core.c revision 7aa2c016db2162defff77f6f5731bff3f25e5175
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 __visible 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 __visible 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 __visible 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 __visible 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 __visible 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 = nice_to_rlimit(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 increment = clamp(increment, -NICE_WIDTH, NICE_WIDTH); 3061 nice = task_nice(current) + increment; 3062 3063 nice = clamp_val(nice, MIN_NICE, MAX_NICE); 3064 if (increment < 0 && !can_nice(current, nice)) 3065 return -EPERM; 3066 3067 retval = security_task_setnice(current, nice); 3068 if (retval) 3069 return retval; 3070 3071 set_user_nice(current, nice); 3072 return 0; 3073} 3074 3075#endif 3076 3077/** 3078 * task_prio - return the priority value of a given task. 3079 * @p: the task in question. 3080 * 3081 * Return: The priority value as seen by users in /proc. 3082 * RT tasks are offset by -200. Normal tasks are centered 3083 * around 0, value goes from -16 to +15. 3084 */ 3085int task_prio(const struct task_struct *p) 3086{ 3087 return p->prio - MAX_RT_PRIO; 3088} 3089 3090/** 3091 * idle_cpu - is a given cpu idle currently? 3092 * @cpu: the processor in question. 3093 * 3094 * Return: 1 if the CPU is currently idle. 0 otherwise. 3095 */ 3096int idle_cpu(int cpu) 3097{ 3098 struct rq *rq = cpu_rq(cpu); 3099 3100 if (rq->curr != rq->idle) 3101 return 0; 3102 3103 if (rq->nr_running) 3104 return 0; 3105 3106#ifdef CONFIG_SMP 3107 if (!llist_empty(&rq->wake_list)) 3108 return 0; 3109#endif 3110 3111 return 1; 3112} 3113 3114/** 3115 * idle_task - return the idle task for a given cpu. 3116 * @cpu: the processor in question. 3117 * 3118 * Return: The idle task for the cpu @cpu. 3119 */ 3120struct task_struct *idle_task(int cpu) 3121{ 3122 return cpu_rq(cpu)->idle; 3123} 3124 3125/** 3126 * find_process_by_pid - find a process with a matching PID value. 3127 * @pid: the pid in question. 3128 * 3129 * The task of @pid, if found. %NULL otherwise. 3130 */ 3131static struct task_struct *find_process_by_pid(pid_t pid) 3132{ 3133 return pid ? find_task_by_vpid(pid) : current; 3134} 3135 3136/* 3137 * This function initializes the sched_dl_entity of a newly becoming 3138 * SCHED_DEADLINE task. 3139 * 3140 * Only the static values are considered here, the actual runtime and the 3141 * absolute deadline will be properly calculated when the task is enqueued 3142 * for the first time with its new policy. 3143 */ 3144static void 3145__setparam_dl(struct task_struct *p, const struct sched_attr *attr) 3146{ 3147 struct sched_dl_entity *dl_se = &p->dl; 3148 3149 init_dl_task_timer(dl_se); 3150 dl_se->dl_runtime = attr->sched_runtime; 3151 dl_se->dl_deadline = attr->sched_deadline; 3152 dl_se->dl_period = attr->sched_period ?: dl_se->dl_deadline; 3153 dl_se->flags = attr->sched_flags; 3154 dl_se->dl_bw = to_ratio(dl_se->dl_period, dl_se->dl_runtime); 3155 dl_se->dl_throttled = 0; 3156 dl_se->dl_new = 1; 3157 dl_se->dl_yielded = 0; 3158} 3159 3160static void __setscheduler_params(struct task_struct *p, 3161 const struct sched_attr *attr) 3162{ 3163 int policy = attr->sched_policy; 3164 3165 if (policy == -1) /* setparam */ 3166 policy = p->policy; 3167 3168 p->policy = policy; 3169 3170 if (dl_policy(policy)) 3171 __setparam_dl(p, attr); 3172 else if (fair_policy(policy)) 3173 p->static_prio = NICE_TO_PRIO(attr->sched_nice); 3174 3175 /* 3176 * __sched_setscheduler() ensures attr->sched_priority == 0 when 3177 * !rt_policy. Always setting this ensures that things like 3178 * getparam()/getattr() don't report silly values for !rt tasks. 3179 */ 3180 p->rt_priority = attr->sched_priority; 3181 p->normal_prio = normal_prio(p); 3182 set_load_weight(p); 3183} 3184 3185/* Actually do priority change: must hold pi & rq lock. */ 3186static void __setscheduler(struct rq *rq, struct task_struct *p, 3187 const struct sched_attr *attr) 3188{ 3189 __setscheduler_params(p, attr); 3190 3191 /* 3192 * If we get here, there was no pi waiters boosting the 3193 * task. It is safe to use the normal prio. 3194 */ 3195 p->prio = normal_prio(p); 3196 3197 if (dl_prio(p->prio)) 3198 p->sched_class = &dl_sched_class; 3199 else if (rt_prio(p->prio)) 3200 p->sched_class = &rt_sched_class; 3201 else 3202 p->sched_class = &fair_sched_class; 3203} 3204 3205static void 3206__getparam_dl(struct task_struct *p, struct sched_attr *attr) 3207{ 3208 struct sched_dl_entity *dl_se = &p->dl; 3209 3210 attr->sched_priority = p->rt_priority; 3211 attr->sched_runtime = dl_se->dl_runtime; 3212 attr->sched_deadline = dl_se->dl_deadline; 3213 attr->sched_period = dl_se->dl_period; 3214 attr->sched_flags = dl_se->flags; 3215} 3216 3217/* 3218 * This function validates the new parameters of a -deadline task. 3219 * We ask for the deadline not being zero, and greater or equal 3220 * than the runtime, as well as the period of being zero or 3221 * greater than deadline. Furthermore, we have to be sure that 3222 * user parameters are above the internal resolution of 1us (we 3223 * check sched_runtime only since it is always the smaller one) and 3224 * below 2^63 ns (we have to check both sched_deadline and 3225 * sched_period, as the latter can be zero). 3226 */ 3227static bool 3228__checkparam_dl(const struct sched_attr *attr) 3229{ 3230 /* deadline != 0 */ 3231 if (attr->sched_deadline == 0) 3232 return false; 3233 3234 /* 3235 * Since we truncate DL_SCALE bits, make sure we're at least 3236 * that big. 3237 */ 3238 if (attr->sched_runtime < (1ULL << DL_SCALE)) 3239 return false; 3240 3241 /* 3242 * Since we use the MSB for wrap-around and sign issues, make 3243 * sure it's not set (mind that period can be equal to zero). 3244 */ 3245 if (attr->sched_deadline & (1ULL << 63) || 3246 attr->sched_period & (1ULL << 63)) 3247 return false; 3248 3249 /* runtime <= deadline <= period (if period != 0) */ 3250 if ((attr->sched_period != 0 && 3251 attr->sched_period < attr->sched_deadline) || 3252 attr->sched_deadline < attr->sched_runtime) 3253 return false; 3254 3255 return true; 3256} 3257 3258/* 3259 * check the target process has a UID that matches the current process's 3260 */ 3261static bool check_same_owner(struct task_struct *p) 3262{ 3263 const struct cred *cred = current_cred(), *pcred; 3264 bool match; 3265 3266 rcu_read_lock(); 3267 pcred = __task_cred(p); 3268 match = (uid_eq(cred->euid, pcred->euid) || 3269 uid_eq(cred->euid, pcred->uid)); 3270 rcu_read_unlock(); 3271 return match; 3272} 3273 3274static int __sched_setscheduler(struct task_struct *p, 3275 const struct sched_attr *attr, 3276 bool user) 3277{ 3278 int newprio = dl_policy(attr->sched_policy) ? MAX_DL_PRIO - 1 : 3279 MAX_RT_PRIO - 1 - attr->sched_priority; 3280 int retval, oldprio, oldpolicy = -1, on_rq, running; 3281 int policy = attr->sched_policy; 3282 unsigned long flags; 3283 const struct sched_class *prev_class; 3284 struct rq *rq; 3285 int reset_on_fork; 3286 3287 /* may grab non-irq protected spin_locks */ 3288 BUG_ON(in_interrupt()); 3289recheck: 3290 /* double check policy once rq lock held */ 3291 if (policy < 0) { 3292 reset_on_fork = p->sched_reset_on_fork; 3293 policy = oldpolicy = p->policy; 3294 } else { 3295 reset_on_fork = !!(attr->sched_flags & SCHED_FLAG_RESET_ON_FORK); 3296 3297 if (policy != SCHED_DEADLINE && 3298 policy != SCHED_FIFO && policy != SCHED_RR && 3299 policy != SCHED_NORMAL && policy != SCHED_BATCH && 3300 policy != SCHED_IDLE) 3301 return -EINVAL; 3302 } 3303 3304 if (attr->sched_flags & ~(SCHED_FLAG_RESET_ON_FORK)) 3305 return -EINVAL; 3306 3307 /* 3308 * Valid priorities for SCHED_FIFO and SCHED_RR are 3309 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL, 3310 * SCHED_BATCH and SCHED_IDLE is 0. 3311 */ 3312 if ((p->mm && attr->sched_priority > MAX_USER_RT_PRIO-1) || 3313 (!p->mm && attr->sched_priority > MAX_RT_PRIO-1)) 3314 return -EINVAL; 3315 if ((dl_policy(policy) && !__checkparam_dl(attr)) || 3316 (rt_policy(policy) != (attr->sched_priority != 0))) 3317 return -EINVAL; 3318 3319 /* 3320 * Allow unprivileged RT tasks to decrease priority: 3321 */ 3322 if (user && !capable(CAP_SYS_NICE)) { 3323 if (fair_policy(policy)) { 3324 if (attr->sched_nice < task_nice(p) && 3325 !can_nice(p, attr->sched_nice)) 3326 return -EPERM; 3327 } 3328 3329 if (rt_policy(policy)) { 3330 unsigned long rlim_rtprio = 3331 task_rlimit(p, RLIMIT_RTPRIO); 3332 3333 /* can't set/change the rt policy */ 3334 if (policy != p->policy && !rlim_rtprio) 3335 return -EPERM; 3336 3337 /* can't increase priority */ 3338 if (attr->sched_priority > p->rt_priority && 3339 attr->sched_priority > rlim_rtprio) 3340 return -EPERM; 3341 } 3342 3343 /* 3344 * Can't set/change SCHED_DEADLINE policy at all for now 3345 * (safest behavior); in the future we would like to allow 3346 * unprivileged DL tasks to increase their relative deadline 3347 * or reduce their runtime (both ways reducing utilization) 3348 */ 3349 if (dl_policy(policy)) 3350 return -EPERM; 3351 3352 /* 3353 * Treat SCHED_IDLE as nice 20. Only allow a switch to 3354 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it. 3355 */ 3356 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) { 3357 if (!can_nice(p, task_nice(p))) 3358 return -EPERM; 3359 } 3360 3361 /* can't change other user's priorities */ 3362 if (!check_same_owner(p)) 3363 return -EPERM; 3364 3365 /* Normal users shall not reset the sched_reset_on_fork flag */ 3366 if (p->sched_reset_on_fork && !reset_on_fork) 3367 return -EPERM; 3368 } 3369 3370 if (user) { 3371 retval = security_task_setscheduler(p); 3372 if (retval) 3373 return retval; 3374 } 3375 3376 /* 3377 * make sure no PI-waiters arrive (or leave) while we are 3378 * changing the priority of the task: 3379 * 3380 * To be able to change p->policy safely, the appropriate 3381 * runqueue lock must be held. 3382 */ 3383 rq = task_rq_lock(p, &flags); 3384 3385 /* 3386 * Changing the policy of the stop threads its a very bad idea 3387 */ 3388 if (p == rq->stop) { 3389 task_rq_unlock(rq, p, &flags); 3390 return -EINVAL; 3391 } 3392 3393 /* 3394 * If not changing anything there's no need to proceed further, 3395 * but store a possible modification of reset_on_fork. 3396 */ 3397 if (unlikely(policy == p->policy)) { 3398 if (fair_policy(policy) && attr->sched_nice != task_nice(p)) 3399 goto change; 3400 if (rt_policy(policy) && attr->sched_priority != p->rt_priority) 3401 goto change; 3402 if (dl_policy(policy)) 3403 goto change; 3404 3405 p->sched_reset_on_fork = reset_on_fork; 3406 task_rq_unlock(rq, p, &flags); 3407 return 0; 3408 } 3409change: 3410 3411 if (user) { 3412#ifdef CONFIG_RT_GROUP_SCHED 3413 /* 3414 * Do not allow realtime tasks into groups that have no runtime 3415 * assigned. 3416 */ 3417 if (rt_bandwidth_enabled() && rt_policy(policy) && 3418 task_group(p)->rt_bandwidth.rt_runtime == 0 && 3419 !task_group_is_autogroup(task_group(p))) { 3420 task_rq_unlock(rq, p, &flags); 3421 return -EPERM; 3422 } 3423#endif 3424#ifdef CONFIG_SMP 3425 if (dl_bandwidth_enabled() && dl_policy(policy)) { 3426 cpumask_t *span = rq->rd->span; 3427 3428 /* 3429 * Don't allow tasks with an affinity mask smaller than 3430 * the entire root_domain to become SCHED_DEADLINE. We 3431 * will also fail if there's no bandwidth available. 3432 */ 3433 if (!cpumask_subset(span, &p->cpus_allowed) || 3434 rq->rd->dl_bw.bw == 0) { 3435 task_rq_unlock(rq, p, &flags); 3436 return -EPERM; 3437 } 3438 } 3439#endif 3440 } 3441 3442 /* recheck policy now with rq lock held */ 3443 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) { 3444 policy = oldpolicy = -1; 3445 task_rq_unlock(rq, p, &flags); 3446 goto recheck; 3447 } 3448 3449 /* 3450 * If setscheduling to SCHED_DEADLINE (or changing the parameters 3451 * of a SCHED_DEADLINE task) we need to check if enough bandwidth 3452 * is available. 3453 */ 3454 if ((dl_policy(policy) || dl_task(p)) && dl_overflow(p, policy, attr)) { 3455 task_rq_unlock(rq, p, &flags); 3456 return -EBUSY; 3457 } 3458 3459 p->sched_reset_on_fork = reset_on_fork; 3460 oldprio = p->prio; 3461 3462 /* 3463 * Special case for priority boosted tasks. 3464 * 3465 * If the new priority is lower or equal (user space view) 3466 * than the current (boosted) priority, we just store the new 3467 * normal parameters and do not touch the scheduler class and 3468 * the runqueue. This will be done when the task deboost 3469 * itself. 3470 */ 3471 if (rt_mutex_check_prio(p, newprio)) { 3472 __setscheduler_params(p, attr); 3473 task_rq_unlock(rq, p, &flags); 3474 return 0; 3475 } 3476 3477 on_rq = p->on_rq; 3478 running = task_current(rq, p); 3479 if (on_rq) 3480 dequeue_task(rq, p, 0); 3481 if (running) 3482 p->sched_class->put_prev_task(rq, p); 3483 3484 prev_class = p->sched_class; 3485 __setscheduler(rq, p, attr); 3486 3487 if (running) 3488 p->sched_class->set_curr_task(rq); 3489 if (on_rq) { 3490 /* 3491 * We enqueue to tail when the priority of a task is 3492 * increased (user space view). 3493 */ 3494 enqueue_task(rq, p, oldprio <= p->prio ? ENQUEUE_HEAD : 0); 3495 } 3496 3497 check_class_changed(rq, p, prev_class, oldprio); 3498 task_rq_unlock(rq, p, &flags); 3499 3500 rt_mutex_adjust_pi(p); 3501 3502 return 0; 3503} 3504 3505static int _sched_setscheduler(struct task_struct *p, int policy, 3506 const struct sched_param *param, bool check) 3507{ 3508 struct sched_attr attr = { 3509 .sched_policy = policy, 3510 .sched_priority = param->sched_priority, 3511 .sched_nice = PRIO_TO_NICE(p->static_prio), 3512 }; 3513 3514 /* 3515 * Fixup the legacy SCHED_RESET_ON_FORK hack 3516 */ 3517 if (policy & SCHED_RESET_ON_FORK) { 3518 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK; 3519 policy &= ~SCHED_RESET_ON_FORK; 3520 attr.sched_policy = policy; 3521 } 3522 3523 return __sched_setscheduler(p, &attr, check); 3524} 3525/** 3526 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread. 3527 * @p: the task in question. 3528 * @policy: new policy. 3529 * @param: structure containing the new RT priority. 3530 * 3531 * Return: 0 on success. An error code otherwise. 3532 * 3533 * NOTE that the task may be already dead. 3534 */ 3535int sched_setscheduler(struct task_struct *p, int policy, 3536 const struct sched_param *param) 3537{ 3538 return _sched_setscheduler(p, policy, param, true); 3539} 3540EXPORT_SYMBOL_GPL(sched_setscheduler); 3541 3542int sched_setattr(struct task_struct *p, const struct sched_attr *attr) 3543{ 3544 return __sched_setscheduler(p, attr, true); 3545} 3546EXPORT_SYMBOL_GPL(sched_setattr); 3547 3548/** 3549 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace. 3550 * @p: the task in question. 3551 * @policy: new policy. 3552 * @param: structure containing the new RT priority. 3553 * 3554 * Just like sched_setscheduler, only don't bother checking if the 3555 * current context has permission. For example, this is needed in 3556 * stop_machine(): we create temporary high priority worker threads, 3557 * but our caller might not have that capability. 3558 * 3559 * Return: 0 on success. An error code otherwise. 3560 */ 3561int sched_setscheduler_nocheck(struct task_struct *p, int policy, 3562 const struct sched_param *param) 3563{ 3564 return _sched_setscheduler(p, policy, param, false); 3565} 3566 3567static int 3568do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param) 3569{ 3570 struct sched_param lparam; 3571 struct task_struct *p; 3572 int retval; 3573 3574 if (!param || pid < 0) 3575 return -EINVAL; 3576 if (copy_from_user(&lparam, param, sizeof(struct sched_param))) 3577 return -EFAULT; 3578 3579 rcu_read_lock(); 3580 retval = -ESRCH; 3581 p = find_process_by_pid(pid); 3582 if (p != NULL) 3583 retval = sched_setscheduler(p, policy, &lparam); 3584 rcu_read_unlock(); 3585 3586 return retval; 3587} 3588 3589/* 3590 * Mimics kernel/events/core.c perf_copy_attr(). 3591 */ 3592static int sched_copy_attr(struct sched_attr __user *uattr, 3593 struct sched_attr *attr) 3594{ 3595 u32 size; 3596 int ret; 3597 3598 if (!access_ok(VERIFY_WRITE, uattr, SCHED_ATTR_SIZE_VER0)) 3599 return -EFAULT; 3600 3601 /* 3602 * zero the full structure, so that a short copy will be nice. 3603 */ 3604 memset(attr, 0, sizeof(*attr)); 3605 3606 ret = get_user(size, &uattr->size); 3607 if (ret) 3608 return ret; 3609 3610 if (size > PAGE_SIZE) /* silly large */ 3611 goto err_size; 3612 3613 if (!size) /* abi compat */ 3614 size = SCHED_ATTR_SIZE_VER0; 3615 3616 if (size < SCHED_ATTR_SIZE_VER0) 3617 goto err_size; 3618 3619 /* 3620 * If we're handed a bigger struct than we know of, 3621 * ensure all the unknown bits are 0 - i.e. new 3622 * user-space does not rely on any kernel feature 3623 * extensions we dont know about yet. 3624 */ 3625 if (size > sizeof(*attr)) { 3626 unsigned char __user *addr; 3627 unsigned char __user *end; 3628 unsigned char val; 3629 3630 addr = (void __user *)uattr + sizeof(*attr); 3631 end = (void __user *)uattr + size; 3632 3633 for (; addr < end; addr++) { 3634 ret = get_user(val, addr); 3635 if (ret) 3636 return ret; 3637 if (val) 3638 goto err_size; 3639 } 3640 size = sizeof(*attr); 3641 } 3642 3643 ret = copy_from_user(attr, uattr, size); 3644 if (ret) 3645 return -EFAULT; 3646 3647 /* 3648 * XXX: do we want to be lenient like existing syscalls; or do we want 3649 * to be strict and return an error on out-of-bounds values? 3650 */ 3651 attr->sched_nice = clamp(attr->sched_nice, MIN_NICE, MAX_NICE); 3652 3653 return 0; 3654 3655err_size: 3656 put_user(sizeof(*attr), &uattr->size); 3657 return -E2BIG; 3658} 3659 3660/** 3661 * sys_sched_setscheduler - set/change the scheduler policy and RT priority 3662 * @pid: the pid in question. 3663 * @policy: new policy. 3664 * @param: structure containing the new RT priority. 3665 * 3666 * Return: 0 on success. An error code otherwise. 3667 */ 3668SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy, 3669 struct sched_param __user *, param) 3670{ 3671 /* negative values for policy are not valid */ 3672 if (policy < 0) 3673 return -EINVAL; 3674 3675 return do_sched_setscheduler(pid, policy, param); 3676} 3677 3678/** 3679 * sys_sched_setparam - set/change the RT priority of a thread 3680 * @pid: the pid in question. 3681 * @param: structure containing the new RT priority. 3682 * 3683 * Return: 0 on success. An error code otherwise. 3684 */ 3685SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param) 3686{ 3687 return do_sched_setscheduler(pid, -1, param); 3688} 3689 3690/** 3691 * sys_sched_setattr - same as above, but with extended sched_attr 3692 * @pid: the pid in question. 3693 * @uattr: structure containing the extended parameters. 3694 * @flags: for future extension. 3695 */ 3696SYSCALL_DEFINE3(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr, 3697 unsigned int, flags) 3698{ 3699 struct sched_attr attr; 3700 struct task_struct *p; 3701 int retval; 3702 3703 if (!uattr || pid < 0 || flags) 3704 return -EINVAL; 3705 3706 retval = sched_copy_attr(uattr, &attr); 3707 if (retval) 3708 return retval; 3709 3710 if (attr.sched_policy < 0) 3711 return -EINVAL; 3712 3713 rcu_read_lock(); 3714 retval = -ESRCH; 3715 p = find_process_by_pid(pid); 3716 if (p != NULL) 3717 retval = sched_setattr(p, &attr); 3718 rcu_read_unlock(); 3719 3720 return retval; 3721} 3722 3723/** 3724 * sys_sched_getscheduler - get the policy (scheduling class) of a thread 3725 * @pid: the pid in question. 3726 * 3727 * Return: On success, the policy of the thread. Otherwise, a negative error 3728 * code. 3729 */ 3730SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid) 3731{ 3732 struct task_struct *p; 3733 int retval; 3734 3735 if (pid < 0) 3736 return -EINVAL; 3737 3738 retval = -ESRCH; 3739 rcu_read_lock(); 3740 p = find_process_by_pid(pid); 3741 if (p) { 3742 retval = security_task_getscheduler(p); 3743 if (!retval) 3744 retval = p->policy 3745 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0); 3746 } 3747 rcu_read_unlock(); 3748 return retval; 3749} 3750 3751/** 3752 * sys_sched_getparam - get the RT priority of a thread 3753 * @pid: the pid in question. 3754 * @param: structure containing the RT priority. 3755 * 3756 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error 3757 * code. 3758 */ 3759SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param) 3760{ 3761 struct sched_param lp = { .sched_priority = 0 }; 3762 struct task_struct *p; 3763 int retval; 3764 3765 if (!param || pid < 0) 3766 return -EINVAL; 3767 3768 rcu_read_lock(); 3769 p = find_process_by_pid(pid); 3770 retval = -ESRCH; 3771 if (!p) 3772 goto out_unlock; 3773 3774 retval = security_task_getscheduler(p); 3775 if (retval) 3776 goto out_unlock; 3777 3778 if (task_has_rt_policy(p)) 3779 lp.sched_priority = p->rt_priority; 3780 rcu_read_unlock(); 3781 3782 /* 3783 * This one might sleep, we cannot do it with a spinlock held ... 3784 */ 3785 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0; 3786 3787 return retval; 3788 3789out_unlock: 3790 rcu_read_unlock(); 3791 return retval; 3792} 3793 3794static int sched_read_attr(struct sched_attr __user *uattr, 3795 struct sched_attr *attr, 3796 unsigned int usize) 3797{ 3798 int ret; 3799 3800 if (!access_ok(VERIFY_WRITE, uattr, usize)) 3801 return -EFAULT; 3802 3803 /* 3804 * If we're handed a smaller struct than we know of, 3805 * ensure all the unknown bits are 0 - i.e. old 3806 * user-space does not get uncomplete information. 3807 */ 3808 if (usize < sizeof(*attr)) { 3809 unsigned char *addr; 3810 unsigned char *end; 3811 3812 addr = (void *)attr + usize; 3813 end = (void *)attr + sizeof(*attr); 3814 3815 for (; addr < end; addr++) { 3816 if (*addr) 3817 return -EFBIG; 3818 } 3819 3820 attr->size = usize; 3821 } 3822 3823 ret = copy_to_user(uattr, attr, attr->size); 3824 if (ret) 3825 return -EFAULT; 3826 3827 return 0; 3828} 3829 3830/** 3831 * sys_sched_getattr - similar to sched_getparam, but with sched_attr 3832 * @pid: the pid in question. 3833 * @uattr: structure containing the extended parameters. 3834 * @size: sizeof(attr) for fwd/bwd comp. 3835 * @flags: for future extension. 3836 */ 3837SYSCALL_DEFINE4(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr, 3838 unsigned int, size, unsigned int, flags) 3839{ 3840 struct sched_attr attr = { 3841 .size = sizeof(struct sched_attr), 3842 }; 3843 struct task_struct *p; 3844 int retval; 3845 3846 if (!uattr || pid < 0 || size > PAGE_SIZE || 3847 size < SCHED_ATTR_SIZE_VER0 || flags) 3848 return -EINVAL; 3849 3850 rcu_read_lock(); 3851 p = find_process_by_pid(pid); 3852 retval = -ESRCH; 3853 if (!p) 3854 goto out_unlock; 3855 3856 retval = security_task_getscheduler(p); 3857 if (retval) 3858 goto out_unlock; 3859 3860 attr.sched_policy = p->policy; 3861 if (p->sched_reset_on_fork) 3862 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK; 3863 if (task_has_dl_policy(p)) 3864 __getparam_dl(p, &attr); 3865 else if (task_has_rt_policy(p)) 3866 attr.sched_priority = p->rt_priority; 3867 else 3868 attr.sched_nice = task_nice(p); 3869 3870 rcu_read_unlock(); 3871 3872 retval = sched_read_attr(uattr, &attr, size); 3873 return retval; 3874 3875out_unlock: 3876 rcu_read_unlock(); 3877 return retval; 3878} 3879 3880long sched_setaffinity(pid_t pid, const struct cpumask *in_mask) 3881{ 3882 cpumask_var_t cpus_allowed, new_mask; 3883 struct task_struct *p; 3884 int retval; 3885 3886 rcu_read_lock(); 3887 3888 p = find_process_by_pid(pid); 3889 if (!p) { 3890 rcu_read_unlock(); 3891 return -ESRCH; 3892 } 3893 3894 /* Prevent p going away */ 3895 get_task_struct(p); 3896 rcu_read_unlock(); 3897 3898 if (p->flags & PF_NO_SETAFFINITY) { 3899 retval = -EINVAL; 3900 goto out_put_task; 3901 } 3902 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) { 3903 retval = -ENOMEM; 3904 goto out_put_task; 3905 } 3906 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) { 3907 retval = -ENOMEM; 3908 goto out_free_cpus_allowed; 3909 } 3910 retval = -EPERM; 3911 if (!check_same_owner(p)) { 3912 rcu_read_lock(); 3913 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) { 3914 rcu_read_unlock(); 3915 goto out_unlock; 3916 } 3917 rcu_read_unlock(); 3918 } 3919 3920 retval = security_task_setscheduler(p); 3921 if (retval) 3922 goto out_unlock; 3923 3924 3925 cpuset_cpus_allowed(p, cpus_allowed); 3926 cpumask_and(new_mask, in_mask, cpus_allowed); 3927 3928 /* 3929 * Since bandwidth control happens on root_domain basis, 3930 * if admission test is enabled, we only admit -deadline 3931 * tasks allowed to run on all the CPUs in the task's 3932 * root_domain. 3933 */ 3934#ifdef CONFIG_SMP 3935 if (task_has_dl_policy(p)) { 3936 const struct cpumask *span = task_rq(p)->rd->span; 3937 3938 if (dl_bandwidth_enabled() && !cpumask_subset(span, new_mask)) { 3939 retval = -EBUSY; 3940 goto out_unlock; 3941 } 3942 } 3943#endif 3944again: 3945 retval = set_cpus_allowed_ptr(p, new_mask); 3946 3947 if (!retval) { 3948 cpuset_cpus_allowed(p, cpus_allowed); 3949 if (!cpumask_subset(new_mask, cpus_allowed)) { 3950 /* 3951 * We must have raced with a concurrent cpuset 3952 * update. Just reset the cpus_allowed to the 3953 * cpuset's cpus_allowed 3954 */ 3955 cpumask_copy(new_mask, cpus_allowed); 3956 goto again; 3957 } 3958 } 3959out_unlock: 3960 free_cpumask_var(new_mask); 3961out_free_cpus_allowed: 3962 free_cpumask_var(cpus_allowed); 3963out_put_task: 3964 put_task_struct(p); 3965 return retval; 3966} 3967 3968static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len, 3969 struct cpumask *new_mask) 3970{ 3971 if (len < cpumask_size()) 3972 cpumask_clear(new_mask); 3973 else if (len > cpumask_size()) 3974 len = cpumask_size(); 3975 3976 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0; 3977} 3978 3979/** 3980 * sys_sched_setaffinity - set the cpu affinity of a process 3981 * @pid: pid of the process 3982 * @len: length in bytes of the bitmask pointed to by user_mask_ptr 3983 * @user_mask_ptr: user-space pointer to the new cpu mask 3984 * 3985 * Return: 0 on success. An error code otherwise. 3986 */ 3987SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len, 3988 unsigned long __user *, user_mask_ptr) 3989{ 3990 cpumask_var_t new_mask; 3991 int retval; 3992 3993 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) 3994 return -ENOMEM; 3995 3996 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask); 3997 if (retval == 0) 3998 retval = sched_setaffinity(pid, new_mask); 3999 free_cpumask_var(new_mask); 4000 return retval; 4001} 4002 4003long sched_getaffinity(pid_t pid, struct cpumask *mask) 4004{ 4005 struct task_struct *p; 4006 unsigned long flags; 4007 int retval; 4008 4009 rcu_read_lock(); 4010 4011 retval = -ESRCH; 4012 p = find_process_by_pid(pid); 4013 if (!p) 4014 goto out_unlock; 4015 4016 retval = security_task_getscheduler(p); 4017 if (retval) 4018 goto out_unlock; 4019 4020 raw_spin_lock_irqsave(&p->pi_lock, flags); 4021 cpumask_and(mask, &p->cpus_allowed, cpu_active_mask); 4022 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 4023 4024out_unlock: 4025 rcu_read_unlock(); 4026 4027 return retval; 4028} 4029 4030/** 4031 * sys_sched_getaffinity - get the cpu affinity of a process 4032 * @pid: pid of the process 4033 * @len: length in bytes of the bitmask pointed to by user_mask_ptr 4034 * @user_mask_ptr: user-space pointer to hold the current cpu mask 4035 * 4036 * Return: 0 on success. An error code otherwise. 4037 */ 4038SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len, 4039 unsigned long __user *, user_mask_ptr) 4040{ 4041 int ret; 4042 cpumask_var_t mask; 4043 4044 if ((len * BITS_PER_BYTE) < nr_cpu_ids) 4045 return -EINVAL; 4046 if (len & (sizeof(unsigned long)-1)) 4047 return -EINVAL; 4048 4049 if (!alloc_cpumask_var(&mask, GFP_KERNEL)) 4050 return -ENOMEM; 4051 4052 ret = sched_getaffinity(pid, mask); 4053 if (ret == 0) { 4054 size_t retlen = min_t(size_t, len, cpumask_size()); 4055 4056 if (copy_to_user(user_mask_ptr, mask, retlen)) 4057 ret = -EFAULT; 4058 else 4059 ret = retlen; 4060 } 4061 free_cpumask_var(mask); 4062 4063 return ret; 4064} 4065 4066/** 4067 * sys_sched_yield - yield the current processor to other threads. 4068 * 4069 * This function yields the current CPU to other tasks. If there are no 4070 * other threads running on this CPU then this function will return. 4071 * 4072 * Return: 0. 4073 */ 4074SYSCALL_DEFINE0(sched_yield) 4075{ 4076 struct rq *rq = this_rq_lock(); 4077 4078 schedstat_inc(rq, yld_count); 4079 current->sched_class->yield_task(rq); 4080 4081 /* 4082 * Since we are going to call schedule() anyway, there's 4083 * no need to preempt or enable interrupts: 4084 */ 4085 __release(rq->lock); 4086 spin_release(&rq->lock.dep_map, 1, _THIS_IP_); 4087 do_raw_spin_unlock(&rq->lock); 4088 sched_preempt_enable_no_resched(); 4089 4090 schedule(); 4091 4092 return 0; 4093} 4094 4095static void __cond_resched(void) 4096{ 4097 __preempt_count_add(PREEMPT_ACTIVE); 4098 __schedule(); 4099 __preempt_count_sub(PREEMPT_ACTIVE); 4100} 4101 4102int __sched _cond_resched(void) 4103{ 4104 if (should_resched()) { 4105 __cond_resched(); 4106 return 1; 4107 } 4108 return 0; 4109} 4110EXPORT_SYMBOL(_cond_resched); 4111 4112/* 4113 * __cond_resched_lock() - if a reschedule is pending, drop the given lock, 4114 * call schedule, and on return reacquire the lock. 4115 * 4116 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level 4117 * operations here to prevent schedule() from being called twice (once via 4118 * spin_unlock(), once by hand). 4119 */ 4120int __cond_resched_lock(spinlock_t *lock) 4121{ 4122 int resched = should_resched(); 4123 int ret = 0; 4124 4125 lockdep_assert_held(lock); 4126 4127 if (spin_needbreak(lock) || resched) { 4128 spin_unlock(lock); 4129 if (resched) 4130 __cond_resched(); 4131 else 4132 cpu_relax(); 4133 ret = 1; 4134 spin_lock(lock); 4135 } 4136 return ret; 4137} 4138EXPORT_SYMBOL(__cond_resched_lock); 4139 4140int __sched __cond_resched_softirq(void) 4141{ 4142 BUG_ON(!in_softirq()); 4143 4144 if (should_resched()) { 4145 local_bh_enable(); 4146 __cond_resched(); 4147 local_bh_disable(); 4148 return 1; 4149 } 4150 return 0; 4151} 4152EXPORT_SYMBOL(__cond_resched_softirq); 4153 4154/** 4155 * yield - yield the current processor to other threads. 4156 * 4157 * Do not ever use this function, there's a 99% chance you're doing it wrong. 4158 * 4159 * The scheduler is at all times free to pick the calling task as the most 4160 * eligible task to run, if removing the yield() call from your code breaks 4161 * it, its already broken. 4162 * 4163 * Typical broken usage is: 4164 * 4165 * while (!event) 4166 * yield(); 4167 * 4168 * where one assumes that yield() will let 'the other' process run that will 4169 * make event true. If the current task is a SCHED_FIFO task that will never 4170 * happen. Never use yield() as a progress guarantee!! 4171 * 4172 * If you want to use yield() to wait for something, use wait_event(). 4173 * If you want to use yield() to be 'nice' for others, use cond_resched(). 4174 * If you still want to use yield(), do not! 4175 */ 4176void __sched yield(void) 4177{ 4178 set_current_state(TASK_RUNNING); 4179 sys_sched_yield(); 4180} 4181EXPORT_SYMBOL(yield); 4182 4183/** 4184 * yield_to - yield the current processor to another thread in 4185 * your thread group, or accelerate that thread toward the 4186 * processor it's on. 4187 * @p: target task 4188 * @preempt: whether task preemption is allowed or not 4189 * 4190 * It's the caller's job to ensure that the target task struct 4191 * can't go away on us before we can do any checks. 4192 * 4193 * Return: 4194 * true (>0) if we indeed boosted the target task. 4195 * false (0) if we failed to boost the target. 4196 * -ESRCH if there's no task to yield to. 4197 */ 4198bool __sched yield_to(struct task_struct *p, bool preempt) 4199{ 4200 struct task_struct *curr = current; 4201 struct rq *rq, *p_rq; 4202 unsigned long flags; 4203 int yielded = 0; 4204 4205 local_irq_save(flags); 4206 rq = this_rq(); 4207 4208again: 4209 p_rq = task_rq(p); 4210 /* 4211 * If we're the only runnable task on the rq and target rq also 4212 * has only one task, there's absolutely no point in yielding. 4213 */ 4214 if (rq->nr_running == 1 && p_rq->nr_running == 1) { 4215 yielded = -ESRCH; 4216 goto out_irq; 4217 } 4218 4219 double_rq_lock(rq, p_rq); 4220 if (task_rq(p) != p_rq) { 4221 double_rq_unlock(rq, p_rq); 4222 goto again; 4223 } 4224 4225 if (!curr->sched_class->yield_to_task) 4226 goto out_unlock; 4227 4228 if (curr->sched_class != p->sched_class) 4229 goto out_unlock; 4230 4231 if (task_running(p_rq, p) || p->state) 4232 goto out_unlock; 4233 4234 yielded = curr->sched_class->yield_to_task(rq, p, preempt); 4235 if (yielded) { 4236 schedstat_inc(rq, yld_count); 4237 /* 4238 * Make p's CPU reschedule; pick_next_entity takes care of 4239 * fairness. 4240 */ 4241 if (preempt && rq != p_rq) 4242 resched_task(p_rq->curr); 4243 } 4244 4245out_unlock: 4246 double_rq_unlock(rq, p_rq); 4247out_irq: 4248 local_irq_restore(flags); 4249 4250 if (yielded > 0) 4251 schedule(); 4252 4253 return yielded; 4254} 4255EXPORT_SYMBOL_GPL(yield_to); 4256 4257/* 4258 * This task is about to go to sleep on IO. Increment rq->nr_iowait so 4259 * that process accounting knows that this is a task in IO wait state. 4260 */ 4261void __sched io_schedule(void) 4262{ 4263 struct rq *rq = raw_rq(); 4264 4265 delayacct_blkio_start(); 4266 atomic_inc(&rq->nr_iowait); 4267 blk_flush_plug(current); 4268 current->in_iowait = 1; 4269 schedule(); 4270 current->in_iowait = 0; 4271 atomic_dec(&rq->nr_iowait); 4272 delayacct_blkio_end(); 4273} 4274EXPORT_SYMBOL(io_schedule); 4275 4276long __sched io_schedule_timeout(long timeout) 4277{ 4278 struct rq *rq = raw_rq(); 4279 long ret; 4280 4281 delayacct_blkio_start(); 4282 atomic_inc(&rq->nr_iowait); 4283 blk_flush_plug(current); 4284 current->in_iowait = 1; 4285 ret = schedule_timeout(timeout); 4286 current->in_iowait = 0; 4287 atomic_dec(&rq->nr_iowait); 4288 delayacct_blkio_end(); 4289 return ret; 4290} 4291 4292/** 4293 * sys_sched_get_priority_max - return maximum RT priority. 4294 * @policy: scheduling class. 4295 * 4296 * Return: On success, this syscall returns the maximum 4297 * rt_priority that can be used by a given scheduling class. 4298 * On failure, a negative error code is returned. 4299 */ 4300SYSCALL_DEFINE1(sched_get_priority_max, int, policy) 4301{ 4302 int ret = -EINVAL; 4303 4304 switch (policy) { 4305 case SCHED_FIFO: 4306 case SCHED_RR: 4307 ret = MAX_USER_RT_PRIO-1; 4308 break; 4309 case SCHED_DEADLINE: 4310 case SCHED_NORMAL: 4311 case SCHED_BATCH: 4312 case SCHED_IDLE: 4313 ret = 0; 4314 break; 4315 } 4316 return ret; 4317} 4318 4319/** 4320 * sys_sched_get_priority_min - return minimum RT priority. 4321 * @policy: scheduling class. 4322 * 4323 * Return: On success, this syscall returns the minimum 4324 * rt_priority that can be used by a given scheduling class. 4325 * On failure, a negative error code is returned. 4326 */ 4327SYSCALL_DEFINE1(sched_get_priority_min, int, policy) 4328{ 4329 int ret = -EINVAL; 4330 4331 switch (policy) { 4332 case SCHED_FIFO: 4333 case SCHED_RR: 4334 ret = 1; 4335 break; 4336 case SCHED_DEADLINE: 4337 case SCHED_NORMAL: 4338 case SCHED_BATCH: 4339 case SCHED_IDLE: 4340 ret = 0; 4341 } 4342 return ret; 4343} 4344 4345/** 4346 * sys_sched_rr_get_interval - return the default timeslice of a process. 4347 * @pid: pid of the process. 4348 * @interval: userspace pointer to the timeslice value. 4349 * 4350 * this syscall writes the default timeslice value of a given process 4351 * into the user-space timespec buffer. A value of '0' means infinity. 4352 * 4353 * Return: On success, 0 and the timeslice is in @interval. Otherwise, 4354 * an error code. 4355 */ 4356SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid, 4357 struct timespec __user *, interval) 4358{ 4359 struct task_struct *p; 4360 unsigned int time_slice; 4361 unsigned long flags; 4362 struct rq *rq; 4363 int retval; 4364 struct timespec t; 4365 4366 if (pid < 0) 4367 return -EINVAL; 4368 4369 retval = -ESRCH; 4370 rcu_read_lock(); 4371 p = find_process_by_pid(pid); 4372 if (!p) 4373 goto out_unlock; 4374 4375 retval = security_task_getscheduler(p); 4376 if (retval) 4377 goto out_unlock; 4378 4379 rq = task_rq_lock(p, &flags); 4380 time_slice = 0; 4381 if (p->sched_class->get_rr_interval) 4382 time_slice = p->sched_class->get_rr_interval(rq, p); 4383 task_rq_unlock(rq, p, &flags); 4384 4385 rcu_read_unlock(); 4386 jiffies_to_timespec(time_slice, &t); 4387 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0; 4388 return retval; 4389 4390out_unlock: 4391 rcu_read_unlock(); 4392 return retval; 4393} 4394 4395static const char stat_nam[] = TASK_STATE_TO_CHAR_STR; 4396 4397void sched_show_task(struct task_struct *p) 4398{ 4399 unsigned long free = 0; 4400 int ppid; 4401 unsigned state; 4402 4403 state = p->state ? __ffs(p->state) + 1 : 0; 4404 printk(KERN_INFO "%-15.15s %c", p->comm, 4405 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?'); 4406#if BITS_PER_LONG == 32 4407 if (state == TASK_RUNNING) 4408 printk(KERN_CONT " running "); 4409 else 4410 printk(KERN_CONT " %08lx ", thread_saved_pc(p)); 4411#else 4412 if (state == TASK_RUNNING) 4413 printk(KERN_CONT " running task "); 4414 else 4415 printk(KERN_CONT " %016lx ", thread_saved_pc(p)); 4416#endif 4417#ifdef CONFIG_DEBUG_STACK_USAGE 4418 free = stack_not_used(p); 4419#endif 4420 rcu_read_lock(); 4421 ppid = task_pid_nr(rcu_dereference(p->real_parent)); 4422 rcu_read_unlock(); 4423 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free, 4424 task_pid_nr(p), ppid, 4425 (unsigned long)task_thread_info(p)->flags); 4426 4427 print_worker_info(KERN_INFO, p); 4428 show_stack(p, NULL); 4429} 4430 4431void show_state_filter(unsigned long state_filter) 4432{ 4433 struct task_struct *g, *p; 4434 4435#if BITS_PER_LONG == 32 4436 printk(KERN_INFO 4437 " task PC stack pid father\n"); 4438#else 4439 printk(KERN_INFO 4440 " task PC stack pid father\n"); 4441#endif 4442 rcu_read_lock(); 4443 do_each_thread(g, p) { 4444 /* 4445 * reset the NMI-timeout, listing all files on a slow 4446 * console might take a lot of time: 4447 */ 4448 touch_nmi_watchdog(); 4449 if (!state_filter || (p->state & state_filter)) 4450 sched_show_task(p); 4451 } while_each_thread(g, p); 4452 4453 touch_all_softlockup_watchdogs(); 4454 4455#ifdef CONFIG_SCHED_DEBUG 4456 sysrq_sched_debug_show(); 4457#endif 4458 rcu_read_unlock(); 4459 /* 4460 * Only show locks if all tasks are dumped: 4461 */ 4462 if (!state_filter) 4463 debug_show_all_locks(); 4464} 4465 4466void init_idle_bootup_task(struct task_struct *idle) 4467{ 4468 idle->sched_class = &idle_sched_class; 4469} 4470 4471/** 4472 * init_idle - set up an idle thread for a given CPU 4473 * @idle: task in question 4474 * @cpu: cpu the idle task belongs to 4475 * 4476 * NOTE: this function does not set the idle thread's NEED_RESCHED 4477 * flag, to make booting more robust. 4478 */ 4479void init_idle(struct task_struct *idle, int cpu) 4480{ 4481 struct rq *rq = cpu_rq(cpu); 4482 unsigned long flags; 4483 4484 raw_spin_lock_irqsave(&rq->lock, flags); 4485 4486 __sched_fork(0, idle); 4487 idle->state = TASK_RUNNING; 4488 idle->se.exec_start = sched_clock(); 4489 4490 do_set_cpus_allowed(idle, cpumask_of(cpu)); 4491 /* 4492 * We're having a chicken and egg problem, even though we are 4493 * holding rq->lock, the cpu isn't yet set to this cpu so the 4494 * lockdep check in task_group() will fail. 4495 * 4496 * Similar case to sched_fork(). / Alternatively we could 4497 * use task_rq_lock() here and obtain the other rq->lock. 4498 * 4499 * Silence PROVE_RCU 4500 */ 4501 rcu_read_lock(); 4502 __set_task_cpu(idle, cpu); 4503 rcu_read_unlock(); 4504 4505 rq->curr = rq->idle = idle; 4506 idle->on_rq = 1; 4507#if defined(CONFIG_SMP) 4508 idle->on_cpu = 1; 4509#endif 4510 raw_spin_unlock_irqrestore(&rq->lock, flags); 4511 4512 /* Set the preempt count _outside_ the spinlocks! */ 4513 init_idle_preempt_count(idle, cpu); 4514 4515 /* 4516 * The idle tasks have their own, simple scheduling class: 4517 */ 4518 idle->sched_class = &idle_sched_class; 4519 ftrace_graph_init_idle_task(idle, cpu); 4520 vtime_init_idle(idle, cpu); 4521#if defined(CONFIG_SMP) 4522 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu); 4523#endif 4524} 4525 4526#ifdef CONFIG_SMP 4527void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask) 4528{ 4529 if (p->sched_class && p->sched_class->set_cpus_allowed) 4530 p->sched_class->set_cpus_allowed(p, new_mask); 4531 4532 cpumask_copy(&p->cpus_allowed, new_mask); 4533 p->nr_cpus_allowed = cpumask_weight(new_mask); 4534} 4535 4536/* 4537 * This is how migration works: 4538 * 4539 * 1) we invoke migration_cpu_stop() on the target CPU using 4540 * stop_one_cpu(). 4541 * 2) stopper starts to run (implicitly forcing the migrated thread 4542 * off the CPU) 4543 * 3) it checks whether the migrated task is still in the wrong runqueue. 4544 * 4) if it's in the wrong runqueue then the migration thread removes 4545 * it and puts it into the right queue. 4546 * 5) stopper completes and stop_one_cpu() returns and the migration 4547 * is done. 4548 */ 4549 4550/* 4551 * Change a given task's CPU affinity. Migrate the thread to a 4552 * proper CPU and schedule it away if the CPU it's executing on 4553 * is removed from the allowed bitmask. 4554 * 4555 * NOTE: the caller must have a valid reference to the task, the 4556 * task must not exit() & deallocate itself prematurely. The 4557 * call is not atomic; no spinlocks may be held. 4558 */ 4559int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask) 4560{ 4561 unsigned long flags; 4562 struct rq *rq; 4563 unsigned int dest_cpu; 4564 int ret = 0; 4565 4566 rq = task_rq_lock(p, &flags); 4567 4568 if (cpumask_equal(&p->cpus_allowed, new_mask)) 4569 goto out; 4570 4571 if (!cpumask_intersects(new_mask, cpu_active_mask)) { 4572 ret = -EINVAL; 4573 goto out; 4574 } 4575 4576 do_set_cpus_allowed(p, new_mask); 4577 4578 /* Can the task run on the task's current CPU? If so, we're done */ 4579 if (cpumask_test_cpu(task_cpu(p), new_mask)) 4580 goto out; 4581 4582 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask); 4583 if (p->on_rq) { 4584 struct migration_arg arg = { p, dest_cpu }; 4585 /* Need help from migration thread: drop lock and wait. */ 4586 task_rq_unlock(rq, p, &flags); 4587 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg); 4588 tlb_migrate_finish(p->mm); 4589 return 0; 4590 } 4591out: 4592 task_rq_unlock(rq, p, &flags); 4593 4594 return ret; 4595} 4596EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr); 4597 4598/* 4599 * Move (not current) task off this cpu, onto dest cpu. We're doing 4600 * this because either it can't run here any more (set_cpus_allowed() 4601 * away from this CPU, or CPU going down), or because we're 4602 * attempting to rebalance this task on exec (sched_exec). 4603 * 4604 * So we race with normal scheduler movements, but that's OK, as long 4605 * as the task is no longer on this CPU. 4606 * 4607 * Returns non-zero if task was successfully migrated. 4608 */ 4609static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu) 4610{ 4611 struct rq *rq_dest, *rq_src; 4612 int ret = 0; 4613 4614 if (unlikely(!cpu_active(dest_cpu))) 4615 return ret; 4616 4617 rq_src = cpu_rq(src_cpu); 4618 rq_dest = cpu_rq(dest_cpu); 4619 4620 raw_spin_lock(&p->pi_lock); 4621 double_rq_lock(rq_src, rq_dest); 4622 /* Already moved. */ 4623 if (task_cpu(p) != src_cpu) 4624 goto done; 4625 /* Affinity changed (again). */ 4626 if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p))) 4627 goto fail; 4628 4629 /* 4630 * If we're not on a rq, the next wake-up will ensure we're 4631 * placed properly. 4632 */ 4633 if (p->on_rq) { 4634 dequeue_task(rq_src, p, 0); 4635 set_task_cpu(p, dest_cpu); 4636 enqueue_task(rq_dest, p, 0); 4637 check_preempt_curr(rq_dest, p, 0); 4638 } 4639done: 4640 ret = 1; 4641fail: 4642 double_rq_unlock(rq_src, rq_dest); 4643 raw_spin_unlock(&p->pi_lock); 4644 return ret; 4645} 4646 4647#ifdef CONFIG_NUMA_BALANCING 4648/* Migrate current task p to target_cpu */ 4649int migrate_task_to(struct task_struct *p, int target_cpu) 4650{ 4651 struct migration_arg arg = { p, target_cpu }; 4652 int curr_cpu = task_cpu(p); 4653 4654 if (curr_cpu == target_cpu) 4655 return 0; 4656 4657 if (!cpumask_test_cpu(target_cpu, tsk_cpus_allowed(p))) 4658 return -EINVAL; 4659 4660 /* TODO: This is not properly updating schedstats */ 4661 4662 trace_sched_move_numa(p, curr_cpu, target_cpu); 4663 return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg); 4664} 4665 4666/* 4667 * Requeue a task on a given node and accurately track the number of NUMA 4668 * tasks on the runqueues 4669 */ 4670void sched_setnuma(struct task_struct *p, int nid) 4671{ 4672 struct rq *rq; 4673 unsigned long flags; 4674 bool on_rq, running; 4675 4676 rq = task_rq_lock(p, &flags); 4677 on_rq = p->on_rq; 4678 running = task_current(rq, p); 4679 4680 if (on_rq) 4681 dequeue_task(rq, p, 0); 4682 if (running) 4683 p->sched_class->put_prev_task(rq, p); 4684 4685 p->numa_preferred_nid = nid; 4686 4687 if (running) 4688 p->sched_class->set_curr_task(rq); 4689 if (on_rq) 4690 enqueue_task(rq, p, 0); 4691 task_rq_unlock(rq, p, &flags); 4692} 4693#endif 4694 4695/* 4696 * migration_cpu_stop - this will be executed by a highprio stopper thread 4697 * and performs thread migration by bumping thread off CPU then 4698 * 'pushing' onto another runqueue. 4699 */ 4700static int migration_cpu_stop(void *data) 4701{ 4702 struct migration_arg *arg = data; 4703 4704 /* 4705 * The original target cpu might have gone down and we might 4706 * be on another cpu but it doesn't matter. 4707 */ 4708 local_irq_disable(); 4709 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu); 4710 local_irq_enable(); 4711 return 0; 4712} 4713 4714#ifdef CONFIG_HOTPLUG_CPU 4715 4716/* 4717 * Ensures that the idle task is using init_mm right before its cpu goes 4718 * offline. 4719 */ 4720void idle_task_exit(void) 4721{ 4722 struct mm_struct *mm = current->active_mm; 4723 4724 BUG_ON(cpu_online(smp_processor_id())); 4725 4726 if (mm != &init_mm) { 4727 switch_mm(mm, &init_mm, current); 4728 finish_arch_post_lock_switch(); 4729 } 4730 mmdrop(mm); 4731} 4732 4733/* 4734 * Since this CPU is going 'away' for a while, fold any nr_active delta 4735 * we might have. Assumes we're called after migrate_tasks() so that the 4736 * nr_active count is stable. 4737 * 4738 * Also see the comment "Global load-average calculations". 4739 */ 4740static void calc_load_migrate(struct rq *rq) 4741{ 4742 long delta = calc_load_fold_active(rq); 4743 if (delta) 4744 atomic_long_add(delta, &calc_load_tasks); 4745} 4746 4747static void put_prev_task_fake(struct rq *rq, struct task_struct *prev) 4748{ 4749} 4750 4751static const struct sched_class fake_sched_class = { 4752 .put_prev_task = put_prev_task_fake, 4753}; 4754 4755static struct task_struct fake_task = { 4756 /* 4757 * Avoid pull_{rt,dl}_task() 4758 */ 4759 .prio = MAX_PRIO + 1, 4760 .sched_class = &fake_sched_class, 4761}; 4762 4763/* 4764 * Migrate all tasks from the rq, sleeping tasks will be migrated by 4765 * try_to_wake_up()->select_task_rq(). 4766 * 4767 * Called with rq->lock held even though we'er in stop_machine() and 4768 * there's no concurrency possible, we hold the required locks anyway 4769 * because of lock validation efforts. 4770 */ 4771static void migrate_tasks(unsigned int dead_cpu) 4772{ 4773 struct rq *rq = cpu_rq(dead_cpu); 4774 struct task_struct *next, *stop = rq->stop; 4775 int dest_cpu; 4776 4777 /* 4778 * Fudge the rq selection such that the below task selection loop 4779 * doesn't get stuck on the currently eligible stop task. 4780 * 4781 * We're currently inside stop_machine() and the rq is either stuck 4782 * in the stop_machine_cpu_stop() loop, or we're executing this code, 4783 * either way we should never end up calling schedule() until we're 4784 * done here. 4785 */ 4786 rq->stop = NULL; 4787 4788 /* 4789 * put_prev_task() and pick_next_task() sched 4790 * class method both need to have an up-to-date 4791 * value of rq->clock[_task] 4792 */ 4793 update_rq_clock(rq); 4794 4795 for ( ; ; ) { 4796 /* 4797 * There's this thread running, bail when that's the only 4798 * remaining thread. 4799 */ 4800 if (rq->nr_running == 1) 4801 break; 4802 4803 next = pick_next_task(rq, &fake_task); 4804 BUG_ON(!next); 4805 next->sched_class->put_prev_task(rq, next); 4806 4807 /* Find suitable destination for @next, with force if needed. */ 4808 dest_cpu = select_fallback_rq(dead_cpu, next); 4809 raw_spin_unlock(&rq->lock); 4810 4811 __migrate_task(next, dead_cpu, dest_cpu); 4812 4813 raw_spin_lock(&rq->lock); 4814 } 4815 4816 rq->stop = stop; 4817} 4818 4819#endif /* CONFIG_HOTPLUG_CPU */ 4820 4821#if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL) 4822 4823static struct ctl_table sd_ctl_dir[] = { 4824 { 4825 .procname = "sched_domain", 4826 .mode = 0555, 4827 }, 4828 {} 4829}; 4830 4831static struct ctl_table sd_ctl_root[] = { 4832 { 4833 .procname = "kernel", 4834 .mode = 0555, 4835 .child = sd_ctl_dir, 4836 }, 4837 {} 4838}; 4839 4840static struct ctl_table *sd_alloc_ctl_entry(int n) 4841{ 4842 struct ctl_table *entry = 4843 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL); 4844 4845 return entry; 4846} 4847 4848static void sd_free_ctl_entry(struct ctl_table **tablep) 4849{ 4850 struct ctl_table *entry; 4851 4852 /* 4853 * In the intermediate directories, both the child directory and 4854 * procname are dynamically allocated and could fail but the mode 4855 * will always be set. In the lowest directory the names are 4856 * static strings and all have proc handlers. 4857 */ 4858 for (entry = *tablep; entry->mode; entry++) { 4859 if (entry->child) 4860 sd_free_ctl_entry(&entry->child); 4861 if (entry->proc_handler == NULL) 4862 kfree(entry->procname); 4863 } 4864 4865 kfree(*tablep); 4866 *tablep = NULL; 4867} 4868 4869static int min_load_idx = 0; 4870static int max_load_idx = CPU_LOAD_IDX_MAX-1; 4871 4872static void 4873set_table_entry(struct ctl_table *entry, 4874 const char *procname, void *data, int maxlen, 4875 umode_t mode, proc_handler *proc_handler, 4876 bool load_idx) 4877{ 4878 entry->procname = procname; 4879 entry->data = data; 4880 entry->maxlen = maxlen; 4881 entry->mode = mode; 4882 entry->proc_handler = proc_handler; 4883 4884 if (load_idx) { 4885 entry->extra1 = &min_load_idx; 4886 entry->extra2 = &max_load_idx; 4887 } 4888} 4889 4890static struct ctl_table * 4891sd_alloc_ctl_domain_table(struct sched_domain *sd) 4892{ 4893 struct ctl_table *table = sd_alloc_ctl_entry(14); 4894 4895 if (table == NULL) 4896 return NULL; 4897 4898 set_table_entry(&table[0], "min_interval", &sd->min_interval, 4899 sizeof(long), 0644, proc_doulongvec_minmax, false); 4900 set_table_entry(&table[1], "max_interval", &sd->max_interval, 4901 sizeof(long), 0644, proc_doulongvec_minmax, false); 4902 set_table_entry(&table[2], "busy_idx", &sd->busy_idx, 4903 sizeof(int), 0644, proc_dointvec_minmax, true); 4904 set_table_entry(&table[3], "idle_idx", &sd->idle_idx, 4905 sizeof(int), 0644, proc_dointvec_minmax, true); 4906 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx, 4907 sizeof(int), 0644, proc_dointvec_minmax, true); 4908 set_table_entry(&table[5], "wake_idx", &sd->wake_idx, 4909 sizeof(int), 0644, proc_dointvec_minmax, true); 4910 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx, 4911 sizeof(int), 0644, proc_dointvec_minmax, true); 4912 set_table_entry(&table[7], "busy_factor", &sd->busy_factor, 4913 sizeof(int), 0644, proc_dointvec_minmax, false); 4914 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct, 4915 sizeof(int), 0644, proc_dointvec_minmax, false); 4916 set_table_entry(&table[9], "cache_nice_tries", 4917 &sd->cache_nice_tries, 4918 sizeof(int), 0644, proc_dointvec_minmax, false); 4919 set_table_entry(&table[10], "flags", &sd->flags, 4920 sizeof(int), 0644, proc_dointvec_minmax, false); 4921 set_table_entry(&table[11], "max_newidle_lb_cost", 4922 &sd->max_newidle_lb_cost, 4923 sizeof(long), 0644, proc_doulongvec_minmax, false); 4924 set_table_entry(&table[12], "name", sd->name, 4925 CORENAME_MAX_SIZE, 0444, proc_dostring, false); 4926 /* &table[13] is terminator */ 4927 4928 return table; 4929} 4930 4931static struct ctl_table *sd_alloc_ctl_cpu_table(int cpu) 4932{ 4933 struct ctl_table *entry, *table; 4934 struct sched_domain *sd; 4935 int domain_num = 0, i; 4936 char buf[32]; 4937 4938 for_each_domain(cpu, sd) 4939 domain_num++; 4940 entry = table = sd_alloc_ctl_entry(domain_num + 1); 4941 if (table == NULL) 4942 return NULL; 4943 4944 i = 0; 4945 for_each_domain(cpu, sd) { 4946 snprintf(buf, 32, "domain%d", i); 4947 entry->procname = kstrdup(buf, GFP_KERNEL); 4948 entry->mode = 0555; 4949 entry->child = sd_alloc_ctl_domain_table(sd); 4950 entry++; 4951 i++; 4952 } 4953 return table; 4954} 4955 4956static struct ctl_table_header *sd_sysctl_header; 4957static void register_sched_domain_sysctl(void) 4958{ 4959 int i, cpu_num = num_possible_cpus(); 4960 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1); 4961 char buf[32]; 4962 4963 WARN_ON(sd_ctl_dir[0].child); 4964 sd_ctl_dir[0].child = entry; 4965 4966 if (entry == NULL) 4967 return; 4968 4969 for_each_possible_cpu(i) { 4970 snprintf(buf, 32, "cpu%d", i); 4971 entry->procname = kstrdup(buf, GFP_KERNEL); 4972 entry->mode = 0555; 4973 entry->child = sd_alloc_ctl_cpu_table(i); 4974 entry++; 4975 } 4976 4977 WARN_ON(sd_sysctl_header); 4978 sd_sysctl_header = register_sysctl_table(sd_ctl_root); 4979} 4980 4981/* may be called multiple times per register */ 4982static void unregister_sched_domain_sysctl(void) 4983{ 4984 if (sd_sysctl_header) 4985 unregister_sysctl_table(sd_sysctl_header); 4986 sd_sysctl_header = NULL; 4987 if (sd_ctl_dir[0].child) 4988 sd_free_ctl_entry(&sd_ctl_dir[0].child); 4989} 4990#else 4991static void register_sched_domain_sysctl(void) 4992{ 4993} 4994static void unregister_sched_domain_sysctl(void) 4995{ 4996} 4997#endif 4998 4999static void set_rq_online(struct rq *rq) 5000{ 5001 if (!rq->online) { 5002 const struct sched_class *class; 5003 5004 cpumask_set_cpu(rq->cpu, rq->rd->online); 5005 rq->online = 1; 5006 5007 for_each_class(class) { 5008 if (class->rq_online) 5009 class->rq_online(rq); 5010 } 5011 } 5012} 5013 5014static void set_rq_offline(struct rq *rq) 5015{ 5016 if (rq->online) { 5017 const struct sched_class *class; 5018 5019 for_each_class(class) { 5020 if (class->rq_offline) 5021 class->rq_offline(rq); 5022 } 5023 5024 cpumask_clear_cpu(rq->cpu, rq->rd->online); 5025 rq->online = 0; 5026 } 5027} 5028 5029/* 5030 * migration_call - callback that gets triggered when a CPU is added. 5031 * Here we can start up the necessary migration thread for the new CPU. 5032 */ 5033static int 5034migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu) 5035{ 5036 int cpu = (long)hcpu; 5037 unsigned long flags; 5038 struct rq *rq = cpu_rq(cpu); 5039 5040 switch (action & ~CPU_TASKS_FROZEN) { 5041 5042 case CPU_UP_PREPARE: 5043 rq->calc_load_update = calc_load_update; 5044 break; 5045 5046 case CPU_ONLINE: 5047 /* Update our root-domain */ 5048 raw_spin_lock_irqsave(&rq->lock, flags); 5049 if (rq->rd) { 5050 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span)); 5051 5052 set_rq_online(rq); 5053 } 5054 raw_spin_unlock_irqrestore(&rq->lock, flags); 5055 break; 5056 5057#ifdef CONFIG_HOTPLUG_CPU 5058 case CPU_DYING: 5059 sched_ttwu_pending(); 5060 /* Update our root-domain */ 5061 raw_spin_lock_irqsave(&rq->lock, flags); 5062 if (rq->rd) { 5063 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span)); 5064 set_rq_offline(rq); 5065 } 5066 migrate_tasks(cpu); 5067 BUG_ON(rq->nr_running != 1); /* the migration thread */ 5068 raw_spin_unlock_irqrestore(&rq->lock, flags); 5069 break; 5070 5071 case CPU_DEAD: 5072 calc_load_migrate(rq); 5073 break; 5074#endif 5075 } 5076 5077 update_max_interval(); 5078 5079 return NOTIFY_OK; 5080} 5081 5082/* 5083 * Register at high priority so that task migration (migrate_all_tasks) 5084 * happens before everything else. This has to be lower priority than 5085 * the notifier in the perf_event subsystem, though. 5086 */ 5087static struct notifier_block migration_notifier = { 5088 .notifier_call = migration_call, 5089 .priority = CPU_PRI_MIGRATION, 5090}; 5091 5092static void __cpuinit set_cpu_rq_start_time(void) 5093{ 5094 int cpu = smp_processor_id(); 5095 struct rq *rq = cpu_rq(cpu); 5096 rq->age_stamp = sched_clock_cpu(cpu); 5097} 5098 5099static int sched_cpu_active(struct notifier_block *nfb, 5100 unsigned long action, void *hcpu) 5101{ 5102 switch (action & ~CPU_TASKS_FROZEN) { 5103 case CPU_STARTING: 5104 set_cpu_rq_start_time(); 5105 return NOTIFY_OK; 5106 case CPU_DOWN_FAILED: 5107 set_cpu_active((long)hcpu, true); 5108 return NOTIFY_OK; 5109 default: 5110 return NOTIFY_DONE; 5111 } 5112} 5113 5114static int sched_cpu_inactive(struct notifier_block *nfb, 5115 unsigned long action, void *hcpu) 5116{ 5117 unsigned long flags; 5118 long cpu = (long)hcpu; 5119 5120 switch (action & ~CPU_TASKS_FROZEN) { 5121 case CPU_DOWN_PREPARE: 5122 set_cpu_active(cpu, false); 5123 5124 /* explicitly allow suspend */ 5125 if (!(action & CPU_TASKS_FROZEN)) { 5126 struct dl_bw *dl_b = dl_bw_of(cpu); 5127 bool overflow; 5128 int cpus; 5129 5130 raw_spin_lock_irqsave(&dl_b->lock, flags); 5131 cpus = dl_bw_cpus(cpu); 5132 overflow = __dl_overflow(dl_b, cpus, 0, 0); 5133 raw_spin_unlock_irqrestore(&dl_b->lock, flags); 5134 5135 if (overflow) 5136 return notifier_from_errno(-EBUSY); 5137 } 5138 return NOTIFY_OK; 5139 } 5140 5141 return NOTIFY_DONE; 5142} 5143 5144static int __init migration_init(void) 5145{ 5146 void *cpu = (void *)(long)smp_processor_id(); 5147 int err; 5148 5149 /* Initialize migration for the boot CPU */ 5150 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu); 5151 BUG_ON(err == NOTIFY_BAD); 5152 migration_call(&migration_notifier, CPU_ONLINE, cpu); 5153 register_cpu_notifier(&migration_notifier); 5154 5155 /* Register cpu active notifiers */ 5156 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE); 5157 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE); 5158 5159 return 0; 5160} 5161early_initcall(migration_init); 5162#endif 5163 5164#ifdef CONFIG_SMP 5165 5166static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */ 5167 5168#ifdef CONFIG_SCHED_DEBUG 5169 5170static __read_mostly int sched_debug_enabled; 5171 5172static int __init sched_debug_setup(char *str) 5173{ 5174 sched_debug_enabled = 1; 5175 5176 return 0; 5177} 5178early_param("sched_debug", sched_debug_setup); 5179 5180static inline bool sched_debug(void) 5181{ 5182 return sched_debug_enabled; 5183} 5184 5185static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level, 5186 struct cpumask *groupmask) 5187{ 5188 struct sched_group *group = sd->groups; 5189 char str[256]; 5190 5191 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd)); 5192 cpumask_clear(groupmask); 5193 5194 printk(KERN_DEBUG "%*s domain %d: ", level, "", level); 5195 5196 if (!(sd->flags & SD_LOAD_BALANCE)) { 5197 printk("does not load-balance\n"); 5198 if (sd->parent) 5199 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain" 5200 " has parent"); 5201 return -1; 5202 } 5203 5204 printk(KERN_CONT "span %s level %s\n", str, sd->name); 5205 5206 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) { 5207 printk(KERN_ERR "ERROR: domain->span does not contain " 5208 "CPU%d\n", cpu); 5209 } 5210 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) { 5211 printk(KERN_ERR "ERROR: domain->groups does not contain" 5212 " CPU%d\n", cpu); 5213 } 5214 5215 printk(KERN_DEBUG "%*s groups:", level + 1, ""); 5216 do { 5217 if (!group) { 5218 printk("\n"); 5219 printk(KERN_ERR "ERROR: group is NULL\n"); 5220 break; 5221 } 5222 5223 /* 5224 * Even though we initialize ->power to something semi-sane, 5225 * we leave power_orig unset. This allows us to detect if 5226 * domain iteration is still funny without causing /0 traps. 5227 */ 5228 if (!group->sgp->power_orig) { 5229 printk(KERN_CONT "\n"); 5230 printk(KERN_ERR "ERROR: domain->cpu_power not " 5231 "set\n"); 5232 break; 5233 } 5234 5235 if (!cpumask_weight(sched_group_cpus(group))) { 5236 printk(KERN_CONT "\n"); 5237 printk(KERN_ERR "ERROR: empty group\n"); 5238 break; 5239 } 5240 5241 if (!(sd->flags & SD_OVERLAP) && 5242 cpumask_intersects(groupmask, sched_group_cpus(group))) { 5243 printk(KERN_CONT "\n"); 5244 printk(KERN_ERR "ERROR: repeated CPUs\n"); 5245 break; 5246 } 5247 5248 cpumask_or(groupmask, groupmask, sched_group_cpus(group)); 5249 5250 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group)); 5251 5252 printk(KERN_CONT " %s", str); 5253 if (group->sgp->power != SCHED_POWER_SCALE) { 5254 printk(KERN_CONT " (cpu_power = %d)", 5255 group->sgp->power); 5256 } 5257 5258 group = group->next; 5259 } while (group != sd->groups); 5260 printk(KERN_CONT "\n"); 5261 5262 if (!cpumask_equal(sched_domain_span(sd), groupmask)) 5263 printk(KERN_ERR "ERROR: groups don't span domain->span\n"); 5264 5265 if (sd->parent && 5266 !cpumask_subset(groupmask, sched_domain_span(sd->parent))) 5267 printk(KERN_ERR "ERROR: parent span is not a superset " 5268 "of domain->span\n"); 5269 return 0; 5270} 5271 5272static void sched_domain_debug(struct sched_domain *sd, int cpu) 5273{ 5274 int level = 0; 5275 5276 if (!sched_debug_enabled) 5277 return; 5278 5279 if (!sd) { 5280 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu); 5281 return; 5282 } 5283 5284 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu); 5285 5286 for (;;) { 5287 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask)) 5288 break; 5289 level++; 5290 sd = sd->parent; 5291 if (!sd) 5292 break; 5293 } 5294} 5295#else /* !CONFIG_SCHED_DEBUG */ 5296# define sched_domain_debug(sd, cpu) do { } while (0) 5297static inline bool sched_debug(void) 5298{ 5299 return false; 5300} 5301#endif /* CONFIG_SCHED_DEBUG */ 5302 5303static int sd_degenerate(struct sched_domain *sd) 5304{ 5305 if (cpumask_weight(sched_domain_span(sd)) == 1) 5306 return 1; 5307 5308 /* Following flags need at least 2 groups */ 5309 if (sd->flags & (SD_LOAD_BALANCE | 5310 SD_BALANCE_NEWIDLE | 5311 SD_BALANCE_FORK | 5312 SD_BALANCE_EXEC | 5313 SD_SHARE_CPUPOWER | 5314 SD_SHARE_PKG_RESOURCES | 5315 SD_SHARE_POWERDOMAIN)) { 5316 if (sd->groups != sd->groups->next) 5317 return 0; 5318 } 5319 5320 /* Following flags don't use groups */ 5321 if (sd->flags & (SD_WAKE_AFFINE)) 5322 return 0; 5323 5324 return 1; 5325} 5326 5327static int 5328sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent) 5329{ 5330 unsigned long cflags = sd->flags, pflags = parent->flags; 5331 5332 if (sd_degenerate(parent)) 5333 return 1; 5334 5335 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent))) 5336 return 0; 5337 5338 /* Flags needing groups don't count if only 1 group in parent */ 5339 if (parent->groups == parent->groups->next) { 5340 pflags &= ~(SD_LOAD_BALANCE | 5341 SD_BALANCE_NEWIDLE | 5342 SD_BALANCE_FORK | 5343 SD_BALANCE_EXEC | 5344 SD_SHARE_CPUPOWER | 5345 SD_SHARE_PKG_RESOURCES | 5346 SD_PREFER_SIBLING | 5347 SD_SHARE_POWERDOMAIN); 5348 if (nr_node_ids == 1) 5349 pflags &= ~SD_SERIALIZE; 5350 } 5351 if (~cflags & pflags) 5352 return 0; 5353 5354 return 1; 5355} 5356 5357static void free_rootdomain(struct rcu_head *rcu) 5358{ 5359 struct root_domain *rd = container_of(rcu, struct root_domain, rcu); 5360 5361 cpupri_cleanup(&rd->cpupri); 5362 cpudl_cleanup(&rd->cpudl); 5363 free_cpumask_var(rd->dlo_mask); 5364 free_cpumask_var(rd->rto_mask); 5365 free_cpumask_var(rd->online); 5366 free_cpumask_var(rd->span); 5367 kfree(rd); 5368} 5369 5370static void rq_attach_root(struct rq *rq, struct root_domain *rd) 5371{ 5372 struct root_domain *old_rd = NULL; 5373 unsigned long flags; 5374 5375 raw_spin_lock_irqsave(&rq->lock, flags); 5376 5377 if (rq->rd) { 5378 old_rd = rq->rd; 5379 5380 if (cpumask_test_cpu(rq->cpu, old_rd->online)) 5381 set_rq_offline(rq); 5382 5383 cpumask_clear_cpu(rq->cpu, old_rd->span); 5384 5385 /* 5386 * If we dont want to free the old_rd yet then 5387 * set old_rd to NULL to skip the freeing later 5388 * in this function: 5389 */ 5390 if (!atomic_dec_and_test(&old_rd->refcount)) 5391 old_rd = NULL; 5392 } 5393 5394 atomic_inc(&rd->refcount); 5395 rq->rd = rd; 5396 5397 cpumask_set_cpu(rq->cpu, rd->span); 5398 if (cpumask_test_cpu(rq->cpu, cpu_active_mask)) 5399 set_rq_online(rq); 5400 5401 raw_spin_unlock_irqrestore(&rq->lock, flags); 5402 5403 if (old_rd) 5404 call_rcu_sched(&old_rd->rcu, free_rootdomain); 5405} 5406 5407static int init_rootdomain(struct root_domain *rd) 5408{ 5409 memset(rd, 0, sizeof(*rd)); 5410 5411 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL)) 5412 goto out; 5413 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL)) 5414 goto free_span; 5415 if (!alloc_cpumask_var(&rd->dlo_mask, GFP_KERNEL)) 5416 goto free_online; 5417 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL)) 5418 goto free_dlo_mask; 5419 5420 init_dl_bw(&rd->dl_bw); 5421 if (cpudl_init(&rd->cpudl) != 0) 5422 goto free_dlo_mask; 5423 5424 if (cpupri_init(&rd->cpupri) != 0) 5425 goto free_rto_mask; 5426 return 0; 5427 5428free_rto_mask: 5429 free_cpumask_var(rd->rto_mask); 5430free_dlo_mask: 5431 free_cpumask_var(rd->dlo_mask); 5432free_online: 5433 free_cpumask_var(rd->online); 5434free_span: 5435 free_cpumask_var(rd->span); 5436out: 5437 return -ENOMEM; 5438} 5439 5440/* 5441 * By default the system creates a single root-domain with all cpus as 5442 * members (mimicking the global state we have today). 5443 */ 5444struct root_domain def_root_domain; 5445 5446static void init_defrootdomain(void) 5447{ 5448 init_rootdomain(&def_root_domain); 5449 5450 atomic_set(&def_root_domain.refcount, 1); 5451} 5452 5453static struct root_domain *alloc_rootdomain(void) 5454{ 5455 struct root_domain *rd; 5456 5457 rd = kmalloc(sizeof(*rd), GFP_KERNEL); 5458 if (!rd) 5459 return NULL; 5460 5461 if (init_rootdomain(rd) != 0) { 5462 kfree(rd); 5463 return NULL; 5464 } 5465 5466 return rd; 5467} 5468 5469static void free_sched_groups(struct sched_group *sg, int free_sgp) 5470{ 5471 struct sched_group *tmp, *first; 5472 5473 if (!sg) 5474 return; 5475 5476 first = sg; 5477 do { 5478 tmp = sg->next; 5479 5480 if (free_sgp && atomic_dec_and_test(&sg->sgp->ref)) 5481 kfree(sg->sgp); 5482 5483 kfree(sg); 5484 sg = tmp; 5485 } while (sg != first); 5486} 5487 5488static void free_sched_domain(struct rcu_head *rcu) 5489{ 5490 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu); 5491 5492 /* 5493 * If its an overlapping domain it has private groups, iterate and 5494 * nuke them all. 5495 */ 5496 if (sd->flags & SD_OVERLAP) { 5497 free_sched_groups(sd->groups, 1); 5498 } else if (atomic_dec_and_test(&sd->groups->ref)) { 5499 kfree(sd->groups->sgp); 5500 kfree(sd->groups); 5501 } 5502 kfree(sd); 5503} 5504 5505static void destroy_sched_domain(struct sched_domain *sd, int cpu) 5506{ 5507 call_rcu(&sd->rcu, free_sched_domain); 5508} 5509 5510static void destroy_sched_domains(struct sched_domain *sd, int cpu) 5511{ 5512 for (; sd; sd = sd->parent) 5513 destroy_sched_domain(sd, cpu); 5514} 5515 5516/* 5517 * Keep a special pointer to the highest sched_domain that has 5518 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this 5519 * allows us to avoid some pointer chasing select_idle_sibling(). 5520 * 5521 * Also keep a unique ID per domain (we use the first cpu number in 5522 * the cpumask of the domain), this allows us to quickly tell if 5523 * two cpus are in the same cache domain, see cpus_share_cache(). 5524 */ 5525DEFINE_PER_CPU(struct sched_domain *, sd_llc); 5526DEFINE_PER_CPU(int, sd_llc_size); 5527DEFINE_PER_CPU(int, sd_llc_id); 5528DEFINE_PER_CPU(struct sched_domain *, sd_numa); 5529DEFINE_PER_CPU(struct sched_domain *, sd_busy); 5530DEFINE_PER_CPU(struct sched_domain *, sd_asym); 5531 5532static void update_top_cache_domain(int cpu) 5533{ 5534 struct sched_domain *sd; 5535 struct sched_domain *busy_sd = NULL; 5536 int id = cpu; 5537 int size = 1; 5538 5539 sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES); 5540 if (sd) { 5541 id = cpumask_first(sched_domain_span(sd)); 5542 size = cpumask_weight(sched_domain_span(sd)); 5543 busy_sd = sd->parent; /* sd_busy */ 5544 } 5545 rcu_assign_pointer(per_cpu(sd_busy, cpu), busy_sd); 5546 5547 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd); 5548 per_cpu(sd_llc_size, cpu) = size; 5549 per_cpu(sd_llc_id, cpu) = id; 5550 5551 sd = lowest_flag_domain(cpu, SD_NUMA); 5552 rcu_assign_pointer(per_cpu(sd_numa, cpu), sd); 5553 5554 sd = highest_flag_domain(cpu, SD_ASYM_PACKING); 5555 rcu_assign_pointer(per_cpu(sd_asym, cpu), sd); 5556} 5557 5558/* 5559 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must 5560 * hold the hotplug lock. 5561 */ 5562static void 5563cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu) 5564{ 5565 struct rq *rq = cpu_rq(cpu); 5566 struct sched_domain *tmp; 5567 5568 /* Remove the sched domains which do not contribute to scheduling. */ 5569 for (tmp = sd; tmp; ) { 5570 struct sched_domain *parent = tmp->parent; 5571 if (!parent) 5572 break; 5573 5574 if (sd_parent_degenerate(tmp, parent)) { 5575 tmp->parent = parent->parent; 5576 if (parent->parent) 5577 parent->parent->child = tmp; 5578 /* 5579 * Transfer SD_PREFER_SIBLING down in case of a 5580 * degenerate parent; the spans match for this 5581 * so the property transfers. 5582 */ 5583 if (parent->flags & SD_PREFER_SIBLING) 5584 tmp->flags |= SD_PREFER_SIBLING; 5585 destroy_sched_domain(parent, cpu); 5586 } else 5587 tmp = tmp->parent; 5588 } 5589 5590 if (sd && sd_degenerate(sd)) { 5591 tmp = sd; 5592 sd = sd->parent; 5593 destroy_sched_domain(tmp, cpu); 5594 if (sd) 5595 sd->child = NULL; 5596 } 5597 5598 sched_domain_debug(sd, cpu); 5599 5600 rq_attach_root(rq, rd); 5601 tmp = rq->sd; 5602 rcu_assign_pointer(rq->sd, sd); 5603 destroy_sched_domains(tmp, cpu); 5604 5605 update_top_cache_domain(cpu); 5606} 5607 5608/* cpus with isolated domains */ 5609static cpumask_var_t cpu_isolated_map; 5610 5611/* Setup the mask of cpus configured for isolated domains */ 5612static int __init isolated_cpu_setup(char *str) 5613{ 5614 alloc_bootmem_cpumask_var(&cpu_isolated_map); 5615 cpulist_parse(str, cpu_isolated_map); 5616 return 1; 5617} 5618 5619__setup("isolcpus=", isolated_cpu_setup); 5620 5621struct s_data { 5622 struct sched_domain ** __percpu sd; 5623 struct root_domain *rd; 5624}; 5625 5626enum s_alloc { 5627 sa_rootdomain, 5628 sa_sd, 5629 sa_sd_storage, 5630 sa_none, 5631}; 5632 5633/* 5634 * Build an iteration mask that can exclude certain CPUs from the upwards 5635 * domain traversal. 5636 * 5637 * Asymmetric node setups can result in situations where the domain tree is of 5638 * unequal depth, make sure to skip domains that already cover the entire 5639 * range. 5640 * 5641 * In that case build_sched_domains() will have terminated the iteration early 5642 * and our sibling sd spans will be empty. Domains should always include the 5643 * cpu they're built on, so check that. 5644 * 5645 */ 5646static void build_group_mask(struct sched_domain *sd, struct sched_group *sg) 5647{ 5648 const struct cpumask *span = sched_domain_span(sd); 5649 struct sd_data *sdd = sd->private; 5650 struct sched_domain *sibling; 5651 int i; 5652 5653 for_each_cpu(i, span) { 5654 sibling = *per_cpu_ptr(sdd->sd, i); 5655 if (!cpumask_test_cpu(i, sched_domain_span(sibling))) 5656 continue; 5657 5658 cpumask_set_cpu(i, sched_group_mask(sg)); 5659 } 5660} 5661 5662/* 5663 * Return the canonical balance cpu for this group, this is the first cpu 5664 * of this group that's also in the iteration mask. 5665 */ 5666int group_balance_cpu(struct sched_group *sg) 5667{ 5668 return cpumask_first_and(sched_group_cpus(sg), sched_group_mask(sg)); 5669} 5670 5671static int 5672build_overlap_sched_groups(struct sched_domain *sd, int cpu) 5673{ 5674 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg; 5675 const struct cpumask *span = sched_domain_span(sd); 5676 struct cpumask *covered = sched_domains_tmpmask; 5677 struct sd_data *sdd = sd->private; 5678 struct sched_domain *child; 5679 int i; 5680 5681 cpumask_clear(covered); 5682 5683 for_each_cpu(i, span) { 5684 struct cpumask *sg_span; 5685 5686 if (cpumask_test_cpu(i, covered)) 5687 continue; 5688 5689 child = *per_cpu_ptr(sdd->sd, i); 5690 5691 /* See the comment near build_group_mask(). */ 5692 if (!cpumask_test_cpu(i, sched_domain_span(child))) 5693 continue; 5694 5695 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(), 5696 GFP_KERNEL, cpu_to_node(cpu)); 5697 5698 if (!sg) 5699 goto fail; 5700 5701 sg_span = sched_group_cpus(sg); 5702 if (child->child) { 5703 child = child->child; 5704 cpumask_copy(sg_span, sched_domain_span(child)); 5705 } else 5706 cpumask_set_cpu(i, sg_span); 5707 5708 cpumask_or(covered, covered, sg_span); 5709 5710 sg->sgp = *per_cpu_ptr(sdd->sgp, i); 5711 if (atomic_inc_return(&sg->sgp->ref) == 1) 5712 build_group_mask(sd, sg); 5713 5714 /* 5715 * Initialize sgp->power such that even if we mess up the 5716 * domains and no possible iteration will get us here, we won't 5717 * die on a /0 trap. 5718 */ 5719 sg->sgp->power = SCHED_POWER_SCALE * cpumask_weight(sg_span); 5720 sg->sgp->power_orig = sg->sgp->power; 5721 5722 /* 5723 * Make sure the first group of this domain contains the 5724 * canonical balance cpu. Otherwise the sched_domain iteration 5725 * breaks. See update_sg_lb_stats(). 5726 */ 5727 if ((!groups && cpumask_test_cpu(cpu, sg_span)) || 5728 group_balance_cpu(sg) == cpu) 5729 groups = sg; 5730 5731 if (!first) 5732 first = sg; 5733 if (last) 5734 last->next = sg; 5735 last = sg; 5736 last->next = first; 5737 } 5738 sd->groups = groups; 5739 5740 return 0; 5741 5742fail: 5743 free_sched_groups(first, 0); 5744 5745 return -ENOMEM; 5746} 5747 5748static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg) 5749{ 5750 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu); 5751 struct sched_domain *child = sd->child; 5752 5753 if (child) 5754 cpu = cpumask_first(sched_domain_span(child)); 5755 5756 if (sg) { 5757 *sg = *per_cpu_ptr(sdd->sg, cpu); 5758 (*sg)->sgp = *per_cpu_ptr(sdd->sgp, cpu); 5759 atomic_set(&(*sg)->sgp->ref, 1); /* for claim_allocations */ 5760 } 5761 5762 return cpu; 5763} 5764 5765/* 5766 * build_sched_groups will build a circular linked list of the groups 5767 * covered by the given span, and will set each group's ->cpumask correctly, 5768 * and ->cpu_power to 0. 5769 * 5770 * Assumes the sched_domain tree is fully constructed 5771 */ 5772static int 5773build_sched_groups(struct sched_domain *sd, int cpu) 5774{ 5775 struct sched_group *first = NULL, *last = NULL; 5776 struct sd_data *sdd = sd->private; 5777 const struct cpumask *span = sched_domain_span(sd); 5778 struct cpumask *covered; 5779 int i; 5780 5781 get_group(cpu, sdd, &sd->groups); 5782 atomic_inc(&sd->groups->ref); 5783 5784 if (cpu != cpumask_first(span)) 5785 return 0; 5786 5787 lockdep_assert_held(&sched_domains_mutex); 5788 covered = sched_domains_tmpmask; 5789 5790 cpumask_clear(covered); 5791 5792 for_each_cpu(i, span) { 5793 struct sched_group *sg; 5794 int group, j; 5795 5796 if (cpumask_test_cpu(i, covered)) 5797 continue; 5798 5799 group = get_group(i, sdd, &sg); 5800 cpumask_setall(sched_group_mask(sg)); 5801 5802 for_each_cpu(j, span) { 5803 if (get_group(j, sdd, NULL) != group) 5804 continue; 5805 5806 cpumask_set_cpu(j, covered); 5807 cpumask_set_cpu(j, sched_group_cpus(sg)); 5808 } 5809 5810 if (!first) 5811 first = sg; 5812 if (last) 5813 last->next = sg; 5814 last = sg; 5815 } 5816 last->next = first; 5817 5818 return 0; 5819} 5820 5821/* 5822 * Initialize sched groups cpu_power. 5823 * 5824 * cpu_power indicates the capacity of sched group, which is used while 5825 * distributing the load between different sched groups in a sched domain. 5826 * Typically cpu_power for all the groups in a sched domain will be same unless 5827 * there are asymmetries in the topology. If there are asymmetries, group 5828 * having more cpu_power will pickup more load compared to the group having 5829 * less cpu_power. 5830 */ 5831static void init_sched_groups_power(int cpu, struct sched_domain *sd) 5832{ 5833 struct sched_group *sg = sd->groups; 5834 5835 WARN_ON(!sg); 5836 5837 do { 5838 sg->group_weight = cpumask_weight(sched_group_cpus(sg)); 5839 sg = sg->next; 5840 } while (sg != sd->groups); 5841 5842 if (cpu != group_balance_cpu(sg)) 5843 return; 5844 5845 update_group_power(sd, cpu); 5846 atomic_set(&sg->sgp->nr_busy_cpus, sg->group_weight); 5847} 5848 5849/* 5850 * Initializers for schedule domains 5851 * Non-inlined to reduce accumulated stack pressure in build_sched_domains() 5852 */ 5853 5854static int default_relax_domain_level = -1; 5855int sched_domain_level_max; 5856 5857static int __init setup_relax_domain_level(char *str) 5858{ 5859 if (kstrtoint(str, 0, &default_relax_domain_level)) 5860 pr_warn("Unable to set relax_domain_level\n"); 5861 5862 return 1; 5863} 5864__setup("relax_domain_level=", setup_relax_domain_level); 5865 5866static void set_domain_attribute(struct sched_domain *sd, 5867 struct sched_domain_attr *attr) 5868{ 5869 int request; 5870 5871 if (!attr || attr->relax_domain_level < 0) { 5872 if (default_relax_domain_level < 0) 5873 return; 5874 else 5875 request = default_relax_domain_level; 5876 } else 5877 request = attr->relax_domain_level; 5878 if (request < sd->level) { 5879 /* turn off idle balance on this domain */ 5880 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE); 5881 } else { 5882 /* turn on idle balance on this domain */ 5883 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE); 5884 } 5885} 5886 5887static void __sdt_free(const struct cpumask *cpu_map); 5888static int __sdt_alloc(const struct cpumask *cpu_map); 5889 5890static void __free_domain_allocs(struct s_data *d, enum s_alloc what, 5891 const struct cpumask *cpu_map) 5892{ 5893 switch (what) { 5894 case sa_rootdomain: 5895 if (!atomic_read(&d->rd->refcount)) 5896 free_rootdomain(&d->rd->rcu); /* fall through */ 5897 case sa_sd: 5898 free_percpu(d->sd); /* fall through */ 5899 case sa_sd_storage: 5900 __sdt_free(cpu_map); /* fall through */ 5901 case sa_none: 5902 break; 5903 } 5904} 5905 5906static enum s_alloc __visit_domain_allocation_hell(struct s_data *d, 5907 const struct cpumask *cpu_map) 5908{ 5909 memset(d, 0, sizeof(*d)); 5910 5911 if (__sdt_alloc(cpu_map)) 5912 return sa_sd_storage; 5913 d->sd = alloc_percpu(struct sched_domain *); 5914 if (!d->sd) 5915 return sa_sd_storage; 5916 d->rd = alloc_rootdomain(); 5917 if (!d->rd) 5918 return sa_sd; 5919 return sa_rootdomain; 5920} 5921 5922/* 5923 * NULL the sd_data elements we've used to build the sched_domain and 5924 * sched_group structure so that the subsequent __free_domain_allocs() 5925 * will not free the data we're using. 5926 */ 5927static void claim_allocations(int cpu, struct sched_domain *sd) 5928{ 5929 struct sd_data *sdd = sd->private; 5930 5931 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd); 5932 *per_cpu_ptr(sdd->sd, cpu) = NULL; 5933 5934 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref)) 5935 *per_cpu_ptr(sdd->sg, cpu) = NULL; 5936 5937 if (atomic_read(&(*per_cpu_ptr(sdd->sgp, cpu))->ref)) 5938 *per_cpu_ptr(sdd->sgp, cpu) = NULL; 5939} 5940 5941#ifdef CONFIG_NUMA 5942static int sched_domains_numa_levels; 5943static int *sched_domains_numa_distance; 5944static struct cpumask ***sched_domains_numa_masks; 5945static int sched_domains_curr_level; 5946#endif 5947 5948/* 5949 * SD_flags allowed in topology descriptions. 5950 * 5951 * SD_SHARE_CPUPOWER - describes SMT topologies 5952 * SD_SHARE_PKG_RESOURCES - describes shared caches 5953 * SD_NUMA - describes NUMA topologies 5954 * SD_SHARE_POWERDOMAIN - describes shared power domain 5955 * 5956 * Odd one out: 5957 * SD_ASYM_PACKING - describes SMT quirks 5958 */ 5959#define TOPOLOGY_SD_FLAGS \ 5960 (SD_SHARE_CPUPOWER | \ 5961 SD_SHARE_PKG_RESOURCES | \ 5962 SD_NUMA | \ 5963 SD_ASYM_PACKING | \ 5964 SD_SHARE_POWERDOMAIN) 5965 5966static struct sched_domain * 5967sd_init(struct sched_domain_topology_level *tl, int cpu) 5968{ 5969 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); 5970 int sd_weight, sd_flags = 0; 5971 5972#ifdef CONFIG_NUMA 5973 /* 5974 * Ugly hack to pass state to sd_numa_mask()... 5975 */ 5976 sched_domains_curr_level = tl->numa_level; 5977#endif 5978 5979 sd_weight = cpumask_weight(tl->mask(cpu)); 5980 5981 if (tl->sd_flags) 5982 sd_flags = (*tl->sd_flags)(); 5983 if (WARN_ONCE(sd_flags & ~TOPOLOGY_SD_FLAGS, 5984 "wrong sd_flags in topology description\n")) 5985 sd_flags &= ~TOPOLOGY_SD_FLAGS; 5986 5987 *sd = (struct sched_domain){ 5988 .min_interval = sd_weight, 5989 .max_interval = 2*sd_weight, 5990 .busy_factor = 32, 5991 .imbalance_pct = 125, 5992 5993 .cache_nice_tries = 0, 5994 .busy_idx = 0, 5995 .idle_idx = 0, 5996 .newidle_idx = 0, 5997 .wake_idx = 0, 5998 .forkexec_idx = 0, 5999 6000 .flags = 1*SD_LOAD_BALANCE 6001 | 1*SD_BALANCE_NEWIDLE 6002 | 1*SD_BALANCE_EXEC 6003 | 1*SD_BALANCE_FORK 6004 | 0*SD_BALANCE_WAKE 6005 | 1*SD_WAKE_AFFINE 6006 | 0*SD_SHARE_CPUPOWER 6007 | 0*SD_SHARE_PKG_RESOURCES 6008 | 0*SD_SERIALIZE 6009 | 0*SD_PREFER_SIBLING 6010 | 0*SD_NUMA 6011 | sd_flags 6012 , 6013 6014 .last_balance = jiffies, 6015 .balance_interval = sd_weight, 6016 .smt_gain = 0, 6017 .max_newidle_lb_cost = 0, 6018 .next_decay_max_lb_cost = jiffies, 6019#ifdef CONFIG_SCHED_DEBUG 6020 .name = tl->name, 6021#endif 6022 }; 6023 6024 /* 6025 * Convert topological properties into behaviour. 6026 */ 6027 6028 if (sd->flags & SD_SHARE_CPUPOWER) { 6029 sd->imbalance_pct = 110; 6030 sd->smt_gain = 1178; /* ~15% */ 6031 6032 } else if (sd->flags & SD_SHARE_PKG_RESOURCES) { 6033 sd->imbalance_pct = 117; 6034 sd->cache_nice_tries = 1; 6035 sd->busy_idx = 2; 6036 6037#ifdef CONFIG_NUMA 6038 } else if (sd->flags & SD_NUMA) { 6039 sd->cache_nice_tries = 2; 6040 sd->busy_idx = 3; 6041 sd->idle_idx = 2; 6042 6043 sd->flags |= SD_SERIALIZE; 6044 if (sched_domains_numa_distance[tl->numa_level] > RECLAIM_DISTANCE) { 6045 sd->flags &= ~(SD_BALANCE_EXEC | 6046 SD_BALANCE_FORK | 6047 SD_WAKE_AFFINE); 6048 } 6049 6050#endif 6051 } else { 6052 sd->flags |= SD_PREFER_SIBLING; 6053 sd->cache_nice_tries = 1; 6054 sd->busy_idx = 2; 6055 sd->idle_idx = 1; 6056 } 6057 6058 sd->private = &tl->data; 6059 6060 return sd; 6061} 6062 6063/* 6064 * Topology list, bottom-up. 6065 */ 6066static struct sched_domain_topology_level default_topology[] = { 6067#ifdef CONFIG_SCHED_SMT 6068 { cpu_smt_mask, cpu_smt_flags, SD_INIT_NAME(SMT) }, 6069#endif 6070#ifdef CONFIG_SCHED_MC 6071 { cpu_coregroup_mask, cpu_core_flags, SD_INIT_NAME(MC) }, 6072#endif 6073 { cpu_cpu_mask, SD_INIT_NAME(DIE) }, 6074 { NULL, }, 6075}; 6076 6077struct sched_domain_topology_level *sched_domain_topology = default_topology; 6078 6079#define for_each_sd_topology(tl) \ 6080 for (tl = sched_domain_topology; tl->mask; tl++) 6081 6082void set_sched_topology(struct sched_domain_topology_level *tl) 6083{ 6084 sched_domain_topology = tl; 6085} 6086 6087#ifdef CONFIG_NUMA 6088 6089static const struct cpumask *sd_numa_mask(int cpu) 6090{ 6091 return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)]; 6092} 6093 6094static void sched_numa_warn(const char *str) 6095{ 6096 static int done = false; 6097 int i,j; 6098 6099 if (done) 6100 return; 6101 6102 done = true; 6103 6104 printk(KERN_WARNING "ERROR: %s\n\n", str); 6105 6106 for (i = 0; i < nr_node_ids; i++) { 6107 printk(KERN_WARNING " "); 6108 for (j = 0; j < nr_node_ids; j++) 6109 printk(KERN_CONT "%02d ", node_distance(i,j)); 6110 printk(KERN_CONT "\n"); 6111 } 6112 printk(KERN_WARNING "\n"); 6113} 6114 6115static bool find_numa_distance(int distance) 6116{ 6117 int i; 6118 6119 if (distance == node_distance(0, 0)) 6120 return true; 6121 6122 for (i = 0; i < sched_domains_numa_levels; i++) { 6123 if (sched_domains_numa_distance[i] == distance) 6124 return true; 6125 } 6126 6127 return false; 6128} 6129 6130static void sched_init_numa(void) 6131{ 6132 int next_distance, curr_distance = node_distance(0, 0); 6133 struct sched_domain_topology_level *tl; 6134 int level = 0; 6135 int i, j, k; 6136 6137 sched_domains_numa_distance = kzalloc(sizeof(int) * nr_node_ids, GFP_KERNEL); 6138 if (!sched_domains_numa_distance) 6139 return; 6140 6141 /* 6142 * O(nr_nodes^2) deduplicating selection sort -- in order to find the 6143 * unique distances in the node_distance() table. 6144 * 6145 * Assumes node_distance(0,j) includes all distances in 6146 * node_distance(i,j) in order to avoid cubic time. 6147 */ 6148 next_distance = curr_distance; 6149 for (i = 0; i < nr_node_ids; i++) { 6150 for (j = 0; j < nr_node_ids; j++) { 6151 for (k = 0; k < nr_node_ids; k++) { 6152 int distance = node_distance(i, k); 6153 6154 if (distance > curr_distance && 6155 (distance < next_distance || 6156 next_distance == curr_distance)) 6157 next_distance = distance; 6158 6159 /* 6160 * While not a strong assumption it would be nice to know 6161 * about cases where if node A is connected to B, B is not 6162 * equally connected to A. 6163 */ 6164 if (sched_debug() && node_distance(k, i) != distance) 6165 sched_numa_warn("Node-distance not symmetric"); 6166 6167 if (sched_debug() && i && !find_numa_distance(distance)) 6168 sched_numa_warn("Node-0 not representative"); 6169 } 6170 if (next_distance != curr_distance) { 6171 sched_domains_numa_distance[level++] = next_distance; 6172 sched_domains_numa_levels = level; 6173 curr_distance = next_distance; 6174 } else break; 6175 } 6176 6177 /* 6178 * In case of sched_debug() we verify the above assumption. 6179 */ 6180 if (!sched_debug()) 6181 break; 6182 } 6183 /* 6184 * 'level' contains the number of unique distances, excluding the 6185 * identity distance node_distance(i,i). 6186 * 6187 * The sched_domains_numa_distance[] array includes the actual distance 6188 * numbers. 6189 */ 6190 6191 /* 6192 * Here, we should temporarily reset sched_domains_numa_levels to 0. 6193 * If it fails to allocate memory for array sched_domains_numa_masks[][], 6194 * the array will contain less then 'level' members. This could be 6195 * dangerous when we use it to iterate array sched_domains_numa_masks[][] 6196 * in other functions. 6197 * 6198 * We reset it to 'level' at the end of this function. 6199 */ 6200 sched_domains_numa_levels = 0; 6201 6202 sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL); 6203 if (!sched_domains_numa_masks) 6204 return; 6205 6206 /* 6207 * Now for each level, construct a mask per node which contains all 6208 * cpus of nodes that are that many hops away from us. 6209 */ 6210 for (i = 0; i < level; i++) { 6211 sched_domains_numa_masks[i] = 6212 kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL); 6213 if (!sched_domains_numa_masks[i]) 6214 return; 6215 6216 for (j = 0; j < nr_node_ids; j++) { 6217 struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL); 6218 if (!mask) 6219 return; 6220 6221 sched_domains_numa_masks[i][j] = mask; 6222 6223 for (k = 0; k < nr_node_ids; k++) { 6224 if (node_distance(j, k) > sched_domains_numa_distance[i]) 6225 continue; 6226 6227 cpumask_or(mask, mask, cpumask_of_node(k)); 6228 } 6229 } 6230 } 6231 6232 /* Compute default topology size */ 6233 for (i = 0; sched_domain_topology[i].mask; i++); 6234 6235 tl = kzalloc((i + level + 1) * 6236 sizeof(struct sched_domain_topology_level), GFP_KERNEL); 6237 if (!tl) 6238 return; 6239 6240 /* 6241 * Copy the default topology bits.. 6242 */ 6243 for (i = 0; sched_domain_topology[i].mask; i++) 6244 tl[i] = sched_domain_topology[i]; 6245 6246 /* 6247 * .. and append 'j' levels of NUMA goodness. 6248 */ 6249 for (j = 0; j < level; i++, j++) { 6250 tl[i] = (struct sched_domain_topology_level){ 6251 .mask = sd_numa_mask, 6252 .sd_flags = cpu_numa_flags, 6253 .flags = SDTL_OVERLAP, 6254 .numa_level = j, 6255 SD_INIT_NAME(NUMA) 6256 }; 6257 } 6258 6259 sched_domain_topology = tl; 6260 6261 sched_domains_numa_levels = level; 6262} 6263 6264static void sched_domains_numa_masks_set(int cpu) 6265{ 6266 int i, j; 6267 int node = cpu_to_node(cpu); 6268 6269 for (i = 0; i < sched_domains_numa_levels; i++) { 6270 for (j = 0; j < nr_node_ids; j++) { 6271 if (node_distance(j, node) <= sched_domains_numa_distance[i]) 6272 cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]); 6273 } 6274 } 6275} 6276 6277static void sched_domains_numa_masks_clear(int cpu) 6278{ 6279 int i, j; 6280 for (i = 0; i < sched_domains_numa_levels; i++) { 6281 for (j = 0; j < nr_node_ids; j++) 6282 cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]); 6283 } 6284} 6285 6286/* 6287 * Update sched_domains_numa_masks[level][node] array when new cpus 6288 * are onlined. 6289 */ 6290static int sched_domains_numa_masks_update(struct notifier_block *nfb, 6291 unsigned long action, 6292 void *hcpu) 6293{ 6294 int cpu = (long)hcpu; 6295 6296 switch (action & ~CPU_TASKS_FROZEN) { 6297 case CPU_ONLINE: 6298 sched_domains_numa_masks_set(cpu); 6299 break; 6300 6301 case CPU_DEAD: 6302 sched_domains_numa_masks_clear(cpu); 6303 break; 6304 6305 default: 6306 return NOTIFY_DONE; 6307 } 6308 6309 return NOTIFY_OK; 6310} 6311#else 6312static inline void sched_init_numa(void) 6313{ 6314} 6315 6316static int sched_domains_numa_masks_update(struct notifier_block *nfb, 6317 unsigned long action, 6318 void *hcpu) 6319{ 6320 return 0; 6321} 6322#endif /* CONFIG_NUMA */ 6323 6324static int __sdt_alloc(const struct cpumask *cpu_map) 6325{ 6326 struct sched_domain_topology_level *tl; 6327 int j; 6328 6329 for_each_sd_topology(tl) { 6330 struct sd_data *sdd = &tl->data; 6331 6332 sdd->sd = alloc_percpu(struct sched_domain *); 6333 if (!sdd->sd) 6334 return -ENOMEM; 6335 6336 sdd->sg = alloc_percpu(struct sched_group *); 6337 if (!sdd->sg) 6338 return -ENOMEM; 6339 6340 sdd->sgp = alloc_percpu(struct sched_group_power *); 6341 if (!sdd->sgp) 6342 return -ENOMEM; 6343 6344 for_each_cpu(j, cpu_map) { 6345 struct sched_domain *sd; 6346 struct sched_group *sg; 6347 struct sched_group_power *sgp; 6348 6349 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(), 6350 GFP_KERNEL, cpu_to_node(j)); 6351 if (!sd) 6352 return -ENOMEM; 6353 6354 *per_cpu_ptr(sdd->sd, j) = sd; 6355 6356 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(), 6357 GFP_KERNEL, cpu_to_node(j)); 6358 if (!sg) 6359 return -ENOMEM; 6360 6361 sg->next = sg; 6362 6363 *per_cpu_ptr(sdd->sg, j) = sg; 6364 6365 sgp = kzalloc_node(sizeof(struct sched_group_power) + cpumask_size(), 6366 GFP_KERNEL, cpu_to_node(j)); 6367 if (!sgp) 6368 return -ENOMEM; 6369 6370 *per_cpu_ptr(sdd->sgp, j) = sgp; 6371 } 6372 } 6373 6374 return 0; 6375} 6376 6377static void __sdt_free(const struct cpumask *cpu_map) 6378{ 6379 struct sched_domain_topology_level *tl; 6380 int j; 6381 6382 for_each_sd_topology(tl) { 6383 struct sd_data *sdd = &tl->data; 6384 6385 for_each_cpu(j, cpu_map) { 6386 struct sched_domain *sd; 6387 6388 if (sdd->sd) { 6389 sd = *per_cpu_ptr(sdd->sd, j); 6390 if (sd && (sd->flags & SD_OVERLAP)) 6391 free_sched_groups(sd->groups, 0); 6392 kfree(*per_cpu_ptr(sdd->sd, j)); 6393 } 6394 6395 if (sdd->sg) 6396 kfree(*per_cpu_ptr(sdd->sg, j)); 6397 if (sdd->sgp) 6398 kfree(*per_cpu_ptr(sdd->sgp, j)); 6399 } 6400 free_percpu(sdd->sd); 6401 sdd->sd = NULL; 6402 free_percpu(sdd->sg); 6403 sdd->sg = NULL; 6404 free_percpu(sdd->sgp); 6405 sdd->sgp = NULL; 6406 } 6407} 6408 6409struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl, 6410 const struct cpumask *cpu_map, struct sched_domain_attr *attr, 6411 struct sched_domain *child, int cpu) 6412{ 6413 struct sched_domain *sd = sd_init(tl, cpu); 6414 if (!sd) 6415 return child; 6416 6417 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu)); 6418 if (child) { 6419 sd->level = child->level + 1; 6420 sched_domain_level_max = max(sched_domain_level_max, sd->level); 6421 child->parent = sd; 6422 sd->child = child; 6423 } 6424 set_domain_attribute(sd, attr); 6425 6426 return sd; 6427} 6428 6429/* 6430 * Build sched domains for a given set of cpus and attach the sched domains 6431 * to the individual cpus 6432 */ 6433static int build_sched_domains(const struct cpumask *cpu_map, 6434 struct sched_domain_attr *attr) 6435{ 6436 enum s_alloc alloc_state; 6437 struct sched_domain *sd; 6438 struct s_data d; 6439 int i, ret = -ENOMEM; 6440 6441 alloc_state = __visit_domain_allocation_hell(&d, cpu_map); 6442 if (alloc_state != sa_rootdomain) 6443 goto error; 6444 6445 /* Set up domains for cpus specified by the cpu_map. */ 6446 for_each_cpu(i, cpu_map) { 6447 struct sched_domain_topology_level *tl; 6448 6449 sd = NULL; 6450 for_each_sd_topology(tl) { 6451 sd = build_sched_domain(tl, cpu_map, attr, sd, i); 6452 if (tl == sched_domain_topology) 6453 *per_cpu_ptr(d.sd, i) = sd; 6454 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP)) 6455 sd->flags |= SD_OVERLAP; 6456 if (cpumask_equal(cpu_map, sched_domain_span(sd))) 6457 break; 6458 } 6459 } 6460 6461 /* Build the groups for the domains */ 6462 for_each_cpu(i, cpu_map) { 6463 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) { 6464 sd->span_weight = cpumask_weight(sched_domain_span(sd)); 6465 if (sd->flags & SD_OVERLAP) { 6466 if (build_overlap_sched_groups(sd, i)) 6467 goto error; 6468 } else { 6469 if (build_sched_groups(sd, i)) 6470 goto error; 6471 } 6472 } 6473 } 6474 6475 /* Calculate CPU power for physical packages and nodes */ 6476 for (i = nr_cpumask_bits-1; i >= 0; i--) { 6477 if (!cpumask_test_cpu(i, cpu_map)) 6478 continue; 6479 6480 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) { 6481 claim_allocations(i, sd); 6482 init_sched_groups_power(i, sd); 6483 } 6484 } 6485 6486 /* Attach the domains */ 6487 rcu_read_lock(); 6488 for_each_cpu(i, cpu_map) { 6489 sd = *per_cpu_ptr(d.sd, i); 6490 cpu_attach_domain(sd, d.rd, i); 6491 } 6492 rcu_read_unlock(); 6493 6494 ret = 0; 6495error: 6496 __free_domain_allocs(&d, alloc_state, cpu_map); 6497 return ret; 6498} 6499 6500static cpumask_var_t *doms_cur; /* current sched domains */ 6501static int ndoms_cur; /* number of sched domains in 'doms_cur' */ 6502static struct sched_domain_attr *dattr_cur; 6503 /* attribues of custom domains in 'doms_cur' */ 6504 6505/* 6506 * Special case: If a kmalloc of a doms_cur partition (array of 6507 * cpumask) fails, then fallback to a single sched domain, 6508 * as determined by the single cpumask fallback_doms. 6509 */ 6510static cpumask_var_t fallback_doms; 6511 6512/* 6513 * arch_update_cpu_topology lets virtualized architectures update the 6514 * cpu core maps. It is supposed to return 1 if the topology changed 6515 * or 0 if it stayed the same. 6516 */ 6517int __weak arch_update_cpu_topology(void) 6518{ 6519 return 0; 6520} 6521 6522cpumask_var_t *alloc_sched_domains(unsigned int ndoms) 6523{ 6524 int i; 6525 cpumask_var_t *doms; 6526 6527 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL); 6528 if (!doms) 6529 return NULL; 6530 for (i = 0; i < ndoms; i++) { 6531 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) { 6532 free_sched_domains(doms, i); 6533 return NULL; 6534 } 6535 } 6536 return doms; 6537} 6538 6539void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms) 6540{ 6541 unsigned int i; 6542 for (i = 0; i < ndoms; i++) 6543 free_cpumask_var(doms[i]); 6544 kfree(doms); 6545} 6546 6547/* 6548 * Set up scheduler domains and groups. Callers must hold the hotplug lock. 6549 * For now this just excludes isolated cpus, but could be used to 6550 * exclude other special cases in the future. 6551 */ 6552static int init_sched_domains(const struct cpumask *cpu_map) 6553{ 6554 int err; 6555 6556 arch_update_cpu_topology(); 6557 ndoms_cur = 1; 6558 doms_cur = alloc_sched_domains(ndoms_cur); 6559 if (!doms_cur) 6560 doms_cur = &fallback_doms; 6561 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map); 6562 err = build_sched_domains(doms_cur[0], NULL); 6563 register_sched_domain_sysctl(); 6564 6565 return err; 6566} 6567 6568/* 6569 * Detach sched domains from a group of cpus specified in cpu_map 6570 * These cpus will now be attached to the NULL domain 6571 */ 6572static void detach_destroy_domains(const struct cpumask *cpu_map) 6573{ 6574 int i; 6575 6576 rcu_read_lock(); 6577 for_each_cpu(i, cpu_map) 6578 cpu_attach_domain(NULL, &def_root_domain, i); 6579 rcu_read_unlock(); 6580} 6581 6582/* handle null as "default" */ 6583static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur, 6584 struct sched_domain_attr *new, int idx_new) 6585{ 6586 struct sched_domain_attr tmp; 6587 6588 /* fast path */ 6589 if (!new && !cur) 6590 return 1; 6591 6592 tmp = SD_ATTR_INIT; 6593 return !memcmp(cur ? (cur + idx_cur) : &tmp, 6594 new ? (new + idx_new) : &tmp, 6595 sizeof(struct sched_domain_attr)); 6596} 6597 6598/* 6599 * Partition sched domains as specified by the 'ndoms_new' 6600 * cpumasks in the array doms_new[] of cpumasks. This compares 6601 * doms_new[] to the current sched domain partitioning, doms_cur[]. 6602 * It destroys each deleted domain and builds each new domain. 6603 * 6604 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'. 6605 * The masks don't intersect (don't overlap.) We should setup one 6606 * sched domain for each mask. CPUs not in any of the cpumasks will 6607 * not be load balanced. If the same cpumask appears both in the 6608 * current 'doms_cur' domains and in the new 'doms_new', we can leave 6609 * it as it is. 6610 * 6611 * The passed in 'doms_new' should be allocated using 6612 * alloc_sched_domains. This routine takes ownership of it and will 6613 * free_sched_domains it when done with it. If the caller failed the 6614 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1, 6615 * and partition_sched_domains() will fallback to the single partition 6616 * 'fallback_doms', it also forces the domains to be rebuilt. 6617 * 6618 * If doms_new == NULL it will be replaced with cpu_online_mask. 6619 * ndoms_new == 0 is a special case for destroying existing domains, 6620 * and it will not create the default domain. 6621 * 6622 * Call with hotplug lock held 6623 */ 6624void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[], 6625 struct sched_domain_attr *dattr_new) 6626{ 6627 int i, j, n; 6628 int new_topology; 6629 6630 mutex_lock(&sched_domains_mutex); 6631 6632 /* always unregister in case we don't destroy any domains */ 6633 unregister_sched_domain_sysctl(); 6634 6635 /* Let architecture update cpu core mappings. */ 6636 new_topology = arch_update_cpu_topology(); 6637 6638 n = doms_new ? ndoms_new : 0; 6639 6640 /* Destroy deleted domains */ 6641 for (i = 0; i < ndoms_cur; i++) { 6642 for (j = 0; j < n && !new_topology; j++) { 6643 if (cpumask_equal(doms_cur[i], doms_new[j]) 6644 && dattrs_equal(dattr_cur, i, dattr_new, j)) 6645 goto match1; 6646 } 6647 /* no match - a current sched domain not in new doms_new[] */ 6648 detach_destroy_domains(doms_cur[i]); 6649match1: 6650 ; 6651 } 6652 6653 n = ndoms_cur; 6654 if (doms_new == NULL) { 6655 n = 0; 6656 doms_new = &fallback_doms; 6657 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map); 6658 WARN_ON_ONCE(dattr_new); 6659 } 6660 6661 /* Build new domains */ 6662 for (i = 0; i < ndoms_new; i++) { 6663 for (j = 0; j < n && !new_topology; j++) { 6664 if (cpumask_equal(doms_new[i], doms_cur[j]) 6665 && dattrs_equal(dattr_new, i, dattr_cur, j)) 6666 goto match2; 6667 } 6668 /* no match - add a new doms_new */ 6669 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL); 6670match2: 6671 ; 6672 } 6673 6674 /* Remember the new sched domains */ 6675 if (doms_cur != &fallback_doms) 6676 free_sched_domains(doms_cur, ndoms_cur); 6677 kfree(dattr_cur); /* kfree(NULL) is safe */ 6678 doms_cur = doms_new; 6679 dattr_cur = dattr_new; 6680 ndoms_cur = ndoms_new; 6681 6682 register_sched_domain_sysctl(); 6683 6684 mutex_unlock(&sched_domains_mutex); 6685} 6686 6687static int num_cpus_frozen; /* used to mark begin/end of suspend/resume */ 6688 6689/* 6690 * Update cpusets according to cpu_active mask. If cpusets are 6691 * disabled, cpuset_update_active_cpus() becomes a simple wrapper 6692 * around partition_sched_domains(). 6693 * 6694 * If we come here as part of a suspend/resume, don't touch cpusets because we 6695 * want to restore it back to its original state upon resume anyway. 6696 */ 6697static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action, 6698 void *hcpu) 6699{ 6700 switch (action) { 6701 case CPU_ONLINE_FROZEN: 6702 case CPU_DOWN_FAILED_FROZEN: 6703 6704 /* 6705 * num_cpus_frozen tracks how many CPUs are involved in suspend 6706 * resume sequence. As long as this is not the last online 6707 * operation in the resume sequence, just build a single sched 6708 * domain, ignoring cpusets. 6709 */ 6710 num_cpus_frozen--; 6711 if (likely(num_cpus_frozen)) { 6712 partition_sched_domains(1, NULL, NULL); 6713 break; 6714 } 6715 6716 /* 6717 * This is the last CPU online operation. So fall through and 6718 * restore the original sched domains by considering the 6719 * cpuset configurations. 6720 */ 6721 6722 case CPU_ONLINE: 6723 case CPU_DOWN_FAILED: 6724 cpuset_update_active_cpus(true); 6725 break; 6726 default: 6727 return NOTIFY_DONE; 6728 } 6729 return NOTIFY_OK; 6730} 6731 6732static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action, 6733 void *hcpu) 6734{ 6735 switch (action) { 6736 case CPU_DOWN_PREPARE: 6737 cpuset_update_active_cpus(false); 6738 break; 6739 case CPU_DOWN_PREPARE_FROZEN: 6740 num_cpus_frozen++; 6741 partition_sched_domains(1, NULL, NULL); 6742 break; 6743 default: 6744 return NOTIFY_DONE; 6745 } 6746 return NOTIFY_OK; 6747} 6748 6749void __init sched_init_smp(void) 6750{ 6751 cpumask_var_t non_isolated_cpus; 6752 6753 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL); 6754 alloc_cpumask_var(&fallback_doms, GFP_KERNEL); 6755 6756 sched_init_numa(); 6757 6758 /* 6759 * There's no userspace yet to cause hotplug operations; hence all the 6760 * cpu masks are stable and all blatant races in the below code cannot 6761 * happen. 6762 */ 6763 mutex_lock(&sched_domains_mutex); 6764 init_sched_domains(cpu_active_mask); 6765 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map); 6766 if (cpumask_empty(non_isolated_cpus)) 6767 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus); 6768 mutex_unlock(&sched_domains_mutex); 6769 6770 hotcpu_notifier(sched_domains_numa_masks_update, CPU_PRI_SCHED_ACTIVE); 6771 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE); 6772 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE); 6773 6774 init_hrtick(); 6775 6776 /* Move init over to a non-isolated CPU */ 6777 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0) 6778 BUG(); 6779 sched_init_granularity(); 6780 free_cpumask_var(non_isolated_cpus); 6781 6782 init_sched_rt_class(); 6783 init_sched_dl_class(); 6784} 6785#else 6786void __init sched_init_smp(void) 6787{ 6788 sched_init_granularity(); 6789} 6790#endif /* CONFIG_SMP */ 6791 6792const_debug unsigned int sysctl_timer_migration = 1; 6793 6794int in_sched_functions(unsigned long addr) 6795{ 6796 return in_lock_functions(addr) || 6797 (addr >= (unsigned long)__sched_text_start 6798 && addr < (unsigned long)__sched_text_end); 6799} 6800 6801#ifdef CONFIG_CGROUP_SCHED 6802/* 6803 * Default task group. 6804 * Every task in system belongs to this group at bootup. 6805 */ 6806struct task_group root_task_group; 6807LIST_HEAD(task_groups); 6808#endif 6809 6810DECLARE_PER_CPU(cpumask_var_t, load_balance_mask); 6811 6812void __init sched_init(void) 6813{ 6814 int i, j; 6815 unsigned long alloc_size = 0, ptr; 6816 6817#ifdef CONFIG_FAIR_GROUP_SCHED 6818 alloc_size += 2 * nr_cpu_ids * sizeof(void **); 6819#endif 6820#ifdef CONFIG_RT_GROUP_SCHED 6821 alloc_size += 2 * nr_cpu_ids * sizeof(void **); 6822#endif 6823#ifdef CONFIG_CPUMASK_OFFSTACK 6824 alloc_size += num_possible_cpus() * cpumask_size(); 6825#endif 6826 if (alloc_size) { 6827 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT); 6828 6829#ifdef CONFIG_FAIR_GROUP_SCHED 6830 root_task_group.se = (struct sched_entity **)ptr; 6831 ptr += nr_cpu_ids * sizeof(void **); 6832 6833 root_task_group.cfs_rq = (struct cfs_rq **)ptr; 6834 ptr += nr_cpu_ids * sizeof(void **); 6835 6836#endif /* CONFIG_FAIR_GROUP_SCHED */ 6837#ifdef CONFIG_RT_GROUP_SCHED 6838 root_task_group.rt_se = (struct sched_rt_entity **)ptr; 6839 ptr += nr_cpu_ids * sizeof(void **); 6840 6841 root_task_group.rt_rq = (struct rt_rq **)ptr; 6842 ptr += nr_cpu_ids * sizeof(void **); 6843 6844#endif /* CONFIG_RT_GROUP_SCHED */ 6845#ifdef CONFIG_CPUMASK_OFFSTACK 6846 for_each_possible_cpu(i) { 6847 per_cpu(load_balance_mask, i) = (void *)ptr; 6848 ptr += cpumask_size(); 6849 } 6850#endif /* CONFIG_CPUMASK_OFFSTACK */ 6851 } 6852 6853 init_rt_bandwidth(&def_rt_bandwidth, 6854 global_rt_period(), global_rt_runtime()); 6855 init_dl_bandwidth(&def_dl_bandwidth, 6856 global_rt_period(), global_rt_runtime()); 6857 6858#ifdef CONFIG_SMP 6859 init_defrootdomain(); 6860#endif 6861 6862#ifdef CONFIG_RT_GROUP_SCHED 6863 init_rt_bandwidth(&root_task_group.rt_bandwidth, 6864 global_rt_period(), global_rt_runtime()); 6865#endif /* CONFIG_RT_GROUP_SCHED */ 6866 6867#ifdef CONFIG_CGROUP_SCHED 6868 list_add(&root_task_group.list, &task_groups); 6869 INIT_LIST_HEAD(&root_task_group.children); 6870 INIT_LIST_HEAD(&root_task_group.siblings); 6871 autogroup_init(&init_task); 6872 6873#endif /* CONFIG_CGROUP_SCHED */ 6874 6875 for_each_possible_cpu(i) { 6876 struct rq *rq; 6877 6878 rq = cpu_rq(i); 6879 raw_spin_lock_init(&rq->lock); 6880 rq->nr_running = 0; 6881 rq->calc_load_active = 0; 6882 rq->calc_load_update = jiffies + LOAD_FREQ; 6883 init_cfs_rq(&rq->cfs); 6884 init_rt_rq(&rq->rt, rq); 6885 init_dl_rq(&rq->dl, rq); 6886#ifdef CONFIG_FAIR_GROUP_SCHED 6887 root_task_group.shares = ROOT_TASK_GROUP_LOAD; 6888 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list); 6889 /* 6890 * How much cpu bandwidth does root_task_group get? 6891 * 6892 * In case of task-groups formed thr' the cgroup filesystem, it 6893 * gets 100% of the cpu resources in the system. This overall 6894 * system cpu resource is divided among the tasks of 6895 * root_task_group and its child task-groups in a fair manner, 6896 * based on each entity's (task or task-group's) weight 6897 * (se->load.weight). 6898 * 6899 * In other words, if root_task_group has 10 tasks of weight 6900 * 1024) and two child groups A0 and A1 (of weight 1024 each), 6901 * then A0's share of the cpu resource is: 6902 * 6903 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33% 6904 * 6905 * We achieve this by letting root_task_group's tasks sit 6906 * directly in rq->cfs (i.e root_task_group->se[] = NULL). 6907 */ 6908 init_cfs_bandwidth(&root_task_group.cfs_bandwidth); 6909 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL); 6910#endif /* CONFIG_FAIR_GROUP_SCHED */ 6911 6912 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime; 6913#ifdef CONFIG_RT_GROUP_SCHED 6914 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL); 6915#endif 6916 6917 for (j = 0; j < CPU_LOAD_IDX_MAX; j++) 6918 rq->cpu_load[j] = 0; 6919 6920 rq->last_load_update_tick = jiffies; 6921 6922#ifdef CONFIG_SMP 6923 rq->sd = NULL; 6924 rq->rd = NULL; 6925 rq->cpu_power = SCHED_POWER_SCALE; 6926 rq->post_schedule = 0; 6927 rq->active_balance = 0; 6928 rq->next_balance = jiffies; 6929 rq->push_cpu = 0; 6930 rq->cpu = i; 6931 rq->online = 0; 6932 rq->idle_stamp = 0; 6933 rq->avg_idle = 2*sysctl_sched_migration_cost; 6934 rq->max_idle_balance_cost = sysctl_sched_migration_cost; 6935 6936 INIT_LIST_HEAD(&rq->cfs_tasks); 6937 6938 rq_attach_root(rq, &def_root_domain); 6939#ifdef CONFIG_NO_HZ_COMMON 6940 rq->nohz_flags = 0; 6941#endif 6942#ifdef CONFIG_NO_HZ_FULL 6943 rq->last_sched_tick = 0; 6944#endif 6945#endif 6946 init_rq_hrtick(rq); 6947 atomic_set(&rq->nr_iowait, 0); 6948 } 6949 6950 set_load_weight(&init_task); 6951 6952#ifdef CONFIG_PREEMPT_NOTIFIERS 6953 INIT_HLIST_HEAD(&init_task.preempt_notifiers); 6954#endif 6955 6956 /* 6957 * The boot idle thread does lazy MMU switching as well: 6958 */ 6959 atomic_inc(&init_mm.mm_count); 6960 enter_lazy_tlb(&init_mm, current); 6961 6962 /* 6963 * Make us the idle thread. Technically, schedule() should not be 6964 * called from this thread, however somewhere below it might be, 6965 * but because we are the idle thread, we just pick up running again 6966 * when this runqueue becomes "idle". 6967 */ 6968 init_idle(current, smp_processor_id()); 6969 6970 calc_load_update = jiffies + LOAD_FREQ; 6971 6972 /* 6973 * During early bootup we pretend to be a normal task: 6974 */ 6975 current->sched_class = &fair_sched_class; 6976 6977#ifdef CONFIG_SMP 6978 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT); 6979 /* May be allocated at isolcpus cmdline parse time */ 6980 if (cpu_isolated_map == NULL) 6981 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT); 6982 idle_thread_set_boot_cpu(); 6983 set_cpu_rq_start_time(); 6984#endif 6985 init_sched_fair_class(); 6986 6987 scheduler_running = 1; 6988} 6989 6990#ifdef CONFIG_DEBUG_ATOMIC_SLEEP 6991static inline int preempt_count_equals(int preempt_offset) 6992{ 6993 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth(); 6994 6995 return (nested == preempt_offset); 6996} 6997 6998void __might_sleep(const char *file, int line, int preempt_offset) 6999{ 7000 static unsigned long prev_jiffy; /* ratelimiting */ 7001 7002 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */ 7003 if ((preempt_count_equals(preempt_offset) && !irqs_disabled() && 7004 !is_idle_task(current)) || 7005 system_state != SYSTEM_RUNNING || oops_in_progress) 7006 return; 7007 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy) 7008 return; 7009 prev_jiffy = jiffies; 7010 7011 printk(KERN_ERR 7012 "BUG: sleeping function called from invalid context at %s:%d\n", 7013 file, line); 7014 printk(KERN_ERR 7015 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n", 7016 in_atomic(), irqs_disabled(), 7017 current->pid, current->comm); 7018 7019 debug_show_held_locks(current); 7020 if (irqs_disabled()) 7021 print_irqtrace_events(current); 7022#ifdef CONFIG_DEBUG_PREEMPT 7023 if (!preempt_count_equals(preempt_offset)) { 7024 pr_err("Preemption disabled at:"); 7025 print_ip_sym(current->preempt_disable_ip); 7026 pr_cont("\n"); 7027 } 7028#endif 7029 dump_stack(); 7030} 7031EXPORT_SYMBOL(__might_sleep); 7032#endif 7033 7034#ifdef CONFIG_MAGIC_SYSRQ 7035static void normalize_task(struct rq *rq, struct task_struct *p) 7036{ 7037 const struct sched_class *prev_class = p->sched_class; 7038 struct sched_attr attr = { 7039 .sched_policy = SCHED_NORMAL, 7040 }; 7041 int old_prio = p->prio; 7042 int on_rq; 7043 7044 on_rq = p->on_rq; 7045 if (on_rq) 7046 dequeue_task(rq, p, 0); 7047 __setscheduler(rq, p, &attr); 7048 if (on_rq) { 7049 enqueue_task(rq, p, 0); 7050 resched_task(rq->curr); 7051 } 7052 7053 check_class_changed(rq, p, prev_class, old_prio); 7054} 7055 7056void normalize_rt_tasks(void) 7057{ 7058 struct task_struct *g, *p; 7059 unsigned long flags; 7060 struct rq *rq; 7061 7062 read_lock_irqsave(&tasklist_lock, flags); 7063 do_each_thread(g, p) { 7064 /* 7065 * Only normalize user tasks: 7066 */ 7067 if (!p->mm) 7068 continue; 7069 7070 p->se.exec_start = 0; 7071#ifdef CONFIG_SCHEDSTATS 7072 p->se.statistics.wait_start = 0; 7073 p->se.statistics.sleep_start = 0; 7074 p->se.statistics.block_start = 0; 7075#endif 7076 7077 if (!dl_task(p) && !rt_task(p)) { 7078 /* 7079 * Renice negative nice level userspace 7080 * tasks back to 0: 7081 */ 7082 if (task_nice(p) < 0 && p->mm) 7083 set_user_nice(p, 0); 7084 continue; 7085 } 7086 7087 raw_spin_lock(&p->pi_lock); 7088 rq = __task_rq_lock(p); 7089 7090 normalize_task(rq, p); 7091 7092 __task_rq_unlock(rq); 7093 raw_spin_unlock(&p->pi_lock); 7094 } while_each_thread(g, p); 7095 7096 read_unlock_irqrestore(&tasklist_lock, flags); 7097} 7098 7099#endif /* CONFIG_MAGIC_SYSRQ */ 7100 7101#if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) 7102/* 7103 * These functions are only useful for the IA64 MCA handling, or kdb. 7104 * 7105 * They can only be called when the whole system has been 7106 * stopped - every CPU needs to be quiescent, and no scheduling 7107 * activity can take place. Using them for anything else would 7108 * be a serious bug, and as a result, they aren't even visible 7109 * under any other configuration. 7110 */ 7111 7112/** 7113 * curr_task - return the current task for a given cpu. 7114 * @cpu: the processor in question. 7115 * 7116 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED! 7117 * 7118 * Return: The current task for @cpu. 7119 */ 7120struct task_struct *curr_task(int cpu) 7121{ 7122 return cpu_curr(cpu); 7123} 7124 7125#endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */ 7126 7127#ifdef CONFIG_IA64 7128/** 7129 * set_curr_task - set the current task for a given cpu. 7130 * @cpu: the processor in question. 7131 * @p: the task pointer to set. 7132 * 7133 * Description: This function must only be used when non-maskable interrupts 7134 * are serviced on a separate stack. It allows the architecture to switch the 7135 * notion of the current task on a cpu in a non-blocking manner. This function 7136 * must be called with all CPU's synchronized, and interrupts disabled, the 7137 * and caller must save the original value of the current task (see 7138 * curr_task() above) and restore that value before reenabling interrupts and 7139 * re-starting the system. 7140 * 7141 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED! 7142 */ 7143void set_curr_task(int cpu, struct task_struct *p) 7144{ 7145 cpu_curr(cpu) = p; 7146} 7147 7148#endif 7149 7150#ifdef CONFIG_CGROUP_SCHED 7151/* task_group_lock serializes the addition/removal of task groups */ 7152static DEFINE_SPINLOCK(task_group_lock); 7153 7154static void free_sched_group(struct task_group *tg) 7155{ 7156 free_fair_sched_group(tg); 7157 free_rt_sched_group(tg); 7158 autogroup_free(tg); 7159 kfree(tg); 7160} 7161 7162/* allocate runqueue etc for a new task group */ 7163struct task_group *sched_create_group(struct task_group *parent) 7164{ 7165 struct task_group *tg; 7166 7167 tg = kzalloc(sizeof(*tg), GFP_KERNEL); 7168 if (!tg) 7169 return ERR_PTR(-ENOMEM); 7170 7171 if (!alloc_fair_sched_group(tg, parent)) 7172 goto err; 7173 7174 if (!alloc_rt_sched_group(tg, parent)) 7175 goto err; 7176 7177 return tg; 7178 7179err: 7180 free_sched_group(tg); 7181 return ERR_PTR(-ENOMEM); 7182} 7183 7184void sched_online_group(struct task_group *tg, struct task_group *parent) 7185{ 7186 unsigned long flags; 7187 7188 spin_lock_irqsave(&task_group_lock, flags); 7189 list_add_rcu(&tg->list, &task_groups); 7190 7191 WARN_ON(!parent); /* root should already exist */ 7192 7193 tg->parent = parent; 7194 INIT_LIST_HEAD(&tg->children); 7195 list_add_rcu(&tg->siblings, &parent->children); 7196 spin_unlock_irqrestore(&task_group_lock, flags); 7197} 7198 7199/* rcu callback to free various structures associated with a task group */ 7200static void free_sched_group_rcu(struct rcu_head *rhp) 7201{ 7202 /* now it should be safe to free those cfs_rqs */ 7203 free_sched_group(container_of(rhp, struct task_group, rcu)); 7204} 7205 7206/* Destroy runqueue etc associated with a task group */ 7207void sched_destroy_group(struct task_group *tg) 7208{ 7209 /* wait for possible concurrent references to cfs_rqs complete */ 7210 call_rcu(&tg->rcu, free_sched_group_rcu); 7211} 7212 7213void sched_offline_group(struct task_group *tg) 7214{ 7215 unsigned long flags; 7216 int i; 7217 7218 /* end participation in shares distribution */ 7219 for_each_possible_cpu(i) 7220 unregister_fair_sched_group(tg, i); 7221 7222 spin_lock_irqsave(&task_group_lock, flags); 7223 list_del_rcu(&tg->list); 7224 list_del_rcu(&tg->siblings); 7225 spin_unlock_irqrestore(&task_group_lock, flags); 7226} 7227 7228/* change task's runqueue when it moves between groups. 7229 * The caller of this function should have put the task in its new group 7230 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to 7231 * reflect its new group. 7232 */ 7233void sched_move_task(struct task_struct *tsk) 7234{ 7235 struct task_group *tg; 7236 int on_rq, running; 7237 unsigned long flags; 7238 struct rq *rq; 7239 7240 rq = task_rq_lock(tsk, &flags); 7241 7242 running = task_current(rq, tsk); 7243 on_rq = tsk->on_rq; 7244 7245 if (on_rq) 7246 dequeue_task(rq, tsk, 0); 7247 if (unlikely(running)) 7248 tsk->sched_class->put_prev_task(rq, tsk); 7249 7250 tg = container_of(task_css_check(tsk, cpu_cgrp_id, 7251 lockdep_is_held(&tsk->sighand->siglock)), 7252 struct task_group, css); 7253 tg = autogroup_task_group(tsk, tg); 7254 tsk->sched_task_group = tg; 7255 7256#ifdef CONFIG_FAIR_GROUP_SCHED 7257 if (tsk->sched_class->task_move_group) 7258 tsk->sched_class->task_move_group(tsk, on_rq); 7259 else 7260#endif 7261 set_task_rq(tsk, task_cpu(tsk)); 7262 7263 if (unlikely(running)) 7264 tsk->sched_class->set_curr_task(rq); 7265 if (on_rq) 7266 enqueue_task(rq, tsk, 0); 7267 7268 task_rq_unlock(rq, tsk, &flags); 7269} 7270#endif /* CONFIG_CGROUP_SCHED */ 7271 7272#ifdef CONFIG_RT_GROUP_SCHED 7273/* 7274 * Ensure that the real time constraints are schedulable. 7275 */ 7276static DEFINE_MUTEX(rt_constraints_mutex); 7277 7278/* Must be called with tasklist_lock held */ 7279static inline int tg_has_rt_tasks(struct task_group *tg) 7280{ 7281 struct task_struct *g, *p; 7282 7283 do_each_thread(g, p) { 7284 if (rt_task(p) && task_rq(p)->rt.tg == tg) 7285 return 1; 7286 } while_each_thread(g, p); 7287 7288 return 0; 7289} 7290 7291struct rt_schedulable_data { 7292 struct task_group *tg; 7293 u64 rt_period; 7294 u64 rt_runtime; 7295}; 7296 7297static int tg_rt_schedulable(struct task_group *tg, void *data) 7298{ 7299 struct rt_schedulable_data *d = data; 7300 struct task_group *child; 7301 unsigned long total, sum = 0; 7302 u64 period, runtime; 7303 7304 period = ktime_to_ns(tg->rt_bandwidth.rt_period); 7305 runtime = tg->rt_bandwidth.rt_runtime; 7306 7307 if (tg == d->tg) { 7308 period = d->rt_period; 7309 runtime = d->rt_runtime; 7310 } 7311 7312 /* 7313 * Cannot have more runtime than the period. 7314 */ 7315 if (runtime > period && runtime != RUNTIME_INF) 7316 return -EINVAL; 7317 7318 /* 7319 * Ensure we don't starve existing RT tasks. 7320 */ 7321 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg)) 7322 return -EBUSY; 7323 7324 total = to_ratio(period, runtime); 7325 7326 /* 7327 * Nobody can have more than the global setting allows. 7328 */ 7329 if (total > to_ratio(global_rt_period(), global_rt_runtime())) 7330 return -EINVAL; 7331 7332 /* 7333 * The sum of our children's runtime should not exceed our own. 7334 */ 7335 list_for_each_entry_rcu(child, &tg->children, siblings) { 7336 period = ktime_to_ns(child->rt_bandwidth.rt_period); 7337 runtime = child->rt_bandwidth.rt_runtime; 7338 7339 if (child == d->tg) { 7340 period = d->rt_period; 7341 runtime = d->rt_runtime; 7342 } 7343 7344 sum += to_ratio(period, runtime); 7345 } 7346 7347 if (sum > total) 7348 return -EINVAL; 7349 7350 return 0; 7351} 7352 7353static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime) 7354{ 7355 int ret; 7356 7357 struct rt_schedulable_data data = { 7358 .tg = tg, 7359 .rt_period = period, 7360 .rt_runtime = runtime, 7361 }; 7362 7363 rcu_read_lock(); 7364 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data); 7365 rcu_read_unlock(); 7366 7367 return ret; 7368} 7369 7370static int tg_set_rt_bandwidth(struct task_group *tg, 7371 u64 rt_period, u64 rt_runtime) 7372{ 7373 int i, err = 0; 7374 7375 mutex_lock(&rt_constraints_mutex); 7376 read_lock(&tasklist_lock); 7377 err = __rt_schedulable(tg, rt_period, rt_runtime); 7378 if (err) 7379 goto unlock; 7380 7381 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock); 7382 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period); 7383 tg->rt_bandwidth.rt_runtime = rt_runtime; 7384 7385 for_each_possible_cpu(i) { 7386 struct rt_rq *rt_rq = tg->rt_rq[i]; 7387 7388 raw_spin_lock(&rt_rq->rt_runtime_lock); 7389 rt_rq->rt_runtime = rt_runtime; 7390 raw_spin_unlock(&rt_rq->rt_runtime_lock); 7391 } 7392 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock); 7393unlock: 7394 read_unlock(&tasklist_lock); 7395 mutex_unlock(&rt_constraints_mutex); 7396 7397 return err; 7398} 7399 7400static int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us) 7401{ 7402 u64 rt_runtime, rt_period; 7403 7404 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period); 7405 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC; 7406 if (rt_runtime_us < 0) 7407 rt_runtime = RUNTIME_INF; 7408 7409 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime); 7410} 7411 7412static long sched_group_rt_runtime(struct task_group *tg) 7413{ 7414 u64 rt_runtime_us; 7415 7416 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF) 7417 return -1; 7418 7419 rt_runtime_us = tg->rt_bandwidth.rt_runtime; 7420 do_div(rt_runtime_us, NSEC_PER_USEC); 7421 return rt_runtime_us; 7422} 7423 7424static int sched_group_set_rt_period(struct task_group *tg, long rt_period_us) 7425{ 7426 u64 rt_runtime, rt_period; 7427 7428 rt_period = (u64)rt_period_us * NSEC_PER_USEC; 7429 rt_runtime = tg->rt_bandwidth.rt_runtime; 7430 7431 if (rt_period == 0) 7432 return -EINVAL; 7433 7434 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime); 7435} 7436 7437static long sched_group_rt_period(struct task_group *tg) 7438{ 7439 u64 rt_period_us; 7440 7441 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period); 7442 do_div(rt_period_us, NSEC_PER_USEC); 7443 return rt_period_us; 7444} 7445#endif /* CONFIG_RT_GROUP_SCHED */ 7446 7447#ifdef CONFIG_RT_GROUP_SCHED 7448static int sched_rt_global_constraints(void) 7449{ 7450 int ret = 0; 7451 7452 mutex_lock(&rt_constraints_mutex); 7453 read_lock(&tasklist_lock); 7454 ret = __rt_schedulable(NULL, 0, 0); 7455 read_unlock(&tasklist_lock); 7456 mutex_unlock(&rt_constraints_mutex); 7457 7458 return ret; 7459} 7460 7461static int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk) 7462{ 7463 /* Don't accept realtime tasks when there is no way for them to run */ 7464 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0) 7465 return 0; 7466 7467 return 1; 7468} 7469 7470#else /* !CONFIG_RT_GROUP_SCHED */ 7471static int sched_rt_global_constraints(void) 7472{ 7473 unsigned long flags; 7474 int i, ret = 0; 7475 7476 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags); 7477 for_each_possible_cpu(i) { 7478 struct rt_rq *rt_rq = &cpu_rq(i)->rt; 7479 7480 raw_spin_lock(&rt_rq->rt_runtime_lock); 7481 rt_rq->rt_runtime = global_rt_runtime(); 7482 raw_spin_unlock(&rt_rq->rt_runtime_lock); 7483 } 7484 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags); 7485 7486 return ret; 7487} 7488#endif /* CONFIG_RT_GROUP_SCHED */ 7489 7490static int sched_dl_global_constraints(void) 7491{ 7492 u64 runtime = global_rt_runtime(); 7493 u64 period = global_rt_period(); 7494 u64 new_bw = to_ratio(period, runtime); 7495 int cpu, ret = 0; 7496 unsigned long flags; 7497 7498 /* 7499 * Here we want to check the bandwidth not being set to some 7500 * value smaller than the currently allocated bandwidth in 7501 * any of the root_domains. 7502 * 7503 * FIXME: Cycling on all the CPUs is overdoing, but simpler than 7504 * cycling on root_domains... Discussion on different/better 7505 * solutions is welcome! 7506 */ 7507 for_each_possible_cpu(cpu) { 7508 struct dl_bw *dl_b = dl_bw_of(cpu); 7509 7510 raw_spin_lock_irqsave(&dl_b->lock, flags); 7511 if (new_bw < dl_b->total_bw) 7512 ret = -EBUSY; 7513 raw_spin_unlock_irqrestore(&dl_b->lock, flags); 7514 7515 if (ret) 7516 break; 7517 } 7518 7519 return ret; 7520} 7521 7522static void sched_dl_do_global(void) 7523{ 7524 u64 new_bw = -1; 7525 int cpu; 7526 unsigned long flags; 7527 7528 def_dl_bandwidth.dl_period = global_rt_period(); 7529 def_dl_bandwidth.dl_runtime = global_rt_runtime(); 7530 7531 if (global_rt_runtime() != RUNTIME_INF) 7532 new_bw = to_ratio(global_rt_period(), global_rt_runtime()); 7533 7534 /* 7535 * FIXME: As above... 7536 */ 7537 for_each_possible_cpu(cpu) { 7538 struct dl_bw *dl_b = dl_bw_of(cpu); 7539 7540 raw_spin_lock_irqsave(&dl_b->lock, flags); 7541 dl_b->bw = new_bw; 7542 raw_spin_unlock_irqrestore(&dl_b->lock, flags); 7543 } 7544} 7545 7546static int sched_rt_global_validate(void) 7547{ 7548 if (sysctl_sched_rt_period <= 0) 7549 return -EINVAL; 7550 7551 if ((sysctl_sched_rt_runtime != RUNTIME_INF) && 7552 (sysctl_sched_rt_runtime > sysctl_sched_rt_period)) 7553 return -EINVAL; 7554 7555 return 0; 7556} 7557 7558static void sched_rt_do_global(void) 7559{ 7560 def_rt_bandwidth.rt_runtime = global_rt_runtime(); 7561 def_rt_bandwidth.rt_period = ns_to_ktime(global_rt_period()); 7562} 7563 7564int sched_rt_handler(struct ctl_table *table, int write, 7565 void __user *buffer, size_t *lenp, 7566 loff_t *ppos) 7567{ 7568 int old_period, old_runtime; 7569 static DEFINE_MUTEX(mutex); 7570 int ret; 7571 7572 mutex_lock(&mutex); 7573 old_period = sysctl_sched_rt_period; 7574 old_runtime = sysctl_sched_rt_runtime; 7575 7576 ret = proc_dointvec(table, write, buffer, lenp, ppos); 7577 7578 if (!ret && write) { 7579 ret = sched_rt_global_validate(); 7580 if (ret) 7581 goto undo; 7582 7583 ret = sched_rt_global_constraints(); 7584 if (ret) 7585 goto undo; 7586 7587 ret = sched_dl_global_constraints(); 7588 if (ret) 7589 goto undo; 7590 7591 sched_rt_do_global(); 7592 sched_dl_do_global(); 7593 } 7594 if (0) { 7595undo: 7596 sysctl_sched_rt_period = old_period; 7597 sysctl_sched_rt_runtime = old_runtime; 7598 } 7599 mutex_unlock(&mutex); 7600 7601 return ret; 7602} 7603 7604int sched_rr_handler(struct ctl_table *table, int write, 7605 void __user *buffer, size_t *lenp, 7606 loff_t *ppos) 7607{ 7608 int ret; 7609 static DEFINE_MUTEX(mutex); 7610 7611 mutex_lock(&mutex); 7612 ret = proc_dointvec(table, write, buffer, lenp, ppos); 7613 /* make sure that internally we keep jiffies */ 7614 /* also, writing zero resets timeslice to default */ 7615 if (!ret && write) { 7616 sched_rr_timeslice = sched_rr_timeslice <= 0 ? 7617 RR_TIMESLICE : msecs_to_jiffies(sched_rr_timeslice); 7618 } 7619 mutex_unlock(&mutex); 7620 return ret; 7621} 7622 7623#ifdef CONFIG_CGROUP_SCHED 7624 7625static inline struct task_group *css_tg(struct cgroup_subsys_state *css) 7626{ 7627 return css ? container_of(css, struct task_group, css) : NULL; 7628} 7629 7630static struct cgroup_subsys_state * 7631cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css) 7632{ 7633 struct task_group *parent = css_tg(parent_css); 7634 struct task_group *tg; 7635 7636 if (!parent) { 7637 /* This is early initialization for the top cgroup */ 7638 return &root_task_group.css; 7639 } 7640 7641 tg = sched_create_group(parent); 7642 if (IS_ERR(tg)) 7643 return ERR_PTR(-ENOMEM); 7644 7645 return &tg->css; 7646} 7647 7648static int cpu_cgroup_css_online(struct cgroup_subsys_state *css) 7649{ 7650 struct task_group *tg = css_tg(css); 7651 struct task_group *parent = css_tg(css_parent(css)); 7652 7653 if (parent) 7654 sched_online_group(tg, parent); 7655 return 0; 7656} 7657 7658static void cpu_cgroup_css_free(struct cgroup_subsys_state *css) 7659{ 7660 struct task_group *tg = css_tg(css); 7661 7662 sched_destroy_group(tg); 7663} 7664 7665static void cpu_cgroup_css_offline(struct cgroup_subsys_state *css) 7666{ 7667 struct task_group *tg = css_tg(css); 7668 7669 sched_offline_group(tg); 7670} 7671 7672static int cpu_cgroup_can_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#ifdef CONFIG_RT_GROUP_SCHED 7679 if (!sched_rt_can_attach(css_tg(css), task)) 7680 return -EINVAL; 7681#else 7682 /* We don't support RT-tasks being in separate groups */ 7683 if (task->sched_class != &fair_sched_class) 7684 return -EINVAL; 7685#endif 7686 } 7687 return 0; 7688} 7689 7690static void cpu_cgroup_attach(struct cgroup_subsys_state *css, 7691 struct cgroup_taskset *tset) 7692{ 7693 struct task_struct *task; 7694 7695 cgroup_taskset_for_each(task, tset) 7696 sched_move_task(task); 7697} 7698 7699static void cpu_cgroup_exit(struct cgroup_subsys_state *css, 7700 struct cgroup_subsys_state *old_css, 7701 struct task_struct *task) 7702{ 7703 /* 7704 * cgroup_exit() is called in the copy_process() failure path. 7705 * Ignore this case since the task hasn't ran yet, this avoids 7706 * trying to poke a half freed task state from generic code. 7707 */ 7708 if (!(task->flags & PF_EXITING)) 7709 return; 7710 7711 sched_move_task(task); 7712} 7713 7714#ifdef CONFIG_FAIR_GROUP_SCHED 7715static int cpu_shares_write_u64(struct cgroup_subsys_state *css, 7716 struct cftype *cftype, u64 shareval) 7717{ 7718 return sched_group_set_shares(css_tg(css), scale_load(shareval)); 7719} 7720 7721static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css, 7722 struct cftype *cft) 7723{ 7724 struct task_group *tg = css_tg(css); 7725 7726 return (u64) scale_load_down(tg->shares); 7727} 7728 7729#ifdef CONFIG_CFS_BANDWIDTH 7730static DEFINE_MUTEX(cfs_constraints_mutex); 7731 7732const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */ 7733const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */ 7734 7735static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime); 7736 7737static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota) 7738{ 7739 int i, ret = 0, runtime_enabled, runtime_was_enabled; 7740 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth; 7741 7742 if (tg == &root_task_group) 7743 return -EINVAL; 7744 7745 /* 7746 * Ensure we have at some amount of bandwidth every period. This is 7747 * to prevent reaching a state of large arrears when throttled via 7748 * entity_tick() resulting in prolonged exit starvation. 7749 */ 7750 if (quota < min_cfs_quota_period || period < min_cfs_quota_period) 7751 return -EINVAL; 7752 7753 /* 7754 * Likewise, bound things on the otherside by preventing insane quota 7755 * periods. This also allows us to normalize in computing quota 7756 * feasibility. 7757 */ 7758 if (period > max_cfs_quota_period) 7759 return -EINVAL; 7760 7761 mutex_lock(&cfs_constraints_mutex); 7762 ret = __cfs_schedulable(tg, period, quota); 7763 if (ret) 7764 goto out_unlock; 7765 7766 runtime_enabled = quota != RUNTIME_INF; 7767 runtime_was_enabled = cfs_b->quota != RUNTIME_INF; 7768 /* 7769 * If we need to toggle cfs_bandwidth_used, off->on must occur 7770 * before making related changes, and on->off must occur afterwards 7771 */ 7772 if (runtime_enabled && !runtime_was_enabled) 7773 cfs_bandwidth_usage_inc(); 7774 raw_spin_lock_irq(&cfs_b->lock); 7775 cfs_b->period = ns_to_ktime(period); 7776 cfs_b->quota = quota; 7777 7778 __refill_cfs_bandwidth_runtime(cfs_b); 7779 /* restart the period timer (if active) to handle new period expiry */ 7780 if (runtime_enabled && cfs_b->timer_active) { 7781 /* force a reprogram */ 7782 cfs_b->timer_active = 0; 7783 __start_cfs_bandwidth(cfs_b); 7784 } 7785 raw_spin_unlock_irq(&cfs_b->lock); 7786 7787 for_each_possible_cpu(i) { 7788 struct cfs_rq *cfs_rq = tg->cfs_rq[i]; 7789 struct rq *rq = cfs_rq->rq; 7790 7791 raw_spin_lock_irq(&rq->lock); 7792 cfs_rq->runtime_enabled = runtime_enabled; 7793 cfs_rq->runtime_remaining = 0; 7794 7795 if (cfs_rq->throttled) 7796 unthrottle_cfs_rq(cfs_rq); 7797 raw_spin_unlock_irq(&rq->lock); 7798 } 7799 if (runtime_was_enabled && !runtime_enabled) 7800 cfs_bandwidth_usage_dec(); 7801out_unlock: 7802 mutex_unlock(&cfs_constraints_mutex); 7803 7804 return ret; 7805} 7806 7807int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us) 7808{ 7809 u64 quota, period; 7810 7811 period = ktime_to_ns(tg->cfs_bandwidth.period); 7812 if (cfs_quota_us < 0) 7813 quota = RUNTIME_INF; 7814 else 7815 quota = (u64)cfs_quota_us * NSEC_PER_USEC; 7816 7817 return tg_set_cfs_bandwidth(tg, period, quota); 7818} 7819 7820long tg_get_cfs_quota(struct task_group *tg) 7821{ 7822 u64 quota_us; 7823 7824 if (tg->cfs_bandwidth.quota == RUNTIME_INF) 7825 return -1; 7826 7827 quota_us = tg->cfs_bandwidth.quota; 7828 do_div(quota_us, NSEC_PER_USEC); 7829 7830 return quota_us; 7831} 7832 7833int tg_set_cfs_period(struct task_group *tg, long cfs_period_us) 7834{ 7835 u64 quota, period; 7836 7837 period = (u64)cfs_period_us * NSEC_PER_USEC; 7838 quota = tg->cfs_bandwidth.quota; 7839 7840 return tg_set_cfs_bandwidth(tg, period, quota); 7841} 7842 7843long tg_get_cfs_period(struct task_group *tg) 7844{ 7845 u64 cfs_period_us; 7846 7847 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period); 7848 do_div(cfs_period_us, NSEC_PER_USEC); 7849 7850 return cfs_period_us; 7851} 7852 7853static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css, 7854 struct cftype *cft) 7855{ 7856 return tg_get_cfs_quota(css_tg(css)); 7857} 7858 7859static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css, 7860 struct cftype *cftype, s64 cfs_quota_us) 7861{ 7862 return tg_set_cfs_quota(css_tg(css), cfs_quota_us); 7863} 7864 7865static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css, 7866 struct cftype *cft) 7867{ 7868 return tg_get_cfs_period(css_tg(css)); 7869} 7870 7871static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css, 7872 struct cftype *cftype, u64 cfs_period_us) 7873{ 7874 return tg_set_cfs_period(css_tg(css), cfs_period_us); 7875} 7876 7877struct cfs_schedulable_data { 7878 struct task_group *tg; 7879 u64 period, quota; 7880}; 7881 7882/* 7883 * normalize group quota/period to be quota/max_period 7884 * note: units are usecs 7885 */ 7886static u64 normalize_cfs_quota(struct task_group *tg, 7887 struct cfs_schedulable_data *d) 7888{ 7889 u64 quota, period; 7890 7891 if (tg == d->tg) { 7892 period = d->period; 7893 quota = d->quota; 7894 } else { 7895 period = tg_get_cfs_period(tg); 7896 quota = tg_get_cfs_quota(tg); 7897 } 7898 7899 /* note: these should typically be equivalent */ 7900 if (quota == RUNTIME_INF || quota == -1) 7901 return RUNTIME_INF; 7902 7903 return to_ratio(period, quota); 7904} 7905 7906static int tg_cfs_schedulable_down(struct task_group *tg, void *data) 7907{ 7908 struct cfs_schedulable_data *d = data; 7909 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth; 7910 s64 quota = 0, parent_quota = -1; 7911 7912 if (!tg->parent) { 7913 quota = RUNTIME_INF; 7914 } else { 7915 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth; 7916 7917 quota = normalize_cfs_quota(tg, d); 7918 parent_quota = parent_b->hierarchal_quota; 7919 7920 /* 7921 * ensure max(child_quota) <= parent_quota, inherit when no 7922 * limit is set 7923 */ 7924 if (quota == RUNTIME_INF) 7925 quota = parent_quota; 7926 else if (parent_quota != RUNTIME_INF && quota > parent_quota) 7927 return -EINVAL; 7928 } 7929 cfs_b->hierarchal_quota = quota; 7930 7931 return 0; 7932} 7933 7934static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota) 7935{ 7936 int ret; 7937 struct cfs_schedulable_data data = { 7938 .tg = tg, 7939 .period = period, 7940 .quota = quota, 7941 }; 7942 7943 if (quota != RUNTIME_INF) { 7944 do_div(data.period, NSEC_PER_USEC); 7945 do_div(data.quota, NSEC_PER_USEC); 7946 } 7947 7948 rcu_read_lock(); 7949 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data); 7950 rcu_read_unlock(); 7951 7952 return ret; 7953} 7954 7955static int cpu_stats_show(struct seq_file *sf, void *v) 7956{ 7957 struct task_group *tg = css_tg(seq_css(sf)); 7958 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth; 7959 7960 seq_printf(sf, "nr_periods %d\n", cfs_b->nr_periods); 7961 seq_printf(sf, "nr_throttled %d\n", cfs_b->nr_throttled); 7962 seq_printf(sf, "throttled_time %llu\n", cfs_b->throttled_time); 7963 7964 return 0; 7965} 7966#endif /* CONFIG_CFS_BANDWIDTH */ 7967#endif /* CONFIG_FAIR_GROUP_SCHED */ 7968 7969#ifdef CONFIG_RT_GROUP_SCHED 7970static int cpu_rt_runtime_write(struct cgroup_subsys_state *css, 7971 struct cftype *cft, s64 val) 7972{ 7973 return sched_group_set_rt_runtime(css_tg(css), val); 7974} 7975 7976static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css, 7977 struct cftype *cft) 7978{ 7979 return sched_group_rt_runtime(css_tg(css)); 7980} 7981 7982static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css, 7983 struct cftype *cftype, u64 rt_period_us) 7984{ 7985 return sched_group_set_rt_period(css_tg(css), rt_period_us); 7986} 7987 7988static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css, 7989 struct cftype *cft) 7990{ 7991 return sched_group_rt_period(css_tg(css)); 7992} 7993#endif /* CONFIG_RT_GROUP_SCHED */ 7994 7995static struct cftype cpu_files[] = { 7996#ifdef CONFIG_FAIR_GROUP_SCHED 7997 { 7998 .name = "shares", 7999 .read_u64 = cpu_shares_read_u64, 8000 .write_u64 = cpu_shares_write_u64, 8001 }, 8002#endif 8003#ifdef CONFIG_CFS_BANDWIDTH 8004 { 8005 .name = "cfs_quota_us", 8006 .read_s64 = cpu_cfs_quota_read_s64, 8007 .write_s64 = cpu_cfs_quota_write_s64, 8008 }, 8009 { 8010 .name = "cfs_period_us", 8011 .read_u64 = cpu_cfs_period_read_u64, 8012 .write_u64 = cpu_cfs_period_write_u64, 8013 }, 8014 { 8015 .name = "stat", 8016 .seq_show = cpu_stats_show, 8017 }, 8018#endif 8019#ifdef CONFIG_RT_GROUP_SCHED 8020 { 8021 .name = "rt_runtime_us", 8022 .read_s64 = cpu_rt_runtime_read, 8023 .write_s64 = cpu_rt_runtime_write, 8024 }, 8025 { 8026 .name = "rt_period_us", 8027 .read_u64 = cpu_rt_period_read_uint, 8028 .write_u64 = cpu_rt_period_write_uint, 8029 }, 8030#endif 8031 { } /* terminate */ 8032}; 8033 8034struct cgroup_subsys cpu_cgrp_subsys = { 8035 .css_alloc = cpu_cgroup_css_alloc, 8036 .css_free = cpu_cgroup_css_free, 8037 .css_online = cpu_cgroup_css_online, 8038 .css_offline = cpu_cgroup_css_offline, 8039 .can_attach = cpu_cgroup_can_attach, 8040 .attach = cpu_cgroup_attach, 8041 .exit = cpu_cgroup_exit, 8042 .base_cftypes = cpu_files, 8043 .early_init = 1, 8044}; 8045 8046#endif /* CONFIG_CGROUP_SCHED */ 8047 8048void dump_cpu_task(int cpu) 8049{ 8050 pr_info("Task dump for CPU %d:\n", cpu); 8051 sched_show_task(cpu_curr(cpu)); 8052} 8053