 4642 tlb_migrate_finish(p->mm); 4643 return 0; 4644 } 4645out: 4646 task_rq_unlock(rq, p, &flags); 4647 4648 return ret; 4649} 4650EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr); 4651 4652/* 4653 * Move (not current) task off this cpu, onto dest cpu. We're doing 4654 * this because either it can't run here any more (set_cpus_allowed() 4655 * away from this CPU, or CPU going down), or because we're 4656 * attempting to rebalance this task on exec (sched_exec). 4657 * 4658 * So we race with normal scheduler movements, but that's OK, as long 4659 * as the task is no longer on this CPU. 4660 * 4661 * Returns non-zero if task was successfully migrated. 4662 */ 4663static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu) 4664{ 4665 struct rq *rq_dest, *rq_src; 4666 int ret = 0; 4667 4668 if (unlikely(!cpu_active(dest_cpu))) 4669 return ret; 4670 4671 rq_src = cpu_rq(src_cpu); 4672 rq_dest = cpu_rq(dest_cpu); 4673 4674 raw_spin_lock(&p->pi_lock); 4675 double_rq_lock(rq_src, rq_dest); 4676 /* Already moved. */ 4677 if (task_cpu(p) != src_cpu) 4678 goto done; 4679 /* Affinity changed (again). */ 4680 if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p))) 4681 goto fail; 4682 4683 /* 4684 * If we're not on a rq, the next wake-up will ensure we're 4685 * placed properly. 4686 */ 4687 if (p->on_rq) { 4688 dequeue_task(rq_src, p, 0); 4689 set_task_cpu(p, dest_cpu); 4690 enqueue_task(rq_dest, p, 0); 4691 check_preempt_curr(rq_dest, p, 0); 4692 } 4693done: 4694 ret = 1; 4695fail: 4696 double_rq_unlock(rq_src, rq_dest); 4697 raw_spin_unlock(&p->pi_lock); 4698 return ret; 4699} 4700 4701#ifdef CONFIG_NUMA_BALANCING 4702/* Migrate current task p to target_cpu */ 4703int migrate_task_to(struct task_struct *p, int target_cpu) 4704{ 4705 struct migration_arg arg = { p, target_cpu }; 4706 int curr_cpu = task_cpu(p); 4707 4708 if (curr_cpu == target_cpu) 4709 return 0; 4710 4711 if (!cpumask_test_cpu(target_cpu, tsk_cpus_allowed(p))) 4712 return -EINVAL; 4713 4714 /* TODO: This is not properly updating schedstats */ 4715 4716 trace_sched_move_numa(p, curr_cpu, target_cpu); 4717 return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg); 4718} 4719 4720/* 4721 * Requeue a task on a given node and accurately track the number of NUMA 4722 * tasks on the runqueues 4723 */ 4724void sched_setnuma(struct task_struct *p, int nid) 4725{ 4726 struct rq *rq; 4727 unsigned long flags; 4728 bool on_rq, running; 4729 4730 rq = task_rq_lock(p, &flags); 4731 on_rq = p->on_rq; 4732 running = task_current(rq, p); 4733 4734 if (on_rq) 4735 dequeue_task(rq, p, 0); 4736 if (running) 4737 p->sched_class->put_prev_task(rq, p); 4738 4739 p->numa_preferred_nid = nid; 4740 4741 if (running) 4742 p->sched_class->set_curr_task(rq); 4743 if (on_rq) 4744 enqueue_task(rq, p, 0); 4745 task_rq_unlock(rq, p, &flags); 4746} 4747#endif 4748 4749/* 4750 * migration_cpu_stop - this will be executed by a highprio stopper thread 4751 * and performs thread migration by bumping thread off CPU then 4752 * 'pushing' onto another runqueue. 4753 */ 4754static int migration_cpu_stop(void *data) 4755{ 4756 struct migration_arg *arg = data; 4757 4758 /* 4759 * The original target cpu might have gone down and we might 4760 * be on another cpu but it doesn't matter. 4761 */ 4762 local_irq_disable(); 4763 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu); 4764 local_irq_enable(); 4765 return 0; 4766} 4767 4768#ifdef CONFIG_HOTPLUG_CPU 4769 4770/* 4771 * Ensures that the idle task is using init_mm right before its cpu goes 4772 * offline. 4773 */ 4774void idle_task_exit(void) 4775{ 4776 struct mm_struct *mm = current->active_mm; 4777 4778 BUG_ON(cpu_online(smp_processor_id())); 4779 4780 if (mm != &init_mm) { 4781 switch_mm(mm, &init_mm, current); 4782 finish_arch_post_lock_switch(); 4783 } 4784 mmdrop(mm); 4785} 4786 4787/* 4788 * Since this CPU is going 'away' for a while, fold any nr_active delta 4789 * we might have. Assumes we're called after migrate_tasks() so that the 4790 * nr_active count is stable. 4791 * 4792 * Also see the comment "Global load-average calculations". 4793 */ 4794static void calc_load_migrate(struct rq *rq) 4795{ 4796 long delta = calc_load_fold_active(rq); 4797 if (delta) 4798 atomic_long_add(delta, &calc_load_tasks); 4799} 4800 4801static void put_prev_task_fake(struct rq *rq, struct task_struct *prev) 4802{ 4803} 4804 4805static const struct sched_class fake_sched_class = { 4806 .put_prev_task = put_prev_task_fake, 4807}; 4808 4809static struct task_struct fake_task = { 4810 /* 4811 * Avoid pull_{rt,dl}_task() 4812 */ 4813 .prio = MAX_PRIO + 1, 4814 .sched_class = &fake_sched_class, 4815}; 4816 4817/* 4818 * Migrate all tasks from the rq, sleeping tasks will be migrated by 4819 * try_to_wake_up()->select_task_rq(). 4820 * 4821 * Called with rq->lock held even though we'er in stop_machine() and 4822 * there's no concurrency possible, we hold the required locks anyway 4823 * because of lock validation efforts. 4824 */ 4825static void migrate_tasks(unsigned int dead_cpu) 4826{ 4827 struct rq *rq = cpu_rq(dead_cpu); 4828 struct task_struct *next, *stop = rq->stop; 4829 int dest_cpu; 4830 4831 /* 4832 * Fudge the rq selection such that the below task selection loop 4833 * doesn't get stuck on the currently eligible stop task. 4834 * 4835 * We're currently inside stop_machine() and the rq is either stuck 4836 * in the stop_machine_cpu_stop() loop, or we're executing this code, 4837 * either way we should never end up calling schedule() until we're 4838 * done here. 4839 */ 4840 rq->stop = NULL; 4841 4842 /* 4843 * put_prev_task() and pick_next_task() sched 4844 * class method both need to have an up-to-date 4845 * value of rq->clock[_task] 4846 */ 4847 update_rq_clock(rq); 4848 4849 for ( ; ; ) { 4850 /* 4851 * There's this thread running, bail when that's the only 4852 * remaining thread. 4853 */ 4854 if (rq->nr_running == 1) 4855 break; 4856 4857 next = pick_next_task(rq, &fake_task); 4858 BUG_ON(!next); 4859 next->sched_class->put_prev_task(rq, next); 4860 4861 /* Find suitable destination for @next, with force if needed. */ 4862 dest_cpu = select_fallback_rq(dead_cpu, next); 4863 raw_spin_unlock(&rq->lock); 4864 4865 __migrate_task(next, dead_cpu, dest_cpu); 4866 4867 raw_spin_lock(&rq->lock); 4868 } 4869 4870 rq->stop = stop; 4871} 4872 4873#endif /* CONFIG_HOTPLUG_CPU */ 4874 4875#if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL) 4876 4877static struct ctl_table sd_ctl_dir[] = { 4878 { 4879 .procname = "sched_domain", 4880 .mode = 0555, 4881 }, 4882 {} 4883}; 4884 4885static struct ctl_table sd_ctl_root[] = { 4886 { 4887 .procname = "kernel", 4888 .mode = 0555, 4889 .child = sd_ctl_dir, 4890 }, 4891 {} 4892}; 4893 4894static struct ctl_table *sd_alloc_ctl_entry(int n) 4895{ 4896 struct ctl_table *entry = 4897 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL); 4898 4899 return entry; 4900} 4901 4902static void sd_free_ctl_entry(struct ctl_table **tablep) 4903{ 4904 struct ctl_table *entry; 4905 4906 /* 4907 * In the intermediate directories, both the child directory and 4908 * procname are dynamically allocated and could fail but the mode 4909 * will always be set. In the lowest directory the names are 4910 * static strings and all have proc handlers. 4911 */ 4912 for (entry = *tablep; entry->mode; entry++) { 4913 if (entry->child) 4914 sd_free_ctl_entry(&entry->child); 4915 if (entry->proc_handler == NULL) 4916 kfree(entry->procname); 4917 } 4918 4919 kfree(*tablep); 4920 *tablep = NULL; 4921} 4922 4923static int min_load_idx = 0; 4924static int max_load_idx = CPU_LOAD_IDX_MAX-1; 4925 4926static void 4927set_table_entry(struct ctl_table *entry, 4928 const char *procname, void *data, int maxlen, 4929 umode_t mode, proc_handler *proc_handler, 4930 bool load_idx) 4931{ 4932 entry->procname = procname; 4933 entry->data = data; 4934 entry->maxlen = maxlen; 4935 entry->mode = mode; 4936 entry->proc_handler = proc_handler; 4937 4938 if (load_idx) { 4939 entry->extra1 = &min_load_idx; 4940 entry->extra2 = &max_load_idx; 4941 } 4942} 4943 4944static struct ctl_table * 4945sd_alloc_ctl_domain_table(struct sched_domain *sd) 4946{ 4947 struct ctl_table *table = sd_alloc_ctl_entry(14); 4948 4949 if (table == NULL) 4950 return NULL; 4951 4952 set_table_entry(&table[0], "min_interval", &sd->min_interval, 4953 sizeof(long), 0644, proc_doulongvec_minmax, false); 4954 set_table_entry(&table[1], "max_interval", &sd->max_interval, 4955 sizeof(long), 0644, proc_doulongvec_minmax, false); 4956 set_table_entry(&table[2], "busy_idx", &sd->busy_idx, 4957 sizeof(int), 0644, proc_dointvec_minmax, true); 4958 set_table_entry(&table[3], "idle_idx", &sd->idle_idx, 4959 sizeof(int), 0644, proc_dointvec_minmax, true); 4960 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx, 4961 sizeof(int), 0644, proc_dointvec_minmax, true); 4962 set_table_entry(&table[5], "wake_idx", &sd->wake_idx, 4963 sizeof(int), 0644, proc_dointvec_minmax, true); 4964 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx, 4965 sizeof(int), 0644, proc_dointvec_minmax, true); 4966 set_table_entry(&table[7], "busy_factor", &sd->busy_factor, 4967 sizeof(int), 0644, proc_dointvec_minmax, false); 4968 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct, 4969 sizeof(int), 0644, proc_dointvec_minmax, false); 4970 set_table_entry(&table[9], "cache_nice_tries", 4971 &sd->cache_nice_tries, 4972 sizeof(int), 0644, proc_dointvec_minmax, false); 4973 set_table_entry(&table[10], "flags", &sd->flags, 4974 sizeof(int), 0644, proc_dointvec_minmax, false); 4975 set_table_entry(&table[11], "max_newidle_lb_cost", 4976 &sd->max_newidle_lb_cost, 4977 sizeof(long), 0644, proc_doulongvec_minmax, false); 4978 set_table_entry(&table[12], "name", sd->name, 4979 CORENAME_MAX_SIZE, 0444, proc_dostring, false); 4980 /* &table[13] is terminator */ 4981 4982 return table; 4983} 4984 4985static struct ctl_table *sd_alloc_ctl_cpu_table(int cpu) 4986{ 4987 struct ctl_table *entry, *table; 4988 struct sched_domain *sd; 4989 int domain_num = 0, i; 4990 char buf[32]; 4991 4992 for_each_domain(cpu, sd) 4993 domain_num++; 4994 entry = table = sd_alloc_ctl_entry(domain_num + 1); 4995 if (table == NULL) 4996 return NULL; 4997 4998 i = 0; 4999 for_each_domain(cpu, sd) { 5000 snprintf(buf, 32, "domain%d", i); 5001 entry->procname = kstrdup(buf, GFP_KERNEL); 5002 entry->mode = 0555; 5003 entry->child = sd_alloc_ctl_domain_table(sd); 5004 entry++; 5005 i++; 5006 } 5007 return table; 5008} 5009 5010static struct ctl_table_header *sd_sysctl_header; 5011static void register_sched_domain_sysctl(void) 5012{ 5013 int i, cpu_num = num_possible_cpus(); 5014 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1); 5015 char buf[32]; 5016 5017 WARN_ON(sd_ctl_dir[0].child); 5018 sd_ctl_dir[0].child = entry; 5019 5020 if (entry == NULL) 5021 return; 5022 5023 for_each_possible_cpu(i) { 5024 snprintf(buf, 32, "cpu%d", i); 5025 entry->procname = kstrdup(buf, GFP_KERNEL); 5026 entry->mode = 0555; 5027 entry->child = sd_alloc_ctl_cpu_table(i); 5028 entry++; 5029 } 5030 5031 WARN_ON(sd_sysctl_header); 5032 sd_sysctl_header = register_sysctl_table(sd_ctl_root); 5033} 5034 5035/* may be called multiple times per register */ 5036static void unregister_sched_domain_sysctl(void) 5037{ 5038 if (sd_sysctl_header) 5039 unregister_sysctl_table(sd_sysctl_header); 5040 sd_sysctl_header = NULL; 5041 if (sd_ctl_dir[0].child) 5042 sd_free_ctl_entry(&sd_ctl_dir[0].child); 5043} 5044#else 5045static void register_sched_domain_sysctl(void) 5046{ 5047} 5048static void unregister_sched_domain_sysctl(void) 5049{ 5050} 5051#endif 5052 5053static void set_rq_online(struct rq *rq) 5054{ 5055 if (!rq->online) { 5056 const struct sched_class *class; 5057 5058 cpumask_set_cpu(rq->cpu, rq->rd->online); 5059 rq->online = 1; 5060 5061 for_each_class(class) { 5062 if (class->rq_online) 5063 class->rq_online(rq); 5064 } 5065 } 5066} 5067 5068static void set_rq_offline(struct rq *rq) 5069{ 5070 if (rq->online) { 5071 const struct sched_class *class; 5072 5073 for_each_class(class) { 5074 if (class->rq_offline) 5075 class->rq_offline(rq); 5076 } 5077 5078 cpumask_clear_cpu(rq->cpu, rq->rd->online); 5079 rq->online = 0; 5080 } 5081} 5082 5083/* 5084 * migration_call - callback that gets triggered when a CPU is added. 5085 * Here we can start up the necessary migration thread for the new CPU. 5086 */ 5087static int 5088migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu) 5089{ 5090 int cpu = (long)hcpu; 5091 unsigned long flags; 5092 struct rq *rq = cpu_rq(cpu); 5093 5094 switch (action & ~CPU_TASKS_FROZEN) { 5095 5096 case CPU_UP_PREPARE: 5097 rq->calc_load_update = calc_load_update; 5098 break; 5099 5100 case CPU_ONLINE: 5101 /* Update our root-domain */ 5102 raw_spin_lock_irqsave(&rq->lock, flags); 5103 if (rq->rd) { 5104 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span)); 5105 5106 set_rq_online(rq); 5107 } 5108 raw_spin_unlock_irqrestore(&rq->lock, flags); 5109 break; 5110 5111#ifdef CONFIG_HOTPLUG_CPU 5112 case CPU_DYING: 5113 sched_ttwu_pending(); 5114 /* Update our root-domain */ 5115 raw_spin_lock_irqsave(&rq->lock, flags); 5116 if (rq->rd) { 5117 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span)); 5118 set_rq_offline(rq); 5119 } 5120 migrate_tasks(cpu); 5121 BUG_ON(rq->nr_running != 1); /* the migration thread */ 5122 raw_spin_unlock_irqrestore(&rq->lock, flags); 5123 break; 5124 5125 case CPU_DEAD: 5126 calc_load_migrate(rq); 5127 break; 5128#endif 5129 } 5130 5131 update_max_interval(); 5132 5133 return NOTIFY_OK; 5134} 5135 5136/* 5137 * Register at high priority so that task migration (migrate_all_tasks) 5138 * happens before everything else. This has to be lower priority than 5139 * the notifier in the perf_event subsystem, though. 5140 */ 5141static struct notifier_block migration_notifier = { 5142 .notifier_call = migration_call, 5143 .priority = CPU_PRI_MIGRATION, 5144}; 5145 5146static void __cpuinit set_cpu_rq_start_time(void) 5147{ 5148 int cpu = smp_processor_id(); 5149 struct rq *rq = cpu_rq(cpu); 5150 rq->age_stamp = sched_clock_cpu(cpu); 5151} 5152 5153static int sched_cpu_active(struct notifier_block *nfb, 5154 unsigned long action, void *hcpu) 5155{ 5156 switch (action & ~CPU_TASKS_FROZEN) { 5157 case CPU_STARTING: 5158 set_cpu_rq_start_time(); 5159 return NOTIFY_OK; 5160 case CPU_DOWN_FAILED: 5161 set_cpu_active((long)hcpu, true); 5162 return NOTIFY_OK; 5163 default: 5164 return NOTIFY_DONE; 5165 } 5166} 5167 5168static int sched_cpu_inactive(struct notifier_block *nfb, 5169 unsigned long action, void *hcpu) 5170{ 5171 unsigned long flags; 5172 long cpu = (long)hcpu; 5173 5174 switch (action & ~CPU_TASKS_FROZEN) { 5175 case CPU_DOWN_PREPARE: 5176 set_cpu_active(cpu, false); 5177 5178 /* explicitly allow suspend */ 5179 if (!(action & CPU_TASKS_FROZEN)) { 5180 struct dl_bw *dl_b = dl_bw_of(cpu); 5181 bool overflow; 5182 int cpus; 5183 5184 raw_spin_lock_irqsave(&dl_b->lock, flags); 5185 cpus = dl_bw_cpus(cpu); 5186 overflow = __dl_overflow(dl_b, cpus, 0, 0); 5187 raw_spin_unlock_irqrestore(&dl_b->lock, flags); 5188 5189 if (overflow) 5190 return notifier_from_errno(-EBUSY); 5191 } 5192 return NOTIFY_OK; 5193 } 5194 5195 return NOTIFY_DONE; 5196} 5197 5198static int __init migration_init(void) 5199{ 5200 void *cpu = (void *)(long)smp_processor_id(); 5201 int err; 5202 5203 /* Initialize migration for the boot CPU */ 5204 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu); 5205 BUG_ON(err == NOTIFY_BAD); 5206 migration_call(&migration_notifier, CPU_ONLINE, cpu); 5207 register_cpu_notifier(&migration_notifier); 5208 5209 /* Register cpu active notifiers */ 5210 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE); 5211 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE); 5212 5213 return 0; 5214} 5215early_initcall(migration_init); 5216#endif 5217 5218#ifdef CONFIG_SMP 5219 5220static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */ 5221 5222#ifdef CONFIG_SCHED_DEBUG 5223 5224static __read_mostly int sched_debug_enabled; 5225 5226static int __init sched_debug_setup(char *str) 5227{ 5228 sched_debug_enabled = 1; 5229 5230 return 0; 5231} 5232early_param("sched_debug", sched_debug_setup); 5233 5234static inline bool sched_debug(void) 5235{ 5236 return sched_debug_enabled; 5237} 5238 5239static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level, 5240 struct cpumask *groupmask) 5241{ 5242 struct sched_group *group = sd->groups; 5243 char str[256]; 5244 5245 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd)); 5246 cpumask_clear(groupmask); 5247 5248 printk(KERN_DEBUG "%*s domain %d: ", level, "", level); 5249 5250 if (!(sd->flags & SD_LOAD_BALANCE)) { 5251 printk("does not load-balance\n"); 5252 if (sd->parent) 5253 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain" 5254 " has parent"); 5255 return -1; 5256 } 5257 5258 printk(KERN_CONT "span %s level %s\n", str, sd->name); 5259 5260 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) { 5261 printk(KERN_ERR "ERROR: domain->span does not contain " 5262 "CPU%d\n", cpu); 5263 } 5264 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) { 5265 printk(KERN_ERR "ERROR: domain->groups does not contain" 5266 " CPU%d\n", cpu); 5267 } 5268 5269 printk(KERN_DEBUG "%*s groups:", level + 1, ""); 5270 do { 5271 if (!group) { 5272 printk("\n"); 5273 printk(KERN_ERR "ERROR: group is NULL\n"); 5274 break; 5275 } 5276 5277 /* 5278 * Even though we initialize ->capacity to something semi-sane, 5279 * we leave capacity_orig unset. This allows us to detect if 5280 * domain iteration is still funny without causing /0 traps. 5281 */ 5282 if (!group->sgc->capacity_orig) { 5283 printk(KERN_CONT "\n"); 5284 printk(KERN_ERR "ERROR: domain->cpu_capacity not set\n"); 5285 break; 5286 } 5287 5288 if (!cpumask_weight(sched_group_cpus(group))) { 5289 printk(KERN_CONT "\n"); 5290 printk(KERN_ERR "ERROR: empty group\n"); 5291 break; 5292 } 5293 5294 if (!(sd->flags & SD_OVERLAP) && 5295 cpumask_intersects(groupmask, sched_group_cpus(group))) { 5296 printk(KERN_CONT "\n"); 5297 printk(KERN_ERR "ERROR: repeated CPUs\n"); 5298 break; 5299 } 5300 5301 cpumask_or(groupmask, groupmask, sched_group_cpus(group)); 5302 5303 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group)); 5304 5305 printk(KERN_CONT " %s", str); 5306 if (group->sgc->capacity != SCHED_CAPACITY_SCALE) { 5307 printk(KERN_CONT " (cpu_capacity = %d)", 5308 group->sgc->capacity); 5309 } 5310 5311 group = group->next; 5312 } while (group != sd->groups); 5313 printk(KERN_CONT "\n"); 5314 5315 if (!cpumask_equal(sched_domain_span(sd), groupmask)) 5316 printk(KERN_ERR "ERROR: groups don't span domain->span\n"); 5317 5318 if (sd->parent && 5319 !cpumask_subset(groupmask, sched_domain_span(sd->parent))) 5320 printk(KERN_ERR "ERROR: parent span is not a superset " 5321 "of domain->span\n"); 5322 return 0; 5323} 5324 5325static void sched_domain_debug(struct sched_domain *sd, int cpu) 5326{ 5327 int level = 0; 5328 5329 if (!sched_debug_enabled) 5330 return; 5331 5332 if (!sd) { 5333 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu); 5334 return; 5335 } 5336 5337 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu); 5338 5339 for (;;) { 5340 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask)) 5341 break; 5342 level++; 5343 sd = sd->parent; 5344 if (!sd) 5345 break; 5346 } 5347} 5348#else /* !CONFIG_SCHED_DEBUG */ 5349# define sched_domain_debug(sd, cpu) do { } while (0) 5350static inline bool sched_debug(void) 5351{ 5352 return false; 5353} 5354#endif /* CONFIG_SCHED_DEBUG */ 5355 5356static int sd_degenerate(struct sched_domain *sd) 5357{ 5358 if (cpumask_weight(sched_domain_span(sd)) == 1) 5359 return 1; 5360 5361 /* Following flags need at least 2 groups */ 5362 if (sd->flags & (SD_LOAD_BALANCE | 5363 SD_BALANCE_NEWIDLE | 5364 SD_BALANCE_FORK | 5365 SD_BALANCE_EXEC | 5366 SD_SHARE_CPUCAPACITY | 5367 SD_SHARE_PKG_RESOURCES | 5368 SD_SHARE_POWERDOMAIN)) { 5369 if (sd->groups != sd->groups->next) 5370 return 0; 5371 } 5372 5373 /* Following flags don't use groups */ 5374 if (sd->flags & (SD_WAKE_AFFINE)) 5375 return 0; 5376 5377 return 1; 5378} 5379 5380static int 5381sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent) 5382{ 5383 unsigned long cflags = sd->flags, pflags = parent->flags; 5384 5385 if (sd_degenerate(parent)) 5386 return 1; 5387 5388 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent))) 5389 return 0; 5390 5391 /* Flags needing groups don't count if only 1 group in parent */ 5392 if (parent->groups == parent->groups->next) { 5393 pflags &= ~(SD_LOAD_BALANCE | 5394 SD_BALANCE_NEWIDLE | 5395 SD_BALANCE_FORK | 5396 SD_BALANCE_EXEC | 5397 SD_SHARE_CPUCAPACITY | 5398 SD_SHARE_PKG_RESOURCES | 5399 SD_PREFER_SIBLING | 5400 SD_SHARE_POWERDOMAIN); 5401 if (nr_node_ids == 1) 5402 pflags &= ~SD_SERIALIZE; 5403 } 5404 if (~cflags & pflags) 5405 return 0; 5406 5407 return 1; 5408} 5409 5410static void free_rootdomain(struct rcu_head *rcu) 5411{ 5412 struct root_domain *rd = container_of(rcu, struct root_domain, rcu); 5413 5414 cpupri_cleanup(&rd->cpupri); 5415 cpudl_cleanup(&rd->cpudl); 5416 free_cpumask_var(rd->dlo_mask); 5417 free_cpumask_var(rd->rto_mask); 5418 free_cpumask_var(rd->online); 5419 free_cpumask_var(rd->span); 5420 kfree(rd); 5421} 5422 5423static void rq_attach_root(struct rq *rq, struct root_domain *rd) 5424{ 5425 struct root_domain *old_rd = NULL; 5426 unsigned long flags; 5427 5428 raw_spin_lock_irqsave(&rq->lock, flags); 5429 5430 if (rq->rd) { 5431 old_rd = rq->rd; 5432 5433 if (cpumask_test_cpu(rq->cpu, old_rd->online)) 5434 set_rq_offline(rq); 5435 5436 cpumask_clear_cpu(rq->cpu, old_rd->span); 5437 5438 /* 5439 * If we dont want to free the old_rd yet then 5440 * set old_rd to NULL to skip the freeing later 5441 * in this function: 5442 */ 5443 if (!atomic_dec_and_test(&old_rd->refcount)) 5444 old_rd = NULL; 5445 } 5446 5447 atomic_inc(&rd->refcount); 5448 rq->rd = rd; 5449 5450 cpumask_set_cpu(rq->cpu, rd->span); 5451 if (cpumask_test_cpu(rq->cpu, cpu_active_mask)) 5452 set_rq_online(rq); 5453 5454 raw_spin_unlock_irqrestore(&rq->lock, flags); 5455 5456 if (old_rd) 5457 call_rcu_sched(&old_rd->rcu, free_rootdomain); 5458} 5459 5460static int init_rootdomain(struct root_domain *rd) 5461{ 5462 memset(rd, 0, sizeof(*rd)); 5463 5464 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL)) 5465 goto out; 5466 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL)) 5467 goto free_span; 5468 if (!alloc_cpumask_var(&rd->dlo_mask, GFP_KERNEL)) 5469 goto free_online; 5470 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL)) 5471 goto free_dlo_mask; 5472 5473 init_dl_bw(&rd->dl_bw); 5474 if (cpudl_init(&rd->cpudl) != 0) 5475 goto free_dlo_mask; 5476 5477 if (cpupri_init(&rd->cpupri) != 0) 5478 goto free_rto_mask; 5479 return 0; 5480 5481free_rto_mask: 5482 free_cpumask_var(rd->rto_mask); 5483free_dlo_mask: 5484 free_cpumask_var(rd->dlo_mask); 5485free_online: 5486 free_cpumask_var(rd->online); 5487free_span: 5488 free_cpumask_var(rd->span); 5489out: 5490 return -ENOMEM; 5491} 5492 5493/* 5494 * By default the system creates a single root-domain with all cpus as 5495 * members (mimicking the global state we have today). 5496 */ 5497struct root_domain def_root_domain; 5498 5499static void init_defrootdomain(void) 5500{ 5501 init_rootdomain(&def_root_domain); 5502 5503 atomic_set(&def_root_domain.refcount, 1); 5504} 5505 5506static struct root_domain *alloc_rootdomain(void) 5507{ 5508 struct root_domain *rd; 5509 5510 rd = kmalloc(sizeof(*rd), GFP_KERNEL); 5511 if (!rd) 5512 return NULL; 5513 5514 if (init_rootdomain(rd) != 0) { 5515 kfree(rd); 5516 return NULL; 5517 } 5518 5519 return rd; 5520} 5521 5522static void free_sched_groups(struct sched_group *sg, int free_sgc) 5523{ 5524 struct sched_group *tmp, *first; 5525 5526 if (!sg) 5527 return; 5528 5529 first = sg; 5530 do { 5531 tmp = sg->next; 5532 5533 if (free_sgc && atomic_dec_and_test(&sg->sgc->ref)) 5534 kfree(sg->sgc); 5535 5536 kfree(sg); 5537 sg = tmp; 5538 } while (sg != first); 5539} 5540 5541static void free_sched_domain(struct rcu_head *rcu) 5542{ 5543 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu); 5544 5545 /* 5546 * If its an overlapping domain it has private groups, iterate and 5547 * nuke them all. 5548 */ 5549 if (sd->flags & SD_OVERLAP) { 5550 free_sched_groups(sd->groups, 1); 5551 } else if (atomic_dec_and_test(&sd->groups->ref)) { 5552 kfree(sd->groups->sgc); 5553 kfree(sd->groups); 5554 } 5555 kfree(sd); 5556} 5557 5558static void destroy_sched_domain(struct sched_domain *sd, int cpu) 5559{ 5560 call_rcu(&sd->rcu, free_sched_domain); 5561} 5562 5563static void destroy_sched_domains(struct sched_domain *sd, int cpu) 5564{ 5565 for (; sd; sd = sd->parent) 5566 destroy_sched_domain(sd, cpu); 5567} 5568 5569/* 5570 * Keep a special pointer to the highest sched_domain that has 5571 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this 5572 * allows us to avoid some pointer chasing select_idle_sibling(). 5573 * 5574 * Also keep a unique ID per domain (we use the first cpu number in 5575 * the cpumask of the domain), this allows us to quickly tell if 5576 * two cpus are in the same cache domain, see cpus_share_cache(). 5577 */ 5578DEFINE_PER_CPU(struct sched_domain *, sd_llc); 5579DEFINE_PER_CPU(int, sd_llc_size); 5580DEFINE_PER_CPU(int, sd_llc_id); 5581DEFINE_PER_CPU(struct sched_domain *, sd_numa); 5582DEFINE_PER_CPU(struct sched_domain *, sd_busy); 5583DEFINE_PER_CPU(struct sched_domain *, sd_asym); 5584 5585static void update_top_cache_domain(int cpu) 5586{ 5587 struct sched_domain *sd; 5588 struct sched_domain *busy_sd = NULL; 5589 int id = cpu; 5590 int size = 1; 5591 5592 sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES); 5593 if (sd) { 5594 id = cpumask_first(sched_domain_span(sd)); 5595 size = cpumask_weight(sched_domain_span(sd)); 5596 busy_sd = sd->parent; /* sd_busy */ 5597 } 5598 rcu_assign_pointer(per_cpu(sd_busy, cpu), busy_sd); 5599 5600 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd); 5601 per_cpu(sd_llc_size, cpu) = size; 5602 per_cpu(sd_llc_id, cpu) = id; 5603 5604 sd = lowest_flag_domain(cpu, SD_NUMA); 5605 rcu_assign_pointer(per_cpu(sd_numa, cpu), sd); 5606 5607 sd = highest_flag_domain(cpu, SD_ASYM_PACKING); 5608 rcu_assign_pointer(per_cpu(sd_asym, cpu), sd); 5609} 5610 5611/* 5612 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must 5613 * hold the hotplug lock. 5614 */ 5615static void 5616cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu) 5617{ 5618 struct rq *rq = cpu_rq(cpu); 5619 struct sched_domain *tmp; 5620 5621 /* Remove the sched domains which do not contribute to scheduling. */ 5622 for (tmp = sd; tmp; ) { 5623 struct sched_domain *parent = tmp->parent; 5624 if (!parent) 5625 break; 5626 5627 if (sd_parent_degenerate(tmp, parent)) { 5628 tmp->parent = parent->parent; 5629 if (parent->parent) 5630 parent->parent->child = tmp; 5631 /* 5632 * Transfer SD_PREFER_SIBLING down in case of a 5633 * degenerate parent; the spans match for this 5634 * so the property transfers. 5635 */ 5636 if (parent->flags & SD_PREFER_SIBLING) 5637 tmp->flags |= SD_PREFER_SIBLING; 5638 destroy_sched_domain(parent, cpu); 5639 } else 5640 tmp = tmp->parent; 5641 } 5642 5643 if (sd && sd_degenerate(sd)) { 5644 tmp = sd; 5645 sd = sd->parent; 5646 destroy_sched_domain(tmp, cpu); 5647 if (sd) 5648 sd->child = NULL; 5649 } 5650 5651 sched_domain_debug(sd, cpu); 5652 5653 rq_attach_root(rq, rd); 5654 tmp = rq->sd; 5655 rcu_assign_pointer(rq->sd, sd); 5656 destroy_sched_domains(tmp, cpu); 5657 5658 update_top_cache_domain(cpu); 5659} 5660 5661/* cpus with isolated domains */ 5662static cpumask_var_t cpu_isolated_map; 5663 5664/* Setup the mask of cpus configured for isolated domains */ 5665static int __init isolated_cpu_setup(char *str) 5666{ 5667 alloc_bootmem_cpumask_var(&cpu_isolated_map); 5668 cpulist_parse(str, cpu_isolated_map); 5669 return 1; 5670} 5671 5672__setup("isolcpus=", isolated_cpu_setup); 5673 5674struct s_data { 5675 struct sched_domain ** __percpu sd; 5676 struct root_domain *rd; 5677}; 5678 5679enum s_alloc { 5680 sa_rootdomain, 5681 sa_sd, 5682 sa_sd_storage, 5683 sa_none, 5684}; 5685 5686/* 5687 * Build an iteration mask that can exclude certain CPUs from the upwards 5688 * domain traversal. 5689 * 5690 * Asymmetric node setups can result in situations where the domain tree is of 5691 * unequal depth, make sure to skip domains that already cover the entire 5692 * range. 5693 * 5694 * In that case build_sched_domains() will have terminated the iteration early 5695 * and our sibling sd spans will be empty. Domains should always include the 5696 * cpu they're built on, so check that. 5697 * 5698 */ 5699static void build_group_mask(struct sched_domain *sd, struct sched_group *sg) 5700{ 5701 const struct cpumask *span = sched_domain_span(sd); 5702 struct sd_data *sdd = sd->private; 5703 struct sched_domain *sibling; 5704 int i; 5705 5706 for_each_cpu(i, span) { 5707 sibling = *per_cpu_ptr(sdd->sd, i); 5708 if (!cpumask_test_cpu(i, sched_domain_span(sibling))) 5709 continue; 5710 5711 cpumask_set_cpu(i, sched_group_mask(sg)); 5712 } 5713} 5714 5715/* 5716 * Return the canonical balance cpu for this group, this is the first cpu 5717 * of this group that's also in the iteration mask. 5718 */ 5719int group_balance_cpu(struct sched_group *sg) 5720{ 5721 return cpumask_first_and(sched_group_cpus(sg), sched_group_mask(sg)); 5722} 5723 5724static int 5725build_overlap_sched_groups(struct sched_domain *sd, int cpu) 5726{ 5727 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg; 5728 const struct cpumask *span = sched_domain_span(sd); 5729 struct cpumask *covered = sched_domains_tmpmask; 5730 struct sd_data *sdd = sd->private; 5731 struct sched_domain *child; 5732 int i; 5733 5734 cpumask_clear(covered); 5735 5736 for_each_cpu(i, span) { 5737 struct cpumask *sg_span; 5738 5739 if (cpumask_test_cpu(i, covered)) 5740 continue; 5741 5742 child = *per_cpu_ptr(sdd->sd, i); 5743 5744 /* See the comment near build_group_mask(). */ 5745 if (!cpumask_test_cpu(i, sched_domain_span(child))) 5746 continue; 5747 5748 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(), 5749 GFP_KERNEL, cpu_to_node(cpu)); 5750 5751 if (!sg) 5752 goto fail; 5753 5754 sg_span = sched_group_cpus(sg); 5755 if (child->child) { 5756 child = child->child; 5757 cpumask_copy(sg_span, sched_domain_span(child)); 5758 } else 5759 cpumask_set_cpu(i, sg_span); 5760 5761 cpumask_or(covered, covered, sg_span); 5762 5763 sg->sgc = *per_cpu_ptr(sdd->sgc, i); 5764 if (atomic_inc_return(&sg->sgc->ref) == 1) 5765 build_group_mask(sd, sg); 5766 5767 /* 5768 * Initialize sgc->capacity such that even if we mess up the 5769 * domains and no possible iteration will get us here, we won't 5770 * die on a /0 trap. 5771 */ 5772 sg->sgc->capacity = SCHED_CAPACITY_SCALE * cpumask_weight(sg_span); 5773 sg->sgc->capacity_orig = sg->sgc->capacity; 5774 5775 /* 5776 * Make sure the first group of this domain contains the 5777 * canonical balance cpu. Otherwise the sched_domain iteration 5778 * breaks. See update_sg_lb_stats(). 5779 */ 5780 if ((!groups && cpumask_test_cpu(cpu, sg_span)) || 5781 group_balance_cpu(sg) == cpu) 5782 groups = sg; 5783 5784 if (!first) 5785 first = sg; 5786 if (last) 5787 last->next = sg; 5788 last = sg; 5789 last->next = first; 5790 } 5791 sd->groups = groups; 5792 5793 return 0; 5794 5795fail: 5796 free_sched_groups(first, 0); 5797 5798 return -ENOMEM; 5799} 5800 5801static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg) 5802{ 5803 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu); 5804 struct sched_domain *child = sd->child; 5805 5806 if (child) 5807 cpu = cpumask_first(sched_domain_span(child)); 5808 5809 if (sg) { 5810 *sg = *per_cpu_ptr(sdd->sg, cpu); 5811 (*sg)->sgc = *per_cpu_ptr(sdd->sgc, cpu); 5812 atomic_set(&(*sg)->sgc->ref, 1); /* for claim_allocations */ 5813 } 5814 5815 return cpu; 5816} 5817 5818/* 5819 * build_sched_groups will build a circular linked list of the groups 5820 * covered by the given span, and will set each group's ->cpumask correctly, 5821 * and ->cpu_capacity to 0. 5822 * 5823 * Assumes the sched_domain tree is fully constructed 5824 */ 5825static int 5826build_sched_groups(struct sched_domain *sd, int cpu) 5827{ 5828 struct sched_group *first = NULL, *last = NULL; 5829 struct sd_data *sdd = sd->private; 5830 const struct cpumask *span = sched_domain_span(sd); 5831 struct cpumask *covered; 5832 int i; 5833 5834 get_group(cpu, sdd, &sd->groups); 5835 atomic_inc(&sd->groups->ref); 5836 5837 if (cpu != cpumask_first(span)) 5838 return 0; 5839 5840 lockdep_assert_held(&sched_domains_mutex); 5841 covered = sched_domains_tmpmask; 5842 5843 cpumask_clear(covered); 5844 5845 for_each_cpu(i, span) { 5846 struct sched_group *sg; 5847 int group, j; 5848 5849 if (cpumask_test_cpu(i, covered)) 5850 continue; 5851 5852 group = get_group(i, sdd, &sg); 5853 cpumask_setall(sched_group_mask(sg)); 5854 5855 for_each_cpu(j, span) { 5856 if (get_group(j, sdd, NULL) != group) 5857 continue; 5858 5859 cpumask_set_cpu(j, covered); 5860 cpumask_set_cpu(j, sched_group_cpus(sg)); 5861 } 5862 5863 if (!first) 5864 first = sg; 5865 if (last) 5866 last->next = sg; 5867 last = sg; 5868 } 5869 last->next = first; 5870 5871 return 0; 5872} 5873 5874/* 5875 * Initialize sched groups cpu_capacity. 5876 * 5877 * cpu_capacity indicates the capacity of sched group, which is used while 5878 * distributing the load between different sched groups in a sched domain. 5879 * Typically cpu_capacity for all the groups in a sched domain will be same 5880 * unless there are asymmetries in the topology. If there are asymmetries, 5881 * group having more cpu_capacity will pickup more load compared to the 5882 * group having less cpu_capacity. 5883 */ 5884static void init_sched_groups_capacity(int cpu, struct sched_domain *sd) 5885{ 5886 struct sched_group *sg = sd->groups; 5887 5888 WARN_ON(!sg); 5889 5890 do { 5891 sg->group_weight = cpumask_weight(sched_group_cpus(sg)); 5892 sg = sg->next; 5893 } while (sg != sd->groups); 5894 5895 if (cpu != group_balance_cpu(sg)) 5896 return; 5897 5898 update_group_capacity(sd, cpu); 5899 atomic_set(&sg->sgc->nr_busy_cpus, sg->group_weight); 5900} 5901 5902/* 5903 * Initializers for schedule domains 5904 * Non-inlined to reduce accumulated stack pressure in build_sched_domains() 5905 */ 5906 5907static int default_relax_domain_level = -1; 5908int sched_domain_level_max; 5909 5910static int __init setup_relax_domain_level(char *str) 5911{ 5912 if (kstrtoint(str, 0, &default_relax_domain_level)) 5913 pr_warn("Unable to set relax_domain_level\n"); 5914 5915 return 1; 5916} 5917__setup("relax_domain_level=", setup_relax_domain_level); 5918 5919static void set_domain_attribute(struct sched_domain *sd, 5920 struct sched_domain_attr *attr) 5921{ 5922 int request; 5923 5924 if (!attr || attr->relax_domain_level < 0) { 5925 if (default_relax_domain_level < 0) 5926 return; 5927 else 5928 request = default_relax_domain_level; 5929 } else 5930 request = attr->relax_domain_level; 5931 if (request < sd->level) { 5932 /* turn off idle balance on this domain */ 5933 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE); 5934 } else { 5935 /* turn on idle balance on this domain */ 5936 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE); 5937 } 5938} 5939 5940static void __sdt_free(const struct cpumask *cpu_map); 5941static int __sdt_alloc(const struct cpumask *cpu_map); 5942 5943static void __free_domain_allocs(struct s_data *d, enum s_alloc what, 5944 const struct cpumask *cpu_map) 5945{ 5946 switch (what) { 5947 case sa_rootdomain: 5948 if (!atomic_read(&d->rd->refcount)) 5949 free_rootdomain(&d->rd->rcu); /* fall through */ 5950 case sa_sd: 5951 free_percpu(d->sd); /* fall through */ 5952 case sa_sd_storage: 5953 __sdt_free(cpu_map); /* fall through */ 5954 case sa_none: 5955 break; 5956 } 5957} 5958 5959static enum s_alloc __visit_domain_allocation_hell(struct s_data *d, 5960 const struct cpumask *cpu_map) 5961{ 5962 memset(d, 0, sizeof(*d)); 5963 5964 if (__sdt_alloc(cpu_map)) 5965 return sa_sd_storage; 5966 d->sd = alloc_percpu(struct sched_domain *); 5967 if (!d->sd) 5968 return sa_sd_storage; 5969 d->rd = alloc_rootdomain(); 5970 if (!d->rd) 5971 return sa_sd; 5972 return sa_rootdomain; 5973} 5974 5975/* 5976 * NULL the sd_data elements we've used to build the sched_domain and 5977 * sched_group structure so that the subsequent __free_domain_allocs() 5978 * will not free the data we're using. 5979 */ 5980static void claim_allocations(int cpu, struct sched_domain *sd) 5981{ 5982 struct sd_data *sdd = sd->private; 5983 5984 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd); 5985 *per_cpu_ptr(sdd->sd, cpu) = NULL; 5986 5987 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref)) 5988 *per_cpu_ptr(sdd->sg, cpu) = NULL; 5989 5990 if (atomic_read(&(*per_cpu_ptr(sdd->sgc, cpu))->ref)) 5991 *per_cpu_ptr(sdd->sgc, cpu) = NULL; 5992} 5993 5994#ifdef CONFIG_NUMA 5995static int sched_domains_numa_levels; 5996static int *sched_domains_numa_distance; 5997static struct cpumask ***sched_domains_numa_masks; 5998static int sched_domains_curr_level; 5999#endif 6000 6001/* 6002 * SD_flags allowed in topology descriptions. 6003 * 6004 * SD_SHARE_CPUCAPACITY - describes SMT topologies 6005 * SD_SHARE_PKG_RESOURCES - describes shared caches 6006 * SD_NUMA - describes NUMA topologies 6007 * SD_SHARE_POWERDOMAIN - describes shared power domain 6008 * 6009 * Odd one out: 6010 * SD_ASYM_PACKING - describes SMT quirks 6011 */ 6012#define TOPOLOGY_SD_FLAGS \ 6013 (SD_SHARE_CPUCAPACITY | \ 6014 SD_SHARE_PKG_RESOURCES | \ 6015 SD_NUMA | \ 6016 SD_ASYM_PACKING | \ 6017 SD_SHARE_POWERDOMAIN) 6018 6019static struct sched_domain * 6020sd_init(struct sched_domain_topology_level *tl, int cpu) 6021{ 6022 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); 6023 int sd_weight, sd_flags = 0; 6024 6025#ifdef CONFIG_NUMA 6026 /* 6027 * Ugly hack to pass state to sd_numa_mask()... 6028 */ 6029 sched_domains_curr_level = tl->numa_level; 6030#endif 6031 6032 sd_weight = cpumask_weight(tl->mask(cpu)); 6033 6034 if (tl->sd_flags) 6035 sd_flags = (*tl->sd_flags)(); 6036 if (WARN_ONCE(sd_flags & ~TOPOLOGY_SD_FLAGS, 6037 "wrong sd_flags in topology description\n")) 6038 sd_flags &= ~TOPOLOGY_SD_FLAGS; 6039 6040 *sd = (struct sched_domain){ 6041 .min_interval = sd_weight, 6042 .max_interval = 2*sd_weight, 6043 .busy_factor = 32, 6044 .imbalance_pct = 125, 6045 6046 .cache_nice_tries = 0, 6047 .busy_idx = 0, 6048 .idle_idx = 0, 6049 .newidle_idx = 0, 6050 .wake_idx = 0, 6051 .forkexec_idx = 0, 6052 6053 .flags = 1*SD_LOAD_BALANCE 6054 | 1*SD_BALANCE_NEWIDLE 6055 | 1*SD_BALANCE_EXEC 6056 | 1*SD_BALANCE_FORK 6057 | 0*SD_BALANCE_WAKE 6058 | 1*SD_WAKE_AFFINE 6059 | 0*SD_SHARE_CPUCAPACITY 6060 | 0*SD_SHARE_PKG_RESOURCES 6061 | 0*SD_SERIALIZE 6062 | 0*SD_PREFER_SIBLING 6063 | 0*SD_NUMA 6064 | sd_flags 6065 , 6066 6067 .last_balance = jiffies, 6068 .balance_interval = sd_weight, 6069 .smt_gain = 0, 6070 .max_newidle_lb_cost = 0, 6071 .next_decay_max_lb_cost = jiffies, 6072#ifdef CONFIG_SCHED_DEBUG 6073 .name = tl->name, 6074#endif 6075 }; 6076 6077 /* 6078 * Convert topological properties into behaviour. 6079 */ 6080 6081 if (sd->flags & SD_SHARE_CPUCAPACITY) { 6082 sd->imbalance_pct = 110; 6083 sd->smt_gain = 1178; /* ~15% */ 6084 6085 } else if (sd->flags & SD_SHARE_PKG_RESOURCES) { 6086 sd->imbalance_pct = 117; 6087 sd->cache_nice_tries = 1; 6088 sd->busy_idx = 2; 6089 6090#ifdef CONFIG_NUMA 6091 } else if (sd->flags & SD_NUMA) { 6092 sd->cache_nice_tries = 2; 6093 sd->busy_idx = 3; 6094 sd->idle_idx = 2; 6095 6096 sd->flags |= SD_SERIALIZE; 6097 if (sched_domains_numa_distance[tl->numa_level] > RECLAIM_DISTANCE) { 6098 sd->flags &= ~(SD_BALANCE_EXEC | 6099 SD_BALANCE_FORK | 6100 SD_WAKE_AFFINE); 6101 } 6102 6103#endif 6104 } else { 6105 sd->flags |= SD_PREFER_SIBLING; 6106 sd->cache_nice_tries = 1; 6107 sd->busy_idx = 2; 6108 sd->idle_idx = 1; 6109 } 6110 6111 sd->private = &tl->data; 6112 6113 return sd; 6114} 6115 6116/* 6117 * Topology list, bottom-up. 6118 */ 6119static struct sched_domain_topology_level default_topology[] = { 6120#ifdef CONFIG_SCHED_SMT 6121 { cpu_smt_mask, cpu_smt_flags, SD_INIT_NAME(SMT) }, 6122#endif 6123#ifdef CONFIG_SCHED_MC 6124 { cpu_coregroup_mask, cpu_core_flags, SD_INIT_NAME(MC) }, 6125#endif 6126 { cpu_cpu_mask, SD_INIT_NAME(DIE) }, 6127 { NULL, }, 6128}; 6129 6130struct sched_domain_topology_level *sched_domain_topology = default_topology; 6131 6132#define for_each_sd_topology(tl) \ 6133 for (tl = sched_domain_topology; tl->mask; tl++) 6134 6135void set_sched_topology(struct sched_domain_topology_level *tl) 6136{ 6137 sched_domain_topology = tl; 6138} 6139 6140#ifdef CONFIG_NUMA 6141 6142static const struct cpumask *sd_numa_mask(int cpu) 6143{ 6144 return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)]; 6145} 6146 6147static void sched_numa_warn(const char *str) 6148{ 6149 static int done = false; 6150 int i,j; 6151 6152 if (done) 6153 return; 6154 6155 done = true; 6156 6157 printk(KERN_WARNING "ERROR: %s\n\n", str); 6158 6159 for (i = 0; i < nr_node_ids; i++) { 6160 printk(KERN_WARNING " "); 6161 for (j = 0; j < nr_node_ids; j++) 6162 printk(KERN_CONT "%02d ", node_distance(i,j)); 6163 printk(KERN_CONT "\n"); 6164 } 6165 printk(KERN_WARNING "\n"); 6166} 6167 6168static bool find_numa_distance(int distance) 6169{ 6170 int i; 6171 6172 if (distance == node_distance(0, 0)) 6173 return true; 6174 6175 for (i = 0; i < sched_domains_numa_levels; i++) { 6176 if (sched_domains_numa_distance[i] == distance) 6177 return true; 6178 } 6179 6180 return false; 6181} 6182 6183static void sched_init_numa(void) 6184{ 6185 int next_distance, curr_distance = node_distance(0, 0); 6186 struct sched_domain_topology_level *tl; 6187 int level = 0; 6188 int i, j, k; 6189 6190 sched_domains_numa_distance = kzalloc(sizeof(int) * nr_node_ids, GFP_KERNEL); 6191 if (!sched_domains_numa_distance) 6192 return; 6193 6194 /* 6195 * O(nr_nodes^2) deduplicating selection sort -- in order to find the 6196 * unique distances in the node_distance() table. 6197 * 6198 * Assumes node_distance(0,j) includes all distances in 6199 * node_distance(i,j) in order to avoid cubic time. 6200 */ 6201 next_distance = curr_distance; 6202 for (i = 0; i < nr_node_ids; i++) { 6203 for (j = 0; j < nr_node_ids; j++) { 6204 for (k = 0; k < nr_node_ids; k++) { 6205 int distance = node_distance(i, k); 6206 6207 if (distance > curr_distance && 6208 (distance < next_distance || 6209 next_distance == curr_distance)) 6210 next_distance = distance; 6211 6212 /* 6213 * While not a strong assumption it would be nice to know 6214 * about cases where if node A is connected to B, B is not 6215 * equally connected to A. 6216 */ 6217 if (sched_debug() && node_distance(k, i) != distance) 6218 sched_numa_warn("Node-distance not symmetric"); 6219 6220 if (sched_debug() && i && !find_numa_distance(distance)) 6221 sched_numa_warn("Node-0 not representative"); 6222 } 6223 if (next_distance != curr_distance) { 6224 sched_domains_numa_distance[level++] = next_distance; 6225 sched_domains_numa_levels = level; 6226 curr_distance = next_distance; 6227 } else break; 6228 } 6229 6230 /* 6231 * In case of sched_debug() we verify the above assumption. 6232 */ 6233 if (!sched_debug()) 6234 break; 6235 } 6236 /* 6237 * 'level' contains the number of unique distances, excluding the 6238 * identity distance node_distance(i,i). 6239 * 6240 * The sched_domains_numa_distance[] array includes the actual distance 6241 * numbers. 6242 */ 6243 6244 /* 6245 * Here, we should temporarily reset sched_domains_numa_levels to 0. 6246 * If it fails to allocate memory for array sched_domains_numa_masks[][], 6247 * the array will contain less then 'level' members. This could be 6248 * dangerous when we use it to iterate array sched_domains_numa_masks[][] 6249 * in other functions. 6250 * 6251 * We reset it to 'level' at the end of this function. 6252 */ 6253 sched_domains_numa_levels = 0; 6254 6255 sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL); 6256 if (!sched_domains_numa_masks) 6257 return; 6258 6259 /* 6260 * Now for each level, construct a mask per node which contains all 6261 * cpus of nodes that are that many hops away from us. 6262 */ 6263 for (i = 0; i < level; i++) { 6264 sched_domains_numa_masks[i] = 6265 kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL); 6266 if (!sched_domains_numa_masks[i]) 6267 return; 6268 6269 for (j = 0; j < nr_node_ids; j++) { 6270 struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL); 6271 if (!mask) 6272 return; 6273 6274 sched_domains_numa_masks[i][j] = mask; 6275 6276 for (k = 0; k < nr_node_ids; k++) { 6277 if (node_distance(j, k) > sched_domains_numa_distance[i]) 6278 continue; 6279 6280 cpumask_or(mask, mask, cpumask_of_node(k)); 6281 } 6282 } 6283 } 6284 6285 /* Compute default topology size */ 6286 for (i = 0; sched_domain_topology[i].mask; i++); 6287 6288 tl = kzalloc((i + level + 1) * 6289 sizeof(struct sched_domain_topology_level), GFP_KERNEL); 6290 if (!tl) 6291 return; 6292 6293 /* 6294 * Copy the default topology bits.. 6295 */ 6296 for (i = 0; sched_domain_topology[i].mask; i++) 6297 tl[i] = sched_domain_topology[i]; 6298 6299 /* 6300 * .. and append 'j' levels of NUMA goodness. 6301 */ 6302 for (j = 0; j < level; i++, j++) { 6303 tl[i] = (struct sched_domain_topology_level){ 6304 .mask = sd_numa_mask, 6305 .sd_flags = cpu_numa_flags, 6306 .flags = SDTL_OVERLAP, 6307 .numa_level = j, 6308 SD_INIT_NAME(NUMA) 6309 }; 6310 } 6311 6312 sched_domain_topology = tl; 6313 6314 sched_domains_numa_levels = level; 6315} 6316 6317static void sched_domains_numa_masks_set(int cpu) 6318{ 6319 int i, j; 6320 int node = cpu_to_node(cpu); 6321 6322 for (i = 0; i < sched_domains_numa_levels; i++) { 6323 for (j = 0; j < nr_node_ids; j++) { 6324 if (node_distance(j, node) <= sched_domains_numa_distance[i]) 6325 cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]); 6326 } 6327 } 6328} 6329 6330static void sched_domains_numa_masks_clear(int cpu) 6331{ 6332 int i, j; 6333 for (i = 0; i < sched_domains_numa_levels; i++) { 6334 for (j = 0; j < nr_node_ids; j++) 6335 cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]); 6336 } 6337} 6338 6339/* 6340 * Update sched_domains_numa_masks[level][node] array when new cpus 6341 * are onlined. 6342 */ 6343static int sched_domains_numa_masks_update(struct notifier_block *nfb, 6344 unsigned long action, 6345 void *hcpu) 6346{ 6347 int cpu = (long)hcpu; 6348 6349 switch (action & ~CPU_TASKS_FROZEN) { 6350 case CPU_ONLINE: 6351 sched_domains_numa_masks_set(cpu); 6352 break; 6353 6354 case CPU_DEAD: 6355 sched_domains_numa_masks_clear(cpu); 6356 break; 6357 6358 default: 6359 return NOTIFY_DONE; 6360 } 6361 6362 return NOTIFY_OK; 6363} 6364#else 6365static inline void sched_init_numa(void) 6366{ 6367} 6368 6369static int sched_domains_numa_masks_update(struct notifier_block *nfb, 6370 unsigned long action, 6371 void *hcpu) 6372{ 6373 return 0; 6374} 6375#endif /* CONFIG_NUMA */ 6376 6377static int __sdt_alloc(const struct cpumask *cpu_map) 6378{ 6379 struct sched_domain_topology_level *tl; 6380 int j; 6381 6382 for_each_sd_topology(tl) { 6383 struct sd_data *sdd = &tl->data; 6384 6385 sdd->sd = alloc_percpu(struct sched_domain *); 6386 if (!sdd->sd) 6387 return -ENOMEM; 6388 6389 sdd->sg = alloc_percpu(struct sched_group *); 6390 if (!sdd->sg) 6391 return -ENOMEM; 6392 6393 sdd->sgc = alloc_percpu(struct sched_group_capacity *); 6394 if (!sdd->sgc) 6395 return -ENOMEM; 6396 6397 for_each_cpu(j, cpu_map) { 6398 struct sched_domain *sd; 6399 struct sched_group *sg; 6400 struct sched_group_capacity *sgc; 6401 6402 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(), 6403 GFP_KERNEL, cpu_to_node(j)); 6404 if (!sd) 6405 return -ENOMEM; 6406 6407 *per_cpu_ptr(sdd->sd, j) = sd; 6408 6409 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(), 6410 GFP_KERNEL, cpu_to_node(j)); 6411 if (!sg) 6412 return -ENOMEM; 6413 6414 sg->next = sg; 6415 6416 *per_cpu_ptr(sdd->sg, j) = sg; 6417 6418 sgc = kzalloc_node(sizeof(struct sched_group_capacity) + cpumask_size(), 6419 GFP_KERNEL, cpu_to_node(j)); 6420 if (!sgc) 6421 return -ENOMEM; 6422 6423 *per_cpu_ptr(sdd->sgc, j) = sgc; 6424 } 6425 } 6426 6427 return 0; 6428} 6429 6430static void __sdt_free(const struct cpumask *cpu_map) 6431{ 6432 struct sched_domain_topology_level *tl; 6433 int j; 6434 6435 for_each_sd_topology(tl) { 6436 struct sd_data *sdd = &tl->data; 6437 6438 for_each_cpu(j, cpu_map) { 6439 struct sched_domain *sd; 6440 6441 if (sdd->sd) { 6442 sd = *per_cpu_ptr(sdd->sd, j); 6443 if (sd && (sd->flags & SD_OVERLAP)) 6444 free_sched_groups(sd->groups, 0); 6445 kfree(*per_cpu_ptr(sdd->sd, j)); 6446 } 6447 6448 if (sdd->sg) 6449 kfree(*per_cpu_ptr(sdd->sg, j)); 6450 if (sdd->sgc) 6451 kfree(*per_cpu_ptr(sdd->sgc, j)); 6452 } 6453 free_percpu(sdd->sd); 6454 sdd->sd = NULL; 6455 free_percpu(sdd->sg); 6456 sdd->sg = NULL; 6457 free_percpu(sdd->sgc); 6458 sdd->sgc = NULL; 6459 } 6460} 6461 6462struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl, 6463 const struct cpumask *cpu_map, struct sched_domain_attr *attr, 6464 struct sched_domain *child, int cpu) 6465{ 6466 struct sched_domain *sd = sd_init(tl, cpu); 6467 if (!sd) 6468 return child; 6469 6470 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu)); 6471 if (child) { 6472 sd->level = child->level + 1; 6473 sched_domain_level_max = max(sched_domain_level_max, sd->level); 6474 child->parent = sd; 6475 sd->child = child; 6476 } 6477 set_domain_attribute(sd, attr); 6478 6479 return sd; 6480} 6481 6482/* 6483 * Build sched domains for a given set of cpus and attach the sched domains 6484 * to the individual cpus 6485 */ 6486static int build_sched_domains(const struct cpumask *cpu_map, 6487 struct sched_domain_attr *attr) 6488{ 6489 enum s_alloc alloc_state; 6490 struct sched_domain *sd; 6491 struct s_data d; 6492 int i, ret = -ENOMEM; 6493 6494 alloc_state = __visit_domain_allocation_hell(&d, cpu_map); 6495 if (alloc_state != sa_rootdomain) 6496 goto error; 6497 6498 /* Set up domains for cpus specified by the cpu_map. */ 6499 for_each_cpu(i, cpu_map) { 6500 struct sched_domain_topology_level *tl; 6501 6502 sd = NULL; 6503 for_each_sd_topology(tl) { 6504 sd = build_sched_domain(tl, cpu_map, attr, sd, i); 6505 if (tl == sched_domain_topology) 6506 *per_cpu_ptr(d.sd, i) = sd; 6507 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP)) 6508 sd->flags |= SD_OVERLAP; 6509 if (cpumask_equal(cpu_map, sched_domain_span(sd))) 6510 break; 6511 } 6512 } 6513 6514 /* Build the groups for the domains */ 6515 for_each_cpu(i, cpu_map) { 6516 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) { 6517 sd->span_weight = cpumask_weight(sched_domain_span(sd)); 6518 if (sd->flags & SD_OVERLAP) { 6519 if (build_overlap_sched_groups(sd, i)) 6520 goto error; 6521 } else { 6522 if (build_sched_groups(sd, i)) 6523 goto error; 6524 } 6525 } 6526 } 6527 6528 /* Calculate CPU capacity for physical packages and nodes */ 6529 for (i = nr_cpumask_bits-1; i >= 0; i--) { 6530 if (!cpumask_test_cpu(i, cpu_map)) 6531 continue; 6532 6533 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) { 6534 claim_allocations(i, sd); 6535 init_sched_groups_capacity(i, sd); 6536 } 6537 } 6538 6539 /* Attach the domains */ 6540 rcu_read_lock(); 6541 for_each_cpu(i, cpu_map) { 6542 sd = *per_cpu_ptr(d.sd, i); 6543 cpu_attach_domain(sd, d.rd, i); 6544 } 6545 rcu_read_unlock(); 6546 6547 ret = 0; 6548error: 6549 __free_domain_allocs(&d, alloc_state, cpu_map); 6550 return ret; 6551} 6552 6553static cpumask_var_t *doms_cur; /* current sched domains */ 6554static int ndoms_cur; /* number of sched domains in 'doms_cur' */ 6555static struct sched_domain_attr *dattr_cur; 6556 /* attribues of custom domains in 'doms_cur' */ 6557 6558/* 6559 * Special case: If a kmalloc of a doms_cur partition (array of 6560 * cpumask) fails, then fallback to a single sched domain, 6561 * as determined by the single cpumask fallback_doms. 6562 */ 6563static cpumask_var_t fallback_doms; 6564 6565/* 6566 * arch_update_cpu_topology lets virtualized architectures update the 6567 * cpu core maps. It is supposed to return 1 if the topology changed 6568 * or 0 if it stayed the same. 6569 */ 6570int __weak arch_update_cpu_topology(void) 6571{ 6572 return 0; 6573} 6574 6575cpumask_var_t *alloc_sched_domains(unsigned int ndoms) 6576{ 6577 int i; 6578 cpumask_var_t *doms; 6579 6580 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL); 6581 if (!doms) 6582 return NULL; 6583 for (i = 0; i < ndoms; i++) { 6584 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) { 6585 free_sched_domains(doms, i); 6586 return NULL; 6587 } 6588 } 6589 return doms; 6590} 6591 6592void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms) 6593{ 6594 unsigned int i; 6595 for (i = 0; i < ndoms; i++) 6596 free_cpumask_var(doms[i]); 6597 kfree(doms); 6598} 6599 6600/* 6601 * Set up scheduler domains and groups. Callers must hold the hotplug lock. 6602 * For now this just excludes isolated cpus, but could be used to 6603 * exclude other special cases in the future. 6604 */ 6605static int init_sched_domains(const struct cpumask *cpu_map) 6606{ 6607 int err; 6608 6609 arch_update_cpu_topology(); 6610 ndoms_cur = 1; 6611 doms_cur = alloc_sched_domains(ndoms_cur); 6612 if (!doms_cur) 6613 doms_cur = &fallback_doms; 6614 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map); 6615 err = build_sched_domains(doms_cur[0], NULL); 6616 register_sched_domain_sysctl(); 6617 6618 return err; 6619} 6620 6621/* 6622 * Detach sched domains from a group of cpus specified in cpu_map 6623 * These cpus will now be attached to the NULL domain 6624 */ 6625static void detach_destroy_domains(const struct cpumask *cpu_map) 6626{ 6627 int i; 6628 6629 rcu_read_lock(); 6630 for_each_cpu(i, cpu_map) 6631 cpu_attach_domain(NULL, &def_root_domain, i); 6632 rcu_read_unlock(); 6633} 6634 6635/* handle null as "default" */ 6636static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur, 6637 struct sched_domain_attr *new, int idx_new) 6638{ 6639 struct sched_domain_attr tmp; 6640 6641 /* fast path */ 6642 if (!new && !cur) 6643 return 1; 6644 6645 tmp = SD_ATTR_INIT; 6646 return !memcmp(cur ? (cur + idx_cur) : &tmp, 6647 new ? (new + idx_new) : &tmp, 6648 sizeof(struct sched_domain_attr)); 6649} 6650 6651/* 6652 * Partition sched domains as specified by the 'ndoms_new' 6653 * cpumasks in the array doms_new[] of cpumasks. This compares 6654 * doms_new[] to the current sched domain partitioning, doms_cur[]. 6655 * It destroys each deleted domain and builds each new domain. 6656 * 6657 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'. 6658 * The masks don't intersect (don't overlap.) We should setup one 6659 * sched domain for each mask. CPUs not in any of the cpumasks will 6660 * not be load balanced. If the same cpumask appears both in the 6661 * current 'doms_cur' domains and in the new 'doms_new', we can leave 6662 * it as it is. 6663 * 6664 * The passed in 'doms_new' should be allocated using 6665 * alloc_sched_domains. This routine takes ownership of it and will 6666 * free_sched_domains it when done with it. If the caller failed the 6667 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1, 6668 * and partition_sched_domains() will fallback to the single partition 6669 * 'fallback_doms', it also forces the domains to be rebuilt. 6670 * 6671 * If doms_new == NULL it will be replaced with cpu_online_mask. 6672 * ndoms_new == 0 is a special case for destroying existing domains, 6673 * and it will not create the default domain. 6674 * 6675 * Call with hotplug lock held 6676 */ 6677void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[], 6678 struct sched_domain_attr *dattr_new) 6679{ 6680 int i, j, n; 6681 int new_topology; 6682 6683 mutex_lock(&sched_domains_mutex); 6684 6685 /* always unregister in case we don't destroy any domains */ 6686 unregister_sched_domain_sysctl(); 6687 6688 /* Let architecture update cpu core mappings. */ 6689 new_topology = arch_update_cpu_topology(); 6690 6691 n = doms_new ? ndoms_new : 0; 6692 6693 /* Destroy deleted domains */ 6694 for (i = 0; i < ndoms_cur; i++) { 6695 for (j = 0; j < n && !new_topology; j++) { 6696 if (cpumask_equal(doms_cur[i], doms_new[j]) 6697 && dattrs_equal(dattr_cur, i, dattr_new, j)) 6698 goto match1; 6699 } 6700 /* no match - a current sched domain not in new doms_new[] */ 6701 detach_destroy_domains(doms_cur[i]); 6702match1: 6703 ; 6704 } 6705 6706 n = ndoms_cur; 6707 if (doms_new == NULL) { 6708 n = 0; 6709 doms_new = &fallback_doms; 6710 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map); 6711 WARN_ON_ONCE(dattr_new); 6712 } 6713 6714 /* Build new domains */ 6715 for (i = 0; i < ndoms_new; i++) { 6716 for (j = 0; j < n && !new_topology; j++) { 6717 if (cpumask_equal(doms_new[i], doms_cur[j]) 6718 && dattrs_equal(dattr_new, i, dattr_cur, j)) 6719 goto match2; 6720 } 6721 /* no match - add a new doms_new */ 6722 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL); 6723match2: 6724 ; 6725 } 6726 6727 /* Remember the new sched domains */ 6728 if (doms_cur != &fallback_doms) 6729 free_sched_domains(doms_cur, ndoms_cur); 6730 kfree(dattr_cur); /* kfree(NULL) is safe */ 6731 doms_cur = doms_new; 6732 dattr_cur = dattr_new; 6733 ndoms_cur = ndoms_new; 6734 6735 register_sched_domain_sysctl(); 6736 6737 mutex_unlock(&sched_domains_mutex); 6738} 6739 6740static int num_cpus_frozen; /* used to mark begin/end of suspend/resume */ 6741 6742/* 6743 * Update cpusets according to cpu_active mask. If cpusets are 6744 * disabled, cpuset_update_active_cpus() becomes a simple wrapper 6745 * around partition_sched_domains(). 6746 * 6747 * If we come here as part of a suspend/resume, don't touch cpusets because we 6748 * want to restore it back to its original state upon resume anyway. 6749 */ 6750static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action, 6751 void *hcpu) 6752{ 6753 switch (action) { 6754 case CPU_ONLINE_FROZEN: 6755 case CPU_DOWN_FAILED_FROZEN: 6756 6757 /* 6758 * num_cpus_frozen tracks how many CPUs are involved in suspend 6759 * resume sequence. As long as this is not the last online 6760 * operation in the resume sequence, just build a single sched 6761 * domain, ignoring cpusets. 6762 */ 6763 num_cpus_frozen--; 6764 if (likely(num_cpus_frozen)) { 6765 partition_sched_domains(1, NULL, NULL); 6766 break; 6767 } 6768 6769 /* 6770 * This is the last CPU online operation. So fall through and 6771 * restore the original sched domains by considering the 6772 * cpuset configurations. 6773 */ 6774 6775 case CPU_ONLINE: 6776 case CPU_DOWN_FAILED: 6777 cpuset_update_active_cpus(true); 6778 break; 6779 default: 6780 return NOTIFY_DONE; 6781 } 6782 return NOTIFY_OK; 6783} 6784 6785static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action, 6786 void *hcpu) 6787{ 6788 switch (action) { 6789 case CPU_DOWN_PREPARE: 6790 cpuset_update_active_cpus(false); 6791 break; 6792 case CPU_DOWN_PREPARE_FROZEN: 6793 num_cpus_frozen++; 6794 partition_sched_domains(1, NULL, NULL); 6795 break; 6796 default: 6797 return NOTIFY_DONE; 6798 } 6799 return NOTIFY_OK; 6800} 6801 6802void __init sched_init_smp(void) 6803{ 6804 cpumask_var_t non_isolated_cpus; 6805 6806 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL); 6807 alloc_cpumask_var(&fallback_doms, GFP_KERNEL); 6808 6809 sched_init_numa(); 6810 6811 /* 6812 * There's no userspace yet to cause hotplug operations; hence all the 6813 * cpu masks are stable and all blatant races in the below code cannot 6814 * happen. 6815 */ 6816 mutex_lock(&sched_domains_mutex); 6817 init_sched_domains(cpu_active_mask); 6818 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map); 6819 if (cpumask_empty(non_isolated_cpus)) 6820 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus); 6821 mutex_unlock(&sched_domains_mutex); 6822 6823 hotcpu_notifier(sched_domains_numa_masks_update, CPU_PRI_SCHED_ACTIVE); 6824 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE); 6825 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE); 6826 6827 init_hrtick(); 6828 6829 /* Move init over to a non-isolated CPU */ 6830 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0) 6831 BUG(); 6832 sched_init_granularity(); 6833 free_cpumask_var(non_isolated_cpus); 6834 6835 init_sched_rt_class(); 6836 init_sched_dl_class(); 6837} 6838#else 6839void __init sched_init_smp(void) 6840{ 6841 sched_init_granularity(); 6842} 6843#endif /* CONFIG_SMP */ 6844 6845const_debug unsigned int sysctl_timer_migration = 1; 6846 6847int in_sched_functions(unsigned long addr) 6848{ 6849 return in_lock_functions(addr) || 6850 (addr >= (unsigned long)__sched_text_start 6851 && addr < (unsigned long)__sched_text_end); 6852} 6853 6854#ifdef CONFIG_CGROUP_SCHED 6855/* 6856 * Default task group. 6857 * Every task in system belongs to this group at bootup. 6858 */ 6859struct task_group root_task_group; 6860LIST_HEAD(task_groups); 6861#endif 6862 6863DECLARE_PER_CPU(cpumask_var_t, load_balance_mask); 6864 6865void __init sched_init(void) 6866{ 6867 int i, j; 6868 unsigned long alloc_size = 0, ptr; 6869 6870#ifdef CONFIG_FAIR_GROUP_SCHED 6871 alloc_size += 2 * nr_cpu_ids * sizeof(void **); 6872#endif 6873#ifdef CONFIG_RT_GROUP_SCHED 6874 alloc_size += 2 * nr_cpu_ids * sizeof(void **); 6875#endif 6876#ifdef CONFIG_CPUMASK_OFFSTACK 6877 alloc_size += num_possible_cpus() * cpumask_size(); 6878#endif 6879 if (alloc_size) { 6880 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT); 6881 6882#ifdef CONFIG_FAIR_GROUP_SCHED 6883 root_task_group.se = (struct sched_entity **)ptr; 6884 ptr += nr_cpu_ids * sizeof(void **); 6885 6886 root_task_group.cfs_rq = (struct cfs_rq **)ptr; 6887 ptr += nr_cpu_ids * sizeof(void **); 6888 6889#endif /* CONFIG_FAIR_GROUP_SCHED */ 6890#ifdef CONFIG_RT_GROUP_SCHED 6891 root_task_group.rt_se = (struct sched_rt_entity **)ptr; 6892 ptr += nr_cpu_ids * sizeof(void **); 6893 6894 root_task_group.rt_rq = (struct rt_rq **)ptr; 6895 ptr += nr_cpu_ids * sizeof(void **); 6896 6897#endif /* CONFIG_RT_GROUP_SCHED */ 6898#ifdef CONFIG_CPUMASK_OFFSTACK 6899 for_each_possible_cpu(i) { 6900 per_cpu(load_balance_mask, i) = (void *)ptr; 6901 ptr += cpumask_size(); 6902 } 6903#endif /* CONFIG_CPUMASK_OFFSTACK */ 6904 } 6905 6906 init_rt_bandwidth(&def_rt_bandwidth, 6907 global_rt_period(), global_rt_runtime()); 6908 init_dl_bandwidth(&def_dl_bandwidth, 6909 global_rt_period(), global_rt_runtime()); 6910 6911#ifdef CONFIG_SMP 6912 init_defrootdomain(); 6913#endif 6914 6915#ifdef CONFIG_RT_GROUP_SCHED 6916 init_rt_bandwidth(&root_task_group.rt_bandwidth, 6917 global_rt_period(), global_rt_runtime()); 6918#endif /* CONFIG_RT_GROUP_SCHED */ 6919 6920#ifdef CONFIG_CGROUP_SCHED 6921 list_add(&root_task_group.list, &task_groups); 6922 INIT_LIST_HEAD(&root_task_group.children); 6923 INIT_LIST_HEAD(&root_task_group.siblings); 6924 autogroup_init(&init_task); 6925 6926#endif /* CONFIG_CGROUP_SCHED */ 6927 6928 for_each_possible_cpu(i) { 6929 struct rq *rq; 6930 6931 rq = cpu_rq(i); 6932 raw_spin_lock_init(&rq->lock); 6933 rq->nr_running = 0; 6934 rq->calc_load_active = 0; 6935 rq->calc_load_update = jiffies + LOAD_FREQ; 6936 init_cfs_rq(&rq->cfs); 6937 init_rt_rq(&rq->rt, rq); 6938 init_dl_rq(&rq->dl, rq); 6939#ifdef CONFIG_FAIR_GROUP_SCHED 6940 root_task_group.shares = ROOT_TASK_GROUP_LOAD; 6941 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list); 6942 /* 6943 * How much cpu bandwidth does root_task_group get? 6944 * 6945 * In case of task-groups formed thr' the cgroup filesystem, it 6946 * gets 100% of the cpu resources in the system. This overall 6947 * system cpu resource is divided among the tasks of 6948 * root_task_group and its child task-groups in a fair manner, 6949 * based on each entity's (task or task-group's) weight 6950 * (se->load.weight). 6951 * 6952 * In other words, if root_task_group has 10 tasks of weight 6953 * 1024) and two child groups A0 and A1 (of weight 1024 each), 6954 * then A0's share of the cpu resource is: 6955 * 6956 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33% 6957 * 6958 * We achieve this by letting root_task_group's tasks sit 6959 * directly in rq->cfs (i.e root_task_group->se[] = NULL). 6960 */ 6961 init_cfs_bandwidth(&root_task_group.cfs_bandwidth); 6962 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL); 6963#endif /* CONFIG_FAIR_GROUP_SCHED */ 6964 6965 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime; 6966#ifdef CONFIG_RT_GROUP_SCHED 6967 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL); 6968#endif 6969 6970 for (j = 0; j < CPU_LOAD_IDX_MAX; j++) 6971 rq->cpu_load[j] = 0; 6972 6973 rq->last_load_update_tick = jiffies; 6974 6975#ifdef CONFIG_SMP 6976 rq->sd = NULL; 6977 rq->rd = NULL; 6978 rq->cpu_capacity = SCHED_CAPACITY_SCALE; 6979 rq->post_schedule = 0; 6980 rq->active_balance = 0; 6981 rq->next_balance = jiffies; 6982 rq->push_cpu = 0; 6983 rq->cpu = i; 6984 rq->online = 0; 6985 rq->idle_stamp = 0; 6986 rq->avg_idle = 2*sysctl_sched_migration_cost; 6987 rq->max_idle_balance_cost = sysctl_sched_migration_cost; 6988 6989 INIT_LIST_HEAD(&rq->cfs_tasks); 6990 6991 rq_attach_root(rq, &def_root_domain); 6992#ifdef CONFIG_NO_HZ_COMMON 6993 rq->nohz_flags = 0; 6994#endif 6995#ifdef CONFIG_NO_HZ_FULL 6996 rq->last_sched_tick = 0; 6997#endif 6998#endif 6999 init_rq_hrtick(rq); 7000 atomic_set(&rq->nr_iowait, 0); 7001 } 7002 7003 set_load_weight(&init_task); 7004 7005#ifdef CONFIG_PREEMPT_NOTIFIERS 7006 INIT_HLIST_HEAD(&init_task.preempt_notifiers); 7007#endif 7008 7009 /* 7010 * The boot idle thread does lazy MMU switching as well: 7011 */ 7012 atomic_inc(&init_mm.mm_count); 7013 enter_lazy_tlb(&init_mm, current); 7014 7015 /* 7016 * Make us the idle thread. Technically, schedule() should not be 7017 * called from this thread, however somewhere below it might be, 7018 * but because we are the idle thread, we just pick up running again 7019 * when this runqueue becomes "idle". 7020 */ 7021 init_idle(current, smp_processor_id()); 7022 7023 calc_load_update = jiffies + LOAD_FREQ; 7024 7025 /* 7026 * During early bootup we pretend to be a normal task: 7027 */ 7028 current->sched_class = &fair_sched_class; 7029 7030#ifdef CONFIG_SMP 7031 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT); 7032 /* May be allocated at isolcpus cmdline parse time */ 7033 if (cpu_isolated_map == NULL) 7034 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT); 7035 idle_thread_set_boot_cpu(); 7036 set_cpu_rq_start_time(); 7037#endif 7038 init_sched_fair_class(); 7039 7040 scheduler_running = 1; 7041} 7042 7043#ifdef CONFIG_DEBUG_ATOMIC_SLEEP 7044static inline int preempt_count_equals(int preempt_offset) 7045{ 7046 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth(); 7047 7048 return (nested == preempt_offset); 7049} 7050 7051void __might_sleep(const char *file, int line, int preempt_offset) 7052{ 7053 static unsigned long prev_jiffy; /* ratelimiting */ 7054 7055 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */ 7056 if ((preempt_count_equals(preempt_offset) && !irqs_disabled() && 7057 !is_idle_task(current)) || 7058 system_state != SYSTEM_RUNNING || oops_in_progress) 7059 return; 7060 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy) 7061 return; 7062 prev_jiffy = jiffies; 7063 7064 printk(KERN_ERR 7065 "BUG: sleeping function called from invalid context at %s:%d\n", 7066 file, line); 7067 printk(KERN_ERR 7068 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n", 7069 in_atomic(), irqs_disabled(), 7070 current->pid, current->comm); 7071 7072 debug_show_held_locks(current); 7073 if (irqs_disabled()) 7074 print_irqtrace_events(current); 7075#ifdef CONFIG_DEBUG_PREEMPT 7076 if (!preempt_count_equals(preempt_offset)) { 7077 pr_err("Preemption disabled at:"); 7078 print_ip_sym(current->preempt_disable_ip); 7079 pr_cont("\n"); 7080 } 7081#endif 7082 dump_stack(); 7083} 7084EXPORT_SYMBOL(__might_sleep); 7085#endif 7086 7087#ifdef CONFIG_MAGIC_SYSRQ 7088static void normalize_task(struct rq *rq, struct task_struct *p) 7089{ 7090 const struct sched_class *prev_class = p->sched_class; 7091 struct sched_attr attr = { 7092 .sched_policy = SCHED_NORMAL, 7093 }; 7094 int old_prio = p->prio; 7095 int on_rq; 7096 7097 on_rq = p->on_rq; 7098 if (on_rq) 7099 dequeue_task(rq, p, 0); 7100 __setscheduler(rq, p, &attr); 7101 if (on_rq) { 7102 enqueue_task(rq, p, 0); 7103 resched_task(rq->curr); 7104 } 7105 7106 check_class_changed(rq, p, prev_class, old_prio); 7107} 7108 7109void normalize_rt_tasks(void) 7110{ 7111 struct task_struct *g, *p; 7112 unsigned long flags; 7113 struct rq *rq; 7114 7115 read_lock_irqsave(&tasklist_lock, flags); 7116 do_each_thread(g, p) { 7117 /* 7118 * Only normalize user tasks: 7119 */ 7120 if (!p->mm) 7121 continue; 7122 7123 p->se.exec_start = 0; 7124#ifdef CONFIG_SCHEDSTATS 7125 p->se.statistics.wait_start = 0; 7126 p->se.statistics.sleep_start = 0; 7127 p->se.statistics.block_start = 0; 7128#endif 7129 7130 if (!dl_task(p) && !rt_task(p)) { 7131 /* 7132 * Renice negative nice level userspace 7133 * tasks back to 0: 7134 */ 7135 if (task_nice(p) < 0 && p->mm) 7136 set_user_nice(p, 0); 7137 continue; 7138 } 7139 7140 raw_spin_lock(&p->pi_lock); 7141 rq = __task_rq_lock(p); 7142 7143 normalize_task(rq, p); 7144 7145 __task_rq_unlock(rq); 7146 raw_spin_unlock(&p->pi_lock); 7147 } while_each_thread(g, p); 7148 7149 read_unlock_irqrestore(&tasklist_lock, flags); 7150} 7151 7152#endif /* CONFIG_MAGIC_SYSRQ */ 7153 7154#if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) 7155/* 7156 * These functions are only useful for the IA64 MCA handling, or kdb. 7157 * 7158 * They can only be called when the whole system has been 7159 * stopped - every CPU needs to be quiescent, and no scheduling 7160 * activity can take place. Using them for anything else would 7161 * be a serious bug, and as a result, they aren't even visible 7162 * under any other configuration. 7163 */ 7164 7165/** 7166 * curr_task - return the current task for a given cpu. 7167 * @cpu: the processor in question. 7168 * 7169 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED! 7170 * 7171 * Return: The current task for @cpu. 7172 */ 7173struct task_struct *curr_task(int cpu) 7174{ 7175 return cpu_curr(cpu); 7176} 7177 7178#endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */ 7179 7180#ifdef CONFIG_IA64 7181/** 7182 * set_curr_task - set the current task for a given cpu. 7183 * @cpu: the processor in question. 7184 * @p: the task pointer to set. 7185 * 7186 * Description: This function must only be used when non-maskable interrupts 7187 * are serviced on a separate stack. It allows the architecture to switch the 7188 * notion of the current task on a cpu in a non-blocking manner. This function 7189 * must be called with all CPU's synchronized, and interrupts disabled, the 7190 * and caller must save the original value of the current task (see 7191 * curr_task() above) and restore that value before reenabling interrupts and 7192 * re-starting the system. 7193 * 7194 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED! 7195 */ 7196void set_curr_task(int cpu, struct task_struct *p) 7197{ 7198 cpu_curr(cpu) = p; 7199} 7200 7201#endif 7202 7203#ifdef CONFIG_CGROUP_SCHED 7204/* task_group_lock serializes the addition/removal of task groups */ 7205static DEFINE_SPINLOCK(task_group_lock); 7206 7207static void free_sched_group(struct task_group *tg) 7208{ 7209 free_fair_sched_group(tg); 7210 free_rt_sched_group(tg); 7211 autogroup_free(tg); 7212 kfree(tg); 7213} 7214 7215/* allocate runqueue etc for a new task group */ 7216struct task_group *sched_create_group(struct task_group *parent) 7217{ 7218 struct task_group *tg; 7219 7220 tg = kzalloc(sizeof(*tg), GFP_KERNEL); 7221 if (!tg) 7222 return ERR_PTR(-ENOMEM); 7223 7224 if (!alloc_fair_sched_group(tg, parent)) 7225 goto err; 7226 7227 if (!alloc_rt_sched_group(tg, parent)) 7228 goto err; 7229 7230 return tg; 7231 7232err: 7233 free_sched_group(tg); 7234 return ERR_PTR(-ENOMEM); 7235} 7236 7237void sched_online_group(struct task_group *tg, struct task_group *parent) 7238{ 7239 unsigned long flags; 7240 7241 spin_lock_irqsave(&task_group_lock, flags); 7242 list_add_rcu(&tg->list, &task_groups); 7243 7244 WARN_ON(!parent); /* root should already exist */ 7245 7246 tg->parent = parent; 7247 INIT_LIST_HEAD(&tg->children); 7248 list_add_rcu(&tg->siblings, &parent->children); 7249 spin_unlock_irqrestore(&task_group_lock, flags); 7250} 7251 7252/* rcu callback to free various structures associated with a task group */ 7253static void free_sched_group_rcu(struct rcu_head *rhp) 7254{ 7255 /* now it should be safe to free those cfs_rqs */ 7256 free_sched_group(container_of(rhp, struct task_group, rcu)); 7257} 7258 7259/* Destroy runqueue etc associated with a task group */ 7260void sched_destroy_group(struct task_group *tg) 7261{ 7262 /* wait for possible concurrent references to cfs_rqs complete */ 7263 call_rcu(&tg->rcu, free_sched_group_rcu); 7264} 7265 7266void sched_offline_group(struct task_group *tg) 7267{ 7268 unsigned long flags; 7269 int i; 7270 7271 /* end participation in shares distribution */ 7272 for_each_possible_cpu(i) 7273 unregister_fair_sched_group(tg, i); 7274 7275 spin_lock_irqsave(&task_group_lock, flags); 7276 list_del_rcu(&tg->list); 7277 list_del_rcu(&tg->siblings); 7278 spin_unlock_irqrestore(&task_group_lock, flags); 7279} 7280 7281/* change task's runqueue when it moves between groups. 7282 * The caller of this function should have put the task in its new group 7283 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to 7284 * reflect its new group. 7285 */ 7286void sched_move_task(struct task_struct *tsk) 7287{ 7288 struct task_group *tg; 7289 int on_rq, running; 7290 unsigned long flags; 7291 struct rq *rq; 7292 7293 rq = task_rq_lock(tsk, &flags); 7294 7295 running = task_current(rq, tsk); 7296 on_rq = tsk->on_rq; 7297 7298 if (on_rq) 7299 dequeue_task(rq, tsk, 0); 7300 if (unlikely(running)) 7301 tsk->sched_class->put_prev_task(rq, tsk); 7302 7303 tg = container_of(task_css_check(tsk, cpu_cgrp_id, 7304 lockdep_is_held(&tsk->sighand->siglock)), 7305 struct task_group, css); 7306 tg = autogroup_task_group(tsk, tg); 7307 tsk->sched_task_group = tg; 7308 7309#ifdef CONFIG_FAIR_GROUP_SCHED 7310 if (tsk->sched_class->task_move_group) 7311 tsk->sched_class->task_move_group(tsk, on_rq); 7312 else 7313#endif 7314 set_task_rq(tsk, task_cpu(tsk)); 7315 7316 if (unlikely(running)) 7317 tsk->sched_class->set_curr_task(rq); 7318 if (on_rq) 7319 enqueue_task(rq, tsk, 0); 7320 7321 task_rq_unlock(rq, tsk, &flags); 7322} 7323#endif /* CONFIG_CGROUP_SCHED */ 7324 7325#ifdef CONFIG_RT_GROUP_SCHED 7326/* 7327 * Ensure that the real time constraints are schedulable. 7328 */ 7329static DEFINE_MUTEX(rt_constraints_mutex); 7330 7331/* Must be called with tasklist_lock held */ 7332static inline int tg_has_rt_tasks(struct task_group *tg) 7333{ 7334 struct task_struct *g, *p; 7335 7336 do_each_thread(g, p) { 7337 if (rt_task(p) && task_rq(p)->rt.tg == tg) 7338 return 1; 7339 } while_each_thread(g, p); 7340 7341 return 0; 7342} 7343 7344struct rt_schedulable_data { 7345 struct task_group *tg; 7346 u64 rt_period; 7347 u64 rt_runtime; 7348}; 7349 7350static int tg_rt_schedulable(struct task_group *tg, void *data) 7351{ 7352 struct rt_schedulable_data *d = data; 7353 struct task_group *child; 7354 unsigned long total, sum = 0; 7355 u64 period, runtime; 7356 7357 period = ktime_to_ns(tg->rt_bandwidth.rt_period); 7358 runtime = tg->rt_bandwidth.rt_runtime; 7359 7360 if (tg == d->tg) { 7361 period = d->rt_period; 7362 runtime = d->rt_runtime; 7363 } 7364 7365 /* 7366 * Cannot have more runtime than the period. 7367 */ 7368 if (runtime > period && runtime != RUNTIME_INF) 7369 return -EINVAL; 7370 7371 /* 7372 * Ensure we don't starve existing RT tasks. 7373 */ 7374 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg)) 7375 return -EBUSY; 7376 7377 total = to_ratio(period, runtime); 7378 7379 /* 7380 * Nobody can have more than the global setting allows. 7381 */ 7382 if (total > to_ratio(global_rt_period(), global_rt_runtime())) 7383 return -EINVAL; 7384 7385 /* 7386 * The sum of our children's runtime should not exceed our own. 7387 */ 7388 list_for_each_entry_rcu(child, &tg->children, siblings) { 7389 period = ktime_to_ns(child->rt_bandwidth.rt_period); 7390 runtime = child->rt_bandwidth.rt_runtime; 7391 7392 if (child == d->tg) { 7393 period = d->rt_period; 7394 runtime = d->rt_runtime; 7395 } 7396 7397 sum += to_ratio(period, runtime); 7398 } 7399 7400 if (sum > total) 7401 return -EINVAL; 7402 7403 return 0; 7404} 7405 7406static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime) 7407{ 7408 int ret; 7409 7410 struct rt_schedulable_data data = { 7411 .tg = tg, 7412 .rt_period = period, 7413 .rt_runtime = runtime, 7414 }; 7415 7416 rcu_read_lock(); 7417 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data); 7418 rcu_read_unlock(); 7419 7420 return ret; 7421} 7422 7423static int tg_set_rt_bandwidth(struct task_group *tg, 7424 u64 rt_period, u64 rt_runtime) 7425{ 7426 int i, err = 0; 7427 7428 mutex_lock(&rt_constraints_mutex); 7429 read_lock(&tasklist_lock); 7430 err = __rt_schedulable(tg, rt_period, rt_runtime); 7431 if (err) 7432 goto unlock; 7433 7434 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock); 7435 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period); 7436 tg->rt_bandwidth.rt_runtime = rt_runtime; 7437 7438 for_each_possible_cpu(i) { 7439 struct rt_rq *rt_rq = tg->rt_rq[i]; 7440 7441 raw_spin_lock(&rt_rq->rt_runtime_lock); 7442 rt_rq->rt_runtime = rt_runtime; 7443 raw_spin_unlock(&rt_rq->rt_runtime_lock); 7444 } 7445 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock); 7446unlock: 7447 read_unlock(&tasklist_lock); 7448 mutex_unlock(&rt_constraints_mutex); 7449 7450 return err; 7451} 7452 7453static int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us) 7454{ 7455 u64 rt_runtime, rt_period; 7456 7457 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period); 7458 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC; 7459 if (rt_runtime_us < 0) 7460 rt_runtime = RUNTIME_INF; 7461 7462 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime); 7463} 7464 7465static long sched_group_rt_runtime(struct task_group *tg) 7466{ 7467 u64 rt_runtime_us; 7468 7469 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF) 7470 return -1; 7471 7472 rt_runtime_us = tg->rt_bandwidth.rt_runtime; 7473 do_div(rt_runtime_us, NSEC_PER_USEC); 7474 return rt_runtime_us; 7475} 7476 7477static int sched_group_set_rt_period(struct task_group *tg, long rt_period_us) 7478{ 7479 u64 rt_runtime, rt_period; 7480 7481 rt_period = (u64)rt_period_us * NSEC_PER_USEC; 7482 rt_runtime = tg->rt_bandwidth.rt_runtime; 7483 7484 if (rt_period == 0) 7485 return -EINVAL; 7486 7487 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime); 7488} 7489 7490static long sched_group_rt_period(struct task_group *tg) 7491{ 7492 u64 rt_period_us; 7493 7494 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period); 7495 do_div(rt_period_us, NSEC_PER_USEC); 7496 return rt_period_us; 7497} 7498#endif /* CONFIG_RT_GROUP_SCHED */ 7499 7500#ifdef CONFIG_RT_GROUP_SCHED 7501static int sched_rt_global_constraints(void) 7502{ 7503 int ret = 0; 7504 7505 mutex_lock(&rt_constraints_mutex); 7506 read_lock(&tasklist_lock); 7507 ret = __rt_schedulable(NULL, 0, 0); 7508 read_unlock(&tasklist_lock); 7509 mutex_unlock(&rt_constraints_mutex); 7510 7511 return ret; 7512} 7513 7514static int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk) 7515{ 7516 /* Don't accept realtime tasks when there is no way for them to run */ 7517 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0) 7518 return 0; 7519 7520 return 1; 7521} 7522 7523#else /* !CONFIG_RT_GROUP_SCHED */ 7524static int sched_rt_global_constraints(void) 7525{ 7526 unsigned long flags; 7527 int i, ret = 0; 7528 7529 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags); 7530 for_each_possible_cpu(i) { 7531 struct rt_rq *rt_rq = &cpu_rq(i)->rt; 7532 7533 raw_spin_lock(&rt_rq->rt_runtime_lock); 7534 rt_rq->rt_runtime = global_rt_runtime(); 7535 raw_spin_unlock(&rt_rq->rt_runtime_lock); 7536 } 7537 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags); 7538 7539 return ret; 7540} 7541#endif /* CONFIG_RT_GROUP_SCHED */ 7542 7543static int sched_dl_global_constraints(void) 7544{ 7545 u64 runtime = global_rt_runtime(); 7546 u64 period = global_rt_period(); 7547 u64 new_bw = to_ratio(period, runtime); 7548 int cpu, ret = 0; 7549 unsigned long flags; 7550 7551 /* 7552 * Here we want to check the bandwidth not being set to some 7553 * value smaller than the currently allocated bandwidth in 7554 * any of the root_domains. 7555 * 7556 * FIXME: Cycling on all the CPUs is overdoing, but simpler than 7557 * cycling on root_domains... Discussion on different/better 7558 * solutions is welcome! 7559 */ 7560 for_each_possible_cpu(cpu) { 7561 struct dl_bw *dl_b = dl_bw_of(cpu); 7562 7563 raw_spin_lock_irqsave(&dl_b->lock, flags); 7564 if (new_bw < dl_b->total_bw) 7565 ret = -EBUSY; 7566 raw_spin_unlock_irqrestore(&dl_b->lock, flags); 7567 7568 if (ret) 7569 break; 7570 } 7571 7572 return ret; 7573} 7574 7575static void sched_dl_do_global(void) 7576{ 7577 u64 new_bw = -1; 7578 int cpu; 7579 unsigned long flags; 7580 7581 def_dl_bandwidth.dl_period = global_rt_period(); 7582 def_dl_bandwidth.dl_runtime = global_rt_runtime(); 7583 7584 if (global_rt_runtime() != RUNTIME_INF) 7585 new_bw = to_ratio(global_rt_period(), global_rt_runtime()); 7586 7587 /* 7588 * FIXME: As above... 7589 */ 7590 for_each_possible_cpu(cpu) { 7591 struct dl_bw *dl_b = dl_bw_of(cpu); 7592 7593 raw_spin_lock_irqsave(&dl_b->lock, flags); 7594 dl_b->bw = new_bw; 7595 raw_spin_unlock_irqrestore(&dl_b->lock, flags); 7596 } 7597} 7598 7599static int sched_rt_global_validate(void) 7600{ 7601 if (sysctl_sched_rt_period <= 0) 7602 return -EINVAL; 7603 7604 if ((sysctl_sched_rt_runtime != RUNTIME_INF) && 7605 (sysctl_sched_rt_runtime > sysctl_sched_rt_period)) 7606 return -EINVAL; 7607 7608 return 0; 7609} 7610 7611static void sched_rt_do_global(void) 7612{ 7613 def_rt_bandwidth.rt_runtime = global_rt_runtime(); 7614 def_rt_bandwidth.rt_period = ns_to_ktime(global_rt_period()); 7615} 7616 7617int sched_rt_handler(struct ctl_table *table, int write, 7618 void __user *buffer, size_t *lenp, 7619 loff_t *ppos) 7620{ 7621 int old_period, old_runtime; 7622 static DEFINE_MUTEX(mutex); 7623 int ret; 7624 7625 mutex_lock(&mutex); 7626 old_period = sysctl_sched_rt_period; 7627 old_runtime = sysctl_sched_rt_runtime; 7628 7629 ret = proc_dointvec(table, write, buffer, lenp, ppos); 7630 7631 if (!ret && write) { 7632 ret = sched_rt_global_validate(); 7633 if (ret) 7634 goto undo; 7635 7636 ret = sched_rt_global_constraints(); 7637 if (ret) 7638 goto undo; 7639 7640 ret = sched_dl_global_constraints(); 7641 if (ret) 7642 goto undo; 7643 7644 sched_rt_do_global(); 7645 sched_dl_do_global(); 7646 } 7647 if (0) { 7648undo: 7649 sysctl_sched_rt_period = old_period; 7650 sysctl_sched_rt_runtime = old_runtime; 7651 } 7652 mutex_unlock(&mutex); 7653 7654 return ret; 7655} 7656 7657int sched_rr_handler(struct ctl_table *table, int write, 7658 void __user *buffer, size_t *lenp, 7659 loff_t *ppos) 7660{ 7661 int ret; 7662 static DEFINE_MUTEX(mutex); 7663 7664 mutex_lock(&mutex); 7665 ret = proc_dointvec(table, write, buffer, lenp, ppos); 7666 /* make sure that internally we keep jiffies */ 7667 /* also, writing zero resets timeslice to default */ 7668 if (!ret && write) { 7669 sched_rr_timeslice = sched_rr_timeslice <= 0 ? 7670 RR_TIMESLICE : msecs_to_jiffies(sched_rr_timeslice); 7671 } 7672 mutex_unlock(&mutex); 7673 return ret; 7674} 7675 7676#ifdef CONFIG_CGROUP_SCHED 7677 7678static inline struct task_group *css_tg(struct cgroup_subsys_state *css) 7679{ 7680 return css ? container_of(css, struct task_group, css) : NULL; 7681} 7682 7683static struct cgroup_subsys_state * 7684cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css) 7685{ 7686 struct task_group *parent = css_tg(parent_css); 7687 struct task_group *tg; 7688 7689 if (!parent) { 7690 /* This is early initialization for the top cgroup */ 7691 return &root_task_group.css; 7692 } 7693 7694 tg = sched_create_group(parent); 7695 if (IS_ERR(tg)) 7696 return ERR_PTR(-ENOMEM); 7697 7698 return &tg->css; 7699} 7700 7701static int cpu_cgroup_css_online(struct cgroup_subsys_state *css) 7702{ 7703 struct task_group *tg = css_tg(css); 7704 struct task_group *parent = css_tg(css->parent); 7705 7706 if (parent) 7707 sched_online_group(tg, parent); 7708 return 0; 7709} 7710 7711static void cpu_cgroup_css_free(struct cgroup_subsys_state *css) 7712{ 7713 struct task_group *tg = css_tg(css); 7714 7715 sched_destroy_group(tg); 7716} 7717 7718static void cpu_cgroup_css_offline(struct cgroup_subsys_state *css) 7719{ 7720 struct task_group *tg = css_tg(css); 7721 7722 sched_offline_group(tg); 7723} 7724 7725static int cpu_cgroup_can_attach(struct cgroup_subsys_state *css, 7726 struct cgroup_taskset *tset) 7727{ 7728 struct task_struct *task; 7729 7730 cgroup_taskset_for_each(task, tset) { 7731#ifdef CONFIG_RT_GROUP_SCHED 7732 if (!sched_rt_can_attach(css_tg(css), task)) 7733 return -EINVAL; 7734#else 7735 /* We don't support RT-tasks being in separate groups */ 7736 if (task->sched_class != &fair_sched_class) 7737 return -EINVAL; 7738#endif 7739 } 7740 return 0; 7741} 7742 7743static void cpu_cgroup_attach(struct cgroup_subsys_state *css, 7744 struct cgroup_taskset *tset) 7745{ 7746 struct task_struct *task; 7747 7748 cgroup_taskset_for_each(task, tset) 7749 sched_move_task(task); 7750} 7751 7752static void cpu_cgroup_exit(struct cgroup_subsys_state *css, 7753 struct cgroup_subsys_state *old_css, 7754 struct task_struct *task) 7755{ 7756 /* 7757 * cgroup_exit() is called in the copy_process() failure path. 7758 * Ignore this case since the task hasn't ran yet, this avoids 7759 * trying to poke a half freed task state from generic code. 7760 */ 7761 if (!(task->flags & PF_EXITING)) 7762 return; 7763 7764 sched_move_task(task); 7765} 7766 7767#ifdef CONFIG_FAIR_GROUP_SCHED 7768static int cpu_shares_write_u64(struct cgroup_subsys_state *css, 7769 struct cftype *cftype, u64 shareval) 7770{ 7771 return sched_group_set_shares(css_tg(css), scale_load(shareval)); 7772} 7773 7774static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css, 7775 struct cftype *cft) 7776{ 7777 struct task_group *tg = css_tg(css); 7778 7779 return (u64) scale_load_down(tg->shares); 7780} 7781 7782#ifdef CONFIG_CFS_BANDWIDTH 7783static DEFINE_MUTEX(cfs_constraints_mutex); 7784 7785const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */ 7786const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */ 7787 7788static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime); 7789 7790static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota) 7791{ 7792 int i, ret = 0, runtime_enabled, runtime_was_enabled; 7793 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth; 7794 7795 if (tg == &root_task_group) 7796 return -EINVAL; 7797 7798 /* 7799 * Ensure we have at some amount of bandwidth every period. This is 7800 * to prevent reaching a state of large arrears when throttled via 7801 * entity_tick() resulting in prolonged exit starvation. 7802 */ 7803 if (quota < min_cfs_quota_period || period < min_cfs_quota_period) 7804 return -EINVAL; 7805 7806 /* 7807 * Likewise, bound things on the otherside by preventing insane quota 7808 * periods. This also allows us to normalize in computing quota 7809 * feasibility. 7810 */ 7811 if (period > max_cfs_quota_period) 7812 return -EINVAL; 7813 7814 mutex_lock(&cfs_constraints_mutex); 7815 ret = __cfs_schedulable(tg, period, quota); 7816 if (ret) 7817 goto out_unlock; 7818 7819 runtime_enabled = quota != RUNTIME_INF; 7820 runtime_was_enabled = cfs_b->quota != RUNTIME_INF; 7821 /* 7822 * If we need to toggle cfs_bandwidth_used, off->on must occur 7823 * before making related changes, and on->off must occur afterwards 7824 */ 7825 if (runtime_enabled && !runtime_was_enabled) 7826 cfs_bandwidth_usage_inc(); 7827 raw_spin_lock_irq(&cfs_b->lock); 7828 cfs_b->period = ns_to_ktime(period); 7829 cfs_b->quota = quota; 7830 7831 __refill_cfs_bandwidth_runtime(cfs_b); 7832 /* restart the period timer (if active) to handle new period expiry */ 7833 if (runtime_enabled && cfs_b->timer_active) { 7834 /* force a reprogram */ 7835 __start_cfs_bandwidth(cfs_b, true); 7836 } 7837 raw_spin_unlock_irq(&cfs_b->lock); 7838 7839 for_each_possible_cpu(i) { 7840 struct cfs_rq *cfs_rq = tg->cfs_rq[i]; 7841 struct rq *rq = cfs_rq->rq; 7842 7843 raw_spin_lock_irq(&rq->lock); 7844 cfs_rq->runtime_enabled = runtime_enabled; 7845 cfs_rq->runtime_remaining = 0; 7846 7847 if (cfs_rq->throttled) 7848 unthrottle_cfs_rq(cfs_rq); 7849 raw_spin_unlock_irq(&rq->lock); 7850 } 7851 if (runtime_was_enabled && !runtime_enabled) 7852 cfs_bandwidth_usage_dec(); 7853out_unlock: 7854 mutex_unlock(&cfs_constraints_mutex); 7855 7856 return ret; 7857} 7858 7859int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us) 7860{ 7861 u64 quota, period; 7862 7863 period = ktime_to_ns(tg->cfs_bandwidth.period); 7864 if (cfs_quota_us < 0) 7865 quota = RUNTIME_INF; 7866 else 7867 quota = (u64)cfs_quota_us * NSEC_PER_USEC; 7868 7869 return tg_set_cfs_bandwidth(tg, period, quota); 7870} 7871 7872long tg_get_cfs_quota(struct task_group *tg) 7873{ 7874 u64 quota_us; 7875 7876 if (tg->cfs_bandwidth.quota == RUNTIME_INF) 7877 return -1; 7878 7879 quota_us = tg->cfs_bandwidth.quota; 7880 do_div(quota_us, NSEC_PER_USEC); 7881 7882 return quota_us; 7883} 7884 7885int tg_set_cfs_period(struct task_group *tg, long cfs_period_us) 7886{ 7887 u64 quota, period; 7888 7889 period = (u64)cfs_period_us * NSEC_PER_USEC; 7890 quota = tg->cfs_bandwidth.quota; 7891 7892 return tg_set_cfs_bandwidth(tg, period, quota); 7893} 7894 7895long tg_get_cfs_period(struct task_group *tg) 7896{ 7897 u64 cfs_period_us; 7898 7899 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period); 7900 do_div(cfs_period_us, NSEC_PER_USEC); 7901 7902 return cfs_period_us; 7903} 7904 7905static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css, 7906 struct cftype *cft) 7907{ 7908 return tg_get_cfs_quota(css_tg(css)); 7909} 7910 7911static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css, 7912 struct cftype *cftype, s64 cfs_quota_us) 7913{ 7914 return tg_set_cfs_quota(css_tg(css), cfs_quota_us); 7915} 7916 7917static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css, 7918 struct cftype *cft) 7919{ 7920 return tg_get_cfs_period(css_tg(css)); 7921} 7922 7923static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css, 7924 struct cftype *cftype, u64 cfs_period_us) 7925{ 7926 return tg_set_cfs_period(css_tg(css), cfs_period_us); 7927} 7928 7929struct cfs_schedulable_data { 7930 struct task_group *tg; 7931 u64 period, quota; 7932}; 7933 7934/* 7935 * normalize group quota/period to be quota/max_period 7936 * note: units are usecs 7937 */ 7938static u64 normalize_cfs_quota(struct task_group *tg, 7939 struct cfs_schedulable_data *d) 7940{ 7941 u64 quota, period; 7942 7943 if (tg == d->tg) { 7944 period = d->period; 7945 quota = d->quota; 7946 } else { 7947 period = tg_get_cfs_period(tg); 7948 quota = tg_get_cfs_quota(tg); 7949 } 7950 7951 /* note: these should typically be equivalent */ 7952 if (quota == RUNTIME_INF || quota == -1) 7953 return RUNTIME_INF; 7954 7955 return to_ratio(period, quota); 7956} 7957 7958static int tg_cfs_schedulable_down(struct task_group *tg, void *data) 7959{ 7960 struct cfs_schedulable_data *d = data; 7961 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth; 7962 s64 quota = 0, parent_quota = -1; 7963 7964 if (!tg->parent) { 7965 quota = RUNTIME_INF; 7966 } else { 7967 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth; 7968 7969 quota = normalize_cfs_quota(tg, d); 7970 parent_quota = parent_b->hierarchal_quota; 7971 7972 /* 7973 * ensure max(child_quota) <= parent_quota, inherit when no 7974 * limit is set 7975 */ 7976 if (quota == RUNTIME_INF) 7977 quota = parent_quota; 7978 else if (parent_quota != RUNTIME_INF && quota > parent_quota) 7979 return -EINVAL; 7980 } 7981 cfs_b->hierarchal_quota = quota; 7982 7983 return 0; 7984} 7985 7986static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota) 7987{ 7988 int ret; 7989 struct cfs_schedulable_data data = { 7990 .tg = tg, 7991 .period = period, 7992 .quota = quota, 7993 }; 7994 7995 if (quota != RUNTIME_INF) { 7996 do_div(data.period, NSEC_PER_USEC); 7997 do_div(data.quota, NSEC_PER_USEC); 7998 } 7999 8000 rcu_read_lock(); 8001 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data); 8002 rcu_read_unlock(); 8003 8004 return ret; 8005} 8006 8007static int cpu_stats_show(struct seq_file *sf, void *v) 8008{ 8009 struct task_group *tg = css_tg(seq_css(sf)); 8010 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth; 8011 8012 seq_printf(sf, "nr_periods %d\n", cfs_b->nr_periods); 8013 seq_printf(sf, "nr_throttled %d\n", cfs_b->nr_throttled); 8014 seq_printf(sf, "throttled_time %llu\n", cfs_b->throttled_time); 8015 8016 return 0; 8017} 8018#endif /* CONFIG_CFS_BANDWIDTH */ 8019#endif /* CONFIG_FAIR_GROUP_SCHED */ 8020 8021#ifdef CONFIG_RT_GROUP_SCHED 8022static int cpu_rt_runtime_write(struct cgroup_subsys_state *css, 8023 struct cftype *cft, s64 val) 8024{ 8025 return sched_group_set_rt_runtime(css_tg(css), val); 8026} 8027 8028static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css, 8029 struct cftype *cft) 8030{ 8031 return sched_group_rt_runtime(css_tg(css)); 8032} 8033 8034static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css, 8035 struct cftype *cftype, u64 rt_period_us) 8036{ 8037 return sched_group_set_rt_period(css_tg(css), rt_period_us); 8038} 8039 8040static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css, 8041 struct cftype *cft) 8042{ 8043 return sched_group_rt_period(css_tg(css)); 8044} 8045#endif /* CONFIG_RT_GROUP_SCHED */ 8046 8047static struct cftype cpu_files[] = { 8048#ifdef CONFIG_FAIR_GROUP_SCHED 8049 { 8050 .name = "shares", 8051 .read_u64 = cpu_shares_read_u64, 8052 .write_u64 = cpu_shares_write_u64, 8053 }, 8054#endif 8055#ifdef CONFIG_CFS_BANDWIDTH 8056 { 8057 .name = "cfs_quota_us", 8058 .read_s64 = cpu_cfs_quota_read_s64, 8059 .write_s64 = cpu_cfs_quota_write_s64, 8060 }, 8061 { 8062 .name = "cfs_period_us", 8063 .read_u64 = cpu_cfs_period_read_u64, 8064 .write_u64 = cpu_cfs_period_write_u64, 8065 }, 8066 { 8067 .name = "stat", 8068 .seq_show = cpu_stats_show, 8069 }, 8070#endif 8071#ifdef CONFIG_RT_GROUP_SCHED 8072 { 8073 .name = "rt_runtime_us", 8074 .read_s64 = cpu_rt_runtime_read, 8075 .write_s64 = cpu_rt_runtime_write, 8076 }, 8077 { 8078 .name = "rt_period_us", 8079 .read_u64 = cpu_rt_period_read_uint, 8080 .write_u64 = cpu_rt_period_write_uint, 8081 }, 8082#endif 8083 { } /* terminate */ 8084}; 8085 8086struct cgroup_subsys cpu_cgrp_subsys = { 8087 .css_alloc = cpu_cgroup_css_alloc, 8088 .css_free = cpu_cgroup_css_free, 8089 .css_online = cpu_cgroup_css_online, 8090 .css_offline = cpu_cgroup_css_offline, 8091 .can_attach = cpu_cgroup_can_attach, 8092 .attach = cpu_cgroup_attach, 8093 .exit = cpu_cgroup_exit, 8094 .base_cftypes = cpu_files, 8095 .early_init = 1, 8096}; 8097 8098#endif /* CONFIG_CGROUP_SCHED */ 8099 8100void dump_cpu_task(int cpu) 8101{ 8102 pr_info("Task dump for CPU %d:\n", cpu); 8103 sched_show_task(cpu_curr(cpu)); 8104} 8105