core.c revision fd99f91aa007ba255aac44fe6cf21c1db398243a
1/* 2 * kernel/sched/core.c 3 * 4 * Kernel scheduler and related syscalls 5 * 6 * Copyright (C) 1991-2002 Linus Torvalds 7 * 8 * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and 9 * make semaphores SMP safe 10 * 1998-11-19 Implemented schedule_timeout() and related stuff 11 * by Andrea Arcangeli 12 * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar: 13 * hybrid priority-list and round-robin design with 14 * an array-switch method of distributing timeslices 15 * and per-CPU runqueues. Cleanups and useful suggestions 16 * by Davide Libenzi, preemptible kernel bits by Robert Love. 17 * 2003-09-03 Interactivity tuning by Con Kolivas. 18 * 2004-04-02 Scheduler domains code by Nick Piggin 19 * 2007-04-15 Work begun on replacing all interactivity tuning with a 20 * fair scheduling design by Con Kolivas. 21 * 2007-05-05 Load balancing (smp-nice) and other improvements 22 * by Peter Williams 23 * 2007-05-06 Interactivity improvements to CFS by Mike Galbraith 24 * 2007-07-01 Group scheduling enhancements by Srivatsa Vaddagiri 25 * 2007-11-29 RT balancing improvements by Steven Rostedt, Gregory Haskins, 26 * Thomas Gleixner, Mike Kravetz 27 */ 28 29#include <linux/mm.h> 30#include <linux/module.h> 31#include <linux/nmi.h> 32#include <linux/init.h> 33#include <linux/uaccess.h> 34#include <linux/highmem.h> 35#include <asm/mmu_context.h> 36#include <linux/interrupt.h> 37#include <linux/capability.h> 38#include <linux/completion.h> 39#include <linux/kernel_stat.h> 40#include <linux/debug_locks.h> 41#include <linux/perf_event.h> 42#include <linux/security.h> 43#include <linux/notifier.h> 44#include <linux/profile.h> 45#include <linux/freezer.h> 46#include <linux/vmalloc.h> 47#include <linux/blkdev.h> 48#include <linux/delay.h> 49#include <linux/pid_namespace.h> 50#include <linux/smp.h> 51#include <linux/threads.h> 52#include <linux/timer.h> 53#include <linux/rcupdate.h> 54#include <linux/cpu.h> 55#include <linux/cpuset.h> 56#include <linux/percpu.h> 57#include <linux/proc_fs.h> 58#include <linux/seq_file.h> 59#include <linux/sysctl.h> 60#include <linux/syscalls.h> 61#include <linux/times.h> 62#include <linux/tsacct_kern.h> 63#include <linux/kprobes.h> 64#include <linux/delayacct.h> 65#include <linux/unistd.h> 66#include <linux/pagemap.h> 67#include <linux/hrtimer.h> 68#include <linux/tick.h> 69#include <linux/debugfs.h> 70#include <linux/ctype.h> 71#include <linux/ftrace.h> 72#include <linux/slab.h> 73#include <linux/init_task.h> 74#include <linux/binfmts.h> 75#include <linux/context_tracking.h> 76#include <linux/compiler.h> 77 78#include <asm/switch_to.h> 79#include <asm/tlb.h> 80#include <asm/irq_regs.h> 81#include <asm/mutex.h> 82#ifdef CONFIG_PARAVIRT 83#include <asm/paravirt.h> 84#endif 85 86#include "sched.h" 87#include "../workqueue_internal.h" 88#include "../smpboot.h" 89 90#define CREATE_TRACE_POINTS 91#include <trace/events/sched.h> 92 93void start_bandwidth_timer(struct hrtimer *period_timer, ktime_t period) 94{ 95 unsigned long delta; 96 ktime_t soft, hard, now; 97 98 for (;;) { 99 if (hrtimer_active(period_timer)) 100 break; 101 102 now = hrtimer_cb_get_time(period_timer); 103 hrtimer_forward(period_timer, now, period); 104 105 soft = hrtimer_get_softexpires(period_timer); 106 hard = hrtimer_get_expires(period_timer); 107 delta = ktime_to_ns(ktime_sub(hard, soft)); 108 __hrtimer_start_range_ns(period_timer, soft, delta, 109 HRTIMER_MODE_ABS_PINNED, 0); 110 } 111} 112 113DEFINE_MUTEX(sched_domains_mutex); 114DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues); 115 116static void update_rq_clock_task(struct rq *rq, s64 delta); 117 118void update_rq_clock(struct rq *rq) 119{ 120 s64 delta; 121 122 if (rq->skip_clock_update > 0) 123 return; 124 125 delta = sched_clock_cpu(cpu_of(rq)) - rq->clock; 126 rq->clock += delta; 127 update_rq_clock_task(rq, delta); 128} 129 130/* 131 * Debugging: various feature bits 132 */ 133 134#define SCHED_FEAT(name, enabled) \ 135 (1UL << __SCHED_FEAT_##name) * enabled | 136 137const_debug unsigned int sysctl_sched_features = 138#include "features.h" 139 0; 140 141#undef SCHED_FEAT 142 143#ifdef CONFIG_SCHED_DEBUG 144#define SCHED_FEAT(name, enabled) \ 145 #name , 146 147static const char * const sched_feat_names[] = { 148#include "features.h" 149}; 150 151#undef SCHED_FEAT 152 153static int sched_feat_show(struct seq_file *m, void *v) 154{ 155 int i; 156 157 for (i = 0; i < __SCHED_FEAT_NR; i++) { 158 if (!(sysctl_sched_features & (1UL << i))) 159 seq_puts(m, "NO_"); 160 seq_printf(m, "%s ", sched_feat_names[i]); 161 } 162 seq_puts(m, "\n"); 163 164 return 0; 165} 166 167#ifdef HAVE_JUMP_LABEL 168 169#define jump_label_key__true STATIC_KEY_INIT_TRUE 170#define jump_label_key__false STATIC_KEY_INIT_FALSE 171 172#define SCHED_FEAT(name, enabled) \ 173 jump_label_key__##enabled , 174 175struct static_key sched_feat_keys[__SCHED_FEAT_NR] = { 176#include "features.h" 177}; 178 179#undef SCHED_FEAT 180 181static void sched_feat_disable(int i) 182{ 183 if (static_key_enabled(&sched_feat_keys[i])) 184 static_key_slow_dec(&sched_feat_keys[i]); 185} 186 187static void sched_feat_enable(int i) 188{ 189 if (!static_key_enabled(&sched_feat_keys[i])) 190 static_key_slow_inc(&sched_feat_keys[i]); 191} 192#else 193static void sched_feat_disable(int i) { }; 194static void sched_feat_enable(int i) { }; 195#endif /* HAVE_JUMP_LABEL */ 196 197static int sched_feat_set(char *cmp) 198{ 199 int i; 200 int neg = 0; 201 202 if (strncmp(cmp, "NO_", 3) == 0) { 203 neg = 1; 204 cmp += 3; 205 } 206 207 for (i = 0; i < __SCHED_FEAT_NR; i++) { 208 if (strcmp(cmp, sched_feat_names[i]) == 0) { 209 if (neg) { 210 sysctl_sched_features &= ~(1UL << i); 211 sched_feat_disable(i); 212 } else { 213 sysctl_sched_features |= (1UL << i); 214 sched_feat_enable(i); 215 } 216 break; 217 } 218 } 219 220 return i; 221} 222 223static ssize_t 224sched_feat_write(struct file *filp, const char __user *ubuf, 225 size_t cnt, loff_t *ppos) 226{ 227 char buf[64]; 228 char *cmp; 229 int i; 230 231 if (cnt > 63) 232 cnt = 63; 233 234 if (copy_from_user(&buf, ubuf, cnt)) 235 return -EFAULT; 236 237 buf[cnt] = 0; 238 cmp = strstrip(buf); 239 240 i = sched_feat_set(cmp); 241 if (i == __SCHED_FEAT_NR) 242 return -EINVAL; 243 244 *ppos += cnt; 245 246 return cnt; 247} 248 249static int sched_feat_open(struct inode *inode, struct file *filp) 250{ 251 return single_open(filp, sched_feat_show, NULL); 252} 253 254static const struct file_operations sched_feat_fops = { 255 .open = sched_feat_open, 256 .write = sched_feat_write, 257 .read = seq_read, 258 .llseek = seq_lseek, 259 .release = single_release, 260}; 261 262static __init int sched_init_debug(void) 263{ 264 debugfs_create_file("sched_features", 0644, NULL, NULL, 265 &sched_feat_fops); 266 267 return 0; 268} 269late_initcall(sched_init_debug); 270#endif /* CONFIG_SCHED_DEBUG */ 271 272/* 273 * Number of tasks to iterate in a single balance run. 274 * Limited because this is done with IRQs disabled. 275 */ 276const_debug unsigned int sysctl_sched_nr_migrate = 32; 277 278/* 279 * period over which we average the RT time consumption, measured 280 * in ms. 281 * 282 * default: 1s 283 */ 284const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC; 285 286/* 287 * period over which we measure -rt task cpu usage in us. 288 * default: 1s 289 */ 290unsigned int sysctl_sched_rt_period = 1000000; 291 292__read_mostly int scheduler_running; 293 294/* 295 * part of the period that we allow rt tasks to run in us. 296 * default: 0.95s 297 */ 298int sysctl_sched_rt_runtime = 950000; 299 300/* 301 * __task_rq_lock - lock the rq @p resides on. 302 */ 303static inline struct rq *__task_rq_lock(struct task_struct *p) 304 __acquires(rq->lock) 305{ 306 struct rq *rq; 307 308 lockdep_assert_held(&p->pi_lock); 309 310 for (;;) { 311 rq = task_rq(p); 312 raw_spin_lock(&rq->lock); 313 if (likely(rq == task_rq(p))) 314 return rq; 315 raw_spin_unlock(&rq->lock); 316 } 317} 318 319/* 320 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on. 321 */ 322static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags) 323 __acquires(p->pi_lock) 324 __acquires(rq->lock) 325{ 326 struct rq *rq; 327 328 for (;;) { 329 raw_spin_lock_irqsave(&p->pi_lock, *flags); 330 rq = task_rq(p); 331 raw_spin_lock(&rq->lock); 332 if (likely(rq == task_rq(p))) 333 return rq; 334 raw_spin_unlock(&rq->lock); 335 raw_spin_unlock_irqrestore(&p->pi_lock, *flags); 336 } 337} 338 339static void __task_rq_unlock(struct rq *rq) 340 __releases(rq->lock) 341{ 342 raw_spin_unlock(&rq->lock); 343} 344 345static inline void 346task_rq_unlock(struct rq *rq, struct task_struct *p, unsigned long *flags) 347 __releases(rq->lock) 348 __releases(p->pi_lock) 349{ 350 raw_spin_unlock(&rq->lock); 351 raw_spin_unlock_irqrestore(&p->pi_lock, *flags); 352} 353 354/* 355 * this_rq_lock - lock this runqueue and disable interrupts. 356 */ 357static struct rq *this_rq_lock(void) 358 __acquires(rq->lock) 359{ 360 struct rq *rq; 361 362 local_irq_disable(); 363 rq = this_rq(); 364 raw_spin_lock(&rq->lock); 365 366 return rq; 367} 368 369#ifdef CONFIG_SCHED_HRTICK 370/* 371 * Use HR-timers to deliver accurate preemption points. 372 */ 373 374static void hrtick_clear(struct rq *rq) 375{ 376 if (hrtimer_active(&rq->hrtick_timer)) 377 hrtimer_cancel(&rq->hrtick_timer); 378} 379 380/* 381 * High-resolution timer tick. 382 * Runs from hardirq context with interrupts disabled. 383 */ 384static enum hrtimer_restart hrtick(struct hrtimer *timer) 385{ 386 struct rq *rq = container_of(timer, struct rq, hrtick_timer); 387 388 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id()); 389 390 raw_spin_lock(&rq->lock); 391 update_rq_clock(rq); 392 rq->curr->sched_class->task_tick(rq, rq->curr, 1); 393 raw_spin_unlock(&rq->lock); 394 395 return HRTIMER_NORESTART; 396} 397 398#ifdef CONFIG_SMP 399 400static int __hrtick_restart(struct rq *rq) 401{ 402 struct hrtimer *timer = &rq->hrtick_timer; 403 ktime_t time = hrtimer_get_softexpires(timer); 404 405 return __hrtimer_start_range_ns(timer, time, 0, HRTIMER_MODE_ABS_PINNED, 0); 406} 407 408/* 409 * called from hardirq (IPI) context 410 */ 411static void __hrtick_start(void *arg) 412{ 413 struct rq *rq = arg; 414 415 raw_spin_lock(&rq->lock); 416 __hrtick_restart(rq); 417 rq->hrtick_csd_pending = 0; 418 raw_spin_unlock(&rq->lock); 419} 420 421/* 422 * Called to set the hrtick timer state. 423 * 424 * called with rq->lock held and irqs disabled 425 */ 426void hrtick_start(struct rq *rq, u64 delay) 427{ 428 struct hrtimer *timer = &rq->hrtick_timer; 429 ktime_t time = ktime_add_ns(timer->base->get_time(), delay); 430 431 hrtimer_set_expires(timer, time); 432 433 if (rq == this_rq()) { 434 __hrtick_restart(rq); 435 } else if (!rq->hrtick_csd_pending) { 436 smp_call_function_single_async(cpu_of(rq), &rq->hrtick_csd); 437 rq->hrtick_csd_pending = 1; 438 } 439} 440 441static int 442hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu) 443{ 444 int cpu = (int)(long)hcpu; 445 446 switch (action) { 447 case CPU_UP_CANCELED: 448 case CPU_UP_CANCELED_FROZEN: 449 case CPU_DOWN_PREPARE: 450 case CPU_DOWN_PREPARE_FROZEN: 451 case CPU_DEAD: 452 case CPU_DEAD_FROZEN: 453 hrtick_clear(cpu_rq(cpu)); 454 return NOTIFY_OK; 455 } 456 457 return NOTIFY_DONE; 458} 459 460static __init void init_hrtick(void) 461{ 462 hotcpu_notifier(hotplug_hrtick, 0); 463} 464#else 465/* 466 * Called to set the hrtick timer state. 467 * 468 * called with rq->lock held and irqs disabled 469 */ 470void hrtick_start(struct rq *rq, u64 delay) 471{ 472 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0, 473 HRTIMER_MODE_REL_PINNED, 0); 474} 475 476static inline void init_hrtick(void) 477{ 478} 479#endif /* CONFIG_SMP */ 480 481static void init_rq_hrtick(struct rq *rq) 482{ 483#ifdef CONFIG_SMP 484 rq->hrtick_csd_pending = 0; 485 486 rq->hrtick_csd.flags = 0; 487 rq->hrtick_csd.func = __hrtick_start; 488 rq->hrtick_csd.info = rq; 489#endif 490 491 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); 492 rq->hrtick_timer.function = hrtick; 493} 494#else /* CONFIG_SCHED_HRTICK */ 495static inline void hrtick_clear(struct rq *rq) 496{ 497} 498 499static inline void init_rq_hrtick(struct rq *rq) 500{ 501} 502 503static inline void init_hrtick(void) 504{ 505} 506#endif /* CONFIG_SCHED_HRTICK */ 507 508/* 509 * cmpxchg based fetch_or, macro so it works for different integer types 510 */ 511#define fetch_or(ptr, val) \ 512({ typeof(*(ptr)) __old, __val = *(ptr); \ 513 for (;;) { \ 514 __old = cmpxchg((ptr), __val, __val | (val)); \ 515 if (__old == __val) \ 516 break; \ 517 __val = __old; \ 518 } \ 519 __old; \ 520}) 521 522#ifdef TIF_POLLING_NRFLAG 523/* 524 * Atomically set TIF_NEED_RESCHED and test for TIF_POLLING_NRFLAG, 525 * this avoids any races wrt polling state changes and thereby avoids 526 * spurious IPIs. 527 */ 528static bool set_nr_and_not_polling(struct task_struct *p) 529{ 530 struct thread_info *ti = task_thread_info(p); 531 return !(fetch_or(&ti->flags, _TIF_NEED_RESCHED) & _TIF_POLLING_NRFLAG); 532} 533#else 534static bool set_nr_and_not_polling(struct task_struct *p) 535{ 536 set_tsk_need_resched(p); 537 return true; 538} 539#endif 540 541/* 542 * resched_task - mark a task 'to be rescheduled now'. 543 * 544 * On UP this means the setting of the need_resched flag, on SMP it 545 * might also involve a cross-CPU call to trigger the scheduler on 546 * the target CPU. 547 */ 548void resched_task(struct task_struct *p) 549{ 550 int cpu; 551 552 lockdep_assert_held(&task_rq(p)->lock); 553 554 if (test_tsk_need_resched(p)) 555 return; 556 557 cpu = task_cpu(p); 558 559 if (cpu == smp_processor_id()) { 560 set_tsk_need_resched(p); 561 set_preempt_need_resched(); 562 return; 563 } 564 565 if (set_nr_and_not_polling(p)) 566 smp_send_reschedule(cpu); 567} 568 569void resched_cpu(int cpu) 570{ 571 struct rq *rq = cpu_rq(cpu); 572 unsigned long flags; 573 574 if (!raw_spin_trylock_irqsave(&rq->lock, flags)) 575 return; 576 resched_task(cpu_curr(cpu)); 577 raw_spin_unlock_irqrestore(&rq->lock, flags); 578} 579 580#ifdef CONFIG_SMP 581#ifdef CONFIG_NO_HZ_COMMON 582/* 583 * In the semi idle case, use the nearest busy cpu for migrating timers 584 * from an idle cpu. This is good for power-savings. 585 * 586 * We don't do similar optimization for completely idle system, as 587 * selecting an idle cpu will add more delays to the timers than intended 588 * (as that cpu's timer base may not be uptodate wrt jiffies etc). 589 */ 590int get_nohz_timer_target(int pinned) 591{ 592 int cpu = smp_processor_id(); 593 int i; 594 struct sched_domain *sd; 595 596 if (pinned || !get_sysctl_timer_migration() || !idle_cpu(cpu)) 597 return cpu; 598 599 rcu_read_lock(); 600 for_each_domain(cpu, sd) { 601 for_each_cpu(i, sched_domain_span(sd)) { 602 if (!idle_cpu(i)) { 603 cpu = i; 604 goto unlock; 605 } 606 } 607 } 608unlock: 609 rcu_read_unlock(); 610 return cpu; 611} 612/* 613 * When add_timer_on() enqueues a timer into the timer wheel of an 614 * idle CPU then this timer might expire before the next timer event 615 * which is scheduled to wake up that CPU. In case of a completely 616 * idle system the next event might even be infinite time into the 617 * future. wake_up_idle_cpu() ensures that the CPU is woken up and 618 * leaves the inner idle loop so the newly added timer is taken into 619 * account when the CPU goes back to idle and evaluates the timer 620 * wheel for the next timer event. 621 */ 622static void wake_up_idle_cpu(int cpu) 623{ 624 struct rq *rq = cpu_rq(cpu); 625 626 if (cpu == smp_processor_id()) 627 return; 628 629 /* 630 * This is safe, as this function is called with the timer 631 * wheel base lock of (cpu) held. When the CPU is on the way 632 * to idle and has not yet set rq->curr to idle then it will 633 * be serialized on the timer wheel base lock and take the new 634 * timer into account automatically. 635 */ 636 if (rq->curr != rq->idle) 637 return; 638 639 /* 640 * We can set TIF_RESCHED on the idle task of the other CPU 641 * lockless. The worst case is that the other CPU runs the 642 * idle task through an additional NOOP schedule() 643 */ 644 set_tsk_need_resched(rq->idle); 645 646 /* NEED_RESCHED must be visible before we test polling */ 647 smp_mb(); 648 if (!tsk_is_polling(rq->idle)) 649 smp_send_reschedule(cpu); 650} 651 652static bool wake_up_full_nohz_cpu(int cpu) 653{ 654 if (tick_nohz_full_cpu(cpu)) { 655 if (cpu != smp_processor_id() || 656 tick_nohz_tick_stopped()) 657 smp_send_reschedule(cpu); 658 return true; 659 } 660 661 return false; 662} 663 664void wake_up_nohz_cpu(int cpu) 665{ 666 if (!wake_up_full_nohz_cpu(cpu)) 667 wake_up_idle_cpu(cpu); 668} 669 670static inline bool got_nohz_idle_kick(void) 671{ 672 int cpu = smp_processor_id(); 673 674 if (!test_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu))) 675 return false; 676 677 if (idle_cpu(cpu) && !need_resched()) 678 return true; 679 680 /* 681 * We can't run Idle Load Balance on this CPU for this time so we 682 * cancel it and clear NOHZ_BALANCE_KICK 683 */ 684 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu)); 685 return false; 686} 687 688#else /* CONFIG_NO_HZ_COMMON */ 689 690static inline bool got_nohz_idle_kick(void) 691{ 692 return false; 693} 694 695#endif /* CONFIG_NO_HZ_COMMON */ 696 697#ifdef CONFIG_NO_HZ_FULL 698bool sched_can_stop_tick(void) 699{ 700 struct rq *rq; 701 702 rq = this_rq(); 703 704 /* Make sure rq->nr_running update is visible after the IPI */ 705 smp_rmb(); 706 707 /* More than one running task need preemption */ 708 if (rq->nr_running > 1) 709 return false; 710 711 return true; 712} 713#endif /* CONFIG_NO_HZ_FULL */ 714 715void sched_avg_update(struct rq *rq) 716{ 717 s64 period = sched_avg_period(); 718 719 while ((s64)(rq_clock(rq) - rq->age_stamp) > period) { 720 /* 721 * Inline assembly required to prevent the compiler 722 * optimising this loop into a divmod call. 723 * See __iter_div_u64_rem() for another example of this. 724 */ 725 asm("" : "+rm" (rq->age_stamp)); 726 rq->age_stamp += period; 727 rq->rt_avg /= 2; 728 } 729} 730 731#endif /* CONFIG_SMP */ 732 733#if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \ 734 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH))) 735/* 736 * Iterate task_group tree rooted at *from, calling @down when first entering a 737 * node and @up when leaving it for the final time. 738 * 739 * Caller must hold rcu_lock or sufficient equivalent. 740 */ 741int walk_tg_tree_from(struct task_group *from, 742 tg_visitor down, tg_visitor up, void *data) 743{ 744 struct task_group *parent, *child; 745 int ret; 746 747 parent = from; 748 749down: 750 ret = (*down)(parent, data); 751 if (ret) 752 goto out; 753 list_for_each_entry_rcu(child, &parent->children, siblings) { 754 parent = child; 755 goto down; 756 757up: 758 continue; 759 } 760 ret = (*up)(parent, data); 761 if (ret || parent == from) 762 goto out; 763 764 child = parent; 765 parent = parent->parent; 766 if (parent) 767 goto up; 768out: 769 return ret; 770} 771 772int tg_nop(struct task_group *tg, void *data) 773{ 774 return 0; 775} 776#endif 777 778static void set_load_weight(struct task_struct *p) 779{ 780 int prio = p->static_prio - MAX_RT_PRIO; 781 struct load_weight *load = &p->se.load; 782 783 /* 784 * SCHED_IDLE tasks get minimal weight: 785 */ 786 if (p->policy == SCHED_IDLE) { 787 load->weight = scale_load(WEIGHT_IDLEPRIO); 788 load->inv_weight = WMULT_IDLEPRIO; 789 return; 790 } 791 792 load->weight = scale_load(prio_to_weight[prio]); 793 load->inv_weight = prio_to_wmult[prio]; 794} 795 796static void enqueue_task(struct rq *rq, struct task_struct *p, int flags) 797{ 798 update_rq_clock(rq); 799 sched_info_queued(rq, p); 800 p->sched_class->enqueue_task(rq, p, flags); 801} 802 803static void dequeue_task(struct rq *rq, struct task_struct *p, int flags) 804{ 805 update_rq_clock(rq); 806 sched_info_dequeued(rq, p); 807 p->sched_class->dequeue_task(rq, p, flags); 808} 809 810void activate_task(struct rq *rq, struct task_struct *p, int flags) 811{ 812 if (task_contributes_to_load(p)) 813 rq->nr_uninterruptible--; 814 815 enqueue_task(rq, p, flags); 816} 817 818void deactivate_task(struct rq *rq, struct task_struct *p, int flags) 819{ 820 if (task_contributes_to_load(p)) 821 rq->nr_uninterruptible++; 822 823 dequeue_task(rq, p, flags); 824} 825 826static void update_rq_clock_task(struct rq *rq, s64 delta) 827{ 828/* 829 * In theory, the compile should just see 0 here, and optimize out the call 830 * to sched_rt_avg_update. But I don't trust it... 831 */ 832#if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING) 833 s64 steal = 0, irq_delta = 0; 834#endif 835#ifdef CONFIG_IRQ_TIME_ACCOUNTING 836 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time; 837 838 /* 839 * Since irq_time is only updated on {soft,}irq_exit, we might run into 840 * this case when a previous update_rq_clock() happened inside a 841 * {soft,}irq region. 842 * 843 * When this happens, we stop ->clock_task and only update the 844 * prev_irq_time stamp to account for the part that fit, so that a next 845 * update will consume the rest. This ensures ->clock_task is 846 * monotonic. 847 * 848 * It does however cause some slight miss-attribution of {soft,}irq 849 * time, a more accurate solution would be to update the irq_time using 850 * the current rq->clock timestamp, except that would require using 851 * atomic ops. 852 */ 853 if (irq_delta > delta) 854 irq_delta = delta; 855 856 rq->prev_irq_time += irq_delta; 857 delta -= irq_delta; 858#endif 859#ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING 860 if (static_key_false((¶virt_steal_rq_enabled))) { 861 steal = paravirt_steal_clock(cpu_of(rq)); 862 steal -= rq->prev_steal_time_rq; 863 864 if (unlikely(steal > delta)) 865 steal = delta; 866 867 rq->prev_steal_time_rq += steal; 868 delta -= steal; 869 } 870#endif 871 872 rq->clock_task += delta; 873 874#if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING) 875 if ((irq_delta + steal) && sched_feat(NONTASK_POWER)) 876 sched_rt_avg_update(rq, irq_delta + steal); 877#endif 878} 879 880void sched_set_stop_task(int cpu, struct task_struct *stop) 881{ 882 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 }; 883 struct task_struct *old_stop = cpu_rq(cpu)->stop; 884 885 if (stop) { 886 /* 887 * Make it appear like a SCHED_FIFO task, its something 888 * userspace knows about and won't get confused about. 889 * 890 * Also, it will make PI more or less work without too 891 * much confusion -- but then, stop work should not 892 * rely on PI working anyway. 893 */ 894 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m); 895 896 stop->sched_class = &stop_sched_class; 897 } 898 899 cpu_rq(cpu)->stop = stop; 900 901 if (old_stop) { 902 /* 903 * Reset it back to a normal scheduling class so that 904 * it can die in pieces. 905 */ 906 old_stop->sched_class = &rt_sched_class; 907 } 908} 909 910/* 911 * __normal_prio - return the priority that is based on the static prio 912 */ 913static inline int __normal_prio(struct task_struct *p) 914{ 915 return p->static_prio; 916} 917 918/* 919 * Calculate the expected normal priority: i.e. priority 920 * without taking RT-inheritance into account. Might be 921 * boosted by interactivity modifiers. Changes upon fork, 922 * setprio syscalls, and whenever the interactivity 923 * estimator recalculates. 924 */ 925static inline int normal_prio(struct task_struct *p) 926{ 927 int prio; 928 929 if (task_has_dl_policy(p)) 930 prio = MAX_DL_PRIO-1; 931 else if (task_has_rt_policy(p)) 932 prio = MAX_RT_PRIO-1 - p->rt_priority; 933 else 934 prio = __normal_prio(p); 935 return prio; 936} 937 938/* 939 * Calculate the current priority, i.e. the priority 940 * taken into account by the scheduler. This value might 941 * be boosted by RT tasks, or might be boosted by 942 * interactivity modifiers. Will be RT if the task got 943 * RT-boosted. If not then it returns p->normal_prio. 944 */ 945static int effective_prio(struct task_struct *p) 946{ 947 p->normal_prio = normal_prio(p); 948 /* 949 * If we are RT tasks or we were boosted to RT priority, 950 * keep the priority unchanged. Otherwise, update priority 951 * to the normal priority: 952 */ 953 if (!rt_prio(p->prio)) 954 return p->normal_prio; 955 return p->prio; 956} 957 958/** 959 * task_curr - is this task currently executing on a CPU? 960 * @p: the task in question. 961 * 962 * Return: 1 if the task is currently executing. 0 otherwise. 963 */ 964inline int task_curr(const struct task_struct *p) 965{ 966 return cpu_curr(task_cpu(p)) == p; 967} 968 969static inline void check_class_changed(struct rq *rq, struct task_struct *p, 970 const struct sched_class *prev_class, 971 int oldprio) 972{ 973 if (prev_class != p->sched_class) { 974 if (prev_class->switched_from) 975 prev_class->switched_from(rq, p); 976 p->sched_class->switched_to(rq, p); 977 } else if (oldprio != p->prio || dl_task(p)) 978 p->sched_class->prio_changed(rq, p, oldprio); 979} 980 981void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags) 982{ 983 const struct sched_class *class; 984 985 if (p->sched_class == rq->curr->sched_class) { 986 rq->curr->sched_class->check_preempt_curr(rq, p, flags); 987 } else { 988 for_each_class(class) { 989 if (class == rq->curr->sched_class) 990 break; 991 if (class == p->sched_class) { 992 resched_task(rq->curr); 993 break; 994 } 995 } 996 } 997 998 /* 999 * A queue event has occurred, and we're going to schedule. In 1000 * this case, we can save a useless back to back clock update. 1001 */ 1002 if (rq->curr->on_rq && test_tsk_need_resched(rq->curr)) 1003 rq->skip_clock_update = 1; 1004} 1005 1006#ifdef CONFIG_SMP 1007void set_task_cpu(struct task_struct *p, unsigned int new_cpu) 1008{ 1009#ifdef CONFIG_SCHED_DEBUG 1010 /* 1011 * We should never call set_task_cpu() on a blocked task, 1012 * ttwu() will sort out the placement. 1013 */ 1014 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING && 1015 !(task_preempt_count(p) & PREEMPT_ACTIVE)); 1016 1017#ifdef CONFIG_LOCKDEP 1018 /* 1019 * The caller should hold either p->pi_lock or rq->lock, when changing 1020 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks. 1021 * 1022 * sched_move_task() holds both and thus holding either pins the cgroup, 1023 * see task_group(). 1024 * 1025 * Furthermore, all task_rq users should acquire both locks, see 1026 * task_rq_lock(). 1027 */ 1028 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) || 1029 lockdep_is_held(&task_rq(p)->lock))); 1030#endif 1031#endif 1032 1033 trace_sched_migrate_task(p, new_cpu); 1034 1035 if (task_cpu(p) != new_cpu) { 1036 if (p->sched_class->migrate_task_rq) 1037 p->sched_class->migrate_task_rq(p, new_cpu); 1038 p->se.nr_migrations++; 1039 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, NULL, 0); 1040 } 1041 1042 __set_task_cpu(p, new_cpu); 1043} 1044 1045static void __migrate_swap_task(struct task_struct *p, int cpu) 1046{ 1047 if (p->on_rq) { 1048 struct rq *src_rq, *dst_rq; 1049 1050 src_rq = task_rq(p); 1051 dst_rq = cpu_rq(cpu); 1052 1053 deactivate_task(src_rq, p, 0); 1054 set_task_cpu(p, cpu); 1055 activate_task(dst_rq, p, 0); 1056 check_preempt_curr(dst_rq, p, 0); 1057 } else { 1058 /* 1059 * Task isn't running anymore; make it appear like we migrated 1060 * it before it went to sleep. This means on wakeup we make the 1061 * previous cpu our targer instead of where it really is. 1062 */ 1063 p->wake_cpu = cpu; 1064 } 1065} 1066 1067struct migration_swap_arg { 1068 struct task_struct *src_task, *dst_task; 1069 int src_cpu, dst_cpu; 1070}; 1071 1072static int migrate_swap_stop(void *data) 1073{ 1074 struct migration_swap_arg *arg = data; 1075 struct rq *src_rq, *dst_rq; 1076 int ret = -EAGAIN; 1077 1078 src_rq = cpu_rq(arg->src_cpu); 1079 dst_rq = cpu_rq(arg->dst_cpu); 1080 1081 double_raw_lock(&arg->src_task->pi_lock, 1082 &arg->dst_task->pi_lock); 1083 double_rq_lock(src_rq, dst_rq); 1084 if (task_cpu(arg->dst_task) != arg->dst_cpu) 1085 goto unlock; 1086 1087 if (task_cpu(arg->src_task) != arg->src_cpu) 1088 goto unlock; 1089 1090 if (!cpumask_test_cpu(arg->dst_cpu, tsk_cpus_allowed(arg->src_task))) 1091 goto unlock; 1092 1093 if (!cpumask_test_cpu(arg->src_cpu, tsk_cpus_allowed(arg->dst_task))) 1094 goto unlock; 1095 1096 __migrate_swap_task(arg->src_task, arg->dst_cpu); 1097 __migrate_swap_task(arg->dst_task, arg->src_cpu); 1098 1099 ret = 0; 1100 1101unlock: 1102 double_rq_unlock(src_rq, dst_rq); 1103 raw_spin_unlock(&arg->dst_task->pi_lock); 1104 raw_spin_unlock(&arg->src_task->pi_lock); 1105 1106 return ret; 1107} 1108 1109/* 1110 * Cross migrate two tasks 1111 */ 1112int migrate_swap(struct task_struct *cur, struct task_struct *p) 1113{ 1114 struct migration_swap_arg arg; 1115 int ret = -EINVAL; 1116 1117 arg = (struct migration_swap_arg){ 1118 .src_task = cur, 1119 .src_cpu = task_cpu(cur), 1120 .dst_task = p, 1121 .dst_cpu = task_cpu(p), 1122 }; 1123 1124 if (arg.src_cpu == arg.dst_cpu) 1125 goto out; 1126 1127 /* 1128 * These three tests are all lockless; this is OK since all of them 1129 * will be re-checked with proper locks held further down the line. 1130 */ 1131 if (!cpu_active(arg.src_cpu) || !cpu_active(arg.dst_cpu)) 1132 goto out; 1133 1134 if (!cpumask_test_cpu(arg.dst_cpu, tsk_cpus_allowed(arg.src_task))) 1135 goto out; 1136 1137 if (!cpumask_test_cpu(arg.src_cpu, tsk_cpus_allowed(arg.dst_task))) 1138 goto out; 1139 1140 trace_sched_swap_numa(cur, arg.src_cpu, p, arg.dst_cpu); 1141 ret = stop_two_cpus(arg.dst_cpu, arg.src_cpu, migrate_swap_stop, &arg); 1142 1143out: 1144 return ret; 1145} 1146 1147struct migration_arg { 1148 struct task_struct *task; 1149 int dest_cpu; 1150}; 1151 1152static int migration_cpu_stop(void *data); 1153 1154/* 1155 * wait_task_inactive - wait for a thread to unschedule. 1156 * 1157 * If @match_state is nonzero, it's the @p->state value just checked and 1158 * not expected to change. If it changes, i.e. @p might have woken up, 1159 * then return zero. When we succeed in waiting for @p to be off its CPU, 1160 * we return a positive number (its total switch count). If a second call 1161 * a short while later returns the same number, the caller can be sure that 1162 * @p has remained unscheduled the whole time. 1163 * 1164 * The caller must ensure that the task *will* unschedule sometime soon, 1165 * else this function might spin for a *long* time. This function can't 1166 * be called with interrupts off, or it may introduce deadlock with 1167 * smp_call_function() if an IPI is sent by the same process we are 1168 * waiting to become inactive. 1169 */ 1170unsigned long wait_task_inactive(struct task_struct *p, long match_state) 1171{ 1172 unsigned long flags; 1173 int running, on_rq; 1174 unsigned long ncsw; 1175 struct rq *rq; 1176 1177 for (;;) { 1178 /* 1179 * We do the initial early heuristics without holding 1180 * any task-queue locks at all. We'll only try to get 1181 * the runqueue lock when things look like they will 1182 * work out! 1183 */ 1184 rq = task_rq(p); 1185 1186 /* 1187 * If the task is actively running on another CPU 1188 * still, just relax and busy-wait without holding 1189 * any locks. 1190 * 1191 * NOTE! Since we don't hold any locks, it's not 1192 * even sure that "rq" stays as the right runqueue! 1193 * But we don't care, since "task_running()" will 1194 * return false if the runqueue has changed and p 1195 * is actually now running somewhere else! 1196 */ 1197 while (task_running(rq, p)) { 1198 if (match_state && unlikely(p->state != match_state)) 1199 return 0; 1200 cpu_relax(); 1201 } 1202 1203 /* 1204 * Ok, time to look more closely! We need the rq 1205 * lock now, to be *sure*. If we're wrong, we'll 1206 * just go back and repeat. 1207 */ 1208 rq = task_rq_lock(p, &flags); 1209 trace_sched_wait_task(p); 1210 running = task_running(rq, p); 1211 on_rq = p->on_rq; 1212 ncsw = 0; 1213 if (!match_state || p->state == match_state) 1214 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */ 1215 task_rq_unlock(rq, p, &flags); 1216 1217 /* 1218 * If it changed from the expected state, bail out now. 1219 */ 1220 if (unlikely(!ncsw)) 1221 break; 1222 1223 /* 1224 * Was it really running after all now that we 1225 * checked with the proper locks actually held? 1226 * 1227 * Oops. Go back and try again.. 1228 */ 1229 if (unlikely(running)) { 1230 cpu_relax(); 1231 continue; 1232 } 1233 1234 /* 1235 * It's not enough that it's not actively running, 1236 * it must be off the runqueue _entirely_, and not 1237 * preempted! 1238 * 1239 * So if it was still runnable (but just not actively 1240 * running right now), it's preempted, and we should 1241 * yield - it could be a while. 1242 */ 1243 if (unlikely(on_rq)) { 1244 ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ); 1245 1246 set_current_state(TASK_UNINTERRUPTIBLE); 1247 schedule_hrtimeout(&to, HRTIMER_MODE_REL); 1248 continue; 1249 } 1250 1251 /* 1252 * Ahh, all good. It wasn't running, and it wasn't 1253 * runnable, which means that it will never become 1254 * running in the future either. We're all done! 1255 */ 1256 break; 1257 } 1258 1259 return ncsw; 1260} 1261 1262/*** 1263 * kick_process - kick a running thread to enter/exit the kernel 1264 * @p: the to-be-kicked thread 1265 * 1266 * Cause a process which is running on another CPU to enter 1267 * kernel-mode, without any delay. (to get signals handled.) 1268 * 1269 * NOTE: this function doesn't have to take the runqueue lock, 1270 * because all it wants to ensure is that the remote task enters 1271 * the kernel. If the IPI races and the task has been migrated 1272 * to another CPU then no harm is done and the purpose has been 1273 * achieved as well. 1274 */ 1275void kick_process(struct task_struct *p) 1276{ 1277 int cpu; 1278 1279 preempt_disable(); 1280 cpu = task_cpu(p); 1281 if ((cpu != smp_processor_id()) && task_curr(p)) 1282 smp_send_reschedule(cpu); 1283 preempt_enable(); 1284} 1285EXPORT_SYMBOL_GPL(kick_process); 1286#endif /* CONFIG_SMP */ 1287 1288#ifdef CONFIG_SMP 1289/* 1290 * ->cpus_allowed is protected by both rq->lock and p->pi_lock 1291 */ 1292static int select_fallback_rq(int cpu, struct task_struct *p) 1293{ 1294 int nid = cpu_to_node(cpu); 1295 const struct cpumask *nodemask = NULL; 1296 enum { cpuset, possible, fail } state = cpuset; 1297 int dest_cpu; 1298 1299 /* 1300 * If the node that the cpu is on has been offlined, cpu_to_node() 1301 * will return -1. There is no cpu on the node, and we should 1302 * select the cpu on the other node. 1303 */ 1304 if (nid != -1) { 1305 nodemask = cpumask_of_node(nid); 1306 1307 /* Look for allowed, online CPU in same node. */ 1308 for_each_cpu(dest_cpu, nodemask) { 1309 if (!cpu_online(dest_cpu)) 1310 continue; 1311 if (!cpu_active(dest_cpu)) 1312 continue; 1313 if (cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p))) 1314 return dest_cpu; 1315 } 1316 } 1317 1318 for (;;) { 1319 /* Any allowed, online CPU? */ 1320 for_each_cpu(dest_cpu, tsk_cpus_allowed(p)) { 1321 if (!cpu_online(dest_cpu)) 1322 continue; 1323 if (!cpu_active(dest_cpu)) 1324 continue; 1325 goto out; 1326 } 1327 1328 switch (state) { 1329 case cpuset: 1330 /* No more Mr. Nice Guy. */ 1331 cpuset_cpus_allowed_fallback(p); 1332 state = possible; 1333 break; 1334 1335 case possible: 1336 do_set_cpus_allowed(p, cpu_possible_mask); 1337 state = fail; 1338 break; 1339 1340 case fail: 1341 BUG(); 1342 break; 1343 } 1344 } 1345 1346out: 1347 if (state != cpuset) { 1348 /* 1349 * Don't tell them about moving exiting tasks or 1350 * kernel threads (both mm NULL), since they never 1351 * leave kernel. 1352 */ 1353 if (p->mm && printk_ratelimit()) { 1354 printk_sched("process %d (%s) no longer affine to cpu%d\n", 1355 task_pid_nr(p), p->comm, cpu); 1356 } 1357 } 1358 1359 return dest_cpu; 1360} 1361 1362/* 1363 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable. 1364 */ 1365static inline 1366int select_task_rq(struct task_struct *p, int cpu, int sd_flags, int wake_flags) 1367{ 1368 cpu = p->sched_class->select_task_rq(p, cpu, sd_flags, wake_flags); 1369 1370 /* 1371 * In order not to call set_task_cpu() on a blocking task we need 1372 * to rely on ttwu() to place the task on a valid ->cpus_allowed 1373 * cpu. 1374 * 1375 * Since this is common to all placement strategies, this lives here. 1376 * 1377 * [ this allows ->select_task() to simply return task_cpu(p) and 1378 * not worry about this generic constraint ] 1379 */ 1380 if (unlikely(!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)) || 1381 !cpu_online(cpu))) 1382 cpu = select_fallback_rq(task_cpu(p), p); 1383 1384 return cpu; 1385} 1386 1387static void update_avg(u64 *avg, u64 sample) 1388{ 1389 s64 diff = sample - *avg; 1390 *avg += diff >> 3; 1391} 1392#endif 1393 1394static void 1395ttwu_stat(struct task_struct *p, int cpu, int wake_flags) 1396{ 1397#ifdef CONFIG_SCHEDSTATS 1398 struct rq *rq = this_rq(); 1399 1400#ifdef CONFIG_SMP 1401 int this_cpu = smp_processor_id(); 1402 1403 if (cpu == this_cpu) { 1404 schedstat_inc(rq, ttwu_local); 1405 schedstat_inc(p, se.statistics.nr_wakeups_local); 1406 } else { 1407 struct sched_domain *sd; 1408 1409 schedstat_inc(p, se.statistics.nr_wakeups_remote); 1410 rcu_read_lock(); 1411 for_each_domain(this_cpu, sd) { 1412 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) { 1413 schedstat_inc(sd, ttwu_wake_remote); 1414 break; 1415 } 1416 } 1417 rcu_read_unlock(); 1418 } 1419 1420 if (wake_flags & WF_MIGRATED) 1421 schedstat_inc(p, se.statistics.nr_wakeups_migrate); 1422 1423#endif /* CONFIG_SMP */ 1424 1425 schedstat_inc(rq, ttwu_count); 1426 schedstat_inc(p, se.statistics.nr_wakeups); 1427 1428 if (wake_flags & WF_SYNC) 1429 schedstat_inc(p, se.statistics.nr_wakeups_sync); 1430 1431#endif /* CONFIG_SCHEDSTATS */ 1432} 1433 1434static void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags) 1435{ 1436 activate_task(rq, p, en_flags); 1437 p->on_rq = 1; 1438 1439 /* if a worker is waking up, notify workqueue */ 1440 if (p->flags & PF_WQ_WORKER) 1441 wq_worker_waking_up(p, cpu_of(rq)); 1442} 1443 1444/* 1445 * Mark the task runnable and perform wakeup-preemption. 1446 */ 1447static void 1448ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags) 1449{ 1450 check_preempt_curr(rq, p, wake_flags); 1451 trace_sched_wakeup(p, true); 1452 1453 p->state = TASK_RUNNING; 1454#ifdef CONFIG_SMP 1455 if (p->sched_class->task_woken) 1456 p->sched_class->task_woken(rq, p); 1457 1458 if (rq->idle_stamp) { 1459 u64 delta = rq_clock(rq) - rq->idle_stamp; 1460 u64 max = 2*rq->max_idle_balance_cost; 1461 1462 update_avg(&rq->avg_idle, delta); 1463 1464 if (rq->avg_idle > max) 1465 rq->avg_idle = max; 1466 1467 rq->idle_stamp = 0; 1468 } 1469#endif 1470} 1471 1472static void 1473ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags) 1474{ 1475#ifdef CONFIG_SMP 1476 if (p->sched_contributes_to_load) 1477 rq->nr_uninterruptible--; 1478#endif 1479 1480 ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_WAKING); 1481 ttwu_do_wakeup(rq, p, wake_flags); 1482} 1483 1484/* 1485 * Called in case the task @p isn't fully descheduled from its runqueue, 1486 * in this case we must do a remote wakeup. Its a 'light' wakeup though, 1487 * since all we need to do is flip p->state to TASK_RUNNING, since 1488 * the task is still ->on_rq. 1489 */ 1490static int ttwu_remote(struct task_struct *p, int wake_flags) 1491{ 1492 struct rq *rq; 1493 int ret = 0; 1494 1495 rq = __task_rq_lock(p); 1496 if (p->on_rq) { 1497 /* check_preempt_curr() may use rq clock */ 1498 update_rq_clock(rq); 1499 ttwu_do_wakeup(rq, p, wake_flags); 1500 ret = 1; 1501 } 1502 __task_rq_unlock(rq); 1503 1504 return ret; 1505} 1506 1507#ifdef CONFIG_SMP 1508static void sched_ttwu_pending(void) 1509{ 1510 struct rq *rq = this_rq(); 1511 struct llist_node *llist = llist_del_all(&rq->wake_list); 1512 struct task_struct *p; 1513 1514 raw_spin_lock(&rq->lock); 1515 1516 while (llist) { 1517 p = llist_entry(llist, struct task_struct, wake_entry); 1518 llist = llist_next(llist); 1519 ttwu_do_activate(rq, p, 0); 1520 } 1521 1522 raw_spin_unlock(&rq->lock); 1523} 1524 1525void scheduler_ipi(void) 1526{ 1527 /* 1528 * Fold TIF_NEED_RESCHED into the preempt_count; anybody setting 1529 * TIF_NEED_RESCHED remotely (for the first time) will also send 1530 * this IPI. 1531 */ 1532 preempt_fold_need_resched(); 1533 1534 if (llist_empty(&this_rq()->wake_list) 1535 && !tick_nohz_full_cpu(smp_processor_id()) 1536 && !got_nohz_idle_kick()) 1537 return; 1538 1539 /* 1540 * Not all reschedule IPI handlers call irq_enter/irq_exit, since 1541 * traditionally all their work was done from the interrupt return 1542 * path. Now that we actually do some work, we need to make sure 1543 * we do call them. 1544 * 1545 * Some archs already do call them, luckily irq_enter/exit nest 1546 * properly. 1547 * 1548 * Arguably we should visit all archs and update all handlers, 1549 * however a fair share of IPIs are still resched only so this would 1550 * somewhat pessimize the simple resched case. 1551 */ 1552 irq_enter(); 1553 tick_nohz_full_check(); 1554 sched_ttwu_pending(); 1555 1556 /* 1557 * Check if someone kicked us for doing the nohz idle load balance. 1558 */ 1559 if (unlikely(got_nohz_idle_kick())) { 1560 this_rq()->idle_balance = 1; 1561 raise_softirq_irqoff(SCHED_SOFTIRQ); 1562 } 1563 irq_exit(); 1564} 1565 1566static void ttwu_queue_remote(struct task_struct *p, int cpu) 1567{ 1568 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list)) 1569 smp_send_reschedule(cpu); 1570} 1571 1572bool cpus_share_cache(int this_cpu, int that_cpu) 1573{ 1574 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu); 1575} 1576#endif /* CONFIG_SMP */ 1577 1578static void ttwu_queue(struct task_struct *p, int cpu) 1579{ 1580 struct rq *rq = cpu_rq(cpu); 1581 1582#if defined(CONFIG_SMP) 1583 if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) { 1584 sched_clock_cpu(cpu); /* sync clocks x-cpu */ 1585 ttwu_queue_remote(p, cpu); 1586 return; 1587 } 1588#endif 1589 1590 raw_spin_lock(&rq->lock); 1591 ttwu_do_activate(rq, p, 0); 1592 raw_spin_unlock(&rq->lock); 1593} 1594 1595/** 1596 * try_to_wake_up - wake up a thread 1597 * @p: the thread to be awakened 1598 * @state: the mask of task states that can be woken 1599 * @wake_flags: wake modifier flags (WF_*) 1600 * 1601 * Put it on the run-queue if it's not already there. The "current" 1602 * thread is always on the run-queue (except when the actual 1603 * re-schedule is in progress), and as such you're allowed to do 1604 * the simpler "current->state = TASK_RUNNING" to mark yourself 1605 * runnable without the overhead of this. 1606 * 1607 * Return: %true if @p was woken up, %false if it was already running. 1608 * or @state didn't match @p's state. 1609 */ 1610static int 1611try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags) 1612{ 1613 unsigned long flags; 1614 int cpu, success = 0; 1615 1616 /* 1617 * If we are going to wake up a thread waiting for CONDITION we 1618 * need to ensure that CONDITION=1 done by the caller can not be 1619 * reordered with p->state check below. This pairs with mb() in 1620 * set_current_state() the waiting thread does. 1621 */ 1622 smp_mb__before_spinlock(); 1623 raw_spin_lock_irqsave(&p->pi_lock, flags); 1624 if (!(p->state & state)) 1625 goto out; 1626 1627 success = 1; /* we're going to change ->state */ 1628 cpu = task_cpu(p); 1629 1630 if (p->on_rq && ttwu_remote(p, wake_flags)) 1631 goto stat; 1632 1633#ifdef CONFIG_SMP 1634 /* 1635 * If the owning (remote) cpu is still in the middle of schedule() with 1636 * this task as prev, wait until its done referencing the task. 1637 */ 1638 while (p->on_cpu) 1639 cpu_relax(); 1640 /* 1641 * Pairs with the smp_wmb() in finish_lock_switch(). 1642 */ 1643 smp_rmb(); 1644 1645 p->sched_contributes_to_load = !!task_contributes_to_load(p); 1646 p->state = TASK_WAKING; 1647 1648 if (p->sched_class->task_waking) 1649 p->sched_class->task_waking(p); 1650 1651 cpu = select_task_rq(p, p->wake_cpu, SD_BALANCE_WAKE, wake_flags); 1652 if (task_cpu(p) != cpu) { 1653 wake_flags |= WF_MIGRATED; 1654 set_task_cpu(p, cpu); 1655 } 1656#endif /* CONFIG_SMP */ 1657 1658 ttwu_queue(p, cpu); 1659stat: 1660 ttwu_stat(p, cpu, wake_flags); 1661out: 1662 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 1663 1664 return success; 1665} 1666 1667/** 1668 * try_to_wake_up_local - try to wake up a local task with rq lock held 1669 * @p: the thread to be awakened 1670 * 1671 * Put @p on the run-queue if it's not already there. The caller must 1672 * ensure that this_rq() is locked, @p is bound to this_rq() and not 1673 * the current task. 1674 */ 1675static void try_to_wake_up_local(struct task_struct *p) 1676{ 1677 struct rq *rq = task_rq(p); 1678 1679 if (WARN_ON_ONCE(rq != this_rq()) || 1680 WARN_ON_ONCE(p == current)) 1681 return; 1682 1683 lockdep_assert_held(&rq->lock); 1684 1685 if (!raw_spin_trylock(&p->pi_lock)) { 1686 raw_spin_unlock(&rq->lock); 1687 raw_spin_lock(&p->pi_lock); 1688 raw_spin_lock(&rq->lock); 1689 } 1690 1691 if (!(p->state & TASK_NORMAL)) 1692 goto out; 1693 1694 if (!p->on_rq) 1695 ttwu_activate(rq, p, ENQUEUE_WAKEUP); 1696 1697 ttwu_do_wakeup(rq, p, 0); 1698 ttwu_stat(p, smp_processor_id(), 0); 1699out: 1700 raw_spin_unlock(&p->pi_lock); 1701} 1702 1703/** 1704 * wake_up_process - Wake up a specific process 1705 * @p: The process to be woken up. 1706 * 1707 * Attempt to wake up the nominated process and move it to the set of runnable 1708 * processes. 1709 * 1710 * Return: 1 if the process was woken up, 0 if it was already running. 1711 * 1712 * It may be assumed that this function implies a write memory barrier before 1713 * changing the task state if and only if any tasks are woken up. 1714 */ 1715int wake_up_process(struct task_struct *p) 1716{ 1717 WARN_ON(task_is_stopped_or_traced(p)); 1718 return try_to_wake_up(p, TASK_NORMAL, 0); 1719} 1720EXPORT_SYMBOL(wake_up_process); 1721 1722int wake_up_state(struct task_struct *p, unsigned int state) 1723{ 1724 return try_to_wake_up(p, state, 0); 1725} 1726 1727/* 1728 * Perform scheduler related setup for a newly forked process p. 1729 * p is forked by current. 1730 * 1731 * __sched_fork() is basic setup used by init_idle() too: 1732 */ 1733static void __sched_fork(unsigned long clone_flags, struct task_struct *p) 1734{ 1735 p->on_rq = 0; 1736 1737 p->se.on_rq = 0; 1738 p->se.exec_start = 0; 1739 p->se.sum_exec_runtime = 0; 1740 p->se.prev_sum_exec_runtime = 0; 1741 p->se.nr_migrations = 0; 1742 p->se.vruntime = 0; 1743 INIT_LIST_HEAD(&p->se.group_node); 1744 1745#ifdef CONFIG_SCHEDSTATS 1746 memset(&p->se.statistics, 0, sizeof(p->se.statistics)); 1747#endif 1748 1749 RB_CLEAR_NODE(&p->dl.rb_node); 1750 hrtimer_init(&p->dl.dl_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); 1751 p->dl.dl_runtime = p->dl.runtime = 0; 1752 p->dl.dl_deadline = p->dl.deadline = 0; 1753 p->dl.dl_period = 0; 1754 p->dl.flags = 0; 1755 1756 INIT_LIST_HEAD(&p->rt.run_list); 1757 1758#ifdef CONFIG_PREEMPT_NOTIFIERS 1759 INIT_HLIST_HEAD(&p->preempt_notifiers); 1760#endif 1761 1762#ifdef CONFIG_NUMA_BALANCING 1763 if (p->mm && atomic_read(&p->mm->mm_users) == 1) { 1764 p->mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay); 1765 p->mm->numa_scan_seq = 0; 1766 } 1767 1768 if (clone_flags & CLONE_VM) 1769 p->numa_preferred_nid = current->numa_preferred_nid; 1770 else 1771 p->numa_preferred_nid = -1; 1772 1773 p->node_stamp = 0ULL; 1774 p->numa_scan_seq = p->mm ? p->mm->numa_scan_seq : 0; 1775 p->numa_scan_period = sysctl_numa_balancing_scan_delay; 1776 p->numa_work.next = &p->numa_work; 1777 p->numa_faults_memory = NULL; 1778 p->numa_faults_buffer_memory = NULL; 1779 p->last_task_numa_placement = 0; 1780 p->last_sum_exec_runtime = 0; 1781 1782 INIT_LIST_HEAD(&p->numa_entry); 1783 p->numa_group = NULL; 1784#endif /* CONFIG_NUMA_BALANCING */ 1785} 1786 1787#ifdef CONFIG_NUMA_BALANCING 1788#ifdef CONFIG_SCHED_DEBUG 1789void set_numabalancing_state(bool enabled) 1790{ 1791 if (enabled) 1792 sched_feat_set("NUMA"); 1793 else 1794 sched_feat_set("NO_NUMA"); 1795} 1796#else 1797__read_mostly bool numabalancing_enabled; 1798 1799void set_numabalancing_state(bool enabled) 1800{ 1801 numabalancing_enabled = enabled; 1802} 1803#endif /* CONFIG_SCHED_DEBUG */ 1804 1805#ifdef CONFIG_PROC_SYSCTL 1806int sysctl_numa_balancing(struct ctl_table *table, int write, 1807 void __user *buffer, size_t *lenp, loff_t *ppos) 1808{ 1809 struct ctl_table t; 1810 int err; 1811 int state = numabalancing_enabled; 1812 1813 if (write && !capable(CAP_SYS_ADMIN)) 1814 return -EPERM; 1815 1816 t = *table; 1817 t.data = &state; 1818 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos); 1819 if (err < 0) 1820 return err; 1821 if (write) 1822 set_numabalancing_state(state); 1823 return err; 1824} 1825#endif 1826#endif 1827 1828/* 1829 * fork()/clone()-time setup: 1830 */ 1831int sched_fork(unsigned long clone_flags, struct task_struct *p) 1832{ 1833 unsigned long flags; 1834 int cpu = get_cpu(); 1835 1836 __sched_fork(clone_flags, p); 1837 /* 1838 * We mark the process as running here. This guarantees that 1839 * nobody will actually run it, and a signal or other external 1840 * event cannot wake it up and insert it on the runqueue either. 1841 */ 1842 p->state = TASK_RUNNING; 1843 1844 /* 1845 * Make sure we do not leak PI boosting priority to the child. 1846 */ 1847 p->prio = current->normal_prio; 1848 1849 /* 1850 * Revert to default priority/policy on fork if requested. 1851 */ 1852 if (unlikely(p->sched_reset_on_fork)) { 1853 if (task_has_dl_policy(p) || task_has_rt_policy(p)) { 1854 p->policy = SCHED_NORMAL; 1855 p->static_prio = NICE_TO_PRIO(0); 1856 p->rt_priority = 0; 1857 } else if (PRIO_TO_NICE(p->static_prio) < 0) 1858 p->static_prio = NICE_TO_PRIO(0); 1859 1860 p->prio = p->normal_prio = __normal_prio(p); 1861 set_load_weight(p); 1862 1863 /* 1864 * We don't need the reset flag anymore after the fork. It has 1865 * fulfilled its duty: 1866 */ 1867 p->sched_reset_on_fork = 0; 1868 } 1869 1870 if (dl_prio(p->prio)) { 1871 put_cpu(); 1872 return -EAGAIN; 1873 } else if (rt_prio(p->prio)) { 1874 p->sched_class = &rt_sched_class; 1875 } else { 1876 p->sched_class = &fair_sched_class; 1877 } 1878 1879 if (p->sched_class->task_fork) 1880 p->sched_class->task_fork(p); 1881 1882 /* 1883 * The child is not yet in the pid-hash so no cgroup attach races, 1884 * and the cgroup is pinned to this child due to cgroup_fork() 1885 * is ran before sched_fork(). 1886 * 1887 * Silence PROVE_RCU. 1888 */ 1889 raw_spin_lock_irqsave(&p->pi_lock, flags); 1890 set_task_cpu(p, cpu); 1891 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 1892 1893#if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT) 1894 if (likely(sched_info_on())) 1895 memset(&p->sched_info, 0, sizeof(p->sched_info)); 1896#endif 1897#if defined(CONFIG_SMP) 1898 p->on_cpu = 0; 1899#endif 1900 init_task_preempt_count(p); 1901#ifdef CONFIG_SMP 1902 plist_node_init(&p->pushable_tasks, MAX_PRIO); 1903 RB_CLEAR_NODE(&p->pushable_dl_tasks); 1904#endif 1905 1906 put_cpu(); 1907 return 0; 1908} 1909 1910unsigned long to_ratio(u64 period, u64 runtime) 1911{ 1912 if (runtime == RUNTIME_INF) 1913 return 1ULL << 20; 1914 1915 /* 1916 * Doing this here saves a lot of checks in all 1917 * the calling paths, and returning zero seems 1918 * safe for them anyway. 1919 */ 1920 if (period == 0) 1921 return 0; 1922 1923 return div64_u64(runtime << 20, period); 1924} 1925 1926#ifdef CONFIG_SMP 1927inline struct dl_bw *dl_bw_of(int i) 1928{ 1929 return &cpu_rq(i)->rd->dl_bw; 1930} 1931 1932static inline int dl_bw_cpus(int i) 1933{ 1934 struct root_domain *rd = cpu_rq(i)->rd; 1935 int cpus = 0; 1936 1937 for_each_cpu_and(i, rd->span, cpu_active_mask) 1938 cpus++; 1939 1940 return cpus; 1941} 1942#else 1943inline struct dl_bw *dl_bw_of(int i) 1944{ 1945 return &cpu_rq(i)->dl.dl_bw; 1946} 1947 1948static inline int dl_bw_cpus(int i) 1949{ 1950 return 1; 1951} 1952#endif 1953 1954static inline 1955void __dl_clear(struct dl_bw *dl_b, u64 tsk_bw) 1956{ 1957 dl_b->total_bw -= tsk_bw; 1958} 1959 1960static inline 1961void __dl_add(struct dl_bw *dl_b, u64 tsk_bw) 1962{ 1963 dl_b->total_bw += tsk_bw; 1964} 1965 1966static inline 1967bool __dl_overflow(struct dl_bw *dl_b, int cpus, u64 old_bw, u64 new_bw) 1968{ 1969 return dl_b->bw != -1 && 1970 dl_b->bw * cpus < dl_b->total_bw - old_bw + new_bw; 1971} 1972 1973/* 1974 * We must be sure that accepting a new task (or allowing changing the 1975 * parameters of an existing one) is consistent with the bandwidth 1976 * constraints. If yes, this function also accordingly updates the currently 1977 * allocated bandwidth to reflect the new situation. 1978 * 1979 * This function is called while holding p's rq->lock. 1980 */ 1981static int dl_overflow(struct task_struct *p, int policy, 1982 const struct sched_attr *attr) 1983{ 1984 1985 struct dl_bw *dl_b = dl_bw_of(task_cpu(p)); 1986 u64 period = attr->sched_period ?: attr->sched_deadline; 1987 u64 runtime = attr->sched_runtime; 1988 u64 new_bw = dl_policy(policy) ? to_ratio(period, runtime) : 0; 1989 int cpus, err = -1; 1990 1991 if (new_bw == p->dl.dl_bw) 1992 return 0; 1993 1994 /* 1995 * Either if a task, enters, leave, or stays -deadline but changes 1996 * its parameters, we may need to update accordingly the total 1997 * allocated bandwidth of the container. 1998 */ 1999 raw_spin_lock(&dl_b->lock); 2000 cpus = dl_bw_cpus(task_cpu(p)); 2001 if (dl_policy(policy) && !task_has_dl_policy(p) && 2002 !__dl_overflow(dl_b, cpus, 0, new_bw)) { 2003 __dl_add(dl_b, new_bw); 2004 err = 0; 2005 } else if (dl_policy(policy) && task_has_dl_policy(p) && 2006 !__dl_overflow(dl_b, cpus, p->dl.dl_bw, new_bw)) { 2007 __dl_clear(dl_b, p->dl.dl_bw); 2008 __dl_add(dl_b, new_bw); 2009 err = 0; 2010 } else if (!dl_policy(policy) && task_has_dl_policy(p)) { 2011 __dl_clear(dl_b, p->dl.dl_bw); 2012 err = 0; 2013 } 2014 raw_spin_unlock(&dl_b->lock); 2015 2016 return err; 2017} 2018 2019extern void init_dl_bw(struct dl_bw *dl_b); 2020 2021/* 2022 * wake_up_new_task - wake up a newly created task for the first time. 2023 * 2024 * This function will do some initial scheduler statistics housekeeping 2025 * that must be done for every newly created context, then puts the task 2026 * on the runqueue and wakes it. 2027 */ 2028void wake_up_new_task(struct task_struct *p) 2029{ 2030 unsigned long flags; 2031 struct rq *rq; 2032 2033 raw_spin_lock_irqsave(&p->pi_lock, flags); 2034#ifdef CONFIG_SMP 2035 /* 2036 * Fork balancing, do it here and not earlier because: 2037 * - cpus_allowed can change in the fork path 2038 * - any previously selected cpu might disappear through hotplug 2039 */ 2040 set_task_cpu(p, select_task_rq(p, task_cpu(p), SD_BALANCE_FORK, 0)); 2041#endif 2042 2043 /* Initialize new task's runnable average */ 2044 init_task_runnable_average(p); 2045 rq = __task_rq_lock(p); 2046 activate_task(rq, p, 0); 2047 p->on_rq = 1; 2048 trace_sched_wakeup_new(p, true); 2049 check_preempt_curr(rq, p, WF_FORK); 2050#ifdef CONFIG_SMP 2051 if (p->sched_class->task_woken) 2052 p->sched_class->task_woken(rq, p); 2053#endif 2054 task_rq_unlock(rq, p, &flags); 2055} 2056 2057#ifdef CONFIG_PREEMPT_NOTIFIERS 2058 2059/** 2060 * preempt_notifier_register - tell me when current is being preempted & rescheduled 2061 * @notifier: notifier struct to register 2062 */ 2063void preempt_notifier_register(struct preempt_notifier *notifier) 2064{ 2065 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers); 2066} 2067EXPORT_SYMBOL_GPL(preempt_notifier_register); 2068 2069/** 2070 * preempt_notifier_unregister - no longer interested in preemption notifications 2071 * @notifier: notifier struct to unregister 2072 * 2073 * This is safe to call from within a preemption notifier. 2074 */ 2075void preempt_notifier_unregister(struct preempt_notifier *notifier) 2076{ 2077 hlist_del(¬ifier->link); 2078} 2079EXPORT_SYMBOL_GPL(preempt_notifier_unregister); 2080 2081static void fire_sched_in_preempt_notifiers(struct task_struct *curr) 2082{ 2083 struct preempt_notifier *notifier; 2084 2085 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link) 2086 notifier->ops->sched_in(notifier, raw_smp_processor_id()); 2087} 2088 2089static void 2090fire_sched_out_preempt_notifiers(struct task_struct *curr, 2091 struct task_struct *next) 2092{ 2093 struct preempt_notifier *notifier; 2094 2095 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link) 2096 notifier->ops->sched_out(notifier, next); 2097} 2098 2099#else /* !CONFIG_PREEMPT_NOTIFIERS */ 2100 2101static void fire_sched_in_preempt_notifiers(struct task_struct *curr) 2102{ 2103} 2104 2105static void 2106fire_sched_out_preempt_notifiers(struct task_struct *curr, 2107 struct task_struct *next) 2108{ 2109} 2110 2111#endif /* CONFIG_PREEMPT_NOTIFIERS */ 2112 2113/** 2114 * prepare_task_switch - prepare to switch tasks 2115 * @rq: the runqueue preparing to switch 2116 * @prev: the current task that is being switched out 2117 * @next: the task we are going to switch to. 2118 * 2119 * This is called with the rq lock held and interrupts off. It must 2120 * be paired with a subsequent finish_task_switch after the context 2121 * switch. 2122 * 2123 * prepare_task_switch sets up locking and calls architecture specific 2124 * hooks. 2125 */ 2126static inline void 2127prepare_task_switch(struct rq *rq, struct task_struct *prev, 2128 struct task_struct *next) 2129{ 2130 trace_sched_switch(prev, next); 2131 sched_info_switch(rq, prev, next); 2132 perf_event_task_sched_out(prev, next); 2133 fire_sched_out_preempt_notifiers(prev, next); 2134 prepare_lock_switch(rq, next); 2135 prepare_arch_switch(next); 2136} 2137 2138/** 2139 * finish_task_switch - clean up after a task-switch 2140 * @rq: runqueue associated with task-switch 2141 * @prev: the thread we just switched away from. 2142 * 2143 * finish_task_switch must be called after the context switch, paired 2144 * with a prepare_task_switch call before the context switch. 2145 * finish_task_switch will reconcile locking set up by prepare_task_switch, 2146 * and do any other architecture-specific cleanup actions. 2147 * 2148 * Note that we may have delayed dropping an mm in context_switch(). If 2149 * so, we finish that here outside of the runqueue lock. (Doing it 2150 * with the lock held can cause deadlocks; see schedule() for 2151 * details.) 2152 */ 2153static void finish_task_switch(struct rq *rq, struct task_struct *prev) 2154 __releases(rq->lock) 2155{ 2156 struct mm_struct *mm = rq->prev_mm; 2157 long prev_state; 2158 2159 rq->prev_mm = NULL; 2160 2161 /* 2162 * A task struct has one reference for the use as "current". 2163 * If a task dies, then it sets TASK_DEAD in tsk->state and calls 2164 * schedule one last time. The schedule call will never return, and 2165 * the scheduled task must drop that reference. 2166 * The test for TASK_DEAD must occur while the runqueue locks are 2167 * still held, otherwise prev could be scheduled on another cpu, die 2168 * there before we look at prev->state, and then the reference would 2169 * be dropped twice. 2170 * Manfred Spraul <manfred@colorfullife.com> 2171 */ 2172 prev_state = prev->state; 2173 vtime_task_switch(prev); 2174 finish_arch_switch(prev); 2175 perf_event_task_sched_in(prev, current); 2176 finish_lock_switch(rq, prev); 2177 finish_arch_post_lock_switch(); 2178 2179 fire_sched_in_preempt_notifiers(current); 2180 if (mm) 2181 mmdrop(mm); 2182 if (unlikely(prev_state == TASK_DEAD)) { 2183 if (prev->sched_class->task_dead) 2184 prev->sched_class->task_dead(prev); 2185 2186 /* 2187 * Remove function-return probe instances associated with this 2188 * task and put them back on the free list. 2189 */ 2190 kprobe_flush_task(prev); 2191 put_task_struct(prev); 2192 } 2193 2194 tick_nohz_task_switch(current); 2195} 2196 2197#ifdef CONFIG_SMP 2198 2199/* rq->lock is NOT held, but preemption is disabled */ 2200static inline void post_schedule(struct rq *rq) 2201{ 2202 if (rq->post_schedule) { 2203 unsigned long flags; 2204 2205 raw_spin_lock_irqsave(&rq->lock, flags); 2206 if (rq->curr->sched_class->post_schedule) 2207 rq->curr->sched_class->post_schedule(rq); 2208 raw_spin_unlock_irqrestore(&rq->lock, flags); 2209 2210 rq->post_schedule = 0; 2211 } 2212} 2213 2214#else 2215 2216static inline void post_schedule(struct rq *rq) 2217{ 2218} 2219 2220#endif 2221 2222/** 2223 * schedule_tail - first thing a freshly forked thread must call. 2224 * @prev: the thread we just switched away from. 2225 */ 2226asmlinkage void schedule_tail(struct task_struct *prev) 2227 __releases(rq->lock) 2228{ 2229 struct rq *rq = this_rq(); 2230 2231 finish_task_switch(rq, prev); 2232 2233 /* 2234 * FIXME: do we need to worry about rq being invalidated by the 2235 * task_switch? 2236 */ 2237 post_schedule(rq); 2238 2239#ifdef __ARCH_WANT_UNLOCKED_CTXSW 2240 /* In this case, finish_task_switch does not reenable preemption */ 2241 preempt_enable(); 2242#endif 2243 if (current->set_child_tid) 2244 put_user(task_pid_vnr(current), current->set_child_tid); 2245} 2246 2247/* 2248 * context_switch - switch to the new MM and the new 2249 * thread's register state. 2250 */ 2251static inline void 2252context_switch(struct rq *rq, struct task_struct *prev, 2253 struct task_struct *next) 2254{ 2255 struct mm_struct *mm, *oldmm; 2256 2257 prepare_task_switch(rq, prev, next); 2258 2259 mm = next->mm; 2260 oldmm = prev->active_mm; 2261 /* 2262 * For paravirt, this is coupled with an exit in switch_to to 2263 * combine the page table reload and the switch backend into 2264 * one hypercall. 2265 */ 2266 arch_start_context_switch(prev); 2267 2268 if (!mm) { 2269 next->active_mm = oldmm; 2270 atomic_inc(&oldmm->mm_count); 2271 enter_lazy_tlb(oldmm, next); 2272 } else 2273 switch_mm(oldmm, mm, next); 2274 2275 if (!prev->mm) { 2276 prev->active_mm = NULL; 2277 rq->prev_mm = oldmm; 2278 } 2279 /* 2280 * Since the runqueue lock will be released by the next 2281 * task (which is an invalid locking op but in the case 2282 * of the scheduler it's an obvious special-case), so we 2283 * do an early lockdep release here: 2284 */ 2285#ifndef __ARCH_WANT_UNLOCKED_CTXSW 2286 spin_release(&rq->lock.dep_map, 1, _THIS_IP_); 2287#endif 2288 2289 context_tracking_task_switch(prev, next); 2290 /* Here we just switch the register state and the stack. */ 2291 switch_to(prev, next, prev); 2292 2293 barrier(); 2294 /* 2295 * this_rq must be evaluated again because prev may have moved 2296 * CPUs since it called schedule(), thus the 'rq' on its stack 2297 * frame will be invalid. 2298 */ 2299 finish_task_switch(this_rq(), prev); 2300} 2301 2302/* 2303 * nr_running and nr_context_switches: 2304 * 2305 * externally visible scheduler statistics: current number of runnable 2306 * threads, total number of context switches performed since bootup. 2307 */ 2308unsigned long nr_running(void) 2309{ 2310 unsigned long i, sum = 0; 2311 2312 for_each_online_cpu(i) 2313 sum += cpu_rq(i)->nr_running; 2314 2315 return sum; 2316} 2317 2318unsigned long long nr_context_switches(void) 2319{ 2320 int i; 2321 unsigned long long sum = 0; 2322 2323 for_each_possible_cpu(i) 2324 sum += cpu_rq(i)->nr_switches; 2325 2326 return sum; 2327} 2328 2329unsigned long nr_iowait(void) 2330{ 2331 unsigned long i, sum = 0; 2332 2333 for_each_possible_cpu(i) 2334 sum += atomic_read(&cpu_rq(i)->nr_iowait); 2335 2336 return sum; 2337} 2338 2339unsigned long nr_iowait_cpu(int cpu) 2340{ 2341 struct rq *this = cpu_rq(cpu); 2342 return atomic_read(&this->nr_iowait); 2343} 2344 2345#ifdef CONFIG_SMP 2346 2347/* 2348 * sched_exec - execve() is a valuable balancing opportunity, because at 2349 * this point the task has the smallest effective memory and cache footprint. 2350 */ 2351void sched_exec(void) 2352{ 2353 struct task_struct *p = current; 2354 unsigned long flags; 2355 int dest_cpu; 2356 2357 raw_spin_lock_irqsave(&p->pi_lock, flags); 2358 dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), SD_BALANCE_EXEC, 0); 2359 if (dest_cpu == smp_processor_id()) 2360 goto unlock; 2361 2362 if (likely(cpu_active(dest_cpu))) { 2363 struct migration_arg arg = { p, dest_cpu }; 2364 2365 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 2366 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg); 2367 return; 2368 } 2369unlock: 2370 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 2371} 2372 2373#endif 2374 2375DEFINE_PER_CPU(struct kernel_stat, kstat); 2376DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat); 2377 2378EXPORT_PER_CPU_SYMBOL(kstat); 2379EXPORT_PER_CPU_SYMBOL(kernel_cpustat); 2380 2381/* 2382 * Return any ns on the sched_clock that have not yet been accounted in 2383 * @p in case that task is currently running. 2384 * 2385 * Called with task_rq_lock() held on @rq. 2386 */ 2387static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq) 2388{ 2389 u64 ns = 0; 2390 2391 if (task_current(rq, p)) { 2392 update_rq_clock(rq); 2393 ns = rq_clock_task(rq) - p->se.exec_start; 2394 if ((s64)ns < 0) 2395 ns = 0; 2396 } 2397 2398 return ns; 2399} 2400 2401unsigned long long task_delta_exec(struct task_struct *p) 2402{ 2403 unsigned long flags; 2404 struct rq *rq; 2405 u64 ns = 0; 2406 2407 rq = task_rq_lock(p, &flags); 2408 ns = do_task_delta_exec(p, rq); 2409 task_rq_unlock(rq, p, &flags); 2410 2411 return ns; 2412} 2413 2414/* 2415 * Return accounted runtime for the task. 2416 * In case the task is currently running, return the runtime plus current's 2417 * pending runtime that have not been accounted yet. 2418 */ 2419unsigned long long task_sched_runtime(struct task_struct *p) 2420{ 2421 unsigned long flags; 2422 struct rq *rq; 2423 u64 ns = 0; 2424 2425#if defined(CONFIG_64BIT) && defined(CONFIG_SMP) 2426 /* 2427 * 64-bit doesn't need locks to atomically read a 64bit value. 2428 * So we have a optimization chance when the task's delta_exec is 0. 2429 * Reading ->on_cpu is racy, but this is ok. 2430 * 2431 * If we race with it leaving cpu, we'll take a lock. So we're correct. 2432 * If we race with it entering cpu, unaccounted time is 0. This is 2433 * indistinguishable from the read occurring a few cycles earlier. 2434 */ 2435 if (!p->on_cpu) 2436 return p->se.sum_exec_runtime; 2437#endif 2438 2439 rq = task_rq_lock(p, &flags); 2440 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq); 2441 task_rq_unlock(rq, p, &flags); 2442 2443 return ns; 2444} 2445 2446/* 2447 * This function gets called by the timer code, with HZ frequency. 2448 * We call it with interrupts disabled. 2449 */ 2450void scheduler_tick(void) 2451{ 2452 int cpu = smp_processor_id(); 2453 struct rq *rq = cpu_rq(cpu); 2454 struct task_struct *curr = rq->curr; 2455 2456 sched_clock_tick(); 2457 2458 raw_spin_lock(&rq->lock); 2459 update_rq_clock(rq); 2460 curr->sched_class->task_tick(rq, curr, 0); 2461 update_cpu_load_active(rq); 2462 raw_spin_unlock(&rq->lock); 2463 2464 perf_event_task_tick(); 2465 2466#ifdef CONFIG_SMP 2467 rq->idle_balance = idle_cpu(cpu); 2468 trigger_load_balance(rq); 2469#endif 2470 rq_last_tick_reset(rq); 2471} 2472 2473#ifdef CONFIG_NO_HZ_FULL 2474/** 2475 * scheduler_tick_max_deferment 2476 * 2477 * Keep at least one tick per second when a single 2478 * active task is running because the scheduler doesn't 2479 * yet completely support full dynticks environment. 2480 * 2481 * This makes sure that uptime, CFS vruntime, load 2482 * balancing, etc... continue to move forward, even 2483 * with a very low granularity. 2484 * 2485 * Return: Maximum deferment in nanoseconds. 2486 */ 2487u64 scheduler_tick_max_deferment(void) 2488{ 2489 struct rq *rq = this_rq(); 2490 unsigned long next, now = ACCESS_ONCE(jiffies); 2491 2492 next = rq->last_sched_tick + HZ; 2493 2494 if (time_before_eq(next, now)) 2495 return 0; 2496 2497 return jiffies_to_nsecs(next - now); 2498} 2499#endif 2500 2501notrace unsigned long get_parent_ip(unsigned long addr) 2502{ 2503 if (in_lock_functions(addr)) { 2504 addr = CALLER_ADDR2; 2505 if (in_lock_functions(addr)) 2506 addr = CALLER_ADDR3; 2507 } 2508 return addr; 2509} 2510 2511#if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \ 2512 defined(CONFIG_PREEMPT_TRACER)) 2513 2514void __kprobes preempt_count_add(int val) 2515{ 2516#ifdef CONFIG_DEBUG_PREEMPT 2517 /* 2518 * Underflow? 2519 */ 2520 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0))) 2521 return; 2522#endif 2523 __preempt_count_add(val); 2524#ifdef CONFIG_DEBUG_PREEMPT 2525 /* 2526 * Spinlock count overflowing soon? 2527 */ 2528 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >= 2529 PREEMPT_MASK - 10); 2530#endif 2531 if (preempt_count() == val) { 2532 unsigned long ip = get_parent_ip(CALLER_ADDR1); 2533#ifdef CONFIG_DEBUG_PREEMPT 2534 current->preempt_disable_ip = ip; 2535#endif 2536 trace_preempt_off(CALLER_ADDR0, ip); 2537 } 2538} 2539EXPORT_SYMBOL(preempt_count_add); 2540 2541void __kprobes preempt_count_sub(int val) 2542{ 2543#ifdef CONFIG_DEBUG_PREEMPT 2544 /* 2545 * Underflow? 2546 */ 2547 if (DEBUG_LOCKS_WARN_ON(val > preempt_count())) 2548 return; 2549 /* 2550 * Is the spinlock portion underflowing? 2551 */ 2552 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) && 2553 !(preempt_count() & PREEMPT_MASK))) 2554 return; 2555#endif 2556 2557 if (preempt_count() == val) 2558 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1)); 2559 __preempt_count_sub(val); 2560} 2561EXPORT_SYMBOL(preempt_count_sub); 2562 2563#endif 2564 2565/* 2566 * Print scheduling while atomic bug: 2567 */ 2568static noinline void __schedule_bug(struct task_struct *prev) 2569{ 2570 if (oops_in_progress) 2571 return; 2572 2573 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n", 2574 prev->comm, prev->pid, preempt_count()); 2575 2576 debug_show_held_locks(prev); 2577 print_modules(); 2578 if (irqs_disabled()) 2579 print_irqtrace_events(prev); 2580#ifdef CONFIG_DEBUG_PREEMPT 2581 if (in_atomic_preempt_off()) { 2582 pr_err("Preemption disabled at:"); 2583 print_ip_sym(current->preempt_disable_ip); 2584 pr_cont("\n"); 2585 } 2586#endif 2587 dump_stack(); 2588 add_taint(TAINT_WARN, LOCKDEP_STILL_OK); 2589} 2590 2591/* 2592 * Various schedule()-time debugging checks and statistics: 2593 */ 2594static inline void schedule_debug(struct task_struct *prev) 2595{ 2596 /* 2597 * Test if we are atomic. Since do_exit() needs to call into 2598 * schedule() atomically, we ignore that path. Otherwise whine 2599 * if we are scheduling when we should not. 2600 */ 2601 if (unlikely(in_atomic_preempt_off() && prev->state != TASK_DEAD)) 2602 __schedule_bug(prev); 2603 rcu_sleep_check(); 2604 2605 profile_hit(SCHED_PROFILING, __builtin_return_address(0)); 2606 2607 schedstat_inc(this_rq(), sched_count); 2608} 2609 2610/* 2611 * Pick up the highest-prio task: 2612 */ 2613static inline struct task_struct * 2614pick_next_task(struct rq *rq, struct task_struct *prev) 2615{ 2616 const struct sched_class *class = &fair_sched_class; 2617 struct task_struct *p; 2618 2619 /* 2620 * Optimization: we know that if all tasks are in 2621 * the fair class we can call that function directly: 2622 */ 2623 if (likely(prev->sched_class == class && 2624 rq->nr_running == rq->cfs.h_nr_running)) { 2625 p = fair_sched_class.pick_next_task(rq, prev); 2626 if (unlikely(p == RETRY_TASK)) 2627 goto again; 2628 2629 /* assumes fair_sched_class->next == idle_sched_class */ 2630 if (unlikely(!p)) 2631 p = idle_sched_class.pick_next_task(rq, prev); 2632 2633 return p; 2634 } 2635 2636again: 2637 for_each_class(class) { 2638 p = class->pick_next_task(rq, prev); 2639 if (p) { 2640 if (unlikely(p == RETRY_TASK)) 2641 goto again; 2642 return p; 2643 } 2644 } 2645 2646 BUG(); /* the idle class will always have a runnable task */ 2647} 2648 2649/* 2650 * __schedule() is the main scheduler function. 2651 * 2652 * The main means of driving the scheduler and thus entering this function are: 2653 * 2654 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc. 2655 * 2656 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return 2657 * paths. For example, see arch/x86/entry_64.S. 2658 * 2659 * To drive preemption between tasks, the scheduler sets the flag in timer 2660 * interrupt handler scheduler_tick(). 2661 * 2662 * 3. Wakeups don't really cause entry into schedule(). They add a 2663 * task to the run-queue and that's it. 2664 * 2665 * Now, if the new task added to the run-queue preempts the current 2666 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets 2667 * called on the nearest possible occasion: 2668 * 2669 * - If the kernel is preemptible (CONFIG_PREEMPT=y): 2670 * 2671 * - in syscall or exception context, at the next outmost 2672 * preempt_enable(). (this might be as soon as the wake_up()'s 2673 * spin_unlock()!) 2674 * 2675 * - in IRQ context, return from interrupt-handler to 2676 * preemptible context 2677 * 2678 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set) 2679 * then at the next: 2680 * 2681 * - cond_resched() call 2682 * - explicit schedule() call 2683 * - return from syscall or exception to user-space 2684 * - return from interrupt-handler to user-space 2685 */ 2686static void __sched __schedule(void) 2687{ 2688 struct task_struct *prev, *next; 2689 unsigned long *switch_count; 2690 struct rq *rq; 2691 int cpu; 2692 2693need_resched: 2694 preempt_disable(); 2695 cpu = smp_processor_id(); 2696 rq = cpu_rq(cpu); 2697 rcu_note_context_switch(cpu); 2698 prev = rq->curr; 2699 2700 schedule_debug(prev); 2701 2702 if (sched_feat(HRTICK)) 2703 hrtick_clear(rq); 2704 2705 /* 2706 * Make sure that signal_pending_state()->signal_pending() below 2707 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE) 2708 * done by the caller to avoid the race with signal_wake_up(). 2709 */ 2710 smp_mb__before_spinlock(); 2711 raw_spin_lock_irq(&rq->lock); 2712 2713 switch_count = &prev->nivcsw; 2714 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) { 2715 if (unlikely(signal_pending_state(prev->state, prev))) { 2716 prev->state = TASK_RUNNING; 2717 } else { 2718 deactivate_task(rq, prev, DEQUEUE_SLEEP); 2719 prev->on_rq = 0; 2720 2721 /* 2722 * If a worker went to sleep, notify and ask workqueue 2723 * whether it wants to wake up a task to maintain 2724 * concurrency. 2725 */ 2726 if (prev->flags & PF_WQ_WORKER) { 2727 struct task_struct *to_wakeup; 2728 2729 to_wakeup = wq_worker_sleeping(prev, cpu); 2730 if (to_wakeup) 2731 try_to_wake_up_local(to_wakeup); 2732 } 2733 } 2734 switch_count = &prev->nvcsw; 2735 } 2736 2737 if (prev->on_rq || rq->skip_clock_update < 0) 2738 update_rq_clock(rq); 2739 2740 next = pick_next_task(rq, prev); 2741 clear_tsk_need_resched(prev); 2742 clear_preempt_need_resched(); 2743 rq->skip_clock_update = 0; 2744 2745 if (likely(prev != next)) { 2746 rq->nr_switches++; 2747 rq->curr = next; 2748 ++*switch_count; 2749 2750 context_switch(rq, prev, next); /* unlocks the rq */ 2751 /* 2752 * The context switch have flipped the stack from under us 2753 * and restored the local variables which were saved when 2754 * this task called schedule() in the past. prev == current 2755 * is still correct, but it can be moved to another cpu/rq. 2756 */ 2757 cpu = smp_processor_id(); 2758 rq = cpu_rq(cpu); 2759 } else 2760 raw_spin_unlock_irq(&rq->lock); 2761 2762 post_schedule(rq); 2763 2764 sched_preempt_enable_no_resched(); 2765 if (need_resched()) 2766 goto need_resched; 2767} 2768 2769static inline void sched_submit_work(struct task_struct *tsk) 2770{ 2771 if (!tsk->state || tsk_is_pi_blocked(tsk)) 2772 return; 2773 /* 2774 * If we are going to sleep and we have plugged IO queued, 2775 * make sure to submit it to avoid deadlocks. 2776 */ 2777 if (blk_needs_flush_plug(tsk)) 2778 blk_schedule_flush_plug(tsk); 2779} 2780 2781asmlinkage void __sched schedule(void) 2782{ 2783 struct task_struct *tsk = current; 2784 2785 sched_submit_work(tsk); 2786 __schedule(); 2787} 2788EXPORT_SYMBOL(schedule); 2789 2790#ifdef CONFIG_CONTEXT_TRACKING 2791asmlinkage void __sched schedule_user(void) 2792{ 2793 /* 2794 * If we come here after a random call to set_need_resched(), 2795 * or we have been woken up remotely but the IPI has not yet arrived, 2796 * we haven't yet exited the RCU idle mode. Do it here manually until 2797 * we find a better solution. 2798 */ 2799 user_exit(); 2800 schedule(); 2801 user_enter(); 2802} 2803#endif 2804 2805/** 2806 * schedule_preempt_disabled - called with preemption disabled 2807 * 2808 * Returns with preemption disabled. Note: preempt_count must be 1 2809 */ 2810void __sched schedule_preempt_disabled(void) 2811{ 2812 sched_preempt_enable_no_resched(); 2813 schedule(); 2814 preempt_disable(); 2815} 2816 2817#ifdef CONFIG_PREEMPT 2818/* 2819 * this is the entry point to schedule() from in-kernel preemption 2820 * off of preempt_enable. Kernel preemptions off return from interrupt 2821 * occur there and call schedule directly. 2822 */ 2823asmlinkage void __sched notrace preempt_schedule(void) 2824{ 2825 /* 2826 * If there is a non-zero preempt_count or interrupts are disabled, 2827 * we do not want to preempt the current task. Just return.. 2828 */ 2829 if (likely(!preemptible())) 2830 return; 2831 2832 do { 2833 __preempt_count_add(PREEMPT_ACTIVE); 2834 __schedule(); 2835 __preempt_count_sub(PREEMPT_ACTIVE); 2836 2837 /* 2838 * Check again in case we missed a preemption opportunity 2839 * between schedule and now. 2840 */ 2841 barrier(); 2842 } while (need_resched()); 2843} 2844EXPORT_SYMBOL(preempt_schedule); 2845#endif /* CONFIG_PREEMPT */ 2846 2847/* 2848 * this is the entry point to schedule() from kernel preemption 2849 * off of irq context. 2850 * Note, that this is called and return with irqs disabled. This will 2851 * protect us against recursive calling from irq. 2852 */ 2853asmlinkage void __sched preempt_schedule_irq(void) 2854{ 2855 enum ctx_state prev_state; 2856 2857 /* Catch callers which need to be fixed */ 2858 BUG_ON(preempt_count() || !irqs_disabled()); 2859 2860 prev_state = exception_enter(); 2861 2862 do { 2863 __preempt_count_add(PREEMPT_ACTIVE); 2864 local_irq_enable(); 2865 __schedule(); 2866 local_irq_disable(); 2867 __preempt_count_sub(PREEMPT_ACTIVE); 2868 2869 /* 2870 * Check again in case we missed a preemption opportunity 2871 * between schedule and now. 2872 */ 2873 barrier(); 2874 } while (need_resched()); 2875 2876 exception_exit(prev_state); 2877} 2878 2879int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags, 2880 void *key) 2881{ 2882 return try_to_wake_up(curr->private, mode, wake_flags); 2883} 2884EXPORT_SYMBOL(default_wake_function); 2885 2886#ifdef CONFIG_RT_MUTEXES 2887 2888/* 2889 * rt_mutex_setprio - set the current priority of a task 2890 * @p: task 2891 * @prio: prio value (kernel-internal form) 2892 * 2893 * This function changes the 'effective' priority of a task. It does 2894 * not touch ->normal_prio like __setscheduler(). 2895 * 2896 * Used by the rt_mutex code to implement priority inheritance 2897 * logic. Call site only calls if the priority of the task changed. 2898 */ 2899void rt_mutex_setprio(struct task_struct *p, int prio) 2900{ 2901 int oldprio, on_rq, running, enqueue_flag = 0; 2902 struct rq *rq; 2903 const struct sched_class *prev_class; 2904 2905 BUG_ON(prio > MAX_PRIO); 2906 2907 rq = __task_rq_lock(p); 2908 2909 /* 2910 * Idle task boosting is a nono in general. There is one 2911 * exception, when PREEMPT_RT and NOHZ is active: 2912 * 2913 * The idle task calls get_next_timer_interrupt() and holds 2914 * the timer wheel base->lock on the CPU and another CPU wants 2915 * to access the timer (probably to cancel it). We can safely 2916 * ignore the boosting request, as the idle CPU runs this code 2917 * with interrupts disabled and will complete the lock 2918 * protected section without being interrupted. So there is no 2919 * real need to boost. 2920 */ 2921 if (unlikely(p == rq->idle)) { 2922 WARN_ON(p != rq->curr); 2923 WARN_ON(p->pi_blocked_on); 2924 goto out_unlock; 2925 } 2926 2927 trace_sched_pi_setprio(p, prio); 2928 p->pi_top_task = rt_mutex_get_top_task(p); 2929 oldprio = p->prio; 2930 prev_class = p->sched_class; 2931 on_rq = p->on_rq; 2932 running = task_current(rq, p); 2933 if (on_rq) 2934 dequeue_task(rq, p, 0); 2935 if (running) 2936 p->sched_class->put_prev_task(rq, p); 2937 2938 /* 2939 * Boosting condition are: 2940 * 1. -rt task is running and holds mutex A 2941 * --> -dl task blocks on mutex A 2942 * 2943 * 2. -dl task is running and holds mutex A 2944 * --> -dl task blocks on mutex A and could preempt the 2945 * running task 2946 */ 2947 if (dl_prio(prio)) { 2948 if (!dl_prio(p->normal_prio) || (p->pi_top_task && 2949 dl_entity_preempt(&p->pi_top_task->dl, &p->dl))) { 2950 p->dl.dl_boosted = 1; 2951 p->dl.dl_throttled = 0; 2952 enqueue_flag = ENQUEUE_REPLENISH; 2953 } else 2954 p->dl.dl_boosted = 0; 2955 p->sched_class = &dl_sched_class; 2956 } else if (rt_prio(prio)) { 2957 if (dl_prio(oldprio)) 2958 p->dl.dl_boosted = 0; 2959 if (oldprio < prio) 2960 enqueue_flag = ENQUEUE_HEAD; 2961 p->sched_class = &rt_sched_class; 2962 } else { 2963 if (dl_prio(oldprio)) 2964 p->dl.dl_boosted = 0; 2965 p->sched_class = &fair_sched_class; 2966 } 2967 2968 p->prio = prio; 2969 2970 if (running) 2971 p->sched_class->set_curr_task(rq); 2972 if (on_rq) 2973 enqueue_task(rq, p, enqueue_flag); 2974 2975 check_class_changed(rq, p, prev_class, oldprio); 2976out_unlock: 2977 __task_rq_unlock(rq); 2978} 2979#endif 2980 2981void set_user_nice(struct task_struct *p, long nice) 2982{ 2983 int old_prio, delta, on_rq; 2984 unsigned long flags; 2985 struct rq *rq; 2986 2987 if (task_nice(p) == nice || nice < MIN_NICE || nice > MAX_NICE) 2988 return; 2989 /* 2990 * We have to be careful, if called from sys_setpriority(), 2991 * the task might be in the middle of scheduling on another CPU. 2992 */ 2993 rq = task_rq_lock(p, &flags); 2994 /* 2995 * The RT priorities are set via sched_setscheduler(), but we still 2996 * allow the 'normal' nice value to be set - but as expected 2997 * it wont have any effect on scheduling until the task is 2998 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR: 2999 */ 3000 if (task_has_dl_policy(p) || task_has_rt_policy(p)) { 3001 p->static_prio = NICE_TO_PRIO(nice); 3002 goto out_unlock; 3003 } 3004 on_rq = p->on_rq; 3005 if (on_rq) 3006 dequeue_task(rq, p, 0); 3007 3008 p->static_prio = NICE_TO_PRIO(nice); 3009 set_load_weight(p); 3010 old_prio = p->prio; 3011 p->prio = effective_prio(p); 3012 delta = p->prio - old_prio; 3013 3014 if (on_rq) { 3015 enqueue_task(rq, p, 0); 3016 /* 3017 * If the task increased its priority or is running and 3018 * lowered its priority, then reschedule its CPU: 3019 */ 3020 if (delta < 0 || (delta > 0 && task_running(rq, p))) 3021 resched_task(rq->curr); 3022 } 3023out_unlock: 3024 task_rq_unlock(rq, p, &flags); 3025} 3026EXPORT_SYMBOL(set_user_nice); 3027 3028/* 3029 * can_nice - check if a task can reduce its nice value 3030 * @p: task 3031 * @nice: nice value 3032 */ 3033int can_nice(const struct task_struct *p, const int nice) 3034{ 3035 /* convert nice value [19,-20] to rlimit style value [1,40] */ 3036 int nice_rlim = 20 - nice; 3037 3038 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) || 3039 capable(CAP_SYS_NICE)); 3040} 3041 3042#ifdef __ARCH_WANT_SYS_NICE 3043 3044/* 3045 * sys_nice - change the priority of the current process. 3046 * @increment: priority increment 3047 * 3048 * sys_setpriority is a more generic, but much slower function that 3049 * does similar things. 3050 */ 3051SYSCALL_DEFINE1(nice, int, increment) 3052{ 3053 long nice, retval; 3054 3055 /* 3056 * Setpriority might change our priority at the same moment. 3057 * We don't have to worry. Conceptually one call occurs first 3058 * and we have a single winner. 3059 */ 3060 if (increment < -40) 3061 increment = -40; 3062 if (increment > 40) 3063 increment = 40; 3064 3065 nice = task_nice(current) + increment; 3066 if (nice < MIN_NICE) 3067 nice = MIN_NICE; 3068 if (nice > MAX_NICE) 3069 nice = MAX_NICE; 3070 3071 if (increment < 0 && !can_nice(current, nice)) 3072 return -EPERM; 3073 3074 retval = security_task_setnice(current, nice); 3075 if (retval) 3076 return retval; 3077 3078 set_user_nice(current, nice); 3079 return 0; 3080} 3081 3082#endif 3083 3084/** 3085 * task_prio - return the priority value of a given task. 3086 * @p: the task in question. 3087 * 3088 * Return: The priority value as seen by users in /proc. 3089 * RT tasks are offset by -200. Normal tasks are centered 3090 * around 0, value goes from -16 to +15. 3091 */ 3092int task_prio(const struct task_struct *p) 3093{ 3094 return p->prio - MAX_RT_PRIO; 3095} 3096 3097/** 3098 * idle_cpu - is a given cpu idle currently? 3099 * @cpu: the processor in question. 3100 * 3101 * Return: 1 if the CPU is currently idle. 0 otherwise. 3102 */ 3103int idle_cpu(int cpu) 3104{ 3105 struct rq *rq = cpu_rq(cpu); 3106 3107 if (rq->curr != rq->idle) 3108 return 0; 3109 3110 if (rq->nr_running) 3111 return 0; 3112 3113#ifdef CONFIG_SMP 3114 if (!llist_empty(&rq->wake_list)) 3115 return 0; 3116#endif 3117 3118 return 1; 3119} 3120 3121/** 3122 * idle_task - return the idle task for a given cpu. 3123 * @cpu: the processor in question. 3124 * 3125 * Return: The idle task for the cpu @cpu. 3126 */ 3127struct task_struct *idle_task(int cpu) 3128{ 3129 return cpu_rq(cpu)->idle; 3130} 3131 3132/** 3133 * find_process_by_pid - find a process with a matching PID value. 3134 * @pid: the pid in question. 3135 * 3136 * The task of @pid, if found. %NULL otherwise. 3137 */ 3138static struct task_struct *find_process_by_pid(pid_t pid) 3139{ 3140 return pid ? find_task_by_vpid(pid) : current; 3141} 3142 3143/* 3144 * This function initializes the sched_dl_entity of a newly becoming 3145 * SCHED_DEADLINE task. 3146 * 3147 * Only the static values are considered here, the actual runtime and the 3148 * absolute deadline will be properly calculated when the task is enqueued 3149 * for the first time with its new policy. 3150 */ 3151static void 3152__setparam_dl(struct task_struct *p, const struct sched_attr *attr) 3153{ 3154 struct sched_dl_entity *dl_se = &p->dl; 3155 3156 init_dl_task_timer(dl_se); 3157 dl_se->dl_runtime = attr->sched_runtime; 3158 dl_se->dl_deadline = attr->sched_deadline; 3159 dl_se->dl_period = attr->sched_period ?: dl_se->dl_deadline; 3160 dl_se->flags = attr->sched_flags; 3161 dl_se->dl_bw = to_ratio(dl_se->dl_period, dl_se->dl_runtime); 3162 dl_se->dl_throttled = 0; 3163 dl_se->dl_new = 1; 3164 dl_se->dl_yielded = 0; 3165} 3166 3167static void __setscheduler_params(struct task_struct *p, 3168 const struct sched_attr *attr) 3169{ 3170 int policy = attr->sched_policy; 3171 3172 if (policy == -1) /* setparam */ 3173 policy = p->policy; 3174 3175 p->policy = policy; 3176 3177 if (dl_policy(policy)) 3178 __setparam_dl(p, attr); 3179 else if (fair_policy(policy)) 3180 p->static_prio = NICE_TO_PRIO(attr->sched_nice); 3181 3182 /* 3183 * __sched_setscheduler() ensures attr->sched_priority == 0 when 3184 * !rt_policy. Always setting this ensures that things like 3185 * getparam()/getattr() don't report silly values for !rt tasks. 3186 */ 3187 p->rt_priority = attr->sched_priority; 3188 p->normal_prio = normal_prio(p); 3189 set_load_weight(p); 3190} 3191 3192/* Actually do priority change: must hold pi & rq lock. */ 3193static void __setscheduler(struct rq *rq, struct task_struct *p, 3194 const struct sched_attr *attr) 3195{ 3196 __setscheduler_params(p, attr); 3197 3198 /* 3199 * If we get here, there was no pi waiters boosting the 3200 * task. It is safe to use the normal prio. 3201 */ 3202 p->prio = normal_prio(p); 3203 3204 if (dl_prio(p->prio)) 3205 p->sched_class = &dl_sched_class; 3206 else if (rt_prio(p->prio)) 3207 p->sched_class = &rt_sched_class; 3208 else 3209 p->sched_class = &fair_sched_class; 3210} 3211 3212static void 3213__getparam_dl(struct task_struct *p, struct sched_attr *attr) 3214{ 3215 struct sched_dl_entity *dl_se = &p->dl; 3216 3217 attr->sched_priority = p->rt_priority; 3218 attr->sched_runtime = dl_se->dl_runtime; 3219 attr->sched_deadline = dl_se->dl_deadline; 3220 attr->sched_period = dl_se->dl_period; 3221 attr->sched_flags = dl_se->flags; 3222} 3223 3224/* 3225 * This function validates the new parameters of a -deadline task. 3226 * We ask for the deadline not being zero, and greater or equal 3227 * than the runtime, as well as the period of being zero or 3228 * greater than deadline. Furthermore, we have to be sure that 3229 * user parameters are above the internal resolution (1us); we 3230 * check sched_runtime only since it is always the smaller one. 3231 */ 3232static bool 3233__checkparam_dl(const struct sched_attr *attr) 3234{ 3235 return attr && attr->sched_deadline != 0 && 3236 (attr->sched_period == 0 || 3237 (s64)(attr->sched_period - attr->sched_deadline) >= 0) && 3238 (s64)(attr->sched_deadline - attr->sched_runtime ) >= 0 && 3239 attr->sched_runtime >= (2 << (DL_SCALE - 1)); 3240} 3241 3242/* 3243 * check the target process has a UID that matches the current process's 3244 */ 3245static bool check_same_owner(struct task_struct *p) 3246{ 3247 const struct cred *cred = current_cred(), *pcred; 3248 bool match; 3249 3250 rcu_read_lock(); 3251 pcred = __task_cred(p); 3252 match = (uid_eq(cred->euid, pcred->euid) || 3253 uid_eq(cred->euid, pcred->uid)); 3254 rcu_read_unlock(); 3255 return match; 3256} 3257 3258static int __sched_setscheduler(struct task_struct *p, 3259 const struct sched_attr *attr, 3260 bool user) 3261{ 3262 int newprio = dl_policy(attr->sched_policy) ? MAX_DL_PRIO - 1 : 3263 MAX_RT_PRIO - 1 - attr->sched_priority; 3264 int retval, oldprio, oldpolicy = -1, on_rq, running; 3265 int policy = attr->sched_policy; 3266 unsigned long flags; 3267 const struct sched_class *prev_class; 3268 struct rq *rq; 3269 int reset_on_fork; 3270 3271 /* may grab non-irq protected spin_locks */ 3272 BUG_ON(in_interrupt()); 3273recheck: 3274 /* double check policy once rq lock held */ 3275 if (policy < 0) { 3276 reset_on_fork = p->sched_reset_on_fork; 3277 policy = oldpolicy = p->policy; 3278 } else { 3279 reset_on_fork = !!(attr->sched_flags & SCHED_FLAG_RESET_ON_FORK); 3280 3281 if (policy != SCHED_DEADLINE && 3282 policy != SCHED_FIFO && policy != SCHED_RR && 3283 policy != SCHED_NORMAL && policy != SCHED_BATCH && 3284 policy != SCHED_IDLE) 3285 return -EINVAL; 3286 } 3287 3288 if (attr->sched_flags & ~(SCHED_FLAG_RESET_ON_FORK)) 3289 return -EINVAL; 3290 3291 /* 3292 * Valid priorities for SCHED_FIFO and SCHED_RR are 3293 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL, 3294 * SCHED_BATCH and SCHED_IDLE is 0. 3295 */ 3296 if ((p->mm && attr->sched_priority > MAX_USER_RT_PRIO-1) || 3297 (!p->mm && attr->sched_priority > MAX_RT_PRIO-1)) 3298 return -EINVAL; 3299 if ((dl_policy(policy) && !__checkparam_dl(attr)) || 3300 (rt_policy(policy) != (attr->sched_priority != 0))) 3301 return -EINVAL; 3302 3303 /* 3304 * Allow unprivileged RT tasks to decrease priority: 3305 */ 3306 if (user && !capable(CAP_SYS_NICE)) { 3307 if (fair_policy(policy)) { 3308 if (attr->sched_nice < task_nice(p) && 3309 !can_nice(p, attr->sched_nice)) 3310 return -EPERM; 3311 } 3312 3313 if (rt_policy(policy)) { 3314 unsigned long rlim_rtprio = 3315 task_rlimit(p, RLIMIT_RTPRIO); 3316 3317 /* can't set/change the rt policy */ 3318 if (policy != p->policy && !rlim_rtprio) 3319 return -EPERM; 3320 3321 /* can't increase priority */ 3322 if (attr->sched_priority > p->rt_priority && 3323 attr->sched_priority > rlim_rtprio) 3324 return -EPERM; 3325 } 3326 3327 /* 3328 * Can't set/change SCHED_DEADLINE policy at all for now 3329 * (safest behavior); in the future we would like to allow 3330 * unprivileged DL tasks to increase their relative deadline 3331 * or reduce their runtime (both ways reducing utilization) 3332 */ 3333 if (dl_policy(policy)) 3334 return -EPERM; 3335 3336 /* 3337 * Treat SCHED_IDLE as nice 20. Only allow a switch to 3338 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it. 3339 */ 3340 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) { 3341 if (!can_nice(p, task_nice(p))) 3342 return -EPERM; 3343 } 3344 3345 /* can't change other user's priorities */ 3346 if (!check_same_owner(p)) 3347 return -EPERM; 3348 3349 /* Normal users shall not reset the sched_reset_on_fork flag */ 3350 if (p->sched_reset_on_fork && !reset_on_fork) 3351 return -EPERM; 3352 } 3353 3354 if (user) { 3355 retval = security_task_setscheduler(p); 3356 if (retval) 3357 return retval; 3358 } 3359 3360 /* 3361 * make sure no PI-waiters arrive (or leave) while we are 3362 * changing the priority of the task: 3363 * 3364 * To be able to change p->policy safely, the appropriate 3365 * runqueue lock must be held. 3366 */ 3367 rq = task_rq_lock(p, &flags); 3368 3369 /* 3370 * Changing the policy of the stop threads its a very bad idea 3371 */ 3372 if (p == rq->stop) { 3373 task_rq_unlock(rq, p, &flags); 3374 return -EINVAL; 3375 } 3376 3377 /* 3378 * If not changing anything there's no need to proceed further, 3379 * but store a possible modification of reset_on_fork. 3380 */ 3381 if (unlikely(policy == p->policy)) { 3382 if (fair_policy(policy) && attr->sched_nice != task_nice(p)) 3383 goto change; 3384 if (rt_policy(policy) && attr->sched_priority != p->rt_priority) 3385 goto change; 3386 if (dl_policy(policy)) 3387 goto change; 3388 3389 p->sched_reset_on_fork = reset_on_fork; 3390 task_rq_unlock(rq, p, &flags); 3391 return 0; 3392 } 3393change: 3394 3395 if (user) { 3396#ifdef CONFIG_RT_GROUP_SCHED 3397 /* 3398 * Do not allow realtime tasks into groups that have no runtime 3399 * assigned. 3400 */ 3401 if (rt_bandwidth_enabled() && rt_policy(policy) && 3402 task_group(p)->rt_bandwidth.rt_runtime == 0 && 3403 !task_group_is_autogroup(task_group(p))) { 3404 task_rq_unlock(rq, p, &flags); 3405 return -EPERM; 3406 } 3407#endif 3408#ifdef CONFIG_SMP 3409 if (dl_bandwidth_enabled() && dl_policy(policy)) { 3410 cpumask_t *span = rq->rd->span; 3411 3412 /* 3413 * Don't allow tasks with an affinity mask smaller than 3414 * the entire root_domain to become SCHED_DEADLINE. We 3415 * will also fail if there's no bandwidth available. 3416 */ 3417 if (!cpumask_subset(span, &p->cpus_allowed) || 3418 rq->rd->dl_bw.bw == 0) { 3419 task_rq_unlock(rq, p, &flags); 3420 return -EPERM; 3421 } 3422 } 3423#endif 3424 } 3425 3426 /* recheck policy now with rq lock held */ 3427 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) { 3428 policy = oldpolicy = -1; 3429 task_rq_unlock(rq, p, &flags); 3430 goto recheck; 3431 } 3432 3433 /* 3434 * If setscheduling to SCHED_DEADLINE (or changing the parameters 3435 * of a SCHED_DEADLINE task) we need to check if enough bandwidth 3436 * is available. 3437 */ 3438 if ((dl_policy(policy) || dl_task(p)) && dl_overflow(p, policy, attr)) { 3439 task_rq_unlock(rq, p, &flags); 3440 return -EBUSY; 3441 } 3442 3443 p->sched_reset_on_fork = reset_on_fork; 3444 oldprio = p->prio; 3445 3446 /* 3447 * Special case for priority boosted tasks. 3448 * 3449 * If the new priority is lower or equal (user space view) 3450 * than the current (boosted) priority, we just store the new 3451 * normal parameters and do not touch the scheduler class and 3452 * the runqueue. This will be done when the task deboost 3453 * itself. 3454 */ 3455 if (rt_mutex_check_prio(p, newprio)) { 3456 __setscheduler_params(p, attr); 3457 task_rq_unlock(rq, p, &flags); 3458 return 0; 3459 } 3460 3461 on_rq = p->on_rq; 3462 running = task_current(rq, p); 3463 if (on_rq) 3464 dequeue_task(rq, p, 0); 3465 if (running) 3466 p->sched_class->put_prev_task(rq, p); 3467 3468 prev_class = p->sched_class; 3469 __setscheduler(rq, p, attr); 3470 3471 if (running) 3472 p->sched_class->set_curr_task(rq); 3473 if (on_rq) { 3474 /* 3475 * We enqueue to tail when the priority of a task is 3476 * increased (user space view). 3477 */ 3478 enqueue_task(rq, p, oldprio <= p->prio ? ENQUEUE_HEAD : 0); 3479 } 3480 3481 check_class_changed(rq, p, prev_class, oldprio); 3482 task_rq_unlock(rq, p, &flags); 3483 3484 rt_mutex_adjust_pi(p); 3485 3486 return 0; 3487} 3488 3489static int _sched_setscheduler(struct task_struct *p, int policy, 3490 const struct sched_param *param, bool check) 3491{ 3492 struct sched_attr attr = { 3493 .sched_policy = policy, 3494 .sched_priority = param->sched_priority, 3495 .sched_nice = PRIO_TO_NICE(p->static_prio), 3496 }; 3497 3498 /* 3499 * Fixup the legacy SCHED_RESET_ON_FORK hack 3500 */ 3501 if (policy & SCHED_RESET_ON_FORK) { 3502 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK; 3503 policy &= ~SCHED_RESET_ON_FORK; 3504 attr.sched_policy = policy; 3505 } 3506 3507 return __sched_setscheduler(p, &attr, check); 3508} 3509/** 3510 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread. 3511 * @p: the task in question. 3512 * @policy: new policy. 3513 * @param: structure containing the new RT priority. 3514 * 3515 * Return: 0 on success. An error code otherwise. 3516 * 3517 * NOTE that the task may be already dead. 3518 */ 3519int sched_setscheduler(struct task_struct *p, int policy, 3520 const struct sched_param *param) 3521{ 3522 return _sched_setscheduler(p, policy, param, true); 3523} 3524EXPORT_SYMBOL_GPL(sched_setscheduler); 3525 3526int sched_setattr(struct task_struct *p, const struct sched_attr *attr) 3527{ 3528 return __sched_setscheduler(p, attr, true); 3529} 3530EXPORT_SYMBOL_GPL(sched_setattr); 3531 3532/** 3533 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace. 3534 * @p: the task in question. 3535 * @policy: new policy. 3536 * @param: structure containing the new RT priority. 3537 * 3538 * Just like sched_setscheduler, only don't bother checking if the 3539 * current context has permission. For example, this is needed in 3540 * stop_machine(): we create temporary high priority worker threads, 3541 * but our caller might not have that capability. 3542 * 3543 * Return: 0 on success. An error code otherwise. 3544 */ 3545int sched_setscheduler_nocheck(struct task_struct *p, int policy, 3546 const struct sched_param *param) 3547{ 3548 return _sched_setscheduler(p, policy, param, false); 3549} 3550 3551static int 3552do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param) 3553{ 3554 struct sched_param lparam; 3555 struct task_struct *p; 3556 int retval; 3557 3558 if (!param || pid < 0) 3559 return -EINVAL; 3560 if (copy_from_user(&lparam, param, sizeof(struct sched_param))) 3561 return -EFAULT; 3562 3563 rcu_read_lock(); 3564 retval = -ESRCH; 3565 p = find_process_by_pid(pid); 3566 if (p != NULL) 3567 retval = sched_setscheduler(p, policy, &lparam); 3568 rcu_read_unlock(); 3569 3570 return retval; 3571} 3572 3573/* 3574 * Mimics kernel/events/core.c perf_copy_attr(). 3575 */ 3576static int sched_copy_attr(struct sched_attr __user *uattr, 3577 struct sched_attr *attr) 3578{ 3579 u32 size; 3580 int ret; 3581 3582 if (!access_ok(VERIFY_WRITE, uattr, SCHED_ATTR_SIZE_VER0)) 3583 return -EFAULT; 3584 3585 /* 3586 * zero the full structure, so that a short copy will be nice. 3587 */ 3588 memset(attr, 0, sizeof(*attr)); 3589 3590 ret = get_user(size, &uattr->size); 3591 if (ret) 3592 return ret; 3593 3594 if (size > PAGE_SIZE) /* silly large */ 3595 goto err_size; 3596 3597 if (!size) /* abi compat */ 3598 size = SCHED_ATTR_SIZE_VER0; 3599 3600 if (size < SCHED_ATTR_SIZE_VER0) 3601 goto err_size; 3602 3603 /* 3604 * If we're handed a bigger struct than we know of, 3605 * ensure all the unknown bits are 0 - i.e. new 3606 * user-space does not rely on any kernel feature 3607 * extensions we dont know about yet. 3608 */ 3609 if (size > sizeof(*attr)) { 3610 unsigned char __user *addr; 3611 unsigned char __user *end; 3612 unsigned char val; 3613 3614 addr = (void __user *)uattr + sizeof(*attr); 3615 end = (void __user *)uattr + size; 3616 3617 for (; addr < end; addr++) { 3618 ret = get_user(val, addr); 3619 if (ret) 3620 return ret; 3621 if (val) 3622 goto err_size; 3623 } 3624 size = sizeof(*attr); 3625 } 3626 3627 ret = copy_from_user(attr, uattr, size); 3628 if (ret) 3629 return -EFAULT; 3630 3631 /* 3632 * XXX: do we want to be lenient like existing syscalls; or do we want 3633 * to be strict and return an error on out-of-bounds values? 3634 */ 3635 attr->sched_nice = clamp(attr->sched_nice, MIN_NICE, MAX_NICE); 3636 3637out: 3638 return ret; 3639 3640err_size: 3641 put_user(sizeof(*attr), &uattr->size); 3642 ret = -E2BIG; 3643 goto out; 3644} 3645 3646/** 3647 * sys_sched_setscheduler - set/change the scheduler policy and RT priority 3648 * @pid: the pid in question. 3649 * @policy: new policy. 3650 * @param: structure containing the new RT priority. 3651 * 3652 * Return: 0 on success. An error code otherwise. 3653 */ 3654SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy, 3655 struct sched_param __user *, param) 3656{ 3657 /* negative values for policy are not valid */ 3658 if (policy < 0) 3659 return -EINVAL; 3660 3661 return do_sched_setscheduler(pid, policy, param); 3662} 3663 3664/** 3665 * sys_sched_setparam - set/change the RT priority of a thread 3666 * @pid: the pid in question. 3667 * @param: structure containing the new RT priority. 3668 * 3669 * Return: 0 on success. An error code otherwise. 3670 */ 3671SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param) 3672{ 3673 return do_sched_setscheduler(pid, -1, param); 3674} 3675 3676/** 3677 * sys_sched_setattr - same as above, but with extended sched_attr 3678 * @pid: the pid in question. 3679 * @uattr: structure containing the extended parameters. 3680 * @flags: for future extension. 3681 */ 3682SYSCALL_DEFINE3(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr, 3683 unsigned int, flags) 3684{ 3685 struct sched_attr attr; 3686 struct task_struct *p; 3687 int retval; 3688 3689 if (!uattr || pid < 0 || flags) 3690 return -EINVAL; 3691 3692 if (sched_copy_attr(uattr, &attr)) 3693 return -EFAULT; 3694 3695 rcu_read_lock(); 3696 retval = -ESRCH; 3697 p = find_process_by_pid(pid); 3698 if (p != NULL) 3699 retval = sched_setattr(p, &attr); 3700 rcu_read_unlock(); 3701 3702 return retval; 3703} 3704 3705/** 3706 * sys_sched_getscheduler - get the policy (scheduling class) of a thread 3707 * @pid: the pid in question. 3708 * 3709 * Return: On success, the policy of the thread. Otherwise, a negative error 3710 * code. 3711 */ 3712SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid) 3713{ 3714 struct task_struct *p; 3715 int retval; 3716 3717 if (pid < 0) 3718 return -EINVAL; 3719 3720 retval = -ESRCH; 3721 rcu_read_lock(); 3722 p = find_process_by_pid(pid); 3723 if (p) { 3724 retval = security_task_getscheduler(p); 3725 if (!retval) 3726 retval = p->policy 3727 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0); 3728 } 3729 rcu_read_unlock(); 3730 return retval; 3731} 3732 3733/** 3734 * sys_sched_getparam - get the RT priority of a thread 3735 * @pid: the pid in question. 3736 * @param: structure containing the RT priority. 3737 * 3738 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error 3739 * code. 3740 */ 3741SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param) 3742{ 3743 struct sched_param lp; 3744 struct task_struct *p; 3745 int retval; 3746 3747 if (!param || pid < 0) 3748 return -EINVAL; 3749 3750 rcu_read_lock(); 3751 p = find_process_by_pid(pid); 3752 retval = -ESRCH; 3753 if (!p) 3754 goto out_unlock; 3755 3756 retval = security_task_getscheduler(p); 3757 if (retval) 3758 goto out_unlock; 3759 3760 if (task_has_dl_policy(p)) { 3761 retval = -EINVAL; 3762 goto out_unlock; 3763 } 3764 lp.sched_priority = p->rt_priority; 3765 rcu_read_unlock(); 3766 3767 /* 3768 * This one might sleep, we cannot do it with a spinlock held ... 3769 */ 3770 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0; 3771 3772 return retval; 3773 3774out_unlock: 3775 rcu_read_unlock(); 3776 return retval; 3777} 3778 3779static int sched_read_attr(struct sched_attr __user *uattr, 3780 struct sched_attr *attr, 3781 unsigned int usize) 3782{ 3783 int ret; 3784 3785 if (!access_ok(VERIFY_WRITE, uattr, usize)) 3786 return -EFAULT; 3787 3788 /* 3789 * If we're handed a smaller struct than we know of, 3790 * ensure all the unknown bits are 0 - i.e. old 3791 * user-space does not get uncomplete information. 3792 */ 3793 if (usize < sizeof(*attr)) { 3794 unsigned char *addr; 3795 unsigned char *end; 3796 3797 addr = (void *)attr + usize; 3798 end = (void *)attr + sizeof(*attr); 3799 3800 for (; addr < end; addr++) { 3801 if (*addr) 3802 goto err_size; 3803 } 3804 3805 attr->size = usize; 3806 } 3807 3808 ret = copy_to_user(uattr, attr, attr->size); 3809 if (ret) 3810 return -EFAULT; 3811 3812out: 3813 return ret; 3814 3815err_size: 3816 ret = -E2BIG; 3817 goto out; 3818} 3819 3820/** 3821 * sys_sched_getattr - similar to sched_getparam, but with sched_attr 3822 * @pid: the pid in question. 3823 * @uattr: structure containing the extended parameters. 3824 * @size: sizeof(attr) for fwd/bwd comp. 3825 * @flags: for future extension. 3826 */ 3827SYSCALL_DEFINE4(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr, 3828 unsigned int, size, unsigned int, flags) 3829{ 3830 struct sched_attr attr = { 3831 .size = sizeof(struct sched_attr), 3832 }; 3833 struct task_struct *p; 3834 int retval; 3835 3836 if (!uattr || pid < 0 || size > PAGE_SIZE || 3837 size < SCHED_ATTR_SIZE_VER0 || flags) 3838 return -EINVAL; 3839 3840 rcu_read_lock(); 3841 p = find_process_by_pid(pid); 3842 retval = -ESRCH; 3843 if (!p) 3844 goto out_unlock; 3845 3846 retval = security_task_getscheduler(p); 3847 if (retval) 3848 goto out_unlock; 3849 3850 attr.sched_policy = p->policy; 3851 if (p->sched_reset_on_fork) 3852 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK; 3853 if (task_has_dl_policy(p)) 3854 __getparam_dl(p, &attr); 3855 else if (task_has_rt_policy(p)) 3856 attr.sched_priority = p->rt_priority; 3857 else 3858 attr.sched_nice = task_nice(p); 3859 3860 rcu_read_unlock(); 3861 3862 retval = sched_read_attr(uattr, &attr, size); 3863 return retval; 3864 3865out_unlock: 3866 rcu_read_unlock(); 3867 return retval; 3868} 3869 3870long sched_setaffinity(pid_t pid, const struct cpumask *in_mask) 3871{ 3872 cpumask_var_t cpus_allowed, new_mask; 3873 struct task_struct *p; 3874 int retval; 3875 3876 rcu_read_lock(); 3877 3878 p = find_process_by_pid(pid); 3879 if (!p) { 3880 rcu_read_unlock(); 3881 return -ESRCH; 3882 } 3883 3884 /* Prevent p going away */ 3885 get_task_struct(p); 3886 rcu_read_unlock(); 3887 3888 if (p->flags & PF_NO_SETAFFINITY) { 3889 retval = -EINVAL; 3890 goto out_put_task; 3891 } 3892 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) { 3893 retval = -ENOMEM; 3894 goto out_put_task; 3895 } 3896 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) { 3897 retval = -ENOMEM; 3898 goto out_free_cpus_allowed; 3899 } 3900 retval = -EPERM; 3901 if (!check_same_owner(p)) { 3902 rcu_read_lock(); 3903 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) { 3904 rcu_read_unlock(); 3905 goto out_unlock; 3906 } 3907 rcu_read_unlock(); 3908 } 3909 3910 retval = security_task_setscheduler(p); 3911 if (retval) 3912 goto out_unlock; 3913 3914 3915 cpuset_cpus_allowed(p, cpus_allowed); 3916 cpumask_and(new_mask, in_mask, cpus_allowed); 3917 3918 /* 3919 * Since bandwidth control happens on root_domain basis, 3920 * if admission test is enabled, we only admit -deadline 3921 * tasks allowed to run on all the CPUs in the task's 3922 * root_domain. 3923 */ 3924#ifdef CONFIG_SMP 3925 if (task_has_dl_policy(p)) { 3926 const struct cpumask *span = task_rq(p)->rd->span; 3927 3928 if (dl_bandwidth_enabled() && !cpumask_subset(span, new_mask)) { 3929 retval = -EBUSY; 3930 goto out_unlock; 3931 } 3932 } 3933#endif 3934again: 3935 retval = set_cpus_allowed_ptr(p, new_mask); 3936 3937 if (!retval) { 3938 cpuset_cpus_allowed(p, cpus_allowed); 3939 if (!cpumask_subset(new_mask, cpus_allowed)) { 3940 /* 3941 * We must have raced with a concurrent cpuset 3942 * update. Just reset the cpus_allowed to the 3943 * cpuset's cpus_allowed 3944 */ 3945 cpumask_copy(new_mask, cpus_allowed); 3946 goto again; 3947 } 3948 } 3949out_unlock: 3950 free_cpumask_var(new_mask); 3951out_free_cpus_allowed: 3952 free_cpumask_var(cpus_allowed); 3953out_put_task: 3954 put_task_struct(p); 3955 return retval; 3956} 3957 3958static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len, 3959 struct cpumask *new_mask) 3960{ 3961 if (len < cpumask_size()) 3962 cpumask_clear(new_mask); 3963 else if (len > cpumask_size()) 3964 len = cpumask_size(); 3965 3966 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0; 3967} 3968 3969/** 3970 * sys_sched_setaffinity - set the cpu affinity of a process 3971 * @pid: pid of the process 3972 * @len: length in bytes of the bitmask pointed to by user_mask_ptr 3973 * @user_mask_ptr: user-space pointer to the new cpu mask 3974 * 3975 * Return: 0 on success. An error code otherwise. 3976 */ 3977SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len, 3978 unsigned long __user *, user_mask_ptr) 3979{ 3980 cpumask_var_t new_mask; 3981 int retval; 3982 3983 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) 3984 return -ENOMEM; 3985 3986 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask); 3987 if (retval == 0) 3988 retval = sched_setaffinity(pid, new_mask); 3989 free_cpumask_var(new_mask); 3990 return retval; 3991} 3992 3993long sched_getaffinity(pid_t pid, struct cpumask *mask) 3994{ 3995 struct task_struct *p; 3996 unsigned long flags; 3997 int retval; 3998 3999 rcu_read_lock(); 4000 4001 retval = -ESRCH; 4002 p = find_process_by_pid(pid); 4003 if (!p) 4004 goto out_unlock; 4005 4006 retval = security_task_getscheduler(p); 4007 if (retval) 4008 goto out_unlock; 4009 4010 raw_spin_lock_irqsave(&p->pi_lock, flags); 4011 cpumask_and(mask, &p->cpus_allowed, cpu_active_mask); 4012 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 4013 4014out_unlock: 4015 rcu_read_unlock(); 4016 4017 return retval; 4018} 4019 4020/** 4021 * sys_sched_getaffinity - get the cpu affinity of a process 4022 * @pid: pid of the process 4023 * @len: length in bytes of the bitmask pointed to by user_mask_ptr 4024 * @user_mask_ptr: user-space pointer to hold the current cpu mask 4025 * 4026 * Return: 0 on success. An error code otherwise. 4027 */ 4028SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len, 4029 unsigned long __user *, user_mask_ptr) 4030{ 4031 int ret; 4032 cpumask_var_t mask; 4033 4034 if ((len * BITS_PER_BYTE) < nr_cpu_ids) 4035 return -EINVAL; 4036 if (len & (sizeof(unsigned long)-1)) 4037 return -EINVAL; 4038 4039 if (!alloc_cpumask_var(&mask, GFP_KERNEL)) 4040 return -ENOMEM; 4041 4042 ret = sched_getaffinity(pid, mask); 4043 if (ret == 0) { 4044 size_t retlen = min_t(size_t, len, cpumask_size()); 4045 4046 if (copy_to_user(user_mask_ptr, mask, retlen)) 4047 ret = -EFAULT; 4048 else 4049 ret = retlen; 4050 } 4051 free_cpumask_var(mask); 4052 4053 return ret; 4054} 4055 4056/** 4057 * sys_sched_yield - yield the current processor to other threads. 4058 * 4059 * This function yields the current CPU to other tasks. If there are no 4060 * other threads running on this CPU then this function will return. 4061 * 4062 * Return: 0. 4063 */ 4064SYSCALL_DEFINE0(sched_yield) 4065{ 4066 struct rq *rq = this_rq_lock(); 4067 4068 schedstat_inc(rq, yld_count); 4069 current->sched_class->yield_task(rq); 4070 4071 /* 4072 * Since we are going to call schedule() anyway, there's 4073 * no need to preempt or enable interrupts: 4074 */ 4075 __release(rq->lock); 4076 spin_release(&rq->lock.dep_map, 1, _THIS_IP_); 4077 do_raw_spin_unlock(&rq->lock); 4078 sched_preempt_enable_no_resched(); 4079 4080 schedule(); 4081 4082 return 0; 4083} 4084 4085static void __cond_resched(void) 4086{ 4087 __preempt_count_add(PREEMPT_ACTIVE); 4088 __schedule(); 4089 __preempt_count_sub(PREEMPT_ACTIVE); 4090} 4091 4092int __sched _cond_resched(void) 4093{ 4094 if (should_resched()) { 4095 __cond_resched(); 4096 return 1; 4097 } 4098 return 0; 4099} 4100EXPORT_SYMBOL(_cond_resched); 4101 4102/* 4103 * __cond_resched_lock() - if a reschedule is pending, drop the given lock, 4104 * call schedule, and on return reacquire the lock. 4105 * 4106 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level 4107 * operations here to prevent schedule() from being called twice (once via 4108 * spin_unlock(), once by hand). 4109 */ 4110int __cond_resched_lock(spinlock_t *lock) 4111{ 4112 int resched = should_resched(); 4113 int ret = 0; 4114 4115 lockdep_assert_held(lock); 4116 4117 if (spin_needbreak(lock) || resched) { 4118 spin_unlock(lock); 4119 if (resched) 4120 __cond_resched(); 4121 else 4122 cpu_relax(); 4123 ret = 1; 4124 spin_lock(lock); 4125 } 4126 return ret; 4127} 4128EXPORT_SYMBOL(__cond_resched_lock); 4129 4130int __sched __cond_resched_softirq(void) 4131{ 4132 BUG_ON(!in_softirq()); 4133 4134 if (should_resched()) { 4135 local_bh_enable(); 4136 __cond_resched(); 4137 local_bh_disable(); 4138 return 1; 4139 } 4140 return 0; 4141} 4142EXPORT_SYMBOL(__cond_resched_softirq); 4143 4144/** 4145 * yield - yield the current processor to other threads. 4146 * 4147 * Do not ever use this function, there's a 99% chance you're doing it wrong. 4148 * 4149 * The scheduler is at all times free to pick the calling task as the most 4150 * eligible task to run, if removing the yield() call from your code breaks 4151 * it, its already broken. 4152 * 4153 * Typical broken usage is: 4154 * 4155 * while (!event) 4156 * yield(); 4157 * 4158 * where one assumes that yield() will let 'the other' process run that will 4159 * make event true. If the current task is a SCHED_FIFO task that will never 4160 * happen. Never use yield() as a progress guarantee!! 4161 * 4162 * If you want to use yield() to wait for something, use wait_event(). 4163 * If you want to use yield() to be 'nice' for others, use cond_resched(). 4164 * If you still want to use yield(), do not! 4165 */ 4166void __sched yield(void) 4167{ 4168 set_current_state(TASK_RUNNING); 4169 sys_sched_yield(); 4170} 4171EXPORT_SYMBOL(yield); 4172 4173/** 4174 * yield_to - yield the current processor to another thread in 4175 * your thread group, or accelerate that thread toward the 4176 * processor it's on. 4177 * @p: target task 4178 * @preempt: whether task preemption is allowed or not 4179 * 4180 * It's the caller's job to ensure that the target task struct 4181 * can't go away on us before we can do any checks. 4182 * 4183 * Return: 4184 * true (>0) if we indeed boosted the target task. 4185 * false (0) if we failed to boost the target. 4186 * -ESRCH if there's no task to yield to. 4187 */ 4188bool __sched yield_to(struct task_struct *p, bool preempt) 4189{ 4190 struct task_struct *curr = current; 4191 struct rq *rq, *p_rq; 4192 unsigned long flags; 4193 int yielded = 0; 4194 4195 local_irq_save(flags); 4196 rq = this_rq(); 4197 4198again: 4199 p_rq = task_rq(p); 4200 /* 4201 * If we're the only runnable task on the rq and target rq also 4202 * has only one task, there's absolutely no point in yielding. 4203 */ 4204 if (rq->nr_running == 1 && p_rq->nr_running == 1) { 4205 yielded = -ESRCH; 4206 goto out_irq; 4207 } 4208 4209 double_rq_lock(rq, p_rq); 4210 if (task_rq(p) != p_rq) { 4211 double_rq_unlock(rq, p_rq); 4212 goto again; 4213 } 4214 4215 if (!curr->sched_class->yield_to_task) 4216 goto out_unlock; 4217 4218 if (curr->sched_class != p->sched_class) 4219 goto out_unlock; 4220 4221 if (task_running(p_rq, p) || p->state) 4222 goto out_unlock; 4223 4224 yielded = curr->sched_class->yield_to_task(rq, p, preempt); 4225 if (yielded) { 4226 schedstat_inc(rq, yld_count); 4227 /* 4228 * Make p's CPU reschedule; pick_next_entity takes care of 4229 * fairness. 4230 */ 4231 if (preempt && rq != p_rq) 4232 resched_task(p_rq->curr); 4233 } 4234 4235out_unlock: 4236 double_rq_unlock(rq, p_rq); 4237out_irq: 4238 local_irq_restore(flags); 4239 4240 if (yielded > 0) 4241 schedule(); 4242 4243 return yielded; 4244} 4245EXPORT_SYMBOL_GPL(yield_to); 4246 4247/* 4248 * This task is about to go to sleep on IO. Increment rq->nr_iowait so 4249 * that process accounting knows that this is a task in IO wait state. 4250 */ 4251void __sched io_schedule(void) 4252{ 4253 struct rq *rq = raw_rq(); 4254 4255 delayacct_blkio_start(); 4256 atomic_inc(&rq->nr_iowait); 4257 blk_flush_plug(current); 4258 current->in_iowait = 1; 4259 schedule(); 4260 current->in_iowait = 0; 4261 atomic_dec(&rq->nr_iowait); 4262 delayacct_blkio_end(); 4263} 4264EXPORT_SYMBOL(io_schedule); 4265 4266long __sched io_schedule_timeout(long timeout) 4267{ 4268 struct rq *rq = raw_rq(); 4269 long ret; 4270 4271 delayacct_blkio_start(); 4272 atomic_inc(&rq->nr_iowait); 4273 blk_flush_plug(current); 4274 current->in_iowait = 1; 4275 ret = schedule_timeout(timeout); 4276 current->in_iowait = 0; 4277 atomic_dec(&rq->nr_iowait); 4278 delayacct_blkio_end(); 4279 return ret; 4280} 4281 4282/** 4283 * sys_sched_get_priority_max - return maximum RT priority. 4284 * @policy: scheduling class. 4285 * 4286 * Return: On success, this syscall returns the maximum 4287 * rt_priority that can be used by a given scheduling class. 4288 * On failure, a negative error code is returned. 4289 */ 4290SYSCALL_DEFINE1(sched_get_priority_max, int, policy) 4291{ 4292 int ret = -EINVAL; 4293 4294 switch (policy) { 4295 case SCHED_FIFO: 4296 case SCHED_RR: 4297 ret = MAX_USER_RT_PRIO-1; 4298 break; 4299 case SCHED_DEADLINE: 4300 case SCHED_NORMAL: 4301 case SCHED_BATCH: 4302 case SCHED_IDLE: 4303 ret = 0; 4304 break; 4305 } 4306 return ret; 4307} 4308 4309/** 4310 * sys_sched_get_priority_min - return minimum RT priority. 4311 * @policy: scheduling class. 4312 * 4313 * Return: On success, this syscall returns the minimum 4314 * rt_priority that can be used by a given scheduling class. 4315 * On failure, a negative error code is returned. 4316 */ 4317SYSCALL_DEFINE1(sched_get_priority_min, int, policy) 4318{ 4319 int ret = -EINVAL; 4320 4321 switch (policy) { 4322 case SCHED_FIFO: 4323 case SCHED_RR: 4324 ret = 1; 4325 break; 4326 case SCHED_DEADLINE: 4327 case SCHED_NORMAL: 4328 case SCHED_BATCH: 4329 case SCHED_IDLE: 4330 ret = 0; 4331 } 4332 return ret; 4333} 4334 4335/** 4336 * sys_sched_rr_get_interval - return the default timeslice of a process. 4337 * @pid: pid of the process. 4338 * @interval: userspace pointer to the timeslice value. 4339 * 4340 * this syscall writes the default timeslice value of a given process 4341 * into the user-space timespec buffer. A value of '0' means infinity. 4342 * 4343 * Return: On success, 0 and the timeslice is in @interval. Otherwise, 4344 * an error code. 4345 */ 4346SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid, 4347 struct timespec __user *, interval) 4348{ 4349 struct task_struct *p; 4350 unsigned int time_slice; 4351 unsigned long flags; 4352 struct rq *rq; 4353 int retval; 4354 struct timespec t; 4355 4356 if (pid < 0) 4357 return -EINVAL; 4358 4359 retval = -ESRCH; 4360 rcu_read_lock(); 4361 p = find_process_by_pid(pid); 4362 if (!p) 4363 goto out_unlock; 4364 4365 retval = security_task_getscheduler(p); 4366 if (retval) 4367 goto out_unlock; 4368 4369 rq = task_rq_lock(p, &flags); 4370 time_slice = 0; 4371 if (p->sched_class->get_rr_interval) 4372 time_slice = p->sched_class->get_rr_interval(rq, p); 4373 task_rq_unlock(rq, p, &flags); 4374 4375 rcu_read_unlock(); 4376 jiffies_to_timespec(time_slice, &t); 4377 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0; 4378 return retval; 4379 4380out_unlock: 4381 rcu_read_unlock(); 4382 return retval; 4383} 4384 4385static const char stat_nam[] = TASK_STATE_TO_CHAR_STR; 4386 4387void sched_show_task(struct task_struct *p) 4388{ 4389 unsigned long free = 0; 4390 int ppid; 4391 unsigned state; 4392 4393 state = p->state ? __ffs(p->state) + 1 : 0; 4394 printk(KERN_INFO "%-15.15s %c", p->comm, 4395 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?'); 4396#if BITS_PER_LONG == 32 4397 if (state == TASK_RUNNING) 4398 printk(KERN_CONT " running "); 4399 else 4400 printk(KERN_CONT " %08lx ", thread_saved_pc(p)); 4401#else 4402 if (state == TASK_RUNNING) 4403 printk(KERN_CONT " running task "); 4404 else 4405 printk(KERN_CONT " %016lx ", thread_saved_pc(p)); 4406#endif 4407#ifdef CONFIG_DEBUG_STACK_USAGE 4408 free = stack_not_used(p); 4409#endif 4410 rcu_read_lock(); 4411 ppid = task_pid_nr(rcu_dereference(p->real_parent)); 4412 rcu_read_unlock(); 4413 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free, 4414 task_pid_nr(p), ppid, 4415 (unsigned long)task_thread_info(p)->flags); 4416 4417 print_worker_info(KERN_INFO, p); 4418 show_stack(p, NULL); 4419} 4420 4421void show_state_filter(unsigned long state_filter) 4422{ 4423 struct task_struct *g, *p; 4424 4425#if BITS_PER_LONG == 32 4426 printk(KERN_INFO 4427 " task PC stack pid father\n"); 4428#else 4429 printk(KERN_INFO 4430 " task PC stack pid father\n"); 4431#endif 4432 rcu_read_lock(); 4433 do_each_thread(g, p) { 4434 /* 4435 * reset the NMI-timeout, listing all files on a slow 4436 * console might take a lot of time: 4437 */ 4438 touch_nmi_watchdog(); 4439 if (!state_filter || (p->state & state_filter)) 4440 sched_show_task(p); 4441 } while_each_thread(g, p); 4442 4443 touch_all_softlockup_watchdogs(); 4444 4445#ifdef CONFIG_SCHED_DEBUG 4446 sysrq_sched_debug_show(); 4447#endif 4448 rcu_read_unlock(); 4449 /* 4450 * Only show locks if all tasks are dumped: 4451 */ 4452 if (!state_filter) 4453 debug_show_all_locks(); 4454} 4455 4456void init_idle_bootup_task(struct task_struct *idle) 4457{ 4458 idle->sched_class = &idle_sched_class; 4459} 4460 4461/** 4462 * init_idle - set up an idle thread for a given CPU 4463 * @idle: task in question 4464 * @cpu: cpu the idle task belongs to 4465 * 4466 * NOTE: this function does not set the idle thread's NEED_RESCHED 4467 * flag, to make booting more robust. 4468 */ 4469void init_idle(struct task_struct *idle, int cpu) 4470{ 4471 struct rq *rq = cpu_rq(cpu); 4472 unsigned long flags; 4473 4474 raw_spin_lock_irqsave(&rq->lock, flags); 4475 4476 __sched_fork(0, idle); 4477 idle->state = TASK_RUNNING; 4478 idle->se.exec_start = sched_clock(); 4479 4480 do_set_cpus_allowed(idle, cpumask_of(cpu)); 4481 /* 4482 * We're having a chicken and egg problem, even though we are 4483 * holding rq->lock, the cpu isn't yet set to this cpu so the 4484 * lockdep check in task_group() will fail. 4485 * 4486 * Similar case to sched_fork(). / Alternatively we could 4487 * use task_rq_lock() here and obtain the other rq->lock. 4488 * 4489 * Silence PROVE_RCU 4490 */ 4491 rcu_read_lock(); 4492 __set_task_cpu(idle, cpu); 4493 rcu_read_unlock(); 4494 4495 rq->curr = rq->idle = idle; 4496 idle->on_rq = 1; 4497#if defined(CONFIG_SMP) 4498 idle->on_cpu = 1; 4499#endif 4500 raw_spin_unlock_irqrestore(&rq->lock, flags); 4501 4502 /* Set the preempt count _outside_ the spinlocks! */ 4503 init_idle_preempt_count(idle, cpu); 4504 4505 /* 4506 * The idle tasks have their own, simple scheduling class: 4507 */ 4508 idle->sched_class = &idle_sched_class; 4509 ftrace_graph_init_idle_task(idle, cpu); 4510 vtime_init_idle(idle, cpu); 4511#if defined(CONFIG_SMP) 4512 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu); 4513#endif 4514} 4515 4516#ifdef CONFIG_SMP 4517void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask) 4518{ 4519 if (p->sched_class && p->sched_class->set_cpus_allowed) 4520 p->sched_class->set_cpus_allowed(p, new_mask); 4521 4522 cpumask_copy(&p->cpus_allowed, new_mask); 4523 p->nr_cpus_allowed = cpumask_weight(new_mask); 4524} 4525 4526/* 4527 * This is how migration works: 4528 * 4529 * 1) we invoke migration_cpu_stop() on the target CPU using 4530 * stop_one_cpu(). 4531 * 2) stopper starts to run (implicitly forcing the migrated thread 4532 * off the CPU) 4533 * 3) it checks whether the migrated task is still in the wrong runqueue. 4534 * 4) if it's in the wrong runqueue then the migration thread removes 4535 * it and puts it into the right queue. 4536 * 5) stopper completes and stop_one_cpu() returns and the migration 4537 * is done. 4538 */ 4539 4540/* 4541 * Change a given task's CPU affinity. Migrate the thread to a 4542 * proper CPU and schedule it away if the CPU it's executing on 4543 * is removed from the allowed bitmask. 4544 * 4545 * NOTE: the caller must have a valid reference to the task, the 4546 * task must not exit() & deallocate itself prematurely. The 4547 * call is not atomic; no spinlocks may be held. 4548 */ 4549int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask) 4550{ 4551 unsigned long flags; 4552 struct rq *rq; 4553 unsigned int dest_cpu; 4554 int ret = 0; 4555 4556 rq = task_rq_lock(p, &flags); 4557 4558 if (cpumask_equal(&p->cpus_allowed, new_mask)) 4559 goto out; 4560 4561 if (!cpumask_intersects(new_mask, cpu_active_mask)) { 4562 ret = -EINVAL; 4563 goto out; 4564 } 4565 4566 do_set_cpus_allowed(p, new_mask); 4567 4568 /* Can the task run on the task's current CPU? If so, we're done */ 4569 if (cpumask_test_cpu(task_cpu(p), new_mask)) 4570 goto out; 4571 4572 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask); 4573 if (p->on_rq) { 4574 struct migration_arg arg = { p, dest_cpu }; 4575 /* Need help from migration thread: drop lock and wait. */ 4576 task_rq_unlock(rq, p, &flags); 4577 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg); 4578 tlb_migrate_finish(p->mm); 4579 return 0; 4580 } 4581out: 4582 task_rq_unlock(rq, p, &flags); 4583 4584 return ret; 4585} 4586EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr); 4587 4588/* 4589 * Move (not current) task off this cpu, onto dest cpu. We're doing 4590 * this because either it can't run here any more (set_cpus_allowed() 4591 * away from this CPU, or CPU going down), or because we're 4592 * attempting to rebalance this task on exec (sched_exec). 4593 * 4594 * So we race with normal scheduler movements, but that's OK, as long 4595 * as the task is no longer on this CPU. 4596 * 4597 * Returns non-zero if task was successfully migrated. 4598 */ 4599static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu) 4600{ 4601 struct rq *rq_dest, *rq_src; 4602 int ret = 0; 4603 4604 if (unlikely(!cpu_active(dest_cpu))) 4605 return ret; 4606 4607 rq_src = cpu_rq(src_cpu); 4608 rq_dest = cpu_rq(dest_cpu); 4609 4610 raw_spin_lock(&p->pi_lock); 4611 double_rq_lock(rq_src, rq_dest); 4612 /* Already moved. */ 4613 if (task_cpu(p) != src_cpu) 4614 goto done; 4615 /* Affinity changed (again). */ 4616 if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p))) 4617 goto fail; 4618 4619 /* 4620 * If we're not on a rq, the next wake-up will ensure we're 4621 * placed properly. 4622 */ 4623 if (p->on_rq) { 4624 dequeue_task(rq_src, p, 0); 4625 set_task_cpu(p, dest_cpu); 4626 enqueue_task(rq_dest, p, 0); 4627 check_preempt_curr(rq_dest, p, 0); 4628 } 4629done: 4630 ret = 1; 4631fail: 4632 double_rq_unlock(rq_src, rq_dest); 4633 raw_spin_unlock(&p->pi_lock); 4634 return ret; 4635} 4636 4637#ifdef CONFIG_NUMA_BALANCING 4638/* Migrate current task p to target_cpu */ 4639int migrate_task_to(struct task_struct *p, int target_cpu) 4640{ 4641 struct migration_arg arg = { p, target_cpu }; 4642 int curr_cpu = task_cpu(p); 4643 4644 if (curr_cpu == target_cpu) 4645 return 0; 4646 4647 if (!cpumask_test_cpu(target_cpu, tsk_cpus_allowed(p))) 4648 return -EINVAL; 4649 4650 /* TODO: This is not properly updating schedstats */ 4651 4652 trace_sched_move_numa(p, curr_cpu, target_cpu); 4653 return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg); 4654} 4655 4656/* 4657 * Requeue a task on a given node and accurately track the number of NUMA 4658 * tasks on the runqueues 4659 */ 4660void sched_setnuma(struct task_struct *p, int nid) 4661{ 4662 struct rq *rq; 4663 unsigned long flags; 4664 bool on_rq, running; 4665 4666 rq = task_rq_lock(p, &flags); 4667 on_rq = p->on_rq; 4668 running = task_current(rq, p); 4669 4670 if (on_rq) 4671 dequeue_task(rq, p, 0); 4672 if (running) 4673 p->sched_class->put_prev_task(rq, p); 4674 4675 p->numa_preferred_nid = nid; 4676 4677 if (running) 4678 p->sched_class->set_curr_task(rq); 4679 if (on_rq) 4680 enqueue_task(rq, p, 0); 4681 task_rq_unlock(rq, p, &flags); 4682} 4683#endif 4684 4685/* 4686 * migration_cpu_stop - this will be executed by a highprio stopper thread 4687 * and performs thread migration by bumping thread off CPU then 4688 * 'pushing' onto another runqueue. 4689 */ 4690static int migration_cpu_stop(void *data) 4691{ 4692 struct migration_arg *arg = data; 4693 4694 /* 4695 * The original target cpu might have gone down and we might 4696 * be on another cpu but it doesn't matter. 4697 */ 4698 local_irq_disable(); 4699 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu); 4700 local_irq_enable(); 4701 return 0; 4702} 4703 4704#ifdef CONFIG_HOTPLUG_CPU 4705 4706/* 4707 * Ensures that the idle task is using init_mm right before its cpu goes 4708 * offline. 4709 */ 4710void idle_task_exit(void) 4711{ 4712 struct mm_struct *mm = current->active_mm; 4713 4714 BUG_ON(cpu_online(smp_processor_id())); 4715 4716 if (mm != &init_mm) { 4717 switch_mm(mm, &init_mm, current); 4718 finish_arch_post_lock_switch(); 4719 } 4720 mmdrop(mm); 4721} 4722 4723/* 4724 * Since this CPU is going 'away' for a while, fold any nr_active delta 4725 * we might have. Assumes we're called after migrate_tasks() so that the 4726 * nr_active count is stable. 4727 * 4728 * Also see the comment "Global load-average calculations". 4729 */ 4730static void calc_load_migrate(struct rq *rq) 4731{ 4732 long delta = calc_load_fold_active(rq); 4733 if (delta) 4734 atomic_long_add(delta, &calc_load_tasks); 4735} 4736 4737static void put_prev_task_fake(struct rq *rq, struct task_struct *prev) 4738{ 4739} 4740 4741static const struct sched_class fake_sched_class = { 4742 .put_prev_task = put_prev_task_fake, 4743}; 4744 4745static struct task_struct fake_task = { 4746 /* 4747 * Avoid pull_{rt,dl}_task() 4748 */ 4749 .prio = MAX_PRIO + 1, 4750 .sched_class = &fake_sched_class, 4751}; 4752 4753/* 4754 * Migrate all tasks from the rq, sleeping tasks will be migrated by 4755 * try_to_wake_up()->select_task_rq(). 4756 * 4757 * Called with rq->lock held even though we'er in stop_machine() and 4758 * there's no concurrency possible, we hold the required locks anyway 4759 * because of lock validation efforts. 4760 */ 4761static void migrate_tasks(unsigned int dead_cpu) 4762{ 4763 struct rq *rq = cpu_rq(dead_cpu); 4764 struct task_struct *next, *stop = rq->stop; 4765 int dest_cpu; 4766 4767 /* 4768 * Fudge the rq selection such that the below task selection loop 4769 * doesn't get stuck on the currently eligible stop task. 4770 * 4771 * We're currently inside stop_machine() and the rq is either stuck 4772 * in the stop_machine_cpu_stop() loop, or we're executing this code, 4773 * either way we should never end up calling schedule() until we're 4774 * done here. 4775 */ 4776 rq->stop = NULL; 4777 4778 /* 4779 * put_prev_task() and pick_next_task() sched 4780 * class method both need to have an up-to-date 4781 * value of rq->clock[_task] 4782 */ 4783 update_rq_clock(rq); 4784 4785 for ( ; ; ) { 4786 /* 4787 * There's this thread running, bail when that's the only 4788 * remaining thread. 4789 */ 4790 if (rq->nr_running == 1) 4791 break; 4792 4793 next = pick_next_task(rq, &fake_task); 4794 BUG_ON(!next); 4795 next->sched_class->put_prev_task(rq, next); 4796 4797 /* Find suitable destination for @next, with force if needed. */ 4798 dest_cpu = select_fallback_rq(dead_cpu, next); 4799 raw_spin_unlock(&rq->lock); 4800 4801 __migrate_task(next, dead_cpu, dest_cpu); 4802 4803 raw_spin_lock(&rq->lock); 4804 } 4805 4806 rq->stop = stop; 4807} 4808 4809#endif /* CONFIG_HOTPLUG_CPU */ 4810 4811#if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL) 4812 4813static struct ctl_table sd_ctl_dir[] = { 4814 { 4815 .procname = "sched_domain", 4816 .mode = 0555, 4817 }, 4818 {} 4819}; 4820 4821static struct ctl_table sd_ctl_root[] = { 4822 { 4823 .procname = "kernel", 4824 .mode = 0555, 4825 .child = sd_ctl_dir, 4826 }, 4827 {} 4828}; 4829 4830static struct ctl_table *sd_alloc_ctl_entry(int n) 4831{ 4832 struct ctl_table *entry = 4833 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL); 4834 4835 return entry; 4836} 4837 4838static void sd_free_ctl_entry(struct ctl_table **tablep) 4839{ 4840 struct ctl_table *entry; 4841 4842 /* 4843 * In the intermediate directories, both the child directory and 4844 * procname are dynamically allocated and could fail but the mode 4845 * will always be set. In the lowest directory the names are 4846 * static strings and all have proc handlers. 4847 */ 4848 for (entry = *tablep; entry->mode; entry++) { 4849 if (entry->child) 4850 sd_free_ctl_entry(&entry->child); 4851 if (entry->proc_handler == NULL) 4852 kfree(entry->procname); 4853 } 4854 4855 kfree(*tablep); 4856 *tablep = NULL; 4857} 4858 4859static int min_load_idx = 0; 4860static int max_load_idx = CPU_LOAD_IDX_MAX-1; 4861 4862static void 4863set_table_entry(struct ctl_table *entry, 4864 const char *procname, void *data, int maxlen, 4865 umode_t mode, proc_handler *proc_handler, 4866 bool load_idx) 4867{ 4868 entry->procname = procname; 4869 entry->data = data; 4870 entry->maxlen = maxlen; 4871 entry->mode = mode; 4872 entry->proc_handler = proc_handler; 4873 4874 if (load_idx) { 4875 entry->extra1 = &min_load_idx; 4876 entry->extra2 = &max_load_idx; 4877 } 4878} 4879 4880static struct ctl_table * 4881sd_alloc_ctl_domain_table(struct sched_domain *sd) 4882{ 4883 struct ctl_table *table = sd_alloc_ctl_entry(14); 4884 4885 if (table == NULL) 4886 return NULL; 4887 4888 set_table_entry(&table[0], "min_interval", &sd->min_interval, 4889 sizeof(long), 0644, proc_doulongvec_minmax, false); 4890 set_table_entry(&table[1], "max_interval", &sd->max_interval, 4891 sizeof(long), 0644, proc_doulongvec_minmax, false); 4892 set_table_entry(&table[2], "busy_idx", &sd->busy_idx, 4893 sizeof(int), 0644, proc_dointvec_minmax, true); 4894 set_table_entry(&table[3], "idle_idx", &sd->idle_idx, 4895 sizeof(int), 0644, proc_dointvec_minmax, true); 4896 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx, 4897 sizeof(int), 0644, proc_dointvec_minmax, true); 4898 set_table_entry(&table[5], "wake_idx", &sd->wake_idx, 4899 sizeof(int), 0644, proc_dointvec_minmax, true); 4900 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx, 4901 sizeof(int), 0644, proc_dointvec_minmax, true); 4902 set_table_entry(&table[7], "busy_factor", &sd->busy_factor, 4903 sizeof(int), 0644, proc_dointvec_minmax, false); 4904 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct, 4905 sizeof(int), 0644, proc_dointvec_minmax, false); 4906 set_table_entry(&table[9], "cache_nice_tries", 4907 &sd->cache_nice_tries, 4908 sizeof(int), 0644, proc_dointvec_minmax, false); 4909 set_table_entry(&table[10], "flags", &sd->flags, 4910 sizeof(int), 0644, proc_dointvec_minmax, false); 4911 set_table_entry(&table[11], "max_newidle_lb_cost", 4912 &sd->max_newidle_lb_cost, 4913 sizeof(long), 0644, proc_doulongvec_minmax, false); 4914 set_table_entry(&table[12], "name", sd->name, 4915 CORENAME_MAX_SIZE, 0444, proc_dostring, false); 4916 /* &table[13] is terminator */ 4917 4918 return table; 4919} 4920 4921static struct ctl_table *sd_alloc_ctl_cpu_table(int cpu) 4922{ 4923 struct ctl_table *entry, *table; 4924 struct sched_domain *sd; 4925 int domain_num = 0, i; 4926 char buf[32]; 4927 4928 for_each_domain(cpu, sd) 4929 domain_num++; 4930 entry = table = sd_alloc_ctl_entry(domain_num + 1); 4931 if (table == NULL) 4932 return NULL; 4933 4934 i = 0; 4935 for_each_domain(cpu, sd) { 4936 snprintf(buf, 32, "domain%d", i); 4937 entry->procname = kstrdup(buf, GFP_KERNEL); 4938 entry->mode = 0555; 4939 entry->child = sd_alloc_ctl_domain_table(sd); 4940 entry++; 4941 i++; 4942 } 4943 return table; 4944} 4945 4946static struct ctl_table_header *sd_sysctl_header; 4947static void register_sched_domain_sysctl(void) 4948{ 4949 int i, cpu_num = num_possible_cpus(); 4950 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1); 4951 char buf[32]; 4952 4953 WARN_ON(sd_ctl_dir[0].child); 4954 sd_ctl_dir[0].child = entry; 4955 4956 if (entry == NULL) 4957 return; 4958 4959 for_each_possible_cpu(i) { 4960 snprintf(buf, 32, "cpu%d", i); 4961 entry->procname = kstrdup(buf, GFP_KERNEL); 4962 entry->mode = 0555; 4963 entry->child = sd_alloc_ctl_cpu_table(i); 4964 entry++; 4965 } 4966 4967 WARN_ON(sd_sysctl_header); 4968 sd_sysctl_header = register_sysctl_table(sd_ctl_root); 4969} 4970 4971/* may be called multiple times per register */ 4972static void unregister_sched_domain_sysctl(void) 4973{ 4974 if (sd_sysctl_header) 4975 unregister_sysctl_table(sd_sysctl_header); 4976 sd_sysctl_header = NULL; 4977 if (sd_ctl_dir[0].child) 4978 sd_free_ctl_entry(&sd_ctl_dir[0].child); 4979} 4980#else 4981static void register_sched_domain_sysctl(void) 4982{ 4983} 4984static void unregister_sched_domain_sysctl(void) 4985{ 4986} 4987#endif 4988 4989static void set_rq_online(struct rq *rq) 4990{ 4991 if (!rq->online) { 4992 const struct sched_class *class; 4993 4994 cpumask_set_cpu(rq->cpu, rq->rd->online); 4995 rq->online = 1; 4996 4997 for_each_class(class) { 4998 if (class->rq_online) 4999 class->rq_online(rq); 5000 } 5001 } 5002} 5003 5004static void set_rq_offline(struct rq *rq) 5005{ 5006 if (rq->online) { 5007 const struct sched_class *class; 5008 5009 for_each_class(class) { 5010 if (class->rq_offline) 5011 class->rq_offline(rq); 5012 } 5013 5014 cpumask_clear_cpu(rq->cpu, rq->rd->online); 5015 rq->online = 0; 5016 } 5017} 5018 5019/* 5020 * migration_call - callback that gets triggered when a CPU is added. 5021 * Here we can start up the necessary migration thread for the new CPU. 5022 */ 5023static int 5024migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu) 5025{ 5026 int cpu = (long)hcpu; 5027 unsigned long flags; 5028 struct rq *rq = cpu_rq(cpu); 5029 5030 switch (action & ~CPU_TASKS_FROZEN) { 5031 5032 case CPU_UP_PREPARE: 5033 rq->calc_load_update = calc_load_update; 5034 break; 5035 5036 case CPU_ONLINE: 5037 /* Update our root-domain */ 5038 raw_spin_lock_irqsave(&rq->lock, flags); 5039 if (rq->rd) { 5040 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span)); 5041 5042 set_rq_online(rq); 5043 } 5044 raw_spin_unlock_irqrestore(&rq->lock, flags); 5045 break; 5046 5047#ifdef CONFIG_HOTPLUG_CPU 5048 case CPU_DYING: 5049 sched_ttwu_pending(); 5050 /* Update our root-domain */ 5051 raw_spin_lock_irqsave(&rq->lock, flags); 5052 if (rq->rd) { 5053 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span)); 5054 set_rq_offline(rq); 5055 } 5056 migrate_tasks(cpu); 5057 BUG_ON(rq->nr_running != 1); /* the migration thread */ 5058 raw_spin_unlock_irqrestore(&rq->lock, flags); 5059 break; 5060 5061 case CPU_DEAD: 5062 calc_load_migrate(rq); 5063 break; 5064#endif 5065 } 5066 5067 update_max_interval(); 5068 5069 return NOTIFY_OK; 5070} 5071 5072/* 5073 * Register at high priority so that task migration (migrate_all_tasks) 5074 * happens before everything else. This has to be lower priority than 5075 * the notifier in the perf_event subsystem, though. 5076 */ 5077static struct notifier_block migration_notifier = { 5078 .notifier_call = migration_call, 5079 .priority = CPU_PRI_MIGRATION, 5080}; 5081 5082static int sched_cpu_active(struct notifier_block *nfb, 5083 unsigned long action, void *hcpu) 5084{ 5085 switch (action & ~CPU_TASKS_FROZEN) { 5086 case CPU_STARTING: 5087 case CPU_DOWN_FAILED: 5088 set_cpu_active((long)hcpu, true); 5089 return NOTIFY_OK; 5090 default: 5091 return NOTIFY_DONE; 5092 } 5093} 5094 5095static int sched_cpu_inactive(struct notifier_block *nfb, 5096 unsigned long action, void *hcpu) 5097{ 5098 unsigned long flags; 5099 long cpu = (long)hcpu; 5100 5101 switch (action & ~CPU_TASKS_FROZEN) { 5102 case CPU_DOWN_PREPARE: 5103 set_cpu_active(cpu, false); 5104 5105 /* explicitly allow suspend */ 5106 if (!(action & CPU_TASKS_FROZEN)) { 5107 struct dl_bw *dl_b = dl_bw_of(cpu); 5108 bool overflow; 5109 int cpus; 5110 5111 raw_spin_lock_irqsave(&dl_b->lock, flags); 5112 cpus = dl_bw_cpus(cpu); 5113 overflow = __dl_overflow(dl_b, cpus, 0, 0); 5114 raw_spin_unlock_irqrestore(&dl_b->lock, flags); 5115 5116 if (overflow) 5117 return notifier_from_errno(-EBUSY); 5118 } 5119 return NOTIFY_OK; 5120 } 5121 5122 return NOTIFY_DONE; 5123} 5124 5125static int __init migration_init(void) 5126{ 5127 void *cpu = (void *)(long)smp_processor_id(); 5128 int err; 5129 5130 /* Initialize migration for the boot CPU */ 5131 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu); 5132 BUG_ON(err == NOTIFY_BAD); 5133 migration_call(&migration_notifier, CPU_ONLINE, cpu); 5134 register_cpu_notifier(&migration_notifier); 5135 5136 /* Register cpu active notifiers */ 5137 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE); 5138 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE); 5139 5140 return 0; 5141} 5142early_initcall(migration_init); 5143#endif 5144 5145#ifdef CONFIG_SMP 5146 5147static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */ 5148 5149#ifdef CONFIG_SCHED_DEBUG 5150 5151static __read_mostly int sched_debug_enabled; 5152 5153static int __init sched_debug_setup(char *str) 5154{ 5155 sched_debug_enabled = 1; 5156 5157 return 0; 5158} 5159early_param("sched_debug", sched_debug_setup); 5160 5161static inline bool sched_debug(void) 5162{ 5163 return sched_debug_enabled; 5164} 5165 5166static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level, 5167 struct cpumask *groupmask) 5168{ 5169 struct sched_group *group = sd->groups; 5170 char str[256]; 5171 5172 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd)); 5173 cpumask_clear(groupmask); 5174 5175 printk(KERN_DEBUG "%*s domain %d: ", level, "", level); 5176 5177 if (!(sd->flags & SD_LOAD_BALANCE)) { 5178 printk("does not load-balance\n"); 5179 if (sd->parent) 5180 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain" 5181 " has parent"); 5182 return -1; 5183 } 5184 5185 printk(KERN_CONT "span %s level %s\n", str, sd->name); 5186 5187 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) { 5188 printk(KERN_ERR "ERROR: domain->span does not contain " 5189 "CPU%d\n", cpu); 5190 } 5191 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) { 5192 printk(KERN_ERR "ERROR: domain->groups does not contain" 5193 " CPU%d\n", cpu); 5194 } 5195 5196 printk(KERN_DEBUG "%*s groups:", level + 1, ""); 5197 do { 5198 if (!group) { 5199 printk("\n"); 5200 printk(KERN_ERR "ERROR: group is NULL\n"); 5201 break; 5202 } 5203 5204 /* 5205 * Even though we initialize ->power to something semi-sane, 5206 * we leave power_orig unset. This allows us to detect if 5207 * domain iteration is still funny without causing /0 traps. 5208 */ 5209 if (!group->sgp->power_orig) { 5210 printk(KERN_CONT "\n"); 5211 printk(KERN_ERR "ERROR: domain->cpu_power not " 5212 "set\n"); 5213 break; 5214 } 5215 5216 if (!cpumask_weight(sched_group_cpus(group))) { 5217 printk(KERN_CONT "\n"); 5218 printk(KERN_ERR "ERROR: empty group\n"); 5219 break; 5220 } 5221 5222 if (!(sd->flags & SD_OVERLAP) && 5223 cpumask_intersects(groupmask, sched_group_cpus(group))) { 5224 printk(KERN_CONT "\n"); 5225 printk(KERN_ERR "ERROR: repeated CPUs\n"); 5226 break; 5227 } 5228 5229 cpumask_or(groupmask, groupmask, sched_group_cpus(group)); 5230 5231 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group)); 5232 5233 printk(KERN_CONT " %s", str); 5234 if (group->sgp->power != SCHED_POWER_SCALE) { 5235 printk(KERN_CONT " (cpu_power = %d)", 5236 group->sgp->power); 5237 } 5238 5239 group = group->next; 5240 } while (group != sd->groups); 5241 printk(KERN_CONT "\n"); 5242 5243 if (!cpumask_equal(sched_domain_span(sd), groupmask)) 5244 printk(KERN_ERR "ERROR: groups don't span domain->span\n"); 5245 5246 if (sd->parent && 5247 !cpumask_subset(groupmask, sched_domain_span(sd->parent))) 5248 printk(KERN_ERR "ERROR: parent span is not a superset " 5249 "of domain->span\n"); 5250 return 0; 5251} 5252 5253static void sched_domain_debug(struct sched_domain *sd, int cpu) 5254{ 5255 int level = 0; 5256 5257 if (!sched_debug_enabled) 5258 return; 5259 5260 if (!sd) { 5261 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu); 5262 return; 5263 } 5264 5265 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu); 5266 5267 for (;;) { 5268 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask)) 5269 break; 5270 level++; 5271 sd = sd->parent; 5272 if (!sd) 5273 break; 5274 } 5275} 5276#else /* !CONFIG_SCHED_DEBUG */ 5277# define sched_domain_debug(sd, cpu) do { } while (0) 5278static inline bool sched_debug(void) 5279{ 5280 return false; 5281} 5282#endif /* CONFIG_SCHED_DEBUG */ 5283 5284static int sd_degenerate(struct sched_domain *sd) 5285{ 5286 if (cpumask_weight(sched_domain_span(sd)) == 1) 5287 return 1; 5288 5289 /* Following flags need at least 2 groups */ 5290 if (sd->flags & (SD_LOAD_BALANCE | 5291 SD_BALANCE_NEWIDLE | 5292 SD_BALANCE_FORK | 5293 SD_BALANCE_EXEC | 5294 SD_SHARE_CPUPOWER | 5295 SD_SHARE_PKG_RESOURCES | 5296 SD_SHARE_POWERDOMAIN)) { 5297 if (sd->groups != sd->groups->next) 5298 return 0; 5299 } 5300 5301 /* Following flags don't use groups */ 5302 if (sd->flags & (SD_WAKE_AFFINE)) 5303 return 0; 5304 5305 return 1; 5306} 5307 5308static int 5309sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent) 5310{ 5311 unsigned long cflags = sd->flags, pflags = parent->flags; 5312 5313 if (sd_degenerate(parent)) 5314 return 1; 5315 5316 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent))) 5317 return 0; 5318 5319 /* Flags needing groups don't count if only 1 group in parent */ 5320 if (parent->groups == parent->groups->next) { 5321 pflags &= ~(SD_LOAD_BALANCE | 5322 SD_BALANCE_NEWIDLE | 5323 SD_BALANCE_FORK | 5324 SD_BALANCE_EXEC | 5325 SD_SHARE_CPUPOWER | 5326 SD_SHARE_PKG_RESOURCES | 5327 SD_PREFER_SIBLING | 5328 SD_SHARE_POWERDOMAIN); 5329 if (nr_node_ids == 1) 5330 pflags &= ~SD_SERIALIZE; 5331 } 5332 if (~cflags & pflags) 5333 return 0; 5334 5335 return 1; 5336} 5337 5338static void free_rootdomain(struct rcu_head *rcu) 5339{ 5340 struct root_domain *rd = container_of(rcu, struct root_domain, rcu); 5341 5342 cpupri_cleanup(&rd->cpupri); 5343 cpudl_cleanup(&rd->cpudl); 5344 free_cpumask_var(rd->dlo_mask); 5345 free_cpumask_var(rd->rto_mask); 5346 free_cpumask_var(rd->online); 5347 free_cpumask_var(rd->span); 5348 kfree(rd); 5349} 5350 5351static void rq_attach_root(struct rq *rq, struct root_domain *rd) 5352{ 5353 struct root_domain *old_rd = NULL; 5354 unsigned long flags; 5355 5356 raw_spin_lock_irqsave(&rq->lock, flags); 5357 5358 if (rq->rd) { 5359 old_rd = rq->rd; 5360 5361 if (cpumask_test_cpu(rq->cpu, old_rd->online)) 5362 set_rq_offline(rq); 5363 5364 cpumask_clear_cpu(rq->cpu, old_rd->span); 5365 5366 /* 5367 * If we dont want to free the old_rd yet then 5368 * set old_rd to NULL to skip the freeing later 5369 * in this function: 5370 */ 5371 if (!atomic_dec_and_test(&old_rd->refcount)) 5372 old_rd = NULL; 5373 } 5374 5375 atomic_inc(&rd->refcount); 5376 rq->rd = rd; 5377 5378 cpumask_set_cpu(rq->cpu, rd->span); 5379 if (cpumask_test_cpu(rq->cpu, cpu_active_mask)) 5380 set_rq_online(rq); 5381 5382 raw_spin_unlock_irqrestore(&rq->lock, flags); 5383 5384 if (old_rd) 5385 call_rcu_sched(&old_rd->rcu, free_rootdomain); 5386} 5387 5388static int init_rootdomain(struct root_domain *rd) 5389{ 5390 memset(rd, 0, sizeof(*rd)); 5391 5392 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL)) 5393 goto out; 5394 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL)) 5395 goto free_span; 5396 if (!alloc_cpumask_var(&rd->dlo_mask, GFP_KERNEL)) 5397 goto free_online; 5398 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL)) 5399 goto free_dlo_mask; 5400 5401 init_dl_bw(&rd->dl_bw); 5402 if (cpudl_init(&rd->cpudl) != 0) 5403 goto free_dlo_mask; 5404 5405 if (cpupri_init(&rd->cpupri) != 0) 5406 goto free_rto_mask; 5407 return 0; 5408 5409free_rto_mask: 5410 free_cpumask_var(rd->rto_mask); 5411free_dlo_mask: 5412 free_cpumask_var(rd->dlo_mask); 5413free_online: 5414 free_cpumask_var(rd->online); 5415free_span: 5416 free_cpumask_var(rd->span); 5417out: 5418 return -ENOMEM; 5419} 5420 5421/* 5422 * By default the system creates a single root-domain with all cpus as 5423 * members (mimicking the global state we have today). 5424 */ 5425struct root_domain def_root_domain; 5426 5427static void init_defrootdomain(void) 5428{ 5429 init_rootdomain(&def_root_domain); 5430 5431 atomic_set(&def_root_domain.refcount, 1); 5432} 5433 5434static struct root_domain *alloc_rootdomain(void) 5435{ 5436 struct root_domain *rd; 5437 5438 rd = kmalloc(sizeof(*rd), GFP_KERNEL); 5439 if (!rd) 5440 return NULL; 5441 5442 if (init_rootdomain(rd) != 0) { 5443 kfree(rd); 5444 return NULL; 5445 } 5446 5447 return rd; 5448} 5449 5450static void free_sched_groups(struct sched_group *sg, int free_sgp) 5451{ 5452 struct sched_group *tmp, *first; 5453 5454 if (!sg) 5455 return; 5456 5457 first = sg; 5458 do { 5459 tmp = sg->next; 5460 5461 if (free_sgp && atomic_dec_and_test(&sg->sgp->ref)) 5462 kfree(sg->sgp); 5463 5464 kfree(sg); 5465 sg = tmp; 5466 } while (sg != first); 5467} 5468 5469static void free_sched_domain(struct rcu_head *rcu) 5470{ 5471 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu); 5472 5473 /* 5474 * If its an overlapping domain it has private groups, iterate and 5475 * nuke them all. 5476 */ 5477 if (sd->flags & SD_OVERLAP) { 5478 free_sched_groups(sd->groups, 1); 5479 } else if (atomic_dec_and_test(&sd->groups->ref)) { 5480 kfree(sd->groups->sgp); 5481 kfree(sd->groups); 5482 } 5483 kfree(sd); 5484} 5485 5486static void destroy_sched_domain(struct sched_domain *sd, int cpu) 5487{ 5488 call_rcu(&sd->rcu, free_sched_domain); 5489} 5490 5491static void destroy_sched_domains(struct sched_domain *sd, int cpu) 5492{ 5493 for (; sd; sd = sd->parent) 5494 destroy_sched_domain(sd, cpu); 5495} 5496 5497/* 5498 * Keep a special pointer to the highest sched_domain that has 5499 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this 5500 * allows us to avoid some pointer chasing select_idle_sibling(). 5501 * 5502 * Also keep a unique ID per domain (we use the first cpu number in 5503 * the cpumask of the domain), this allows us to quickly tell if 5504 * two cpus are in the same cache domain, see cpus_share_cache(). 5505 */ 5506DEFINE_PER_CPU(struct sched_domain *, sd_llc); 5507DEFINE_PER_CPU(int, sd_llc_size); 5508DEFINE_PER_CPU(int, sd_llc_id); 5509DEFINE_PER_CPU(struct sched_domain *, sd_numa); 5510DEFINE_PER_CPU(struct sched_domain *, sd_busy); 5511DEFINE_PER_CPU(struct sched_domain *, sd_asym); 5512 5513static void update_top_cache_domain(int cpu) 5514{ 5515 struct sched_domain *sd; 5516 struct sched_domain *busy_sd = NULL; 5517 int id = cpu; 5518 int size = 1; 5519 5520 sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES); 5521 if (sd) { 5522 id = cpumask_first(sched_domain_span(sd)); 5523 size = cpumask_weight(sched_domain_span(sd)); 5524 busy_sd = sd->parent; /* sd_busy */ 5525 } 5526 rcu_assign_pointer(per_cpu(sd_busy, cpu), busy_sd); 5527 5528 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd); 5529 per_cpu(sd_llc_size, cpu) = size; 5530 per_cpu(sd_llc_id, cpu) = id; 5531 5532 sd = lowest_flag_domain(cpu, SD_NUMA); 5533 rcu_assign_pointer(per_cpu(sd_numa, cpu), sd); 5534 5535 sd = highest_flag_domain(cpu, SD_ASYM_PACKING); 5536 rcu_assign_pointer(per_cpu(sd_asym, cpu), sd); 5537} 5538 5539/* 5540 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must 5541 * hold the hotplug lock. 5542 */ 5543static void 5544cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu) 5545{ 5546 struct rq *rq = cpu_rq(cpu); 5547 struct sched_domain *tmp; 5548 5549 /* Remove the sched domains which do not contribute to scheduling. */ 5550 for (tmp = sd; tmp; ) { 5551 struct sched_domain *parent = tmp->parent; 5552 if (!parent) 5553 break; 5554 5555 if (sd_parent_degenerate(tmp, parent)) { 5556 tmp->parent = parent->parent; 5557 if (parent->parent) 5558 parent->parent->child = tmp; 5559 /* 5560 * Transfer SD_PREFER_SIBLING down in case of a 5561 * degenerate parent; the spans match for this 5562 * so the property transfers. 5563 */ 5564 if (parent->flags & SD_PREFER_SIBLING) 5565 tmp->flags |= SD_PREFER_SIBLING; 5566 destroy_sched_domain(parent, cpu); 5567 } else 5568 tmp = tmp->parent; 5569 } 5570 5571 if (sd && sd_degenerate(sd)) { 5572 tmp = sd; 5573 sd = sd->parent; 5574 destroy_sched_domain(tmp, cpu); 5575 if (sd) 5576 sd->child = NULL; 5577 } 5578 5579 sched_domain_debug(sd, cpu); 5580 5581 rq_attach_root(rq, rd); 5582 tmp = rq->sd; 5583 rcu_assign_pointer(rq->sd, sd); 5584 destroy_sched_domains(tmp, cpu); 5585 5586 update_top_cache_domain(cpu); 5587} 5588 5589/* cpus with isolated domains */ 5590static cpumask_var_t cpu_isolated_map; 5591 5592/* Setup the mask of cpus configured for isolated domains */ 5593static int __init isolated_cpu_setup(char *str) 5594{ 5595 alloc_bootmem_cpumask_var(&cpu_isolated_map); 5596 cpulist_parse(str, cpu_isolated_map); 5597 return 1; 5598} 5599 5600__setup("isolcpus=", isolated_cpu_setup); 5601 5602struct s_data { 5603 struct sched_domain ** __percpu sd; 5604 struct root_domain *rd; 5605}; 5606 5607enum s_alloc { 5608 sa_rootdomain, 5609 sa_sd, 5610 sa_sd_storage, 5611 sa_none, 5612}; 5613 5614/* 5615 * Build an iteration mask that can exclude certain CPUs from the upwards 5616 * domain traversal. 5617 * 5618 * Asymmetric node setups can result in situations where the domain tree is of 5619 * unequal depth, make sure to skip domains that already cover the entire 5620 * range. 5621 * 5622 * In that case build_sched_domains() will have terminated the iteration early 5623 * and our sibling sd spans will be empty. Domains should always include the 5624 * cpu they're built on, so check that. 5625 * 5626 */ 5627static void build_group_mask(struct sched_domain *sd, struct sched_group *sg) 5628{ 5629 const struct cpumask *span = sched_domain_span(sd); 5630 struct sd_data *sdd = sd->private; 5631 struct sched_domain *sibling; 5632 int i; 5633 5634 for_each_cpu(i, span) { 5635 sibling = *per_cpu_ptr(sdd->sd, i); 5636 if (!cpumask_test_cpu(i, sched_domain_span(sibling))) 5637 continue; 5638 5639 cpumask_set_cpu(i, sched_group_mask(sg)); 5640 } 5641} 5642 5643/* 5644 * Return the canonical balance cpu for this group, this is the first cpu 5645 * of this group that's also in the iteration mask. 5646 */ 5647int group_balance_cpu(struct sched_group *sg) 5648{ 5649 return cpumask_first_and(sched_group_cpus(sg), sched_group_mask(sg)); 5650} 5651 5652static int 5653build_overlap_sched_groups(struct sched_domain *sd, int cpu) 5654{ 5655 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg; 5656 const struct cpumask *span = sched_domain_span(sd); 5657 struct cpumask *covered = sched_domains_tmpmask; 5658 struct sd_data *sdd = sd->private; 5659 struct sched_domain *child; 5660 int i; 5661 5662 cpumask_clear(covered); 5663 5664 for_each_cpu(i, span) { 5665 struct cpumask *sg_span; 5666 5667 if (cpumask_test_cpu(i, covered)) 5668 continue; 5669 5670 child = *per_cpu_ptr(sdd->sd, i); 5671 5672 /* See the comment near build_group_mask(). */ 5673 if (!cpumask_test_cpu(i, sched_domain_span(child))) 5674 continue; 5675 5676 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(), 5677 GFP_KERNEL, cpu_to_node(cpu)); 5678 5679 if (!sg) 5680 goto fail; 5681 5682 sg_span = sched_group_cpus(sg); 5683 if (child->child) { 5684 child = child->child; 5685 cpumask_copy(sg_span, sched_domain_span(child)); 5686 } else 5687 cpumask_set_cpu(i, sg_span); 5688 5689 cpumask_or(covered, covered, sg_span); 5690 5691 sg->sgp = *per_cpu_ptr(sdd->sgp, i); 5692 if (atomic_inc_return(&sg->sgp->ref) == 1) 5693 build_group_mask(sd, sg); 5694 5695 /* 5696 * Initialize sgp->power such that even if we mess up the 5697 * domains and no possible iteration will get us here, we won't 5698 * die on a /0 trap. 5699 */ 5700 sg->sgp->power = SCHED_POWER_SCALE * cpumask_weight(sg_span); 5701 sg->sgp->power_orig = sg->sgp->power; 5702 5703 /* 5704 * Make sure the first group of this domain contains the 5705 * canonical balance cpu. Otherwise the sched_domain iteration 5706 * breaks. See update_sg_lb_stats(). 5707 */ 5708 if ((!groups && cpumask_test_cpu(cpu, sg_span)) || 5709 group_balance_cpu(sg) == cpu) 5710 groups = sg; 5711 5712 if (!first) 5713 first = sg; 5714 if (last) 5715 last->next = sg; 5716 last = sg; 5717 last->next = first; 5718 } 5719 sd->groups = groups; 5720 5721 return 0; 5722 5723fail: 5724 free_sched_groups(first, 0); 5725 5726 return -ENOMEM; 5727} 5728 5729static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg) 5730{ 5731 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu); 5732 struct sched_domain *child = sd->child; 5733 5734 if (child) 5735 cpu = cpumask_first(sched_domain_span(child)); 5736 5737 if (sg) { 5738 *sg = *per_cpu_ptr(sdd->sg, cpu); 5739 (*sg)->sgp = *per_cpu_ptr(sdd->sgp, cpu); 5740 atomic_set(&(*sg)->sgp->ref, 1); /* for claim_allocations */ 5741 } 5742 5743 return cpu; 5744} 5745 5746/* 5747 * build_sched_groups will build a circular linked list of the groups 5748 * covered by the given span, and will set each group's ->cpumask correctly, 5749 * and ->cpu_power to 0. 5750 * 5751 * Assumes the sched_domain tree is fully constructed 5752 */ 5753static int 5754build_sched_groups(struct sched_domain *sd, int cpu) 5755{ 5756 struct sched_group *first = NULL, *last = NULL; 5757 struct sd_data *sdd = sd->private; 5758 const struct cpumask *span = sched_domain_span(sd); 5759 struct cpumask *covered; 5760 int i; 5761 5762 get_group(cpu, sdd, &sd->groups); 5763 atomic_inc(&sd->groups->ref); 5764 5765 if (cpu != cpumask_first(span)) 5766 return 0; 5767 5768 lockdep_assert_held(&sched_domains_mutex); 5769 covered = sched_domains_tmpmask; 5770 5771 cpumask_clear(covered); 5772 5773 for_each_cpu(i, span) { 5774 struct sched_group *sg; 5775 int group, j; 5776 5777 if (cpumask_test_cpu(i, covered)) 5778 continue; 5779 5780 group = get_group(i, sdd, &sg); 5781 cpumask_clear(sched_group_cpus(sg)); 5782 sg->sgp->power = 0; 5783 cpumask_setall(sched_group_mask(sg)); 5784 5785 for_each_cpu(j, span) { 5786 if (get_group(j, sdd, NULL) != group) 5787 continue; 5788 5789 cpumask_set_cpu(j, covered); 5790 cpumask_set_cpu(j, sched_group_cpus(sg)); 5791 } 5792 5793 if (!first) 5794 first = sg; 5795 if (last) 5796 last->next = sg; 5797 last = sg; 5798 } 5799 last->next = first; 5800 5801 return 0; 5802} 5803 5804/* 5805 * Initialize sched groups cpu_power. 5806 * 5807 * cpu_power indicates the capacity of sched group, which is used while 5808 * distributing the load between different sched groups in a sched domain. 5809 * Typically cpu_power for all the groups in a sched domain will be same unless 5810 * there are asymmetries in the topology. If there are asymmetries, group 5811 * having more cpu_power will pickup more load compared to the group having 5812 * less cpu_power. 5813 */ 5814static void init_sched_groups_power(int cpu, struct sched_domain *sd) 5815{ 5816 struct sched_group *sg = sd->groups; 5817 5818 WARN_ON(!sg); 5819 5820 do { 5821 sg->group_weight = cpumask_weight(sched_group_cpus(sg)); 5822 sg = sg->next; 5823 } while (sg != sd->groups); 5824 5825 if (cpu != group_balance_cpu(sg)) 5826 return; 5827 5828 update_group_power(sd, cpu); 5829 atomic_set(&sg->sgp->nr_busy_cpus, sg->group_weight); 5830} 5831 5832/* 5833 * Initializers for schedule domains 5834 * Non-inlined to reduce accumulated stack pressure in build_sched_domains() 5835 */ 5836 5837static int default_relax_domain_level = -1; 5838int sched_domain_level_max; 5839 5840static int __init setup_relax_domain_level(char *str) 5841{ 5842 if (kstrtoint(str, 0, &default_relax_domain_level)) 5843 pr_warn("Unable to set relax_domain_level\n"); 5844 5845 return 1; 5846} 5847__setup("relax_domain_level=", setup_relax_domain_level); 5848 5849static void set_domain_attribute(struct sched_domain *sd, 5850 struct sched_domain_attr *attr) 5851{ 5852 int request; 5853 5854 if (!attr || attr->relax_domain_level < 0) { 5855 if (default_relax_domain_level < 0) 5856 return; 5857 else 5858 request = default_relax_domain_level; 5859 } else 5860 request = attr->relax_domain_level; 5861 if (request < sd->level) { 5862 /* turn off idle balance on this domain */ 5863 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE); 5864 } else { 5865 /* turn on idle balance on this domain */ 5866 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE); 5867 } 5868} 5869 5870static void __sdt_free(const struct cpumask *cpu_map); 5871static int __sdt_alloc(const struct cpumask *cpu_map); 5872 5873static void __free_domain_allocs(struct s_data *d, enum s_alloc what, 5874 const struct cpumask *cpu_map) 5875{ 5876 switch (what) { 5877 case sa_rootdomain: 5878 if (!atomic_read(&d->rd->refcount)) 5879 free_rootdomain(&d->rd->rcu); /* fall through */ 5880 case sa_sd: 5881 free_percpu(d->sd); /* fall through */ 5882 case sa_sd_storage: 5883 __sdt_free(cpu_map); /* fall through */ 5884 case sa_none: 5885 break; 5886 } 5887} 5888 5889static enum s_alloc __visit_domain_allocation_hell(struct s_data *d, 5890 const struct cpumask *cpu_map) 5891{ 5892 memset(d, 0, sizeof(*d)); 5893 5894 if (__sdt_alloc(cpu_map)) 5895 return sa_sd_storage; 5896 d->sd = alloc_percpu(struct sched_domain *); 5897 if (!d->sd) 5898 return sa_sd_storage; 5899 d->rd = alloc_rootdomain(); 5900 if (!d->rd) 5901 return sa_sd; 5902 return sa_rootdomain; 5903} 5904 5905/* 5906 * NULL the sd_data elements we've used to build the sched_domain and 5907 * sched_group structure so that the subsequent __free_domain_allocs() 5908 * will not free the data we're using. 5909 */ 5910static void claim_allocations(int cpu, struct sched_domain *sd) 5911{ 5912 struct sd_data *sdd = sd->private; 5913 5914 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd); 5915 *per_cpu_ptr(sdd->sd, cpu) = NULL; 5916 5917 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref)) 5918 *per_cpu_ptr(sdd->sg, cpu) = NULL; 5919 5920 if (atomic_read(&(*per_cpu_ptr(sdd->sgp, cpu))->ref)) 5921 *per_cpu_ptr(sdd->sgp, cpu) = NULL; 5922} 5923 5924#ifdef CONFIG_NUMA 5925static int sched_domains_numa_levels; 5926static int *sched_domains_numa_distance; 5927static struct cpumask ***sched_domains_numa_masks; 5928static int sched_domains_curr_level; 5929#endif 5930 5931/* 5932 * SD_flags allowed in topology descriptions. 5933 * 5934 * SD_SHARE_CPUPOWER - describes SMT topologies 5935 * SD_SHARE_PKG_RESOURCES - describes shared caches 5936 * SD_NUMA - describes NUMA topologies 5937 * SD_SHARE_POWERDOMAIN - describes shared power domain 5938 * 5939 * Odd one out: 5940 * SD_ASYM_PACKING - describes SMT quirks 5941 */ 5942#define TOPOLOGY_SD_FLAGS \ 5943 (SD_SHARE_CPUPOWER | \ 5944 SD_SHARE_PKG_RESOURCES | \ 5945 SD_NUMA | \ 5946 SD_ASYM_PACKING | \ 5947 SD_SHARE_POWERDOMAIN) 5948 5949static struct sched_domain * 5950sd_init(struct sched_domain_topology_level *tl, int cpu) 5951{ 5952 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); 5953 int sd_weight, sd_flags = 0; 5954 5955#ifdef CONFIG_NUMA 5956 /* 5957 * Ugly hack to pass state to sd_numa_mask()... 5958 */ 5959 sched_domains_curr_level = tl->numa_level; 5960#endif 5961 5962 sd_weight = cpumask_weight(tl->mask(cpu)); 5963 5964 if (tl->sd_flags) 5965 sd_flags = (*tl->sd_flags)(); 5966 if (WARN_ONCE(sd_flags & ~TOPOLOGY_SD_FLAGS, 5967 "wrong sd_flags in topology description\n")) 5968 sd_flags &= ~TOPOLOGY_SD_FLAGS; 5969 5970 *sd = (struct sched_domain){ 5971 .min_interval = sd_weight, 5972 .max_interval = 2*sd_weight, 5973 .busy_factor = 32, 5974 .imbalance_pct = 125, 5975 5976 .cache_nice_tries = 0, 5977 .busy_idx = 0, 5978 .idle_idx = 0, 5979 .newidle_idx = 0, 5980 .wake_idx = 0, 5981 .forkexec_idx = 0, 5982 5983 .flags = 1*SD_LOAD_BALANCE 5984 | 1*SD_BALANCE_NEWIDLE 5985 | 1*SD_BALANCE_EXEC 5986 | 1*SD_BALANCE_FORK 5987 | 0*SD_BALANCE_WAKE 5988 | 1*SD_WAKE_AFFINE 5989 | 0*SD_SHARE_CPUPOWER 5990 | 0*SD_SHARE_PKG_RESOURCES 5991 | 0*SD_SERIALIZE 5992 | 0*SD_PREFER_SIBLING 5993 | 0*SD_NUMA 5994 | sd_flags 5995 , 5996 5997 .last_balance = jiffies, 5998 .balance_interval = sd_weight, 5999 .smt_gain = 0, 6000 .max_newidle_lb_cost = 0, 6001 .next_decay_max_lb_cost = jiffies, 6002#ifdef CONFIG_SCHED_DEBUG 6003 .name = tl->name, 6004#endif 6005 }; 6006 6007 /* 6008 * Convert topological properties into behaviour. 6009 */ 6010 6011 if (sd->flags & SD_SHARE_CPUPOWER) { 6012 sd->imbalance_pct = 110; 6013 sd->smt_gain = 1178; /* ~15% */ 6014 6015 } else if (sd->flags & SD_SHARE_PKG_RESOURCES) { 6016 sd->imbalance_pct = 117; 6017 sd->cache_nice_tries = 1; 6018 sd->busy_idx = 2; 6019 6020#ifdef CONFIG_NUMA 6021 } else if (sd->flags & SD_NUMA) { 6022 sd->cache_nice_tries = 2; 6023 sd->busy_idx = 3; 6024 sd->idle_idx = 2; 6025 6026 sd->flags |= SD_SERIALIZE; 6027 if (sched_domains_numa_distance[tl->numa_level] > RECLAIM_DISTANCE) { 6028 sd->flags &= ~(SD_BALANCE_EXEC | 6029 SD_BALANCE_FORK | 6030 SD_WAKE_AFFINE); 6031 } 6032 6033#endif 6034 } else { 6035 sd->flags |= SD_PREFER_SIBLING; 6036 sd->cache_nice_tries = 1; 6037 sd->busy_idx = 2; 6038 sd->idle_idx = 1; 6039 } 6040 6041 sd->private = &tl->data; 6042 6043 return sd; 6044} 6045 6046/* 6047 * Topology list, bottom-up. 6048 */ 6049static struct sched_domain_topology_level default_topology[] = { 6050#ifdef CONFIG_SCHED_SMT 6051 { cpu_smt_mask, cpu_smt_flags, SD_INIT_NAME(SMT) }, 6052#endif 6053#ifdef CONFIG_SCHED_MC 6054 { cpu_coregroup_mask, cpu_core_flags, SD_INIT_NAME(MC) }, 6055#endif 6056 { cpu_cpu_mask, SD_INIT_NAME(DIE) }, 6057 { NULL, }, 6058}; 6059 6060struct sched_domain_topology_level *sched_domain_topology = default_topology; 6061 6062#define for_each_sd_topology(tl) \ 6063 for (tl = sched_domain_topology; tl->mask; tl++) 6064 6065void set_sched_topology(struct sched_domain_topology_level *tl) 6066{ 6067 sched_domain_topology = tl; 6068} 6069 6070#ifdef CONFIG_NUMA 6071 6072static const struct cpumask *sd_numa_mask(int cpu) 6073{ 6074 return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)]; 6075} 6076 6077static void sched_numa_warn(const char *str) 6078{ 6079 static int done = false; 6080 int i,j; 6081 6082 if (done) 6083 return; 6084 6085 done = true; 6086 6087 printk(KERN_WARNING "ERROR: %s\n\n", str); 6088 6089 for (i = 0; i < nr_node_ids; i++) { 6090 printk(KERN_WARNING " "); 6091 for (j = 0; j < nr_node_ids; j++) 6092 printk(KERN_CONT "%02d ", node_distance(i,j)); 6093 printk(KERN_CONT "\n"); 6094 } 6095 printk(KERN_WARNING "\n"); 6096} 6097 6098static bool find_numa_distance(int distance) 6099{ 6100 int i; 6101 6102 if (distance == node_distance(0, 0)) 6103 return true; 6104 6105 for (i = 0; i < sched_domains_numa_levels; i++) { 6106 if (sched_domains_numa_distance[i] == distance) 6107 return true; 6108 } 6109 6110 return false; 6111} 6112 6113static void sched_init_numa(void) 6114{ 6115 int next_distance, curr_distance = node_distance(0, 0); 6116 struct sched_domain_topology_level *tl; 6117 int level = 0; 6118 int i, j, k; 6119 6120 sched_domains_numa_distance = kzalloc(sizeof(int) * nr_node_ids, GFP_KERNEL); 6121 if (!sched_domains_numa_distance) 6122 return; 6123 6124 /* 6125 * O(nr_nodes^2) deduplicating selection sort -- in order to find the 6126 * unique distances in the node_distance() table. 6127 * 6128 * Assumes node_distance(0,j) includes all distances in 6129 * node_distance(i,j) in order to avoid cubic time. 6130 */ 6131 next_distance = curr_distance; 6132 for (i = 0; i < nr_node_ids; i++) { 6133 for (j = 0; j < nr_node_ids; j++) { 6134 for (k = 0; k < nr_node_ids; k++) { 6135 int distance = node_distance(i, k); 6136 6137 if (distance > curr_distance && 6138 (distance < next_distance || 6139 next_distance == curr_distance)) 6140 next_distance = distance; 6141 6142 /* 6143 * While not a strong assumption it would be nice to know 6144 * about cases where if node A is connected to B, B is not 6145 * equally connected to A. 6146 */ 6147 if (sched_debug() && node_distance(k, i) != distance) 6148 sched_numa_warn("Node-distance not symmetric"); 6149 6150 if (sched_debug() && i && !find_numa_distance(distance)) 6151 sched_numa_warn("Node-0 not representative"); 6152 } 6153 if (next_distance != curr_distance) { 6154 sched_domains_numa_distance[level++] = next_distance; 6155 sched_domains_numa_levels = level; 6156 curr_distance = next_distance; 6157 } else break; 6158 } 6159 6160 /* 6161 * In case of sched_debug() we verify the above assumption. 6162 */ 6163 if (!sched_debug()) 6164 break; 6165 } 6166 /* 6167 * 'level' contains the number of unique distances, excluding the 6168 * identity distance node_distance(i,i). 6169 * 6170 * The sched_domains_numa_distance[] array includes the actual distance 6171 * numbers. 6172 */ 6173 6174 /* 6175 * Here, we should temporarily reset sched_domains_numa_levels to 0. 6176 * If it fails to allocate memory for array sched_domains_numa_masks[][], 6177 * the array will contain less then 'level' members. This could be 6178 * dangerous when we use it to iterate array sched_domains_numa_masks[][] 6179 * in other functions. 6180 * 6181 * We reset it to 'level' at the end of this function. 6182 */ 6183 sched_domains_numa_levels = 0; 6184 6185 sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL); 6186 if (!sched_domains_numa_masks) 6187 return; 6188 6189 /* 6190 * Now for each level, construct a mask per node which contains all 6191 * cpus of nodes that are that many hops away from us. 6192 */ 6193 for (i = 0; i < level; i++) { 6194 sched_domains_numa_masks[i] = 6195 kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL); 6196 if (!sched_domains_numa_masks[i]) 6197 return; 6198 6199 for (j = 0; j < nr_node_ids; j++) { 6200 struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL); 6201 if (!mask) 6202 return; 6203 6204 sched_domains_numa_masks[i][j] = mask; 6205 6206 for (k = 0; k < nr_node_ids; k++) { 6207 if (node_distance(j, k) > sched_domains_numa_distance[i]) 6208 continue; 6209 6210 cpumask_or(mask, mask, cpumask_of_node(k)); 6211 } 6212 } 6213 } 6214 6215 /* Compute default topology size */ 6216 for (i = 0; sched_domain_topology[i].mask; i++); 6217 6218 tl = kzalloc((i + level) * 6219 sizeof(struct sched_domain_topology_level), GFP_KERNEL); 6220 if (!tl) 6221 return; 6222 6223 /* 6224 * Copy the default topology bits.. 6225 */ 6226 for (i = 0; sched_domain_topology[i].mask; i++) 6227 tl[i] = sched_domain_topology[i]; 6228 6229 /* 6230 * .. and append 'j' levels of NUMA goodness. 6231 */ 6232 for (j = 0; j < level; i++, j++) { 6233 tl[i] = (struct sched_domain_topology_level){ 6234 .mask = sd_numa_mask, 6235 .sd_flags = cpu_numa_flags, 6236 .flags = SDTL_OVERLAP, 6237 .numa_level = j, 6238 SD_INIT_NAME(NUMA) 6239 }; 6240 } 6241 6242 sched_domain_topology = tl; 6243 6244 sched_domains_numa_levels = level; 6245} 6246 6247static void sched_domains_numa_masks_set(int cpu) 6248{ 6249 int i, j; 6250 int node = cpu_to_node(cpu); 6251 6252 for (i = 0; i < sched_domains_numa_levels; i++) { 6253 for (j = 0; j < nr_node_ids; j++) { 6254 if (node_distance(j, node) <= sched_domains_numa_distance[i]) 6255 cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]); 6256 } 6257 } 6258} 6259 6260static void sched_domains_numa_masks_clear(int cpu) 6261{ 6262 int i, j; 6263 for (i = 0; i < sched_domains_numa_levels; i++) { 6264 for (j = 0; j < nr_node_ids; j++) 6265 cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]); 6266 } 6267} 6268 6269/* 6270 * Update sched_domains_numa_masks[level][node] array when new cpus 6271 * are onlined. 6272 */ 6273static int sched_domains_numa_masks_update(struct notifier_block *nfb, 6274 unsigned long action, 6275 void *hcpu) 6276{ 6277 int cpu = (long)hcpu; 6278 6279 switch (action & ~CPU_TASKS_FROZEN) { 6280 case CPU_ONLINE: 6281 sched_domains_numa_masks_set(cpu); 6282 break; 6283 6284 case CPU_DEAD: 6285 sched_domains_numa_masks_clear(cpu); 6286 break; 6287 6288 default: 6289 return NOTIFY_DONE; 6290 } 6291 6292 return NOTIFY_OK; 6293} 6294#else 6295static inline void sched_init_numa(void) 6296{ 6297} 6298 6299static int sched_domains_numa_masks_update(struct notifier_block *nfb, 6300 unsigned long action, 6301 void *hcpu) 6302{ 6303 return 0; 6304} 6305#endif /* CONFIG_NUMA */ 6306 6307static int __sdt_alloc(const struct cpumask *cpu_map) 6308{ 6309 struct sched_domain_topology_level *tl; 6310 int j; 6311 6312 for_each_sd_topology(tl) { 6313 struct sd_data *sdd = &tl->data; 6314 6315 sdd->sd = alloc_percpu(struct sched_domain *); 6316 if (!sdd->sd) 6317 return -ENOMEM; 6318 6319 sdd->sg = alloc_percpu(struct sched_group *); 6320 if (!sdd->sg) 6321 return -ENOMEM; 6322 6323 sdd->sgp = alloc_percpu(struct sched_group_power *); 6324 if (!sdd->sgp) 6325 return -ENOMEM; 6326 6327 for_each_cpu(j, cpu_map) { 6328 struct sched_domain *sd; 6329 struct sched_group *sg; 6330 struct sched_group_power *sgp; 6331 6332 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(), 6333 GFP_KERNEL, cpu_to_node(j)); 6334 if (!sd) 6335 return -ENOMEM; 6336 6337 *per_cpu_ptr(sdd->sd, j) = sd; 6338 6339 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(), 6340 GFP_KERNEL, cpu_to_node(j)); 6341 if (!sg) 6342 return -ENOMEM; 6343 6344 sg->next = sg; 6345 6346 *per_cpu_ptr(sdd->sg, j) = sg; 6347 6348 sgp = kzalloc_node(sizeof(struct sched_group_power) + cpumask_size(), 6349 GFP_KERNEL, cpu_to_node(j)); 6350 if (!sgp) 6351 return -ENOMEM; 6352 6353 *per_cpu_ptr(sdd->sgp, j) = sgp; 6354 } 6355 } 6356 6357 return 0; 6358} 6359 6360static void __sdt_free(const struct cpumask *cpu_map) 6361{ 6362 struct sched_domain_topology_level *tl; 6363 int j; 6364 6365 for_each_sd_topology(tl) { 6366 struct sd_data *sdd = &tl->data; 6367 6368 for_each_cpu(j, cpu_map) { 6369 struct sched_domain *sd; 6370 6371 if (sdd->sd) { 6372 sd = *per_cpu_ptr(sdd->sd, j); 6373 if (sd && (sd->flags & SD_OVERLAP)) 6374 free_sched_groups(sd->groups, 0); 6375 kfree(*per_cpu_ptr(sdd->sd, j)); 6376 } 6377 6378 if (sdd->sg) 6379 kfree(*per_cpu_ptr(sdd->sg, j)); 6380 if (sdd->sgp) 6381 kfree(*per_cpu_ptr(sdd->sgp, j)); 6382 } 6383 free_percpu(sdd->sd); 6384 sdd->sd = NULL; 6385 free_percpu(sdd->sg); 6386 sdd->sg = NULL; 6387 free_percpu(sdd->sgp); 6388 sdd->sgp = NULL; 6389 } 6390} 6391 6392struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl, 6393 const struct cpumask *cpu_map, struct sched_domain_attr *attr, 6394 struct sched_domain *child, int cpu) 6395{ 6396 struct sched_domain *sd = sd_init(tl, cpu); 6397 if (!sd) 6398 return child; 6399 6400 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu)); 6401 if (child) { 6402 sd->level = child->level + 1; 6403 sched_domain_level_max = max(sched_domain_level_max, sd->level); 6404 child->parent = sd; 6405 sd->child = child; 6406 } 6407 set_domain_attribute(sd, attr); 6408 6409 return sd; 6410} 6411 6412/* 6413 * Build sched domains for a given set of cpus and attach the sched domains 6414 * to the individual cpus 6415 */ 6416static int build_sched_domains(const struct cpumask *cpu_map, 6417 struct sched_domain_attr *attr) 6418{ 6419 enum s_alloc alloc_state; 6420 struct sched_domain *sd; 6421 struct s_data d; 6422 int i, ret = -ENOMEM; 6423 6424 alloc_state = __visit_domain_allocation_hell(&d, cpu_map); 6425 if (alloc_state != sa_rootdomain) 6426 goto error; 6427 6428 /* Set up domains for cpus specified by the cpu_map. */ 6429 for_each_cpu(i, cpu_map) { 6430 struct sched_domain_topology_level *tl; 6431 6432 sd = NULL; 6433 for_each_sd_topology(tl) { 6434 sd = build_sched_domain(tl, cpu_map, attr, sd, i); 6435 if (tl == sched_domain_topology) 6436 *per_cpu_ptr(d.sd, i) = sd; 6437 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP)) 6438 sd->flags |= SD_OVERLAP; 6439 if (cpumask_equal(cpu_map, sched_domain_span(sd))) 6440 break; 6441 } 6442 } 6443 6444 /* Build the groups for the domains */ 6445 for_each_cpu(i, cpu_map) { 6446 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) { 6447 sd->span_weight = cpumask_weight(sched_domain_span(sd)); 6448 if (sd->flags & SD_OVERLAP) { 6449 if (build_overlap_sched_groups(sd, i)) 6450 goto error; 6451 } else { 6452 if (build_sched_groups(sd, i)) 6453 goto error; 6454 } 6455 } 6456 } 6457 6458 /* Calculate CPU power for physical packages and nodes */ 6459 for (i = nr_cpumask_bits-1; i >= 0; i--) { 6460 if (!cpumask_test_cpu(i, cpu_map)) 6461 continue; 6462 6463 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) { 6464 claim_allocations(i, sd); 6465 init_sched_groups_power(i, sd); 6466 } 6467 } 6468 6469 /* Attach the domains */ 6470 rcu_read_lock(); 6471 for_each_cpu(i, cpu_map) { 6472 sd = *per_cpu_ptr(d.sd, i); 6473 cpu_attach_domain(sd, d.rd, i); 6474 } 6475 rcu_read_unlock(); 6476 6477 ret = 0; 6478error: 6479 __free_domain_allocs(&d, alloc_state, cpu_map); 6480 return ret; 6481} 6482 6483static cpumask_var_t *doms_cur; /* current sched domains */ 6484static int ndoms_cur; /* number of sched domains in 'doms_cur' */ 6485static struct sched_domain_attr *dattr_cur; 6486 /* attribues of custom domains in 'doms_cur' */ 6487 6488/* 6489 * Special case: If a kmalloc of a doms_cur partition (array of 6490 * cpumask) fails, then fallback to a single sched domain, 6491 * as determined by the single cpumask fallback_doms. 6492 */ 6493static cpumask_var_t fallback_doms; 6494 6495/* 6496 * arch_update_cpu_topology lets virtualized architectures update the 6497 * cpu core maps. It is supposed to return 1 if the topology changed 6498 * or 0 if it stayed the same. 6499 */ 6500int __weak arch_update_cpu_topology(void) 6501{ 6502 return 0; 6503} 6504 6505cpumask_var_t *alloc_sched_domains(unsigned int ndoms) 6506{ 6507 int i; 6508 cpumask_var_t *doms; 6509 6510 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL); 6511 if (!doms) 6512 return NULL; 6513 for (i = 0; i < ndoms; i++) { 6514 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) { 6515 free_sched_domains(doms, i); 6516 return NULL; 6517 } 6518 } 6519 return doms; 6520} 6521 6522void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms) 6523{ 6524 unsigned int i; 6525 for (i = 0; i < ndoms; i++) 6526 free_cpumask_var(doms[i]); 6527 kfree(doms); 6528} 6529 6530/* 6531 * Set up scheduler domains and groups. Callers must hold the hotplug lock. 6532 * For now this just excludes isolated cpus, but could be used to 6533 * exclude other special cases in the future. 6534 */ 6535static int init_sched_domains(const struct cpumask *cpu_map) 6536{ 6537 int err; 6538 6539 arch_update_cpu_topology(); 6540 ndoms_cur = 1; 6541 doms_cur = alloc_sched_domains(ndoms_cur); 6542 if (!doms_cur) 6543 doms_cur = &fallback_doms; 6544 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map); 6545 err = build_sched_domains(doms_cur[0], NULL); 6546 register_sched_domain_sysctl(); 6547 6548 return err; 6549} 6550 6551/* 6552 * Detach sched domains from a group of cpus specified in cpu_map 6553 * These cpus will now be attached to the NULL domain 6554 */ 6555static void detach_destroy_domains(const struct cpumask *cpu_map) 6556{ 6557 int i; 6558 6559 rcu_read_lock(); 6560 for_each_cpu(i, cpu_map) 6561 cpu_attach_domain(NULL, &def_root_domain, i); 6562 rcu_read_unlock(); 6563} 6564 6565/* handle null as "default" */ 6566static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur, 6567 struct sched_domain_attr *new, int idx_new) 6568{ 6569 struct sched_domain_attr tmp; 6570 6571 /* fast path */ 6572 if (!new && !cur) 6573 return 1; 6574 6575 tmp = SD_ATTR_INIT; 6576 return !memcmp(cur ? (cur + idx_cur) : &tmp, 6577 new ? (new + idx_new) : &tmp, 6578 sizeof(struct sched_domain_attr)); 6579} 6580 6581/* 6582 * Partition sched domains as specified by the 'ndoms_new' 6583 * cpumasks in the array doms_new[] of cpumasks. This compares 6584 * doms_new[] to the current sched domain partitioning, doms_cur[]. 6585 * It destroys each deleted domain and builds each new domain. 6586 * 6587 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'. 6588 * The masks don't intersect (don't overlap.) We should setup one 6589 * sched domain for each mask. CPUs not in any of the cpumasks will 6590 * not be load balanced. If the same cpumask appears both in the 6591 * current 'doms_cur' domains and in the new 'doms_new', we can leave 6592 * it as it is. 6593 * 6594 * The passed in 'doms_new' should be allocated using 6595 * alloc_sched_domains. This routine takes ownership of it and will 6596 * free_sched_domains it when done with it. If the caller failed the 6597 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1, 6598 * and partition_sched_domains() will fallback to the single partition 6599 * 'fallback_doms', it also forces the domains to be rebuilt. 6600 * 6601 * If doms_new == NULL it will be replaced with cpu_online_mask. 6602 * ndoms_new == 0 is a special case for destroying existing domains, 6603 * and it will not create the default domain. 6604 * 6605 * Call with hotplug lock held 6606 */ 6607void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[], 6608 struct sched_domain_attr *dattr_new) 6609{ 6610 int i, j, n; 6611 int new_topology; 6612 6613 mutex_lock(&sched_domains_mutex); 6614 6615 /* always unregister in case we don't destroy any domains */ 6616 unregister_sched_domain_sysctl(); 6617 6618 /* Let architecture update cpu core mappings. */ 6619 new_topology = arch_update_cpu_topology(); 6620 6621 n = doms_new ? ndoms_new : 0; 6622 6623 /* Destroy deleted domains */ 6624 for (i = 0; i < ndoms_cur; i++) { 6625 for (j = 0; j < n && !new_topology; j++) { 6626 if (cpumask_equal(doms_cur[i], doms_new[j]) 6627 && dattrs_equal(dattr_cur, i, dattr_new, j)) 6628 goto match1; 6629 } 6630 /* no match - a current sched domain not in new doms_new[] */ 6631 detach_destroy_domains(doms_cur[i]); 6632match1: 6633 ; 6634 } 6635 6636 n = ndoms_cur; 6637 if (doms_new == NULL) { 6638 n = 0; 6639 doms_new = &fallback_doms; 6640 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map); 6641 WARN_ON_ONCE(dattr_new); 6642 } 6643 6644 /* Build new domains */ 6645 for (i = 0; i < ndoms_new; i++) { 6646 for (j = 0; j < n && !new_topology; j++) { 6647 if (cpumask_equal(doms_new[i], doms_cur[j]) 6648 && dattrs_equal(dattr_new, i, dattr_cur, j)) 6649 goto match2; 6650 } 6651 /* no match - add a new doms_new */ 6652 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL); 6653match2: 6654 ; 6655 } 6656 6657 /* Remember the new sched domains */ 6658 if (doms_cur != &fallback_doms) 6659 free_sched_domains(doms_cur, ndoms_cur); 6660 kfree(dattr_cur); /* kfree(NULL) is safe */ 6661 doms_cur = doms_new; 6662 dattr_cur = dattr_new; 6663 ndoms_cur = ndoms_new; 6664 6665 register_sched_domain_sysctl(); 6666 6667 mutex_unlock(&sched_domains_mutex); 6668} 6669 6670static int num_cpus_frozen; /* used to mark begin/end of suspend/resume */ 6671 6672/* 6673 * Update cpusets according to cpu_active mask. If cpusets are 6674 * disabled, cpuset_update_active_cpus() becomes a simple wrapper 6675 * around partition_sched_domains(). 6676 * 6677 * If we come here as part of a suspend/resume, don't touch cpusets because we 6678 * want to restore it back to its original state upon resume anyway. 6679 */ 6680static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action, 6681 void *hcpu) 6682{ 6683 switch (action) { 6684 case CPU_ONLINE_FROZEN: 6685 case CPU_DOWN_FAILED_FROZEN: 6686 6687 /* 6688 * num_cpus_frozen tracks how many CPUs are involved in suspend 6689 * resume sequence. As long as this is not the last online 6690 * operation in the resume sequence, just build a single sched 6691 * domain, ignoring cpusets. 6692 */ 6693 num_cpus_frozen--; 6694 if (likely(num_cpus_frozen)) { 6695 partition_sched_domains(1, NULL, NULL); 6696 break; 6697 } 6698 6699 /* 6700 * This is the last CPU online operation. So fall through and 6701 * restore the original sched domains by considering the 6702 * cpuset configurations. 6703 */ 6704 6705 case CPU_ONLINE: 6706 case CPU_DOWN_FAILED: 6707 cpuset_update_active_cpus(true); 6708 break; 6709 default: 6710 return NOTIFY_DONE; 6711 } 6712 return NOTIFY_OK; 6713} 6714 6715static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action, 6716 void *hcpu) 6717{ 6718 switch (action) { 6719 case CPU_DOWN_PREPARE: 6720 cpuset_update_active_cpus(false); 6721 break; 6722 case CPU_DOWN_PREPARE_FROZEN: 6723 num_cpus_frozen++; 6724 partition_sched_domains(1, NULL, NULL); 6725 break; 6726 default: 6727 return NOTIFY_DONE; 6728 } 6729 return NOTIFY_OK; 6730} 6731 6732void __init sched_init_smp(void) 6733{ 6734 cpumask_var_t non_isolated_cpus; 6735 6736 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL); 6737 alloc_cpumask_var(&fallback_doms, GFP_KERNEL); 6738 6739 sched_init_numa(); 6740 6741 /* 6742 * There's no userspace yet to cause hotplug operations; hence all the 6743 * cpu masks are stable and all blatant races in the below code cannot 6744 * happen. 6745 */ 6746 mutex_lock(&sched_domains_mutex); 6747 init_sched_domains(cpu_active_mask); 6748 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map); 6749 if (cpumask_empty(non_isolated_cpus)) 6750 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus); 6751 mutex_unlock(&sched_domains_mutex); 6752 6753 hotcpu_notifier(sched_domains_numa_masks_update, CPU_PRI_SCHED_ACTIVE); 6754 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE); 6755 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE); 6756 6757 init_hrtick(); 6758 6759 /* Move init over to a non-isolated CPU */ 6760 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0) 6761 BUG(); 6762 sched_init_granularity(); 6763 free_cpumask_var(non_isolated_cpus); 6764 6765 init_sched_rt_class(); 6766 init_sched_dl_class(); 6767} 6768#else 6769void __init sched_init_smp(void) 6770{ 6771 sched_init_granularity(); 6772} 6773#endif /* CONFIG_SMP */ 6774 6775const_debug unsigned int sysctl_timer_migration = 1; 6776 6777int in_sched_functions(unsigned long addr) 6778{ 6779 return in_lock_functions(addr) || 6780 (addr >= (unsigned long)__sched_text_start 6781 && addr < (unsigned long)__sched_text_end); 6782} 6783 6784#ifdef CONFIG_CGROUP_SCHED 6785/* 6786 * Default task group. 6787 * Every task in system belongs to this group at bootup. 6788 */ 6789struct task_group root_task_group; 6790LIST_HEAD(task_groups); 6791#endif 6792 6793DECLARE_PER_CPU(cpumask_var_t, load_balance_mask); 6794 6795void __init sched_init(void) 6796{ 6797 int i, j; 6798 unsigned long alloc_size = 0, ptr; 6799 6800#ifdef CONFIG_FAIR_GROUP_SCHED 6801 alloc_size += 2 * nr_cpu_ids * sizeof(void **); 6802#endif 6803#ifdef CONFIG_RT_GROUP_SCHED 6804 alloc_size += 2 * nr_cpu_ids * sizeof(void **); 6805#endif 6806#ifdef CONFIG_CPUMASK_OFFSTACK 6807 alloc_size += num_possible_cpus() * cpumask_size(); 6808#endif 6809 if (alloc_size) { 6810 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT); 6811 6812#ifdef CONFIG_FAIR_GROUP_SCHED 6813 root_task_group.se = (struct sched_entity **)ptr; 6814 ptr += nr_cpu_ids * sizeof(void **); 6815 6816 root_task_group.cfs_rq = (struct cfs_rq **)ptr; 6817 ptr += nr_cpu_ids * sizeof(void **); 6818 6819#endif /* CONFIG_FAIR_GROUP_SCHED */ 6820#ifdef CONFIG_RT_GROUP_SCHED 6821 root_task_group.rt_se = (struct sched_rt_entity **)ptr; 6822 ptr += nr_cpu_ids * sizeof(void **); 6823 6824 root_task_group.rt_rq = (struct rt_rq **)ptr; 6825 ptr += nr_cpu_ids * sizeof(void **); 6826 6827#endif /* CONFIG_RT_GROUP_SCHED */ 6828#ifdef CONFIG_CPUMASK_OFFSTACK 6829 for_each_possible_cpu(i) { 6830 per_cpu(load_balance_mask, i) = (void *)ptr; 6831 ptr += cpumask_size(); 6832 } 6833#endif /* CONFIG_CPUMASK_OFFSTACK */ 6834 } 6835 6836 init_rt_bandwidth(&def_rt_bandwidth, 6837 global_rt_period(), global_rt_runtime()); 6838 init_dl_bandwidth(&def_dl_bandwidth, 6839 global_rt_period(), global_rt_runtime()); 6840 6841#ifdef CONFIG_SMP 6842 init_defrootdomain(); 6843#endif 6844 6845#ifdef CONFIG_RT_GROUP_SCHED 6846 init_rt_bandwidth(&root_task_group.rt_bandwidth, 6847 global_rt_period(), global_rt_runtime()); 6848#endif /* CONFIG_RT_GROUP_SCHED */ 6849 6850#ifdef CONFIG_CGROUP_SCHED 6851 list_add(&root_task_group.list, &task_groups); 6852 INIT_LIST_HEAD(&root_task_group.children); 6853 INIT_LIST_HEAD(&root_task_group.siblings); 6854 autogroup_init(&init_task); 6855 6856#endif /* CONFIG_CGROUP_SCHED */ 6857 6858 for_each_possible_cpu(i) { 6859 struct rq *rq; 6860 6861 rq = cpu_rq(i); 6862 raw_spin_lock_init(&rq->lock); 6863 rq->nr_running = 0; 6864 rq->calc_load_active = 0; 6865 rq->calc_load_update = jiffies + LOAD_FREQ; 6866 init_cfs_rq(&rq->cfs); 6867 init_rt_rq(&rq->rt, rq); 6868 init_dl_rq(&rq->dl, rq); 6869#ifdef CONFIG_FAIR_GROUP_SCHED 6870 root_task_group.shares = ROOT_TASK_GROUP_LOAD; 6871 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list); 6872 /* 6873 * How much cpu bandwidth does root_task_group get? 6874 * 6875 * In case of task-groups formed thr' the cgroup filesystem, it 6876 * gets 100% of the cpu resources in the system. This overall 6877 * system cpu resource is divided among the tasks of 6878 * root_task_group and its child task-groups in a fair manner, 6879 * based on each entity's (task or task-group's) weight 6880 * (se->load.weight). 6881 * 6882 * In other words, if root_task_group has 10 tasks of weight 6883 * 1024) and two child groups A0 and A1 (of weight 1024 each), 6884 * then A0's share of the cpu resource is: 6885 * 6886 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33% 6887 * 6888 * We achieve this by letting root_task_group's tasks sit 6889 * directly in rq->cfs (i.e root_task_group->se[] = NULL). 6890 */ 6891 init_cfs_bandwidth(&root_task_group.cfs_bandwidth); 6892 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL); 6893#endif /* CONFIG_FAIR_GROUP_SCHED */ 6894 6895 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime; 6896#ifdef CONFIG_RT_GROUP_SCHED 6897 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL); 6898#endif 6899 6900 for (j = 0; j < CPU_LOAD_IDX_MAX; j++) 6901 rq->cpu_load[j] = 0; 6902 6903 rq->last_load_update_tick = jiffies; 6904 6905#ifdef CONFIG_SMP 6906 rq->sd = NULL; 6907 rq->rd = NULL; 6908 rq->cpu_power = SCHED_POWER_SCALE; 6909 rq->post_schedule = 0; 6910 rq->active_balance = 0; 6911 rq->next_balance = jiffies; 6912 rq->push_cpu = 0; 6913 rq->cpu = i; 6914 rq->online = 0; 6915 rq->idle_stamp = 0; 6916 rq->avg_idle = 2*sysctl_sched_migration_cost; 6917 rq->max_idle_balance_cost = sysctl_sched_migration_cost; 6918 6919 INIT_LIST_HEAD(&rq->cfs_tasks); 6920 6921 rq_attach_root(rq, &def_root_domain); 6922#ifdef CONFIG_NO_HZ_COMMON 6923 rq->nohz_flags = 0; 6924#endif 6925#ifdef CONFIG_NO_HZ_FULL 6926 rq->last_sched_tick = 0; 6927#endif 6928#endif 6929 init_rq_hrtick(rq); 6930 atomic_set(&rq->nr_iowait, 0); 6931 } 6932 6933 set_load_weight(&init_task); 6934 6935#ifdef CONFIG_PREEMPT_NOTIFIERS 6936 INIT_HLIST_HEAD(&init_task.preempt_notifiers); 6937#endif 6938 6939 /* 6940 * The boot idle thread does lazy MMU switching as well: 6941 */ 6942 atomic_inc(&init_mm.mm_count); 6943 enter_lazy_tlb(&init_mm, current); 6944 6945 /* 6946 * Make us the idle thread. Technically, schedule() should not be 6947 * called from this thread, however somewhere below it might be, 6948 * but because we are the idle thread, we just pick up running again 6949 * when this runqueue becomes "idle". 6950 */ 6951 init_idle(current, smp_processor_id()); 6952 6953 calc_load_update = jiffies + LOAD_FREQ; 6954 6955 /* 6956 * During early bootup we pretend to be a normal task: 6957 */ 6958 current->sched_class = &fair_sched_class; 6959 6960#ifdef CONFIG_SMP 6961 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT); 6962 /* May be allocated at isolcpus cmdline parse time */ 6963 if (cpu_isolated_map == NULL) 6964 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT); 6965 idle_thread_set_boot_cpu(); 6966#endif 6967 init_sched_fair_class(); 6968 6969 scheduler_running = 1; 6970} 6971 6972#ifdef CONFIG_DEBUG_ATOMIC_SLEEP 6973static inline int preempt_count_equals(int preempt_offset) 6974{ 6975 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth(); 6976 6977 return (nested == preempt_offset); 6978} 6979 6980void __might_sleep(const char *file, int line, int preempt_offset) 6981{ 6982 static unsigned long prev_jiffy; /* ratelimiting */ 6983 6984 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */ 6985 if ((preempt_count_equals(preempt_offset) && !irqs_disabled() && 6986 !is_idle_task(current)) || 6987 system_state != SYSTEM_RUNNING || oops_in_progress) 6988 return; 6989 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy) 6990 return; 6991 prev_jiffy = jiffies; 6992 6993 printk(KERN_ERR 6994 "BUG: sleeping function called from invalid context at %s:%d\n", 6995 file, line); 6996 printk(KERN_ERR 6997 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n", 6998 in_atomic(), irqs_disabled(), 6999 current->pid, current->comm); 7000 7001 debug_show_held_locks(current); 7002 if (irqs_disabled()) 7003 print_irqtrace_events(current); 7004#ifdef CONFIG_DEBUG_PREEMPT 7005 if (!preempt_count_equals(preempt_offset)) { 7006 pr_err("Preemption disabled at:"); 7007 print_ip_sym(current->preempt_disable_ip); 7008 pr_cont("\n"); 7009 } 7010#endif 7011 dump_stack(); 7012} 7013EXPORT_SYMBOL(__might_sleep); 7014#endif 7015 7016#ifdef CONFIG_MAGIC_SYSRQ 7017static void normalize_task(struct rq *rq, struct task_struct *p) 7018{ 7019 const struct sched_class *prev_class = p->sched_class; 7020 struct sched_attr attr = { 7021 .sched_policy = SCHED_NORMAL, 7022 }; 7023 int old_prio = p->prio; 7024 int on_rq; 7025 7026 on_rq = p->on_rq; 7027 if (on_rq) 7028 dequeue_task(rq, p, 0); 7029 __setscheduler(rq, p, &attr); 7030 if (on_rq) { 7031 enqueue_task(rq, p, 0); 7032 resched_task(rq->curr); 7033 } 7034 7035 check_class_changed(rq, p, prev_class, old_prio); 7036} 7037 7038void normalize_rt_tasks(void) 7039{ 7040 struct task_struct *g, *p; 7041 unsigned long flags; 7042 struct rq *rq; 7043 7044 read_lock_irqsave(&tasklist_lock, flags); 7045 do_each_thread(g, p) { 7046 /* 7047 * Only normalize user tasks: 7048 */ 7049 if (!p->mm) 7050 continue; 7051 7052 p->se.exec_start = 0; 7053#ifdef CONFIG_SCHEDSTATS 7054 p->se.statistics.wait_start = 0; 7055 p->se.statistics.sleep_start = 0; 7056 p->se.statistics.block_start = 0; 7057#endif 7058 7059 if (!dl_task(p) && !rt_task(p)) { 7060 /* 7061 * Renice negative nice level userspace 7062 * tasks back to 0: 7063 */ 7064 if (task_nice(p) < 0 && p->mm) 7065 set_user_nice(p, 0); 7066 continue; 7067 } 7068 7069 raw_spin_lock(&p->pi_lock); 7070 rq = __task_rq_lock(p); 7071 7072 normalize_task(rq, p); 7073 7074 __task_rq_unlock(rq); 7075 raw_spin_unlock(&p->pi_lock); 7076 } while_each_thread(g, p); 7077 7078 read_unlock_irqrestore(&tasklist_lock, flags); 7079} 7080 7081#endif /* CONFIG_MAGIC_SYSRQ */ 7082 7083#if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) 7084/* 7085 * These functions are only useful for the IA64 MCA handling, or kdb. 7086 * 7087 * They can only be called when the whole system has been 7088 * stopped - every CPU needs to be quiescent, and no scheduling 7089 * activity can take place. Using them for anything else would 7090 * be a serious bug, and as a result, they aren't even visible 7091 * under any other configuration. 7092 */ 7093 7094/** 7095 * curr_task - return the current task for a given cpu. 7096 * @cpu: the processor in question. 7097 * 7098 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED! 7099 * 7100 * Return: The current task for @cpu. 7101 */ 7102struct task_struct *curr_task(int cpu) 7103{ 7104 return cpu_curr(cpu); 7105} 7106 7107#endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */ 7108 7109#ifdef CONFIG_IA64 7110/** 7111 * set_curr_task - set the current task for a given cpu. 7112 * @cpu: the processor in question. 7113 * @p: the task pointer to set. 7114 * 7115 * Description: This function must only be used when non-maskable interrupts 7116 * are serviced on a separate stack. It allows the architecture to switch the 7117 * notion of the current task on a cpu in a non-blocking manner. This function 7118 * must be called with all CPU's synchronized, and interrupts disabled, the 7119 * and caller must save the original value of the current task (see 7120 * curr_task() above) and restore that value before reenabling interrupts and 7121 * re-starting the system. 7122 * 7123 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED! 7124 */ 7125void set_curr_task(int cpu, struct task_struct *p) 7126{ 7127 cpu_curr(cpu) = p; 7128} 7129 7130#endif 7131 7132#ifdef CONFIG_CGROUP_SCHED 7133/* task_group_lock serializes the addition/removal of task groups */ 7134static DEFINE_SPINLOCK(task_group_lock); 7135 7136static void free_sched_group(struct task_group *tg) 7137{ 7138 free_fair_sched_group(tg); 7139 free_rt_sched_group(tg); 7140 autogroup_free(tg); 7141 kfree(tg); 7142} 7143 7144/* allocate runqueue etc for a new task group */ 7145struct task_group *sched_create_group(struct task_group *parent) 7146{ 7147 struct task_group *tg; 7148 7149 tg = kzalloc(sizeof(*tg), GFP_KERNEL); 7150 if (!tg) 7151 return ERR_PTR(-ENOMEM); 7152 7153 if (!alloc_fair_sched_group(tg, parent)) 7154 goto err; 7155 7156 if (!alloc_rt_sched_group(tg, parent)) 7157 goto err; 7158 7159 return tg; 7160 7161err: 7162 free_sched_group(tg); 7163 return ERR_PTR(-ENOMEM); 7164} 7165 7166void sched_online_group(struct task_group *tg, struct task_group *parent) 7167{ 7168 unsigned long flags; 7169 7170 spin_lock_irqsave(&task_group_lock, flags); 7171 list_add_rcu(&tg->list, &task_groups); 7172 7173 WARN_ON(!parent); /* root should already exist */ 7174 7175 tg->parent = parent; 7176 INIT_LIST_HEAD(&tg->children); 7177 list_add_rcu(&tg->siblings, &parent->children); 7178 spin_unlock_irqrestore(&task_group_lock, flags); 7179} 7180 7181/* rcu callback to free various structures associated with a task group */ 7182static void free_sched_group_rcu(struct rcu_head *rhp) 7183{ 7184 /* now it should be safe to free those cfs_rqs */ 7185 free_sched_group(container_of(rhp, struct task_group, rcu)); 7186} 7187 7188/* Destroy runqueue etc associated with a task group */ 7189void sched_destroy_group(struct task_group *tg) 7190{ 7191 /* wait for possible concurrent references to cfs_rqs complete */ 7192 call_rcu(&tg->rcu, free_sched_group_rcu); 7193} 7194 7195void sched_offline_group(struct task_group *tg) 7196{ 7197 unsigned long flags; 7198 int i; 7199 7200 /* end participation in shares distribution */ 7201 for_each_possible_cpu(i) 7202 unregister_fair_sched_group(tg, i); 7203 7204 spin_lock_irqsave(&task_group_lock, flags); 7205 list_del_rcu(&tg->list); 7206 list_del_rcu(&tg->siblings); 7207 spin_unlock_irqrestore(&task_group_lock, flags); 7208} 7209 7210/* change task's runqueue when it moves between groups. 7211 * The caller of this function should have put the task in its new group 7212 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to 7213 * reflect its new group. 7214 */ 7215void sched_move_task(struct task_struct *tsk) 7216{ 7217 struct task_group *tg; 7218 int on_rq, running; 7219 unsigned long flags; 7220 struct rq *rq; 7221 7222 rq = task_rq_lock(tsk, &flags); 7223 7224 running = task_current(rq, tsk); 7225 on_rq = tsk->on_rq; 7226 7227 if (on_rq) 7228 dequeue_task(rq, tsk, 0); 7229 if (unlikely(running)) 7230 tsk->sched_class->put_prev_task(rq, tsk); 7231 7232 tg = container_of(task_css_check(tsk, cpu_cgrp_id, 7233 lockdep_is_held(&tsk->sighand->siglock)), 7234 struct task_group, css); 7235 tg = autogroup_task_group(tsk, tg); 7236 tsk->sched_task_group = tg; 7237 7238#ifdef CONFIG_FAIR_GROUP_SCHED 7239 if (tsk->sched_class->task_move_group) 7240 tsk->sched_class->task_move_group(tsk, on_rq); 7241 else 7242#endif 7243 set_task_rq(tsk, task_cpu(tsk)); 7244 7245 if (unlikely(running)) 7246 tsk->sched_class->set_curr_task(rq); 7247 if (on_rq) 7248 enqueue_task(rq, tsk, 0); 7249 7250 task_rq_unlock(rq, tsk, &flags); 7251} 7252#endif /* CONFIG_CGROUP_SCHED */ 7253 7254#ifdef CONFIG_RT_GROUP_SCHED 7255/* 7256 * Ensure that the real time constraints are schedulable. 7257 */ 7258static DEFINE_MUTEX(rt_constraints_mutex); 7259 7260/* Must be called with tasklist_lock held */ 7261static inline int tg_has_rt_tasks(struct task_group *tg) 7262{ 7263 struct task_struct *g, *p; 7264 7265 do_each_thread(g, p) { 7266 if (rt_task(p) && task_rq(p)->rt.tg == tg) 7267 return 1; 7268 } while_each_thread(g, p); 7269 7270 return 0; 7271} 7272 7273struct rt_schedulable_data { 7274 struct task_group *tg; 7275 u64 rt_period; 7276 u64 rt_runtime; 7277}; 7278 7279static int tg_rt_schedulable(struct task_group *tg, void *data) 7280{ 7281 struct rt_schedulable_data *d = data; 7282 struct task_group *child; 7283 unsigned long total, sum = 0; 7284 u64 period, runtime; 7285 7286 period = ktime_to_ns(tg->rt_bandwidth.rt_period); 7287 runtime = tg->rt_bandwidth.rt_runtime; 7288 7289 if (tg == d->tg) { 7290 period = d->rt_period; 7291 runtime = d->rt_runtime; 7292 } 7293 7294 /* 7295 * Cannot have more runtime than the period. 7296 */ 7297 if (runtime > period && runtime != RUNTIME_INF) 7298 return -EINVAL; 7299 7300 /* 7301 * Ensure we don't starve existing RT tasks. 7302 */ 7303 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg)) 7304 return -EBUSY; 7305 7306 total = to_ratio(period, runtime); 7307 7308 /* 7309 * Nobody can have more than the global setting allows. 7310 */ 7311 if (total > to_ratio(global_rt_period(), global_rt_runtime())) 7312 return -EINVAL; 7313 7314 /* 7315 * The sum of our children's runtime should not exceed our own. 7316 */ 7317 list_for_each_entry_rcu(child, &tg->children, siblings) { 7318 period = ktime_to_ns(child->rt_bandwidth.rt_period); 7319 runtime = child->rt_bandwidth.rt_runtime; 7320 7321 if (child == d->tg) { 7322 period = d->rt_period; 7323 runtime = d->rt_runtime; 7324 } 7325 7326 sum += to_ratio(period, runtime); 7327 } 7328 7329 if (sum > total) 7330 return -EINVAL; 7331 7332 return 0; 7333} 7334 7335static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime) 7336{ 7337 int ret; 7338 7339 struct rt_schedulable_data data = { 7340 .tg = tg, 7341 .rt_period = period, 7342 .rt_runtime = runtime, 7343 }; 7344 7345 rcu_read_lock(); 7346 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data); 7347 rcu_read_unlock(); 7348 7349 return ret; 7350} 7351 7352static int tg_set_rt_bandwidth(struct task_group *tg, 7353 u64 rt_period, u64 rt_runtime) 7354{ 7355 int i, err = 0; 7356 7357 mutex_lock(&rt_constraints_mutex); 7358 read_lock(&tasklist_lock); 7359 err = __rt_schedulable(tg, rt_period, rt_runtime); 7360 if (err) 7361 goto unlock; 7362 7363 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock); 7364 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period); 7365 tg->rt_bandwidth.rt_runtime = rt_runtime; 7366 7367 for_each_possible_cpu(i) { 7368 struct rt_rq *rt_rq = tg->rt_rq[i]; 7369 7370 raw_spin_lock(&rt_rq->rt_runtime_lock); 7371 rt_rq->rt_runtime = rt_runtime; 7372 raw_spin_unlock(&rt_rq->rt_runtime_lock); 7373 } 7374 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock); 7375unlock: 7376 read_unlock(&tasklist_lock); 7377 mutex_unlock(&rt_constraints_mutex); 7378 7379 return err; 7380} 7381 7382static int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us) 7383{ 7384 u64 rt_runtime, rt_period; 7385 7386 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period); 7387 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC; 7388 if (rt_runtime_us < 0) 7389 rt_runtime = RUNTIME_INF; 7390 7391 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime); 7392} 7393 7394static long sched_group_rt_runtime(struct task_group *tg) 7395{ 7396 u64 rt_runtime_us; 7397 7398 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF) 7399 return -1; 7400 7401 rt_runtime_us = tg->rt_bandwidth.rt_runtime; 7402 do_div(rt_runtime_us, NSEC_PER_USEC); 7403 return rt_runtime_us; 7404} 7405 7406static int sched_group_set_rt_period(struct task_group *tg, long rt_period_us) 7407{ 7408 u64 rt_runtime, rt_period; 7409 7410 rt_period = (u64)rt_period_us * NSEC_PER_USEC; 7411 rt_runtime = tg->rt_bandwidth.rt_runtime; 7412 7413 if (rt_period == 0) 7414 return -EINVAL; 7415 7416 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime); 7417} 7418 7419static long sched_group_rt_period(struct task_group *tg) 7420{ 7421 u64 rt_period_us; 7422 7423 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period); 7424 do_div(rt_period_us, NSEC_PER_USEC); 7425 return rt_period_us; 7426} 7427#endif /* CONFIG_RT_GROUP_SCHED */ 7428 7429#ifdef CONFIG_RT_GROUP_SCHED 7430static int sched_rt_global_constraints(void) 7431{ 7432 int ret = 0; 7433 7434 mutex_lock(&rt_constraints_mutex); 7435 read_lock(&tasklist_lock); 7436 ret = __rt_schedulable(NULL, 0, 0); 7437 read_unlock(&tasklist_lock); 7438 mutex_unlock(&rt_constraints_mutex); 7439 7440 return ret; 7441} 7442 7443static int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk) 7444{ 7445 /* Don't accept realtime tasks when there is no way for them to run */ 7446 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0) 7447 return 0; 7448 7449 return 1; 7450} 7451 7452#else /* !CONFIG_RT_GROUP_SCHED */ 7453static int sched_rt_global_constraints(void) 7454{ 7455 unsigned long flags; 7456 int i, ret = 0; 7457 7458 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags); 7459 for_each_possible_cpu(i) { 7460 struct rt_rq *rt_rq = &cpu_rq(i)->rt; 7461 7462 raw_spin_lock(&rt_rq->rt_runtime_lock); 7463 rt_rq->rt_runtime = global_rt_runtime(); 7464 raw_spin_unlock(&rt_rq->rt_runtime_lock); 7465 } 7466 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags); 7467 7468 return ret; 7469} 7470#endif /* CONFIG_RT_GROUP_SCHED */ 7471 7472static int sched_dl_global_constraints(void) 7473{ 7474 u64 runtime = global_rt_runtime(); 7475 u64 period = global_rt_period(); 7476 u64 new_bw = to_ratio(period, runtime); 7477 int cpu, ret = 0; 7478 unsigned long flags; 7479 7480 /* 7481 * Here we want to check the bandwidth not being set to some 7482 * value smaller than the currently allocated bandwidth in 7483 * any of the root_domains. 7484 * 7485 * FIXME: Cycling on all the CPUs is overdoing, but simpler than 7486 * cycling on root_domains... Discussion on different/better 7487 * solutions is welcome! 7488 */ 7489 for_each_possible_cpu(cpu) { 7490 struct dl_bw *dl_b = dl_bw_of(cpu); 7491 7492 raw_spin_lock_irqsave(&dl_b->lock, flags); 7493 if (new_bw < dl_b->total_bw) 7494 ret = -EBUSY; 7495 raw_spin_unlock_irqrestore(&dl_b->lock, flags); 7496 7497 if (ret) 7498 break; 7499 } 7500 7501 return ret; 7502} 7503 7504static void sched_dl_do_global(void) 7505{ 7506 u64 new_bw = -1; 7507 int cpu; 7508 unsigned long flags; 7509 7510 def_dl_bandwidth.dl_period = global_rt_period(); 7511 def_dl_bandwidth.dl_runtime = global_rt_runtime(); 7512 7513 if (global_rt_runtime() != RUNTIME_INF) 7514 new_bw = to_ratio(global_rt_period(), global_rt_runtime()); 7515 7516 /* 7517 * FIXME: As above... 7518 */ 7519 for_each_possible_cpu(cpu) { 7520 struct dl_bw *dl_b = dl_bw_of(cpu); 7521 7522 raw_spin_lock_irqsave(&dl_b->lock, flags); 7523 dl_b->bw = new_bw; 7524 raw_spin_unlock_irqrestore(&dl_b->lock, flags); 7525 } 7526} 7527 7528static int sched_rt_global_validate(void) 7529{ 7530 if (sysctl_sched_rt_period <= 0) 7531 return -EINVAL; 7532 7533 if ((sysctl_sched_rt_runtime != RUNTIME_INF) && 7534 (sysctl_sched_rt_runtime > sysctl_sched_rt_period)) 7535 return -EINVAL; 7536 7537 return 0; 7538} 7539 7540static void sched_rt_do_global(void) 7541{ 7542 def_rt_bandwidth.rt_runtime = global_rt_runtime(); 7543 def_rt_bandwidth.rt_period = ns_to_ktime(global_rt_period()); 7544} 7545 7546int sched_rt_handler(struct ctl_table *table, int write, 7547 void __user *buffer, size_t *lenp, 7548 loff_t *ppos) 7549{ 7550 int old_period, old_runtime; 7551 static DEFINE_MUTEX(mutex); 7552 int ret; 7553 7554 mutex_lock(&mutex); 7555 old_period = sysctl_sched_rt_period; 7556 old_runtime = sysctl_sched_rt_runtime; 7557 7558 ret = proc_dointvec(table, write, buffer, lenp, ppos); 7559 7560 if (!ret && write) { 7561 ret = sched_rt_global_validate(); 7562 if (ret) 7563 goto undo; 7564 7565 ret = sched_rt_global_constraints(); 7566 if (ret) 7567 goto undo; 7568 7569 ret = sched_dl_global_constraints(); 7570 if (ret) 7571 goto undo; 7572 7573 sched_rt_do_global(); 7574 sched_dl_do_global(); 7575 } 7576 if (0) { 7577undo: 7578 sysctl_sched_rt_period = old_period; 7579 sysctl_sched_rt_runtime = old_runtime; 7580 } 7581 mutex_unlock(&mutex); 7582 7583 return ret; 7584} 7585 7586int sched_rr_handler(struct ctl_table *table, int write, 7587 void __user *buffer, size_t *lenp, 7588 loff_t *ppos) 7589{ 7590 int ret; 7591 static DEFINE_MUTEX(mutex); 7592 7593 mutex_lock(&mutex); 7594 ret = proc_dointvec(table, write, buffer, lenp, ppos); 7595 /* make sure that internally we keep jiffies */ 7596 /* also, writing zero resets timeslice to default */ 7597 if (!ret && write) { 7598 sched_rr_timeslice = sched_rr_timeslice <= 0 ? 7599 RR_TIMESLICE : msecs_to_jiffies(sched_rr_timeslice); 7600 } 7601 mutex_unlock(&mutex); 7602 return ret; 7603} 7604 7605#ifdef CONFIG_CGROUP_SCHED 7606 7607static inline struct task_group *css_tg(struct cgroup_subsys_state *css) 7608{ 7609 return css ? container_of(css, struct task_group, css) : NULL; 7610} 7611 7612static struct cgroup_subsys_state * 7613cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css) 7614{ 7615 struct task_group *parent = css_tg(parent_css); 7616 struct task_group *tg; 7617 7618 if (!parent) { 7619 /* This is early initialization for the top cgroup */ 7620 return &root_task_group.css; 7621 } 7622 7623 tg = sched_create_group(parent); 7624 if (IS_ERR(tg)) 7625 return ERR_PTR(-ENOMEM); 7626 7627 return &tg->css; 7628} 7629 7630static int cpu_cgroup_css_online(struct cgroup_subsys_state *css) 7631{ 7632 struct task_group *tg = css_tg(css); 7633 struct task_group *parent = css_tg(css_parent(css)); 7634 7635 if (parent) 7636 sched_online_group(tg, parent); 7637 return 0; 7638} 7639 7640static void cpu_cgroup_css_free(struct cgroup_subsys_state *css) 7641{ 7642 struct task_group *tg = css_tg(css); 7643 7644 sched_destroy_group(tg); 7645} 7646 7647static void cpu_cgroup_css_offline(struct cgroup_subsys_state *css) 7648{ 7649 struct task_group *tg = css_tg(css); 7650 7651 sched_offline_group(tg); 7652} 7653 7654static int cpu_cgroup_can_attach(struct cgroup_subsys_state *css, 7655 struct cgroup_taskset *tset) 7656{ 7657 struct task_struct *task; 7658 7659 cgroup_taskset_for_each(task, tset) { 7660#ifdef CONFIG_RT_GROUP_SCHED 7661 if (!sched_rt_can_attach(css_tg(css), task)) 7662 return -EINVAL; 7663#else 7664 /* We don't support RT-tasks being in separate groups */ 7665 if (task->sched_class != &fair_sched_class) 7666 return -EINVAL; 7667#endif 7668 } 7669 return 0; 7670} 7671 7672static void cpu_cgroup_attach(struct cgroup_subsys_state *css, 7673 struct cgroup_taskset *tset) 7674{ 7675 struct task_struct *task; 7676 7677 cgroup_taskset_for_each(task, tset) 7678 sched_move_task(task); 7679} 7680 7681static void cpu_cgroup_exit(struct cgroup_subsys_state *css, 7682 struct cgroup_subsys_state *old_css, 7683 struct task_struct *task) 7684{ 7685 /* 7686 * cgroup_exit() is called in the copy_process() failure path. 7687 * Ignore this case since the task hasn't ran yet, this avoids 7688 * trying to poke a half freed task state from generic code. 7689 */ 7690 if (!(task->flags & PF_EXITING)) 7691 return; 7692 7693 sched_move_task(task); 7694} 7695 7696#ifdef CONFIG_FAIR_GROUP_SCHED 7697static int cpu_shares_write_u64(struct cgroup_subsys_state *css, 7698 struct cftype *cftype, u64 shareval) 7699{ 7700 return sched_group_set_shares(css_tg(css), scale_load(shareval)); 7701} 7702 7703static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css, 7704 struct cftype *cft) 7705{ 7706 struct task_group *tg = css_tg(css); 7707 7708 return (u64) scale_load_down(tg->shares); 7709} 7710 7711#ifdef CONFIG_CFS_BANDWIDTH 7712static DEFINE_MUTEX(cfs_constraints_mutex); 7713 7714const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */ 7715const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */ 7716 7717static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime); 7718 7719static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota) 7720{ 7721 int i, ret = 0, runtime_enabled, runtime_was_enabled; 7722 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth; 7723 7724 if (tg == &root_task_group) 7725 return -EINVAL; 7726 7727 /* 7728 * Ensure we have at some amount of bandwidth every period. This is 7729 * to prevent reaching a state of large arrears when throttled via 7730 * entity_tick() resulting in prolonged exit starvation. 7731 */ 7732 if (quota < min_cfs_quota_period || period < min_cfs_quota_period) 7733 return -EINVAL; 7734 7735 /* 7736 * Likewise, bound things on the otherside by preventing insane quota 7737 * periods. This also allows us to normalize in computing quota 7738 * feasibility. 7739 */ 7740 if (period > max_cfs_quota_period) 7741 return -EINVAL; 7742 7743 mutex_lock(&cfs_constraints_mutex); 7744 ret = __cfs_schedulable(tg, period, quota); 7745 if (ret) 7746 goto out_unlock; 7747 7748 runtime_enabled = quota != RUNTIME_INF; 7749 runtime_was_enabled = cfs_b->quota != RUNTIME_INF; 7750 /* 7751 * If we need to toggle cfs_bandwidth_used, off->on must occur 7752 * before making related changes, and on->off must occur afterwards 7753 */ 7754 if (runtime_enabled && !runtime_was_enabled) 7755 cfs_bandwidth_usage_inc(); 7756 raw_spin_lock_irq(&cfs_b->lock); 7757 cfs_b->period = ns_to_ktime(period); 7758 cfs_b->quota = quota; 7759 7760 __refill_cfs_bandwidth_runtime(cfs_b); 7761 /* restart the period timer (if active) to handle new period expiry */ 7762 if (runtime_enabled && cfs_b->timer_active) { 7763 /* force a reprogram */ 7764 cfs_b->timer_active = 0; 7765 __start_cfs_bandwidth(cfs_b); 7766 } 7767 raw_spin_unlock_irq(&cfs_b->lock); 7768 7769 for_each_possible_cpu(i) { 7770 struct cfs_rq *cfs_rq = tg->cfs_rq[i]; 7771 struct rq *rq = cfs_rq->rq; 7772 7773 raw_spin_lock_irq(&rq->lock); 7774 cfs_rq->runtime_enabled = runtime_enabled; 7775 cfs_rq->runtime_remaining = 0; 7776 7777 if (cfs_rq->throttled) 7778 unthrottle_cfs_rq(cfs_rq); 7779 raw_spin_unlock_irq(&rq->lock); 7780 } 7781 if (runtime_was_enabled && !runtime_enabled) 7782 cfs_bandwidth_usage_dec(); 7783out_unlock: 7784 mutex_unlock(&cfs_constraints_mutex); 7785 7786 return ret; 7787} 7788 7789int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us) 7790{ 7791 u64 quota, period; 7792 7793 period = ktime_to_ns(tg->cfs_bandwidth.period); 7794 if (cfs_quota_us < 0) 7795 quota = RUNTIME_INF; 7796 else 7797 quota = (u64)cfs_quota_us * NSEC_PER_USEC; 7798 7799 return tg_set_cfs_bandwidth(tg, period, quota); 7800} 7801 7802long tg_get_cfs_quota(struct task_group *tg) 7803{ 7804 u64 quota_us; 7805 7806 if (tg->cfs_bandwidth.quota == RUNTIME_INF) 7807 return -1; 7808 7809 quota_us = tg->cfs_bandwidth.quota; 7810 do_div(quota_us, NSEC_PER_USEC); 7811 7812 return quota_us; 7813} 7814 7815int tg_set_cfs_period(struct task_group *tg, long cfs_period_us) 7816{ 7817 u64 quota, period; 7818 7819 period = (u64)cfs_period_us * NSEC_PER_USEC; 7820 quota = tg->cfs_bandwidth.quota; 7821 7822 return tg_set_cfs_bandwidth(tg, period, quota); 7823} 7824 7825long tg_get_cfs_period(struct task_group *tg) 7826{ 7827 u64 cfs_period_us; 7828 7829 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period); 7830 do_div(cfs_period_us, NSEC_PER_USEC); 7831 7832 return cfs_period_us; 7833} 7834 7835static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css, 7836 struct cftype *cft) 7837{ 7838 return tg_get_cfs_quota(css_tg(css)); 7839} 7840 7841static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css, 7842 struct cftype *cftype, s64 cfs_quota_us) 7843{ 7844 return tg_set_cfs_quota(css_tg(css), cfs_quota_us); 7845} 7846 7847static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css, 7848 struct cftype *cft) 7849{ 7850 return tg_get_cfs_period(css_tg(css)); 7851} 7852 7853static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css, 7854 struct cftype *cftype, u64 cfs_period_us) 7855{ 7856 return tg_set_cfs_period(css_tg(css), cfs_period_us); 7857} 7858 7859struct cfs_schedulable_data { 7860 struct task_group *tg; 7861 u64 period, quota; 7862}; 7863 7864/* 7865 * normalize group quota/period to be quota/max_period 7866 * note: units are usecs 7867 */ 7868static u64 normalize_cfs_quota(struct task_group *tg, 7869 struct cfs_schedulable_data *d) 7870{ 7871 u64 quota, period; 7872 7873 if (tg == d->tg) { 7874 period = d->period; 7875 quota = d->quota; 7876 } else { 7877 period = tg_get_cfs_period(tg); 7878 quota = tg_get_cfs_quota(tg); 7879 } 7880 7881 /* note: these should typically be equivalent */ 7882 if (quota == RUNTIME_INF || quota == -1) 7883 return RUNTIME_INF; 7884 7885 return to_ratio(period, quota); 7886} 7887 7888static int tg_cfs_schedulable_down(struct task_group *tg, void *data) 7889{ 7890 struct cfs_schedulable_data *d = data; 7891 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth; 7892 s64 quota = 0, parent_quota = -1; 7893 7894 if (!tg->parent) { 7895 quota = RUNTIME_INF; 7896 } else { 7897 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth; 7898 7899 quota = normalize_cfs_quota(tg, d); 7900 parent_quota = parent_b->hierarchal_quota; 7901 7902 /* 7903 * ensure max(child_quota) <= parent_quota, inherit when no 7904 * limit is set 7905 */ 7906 if (quota == RUNTIME_INF) 7907 quota = parent_quota; 7908 else if (parent_quota != RUNTIME_INF && quota > parent_quota) 7909 return -EINVAL; 7910 } 7911 cfs_b->hierarchal_quota = quota; 7912 7913 return 0; 7914} 7915 7916static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota) 7917{ 7918 int ret; 7919 struct cfs_schedulable_data data = { 7920 .tg = tg, 7921 .period = period, 7922 .quota = quota, 7923 }; 7924 7925 if (quota != RUNTIME_INF) { 7926 do_div(data.period, NSEC_PER_USEC); 7927 do_div(data.quota, NSEC_PER_USEC); 7928 } 7929 7930 rcu_read_lock(); 7931 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data); 7932 rcu_read_unlock(); 7933 7934 return ret; 7935} 7936 7937static int cpu_stats_show(struct seq_file *sf, void *v) 7938{ 7939 struct task_group *tg = css_tg(seq_css(sf)); 7940 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth; 7941 7942 seq_printf(sf, "nr_periods %d\n", cfs_b->nr_periods); 7943 seq_printf(sf, "nr_throttled %d\n", cfs_b->nr_throttled); 7944 seq_printf(sf, "throttled_time %llu\n", cfs_b->throttled_time); 7945 7946 return 0; 7947} 7948#endif /* CONFIG_CFS_BANDWIDTH */ 7949#endif /* CONFIG_FAIR_GROUP_SCHED */ 7950 7951#ifdef CONFIG_RT_GROUP_SCHED 7952static int cpu_rt_runtime_write(struct cgroup_subsys_state *css, 7953 struct cftype *cft, s64 val) 7954{ 7955 return sched_group_set_rt_runtime(css_tg(css), val); 7956} 7957 7958static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css, 7959 struct cftype *cft) 7960{ 7961 return sched_group_rt_runtime(css_tg(css)); 7962} 7963 7964static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css, 7965 struct cftype *cftype, u64 rt_period_us) 7966{ 7967 return sched_group_set_rt_period(css_tg(css), rt_period_us); 7968} 7969 7970static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css, 7971 struct cftype *cft) 7972{ 7973 return sched_group_rt_period(css_tg(css)); 7974} 7975#endif /* CONFIG_RT_GROUP_SCHED */ 7976 7977static struct cftype cpu_files[] = { 7978#ifdef CONFIG_FAIR_GROUP_SCHED 7979 { 7980 .name = "shares", 7981 .read_u64 = cpu_shares_read_u64, 7982 .write_u64 = cpu_shares_write_u64, 7983 }, 7984#endif 7985#ifdef CONFIG_CFS_BANDWIDTH 7986 { 7987 .name = "cfs_quota_us", 7988 .read_s64 = cpu_cfs_quota_read_s64, 7989 .write_s64 = cpu_cfs_quota_write_s64, 7990 }, 7991 { 7992 .name = "cfs_period_us", 7993 .read_u64 = cpu_cfs_period_read_u64, 7994 .write_u64 = cpu_cfs_period_write_u64, 7995 }, 7996 { 7997 .name = "stat", 7998 .seq_show = cpu_stats_show, 7999 }, 8000#endif 8001#ifdef CONFIG_RT_GROUP_SCHED 8002 { 8003 .name = "rt_runtime_us", 8004 .read_s64 = cpu_rt_runtime_read, 8005 .write_s64 = cpu_rt_runtime_write, 8006 }, 8007 { 8008 .name = "rt_period_us", 8009 .read_u64 = cpu_rt_period_read_uint, 8010 .write_u64 = cpu_rt_period_write_uint, 8011 }, 8012#endif 8013 { } /* terminate */ 8014}; 8015 8016struct cgroup_subsys cpu_cgrp_subsys = { 8017 .css_alloc = cpu_cgroup_css_alloc, 8018 .css_free = cpu_cgroup_css_free, 8019 .css_online = cpu_cgroup_css_online, 8020 .css_offline = cpu_cgroup_css_offline, 8021 .can_attach = cpu_cgroup_can_attach, 8022 .attach = cpu_cgroup_attach, 8023 .exit = cpu_cgroup_exit, 8024 .base_cftypes = cpu_files, 8025 .early_init = 1, 8026}; 8027 8028#endif /* CONFIG_CGROUP_SCHED */ 8029 8030void dump_cpu_task(int cpu) 8031{ 8032 pr_info("Task dump for CPU %d:\n", cpu); 8033 sched_show_task(cpu_curr(cpu)); 8034} 8035